diff --git a/.gitattributes b/.gitattributes index 907023af39d0664ea3c447a14961373c535f61e6..e94d69a753a29f2b764b9d35fd68a371e462ecb4 100644 --- a/.gitattributes +++ b/.gitattributes @@ -2268,3 +2268,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2025/SpatialTrackerV2_[[:space:]]Advancing[[:space:]]3D[[:space:]]Point[[:space:]]Tracking[[:space:]]with[[:space:]]Explicit[[:space:]]Camera[[:space:]]Motion/925f4fd6-2e60-44c7-a108-57075a837941_origin.pdf filter=lfs diff=lfs merge=lfs -text 2025/Spatially-Varying[[:space:]]Autofocus/812fa83e-886a-4a5e-b27e-3f560550ef19_origin.pdf filter=lfs diff=lfs merge=lfs -text 2025/Spatio-Spectral[[:space:]]Pattern[[:space:]]Illumination[[:space:]]for[[:space:]]Direct[[:space:]]and[[:space:]]Indirect[[:space:]]Separation[[:space:]]from[[:space:]]a[[:space:]]Single[[:space:]]Hyperspectral[[:space:]]Image/c416257e-42cc-4603-837a-a5cb6ea0c724_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SpecGuard_[[:space:]]Spectral[[:space:]]Projection-based[[:space:]]Advanced[[:space:]]Invisible[[:space:]]Watermarking/1fcb8d1e-6633-4c96-8cd9-9ceeae669d64_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Spectral[[:space:]]Image[[:space:]]Tokenizer/2b3586c0-b936-4d11-a38a-2ff8e5859a65_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Spectral[[:space:]]Sensitivity[[:space:]]Estimation[[:space:]]with[[:space:]]an[[:space:]]Uncalibrated[[:space:]]Diffraction[[:space:]]Grating/79b5704f-97ff-445c-93ff-ff5edaa3c620_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SpectralAR_[[:space:]]Spectral[[:space:]]Autoregressive[[:space:]]Visual[[:space:]]Generation/ddfacecc-ae50-4e90-8053-8ccf2e5c55b7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Spherical[[:space:]]Epipolar[[:space:]]Rectification[[:space:]]for[[:space:]]Deep[[:space:]]Two-View[[:space:]]Absolute[[:space:]]Depth[[:space:]]Estimation/580cf2a2-db43-4692-a58d-cadb1e36cc3b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SpiLiFormer_[[:space:]]Enhancing[[:space:]]Spiking[[:space:]]Transformers[[:space:]]with[[:space:]]Lateral[[:space:]]Inhibition/75735d8f-465e-40da-8f9a-a276e0092626_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SpikeDiff_[[:space:]]Zero-shot[[:space:]]High-Quality[[:space:]]Video[[:space:]]Reconstruction[[:space:]]from[[:space:]]Chromatic[[:space:]]Spike[[:space:]]Camera[[:space:]]and[[:space:]]Sub-millisecond[[:space:]]Spike[[:space:]]Streams/3f42c09f-75a5-4dd4-bd4a-8b0a20ea82c7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SpikePack_[[:space:]]Enhanced[[:space:]]Information[[:space:]]Flow[[:space:]]in[[:space:]]Spiking[[:space:]]Neural[[:space:]]Networks[[:space:]]with[[:space:]]High[[:space:]]Hardware[[:space:]]Compatibility/f9abc2f2-d2b5-43b8-9f56-aa36381da767_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SpinMeRound_[[:space:]]Consistent[[:space:]]Multi-View[[:space:]]Identity[[:space:]]Generation[[:space:]]Using[[:space:]]Diffusion[[:space:]]Models/d46ce965-2153-45b4-9566-3f0d4d7d97a6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SplArt_[[:space:]]Articulation[[:space:]]Estimation[[:space:]]and[[:space:]]Part-Level[[:space:]]Reconstruction[[:space:]]with[[:space:]]3D[[:space:]]Gaussian[[:space:]]Splatting/99f05b1c-e917-486f-913a-4812f9c2d857_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Splat-LOAM_[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]LiDAR[[:space:]]Odometry[[:space:]]and[[:space:]]Mapping/21f438cd-a199-46e8-b668-60b759526f59_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Splat-based[[:space:]]3D[[:space:]]Scene[[:space:]]Reconstruction[[:space:]]with[[:space:]]Extreme[[:space:]]Motion-blur/73a5b991-b304-4700-8590-8334fbf2bdee_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SplatTalk_[[:space:]]3D[[:space:]]VQA[[:space:]]with[[:space:]]Gaussian[[:space:]]Splatting/71c4eb8c-9be8-49e8-bab3-c4048f8e7d2a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Split-and-Combine_[[:space:]]Enhancing[[:space:]]Style[[:space:]]Augmentation[[:space:]]for[[:space:]]Single[[:space:]]Domain[[:space:]]Generalization/3714ebba-fcba-4212-aaf3-03100ed83e52_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/St4RTrack_[[:space:]]Simultaneous[[:space:]]4D[[:space:]]Reconstruction[[:space:]]and[[:space:]]Tracking[[:space:]]in[[:space:]]the[[:space:]]World/8996ce18-d088-44bb-a34b-53f80b972348_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Stable[[:space:]]Diffusion[[:space:]]Models[[:space:]]are[[:space:]]Secretly[[:space:]]Good[[:space:]]at[[:space:]]Visual[[:space:]]In-Context[[:space:]]Learning/94d38526-8561-492c-b6db-5bcfedbe9083_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Stable[[:space:]]Score[[:space:]]Distillation/62930789-4f50-45e0-b2ce-8d1494eced86_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Stable[[:space:]]Virtual[[:space:]]Camera_[[:space:]]Generative[[:space:]]View[[:space:]]Synthesis[[:space:]]with[[:space:]]Diffusion[[:space:]]Models/8df1f571-dcd1-4fc3-a74d-5f6f5bf1e485_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Stable-Sim2Real_[[:space:]]Exploring[[:space:]]Simulation[[:space:]]of[[:space:]]Real-Captured[[:space:]]3D[[:space:]]Data[[:space:]]with[[:space:]]Two-Stage[[:space:]]Depth[[:space:]]Diffusion/ac3096b4-f6df-4b0e-b9d1-43e46507be43_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/StableCodec_[[:space:]]Taming[[:space:]]One-Step[[:space:]]Diffusion[[:space:]]for[[:space:]]Extreme[[:space:]]Image[[:space:]]Compression/5269d8ed-d6d6-4961-a311-a0e841365f77_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/StableDepth_[[:space:]]Scene-Consistent[[:space:]]and[[:space:]]Scale-Invariant[[:space:]]Monocular[[:space:]]Depth/73601462-88ee-4d47-a92f-ac7545019662_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Staining[[:space:]]and[[:space:]]Locking[[:space:]]Computer[[:space:]]Vision[[:space:]]Models[[:space:]]Without[[:space:]]Retraining/6d43a5d3-0678-494f-aae6-04c44e76e1c5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Statistical[[:space:]]Confidence[[:space:]]Rescoring[[:space:]]for[[:space:]]Robust[[:space:]]3D[[:space:]]Scene[[:space:]]Graph[[:space:]]Generation[[:space:]]from[[:space:]]Multi-View[[:space:]]Images/00c53575-a21f-463b-8889-7a2370d2bff7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/StealthAttack_[[:space:]]Robust[[:space:]]3D[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]Poisoning[[:space:]]via[[:space:]]Density-Guided[[:space:]]Illusions/b6c26c82-2907-466d-9aed-934611879ad9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Stealthy[[:space:]]Backdoor[[:space:]]Attack[[:space:]]in[[:space:]]Federated[[:space:]]Learning[[:space:]]via[[:space:]]Adaptive[[:space:]]Layer-wise[[:space:]]Gradient[[:space:]]Alignment/f7dfb2e6-c547-4b0c-bb2e-4bfc0b02a71f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SteerX_[[:space:]]Creating[[:space:]]Any[[:space:]]Camera-Free[[:space:]]3D[[:space:]]and[[:space:]]4D[[:space:]]Scenes[[:space:]]with[[:space:]]Geometric[[:space:]]Steering/a06f04c2-df21-4c58-83ac-45d3ea205c44_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Steering[[:space:]]Guidance[[:space:]]for[[:space:]]Personalized[[:space:]]Text-to-Image[[:space:]]Diffusion[[:space:]]Models/34e2c957-5535-42df-a5b2-c4f7cd9d6450_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Stepping[[:space:]]Out[[:space:]]of[[:space:]]Similar[[:space:]]Semantic[[:space:]]Space[[:space:]]for[[:space:]]Open-Vocabulary[[:space:]]Segmentation/f495e889-2c8e-4e64-ba9d-86b1aec5716d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Stereo[[:space:]]Any[[:space:]]Video_[[:space:]]Temporally[[:space:]]Consistent[[:space:]]Stereo[[:space:]]Matching/dc20cc87-36ce-4c69-8c57-afd220728c97_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Stochastic[[:space:]]Gradient[[:space:]]Estimation[[:space:]]for[[:space:]]Higher-Order[[:space:]]Differentiable[[:space:]]Rendering/95d3584a-ba8b-4c05-9f1b-c00fa69cfc88_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Stochastic[[:space:]]Interpolants[[:space:]]for[[:space:]]Revealing[[:space:]]Stylistic[[:space:]]Flows[[:space:]]across[[:space:]]the[[:space:]]History[[:space:]]of[[:space:]]Art/16c23549-7e7f-4e3a-8944-314b6bb00228_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/StochasticSplats_[[:space:]]Stochastic[[:space:]]Rasterization[[:space:]]for[[:space:]]Sorting-Free[[:space:]]3D[[:space:]]Gaussian[[:space:]]Splatting/d963f664-7a0a-43e1-95cb-e52f6169cd94_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/StolenLoRA_[[:space:]]Exploring[[:space:]]LoRA[[:space:]]Extraction[[:space:]]Attacks[[:space:]]via[[:space:]]Synthetic[[:space:]]Data/f4ddd460-9cf1-429e-89c2-3d861bf9c661_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Straighten[[:space:]]Viscous[[:space:]]Rectified[[:space:]]Flow[[:space:]]via[[:space:]]Noise[[:space:]]Optimization/7bcafc85-3856-49c2-a845-9dc4fd667b05_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/StrandHead_[[:space:]]Text[[:space:]]to[[:space:]]Hair-Disentangled[[:space:]]3D[[:space:]]Head[[:space:]]Avatars[[:space:]]Using[[:space:]]Human-Centric[[:space:]]Priors/5c673a06-82fc-4d7f-915c-54daefec29d8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/StreamDiffusion_[[:space:]]A[[:space:]]Pipeline-level[[:space:]]Solution[[:space:]]for[[:space:]]Real-Time[[:space:]]Interactive[[:space:]]Generation/522f37ba-5716-44cf-af98-f93d9030fe6d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/StreamGS_[[:space:]]Online[[:space:]]Generalizable[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]Reconstruction[[:space:]]for[[:space:]]Unposed[[:space:]]Image[[:space:]]Streams/9d1940fe-237f-4495-9085-9441928f0c47_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/StreamMind_[[:space:]]Unlocking[[:space:]]Full[[:space:]]Frame[[:space:]]Rate[[:space:]]Streaming[[:space:]]Video[[:space:]]Dialogue[[:space:]]through[[:space:]]Event-Gated[[:space:]]Cognition/61e82f55-a3d1-4244-b3cc-22632c8cab6f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Streaming[[:space:]]VideoLLMs[[:space:]]for[[:space:]]Real-Time[[:space:]]Procedural[[:space:]]Video[[:space:]]Understanding/e3ae8cdd-650d-48b8-9247-fe7363ebedec_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Streamlining[[:space:]]Image[[:space:]]Editing[[:space:]]with[[:space:]]Layered[[:space:]]Diffusion[[:space:]]Brushes/6116442a-236c-4dff-8b3c-c87a9184cffb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Street[[:space:]]Gaussians[[:space:]]without[[:space:]]3D[[:space:]]Object[[:space:]]Tracker/ceea65cc-c6ed-4695-b497-f6fc77f3a4c6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Stroke2Sketch_[[:space:]]Harnessing[[:space:]]Stroke[[:space:]]Attributes[[:space:]]for[[:space:]]Training-Free[[:space:]]Sketch[[:space:]]Generation/88f15fb1-f5ab-4157-a23c-15d3cb3bc8f3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Stronger,[[:space:]]Steadier[[:space:]]&[[:space:]]Superior_[[:space:]]Geometric[[:space:]]Consistency[[:space:]]in[[:space:]]Depth[[:space:]]VFM[[:space:]]Forges[[:space:]]Domain[[:space:]]Generalized[[:space:]]Semantic[[:space:]]Segmentation/2e6ee106-ae7d-42e3-b798-bd6f97fafe0a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/StruMamba3D_[[:space:]]Exploring[[:space:]]Structural[[:space:]]Mamba[[:space:]]for[[:space:]]Self-supervised[[:space:]]Point[[:space:]]Cloud[[:space:]]Representation[[:space:]]Learning/9ae9c283-762a-4256-b4de-303cb73e1e2c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Structure[[:space:]]Matters_[[:space:]]Revisiting[[:space:]]Boundary[[:space:]]Refinement[[:space:]]in[[:space:]]Video[[:space:]]Object[[:space:]]Segmentation/bf659278-d12b-4982-824e-fbf802268f8f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Structure-Guided[[:space:]]Diffusion[[:space:]]Models[[:space:]]for[[:space:]]High-Fidelity[[:space:]]Portrait[[:space:]]Shadow[[:space:]]Removal/73c9755b-db1d-47f9-8857-79c13b240555_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Structure-aware[[:space:]]Semantic[[:space:]]Discrepancy[[:space:]]and[[:space:]]Consistency[[:space:]]for[[:space:]]3D[[:space:]]Medical[[:space:]]Image[[:space:]]Self-supervised[[:space:]]Learning/d73adb04-404d-4f0b-9d9c-095f1f245a9b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Structured[[:space:]]Policy[[:space:]]Optimization_[[:space:]]Enhance[[:space:]]Large[[:space:]]Vision-Language[[:space:]]Model[[:space:]]via[[:space:]]Self-referenced[[:space:]]Dialogue/5161fa7a-2cd7-4a41-959a-e1ddb02c88dc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/StyleKeeper_[[:space:]]Prevent[[:space:]]Content[[:space:]]Leakage[[:space:]]using[[:space:]]Negative[[:space:]]Visual[[:space:]]Query[[:space:]]Guidance/45fcd356-f226-4ed7-980b-85432724bf8a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/StyleMotif_[[:space:]]Multi-Modal[[:space:]]Motion[[:space:]]Stylization[[:space:]]using[[:space:]]Style-Content[[:space:]]Cross[[:space:]]Fusion/ffae00f3-5394-4dd8-b9a4-bc3713aa96e3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/StyleSRN_[[:space:]]Scene[[:space:]]Text[[:space:]]Image[[:space:]]Super-Resolution[[:space:]]with[[:space:]]Text[[:space:]]Style[[:space:]]Embedding/34baee7e-bb1d-4173-bb75-506b85eb6a0f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Stylized-Face_[[:space:]]A[[:space:]]Million-level[[:space:]]Stylized[[:space:]]Face[[:space:]]Dataset[[:space:]]for[[:space:]]Face[[:space:]]Recognition/fc351eb8-6a0f-465e-ac2f-6fa6da65b294_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SuMa_[[:space:]]A[[:space:]]Subspace[[:space:]]Mapping[[:space:]]Approach[[:space:]]for[[:space:]]Robust[[:space:]]and[[:space:]]Effective[[:space:]]Concept[[:space:]]Erasure[[:space:]]in[[:space:]]Text-to-Image[[:space:]]Diffusion[[:space:]]Models/51080fa1-f759-4df5-8927-63f76b232e8f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Subjective[[:space:]]Camera[[:space:]]1.0_[[:space:]]Bridging[[:space:]]Human[[:space:]]Cognition[[:space:]]and[[:space:]]Visual[[:space:]]Reconstruction[[:space:]]through[[:space:]]Sequence-Aware[[:space:]]Sketch-Guided[[:space:]]Diffusion/cae6ed8f-de57-4610-b479-1652f9260751_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SummDiff_[[:space:]]Generative[[:space:]]Modeling[[:space:]]of[[:space:]]Video[[:space:]]Summarization[[:space:]]with[[:space:]]Diffusion/314d20e1-b991-4676-b961-e90cddf3f9d6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Super[[:space:]]Resolved[[:space:]]Imaging[[:space:]]with[[:space:]]Adaptive[[:space:]]Optics/330a7bf8-492d-4b73-b28f-8e96f1e5081a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SuperDec_[[:space:]]3D[[:space:]]Scene[[:space:]]Decomposition[[:space:]]with[[:space:]]Superquadrics[[:space:]]Primitives/ed2eb095-495d-4dd3-9f36-9c1449d1ce9d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SuperEdit_[[:space:]]Rectifying[[:space:]]and[[:space:]]Facilitating[[:space:]]Supervision[[:space:]]for[[:space:]]Instruction-Based[[:space:]]Image[[:space:]]Editing/395fe9d1-b3f1-4c45-afff-e5eaba8e26ad_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SuperEvent_[[:space:]]Cross-Modal[[:space:]]Learning[[:space:]]of[[:space:]]Event-based[[:space:]]Keypoint[[:space:]]Detection[[:space:]]for[[:space:]]SLAM/2e684548-d342-4d0d-9d82-edd690c021a1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SuperMat_[[:space:]]Physically[[:space:]]Consistent[[:space:]]PBR[[:space:]]Material[[:space:]]Estimation[[:space:]]at[[:space:]]Interactive[[:space:]]Rates/4daa8d06-6978-409c-9337-f3e4e925bd23_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Supercharged[[:space:]]One-step[[:space:]]Text-to-Image[[:space:]]Diffusion[[:space:]]Models[[:space:]]with[[:space:]]Negative[[:space:]]Prompts/1564dd9e-386a-48aa-bdfb-0018fc932f68_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Supercharging[[:space:]]Floorplan[[:space:]]Localization[[:space:]]with[[:space:]]Semantic[[:space:]]Rays/834ff1eb-78d1-464a-8d20-1e5cbed96504_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Superpowering[[:space:]]Open-Vocabulary[[:space:]]Object[[:space:]]Detectors[[:space:]]for[[:space:]]X-ray[[:space:]]Vision/60987d09-4da2-484e-8cc1-bea5f240eaf8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Supervised[[:space:]]Exploratory[[:space:]]Learning[[:space:]]for[[:space:]]Long-Tailed[[:space:]]Visual[[:space:]]Recognition/3fb70b61-418e-4890-be11-1e54f33e6902_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2025/SpecGuard_ Spectral Projection-based Advanced Invisible Watermarking/1fcb8d1e-6633-4c96-8cd9-9ceeae669d64_content_list.json b/2025/SpecGuard_ Spectral Projection-based Advanced Invisible Watermarking/1fcb8d1e-6633-4c96-8cd9-9ceeae669d64_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8800657265302945c27d8f7f5a96ad7cb56d5af7 --- /dev/null +++ b/2025/SpecGuard_ Spectral Projection-based Advanced Invisible Watermarking/1fcb8d1e-6633-4c96-8cd9-9ceeae669d64_content_list.json @@ -0,0 +1,1767 @@ +[ + { + "type": "text", + "text": "SpecGuard: Spectral Projection-based Advanced Invisible Watermarking", + "text_level": 1, + "bbox": [ + 125, + 130, + 870, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inzamamul Alam, Md Tanvir Islam, Simon S. Woo*, Khan Muhammad Sungkyunkwan University", + "bbox": [ + 220, + 180, + 776, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{inzil5, tanvirnwu, swoo, khanmuhammad}@g.skku.edu", + "bbox": [ + 274, + 219, + 718, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Watermarking embeds imperceptible patterns into images for authenticity verification. However, existing methods often lack robustness against various transformations primarily including distortions, image regeneration, and adversarial perturbation, creating real-world challenges. In this work, we introduce SpecGuard, a novel watermarking approach for robust and invisible image watermarking. Unlike prior approaches, we embed the message inside hidden convolution layers by converting from the spatial domain to the frequency domain using spectral projection of a higher frequency band that is decomposed by wavelet projection. Spectral projection employs Fast Fourier Transform approximation to transform spatial data into the frequency domain efficiently. In the encoding phase, a strength factor enhances resilience against diverse attacks, including adversarial, geometric, and regeneration-based distortions, ensuring the preservation of copyrighted information. Meanwhile, the decoder leverages Parseval's theorem to effectively learn and extract the watermark pattern, enabling accurate retrieval under challenging transformations. We evaluate the proposed SpecGuard based on the embedded watermark's invisibility, capacity, and robustness. Comprehensive experiments demonstrate the proposed SpecGuard outperforms the state-of-the-art models. To ensure reproducibility, the full code is released on GitHub.", + "bbox": [ + 86, + 301, + 485, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 696, + 220, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the rapid advancement of digital media and artificial intelligence, concerns regarding image authenticity, copyright protection, and content integrity have become more challenging than ever [9, 16, 32]. Moreover, the widespread availability of the latest image manipulation tools [4, 5, 14] enables malicious tamperers to easily forge and redistribute digital content without authorization, posing a significant threat to ownership verification [29]. This growing risk emphasizes the need for reliable techniques for secure authentication and detection of unauthorized manipulation.", + "bbox": [ + 89, + 722, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a38e60f474c97ee8790f581a20a8d3b9e9e3db28ed3cebb62e1971ff215af1a0.jpg", + "image_caption": [ + "Figure 1. Image authentication using our proposed SpecGuard." + ], + "image_footnote": [], + "bbox": [ + 516, + 268, + 906, + 426 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, invisible watermarking has gained significant attention as a prominent defense mechanism for media authentication by embedding invisible messages into images to verify authenticity [35, 48]. In fact, invisible watermarks are preferred for preserving image quality and resisting tampering. These watermarks are unique to the creator and enable tamper verification by comparing the retrieved watermark to the original, as the high-level process is presented in Fig. 1. Traditional watermarking methods often rely on transformation techniques [18, 33]. Deep learning approaches like StegaStamp [41], Stable Signature [35], and HiDDeN [53] provide end-to-end solutions for message embedding. However, these methods often struggle with fragility in handling common image processing operations such as resizing, cropping, compression, and noise addition, which can distort or erase the embedded watermark. Additionally, the performance of watermark embedding and extraction often remains vulnerable to attacks with noise injection, blurring, contrasting, and rotation [8].", + "bbox": [ + 511, + 462, + 906, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address the aforementioned challenges, we introduce a novel robust, and invisible image watermarking method named SpecGuard. SpecGuard is designed to overcome the fundamental trade-offs [8] between imperceptibility, and robustness. Our proposed SpecGuard strategically embeds watermark information in the spectral domain, leveraging wavelet-based decomposition to distribute the watermark across high-frequency components. Unlike traditional frequency domain watermarking techniques [47, 49] that are easily disrupted by common image manipulations, Spec-", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: swoo@g.skku.edu (Simon S. Woo)", + "bbox": [ + 107, + 886, + 424, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17984", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Guard maintains imperceptibility while significantly improving robustness against a wide range of transformations.", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overall, our proposed SpecGuard addresses the current limitations of the previous watermarking methods by providing a robust, imperceptible watermarking technique that maintains integrity under diverse manipulations, significantly enhancing digital content security and authenticity verification. Our key contributions are as follows:", + "bbox": [ + 89, + 125, + 483, + 218 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a novel watermarking approach that embeds message bits in high-frequency spectral components via wavelet and spectral projection inside hidden convolutional layers, ensuring robustness against various transformations and adversarial attacks.", + "- We adapt Parseval's theorem [20] as a learnable threshold to optimize SpecGuard and spectral masking for robust watermark bit recovery under diverse transformations including distortions, regeneration, and adversarial attacks, proven through the experimental results.", + "- Our extensive evaluations demonstrate SpecGuard's superior bit embedding capacity and producing better invisible watermarked images, surpassing the performance of state-of-the-art (SOTA) methods." + ], + "bbox": [ + 89, + 226, + 483, + 438 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 463, + 240, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Watermarking an image has been a widely researched topic for securing the ownership and verifying authenticity of digital content [37]. Traditional watermarking techniques typically embed invisible [42] or visible [7] watermarks into images, which can later be extracted or detected to verify the content's originality. These methods can be broadly classified into spatial-domain [39, 44] and frequency-domain [13] watermarking, while some are based on combined methods [38, 50]. However, researchers recently proposed many advanced models [23, 27, 31, 40] for effective watermark removal. To face this growing challenge, researchers introduced different methods [2, 3, 15, 22, 28, 51, 53] as alternatives to deep learning-based encoders or decoders to produce more robust image watermarking. Furthermore, iterative models have demonstrated competitive performance [21, 34], particularly in robustness against a wide range of transformations. In addition, with the rise of generative methods, researchers used the watermark-labeled data for training to learn how to produce watermarks [11, 24]. Also, models that combine generative methods with watermarking techniques show promise in effective image watermarking [25, 30, 36]. However, such approaches face limitations such as increased computational complexity and longer processing times. These approaches are also more vulnerable to adversarial attacks that can target and distort the embedded watermark without altering the content visibly.", + "bbox": [ + 89, + 493, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Proposed Method: SpecGuard", + "text_level": 1, + "bbox": [ + 513, + 89, + 794, + 107 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We introduce SpecGuard, as illustrated in Fig. 2, which involves two fundamental modules: an \"Encoder\" for embedding the watermark and a \"Decoder\" for accurately extracting the watermark detailed in the following sections.", + "bbox": [ + 511, + 114, + 906, + 176 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Encoder", + "text_level": 1, + "bbox": [ + 513, + 184, + 614, + 199 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By targeting high-frequency components, the encoder integrates a binary message $M$ into the cover image $I$ . Using wavelet projection (WP) [33] and a Fast Fourier Transform (FFT)-based spectral projection (SP) [18] approximation, the message $M$ is inserted into specific frequency bands, minimizing perceptual impact.", + "bbox": [ + 511, + 205, + 905, + 296 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Wavelet Projection. We use a wavelet projection to capture frequency and spatial localization features that describe an image across different scales, as shown in Eq. (1):", + "bbox": [ + 511, + 297, + 905, + 342 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nW (a, b) = \\frac {1}{\\sqrt {| a |}} \\int_ {- \\infty} ^ {\\infty} f (x) \\psi \\left(\\frac {x - b}{a}\\right) d x, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 349, + 906, + 382 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $a \\in \\mathbb{R} \\setminus \\{0\\}$ , and $b \\in \\mathbb{R}$ denote the scaling and translation parameters, respectively. Here, $\\psi_{a,b}(x)$ represents a rescaled and translated form of the mother wavelet $\\psi$ , defined as follows:", + "bbox": [ + 511, + 388, + 905, + 448 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\psi_ {a, b} (x) = \\psi \\left(\\frac {x - b}{a}\\right) \\cdot \\frac {1}{\\sqrt {| a |}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 454, + 905, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\frac{1}{\\sqrt{|a|}}$ functions as a normalization factor, guaranteeing that the energy of the wavelet is invariant to the scaling parameter $a$ . Minimal values of $a$ compress the wavelet, enabling the inspection of high-frequency components, whereas greater values of $a$ elongate the wavelet, promoting low-frequency analysis. Since each mother wavelet $\\psi$ is built with zero mean and finite energy [12], it guarantees to maintain stability as follows:", + "bbox": [ + 511, + 494, + 905, + 621 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\int_ {- \\infty} ^ {\\infty} \\psi (x) d x = 0, \\quad \\int_ {- \\infty} ^ {\\infty} | \\psi (x) | ^ {2} d x < \\infty , \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 628, + 903, + 659 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where the wavelet projection from Eq. (1) decomposes the input into orthogonal wavelet sets using discrete scales and translations. For $2D$ inputs, the scaled and translated basis elements [1] are defined for each coordinate pair $(u,v)$ :", + "bbox": [ + 511, + 666, + 905, + 727 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} _ {L L} = \\phi (u, v) = \\phi (u) \\phi (v), \\qquad \\mathbf {S} _ {L H} = \\psi_ {H} (u, v) = \\psi (u) \\phi (v),\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 734, + 913, + 751 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} _ {H L} = \\psi_ {V} (u, v) = \\phi (u) \\psi (v), \\quad \\mathbf {S} _ {H H} = \\psi_ {D} (u, v) = \\psi (u) \\psi (v), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 753, + 913, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where, $H$ , $V$ , and $D$ represent the horizontal, vertical, and diagonal decomposition direction, respectively. To depict the image at different resolutions, we define scaling and wavelet functions at scale $j$ as shown below:", + "bbox": [ + 511, + 781, + 906, + 840 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {j, m, n} (u, v) = 2 ^ {j / 2} \\phi \\left(u - \\frac {m}{2 ^ {j}}, v - \\frac {n}{2 ^ {j}}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 848, + 903, + 883 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\psi_ {j, m, n} ^ {d} (u, v) = 2 ^ {j / 2} \\psi^ {d} \\left(u - \\frac {m}{2 ^ {j}}, v - \\frac {n}{2 ^ {j}}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 878, + 852, + 904 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17985", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d547e59ba83c2de38ccd65c824e949168c58fb8718a8d8f0e3fe187603ffd931.jpg", + "image_caption": [ + "Figure 2. Architecture of the proposed SpecGuard watermarking method involves encoding a binary message $M$ into the high-frequency band of the cover image $I$ using wavelet and spectral projection and learning to decode the embedded message." + ], + "image_footnote": [], + "bbox": [ + 96, + 87, + 903, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $d \\in \\{H, V, D\\}$ is the wavelet function direction that serves as discrete basis elements for multi-resolution analysis, capturing details across frequency bands and spatial locations. In Eq. (6), $T_{m,n}$ denotes the intensity or pixel value of the cover image $I$ at spatial coordinates $(m, n)$ . The discrete scaling function $W_{\\phi}(j, u, v)$ (approximation at scale $j$ ) and the detail coefficients $W_{\\psi}^{d}(j, u, v)$ for each direction are computed accordingly as follows:", + "bbox": [ + 89, + 354, + 483, + 476 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nW _ {\\phi} (j, u, v) = \\frac {1}{l} \\sum_ {m = 0} ^ {l - 1} \\sum_ {n = 0} ^ {l - 1} T _ {m, n} \\phi \\left(m - u \\cdot 2 ^ {- j}, n - v \\cdot 2 ^ {- j}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 484, + 475, + 521 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nW _ {\\psi} ^ {d} (j, u, v) = \\frac {1}{l} \\sum_ {m = 0} ^ {l - 1} \\sum_ {n = 0} ^ {l - 1} T _ {m, n} \\psi^ {d} \\left(m - u \\cdot 2 ^ {- j}, n - v \\cdot 2 ^ {- j}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 523, + 483, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with $l$ as the discrete region dimension, these coefficients capture multi-scale, multi-orientation image information, forming the basis of spectral features as follows:", + "bbox": [ + 89, + 573, + 483, + 619 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\beta_ {j} = \\bigcup_ {d \\in \\{H, V, D \\}} \\left(W _ {\\phi} (j, u, v) \\cup W _ {\\psi} ^ {d} (j, u, v)\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 626, + 482, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This feature set $\\beta_{j}$ captures key frequency and spatial details across resolutions, forming the foundation for the watermark embedding process of our SpecGuard.", + "bbox": [ + 89, + 666, + 482, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Selective Frequency Band Decomposition. To refine the embedding process, we segment the data into distinct frequency bands. The decomposition level $\\kappa$ is determined by the image complexity, calculated as follows:", + "bbox": [ + 89, + 710, + 482, + 771 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\kappa = \\left\\lfloor \\sqrt {\\log (1 + N)} \\right\\rfloor , \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 780, + 482, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $N$ denotes the total pixel count in the cover image $I$ . And, each component $\\beta_{j}$ falls within a unique frequency band, yielding a total of $1 + 3\\kappa$ distinct frequency bands as follows:", + "bbox": [ + 89, + 806, + 483, + 864 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\beta_ {j} = \\phi_ {j} (u, v) \\cup \\bigcup_ {d \\in \\{H, V, D \\}} \\psi_ {j} ^ {d} (u, v). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 873, + 482, + 904 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The components $\\beta_{j}$ , consisting of scaling functions $\\phi_j(u,v)$ and wavelet functions $\\psi_j^d (u,v)$ , capture specific spatial frequency bands, enabling targeted high-frequency embedding. We translate the WP into disjoint intervals representing a unique frequency range to approximate the segmentation in the frequency domain:", + "bbox": [ + 511, + 354, + 906, + 445 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\beta_ {j} = \\left\\{W _ {\\psi} ^ {d} (u, v) \\mid u, v \\in \\left(\\frac {j \\cdot L}{\\kappa}, \\frac {(j + 1) \\cdot L}{\\kappa}\\right) \\right\\}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 454, + 906, + 486 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where, $L$ is the dimension of $S_{HH}$ , and $W_{\\psi}^{d}(u,v)$ represents wavelet values within segmented intervals. This frequency band partitioning mimics the frequency selectivity of wavelet sub-bands, enabling effective targeting of high-frequency regions for optimal embedding.", + "bbox": [ + 511, + 494, + 905, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Approximation of Spectral Projection. We first apply spectral projection on the $S_{HH}$ sub-band, transforming it into the spectral domain. Given a matrix $T(x,y)$ representing pixel intensities in $S_{HH}$ , the spectral projection computes the spectral components $\\zeta(u,v)$ as follows:", + "bbox": [ + 511, + 571, + 905, + 647 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\zeta (u, v) = \\frac {1}{L ^ {2}} \\sum_ {x} \\sum_ {y} T (x, y) \\cdot \\exp \\left(- i \\frac {2 \\pi}{L} (x \\cdot u + y \\cdot v)\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 657, + 903, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $L$ denotes the dimension of $S_{HH}$ , $T(x,y)$ provides the intensity at each coordinate $(x,y)$ which is equivalent to $W_{\\psi}^{d}(u,v)$ in Eq. (6), $i$ is the imaginary unit, and $(u,v)$ are the spectral coordinates.", + "bbox": [ + 511, + 704, + 905, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To approximate the spectral components using the FFT, we create a symmetrically extended version $\\tilde{T}(x,y)$ of the original $N\\times N$ matrix $T(x,y)$ . This extension is achieved by mirroring $T(x,y)$ along its boundaries, doubling its size to $2N\\times 2N$ . Specifically, the original matrix occupies the top-left quadrant, with the remaining quadrants filled by reflecting $T(x,y)$ horizontally, vertically, and diagonally, respectively. This symmetric structure ensures that the FFT yields only real values, allowing the spectral coefficients to", + "bbox": [ + 511, + 765, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17986", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "be extracted directly from the real part of the FFT operation. Then, we apply the 2D FFT to $\\tilde{T} (x,y)$ as follows:", + "bbox": [ + 89, + 90, + 480, + 122 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF (u, v) = \\frac {1}{(2 N) ^ {2}} \\sum_ {x} \\sum_ {y} \\tilde {T} (x, y) \\cdot \\exp \\left(- i \\frac {2 \\pi}{2 N} (x \\cdot u + y \\cdot v)\\right) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 130, + 486, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The SP coefficients are then approximated by taking the real part (Re) of $F$ in the original $N \\times N$ region as follows:", + "bbox": [ + 89, + 178, + 483, + 208 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\zeta (u, v) \\approx \\operatorname {R e} (F (u, v)), \\quad 0 \\leq u, v < N. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 222, + 482, + 238 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Applying Eq. (13) to the sub-bands extracted from wavelet projection in Eq. (6), we achieve a computationally efficient spectral projection by leveraging the FFT approximation on a symmetrically extended matrix, maintaining effective embedding properties within the spectral domain.", + "bbox": [ + 89, + 241, + 483, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SpecGuard Embedding Process. The embedding process integrates the binary message $M$ into the high-frequency band $S_{HH}$ of the cover image $I$ , enhancing robustness and imperceptibility through wavelet and spectral projection. Using the Eq. (6) and Eq. (13), the cover image $I$ is decomposed into sub-bands $S_{LL}, S_{LH}, S_{HL}$ , and $S_{HH}$ within spectral domain, with $S_{HH}$ providing high-frequency details for embedding. A variable number $k$ of convolutional layers with a $K \\times K$ kernel, followed by LeakyReLU activation, are recursively applied to $S_{HH}$ to refine features:", + "bbox": [ + 89, + 316, + 483, + 468 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {H H} ^ {(n + 1)} = \\text {L e a k y R e L U} \\left(\\operatorname {C o n v} _ {2 D} \\left(S _ {H H} ^ {(n)}, K\\right)\\right), \\quad n = 1, \\dots , k. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 486, + 482, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The final output $S_{HH}^{(n + 1)}$ from Eq. (14) represents the modified high-frequency band, primed for embedding.", + "bbox": [ + 89, + 511, + 482, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The message $M$ , represented as a binary vector of length $l$ ( $M \\in \\{0,1\\}^l$ ), with batch size $b$ and message length $l$ , is reshaped and expanded across channels $c$ to align with $S_{HH}^{(n+1)}$ . This ensures $M_{\\mathrm{expanded}}$ conforms to the dimension $[b,c,l]$ , where each message is structured accordingly.", + "bbox": [ + 89, + 542, + 482, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To localize the embedding, we create a radial mask centered at $(c_x, c_y) = \\left(\\frac{h}{2}, \\frac{w}{2}\\right)$ , where $h$ and $w$ represent the height and width of $S_{HH}$ . The Euclidean distance $D(x_i, y_i)$ from the center $(c_x, c_y)$ is computed for each coefficient $(x_i, y_i)$ . A binary mask is then generated within the predefined radius $r$ based on the distance $D(x_i, y_i)$ , such that if $D(x_i, y_i) \\leq r$ , the mask value is set to 1, allowing embedding in the corresponding region. Otherwise, the mask value is 0, restricting embedding to areas within a specified radius $r$ , ensuring focus on high-frequency regions.", + "bbox": [ + 89, + 621, + 482, + 770 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each coordinate $(x_{i},y_{i})$ where mask $(x_{i},y_{i})$ is 1 and $W_{c}\\in c$ , the embedding operation is performed as follows:", + "bbox": [ + 89, + 771, + 482, + 801 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {H H} ^ {(n + 1)} [:, W _ {c}, x _ {i}, y _ {i} ] + = M _ {\\text {e x p a n d e d}} [:, W _ {c}, i ] \\cdot s, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 810, + 482, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $s$ is the strength factor controlling embedding intensity and invisibility. After embedding, the modified coefficients $S_{HH}^{(n + 1)}$ undergo a final convolution and LeakyReLU using Eq. (14), by setting the value of $k = 1$ to harmonize", + "bbox": [ + 89, + 838, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the embedded message. Following this approach, SpecGuard embeds the message into the spectral domain in a transformed form, differing from its original input representation. By blending the message seamlessly into the spectral space based on the $r$ , $s$ , and $W_{c}$ , it becomes inherently concealed within the domain, rendering its presence imperceptible. Without knowledge of $r$ , $s$ , and $W_{c}$ , it becomes exceedingly challenging to localize the embedded message, further enhancing the security of the system. This transformation ensures the embedding process remains opaque to any adversarial attacker, making SpecGuard black-box.", + "bbox": [ + 511, + 90, + 903, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reconstruction. SpecGuard encoder reconstructs the watermarked image $I_{\\mathrm{embedded}}$ by inverse transformation restoring $S_{HH}$ back into the spatial domain. The reconstruction process integrates the inverse wavelet projection (IWP) [33] and inverse spectral projection (ISP) [18], ensuring the embedded modifications are correctly translated into the spatial domain. To reconstruct the spatial domain image, $S_{HH}$ is combined with the other sub-bands $S_{LL}, S_{LH}$ , and $S_{HL}$ . For the SP embedded in $S_{HH}$ , the ISP is applied to reconstruct $S_{HH}$ to spatial domain as follows:", + "bbox": [ + 511, + 257, + 905, + 407 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {H H} (x, y) = \\sum_ {u = 0} ^ {L - 1} \\sum_ {v = 0} ^ {L - 1} \\zeta (u, v) \\cdot \\exp \\left(i \\frac {2 \\pi}{L} (x \\cdot u + y \\cdot v)\\right), \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 412, + 903, + 448 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\zeta(u, v)$ represents spectral coefficients from the embedding process, $L$ denotes the dimension of $S_{HH}$ , and $(x, y)$ are spatial coordinates. SpecGuard then reconstructs the watermarked image $I_{\\mathrm{embedded}}$ using the IWP as follows:", + "bbox": [ + 511, + 452, + 905, + 513 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {\\text {e m b e d d e d}} (x, y) = \\operatorname {I W P} \\left(S _ {L L}, S _ {L H}, S _ {H L}, S _ {H H}\\right). \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 517, + 903, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This process seamlessly embeds the watermark message $M$ in the spectral domain, preserving the cover image $I$ 's integrity. The inverse transformations that are expressed in Eq. (16) and Eq. (17) fully restore visual quality, maintaining all frequency components.", + "bbox": [ + 511, + 537, + 905, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.Decoder", + "text_level": 1, + "bbox": [ + 513, + 622, + 614, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Algorithm 1, SpecGuard decoding process starts by applying wavelet projection ( Eq. (1)) to the watermarked image $I_{\\mathrm{embedded}}$ , separating it into low and high-frequency bands, where the high-frequency band $S_{D_{HH}}^{high}$ contains the embedded message similar to the process in the encoding phase, particularly in Eq. (6). An approximation of the spectral projection using FFT as shown in Eq. (13) is then applied to $S_{D_{HH}}^{high}$ returning the transformed data $S_{D_{HH}}^{\\mathrm{sp}}$ . Then, $S_{D_{HH}}$ is further refined through convolutional layers that captures the local features for message extraction.", + "bbox": [ + 511, + 643, + 905, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To extract the message, a radial mask is created to isolate high-frequency areas within $S_{D_{HH}}$ , targeting the embedded regions based on their distance from the center. The masked values are compared against a learnable threshold $\\theta$ to decode each bit of the hidden message $D_{\\mathrm{M}}$ . Here, $\\theta$ serves as a threshold that adapts to the spectral patterns across the entire image, learning the distinct characteristics", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17987", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 SpecGuard decoder with wavelet, spectral projection with FFT approximation, and learnable threshold." + ], + "code_body": "1: Input: Watermarked image $I_{\\mathrm{embedded}}$ , learnable $\\theta$ message length $l$ \nradius $r$ ,watermark channel $W_{c}$ \n2: Output: Decoded binary message $D_{\\mathrm{M}}$ \n3: Procedure: Apply Wavelet Projection on $I_{\\mathrm{embedded}}$ to obtain $S_{D_{LL}}$ (low-frequency) and $S_{D_{HH}}^{\\mathrm{high}}$ (high-frequency) \n4: Procedure: Spectral approximation with FFT $(S_{D_{HH}}^{\\mathrm{high}})$ .. \n5: Separate even and odd indices: $v = [x_{\\mathrm{even}},$ reverse $(x_{\\mathrm{odd}})]$ \n6: Compute FFT on $v$ .. $V_{\\mathrm{complex}} = \\mathrm{FFT}(v)$ \n7: $V_{\\mathrm{real}} = V_{\\mathrm{complex}}\\cdot \\left[\\cos \\left(\\frac{-\\pi k}{2N}\\right),\\sin \\left(\\frac{-\\pi k}{2N}\\right)\\right] / /$ Calculate Real \n8: $V_{\\mathrm{real}}[0]\\gets \\frac{V_{\\mathrm{real}}[0]}{\\sqrt{N\\cdot 2}},V_{\\mathrm{real}}[1:\\gets \\frac{V_{\\mathrm{real}}[1:]}{\\sqrt{\\frac{N}{2}\\cdot 2}} /$ Energy preservation \n9: Transpose result and repeat to obtain $S_{D_{HH}}^{\\mathrm{sp}}$ \n10: Return $S_{D_{HH}}^{\\mathrm{sp}}$ \n11: Procedure: Pass $S_{D_{HH}}^{\\mathrm{sp}}$ through sequential layers as: \n $S_{D_{HH}}^{(n + 1)} =$ LeakyReLU $\\left(\\operatorname {Conv}_{2D}\\left(S_{HH}^{sp(n)},K\\right)\\right)$ $n = 1,\\dots,k,$ \n12: Return $S_{D_{HH}}^{(n + 1)}$ \n13: Procedure: Extraction $(S_{D_{HH}}^{(n + 1)},l)$ . \n14: Set $(c_x,c_y) = \\left(\\frac{H}{2},\\frac{W}{2}\\right)$ \n15: Generate mask for high-frequency region within radius $r$ \nfor each coordinate $(i,j)$ do: \n $D(x_i,y_i) = \\sqrt{(x_i - c_x)^2 + (y_i - c_y)^2}$ // Euclidian Distance \n17: if $D(x_i,y_i)\\leq r$ then \n18: Set mask[i, $j] = 1$ \n19: end if \n20: end for \n21: Extract mask: $S_{D_{HH}}[:,W_c,$ mask[i, $j]$ \n22: Decode message using learnable $\\theta$ ..", + "bbox": [ + 94, + 123, + 483, + 526 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nD _ {\\mathbf {M}} [ i ] = \\left\\{ \\begin{array}{l l} 1 & \\text {i f E x t r a c t e d} [ i ] > \\theta \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 532, + 392, + 566 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [], + "code_body": "23: Update $\\theta$ dynamically: $\\theta \\gets \\theta -\\eta \\cdot \\frac{\\partial L_{\\mathrm{dec}}}{\\partial\\theta}$ // Optimizes robustness \n24: Return $D_{\\mathrm{M}}$", + "bbox": [ + 91, + 574, + 482, + 601 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of the embedded watermark. From Parseval's theorem [20] ensures overall spectral and spatial energies remain equivalent, though local spectral energy distributions are altered by the watermark strength factor $s$ .", + "bbox": [ + 89, + 638, + 483, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The watermark's strength factor $s$ ensures that the high-energy areas where the message $M$ is embedded as \"1\" remain robust, experiencing a minimum distortion in such conditions. Moreover, this threshold can be optimized for better bit recovery accuracy during training. As $\\theta$ learns, it recognizes that areas encoded as \"1\" carry higher energy and impact due to the strength factor $s$ of Eq. (15), while areas marked as \"0\", softened by the LeakyReLU's minimal negative slope, have a lower intensity. Such a dynamic approach enables $\\theta$ to identify and protect the embedded message $M$ even when external disturbances occur, preserving the watermark's structure within the watermarked image $I_{\\text{embedded}}$ . And, $\\theta$ effectively learns to distin", + "bbox": [ + 89, + 704, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "guish high-energy watermark regions. Therefore, the embedded message is more recoverable under diverse attacks, and SpecGuard's decoder ensures valid watermark bit extraction. Theoretical explanation of Parseval theorem's [20] impact on message extraction is in the Supplementary.", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Loss Calculation for SpecGuard", + "text_level": 1, + "bbox": [ + 511, + 172, + 797, + 190 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To achieve the training objective of robust and invisible watermark embedding, a composite loss function is defined with two terms: encoder loss $L_{\\mathrm{enc}}$ as expressed in Eq. (18) and decoder loss $L_{\\mathrm{dec}}$ as expressed in Eq. (19).", + "bbox": [ + 511, + 195, + 905, + 256 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathbb {E} _ {(I, M) \\sim D} L _ {\\text {e n c}} (I, I _ {\\text {e m b e d d e d}}) = \\| E _ {\\theta} (I, M) - I \\| ^ {2}, \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 258, + 905, + 281 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathbb {E} _ {(I, M) \\sim D} L _ {\\mathrm {d e c}} (M, D _ {\\mathrm {M}}) = \\| D _ {\\theta} \\left(I _ {\\text {e m b e d d e d}}\\right) - M \\| ^ {2}, \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 301, + 905, + 323 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $E_{\\theta}(I,M)$ denotes the encoder output, embedding the message $M$ into the cover image $I$ to produce $I_{\\mathrm{embedded}}$ . By minimizing $L_{\\mathrm{enc}}$ , the encoder learns to embed the watermark invisibly, preserving the fidelity of the cover image. $D_{\\theta}(I_{\\mathrm{embedded}})$ denotes the decoder's output from the watermarked image $I_{\\mathrm{embedded}}$ . Minimizing $L_{\\mathrm{dec}}$ enables the decoder to reliably retrieve the embedded message under varying conditions, such as noise and transformation.", + "bbox": [ + 511, + 327, + 905, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The total loss $L$ as shown in Eq. (20) used for optimizing the model combines these terms, balancing invisibility and robustness through weighted coefficients as follows:", + "bbox": [ + 511, + 448, + 905, + 493 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} L = \\lambda_ {\\text {e n c}} L _ {\\text {e n c}} + \\lambda_ {\\text {d e c}} L _ {\\text {d e c}}, \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 498, + 903, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{\\mathrm{enc}}$ and $\\lambda_{\\mathrm{dec}}$ control the relative importance of visual fidelity and message recoverability.", + "bbox": [ + 511, + 522, + 905, + 554 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 564, + 718, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Dataset", + "text_level": 1, + "bbox": [ + 511, + 589, + 607, + 603 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SpecGuard is trained on the MS-COCO dataset [26], which contains 25K images. To evaluate the robustness of the watermarking methods including our SpecGuard against different types of attacks, such as distortions, regenerations, and adversarial attacks, we used three datasets: DiffusionDB [46], MS-COCO [26], and DALL-E31. Each of these datasets has a unique distribution of prompt words. We also ensured that no unethical or violent terms were included in the prompts. We randomly picked 200 images from MS-COCO [26] and applied watermark using SpecGuard for further verifying the robustness after uploading on various social media platforms and applying AI-based Photoshop Neural Filters (PNFs)2. The PNFs include depth blur, artistic style transfer, super zoom, JPEG artifact reduction, and colorization. For the super zoom filter, we set the 'Sharpen' and 'Noise Reduction' parameters to 15. For all other filters, we used the default settings.", + "bbox": [ + 511, + 609, + 906, + 868 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "$^{1}$ https://huggingface.co/datasets/OpenDatasets/dalle-3-dataset", + "bbox": [ + 529, + 875, + 862, + 886 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://www.adobe.com/products/photoshop/neural-filter.html", + "bbox": [ + 532, + 887, + 864, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17988", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a0aa4adfdd645083602c529592e0a888bef0b968035d569a2332b2aa6bacde2f.jpg", + "image_caption": [ + "Figure 3. Some best results for cover vs watermarked images with PSNR/SSIM $(\\uparrow)$ scores showing minimal visual degradation when watermarked using proposed SpecGuard." + ], + "image_footnote": [], + "bbox": [ + 112, + 87, + 460, + 246 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5f45c501721dd458500ffdb994d641de3800a94dae010fd0a877fc775eb9f492.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Metrics256 × 256512 × 5121024 × 1024
CelebA-HQMS-COCOCelebA-HQMS-COCOCelebA-HQMS-COCO
PSNR↑40.36140.32044.65144.68048.17048.081
SSIM↑0.98890.98880.99270.99270.99370.9936
FID↓16.45116.69016.97217.02017.44616.955
MSE↓0.00020.00020.00010.00010.00010.0001
", + "bbox": [ + 91, + 316, + 495, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation", + "text_level": 1, + "bbox": [ + 89, + 452, + 250, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We used CUDA v11.3 and PyTorch with a batch size of 32 and the Adam optimizer on a multiple NVIDIA RTX 2080-equipped server. Mean Squared Error (MSE) and Bit Recovery Accuracy (BRA) are used for loss and accuracy calculation. We used Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), Fréchet Inception Distance (FID), and MSE to evaluate perceptual quality. Our model is trained for 300 epochs, with the decoder learning rate set to $1 \\times 10^{-3}$ , reduced by half every 100 steps, and the encoder learning rate is set to $1 \\times 10^{-2}$ without scheduling. We set our watermark radius $(r)$ , strength factor $(s)$ , initial learning parameter $(\\theta)$ , and the number of convolutional layers $(k)$ to 100, 20, 0.001, and 32, respectively. This setup is applied with a message bit length (BL) of 48, 64, 128, and 256. Initially, decoder loss weight $(\\lambda_{\\mathrm{dec}})$ and encoder loss weight $(\\lambda_{\\mathrm{enc}})$ are set to 1.0 and 0.7, respectively. For assessing the robustness of watermarking methods against diverse attacks, we inherited the experimental setups from Waves [8] and used effective metrics such as \"Quality at 95% Performance (Q@0.95P)\", \"Quality at 70% Performance (Q@0.7P)\", \"Avg P\" and \"Avg Q.\" Here, Q@0.95P and Q@0.7P indicate the level of image quality degradation required for watermark detection accuracy to reach 95% and 70%, respectively. The average performance (Avg P) metric represents the mean detection accuracy across various attack strengths, while the average quality degradation (Avg Q) measures the overall impact of attacks on image quality [8, 35].", + "bbox": [ + 89, + 477, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e26592245971c4568b85a395ec79c4a5ee94aba7e1f88ef07e988612e10730f6.jpg", + "table_caption": [ + "Table 1. Perceptual quality of SpecGuard watermarked images across resolutions and datasets using a fixed 30-bit message length." + ], + "table_footnote": [], + "table_body": "
MethodsVenueBLPSNR↑SSIM↑FID↓BRA↑
Pre-processing methodsTree-Ring [47]NeurIPS'236432.330.9117.70.98
12832.100.9017.80.96
25631.850.8917.90.94
Stable Signature [35]ICCV'236430.000.8919.60.98
12829.800.8819.70.96
25629.500.8719.80.96
Yang et al. [49]CVPR'246431.450.9018.20.98
12831.200.8918.30.93
25630.950.8818.40.89
SleeperMark [45]CVPR'256431.800.9218.00.97
12831.600.9118.10.93
25631.350.9018.20.87
Post-processing methodsHiDDeN [53]ECCV'186432.010.8819.70.98
12831.800.8719.80.85
25631.500.8619.90.82
StegaStamp [41]CVPR'206428.500.9117.90.99
12828.200.9018.00.98
25628.000.8918.10.94
MBRS [17]ACM MM'216438.200.9617.90.98
12837.900.9518.00.96
25637.500.9418.20.94
FIN [10]AAAI'236436.700.9518.30.97
12836.400.9418.40.96
25636.100.9318.50.96
MuST [43]AAAI'246441.200.9717.50.98
12840.900.9617.60.93
25640.500.9517.80.90
EditGuard [52]CVPR'246441.560.9717.80.98
12841.300.9617.90.97
25640.900.9518.00.97
SpecGuard (Ours)ICCV'256442.590.9817.20.99
12842.890.9917.00.99
25640.860.9917.60.98
", + "bbox": [ + 516, + 88, + 901, + 448 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison of SOTA pre-processing and post-processing watermarking methods with SpecGuard without attacks.", + "bbox": [ + 511, + 458, + 906, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Watermark Invisibility", + "text_level": 1, + "bbox": [ + 511, + 510, + 725, + 526 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the invisibility of the embedded watermark, we conducted perceptual and quantitative assessments using SpecGuard. As shown in Fig. 3, there is no noticeable perceptual degradation between the cover and watermarked images, confirming that the watermark remains imperceptible to the human eye. For a more comprehensive evaluation, we created three subsets of different image sizes ranging between 256 to 1024 with images from the MS-COCO [26] and CelebA-HQ [19] datasets and applied the SpecGuard watermarking method to compare the average PSNR values between the cover and watermarked images, as in Tab. 1.", + "bbox": [ + 511, + 535, + 906, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For quantitative evaluation, we further compare the performance of SpecGuard with the SOTA pre-processing and post-processing watermarking methods. As presented in Tab. 2, SpecGuard achieves the highest PSNR of 42.89 when the bit length was 128. Additionally, it attains the highest SSIM of 0.99 at a BL of 128 and 256 among all compared methods, indicating minimal visual distortion. Additionally, SpecGuard achieved the lowest FID of 17.0 and the highest BRA of 0.99, ensuring strong robustness while maintaining imperceptibility. Overall, our results demonstrate that SpecGuard outperforms both pre-processing and post-processing watermarking methods, achieving superior imperceptibility and robustness.", + "bbox": [ + 511, + 704, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17989", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2a71067f8a7cecffde21a5c91f6e92b06087c2314d3e6580ab8f3c587d78eb5f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Attack TypeTree-Ring [47]Stable Signature [35]StegaStamp [41]SpecGuard (Ours)
Q@0.95PQ@0.7PAvg PAvg QQ@0.95PQ@0.7PAvg PAvg QQ@0.95PQ@0.7PAvg PAvg QQ@0.95PQ@0.7PAvg PAvg Q
DistortionsRotation0.4640.5210.3750.6480.6240.7020.5940.6500.4230.4980.3570.6160.8630.8630.6870.653
Crop0.5920.5920.3320.463infinf0.9950.4610.6020.6020.5400.4510.8120.8120.9980.742
Brightinfinfinf0.304infinf0.9980.305infinf0.9980.317infinf0.9980.466
Contrastinfinf0.9980.243infinf0.9980.243infinf0.9980.231infinf0.9980.556
Blur0.8611.1120.5631.221- inf- inf0.0001.2040.8480.9620.4141.0000.921inf1.0001.452
Noise0.548inf0.9800.3950.4020.5200.8700.390infinf1.0000.360infinf0.9990.568
JPEG0.4990.4990.9290.2840.4850.4850.7930.284infinf0.9980.263infinf1.0000.495
Geo0.5250.5930.2770.7680.850inf0.9370.7670.6630.6930.3960.7330.8690.8690.8650.623
Deg0.620inf0.8920.6940.2060.3690.3000.6790.8260.9750.8520.6640.8951.1410.9150.749
Combine0.5390.7510.4030.9080.5380.6910.3340.9000.9451.1010.7950.8700.9791.2560.9110.952
RegenerationRegen-Diff- inf0.3070.6120.323- inf- inf0.0010.3000.331inf0.9430.327infinf0.9820.477
Regen-DiffPinf0.3070.6010.327- inf- inf0.0010.3030.333inf0.9400.329infinf0.9820.562
Regen-VAE0.5780.5780.8320.3480.5450.5450.5160.339infinf1.0000.343infinf0.9950.521
Regen-KLVAEinfinf0.9900.233 6- inf0.1760.2170.206infinf1.0000.240infinf0.9900.492
Rinse-2xDiff- inf0.3330.5100.357- inf- inf0.0010.3320.391inf0.9410.366infinf0.9930.561
Rinse-4xDiff- inf0.3550.4430.466- inf- inf0.0000.4380.388inf0.9090.477infinf0.9920.533
AdversarialAdvEmbG-KLVAE8- inf0.1640.4480.253infinf0.9980.249infinf1.0000.232infinf1.0000.456
AdvEmbB-RN180.241inf0.9530.218infinf0.9990.212infinf1.0000.196infinf1.0000.467
AdvEmbB-CLIP0.541inf0.9320.549infinf0.9990.541infinf1.0000.488infinf1.0000.436
AdvEmbB-KLVAE160.195inf0.8880.238infinf0.9970.233infinf1.0000.206infinf1.0000.482
AdvEmbB-SdxlVAE0.222inf0.9340.221infinf0.9980.219infinf1.0000.204infinf1.0000.492
AdvCls-UnWM&WM- inf0.1020.4990.145infinf0.9990.101infinf1.0000.101infinf1.0000.497
AdvCls-Real&WMinfinf1.0000.047infinf0.9980.092infinf1.0000.106infinf1.0000.427
AdvCls-WM1&WM2- inf0.1010.4920.139infinf0.9990.084infinf1.0000.129infinf1.0000.441
", + "bbox": [ + 89, + 87, + 906, + 381 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Robustness comparison various across attacks using Q@0.95P(↑), Q@0.7P(↑), Avg P(↑) and Avg Q(↑). Here, 'inf' denotes that no attack was sufficient to degrade performance below the threshold, indicating strong robustness, whereas '-inf' signifies that even the weakest attack caused detection to fall below the threshold, reflecting weak robustness.", + "bbox": [ + 89, + 392, + 906, + 434 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c3af77e50ef9cbb097edd6fa72e5f5dc1f03dcdb43ce23ff8d8a8d959ba8cd77.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 109, + 446, + 267, + 578 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/241c17b3a2b9fbfba53c2704606421d5b419030a809fb05f6cd4584c9e9e8191.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 267, + 446, + 419, + 578 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3df03b9919033d0e53d1f6dd7d95dbe480ad075013eae83e0ac50a1c53f49898.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 446, + 575, + 578 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4b653ed4ee169dd20947b2d3dc9e8d5e0b593ccc7fedcef8cab76ddd2d56d318.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 446, + 730, + 578 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5270cac58775daa7649a400c4517462d0dd4897b733eecdbe49d58293744c7ee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 730, + 446, + 885, + 578 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b65904e3bc4503ca2ba4bee9c230545e7542f295729e50187ca6365eaafd7b81.jpg", + "image_caption": [ + "Figure 4. Robustness validation of our proposed SpecGuard under different distortion attacks, including geometric transformations: Geo (rotation, cropping), photometric modifications (brightness, contrast), and degradations: Deg (blur, noise, JPEG compression)." + ], + "image_footnote": [], + "bbox": [ + 109, + 579, + 264, + 710 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/789232424565e24e2ea17562c99484682a1733df209991e89e396a6149503d4e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 267, + 579, + 419, + 710 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/64eea155e5ca4eeaa015eb0b296aef43acd61b56cdf4f38fc6fa2b72180833f1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 579, + 575, + 710 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b46ee883b746c9f9e646c4abf09ba7c7cfcca0168764f5b5c5f5bd6ea3f5e85a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 579, + 730, + 710 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5aacc06a0c1b739db3d3ea1c7bb773eadce67648e90eac85bd0025e7b27edb56.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 730, + 579, + 885, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Capacity", + "text_level": 1, + "bbox": [ + 89, + 763, + 196, + 780 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate embedding capacity, we examined SpecGuard across different bit lengths and compared it with SOTA watermarking methods. Our experiments with 64, 128, and 256 bits demonstrate SpecGuard's high capacity of bit embedding while maintaining perceptual quality and robustness, as shown in Tab. 2. Notably, it achieves a PSNR of 42.89, the highest among all methods, along with the high", + "bbox": [ + 89, + 794, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "est BRA of 0.99 and the lowest FID of 17.0 at 128 bits, ensuring minimal visual impact. This adaptability to different bit lengths without quality loss makes SpecGuard ideal for applications requiring flexible watermark sizes. Unlike StegaStamp and HiDDeN, which suffer reduced BRA for higher message bits, SpecGuard consistently extracts bits across all tested lengths. SpecGuard's theoretical watermark capacity is provided in the Supplementary.", + "bbox": [ + 511, + 763, + 906, + 886 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17990", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3c1378a268f6043bfc033716b045ad086583b54b6b50a4d5ece5a6c2d41dff48.jpg", + "table_caption": [], + "table_footnote": [ + "*WP: Wavelet Projection, SP: Spectral Projection, FA: FFT Approximation" + ], + "table_body": "
ModulesPSNR/SSIM↑BRA↑ModulesPSNR/SSIM↑BRA↑
WP(L1)40.51/0.960.92WP(L1)+SPFA42.89/0.990.99
WP(L2)38.15/0.930.87WP(L2)+SPFA36.25/0.920.89
AttacksPSNR/SSIM↑BRA↑AttacksPSNR/SSIM↑BRA↑
Rotate (45°)12.15/21.310.82Rotate (90°)11.15/19.310.65
Blur (0.3)35.01/0.950.98Blur (0.6)30.11/0.910.98
Geo (0.3)12.08/0.500.93Geo (0.6)10.25/0.450.86
", + "bbox": [ + 102, + 88, + 470, + 165 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a57aba76b9cf700dafb4d665320ab98872b6d12a71aba9b2fc4ac70cb83ee599.jpg", + "table_caption": [ + "Table 4. Ablation studies on the proposed SpecGuard for across various configurations, setting $M = 128$ , $r = 100$ , and $s = 20$ ." + ], + "table_footnote": [], + "table_body": "
PlatformPSNR/SSIM↑BRA↑PS FiltersPSNR/SSIM↑BRA↑
Facebook48.56/0.970.97Depth Blur25.25/0.890.85
LinkedIn47.55/0.970.96StyleT.25.12/0.840.85
Instagram48.56/0.980.98Super Zoom36.15/0.880.95
WhatsApp42.10/0.960.97JPEG Artifacts31.01/0.850.94
X (Twitter)49.25/1.000.99Colorize23.15/0.820.92
", + "bbox": [ + 91, + 215, + 483, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Evaluation of SpecGuard's robustness across Photoshop filters and while uploaded on different social media platforms.", + "bbox": [ + 89, + 292, + 482, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Robustness", + "text_level": 1, + "bbox": [ + 89, + 333, + 215, + 348 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate watermarking robustness by analyzing detection performance against a range of diverse and challenging real-world attacks. Results demonstrate the strong robustness of SpecGuard across various attacks. For example, as presented in Tab. 3, against geometric distortions such as cropping and rotation, SpecGuard achieved an Avg P of 0.998 and 0.687, respectively. Similarly, across the combined distortion-based attacks, SpecGuard achieves an overall Avg P of 0.911 and Avg Q of 0.952, ensuring minimal quality loss while maintaining high detection accuracy. Notably, the high values of Q@0.95P and Q@0.7P indicate that SpecGuard can sustain reliable detection at strict performance thresholds, even under aggressive perturbations. Unlike prior methods that struggle with extreme transformations, SpecGuard shows remarkable robustness against regeneration-based attacks like Rinse-2xDiff [6] (an image is noised then denoised by Stable Diffusion v1.4 two times with strength as a number of timesteps, 20-100) and RegenVAE [6], maintaining high Avg P. Similarly, under adversarial attacks, SpecGuard consistently secures watermark detectability, outperforming existing techniques across all tested scenarios. These results establish SpecGuard as a highly robust watermarking approach capable of preserving image integrity even under severe distortions and adversarial manipulations, ensuring watermark reliability across diverse attack types. More details about how the attacks are performed are provided in supplementary material. Further, our results in Fig. 4 highlight the strong robustness of SpecGuard against various distortion attacks compared to other SOTA watermarking methods.", + "bbox": [ + 89, + 356, + 482, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Social Platforms and Photoshop Filters. SpecGuard's robustness when images are shared across social media platforms and subjected to common Photoshop Neural Filters (PNFs) is shown in Tab. 5. SpecGuard consistently maintains high PSNR and SSIM values, with BRA values close to 0.99 on platforms such as X (formally Twitter), Insta", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "gram, and Facebook. Also, it shown strong resilience to various PNFs, such as Super Zoom and JPEG Artifacts achieving BRA of 0.95 and 0.94. The PSNR, SSIM, and BRA values are expected to decrease with the severity of image manipulation, as increased manipulation leads to loss of image authenticity. For example, as we applied $60\\%$ style transfer the PSNR and BRA decreased to 25.12 and 0.85. Similarly, the depth blur which excessively reduces the image clarity also causes the decrease of BRA to 0.85.", + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 238, + 666, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We examined the impact of WP at different levels ( $L_{1}$ and $L_{2}$ ) and its combination with SP using $FA$ in Tab. 4. As observed, the $\\mathrm{WP}(L_1) + \\mathrm{SP}_{FA}$ configuration achieved the highest PSNR and SSIM values of 42.89 and 0.99, respectively, and BRA of 0.99, indicating improved watermark invisibility and robustness. In contrast, using WP alone at either $L_{1}$ or $L_{2}$ resulted in lower BRA, with values of 0.92 and 0.87, respectively, demonstrating that the combined $\\mathrm{WP} + \\mathrm{SP}_{FA}$ approach significantly enhances performance. We also evaluated the robustness of SpecGuard under strong adversarial attacks identified in Tab. 4, such as rotation, blur, and geometric transformations. The results indicate that higher levels of attack severity, such as $90^{\\circ}$ rotation, lead to a more significant drop in PSNR, SSIM, and BRA, with values dropping to 11.15, 19.31, and 0.65, respectively. Despite this, the model shows relatively high resilience under moderate attack intensities, such as $45^{\\circ}$ rotation and low levels of blur and geometric distortion, achieving BRA values as high as 0.93 under geometric transformations at the 0.3 thresholds. More ablations are in the supplementary.", + "bbox": [ + 511, + 258, + 906, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 577, + 633, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose SpecGuard, a novel invisible watermarking method that ensures secure and robust information concealment. Unlike traditional approaches, SpecGuard remains highly resilient against diverse distortions, adversarial attacks, and regeneration-based transformations. Experimental results demonstrate its superior bit recovery accuracy of $99\\%$ maintaining high PSNR. By outperforming SOTA watermarking methods in both detection reliability and imperceptibility, SpecGuard establishes a new benchmark for watermarking under real-world constraints.", + "bbox": [ + 511, + 603, + 906, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 768, + 671, + 786 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was partly supported by Institute for Information & communication Technology Planning & evaluation (IITP) grants funded by the Korean government MSIT: (RS-2022-II221199, RS-2022-II220688, RS-2019-II190421, RS-2023-00230337, RS-2024-00356293, RS-2024-00437849, RS-2021-II212068, RS-2025-02304983, and RS-2025-02263841).", + "bbox": [ + 511, + 794, + 903, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17991", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Edward H Adelson, Eero Simoncelli, and Rajesh Hingorani. Orthogonal pyramid transforms for image coding. In Visual Communications and image processing II, pages 50-58. SPIE, 1987. 2", + "[2] Mahdi Ahmadi, Alireza Norouzi, Nader Karimi, Shadrokh Samavi, and Ali Emami. Redmark: Framework for residual diffusion watermarking based on deep networks. Expert Systems with Applications, 146:113157, 2020. 2", + "[3] Inzamamul Alam, Md Tanvir Islam, and Simon S Woo. Specxnet: A dual-domain convolutional network for robust deepfake detection. In Proceedings of the 33nd ACM International Conference on Multimedia, 2025. 2", + "[4] Inzamamul Alam, Md Tanvir Islam, and Simon S Woo. Saliency-aware diffusion reconstruction for effective invisible watermark removal. In Companion Proceedings of the ACM on Web Conference 2025, pages 849-853, 2025. 1", + "[5] Aashutosh AV, Srijan Das, Abhijit Das, et al. Latent flow diffusion for deepfake video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3781-3790, 2024. 1", + "[6] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436, 2018. 8", + "[7] Tali Dekel, Michael Rubinstein, Ce Liu, and William T Freeman. On the effectiveness of visible watermarks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2146-2154, 2017. 2", + "[8] Mucong Ding, Tahseen Rabbani, Bang An, Aakriti Agrawal, Yuancheng Xu, Chenghao Deng, Sicheng Zhu, Abdirisak Mohamed, Yuxin Wen, Tom Goldstein, et al. Waves: Benchmarking the robustness of image watermarks. In ICLR 2024 Workshop on Reliable and Responsible Foundation Models, 2024. 1, 6", + "[9] Hubert Etienne. The future of online trust (and why deepfake is advancing it). AI and Ethics, 1(4):553-562, 2021. 1", + "[10] Han Fang, Yupeng Qiu, Kejiang Chen, Jiyi Zhang, Weiming Zhang, and Ee-Chien Chang. Flow-based robust watermarking with invertible noise layer for black-box distortions. In Proceedings of the AAAI conference on artificial intelligence, pages 5054-5061, 2023. 6", + "[11] Jianwei Fei, Zhihua Xia, Benedetta Tondi, and Mauro Barni. Supervised gan watermarking for intellectual property protection. In 2022 IEEE International Workshop on Information Forensics and Security (WIFS), pages 1-6. IEEE, 2022. 2", + "[12] S. A. Fulling. The local geometric asymptotics of continuum eigenfunction expansions. ii. one-dimensional systems. SIAM Journal on Mathematical Analysis, 14(4):605-623, 1983. 2", + "[13] Mahdieh Ghazvini, Elham Mohamadi Hachrood, and Mojdeh Mirzadi. An improved image watermarking method in frequency domain. Journal of Applied Security Research, 12 (2):260-275, 2017. 2", + "[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1", + "[15] Md Tanvir Islam, Ik Hyun Lee, Ahmed Ibrahim Alzahrani, and Khan Muhammad. Mexic: A meta ensemble explainable approach for ai-synthesized fake image classification. *Alexandria Engineering Journal*, 116:351–363, 2025. 2", + "[16] K Jayashre and M Amsaprabhaa. Safeguarding media integrity: A hybrid optimized deep feature fusion based deepfake detection in videos. Computers & Security, 142:103860, 2024. 1", + "[17] Zhaoyang Jia, Han Fang, and Weiming Zhang. Mbrs: Enhancing robustness of dnn-based watermarking by minibatch of real and simulatedJPEG compression. In Proceedings of the 29th ACM international conference on multimedia, pages 41-49, 2021. 6", + "[18] Hoon Kang and Joonsoo Ha. Projection spectral analysis. International Journal of Control, Automation and Systems, 13(6):1530-1537, 2015. 1, 2, 4", + "[19] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations (ICLR), 2018. 6", + "[20] SS Kelkar, LL Grigsby, and J Langsner. An extension of parseval's theorem and its use in calculating transient energy in the frequency domain. IEEE Transactions on Industrial Electronics, (1):42-45, 1983. 2, 5", + "[21] Varsha Kishore, Xiangyu Chen, Yan Wang, Boyi Li, and Kilian Q Weinberger. Fixed neural network steganography: Train the images, not the network. In International Conference on Learning Representations, 2021. 2", + "[22] Jae-Eun Lee, Young-Ho Seo, and Dong-Wook Kim. Convolutional neural network-based digital image watermarking adaptive to the resolution of image and watermark. Applied Sciences, 10(19):6854, 2020. 2", + "[23] Yicheng Leng, Chaowei Fang, Gen Li, Yixiang Fang, and Guanbin Li. Removing interference and recovering content imaginatively for visible watermark removal. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2983-2990, 2024. 2", + "[24] Xiao Li, Liquan Chen, Ju Jia, Zhongyuan Qin, and Zhangjie Fu. A lightweight image forgery prevention scheme for IoT using gan-based steganography. IEEE Transactions on Industrial Informatics, 2024. 2", + "[25] Dongdong Lin, Benedetta Tondi, Bin Li, and Mauro Barni. A cyclegan watermarking method for ownership verification. IEEE Transactions on Dependable and Secure Computing, 2024. 2", + "[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5, 6", + "[27] Yang Liu, Zhen Zhu, and Xiang Bai. Wdnet: Watermark-decomposition network for visible watermark removal. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 3685-3693, 2021. 2" + ], + "bbox": [ + 517, + 93, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17992", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Xiyang Luo, Ruohan Zhan, Huiwen Chang, Feng Yang, and Peyman Milanfar. Distortion agnostic deep watermarking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13548-13557, 2020. 2", + "[29] Thanh Thi Nguyen, Quoc Viet Hung Nguyen, Dung Tien Nguyen, Duc Thanh Nguyen, Thien Huynh-The, Saeid Nahavandi, Thanh Tam Nguyen, Quoc-Viet Pham, and Cuong M Nguyen. Deep learning for deepfakes creation and detection: A survey. Computer Vision and Image Understanding, 223:103525, 2022. 1", + "[30] Guangyu Nie, Changhoon Kim, Yezhou Yang, and Yi Ren. Attributing image generative models using latent fingerprints. In International Conference on Machine Learning, pages 26150-26165. PMLR, 2023. 2", + "[31] Li Niu, Xing Zhao, Bo Zhang, and Liqing Zhang. Fine-grained visible watermark removal. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12770-12779, 2023. 2", + "[32] Konstantin A Pantserev. The malicious use of ai-based deepfake technology as the new threat to psychological security and political stability. Cyber defence in the age of AI, smart societies and augmented humanity, pages 37-55, 2020. 1", + "[33] Ram Shankar Pathak. The wavelet transform. Springer Science & Business Media, 2009. 1, 2, 4", + "[34] Fernandez Pierre, Alexandre Sablayrolles, Teddy Furon, Hervé Jégou, and Matthijs Douze. Watermarking images in self-supervised latent spaces. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3054-3058. IEEE, 2022. 2", + "[35] Fernandez Pierre, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22466-22477, 2023. 1, 6, 7", + "[36] Tong Qiao, Yuyan Ma, Ning Zheng, Hanzhou Wu, Yanli Chen, Ming Xu, and Xiangyang Luo. A novel model watermarking for protecting generative adversarial network. Computers & Security, 127:103102, 2023. 2", + "[37] Sunpreet Sharma, Ju Jia Zou, Gu Fang, Pancham Shukla, and Weidong Cai. A review of image watermarking for identity protection and verification. Multimedia Tools and Applications, 83(11):31829-31891, 2024. 2", + "[38] Qingtang Su, Huanying Wang, DeCheng Liu, Zihan Yuan, and Xueting Zhang. A combined domain watermarking algorithm of color image. Multimedia Tools and Applications, 79(39):30023-30043, 2020. 2", + "[39] Qingtang Su, Xueting Zhang, and Huanying Wang. A blind color image watermarking algorithm combined spatial domain and svd. International Journal of Intelligent Systems, 37(8):4747-4771, 2022. 2", + "[40] Ruizhou Sun, Yukun Su, and Qingyao Wu. Denet: disentangled embedding network for visible watermark removal. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2411-2419, 2023. 2", + "[41] Matthew Tancik, Ben Mildenhall, and Ren Ng. Stegastamp: Invisible hyperlinks in physical photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2117-2126, 2020. 1, 6, 7" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] D Vaishnavi and TS Subashini. Robust and invisible image watermarking in rgb color space using svd. Procedia Computer Science, 46:1770-1777, 2015. 2", + "[43] Guanjie Wang, Zehua Ma, Chang Liu, Xi Yang, Han Fang, Weiming Zhang, and Nenghai Yu. Must: Robust image watermarking for multi-source tracing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5364-5371, 2024. 6", + "[44] Huanying Wang and Qingtang Su. A color image watermarking method combined qr decomposition and spatial domain. Multimedia Tools and Applications, 81(26):37895-37916, 2022. 2", + "[45] Zilan Wang, Junfeng Guo, Jiacheng Zhu, Yiming Li, Heng Huang, Muhao Chen, and Zhengzhong Tu. Sleepermark: Towards robust watermark against fine-tuning text-to-image diffusion models. arXiv preprint arXiv:2412.04852, 2024. 6", + "[46] Zijie J Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. Diffusional: A large-scale prompt gallery dataset for text-to-image generative models. arXiv preprint arXiv:2210.14896, 2022.5", + "[47] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-rings watermarks: Invisible fingerprints for diffusion images. In Advances in Neural Information Processing Systems, pages 58047-58063. Curran Associates, Inc., 2023. 1, 6, 7", + "[48] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-rings watermarks: Invisible fingerprints for diffusion images. Advances in Neural Information Processing Systems, 36, 2024. 1", + "[49] Zijin Yang, Kai Zeng, Kejiang Chen, Han Fang, Weiming Zhang, and Nenghai Yu. Gaussian shading: Provable performance-lossless image watermarking for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12162-12171, 2024. 1, 6", + "[50] Zihan Yuan, Qingtang Su, Decheng Liu, and Xueting Zhang. A blind image watermarking scheme combining spatial domain and frequency domain. The visual computer, 37:1867-1881, 2021. 2", + "[51] Chaoning Zhang, Philipp Benz, Adil Karjauv, Geng Sun, and In So Kweon. Udh: Universal deep hiding for steganography, watermarking, and light field messaging. Advances in Neural Information Processing Systems, 33:10223-10234, 2020. 2", + "[52] Xuanyu Zhang, Runyi Li, Jiwen Yu, Youmin Xu, Weiqi Li, and Jian Zhang. Editguard: Versatile image watermarking for tamper localization and copyright protection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11964-11974, 2024. 6", + "[53] Jiren Zhu, Russell Kaplan, Justin Johnson, and Li Fei-Fei. Hidden: Hiding data with deep networks. European Conference on Computer Vision (ECCV), 2018. 1, 2, 6" + ], + "bbox": [ + 516, + 92, + 906, + 825 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17993", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SpecGuard_ Spectral Projection-based Advanced Invisible Watermarking/1fcb8d1e-6633-4c96-8cd9-9ceeae669d64_model.json b/2025/SpecGuard_ Spectral Projection-based Advanced Invisible Watermarking/1fcb8d1e-6633-4c96-8cd9-9ceeae669d64_model.json new file mode 100644 index 0000000000000000000000000000000000000000..86dcc9fa299a33fbea0004d428f4f4d228e6a523 --- /dev/null +++ b/2025/SpecGuard_ Spectral Projection-based Advanced Invisible Watermarking/1fcb8d1e-6633-4c96-8cd9-9ceeae669d64_model.json @@ -0,0 +1,2343 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.131, + 0.872, + 0.154 + ], + "angle": 0, + "content": "SpecGuard: Spectral Projection-based Advanced Invisible Watermarking" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.181, + 0.777, + 0.218 + ], + "angle": 0, + "content": "Inzamamul Alam, Md Tanvir Islam, Simon S. Woo*, Khan Muhammad Sungkyunkwan University" + }, + { + "type": "text", + "bbox": [ + 0.275, + 0.22, + 0.719, + 0.234 + ], + "angle": 0, + "content": "{inzil5, tanvirnwu, swoo, khanmuhammad}@g.skku.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.302, + 0.486, + 0.681 + ], + "angle": 0, + "content": "Watermarking embeds imperceptible patterns into images for authenticity verification. However, existing methods often lack robustness against various transformations primarily including distortions, image regeneration, and adversarial perturbation, creating real-world challenges. In this work, we introduce SpecGuard, a novel watermarking approach for robust and invisible image watermarking. Unlike prior approaches, we embed the message inside hidden convolution layers by converting from the spatial domain to the frequency domain using spectral projection of a higher frequency band that is decomposed by wavelet projection. Spectral projection employs Fast Fourier Transform approximation to transform spatial data into the frequency domain efficiently. In the encoding phase, a strength factor enhances resilience against diverse attacks, including adversarial, geometric, and regeneration-based distortions, ensuring the preservation of copyrighted information. Meanwhile, the decoder leverages Parseval's theorem to effectively learn and extract the watermark pattern, enabling accurate retrieval under challenging transformations. We evaluate the proposed SpecGuard based on the embedded watermark's invisibility, capacity, and robustness. Comprehensive experiments demonstrate the proposed SpecGuard outperforms the state-of-the-art models. To ensure reproducibility, the full code is released on GitHub." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.697, + 0.222, + 0.712 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.723, + 0.485, + 0.875 + ], + "angle": 0, + "content": "With the rapid advancement of digital media and artificial intelligence, concerns regarding image authenticity, copyright protection, and content integrity have become more challenging than ever [9, 16, 32]. Moreover, the widespread availability of the latest image manipulation tools [4, 5, 14] enables malicious tamperers to easily forge and redistribute digital content without authorization, posing a significant threat to ownership verification [29]. This growing risk emphasizes the need for reliable techniques for secure authentication and detection of unauthorized manipulation." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.269, + 0.907, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.436, + 0.898, + 0.451 + ], + "angle": 0, + "content": "Figure 1. Image authentication using our proposed SpecGuard." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.907, + 0.75 + ], + "angle": 0, + "content": "Recently, invisible watermarking has gained significant attention as a prominent defense mechanism for media authentication by embedding invisible messages into images to verify authenticity [35, 48]. In fact, invisible watermarks are preferred for preserving image quality and resisting tampering. These watermarks are unique to the creator and enable tamper verification by comparing the retrieved watermark to the original, as the high-level process is presented in Fig. 1. Traditional watermarking methods often rely on transformation techniques [18, 33]. Deep learning approaches like StegaStamp [41], Stable Signature [35], and HiDDeN [53] provide end-to-end solutions for message embedding. However, these methods often struggle with fragility in handling common image processing operations such as resizing, cropping, compression, and noise addition, which can distort or erase the embedded watermark. Additionally, the performance of watermark embedding and extraction often remains vulnerable to attacks with noise injection, blurring, contrasting, and rotation [8]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To address the aforementioned challenges, we introduce a novel robust, and invisible image watermarking method named SpecGuard. SpecGuard is designed to overcome the fundamental trade-offs [8] between imperceptibility, and robustness. Our proposed SpecGuard strategically embeds watermark information in the spectral domain, leveraging wavelet-based decomposition to distribute the watermark across high-frequency components. Unlike traditional frequency domain watermarking techniques [47, 49] that are easily disrupted by common image manipulations, Spec-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.425, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author: swoo@g.skku.edu (Simon S. Woo)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17984" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "Guard maintains imperceptibility while significantly improving robustness against a wide range of transformations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.126, + 0.484, + 0.219 + ], + "angle": 0, + "content": "Overall, our proposed SpecGuard addresses the current limitations of the previous watermarking methods by providing a robust, imperceptible watermarking technique that maintains integrity under diverse manipulations, significantly enhancing digital content security and authenticity verification. Our key contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.227, + 0.483, + 0.301 + ], + "angle": 0, + "content": "- We introduce a novel watermarking approach that embeds message bits in high-frequency spectral components via wavelet and spectral projection inside hidden convolutional layers, ensuring robustness against various transformations and adversarial attacks." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.302, + 0.484, + 0.378 + ], + "angle": 0, + "content": "- We adapt Parseval's theorem [20] as a learnable threshold to optimize SpecGuard and spectral masking for robust watermark bit recovery under diverse transformations including distortions, regeneration, and adversarial attacks, proven through the experimental results." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.379, + 0.484, + 0.439 + ], + "angle": 0, + "content": "- Our extensive evaluations demonstrate SpecGuard's superior bit embedding capacity and producing better invisible watermarked images, surpassing the performance of state-of-the-art (SOTA) methods." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.227, + 0.484, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.464, + 0.241, + 0.48 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.494, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Watermarking an image has been a widely researched topic for securing the ownership and verifying authenticity of digital content [37]. Traditional watermarking techniques typically embed invisible [42] or visible [7] watermarks into images, which can later be extracted or detected to verify the content's originality. These methods can be broadly classified into spatial-domain [39, 44] and frequency-domain [13] watermarking, while some are based on combined methods [38, 50]. However, researchers recently proposed many advanced models [23, 27, 31, 40] for effective watermark removal. To face this growing challenge, researchers introduced different methods [2, 3, 15, 22, 28, 51, 53] as alternatives to deep learning-based encoders or decoders to produce more robust image watermarking. Furthermore, iterative models have demonstrated competitive performance [21, 34], particularly in robustness against a wide range of transformations. In addition, with the rise of generative methods, researchers used the watermark-labeled data for training to learn how to produce watermarks [11, 24]. Also, models that combine generative methods with watermarking techniques show promise in effective image watermarking [25, 30, 36]. However, such approaches face limitations such as increased computational complexity and longer processing times. These approaches are also more vulnerable to adversarial attacks that can target and distort the embedded watermark without altering the content visibly." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.795, + 0.108 + ], + "angle": 0, + "content": "3. Proposed Method: SpecGuard" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.115, + 0.907, + 0.178 + ], + "angle": 0, + "content": "We introduce SpecGuard, as illustrated in Fig. 2, which involves two fundamental modules: an \"Encoder\" for embedding the watermark and a \"Decoder\" for accurately extracting the watermark detailed in the following sections." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.185, + 0.616, + 0.2 + ], + "angle": 0, + "content": "3.1. Encoder" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.207, + 0.906, + 0.297 + ], + "angle": 0, + "content": "By targeting high-frequency components, the encoder integrates a binary message \\( M \\) into the cover image \\( I \\). Using wavelet projection (WP) [33] and a Fast Fourier Transform (FFT)-based spectral projection (SP) [18] approximation, the message \\( M \\) is inserted into specific frequency bands, minimizing perceptual impact." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.298, + 0.906, + 0.343 + ], + "angle": 0, + "content": "Wavelet Projection. We use a wavelet projection to capture frequency and spatial localization features that describe an image across different scales, as shown in Eq. (1):" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.35, + 0.907, + 0.383 + ], + "angle": 0, + "content": "\\[\nW (a, b) = \\frac {1}{\\sqrt {| a |}} \\int_ {- \\infty} ^ {\\infty} f (x) \\psi \\left(\\frac {x - b}{a}\\right) d x, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.389, + 0.906, + 0.449 + ], + "angle": 0, + "content": "where \\(a \\in \\mathbb{R} \\setminus \\{0\\}\\), and \\(b \\in \\mathbb{R}\\) denote the scaling and translation parameters, respectively. Here, \\(\\psi_{a,b}(x)\\) represents a rescaled and translated form of the mother wavelet \\(\\psi\\), defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.455, + 0.906, + 0.488 + ], + "angle": 0, + "content": "\\[\n\\psi_ {a, b} (x) = \\psi \\left(\\frac {x - b}{a}\\right) \\cdot \\frac {1}{\\sqrt {| a |}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.496, + 0.906, + 0.622 + ], + "angle": 0, + "content": "where \\(\\frac{1}{\\sqrt{|a|}}\\) functions as a normalization factor, guaranteeing that the energy of the wavelet is invariant to the scaling parameter \\(a\\). Minimal values of \\(a\\) compress the wavelet, enabling the inspection of high-frequency components, whereas greater values of \\(a\\) elongate the wavelet, promoting low-frequency analysis. Since each mother wavelet \\(\\psi\\) is built with zero mean and finite energy [12], it guarantees to maintain stability as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.63, + 0.905, + 0.66 + ], + "angle": 0, + "content": "\\[\n\\int_ {- \\infty} ^ {\\infty} \\psi (x) d x = 0, \\quad \\int_ {- \\infty} ^ {\\infty} | \\psi (x) | ^ {2} d x < \\infty , \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.667, + 0.906, + 0.728 + ], + "angle": 0, + "content": "where the wavelet projection from Eq. (1) decomposes the input into orthogonal wavelet sets using discrete scales and translations. For \\(2D\\) inputs, the scaled and translated basis elements [1] are defined for each coordinate pair \\((u,v)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.736, + 0.915, + 0.752 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} _ {L L} = \\phi (u, v) = \\phi (u) \\phi (v), \\qquad \\mathbf {S} _ {L H} = \\psi_ {H} (u, v) = \\psi (u) \\phi (v),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.754, + 0.915, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} _ {H L} = \\psi_ {V} (u, v) = \\phi (u) \\psi (v), \\quad \\mathbf {S} _ {H H} = \\psi_ {D} (u, v) = \\psi (u) \\psi (v), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.782, + 0.907, + 0.841 + ], + "angle": 0, + "content": "where, \\(H\\), \\(V\\), and \\(D\\) represent the horizontal, vertical, and diagonal decomposition direction, respectively. To depict the image at different resolutions, we define scaling and wavelet functions at scale \\(j\\) as shown below:" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.849, + 0.905, + 0.884 + ], + "angle": 0, + "content": "\\[\n\\phi_ {j, m, n} (u, v) = 2 ^ {j / 2} \\phi \\left(u - \\frac {m}{2 ^ {j}}, v - \\frac {n}{2 ^ {j}}\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.88, + 0.853, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\psi_ {j, m, n} ^ {d} (u, v) = 2 ^ {j / 2} \\psi^ {d} \\left(u - \\frac {m}{2 ^ {j}}, v - \\frac {n}{2 ^ {j}}\\right),\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17985" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.088, + 0.905, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.311, + 0.908, + 0.342 + ], + "angle": 0, + "content": "Figure 2. Architecture of the proposed SpecGuard watermarking method involves encoding a binary message \\( M \\) into the high-frequency band of the cover image \\( I \\) using wavelet and spectral projection and learning to decode the embedded message." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.355, + 0.484, + 0.477 + ], + "angle": 0, + "content": "where \\(d \\in \\{H, V, D\\}\\) is the wavelet function direction that serves as discrete basis elements for multi-resolution analysis, capturing details across frequency bands and spatial locations. In Eq. (6), \\(T_{m,n}\\) denotes the intensity or pixel value of the cover image \\(I\\) at spatial coordinates \\((m, n)\\). The discrete scaling function \\(W_{\\phi}(j, u, v)\\) (approximation at scale \\(j\\)) and the detail coefficients \\(W_{\\psi}^{d}(j, u, v)\\) for each direction are computed accordingly as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.485, + 0.476, + 0.522 + ], + "angle": 0, + "content": "\\[\nW _ {\\phi} (j, u, v) = \\frac {1}{l} \\sum_ {m = 0} ^ {l - 1} \\sum_ {n = 0} ^ {l - 1} T _ {m, n} \\phi \\left(m - u \\cdot 2 ^ {- j}, n - v \\cdot 2 ^ {- j}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.524, + 0.484, + 0.572 + ], + "angle": 0, + "content": "\\[\nW _ {\\psi} ^ {d} (j, u, v) = \\frac {1}{l} \\sum_ {m = 0} ^ {l - 1} \\sum_ {n = 0} ^ {l - 1} T _ {m, n} \\psi^ {d} \\left(m - u \\cdot 2 ^ {- j}, n - v \\cdot 2 ^ {- j}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.574, + 0.484, + 0.62 + ], + "angle": 0, + "content": "with \\(l\\) as the discrete region dimension, these coefficients capture multi-scale, multi-orientation image information, forming the basis of spectral features as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.627, + 0.483, + 0.658 + ], + "angle": 0, + "content": "\\[\n\\beta_ {j} = \\bigcup_ {d \\in \\{H, V, D \\}} \\left(W _ {\\phi} (j, u, v) \\cup W _ {\\psi} ^ {d} (j, u, v)\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.667, + 0.483, + 0.711 + ], + "angle": 0, + "content": "This feature set \\(\\beta_{j}\\) captures key frequency and spatial details across resolutions, forming the foundation for the watermark embedding process of our SpecGuard." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.712, + 0.483, + 0.772 + ], + "angle": 0, + "content": "Selective Frequency Band Decomposition. To refine the embedding process, we segment the data into distinct frequency bands. The decomposition level \\(\\kappa\\) is determined by the image complexity, calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.781, + 0.483, + 0.799 + ], + "angle": 0, + "content": "\\[\n\\kappa = \\left\\lfloor \\sqrt {\\log (1 + N)} \\right\\rfloor , \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.807, + 0.484, + 0.866 + ], + "angle": 0, + "content": "where \\(N\\) denotes the total pixel count in the cover image \\(I\\). And, each component \\(\\beta_{j}\\) falls within a unique frequency band, yielding a total of \\(1 + 3\\kappa\\) distinct frequency bands as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.874, + 0.483, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\beta_ {j} = \\phi_ {j} (u, v) \\cup \\bigcup_ {d \\in \\{H, V, D \\}} \\psi_ {j} ^ {d} (u, v). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.355, + 0.907, + 0.446 + ], + "angle": 0, + "content": "The components \\(\\beta_{j}\\), consisting of scaling functions \\(\\phi_j(u,v)\\) and wavelet functions \\(\\psi_j^d (u,v)\\), capture specific spatial frequency bands, enabling targeted high-frequency embedding. We translate the WP into disjoint intervals representing a unique frequency range to approximate the segmentation in the frequency domain:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.455, + 0.907, + 0.487 + ], + "angle": 0, + "content": "\\[\n\\beta_ {j} = \\left\\{W _ {\\psi} ^ {d} (u, v) \\mid u, v \\in \\left(\\frac {j \\cdot L}{\\kappa}, \\frac {(j + 1) \\cdot L}{\\kappa}\\right) \\right\\}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.496, + 0.906, + 0.572 + ], + "angle": 0, + "content": "where, \\(L\\) is the dimension of \\(S_{HH}\\), and \\(W_{\\psi}^{d}(u,v)\\) represents wavelet values within segmented intervals. This frequency band partitioning mimics the frequency selectivity of wavelet sub-bands, enabling effective targeting of high-frequency regions for optimal embedding." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.572, + 0.906, + 0.648 + ], + "angle": 0, + "content": "Approximation of Spectral Projection. We first apply spectral projection on the \\(S_{HH}\\) sub-band, transforming it into the spectral domain. Given a matrix \\(T(x,y)\\) representing pixel intensities in \\(S_{HH}\\), the spectral projection computes the spectral components \\(\\zeta(u,v)\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.658, + 0.905, + 0.704 + ], + "angle": 0, + "content": "\\[\n\\zeta (u, v) = \\frac {1}{L ^ {2}} \\sum_ {x} \\sum_ {y} T (x, y) \\cdot \\exp \\left(- i \\frac {2 \\pi}{L} (x \\cdot u + y \\cdot v)\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.705, + 0.906, + 0.764 + ], + "angle": 0, + "content": "where \\(L\\) denotes the dimension of \\(S_{HH}\\), \\(T(x,y)\\) provides the intensity at each coordinate \\((x,y)\\) which is equivalent to \\(W_{\\psi}^{d}(u,v)\\) in Eq. (6), \\(i\\) is the imaginary unit, and \\((u,v)\\) are the spectral coordinates." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.907, + 0.902 + ], + "angle": 0, + "content": "To approximate the spectral components using the FFT, we create a symmetrically extended version \\(\\tilde{T}(x,y)\\) of the original \\(N\\times N\\) matrix \\(T(x,y)\\). This extension is achieved by mirroring \\(T(x,y)\\) along its boundaries, doubling its size to \\(2N\\times 2N\\). Specifically, the original matrix occupies the top-left quadrant, with the remaining quadrants filled by reflecting \\(T(x,y)\\) horizontally, vertically, and diagonally, respectively. This symmetric structure ensures that the FFT yields only real values, allowing the spectral coefficients to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17986" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.123 + ], + "angle": 0, + "content": "be extracted directly from the real part of the FFT operation. Then, we apply the 2D FFT to \\(\\tilde{T} (x,y)\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.131, + 0.487, + 0.177 + ], + "angle": 0, + "content": "\\[\nF (u, v) = \\frac {1}{(2 N) ^ {2}} \\sum_ {x} \\sum_ {y} \\tilde {T} (x, y) \\cdot \\exp \\left(- i \\frac {2 \\pi}{2 N} (x \\cdot u + y \\cdot v)\\right) \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.179, + 0.484, + 0.209 + ], + "angle": 0, + "content": "The SP coefficients are then approximated by taking the real part (Re) of \\( F \\) in the original \\( N \\times N \\) region as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.223, + 0.483, + 0.239 + ], + "angle": 0, + "content": "\\[\n\\zeta (u, v) \\approx \\operatorname {R e} (F (u, v)), \\quad 0 \\leq u, v < N. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.242, + 0.484, + 0.317 + ], + "angle": 0, + "content": "Applying Eq. (13) to the sub-bands extracted from wavelet projection in Eq. (6), we achieve a computationally efficient spectral projection by leveraging the FFT approximation on a symmetrically extended matrix, maintaining effective embedding properties within the spectral domain." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.484, + 0.469 + ], + "angle": 0, + "content": "SpecGuard Embedding Process. The embedding process integrates the binary message \\(M\\) into the high-frequency band \\(S_{HH}\\) of the cover image \\(I\\), enhancing robustness and imperceptibility through wavelet and spectral projection. Using the Eq. (6) and Eq. (13), the cover image \\(I\\) is decomposed into sub-bands \\(S_{LL}, S_{LH}, S_{HL}\\), and \\(S_{HH}\\) within spectral domain, with \\(S_{HH}\\) providing high-frequency details for embedding. A variable number \\(k\\) of convolutional layers with a \\(K \\times K\\) kernel, followed by LeakyReLU activation, are recursively applied to \\(S_{HH}\\) to refine features:" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.487, + 0.483, + 0.505 + ], + "angle": 0, + "content": "\\[\nS _ {H H} ^ {(n + 1)} = \\text {L e a k y R e L U} \\left(\\operatorname {C o n v} _ {2 D} \\left(S _ {H H} ^ {(n)}, K\\right)\\right), \\quad n = 1, \\dots , k. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.512, + 0.483, + 0.543 + ], + "angle": 0, + "content": "The final output \\( S_{HH}^{(n + 1)} \\) from Eq. (14) represents the modified high-frequency band, primed for embedding." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.544, + 0.483, + 0.621 + ], + "angle": 0, + "content": "The message \\(M\\), represented as a binary vector of length \\(l\\) (\\(M \\in \\{0,1\\}^l\\)), with batch size \\(b\\) and message length \\(l\\), is reshaped and expanded across channels \\(c\\) to align with \\(S_{HH}^{(n+1)}\\). This ensures \\(M_{\\mathrm{expanded}}\\) conforms to the dimension \\([b,c,l]\\), where each message is structured accordingly." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.622, + 0.483, + 0.771 + ], + "angle": 0, + "content": "To localize the embedding, we create a radial mask centered at \\( (c_x, c_y) = \\left(\\frac{h}{2}, \\frac{w}{2}\\right) \\), where \\( h \\) and \\( w \\) represent the height and width of \\( S_{HH} \\). The Euclidean distance \\( D(x_i, y_i) \\) from the center \\( (c_x, c_y) \\) is computed for each coefficient \\( (x_i, y_i) \\). A binary mask is then generated within the predefined radius \\( r \\) based on the distance \\( D(x_i, y_i) \\), such that if \\( D(x_i, y_i) \\leq r \\), the mask value is set to 1, allowing embedding in the corresponding region. Otherwise, the mask value is 0, restricting embedding to areas within a specified radius \\( r \\), ensuring focus on high-frequency regions." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.772, + 0.483, + 0.803 + ], + "angle": 0, + "content": "For each coordinate \\((x_{i},y_{i})\\) where mask \\((x_{i},y_{i})\\) is 1 and \\(W_{c}\\in c\\), the embedding operation is performed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.811, + 0.483, + 0.83 + ], + "angle": 0, + "content": "\\[\nS _ {H H} ^ {(n + 1)} [:, W _ {c}, x _ {i}, y _ {i} ] + = M _ {\\text {e x p a n d e d}} [:, W _ {c}, i ] \\cdot s, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.839, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\( s \\) is the strength factor controlling embedding intensity and invisibility. After embedding, the modified coefficients \\( S_{HH}^{(n + 1)} \\) undergo a final convolution and LeakyReLU using Eq. (14), by setting the value of \\( k = 1 \\) to harmonize" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.257 + ], + "angle": 0, + "content": "the embedded message. Following this approach, SpecGuard embeds the message into the spectral domain in a transformed form, differing from its original input representation. By blending the message seamlessly into the spectral space based on the \\(r\\), \\(s\\), and \\(W_{c}\\), it becomes inherently concealed within the domain, rendering its presence imperceptible. Without knowledge of \\(r\\), \\(s\\), and \\(W_{c}\\), it becomes exceedingly challenging to localize the embedded message, further enhancing the security of the system. This transformation ensures the embedding process remains opaque to any adversarial attacker, making SpecGuard black-box." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.258, + 0.906, + 0.409 + ], + "angle": 0, + "content": "Reconstruction. SpecGuard encoder reconstructs the watermarked image \\( I_{\\mathrm{embedded}} \\) by inverse transformation restoring \\( S_{HH} \\) back into the spatial domain. The reconstruction process integrates the inverse wavelet projection (IWP) [33] and inverse spectral projection (ISP) [18], ensuring the embedded modifications are correctly translated into the spatial domain. To reconstruct the spatial domain image, \\( S_{HH} \\) is combined with the other sub-bands \\( S_{LL}, S_{LH} \\), and \\( S_{HL} \\). For the SP embedded in \\( S_{HH} \\), the ISP is applied to reconstruct \\( S_{HH} \\) to spatial domain as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.414, + 0.905, + 0.449 + ], + "angle": 0, + "content": "\\[\nS _ {H H} (x, y) = \\sum_ {u = 0} ^ {L - 1} \\sum_ {v = 0} ^ {L - 1} \\zeta (u, v) \\cdot \\exp \\left(i \\frac {2 \\pi}{L} (x \\cdot u + y \\cdot v)\\right), \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.453, + 0.906, + 0.514 + ], + "angle": 0, + "content": "where \\(\\zeta(u, v)\\) represents spectral coefficients from the embedding process, \\(L\\) denotes the dimension of \\(S_{HH}\\), and \\((x, y)\\) are spatial coordinates. SpecGuard then reconstructs the watermarked image \\(I_{\\mathrm{embedded}}\\) using the IWP as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.518, + 0.905, + 0.533 + ], + "angle": 0, + "content": "\\[\nI _ {\\text {e m b e d d e d}} (x, y) = \\operatorname {I W P} \\left(S _ {L L}, S _ {L H}, S _ {H L}, S _ {H H}\\right). \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.538, + 0.906, + 0.614 + ], + "angle": 0, + "content": "This process seamlessly embeds the watermark message \\(M\\) in the spectral domain, preserving the cover image \\(I\\)'s integrity. The inverse transformations that are expressed in Eq. (16) and Eq. (17) fully restore visual quality, maintaining all frequency components." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.623, + 0.615, + 0.637 + ], + "angle": 0, + "content": "3.2.Decoder" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.906, + 0.795 + ], + "angle": 0, + "content": "As shown in Algorithm 1, SpecGuard decoding process starts by applying wavelet projection ( Eq. (1)) to the watermarked image \\( I_{\\mathrm{embedded}} \\), separating it into low and high-frequency bands, where the high-frequency band \\( S_{D_{HH}}^{high} \\) contains the embedded message similar to the process in the encoding phase, particularly in Eq. (6). An approximation of the spectral projection using FFT as shown in Eq. (13) is then applied to \\( S_{D_{HH}}^{high} \\) returning the transformed data \\( S_{D_{HH}}^{\\mathrm{sp}} \\). Then, \\( S_{D_{HH}} \\) is further refined through convolutional layers that captures the local features for message extraction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To extract the message, a radial mask is created to isolate high-frequency areas within \\( S_{D_{HH}} \\), targeting the embedded regions based on their distance from the center. The masked values are compared against a learnable threshold \\( \\theta \\) to decode each bit of the hidden message \\( D_{\\mathrm{M}} \\). Here, \\( \\theta \\) serves as a threshold that adapts to the spectral patterns across the entire image, learning the distinct characteristics" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17987" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.092, + 0.092, + 0.483, + 0.12 + ], + "angle": 0, + "content": "Algorithm 1 SpecGuard decoder with wavelet, spectral projection with FFT approximation, and learnable threshold." + }, + { + "type": "algorithm", + "bbox": [ + 0.095, + 0.124, + 0.484, + 0.527 + ], + "angle": 0, + "content": "1: Input: Watermarked image \\(I_{\\mathrm{embedded}}\\) , learnable \\(\\theta\\) message length \\(l\\) \nradius \\(r\\) ,watermark channel \\(W_{c}\\) \n2: Output: Decoded binary message \\(D_{\\mathrm{M}}\\) \n3: Procedure: Apply Wavelet Projection on \\(I_{\\mathrm{embedded}}\\) to obtain \\(S_{D_{LL}}\\) (low-frequency) and \\(S_{D_{HH}}^{\\mathrm{high}}\\) (high-frequency) \n4: Procedure: Spectral approximation with FFT \\((S_{D_{HH}}^{\\mathrm{high}})\\) .. \n5: Separate even and odd indices: \\(v = [x_{\\mathrm{even}},\\) reverse \\((x_{\\mathrm{odd}})]\\) \n6: Compute FFT on \\(v\\) .. \\(V_{\\mathrm{complex}} = \\mathrm{FFT}(v)\\) \n7: \\(V_{\\mathrm{real}} = V_{\\mathrm{complex}}\\cdot \\left[\\cos \\left(\\frac{-\\pi k}{2N}\\right),\\sin \\left(\\frac{-\\pi k}{2N}\\right)\\right] / /\\) Calculate Real \n8: \\(V_{\\mathrm{real}}[0]\\gets \\frac{V_{\\mathrm{real}}[0]}{\\sqrt{N\\cdot 2}},V_{\\mathrm{real}}[1:\\gets \\frac{V_{\\mathrm{real}}[1:]}{\\sqrt{\\frac{N}{2}\\cdot 2}} /\\) Energy preservation \n9: Transpose result and repeat to obtain \\(S_{D_{HH}}^{\\mathrm{sp}}\\) \n10: Return \\(S_{D_{HH}}^{\\mathrm{sp}}\\) \n11: Procedure: Pass \\(S_{D_{HH}}^{\\mathrm{sp}}\\) through sequential layers as: \n\\(S_{D_{HH}}^{(n + 1)} =\\) LeakyReLU \\(\\left(\\operatorname {Conv}_{2D}\\left(S_{HH}^{sp(n)},K\\right)\\right)\\) \\(n = 1,\\dots,k,\\) \n12: Return \\(S_{D_{HH}}^{(n + 1)}\\) \n13: Procedure: Extraction \\((S_{D_{HH}}^{(n + 1)},l)\\) . \n14: Set \\((c_x,c_y) = \\left(\\frac{H}{2},\\frac{W}{2}\\right)\\) \n15: Generate mask for high-frequency region within radius \\(r\\) \nfor each coordinate \\((i,j)\\) do: \n\\(D(x_i,y_i) = \\sqrt{(x_i - c_x)^2 + (y_i - c_y)^2}\\) // Euclidian Distance \n17: if \\(D(x_i,y_i)\\leq r\\) then \n18: Set mask[i, \\(j] = 1\\) \n19: end if \n20: end for \n21: Extract mask: \\(S_{D_{HH}}[:,W_c,\\) mask[i, \\(j]\\) \n22: Decode message using learnable \\(\\theta\\) .." + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.534, + 0.393, + 0.567 + ], + "angle": 0, + "content": "\\[\nD _ {\\mathbf {M}} [ i ] = \\left\\{ \\begin{array}{l l} 1 & \\text {i f E x t r a c t e d} [ i ] > \\theta \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right.\n\\]" + }, + { + "type": "algorithm", + "bbox": [ + 0.092, + 0.575, + 0.483, + 0.602 + ], + "angle": 0, + "content": "23: Update \\(\\theta\\) dynamically: \\(\\theta \\gets \\theta -\\eta \\cdot \\frac{\\partial L_{\\mathrm{dec}}}{\\partial\\theta}\\) // Optimizes robustness \n24: Return \\(D_{\\mathrm{M}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.64, + 0.484, + 0.701 + ], + "angle": 0, + "content": "of the embedded watermark. From Parseval's theorem [20] ensures overall spectral and spatial energies remain equivalent, though local spectral energy distributions are altered by the watermark strength factor \\( s \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The watermark's strength factor \\( s \\) ensures that the high-energy areas where the message \\( M \\) is embedded as \"1\" remain robust, experiencing a minimum distortion in such conditions. Moreover, this threshold can be optimized for better bit recovery accuracy during training. As \\( \\theta \\) learns, it recognizes that areas encoded as \"1\" carry higher energy and impact due to the strength factor \\( s \\) of Eq. (15), while areas marked as \"0\", softened by the LeakyReLU's minimal negative slope, have a lower intensity. Such a dynamic approach enables \\( \\theta \\) to identify and protect the embedded message \\( M \\) even when external disturbances occur, preserving the watermark's structure within the watermarked image \\( I_{\\text{embedded}} \\). And, \\( \\theta \\) effectively learns to distin" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "guish high-energy watermark regions. Therefore, the embedded message is more recoverable under diverse attacks, and SpecGuard's decoder ensures valid watermark bit extraction. Theoretical explanation of Parseval theorem's [20] impact on message extraction is in the Supplementary." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.174, + 0.798, + 0.191 + ], + "angle": 0, + "content": "3.3. Loss Calculation for SpecGuard" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.196, + 0.906, + 0.257 + ], + "angle": 0, + "content": "To achieve the training objective of robust and invisible watermark embedding, a composite loss function is defined with two terms: encoder loss \\( L_{\\mathrm{enc}} \\) as expressed in Eq. (18) and decoder loss \\( L_{\\mathrm{dec}} \\) as expressed in Eq. (19)." + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.26, + 0.906, + 0.282 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathbb {E} _ {(I, M) \\sim D} L _ {\\text {e n c}} (I, I _ {\\text {e m b e d d e d}}) = \\| E _ {\\theta} (I, M) - I \\| ^ {2}, \\tag {18}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.302, + 0.906, + 0.324 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathbb {E} _ {(I, M) \\sim D} L _ {\\mathrm {d e c}} (M, D _ {\\mathrm {M}}) = \\| D _ {\\theta} \\left(I _ {\\text {e m b e d d e d}}\\right) - M \\| ^ {2}, \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.328, + 0.906, + 0.448 + ], + "angle": 0, + "content": "where \\( E_{\\theta}(I,M) \\) denotes the encoder output, embedding the message \\( M \\) into the cover image \\( I \\) to produce \\( I_{\\mathrm{embedded}} \\). By minimizing \\( L_{\\mathrm{enc}} \\), the encoder learns to embed the watermark invisibly, preserving the fidelity of the cover image. \\( D_{\\theta}(I_{\\mathrm{embedded}}) \\) denotes the decoder's output from the watermarked image \\( I_{\\mathrm{embedded}} \\). Minimizing \\( L_{\\mathrm{dec}} \\) enables the decoder to reliably retrieve the embedded message under varying conditions, such as noise and transformation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.449, + 0.906, + 0.494 + ], + "angle": 0, + "content": "The total loss \\( L \\) as shown in Eq. (20) used for optimizing the model combines these terms, balancing invisibility and robustness through weighted coefficients as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.499, + 0.905, + 0.518 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} L = \\lambda_ {\\text {e n c}} L _ {\\text {e n c}} + \\lambda_ {\\text {d e c}} L _ {\\text {d e c}}, \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.906, + 0.555 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\mathrm{enc}}\\) and \\(\\lambda_{\\mathrm{dec}}\\) control the relative importance of visual fidelity and message recoverability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.565, + 0.719, + 0.582 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.59, + 0.609, + 0.604 + ], + "angle": 0, + "content": "4.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.907, + 0.869 + ], + "angle": 0, + "content": "SpecGuard is trained on the MS-COCO dataset [26], which contains 25K images. To evaluate the robustness of the watermarking methods including our SpecGuard against different types of attacks, such as distortions, regenerations, and adversarial attacks, we used three datasets: DiffusionDB [46], MS-COCO [26], and DALL-E31. Each of these datasets has a unique distribution of prompt words. We also ensured that no unethical or violent terms were included in the prompts. We randomly picked 200 images from MS-COCO [26] and applied watermark using SpecGuard for further verifying the robustness after uploading on various social media platforms and applying AI-based Photoshop Neural Filters (PNFs)2. The PNFs include depth blur, artistic style transfer, super zoom, JPEG artifact reduction, and colorization. For the super zoom filter, we set the 'Sharpen' and 'Noise Reduction' parameters to 15. For all other filters, we used the default settings." + }, + { + "type": "page_footnote", + "bbox": [ + 0.53, + 0.875, + 0.863, + 0.887 + ], + "angle": 0, + "content": "\\(^{1}\\)https://huggingface.co/datasets/OpenDatasets/dalle-3-dataset" + }, + { + "type": "page_footnote", + "bbox": [ + 0.533, + 0.888, + 0.865, + 0.9 + ], + "angle": 0, + "content": "\\(^{2}\\)https://www.adobe.com/products/photoshop/neural-filter.html" + }, + { + "type": "list", + "bbox": [ + 0.53, + 0.875, + 0.865, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17988" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.088, + 0.462, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.256, + 0.483, + 0.298 + ], + "angle": 0, + "content": "Figure 3. Some best results for cover vs watermarked images with PSNR/SSIM \\((\\uparrow)\\) scores showing minimal visual degradation when watermarked using proposed SpecGuard." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.318, + 0.496, + 0.384 + ], + "angle": 0, + "content": "
Metrics256 × 256512 × 5121024 × 1024
CelebA-HQMS-COCOCelebA-HQMS-COCOCelebA-HQMS-COCO
PSNR↑40.36140.32044.65144.68048.17048.081
SSIM↑0.98890.98880.99270.99270.99370.9936
FID↓16.45116.69016.97217.02017.44616.955
MSE↓0.00020.00020.00010.00010.00010.0001
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.393, + 0.484, + 0.422 + ], + "angle": 0, + "content": "Table 1. Perceptual quality of SpecGuard watermarked images across resolutions and datasets using a fixed 30-bit message length." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.453, + 0.251, + 0.469 + ], + "angle": 0, + "content": "4.2. Implementation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We used CUDA v11.3 and PyTorch with a batch size of 32 and the Adam optimizer on a multiple NVIDIA RTX 2080-equipped server. Mean Squared Error (MSE) and Bit Recovery Accuracy (BRA) are used for loss and accuracy calculation. We used Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), Fréchet Inception Distance (FID), and MSE to evaluate perceptual quality. Our model is trained for 300 epochs, with the decoder learning rate set to \\(1 \\times 10^{-3}\\), reduced by half every 100 steps, and the encoder learning rate is set to \\(1 \\times 10^{-2}\\) without scheduling. We set our watermark radius \\((r)\\), strength factor \\((s)\\), initial learning parameter \\((\\theta)\\), and the number of convolutional layers \\((k)\\) to 100, 20, 0.001, and 32, respectively. This setup is applied with a message bit length (BL) of 48, 64, 128, and 256. Initially, decoder loss weight \\((\\lambda_{\\mathrm{dec}})\\) and encoder loss weight \\((\\lambda_{\\mathrm{enc}})\\) are set to 1.0 and 0.7, respectively. For assessing the robustness of watermarking methods against diverse attacks, we inherited the experimental setups from Waves [8] and used effective metrics such as \"Quality at 95% Performance (Q@0.95P)\", \"Quality at 70% Performance (Q@0.7P)\", \"Avg P\" and \"Avg Q.\" Here, Q@0.95P and Q@0.7P indicate the level of image quality degradation required for watermark detection accuracy to reach 95% and 70%, respectively. The average performance (Avg P) metric represents the mean detection accuracy across various attack strengths, while the average quality degradation (Avg Q) measures the overall impact of attacks on image quality [8, 35]." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.089, + 0.903, + 0.449 + ], + "angle": 0, + "content": "
MethodsVenueBLPSNR↑SSIM↑FID↓BRA↑
Pre-processing methodsTree-Ring [47]NeurIPS'236432.330.9117.70.98
12832.100.9017.80.96
25631.850.8917.90.94
Stable Signature [35]ICCV'236430.000.8919.60.98
12829.800.8819.70.96
25629.500.8719.80.96
Yang et al. [49]CVPR'246431.450.9018.20.98
12831.200.8918.30.93
25630.950.8818.40.89
SleeperMark [45]CVPR'256431.800.9218.00.97
12831.600.9118.10.93
25631.350.9018.20.87
Post-processing methodsHiDDeN [53]ECCV'186432.010.8819.70.98
12831.800.8719.80.85
25631.500.8619.90.82
StegaStamp [41]CVPR'206428.500.9117.90.99
12828.200.9018.00.98
25628.000.8918.10.94
MBRS [17]ACM MM'216438.200.9617.90.98
12837.900.9518.00.96
25637.500.9418.20.94
FIN [10]AAAI'236436.700.9518.30.97
12836.400.9418.40.96
25636.100.9318.50.96
MuST [43]AAAI'246441.200.9717.50.98
12840.900.9617.60.93
25640.500.9517.80.90
EditGuard [52]CVPR'246441.560.9717.80.98
12841.300.9617.90.97
25640.900.9518.00.97
SpecGuard (Ours)ICCV'256442.590.9817.20.99
12842.890.9917.00.99
25640.860.9917.60.98
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.459, + 0.907, + 0.488 + ], + "angle": 0, + "content": "Table 2. Comparison of SOTA pre-processing and post-processing watermarking methods with SpecGuard without attacks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.511, + 0.727, + 0.527 + ], + "angle": 0, + "content": "4.3. Watermark Invisibility" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.536, + 0.907, + 0.702 + ], + "angle": 0, + "content": "To evaluate the invisibility of the embedded watermark, we conducted perceptual and quantitative assessments using SpecGuard. As shown in Fig. 3, there is no noticeable perceptual degradation between the cover and watermarked images, confirming that the watermark remains imperceptible to the human eye. For a more comprehensive evaluation, we created three subsets of different image sizes ranging between 256 to 1024 with images from the MS-COCO [26] and CelebA-HQ [19] datasets and applied the SpecGuard watermarking method to compare the average PSNR values between the cover and watermarked images, as in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.903 + ], + "angle": 0, + "content": "For quantitative evaluation, we further compare the performance of SpecGuard with the SOTA pre-processing and post-processing watermarking methods. As presented in Tab. 2, SpecGuard achieves the highest PSNR of 42.89 when the bit length was 128. Additionally, it attains the highest SSIM of 0.99 at a BL of 128 and 256 among all compared methods, indicating minimal visual distortion. Additionally, SpecGuard achieved the lowest FID of 17.0 and the highest BRA of 0.99, ensuring strong robustness while maintaining imperceptibility. Overall, our results demonstrate that SpecGuard outperforms both pre-processing and post-processing watermarking methods, achieving superior imperceptibility and robustness." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17989" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.091, + 0.088, + 0.907, + 0.382 + ], + "angle": 0, + "content": "
Attack TypeTree-Ring [47]Stable Signature [35]StegaStamp [41]SpecGuard (Ours)
Q@0.95PQ@0.7PAvg PAvg QQ@0.95PQ@0.7PAvg PAvg QQ@0.95PQ@0.7PAvg PAvg QQ@0.95PQ@0.7PAvg PAvg Q
DistortionsRotation0.4640.5210.3750.6480.6240.7020.5940.6500.4230.4980.3570.6160.8630.8630.6870.653
Crop0.5920.5920.3320.463infinf0.9950.4610.6020.6020.5400.4510.8120.8120.9980.742
Brightinfinfinf0.304infinf0.9980.305infinf0.9980.317infinf0.9980.466
Contrastinfinf0.9980.243infinf0.9980.243infinf0.9980.231infinf0.9980.556
Blur0.8611.1120.5631.221- inf- inf0.0001.2040.8480.9620.4141.0000.921inf1.0001.452
Noise0.548inf0.9800.3950.4020.5200.8700.390infinf1.0000.360infinf0.9990.568
JPEG0.4990.4990.9290.2840.4850.4850.7930.284infinf0.9980.263infinf1.0000.495
Geo0.5250.5930.2770.7680.850inf0.9370.7670.6630.6930.3960.7330.8690.8690.8650.623
Deg0.620inf0.8920.6940.2060.3690.3000.6790.8260.9750.8520.6640.8951.1410.9150.749
Combine0.5390.7510.4030.9080.5380.6910.3340.9000.9451.1010.7950.8700.9791.2560.9110.952
RegenerationRegen-Diff- inf0.3070.6120.323- inf- inf0.0010.3000.331inf0.9430.327infinf0.9820.477
Regen-DiffPinf0.3070.6010.327- inf- inf0.0010.3030.333inf0.9400.329infinf0.9820.562
Regen-VAE0.5780.5780.8320.3480.5450.5450.5160.339infinf1.0000.343infinf0.9950.521
Regen-KLVAEinfinf0.9900.233 6- inf0.1760.2170.206infinf1.0000.240infinf0.9900.492
Rinse-2xDiff- inf0.3330.5100.357- inf- inf0.0010.3320.391inf0.9410.366infinf0.9930.561
Rinse-4xDiff- inf0.3550.4430.466- inf- inf0.0000.4380.388inf0.9090.477infinf0.9920.533
AdversarialAdvEmbG-KLVAE8- inf0.1640.4480.253infinf0.9980.249infinf1.0000.232infinf1.0000.456
AdvEmbB-RN180.241inf0.9530.218infinf0.9990.212infinf1.0000.196infinf1.0000.467
AdvEmbB-CLIP0.541inf0.9320.549infinf0.9990.541infinf1.0000.488infinf1.0000.436
AdvEmbB-KLVAE160.195inf0.8880.238infinf0.9970.233infinf1.0000.206infinf1.0000.482
AdvEmbB-SdxlVAE0.222inf0.9340.221infinf0.9980.219infinf1.0000.204infinf1.0000.492
AdvCls-UnWM&WM- inf0.1020.4990.145infinf0.9990.101infinf1.0000.101infinf1.0000.497
AdvCls-Real&WMinfinf1.0000.047infinf0.9980.092infinf1.0000.106infinf1.0000.427
AdvCls-WM1&WM2- inf0.1010.4920.139infinf0.9990.084infinf1.0000.129infinf1.0000.441
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.393, + 0.907, + 0.435 + ], + "angle": 0, + "content": "Table 3. Robustness comparison various across attacks using Q@0.95P(↑), Q@0.7P(↑), Avg P(↑) and Avg Q(↑). Here, 'inf' denotes that no attack was sufficient to degrade performance below the threshold, indicating strong robustness, whereas '-inf' signifies that even the weakest attack caused detection to fall below the threshold, reflecting weak robustness." + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.447, + 0.268, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.268, + 0.447, + 0.421, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.421, + 0.447, + 0.576, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.447, + 0.731, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.447, + 0.886, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.58, + 0.266, + 0.711 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.268, + 0.58, + 0.421, + 0.711 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.421, + 0.58, + 0.576, + 0.711 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.58, + 0.731, + 0.711 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.58, + 0.886, + 0.711 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.721, + 0.907, + 0.75 + ], + "angle": 0, + "content": "Figure 4. Robustness validation of our proposed SpecGuard under different distortion attacks, including geometric transformations: Geo (rotation, cropping), photometric modifications (brightness, contrast), and degradations: Deg (blur, noise, JPEG compression)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.764, + 0.197, + 0.781 + ], + "angle": 0, + "content": "4.4. Capacity" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.485, + 0.902 + ], + "angle": 0, + "content": "To evaluate embedding capacity, we examined SpecGuard across different bit lengths and compared it with SOTA watermarking methods. Our experiments with 64, 128, and 256 bits demonstrate SpecGuard's high capacity of bit embedding while maintaining perceptual quality and robustness, as shown in Tab. 2. Notably, it achieves a PSNR of 42.89, the highest among all methods, along with the high" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.887 + ], + "angle": 0, + "content": "est BRA of 0.99 and the lowest FID of 17.0 at 128 bits, ensuring minimal visual impact. This adaptability to different bit lengths without quality loss makes SpecGuard ideal for applications requiring flexible watermark sizes. Unlike StegaStamp and HiDDeN, which suffer reduced BRA for higher message bits, SpecGuard consistently extracts bits across all tested lengths. SpecGuard's theoretical watermark capacity is provided in the Supplementary." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17990" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.089, + 0.472, + 0.166 + ], + "angle": 0, + "content": "
ModulesPSNR/SSIM↑BRA↑ModulesPSNR/SSIM↑BRA↑
WP(L1)40.51/0.960.92WP(L1)+SPFA42.89/0.990.99
WP(L2)38.15/0.930.87WP(L2)+SPFA36.25/0.920.89
AttacksPSNR/SSIM↑BRA↑AttacksPSNR/SSIM↑BRA↑
Rotate (45°)12.15/21.310.82Rotate (90°)11.15/19.310.65
Blur (0.3)35.01/0.950.98Blur (0.6)30.11/0.910.98
Geo (0.3)12.08/0.500.93Geo (0.6)10.25/0.450.86
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.11, + 0.166, + 0.468, + 0.176 + ], + "angle": 0, + "content": "*WP: Wavelet Projection, SP: Spectral Projection, FA: FFT Approximation" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.188, + 0.483, + 0.216 + ], + "angle": 0, + "content": "Table 4. Ablation studies on the proposed SpecGuard for across various configurations, setting \\( M = 128 \\), \\( r = 100 \\), and \\( s = 20 \\)." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.217, + 0.485, + 0.284 + ], + "angle": 0, + "content": "
PlatformPSNR/SSIM↑BRA↑PS FiltersPSNR/SSIM↑BRA↑
Facebook48.56/0.970.97Depth Blur25.25/0.890.85
LinkedIn47.55/0.970.96StyleT.25.12/0.840.85
Instagram48.56/0.980.98Super Zoom36.15/0.880.95
WhatsApp42.10/0.960.97JPEG Artifacts31.01/0.850.94
X (Twitter)49.25/1.000.99Colorize23.15/0.820.92
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.293, + 0.483, + 0.322 + ], + "angle": 0, + "content": "Table 5. Evaluation of SpecGuard's robustness across Photoshop filters and while uploaded on different social media platforms." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.334, + 0.216, + 0.349 + ], + "angle": 0, + "content": "4.5. Robustness" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.357, + 0.483, + 0.809 + ], + "angle": 0, + "content": "We evaluate watermarking robustness by analyzing detection performance against a range of diverse and challenging real-world attacks. Results demonstrate the strong robustness of SpecGuard across various attacks. For example, as presented in Tab. 3, against geometric distortions such as cropping and rotation, SpecGuard achieved an Avg P of 0.998 and 0.687, respectively. Similarly, across the combined distortion-based attacks, SpecGuard achieves an overall Avg P of 0.911 and Avg Q of 0.952, ensuring minimal quality loss while maintaining high detection accuracy. Notably, the high values of Q@0.95P and Q@0.7P indicate that SpecGuard can sustain reliable detection at strict performance thresholds, even under aggressive perturbations. Unlike prior methods that struggle with extreme transformations, SpecGuard shows remarkable robustness against regeneration-based attacks like Rinse-2xDiff [6] (an image is noised then denoised by Stable Diffusion v1.4 two times with strength as a number of timesteps, 20-100) and RegenVAE [6], maintaining high Avg P. Similarly, under adversarial attacks, SpecGuard consistently secures watermark detectability, outperforming existing techniques across all tested scenarios. These results establish SpecGuard as a highly robust watermarking approach capable of preserving image integrity even under severe distortions and adversarial manipulations, ensuring watermark reliability across diverse attack types. More details about how the attacks are performed are provided in supplementary material. Further, our results in Fig. 4 highlight the strong robustness of SpecGuard against various distortion attacks compared to other SOTA watermarking methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Social Platforms and Photoshop Filters. SpecGuard's robustness when images are shared across social media platforms and subjected to common Photoshop Neural Filters (PNFs) is shown in Tab. 5. SpecGuard consistently maintains high PSNR and SSIM values, with BRA values close to 0.99 on platforms such as X (formally Twitter), Insta" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": "gram, and Facebook. Also, it shown strong resilience to various PNFs, such as Super Zoom and JPEG Artifacts achieving BRA of 0.95 and 0.94. The PSNR, SSIM, and BRA values are expected to decrease with the severity of image manipulation, as increased manipulation leads to loss of image authenticity. For example, as we applied \\(60\\%\\) style transfer the PSNR and BRA decreased to 25.12 and 0.85. Similarly, the depth blur which excessively reduces the image clarity also causes the decrease of BRA to 0.85." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.239, + 0.667, + 0.255 + ], + "angle": 0, + "content": "4.6. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.26, + 0.907, + 0.563 + ], + "angle": 0, + "content": "We examined the impact of WP at different levels (\\(L_{1}\\) and \\(L_{2}\\)) and its combination with SP using \\(FA\\) in Tab. 4. As observed, the \\(\\mathrm{WP}(L_1) + \\mathrm{SP}_{FA}\\) configuration achieved the highest PSNR and SSIM values of 42.89 and 0.99, respectively, and BRA of 0.99, indicating improved watermark invisibility and robustness. In contrast, using WP alone at either \\(L_{1}\\) or \\(L_{2}\\) resulted in lower BRA, with values of 0.92 and 0.87, respectively, demonstrating that the combined \\(\\mathrm{WP} + \\mathrm{SP}_{FA}\\) approach significantly enhances performance. We also evaluated the robustness of SpecGuard under strong adversarial attacks identified in Tab. 4, such as rotation, blur, and geometric transformations. The results indicate that higher levels of attack severity, such as \\(90^{\\circ}\\) rotation, lead to a more significant drop in PSNR, SSIM, and BRA, with values dropping to 11.15, 19.31, and 0.65, respectively. Despite this, the model shows relatively high resilience under moderate attack intensities, such as \\(45^{\\circ}\\) rotation and low levels of blur and geometric distortion, achieving BRA values as high as 0.93 under geometric transformations at the 0.3 thresholds. More ablations are in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.578, + 0.634, + 0.594 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.604, + 0.907, + 0.755 + ], + "angle": 0, + "content": "We propose SpecGuard, a novel invisible watermarking method that ensures secure and robust information concealment. Unlike traditional approaches, SpecGuard remains highly resilient against diverse distortions, adversarial attacks, and regeneration-based transformations. Experimental results demonstrate its superior bit recovery accuracy of \\(99\\%\\) maintaining high PSNR. By outperforming SOTA watermarking methods in both detection reliability and imperceptibility, SpecGuard establishes a new benchmark for watermarking under real-world constraints." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.77, + 0.673, + 0.787 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.905, + 0.901 + ], + "angle": 0, + "content": "This work was partly supported by Institute for Information & communication Technology Planning & evaluation (IITP) grants funded by the Korean government MSIT: (RS-2022-II221199, RS-2022-II220688, RS-2019-II190421, RS-2023-00230337, RS-2024-00356293, RS-2024-00437849, RS-2021-II212068, RS-2025-02304983, and RS-2025-02263841)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17991" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.169 + ], + "angle": 0, + "content": "[1] Edward H Adelson, Eero Simoncelli, and Rajesh Hingorani. Orthogonal pyramid transforms for image coding. In Visual Communications and image processing II, pages 50-58. SPIE, 1987. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.226 + ], + "angle": 0, + "content": "[2] Mahdi Ahmadi, Alireza Norouzi, Nader Karimi, Shadrokh Samavi, and Ali Emami. Redmark: Framework for residual diffusion watermarking based on deep networks. Expert Systems with Applications, 146:113157, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.483, + 0.282 + ], + "angle": 0, + "content": "[3] Inzamamul Alam, Md Tanvir Islam, and Simon S Woo. Specxnet: A dual-domain convolutional network for robust deepfake detection. In Proceedings of the 33nd ACM International Conference on Multimedia, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.284, + 0.483, + 0.338 + ], + "angle": 0, + "content": "[4] Inzamamul Alam, Md Tanvir Islam, and Simon S Woo. Saliency-aware diffusion reconstruction for effective invisible watermark removal. In Companion Proceedings of the ACM on Web Conference 2025, pages 849-853, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.341, + 0.484, + 0.395 + ], + "angle": 0, + "content": "[5] Aashutosh AV, Srijan Das, Abhijit Das, et al. Latent flow diffusion for deepfake video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3781-3790, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.397, + 0.482, + 0.45 + ], + "angle": 0, + "content": "[6] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.453, + 0.482, + 0.507 + ], + "angle": 0, + "content": "[7] Tali Dekel, Michael Rubinstein, Ce Liu, and William T Freeman. On the effectiveness of visible watermarks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2146-2154, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.509, + 0.482, + 0.59 + ], + "angle": 0, + "content": "[8] Mucong Ding, Tahseen Rabbani, Bang An, Aakriti Agrawal, Yuancheng Xu, Chenghao Deng, Sicheng Zhu, Abdirisak Mohamed, Yuxin Wen, Tom Goldstein, et al. Waves: Benchmarking the robustness of image watermarks. In ICLR 2024 Workshop on Reliable and Responsible Foundation Models, 2024. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.593, + 0.482, + 0.62 + ], + "angle": 0, + "content": "[9] Hubert Etienne. The future of online trust (and why deepfake is advancing it). AI and Ethics, 1(4):553-562, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.622, + 0.482, + 0.689 + ], + "angle": 0, + "content": "[10] Han Fang, Yupeng Qiu, Kejiang Chen, Jiyi Zhang, Weiming Zhang, and Ee-Chien Chang. Flow-based robust watermarking with invertible noise layer for black-box distortions. In Proceedings of the AAAI conference on artificial intelligence, pages 5054-5061, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[11] Jianwei Fei, Zhihua Xia, Benedetta Tondi, and Mauro Barni. Supervised gan watermarking for intellectual property protection. In 2022 IEEE International Workshop on Information Forensics and Security (WIFS), pages 1-6. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[12] S. A. Fulling. The local geometric asymptotics of continuum eigenfunction expansions. ii. one-dimensional systems. SIAM Journal on Mathematical Analysis, 14(4):605-623, 1983. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[13] Mahdieh Ghazvini, Elham Mohamadi Hachrood, and Mojdeh Mirzadi. An improved image watermarking method in frequency domain. Journal of Applied Security Research, 12 (2):260-275, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[15] Md Tanvir Islam, Ik Hyun Lee, Ahmed Ibrahim Alzahrani, and Khan Muhammad. Mexic: A meta ensemble explainable approach for ai-synthesized fake image classification. *Alexandria Engineering Journal*, 116:351–363, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[16] K Jayashre and M Amsaprabhaa. Safeguarding media integrity: A hybrid optimized deep feature fusion based deepfake detection in videos. Computers & Security, 142:103860, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[17] Zhaoyang Jia, Han Fang, and Weiming Zhang. Mbrs: Enhancing robustness of dnn-based watermarking by minibatch of real and simulatedJPEG compression. In Proceedings of the 29th ACM international conference on multimedia, pages 41-49, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.307, + 0.904, + 0.347 + ], + "angle": 0, + "content": "[18] Hoon Kang and Joonsoo Ha. Projection spectral analysis. International Journal of Control, Automation and Systems, 13(6):1530-1537, 2015. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.35, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[19] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations (ICLR), 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.407, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[20] SS Kelkar, LL Grigsby, and J Langsner. An extension of parseval's theorem and its use in calculating transient energy in the frequency domain. IEEE Transactions on Industrial Electronics, (1):42-45, 1983. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.463, + 0.904, + 0.517 + ], + "angle": 0, + "content": "[21] Varsha Kishore, Xiangyu Chen, Yan Wang, Boyi Li, and Kilian Q Weinberger. Fixed neural network steganography: Train the images, not the network. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.52, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[22] Jae-Eun Lee, Young-Ho Seo, and Dong-Wook Kim. Convolutional neural network-based digital image watermarking adaptive to the resolution of image and watermark. Applied Sciences, 10(19):6854, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.577, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[23] Yicheng Leng, Chaowei Fang, Gen Li, Yixiang Fang, and Guanbin Li. Removing interference and recovering content imaginatively for visible watermark removal. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2983-2990, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.648, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[24] Xiao Li, Liquan Chen, Ju Jia, Zhongyuan Qin, and Zhangjie Fu. A lightweight image forgery prevention scheme for IoT using gan-based steganography. IEEE Transactions on Industrial Informatics, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.705, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[25] Dongdong Lin, Benedetta Tondi, Bin Li, and Mauro Barni. A cyclegan watermarking method for ownership verification. IEEE Transactions on Dependable and Secure Computing, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.761, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Yang Liu, Zhen Zhu, and Xiang Bai. Wdnet: Watermark-decomposition network for visible watermark removal. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 3685-3693, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17992" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[28] Xiyang Luo, Ruohan Zhan, Huiwen Chang, Feng Yang, and Peyman Milanfar. Distortion agnostic deep watermarking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13548-13557, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.231 + ], + "angle": 0, + "content": "[29] Thanh Thi Nguyen, Quoc Viet Hung Nguyen, Dung Tien Nguyen, Duc Thanh Nguyen, Thien Huynh-The, Saeid Nahavandi, Thanh Tam Nguyen, Quoc-Viet Pham, and Cuong M Nguyen. Deep learning for deepfakes creation and detection: A survey. Computer Vision and Image Understanding, 223:103525, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.287 + ], + "angle": 0, + "content": "[30] Guangyu Nie, Changhoon Kim, Yezhou Yang, and Yi Ren. Attributing image generative models using latent fingerprints. In International Conference on Machine Learning, pages 26150-26165. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.289, + 0.482, + 0.343 + ], + "angle": 0, + "content": "[31] Li Niu, Xing Zhao, Bo Zhang, and Liqing Zhang. Fine-grained visible watermark removal. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12770-12779, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.345, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[32] Konstantin A Pantserev. The malicious use of ai-based deepfake technology as the new threat to psychological security and political stability. Cyber defence in the age of AI, smart societies and augmented humanity, pages 37-55, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.401, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[33] Ram Shankar Pathak. The wavelet transform. Springer Science & Business Media, 2009. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[34] Fernandez Pierre, Alexandre Sablayrolles, Teddy Furon, Hervé Jégou, and Matthijs Douze. Watermarking images in self-supervised latent spaces. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3054-3058. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.482, + 0.566 + ], + "angle": 0, + "content": "[35] Fernandez Pierre, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22466-22477, 2023. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.568, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[36] Tong Qiao, Yuyan Ma, Ning Zheng, Hanzhou Wu, Yanli Chen, Ming Xu, and Xiangyang Luo. A novel model watermarking for protecting generative adversarial network. Computers & Security, 127:103102, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.677 + ], + "angle": 0, + "content": "[37] Sunpreet Sharma, Ju Jia Zou, Gu Fang, Pancham Shukla, and Weidong Cai. A review of image watermarking for identity protection and verification. Multimedia Tools and Applications, 83(11):31829-31891, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.733 + ], + "angle": 0, + "content": "[38] Qingtang Su, Huanying Wang, DeCheng Liu, Zihan Yuan, and Xueting Zhang. A combined domain watermarking algorithm of color image. Multimedia Tools and Applications, 79(39):30023-30043, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[39] Qingtang Su, Xueting Zhang, and Huanying Wang. A blind color image watermarking algorithm combined spatial domain and svd. International Journal of Intelligent Systems, 37(8):4747-4771, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.791, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[40] Ruizhou Sun, Yukun Su, and Qingyao Wu. Denet: disentangled embedding network for visible watermark removal. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2411-2419, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.847, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[41] Matthew Tancik, Ben Mildenhall, and Ren Ng. Stegastamp: Invisible hyperlinks in physical photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2117-2126, 2020. 1, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[42] D Vaishnavi and TS Subashini. Robust and invisible image watermarking in rgb color space using svd. Procedia Computer Science, 46:1770-1777, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.907, + 0.204 + ], + "angle": 0, + "content": "[43] Guanjie Wang, Zehua Ma, Chang Liu, Xi Yang, Han Fang, Weiming Zhang, and Nenghai Yu. Must: Robust image watermarking for multi-source tracing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5364-5371, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[44] Huanying Wang and Qingtang Su. A color image watermarking method combined qr decomposition and spatial domain. Multimedia Tools and Applications, 81(26):37895-37916, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.263, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[45] Zilan Wang, Junfeng Guo, Jiacheng Zhu, Yiming Li, Heng Huang, Muhao Chen, and Zhengzhong Tu. Sleepermark: Towards robust watermark against fine-tuning text-to-image diffusion models. arXiv preprint arXiv:2412.04852, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.905, + 0.386 + ], + "angle": 0, + "content": "[46] Zijie J Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. Diffusional: A large-scale prompt gallery dataset for text-to-image generative models. arXiv preprint arXiv:2210.14896, 2022.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[47] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-rings watermarks: Invisible fingerprints for diffusion images. In Advances in Neural Information Processing Systems, pages 58047-58063. Curran Associates, Inc., 2023. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.515 + ], + "angle": 0, + "content": "[48] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-rings watermarks: Invisible fingerprints for diffusion images. Advances in Neural Information Processing Systems, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.517, + 0.905, + 0.599 + ], + "angle": 0, + "content": "[49] Zijin Yang, Kai Zeng, Kejiang Chen, Han Fang, Weiming Zhang, and Nenghai Yu. Gaussian shading: Provable performance-lossless image watermarking for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12162-12171, 2024. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.602, + 0.905, + 0.655 + ], + "angle": 0, + "content": "[50] Zihan Yuan, Qingtang Su, Decheng Liu, and Xueting Zhang. A blind image watermarking scheme combining spatial domain and frequency domain. The visual computer, 37:1867-1881, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.657, + 0.905, + 0.713 + ], + "angle": 0, + "content": "[51] Chaoning Zhang, Philipp Benz, Adil Karjauv, Geng Sun, and In So Kweon. Udh: Universal deep hiding for steganography, watermarking, and light field messaging. Advances in Neural Information Processing Systems, 33:10223-10234, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.715, + 0.905, + 0.783 + ], + "angle": 0, + "content": "[52] Xuanyu Zhang, Runyi Li, Jiwen Yu, Youmin Xu, Weiqi Li, and Jian Zhang. Editguard: Versatile image watermarking for tamper localization and copyright protection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11964-11974, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.786, + 0.905, + 0.827 + ], + "angle": 0, + "content": "[53] Jiren Zhu, Russell Kaplan, Justin Johnson, and Li Fei-Fei. Hidden: Hiding data with deep networks. European Conference on Computer Vision (ECCV), 2018. 1, 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.827 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17993" + } + ] +] \ No newline at end of file diff --git a/2025/SpecGuard_ Spectral Projection-based Advanced Invisible Watermarking/1fcb8d1e-6633-4c96-8cd9-9ceeae669d64_origin.pdf b/2025/SpecGuard_ Spectral Projection-based Advanced Invisible Watermarking/1fcb8d1e-6633-4c96-8cd9-9ceeae669d64_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..481558d336d082380a42ec91a57ae6d928d06dca --- /dev/null +++ b/2025/SpecGuard_ Spectral Projection-based Advanced Invisible Watermarking/1fcb8d1e-6633-4c96-8cd9-9ceeae669d64_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1bd493aa9bc5f6c6a03d70814ffe9100011bc87745d16ad940c3300bbf3ed231 +size 2595094 diff --git a/2025/SpecGuard_ Spectral Projection-based Advanced Invisible Watermarking/full.md b/2025/SpecGuard_ Spectral Projection-based Advanced Invisible Watermarking/full.md new file mode 100644 index 0000000000000000000000000000000000000000..01750abc78ae45ef8bade37836395a4b4d6f7639 --- /dev/null +++ b/2025/SpecGuard_ Spectral Projection-based Advanced Invisible Watermarking/full.md @@ -0,0 +1,395 @@ +# SpecGuard: Spectral Projection-based Advanced Invisible Watermarking + +Inzamamul Alam, Md Tanvir Islam, Simon S. Woo*, Khan Muhammad Sungkyunkwan University + +{inzil5, tanvirnwu, swoo, khanmuhammad}@g.skku.edu + +# Abstract + +Watermarking embeds imperceptible patterns into images for authenticity verification. However, existing methods often lack robustness against various transformations primarily including distortions, image regeneration, and adversarial perturbation, creating real-world challenges. In this work, we introduce SpecGuard, a novel watermarking approach for robust and invisible image watermarking. Unlike prior approaches, we embed the message inside hidden convolution layers by converting from the spatial domain to the frequency domain using spectral projection of a higher frequency band that is decomposed by wavelet projection. Spectral projection employs Fast Fourier Transform approximation to transform spatial data into the frequency domain efficiently. In the encoding phase, a strength factor enhances resilience against diverse attacks, including adversarial, geometric, and regeneration-based distortions, ensuring the preservation of copyrighted information. Meanwhile, the decoder leverages Parseval's theorem to effectively learn and extract the watermark pattern, enabling accurate retrieval under challenging transformations. We evaluate the proposed SpecGuard based on the embedded watermark's invisibility, capacity, and robustness. Comprehensive experiments demonstrate the proposed SpecGuard outperforms the state-of-the-art models. To ensure reproducibility, the full code is released on GitHub. + +# 1. Introduction + +With the rapid advancement of digital media and artificial intelligence, concerns regarding image authenticity, copyright protection, and content integrity have become more challenging than ever [9, 16, 32]. Moreover, the widespread availability of the latest image manipulation tools [4, 5, 14] enables malicious tamperers to easily forge and redistribute digital content without authorization, posing a significant threat to ownership verification [29]. This growing risk emphasizes the need for reliable techniques for secure authentication and detection of unauthorized manipulation. + +![](images/a38e60f474c97ee8790f581a20a8d3b9e9e3db28ed3cebb62e1971ff215af1a0.jpg) +Figure 1. Image authentication using our proposed SpecGuard. + +Recently, invisible watermarking has gained significant attention as a prominent defense mechanism for media authentication by embedding invisible messages into images to verify authenticity [35, 48]. In fact, invisible watermarks are preferred for preserving image quality and resisting tampering. These watermarks are unique to the creator and enable tamper verification by comparing the retrieved watermark to the original, as the high-level process is presented in Fig. 1. Traditional watermarking methods often rely on transformation techniques [18, 33]. Deep learning approaches like StegaStamp [41], Stable Signature [35], and HiDDeN [53] provide end-to-end solutions for message embedding. However, these methods often struggle with fragility in handling common image processing operations such as resizing, cropping, compression, and noise addition, which can distort or erase the embedded watermark. Additionally, the performance of watermark embedding and extraction often remains vulnerable to attacks with noise injection, blurring, contrasting, and rotation [8]. + +To address the aforementioned challenges, we introduce a novel robust, and invisible image watermarking method named SpecGuard. SpecGuard is designed to overcome the fundamental trade-offs [8] between imperceptibility, and robustness. Our proposed SpecGuard strategically embeds watermark information in the spectral domain, leveraging wavelet-based decomposition to distribute the watermark across high-frequency components. Unlike traditional frequency domain watermarking techniques [47, 49] that are easily disrupted by common image manipulations, Spec- + +Guard maintains imperceptibility while significantly improving robustness against a wide range of transformations. + +Overall, our proposed SpecGuard addresses the current limitations of the previous watermarking methods by providing a robust, imperceptible watermarking technique that maintains integrity under diverse manipulations, significantly enhancing digital content security and authenticity verification. Our key contributions are as follows: + +- We introduce a novel watermarking approach that embeds message bits in high-frequency spectral components via wavelet and spectral projection inside hidden convolutional layers, ensuring robustness against various transformations and adversarial attacks. +- We adapt Parseval's theorem [20] as a learnable threshold to optimize SpecGuard and spectral masking for robust watermark bit recovery under diverse transformations including distortions, regeneration, and adversarial attacks, proven through the experimental results. +- Our extensive evaluations demonstrate SpecGuard's superior bit embedding capacity and producing better invisible watermarked images, surpassing the performance of state-of-the-art (SOTA) methods. + +# 2. Related Works + +Watermarking an image has been a widely researched topic for securing the ownership and verifying authenticity of digital content [37]. Traditional watermarking techniques typically embed invisible [42] or visible [7] watermarks into images, which can later be extracted or detected to verify the content's originality. These methods can be broadly classified into spatial-domain [39, 44] and frequency-domain [13] watermarking, while some are based on combined methods [38, 50]. However, researchers recently proposed many advanced models [23, 27, 31, 40] for effective watermark removal. To face this growing challenge, researchers introduced different methods [2, 3, 15, 22, 28, 51, 53] as alternatives to deep learning-based encoders or decoders to produce more robust image watermarking. Furthermore, iterative models have demonstrated competitive performance [21, 34], particularly in robustness against a wide range of transformations. In addition, with the rise of generative methods, researchers used the watermark-labeled data for training to learn how to produce watermarks [11, 24]. Also, models that combine generative methods with watermarking techniques show promise in effective image watermarking [25, 30, 36]. However, such approaches face limitations such as increased computational complexity and longer processing times. These approaches are also more vulnerable to adversarial attacks that can target and distort the embedded watermark without altering the content visibly. + +# 3. Proposed Method: SpecGuard + +We introduce SpecGuard, as illustrated in Fig. 2, which involves two fundamental modules: an "Encoder" for embedding the watermark and a "Decoder" for accurately extracting the watermark detailed in the following sections. + +# 3.1. Encoder + +By targeting high-frequency components, the encoder integrates a binary message $M$ into the cover image $I$ . Using wavelet projection (WP) [33] and a Fast Fourier Transform (FFT)-based spectral projection (SP) [18] approximation, the message $M$ is inserted into specific frequency bands, minimizing perceptual impact. + +Wavelet Projection. We use a wavelet projection to capture frequency and spatial localization features that describe an image across different scales, as shown in Eq. (1): + +$$ +W (a, b) = \frac {1}{\sqrt {| a |}} \int_ {- \infty} ^ {\infty} f (x) \psi \left(\frac {x - b}{a}\right) d x, \tag {1} +$$ + +where $a \in \mathbb{R} \setminus \{0\}$ , and $b \in \mathbb{R}$ denote the scaling and translation parameters, respectively. Here, $\psi_{a,b}(x)$ represents a rescaled and translated form of the mother wavelet $\psi$ , defined as follows: + +$$ +\psi_ {a, b} (x) = \psi \left(\frac {x - b}{a}\right) \cdot \frac {1}{\sqrt {| a |}}, \tag {2} +$$ + +where $\frac{1}{\sqrt{|a|}}$ functions as a normalization factor, guaranteeing that the energy of the wavelet is invariant to the scaling parameter $a$ . Minimal values of $a$ compress the wavelet, enabling the inspection of high-frequency components, whereas greater values of $a$ elongate the wavelet, promoting low-frequency analysis. Since each mother wavelet $\psi$ is built with zero mean and finite energy [12], it guarantees to maintain stability as follows: + +$$ +\int_ {- \infty} ^ {\infty} \psi (x) d x = 0, \quad \int_ {- \infty} ^ {\infty} | \psi (x) | ^ {2} d x < \infty , \tag {3} +$$ + +where the wavelet projection from Eq. (1) decomposes the input into orthogonal wavelet sets using discrete scales and translations. For $2D$ inputs, the scaled and translated basis elements [1] are defined for each coordinate pair $(u,v)$ : + +$$ +\mathbf {S} _ {L L} = \phi (u, v) = \phi (u) \phi (v), \qquad \mathbf {S} _ {L H} = \psi_ {H} (u, v) = \psi (u) \phi (v), +$$ + +$$ +\mathbf {S} _ {H L} = \psi_ {V} (u, v) = \phi (u) \psi (v), \quad \mathbf {S} _ {H H} = \psi_ {D} (u, v) = \psi (u) \psi (v), \tag {4} +$$ + +where, $H$ , $V$ , and $D$ represent the horizontal, vertical, and diagonal decomposition direction, respectively. To depict the image at different resolutions, we define scaling and wavelet functions at scale $j$ as shown below: + +$$ +\phi_ {j, m, n} (u, v) = 2 ^ {j / 2} \phi \left(u - \frac {m}{2 ^ {j}}, v - \frac {n}{2 ^ {j}}\right), \tag {5} +$$ + +$$ +\psi_ {j, m, n} ^ {d} (u, v) = 2 ^ {j / 2} \psi^ {d} \left(u - \frac {m}{2 ^ {j}}, v - \frac {n}{2 ^ {j}}\right), +$$ + +![](images/d547e59ba83c2de38ccd65c824e949168c58fb8718a8d8f0e3fe187603ffd931.jpg) +Figure 2. Architecture of the proposed SpecGuard watermarking method involves encoding a binary message $M$ into the high-frequency band of the cover image $I$ using wavelet and spectral projection and learning to decode the embedded message. + +where $d \in \{H, V, D\}$ is the wavelet function direction that serves as discrete basis elements for multi-resolution analysis, capturing details across frequency bands and spatial locations. In Eq. (6), $T_{m,n}$ denotes the intensity or pixel value of the cover image $I$ at spatial coordinates $(m, n)$ . The discrete scaling function $W_{\phi}(j, u, v)$ (approximation at scale $j$ ) and the detail coefficients $W_{\psi}^{d}(j, u, v)$ for each direction are computed accordingly as follows: + +$$ +W _ {\phi} (j, u, v) = \frac {1}{l} \sum_ {m = 0} ^ {l - 1} \sum_ {n = 0} ^ {l - 1} T _ {m, n} \phi \left(m - u \cdot 2 ^ {- j}, n - v \cdot 2 ^ {- j}\right), +$$ + +$$ +W _ {\psi} ^ {d} (j, u, v) = \frac {1}{l} \sum_ {m = 0} ^ {l - 1} \sum_ {n = 0} ^ {l - 1} T _ {m, n} \psi^ {d} \left(m - u \cdot 2 ^ {- j}, n - v \cdot 2 ^ {- j}\right), \tag {6} +$$ + +with $l$ as the discrete region dimension, these coefficients capture multi-scale, multi-orientation image information, forming the basis of spectral features as follows: + +$$ +\beta_ {j} = \bigcup_ {d \in \{H, V, D \}} \left(W _ {\phi} (j, u, v) \cup W _ {\psi} ^ {d} (j, u, v)\right). \tag {7} +$$ + +This feature set $\beta_{j}$ captures key frequency and spatial details across resolutions, forming the foundation for the watermark embedding process of our SpecGuard. + +Selective Frequency Band Decomposition. To refine the embedding process, we segment the data into distinct frequency bands. The decomposition level $\kappa$ is determined by the image complexity, calculated as follows: + +$$ +\kappa = \left\lfloor \sqrt {\log (1 + N)} \right\rfloor , \tag {8} +$$ + +where $N$ denotes the total pixel count in the cover image $I$ . And, each component $\beta_{j}$ falls within a unique frequency band, yielding a total of $1 + 3\kappa$ distinct frequency bands as follows: + +$$ +\beta_ {j} = \phi_ {j} (u, v) \cup \bigcup_ {d \in \{H, V, D \}} \psi_ {j} ^ {d} (u, v). \tag {9} +$$ + +The components $\beta_{j}$ , consisting of scaling functions $\phi_j(u,v)$ and wavelet functions $\psi_j^d (u,v)$ , capture specific spatial frequency bands, enabling targeted high-frequency embedding. We translate the WP into disjoint intervals representing a unique frequency range to approximate the segmentation in the frequency domain: + +$$ +\beta_ {j} = \left\{W _ {\psi} ^ {d} (u, v) \mid u, v \in \left(\frac {j \cdot L}{\kappa}, \frac {(j + 1) \cdot L}{\kappa}\right) \right\}, \tag {10} +$$ + +where, $L$ is the dimension of $S_{HH}$ , and $W_{\psi}^{d}(u,v)$ represents wavelet values within segmented intervals. This frequency band partitioning mimics the frequency selectivity of wavelet sub-bands, enabling effective targeting of high-frequency regions for optimal embedding. + +Approximation of Spectral Projection. We first apply spectral projection on the $S_{HH}$ sub-band, transforming it into the spectral domain. Given a matrix $T(x,y)$ representing pixel intensities in $S_{HH}$ , the spectral projection computes the spectral components $\zeta(u,v)$ as follows: + +$$ +\zeta (u, v) = \frac {1}{L ^ {2}} \sum_ {x} \sum_ {y} T (x, y) \cdot \exp \left(- i \frac {2 \pi}{L} (x \cdot u + y \cdot v)\right), \tag {11} +$$ + +where $L$ denotes the dimension of $S_{HH}$ , $T(x,y)$ provides the intensity at each coordinate $(x,y)$ which is equivalent to $W_{\psi}^{d}(u,v)$ in Eq. (6), $i$ is the imaginary unit, and $(u,v)$ are the spectral coordinates. + +To approximate the spectral components using the FFT, we create a symmetrically extended version $\tilde{T}(x,y)$ of the original $N\times N$ matrix $T(x,y)$ . This extension is achieved by mirroring $T(x,y)$ along its boundaries, doubling its size to $2N\times 2N$ . Specifically, the original matrix occupies the top-left quadrant, with the remaining quadrants filled by reflecting $T(x,y)$ horizontally, vertically, and diagonally, respectively. This symmetric structure ensures that the FFT yields only real values, allowing the spectral coefficients to + +be extracted directly from the real part of the FFT operation. Then, we apply the 2D FFT to $\tilde{T} (x,y)$ as follows: + +$$ +F (u, v) = \frac {1}{(2 N) ^ {2}} \sum_ {x} \sum_ {y} \tilde {T} (x, y) \cdot \exp \left(- i \frac {2 \pi}{2 N} (x \cdot u + y \cdot v)\right) \tag {12} +$$ + +The SP coefficients are then approximated by taking the real part (Re) of $F$ in the original $N \times N$ region as follows: + +$$ +\zeta (u, v) \approx \operatorname {R e} (F (u, v)), \quad 0 \leq u, v < N. \tag {13} +$$ + +Applying Eq. (13) to the sub-bands extracted from wavelet projection in Eq. (6), we achieve a computationally efficient spectral projection by leveraging the FFT approximation on a symmetrically extended matrix, maintaining effective embedding properties within the spectral domain. + +SpecGuard Embedding Process. The embedding process integrates the binary message $M$ into the high-frequency band $S_{HH}$ of the cover image $I$ , enhancing robustness and imperceptibility through wavelet and spectral projection. Using the Eq. (6) and Eq. (13), the cover image $I$ is decomposed into sub-bands $S_{LL}, S_{LH}, S_{HL}$ , and $S_{HH}$ within spectral domain, with $S_{HH}$ providing high-frequency details for embedding. A variable number $k$ of convolutional layers with a $K \times K$ kernel, followed by LeakyReLU activation, are recursively applied to $S_{HH}$ to refine features: + +$$ +S _ {H H} ^ {(n + 1)} = \text {L e a k y R e L U} \left(\operatorname {C o n v} _ {2 D} \left(S _ {H H} ^ {(n)}, K\right)\right), \quad n = 1, \dots , k. \tag {14} +$$ + +The final output $S_{HH}^{(n + 1)}$ from Eq. (14) represents the modified high-frequency band, primed for embedding. + +The message $M$ , represented as a binary vector of length $l$ ( $M \in \{0,1\}^l$ ), with batch size $b$ and message length $l$ , is reshaped and expanded across channels $c$ to align with $S_{HH}^{(n+1)}$ . This ensures $M_{\mathrm{expanded}}$ conforms to the dimension $[b,c,l]$ , where each message is structured accordingly. + +To localize the embedding, we create a radial mask centered at $(c_x, c_y) = \left(\frac{h}{2}, \frac{w}{2}\right)$ , where $h$ and $w$ represent the height and width of $S_{HH}$ . The Euclidean distance $D(x_i, y_i)$ from the center $(c_x, c_y)$ is computed for each coefficient $(x_i, y_i)$ . A binary mask is then generated within the predefined radius $r$ based on the distance $D(x_i, y_i)$ , such that if $D(x_i, y_i) \leq r$ , the mask value is set to 1, allowing embedding in the corresponding region. Otherwise, the mask value is 0, restricting embedding to areas within a specified radius $r$ , ensuring focus on high-frequency regions. + +For each coordinate $(x_{i},y_{i})$ where mask $(x_{i},y_{i})$ is 1 and $W_{c}\in c$ , the embedding operation is performed as follows: + +$$ +S _ {H H} ^ {(n + 1)} [:, W _ {c}, x _ {i}, y _ {i} ] + = M _ {\text {e x p a n d e d}} [:, W _ {c}, i ] \cdot s, \tag {15} +$$ + +where $s$ is the strength factor controlling embedding intensity and invisibility. After embedding, the modified coefficients $S_{HH}^{(n + 1)}$ undergo a final convolution and LeakyReLU using Eq. (14), by setting the value of $k = 1$ to harmonize + +the embedded message. Following this approach, SpecGuard embeds the message into the spectral domain in a transformed form, differing from its original input representation. By blending the message seamlessly into the spectral space based on the $r$ , $s$ , and $W_{c}$ , it becomes inherently concealed within the domain, rendering its presence imperceptible. Without knowledge of $r$ , $s$ , and $W_{c}$ , it becomes exceedingly challenging to localize the embedded message, further enhancing the security of the system. This transformation ensures the embedding process remains opaque to any adversarial attacker, making SpecGuard black-box. + +Reconstruction. SpecGuard encoder reconstructs the watermarked image $I_{\mathrm{embedded}}$ by inverse transformation restoring $S_{HH}$ back into the spatial domain. The reconstruction process integrates the inverse wavelet projection (IWP) [33] and inverse spectral projection (ISP) [18], ensuring the embedded modifications are correctly translated into the spatial domain. To reconstruct the spatial domain image, $S_{HH}$ is combined with the other sub-bands $S_{LL}, S_{LH}$ , and $S_{HL}$ . For the SP embedded in $S_{HH}$ , the ISP is applied to reconstruct $S_{HH}$ to spatial domain as follows: + +$$ +S _ {H H} (x, y) = \sum_ {u = 0} ^ {L - 1} \sum_ {v = 0} ^ {L - 1} \zeta (u, v) \cdot \exp \left(i \frac {2 \pi}{L} (x \cdot u + y \cdot v)\right), \tag {16} +$$ + +where $\zeta(u, v)$ represents spectral coefficients from the embedding process, $L$ denotes the dimension of $S_{HH}$ , and $(x, y)$ are spatial coordinates. SpecGuard then reconstructs the watermarked image $I_{\mathrm{embedded}}$ using the IWP as follows: + +$$ +I _ {\text {e m b e d d e d}} (x, y) = \operatorname {I W P} \left(S _ {L L}, S _ {L H}, S _ {H L}, S _ {H H}\right). \tag {17} +$$ + +This process seamlessly embeds the watermark message $M$ in the spectral domain, preserving the cover image $I$ 's integrity. The inverse transformations that are expressed in Eq. (16) and Eq. (17) fully restore visual quality, maintaining all frequency components. + +# 3.2.Decoder + +As shown in Algorithm 1, SpecGuard decoding process starts by applying wavelet projection ( Eq. (1)) to the watermarked image $I_{\mathrm{embedded}}$ , separating it into low and high-frequency bands, where the high-frequency band $S_{D_{HH}}^{high}$ contains the embedded message similar to the process in the encoding phase, particularly in Eq. (6). An approximation of the spectral projection using FFT as shown in Eq. (13) is then applied to $S_{D_{HH}}^{high}$ returning the transformed data $S_{D_{HH}}^{\mathrm{sp}}$ . Then, $S_{D_{HH}}$ is further refined through convolutional layers that captures the local features for message extraction. + +To extract the message, a radial mask is created to isolate high-frequency areas within $S_{D_{HH}}$ , targeting the embedded regions based on their distance from the center. The masked values are compared against a learnable threshold $\theta$ to decode each bit of the hidden message $D_{\mathrm{M}}$ . Here, $\theta$ serves as a threshold that adapts to the spectral patterns across the entire image, learning the distinct characteristics + +Algorithm 1 SpecGuard decoder with wavelet, spectral projection with FFT approximation, and learnable threshold. +1: Input: Watermarked image $I_{\mathrm{embedded}}$ , learnable $\theta$ message length $l$ +radius $r$ ,watermark channel $W_{c}$ +2: Output: Decoded binary message $D_{\mathrm{M}}$ +3: Procedure: Apply Wavelet Projection on $I_{\mathrm{embedded}}$ to obtain $S_{D_{LL}}$ (low-frequency) and $S_{D_{HH}}^{\mathrm{high}}$ (high-frequency) +4: Procedure: Spectral approximation with FFT $(S_{D_{HH}}^{\mathrm{high}})$ .. +5: Separate even and odd indices: $v = [x_{\mathrm{even}},$ reverse $(x_{\mathrm{odd}})]$ +6: Compute FFT on $v$ .. $V_{\mathrm{complex}} = \mathrm{FFT}(v)$ +7: $V_{\mathrm{real}} = V_{\mathrm{complex}}\cdot \left[\cos \left(\frac{-\pi k}{2N}\right),\sin \left(\frac{-\pi k}{2N}\right)\right] / /$ Calculate Real +8: $V_{\mathrm{real}}[0]\gets \frac{V_{\mathrm{real}}[0]}{\sqrt{N\cdot 2}},V_{\mathrm{real}}[1:\gets \frac{V_{\mathrm{real}}[1:]}{\sqrt{\frac{N}{2}\cdot 2}} /$ Energy preservation +9: Transpose result and repeat to obtain $S_{D_{HH}}^{\mathrm{sp}}$ +10: Return $S_{D_{HH}}^{\mathrm{sp}}$ +11: Procedure: Pass $S_{D_{HH}}^{\mathrm{sp}}$ through sequential layers as: + $S_{D_{HH}}^{(n + 1)} =$ LeakyReLU $\left(\operatorname {Conv}_{2D}\left(S_{HH}^{sp(n)},K\right)\right)$ $n = 1,\dots,k,$ +12: Return $S_{D_{HH}}^{(n + 1)}$ +13: Procedure: Extraction $(S_{D_{HH}}^{(n + 1)},l)$ . +14: Set $(c_x,c_y) = \left(\frac{H}{2},\frac{W}{2}\right)$ +15: Generate mask for high-frequency region within radius $r$ +for each coordinate $(i,j)$ do: + $D(x_i,y_i) = \sqrt{(x_i - c_x)^2 + (y_i - c_y)^2}$ // Euclidian Distance +17: if $D(x_i,y_i)\leq r$ then +18: Set mask[i, $j] = 1$ +19: end if +20: end for +21: Extract mask: $S_{D_{HH}}[:,W_c,$ mask[i, $j]$ +22: Decode message using learnable $\theta$ .. + +$$ +D _ {\mathbf {M}} [ i ] = \left\{ \begin{array}{l l} 1 & \text {i f E x t r a c t e d} [ i ] > \theta \\ 0 & \text {o t h e r w i s e} \end{array} \right. +$$ + +23: Update $\theta$ dynamically: $\theta \gets \theta -\eta \cdot \frac{\partial L_{\mathrm{dec}}}{\partial\theta}$ // Optimizes robustness +24: Return $D_{\mathrm{M}}$ + +of the embedded watermark. From Parseval's theorem [20] ensures overall spectral and spatial energies remain equivalent, though local spectral energy distributions are altered by the watermark strength factor $s$ . + +The watermark's strength factor $s$ ensures that the high-energy areas where the message $M$ is embedded as "1" remain robust, experiencing a minimum distortion in such conditions. Moreover, this threshold can be optimized for better bit recovery accuracy during training. As $\theta$ learns, it recognizes that areas encoded as "1" carry higher energy and impact due to the strength factor $s$ of Eq. (15), while areas marked as "0", softened by the LeakyReLU's minimal negative slope, have a lower intensity. Such a dynamic approach enables $\theta$ to identify and protect the embedded message $M$ even when external disturbances occur, preserving the watermark's structure within the watermarked image $I_{\text{embedded}}$ . And, $\theta$ effectively learns to distin + +guish high-energy watermark regions. Therefore, the embedded message is more recoverable under diverse attacks, and SpecGuard's decoder ensures valid watermark bit extraction. Theoretical explanation of Parseval theorem's [20] impact on message extraction is in the Supplementary. + +# 3.3. Loss Calculation for SpecGuard + +To achieve the training objective of robust and invisible watermark embedding, a composite loss function is defined with two terms: encoder loss $L_{\mathrm{enc}}$ as expressed in Eq. (18) and decoder loss $L_{\mathrm{dec}}$ as expressed in Eq. (19). + +$$ +\min _ {\theta} \mathbb {E} _ {(I, M) \sim D} L _ {\text {e n c}} (I, I _ {\text {e m b e d d e d}}) = \| E _ {\theta} (I, M) - I \| ^ {2}, \tag {18} +$$ + +$$ +\min _ {\theta} \mathbb {E} _ {(I, M) \sim D} L _ {\mathrm {d e c}} (M, D _ {\mathrm {M}}) = \| D _ {\theta} \left(I _ {\text {e m b e d d e d}}\right) - M \| ^ {2}, \tag {19} +$$ + +where $E_{\theta}(I,M)$ denotes the encoder output, embedding the message $M$ into the cover image $I$ to produce $I_{\mathrm{embedded}}$ . By minimizing $L_{\mathrm{enc}}$ , the encoder learns to embed the watermark invisibly, preserving the fidelity of the cover image. $D_{\theta}(I_{\mathrm{embedded}})$ denotes the decoder's output from the watermarked image $I_{\mathrm{embedded}}$ . Minimizing $L_{\mathrm{dec}}$ enables the decoder to reliably retrieve the embedded message under varying conditions, such as noise and transformation. + +The total loss $L$ as shown in Eq. (20) used for optimizing the model combines these terms, balancing invisibility and robustness through weighted coefficients as follows: + +$$ +\min _ {\theta} L = \lambda_ {\text {e n c}} L _ {\text {e n c}} + \lambda_ {\text {d e c}} L _ {\text {d e c}}, \tag {20} +$$ + +where $\lambda_{\mathrm{enc}}$ and $\lambda_{\mathrm{dec}}$ control the relative importance of visual fidelity and message recoverability. + +# 4. Experimental Results + +# 4.1. Dataset + +SpecGuard is trained on the MS-COCO dataset [26], which contains 25K images. To evaluate the robustness of the watermarking methods including our SpecGuard against different types of attacks, such as distortions, regenerations, and adversarial attacks, we used three datasets: DiffusionDB [46], MS-COCO [26], and DALL-E31. Each of these datasets has a unique distribution of prompt words. We also ensured that no unethical or violent terms were included in the prompts. We randomly picked 200 images from MS-COCO [26] and applied watermark using SpecGuard for further verifying the robustness after uploading on various social media platforms and applying AI-based Photoshop Neural Filters (PNFs)2. The PNFs include depth blur, artistic style transfer, super zoom, JPEG artifact reduction, and colorization. For the super zoom filter, we set the 'Sharpen' and 'Noise Reduction' parameters to 15. For all other filters, we used the default settings. + +![](images/a0aa4adfdd645083602c529592e0a888bef0b968035d569a2332b2aa6bacde2f.jpg) +Figure 3. Some best results for cover vs watermarked images with PSNR/SSIM $(\uparrow)$ scores showing minimal visual degradation when watermarked using proposed SpecGuard. + +
Metrics256 × 256512 × 5121024 × 1024
CelebA-HQMS-COCOCelebA-HQMS-COCOCelebA-HQMS-COCO
PSNR↑40.36140.32044.65144.68048.17048.081
SSIM↑0.98890.98880.99270.99270.99370.9936
FID↓16.45116.69016.97217.02017.44616.955
MSE↓0.00020.00020.00010.00010.00010.0001
+ +# 4.2. Implementation + +We used CUDA v11.3 and PyTorch with a batch size of 32 and the Adam optimizer on a multiple NVIDIA RTX 2080-equipped server. Mean Squared Error (MSE) and Bit Recovery Accuracy (BRA) are used for loss and accuracy calculation. We used Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), Fréchet Inception Distance (FID), and MSE to evaluate perceptual quality. Our model is trained for 300 epochs, with the decoder learning rate set to $1 \times 10^{-3}$ , reduced by half every 100 steps, and the encoder learning rate is set to $1 \times 10^{-2}$ without scheduling. We set our watermark radius $(r)$ , strength factor $(s)$ , initial learning parameter $(\theta)$ , and the number of convolutional layers $(k)$ to 100, 20, 0.001, and 32, respectively. This setup is applied with a message bit length (BL) of 48, 64, 128, and 256. Initially, decoder loss weight $(\lambda_{\mathrm{dec}})$ and encoder loss weight $(\lambda_{\mathrm{enc}})$ are set to 1.0 and 0.7, respectively. For assessing the robustness of watermarking methods against diverse attacks, we inherited the experimental setups from Waves [8] and used effective metrics such as "Quality at 95% Performance (Q@0.95P)", "Quality at 70% Performance (Q@0.7P)", "Avg P" and "Avg Q." Here, Q@0.95P and Q@0.7P indicate the level of image quality degradation required for watermark detection accuracy to reach 95% and 70%, respectively. The average performance (Avg P) metric represents the mean detection accuracy across various attack strengths, while the average quality degradation (Avg Q) measures the overall impact of attacks on image quality [8, 35]. + +Table 1. Perceptual quality of SpecGuard watermarked images across resolutions and datasets using a fixed 30-bit message length. + +
MethodsVenueBLPSNR↑SSIM↑FID↓BRA↑
Pre-processing methodsTree-Ring [47]NeurIPS'236432.330.9117.70.98
12832.100.9017.80.96
25631.850.8917.90.94
Stable Signature [35]ICCV'236430.000.8919.60.98
12829.800.8819.70.96
25629.500.8719.80.96
Yang et al. [49]CVPR'246431.450.9018.20.98
12831.200.8918.30.93
25630.950.8818.40.89
SleeperMark [45]CVPR'256431.800.9218.00.97
12831.600.9118.10.93
25631.350.9018.20.87
Post-processing methodsHiDDeN [53]ECCV'186432.010.8819.70.98
12831.800.8719.80.85
25631.500.8619.90.82
StegaStamp [41]CVPR'206428.500.9117.90.99
12828.200.9018.00.98
25628.000.8918.10.94
MBRS [17]ACM MM'216438.200.9617.90.98
12837.900.9518.00.96
25637.500.9418.20.94
FIN [10]AAAI'236436.700.9518.30.97
12836.400.9418.40.96
25636.100.9318.50.96
MuST [43]AAAI'246441.200.9717.50.98
12840.900.9617.60.93
25640.500.9517.80.90
EditGuard [52]CVPR'246441.560.9717.80.98
12841.300.9617.90.97
25640.900.9518.00.97
SpecGuard (Ours)ICCV'256442.590.9817.20.99
12842.890.9917.00.99
25640.860.9917.60.98
+ +Table 2. Comparison of SOTA pre-processing and post-processing watermarking methods with SpecGuard without attacks. + +# 4.3. Watermark Invisibility + +To evaluate the invisibility of the embedded watermark, we conducted perceptual and quantitative assessments using SpecGuard. As shown in Fig. 3, there is no noticeable perceptual degradation between the cover and watermarked images, confirming that the watermark remains imperceptible to the human eye. For a more comprehensive evaluation, we created three subsets of different image sizes ranging between 256 to 1024 with images from the MS-COCO [26] and CelebA-HQ [19] datasets and applied the SpecGuard watermarking method to compare the average PSNR values between the cover and watermarked images, as in Tab. 1. + +For quantitative evaluation, we further compare the performance of SpecGuard with the SOTA pre-processing and post-processing watermarking methods. As presented in Tab. 2, SpecGuard achieves the highest PSNR of 42.89 when the bit length was 128. Additionally, it attains the highest SSIM of 0.99 at a BL of 128 and 256 among all compared methods, indicating minimal visual distortion. Additionally, SpecGuard achieved the lowest FID of 17.0 and the highest BRA of 0.99, ensuring strong robustness while maintaining imperceptibility. Overall, our results demonstrate that SpecGuard outperforms both pre-processing and post-processing watermarking methods, achieving superior imperceptibility and robustness. + +
Attack TypeTree-Ring [47]Stable Signature [35]StegaStamp [41]SpecGuard (Ours)
Q@0.95PQ@0.7PAvg PAvg QQ@0.95PQ@0.7PAvg PAvg QQ@0.95PQ@0.7PAvg PAvg QQ@0.95PQ@0.7PAvg PAvg Q
DistortionsRotation0.4640.5210.3750.6480.6240.7020.5940.6500.4230.4980.3570.6160.8630.8630.6870.653
Crop0.5920.5920.3320.463infinf0.9950.4610.6020.6020.5400.4510.8120.8120.9980.742
Brightinfinfinf0.304infinf0.9980.305infinf0.9980.317infinf0.9980.466
Contrastinfinf0.9980.243infinf0.9980.243infinf0.9980.231infinf0.9980.556
Blur0.8611.1120.5631.221- inf- inf0.0001.2040.8480.9620.4141.0000.921inf1.0001.452
Noise0.548inf0.9800.3950.4020.5200.8700.390infinf1.0000.360infinf0.9990.568
JPEG0.4990.4990.9290.2840.4850.4850.7930.284infinf0.9980.263infinf1.0000.495
Geo0.5250.5930.2770.7680.850inf0.9370.7670.6630.6930.3960.7330.8690.8690.8650.623
Deg0.620inf0.8920.6940.2060.3690.3000.6790.8260.9750.8520.6640.8951.1410.9150.749
Combine0.5390.7510.4030.9080.5380.6910.3340.9000.9451.1010.7950.8700.9791.2560.9110.952
RegenerationRegen-Diff- inf0.3070.6120.323- inf- inf0.0010.3000.331inf0.9430.327infinf0.9820.477
Regen-DiffPinf0.3070.6010.327- inf- inf0.0010.3030.333inf0.9400.329infinf0.9820.562
Regen-VAE0.5780.5780.8320.3480.5450.5450.5160.339infinf1.0000.343infinf0.9950.521
Regen-KLVAEinfinf0.9900.233 6- inf0.1760.2170.206infinf1.0000.240infinf0.9900.492
Rinse-2xDiff- inf0.3330.5100.357- inf- inf0.0010.3320.391inf0.9410.366infinf0.9930.561
Rinse-4xDiff- inf0.3550.4430.466- inf- inf0.0000.4380.388inf0.9090.477infinf0.9920.533
AdversarialAdvEmbG-KLVAE8- inf0.1640.4480.253infinf0.9980.249infinf1.0000.232infinf1.0000.456
AdvEmbB-RN180.241inf0.9530.218infinf0.9990.212infinf1.0000.196infinf1.0000.467
AdvEmbB-CLIP0.541inf0.9320.549infinf0.9990.541infinf1.0000.488infinf1.0000.436
AdvEmbB-KLVAE160.195inf0.8880.238infinf0.9970.233infinf1.0000.206infinf1.0000.482
AdvEmbB-SdxlVAE0.222inf0.9340.221infinf0.9980.219infinf1.0000.204infinf1.0000.492
AdvCls-UnWM&WM- inf0.1020.4990.145infinf0.9990.101infinf1.0000.101infinf1.0000.497
AdvCls-Real&WMinfinf1.0000.047infinf0.9980.092infinf1.0000.106infinf1.0000.427
AdvCls-WM1&WM2- inf0.1010.4920.139infinf0.9990.084infinf1.0000.129infinf1.0000.441
+ +Table 3. Robustness comparison various across attacks using Q@0.95P(↑), Q@0.7P(↑), Avg P(↑) and Avg Q(↑). Here, 'inf' denotes that no attack was sufficient to degrade performance below the threshold, indicating strong robustness, whereas '-inf' signifies that even the weakest attack caused detection to fall below the threshold, reflecting weak robustness. + +![](images/c3af77e50ef9cbb097edd6fa72e5f5dc1f03dcdb43ce23ff8d8a8d959ba8cd77.jpg) + +![](images/241c17b3a2b9fbfba53c2704606421d5b419030a809fb05f6cd4584c9e9e8191.jpg) + +![](images/3df03b9919033d0e53d1f6dd7d95dbe480ad075013eae83e0ac50a1c53f49898.jpg) + +![](images/4b653ed4ee169dd20947b2d3dc9e8d5e0b593ccc7fedcef8cab76ddd2d56d318.jpg) + +![](images/5270cac58775daa7649a400c4517462d0dd4897b733eecdbe49d58293744c7ee.jpg) + +![](images/b65904e3bc4503ca2ba4bee9c230545e7542f295729e50187ca6365eaafd7b81.jpg) +Figure 4. Robustness validation of our proposed SpecGuard under different distortion attacks, including geometric transformations: Geo (rotation, cropping), photometric modifications (brightness, contrast), and degradations: Deg (blur, noise, JPEG compression). + +![](images/789232424565e24e2ea17562c99484682a1733df209991e89e396a6149503d4e.jpg) + +![](images/64eea155e5ca4eeaa015eb0b296aef43acd61b56cdf4f38fc6fa2b72180833f1.jpg) + +![](images/b46ee883b746c9f9e646c4abf09ba7c7cfcca0168764f5b5c5f5bd6ea3f5e85a.jpg) + +![](images/5aacc06a0c1b739db3d3ea1c7bb773eadce67648e90eac85bd0025e7b27edb56.jpg) + +# 4.4. Capacity + +To evaluate embedding capacity, we examined SpecGuard across different bit lengths and compared it with SOTA watermarking methods. Our experiments with 64, 128, and 256 bits demonstrate SpecGuard's high capacity of bit embedding while maintaining perceptual quality and robustness, as shown in Tab. 2. Notably, it achieves a PSNR of 42.89, the highest among all methods, along with the high + +est BRA of 0.99 and the lowest FID of 17.0 at 128 bits, ensuring minimal visual impact. This adaptability to different bit lengths without quality loss makes SpecGuard ideal for applications requiring flexible watermark sizes. Unlike StegaStamp and HiDDeN, which suffer reduced BRA for higher message bits, SpecGuard consistently extracts bits across all tested lengths. SpecGuard's theoretical watermark capacity is provided in the Supplementary. + +
ModulesPSNR/SSIM↑BRA↑ModulesPSNR/SSIM↑BRA↑
WP(L1)40.51/0.960.92WP(L1)+SPFA42.89/0.990.99
WP(L2)38.15/0.930.87WP(L2)+SPFA36.25/0.920.89
AttacksPSNR/SSIM↑BRA↑AttacksPSNR/SSIM↑BRA↑
Rotate (45°)12.15/21.310.82Rotate (90°)11.15/19.310.65
Blur (0.3)35.01/0.950.98Blur (0.6)30.11/0.910.98
Geo (0.3)12.08/0.500.93Geo (0.6)10.25/0.450.86
+ +*WP: Wavelet Projection, SP: Spectral Projection, FA: FFT Approximation + +Table 4. Ablation studies on the proposed SpecGuard for across various configurations, setting $M = 128$ , $r = 100$ , and $s = 20$ . + +
PlatformPSNR/SSIM↑BRA↑PS FiltersPSNR/SSIM↑BRA↑
Facebook48.56/0.970.97Depth Blur25.25/0.890.85
LinkedIn47.55/0.970.96StyleT.25.12/0.840.85
Instagram48.56/0.980.98Super Zoom36.15/0.880.95
WhatsApp42.10/0.960.97JPEG Artifacts31.01/0.850.94
X (Twitter)49.25/1.000.99Colorize23.15/0.820.92
+ +Table 5. Evaluation of SpecGuard's robustness across Photoshop filters and while uploaded on different social media platforms. + +# 4.5. Robustness + +We evaluate watermarking robustness by analyzing detection performance against a range of diverse and challenging real-world attacks. Results demonstrate the strong robustness of SpecGuard across various attacks. For example, as presented in Tab. 3, against geometric distortions such as cropping and rotation, SpecGuard achieved an Avg P of 0.998 and 0.687, respectively. Similarly, across the combined distortion-based attacks, SpecGuard achieves an overall Avg P of 0.911 and Avg Q of 0.952, ensuring minimal quality loss while maintaining high detection accuracy. Notably, the high values of Q@0.95P and Q@0.7P indicate that SpecGuard can sustain reliable detection at strict performance thresholds, even under aggressive perturbations. Unlike prior methods that struggle with extreme transformations, SpecGuard shows remarkable robustness against regeneration-based attacks like Rinse-2xDiff [6] (an image is noised then denoised by Stable Diffusion v1.4 two times with strength as a number of timesteps, 20-100) and RegenVAE [6], maintaining high Avg P. Similarly, under adversarial attacks, SpecGuard consistently secures watermark detectability, outperforming existing techniques across all tested scenarios. These results establish SpecGuard as a highly robust watermarking approach capable of preserving image integrity even under severe distortions and adversarial manipulations, ensuring watermark reliability across diverse attack types. More details about how the attacks are performed are provided in supplementary material. Further, our results in Fig. 4 highlight the strong robustness of SpecGuard against various distortion attacks compared to other SOTA watermarking methods. + +Social Platforms and Photoshop Filters. SpecGuard's robustness when images are shared across social media platforms and subjected to common Photoshop Neural Filters (PNFs) is shown in Tab. 5. SpecGuard consistently maintains high PSNR and SSIM values, with BRA values close to 0.99 on platforms such as X (formally Twitter), Insta + +gram, and Facebook. Also, it shown strong resilience to various PNFs, such as Super Zoom and JPEG Artifacts achieving BRA of 0.95 and 0.94. The PSNR, SSIM, and BRA values are expected to decrease with the severity of image manipulation, as increased manipulation leads to loss of image authenticity. For example, as we applied $60\%$ style transfer the PSNR and BRA decreased to 25.12 and 0.85. Similarly, the depth blur which excessively reduces the image clarity also causes the decrease of BRA to 0.85. + +# 4.6. Ablation Study + +We examined the impact of WP at different levels ( $L_{1}$ and $L_{2}$ ) and its combination with SP using $FA$ in Tab. 4. As observed, the $\mathrm{WP}(L_1) + \mathrm{SP}_{FA}$ configuration achieved the highest PSNR and SSIM values of 42.89 and 0.99, respectively, and BRA of 0.99, indicating improved watermark invisibility and robustness. In contrast, using WP alone at either $L_{1}$ or $L_{2}$ resulted in lower BRA, with values of 0.92 and 0.87, respectively, demonstrating that the combined $\mathrm{WP} + \mathrm{SP}_{FA}$ approach significantly enhances performance. We also evaluated the robustness of SpecGuard under strong adversarial attacks identified in Tab. 4, such as rotation, blur, and geometric transformations. The results indicate that higher levels of attack severity, such as $90^{\circ}$ rotation, lead to a more significant drop in PSNR, SSIM, and BRA, with values dropping to 11.15, 19.31, and 0.65, respectively. Despite this, the model shows relatively high resilience under moderate attack intensities, such as $45^{\circ}$ rotation and low levels of blur and geometric distortion, achieving BRA values as high as 0.93 under geometric transformations at the 0.3 thresholds. More ablations are in the supplementary. + +# 5. Conclusion + +We propose SpecGuard, a novel invisible watermarking method that ensures secure and robust information concealment. Unlike traditional approaches, SpecGuard remains highly resilient against diverse distortions, adversarial attacks, and regeneration-based transformations. Experimental results demonstrate its superior bit recovery accuracy of $99\%$ maintaining high PSNR. By outperforming SOTA watermarking methods in both detection reliability and imperceptibility, SpecGuard establishes a new benchmark for watermarking under real-world constraints. + +# Acknowledgments + +This work was partly supported by Institute for Information & communication Technology Planning & evaluation (IITP) grants funded by the Korean government MSIT: (RS-2022-II221199, RS-2022-II220688, RS-2019-II190421, RS-2023-00230337, RS-2024-00356293, RS-2024-00437849, RS-2021-II212068, RS-2025-02304983, and RS-2025-02263841). + +# References + +[1] Edward H Adelson, Eero Simoncelli, and Rajesh Hingorani. Orthogonal pyramid transforms for image coding. In Visual Communications and image processing II, pages 50-58. SPIE, 1987. 2 +[2] Mahdi Ahmadi, Alireza Norouzi, Nader Karimi, Shadrokh Samavi, and Ali Emami. Redmark: Framework for residual diffusion watermarking based on deep networks. Expert Systems with Applications, 146:113157, 2020. 2 +[3] Inzamamul Alam, Md Tanvir Islam, and Simon S Woo. Specxnet: A dual-domain convolutional network for robust deepfake detection. In Proceedings of the 33nd ACM International Conference on Multimedia, 2025. 2 +[4] Inzamamul Alam, Md Tanvir Islam, and Simon S Woo. Saliency-aware diffusion reconstruction for effective invisible watermark removal. In Companion Proceedings of the ACM on Web Conference 2025, pages 849-853, 2025. 1 +[5] Aashutosh AV, Srijan Das, Abhijit Das, et al. Latent flow diffusion for deepfake video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3781-3790, 2024. 1 +[6] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436, 2018. 8 +[7] Tali Dekel, Michael Rubinstein, Ce Liu, and William T Freeman. On the effectiveness of visible watermarks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2146-2154, 2017. 2 +[8] Mucong Ding, Tahseen Rabbani, Bang An, Aakriti Agrawal, Yuancheng Xu, Chenghao Deng, Sicheng Zhu, Abdirisak Mohamed, Yuxin Wen, Tom Goldstein, et al. Waves: Benchmarking the robustness of image watermarks. In ICLR 2024 Workshop on Reliable and Responsible Foundation Models, 2024. 1, 6 +[9] Hubert Etienne. The future of online trust (and why deepfake is advancing it). AI and Ethics, 1(4):553-562, 2021. 1 +[10] Han Fang, Yupeng Qiu, Kejiang Chen, Jiyi Zhang, Weiming Zhang, and Ee-Chien Chang. Flow-based robust watermarking with invertible noise layer for black-box distortions. In Proceedings of the AAAI conference on artificial intelligence, pages 5054-5061, 2023. 6 +[11] Jianwei Fei, Zhihua Xia, Benedetta Tondi, and Mauro Barni. Supervised gan watermarking for intellectual property protection. In 2022 IEEE International Workshop on Information Forensics and Security (WIFS), pages 1-6. IEEE, 2022. 2 +[12] S. A. Fulling. The local geometric asymptotics of continuum eigenfunction expansions. ii. one-dimensional systems. SIAM Journal on Mathematical Analysis, 14(4):605-623, 1983. 2 +[13] Mahdieh Ghazvini, Elham Mohamadi Hachrood, and Mojdeh Mirzadi. An improved image watermarking method in frequency domain. Journal of Applied Security Research, 12 (2):260-275, 2017. 2 +[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and + +Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1 +[15] Md Tanvir Islam, Ik Hyun Lee, Ahmed Ibrahim Alzahrani, and Khan Muhammad. Mexic: A meta ensemble explainable approach for ai-synthesized fake image classification. *Alexandria Engineering Journal*, 116:351–363, 2025. 2 +[16] K Jayashre and M Amsaprabhaa. Safeguarding media integrity: A hybrid optimized deep feature fusion based deepfake detection in videos. Computers & Security, 142:103860, 2024. 1 +[17] Zhaoyang Jia, Han Fang, and Weiming Zhang. Mbrs: Enhancing robustness of dnn-based watermarking by minibatch of real and simulatedJPEG compression. In Proceedings of the 29th ACM international conference on multimedia, pages 41-49, 2021. 6 +[18] Hoon Kang and Joonsoo Ha. Projection spectral analysis. International Journal of Control, Automation and Systems, 13(6):1530-1537, 2015. 1, 2, 4 +[19] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations (ICLR), 2018. 6 +[20] SS Kelkar, LL Grigsby, and J Langsner. An extension of parseval's theorem and its use in calculating transient energy in the frequency domain. IEEE Transactions on Industrial Electronics, (1):42-45, 1983. 2, 5 +[21] Varsha Kishore, Xiangyu Chen, Yan Wang, Boyi Li, and Kilian Q Weinberger. Fixed neural network steganography: Train the images, not the network. In International Conference on Learning Representations, 2021. 2 +[22] Jae-Eun Lee, Young-Ho Seo, and Dong-Wook Kim. Convolutional neural network-based digital image watermarking adaptive to the resolution of image and watermark. Applied Sciences, 10(19):6854, 2020. 2 +[23] Yicheng Leng, Chaowei Fang, Gen Li, Yixiang Fang, and Guanbin Li. Removing interference and recovering content imaginatively for visible watermark removal. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2983-2990, 2024. 2 +[24] Xiao Li, Liquan Chen, Ju Jia, Zhongyuan Qin, and Zhangjie Fu. A lightweight image forgery prevention scheme for IoT using gan-based steganography. IEEE Transactions on Industrial Informatics, 2024. 2 +[25] Dongdong Lin, Benedetta Tondi, Bin Li, and Mauro Barni. A cyclegan watermarking method for ownership verification. IEEE Transactions on Dependable and Secure Computing, 2024. 2 +[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5, 6 +[27] Yang Liu, Zhen Zhu, and Xiang Bai. Wdnet: Watermark-decomposition network for visible watermark removal. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 3685-3693, 2021. 2 + +[28] Xiyang Luo, Ruohan Zhan, Huiwen Chang, Feng Yang, and Peyman Milanfar. Distortion agnostic deep watermarking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13548-13557, 2020. 2 +[29] Thanh Thi Nguyen, Quoc Viet Hung Nguyen, Dung Tien Nguyen, Duc Thanh Nguyen, Thien Huynh-The, Saeid Nahavandi, Thanh Tam Nguyen, Quoc-Viet Pham, and Cuong M Nguyen. Deep learning for deepfakes creation and detection: A survey. Computer Vision and Image Understanding, 223:103525, 2022. 1 +[30] Guangyu Nie, Changhoon Kim, Yezhou Yang, and Yi Ren. Attributing image generative models using latent fingerprints. In International Conference on Machine Learning, pages 26150-26165. PMLR, 2023. 2 +[31] Li Niu, Xing Zhao, Bo Zhang, and Liqing Zhang. Fine-grained visible watermark removal. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12770-12779, 2023. 2 +[32] Konstantin A Pantserev. The malicious use of ai-based deepfake technology as the new threat to psychological security and political stability. Cyber defence in the age of AI, smart societies and augmented humanity, pages 37-55, 2020. 1 +[33] Ram Shankar Pathak. The wavelet transform. Springer Science & Business Media, 2009. 1, 2, 4 +[34] Fernandez Pierre, Alexandre Sablayrolles, Teddy Furon, Hervé Jégou, and Matthijs Douze. Watermarking images in self-supervised latent spaces. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3054-3058. IEEE, 2022. 2 +[35] Fernandez Pierre, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22466-22477, 2023. 1, 6, 7 +[36] Tong Qiao, Yuyan Ma, Ning Zheng, Hanzhou Wu, Yanli Chen, Ming Xu, and Xiangyang Luo. A novel model watermarking for protecting generative adversarial network. Computers & Security, 127:103102, 2023. 2 +[37] Sunpreet Sharma, Ju Jia Zou, Gu Fang, Pancham Shukla, and Weidong Cai. A review of image watermarking for identity protection and verification. Multimedia Tools and Applications, 83(11):31829-31891, 2024. 2 +[38] Qingtang Su, Huanying Wang, DeCheng Liu, Zihan Yuan, and Xueting Zhang. A combined domain watermarking algorithm of color image. Multimedia Tools and Applications, 79(39):30023-30043, 2020. 2 +[39] Qingtang Su, Xueting Zhang, and Huanying Wang. A blind color image watermarking algorithm combined spatial domain and svd. International Journal of Intelligent Systems, 37(8):4747-4771, 2022. 2 +[40] Ruizhou Sun, Yukun Su, and Qingyao Wu. Denet: disentangled embedding network for visible watermark removal. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2411-2419, 2023. 2 +[41] Matthew Tancik, Ben Mildenhall, and Ren Ng. Stegastamp: Invisible hyperlinks in physical photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2117-2126, 2020. 1, 6, 7 + +[42] D Vaishnavi and TS Subashini. Robust and invisible image watermarking in rgb color space using svd. Procedia Computer Science, 46:1770-1777, 2015. 2 +[43] Guanjie Wang, Zehua Ma, Chang Liu, Xi Yang, Han Fang, Weiming Zhang, and Nenghai Yu. Must: Robust image watermarking for multi-source tracing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5364-5371, 2024. 6 +[44] Huanying Wang and Qingtang Su. A color image watermarking method combined qr decomposition and spatial domain. Multimedia Tools and Applications, 81(26):37895-37916, 2022. 2 +[45] Zilan Wang, Junfeng Guo, Jiacheng Zhu, Yiming Li, Heng Huang, Muhao Chen, and Zhengzhong Tu. Sleepermark: Towards robust watermark against fine-tuning text-to-image diffusion models. arXiv preprint arXiv:2412.04852, 2024. 6 +[46] Zijie J Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. Diffusional: A large-scale prompt gallery dataset for text-to-image generative models. arXiv preprint arXiv:2210.14896, 2022.5 +[47] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-rings watermarks: Invisible fingerprints for diffusion images. In Advances in Neural Information Processing Systems, pages 58047-58063. Curran Associates, Inc., 2023. 1, 6, 7 +[48] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-rings watermarks: Invisible fingerprints for diffusion images. Advances in Neural Information Processing Systems, 36, 2024. 1 +[49] Zijin Yang, Kai Zeng, Kejiang Chen, Han Fang, Weiming Zhang, and Nenghai Yu. Gaussian shading: Provable performance-lossless image watermarking for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12162-12171, 2024. 1, 6 +[50] Zihan Yuan, Qingtang Su, Decheng Liu, and Xueting Zhang. A blind image watermarking scheme combining spatial domain and frequency domain. The visual computer, 37:1867-1881, 2021. 2 +[51] Chaoning Zhang, Philipp Benz, Adil Karjauv, Geng Sun, and In So Kweon. Udh: Universal deep hiding for steganography, watermarking, and light field messaging. Advances in Neural Information Processing Systems, 33:10223-10234, 2020. 2 +[52] Xuanyu Zhang, Runyi Li, Jiwen Yu, Youmin Xu, Weiqi Li, and Jian Zhang. Editguard: Versatile image watermarking for tamper localization and copyright protection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11964-11974, 2024. 6 +[53] Jiren Zhu, Russell Kaplan, Justin Johnson, and Li Fei-Fei. Hidden: Hiding data with deep networks. European Conference on Computer Vision (ECCV), 2018. 1, 2, 6 \ No newline at end of file diff --git a/2025/SpecGuard_ Spectral Projection-based Advanced Invisible Watermarking/images.zip b/2025/SpecGuard_ Spectral Projection-based Advanced Invisible Watermarking/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ccd65454a058c1ee4a6e3784a2fe9545ea41344d --- /dev/null +++ b/2025/SpecGuard_ Spectral Projection-based Advanced Invisible Watermarking/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a0828846bc1fad3ac023426eee74c0c75cac241a0ad6f6af539b95b3c5d5a4ec +size 805514 diff --git a/2025/SpecGuard_ Spectral Projection-based Advanced Invisible Watermarking/layout.json b/2025/SpecGuard_ Spectral Projection-based Advanced Invisible Watermarking/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..66729d648a809bf5f35bf5ad2efcddf6f92c7fea --- /dev/null +++ b/2025/SpecGuard_ Spectral Projection-based Advanced Invisible Watermarking/layout.json @@ -0,0 +1,11362 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 77, + 103, + 533, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 103, + 533, + 121 + ], + "spans": [ + { + "bbox": [ + 77, + 103, + 533, + 121 + ], + "type": "text", + "content": "SpecGuard: Spectral Projection-based Advanced Invisible Watermarking" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 143, + 475, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 143, + 475, + 172 + ], + "spans": [ + { + "bbox": [ + 135, + 143, + 475, + 172 + ], + "type": "text", + "content": "Inzamamul Alam, Md Tanvir Islam, Simon S. Woo*, Khan Muhammad Sungkyunkwan University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 168, + 174, + 440, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 174, + 440, + 185 + ], + "spans": [ + { + "bbox": [ + 168, + 174, + 440, + 185 + ], + "type": "text", + "content": "{inzil5, tanvirnwu, swoo, khanmuhammad}@g.skku.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 239, + 297, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 297, + 539 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 297, + 539 + ], + "type": "text", + "content": "Watermarking embeds imperceptible patterns into images for authenticity verification. However, existing methods often lack robustness against various transformations primarily including distortions, image regeneration, and adversarial perturbation, creating real-world challenges. In this work, we introduce SpecGuard, a novel watermarking approach for robust and invisible image watermarking. Unlike prior approaches, we embed the message inside hidden convolution layers by converting from the spatial domain to the frequency domain using spectral projection of a higher frequency band that is decomposed by wavelet projection. Spectral projection employs Fast Fourier Transform approximation to transform spatial data into the frequency domain efficiently. In the encoding phase, a strength factor enhances resilience against diverse attacks, including adversarial, geometric, and regeneration-based distortions, ensuring the preservation of copyrighted information. Meanwhile, the decoder leverages Parseval's theorem to effectively learn and extract the watermark pattern, enabling accurate retrieval under challenging transformations. We evaluate the proposed SpecGuard based on the embedded watermark's invisibility, capacity, and robustness. Comprehensive experiments demonstrate the proposed SpecGuard outperforms the state-of-the-art models. To ensure reproducibility, the full code is released on GitHub." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 552, + 135, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 135, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 135, + 563 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 572, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 572, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 572, + 296, + 693 + ], + "type": "text", + "content": "With the rapid advancement of digital media and artificial intelligence, concerns regarding image authenticity, copyright protection, and content integrity have become more challenging than ever [9, 16, 32]. Moreover, the widespread availability of the latest image manipulation tools [4, 5, 14] enables malicious tamperers to easily forge and redistribute digital content without authorization, posing a significant threat to ownership verification [29]. This growing risk emphasizes the need for reliable techniques for secure authentication and detection of unauthorized manipulation." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 213, + 555, + 338 + ], + "blocks": [ + { + "bbox": [ + 316, + 213, + 555, + 338 + ], + "lines": [ + { + "bbox": [ + 316, + 213, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 213, + 555, + 338 + ], + "type": "image", + "image_path": "a38e60f474c97ee8790f581a20a8d3b9e9e3db28ed3cebb62e1971ff215af1a0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 319, + 345, + 549, + 357 + ], + "lines": [ + { + "bbox": [ + 319, + 345, + 549, + 357 + ], + "spans": [ + { + "bbox": [ + 319, + 345, + 549, + 357 + ], + "type": "text", + "content": "Figure 1. Image authentication using our proposed SpecGuard." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 366, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 555, + 594 + ], + "type": "text", + "content": "Recently, invisible watermarking has gained significant attention as a prominent defense mechanism for media authentication by embedding invisible messages into images to verify authenticity [35, 48]. In fact, invisible watermarks are preferred for preserving image quality and resisting tampering. These watermarks are unique to the creator and enable tamper verification by comparing the retrieved watermark to the original, as the high-level process is presented in Fig. 1. Traditional watermarking methods often rely on transformation techniques [18, 33]. Deep learning approaches like StegaStamp [41], Stable Signature [35], and HiDDeN [53] provide end-to-end solutions for message embedding. However, these methods often struggle with fragility in handling common image processing operations such as resizing, cropping, compression, and noise addition, which can distort or erase the embedded watermark. Additionally, the performance of watermark embedding and extraction often remains vulnerable to attacks with noise injection, blurring, contrasting, and rotation [8]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "To address the aforementioned challenges, we introduce a novel robust, and invisible image watermarking method named SpecGuard. SpecGuard is designed to overcome the fundamental trade-offs [8] between imperceptibility, and robustness. Our proposed SpecGuard strategically embeds watermark information in the spectral domain, leveraging wavelet-based decomposition to distribute the watermark across high-frequency components. Unlike traditional frequency domain watermarking techniques [47, 49] that are easily disrupted by common image manipulations, Spec-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 260, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 260, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 260, + 713 + ], + "type": "text", + "content": "*Corresponding author: swoo@g.skku.edu (Simon S. Woo)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17984" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "Guard maintains imperceptibility while significantly improving robustness against a wide range of transformations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 99, + 296, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 99, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 55, + 99, + 296, + 173 + ], + "type": "text", + "content": "Overall, our proposed SpecGuard addresses the current limitations of the previous watermarking methods by providing a robust, imperceptible watermarking technique that maintains integrity under diverse manipulations, significantly enhancing digital content security and authenticity verification. Our key contributions are as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 179, + 296, + 347 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 179, + 295, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 179, + 295, + 238 + ], + "spans": [ + { + "bbox": [ + 55, + 179, + 295, + 238 + ], + "type": "text", + "content": "- We introduce a novel watermarking approach that embeds message bits in high-frequency spectral components via wavelet and spectral projection inside hidden convolutional layers, ensuring robustness against various transformations and adversarial attacks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 239, + 296, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 299 + ], + "type": "text", + "content": "- We adapt Parseval's theorem [20] as a learnable threshold to optimize SpecGuard and spectral masking for robust watermark bit recovery under diverse transformations including distortions, regeneration, and adversarial attacks, proven through the experimental results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 300, + 296, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 347 + ], + "type": "text", + "content": "- Our extensive evaluations demonstrate SpecGuard's superior bit embedding capacity and producing better invisible watermarked images, surpassing the performance of state-of-the-art (SOTA) methods." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 367, + 147, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 147, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 147, + 380 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 391, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 296, + 714 + ], + "type": "text", + "content": "Watermarking an image has been a widely researched topic for securing the ownership and verifying authenticity of digital content [37]. Traditional watermarking techniques typically embed invisible [42] or visible [7] watermarks into images, which can later be extracted or detected to verify the content's originality. These methods can be broadly classified into spatial-domain [39, 44] and frequency-domain [13] watermarking, while some are based on combined methods [38, 50]. However, researchers recently proposed many advanced models [23, 27, 31, 40] for effective watermark removal. To face this growing challenge, researchers introduced different methods [2, 3, 15, 22, 28, 51, 53] as alternatives to deep learning-based encoders or decoders to produce more robust image watermarking. Furthermore, iterative models have demonstrated competitive performance [21, 34], particularly in robustness against a wide range of transformations. In addition, with the rise of generative methods, researchers used the watermark-labeled data for training to learn how to produce watermarks [11, 24]. Also, models that combine generative methods with watermarking techniques show promise in effective image watermarking [25, 30, 36]. However, such approaches face limitations such as increased computational complexity and longer processing times. These approaches are also more vulnerable to adversarial attacks that can target and distort the embedded watermark without altering the content visibly." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 71, + 486, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 486, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 486, + 85 + ], + "type": "text", + "content": "3. Proposed Method: SpecGuard" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 91, + 555, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 140 + ], + "type": "text", + "content": "We introduce SpecGuard, as illustrated in Fig. 2, which involves two fundamental modules: an \"Encoder\" for embedding the watermark and a \"Decoder\" for accurately extracting the watermark detailed in the following sections." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 146, + 376, + 158 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 146, + 376, + 158 + ], + "spans": [ + { + "bbox": [ + 314, + 146, + 376, + 158 + ], + "type": "text", + "content": "3.1. Encoder" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 163, + 554, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 554, + 235 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 554, + 235 + ], + "type": "text", + "content": "By targeting high-frequency components, the encoder integrates a binary message " + }, + { + "bbox": [ + 313, + 163, + 554, + 235 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 163, + 554, + 235 + ], + "type": "text", + "content": " into the cover image " + }, + { + "bbox": [ + 313, + 163, + 554, + 235 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 163, + 554, + 235 + ], + "type": "text", + "content": ". Using wavelet projection (WP) [33] and a Fast Fourier Transform (FFT)-based spectral projection (SP) [18] approximation, the message " + }, + { + "bbox": [ + 313, + 163, + 554, + 235 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 163, + 554, + 235 + ], + "type": "text", + "content": " is inserted into specific frequency bands, minimizing perceptual impact." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 236, + 554, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 554, + 271 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 554, + 271 + ], + "type": "text", + "content": "Wavelet Projection. We use a wavelet projection to capture frequency and spatial localization features that describe an image across different scales, as shown in Eq. (1):" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 350, + 277, + 555, + 303 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 277, + 555, + 303 + ], + "spans": [ + { + "bbox": [ + 350, + 277, + 555, + 303 + ], + "type": "interline_equation", + "content": "W (a, b) = \\frac {1}{\\sqrt {| a |}} \\int_ {- \\infty} ^ {\\infty} f (x) \\psi \\left(\\frac {x - b}{a}\\right) d x, \\tag {1}", + "image_path": "d34f82ab3506328654af1d8aeb8503e7c1bad490219dd92d08f6c3662aaea307.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 308, + 554, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 554, + 355 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 554, + 355 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 308, + 554, + 355 + ], + "type": "inline_equation", + "content": "a \\in \\mathbb{R} \\setminus \\{0\\}" + }, + { + "bbox": [ + 313, + 308, + 554, + 355 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 308, + 554, + 355 + ], + "type": "inline_equation", + "content": "b \\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 308, + 554, + 355 + ], + "type": "text", + "content": " denote the scaling and translation parameters, respectively. Here, " + }, + { + "bbox": [ + 313, + 308, + 554, + 355 + ], + "type": "inline_equation", + "content": "\\psi_{a,b}(x)" + }, + { + "bbox": [ + 313, + 308, + 554, + 355 + ], + "type": "text", + "content": " represents a rescaled and translated form of the mother wavelet " + }, + { + "bbox": [ + 313, + 308, + 554, + 355 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 313, + 308, + 554, + 355 + ], + "type": "text", + "content": ", defined as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 374, + 360, + 554, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 360, + 554, + 386 + ], + "spans": [ + { + "bbox": [ + 374, + 360, + 554, + 386 + ], + "type": "interline_equation", + "content": "\\psi_ {a, b} (x) = \\psi \\left(\\frac {x - b}{a}\\right) \\cdot \\frac {1}{\\sqrt {| a |}}, \\tag {2}", + "image_path": "7e45182c6883fecf5c64f17f71ce396cfeb398fd1c3f9441528e2108d1a4f3c9.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 392, + 554, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 554, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 554, + 492 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 392, + 554, + 492 + ], + "type": "inline_equation", + "content": "\\frac{1}{\\sqrt{|a|}}" + }, + { + "bbox": [ + 313, + 392, + 554, + 492 + ], + "type": "text", + "content": " functions as a normalization factor, guaranteeing that the energy of the wavelet is invariant to the scaling parameter " + }, + { + "bbox": [ + 313, + 392, + 554, + 492 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 392, + 554, + 492 + ], + "type": "text", + "content": ". Minimal values of " + }, + { + "bbox": [ + 313, + 392, + 554, + 492 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 392, + 554, + 492 + ], + "type": "text", + "content": " compress the wavelet, enabling the inspection of high-frequency components, whereas greater values of " + }, + { + "bbox": [ + 313, + 392, + 554, + 492 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 392, + 554, + 492 + ], + "type": "text", + "content": " elongate the wavelet, promoting low-frequency analysis. Since each mother wavelet " + }, + { + "bbox": [ + 313, + 392, + 554, + 492 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 313, + 392, + 554, + 492 + ], + "type": "text", + "content": " is built with zero mean and finite energy [12], it guarantees to maintain stability as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 350, + 498, + 553, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 498, + 553, + 522 + ], + "spans": [ + { + "bbox": [ + 350, + 498, + 553, + 522 + ], + "type": "interline_equation", + "content": "\\int_ {- \\infty} ^ {\\infty} \\psi (x) d x = 0, \\quad \\int_ {- \\infty} ^ {\\infty} | \\psi (x) | ^ {2} d x < \\infty , \\tag {3}", + "image_path": "e78e6a1fadf5c3ddae39f07fd6c99ddb4e0e4452f3d914c86fd0ad72064026a2.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 528, + 554, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 554, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 554, + 576 + ], + "type": "text", + "content": "where the wavelet projection from Eq. (1) decomposes the input into orthogonal wavelet sets using discrete scales and translations. For " + }, + { + "bbox": [ + 313, + 528, + 554, + 576 + ], + "type": "inline_equation", + "content": "2D" + }, + { + "bbox": [ + 313, + 528, + 554, + 576 + ], + "type": "text", + "content": " inputs, the scaled and translated basis elements [1] are defined for each coordinate pair " + }, + { + "bbox": [ + 313, + 528, + 554, + 576 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 313, + 528, + 554, + 576 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 582, + 559, + 595 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 582, + 559, + 595 + ], + "spans": [ + { + "bbox": [ + 314, + 582, + 559, + 595 + ], + "type": "interline_equation", + "content": "\\mathbf {S} _ {L L} = \\phi (u, v) = \\phi (u) \\phi (v), \\qquad \\mathbf {S} _ {L H} = \\psi_ {H} (u, v) = \\psi (u) \\phi (v),", + "image_path": "df30bed085da371aac41454a9f44cc9967b65b9b7f438618211b7ed8cfc59c9f.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 597, + 559, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 597, + 559, + 617 + ], + "spans": [ + { + "bbox": [ + 315, + 597, + 559, + 617 + ], + "type": "interline_equation", + "content": "\\mathbf {S} _ {H L} = \\psi_ {V} (u, v) = \\phi (u) \\psi (v), \\quad \\mathbf {S} _ {H H} = \\psi_ {D} (u, v) = \\psi (u) \\psi (v), \\tag {4}", + "image_path": "9e92a0438b2d3c00b53796680bbf99aabb039d5e14c4ed802822df25175cfce3.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 619, + 555, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 619, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 619, + 555, + 666 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 313, + 619, + 555, + 666 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 619, + 555, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 619, + 555, + 666 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 619, + 555, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 619, + 555, + 666 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 619, + 555, + 666 + ], + "type": "text", + "content": " represent the horizontal, vertical, and diagonal decomposition direction, respectively. To depict the image at different resolutions, we define scaling and wavelet functions at scale " + }, + { + "bbox": [ + 313, + 619, + 555, + 666 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 619, + 555, + 666 + ], + "type": "text", + "content": " as shown below:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 345, + 672, + 553, + 700 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 672, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 345, + 672, + 553, + 700 + ], + "type": "interline_equation", + "content": "\\phi_ {j, m, n} (u, v) = 2 ^ {j / 2} \\phi \\left(u - \\frac {m}{2 ^ {j}}, v - \\frac {n}{2 ^ {j}}\\right), \\tag {5}", + "image_path": "c5a621d5ee7cd840e9acb0b53e6e2d2421a18792a6fd2701c603f80ed5f1b4e9.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 347, + 696, + 522, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 696, + 522, + 716 + ], + "spans": [ + { + "bbox": [ + 347, + 696, + 522, + 716 + ], + "type": "interline_equation", + "content": "\\psi_ {j, m, n} ^ {d} (u, v) = 2 ^ {j / 2} \\psi^ {d} \\left(u - \\frac {m}{2 ^ {j}}, v - \\frac {n}{2 ^ {j}}\\right),", + "image_path": "e0a3f12e0210c5425b89d2f76c930b137c8182adc04e4aa58697433becaf34a4.jpg" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17985" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 69, + 553, + 240 + ], + "blocks": [ + { + "bbox": [ + 59, + 69, + 553, + 240 + ], + "lines": [ + { + "bbox": [ + 59, + 69, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 59, + 69, + 553, + 240 + ], + "type": "image", + "image_path": "d547e59ba83c2de38ccd65c824e949168c58fb8718a8d8f0e3fe187603ffd931.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 246, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 54, + 246, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 54, + 246, + 555, + 270 + ], + "type": "text", + "content": "Figure 2. Architecture of the proposed SpecGuard watermarking method involves encoding a binary message " + }, + { + "bbox": [ + 54, + 246, + 555, + 270 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 54, + 246, + 555, + 270 + ], + "type": "text", + "content": " into the high-frequency band of the cover image " + }, + { + "bbox": [ + 54, + 246, + 555, + 270 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 54, + 246, + 555, + 270 + ], + "type": "text", + "content": " using wavelet and spectral projection and learning to decode the embedded message." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 281, + 296, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 296, + 377 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 296, + 377 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 281, + 296, + 377 + ], + "type": "inline_equation", + "content": "d \\in \\{H, V, D\\}" + }, + { + "bbox": [ + 55, + 281, + 296, + 377 + ], + "type": "text", + "content": " is the wavelet function direction that serves as discrete basis elements for multi-resolution analysis, capturing details across frequency bands and spatial locations. In Eq. (6), " + }, + { + "bbox": [ + 55, + 281, + 296, + 377 + ], + "type": "inline_equation", + "content": "T_{m,n}" + }, + { + "bbox": [ + 55, + 281, + 296, + 377 + ], + "type": "text", + "content": " denotes the intensity or pixel value of the cover image " + }, + { + "bbox": [ + 55, + 281, + 296, + 377 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 281, + 296, + 377 + ], + "type": "text", + "content": " at spatial coordinates " + }, + { + "bbox": [ + 55, + 281, + 296, + 377 + ], + "type": "inline_equation", + "content": "(m, n)" + }, + { + "bbox": [ + 55, + 281, + 296, + 377 + ], + "type": "text", + "content": ". The discrete scaling function " + }, + { + "bbox": [ + 55, + 281, + 296, + 377 + ], + "type": "inline_equation", + "content": "W_{\\phi}(j, u, v)" + }, + { + "bbox": [ + 55, + 281, + 296, + 377 + ], + "type": "text", + "content": " (approximation at scale " + }, + { + "bbox": [ + 55, + 281, + 296, + 377 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 281, + 296, + 377 + ], + "type": "text", + "content": ") and the detail coefficients " + }, + { + "bbox": [ + 55, + 281, + 296, + 377 + ], + "type": "inline_equation", + "content": "W_{\\psi}^{d}(j, u, v)" + }, + { + "bbox": [ + 55, + 281, + 296, + 377 + ], + "type": "text", + "content": " for each direction are computed accordingly as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 384, + 291, + 413 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 291, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 291, + 413 + ], + "type": "interline_equation", + "content": "W _ {\\phi} (j, u, v) = \\frac {1}{l} \\sum_ {m = 0} ^ {l - 1} \\sum_ {n = 0} ^ {l - 1} T _ {m, n} \\phi \\left(m - u \\cdot 2 ^ {- j}, n - v \\cdot 2 ^ {- j}\\right),", + "image_path": "d5e21ba4df08262e4e37b0ae15fdc9715258c6a4761b6e9986d0c1b5e7a47fa3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 415, + 296, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 453 + ], + "type": "interline_equation", + "content": "W _ {\\psi} ^ {d} (j, u, v) = \\frac {1}{l} \\sum_ {m = 0} ^ {l - 1} \\sum_ {n = 0} ^ {l - 1} T _ {m, n} \\psi^ {d} \\left(m - u \\cdot 2 ^ {- j}, n - v \\cdot 2 ^ {- j}\\right), \\tag {6}", + "image_path": "a27141e64a61f40615e79f75f7b45f2b44491f2d0cf1bb9a7432e3792d071307.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 454, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 296, + 491 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 55, + 454, + 296, + 491 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 454, + 296, + 491 + ], + "type": "text", + "content": " as the discrete region dimension, these coefficients capture multi-scale, multi-orientation image information, forming the basis of spectral features as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 86, + 496, + 295, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 496, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 86, + 496, + 295, + 521 + ], + "type": "interline_equation", + "content": "\\beta_ {j} = \\bigcup_ {d \\in \\{H, V, D \\}} \\left(W _ {\\phi} (j, u, v) \\cup W _ {\\psi} ^ {d} (j, u, v)\\right). \\tag {7}", + "image_path": "4bdd29a935edf767a61feb425be13b640965ab0c2c4eb4b31f73e491216600f3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 528, + 295, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 295, + 563 + ], + "type": "text", + "content": "This feature set " + }, + { + "bbox": [ + 55, + 528, + 295, + 563 + ], + "type": "inline_equation", + "content": "\\beta_{j}" + }, + { + "bbox": [ + 55, + 528, + 295, + 563 + ], + "type": "text", + "content": " captures key frequency and spatial details across resolutions, forming the foundation for the watermark embedding process of our SpecGuard." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 563, + 295, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 295, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 295, + 611 + ], + "type": "text", + "content": "Selective Frequency Band Decomposition. To refine the embedding process, we segment the data into distinct frequency bands. The decomposition level " + }, + { + "bbox": [ + 55, + 563, + 295, + 611 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 55, + 563, + 295, + 611 + ], + "type": "text", + "content": " is determined by the image complexity, calculated as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 618, + 295, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 618, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 618, + 295, + 632 + ], + "type": "interline_equation", + "content": "\\kappa = \\left\\lfloor \\sqrt {\\log (1 + N)} \\right\\rfloor , \\tag {8}", + "image_path": "57adc211274c41bb4e5f618fca6a851ff5afa7a001c33d9beb2e709315ee440c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 639, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 296, + 685 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 639, + 296, + 685 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 639, + 296, + 685 + ], + "type": "text", + "content": " denotes the total pixel count in the cover image " + }, + { + "bbox": [ + 55, + 639, + 296, + 685 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 639, + 296, + 685 + ], + "type": "text", + "content": ". And, each component " + }, + { + "bbox": [ + 55, + 639, + 296, + 685 + ], + "type": "inline_equation", + "content": "\\beta_{j}" + }, + { + "bbox": [ + 55, + 639, + 296, + 685 + ], + "type": "text", + "content": " falls within a unique frequency band, yielding a total of " + }, + { + "bbox": [ + 55, + 639, + 296, + 685 + ], + "type": "inline_equation", + "content": "1 + 3\\kappa" + }, + { + "bbox": [ + 55, + 639, + 296, + 685 + ], + "type": "text", + "content": " distinct frequency bands as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 692, + 295, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 692, + 295, + 716 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 295, + 716 + ], + "type": "interline_equation", + "content": "\\beta_ {j} = \\phi_ {j} (u, v) \\cup \\bigcup_ {d \\in \\{H, V, D \\}} \\psi_ {j} ^ {d} (u, v). \\tag {9}", + "image_path": "d43f807fbdb4bf066ac15ee44fe485b41442492e1735f1ef5ab925ff25bc86c5.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 281, + 555, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 353 + ], + "type": "text", + "content": "The components " + }, + { + "bbox": [ + 313, + 281, + 555, + 353 + ], + "type": "inline_equation", + "content": "\\beta_{j}" + }, + { + "bbox": [ + 313, + 281, + 555, + 353 + ], + "type": "text", + "content": ", consisting of scaling functions " + }, + { + "bbox": [ + 313, + 281, + 555, + 353 + ], + "type": "inline_equation", + "content": "\\phi_j(u,v)" + }, + { + "bbox": [ + 313, + 281, + 555, + 353 + ], + "type": "text", + "content": " and wavelet functions " + }, + { + "bbox": [ + 313, + 281, + 555, + 353 + ], + "type": "inline_equation", + "content": "\\psi_j^d (u,v)" + }, + { + "bbox": [ + 313, + 281, + 555, + 353 + ], + "type": "text", + "content": ", capture specific spatial frequency bands, enabling targeted high-frequency embedding. We translate the WP into disjoint intervals representing a unique frequency range to approximate the segmentation in the frequency domain:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 330, + 360, + 555, + 385 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 360, + 555, + 385 + ], + "spans": [ + { + "bbox": [ + 330, + 360, + 555, + 385 + ], + "type": "interline_equation", + "content": "\\beta_ {j} = \\left\\{W _ {\\psi} ^ {d} (u, v) \\mid u, v \\in \\left(\\frac {j \\cdot L}{\\kappa}, \\frac {(j + 1) \\cdot L}{\\kappa}\\right) \\right\\}, \\tag {10}", + "image_path": "d073fe2084ed4b7f71ffb23cdeefc865b8eebc6cdfdef0b296c7f6685e01d821.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 392, + 554, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 554, + 453 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 313, + 392, + 554, + 453 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 392, + 554, + 453 + ], + "type": "text", + "content": " is the dimension of " + }, + { + "bbox": [ + 313, + 392, + 554, + 453 + ], + "type": "inline_equation", + "content": "S_{HH}" + }, + { + "bbox": [ + 313, + 392, + 554, + 453 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 392, + 554, + 453 + ], + "type": "inline_equation", + "content": "W_{\\psi}^{d}(u,v)" + }, + { + "bbox": [ + 313, + 392, + 554, + 453 + ], + "type": "text", + "content": " represents wavelet values within segmented intervals. This frequency band partitioning mimics the frequency selectivity of wavelet sub-bands, enabling effective targeting of high-frequency regions for optimal embedding." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 453, + 554, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 554, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 554, + 513 + ], + "type": "text", + "content": "Approximation of Spectral Projection. We first apply spectral projection on the " + }, + { + "bbox": [ + 313, + 453, + 554, + 513 + ], + "type": "inline_equation", + "content": "S_{HH}" + }, + { + "bbox": [ + 313, + 453, + 554, + 513 + ], + "type": "text", + "content": " sub-band, transforming it into the spectral domain. Given a matrix " + }, + { + "bbox": [ + 313, + 453, + 554, + 513 + ], + "type": "inline_equation", + "content": "T(x,y)" + }, + { + "bbox": [ + 313, + 453, + 554, + 513 + ], + "type": "text", + "content": " representing pixel intensities in " + }, + { + "bbox": [ + 313, + 453, + 554, + 513 + ], + "type": "inline_equation", + "content": "S_{HH}" + }, + { + "bbox": [ + 313, + 453, + 554, + 513 + ], + "type": "text", + "content": ", the spectral projection computes the spectral components " + }, + { + "bbox": [ + 313, + 453, + 554, + 513 + ], + "type": "inline_equation", + "content": "\\zeta(u,v)" + }, + { + "bbox": [ + 313, + 453, + 554, + 513 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 318, + 521, + 553, + 557 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 521, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 318, + 521, + 553, + 557 + ], + "type": "interline_equation", + "content": "\\zeta (u, v) = \\frac {1}{L ^ {2}} \\sum_ {x} \\sum_ {y} T (x, y) \\cdot \\exp \\left(- i \\frac {2 \\pi}{L} (x \\cdot u + y \\cdot v)\\right), \\tag {11}", + "image_path": "2a05e491d803fd9707677d4a5cc30f25c03e31569f12b5fb3f0343f753a65e30.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 558, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 558, + 554, + 605 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 558, + 554, + 605 + ], + "type": "text", + "content": " denotes the dimension of " + }, + { + "bbox": [ + 313, + 558, + 554, + 605 + ], + "type": "inline_equation", + "content": "S_{HH}" + }, + { + "bbox": [ + 313, + 558, + 554, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 558, + 554, + 605 + ], + "type": "inline_equation", + "content": "T(x,y)" + }, + { + "bbox": [ + 313, + 558, + 554, + 605 + ], + "type": "text", + "content": " provides the intensity at each coordinate " + }, + { + "bbox": [ + 313, + 558, + 554, + 605 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 313, + 558, + 554, + 605 + ], + "type": "text", + "content": " which is equivalent to " + }, + { + "bbox": [ + 313, + 558, + 554, + 605 + ], + "type": "inline_equation", + "content": "W_{\\psi}^{d}(u,v)" + }, + { + "bbox": [ + 313, + 558, + 554, + 605 + ], + "type": "text", + "content": " in Eq. (6), " + }, + { + "bbox": [ + 313, + 558, + 554, + 605 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 558, + 554, + 605 + ], + "type": "text", + "content": " is the imaginary unit, and " + }, + { + "bbox": [ + 313, + 558, + 554, + 605 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 313, + 558, + 554, + 605 + ], + "type": "text", + "content": " are the spectral coordinates." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 606, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 555, + 714 + ], + "type": "text", + "content": "To approximate the spectral components using the FFT, we create a symmetrically extended version " + }, + { + "bbox": [ + 313, + 606, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{T}(x,y)" + }, + { + "bbox": [ + 313, + 606, + 555, + 714 + ], + "type": "text", + "content": " of the original " + }, + { + "bbox": [ + 313, + 606, + 555, + 714 + ], + "type": "inline_equation", + "content": "N\\times N" + }, + { + "bbox": [ + 313, + 606, + 555, + 714 + ], + "type": "text", + "content": " matrix " + }, + { + "bbox": [ + 313, + 606, + 555, + 714 + ], + "type": "inline_equation", + "content": "T(x,y)" + }, + { + "bbox": [ + 313, + 606, + 555, + 714 + ], + "type": "text", + "content": ". This extension is achieved by mirroring " + }, + { + "bbox": [ + 313, + 606, + 555, + 714 + ], + "type": "inline_equation", + "content": "T(x,y)" + }, + { + "bbox": [ + 313, + 606, + 555, + 714 + ], + "type": "text", + "content": " along its boundaries, doubling its size to " + }, + { + "bbox": [ + 313, + 606, + 555, + 714 + ], + "type": "inline_equation", + "content": "2N\\times 2N" + }, + { + "bbox": [ + 313, + 606, + 555, + 714 + ], + "type": "text", + "content": ". Specifically, the original matrix occupies the top-left quadrant, with the remaining quadrants filled by reflecting " + }, + { + "bbox": [ + 313, + 606, + 555, + 714 + ], + "type": "inline_equation", + "content": "T(x,y)" + }, + { + "bbox": [ + 313, + 606, + 555, + 714 + ], + "type": "text", + "content": " horizontally, vertically, and diagonally, respectively. This symmetric structure ensures that the FFT yields only real values, allowing the spectral coefficients to" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17986" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 97 + ], + "type": "text", + "content": "be extracted directly from the real part of the FFT operation. Then, we apply the 2D FFT to " + }, + { + "bbox": [ + 55, + 72, + 294, + 97 + ], + "type": "inline_equation", + "content": "\\tilde{T} (x,y)" + }, + { + "bbox": [ + 55, + 72, + 294, + 97 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 103, + 298, + 140 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 103, + 298, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 103, + 298, + 140 + ], + "type": "interline_equation", + "content": "F (u, v) = \\frac {1}{(2 N) ^ {2}} \\sum_ {x} \\sum_ {y} \\tilde {T} (x, y) \\cdot \\exp \\left(- i \\frac {2 \\pi}{2 N} (x \\cdot u + y \\cdot v)\\right) \\tag {12}", + "image_path": "d1bc9ed75128393a71f21595d116f4b84866c68d18fa5aa9f3b1f6c85f074537.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 141, + 296, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 141, + 296, + 165 + ], + "spans": [ + { + "bbox": [ + 55, + 141, + 296, + 165 + ], + "type": "text", + "content": "The SP coefficients are then approximated by taking the real part (Re) of " + }, + { + "bbox": [ + 55, + 141, + 296, + 165 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 141, + 296, + 165 + ], + "type": "text", + "content": " in the original " + }, + { + "bbox": [ + 55, + 141, + 296, + 165 + ], + "type": "inline_equation", + "content": "N \\times N" + }, + { + "bbox": [ + 55, + 141, + 296, + 165 + ], + "type": "text", + "content": " region as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 97, + 176, + 295, + 189 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 176, + 295, + 189 + ], + "spans": [ + { + "bbox": [ + 97, + 176, + 295, + 189 + ], + "type": "interline_equation", + "content": "\\zeta (u, v) \\approx \\operatorname {R e} (F (u, v)), \\quad 0 \\leq u, v < N. \\tag {13}", + "image_path": "febe9345d7cbf5c682f5cbd858ce1018ea7f63ed6a33ab85ff265aece767df78.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 191, + 296, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 191, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 296, + 251 + ], + "type": "text", + "content": "Applying Eq. (13) to the sub-bands extracted from wavelet projection in Eq. (6), we achieve a computationally efficient spectral projection by leveraging the FFT approximation on a symmetrically extended matrix, maintaining effective embedding properties within the spectral domain." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 251, + 296, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 371 + ], + "type": "text", + "content": "SpecGuard Embedding Process. The embedding process integrates the binary message " + }, + { + "bbox": [ + 55, + 251, + 296, + 371 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 251, + 296, + 371 + ], + "type": "text", + "content": " into the high-frequency band " + }, + { + "bbox": [ + 55, + 251, + 296, + 371 + ], + "type": "inline_equation", + "content": "S_{HH}" + }, + { + "bbox": [ + 55, + 251, + 296, + 371 + ], + "type": "text", + "content": " of the cover image " + }, + { + "bbox": [ + 55, + 251, + 296, + 371 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 251, + 296, + 371 + ], + "type": "text", + "content": ", enhancing robustness and imperceptibility through wavelet and spectral projection. Using the Eq. (6) and Eq. (13), the cover image " + }, + { + "bbox": [ + 55, + 251, + 296, + 371 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 251, + 296, + 371 + ], + "type": "text", + "content": " is decomposed into sub-bands " + }, + { + "bbox": [ + 55, + 251, + 296, + 371 + ], + "type": "inline_equation", + "content": "S_{LL}, S_{LH}, S_{HL}" + }, + { + "bbox": [ + 55, + 251, + 296, + 371 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 251, + 296, + 371 + ], + "type": "inline_equation", + "content": "S_{HH}" + }, + { + "bbox": [ + 55, + 251, + 296, + 371 + ], + "type": "text", + "content": " within spectral domain, with " + }, + { + "bbox": [ + 55, + 251, + 296, + 371 + ], + "type": "inline_equation", + "content": "S_{HH}" + }, + { + "bbox": [ + 55, + 251, + 296, + 371 + ], + "type": "text", + "content": " providing high-frequency details for embedding. A variable number " + }, + { + "bbox": [ + 55, + 251, + 296, + 371 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 251, + 296, + 371 + ], + "type": "text", + "content": " of convolutional layers with a " + }, + { + "bbox": [ + 55, + 251, + 296, + 371 + ], + "type": "inline_equation", + "content": "K \\times K" + }, + { + "bbox": [ + 55, + 251, + 296, + 371 + ], + "type": "text", + "content": " kernel, followed by LeakyReLU activation, are recursively applied to " + }, + { + "bbox": [ + 55, + 251, + 296, + 371 + ], + "type": "inline_equation", + "content": "S_{HH}" + }, + { + "bbox": [ + 55, + 251, + 296, + 371 + ], + "type": "text", + "content": " to refine features:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 64, + 385, + 295, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 385, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 64, + 385, + 295, + 399 + ], + "type": "interline_equation", + "content": "S _ {H H} ^ {(n + 1)} = \\text {L e a k y R e L U} \\left(\\operatorname {C o n v} _ {2 D} \\left(S _ {H H} ^ {(n)}, K\\right)\\right), \\quad n = 1, \\dots , k. \\tag {14}", + "image_path": "dbe40248ef726d795294450ea03876caa0e83f465d32e7596362570f3269aaa1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 405, + 295, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 405, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 295, + 430 + ], + "type": "text", + "content": "The final output " + }, + { + "bbox": [ + 55, + 405, + 295, + 430 + ], + "type": "inline_equation", + "content": "S_{HH}^{(n + 1)}" + }, + { + "bbox": [ + 55, + 405, + 295, + 430 + ], + "type": "text", + "content": " from Eq. (14) represents the modified high-frequency band, primed for embedding." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 430, + 295, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 295, + 491 + ], + "type": "text", + "content": "The message " + }, + { + "bbox": [ + 55, + 430, + 295, + 491 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 430, + 295, + 491 + ], + "type": "text", + "content": ", represented as a binary vector of length " + }, + { + "bbox": [ + 55, + 430, + 295, + 491 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 430, + 295, + 491 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 430, + 295, + 491 + ], + "type": "inline_equation", + "content": "M \\in \\{0,1\\}^l" + }, + { + "bbox": [ + 55, + 430, + 295, + 491 + ], + "type": "text", + "content": "), with batch size " + }, + { + "bbox": [ + 55, + 430, + 295, + 491 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 430, + 295, + 491 + ], + "type": "text", + "content": " and message length " + }, + { + "bbox": [ + 55, + 430, + 295, + 491 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 430, + 295, + 491 + ], + "type": "text", + "content": ", is reshaped and expanded across channels " + }, + { + "bbox": [ + 55, + 430, + 295, + 491 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 430, + 295, + 491 + ], + "type": "text", + "content": " to align with " + }, + { + "bbox": [ + 55, + 430, + 295, + 491 + ], + "type": "inline_equation", + "content": "S_{HH}^{(n+1)}" + }, + { + "bbox": [ + 55, + 430, + 295, + 491 + ], + "type": "text", + "content": ". This ensures " + }, + { + "bbox": [ + 55, + 430, + 295, + 491 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{expanded}}" + }, + { + "bbox": [ + 55, + 430, + 295, + 491 + ], + "type": "text", + "content": " conforms to the dimension " + }, + { + "bbox": [ + 55, + 430, + 295, + 491 + ], + "type": "inline_equation", + "content": "[b,c,l]" + }, + { + "bbox": [ + 55, + 430, + 295, + 491 + ], + "type": "text", + "content": ", where each message is structured accordingly." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 492, + 295, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 295, + 610 + ], + "type": "text", + "content": "To localize the embedding, we create a radial mask centered at " + }, + { + "bbox": [ + 55, + 492, + 295, + 610 + ], + "type": "inline_equation", + "content": "(c_x, c_y) = \\left(\\frac{h}{2}, \\frac{w}{2}\\right)" + }, + { + "bbox": [ + 55, + 492, + 295, + 610 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 492, + 295, + 610 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 492, + 295, + 610 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 492, + 295, + 610 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 492, + 295, + 610 + ], + "type": "text", + "content": " represent the height and width of " + }, + { + "bbox": [ + 55, + 492, + 295, + 610 + ], + "type": "inline_equation", + "content": "S_{HH}" + }, + { + "bbox": [ + 55, + 492, + 295, + 610 + ], + "type": "text", + "content": ". The Euclidean distance " + }, + { + "bbox": [ + 55, + 492, + 295, + 610 + ], + "type": "inline_equation", + "content": "D(x_i, y_i)" + }, + { + "bbox": [ + 55, + 492, + 295, + 610 + ], + "type": "text", + "content": " from the center " + }, + { + "bbox": [ + 55, + 492, + 295, + 610 + ], + "type": "inline_equation", + "content": "(c_x, c_y)" + }, + { + "bbox": [ + 55, + 492, + 295, + 610 + ], + "type": "text", + "content": " is computed for each coefficient " + }, + { + "bbox": [ + 55, + 492, + 295, + 610 + ], + "type": "inline_equation", + "content": "(x_i, y_i)" + }, + { + "bbox": [ + 55, + 492, + 295, + 610 + ], + "type": "text", + "content": ". A binary mask is then generated within the predefined radius " + }, + { + "bbox": [ + 55, + 492, + 295, + 610 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 492, + 295, + 610 + ], + "type": "text", + "content": " based on the distance " + }, + { + "bbox": [ + 55, + 492, + 295, + 610 + ], + "type": "inline_equation", + "content": "D(x_i, y_i)" + }, + { + "bbox": [ + 55, + 492, + 295, + 610 + ], + "type": "text", + "content": ", such that if " + }, + { + "bbox": [ + 55, + 492, + 295, + 610 + ], + "type": "inline_equation", + "content": "D(x_i, y_i) \\leq r" + }, + { + "bbox": [ + 55, + 492, + 295, + 610 + ], + "type": "text", + "content": ", the mask value is set to 1, allowing embedding in the corresponding region. Otherwise, the mask value is 0, restricting embedding to areas within a specified radius " + }, + { + "bbox": [ + 55, + 492, + 295, + 610 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 492, + 295, + 610 + ], + "type": "text", + "content": ", ensuring focus on high-frequency regions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 611, + 295, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 295, + 635 + ], + "type": "text", + "content": "For each coordinate " + }, + { + "bbox": [ + 55, + 611, + 295, + 635 + ], + "type": "inline_equation", + "content": "(x_{i},y_{i})" + }, + { + "bbox": [ + 55, + 611, + 295, + 635 + ], + "type": "text", + "content": " where mask " + }, + { + "bbox": [ + 55, + 611, + 295, + 635 + ], + "type": "inline_equation", + "content": "(x_{i},y_{i})" + }, + { + "bbox": [ + 55, + 611, + 295, + 635 + ], + "type": "text", + "content": " is 1 and " + }, + { + "bbox": [ + 55, + 611, + 295, + 635 + ], + "type": "inline_equation", + "content": "W_{c}\\in c" + }, + { + "bbox": [ + 55, + 611, + 295, + 635 + ], + "type": "text", + "content": ", the embedding operation is performed as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 86, + 642, + 295, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 642, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 86, + 642, + 295, + 657 + ], + "type": "interline_equation", + "content": "S _ {H H} ^ {(n + 1)} [:, W _ {c}, x _ {i}, y _ {i} ] + = M _ {\\text {e x p a n d e d}} [:, W _ {c}, i ] \\cdot s, \\tag {15}", + "image_path": "0d831b7828d8a0e6e94c85158c155e5dbef397330206f8acd6453419a10a92b2.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 664, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 664, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 664, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 664, + 296, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 664, + 296, + 714 + ], + "type": "text", + "content": " is the strength factor controlling embedding intensity and invisibility. After embedding, the modified coefficients " + }, + { + "bbox": [ + 55, + 664, + 296, + 714 + ], + "type": "inline_equation", + "content": "S_{HH}^{(n + 1)}" + }, + { + "bbox": [ + 55, + 664, + 296, + 714 + ], + "type": "text", + "content": " undergo a final convolution and LeakyReLU using Eq. (14), by setting the value of " + }, + { + "bbox": [ + 55, + 664, + 296, + 714 + ], + "type": "inline_equation", + "content": "k = 1" + }, + { + "bbox": [ + 55, + 664, + 296, + 714 + ], + "type": "text", + "content": " to harmonize" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "content": "the embedded message. Following this approach, SpecGuard embeds the message into the spectral domain in a transformed form, differing from its original input representation. By blending the message seamlessly into the spectral space based on the " + }, + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "inline_equation", + "content": "W_{c}" + }, + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "content": ", it becomes inherently concealed within the domain, rendering its presence imperceptible. Without knowledge of " + }, + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "inline_equation", + "content": "W_{c}" + }, + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "content": ", it becomes exceedingly challenging to localize the embedded message, further enhancing the security of the system. This transformation ensures the embedding process remains opaque to any adversarial attacker, making SpecGuard black-box." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 204, + 554, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 554, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 554, + 323 + ], + "type": "text", + "content": "Reconstruction. SpecGuard encoder reconstructs the watermarked image " + }, + { + "bbox": [ + 313, + 204, + 554, + 323 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{embedded}}" + }, + { + "bbox": [ + 313, + 204, + 554, + 323 + ], + "type": "text", + "content": " by inverse transformation restoring " + }, + { + "bbox": [ + 313, + 204, + 554, + 323 + ], + "type": "inline_equation", + "content": "S_{HH}" + }, + { + "bbox": [ + 313, + 204, + 554, + 323 + ], + "type": "text", + "content": " back into the spatial domain. The reconstruction process integrates the inverse wavelet projection (IWP) [33] and inverse spectral projection (ISP) [18], ensuring the embedded modifications are correctly translated into the spatial domain. To reconstruct the spatial domain image, " + }, + { + "bbox": [ + 313, + 204, + 554, + 323 + ], + "type": "inline_equation", + "content": "S_{HH}" + }, + { + "bbox": [ + 313, + 204, + 554, + 323 + ], + "type": "text", + "content": " is combined with the other sub-bands " + }, + { + "bbox": [ + 313, + 204, + 554, + 323 + ], + "type": "inline_equation", + "content": "S_{LL}, S_{LH}" + }, + { + "bbox": [ + 313, + 204, + 554, + 323 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 204, + 554, + 323 + ], + "type": "inline_equation", + "content": "S_{HL}" + }, + { + "bbox": [ + 313, + 204, + 554, + 323 + ], + "type": "text", + "content": ". For the SP embedded in " + }, + { + "bbox": [ + 313, + 204, + 554, + 323 + ], + "type": "inline_equation", + "content": "S_{HH}" + }, + { + "bbox": [ + 313, + 204, + 554, + 323 + ], + "type": "text", + "content": ", the ISP is applied to reconstruct " + }, + { + "bbox": [ + 313, + 204, + 554, + 323 + ], + "type": "inline_equation", + "content": "S_{HH}" + }, + { + "bbox": [ + 313, + 204, + 554, + 323 + ], + "type": "text", + "content": " to spatial domain as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 323, + 327, + 553, + 355 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 327, + 553, + 355 + ], + "spans": [ + { + "bbox": [ + 323, + 327, + 553, + 355 + ], + "type": "interline_equation", + "content": "S _ {H H} (x, y) = \\sum_ {u = 0} ^ {L - 1} \\sum_ {v = 0} ^ {L - 1} \\zeta (u, v) \\cdot \\exp \\left(i \\frac {2 \\pi}{L} (x \\cdot u + y \\cdot v)\\right), \\tag {16}", + "image_path": "1c3617d61dd0682250d643b38ea73774d88d8bff6b70d8c5ea5ac799bfb0e153.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "inline_equation", + "content": "\\zeta(u, v)" + }, + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "text", + "content": " represents spectral coefficients from the embedding process, " + }, + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "text", + "content": " denotes the dimension of " + }, + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "inline_equation", + "content": "S_{HH}" + }, + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "inline_equation", + "content": "(x, y)" + }, + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "text", + "content": " are spatial coordinates. SpecGuard then reconstructs the watermarked image " + }, + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{embedded}}" + }, + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "text", + "content": " using the IWP as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 346, + 410, + 553, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 410, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 346, + 410, + 553, + 422 + ], + "type": "interline_equation", + "content": "I _ {\\text {e m b e d d e d}} (x, y) = \\operatorname {I W P} \\left(S _ {L L}, S _ {L H}, S _ {H L}, S _ {H H}\\right). \\tag {17}", + "image_path": "217d32d9a3524b8df0a5b644f1033ed41a261383be8ad7a6814e9f66963805b2.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 426, + 554, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 554, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 554, + 486 + ], + "type": "text", + "content": "This process seamlessly embeds the watermark message " + }, + { + "bbox": [ + 313, + 426, + 554, + 486 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 426, + 554, + 486 + ], + "type": "text", + "content": " in the spectral domain, preserving the cover image " + }, + { + "bbox": [ + 313, + 426, + 554, + 486 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 426, + 554, + 486 + ], + "type": "text", + "content": "'s integrity. The inverse transformations that are expressed in Eq. (16) and Eq. (17) fully restore visual quality, maintaining all frequency components." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 493, + 376, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 493, + 376, + 504 + ], + "spans": [ + { + "bbox": [ + 314, + 493, + 376, + 504 + ], + "type": "text", + "content": "3.2.Decoder" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 510, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 629 + ], + "type": "text", + "content": "As shown in Algorithm 1, SpecGuard decoding process starts by applying wavelet projection ( Eq. (1)) to the watermarked image " + }, + { + "bbox": [ + 313, + 510, + 554, + 629 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{embedded}}" + }, + { + "bbox": [ + 313, + 510, + 554, + 629 + ], + "type": "text", + "content": ", separating it into low and high-frequency bands, where the high-frequency band " + }, + { + "bbox": [ + 313, + 510, + 554, + 629 + ], + "type": "inline_equation", + "content": "S_{D_{HH}}^{high}" + }, + { + "bbox": [ + 313, + 510, + 554, + 629 + ], + "type": "text", + "content": " contains the embedded message similar to the process in the encoding phase, particularly in Eq. (6). An approximation of the spectral projection using FFT as shown in Eq. (13) is then applied to " + }, + { + "bbox": [ + 313, + 510, + 554, + 629 + ], + "type": "inline_equation", + "content": "S_{D_{HH}}^{high}" + }, + { + "bbox": [ + 313, + 510, + 554, + 629 + ], + "type": "text", + "content": " returning the transformed data " + }, + { + "bbox": [ + 313, + 510, + 554, + 629 + ], + "type": "inline_equation", + "content": "S_{D_{HH}}^{\\mathrm{sp}}" + }, + { + "bbox": [ + 313, + 510, + 554, + 629 + ], + "type": "text", + "content": ". Then, " + }, + { + "bbox": [ + 313, + 510, + 554, + 629 + ], + "type": "inline_equation", + "content": "S_{D_{HH}}" + }, + { + "bbox": [ + 313, + 510, + 554, + 629 + ], + "type": "text", + "content": " is further refined through convolutional layers that captures the local features for message extraction." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "To extract the message, a radial mask is created to isolate high-frequency areas within " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "S_{D_{HH}}" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": ", targeting the embedded regions based on their distance from the center. The masked values are compared against a learnable threshold " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": " to decode each bit of the hidden message " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{M}}" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": " serves as a threshold that adapts to the spectral patterns across the entire image, learning the distinct characteristics" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17987" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 58, + 98, + 296, + 417 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 95 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 295, + 95 + ], + "type": "text", + "content": "Algorithm 1 SpecGuard decoder with wavelet, spectral projection with FFT approximation, and learnable threshold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "lines": [ + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "spans": [ + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "text", + "content": "1: Input: Watermarked image " + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{embedded}}" + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "text", + "content": " , learnable " + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "text", + "content": " message length " + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "text", + "content": " \nradius " + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "text", + "content": " ,watermark channel " + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "inline_equation", + "content": "W_{c}" + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "text", + "content": " \n2: Output: Decoded binary message " + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{M}}" + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "text", + "content": " \n3: Procedure: Apply Wavelet Projection on " + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{embedded}}" + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "inline_equation", + "content": "S_{D_{LL}}" + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "text", + "content": " (low-frequency) and " + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "inline_equation", + "content": "S_{D_{HH}}^{\\mathrm{high}}" + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "text", + "content": " (high-frequency) \n4: Procedure: Spectral approximation with FFT " + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "inline_equation", + "content": "(S_{D_{HH}}^{\\mathrm{high}})" + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "text", + "content": " .. \n5: Separate even and odd indices: " + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "inline_equation", + "content": "v = [x_{\\mathrm{even}}," + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "text", + "content": " reverse " + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "inline_equation", + "content": "(x_{\\mathrm{odd}})]" + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "text", + "content": " \n6: Compute FFT on " + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "text", + "content": " .. " + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{complex}} = \\mathrm{FFT}(v)" + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "text", + "content": " \n7: " + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{real}} = V_{\\mathrm{complex}}\\cdot \\left[\\cos \\left(\\frac{-\\pi k}{2N}\\right),\\sin \\left(\\frac{-\\pi k}{2N}\\right)\\right] / /" + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "text", + "content": " Calculate Real \n8: " + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{real}}[0]\\gets \\frac{V_{\\mathrm{real}}[0]}{\\sqrt{N\\cdot 2}},V_{\\mathrm{real}}[1:\\gets \\frac{V_{\\mathrm{real}}[1:]}{\\sqrt{\\frac{N}{2}\\cdot 2}} /" + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "text", + "content": " Energy preservation \n9: Transpose result and repeat to obtain " + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "inline_equation", + "content": "S_{D_{HH}}^{\\mathrm{sp}}" + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "text", + "content": " \n10: Return " + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "inline_equation", + "content": "S_{D_{HH}}^{\\mathrm{sp}}" + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "text", + "content": " \n11: Procedure: Pass " + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "inline_equation", + "content": "S_{D_{HH}}^{\\mathrm{sp}}" + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "text", + "content": " through sequential layers as: \n" + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "inline_equation", + "content": "S_{D_{HH}}^{(n + 1)} =" + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "text", + "content": " LeakyReLU " + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "inline_equation", + "content": "\\left(\\operatorname {Conv}_{2D}\\left(S_{HH}^{sp(n)},K\\right)\\right)" + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "inline_equation", + "content": "n = 1,\\dots,k," + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "text", + "content": " \n12: Return " + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "inline_equation", + "content": "S_{D_{HH}}^{(n + 1)}" + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "text", + "content": " \n13: Procedure: Extraction " + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "inline_equation", + "content": "(S_{D_{HH}}^{(n + 1)},l)" + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "text", + "content": " . \n14: Set " + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "inline_equation", + "content": "(c_x,c_y) = \\left(\\frac{H}{2},\\frac{W}{2}\\right)" + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "text", + "content": " \n15: Generate mask for high-frequency region within radius " + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "text", + "content": " \nfor each coordinate " + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "text", + "content": " do: \n" + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "inline_equation", + "content": "D(x_i,y_i) = \\sqrt{(x_i - c_x)^2 + (y_i - c_y)^2}" + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "text", + "content": " // Euclidian Distance \n17: if " + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "inline_equation", + "content": "D(x_i,y_i)\\leq r" + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "text", + "content": " then \n18: Set mask[i, " + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "inline_equation", + "content": "j] = 1" + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "text", + "content": " \n19: end if \n20: end for \n21: Extract mask: " + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "inline_equation", + "content": "S_{D_{HH}}[:,W_c," + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "text", + "content": " mask[i, " + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "inline_equation", + "content": "j]" + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "text", + "content": " \n22: Decode message using learnable " + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 58, + 98, + 296, + 417 + ], + "type": "text", + "content": " .." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 125, + 422, + 240, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 422, + 240, + 449 + ], + "spans": [ + { + "bbox": [ + 125, + 422, + 240, + 449 + ], + "type": "interline_equation", + "content": "D _ {\\mathbf {M}} [ i ] = \\left\\{ \\begin{array}{l l} 1 & \\text {i f E x t r a c t e d} [ i ] > \\theta \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right.", + "image_path": "53d7a72a1786b6d037bdfa2d5e8265fe8c74846eed65d78cfe5ce6016d2a9a19.jpg" + } + ] + } + ], + "index": 2 + }, + { + "type": "code", + "bbox": [ + 56, + 455, + 295, + 476 + ], + "blocks": [ + { + "bbox": [ + 56, + 455, + 295, + 476 + ], + "lines": [ + { + "bbox": [ + 56, + 455, + 295, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 295, + 476 + ], + "type": "text", + "content": "23: Update " + }, + { + "bbox": [ + 56, + 455, + 295, + 476 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 56, + 455, + 295, + 476 + ], + "type": "text", + "content": " dynamically: " + }, + { + "bbox": [ + 56, + 455, + 295, + 476 + ], + "type": "inline_equation", + "content": "\\theta \\gets \\theta -\\eta \\cdot \\frac{\\partial L_{\\mathrm{dec}}}{\\partial\\theta}" + }, + { + "bbox": [ + 56, + 455, + 295, + 476 + ], + "type": "text", + "content": " // Optimizes robustness \n24: Return " + }, + { + "bbox": [ + 56, + 455, + 295, + 476 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{M}}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 506, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 296, + 555 + ], + "type": "text", + "content": "of the embedded watermark. From Parseval's theorem [20] ensures overall spectral and spatial energies remain equivalent, though local spectral energy distributions are altered by the watermark strength factor " + }, + { + "bbox": [ + 55, + 506, + 296, + 555 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 506, + 296, + 555 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": "The watermark's strength factor " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " ensures that the high-energy areas where the message " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " is embedded as \"1\" remain robust, experiencing a minimum distortion in such conditions. Moreover, this threshold can be optimized for better bit recovery accuracy during training. As " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " learns, it recognizes that areas encoded as \"1\" carry higher energy and impact due to the strength factor " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " of Eq. (15), while areas marked as \"0\", softened by the LeakyReLU's minimal negative slope, have a lower intensity. Such a dynamic approach enables " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " to identify and protect the embedded message " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " even when external disturbances occur, preserving the watermark's structure within the watermarked image " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "I_{\\text{embedded}}" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": ". And, " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " effectively learns to distin" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "guish high-energy watermark regions. Therefore, the embedded message is more recoverable under diverse attacks, and SpecGuard's decoder ensures valid watermark bit extraction. Theoretical explanation of Parseval theorem's [20] impact on message extraction is in the Supplementary." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 137, + 488, + 151 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 137, + 488, + 151 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 488, + 151 + ], + "type": "text", + "content": "3.3. Loss Calculation for SpecGuard" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 155, + 554, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 155, + 554, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 155, + 554, + 203 + ], + "type": "text", + "content": "To achieve the training objective of robust and invisible watermark embedding, a composite loss function is defined with two terms: encoder loss " + }, + { + "bbox": [ + 313, + 155, + 554, + 203 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{enc}}" + }, + { + "bbox": [ + 313, + 155, + 554, + 203 + ], + "type": "text", + "content": " as expressed in Eq. (18) and decoder loss " + }, + { + "bbox": [ + 313, + 155, + 554, + 203 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{dec}}" + }, + { + "bbox": [ + 313, + 155, + 554, + 203 + ], + "type": "text", + "content": " as expressed in Eq. (19)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 322, + 205, + 554, + 223 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 205, + 554, + 223 + ], + "spans": [ + { + "bbox": [ + 322, + 205, + 554, + 223 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathbb {E} _ {(I, M) \\sim D} L _ {\\text {e n c}} (I, I _ {\\text {e m b e d d e d}}) = \\| E _ {\\theta} (I, M) - I \\| ^ {2}, \\tag {18}", + "image_path": "fdff86516b048b45cda9d9ceb0b52614901e046b09eda45b83da24d2977bb26f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 319, + 239, + 554, + 256 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 239, + 554, + 256 + ], + "spans": [ + { + "bbox": [ + 319, + 239, + 554, + 256 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathbb {E} _ {(I, M) \\sim D} L _ {\\mathrm {d e c}} (M, D _ {\\mathrm {M}}) = \\| D _ {\\theta} \\left(I _ {\\text {e m b e d d e d}}\\right) - M \\| ^ {2}, \\tag {19}", + "image_path": "5f433fe24b79dd11580bd74a76cb4a8f42ff7ac85593e97d648d0124daa4136e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 259, + 554, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 259, + 554, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 554, + 354 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 259, + 554, + 354 + ], + "type": "inline_equation", + "content": "E_{\\theta}(I,M)" + }, + { + "bbox": [ + 313, + 259, + 554, + 354 + ], + "type": "text", + "content": " denotes the encoder output, embedding the message " + }, + { + "bbox": [ + 313, + 259, + 554, + 354 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 259, + 554, + 354 + ], + "type": "text", + "content": " into the cover image " + }, + { + "bbox": [ + 313, + 259, + 554, + 354 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 259, + 554, + 354 + ], + "type": "text", + "content": " to produce " + }, + { + "bbox": [ + 313, + 259, + 554, + 354 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{embedded}}" + }, + { + "bbox": [ + 313, + 259, + 554, + 354 + ], + "type": "text", + "content": ". By minimizing " + }, + { + "bbox": [ + 313, + 259, + 554, + 354 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{enc}}" + }, + { + "bbox": [ + 313, + 259, + 554, + 354 + ], + "type": "text", + "content": ", the encoder learns to embed the watermark invisibly, preserving the fidelity of the cover image. " + }, + { + "bbox": [ + 313, + 259, + 554, + 354 + ], + "type": "inline_equation", + "content": "D_{\\theta}(I_{\\mathrm{embedded}})" + }, + { + "bbox": [ + 313, + 259, + 554, + 354 + ], + "type": "text", + "content": " denotes the decoder's output from the watermarked image " + }, + { + "bbox": [ + 313, + 259, + 554, + 354 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{embedded}}" + }, + { + "bbox": [ + 313, + 259, + 554, + 354 + ], + "type": "text", + "content": ". Minimizing " + }, + { + "bbox": [ + 313, + 259, + 554, + 354 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{dec}}" + }, + { + "bbox": [ + 313, + 259, + 554, + 354 + ], + "type": "text", + "content": " enables the decoder to reliably retrieve the embedded message under varying conditions, such as noise and transformation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 355, + 554, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 554, + 391 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 554, + 391 + ], + "type": "text", + "content": "The total loss " + }, + { + "bbox": [ + 313, + 355, + 554, + 391 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 355, + 554, + 391 + ], + "type": "text", + "content": " as shown in Eq. (20) used for optimizing the model combines these terms, balancing invisibility and robustness through weighted coefficients as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 375, + 395, + 553, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 395, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 375, + 395, + 553, + 410 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} L = \\lambda_ {\\text {e n c}} L _ {\\text {e n c}} + \\lambda_ {\\text {d e c}} L _ {\\text {d e c}}, \\tag {20}", + "image_path": "9cbf7dc716bdbe660f7553db7c94b991078e3067267f97a94fac871b1b858b6d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 414, + 554, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 554, + 439 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 554, + 439 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 414, + 554, + 439 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{enc}}" + }, + { + "bbox": [ + 313, + 414, + 554, + 439 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 414, + 554, + 439 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{dec}}" + }, + { + "bbox": [ + 313, + 414, + 554, + 439 + ], + "type": "text", + "content": " control the relative importance of visual fidelity and message recoverability." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 447, + 440, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 440, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 440, + 460 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 467, + 372, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 372, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 372, + 478 + ], + "type": "text", + "content": "4.1. Dataset" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 483, + 555, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 688 + ], + "type": "text", + "content": "SpecGuard is trained on the MS-COCO dataset [26], which contains 25K images. To evaluate the robustness of the watermarking methods including our SpecGuard against different types of attacks, such as distortions, regenerations, and adversarial attacks, we used three datasets: DiffusionDB [46], MS-COCO [26], and DALL-E31. Each of these datasets has a unique distribution of prompt words. We also ensured that no unethical or violent terms were included in the prompts. We randomly picked 200 images from MS-COCO [26] and applied watermark using SpecGuard for further verifying the robustness after uploading on various social media platforms and applying AI-based Photoshop Neural Filters (PNFs)2. The PNFs include depth blur, artistic style transfer, super zoom, JPEG artifact reduction, and colorization. For the super zoom filter, we set the 'Sharpen' and 'Noise Reduction' parameters to 15. For all other filters, we used the default settings." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 693, + 528, + 702 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 693, + 528, + 702 + ], + "spans": [ + { + "bbox": [ + 324, + 693, + 528, + 702 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 324, + 693, + 528, + 702 + ], + "type": "text", + "content": "https://huggingface.co/datasets/OpenDatasets/dalle-3-dataset" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 326, + 703, + 529, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 703, + 529, + 712 + ], + "spans": [ + { + "bbox": [ + 326, + 703, + 529, + 712 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 326, + 703, + 529, + 712 + ], + "type": "text", + "content": "https://www.adobe.com/products/photoshop/neural-filter.html" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17988" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 69, + 282, + 195 + ], + "blocks": [ + { + "bbox": [ + 69, + 69, + 282, + 195 + ], + "lines": [ + { + "bbox": [ + 69, + 69, + 282, + 195 + ], + "spans": [ + { + "bbox": [ + 69, + 69, + 282, + 195 + ], + "type": "image", + "image_path": "a0aa4adfdd645083602c529592e0a888bef0b968035d569a2332b2aa6bacde2f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 202, + 295, + 236 + ], + "lines": [ + { + "bbox": [ + 55, + 202, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 295, + 236 + ], + "type": "text", + "content": "Figure 3. Some best results for cover vs watermarked images with PSNR/SSIM " + }, + { + "bbox": [ + 55, + 202, + 295, + 236 + ], + "type": "inline_equation", + "content": "(\\uparrow)" + }, + { + "bbox": [ + 55, + 202, + 295, + 236 + ], + "type": "text", + "content": " scores showing minimal visual degradation when watermarked using proposed SpecGuard." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 251, + 303, + 304 + ], + "blocks": [ + { + "bbox": [ + 56, + 251, + 303, + 304 + ], + "lines": [ + { + "bbox": [ + 56, + 251, + 303, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 303, + 304 + ], + "type": "table", + "html": "
Metrics256 × 256512 × 5121024 × 1024
CelebA-HQMS-COCOCelebA-HQMS-COCOCelebA-HQMS-COCO
PSNR↑40.36140.32044.65144.68048.17048.081
SSIM↑0.98890.98880.99270.99270.99370.9936
FID↓16.45116.69016.97217.02017.44616.955
MSE↓0.00020.00020.00010.00010.00010.0001
", + "image_path": "5f45c501721dd458500ffdb994d641de3800a94dae010fd0a877fc775eb9f492.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 358, + 153, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 153, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 153, + 371 + ], + "type": "text", + "content": "4.2. Implementation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": "We used CUDA v11.3 and PyTorch with a batch size of 32 and the Adam optimizer on a multiple NVIDIA RTX 2080-equipped server. Mean Squared Error (MSE) and Bit Recovery Accuracy (BRA) are used for loss and accuracy calculation. We used Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), Fréchet Inception Distance (FID), and MSE to evaluate perceptual quality. Our model is trained for 300 epochs, with the decoder learning rate set to " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-3}" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": ", reduced by half every 100 steps, and the encoder learning rate is set to " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-2}" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": " without scheduling. We set our watermark radius " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": "(r)" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": ", strength factor " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": "(s)" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": ", initial learning parameter " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": "(\\theta)" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": ", and the number of convolutional layers " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": "(k)" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": " to 100, 20, 0.001, and 32, respectively. This setup is applied with a message bit length (BL) of 48, 64, 128, and 256. Initially, decoder loss weight " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": "(\\lambda_{\\mathrm{dec}})" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": " and encoder loss weight " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": "(\\lambda_{\\mathrm{enc}})" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": " are set to 1.0 and 0.7, respectively. For assessing the robustness of watermarking methods against diverse attacks, we inherited the experimental setups from Waves [8] and used effective metrics such as \"Quality at 95% Performance (Q@0.95P)\", \"Quality at 70% Performance (Q@0.7P)\", \"Avg P\" and \"Avg Q.\" Here, Q@0.95P and Q@0.7P indicate the level of image quality degradation required for watermark detection accuracy to reach 95% and 70%, respectively. The average performance (Avg P) metric represents the mean detection accuracy across various attack strengths, while the average quality degradation (Avg Q) measures the overall impact of attacks on image quality [8, 35]." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 316, + 70, + 552, + 355 + ], + "blocks": [ + { + "bbox": [ + 55, + 311, + 296, + 334 + ], + "lines": [ + { + "bbox": [ + 55, + 311, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 296, + 334 + ], + "type": "text", + "content": "Table 1. Perceptual quality of SpecGuard watermarked images across resolutions and datasets using a fixed 30-bit message length." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 70, + 552, + 355 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 552, + 355 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 552, + 355 + ], + "type": "table", + "html": "
MethodsVenueBLPSNR↑SSIM↑FID↓BRA↑
Pre-processing methodsTree-Ring [47]NeurIPS'236432.330.9117.70.98
12832.100.9017.80.96
25631.850.8917.90.94
Stable Signature [35]ICCV'236430.000.8919.60.98
12829.800.8819.70.96
25629.500.8719.80.96
Yang et al. [49]CVPR'246431.450.9018.20.98
12831.200.8918.30.93
25630.950.8818.40.89
SleeperMark [45]CVPR'256431.800.9218.00.97
12831.600.9118.10.93
25631.350.9018.20.87
Post-processing methodsHiDDeN [53]ECCV'186432.010.8819.70.98
12831.800.8719.80.85
25631.500.8619.90.82
StegaStamp [41]CVPR'206428.500.9117.90.99
12828.200.9018.00.98
25628.000.8918.10.94
MBRS [17]ACM MM'216438.200.9617.90.98
12837.900.9518.00.96
25637.500.9418.20.94
FIN [10]AAAI'236436.700.9518.30.97
12836.400.9418.40.96
25636.100.9318.50.96
MuST [43]AAAI'246441.200.9717.50.98
12840.900.9617.60.93
25640.500.9517.80.90
EditGuard [52]CVPR'246441.560.9717.80.98
12841.300.9617.90.97
25640.900.9518.00.97
SpecGuard (Ours)ICCV'256442.590.9817.20.99
12842.890.9917.00.99
25640.860.9917.60.98
", + "image_path": "e26592245971c4568b85a395ec79c4a5ee94aba7e1f88ef07e988612e10730f6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 363, + 555, + 386 + ], + "lines": [ + { + "bbox": [ + 313, + 363, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 555, + 386 + ], + "type": "text", + "content": "Table 2. Comparison of SOTA pre-processing and post-processing watermarking methods with SpecGuard without attacks." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 404, + 444, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 404, + 444, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 404, + 444, + 417 + ], + "type": "text", + "content": "4.3. Watermark Invisibility" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 424, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 555, + 555 + ], + "type": "text", + "content": "To evaluate the invisibility of the embedded watermark, we conducted perceptual and quantitative assessments using SpecGuard. As shown in Fig. 3, there is no noticeable perceptual degradation between the cover and watermarked images, confirming that the watermark remains imperceptible to the human eye. For a more comprehensive evaluation, we created three subsets of different image sizes ranging between 256 to 1024 with images from the MS-COCO [26] and CelebA-HQ [19] datasets and applied the SpecGuard watermarking method to compare the average PSNR values between the cover and watermarked images, as in Tab. 1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": "For quantitative evaluation, we further compare the performance of SpecGuard with the SOTA pre-processing and post-processing watermarking methods. As presented in Tab. 2, SpecGuard achieves the highest PSNR of 42.89 when the bit length was 128. Additionally, it attains the highest SSIM of 0.99 at a BL of 128 and 256 among all compared methods, indicating minimal visual distortion. Additionally, SpecGuard achieved the lowest FID of 17.0 and the highest BRA of 0.99, ensuring strong robustness while maintaining imperceptibility. Overall, our results demonstrate that SpecGuard outperforms both pre-processing and post-processing watermarking methods, achieving superior imperceptibility and robustness." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17989" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 69, + 555, + 302 + ], + "blocks": [ + { + "bbox": [ + 55, + 69, + 555, + 302 + ], + "lines": [ + { + "bbox": [ + 55, + 69, + 555, + 302 + ], + "spans": [ + { + "bbox": [ + 55, + 69, + 555, + 302 + ], + "type": "table", + "html": "
Attack TypeTree-Ring [47]Stable Signature [35]StegaStamp [41]SpecGuard (Ours)
Q@0.95PQ@0.7PAvg PAvg QQ@0.95PQ@0.7PAvg PAvg QQ@0.95PQ@0.7PAvg PAvg QQ@0.95PQ@0.7PAvg PAvg Q
DistortionsRotation0.4640.5210.3750.6480.6240.7020.5940.6500.4230.4980.3570.6160.8630.8630.6870.653
Crop0.5920.5920.3320.463infinf0.9950.4610.6020.6020.5400.4510.8120.8120.9980.742
Brightinfinfinf0.304infinf0.9980.305infinf0.9980.317infinf0.9980.466
Contrastinfinf0.9980.243infinf0.9980.243infinf0.9980.231infinf0.9980.556
Blur0.8611.1120.5631.221- inf- inf0.0001.2040.8480.9620.4141.0000.921inf1.0001.452
Noise0.548inf0.9800.3950.4020.5200.8700.390infinf1.0000.360infinf0.9990.568
JPEG0.4990.4990.9290.2840.4850.4850.7930.284infinf0.9980.263infinf1.0000.495
Geo0.5250.5930.2770.7680.850inf0.9370.7670.6630.6930.3960.7330.8690.8690.8650.623
Deg0.620inf0.8920.6940.2060.3690.3000.6790.8260.9750.8520.6640.8951.1410.9150.749
Combine0.5390.7510.4030.9080.5380.6910.3340.9000.9451.1010.7950.8700.9791.2560.9110.952
RegenerationRegen-Diff- inf0.3070.6120.323- inf- inf0.0010.3000.331inf0.9430.327infinf0.9820.477
Regen-DiffPinf0.3070.6010.327- inf- inf0.0010.3030.333inf0.9400.329infinf0.9820.562
Regen-VAE0.5780.5780.8320.3480.5450.5450.5160.339infinf1.0000.343infinf0.9950.521
Regen-KLVAEinfinf0.9900.233 6- inf0.1760.2170.206infinf1.0000.240infinf0.9900.492
Rinse-2xDiff- inf0.3330.5100.357- inf- inf0.0010.3320.391inf0.9410.366infinf0.9930.561
Rinse-4xDiff- inf0.3550.4430.466- inf- inf0.0000.4380.388inf0.9090.477infinf0.9920.533
AdversarialAdvEmbG-KLVAE8- inf0.1640.4480.253infinf0.9980.249infinf1.0000.232infinf1.0000.456
AdvEmbB-RN180.241inf0.9530.218infinf0.9990.212infinf1.0000.196infinf1.0000.467
AdvEmbB-CLIP0.541inf0.9320.549infinf0.9990.541infinf1.0000.488infinf1.0000.436
AdvEmbB-KLVAE160.195inf0.8880.238infinf0.9970.233infinf1.0000.206infinf1.0000.482
AdvEmbB-SdxlVAE0.222inf0.9340.221infinf0.9980.219infinf1.0000.204infinf1.0000.492
AdvCls-UnWM&WM- inf0.1020.4990.145infinf0.9990.101infinf1.0000.101infinf1.0000.497
AdvCls-Real&WMinfinf1.0000.047infinf0.9980.092infinf1.0000.106infinf1.0000.427
AdvCls-WM1&WM2- inf0.1010.4920.139infinf0.9990.084infinf1.0000.129infinf1.0000.441
", + "image_path": "2a71067f8a7cecffde21a5c91f6e92b06087c2314d3e6580ab8f3c587d78eb5f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 311, + 555, + 344 + ], + "lines": [ + { + "bbox": [ + 55, + 311, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 555, + 344 + ], + "type": "text", + "content": "Table 3. Robustness comparison various across attacks using Q@0.95P(↑), Q@0.7P(↑), Avg P(↑) and Avg Q(↑). Here, 'inf' denotes that no attack was sufficient to degrade performance below the threshold, indicating strong robustness, whereas '-inf' signifies that even the weakest attack caused detection to fall below the threshold, reflecting weak robustness." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 67, + 354, + 164, + 458 + ], + "blocks": [ + { + "bbox": [ + 67, + 354, + 164, + 458 + ], + "lines": [ + { + "bbox": [ + 67, + 354, + 164, + 458 + ], + "spans": [ + { + "bbox": [ + 67, + 354, + 164, + 458 + ], + "type": "image", + "image_path": "c3af77e50ef9cbb097edd6fa72e5f5dc1f03dcdb43ce23ff8d8a8d959ba8cd77.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 164, + 354, + 257, + 458 + ], + "blocks": [ + { + "bbox": [ + 164, + 354, + 257, + 458 + ], + "lines": [ + { + "bbox": [ + 164, + 354, + 257, + 458 + ], + "spans": [ + { + "bbox": [ + 164, + 354, + 257, + 458 + ], + "type": "image", + "image_path": "241c17b3a2b9fbfba53c2704606421d5b419030a809fb05f6cd4584c9e9e8191.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 257, + 354, + 352, + 458 + ], + "blocks": [ + { + "bbox": [ + 257, + 354, + 352, + 458 + ], + "lines": [ + { + "bbox": [ + 257, + 354, + 352, + 458 + ], + "spans": [ + { + "bbox": [ + 257, + 354, + 352, + 458 + ], + "type": "image", + "image_path": "3df03b9919033d0e53d1f6dd7d95dbe480ad075013eae83e0ac50a1c53f49898.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 353, + 354, + 447, + 458 + ], + "blocks": [ + { + "bbox": [ + 353, + 354, + 447, + 458 + ], + "lines": [ + { + "bbox": [ + 353, + 354, + 447, + 458 + ], + "spans": [ + { + "bbox": [ + 353, + 354, + 447, + 458 + ], + "type": "image", + "image_path": "4b653ed4ee169dd20947b2d3dc9e8d5e0b593ccc7fedcef8cab76ddd2d56d318.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 447, + 354, + 542, + 458 + ], + "blocks": [ + { + "bbox": [ + 447, + 354, + 542, + 458 + ], + "lines": [ + { + "bbox": [ + 447, + 354, + 542, + 458 + ], + "spans": [ + { + "bbox": [ + 447, + 354, + 542, + 458 + ], + "type": "image", + "image_path": "5270cac58775daa7649a400c4517462d0dd4897b733eecdbe49d58293744c7ee.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 67, + 459, + 162, + 563 + ], + "blocks": [ + { + "bbox": [ + 67, + 459, + 162, + 563 + ], + "lines": [ + { + "bbox": [ + 67, + 459, + 162, + 563 + ], + "spans": [ + { + "bbox": [ + 67, + 459, + 162, + 563 + ], + "type": "image", + "image_path": "b65904e3bc4503ca2ba4bee9c230545e7542f295729e50187ca6365eaafd7b81.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 571, + 555, + 594 + ], + "lines": [ + { + "bbox": [ + 55, + 571, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 555, + 594 + ], + "type": "text", + "content": "Figure 4. Robustness validation of our proposed SpecGuard under different distortion attacks, including geometric transformations: Geo (rotation, cropping), photometric modifications (brightness, contrast), and degradations: Deg (blur, noise, JPEG compression)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 164, + 459, + 257, + 563 + ], + "blocks": [ + { + "bbox": [ + 164, + 459, + 257, + 563 + ], + "lines": [ + { + "bbox": [ + 164, + 459, + 257, + 563 + ], + "spans": [ + { + "bbox": [ + 164, + 459, + 257, + 563 + ], + "type": "image", + "image_path": "789232424565e24e2ea17562c99484682a1733df209991e89e396a6149503d4e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 257, + 459, + 352, + 563 + ], + "blocks": [ + { + "bbox": [ + 257, + 459, + 352, + 563 + ], + "lines": [ + { + "bbox": [ + 257, + 459, + 352, + 563 + ], + "spans": [ + { + "bbox": [ + 257, + 459, + 352, + 563 + ], + "type": "image", + "image_path": "64eea155e5ca4eeaa015eb0b296aef43acd61b56cdf4f38fc6fa2b72180833f1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 353, + 459, + 447, + 563 + ], + "blocks": [ + { + "bbox": [ + 353, + 459, + 447, + 563 + ], + "lines": [ + { + "bbox": [ + 353, + 459, + 447, + 563 + ], + "spans": [ + { + "bbox": [ + 353, + 459, + 447, + 563 + ], + "type": "image", + "image_path": "b46ee883b746c9f9e646c4abf09ba7c7cfcca0168764f5b5c5f5bd6ea3f5e85a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 447, + 459, + 542, + 563 + ], + "blocks": [ + { + "bbox": [ + 447, + 459, + 542, + 563 + ], + "lines": [ + { + "bbox": [ + 447, + 459, + 542, + 563 + ], + "spans": [ + { + "bbox": [ + 447, + 459, + 542, + 563 + ], + "type": "image", + "image_path": "5aacc06a0c1b739db3d3ea1c7bb773eadce67648e90eac85bd0025e7b27edb56.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 605, + 120, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 120, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 120, + 618 + ], + "type": "text", + "content": "4.4. Capacity" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": "To evaluate embedding capacity, we examined SpecGuard across different bit lengths and compared it with SOTA watermarking methods. Our experiments with 64, 128, and 256 bits demonstrate SpecGuard's high capacity of bit embedding while maintaining perceptual quality and robustness, as shown in Tab. 2. Notably, it achieves a PSNR of 42.89, the highest among all methods, along with the high" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 555, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 702 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 702 + ], + "type": "text", + "content": "est BRA of 0.99 and the lowest FID of 17.0 at 128 bits, ensuring minimal visual impact. This adaptability to different bit lengths without quality loss makes SpecGuard ideal for applications requiring flexible watermark sizes. Unlike StegaStamp and HiDDeN, which suffer reduced BRA for higher message bits, SpecGuard consistently extracts bits across all tested lengths. SpecGuard's theoretical watermark capacity is provided in the Supplementary." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "17990" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 288, + 131 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 288, + 131 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 288, + 131 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 288, + 131 + ], + "type": "table", + "html": "
ModulesPSNR/SSIM↑BRA↑ModulesPSNR/SSIM↑BRA↑
WP(L1)40.51/0.960.92WP(L1)+SPFA42.89/0.990.99
WP(L2)38.15/0.930.87WP(L2)+SPFA36.25/0.920.89
AttacksPSNR/SSIM↑BRA↑AttacksPSNR/SSIM↑BRA↑
Rotate (45°)12.15/21.310.82Rotate (90°)11.15/19.310.65
Blur (0.3)35.01/0.950.98Blur (0.6)30.11/0.910.98
Geo (0.3)12.08/0.500.93Geo (0.6)10.25/0.450.86
", + "image_path": "3c1378a268f6043bfc033716b045ad086583b54b6b50a4d5ece5a6c2d41dff48.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 67, + 131, + 286, + 139 + ], + "lines": [ + { + "bbox": [ + 67, + 131, + 286, + 139 + ], + "spans": [ + { + "bbox": [ + 67, + 131, + 286, + 139 + ], + "type": "text", + "content": "*WP: Wavelet Projection, SP: Spectral Projection, FA: FFT Approximation" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 171, + 296, + 224 + ], + "blocks": [ + { + "bbox": [ + 55, + 148, + 295, + 171 + ], + "lines": [ + { + "bbox": [ + 55, + 148, + 295, + 171 + ], + "spans": [ + { + "bbox": [ + 55, + 148, + 295, + 171 + ], + "type": "text", + "content": "Table 4. Ablation studies on the proposed SpecGuard for across various configurations, setting " + }, + { + "bbox": [ + 55, + 148, + 295, + 171 + ], + "type": "inline_equation", + "content": "M = 128" + }, + { + "bbox": [ + 55, + 148, + 295, + 171 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 148, + 295, + 171 + ], + "type": "inline_equation", + "content": "r = 100" + }, + { + "bbox": [ + 55, + 148, + 295, + 171 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 148, + 295, + 171 + ], + "type": "inline_equation", + "content": "s = 20" + }, + { + "bbox": [ + 55, + 148, + 295, + 171 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 171, + 296, + 224 + ], + "lines": [ + { + "bbox": [ + 56, + 171, + 296, + 224 + ], + "spans": [ + { + "bbox": [ + 56, + 171, + 296, + 224 + ], + "type": "table", + "html": "
PlatformPSNR/SSIM↑BRA↑PS FiltersPSNR/SSIM↑BRA↑
Facebook48.56/0.970.97Depth Blur25.25/0.890.85
LinkedIn47.55/0.970.96StyleT.25.12/0.840.85
Instagram48.56/0.980.98Super Zoom36.15/0.880.95
WhatsApp42.10/0.960.97JPEG Artifacts31.01/0.850.94
X (Twitter)49.25/1.000.99Colorize23.15/0.820.92
", + "image_path": "a57aba76b9cf700dafb4d665320ab98872b6d12a71aba9b2fc4ac70cb83ee599.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 232, + 295, + 255 + ], + "lines": [ + { + "bbox": [ + 55, + 232, + 295, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 295, + 255 + ], + "type": "text", + "content": "Table 5. Evaluation of SpecGuard's robustness across Photoshop filters and while uploaded on different social media platforms." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 264, + 132, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 264, + 132, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 132, + 276 + ], + "type": "text", + "content": "4.5. Robustness" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 282, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 282, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 295, + 640 + ], + "type": "text", + "content": "We evaluate watermarking robustness by analyzing detection performance against a range of diverse and challenging real-world attacks. Results demonstrate the strong robustness of SpecGuard across various attacks. For example, as presented in Tab. 3, against geometric distortions such as cropping and rotation, SpecGuard achieved an Avg P of 0.998 and 0.687, respectively. Similarly, across the combined distortion-based attacks, SpecGuard achieves an overall Avg P of 0.911 and Avg Q of 0.952, ensuring minimal quality loss while maintaining high detection accuracy. Notably, the high values of Q@0.95P and Q@0.7P indicate that SpecGuard can sustain reliable detection at strict performance thresholds, even under aggressive perturbations. Unlike prior methods that struggle with extreme transformations, SpecGuard shows remarkable robustness against regeneration-based attacks like Rinse-2xDiff [6] (an image is noised then denoised by Stable Diffusion v1.4 two times with strength as a number of timesteps, 20-100) and RegenVAE [6], maintaining high Avg P. Similarly, under adversarial attacks, SpecGuard consistently secures watermark detectability, outperforming existing techniques across all tested scenarios. These results establish SpecGuard as a highly robust watermarking approach capable of preserving image integrity even under severe distortions and adversarial manipulations, ensuring watermark reliability across diverse attack types. More details about how the attacks are performed are provided in supplementary material. Further, our results in Fig. 4 highlight the strong robustness of SpecGuard against various distortion attacks compared to other SOTA watermarking methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "Social Platforms and Photoshop Filters. SpecGuard's robustness when images are shared across social media platforms and subjected to common Photoshop Neural Filters (PNFs) is shown in Tab. 5. SpecGuard consistently maintains high PSNR and SSIM values, with BRA values close to 0.99 on platforms such as X (formally Twitter), Insta" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": "gram, and Facebook. Also, it shown strong resilience to various PNFs, such as Super Zoom and JPEG Artifacts achieving BRA of 0.95 and 0.94. The PSNR, SSIM, and BRA values are expected to decrease with the severity of image manipulation, as increased manipulation leads to loss of image authenticity. For example, as we applied " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": " style transfer the PSNR and BRA decreased to 25.12 and 0.85. Similarly, the depth blur which excessively reduces the image clarity also causes the decrease of BRA to 0.85." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 189, + 408, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 189, + 408, + 201 + ], + "spans": [ + { + "bbox": [ + 313, + 189, + 408, + 201 + ], + "type": "text", + "content": "4.6. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 205, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 555, + 445 + ], + "type": "text", + "content": "We examined the impact of WP at different levels (" + }, + { + "bbox": [ + 313, + 205, + 555, + 445 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 313, + 205, + 555, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 205, + 555, + 445 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 313, + 205, + 555, + 445 + ], + "type": "text", + "content": ") and its combination with SP using " + }, + { + "bbox": [ + 313, + 205, + 555, + 445 + ], + "type": "inline_equation", + "content": "FA" + }, + { + "bbox": [ + 313, + 205, + 555, + 445 + ], + "type": "text", + "content": " in Tab. 4. As observed, the " + }, + { + "bbox": [ + 313, + 205, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\mathrm{WP}(L_1) + \\mathrm{SP}_{FA}" + }, + { + "bbox": [ + 313, + 205, + 555, + 445 + ], + "type": "text", + "content": " configuration achieved the highest PSNR and SSIM values of 42.89 and 0.99, respectively, and BRA of 0.99, indicating improved watermark invisibility and robustness. In contrast, using WP alone at either " + }, + { + "bbox": [ + 313, + 205, + 555, + 445 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 313, + 205, + 555, + 445 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 205, + 555, + 445 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 313, + 205, + 555, + 445 + ], + "type": "text", + "content": " resulted in lower BRA, with values of 0.92 and 0.87, respectively, demonstrating that the combined " + }, + { + "bbox": [ + 313, + 205, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\mathrm{WP} + \\mathrm{SP}_{FA}" + }, + { + "bbox": [ + 313, + 205, + 555, + 445 + ], + "type": "text", + "content": " approach significantly enhances performance. We also evaluated the robustness of SpecGuard under strong adversarial attacks identified in Tab. 4, such as rotation, blur, and geometric transformations. The results indicate that higher levels of attack severity, such as " + }, + { + "bbox": [ + 313, + 205, + 555, + 445 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 313, + 205, + 555, + 445 + ], + "type": "text", + "content": " rotation, lead to a more significant drop in PSNR, SSIM, and BRA, with values dropping to 11.15, 19.31, and 0.65, respectively. Despite this, the model shows relatively high resilience under moderate attack intensities, such as " + }, + { + "bbox": [ + 313, + 205, + 555, + 445 + ], + "type": "inline_equation", + "content": "45^{\\circ}" + }, + { + "bbox": [ + 313, + 205, + 555, + 445 + ], + "type": "text", + "content": " rotation and low levels of blur and geometric distortion, achieving BRA values as high as 0.93 under geometric transformations at the 0.3 thresholds. More ablations are in the supplementary." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 457, + 388, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 388, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 388, + 470 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 478, + 555, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 555, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 555, + 597 + ], + "type": "text", + "content": "We propose SpecGuard, a novel invisible watermarking method that ensures secure and robust information concealment. Unlike traditional approaches, SpecGuard remains highly resilient against diverse distortions, adversarial attacks, and regeneration-based transformations. Experimental results demonstrate its superior bit recovery accuracy of " + }, + { + "bbox": [ + 313, + 478, + 555, + 597 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 313, + 478, + 555, + 597 + ], + "type": "text", + "content": " maintaining high PSNR. By outperforming SOTA watermarking methods in both detection reliability and imperceptibility, SpecGuard establishes a new benchmark for watermarking under real-world constraints." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 609, + 411, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 411, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 411, + 623 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "content": "This work was partly supported by Institute for Information & communication Technology Planning & evaluation (IITP) grants funded by the Korean government MSIT: (RS-2022-II221199, RS-2022-II220688, RS-2019-II190421, RS-2023-00230337, RS-2024-00356293, RS-2024-00437849, RS-2021-II212068, RS-2025-02304983, and RS-2025-02263841)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17991" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "text", + "content": "[1] Edward H Adelson, Eero Simoncelli, and Rajesh Hingorani. Orthogonal pyramid transforms for image coding. In Visual Communications and image processing II, pages 50-58. SPIE, 1987. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "type": "text", + "content": "[2] Mahdi Ahmadi, Alireza Norouzi, Nader Karimi, Shadrokh Samavi, and Ali Emami. Redmark: Framework for residual diffusion watermarking based on deep networks. Expert Systems with Applications, 146:113157, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 180, + 295, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 295, + 223 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 295, + 223 + ], + "type": "text", + "content": "[3] Inzamamul Alam, Md Tanvir Islam, and Simon S Woo. Specxnet: A dual-domain convolutional network for robust deepfake detection. In Proceedings of the 33nd ACM International Conference on Multimedia, 2025. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 224, + 295, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 224, + 295, + 267 + ], + "spans": [ + { + "bbox": [ + 62, + 224, + 295, + 267 + ], + "type": "text", + "content": "[4] Inzamamul Alam, Md Tanvir Islam, and Simon S Woo. Saliency-aware diffusion reconstruction for effective invisible watermark removal. In Companion Proceedings of the ACM on Web Conference 2025, pages 849-853, 2025. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 270, + 296, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 296, + 312 + ], + "type": "text", + "content": "[5] Aashutosh AV, Srijan Das, Abhijit Das, et al. Latent flow diffusion for deepfake video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3781-3790, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 314, + 294, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 314, + 294, + 356 + ], + "spans": [ + { + "bbox": [ + 62, + 314, + 294, + 356 + ], + "type": "text", + "content": "[6] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436, 2018. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 358, + 294, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 358, + 294, + 401 + ], + "spans": [ + { + "bbox": [ + 62, + 358, + 294, + 401 + ], + "type": "text", + "content": "[7] Tali Dekel, Michael Rubinstein, Ce Liu, and William T Freeman. On the effectiveness of visible watermarks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2146-2154, 2017. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 403, + 294, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 294, + 467 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 294, + 467 + ], + "type": "text", + "content": "[8] Mucong Ding, Tahseen Rabbani, Bang An, Aakriti Agrawal, Yuancheng Xu, Chenghao Deng, Sicheng Zhu, Abdirisak Mohamed, Yuxin Wen, Tom Goldstein, et al. Waves: Benchmarking the robustness of image watermarks. In ICLR 2024 Workshop on Reliable and Responsible Foundation Models, 2024. 1, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 469, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 469, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 469, + 294, + 491 + ], + "type": "text", + "content": "[9] Hubert Etienne. The future of online trust (and why deepfake is advancing it). AI and Ethics, 1(4):553-562, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 492, + 294, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 492, + 294, + 545 + ], + "spans": [ + { + "bbox": [ + 57, + 492, + 294, + 545 + ], + "type": "text", + "content": "[10] Han Fang, Yupeng Qiu, Kejiang Chen, Jiyi Zhang, Weiming Zhang, and Ee-Chien Chang. Flow-based robust watermarking with invertible noise layer for black-box distortions. In Proceedings of the AAAI conference on artificial intelligence, pages 5054-5061, 2023. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 547, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 294, + 600 + ], + "type": "text", + "content": "[11] Jianwei Fei, Zhihua Xia, Benedetta Tondi, and Mauro Barni. Supervised gan watermarking for intellectual property protection. In 2022 IEEE International Workshop on Information Forensics and Security (WIFS), pages 1-6. IEEE, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 602, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 294, + 644 + ], + "type": "text", + "content": "[12] S. A. Fulling. The local geometric asymptotics of continuum eigenfunction expansions. ii. one-dimensional systems. SIAM Journal on Mathematical Analysis, 14(4):605-623, 1983. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "type": "text", + "content": "[13] Mahdieh Ghazvini, Elham Mohamadi Hachrood, and Mojdeh Mirzadi. An improved image watermarking method in frequency domain. Journal of Applied Security Research, 12 (2):260-275, 2017. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "text", + "content": "[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "text", + "content": "Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 97, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 140 + ], + "type": "text", + "content": "[15] Md Tanvir Islam, Ik Hyun Lee, Ahmed Ibrahim Alzahrani, and Khan Muhammad. Mexic: A meta ensemble explainable approach for ai-synthesized fake image classification. *Alexandria Engineering Journal*, 116:351–363, 2025. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 184 + ], + "type": "text", + "content": "[16] K Jayashre and M Amsaprabhaa. Safeguarding media integrity: A hybrid optimized deep feature fusion based deepfake detection in videos. Computers & Security, 142:103860, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 186, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 241 + ], + "type": "text", + "content": "[17] Zhaoyang Jia, Han Fang, and Weiming Zhang. Mbrs: Enhancing robustness of dnn-based watermarking by minibatch of real and simulatedJPEG compression. In Proceedings of the 29th ACM international conference on multimedia, pages 41-49, 2021. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 243, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 243, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 243, + 553, + 274 + ], + "type": "text", + "content": "[18] Hoon Kang and Joonsoo Ha. Projection spectral analysis. International Journal of Control, Automation and Systems, 13(6):1530-1537, 2015. 1, 2, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 277, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 277, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 277, + 553, + 319 + ], + "type": "text", + "content": "[19] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations (ICLR), 2018. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 322, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 322, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 317, + 322, + 553, + 364 + ], + "type": "text", + "content": "[20] SS Kelkar, LL Grigsby, and J Langsner. An extension of parseval's theorem and its use in calculating transient energy in the frequency domain. IEEE Transactions on Industrial Electronics, (1):42-45, 1983. 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 366, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 366, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 317, + 366, + 553, + 409 + ], + "type": "text", + "content": "[21] Varsha Kishore, Xiangyu Chen, Yan Wang, Boyi Li, and Kilian Q Weinberger. Fixed neural network steganography: Train the images, not the network. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 411, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 411, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 317, + 411, + 553, + 454 + ], + "type": "text", + "content": "[22] Jae-Eun Lee, Young-Ho Seo, and Dong-Wook Kim. Convolutional neural network-based digital image watermarking adaptive to the resolution of image and watermark. Applied Sciences, 10(19):6854, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 456, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 553, + 510 + ], + "type": "text", + "content": "[23] Yicheng Leng, Chaowei Fang, Gen Li, Yixiang Fang, and Guanbin Li. Removing interference and recovering content imaginatively for visible watermark removal. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2983-2990, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 513, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 513, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 513, + 553, + 555 + ], + "type": "text", + "content": "[24] Xiao Li, Liquan Chen, Ju Jia, Zhongyuan Qin, and Zhangjie Fu. A lightweight image forgery prevention scheme for IoT using gan-based steganography. IEEE Transactions on Industrial Informatics, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 558, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 558, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 558, + 553, + 600 + ], + "type": "text", + "content": "[25] Dongdong Lin, Benedetta Tondi, Bin Li, and Mauro Barni. A cyclegan watermarking method for ownership verification. IEEE Transactions on Dependable and Secure Computing, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 602, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 602, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 602, + 553, + 668 + ], + "type": "text", + "content": "[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "type": "text", + "content": "[27] Yang Liu, Zhen Zhu, and Xiang Bai. Wdnet: Watermark-decomposition network for visible watermark removal. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 3685-3693, 2021. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17992" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[28] Xiyang Luo, Ruohan Zhan, Huiwen Chang, Feng Yang, and Peyman Milanfar. Distortion agnostic deep watermarking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13548-13557, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 182 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 182 + ], + "type": "text", + "content": "[29] Thanh Thi Nguyen, Quoc Viet Hung Nguyen, Dung Tien Nguyen, Duc Thanh Nguyen, Thien Huynh-The, Saeid Nahavandi, Thanh Tam Nguyen, Quoc-Viet Pham, and Cuong M Nguyen. Deep learning for deepfakes creation and detection: A survey. Computer Vision and Image Understanding, 223:103525, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "type": "text", + "content": "[30] Guangyu Nie, Changhoon Kim, Yezhou Yang, and Yi Ren. Attributing image generative models using latent fingerprints. In International Conference on Machine Learning, pages 26150-26165. PMLR, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 294, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 271 + ], + "type": "text", + "content": "[31] Li Niu, Xing Zhao, Bo Zhang, and Liqing Zhang. Fine-grained visible watermark removal. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12770-12779, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 273, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 273, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 273, + 294, + 316 + ], + "type": "text", + "content": "[32] Konstantin A Pantserev. The malicious use of ai-based deepfake technology as the new threat to psychological security and political stability. Cyber defence in the age of AI, smart societies and augmented humanity, pages 37-55, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 317, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 317, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 294, + 337 + ], + "type": "text", + "content": "[33] Ram Shankar Pathak. The wavelet transform. Springer Science & Business Media, 2009. 1, 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "type": "text", + "content": "[34] Fernandez Pierre, Alexandre Sablayrolles, Teddy Furon, Hervé Jégou, and Matthijs Douze. Watermarking images in self-supervised latent spaces. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3054-3058. IEEE, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 394, + 294, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 294, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 294, + 448 + ], + "type": "text", + "content": "[35] Fernandez Pierre, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22466-22477, 2023. 1, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "type": "text", + "content": "[36] Tong Qiao, Yuyan Ma, Ning Zheng, Hanzhou Wu, Yanli Chen, Ming Xu, and Xiangyang Luo. A novel model watermarking for protecting generative adversarial network. Computers & Security, 127:103102, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "type": "text", + "content": "[37] Sunpreet Sharma, Ju Jia Zou, Gu Fang, Pancham Shukla, and Weidong Cai. A review of image watermarking for identity protection and verification. Multimedia Tools and Applications, 83(11):31829-31891, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 537, + 294, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 580 + ], + "type": "text", + "content": "[38] Qingtang Su, Huanying Wang, DeCheng Liu, Zihan Yuan, and Xueting Zhang. A combined domain watermarking algorithm of color image. Multimedia Tools and Applications, 79(39):30023-30043, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "type": "text", + "content": "[39] Qingtang Su, Xueting Zhang, and Huanying Wang. A blind color image watermarking algorithm combined spatial domain and svd. International Journal of Intelligent Systems, 37(8):4747-4771, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 626, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 626, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 626, + 294, + 669 + ], + "type": "text", + "content": "[40] Ruizhou Sun, Yukun Su, and Qingyao Wu. Denet: disentangled embedding network for visible watermark removal. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2411-2419, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[41] Matthew Tancik, Ben Mildenhall, and Ren Ng. Stegastamp: Invisible hyperlinks in physical photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2117-2126, 2020. 1, 6, 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 654 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[42] D Vaishnavi and TS Subashini. Robust and invisible image watermarking in rgb color space using svd. Procedia Computer Science, 46:1770-1777, 2015. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 108, + 555, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 555, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 555, + 161 + ], + "type": "text", + "content": "[43] Guanjie Wang, Zehua Ma, Chang Liu, Xi Yang, Han Fang, Weiming Zhang, and Nenghai Yu. Must: Robust image watermarking for multi-source tracing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5364-5371, 2024. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 163, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 205 + ], + "type": "text", + "content": "[44] Huanying Wang and Qingtang Su. A color image watermarking method combined qr decomposition and spatial domain. Multimedia Tools and Applications, 81(26):37895-37916, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 208, + 553, + 251 + ], + "type": "text", + "content": "[45] Zilan Wang, Junfeng Guo, Jiacheng Zhu, Yiming Li, Heng Huang, Muhao Chen, and Zhengzhong Tu. Sleepermark: Towards robust watermark against fine-tuning text-to-image diffusion models. arXiv preprint arXiv:2412.04852, 2024. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 254, + 553, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 553, + 305 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 553, + 305 + ], + "type": "text", + "content": "[46] Zijie J Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. Diffusional: A large-scale prompt gallery dataset for text-to-image generative models. arXiv preprint arXiv:2210.14896, 2022.5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 308, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 363 + ], + "type": "text", + "content": "[47] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-rings watermarks: Invisible fingerprints for diffusion images. In Advances in Neural Information Processing Systems, pages 58047-58063. Curran Associates, Inc., 2023. 1, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 365, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 407 + ], + "type": "text", + "content": "[48] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-rings watermarks: Invisible fingerprints for diffusion images. Advances in Neural Information Processing Systems, 36, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 409, + 553, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 409, + 553, + 474 + ], + "spans": [ + { + "bbox": [ + 316, + 409, + 553, + 474 + ], + "type": "text", + "content": "[49] Zijin Yang, Kai Zeng, Kejiang Chen, Han Fang, Weiming Zhang, and Nenghai Yu. Gaussian shading: Provable performance-lossless image watermarking for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12162-12171, 2024. 1, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 476, + 553, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 476, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 316, + 476, + 553, + 518 + ], + "type": "text", + "content": "[50] Zihan Yuan, Qingtang Su, Decheng Liu, and Xueting Zhang. A blind image watermarking scheme combining spatial domain and frequency domain. The visual computer, 37:1867-1881, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 520, + 553, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 520, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 316, + 520, + 553, + 564 + ], + "type": "text", + "content": "[51] Chaoning Zhang, Philipp Benz, Adil Karjauv, Geng Sun, and In So Kweon. Udh: Universal deep hiding for steganography, watermarking, and light field messaging. Advances in Neural Information Processing Systems, 33:10223-10234, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 566, + 553, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 566, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 317, + 566, + 553, + 620 + ], + "type": "text", + "content": "[52] Xuanyu Zhang, Runyi Li, Jiwen Yu, Youmin Xu, Weiqi Li, and Jian Zhang. Editguard: Versatile image watermarking for tamper localization and copyright protection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11964-11974, 2024. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 622, + 553, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 622, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 316, + 622, + 553, + 654 + ], + "type": "text", + "content": "[53] Jiren Zhu, Russell Kaplan, Justin Johnson, and Li Fei-Fei. Hidden: Hiding data with deep networks. European Conference on Computer Vision (ECCV), 2018. 1, 2, 6" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "17993" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Spectral Image Tokenizer/2b3586c0-b936-4d11-a38a-2ff8e5859a65_content_list.json b/2025/Spectral Image Tokenizer/2b3586c0-b936-4d11-a38a-2ff8e5859a65_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3dc405786981c77b8dde71ec81cc429744623070 --- /dev/null +++ b/2025/Spectral Image Tokenizer/2b3586c0-b936-4d11-a38a-2ff8e5859a65_content_list.json @@ -0,0 +1,2064 @@ +[ + { + "type": "text", + "text": "Spectral Image Tokenizer", + "text_level": 1, + "bbox": [ + 367, + 130, + 630, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Carlos Esteves", + "bbox": [ + 163, + 189, + 282, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mohammed Suhail", + "bbox": [ + 408, + 189, + 560, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ameesh Makadia", + "bbox": [ + 687, + 189, + 828, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{machc, suhailmhd, makadia}@google.com", + "bbox": [ + 326, + 213, + 643, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Google Research", + "bbox": [ + 416, + 231, + 553, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 306, + 326, + 321 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image tokenizers map images to sequences of discrete tokens, and are a crucial component of autoregressive transformer-based image generation. The tokens are typically associated with spatial locations in the input image, arranged in raster scan order, which is not ideal for autoregressive modeling. In this paper, we propose to tokenize the image spectrum instead, obtained from a discrete wavelet transform (DWT), such that the sequence of tokens represents the image in a coarse-to-fine fashion. Our tokenizer brings several advantages: 1) it leverages that natural images are more compressible at high frequencies, 2) it can take and reconstruct images of different resolutions without retraining, 3) it improves the conditioning for next-token prediction – instead of conditioning on a partial line-by-line reconstruction of the image, it takes a coarse reconstruction of the full image, 4) it enables partial decoding where the first few generated tokens can reconstruct a coarse version of the image, 5) it enables autoregressive models to be used for image upsampling. We evaluate the tokenizer reconstruction metrics as well as multiscale image generation, text-guided image upsampling and editing.", + "bbox": [ + 88, + 339, + 485, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 690, + 220, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In natural language processing, tokenization associates sets of characters to entries in a vocabulary, which is straightforward since language is discrete. The sequence of tokens is suitably modeled as a categorical distribution with autoregressive (AR) transformers, which is the foundation of modern large language models (LLMs) [33, 46, 47].", + "bbox": [ + 89, + 717, + 483, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While natural images are represented by discrete pixel values, they exhibit high dimensionality, redundancies, and noise so it's impractical to associate one token per pixel. This motivated a long line of research of learnable image tokenizers [15, 37, 37, 51, 57]. While there are successful autoregressive image generation models [15, 57, 58],", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "images are not sequential like language, which motivated developing alternatives such as denoising diffusion models [11, 38, 43] and masked transformers [3, 4, 60]. Nevertheless, most of these alternatives also operate on the latent space of tokenizers like VQGAN [15] instead of raw pixels.", + "bbox": [ + 511, + 306, + 906, + 383 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we revisit AR transformer-based image generation. Our main contribution is a tokenizer operating on the image spectrum, specifically on DWT coefficients, where the coarse-to-fine representation lends itself more naturally to a sequential interpretation. Our Spectral Image Tokenizer (SIT) has several useful properties and enables different applications:", + "bbox": [ + 511, + 383, + 906, + 489 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "P.1 Since the power spectrum of natural images decreases with frequency, high frequencies can be more heavily compressed with little effect in visual quality. SIT leverages this by associating tokens to larger patches at higher wavelet scales than at lower scales (see Fig. 2).", + "P.2 SIT is transformer-based [52]; by using an attention mask where each scale depends on itself and lower scales (\"Scale-Causal attention\"), SIT can be trained at a single resolution and used to tokenize images of multiple resolutions (any number of scales up to the trained maximum), and detokenize partial token sequences (up to some scale), reconstructing a coarse image.", + "P.3 Using SIT, we train an autoregressive generative transformer (AR-SIT) that models images coarse-to-fine. The next-token prediction is then conditioned on a coarse reconstruction of the image given by the partial token sequence, instead of the usual conditioning on the partial reconstruction of the previous rows of the image.", + "P.4 AR-SIT can quickly generate only the first few tokens and reconstruct a coarse version of the image, enabling applications like quickly showing multiple coarse generations and letting the user select which ones to refine.", + "P.5 AR-SIT can be used for text-based upsampling of an input low resolution image, by starting the decoding process with the few tokens output by SIT, and generating the rest of the sequence up to a desired resolution.", + "P.6 AR-SIT can be used for text-guided image editing, by" + ], + "bbox": [ + 514, + 493, + 905, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17181", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2bfa86016e0efaaf3b1d858e61d41b14be8218e91de76c2783d796f1bade2604.jpg", + "image_caption": [ + "Figure 1. Left: we introduce a Spectral Image Tokenizer (SIT), that learns to encode and decode discrete wavelet transform (DWT) coefficients to and from a small set of discrete tokens, such that the sequence represents the image in a coarse-fine fashion. SIT is naturally multiscale and enables coarse-to-fine autoregressive image generation with our AR-SIT model. SIT also leverages the sparsity of high frequency coefficients in natural images. Right: details of the encoder/decoder transformer layers. The main architectural difference with respect to previous tokenizers is that the distributions of DWT approximation and details coefficients are distinct, hence we use specialized parameters for each in the quantizer codebooks and inner transformer layers. We also introduce a scale-causal attention where each token attends to its own scale and lower scales, which enables encoding, decoding, generating, and upsampling images at different resolutions." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 380, + 309 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/db8c6463a1c65f8aaf7b1abc40644785cb3a11c1ae22bbb2fa368a367c846a68.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 85, + 908, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "encoding a given image up to a coarse scale, and generating the finer details conditioned on a new caption.", + "bbox": [ + 116, + 428, + 482, + 458 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Currently, image generation is dominated by diffusion models such as Imagen 3 [23], DALL-E-3 [1] and Stable Diffusion 3 [16]. On the other hand, LLMs such as Gemini [46], GPT-4 [33], and Llama 3 [47] are based on autoregressive transformers. We believe autoregressive image generation is still worth pursuing because it might benefit from advances in LLMs, and multimodal applications might benefit of having a single architecture for all modalities. Recent work on image and video generation support this point [45, 48, 59]. Dieleman [12] recently interpreted denoising diffusion models as spectral autoregression, since, when looking at image spectra, the denoising procedure uncovers frequencies from low to high. In contrast, our method does literal spectral autoregression.", + "bbox": [ + 89, + 462, + 483, + 675 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 89, + 691, + 227, + 707 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image tokenization Several methods have been developed to map images to a small set of discrete tokens suitable for generative modeling. VQ-VAE [51] introduced vector-quantization in the latent space of a Variational AutoEncoder to map images, audio and video to a set of discrete values. VQGAN [15] improved upon VQVAE by using perceptual and adversarial losses. We build on ViT-VQGAN [57], which improved upon VQGAN by using a Vision Transformer (ViT) [13] instead of convolutions, as well as codebook factorization and feature normalization.", + "bbox": [ + 89, + 718, + 482, + 868 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we are interested in multiscale image representations. VQ-VAE-2 [37] introduced multiscale latents", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "by keeping and quantizing intermediate downsampled convolutional features. RQ-VAE [26] quantized a set of residuals such that the latent vector is represented in a coarse-to-fine fashion and reconstructed by adding the code embeddings for each residual. Similarly, VAR [48] and STAR [28] split the latent space in multi-scale quantized residuals, so that it can be reconstructed by upsampling and summing. FAR [56] splits the latents into frequencies to train a next-frequency generative model. DQ-VAE [22] encoded smooth regions with fewer tokens that are generated first, while QG-VAE [14] and SEMANTICIST [53] learn non-local tokens that are ordered by importance.", + "bbox": [ + 511, + 426, + 906, + 608 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The crucial difference between our approach and the aforementioned is that we operate on spectral coefficients of the input and not on latent features. This allows our tokenizer to be truly multiscale; it can take inputs at different scales and reconstruct up to some scale, and leverage that higher frequencies are more compressible in natural images.", + "bbox": [ + 511, + 609, + 906, + 699 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Tangentially related to our work, Wave-ViT [55] modified the ViT self-attention by applying a DWT to the input and concatenating coefficients, effectively exchanging space for depth to reduce the sequence length. Zhu and Soricut [62] modified the ViT patchifier in a similar way, and introduced patch embeddings that leverage the sparsity of high frequency coefficients. Both methods are for discriminative tasks such as image classification and segmentation, while we focus on generation.", + "bbox": [ + 511, + 700, + 905, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Autoregressive image generation Early approaches PixelRNN and PixelCNN [49, 50] generate images pixel by pixel by modeling the conditional distribution of the pixel given the previous pixels with recurrent layers of causal", + "bbox": [ + 511, + 840, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17182", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "convolutions. PixelSNAIL [5] improved on this model by introducing self-attention layer to better model long-range dependencies. VQ-VAE [51] introduced a two-stage approach with a tokenizer and a separate stage to model the distribution of tokens. VQGAN [15] greatly improved these results by both improving the tokenizer and using a transformer to model the distribution. Finally, ViT-VQGAN [57] proposed a transformer-based tokenizer, which Parti [58] used with a large autoregressive transformer capable of high-quality text-to-image generation.", + "bbox": [ + 89, + 90, + 483, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MaskGIT [3] and Muse [4] highlighted the disadvantages of the typical autoregressive models raster-order conditioning, and proposed to generate all tokens in parallel iteratively, where each iteration keeps the highest confidence tokens. We address the same problem with a tokenizer whose sequence represents the image in a coarse-to-fine order instead of raster-order.", + "bbox": [ + 89, + 242, + 483, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multiscale image generation Multiscale image generation ideas have appeared in the context of VAEs [2], GANs [24, 39], and diffusion models [18, 25, 40], but have not been sufficiently explored with AR transformers.", + "bbox": [ + 89, + 349, + 483, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Nash et al. [32] represented an image as a sequence of quantized and thresholded DCT coefficients, where the compression comes from the fact that many coefficients are zero and are omitted. By sorting the sequence by frequency, the method can model images from coarse-to-fine like we do. However, the compressed representation is handcrafted and results in long sequences. In a similar vein, Mattar et al. [30] handcrafted a tokenizer based on DWT coefficients, introducing tokens to represent large chunks of zeros. It also results in long sequences and is only applied to generations of small grayscale images. In contrast with these methods, instead of handcrafting a compressed input representation, our tokenizer learns to encode to, and decode from, a short compressed coarse-to-fine sequence.", + "bbox": [ + 89, + 410, + 483, + 621 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Background", + "text_level": 1, + "bbox": [ + 89, + 635, + 218, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A discrete wavelet transform (DWT) is based on successive convolutions of the signal $f$ with a pair of lowpass $g$ and highpass $h$ filters with the first step as follows", + "bbox": [ + 89, + 660, + 483, + 705 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\operatorname {l o w} _ {1}} [ n ] = f \\star g = \\sum_ {k} f [ k ] g [ n - k ], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 717, + 480, + 750 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\operatorname {h i g h} _ {1}} [ n ] = f \\star h = \\sum_ {k} f [ k ] h [ n - k ]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 752, + 480, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The high and low outputs are subsampled by a factor of two, and the operation is repeated for the low channel, such that at level $L$ we compute $f_{\\mathrm{low}_L} = f_{\\mathrm{low}_{L - 1}} \\star g$ and $f_{\\mathrm{high}_L} = f_{\\mathrm{low}_{L - 1}} \\star h$ , subsample again and drop $f_{\\mathrm{low}_{L - 1}}$ . The output at level $L$ comprises the approximation coefficients $f_{\\mathrm{low}_L}$ and the detail coefficients $\\{f_{\\mathrm{high}_k}\\}$ for $1 \\leq k \\leq L$ . This output has the same cardinality as the input and the", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/863d7df0e87b9ef660d64f4c69a21c033d45e2e444801ba5802f9b521fe6710d.jpg", + "image_caption": [ + "Figure 2. Input patchification. Left: typical patchification for Vision Transformers (ViT) [13], where the image is split in equal-sized patches. Right: we propose to patchify the coefficients of a discrete wavelet transform (DWT) instead. Each scale is shown in a different color. Scales other than the lowest contain three blocks representing horizontal, vertical and diagonal details; we concatenate the spatially correspondent patches of each block such that each scale is represented by the same sequence length. This results in larger patch sizes for higher scales, which are more compressible. The figure shows 3 scales and 16 tokens per scale; in our experiments we use 4 or 5 scales and 256 tokens per scale." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 709, + 238 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3d53cd99caa83e3e495af4cbd9515dbd49fa7a9103b690f6877a7ea3f9b4392f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 90, + 901, + 237 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "transformation is invertible, where the forward transform is typically called \"analysis\" and the backward \"synthesis\". The simplest wavelet family is the Haar, where $g = [1,1]^{\\top}$ and $h = [1, -1]^{\\top}$ (optionally scaled to unit norm).", + "bbox": [ + 511, + 416, + 905, + 476 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For image analysis we use a 2D DWT, which is obtained by simply convolving the rows and columns with $g$ and $h$ . The approximation coefficients $f_{\\mathrm{low}_1} = f \\star (gg^\\top)$ , and the details are divided into horizontal $f_{\\mathrm{H}_1} = f \\star (gh^\\top)$ , vertical $f_{\\mathrm{V}_1} = f \\star (hg^\\top)$ and diagonal $f_{\\mathrm{D}_1} = f \\star (hh^\\top)$ details. Subsequent levels apply the same operations to the approximation coefficients. Fig. 2 shows a two-level transform, where we can see that the approximation coefficients correspond to a coarse version of the image.", + "bbox": [ + 511, + 477, + 905, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We refer to the textbooks by Mallat [29] and Daubechies [9] for more information about wavelets.", + "bbox": [ + 511, + 613, + 905, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 513, + 659, + 604, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our main contribution, SIT, is an image tokenizer that takes discrete wavelet transform (DWT) coefficients. The model follows ViT-VQGAN [57], with important changes that we describe in this section and visualize in Fig. 1. For image generation, we introduce AR-SIT which is based on Parti [58] with minor changes, using SIT as the tokenizer.", + "bbox": [ + 511, + 685, + 905, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Tokenizer", + "text_level": 1, + "bbox": [ + 511, + 787, + 625, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Patchification and featurization The first step is to map the input to a sequence of patches. We apply the Haar DWT on the input image and patchify each scale separately. While Haar is the simplest wavelet and lacks properties found in other wavelets useful for compression, we found no benefits of using other wavelet families such as CDF [8].", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17183", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our design choice is to use the same number of patches for each scale. In a DWT, the higher scales correspond to high frequency details which are represented by more coefficients than the lower scales, but contribute less to the spatial pixel values. In other words, in natural images, most of the power spectrum is concentrated on lower frequencies. By representing each scale with the same number of tokens, we are compressing more the higher frequencies (since they are represented by more coefficients), similarly to what is done in image compression methods such as JPEG2000 [7].", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The approximation (or lowpass) DWT coefficients correspond to a coarse version of the input image, which we consider the first scale. The following scales are divided in three blocks corresponding to horizontal, vertical, and diagonal details, where each coefficient relates to a specific spatial location. Thus, we can concatenate the three blocks such that each entry corresponds to the same spatial location. For example, the first scale will typically be split in patches that are $32 \\times 32 \\times 3$ , with channels corresponding to RGB, the second scale will be $32 \\times 32 \\times 9$ , where the channels correspond to the RGB of horizontal, vertical, and diagonal details, the third will be $64 \\times 64 \\times 9$ and so on. Fig. 2 shows an example of our patchification scheme.", + "bbox": [ + 89, + 242, + 482, + 438 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since patches have different resolutions (higher scales will have larger patches), the usual ViT linear embedding to map patches to features cannot be shared across all patches so we have different parameters per scale. Formally, given an image $f$ , our patchifiers compute", + "bbox": [ + 89, + 439, + 482, + 515 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\text {l o w} L}, \\left\\{f _ {\\mathrm {H} _ {i}} \\right\\} _ {i \\leq L}, \\left\\{f _ {\\mathrm {V} _ {i}} \\right\\} _ {i \\leq L}, \\left\\{f _ {\\mathrm {D} _ {i}} \\right\\} _ {i \\leq L} = \\mathrm {D W T} (f), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 527, + 482, + 545 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nc _ {1} = P _ {0} \\left(f _ {\\text {l o w} L}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 547, + 480, + 563 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nc _ {s} = P _ {s} \\left(f _ {\\mathrm {H} _ {L - s + 2}}, f _ {\\mathrm {V} _ {L - s + 2}}, f _ {\\mathrm {D} _ {L - s + 2}}\\right), 1 < s \\leq L, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 566, + 480, + 582 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c_{s} = \\{c_{s}^{n}\\}_{1\\leq n\\leq N}$ is the sequence of token embeddings at scale $s$ , and $c_{s}^{n}\\in \\mathbb{R}^{C}$ is the embedding of the $n$ -th token at the $s$ -th scale, while $L$ is the number of DWT levels, $S = L + 1$ is the number of scales and $N$ the number of tokens per scale. For brevity, when there is no ambiguity, we may omit the set indexing and use $\\{c_s\\}$ to denote the set of tokens of all scales, for example.", + "bbox": [ + 89, + 594, + 482, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The final projection after the decoder needs to map the features back to different sized patches, so it also has different parameters per scale. Those patches still represent DWT coefficients so an inverse DWT (IDWT) is finally applied to obtain an image output.", + "bbox": [ + 89, + 700, + 482, + 776 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Flexible sequence length The tokenizer encoder and decoder transformers operate on the sequence of patch features of length $SN$ . The sequence length is a major factor in resource utilization so we want to keep it constrained. Our method is flexible since we can choose the number of scales and the patch size per scale, while most ViT-based models such as ViT-VQGAN [57] are more restricted. They use the same patch size for the whole image; thus, keeping the", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "same patch size and doubling each image dimension would increase the sequence length by a factor of 4, where our method is capable of including additional scales which only increases the sequence length by multiples of $N$ .", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For example, for a $256 \\times 256$ input, ViT-VQGAN uses $8 \\times 8$ patches to obtain a sequence of 1024 elements. Our SIT can use 4 scales and 256 tokens per scale for the same sequence length. When increasing the resolution to $512 \\times 512$ , the baseline can either increase the patch size to $16 \\times 16$ , resulting in the same sequence length, or keep it $8 \\times 8$ , resulting in a $4 \\times$ longer sequence. SIT, for example, can vary the number of scales from 4 to 6, resulting in sequence lengths of 1024, 1280, 1536 which are more manageable.", + "bbox": [ + 511, + 152, + 903, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Transformers After featurization, the single sequence containing all scales passes through a transformer encoder, followed by quantization and a transformer decoder. We propose two optional modifications to the transformers, which are otherwise identical to the ones used in ViT [13].", + "bbox": [ + 511, + 294, + 903, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, we propose a scale-causal attention mask, where an element at some scale attends to all elements of its own scale and lower scales, represented by a block-triangular mask pattern. With dense attention, we write the application of the encoding transformer $T_{\\mathrm{enc}}$ as $\\{z_s\\} = T_{\\mathrm{enc}}(\\{c_s\\})$ , and for scale-causal we write $z_s = T_{\\mathrm{enc}}(\\{c_k\\}_{1 \\leq k \\leq s})$ for each $s$ . The scale-causal attention can be applied independently to the encoder and decoder, enabling different applications. For the multiscale reconstruction experiments in Sec. 5.1, we need to both encode and decode multiple resolutions, so both encoder and decoder use scale-causal masks. For coarse-to-fine image generation in Sec. 5.2, only the decoder needs to be scale-causal in order to decode the partially generated sequence. For the text-guided image upsampling in Sec. 5.3, only the encoder needs to be scale-causal to encode the lower resolution inputs. For the image editing experiments in Sec. 5.4, the scale-causal encoder prevents information leaking from high to low scale.", + "bbox": [ + 511, + 371, + 903, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Second, we propose to use different transformer parameters for the approximation (first scale) and details coefficients (other scales). In our model, the subsequences corresponding to coefficients of each type of coefficient come from quite distinct distributions, so it makes sense to treat them differently. This contrasts with the spatial representation of images where each patch can be considered as coming from the same distribution. Thus, the parameters of the key, query, and value embeddings, the layer norms, and the MLP on each transformer layer are not shared between the approximation and details coefficients. This Approximation-Details Transformer (ADTransformer) still takes a single sequence composed of all coefficients. We experimented with different transformers per sequence with cross-attention for information sharing, but it performed worse. This change leads to more memory utilization to store the extra parameters, but the training/inference speed", + "bbox": [ + 511, + 643, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17184", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/94af4ecd39a3b10381c810b05a753d77349002dbaf0731b256169d374b3e1895.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LPIPS ↓PSNR ↑L1 ↓FID ↓IS ↑images/s ↑
Resolution: 512 × 512
ViT-VQGAN0.32022.40.0426.92151.5593
SIT-5 (Ours)0.26022.00.0462.65192.0410
SIT-6 (Ours)0.23923.10.0401.74203.7320
Resolution: 256 × 256
ViT-VQGAN (reported)-24.80.0321.99184.4-
ViT-VQGAN (reproduced)0.16725.00.0312.33184.0-
ViT-VQGAN (no LL)0.16323.80.0381.20194.6626
SIT-4 (Ours)0.14424.00.0371.20199.5596
SIT-5 (Ours)0.13524.50.0350.97202.3411
SIT-SC-5 (Ours)0.16124.10.0371.33193.7411
Resolution: 128 × 128
ViT-VQGAN0.18526.30.0303.77117.3626
SIT-SC-5 (ours)0.15927.10.0272.13129.3582
Resolution: 64 × 64
ViT-VQGAN0.12928.80.0233.5321.0627
SIT-SC-5 (ours)0.11131.30.0171.3930.1847
Resolution: 32 × 32
ViT-VQGAN0.21423.30.045-3.7627
SIT-SC-5 (ours)0.02936.80.0100.313.5825
Resolution: 16 × 16
ViT-VQGAN0.12724.90.039-1.7627
SIT-SC-5 (ours)0.01341.30.0060.091.82620
", + "bbox": [ + 93, + 89, + 480, + 404 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "is similar because the number of operations is unchanged.", + "bbox": [ + 89, + 593, + 473, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantizer The encoder outputs a sequence of features that are quantized to a fixed-sized codebook, similarly to ViT-VQGAN [57], VQVAE [51], and VQGAN [15]. We modify the quantizer such that approximation and details have different codebooks for the same reasons discussed previously. Thus, codebook sizes and feature dimensions are the same as the ViT-VQGAN baseline, but we have different features for the same code at different positions. Formally, we apply $q_{s}^{n} = Q_{\\mathrm{approx}}(z_{s}^{n})$ for $s = 1$ and $q_{s}^{n} = Q_{\\mathrm{details}}(z_{s}^{n})$ for $s \\geq 2$ for each pair $(s,n)$ , where $q_{s}^{n}$ is chosen from the codebook for $s$ and can be associated with its discrete position in the codebook, denoted $\\lfloor q_s^n \\rfloor$ .", + "bbox": [ + 89, + 612, + 483, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training We follow the ViT-VQGAN training protocol and use the same weighting for the L2, perceptual, adversarial, and quantization losses. We remove the logit-laplace loss that was shown detrimental in follow-up work [58].", + "bbox": [ + 89, + 794, + 483, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "All losses are applied to the spatial domain images reconstructed by the inverse discrete wavelet transform on the decoder output: $\\hat{f} = \\mathrm{IDWT}(T_{\\mathrm{dec}}(\\{q_s\\}))$ . We noticed insta", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4a11b64ad2dc3a729b5f5fed6063400c38fa36b2e38f86efa3d6d25bf0754280.jpg", + "table_caption": [ + "Table 1. Multiscale reconstruction on ImageNet. \"SC\" denotes scale-causal attention, which slightly reduces performance at the highest resolution but enables multiscale reconstruction without downsampling/upsampling or retraining. The ViT-VQGAN values from Yu et al. [57] used a logit-laplace loss (LL) which was later considered harmful [58], so we retrain without it. Our SIT improves reconstruction metrics, is significantly faster at lower resolutions, and robust when increasing resolution. The ViT-VQGAN baseline suffered from instability during training on $512 \\times 512$ inputs so we selected the best values before divergence. We report test time throughput of an encoding/decoding cycle for the max batch size that fits on a TPU v5e." + ], + "table_footnote": [], + "table_body": "
FID ↓IS ↑images/s ↑images/Gb ↑
Resolution: 256 × 256
Parti350M (reported)14.1---
Parti350M12.436.57.812.0
AR-SIT-SCD-4 (Ours)12.637.36.58.0
Resolution: 128 × 128
Parti350M11.233.57.612.0
AR-SIT-SCD-4 (Ours)11.433.212.612.0
Resolution: 64 × 64
Parti350M10.516.97.612.0
AR-SIT-SCD-4 (Ours)11.418.624.516.0
Resolution: 32 × 32
Parti350M5.82.97.77.7
AR-SIT-SCD-4 (Ours)7.63.274.728.0
", + "bbox": [ + 516, + 90, + 903, + 327 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Coarse-to-fine image generation on MS-COCO [27]. With SIT, the autoregressive generation is stopped early for a coarse version of the image. The Parti350M baseline does not have this property, so we generate at full resolution and downsample for comparison. AR-SIT matches the baseline performance at the training resolution but is several times faster and more memory efficient at lower resolutions, even when trained only on higher resolution data. We report throughput and memory utilization during generation given the max batch size that fits on a TPU v6e.", + "bbox": [ + 511, + 337, + 906, + 464 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "bility during training due to the adversarial loss, which was fixed by applying spectral normalization following Miyato et al. [31], which simply divides the discriminator weight matrices by their largest singular value.", + "bbox": [ + 511, + 474, + 906, + 536 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Autoregressive image generation", + "text_level": 1, + "bbox": [ + 511, + 542, + 799, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use our tokenizer for autoregressive image generation, by training a second stage transformer model similar to Parti [58], with some modifications. Formally, the autoregressive transformer $T$ models categorical distributions over the discrete codes", + "bbox": [ + 511, + 565, + 905, + 640 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP \\left(\\left\\lfloor q _ {s} ^ {n} \\right\\rfloor \\mid \\left\\{\\left\\lfloor q _ {i} \\right\\rfloor \\right\\} _ {1 \\leq i < s} \\cup \\left\\{\\left\\lfloor q _ {s} ^ {i} \\right\\rfloor \\right\\} ^ {1 \\leq i < n}\\right) = T (c) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 648, + 903, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which can be sampled one code at a time for generation. $T$ can be conditioned on a textual description processed by a transformer encoder for text-to-image generation. For training, we model the distribution of input codes as $P(\\{\\lfloor q_s\\rfloor \\}) = \\prod_{1\\leq s\\leq S}^{1\\leq n\\leq N}P(\\lfloor q_s^n\\rfloor)$ and minimize the negative log-likelihood $-\\log P(\\{\\lfloor q_s\\rfloor \\})$ over the training set.", + "bbox": [ + 511, + 672, + 905, + 765 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With different codebooks for approximation and details, the same token id might have a different meaning depending on its position. Thus, the AR model has different token embeddings for approximation and details tokens. The last layer for logit prediction is also different.", + "bbox": [ + 511, + 765, + 905, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the generative applications in Secs. 5.2 and 5.3, we introduce mild changes in order to interrupt the generation after all tokens up to a certain scale are generated, and to start the generation with given tokens up to a certain scale.", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17185", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dc43ffe6e96642fcda4ec765bcd42878e8704676603cf14fde48296916c4cb76.jpg", + "image_caption": [ + "\"a zebra standing on dirty area with trees in the background.\"" + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 488, + 164 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e248b9896775e4fae5bad5b558a325a77445733fe2d81b49fc5fd2019b6f9431.jpg", + "image_caption": [ + "\"an elaborate metal vase holds a decorative bouquet of flowers.\"" + ], + "image_footnote": [], + "bbox": [ + 94, + 176, + 488, + 253 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/de3eb6007c31642354f152e1bd512c4eea36619650b3f8dfad3b853f0ee1e278.jpg", + "image_caption": [ + "\"the wing of a plane located in the sky above a sea and city.\"", + "Figure 3. Coarse-to-fine text-to-image on MS-COCO [27] prompts. Each quadruple shows generations from AR-SIT-SCD for the given prompt, with only the first $25\\%$ , $50\\%$ , $75\\%$ and $100\\%$ of tokens generated, corresponding to resolutions of $32 \\times 32$ to $256 \\times 256$ . Training uses only $256 \\times 256$ images. Our model enables quick generation of coarse image candidates that can be further improved if needed." + ], + "image_footnote": [], + "bbox": [ + 94, + 263, + 488, + 340 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/25862973e1bf70f60bd7575e0f87c543fe55309df8ecab772e174e9998dc63f2.jpg", + "image_caption": [ + "\"a modern bathroom is shown with a square sink.\"" + ], + "image_footnote": [], + "bbox": [ + 506, + 88, + 903, + 165 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f33eabf0788eab36bb4215f991b3e7ffd68e526e0ff44c2644f486c2dba10819.jpg", + "image_caption": [ + "\"a toy boat sits on the ground.\"" + ], + "image_footnote": [], + "bbox": [ + 508, + 176, + 903, + 253 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/710941c16491f7e36dfe7d85aeba52cd9c1e9fd64813e64988728e31bbccaf38.jpg", + "image_caption": [ + "\"This brown and white cat is standing on a desk in front of a computer.\"" + ], + "image_footnote": [], + "bbox": [ + 508, + 263, + 903, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 410, + 222, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We focus our experiments on demonstrating the properties and applications enumerated in P.1-P.6 (Sec. 1). Our ideas build on ViT-based tokenizers, so the most fair comparisons are against ViT tokenizers and AR generative models; namely ViT-VQGAN [57] and Parti [58]. In these comparisons we are able to match the architectures and training protocol exactly. Nevertheless we also compare with a broader set of methods in Sec. 5.5.", + "bbox": [ + 89, + 436, + 483, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Multiscale reconstruction", + "text_level": 1, + "bbox": [ + 89, + 571, + 323, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train our tokenizer on ImageNet [10] and evaluate its reconstruction performance. We follow the ViT-VQGAN [57] \"Small\" architecture and training protocol.", + "bbox": [ + 89, + 594, + 482, + 640 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The model \"SIT-5\" is the Spectral Image Tokenizer with 5 scales from $16 \\times 16$ to $128 \\times 128$ resolutions, where each scale is represented by 256 tokens. The variation \"SIT-SC\" uses scale-causal attention on both the encoder and decoder, which enables handling inputs and outputs of different resolutions, even though it is only trained at $256 \\times 256$ . For example, when the input image is $64 \\times 64$ , only the first three scales are used, resulting in shorter sequence lengths which reduces memory utilization and processing time. Both models employ the ADTransformer described in Sec. 4.1.", + "bbox": [ + 89, + 641, + 482, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The ViT-VQGAN baseline only works for $256 \\times 256$ inputs, so to evaluate against it fairly, we upsampled the low-resolution inputs to that resolution. This brings the input patches away from the training distribution, which might explain the drop in reconstruction quality.", + "bbox": [ + 89, + 792, + 482, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also evaluate at $512 \\times 512$ . The ViT-VQGAN baseline increases its patch size from $8 \\times 8$ to $16 \\times 16$ , keep-", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing the sequence length constant. Interestingly, the baseline suffered heavy instability during training, which was not resolved by reducing the learning rate, using spectral normalization, or the logit-laplace loss. In contrast, our SIT variations trained successfully with no hyperparameter changes.", + "bbox": [ + 511, + 411, + 903, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 1 shows the reconstruction metrics; Fig. 7 shows reconstruction samples at multiple resolutions. Results provide evidence for P.1 and demonstrate P.2 (Sec. 1).", + "bbox": [ + 511, + 488, + 905, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Coarse-to-fine text-to-image generation", + "text_level": 1, + "bbox": [ + 511, + 544, + 852, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We tackle text-to-image generation by using an autoregressive transformer to model the distribution of discrete tokens output by our tokenizer. Since the SIT sequence of tokens represents an image in a coarse-fine fashion, the autoregressive model generation has the same property, which means that we can interrupt the generation after a certain number of tokens and decode a coarse version of the generation.", + "bbox": [ + 511, + 566, + 905, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For this to work, the SIT decoder must be scale-causal, while there is no constraint for the encoder so we use dense attention (we denote this variant Scale-Causal Decoder, or \"SIT-SCD\"). It uses a \"small\" encoder and \"large\" decoder as described in ViT-VQGAN [57], and 4 scales.", + "bbox": [ + 511, + 672, + 905, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We follow the Parti350M [58] architecture and training protocol and name our generative model auto-regressive SIT, or \"AR-SIT\". SIT and AR-SIT are trained on a subset of WebLI [6] of around 128M images, where each image is seen at most once during training. Evaluation is on 30k examples of MS-COCO [27]. We show metrics and generations at different resolutions in Fig. 3 and Tab. 2. AR-SIT matches the baseline performance but is several times faster and more memory efficient at lower resolutions. These results demonstrate the property P.4 (Sec. 1).", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17186", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f6682868634b64cec23c47bc161205192494d27dd60c2e7eb70a3346c8aad53c.jpg", + "image_caption": [ + "\"an assortment of some colorful vases on display on a table\"" + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 222, + 189 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6864d3bfd29c46ef47320f2be74f6fdfb48f5d5df51d90f044cab6eda868f28a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 89, + 354, + 189 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8b439c1c4445e8bb03f450c3747720ece9b8ca68277d231315d84b1d2d995342.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 89, + 486, + 189 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/517ebeae81c58a9d8c5eff6784c15ff05c0996997375cf294898fd3a0f22c118.jpg", + "image_caption": [ + "\"one train riding on the multiple train tracks side by side.\"" + ], + "image_footnote": [], + "bbox": [ + 94, + 205, + 223, + 306 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/83ed571691f19876c737ebdc8743125915e5cd1d2239ea506d6fe5a93bb4bb8f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 205, + 354, + 306 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1a458d4fa8ab0ab3e5ed942ae2f094554c544dd10955bb5af9dcdb9da4d57813.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 205, + 486, + 306 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3dd72cc01a64562a44a1ef23d88bfbc31770459cf258ae88a315876675ba6fcd.jpg", + "image_caption": [ + "\"an image of a fired hydrant on the street\"", + "Figure 4. Text-guided image upsampling on MS-COCO [27]. Our coarse-to-fine generative models can take a low-resolution image, encode it as the first few tokens of a sequence, and generate the rest of sequence, which, when decoded, effectively upsamples the input. Each triplet shows the given $32 \\times 32$ image, our $256 \\times 256$ reconstruction and the ground truth." + ], + "image_footnote": [], + "bbox": [ + 94, + 323, + 223, + 422 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/841d651e744429d21dc2987e5397e2c7a88724fd690efd869b771cbe44f39aa7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 323, + 354, + 422 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ac38bdf79291ffd3ab1d63ed72d49ddcd4bdc91df8eaf45bdd23a839dc5e1b20.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 323, + 486, + 422 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7d60852496ac37397b718e1b30f498f7a70098acac0efb310597ad1dab13ca2a.jpg", + "image_caption": [ + "\"a large elephant standing on top of a grass covered field.\"" + ], + "image_footnote": [], + "bbox": [ + 509, + 88, + 638, + 189 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7147839eb4f3185aab90b428d0b70471f566d9b819fde23f2cde8f026ed5c2d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 89, + 771, + 188 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7ca9cd0198cae27ae5efdec7451fd2e87c474908da76f7885961a5e3d24e50d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 772, + 89, + 903, + 189 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/09890d806c9622da8906b588b07da60bbe2e63f1dbd4ee7306ab2a57116e4b3d.jpg", + "image_caption": [ + "\"three tropical bird perched on top of high bare branches\"" + ], + "image_footnote": [], + "bbox": [ + 509, + 205, + 640, + 306 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/59ac6e58680042f11f5f3a0119b5af6091a722258fd001e25739170221fc02b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 205, + 771, + 306 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3587c8b075c0638cfffa457a21b7a195201df8f66d0c9cf958a65b203f57af7a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 772, + 205, + 903, + 306 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bea17c977521287e89bcf60654bbe228b5d095d91272b5755f17af6822cf67d5.jpg", + "image_caption": [ + "\"a chair with a pink blanket is sitting on a red floor.\"" + ], + "image_footnote": [], + "bbox": [ + 509, + 323, + 638, + 422 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/843e97f846207e282a99007ba5204d3856f4d925c573a79c3727719130dec0b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 323, + 771, + 422 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f367404a2ed76834b55032af67c3781a729a479906a959078f6746c6164bdd31.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 772, + 323, + 903, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Text-guided image upsampling", + "text_level": 1, + "bbox": [ + 89, + 497, + 362, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We leverage the coarse-to-fine nature of SIT to apply a pretrained AR-SIT for text-guided image upsampling, where we are given a low resolution image and a caption.", + "bbox": [ + 89, + 518, + 482, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The idea is to encode the low-resolution image, which will give the first tokens of our high resolution output. AR-SIT then takes these tokens and generates the rest of the sequence. For this to work, the SIT encoder must be scale-causal to properly tokenize low-resolution inputs, while there is no constraint for the decoder so we use dense attention. We denote this variant Scale-Causal Encoder, or \"SIT-SCE\". The model is otherwise as the one used in Sec. 5.2.", + "bbox": [ + 89, + 565, + 482, + 685 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here the inputs are $32 \\times 32$ . Fig. 4 shows text-guided upsampling examples on MS-COCO [27]. We obtain an FID of 6.2 when evaluating these generations, compared to 12.6 when given only the prompts. Fig. 8 and Tab. 4 show additional results upsampling from $16 \\times 16$ .", + "bbox": [ + 89, + 686, + 482, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Text-guided image editing", + "text_level": 1, + "bbox": [ + 89, + 771, + 326, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our coarse-to-fine AR-SIT enables a text-guided image editing application, where we want to change image details while keeping the same overall appearance, which corresponds to freezing lower scales while generating higher.", + "bbox": [ + 89, + 794, + 482, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We apply an AR-SIT trained only on the maximum likelihood objective as follows. The given image is tokenized only up to the first scale, corresponding to a coarse repre", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sensation. The tokenizer encoder must be scale-causal in order to avoid leakage from high to low scales, so we use SIT-SCE here. Now we use AR-SIT to generate the rest of the sequence, conditioning on the textual caption.", + "bbox": [ + 511, + 498, + 906, + 558 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this experiment, we use a 5-scale model such that the lowest resolution is $16 \\times 16$ ; we found that starting with higher resolutions limits the changes the model can generate. Fig. 5 shows some examples of text-guided editing on MS-COCO [27], where we lightly modify the original captions, for example by swapping \"elephants\" with \"cows\". We show additional results in Fig. 6 in the supplemental.", + "bbox": [ + 511, + 560, + 906, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5. Class-conditional image generation", + "text_level": 1, + "bbox": [ + 511, + 679, + 818, + 695 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here we evaluate class-conditional generation on $256 \\times 256$ ImageNet. We use the SIT-4 tokenizer from Tab. 1 and follow ViT-VQGAN's [57] VIM-Base AR architecture and training protocol. Tab. 3 shows we outperform the baseline in this fair comparison which is evidence for the improved conditioning described in P.3. Fig. 9 shows samples.", + "bbox": [ + 511, + 702, + 906, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "AR-SIT-4* departs from ViT-VQGAN and Parti hyperparameters to compare with more recent works, by increasing the perceptual and adversarial loss weights, doubling the tokenizer codebook, reducing its latent dimension, training with a constant learning rate, including 2D RoPE [44] and GeGLU [42], in both the tokenizer and generator. We train the AR model longer and since there is no text encoder, we", + "bbox": [ + 511, + 795, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17187", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/36ecb5f118fe41bcc2c59633b2689ae5406defc5a4d24fa7a3a57028d53d54c0.jpg", + "image_caption": [ + "\"a close-up of a dog face.\"" + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 223, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eaa16f39af62397d6cb9924f8124b28531b5cf0c7837c487bb5b87773e0227e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 89, + 354, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8b7a5d722eedde08ddd90ded0d73e568da830eff1aa825e3f59c8feafb447751.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 89, + 486, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e1bdf2e9ecfc4fc44af3a82b3207598dc95b2058bc453fd53d75c2e575b26e44.jpg", + "image_caption": [ + "\"a couple of cows are standing in a field\"" + ], + "image_footnote": [], + "bbox": [ + 94, + 210, + 223, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4b686611eccd2b3eb076ee9970ff9de5fa2e2ee260cf20e00d6c7b996ea6c511.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 210, + 354, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fba81ff6c5e4cb8a1d079101361afbc84072e7398dc85afe295e398c9d708fd4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 210, + 486, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/17dafd257af776ffd48630fd378c65e9aa0eae746acc11a8c705d00d85eeb144.jpg", + "image_caption": [ + "\"a robot on a grass covered field.\"", + "Figure 5. Text-guided image editing on MS-COCO [27]. Our coarse-to-fine generative models can do text-guided editing by encoding a given image but keeping only the lower scales, and using a pre-trained AR-SIT to re-generate the higher scales conditioned on the textual prompt. Each triplet shows the given image, its $16 \\times 16$ reconstruction using only the coefficients used to start the generation, and the edited image after generating the whole sequence." + ], + "image_footnote": [], + "bbox": [ + 94, + 323, + 223, + 422 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/15650ebe8f5fb1e14f5cbff4550b8905037edd0aa3d798bc3ff42b08b442e51b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 323, + 354, + 422 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5ee81e4ebd6fea49160ac75d835979d3bf656adeb5b1e4f03fcd0468c8bb7029.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 323, + 486, + 422 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a4f5f01bcb766309cac7140a75cfd6400b2de4c276273946f37403fb1790c353.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 89, + 638, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a269ec68a99d50ebee0f6707b77c78d00284cdcb065d5851444d20dad58912dd.jpg", + "image_caption": [ + "\"some boats on a lake.\"" + ], + "image_footnote": [], + "bbox": [ + 638, + 89, + 771, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ea23faefa17f90054e42473d5aea37b128c8b8013a982acaeddc599c646f77d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 771, + 89, + 903, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0d30afa8e30a582f30149e98842e7a669a3e05b5821eb8de5f7de5dd36265abf.jpg", + "image_caption": [ + "\"a green pasture filled with wildflowers.\"" + ], + "image_footnote": [], + "bbox": [ + 509, + 205, + 638, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ff89ab629159c0c08a685fa6f925aba8900b484de8f29dc646b6ec676b76173c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 205, + 771, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b5f7951518fe619b80069c74c9078348110e9869631ac65634c35b482c067907.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 771, + 205, + 903, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/acb38a9ac3421cb4c5aa1397614270a07ffe99a30a75b4517d1a1f8d2e534564.jpg", + "image_caption": [ + "\"a very cute dog laying in a very small sink.\"" + ], + "image_footnote": [], + "bbox": [ + 509, + 323, + 638, + 422 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1a1b11602e9329617c11e9f5b02cf0445c2342b98f5671840a234d1a3534b577.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 323, + 771, + 422 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/caa571ccd6daeecd2b24d2945ccbd8468bc354ee21baf542d4cdf8af06cfa34d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 771, + 323, + 903, + 422 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fe7dc44f4c4262e5cf1e48adcced02f419151f31341155e9b61616f04ff51987.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
paramsseq lenCFGFID ↓IS ↑
AR-ViT-VQGAN (reported)650M1024-8.81110.8
AR-ViT-VQGAN (reproduced)650M1024-8.37111.8
AR-SIT-4 (Ours)650M1024-6.95138.3
RQ-VAE [26]480M256-15.7286.8
DQ-VAE [22]355M640-7.34152.8
LlamaGen-L [45]343M2561.54.08198.5
VAR [48]310M6802.03.30274.4
AR-SIT-4* (Ours)350M10241.54.06190.9
", + "bbox": [ + 93, + 510, + 480, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. ImageNet $256 \\times 256$ class conditional generation. Top half shows a fair comparison against ViT-VQGAN, which we clearly outperform. AR-SIT-4* uses different hyperparameters for comparison against a broader class of methods. While we do not outperform all of them, there are differences in the architectures and training schedules that influence the results and are orthogonal to our contributions. Moreover, none of the baselines has the multiscale capabilities described in P.2 and P.4 (Sec. 1).", + "bbox": [ + 89, + 660, + 482, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "double the number of decoder layers. The bottom half of Tab. 3 compares with similarly-sized AR models including some with multiscale/residual latents (although none exhibit properties P.2 and P.4). AR-SIT-4* performance matches LlamaGen but not to VAR, but the models are not exactly comparable since they use a convolutional tokenizers and other improvements such as AdaLN [34] and attention normalization; VAR also trains the tokenizer on larger", + "bbox": [ + 89, + 779, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "data than ImageNet. These are orthogonal to our contributions and can be incorporated in future work.", + "bbox": [ + 511, + 512, + 903, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also train an AR-SIT-5* on $512 \\times 512$ ImageNet and obtain an FID of 5.97 and IS of 215.22, showing robustness at higher resolutions, while most prior work requires separate stages for upsampling. Fig. 10 shows samples.", + "bbox": [ + 511, + 544, + 905, + 606 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion and limitations", + "text_level": 1, + "bbox": [ + 511, + 627, + 763, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented a spectral image tokenizer (SIT), and an autoregressive generative transformer trained with it (AR-SIT). SIT is naturally multiscale and leverages spectral properties of natural images for improved reconstruction quality. AR-SIT enables applications such as rapid generation of coarse images that can be refined later, and text-guided image upsampling and editing.", + "bbox": [ + 511, + 656, + 905, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While our methods improved tokenizer reconstruction accuracy and class-conditional generation, the text-to-image metrics were similar to the fair baseline. Chang et al. [4] similarly observed that a better tokenizer does not necessarily lead to a better generative model. Nevertheless, our method has multiscale properties and enables new applications not possible with prior work. We only experiment with realtively small AR-SITs of 350-650M parameters, while the Parti [58] baseline goes up to 22B parameters.", + "bbox": [ + 511, + 763, + 906, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17188", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank Leonardo Zepeda-Nuñez for reviewing this manuscript and offering interesting discussions and suggestions, and Jon Barron for sharing useful code we relied on.", + "bbox": [ + 89, + 114, + 483, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 176, + 187, + 191 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] James Betker, Gabriel Goh, Li Jing, † TimBrooks, Jianfeng Wang, Linjie Li, † LongOuyang, † JuntangZhuang, † JoyceLee, † YufeiGuo, † WesamManassra, † PrafullaDhariwal, † CaseyChu, † YunxinJiao, and Aditya Ramesh. Improving image generation with better captions. 2", + "[2] Lei Cai, Hongyang Gao, and Shuiwang Ji. Multi-Stage Variational Auto-Encoders for Coarse-to-Fine Image Generation, page 630–638. Society for Industrial and Applied Mathematics, 2019. 3", + "[3] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T. Freeman. Maskgit: Masked generative image transformer. In CVPR, 2022. 1, 3", + "[4] Huiwen Chang, Han Zhang, Jarred Barber, Aaron Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Patrick Murphy, William T. Freeman, Michael Rubinstein, Yuanzhen Li, and Dilip Krishnan. Muse: Text-to-image generation via masked generative transformers. In ICML, 2023. 1, 3, 8", + "[5] XI Chen, Nikhil Mishra, Mostafa Rohaninejad, and Pieter Abbeel. PixelSNAIL: An improved autoregressive generative model. In ICML, 2018. 3", + "[6] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. Pali: A jointly-scaled multilingual language-image model, 2022. 6", + "[7] C. A. Christopoulos, T. Ebrahimi, and A. N. Skodras. Jpeg2000: the new still picture compression standard. In Proceedings of the 2000 ACM Workshops on Multimedia, page 45-49, New York, NY, USA, 2000. Association for Computing Machinery. 4", + "[8] A. Cohen, Ingrid Daubechies, and J.-C. Feauveau. Biorthogonal bases of compactly supported wavelets. Communications on Pure and Applied Mathematics, 45(5):485-560, 1992. 3", + "[9] Ingrid Daubechies. Ten lectures on wavelets. Society for Industrial and Applied Mathematics, USA, 1992. 3", + "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6", + "[11] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 1", + "[12] Sander Dieleman. Diffusion is spectral autoregression, 2024. 2" + ], + "bbox": [ + 93, + 200, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2, 3, 4", + "[14] Tim Elsner, Paula Usinger, Victor Czech, Gregor Kobsik, Yanjiang He, Isaak Lim, and Leif Kobbelt. Quantised global autoencoder: A holistic approach to representing visual data, 2024. 2", + "[15] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 1, 2, 3, 5", + "[16] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis, 2024. 2", + "[17] Rinon Gal, Dana Cohen Hochberg, Amit Bermano, and Daniel Cohen-Or. Swagan: a style-based wavelet-driven generative model. ACM TOG, 2021. 1", + "[18] Jiatao Gu, Shuangfei Zhai, Yizhe Zhang, Joshua M. Susskind, and Navdeep Jaitly. Matryoshka diffusion models. In ICLR, 2024. 3", + "[19] Philippe Hansen-Estruch, David Yan, Ching-Yao Chung, Orr Zohar, Jialiang Wang, Tingbo Hou, Tao Xu, Sriram Vishwanath, Peter Vajda, and Xinlei Chen. Learnings from scaling visual tokenizers for reconstruction and generation, 2025. 1", + "[20] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 2", + "[21] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022. 1", + "[22] Mengqi Huang, Zhendong Mao, Zhuowei Chen, and Yongdong Zhang. Towards accurate image coding: Improved autoregressive image generation with dynamic vector quantization. In CVPR, 2023. 2, 8", + "[23]Imagen-Team-Google. Imagen 3, 2024. 2", + "[24] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 3", + "[25] Vladimir Kulikov, Shahar Yadin, Matan Kleiner, and Tomer Michaeli. Sinddm: A single image denoising diffusion model. In ICML, pages 17920-17930, 2023. 3", + "[26] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In CVPR, 2022. 2, 8", + "[27] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In ECCV. 5, 6, 7, 8, 2, 3", + "[28] Xiaoxiao Ma, Mohan Zhou, Tao Liang, Yalong Bai, Tiejun Zhao, Huaian Chen, and Yi Jin. Star: Scale-wise text-to-image generation via auto-regressive representations, 2024." + ], + "bbox": [ + 516, + 92, + 905, + 886 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17189", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Stephane Mallat. A Wavelet Tour of Signal Processing, Third Edition: The Sparse Way. Academic Press, Inc., USA, 3rd edition, 2008. 3", + "[30] Wael Mattar, Idan Levy, Nir Sharon, and Shai Dekel. Wavelets are all you need for autoregressive image generation, 2024. 3", + "[31] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. In ICLR, 2018. 5", + "[32] Charlie Nash, Jacob Menick, Sander Dieleman, and Peter W. Battaglia. Generating images with sparse representations. In ICML, 2021. 3", + "[33] OpenAI. Gpt-4 technical report, 2023. 1, 2", + "[34] William Peebles and Saining Xie. Scalable diffusion models with transformers. 2023. 8", + "[35] Hao Phung, Quan Dao, and Anh Tran. Wavelet diffusion models are fast and scalable image generators. In CVPR, 2023. 1", + "[36] Vivek Ramanujan, Kushal Tirumala, Armen Aghajanyan, Luke Zettlemoyer, and Ali Farhadi. When worse is better: Navigating the compression-generation tradeoff in visual tokenization, 2024. 1", + "[37] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. In NeurIPS, 2019. 1, 2", + "[38] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2021. 1", + "[39] Tamar Rott Shaham, Tali Dekel, and Tomer Michaeli. Singan: Learning a generative model from a single natural image. In ICCV, 2019. 3", + "[40] Dohoon Ryu and Jong Chul Ye. Pyramidal denoising diffusion probabilistic models, 2022. 3", + "[41] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, Xi Chen, and Xi Chen. Improved techniques for training gans. In NeurIPS, 2016. 2", + "[42] Noam Shazeer. Glu variants improve transformer, 2020. 7", + "[43] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 1", + "[44] Jianlin Su, Yu Lu, Shengfeng Pan, Bo Wen, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding, 2021. 7", + "[45] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 2, 8", + "[46] Gemini Team. Gemini: A family of highly capable multimodal models, 2023. 1, 2", + "[47] Llama team. The llama 3 herd of models, 2024. 1, 2", + "[48] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. In NeurIPS, 2024. 2, 8", + "[49] Aaron van den Oord, Nal Kalchbrenner, Lasse Espeholt, kory kavukcuoglu, Oriol Vinyals, and Alex Graves. Conditional image generation with pixelCNN decoders. In NeurIPS, 2016. 2" + ], + "bbox": [ + 91, + 92, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Aäron van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In ICML, 2016. 2", + "[51] Aaron van den Oord, Oriol Vinyals, and koray kavukcuoglu. Neural discrete representation learning. In NeurIPS, 2017. 1, 2, 3, 5", + "[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 1", + "[53] Xin Wen, Bingchen Zhao, Ismail Elezi, Jiankang Deng, and Xiaojuan Qi. \"principal components\" enable a new language of images, 2025. 2", + "[54] Jingfeng Yao, Bin Yang, and Xinggang Wang. Reconstruction vs. generation: Taming optimization dilemma in latent diffusion models. In CVPR, 2025. 1", + "[55] Ting Yao, Yingwei Pan, Yehao Li, Chong-Wah Ngo, and Tao Mei. Wave-vit: Unifying wavelet and transformers for visual representation learning. In ECCV, 2022. 2", + "[56] Hu Yu, Hao Luo, Hangjie Yuan, Yu Rong, and Feng Zhao. Frequency autoregressive image generation with continuous tokens, 2025. 2", + "[57] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved VQGAN. In ICLR, 2022. 1, 2, 3, 4, 5, 6, 7", + "[58] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research, 2022. Featured Certification. 1, 3, 5, 6, 8", + "[59] Lijun Yu, Jose Lezama, Nitesh Bharadwaj Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G Hauptmann, Boqing Gong, Ming-Hsuan Yang, Irfan Essa, David A Ross, and Lu Jiang. Language model beats diffusion - tokenizer is key to visual generation. In ICLR, 2024. 2", + "[60] Qihang Yu, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An image is worth 32 tokens for reconstruction and generation. In NeurIPS, 2024. 1", + "[61] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 2", + "[62] Zhenhai Zhu and Radu Soricut. Wavelet-based image tokenizer for vision transformers, 2024. 2" + ], + "bbox": [ + 516, + 92, + 903, + 756 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17190", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Spectral Image Tokenizer/2b3586c0-b936-4d11-a38a-2ff8e5859a65_model.json b/2025/Spectral Image Tokenizer/2b3586c0-b936-4d11-a38a-2ff8e5859a65_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7b74d11b03a738080b26da5f98411488f4c47c47 --- /dev/null +++ b/2025/Spectral Image Tokenizer/2b3586c0-b936-4d11-a38a-2ff8e5859a65_model.json @@ -0,0 +1,2849 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.368, + 0.131, + 0.631, + 0.154 + ], + "angle": 0, + "content": "Spectral Image Tokenizer" + }, + { + "type": "text", + "bbox": [ + 0.164, + 0.19, + 0.284, + 0.206 + ], + "angle": 0, + "content": "Carlos Esteves" + }, + { + "type": "text", + "bbox": [ + 0.409, + 0.19, + 0.562, + 0.206 + ], + "angle": 0, + "content": "Mohammed Suhail" + }, + { + "type": "text", + "bbox": [ + 0.688, + 0.19, + 0.829, + 0.206 + ], + "angle": 0, + "content": "Ameesh Makadia" + }, + { + "type": "text", + "bbox": [ + 0.327, + 0.214, + 0.644, + 0.23 + ], + "angle": 0, + "content": "{machc, suhailmhd, makadia}@google.com" + }, + { + "type": "text", + "bbox": [ + 0.417, + 0.232, + 0.555, + 0.25 + ], + "angle": 0, + "content": "Google Research" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.307, + 0.327, + 0.323 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.34, + 0.486, + 0.658 + ], + "angle": 0, + "content": "Image tokenizers map images to sequences of discrete tokens, and are a crucial component of autoregressive transformer-based image generation. The tokens are typically associated with spatial locations in the input image, arranged in raster scan order, which is not ideal for autoregressive modeling. In this paper, we propose to tokenize the image spectrum instead, obtained from a discrete wavelet transform (DWT), such that the sequence of tokens represents the image in a coarse-to-fine fashion. Our tokenizer brings several advantages: 1) it leverages that natural images are more compressible at high frequencies, 2) it can take and reconstruct images of different resolutions without retraining, 3) it improves the conditioning for next-token prediction – instead of conditioning on a partial line-by-line reconstruction of the image, it takes a coarse reconstruction of the full image, 4) it enables partial decoding where the first few generated tokens can reconstruct a coarse version of the image, 5) it enables autoregressive models to be used for image upsampling. We evaluate the tokenizer reconstruction metrics as well as multiscale image generation, text-guided image upsampling and editing." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.691, + 0.222, + 0.707 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.718, + 0.484, + 0.809 + ], + "angle": 0, + "content": "In natural language processing, tokenization associates sets of characters to entries in a vocabulary, which is straightforward since language is discrete. The sequence of tokens is suitably modeled as a categorical distribution with autoregressive (AR) transformers, which is the foundation of modern large language models (LLMs) [33, 46, 47]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "While natural images are represented by discrete pixel values, they exhibit high dimensionality, redundancies, and noise so it's impractical to associate one token per pixel. This motivated a long line of research of learnable image tokenizers [15, 37, 37, 51, 57]. While there are successful autoregressive image generation models [15, 57, 58]," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.308, + 0.907, + 0.384 + ], + "angle": 0, + "content": "images are not sequential like language, which motivated developing alternatives such as denoising diffusion models [11, 38, 43] and masked transformers [3, 4, 60]. Nevertheless, most of these alternatives also operate on the latent space of tokenizers like VQGAN [15] instead of raw pixels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.384, + 0.907, + 0.49 + ], + "angle": 0, + "content": "In this work, we revisit AR transformer-based image generation. Our main contribution is a tokenizer operating on the image spectrum, specifically on DWT coefficients, where the coarse-to-fine representation lends itself more naturally to a sequential interpretation. Our Spectral Image Tokenizer (SIT) has several useful properties and enables different applications:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.494, + 0.906, + 0.569 + ], + "angle": 0, + "content": "P.1 Since the power spectrum of natural images decreases with frequency, high frequencies can be more heavily compressed with little effect in visual quality. SIT leverages this by associating tokens to larger patches at higher wavelet scales than at lower scales (see Fig. 2)." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.569, + 0.906, + 0.674 + ], + "angle": 0, + "content": "P.2 SIT is transformer-based [52]; by using an attention mask where each scale depends on itself and lower scales (\"Scale-Causal attention\"), SIT can be trained at a single resolution and used to tokenize images of multiple resolutions (any number of scales up to the trained maximum), and detokenize partial token sequences (up to some scale), reconstructing a coarse image." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.675, + 0.906, + 0.765 + ], + "angle": 0, + "content": "P.3 Using SIT, we train an autoregressive generative transformer (AR-SIT) that models images coarse-to-fine. The next-token prediction is then conditioned on a coarse reconstruction of the image given by the partial token sequence, instead of the usual conditioning on the partial reconstruction of the previous rows of the image." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.765, + 0.905, + 0.825 + ], + "angle": 0, + "content": "P.4 AR-SIT can quickly generate only the first few tokens and reconstruct a coarse version of the image, enabling applications like quickly showing multiple coarse generations and letting the user select which ones to refine." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.825, + 0.905, + 0.885 + ], + "angle": 0, + "content": "P.5 AR-SIT can be used for text-based upsampling of an input low resolution image, by starting the decoding process with the few tokens output by SIT, and generating the rest of the sequence up to a desired resolution." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.885, + 0.905, + 0.902 + ], + "angle": 0, + "content": "P.6 AR-SIT can be used for text-guided image editing, by" + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.494, + 0.906, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17181" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.087, + 0.382, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.426, + 0.087, + 0.909, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.319, + 0.907, + 0.417 + ], + "angle": 0, + "content": "Figure 1. Left: we introduce a Spectral Image Tokenizer (SIT), that learns to encode and decode discrete wavelet transform (DWT) coefficients to and from a small set of discrete tokens, such that the sequence represents the image in a coarse-fine fashion. SIT is naturally multiscale and enables coarse-to-fine autoregressive image generation with our AR-SIT model. SIT also leverages the sparsity of high frequency coefficients in natural images. Right: details of the encoder/decoder transformer layers. The main architectural difference with respect to previous tokenizers is that the distributions of DWT approximation and details coefficients are distinct, hence we use specialized parameters for each in the quantizer codebooks and inner transformer layers. We also introduce a scale-causal attention where each token attends to its own scale and lower scales, which enables encoding, decoding, generating, and upsampling images at different resolutions." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.429, + 0.483, + 0.459 + ], + "angle": 0, + "content": "encoding a given image up to a coarse scale, and generating the finer details conditioned on a new caption." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.463, + 0.484, + 0.676 + ], + "angle": 0, + "content": "Currently, image generation is dominated by diffusion models such as Imagen 3 [23], DALL-E-3 [1] and Stable Diffusion 3 [16]. On the other hand, LLMs such as Gemini [46], GPT-4 [33], and Llama 3 [47] are based on autoregressive transformers. We believe autoregressive image generation is still worth pursuing because it might benefit from advances in LLMs, and multimodal applications might benefit of having a single architecture for all modalities. Recent work on image and video generation support this point [45, 48, 59]. Dieleman [12] recently interpreted denoising diffusion models as spectral autoregression, since, when looking at image spectra, the denoising procedure uncovers frequencies from low to high. In contrast, our method does literal spectral autoregression." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.692, + 0.228, + 0.708 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.869 + ], + "angle": 0, + "content": "Image tokenization Several methods have been developed to map images to a small set of discrete tokens suitable for generative modeling. VQ-VAE [51] introduced vector-quantization in the latent space of a Variational AutoEncoder to map images, audio and video to a set of discrete values. VQGAN [15] improved upon VQVAE by using perceptual and adversarial losses. We build on ViT-VQGAN [57], which improved upon VQGAN by using a Vision Transformer (ViT) [13] instead of convolutions, as well as codebook factorization and feature normalization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this paper, we are interested in multiscale image representations. VQ-VAE-2 [37] introduced multiscale latents" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.428, + 0.907, + 0.609 + ], + "angle": 0, + "content": "by keeping and quantizing intermediate downsampled convolutional features. RQ-VAE [26] quantized a set of residuals such that the latent vector is represented in a coarse-to-fine fashion and reconstructed by adding the code embeddings for each residual. Similarly, VAR [48] and STAR [28] split the latent space in multi-scale quantized residuals, so that it can be reconstructed by upsampling and summing. FAR [56] splits the latents into frequencies to train a next-frequency generative model. DQ-VAE [22] encoded smooth regions with fewer tokens that are generated first, while QG-VAE [14] and SEMANTICIST [53] learn non-local tokens that are ordered by importance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.907, + 0.7 + ], + "angle": 0, + "content": "The crucial difference between our approach and the aforementioned is that we operate on spectral coefficients of the input and not on latent features. This allows our tokenizer to be truly multiscale; it can take inputs at different scales and reconstruct up to some scale, and leverage that higher frequencies are more compressible in natural images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.701, + 0.906, + 0.837 + ], + "angle": 0, + "content": "Tangentially related to our work, Wave-ViT [55] modified the ViT self-attention by applying a DWT to the input and concatenating coefficients, effectively exchanging space for depth to reduce the sequence length. Zhu and Soricut [62] modified the ViT patchifier in a similar way, and introduced patch embeddings that leverage the sparsity of high frequency coefficients. Both methods are for discriminative tasks such as image classification and segmentation, while we focus on generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Autoregressive image generation Early approaches PixelRNN and PixelCNN [49, 50] generate images pixel by pixel by modeling the conditional distribution of the pixel given the previous pixels with recurrent layers of causal" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17182" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.242 + ], + "angle": 0, + "content": "convolutions. PixelSNAIL [5] improved on this model by introducing self-attention layer to better model long-range dependencies. VQ-VAE [51] introduced a two-stage approach with a tokenizer and a separate stage to model the distribution of tokens. VQGAN [15] greatly improved these results by both improving the tokenizer and using a transformer to model the distribution. Finally, ViT-VQGAN [57] proposed a transformer-based tokenizer, which Parti [58] used with a large autoregressive transformer capable of high-quality text-to-image generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.484, + 0.348 + ], + "angle": 0, + "content": "MaskGIT [3] and Muse [4] highlighted the disadvantages of the typical autoregressive models raster-order conditioning, and proposed to generate all tokens in parallel iteratively, where each iteration keeps the highest confidence tokens. We address the same problem with a tokenizer whose sequence represents the image in a coarse-to-fine order instead of raster-order." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.35, + 0.484, + 0.41 + ], + "angle": 0, + "content": "Multiscale image generation Multiscale image generation ideas have appeared in the context of VAEs [2], GANs [24, 39], and diffusion models [18, 25, 40], but have not been sufficiently explored with AR transformers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.411, + 0.484, + 0.622 + ], + "angle": 0, + "content": "Nash et al. [32] represented an image as a sequence of quantized and thresholded DCT coefficients, where the compression comes from the fact that many coefficients are zero and are omitted. By sorting the sequence by frequency, the method can model images from coarse-to-fine like we do. However, the compressed representation is handcrafted and results in long sequences. In a similar vein, Mattar et al. [30] handcrafted a tokenizer based on DWT coefficients, introducing tokens to represent large chunks of zeros. It also results in long sequences and is only applied to generations of small grayscale images. In contrast with these methods, instead of handcrafting a compressed input representation, our tokenizer learns to encode to, and decode from, a short compressed coarse-to-fine sequence." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.636, + 0.22, + 0.654 + ], + "angle": 0, + "content": "3. Background" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.661, + 0.484, + 0.707 + ], + "angle": 0, + "content": "A discrete wavelet transform (DWT) is based on successive convolutions of the signal \\( f \\) with a pair of lowpass \\( g \\) and highpass \\( h \\) filters with the first step as follows" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.718, + 0.482, + 0.75 + ], + "angle": 0, + "content": "\\[\nf _ {\\operatorname {l o w} _ {1}} [ n ] = f \\star g = \\sum_ {k} f [ k ] g [ n - k ], \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.753, + 0.482, + 0.784 + ], + "angle": 0, + "content": "\\[\nf _ {\\operatorname {h i g h} _ {1}} [ n ] = f \\star h = \\sum_ {k} f [ k ] h [ n - k ]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The high and low outputs are subsampled by a factor of two, and the operation is repeated for the low channel, such that at level \\( L \\) we compute \\( f_{\\mathrm{low}_L} = f_{\\mathrm{low}_{L - 1}} \\star g \\) and \\( f_{\\mathrm{high}_L} = f_{\\mathrm{low}_{L - 1}} \\star h \\), subsample again and drop \\( f_{\\mathrm{low}_{L - 1}} \\). The output at level \\( L \\) comprises the approximation coefficients \\( f_{\\mathrm{low}_L} \\) and the detail coefficients \\( \\{f_{\\mathrm{high}_k}\\} \\) for \\( 1 \\leq k \\leq L \\). This output has the same cardinality as the input and the" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.71, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.091, + 0.903, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.251, + 0.907, + 0.405 + ], + "angle": 0, + "content": "Figure 2. Input patchification. Left: typical patchification for Vision Transformers (ViT) [13], where the image is split in equal-sized patches. Right: we propose to patchify the coefficients of a discrete wavelet transform (DWT) instead. Each scale is shown in a different color. Scales other than the lowest contain three blocks representing horizontal, vertical and diagonal details; we concatenate the spatially correspondent patches of each block such that each scale is represented by the same sequence length. This results in larger patch sizes for higher scales, which are more compressible. The figure shows 3 scales and 16 tokens per scale; in our experiments we use 4 or 5 scales and 256 tokens per scale." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.417, + 0.906, + 0.477 + ], + "angle": 0, + "content": "transformation is invertible, where the forward transform is typically called \"analysis\" and the backward \"synthesis\". The simplest wavelet family is the Haar, where \\( g = [1,1]^{\\top} \\) and \\( h = [1, -1]^{\\top} \\) (optionally scaled to unit norm)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.906, + 0.613 + ], + "angle": 0, + "content": "For image analysis we use a 2D DWT, which is obtained by simply convolving the rows and columns with \\( g \\) and \\( h \\). The approximation coefficients \\( f_{\\mathrm{low}_1} = f \\star (gg^\\top) \\), and the details are divided into horizontal \\( f_{\\mathrm{H}_1} = f \\star (gh^\\top) \\), vertical \\( f_{\\mathrm{V}_1} = f \\star (hg^\\top) \\) and diagonal \\( f_{\\mathrm{D}_1} = f \\star (hh^\\top) \\) details. Subsequent levels apply the same operations to the approximation coefficients. Fig. 2 shows a two-level transform, where we can see that the approximation coefficients correspond to a coarse version of the image." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.614, + 0.906, + 0.644 + ], + "angle": 0, + "content": "We refer to the textbooks by Mallat [29] and Daubechies [9] for more information about wavelets." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.66, + 0.605, + 0.675 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.686, + 0.906, + 0.777 + ], + "angle": 0, + "content": "Our main contribution, SIT, is an image tokenizer that takes discrete wavelet transform (DWT) coefficients. The model follows ViT-VQGAN [57], with important changes that we describe in this section and visualize in Fig. 1. For image generation, we introduce AR-SIT which is based on Parti [58] with minor changes, using SIT as the tokenizer." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.627, + 0.802 + ], + "angle": 0, + "content": "4.1. Tokenizer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Patchification and featurization The first step is to map the input to a sequence of patches. We apply the Haar DWT on the input image and patchify each scale separately. While Haar is the simplest wavelet and lacks properties found in other wavelets useful for compression, we found no benefits of using other wavelet families such as CDF [8]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17183" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "Our design choice is to use the same number of patches for each scale. In a DWT, the higher scales correspond to high frequency details which are represented by more coefficients than the lower scales, but contribute less to the spatial pixel values. In other words, in natural images, most of the power spectrum is concentrated on lower frequencies. By representing each scale with the same number of tokens, we are compressing more the higher frequencies (since they are represented by more coefficients), similarly to what is done in image compression methods such as JPEG2000 [7]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.483, + 0.439 + ], + "angle": 0, + "content": "The approximation (or lowpass) DWT coefficients correspond to a coarse version of the input image, which we consider the first scale. The following scales are divided in three blocks corresponding to horizontal, vertical, and diagonal details, where each coefficient relates to a specific spatial location. Thus, we can concatenate the three blocks such that each entry corresponds to the same spatial location. For example, the first scale will typically be split in patches that are \\(32 \\times 32 \\times 3\\), with channels corresponding to RGB, the second scale will be \\(32 \\times 32 \\times 9\\), where the channels correspond to the RGB of horizontal, vertical, and diagonal details, the third will be \\(64 \\times 64 \\times 9\\) and so on. Fig. 2 shows an example of our patchification scheme." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.44, + 0.483, + 0.516 + ], + "angle": 0, + "content": "Since patches have different resolutions (higher scales will have larger patches), the usual ViT linear embedding to map patches to features cannot be shared across all patches so we have different parameters per scale. Formally, given an image \\( f \\), our patchifiers compute" + }, + { + "type": "equation", + "bbox": [ + 0.109, + 0.529, + 0.483, + 0.546 + ], + "angle": 0, + "content": "\\[\nf _ {\\text {l o w} L}, \\left\\{f _ {\\mathrm {H} _ {i}} \\right\\} _ {i \\leq L}, \\left\\{f _ {\\mathrm {V} _ {i}} \\right\\} _ {i \\leq L}, \\left\\{f _ {\\mathrm {D} _ {i}} \\right\\} _ {i \\leq L} = \\mathrm {D W T} (f), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.109, + 0.548, + 0.482, + 0.564 + ], + "angle": 0, + "content": "\\[\nc _ {1} = P _ {0} \\left(f _ {\\text {l o w} L}\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.109, + 0.567, + 0.482, + 0.583 + ], + "angle": 0, + "content": "\\[\nc _ {s} = P _ {s} \\left(f _ {\\mathrm {H} _ {L - s + 2}}, f _ {\\mathrm {V} _ {L - s + 2}}, f _ {\\mathrm {D} _ {L - s + 2}}\\right), 1 < s \\leq L, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.595, + 0.483, + 0.701 + ], + "angle": 0, + "content": "where \\( c_{s} = \\{c_{s}^{n}\\}_{1\\leq n\\leq N} \\) is the sequence of token embeddings at scale \\( s \\), and \\( c_{s}^{n}\\in \\mathbb{R}^{C} \\) is the embedding of the \\( n \\)-th token at the \\( s \\)-th scale, while \\( L \\) is the number of DWT levels, \\( S = L + 1 \\) is the number of scales and \\( N \\) the number of tokens per scale. For brevity, when there is no ambiguity, we may omit the set indexing and use \\( \\{c_s\\} \\) to denote the set of tokens of all scales, for example." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.702, + 0.483, + 0.777 + ], + "angle": 0, + "content": "The final projection after the decoder needs to map the features back to different sized patches, so it also has different parameters per scale. Those patches still represent DWT coefficients so an inverse DWT (IDWT) is finally applied to obtain an image output." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Flexible sequence length The tokenizer encoder and decoder transformers operate on the sequence of patch features of length \\(SN\\). The sequence length is a major factor in resource utilization so we want to keep it constrained. Our method is flexible since we can choose the number of scales and the patch size per scale, while most ViT-based models such as ViT-VQGAN [57] are more restricted. They use the same patch size for the whole image; thus, keeping the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "same patch size and doubling each image dimension would increase the sequence length by a factor of 4, where our method is capable of including additional scales which only increases the sequence length by multiples of \\( N \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.905, + 0.289 + ], + "angle": 0, + "content": "For example, for a \\(256 \\times 256\\) input, ViT-VQGAN uses \\(8 \\times 8\\) patches to obtain a sequence of 1024 elements. Our SIT can use 4 scales and 256 tokens per scale for the same sequence length. When increasing the resolution to \\(512 \\times 512\\), the baseline can either increase the patch size to \\(16 \\times 16\\), resulting in the same sequence length, or keep it \\(8 \\times 8\\), resulting in a \\(4 \\times\\) longer sequence. SIT, for example, can vary the number of scales from 4 to 6, resulting in sequence lengths of 1024, 1280, 1536 which are more manageable." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.295, + 0.905, + 0.37 + ], + "angle": 0, + "content": "Transformers After featurization, the single sequence containing all scales passes through a transformer encoder, followed by quantization and a transformer decoder. We propose two optional modifications to the transformers, which are otherwise identical to the ones used in ViT [13]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.905, + 0.644 + ], + "angle": 0, + "content": "First, we propose a scale-causal attention mask, where an element at some scale attends to all elements of its own scale and lower scales, represented by a block-triangular mask pattern. With dense attention, we write the application of the encoding transformer \\( T_{\\mathrm{enc}} \\) as \\( \\{z_s\\} = T_{\\mathrm{enc}}(\\{c_s\\}) \\), and for scale-causal we write \\( z_s = T_{\\mathrm{enc}}(\\{c_k\\}_{1 \\leq k \\leq s}) \\) for each \\( s \\). The scale-causal attention can be applied independently to the encoder and decoder, enabling different applications. For the multiscale reconstruction experiments in Sec. 5.1, we need to both encode and decode multiple resolutions, so both encoder and decoder use scale-causal masks. For coarse-to-fine image generation in Sec. 5.2, only the decoder needs to be scale-causal in order to decode the partially generated sequence. For the text-guided image upsampling in Sec. 5.3, only the encoder needs to be scale-causal to encode the lower resolution inputs. For the image editing experiments in Sec. 5.4, the scale-causal encoder prevents information leaking from high to low scale." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Second, we propose to use different transformer parameters for the approximation (first scale) and details coefficients (other scales). In our model, the subsequences corresponding to coefficients of each type of coefficient come from quite distinct distributions, so it makes sense to treat them differently. This contrasts with the spatial representation of images where each patch can be considered as coming from the same distribution. Thus, the parameters of the key, query, and value embeddings, the layer norms, and the MLP on each transformer layer are not shared between the approximation and details coefficients. This Approximation-Details Transformer (ADTransformer) still takes a single sequence composed of all coefficients. We experimented with different transformers per sequence with cross-attention for information sharing, but it performed worse. This change leads to more memory utilization to store the extra parameters, but the training/inference speed" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17184" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.09, + 0.482, + 0.405 + ], + "angle": 0, + "content": "
LPIPS ↓PSNR ↑L1 ↓FID ↓IS ↑images/s ↑
Resolution: 512 × 512
ViT-VQGAN0.32022.40.0426.92151.5593
SIT-5 (Ours)0.26022.00.0462.65192.0410
SIT-6 (Ours)0.23923.10.0401.74203.7320
Resolution: 256 × 256
ViT-VQGAN (reported)-24.80.0321.99184.4-
ViT-VQGAN (reproduced)0.16725.00.0312.33184.0-
ViT-VQGAN (no LL)0.16323.80.0381.20194.6626
SIT-4 (Ours)0.14424.00.0371.20199.5596
SIT-5 (Ours)0.13524.50.0350.97202.3411
SIT-SC-5 (Ours)0.16124.10.0371.33193.7411
Resolution: 128 × 128
ViT-VQGAN0.18526.30.0303.77117.3626
SIT-SC-5 (ours)0.15927.10.0272.13129.3582
Resolution: 64 × 64
ViT-VQGAN0.12928.80.0233.5321.0627
SIT-SC-5 (ours)0.11131.30.0171.3930.1847
Resolution: 32 × 32
ViT-VQGAN0.21423.30.045-3.7627
SIT-SC-5 (ours)0.02936.80.0100.313.5825
Resolution: 16 × 16
ViT-VQGAN0.12724.90.039-1.7627
SIT-SC-5 (ours)0.01341.30.0060.091.82620
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.414, + 0.484, + 0.581 + ], + "angle": 0, + "content": "Table 1. Multiscale reconstruction on ImageNet. \"SC\" denotes scale-causal attention, which slightly reduces performance at the highest resolution but enables multiscale reconstruction without downsampling/upsampling or retraining. The ViT-VQGAN values from Yu et al. [57] used a logit-laplace loss (LL) which was later considered harmful [58], so we retrain without it. Our SIT improves reconstruction metrics, is significantly faster at lower resolutions, and robust when increasing resolution. The ViT-VQGAN baseline suffered from instability during training on \\(512 \\times 512\\) inputs so we selected the best values before divergence. We report test time throughput of an encoding/decoding cycle for the max batch size that fits on a TPU v5e." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.594, + 0.474, + 0.61 + ], + "angle": 0, + "content": "is similar because the number of operations is unchanged." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.484, + 0.794 + ], + "angle": 0, + "content": "Quantizer The encoder outputs a sequence of features that are quantized to a fixed-sized codebook, similarly to ViT-VQGAN [57], VQVAE [51], and VQGAN [15]. We modify the quantizer such that approximation and details have different codebooks for the same reasons discussed previously. Thus, codebook sizes and feature dimensions are the same as the ViT-VQGAN baseline, but we have different features for the same code at different positions. Formally, we apply \\( q_{s}^{n} = Q_{\\mathrm{approx}}(z_{s}^{n}) \\) for \\( s = 1 \\) and \\( q_{s}^{n} = Q_{\\mathrm{details}}(z_{s}^{n}) \\) for \\( s \\geq 2 \\) for each pair \\( (s,n) \\), where \\( q_{s}^{n} \\) is chosen from the codebook for \\( s \\) and can be associated with its discrete position in the codebook, denoted \\( \\lfloor q_s^n \\rfloor \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.856 + ], + "angle": 0, + "content": "Training We follow the ViT-VQGAN training protocol and use the same weighting for the L2, perceptual, adversarial, and quantization losses. We remove the logit-laplace loss that was shown detrimental in follow-up work [58]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "All losses are applied to the spatial domain images reconstructed by the inverse discrete wavelet transform on the decoder output: \\(\\hat{f} = \\mathrm{IDWT}(T_{\\mathrm{dec}}(\\{q_s\\}))\\). We noticed insta" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.091, + 0.905, + 0.328 + ], + "angle": 0, + "content": "
FID ↓IS ↑images/s ↑images/Gb ↑
Resolution: 256 × 256
Parti350M (reported)14.1---
Parti350M12.436.57.812.0
AR-SIT-SCD-4 (Ours)12.637.36.58.0
Resolution: 128 × 128
Parti350M11.233.57.612.0
AR-SIT-SCD-4 (Ours)11.433.212.612.0
Resolution: 64 × 64
Parti350M10.516.97.612.0
AR-SIT-SCD-4 (Ours)11.418.624.516.0
Resolution: 32 × 32
Parti350M5.82.97.77.7
AR-SIT-SCD-4 (Ours)7.63.274.728.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.338, + 0.907, + 0.465 + ], + "angle": 0, + "content": "Table 2. Coarse-to-fine image generation on MS-COCO [27]. With SIT, the autoregressive generation is stopped early for a coarse version of the image. The Parti350M baseline does not have this property, so we generate at full resolution and downsample for comparison. AR-SIT matches the baseline performance at the training resolution but is several times faster and more memory efficient at lower resolutions, even when trained only on higher resolution data. We report throughput and memory utilization during generation given the max batch size that fits on a TPU v6e." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.475, + 0.907, + 0.537 + ], + "angle": 0, + "content": "bility during training due to the adversarial loss, which was fixed by applying spectral normalization following Miyato et al. [31], which simply divides the discriminator weight matrices by their largest singular value." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.544, + 0.8, + 0.56 + ], + "angle": 0, + "content": "4.2. Autoregressive image generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.566, + 0.906, + 0.641 + ], + "angle": 0, + "content": "We use our tokenizer for autoregressive image generation, by training a second stage transformer model similar to Parti [58], with some modifications. Formally, the autoregressive transformer \\(T\\) models categorical distributions over the discrete codes" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.649, + 0.905, + 0.668 + ], + "angle": 0, + "content": "\\[\nP \\left(\\left\\lfloor q _ {s} ^ {n} \\right\\rfloor \\mid \\left\\{\\left\\lfloor q _ {i} \\right\\rfloor \\right\\} _ {1 \\leq i < s} \\cup \\left\\{\\left\\lfloor q _ {s} ^ {i} \\right\\rfloor \\right\\} ^ {1 \\leq i < n}\\right) = T (c) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.906, + 0.766 + ], + "angle": 0, + "content": "which can be sampled one code at a time for generation. \\(T\\) can be conditioned on a textual description processed by a transformer encoder for text-to-image generation. For training, we model the distribution of input codes as \\(P(\\{\\lfloor q_s\\rfloor \\}) = \\prod_{1\\leq s\\leq S}^{1\\leq n\\leq N}P(\\lfloor q_s^n\\rfloor)\\) and minimize the negative log-likelihood \\(-\\log P(\\{\\lfloor q_s\\rfloor \\})\\) over the training set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.906, + 0.84 + ], + "angle": 0, + "content": "With different codebooks for approximation and details, the same token id might have a different meaning depending on its position. Thus, the AR model has different token embeddings for approximation and details tokens. The last layer for logit prediction is also different." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "For the generative applications in Secs. 5.2 and 5.3, we introduce mild changes in order to interrupt the generation after all tokens up to a certain scale are generated, and to start the generation with given tokens up to a certain scale." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17185" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.49, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.166, + 0.413, + 0.175 + ], + "angle": 0, + "content": "\"a zebra standing on dirty area with trees in the background.\"" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.177, + 0.49, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.165, + 0.254, + 0.419, + 0.262 + ], + "angle": 0, + "content": "\"an elaborate metal vase holds a decorative bouquet of flowers.\"" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.265, + 0.49, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.342, + 0.412, + 0.351 + ], + "angle": 0, + "content": "\"the wing of a plane located in the sky above a sea and city.\"" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.089, + 0.905, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.608, + 0.166, + 0.805, + 0.175 + ], + "angle": 0, + "content": "\"a modern bathroom is shown with a square sink.\"" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.177, + 0.905, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.646, + 0.254, + 0.768, + 0.263 + ], + "angle": 0, + "content": "\"a toy boat sits on the ground.\"" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.265, + 0.905, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.342, + 0.845, + 0.351 + ], + "angle": 0, + "content": "\"This brown and white cat is standing on a desk in front of a computer.\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.362, + 0.908, + 0.404 + ], + "angle": 0, + "content": "Figure 3. Coarse-to-fine text-to-image on MS-COCO [27] prompts. Each quadruple shows generations from AR-SIT-SCD for the given prompt, with only the first \\(25\\%\\), \\(50\\%\\), \\(75\\%\\) and \\(100\\%\\) of tokens generated, corresponding to resolutions of \\(32 \\times 32\\) to \\(256 \\times 256\\). Training uses only \\(256 \\times 256\\) images. Our model enables quick generation of coarse image candidates that can be further improved if needed." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.411, + 0.223, + 0.429 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.438, + 0.484, + 0.558 + ], + "angle": 0, + "content": "We focus our experiments on demonstrating the properties and applications enumerated in P.1-P.6 (Sec. 1). Our ideas build on ViT-based tokenizers, so the most fair comparisons are against ViT tokenizers and AR generative models; namely ViT-VQGAN [57] and Parti [58]. In these comparisons we are able to match the architectures and training protocol exactly. Nevertheless we also compare with a broader set of methods in Sec. 5.5." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.572, + 0.325, + 0.587 + ], + "angle": 0, + "content": "5.1. Multiscale reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.595, + 0.483, + 0.641 + ], + "angle": 0, + "content": "We train our tokenizer on ImageNet [10] and evaluate its reconstruction performance. We follow the ViT-VQGAN [57] \"Small\" architecture and training protocol." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.483, + 0.793 + ], + "angle": 0, + "content": "The model \"SIT-5\" is the Spectral Image Tokenizer with 5 scales from \\(16 \\times 16\\) to \\(128 \\times 128\\) resolutions, where each scale is represented by 256 tokens. The variation \"SIT-SC\" uses scale-causal attention on both the encoder and decoder, which enables handling inputs and outputs of different resolutions, even though it is only trained at \\(256 \\times 256\\). For example, when the input image is \\(64 \\times 64\\), only the first three scales are used, resulting in shorter sequence lengths which reduces memory utilization and processing time. Both models employ the ADTransformer described in Sec. 4.1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.794, + 0.483, + 0.87 + ], + "angle": 0, + "content": "The ViT-VQGAN baseline only works for \\(256 \\times 256\\) inputs, so to evaluate against it fairly, we upsampled the low-resolution inputs to that resolution. This brings the input patches away from the training distribution, which might explain the drop in reconstruction quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We also evaluate at \\(512 \\times 512\\). The ViT-VQGAN baseline increases its patch size from \\(8 \\times 8\\) to \\(16 \\times 16\\), keep-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.412, + 0.905, + 0.488 + ], + "angle": 0, + "content": "ing the sequence length constant. Interestingly, the baseline suffered heavy instability during training, which was not resolved by reducing the learning rate, using spectral normalization, or the logit-laplace loss. In contrast, our SIT variations trained successfully with no hyperparameter changes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.906, + 0.534 + ], + "angle": 0, + "content": "Tab. 1 shows the reconstruction metrics; Fig. 7 shows reconstruction samples at multiple resolutions. Results provide evidence for P.1 and demonstrate P.2 (Sec. 1)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.545, + 0.853, + 0.562 + ], + "angle": 0, + "content": "5.2. Coarse-to-fine text-to-image generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.906, + 0.674 + ], + "angle": 0, + "content": "We tackle text-to-image generation by using an autoregressive transformer to model the distribution of discrete tokens output by our tokenizer. Since the SIT sequence of tokens represents an image in a coarse-fine fashion, the autoregressive model generation has the same property, which means that we can interrupt the generation after a certain number of tokens and decode a coarse version of the generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.906, + 0.75 + ], + "angle": 0, + "content": "For this to work, the SIT decoder must be scale-causal, while there is no constraint for the encoder so we use dense attention (we denote this variant Scale-Causal Decoder, or \"SIT-SCD\"). It uses a \"small\" encoder and \"large\" decoder as described in ViT-VQGAN [57], and 4 scales." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "We follow the Parti350M [58] architecture and training protocol and name our generative model auto-regressive SIT, or \"AR-SIT\". SIT and AR-SIT are trained on a subset of WebLI [6] of around 128M images, where each image is seen at most once during training. Evaluation is on 30k examples of MS-COCO [27]. We show metrics and generations at different resolutions in Fig. 3 and Tab. 2. AR-SIT matches the baseline performance but is several times faster and more memory efficient at lower resolutions. These results demonstrate the property P.4 (Sec. 1)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17186" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.223, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.09, + 0.356, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.09, + 0.488, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.191, + 0.451, + 0.203 + ], + "angle": 0, + "content": "\"an assortment of some colorful vases on display on a table\"" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.207, + 0.225, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.207, + 0.356, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.207, + 0.488, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.308, + 0.442, + 0.32 + ], + "angle": 0, + "content": "\"one train riding on the multiple train tracks side by side.\"" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.324, + 0.225, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.324, + 0.356, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.324, + 0.488, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.425, + 0.402, + 0.437 + ], + "angle": 0, + "content": "\"an image of a fired hydrant on the street\"" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.089, + 0.64, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.09, + 0.772, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.774, + 0.09, + 0.904, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.553, + 0.191, + 0.862, + 0.203 + ], + "angle": 0, + "content": "\"a large elephant standing on top of a grass covered field.\"" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.207, + 0.641, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.207, + 0.772, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.774, + 0.207, + 0.904, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.557, + 0.308, + 0.858, + 0.32 + ], + "angle": 0, + "content": "\"three tropical bird perched on top of high bare branches\"" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.324, + 0.64, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.324, + 0.772, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.774, + 0.324, + 0.904, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.573, + 0.425, + 0.842, + 0.437 + ], + "angle": 0, + "content": "\"a chair with a pink blanket is sitting on a red floor.\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.448, + 0.908, + 0.492 + ], + "angle": 0, + "content": "Figure 4. Text-guided image upsampling on MS-COCO [27]. Our coarse-to-fine generative models can take a low-resolution image, encode it as the first few tokens of a sequence, and generate the rest of sequence, which, when decoded, effectively upsamples the input. Each triplet shows the given \\(32 \\times 32\\) image, our \\(256 \\times 256\\) reconstruction and the ground truth." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.498, + 0.364, + 0.515 + ], + "angle": 0, + "content": "5.3. Text-guided image upsampling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.483, + 0.566 + ], + "angle": 0, + "content": "We leverage the coarse-to-fine nature of SIT to apply a pretrained AR-SIT for text-guided image upsampling, where we are given a low resolution image and a caption." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.566, + 0.483, + 0.686 + ], + "angle": 0, + "content": "The idea is to encode the low-resolution image, which will give the first tokens of our high resolution output. AR-SIT then takes these tokens and generates the rest of the sequence. For this to work, the SIT encoder must be scale-causal to properly tokenize low-resolution inputs, while there is no constraint for the decoder so we use dense attention. We denote this variant Scale-Causal Encoder, or \"SIT-SCE\". The model is otherwise as the one used in Sec. 5.2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Here the inputs are \\(32 \\times 32\\). Fig. 4 shows text-guided upsampling examples on MS-COCO [27]. We obtain an FID of 6.2 when evaluating these generations, compared to 12.6 when given only the prompts. Fig. 8 and Tab. 4 show additional results upsampling from \\(16 \\times 16\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.772, + 0.327, + 0.789 + ], + "angle": 0, + "content": "5.4. Text-guided image editing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Our coarse-to-fine AR-SIT enables a text-guided image editing application, where we want to change image details while keeping the same overall appearance, which corresponds to freezing lower scales while generating higher." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We apply an AR-SIT trained only on the maximum likelihood objective as follows. The given image is tokenized only up to the first scale, corresponding to a coarse repre" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.499, + 0.907, + 0.559 + ], + "angle": 0, + "content": "sensation. The tokenizer encoder must be scale-causal in order to avoid leakage from high to low scales, so we use SIT-SCE here. Now we use AR-SIT to generate the rest of the sequence, conditioning on the textual caption." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.561, + 0.907, + 0.667 + ], + "angle": 0, + "content": "In this experiment, we use a 5-scale model such that the lowest resolution is \\(16 \\times 16\\); we found that starting with higher resolutions limits the changes the model can generate. Fig. 5 shows some examples of text-guided editing on MS-COCO [27], where we lightly modify the original captions, for example by swapping \"elephants\" with \"cows\". We show additional results in Fig. 6 in the supplemental." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.82, + 0.696 + ], + "angle": 0, + "content": "5.5. Class-conditional image generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.907, + 0.794 + ], + "angle": 0, + "content": "Here we evaluate class-conditional generation on \\(256 \\times 256\\) ImageNet. We use the SIT-4 tokenizer from Tab. 1 and follow ViT-VQGAN's [57] VIM-Base AR architecture and training protocol. Tab. 3 shows we outperform the baseline in this fair comparison which is evidence for the improved conditioning described in P.3. Fig. 9 shows samples." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.901 + ], + "angle": 0, + "content": "AR-SIT-4* departs from ViT-VQGAN and Parti hyperparameters to compare with more recent works, by increasing the perceptual and adversarial loss weights, doubling the tokenizer codebook, reducing its latent dimension, training with a constant learning rate, including 2D RoPE [44] and GeGLU [42], in both the tokenizer and generator. We train the AR model longer and since there is no text encoder, we" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17187" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.225, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.09, + 0.356, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.09, + 0.488, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.191, + 0.362, + 0.203 + ], + "angle": 0, + "content": "\"a close-up of a dog face.\"" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.212, + 0.225, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.212, + 0.356, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.212, + 0.488, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.184, + 0.308, + 0.399, + 0.32 + ], + "angle": 0, + "content": "\"a couple of cows are standing in a field\"" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.324, + 0.225, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.324, + 0.356, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.324, + 0.488, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.201, + 0.426, + 0.381, + 0.437 + ], + "angle": 0, + "content": "\"a robot on a grass covered field.\"" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.09, + 0.64, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.09, + 0.772, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.09, + 0.904, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.191, + 0.771, + 0.202 + ], + "angle": 0, + "content": "\"some boats on a lake.\"" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.207, + 0.64, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.207, + 0.772, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.207, + 0.904, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.602, + 0.308, + 0.812, + 0.32 + ], + "angle": 0, + "content": "\"a green pasture filled with wildflowers.\"" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.324, + 0.64, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.324, + 0.772, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.324, + 0.904, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.59, + 0.425, + 0.824, + 0.438 + ], + "angle": 0, + "content": "\"a very cute dog laying in a very small sink.\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.448, + 0.907, + 0.504 + ], + "angle": 0, + "content": "Figure 5. Text-guided image editing on MS-COCO [27]. Our coarse-to-fine generative models can do text-guided editing by encoding a given image but keeping only the lower scales, and using a pre-trained AR-SIT to re-generate the higher scales conditioned on the textual prompt. Each triplet shows the given image, its \\(16 \\times 16\\) reconstruction using only the coefficients used to start the generation, and the edited image after generating the whole sequence." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.511, + 0.482, + 0.649 + ], + "angle": 0, + "content": "
paramsseq lenCFGFID ↓IS ↑
AR-ViT-VQGAN (reported)650M1024-8.81110.8
AR-ViT-VQGAN (reproduced)650M1024-8.37111.8
AR-SIT-4 (Ours)650M1024-6.95138.3
RQ-VAE [26]480M256-15.7286.8
DQ-VAE [22]355M640-7.34152.8
LlamaGen-L [45]343M2561.54.08198.5
VAR [48]310M6802.03.30274.4
AR-SIT-4* (Ours)350M10241.54.06190.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.661, + 0.483, + 0.773 + ], + "angle": 0, + "content": "Table 3. ImageNet \\(256 \\times 256\\) class conditional generation. Top half shows a fair comparison against ViT-VQGAN, which we clearly outperform. AR-SIT-4* uses different hyperparameters for comparison against a broader class of methods. While we do not outperform all of them, there are differences in the architectures and training schedules that influence the results and are orthogonal to our contributions. Moreover, none of the baselines has the multiscale capabilities described in P.2 and P.4 (Sec. 1)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.902 + ], + "angle": 0, + "content": "double the number of decoder layers. The bottom half of Tab. 3 compares with similarly-sized AR models including some with multiscale/residual latents (although none exhibit properties P.2 and P.4). AR-SIT-4* performance matches LlamaGen but not to VAR, but the models are not exactly comparable since they use a convolutional tokenizers and other improvements such as AdaLN [34] and attention normalization; VAR also trains the tokenizer on larger" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.513, + 0.905, + 0.543 + ], + "angle": 0, + "content": "data than ImageNet. These are orthogonal to our contributions and can be incorporated in future work." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.906, + 0.607 + ], + "angle": 0, + "content": "We also train an AR-SIT-5* on \\( 512 \\times 512 \\) ImageNet and obtain an FID of 5.97 and IS of 215.22, showing robustness at higher resolutions, while most prior work requires separate stages for upsampling. Fig. 10 shows samples." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.628, + 0.764, + 0.644 + ], + "angle": 0, + "content": "6. Conclusion and limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.906, + 0.763 + ], + "angle": 0, + "content": "We presented a spectral image tokenizer (SIT), and an autoregressive generative transformer trained with it (AR-SIT). SIT is naturally multiscale and leverages spectral properties of natural images for improved reconstruction quality. AR-SIT enables applications such as rapid generation of coarse images that can be refined later, and text-guided image upsampling and editing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.902 + ], + "angle": 0, + "content": "While our methods improved tokenizer reconstruction accuracy and class-conditional generation, the text-to-image metrics were similar to the fair baseline. Chang et al. [4] similarly observed that a better tokenizer does not necessarily lead to a better generative model. Nevertheless, our method has multiscale properties and enables new applications not possible with prior work. We only experiment with realtively small AR-SITs of 350-650M parameters, while the Parti [58] baseline goes up to 22B parameters." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17188" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.484, + 0.162 + ], + "angle": 0, + "content": "We thank Leonardo Zepeda-Nuñez for reviewing this manuscript and offering interesting discussions and suggestions, and Jon Barron for sharing useful code we relied on." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.177, + 0.188, + 0.193 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.202, + 0.484, + 0.272 + ], + "angle": 0, + "content": "[1] James Betker, Gabriel Goh, Li Jing, † TimBrooks, Jianfeng Wang, Linjie Li, † LongOuyang, † JuntangZhuang, † JoyceLee, † YufeiGuo, † WesamManassra, † PrafullaDhariwal, † CaseyChu, † YunxinJiao, and Aditya Ramesh. Improving image generation with better captions. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.274, + 0.484, + 0.327 + ], + "angle": 0, + "content": "[2] Lei Cai, Hongyang Gao, and Shuiwang Ji. Multi-Stage Variational Auto-Encoders for Coarse-to-Fine Image Generation, page 630–638. Society for Industrial and Applied Mathematics, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.33, + 0.483, + 0.371 + ], + "angle": 0, + "content": "[3] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T. Freeman. Maskgit: Masked generative image transformer. In CVPR, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.374, + 0.484, + 0.456 + ], + "angle": 0, + "content": "[4] Huiwen Chang, Han Zhang, Jarred Barber, Aaron Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Patrick Murphy, William T. Freeman, Michael Rubinstein, Yuanzhen Li, and Dilip Krishnan. Muse: Text-to-image generation via masked generative transformers. In ICML, 2023. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.459, + 0.484, + 0.5 + ], + "angle": 0, + "content": "[5] XI Chen, Nikhil Mishra, Mostafa Rohaninejad, and Pieter Abbeel. PixelSNAIL: An improved autoregressive generative model. In ICML, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.502, + 0.485, + 0.64 + ], + "angle": 0, + "content": "[6] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. Pali: A jointly-scaled multilingual language-image model, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.643, + 0.484, + 0.712 + ], + "angle": 0, + "content": "[7] C. A. Christopoulos, T. Ebrahimi, and A. N. Skodras. Jpeg2000: the new still picture compression standard. In Proceedings of the 2000 ACM Workshops on Multimedia, page 45-49, New York, NY, USA, 2000. Association for Computing Machinery. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.714, + 0.484, + 0.768 + ], + "angle": 0, + "content": "[8] A. Cohen, Ingrid Daubechies, and J.-C. Feauveau. Biorthogonal bases of compactly supported wavelets. Communications on Pure and Applied Mathematics, 45(5):485-560, 1992. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.771, + 0.483, + 0.799 + ], + "angle": 0, + "content": "[9] Ingrid Daubechies. Ten lectures on wavelets. Society for Industrial and Applied Mathematics, USA, 1992. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.801, + 0.483, + 0.841 + ], + "angle": 0, + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.844, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[11] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[12] Sander Dieleman. Diffusion is spectral autoregression, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.202, + 0.485, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.175 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.23 + ], + "angle": 0, + "content": "[14] Tim Elsner, Paula Usinger, Victor Czech, Gregor Kobsik, Yanjiang He, Isaak Lim, and Leif Kobbelt. Quantised global autoencoder: A holistic approach to representing visual data, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[15] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.356 + ], + "angle": 0, + "content": "[16] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[17] Rinon Gal, Dana Cohen Hochberg, Amit Bermano, and Daniel Cohen-Or. Swagan: a style-based wavelet-driven generative model. ACM TOG, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.439 + ], + "angle": 0, + "content": "[18] Jiatao Gu, Shuangfei Zhai, Yizhe Zhang, Joshua M. Susskind, and Navdeep Jaitly. Matryoshka diffusion models. In ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.508 + ], + "angle": 0, + "content": "[19] Philippe Hansen-Estruch, David Yan, Ching-Yao Chung, Orr Zohar, Jialiang Wang, Tingbo Hou, Tao Xu, Sriram Vishwanath, Peter Vajda, and Xinlei Chen. Learnings from scaling visual tokenizers for reconstruction and generation, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.565 + ], + "angle": 0, + "content": "[20] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[21] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.597, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[22] Mengqi Huang, Zhendong Mao, Zhuowei Chen, and Yongdong Zhang. Towards accurate image coding: Improved autoregressive image generation with dynamic vector quantization. In CVPR, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.799, + 0.664 + ], + "angle": 0, + "content": "[23]Imagen-Team-Google. Imagen 3, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[24] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[25] Vladimir Kulikov, Shahar Yadin, Matan Kleiner, and Tomer Michaeli. Sinddm: A single image denoising diffusion model. In ICML, pages 17920-17930, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[26] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In CVPR, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[27] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In ECCV. 5, 6, 7, 8, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.887 + ], + "angle": 0, + "content": "[28] Xiaoxiao Ma, Mohan Zhou, Tao Liang, Yalong Bai, Tiejun Zhao, Huaian Chen, and Yi Jin. Star: Scale-wise text-to-image generation via auto-regressive representations, 2024." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.887 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17189" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.134 + ], + "angle": 0, + "content": "[29] Stephane Mallat. A Wavelet Tour of Signal Processing, Third Edition: The Sparse Way. Academic Press, Inc., USA, 3rd edition, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.484, + 0.176 + ], + "angle": 0, + "content": "[30] Wael Mattar, Idan Levy, Nir Sharon, and Shai Dekel. Wavelets are all you need for autoregressive image generation, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.484, + 0.219 + ], + "angle": 0, + "content": "[31] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. In ICLR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.484, + 0.261 + ], + "angle": 0, + "content": "[32] Charlie Nash, Jacob Menick, Sander Dieleman, and Peter W. Battaglia. Generating images with sparse representations. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.262, + 0.384, + 0.276 + ], + "angle": 0, + "content": "[33] OpenAI. Gpt-4 technical report, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.484, + 0.304 + ], + "angle": 0, + "content": "[34] William Peebles and Saining Xie. Scalable diffusion models with transformers. 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.484, + 0.347 + ], + "angle": 0, + "content": "[35] Hao Phung, Quan Dao, and Anh Tran. Wavelet diffusion models are fast and scalable image generators. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.484, + 0.403 + ], + "angle": 0, + "content": "[36] Vivek Ramanujan, Kushal Tirumala, Armen Aghajanyan, Luke Zettlemoyer, and Ali Farhadi. When worse is better: Navigating the compression-generation tradeoff in visual tokenization, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.484, + 0.445 + ], + "angle": 0, + "content": "[37] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. In NeurIPS, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.447, + 0.484, + 0.489 + ], + "angle": 0, + "content": "[38] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.484, + 0.531 + ], + "angle": 0, + "content": "[39] Tamar Rott Shaham, Tali Dekel, and Tomer Michaeli. Singan: Learning a generative model from a single natural image. In ICCV, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.532, + 0.484, + 0.56 + ], + "angle": 0, + "content": "[40] Dohoon Ryu and Jong Chul Ye. Pyramidal denoising diffusion probabilistic models, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.484, + 0.603 + ], + "angle": 0, + "content": "[41] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, Xi Chen, and Xi Chen. Improved techniques for training gans. In NeurIPS, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.472, + 0.617 + ], + "angle": 0, + "content": "[42] Noam Shazeer. Glu variants improve transformer, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.484, + 0.66 + ], + "angle": 0, + "content": "[43] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.66, + 0.484, + 0.702 + ], + "angle": 0, + "content": "[44] Jianlin Su, Yu Lu, Shengfeng Pan, Bo Wen, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.484, + 0.759 + ], + "angle": 0, + "content": "[45] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.484, + 0.787 + ], + "angle": 0, + "content": "[46] Gemini Team. Gemini: A family of highly capable multimodal models, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.437, + 0.802 + ], + "angle": 0, + "content": "[47] Llama team. The llama 3 herd of models, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[48] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. In NeurIPS, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[49] Aaron van den Oord, Nal Kalchbrenner, Lasse Espeholt, kory kavukcuoglu, Oriol Vinyals, and Alex Graves. Conditional image generation with pixelCNN decoders. In NeurIPS, 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[50] Aäron van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In ICML, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[51] Aaron van den Oord, Oriol Vinyals, and koray kavukcuoglu. Neural discrete representation learning. In NeurIPS, 2017. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[53] Xin Wen, Bingchen Zhao, Ismail Elezi, Jiankang Deng, and Xiaojuan Qi. \"principal components\" enable a new language of images, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[54] Jingfeng Yao, Bin Yang, and Xinggang Wang. Reconstruction vs. generation: Taming optimization dilemma in latent diffusion models. In CVPR, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.306, + 0.905, + 0.349 + ], + "angle": 0, + "content": "[55] Ting Yao, Yingwei Pan, Yehao Li, Chong-Wah Ngo, and Tao Mei. Wave-vit: Unifying wavelet and transformers for visual representation learning. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[56] Hu Yu, Hao Luo, Hangjie Yuan, Yu Rong, and Feng Zhao. Frequency autoregressive image generation with continuous tokens, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.448 + ], + "angle": 0, + "content": "[57] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved VQGAN. In ICLR, 2022. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[58] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research, 2022. Featured Certification. 1, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[59] Lijun Yu, Jose Lezama, Nitesh Bharadwaj Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G Hauptmann, Boqing Gong, Ming-Hsuan Yang, Irfan Essa, David A Ross, and Lu Jiang. Language model beats diffusion - tokenizer is key to visual generation. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.631, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[60] Qihang Yu, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An image is worth 32 tokens for reconstruction and generation. In NeurIPS, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.687, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[61] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.731, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[62] Zhenhai Zhu and Radu Soricut. Wavelet-based image tokenizer for vision transformers, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.757 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17190" + } + ] +] \ No newline at end of file diff --git a/2025/Spectral Image Tokenizer/2b3586c0-b936-4d11-a38a-2ff8e5859a65_origin.pdf b/2025/Spectral Image Tokenizer/2b3586c0-b936-4d11-a38a-2ff8e5859a65_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..05df9df0d45e20aa528b6119f7930cbbd9fd24c2 --- /dev/null +++ b/2025/Spectral Image Tokenizer/2b3586c0-b936-4d11-a38a-2ff8e5859a65_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:74f0e24dbee38005fa556fd1978966797730982b982215866b85c4b719f0fa96 +size 17031196 diff --git a/2025/Spectral Image Tokenizer/full.md b/2025/Spectral Image Tokenizer/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c03c8ae60d97a66fc600438d9c15b09b59b356fe --- /dev/null +++ b/2025/Spectral Image Tokenizer/full.md @@ -0,0 +1,405 @@ +# Spectral Image Tokenizer + +Carlos Esteves + +Mohammed Suhail + +Ameesh Makadia + +{machc, suhailmhd, makadia}@google.com + +Google Research + +# Abstract + +Image tokenizers map images to sequences of discrete tokens, and are a crucial component of autoregressive transformer-based image generation. The tokens are typically associated with spatial locations in the input image, arranged in raster scan order, which is not ideal for autoregressive modeling. In this paper, we propose to tokenize the image spectrum instead, obtained from a discrete wavelet transform (DWT), such that the sequence of tokens represents the image in a coarse-to-fine fashion. Our tokenizer brings several advantages: 1) it leverages that natural images are more compressible at high frequencies, 2) it can take and reconstruct images of different resolutions without retraining, 3) it improves the conditioning for next-token prediction – instead of conditioning on a partial line-by-line reconstruction of the image, it takes a coarse reconstruction of the full image, 4) it enables partial decoding where the first few generated tokens can reconstruct a coarse version of the image, 5) it enables autoregressive models to be used for image upsampling. We evaluate the tokenizer reconstruction metrics as well as multiscale image generation, text-guided image upsampling and editing. + +# 1. Introduction + +In natural language processing, tokenization associates sets of characters to entries in a vocabulary, which is straightforward since language is discrete. The sequence of tokens is suitably modeled as a categorical distribution with autoregressive (AR) transformers, which is the foundation of modern large language models (LLMs) [33, 46, 47]. + +While natural images are represented by discrete pixel values, they exhibit high dimensionality, redundancies, and noise so it's impractical to associate one token per pixel. This motivated a long line of research of learnable image tokenizers [15, 37, 37, 51, 57]. While there are successful autoregressive image generation models [15, 57, 58], + +images are not sequential like language, which motivated developing alternatives such as denoising diffusion models [11, 38, 43] and masked transformers [3, 4, 60]. Nevertheless, most of these alternatives also operate on the latent space of tokenizers like VQGAN [15] instead of raw pixels. + +In this work, we revisit AR transformer-based image generation. Our main contribution is a tokenizer operating on the image spectrum, specifically on DWT coefficients, where the coarse-to-fine representation lends itself more naturally to a sequential interpretation. Our Spectral Image Tokenizer (SIT) has several useful properties and enables different applications: + +P.1 Since the power spectrum of natural images decreases with frequency, high frequencies can be more heavily compressed with little effect in visual quality. SIT leverages this by associating tokens to larger patches at higher wavelet scales than at lower scales (see Fig. 2). +P.2 SIT is transformer-based [52]; by using an attention mask where each scale depends on itself and lower scales ("Scale-Causal attention"), SIT can be trained at a single resolution and used to tokenize images of multiple resolutions (any number of scales up to the trained maximum), and detokenize partial token sequences (up to some scale), reconstructing a coarse image. +P.3 Using SIT, we train an autoregressive generative transformer (AR-SIT) that models images coarse-to-fine. The next-token prediction is then conditioned on a coarse reconstruction of the image given by the partial token sequence, instead of the usual conditioning on the partial reconstruction of the previous rows of the image. +P.4 AR-SIT can quickly generate only the first few tokens and reconstruct a coarse version of the image, enabling applications like quickly showing multiple coarse generations and letting the user select which ones to refine. +P.5 AR-SIT can be used for text-based upsampling of an input low resolution image, by starting the decoding process with the few tokens output by SIT, and generating the rest of the sequence up to a desired resolution. +P.6 AR-SIT can be used for text-guided image editing, by + +![](images/2bfa86016e0efaaf3b1d858e61d41b14be8218e91de76c2783d796f1bade2604.jpg) +Figure 1. Left: we introduce a Spectral Image Tokenizer (SIT), that learns to encode and decode discrete wavelet transform (DWT) coefficients to and from a small set of discrete tokens, such that the sequence represents the image in a coarse-fine fashion. SIT is naturally multiscale and enables coarse-to-fine autoregressive image generation with our AR-SIT model. SIT also leverages the sparsity of high frequency coefficients in natural images. Right: details of the encoder/decoder transformer layers. The main architectural difference with respect to previous tokenizers is that the distributions of DWT approximation and details coefficients are distinct, hence we use specialized parameters for each in the quantizer codebooks and inner transformer layers. We also introduce a scale-causal attention where each token attends to its own scale and lower scales, which enables encoding, decoding, generating, and upsampling images at different resolutions. + +![](images/db8c6463a1c65f8aaf7b1abc40644785cb3a11c1ae22bbb2fa368a367c846a68.jpg) + +encoding a given image up to a coarse scale, and generating the finer details conditioned on a new caption. + +Currently, image generation is dominated by diffusion models such as Imagen 3 [23], DALL-E-3 [1] and Stable Diffusion 3 [16]. On the other hand, LLMs such as Gemini [46], GPT-4 [33], and Llama 3 [47] are based on autoregressive transformers. We believe autoregressive image generation is still worth pursuing because it might benefit from advances in LLMs, and multimodal applications might benefit of having a single architecture for all modalities. Recent work on image and video generation support this point [45, 48, 59]. Dieleman [12] recently interpreted denoising diffusion models as spectral autoregression, since, when looking at image spectra, the denoising procedure uncovers frequencies from low to high. In contrast, our method does literal spectral autoregression. + +# 2. Related work + +Image tokenization Several methods have been developed to map images to a small set of discrete tokens suitable for generative modeling. VQ-VAE [51] introduced vector-quantization in the latent space of a Variational AutoEncoder to map images, audio and video to a set of discrete values. VQGAN [15] improved upon VQVAE by using perceptual and adversarial losses. We build on ViT-VQGAN [57], which improved upon VQGAN by using a Vision Transformer (ViT) [13] instead of convolutions, as well as codebook factorization and feature normalization. + +In this paper, we are interested in multiscale image representations. VQ-VAE-2 [37] introduced multiscale latents + +by keeping and quantizing intermediate downsampled convolutional features. RQ-VAE [26] quantized a set of residuals such that the latent vector is represented in a coarse-to-fine fashion and reconstructed by adding the code embeddings for each residual. Similarly, VAR [48] and STAR [28] split the latent space in multi-scale quantized residuals, so that it can be reconstructed by upsampling and summing. FAR [56] splits the latents into frequencies to train a next-frequency generative model. DQ-VAE [22] encoded smooth regions with fewer tokens that are generated first, while QG-VAE [14] and SEMANTICIST [53] learn non-local tokens that are ordered by importance. + +The crucial difference between our approach and the aforementioned is that we operate on spectral coefficients of the input and not on latent features. This allows our tokenizer to be truly multiscale; it can take inputs at different scales and reconstruct up to some scale, and leverage that higher frequencies are more compressible in natural images. + +Tangentially related to our work, Wave-ViT [55] modified the ViT self-attention by applying a DWT to the input and concatenating coefficients, effectively exchanging space for depth to reduce the sequence length. Zhu and Soricut [62] modified the ViT patchifier in a similar way, and introduced patch embeddings that leverage the sparsity of high frequency coefficients. Both methods are for discriminative tasks such as image classification and segmentation, while we focus on generation. + +Autoregressive image generation Early approaches PixelRNN and PixelCNN [49, 50] generate images pixel by pixel by modeling the conditional distribution of the pixel given the previous pixels with recurrent layers of causal + +convolutions. PixelSNAIL [5] improved on this model by introducing self-attention layer to better model long-range dependencies. VQ-VAE [51] introduced a two-stage approach with a tokenizer and a separate stage to model the distribution of tokens. VQGAN [15] greatly improved these results by both improving the tokenizer and using a transformer to model the distribution. Finally, ViT-VQGAN [57] proposed a transformer-based tokenizer, which Parti [58] used with a large autoregressive transformer capable of high-quality text-to-image generation. + +MaskGIT [3] and Muse [4] highlighted the disadvantages of the typical autoregressive models raster-order conditioning, and proposed to generate all tokens in parallel iteratively, where each iteration keeps the highest confidence tokens. We address the same problem with a tokenizer whose sequence represents the image in a coarse-to-fine order instead of raster-order. + +Multiscale image generation Multiscale image generation ideas have appeared in the context of VAEs [2], GANs [24, 39], and diffusion models [18, 25, 40], but have not been sufficiently explored with AR transformers. + +Nash et al. [32] represented an image as a sequence of quantized and thresholded DCT coefficients, where the compression comes from the fact that many coefficients are zero and are omitted. By sorting the sequence by frequency, the method can model images from coarse-to-fine like we do. However, the compressed representation is handcrafted and results in long sequences. In a similar vein, Mattar et al. [30] handcrafted a tokenizer based on DWT coefficients, introducing tokens to represent large chunks of zeros. It also results in long sequences and is only applied to generations of small grayscale images. In contrast with these methods, instead of handcrafting a compressed input representation, our tokenizer learns to encode to, and decode from, a short compressed coarse-to-fine sequence. + +# 3. Background + +A discrete wavelet transform (DWT) is based on successive convolutions of the signal $f$ with a pair of lowpass $g$ and highpass $h$ filters with the first step as follows + +$$ +f _ {\operatorname {l o w} _ {1}} [ n ] = f \star g = \sum_ {k} f [ k ] g [ n - k ], \tag {1} +$$ + +$$ +f _ {\operatorname {h i g h} _ {1}} [ n ] = f \star h = \sum_ {k} f [ k ] h [ n - k ]. \tag {2} +$$ + +The high and low outputs are subsampled by a factor of two, and the operation is repeated for the low channel, such that at level $L$ we compute $f_{\mathrm{low}_L} = f_{\mathrm{low}_{L - 1}} \star g$ and $f_{\mathrm{high}_L} = f_{\mathrm{low}_{L - 1}} \star h$ , subsample again and drop $f_{\mathrm{low}_{L - 1}}$ . The output at level $L$ comprises the approximation coefficients $f_{\mathrm{low}_L}$ and the detail coefficients $\{f_{\mathrm{high}_k}\}$ for $1 \leq k \leq L$ . This output has the same cardinality as the input and the + +![](images/863d7df0e87b9ef660d64f4c69a21c033d45e2e444801ba5802f9b521fe6710d.jpg) +Figure 2. Input patchification. Left: typical patchification for Vision Transformers (ViT) [13], where the image is split in equal-sized patches. Right: we propose to patchify the coefficients of a discrete wavelet transform (DWT) instead. Each scale is shown in a different color. Scales other than the lowest contain three blocks representing horizontal, vertical and diagonal details; we concatenate the spatially correspondent patches of each block such that each scale is represented by the same sequence length. This results in larger patch sizes for higher scales, which are more compressible. The figure shows 3 scales and 16 tokens per scale; in our experiments we use 4 or 5 scales and 256 tokens per scale. + +![](images/3d53cd99caa83e3e495af4cbd9515dbd49fa7a9103b690f6877a7ea3f9b4392f.jpg) + +transformation is invertible, where the forward transform is typically called "analysis" and the backward "synthesis". The simplest wavelet family is the Haar, where $g = [1,1]^{\top}$ and $h = [1, -1]^{\top}$ (optionally scaled to unit norm). + +For image analysis we use a 2D DWT, which is obtained by simply convolving the rows and columns with $g$ and $h$ . The approximation coefficients $f_{\mathrm{low}_1} = f \star (gg^\top)$ , and the details are divided into horizontal $f_{\mathrm{H}_1} = f \star (gh^\top)$ , vertical $f_{\mathrm{V}_1} = f \star (hg^\top)$ and diagonal $f_{\mathrm{D}_1} = f \star (hh^\top)$ details. Subsequent levels apply the same operations to the approximation coefficients. Fig. 2 shows a two-level transform, where we can see that the approximation coefficients correspond to a coarse version of the image. + +We refer to the textbooks by Mallat [29] and Daubechies [9] for more information about wavelets. + +# 4. Method + +Our main contribution, SIT, is an image tokenizer that takes discrete wavelet transform (DWT) coefficients. The model follows ViT-VQGAN [57], with important changes that we describe in this section and visualize in Fig. 1. For image generation, we introduce AR-SIT which is based on Parti [58] with minor changes, using SIT as the tokenizer. + +# 4.1. Tokenizer + +Patchification and featurization The first step is to map the input to a sequence of patches. We apply the Haar DWT on the input image and patchify each scale separately. While Haar is the simplest wavelet and lacks properties found in other wavelets useful for compression, we found no benefits of using other wavelet families such as CDF [8]. + +Our design choice is to use the same number of patches for each scale. In a DWT, the higher scales correspond to high frequency details which are represented by more coefficients than the lower scales, but contribute less to the spatial pixel values. In other words, in natural images, most of the power spectrum is concentrated on lower frequencies. By representing each scale with the same number of tokens, we are compressing more the higher frequencies (since they are represented by more coefficients), similarly to what is done in image compression methods such as JPEG2000 [7]. + +The approximation (or lowpass) DWT coefficients correspond to a coarse version of the input image, which we consider the first scale. The following scales are divided in three blocks corresponding to horizontal, vertical, and diagonal details, where each coefficient relates to a specific spatial location. Thus, we can concatenate the three blocks such that each entry corresponds to the same spatial location. For example, the first scale will typically be split in patches that are $32 \times 32 \times 3$ , with channels corresponding to RGB, the second scale will be $32 \times 32 \times 9$ , where the channels correspond to the RGB of horizontal, vertical, and diagonal details, the third will be $64 \times 64 \times 9$ and so on. Fig. 2 shows an example of our patchification scheme. + +Since patches have different resolutions (higher scales will have larger patches), the usual ViT linear embedding to map patches to features cannot be shared across all patches so we have different parameters per scale. Formally, given an image $f$ , our patchifiers compute + +$$ +f _ {\text {l o w} L}, \left\{f _ {\mathrm {H} _ {i}} \right\} _ {i \leq L}, \left\{f _ {\mathrm {V} _ {i}} \right\} _ {i \leq L}, \left\{f _ {\mathrm {D} _ {i}} \right\} _ {i \leq L} = \mathrm {D W T} (f), \tag {3} +$$ + +$$ +c _ {1} = P _ {0} \left(f _ {\text {l o w} L}\right), \tag {4} +$$ + +$$ +c _ {s} = P _ {s} \left(f _ {\mathrm {H} _ {L - s + 2}}, f _ {\mathrm {V} _ {L - s + 2}}, f _ {\mathrm {D} _ {L - s + 2}}\right), 1 < s \leq L, \tag {5} +$$ + +where $c_{s} = \{c_{s}^{n}\}_{1\leq n\leq N}$ is the sequence of token embeddings at scale $s$ , and $c_{s}^{n}\in \mathbb{R}^{C}$ is the embedding of the $n$ -th token at the $s$ -th scale, while $L$ is the number of DWT levels, $S = L + 1$ is the number of scales and $N$ the number of tokens per scale. For brevity, when there is no ambiguity, we may omit the set indexing and use $\{c_s\}$ to denote the set of tokens of all scales, for example. + +The final projection after the decoder needs to map the features back to different sized patches, so it also has different parameters per scale. Those patches still represent DWT coefficients so an inverse DWT (IDWT) is finally applied to obtain an image output. + +Flexible sequence length The tokenizer encoder and decoder transformers operate on the sequence of patch features of length $SN$ . The sequence length is a major factor in resource utilization so we want to keep it constrained. Our method is flexible since we can choose the number of scales and the patch size per scale, while most ViT-based models such as ViT-VQGAN [57] are more restricted. They use the same patch size for the whole image; thus, keeping the + +same patch size and doubling each image dimension would increase the sequence length by a factor of 4, where our method is capable of including additional scales which only increases the sequence length by multiples of $N$ . + +For example, for a $256 \times 256$ input, ViT-VQGAN uses $8 \times 8$ patches to obtain a sequence of 1024 elements. Our SIT can use 4 scales and 256 tokens per scale for the same sequence length. When increasing the resolution to $512 \times 512$ , the baseline can either increase the patch size to $16 \times 16$ , resulting in the same sequence length, or keep it $8 \times 8$ , resulting in a $4 \times$ longer sequence. SIT, for example, can vary the number of scales from 4 to 6, resulting in sequence lengths of 1024, 1280, 1536 which are more manageable. + +Transformers After featurization, the single sequence containing all scales passes through a transformer encoder, followed by quantization and a transformer decoder. We propose two optional modifications to the transformers, which are otherwise identical to the ones used in ViT [13]. + +First, we propose a scale-causal attention mask, where an element at some scale attends to all elements of its own scale and lower scales, represented by a block-triangular mask pattern. With dense attention, we write the application of the encoding transformer $T_{\mathrm{enc}}$ as $\{z_s\} = T_{\mathrm{enc}}(\{c_s\})$ , and for scale-causal we write $z_s = T_{\mathrm{enc}}(\{c_k\}_{1 \leq k \leq s})$ for each $s$ . The scale-causal attention can be applied independently to the encoder and decoder, enabling different applications. For the multiscale reconstruction experiments in Sec. 5.1, we need to both encode and decode multiple resolutions, so both encoder and decoder use scale-causal masks. For coarse-to-fine image generation in Sec. 5.2, only the decoder needs to be scale-causal in order to decode the partially generated sequence. For the text-guided image upsampling in Sec. 5.3, only the encoder needs to be scale-causal to encode the lower resolution inputs. For the image editing experiments in Sec. 5.4, the scale-causal encoder prevents information leaking from high to low scale. + +Second, we propose to use different transformer parameters for the approximation (first scale) and details coefficients (other scales). In our model, the subsequences corresponding to coefficients of each type of coefficient come from quite distinct distributions, so it makes sense to treat them differently. This contrasts with the spatial representation of images where each patch can be considered as coming from the same distribution. Thus, the parameters of the key, query, and value embeddings, the layer norms, and the MLP on each transformer layer are not shared between the approximation and details coefficients. This Approximation-Details Transformer (ADTransformer) still takes a single sequence composed of all coefficients. We experimented with different transformers per sequence with cross-attention for information sharing, but it performed worse. This change leads to more memory utilization to store the extra parameters, but the training/inference speed + +
LPIPS ↓PSNR ↑L1 ↓FID ↓IS ↑images/s ↑
Resolution: 512 × 512
ViT-VQGAN0.32022.40.0426.92151.5593
SIT-5 (Ours)0.26022.00.0462.65192.0410
SIT-6 (Ours)0.23923.10.0401.74203.7320
Resolution: 256 × 256
ViT-VQGAN (reported)-24.80.0321.99184.4-
ViT-VQGAN (reproduced)0.16725.00.0312.33184.0-
ViT-VQGAN (no LL)0.16323.80.0381.20194.6626
SIT-4 (Ours)0.14424.00.0371.20199.5596
SIT-5 (Ours)0.13524.50.0350.97202.3411
SIT-SC-5 (Ours)0.16124.10.0371.33193.7411
Resolution: 128 × 128
ViT-VQGAN0.18526.30.0303.77117.3626
SIT-SC-5 (ours)0.15927.10.0272.13129.3582
Resolution: 64 × 64
ViT-VQGAN0.12928.80.0233.5321.0627
SIT-SC-5 (ours)0.11131.30.0171.3930.1847
Resolution: 32 × 32
ViT-VQGAN0.21423.30.045-3.7627
SIT-SC-5 (ours)0.02936.80.0100.313.5825
Resolution: 16 × 16
ViT-VQGAN0.12724.90.039-1.7627
SIT-SC-5 (ours)0.01341.30.0060.091.82620
+ +is similar because the number of operations is unchanged. + +Quantizer The encoder outputs a sequence of features that are quantized to a fixed-sized codebook, similarly to ViT-VQGAN [57], VQVAE [51], and VQGAN [15]. We modify the quantizer such that approximation and details have different codebooks for the same reasons discussed previously. Thus, codebook sizes and feature dimensions are the same as the ViT-VQGAN baseline, but we have different features for the same code at different positions. Formally, we apply $q_{s}^{n} = Q_{\mathrm{approx}}(z_{s}^{n})$ for $s = 1$ and $q_{s}^{n} = Q_{\mathrm{details}}(z_{s}^{n})$ for $s \geq 2$ for each pair $(s,n)$ , where $q_{s}^{n}$ is chosen from the codebook for $s$ and can be associated with its discrete position in the codebook, denoted $\lfloor q_s^n \rfloor$ . + +Training We follow the ViT-VQGAN training protocol and use the same weighting for the L2, perceptual, adversarial, and quantization losses. We remove the logit-laplace loss that was shown detrimental in follow-up work [58]. + +All losses are applied to the spatial domain images reconstructed by the inverse discrete wavelet transform on the decoder output: $\hat{f} = \mathrm{IDWT}(T_{\mathrm{dec}}(\{q_s\}))$ . We noticed insta + +Table 1. Multiscale reconstruction on ImageNet. "SC" denotes scale-causal attention, which slightly reduces performance at the highest resolution but enables multiscale reconstruction without downsampling/upsampling or retraining. The ViT-VQGAN values from Yu et al. [57] used a logit-laplace loss (LL) which was later considered harmful [58], so we retrain without it. Our SIT improves reconstruction metrics, is significantly faster at lower resolutions, and robust when increasing resolution. The ViT-VQGAN baseline suffered from instability during training on $512 \times 512$ inputs so we selected the best values before divergence. We report test time throughput of an encoding/decoding cycle for the max batch size that fits on a TPU v5e. + +
FID ↓IS ↑images/s ↑images/Gb ↑
Resolution: 256 × 256
Parti350M (reported)14.1---
Parti350M12.436.57.812.0
AR-SIT-SCD-4 (Ours)12.637.36.58.0
Resolution: 128 × 128
Parti350M11.233.57.612.0
AR-SIT-SCD-4 (Ours)11.433.212.612.0
Resolution: 64 × 64
Parti350M10.516.97.612.0
AR-SIT-SCD-4 (Ours)11.418.624.516.0
Resolution: 32 × 32
Parti350M5.82.97.77.7
AR-SIT-SCD-4 (Ours)7.63.274.728.0
+ +Table 2. Coarse-to-fine image generation on MS-COCO [27]. With SIT, the autoregressive generation is stopped early for a coarse version of the image. The Parti350M baseline does not have this property, so we generate at full resolution and downsample for comparison. AR-SIT matches the baseline performance at the training resolution but is several times faster and more memory efficient at lower resolutions, even when trained only on higher resolution data. We report throughput and memory utilization during generation given the max batch size that fits on a TPU v6e. + +bility during training due to the adversarial loss, which was fixed by applying spectral normalization following Miyato et al. [31], which simply divides the discriminator weight matrices by their largest singular value. + +# 4.2. Autoregressive image generation + +We use our tokenizer for autoregressive image generation, by training a second stage transformer model similar to Parti [58], with some modifications. Formally, the autoregressive transformer $T$ models categorical distributions over the discrete codes + +$$ +P \left(\left\lfloor q _ {s} ^ {n} \right\rfloor \mid \left\{\left\lfloor q _ {i} \right\rfloor \right\} _ {1 \leq i < s} \cup \left\{\left\lfloor q _ {s} ^ {i} \right\rfloor \right\} ^ {1 \leq i < n}\right) = T (c) \tag {6} +$$ + +which can be sampled one code at a time for generation. $T$ can be conditioned on a textual description processed by a transformer encoder for text-to-image generation. For training, we model the distribution of input codes as $P(\{\lfloor q_s\rfloor \}) = \prod_{1\leq s\leq S}^{1\leq n\leq N}P(\lfloor q_s^n\rfloor)$ and minimize the negative log-likelihood $-\log P(\{\lfloor q_s\rfloor \})$ over the training set. + +With different codebooks for approximation and details, the same token id might have a different meaning depending on its position. Thus, the AR model has different token embeddings for approximation and details tokens. The last layer for logit prediction is also different. + +For the generative applications in Secs. 5.2 and 5.3, we introduce mild changes in order to interrupt the generation after all tokens up to a certain scale are generated, and to start the generation with given tokens up to a certain scale. + +![](images/dc43ffe6e96642fcda4ec765bcd42878e8704676603cf14fde48296916c4cb76.jpg) +"a zebra standing on dirty area with trees in the background." + +![](images/e248b9896775e4fae5bad5b558a325a77445733fe2d81b49fc5fd2019b6f9431.jpg) +"an elaborate metal vase holds a decorative bouquet of flowers." + +![](images/de3eb6007c31642354f152e1bd512c4eea36619650b3f8dfad3b853f0ee1e278.jpg) +"the wing of a plane located in the sky above a sea and city." +Figure 3. Coarse-to-fine text-to-image on MS-COCO [27] prompts. Each quadruple shows generations from AR-SIT-SCD for the given prompt, with only the first $25\%$ , $50\%$ , $75\%$ and $100\%$ of tokens generated, corresponding to resolutions of $32 \times 32$ to $256 \times 256$ . Training uses only $256 \times 256$ images. Our model enables quick generation of coarse image candidates that can be further improved if needed. + +![](images/25862973e1bf70f60bd7575e0f87c543fe55309df8ecab772e174e9998dc63f2.jpg) +"a modern bathroom is shown with a square sink." + +![](images/f33eabf0788eab36bb4215f991b3e7ffd68e526e0ff44c2644f486c2dba10819.jpg) +"a toy boat sits on the ground." + +![](images/710941c16491f7e36dfe7d85aeba52cd9c1e9fd64813e64988728e31bbccaf38.jpg) +"This brown and white cat is standing on a desk in front of a computer." + +# 5. Experiments + +We focus our experiments on demonstrating the properties and applications enumerated in P.1-P.6 (Sec. 1). Our ideas build on ViT-based tokenizers, so the most fair comparisons are against ViT tokenizers and AR generative models; namely ViT-VQGAN [57] and Parti [58]. In these comparisons we are able to match the architectures and training protocol exactly. Nevertheless we also compare with a broader set of methods in Sec. 5.5. + +# 5.1. Multiscale reconstruction + +We train our tokenizer on ImageNet [10] and evaluate its reconstruction performance. We follow the ViT-VQGAN [57] "Small" architecture and training protocol. + +The model "SIT-5" is the Spectral Image Tokenizer with 5 scales from $16 \times 16$ to $128 \times 128$ resolutions, where each scale is represented by 256 tokens. The variation "SIT-SC" uses scale-causal attention on both the encoder and decoder, which enables handling inputs and outputs of different resolutions, even though it is only trained at $256 \times 256$ . For example, when the input image is $64 \times 64$ , only the first three scales are used, resulting in shorter sequence lengths which reduces memory utilization and processing time. Both models employ the ADTransformer described in Sec. 4.1. + +The ViT-VQGAN baseline only works for $256 \times 256$ inputs, so to evaluate against it fairly, we upsampled the low-resolution inputs to that resolution. This brings the input patches away from the training distribution, which might explain the drop in reconstruction quality. + +We also evaluate at $512 \times 512$ . The ViT-VQGAN baseline increases its patch size from $8 \times 8$ to $16 \times 16$ , keep- + +ing the sequence length constant. Interestingly, the baseline suffered heavy instability during training, which was not resolved by reducing the learning rate, using spectral normalization, or the logit-laplace loss. In contrast, our SIT variations trained successfully with no hyperparameter changes. + +Tab. 1 shows the reconstruction metrics; Fig. 7 shows reconstruction samples at multiple resolutions. Results provide evidence for P.1 and demonstrate P.2 (Sec. 1). + +# 5.2. Coarse-to-fine text-to-image generation + +We tackle text-to-image generation by using an autoregressive transformer to model the distribution of discrete tokens output by our tokenizer. Since the SIT sequence of tokens represents an image in a coarse-fine fashion, the autoregressive model generation has the same property, which means that we can interrupt the generation after a certain number of tokens and decode a coarse version of the generation. + +For this to work, the SIT decoder must be scale-causal, while there is no constraint for the encoder so we use dense attention (we denote this variant Scale-Causal Decoder, or "SIT-SCD"). It uses a "small" encoder and "large" decoder as described in ViT-VQGAN [57], and 4 scales. + +We follow the Parti350M [58] architecture and training protocol and name our generative model auto-regressive SIT, or "AR-SIT". SIT and AR-SIT are trained on a subset of WebLI [6] of around 128M images, where each image is seen at most once during training. Evaluation is on 30k examples of MS-COCO [27]. We show metrics and generations at different resolutions in Fig. 3 and Tab. 2. AR-SIT matches the baseline performance but is several times faster and more memory efficient at lower resolutions. These results demonstrate the property P.4 (Sec. 1). + +![](images/f6682868634b64cec23c47bc161205192494d27dd60c2e7eb70a3346c8aad53c.jpg) +"an assortment of some colorful vases on display on a table" + +![](images/6864d3bfd29c46ef47320f2be74f6fdfb48f5d5df51d90f044cab6eda868f28a.jpg) + +![](images/8b439c1c4445e8bb03f450c3747720ece9b8ca68277d231315d84b1d2d995342.jpg) + +![](images/517ebeae81c58a9d8c5eff6784c15ff05c0996997375cf294898fd3a0f22c118.jpg) +"one train riding on the multiple train tracks side by side." + +![](images/83ed571691f19876c737ebdc8743125915e5cd1d2239ea506d6fe5a93bb4bb8f.jpg) + +![](images/1a458d4fa8ab0ab3e5ed942ae2f094554c544dd10955bb5af9dcdb9da4d57813.jpg) + +![](images/3dd72cc01a64562a44a1ef23d88bfbc31770459cf258ae88a315876675ba6fcd.jpg) +"an image of a fired hydrant on the street" +Figure 4. Text-guided image upsampling on MS-COCO [27]. Our coarse-to-fine generative models can take a low-resolution image, encode it as the first few tokens of a sequence, and generate the rest of sequence, which, when decoded, effectively upsamples the input. Each triplet shows the given $32 \times 32$ image, our $256 \times 256$ reconstruction and the ground truth. + +![](images/841d651e744429d21dc2987e5397e2c7a88724fd690efd869b771cbe44f39aa7.jpg) + +![](images/ac38bdf79291ffd3ab1d63ed72d49ddcd4bdc91df8eaf45bdd23a839dc5e1b20.jpg) + +![](images/7d60852496ac37397b718e1b30f498f7a70098acac0efb310597ad1dab13ca2a.jpg) +"a large elephant standing on top of a grass covered field." + +![](images/7147839eb4f3185aab90b428d0b70471f566d9b819fde23f2cde8f026ed5c2d3.jpg) + +![](images/7ca9cd0198cae27ae5efdec7451fd2e87c474908da76f7885961a5e3d24e50d4.jpg) + +![](images/09890d806c9622da8906b588b07da60bbe2e63f1dbd4ee7306ab2a57116e4b3d.jpg) +"three tropical bird perched on top of high bare branches" + +![](images/59ac6e58680042f11f5f3a0119b5af6091a722258fd001e25739170221fc02b0.jpg) + +![](images/3587c8b075c0638cfffa457a21b7a195201df8f66d0c9cf958a65b203f57af7a.jpg) + +![](images/bea17c977521287e89bcf60654bbe228b5d095d91272b5755f17af6822cf67d5.jpg) +"a chair with a pink blanket is sitting on a red floor." + +![](images/843e97f846207e282a99007ba5204d3856f4d925c573a79c3727719130dec0b0.jpg) + +![](images/f367404a2ed76834b55032af67c3781a729a479906a959078f6746c6164bdd31.jpg) + +# 5.3. Text-guided image upsampling + +We leverage the coarse-to-fine nature of SIT to apply a pretrained AR-SIT for text-guided image upsampling, where we are given a low resolution image and a caption. + +The idea is to encode the low-resolution image, which will give the first tokens of our high resolution output. AR-SIT then takes these tokens and generates the rest of the sequence. For this to work, the SIT encoder must be scale-causal to properly tokenize low-resolution inputs, while there is no constraint for the decoder so we use dense attention. We denote this variant Scale-Causal Encoder, or "SIT-SCE". The model is otherwise as the one used in Sec. 5.2. + +Here the inputs are $32 \times 32$ . Fig. 4 shows text-guided upsampling examples on MS-COCO [27]. We obtain an FID of 6.2 when evaluating these generations, compared to 12.6 when given only the prompts. Fig. 8 and Tab. 4 show additional results upsampling from $16 \times 16$ . + +# 5.4. Text-guided image editing + +Our coarse-to-fine AR-SIT enables a text-guided image editing application, where we want to change image details while keeping the same overall appearance, which corresponds to freezing lower scales while generating higher. + +We apply an AR-SIT trained only on the maximum likelihood objective as follows. The given image is tokenized only up to the first scale, corresponding to a coarse repre + +sensation. The tokenizer encoder must be scale-causal in order to avoid leakage from high to low scales, so we use SIT-SCE here. Now we use AR-SIT to generate the rest of the sequence, conditioning on the textual caption. + +In this experiment, we use a 5-scale model such that the lowest resolution is $16 \times 16$ ; we found that starting with higher resolutions limits the changes the model can generate. Fig. 5 shows some examples of text-guided editing on MS-COCO [27], where we lightly modify the original captions, for example by swapping "elephants" with "cows". We show additional results in Fig. 6 in the supplemental. + +# 5.5. Class-conditional image generation + +Here we evaluate class-conditional generation on $256 \times 256$ ImageNet. We use the SIT-4 tokenizer from Tab. 1 and follow ViT-VQGAN's [57] VIM-Base AR architecture and training protocol. Tab. 3 shows we outperform the baseline in this fair comparison which is evidence for the improved conditioning described in P.3. Fig. 9 shows samples. + +AR-SIT-4* departs from ViT-VQGAN and Parti hyperparameters to compare with more recent works, by increasing the perceptual and adversarial loss weights, doubling the tokenizer codebook, reducing its latent dimension, training with a constant learning rate, including 2D RoPE [44] and GeGLU [42], in both the tokenizer and generator. We train the AR model longer and since there is no text encoder, we + +![](images/36ecb5f118fe41bcc2c59633b2689ae5406defc5a4d24fa7a3a57028d53d54c0.jpg) +"a close-up of a dog face." + +![](images/eaa16f39af62397d6cb9924f8124b28531b5cf0c7837c487bb5b87773e0227e3.jpg) + +![](images/8b7a5d722eedde08ddd90ded0d73e568da830eff1aa825e3f59c8feafb447751.jpg) + +![](images/e1bdf2e9ecfc4fc44af3a82b3207598dc95b2058bc453fd53d75c2e575b26e44.jpg) +"a couple of cows are standing in a field" + +![](images/4b686611eccd2b3eb076ee9970ff9de5fa2e2ee260cf20e00d6c7b996ea6c511.jpg) + +![](images/fba81ff6c5e4cb8a1d079101361afbc84072e7398dc85afe295e398c9d708fd4.jpg) + +![](images/17dafd257af776ffd48630fd378c65e9aa0eae746acc11a8c705d00d85eeb144.jpg) +"a robot on a grass covered field." +Figure 5. Text-guided image editing on MS-COCO [27]. Our coarse-to-fine generative models can do text-guided editing by encoding a given image but keeping only the lower scales, and using a pre-trained AR-SIT to re-generate the higher scales conditioned on the textual prompt. Each triplet shows the given image, its $16 \times 16$ reconstruction using only the coefficients used to start the generation, and the edited image after generating the whole sequence. + +![](images/15650ebe8f5fb1e14f5cbff4550b8905037edd0aa3d798bc3ff42b08b442e51b.jpg) + +![](images/5ee81e4ebd6fea49160ac75d835979d3bf656adeb5b1e4f03fcd0468c8bb7029.jpg) + +![](images/a4f5f01bcb766309cac7140a75cfd6400b2de4c276273946f37403fb1790c353.jpg) + +![](images/a269ec68a99d50ebee0f6707b77c78d00284cdcb065d5851444d20dad58912dd.jpg) +"some boats on a lake." + +![](images/ea23faefa17f90054e42473d5aea37b128c8b8013a982acaeddc599c646f77d1.jpg) + +![](images/0d30afa8e30a582f30149e98842e7a669a3e05b5821eb8de5f7de5dd36265abf.jpg) +"a green pasture filled with wildflowers." + +![](images/ff89ab629159c0c08a685fa6f925aba8900b484de8f29dc646b6ec676b76173c.jpg) + +![](images/b5f7951518fe619b80069c74c9078348110e9869631ac65634c35b482c067907.jpg) + +![](images/acb38a9ac3421cb4c5aa1397614270a07ffe99a30a75b4517d1a1f8d2e534564.jpg) +"a very cute dog laying in a very small sink." + +![](images/1a1b11602e9329617c11e9f5b02cf0445c2342b98f5671840a234d1a3534b577.jpg) + +![](images/caa571ccd6daeecd2b24d2945ccbd8468bc354ee21baf542d4cdf8af06cfa34d.jpg) + +
paramsseq lenCFGFID ↓IS ↑
AR-ViT-VQGAN (reported)650M1024-8.81110.8
AR-ViT-VQGAN (reproduced)650M1024-8.37111.8
AR-SIT-4 (Ours)650M1024-6.95138.3
RQ-VAE [26]480M256-15.7286.8
DQ-VAE [22]355M640-7.34152.8
LlamaGen-L [45]343M2561.54.08198.5
VAR [48]310M6802.03.30274.4
AR-SIT-4* (Ours)350M10241.54.06190.9
+ +Table 3. ImageNet $256 \times 256$ class conditional generation. Top half shows a fair comparison against ViT-VQGAN, which we clearly outperform. AR-SIT-4* uses different hyperparameters for comparison against a broader class of methods. While we do not outperform all of them, there are differences in the architectures and training schedules that influence the results and are orthogonal to our contributions. Moreover, none of the baselines has the multiscale capabilities described in P.2 and P.4 (Sec. 1). + +double the number of decoder layers. The bottom half of Tab. 3 compares with similarly-sized AR models including some with multiscale/residual latents (although none exhibit properties P.2 and P.4). AR-SIT-4* performance matches LlamaGen but not to VAR, but the models are not exactly comparable since they use a convolutional tokenizers and other improvements such as AdaLN [34] and attention normalization; VAR also trains the tokenizer on larger + +data than ImageNet. These are orthogonal to our contributions and can be incorporated in future work. + +We also train an AR-SIT-5* on $512 \times 512$ ImageNet and obtain an FID of 5.97 and IS of 215.22, showing robustness at higher resolutions, while most prior work requires separate stages for upsampling. Fig. 10 shows samples. + +# 6. Conclusion and limitations + +We presented a spectral image tokenizer (SIT), and an autoregressive generative transformer trained with it (AR-SIT). SIT is naturally multiscale and leverages spectral properties of natural images for improved reconstruction quality. AR-SIT enables applications such as rapid generation of coarse images that can be refined later, and text-guided image upsampling and editing. + +While our methods improved tokenizer reconstruction accuracy and class-conditional generation, the text-to-image metrics were similar to the fair baseline. Chang et al. [4] similarly observed that a better tokenizer does not necessarily lead to a better generative model. Nevertheless, our method has multiscale properties and enables new applications not possible with prior work. We only experiment with realtively small AR-SITs of 350-650M parameters, while the Parti [58] baseline goes up to 22B parameters. + +# 7. Acknowledgments + +We thank Leonardo Zepeda-Nuñez for reviewing this manuscript and offering interesting discussions and suggestions, and Jon Barron for sharing useful code we relied on. + +# References + +[1] James Betker, Gabriel Goh, Li Jing, † TimBrooks, Jianfeng Wang, Linjie Li, † LongOuyang, † JuntangZhuang, † JoyceLee, † YufeiGuo, † WesamManassra, † PrafullaDhariwal, † CaseyChu, † YunxinJiao, and Aditya Ramesh. Improving image generation with better captions. 2 +[2] Lei Cai, Hongyang Gao, and Shuiwang Ji. Multi-Stage Variational Auto-Encoders for Coarse-to-Fine Image Generation, page 630–638. Society for Industrial and Applied Mathematics, 2019. 3 +[3] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T. Freeman. Maskgit: Masked generative image transformer. In CVPR, 2022. 1, 3 +[4] Huiwen Chang, Han Zhang, Jarred Barber, Aaron Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Patrick Murphy, William T. Freeman, Michael Rubinstein, Yuanzhen Li, and Dilip Krishnan. Muse: Text-to-image generation via masked generative transformers. In ICML, 2023. 1, 3, 8 +[5] XI Chen, Nikhil Mishra, Mostafa Rohaninejad, and Pieter Abbeel. PixelSNAIL: An improved autoregressive generative model. In ICML, 2018. 3 +[6] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. Pali: A jointly-scaled multilingual language-image model, 2022. 6 +[7] C. A. Christopoulos, T. Ebrahimi, and A. N. Skodras. Jpeg2000: the new still picture compression standard. In Proceedings of the 2000 ACM Workshops on Multimedia, page 45-49, New York, NY, USA, 2000. Association for Computing Machinery. 4 +[8] A. Cohen, Ingrid Daubechies, and J.-C. Feauveau. Biorthogonal bases of compactly supported wavelets. Communications on Pure and Applied Mathematics, 45(5):485-560, 1992. 3 +[9] Ingrid Daubechies. Ten lectures on wavelets. Society for Industrial and Applied Mathematics, USA, 1992. 3 +[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6 +[11] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 1 +[12] Sander Dieleman. Diffusion is spectral autoregression, 2024. 2 + +[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2, 3, 4 +[14] Tim Elsner, Paula Usinger, Victor Czech, Gregor Kobsik, Yanjiang He, Isaak Lim, and Leif Kobbelt. Quantised global autoencoder: A holistic approach to representing visual data, 2024. 2 +[15] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 1, 2, 3, 5 +[16] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis, 2024. 2 +[17] Rinon Gal, Dana Cohen Hochberg, Amit Bermano, and Daniel Cohen-Or. Swagan: a style-based wavelet-driven generative model. ACM TOG, 2021. 1 +[18] Jiatao Gu, Shuangfei Zhai, Yizhe Zhang, Joshua M. Susskind, and Navdeep Jaitly. Matryoshka diffusion models. In ICLR, 2024. 3 +[19] Philippe Hansen-Estruch, David Yan, Ching-Yao Chung, Orr Zohar, Jialiang Wang, Tingbo Hou, Tao Xu, Sriram Vishwanath, Peter Vajda, and Xinlei Chen. Learnings from scaling visual tokenizers for reconstruction and generation, 2025. 1 +[20] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 2 +[21] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022. 1 +[22] Mengqi Huang, Zhendong Mao, Zhuowei Chen, and Yongdong Zhang. Towards accurate image coding: Improved autoregressive image generation with dynamic vector quantization. In CVPR, 2023. 2, 8 +[23]Imagen-Team-Google. Imagen 3, 2024. 2 +[24] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 3 +[25] Vladimir Kulikov, Shahar Yadin, Matan Kleiner, and Tomer Michaeli. Sinddm: A single image denoising diffusion model. In ICML, pages 17920-17930, 2023. 3 +[26] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In CVPR, 2022. 2, 8 +[27] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In ECCV. 5, 6, 7, 8, 2, 3 +[28] Xiaoxiao Ma, Mohan Zhou, Tao Liang, Yalong Bai, Tiejun Zhao, Huaian Chen, and Yi Jin. Star: Scale-wise text-to-image generation via auto-regressive representations, 2024. + +[29] Stephane Mallat. A Wavelet Tour of Signal Processing, Third Edition: The Sparse Way. Academic Press, Inc., USA, 3rd edition, 2008. 3 +[30] Wael Mattar, Idan Levy, Nir Sharon, and Shai Dekel. Wavelets are all you need for autoregressive image generation, 2024. 3 +[31] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. In ICLR, 2018. 5 +[32] Charlie Nash, Jacob Menick, Sander Dieleman, and Peter W. Battaglia. Generating images with sparse representations. In ICML, 2021. 3 +[33] OpenAI. Gpt-4 technical report, 2023. 1, 2 +[34] William Peebles and Saining Xie. Scalable diffusion models with transformers. 2023. 8 +[35] Hao Phung, Quan Dao, and Anh Tran. Wavelet diffusion models are fast and scalable image generators. In CVPR, 2023. 1 +[36] Vivek Ramanujan, Kushal Tirumala, Armen Aghajanyan, Luke Zettlemoyer, and Ali Farhadi. When worse is better: Navigating the compression-generation tradeoff in visual tokenization, 2024. 1 +[37] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. In NeurIPS, 2019. 1, 2 +[38] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2021. 1 +[39] Tamar Rott Shaham, Tali Dekel, and Tomer Michaeli. Singan: Learning a generative model from a single natural image. In ICCV, 2019. 3 +[40] Dohoon Ryu and Jong Chul Ye. Pyramidal denoising diffusion probabilistic models, 2022. 3 +[41] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, Xi Chen, and Xi Chen. Improved techniques for training gans. In NeurIPS, 2016. 2 +[42] Noam Shazeer. Glu variants improve transformer, 2020. 7 +[43] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 1 +[44] Jianlin Su, Yu Lu, Shengfeng Pan, Bo Wen, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding, 2021. 7 +[45] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 2, 8 +[46] Gemini Team. Gemini: A family of highly capable multimodal models, 2023. 1, 2 +[47] Llama team. The llama 3 herd of models, 2024. 1, 2 +[48] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. In NeurIPS, 2024. 2, 8 +[49] Aaron van den Oord, Nal Kalchbrenner, Lasse Espeholt, kory kavukcuoglu, Oriol Vinyals, and Alex Graves. Conditional image generation with pixelCNN decoders. In NeurIPS, 2016. 2 + +[50] Aäron van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In ICML, 2016. 2 +[51] Aaron van den Oord, Oriol Vinyals, and koray kavukcuoglu. Neural discrete representation learning. In NeurIPS, 2017. 1, 2, 3, 5 +[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 1 +[53] Xin Wen, Bingchen Zhao, Ismail Elezi, Jiankang Deng, and Xiaojuan Qi. "principal components" enable a new language of images, 2025. 2 +[54] Jingfeng Yao, Bin Yang, and Xinggang Wang. Reconstruction vs. generation: Taming optimization dilemma in latent diffusion models. In CVPR, 2025. 1 +[55] Ting Yao, Yingwei Pan, Yehao Li, Chong-Wah Ngo, and Tao Mei. Wave-vit: Unifying wavelet and transformers for visual representation learning. In ECCV, 2022. 2 +[56] Hu Yu, Hao Luo, Hangjie Yuan, Yu Rong, and Feng Zhao. Frequency autoregressive image generation with continuous tokens, 2025. 2 +[57] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved VQGAN. In ICLR, 2022. 1, 2, 3, 4, 5, 6, 7 +[58] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research, 2022. Featured Certification. 1, 3, 5, 6, 8 +[59] Lijun Yu, Jose Lezama, Nitesh Bharadwaj Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G Hauptmann, Boqing Gong, Ming-Hsuan Yang, Irfan Essa, David A Ross, and Lu Jiang. Language model beats diffusion - tokenizer is key to visual generation. In ICLR, 2024. 2 +[60] Qihang Yu, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An image is worth 32 tokens for reconstruction and generation. In NeurIPS, 2024. 1 +[61] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 2 +[62] Zhenhai Zhu and Radu Soricut. Wavelet-based image tokenizer for vision transformers, 2024. 2 \ No newline at end of file diff --git a/2025/Spectral Image Tokenizer/images.zip b/2025/Spectral Image Tokenizer/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7fe1f83553c8f1c0d4e2c57273a1a14fecf9e34d --- /dev/null +++ b/2025/Spectral Image Tokenizer/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9d8b617a0df5b3b9cab953c9ea3c0f30b12908ac06c151c5eff058aa0def58e5 +size 708229 diff --git a/2025/Spectral Image Tokenizer/layout.json b/2025/Spectral Image Tokenizer/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0159a74e3037ebcc921153bcb16407b277c48bff --- /dev/null +++ b/2025/Spectral Image Tokenizer/layout.json @@ -0,0 +1,10739 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 225, + 103, + 386, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 103, + 386, + 121 + ], + "spans": [ + { + "bbox": [ + 225, + 103, + 386, + 121 + ], + "type": "text", + "content": "Spectral Image Tokenizer" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 100, + 150, + 173, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 150, + 173, + 163 + ], + "spans": [ + { + "bbox": [ + 100, + 150, + 173, + 163 + ], + "type": "text", + "content": "Carlos Esteves" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 250, + 150, + 343, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 150, + 343, + 163 + ], + "spans": [ + { + "bbox": [ + 250, + 150, + 343, + 163 + ], + "type": "text", + "content": "Mohammed Suhail" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 421, + 150, + 507, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 150, + 507, + 163 + ], + "spans": [ + { + "bbox": [ + 421, + 150, + 507, + 163 + ], + "type": "text", + "content": "Ameesh Makadia" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 200, + 169, + 394, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 169, + 394, + 182 + ], + "spans": [ + { + "bbox": [ + 200, + 169, + 394, + 182 + ], + "type": "text", + "content": "{machc, suhailmhd, makadia}@google.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 255, + 183, + 339, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 183, + 339, + 198 + ], + "spans": [ + { + "bbox": [ + 255, + 183, + 339, + 198 + ], + "type": "text", + "content": "Google Research" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 152, + 243, + 200, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 243, + 200, + 255 + ], + "spans": [ + { + "bbox": [ + 152, + 243, + 200, + 255 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 269, + 297, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 269, + 297, + 521 + ], + "spans": [ + { + "bbox": [ + 54, + 269, + 297, + 521 + ], + "type": "text", + "content": "Image tokenizers map images to sequences of discrete tokens, and are a crucial component of autoregressive transformer-based image generation. The tokens are typically associated with spatial locations in the input image, arranged in raster scan order, which is not ideal for autoregressive modeling. In this paper, we propose to tokenize the image spectrum instead, obtained from a discrete wavelet transform (DWT), such that the sequence of tokens represents the image in a coarse-to-fine fashion. Our tokenizer brings several advantages: 1) it leverages that natural images are more compressible at high frequencies, 2) it can take and reconstruct images of different resolutions without retraining, 3) it improves the conditioning for next-token prediction – instead of conditioning on a partial line-by-line reconstruction of the image, it takes a coarse reconstruction of the full image, 4) it enables partial decoding where the first few generated tokens can reconstruct a coarse version of the image, 5) it enables autoregressive models to be used for image upsampling. We evaluate the tokenizer reconstruction metrics as well as multiscale image generation, text-guided image upsampling and editing." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 547, + 135, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 547, + 135, + 559 + ], + "spans": [ + { + "bbox": [ + 56, + 547, + 135, + 559 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 568, + 296, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 296, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 296, + 640 + ], + "type": "text", + "content": "In natural language processing, tokenization associates sets of characters to entries in a vocabulary, which is straightforward since language is discrete. The sequence of tokens is suitably modeled as a categorical distribution with autoregressive (AR) transformers, which is the foundation of modern large language models (LLMs) [33, 46, 47]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "While natural images are represented by discrete pixel values, they exhibit high dimensionality, redundancies, and noise so it's impractical to associate one token per pixel. This motivated a long line of research of learnable image tokenizers [15, 37, 37, 51, 57]. While there are successful autoregressive image generation models [15, 57, 58]," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 243, + 555, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 304 + ], + "type": "text", + "content": "images are not sequential like language, which motivated developing alternatives such as denoising diffusion models [11, 38, 43] and masked transformers [3, 4, 60]. Nevertheless, most of these alternatives also operate on the latent space of tokenizers like VQGAN [15] instead of raw pixels." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 304, + 555, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 555, + 388 + ], + "type": "text", + "content": "In this work, we revisit AR transformer-based image generation. Our main contribution is a tokenizer operating on the image spectrum, specifically on DWT coefficients, where the coarse-to-fine representation lends itself more naturally to a sequential interpretation. Our Spectral Image Tokenizer (SIT) has several useful properties and enables different applications:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 391, + 554, + 714 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 315, + 391, + 554, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 391, + 554, + 450 + ], + "spans": [ + { + "bbox": [ + 315, + 391, + 554, + 450 + ], + "type": "text", + "content": "P.1 Since the power spectrum of natural images decreases with frequency, high frequencies can be more heavily compressed with little effect in visual quality. SIT leverages this by associating tokens to larger patches at higher wavelet scales than at lower scales (see Fig. 2)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 450, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 450, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 315, + 450, + 554, + 533 + ], + "type": "text", + "content": "P.2 SIT is transformer-based [52]; by using an attention mask where each scale depends on itself and lower scales (\"Scale-Causal attention\"), SIT can be trained at a single resolution and used to tokenize images of multiple resolutions (any number of scales up to the trained maximum), and detokenize partial token sequences (up to some scale), reconstructing a coarse image." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 534, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 534, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 315, + 534, + 554, + 605 + ], + "type": "text", + "content": "P.3 Using SIT, we train an autoregressive generative transformer (AR-SIT) that models images coarse-to-fine. The next-token prediction is then conditioned on a coarse reconstruction of the image given by the partial token sequence, instead of the usual conditioning on the partial reconstruction of the previous rows of the image." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 605, + 553, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 605, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 315, + 605, + 553, + 653 + ], + "type": "text", + "content": "P.4 AR-SIT can quickly generate only the first few tokens and reconstruct a coarse version of the image, enabling applications like quickly showing multiple coarse generations and letting the user select which ones to refine." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 653, + 553, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 653, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 315, + 653, + 553, + 700 + ], + "type": "text", + "content": "P.5 AR-SIT can be used for text-based upsampling of an input low resolution image, by starting the decoding process with the few tokens output by SIT, and generating the rest of the sequence up to a desired resolution." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 700, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 700, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 315, + 700, + 553, + 714 + ], + "type": "text", + "content": "P.6 AR-SIT can be used for text-guided image editing, by" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17181" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 233, + 245 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 233, + 245 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 233, + 245 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 233, + 245 + ], + "type": "image", + "image_path": "2bfa86016e0efaaf3b1d858e61d41b14be8218e91de76c2783d796f1bade2604.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 252, + 555, + 330 + ], + "lines": [ + { + "bbox": [ + 55, + 252, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 555, + 330 + ], + "type": "text", + "content": "Figure 1. Left: we introduce a Spectral Image Tokenizer (SIT), that learns to encode and decode discrete wavelet transform (DWT) coefficients to and from a small set of discrete tokens, such that the sequence represents the image in a coarse-fine fashion. SIT is naturally multiscale and enables coarse-to-fine autoregressive image generation with our AR-SIT model. SIT also leverages the sparsity of high frequency coefficients in natural images. Right: details of the encoder/decoder transformer layers. The main architectural difference with respect to previous tokenizers is that the distributions of DWT approximation and details coefficients are distinct, hence we use specialized parameters for each in the quantizer codebooks and inner transformer layers. We also introduce a scale-causal attention where each token attends to its own scale and lower scales, which enables encoding, decoding, generating, and upsampling images at different resolutions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 260, + 68, + 556, + 242 + ], + "blocks": [ + { + "bbox": [ + 260, + 68, + 556, + 242 + ], + "lines": [ + { + "bbox": [ + 260, + 68, + 556, + 242 + ], + "spans": [ + { + "bbox": [ + 260, + 68, + 556, + 242 + ], + "type": "image", + "image_path": "db8c6463a1c65f8aaf7b1abc40644785cb3a11c1ae22bbb2fa368a367c846a68.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 71, + 339, + 295, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 339, + 295, + 363 + ], + "spans": [ + { + "bbox": [ + 71, + 339, + 295, + 363 + ], + "type": "text", + "content": "encoding a given image up to a coarse scale, and generating the finer details conditioned on a new caption." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 366, + 296, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 296, + 535 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 296, + 535 + ], + "type": "text", + "content": "Currently, image generation is dominated by diffusion models such as Imagen 3 [23], DALL-E-3 [1] and Stable Diffusion 3 [16]. On the other hand, LLMs such as Gemini [46], GPT-4 [33], and Llama 3 [47] are based on autoregressive transformers. We believe autoregressive image generation is still worth pursuing because it might benefit from advances in LLMs, and multimodal applications might benefit of having a single architecture for all modalities. Recent work on image and video generation support this point [45, 48, 59]. Dieleman [12] recently interpreted denoising diffusion models as spectral autoregression, since, when looking at image spectra, the denoising procedure uncovers frequencies from low to high. In contrast, our method does literal spectral autoregression." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 548, + 139, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 139, + 560 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 139, + 560 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 569, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 688 + ], + "type": "text", + "content": "Image tokenization Several methods have been developed to map images to a small set of discrete tokens suitable for generative modeling. VQ-VAE [51] introduced vector-quantization in the latent space of a Variational AutoEncoder to map images, audio and video to a set of discrete values. VQGAN [15] improved upon VQVAE by using perceptual and adversarial losses. We build on ViT-VQGAN [57], which improved upon VQGAN by using a Vision Transformer (ViT) [13] instead of convolutions, as well as codebook factorization and feature normalization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "In this paper, we are interested in multiscale image representations. VQ-VAE-2 [37] introduced multiscale latents" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 338, + 555, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 482 + ], + "type": "text", + "content": "by keeping and quantizing intermediate downsampled convolutional features. RQ-VAE [26] quantized a set of residuals such that the latent vector is represented in a coarse-to-fine fashion and reconstructed by adding the code embeddings for each residual. Similarly, VAR [48] and STAR [28] split the latent space in multi-scale quantized residuals, so that it can be reconstructed by upsampling and summing. FAR [56] splits the latents into frequencies to train a next-frequency generative model. DQ-VAE [22] encoded smooth regions with fewer tokens that are generated first, while QG-VAE [14] and SEMANTICIST [53] learn non-local tokens that are ordered by importance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 483, + 555, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 554 + ], + "type": "text", + "content": "The crucial difference between our approach and the aforementioned is that we operate on spectral coefficients of the input and not on latent features. This allows our tokenizer to be truly multiscale; it can take inputs at different scales and reconstruct up to some scale, and leverage that higher frequencies are more compressible in natural images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 555, + 554, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 554, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 554, + 662 + ], + "type": "text", + "content": "Tangentially related to our work, Wave-ViT [55] modified the ViT self-attention by applying a DWT to the input and concatenating coefficients, effectively exchanging space for depth to reduce the sequence length. Zhu and Soricut [62] modified the ViT patchifier in a similar way, and introduced patch embeddings that leverage the sparsity of high frequency coefficients. Both methods are for discriminative tasks such as image classification and segmentation, while we focus on generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": "Autoregressive image generation Early approaches PixelRNN and PixelCNN [49, 50] generate images pixel by pixel by modeling the conditional distribution of the pixel given the previous pixels with recurrent layers of causal" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17182" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 191 + ], + "type": "text", + "content": "convolutions. PixelSNAIL [5] improved on this model by introducing self-attention layer to better model long-range dependencies. VQ-VAE [51] introduced a two-stage approach with a tokenizer and a separate stage to model the distribution of tokens. VQGAN [15] greatly improved these results by both improving the tokenizer and using a transformer to model the distribution. Finally, ViT-VQGAN [57] proposed a transformer-based tokenizer, which Parti [58] used with a large autoregressive transformer capable of high-quality text-to-image generation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 192, + 296, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 296, + 275 + ], + "type": "text", + "content": "MaskGIT [3] and Muse [4] highlighted the disadvantages of the typical autoregressive models raster-order conditioning, and proposed to generate all tokens in parallel iteratively, where each iteration keeps the highest confidence tokens. We address the same problem with a tokenizer whose sequence represents the image in a coarse-to-fine order instead of raster-order." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 277, + 296, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 296, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 296, + 324 + ], + "type": "text", + "content": "Multiscale image generation Multiscale image generation ideas have appeared in the context of VAEs [2], GANs [24, 39], and diffusion models [18, 25, 40], but have not been sufficiently explored with AR transformers." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 325, + 296, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 325, + 296, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 325, + 296, + 492 + ], + "type": "text", + "content": "Nash et al. [32] represented an image as a sequence of quantized and thresholded DCT coefficients, where the compression comes from the fact that many coefficients are zero and are omitted. By sorting the sequence by frequency, the method can model images from coarse-to-fine like we do. However, the compressed representation is handcrafted and results in long sequences. In a similar vein, Mattar et al. [30] handcrafted a tokenizer based on DWT coefficients, introducing tokens to represent large chunks of zeros. It also results in long sequences and is only applied to generations of small grayscale images. In contrast with these methods, instead of handcrafting a compressed input representation, our tokenizer learns to encode to, and decode from, a short compressed coarse-to-fine sequence." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 503, + 134, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 134, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 134, + 517 + ], + "type": "text", + "content": "3. Background" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 523, + 296, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 523, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 523, + 296, + 559 + ], + "type": "text", + "content": "A discrete wavelet transform (DWT) is based on successive convolutions of the signal " + }, + { + "bbox": [ + 55, + 523, + 296, + 559 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 523, + 296, + 559 + ], + "type": "text", + "content": " with a pair of lowpass " + }, + { + "bbox": [ + 55, + 523, + 296, + 559 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 523, + 296, + 559 + ], + "type": "text", + "content": " and highpass " + }, + { + "bbox": [ + 55, + 523, + 296, + 559 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 523, + 296, + 559 + ], + "type": "text", + "content": " filters with the first step as follows" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 99, + 568, + 294, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 568, + 294, + 594 + ], + "spans": [ + { + "bbox": [ + 99, + 568, + 294, + 594 + ], + "type": "interline_equation", + "content": "f _ {\\operatorname {l o w} _ {1}} [ n ] = f \\star g = \\sum_ {k} f [ k ] g [ n - k ], \\tag {1}", + "image_path": "8740d0c8dba5d3a0c82b9e8125881e4cc0416bf0036228da73d5aab684082fe4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 97, + 596, + 294, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 596, + 294, + 620 + ], + "spans": [ + { + "bbox": [ + 97, + 596, + 294, + 620 + ], + "type": "interline_equation", + "content": "f _ {\\operatorname {h i g h} _ {1}} [ n ] = f \\star h = \\sum_ {k} f [ k ] h [ n - k ]. \\tag {2}", + "image_path": "91a0b6ce73bcb3e3816add88d171897a71134f924a6ab485520e297b3061e72d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "The high and low outputs are subsampled by a factor of two, and the operation is repeated for the low channel, such that at level " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " we compute " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{low}_L} = f_{\\mathrm{low}_{L - 1}} \\star g" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{high}_L} = f_{\\mathrm{low}_{L - 1}} \\star h" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": ", subsample again and drop " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{low}_{L - 1}}" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": ". The output at level " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " comprises the approximation coefficients " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{low}_L}" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " and the detail coefficients " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\{f_{\\mathrm{high}_k}\\}" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "1 \\leq k \\leq L" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": ". This output has the same cardinality as the input and the" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 434, + 189 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 434, + 189 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 434, + 189 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 434, + 189 + ], + "type": "image", + "image_path": "863d7df0e87b9ef660d64f4c69a21c033d45e2e444801ba5802f9b521fe6710d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 198, + 555, + 320 + ], + "lines": [ + { + "bbox": [ + 313, + 198, + 555, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 555, + 320 + ], + "type": "text", + "content": "Figure 2. Input patchification. Left: typical patchification for Vision Transformers (ViT) [13], where the image is split in equal-sized patches. Right: we propose to patchify the coefficients of a discrete wavelet transform (DWT) instead. Each scale is shown in a different color. Scales other than the lowest contain three blocks representing horizontal, vertical and diagonal details; we concatenate the spatially correspondent patches of each block such that each scale is represented by the same sequence length. This results in larger patch sizes for higher scales, which are more compressible. The figure shows 3 scales and 16 tokens per scale; in our experiments we use 4 or 5 scales and 256 tokens per scale." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 435, + 72, + 552, + 188 + ], + "blocks": [ + { + "bbox": [ + 435, + 72, + 552, + 188 + ], + "lines": [ + { + "bbox": [ + 435, + 72, + 552, + 188 + ], + "spans": [ + { + "bbox": [ + 435, + 72, + 552, + 188 + ], + "type": "image", + "image_path": "3d53cd99caa83e3e495af4cbd9515dbd49fa7a9103b690f6877a7ea3f9b4392f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 330, + 554, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 554, + 377 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 554, + 377 + ], + "type": "text", + "content": "transformation is invertible, where the forward transform is typically called \"analysis\" and the backward \"synthesis\". The simplest wavelet family is the Haar, where " + }, + { + "bbox": [ + 313, + 330, + 554, + 377 + ], + "type": "inline_equation", + "content": "g = [1,1]^{\\top}" + }, + { + "bbox": [ + 313, + 330, + 554, + 377 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 330, + 554, + 377 + ], + "type": "inline_equation", + "content": "h = [1, -1]^{\\top}" + }, + { + "bbox": [ + 313, + 330, + 554, + 377 + ], + "type": "text", + "content": " (optionally scaled to unit norm)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 378, + 554, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 554, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 554, + 485 + ], + "type": "text", + "content": "For image analysis we use a 2D DWT, which is obtained by simply convolving the rows and columns with " + }, + { + "bbox": [ + 313, + 378, + 554, + 485 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 378, + 554, + 485 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 378, + 554, + 485 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 378, + 554, + 485 + ], + "type": "text", + "content": ". The approximation coefficients " + }, + { + "bbox": [ + 313, + 378, + 554, + 485 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{low}_1} = f \\star (gg^\\top)" + }, + { + "bbox": [ + 313, + 378, + 554, + 485 + ], + "type": "text", + "content": ", and the details are divided into horizontal " + }, + { + "bbox": [ + 313, + 378, + 554, + 485 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{H}_1} = f \\star (gh^\\top)" + }, + { + "bbox": [ + 313, + 378, + 554, + 485 + ], + "type": "text", + "content": ", vertical " + }, + { + "bbox": [ + 313, + 378, + 554, + 485 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{V}_1} = f \\star (hg^\\top)" + }, + { + "bbox": [ + 313, + 378, + 554, + 485 + ], + "type": "text", + "content": " and diagonal " + }, + { + "bbox": [ + 313, + 378, + 554, + 485 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{D}_1} = f \\star (hh^\\top)" + }, + { + "bbox": [ + 313, + 378, + 554, + 485 + ], + "type": "text", + "content": " details. Subsequent levels apply the same operations to the approximation coefficients. Fig. 2 shows a two-level transform, where we can see that the approximation coefficients correspond to a coarse version of the image." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 486, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 554, + 510 + ], + "type": "text", + "content": "We refer to the textbooks by Mallat [29] and Daubechies [9] for more information about wavelets." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 522, + 370, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 522, + 370, + 534 + ], + "spans": [ + { + "bbox": [ + 314, + 522, + 370, + 534 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 543, + 554, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 554, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 554, + 615 + ], + "type": "text", + "content": "Our main contribution, SIT, is an image tokenizer that takes discrete wavelet transform (DWT) coefficients. The model follows ViT-VQGAN [57], with important changes that we describe in this section and visualize in Fig. 1. For image generation, we introduce AR-SIT which is based on Parti [58] with minor changes, using SIT as the tokenizer." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 624, + 383, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 383, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 383, + 635 + ], + "type": "text", + "content": "4.1. Tokenizer" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "Patchification and featurization The first step is to map the input to a sequence of patches. We apply the Haar DWT on the input image and patchify each scale separately. While Haar is the simplest wavelet and lacks properties found in other wavelets useful for compression, we found no benefits of using other wavelet families such as CDF [8]." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17183" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "Our design choice is to use the same number of patches for each scale. In a DWT, the higher scales correspond to high frequency details which are represented by more coefficients than the lower scales, but contribute less to the spatial pixel values. In other words, in natural images, most of the power spectrum is concentrated on lower frequencies. By representing each scale with the same number of tokens, we are compressing more the higher frequencies (since they are represented by more coefficients), similarly to what is done in image compression methods such as JPEG2000 [7]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 192, + 295, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 295, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 295, + 347 + ], + "type": "text", + "content": "The approximation (or lowpass) DWT coefficients correspond to a coarse version of the input image, which we consider the first scale. The following scales are divided in three blocks corresponding to horizontal, vertical, and diagonal details, where each coefficient relates to a specific spatial location. Thus, we can concatenate the three blocks such that each entry corresponds to the same spatial location. For example, the first scale will typically be split in patches that are " + }, + { + "bbox": [ + 55, + 192, + 295, + 347 + ], + "type": "inline_equation", + "content": "32 \\times 32 \\times 3" + }, + { + "bbox": [ + 55, + 192, + 295, + 347 + ], + "type": "text", + "content": ", with channels corresponding to RGB, the second scale will be " + }, + { + "bbox": [ + 55, + 192, + 295, + 347 + ], + "type": "inline_equation", + "content": "32 \\times 32 \\times 9" + }, + { + "bbox": [ + 55, + 192, + 295, + 347 + ], + "type": "text", + "content": ", where the channels correspond to the RGB of horizontal, vertical, and diagonal details, the third will be " + }, + { + "bbox": [ + 55, + 192, + 295, + 347 + ], + "type": "inline_equation", + "content": "64 \\times 64 \\times 9" + }, + { + "bbox": [ + 55, + 192, + 295, + 347 + ], + "type": "text", + "content": " and so on. Fig. 2 shows an example of our patchification scheme." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 348, + 295, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 295, + 408 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 295, + 408 + ], + "type": "text", + "content": "Since patches have different resolutions (higher scales will have larger patches), the usual ViT linear embedding to map patches to features cannot be shared across all patches so we have different parameters per scale. Formally, given an image " + }, + { + "bbox": [ + 55, + 348, + 295, + 408 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 348, + 295, + 408 + ], + "type": "text", + "content": ", our patchifiers compute" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 418, + 295, + 432 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 418, + 295, + 432 + ], + "spans": [ + { + "bbox": [ + 66, + 418, + 295, + 432 + ], + "type": "interline_equation", + "content": "f _ {\\text {l o w} L}, \\left\\{f _ {\\mathrm {H} _ {i}} \\right\\} _ {i \\leq L}, \\left\\{f _ {\\mathrm {V} _ {i}} \\right\\} _ {i \\leq L}, \\left\\{f _ {\\mathrm {D} _ {i}} \\right\\} _ {i \\leq L} = \\mathrm {D W T} (f), \\tag {3}", + "image_path": "6740c7e5657068b85c7352e7a6f4748ef4596cd6a50a8b486ff7eebd79bcfee0.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 434, + 294, + 446 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 434, + 294, + 446 + ], + "spans": [ + { + "bbox": [ + 66, + 434, + 294, + 446 + ], + "type": "interline_equation", + "content": "c _ {1} = P _ {0} \\left(f _ {\\text {l o w} L}\\right), \\tag {4}", + "image_path": "a5c47fc5e73378b7d5d3c7a3d5f09a99d052e7180b0d0fcf86c6e656203daa58.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 449, + 294, + 461 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 449, + 294, + 461 + ], + "spans": [ + { + "bbox": [ + 66, + 449, + 294, + 461 + ], + "type": "interline_equation", + "content": "c _ {s} = P _ {s} \\left(f _ {\\mathrm {H} _ {L - s + 2}}, f _ {\\mathrm {V} _ {L - s + 2}}, f _ {\\mathrm {D} _ {L - s + 2}}\\right), 1 < s \\leq L, \\tag {5}", + "image_path": "dddbf0fd25b8e36f26eef543de56a9c2a9d50be3ad5512345854f1aafaa456e2.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 471, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 295, + 555 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 471, + 295, + 555 + ], + "type": "inline_equation", + "content": "c_{s} = \\{c_{s}^{n}\\}_{1\\leq n\\leq N}" + }, + { + "bbox": [ + 55, + 471, + 295, + 555 + ], + "type": "text", + "content": " is the sequence of token embeddings at scale " + }, + { + "bbox": [ + 55, + 471, + 295, + 555 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 471, + 295, + 555 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 471, + 295, + 555 + ], + "type": "inline_equation", + "content": "c_{s}^{n}\\in \\mathbb{R}^{C}" + }, + { + "bbox": [ + 55, + 471, + 295, + 555 + ], + "type": "text", + "content": " is the embedding of the " + }, + { + "bbox": [ + 55, + 471, + 295, + 555 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 471, + 295, + 555 + ], + "type": "text", + "content": "-th token at the " + }, + { + "bbox": [ + 55, + 471, + 295, + 555 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 471, + 295, + 555 + ], + "type": "text", + "content": "-th scale, while " + }, + { + "bbox": [ + 55, + 471, + 295, + 555 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 471, + 295, + 555 + ], + "type": "text", + "content": " is the number of DWT levels, " + }, + { + "bbox": [ + 55, + 471, + 295, + 555 + ], + "type": "inline_equation", + "content": "S = L + 1" + }, + { + "bbox": [ + 55, + 471, + 295, + 555 + ], + "type": "text", + "content": " is the number of scales and " + }, + { + "bbox": [ + 55, + 471, + 295, + 555 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 471, + 295, + 555 + ], + "type": "text", + "content": " the number of tokens per scale. For brevity, when there is no ambiguity, we may omit the set indexing and use " + }, + { + "bbox": [ + 55, + 471, + 295, + 555 + ], + "type": "inline_equation", + "content": "\\{c_s\\}" + }, + { + "bbox": [ + 55, + 471, + 295, + 555 + ], + "type": "text", + "content": " to denote the set of tokens of all scales, for example." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 555, + 295, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 295, + 615 + ], + "type": "text", + "content": "The final projection after the decoder needs to map the features back to different sized patches, so it also has different parameters per scale. Those patches still represent DWT coefficients so an inverse DWT (IDWT) is finally applied to obtain an image output." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "Flexible sequence length The tokenizer encoder and decoder transformers operate on the sequence of patch features of length " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "SN" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": ". The sequence length is a major factor in resource utilization so we want to keep it constrained. Our method is flexible since we can choose the number of scales and the patch size per scale, while most ViT-based models such as ViT-VQGAN [57] are more restricted. They use the same patch size for the whole image; thus, keeping the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "same patch size and doubling each image dimension would increase the sequence length by a factor of 4, where our method is capable of including additional scales which only increases the sequence length by multiples of " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 121, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 553, + 228 + ], + "type": "text", + "content": "For example, for a " + }, + { + "bbox": [ + 313, + 121, + 553, + 228 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 121, + 553, + 228 + ], + "type": "text", + "content": " input, ViT-VQGAN uses " + }, + { + "bbox": [ + 313, + 121, + 553, + 228 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 313, + 121, + 553, + 228 + ], + "type": "text", + "content": " patches to obtain a sequence of 1024 elements. Our SIT can use 4 scales and 256 tokens per scale for the same sequence length. When increasing the resolution to " + }, + { + "bbox": [ + 313, + 121, + 553, + 228 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 121, + 553, + 228 + ], + "type": "text", + "content": ", the baseline can either increase the patch size to " + }, + { + "bbox": [ + 313, + 121, + 553, + 228 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 313, + 121, + 553, + 228 + ], + "type": "text", + "content": ", resulting in the same sequence length, or keep it " + }, + { + "bbox": [ + 313, + 121, + 553, + 228 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 313, + 121, + 553, + 228 + ], + "type": "text", + "content": ", resulting in a " + }, + { + "bbox": [ + 313, + 121, + 553, + 228 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 313, + 121, + 553, + 228 + ], + "type": "text", + "content": " longer sequence. SIT, for example, can vary the number of scales from 4 to 6, resulting in sequence lengths of 1024, 1280, 1536 which are more manageable." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 233, + 553, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 233, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 233, + 553, + 293 + ], + "type": "text", + "content": "Transformers After featurization, the single sequence containing all scales passes through a transformer encoder, followed by quantization and a transformer decoder. We propose two optional modifications to the transformers, which are otherwise identical to the ones used in ViT [13]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 294, + 553, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 553, + 510 + ], + "type": "text", + "content": "First, we propose a scale-causal attention mask, where an element at some scale attends to all elements of its own scale and lower scales, represented by a block-triangular mask pattern. With dense attention, we write the application of the encoding transformer " + }, + { + "bbox": [ + 313, + 294, + 553, + 510 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{enc}}" + }, + { + "bbox": [ + 313, + 294, + 553, + 510 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 294, + 553, + 510 + ], + "type": "inline_equation", + "content": "\\{z_s\\} = T_{\\mathrm{enc}}(\\{c_s\\})" + }, + { + "bbox": [ + 313, + 294, + 553, + 510 + ], + "type": "text", + "content": ", and for scale-causal we write " + }, + { + "bbox": [ + 313, + 294, + 553, + 510 + ], + "type": "inline_equation", + "content": "z_s = T_{\\mathrm{enc}}(\\{c_k\\}_{1 \\leq k \\leq s})" + }, + { + "bbox": [ + 313, + 294, + 553, + 510 + ], + "type": "text", + "content": " for each " + }, + { + "bbox": [ + 313, + 294, + 553, + 510 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 294, + 553, + 510 + ], + "type": "text", + "content": ". The scale-causal attention can be applied independently to the encoder and decoder, enabling different applications. For the multiscale reconstruction experiments in Sec. 5.1, we need to both encode and decode multiple resolutions, so both encoder and decoder use scale-causal masks. For coarse-to-fine image generation in Sec. 5.2, only the decoder needs to be scale-causal in order to decode the partially generated sequence. For the text-guided image upsampling in Sec. 5.3, only the encoder needs to be scale-causal to encode the lower resolution inputs. For the image editing experiments in Sec. 5.4, the scale-causal encoder prevents information leaking from high to low scale." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 510, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 713 + ], + "type": "text", + "content": "Second, we propose to use different transformer parameters for the approximation (first scale) and details coefficients (other scales). In our model, the subsequences corresponding to coefficients of each type of coefficient come from quite distinct distributions, so it makes sense to treat them differently. This contrasts with the spatial representation of images where each patch can be considered as coming from the same distribution. Thus, the parameters of the key, query, and value embeddings, the layer norms, and the MLP on each transformer layer are not shared between the approximation and details coefficients. This Approximation-Details Transformer (ADTransformer) still takes a single sequence composed of all coefficients. We experimented with different transformers per sequence with cross-attention for information sharing, but it performed worse. This change leads to more memory utilization to store the extra parameters, but the training/inference speed" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17184" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 71, + 294, + 320 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 294, + 320 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 294, + 320 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 294, + 320 + ], + "type": "table", + "html": "
LPIPS ↓PSNR ↑L1 ↓FID ↓IS ↑images/s ↑
Resolution: 512 × 512
ViT-VQGAN0.32022.40.0426.92151.5593
SIT-5 (Ours)0.26022.00.0462.65192.0410
SIT-6 (Ours)0.23923.10.0401.74203.7320
Resolution: 256 × 256
ViT-VQGAN (reported)-24.80.0321.99184.4-
ViT-VQGAN (reproduced)0.16725.00.0312.33184.0-
ViT-VQGAN (no LL)0.16323.80.0381.20194.6626
SIT-4 (Ours)0.14424.00.0371.20199.5596
SIT-5 (Ours)0.13524.50.0350.97202.3411
SIT-SC-5 (Ours)0.16124.10.0371.33193.7411
Resolution: 128 × 128
ViT-VQGAN0.18526.30.0303.77117.3626
SIT-SC-5 (ours)0.15927.10.0272.13129.3582
Resolution: 64 × 64
ViT-VQGAN0.12928.80.0233.5321.0627
SIT-SC-5 (ours)0.11131.30.0171.3930.1847
Resolution: 32 × 32
ViT-VQGAN0.21423.30.045-3.7627
SIT-SC-5 (ours)0.02936.80.0100.313.5825
Resolution: 16 × 16
ViT-VQGAN0.12724.90.039-1.7627
SIT-SC-5 (ours)0.01341.30.0060.091.82620
", + "image_path": "94af4ecd39a3b10381c810b05a753d77349002dbaf0731b256169d374b3e1895.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 470, + 290, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 470, + 290, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 290, + 483 + ], + "type": "text", + "content": "is similar because the number of operations is unchanged." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 485, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 296, + 628 + ], + "type": "text", + "content": "Quantizer The encoder outputs a sequence of features that are quantized to a fixed-sized codebook, similarly to ViT-VQGAN [57], VQVAE [51], and VQGAN [15]. We modify the quantizer such that approximation and details have different codebooks for the same reasons discussed previously. Thus, codebook sizes and feature dimensions are the same as the ViT-VQGAN baseline, but we have different features for the same code at different positions. Formally, we apply " + }, + { + "bbox": [ + 55, + 485, + 296, + 628 + ], + "type": "inline_equation", + "content": "q_{s}^{n} = Q_{\\mathrm{approx}}(z_{s}^{n})" + }, + { + "bbox": [ + 55, + 485, + 296, + 628 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 485, + 296, + 628 + ], + "type": "inline_equation", + "content": "s = 1" + }, + { + "bbox": [ + 55, + 485, + 296, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 485, + 296, + 628 + ], + "type": "inline_equation", + "content": "q_{s}^{n} = Q_{\\mathrm{details}}(z_{s}^{n})" + }, + { + "bbox": [ + 55, + 485, + 296, + 628 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 485, + 296, + 628 + ], + "type": "inline_equation", + "content": "s \\geq 2" + }, + { + "bbox": [ + 55, + 485, + 296, + 628 + ], + "type": "text", + "content": " for each pair " + }, + { + "bbox": [ + 55, + 485, + 296, + 628 + ], + "type": "inline_equation", + "content": "(s,n)" + }, + { + "bbox": [ + 55, + 485, + 296, + 628 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 485, + 296, + 628 + ], + "type": "inline_equation", + "content": "q_{s}^{n}" + }, + { + "bbox": [ + 55, + 485, + 296, + 628 + ], + "type": "text", + "content": " is chosen from the codebook for " + }, + { + "bbox": [ + 55, + 485, + 296, + 628 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 485, + 296, + 628 + ], + "type": "text", + "content": " and can be associated with its discrete position in the codebook, denoted " + }, + { + "bbox": [ + 55, + 485, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\lfloor q_s^n \\rfloor" + }, + { + "bbox": [ + 55, + 485, + 296, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 629, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 677 + ], + "type": "text", + "content": "Training We follow the ViT-VQGAN training protocol and use the same weighting for the L2, perceptual, adversarial, and quantization losses. We remove the logit-laplace loss that was shown detrimental in follow-up work [58]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "All losses are applied to the spatial domain images reconstructed by the inverse discrete wavelet transform on the decoder output: " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{f} = \\mathrm{IDWT}(T_{\\mathrm{dec}}(\\{q_s\\}))" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ". We noticed insta" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 316, + 72, + 553, + 259 + ], + "blocks": [ + { + "bbox": [ + 55, + 327, + 296, + 460 + ], + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 460 + ], + "type": "text", + "content": "Table 1. Multiscale reconstruction on ImageNet. \"SC\" denotes scale-causal attention, which slightly reduces performance at the highest resolution but enables multiscale reconstruction without downsampling/upsampling or retraining. The ViT-VQGAN values from Yu et al. [57] used a logit-laplace loss (LL) which was later considered harmful [58], so we retrain without it. Our SIT improves reconstruction metrics, is significantly faster at lower resolutions, and robust when increasing resolution. The ViT-VQGAN baseline suffered from instability during training on " + }, + { + "bbox": [ + 55, + 327, + 296, + 460 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 327, + 296, + 460 + ], + "type": "text", + "content": " inputs so we selected the best values before divergence. We report test time throughput of an encoding/decoding cycle for the max batch size that fits on a TPU v5e." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 72, + 553, + 259 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 259 + ], + "type": "table", + "html": "
FID ↓IS ↑images/s ↑images/Gb ↑
Resolution: 256 × 256
Parti350M (reported)14.1---
Parti350M12.436.57.812.0
AR-SIT-SCD-4 (Ours)12.637.36.58.0
Resolution: 128 × 128
Parti350M11.233.57.612.0
AR-SIT-SCD-4 (Ours)11.433.212.612.0
Resolution: 64 × 64
Parti350M10.516.97.612.0
AR-SIT-SCD-4 (Ours)11.418.624.516.0
Resolution: 32 × 32
Parti350M5.82.97.77.7
AR-SIT-SCD-4 (Ours)7.63.274.728.0
", + "image_path": "4a11b64ad2dc3a729b5f5fed6063400c38fa36b2e38f86efa3d6d25bf0754280.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 267, + 555, + 368 + ], + "lines": [ + { + "bbox": [ + 313, + 267, + 555, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 267, + 555, + 368 + ], + "type": "text", + "content": "Table 2. Coarse-to-fine image generation on MS-COCO [27]. With SIT, the autoregressive generation is stopped early for a coarse version of the image. The Parti350M baseline does not have this property, so we generate at full resolution and downsample for comparison. AR-SIT matches the baseline performance at the training resolution but is several times faster and more memory efficient at lower resolutions, even when trained only on higher resolution data. We report throughput and memory utilization during generation given the max batch size that fits on a TPU v6e." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 376, + 555, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 425 + ], + "type": "text", + "content": "bility during training due to the adversarial loss, which was fixed by applying spectral normalization following Miyato et al. [31], which simply divides the discriminator weight matrices by their largest singular value." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 430, + 489, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 489, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 489, + 443 + ], + "type": "text", + "content": "4.2. Autoregressive image generation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 448, + 554, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 554, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 554, + 507 + ], + "type": "text", + "content": "We use our tokenizer for autoregressive image generation, by training a second stage transformer model similar to Parti [58], with some modifications. Formally, the autoregressive transformer " + }, + { + "bbox": [ + 313, + 448, + 554, + 507 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 448, + 554, + 507 + ], + "type": "text", + "content": " models categorical distributions over the discrete codes" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 337, + 514, + 553, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 514, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 337, + 514, + 553, + 529 + ], + "type": "interline_equation", + "content": "P \\left(\\left\\lfloor q _ {s} ^ {n} \\right\\rfloor \\mid \\left\\{\\left\\lfloor q _ {i} \\right\\rfloor \\right\\} _ {1 \\leq i < s} \\cup \\left\\{\\left\\lfloor q _ {s} ^ {i} \\right\\rfloor \\right\\} ^ {1 \\leq i < n}\\right) = T (c) \\tag {6}", + "image_path": "93b268194fac8120de06bff8b968dc9275e4e83d3875b90392881ee89227a293.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 533, + 554, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 554, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 554, + 606 + ], + "type": "text", + "content": "which can be sampled one code at a time for generation. " + }, + { + "bbox": [ + 313, + 533, + 554, + 606 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 533, + 554, + 606 + ], + "type": "text", + "content": " can be conditioned on a textual description processed by a transformer encoder for text-to-image generation. For training, we model the distribution of input codes as " + }, + { + "bbox": [ + 313, + 533, + 554, + 606 + ], + "type": "inline_equation", + "content": "P(\\{\\lfloor q_s\\rfloor \\}) = \\prod_{1\\leq s\\leq S}^{1\\leq n\\leq N}P(\\lfloor q_s^n\\rfloor)" + }, + { + "bbox": [ + 313, + 533, + 554, + 606 + ], + "type": "text", + "content": " and minimize the negative log-likelihood " + }, + { + "bbox": [ + 313, + 533, + 554, + 606 + ], + "type": "inline_equation", + "content": "-\\log P(\\{\\lfloor q_s\\rfloor \\})" + }, + { + "bbox": [ + 313, + 533, + 554, + 606 + ], + "type": "text", + "content": " over the training set." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 606, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 554, + 665 + ], + "type": "text", + "content": "With different codebooks for approximation and details, the same token id might have a different meaning depending on its position. Thus, the AR model has different token embeddings for approximation and details tokens. The last layer for logit prediction is also different." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "For the generative applications in Secs. 5.2 and 5.3, we introduce mild changes in order to interrupt the generation after all tokens up to a certain scale are generated, and to start the generation with given tokens up to a certain scale." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17185" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 299, + 130 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 299, + 130 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 299, + 130 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 299, + 130 + ], + "type": "image", + "image_path": "dc43ffe6e96642fcda4ec765bcd42878e8704676603cf14fde48296916c4cb76.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 131, + 252, + 138 + ], + "lines": [ + { + "bbox": [ + 104, + 131, + 252, + 138 + ], + "spans": [ + { + "bbox": [ + 104, + 131, + 252, + 138 + ], + "type": "text", + "content": "\"a zebra standing on dirty area with trees in the background.\"" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 140, + 299, + 201 + ], + "blocks": [ + { + "bbox": [ + 58, + 140, + 299, + 201 + ], + "lines": [ + { + "bbox": [ + 58, + 140, + 299, + 201 + ], + "spans": [ + { + "bbox": [ + 58, + 140, + 299, + 201 + ], + "type": "image", + "image_path": "e248b9896775e4fae5bad5b558a325a77445733fe2d81b49fc5fd2019b6f9431.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 100, + 201, + 256, + 207 + ], + "lines": [ + { + "bbox": [ + 100, + 201, + 256, + 207 + ], + "spans": [ + { + "bbox": [ + 100, + 201, + 256, + 207 + ], + "type": "text", + "content": "\"an elaborate metal vase holds a decorative bouquet of flowers.\"" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 209, + 299, + 270 + ], + "blocks": [ + { + "bbox": [ + 58, + 209, + 299, + 270 + ], + "lines": [ + { + "bbox": [ + 58, + 209, + 299, + 270 + ], + "spans": [ + { + "bbox": [ + 58, + 209, + 299, + 270 + ], + "type": "image", + "image_path": "de3eb6007c31642354f152e1bd512c4eea36619650b3f8dfad3b853f0ee1e278.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 270, + 252, + 277 + ], + "lines": [ + { + "bbox": [ + 104, + 270, + 252, + 277 + ], + "spans": [ + { + "bbox": [ + 104, + 270, + 252, + 277 + ], + "type": "text", + "content": "\"the wing of a plane located in the sky above a sea and city.\"" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 286, + 555, + 319 + ], + "lines": [ + { + "bbox": [ + 54, + 286, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 54, + 286, + 555, + 319 + ], + "type": "text", + "content": "Figure 3. Coarse-to-fine text-to-image on MS-COCO [27] prompts. Each quadruple shows generations from AR-SIT-SCD for the given prompt, with only the first " + }, + { + "bbox": [ + 54, + 286, + 555, + 319 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 54, + 286, + 555, + 319 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 286, + 555, + 319 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 54, + 286, + 555, + 319 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 286, + 555, + 319 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 54, + 286, + 555, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 286, + 555, + 319 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 54, + 286, + 555, + 319 + ], + "type": "text", + "content": " of tokens generated, corresponding to resolutions of " + }, + { + "bbox": [ + 54, + 286, + 555, + 319 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 54, + 286, + 555, + 319 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 54, + 286, + 555, + 319 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 54, + 286, + 555, + 319 + ], + "type": "text", + "content": ". Training uses only " + }, + { + "bbox": [ + 54, + 286, + 555, + 319 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 54, + 286, + 555, + 319 + ], + "type": "text", + "content": " images. Our model enables quick generation of coarse image candidates that can be further improved if needed." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 70, + 553, + 131 + ], + "blocks": [ + { + "bbox": [ + 310, + 70, + 553, + 131 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 553, + 131 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 553, + 131 + ], + "type": "image", + "image_path": "25862973e1bf70f60bd7575e0f87c543fe55309df8ecab772e174e9998dc63f2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 131, + 492, + 138 + ], + "lines": [ + { + "bbox": [ + 372, + 131, + 492, + 138 + ], + "spans": [ + { + "bbox": [ + 372, + 131, + 492, + 138 + ], + "type": "text", + "content": "\"a modern bathroom is shown with a square sink.\"" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 311, + 140, + 553, + 201 + ], + "blocks": [ + { + "bbox": [ + 311, + 140, + 553, + 201 + ], + "lines": [ + { + "bbox": [ + 311, + 140, + 553, + 201 + ], + "spans": [ + { + "bbox": [ + 311, + 140, + 553, + 201 + ], + "type": "image", + "image_path": "f33eabf0788eab36bb4215f991b3e7ffd68e526e0ff44c2644f486c2dba10819.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 201, + 470, + 208 + ], + "lines": [ + { + "bbox": [ + 395, + 201, + 470, + 208 + ], + "spans": [ + { + "bbox": [ + 395, + 201, + 470, + 208 + ], + "type": "text", + "content": "\"a toy boat sits on the ground.\"" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 311, + 209, + 553, + 270 + ], + "blocks": [ + { + "bbox": [ + 311, + 209, + 553, + 270 + ], + "lines": [ + { + "bbox": [ + 311, + 209, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 311, + 209, + 553, + 270 + ], + "type": "image", + "image_path": "710941c16491f7e36dfe7d85aeba52cd9c1e9fd64813e64988728e31bbccaf38.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 270, + 517, + 277 + ], + "lines": [ + { + "bbox": [ + 347, + 270, + 517, + 277 + ], + "spans": [ + { + "bbox": [ + 347, + 270, + 517, + 277 + ], + "type": "text", + "content": "\"This brown and white cat is standing on a desk in front of a computer.\"" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 325, + 136, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 325, + 136, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 325, + 136, + 339 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 346, + 296, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 296, + 441 + ], + "type": "text", + "content": "We focus our experiments on demonstrating the properties and applications enumerated in P.1-P.6 (Sec. 1). Our ideas build on ViT-based tokenizers, so the most fair comparisons are against ViT tokenizers and AR generative models; namely ViT-VQGAN [57] and Parti [58]. In these comparisons we are able to match the architectures and training protocol exactly. Nevertheless we also compare with a broader set of methods in Sec. 5.5." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 453, + 198, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 198, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 198, + 464 + ], + "type": "text", + "content": "5.1. Multiscale reconstruction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 471, + 295, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 295, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 295, + 507 + ], + "type": "text", + "content": "We train our tokenizer on ImageNet [10] and evaluate its reconstruction performance. We follow the ViT-VQGAN [57] \"Small\" architecture and training protocol." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 508, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 295, + 628 + ], + "type": "text", + "content": "The model \"SIT-5\" is the Spectral Image Tokenizer with 5 scales from " + }, + { + "bbox": [ + 55, + 508, + 295, + 628 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 55, + 508, + 295, + 628 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 508, + 295, + 628 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 55, + 508, + 295, + 628 + ], + "type": "text", + "content": " resolutions, where each scale is represented by 256 tokens. The variation \"SIT-SC\" uses scale-causal attention on both the encoder and decoder, which enables handling inputs and outputs of different resolutions, even though it is only trained at " + }, + { + "bbox": [ + 55, + 508, + 295, + 628 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 508, + 295, + 628 + ], + "type": "text", + "content": ". For example, when the input image is " + }, + { + "bbox": [ + 55, + 508, + 295, + 628 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 55, + 508, + 295, + 628 + ], + "type": "text", + "content": ", only the first three scales are used, resulting in shorter sequence lengths which reduces memory utilization and processing time. Both models employ the ADTransformer described in Sec. 4.1." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 628, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 295, + 689 + ], + "type": "text", + "content": "The ViT-VQGAN baseline only works for " + }, + { + "bbox": [ + 55, + 628, + 295, + 689 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 628, + 295, + 689 + ], + "type": "text", + "content": " inputs, so to evaluate against it fairly, we upsampled the low-resolution inputs to that resolution. This brings the input patches away from the training distribution, which might explain the drop in reconstruction quality." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "We also evaluate at " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": ". The ViT-VQGAN baseline increases its patch size from " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": ", keep-" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 326, + 553, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 553, + 386 + ], + "type": "text", + "content": "ing the sequence length constant. Interestingly, the baseline suffered heavy instability during training, which was not resolved by reducing the learning rate, using spectral normalization, or the logit-laplace loss. In contrast, our SIT variations trained successfully with no hyperparameter changes." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 387, + 554, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 554, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 554, + 422 + ], + "type": "text", + "content": "Tab. 1 shows the reconstruction metrics; Fig. 7 shows reconstruction samples at multiple resolutions. Results provide evidence for P.1 and demonstrate P.2 (Sec. 1)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 431, + 522, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 522, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 522, + 445 + ], + "type": "text", + "content": "5.2. Coarse-to-fine text-to-image generation" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 449, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 554, + 533 + ], + "type": "text", + "content": "We tackle text-to-image generation by using an autoregressive transformer to model the distribution of discrete tokens output by our tokenizer. Since the SIT sequence of tokens represents an image in a coarse-fine fashion, the autoregressive model generation has the same property, which means that we can interrupt the generation after a certain number of tokens and decode a coarse version of the generation." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 533, + 554, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 554, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 554, + 594 + ], + "type": "text", + "content": "For this to work, the SIT decoder must be scale-causal, while there is no constraint for the encoder so we use dense attention (we denote this variant Scale-Causal Decoder, or \"SIT-SCD\"). It uses a \"small\" encoder and \"large\" decoder as described in ViT-VQGAN [57], and 4 scales." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "We follow the Parti350M [58] architecture and training protocol and name our generative model auto-regressive SIT, or \"AR-SIT\". SIT and AR-SIT are trained on a subset of WebLI [6] of around 128M images, where each image is seen at most once during training. Evaluation is on 30k examples of MS-COCO [27]. We show metrics and generations at different resolutions in Fig. 3 and Tab. 2. AR-SIT matches the baseline performance but is several times faster and more memory efficient at lower resolutions. These results demonstrate the property P.4 (Sec. 1)." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17186" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 136, + 150 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 136, + 150 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 136, + 150 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 136, + 150 + ], + "type": "image", + "image_path": "f6682868634b64cec23c47bc161205192494d27dd60c2e7eb70a3346c8aad53c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 151, + 276, + 160 + ], + "lines": [ + { + "bbox": [ + 80, + 151, + 276, + 160 + ], + "spans": [ + { + "bbox": [ + 80, + 151, + 276, + 160 + ], + "type": "text", + "content": "\"an assortment of some colorful vases on display on a table\"" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 138, + 71, + 217, + 150 + ], + "blocks": [ + { + "bbox": [ + 138, + 71, + 217, + 150 + ], + "lines": [ + { + "bbox": [ + 138, + 71, + 217, + 150 + ], + "spans": [ + { + "bbox": [ + 138, + 71, + 217, + 150 + ], + "type": "image", + "image_path": "6864d3bfd29c46ef47320f2be74f6fdfb48f5d5df51d90f044cab6eda868f28a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 219, + 71, + 298, + 150 + ], + "blocks": [ + { + "bbox": [ + 219, + 71, + 298, + 150 + ], + "lines": [ + { + "bbox": [ + 219, + 71, + 298, + 150 + ], + "spans": [ + { + "bbox": [ + 219, + 71, + 298, + 150 + ], + "type": "image", + "image_path": "8b439c1c4445e8bb03f450c3747720ece9b8ca68277d231315d84b1d2d995342.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 163, + 137, + 243 + ], + "blocks": [ + { + "bbox": [ + 58, + 163, + 137, + 243 + ], + "lines": [ + { + "bbox": [ + 58, + 163, + 137, + 243 + ], + "spans": [ + { + "bbox": [ + 58, + 163, + 137, + 243 + ], + "type": "image", + "image_path": "517ebeae81c58a9d8c5eff6784c15ff05c0996997375cf294898fd3a0f22c118.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 243, + 270, + 253 + ], + "lines": [ + { + "bbox": [ + 86, + 243, + 270, + 253 + ], + "spans": [ + { + "bbox": [ + 86, + 243, + 270, + 253 + ], + "type": "text", + "content": "\"one train riding on the multiple train tracks side by side.\"" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 138, + 163, + 217, + 243 + ], + "blocks": [ + { + "bbox": [ + 138, + 163, + 217, + 243 + ], + "lines": [ + { + "bbox": [ + 138, + 163, + 217, + 243 + ], + "spans": [ + { + "bbox": [ + 138, + 163, + 217, + 243 + ], + "type": "image", + "image_path": "83ed571691f19876c737ebdc8743125915e5cd1d2239ea506d6fe5a93bb4bb8f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 219, + 163, + 298, + 243 + ], + "blocks": [ + { + "bbox": [ + 219, + 163, + 298, + 243 + ], + "lines": [ + { + "bbox": [ + 219, + 163, + 298, + 243 + ], + "spans": [ + { + "bbox": [ + 219, + 163, + 298, + 243 + ], + "type": "image", + "image_path": "1a458d4fa8ab0ab3e5ed942ae2f094554c544dd10955bb5af9dcdb9da4d57813.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 58, + 256, + 137, + 335 + ], + "blocks": [ + { + "bbox": [ + 58, + 256, + 137, + 335 + ], + "lines": [ + { + "bbox": [ + 58, + 256, + 137, + 335 + ], + "spans": [ + { + "bbox": [ + 58, + 256, + 137, + 335 + ], + "type": "image", + "image_path": "3dd72cc01a64562a44a1ef23d88bfbc31770459cf258ae88a315876675ba6fcd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 110, + 336, + 246, + 346 + ], + "lines": [ + { + "bbox": [ + 110, + 336, + 246, + 346 + ], + "spans": [ + { + "bbox": [ + 110, + 336, + 246, + 346 + ], + "type": "text", + "content": "\"an image of a fired hydrant on the street\"" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 354, + 555, + 389 + ], + "lines": [ + { + "bbox": [ + 54, + 354, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 54, + 354, + 555, + 389 + ], + "type": "text", + "content": "Figure 4. Text-guided image upsampling on MS-COCO [27]. Our coarse-to-fine generative models can take a low-resolution image, encode it as the first few tokens of a sequence, and generate the rest of sequence, which, when decoded, effectively upsamples the input. Each triplet shows the given " + }, + { + "bbox": [ + 54, + 354, + 555, + 389 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 54, + 354, + 555, + 389 + ], + "type": "text", + "content": " image, our " + }, + { + "bbox": [ + 54, + 354, + 555, + 389 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 54, + 354, + 555, + 389 + ], + "type": "text", + "content": " reconstruction and the ground truth." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 138, + 256, + 217, + 335 + ], + "blocks": [ + { + "bbox": [ + 138, + 256, + 217, + 335 + ], + "lines": [ + { + "bbox": [ + 138, + 256, + 217, + 335 + ], + "spans": [ + { + "bbox": [ + 138, + 256, + 217, + 335 + ], + "type": "image", + "image_path": "841d651e744429d21dc2987e5397e2c7a88724fd690efd869b771cbe44f39aa7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 219, + 256, + 298, + 335 + ], + "blocks": [ + { + "bbox": [ + 219, + 256, + 298, + 335 + ], + "lines": [ + { + "bbox": [ + 219, + 256, + 298, + 335 + ], + "spans": [ + { + "bbox": [ + 219, + 256, + 298, + 335 + ], + "type": "image", + "image_path": "ac38bdf79291ffd3ab1d63ed72d49ddcd4bdc91df8eaf45bdd23a839dc5e1b20.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 312, + 70, + 391, + 150 + ], + "blocks": [ + { + "bbox": [ + 312, + 70, + 391, + 150 + ], + "lines": [ + { + "bbox": [ + 312, + 70, + 391, + 150 + ], + "spans": [ + { + "bbox": [ + 312, + 70, + 391, + 150 + ], + "type": "image", + "image_path": "7d60852496ac37397b718e1b30f498f7a70098acac0efb310597ad1dab13ca2a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 338, + 151, + 527, + 160 + ], + "lines": [ + { + "bbox": [ + 338, + 151, + 527, + 160 + ], + "spans": [ + { + "bbox": [ + 338, + 151, + 527, + 160 + ], + "type": "text", + "content": "\"a large elephant standing on top of a grass covered field.\"" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 392, + 71, + 472, + 149 + ], + "blocks": [ + { + "bbox": [ + 392, + 71, + 472, + 149 + ], + "lines": [ + { + "bbox": [ + 392, + 71, + 472, + 149 + ], + "spans": [ + { + "bbox": [ + 392, + 71, + 472, + 149 + ], + "type": "image", + "image_path": "7147839eb4f3185aab90b428d0b70471f566d9b819fde23f2cde8f026ed5c2d3.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 473, + 71, + 553, + 150 + ], + "blocks": [ + { + "bbox": [ + 473, + 71, + 553, + 150 + ], + "lines": [ + { + "bbox": [ + 473, + 71, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 473, + 71, + 553, + 150 + ], + "type": "image", + "image_path": "7ca9cd0198cae27ae5efdec7451fd2e87c474908da76f7885961a5e3d24e50d4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 312, + 163, + 392, + 243 + ], + "blocks": [ + { + "bbox": [ + 312, + 163, + 392, + 243 + ], + "lines": [ + { + "bbox": [ + 312, + 163, + 392, + 243 + ], + "spans": [ + { + "bbox": [ + 312, + 163, + 392, + 243 + ], + "type": "image", + "image_path": "09890d806c9622da8906b588b07da60bbe2e63f1dbd4ee7306ab2a57116e4b3d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 243, + 525, + 253 + ], + "lines": [ + { + "bbox": [ + 340, + 243, + 525, + 253 + ], + "spans": [ + { + "bbox": [ + 340, + 243, + 525, + 253 + ], + "type": "text", + "content": "\"three tropical bird perched on top of high bare branches\"" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 392, + 163, + 472, + 243 + ], + "blocks": [ + { + "bbox": [ + 392, + 163, + 472, + 243 + ], + "lines": [ + { + "bbox": [ + 392, + 163, + 472, + 243 + ], + "spans": [ + { + "bbox": [ + 392, + 163, + 472, + 243 + ], + "type": "image", + "image_path": "59ac6e58680042f11f5f3a0119b5af6091a722258fd001e25739170221fc02b0.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 473, + 163, + 553, + 243 + ], + "blocks": [ + { + "bbox": [ + 473, + 163, + 553, + 243 + ], + "lines": [ + { + "bbox": [ + 473, + 163, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 473, + 163, + 553, + 243 + ], + "type": "image", + "image_path": "3587c8b075c0638cfffa457a21b7a195201df8f66d0c9cf958a65b203f57af7a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 312, + 256, + 391, + 335 + ], + "blocks": [ + { + "bbox": [ + 312, + 256, + 391, + 335 + ], + "lines": [ + { + "bbox": [ + 312, + 256, + 391, + 335 + ], + "spans": [ + { + "bbox": [ + 312, + 256, + 391, + 335 + ], + "type": "image", + "image_path": "bea17c977521287e89bcf60654bbe228b5d095d91272b5755f17af6822cf67d5.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 336, + 515, + 346 + ], + "lines": [ + { + "bbox": [ + 350, + 336, + 515, + 346 + ], + "spans": [ + { + "bbox": [ + 350, + 336, + 515, + 346 + ], + "type": "text", + "content": "\"a chair with a pink blanket is sitting on a red floor.\"" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 392, + 256, + 472, + 335 + ], + "blocks": [ + { + "bbox": [ + 392, + 256, + 472, + 335 + ], + "lines": [ + { + "bbox": [ + 392, + 256, + 472, + 335 + ], + "spans": [ + { + "bbox": [ + 392, + 256, + 472, + 335 + ], + "type": "image", + "image_path": "843e97f846207e282a99007ba5204d3856f4d925c573a79c3727719130dec0b0.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 473, + 256, + 553, + 335 + ], + "blocks": [ + { + "bbox": [ + 473, + 256, + 553, + 335 + ], + "lines": [ + { + "bbox": [ + 473, + 256, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 473, + 256, + 553, + 335 + ], + "type": "image", + "image_path": "f367404a2ed76834b55032af67c3781a729a479906a959078f6746c6164bdd31.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 394, + 222, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 394, + 222, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 222, + 407 + ], + "type": "text", + "content": "5.3. Text-guided image upsampling" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 411, + 295, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 295, + 448 + ], + "type": "text", + "content": "We leverage the coarse-to-fine nature of SIT to apply a pretrained AR-SIT for text-guided image upsampling, where we are given a low resolution image and a caption." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 448, + 295, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 295, + 543 + ], + "type": "text", + "content": "The idea is to encode the low-resolution image, which will give the first tokens of our high resolution output. AR-SIT then takes these tokens and generates the rest of the sequence. For this to work, the SIT encoder must be scale-causal to properly tokenize low-resolution inputs, while there is no constraint for the decoder so we use dense attention. We denote this variant Scale-Causal Encoder, or \"SIT-SCE\". The model is otherwise as the one used in Sec. 5.2." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 544, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 604 + ], + "type": "text", + "content": "Here the inputs are " + }, + { + "bbox": [ + 55, + 544, + 295, + 604 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 55, + 544, + 295, + 604 + ], + "type": "text", + "content": ". Fig. 4 shows text-guided upsampling examples on MS-COCO [27]. We obtain an FID of 6.2 when evaluating these generations, compared to 12.6 when given only the prompts. Fig. 8 and Tab. 4 show additional results upsampling from " + }, + { + "bbox": [ + 55, + 544, + 295, + 604 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 55, + 544, + 295, + 604 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 55, + 611, + 200, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 200, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 200, + 624 + ], + "type": "text", + "content": "5.4. Text-guided image editing" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 55, + 629, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 677 + ], + "type": "text", + "content": "Our coarse-to-fine AR-SIT enables a text-guided image editing application, where we want to change image details while keeping the same overall appearance, which corresponds to freezing lower scales while generating higher." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "We apply an AR-SIT trained only on the maximum likelihood objective as follows. The given image is tokenized only up to the first scale, corresponding to a coarse repre" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 395, + 555, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 555, + 442 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 555, + 442 + ], + "type": "text", + "content": "sensation. The tokenizer encoder must be scale-causal in order to avoid leakage from high to low scales, so we use SIT-SCE here. Now we use AR-SIT to generate the rest of the sequence, conditioning on the textual caption." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 444, + 555, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 555, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 555, + 528 + ], + "type": "text", + "content": "In this experiment, we use a 5-scale model such that the lowest resolution is " + }, + { + "bbox": [ + 313, + 444, + 555, + 528 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 313, + 444, + 555, + 528 + ], + "type": "text", + "content": "; we found that starting with higher resolutions limits the changes the model can generate. Fig. 5 shows some examples of text-guided editing on MS-COCO [27], where we lightly modify the original captions, for example by swapping \"elephants\" with \"cows\". We show additional results in Fig. 6 in the supplemental." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 538, + 501, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 501, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 501, + 551 + ], + "type": "text", + "content": "5.5. Class-conditional image generation" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 556, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 555, + 628 + ], + "type": "text", + "content": "Here we evaluate class-conditional generation on " + }, + { + "bbox": [ + 313, + 556, + 555, + 628 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 556, + 555, + 628 + ], + "type": "text", + "content": " ImageNet. We use the SIT-4 tokenizer from Tab. 1 and follow ViT-VQGAN's [57] VIM-Base AR architecture and training protocol. Tab. 3 shows we outperform the baseline in this fair comparison which is evidence for the improved conditioning described in P.3. Fig. 9 shows samples." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "content": "AR-SIT-4* departs from ViT-VQGAN and Parti hyperparameters to compare with more recent works, by increasing the perceptual and adversarial loss weights, doubling the tokenizer codebook, reducing its latent dimension, training with a constant learning rate, including 2D RoPE [44] and GeGLU [42], in both the tokenizer and generator. We train the AR model longer and since there is no text encoder, we" + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17187" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 137, + 150 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 137, + 150 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 137, + 150 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 137, + 150 + ], + "type": "image", + "image_path": "36ecb5f118fe41bcc2c59633b2689ae5406defc5a4d24fa7a3a57028d53d54c0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 151, + 221, + 160 + ], + "lines": [ + { + "bbox": [ + 135, + 151, + 221, + 160 + ], + "spans": [ + { + "bbox": [ + 135, + 151, + 221, + 160 + ], + "type": "text", + "content": "\"a close-up of a dog face.\"" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 137, + 71, + 217, + 150 + ], + "blocks": [ + { + "bbox": [ + 137, + 71, + 217, + 150 + ], + "lines": [ + { + "bbox": [ + 137, + 71, + 217, + 150 + ], + "spans": [ + { + "bbox": [ + 137, + 71, + 217, + 150 + ], + "type": "image", + "image_path": "eaa16f39af62397d6cb9924f8124b28531b5cf0c7837c487bb5b87773e0227e3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 217, + 71, + 298, + 150 + ], + "blocks": [ + { + "bbox": [ + 217, + 71, + 298, + 150 + ], + "lines": [ + { + "bbox": [ + 217, + 71, + 298, + 150 + ], + "spans": [ + { + "bbox": [ + 217, + 71, + 298, + 150 + ], + "type": "image", + "image_path": "8b7a5d722eedde08ddd90ded0d73e568da830eff1aa825e3f59c8feafb447751.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 167, + 137, + 243 + ], + "blocks": [ + { + "bbox": [ + 58, + 167, + 137, + 243 + ], + "lines": [ + { + "bbox": [ + 58, + 167, + 137, + 243 + ], + "spans": [ + { + "bbox": [ + 58, + 167, + 137, + 243 + ], + "type": "image", + "image_path": "e1bdf2e9ecfc4fc44af3a82b3207598dc95b2058bc453fd53d75c2e575b26e44.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 112, + 243, + 244, + 253 + ], + "lines": [ + { + "bbox": [ + 112, + 243, + 244, + 253 + ], + "spans": [ + { + "bbox": [ + 112, + 243, + 244, + 253 + ], + "type": "text", + "content": "\"a couple of cows are standing in a field\"" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 137, + 167, + 217, + 243 + ], + "blocks": [ + { + "bbox": [ + 137, + 167, + 217, + 243 + ], + "lines": [ + { + "bbox": [ + 137, + 167, + 217, + 243 + ], + "spans": [ + { + "bbox": [ + 137, + 167, + 217, + 243 + ], + "type": "image", + "image_path": "4b686611eccd2b3eb076ee9970ff9de5fa2e2ee260cf20e00d6c7b996ea6c511.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 217, + 167, + 298, + 243 + ], + "blocks": [ + { + "bbox": [ + 217, + 167, + 298, + 243 + ], + "lines": [ + { + "bbox": [ + 217, + 167, + 298, + 243 + ], + "spans": [ + { + "bbox": [ + 217, + 167, + 298, + 243 + ], + "type": "image", + "image_path": "fba81ff6c5e4cb8a1d079101361afbc84072e7398dc85afe295e398c9d708fd4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 58, + 256, + 137, + 335 + ], + "blocks": [ + { + "bbox": [ + 58, + 256, + 137, + 335 + ], + "lines": [ + { + "bbox": [ + 58, + 256, + 137, + 335 + ], + "spans": [ + { + "bbox": [ + 58, + 256, + 137, + 335 + ], + "type": "image", + "image_path": "17dafd257af776ffd48630fd378c65e9aa0eae746acc11a8c705d00d85eeb144.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 123, + 337, + 233, + 346 + ], + "lines": [ + { + "bbox": [ + 123, + 337, + 233, + 346 + ], + "spans": [ + { + "bbox": [ + 123, + 337, + 233, + 346 + ], + "type": "text", + "content": "\"a robot on a grass covered field.\"" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 354, + 555, + 399 + ], + "lines": [ + { + "bbox": [ + 55, + 354, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 555, + 399 + ], + "type": "text", + "content": "Figure 5. Text-guided image editing on MS-COCO [27]. Our coarse-to-fine generative models can do text-guided editing by encoding a given image but keeping only the lower scales, and using a pre-trained AR-SIT to re-generate the higher scales conditioned on the textual prompt. Each triplet shows the given image, its " + }, + { + "bbox": [ + 55, + 354, + 555, + 399 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 55, + 354, + 555, + 399 + ], + "type": "text", + "content": " reconstruction using only the coefficients used to start the generation, and the edited image after generating the whole sequence." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 137, + 256, + 217, + 335 + ], + "blocks": [ + { + "bbox": [ + 137, + 256, + 217, + 335 + ], + "lines": [ + { + "bbox": [ + 137, + 256, + 217, + 335 + ], + "spans": [ + { + "bbox": [ + 137, + 256, + 217, + 335 + ], + "type": "image", + "image_path": "15650ebe8f5fb1e14f5cbff4550b8905037edd0aa3d798bc3ff42b08b442e51b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 217, + 256, + 298, + 335 + ], + "blocks": [ + { + "bbox": [ + 217, + 256, + 298, + 335 + ], + "lines": [ + { + "bbox": [ + 217, + 256, + 298, + 335 + ], + "spans": [ + { + "bbox": [ + 217, + 256, + 298, + 335 + ], + "type": "image", + "image_path": "5ee81e4ebd6fea49160ac75d835979d3bf656adeb5b1e4f03fcd0468c8bb7029.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 312, + 71, + 391, + 150 + ], + "blocks": [ + { + "bbox": [ + 312, + 71, + 391, + 150 + ], + "lines": [ + { + "bbox": [ + 312, + 71, + 391, + 150 + ], + "spans": [ + { + "bbox": [ + 312, + 71, + 391, + 150 + ], + "type": "image", + "image_path": "a4f5f01bcb766309cac7140a75cfd6400b2de4c276273946f37403fb1790c353.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 391, + 71, + 472, + 150 + ], + "blocks": [ + { + "bbox": [ + 391, + 71, + 472, + 150 + ], + "lines": [ + { + "bbox": [ + 391, + 71, + 472, + 150 + ], + "spans": [ + { + "bbox": [ + 391, + 71, + 472, + 150 + ], + "type": "image", + "image_path": "a269ec68a99d50ebee0f6707b77c78d00284cdcb065d5851444d20dad58912dd.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 151, + 471, + 159 + ], + "lines": [ + { + "bbox": [ + 394, + 151, + 471, + 159 + ], + "spans": [ + { + "bbox": [ + 394, + 151, + 471, + 159 + ], + "type": "text", + "content": "\"some boats on a lake.\"" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 472, + 71, + 553, + 150 + ], + "blocks": [ + { + "bbox": [ + 472, + 71, + 553, + 150 + ], + "lines": [ + { + "bbox": [ + 472, + 71, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 472, + 71, + 553, + 150 + ], + "type": "image", + "image_path": "ea23faefa17f90054e42473d5aea37b128c8b8013a982acaeddc599c646f77d1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 312, + 163, + 391, + 243 + ], + "blocks": [ + { + "bbox": [ + 312, + 163, + 391, + 243 + ], + "lines": [ + { + "bbox": [ + 312, + 163, + 391, + 243 + ], + "spans": [ + { + "bbox": [ + 312, + 163, + 391, + 243 + ], + "type": "image", + "image_path": "0d30afa8e30a582f30149e98842e7a669a3e05b5821eb8de5f7de5dd36265abf.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 368, + 243, + 496, + 253 + ], + "lines": [ + { + "bbox": [ + 368, + 243, + 496, + 253 + ], + "spans": [ + { + "bbox": [ + 368, + 243, + 496, + 253 + ], + "type": "text", + "content": "\"a green pasture filled with wildflowers.\"" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 391, + 163, + 472, + 243 + ], + "blocks": [ + { + "bbox": [ + 391, + 163, + 472, + 243 + ], + "lines": [ + { + "bbox": [ + 391, + 163, + 472, + 243 + ], + "spans": [ + { + "bbox": [ + 391, + 163, + 472, + 243 + ], + "type": "image", + "image_path": "ff89ab629159c0c08a685fa6f925aba8900b484de8f29dc646b6ec676b76173c.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 472, + 163, + 553, + 243 + ], + "blocks": [ + { + "bbox": [ + 472, + 163, + 553, + 243 + ], + "lines": [ + { + "bbox": [ + 472, + 163, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 472, + 163, + 553, + 243 + ], + "type": "image", + "image_path": "b5f7951518fe619b80069c74c9078348110e9869631ac65634c35b482c067907.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 312, + 256, + 391, + 335 + ], + "blocks": [ + { + "bbox": [ + 312, + 256, + 391, + 335 + ], + "lines": [ + { + "bbox": [ + 312, + 256, + 391, + 335 + ], + "spans": [ + { + "bbox": [ + 312, + 256, + 391, + 335 + ], + "type": "image", + "image_path": "acb38a9ac3421cb4c5aa1397614270a07ffe99a30a75b4517d1a1f8d2e534564.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 336, + 504, + 346 + ], + "lines": [ + { + "bbox": [ + 361, + 336, + 504, + 346 + ], + "spans": [ + { + "bbox": [ + 361, + 336, + 504, + 346 + ], + "type": "text", + "content": "\"a very cute dog laying in a very small sink.\"" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 391, + 256, + 472, + 335 + ], + "blocks": [ + { + "bbox": [ + 391, + 256, + 472, + 335 + ], + "lines": [ + { + "bbox": [ + 391, + 256, + 472, + 335 + ], + "spans": [ + { + "bbox": [ + 391, + 256, + 472, + 335 + ], + "type": "image", + "image_path": "1a1b11602e9329617c11e9f5b02cf0445c2342b98f5671840a234d1a3534b577.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 472, + 256, + 553, + 335 + ], + "blocks": [ + { + "bbox": [ + 472, + 256, + 553, + 335 + ], + "lines": [ + { + "bbox": [ + 472, + 256, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 472, + 256, + 553, + 335 + ], + "type": "image", + "image_path": "caa571ccd6daeecd2b24d2945ccbd8468bc354ee21baf542d4cdf8af06cfa34d.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "table", + "bbox": [ + 57, + 404, + 294, + 514 + ], + "blocks": [ + { + "bbox": [ + 57, + 404, + 294, + 514 + ], + "lines": [ + { + "bbox": [ + 57, + 404, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 57, + 404, + 294, + 514 + ], + "type": "table", + "html": "
paramsseq lenCFGFID ↓IS ↑
AR-ViT-VQGAN (reported)650M1024-8.81110.8
AR-ViT-VQGAN (reproduced)650M1024-8.37111.8
AR-SIT-4 (Ours)650M1024-6.95138.3
RQ-VAE [26]480M256-15.7286.8
DQ-VAE [22]355M640-7.34152.8
LlamaGen-L [45]343M2561.54.08198.5
VAR [48]310M6802.03.30274.4
AR-SIT-4* (Ours)350M10241.54.06190.9
", + "image_path": "fe7dc44f4c4262e5cf1e48adcced02f419151f31341155e9b61616f04ff51987.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 523, + 295, + 612 + ], + "lines": [ + { + "bbox": [ + 55, + 523, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 523, + 295, + 612 + ], + "type": "text", + "content": "Table 3. ImageNet " + }, + { + "bbox": [ + 55, + 523, + 295, + 612 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 523, + 295, + 612 + ], + "type": "text", + "content": " class conditional generation. Top half shows a fair comparison against ViT-VQGAN, which we clearly outperform. AR-SIT-4* uses different hyperparameters for comparison against a broader class of methods. While we do not outperform all of them, there are differences in the architectures and training schedules that influence the results and are orthogonal to our contributions. Moreover, none of the baselines has the multiscale capabilities described in P.2 and P.4 (Sec. 1)." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": "double the number of decoder layers. The bottom half of Tab. 3 compares with similarly-sized AR models including some with multiscale/residual latents (although none exhibit properties P.2 and P.4). AR-SIT-4* performance matches LlamaGen but not to VAR, but the models are not exactly comparable since they use a convolutional tokenizers and other improvements such as AdaLN [34] and attention normalization; VAR also trains the tokenizer on larger" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 406, + 553, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 553, + 430 + ], + "type": "text", + "content": "data than ImageNet. These are orthogonal to our contributions and can be incorporated in future work." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 431, + 554, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 554, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 554, + 480 + ], + "type": "text", + "content": "We also train an AR-SIT-5* on " + }, + { + "bbox": [ + 313, + 431, + 554, + 480 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 431, + 554, + 480 + ], + "type": "text", + "content": " ImageNet and obtain an FID of 5.97 and IS of 215.22, showing robustness at higher resolutions, while most prior work requires separate stages for upsampling. Fig. 10 shows samples." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 497, + 467, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 467, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 467, + 510 + ], + "type": "text", + "content": "6. Conclusion and limitations" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 520, + 554, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 554, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 554, + 604 + ], + "type": "text", + "content": "We presented a spectral image tokenizer (SIT), and an autoregressive generative transformer trained with it (AR-SIT). SIT is naturally multiscale and leverages spectral properties of natural images for improved reconstruction quality. AR-SIT enables applications such as rapid generation of coarse images that can be refined later, and text-guided image upsampling and editing." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": "While our methods improved tokenizer reconstruction accuracy and class-conditional generation, the text-to-image metrics were similar to the fair baseline. Chang et al. [4] similarly observed that a better tokenizer does not necessarily lead to a better generative model. Nevertheless, our method has multiscale properties and enables new applications not possible with prior work. We only experiment with realtively small AR-SITs of 350-650M parameters, while the Parti [58] baseline goes up to 22B parameters." + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17188" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "7. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 128 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 128 + ], + "type": "text", + "content": "We thank Leonardo Zepeda-Nuñez for reviewing this manuscript and offering interesting discussions and suggestions, and Jon Barron for sharing useful code we relied on." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 115, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 115, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 115, + 152 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 159, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 159, + 296, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 159, + 296, + 215 + ], + "spans": [ + { + "bbox": [ + 61, + 159, + 296, + 215 + ], + "type": "text", + "content": "[1] James Betker, Gabriel Goh, Li Jing, † TimBrooks, Jianfeng Wang, Linjie Li, † LongOuyang, † JuntangZhuang, † JoyceLee, † YufeiGuo, † WesamManassra, † PrafullaDhariwal, † CaseyChu, † YunxinJiao, and Aditya Ramesh. Improving image generation with better captions. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 217, + 296, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 217, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 61, + 217, + 296, + 258 + ], + "type": "text", + "content": "[2] Lei Cai, Hongyang Gao, and Shuiwang Ji. Multi-Stage Variational Auto-Encoders for Coarse-to-Fine Image Generation, page 630–638. Society for Industrial and Applied Mathematics, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 261, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 261, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 61, + 261, + 295, + 293 + ], + "type": "text", + "content": "[3] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T. Freeman. Maskgit: Masked generative image transformer. In CVPR, 2022. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 296, + 296, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 296, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 61, + 296, + 296, + 361 + ], + "type": "text", + "content": "[4] Huiwen Chang, Han Zhang, Jarred Barber, Aaron Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Patrick Murphy, William T. Freeman, Michael Rubinstein, Yuanzhen Li, and Dilip Krishnan. Muse: Text-to-image generation via masked generative transformers. In ICML, 2023. 1, 3, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 363, + 296, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 363, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 363, + 296, + 396 + ], + "type": "text", + "content": "[5] XI Chen, Nikhil Mishra, Mostafa Rohaninejad, and Pieter Abbeel. PixelSNAIL: An improved autoregressive generative model. In ICML, 2018. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 397, + 296, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 397, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 61, + 397, + 296, + 506 + ], + "type": "text", + "content": "[6] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. Pali: A jointly-scaled multilingual language-image model, 2022. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 509, + 296, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 509, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 62, + 509, + 296, + 563 + ], + "type": "text", + "content": "[7] C. A. Christopoulos, T. Ebrahimi, and A. N. Skodras. Jpeg2000: the new still picture compression standard. In Proceedings of the 2000 ACM Workshops on Multimedia, page 45-49, New York, NY, USA, 2000. Association for Computing Machinery. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 565, + 296, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 565, + 296, + 608 + ], + "spans": [ + { + "bbox": [ + 62, + 565, + 296, + 608 + ], + "type": "text", + "content": "[8] A. Cohen, Ingrid Daubechies, and J.-C. Feauveau. Biorthogonal bases of compactly supported wavelets. Communications on Pure and Applied Mathematics, 45(5):485-560, 1992. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 610, + 295, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 610, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 62, + 610, + 295, + 632 + ], + "type": "text", + "content": "[9] Ingrid Daubechies. Ten lectures on wavelets. Society for Industrial and Applied Mathematics, USA, 1992. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 634, + 295, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 634, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 57, + 634, + 295, + 666 + ], + "type": "text", + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 668, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 668, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 668, + 295, + 690 + ], + "type": "text", + "content": "[11] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 692, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 712 + ], + "type": "text", + "content": "[12] Sander Dieleman. Diffusion is spectral autoregression, 2024. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 702 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 138 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2, 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "type": "text", + "content": "[14] Tim Elsner, Paula Usinger, Victor Czech, Gregor Kobsik, Yanjiang He, Isaak Lim, and Leif Kobbelt. Quantised global autoencoder: A holistic approach to representing visual data, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 184, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 553, + 215 + ], + "type": "text", + "content": "[15] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 1, 2, 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 217, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 281 + ], + "type": "text", + "content": "[16] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 283, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 315 + ], + "type": "text", + "content": "[17] Rinon Gal, Dana Cohen Hochberg, Amit Bermano, and Daniel Cohen-Or. Swagan: a style-based wavelet-driven generative model. ACM TOG, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 316, + 553, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 347 + ], + "type": "text", + "content": "[18] Jiatao Gu, Shuangfei Zhai, Yizhe Zhang, Joshua M. Susskind, and Navdeep Jaitly. Matryoshka diffusion models. In ICLR, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 350, + 553, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 402 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 402 + ], + "type": "text", + "content": "[19] Philippe Hansen-Estruch, David Yan, Ching-Yao Chung, Orr Zohar, Jialiang Wang, Tingbo Hou, Tao Xu, Sriram Vishwanath, Peter Vajda, and Xinlei Chen. Learnings from scaling visual tokenizers for reconstruction and generation, 2025. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 404, + 553, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 447 + ], + "type": "text", + "content": "[20] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 449, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 470 + ], + "type": "text", + "content": "[21] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "type": "text", + "content": "[22] Mengqi Huang, Zhendong Mao, Zhuowei Chen, and Yongdong Zhang. Towards accurate image coding: Improved autoregressive image generation with dynamic vector quantization. In CVPR, 2023. 2, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 515, + 488, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 488, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 488, + 525 + ], + "type": "text", + "content": "[23]Imagen-Team-Google. Imagen 3, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 526, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 557 + ], + "type": "text", + "content": "[24] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 559, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 592 + ], + "type": "text", + "content": "[25] Vladimir Kulikov, Shahar Yadin, Matan Kleiner, and Tomer Michaeli. Sinddm: A single image denoising diffusion model. In ICML, pages 17920-17930, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 593, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 553, + 624 + ], + "type": "text", + "content": "[26] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In CVPR, 2022. 2, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 625, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 553, + 668 + ], + "type": "text", + "content": "[27] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In ECCV. 5, 6, 7, 8, 2, 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 670, + 553, + 702 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 702 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 702 + ], + "type": "text", + "content": "[28] Xiaoxiao Ma, Mohan Zhou, Tao Liang, Yalong Bai, Tiejun Zhao, Huaian Chen, and Yi Jin. Star: Scale-wise text-to-image generation via auto-regressive representations, 2024." + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17189" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "type": "text", + "content": "[29] Stephane Mallat. A Wavelet Tour of Signal Processing, Third Edition: The Sparse Way. Academic Press, Inc., USA, 3rd edition, 2008. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 296, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 296, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 296, + 139 + ], + "type": "text", + "content": "[30] Wael Mattar, Idan Levy, Nir Sharon, and Shai Dekel. Wavelets are all you need for autoregressive image generation, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 296, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 296, + 173 + ], + "type": "text", + "content": "[31] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. In ICLR, 2018. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 174, + 296, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 296, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 296, + 206 + ], + "type": "text", + "content": "[32] Charlie Nash, Jacob Menick, Sander Dieleman, and Peter W. Battaglia. Generating images with sparse representations. In ICML, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 207, + 235, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 207, + 235, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 207, + 235, + 218 + ], + "type": "text", + "content": "[33] OpenAI. Gpt-4 technical report, 2023. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 219, + 296, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 296, + 240 + ], + "type": "text", + "content": "[34] William Peebles and Saining Xie. Scalable diffusion models with transformers. 2023. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 241, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 296, + 274 + ], + "type": "text", + "content": "[35] Hao Phung, Quan Dao, and Anh Tran. Wavelet diffusion models are fast and scalable image generators. In CVPR, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 276, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 296, + 319 + ], + "type": "text", + "content": "[36] Vivek Ramanujan, Kushal Tirumala, Armen Aghajanyan, Luke Zettlemoyer, and Ali Farhadi. When worse is better: Navigating the compression-generation tradeoff in visual tokenization, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 320, + 296, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 296, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 296, + 352 + ], + "type": "text", + "content": "[37] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. In NeurIPS, 2019. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 354, + 296, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 296, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 296, + 387 + ], + "type": "text", + "content": "[38] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 388, + 296, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 296, + 420 + ], + "type": "text", + "content": "[39] Tamar Rott Shaham, Tali Dekel, and Tomer Michaeli. Singan: Learning a generative model from a single natural image. In ICCV, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 421, + 296, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 296, + 443 + ], + "type": "text", + "content": "[40] Dohoon Ryu and Jong Chul Ye. Pyramidal denoising diffusion probabilistic models, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 444, + 296, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 296, + 477 + ], + "type": "text", + "content": "[41] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, Xi Chen, and Xi Chen. Improved techniques for training gans. In NeurIPS, 2016. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 478, + 288, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 288, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 288, + 488 + ], + "type": "text", + "content": "[42] Noam Shazeer. Glu variants improve transformer, 2020. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 489, + 296, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 296, + 522 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 296, + 522 + ], + "type": "text", + "content": "[43] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 522, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 522, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 522, + 296, + 555 + ], + "type": "text", + "content": "[44] Jianlin Su, Yu Lu, Shengfeng Pan, Bo Wen, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding, 2021. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 556, + 296, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 296, + 601 + ], + "type": "text", + "content": "[45] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 2, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 56, + 601, + 296, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 296, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 296, + 623 + ], + "type": "text", + "content": "[46] Gemini Team. Gemini: A family of highly capable multimodal models, 2023. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 56, + 624, + 267, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 267, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 267, + 635 + ], + "type": "text", + "content": "[47] Llama team. The llama 3 herd of models, 2024. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 56, + 635, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 296, + 669 + ], + "type": "text", + "content": "[48] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. In NeurIPS, 2024. 2, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "type": "text", + "content": "[49] Aaron van den Oord, Nal Kalchbrenner, Lasse Espeholt, kory kavukcuoglu, Oriol Vinyals, and Alex Graves. Conditional image generation with pixelCNN decoders. In NeurIPS, 2016. 2" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 599 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[50] Aäron van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In ICML, 2016. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "type": "text", + "content": "[51] Aaron van den Oord, Oriol Vinyals, and koray kavukcuoglu. Neural discrete representation learning. In NeurIPS, 2017. 1, 2, 3, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 140, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 174 + ], + "type": "text", + "content": "[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "type": "text", + "content": "[53] Xin Wen, Bingchen Zhao, Ismail Elezi, Jiankang Deng, and Xiaojuan Qi. \"principal components\" enable a new language of images, 2025. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 208, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 241 + ], + "type": "text", + "content": "[54] Jingfeng Yao, Bin Yang, and Xinggang Wang. Reconstruction vs. generation: Taming optimization dilemma in latent diffusion models. In CVPR, 2025. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 242, + 553, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 242, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 316, + 242, + 553, + 276 + ], + "type": "text", + "content": "[55] Ting Yao, Yingwei Pan, Yehao Li, Chong-Wah Ngo, and Tao Mei. Wave-vit: Unifying wavelet and transformers for visual representation learning. In ECCV, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 276, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 308 + ], + "type": "text", + "content": "[56] Hu Yu, Hao Luo, Hangjie Yuan, Yu Rong, and Feng Zhao. Frequency autoregressive image generation with continuous tokens, 2025. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 309, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 354 + ], + "type": "text", + "content": "[57] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved VQGAN. In ICLR, 2022. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 355, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 431 + ], + "type": "text", + "content": "[58] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research, 2022. Featured Certification. 1, 3, 5, 6, 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 433, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 498 + ], + "type": "text", + "content": "[59] Lijun Yu, Jose Lezama, Nitesh Bharadwaj Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G Hauptmann, Boqing Gong, Ming-Hsuan Yang, Irfan Essa, David A Ross, and Lu Jiang. Language model beats diffusion - tokenizer is key to visual generation. In ICLR, 2024. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 499, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 499, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 316, + 499, + 553, + 542 + ], + "type": "text", + "content": "[60] Qihang Yu, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An image is worth 32 tokens for reconstruction and generation. In NeurIPS, 2024. 1" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 544, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 578 + ], + "type": "text", + "content": "[61] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 578, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 553, + 599 + ], + "type": "text", + "content": "[62] Zhenhai Zhu and Radu Soricut. Wavelet-based image tokenizer for vision transformers, 2024. 2" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17190" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Spectral Sensitivity Estimation with an Uncalibrated Diffraction Grating/79b5704f-97ff-445c-93ff-ff5edaa3c620_content_list.json b/2025/Spectral Sensitivity Estimation with an Uncalibrated Diffraction Grating/79b5704f-97ff-445c-93ff-ff5edaa3c620_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..da41a98f110a795632a23fdf17f20648f05d3394 --- /dev/null +++ b/2025/Spectral Sensitivity Estimation with an Uncalibrated Diffraction Grating/79b5704f-97ff-445c-93ff-ff5edaa3c620_content_list.json @@ -0,0 +1,1748 @@ +[ + { + "type": "text", + "text": "Spectral Sensitivity Estimation with an Uncalibrated Diffraction Grating", + "text_level": 1, + "bbox": [ + 130, + 130, + 867, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lilika Makabe1 Hiroaki Santo1 Fumio Okura1 Michael S. Brown2 Yasuyuki Matsushita1,3 \n1The University of Osaka 2York University 3Microsoft Research Asia - Tokyo", + "bbox": [ + 106, + 179, + 888, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{makabe.lilika, santo.hiroaki, okura, yasumat}@ist.osaka-u.ac.jp mbrown@yorku.ca", + "bbox": [ + 148, + 218, + 843, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper introduces a practical and accurate calibration method for camera spectral sensitivity using a diffraction grating. Accurate calibration of camera spectral sensitivity is crucial for various computer vision tasks, including color correction, illumination estimation, and material analysis. Unlike existing approaches that require specialized narrow-band filters or reference targets with known spectral reflectances, our method only requires an uncalibrated diffraction grating sheet, readily available off-the-shelf. By capturing images of the direct illumination and its diffracted pattern through the grating sheet, our method estimates both the camera spectral sensitivity and the diffraction grating parameters in a closed-form manner. Experiments on synthetic and real-world data demonstrate that our method outperforms conventional reference target-based methods, underscoring its effectiveness and practicality.", + "bbox": [ + 89, + 300, + 485, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction and Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 571, + 375, + 588 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Estimating a camera's spectral sensitivity—its response to each wavelength of incident light—is essential for tasks such as color measurement, illumination and reflectance estimation, and material recognition. Spectral sensitivity is governed by the camera's optical system, including the sensor's color filter array and lens.", + "bbox": [ + 89, + 598, + 482, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Conventional methods [2, 4, 16, 22] for calibrating spectral sensitivity rely on specialized equipment to emit light with known spectra toward the camera and record the corresponding observed intensities. While accurate, these approaches are costly due to the need for precisely designed optical systems and extensive measurement time. Rather than directly observing light sources with known spectra, prior works [12, 21] have proposed using reference targets with known spectral reflectances under known spectral illuminations, such as sunlight or standard illuminants like D65. While these methods simplify the calibration process, a key challenge is the inherently low-frequency nature of spectral reflectances observed on natural objects [18] and high correlation between the spectral reflectances of indi", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/44fc7825dd7177e6b5c071c764881091c4d5e2f96eaae2ae37270c0f033ed055.jpg", + "image_caption": [ + "Figure 1. Overview of the proposed method. By capturing images of light with a known spectrum through an uncalibrated diffraction grating sheet, we observe both the direct and diffracted light. This practical and easy-to-use setup enables accurate estimation of camera spectral sensitivity." + ], + "image_footnote": [], + "bbox": [ + 517, + 268, + 906, + 435 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "vidual patches [5], which makes it difficult to accurately determine sensitivity at individual wavelengths.", + "bbox": [ + 511, + 535, + 905, + 565 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To enhance calibration accuracy, the use of diffraction grating sheets [13, 24] has been explored in the past. Diffraction occurs when light waves bend and spread as they pass through narrow apertures, with the degree of bending dependent on wavelength. This effect can be observed using an off-the-shelf diffraction grating sheet, available for less than 5 USD. Each diffracted light beam produces distinct spectral peaks, enabling precise estimation of spectral sensitivities at individual wavelengths. A key challenge in using a diffraction grating sheet is calibrating its diffraction parameter, the grating efficiency that describes the wavelength-dependent attenuation. To calibrate the efficiency, current methods require measurements of a reference target with known reflectances along with the diffracted light beam observation. This involves capturing multiple calibration scenes and complicates the process.", + "bbox": [ + 509, + 566, + 906, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For an accurate and easy-to-use calibration method, we propose using an uncalibrated diffraction grating sheet without needing additional captures of reference targets. As shown in Fig. 1, the proposed method observes light with known spectra through the diffraction grating sheet, recording both the direct light and the diffraction pattern.", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27252", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/af84459a4c8cc3381ae4c902f5d33b07805d1a19403cb2d54fa9937eb1d5a31e.jpg", + "image_caption": [ + "Figure 2. Illustration of a diffraction grating. When light of a certain wavelength passes through a diffraction grating, its output intensity is spatially distributed in a characteristic pattern with multiple sharp peaks, known as intensity principal maxima. The pixel positions of the intensity principal maxima shift outward from the center for longer wavelengths. When a broadband light source passes, light beams of different wavelengths appear at distinct shifted positions. In practice, we also need to consider nonuniform transmittance, known as grating efficiency. As a result, the observed intensity at a given pixel position is the product of the incoming light spectrum, the light dispersion function, and the grating efficiency. See Fig. 3 for examples of the observations." + ], + "image_footnote": [], + "bbox": [ + 96, + 78, + 480, + 334 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The proposed method simultaneously estimates the camera spectral sensitivity and grating efficiency. Although the original problem is a bilinear problem where the unknowns appear in their product form, we show that it can be turned into a tractable linear problem by introducing a basis representation for both camera spectral sensitivity and grating efficiency. We further introduce a closed-form solution method to the problem, given the pixel-to-wavelength mapping, i.e., the mapping between the pixel locations in the captured image and the wavelengths of the diffracted light.", + "bbox": [ + 89, + 529, + 483, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct both synthetic and real-world experiments to demonstrate the effectiveness of the proposed method. The results show that our method outperforms existing leading methods using calibrated reflectance targets [12].", + "bbox": [ + 89, + 681, + 483, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions Our chief contributions are as follows:", + "text_level": 1, + "bbox": [ + 89, + 758, + 460, + 773 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a camera spectral sensitivity estimation method that uses an uncalibrated diffraction grating sheet, without the need for capturing reference targets.", + "- The method uses an inexpensive diffraction grating sheet, making it practical, cost-effective, and easy to use.", + "- Accuracy of the proposed method significantly outperforms those that rely on reference targets, with a simpler setup and fewer captures." + ], + "bbox": [ + 91, + 777, + 482, + 898 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/318f9121b60f95ab9b44af1aaa6ccf9eb18a33397a84caf6945e082beef49e9b.jpg", + "image_caption": [ + "Figure 3. Capturing setup and the observed image of direct and diffracted lights. As the diffracted light exhibits multiple principal maxima, multiple diffraction patterns can be observed. The $k$ -th diffraction image, counted from the center $(k = \\pm 1, \\pm 2, \\ldots)$ , is referred to as the $k$ -th order diffraction image. The image intensity is adjusted for visualization." + ], + "image_footnote": [], + "bbox": [ + 516, + 90, + 697, + 181 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f7815dc79db77e4b882253d7d919f8530720a9f1c26e2fe53e76f460fa76a2ea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 90, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Preliminaries", + "text_level": 1, + "bbox": [ + 513, + 301, + 633, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we provide the foundational background on spectral image formation. We introduce the general spectral image formation model, followed by an overview of how diffraction gratings separate light into its constituent wavelengths. Finally, we present the image formation model for observations captured through a diffraction grating.", + "bbox": [ + 511, + 327, + 906, + 419 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Spectral image formation Given camera spectral sensitivity $s(\\lambda_i)$ at discretely sampled wavelengths $\\{\\lambda_i\\}$ and spectrum of the incoming light $e(\\lambda_i)$ to the camera, the intensity observation $m \\in \\mathbb{R}_+$ can be expressed as:", + "bbox": [ + 511, + 424, + 905, + 484 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nm \\propto \\sum_ {i} e (\\lambda_ {i}) s (\\lambda_ {i}). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 496, + 906, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffraction grating A diffraction grating is an optical device with multiple periodic grooves that separates incoming light into its constituent wavelengths, directing each along a distinct path. Due to interference from the multi-groove structure, each wavelength produces multiple principal intensity maxima, corresponding to different diffraction orders $k = \\pm 1, \\pm 2, \\ldots$ , as illustrated in Fig. 2. This results in multiple dispersed observations symmetrically aligned with the direction of incoming light, with the central observation representing the mixture of all constituent wavelengths—namely, the direct light observation.", + "bbox": [ + 511, + 544, + 906, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In general, diffraction observation $m_c^{(k)}(p)$ around the $k^{\\mathrm{th}}$ intensity principal maxima (i.o.w., $k^{\\mathrm{th}}$ order diffraction observation) at pixel location $p$ in channel $c\\in \\{R,G,B\\}$ can be written as", + "bbox": [ + 511, + 710, + 905, + 771 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nm _ {c} ^ {(k)} (p) \\propto \\sum_ {i} w (\\lambda_ {i}; p) e (\\lambda_ {i}) \\eta_ {k} (\\lambda_ {i}) s _ {c} (\\lambda_ {i}), \\qquad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 782, + 906, + 813 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $s_c$ is the camera spectral sensitivity for channel $c$ , and $\\eta_k(\\lambda)$ represents the grating efficiency for each wavelength $\\lambda$ at $k^{\\mathrm{th}}$ order diffraction. The term $w(\\lambda; p)$ is the weight induced by the light dispersion function, which disperses the light into each wavelength $\\lambda$ that is observed at a", + "bbox": [ + 511, + 824, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27253", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "corresponding peak pixel position $p$ depending on the wavelength. We refer to $w(\\lambda ;p)$ as a weight function.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The weight function $w(\\lambda; p)$ approximates a delta function with a sufficient number of slits in the grating, effectively acting as a one-to-one mapping between wavelength $\\lambda$ and pixel position $p$ in the diffraction image [13, 24]. Ideally, the grating efficiency $\\eta(\\lambda)$ should be all 1. However, in practice, the grating efficiency is known to vary depending on the frequency and the material of the grating [17].", + "bbox": [ + 89, + 121, + 483, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 239, + 232, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we provide an overview of existing methods for camera spectral sensitivity calibration. A straightforward approach to recovering spectral sensitivity is to measure the observed intensities $m$ in the image while varying the incoming light with known spectra $e(\\lambda_i)$ . To achieve such controlled illuminations, specialized light sources [9] and light sources with narrow-band filters [2, 4, 16, 22] have been used. However, the equipment is expensive, limiting their practical applications in real-world settings.", + "bbox": [ + 89, + 266, + 482, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To eliminate the need for specialized light sources, methods using reference targets, such as color checker charts [1, 7, 8, 11, 15, 19-21, 25], have been proposed. These methods recover camera spectral sensitivity by observing reference targets composed of multiple surface patches with known spectral reflectances under calibrated spectral illumination that uniformly illuminates the entire target.", + "bbox": [ + 89, + 402, + 483, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To reduce the cost of spectral illumination calibration, several methods [6, 14] assume specific illumination types, such as sunlight [14] or low-frequency spectra [10, 26], and simultaneously estimate camera spectral sensitivity and illumination spectra. While these methods can measure reflected light with different spectra from multiple patches in a single image, they rely on reference targets with known spectral reflectance. Unfortunately, off-the-shelf color checker charts do not provide such data, so one must either rely on externally measured data or perform their own calibration [27]. In addition, due to the low-frequency spectral reflectance of natural objects, the spectral resolution of these methods is limited. To address this challenge, methods using emissive charts [5] have been proposed; however, reference targets employing fluorescent materials are rare and have limited practical applicability.", + "bbox": [ + 89, + 508, + 482, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instead of observing reflectances, methods using a diffraction grating, which makes light beams with distinct spectral peaks at different wavelengths, have been proposed [13, 24]. Karge et al. [13] uses a fluorescent lamp, which has a spiky spectrum, to calibrate the pixel-to-wavelength mapping. They then replace the fluorescent lamp with a halogen lamp, which has a flatter spectrum, to observe the diffraction pattern. They additionally need to capture reference targets with known spectral reflectances under controlled spectral illumination to account for the at", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tenuation due to grating efficiency. Toivonen et al. [24] use multiple light sources with different spectra, generated by a halogen lamp equipped with calibrated color filters, to capture the multiple diffraction patterns. In addition, they also record images of a transmissive color chart, comprising multiple patches of calibrated color filters, through different light sources. By recording the diffraction patterns of different light sources and images of the transmissive color chart as input, they achieve accurate calibration of the spectral sensitivity. These methods use inexpensive setups compared to the specialized equipment and achieve more accurate calibration than those relying solely on a color checker chart. However, these approaches require multiple setups and complex capture procedures, which remain a significant practical challenge.", + "bbox": [ + 511, + 90, + 903, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike the above approaches, Solomatov and Akkaynak [23] use metadata recorded in the Exif, specifically the color matrices that transforms sensor responses to the CIE XYZ color space, to train a neural network for estimating spectral sensitivity. This approach does not require any additional captures for calibration; however, it cannot account for external factors, such as lens and camera filters. Moreover, their estimation exhibits white-balance ambiguity, and recovering the full spectral sensitivity requires white-balance calibration using a color checker chart.", + "bbox": [ + 511, + 316, + 903, + 468 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use an uncalibrated diffraction grating sheet, where the grating efficiency is unknown. Unlike previous diffraction grating-based methods, our approach only requires capturing a light source through the diffraction grating sheet, providing both the direct light and the diffraction pattern, enabling practical and accurate estimation. Notably, our minimal setting uses one light and two images under the light, and does not require any scene interaction after setup, whereas existing methods require at least two scene changes and three lights with known spectra [13].", + "bbox": [ + 511, + 469, + 905, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 513, + 633, + 687, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first outline the problem setting for determining the camera spectral sensitivity using a diffraction grating. We then present the proposed method.", + "bbox": [ + 511, + 657, + 905, + 704 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem setting Our image formation models of the direct light observation $m_{\\mathrm{dir}}$ and the first-order diffracted light observation $m_{\\mathrm{dif}}$ are written as:", + "bbox": [ + 511, + 707, + 905, + 753 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nm _ {\\mathrm {d i r}} \\propto E \\sum_ {i} e \\left(\\lambda_ {i}\\right) s \\left(\\lambda_ {i}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 763, + 903, + 803 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nm _ {\\mathrm {d i f}} (p) \\propto E \\sum_ {i} w (\\lambda_ {i}; p) e (\\lambda_ {i}) \\eta (\\lambda_ {i}) s (\\lambda_ {i}),\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 799, + 852, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $E$ is an unknown scale of the incoming light spectrum and $e(\\lambda)$ is a normalized spectrum of the incoming light. We omit the pixel position dependency for the direct light observation $m_{\\mathrm{dir}}$ as it remains constant along the", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27254", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fc315c169759536254fe61b419fc8c606d9e569985d78c36ab453ed2baae709d.jpg", + "image_caption": [ + "Figure 4. Illustration of the image formation models and overview of our estimation. The observation should vary only along the diffraction direction. To obtain the mean observation, we average along the vertical direction in the figure, resulting in the direct light observation in $\\mathbb{R}^3$ and the diffracted light observation in $\\mathbb{R}^{n\\times 3}$ in RGB channels, respectively. The sampled observation plot in the bottom figure shows a per-channel vector plot along the diffraction direction for each channel. This plot should have the same outline as the wavelength-wise multiplication of the grating efficiency, camera spectral sensitivity, and the incoming light spectrum, leaving the unknown mapping along the horizontal axis, referred to as the pixel-to-wavelength mapping. Using both the direct and diffracted light observation, given the pixel-to-wavelength mapping, we estimate the camera spectral sensitivity and the grating efficiency as described in Sec 3.1. Section 3.2 covers the estimation of the pixel-to-wavelength mapping." + ], + "image_footnote": [], + "bbox": [ + 94, + 78, + 903, + 255 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "diffraction (horizontal) direction, unlike the diffracted light. We do not use $k \\geq 2$ diffracted light observations in our method because, in practice, the intensity drastically decreases for higher-order diffraction.", + "bbox": [ + 89, + 406, + 483, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a known incoming light spectrum $e(\\lambda)$ and by observing both the direct and first-order diffracted lights, $m_{\\mathrm{dir}}$ and $m_{\\mathrm{dif}}(p)$ , our goal is to estimate the camera's spectral sensitivity $s(\\lambda)$ using an uncalibrated diffraction grating, where both the weight function $w(\\lambda; p)$ and the grating efficiency $\\eta(\\lambda)$ are unknown.", + "bbox": [ + 89, + 472, + 483, + 563 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the following, we first describe the method to jointly estimate the camera spectral sensitivity $s(\\lambda)$ and the grating efficiency $\\eta(\\lambda)$ , assuming that the weight function $w(\\lambda; p)$ is known beforehand. We then present the method to determine the weight function $w(\\lambda; p)$ .", + "bbox": [ + 89, + 569, + 483, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Estimating camera spectral sensitivity $s$ and grating efficiency $\\eta$", + "text_level": 1, + "bbox": [ + 89, + 670, + 483, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We now estimate the camera spectral sensitivity $s(\\lambda)$ and the grating efficiency $\\eta(\\lambda)$ , given the weight function $w(\\lambda; p)$ . We show, for the first time, that the problem has a closed-form solution when the unknowns are expressed as linear combinations of their bases. To this end, we first explain the constraints derived from the image formation models of the direct and diffracted light observations, and then present the solution method.", + "bbox": [ + 89, + 713, + 483, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let us assume camera spectral sensitivity $\\mathbf{s}$ and the inverse of grating efficiency $\\eta^{-1}$ at discretely sampled wavelengths $\\lambda_1, \\ldots, \\lambda_f$ can be represented by linear combinations of bases whose numbers of bases are $b_s$ and $b_\\eta$ , re", + "bbox": [ + 89, + 840, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "spectively. Namely,", + "bbox": [ + 513, + 406, + 648, + 421 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} \\triangleq \\left[ \\begin{array}{c c c} s (\\lambda_ {1}), & \\ldots , & s (\\lambda_ {f}) \\end{array} \\right] ^ {\\mathsf {T}} = \\mathbf {B} _ {s} \\mathbf {c} _ {s} \\in \\mathbb {R} _ {+} ^ {f},\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 431, + 859, + 452 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\eta} ^ {- 1} \\triangleq \\left[ \\begin{array}{l l l} \\frac {1}{\\eta \\left(\\lambda_ {1}\\right)}, & \\dots , & \\frac {1}{\\eta \\left(\\lambda_ {f}\\right)} \\end{array} \\right] ^ {\\intercal} = \\mathbf {B} _ {\\eta} \\mathbf {c} _ {\\eta} \\in \\mathbb {R} _ {+} ^ {f}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 450, + 903, + 483 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{B}_* \\in \\mathbb{R}^{f \\times b_*}$ is a basis matrix, and $\\mathbf{c}_* \\in \\mathbb{R}^{b_*}$ is a coefficient vector. We obtained the basis matrix $\\mathbf{B}_s$ of camera spectral sensitivity $s$ at a channel by singular value decomposition (SVD) on the same set of data used in [23]. For the inverse grating efficiency function $\\eta^{-1}$ , we use a Fourier basis representation. Further details can be found in the supplementary material.", + "bbox": [ + 511, + 494, + 906, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Linear constraint from direct light observation By introducing the basis representation, the aforementioned image formation model of direct light observations in Eq. (3) can be rewritten in a vector form as", + "bbox": [ + 511, + 604, + 905, + 664 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nm _ {\\mathrm {d i r}} \\propto E \\mathbf {e} ^ {\\mathsf {T}} \\mathbf {s} = E \\mathbf {e} ^ {\\mathsf {T}} \\mathbf {B} _ {s} \\mathbf {c} _ {s}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 675, + 903, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{e} \\in \\mathbb{R}_+^f$ is the normalized incoming light spectrum. The light scale $E$ can be set to 1, and the proportionality can be turned into equality without loss of generality, as the global scale can be accounted for in camera spectral sensitivity s. As a result, we have", + "bbox": [ + 511, + 707, + 905, + 782 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nm _ {\\mathrm {d i r}} = \\underbrace {\\mathbf {e} ^ {\\mathsf {T}} \\mathbf {B} _ {s}} _ {\\mathbf {A} _ {\\mathrm {d i r}}} \\mathbf {c} _ {s}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 650, + 792, + 903, + 828 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{A}_{\\mathrm{dir}} \\in \\mathbb{R}^{1 \\times b_s}$ is a known row vector. Similar equations can be derived for all RGB channels to form a vector of direct light observations $\\mathbf{m}_{\\mathrm{dir}} = [m_{\\mathrm{dir}(R)}, m_{\\mathrm{dir}(G)}, m_{\\mathrm{dir}(B)}]^{\\top}$ on the left-hand side", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27255", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of the equation, resulting in 3 linear equations with unknown vector $[\\mathbf{c}_{s(R)}^{\\intercal},\\mathbf{c}_{s(G)}^{\\intercal},\\mathbf{c}_{s(B)}^{\\intercal}]^{\\intercal}$ , for which $\\mathbf{A}_{\\mathrm{dir}}$ becomes a $3\\times 3b_{s}$ matrix.", + "bbox": [ + 89, + 90, + 482, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Linear constraint from diffracted light observation Like the direct light observation case, the diffracted light's image formation model in Eq. (3) can be rewritten in a matrix form as", + "bbox": [ + 89, + 138, + 483, + 199 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {m} _ {\\text {d i f}} = \\left[ m _ {\\text {d i f}} \\left(p _ {1}\\right), \\dots , m _ {\\text {d i f}} \\left(p _ {n}\\right) \\right] ^ {\\top} \\tag {7} \\\\ = \\mathbf {W} \\operatorname {d i a g} (\\mathbf {e}) \\operatorname {d i a g} (\\boldsymbol {\\eta}) \\mathbf {s}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 204, + 480, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{m}_{\\mathrm{dif}} \\in \\mathbb{R}_+^n$ is a stack of the diffracted light observations at pixels $p_1, \\ldots, p_n$ . $\\mathbf{W} \\in \\mathbb{R}_+^{n \\times f}$ denotes the known weight function for all $n$ pixels, written in a matrix form as", + "bbox": [ + 89, + 252, + 483, + 300 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {W} = \\left[ \\begin{array}{c c c} w \\left(\\lambda_ {1}, p _ {1}\\right) & \\dots & w \\left(\\lambda_ {f}, p _ {1}\\right) \\\\ \\vdots & & \\\\ w \\left(\\lambda_ {1}, p _ {n}\\right) & \\dots & w \\left(\\lambda_ {f}, p _ {n}\\right) \\end{array} \\right] \\in \\mathbb {R} _ {+} ^ {n \\times f}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 319, + 482, + 377 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which we refer to as the weight matrix.", + "bbox": [ + 89, + 383, + 349, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The above image formation model of diffracted light can be equivalently turned into", + "bbox": [ + 89, + 398, + 482, + 428 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} = \\operatorname {d i a g} \\left(\\boldsymbol {\\eta} ^ {- 1}\\right) \\underbrace {\\operatorname {d i a g} \\left(\\mathbf {e} ^ {- 1}\\right) \\mathbf {W} ^ {\\dagger} \\mathbf {m} _ {\\mathrm {d i f}}} _ {\\mathbf {a}} = \\operatorname {d i a g} (\\mathbf {a}) \\boldsymbol {\\eta} ^ {- 1}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 434, + 482, + 483 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the inverse of a vector denotes an element-wise inverse operator, $\\mathbf{W}^{\\dagger}$ is a left pseudo-inverse of $\\mathbf{W}$ , and $\\mathbf{a} \\in \\mathbb{R}^{f}$ is a known channel-and-light-dependent vector. Since we can confidently assume $n > f$ in our setting, the weight matrix $\\mathbf{W}$ should have full column rank when the diffracted observation spans the entire visible wavelength domain, and thus it has a left pseudo-inverse.", + "bbox": [ + 89, + 484, + 483, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "From Eq. (4), Eq. (9) can be written as", + "bbox": [ + 109, + 590, + 370, + 604 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {B} _ {s} \\mathbf {c} _ {s} = \\operatorname {d i a g} (\\mathbf {a}) \\mathbf {B} _ {\\eta} \\mathbf {c} _ {\\eta}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 613, + 480, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Putting together the unknowns $\\mathbf{c}_{\\eta}$ and $\\mathbf{c}_s$ , we have the following homogeneous system of equations:", + "bbox": [ + 89, + 638, + 483, + 667 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\underbrace {\\left[ \\begin{array}{l l} \\operatorname {d i a g} (\\mathbf {a}) \\mathbf {B} _ {\\eta} & - \\mathbf {B} _ {s} \\end{array} \\right]} _ {\\mathbf {A} _ {\\mathrm {d i f}}} \\left[ \\begin{array}{l} \\mathbf {c} _ {\\eta} \\\\ \\mathbf {c} _ {s} \\end{array} \\right] = \\mathbf {0}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 675, + 482, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{A}_{\\mathrm{dif}} \\in \\mathbb{R}^{f \\times (b_{\\eta} + b_{s})}$ is a known matrix. For the case of RGB images, stacking similar equations from all RGB channels, we obtain $3f$ linear equations from the diffracted light observations for unknown vector $[\\mathbf{c}_{\\eta}^{\\top}, \\mathbf{c}_{s(R)}^{\\top}, \\mathbf{c}_{s(G)}^{\\top}, \\mathbf{c}_{s(B)}^{\\top}]^{\\top} \\in \\mathbb{R}^{b_{\\eta} + 3b_{s}}$ , and $\\mathbf{A}_{\\mathrm{dif}}$ becomes a $3f \\times (b_{\\eta} + 3b_{s})$ matrix.", + "bbox": [ + 89, + 728, + 483, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Solution method We can estimate the camera spectral sensitivity $s$ and grating efficiency $\\eta$ by solving the following optimization problem using RGB channels:", + "bbox": [ + 89, + 825, + 483, + 869 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} ^ {*} = \\underset {\\mathbf {x}} {\\operatorname {a r g m i n}} \\| \\mathbf {A} _ {\\mathrm {d i f}} \\mathbf {x} \\| _ {2} ^ {2} \\text {s . t .} \\left[ \\begin{array}{l l} \\mathbf {0} & \\mathbf {A} _ {\\mathrm {d i r}} \\end{array} \\right] \\mathbf {x} = \\mathbf {m} _ {\\mathrm {d i r}}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 877, + 483, + 904 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{x} = \\left[\\mathbf{c}_{\\eta}^{\\mathsf{T}},\\mathbf{c}_{s(R)}^{\\mathsf{T}},\\mathbf{c}_{s(G)}^{\\mathsf{T}},\\mathbf{c}_{s(B)}^{\\mathsf{T}}\\right]^{\\mathsf{T}}\\in \\mathbb{R}^{b_{\\eta} + 3b_s}$ is the vector of unknowns. The above optimization problem has a closed-form solution by introducing a Lagrange multiplier when the number bases $b_{*}$ and the wavelength resolution $f$ satisfy $3f\\geq b_{\\eta} + 3b_{s}$ , and $\\mathbf{A}_{\\mathrm{dif}}$ and $\\mathbf{A}_{\\mathrm{dir}}$ are full-rank. The details of the solution method are provided in the supplementary material.", + "bbox": [ + 511, + 88, + 906, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Estimation of weight function $w$", + "text_level": 1, + "bbox": [ + 511, + 215, + 794, + 232 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As described in Sec. 1, the weight function $w(\\lambda; p)$ has a peaky shape using diffraction gratings; therefore, we assume there is a one-to-one mapping between the pixel location $p$ in the diffraction image and the wavelength $\\lambda$ as in previous methods [13, 24]. To estimate the weight function $w(\\lambda; p)$ , it is necessary to establish the pixel-to-wavelength mapping, namely, relating the pixel locations of the diffraction image to the wavelength as illustrated in Fig. 6.", + "bbox": [ + 511, + 238, + 906, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose to use a quadratic function to represent the one-to-one pixel-to-wavelength mapping $g: p \\to \\lambda$ as", + "bbox": [ + 511, + 359, + 905, + 391 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda = g (p) = a p ^ {2} + b p + c, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 402, + 903, + 420 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $a, b, c \\in \\mathbb{R}$ are the parameters. This quadratic representation allows us stably determining the mapping with retaining accuracy. For further discussion about this parameterization, please refer to our supplementary.", + "bbox": [ + 511, + 431, + 905, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Once the pixel-to-wavelength mapping $g(p)$ is determined, the elements of the weight matrix $\\mathbf{W}$ is simply obtained by linear interpolation as", + "bbox": [ + 511, + 494, + 905, + 540 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw (\\lambda , p) = \\left\\{ \\begin{array}{l l} \\frac {\\lambda_ {i + 1} - g (p)}{\\lambda_ {i + 1} - \\lambda_ {i}} & \\text {i f} \\lambda = \\lambda_ {i} \\\\ \\frac {g (p) - \\lambda_ {i}}{\\lambda_ {i + 1} - \\lambda_ {i}} & \\text {i f} \\lambda = \\lambda_ {i + 1}, \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 551, + 905, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{i}$ is the nearest-neighbor wavelength to $g(p)$ that satisfies $\\lambda_{i} \\leq g(p) \\leq \\lambda_{i+1}$ . Therefore, the problem of estimating the weight function boils down to the estimation of pixel-to-wavelength mapping. The challenge, though, is that we only have access to the diffraction image and illumination spectra, but not to camera spectral sensitivity and grating efficiency. Now we describe two settings for determining the pixel-to-wavelength mapping; one using fluorescent and LED lights (denoted as Fluorescent+LED), and another using only LED light (denoted as LED only).", + "bbox": [ + 511, + 651, + 906, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fluorescent+LED case When light with spiky spectral peaks, such as fluorescent light, passes through the diffraction grating, the resulting diffracted observation exhibits a similar spiky pattern, as shown in Fig. 5. Therefore, the fluorescent light provides a strong cue for determining the pixel-to-wavelength mapping. In this case, we use a similar", + "bbox": [ + 511, + 809, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27256", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/95a9c140f1146aa67cb9110d4a2d3be9199b3d69cf7a88772ca29ea55de53de3.jpg", + "image_caption": [ + "Figure 5. The illuminant spectrum of the light source used in our real-world experiment and an example of the corresponding direct and diffracted light observations. The observations shown here are adjusted for visualization." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 478, + 186 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "approach to [13] that associates corresponding peak positions in the pixel and wavelength domains. Since the spectral sensitivity $\\mathbf{s}$ is unknown, we weigh the illumination spectra $\\mathbf{e}$ by the mean spectral sensitivity $\\bar{\\mathbf{s}}$ as $\\bar{\\mathbf{s}}\\odot \\mathbf{e}$ , where $\\bar{\\mathbf{s}}$ is obtained from the same dataset used for the camera spectral sensitivity basis [23], and $\\odot$ represents an elementwise multiplication operator. We have observed that the unknown grating efficiency $\\eta$ is negligible in determining the peak positions due to its low-frequency distribution.", + "bbox": [ + 88, + 266, + 483, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We select the top two peaks from each of the pixel and wavelength domains for each channel to determine the polynomial parameters $a$ , $b$ , and $c$ of the pixel-to-wavelength mapping $g$ via linear least squares. Once the pixel-to-wavelength mapping $g$ and weight matrix $\\mathbf{W}$ are obtained, we take another capture using an LED light for obtaining the direct and diffracted measurements used in Sec. 3.1.", + "bbox": [ + 89, + 402, + 483, + 507 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LED only case To simplify the recording process, it is preferable if we can work only with an LED. In this case, due to the low-frequency incoming spectra without obvious peaks, we cannot take the strategy of relating peak positions used in the Fluorescent+LED case.", + "bbox": [ + 89, + 511, + 483, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Therefore, instead of finding peak positions, we frame the problem as finding the optimal horizontal mapping (pixel-to-wavelength) between two plots—the normalized diffraction observations $\\mathbf{u}$ and the normalized mean camera spectral sensitivity $\\mathbf{v}$ that is scaled by illumination spectrum, expressed as:", + "bbox": [ + 89, + 587, + 483, + 678 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} \\triangleq \\nu (\\mathbf {e} \\odot \\bar {\\mathbf {s}}) = [ v (\\lambda_ {1}), \\dots , v (\\lambda_ {f}) ] \\in \\mathbb {R} _ {+} ^ {f}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 686, + 449, + 705 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {u} \\triangleq \\nu (\\mathbf {m} _ {\\mathrm {d i f}}) = [ u (p _ {1}), \\dots , u (p _ {n}) ] \\in \\mathbb {R} _ {+} ^ {n},\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 708, + 454, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\nu$ is a vector normalization as $\\nu (\\mathbf{x}) = \\mathbf{x} / \\| \\mathbf{x}\\| _2$ .", + "bbox": [ + 89, + 734, + 449, + 751 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We solve the point-to-plane iterative closest point (ICP) problem [3] to determine the optimal polynomial parameters $a$ , $b$ , and $c$ along the horizontal axis by minimizing the following objective function:", + "bbox": [ + 89, + 751, + 483, + 811 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {I C P}} = \\sum_ {i = 1} ^ {n} \\sum_ {j \\in \\mathcal {N} (i)} \\left(\\left(\\left[ \\begin{array}{c} a p _ {i} ^ {2} + b p _ {i} + c \\\\ u (p _ {i}) \\end{array} \\right] - \\left[ \\begin{array}{c} \\lambda_ {j} \\\\ v (\\lambda_ {j}) \\end{array} \\right]\\right) ^ {\\top} \\mathbf {n _ {v}} (j)\\right) ^ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 818, + 488, + 859 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{N}(i)$ is the set of indices for the current neighboring points of $[ap_i^2 +bp_i + c,u(p_i)]^\\top$ , and $\\mathbf{n_v}(j)\\in \\mathbb{S}\\subset \\mathbb{R}^2$", + "bbox": [ + 89, + 867, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/83b58e14ecc95c28607285e0bec20125b826946c5fe94919d2acb16d329f433f.jpg", + "image_caption": [ + "Figure 6. Illustration of our pixel-to-wavelength mapping estimation. Given the normalized diffraction observations $\\mathbf{u}_c$ and normalized illumination-scaled mean camera spectral sensitivity $\\mathbf{v}_c$ for all $c \\in \\{R, G, B\\}$ channels, we estimate the optimal pixel-to-wavelength mapping (or horizontal warping, as shown in the figure), which minimizes the distance between the two plots." + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 903, + 280 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "is a unit normal vector to $[\\lambda_j,v(\\lambda_j)]^\\top$ which satisfies", + "bbox": [ + 511, + 386, + 864, + 406 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\begin{array}{c} \\lambda_ {j} - \\lambda_ {j + 1} \\\\ v (\\lambda_ {j}) - v (\\lambda_ {j + 1}) \\end{array} \\right] ^ {\\mathsf {T}} \\mathbf {n} _ {\\mathrm {v}} (j) = 0. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 415, + 903, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We iteratively minimize the objective function for 500 iterations, with the initial guesses set to $a = 0$ , $b = f / n$ , $c = 0$ , where $b$ is given as the ratio of the numbers of elements of $\\mathbf{u}$ and $\\mathbf{v}$ , to determine the pixel-to-wavelength mapping $g$ . This setting allows for a convenient one-time capture for both determining the pixel-to-wavelength mapping and obtaining the direct and diffracted measurements used in Sec. 3.1.", + "bbox": [ + 511, + 462, + 906, + 582 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 597, + 645, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we evaluate our method on both synthetic and real-world scenes. We first describe our experimental setup and then present the evaluation results.", + "bbox": [ + 511, + 623, + 905, + 669 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines We compare the proposed approach with the Exif-based method (Exif+CC) [23] and the color chart-based method (CC) [12]. We present two variants of the proposed method as detailed in Sec. 3.2: Ours (LED+Flu), which employs a fluorescent lamp observation for pixel-to-wavelength mapping estimation, and Ours (LED), which uses only an LED light observation. For the color chart-based method, we use 2 basis functions, as mentioned in the original paper, which yielded the best results in our preliminary experiments. Since the Exif-based method cannot estimate white balance and introduces a $3 \\times 3$ ambiguity, we use the color checker observation to resolve this ambiguity.", + "bbox": [ + 511, + 674, + 906, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also compared ours with the existing diffraction grating-based approach by Toivonen et al. [24], adjusted to our experimental setup. However, their method produced", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27257", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "unstable estimations due to the limited observations available in our setup compared to their original conditions. Further details are provided in the supplemental material.", + "bbox": [ + 89, + 90, + 480, + 137 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Basis representations To compute the basis for camera spectral sensitivity $s(\\lambda)$ , we use the dataset from the Exif-based method [23], which includes 44 cameras for training and 5 for testing. Following the detailed analysis on basis selection for a camera spectral sensitivity in [12], we apply SVD to the sensitivities in the training dataset in a per-channel manner. For the grating efficiency, we use a Fourier basis. Further discussion on the grating efficiency basis can be found in the supplementary material. The spectral resolution $f$ in the dataset of camera spectral sensitivities is $f = 31$ from $400\\mathrm{nm}$ to $700\\mathrm{nm}$ . Assuming the number of basis for the camera spectral sensitivity and the grating efficiency to be the same $b_{\\eta} = b_{s}$ , the required condition for the number of bases for our method is $b_{\\eta} = b_{s} \\leq 23$ based on the discussion in Sec. 3.1. For robust estimation, the number of basis functions is set to 7 for each channel of the camera spectral sensitivity and the grating efficiency across all experiments.", + "bbox": [ + 89, + 141, + 482, + 414 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation metrics Following [23], we use the average relative full-scale error across the three color channels, RE, which is defined as:", + "bbox": [ + 89, + 419, + 482, + 464 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {R E} = \\frac {1}{3} \\sum_ {c} \\frac {\\operatorname {R M S E} (c)}{\\max \\left(\\mathbf {s} _ {c}\\right)}, \\operatorname {R M S E} (c) = \\sqrt {\\frac {\\left\\| \\mathbf {s} _ {c} ^ {*} - \\mathbf {s} _ {c} \\right\\| _ {2} ^ {2}}{f}}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 474, + 482, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{s}_c^*$ and $\\mathbf{s}_c$ denote the estimated and ground-truth camera spectral sensitivity for each channel $c\\in \\{R,G,B\\}$ .", + "bbox": [ + 89, + 523, + 482, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Evaluation with synthetic scenes", + "text_level": 1, + "bbox": [ + 89, + 565, + 375, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we quantitatively evaluate the comparison methods and ours on a synthetic dataset.", + "bbox": [ + 89, + 587, + 482, + 617 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dataset We render the direct and diffracted images using Eq. (2), with parameters as follows. We use five camera spectral sensitivities: EOS 600D, iPhone12ProMax, Olympus EPL2, Pentax K5, and Sumsung GalaxyS20, all from the same test dataset as the Exif-based method. The D65 illuminant, as defined by the CIE Standard Illuminants, is used as the incoming light. To synthesize the pixel-to-wavelength mapping, we randomly define the mapping while ensuring that the visible wavelength range is fully covered in the diffraction image. The grating efficiency function is set to a low-frequency response as illustrated in Fig. 4, parameterized by 7 Fourier bases. Since the Exif-based method relies on metadata rather than images, we evaluate their method using real-world data.", + "bbox": [ + 89, + 623, + 482, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results Figure 7(a) shows the estimated results of the proposed and comparison methods for our synthetic scenes. For ideal observations, both Ours (LED+Flu) and the color chart-based method, CC, yield low estimation errors, as", + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "both methods are guaranteed to produce optimal results under the basis representation assumption. The remaining errors stem from the gap between the assumption and the synthetic data. The Exif-based method, Exif+CC, produces a generally reasonable overall shape, but it fails to accurately capture finer details. Ours (LED) yields a larger error than Ours (LED+Flu) due to estimation errors in the pixel-to-wavelength mapping when using only LED observations; however, overall, we can still obtain a reasonable estimation even by capturing just a single calibration scene.", + "bbox": [ + 511, + 90, + 903, + 242 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Evaluation with real-world scenes", + "text_level": 1, + "bbox": [ + 511, + 252, + 808, + 267 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here, we present evaluations using our real-world dataset.", + "bbox": [ + 511, + 275, + 895, + 290 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dataset 3 illustrates our experimental setup, which consists of a light source, a diffraction grating sheet, and a camera. We capture the images in a dark room for both our method and the comparison methods, ensuring that the only light in the scene comes from illumination with known spectra (LED or fluorescent). We use a 3D-printed slit for the incoming light, positioned in front of the light source, and place a diffraction grating in front of the camera, manually aligned roughly parallel to the image plane. For the spiky light source, we use a Pa-Look Ball (EFD15EL)1 fluorescent lamp. We use the following four cameras: EOS RP, iPhone 15ProMax, Sony $\\alpha$ 1, and DJI Pocket3. For the non-spiky light sources and diffraction sheets, we use three different products for each. In this section, we present the results for one setup. Additional experiments with other setups are provided in the supplementary material.", + "bbox": [ + 511, + 295, + 905, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Due to the high dynamic range of the direct and diffracted light, we capture HDR images by taking two images with different exposure times for each. The extra diffracted image under fluorescent light is captured for LED+Flu, solely for pixel-to-wavelength mapping. After capturing the direct and diffracted light observations, we roughly crop them with a sufficient margin at both ends of the diffraction pattern along the diffraction direction. The diffracted light observation is cropped to approximately $200 \\times 1000$ pix and aggregated to $1 \\times 1000$ pix.", + "bbox": [ + 511, + 551, + 905, + 703 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the comparison methods, we use a 24-patch color checker chart with spectral reflectance data from Chromaxion. The color checker chart is captured under the same non-spiky light source used in our experiments. We carefully position the color checker chart so that it is parallel to the camera and uniformly illuminated by the light source, ensuring that no shadows are cast. The metadata for the Exif-based method is extracted from our captured data following the procedure described in their paper [23].", + "bbox": [ + 511, + 704, + 905, + 840 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "1 Panasonic Corporation", + "bbox": [ + 532, + 851, + 661, + 862 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "2ColorChecker Classic, Calibrite", + "bbox": [ + 532, + 863, + 707, + 875 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "3Chromaxion, https://www.chromaxion.com, last accessed on March 7, 2025.", + "bbox": [ + 514, + 875, + 903, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27258", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/248692476abfeef63fd2fccf00e5837067470deea016905411937e1014429e3c.jpg", + "image_caption": [ + "(a) Synthetic experiment" + ], + "image_footnote": [], + "bbox": [ + 86, + 104, + 509, + 479 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/af89459e082555441ea3e987d5cbb408993d84a1567429bdae276bce4bd19096.jpg", + "image_caption": [ + "(b) Real-world experiment", + "Figure 7. Results of the proposed and comparison methods for our (a) synthetic and (b) real-world datasets. For each setting and method, we show plots of the estimated (solid line) and ground-truth (dotted line) spectral sensitivity for each color channel. All plots are normalized by the maximum response across all channels. The values below the plots indicate the RE defined in Eq. (16). For simplicity, the factor of $10^{-2}$ is omitted. Underscored bold and underscored score are the best and second-best result for each setting, respectively." + ], + "image_footnote": [], + "bbox": [ + 522, + 104, + 916, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The ground-truth spectral sensitivities are obtained using a calibration device with a lens attached. The illuminant spectra of the light sources, shown in Fig. 5, are measured through the diffraction grating using a spectrometer.", + "bbox": [ + 88, + 587, + 482, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results Figure 7(b) presents the estimated results of the proposed and comparison methods. Overall, the proposed methods achieve the lowest estimation errors for most scenes, and Ours (LED) produces results nearly comparable to Ours (LED+Flu). For the EOS RP scene, Ours (LED) yields a larger error due to the failure of pixel-to-wavelength mapping estimation using a non-spiky light source. The color checker-based method, CC, is relatively sensitive to real-world noise due to the low dimensionality of the spectral reflectances in the color patches, as discussed in previous works [5, 24]. While the Exif-based method, Exif+CC, has an advantage of being easy to use, as it only relies on metadata for recovering sensitivity, it yields larger estimation errors across all scenes.", + "bbox": [ + 88, + 655, + 482, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 585, + 638, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a method for estimating camera spectral sensitivity using an uncalibrated diffraction grating sheet. Compared to previous methods that use reference targets, such as a color checker chart, our approach achieves highly accurate estimation by leveraging the diffraction grating. Unlike previous diffraction grating-based methods, the proposed method jointly estimates both camera spectral sensitivity and grating efficiency without needing additional recordings, making it practical, cost-efficient, and easy-to-use.", + "bbox": [ + 509, + 612, + 906, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "One of the limitations of our method is that pixel-to-wavelength mapping estimation breaks down if the illumination has uniform spectra. In this case, the diffraction image simply shows the camera spectral sensitivity scaled by unknown grating efficiency, providing no clues for the mapping estimation. However, in practice, readily-available light sources do not have such uniform spectra. LEDs, for example, tend to have a non-uniform spectral distribution, enabling us to estimate the mapping and camera spectral sensitivities, as demonstrated in our real-world experiments.", + "bbox": [ + 511, + 750, + 908, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "4camSPECS XL, Image Engineering GmbH & Co. KG", + "bbox": [ + 107, + 875, + 403, + 886 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "$^{5}$ CS-2000S, KONICA MINOLTA, INC.", + "bbox": [ + 107, + 886, + 321, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27259", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by JSPS KAKENHI Grant Numbers JP23H05491 and JP25K00142, JST ASPIRE Grant Number JPMJAP2304, the Canada First Research Excellence Fund program (VISTA), an NSERC Discovery Grant, and the Canada Research Chair program.", + "bbox": [ + 89, + 113, + 483, + 200 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 210, + 187, + 227 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ali Alsam and Graham Finlayson. Recovering spectral sensitivities with uncertainty. In Proceedings of Conference on Colour in Graphics, Imaging, and Vision (CGIV), pages 22-26, 2002. 3", + "[2] E Berra, S Gibson-Poole, A MacArthur, R Gaulton, and A Hamilton. Estimation of the spectral sensitivity functions of un-modified and modified commercial off-the-shelf digital cameras to enable their use as a multispectral imaging system for UAVs. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 40: 207-214, 2015. 1, 3", + "[3] Yang Chen and Gérard Medioni. Object modelling by registration of multiple range images. Image and Vision Computing, 10(3):145-155, 1992. 6", + "[4] Maryam Mohammadzadeh Darrodi, Graham Finlayson, Teresa Goodman, and Michal Mackiewicz. Reference data set for camera spectral sensitivity estimation. Journal of the Optical Society of America A, 32(3):381-391, 2015. 1, 3", + "[5] Jeffrey M DiCarlo, Glen Eric Montgomery, and Steven W Trovinger. Emissive chart for imager calibration. In Proceedings of IS&T Color and Imaging Conference (CIC), pages 295-301, 2004. 1, 3, 8", + "[6] Bob Dyas. Robust color sensor response characterization. In Proceedings of IS&T Color and Imaging Conference (CIC), pages 144-148, 2000. 3", + "[7] Marc Ebner. Estimating the spectral sensitivity of a digital sensor using calibration targets. In Proceedings of Annual Conference on Genetic and Evolutionary Computation, pages 642-649, 2007. 3", + "[8] Graham Finlayson, Maryam Mohammadzadeh Darrodi, and Michal Mackiewicz. Rank-based camera spectral sensitivity estimation. 33(4):589-599, 2016. 3", + "[9] Paul M Hubel, Doron Sherman, and Joyce E Farrell. A comparison of methods of sensor spectral sensitivity estimation. In Proceedings of IS&T Color and Imaging Conference (CIC), pages 45-48, 1994. 3", + "[10] Cong Phuoc Huynh and Antonio Robles-Kelly. Recovery of spectral sensitivity functions from a colour chart image under unknown spectrally smooth illumination. In Proceedings of IEEE International Conference on Pattern Recognition (ICPR), pages 708-713, 2014. 3", + "[11] Yuhyun Ji, Yunsang Kwak, Sang Mok Park, and Young L Kim. Compressive recovery of smartphone RGB spectral sensitivity functions. Optics Express, 29(8):11947-11961, 2021. 3", + "[12] Jun Jiang, Dengyu Liu, Jinwei Gu, and Sabine Süssstrunk. What is the space of spectral sensitivity functions for digital" + ], + "bbox": [ + 93, + 236, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "color cameras? In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 168-179, 2013. 1, 2, 6, 7, 8", + "[13] Andreas Karge, Ingmar Rieger, Bernhard Eberhardt, and Andreas Schilling. Using chromaticity error minimisation for fast camera spectral responsivity measurement. In Proceedings of IS&T Color and Imaging Conference (CIC), pages 67-74, 2018. 1, 3, 5, 6", + "[14] Rei Kawakami, Hongxun Zhao, Robby T Tan, and Katsushi Ikeuchi. Camera spectral sensitivity and white balance estimation from sky images. International Journal of Computer Vision (IJCV), 105:187-204, 2013. 3", + "[15] Berk Kaya, Yigit Baran Can, and Radu Timofte. Towards spectral estimation from a single RGB image in the wild. In Proceedings of IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 3546-3555, 2019. 3", + "[16] Lindsay W MacDonald. Determining camera spectral responsivity with multispectral transmission filters. In Proceedings of IS&T Color and Imaging Conference (CIC), pages 12-17, 2015. 1, 3", + "[17] Christopher Palmer and Erwin G Loewen. Diffraction Grating Handbook. Newport Corporation New York, 2005. 3", + "[18] J. P. S. Parkkinen, J. Hallikainen, and T. Jaaskelainen. Characteristic spectra of munsell colors. Journal of the Optical Society of America A, 6(2):318-322, 1989. 1", + "[19] Dilip Prasad, Rang Nguyen, and Michael Brown. Quick approximation of camera's spectral response from casual lighting. In Proceedings of IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 844-851, 2013. 3", + "[20] Shuxue Quan, Noboru Ohta, and Xiaoyun Jiang. Comparative study on sensor spectral sensitivity estimation. In Proceedings of SPIE Color Imaging VIII: Processing, Hardcopy, and Applications, pages 209-220, 2003.", + "[21] Martin Rump, Arno Zinke, and Reinhard Klein. Practical spectral characterization of trichromatic cameras. ACM Transactions on Graphics (TOG), 30(6):170, 2011. 1, 3", + "[22] Yuri Rzhanov. Accurate determination of camera quantum efficiency from a single image. Journal of Imaging, 10(7): 169, 2024. 1, 3", + "[23] Grigory Solomatov and Derya Akkaynak. Spectral sensitivity estimation without a camera. In International Conference on Computational Photography (ICCP), pages 1-12, 2023. 3, 4, 6, 7, 8", + "[24] Mikko E Toivonen and Arto Klami. Practical camera sensor spectral response and uncertainty estimation. Journal of Imaging, 6(8):79, 2020. 1, 3, 5, 6, 8", + "[25] Shoji Tominaga, Shogo Nishi, and Ryo Ohtera. Measurement and estimation of spectral sensitivity functions for mobile phone cameras. Sensors, 21(15):4985, 2021. 3", + "[26] Lin Zhang, Ying Fu, Yinqiang Zheng, and Hua Huang. Camera spectral sensitivity, illumination and spectral reflectance estimation for a hybrid hyperspectral image capture system. In Proceedings of IEEE International Conference on Image Processing (ICIP), pages 545-549, 2017. 3" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27260", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "ref_text", + "text": "[27] Jieyu Zhu, Xufen Xie, Ningfang Liao, Zhuangzhi Zhang, Wenmin Wu, and Lvming Lv. Spectral sensitivity estimation of trichromatic camera based on orthogonal test and window filtering. Optics Express, 28(19):28085-28100, 2020. 3", + "bbox": [ + 91, + 90, + 482, + 147 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27261", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Spectral Sensitivity Estimation with an Uncalibrated Diffraction Grating/79b5704f-97ff-445c-93ff-ff5edaa3c620_model.json b/2025/Spectral Sensitivity Estimation with an Uncalibrated Diffraction Grating/79b5704f-97ff-445c-93ff-ff5edaa3c620_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f3a70bffdb8baf524fd39454dff5c58c1bdf962b --- /dev/null +++ b/2025/Spectral Sensitivity Estimation with an Uncalibrated Diffraction Grating/79b5704f-97ff-445c-93ff-ff5edaa3c620_model.json @@ -0,0 +1,2112 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.131, + 0.131, + 0.868, + 0.154 + ], + "angle": 0, + "content": "Spectral Sensitivity Estimation with an Uncalibrated Diffraction Grating" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.18, + 0.89, + 0.217 + ], + "angle": 0, + "content": "Lilika Makabe1 Hiroaki Santo1 Fumio Okura1 Michael S. Brown2 Yasuyuki Matsushita1,3 \n1The University of Osaka 2York University 3Microsoft Research Asia - Tokyo" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.219, + 0.844, + 0.235 + ], + "angle": 0, + "content": "{makabe.lilika, santo.hiroaki, okura, yasumat}@ist.osaka-u.ac.jp mbrown@yorku.ca" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.486, + 0.543 + ], + "angle": 0, + "content": "This paper introduces a practical and accurate calibration method for camera spectral sensitivity using a diffraction grating. Accurate calibration of camera spectral sensitivity is crucial for various computer vision tasks, including color correction, illumination estimation, and material analysis. Unlike existing approaches that require specialized narrow-band filters or reference targets with known spectral reflectances, our method only requires an uncalibrated diffraction grating sheet, readily available off-the-shelf. By capturing images of the direct illumination and its diffracted pattern through the grating sheet, our method estimates both the camera spectral sensitivity and the diffraction grating parameters in a closed-form manner. Experiments on synthetic and real-world data demonstrate that our method outperforms conventional reference target-based methods, underscoring its effectiveness and practicality." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.573, + 0.377, + 0.589 + ], + "angle": 0, + "content": "1. Introduction and Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.688 + ], + "angle": 0, + "content": "Estimating a camera's spectral sensitivity—its response to each wavelength of incident light—is essential for tasks such as color measurement, illumination and reflectance estimation, and material recognition. Spectral sensitivity is governed by the camera's optical system, including the sensor's color filter array and lens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Conventional methods [2, 4, 16, 22] for calibrating spectral sensitivity rely on specialized equipment to emit light with known spectra toward the camera and record the corresponding observed intensities. While accurate, these approaches are costly due to the need for precisely designed optical systems and extensive measurement time. Rather than directly observing light sources with known spectra, prior works [12, 21] have proposed using reference targets with known spectral reflectances under known spectral illuminations, such as sunlight or standard illuminants like D65. While these methods simplify the calibration process, a key challenge is the inherently low-frequency nature of spectral reflectances observed on natural objects [18] and high correlation between the spectral reflectances of indi" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.269, + 0.907, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.448, + 0.907, + 0.518 + ], + "angle": 0, + "content": "Figure 1. Overview of the proposed method. By capturing images of light with a known spectrum through an uncalibrated diffraction grating sheet, we observe both the direct and diffracted light. This practical and easy-to-use setup enables accurate estimation of camera spectral sensitivity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.536, + 0.906, + 0.566 + ], + "angle": 0, + "content": "vidual patches [5], which makes it difficult to accurately determine sensitivity at individual wavelengths." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.568, + 0.907, + 0.809 + ], + "angle": 0, + "content": "To enhance calibration accuracy, the use of diffraction grating sheets [13, 24] has been explored in the past. Diffraction occurs when light waves bend and spread as they pass through narrow apertures, with the degree of bending dependent on wavelength. This effect can be observed using an off-the-shelf diffraction grating sheet, available for less than 5 USD. Each diffracted light beam produces distinct spectral peaks, enabling precise estimation of spectral sensitivities at individual wavelengths. A key challenge in using a diffraction grating sheet is calibrating its diffraction parameter, the grating efficiency that describes the wavelength-dependent attenuation. To calibrate the efficiency, current methods require measurements of a reference target with known reflectances along with the diffracted light beam observation. This involves capturing multiple calibration scenes and complicates the process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "For an accurate and easy-to-use calibration method, we propose using an uncalibrated diffraction grating sheet without needing additional captures of reference targets. As shown in Fig. 1, the proposed method observes light with known spectra through the diffraction grating sheet, recording both the direct light and the diffraction pattern." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27252" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.079, + 0.482, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.339, + 0.485, + 0.507 + ], + "angle": 0, + "content": "Figure 2. Illustration of a diffraction grating. When light of a certain wavelength passes through a diffraction grating, its output intensity is spatially distributed in a characteristic pattern with multiple sharp peaks, known as intensity principal maxima. The pixel positions of the intensity principal maxima shift outward from the center for longer wavelengths. When a broadband light source passes, light beams of different wavelengths appear at distinct shifted positions. In practice, we also need to consider nonuniform transmittance, known as grating efficiency. As a result, the observed intensity at a given pixel position is the product of the incoming light spectrum, the light dispersion function, and the grating efficiency. See Fig. 3 for examples of the observations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.484, + 0.682 + ], + "angle": 0, + "content": "The proposed method simultaneously estimates the camera spectral sensitivity and grating efficiency. Although the original problem is a bilinear problem where the unknowns appear in their product form, we show that it can be turned into a tractable linear problem by introducing a basis representation for both camera spectral sensitivity and grating efficiency. We further introduce a closed-form solution method to the problem, given the pixel-to-wavelength mapping, i.e., the mapping between the pixel locations in the captured image and the wavelengths of the diffracted light." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.683, + 0.484, + 0.745 + ], + "angle": 0, + "content": "We conduct both synthetic and real-world experiments to demonstrate the effectiveness of the proposed method. The results show that our method outperforms existing leading methods using calibrated reflectance targets [12]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.76, + 0.462, + 0.775 + ], + "angle": 0, + "content": "Contributions Our chief contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.778, + 0.483, + 0.822 + ], + "angle": 0, + "content": "- We propose a camera spectral sensitivity estimation method that uses an uncalibrated diffraction grating sheet, without the need for capturing reference targets." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.823, + 0.482, + 0.853 + ], + "angle": 0, + "content": "- The method uses an inexpensive diffraction grating sheet, making it practical, cost-effective, and easy to use." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.853, + 0.483, + 0.899 + ], + "angle": 0, + "content": "- Accuracy of the proposed method significantly outperforms those that rely on reference targets, with a simpler setup and fewer captures." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.778, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.091, + 0.698, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.091, + 0.905, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.194, + 0.907, + 0.278 + ], + "angle": 0, + "content": "Figure 3. Capturing setup and the observed image of direct and diffracted lights. As the diffracted light exhibits multiple principal maxima, multiple diffraction patterns can be observed. The \\(k\\)-th diffraction image, counted from the center \\((k = \\pm 1, \\pm 2, \\ldots)\\), is referred to as the \\(k\\)-th order diffraction image. The image intensity is adjusted for visualization." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.303, + 0.634, + 0.318 + ], + "angle": 0, + "content": "Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.328, + 0.907, + 0.42 + ], + "angle": 0, + "content": "In this section, we provide the foundational background on spectral image formation. We introduce the general spectral image formation model, followed by an overview of how diffraction gratings separate light into its constituent wavelengths. Finally, we present the image formation model for observations captured through a diffraction grating." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.425, + 0.906, + 0.485 + ], + "angle": 0, + "content": "Spectral image formation Given camera spectral sensitivity \\( s(\\lambda_i) \\) at discretely sampled wavelengths \\( \\{\\lambda_i\\} \\) and spectrum of the incoming light \\( e(\\lambda_i) \\) to the camera, the intensity observation \\( m \\in \\mathbb{R}_+ \\) can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.497, + 0.907, + 0.528 + ], + "angle": 0, + "content": "\\[\nm \\propto \\sum_ {i} e (\\lambda_ {i}) s (\\lambda_ {i}). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.907, + 0.711 + ], + "angle": 0, + "content": "Diffraction grating A diffraction grating is an optical device with multiple periodic grooves that separates incoming light into its constituent wavelengths, directing each along a distinct path. Due to interference from the multi-groove structure, each wavelength produces multiple principal intensity maxima, corresponding to different diffraction orders \\( k = \\pm 1, \\pm 2, \\ldots \\), as illustrated in Fig. 2. This results in multiple dispersed observations symmetrically aligned with the direction of incoming light, with the central observation representing the mixture of all constituent wavelengths—namely, the direct light observation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.711, + 0.906, + 0.772 + ], + "angle": 0, + "content": "In general, diffraction observation \\(m_c^{(k)}(p)\\) around the \\(k^{\\mathrm{th}}\\) intensity principal maxima (i.o.w., \\(k^{\\mathrm{th}}\\) order diffraction observation) at pixel location \\(p\\) in channel \\(c\\in \\{R,G,B\\}\\) can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.783, + 0.907, + 0.814 + ], + "angle": 0, + "content": "\\[\nm _ {c} ^ {(k)} (p) \\propto \\sum_ {i} w (\\lambda_ {i}; p) e (\\lambda_ {i}) \\eta_ {k} (\\lambda_ {i}) s _ {c} (\\lambda_ {i}), \\qquad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\( s_c \\) is the camera spectral sensitivity for channel \\( c \\), and \\( \\eta_k(\\lambda) \\) represents the grating efficiency for each wavelength \\( \\lambda \\) at \\( k^{\\mathrm{th}} \\) order diffraction. The term \\( w(\\lambda; p) \\) is the weight induced by the light dispersion function, which disperses the light into each wavelength \\( \\lambda \\) that is observed at a" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27253" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "corresponding peak pixel position \\( p \\) depending on the wavelength. We refer to \\( w(\\lambda ;p) \\) as a weight function." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.484, + 0.228 + ], + "angle": 0, + "content": "The weight function \\( w(\\lambda; p) \\) approximates a delta function with a sufficient number of slits in the grating, effectively acting as a one-to-one mapping between wavelength \\( \\lambda \\) and pixel position \\( p \\) in the diffraction image [13, 24]. Ideally, the grating efficiency \\( \\eta(\\lambda) \\) should be all 1. However, in practice, the grating efficiency is known to vary depending on the frequency and the material of the grating [17]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.241, + 0.233, + 0.258 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.267, + 0.483, + 0.403 + ], + "angle": 0, + "content": "In this section, we provide an overview of existing methods for camera spectral sensitivity calibration. A straightforward approach to recovering spectral sensitivity is to measure the observed intensities \\( m \\) in the image while varying the incoming light with known spectra \\( e(\\lambda_i) \\). To achieve such controlled illuminations, specialized light sources [9] and light sources with narrow-band filters [2, 4, 16, 22] have been used. However, the equipment is expensive, limiting their practical applications in real-world settings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.484, + 0.508 + ], + "angle": 0, + "content": "To eliminate the need for specialized light sources, methods using reference targets, such as color checker charts [1, 7, 8, 11, 15, 19-21, 25], have been proposed. These methods recover camera spectral sensitivity by observing reference targets composed of multiple surface patches with known spectral reflectances under calibrated spectral illumination that uniformly illuminates the entire target." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.483, + 0.75 + ], + "angle": 0, + "content": "To reduce the cost of spectral illumination calibration, several methods [6, 14] assume specific illumination types, such as sunlight [14] or low-frequency spectra [10, 26], and simultaneously estimate camera spectral sensitivity and illumination spectra. While these methods can measure reflected light with different spectra from multiple patches in a single image, they rely on reference targets with known spectral reflectance. Unfortunately, off-the-shelf color checker charts do not provide such data, so one must either rely on externally measured data or perform their own calibration [27]. In addition, due to the low-frequency spectral reflectance of natural objects, the spectral resolution of these methods is limited. To address this challenge, methods using emissive charts [5] have been proposed; however, reference targets employing fluorescent materials are rare and have limited practical applicability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.751, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Instead of observing reflectances, methods using a diffraction grating, which makes light beams with distinct spectral peaks at different wavelengths, have been proposed [13, 24]. Karge et al. [13] uses a fluorescent lamp, which has a spiky spectrum, to calibrate the pixel-to-wavelength mapping. They then replace the fluorescent lamp with a halogen lamp, which has a flatter spectrum, to observe the diffraction pattern. They additionally need to capture reference targets with known spectral reflectances under controlled spectral illumination to account for the at" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.318 + ], + "angle": 0, + "content": "tenuation due to grating efficiency. Toivonen et al. [24] use multiple light sources with different spectra, generated by a halogen lamp equipped with calibrated color filters, to capture the multiple diffraction patterns. In addition, they also record images of a transmissive color chart, comprising multiple patches of calibrated color filters, through different light sources. By recording the diffraction patterns of different light sources and images of the transmissive color chart as input, they achieve accurate calibration of the spectral sensitivity. These methods use inexpensive setups compared to the specialized equipment and achieve more accurate calibration than those relying solely on a color checker chart. However, these approaches require multiple setups and complex capture procedures, which remain a significant practical challenge." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.905, + 0.469 + ], + "angle": 0, + "content": "Unlike the above approaches, Solomatov and Akkaynak [23] use metadata recorded in the Exif, specifically the color matrices that transforms sensor responses to the CIE XYZ color space, to train a neural network for estimating spectral sensitivity. This approach does not require any additional captures for calibration; however, it cannot account for external factors, such as lens and camera filters. Moreover, their estimation exhibits white-balance ambiguity, and recovering the full spectral sensitivity requires white-balance calibration using a color checker chart." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.906, + 0.621 + ], + "angle": 0, + "content": "We use an uncalibrated diffraction grating sheet, where the grating efficiency is unknown. Unlike previous diffraction grating-based methods, our approach only requires capturing a light source through the diffraction grating sheet, providing both the direct light and the diffraction pattern, enabling practical and accurate estimation. Notably, our minimal setting uses one light and two images under the light, and does not require any scene interaction after setup, whereas existing methods require at least two scene changes and three lights with known spectra [13]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.634, + 0.688, + 0.65 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.906, + 0.705 + ], + "angle": 0, + "content": "In this section, we first outline the problem setting for determining the camera spectral sensitivity using a diffraction grating. We then present the proposed method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.708, + 0.906, + 0.755 + ], + "angle": 0, + "content": "Problem setting Our image formation models of the direct light observation \\( m_{\\mathrm{dir}} \\) and the first-order diffracted light observation \\( m_{\\mathrm{dif}} \\) are written as:" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.765, + 0.905, + 0.804 + ], + "angle": 0, + "content": "\\[\nm _ {\\mathrm {d i r}} \\propto E \\sum_ {i} e \\left(\\lambda_ {i}\\right) s \\left(\\lambda_ {i}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.8, + 0.853, + 0.83 + ], + "angle": 0, + "content": "\\[\nm _ {\\mathrm {d i f}} (p) \\propto E \\sum_ {i} w (\\lambda_ {i}; p) e (\\lambda_ {i}) \\eta (\\lambda_ {i}) s (\\lambda_ {i}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where \\(E\\) is an unknown scale of the incoming light spectrum and \\(e(\\lambda)\\) is a normalized spectrum of the incoming light. We omit the pixel position dependency for the direct light observation \\(m_{\\mathrm{dir}}\\) as it remains constant along the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27254" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.079, + 0.904, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.269, + 0.907, + 0.381 + ], + "angle": 0, + "content": "Figure 4. Illustration of the image formation models and overview of our estimation. The observation should vary only along the diffraction direction. To obtain the mean observation, we average along the vertical direction in the figure, resulting in the direct light observation in \\(\\mathbb{R}^3\\) and the diffracted light observation in \\(\\mathbb{R}^{n\\times 3}\\) in RGB channels, respectively. The sampled observation plot in the bottom figure shows a per-channel vector plot along the diffraction direction for each channel. This plot should have the same outline as the wavelength-wise multiplication of the grating efficiency, camera spectral sensitivity, and the incoming light spectrum, leaving the unknown mapping along the horizontal axis, referred to as the pixel-to-wavelength mapping. Using both the direct and diffracted light observation, given the pixel-to-wavelength mapping, we estimate the camera spectral sensitivity and the grating efficiency as described in Sec 3.1. Section 3.2 covers the estimation of the pixel-to-wavelength mapping." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.407, + 0.484, + 0.467 + ], + "angle": 0, + "content": "diffraction (horizontal) direction, unlike the diffracted light. We do not use \\( k \\geq 2 \\) diffracted light observations in our method because, in practice, the intensity drastically decreases for higher-order diffraction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.473, + 0.484, + 0.564 + ], + "angle": 0, + "content": "Given a known incoming light spectrum \\( e(\\lambda) \\) and by observing both the direct and first-order diffracted lights, \\( m_{\\mathrm{dir}} \\) and \\( m_{\\mathrm{dif}}(p) \\), our goal is to estimate the camera's spectral sensitivity \\( s(\\lambda) \\) using an uncalibrated diffraction grating, where both the weight function \\( w(\\lambda; p) \\) and the grating efficiency \\( \\eta(\\lambda) \\) are unknown." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.484, + 0.645 + ], + "angle": 0, + "content": "In the following, we first describe the method to jointly estimate the camera spectral sensitivity \\( s(\\lambda) \\) and the grating efficiency \\( \\eta(\\lambda) \\), assuming that the weight function \\( w(\\lambda; p) \\) is known beforehand. We then present the method to determine the weight function \\( w(\\lambda; p) \\)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.671, + 0.484, + 0.703 + ], + "angle": 0, + "content": "3.1. Estimating camera spectral sensitivity \\( s \\) and grating efficiency \\( \\eta \\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.714, + 0.484, + 0.835 + ], + "angle": 0, + "content": "We now estimate the camera spectral sensitivity \\( s(\\lambda) \\) and the grating efficiency \\( \\eta(\\lambda) \\), given the weight function \\( w(\\lambda; p) \\). We show, for the first time, that the problem has a closed-form solution when the unknowns are expressed as linear combinations of their bases. To this end, we first explain the constraints derived from the image formation models of the direct and diffracted light observations, and then present the solution method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Let us assume camera spectral sensitivity \\( \\mathbf{s} \\) and the inverse of grating efficiency \\( \\eta^{-1} \\) at discretely sampled wavelengths \\( \\lambda_1, \\ldots, \\lambda_f \\) can be represented by linear combinations of bases whose numbers of bases are \\( b_s \\) and \\( b_\\eta \\), re" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.407, + 0.649, + 0.422 + ], + "angle": 0, + "content": "spectively. Namely," + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.432, + 0.86, + 0.453 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} \\triangleq \\left[ \\begin{array}{c c c} s (\\lambda_ {1}), & \\ldots , & s (\\lambda_ {f}) \\end{array} \\right] ^ {\\mathsf {T}} = \\mathbf {B} _ {s} \\mathbf {c} _ {s} \\in \\mathbb {R} _ {+} ^ {f},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.452, + 0.905, + 0.484 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\eta} ^ {- 1} \\triangleq \\left[ \\begin{array}{l l l} \\frac {1}{\\eta \\left(\\lambda_ {1}\\right)}, & \\dots , & \\frac {1}{\\eta \\left(\\lambda_ {f}\\right)} \\end{array} \\right] ^ {\\intercal} = \\mathbf {B} _ {\\eta} \\mathbf {c} _ {\\eta} \\in \\mathbb {R} _ {+} ^ {f}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.495, + 0.907, + 0.601 + ], + "angle": 0, + "content": "where \\(\\mathbf{B}_* \\in \\mathbb{R}^{f \\times b_*}\\) is a basis matrix, and \\(\\mathbf{c}_* \\in \\mathbb{R}^{b_*}\\) is a coefficient vector. We obtained the basis matrix \\(\\mathbf{B}_s\\) of camera spectral sensitivity \\(s\\) at a channel by singular value decomposition (SVD) on the same set of data used in [23]. For the inverse grating efficiency function \\(\\eta^{-1}\\), we use a Fourier basis representation. Further details can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.606, + 0.906, + 0.665 + ], + "angle": 0, + "content": "Linear constraint from direct light observation By introducing the basis representation, the aforementioned image formation model of direct light observations in Eq. (3) can be rewritten in a vector form as" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.676, + 0.905, + 0.695 + ], + "angle": 0, + "content": "\\[\nm _ {\\mathrm {d i r}} \\propto E \\mathbf {e} ^ {\\mathsf {T}} \\mathbf {s} = E \\mathbf {e} ^ {\\mathsf {T}} \\mathbf {B} _ {s} \\mathbf {c} _ {s}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.708, + 0.906, + 0.784 + ], + "angle": 0, + "content": "where \\(\\mathbf{e} \\in \\mathbb{R}_+^f\\) is the normalized incoming light spectrum. The light scale \\(E\\) can be set to 1, and the proportionality can be turned into equality without loss of generality, as the global scale can be accounted for in camera spectral sensitivity s. As a result, we have" + }, + { + "type": "equation", + "bbox": [ + 0.651, + 0.793, + 0.905, + 0.829 + ], + "angle": 0, + "content": "\\[\nm _ {\\mathrm {d i r}} = \\underbrace {\\mathbf {e} ^ {\\mathsf {T}} \\mathbf {B} _ {s}} _ {\\mathbf {A} _ {\\mathrm {d i r}}} \\mathbf {c} _ {s}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathbf{A}_{\\mathrm{dir}} \\in \\mathbb{R}^{1 \\times b_s}\\) is a known row vector. Similar equations can be derived for all RGB channels to form a vector of direct light observations \\(\\mathbf{m}_{\\mathrm{dir}} = [m_{\\mathrm{dir}(R)}, m_{\\mathrm{dir}(G)}, m_{\\mathrm{dir}(B)}]^{\\top}\\) on the left-hand side" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27255" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.137 + ], + "angle": 0, + "content": "of the equation, resulting in 3 linear equations with unknown vector \\([\\mathbf{c}_{s(R)}^{\\intercal},\\mathbf{c}_{s(G)}^{\\intercal},\\mathbf{c}_{s(B)}^{\\intercal}]^{\\intercal}\\), for which \\(\\mathbf{A}_{\\mathrm{dir}}\\) becomes a \\(3\\times 3b_{s}\\) matrix." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.14, + 0.484, + 0.2 + ], + "angle": 0, + "content": "Linear constraint from diffracted light observation Like the direct light observation case, the diffracted light's image formation model in Eq. (3) can be rewritten in a matrix form as" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.205, + 0.482, + 0.245 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {m} _ {\\text {d i f}} = \\left[ m _ {\\text {d i f}} \\left(p _ {1}\\right), \\dots , m _ {\\text {d i f}} \\left(p _ {n}\\right) \\right] ^ {\\top} \\tag {7} \\\\ = \\mathbf {W} \\operatorname {d i a g} (\\mathbf {e}) \\operatorname {d i a g} (\\boldsymbol {\\eta}) \\mathbf {s}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.253, + 0.484, + 0.301 + ], + "angle": 0, + "content": "where \\(\\mathbf{m}_{\\mathrm{dif}} \\in \\mathbb{R}_+^n\\) is a stack of the diffracted light observations at pixels \\(p_1, \\ldots, p_n\\). \\(\\mathbf{W} \\in \\mathbb{R}_+^{n \\times f}\\) denotes the known weight function for all \\(n\\) pixels, written in a matrix form as" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.32, + 0.483, + 0.378 + ], + "angle": 0, + "content": "\\[\n\\mathbf {W} = \\left[ \\begin{array}{c c c} w \\left(\\lambda_ {1}, p _ {1}\\right) & \\dots & w \\left(\\lambda_ {f}, p _ {1}\\right) \\\\ \\vdots & & \\\\ w \\left(\\lambda_ {1}, p _ {n}\\right) & \\dots & w \\left(\\lambda_ {f}, p _ {n}\\right) \\end{array} \\right] \\in \\mathbb {R} _ {+} ^ {n \\times f}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.384, + 0.351, + 0.398 + ], + "angle": 0, + "content": "which we refer to as the weight matrix." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.399, + 0.483, + 0.429 + ], + "angle": 0, + "content": "The above image formation model of diffracted light can be equivalently turned into" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.435, + 0.483, + 0.484 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} = \\operatorname {d i a g} \\left(\\boldsymbol {\\eta} ^ {- 1}\\right) \\underbrace {\\operatorname {d i a g} \\left(\\mathbf {e} ^ {- 1}\\right) \\mathbf {W} ^ {\\dagger} \\mathbf {m} _ {\\mathrm {d i f}}} _ {\\mathbf {a}} = \\operatorname {d i a g} (\\mathbf {a}) \\boldsymbol {\\eta} ^ {- 1}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.485, + 0.484, + 0.59 + ], + "angle": 0, + "content": "where the inverse of a vector denotes an element-wise inverse operator, \\(\\mathbf{W}^{\\dagger}\\) is a left pseudo-inverse of \\(\\mathbf{W}\\), and \\(\\mathbf{a} \\in \\mathbb{R}^{f}\\) is a known channel-and-light-dependent vector. Since we can confidently assume \\(n > f\\) in our setting, the weight matrix \\(\\mathbf{W}\\) should have full column rank when the diffracted observation spans the entire visible wavelength domain, and thus it has a left pseudo-inverse." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.591, + 0.371, + 0.606 + ], + "angle": 0, + "content": "From Eq. (4), Eq. (9) can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.614, + 0.482, + 0.631 + ], + "angle": 0, + "content": "\\[\n\\mathbf {B} _ {s} \\mathbf {c} _ {s} = \\operatorname {d i a g} (\\mathbf {a}) \\mathbf {B} _ {\\eta} \\mathbf {c} _ {\\eta}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.639, + 0.484, + 0.669 + ], + "angle": 0, + "content": "Putting together the unknowns \\(\\mathbf{c}_{\\eta}\\) and \\(\\mathbf{c}_s\\), we have the following homogeneous system of equations:" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.676, + 0.483, + 0.72 + ], + "angle": 0, + "content": "\\[\n\\underbrace {\\left[ \\begin{array}{l l} \\operatorname {d i a g} (\\mathbf {a}) \\mathbf {B} _ {\\eta} & - \\mathbf {B} _ {s} \\end{array} \\right]} _ {\\mathbf {A} _ {\\mathrm {d i f}}} \\left[ \\begin{array}{l} \\mathbf {c} _ {\\eta} \\\\ \\mathbf {c} _ {s} \\end{array} \\right] = \\mathbf {0}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.729, + 0.484, + 0.822 + ], + "angle": 0, + "content": "where \\(\\mathbf{A}_{\\mathrm{dif}} \\in \\mathbb{R}^{f \\times (b_{\\eta} + b_{s})}\\) is a known matrix. For the case of RGB images, stacking similar equations from all RGB channels, we obtain \\(3f\\) linear equations from the diffracted light observations for unknown vector \\([\\mathbf{c}_{\\eta}^{\\top}, \\mathbf{c}_{s(R)}^{\\top}, \\mathbf{c}_{s(G)}^{\\top}, \\mathbf{c}_{s(B)}^{\\top}]^{\\top} \\in \\mathbb{R}^{b_{\\eta} + 3b_{s}}\\), and \\(\\mathbf{A}_{\\mathrm{dif}}\\) becomes a \\(3f \\times (b_{\\eta} + 3b_{s})\\) matrix." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.871 + ], + "angle": 0, + "content": "Solution method We can estimate the camera spectral sensitivity \\( s \\) and grating efficiency \\( \\eta \\) by solving the following optimization problem using RGB channels:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.878, + 0.484, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} ^ {*} = \\underset {\\mathbf {x}} {\\operatorname {a r g m i n}} \\| \\mathbf {A} _ {\\mathrm {d i f}} \\mathbf {x} \\| _ {2} ^ {2} \\text {s . t .} \\left[ \\begin{array}{l l} \\mathbf {0} & \\mathbf {A} _ {\\mathrm {d i r}} \\end{array} \\right] \\mathbf {x} = \\mathbf {m} _ {\\mathrm {d i r}}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.089, + 0.907, + 0.206 + ], + "angle": 0, + "content": "where \\(\\mathbf{x} = \\left[\\mathbf{c}_{\\eta}^{\\mathsf{T}},\\mathbf{c}_{s(R)}^{\\mathsf{T}},\\mathbf{c}_{s(G)}^{\\mathsf{T}},\\mathbf{c}_{s(B)}^{\\mathsf{T}}\\right]^{\\mathsf{T}}\\in \\mathbb{R}^{b_{\\eta} + 3b_s}\\) is the vector of unknowns. The above optimization problem has a closed-form solution by introducing a Lagrange multiplier when the number bases \\(b_{*}\\) and the wavelength resolution \\(f\\) satisfy \\(3f\\geq b_{\\eta} + 3b_{s}\\), and \\(\\mathbf{A}_{\\mathrm{dif}}\\) and \\(\\mathbf{A}_{\\mathrm{dir}}\\) are full-rank. The details of the solution method are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.217, + 0.795, + 0.233 + ], + "angle": 0, + "content": "3.2. Estimation of weight function \\( w \\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.239, + 0.907, + 0.359 + ], + "angle": 0, + "content": "As described in Sec. 1, the weight function \\( w(\\lambda; p) \\) has a peaky shape using diffraction gratings; therefore, we assume there is a one-to-one mapping between the pixel location \\( p \\) in the diffraction image and the wavelength \\( \\lambda \\) as in previous methods [13, 24]. To estimate the weight function \\( w(\\lambda; p) \\), it is necessary to establish the pixel-to-wavelength mapping, namely, relating the pixel locations of the diffraction image to the wavelength as illustrated in Fig. 6." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.361, + 0.906, + 0.392 + ], + "angle": 0, + "content": "We propose to use a quadratic function to represent the one-to-one pixel-to-wavelength mapping \\( g: p \\to \\lambda \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.403, + 0.905, + 0.421 + ], + "angle": 0, + "content": "\\[\n\\lambda = g (p) = a p ^ {2} + b p + c, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.906, + 0.495 + ], + "angle": 0, + "content": "where \\(a, b, c \\in \\mathbb{R}\\) are the parameters. This quadratic representation allows us stably determining the mapping with retaining accuracy. For further discussion about this parameterization, please refer to our supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.495, + 0.906, + 0.541 + ], + "angle": 0, + "content": "Once the pixel-to-wavelength mapping \\( g(p) \\) is determined, the elements of the weight matrix \\( \\mathbf{W} \\) is simply obtained by linear interpolation as" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.552, + 0.906, + 0.64 + ], + "angle": 0, + "content": "\\[\nw (\\lambda , p) = \\left\\{ \\begin{array}{l l} \\frac {\\lambda_ {i + 1} - g (p)}{\\lambda_ {i + 1} - \\lambda_ {i}} & \\text {i f} \\lambda = \\lambda_ {i} \\\\ \\frac {g (p) - \\lambda_ {i}}{\\lambda_ {i + 1} - \\lambda_ {i}} & \\text {i f} \\lambda = \\lambda_ {i + 1}, \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.652, + 0.907, + 0.803 + ], + "angle": 0, + "content": "where \\(\\lambda_{i}\\) is the nearest-neighbor wavelength to \\(g(p)\\) that satisfies \\(\\lambda_{i} \\leq g(p) \\leq \\lambda_{i+1}\\). Therefore, the problem of estimating the weight function boils down to the estimation of pixel-to-wavelength mapping. The challenge, though, is that we only have access to the diffraction image and illumination spectra, but not to camera spectral sensitivity and grating efficiency. Now we describe two settings for determining the pixel-to-wavelength mapping; one using fluorescent and LED lights (denoted as Fluorescent+LED), and another using only LED light (denoted as LED only)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Fluorescent+LED case When light with spiky spectral peaks, such as fluorescent light, passes through the diffraction grating, the resulting diffracted observation exhibits a similar spiky pattern, as shown in Fig. 5. Therefore, the fluorescent light provides a strong cue for determining the pixel-to-wavelength mapping. In this case, we use a similar" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27256" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.48, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.199, + 0.484, + 0.255 + ], + "angle": 0, + "content": "Figure 5. The illuminant spectrum of the light source used in our real-world experiment and an example of the corresponding direct and diffracted light observations. The observations shown here are adjusted for visualization." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.267, + 0.484, + 0.403 + ], + "angle": 0, + "content": "approach to [13] that associates corresponding peak positions in the pixel and wavelength domains. Since the spectral sensitivity \\(\\mathbf{s}\\) is unknown, we weigh the illumination spectra \\(\\mathbf{e}\\) by the mean spectral sensitivity \\(\\bar{\\mathbf{s}}\\) as \\(\\bar{\\mathbf{s}}\\odot \\mathbf{e}\\), where \\(\\bar{\\mathbf{s}}\\) is obtained from the same dataset used for the camera spectral sensitivity basis [23], and \\(\\odot\\) represents an elementwise multiplication operator. We have observed that the unknown grating efficiency \\(\\eta\\) is negligible in determining the peak positions due to its low-frequency distribution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.484, + 0.508 + ], + "angle": 0, + "content": "We select the top two peaks from each of the pixel and wavelength domains for each channel to determine the polynomial parameters \\(a\\), \\(b\\), and \\(c\\) of the pixel-to-wavelength mapping \\(g\\) via linear least squares. Once the pixel-to-wavelength mapping \\(g\\) and weight matrix \\(\\mathbf{W}\\) are obtained, we take another capture using an LED light for obtaining the direct and diffracted measurements used in Sec. 3.1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.512, + 0.484, + 0.587 + ], + "angle": 0, + "content": "LED only case To simplify the recording process, it is preferable if we can work only with an LED. In this case, due to the low-frequency incoming spectra without obvious peaks, we cannot take the strategy of relating peak positions used in the Fluorescent+LED case." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.588, + 0.484, + 0.679 + ], + "angle": 0, + "content": "Therefore, instead of finding peak positions, we frame the problem as finding the optimal horizontal mapping (pixel-to-wavelength) between two plots—the normalized diffraction observations \\(\\mathbf{u}\\) and the normalized mean camera spectral sensitivity \\(\\mathbf{v}\\) that is scaled by illumination spectrum, expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.687, + 0.45, + 0.707 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} \\triangleq \\nu (\\mathbf {e} \\odot \\bar {\\mathbf {s}}) = [ v (\\lambda_ {1}), \\dots , v (\\lambda_ {f}) ] \\in \\mathbb {R} _ {+} ^ {f}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.709, + 0.455, + 0.729 + ], + "angle": 0, + "content": "\\[\n\\mathbf {u} \\triangleq \\nu (\\mathbf {m} _ {\\mathrm {d i f}}) = [ u (p _ {1}), \\dots , u (p _ {n}) ] \\in \\mathbb {R} _ {+} ^ {n},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.736, + 0.45, + 0.752 + ], + "angle": 0, + "content": "where \\(\\nu\\) is a vector normalization as \\(\\nu (\\mathbf{x}) = \\mathbf{x} / \\| \\mathbf{x}\\| _2\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.752, + 0.484, + 0.813 + ], + "angle": 0, + "content": "We solve the point-to-plane iterative closest point (ICP) problem [3] to determine the optimal polynomial parameters \\( a \\), \\( b \\), and \\( c \\) along the horizontal axis by minimizing the following objective function:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.819, + 0.489, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {I C P}} = \\sum_ {i = 1} ^ {n} \\sum_ {j \\in \\mathcal {N} (i)} \\left(\\left(\\left[ \\begin{array}{c} a p _ {i} ^ {2} + b p _ {i} + c \\\\ u (p _ {i}) \\end{array} \\right] - \\left[ \\begin{array}{c} \\lambda_ {j} \\\\ v (\\lambda_ {j}) \\end{array} \\right]\\right) ^ {\\top} \\mathbf {n _ {v}} (j)\\right) ^ {2},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.868, + 0.484, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathcal{N}(i)\\) is the set of indices for the current neighboring points of \\([ap_i^2 +bp_i + c,u(p_i)]^\\top\\), and \\(\\mathbf{n_v}(j)\\in \\mathbb{S}\\subset \\mathbb{R}^2\\)" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.905, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.291, + 0.906, + 0.376 + ], + "angle": 0, + "content": "Figure 6. Illustration of our pixel-to-wavelength mapping estimation. Given the normalized diffraction observations \\(\\mathbf{u}_c\\) and normalized illumination-scaled mean camera spectral sensitivity \\(\\mathbf{v}_c\\) for all \\(c \\in \\{R, G, B\\}\\) channels, we estimate the optimal pixel-to-wavelength mapping (or horizontal warping, as shown in the figure), which minimizes the distance between the two plots." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.387, + 0.866, + 0.407 + ], + "angle": 0, + "content": "is a unit normal vector to \\([\\lambda_j,v(\\lambda_j)]^\\top\\) which satisfies" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.416, + 0.905, + 0.455 + ], + "angle": 0, + "content": "\\[\n\\left[ \\begin{array}{c} \\lambda_ {j} - \\lambda_ {j + 1} \\\\ v (\\lambda_ {j}) - v (\\lambda_ {j + 1}) \\end{array} \\right] ^ {\\mathsf {T}} \\mathbf {n} _ {\\mathrm {v}} (j) = 0. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.907, + 0.583 + ], + "angle": 0, + "content": "We iteratively minimize the objective function for 500 iterations, with the initial guesses set to \\( a = 0 \\), \\( b = f / n \\), \\( c = 0 \\), where \\( b \\) is given as the ratio of the numbers of elements of \\( \\mathbf{u} \\) and \\( \\mathbf{v} \\), to determine the pixel-to-wavelength mapping \\( g \\). This setting allows for a convenient one-time capture for both determining the pixel-to-wavelength mapping and obtaining the direct and diffracted measurements used in Sec. 3.1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.598, + 0.646, + 0.615 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.624, + 0.906, + 0.67 + ], + "angle": 0, + "content": "In this section, we evaluate our method on both synthetic and real-world scenes. We first describe our experimental setup and then present the evaluation results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.856 + ], + "angle": 0, + "content": "Baselines We compare the proposed approach with the Exif-based method (Exif+CC) [23] and the color chart-based method (CC) [12]. We present two variants of the proposed method as detailed in Sec. 3.2: Ours (LED+Flu), which employs a fluorescent lamp observation for pixel-to-wavelength mapping estimation, and Ours (LED), which uses only an LED light observation. For the color chart-based method, we use 2 basis functions, as mentioned in the original paper, which yielded the best results in our preliminary experiments. Since the Exif-based method cannot estimate white balance and introduces a \\(3 \\times 3\\) ambiguity, we use the color checker observation to resolve this ambiguity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We also compared ours with the existing diffraction grating-based approach by Toivonen et al. [24], adjusted to our experimental setup. However, their method produced" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27257" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.138 + ], + "angle": 0, + "content": "unstable estimations due to the limited observations available in our setup compared to their original conditions. Further details are provided in the supplemental material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.142, + 0.483, + 0.415 + ], + "angle": 0, + "content": "Basis representations To compute the basis for camera spectral sensitivity \\( s(\\lambda) \\), we use the dataset from the Exif-based method [23], which includes 44 cameras for training and 5 for testing. Following the detailed analysis on basis selection for a camera spectral sensitivity in [12], we apply SVD to the sensitivities in the training dataset in a per-channel manner. For the grating efficiency, we use a Fourier basis. Further discussion on the grating efficiency basis can be found in the supplementary material. The spectral resolution \\( f \\) in the dataset of camera spectral sensitivities is \\( f = 31 \\) from \\( 400\\mathrm{nm} \\) to \\( 700\\mathrm{nm} \\). Assuming the number of basis for the camera spectral sensitivity and the grating efficiency to be the same \\( b_{\\eta} = b_{s} \\), the required condition for the number of bases for our method is \\( b_{\\eta} = b_{s} \\leq 23 \\) based on the discussion in Sec. 3.1. For robust estimation, the number of basis functions is set to 7 for each channel of the camera spectral sensitivity and the grating efficiency across all experiments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.42, + 0.483, + 0.465 + ], + "angle": 0, + "content": "Evaluation metrics Following [23], we use the average relative full-scale error across the three color channels, RE, which is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.475, + 0.483, + 0.514 + ], + "angle": 0, + "content": "\\[\n\\mathrm {R E} = \\frac {1}{3} \\sum_ {c} \\frac {\\operatorname {R M S E} (c)}{\\max \\left(\\mathbf {s} _ {c}\\right)}, \\operatorname {R M S E} (c) = \\sqrt {\\frac {\\left\\| \\mathbf {s} _ {c} ^ {*} - \\mathbf {s} _ {c} \\right\\| _ {2} ^ {2}}{f}}, \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.525, + 0.483, + 0.556 + ], + "angle": 0, + "content": "where \\(\\mathbf{s}_c^*\\) and \\(\\mathbf{s}_c\\) denote the estimated and ground-truth camera spectral sensitivity for each channel \\(c\\in \\{R,G,B\\}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.566, + 0.377, + 0.581 + ], + "angle": 0, + "content": "4.1. Evaluation with synthetic scenes" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.588, + 0.483, + 0.618 + ], + "angle": 0, + "content": "In this section, we quantitatively evaluate the comparison methods and ours on a synthetic dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.624, + 0.483, + 0.835 + ], + "angle": 0, + "content": "Dataset We render the direct and diffracted images using Eq. (2), with parameters as follows. We use five camera spectral sensitivities: EOS 600D, iPhone12ProMax, Olympus EPL2, Pentax K5, and Sumsung GalaxyS20, all from the same test dataset as the Exif-based method. The D65 illuminant, as defined by the CIE Standard Illuminants, is used as the incoming light. To synthesize the pixel-to-wavelength mapping, we randomly define the mapping while ensuring that the visible wavelength range is fully covered in the diffraction image. The grating efficiency function is set to a low-frequency response as illustrated in Fig. 4, parameterized by 7 Fourier bases. Since the Exif-based method relies on metadata rather than images, we evaluate their method using real-world data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Results Figure 7(a) shows the estimated results of the proposed and comparison methods for our synthetic scenes. For ideal observations, both Ours (LED+Flu) and the color chart-based method, CC, yield low estimation errors, as" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.243 + ], + "angle": 0, + "content": "both methods are guaranteed to produce optimal results under the basis representation assumption. The remaining errors stem from the gap between the assumption and the synthetic data. The Exif-based method, Exif+CC, produces a generally reasonable overall shape, but it fails to accurately capture finer details. Ours (LED) yields a larger error than Ours (LED+Flu) due to estimation errors in the pixel-to-wavelength mapping when using only LED observations; however, overall, we can still obtain a reasonable estimation even by capturing just a single calibration scene." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.253, + 0.81, + 0.268 + ], + "angle": 0, + "content": "4.2. Evaluation with real-world scenes" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.276, + 0.896, + 0.291 + ], + "angle": 0, + "content": "Here, we present evaluations using our real-world dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.296, + 0.906, + 0.552 + ], + "angle": 0, + "content": "Dataset 3 illustrates our experimental setup, which consists of a light source, a diffraction grating sheet, and a camera. We capture the images in a dark room for both our method and the comparison methods, ensuring that the only light in the scene comes from illumination with known spectra (LED or fluorescent). We use a 3D-printed slit for the incoming light, positioned in front of the light source, and place a diffraction grating in front of the camera, manually aligned roughly parallel to the image plane. For the spiky light source, we use a Pa-Look Ball (EFD15EL)1 fluorescent lamp. We use the following four cameras: EOS RP, iPhone 15ProMax, Sony \\(\\alpha\\)1, and DJI Pocket3. For the non-spiky light sources and diffraction sheets, we use three different products for each. In this section, we present the results for one setup. Additional experiments with other setups are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.704 + ], + "angle": 0, + "content": "Due to the high dynamic range of the direct and diffracted light, we capture HDR images by taking two images with different exposure times for each. The extra diffracted image under fluorescent light is captured for LED+Flu, solely for pixel-to-wavelength mapping. After capturing the direct and diffracted light observations, we roughly crop them with a sufficient margin at both ends of the diffraction pattern along the diffraction direction. The diffracted light observation is cropped to approximately \\(200 \\times 1000\\) pix and aggregated to \\(1 \\times 1000\\) pix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.906, + 0.841 + ], + "angle": 0, + "content": "For the comparison methods, we use a 24-patch color checker chart with spectral reflectance data from Chromaxion. The color checker chart is captured under the same non-spiky light source used in our experiments. We carefully position the color checker chart so that it is parallel to the camera and uniformly illuminated by the light source, ensuring that no shadows are cast. The metadata for the Exif-based method is extracted from our captured data following the procedure described in their paper [23]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.533, + 0.852, + 0.663, + 0.863 + ], + "angle": 0, + "content": "1 Panasonic Corporation" + }, + { + "type": "page_footnote", + "bbox": [ + 0.533, + 0.864, + 0.709, + 0.875 + ], + "angle": 0, + "content": "2ColorChecker Classic, Calibrite" + }, + { + "type": "page_footnote", + "bbox": [ + 0.515, + 0.876, + 0.905, + 0.899 + ], + "angle": 0, + "content": "3Chromaxion, https://www.chromaxion.com, last accessed on March 7, 2025." + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.852, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27258" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.243, + 0.09, + 0.393, + 0.104 + ], + "angle": 0, + "content": "(a) Synthetic experiment" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.106, + 0.511, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.637, + 0.09, + 0.796, + 0.104 + ], + "angle": 0, + "content": "(b) Real-world experiment" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.106, + 0.918, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.506, + 0.908, + 0.562 + ], + "angle": 0, + "content": "Figure 7. Results of the proposed and comparison methods for our (a) synthetic and (b) real-world datasets. For each setting and method, we show plots of the estimated (solid line) and ground-truth (dotted line) spectral sensitivity for each color channel. All plots are normalized by the maximum response across all channels. The values below the plots indicate the RE defined in Eq. (16). For simplicity, the factor of \\(10^{-2}\\) is omitted. Underscored bold and underscored score are the best and second-best result for each setting, respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.588, + 0.483, + 0.65 + ], + "angle": 0, + "content": "The ground-truth spectral sensitivities are obtained using a calibration device with a lens attached. The illuminant spectra of the light sources, shown in Fig. 5, are measured through the diffraction grating using a spectrometer." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.656, + 0.483, + 0.866 + ], + "angle": 0, + "content": "Results Figure 7(b) presents the estimated results of the proposed and comparison methods. Overall, the proposed methods achieve the lowest estimation errors for most scenes, and Ours (LED) produces results nearly comparable to Ours (LED+Flu). For the EOS RP scene, Ours (LED) yields a larger error due to the failure of pixel-to-wavelength mapping estimation using a non-spiky light source. The color checker-based method, CC, is relatively sensitive to real-world noise due to the low dimensionality of the spectral reflectances in the color patches, as discussed in previous works [5, 24]. While the Exif-based method, Exif+CC, has an advantage of being easy to use, as it only relies on metadata for recovering sensitivity, it yields larger estimation errors across all scenes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.587, + 0.64, + 0.603 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.613, + 0.907, + 0.749 + ], + "angle": 0, + "content": "We present a method for estimating camera spectral sensitivity using an uncalibrated diffraction grating sheet. Compared to previous methods that use reference targets, such as a color checker chart, our approach achieves highly accurate estimation by leveraging the diffraction grating. Unlike previous diffraction grating-based methods, the proposed method jointly estimates both camera spectral sensitivity and grating efficiency without needing additional recordings, making it practical, cost-efficient, and easy-to-use." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.902 + ], + "angle": 0, + "content": "One of the limitations of our method is that pixel-to-wavelength mapping estimation breaks down if the illumination has uniform spectra. In this case, the diffraction image simply shows the camera spectral sensitivity scaled by unknown grating efficiency, providing no clues for the mapping estimation. However, in practice, readily-available light sources do not have such uniform spectra. LEDs, for example, tend to have a non-uniform spectral distribution, enabling us to estimate the mapping and camera spectral sensitivities, as demonstrated in our real-world experiments." + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.875, + 0.404, + 0.887 + ], + "angle": 0, + "content": "4camSPECS XL, Image Engineering GmbH & Co. KG" + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.887, + 0.323, + 0.9 + ], + "angle": 0, + "content": "\\(^{5}\\)CS-2000S, KONICA MINOLTA, INC." + }, + { + "type": "list", + "bbox": [ + 0.108, + 0.875, + 0.404, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27259" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.201 + ], + "angle": 0, + "content": "This work was supported by JSPS KAKENHI Grant Numbers JP23H05491 and JP25K00142, JST ASPIRE Grant Number JPMJAP2304, the Canada First Research Excellence Fund program (VISTA), an NSERC Discovery Grant, and the Canada Research Chair program." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.212, + 0.188, + 0.228 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.237, + 0.484, + 0.293 + ], + "angle": 0, + "content": "[1] Ali Alsam and Graham Finlayson. Recovering spectral sensitivities with uncertainty. In Proceedings of Conference on Colour in Graphics, Imaging, and Vision (CGIV), pages 22-26, 2002. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.293, + 0.484, + 0.39 + ], + "angle": 0, + "content": "[2] E Berra, S Gibson-Poole, A MacArthur, R Gaulton, and A Hamilton. Estimation of the spectral sensitivity functions of un-modified and modified commercial off-the-shelf digital cameras to enable their use as a multispectral imaging system for UAVs. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 40: 207-214, 2015. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.392, + 0.483, + 0.434 + ], + "angle": 0, + "content": "[3] Yang Chen and Gérard Medioni. Object modelling by registration of multiple range images. Image and Vision Computing, 10(3):145-155, 1992. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.435, + 0.483, + 0.49 + ], + "angle": 0, + "content": "[4] Maryam Mohammadzadeh Darrodi, Graham Finlayson, Teresa Goodman, and Michal Mackiewicz. Reference data set for camera spectral sensitivity estimation. Journal of the Optical Society of America A, 32(3):381-391, 2015. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.491, + 0.484, + 0.546 + ], + "angle": 0, + "content": "[5] Jeffrey M DiCarlo, Glen Eric Montgomery, and Steven W Trovinger. Emissive chart for imager calibration. In Proceedings of IS&T Color and Imaging Conference (CIC), pages 295-301, 2004. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.548, + 0.483, + 0.589 + ], + "angle": 0, + "content": "[6] Bob Dyas. Robust color sensor response characterization. In Proceedings of IS&T Color and Imaging Conference (CIC), pages 144-148, 2000. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.591, + 0.484, + 0.646 + ], + "angle": 0, + "content": "[7] Marc Ebner. Estimating the spectral sensitivity of a digital sensor using calibration targets. In Proceedings of Annual Conference on Genetic and Evolutionary Computation, pages 642-649, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.647, + 0.483, + 0.688 + ], + "angle": 0, + "content": "[8] Graham Finlayson, Maryam Mohammadzadeh Darrodi, and Michal Mackiewicz. Rank-based camera spectral sensitivity estimation. 33(4):589-599, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.689, + 0.484, + 0.745 + ], + "angle": 0, + "content": "[9] Paul M Hubel, Doron Sherman, and Joyce E Farrell. A comparison of methods of sensor spectral sensitivity estimation. In Proceedings of IS&T Color and Imaging Conference (CIC), pages 45-48, 1994. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[10] Cong Phuoc Huynh and Antonio Robles-Kelly. Recovery of spectral sensitivity functions from a colour chart image under unknown spectrally smooth illumination. In Proceedings of IEEE International Conference on Pattern Recognition (ICPR), pages 708-713, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[11] Yuhyun Ji, Yunsang Kwak, Sang Mok Park, and Young L Kim. Compressive recovery of smartphone RGB spectral sensitivity functions. Optics Express, 29(8):11947-11961, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.872, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[12] Jun Jiang, Dengyu Liu, Jinwei Gu, and Sabine Süssstrunk. What is the space of spectral sensitivity functions for digital" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.237, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.906, + 0.134 + ], + "angle": 0, + "content": "color cameras? In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 168-179, 2013. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.906, + 0.206 + ], + "angle": 0, + "content": "[13] Andreas Karge, Ingmar Rieger, Bernhard Eberhardt, and Andreas Schilling. Using chromaticity error minimisation for fast camera spectral responsivity measurement. In Proceedings of IS&T Color and Imaging Conference (CIC), pages 67-74, 2018. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.906, + 0.263 + ], + "angle": 0, + "content": "[14] Rei Kawakami, Hongxun Zhao, Robby T Tan, and Katsushi Ikeuchi. Camera spectral sensitivity and white balance estimation from sky images. International Journal of Computer Vision (IJCV), 105:187-204, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.265, + 0.906, + 0.333 + ], + "angle": 0, + "content": "[15] Berk Kaya, Yigit Baran Can, and Radu Timofte. Towards spectral estimation from a single RGB image in the wild. In Proceedings of IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 3546-3555, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.337, + 0.906, + 0.393 + ], + "angle": 0, + "content": "[16] Lindsay W MacDonald. Determining camera spectral responsivity with multispectral transmission filters. In Proceedings of IS&T Color and Imaging Conference (CIC), pages 12-17, 2015. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.395, + 0.906, + 0.423 + ], + "angle": 0, + "content": "[17] Christopher Palmer and Erwin G Loewen. Diffraction Grating Handbook. Newport Corporation New York, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.425, + 0.906, + 0.467 + ], + "angle": 0, + "content": "[18] J. P. S. Parkkinen, J. Hallikainen, and T. Jaaskelainen. Characteristic spectra of munsell colors. Journal of the Optical Society of America A, 6(2):318-322, 1989. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.906, + 0.537 + ], + "angle": 0, + "content": "[19] Dilip Prasad, Rang Nguyen, and Michael Brown. Quick approximation of camera's spectral response from casual lighting. In Proceedings of IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 844-851, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.906, + 0.596 + ], + "angle": 0, + "content": "[20] Shuxue Quan, Noboru Ohta, and Xiaoyun Jiang. Comparative study on sensor spectral sensitivity estimation. In Proceedings of SPIE Color Imaging VIII: Processing, Hardcopy, and Applications, pages 209-220, 2003." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.598, + 0.906, + 0.64 + ], + "angle": 0, + "content": "[21] Martin Rump, Arno Zinke, and Reinhard Klein. Practical spectral characterization of trichromatic cameras. ACM Transactions on Graphics (TOG), 30(6):170, 2011. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.642, + 0.906, + 0.683 + ], + "angle": 0, + "content": "[22] Yuri Rzhanov. Accurate determination of camera quantum efficiency from a single image. Journal of Imaging, 10(7): 169, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.686, + 0.906, + 0.741 + ], + "angle": 0, + "content": "[23] Grigory Solomatov and Derya Akkaynak. Spectral sensitivity estimation without a camera. In International Conference on Computational Photography (ICCP), pages 1-12, 2023. 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.743, + 0.908, + 0.786 + ], + "angle": 0, + "content": "[24] Mikko E Toivonen and Arto Klami. Practical camera sensor spectral response and uncertainty estimation. Journal of Imaging, 6(8):79, 2020. 1, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.906, + 0.83 + ], + "angle": 0, + "content": "[25] Shoji Tominaga, Shogo Nishi, and Ryo Ohtera. Measurement and estimation of spectral sensitivity functions for mobile phone cameras. Sensors, 21(15):4985, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[26] Lin Zhang, Ying Fu, Yinqiang Zheng, and Hua Huang. Camera spectral sensitivity, illumination and spectral reflectance estimation for a hybrid hyperspectral image capture system. In Proceedings of IEEE International Conference on Image Processing (ICIP), pages 545-549, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27260" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.148 + ], + "angle": 0, + "content": "[27] Jieyu Zhu, Xufen Xie, Ningfang Liao, Zhuangzhi Zhang, Wenmin Wu, and Lvming Lv. Spectral sensitivity estimation of trichromatic camera based on orthogonal test and window filtering. Optics Express, 28(19):28085-28100, 2020. 3" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27261" + } + ] +] \ No newline at end of file diff --git a/2025/Spectral Sensitivity Estimation with an Uncalibrated Diffraction Grating/79b5704f-97ff-445c-93ff-ff5edaa3c620_origin.pdf b/2025/Spectral Sensitivity Estimation with an Uncalibrated Diffraction Grating/79b5704f-97ff-445c-93ff-ff5edaa3c620_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..115ce9a2d29f3ddcbe656aba602a062649dbfef7 --- /dev/null +++ b/2025/Spectral Sensitivity Estimation with an Uncalibrated Diffraction Grating/79b5704f-97ff-445c-93ff-ff5edaa3c620_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:53f59c81eb1cecae9ae2d13426bae358fdba9396850dfe9a56e42c900b69a802 +size 10167673 diff --git a/2025/Spectral Sensitivity Estimation with an Uncalibrated Diffraction Grating/full.md b/2025/Spectral Sensitivity Estimation with an Uncalibrated Diffraction Grating/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6d3924aa453a3067cfbb875c4dc1767a13b23983 --- /dev/null +++ b/2025/Spectral Sensitivity Estimation with an Uncalibrated Diffraction Grating/full.md @@ -0,0 +1,342 @@ +# Spectral Sensitivity Estimation with an Uncalibrated Diffraction Grating + +Lilika Makabe1 Hiroaki Santo1 Fumio Okura1 Michael S. Brown2 Yasuyuki Matsushita1,3 +1The University of Osaka 2York University 3Microsoft Research Asia - Tokyo + +{makabe.lilika, santo.hiroaki, okura, yasumat}@ist.osaka-u.ac.jp mbrown@yorku.ca + +# Abstract + +This paper introduces a practical and accurate calibration method for camera spectral sensitivity using a diffraction grating. Accurate calibration of camera spectral sensitivity is crucial for various computer vision tasks, including color correction, illumination estimation, and material analysis. Unlike existing approaches that require specialized narrow-band filters or reference targets with known spectral reflectances, our method only requires an uncalibrated diffraction grating sheet, readily available off-the-shelf. By capturing images of the direct illumination and its diffracted pattern through the grating sheet, our method estimates both the camera spectral sensitivity and the diffraction grating parameters in a closed-form manner. Experiments on synthetic and real-world data demonstrate that our method outperforms conventional reference target-based methods, underscoring its effectiveness and practicality. + +# 1. Introduction and Preliminaries + +Estimating a camera's spectral sensitivity—its response to each wavelength of incident light—is essential for tasks such as color measurement, illumination and reflectance estimation, and material recognition. Spectral sensitivity is governed by the camera's optical system, including the sensor's color filter array and lens. + +Conventional methods [2, 4, 16, 22] for calibrating spectral sensitivity rely on specialized equipment to emit light with known spectra toward the camera and record the corresponding observed intensities. While accurate, these approaches are costly due to the need for precisely designed optical systems and extensive measurement time. Rather than directly observing light sources with known spectra, prior works [12, 21] have proposed using reference targets with known spectral reflectances under known spectral illuminations, such as sunlight or standard illuminants like D65. While these methods simplify the calibration process, a key challenge is the inherently low-frequency nature of spectral reflectances observed on natural objects [18] and high correlation between the spectral reflectances of indi + +![](images/44fc7825dd7177e6b5c071c764881091c4d5e2f96eaae2ae37270c0f033ed055.jpg) +Figure 1. Overview of the proposed method. By capturing images of light with a known spectrum through an uncalibrated diffraction grating sheet, we observe both the direct and diffracted light. This practical and easy-to-use setup enables accurate estimation of camera spectral sensitivity. + +vidual patches [5], which makes it difficult to accurately determine sensitivity at individual wavelengths. + +To enhance calibration accuracy, the use of diffraction grating sheets [13, 24] has been explored in the past. Diffraction occurs when light waves bend and spread as they pass through narrow apertures, with the degree of bending dependent on wavelength. This effect can be observed using an off-the-shelf diffraction grating sheet, available for less than 5 USD. Each diffracted light beam produces distinct spectral peaks, enabling precise estimation of spectral sensitivities at individual wavelengths. A key challenge in using a diffraction grating sheet is calibrating its diffraction parameter, the grating efficiency that describes the wavelength-dependent attenuation. To calibrate the efficiency, current methods require measurements of a reference target with known reflectances along with the diffracted light beam observation. This involves capturing multiple calibration scenes and complicates the process. + +For an accurate and easy-to-use calibration method, we propose using an uncalibrated diffraction grating sheet without needing additional captures of reference targets. As shown in Fig. 1, the proposed method observes light with known spectra through the diffraction grating sheet, recording both the direct light and the diffraction pattern. + +![](images/af84459a4c8cc3381ae4c902f5d33b07805d1a19403cb2d54fa9937eb1d5a31e.jpg) +Figure 2. Illustration of a diffraction grating. When light of a certain wavelength passes through a diffraction grating, its output intensity is spatially distributed in a characteristic pattern with multiple sharp peaks, known as intensity principal maxima. The pixel positions of the intensity principal maxima shift outward from the center for longer wavelengths. When a broadband light source passes, light beams of different wavelengths appear at distinct shifted positions. In practice, we also need to consider nonuniform transmittance, known as grating efficiency. As a result, the observed intensity at a given pixel position is the product of the incoming light spectrum, the light dispersion function, and the grating efficiency. See Fig. 3 for examples of the observations. + +The proposed method simultaneously estimates the camera spectral sensitivity and grating efficiency. Although the original problem is a bilinear problem where the unknowns appear in their product form, we show that it can be turned into a tractable linear problem by introducing a basis representation for both camera spectral sensitivity and grating efficiency. We further introduce a closed-form solution method to the problem, given the pixel-to-wavelength mapping, i.e., the mapping between the pixel locations in the captured image and the wavelengths of the diffracted light. + +We conduct both synthetic and real-world experiments to demonstrate the effectiveness of the proposed method. The results show that our method outperforms existing leading methods using calibrated reflectance targets [12]. + +# Contributions Our chief contributions are as follows: + +- We propose a camera spectral sensitivity estimation method that uses an uncalibrated diffraction grating sheet, without the need for capturing reference targets. +- The method uses an inexpensive diffraction grating sheet, making it practical, cost-effective, and easy to use. +- Accuracy of the proposed method significantly outperforms those that rely on reference targets, with a simpler setup and fewer captures. + +![](images/318f9121b60f95ab9b44af1aaa6ccf9eb18a33397a84caf6945e082beef49e9b.jpg) +Figure 3. Capturing setup and the observed image of direct and diffracted lights. As the diffracted light exhibits multiple principal maxima, multiple diffraction patterns can be observed. The $k$ -th diffraction image, counted from the center $(k = \pm 1, \pm 2, \ldots)$ , is referred to as the $k$ -th order diffraction image. The image intensity is adjusted for visualization. + +![](images/f7815dc79db77e4b882253d7d919f8530720a9f1c26e2fe53e76f460fa76a2ea.jpg) + +# Preliminaries + +In this section, we provide the foundational background on spectral image formation. We introduce the general spectral image formation model, followed by an overview of how diffraction gratings separate light into its constituent wavelengths. Finally, we present the image formation model for observations captured through a diffraction grating. + +Spectral image formation Given camera spectral sensitivity $s(\lambda_i)$ at discretely sampled wavelengths $\{\lambda_i\}$ and spectrum of the incoming light $e(\lambda_i)$ to the camera, the intensity observation $m \in \mathbb{R}_+$ can be expressed as: + +$$ +m \propto \sum_ {i} e (\lambda_ {i}) s (\lambda_ {i}). \tag {1} +$$ + +Diffraction grating A diffraction grating is an optical device with multiple periodic grooves that separates incoming light into its constituent wavelengths, directing each along a distinct path. Due to interference from the multi-groove structure, each wavelength produces multiple principal intensity maxima, corresponding to different diffraction orders $k = \pm 1, \pm 2, \ldots$ , as illustrated in Fig. 2. This results in multiple dispersed observations symmetrically aligned with the direction of incoming light, with the central observation representing the mixture of all constituent wavelengths—namely, the direct light observation. + +In general, diffraction observation $m_c^{(k)}(p)$ around the $k^{\mathrm{th}}$ intensity principal maxima (i.o.w., $k^{\mathrm{th}}$ order diffraction observation) at pixel location $p$ in channel $c\in \{R,G,B\}$ can be written as + +$$ +m _ {c} ^ {(k)} (p) \propto \sum_ {i} w (\lambda_ {i}; p) e (\lambda_ {i}) \eta_ {k} (\lambda_ {i}) s _ {c} (\lambda_ {i}), \qquad (2) +$$ + +where $s_c$ is the camera spectral sensitivity for channel $c$ , and $\eta_k(\lambda)$ represents the grating efficiency for each wavelength $\lambda$ at $k^{\mathrm{th}}$ order diffraction. The term $w(\lambda; p)$ is the weight induced by the light dispersion function, which disperses the light into each wavelength $\lambda$ that is observed at a + +corresponding peak pixel position $p$ depending on the wavelength. We refer to $w(\lambda ;p)$ as a weight function. + +The weight function $w(\lambda; p)$ approximates a delta function with a sufficient number of slits in the grating, effectively acting as a one-to-one mapping between wavelength $\lambda$ and pixel position $p$ in the diffraction image [13, 24]. Ideally, the grating efficiency $\eta(\lambda)$ should be all 1. However, in practice, the grating efficiency is known to vary depending on the frequency and the material of the grating [17]. + +# 2. Related Work + +In this section, we provide an overview of existing methods for camera spectral sensitivity calibration. A straightforward approach to recovering spectral sensitivity is to measure the observed intensities $m$ in the image while varying the incoming light with known spectra $e(\lambda_i)$ . To achieve such controlled illuminations, specialized light sources [9] and light sources with narrow-band filters [2, 4, 16, 22] have been used. However, the equipment is expensive, limiting their practical applications in real-world settings. + +To eliminate the need for specialized light sources, methods using reference targets, such as color checker charts [1, 7, 8, 11, 15, 19-21, 25], have been proposed. These methods recover camera spectral sensitivity by observing reference targets composed of multiple surface patches with known spectral reflectances under calibrated spectral illumination that uniformly illuminates the entire target. + +To reduce the cost of spectral illumination calibration, several methods [6, 14] assume specific illumination types, such as sunlight [14] or low-frequency spectra [10, 26], and simultaneously estimate camera spectral sensitivity and illumination spectra. While these methods can measure reflected light with different spectra from multiple patches in a single image, they rely on reference targets with known spectral reflectance. Unfortunately, off-the-shelf color checker charts do not provide such data, so one must either rely on externally measured data or perform their own calibration [27]. In addition, due to the low-frequency spectral reflectance of natural objects, the spectral resolution of these methods is limited. To address this challenge, methods using emissive charts [5] have been proposed; however, reference targets employing fluorescent materials are rare and have limited practical applicability. + +Instead of observing reflectances, methods using a diffraction grating, which makes light beams with distinct spectral peaks at different wavelengths, have been proposed [13, 24]. Karge et al. [13] uses a fluorescent lamp, which has a spiky spectrum, to calibrate the pixel-to-wavelength mapping. They then replace the fluorescent lamp with a halogen lamp, which has a flatter spectrum, to observe the diffraction pattern. They additionally need to capture reference targets with known spectral reflectances under controlled spectral illumination to account for the at + +tenuation due to grating efficiency. Toivonen et al. [24] use multiple light sources with different spectra, generated by a halogen lamp equipped with calibrated color filters, to capture the multiple diffraction patterns. In addition, they also record images of a transmissive color chart, comprising multiple patches of calibrated color filters, through different light sources. By recording the diffraction patterns of different light sources and images of the transmissive color chart as input, they achieve accurate calibration of the spectral sensitivity. These methods use inexpensive setups compared to the specialized equipment and achieve more accurate calibration than those relying solely on a color checker chart. However, these approaches require multiple setups and complex capture procedures, which remain a significant practical challenge. + +Unlike the above approaches, Solomatov and Akkaynak [23] use metadata recorded in the Exif, specifically the color matrices that transforms sensor responses to the CIE XYZ color space, to train a neural network for estimating spectral sensitivity. This approach does not require any additional captures for calibration; however, it cannot account for external factors, such as lens and camera filters. Moreover, their estimation exhibits white-balance ambiguity, and recovering the full spectral sensitivity requires white-balance calibration using a color checker chart. + +We use an uncalibrated diffraction grating sheet, where the grating efficiency is unknown. Unlike previous diffraction grating-based methods, our approach only requires capturing a light source through the diffraction grating sheet, providing both the direct light and the diffraction pattern, enabling practical and accurate estimation. Notably, our minimal setting uses one light and two images under the light, and does not require any scene interaction after setup, whereas existing methods require at least two scene changes and three lights with known spectra [13]. + +# 3. Proposed Method + +In this section, we first outline the problem setting for determining the camera spectral sensitivity using a diffraction grating. We then present the proposed method. + +Problem setting Our image formation models of the direct light observation $m_{\mathrm{dir}}$ and the first-order diffracted light observation $m_{\mathrm{dif}}$ are written as: + +$$ +m _ {\mathrm {d i r}} \propto E \sum_ {i} e \left(\lambda_ {i}\right) s \left(\lambda_ {i}\right), \tag {3} +$$ + +$$ +m _ {\mathrm {d i f}} (p) \propto E \sum_ {i} w (\lambda_ {i}; p) e (\lambda_ {i}) \eta (\lambda_ {i}) s (\lambda_ {i}), +$$ + +where $E$ is an unknown scale of the incoming light spectrum and $e(\lambda)$ is a normalized spectrum of the incoming light. We omit the pixel position dependency for the direct light observation $m_{\mathrm{dir}}$ as it remains constant along the + +![](images/fc315c169759536254fe61b419fc8c606d9e569985d78c36ab453ed2baae709d.jpg) +Figure 4. Illustration of the image formation models and overview of our estimation. The observation should vary only along the diffraction direction. To obtain the mean observation, we average along the vertical direction in the figure, resulting in the direct light observation in $\mathbb{R}^3$ and the diffracted light observation in $\mathbb{R}^{n\times 3}$ in RGB channels, respectively. The sampled observation plot in the bottom figure shows a per-channel vector plot along the diffraction direction for each channel. This plot should have the same outline as the wavelength-wise multiplication of the grating efficiency, camera spectral sensitivity, and the incoming light spectrum, leaving the unknown mapping along the horizontal axis, referred to as the pixel-to-wavelength mapping. Using both the direct and diffracted light observation, given the pixel-to-wavelength mapping, we estimate the camera spectral sensitivity and the grating efficiency as described in Sec 3.1. Section 3.2 covers the estimation of the pixel-to-wavelength mapping. + +diffraction (horizontal) direction, unlike the diffracted light. We do not use $k \geq 2$ diffracted light observations in our method because, in practice, the intensity drastically decreases for higher-order diffraction. + +Given a known incoming light spectrum $e(\lambda)$ and by observing both the direct and first-order diffracted lights, $m_{\mathrm{dir}}$ and $m_{\mathrm{dif}}(p)$ , our goal is to estimate the camera's spectral sensitivity $s(\lambda)$ using an uncalibrated diffraction grating, where both the weight function $w(\lambda; p)$ and the grating efficiency $\eta(\lambda)$ are unknown. + +In the following, we first describe the method to jointly estimate the camera spectral sensitivity $s(\lambda)$ and the grating efficiency $\eta(\lambda)$ , assuming that the weight function $w(\lambda; p)$ is known beforehand. We then present the method to determine the weight function $w(\lambda; p)$ . + +# 3.1. Estimating camera spectral sensitivity $s$ and grating efficiency $\eta$ + +We now estimate the camera spectral sensitivity $s(\lambda)$ and the grating efficiency $\eta(\lambda)$ , given the weight function $w(\lambda; p)$ . We show, for the first time, that the problem has a closed-form solution when the unknowns are expressed as linear combinations of their bases. To this end, we first explain the constraints derived from the image formation models of the direct and diffracted light observations, and then present the solution method. + +Let us assume camera spectral sensitivity $\mathbf{s}$ and the inverse of grating efficiency $\eta^{-1}$ at discretely sampled wavelengths $\lambda_1, \ldots, \lambda_f$ can be represented by linear combinations of bases whose numbers of bases are $b_s$ and $b_\eta$ , re + +spectively. Namely, + +$$ +\mathbf {s} \triangleq \left[ \begin{array}{c c c} s (\lambda_ {1}), & \ldots , & s (\lambda_ {f}) \end{array} \right] ^ {\mathsf {T}} = \mathbf {B} _ {s} \mathbf {c} _ {s} \in \mathbb {R} _ {+} ^ {f}, +$$ + +$$ +\boldsymbol {\eta} ^ {- 1} \triangleq \left[ \begin{array}{l l l} \frac {1}{\eta \left(\lambda_ {1}\right)}, & \dots , & \frac {1}{\eta \left(\lambda_ {f}\right)} \end{array} \right] ^ {\intercal} = \mathbf {B} _ {\eta} \mathbf {c} _ {\eta} \in \mathbb {R} _ {+} ^ {f}, \tag {4} +$$ + +where $\mathbf{B}_* \in \mathbb{R}^{f \times b_*}$ is a basis matrix, and $\mathbf{c}_* \in \mathbb{R}^{b_*}$ is a coefficient vector. We obtained the basis matrix $\mathbf{B}_s$ of camera spectral sensitivity $s$ at a channel by singular value decomposition (SVD) on the same set of data used in [23]. For the inverse grating efficiency function $\eta^{-1}$ , we use a Fourier basis representation. Further details can be found in the supplementary material. + +Linear constraint from direct light observation By introducing the basis representation, the aforementioned image formation model of direct light observations in Eq. (3) can be rewritten in a vector form as + +$$ +m _ {\mathrm {d i r}} \propto E \mathbf {e} ^ {\mathsf {T}} \mathbf {s} = E \mathbf {e} ^ {\mathsf {T}} \mathbf {B} _ {s} \mathbf {c} _ {s}, \tag {5} +$$ + +where $\mathbf{e} \in \mathbb{R}_+^f$ is the normalized incoming light spectrum. The light scale $E$ can be set to 1, and the proportionality can be turned into equality without loss of generality, as the global scale can be accounted for in camera spectral sensitivity s. As a result, we have + +$$ +m _ {\mathrm {d i r}} = \underbrace {\mathbf {e} ^ {\mathsf {T}} \mathbf {B} _ {s}} _ {\mathbf {A} _ {\mathrm {d i r}}} \mathbf {c} _ {s}, \tag {6} +$$ + +where $\mathbf{A}_{\mathrm{dir}} \in \mathbb{R}^{1 \times b_s}$ is a known row vector. Similar equations can be derived for all RGB channels to form a vector of direct light observations $\mathbf{m}_{\mathrm{dir}} = [m_{\mathrm{dir}(R)}, m_{\mathrm{dir}(G)}, m_{\mathrm{dir}(B)}]^{\top}$ on the left-hand side + +of the equation, resulting in 3 linear equations with unknown vector $[\mathbf{c}_{s(R)}^{\intercal},\mathbf{c}_{s(G)}^{\intercal},\mathbf{c}_{s(B)}^{\intercal}]^{\intercal}$ , for which $\mathbf{A}_{\mathrm{dir}}$ becomes a $3\times 3b_{s}$ matrix. + +Linear constraint from diffracted light observation Like the direct light observation case, the diffracted light's image formation model in Eq. (3) can be rewritten in a matrix form as + +$$ +\begin{array}{l} \mathbf {m} _ {\text {d i f}} = \left[ m _ {\text {d i f}} \left(p _ {1}\right), \dots , m _ {\text {d i f}} \left(p _ {n}\right) \right] ^ {\top} \tag {7} \\ = \mathbf {W} \operatorname {d i a g} (\mathbf {e}) \operatorname {d i a g} (\boldsymbol {\eta}) \mathbf {s}, \\ \end{array} +$$ + +where $\mathbf{m}_{\mathrm{dif}} \in \mathbb{R}_+^n$ is a stack of the diffracted light observations at pixels $p_1, \ldots, p_n$ . $\mathbf{W} \in \mathbb{R}_+^{n \times f}$ denotes the known weight function for all $n$ pixels, written in a matrix form as + +$$ +\mathbf {W} = \left[ \begin{array}{c c c} w \left(\lambda_ {1}, p _ {1}\right) & \dots & w \left(\lambda_ {f}, p _ {1}\right) \\ \vdots & & \\ w \left(\lambda_ {1}, p _ {n}\right) & \dots & w \left(\lambda_ {f}, p _ {n}\right) \end{array} \right] \in \mathbb {R} _ {+} ^ {n \times f}, \tag {8} +$$ + +which we refer to as the weight matrix. + +The above image formation model of diffracted light can be equivalently turned into + +$$ +\mathbf {s} = \operatorname {d i a g} \left(\boldsymbol {\eta} ^ {- 1}\right) \underbrace {\operatorname {d i a g} \left(\mathbf {e} ^ {- 1}\right) \mathbf {W} ^ {\dagger} \mathbf {m} _ {\mathrm {d i f}}} _ {\mathbf {a}} = \operatorname {d i a g} (\mathbf {a}) \boldsymbol {\eta} ^ {- 1}, \tag {9} +$$ + +where the inverse of a vector denotes an element-wise inverse operator, $\mathbf{W}^{\dagger}$ is a left pseudo-inverse of $\mathbf{W}$ , and $\mathbf{a} \in \mathbb{R}^{f}$ is a known channel-and-light-dependent vector. Since we can confidently assume $n > f$ in our setting, the weight matrix $\mathbf{W}$ should have full column rank when the diffracted observation spans the entire visible wavelength domain, and thus it has a left pseudo-inverse. + +From Eq. (4), Eq. (9) can be written as + +$$ +\mathbf {B} _ {s} \mathbf {c} _ {s} = \operatorname {d i a g} (\mathbf {a}) \mathbf {B} _ {\eta} \mathbf {c} _ {\eta}. \tag {10} +$$ + +Putting together the unknowns $\mathbf{c}_{\eta}$ and $\mathbf{c}_s$ , we have the following homogeneous system of equations: + +$$ +\underbrace {\left[ \begin{array}{l l} \operatorname {d i a g} (\mathbf {a}) \mathbf {B} _ {\eta} & - \mathbf {B} _ {s} \end{array} \right]} _ {\mathbf {A} _ {\mathrm {d i f}}} \left[ \begin{array}{l} \mathbf {c} _ {\eta} \\ \mathbf {c} _ {s} \end{array} \right] = \mathbf {0}, \tag {11} +$$ + +where $\mathbf{A}_{\mathrm{dif}} \in \mathbb{R}^{f \times (b_{\eta} + b_{s})}$ is a known matrix. For the case of RGB images, stacking similar equations from all RGB channels, we obtain $3f$ linear equations from the diffracted light observations for unknown vector $[\mathbf{c}_{\eta}^{\top}, \mathbf{c}_{s(R)}^{\top}, \mathbf{c}_{s(G)}^{\top}, \mathbf{c}_{s(B)}^{\top}]^{\top} \in \mathbb{R}^{b_{\eta} + 3b_{s}}$ , and $\mathbf{A}_{\mathrm{dif}}$ becomes a $3f \times (b_{\eta} + 3b_{s})$ matrix. + +Solution method We can estimate the camera spectral sensitivity $s$ and grating efficiency $\eta$ by solving the following optimization problem using RGB channels: + +$$ +\mathbf {x} ^ {*} = \underset {\mathbf {x}} {\operatorname {a r g m i n}} \| \mathbf {A} _ {\mathrm {d i f}} \mathbf {x} \| _ {2} ^ {2} \text {s . t .} \left[ \begin{array}{l l} \mathbf {0} & \mathbf {A} _ {\mathrm {d i r}} \end{array} \right] \mathbf {x} = \mathbf {m} _ {\mathrm {d i r}}, \tag {12} +$$ + +where $\mathbf{x} = \left[\mathbf{c}_{\eta}^{\mathsf{T}},\mathbf{c}_{s(R)}^{\mathsf{T}},\mathbf{c}_{s(G)}^{\mathsf{T}},\mathbf{c}_{s(B)}^{\mathsf{T}}\right]^{\mathsf{T}}\in \mathbb{R}^{b_{\eta} + 3b_s}$ is the vector of unknowns. The above optimization problem has a closed-form solution by introducing a Lagrange multiplier when the number bases $b_{*}$ and the wavelength resolution $f$ satisfy $3f\geq b_{\eta} + 3b_{s}$ , and $\mathbf{A}_{\mathrm{dif}}$ and $\mathbf{A}_{\mathrm{dir}}$ are full-rank. The details of the solution method are provided in the supplementary material. + +# 3.2. Estimation of weight function $w$ + +As described in Sec. 1, the weight function $w(\lambda; p)$ has a peaky shape using diffraction gratings; therefore, we assume there is a one-to-one mapping between the pixel location $p$ in the diffraction image and the wavelength $\lambda$ as in previous methods [13, 24]. To estimate the weight function $w(\lambda; p)$ , it is necessary to establish the pixel-to-wavelength mapping, namely, relating the pixel locations of the diffraction image to the wavelength as illustrated in Fig. 6. + +We propose to use a quadratic function to represent the one-to-one pixel-to-wavelength mapping $g: p \to \lambda$ as + +$$ +\lambda = g (p) = a p ^ {2} + b p + c, \tag {13} +$$ + +where $a, b, c \in \mathbb{R}$ are the parameters. This quadratic representation allows us stably determining the mapping with retaining accuracy. For further discussion about this parameterization, please refer to our supplementary. + +Once the pixel-to-wavelength mapping $g(p)$ is determined, the elements of the weight matrix $\mathbf{W}$ is simply obtained by linear interpolation as + +$$ +w (\lambda , p) = \left\{ \begin{array}{l l} \frac {\lambda_ {i + 1} - g (p)}{\lambda_ {i + 1} - \lambda_ {i}} & \text {i f} \lambda = \lambda_ {i} \\ \frac {g (p) - \lambda_ {i}}{\lambda_ {i + 1} - \lambda_ {i}} & \text {i f} \lambda = \lambda_ {i + 1}, \\ 0 & \text {o t h e r w i s e} \end{array} \right. \tag {14} +$$ + +where $\lambda_{i}$ is the nearest-neighbor wavelength to $g(p)$ that satisfies $\lambda_{i} \leq g(p) \leq \lambda_{i+1}$ . Therefore, the problem of estimating the weight function boils down to the estimation of pixel-to-wavelength mapping. The challenge, though, is that we only have access to the diffraction image and illumination spectra, but not to camera spectral sensitivity and grating efficiency. Now we describe two settings for determining the pixel-to-wavelength mapping; one using fluorescent and LED lights (denoted as Fluorescent+LED), and another using only LED light (denoted as LED only). + +Fluorescent+LED case When light with spiky spectral peaks, such as fluorescent light, passes through the diffraction grating, the resulting diffracted observation exhibits a similar spiky pattern, as shown in Fig. 5. Therefore, the fluorescent light provides a strong cue for determining the pixel-to-wavelength mapping. In this case, we use a similar + +![](images/95a9c140f1146aa67cb9110d4a2d3be9199b3d69cf7a88772ca29ea55de53de3.jpg) +Figure 5. The illuminant spectrum of the light source used in our real-world experiment and an example of the corresponding direct and diffracted light observations. The observations shown here are adjusted for visualization. + +approach to [13] that associates corresponding peak positions in the pixel and wavelength domains. Since the spectral sensitivity $\mathbf{s}$ is unknown, we weigh the illumination spectra $\mathbf{e}$ by the mean spectral sensitivity $\bar{\mathbf{s}}$ as $\bar{\mathbf{s}}\odot \mathbf{e}$ , where $\bar{\mathbf{s}}$ is obtained from the same dataset used for the camera spectral sensitivity basis [23], and $\odot$ represents an elementwise multiplication operator. We have observed that the unknown grating efficiency $\eta$ is negligible in determining the peak positions due to its low-frequency distribution. + +We select the top two peaks from each of the pixel and wavelength domains for each channel to determine the polynomial parameters $a$ , $b$ , and $c$ of the pixel-to-wavelength mapping $g$ via linear least squares. Once the pixel-to-wavelength mapping $g$ and weight matrix $\mathbf{W}$ are obtained, we take another capture using an LED light for obtaining the direct and diffracted measurements used in Sec. 3.1. + +LED only case To simplify the recording process, it is preferable if we can work only with an LED. In this case, due to the low-frequency incoming spectra without obvious peaks, we cannot take the strategy of relating peak positions used in the Fluorescent+LED case. + +Therefore, instead of finding peak positions, we frame the problem as finding the optimal horizontal mapping (pixel-to-wavelength) between two plots—the normalized diffraction observations $\mathbf{u}$ and the normalized mean camera spectral sensitivity $\mathbf{v}$ that is scaled by illumination spectrum, expressed as: + +$$ +\mathbf {v} \triangleq \nu (\mathbf {e} \odot \bar {\mathbf {s}}) = [ v (\lambda_ {1}), \dots , v (\lambda_ {f}) ] \in \mathbb {R} _ {+} ^ {f} +$$ + +$$ +\mathbf {u} \triangleq \nu (\mathbf {m} _ {\mathrm {d i f}}) = [ u (p _ {1}), \dots , u (p _ {n}) ] \in \mathbb {R} _ {+} ^ {n}, +$$ + +where $\nu$ is a vector normalization as $\nu (\mathbf{x}) = \mathbf{x} / \| \mathbf{x}\| _2$ . + +We solve the point-to-plane iterative closest point (ICP) problem [3] to determine the optimal polynomial parameters $a$ , $b$ , and $c$ along the horizontal axis by minimizing the following objective function: + +$$ +\mathcal {L} _ {\mathrm {I C P}} = \sum_ {i = 1} ^ {n} \sum_ {j \in \mathcal {N} (i)} \left(\left(\left[ \begin{array}{c} a p _ {i} ^ {2} + b p _ {i} + c \\ u (p _ {i}) \end{array} \right] - \left[ \begin{array}{c} \lambda_ {j} \\ v (\lambda_ {j}) \end{array} \right]\right) ^ {\top} \mathbf {n _ {v}} (j)\right) ^ {2}, +$$ + +where $\mathcal{N}(i)$ is the set of indices for the current neighboring points of $[ap_i^2 +bp_i + c,u(p_i)]^\top$ , and $\mathbf{n_v}(j)\in \mathbb{S}\subset \mathbb{R}^2$ + +![](images/83b58e14ecc95c28607285e0bec20125b826946c5fe94919d2acb16d329f433f.jpg) +Figure 6. Illustration of our pixel-to-wavelength mapping estimation. Given the normalized diffraction observations $\mathbf{u}_c$ and normalized illumination-scaled mean camera spectral sensitivity $\mathbf{v}_c$ for all $c \in \{R, G, B\}$ channels, we estimate the optimal pixel-to-wavelength mapping (or horizontal warping, as shown in the figure), which minimizes the distance between the two plots. + +is a unit normal vector to $[\lambda_j,v(\lambda_j)]^\top$ which satisfies + +$$ +\left[ \begin{array}{c} \lambda_ {j} - \lambda_ {j + 1} \\ v (\lambda_ {j}) - v (\lambda_ {j + 1}) \end{array} \right] ^ {\mathsf {T}} \mathbf {n} _ {\mathrm {v}} (j) = 0. \tag {15} +$$ + +We iteratively minimize the objective function for 500 iterations, with the initial guesses set to $a = 0$ , $b = f / n$ , $c = 0$ , where $b$ is given as the ratio of the numbers of elements of $\mathbf{u}$ and $\mathbf{v}$ , to determine the pixel-to-wavelength mapping $g$ . This setting allows for a convenient one-time capture for both determining the pixel-to-wavelength mapping and obtaining the direct and diffracted measurements used in Sec. 3.1. + +# 4. Experiments + +In this section, we evaluate our method on both synthetic and real-world scenes. We first describe our experimental setup and then present the evaluation results. + +Baselines We compare the proposed approach with the Exif-based method (Exif+CC) [23] and the color chart-based method (CC) [12]. We present two variants of the proposed method as detailed in Sec. 3.2: Ours (LED+Flu), which employs a fluorescent lamp observation for pixel-to-wavelength mapping estimation, and Ours (LED), which uses only an LED light observation. For the color chart-based method, we use 2 basis functions, as mentioned in the original paper, which yielded the best results in our preliminary experiments. Since the Exif-based method cannot estimate white balance and introduces a $3 \times 3$ ambiguity, we use the color checker observation to resolve this ambiguity. + +We also compared ours with the existing diffraction grating-based approach by Toivonen et al. [24], adjusted to our experimental setup. However, their method produced + +unstable estimations due to the limited observations available in our setup compared to their original conditions. Further details are provided in the supplemental material. + +Basis representations To compute the basis for camera spectral sensitivity $s(\lambda)$ , we use the dataset from the Exif-based method [23], which includes 44 cameras for training and 5 for testing. Following the detailed analysis on basis selection for a camera spectral sensitivity in [12], we apply SVD to the sensitivities in the training dataset in a per-channel manner. For the grating efficiency, we use a Fourier basis. Further discussion on the grating efficiency basis can be found in the supplementary material. The spectral resolution $f$ in the dataset of camera spectral sensitivities is $f = 31$ from $400\mathrm{nm}$ to $700\mathrm{nm}$ . Assuming the number of basis for the camera spectral sensitivity and the grating efficiency to be the same $b_{\eta} = b_{s}$ , the required condition for the number of bases for our method is $b_{\eta} = b_{s} \leq 23$ based on the discussion in Sec. 3.1. For robust estimation, the number of basis functions is set to 7 for each channel of the camera spectral sensitivity and the grating efficiency across all experiments. + +Evaluation metrics Following [23], we use the average relative full-scale error across the three color channels, RE, which is defined as: + +$$ +\mathrm {R E} = \frac {1}{3} \sum_ {c} \frac {\operatorname {R M S E} (c)}{\max \left(\mathbf {s} _ {c}\right)}, \operatorname {R M S E} (c) = \sqrt {\frac {\left\| \mathbf {s} _ {c} ^ {*} - \mathbf {s} _ {c} \right\| _ {2} ^ {2}}{f}}, \tag {16} +$$ + +where $\mathbf{s}_c^*$ and $\mathbf{s}_c$ denote the estimated and ground-truth camera spectral sensitivity for each channel $c\in \{R,G,B\}$ . + +# 4.1. Evaluation with synthetic scenes + +In this section, we quantitatively evaluate the comparison methods and ours on a synthetic dataset. + +Dataset We render the direct and diffracted images using Eq. (2), with parameters as follows. We use five camera spectral sensitivities: EOS 600D, iPhone12ProMax, Olympus EPL2, Pentax K5, and Sumsung GalaxyS20, all from the same test dataset as the Exif-based method. The D65 illuminant, as defined by the CIE Standard Illuminants, is used as the incoming light. To synthesize the pixel-to-wavelength mapping, we randomly define the mapping while ensuring that the visible wavelength range is fully covered in the diffraction image. The grating efficiency function is set to a low-frequency response as illustrated in Fig. 4, parameterized by 7 Fourier bases. Since the Exif-based method relies on metadata rather than images, we evaluate their method using real-world data. + +Results Figure 7(a) shows the estimated results of the proposed and comparison methods for our synthetic scenes. For ideal observations, both Ours (LED+Flu) and the color chart-based method, CC, yield low estimation errors, as + +both methods are guaranteed to produce optimal results under the basis representation assumption. The remaining errors stem from the gap between the assumption and the synthetic data. The Exif-based method, Exif+CC, produces a generally reasonable overall shape, but it fails to accurately capture finer details. Ours (LED) yields a larger error than Ours (LED+Flu) due to estimation errors in the pixel-to-wavelength mapping when using only LED observations; however, overall, we can still obtain a reasonable estimation even by capturing just a single calibration scene. + +# 4.2. Evaluation with real-world scenes + +Here, we present evaluations using our real-world dataset. + +Dataset 3 illustrates our experimental setup, which consists of a light source, a diffraction grating sheet, and a camera. We capture the images in a dark room for both our method and the comparison methods, ensuring that the only light in the scene comes from illumination with known spectra (LED or fluorescent). We use a 3D-printed slit for the incoming light, positioned in front of the light source, and place a diffraction grating in front of the camera, manually aligned roughly parallel to the image plane. For the spiky light source, we use a Pa-Look Ball (EFD15EL)1 fluorescent lamp. We use the following four cameras: EOS RP, iPhone 15ProMax, Sony $\alpha$ 1, and DJI Pocket3. For the non-spiky light sources and diffraction sheets, we use three different products for each. In this section, we present the results for one setup. Additional experiments with other setups are provided in the supplementary material. + +Due to the high dynamic range of the direct and diffracted light, we capture HDR images by taking two images with different exposure times for each. The extra diffracted image under fluorescent light is captured for LED+Flu, solely for pixel-to-wavelength mapping. After capturing the direct and diffracted light observations, we roughly crop them with a sufficient margin at both ends of the diffraction pattern along the diffraction direction. The diffracted light observation is cropped to approximately $200 \times 1000$ pix and aggregated to $1 \times 1000$ pix. + +For the comparison methods, we use a 24-patch color checker chart with spectral reflectance data from Chromaxion. The color checker chart is captured under the same non-spiky light source used in our experiments. We carefully position the color checker chart so that it is parallel to the camera and uniformly illuminated by the light source, ensuring that no shadows are cast. The metadata for the Exif-based method is extracted from our captured data following the procedure described in their paper [23]. + +![](images/248692476abfeef63fd2fccf00e5837067470deea016905411937e1014429e3c.jpg) +(a) Synthetic experiment + +![](images/af89459e082555441ea3e987d5cbb408993d84a1567429bdae276bce4bd19096.jpg) +(b) Real-world experiment +Figure 7. Results of the proposed and comparison methods for our (a) synthetic and (b) real-world datasets. For each setting and method, we show plots of the estimated (solid line) and ground-truth (dotted line) spectral sensitivity for each color channel. All plots are normalized by the maximum response across all channels. The values below the plots indicate the RE defined in Eq. (16). For simplicity, the factor of $10^{-2}$ is omitted. Underscored bold and underscored score are the best and second-best result for each setting, respectively. + +The ground-truth spectral sensitivities are obtained using a calibration device with a lens attached. The illuminant spectra of the light sources, shown in Fig. 5, are measured through the diffraction grating using a spectrometer. + +Results Figure 7(b) presents the estimated results of the proposed and comparison methods. Overall, the proposed methods achieve the lowest estimation errors for most scenes, and Ours (LED) produces results nearly comparable to Ours (LED+Flu). For the EOS RP scene, Ours (LED) yields a larger error due to the failure of pixel-to-wavelength mapping estimation using a non-spiky light source. The color checker-based method, CC, is relatively sensitive to real-world noise due to the low dimensionality of the spectral reflectances in the color patches, as discussed in previous works [5, 24]. While the Exif-based method, Exif+CC, has an advantage of being easy to use, as it only relies on metadata for recovering sensitivity, it yields larger estimation errors across all scenes. + +# 5. Conclusions + +We present a method for estimating camera spectral sensitivity using an uncalibrated diffraction grating sheet. Compared to previous methods that use reference targets, such as a color checker chart, our approach achieves highly accurate estimation by leveraging the diffraction grating. Unlike previous diffraction grating-based methods, the proposed method jointly estimates both camera spectral sensitivity and grating efficiency without needing additional recordings, making it practical, cost-efficient, and easy-to-use. + +One of the limitations of our method is that pixel-to-wavelength mapping estimation breaks down if the illumination has uniform spectra. In this case, the diffraction image simply shows the camera spectral sensitivity scaled by unknown grating efficiency, providing no clues for the mapping estimation. However, in practice, readily-available light sources do not have such uniform spectra. LEDs, for example, tend to have a non-uniform spectral distribution, enabling us to estimate the mapping and camera spectral sensitivities, as demonstrated in our real-world experiments. + +# Acknowledgments + +This work was supported by JSPS KAKENHI Grant Numbers JP23H05491 and JP25K00142, JST ASPIRE Grant Number JPMJAP2304, the Canada First Research Excellence Fund program (VISTA), an NSERC Discovery Grant, and the Canada Research Chair program. + +# References + +[1] Ali Alsam and Graham Finlayson. Recovering spectral sensitivities with uncertainty. In Proceedings of Conference on Colour in Graphics, Imaging, and Vision (CGIV), pages 22-26, 2002. 3 +[2] E Berra, S Gibson-Poole, A MacArthur, R Gaulton, and A Hamilton. Estimation of the spectral sensitivity functions of un-modified and modified commercial off-the-shelf digital cameras to enable their use as a multispectral imaging system for UAVs. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 40: 207-214, 2015. 1, 3 +[3] Yang Chen and Gérard Medioni. Object modelling by registration of multiple range images. Image and Vision Computing, 10(3):145-155, 1992. 6 +[4] Maryam Mohammadzadeh Darrodi, Graham Finlayson, Teresa Goodman, and Michal Mackiewicz. Reference data set for camera spectral sensitivity estimation. Journal of the Optical Society of America A, 32(3):381-391, 2015. 1, 3 +[5] Jeffrey M DiCarlo, Glen Eric Montgomery, and Steven W Trovinger. Emissive chart for imager calibration. In Proceedings of IS&T Color and Imaging Conference (CIC), pages 295-301, 2004. 1, 3, 8 +[6] Bob Dyas. Robust color sensor response characterization. In Proceedings of IS&T Color and Imaging Conference (CIC), pages 144-148, 2000. 3 +[7] Marc Ebner. Estimating the spectral sensitivity of a digital sensor using calibration targets. In Proceedings of Annual Conference on Genetic and Evolutionary Computation, pages 642-649, 2007. 3 +[8] Graham Finlayson, Maryam Mohammadzadeh Darrodi, and Michal Mackiewicz. Rank-based camera spectral sensitivity estimation. 33(4):589-599, 2016. 3 +[9] Paul M Hubel, Doron Sherman, and Joyce E Farrell. A comparison of methods of sensor spectral sensitivity estimation. In Proceedings of IS&T Color and Imaging Conference (CIC), pages 45-48, 1994. 3 +[10] Cong Phuoc Huynh and Antonio Robles-Kelly. Recovery of spectral sensitivity functions from a colour chart image under unknown spectrally smooth illumination. In Proceedings of IEEE International Conference on Pattern Recognition (ICPR), pages 708-713, 2014. 3 +[11] Yuhyun Ji, Yunsang Kwak, Sang Mok Park, and Young L Kim. Compressive recovery of smartphone RGB spectral sensitivity functions. Optics Express, 29(8):11947-11961, 2021. 3 +[12] Jun Jiang, Dengyu Liu, Jinwei Gu, and Sabine Süssstrunk. What is the space of spectral sensitivity functions for digital + +color cameras? In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 168-179, 2013. 1, 2, 6, 7, 8 +[13] Andreas Karge, Ingmar Rieger, Bernhard Eberhardt, and Andreas Schilling. Using chromaticity error minimisation for fast camera spectral responsivity measurement. In Proceedings of IS&T Color and Imaging Conference (CIC), pages 67-74, 2018. 1, 3, 5, 6 +[14] Rei Kawakami, Hongxun Zhao, Robby T Tan, and Katsushi Ikeuchi. Camera spectral sensitivity and white balance estimation from sky images. International Journal of Computer Vision (IJCV), 105:187-204, 2013. 3 +[15] Berk Kaya, Yigit Baran Can, and Radu Timofte. Towards spectral estimation from a single RGB image in the wild. In Proceedings of IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 3546-3555, 2019. 3 +[16] Lindsay W MacDonald. Determining camera spectral responsivity with multispectral transmission filters. In Proceedings of IS&T Color and Imaging Conference (CIC), pages 12-17, 2015. 1, 3 +[17] Christopher Palmer and Erwin G Loewen. Diffraction Grating Handbook. Newport Corporation New York, 2005. 3 +[18] J. P. S. Parkkinen, J. Hallikainen, and T. Jaaskelainen. Characteristic spectra of munsell colors. Journal of the Optical Society of America A, 6(2):318-322, 1989. 1 +[19] Dilip Prasad, Rang Nguyen, and Michael Brown. Quick approximation of camera's spectral response from casual lighting. In Proceedings of IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 844-851, 2013. 3 +[20] Shuxue Quan, Noboru Ohta, and Xiaoyun Jiang. Comparative study on sensor spectral sensitivity estimation. In Proceedings of SPIE Color Imaging VIII: Processing, Hardcopy, and Applications, pages 209-220, 2003. +[21] Martin Rump, Arno Zinke, and Reinhard Klein. Practical spectral characterization of trichromatic cameras. ACM Transactions on Graphics (TOG), 30(6):170, 2011. 1, 3 +[22] Yuri Rzhanov. Accurate determination of camera quantum efficiency from a single image. Journal of Imaging, 10(7): 169, 2024. 1, 3 +[23] Grigory Solomatov and Derya Akkaynak. Spectral sensitivity estimation without a camera. In International Conference on Computational Photography (ICCP), pages 1-12, 2023. 3, 4, 6, 7, 8 +[24] Mikko E Toivonen and Arto Klami. Practical camera sensor spectral response and uncertainty estimation. Journal of Imaging, 6(8):79, 2020. 1, 3, 5, 6, 8 +[25] Shoji Tominaga, Shogo Nishi, and Ryo Ohtera. Measurement and estimation of spectral sensitivity functions for mobile phone cameras. Sensors, 21(15):4985, 2021. 3 +[26] Lin Zhang, Ying Fu, Yinqiang Zheng, and Hua Huang. Camera spectral sensitivity, illumination and spectral reflectance estimation for a hybrid hyperspectral image capture system. In Proceedings of IEEE International Conference on Image Processing (ICIP), pages 545-549, 2017. 3 + +[27] Jieyu Zhu, Xufen Xie, Ningfang Liao, Zhuangzhi Zhang, Wenmin Wu, and Lvming Lv. Spectral sensitivity estimation of trichromatic camera based on orthogonal test and window filtering. Optics Express, 28(19):28085-28100, 2020. 3 \ No newline at end of file diff --git a/2025/Spectral Sensitivity Estimation with an Uncalibrated Diffraction Grating/images.zip b/2025/Spectral Sensitivity Estimation with an Uncalibrated Diffraction Grating/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ebd1be2f850f2ef547bfb6e7ca352c2274787fc6 --- /dev/null +++ b/2025/Spectral Sensitivity Estimation with an Uncalibrated Diffraction Grating/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3313935d73cad829ba03af46944035808591fe65d8a3bdb609297f9af524946c +size 576263 diff --git a/2025/Spectral Sensitivity Estimation with an Uncalibrated Diffraction Grating/layout.json b/2025/Spectral Sensitivity Estimation with an Uncalibrated Diffraction Grating/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c99c603787a27b9489c3f74e1d51fcf7685342d8 --- /dev/null +++ b/2025/Spectral Sensitivity Estimation with an Uncalibrated Diffraction Grating/layout.json @@ -0,0 +1,9432 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 80, + 103, + 531, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 103, + 531, + 121 + ], + "spans": [ + { + "bbox": [ + 80, + 103, + 531, + 121 + ], + "type": "text", + "content": "Spectral Sensitivity Estimation with an Uncalibrated Diffraction Grating" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 65, + 142, + 544, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 142, + 544, + 171 + ], + "spans": [ + { + "bbox": [ + 65, + 142, + 544, + 171 + ], + "type": "text", + "content": "Lilika Makabe1 Hiroaki Santo1 Fumio Okura1 Michael S. Brown2 Yasuyuki Matsushita1,3 \n1The University of Osaka 2York University 3Microsoft Research Asia - Tokyo" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 91, + 173, + 516, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 173, + 516, + 186 + ], + "spans": [ + { + "bbox": [ + 91, + 173, + 516, + 186 + ], + "type": "text", + "content": "{makabe.lilika, santo.hiroaki, okura, yasumat}@ist.osaka-u.ac.jp mbrown@yorku.ca" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 238, + 297, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 297, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 297, + 430 + ], + "type": "text", + "content": "This paper introduces a practical and accurate calibration method for camera spectral sensitivity using a diffraction grating. Accurate calibration of camera spectral sensitivity is crucial for various computer vision tasks, including color correction, illumination estimation, and material analysis. Unlike existing approaches that require specialized narrow-band filters or reference targets with known spectral reflectances, our method only requires an uncalibrated diffraction grating sheet, readily available off-the-shelf. By capturing images of the direct illumination and its diffracted pattern through the grating sheet, our method estimates both the camera spectral sensitivity and the diffraction grating parameters in a closed-form manner. Experiments on synthetic and real-world data demonstrate that our method outperforms conventional reference target-based methods, underscoring its effectiveness and practicality." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 453, + 230, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 230, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 230, + 466 + ], + "type": "text", + "content": "1. Introduction and Preliminaries" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 474, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 544 + ], + "type": "text", + "content": "Estimating a camera's spectral sensitivity—its response to each wavelength of incident light—is essential for tasks such as color measurement, illumination and reflectance estimation, and material recognition. Spectral sensitivity is governed by the camera's optical system, including the sensor's color filter array and lens." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "Conventional methods [2, 4, 16, 22] for calibrating spectral sensitivity rely on specialized equipment to emit light with known spectra toward the camera and record the corresponding observed intensities. While accurate, these approaches are costly due to the need for precisely designed optical systems and extensive measurement time. Rather than directly observing light sources with known spectra, prior works [12, 21] have proposed using reference targets with known spectral reflectances under known spectral illuminations, such as sunlight or standard illuminants like D65. While these methods simplify the calibration process, a key challenge is the inherently low-frequency nature of spectral reflectances observed on natural objects [18] and high correlation between the spectral reflectances of indi" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 213, + 555, + 345 + ], + "blocks": [ + { + "bbox": [ + 317, + 213, + 555, + 345 + ], + "lines": [ + { + "bbox": [ + 317, + 213, + 555, + 345 + ], + "spans": [ + { + "bbox": [ + 317, + 213, + 555, + 345 + ], + "type": "image", + "image_path": "44fc7825dd7177e6b5c071c764881091c4d5e2f96eaae2ae37270c0f033ed055.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 354, + 555, + 410 + ], + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 410 + ], + "type": "text", + "content": "Figure 1. Overview of the proposed method. By capturing images of light with a known spectrum through an uncalibrated diffraction grating sheet, we observe both the direct and diffracted light. This practical and easy-to-use setup enables accurate estimation of camera spectral sensitivity." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 424, + 554, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 554, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 554, + 448 + ], + "type": "text", + "content": "vidual patches [5], which makes it difficult to accurately determine sensitivity at individual wavelengths." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 449, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 449, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 312, + 449, + 555, + 640 + ], + "type": "text", + "content": "To enhance calibration accuracy, the use of diffraction grating sheets [13, 24] has been explored in the past. Diffraction occurs when light waves bend and spread as they pass through narrow apertures, with the degree of bending dependent on wavelength. This effect can be observed using an off-the-shelf diffraction grating sheet, available for less than 5 USD. Each diffracted light beam produces distinct spectral peaks, enabling precise estimation of spectral sensitivities at individual wavelengths. A key challenge in using a diffraction grating sheet is calibrating its diffraction parameter, the grating efficiency that describes the wavelength-dependent attenuation. To calibrate the efficiency, current methods require measurements of a reference target with known reflectances along with the diffracted light beam observation. This involves capturing multiple calibration scenes and complicates the process." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "For an accurate and easy-to-use calibration method, we propose using an uncalibrated diffraction grating sheet without needing additional captures of reference targets. As shown in Fig. 1, the proposed method observes light with known spectra through the diffraction grating sheet, recording both the direct light and the diffraction pattern." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27252" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 62, + 294, + 265 + ], + "blocks": [ + { + "bbox": [ + 59, + 62, + 294, + 265 + ], + "lines": [ + { + "bbox": [ + 59, + 62, + 294, + 265 + ], + "spans": [ + { + "bbox": [ + 59, + 62, + 294, + 265 + ], + "type": "image", + "image_path": "af84459a4c8cc3381ae4c902f5d33b07805d1a19403cb2d54fa9937eb1d5a31e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 268, + 296, + 401 + ], + "lines": [ + { + "bbox": [ + 55, + 268, + 296, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 268, + 296, + 401 + ], + "type": "text", + "content": "Figure 2. Illustration of a diffraction grating. When light of a certain wavelength passes through a diffraction grating, its output intensity is spatially distributed in a characteristic pattern with multiple sharp peaks, known as intensity principal maxima. The pixel positions of the intensity principal maxima shift outward from the center for longer wavelengths. When a broadband light source passes, light beams of different wavelengths appear at distinct shifted positions. In practice, we also need to consider nonuniform transmittance, known as grating efficiency. As a result, the observed intensity at a given pixel position is the product of the incoming light spectrum, the light dispersion function, and the grating efficiency. See Fig. 3 for examples of the observations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 419, + 296, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 296, + 540 + ], + "type": "text", + "content": "The proposed method simultaneously estimates the camera spectral sensitivity and grating efficiency. Although the original problem is a bilinear problem where the unknowns appear in their product form, we show that it can be turned into a tractable linear problem by introducing a basis representation for both camera spectral sensitivity and grating efficiency. We further introduce a closed-form solution method to the problem, given the pixel-to-wavelength mapping, i.e., the mapping between the pixel locations in the captured image and the wavelengths of the diffracted light." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 540, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 296, + 590 + ], + "type": "text", + "content": "We conduct both synthetic and real-world experiments to demonstrate the effectiveness of the proposed method. The results show that our method outperforms existing leading methods using calibrated reflectance targets [12]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 601, + 282, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 282, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 282, + 613 + ], + "type": "text", + "content": "Contributions Our chief contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 616, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 616, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 616, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 56, + 616, + 295, + 651 + ], + "type": "text", + "content": "- We propose a camera spectral sensitivity estimation method that uses an uncalibrated diffraction grating sheet, without the need for capturing reference targets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 651, + 294, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 651, + 294, + 675 + ], + "spans": [ + { + "bbox": [ + 56, + 651, + 294, + 675 + ], + "type": "text", + "content": "- The method uses an inexpensive diffraction grating sheet, making it practical, cost-effective, and easy to use." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 675, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 675, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 675, + 295, + 712 + ], + "type": "text", + "content": "- Accuracy of the proposed method significantly outperforms those that rely on reference targets, with a simpler setup and fewer captures." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 316, + 72, + 427, + 144 + ], + "blocks": [ + { + "bbox": [ + 316, + 72, + 427, + 144 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 427, + 144 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 427, + 144 + ], + "type": "image", + "image_path": "318f9121b60f95ab9b44af1aaa6ccf9eb18a33397a84caf6945e082beef49e9b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 153, + 555, + 220 + ], + "lines": [ + { + "bbox": [ + 313, + 153, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 153, + 555, + 220 + ], + "type": "text", + "content": "Figure 3. Capturing setup and the observed image of direct and diffracted lights. As the diffracted light exhibits multiple principal maxima, multiple diffraction patterns can be observed. The " + }, + { + "bbox": [ + 313, + 153, + 555, + 220 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 153, + 555, + 220 + ], + "type": "text", + "content": "-th diffraction image, counted from the center " + }, + { + "bbox": [ + 313, + 153, + 555, + 220 + ], + "type": "inline_equation", + "content": "(k = \\pm 1, \\pm 2, \\ldots)" + }, + { + "bbox": [ + 313, + 153, + 555, + 220 + ], + "type": "text", + "content": ", is referred to as the " + }, + { + "bbox": [ + 313, + 153, + 555, + 220 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 153, + 555, + 220 + ], + "type": "text", + "content": "-th order diffraction image. The image intensity is adjusted for visualization." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 427, + 72, + 553, + 144 + ], + "blocks": [ + { + "bbox": [ + 427, + 72, + 553, + 144 + ], + "lines": [ + { + "bbox": [ + 427, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 427, + 72, + 553, + 144 + ], + "type": "image", + "image_path": "f7815dc79db77e4b882253d7d919f8530720a9f1c26e2fe53e76f460fa76a2ea.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 239, + 388, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 239, + 388, + 251 + ], + "spans": [ + { + "bbox": [ + 314, + 239, + 388, + 251 + ], + "type": "text", + "content": "Preliminaries" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 259, + 555, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 259, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 555, + 332 + ], + "type": "text", + "content": "In this section, we provide the foundational background on spectral image formation. We introduce the general spectral image formation model, followed by an overview of how diffraction gratings separate light into its constituent wavelengths. Finally, we present the image formation model for observations captured through a diffraction grating." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 336, + 554, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 554, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 554, + 384 + ], + "type": "text", + "content": "Spectral image formation Given camera spectral sensitivity " + }, + { + "bbox": [ + 313, + 336, + 554, + 384 + ], + "type": "inline_equation", + "content": "s(\\lambda_i)" + }, + { + "bbox": [ + 313, + 336, + 554, + 384 + ], + "type": "text", + "content": " at discretely sampled wavelengths " + }, + { + "bbox": [ + 313, + 336, + 554, + 384 + ], + "type": "inline_equation", + "content": "\\{\\lambda_i\\}" + }, + { + "bbox": [ + 313, + 336, + 554, + 384 + ], + "type": "text", + "content": " and spectrum of the incoming light " + }, + { + "bbox": [ + 313, + 336, + 554, + 384 + ], + "type": "inline_equation", + "content": "e(\\lambda_i)" + }, + { + "bbox": [ + 313, + 336, + 554, + 384 + ], + "type": "text", + "content": " to the camera, the intensity observation " + }, + { + "bbox": [ + 313, + 336, + 554, + 384 + ], + "type": "inline_equation", + "content": "m \\in \\mathbb{R}_+" + }, + { + "bbox": [ + 313, + 336, + 554, + 384 + ], + "type": "text", + "content": " can be expressed as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 390, + 393, + 555, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 393, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 390, + 393, + 555, + 418 + ], + "type": "interline_equation", + "content": "m \\propto \\sum_ {i} e (\\lambda_ {i}) s (\\lambda_ {i}). \\tag {1}", + "image_path": "16b34ad892751d604dddf72f37d0c150df07020be77b5259c0ff01ffc31c7c25.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 431, + 555, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 555, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 555, + 563 + ], + "type": "text", + "content": "Diffraction grating A diffraction grating is an optical device with multiple periodic grooves that separates incoming light into its constituent wavelengths, directing each along a distinct path. Due to interference from the multi-groove structure, each wavelength produces multiple principal intensity maxima, corresponding to different diffraction orders " + }, + { + "bbox": [ + 313, + 431, + 555, + 563 + ], + "type": "inline_equation", + "content": "k = \\pm 1, \\pm 2, \\ldots" + }, + { + "bbox": [ + 313, + 431, + 555, + 563 + ], + "type": "text", + "content": ", as illustrated in Fig. 2. This results in multiple dispersed observations symmetrically aligned with the direction of incoming light, with the central observation representing the mixture of all constituent wavelengths—namely, the direct light observation." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 563, + 554, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 554, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 554, + 611 + ], + "type": "text", + "content": "In general, diffraction observation " + }, + { + "bbox": [ + 313, + 563, + 554, + 611 + ], + "type": "inline_equation", + "content": "m_c^{(k)}(p)" + }, + { + "bbox": [ + 313, + 563, + 554, + 611 + ], + "type": "text", + "content": " around the " + }, + { + "bbox": [ + 313, + 563, + 554, + 611 + ], + "type": "inline_equation", + "content": "k^{\\mathrm{th}}" + }, + { + "bbox": [ + 313, + 563, + 554, + 611 + ], + "type": "text", + "content": " intensity principal maxima (i.o.w., " + }, + { + "bbox": [ + 313, + 563, + 554, + 611 + ], + "type": "inline_equation", + "content": "k^{\\mathrm{th}}" + }, + { + "bbox": [ + 313, + 563, + 554, + 611 + ], + "type": "text", + "content": " order diffraction observation) at pixel location " + }, + { + "bbox": [ + 313, + 563, + 554, + 611 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 563, + 554, + 611 + ], + "type": "text", + "content": " in channel " + }, + { + "bbox": [ + 313, + 563, + 554, + 611 + ], + "type": "inline_equation", + "content": "c\\in \\{R,G,B\\}" + }, + { + "bbox": [ + 313, + 563, + 554, + 611 + ], + "type": "text", + "content": " can be written as" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 345, + 620, + 555, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 620, + 555, + 644 + ], + "spans": [ + { + "bbox": [ + 345, + 620, + 555, + 644 + ], + "type": "interline_equation", + "content": "m _ {c} ^ {(k)} (p) \\propto \\sum_ {i} w (\\lambda_ {i}; p) e (\\lambda_ {i}) \\eta_ {k} (\\lambda_ {i}) s _ {c} (\\lambda_ {i}), \\qquad (2)", + "image_path": "9d82af6878b10e3c998b93df9a502963f9bd4bcd7fdabc4285061f82271138d6.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "inline_equation", + "content": "s_c" + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "content": " is the camera spectral sensitivity for channel " + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\eta_k(\\lambda)" + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "content": " represents the grating efficiency for each wavelength " + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "inline_equation", + "content": "k^{\\mathrm{th}}" + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "content": " order diffraction. The term " + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "inline_equation", + "content": "w(\\lambda; p)" + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "content": " is the weight induced by the light dispersion function, which disperses the light into each wavelength " + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "content": " that is observed at a" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27253" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "corresponding peak pixel position " + }, + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": " depending on the wavelength. We refer to " + }, + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "inline_equation", + "content": "w(\\lambda ;p)" + }, + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": " as a weight function." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "content": "The weight function " + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "inline_equation", + "content": "w(\\lambda; p)" + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "content": " approximates a delta function with a sufficient number of slits in the grating, effectively acting as a one-to-one mapping between wavelength " + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "content": " and pixel position " + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "content": " in the diffraction image [13, 24]. Ideally, the grating efficiency " + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\eta(\\lambda)" + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "content": " should be all 1. However, in practice, the grating efficiency is known to vary depending on the frequency and the material of the grating [17]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 190, + 142, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 142, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 142, + 204 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 211, + 295, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 211, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 295, + 319 + ], + "type": "text", + "content": "In this section, we provide an overview of existing methods for camera spectral sensitivity calibration. A straightforward approach to recovering spectral sensitivity is to measure the observed intensities " + }, + { + "bbox": [ + 55, + 211, + 295, + 319 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 211, + 295, + 319 + ], + "type": "text", + "content": " in the image while varying the incoming light with known spectra " + }, + { + "bbox": [ + 55, + 211, + 295, + 319 + ], + "type": "inline_equation", + "content": "e(\\lambda_i)" + }, + { + "bbox": [ + 55, + 211, + 295, + 319 + ], + "type": "text", + "content": ". To achieve such controlled illuminations, specialized light sources [9] and light sources with narrow-band filters [2, 4, 16, 22] have been used. However, the equipment is expensive, limiting their practical applications in real-world settings." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 319, + 296, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 402 + ], + "type": "text", + "content": "To eliminate the need for specialized light sources, methods using reference targets, such as color checker charts [1, 7, 8, 11, 15, 19-21, 25], have been proposed. These methods recover camera spectral sensitivity by observing reference targets composed of multiple surface patches with known spectral reflectances under calibrated spectral illumination that uniformly illuminates the entire target." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 403, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 295, + 594 + ], + "type": "text", + "content": "To reduce the cost of spectral illumination calibration, several methods [6, 14] assume specific illumination types, such as sunlight [14] or low-frequency spectra [10, 26], and simultaneously estimate camera spectral sensitivity and illumination spectra. While these methods can measure reflected light with different spectra from multiple patches in a single image, they rely on reference targets with known spectral reflectance. Unfortunately, off-the-shelf color checker charts do not provide such data, so one must either rely on externally measured data or perform their own calibration [27]. In addition, due to the low-frequency spectral reflectance of natural objects, the spectral resolution of these methods is limited. To address this challenge, methods using emissive charts [5] have been proposed; however, reference targets employing fluorescent materials are rare and have limited practical applicability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "Instead of observing reflectances, methods using a diffraction grating, which makes light beams with distinct spectral peaks at different wavelengths, have been proposed [13, 24]. Karge et al. [13] uses a fluorescent lamp, which has a spiky spectrum, to calibrate the pixel-to-wavelength mapping. They then replace the fluorescent lamp with a halogen lamp, which has a flatter spectrum, to observe the diffraction pattern. They additionally need to capture reference targets with known spectral reflectances under controlled spectral illumination to account for the at" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": "tenuation due to grating efficiency. Toivonen et al. [24] use multiple light sources with different spectra, generated by a halogen lamp equipped with calibrated color filters, to capture the multiple diffraction patterns. In addition, they also record images of a transmissive color chart, comprising multiple patches of calibrated color filters, through different light sources. By recording the diffraction patterns of different light sources and images of the transmissive color chart as input, they achieve accurate calibration of the spectral sensitivity. These methods use inexpensive setups compared to the specialized equipment and achieve more accurate calibration than those relying solely on a color checker chart. However, these approaches require multiple setups and complex capture procedures, which remain a significant practical challenge." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 251, + 553, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 553, + 371 + ], + "type": "text", + "content": "Unlike the above approaches, Solomatov and Akkaynak [23] use metadata recorded in the Exif, specifically the color matrices that transforms sensor responses to the CIE XYZ color space, to train a neural network for estimating spectral sensitivity. This approach does not require any additional captures for calibration; however, it cannot account for external factors, such as lens and camera filters. Moreover, their estimation exhibits white-balance ambiguity, and recovering the full spectral sensitivity requires white-balance calibration using a color checker chart." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 372, + 554, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 554, + 491 + ], + "type": "text", + "content": "We use an uncalibrated diffraction grating sheet, where the grating efficiency is unknown. Unlike previous diffraction grating-based methods, our approach only requires capturing a light source through the diffraction grating sheet, providing both the direct light and the diffraction pattern, enabling practical and accurate estimation. Notably, our minimal setting uses one light and two images under the light, and does not require any scene interaction after setup, whereas existing methods require at least two scene changes and three lights with known spectra [13]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 502, + 421, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 502, + 421, + 514 + ], + "spans": [ + { + "bbox": [ + 314, + 502, + 421, + 514 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 521, + 554, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 558 + ], + "type": "text", + "content": "In this section, we first outline the problem setting for determining the camera spectral sensitivity using a diffraction grating. We then present the proposed method." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 560, + 554, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 560, + 554, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 560, + 554, + 597 + ], + "type": "text", + "content": "Problem setting Our image formation models of the direct light observation " + }, + { + "bbox": [ + 313, + 560, + 554, + 597 + ], + "type": "inline_equation", + "content": "m_{\\mathrm{dir}}" + }, + { + "bbox": [ + 313, + 560, + 554, + 597 + ], + "type": "text", + "content": " and the first-order diffracted light observation " + }, + { + "bbox": [ + 313, + 560, + 554, + 597 + ], + "type": "inline_equation", + "content": "m_{\\mathrm{dif}}" + }, + { + "bbox": [ + 313, + 560, + 554, + 597 + ], + "type": "text", + "content": " are written as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 358, + 605, + 553, + 636 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 605, + 553, + 636 + ], + "spans": [ + { + "bbox": [ + 358, + 605, + 553, + 636 + ], + "type": "interline_equation", + "content": "m _ {\\mathrm {d i r}} \\propto E \\sum_ {i} e \\left(\\lambda_ {i}\\right) s \\left(\\lambda_ {i}\\right), \\tag {3}", + "image_path": "b79cc2b1d1bb4d306a0b518b4193c6c8c6859e5ff22132d874f5efb6774a7ba9.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 347, + 633, + 522, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 633, + 522, + 657 + ], + "spans": [ + { + "bbox": [ + 347, + 633, + 522, + 657 + ], + "type": "interline_equation", + "content": "m _ {\\mathrm {d i f}} (p) \\propto E \\sum_ {i} w (\\lambda_ {i}; p) e (\\lambda_ {i}) \\eta (\\lambda_ {i}) s (\\lambda_ {i}),", + "image_path": "cdb2d61ea6b518a455d4ac9ba9634a500a80fca3f2849111335110b1a033b979.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": " is an unknown scale of the incoming light spectrum and " + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "inline_equation", + "content": "e(\\lambda)" + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": " is a normalized spectrum of the incoming light. We omit the pixel position dependency for the direct light observation " + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "inline_equation", + "content": "m_{\\mathrm{dir}}" + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": " as it remains constant along the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27254" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 62, + 553, + 202 + ], + "blocks": [ + { + "bbox": [ + 58, + 62, + 553, + 202 + ], + "lines": [ + { + "bbox": [ + 58, + 62, + 553, + 202 + ], + "spans": [ + { + "bbox": [ + 58, + 62, + 553, + 202 + ], + "type": "image", + "image_path": "fc315c169759536254fe61b419fc8c606d9e569985d78c36ab453ed2baae709d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 213, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 54, + 213, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 54, + 213, + 555, + 301 + ], + "type": "text", + "content": "Figure 4. Illustration of the image formation models and overview of our estimation. The observation should vary only along the diffraction direction. To obtain the mean observation, we average along the vertical direction in the figure, resulting in the direct light observation in " + }, + { + "bbox": [ + 54, + 213, + 555, + 301 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^3" + }, + { + "bbox": [ + 54, + 213, + 555, + 301 + ], + "type": "text", + "content": " and the diffracted light observation in " + }, + { + "bbox": [ + 54, + 213, + 555, + 301 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{n\\times 3}" + }, + { + "bbox": [ + 54, + 213, + 555, + 301 + ], + "type": "text", + "content": " in RGB channels, respectively. The sampled observation plot in the bottom figure shows a per-channel vector plot along the diffraction direction for each channel. This plot should have the same outline as the wavelength-wise multiplication of the grating efficiency, camera spectral sensitivity, and the incoming light spectrum, leaving the unknown mapping along the horizontal axis, referred to as the pixel-to-wavelength mapping. Using both the direct and diffracted light observation, given the pixel-to-wavelength mapping, we estimate the camera spectral sensitivity and the grating efficiency as described in Sec 3.1. Section 3.2 covers the estimation of the pixel-to-wavelength mapping." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 322, + 296, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 322, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 296, + 369 + ], + "type": "text", + "content": "diffraction (horizontal) direction, unlike the diffracted light. We do not use " + }, + { + "bbox": [ + 55, + 322, + 296, + 369 + ], + "type": "inline_equation", + "content": "k \\geq 2" + }, + { + "bbox": [ + 55, + 322, + 296, + 369 + ], + "type": "text", + "content": " diffracted light observations in our method because, in practice, the intensity drastically decreases for higher-order diffraction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 374, + 296, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 296, + 446 + ], + "type": "text", + "content": "Given a known incoming light spectrum " + }, + { + "bbox": [ + 55, + 374, + 296, + 446 + ], + "type": "inline_equation", + "content": "e(\\lambda)" + }, + { + "bbox": [ + 55, + 374, + 296, + 446 + ], + "type": "text", + "content": " and by observing both the direct and first-order diffracted lights, " + }, + { + "bbox": [ + 55, + 374, + 296, + 446 + ], + "type": "inline_equation", + "content": "m_{\\mathrm{dir}}" + }, + { + "bbox": [ + 55, + 374, + 296, + 446 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 374, + 296, + 446 + ], + "type": "inline_equation", + "content": "m_{\\mathrm{dif}}(p)" + }, + { + "bbox": [ + 55, + 374, + 296, + 446 + ], + "type": "text", + "content": ", our goal is to estimate the camera's spectral sensitivity " + }, + { + "bbox": [ + 55, + 374, + 296, + 446 + ], + "type": "inline_equation", + "content": "s(\\lambda)" + }, + { + "bbox": [ + 55, + 374, + 296, + 446 + ], + "type": "text", + "content": " using an uncalibrated diffraction grating, where both the weight function " + }, + { + "bbox": [ + 55, + 374, + 296, + 446 + ], + "type": "inline_equation", + "content": "w(\\lambda; p)" + }, + { + "bbox": [ + 55, + 374, + 296, + 446 + ], + "type": "text", + "content": " and the grating efficiency " + }, + { + "bbox": [ + 55, + 374, + 296, + 446 + ], + "type": "inline_equation", + "content": "\\eta(\\lambda)" + }, + { + "bbox": [ + 55, + 374, + 296, + 446 + ], + "type": "text", + "content": " are unknown." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 451, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 296, + 510 + ], + "type": "text", + "content": "In the following, we first describe the method to jointly estimate the camera spectral sensitivity " + }, + { + "bbox": [ + 55, + 451, + 296, + 510 + ], + "type": "inline_equation", + "content": "s(\\lambda)" + }, + { + "bbox": [ + 55, + 451, + 296, + 510 + ], + "type": "text", + "content": " and the grating efficiency " + }, + { + "bbox": [ + 55, + 451, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\eta(\\lambda)" + }, + { + "bbox": [ + 55, + 451, + 296, + 510 + ], + "type": "text", + "content": ", assuming that the weight function " + }, + { + "bbox": [ + 55, + 451, + 296, + 510 + ], + "type": "inline_equation", + "content": "w(\\lambda; p)" + }, + { + "bbox": [ + 55, + 451, + 296, + 510 + ], + "type": "text", + "content": " is known beforehand. We then present the method to determine the weight function " + }, + { + "bbox": [ + 55, + 451, + 296, + 510 + ], + "type": "inline_equation", + "content": "w(\\lambda; p)" + }, + { + "bbox": [ + 55, + 451, + 296, + 510 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 531, + 296, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 296, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 296, + 556 + ], + "type": "text", + "content": "3.1. Estimating camera spectral sensitivity " + }, + { + "bbox": [ + 55, + 531, + 296, + 556 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 531, + 296, + 556 + ], + "type": "text", + "content": " and grating efficiency " + }, + { + "bbox": [ + 55, + 531, + 296, + 556 + ], + "type": "inline_equation", + "content": "\\eta" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 565, + 296, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 565, + 296, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 565, + 296, + 661 + ], + "type": "text", + "content": "We now estimate the camera spectral sensitivity " + }, + { + "bbox": [ + 55, + 565, + 296, + 661 + ], + "type": "inline_equation", + "content": "s(\\lambda)" + }, + { + "bbox": [ + 55, + 565, + 296, + 661 + ], + "type": "text", + "content": " and the grating efficiency " + }, + { + "bbox": [ + 55, + 565, + 296, + 661 + ], + "type": "inline_equation", + "content": "\\eta(\\lambda)" + }, + { + "bbox": [ + 55, + 565, + 296, + 661 + ], + "type": "text", + "content": ", given the weight function " + }, + { + "bbox": [ + 55, + 565, + 296, + 661 + ], + "type": "inline_equation", + "content": "w(\\lambda; p)" + }, + { + "bbox": [ + 55, + 565, + 296, + 661 + ], + "type": "text", + "content": ". We show, for the first time, that the problem has a closed-form solution when the unknowns are expressed as linear combinations of their bases. To this end, we first explain the constraints derived from the image formation models of the direct and diffracted light observations, and then present the solution method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "text", + "content": "Let us assume camera spectral sensitivity " + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "text", + "content": " and the inverse of grating efficiency " + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\eta^{-1}" + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "text", + "content": " at discretely sampled wavelengths " + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_1, \\ldots, \\lambda_f" + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "text", + "content": " can be represented by linear combinations of bases whose numbers of bases are " + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "inline_equation", + "content": "b_s" + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "inline_equation", + "content": "b_\\eta" + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "text", + "content": ", re" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 322, + 397, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 322, + 397, + 334 + ], + "spans": [ + { + "bbox": [ + 314, + 322, + 397, + 334 + ], + "type": "text", + "content": "spectively. Namely," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 343, + 342, + 526, + 358 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 342, + 526, + 358 + ], + "spans": [ + { + "bbox": [ + 343, + 342, + 526, + 358 + ], + "type": "interline_equation", + "content": "\\mathbf {s} \\triangleq \\left[ \\begin{array}{c c c} s (\\lambda_ {1}), & \\ldots , & s (\\lambda_ {f}) \\end{array} \\right] ^ {\\mathsf {T}} = \\mathbf {B} _ {s} \\mathbf {c} _ {s} \\in \\mathbb {R} _ {+} ^ {f},", + "image_path": "89e51cf151116d0e2c5672ff9e698586c7f607b7390903657cd1ef1eee6c8344.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 331, + 357, + 553, + 383 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 357, + 553, + 383 + ], + "spans": [ + { + "bbox": [ + 331, + 357, + 553, + 383 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\eta} ^ {- 1} \\triangleq \\left[ \\begin{array}{l l l} \\frac {1}{\\eta \\left(\\lambda_ {1}\\right)}, & \\dots , & \\frac {1}{\\eta \\left(\\lambda_ {f}\\right)} \\end{array} \\right] ^ {\\intercal} = \\mathbf {B} _ {\\eta} \\mathbf {c} _ {\\eta} \\in \\mathbb {R} _ {+} ^ {f}, \\tag {4}", + "image_path": "a6460a6ffe8a23308795e9f53fb5ed7e261cf425fa55d6feeac7e002b7688c74.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 392, + 555, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 555, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 392, + 555, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_* \\in \\mathbb{R}^{f \\times b_*}" + }, + { + "bbox": [ + 313, + 392, + 555, + 475 + ], + "type": "text", + "content": " is a basis matrix, and " + }, + { + "bbox": [ + 313, + 392, + 555, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_* \\in \\mathbb{R}^{b_*}" + }, + { + "bbox": [ + 313, + 392, + 555, + 475 + ], + "type": "text", + "content": " is a coefficient vector. We obtained the basis matrix " + }, + { + "bbox": [ + 313, + 392, + 555, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_s" + }, + { + "bbox": [ + 313, + 392, + 555, + 475 + ], + "type": "text", + "content": " of camera spectral sensitivity " + }, + { + "bbox": [ + 313, + 392, + 555, + 475 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 392, + 555, + 475 + ], + "type": "text", + "content": " at a channel by singular value decomposition (SVD) on the same set of data used in [23]. For the inverse grating efficiency function " + }, + { + "bbox": [ + 313, + 392, + 555, + 475 + ], + "type": "inline_equation", + "content": "\\eta^{-1}" + }, + { + "bbox": [ + 313, + 392, + 555, + 475 + ], + "type": "text", + "content": ", we use a Fourier basis representation. Further details can be found in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 479, + 554, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 554, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 554, + 526 + ], + "type": "text", + "content": "Linear constraint from direct light observation By introducing the basis representation, the aforementioned image formation model of direct light observations in Eq. (3) can be rewritten in a vector form as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 372, + 535, + 553, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 535, + 553, + 550 + ], + "spans": [ + { + "bbox": [ + 372, + 535, + 553, + 550 + ], + "type": "interline_equation", + "content": "m _ {\\mathrm {d i r}} \\propto E \\mathbf {e} ^ {\\mathsf {T}} \\mathbf {s} = E \\mathbf {e} ^ {\\mathsf {T}} \\mathbf {B} _ {s} \\mathbf {c} _ {s}, \\tag {5}", + "image_path": "47cd68ff509e6662b7c8541dc7e4dce667eb52f2c51b38743068b9c8ca82f5fa.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 560, + 554, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 560, + 554, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 560, + 554, + 620 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 560, + 554, + 620 + ], + "type": "inline_equation", + "content": "\\mathbf{e} \\in \\mathbb{R}_+^f" + }, + { + "bbox": [ + 313, + 560, + 554, + 620 + ], + "type": "text", + "content": " is the normalized incoming light spectrum. The light scale " + }, + { + "bbox": [ + 313, + 560, + 554, + 620 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 313, + 560, + 554, + 620 + ], + "type": "text", + "content": " can be set to 1, and the proportionality can be turned into equality without loss of generality, as the global scale can be accounted for in camera spectral sensitivity s. As a result, we have" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 398, + 628, + 553, + 656 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 398, + 628, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 398, + 628, + 553, + 656 + ], + "type": "interline_equation", + "content": "m _ {\\mathrm {d i r}} = \\underbrace {\\mathbf {e} ^ {\\mathsf {T}} \\mathbf {B} _ {s}} _ {\\mathbf {A} _ {\\mathrm {d i r}}} \\mathbf {c} _ {s}, \\tag {6}", + "image_path": "17ef92289a71f828e8e706498ea7cde81dd5e3d789f5d8c00d55267fd3d37f8d.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{\\mathrm{dir}} \\in \\mathbb{R}^{1 \\times b_s}" + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": " is a known row vector. Similar equations can be derived for all RGB channels to form a vector of direct light observations " + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_{\\mathrm{dir}} = [m_{\\mathrm{dir}(R)}, m_{\\mathrm{dir}(G)}, m_{\\mathrm{dir}(B)}]^{\\top}" + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": " on the left-hand side" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27255" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": "of the equation, resulting in 3 linear equations with unknown vector " + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "inline_equation", + "content": "[\\mathbf{c}_{s(R)}^{\\intercal},\\mathbf{c}_{s(G)}^{\\intercal},\\mathbf{c}_{s(B)}^{\\intercal}]^{\\intercal}" + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": ", for which " + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{\\mathrm{dir}}" + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": " becomes a " + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "inline_equation", + "content": "3\\times 3b_{s}" + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": " matrix." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 110, + 296, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 110, + 296, + 158 + ], + "spans": [ + { + "bbox": [ + 55, + 110, + 296, + 158 + ], + "type": "text", + "content": "Linear constraint from diffracted light observation Like the direct light observation case, the diffracted light's image formation model in Eq. (3) can be rewritten in a matrix form as" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 94, + 162, + 294, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 162, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 94, + 162, + 294, + 194 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {m} _ {\\text {d i f}} = \\left[ m _ {\\text {d i f}} \\left(p _ {1}\\right), \\dots , m _ {\\text {d i f}} \\left(p _ {n}\\right) \\right] ^ {\\top} \\tag {7} \\\\ = \\mathbf {W} \\operatorname {d i a g} (\\mathbf {e}) \\operatorname {d i a g} (\\boldsymbol {\\eta}) \\mathbf {s}, \\\\ \\end{array}", + "image_path": "d27248be5c8e43408c7eab60ed597ef69c07d82499324eb157738138651f6b65.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 200, + 296, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 200, + 296, + 238 + ], + "spans": [ + { + "bbox": [ + 55, + 200, + 296, + 238 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 200, + 296, + 238 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_{\\mathrm{dif}} \\in \\mathbb{R}_+^n" + }, + { + "bbox": [ + 55, + 200, + 296, + 238 + ], + "type": "text", + "content": " is a stack of the diffracted light observations at pixels " + }, + { + "bbox": [ + 55, + 200, + 296, + 238 + ], + "type": "inline_equation", + "content": "p_1, \\ldots, p_n" + }, + { + "bbox": [ + 55, + 200, + 296, + 238 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 200, + 296, + 238 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}_+^{n \\times f}" + }, + { + "bbox": [ + 55, + 200, + 296, + 238 + ], + "type": "text", + "content": " denotes the known weight function for all " + }, + { + "bbox": [ + 55, + 200, + 296, + 238 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 200, + 296, + 238 + ], + "type": "text", + "content": " pixels, written in a matrix form as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 73, + 253, + 295, + 299 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 253, + 295, + 299 + ], + "spans": [ + { + "bbox": [ + 73, + 253, + 295, + 299 + ], + "type": "interline_equation", + "content": "\\mathbf {W} = \\left[ \\begin{array}{c c c} w \\left(\\lambda_ {1}, p _ {1}\\right) & \\dots & w \\left(\\lambda_ {f}, p _ {1}\\right) \\\\ \\vdots & & \\\\ w \\left(\\lambda_ {1}, p _ {n}\\right) & \\dots & w \\left(\\lambda_ {f}, p _ {n}\\right) \\end{array} \\right] \\in \\mathbb {R} _ {+} ^ {n \\times f}, \\tag {8}", + "image_path": "c637c8e0005d6dbd01b41c58f753427d46bf4f4f54f2e6a533e579af0d3b0372.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 304, + 214, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 214, + 315 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 214, + 315 + ], + "type": "text", + "content": "which we refer to as the weight matrix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 316, + 295, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 295, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 295, + 339 + ], + "type": "text", + "content": "The above image formation model of diffracted light can be equivalently turned into" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 60, + 344, + 295, + 383 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 344, + 295, + 383 + ], + "spans": [ + { + "bbox": [ + 60, + 344, + 295, + 383 + ], + "type": "interline_equation", + "content": "\\mathbf {s} = \\operatorname {d i a g} \\left(\\boldsymbol {\\eta} ^ {- 1}\\right) \\underbrace {\\operatorname {d i a g} \\left(\\mathbf {e} ^ {- 1}\\right) \\mathbf {W} ^ {\\dagger} \\mathbf {m} _ {\\mathrm {d i f}}} _ {\\mathbf {a}} = \\operatorname {d i a g} (\\mathbf {a}) \\boldsymbol {\\eta} ^ {- 1}, \\tag {9}", + "image_path": "6da62c4ea1c40dc752715e7f8f95233d4c0d00eb09b07128ea3d80f95f4c99f7.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 384, + 296, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 296, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 296, + 467 + ], + "type": "text", + "content": "where the inverse of a vector denotes an element-wise inverse operator, " + }, + { + "bbox": [ + 55, + 384, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{\\dagger}" + }, + { + "bbox": [ + 55, + 384, + 296, + 467 + ], + "type": "text", + "content": " is a left pseudo-inverse of " + }, + { + "bbox": [ + 55, + 384, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 55, + 384, + 296, + 467 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 384, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{a} \\in \\mathbb{R}^{f}" + }, + { + "bbox": [ + 55, + 384, + 296, + 467 + ], + "type": "text", + "content": " is a known channel-and-light-dependent vector. Since we can confidently assume " + }, + { + "bbox": [ + 55, + 384, + 296, + 467 + ], + "type": "inline_equation", + "content": "n > f" + }, + { + "bbox": [ + 55, + 384, + 296, + 467 + ], + "type": "text", + "content": " in our setting, the weight matrix " + }, + { + "bbox": [ + 55, + 384, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 55, + 384, + 296, + 467 + ], + "type": "text", + "content": " should have full column rank when the diffracted observation spans the entire visible wavelength domain, and thus it has a left pseudo-inverse." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 468, + 227, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 468, + 227, + 479 + ], + "spans": [ + { + "bbox": [ + 67, + 468, + 227, + 479 + ], + "type": "text", + "content": "From Eq. (4), Eq. (9) can be written as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 125, + 486, + 294, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 486, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 125, + 486, + 294, + 499 + ], + "type": "interline_equation", + "content": "\\mathbf {B} _ {s} \\mathbf {c} _ {s} = \\operatorname {d i a g} (\\mathbf {a}) \\mathbf {B} _ {\\eta} \\mathbf {c} _ {\\eta}. \\tag {10}", + "image_path": "17dc39b9308ce004d52f5af25f2611db2796c6cc59b143c1a1aff9e527ccaa4d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 506, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 296, + 529 + ], + "type": "text", + "content": "Putting together the unknowns " + }, + { + "bbox": [ + 55, + 506, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\eta}" + }, + { + "bbox": [ + 55, + 506, + 296, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 506, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_s" + }, + { + "bbox": [ + 55, + 506, + 296, + 529 + ], + "type": "text", + "content": ", we have the following homogeneous system of equations:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 108, + 535, + 295, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 535, + 295, + 570 + ], + "spans": [ + { + "bbox": [ + 108, + 535, + 295, + 570 + ], + "type": "interline_equation", + "content": "\\underbrace {\\left[ \\begin{array}{l l} \\operatorname {d i a g} (\\mathbf {a}) \\mathbf {B} _ {\\eta} & - \\mathbf {B} _ {s} \\end{array} \\right]} _ {\\mathbf {A} _ {\\mathrm {d i f}}} \\left[ \\begin{array}{l} \\mathbf {c} _ {\\eta} \\\\ \\mathbf {c} _ {s} \\end{array} \\right] = \\mathbf {0}, \\tag {11}", + "image_path": "f5712f4adb07ca04a8d896359e3ea1c547e40f01f0ab5d7d912acbbe283997d0.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 577, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 577, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 577, + 296, + 651 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 577, + 296, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{\\mathrm{dif}} \\in \\mathbb{R}^{f \\times (b_{\\eta} + b_{s})}" + }, + { + "bbox": [ + 55, + 577, + 296, + 651 + ], + "type": "text", + "content": " is a known matrix. For the case of RGB images, stacking similar equations from all RGB channels, we obtain " + }, + { + "bbox": [ + 55, + 577, + 296, + 651 + ], + "type": "inline_equation", + "content": "3f" + }, + { + "bbox": [ + 55, + 577, + 296, + 651 + ], + "type": "text", + "content": " linear equations from the diffracted light observations for unknown vector " + }, + { + "bbox": [ + 55, + 577, + 296, + 651 + ], + "type": "inline_equation", + "content": "[\\mathbf{c}_{\\eta}^{\\top}, \\mathbf{c}_{s(R)}^{\\top}, \\mathbf{c}_{s(G)}^{\\top}, \\mathbf{c}_{s(B)}^{\\top}]^{\\top} \\in \\mathbb{R}^{b_{\\eta} + 3b_{s}}" + }, + { + "bbox": [ + 55, + 577, + 296, + 651 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 577, + 296, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{\\mathrm{dif}}" + }, + { + "bbox": [ + 55, + 577, + 296, + 651 + ], + "type": "text", + "content": " becomes a " + }, + { + "bbox": [ + 55, + 577, + 296, + 651 + ], + "type": "inline_equation", + "content": "3f \\times (b_{\\eta} + 3b_{s})" + }, + { + "bbox": [ + 55, + 577, + 296, + 651 + ], + "type": "text", + "content": " matrix." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 654, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 689 + ], + "type": "text", + "content": "Solution method We can estimate the camera spectral sensitivity " + }, + { + "bbox": [ + 55, + 654, + 296, + 689 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 654, + 296, + 689 + ], + "type": "text", + "content": " and grating efficiency " + }, + { + "bbox": [ + 55, + 654, + 296, + 689 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 654, + 296, + 689 + ], + "type": "text", + "content": " by solving the following optimization problem using RGB channels:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 61, + 695, + 296, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 695, + 296, + 716 + ], + "spans": [ + { + "bbox": [ + 61, + 695, + 296, + 716 + ], + "type": "interline_equation", + "content": "\\mathbf {x} ^ {*} = \\underset {\\mathbf {x}} {\\operatorname {a r g m i n}} \\| \\mathbf {A} _ {\\mathrm {d i f}} \\mathbf {x} \\| _ {2} ^ {2} \\text {s . t .} \\left[ \\begin{array}{l l} \\mathbf {0} & \\mathbf {A} _ {\\mathrm {d i r}} \\end{array} \\right] \\mathbf {x} = \\mathbf {m} _ {\\mathrm {d i r}}, \\tag {12}", + "image_path": "3faaaaa01086debc50da3050762270394fc237c9183cc9a529c3d3c6741040cf.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 70, + 555, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 163 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 163 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 70, + 555, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = \\left[\\mathbf{c}_{\\eta}^{\\mathsf{T}},\\mathbf{c}_{s(R)}^{\\mathsf{T}},\\mathbf{c}_{s(G)}^{\\mathsf{T}},\\mathbf{c}_{s(B)}^{\\mathsf{T}}\\right]^{\\mathsf{T}}\\in \\mathbb{R}^{b_{\\eta} + 3b_s}" + }, + { + "bbox": [ + 313, + 70, + 555, + 163 + ], + "type": "text", + "content": " is the vector of unknowns. The above optimization problem has a closed-form solution by introducing a Lagrange multiplier when the number bases " + }, + { + "bbox": [ + 313, + 70, + 555, + 163 + ], + "type": "inline_equation", + "content": "b_{*}" + }, + { + "bbox": [ + 313, + 70, + 555, + 163 + ], + "type": "text", + "content": " and the wavelength resolution " + }, + { + "bbox": [ + 313, + 70, + 555, + 163 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 70, + 555, + 163 + ], + "type": "text", + "content": " satisfy " + }, + { + "bbox": [ + 313, + 70, + 555, + 163 + ], + "type": "inline_equation", + "content": "3f\\geq b_{\\eta} + 3b_{s}" + }, + { + "bbox": [ + 313, + 70, + 555, + 163 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 70, + 555, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{\\mathrm{dif}}" + }, + { + "bbox": [ + 313, + 70, + 555, + 163 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 70, + 555, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{\\mathrm{dir}}" + }, + { + "bbox": [ + 313, + 70, + 555, + 163 + ], + "type": "text", + "content": " are full-rank. The details of the solution method are provided in the supplementary material." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 171, + 486, + 184 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 486, + 184 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 486, + 184 + ], + "type": "text", + "content": "3.2. Estimation of weight function " + }, + { + "bbox": [ + 313, + 171, + 486, + 184 + ], + "type": "inline_equation", + "content": "w" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 189, + 555, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 189, + 555, + 284 + ], + "spans": [ + { + "bbox": [ + 313, + 189, + 555, + 284 + ], + "type": "text", + "content": "As described in Sec. 1, the weight function " + }, + { + "bbox": [ + 313, + 189, + 555, + 284 + ], + "type": "inline_equation", + "content": "w(\\lambda; p)" + }, + { + "bbox": [ + 313, + 189, + 555, + 284 + ], + "type": "text", + "content": " has a peaky shape using diffraction gratings; therefore, we assume there is a one-to-one mapping between the pixel location " + }, + { + "bbox": [ + 313, + 189, + 555, + 284 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 189, + 555, + 284 + ], + "type": "text", + "content": " in the diffraction image and the wavelength " + }, + { + "bbox": [ + 313, + 189, + 555, + 284 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 189, + 555, + 284 + ], + "type": "text", + "content": " as in previous methods [13, 24]. To estimate the weight function " + }, + { + "bbox": [ + 313, + 189, + 555, + 284 + ], + "type": "inline_equation", + "content": "w(\\lambda; p)" + }, + { + "bbox": [ + 313, + 189, + 555, + 284 + ], + "type": "text", + "content": ", it is necessary to establish the pixel-to-wavelength mapping, namely, relating the pixel locations of the diffraction image to the wavelength as illustrated in Fig. 6." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 285, + 554, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 554, + 310 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 554, + 310 + ], + "type": "text", + "content": "We propose to use a quadratic function to represent the one-to-one pixel-to-wavelength mapping " + }, + { + "bbox": [ + 313, + 285, + 554, + 310 + ], + "type": "inline_equation", + "content": "g: p \\to \\lambda" + }, + { + "bbox": [ + 313, + 285, + 554, + 310 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 379, + 319, + 553, + 333 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 319, + 553, + 333 + ], + "spans": [ + { + "bbox": [ + 379, + 319, + 553, + 333 + ], + "type": "interline_equation", + "content": "\\lambda = g (p) = a p ^ {2} + b p + c, \\tag {13}", + "image_path": "fabf526ac8b6a0de4fc8504e44153d6ffd94368f104fefe87078ba8a6f8e52bf.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 342, + 554, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 554, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 554, + 392 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 342, + 554, + 392 + ], + "type": "inline_equation", + "content": "a, b, c \\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 342, + 554, + 392 + ], + "type": "text", + "content": " are the parameters. This quadratic representation allows us stably determining the mapping with retaining accuracy. For further discussion about this parameterization, please refer to our supplementary." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 392, + 554, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 554, + 428 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 554, + 428 + ], + "type": "text", + "content": "Once the pixel-to-wavelength mapping " + }, + { + "bbox": [ + 313, + 392, + 554, + 428 + ], + "type": "inline_equation", + "content": "g(p)" + }, + { + "bbox": [ + 313, + 392, + 554, + 428 + ], + "type": "text", + "content": " is determined, the elements of the weight matrix " + }, + { + "bbox": [ + 313, + 392, + 554, + 428 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 313, + 392, + 554, + 428 + ], + "type": "text", + "content": " is simply obtained by linear interpolation as" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 349, + 437, + 554, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 437, + 554, + 506 + ], + "spans": [ + { + "bbox": [ + 349, + 437, + 554, + 506 + ], + "type": "interline_equation", + "content": "w (\\lambda , p) = \\left\\{ \\begin{array}{l l} \\frac {\\lambda_ {i + 1} - g (p)}{\\lambda_ {i + 1} - \\lambda_ {i}} & \\text {i f} \\lambda = \\lambda_ {i} \\\\ \\frac {g (p) - \\lambda_ {i}}{\\lambda_ {i + 1} - \\lambda_ {i}} & \\text {i f} \\lambda = \\lambda_ {i + 1}, \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {14}", + "image_path": "e88b74beb2ebd226e50eb0ba123130c34b830430875e81c38bcbd4c7fe5de616.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 516, + 555, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 555, + 635 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 516, + 555, + 635 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 313, + 516, + 555, + 635 + ], + "type": "text", + "content": " is the nearest-neighbor wavelength to " + }, + { + "bbox": [ + 313, + 516, + 555, + 635 + ], + "type": "inline_equation", + "content": "g(p)" + }, + { + "bbox": [ + 313, + 516, + 555, + 635 + ], + "type": "text", + "content": " that satisfies " + }, + { + "bbox": [ + 313, + 516, + 555, + 635 + ], + "type": "inline_equation", + "content": "\\lambda_{i} \\leq g(p) \\leq \\lambda_{i+1}" + }, + { + "bbox": [ + 313, + 516, + 555, + 635 + ], + "type": "text", + "content": ". Therefore, the problem of estimating the weight function boils down to the estimation of pixel-to-wavelength mapping. The challenge, though, is that we only have access to the diffraction image and illumination spectra, but not to camera spectral sensitivity and grating efficiency. Now we describe two settings for determining the pixel-to-wavelength mapping; one using fluorescent and LED lights (denoted as Fluorescent+LED), and another using only LED light (denoted as LED only)." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "content": "Fluorescent+LED case When light with spiky spectral peaks, such as fluorescent light, passes through the diffraction grating, the resulting diffracted observation exhibits a similar spiky pattern, as shown in Fig. 5. Therefore, the fluorescent light provides a strong cue for determining the pixel-to-wavelength mapping. In this case, we use a similar" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27256" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 293, + 148 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 293, + 148 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 293, + 148 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 293, + 148 + ], + "type": "image", + "image_path": "95a9c140f1146aa67cb9110d4a2d3be9199b3d69cf7a88772ca29ea55de53de3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 157, + 296, + 201 + ], + "lines": [ + { + "bbox": [ + 55, + 157, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 296, + 201 + ], + "type": "text", + "content": "Figure 5. The illuminant spectrum of the light source used in our real-world experiment and an example of the corresponding direct and diffracted light observations. The observations shown here are adjusted for visualization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 211, + 296, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 211, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 54, + 211, + 296, + 319 + ], + "type": "text", + "content": "approach to [13] that associates corresponding peak positions in the pixel and wavelength domains. Since the spectral sensitivity " + }, + { + "bbox": [ + 54, + 211, + 296, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 54, + 211, + 296, + 319 + ], + "type": "text", + "content": " is unknown, we weigh the illumination spectra " + }, + { + "bbox": [ + 54, + 211, + 296, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{e}" + }, + { + "bbox": [ + 54, + 211, + 296, + 319 + ], + "type": "text", + "content": " by the mean spectral sensitivity " + }, + { + "bbox": [ + 54, + 211, + 296, + 319 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{s}}" + }, + { + "bbox": [ + 54, + 211, + 296, + 319 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 54, + 211, + 296, + 319 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{s}}\\odot \\mathbf{e}" + }, + { + "bbox": [ + 54, + 211, + 296, + 319 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 211, + 296, + 319 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{s}}" + }, + { + "bbox": [ + 54, + 211, + 296, + 319 + ], + "type": "text", + "content": " is obtained from the same dataset used for the camera spectral sensitivity basis [23], and " + }, + { + "bbox": [ + 54, + 211, + 296, + 319 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 54, + 211, + 296, + 319 + ], + "type": "text", + "content": " represents an elementwise multiplication operator. We have observed that the unknown grating efficiency " + }, + { + "bbox": [ + 54, + 211, + 296, + 319 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 54, + 211, + 296, + 319 + ], + "type": "text", + "content": " is negligible in determining the peak positions due to its low-frequency distribution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 319, + 296, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 402 + ], + "type": "text", + "content": "We select the top two peaks from each of the pixel and wavelength domains for each channel to determine the polynomial parameters " + }, + { + "bbox": [ + 55, + 319, + 296, + 402 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 55, + 319, + 296, + 402 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 319, + 296, + 402 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 319, + 296, + 402 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 319, + 296, + 402 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 319, + 296, + 402 + ], + "type": "text", + "content": " of the pixel-to-wavelength mapping " + }, + { + "bbox": [ + 55, + 319, + 296, + 402 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 319, + 296, + 402 + ], + "type": "text", + "content": " via linear least squares. Once the pixel-to-wavelength mapping " + }, + { + "bbox": [ + 55, + 319, + 296, + 402 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 319, + 296, + 402 + ], + "type": "text", + "content": " and weight matrix " + }, + { + "bbox": [ + 55, + 319, + 296, + 402 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 55, + 319, + 296, + 402 + ], + "type": "text", + "content": " are obtained, we take another capture using an LED light for obtaining the direct and diffracted measurements used in Sec. 3.1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 405, + 296, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 405, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 296, + 464 + ], + "type": "text", + "content": "LED only case To simplify the recording process, it is preferable if we can work only with an LED. In this case, due to the low-frequency incoming spectra without obvious peaks, we cannot take the strategy of relating peak positions used in the Fluorescent+LED case." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 465, + 296, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 465, + 296, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 465, + 296, + 537 + ], + "type": "text", + "content": "Therefore, instead of finding peak positions, we frame the problem as finding the optimal horizontal mapping (pixel-to-wavelength) between two plots—the normalized diffraction observations " + }, + { + "bbox": [ + 55, + 465, + 296, + 537 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 55, + 465, + 296, + 537 + ], + "type": "text", + "content": " and the normalized mean camera spectral sensitivity " + }, + { + "bbox": [ + 55, + 465, + 296, + 537 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 55, + 465, + 296, + 537 + ], + "type": "text", + "content": " that is scaled by illumination spectrum, expressed as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 71, + 544, + 275, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 544, + 275, + 559 + ], + "spans": [ + { + "bbox": [ + 71, + 544, + 275, + 559 + ], + "type": "interline_equation", + "content": "\\mathbf {v} \\triangleq \\nu (\\mathbf {e} \\odot \\bar {\\mathbf {s}}) = [ v (\\lambda_ {1}), \\dots , v (\\lambda_ {f}) ] \\in \\mathbb {R} _ {+} ^ {f}", + "image_path": "be4948d983b411ef084fdbc461204f059f202d1a5df801b64e9d419c1c2203e0.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 71, + 561, + 278, + 577 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 561, + 278, + 577 + ], + "spans": [ + { + "bbox": [ + 71, + 561, + 278, + 577 + ], + "type": "interline_equation", + "content": "\\mathbf {u} \\triangleq \\nu (\\mathbf {m} _ {\\mathrm {d i f}}) = [ u (p _ {1}), \\dots , u (p _ {n}) ] \\in \\mathbb {R} _ {+} ^ {n},", + "image_path": "2c17995dfa1bc4815fa99acc417b135ae535f763fcd19b2d8a7762d265cc6cd2.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 275, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 275, + 595 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 275, + 595 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 582, + 275, + 595 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 55, + 582, + 275, + 595 + ], + "type": "text", + "content": " is a vector normalization as " + }, + { + "bbox": [ + 55, + 582, + 275, + 595 + ], + "type": "inline_equation", + "content": "\\nu (\\mathbf{x}) = \\mathbf{x} / \\| \\mathbf{x}\\| _2" + }, + { + "bbox": [ + 55, + 582, + 275, + 595 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 595, + 296, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 595, + 296, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 595, + 296, + 643 + ], + "type": "text", + "content": "We solve the point-to-plane iterative closest point (ICP) problem [3] to determine the optimal polynomial parameters " + }, + { + "bbox": [ + 55, + 595, + 296, + 643 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 55, + 595, + 296, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 595, + 296, + 643 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 595, + 296, + 643 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 595, + 296, + 643 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 595, + 296, + 643 + ], + "type": "text", + "content": " along the horizontal axis by minimizing the following objective function:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 648, + 299, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 299, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 299, + 681 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {I C P}} = \\sum_ {i = 1} ^ {n} \\sum_ {j \\in \\mathcal {N} (i)} \\left(\\left(\\left[ \\begin{array}{c} a p _ {i} ^ {2} + b p _ {i} + c \\\\ u (p _ {i}) \\end{array} \\right] - \\left[ \\begin{array}{c} \\lambda_ {j} \\\\ v (\\lambda_ {j}) \\end{array} \\right]\\right) ^ {\\top} \\mathbf {n _ {v}} (j)\\right) ^ {2},", + "image_path": "0bb6917a0d1b7abefdbedd138c0d2ff3d661ec0be2a8e05f9004078b4baf39ba.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 687, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 687, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 687, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 687, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(i)" + }, + { + "bbox": [ + 55, + 687, + 296, + 715 + ], + "type": "text", + "content": " is the set of indices for the current neighboring points of " + }, + { + "bbox": [ + 55, + 687, + 296, + 715 + ], + "type": "inline_equation", + "content": "[ap_i^2 +bp_i + c,u(p_i)]^\\top" + }, + { + "bbox": [ + 55, + 687, + 296, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 687, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{n_v}(j)\\in \\mathbb{S}\\subset \\mathbb{R}^2" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 553, + 222 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 553, + 222 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 553, + 222 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 553, + 222 + ], + "type": "image", + "image_path": "83b58e14ecc95c28607285e0bec20125b826946c5fe94919d2acb16d329f433f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 230, + 554, + 297 + ], + "lines": [ + { + "bbox": [ + 313, + 230, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 230, + 554, + 297 + ], + "type": "text", + "content": "Figure 6. Illustration of our pixel-to-wavelength mapping estimation. Given the normalized diffraction observations " + }, + { + "bbox": [ + 313, + 230, + 554, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_c" + }, + { + "bbox": [ + 313, + 230, + 554, + 297 + ], + "type": "text", + "content": " and normalized illumination-scaled mean camera spectral sensitivity " + }, + { + "bbox": [ + 313, + 230, + 554, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_c" + }, + { + "bbox": [ + 313, + 230, + 554, + 297 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 313, + 230, + 554, + 297 + ], + "type": "inline_equation", + "content": "c \\in \\{R, G, B\\}" + }, + { + "bbox": [ + 313, + 230, + 554, + 297 + ], + "type": "text", + "content": " channels, we estimate the optimal pixel-to-wavelength mapping (or horizontal warping, as shown in the figure), which minimizes the distance between the two plots." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 306, + 529, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 529, + 322 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 529, + 322 + ], + "type": "text", + "content": "is a unit normal vector to " + }, + { + "bbox": [ + 313, + 306, + 529, + 322 + ], + "type": "inline_equation", + "content": "[\\lambda_j,v(\\lambda_j)]^\\top" + }, + { + "bbox": [ + 313, + 306, + 529, + 322 + ], + "type": "text", + "content": " which satisfies" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 369, + 329, + 553, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 329, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 369, + 329, + 553, + 360 + ], + "type": "interline_equation", + "content": "\\left[ \\begin{array}{c} \\lambda_ {j} - \\lambda_ {j + 1} \\\\ v (\\lambda_ {j}) - v (\\lambda_ {j + 1}) \\end{array} \\right] ^ {\\mathsf {T}} \\mathbf {n} _ {\\mathrm {v}} (j) = 0. \\tag {15}", + "image_path": "4837de423022234c2125b2498ebd82b5fe93e9685eb09c5464898f7a375cb5d7.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 366, + 555, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 555, + 461 + ], + "type": "text", + "content": "We iteratively minimize the objective function for 500 iterations, with the initial guesses set to " + }, + { + "bbox": [ + 313, + 366, + 555, + 461 + ], + "type": "inline_equation", + "content": "a = 0" + }, + { + "bbox": [ + 313, + 366, + 555, + 461 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 366, + 555, + 461 + ], + "type": "inline_equation", + "content": "b = f / n" + }, + { + "bbox": [ + 313, + 366, + 555, + 461 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 366, + 555, + 461 + ], + "type": "inline_equation", + "content": "c = 0" + }, + { + "bbox": [ + 313, + 366, + 555, + 461 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 366, + 555, + 461 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 366, + 555, + 461 + ], + "type": "text", + "content": " is given as the ratio of the numbers of elements of " + }, + { + "bbox": [ + 313, + 366, + 555, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 313, + 366, + 555, + 461 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 366, + 555, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 313, + 366, + 555, + 461 + ], + "type": "text", + "content": ", to determine the pixel-to-wavelength mapping " + }, + { + "bbox": [ + 313, + 366, + 555, + 461 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 366, + 555, + 461 + ], + "type": "text", + "content": ". This setting allows for a convenient one-time capture for both determining the pixel-to-wavelength mapping and obtaining the direct and diffracted measurements used in Sec. 3.1." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 473, + 395, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 395, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 395, + 487 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 494, + 554, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 494, + 554, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 554, + 530 + ], + "type": "text", + "content": "In this section, we evaluate our method on both synthetic and real-world scenes. We first describe our experimental setup and then present the evaluation results." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 534, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 677 + ], + "type": "text", + "content": "Baselines We compare the proposed approach with the Exif-based method (Exif+CC) [23] and the color chart-based method (CC) [12]. We present two variants of the proposed method as detailed in Sec. 3.2: Ours (LED+Flu), which employs a fluorescent lamp observation for pixel-to-wavelength mapping estimation, and Ours (LED), which uses only an LED light observation. For the color chart-based method, we use 2 basis functions, as mentioned in the original paper, which yielded the best results in our preliminary experiments. Since the Exif-based method cannot estimate white balance and introduces a " + }, + { + "bbox": [ + 313, + 534, + 555, + 677 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 313, + 534, + 555, + 677 + ], + "type": "text", + "content": " ambiguity, we use the color checker observation to resolve this ambiguity." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "We also compared ours with the existing diffraction grating-based approach by Toivonen et al. [24], adjusted to our experimental setup. However, their method produced" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27257" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 109 + ], + "type": "text", + "content": "unstable estimations due to the limited observations available in our setup compared to their original conditions. Further details are provided in the supplemental material." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 112, + 295, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 112, + 295, + 328 + ], + "spans": [ + { + "bbox": [ + 55, + 112, + 295, + 328 + ], + "type": "text", + "content": "Basis representations To compute the basis for camera spectral sensitivity " + }, + { + "bbox": [ + 55, + 112, + 295, + 328 + ], + "type": "inline_equation", + "content": "s(\\lambda)" + }, + { + "bbox": [ + 55, + 112, + 295, + 328 + ], + "type": "text", + "content": ", we use the dataset from the Exif-based method [23], which includes 44 cameras for training and 5 for testing. Following the detailed analysis on basis selection for a camera spectral sensitivity in [12], we apply SVD to the sensitivities in the training dataset in a per-channel manner. For the grating efficiency, we use a Fourier basis. Further discussion on the grating efficiency basis can be found in the supplementary material. The spectral resolution " + }, + { + "bbox": [ + 55, + 112, + 295, + 328 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 112, + 295, + 328 + ], + "type": "text", + "content": " in the dataset of camera spectral sensitivities is " + }, + { + "bbox": [ + 55, + 112, + 295, + 328 + ], + "type": "inline_equation", + "content": "f = 31" + }, + { + "bbox": [ + 55, + 112, + 295, + 328 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 112, + 295, + 328 + ], + "type": "inline_equation", + "content": "400\\mathrm{nm}" + }, + { + "bbox": [ + 55, + 112, + 295, + 328 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 112, + 295, + 328 + ], + "type": "inline_equation", + "content": "700\\mathrm{nm}" + }, + { + "bbox": [ + 55, + 112, + 295, + 328 + ], + "type": "text", + "content": ". Assuming the number of basis for the camera spectral sensitivity and the grating efficiency to be the same " + }, + { + "bbox": [ + 55, + 112, + 295, + 328 + ], + "type": "inline_equation", + "content": "b_{\\eta} = b_{s}" + }, + { + "bbox": [ + 55, + 112, + 295, + 328 + ], + "type": "text", + "content": ", the required condition for the number of bases for our method is " + }, + { + "bbox": [ + 55, + 112, + 295, + 328 + ], + "type": "inline_equation", + "content": "b_{\\eta} = b_{s} \\leq 23" + }, + { + "bbox": [ + 55, + 112, + 295, + 328 + ], + "type": "text", + "content": " based on the discussion in Sec. 3.1. For robust estimation, the number of basis functions is set to 7 for each channel of the camera spectral sensitivity and the grating efficiency across all experiments." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 332, + 295, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 332, + 295, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 332, + 295, + 368 + ], + "type": "text", + "content": "Evaluation metrics Following [23], we use the average relative full-scale error across the three color channels, RE, which is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 376, + 295, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 376, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 61, + 376, + 295, + 407 + ], + "type": "interline_equation", + "content": "\\mathrm {R E} = \\frac {1}{3} \\sum_ {c} \\frac {\\operatorname {R M S E} (c)}{\\max \\left(\\mathbf {s} _ {c}\\right)}, \\operatorname {R M S E} (c) = \\sqrt {\\frac {\\left\\| \\mathbf {s} _ {c} ^ {*} - \\mathbf {s} _ {c} \\right\\| _ {2} ^ {2}}{f}}, \\tag {16}", + "image_path": "98365e52d1fee7510dddc771b3ef32ea46a1556c650575e491dec651463981b1.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 415, + 295, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 295, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 295, + 440 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 415, + 295, + 440 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_c^*" + }, + { + "bbox": [ + 55, + 415, + 295, + 440 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 415, + 295, + 440 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_c" + }, + { + "bbox": [ + 55, + 415, + 295, + 440 + ], + "type": "text", + "content": " denote the estimated and ground-truth camera spectral sensitivity for each channel " + }, + { + "bbox": [ + 55, + 415, + 295, + 440 + ], + "type": "inline_equation", + "content": "c\\in \\{R,G,B\\}" + }, + { + "bbox": [ + 55, + 415, + 295, + 440 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 448, + 230, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 230, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 230, + 460 + ], + "type": "text", + "content": "4.1. Evaluation with synthetic scenes" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 465, + 295, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 465, + 295, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 465, + 295, + 489 + ], + "type": "text", + "content": "In this section, we quantitatively evaluate the comparison methods and ours on a synthetic dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 494, + 295, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 494, + 295, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 295, + 661 + ], + "type": "text", + "content": "Dataset We render the direct and diffracted images using Eq. (2), with parameters as follows. We use five camera spectral sensitivities: EOS 600D, iPhone12ProMax, Olympus EPL2, Pentax K5, and Sumsung GalaxyS20, all from the same test dataset as the Exif-based method. The D65 illuminant, as defined by the CIE Standard Illuminants, is used as the incoming light. To synthesize the pixel-to-wavelength mapping, we randomly define the mapping while ensuring that the visible wavelength range is fully covered in the diffraction image. The grating efficiency function is set to a low-frequency response as illustrated in Fig. 4, parameterized by 7 Fourier bases. Since the Exif-based method relies on metadata rather than images, we evaluate their method using real-world data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": "Results Figure 7(a) shows the estimated results of the proposed and comparison methods for our synthetic scenes. For ideal observations, both Ours (LED+Flu) and the color chart-based method, CC, yield low estimation errors, as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": "both methods are guaranteed to produce optimal results under the basis representation assumption. The remaining errors stem from the gap between the assumption and the synthetic data. The Exif-based method, Exif+CC, produces a generally reasonable overall shape, but it fails to accurately capture finer details. Ours (LED) yields a larger error than Ours (LED+Flu) due to estimation errors in the pixel-to-wavelength mapping when using only LED observations; however, overall, we can still obtain a reasonable estimation even by capturing just a single calibration scene." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 200, + 495, + 212 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 200, + 495, + 212 + ], + "spans": [ + { + "bbox": [ + 313, + 200, + 495, + 212 + ], + "type": "text", + "content": "4.2. Evaluation with real-world scenes" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 218, + 548, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 218, + 548, + 230 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 548, + 230 + ], + "type": "text", + "content": "Here, we present evaluations using our real-world dataset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 234, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 554, + 437 + ], + "type": "text", + "content": "Dataset 3 illustrates our experimental setup, which consists of a light source, a diffraction grating sheet, and a camera. We capture the images in a dark room for both our method and the comparison methods, ensuring that the only light in the scene comes from illumination with known spectra (LED or fluorescent). We use a 3D-printed slit for the incoming light, positioned in front of the light source, and place a diffraction grating in front of the camera, manually aligned roughly parallel to the image plane. For the spiky light source, we use a Pa-Look Ball (EFD15EL)1 fluorescent lamp. We use the following four cameras: EOS RP, iPhone 15ProMax, Sony " + }, + { + "bbox": [ + 313, + 234, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 234, + 554, + 437 + ], + "type": "text", + "content": "1, and DJI Pocket3. For the non-spiky light sources and diffraction sheets, we use three different products for each. In this section, we present the results for one setup. Additional experiments with other setups are provided in the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 437, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 557 + ], + "type": "text", + "content": "Due to the high dynamic range of the direct and diffracted light, we capture HDR images by taking two images with different exposure times for each. The extra diffracted image under fluorescent light is captured for LED+Flu, solely for pixel-to-wavelength mapping. After capturing the direct and diffracted light observations, we roughly crop them with a sufficient margin at both ends of the diffraction pattern along the diffraction direction. The diffracted light observation is cropped to approximately " + }, + { + "bbox": [ + 313, + 437, + 554, + 557 + ], + "type": "inline_equation", + "content": "200 \\times 1000" + }, + { + "bbox": [ + 313, + 437, + 554, + 557 + ], + "type": "text", + "content": " pix and aggregated to " + }, + { + "bbox": [ + 313, + 437, + 554, + 557 + ], + "type": "inline_equation", + "content": "1 \\times 1000" + }, + { + "bbox": [ + 313, + 437, + 554, + 557 + ], + "type": "text", + "content": " pix." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 558, + 554, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 666 + ], + "type": "text", + "content": "For the comparison methods, we use a 24-patch color checker chart with spectral reflectance data from Chromaxion. The color checker chart is captured under the same non-spiky light source used in our experiments. We carefully position the color checker chart so that it is parallel to the camera and uniformly illuminated by the light source, ensuring that no shadows are cast. The metadata for the Exif-based method is extracted from our captured data following the procedure described in their paper [23]." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 326, + 674, + 405, + 683 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 674, + 405, + 683 + ], + "spans": [ + { + "bbox": [ + 326, + 674, + 405, + 683 + ], + "type": "text", + "content": "1 Panasonic Corporation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 326, + 684, + 433, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 684, + 433, + 693 + ], + "spans": [ + { + "bbox": [ + 326, + 684, + 433, + 693 + ], + "type": "text", + "content": "2ColorChecker Classic, Calibrite" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 693, + 553, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 693, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 693, + 553, + 712 + ], + "type": "text", + "content": "3Chromaxion, https://www.chromaxion.com, last accessed on March 7, 2025." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27258" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 83, + 312, + 380 + ], + "blocks": [ + { + "bbox": [ + 148, + 71, + 240, + 82 + ], + "lines": [ + { + "bbox": [ + 148, + 71, + 240, + 82 + ], + "spans": [ + { + "bbox": [ + 148, + 71, + 240, + 82 + ], + "type": "text", + "content": "(a) Synthetic experiment" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 83, + 312, + 380 + ], + "lines": [ + { + "bbox": [ + 53, + 83, + 312, + 380 + ], + "spans": [ + { + "bbox": [ + 53, + 83, + 312, + 380 + ], + "type": "image", + "image_path": "248692476abfeef63fd2fccf00e5837067470deea016905411937e1014429e3c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 320, + 83, + 561, + 392 + ], + "blocks": [ + { + "bbox": [ + 389, + 71, + 487, + 82 + ], + "lines": [ + { + "bbox": [ + 389, + 71, + 487, + 82 + ], + "spans": [ + { + "bbox": [ + 389, + 71, + 487, + 82 + ], + "type": "text", + "content": "(b) Real-world experiment" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 320, + 83, + 561, + 392 + ], + "lines": [ + { + "bbox": [ + 320, + 83, + 561, + 392 + ], + "spans": [ + { + "bbox": [ + 320, + 83, + 561, + 392 + ], + "type": "image", + "image_path": "af89459e082555441ea3e987d5cbb408993d84a1567429bdae276bce4bd19096.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 400, + 555, + 445 + ], + "lines": [ + { + "bbox": [ + 54, + 400, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 54, + 400, + 555, + 445 + ], + "type": "text", + "content": "Figure 7. Results of the proposed and comparison methods for our (a) synthetic and (b) real-world datasets. For each setting and method, we show plots of the estimated (solid line) and ground-truth (dotted line) spectral sensitivity for each color channel. All plots are normalized by the maximum response across all channels. The values below the plots indicate the RE defined in Eq. (16). For simplicity, the factor of " + }, + { + "bbox": [ + 54, + 400, + 555, + 445 + ], + "type": "inline_equation", + "content": "10^{-2}" + }, + { + "bbox": [ + 54, + 400, + 555, + 445 + ], + "type": "text", + "content": " is omitted. Underscored bold and underscored score are the best and second-best result for each setting, respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 465, + 295, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 465, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 54, + 465, + 295, + 514 + ], + "type": "text", + "content": "The ground-truth spectral sensitivities are obtained using a calibration device with a lens attached. The illuminant spectra of the light sources, shown in Fig. 5, are measured through the diffraction grating using a spectrometer." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 519, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 519, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 54, + 519, + 295, + 685 + ], + "type": "text", + "content": "Results Figure 7(b) presents the estimated results of the proposed and comparison methods. Overall, the proposed methods achieve the lowest estimation errors for most scenes, and Ours (LED) produces results nearly comparable to Ours (LED+Flu). For the EOS RP scene, Ours (LED) yields a larger error due to the failure of pixel-to-wavelength mapping estimation using a non-spiky light source. The color checker-based method, CC, is relatively sensitive to real-world noise due to the low dimensionality of the spectral reflectances in the color patches, as discussed in previous works [5, 24]. While the Exif-based method, Exif+CC, has an advantage of being easy to use, as it only relies on metadata for recovering sensitivity, it yields larger estimation errors across all scenes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 464, + 391, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 391, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 391, + 477 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 485, + 555, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 485, + 555, + 593 + ], + "spans": [ + { + "bbox": [ + 312, + 485, + 555, + 593 + ], + "type": "text", + "content": "We present a method for estimating camera spectral sensitivity using an uncalibrated diffraction grating sheet. Compared to previous methods that use reference targets, such as a color checker chart, our approach achieves highly accurate estimation by leveraging the diffraction grating. Unlike previous diffraction grating-based methods, the proposed method jointly estimates both camera spectral sensitivity and grating efficiency without needing additional recordings, making it practical, cost-efficient, and easy-to-use." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": "One of the limitations of our method is that pixel-to-wavelength mapping estimation breaks down if the illumination has uniform spectra. In this case, the diffraction image simply shows the camera spectral sensitivity scaled by unknown grating efficiency, providing no clues for the mapping estimation. However, in practice, readily-available light sources do not have such uniform spectra. LEDs, for example, tend to have a non-uniform spectral distribution, enabling us to estimate the mapping and camera spectral sensitivities, as demonstrated in our real-world experiments." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 693, + 247, + 702 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 247, + 702 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 247, + 702 + ], + "type": "text", + "content": "4camSPECS XL, Image Engineering GmbH & Co. KG" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 66, + 702, + 197, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 197, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 197, + 712 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 66, + 702, + 197, + 712 + ], + "type": "text", + "content": "CS-2000S, KONICA MINOLTA, INC." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27259" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 159 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 159 + ], + "type": "text", + "content": "This work was supported by JSPS KAKENHI Grant Numbers JP23H05491 and JP25K00142, JST ASPIRE Grant Number JPMJAP2304, the Canada First Research Excellence Fund program (VISTA), an NSERC Discovery Grant, and the Canada Research Chair program." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 167, + 115, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 167, + 115, + 180 + ], + "spans": [ + { + "bbox": [ + 56, + 167, + 115, + 180 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 187, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 187, + 296, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 187, + 296, + 232 + ], + "spans": [ + { + "bbox": [ + 61, + 187, + 296, + 232 + ], + "type": "text", + "content": "[1] Ali Alsam and Graham Finlayson. Recovering spectral sensitivities with uncertainty. In Proceedings of Conference on Colour in Graphics, Imaging, and Vision (CGIV), pages 22-26, 2002. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 232, + 296, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 232, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 61, + 232, + 296, + 308 + ], + "type": "text", + "content": "[2] E Berra, S Gibson-Poole, A MacArthur, R Gaulton, and A Hamilton. Estimation of the spectral sensitivity functions of un-modified and modified commercial off-the-shelf digital cameras to enable their use as a multispectral imaging system for UAVs. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 40: 207-214, 2015. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 310, + 295, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 310, + 295, + 343 + ], + "spans": [ + { + "bbox": [ + 61, + 310, + 295, + 343 + ], + "type": "text", + "content": "[3] Yang Chen and Gérard Medioni. Object modelling by registration of multiple range images. Image and Vision Computing, 10(3):145-155, 1992. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 344, + 295, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 344, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 61, + 344, + 295, + 388 + ], + "type": "text", + "content": "[4] Maryam Mohammadzadeh Darrodi, Graham Finlayson, Teresa Goodman, and Michal Mackiewicz. Reference data set for camera spectral sensitivity estimation. Journal of the Optical Society of America A, 32(3):381-391, 2015. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 388, + 296, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 296, + 432 + ], + "type": "text", + "content": "[5] Jeffrey M DiCarlo, Glen Eric Montgomery, and Steven W Trovinger. Emissive chart for imager calibration. In Proceedings of IS&T Color and Imaging Conference (CIC), pages 295-301, 2004. 1, 3, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 434, + 295, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 434, + 295, + 466 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 295, + 466 + ], + "type": "text", + "content": "[6] Bob Dyas. Robust color sensor response characterization. In Proceedings of IS&T Color and Imaging Conference (CIC), pages 144-148, 2000. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 468, + 296, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 468, + 296, + 511 + ], + "spans": [ + { + "bbox": [ + 62, + 468, + 296, + 511 + ], + "type": "text", + "content": "[7] Marc Ebner. Estimating the spectral sensitivity of a digital sensor using calibration targets. In Proceedings of Annual Conference on Genetic and Evolutionary Computation, pages 642-649, 2007. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 512, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 512, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 512, + 295, + 544 + ], + "type": "text", + "content": "[8] Graham Finlayson, Maryam Mohammadzadeh Darrodi, and Michal Mackiewicz. Rank-based camera spectral sensitivity estimation. 33(4):589-599, 2016. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 545, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 545, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 545, + 296, + 590 + ], + "type": "text", + "content": "[9] Paul M Hubel, Doron Sherman, and Joyce E Farrell. A comparison of methods of sensor spectral sensitivity estimation. In Proceedings of IS&T Color and Imaging Conference (CIC), pages 45-48, 1994. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 590, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 295, + 646 + ], + "type": "text", + "content": "[10] Cong Phuoc Huynh and Antonio Robles-Kelly. Recovery of spectral sensitivity functions from a colour chart image under unknown spectrally smooth illumination. In Proceedings of IEEE International Conference on Pattern Recognition (ICPR), pages 708-713, 2014. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "text", + "content": "[11] Yuhyun Ji, Yunsang Kwak, Sang Mok Park, and Young L Kim. Compressive recovery of smartphone RGB spectral sensitivity functions. Optics Express, 29(8):11947-11961, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 690, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 690, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 690, + 295, + 714 + ], + "type": "text", + "content": "[12] Jun Jiang, Dengyu Liu, Jinwei Gu, and Sabine Süssstrunk. What is the space of spectral sensitivity functions for digital" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 73, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 554, + 106 + ], + "type": "text", + "content": "color cameras? In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 168-179, 2013. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 108, + 554, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 554, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 554, + 163 + ], + "type": "text", + "content": "[13] Andreas Karge, Ingmar Rieger, Bernhard Eberhardt, and Andreas Schilling. Using chromaticity error minimisation for fast camera spectral responsivity measurement. In Proceedings of IS&T Color and Imaging Conference (CIC), pages 67-74, 2018. 1, 3, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 164, + 554, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 554, + 208 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 554, + 208 + ], + "type": "text", + "content": "[14] Rei Kawakami, Hongxun Zhao, Robby T Tan, and Katsushi Ikeuchi. Camera spectral sensitivity and white balance estimation from sky images. International Journal of Computer Vision (IJCV), 105:187-204, 2013. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 209, + 554, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 554, + 263 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 554, + 263 + ], + "type": "text", + "content": "[15] Berk Kaya, Yigit Baran Can, and Radu Timofte. Towards spectral estimation from a single RGB image in the wild. In Proceedings of IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 3546-3555, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 266, + 554, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 266, + 554, + 311 + ], + "spans": [ + { + "bbox": [ + 316, + 266, + 554, + 311 + ], + "type": "text", + "content": "[16] Lindsay W MacDonald. Determining camera spectral responsivity with multispectral transmission filters. In Proceedings of IS&T Color and Imaging Conference (CIC), pages 12-17, 2015. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 312, + 554, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 312, + 554, + 335 + ], + "spans": [ + { + "bbox": [ + 316, + 312, + 554, + 335 + ], + "type": "text", + "content": "[17] Christopher Palmer and Erwin G Loewen. Diffraction Grating Handbook. Newport Corporation New York, 2005. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 336, + 554, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 336, + 554, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 336, + 554, + 369 + ], + "type": "text", + "content": "[18] J. P. S. Parkkinen, J. Hallikainen, and T. Jaaskelainen. Characteristic spectra of munsell colors. Journal of the Optical Society of America A, 6(2):318-322, 1989. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 371, + 554, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 554, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 554, + 425 + ], + "type": "text", + "content": "[19] Dilip Prasad, Rang Nguyen, and Michael Brown. Quick approximation of camera's spectral response from casual lighting. In Proceedings of IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 844-851, 2013. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 427, + 554, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 554, + 472 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 554, + 472 + ], + "type": "text", + "content": "[20] Shuxue Quan, Noboru Ohta, and Xiaoyun Jiang. Comparative study on sensor spectral sensitivity estimation. In Proceedings of SPIE Color Imaging VIII: Processing, Hardcopy, and Applications, pages 209-220, 2003." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 473, + 554, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 473, + 554, + 506 + ], + "spans": [ + { + "bbox": [ + 316, + 473, + 554, + 506 + ], + "type": "text", + "content": "[21] Martin Rump, Arno Zinke, and Reinhard Klein. Practical spectral characterization of trichromatic cameras. ACM Transactions on Graphics (TOG), 30(6):170, 2011. 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 508, + 554, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 508, + 554, + 540 + ], + "spans": [ + { + "bbox": [ + 316, + 508, + 554, + 540 + ], + "type": "text", + "content": "[22] Yuri Rzhanov. Accurate determination of camera quantum efficiency from a single image. Journal of Imaging, 10(7): 169, 2024. 1, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 543, + 554, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 543, + 554, + 586 + ], + "spans": [ + { + "bbox": [ + 316, + 543, + 554, + 586 + ], + "type": "text", + "content": "[23] Grigory Solomatov and Derya Akkaynak. Spectral sensitivity estimation without a camera. In International Conference on Computational Photography (ICCP), pages 1-12, 2023. 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 588, + 555, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 588, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 588, + 555, + 622 + ], + "type": "text", + "content": "[24] Mikko E Toivonen and Arto Klami. Practical camera sensor spectral response and uncertainty estimation. Journal of Imaging, 6(8):79, 2020. 1, 3, 5, 6, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 624, + 554, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 554, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 554, + 657 + ], + "type": "text", + "content": "[25] Shoji Tominaga, Shogo Nishi, and Ryo Ohtera. Measurement and estimation of spectral sensitivity functions for mobile phone cameras. Sensors, 21(15):4985, 2021. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 658, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 714 + ], + "type": "text", + "content": "[26] Lin Zhang, Ying Fu, Yinqiang Zheng, and Hua Huang. Camera spectral sensitivity, illumination and spectral reflectance estimation for a hybrid hyperspectral image capture system. In Proceedings of IEEE International Conference on Image Processing (ICIP), pages 545-549, 2017. 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27260" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 295, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 295, + 117 + ], + "type": "text", + "content": "[27] Jieyu Zhu, Xufen Xie, Ningfang Liao, Zhuangzhi Zhang, Wenmin Wu, and Lvming Lv. Spectral sensitivity estimation of trichromatic camera based on orthogonal test and window filtering. Optics Express, 28(19):28085-28100, 2020. 3" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "27261" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SpectralAR_ Spectral Autoregressive Visual Generation/ddfacecc-ae50-4e90-8053-8ccf2e5c55b7_content_list.json b/2025/SpectralAR_ Spectral Autoregressive Visual Generation/ddfacecc-ae50-4e90-8053-8ccf2e5c55b7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8e30567621ca037aab0a3558ee9341fc730ee380 --- /dev/null +++ b/2025/SpectralAR_ Spectral Autoregressive Visual Generation/ddfacecc-ae50-4e90-8053-8ccf2e5c55b7_content_list.json @@ -0,0 +1,1386 @@ +[ + { + "type": "text", + "text": "SpectralAR: Spectral Autoregressive Visual Generation", + "text_level": 1, + "bbox": [ + 217, + 94, + 779, + 116 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuanhui Huang Weiliang Chen Wenzhao Zheng Yueqi Duan Jie Zhou Jiwen Lu Tsinghua University", + "bbox": [ + 282, + 143, + 715, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "huangyh22@mails.tsinghua.edu.cn; wenzhao.zheng@outlook.com", + "bbox": [ + 151, + 200, + 839, + 217 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/62697571be1fe7f68503368ee738c8ebad782066ae33a320043d4f33785b2b06.jpg", + "image_caption": [ + "Figure 1. We approach autoregressive visual generation from the spectral perspective and propose SpectralAR which converts images into 1D causal sequences with nested spectral tokenization and generates images in a hierarchical coarse-to-fine manner. In the autoregressive process, each generated token improves the quality of the image by introducing new high-frequency components." + ], + "image_footnote": [], + "bbox": [ + 93, + 227, + 906, + 618 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 672, + 328, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Autoregressive visual generation has garnered increasing attention due to its scalability and compatibility with other modalities compared with diffusion models. Most existing methods construct visual sequences as spatial patches for autoregressive generation. However, image patches are inherently parallel, contradicting the causal nature of autoregressive modeling. To address this, we propose a Spectral AutoRegressive (SpectralAR) visual generation framework, which realizes causality for visual sequences from the spectral perspective. Specifically, we first transform an image into ordered spectral tokens with Nested Spectral Tokeniza", + "bbox": [ + 88, + 705, + 485, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion, representing lower to higher frequency components. We then perform autoregressive generation in a coarse-to-fine manner with the sequences of spectral tokens. By considering different levels of detail in images, our SpectralAR achieves both sequence causality and token efficiency without bells and whistles. We conduct extensive experiments on ImageNet-1K for image reconstruction and autoregressive generation, and SpectralAR achieves 3.02 gFID with only 64 tokens and 310M parameters. Project page: https://huang-yh.github.io/spectralar/.", + "bbox": [ + 511, + 675, + 908, + 827 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 843, + 645, + 858 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models [19, 39, 44, 46] have long been the best performing approach to visual generation. Despite their ex", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\boxtimes$ Corresponding author.", + "bbox": [ + 112, + 886, + 253, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15842", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ceptional generation quality, diffusion models still exhibit deficiencies in multimodal modeling and integration of perception and generation. The advent of autoregressive visual generation methods [10, 29, 36, 42, 48, 50, 65] alleviates these limitations and enables better scalability with the next-token prediction paradigm. It first utilizes a visual tokenizer to convert images into tokens and then generates samples in a sequential manner. This advancement supports a variety of emerging applications, including scalable visual generation [16, 57, 59, 69], mixed-modal foundation models [49, 56, 60], and autoregressive world models [21, 73].", + "bbox": [ + 89, + 90, + 483, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite their dominance in language modeling [40, 53], the performance of autoregressive models in visual generation [10, 14] is still inferior to that of diffusion models [35, 39] and non-autoregressive models [4, 66]. This distinction can be attributed to the inherent difference between text and image modalities. Text, invented for human communication, is discrete and sequential, while image data is continuous and invariant to translation [28], indicating the intrinsic equality among image pixels. This equality makes it a critical issue for autoregressive visual generation to convert images into one-dimensional sequential tokens [65]. Methods based on spatial scanning [10, 34, 47, 62] attempt to discretize the image locally into tokens according to patches, and then perform autoregressive generation following a certain order based on their locations. However, the resulting spatial sequence violates the equality among image patches, making it suboptimal for causal autoregressive modeling. Another line of work [4, 31, 66] introduces bidirectional interaction in the generator as a workaround. Nonetheless, they still tokenize images spatially, thus assuming a causal order among image patches. In addition, the bidirectional design may deviate from the conventional autoregressive paradigm, complicating their integration into omni-modal frameworks [49, 56]. In contrast to spatial tokenization, VAR [50] explores transforming the image into multiple scales and producing a sequence by concatenating tokens from ordered scales. Although scale-wise autoregressive generation indeed satisfies the equality of image pixels, it suffers from inferior token efficiency and parallel generation of multiple tokens from the same scale.", + "bbox": [ + 89, + 261, + 483, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce a spectral autoregressive visual generation framework to achieve causal autoregressive modeling and improve token efficiency, as shown in Figure 2. Frequency is an inherent attribute of all types of signals and has become a significant perspective and methodology that complements the spatial-temporal domain [2]. For visual data, the spectral density often conforms to the power-law distribution [6], with low-frequency components representing the overall structure of an image and high-frequency components focusing on the intricate details. This hierarchical coarse-to-fine nature of the correspondence between spectral and spatial domains indicates a se", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/158aed395cad4765210f69b7ac64517cfebfe9a6db00f599a8d6d0c04516dffa.jpg", + "image_caption": [ + "Token Efficiency", + "Figure 2. Comparison between autoregressive visual generation methods. SpectralAR achieves both token efficiency and sequence causality with nested 1D spectral tokens." + ], + "image_footnote": [], + "bbox": [ + 511, + 101, + 903, + 424 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "quential order of images, motivating us to represent images as causal spectral sequences. Specifically, we first transform images into spectral tokens with Nested Spectral Tokenization (NST), which uses varying sequence lengths to represent images across different frequency bands. The causality of the spectral sequences originates from the coarse-to-fine progression characteristic resembling human visual perception and is strengthened with the application of a causal mask. In addition, inspired by the image compression literature [55, 58], we design a non-uniform token-frequency mapping, which allocates more tokens to represent low-frequency components and fewer for high-frequency components. This mapping technique greatly reduces the number of tokens, while maintaining the quality of reconstructed samples. In the autoregressive process, we begin with the token representing the DC component and progressively predict tokens corresponding to higher frequencies conditioned on previous ones. We conduct extensive experiments on the ImageNet-1K dataset for image reconstruction and class-conditional generation. Our SpectralAR demonstrates comparable performance with state-of-the-art methods.", + "bbox": [ + 511, + 469, + 908, + 785 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 799, + 656, + 814 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Autoregressive visual generation. While diffusion models have achieved exceptional performance, they are fundamentally different from the conventional autoregressive framework [49, 56] for cross-modal and cross-task modeling. Therefore, a body of research aims to advance autore-", + "bbox": [ + 511, + 825, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15843", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "gressive models for visual generation. Early efforts [5, 51] perform pixel-level generation in the row-major raster-scan order, followed by VQGAN [10] which transfers to the latent feature space of VQVAE [52] for autoregressive modeling. Subsequent work improves based on VQGAN with multiple scales [42], residual quantization [29], ViT architecture [63] or textual conditions [64]. However, raster-scan generation violates the equality between image pixels as discussed in Section 1, contradicting the causality premise of autoregressive modeling. Recently, VAR [50] proposes the scale-wise autoregressive generation that aims to predict the next-scale token map conditioned on the previous ones. Despite achieving causality with the multi-scale design, VAR predicts multiple tokens with bidirectional attention in each step, thus deviating from the standard autoregressive framework. In contrast, our method realizes causal autoregressive generation from the spectral perspective and still conforms to the unidirectional scheme.", + "bbox": [ + 89, + 90, + 480, + 361 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Efficient image tokenization. Autoregressive generation requires the conversion of images into token sequences, which is often achieved using autoencoders [18, 54]. Patch-based autoencoders [10, 24, 37, 52, 72] tokenize images spatially, where each token corresponds to a certain patch from the original image. Although this paradigm performs well in image reconstruction and diffusion-based generation [39, 44], it is not suitable for autoregressive modeling due to its spatial design. Also, its token length is proportional to the square of image resolution, which might become the bottleneck in multimodal modeling given limited context length [49]. TiTok [66] proposes a 1D tokenizer which reduces the number of tokens to 32 for $256 \\times 256$ images. However, TiTok is trained with only an overall reconstruction objective, and thus the precise meaning of these 1D tokens remains unclear. VAR [50] introduces a multi-scale tokenizer for causal autoregressive image generation. Nonetheless, the multi-scale strategy requires an even greater number of tokens compared to patch-based tokenization methods, further diminishing token efficiency. Our method converts images into 1D causal sequences of spectral tokens and enhances token efficiency by leveraging the long-tail distribution of spectral density in image data.", + "bbox": [ + 91, + 368, + 483, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spectral visual analysis. Spectral analysis [2] has been a common technique in computer vision, complementing the spatial and temporal domains. Representative applications include image enhancement and denoising [15, 68], texture analysis and feature extraction [41], compression and super-resolution [13, 55], visual generation [61, 67], adversarial attacks and defenses [12, 32, 33]. For autoregressive image generation, CART [43] and SIT [11] propose to transform an image into multiple causal sets of tokens with base-detail decomposition and discrete wavelet transform, respectively. However, these methods still adhere to the multi-scale 2D tokenization paradigm similar to", + "bbox": [ + 89, + 719, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VAR, resulting in suboptimal token efficiency and bidirectional attention to predict multiple tokens per autoregressive step. In contrast, SpectralAR leverages the discrete cosine transform to capture the global information of an image, and compress it into a 1D sequence with high efficiency.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Approach", + "text_level": 1, + "bbox": [ + 511, + 179, + 705, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Revisiting Images from the Spectral Domain", + "text_level": 1, + "bbox": [ + 511, + 204, + 888, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Discrete cosine transform. Spectral analysis investigates how complex signals can be represented with simpler basis functions [2], producing a spectral density distribution that represents the magnitude of corresponding basic components. This spectral density distribution is an equivalent representation of the original signal and provides a distinct perspective from the spatial domain depending on the properties of the basis functions. We employ the Discrete Cosine Transform (DCT) [1] to convert images into the spectral domain. The DCT result $\\mathbf{D}$ shares the same shape with the transformed image $\\mathbf{I} \\in \\mathbb{R}^{H \\times W}$ (for simplicity, we omit the channel dimension):", + "bbox": [ + 511, + 226, + 906, + 407 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {D} = \\left\\{F (u, v) \\right\\} _ {u, v = 1} ^ {W, H}, \\quad \\mathbf {I} = \\left\\{f (x, y) \\right\\} _ {x, y = 1} ^ {W, H}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 415, + 906, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Each $F(u,v)$ represents the intensity of the corresponding basis function $g_{u,v}(x,y)$ in the image $\\mathbf{I}$ , which writes:", + "bbox": [ + 511, + 445, + 903, + 474 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ng _ {u, v} (x, y) = \\frac {2 C (u) C (v)}{\\sqrt {H W}} \\cos \\frac {(2 x + 1) u \\pi}{2 W} \\cos \\frac {(2 y + 1) v \\pi}{2 H}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 482, + 906, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $C(u) = 1 / \\sqrt{2}$ if $u = 0$ and $C(u) = 1$ otherwise. This family of basis functions has the following properties: (1) Given $u$ and $v$ , the basis function $g_{u,v}(x,y)$ exhibits a checkerboard-like pattern in the spatial domain, with periods along the x- and y-axis of $2W / u$ and $2H / v$ , respectively. This pattern suggests that the basis function characterizes the rate of variation of images in the spatial domain. (2) The $F(u,v)$ s also form a 2D matrix together, where the top-left corner represents the low-frequency components (small $u, v$ ), while the bottom-right corner corresponds to the high-frequency components (large $u, v$ ).", + "bbox": [ + 511, + 521, + 905, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Causality from the spectral domain. Since low- and high-frequency components describe overall structures and intricate details, respectively, we can decompose an image into a sequence of sub-images $\\{\\mathbf{I}_i^{\\prime}\\}_{i = 1}^{L}$ with increasing levels of detail by applying inverse-DCT on partially masked spectral density distributions $\\{\\mathbf{D}_i^{\\prime}\\}_{i = 1}^{L}$ . With more high-frequency components, the sub-images will gradually transition from blurred to sharp, as shown in Figure 2. This hierarchical coarse-to-fine sequence aligns with human visual perception and artistic drawing, thus enhancing causality.", + "bbox": [ + 511, + 689, + 905, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Efficiency from the spectral domain. As the image compression literature [1] pointed out, the DCT result $\\mathbf{D}$ of images conforms to the power-law distribution with the absolute values of $F(u,v)\\mathrm{s}$ on the top-left corner substantially", + "bbox": [ + 511, + 840, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15844", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2b39acda8d76dcefead60a529f616d9fbe05af9a72d331a7794a11cc7fff2558.jpg", + "image_caption": [ + "Figure 3. Overall pipeline of the proposed SpectralAR. Left: We convert an input image into a 1D causal sequence efficiently with nested spectral tokenization. Each nested sequence is supervised with distinct reconstruction target in a coarse-to-fine manner, which endows each token with an explicit spectral interpretation. We also apply the causal mask to the spectral tokens in the encoder and decoder to enhance the causality. Right: We use the groundtruth sequences from the tokenization process to train an autoregressive generation model." + ], + "image_footnote": [], + "bbox": [ + 109, + 88, + 883, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "larger than the bottom-right corner, indicating that most energy of an image is concentrated in the low-frequency components. In addition, human visual perception is less sensitive to high-frequency signals, which have little influence on the visual quality of images. Based on these distinctions, we can encode the high-frequency components of an image with coarser granularity to improve token efficiency, similar to the JPEG algorithm [55] which saves more than $90\\%$ storage for images by suppressing high-frequency signals.", + "bbox": [ + 88, + 465, + 482, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Nested Spectral Tokenization", + "text_level": 1, + "bbox": [ + 89, + 619, + 352, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overall framework. In contrast to the 2D spatial tokenization that captures the local correlation between image patches, the basis function $g_{u,v}(x,y)$ of DCT encodes a global frequency pattern. Therefore, we convert images into 1D tokens in spectral tokenization to reflect the global nature of the basis functions. We start with the general framework of a 1D image tokenizer [66]. Given an image $\\mathbf{I}$ , we aim to encode it into $N$ discrete vectors $\\mathbf{S}$ , and also reconstruct the original image with $\\mathbf{S}$ . We first patchify the image into $\\mathbf{P} \\in \\mathbb{R}^{hw\\times C}$ , and concatenate the image features with the initial query vectors $\\mathbf{S}_0$ to form $[\\mathbf{P};\\mathbf{S}_0] \\in \\mathbb{R}^{(hw + N)\\times C}$ , where $h,w,C$ denote the resolution of image features and the channel dimension, respectively. We then employ the vision transformer $\\mathcal{E}$ to enable feature extraction and interaction between the image features and 1D query tokens, resulting in the informative 1D representation $\\hat{\\mathbf{S}}$ . In the vector quantizer $\\mathcal{Q}$ , we match these continuous vectors with the", + "bbox": [ + 89, + 643, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "codebook embeddings to derive the discrete representation $\\mathbf{S}$ , which could serve as the groundtruth for the autoregressive training. At last, we append $\\mathbf{S}$ to a set of mask tokens $\\mathbf{M} \\in \\mathbb{R}^{hw \\times C}$ and process them with the decoder network $\\mathcal{D}$ similar to $\\mathcal{E}$ , in order to reconstruct the original image. This overall framework could be formulated as:", + "bbox": [ + 511, + 465, + 906, + 555 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {S}} = \\mathcal {E} ([ \\mathbf {P}; \\mathbf {S} _ {0} ]), \\quad \\mathbf {S} = \\mathcal {Q} (\\hat {\\mathbf {S}}), \\quad \\hat {\\mathbf {I}} = \\mathcal {D} ([ \\mathbf {M}; \\mathbf {S} ]), \\quad (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 568, + 906, + 587 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{I}}$ denotes the reconstructed image. The training objective typically consists of multiple loss functions:", + "bbox": [ + 511, + 599, + 906, + 631 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t o k} = \\| \\hat {\\mathbf {I}} - \\mathbf {I} \\| _ {2} ^ {2} + \\| \\hat {\\mathbf {S}} - \\mathbf {S} \\| _ {2} ^ {2} + \\mathcal {L} _ {P} (\\hat {\\mathbf {I}}, \\mathbf {I}) + \\mathcal {L} _ {A} (\\hat {\\mathbf {I}}, \\mathbf {I}), (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 642, + 906, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_P$ and $\\mathcal{L}_A$ denote the perceptual loss [22, 70] and the adversarial loss [10, 14], respectively.", + "bbox": [ + 511, + 672, + 905, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Nested spectral decoding. Different from the plain 1D tokenization in Figure 2, we aim to represent an image as its spectral decompositions $\\{\\mathbf{I}_i^{\\prime}\\}_{i = 1}^{L}$ for sequence causality, which requires establishing a mapping from 1D tokens to these sub-images. One naive way to achieve this would be dividing S into disjoint subsets and assigning them to model different sub-images independently. Similar to the multi-scale tokenization [50], this strategy would inevitably involve bidirectional interaction and diminish token efficiency because sub-images with finer detail would require increasingly more tokens to represent. In contrast, we propose a nested mapping scheme for efficient tokenization, as shown in the bottom of Figure 3. We first construct a", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15845", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1: Nested Spectral Tokenization Training", + "text_level": 1, + "bbox": [ + 99, + 95, + 429, + 109 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1 Inputs: raw image $\\mathbf{I}$ , initial spectral tokens $\\mathbf{S}_0\\in \\mathbb{R}^{N\\times C}$", + "2 Hyperparameters: spectral levels $\\{\\omega_i|i = 1,\\dots,N\\}$", + "3 $\\mathbf{P} =$ patchify(I), $\\hat{\\mathbf{S}} = \\mathcal{E}([\\mathbf{P};\\mathbf{S}_0])$ $\\mathbf{S} = \\mathcal{Q}(\\hat{\\mathbf{S}})$", + "4 $idx = \\mathrm{random\\_choice}(N)$", + "5 $\\mathbf{S}^{\\prime} = \\mathbf{S}[:idx],\\hat{\\mathbf{I}}^{\\prime} = \\mathcal{D}([\\mathbf{M};\\mathbf{S}^{\\prime}])$", + "$\\mathbf{6}$ $\\mathbf{D} = \\mathrm{DCT}(\\mathbf{I}),\\mathbf{D}' = \\mathbf{D}\\circ \\mathbf{1}_{\\omega_i},\\mathbf{I}' = \\mathrm{DCT}^{-1}(\\mathbf{D}')$", + "7 loss $= \\mathcal{L}_{tok}(\\hat{\\mathbf{l}}',\\mathbf{l}')$", + "8 Return: loss for optimization;" + ], + "bbox": [ + 96, + 112, + 457, + 231 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sequence of sub-images with increasing detail by progressively preserving larger regions in the spectral density $\\mathbf{D}$ :", + "bbox": [ + 89, + 243, + 482, + 273 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} _ {i} ^ {\\prime} = \\mathrm {D C T} ^ {- 1} \\left(\\mathbf {D} _ {i} ^ {\\prime}\\right), \\quad \\mathbf {D} _ {i} ^ {\\prime} = \\mathbf {D} \\circ \\mathbf {1} _ {\\omega_ {i}}, \\quad \\omega_ {i - 1} < \\omega_ {i}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 282, + 480, + 300 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{DCT}^{-1}$ , $\\circ$ , $\\mathbf{1}_{\\omega_i}$ denote the inverse DCT operation, element-wise multiplication and a $H\\times W$ matrix with the top-left corner of size $\\omega_{i}\\times \\omega_{i}$ filled by ones and the remaining parts being zeros. Therefore, the sub-image $\\mathbf{I}_i^{\\prime}$ contains all frequency components present in the sub-image $\\mathbf{I}_{i - 1}^{\\prime}$ . Based on this inclusion property, we can reuse the tokens representing the previous sub-image $\\mathbf{I}_{i - 1}^{\\prime}$ to represent the next sub-image $\\mathbf{I}_i^\\prime$ . To avoid bidirectional attention, we make $L$ equal to $N$ so that each token $\\mathbf{s}_i$ in the sequence S corresponds to a unique sub-image $\\mathbf{I}_i^\\prime$ :", + "bbox": [ + 89, + 309, + 483, + 462 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {I}} _ {i} ^ {\\prime} = \\mathcal {D} ([ \\mathbf {M}; \\mathbf {s} _ {1}, \\mathbf {s} _ {2}, \\dots , \\mathbf {s} _ {i} ]), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 472, + 480, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{I}}_i^{\\prime}$ is the reconstruction of sub-image $\\mathbf{I}_i^\\prime$ given the nested 1D sequence $\\{\\mathbf{s}_1,\\mathbf{s}_2,\\dots ,\\mathbf{s}_i\\}$ . Nested spectral decoding compresses an image into a causal 1D sequence where each token $\\mathbf{s}_i$ corresponds to a disjoint set of frequencies and achieves token efficiency by reusing previous tokens.", + "bbox": [ + 89, + 500, + 482, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Non-uniform token-frequency mapping. To further enhance token efficiency, we introduce the non-uniform token-frequency mapping technique. Since high-frequency components have low magnitude and minimal impact on the visual quality of images, we can encode them with coarser granularity compared to the low-frequency counterparts. We achieve this by reducing the interval between $\\omega_{i-1}$ and $\\omega_i$ when $i$ is small, and increase it otherwise:", + "bbox": [ + 89, + 577, + 483, + 696 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\omega_ {i} - \\omega_ {i - 1} \\leq \\omega_ {i + 1} - \\omega_ {i}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 709, + 480, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which demonstrates the general case. This non-uniform mapping allocates later tokens to broader frequency ranges, enabling precise modeling of crucial low-frequency components while efficiently representing high-frequency details.", + "bbox": [ + 89, + 734, + 482, + 795 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Spectral causal mask. Although the sequence $\\mathbf{S}$ supervised with (5)(6) already exhibits a certain degree of causality, the encoding and decoding processes, i.e. $\\mathcal{E}$ and $\\mathcal{D}$ , are still bidirectional, which can lead to information leakage from high-frequency to low-frequency components. Therefore, we propose applying causal masks to the spectral tokens $\\mathbf{S}$ in both the encoder $\\mathcal{E}$ and the decoder $\\mathcal{D}$ , as shown", + "bbox": [ + 89, + 795, + 483, + 898 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ff0613b11d160fce07f9b3d309ffbb601833ab23ce38e9895fd698cb9c338d3b.jpg", + "table_caption": [ + "Table 1. Correlation between tokens of different autoregressive paradigms. We use linear correlation as a proxy metric for the causality of sequences. The spectral sequence demonstrates better causality compared with other methods." + ], + "table_footnote": [], + "table_body": "
Correlation TypeRaster-scanScale-wiseSpectral
R2avg(t2; t1)0.4710.8890.916
R2avg(t3; t1, t2)0.3660.9530.977
R2avg(t4; t1, t2, t3)0.5250.9430.994
Average0.4540.9280.962
", + "bbox": [ + 517, + 145, + 903, + 234 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6f0262dc63c5c27aadfc558f7b636ba1aa0a6311fa06b784b9805f52a5c08d96.jpg", + "image_caption": [ + "Raster-scan", + "Figure 4. Frequency of each element having the highest correlation with another one. It represents the source of reference information for predicting the next token. The raster-scan method exhibits excessive dependency on boundaries of image patches." + ], + "image_footnote": [], + "bbox": [ + 514, + 239, + 638, + 335 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a6881fb4872c5f10b14367c193ba56928c19b9afcc405379f1b6e24877c9be6a.jpg", + "image_caption": [ + "Scale-wise" + ], + "image_footnote": [], + "bbox": [ + 638, + 239, + 758, + 333 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3a959dd944bf178c2326bffa2e8707fe4bb684fa09eb448adcf762c190c01129.jpg", + "image_caption": [ + "Spectral" + ], + "image_footnote": [], + "bbox": [ + 761, + 239, + 880, + 335 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/16229f95e468c3d09cb5cefecca1954ce71d5bfc784494c544366abb177f5dd5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 880, + 239, + 903, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in Figure 3. This spectral causal mask restricts each token s to only attend to tokens that represent lower frequencies, enhancing causality from the architectural perspective. We outline the training procedure in Algorithm 1.", + "bbox": [ + 511, + 405, + 906, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Spectral Autoregressive Generation", + "text_level": 1, + "bbox": [ + 511, + 474, + 821, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Autoregressive modeling has gained prominence in computer vision for its scalability, generalization, and effectiveness across multimodal tasks [25, 49, 56]. While conventional autoregressive generation follows a spatial raster-scan order [10, 51], we propose a hierarchical coarse-to-fine approach in the spectral domain to enhance causality, as shown in the right side of Figure 3. We start with the general framework of autoregressive modeling:", + "bbox": [ + 511, + 497, + 905, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {i + 1} = \\mathcal {M} \\left(\\mathbf {t} _ {1}, \\mathbf {t} _ {2}, \\dots , \\mathbf {t} _ {i}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 631, + 903, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{T} = \\{\\mathbf{t}_i\\}_{i=1}^N$ is a sequence of quantized tokens, and $\\mathbf{p}_{i+1}$ , $\\mathcal{M}$ denote the probability logits for the $(i+1)$ th token and the autoregressive model, respectively. The autoregressive process (8) assumes that the generation of token $\\mathbf{t}_{i+1}$ depends solely on the previous ones. The spatial autoregressive paradigm violates this premise because of the equality of image pixels (as discussed in Section 1). In contrast, we take the spectral tokens $\\mathbf{S}$ from the nested spectral tokenziation as the autoregressive targets $\\mathbf{T}$ . Since the spectral tokens are trained in a nested manner to reconstruct sub-images of increasing levels of detail as in (5)(6), each spectral token $\\mathbf{s}_i$ is expected to enhance the quality of the sub-image $\\mathbf{I}_{i-1}'$ represented by previous tokens from the spectral domain. This progressive refinement process aligns with human visual perception and artistic painting, both of which start with the overall structure and gradually focus on", + "bbox": [ + 511, + 657, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15846", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5fa04c165f971d3fa870dae4f6302a1698c646a463373c7ba1d33c8d905d0bd6.jpg", + "table_caption": [ + "Table 2. Comparison between generative models on class-conditional ImageNet ${256} \\times {256}$ benchmark. “ $\\downarrow$ ” or “↑” indicate lower or higher values are better. \"Token\": the number of tokens used in transformer architectures. \"Step\": the number of model runs needed to generate an image. We compute our wall-clock inference time and scale the Time for other methods accordingly. $\\dagger$ : trained on larger datasets including OpenImages [27]. ‡: implemented with the official tokenizer weight [66] and the scripts from VAR." + ], + "table_footnote": [], + "table_body": "
TypeModelrFID↓gFID↓IS↑Pre↑Rec↑#Token#Para#StepTime
GANBigGAN [3]75.246.95224.50.890.38112M1-
GANGigaGAN [23]-3.45225.50.840.61N.A.569M1-
GANStyleGan-XL [45]7.062.30265.10.780.53166M10.4
Diff.ADM [8]125.7810.94101.00.690.63N.A.554M250235
Diff.CDM [20]-4.88158.7--N.A.-8100-
Diff.LDM-4-G [44]0.27†3.60247.7--N.A.400M250-
Diff.DiT-L/2 [39]0.62†5.02167.20.750.57256458M25043
Diff.DiT-XL/2 [39]0.62†2.27278.20.830.57256675M25063
Diff.L-DiT-3B [71]-2.10304.40.820.602563.0B250>63
Diff.L-DiT-7B [71]-2.28316.20.830.582567.0B250>63
Mask.MaskGIT [4]2.286.18182.10.800.51256227M80.7
Mask.RCG (cond.) [30]-3.49215.5---502M202.7
Mask.TiTok-B64 [66]1.702.48214.7--64177M80.4
2D ScanVQVAE-2 [42]-31.11~450.360.57N.A.13.5B5120-
2D ScanVQGAN [10]7.9415.7874.3--N.A.1.4B25634
2D ScanViTVQ [63]1.284.17175.1--10241.7B1024>34
2D ScanRQTran. [29]3.207.55134.0--64,43.8B6829
VARVAR-d16 [50]3.30274.40.840.51310M100.6
VARVAR-d20 [50]2.57302.60.830.56600M100.7
VARVAR-d24 [50]2.09312.90.820.591.0B100.8
VARVAR-d30 [50]1.92323.10.820.592.0B101.4
1D ARTiTok-B64-d16‡ [66]1.706.30190.10.850.4764310M641
1D ARSpectralAR-d164.033.02282.20.810.5564310M641
1D ARSpectralAR-d204.032.49305.4--64600M641.2
1D ARSpectralAR-d244.032.13307.7--641.0B641.4
1D ARSpectralAR-d16-p44.033.13276.1--16×4310M160.4
", + "bbox": [ + 112, + 146, + 883, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "details. This similarity qualitatively validates the rationale for performing causal autoregressive generation in the spectral domain. We further provide some quantitative analysis through a toy experiment in Section 4.2.", + "bbox": [ + 89, + 602, + 482, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Potential applications. The frequencies represented by the token $\\mathbf{s}_i$ become higher as $i$ increases, while its influence on the image quality diminishes accordingly (check Section 3.1 for details). Therefore, we can control the visual quality of sampled images by adjusting the length of generated sequences, similar to the image compression algorithms [55]. In addition, we can achieve super-resolution by dividing images into disjoint parts smaller than $H\\times W$ , and conducting individual spectral autoregressive generation on each part. We provide further results in Section 4.4.", + "bbox": [ + 89, + 662, + 482, + 814 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 823, + 223, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Dataset and Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 848, + 403, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train and evaluate our SpectralAR on the ImageNet1K [7] benchmark, which contains 1,281,167 and 50,000", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "images for training and validation, respectively. We train the tokenizer and generator on the training split. We evaluate the reconstruction performance on the validation set with reconstruction Fréchet inception distance [17] (rFID), and the generation results with generation FID (gFID) using pre-computed statistics and scripts from ADM [8].", + "bbox": [ + 511, + 602, + 906, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For tokenizer training, we follow the exactly same settings of TiTok [66] for a fair comparison. We also employ the two-stage training strategy with proxy codes [4, 66]. We use the ViT-B [9] as the encoder and decoder, and set the number of spectral tokens as $N = 64$ in our main experiments and the sequence $\\omega$ s for $256 \\times 256$ images as:", + "bbox": [ + 511, + 693, + 906, + 785 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\omega_ {i} = \\left\\{ \\begin{array}{l l} i, & \\text {i f} i \\in (0, 3 2 ], \\\\ 2 i - 3 2, & \\text {i f} i \\in (3 2, 4 8 ], \\\\ 1 2 i - 5 1 2, & \\text {i f} i \\in (4 8, 6 4 ]. \\end{array} \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 796, + 906, + 844 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For generator training, we adopt the same architecture and training recipe as VAR [50], which leverages a GPT-2-like transformer architecture [40] for autoregressive modeling.", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15847", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/730d1fcfc5fc12876dca33d6b4a5f03440c9ed2823c9e73856251edf6e8cba57.jpg", + "image_caption": [ + "Figure 5. Visualizations of generated $256 \\times 256$ samples on ImageNet-1K. These samples cover a wide range of categories and styles, demonstrating the ability of SpectralAR to generate both diverse and high-quality images." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 903, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Proof of Concept", + "text_level": 1, + "bbox": [ + 89, + 676, + 261, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct a proof-of-concept experiment to validate the causality of images in the spectral domain. We use the validation set of the CIFAR-100 dataset [26] as a proxy for the distribution of natural images, which contains 10,000 $32 \\times 32$ images from 100 semantic classes. We aim to measure the linear correlation $(R^2)$ [38] between one token $\\mathbf{t}_i$ and its previous ones $\\{\\mathbf{t}_j\\}_{j=1}^{i-1}$ as a metric for the causality of the sequence $\\mathbf{T}$ . We consider three typical autoregressive paradigms, including raster-scan [10], scale-wise [50] and spectral methods which construct sequence $\\mathbf{T}$ in different ways. Specifically, we divide each image spatially into four patches of size $16 \\times 16$ and rearrange them in raster-scan order for the first variant. For the scale-wise method, we", + "bbox": [ + 89, + 703, + 485, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "downsample each image into sizes of $8 \\times 8$ , $16 \\times 16$ , $24 \\times 24$ and $32 \\times 32$ , respectively. We set the frequency thresholds to 8, 16, 24, 32 and generate the sub-images following (5) for the spectral sequence. Since each token $\\mathbf{t}$ is a high dimensional tensor, we calculate $R^2$ between every element of $\\mathbf{t}_i$ and every element of $\\{\\mathbf{t}_j\\}_{j=1}^{i-1}$ . For each element of $\\mathbf{t}_i$ , we then find the element in $\\{\\mathbf{t}_j\\}_{j=1}^{i-1}$ with the highest correlation, and take an average as the final measurement:", + "bbox": [ + 511, + 678, + 908, + 803 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nR _ {a v g} ^ {2} = \\operatorname {A V G} _ {m} \\left(\\max _ {n} R _ {m, n} ^ {2}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 811, + 906, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $m, n$ denote element number of $\\mathbf{t}_i$ and $\\{\\mathbf{t}_j\\}_{j=1}^{i-1}$ , respectively, and $R_{m,n}^2$ is the correlation between the $m$ th and $n$ th elements. We report the quantitative results in Table 1, where the scale-wise and spectral sequences demonstrate", + "bbox": [ + 511, + 837, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15848", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e4a71f84f58b457302b3e323147862330e36e6528379e33fe549c6bd4ef65dd5.jpg", + "table_caption": [ + "Table 3. Applications of SepctralAR. Super-reso., Trunc. represent super-resolution and truncated, respectively." + ], + "table_footnote": [], + "table_body": "
App. TypeModelFID↓IS↑
Super-reso.Upsample3.09286.6
SpectralAR-Stride2.93276.4
SpectralAR-Patch14.76170.0
Trunc.SpectralAR-Trunc.53.34271.5
SpectralAR-Trunc.106.65211.4
SpectralAR-Trunc.1527.6991.69
SpectralAR-Trunc.03.02282.2
", + "bbox": [ + 94, + 119, + 478, + 258 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "considerably higher correlation than the raster-scan counterpart. This could be attributed to the coarse-to-fine nature of the former two paradigms, while the raster scan method lacks adequate reference information to predict the next image patch, as shown in Figure 4.", + "bbox": [ + 89, + 265, + 482, + 340 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Main Results", + "text_level": 1, + "bbox": [ + 89, + 349, + 230, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We report the performance of SpectralAR on the class-conditional ImageNet-1K [7] $256 \\times 256$ generation benchmark in Table 2. We also implement an autoregressive version of TiTok [66] by using the official tokenizer weight and scripts from VAR [50] for a fair comparison. The reconstruction performance of SpectralAR (4.03 rFID) is inferior compared with TiTok because SpectralAR requires to reconstruct the sub-images corresponding to different frequencies with different lengths of tokens, which is much more difficult compared with the overall reconstruction target in TiTok. Despite the lower reconstruction score, SpectralAR-d16 still outperforms TiTok-B64-d16 and VAR-d16 in autoregressive generation with a clear margin due to better sequence causality which eases autoregressive learning. Furthermore, SpectralAR also achieves better or comparable performance against VAR under the $d20$ and $d24$ settings, respectively. In addition, SpectralAR uses only 64 tokens in both reconstruction and generation, demonstrating superior token efficiency compared to VAR [50] and 2D scan-based methods. We also visualize the generated samples in Figure 5, which shows the diversity and quality of the generation process of SpectralAR. In addition, Figure 1 highlights the hierarchical coarse-to-fine refinement of the images, while SpectralAR generates more tokens in an autoregressive way.", + "bbox": [ + 89, + 371, + 482, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Applications", + "text_level": 1, + "bbox": [ + 89, + 757, + 223, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we provide a quantitative analysis for the potential applications of SpectralAR. For super-resolution, we conduct our experiments based on the $256 \\times 256$ images generated by SpectralAR- $d16$ in Table 2, and directly upsample them to $512 \\times 512$ as the baseline. We construct 4 sub-images using stripped and patch-based methods for SpectralAR-Stride and -Patch, respectively. We then upsample the 4 sub-images to $256 \\times 256$ and use SpectralAR", + "bbox": [ + 89, + 780, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/959f2a27fcfb3240f8bd334c97a8238301d8d5f9cc535ab95bc33c3f2e290ee9.jpg", + "table_caption": [ + "Table 4. Ablation on design choices. Spectral Supervision means using sub-image supervision across different frequency bands. Causal Mask and non-uniform mapping refers to the spectral causal mask and the frequency-token mapping, respectively." + ], + "table_footnote": [], + "table_body": "
Spectral SupervisionCausal MaskNon-uniform MappingFID↓IS↑
×××6.30190.1
✓××5.64255.1
✓✓×3.49222.6
✓✓✓3.02282.2
", + "bbox": [ + 517, + 147, + 903, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to refine them in the spectral domain, and finally reassemble them to generate the final result. According to Table 3, SpectralAR can indeed serve as a spectrum completer for the super-resolution task. In addition, we also experiment with truncated autoregressive generation, where we discard the last few tokens. For example, we discard the last 5 tokens in SpectralAR-Trunc.5 in Table 3. The generation performance worsens slowly when the number of truncated tokens is fewer than 10, and therefore it is possible to further improve token efficiency through truncation according to the requirement for generation quality.", + "bbox": [ + 511, + 258, + 906, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 431, + 666, + 449 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct ablation study to validate the effectiveness of our design choices in Table 4. The first row corresponds to the baseline TiTok implementation for autoregressive generation. Note that the spectral supervision alone could improve FID compared with the TiTok counterpart. This is because the spectral design enhances sequence causality compared with the overall reconstruction target which ignores the correlation between tokens and thus complicates the autoregressive modeling process. The spectral causal mask further enhances performance by improving causality in the encoding and decoding architecture. And the non-uniform token-frequency mapping technique guides the model to focus more on the crucial low-frequency components while representing the high-frequency components efficiently, thus further improving performance.", + "bbox": [ + 511, + 455, + 908, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 694, + 633, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have proposed the spectral autoregressive visual generation method for both causal and efficient autoregressive modeling of image data. Specifically, we first convert images into 1D sequences with nested spectral tokenization. In addition, we have adopted causal masks for spectral tokens in the encoder and decoder to further enhance causality from the architectural perspective. We have also designed a non-uniform token-frequency mapping with emphasis on the low-frequency components in order to improve token efficiency. On the ImageNet-1K generation benchmark, our SpectralAR achieves superior performance compared with other autoregressive generation methods.", + "bbox": [ + 511, + 719, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15849", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China under Grant 62125603, Grant 62321005, Grant 62336004, and Grant 62441616, and in part by the Beijing Natural Science Foundation under Grant No. L247009.", + "bbox": [ + 89, + 114, + 483, + 185 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 213, + 187, + 229 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Nasir Ahmed, T. Natarajan, and Kamisetty R Rao. Discrete cosine transform. IEEE transactions on Computers, 100(1): 90-93, 2006. 3", + "[2] Jean Baptiste Joseph Baron Fourier et al. The analytical theory of heat. Courier Corporation, 2003. 2, 3", + "[3] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 6", + "[4] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In CVPR, pages 11315-11325, 2022. 2, 6", + "[5] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In ICML, pages 1691-1703. PMLR, 2020. 3", + "[6] Aaron Clauset, Cosma Rohilla Shalizi, and Mark EJ Newman. Power-law distributions in empirical data. SIAM review, 51(4):661-703, 2009. 2", + "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009. 6, 8", + "[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NIPS, 34:8780-8794, 2021. 6", + "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6", + "[10] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, pages 12873-12883, 2021. 2, 3, 4, 5, 6, 7", + "[11] Carlos Esteves, Mohammed Suhail, and Ameesh Makadia. Spectral image tokenizer. arXiv preprint arXiv:2412.09607, 2024.3", + "[12] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In ICML, pages 3247-3258. PMLR, 2020. 3", + "[13] Manuel Fritsche, Shuhang Gu, and Radu Timofte. Frequency separation for real-world super-resolution. In ICCVW, pages 3599-3608. IEEE, 2019. 3", + "[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. NIPS, 27, 2014. 2, 4" + ], + "bbox": [ + 93, + 239, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Hayit Greenspan, Charles H Anderson, and Sofia Akber. Image enhancement by nonlinear extrapolation in frequency space. TIP, 9(6):1035-1048, 2000. 3", + "[16] Jian Han, Jinlai Liu, Yi Jiang, Bin Yan, Yuqi Zhang, Zehuan Yuan, Bingyue Peng, and Xiaobing Liu. Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis. arXiv preprint arXiv:2412.04431, 2024. 2", + "[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NIPS, 30, 2017. 6", + "[18] Geoffrey E Hinton and Ruslan R Salakhutdinov. Reducing the dimensionality of data with neural networks. science, 313(5786):504-507, 2006. 3", + "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NIPS, 33:6840-6851, 2020. 1", + "[20] Jonathan Ho, Chitwan Saharia, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. JMLR, 23(47):1-33, 2022. 6", + "[21] Yuanhui Huang, Wenzhao Zheng, Yuan Gao, Xin Tao, Pengfei Wan, Di Zhang, Jie Zhou, and Jiwen Lu. Owl1: Omni world model for consistent long video generation. arXiv preprint arXiv:2412.09600, 2024. 2", + "[22] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, pages 694–711. Springer, 2016. 4", + "[23] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In CVPR, pages 10124-10134, 2023. 6", + "[24] Diederik P Kingma, Max Welling, et al. Auto-encoding variational bayes, 2013. 3", + "[25] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Grant Schindler, Rachel Hornung, Vighnesh Birodkar, Jimmy Yan, Ming-Chang Chiu, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 5", + "[26] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 7", + "[27] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 128(7):1956-1981, 2020. 6", + "[28] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 2", + "[29] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In CVPR, pages 11523-11532, 2022. 2, 3, 6", + "[30] Tianhong Li, Dina Katabi, and Kaiming He. Self-conditioned image generation via generating representations. CoRR, 2023. 6" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15850", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. NIPS, 37:56424-56445, 2025. 2", + "[32] Yuyang Long, Qilong Zhang, Boheng Zeng, Lianli Gao, Xianglong Liu, Jian Zhang, and Jingkuan Song. Frequency domain model augmentation for adversarial attack. In ECCV, pages 549-566. Springer, 2022. 3", + "[33] Cheng Luo, Qinliang Lin, Weicheng Xie, Bizhu Wu, Jinheng Xie, and Linlin Shen. Frequency-driven imperceptible adversarial attack on semantic similarity. In CVPR, pages 15315-15324, 2022. 3", + "[34] Zhuoyan Luo, Fengyuan Shi, Yixiao Ge, Yujiu Yang, Limin Wang, and Ying Shan. Open-magvit2: An open-source project toward democratizing auto-regressive visual generation. arXiv preprint arXiv:2409.04410, 2024. 2", + "[35] Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. In ECCV, pages 23-40. Springer, 2024. 2", + "[36] Xiaoxiao Ma, Mohan Zhou, Tao Liang, Yalong Bai, Tiejun Zhao, Huaian Chen, and Yi Jin. Star: Scale-wise text-to-image generation via auto-regressive representations. arXiv preprint arXiv:2406.10797, 2024. 2", + "[37] Fabian Mentzer, David Minnen, Eirikur Agustsson, and Michael Tschannen. Finite scalar quantization: Vq-vae made simple. arXiv preprint arXiv:2309.15505, 2023. 3", + "[38] Daniel J Ozer. Correlation and the coefficient of determination. Psychological bulletin, 97(2):307, 1985. 7", + "[39] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 1, 2, 3, 6", + "[40] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 2, 6", + "[41] Yongming Rao, Wenliang Zhao, Zheng Zhu, Jiwen Lu, and Jie Zhou. Global filter networks for image classification. Advances in neural information processing systems, 34:980-993, 2021. 3", + "[42] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. NIPS, 32, 2019. 2, 3, 6", + "[43] Siddharth Roheda. Cart: Compositional auto-regressive transformer for image generation. arXiv preprint arXiv:2411.10180, 2024. 3", + "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 3, 6", + "[45] Axel Sauer, Katja Schwarz, and Andreas Geiger. Styleganx1: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH 2022 conference proceedings, pages 1-10, 2022. 6", + "[46] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1", + "[47] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 2", + "[48] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 2", + "[49] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 2, 3, 5", + "[50] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. NIPS, 37:84839-84865, 2025. 2, 3, 4, 6, 7, 8", + "[51] Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelCNN decoders. NIPS, 29, 2016. 3, 5", + "[52] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NIPS, 30, 2017. 3", + "[53] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NIPS, 30, 2017. 2", + "[54] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In ICML, pages 1096-1103, 2008. 3", + "[55] Gregory K Wallace. The jpeg still picture compression standard. Communications of the ACM, 34(4):30-44, 1991. 2, 3, 4, 6", + "[56] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 2, 5", + "[57] Yuqing Wang, Tianwei Xiong, Daquan Zhou, Zhijie Lin, Yang Zhao, Bingyi Kang, Jiashi Feng, and Xihui Liu. Loong: Generating minute-level long videos with autoregressive language models. arXiv preprint arXiv:2410.02757, 2024. 2", + "[58] Thomas Wiegand, Gary J Sullivan, Gisle Bjontegaard, and Ajay Luthra. Overview of the h. 264/avc video coding standard. TCSVT, 13(7):560-576, 2003. 2", + "[59] Chengyue Wu, Xiaokang Chen, Zhiyu Wu, Yiyang Ma, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, Chong Ruan, et al. Janus: Decoupling visual encoding for unified multimodal understanding and generation. arXiv preprint arXiv:2410.13848, 2024. 2", + "[60] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 2", + "[61] Mengping Yang, Zhe Wang, Ziqiu Chi, and Wenyi Feng. Wavegan: Frequency-aware gan for high-fidelity few-shot image generation. In ECCV, pages 1-17. Springer, 2022. 3", + "[62] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15851", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[63] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 3, 6", + "[64] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 3", + "[65] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Randomized autoregressive visual generation. arXiv preprint arXiv:2411.00776, 2024. 2", + "[66] Qihang Yu, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An image is worth 32 tokens for reconstruction and generation. NIPS, 37: 128940-128966, 2025. 2, 3, 4, 6, 8", + "[67] Yingchen Yu, Fangneng Zhan, Shijian Lu, Jianxiong Pan, Feiying Ma, Xuansong Xie, and Chunyan Miao. Wavefill: A wavelet-based generation network for image inpainting. In ICCV, pages 14114-14123, 2021. 3", + "[68] Huanjing Yue, Xiaoyan Sun, Jingyu Yang, and Feng Wu. Cid: Combined image denoising in spatial and frequency domains using web images. In CVPR, pages 2933-2940, 2014. 3", + "[69] Qian Zhang, Xiangzi Dai, Ninghua Yang, Xiang An, Ziyong Feng, and Xingyu Ren. Var-clip: Text-to-image generator with visual auto-regressive modeling. arXiv preprint arXiv:2408.01181, 2024. 2", + "[70] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 4", + "[71] Renrui Zhang, Jiaming Han, Chris Liu, Peng Gao, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, and Yu Qiao. Llama-adapter: Efficient finetuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 6", + "[72] Chuanxia Zheng, Tung-Long Vuong, Jianfei Cai, and Dinh Phung. Movq: Modulating quantized vectors for high-fidelity image generation. NIPS, 35:23412-23425, 2022. 3", + "[73] Wenzhao Zheng, Weiliang Chen, Yuanhui Huang, Borui Zhang, Yueqi Duan, and Jiwen Lu. Occworld: Learning a 3d occupancy world model for autonomous driving. In ECCV, pages 55-72. Springer, 2024. 2" + ], + "bbox": [ + 91, + 90, + 480, + 727 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15852", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SpectralAR_ Spectral Autoregressive Visual Generation/ddfacecc-ae50-4e90-8053-8ccf2e5c55b7_model.json b/2025/SpectralAR_ Spectral Autoregressive Visual Generation/ddfacecc-ae50-4e90-8053-8ccf2e5c55b7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6fcac68ecc94527c885f357b3ae83cf9cb2cbabb --- /dev/null +++ b/2025/SpectralAR_ Spectral Autoregressive Visual Generation/ddfacecc-ae50-4e90-8053-8ccf2e5c55b7_model.json @@ -0,0 +1,2279 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.095, + 0.78, + 0.117 + ], + "angle": 0, + "content": "SpectralAR: Spectral Autoregressive Visual Generation" + }, + { + "type": "text", + "bbox": [ + 0.284, + 0.145, + 0.717, + 0.201 + ], + "angle": 0, + "content": "Yuanhui Huang Weiliang Chen Wenzhao Zheng Yueqi Duan Jie Zhou Jiwen Lu Tsinghua University" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.202, + 0.84, + 0.218 + ], + "angle": 0, + "content": "huangyh22@mails.tsinghua.edu.cn; wenzhao.zheng@outlook.com" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.228, + 0.907, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.62, + 0.908, + 0.662 + ], + "angle": 0, + "content": "Figure 1. We approach autoregressive visual generation from the spectral perspective and propose SpectralAR which converts images into 1D causal sequences with nested spectral tokenization and generates images in a hierarchical coarse-to-fine manner. In the autoregressive process, each generated token improves the quality of the image by introducing new high-frequency components." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.674, + 0.33, + 0.691 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.707, + 0.486, + 0.875 + ], + "angle": 0, + "content": "Autoregressive visual generation has garnered increasing attention due to its scalability and compatibility with other modalities compared with diffusion models. Most existing methods construct visual sequences as spatial patches for autoregressive generation. However, image patches are inherently parallel, contradicting the causal nature of autoregressive modeling. To address this, we propose a Spectral AutoRegressive (SpectralAR) visual generation framework, which realizes causality for visual sequences from the spectral perspective. Specifically, we first transform an image into ordered spectral tokens with Nested Spectral Tokeniza" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.676, + 0.909, + 0.828 + ], + "angle": 0, + "content": "tion, representing lower to higher frequency components. We then perform autoregressive generation in a coarse-to-fine manner with the sequences of spectral tokens. By considering different levels of detail in images, our SpectralAR achieves both sequence causality and token efficiency without bells and whistles. We conduct extensive experiments on ImageNet-1K for image reconstruction and autoregressive generation, and SpectralAR achieves 3.02 gFID with only 64 tokens and 310M parameters. Project page: https://huang-yh.github.io/spectralar/." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.844, + 0.646, + 0.859 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Diffusion models [19, 39, 44, 46] have long been the best performing approach to visual generation. Despite their ex" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.887, + 0.254, + 0.901 + ], + "angle": 0, + "content": "\\(\\boxtimes\\) Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15842" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.258 + ], + "angle": 0, + "content": "ceptional generation quality, diffusion models still exhibit deficiencies in multimodal modeling and integration of perception and generation. The advent of autoregressive visual generation methods [10, 29, 36, 42, 48, 50, 65] alleviates these limitations and enables better scalability with the next-token prediction paradigm. It first utilizes a visual tokenizer to convert images into tokens and then generates samples in a sequential manner. This advancement supports a variety of emerging applications, including scalable visual generation [16, 57, 59, 69], mixed-modal foundation models [49, 56, 60], and autoregressive world models [21, 73]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.262, + 0.485, + 0.716 + ], + "angle": 0, + "content": "Despite their dominance in language modeling [40, 53], the performance of autoregressive models in visual generation [10, 14] is still inferior to that of diffusion models [35, 39] and non-autoregressive models [4, 66]. This distinction can be attributed to the inherent difference between text and image modalities. Text, invented for human communication, is discrete and sequential, while image data is continuous and invariant to translation [28], indicating the intrinsic equality among image pixels. This equality makes it a critical issue for autoregressive visual generation to convert images into one-dimensional sequential tokens [65]. Methods based on spatial scanning [10, 34, 47, 62] attempt to discretize the image locally into tokens according to patches, and then perform autoregressive generation following a certain order based on their locations. However, the resulting spatial sequence violates the equality among image patches, making it suboptimal for causal autoregressive modeling. Another line of work [4, 31, 66] introduces bidirectional interaction in the generator as a workaround. Nonetheless, they still tokenize images spatially, thus assuming a causal order among image patches. In addition, the bidirectional design may deviate from the conventional autoregressive paradigm, complicating their integration into omni-modal frameworks [49, 56]. In contrast to spatial tokenization, VAR [50] explores transforming the image into multiple scales and producing a sequence by concatenating tokens from ordered scales. Although scale-wise autoregressive generation indeed satisfies the equality of image pixels, it suffers from inferior token efficiency and parallel generation of multiple tokens from the same scale." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduce a spectral autoregressive visual generation framework to achieve causal autoregressive modeling and improve token efficiency, as shown in Figure 2. Frequency is an inherent attribute of all types of signals and has become a significant perspective and methodology that complements the spatial-temporal domain [2]. For visual data, the spectral density often conforms to the power-law distribution [6], with low-frequency components representing the overall structure of an image and high-frequency components focusing on the intricate details. This hierarchical coarse-to-fine nature of the correspondence between spectral and spatial domains indicates a se" + }, + { + "type": "image_caption", + "bbox": [ + 0.783, + 0.089, + 0.885, + 0.1 + ], + "angle": 0, + "content": "Token Efficiency" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.102, + 0.905, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.426, + 0.907, + 0.468 + ], + "angle": 0, + "content": "Figure 2. Comparison between autoregressive visual generation methods. SpectralAR achieves both token efficiency and sequence causality with nested 1D spectral tokens." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.909, + 0.786 + ], + "angle": 0, + "content": "quential order of images, motivating us to represent images as causal spectral sequences. Specifically, we first transform images into spectral tokens with Nested Spectral Tokenization (NST), which uses varying sequence lengths to represent images across different frequency bands. The causality of the spectral sequences originates from the coarse-to-fine progression characteristic resembling human visual perception and is strengthened with the application of a causal mask. In addition, inspired by the image compression literature [55, 58], we design a non-uniform token-frequency mapping, which allocates more tokens to represent low-frequency components and fewer for high-frequency components. This mapping technique greatly reduces the number of tokens, while maintaining the quality of reconstructed samples. In the autoregressive process, we begin with the token representing the DC component and progressively predict tokens corresponding to higher frequencies conditioned on previous ones. We conduct extensive experiments on the ImageNet-1K dataset for image reconstruction and class-conditional generation. Our SpectralAR demonstrates comparable performance with state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.8, + 0.658, + 0.815 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Autoregressive visual generation. While diffusion models have achieved exceptional performance, they are fundamentally different from the conventional autoregressive framework [49, 56] for cross-modal and cross-task modeling. Therefore, a body of research aims to advance autore-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15843" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.362 + ], + "angle": 0, + "content": "gressive models for visual generation. Early efforts [5, 51] perform pixel-level generation in the row-major raster-scan order, followed by VQGAN [10] which transfers to the latent feature space of VQVAE [52] for autoregressive modeling. Subsequent work improves based on VQGAN with multiple scales [42], residual quantization [29], ViT architecture [63] or textual conditions [64]. However, raster-scan generation violates the equality between image pixels as discussed in Section 1, contradicting the causality premise of autoregressive modeling. Recently, VAR [50] proposes the scale-wise autoregressive generation that aims to predict the next-scale token map conditioned on the previous ones. Despite achieving causality with the multi-scale design, VAR predicts multiple tokens with bidirectional attention in each step, thus deviating from the standard autoregressive framework. In contrast, our method realizes causal autoregressive generation from the spectral perspective and still conforms to the unidirectional scheme." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.369, + 0.485, + 0.716 + ], + "angle": 0, + "content": "Efficient image tokenization. Autoregressive generation requires the conversion of images into token sequences, which is often achieved using autoencoders [18, 54]. Patch-based autoencoders [10, 24, 37, 52, 72] tokenize images spatially, where each token corresponds to a certain patch from the original image. Although this paradigm performs well in image reconstruction and diffusion-based generation [39, 44], it is not suitable for autoregressive modeling due to its spatial design. Also, its token length is proportional to the square of image resolution, which might become the bottleneck in multimodal modeling given limited context length [49]. TiTok [66] proposes a 1D tokenizer which reduces the number of tokens to 32 for \\(256 \\times 256\\) images. However, TiTok is trained with only an overall reconstruction objective, and thus the precise meaning of these 1D tokens remains unclear. VAR [50] introduces a multi-scale tokenizer for causal autoregressive image generation. Nonetheless, the multi-scale strategy requires an even greater number of tokens compared to patch-based tokenization methods, further diminishing token efficiency. Our method converts images into 1D causal sequences of spectral tokens and enhances token efficiency by leveraging the long-tail distribution of spectral density in image data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Spectral visual analysis. Spectral analysis [2] has been a common technique in computer vision, complementing the spatial and temporal domains. Representative applications include image enhancement and denoising [15, 68], texture analysis and feature extraction [41], compression and super-resolution [13, 55], visual generation [61, 67], adversarial attacks and defenses [12, 32, 33]. For autoregressive image generation, CART [43] and SIT [11] propose to transform an image into multiple causal sets of tokens with base-detail decomposition and discrete wavelet transform, respectively. However, these methods still adhere to the multi-scale 2D tokenization paradigm similar to" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "VAR, resulting in suboptimal token efficiency and bidirectional attention to predict multiple tokens per autoregressive step. In contrast, SpectralAR leverages the discrete cosine transform to capture the global information of an image, and compress it into a 1D sequence with high efficiency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.18, + 0.707, + 0.198 + ], + "angle": 0, + "content": "3. Proposed Approach" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.205, + 0.89, + 0.221 + ], + "angle": 0, + "content": "3.1. Revisiting Images from the Spectral Domain" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.227, + 0.907, + 0.408 + ], + "angle": 0, + "content": "Discrete cosine transform. Spectral analysis investigates how complex signals can be represented with simpler basis functions [2], producing a spectral density distribution that represents the magnitude of corresponding basic components. This spectral density distribution is an equivalent representation of the original signal and provides a distinct perspective from the spatial domain depending on the properties of the basis functions. We employ the Discrete Cosine Transform (DCT) [1] to convert images into the spectral domain. The DCT result \\(\\mathbf{D}\\) shares the same shape with the transformed image \\(\\mathbf{I} \\in \\mathbb{R}^{H \\times W}\\) (for simplicity, we omit the channel dimension):" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.416, + 0.907, + 0.438 + ], + "angle": 0, + "content": "\\[\n\\mathbf {D} = \\left\\{F (u, v) \\right\\} _ {u, v = 1} ^ {W, H}, \\quad \\mathbf {I} = \\left\\{f (x, y) \\right\\} _ {x, y = 1} ^ {W, H}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.446, + 0.905, + 0.476 + ], + "angle": 0, + "content": "Each \\(F(u,v)\\) represents the intensity of the corresponding basis function \\(g_{u,v}(x,y)\\) in the image \\(\\mathbf{I}\\), which writes:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.483, + 0.907, + 0.514 + ], + "angle": 0, + "content": "\\[\ng _ {u, v} (x, y) = \\frac {2 C (u) C (v)}{\\sqrt {H W}} \\cos \\frac {(2 x + 1) u \\pi}{2 W} \\cos \\frac {(2 y + 1) v \\pi}{2 H}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.906, + 0.689 + ], + "angle": 0, + "content": "where \\( C(u) = 1 / \\sqrt{2} \\) if \\( u = 0 \\) and \\( C(u) = 1 \\) otherwise. This family of basis functions has the following properties: (1) Given \\( u \\) and \\( v \\), the basis function \\( g_{u,v}(x,y) \\) exhibits a checkerboard-like pattern in the spatial domain, with periods along the x- and y-axis of \\( 2W / u \\) and \\( 2H / v \\), respectively. This pattern suggests that the basis function characterizes the rate of variation of images in the spatial domain. (2) The \\( F(u,v) \\)s also form a 2D matrix together, where the top-left corner represents the low-frequency components (small \\( u, v \\)), while the bottom-right corner corresponds to the high-frequency components (large \\( u, v \\))." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.906, + 0.84 + ], + "angle": 0, + "content": "Causality from the spectral domain. Since low- and high-frequency components describe overall structures and intricate details, respectively, we can decompose an image into a sequence of sub-images \\(\\{\\mathbf{I}_i^{\\prime}\\}_{i = 1}^{L}\\) with increasing levels of detail by applying inverse-DCT on partially masked spectral density distributions \\(\\{\\mathbf{D}_i^{\\prime}\\}_{i = 1}^{L}\\). With more high-frequency components, the sub-images will gradually transition from blurred to sharp, as shown in Figure 2. This hierarchical coarse-to-fine sequence aligns with human visual perception and artistic drawing, thus enhancing causality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Efficiency from the spectral domain. As the image compression literature [1] pointed out, the DCT result \\(\\mathbf{D}\\) of images conforms to the power-law distribution with the absolute values of \\(F(u,v)\\mathrm{s}\\) on the top-left corner substantially" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15844" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.089, + 0.885, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.406, + 0.908, + 0.463 + ], + "angle": 0, + "content": "Figure 3. Overall pipeline of the proposed SpectralAR. Left: We convert an input image into a 1D causal sequence efficiently with nested spectral tokenization. Each nested sequence is supervised with distinct reconstruction target in a coarse-to-fine manner, which endows each token with an explicit spectral interpretation. We also apply the causal mask to the spectral tokens in the encoder and decoder to enhance the causality. Right: We use the groundtruth sequences from the tokenization process to train an autoregressive generation model." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.467, + 0.483, + 0.604 + ], + "angle": 0, + "content": "larger than the bottom-right corner, indicating that most energy of an image is concentrated in the low-frequency components. In addition, human visual perception is less sensitive to high-frequency signals, which have little influence on the visual quality of images. Based on these distinctions, we can encode the high-frequency components of an image with coarser granularity to improve token efficiency, similar to the JPEG algorithm [55] which saves more than \\(90\\%\\) storage for images by suppressing high-frequency signals." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.62, + 0.353, + 0.636 + ], + "angle": 0, + "content": "3.2. Nested Spectral Tokenization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Overall framework. In contrast to the 2D spatial tokenization that captures the local correlation between image patches, the basis function \\( g_{u,v}(x,y) \\) of DCT encodes a global frequency pattern. Therefore, we convert images into 1D tokens in spectral tokenization to reflect the global nature of the basis functions. We start with the general framework of a 1D image tokenizer [66]. Given an image \\( \\mathbf{I} \\), we aim to encode it into \\( N \\) discrete vectors \\( \\mathbf{S} \\), and also reconstruct the original image with \\( \\mathbf{S} \\). We first patchify the image into \\( \\mathbf{P} \\in \\mathbb{R}^{hw\\times C} \\), and concatenate the image features with the initial query vectors \\( \\mathbf{S}_0 \\) to form \\( [\\mathbf{P};\\mathbf{S}_0] \\in \\mathbb{R}^{(hw + N)\\times C} \\), where \\( h,w,C \\) denote the resolution of image features and the channel dimension, respectively. We then employ the vision transformer \\( \\mathcal{E} \\) to enable feature extraction and interaction between the image features and 1D query tokens, resulting in the informative 1D representation \\( \\hat{\\mathbf{S}} \\). In the vector quantizer \\( \\mathcal{Q} \\), we match these continuous vectors with the" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.467, + 0.907, + 0.556 + ], + "angle": 0, + "content": "codebook embeddings to derive the discrete representation \\(\\mathbf{S}\\), which could serve as the groundtruth for the autoregressive training. At last, we append \\(\\mathbf{S}\\) to a set of mask tokens \\(\\mathbf{M} \\in \\mathbb{R}^{hw \\times C}\\) and process them with the decoder network \\(\\mathcal{D}\\) similar to \\(\\mathcal{E}\\), in order to reconstruct the original image. This overall framework could be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.569, + 0.907, + 0.588 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {S}} = \\mathcal {E} ([ \\mathbf {P}; \\mathbf {S} _ {0} ]), \\quad \\mathbf {S} = \\mathcal {Q} (\\hat {\\mathbf {S}}), \\quad \\hat {\\mathbf {I}} = \\mathcal {D} ([ \\mathbf {M}; \\mathbf {S} ]), \\quad (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.601, + 0.907, + 0.632 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{I}}\\) denotes the reconstructed image. The training objective typically consists of multiple loss functions:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.643, + 0.907, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t o k} = \\| \\hat {\\mathbf {I}} - \\mathbf {I} \\| _ {2} ^ {2} + \\| \\hat {\\mathbf {S}} - \\mathbf {S} \\| _ {2} ^ {2} + \\mathcal {L} _ {P} (\\hat {\\mathbf {I}}, \\mathbf {I}) + \\mathcal {L} _ {A} (\\hat {\\mathbf {I}}, \\mathbf {I}), (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.906, + 0.704 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_P\\) and \\(\\mathcal{L}_A\\) denote the perceptual loss [22, 70] and the adversarial loss [10, 14], respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Nested spectral decoding. Different from the plain 1D tokenization in Figure 2, we aim to represent an image as its spectral decompositions \\(\\{\\mathbf{I}_i^{\\prime}\\}_{i = 1}^{L}\\) for sequence causality, which requires establishing a mapping from 1D tokens to these sub-images. One naive way to achieve this would be dividing S into disjoint subsets and assigning them to model different sub-images independently. Similar to the multi-scale tokenization [50], this strategy would inevitably involve bidirectional interaction and diminish token efficiency because sub-images with finer detail would require increasingly more tokens to represent. In contrast, we propose a nested mapping scheme for efficient tokenization, as shown in the bottom of Figure 3. We first construct a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15845" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.101, + 0.096, + 0.43, + 0.111 + ], + "angle": 0, + "content": "Algorithm 1: Nested Spectral Tokenization Training" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.113, + 0.458, + 0.128 + ], + "angle": 0, + "content": "1 Inputs: raw image \\(\\mathbf{I}\\), initial spectral tokens \\(\\mathbf{S}_0\\in \\mathbb{R}^{N\\times C}\\)" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.128, + 0.437, + 0.143 + ], + "angle": 0, + "content": "2 Hyperparameters: spectral levels \\(\\{\\omega_i|i = 1,\\dots,N\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.144, + 0.39, + 0.159 + ], + "angle": 0, + "content": "3 \\(\\mathbf{P} =\\) patchify(I), \\(\\hat{\\mathbf{S}} = \\mathcal{E}([\\mathbf{P};\\mathbf{S}_0])\\) \\(\\mathbf{S} = \\mathcal{Q}(\\hat{\\mathbf{S}})\\)" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.159, + 0.275, + 0.172 + ], + "angle": 0, + "content": "4 \\(idx = \\mathrm{random\\_choice}(N)\\)" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.173, + 0.309, + 0.188 + ], + "angle": 0, + "content": "5 \\(\\mathbf{S}^{\\prime} = \\mathbf{S}[:idx],\\hat{\\mathbf{I}}^{\\prime} = \\mathcal{D}([\\mathbf{M};\\mathbf{S}^{\\prime}])\\)" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.188, + 0.408, + 0.202 + ], + "angle": 0, + "content": "\\(\\mathbf{6}\\) \\(\\mathbf{D} = \\mathrm{DCT}(\\mathbf{I}),\\mathbf{D}' = \\mathbf{D}\\circ \\mathbf{1}_{\\omega_i},\\mathbf{I}' = \\mathrm{DCT}^{-1}(\\mathbf{D}')\\)" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.203, + 0.235, + 0.218 + ], + "angle": 0, + "content": "7 loss \\(= \\mathcal{L}_{tok}(\\hat{\\mathbf{l}}',\\mathbf{l}')\\)" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.219, + 0.303, + 0.232 + ], + "angle": 0, + "content": "8 Return: loss for optimization;" + }, + { + "type": "list", + "bbox": [ + 0.098, + 0.113, + 0.458, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.244, + 0.483, + 0.274 + ], + "angle": 0, + "content": "sequence of sub-images with increasing detail by progressively preserving larger regions in the spectral density \\(\\mathbf{D}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.283, + 0.482, + 0.301 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} _ {i} ^ {\\prime} = \\mathrm {D C T} ^ {- 1} \\left(\\mathbf {D} _ {i} ^ {\\prime}\\right), \\quad \\mathbf {D} _ {i} ^ {\\prime} = \\mathbf {D} \\circ \\mathbf {1} _ {\\omega_ {i}}, \\quad \\omega_ {i - 1} < \\omega_ {i}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.31, + 0.484, + 0.463 + ], + "angle": 0, + "content": "where \\(\\mathrm{DCT}^{-1}\\), \\(\\circ\\), \\(\\mathbf{1}_{\\omega_i}\\) denote the inverse DCT operation, element-wise multiplication and a \\(H\\times W\\) matrix with the top-left corner of size \\(\\omega_{i}\\times \\omega_{i}\\) filled by ones and the remaining parts being zeros. Therefore, the sub-image \\(\\mathbf{I}_i^{\\prime}\\) contains all frequency components present in the sub-image \\(\\mathbf{I}_{i - 1}^{\\prime}\\). Based on this inclusion property, we can reuse the tokens representing the previous sub-image \\(\\mathbf{I}_{i - 1}^{\\prime}\\) to represent the next sub-image \\(\\mathbf{I}_i^\\prime\\). To avoid bidirectional attention, we make \\(L\\) equal to \\(N\\) so that each token \\(\\mathbf{s}_i\\) in the sequence S corresponds to a unique sub-image \\(\\mathbf{I}_i^\\prime\\):" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.473, + 0.482, + 0.49 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {I}} _ {i} ^ {\\prime} = \\mathcal {D} ([ \\mathbf {M}; \\mathbf {s} _ {1}, \\mathbf {s} _ {2}, \\dots , \\mathbf {s} _ {i} ]), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.501, + 0.483, + 0.576 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{I}}_i^{\\prime}\\) is the reconstruction of sub-image \\(\\mathbf{I}_i^\\prime\\) given the nested 1D sequence \\(\\{\\mathbf{s}_1,\\mathbf{s}_2,\\dots ,\\mathbf{s}_i\\}\\). Nested spectral decoding compresses an image into a causal 1D sequence where each token \\(\\mathbf{s}_i\\) corresponds to a disjoint set of frequencies and achieves token efficiency by reusing previous tokens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.578, + 0.484, + 0.698 + ], + "angle": 0, + "content": "Non-uniform token-frequency mapping. To further enhance token efficiency, we introduce the non-uniform token-frequency mapping technique. Since high-frequency components have low magnitude and minimal impact on the visual quality of images, we can encode them with coarser granularity compared to the low-frequency counterparts. We achieve this by reducing the interval between \\(\\omega_{i-1}\\) and \\(\\omega_i\\) when \\(i\\) is small, and increase it otherwise:" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.71, + 0.482, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\omega_ {i} - \\omega_ {i - 1} \\leq \\omega_ {i + 1} - \\omega_ {i}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.796 + ], + "angle": 0, + "content": "which demonstrates the general case. This non-uniform mapping allocates later tokens to broader frequency ranges, enabling precise modeling of crucial low-frequency components while efficiently representing high-frequency details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.9 + ], + "angle": 0, + "content": "Spectral causal mask. Although the sequence \\(\\mathbf{S}\\) supervised with (5)(6) already exhibits a certain degree of causality, the encoding and decoding processes, i.e. \\(\\mathcal{E}\\) and \\(\\mathcal{D}\\), are still bidirectional, which can lead to information leakage from high-frequency to low-frequency components. Therefore, we propose applying causal masks to the spectral tokens \\(\\mathbf{S}\\) in both the encoder \\(\\mathcal{E}\\) and the decoder \\(\\mathcal{D}\\), as shown" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.145 + ], + "angle": 0, + "content": "Table 1. Correlation between tokens of different autoregressive paradigms. We use linear correlation as a proxy metric for the causality of sequences. The spectral sequence demonstrates better causality compared with other methods." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.146, + 0.905, + 0.236 + ], + "angle": 0, + "content": "
Correlation TypeRaster-scanScale-wiseSpectral
R2avg(t2; t1)0.4710.8890.916
R2avg(t3; t1, t2)0.3660.9530.977
R2avg(t4; t1, t2, t3)0.5250.9430.994
Average0.4540.9280.962
" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.24, + 0.639, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.334, + 0.606, + 0.342 + ], + "angle": 0, + "content": "Raster-scan" + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.24, + 0.759, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.673, + 0.334, + 0.725, + 0.342 + ], + "angle": 0, + "content": "Scale-wise" + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.24, + 0.882, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.8, + 0.334, + 0.842, + 0.343 + ], + "angle": 0, + "content": "Spectral" + }, + { + "type": "image", + "bbox": [ + 0.882, + 0.24, + 0.905, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.346, + 0.907, + 0.401 + ], + "angle": 0, + "content": "Figure 4. Frequency of each element having the highest correlation with another one. It represents the source of reference information for predicting the next token. The raster-scan method exhibits excessive dependency on boundaries of image patches." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.406, + 0.907, + 0.467 + ], + "angle": 0, + "content": "in Figure 3. This spectral causal mask restricts each token s to only attend to tokens that represent lower frequencies, enhancing causality from the architectural perspective. We outline the training procedure in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.476, + 0.823, + 0.492 + ], + "angle": 0, + "content": "3.3. Spectral Autoregressive Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.498, + 0.906, + 0.62 + ], + "angle": 0, + "content": "Autoregressive modeling has gained prominence in computer vision for its scalability, generalization, and effectiveness across multimodal tasks [25, 49, 56]. While conventional autoregressive generation follows a spatial raster-scan order [10, 51], we propose a hierarchical coarse-to-fine approach in the spectral domain to enhance causality, as shown in the right side of Figure 3. We start with the general framework of autoregressive modeling:" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.632, + 0.905, + 0.648 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {i + 1} = \\mathcal {M} \\left(\\mathbf {t} _ {1}, \\mathbf {t} _ {2}, \\dots , \\mathbf {t} _ {i}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{T} = \\{\\mathbf{t}_i\\}_{i=1}^N\\) is a sequence of quantized tokens, and \\(\\mathbf{p}_{i+1}\\), \\(\\mathcal{M}\\) denote the probability logits for the \\((i+1)\\)th token and the autoregressive model, respectively. The autoregressive process (8) assumes that the generation of token \\(\\mathbf{t}_{i+1}\\) depends solely on the previous ones. The spatial autoregressive paradigm violates this premise because of the equality of image pixels (as discussed in Section 1). In contrast, we take the spectral tokens \\(\\mathbf{S}\\) from the nested spectral tokenziation as the autoregressive targets \\(\\mathbf{T}\\). Since the spectral tokens are trained in a nested manner to reconstruct sub-images of increasing levels of detail as in (5)(6), each spectral token \\(\\mathbf{s}_i\\) is expected to enhance the quality of the sub-image \\(\\mathbf{I}_{i-1}'\\) represented by previous tokens from the spectral domain. This progressive refinement process aligns with human visual perception and artistic painting, both of which start with the overall structure and gradually focus on" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15846" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.146 + ], + "angle": 0, + "content": "Table 2. Comparison between generative models on class-conditional ImageNet \\( {256} \\times {256} \\) benchmark. “ \\( \\downarrow \\) ” or “↑” indicate lower or higher values are better. \"Token\": the number of tokens used in transformer architectures. \"Step\": the number of model runs needed to generate an image. We compute our wall-clock inference time and scale the Time for other methods accordingly. \\( \\dagger \\) : trained on larger datasets including OpenImages [27]. ‡: implemented with the official tokenizer weight [66] and the scripts from VAR." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.147, + 0.885, + 0.596 + ], + "angle": 0, + "content": "
TypeModelrFID↓gFID↓IS↑Pre↑Rec↑#Token#Para#StepTime
GANBigGAN [3]75.246.95224.50.890.38112M1-
GANGigaGAN [23]-3.45225.50.840.61N.A.569M1-
GANStyleGan-XL [45]7.062.30265.10.780.53166M10.4
Diff.ADM [8]125.7810.94101.00.690.63N.A.554M250235
Diff.CDM [20]-4.88158.7--N.A.-8100-
Diff.LDM-4-G [44]0.27†3.60247.7--N.A.400M250-
Diff.DiT-L/2 [39]0.62†5.02167.20.750.57256458M25043
Diff.DiT-XL/2 [39]0.62†2.27278.20.830.57256675M25063
Diff.L-DiT-3B [71]-2.10304.40.820.602563.0B250>63
Diff.L-DiT-7B [71]-2.28316.20.830.582567.0B250>63
Mask.MaskGIT [4]2.286.18182.10.800.51256227M80.7
Mask.RCG (cond.) [30]-3.49215.5---502M202.7
Mask.TiTok-B64 [66]1.702.48214.7--64177M80.4
2D ScanVQVAE-2 [42]-31.11~450.360.57N.A.13.5B5120-
2D ScanVQGAN [10]7.9415.7874.3--N.A.1.4B25634
2D ScanViTVQ [63]1.284.17175.1--10241.7B1024>34
2D ScanRQTran. [29]3.207.55134.0--64,43.8B6829
VARVAR-d16 [50]3.30274.40.840.51310M100.6
VARVAR-d20 [50]2.57302.60.830.56600M100.7
VARVAR-d24 [50]2.09312.90.820.591.0B100.8
VARVAR-d30 [50]1.92323.10.820.592.0B101.4
1D ARTiTok-B64-d16‡ [66]1.706.30190.10.850.4764310M641
1D ARSpectralAR-d164.033.02282.20.810.5564310M641
1D ARSpectralAR-d204.032.49305.4--64600M641.2
1D ARSpectralAR-d244.032.13307.7--641.0B641.4
1D ARSpectralAR-d16-p44.033.13276.1--16×4310M160.4
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.603, + 0.483, + 0.663 + ], + "angle": 0, + "content": "details. This similarity qualitatively validates the rationale for performing causal autoregressive generation in the spectral domain. We further provide some quantitative analysis through a toy experiment in Section 4.2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.664, + 0.483, + 0.815 + ], + "angle": 0, + "content": "Potential applications. The frequencies represented by the token \\(\\mathbf{s}_i\\) become higher as \\(i\\) increases, while its influence on the image quality diminishes accordingly (check Section 3.1 for details). Therefore, we can control the visual quality of sampled images by adjusting the length of generated sequences, similar to the image compression algorithms [55]. In addition, we can achieve super-resolution by dividing images into disjoint parts smaller than \\(H\\times W\\), and conducting individual spectral autoregressive generation on each part. We provide further results in Section 4.4." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.824, + 0.224, + 0.841 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.405, + 0.866 + ], + "angle": 0, + "content": "4.1. Dataset and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We train and evaluate our SpectralAR on the ImageNet1K [7] benchmark, which contains 1,281,167 and 50,000" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.603, + 0.907, + 0.693 + ], + "angle": 0, + "content": "images for training and validation, respectively. We train the tokenizer and generator on the training split. We evaluate the reconstruction performance on the validation set with reconstruction Fréchet inception distance [17] (rFID), and the generation results with generation FID (gFID) using pre-computed statistics and scripts from ADM [8]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.694, + 0.907, + 0.786 + ], + "angle": 0, + "content": "For tokenizer training, we follow the exactly same settings of TiTok [66] for a fair comparison. We also employ the two-stage training strategy with proxy codes [4, 66]. We use the ViT-B [9] as the encoder and decoder, and set the number of spectral tokens as \\( N = 64 \\) in our main experiments and the sequence \\( \\omega \\)s for \\( 256 \\times 256 \\) images as:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.797, + 0.907, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\omega_ {i} = \\left\\{ \\begin{array}{l l} i, & \\text {i f} i \\in (0, 3 2 ], \\\\ 2 i - 3 2, & \\text {i f} i \\in (3 2, 4 8 ], \\\\ 1 2 i - 5 1 2, & \\text {i f} i \\in (4 8, 6 4 ]. \\end{array} \\right. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "For generator training, we adopt the same architecture and training recipe as VAR [50], which leverages a GPT-2-like transformer architecture [40] for autoregressive modeling." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15847" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.088, + 0.905, + 0.639 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.642, + 0.908, + 0.673 + ], + "angle": 0, + "content": "Figure 5. Visualizations of generated \\(256 \\times 256\\) samples on ImageNet-1K. These samples cover a wide range of categories and styles, demonstrating the ability of SpectralAR to generate both diverse and high-quality images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.678, + 0.262, + 0.695 + ], + "angle": 0, + "content": "4.2. Proof of Concept" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.486, + 0.901 + ], + "angle": 0, + "content": "We conduct a proof-of-concept experiment to validate the causality of images in the spectral domain. We use the validation set of the CIFAR-100 dataset [26] as a proxy for the distribution of natural images, which contains 10,000 \\(32 \\times 32\\) images from 100 semantic classes. We aim to measure the linear correlation \\((R^2)\\) [38] between one token \\(\\mathbf{t}_i\\) and its previous ones \\(\\{\\mathbf{t}_j\\}_{j=1}^{i-1}\\) as a metric for the causality of the sequence \\(\\mathbf{T}\\). We consider three typical autoregressive paradigms, including raster-scan [10], scale-wise [50] and spectral methods which construct sequence \\(\\mathbf{T}\\) in different ways. Specifically, we divide each image spatially into four patches of size \\(16 \\times 16\\) and rearrange them in raster-scan order for the first variant. For the scale-wise method, we" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.679, + 0.909, + 0.804 + ], + "angle": 0, + "content": "downsample each image into sizes of \\(8 \\times 8\\), \\(16 \\times 16\\), \\(24 \\times 24\\) and \\(32 \\times 32\\), respectively. We set the frequency thresholds to 8, 16, 24, 32 and generate the sub-images following (5) for the spectral sequence. Since each token \\(\\mathbf{t}\\) is a high dimensional tensor, we calculate \\(R^2\\) between every element of \\(\\mathbf{t}_i\\) and every element of \\(\\{\\mathbf{t}_j\\}_{j=1}^{i-1}\\). For each element of \\(\\mathbf{t}_i\\), we then find the element in \\(\\{\\mathbf{t}_j\\}_{j=1}^{i-1}\\) with the highest correlation, and take an average as the final measurement:" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.812, + 0.907, + 0.835 + ], + "angle": 0, + "content": "\\[\nR _ {a v g} ^ {2} = \\operatorname {A V G} _ {m} \\left(\\max _ {n} R _ {m, n} ^ {2}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.838, + 0.909, + 0.903 + ], + "angle": 0, + "content": "where \\(m, n\\) denote element number of \\(\\mathbf{t}_i\\) and \\(\\{\\mathbf{t}_j\\}_{j=1}^{i-1}\\), respectively, and \\(R_{m,n}^2\\) is the correlation between the \\(m\\)th and \\(n\\)th elements. We report the quantitative results in Table 1, where the scale-wise and spectral sequences demonstrate" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15848" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.09, + 0.482, + 0.118 + ], + "angle": 0, + "content": "Table 3. Applications of SepctralAR. Super-reso., Trunc. represent super-resolution and truncated, respectively." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.12, + 0.48, + 0.26 + ], + "angle": 0, + "content": "
App. TypeModelFID↓IS↑
Super-reso.Upsample3.09286.6
SpectralAR-Stride2.93276.4
SpectralAR-Patch14.76170.0
Trunc.SpectralAR-Trunc.53.34271.5
SpectralAR-Trunc.106.65211.4
SpectralAR-Trunc.1527.6991.69
SpectralAR-Trunc.03.02282.2
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.266, + 0.483, + 0.342 + ], + "angle": 0, + "content": "considerably higher correlation than the raster-scan counterpart. This could be attributed to the coarse-to-fine nature of the former two paradigms, while the raster scan method lacks adequate reference information to predict the next image patch, as shown in Figure 4." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.35, + 0.231, + 0.365 + ], + "angle": 0, + "content": "4.3. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.372, + 0.483, + 0.751 + ], + "angle": 0, + "content": "We report the performance of SpectralAR on the class-conditional ImageNet-1K [7] \\(256 \\times 256\\) generation benchmark in Table 2. We also implement an autoregressive version of TiTok [66] by using the official tokenizer weight and scripts from VAR [50] for a fair comparison. The reconstruction performance of SpectralAR (4.03 rFID) is inferior compared with TiTok because SpectralAR requires to reconstruct the sub-images corresponding to different frequencies with different lengths of tokens, which is much more difficult compared with the overall reconstruction target in TiTok. Despite the lower reconstruction score, SpectralAR-d16 still outperforms TiTok-B64-d16 and VAR-d16 in autoregressive generation with a clear margin due to better sequence causality which eases autoregressive learning. Furthermore, SpectralAR also achieves better or comparable performance against VAR under the \\(d20\\) and \\(d24\\) settings, respectively. In addition, SpectralAR uses only 64 tokens in both reconstruction and generation, demonstrating superior token efficiency compared to VAR [50] and 2D scan-based methods. We also visualize the generated samples in Figure 5, which shows the diversity and quality of the generation process of SpectralAR. In addition, Figure 1 highlights the hierarchical coarse-to-fine refinement of the images, while SpectralAR generates more tokens in an autoregressive way." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.225, + 0.774 + ], + "angle": 0, + "content": "4.4. Applications" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this section, we provide a quantitative analysis for the potential applications of SpectralAR. For super-resolution, we conduct our experiments based on the \\(256 \\times 256\\) images generated by SpectralAR-\\(d16\\) in Table 2, and directly upsample them to \\(512 \\times 512\\) as the baseline. We construct 4 sub-images using stripped and patch-based methods for SpectralAR-Stride and -Patch, respectively. We then upsample the 4 sub-images to \\(256 \\times 256\\) and use SpectralAR" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.905, + 0.146 + ], + "angle": 0, + "content": "Table 4. Ablation on design choices. Spectral Supervision means using sub-image supervision across different frequency bands. Causal Mask and non-uniform mapping refers to the spectral causal mask and the frequency-token mapping, respectively." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.148, + 0.905, + 0.251 + ], + "angle": 0, + "content": "
Spectral SupervisionCausal MaskNon-uniform MappingFID↓IS↑
×××6.30190.1
✓××5.64255.1
✓✓×3.49222.6
✓✓✓3.02282.2
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.259, + 0.907, + 0.426 + ], + "angle": 0, + "content": "to refine them in the spectral domain, and finally reassemble them to generate the final result. According to Table 3, SpectralAR can indeed serve as a spectrum completer for the super-resolution task. In addition, we also experiment with truncated autoregressive generation, where we discard the last few tokens. For example, we discard the last 5 tokens in SpectralAR-Trunc.5 in Table 3. The generation performance worsens slowly when the number of truncated tokens is fewer than 10, and therefore it is possible to further improve token efficiency through truncation according to the requirement for generation quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.433, + 0.667, + 0.45 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.909, + 0.683 + ], + "angle": 0, + "content": "We conduct ablation study to validate the effectiveness of our design choices in Table 4. The first row corresponds to the baseline TiTok implementation for autoregressive generation. Note that the spectral supervision alone could improve FID compared with the TiTok counterpart. This is because the spectral design enhances sequence causality compared with the overall reconstruction target which ignores the correlation between tokens and thus complicates the autoregressive modeling process. The spectral causal mask further enhances performance by improving causality in the encoding and decoding architecture. And the non-uniform token-frequency mapping technique guides the model to focus more on the crucial low-frequency components while representing the high-frequency components efficiently, thus further improving performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.634, + 0.71 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this paper, we have proposed the spectral autoregressive visual generation method for both causal and efficient autoregressive modeling of image data. Specifically, we first convert images into 1D sequences with nested spectral tokenization. In addition, we have adopted causal masks for spectral tokens in the encoder and decoder to further enhance causality from the architectural perspective. We have also designed a non-uniform token-frequency mapping with emphasis on the low-frequency components in order to improve token efficiency. On the ImageNet-1K generation benchmark, our SpectralAR achieves superior performance compared with other autoregressive generation methods." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15849" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.186 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 62125603, Grant 62321005, Grant 62336004, and Grant 62441616, and in part by the Beijing Natural Science Foundation under Grant No. L247009." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.214, + 0.188, + 0.23 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.24, + 0.483, + 0.281 + ], + "angle": 0, + "content": "[1] Nasir Ahmed, T. Natarajan, and Kamisetty R Rao. Discrete cosine transform. IEEE transactions on Computers, 100(1): 90-93, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.284, + 0.483, + 0.311 + ], + "angle": 0, + "content": "[2] Jean Baptiste Joseph Baron Fourier et al. The analytical theory of heat. Courier Corporation, 2003. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.314, + 0.482, + 0.354 + ], + "angle": 0, + "content": "[3] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.357, + 0.483, + 0.397 + ], + "angle": 0, + "content": "[4] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In CVPR, pages 11315-11325, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.401, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[5] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In ICML, pages 1691-1703. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.458, + 0.482, + 0.498 + ], + "angle": 0, + "content": "[6] Aaron Clauset, Cosma Rohilla Shalizi, and Mark EJ Newman. Power-law distributions in empirical data. SIAM review, 51(4):661-703, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.501, + 0.482, + 0.542 + ], + "angle": 0, + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.545, + 0.483, + 0.572 + ], + "angle": 0, + "content": "[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NIPS, 34:8780-8794, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.574, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.659, + 0.482, + 0.699 + ], + "angle": 0, + "content": "[10] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, pages 12873-12883, 2021. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.702, + 0.482, + 0.741 + ], + "angle": 0, + "content": "[11] Carlos Esteves, Mohammed Suhail, and Ameesh Makadia. Spectral image tokenizer. arXiv preprint arXiv:2412.09607, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[12] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In ICML, pages 3247-3258. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[13] Manuel Fritsche, Shuhang Gu, and Radu Timofte. Frequency separation for real-world super-resolution. In ICCVW, pages 3599-3608. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. NIPS, 27, 2014. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.24, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[15] Hayit Greenspan, Charles H Anderson, and Sofia Akber. Image enhancement by nonlinear extrapolation in frequency space. TIP, 9(6):1035-1048, 2000. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[16] Jian Han, Jinlai Liu, Yi Jiang, Bin Yan, Yuqi Zhang, Zehuan Yuan, Bingyue Peng, and Xiaobing Liu. Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis. arXiv preprint arXiv:2412.04431, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.905, + 0.246 + ], + "angle": 0, + "content": "[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NIPS, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[18] Geoffrey E Hinton and Ruslan R Salakhutdinov. Reducing the dimensionality of data with neural networks. science, 313(5786):504-507, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NIPS, 33:6840-6851, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.322, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[20] Jonathan Ho, Chitwan Saharia, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. JMLR, 23(47):1-33, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.379, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[21] Yuanhui Huang, Wenzhao Zheng, Yuan Gao, Xin Tao, Pengfei Wan, Di Zhang, Jie Zhou, and Jiwen Lu. Owl1: Omni world model for consistent long video generation. arXiv preprint arXiv:2412.09600, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[22] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, pages 694–711. Springer, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.478, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[23] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In CVPR, pages 10124-10134, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.534, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[24] Diederik P Kingma, Max Welling, et al. Auto-encoding variational bayes, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.564, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[25] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Grant Schindler, Rachel Hornung, Vighnesh Birodkar, Jimmy Yan, Ming-Chang Chiu, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.634, + 0.905, + 0.66 + ], + "angle": 0, + "content": "[26] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.663, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[27] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 128(7):1956-1981, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.747, + 0.905, + 0.799 + ], + "angle": 0, + "content": "[28] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[29] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In CVPR, pages 11523-11532, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[30] Tianhong Li, Dina Katabi, and Kaiming He. Self-conditioned image generation via generating representations. CoRR, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15850" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[31] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. NIPS, 37:56424-56445, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[32] Yuyang Long, Qilong Zhang, Boheng Zeng, Lianli Gao, Xianglong Liu, Jian Zhang, and Jingkuan Song. Frequency domain model augmentation for adversarial attack. In ECCV, pages 549-566. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[33] Cheng Luo, Qinliang Lin, Weicheng Xie, Bizhu Wu, Jinheng Xie, and Linlin Shen. Frequency-driven imperceptible adversarial attack on semantic similarity. In CVPR, pages 15315-15324, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[34] Zhuoyan Luo, Fengyuan Shi, Yixiao Ge, Yujiu Yang, Limin Wang, and Ying Shan. Open-magvit2: An open-source project toward democratizing auto-regressive visual generation. arXiv preprint arXiv:2409.04410, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.373 + ], + "angle": 0, + "content": "[35] Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. In ECCV, pages 23-40. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[36] Xiaoxiao Ma, Mohan Zhou, Tao Liang, Yalong Bai, Tiejun Zhao, Huaian Chen, and Yi Jin. Star: Scale-wise text-to-image generation via auto-regressive representations. arXiv preprint arXiv:2406.10797, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[37] Fabian Mentzer, David Minnen, Eirikur Agustsson, and Michael Tschannen. Finite scalar quantization: Vq-vae made simple. arXiv preprint arXiv:2309.15505, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.475, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[38] Daniel J Ozer. Correlation and the coefficient of determination. Psychological bulletin, 97(2):307, 1985. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.482, + 0.544 + ], + "angle": 0, + "content": "[39] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[40] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.643 + ], + "angle": 0, + "content": "[41] Yongming Rao, Wenliang Zhao, Zheng Zhu, Jiwen Lu, and Jie Zhou. Global filter networks for image classification. Advances in neural information processing systems, 34:980-993, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.482, + 0.686 + ], + "angle": 0, + "content": "[42] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. NIPS, 32, 2019. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.688, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[43] Siddharth Roheda. Cart: Compositional auto-regressive transformer for image generation. arXiv preprint arXiv:2411.10180, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.731, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[45] Axel Sauer, Katja Schwarz, and Andreas Geiger. Styleganx1: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH 2022 conference proceedings, pages 1-10, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[46] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[47] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[48] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[49] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[50] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. NIPS, 37:84839-84865, 2025. 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[51] Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelCNN decoders. NIPS, 29, 2016. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[52] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NIPS, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.35, + 0.905, + 0.391 + ], + "angle": 0, + "content": "[53] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NIPS, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.393, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[54] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In ICML, pages 1096-1103, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.905, + 0.49 + ], + "angle": 0, + "content": "[55] Gregory K Wallace. The jpeg still picture compression standard. Communications of the ACM, 34(4):30-44, 1991. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.492, + 0.905, + 0.547 + ], + "angle": 0, + "content": "[56] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.549, + 0.905, + 0.604 + ], + "angle": 0, + "content": "[57] Yuqing Wang, Tianwei Xiong, Daquan Zhou, Zhijie Lin, Yang Zhao, Bingyi Kang, Jiashi Feng, and Xihui Liu. Loong: Generating minute-level long videos with autoregressive language models. arXiv preprint arXiv:2410.02757, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.606, + 0.905, + 0.646 + ], + "angle": 0, + "content": "[58] Thomas Wiegand, Gary J Sullivan, Gisle Bjontegaard, and Ajay Luthra. Overview of the h. 264/avc video coding standard. TCSVT, 13(7):560-576, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.648, + 0.905, + 0.717 + ], + "angle": 0, + "content": "[59] Chengyue Wu, Xiaokang Chen, Zhiyu Wu, Yiyang Ma, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, Chong Ruan, et al. Janus: Decoupling visual encoding for unified multimodal understanding and generation. arXiv preprint arXiv:2410.13848, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.719, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[60] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.842 + ], + "angle": 0, + "content": "[61] Mengping Yang, Zhe Wang, Ziqiu Chi, and Wenyi Feng. Wavegan: Frequency-aware gan for high-fidelity few-shot image generation. In ECCV, pages 1-17. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[62] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15851" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[63] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[64] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[65] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Randomized autoregressive visual generation. arXiv preprint arXiv:2411.00776, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.33 + ], + "angle": 0, + "content": "[66] Qihang Yu, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An image is worth 32 tokens for reconstruction and generation. NIPS, 37: 128940-128966, 2025. 2, 3, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.388 + ], + "angle": 0, + "content": "[67] Yingchen Yu, Fangneng Zhan, Shijian Lu, Jianxiong Pan, Feiying Ma, Xuansong Xie, and Chunyan Miao. Wavefill: A wavelet-based generation network for image inpainting. In ICCV, pages 14114-14123, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.482, + 0.442 + ], + "angle": 0, + "content": "[68] Huanjing Yue, Xiaoyan Sun, Jingyu Yang, and Feng Wu. Cid: Combined image denoising in spatial and frequency domains using web images. In CVPR, pages 2933-2940, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.446, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[69] Qian Zhang, Xiangzi Dai, Ninghua Yang, Xiang An, Ziyong Feng, and Xingyu Ren. Var-clip: Text-to-image generator with visual auto-regressive modeling. arXiv preprint arXiv:2408.01181, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.503, + 0.482, + 0.557 + ], + "angle": 0, + "content": "[70] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.56, + 0.482, + 0.627 + ], + "angle": 0, + "content": "[71] Renrui Zhang, Jiaming Han, Chris Liu, Peng Gao, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, and Yu Qiao. Llama-adapter: Efficient finetuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.63, + 0.482, + 0.671 + ], + "angle": 0, + "content": "[72] Chuanxia Zheng, Tung-Long Vuong, Jianfei Cai, and Dinh Phung. Movq: Modulating quantized vectors for high-fidelity image generation. NIPS, 35:23412-23425, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.673, + 0.482, + 0.728 + ], + "angle": 0, + "content": "[73] Wenzhao Zheng, Weiliang Chen, Yuanhui Huang, Borui Zhang, Yueqi Duan, and Jiwen Lu. Occworld: Learning a 3d occupancy world model for autonomous driving. In ECCV, pages 55-72. Springer, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.728 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "15852" + } + ] +] \ No newline at end of file diff --git a/2025/SpectralAR_ Spectral Autoregressive Visual Generation/ddfacecc-ae50-4e90-8053-8ccf2e5c55b7_origin.pdf b/2025/SpectralAR_ Spectral Autoregressive Visual Generation/ddfacecc-ae50-4e90-8053-8ccf2e5c55b7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c47e86fcebff7e89fdb0ca9e6046e71cbda78dad --- /dev/null +++ b/2025/SpectralAR_ Spectral Autoregressive Visual Generation/ddfacecc-ae50-4e90-8053-8ccf2e5c55b7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:50cf0cf8aea20027f108ab8a0cd0799c4d74865e77d5c20c61cc4f2376647a77 +size 2965579 diff --git a/2025/SpectralAR_ Spectral Autoregressive Visual Generation/full.md b/2025/SpectralAR_ Spectral Autoregressive Visual Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ee520a8c6fff24787e17bdcf327fe2d43201c51b --- /dev/null +++ b/2025/SpectralAR_ Spectral Autoregressive Visual Generation/full.md @@ -0,0 +1,304 @@ +# SpectralAR: Spectral Autoregressive Visual Generation + +Yuanhui Huang Weiliang Chen Wenzhao Zheng Yueqi Duan Jie Zhou Jiwen Lu Tsinghua University + +huangyh22@mails.tsinghua.edu.cn; wenzhao.zheng@outlook.com + +![](images/62697571be1fe7f68503368ee738c8ebad782066ae33a320043d4f33785b2b06.jpg) +Figure 1. We approach autoregressive visual generation from the spectral perspective and propose SpectralAR which converts images into 1D causal sequences with nested spectral tokenization and generates images in a hierarchical coarse-to-fine manner. In the autoregressive process, each generated token improves the quality of the image by introducing new high-frequency components. + +# Abstract + +Autoregressive visual generation has garnered increasing attention due to its scalability and compatibility with other modalities compared with diffusion models. Most existing methods construct visual sequences as spatial patches for autoregressive generation. However, image patches are inherently parallel, contradicting the causal nature of autoregressive modeling. To address this, we propose a Spectral AutoRegressive (SpectralAR) visual generation framework, which realizes causality for visual sequences from the spectral perspective. Specifically, we first transform an image into ordered spectral tokens with Nested Spectral Tokeniza + +tion, representing lower to higher frequency components. We then perform autoregressive generation in a coarse-to-fine manner with the sequences of spectral tokens. By considering different levels of detail in images, our SpectralAR achieves both sequence causality and token efficiency without bells and whistles. We conduct extensive experiments on ImageNet-1K for image reconstruction and autoregressive generation, and SpectralAR achieves 3.02 gFID with only 64 tokens and 310M parameters. Project page: https://huang-yh.github.io/spectralar/. + +# 1. Introduction + +Diffusion models [19, 39, 44, 46] have long been the best performing approach to visual generation. Despite their ex + +ceptional generation quality, diffusion models still exhibit deficiencies in multimodal modeling and integration of perception and generation. The advent of autoregressive visual generation methods [10, 29, 36, 42, 48, 50, 65] alleviates these limitations and enables better scalability with the next-token prediction paradigm. It first utilizes a visual tokenizer to convert images into tokens and then generates samples in a sequential manner. This advancement supports a variety of emerging applications, including scalable visual generation [16, 57, 59, 69], mixed-modal foundation models [49, 56, 60], and autoregressive world models [21, 73]. + +Despite their dominance in language modeling [40, 53], the performance of autoregressive models in visual generation [10, 14] is still inferior to that of diffusion models [35, 39] and non-autoregressive models [4, 66]. This distinction can be attributed to the inherent difference between text and image modalities. Text, invented for human communication, is discrete and sequential, while image data is continuous and invariant to translation [28], indicating the intrinsic equality among image pixels. This equality makes it a critical issue for autoregressive visual generation to convert images into one-dimensional sequential tokens [65]. Methods based on spatial scanning [10, 34, 47, 62] attempt to discretize the image locally into tokens according to patches, and then perform autoregressive generation following a certain order based on their locations. However, the resulting spatial sequence violates the equality among image patches, making it suboptimal for causal autoregressive modeling. Another line of work [4, 31, 66] introduces bidirectional interaction in the generator as a workaround. Nonetheless, they still tokenize images spatially, thus assuming a causal order among image patches. In addition, the bidirectional design may deviate from the conventional autoregressive paradigm, complicating their integration into omni-modal frameworks [49, 56]. In contrast to spatial tokenization, VAR [50] explores transforming the image into multiple scales and producing a sequence by concatenating tokens from ordered scales. Although scale-wise autoregressive generation indeed satisfies the equality of image pixels, it suffers from inferior token efficiency and parallel generation of multiple tokens from the same scale. + +In this paper, we introduce a spectral autoregressive visual generation framework to achieve causal autoregressive modeling and improve token efficiency, as shown in Figure 2. Frequency is an inherent attribute of all types of signals and has become a significant perspective and methodology that complements the spatial-temporal domain [2]. For visual data, the spectral density often conforms to the power-law distribution [6], with low-frequency components representing the overall structure of an image and high-frequency components focusing on the intricate details. This hierarchical coarse-to-fine nature of the correspondence between spectral and spatial domains indicates a se + +![](images/158aed395cad4765210f69b7ac64517cfebfe9a6db00f599a8d6d0c04516dffa.jpg) +Token Efficiency +Figure 2. Comparison between autoregressive visual generation methods. SpectralAR achieves both token efficiency and sequence causality with nested 1D spectral tokens. + +quential order of images, motivating us to represent images as causal spectral sequences. Specifically, we first transform images into spectral tokens with Nested Spectral Tokenization (NST), which uses varying sequence lengths to represent images across different frequency bands. The causality of the spectral sequences originates from the coarse-to-fine progression characteristic resembling human visual perception and is strengthened with the application of a causal mask. In addition, inspired by the image compression literature [55, 58], we design a non-uniform token-frequency mapping, which allocates more tokens to represent low-frequency components and fewer for high-frequency components. This mapping technique greatly reduces the number of tokens, while maintaining the quality of reconstructed samples. In the autoregressive process, we begin with the token representing the DC component and progressively predict tokens corresponding to higher frequencies conditioned on previous ones. We conduct extensive experiments on the ImageNet-1K dataset for image reconstruction and class-conditional generation. Our SpectralAR demonstrates comparable performance with state-of-the-art methods. + +# 2. Related Work + +Autoregressive visual generation. While diffusion models have achieved exceptional performance, they are fundamentally different from the conventional autoregressive framework [49, 56] for cross-modal and cross-task modeling. Therefore, a body of research aims to advance autore- + +gressive models for visual generation. Early efforts [5, 51] perform pixel-level generation in the row-major raster-scan order, followed by VQGAN [10] which transfers to the latent feature space of VQVAE [52] for autoregressive modeling. Subsequent work improves based on VQGAN with multiple scales [42], residual quantization [29], ViT architecture [63] or textual conditions [64]. However, raster-scan generation violates the equality between image pixels as discussed in Section 1, contradicting the causality premise of autoregressive modeling. Recently, VAR [50] proposes the scale-wise autoregressive generation that aims to predict the next-scale token map conditioned on the previous ones. Despite achieving causality with the multi-scale design, VAR predicts multiple tokens with bidirectional attention in each step, thus deviating from the standard autoregressive framework. In contrast, our method realizes causal autoregressive generation from the spectral perspective and still conforms to the unidirectional scheme. + +Efficient image tokenization. Autoregressive generation requires the conversion of images into token sequences, which is often achieved using autoencoders [18, 54]. Patch-based autoencoders [10, 24, 37, 52, 72] tokenize images spatially, where each token corresponds to a certain patch from the original image. Although this paradigm performs well in image reconstruction and diffusion-based generation [39, 44], it is not suitable for autoregressive modeling due to its spatial design. Also, its token length is proportional to the square of image resolution, which might become the bottleneck in multimodal modeling given limited context length [49]. TiTok [66] proposes a 1D tokenizer which reduces the number of tokens to 32 for $256 \times 256$ images. However, TiTok is trained with only an overall reconstruction objective, and thus the precise meaning of these 1D tokens remains unclear. VAR [50] introduces a multi-scale tokenizer for causal autoregressive image generation. Nonetheless, the multi-scale strategy requires an even greater number of tokens compared to patch-based tokenization methods, further diminishing token efficiency. Our method converts images into 1D causal sequences of spectral tokens and enhances token efficiency by leveraging the long-tail distribution of spectral density in image data. + +Spectral visual analysis. Spectral analysis [2] has been a common technique in computer vision, complementing the spatial and temporal domains. Representative applications include image enhancement and denoising [15, 68], texture analysis and feature extraction [41], compression and super-resolution [13, 55], visual generation [61, 67], adversarial attacks and defenses [12, 32, 33]. For autoregressive image generation, CART [43] and SIT [11] propose to transform an image into multiple causal sets of tokens with base-detail decomposition and discrete wavelet transform, respectively. However, these methods still adhere to the multi-scale 2D tokenization paradigm similar to + +VAR, resulting in suboptimal token efficiency and bidirectional attention to predict multiple tokens per autoregressive step. In contrast, SpectralAR leverages the discrete cosine transform to capture the global information of an image, and compress it into a 1D sequence with high efficiency. + +# 3. Proposed Approach + +# 3.1. Revisiting Images from the Spectral Domain + +Discrete cosine transform. Spectral analysis investigates how complex signals can be represented with simpler basis functions [2], producing a spectral density distribution that represents the magnitude of corresponding basic components. This spectral density distribution is an equivalent representation of the original signal and provides a distinct perspective from the spatial domain depending on the properties of the basis functions. We employ the Discrete Cosine Transform (DCT) [1] to convert images into the spectral domain. The DCT result $\mathbf{D}$ shares the same shape with the transformed image $\mathbf{I} \in \mathbb{R}^{H \times W}$ (for simplicity, we omit the channel dimension): + +$$ +\mathbf {D} = \left\{F (u, v) \right\} _ {u, v = 1} ^ {W, H}, \quad \mathbf {I} = \left\{f (x, y) \right\} _ {x, y = 1} ^ {W, H}. \tag {1} +$$ + +Each $F(u,v)$ represents the intensity of the corresponding basis function $g_{u,v}(x,y)$ in the image $\mathbf{I}$ , which writes: + +$$ +g _ {u, v} (x, y) = \frac {2 C (u) C (v)}{\sqrt {H W}} \cos \frac {(2 x + 1) u \pi}{2 W} \cos \frac {(2 y + 1) v \pi}{2 H}, \tag {2} +$$ + +where $C(u) = 1 / \sqrt{2}$ if $u = 0$ and $C(u) = 1$ otherwise. This family of basis functions has the following properties: (1) Given $u$ and $v$ , the basis function $g_{u,v}(x,y)$ exhibits a checkerboard-like pattern in the spatial domain, with periods along the x- and y-axis of $2W / u$ and $2H / v$ , respectively. This pattern suggests that the basis function characterizes the rate of variation of images in the spatial domain. (2) The $F(u,v)$ s also form a 2D matrix together, where the top-left corner represents the low-frequency components (small $u, v$ ), while the bottom-right corner corresponds to the high-frequency components (large $u, v$ ). + +Causality from the spectral domain. Since low- and high-frequency components describe overall structures and intricate details, respectively, we can decompose an image into a sequence of sub-images $\{\mathbf{I}_i^{\prime}\}_{i = 1}^{L}$ with increasing levels of detail by applying inverse-DCT on partially masked spectral density distributions $\{\mathbf{D}_i^{\prime}\}_{i = 1}^{L}$ . With more high-frequency components, the sub-images will gradually transition from blurred to sharp, as shown in Figure 2. This hierarchical coarse-to-fine sequence aligns with human visual perception and artistic drawing, thus enhancing causality. + +Efficiency from the spectral domain. As the image compression literature [1] pointed out, the DCT result $\mathbf{D}$ of images conforms to the power-law distribution with the absolute values of $F(u,v)\mathrm{s}$ on the top-left corner substantially + +![](images/2b39acda8d76dcefead60a529f616d9fbe05af9a72d331a7794a11cc7fff2558.jpg) +Figure 3. Overall pipeline of the proposed SpectralAR. Left: We convert an input image into a 1D causal sequence efficiently with nested spectral tokenization. Each nested sequence is supervised with distinct reconstruction target in a coarse-to-fine manner, which endows each token with an explicit spectral interpretation. We also apply the causal mask to the spectral tokens in the encoder and decoder to enhance the causality. Right: We use the groundtruth sequences from the tokenization process to train an autoregressive generation model. + +larger than the bottom-right corner, indicating that most energy of an image is concentrated in the low-frequency components. In addition, human visual perception is less sensitive to high-frequency signals, which have little influence on the visual quality of images. Based on these distinctions, we can encode the high-frequency components of an image with coarser granularity to improve token efficiency, similar to the JPEG algorithm [55] which saves more than $90\%$ storage for images by suppressing high-frequency signals. + +# 3.2. Nested Spectral Tokenization + +Overall framework. In contrast to the 2D spatial tokenization that captures the local correlation between image patches, the basis function $g_{u,v}(x,y)$ of DCT encodes a global frequency pattern. Therefore, we convert images into 1D tokens in spectral tokenization to reflect the global nature of the basis functions. We start with the general framework of a 1D image tokenizer [66]. Given an image $\mathbf{I}$ , we aim to encode it into $N$ discrete vectors $\mathbf{S}$ , and also reconstruct the original image with $\mathbf{S}$ . We first patchify the image into $\mathbf{P} \in \mathbb{R}^{hw\times C}$ , and concatenate the image features with the initial query vectors $\mathbf{S}_0$ to form $[\mathbf{P};\mathbf{S}_0] \in \mathbb{R}^{(hw + N)\times C}$ , where $h,w,C$ denote the resolution of image features and the channel dimension, respectively. We then employ the vision transformer $\mathcal{E}$ to enable feature extraction and interaction between the image features and 1D query tokens, resulting in the informative 1D representation $\hat{\mathbf{S}}$ . In the vector quantizer $\mathcal{Q}$ , we match these continuous vectors with the + +codebook embeddings to derive the discrete representation $\mathbf{S}$ , which could serve as the groundtruth for the autoregressive training. At last, we append $\mathbf{S}$ to a set of mask tokens $\mathbf{M} \in \mathbb{R}^{hw \times C}$ and process them with the decoder network $\mathcal{D}$ similar to $\mathcal{E}$ , in order to reconstruct the original image. This overall framework could be formulated as: + +$$ +\hat {\mathbf {S}} = \mathcal {E} ([ \mathbf {P}; \mathbf {S} _ {0} ]), \quad \mathbf {S} = \mathcal {Q} (\hat {\mathbf {S}}), \quad \hat {\mathbf {I}} = \mathcal {D} ([ \mathbf {M}; \mathbf {S} ]), \quad (3) +$$ + +where $\hat{\mathbf{I}}$ denotes the reconstructed image. The training objective typically consists of multiple loss functions: + +$$ +\mathcal {L} _ {t o k} = \| \hat {\mathbf {I}} - \mathbf {I} \| _ {2} ^ {2} + \| \hat {\mathbf {S}} - \mathbf {S} \| _ {2} ^ {2} + \mathcal {L} _ {P} (\hat {\mathbf {I}}, \mathbf {I}) + \mathcal {L} _ {A} (\hat {\mathbf {I}}, \mathbf {I}), (4) +$$ + +where $\mathcal{L}_P$ and $\mathcal{L}_A$ denote the perceptual loss [22, 70] and the adversarial loss [10, 14], respectively. + +Nested spectral decoding. Different from the plain 1D tokenization in Figure 2, we aim to represent an image as its spectral decompositions $\{\mathbf{I}_i^{\prime}\}_{i = 1}^{L}$ for sequence causality, which requires establishing a mapping from 1D tokens to these sub-images. One naive way to achieve this would be dividing S into disjoint subsets and assigning them to model different sub-images independently. Similar to the multi-scale tokenization [50], this strategy would inevitably involve bidirectional interaction and diminish token efficiency because sub-images with finer detail would require increasingly more tokens to represent. In contrast, we propose a nested mapping scheme for efficient tokenization, as shown in the bottom of Figure 3. We first construct a + +# Algorithm 1: Nested Spectral Tokenization Training + +1 Inputs: raw image $\mathbf{I}$ , initial spectral tokens $\mathbf{S}_0\in \mathbb{R}^{N\times C}$ +2 Hyperparameters: spectral levels $\{\omega_i|i = 1,\dots,N\}$ +3 $\mathbf{P} =$ patchify(I), $\hat{\mathbf{S}} = \mathcal{E}([\mathbf{P};\mathbf{S}_0])$ $\mathbf{S} = \mathcal{Q}(\hat{\mathbf{S}})$ +4 $idx = \mathrm{random\_choice}(N)$ +5 $\mathbf{S}^{\prime} = \mathbf{S}[:idx],\hat{\mathbf{I}}^{\prime} = \mathcal{D}([\mathbf{M};\mathbf{S}^{\prime}])$ +$\mathbf{6}$ $\mathbf{D} = \mathrm{DCT}(\mathbf{I}),\mathbf{D}' = \mathbf{D}\circ \mathbf{1}_{\omega_i},\mathbf{I}' = \mathrm{DCT}^{-1}(\mathbf{D}')$ +7 loss $= \mathcal{L}_{tok}(\hat{\mathbf{l}}',\mathbf{l}')$ +8 Return: loss for optimization; + +sequence of sub-images with increasing detail by progressively preserving larger regions in the spectral density $\mathbf{D}$ : + +$$ +\mathbf {I} _ {i} ^ {\prime} = \mathrm {D C T} ^ {- 1} \left(\mathbf {D} _ {i} ^ {\prime}\right), \quad \mathbf {D} _ {i} ^ {\prime} = \mathbf {D} \circ \mathbf {1} _ {\omega_ {i}}, \quad \omega_ {i - 1} < \omega_ {i}, \tag {5} +$$ + +where $\mathrm{DCT}^{-1}$ , $\circ$ , $\mathbf{1}_{\omega_i}$ denote the inverse DCT operation, element-wise multiplication and a $H\times W$ matrix with the top-left corner of size $\omega_{i}\times \omega_{i}$ filled by ones and the remaining parts being zeros. Therefore, the sub-image $\mathbf{I}_i^{\prime}$ contains all frequency components present in the sub-image $\mathbf{I}_{i - 1}^{\prime}$ . Based on this inclusion property, we can reuse the tokens representing the previous sub-image $\mathbf{I}_{i - 1}^{\prime}$ to represent the next sub-image $\mathbf{I}_i^\prime$ . To avoid bidirectional attention, we make $L$ equal to $N$ so that each token $\mathbf{s}_i$ in the sequence S corresponds to a unique sub-image $\mathbf{I}_i^\prime$ : + +$$ +\tilde {\mathbf {I}} _ {i} ^ {\prime} = \mathcal {D} ([ \mathbf {M}; \mathbf {s} _ {1}, \mathbf {s} _ {2}, \dots , \mathbf {s} _ {i} ]), \tag {6} +$$ + +where $\hat{\mathbf{I}}_i^{\prime}$ is the reconstruction of sub-image $\mathbf{I}_i^\prime$ given the nested 1D sequence $\{\mathbf{s}_1,\mathbf{s}_2,\dots ,\mathbf{s}_i\}$ . Nested spectral decoding compresses an image into a causal 1D sequence where each token $\mathbf{s}_i$ corresponds to a disjoint set of frequencies and achieves token efficiency by reusing previous tokens. + +Non-uniform token-frequency mapping. To further enhance token efficiency, we introduce the non-uniform token-frequency mapping technique. Since high-frequency components have low magnitude and minimal impact on the visual quality of images, we can encode them with coarser granularity compared to the low-frequency counterparts. We achieve this by reducing the interval between $\omega_{i-1}$ and $\omega_i$ when $i$ is small, and increase it otherwise: + +$$ +\omega_ {i} - \omega_ {i - 1} \leq \omega_ {i + 1} - \omega_ {i}, \tag {7} +$$ + +which demonstrates the general case. This non-uniform mapping allocates later tokens to broader frequency ranges, enabling precise modeling of crucial low-frequency components while efficiently representing high-frequency details. + +Spectral causal mask. Although the sequence $\mathbf{S}$ supervised with (5)(6) already exhibits a certain degree of causality, the encoding and decoding processes, i.e. $\mathcal{E}$ and $\mathcal{D}$ , are still bidirectional, which can lead to information leakage from high-frequency to low-frequency components. Therefore, we propose applying causal masks to the spectral tokens $\mathbf{S}$ in both the encoder $\mathcal{E}$ and the decoder $\mathcal{D}$ , as shown + +Table 1. Correlation between tokens of different autoregressive paradigms. We use linear correlation as a proxy metric for the causality of sequences. The spectral sequence demonstrates better causality compared with other methods. + +
Correlation TypeRaster-scanScale-wiseSpectral
R2avg(t2; t1)0.4710.8890.916
R2avg(t3; t1, t2)0.3660.9530.977
R2avg(t4; t1, t2, t3)0.5250.9430.994
Average0.4540.9280.962
+ +![](images/6f0262dc63c5c27aadfc558f7b636ba1aa0a6311fa06b784b9805f52a5c08d96.jpg) +Raster-scan +Figure 4. Frequency of each element having the highest correlation with another one. It represents the source of reference information for predicting the next token. The raster-scan method exhibits excessive dependency on boundaries of image patches. + +![](images/a6881fb4872c5f10b14367c193ba56928c19b9afcc405379f1b6e24877c9be6a.jpg) +Scale-wise + +![](images/3a959dd944bf178c2326bffa2e8707fe4bb684fa09eb448adcf762c190c01129.jpg) +Spectral + +![](images/16229f95e468c3d09cb5cefecca1954ce71d5bfc784494c544366abb177f5dd5.jpg) + +in Figure 3. This spectral causal mask restricts each token s to only attend to tokens that represent lower frequencies, enhancing causality from the architectural perspective. We outline the training procedure in Algorithm 1. + +# 3.3. Spectral Autoregressive Generation + +Autoregressive modeling has gained prominence in computer vision for its scalability, generalization, and effectiveness across multimodal tasks [25, 49, 56]. While conventional autoregressive generation follows a spatial raster-scan order [10, 51], we propose a hierarchical coarse-to-fine approach in the spectral domain to enhance causality, as shown in the right side of Figure 3. We start with the general framework of autoregressive modeling: + +$$ +\mathbf {p} _ {i + 1} = \mathcal {M} \left(\mathbf {t} _ {1}, \mathbf {t} _ {2}, \dots , \mathbf {t} _ {i}\right), \tag {8} +$$ + +where $\mathbf{T} = \{\mathbf{t}_i\}_{i=1}^N$ is a sequence of quantized tokens, and $\mathbf{p}_{i+1}$ , $\mathcal{M}$ denote the probability logits for the $(i+1)$ th token and the autoregressive model, respectively. The autoregressive process (8) assumes that the generation of token $\mathbf{t}_{i+1}$ depends solely on the previous ones. The spatial autoregressive paradigm violates this premise because of the equality of image pixels (as discussed in Section 1). In contrast, we take the spectral tokens $\mathbf{S}$ from the nested spectral tokenziation as the autoregressive targets $\mathbf{T}$ . Since the spectral tokens are trained in a nested manner to reconstruct sub-images of increasing levels of detail as in (5)(6), each spectral token $\mathbf{s}_i$ is expected to enhance the quality of the sub-image $\mathbf{I}_{i-1}'$ represented by previous tokens from the spectral domain. This progressive refinement process aligns with human visual perception and artistic painting, both of which start with the overall structure and gradually focus on + +Table 2. Comparison between generative models on class-conditional ImageNet ${256} \times {256}$ benchmark. “ $\downarrow$ ” or “↑” indicate lower or higher values are better. "Token": the number of tokens used in transformer architectures. "Step": the number of model runs needed to generate an image. We compute our wall-clock inference time and scale the Time for other methods accordingly. $\dagger$ : trained on larger datasets including OpenImages [27]. ‡: implemented with the official tokenizer weight [66] and the scripts from VAR. + +
TypeModelrFID↓gFID↓IS↑Pre↑Rec↑#Token#Para#StepTime
GANBigGAN [3]75.246.95224.50.890.38112M1-
GANGigaGAN [23]-3.45225.50.840.61N.A.569M1-
GANStyleGan-XL [45]7.062.30265.10.780.53166M10.4
Diff.ADM [8]125.7810.94101.00.690.63N.A.554M250235
Diff.CDM [20]-4.88158.7--N.A.-8100-
Diff.LDM-4-G [44]0.27†3.60247.7--N.A.400M250-
Diff.DiT-L/2 [39]0.62†5.02167.20.750.57256458M25043
Diff.DiT-XL/2 [39]0.62†2.27278.20.830.57256675M25063
Diff.L-DiT-3B [71]-2.10304.40.820.602563.0B250>63
Diff.L-DiT-7B [71]-2.28316.20.830.582567.0B250>63
Mask.MaskGIT [4]2.286.18182.10.800.51256227M80.7
Mask.RCG (cond.) [30]-3.49215.5---502M202.7
Mask.TiTok-B64 [66]1.702.48214.7--64177M80.4
2D ScanVQVAE-2 [42]-31.11~450.360.57N.A.13.5B5120-
2D ScanVQGAN [10]7.9415.7874.3--N.A.1.4B25634
2D ScanViTVQ [63]1.284.17175.1--10241.7B1024>34
2D ScanRQTran. [29]3.207.55134.0--64,43.8B6829
VARVAR-d16 [50]3.30274.40.840.51310M100.6
VARVAR-d20 [50]2.57302.60.830.56600M100.7
VARVAR-d24 [50]2.09312.90.820.591.0B100.8
VARVAR-d30 [50]1.92323.10.820.592.0B101.4
1D ARTiTok-B64-d16‡ [66]1.706.30190.10.850.4764310M641
1D ARSpectralAR-d164.033.02282.20.810.5564310M641
1D ARSpectralAR-d204.032.49305.4--64600M641.2
1D ARSpectralAR-d244.032.13307.7--641.0B641.4
1D ARSpectralAR-d16-p44.033.13276.1--16×4310M160.4
+ +details. This similarity qualitatively validates the rationale for performing causal autoregressive generation in the spectral domain. We further provide some quantitative analysis through a toy experiment in Section 4.2. + +Potential applications. The frequencies represented by the token $\mathbf{s}_i$ become higher as $i$ increases, while its influence on the image quality diminishes accordingly (check Section 3.1 for details). Therefore, we can control the visual quality of sampled images by adjusting the length of generated sequences, similar to the image compression algorithms [55]. In addition, we can achieve super-resolution by dividing images into disjoint parts smaller than $H\times W$ , and conducting individual spectral autoregressive generation on each part. We provide further results in Section 4.4. + +# 4. Experiments + +# 4.1. Dataset and Implementation Details + +We train and evaluate our SpectralAR on the ImageNet1K [7] benchmark, which contains 1,281,167 and 50,000 + +images for training and validation, respectively. We train the tokenizer and generator on the training split. We evaluate the reconstruction performance on the validation set with reconstruction Fréchet inception distance [17] (rFID), and the generation results with generation FID (gFID) using pre-computed statistics and scripts from ADM [8]. + +For tokenizer training, we follow the exactly same settings of TiTok [66] for a fair comparison. We also employ the two-stage training strategy with proxy codes [4, 66]. We use the ViT-B [9] as the encoder and decoder, and set the number of spectral tokens as $N = 64$ in our main experiments and the sequence $\omega$ s for $256 \times 256$ images as: + +$$ +\omega_ {i} = \left\{ \begin{array}{l l} i, & \text {i f} i \in (0, 3 2 ], \\ 2 i - 3 2, & \text {i f} i \in (3 2, 4 8 ], \\ 1 2 i - 5 1 2, & \text {i f} i \in (4 8, 6 4 ]. \end{array} \right. \tag {9} +$$ + +For generator training, we adopt the same architecture and training recipe as VAR [50], which leverages a GPT-2-like transformer architecture [40] for autoregressive modeling. + +![](images/730d1fcfc5fc12876dca33d6b4a5f03440c9ed2823c9e73856251edf6e8cba57.jpg) +Figure 5. Visualizations of generated $256 \times 256$ samples on ImageNet-1K. These samples cover a wide range of categories and styles, demonstrating the ability of SpectralAR to generate both diverse and high-quality images. + +# 4.2. Proof of Concept + +We conduct a proof-of-concept experiment to validate the causality of images in the spectral domain. We use the validation set of the CIFAR-100 dataset [26] as a proxy for the distribution of natural images, which contains 10,000 $32 \times 32$ images from 100 semantic classes. We aim to measure the linear correlation $(R^2)$ [38] between one token $\mathbf{t}_i$ and its previous ones $\{\mathbf{t}_j\}_{j=1}^{i-1}$ as a metric for the causality of the sequence $\mathbf{T}$ . We consider three typical autoregressive paradigms, including raster-scan [10], scale-wise [50] and spectral methods which construct sequence $\mathbf{T}$ in different ways. Specifically, we divide each image spatially into four patches of size $16 \times 16$ and rearrange them in raster-scan order for the first variant. For the scale-wise method, we + +downsample each image into sizes of $8 \times 8$ , $16 \times 16$ , $24 \times 24$ and $32 \times 32$ , respectively. We set the frequency thresholds to 8, 16, 24, 32 and generate the sub-images following (5) for the spectral sequence. Since each token $\mathbf{t}$ is a high dimensional tensor, we calculate $R^2$ between every element of $\mathbf{t}_i$ and every element of $\{\mathbf{t}_j\}_{j=1}^{i-1}$ . For each element of $\mathbf{t}_i$ , we then find the element in $\{\mathbf{t}_j\}_{j=1}^{i-1}$ with the highest correlation, and take an average as the final measurement: + +$$ +R _ {a v g} ^ {2} = \operatorname {A V G} _ {m} \left(\max _ {n} R _ {m, n} ^ {2}\right), \tag {10} +$$ + +where $m, n$ denote element number of $\mathbf{t}_i$ and $\{\mathbf{t}_j\}_{j=1}^{i-1}$ , respectively, and $R_{m,n}^2$ is the correlation between the $m$ th and $n$ th elements. We report the quantitative results in Table 1, where the scale-wise and spectral sequences demonstrate + +Table 3. Applications of SepctralAR. Super-reso., Trunc. represent super-resolution and truncated, respectively. + +
App. TypeModelFID↓IS↑
Super-reso.Upsample3.09286.6
SpectralAR-Stride2.93276.4
SpectralAR-Patch14.76170.0
Trunc.SpectralAR-Trunc.53.34271.5
SpectralAR-Trunc.106.65211.4
SpectralAR-Trunc.1527.6991.69
SpectralAR-Trunc.03.02282.2
+ +considerably higher correlation than the raster-scan counterpart. This could be attributed to the coarse-to-fine nature of the former two paradigms, while the raster scan method lacks adequate reference information to predict the next image patch, as shown in Figure 4. + +# 4.3. Main Results + +We report the performance of SpectralAR on the class-conditional ImageNet-1K [7] $256 \times 256$ generation benchmark in Table 2. We also implement an autoregressive version of TiTok [66] by using the official tokenizer weight and scripts from VAR [50] for a fair comparison. The reconstruction performance of SpectralAR (4.03 rFID) is inferior compared with TiTok because SpectralAR requires to reconstruct the sub-images corresponding to different frequencies with different lengths of tokens, which is much more difficult compared with the overall reconstruction target in TiTok. Despite the lower reconstruction score, SpectralAR-d16 still outperforms TiTok-B64-d16 and VAR-d16 in autoregressive generation with a clear margin due to better sequence causality which eases autoregressive learning. Furthermore, SpectralAR also achieves better or comparable performance against VAR under the $d20$ and $d24$ settings, respectively. In addition, SpectralAR uses only 64 tokens in both reconstruction and generation, demonstrating superior token efficiency compared to VAR [50] and 2D scan-based methods. We also visualize the generated samples in Figure 5, which shows the diversity and quality of the generation process of SpectralAR. In addition, Figure 1 highlights the hierarchical coarse-to-fine refinement of the images, while SpectralAR generates more tokens in an autoregressive way. + +# 4.4. Applications + +In this section, we provide a quantitative analysis for the potential applications of SpectralAR. For super-resolution, we conduct our experiments based on the $256 \times 256$ images generated by SpectralAR- $d16$ in Table 2, and directly upsample them to $512 \times 512$ as the baseline. We construct 4 sub-images using stripped and patch-based methods for SpectralAR-Stride and -Patch, respectively. We then upsample the 4 sub-images to $256 \times 256$ and use SpectralAR + +Table 4. Ablation on design choices. Spectral Supervision means using sub-image supervision across different frequency bands. Causal Mask and non-uniform mapping refers to the spectral causal mask and the frequency-token mapping, respectively. + +
Spectral SupervisionCausal MaskNon-uniform MappingFID↓IS↑
×××6.30190.1
✓××5.64255.1
✓✓×3.49222.6
✓✓✓3.02282.2
+ +to refine them in the spectral domain, and finally reassemble them to generate the final result. According to Table 3, SpectralAR can indeed serve as a spectrum completer for the super-resolution task. In addition, we also experiment with truncated autoregressive generation, where we discard the last few tokens. For example, we discard the last 5 tokens in SpectralAR-Trunc.5 in Table 3. The generation performance worsens slowly when the number of truncated tokens is fewer than 10, and therefore it is possible to further improve token efficiency through truncation according to the requirement for generation quality. + +# 4.5. Ablation Study + +We conduct ablation study to validate the effectiveness of our design choices in Table 4. The first row corresponds to the baseline TiTok implementation for autoregressive generation. Note that the spectral supervision alone could improve FID compared with the TiTok counterpart. This is because the spectral design enhances sequence causality compared with the overall reconstruction target which ignores the correlation between tokens and thus complicates the autoregressive modeling process. The spectral causal mask further enhances performance by improving causality in the encoding and decoding architecture. And the non-uniform token-frequency mapping technique guides the model to focus more on the crucial low-frequency components while representing the high-frequency components efficiently, thus further improving performance. + +# 5. Conclusion + +In this paper, we have proposed the spectral autoregressive visual generation method for both causal and efficient autoregressive modeling of image data. Specifically, we first convert images into 1D sequences with nested spectral tokenization. In addition, we have adopted causal masks for spectral tokens in the encoder and decoder to further enhance causality from the architectural perspective. We have also designed a non-uniform token-frequency mapping with emphasis on the low-frequency components in order to improve token efficiency. On the ImageNet-1K generation benchmark, our SpectralAR achieves superior performance compared with other autoregressive generation methods. + +# Acknowledgement + +This work was supported in part by the National Natural Science Foundation of China under Grant 62125603, Grant 62321005, Grant 62336004, and Grant 62441616, and in part by the Beijing Natural Science Foundation under Grant No. L247009. + +# References + +[1] Nasir Ahmed, T. Natarajan, and Kamisetty R Rao. Discrete cosine transform. IEEE transactions on Computers, 100(1): 90-93, 2006. 3 +[2] Jean Baptiste Joseph Baron Fourier et al. The analytical theory of heat. Courier Corporation, 2003. 2, 3 +[3] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 6 +[4] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In CVPR, pages 11315-11325, 2022. 2, 6 +[5] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In ICML, pages 1691-1703. PMLR, 2020. 3 +[6] Aaron Clauset, Cosma Rohilla Shalizi, and Mark EJ Newman. Power-law distributions in empirical data. SIAM review, 51(4):661-703, 2009. 2 +[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009. 6, 8 +[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NIPS, 34:8780-8794, 2021. 6 +[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6 +[10] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, pages 12873-12883, 2021. 2, 3, 4, 5, 6, 7 +[11] Carlos Esteves, Mohammed Suhail, and Ameesh Makadia. Spectral image tokenizer. arXiv preprint arXiv:2412.09607, 2024.3 +[12] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In ICML, pages 3247-3258. PMLR, 2020. 3 +[13] Manuel Fritsche, Shuhang Gu, and Radu Timofte. Frequency separation for real-world super-resolution. In ICCVW, pages 3599-3608. IEEE, 2019. 3 +[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. NIPS, 27, 2014. 2, 4 + +[15] Hayit Greenspan, Charles H Anderson, and Sofia Akber. Image enhancement by nonlinear extrapolation in frequency space. TIP, 9(6):1035-1048, 2000. 3 +[16] Jian Han, Jinlai Liu, Yi Jiang, Bin Yan, Yuqi Zhang, Zehuan Yuan, Bingyue Peng, and Xiaobing Liu. Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis. arXiv preprint arXiv:2412.04431, 2024. 2 +[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NIPS, 30, 2017. 6 +[18] Geoffrey E Hinton and Ruslan R Salakhutdinov. Reducing the dimensionality of data with neural networks. science, 313(5786):504-507, 2006. 3 +[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NIPS, 33:6840-6851, 2020. 1 +[20] Jonathan Ho, Chitwan Saharia, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. JMLR, 23(47):1-33, 2022. 6 +[21] Yuanhui Huang, Wenzhao Zheng, Yuan Gao, Xin Tao, Pengfei Wan, Di Zhang, Jie Zhou, and Jiwen Lu. Owl1: Omni world model for consistent long video generation. arXiv preprint arXiv:2412.09600, 2024. 2 +[22] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, pages 694–711. Springer, 2016. 4 +[23] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In CVPR, pages 10124-10134, 2023. 6 +[24] Diederik P Kingma, Max Welling, et al. Auto-encoding variational bayes, 2013. 3 +[25] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Grant Schindler, Rachel Hornung, Vighnesh Birodkar, Jimmy Yan, Ming-Chang Chiu, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 5 +[26] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 7 +[27] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 128(7):1956-1981, 2020. 6 +[28] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 2 +[29] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In CVPR, pages 11523-11532, 2022. 2, 3, 6 +[30] Tianhong Li, Dina Katabi, and Kaiming He. Self-conditioned image generation via generating representations. CoRR, 2023. 6 + +[31] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. NIPS, 37:56424-56445, 2025. 2 +[32] Yuyang Long, Qilong Zhang, Boheng Zeng, Lianli Gao, Xianglong Liu, Jian Zhang, and Jingkuan Song. Frequency domain model augmentation for adversarial attack. In ECCV, pages 549-566. Springer, 2022. 3 +[33] Cheng Luo, Qinliang Lin, Weicheng Xie, Bizhu Wu, Jinheng Xie, and Linlin Shen. Frequency-driven imperceptible adversarial attack on semantic similarity. In CVPR, pages 15315-15324, 2022. 3 +[34] Zhuoyan Luo, Fengyuan Shi, Yixiao Ge, Yujiu Yang, Limin Wang, and Ying Shan. Open-magvit2: An open-source project toward democratizing auto-regressive visual generation. arXiv preprint arXiv:2409.04410, 2024. 2 +[35] Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. In ECCV, pages 23-40. Springer, 2024. 2 +[36] Xiaoxiao Ma, Mohan Zhou, Tao Liang, Yalong Bai, Tiejun Zhao, Huaian Chen, and Yi Jin. Star: Scale-wise text-to-image generation via auto-regressive representations. arXiv preprint arXiv:2406.10797, 2024. 2 +[37] Fabian Mentzer, David Minnen, Eirikur Agustsson, and Michael Tschannen. Finite scalar quantization: Vq-vae made simple. arXiv preprint arXiv:2309.15505, 2023. 3 +[38] Daniel J Ozer. Correlation and the coefficient of determination. Psychological bulletin, 97(2):307, 1985. 7 +[39] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 1, 2, 3, 6 +[40] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 2, 6 +[41] Yongming Rao, Wenliang Zhao, Zheng Zhu, Jiwen Lu, and Jie Zhou. Global filter networks for image classification. Advances in neural information processing systems, 34:980-993, 2021. 3 +[42] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. NIPS, 32, 2019. 2, 3, 6 +[43] Siddharth Roheda. Cart: Compositional auto-regressive transformer for image generation. arXiv preprint arXiv:2411.10180, 2024. 3 +[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 3, 6 +[45] Axel Sauer, Katja Schwarz, and Andreas Geiger. Styleganx1: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH 2022 conference proceedings, pages 1-10, 2022. 6 +[46] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1 +[47] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model + +beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 2 +[48] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 2 +[49] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 2, 3, 5 +[50] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. NIPS, 37:84839-84865, 2025. 2, 3, 4, 6, 7, 8 +[51] Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelCNN decoders. NIPS, 29, 2016. 3, 5 +[52] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NIPS, 30, 2017. 3 +[53] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NIPS, 30, 2017. 2 +[54] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In ICML, pages 1096-1103, 2008. 3 +[55] Gregory K Wallace. The jpeg still picture compression standard. Communications of the ACM, 34(4):30-44, 1991. 2, 3, 4, 6 +[56] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 2, 5 +[57] Yuqing Wang, Tianwei Xiong, Daquan Zhou, Zhijie Lin, Yang Zhao, Bingyi Kang, Jiashi Feng, and Xihui Liu. Loong: Generating minute-level long videos with autoregressive language models. arXiv preprint arXiv:2410.02757, 2024. 2 +[58] Thomas Wiegand, Gary J Sullivan, Gisle Bjontegaard, and Ajay Luthra. Overview of the h. 264/avc video coding standard. TCSVT, 13(7):560-576, 2003. 2 +[59] Chengyue Wu, Xiaokang Chen, Zhiyu Wu, Yiyang Ma, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, Chong Ruan, et al. Janus: Decoupling visual encoding for unified multimodal understanding and generation. arXiv preprint arXiv:2410.13848, 2024. 2 +[60] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 2 +[61] Mengping Yang, Zhe Wang, Ziqiu Chi, and Wenyi Feng. Wavegan: Frequency-aware gan for high-fidelity few-shot image generation. In ECCV, pages 1-17. Springer, 2022. 3 +[62] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 2 + +[63] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 3, 6 +[64] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 3 +[65] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Randomized autoregressive visual generation. arXiv preprint arXiv:2411.00776, 2024. 2 +[66] Qihang Yu, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An image is worth 32 tokens for reconstruction and generation. NIPS, 37: 128940-128966, 2025. 2, 3, 4, 6, 8 +[67] Yingchen Yu, Fangneng Zhan, Shijian Lu, Jianxiong Pan, Feiying Ma, Xuansong Xie, and Chunyan Miao. Wavefill: A wavelet-based generation network for image inpainting. In ICCV, pages 14114-14123, 2021. 3 +[68] Huanjing Yue, Xiaoyan Sun, Jingyu Yang, and Feng Wu. Cid: Combined image denoising in spatial and frequency domains using web images. In CVPR, pages 2933-2940, 2014. 3 +[69] Qian Zhang, Xiangzi Dai, Ninghua Yang, Xiang An, Ziyong Feng, and Xingyu Ren. Var-clip: Text-to-image generator with visual auto-regressive modeling. arXiv preprint arXiv:2408.01181, 2024. 2 +[70] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 4 +[71] Renrui Zhang, Jiaming Han, Chris Liu, Peng Gao, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, and Yu Qiao. Llama-adapter: Efficient finetuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 6 +[72] Chuanxia Zheng, Tung-Long Vuong, Jianfei Cai, and Dinh Phung. Movq: Modulating quantized vectors for high-fidelity image generation. NIPS, 35:23412-23425, 2022. 3 +[73] Wenzhao Zheng, Weiliang Chen, Yuanhui Huang, Borui Zhang, Yueqi Duan, and Jiwen Lu. Occworld: Learning a 3d occupancy world model for autonomous driving. In ECCV, pages 55-72. Springer, 2024. 2 \ No newline at end of file diff --git a/2025/SpectralAR_ Spectral Autoregressive Visual Generation/images.zip b/2025/SpectralAR_ Spectral Autoregressive Visual Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6acca91df23a017ccfe7ef844392f472ddcfe3d9 --- /dev/null +++ b/2025/SpectralAR_ Spectral Autoregressive Visual Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6851013abbeea7c77d8bc7dca4518784eff5c266f334c434369a0c0a61d36f1c +size 1029648 diff --git a/2025/SpectralAR_ Spectral Autoregressive Visual Generation/layout.json b/2025/SpectralAR_ Spectral Autoregressive Visual Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e7cb685a8d27f2db6c1fd964b8327d1eaac971c4 --- /dev/null +++ b/2025/SpectralAR_ Spectral Autoregressive Visual Generation/layout.json @@ -0,0 +1,9446 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 75, + 477, + 92 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 75, + 477, + 92 + ], + "spans": [ + { + "bbox": [ + 133, + 75, + 477, + 92 + ], + "type": "text", + "content": "SpectralAR: Spectral Autoregressive Visual Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 173, + 114, + 438, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 114, + 438, + 159 + ], + "spans": [ + { + "bbox": [ + 173, + 114, + 438, + 159 + ], + "type": "text", + "content": "Yuanhui Huang Weiliang Chen Wenzhao Zheng Yueqi Duan Jie Zhou Jiwen Lu Tsinghua University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 93, + 159, + 514, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 159, + 514, + 172 + ], + "spans": [ + { + "bbox": [ + 93, + 159, + 514, + 172 + ], + "type": "text", + "content": "huangyh22@mails.tsinghua.edu.cn; wenzhao.zheng@outlook.com" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 57, + 180, + 555, + 490 + ], + "blocks": [ + { + "bbox": [ + 57, + 180, + 555, + 490 + ], + "lines": [ + { + "bbox": [ + 57, + 180, + 555, + 490 + ], + "spans": [ + { + "bbox": [ + 57, + 180, + 555, + 490 + ], + "type": "image", + "image_path": "62697571be1fe7f68503368ee738c8ebad782066ae33a320043d4f33785b2b06.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 491, + 555, + 524 + ], + "lines": [ + { + "bbox": [ + 55, + 491, + 555, + 524 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 555, + 524 + ], + "type": "text", + "content": "Figure 1. We approach autoregressive visual generation from the spectral perspective and propose SpectralAR which converts images into 1D causal sequences with nested spectral tokenization and generates images in a hierarchical coarse-to-fine manner. In the autoregressive process, each generated token improves the quality of the image by introducing new high-frequency components." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 533, + 201, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 533, + 201, + 547 + ], + "spans": [ + { + "bbox": [ + 151, + 533, + 201, + 547 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 559, + 297, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 559, + 297, + 693 + ], + "spans": [ + { + "bbox": [ + 54, + 559, + 297, + 693 + ], + "type": "text", + "content": "Autoregressive visual generation has garnered increasing attention due to its scalability and compatibility with other modalities compared with diffusion models. Most existing methods construct visual sequences as spatial patches for autoregressive generation. However, image patches are inherently parallel, contradicting the causal nature of autoregressive modeling. To address this, we propose a Spectral AutoRegressive (SpectralAR) visual generation framework, which realizes causality for visual sequences from the spectral perspective. Specifically, we first transform an image into ordered spectral tokens with Nested Spectral Tokeniza" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 535, + 556, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 535, + 556, + 655 + ], + "spans": [ + { + "bbox": [ + 313, + 535, + 556, + 655 + ], + "type": "text", + "content": "tion, representing lower to higher frequency components. We then perform autoregressive generation in a coarse-to-fine manner with the sequences of spectral tokens. By considering different levels of detail in images, our SpectralAR achieves both sequence causality and token efficiency without bells and whistles. We conduct extensive experiments on ImageNet-1K for image reconstruction and autoregressive generation, and SpectralAR achieves 3.02 gFID with only 64 tokens and 310M parameters. Project page: https://huang-yh.github.io/spectralar/." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 668, + 395, + 680 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 668, + 395, + 680 + ], + "spans": [ + { + "bbox": [ + 314, + 668, + 395, + 680 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "Diffusion models [19, 39, 44, 46] have long been the best performing approach to visual generation. Despite their ex" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 702, + 155, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 702, + 155, + 713 + ], + "spans": [ + { + "bbox": [ + 69, + 702, + 155, + 713 + ], + "type": "inline_equation", + "content": "\\boxtimes" + }, + { + "bbox": [ + 69, + 702, + 155, + 713 + ], + "type": "text", + "content": " Corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15842" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": "ceptional generation quality, diffusion models still exhibit deficiencies in multimodal modeling and integration of perception and generation. The advent of autoregressive visual generation methods [10, 29, 36, 42, 48, 50, 65] alleviates these limitations and enables better scalability with the next-token prediction paradigm. It first utilizes a visual tokenizer to convert images into tokens and then generates samples in a sequential manner. This advancement supports a variety of emerging applications, including scalable visual generation [16, 57, 59, 69], mixed-modal foundation models [49, 56, 60], and autoregressive world models [21, 73]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 207, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 207, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 296, + 567 + ], + "type": "text", + "content": "Despite their dominance in language modeling [40, 53], the performance of autoregressive models in visual generation [10, 14] is still inferior to that of diffusion models [35, 39] and non-autoregressive models [4, 66]. This distinction can be attributed to the inherent difference between text and image modalities. Text, invented for human communication, is discrete and sequential, while image data is continuous and invariant to translation [28], indicating the intrinsic equality among image pixels. This equality makes it a critical issue for autoregressive visual generation to convert images into one-dimensional sequential tokens [65]. Methods based on spatial scanning [10, 34, 47, 62] attempt to discretize the image locally into tokens according to patches, and then perform autoregressive generation following a certain order based on their locations. However, the resulting spatial sequence violates the equality among image patches, making it suboptimal for causal autoregressive modeling. Another line of work [4, 31, 66] introduces bidirectional interaction in the generator as a workaround. Nonetheless, they still tokenize images spatially, thus assuming a causal order among image patches. In addition, the bidirectional design may deviate from the conventional autoregressive paradigm, complicating their integration into omni-modal frameworks [49, 56]. In contrast to spatial tokenization, VAR [50] explores transforming the image into multiple scales and producing a sequence by concatenating tokens from ordered scales. Although scale-wise autoregressive generation indeed satisfies the equality of image pixels, it suffers from inferior token efficiency and parallel generation of multiple tokens from the same scale." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "In this paper, we introduce a spectral autoregressive visual generation framework to achieve causal autoregressive modeling and improve token efficiency, as shown in Figure 2. Frequency is an inherent attribute of all types of signals and has become a significant perspective and methodology that complements the spatial-temporal domain [2]. For visual data, the spectral density often conforms to the power-law distribution [6], with low-frequency components representing the overall structure of an image and high-frequency components focusing on the intricate details. This hierarchical coarse-to-fine nature of the correspondence between spectral and spatial domains indicates a se" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 313, + 80, + 553, + 336 + ], + "blocks": [ + { + "bbox": [ + 479, + 70, + 541, + 79 + ], + "lines": [ + { + "bbox": [ + 479, + 70, + 541, + 79 + ], + "spans": [ + { + "bbox": [ + 479, + 70, + 541, + 79 + ], + "type": "text", + "content": "Token Efficiency" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 80, + 553, + 336 + ], + "lines": [ + { + "bbox": [ + 313, + 80, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 313, + 80, + 553, + 336 + ], + "type": "image", + "image_path": "158aed395cad4765210f69b7ac64517cfebfe9a6db00f599a8d6d0c04516dffa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 337, + 555, + 370 + ], + "lines": [ + { + "bbox": [ + 313, + 337, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 555, + 370 + ], + "type": "text", + "content": "Figure 2. Comparison between autoregressive visual generation methods. SpectralAR achieves both token efficiency and sequence causality with nested 1D spectral tokens." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 372, + 556, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 556, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 556, + 622 + ], + "type": "text", + "content": "quential order of images, motivating us to represent images as causal spectral sequences. Specifically, we first transform images into spectral tokens with Nested Spectral Tokenization (NST), which uses varying sequence lengths to represent images across different frequency bands. The causality of the spectral sequences originates from the coarse-to-fine progression characteristic resembling human visual perception and is strengthened with the application of a causal mask. In addition, inspired by the image compression literature [55, 58], we design a non-uniform token-frequency mapping, which allocates more tokens to represent low-frequency components and fewer for high-frequency components. This mapping technique greatly reduces the number of tokens, while maintaining the quality of reconstructed samples. In the autoregressive process, we begin with the token representing the DC component and progressively predict tokens corresponding to higher frequencies conditioned on previous ones. We conduct extensive experiments on the ImageNet-1K dataset for image reconstruction and class-conditional generation. Our SpectralAR demonstrates comparable performance with state-of-the-art methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 633, + 402, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 633, + 402, + 645 + ], + "spans": [ + { + "bbox": [ + 314, + 633, + 402, + 645 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": "Autoregressive visual generation. While diffusion models have achieved exceptional performance, they are fundamentally different from the conventional autoregressive framework [49, 56] for cross-modal and cross-task modeling. Therefore, a body of research aims to advance autore-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15843" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 286 + ], + "type": "text", + "content": "gressive models for visual generation. Early efforts [5, 51] perform pixel-level generation in the row-major raster-scan order, followed by VQGAN [10] which transfers to the latent feature space of VQVAE [52] for autoregressive modeling. Subsequent work improves based on VQGAN with multiple scales [42], residual quantization [29], ViT architecture [63] or textual conditions [64]. However, raster-scan generation violates the equality between image pixels as discussed in Section 1, contradicting the causality premise of autoregressive modeling. Recently, VAR [50] proposes the scale-wise autoregressive generation that aims to predict the next-scale token map conditioned on the previous ones. Despite achieving causality with the multi-scale design, VAR predicts multiple tokens with bidirectional attention in each step, thus deviating from the standard autoregressive framework. In contrast, our method realizes causal autoregressive generation from the spectral perspective and still conforms to the unidirectional scheme." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 292, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 292, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 292, + 296, + 567 + ], + "type": "text", + "content": "Efficient image tokenization. Autoregressive generation requires the conversion of images into token sequences, which is often achieved using autoencoders [18, 54]. Patch-based autoencoders [10, 24, 37, 52, 72] tokenize images spatially, where each token corresponds to a certain patch from the original image. Although this paradigm performs well in image reconstruction and diffusion-based generation [39, 44], it is not suitable for autoregressive modeling due to its spatial design. Also, its token length is proportional to the square of image resolution, which might become the bottleneck in multimodal modeling given limited context length [49]. TiTok [66] proposes a 1D tokenizer which reduces the number of tokens to 32 for " + }, + { + "bbox": [ + 56, + 292, + 296, + 567 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 56, + 292, + 296, + 567 + ], + "type": "text", + "content": " images. However, TiTok is trained with only an overall reconstruction objective, and thus the precise meaning of these 1D tokens remains unclear. VAR [50] introduces a multi-scale tokenizer for causal autoregressive image generation. Nonetheless, the multi-scale strategy requires an even greater number of tokens compared to patch-based tokenization methods, further diminishing token efficiency. Our method converts images into 1D causal sequences of spectral tokens and enhances token efficiency by leveraging the long-tail distribution of spectral density in image data." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": "Spectral visual analysis. Spectral analysis [2] has been a common technique in computer vision, complementing the spatial and temporal domains. Representative applications include image enhancement and denoising [15, 68], texture analysis and feature extraction [41], compression and super-resolution [13, 55], visual generation [61, 67], adversarial attacks and defenses [12, 32, 33]. For autoregressive image generation, CART [43] and SIT [11] propose to transform an image into multiple causal sets of tokens with base-detail decomposition and discrete wavelet transform, respectively. However, these methods still adhere to the multi-scale 2D tokenization paradigm similar to" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "VAR, resulting in suboptimal token efficiency and bidirectional attention to predict multiple tokens per autoregressive step. In contrast, SpectralAR leverages the discrete cosine transform to capture the global information of an image, and compress it into a 1D sequence with high efficiency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 142, + 432, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 142, + 432, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 142, + 432, + 156 + ], + "type": "text", + "content": "3. Proposed Approach" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 162, + 544, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 162, + 544, + 175 + ], + "spans": [ + { + "bbox": [ + 313, + 162, + 544, + 175 + ], + "type": "text", + "content": "3.1. Revisiting Images from the Spectral Domain" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 179, + 555, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 179, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 179, + 555, + 323 + ], + "type": "text", + "content": "Discrete cosine transform. Spectral analysis investigates how complex signals can be represented with simpler basis functions [2], producing a spectral density distribution that represents the magnitude of corresponding basic components. This spectral density distribution is an equivalent representation of the original signal and provides a distinct perspective from the spatial domain depending on the properties of the basis functions. We employ the Discrete Cosine Transform (DCT) [1] to convert images into the spectral domain. The DCT result " + }, + { + "bbox": [ + 313, + 179, + 555, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 313, + 179, + 555, + 323 + ], + "type": "text", + "content": " shares the same shape with the transformed image " + }, + { + "bbox": [ + 313, + 179, + 555, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 313, + 179, + 555, + 323 + ], + "type": "text", + "content": " (for simplicity, we omit the channel dimension):" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 341, + 329, + 555, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 329, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 341, + 329, + 555, + 346 + ], + "type": "interline_equation", + "content": "\\mathbf {D} = \\left\\{F (u, v) \\right\\} _ {u, v = 1} ^ {W, H}, \\quad \\mathbf {I} = \\left\\{f (x, y) \\right\\} _ {x, y = 1} ^ {W, H}. \\tag {1}", + "image_path": "7f4e5cfa7701e47268c403cbd6f8e4d9a8fa851722224039a6142f34f2ef2d17.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 353, + 553, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 553, + 376 + ], + "type": "text", + "content": "Each " + }, + { + "bbox": [ + 313, + 353, + 553, + 376 + ], + "type": "inline_equation", + "content": "F(u,v)" + }, + { + "bbox": [ + 313, + 353, + 553, + 376 + ], + "type": "text", + "content": " represents the intensity of the corresponding basis function " + }, + { + "bbox": [ + 313, + 353, + 553, + 376 + ], + "type": "inline_equation", + "content": "g_{u,v}(x,y)" + }, + { + "bbox": [ + 313, + 353, + 553, + 376 + ], + "type": "text", + "content": " in the image " + }, + { + "bbox": [ + 313, + 353, + 553, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 313, + 353, + 553, + 376 + ], + "type": "text", + "content": ", which writes:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 318, + 382, + 555, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 382, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 318, + 382, + 555, + 407 + ], + "type": "interline_equation", + "content": "g _ {u, v} (x, y) = \\frac {2 C (u) C (v)}{\\sqrt {H W}} \\cos \\frac {(2 x + 1) u \\pi}{2 W} \\cos \\frac {(2 y + 1) v \\pi}{2 H}, \\tag {2}", + "image_path": "a6eb76a502f6b33fbf38afa5b6ed520692c5c6786aefad8de457aff8381586ee.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 413, + 554, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 554, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 554, + 545 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 413, + 554, + 545 + ], + "type": "inline_equation", + "content": "C(u) = 1 / \\sqrt{2}" + }, + { + "bbox": [ + 313, + 413, + 554, + 545 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 313, + 413, + 554, + 545 + ], + "type": "inline_equation", + "content": "u = 0" + }, + { + "bbox": [ + 313, + 413, + 554, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 413, + 554, + 545 + ], + "type": "inline_equation", + "content": "C(u) = 1" + }, + { + "bbox": [ + 313, + 413, + 554, + 545 + ], + "type": "text", + "content": " otherwise. This family of basis functions has the following properties: (1) Given " + }, + { + "bbox": [ + 313, + 413, + 554, + 545 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 313, + 413, + 554, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 413, + 554, + 545 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 413, + 554, + 545 + ], + "type": "text", + "content": ", the basis function " + }, + { + "bbox": [ + 313, + 413, + 554, + 545 + ], + "type": "inline_equation", + "content": "g_{u,v}(x,y)" + }, + { + "bbox": [ + 313, + 413, + 554, + 545 + ], + "type": "text", + "content": " exhibits a checkerboard-like pattern in the spatial domain, with periods along the x- and y-axis of " + }, + { + "bbox": [ + 313, + 413, + 554, + 545 + ], + "type": "inline_equation", + "content": "2W / u" + }, + { + "bbox": [ + 313, + 413, + 554, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 413, + 554, + 545 + ], + "type": "inline_equation", + "content": "2H / v" + }, + { + "bbox": [ + 313, + 413, + 554, + 545 + ], + "type": "text", + "content": ", respectively. This pattern suggests that the basis function characterizes the rate of variation of images in the spatial domain. (2) The " + }, + { + "bbox": [ + 313, + 413, + 554, + 545 + ], + "type": "inline_equation", + "content": "F(u,v)" + }, + { + "bbox": [ + 313, + 413, + 554, + 545 + ], + "type": "text", + "content": "s also form a 2D matrix together, where the top-left corner represents the low-frequency components (small " + }, + { + "bbox": [ + 313, + 413, + 554, + 545 + ], + "type": "inline_equation", + "content": "u, v" + }, + { + "bbox": [ + 313, + 413, + 554, + 545 + ], + "type": "text", + "content": "), while the bottom-right corner corresponds to the high-frequency components (large " + }, + { + "bbox": [ + 313, + 413, + 554, + 545 + ], + "type": "inline_equation", + "content": "u, v" + }, + { + "bbox": [ + 313, + 413, + 554, + 545 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 546, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 554, + 665 + ], + "type": "text", + "content": "Causality from the spectral domain. Since low- and high-frequency components describe overall structures and intricate details, respectively, we can decompose an image into a sequence of sub-images " + }, + { + "bbox": [ + 313, + 546, + 554, + 665 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_i^{\\prime}\\}_{i = 1}^{L}" + }, + { + "bbox": [ + 313, + 546, + 554, + 665 + ], + "type": "text", + "content": " with increasing levels of detail by applying inverse-DCT on partially masked spectral density distributions " + }, + { + "bbox": [ + 313, + 546, + 554, + 665 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{D}_i^{\\prime}\\}_{i = 1}^{L}" + }, + { + "bbox": [ + 313, + 546, + 554, + 665 + ], + "type": "text", + "content": ". With more high-frequency components, the sub-images will gradually transition from blurred to sharp, as shown in Figure 2. This hierarchical coarse-to-fine sequence aligns with human visual perception and artistic drawing, thus enhancing causality." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "text", + "content": "Efficiency from the spectral domain. As the image compression literature [1] pointed out, the DCT result " + }, + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "text", + "content": " of images conforms to the power-law distribution with the absolute values of " + }, + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "inline_equation", + "content": "F(u,v)\\mathrm{s}" + }, + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "text", + "content": " on the top-left corner substantially" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15844" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 70, + 541, + 320 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 541, + 320 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 541, + 320 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 541, + 320 + ], + "type": "image", + "image_path": "2b39acda8d76dcefead60a529f616d9fbe05af9a72d331a7794a11cc7fff2558.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 321, + 555, + 366 + ], + "lines": [ + { + "bbox": [ + 54, + 321, + 555, + 366 + ], + "spans": [ + { + "bbox": [ + 54, + 321, + 555, + 366 + ], + "type": "text", + "content": "Figure 3. Overall pipeline of the proposed SpectralAR. Left: We convert an input image into a 1D causal sequence efficiently with nested spectral tokenization. Each nested sequence is supervised with distinct reconstruction target in a coarse-to-fine manner, which endows each token with an explicit spectral interpretation. We also apply the causal mask to the spectral tokens in the encoder and decoder to enhance the causality. Right: We use the groundtruth sequences from the tokenization process to train an autoregressive generation model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 369, + 295, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 369, + 295, + 478 + ], + "spans": [ + { + "bbox": [ + 54, + 369, + 295, + 478 + ], + "type": "text", + "content": "larger than the bottom-right corner, indicating that most energy of an image is concentrated in the low-frequency components. In addition, human visual perception is less sensitive to high-frequency signals, which have little influence on the visual quality of images. Based on these distinctions, we can encode the high-frequency components of an image with coarser granularity to improve token efficiency, similar to the JPEG algorithm [55] which saves more than " + }, + { + "bbox": [ + 54, + 369, + 295, + 478 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 54, + 369, + 295, + 478 + ], + "type": "text", + "content": " storage for images by suppressing high-frequency signals." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 491, + 216, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 216, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 216, + 503 + ], + "type": "text", + "content": "3.2. Nested Spectral Tokenization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": "Overall framework. In contrast to the 2D spatial tokenization that captures the local correlation between image patches, the basis function " + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "inline_equation", + "content": "g_{u,v}(x,y)" + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": " of DCT encodes a global frequency pattern. Therefore, we convert images into 1D tokens in spectral tokenization to reflect the global nature of the basis functions. We start with the general framework of a 1D image tokenizer [66]. Given an image " + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": ", we aim to encode it into " + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": " discrete vectors " + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": ", and also reconstruct the original image with " + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": ". We first patchify the image into " + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{P} \\in \\mathbb{R}^{hw\\times C}" + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": ", and concatenate the image features with the initial query vectors " + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_0" + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": " to form " + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "inline_equation", + "content": "[\\mathbf{P};\\mathbf{S}_0] \\in \\mathbb{R}^{(hw + N)\\times C}" + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "inline_equation", + "content": "h,w,C" + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": " denote the resolution of image features and the channel dimension, respectively. We then employ the vision transformer " + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": " to enable feature extraction and interaction between the image features and 1D query tokens, resulting in the informative 1D representation " + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{S}}" + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": ". In the vector quantizer " + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": ", we match these continuous vectors with the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 369, + 555, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 440 + ], + "type": "text", + "content": "codebook embeddings to derive the discrete representation " + }, + { + "bbox": [ + 313, + 369, + 555, + 440 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 313, + 369, + 555, + 440 + ], + "type": "text", + "content": ", which could serve as the groundtruth for the autoregressive training. At last, we append " + }, + { + "bbox": [ + 313, + 369, + 555, + 440 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 313, + 369, + 555, + 440 + ], + "type": "text", + "content": " to a set of mask tokens " + }, + { + "bbox": [ + 313, + 369, + 555, + 440 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\in \\mathbb{R}^{hw \\times C}" + }, + { + "bbox": [ + 313, + 369, + 555, + 440 + ], + "type": "text", + "content": " and process them with the decoder network " + }, + { + "bbox": [ + 313, + 369, + 555, + 440 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 369, + 555, + 440 + ], + "type": "text", + "content": " similar to " + }, + { + "bbox": [ + 313, + 369, + 555, + 440 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 369, + 555, + 440 + ], + "type": "text", + "content": ", in order to reconstruct the original image. This overall framework could be formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 328, + 450, + 555, + 465 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 450, + 555, + 465 + ], + "spans": [ + { + "bbox": [ + 328, + 450, + 555, + 465 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {S}} = \\mathcal {E} ([ \\mathbf {P}; \\mathbf {S} _ {0} ]), \\quad \\mathbf {S} = \\mathcal {Q} (\\hat {\\mathbf {S}}), \\quad \\hat {\\mathbf {I}} = \\mathcal {D} ([ \\mathbf {M}; \\mathbf {S} ]), \\quad (3)", + "image_path": "7dd20992559c6894bb3e60b830b9451e68f764ffeba29d73edd1c82a34ca8bf7.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 475, + 555, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 555, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 555, + 500 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 475, + 555, + 500 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}" + }, + { + "bbox": [ + 313, + 475, + 555, + 500 + ], + "type": "text", + "content": " denotes the reconstructed image. The training objective typically consists of multiple loss functions:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 318, + 509, + 555, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 509, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 318, + 509, + 555, + 525 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t o k} = \\| \\hat {\\mathbf {I}} - \\mathbf {I} \\| _ {2} ^ {2} + \\| \\hat {\\mathbf {S}} - \\mathbf {S} \\| _ {2} ^ {2} + \\mathcal {L} _ {P} (\\hat {\\mathbf {I}}, \\mathbf {I}) + \\mathcal {L} _ {A} (\\hat {\\mathbf {I}}, \\mathbf {I}), (4)", + "image_path": "989e96e228608379651d402118e4edf7dcfa205f8bda1caae505267b22c88be9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 533, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 554, + 557 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 533, + 554, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_P" + }, + { + "bbox": [ + 313, + 533, + 554, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 533, + 554, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_A" + }, + { + "bbox": [ + 313, + 533, + 554, + 557 + ], + "type": "text", + "content": " denote the perceptual loss [22, 70] and the adversarial loss [10, 14], respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "Nested spectral decoding. Different from the plain 1D tokenization in Figure 2, we aim to represent an image as its spectral decompositions " + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_i^{\\prime}\\}_{i = 1}^{L}" + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": " for sequence causality, which requires establishing a mapping from 1D tokens to these sub-images. One naive way to achieve this would be dividing S into disjoint subsets and assigning them to model different sub-images independently. Similar to the multi-scale tokenization [50], this strategy would inevitably involve bidirectional interaction and diminish token efficiency because sub-images with finer detail would require increasingly more tokens to represent. In contrast, we propose a nested mapping scheme for efficient tokenization, as shown in the bottom of Figure 3. We first construct a" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15845" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 76, + 263, + 87 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 76, + 263, + 87 + ], + "spans": [ + { + "bbox": [ + 61, + 76, + 263, + 87 + ], + "type": "text", + "content": "Algorithm 1: Nested Spectral Tokenization Training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 89, + 280, + 183 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 59, + 89, + 280, + 101 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 89, + 280, + 101 + ], + "spans": [ + { + "bbox": [ + 59, + 89, + 280, + 101 + ], + "type": "text", + "content": "1 Inputs: raw image " + }, + { + "bbox": [ + 59, + 89, + 280, + 101 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 59, + 89, + 280, + 101 + ], + "type": "text", + "content": ", initial spectral tokens " + }, + { + "bbox": [ + 59, + 89, + 280, + 101 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_0\\in \\mathbb{R}^{N\\times C}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 101, + 267, + 113 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 101, + 267, + 113 + ], + "spans": [ + { + "bbox": [ + 59, + 101, + 267, + 113 + ], + "type": "text", + "content": "2 Hyperparameters: spectral levels " + }, + { + "bbox": [ + 59, + 101, + 267, + 113 + ], + "type": "inline_equation", + "content": "\\{\\omega_i|i = 1,\\dots,N\\}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 60, + 114, + 238, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 114, + 238, + 125 + ], + "spans": [ + { + "bbox": [ + 60, + 114, + 238, + 125 + ], + "type": "text", + "content": "3 " + }, + { + "bbox": [ + 60, + 114, + 238, + 125 + ], + "type": "inline_equation", + "content": "\\mathbf{P} =" + }, + { + "bbox": [ + 60, + 114, + 238, + 125 + ], + "type": "text", + "content": " patchify(I), " + }, + { + "bbox": [ + 60, + 114, + 238, + 125 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{S}} = \\mathcal{E}([\\mathbf{P};\\mathbf{S}_0])" + }, + { + "bbox": [ + 60, + 114, + 238, + 125 + ], + "type": "inline_equation", + "content": "\\mathbf{S} = \\mathcal{Q}(\\hat{\\mathbf{S}})" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 60, + 125, + 168, + 136 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 125, + 168, + 136 + ], + "spans": [ + { + "bbox": [ + 60, + 125, + 168, + 136 + ], + "type": "text", + "content": "4 " + }, + { + "bbox": [ + 60, + 125, + 168, + 136 + ], + "type": "inline_equation", + "content": "idx = \\mathrm{random\\_choice}(N)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 60, + 137, + 189, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 137, + 189, + 148 + ], + "spans": [ + { + "bbox": [ + 60, + 137, + 189, + 148 + ], + "type": "text", + "content": "5 " + }, + { + "bbox": [ + 60, + 137, + 189, + 148 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^{\\prime} = \\mathbf{S}[:idx],\\hat{\\mathbf{I}}^{\\prime} = \\mathcal{D}([\\mathbf{M};\\mathbf{S}^{\\prime}])" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 60, + 148, + 249, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 148, + 249, + 159 + ], + "spans": [ + { + "bbox": [ + 60, + 148, + 249, + 159 + ], + "type": "inline_equation", + "content": "\\mathbf{6}" + }, + { + "bbox": [ + 60, + 148, + 249, + 159 + ], + "type": "inline_equation", + "content": "\\mathbf{D} = \\mathrm{DCT}(\\mathbf{I}),\\mathbf{D}' = \\mathbf{D}\\circ \\mathbf{1}_{\\omega_i},\\mathbf{I}' = \\mathrm{DCT}^{-1}(\\mathbf{D}')" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 60, + 160, + 143, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 160, + 143, + 172 + ], + "spans": [ + { + "bbox": [ + 60, + 160, + 143, + 172 + ], + "type": "text", + "content": "7 loss " + }, + { + "bbox": [ + 60, + 160, + 143, + 172 + ], + "type": "inline_equation", + "content": "= \\mathcal{L}_{tok}(\\hat{\\mathbf{l}}',\\mathbf{l}')" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 60, + 173, + 185, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 173, + 185, + 183 + ], + "spans": [ + { + "bbox": [ + 60, + 173, + 185, + 183 + ], + "type": "text", + "content": "8 Return: loss for optimization;" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 193, + 295, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 193, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 295, + 217 + ], + "type": "text", + "content": "sequence of sub-images with increasing detail by progressively preserving larger regions in the spectral density " + }, + { + "bbox": [ + 55, + 193, + 295, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 55, + 193, + 295, + 217 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 64, + 224, + 294, + 238 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 224, + 294, + 238 + ], + "spans": [ + { + "bbox": [ + 64, + 224, + 294, + 238 + ], + "type": "interline_equation", + "content": "\\mathbf {I} _ {i} ^ {\\prime} = \\mathrm {D C T} ^ {- 1} \\left(\\mathbf {D} _ {i} ^ {\\prime}\\right), \\quad \\mathbf {D} _ {i} ^ {\\prime} = \\mathbf {D} \\circ \\mathbf {1} _ {\\omega_ {i}}, \\quad \\omega_ {i - 1} < \\omega_ {i}, \\tag {5}", + "image_path": "1027417ee43add138830d0520d2712261dcefce725d5576b7f3f602c10f8e75a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 245, + 296, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 245, + 296, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 296, + 366 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 245, + 296, + 366 + ], + "type": "inline_equation", + "content": "\\mathrm{DCT}^{-1}" + }, + { + "bbox": [ + 55, + 245, + 296, + 366 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 245, + 296, + 366 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 55, + 245, + 296, + 366 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 245, + 296, + 366 + ], + "type": "inline_equation", + "content": "\\mathbf{1}_{\\omega_i}" + }, + { + "bbox": [ + 55, + 245, + 296, + 366 + ], + "type": "text", + "content": " denote the inverse DCT operation, element-wise multiplication and a " + }, + { + "bbox": [ + 55, + 245, + 296, + 366 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 55, + 245, + 296, + 366 + ], + "type": "text", + "content": " matrix with the top-left corner of size " + }, + { + "bbox": [ + 55, + 245, + 296, + 366 + ], + "type": "inline_equation", + "content": "\\omega_{i}\\times \\omega_{i}" + }, + { + "bbox": [ + 55, + 245, + 296, + 366 + ], + "type": "text", + "content": " filled by ones and the remaining parts being zeros. Therefore, the sub-image " + }, + { + "bbox": [ + 55, + 245, + 296, + 366 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i^{\\prime}" + }, + { + "bbox": [ + 55, + 245, + 296, + 366 + ], + "type": "text", + "content": " contains all frequency components present in the sub-image " + }, + { + "bbox": [ + 55, + 245, + 296, + 366 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{i - 1}^{\\prime}" + }, + { + "bbox": [ + 55, + 245, + 296, + 366 + ], + "type": "text", + "content": ". Based on this inclusion property, we can reuse the tokens representing the previous sub-image " + }, + { + "bbox": [ + 55, + 245, + 296, + 366 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{i - 1}^{\\prime}" + }, + { + "bbox": [ + 55, + 245, + 296, + 366 + ], + "type": "text", + "content": " to represent the next sub-image " + }, + { + "bbox": [ + 55, + 245, + 296, + 366 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i^\\prime" + }, + { + "bbox": [ + 55, + 245, + 296, + 366 + ], + "type": "text", + "content": ". To avoid bidirectional attention, we make " + }, + { + "bbox": [ + 55, + 245, + 296, + 366 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 245, + 296, + 366 + ], + "type": "text", + "content": " equal to " + }, + { + "bbox": [ + 55, + 245, + 296, + 366 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 245, + 296, + 366 + ], + "type": "text", + "content": " so that each token " + }, + { + "bbox": [ + 55, + 245, + 296, + 366 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i" + }, + { + "bbox": [ + 55, + 245, + 296, + 366 + ], + "type": "text", + "content": " in the sequence S corresponds to a unique sub-image " + }, + { + "bbox": [ + 55, + 245, + 296, + 366 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i^\\prime" + }, + { + "bbox": [ + 55, + 245, + 296, + 366 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 119, + 374, + 294, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 374, + 294, + 388 + ], + "spans": [ + { + "bbox": [ + 119, + 374, + 294, + 388 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {I}} _ {i} ^ {\\prime} = \\mathcal {D} ([ \\mathbf {M}; \\mathbf {s} _ {1}, \\mathbf {s} _ {2}, \\dots , \\mathbf {s} _ {i} ]), \\tag {6}", + "image_path": "3a06ff5bca9d99e6dfd0f0bbbe2f372b8aed71e2d74458c2a0eb3a62b9eb70c1.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 396, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 295, + 456 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 396, + 295, + 456 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_i^{\\prime}" + }, + { + "bbox": [ + 55, + 396, + 295, + 456 + ], + "type": "text", + "content": " is the reconstruction of sub-image " + }, + { + "bbox": [ + 55, + 396, + 295, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i^\\prime" + }, + { + "bbox": [ + 55, + 396, + 295, + 456 + ], + "type": "text", + "content": " given the nested 1D sequence " + }, + { + "bbox": [ + 55, + 396, + 295, + 456 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{s}_1,\\mathbf{s}_2,\\dots ,\\mathbf{s}_i\\}" + }, + { + "bbox": [ + 55, + 396, + 295, + 456 + ], + "type": "text", + "content": ". Nested spectral decoding compresses an image into a causal 1D sequence where each token " + }, + { + "bbox": [ + 55, + 396, + 295, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i" + }, + { + "bbox": [ + 55, + 396, + 295, + 456 + ], + "type": "text", + "content": " corresponds to a disjoint set of frequencies and achieves token efficiency by reusing previous tokens." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 457, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 457, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 296, + 552 + ], + "type": "text", + "content": "Non-uniform token-frequency mapping. To further enhance token efficiency, we introduce the non-uniform token-frequency mapping technique. Since high-frequency components have low magnitude and minimal impact on the visual quality of images, we can encode them with coarser granularity compared to the low-frequency counterparts. We achieve this by reducing the interval between " + }, + { + "bbox": [ + 55, + 457, + 296, + 552 + ], + "type": "inline_equation", + "content": "\\omega_{i-1}" + }, + { + "bbox": [ + 55, + 457, + 296, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 457, + 296, + 552 + ], + "type": "inline_equation", + "content": "\\omega_i" + }, + { + "bbox": [ + 55, + 457, + 296, + 552 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 55, + 457, + 296, + 552 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 457, + 296, + 552 + ], + "type": "text", + "content": " is small, and increase it otherwise:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 123, + 562, + 294, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 562, + 294, + 574 + ], + "spans": [ + { + "bbox": [ + 123, + 562, + 294, + 574 + ], + "type": "interline_equation", + "content": "\\omega_ {i} - \\omega_ {i - 1} \\leq \\omega_ {i + 1} - \\omega_ {i}, \\tag {7}", + "image_path": "19fca3d0a8cb838d28f25c6a1b646ff3cccd555e9d67b062e299771665c6b103.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 582, + 295, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 630 + ], + "type": "text", + "content": "which demonstrates the general case. This non-uniform mapping allocates later tokens to broader frequency ranges, enabling precise modeling of crucial low-frequency components while efficiently representing high-frequency details." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 630, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 712 + ], + "type": "text", + "content": "Spectral causal mask. Although the sequence " + }, + { + "bbox": [ + 55, + 630, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 55, + 630, + 296, + 712 + ], + "type": "text", + "content": " supervised with (5)(6) already exhibits a certain degree of causality, the encoding and decoding processes, i.e. " + }, + { + "bbox": [ + 55, + 630, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 630, + 296, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 630, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 630, + 296, + 712 + ], + "type": "text", + "content": ", are still bidirectional, which can lead to information leakage from high-frequency to low-frequency components. Therefore, we propose applying causal masks to the spectral tokens " + }, + { + "bbox": [ + 55, + 630, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 55, + 630, + 296, + 712 + ], + "type": "text", + "content": " in both the encoder " + }, + { + "bbox": [ + 55, + 630, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 630, + 296, + 712 + ], + "type": "text", + "content": " and the decoder " + }, + { + "bbox": [ + 55, + 630, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 630, + 296, + 712 + ], + "type": "text", + "content": ", as shown" + } + ] + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 317, + 115, + 553, + 186 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 114 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 114 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 114 + ], + "type": "text", + "content": "Table 1. Correlation between tokens of different autoregressive paradigms. We use linear correlation as a proxy metric for the causality of sequences. The spectral sequence demonstrates better causality compared with other methods." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 115, + 553, + 186 + ], + "lines": [ + { + "bbox": [ + 317, + 115, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 317, + 115, + 553, + 186 + ], + "type": "table", + "html": "
Correlation TypeRaster-scanScale-wiseSpectral
R2avg(t2; t1)0.4710.8890.916
R2avg(t3; t1, t2)0.3660.9530.977
R2avg(t4; t1, t2, t3)0.5250.9430.994
Average0.4540.9280.962
", + "image_path": "ff0613b11d160fce07f9b3d309ffbb601833ab23ce38e9895fd698cb9c338d3b.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 315, + 190, + 391, + 266 + ], + "blocks": [ + { + "bbox": [ + 315, + 190, + 391, + 266 + ], + "lines": [ + { + "bbox": [ + 315, + 190, + 391, + 266 + ], + "spans": [ + { + "bbox": [ + 315, + 190, + 391, + 266 + ], + "type": "image", + "image_path": "6f0262dc63c5c27aadfc558f7b636ba1aa0a6311fa06b784b9805f52a5c08d96.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 264, + 370, + 270 + ], + "lines": [ + { + "bbox": [ + 335, + 264, + 370, + 270 + ], + "spans": [ + { + "bbox": [ + 335, + 264, + 370, + 270 + ], + "type": "text", + "content": "Raster-scan" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 274, + 555, + 317 + ], + "lines": [ + { + "bbox": [ + 313, + 274, + 555, + 317 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 555, + 317 + ], + "type": "text", + "content": "Figure 4. Frequency of each element having the highest correlation with another one. It represents the source of reference information for predicting the next token. The raster-scan method exhibits excessive dependency on boundaries of image patches." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 391, + 190, + 464, + 264 + ], + "blocks": [ + { + "bbox": [ + 391, + 190, + 464, + 264 + ], + "lines": [ + { + "bbox": [ + 391, + 190, + 464, + 264 + ], + "spans": [ + { + "bbox": [ + 391, + 190, + 464, + 264 + ], + "type": "image", + "image_path": "a6881fb4872c5f10b14367c193ba56928c19b9afcc405379f1b6e24877c9be6a.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 411, + 264, + 443, + 270 + ], + "lines": [ + { + "bbox": [ + 411, + 264, + 443, + 270 + ], + "spans": [ + { + "bbox": [ + 411, + 264, + 443, + 270 + ], + "type": "text", + "content": "Scale-wise" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 466, + 190, + 539, + 266 + ], + "blocks": [ + { + "bbox": [ + 466, + 190, + 539, + 266 + ], + "lines": [ + { + "bbox": [ + 466, + 190, + 539, + 266 + ], + "spans": [ + { + "bbox": [ + 466, + 190, + 539, + 266 + ], + "type": "image", + "image_path": "3a959dd944bf178c2326bffa2e8707fe4bb684fa09eb448adcf762c190c01129.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 489, + 264, + 515, + 271 + ], + "lines": [ + { + "bbox": [ + 489, + 264, + 515, + 271 + ], + "spans": [ + { + "bbox": [ + 489, + 264, + 515, + 271 + ], + "type": "text", + "content": "Spectral" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 539, + 190, + 553, + 266 + ], + "blocks": [ + { + "bbox": [ + 539, + 190, + 553, + 266 + ], + "lines": [ + { + "bbox": [ + 539, + 190, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 539, + 190, + 553, + 266 + ], + "type": "image", + "image_path": "16229f95e468c3d09cb5cefecca1954ce71d5bfc784494c544366abb177f5dd5.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 321, + 555, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 321, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 555, + 369 + ], + "type": "text", + "content": "in Figure 3. This spectral causal mask restricts each token s to only attend to tokens that represent lower frequencies, enhancing causality from the architectural perspective. We outline the training procedure in Algorithm 1." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 376, + 503, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 503, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 503, + 389 + ], + "type": "text", + "content": "3.3. Spectral Autoregressive Generation" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 394, + 554, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 554, + 491 + ], + "type": "text", + "content": "Autoregressive modeling has gained prominence in computer vision for its scalability, generalization, and effectiveness across multimodal tasks [25, 49, 56]. While conventional autoregressive generation follows a spatial raster-scan order [10, 51], we propose a hierarchical coarse-to-fine approach in the spectral domain to enhance causality, as shown in the right side of Figure 3. We start with the general framework of autoregressive modeling:" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 380, + 500, + 553, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 500, + 553, + 513 + ], + "spans": [ + { + "bbox": [ + 380, + 500, + 553, + 513 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {i + 1} = \\mathcal {M} \\left(\\mathbf {t} _ {1}, \\mathbf {t} _ {2}, \\dots , \\mathbf {t} _ {i}\\right), \\tag {8}", + "image_path": "dd6dce8bc37e50d1430151e3341f0837f5b1d427e66c8dbebae9481a05ba0968.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{T} = \\{\\mathbf{t}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " is a sequence of quantized tokens, and " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{i+1}" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " denote the probability logits for the " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "(i+1)" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": "th token and the autoregressive model, respectively. The autoregressive process (8) assumes that the generation of token " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{i+1}" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " depends solely on the previous ones. The spatial autoregressive paradigm violates this premise because of the equality of image pixels (as discussed in Section 1). In contrast, we take the spectral tokens " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " from the nested spectral tokenziation as the autoregressive targets " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": ". Since the spectral tokens are trained in a nested manner to reconstruct sub-images of increasing levels of detail as in (5)(6), each spectral token " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " is expected to enhance the quality of the sub-image " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{i-1}'" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " represented by previous tokens from the spectral domain. This progressive refinement process aligns with human visual perception and artistic painting, both of which start with the overall structure and gradually focus on" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15846" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 116, + 541, + 472 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": "Table 2. Comparison between generative models on class-conditional ImageNet " + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "inline_equation", + "content": "{256} \\times {256}" + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": " benchmark. “ " + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": " ” or “↑” indicate lower or higher values are better. \"Token\": the number of tokens used in transformer architectures. \"Step\": the number of model runs needed to generate an image. We compute our wall-clock inference time and scale the Time for other methods accordingly. " + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": " : trained on larger datasets including OpenImages [27]. ‡: implemented with the official tokenizer weight [66] and the scripts from VAR." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 116, + 541, + 472 + ], + "lines": [ + { + "bbox": [ + 69, + 116, + 541, + 472 + ], + "spans": [ + { + "bbox": [ + 69, + 116, + 541, + 472 + ], + "type": "table", + "html": "
TypeModelrFID↓gFID↓IS↑Pre↑Rec↑#Token#Para#StepTime
GANBigGAN [3]75.246.95224.50.890.38112M1-
GANGigaGAN [23]-3.45225.50.840.61N.A.569M1-
GANStyleGan-XL [45]7.062.30265.10.780.53166M10.4
Diff.ADM [8]125.7810.94101.00.690.63N.A.554M250235
Diff.CDM [20]-4.88158.7--N.A.-8100-
Diff.LDM-4-G [44]0.27†3.60247.7--N.A.400M250-
Diff.DiT-L/2 [39]0.62†5.02167.20.750.57256458M25043
Diff.DiT-XL/2 [39]0.62†2.27278.20.830.57256675M25063
Diff.L-DiT-3B [71]-2.10304.40.820.602563.0B250>63
Diff.L-DiT-7B [71]-2.28316.20.830.582567.0B250>63
Mask.MaskGIT [4]2.286.18182.10.800.51256227M80.7
Mask.RCG (cond.) [30]-3.49215.5---502M202.7
Mask.TiTok-B64 [66]1.702.48214.7--64177M80.4
2D ScanVQVAE-2 [42]-31.11~450.360.57N.A.13.5B5120-
2D ScanVQGAN [10]7.9415.7874.3--N.A.1.4B25634
2D ScanViTVQ [63]1.284.17175.1--10241.7B1024>34
2D ScanRQTran. [29]3.207.55134.0--64,43.8B6829
VARVAR-d16 [50]3.30274.40.840.51310M100.6
VARVAR-d20 [50]2.57302.60.830.56600M100.7
VARVAR-d24 [50]2.09312.90.820.591.0B100.8
VARVAR-d30 [50]1.92323.10.820.592.0B101.4
1D ARTiTok-B64-d16‡ [66]1.706.30190.10.850.4764310M641
1D ARSpectralAR-d164.033.02282.20.810.5564310M641
1D ARSpectralAR-d204.032.49305.4--64600M641.2
1D ARSpectralAR-d244.032.13307.7--641.0B641.4
1D ARSpectralAR-d16-p44.033.13276.1--16×4310M160.4
", + "image_path": "5fa04c165f971d3fa870dae4f6302a1698c646a463373c7ba1d33c8d905d0bd6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 477, + 295, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 477, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 477, + 295, + 525 + ], + "type": "text", + "content": "details. This similarity qualitatively validates the rationale for performing causal autoregressive generation in the spectral domain. We further provide some quantitative analysis through a toy experiment in Section 4.2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 525, + 295, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 295, + 645 + ], + "type": "text", + "content": "Potential applications. The frequencies represented by the token " + }, + { + "bbox": [ + 55, + 525, + 295, + 645 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i" + }, + { + "bbox": [ + 55, + 525, + 295, + 645 + ], + "type": "text", + "content": " become higher as " + }, + { + "bbox": [ + 55, + 525, + 295, + 645 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 525, + 295, + 645 + ], + "type": "text", + "content": " increases, while its influence on the image quality diminishes accordingly (check Section 3.1 for details). Therefore, we can control the visual quality of sampled images by adjusting the length of generated sequences, similar to the image compression algorithms [55]. In addition, we can achieve super-resolution by dividing images into disjoint parts smaller than " + }, + { + "bbox": [ + 55, + 525, + 295, + 645 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 55, + 525, + 295, + 645 + ], + "type": "text", + "content": ", and conducting individual spectral autoregressive generation on each part. We provide further results in Section 4.4." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 652, + 137, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 652, + 137, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 652, + 137, + 666 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 672, + 247, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 247, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 247, + 685 + ], + "type": "text", + "content": "4.1. Dataset and Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "We train and evaluate our SpectralAR on the ImageNet1K [7] benchmark, which contains 1,281,167 and 50,000" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 477, + 555, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 477, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 477, + 555, + 548 + ], + "type": "text", + "content": "images for training and validation, respectively. We train the tokenizer and generator on the training split. We evaluate the reconstruction performance on the validation set with reconstruction Fréchet inception distance [17] (rFID), and the generation results with generation FID (gFID) using pre-computed statistics and scripts from ADM [8]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 549, + 555, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 549, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 549, + 555, + 622 + ], + "type": "text", + "content": "For tokenizer training, we follow the exactly same settings of TiTok [66] for a fair comparison. We also employ the two-stage training strategy with proxy codes [4, 66]. We use the ViT-B [9] as the encoder and decoder, and set the number of spectral tokens as " + }, + { + "bbox": [ + 313, + 549, + 555, + 622 + ], + "type": "inline_equation", + "content": "N = 64" + }, + { + "bbox": [ + 313, + 549, + 555, + 622 + ], + "type": "text", + "content": " in our main experiments and the sequence " + }, + { + "bbox": [ + 313, + 549, + 555, + 622 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 549, + 555, + 622 + ], + "type": "text", + "content": "s for " + }, + { + "bbox": [ + 313, + 549, + 555, + 622 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 549, + 555, + 622 + ], + "type": "text", + "content": " images as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 351, + 631, + 555, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 631, + 555, + 669 + ], + "spans": [ + { + "bbox": [ + 351, + 631, + 555, + 669 + ], + "type": "interline_equation", + "content": "\\omega_ {i} = \\left\\{ \\begin{array}{l l} i, & \\text {i f} i \\in (0, 3 2 ], \\\\ 2 i - 3 2, & \\text {i f} i \\in (3 2, 4 8 ], \\\\ 1 2 i - 5 1 2, & \\text {i f} i \\in (4 8, 6 4 ]. \\end{array} \\right. \\tag {9}", + "image_path": "556e99a401f5d0017139150e4e7eccff415607db73e5a182325630571367ebcc.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "For generator training, we adopt the same architecture and training recipe as VAR [50], which leverages a GPT-2-like transformer architecture [40] for autoregressive modeling." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15847" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 553, + 506 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 553, + 506 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 553, + 506 + ], + "type": "image", + "image_path": "730d1fcfc5fc12876dca33d6b4a5f03440c9ed2823c9e73856251edf6e8cba57.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 508, + 555, + 533 + ], + "lines": [ + { + "bbox": [ + 55, + 508, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 555, + 533 + ], + "type": "text", + "content": "Figure 5. Visualizations of generated " + }, + { + "bbox": [ + 55, + 508, + 555, + 533 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 508, + 555, + 533 + ], + "type": "text", + "content": " samples on ImageNet-1K. These samples cover a wide range of categories and styles, demonstrating the ability of SpectralAR to generate both diverse and high-quality images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 536, + 160, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 160, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 160, + 550 + ], + "type": "text", + "content": "4.2. Proof of Concept" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "text", + "content": "We conduct a proof-of-concept experiment to validate the causality of images in the spectral domain. We use the validation set of the CIFAR-100 dataset [26] as a proxy for the distribution of natural images, which contains 10,000 " + }, + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "text", + "content": " images from 100 semantic classes. We aim to measure the linear correlation " + }, + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "inline_equation", + "content": "(R^2)" + }, + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "text", + "content": " [38] between one token " + }, + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i" + }, + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "text", + "content": " and its previous ones " + }, + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{t}_j\\}_{j=1}^{i-1}" + }, + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "text", + "content": " as a metric for the causality of the sequence " + }, + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "text", + "content": ". We consider three typical autoregressive paradigms, including raster-scan [10], scale-wise [50] and spectral methods which construct sequence " + }, + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "text", + "content": " in different ways. Specifically, we divide each image spatially into four patches of size " + }, + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "text", + "content": " and rearrange them in raster-scan order for the first variant. For the scale-wise method, we" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 537, + 556, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 556, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 556, + 636 + ], + "type": "text", + "content": "downsample each image into sizes of " + }, + { + "bbox": [ + 313, + 537, + 556, + 636 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 313, + 537, + 556, + 636 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 537, + 556, + 636 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 313, + 537, + 556, + 636 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 537, + 556, + 636 + ], + "type": "inline_equation", + "content": "24 \\times 24" + }, + { + "bbox": [ + 313, + 537, + 556, + 636 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 537, + 556, + 636 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 313, + 537, + 556, + 636 + ], + "type": "text", + "content": ", respectively. We set the frequency thresholds to 8, 16, 24, 32 and generate the sub-images following (5) for the spectral sequence. Since each token " + }, + { + "bbox": [ + 313, + 537, + 556, + 636 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 313, + 537, + 556, + 636 + ], + "type": "text", + "content": " is a high dimensional tensor, we calculate " + }, + { + "bbox": [ + 313, + 537, + 556, + 636 + ], + "type": "inline_equation", + "content": "R^2" + }, + { + "bbox": [ + 313, + 537, + 556, + 636 + ], + "type": "text", + "content": " between every element of " + }, + { + "bbox": [ + 313, + 537, + 556, + 636 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i" + }, + { + "bbox": [ + 313, + 537, + 556, + 636 + ], + "type": "text", + "content": " and every element of " + }, + { + "bbox": [ + 313, + 537, + 556, + 636 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{t}_j\\}_{j=1}^{i-1}" + }, + { + "bbox": [ + 313, + 537, + 556, + 636 + ], + "type": "text", + "content": ". For each element of " + }, + { + "bbox": [ + 313, + 537, + 556, + 636 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i" + }, + { + "bbox": [ + 313, + 537, + 556, + 636 + ], + "type": "text", + "content": ", we then find the element in " + }, + { + "bbox": [ + 313, + 537, + 556, + 636 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{t}_j\\}_{j=1}^{i-1}" + }, + { + "bbox": [ + 313, + 537, + 556, + 636 + ], + "type": "text", + "content": " with the highest correlation, and take an average as the final measurement:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 376, + 643, + 555, + 661 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 643, + 555, + 661 + ], + "spans": [ + { + "bbox": [ + 376, + 643, + 555, + 661 + ], + "type": "interline_equation", + "content": "R _ {a v g} ^ {2} = \\operatorname {A V G} _ {m} \\left(\\max _ {n} R _ {m, n} ^ {2}\\right), \\tag {10}", + "image_path": "85649adfbf9c9f405983f6665561a150ff3670b8382f9e1678c1764ed571e099.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 663, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 663, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 663, + 556, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 663, + 556, + 715 + ], + "type": "inline_equation", + "content": "m, n" + }, + { + "bbox": [ + 313, + 663, + 556, + 715 + ], + "type": "text", + "content": " denote element number of " + }, + { + "bbox": [ + 313, + 663, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i" + }, + { + "bbox": [ + 313, + 663, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 663, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{t}_j\\}_{j=1}^{i-1}" + }, + { + "bbox": [ + 313, + 663, + 556, + 715 + ], + "type": "text", + "content": ", respectively, and " + }, + { + "bbox": [ + 313, + 663, + 556, + 715 + ], + "type": "inline_equation", + "content": "R_{m,n}^2" + }, + { + "bbox": [ + 313, + 663, + 556, + 715 + ], + "type": "text", + "content": " is the correlation between the " + }, + { + "bbox": [ + 313, + 663, + 556, + 715 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 663, + 556, + 715 + ], + "type": "text", + "content": "th and " + }, + { + "bbox": [ + 313, + 663, + 556, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 663, + 556, + 715 + ], + "type": "text", + "content": "th elements. We report the quantitative results in Table 1, where the scale-wise and spectral sequences demonstrate" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15848" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 95, + 293, + 205 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 294, + 93 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 294, + 93 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 294, + 93 + ], + "type": "text", + "content": "Table 3. Applications of SepctralAR. Super-reso., Trunc. represent super-resolution and truncated, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 95, + 293, + 205 + ], + "lines": [ + { + "bbox": [ + 58, + 95, + 293, + 205 + ], + "spans": [ + { + "bbox": [ + 58, + 95, + 293, + 205 + ], + "type": "table", + "html": "
App. TypeModelFID↓IS↑
Super-reso.Upsample3.09286.6
SpectralAR-Stride2.93276.4
SpectralAR-Patch14.76170.0
Trunc.SpectralAR-Trunc.53.34271.5
SpectralAR-Trunc.106.65211.4
SpectralAR-Trunc.1527.6991.69
SpectralAR-Trunc.03.02282.2
", + "image_path": "e4a71f84f58b457302b3e323147862330e36e6528379e33fe549c6bd4ef65dd5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 210, + 295, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 210, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 210, + 295, + 270 + ], + "type": "text", + "content": "considerably higher correlation than the raster-scan counterpart. This could be attributed to the coarse-to-fine nature of the former two paradigms, while the raster scan method lacks adequate reference information to predict the next image patch, as shown in Figure 4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 277, + 141, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 141, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 141, + 289 + ], + "type": "text", + "content": "4.3. Main Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 294, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 294, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 295, + 594 + ], + "type": "text", + "content": "We report the performance of SpectralAR on the class-conditional ImageNet-1K [7] " + }, + { + "bbox": [ + 55, + 294, + 295, + 594 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 294, + 295, + 594 + ], + "type": "text", + "content": " generation benchmark in Table 2. We also implement an autoregressive version of TiTok [66] by using the official tokenizer weight and scripts from VAR [50] for a fair comparison. The reconstruction performance of SpectralAR (4.03 rFID) is inferior compared with TiTok because SpectralAR requires to reconstruct the sub-images corresponding to different frequencies with different lengths of tokens, which is much more difficult compared with the overall reconstruction target in TiTok. Despite the lower reconstruction score, SpectralAR-d16 still outperforms TiTok-B64-d16 and VAR-d16 in autoregressive generation with a clear margin due to better sequence causality which eases autoregressive learning. Furthermore, SpectralAR also achieves better or comparable performance against VAR under the " + }, + { + "bbox": [ + 55, + 294, + 295, + 594 + ], + "type": "inline_equation", + "content": "d20" + }, + { + "bbox": [ + 55, + 294, + 295, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 294, + 295, + 594 + ], + "type": "inline_equation", + "content": "d24" + }, + { + "bbox": [ + 55, + 294, + 295, + 594 + ], + "type": "text", + "content": " settings, respectively. In addition, SpectralAR uses only 64 tokens in both reconstruction and generation, demonstrating superior token efficiency compared to VAR [50] and 2D scan-based methods. We also visualize the generated samples in Figure 5, which shows the diversity and quality of the generation process of SpectralAR. In addition, Figure 1 highlights the hierarchical coarse-to-fine refinement of the images, while SpectralAR generates more tokens in an autoregressive way." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 600, + 137, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 137, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 137, + 613 + ], + "type": "text", + "content": "4.4. Applications" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": "In this section, we provide a quantitative analysis for the potential applications of SpectralAR. For super-resolution, we conduct our experiments based on the " + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": " images generated by SpectralAR-" + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "inline_equation", + "content": "d16" + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": " in Table 2, and directly upsample them to " + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": " as the baseline. We construct 4 sub-images using stripped and patch-based methods for SpectralAR-Stride and -Patch, respectively. We then upsample the 4 sub-images to " + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": " and use SpectralAR" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 117, + 553, + 198 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 553, + 115 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 115 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 115 + ], + "type": "text", + "content": "Table 4. Ablation on design choices. Spectral Supervision means using sub-image supervision across different frequency bands. Causal Mask and non-uniform mapping refers to the spectral causal mask and the frequency-token mapping, respectively." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 117, + 553, + 198 + ], + "lines": [ + { + "bbox": [ + 317, + 117, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 317, + 117, + 553, + 198 + ], + "type": "table", + "html": "
Spectral SupervisionCausal MaskNon-uniform MappingFID↓IS↑
×××6.30190.1
✓××5.64255.1
✓✓×3.49222.6
✓✓✓3.02282.2
", + "image_path": "959f2a27fcfb3240f8bd334c97a8238301d8d5f9cc535ab95bc33c3f2e290ee9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 205, + 555, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 555, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 555, + 337 + ], + "type": "text", + "content": "to refine them in the spectral domain, and finally reassemble them to generate the final result. According to Table 3, SpectralAR can indeed serve as a spectrum completer for the super-resolution task. In addition, we also experiment with truncated autoregressive generation, where we discard the last few tokens. For example, we discard the last 5 tokens in SpectralAR-Trunc.5 in Table 3. The generation performance worsens slowly when the number of truncated tokens is fewer than 10, and therefore it is possible to further improve token efficiency through truncation according to the requirement for generation quality." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 342, + 408, + 356 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 408, + 356 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 408, + 356 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 361, + 556, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 556, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 556, + 540 + ], + "type": "text", + "content": "We conduct ablation study to validate the effectiveness of our design choices in Table 4. The first row corresponds to the baseline TiTok implementation for autoregressive generation. Note that the spectral supervision alone could improve FID compared with the TiTok counterpart. This is because the spectral design enhances sequence causality compared with the overall reconstruction target which ignores the correlation between tokens and thus complicates the autoregressive modeling process. The spectral causal mask further enhances performance by improving causality in the encoding and decoding architecture. And the non-uniform token-frequency mapping technique guides the model to focus more on the crucial low-frequency components while representing the high-frequency components efficiently, thus further improving performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": "In this paper, we have proposed the spectral autoregressive visual generation method for both causal and efficient autoregressive modeling of image data. Specifically, we first convert images into 1D sequences with nested spectral tokenization. In addition, we have adopted causal masks for spectral tokens in the encoder and decoder to further enhance causality from the architectural perspective. We have also designed a non-uniform token-frequency mapping with emphasis on the low-frequency components in order to improve token efficiency. On the ImageNet-1K generation benchmark, our SpectralAR achieves superior performance compared with other autoregressive generation methods." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15849" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 147 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 147 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 62125603, Grant 62321005, Grant 62336004, and Grant 62441616, and in part by the Beijing Natural Science Foundation under Grant No. L247009." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 169, + 115, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 169, + 115, + 182 + ], + "spans": [ + { + "bbox": [ + 56, + 169, + 115, + 182 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 190, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 190, + 295, + 222 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 190, + 295, + 222 + ], + "spans": [ + { + "bbox": [ + 61, + 190, + 295, + 222 + ], + "type": "text", + "content": "[1] Nasir Ahmed, T. Natarajan, and Kamisetty R Rao. Discrete cosine transform. IEEE transactions on Computers, 100(1): 90-93, 2006. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 224, + 295, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 224, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 224, + 295, + 246 + ], + "type": "text", + "content": "[2] Jean Baptiste Joseph Baron Fourier et al. The analytical theory of heat. Courier Corporation, 2003. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 248, + 294, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 248, + 294, + 280 + ], + "spans": [ + { + "bbox": [ + 62, + 248, + 294, + 280 + ], + "type": "text", + "content": "[3] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 282, + 295, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 282, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 62, + 282, + 295, + 314 + ], + "type": "text", + "content": "[4] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In CVPR, pages 11315-11325, 2022. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 317, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 317, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 62, + 317, + 294, + 358 + ], + "type": "text", + "content": "[5] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In ICML, pages 1691-1703. PMLR, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 362, + 294, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 362, + 294, + 394 + ], + "spans": [ + { + "bbox": [ + 62, + 362, + 294, + 394 + ], + "type": "text", + "content": "[6] Aaron Clauset, Cosma Rohilla Shalizi, and Mark EJ Newman. Power-law distributions in empirical data. SIAM review, 51(4):661-703, 2009. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 396, + 294, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 396, + 294, + 429 + ], + "spans": [ + { + "bbox": [ + 62, + 396, + 294, + 429 + ], + "type": "text", + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009. 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 431, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 431, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 431, + 295, + 453 + ], + "type": "text", + "content": "[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NIPS, 34:8780-8794, 2021. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 454, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 454, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 62, + 454, + 294, + 518 + ], + "type": "text", + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 521, + 294, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 521, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 57, + 521, + 294, + 553 + ], + "type": "text", + "content": "[10] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, pages 12873-12883, 2021. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 555, + 294, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 555, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 57, + 555, + 294, + 586 + ], + "type": "text", + "content": "[11] Carlos Esteves, Mohammed Suhail, and Ameesh Makadia. Spectral image tokenizer. arXiv preprint arXiv:2412.09607, 2024.3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 590, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 294, + 633 + ], + "type": "text", + "content": "[12] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In ICML, pages 3247-3258. PMLR, 2020. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 635, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 294, + 666 + ], + "type": "text", + "content": "[13] Manuel Fritsche, Shuhang Gu, and Radu Timofte. Frequency separation for real-world super-resolution. In ICCVW, pages 3599-3608. IEEE, 2019. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "text", + "content": "[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. NIPS, 27, 2014. 2, 4" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[15] Hayit Greenspan, Charles H Anderson, and Sofia Akber. Image enhancement by nonlinear extrapolation in frequency space. TIP, 9(6):1035-1048, 2000. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 108, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 150 + ], + "type": "text", + "content": "[16] Jian Han, Jinlai Liu, Yi Jiang, Bin Yan, Yuqi Zhang, Zehuan Yuan, Bingyue Peng, and Xiaobing Liu. Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis. arXiv preprint arXiv:2412.04431, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 152, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 194 + ], + "type": "text", + "content": "[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NIPS, 30, 2017. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 198, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 228 + ], + "type": "text", + "content": "[18] Geoffrey E Hinton and Ruslan R Salakhutdinov. Reducing the dimensionality of data with neural networks. science, 313(5786):504-507, 2006. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 231, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 252 + ], + "type": "text", + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NIPS, 33:6840-6851, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 255, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 255, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 255, + 553, + 296 + ], + "type": "text", + "content": "[20] Jonathan Ho, Chitwan Saharia, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. JMLR, 23(47):1-33, 2022. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 300, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 300, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 300, + 553, + 342 + ], + "type": "text", + "content": "[21] Yuanhui Huang, Wenzhao Zheng, Yuan Gao, Xin Tao, Pengfei Wan, Di Zhang, Jie Zhou, and Jiwen Lu. Owl1: Omni world model for consistent long video generation. arXiv preprint arXiv:2412.09600, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 344, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 553, + 376 + ], + "type": "text", + "content": "[22] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, pages 694–711. Springer, 2016. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 378, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 378, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 317, + 378, + 553, + 420 + ], + "type": "text", + "content": "[23] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In CVPR, pages 10124-10134, 2023. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 422, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 443 + ], + "type": "text", + "content": "[24] Diederik P Kingma, Max Welling, et al. Auto-encoding variational bayes, 2013. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 446, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 446, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 317, + 446, + 553, + 499 + ], + "type": "text", + "content": "[25] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Grant Schindler, Rachel Hornung, Vighnesh Birodkar, Jimmy Yan, Ming-Chang Chiu, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 502, + 553, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 502, + 553, + 522 + ], + "spans": [ + { + "bbox": [ + 316, + 502, + 553, + 522 + ], + "type": "text", + "content": "[26] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 525, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 525, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 525, + 553, + 588 + ], + "type": "text", + "content": "[27] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 128(7):1956-1981, 2020. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 591, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 591, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 317, + 591, + 553, + 632 + ], + "type": "text", + "content": "[28] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 636, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 553, + 678 + ], + "type": "text", + "content": "[29] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In CVPR, pages 11523-11532, 2022. 2, 3, 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "text", + "content": "[30] Tianhong Li, Dina Katabi, and Kaiming He. Self-conditioned image generation via generating representations. CoRR, 2023. 6" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15850" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[31] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. NIPS, 37:56424-56445, 2025. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "text", + "content": "[32] Yuyang Long, Qilong Zhang, Boheng Zeng, Lianli Gao, Xianglong Liu, Jian Zhang, and Jingkuan Song. Frequency domain model augmentation for adversarial attack. In ECCV, pages 549-566. Springer, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[33] Cheng Luo, Qinliang Lin, Weicheng Xie, Bizhu Wu, Jinheng Xie, and Linlin Shen. Frequency-driven imperceptible adversarial attack on semantic similarity. In CVPR, pages 15315-15324, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "type": "text", + "content": "[34] Zhuoyan Luo, Fengyuan Shi, Yixiao Ge, Yujiu Yang, Limin Wang, and Ying Shan. Open-magvit2: An open-source project toward democratizing auto-regressive visual generation. arXiv preprint arXiv:2409.04410, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 294, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 295 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 295 + ], + "type": "text", + "content": "[35] Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. In ECCV, pages 23-40. Springer, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 341 + ], + "type": "text", + "content": "[36] Xiaoxiao Ma, Mohan Zhou, Tao Liang, Yalong Bai, Tiejun Zhao, Huaian Chen, and Yi Jin. Star: Scale-wise text-to-image generation via auto-regressive representations. arXiv preprint arXiv:2406.10797, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 342, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 374 + ], + "type": "text", + "content": "[37] Fabian Mentzer, David Minnen, Eirikur Agustsson, and Michael Tschannen. Finite scalar quantization: Vq-vae made simple. arXiv preprint arXiv:2309.15505, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 376, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 398 + ], + "type": "text", + "content": "[38] Daniel J Ozer. Correlation and the coefficient of determination. Psychological bulletin, 97(2):307, 1985. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 399, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 430 + ], + "type": "text", + "content": "[39] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 433, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 294, + 465 + ], + "type": "text", + "content": "[40] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 467, + 294, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 509 + ], + "type": "text", + "content": "[41] Yongming Rao, Wenliang Zhao, Zheng Zhu, Jiwen Lu, and Jie Zhou. Global filter networks for image classification. Advances in neural information processing systems, 34:980-993, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 511, + 294, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 294, + 543 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 294, + 543 + ], + "type": "text", + "content": "[42] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. NIPS, 32, 2019. 2, 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 544, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 294, + 577 + ], + "type": "text", + "content": "[43] Siddharth Roheda. Cart: Compositional auto-regressive transformer for image generation. arXiv preprint arXiv:2411.10180, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 578, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 294, + 621 + ], + "type": "text", + "content": "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 3, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 624, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 656 + ], + "type": "text", + "content": "[45] Axel Sauer, Katja Schwarz, and Andreas Geiger. Styleganx1: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH 2022 conference proceedings, pages 1-10, 2022. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "type": "text", + "content": "[46] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[47] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[48] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "type": "text", + "content": "[49] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 2, 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "text", + "content": "[50] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. NIPS, 37:84839-84865, 2025. 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 220, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 252 + ], + "type": "text", + "content": "[51] Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelCNN decoders. NIPS, 29, 2016. 3, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 254, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 275 + ], + "type": "text", + "content": "[52] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NIPS, 30, 2017. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 277, + 553, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 553, + 309 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 553, + 309 + ], + "type": "text", + "content": "[53] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NIPS, 30, 2017. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 311, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 311, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 311, + 553, + 354 + ], + "type": "text", + "content": "[54] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In ICML, pages 1096-1103, 2008. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 355, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 388 + ], + "type": "text", + "content": "[55] Gregory K Wallace. The jpeg still picture compression standard. Communications of the ACM, 34(4):30-44, 1991. 2, 3, 4, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 389, + 553, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 389, + 553, + 433 + ], + "spans": [ + { + "bbox": [ + 316, + 389, + 553, + 433 + ], + "type": "text", + "content": "[56] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 2, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 434, + 553, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 478 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 478 + ], + "type": "text", + "content": "[57] Yuqing Wang, Tianwei Xiong, Daquan Zhou, Zhijie Lin, Yang Zhao, Bingyi Kang, Jiashi Feng, and Xihui Liu. Loong: Generating minute-level long videos with autoregressive language models. arXiv preprint arXiv:2410.02757, 2024. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 479, + 553, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 553, + 511 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 553, + 511 + ], + "type": "text", + "content": "[58] Thomas Wiegand, Gary J Sullivan, Gisle Bjontegaard, and Ajay Luthra. Overview of the h. 264/avc video coding standard. TCSVT, 13(7):560-576, 2003. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 513, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 513, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 513, + 553, + 567 + ], + "type": "text", + "content": "[59] Chengyue Wu, Xiaokang Chen, Zhiyu Wu, Yiyang Ma, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, Chong Ruan, et al. Janus: Decoupling visual encoding for unified multimodal understanding and generation. arXiv preprint arXiv:2410.13848, 2024. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 569, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 569, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 569, + 553, + 624 + ], + "type": "text", + "content": "[60] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 624, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 666 + ], + "type": "text", + "content": "[61] Mengping Yang, Zhe Wang, Ziqiu Chi, and Wenyi Feng. Wavegan: Frequency-aware gan for high-fidelity few-shot image generation. In ECCV, pages 1-17. Springer, 2022. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "text", + "content": "[62] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 2" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15851" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 576 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "text", + "content": "[63] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "text", + "content": "[64] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 217 + ], + "type": "text", + "content": "[65] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Randomized autoregressive visual generation. arXiv preprint arXiv:2411.00776, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 261 + ], + "type": "text", + "content": "[66] Qihang Yu, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An image is worth 32 tokens for reconstruction and generation. NIPS, 37: 128940-128966, 2025. 2, 3, 4, 6, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 294, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 307 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 307 + ], + "type": "text", + "content": "[67] Yingchen Yu, Fangneng Zhan, Shijian Lu, Jianxiong Pan, Feiying Ma, Xuansong Xie, and Chunyan Miao. Wavefill: A wavelet-based generation network for image inpainting. In ICCV, pages 14114-14123, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 350 + ], + "type": "text", + "content": "[68] Huanjing Yue, Xiaoyan Sun, Jingyu Yang, and Feng Wu. Cid: Combined image denoising in spatial and frequency domains using web images. In CVPR, pages 2933-2940, 2014. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 353, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 294, + 396 + ], + "type": "text", + "content": "[69] Qian Zhang, Xiangzi Dai, Ninghua Yang, Xiang An, Ziyong Feng, and Xingyu Ren. Var-clip: Text-to-image generator with visual auto-regressive modeling. arXiv preprint arXiv:2408.01181, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 398, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 398, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 398, + 294, + 441 + ], + "type": "text", + "content": "[70] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 443, + 294, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 443, + 294, + 496 + ], + "spans": [ + { + "bbox": [ + 56, + 443, + 294, + 496 + ], + "type": "text", + "content": "[71] Renrui Zhang, Jiaming Han, Chris Liu, Peng Gao, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, and Yu Qiao. Llama-adapter: Efficient finetuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 498, + 294, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 294, + 531 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 294, + 531 + ], + "type": "text", + "content": "[72] Chuanxia Zheng, Tung-Long Vuong, Jianfei Cai, and Dinh Phung. Movq: Modulating quantized vectors for high-fidelity image generation. NIPS, 35:23412-23425, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 533, + 294, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 533, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 56, + 533, + 294, + 576 + ], + "type": "text", + "content": "[73] Wenzhao Zheng, Weiliang Chen, Yuanhui Huang, Borui Zhang, Yueqi Duan, and Jiwen Lu. Occworld: Learning a 3d occupancy world model for autonomous driving. In ECCV, pages 55-72. Springer, 2024. 2" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15852" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Spherical Epipolar Rectification for Deep Two-View Absolute Depth Estimation/580cf2a2-db43-4692-a58d-cadb1e36cc3b_content_list.json b/2025/Spherical Epipolar Rectification for Deep Two-View Absolute Depth Estimation/580cf2a2-db43-4692-a58d-cadb1e36cc3b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6534e92ec7591d6e940414c0633086976bd1db78 --- /dev/null +++ b/2025/Spherical Epipolar Rectification for Deep Two-View Absolute Depth Estimation/580cf2a2-db43-4692-a58d-cadb1e36cc3b_content_list.json @@ -0,0 +1,1882 @@ +[ + { + "type": "text", + "text": "Spherical Epipolar Rectification for Deep Two-View Absolute Depth Estimation", + "text_level": 1, + "bbox": [ + 96, + 130, + 898, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pierre-André Brousseau \nUniversité de Montréal", + "bbox": [ + 276, + 181, + 470, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pierre-andre.brousseau@umontreal.ca", + "bbox": [ + 217, + 219, + 529, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sébastien Roy \nUniversité de Montréal \nroys@iro.umontreal.ca", + "bbox": [ + 588, + 181, + 774, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Absolute depth estimation from single camera sequence of images is a relevant task given that mobile machines increasingly rely on vision to navigate. Deep learning for stereo matching has been demonstrated to improve performance for stereo rectified depth estimation but these methods require straightforward left-right camera setups. This work proposes to introduce deep stereo matching to two views of a monocular image sequence obtained from a camera in motion in a static scene. This paper introduces a novel and principled spherical epipolar rectification model, which handles all camera motions. This rectification model is differentiable and allows self-supervised deep stereo matching algorithms to compute disparity and recover depth, given known camera pose. This paper also introduces a spherical crop operation which limits rectified image size and allows for competitive absolute depth estimation performance. This results in a spherical rectification model that is demonstrated to provide metric depth and compete favorably with current state-of-the-art monocular depth estimators. The code is available at https://gitlab.com/labv3d/spherical-stereo.git.", + "bbox": [ + 88, + 300, + 485, + 618 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 648, + 220, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reconstructing the world is an important task in the computer vision field. Multiple views with a calibrated camera system makes it possible to solve for metric depth passively. As humans, we capture the 3D world with a binocular camera system, our eyes, but most camera systems that capture the world as photos and videos instead rely on a single camera. This work tackles the problem of recovering absolute metric depth from two image frames of a monocular sequence with known camera pose using self-supervised deep neural networks. It will be shown that depth from two views with known pose, known as stereo depth estimation, can be extended to monocular sequences by spherical epipolar rectification.", + "bbox": [ + 89, + 672, + 482, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The challenge to address with deep networks is to enable stereo depth estimation for devices where only a single cam", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5f707861d5055841b24c0948b9a2c5483b597ed2a0185242789babd19d6e0a6e.jpg", + "image_caption": [ + "Figure 1. Comparing two-view depth estimation from plane sweep and the proposed two-view depth estimation from spherical epipolar rectification." + ], + "image_footnote": [], + "bbox": [ + 521, + 268, + 901, + 431 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "era is available. This case occurs in systems where the cost of the second camera is prohibitive but pose can be recovered reliably (GPS, IMU, EKF). In this situation and given two views, it is worth exploring methods that can directly solve for absolute metric depth rather than scaled relative depth. Because of the known pose, an epipolar geometry links both views and reduces the task of optical flow estimation, which has a 2D search space, to a stereo matching task, which has a 1D search space. This 1D search space is expressed to deep neural networks as rectified images where the match for each pixel lies on the same row in the matching image. This process is referred to as rectification whereby images are warped to fit the same row requirement [27].", + "bbox": [ + 509, + 506, + 906, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Epipolar rectification for two views of a monocular sequence differs from the standard synchronized two-view stereo camera systems because the rectification cannot be precomputed for image pairs. In the case of two-view depth estimation from a monocular sequence, the epipolar rectification must be computed according to the pose and must handle forward translation. Planar rectification, through homography warping, breaks down when the projection of the second camera center, the epipole, is inside the image. This problem can be addressed by using a rectification with a polar coordinate system, but this category of methods leads to increased deformation of the image structure, i.e. dis", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28925", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tortion. Previously proposed methods for polar rectification have not yet been integrated to deep networks because of the constantly changing rectified image dimensions and the large distortion that significantly affects image structure. In this work, we show that when correctly handled, these distorted images lend themselves to deep neural network applications that can learn to solve for stereoscopic correspondences in monocular image pairs.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper will explore spherical epipolar rectification in the case of monocular image sequences for deep neural stereo matching methods. Many devices that have a camera already include an IMU or a GPS for other purposes, and leveraging them for depth estimation adds value at no extra hardware cost. This work proposes a significant step towards metric monocular depth estimation and the contributions of this paper are as follows:", + "bbox": [ + 89, + 212, + 480, + 330 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a differentiable spherical rectification model that enables deep stereo matching for monocular image pairs.", + "- We introduce the spherical crop operator which reduces distortion and therefore allows training of neural networks.", + "- We demonstrate the performance of spherical rectified stereo matching for solving two-view monocular depth estimation." + ], + "bbox": [ + 89, + 333, + 480, + 467 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To our knowledge, this work is the first to integrate polar rectification into deep stereo matching.", + "bbox": [ + 89, + 469, + 480, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 511, + 232, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Solving depth in a static scene from two images with known camera parameters and pose is a longstanding registration problem. A wide category of methods operate on rectified stereo image pairs to solve for disparity to impressive results. These methods cannot be trivially extended to arbitrary camera geometries. To operate on arbitrary camera geometries, most methods represent the world as 3D world points and solve for depth. This work aims to present the theoretical framework to broaden the concept of stereo matching to two views of a monocular image sequences.", + "bbox": [ + 89, + 537, + 482, + 689 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Polar Rectification for Stereo Matching", + "text_level": 1, + "bbox": [ + 89, + 696, + 428, + 713 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Planar rectification has the drawback of creating rectified images that are potentially infinitely large when the epipoles are inside the images. Cylindrical rectification [25] introduces a new rectification method that guarantees that the rectified images are bounded for any stereo geometry. This rectification is improved by parameterizing the image with polar coordinates [6, 24] which guarantees minimal image size and reduces matching ambiguity to half epipolar lines. Google's Depth from Motion [28] reveals that rectified images can be arbitrarily large and that it should be resized according to an expected maximum disparity. The angular sampling can be improved with a linear transformation [18]", + "bbox": [ + 89, + 719, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/47cc29655b7cc97029282d8d1b5ab8e372e72d7f7ad08530a274602c612c9780.jpg", + "image_caption": [ + "Figure 2. Spherical Representation. Any point $(x,y,z)$ along a camera ray can be projected on the unit sphere and be expressed by two angles, $\\alpha$ and $\\beta$ . An epipolar line is selected by $\\alpha$ , and $\\beta$ is the position along that line." + ], + "image_footnote": [], + "bbox": [ + 584, + 92, + 836, + 209 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and the log-polar [9, 23] coordinate system can be used to reduce distortion. Some works use a variant of cylindrical epipolar rectification in the context of large FOV cameras [22] for stereo matching. Auto-Rectify [2] proposes a deep learning compatible planar epipolar rectification formulation. As it is planar, the rectification cannot handle forward translation. Its experiments therefore deliberately ignore the forward translation because it causes extremely deformed images.", + "bbox": [ + 511, + 292, + 903, + 429 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our proposed work improves upon these related works by using a generalized spherical transformation operation, by expressing position along an epipolar line as an angle, that holds when cropping to smaller regions. Spherical crop solves the problem of resolving the optimal scaling and angle intervals. This allows for control over image sizes. Our formulation is differentiable and can be readily integrated with deep neural networks.", + "bbox": [ + 511, + 430, + 905, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Deep Depth from Two Monocular Images", + "text_level": 1, + "bbox": [ + 511, + 561, + 867, + 577 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Depth from Direct Regression. Most methods estimate depth directly with a neural network for a given frame and validate it with other frames using the camera geometry [17, 21]. Feature or image warping provides sufficient constraints to solve for depth [40, 43]. Novel loss functions on the geometry consistency [1] and the epipolar geometry [7] are proposed to isolate general motion from camera motion. Novel frameworks to visual odometry such as a photometric residual [38] or fundamental matrix estimation [44] have also been proposed. Depth Anything V2 [37] solve for depth in image sequences by building a coherent transformer latent space. Most recently, SC-DepthV3 [26] introduces a novel depth ranking loss to distill from a single-image depth estimator.", + "bbox": [ + 511, + 583, + 905, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Depth from Epipolar Geometry Matching. On the other hand, some approaches use the epipolar geometry to form a cost volume for depth candidates. Differentiable homography [34, 39] warping and plane sweep, such as DeepSfm Revisited [29], can be used to map features from two camera views back to a single 3D cost volume representation linked to the reference view. This depth-discretized epipo", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28926", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/72cb1b0521409ba8d16c63061fee2a87d1eef3d6e7b491b8f0782b5806982cb1.jpg", + "image_caption": [ + "Figure 3. Spherical rectification on a Spherical Crop. For a given image crop, the ranges $\\left[\\alpha_{\\min}, \\alpha_{\\max}\\right], \\left[\\beta_{\\min}, \\beta_{\\max}\\right]$ can be computed from the pixel coordinates of the corners using spherical epipolar geometry. This can be used to map an image crop (LEFT) of size $H_C \\times W_C$ to a spherical crop (RIGHT) $H_S \\times W_S$ . This leads to reduced distortion with respect to mapping the full image at once. This rectification process is illustrated in Rectification-Visualization." + ], + "image_footnote": [], + "bbox": [ + 140, + 89, + 630, + 265 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4e5b66173b1a7c936c2dd49596b6000f088881c16a64f1964d68f1d29e12b161.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 89, + 854, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "lar sampling can be used in conjunction with attention [15] to yield cross-attention cost volumes. This has lead to models that unify the formulation for depth and optical flow, GMDepth [36], through a transformer backbone. Most recently, DUSt3R [31] have represented the 3D reconstruction task as an end-to-end 3D pointmap matching problem. This is furthered in MonST3R [42] which integrates semantic segmentation and global pointcloud optimization.", + "bbox": [ + 88, + 353, + 480, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 1, in comparison to the related works, our proposed approach cannot generate depth for a single frame nor does it preemptively discretize depth. It estimates stereoscopic disparity between two frames and computes absolute depth from camera geometry. This disentangles the geometry from the matching process through spherical epipolar rectification.", + "bbox": [ + 88, + 477, + 482, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Self-Supervised Stereo Matching", + "text_level": 1, + "bbox": [ + 89, + 599, + 377, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Deep stereo matching has had significant advancements and has lead to powerful and detailed results [8, 20, 35]. In the context of this work, we focus on deep stereo methods that can also be trained without ground truth as it emphasizes the epipolar rectification capabilities. The self-supervised stereo matching task aims to generate a disparity that best explains image displacement without ground truth information. Neural networks can be trained in a self-supervised fashion by minimizing the photometric consistency [45], disparity consistency [13] and smoothing [41]. Novel cost volume representations [4, 19] and epipolar constraints [30] have lead to improved performance by guiding the disparity prediction along epipolar lines.", + "bbox": [ + 88, + 625, + 482, + 823 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our proposed method does not contribute to stereo matching per se but relies on Permutation Stereo as detailed in [5] and OpenCV SGBM [16]. It applies deep stereo matching to the case of general motion by spherical epipolar rectification.", + "bbox": [ + 88, + 824, + 483, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Spherical Rectification Model", + "text_level": 1, + "bbox": [ + 511, + 351, + 785, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, the proposed spherical rectification model presumes a pair of images taken from a rectilinear camera in motion. This section describes the mathematical basis of spherical epipolar rectification.", + "bbox": [ + 511, + 378, + 906, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Rectilinear Camera Model", + "text_level": 1, + "bbox": [ + 511, + 454, + 754, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let us describe a camera image $I^{i} \\in \\mathbf{R}^{H_{I} \\times W_{I}}$ taken at a moment $i$ . Spherical rectification occurs between images $(I^{i}, I^{j})$ taken by the same camera at two moments in time $(i, j)$ . The image formation model for a point $(x_{w}, y_{w}, z_{w})$ in the 3D world is:", + "bbox": [ + 511, + 477, + 906, + 553 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI ^ {i} (x, y) = K \\cdot \\left[ R _ {c} ^ {i} \\mid t _ {c} ^ {i} \\right] \\cdot \\left( \\begin{array}{l} x _ {w} \\\\ y _ {w} \\\\ z _ {w} \\\\ 1 \\end{array} \\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 566, + 905, + 630 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nK = \\left[ \\begin{array}{c c c} f _ {x} & 0 & p _ {x} \\\\ 0 & f _ {y} & p _ {y} \\\\ 0 & 0 & 1 \\end{array} \\right]\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 631, + 782, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The internal parameters matrix $K$ is described by $(f_x, f_y)$ , the focal length of the camera in terms of pixel dimensions in the x-axis and the y-axis respectively, and $(p_x, p_y)$ , the coordinates of the principal point. The pose at instant $i$ is provided by rotation matrix $R_c^i$ and translation vector $t_c^i$ .", + "bbox": [ + 511, + 695, + 905, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Spherical Geometry from Poses", + "text_level": 1, + "bbox": [ + 511, + 786, + 792, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spherical rectification requires that its main axis $z$ be aligned with the direction of translation. This axis represents the epipoles and contains both camera centers. Let us designate the extrinsic parameter of a camera at a moment $i$ as $[R_c^i |t_c^i ]\\in \\mathbf{R}^{3\\times 4}$ . The spherical orientation is described as a rotation matrix $R_{S}$ and is recovered from the orientation", + "bbox": [ + 511, + 809, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28927", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8b11457911e07e897ba9466337c9f6948f68ee07616ff2cc218f0ae71df0c62a.jpg", + "image_caption": [ + "Figure 4. Spherical Stereo. Crops from image $I^i$ and $I^{i-1}$ are rectified. A shared feature encoder learns stereo features which are then compared with the dot product. Symmetric normalization (N) is applied according to the Permutation Stereo[4] formulation and the estimated disparities are transformed into depths. The reconstructed images and the depths are then derectified back to image space." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 911, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "given by the translation of camera centers.", + "bbox": [ + 89, + 375, + 372, + 388 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nr _ {z} = t ^ {i} - t ^ {j}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 396, + 277, + 412 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nr _ {y} = (0, 0, 1) \\times r _ {z}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 417, + 318, + 434 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nr _ {x} = r _ {y} \\times r _ {z} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 436, + 482, + 452 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR _ {S} ^ {i, j} = \\left(\\frac {r _ {x}}{| | r _ {x} | |}, \\frac {r _ {y}}{| | r _ {y} | |}, \\frac {r _ {z}}{| | r _ {z} | |}\\right) ^ {T}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 454, + 393, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Spherical Representation", + "text_level": 1, + "bbox": [ + 89, + 498, + 321, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The angles $(\\alpha, \\beta)$ designate a 3D point $(x, y, z)$ on the unit sphere, see Fig. 2, which is aligned with the spherical model principal axis according to the rotation $R_S^{i,j}$ . The coordinate change operation $N_{\\alpha \\beta \\rightarrow xyz}$ is:", + "bbox": [ + 89, + 520, + 483, + 580 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} (x, y, z) ^ {T} = N _ {\\alpha \\beta \\rightarrow x y z} (\\alpha , \\beta) \\\\ = \\left( \\begin{array}{c} \\sin (\\alpha) \\sin (\\beta) \\\\ - \\cos (\\alpha) \\sin (\\beta) \\\\ \\cos (\\beta) \\end{array} \\right) \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 589, + 482, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and the inverse operation $N_{xyz\\to \\alpha \\beta}$ is:", + "bbox": [ + 89, + 665, + 351, + 681 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} (\\alpha , \\beta) ^ {T} = N _ {x y z \\rightarrow \\alpha \\beta} (x, y, z) \\\\ = \\left( \\begin{array}{c} \\tan^ {- 1} \\left(\\frac {x}{- y}\\right) \\\\ \\left| \\tan^ {- 1} \\left(\\frac {\\operatorname {s g n} (x) \\sqrt {x ^ {2} + y ^ {2}}}{z}\\right) \\right| \\end{array} \\right) \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 688, + 482, + 767 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Epipolar Rectification", + "text_level": 1, + "bbox": [ + 89, + 772, + 295, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Because the camera field of view is usually small relative to the full sphere, spherical rectification can yield sparse images. We thus propose to partition the input images into a set of corresponding small regions which are then rectified into shared spherical representations, see Fig. 3. Let us designate a rectified image on the unit sphere, where the pixel intensity is $I_{S}(a,b) \\in \\mathbf{R}^{H_{S} \\times W_{S}}$ .", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The indices of the rows and columns $(a,b)$ of a spherical representation image $I_{S}$ represent uniformly sampled angles $(\\alpha ,\\beta)$ on the unit sphere such that $(\\alpha ,\\beta)\\in$ $\\left([\\alpha_{min},\\alpha_{max}],[\\beta_{min},\\beta_{max}]\\right)$ in a rectified image size $(H_S,W_S)$ , the spherical image height and width, see Fig. 3 (RIGHT). Conversion operations can be defined $N_{ab\\to \\alpha \\beta}$ between the pair $(a,b)$ and $(\\alpha ,\\beta)$ .", + "bbox": [ + 511, + 375, + 906, + 481 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} (\\alpha , \\beta) ^ {T} = N _ {a b \\rightarrow \\alpha \\beta} (a, b) \\\\ = \\left( \\begin{array}{c} \\frac {a \\left(\\alpha_ {\\text {m a x}} - \\alpha_ {\\text {m i n}}\\right)}{H _ {S}} + \\alpha_ {\\text {m i n}} \\\\ \\frac {b \\left(\\beta_ {\\text {m a x}} - \\beta_ {\\text {m i n}}\\right)}{W _ {S}} + \\beta_ {\\text {m i n}} \\end{array} \\right) \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 488, + 903, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The spherical epipolar rectification $M_S$ is a rectification function that extracts a pixel at position $(u, v)$ from an unrectified image $I$ given an $(a, b)$ pair. Therefore, pixel intensity in the rectified image is $I_S(a, b) = I(u, v)$ .", + "bbox": [ + 511, + 556, + 905, + 617 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The spherical epipolar rectification $M_S$ from $(u, v)$ coordinates to $(a, b)$ coordinates is:", + "bbox": [ + 511, + 617, + 905, + 647 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n(u, v) = M _ {S} (a, b)\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 656, + 727, + 672 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n(\\alpha , \\beta) ^ {T} = N _ {a b \\rightarrow \\alpha \\beta} (a, b)\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 675, + 761, + 693 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left(x _ {p}, y _ {p}, z _ {p}\\right) ^ {T} = K \\cdot R _ {c} \\cdot R _ {s} \\cdot N _ {\\alpha \\beta \\rightarrow x y z} (\\alpha , \\beta) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 696, + 903, + 715 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n(u, v) = \\left(\\frac {y _ {p}}{z _ {p}}, \\frac {x _ {p}}{z _ {p}}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 718, + 736, + 752 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Epipolar Derectification", + "text_level": 1, + "bbox": [ + 511, + 757, + 735, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The complementary operator to epipolar rectification is the epipolar derectification. Let us define the conversion operation $N_{\\alpha \\beta \\to ab}$ between the pair $(\\alpha, \\beta)$ and $(a, b)$ .", + "bbox": [ + 511, + 780, + 905, + 825 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} (a, b) ^ {T} = N _ {\\alpha \\beta \\rightarrow a b} (\\alpha , \\beta) \\\\ = \\binom {(\\alpha - \\alpha_ {\\min }) H _ {S} / (\\alpha_ {\\max } - \\alpha_ {\\min })} {\\left. (m o d (\\beta , \\pi) - \\beta_ {\\min }\\right) W _ {S} / \\left(\\beta_ {\\max } - \\beta_ {\\min }\\right)} \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 833, + 903, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28928", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the same manner, an unrectified image $I$ with coordinates $(u,v)$ can be generated from a spherical image $I_S$ such that $I(u,v) = I_S(a,b)$ . The spherical epipolar derectification $M_S^{-1}$ operator is an operator that maps a coordinate in the camera coordinate system to a coordinate on the unit sphere.", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n(a, b) = M _ {S} ^ {- 1} (u, v)\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 191, + 321, + 209 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left( \\begin{array}{l} x _ {s} \\\\ y _ {s} \\\\ z _ {s} \\end{array} \\right) = R _ {s} ^ {T} \\cdot R _ {c} ^ {T} \\cdot K ^ {- 1} \\cdot \\left( \\begin{array}{l} u \\\\ v \\\\ 1 \\end{array} \\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 212, + 480, + 260 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\alpha , \\beta\\right) ^ {T} = N _ {x y z \\rightarrow \\alpha \\beta} \\left(x _ {s}, y _ {s}, z _ {s}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 263, + 388, + 280 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n(a, b) ^ {T} = N _ {\\alpha \\beta \\rightarrow a b} (\\alpha , \\beta)\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 282, + 348, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Spherical Stereo for Two Monocular Views", + "text_level": 1, + "bbox": [ + 89, + 315, + 473, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Spherical Crop", + "text_level": 1, + "bbox": [ + 89, + 340, + 243, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Because spherical epipolar rectification introduces some distortion in the rectified image, we propose a novel spherical crop function illustrated in Fig. 3. This function operates similarly to the random crop data augmentation technique with rectified image dimensions $(H_{S}, W_{S})$ and the selected crop size $(H_{C}, W_{C})$ . During rectification, information loss will be proportional to $\\left(\\frac{H_{C}}{H_{S}}, \\frac{W_{C}}{W_{S}}\\right)$ . It is empirically suggested to select a least $(H_{S}, W_{S}) = 1.8(H_{C}, W_{C})$ .", + "bbox": [ + 89, + 363, + 483, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Illustrated in Fig. 3 (LEFT), given a crop in images $I^i$ and $I^j$ , the corners $c$ are represented as", + "bbox": [ + 89, + 484, + 483, + 516 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nc = \\left( \\begin{array}{c c c c} u _ {m i n} & u _ {m i n} & u _ {m a x} & u _ {m a x} \\\\ v _ {m i n} & v _ {m a x} & v _ {m i n} & v _ {m a x} \\\\ 1 & 1 & 1 & 1 \\end{array} \\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 527, + 482, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and are mapped to rectified image space and then used to compute $(\\alpha_{min},\\alpha_{max},\\beta_{min},\\beta_{max})$ ..", + "bbox": [ + 89, + 588, + 483, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left( \\begin{array}{c} c _ {S} ^ {i} \\\\ c _ {S} ^ {j} \\end{array} \\right) = \\left( \\begin{array}{c} R _ {s} ^ {T} \\cdot (R _ {c} ^ {i}) ^ {T} \\cdot K ^ {- 1} \\cdot c \\\\ R _ {s} ^ {T} \\cdot (R _ {c} ^ {j}) ^ {T} \\cdot K ^ {- 1} \\cdot c \\end{array} \\right) \\\\ \\left( \\begin{array}{c} (\\alpha^ {i}, \\beta^ {i}) ^ {T} \\\\ (\\alpha^ {j}, \\beta^ {j}) ^ {T} \\end{array} \\right) = \\left( \\begin{array}{c} N _ {x y z \\to \\alpha \\beta} (c _ {S} ^ {i}) \\\\ N _ {x y z \\to \\alpha \\beta} (c _ {S} ^ {i}) \\end{array} \\right) \\\\ \\left( \\begin{array}{l l} \\alpha_ {m i n} & \\alpha_ {m a x} \\\\ \\beta_ {m i n} & \\beta_ {m a x} \\end{array} \\right) = \\left( \\begin{array}{l l} \\min \\left(\\alpha^ {i}, \\alpha^ {j}\\right) & \\max \\left(\\alpha^ {i}, \\alpha^ {j}\\right) \\\\ \\min \\left(\\beta^ {i}, \\beta^ {j}\\right) & \\max \\left(\\beta^ {i}, \\beta^ {j}\\right) \\end{array} \\right) \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 632, + 480, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Very importantly, it is possible for the alpha range to be around $2\\pi$ , i.e. $\\alpha_{min} < 2\\pi < \\alpha_{max}$ . When this happens, the spherical rectification of the crop will not work. To solve for this, an offset of $-\\pi$ is added to $\\alpha$ . This represents a spherical representation rotated by $\\pi$ around its $z$ axis. This is not reflected in the above equations to simplify reading. Similarly to networks trained with cropping data augmentation, the neural network is mapped over images at inference time.", + "bbox": [ + 89, + 763, + 483, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Overview of Permutation Stereo Matching", + "text_level": 1, + "bbox": [ + 511, + 90, + 875, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 4, the matching process relies on the Permutation Stereo formulation presented in [5]. It is chosen to highlight the contribution of the spherical epipolar rectification.", + "bbox": [ + 511, + 112, + 905, + 171 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} f ^ {i} = F \\left(I _ {S} ^ {i}\\right), \\quad f ^ {j} = F \\left(I _ {S} ^ {j}\\right) \\\\ C _ {m, n, o} ^ {i, j} = \\sum_ {g} f _ {m, n, g} ^ {i} \\cdot f _ {m, n, g} ^ {j}, \\text {w h e r e} C \\in \\mathbb {R} ^ {H _ {S} \\times W _ {S} \\times W _ {S}} \\\\ P ^ {i, j} = \\operatorname {N o r m} \\left(C ^ {i, j}\\right) \\quad , \\text {w h e r e} \\quad P \\in \\mathbb {R} ^ {H _ {S} \\times W _ {S} \\times W _ {S}} \\tag {11} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 181, + 910, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The normalization operation Norm is a softmax on rows and columns simultaneously and is detailed in [4]. Disparity is obtained from the permutation volume $P^{i,j}$ for the spherical images $I_S^i$ and $I_S^j$ according to [4],", + "bbox": [ + 511, + 282, + 905, + 344 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} D _ {S} ^ {i, j} = \\frac {1}{o} \\sum_ {o} P _ {m, n, o} ^ {i, j} * (o - n) \\tag {12} \\\\ D _ {S} ^ {j, i} = \\frac {1}{n} \\sum_ {n} P _ {m, n, o} ^ {i, j} * (o - n) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 354, + 903, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Depth from a Spherical Disparity Map", + "text_level": 1, + "bbox": [ + 511, + 438, + 846, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In spherical representation, given a disparity map $D_S^{i,j}$ , we solve for depth $Q_S^i$ by triangulation and then detectify it back to the respective image representations $Q^i$ . Notice that while the disparity is defined for an image pair, the depth is defined for the image itself. Here we present the triangulation operator between disparity and depth in spherical representation as $T_{D\\to Q}$", + "bbox": [ + 511, + 459, + 906, + 566 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} Q _ {S} ^ {i} (a, b) = T _ {D \\rightarrow Q} (a, b, D _ {S} ^ {i, j} (a, b)) \\\\ = (R _ {c} ^ {i T} \\cdot p _ {s} ^ {i}) | | t ^ {j} - t ^ {i} | | \\frac {\\sin (\\angle (p _ {s} ^ {i} , p _ {s} ^ {j}))}{\\sin (\\angle (p _ {s} ^ {i} , t ^ {i} - t ^ {j}))} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 578, + 893, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 539, + 637, + 584, + 648 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left( \\begin{array}{c} p _ {s} ^ {i} \\\\ p _ {s} ^ {j} \\end{array} \\right) = \\left( \\begin{array}{c} - R _ {s} \\cdot N _ {\\alpha \\beta \\to x y z} (\\alpha^ {i}, \\beta^ {i}) \\\\ - R _ {s} \\cdot N _ {\\alpha \\beta \\to x y z} (\\alpha^ {j}, \\beta^ {j}) \\end{array} \\right) \\\\ \\left( \\begin{array}{c} (\\alpha^ {i}, \\beta^ {i}) ^ {T} \\\\ (\\alpha^ {j}, \\beta^ {j}) ^ {T} \\end{array} \\right) = \\left( \\begin{array}{c} N _ {a b \\to \\alpha \\beta} (a, b) \\\\ N _ {a b \\to \\alpha \\beta} (a, b + D _ {S} ^ {i, j} (a, b)) \\end{array} \\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 654, + 836, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and the vector angle is", + "bbox": [ + 535, + 728, + 687, + 743 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\angle (g, h) = \\cos^ {- 1} \\left(\\frac {g \\cdot h}{| | g | | | | h | |}\\right) \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 746, + 903, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Training Losses", + "text_level": 1, + "bbox": [ + 511, + 805, + 671, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Supervised Setting. The supervised method is trained with the L1 Loss between predicted absolute depth $\\tilde{d}$ for the first image and ground truth absolute depth $d_{GT}$ .", + "bbox": [ + 511, + 828, + 905, + 873 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\left| \\left| \\tilde {d} - d _ {G T} \\right| \\right| _ {1} \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 648, + 883, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28929", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9f0460303845f5464228dd502e358d5f141c90339a5587a212e7f8cb96ada41a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Crop SizeRectified SizeError - Lower is better ↓Accuracy - Higher is better ↑
MethodHC × WCHS × WSAbs RelSq RelRMSERMSE logδ < 1.25δ < 1.252δ < 1.253
No CropSpherical SGM-256 × 2560.6315.224.990.540.510.690.79
Spherical SGM-768 × 7680.398.463.400.420.690.800.86
Spherical Stereo-256 × 2560.5710.094.630.560.470.660.77
Spherical Stereo-768 × 7680.449.053.460.460.650.770.83
CropSpherical SGM120 × 160256 × 2560.254.632.430.220.780.860.91
Spherical Stereo120 × 160256 × 2560.233.322.390.240.750.840.90
", + "bbox": [ + 140, + 88, + 857, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Spherical Crop ablation study on TartanAir carwelding testing set. Spherical SGM and Spherical Permutation Stereo are used for different image crop $H_{C} \\times W_{C}$ and rectified crop $H_{S} \\times W_{S}$ sizes. \"Crop\" indicates the presence of Spherical Crop and \"No Crop\" indicates an absence of Spherical Crop. Spherical Crop improves performance significantly.", + "bbox": [ + 89, + 202, + 906, + 246 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Self-Supervised Setting. The self-supervised method is trained using photometric losses on the reconstructed images. An image can be reconstructed by warping it with the permutation on its pairing image or on itself:", + "bbox": [ + 89, + 271, + 483, + 333 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nI _ {S} ^ {j ^ {\\prime}} = P \\cdot I _ {S} ^ {i} \\quad I _ {S} ^ {i ^ {\\prime}} = P ^ {\\top} \\cdot I _ {S} ^ {j} \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 340, + 480, + 368 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nI _ {S} ^ {i ^ {\\prime \\prime}} = P ^ {\\top} \\cdot P \\cdot I _ {S} ^ {i} \\qquad I _ {S} ^ {j ^ {\\prime \\prime}} = P \\cdot P ^ {\\top} \\cdot I _ {S} ^ {j}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 363, + 426, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The photometric loss between input and target images $\\mathcal{L}_{\\mathrm{Photo}}(I,I^{*})$ uses the structural similarity (SSIM)[33].", + "bbox": [ + 89, + 392, + 482, + 422 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {P h o t o}} (I, I ^ {*}) = \\frac {\\alpha}{2} (1 - \\operatorname {S S I M} (I, I ^ {*})) + (1 - \\alpha) \\| I - I ^ {*} \\| _ {1} \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 93, + 431, + 480, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The self-supervised loss function occurs between the images and their reconstructions. This is done in image representation rather than spherical representation.", + "bbox": [ + 89, + 470, + 483, + 515 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\frac {1}{4} \\left(\\mathcal {L} _ {\\text {P h o t o}} \\left(I ^ {i}, I ^ {i ^ {\\prime}}\\right) + \\mathcal {L} _ {\\text {P h o t o}} \\left(I ^ {i}, I ^ {i ^ {\\prime \\prime}}\\right) \\right. \\tag {17} \\\\ \\left. + \\mathcal {L} _ {\\text {P h o t o}} \\left(I ^ {j}, I ^ {j ^ {\\prime}}\\right) + \\mathcal {L} _ {\\text {P h o t o}} \\left(I ^ {j}, I ^ {j ^ {\\prime \\prime}}\\right)\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 523, + 480, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 588, + 223, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. The presented results evaluate absolute depth estimation on the TartanAir[32] Samples dataset and the KITTI[12] Eigen[10] split. Both datasets provide ground truth pose. The TartanAir Samples is a synthetic dataset that has 6982 images of size $480 \\times 640$ pixels split across 16 scenes of various locations from seaside towns to under the ocean with day and night lighting conditions. As a test sequence, 46 images from the carwelding sequence have been selected. The KITTI Eigen dataset is a real world dataset of driving scenes and has approximately $20k$ images for training and 697 for testing with an image size of $375 \\times 1242$ pixels. Raw KITTI images were undistorted. Error metrics are computed for absolute depth according to [11] and capped at $50\\mathrm{m}$ according to [14]. Implementation details are presented in supplementary. For comparison, a simplified version of SGM [16] is adapted to Spherical Stereo to illustrate independence to Permutation Stereo. The method is called Spherical SGM and is detailed in [5].", + "bbox": [ + 89, + 613, + 483, + 886 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b5bdc3dcff895bac65be87611b325326417699188caaa68bf5b80b1e49e46ad6.jpg", + "image_caption": [ + "Figure 5. Distortion generated on the rectified image for the different polar rectification methods for a $90^{\\circ}$ FOV camera. Planar rectification is the homography based epipolar rectification available among other things in OpenCV [3]. Cylindrical rectification is the method initially proposed in [25] and used in [6, 24, 28]. Spherical rectification increases distortion by adding an additional polar parameter. Spherical Crop applied to spherical rectification limits this weakness and leads to lowest distortion." + ], + "image_footnote": [], + "bbox": [ + 553, + 267, + 867, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Spherical Crop Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 556, + 785, + 573 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "An ablation study is performed for the Spherical Crop as a part of the Spherical Stereo and the results are presented in Tab. 1. With regards to performance, Spherical SGM leads to better accuracy because of its smoothing properties while Spherical Stereo matching has lower error. Irrespective of the method, results illustrate that the Spherical Crop operation leads to significantly improved performances. Spherical Crop not only improves performance, but also makes training the network less challenging. This can be attributed to the fact that the proposed crop significantly reduces distortion. Previous works [24, 25, 28] have identified that the rectified image must be very large to manage distortion for adequate monocular stereo matching. Mapping an image patch to its corresponding $(\\alpha, \\beta)$ range is demonstrated here to be an appropriate way to handle distortion and makes it possible to integrate polar rectification with deep stereo matching.", + "bbox": [ + 511, + 579, + 906, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Epipolar Distortion Comparison", + "text_level": 1, + "bbox": [ + 511, + 847, + 799, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Distortion during spherical rectification is measured and compared against state-of-the-art cylindrical rectification", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28930", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/af5adbf353a60bfd8ccf9b37489b87c1bd0eccc4b2c189ed8341143bc5a3b566.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Error - Lower is better ↓Accuracy - Higher is better ↑Nbr. of Params.
MethodAbs RelSq RelRMSERMSE logδ < 1.25δ < 1.252δ < 1.253
Depth Anything V2 [37]Supervised & Scaled10.331.042.630.300.380.800.9897.47M
GMDepth [36]Supervised0.241.202.550.240.600.850.957.32M
Lite-Mono 8M [43]Supervised0.230.611.410.340.700.780.818.78M
DeepSfm Rev [29]Supervised0.150.411.420.140.800.950.9916.64M
MonST3R [42]Supervised & Scaled10.150.491.300.140.790.950.98577.81M
Spherical Stereo (ours)Supervised0.100.410.980.120.920.960.980.73M
Spherical Stereo (ours)Self-Supervised0.233.322.390.240.750.840.900.73M
", + "bbox": [ + 117, + 88, + 880, + 203 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Absolute Metric Depth, Comparative results with state of the art methods. Spherical Stereo is trained with the self-supervised loss and the supervised loss. Spherical Stereo is competitive with state of the art methods. ${}^{1}$ Per-sequence scale & shift with GT depths.", + "bbox": [ + 89, + 214, + 906, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ee979d8aeafabf7e3762ed5a91f700a929c65811c61f1e5ed9cf5e5c19403f5f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 135, + 256, + 287, + 344 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a0251621fe2b46941ee32a33ca6dd10ba337633313cb38de795b6d35166acc06.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 256, + 439, + 344 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/427a5e56f1c763b2b68a9f9c65380c9f8a770347cea3406e35b6b0aec08d6a74.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 135, + 344, + 287, + 433 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1d08d66b27dc0effdcb96f5b8593e9b9640a525ac6d325f931e0a86fc520abb3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 344, + 439, + 433 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3a9b97b58469f57e8590422638c72d9e7fae99ca05e74e25f469e44115745ff7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 135, + 433, + 287, + 521 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d3bfb3bb36cfbe82b0dac31b74e646ba90325b9caee328d9e91672de2260c5df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 433, + 439, + 520 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/06b4b6df97abd749a0b0ecc9232c26503c172562a0bd366d7683a207f052f802.jpg", + "image_caption": [ + "Figure 6. Qualitative results showing the inverse of the predicted depth with error map on a TartanAir carwelding test image. Depths are rounded to reveal depth planes. Additional visualization results are available in Depth-Visualization." + ], + "image_footnote": [], + "bbox": [ + 135, + 521, + 287, + 609 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0117432f4594e3c7caf5bf17c840ad850fa66ea0a833c73b52a9c40602a4702e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 521, + 439, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "methods [6, 24, 25, 28] in Fig. 5. Distortion can be directly measured as the ratio of area occupied by the biggest and the smallest rectified pixels in the rectified image. Fig. 5 illustrates that planar rectification is not appropriate for general motion and breaks when the epipole is inside the image. As a downside to being differentiable, the proposed spherical rectification model introduces more distortion than the cylindrical rectification model. This is because there is an additional polar parameter which warps the radial direction non-linearly. That said, the proposed Spherical Crop leads to significantly lower distortion with regards to cylindrical rectification especially for geometries where the baseline direction is not completely forward. The forward baseline di", + "bbox": [ + 89, + 703, + 485, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "resection remains the most unfavorable geometry to solve for two-view monocular stereo. This result highlights that the Spherical Crop minimizes distortion during rectification.", + "bbox": [ + 511, + 258, + 906, + 305 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Depth Estimation Comparison", + "text_level": 1, + "bbox": [ + 511, + 315, + 785, + 330 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "An evaluation of Spherical Stereo with regards to absolute depth estimation is performed and results are presented in Tab. 2 in juxtaposition to state-of-the-art methods. These methods are selected as the benchmark for the two types of depth estimators presented in section 2.2. Depth Anything V2 and MonST3R are scaled per-sequence according to the best fit curve using the ground truth depths, a scaling function is illustrated in supp. Sec. 3. Other methods have been supervised on the dataset using the same splits and according to their training protocols. This is to allow for comparison on absolute depth estimation. A qualitative example of some approaches is presented in Fig. 6. Spherical Stereo outperforms other methods at absolute depth estimation. In the self-supervised setting, Spherical Stereo performs competitively with many methods that rely on ground truths. This demonstrates that the proposed Spherical Epipolar Rectification is a proper way to extend deep stereo matching to two-view absolute depth estimation.", + "bbox": [ + 511, + 338, + 908, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Natural Multi-Image Coherence", + "text_level": 1, + "bbox": [ + 511, + 621, + 797, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given the binocular nature of stereo matching, an evaluation of the behaviour with regards to the selected images at inference is presented in Tab. 3. Spherical Stereo is trained self-supervised for pairs of images $\\{i,i - 1\\}$ as a baseline. This means that 2 images, an image and its previous image, are used for stereo matching during training. Results demonstrate that there is no loss in performance when using images $\\{i,i + 1\\}$ , an image and its following image, at inference. Degradation in performance can be observed when using images that are further away in time, e.g. $\\{i,i - 3\\}$ , because of more out-of-image pixels. The multi-frame results in Tab. 3 show improvements with respect to the baseline across all selected groups of 3 images with the overall best performance when using 5 images when compared to 2 frames. This shows that using more frames always improves depth estimation performance. Because spherical epipolar rectification properly aligns epipolar lines, this al", + "bbox": [ + 511, + 643, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28931", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ae354e18444cca6be1c36c69f12de6d7756300002d1c1834ffb49092aebf8637.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Error - Lower is better ↓Accuracy - Higher is better ↑
Nb. of FramesFrame TimestepsAbs RelSq RelRMSERMSE logδ < 1.25δ < 1.252δ < 1.253
2{i,i-1}0.233.322.390.240.750.840.90
2{i,i-3}0.407.002.720.290.700.800.86
2{i,i-2}0.304.872.430.250.740.830.89
2{i,i+3}0.325.032.750.280.700.810.87
2{i,i+2}0.243.392.460.250.740.830.89
2{i,i+1}0.223.022.530.250.740.830.89
3{i,i-1,i-2}0.233.162.260.230.750.850.90
3{i,i-2,i+2}0.202.052.030.210.770.870.92
3{i,i-1,i+1}0.192.082.210.210.770.860.91
5{i,i-1,i+1,i-2,i+2}0.171.582.020.200.780.870.92
", + "bbox": [ + 158, + 88, + 841, + 241 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/aaf61b7087849fda34b797358fea5ffa202cab5de8334530749edc1bef125f07.jpg", + "table_caption": [ + "Table 3. Multi-image coherence evaluation of Spherical Stereo on TartanAir carwelding testing set. Spherical Stereo is trained self-supervised on images $\\{i,i - 1\\}$ and used for inference on various image pairs, triplets and a quintuplet. Increasing the number of images improves the performance. The method naturally handles using future timesteps when trained on previous timesteps." + ], + "table_footnote": [], + "table_body": "
Error - Lower is better ↓Accuracy - Higher is better ↑
MethodBaselineAbs RelSq RelRMSERMSE logδ < 1.25δ < 1.252δ < 1.253
Spherical SGMAll0.6213.827.270.600.420.640.74
Spherical SGM> 0.30.365.625.190.520.520.760.86
Spherical StereoAll0.496.996.620.650.390.610.73
Spherical Stereo> 0.30.374.475.470.450.470.720.84
", + "bbox": [ + 204, + 306, + 794, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Real World Navigation results on KITTI Eigen testing set. Spherical SGM and Spherical Stereo are used on frames with and without filtering the pose from the IMU. Pose is filtered by selecting baseline larger than $0.3\\mathrm{m}$ . Error decreases when baseline is increased, substantiating Spherical rectification as proper epipolar rectification.", + "bbox": [ + 88, + 393, + 906, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "lows Spherical Stereo to take advantage of added information provided by multiple images.", + "bbox": [ + 89, + 452, + 482, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Real World Navigation", + "text_level": 1, + "bbox": [ + 89, + 492, + 303, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results for absolute depth on the real world data of the KITTI Eigen split are presented in Tab. 4. They show that the proposed Spherical Stereo model solves for absolute depth in driving settings when trained without ground truth depth. KITTI Eigen is selected as a very difficult case given that ground truth pose is noisy and that motion is almost completely forward. Noisy poses prove harder to handle in small baseline cases (Fig. 7) and lead to increased error in the estimated depth. Tab. 4 and Fig. 7 show that performance improves when increasing the baseline. This confirms Spherical Stereo behaves as proper stereo matching.", + "bbox": [ + 89, + 515, + 483, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Limitations", + "text_level": 1, + "bbox": [ + 89, + 696, + 212, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Depth from parallax estimates the depth of objects using the displacement along epipolar lines with a static scene assumption. Objects undergoing independent motion break this assumption and induce unreliable depth estimates. This is a limitation of stereo that still holds for our proposed monocular stereo matching. Our method also does not solve disparity at the focus of expansion (Fig. 6) as it is degenerate. Since the location of this point is known given pose, it could be accounted for. The results are presented in this way because absence of matches is important information. At this point, the user can fill the depth map or not according to the task requirements. It is often more important to have", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e15668ef20e1b4610183cfe1bbd312c7d86a39022b789889dd20958195f0637c.jpg", + "image_caption": [ + "Figure 7. Accuracy $\\delta^3$ with respect to the ground truth baseline (the distance between the cameras) for Spherical Stereo trained self-supervised on KITTI Eigen. Average curve is in red. Accuracy improves as a function of baseline which shows that spherical rectification naturally takes advantage of larger disparity range." + ], + "image_footnote": [], + "bbox": [ + 573, + 449, + 848, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "a few accurate and trustworthy matches than have a dense depth map where the uncertainty distribution is unknown.", + "bbox": [ + 511, + 688, + 906, + 719 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 734, + 633, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents a novel spherical epipolar rectification model for solving absolute depth from two views of a monocular image sequence using deep stereo matching. This differentiable model can be effectively integrated with self-supervised deep stereo matching. The proposed Spherical Crop reduces image distortion compared to previous rectification methods. Spherical Stereo allows for depth images from pairs of views that naturally account for scale and multi-view coherence, key problems of absolute depth estimation.", + "bbox": [ + 511, + 760, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28932", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jiawang Bian, Zhichao Li, Naiyan Wang, Huangying Zhan, Chunhua Shen, Ming-Ming Cheng, and Ian Reid. Unsupervised scale-consistent depth and ego-motion learning from monocular video. Advances in neural information processing systems, 32, 2019. 2", + "[2] Jia-Wang Bian, Huangying Zhan, Naiyan Wang, Tat-Jun Chin, Chunhua Shen, and Ian Reid. Auto-rectify network for unsupervised indoor depth estimation. IEEE transactions on pattern analysis and machine intelligence, 44(12):9802-9813, 2021. 2", + "[3] Gary Bradski, Adrian Kaehler, et al. Opencv. Dr. Dobb's journal of software tools, 3(2), 2000. 6", + "[4] Pierre-André Brousseau and Sébastien Roy. A permutation model for the self-supervised stereo matching problem. In 2022 19th Conference on Robots and Vision (CRV), pages 122–131. IEEE, 2022. 3, 4, 5", + "[5] Pierre-André Brousseau and Sébastien Roy. Strong stereo features for self-supervised practical stereo matching. In BMVC, page 89, 2023. 3, 5, 6", + "[6] Raub Camaioni, Robert H Luke III, Andrew Buck, and Derek T Anderson. Epidepth: a real-time monocular dense-depth estimation pipeline using generic image rectification. In Geospatial Informatics XII, pages 101-114. SPIE, 2022. 2, 6, 7", + "[7] Yuhua Chen, Cordelia Schmid, and Cristian Sminchisescu. Self-supervised learning with geometric constraints in monocular video: Connecting flow, depth, and camera. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7063-7072, 2019. 2", + "[8] Xuelian Cheng, Yiran Zhong, Mehrtash Harandi, Yuchao Dai, Xiaojun Chang, Hongdong Li, Tom Drummond, and Zongyuan Ge. Hierarchical neural architecture search for deep stereo matching. Advances in neural information processing systems, 33:22158-22169, 2020. 3", + "[9] François Darmon and Pascal Monasse. The polar epipolar rectification. Image processing on line, 11:56-75, 2021. 2", + "[10] David Eigen and Rob Fergus. Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. In Proceedings of the IEEE international conference on computer vision, pages 2650-2658, 2015. 6", + "[11] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems, 27, 2014. 6", + "[12] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 6", + "[13] Clément Godard, Oisin Mac Aodha, and Gabriel J Brostow. Unsupervised monocular depth estimation with left-right consistency. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 270-279, 2017. 3", + "[14] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "depth estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3828-3838, 2019. 6", + "[15] Vitor Guizilini, Rares Ambrus, Dian Chen, Sergey Zakharov, and Adrien Gaidon. Multi-frame self-supervised depth with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 160-170, 2022. 3", + "[16] Heiko Hirschmuller. Accurate and efficient stereo processing by semi-global matching and mutual information. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), pages 807-814. IEEE, 2005. 3, 6", + "[17] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1621, 2021. 2", + "[18] Yunseok Lee, Kar-Ann Toh, and Sangyoun Lee. Stereo image rectification based on polar transformation. Optical Engineering, 47(8):087205-087205, 2008. 2", + "[19] Ang Li and Zejian Yuan. Occlusion aware stereo matching via cooperative unsupervised learning. In Asian Conference on Computer Vision, pages 197-213. Springer, 2018. 3", + "[20] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In 2021 International Conference on 3D Vision (3DV), pages 218-227. IEEE, 2021. 3", + "[21] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (ToG), 39(4):71-1, 2020. 2", + "[22] Kevin Matzen, Michael F Cohen, Bryce Evans, Johannes Kopf, and Richard Szeliski. Low-cost 360 stereo photography and video capture. ACM Transactions on Graphics (TOG), 36(4):1-12, 2017. 2", + "[23] Kimmo Palander and Sami S Brandt. Epipolar geometry and log-polar transform in wide baseline stereo matching. In 2008 19th International Conference on Pattern Recognition, pages 1-4. IEEE, 2008. 2", + "[24] Marc Pollefeys, Reinhard Koch, and Luc Van Gool. A simple and efficient rectification method for general motion. In Proceedings of the Seventh IEEE International Conference on Computer Vision, pages 496-501. IEEE, 1999. 2, 6, 7", + "[25] Sébastien Roy, Jean Meunier, and Ingemar J Cox. Cylindrical rectification to minimize epipolar distortion. In Proceedings of IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 393-399. IEEE, 1997. 2, 6, 7", + "[26] Libo Sun, Jia-Wang Bian, Huangying Zhan, Wei Yin, Ian Reid, and Chunhua Shen. Sc-depthv3: Robust self-supervised monocular depth estimation for dynamic scenes. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2", + "[27] Richard Szeliski. Computer vision: algorithms and applications. Springer Nature, 2022. 1", + "[28] Julien Valentin, Adarsh Kowdle, Jonathan T Barron, Neal Wadhwa, Max Dzitsiuk, Michael Schoenberg, Vivek Verma, Ambrus Csaszar, Eric Turner, Ivan Dryanovski, et al. Depth" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28933", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "from motion for smartphone ar. ACM Transactions on Graphics (ToG), 37(6):1-19, 2018. 2, 6, 7", + "[29] Jianyuan Wang, Yiran Zhong, Yuchao Dai, Stan Birchfield, Kaihao Zhang, Nikolai Smolyanskiy, and Hongdong Li. Deep two-view structure-from-motion revisited. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 8953-8962, 2021. 2, 7", + "[30] Longguang Wang, Yulan Guo, Yingqian Wang, Zhengfa Liang, Zaiping Lin, Jungang Yang, and Wei An. Parallax attention for unsupervised stereo correspondence learning. IEEE transactions on pattern analysis and machine intelligence, 2020. 3", + "[31] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 3", + "[32] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4909-4916. IEEE, 2020. 6", + "[33] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6", + "[34] Jamie Watson, Oisin Mac Aodha, Victor Prisacariu, Gabriel Brostow, and Michael Firman. The temporal opportunist: Self-supervised multi-frame monocular depth. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1164-1174, 2021. 2", + "[35] Gangwei Xu, Yun Wang, Junda Cheng, Jinhui Tang, and Xin Yang. Accurate and efficient stereo matching via attention concatenation volume. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(4):2461-2474, 2023. 3", + "[36] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, Fisher Yu, Dacheng Tao, and Andreas Geiger. Unifying flow, stereo and depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3, 7", + "[37] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 2, 7", + "[38] Nan Yang, Lukas von Stumberg, Rui Wang, and Daniel Cremers. D3vo: Deep depth, deep pose and deep uncertainty for monocular visual odometry. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1281-1292, 2020. 2", + "[39] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 2", + "[40] Zhichao Yin and Jianping Shi. Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1983-1992, 2018. 2" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Jason J Yu, Adam W Harley, and Konstantinos G Derpanis. Back to basics: Unsupervised learning of optical flow via brightness constancy and motion smoothness. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part III 14, pages 3-10. Springer, 2016. 3", + "[42] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arXiv:2410.03825, 2024. 3, 7", + "[43] Ning Zhang, Francesco Nex, George Vosselman, and Norman Kerle. Lite-mono: A lightweight cnn and transformer architecture for self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18537-18546, 2023. 2, 7", + "[44] Wang Zhao, Shaohui Liu, Yezhi Shu, and Yong-Jin Liu. Towards better generalization: Joint depth-posed learning without posenet. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9151–9161, 2020. 2", + "[45] Yiran Zhong, Yuchao Dai, and Hongdong Li. Self-supervised learning for stereo matching with self-improving ability. arXiv preprint arXiv:1709.00930, 2017. 3" + ], + "bbox": [ + 516, + 92, + 903, + 443 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28934", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Spherical Epipolar Rectification for Deep Two-View Absolute Depth Estimation/580cf2a2-db43-4692-a58d-cadb1e36cc3b_model.json b/2025/Spherical Epipolar Rectification for Deep Two-View Absolute Depth Estimation/580cf2a2-db43-4692-a58d-cadb1e36cc3b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e9855e63261d9ba7ec175f7d4040f8e4aab0743d --- /dev/null +++ b/2025/Spherical Epipolar Rectification for Deep Two-View Absolute Depth Estimation/580cf2a2-db43-4692-a58d-cadb1e36cc3b_model.json @@ -0,0 +1,2365 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.098, + 0.131, + 0.9, + 0.154 + ], + "angle": 0, + "content": "Spherical Epipolar Rectification for Deep Two-View Absolute Depth Estimation" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.182, + 0.471, + 0.216 + ], + "angle": 0, + "content": "Pierre-André Brousseau \nUniversité de Montréal" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.22, + 0.53, + 0.234 + ], + "angle": 0, + "content": "pierre-andre.brousseau@umontreal.ca" + }, + { + "type": "text", + "bbox": [ + 0.589, + 0.182, + 0.776, + 0.234 + ], + "angle": 0, + "content": "Sébastien Roy \nUniversité de Montréal \nroys@iro.umontreal.ca" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.301, + 0.486, + 0.619 + ], + "angle": 0, + "content": "Absolute depth estimation from single camera sequence of images is a relevant task given that mobile machines increasingly rely on vision to navigate. Deep learning for stereo matching has been demonstrated to improve performance for stereo rectified depth estimation but these methods require straightforward left-right camera setups. This work proposes to introduce deep stereo matching to two views of a monocular image sequence obtained from a camera in motion in a static scene. This paper introduces a novel and principled spherical epipolar rectification model, which handles all camera motions. This rectification model is differentiable and allows self-supervised deep stereo matching algorithms to compute disparity and recover depth, given known camera pose. This paper also introduces a spherical crop operation which limits rectified image size and allows for competitive absolute depth estimation performance. This results in a spherical rectification model that is demonstrated to provide metric depth and compete favorably with current state-of-the-art monocular depth estimators. The code is available at https://gitlab.com/labv3d/spherical-stereo.git." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.649, + 0.222, + 0.663 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.869 + ], + "angle": 0, + "content": "Reconstructing the world is an important task in the computer vision field. Multiple views with a calibrated camera system makes it possible to solve for metric depth passively. As humans, we capture the 3D world with a binocular camera system, our eyes, but most camera systems that capture the world as photos and videos instead rely on a single camera. This work tackles the problem of recovering absolute metric depth from two image frames of a monocular sequence with known camera pose using self-supervised deep neural networks. It will be shown that depth from two views with known pose, known as stereo depth estimation, can be extended to monocular sequences by spherical epipolar rectification." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The challenge to address with deep networks is to enable stereo depth estimation for devices where only a single cam" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.269, + 0.902, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.444, + 0.907, + 0.486 + ], + "angle": 0, + "content": "Figure 1. Comparing two-view depth estimation from plane sweep and the proposed two-view depth estimation from spherical epipolar rectification." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.507, + 0.907, + 0.719 + ], + "angle": 0, + "content": "era is available. This case occurs in systems where the cost of the second camera is prohibitive but pose can be recovered reliably (GPS, IMU, EKF). In this situation and given two views, it is worth exploring methods that can directly solve for absolute metric depth rather than scaled relative depth. Because of the known pose, an epipolar geometry links both views and reduces the task of optical flow estimation, which has a 2D search space, to a stereo matching task, which has a 1D search space. This 1D search space is expressed to deep neural networks as rectified images where the match for each pixel lies on the same row in the matching image. This process is referred to as rectification whereby images are warped to fit the same row requirement [27]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Epipolar rectification for two views of a monocular sequence differs from the standard synchronized two-view stereo camera systems because the rectification cannot be precomputed for image pairs. In the case of two-view depth estimation from a monocular sequence, the epipolar rectification must be computed according to the pose and must handle forward translation. Planar rectification, through homography warping, breaks down when the projection of the second camera center, the epipole, is inside the image. This problem can be addressed by using a rectification with a polar coordinate system, but this category of methods leads to increased deformation of the image structure, i.e. dis" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28925" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "tortion. Previously proposed methods for polar rectification have not yet been integrated to deep networks because of the constantly changing rectified image dimensions and the large distortion that significantly affects image structure. In this work, we show that when correctly handled, these distorted images lend themselves to deep neural network applications that can learn to solve for stereoscopic correspondences in monocular image pairs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.213, + 0.482, + 0.332 + ], + "angle": 0, + "content": "This paper will explore spherical epipolar rectification in the case of monocular image sequences for deep neural stereo matching methods. Many devices that have a camera already include an IMU or a GPS for other purposes, and leveraging them for depth estimation adds value at no extra hardware cost. This work proposes a significant step towards metric monocular depth estimation and the contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.334, + 0.481, + 0.378 + ], + "angle": 0, + "content": "- We propose a differentiable spherical rectification model that enables deep stereo matching for monocular image pairs." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.379, + 0.482, + 0.422 + ], + "angle": 0, + "content": "- We introduce the spherical crop operator which reduces distortion and therefore allows training of neural networks." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.424, + 0.482, + 0.468 + ], + "angle": 0, + "content": "- We demonstrate the performance of spherical rectified stereo matching for solving two-view monocular depth estimation." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.334, + 0.482, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.47, + 0.482, + 0.5 + ], + "angle": 0, + "content": "To our knowledge, this work is the first to integrate polar rectification into deep stereo matching." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.512, + 0.233, + 0.528 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.483, + 0.69 + ], + "angle": 0, + "content": "Solving depth in a static scene from two images with known camera parameters and pose is a longstanding registration problem. A wide category of methods operate on rectified stereo image pairs to solve for disparity to impressive results. These methods cannot be trivially extended to arbitrary camera geometries. To operate on arbitrary camera geometries, most methods represent the world as 3D world points and solve for depth. This work aims to present the theoretical framework to broaden the concept of stereo matching to two views of a monocular image sequences." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.698, + 0.429, + 0.714 + ], + "angle": 0, + "content": "2.1. Polar Rectification for Stereo Matching" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Planar rectification has the drawback of creating rectified images that are potentially infinitely large when the epipoles are inside the images. Cylindrical rectification [25] introduces a new rectification method that guarantees that the rectified images are bounded for any stereo geometry. This rectification is improved by parameterizing the image with polar coordinates [6, 24] which guarantees minimal image size and reduces matching ambiguity to half epipolar lines. Google's Depth from Motion [28] reveals that rectified images can be arbitrarily large and that it should be resized according to an expected maximum disparity. The angular sampling can be improved with a linear transformation [18]" + }, + { + "type": "image", + "bbox": [ + 0.586, + 0.093, + 0.837, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.222, + 0.905, + 0.278 + ], + "angle": 0, + "content": "Figure 2. Spherical Representation. Any point \\((x,y,z)\\) along a camera ray can be projected on the unit sphere and be expressed by two angles, \\(\\alpha\\) and \\(\\beta\\). An epipolar line is selected by \\(\\alpha\\), and \\(\\beta\\) is the position along that line." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.294, + 0.905, + 0.43 + ], + "angle": 0, + "content": "and the log-polar [9, 23] coordinate system can be used to reduce distortion. Some works use a variant of cylindrical epipolar rectification in the context of large FOV cameras [22] for stereo matching. Auto-Rectify [2] proposes a deep learning compatible planar epipolar rectification formulation. As it is planar, the rectification cannot handle forward translation. Its experiments therefore deliberately ignore the forward translation because it causes extremely deformed images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.906, + 0.552 + ], + "angle": 0, + "content": "Our proposed work improves upon these related works by using a generalized spherical transformation operation, by expressing position along an epipolar line as an angle, that holds when cropping to smaller regions. Spherical crop solves the problem of resolving the optimal scaling and angle intervals. This allows for control over image sizes. Our formulation is differentiable and can be readily integrated with deep neural networks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.562, + 0.868, + 0.578 + ], + "angle": 0, + "content": "2.2. Deep Depth from Two Monocular Images" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.906, + 0.795 + ], + "angle": 0, + "content": "Depth from Direct Regression. Most methods estimate depth directly with a neural network for a given frame and validate it with other frames using the camera geometry [17, 21]. Feature or image warping provides sufficient constraints to solve for depth [40, 43]. Novel loss functions on the geometry consistency [1] and the epipolar geometry [7] are proposed to isolate general motion from camera motion. Novel frameworks to visual odometry such as a photometric residual [38] or fundamental matrix estimation [44] have also been proposed. Depth Anything V2 [37] solve for depth in image sequences by building a coherent transformer latent space. Most recently, SC-DepthV3 [26] introduces a novel depth ranking loss to distill from a single-image depth estimator." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Depth from Epipolar Geometry Matching. On the other hand, some approaches use the epipolar geometry to form a cost volume for depth candidates. Differentiable homography [34, 39] warping and plane sweep, such as DeepSfm Revisited [29], can be used to map features from two camera views back to a single 3D cost volume representation linked to the reference view. This depth-discretized epipo" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28926" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.09, + 0.631, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.09, + 0.856, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.282, + 0.908, + 0.339 + ], + "angle": 0, + "content": "Figure 3. Spherical rectification on a Spherical Crop. For a given image crop, the ranges \\(\\left[\\alpha_{\\min}, \\alpha_{\\max}\\right], \\left[\\beta_{\\min}, \\beta_{\\max}\\right]\\) can be computed from the pixel coordinates of the corners using spherical epipolar geometry. This can be used to map an image crop (LEFT) of size \\(H_C \\times W_C\\) to a spherical crop (RIGHT) \\(H_S \\times W_S\\). This leads to reduced distortion with respect to mapping the full image at once. This rectification process is illustrated in Rectification-Visualization." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.354, + 0.482, + 0.476 + ], + "angle": 0, + "content": "lar sampling can be used in conjunction with attention [15] to yield cross-attention cost volumes. This has lead to models that unify the formulation for depth and optical flow, GMDepth [36], through a transformer backbone. Most recently, DUSt3R [31] have represented the 3D reconstruction task as an end-to-end 3D pointmap matching problem. This is furthered in MonST3R [42] which integrates semantic segmentation and global pointcloud optimization." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.478, + 0.483, + 0.584 + ], + "angle": 0, + "content": "As shown in Fig. 1, in comparison to the related works, our proposed approach cannot generate depth for a single frame nor does it preemptively discretize depth. It estimates stereoscopic disparity between two frames and computes absolute depth from camera geometry. This disentangles the geometry from the matching process through spherical epipolar rectification." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.601, + 0.379, + 0.618 + ], + "angle": 0, + "content": "2.3. Self-Supervised Stereo Matching" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.626, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Deep stereo matching has had significant advancements and has lead to powerful and detailed results [8, 20, 35]. In the context of this work, we focus on deep stereo methods that can also be trained without ground truth as it emphasizes the epipolar rectification capabilities. The self-supervised stereo matching task aims to generate a disparity that best explains image displacement without ground truth information. Neural networks can be trained in a self-supervised fashion by minimizing the photometric consistency [45], disparity consistency [13] and smoothing [41]. Novel cost volume representations [4, 19] and epipolar constraints [30] have lead to improved performance by guiding the disparity prediction along epipolar lines." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.825, + 0.484, + 0.9 + ], + "angle": 0, + "content": "Our proposed method does not contribute to stereo matching per se but relies on Permutation Stereo as detailed in [5] and OpenCV SGBM [16]. It applies deep stereo matching to the case of general motion by spherical epipolar rectification." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.352, + 0.787, + 0.37 + ], + "angle": 0, + "content": "3. Spherical Rectification Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.907, + 0.442 + ], + "angle": 0, + "content": "In this work, the proposed spherical rectification model presumes a pair of images taken from a rectilinear camera in motion. This section describes the mathematical basis of spherical epipolar rectification." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.455, + 0.756, + 0.47 + ], + "angle": 0, + "content": "3.1. Rectilinear Camera Model" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.478, + 0.907, + 0.554 + ], + "angle": 0, + "content": "Let us describe a camera image \\( I^{i} \\in \\mathbf{R}^{H_{I} \\times W_{I}} \\) taken at a moment \\( i \\). Spherical rectification occurs between images \\( (I^{i}, I^{j}) \\) taken by the same camera at two moments in time \\( (i, j) \\). The image formation model for a point \\( (x_{w}, y_{w}, z_{w}) \\) in the 3D world is:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.567, + 0.906, + 0.631 + ], + "angle": 0, + "content": "\\[\nI ^ {i} (x, y) = K \\cdot \\left[ R _ {c} ^ {i} \\mid t _ {c} ^ {i} \\right] \\cdot \\left( \\begin{array}{l} x _ {w} \\\\ y _ {w} \\\\ z _ {w} \\\\ 1 \\end{array} \\right) \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.642, + 0.632, + 0.783, + 0.682 + ], + "angle": 0, + "content": "\\[\nK = \\left[ \\begin{array}{c c c} f _ {x} & 0 & p _ {x} \\\\ 0 & f _ {y} & p _ {y} \\\\ 0 & 0 & 1 \\end{array} \\right]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.696, + 0.906, + 0.773 + ], + "angle": 0, + "content": "The internal parameters matrix \\(K\\) is described by \\((f_x, f_y)\\), the focal length of the camera in terms of pixel dimensions in the x-axis and the y-axis respectively, and \\((p_x, p_y)\\), the coordinates of the principal point. The pose at instant \\(i\\) is provided by rotation matrix \\(R_c^i\\) and translation vector \\(t_c^i\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.787, + 0.794, + 0.803 + ], + "angle": 0, + "content": "3.2. Spherical Geometry from Poses" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Spherical rectification requires that its main axis \\(z\\) be aligned with the direction of translation. This axis represents the epipoles and contains both camera centers. Let us designate the extrinsic parameter of a camera at a moment \\(i\\) as \\([R_c^i |t_c^i ]\\in \\mathbf{R}^{3\\times 4}\\). The spherical orientation is described as a rotation matrix \\(R_{S}\\) and is recovered from the orientation" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28927" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.09, + 0.912, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.316, + 0.908, + 0.361 + ], + "angle": 0, + "content": "Figure 4. Spherical Stereo. Crops from image \\( I^i \\) and \\( I^{i-1} \\) are rectified. A shared feature encoder learns stereo features which are then compared with the dot product. Symmetric normalization (N) is applied according to the Permutation Stereo[4] formulation and the estimated disparities are transformed into depths. The reconstructed images and the depths are then derectified back to image space." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.375, + 0.373, + 0.39 + ], + "angle": 0, + "content": "given by the translation of camera centers." + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.397, + 0.279, + 0.414 + ], + "angle": 0, + "content": "\\[\nr _ {z} = t ^ {i} - t ^ {j}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.193, + 0.418, + 0.32, + 0.435 + ], + "angle": 0, + "content": "\\[\nr _ {y} = (0, 0, 1) \\times r _ {z}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.438, + 0.483, + 0.453 + ], + "angle": 0, + "content": "\\[\nr _ {x} = r _ {y} \\times r _ {z} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.455, + 0.394, + 0.493 + ], + "angle": 0, + "content": "\\[\nR _ {S} ^ {i, j} = \\left(\\frac {r _ {x}}{| | r _ {x} | |}, \\frac {r _ {y}}{| | r _ {y} | |}, \\frac {r _ {z}}{| | r _ {z} | |}\\right) ^ {T}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.499, + 0.322, + 0.515 + ], + "angle": 0, + "content": "3.3. Spherical Representation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.521, + 0.484, + 0.582 + ], + "angle": 0, + "content": "The angles \\((\\alpha, \\beta)\\) designate a 3D point \\((x, y, z)\\) on the unit sphere, see Fig. 2, which is aligned with the spherical model principal axis according to the rotation \\(R_S^{i,j}\\). The coordinate change operation \\(N_{\\alpha \\beta \\rightarrow xyz}\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.59, + 0.483, + 0.66 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} (x, y, z) ^ {T} = N _ {\\alpha \\beta \\rightarrow x y z} (\\alpha , \\beta) \\\\ = \\left( \\begin{array}{c} \\sin (\\alpha) \\sin (\\beta) \\\\ - \\cos (\\alpha) \\sin (\\beta) \\\\ \\cos (\\beta) \\end{array} \\right) \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.666, + 0.352, + 0.682 + ], + "angle": 0, + "content": "and the inverse operation \\(N_{xyz\\to \\alpha \\beta}\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.689, + 0.483, + 0.768 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} (\\alpha , \\beta) ^ {T} = N _ {x y z \\rightarrow \\alpha \\beta} (x, y, z) \\\\ = \\left( \\begin{array}{c} \\tan^ {- 1} \\left(\\frac {x}{- y}\\right) \\\\ \\left| \\tan^ {- 1} \\left(\\frac {\\operatorname {s g n} (x) \\sqrt {x ^ {2} + y ^ {2}}}{z}\\right) \\right| \\end{array} \\right) \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.297, + 0.79 + ], + "angle": 0, + "content": "3.4. Epipolar Rectification" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Because the camera field of view is usually small relative to the full sphere, spherical rectification can yield sparse images. We thus propose to partition the input images into a set of corresponding small regions which are then rectified into shared spherical representations, see Fig. 3. Let us designate a rectified image on the unit sphere, where the pixel intensity is \\( I_{S}(a,b) \\in \\mathbf{R}^{H_{S} \\times W_{S}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.375, + 0.907, + 0.482 + ], + "angle": 0, + "content": "The indices of the rows and columns \\((a,b)\\) of a spherical representation image \\(I_{S}\\) represent uniformly sampled angles \\((\\alpha ,\\beta)\\) on the unit sphere such that \\((\\alpha ,\\beta)\\in\\) \\(\\left([\\alpha_{min},\\alpha_{max}],[\\beta_{min},\\beta_{max}]\\right)\\) in a rectified image size \\((H_S,W_S)\\), the spherical image height and width, see Fig. 3 (RIGHT). Conversion operations can be defined \\(N_{ab\\to \\alpha \\beta}\\) between the pair \\((a,b)\\) and \\((\\alpha ,\\beta)\\)." + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.489, + 0.905, + 0.552 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} (\\alpha , \\beta) ^ {T} = N _ {a b \\rightarrow \\alpha \\beta} (a, b) \\\\ = \\left( \\begin{array}{c} \\frac {a \\left(\\alpha_ {\\text {m a x}} - \\alpha_ {\\text {m i n}}\\right)}{H _ {S}} + \\alpha_ {\\text {m i n}} \\\\ \\frac {b \\left(\\beta_ {\\text {m a x}} - \\beta_ {\\text {m i n}}\\right)}{W _ {S}} + \\beta_ {\\text {m i n}} \\end{array} \\right) \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.557, + 0.906, + 0.618 + ], + "angle": 0, + "content": "The spherical epipolar rectification \\( M_S \\) is a rectification function that extracts a pixel at position \\( (u, v) \\) from an unrectified image \\( I \\) given an \\( (a, b) \\) pair. Therefore, pixel intensity in the rectified image is \\( I_S(a, b) = I(u, v) \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.618, + 0.906, + 0.648 + ], + "angle": 0, + "content": "The spherical epipolar rectification \\( M_S \\) from \\( (u, v) \\) coordinates to \\( (a, b) \\) coordinates is:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.657, + 0.728, + 0.674 + ], + "angle": 0, + "content": "\\[\n(u, v) = M _ {S} (a, b)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.676, + 0.762, + 0.694 + ], + "angle": 0, + "content": "\\[\n(\\alpha , \\beta) ^ {T} = N _ {a b \\rightarrow \\alpha \\beta} (a, b)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.697, + 0.905, + 0.716 + ], + "angle": 0, + "content": "\\[\n\\left(x _ {p}, y _ {p}, z _ {p}\\right) ^ {T} = K \\cdot R _ {c} \\cdot R _ {s} \\cdot N _ {\\alpha \\beta \\rightarrow x y z} (\\alpha , \\beta) \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.719, + 0.737, + 0.753 + ], + "angle": 0, + "content": "\\[\n(u, v) = \\left(\\frac {y _ {p}}{z _ {p}}, \\frac {x _ {p}}{z _ {p}}\\right)\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.736, + 0.775 + ], + "angle": 0, + "content": "3.5. Epipolar Derectification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.906, + 0.827 + ], + "angle": 0, + "content": "The complementary operator to epipolar rectification is the epipolar derectification. Let us define the conversion operation \\(N_{\\alpha \\beta \\to ab}\\) between the pair \\((\\alpha, \\beta)\\) and \\((a, b)\\)." + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.834, + 0.905, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} (a, b) ^ {T} = N _ {\\alpha \\beta \\rightarrow a b} (\\alpha , \\beta) \\\\ = \\binom {(\\alpha - \\alpha_ {\\min }) H _ {S} / (\\alpha_ {\\max } - \\alpha_ {\\min })} {\\left. (m o d (\\beta , \\pi) - \\beta_ {\\min }\\right) W _ {S} / \\left(\\beta_ {\\max } - \\beta_ {\\min }\\right)} \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28928" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.485, + 0.168 + ], + "angle": 0, + "content": "In the same manner, an unrectified image \\(I\\) with coordinates \\((u,v)\\) can be generated from a spherical image \\(I_S\\) such that \\(I(u,v) = I_S(a,b)\\). The spherical epipolar derectification \\(M_S^{-1}\\) operator is an operator that maps a coordinate in the camera coordinate system to a coordinate on the unit sphere." + }, + { + "type": "equation", + "bbox": [ + 0.193, + 0.193, + 0.322, + 0.21 + ], + "angle": 0, + "content": "\\[\n(a, b) = M _ {S} ^ {- 1} (u, v)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.213, + 0.482, + 0.261 + ], + "angle": 0, + "content": "\\[\n\\left( \\begin{array}{l} x _ {s} \\\\ y _ {s} \\\\ z _ {s} \\end{array} \\right) = R _ {s} ^ {T} \\cdot R _ {c} ^ {T} \\cdot K ^ {- 1} \\cdot \\left( \\begin{array}{l} u \\\\ v \\\\ 1 \\end{array} \\right) \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.264, + 0.389, + 0.281 + ], + "angle": 0, + "content": "\\[\n\\left(\\alpha , \\beta\\right) ^ {T} = N _ {x y z \\rightarrow \\alpha \\beta} \\left(x _ {s}, y _ {s}, z _ {s}\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.284, + 0.349, + 0.302 + ], + "angle": 0, + "content": "\\[\n(a, b) ^ {T} = N _ {\\alpha \\beta \\rightarrow a b} (\\alpha , \\beta)\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.316, + 0.475, + 0.333 + ], + "angle": 0, + "content": "4. Spherical Stereo for Two Monocular Views" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.341, + 0.245, + 0.358 + ], + "angle": 0, + "content": "4.1. Spherical Crop" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.484, + 0.485 + ], + "angle": 0, + "content": "Because spherical epipolar rectification introduces some distortion in the rectified image, we propose a novel spherical crop function illustrated in Fig. 3. This function operates similarly to the random crop data augmentation technique with rectified image dimensions \\((H_{S}, W_{S})\\) and the selected crop size \\((H_{C}, W_{C})\\). During rectification, information loss will be proportional to \\(\\left(\\frac{H_{C}}{H_{S}}, \\frac{W_{C}}{W_{S}}\\right)\\). It is empirically suggested to select a least \\((H_{S}, W_{S}) = 1.8(H_{C}, W_{C})\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.486, + 0.484, + 0.517 + ], + "angle": 0, + "content": "Illustrated in Fig. 3 (LEFT), given a crop in images \\( I^i \\) and \\( I^j \\), the corners \\( c \\) are represented as" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.529, + 0.483, + 0.577 + ], + "angle": 0, + "content": "\\[\nc = \\left( \\begin{array}{c c c c} u _ {m i n} & u _ {m i n} & u _ {m a x} & u _ {m a x} \\\\ v _ {m i n} & v _ {m a x} & v _ {m i n} & v _ {m a x} \\\\ 1 & 1 & 1 & 1 \\end{array} \\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.589, + 0.484, + 0.621 + ], + "angle": 0, + "content": "and are mapped to rectified image space and then used to compute \\((\\alpha_{min},\\alpha_{max},\\beta_{min},\\beta_{max})\\) .." + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.633, + 0.482, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left( \\begin{array}{c} c _ {S} ^ {i} \\\\ c _ {S} ^ {j} \\end{array} \\right) = \\left( \\begin{array}{c} R _ {s} ^ {T} \\cdot (R _ {c} ^ {i}) ^ {T} \\cdot K ^ {- 1} \\cdot c \\\\ R _ {s} ^ {T} \\cdot (R _ {c} ^ {j}) ^ {T} \\cdot K ^ {- 1} \\cdot c \\end{array} \\right) \\\\ \\left( \\begin{array}{c} (\\alpha^ {i}, \\beta^ {i}) ^ {T} \\\\ (\\alpha^ {j}, \\beta^ {j}) ^ {T} \\end{array} \\right) = \\left( \\begin{array}{c} N _ {x y z \\to \\alpha \\beta} (c _ {S} ^ {i}) \\\\ N _ {x y z \\to \\alpha \\beta} (c _ {S} ^ {i}) \\end{array} \\right) \\\\ \\left( \\begin{array}{l l} \\alpha_ {m i n} & \\alpha_ {m a x} \\\\ \\beta_ {m i n} & \\beta_ {m a x} \\end{array} \\right) = \\left( \\begin{array}{l l} \\min \\left(\\alpha^ {i}, \\alpha^ {j}\\right) & \\max \\left(\\alpha^ {i}, \\alpha^ {j}\\right) \\\\ \\min \\left(\\beta^ {i}, \\beta^ {j}\\right) & \\max \\left(\\beta^ {i}, \\beta^ {j}\\right) \\end{array} \\right) \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.9 + ], + "angle": 0, + "content": "Very importantly, it is possible for the alpha range to be around \\(2\\pi\\), i.e. \\(\\alpha_{min} < 2\\pi < \\alpha_{max}\\). When this happens, the spherical rectification of the crop will not work. To solve for this, an offset of \\(-\\pi\\) is added to \\(\\alpha\\). This represents a spherical representation rotated by \\(\\pi\\) around its \\(z\\) axis. This is not reflected in the above equations to simplify reading. Similarly to networks trained with cropping data augmentation, the neural network is mapped over images at inference time." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.091, + 0.877, + 0.108 + ], + "angle": 0, + "content": "4.2. Overview of Permutation Stereo Matching" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.906, + 0.172 + ], + "angle": 0, + "content": "As shown in Fig. 4, the matching process relies on the Permutation Stereo formulation presented in [5]. It is chosen to highlight the contribution of the spherical epipolar rectification." + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.182, + 0.911, + 0.272 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} f ^ {i} = F \\left(I _ {S} ^ {i}\\right), \\quad f ^ {j} = F \\left(I _ {S} ^ {j}\\right) \\\\ C _ {m, n, o} ^ {i, j} = \\sum_ {g} f _ {m, n, g} ^ {i} \\cdot f _ {m, n, g} ^ {j}, \\text {w h e r e} C \\in \\mathbb {R} ^ {H _ {S} \\times W _ {S} \\times W _ {S}} \\\\ P ^ {i, j} = \\operatorname {N o r m} \\left(C ^ {i, j}\\right) \\quad , \\text {w h e r e} \\quad P \\in \\mathbb {R} ^ {H _ {S} \\times W _ {S} \\times W _ {S}} \\tag {11} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.283, + 0.906, + 0.345 + ], + "angle": 0, + "content": "The normalization operation Norm is a softmax on rows and columns simultaneously and is detailed in [4]. Disparity is obtained from the permutation volume \\( P^{i,j} \\) for the spherical images \\( I_S^i \\) and \\( I_S^j \\) according to [4]," + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.355, + 0.905, + 0.427 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} D _ {S} ^ {i, j} = \\frac {1}{o} \\sum_ {o} P _ {m, n, o} ^ {i, j} * (o - n) \\tag {12} \\\\ D _ {S} ^ {j, i} = \\frac {1}{n} \\sum_ {n} P _ {m, n, o} ^ {i, j} * (o - n) \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.439, + 0.847, + 0.456 + ], + "angle": 0, + "content": "4.3. Depth from a Spherical Disparity Map" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.46, + 0.907, + 0.568 + ], + "angle": 0, + "content": "In spherical representation, given a disparity map \\(D_S^{i,j}\\), we solve for depth \\(Q_S^i\\) by triangulation and then detectify it back to the respective image representations \\(Q^i\\). Notice that while the disparity is defined for an image pair, the depth is defined for the image itself. Here we present the triangulation operator between disparity and depth in spherical representation as \\(T_{D\\to Q}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.579, + 0.895, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} Q _ {S} ^ {i} (a, b) = T _ {D \\rightarrow Q} (a, b, D _ {S} ^ {i, j} (a, b)) \\\\ = (R _ {c} ^ {i T} \\cdot p _ {s} ^ {i}) | | t ^ {j} - t ^ {i} | | \\frac {\\sin (\\angle (p _ {s} ^ {i} , p _ {s} ^ {j}))}{\\sin (\\angle (p _ {s} ^ {i} , t ^ {i} - t ^ {j}))} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.54, + 0.638, + 0.585, + 0.65 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.655, + 0.838, + 0.725 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left( \\begin{array}{c} p _ {s} ^ {i} \\\\ p _ {s} ^ {j} \\end{array} \\right) = \\left( \\begin{array}{c} - R _ {s} \\cdot N _ {\\alpha \\beta \\to x y z} (\\alpha^ {i}, \\beta^ {i}) \\\\ - R _ {s} \\cdot N _ {\\alpha \\beta \\to x y z} (\\alpha^ {j}, \\beta^ {j}) \\end{array} \\right) \\\\ \\left( \\begin{array}{c} (\\alpha^ {i}, \\beta^ {i}) ^ {T} \\\\ (\\alpha^ {j}, \\beta^ {j}) ^ {T} \\end{array} \\right) = \\left( \\begin{array}{c} N _ {a b \\to \\alpha \\beta} (a, b) \\\\ N _ {a b \\to \\alpha \\beta} (a, b + D _ {S} ^ {i, j} (a, b)) \\end{array} \\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.537, + 0.729, + 0.689, + 0.744 + ], + "angle": 0, + "content": "and the vector angle is" + }, + { + "type": "equation", + "bbox": [ + 0.562, + 0.747, + 0.905, + 0.794 + ], + "angle": 0, + "content": "\\[\n\\angle (g, h) = \\cos^ {- 1} \\left(\\frac {g \\cdot h}{| | g | | | | h | |}\\right) \\tag {13}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.806, + 0.673, + 0.822 + ], + "angle": 0, + "content": "4.4. Training Losses" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.829, + 0.906, + 0.874 + ], + "angle": 0, + "content": "Supervised Setting. The supervised method is trained with the L1 Loss between predicted absolute depth \\(\\tilde{d}\\) for the first image and ground truth absolute depth \\(d_{GT}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.65, + 0.885, + 0.905, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\left| \\left| \\tilde {d} - d _ {G T} \\right| \\right| _ {1} \\tag {14}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28929" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.142, + 0.089, + 0.858, + 0.193 + ], + "angle": 0, + "content": "
Crop SizeRectified SizeError - Lower is better ↓Accuracy - Higher is better ↑
MethodHC × WCHS × WSAbs RelSq RelRMSERMSE logδ < 1.25δ < 1.252δ < 1.253
No CropSpherical SGM-256 × 2560.6315.224.990.540.510.690.79
Spherical SGM-768 × 7680.398.463.400.420.690.800.86
Spherical Stereo-256 × 2560.5710.094.630.560.470.660.77
Spherical Stereo-768 × 7680.449.053.460.460.650.770.83
CropSpherical SGM120 × 160256 × 2560.254.632.430.220.780.860.91
Spherical Stereo120 × 160256 × 2560.233.322.390.240.750.840.90
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.203, + 0.908, + 0.247 + ], + "angle": 0, + "content": "Table 1. Spherical Crop ablation study on TartanAir carwelding testing set. Spherical SGM and Spherical Permutation Stereo are used for different image crop \\(H_{C} \\times W_{C}\\) and rectified crop \\(H_{S} \\times W_{S}\\) sizes. \"Crop\" indicates the presence of Spherical Crop and \"No Crop\" indicates an absence of Spherical Crop. Spherical Crop improves performance significantly." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.272, + 0.485, + 0.334 + ], + "angle": 0, + "content": "Self-Supervised Setting. The self-supervised method is trained using photometric losses on the reconstructed images. An image can be reconstructed by warping it with the permutation on its pairing image or on itself:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.341, + 0.482, + 0.369 + ], + "angle": 0, + "content": "\\[\nI _ {S} ^ {j ^ {\\prime}} = P \\cdot I _ {S} ^ {i} \\quad I _ {S} ^ {i ^ {\\prime}} = P ^ {\\top} \\cdot I _ {S} ^ {j} \\tag {15}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.364, + 0.427, + 0.385 + ], + "angle": 0, + "content": "\\[\nI _ {S} ^ {i ^ {\\prime \\prime}} = P ^ {\\top} \\cdot P \\cdot I _ {S} ^ {i} \\qquad I _ {S} ^ {j ^ {\\prime \\prime}} = P \\cdot P ^ {\\top} \\cdot I _ {S} ^ {j}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.393, + 0.483, + 0.424 + ], + "angle": 0, + "content": "The photometric loss between input and target images \\(\\mathcal{L}_{\\mathrm{Photo}}(I,I^{*})\\) uses the structural similarity (SSIM)[33]." + }, + { + "type": "equation", + "bbox": [ + 0.094, + 0.432, + 0.482, + 0.47 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {P h o t o}} (I, I ^ {*}) = \\frac {\\alpha}{2} (1 - \\operatorname {S S I M} (I, I ^ {*})) + (1 - \\alpha) \\| I - I ^ {*} \\| _ {1} \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.471, + 0.484, + 0.516 + ], + "angle": 0, + "content": "The self-supervised loss function occurs between the images and their reconstructions. This is done in image representation rather than spherical representation." + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.524, + 0.482, + 0.581 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\frac {1}{4} \\left(\\mathcal {L} _ {\\text {P h o t o}} \\left(I ^ {i}, I ^ {i ^ {\\prime}}\\right) + \\mathcal {L} _ {\\text {P h o t o}} \\left(I ^ {i}, I ^ {i ^ {\\prime \\prime}}\\right) \\right. \\tag {17} \\\\ \\left. + \\mathcal {L} _ {\\text {P h o t o}} \\left(I ^ {j}, I ^ {j ^ {\\prime}}\\right) + \\mathcal {L} _ {\\text {P h o t o}} \\left(I ^ {j}, I ^ {j ^ {\\prime \\prime}}\\right)\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.589, + 0.224, + 0.606 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.887 + ], + "angle": 0, + "content": "Datasets. The presented results evaluate absolute depth estimation on the TartanAir[32] Samples dataset and the KITTI[12] Eigen[10] split. Both datasets provide ground truth pose. The TartanAir Samples is a synthetic dataset that has 6982 images of size \\(480 \\times 640\\) pixels split across 16 scenes of various locations from seaside towns to under the ocean with day and night lighting conditions. As a test sequence, 46 images from the carwelding sequence have been selected. The KITTI Eigen dataset is a real world dataset of driving scenes and has approximately \\(20k\\) images for training and 697 for testing with an image size of \\(375 \\times 1242\\) pixels. Raw KITTI images were undistorted. Error metrics are computed for absolute depth according to [11] and capped at \\(50\\mathrm{m}\\) according to [14]. Implementation details are presented in supplementary. For comparison, a simplified version of SGM [16] is adapted to Spherical Stereo to illustrate independence to Permutation Stereo. The method is called Spherical SGM and is detailed in [5]." + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.268, + 0.868, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.437, + 0.908, + 0.547 + ], + "angle": 0, + "content": "Figure 5. Distortion generated on the rectified image for the different polar rectification methods for a \\(90^{\\circ}\\) FOV camera. Planar rectification is the homography based epipolar rectification available among other things in OpenCV [3]. Cylindrical rectification is the method initially proposed in [25] and used in [6, 24, 28]. Spherical rectification increases distortion by adding an additional polar parameter. Spherical Crop applied to spherical rectification limits this weakness and leads to lowest distortion." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.558, + 0.787, + 0.574 + ], + "angle": 0, + "content": "5.1. Spherical Crop Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.907, + 0.837 + ], + "angle": 0, + "content": "An ablation study is performed for the Spherical Crop as a part of the Spherical Stereo and the results are presented in Tab. 1. With regards to performance, Spherical SGM leads to better accuracy because of its smoothing properties while Spherical Stereo matching has lower error. Irrespective of the method, results illustrate that the Spherical Crop operation leads to significantly improved performances. Spherical Crop not only improves performance, but also makes training the network less challenging. This can be attributed to the fact that the proposed crop significantly reduces distortion. Previous works [24, 25, 28] have identified that the rectified image must be very large to manage distortion for adequate monocular stereo matching. Mapping an image patch to its corresponding \\((\\alpha, \\beta)\\) range is demonstrated here to be an appropriate way to handle distortion and makes it possible to integrate polar rectification with deep stereo matching." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.8, + 0.865 + ], + "angle": 0, + "content": "5.2. Epipolar Distortion Comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Distortion during spherical rectification is measured and compared against state-of-the-art cylindrical rectification" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28930" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.119, + 0.089, + 0.881, + 0.204 + ], + "angle": 0, + "content": "
Error - Lower is better ↓Accuracy - Higher is better ↑Nbr. of Params.
MethodAbs RelSq RelRMSERMSE logδ < 1.25δ < 1.252δ < 1.253
Depth Anything V2 [37]Supervised & Scaled10.331.042.630.300.380.800.9897.47M
GMDepth [36]Supervised0.241.202.550.240.600.850.957.32M
Lite-Mono 8M [43]Supervised0.230.611.410.340.700.780.818.78M
DeepSfm Rev [29]Supervised0.150.411.420.140.800.950.9916.64M
MonST3R [42]Supervised & Scaled10.150.491.300.140.790.950.98577.81M
Spherical Stereo (ours)Supervised0.100.410.980.120.920.960.980.73M
Spherical Stereo (ours)Self-Supervised0.233.322.390.240.750.840.900.73M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.215, + 0.907, + 0.245 + ], + "angle": 0, + "content": "Table 2. Absolute Metric Depth, Comparative results with state of the art methods. Spherical Stereo is trained with the self-supervised loss and the supervised loss. Spherical Stereo is competitive with state of the art methods. \\( {}^{1} \\) Per-sequence scale & shift with GT depths." + }, + { + "type": "image", + "bbox": [ + 0.136, + 0.257, + 0.288, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.257, + 0.44, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.136, + 0.345, + 0.288, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.345, + 0.44, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.136, + 0.434, + 0.288, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.434, + 0.44, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.136, + 0.522, + 0.288, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.522, + 0.44, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.62, + 0.486, + 0.677 + ], + "angle": 0, + "content": "Figure 6. Qualitative results showing the inverse of the predicted depth with error map on a TartanAir carwelding test image. Depths are rounded to reveal depth planes. Additional visualization results are available in Depth-Visualization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.486, + 0.901 + ], + "angle": 0, + "content": "methods [6, 24, 25, 28] in Fig. 5. Distortion can be directly measured as the ratio of area occupied by the biggest and the smallest rectified pixels in the rectified image. Fig. 5 illustrates that planar rectification is not appropriate for general motion and breaks when the epipole is inside the image. As a downside to being differentiable, the proposed spherical rectification model introduces more distortion than the cylindrical rectification model. This is because there is an additional polar parameter which warps the radial direction non-linearly. That said, the proposed Spherical Crop leads to significantly lower distortion with regards to cylindrical rectification especially for geometries where the baseline direction is not completely forward. The forward baseline di" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.26, + 0.907, + 0.306 + ], + "angle": 0, + "content": "resection remains the most unfavorable geometry to solve for two-view monocular stereo. This result highlights that the Spherical Crop minimizes distortion during rectification." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.316, + 0.786, + 0.332 + ], + "angle": 0, + "content": "5.3. Depth Estimation Comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.339, + 0.909, + 0.611 + ], + "angle": 0, + "content": "An evaluation of Spherical Stereo with regards to absolute depth estimation is performed and results are presented in Tab. 2 in juxtaposition to state-of-the-art methods. These methods are selected as the benchmark for the two types of depth estimators presented in section 2.2. Depth Anything V2 and MonST3R are scaled per-sequence according to the best fit curve using the ground truth depths, a scaling function is illustrated in supp. Sec. 3. Other methods have been supervised on the dataset using the same splits and according to their training protocols. This is to allow for comparison on absolute depth estimation. A qualitative example of some approaches is presented in Fig. 6. Spherical Stereo outperforms other methods at absolute depth estimation. In the self-supervised setting, Spherical Stereo performs competitively with many methods that rely on ground truths. This demonstrates that the proposed Spherical Epipolar Rectification is a proper way to extend deep stereo matching to two-view absolute depth estimation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.622, + 0.799, + 0.637 + ], + "angle": 0, + "content": "5.4. Natural Multi-Image Coherence" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Given the binocular nature of stereo matching, an evaluation of the behaviour with regards to the selected images at inference is presented in Tab. 3. Spherical Stereo is trained self-supervised for pairs of images \\(\\{i,i - 1\\}\\) as a baseline. This means that 2 images, an image and its previous image, are used for stereo matching during training. Results demonstrate that there is no loss in performance when using images \\(\\{i,i + 1\\}\\), an image and its following image, at inference. Degradation in performance can be observed when using images that are further away in time, e.g. \\(\\{i,i - 3\\}\\), because of more out-of-image pixels. The multi-frame results in Tab. 3 show improvements with respect to the baseline across all selected groups of 3 images with the overall best performance when using 5 images when compared to 2 frames. This shows that using more frames always improves depth estimation performance. Because spherical epipolar rectification properly aligns epipolar lines, this al" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28931" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.16, + 0.089, + 0.842, + 0.242 + ], + "angle": 0, + "content": "
Error - Lower is better ↓Accuracy - Higher is better ↑
Nb. of FramesFrame TimestepsAbs RelSq RelRMSERMSE logδ < 1.25δ < 1.252δ < 1.253
2{i,i-1}0.233.322.390.240.750.840.90
2{i,i-3}0.407.002.720.290.700.800.86
2{i,i-2}0.304.872.430.250.740.830.89
2{i,i+3}0.325.032.750.280.700.810.87
2{i,i+2}0.243.392.460.250.740.830.89
2{i,i+1}0.223.022.530.250.740.830.89
3{i,i-1,i-2}0.233.162.260.230.750.850.90
3{i,i-2,i+2}0.202.052.030.210.770.870.92
3{i,i-1,i+1}0.192.082.210.210.770.860.91
5{i,i-1,i+1,i-2,i+2}0.171.582.020.200.780.870.92
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.251, + 0.908, + 0.296 + ], + "angle": 0, + "content": "Table 3. Multi-image coherence evaluation of Spherical Stereo on TartanAir carwelding testing set. Spherical Stereo is trained self-supervised on images \\(\\{i,i - 1\\}\\) and used for inference on various image pairs, triplets and a quintuplet. Increasing the number of images improves the performance. The method naturally handles using future timesteps when trained on previous timesteps." + }, + { + "type": "table", + "bbox": [ + 0.205, + 0.307, + 0.795, + 0.387 + ], + "angle": 0, + "content": "
Error - Lower is better ↓Accuracy - Higher is better ↑
MethodBaselineAbs RelSq RelRMSERMSE logδ < 1.25δ < 1.252δ < 1.253
Spherical SGMAll0.6213.827.270.600.420.640.74
Spherical SGM> 0.30.365.625.190.520.520.760.86
Spherical StereoAll0.496.996.620.650.390.610.73
Spherical Stereo> 0.30.374.475.470.450.470.720.84
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.395, + 0.908, + 0.44 + ], + "angle": 0, + "content": "Table 4. Real World Navigation results on KITTI Eigen testing set. Spherical SGM and Spherical Stereo are used on frames with and without filtering the pose from the IMU. Pose is filtered by selecting baseline larger than \\(0.3\\mathrm{m}\\). Error decreases when baseline is increased, substantiating Spherical rectification as proper epipolar rectification." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.453, + 0.483, + 0.485 + ], + "angle": 0, + "content": "lows Spherical Stereo to take advantage of added information provided by multiple images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.493, + 0.304, + 0.51 + ], + "angle": 0, + "content": "5.5. Real World Navigation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.484, + 0.683 + ], + "angle": 0, + "content": "Results for absolute depth on the real world data of the KITTI Eigen split are presented in Tab. 4. They show that the proposed Spherical Stereo model solves for absolute depth in driving settings when trained without ground truth depth. KITTI Eigen is selected as a very difficult case given that ground truth pose is noisy and that motion is almost completely forward. Noisy poses prove harder to handle in small baseline cases (Fig. 7) and lead to increased error in the estimated depth. Tab. 4 and Fig. 7 show that performance improves when increasing the baseline. This confirms Spherical Stereo behaves as proper stereo matching." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.697, + 0.214, + 0.712 + ], + "angle": 0, + "content": "6. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Depth from parallax estimates the depth of objects using the displacement along epipolar lines with a static scene assumption. Objects undergoing independent motion break this assumption and induce unreliable depth estimates. This is a limitation of stereo that still holds for our proposed monocular stereo matching. Our method also does not solve disparity at the focus of expansion (Fig. 6) as it is degenerate. Since the location of this point is known given pose, it could be accounted for. The results are presented in this way because absence of matches is important information. At this point, the user can fill the depth map or not according to the task requirements. It is often more important to have" + }, + { + "type": "image", + "bbox": [ + 0.574, + 0.45, + 0.849, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.601, + 0.907, + 0.673 + ], + "angle": 0, + "content": "Figure 7. Accuracy \\(\\delta^3\\) with respect to the ground truth baseline (the distance between the cameras) for Spherical Stereo trained self-supervised on KITTI Eigen. Average curve is in red. Accuracy improves as a function of baseline which shows that spherical rectification naturally takes advantage of larger disparity range." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.72 + ], + "angle": 0, + "content": "a few accurate and trustworthy matches than have a dense depth map where the uncertainty distribution is unknown." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.736, + 0.634, + 0.751 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.761, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This paper presents a novel spherical epipolar rectification model for solving absolute depth from two views of a monocular image sequence using deep stereo matching. This differentiable model can be effectively integrated with self-supervised deep stereo matching. The proposed Spherical Crop reduces image distortion compared to previous rectification methods. Spherical Stereo allows for depth images from pairs of views that naturally account for scale and multi-view coherence, key problems of absolute depth estimation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28932" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Jiawang Bian, Zhichao Li, Naiyan Wang, Huangying Zhan, Chunhua Shen, Ming-Ming Cheng, and Ian Reid. Unsupervised scale-consistent depth and ego-motion learning from monocular video. Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.254 + ], + "angle": 0, + "content": "[2] Jia-Wang Bian, Huangying Zhan, Naiyan Wang, Tat-Jun Chin, Chunhua Shen, and Ian Reid. Auto-rectify network for unsupervised indoor depth estimation. IEEE transactions on pattern analysis and machine intelligence, 44(12):9802-9813, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.483, + 0.284 + ], + "angle": 0, + "content": "[3] Gary Bradski, Adrian Kaehler, et al. Opencv. Dr. Dobb's journal of software tools, 3(2), 2000. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.285, + 0.483, + 0.338 + ], + "angle": 0, + "content": "[4] Pierre-André Brousseau and Sébastien Roy. A permutation model for the self-supervised stereo matching problem. In 2022 19th Conference on Robots and Vision (CRV), pages 122–131. IEEE, 2022. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.341, + 0.483, + 0.382 + ], + "angle": 0, + "content": "[5] Pierre-André Brousseau and Sébastien Roy. Strong stereo features for self-supervised practical stereo matching. In BMVC, page 89, 2023. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.383, + 0.483, + 0.451 + ], + "angle": 0, + "content": "[6] Raub Camaioni, Robert H Luke III, Andrew Buck, and Derek T Anderson. Epidepth: a real-time monocular dense-depth estimation pipeline using generic image rectification. In Geospatial Informatics XII, pages 101-114. SPIE, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.453, + 0.483, + 0.522 + ], + "angle": 0, + "content": "[7] Yuhua Chen, Cordelia Schmid, and Cristian Sminchisescu. Self-supervised learning with geometric constraints in monocular video: Connecting flow, depth, and camera. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7063-7072, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.523, + 0.483, + 0.592 + ], + "angle": 0, + "content": "[8] Xuelian Cheng, Yiran Zhong, Mehrtash Harandi, Yuchao Dai, Xiaojun Chang, Hongdong Li, Tom Drummond, and Zongyuan Ge. Hierarchical neural architecture search for deep stereo matching. Advances in neural information processing systems, 33:22158-22169, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.593, + 0.483, + 0.621 + ], + "angle": 0, + "content": "[9] François Darmon and Pascal Monasse. The polar epipolar rectification. Image processing on line, 11:56-75, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.621, + 0.483, + 0.689 + ], + "angle": 0, + "content": "[10] David Eigen and Rob Fergus. Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. In Proceedings of the IEEE international conference on computer vision, pages 2650-2658, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[11] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems, 27, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[12] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[13] Clément Godard, Oisin Mac Aodha, and Gabriel J Brostow. Unsupervised monocular depth estimation with left-right consistency. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 270-279, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "depth estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3828-3838, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[15] Vitor Guizilini, Rares Ambrus, Dian Chen, Sergey Zakharov, and Adrien Gaidon. Multi-frame self-supervised depth with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 160-170, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[16] Heiko Hirschmuller. Accurate and efficient stereo processing by semi-global matching and mutual information. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), pages 807-814. IEEE, 2005. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.907, + 0.332 + ], + "angle": 0, + "content": "[17] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1621, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[18] Yunseok Lee, Kar-Ann Toh, and Sangyoun Lee. Stereo image rectification based on polar transformation. Optical Engineering, 47(8):087205-087205, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[19] Ang Li and Zejian Yuan. Occlusion aware stereo matching via cooperative unsupervised learning. In Asian Conference on Computer Vision, pages 197-213. Springer, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[20] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In 2021 International Conference on 3D Vision (3DV), pages 218-227. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.518 + ], + "angle": 0, + "content": "[21] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (ToG), 39(4):71-1, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[22] Kevin Matzen, Michael F Cohen, Bryce Evans, Johannes Kopf, and Richard Szeliski. Low-cost 360 stereo photography and video capture. ACM Transactions on Graphics (TOG), 36(4):1-12, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[23] Kimmo Palander and Sami S Brandt. Epipolar geometry and log-polar transform in wide baseline stereo matching. In 2008 19th International Conference on Pattern Recognition, pages 1-4. IEEE, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.688 + ], + "angle": 0, + "content": "[24] Marc Pollefeys, Reinhard Koch, and Luc Van Gool. A simple and efficient rectification method for general motion. In Proceedings of the Seventh IEEE International Conference on Computer Vision, pages 496-501. IEEE, 1999. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[25] Sébastien Roy, Jean Meunier, and Ingemar J Cox. Cylindrical rectification to minimize epipolar distortion. In Proceedings of IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 393-399. IEEE, 1997. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[26] Libo Sun, Jia-Wang Bian, Huangying Zhan, Wei Yin, Ian Reid, and Chunhua Shen. Sc-depthv3: Robust self-supervised monocular depth estimation for dynamic scenes. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[27] Richard Szeliski. Computer vision: algorithms and applications. Springer Nature, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[28] Julien Valentin, Adarsh Kowdle, Jonathan T Barron, Neal Wadhwa, Max Dzitsiuk, Michael Schoenberg, Vivek Verma, Ambrus Csaszar, Eric Turner, Ivan Dryanovski, et al. Depth" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28933" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "from motion for smartphone ar. ACM Transactions on Graphics (ToG), 37(6):1-19, 2018. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[29] Jianyuan Wang, Yiran Zhong, Yuchao Dai, Stan Birchfield, Kaihao Zhang, Nikolai Smolyanskiy, and Hongdong Li. Deep two-view structure-from-motion revisited. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 8953-8962, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[30] Longguang Wang, Yulan Guo, Yingqian Wang, Zhengfa Liang, Zaiping Lin, Jungang Yang, and Wei An. Parallax attention for unsupervised stereo correspondence learning. IEEE transactions on pattern analysis and machine intelligence, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[31] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.411 + ], + "angle": 0, + "content": "[32] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4909-4916. IEEE, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.467 + ], + "angle": 0, + "content": "[33] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[34] Jamie Watson, Oisin Mac Aodha, Victor Prisacariu, Gabriel Brostow, and Michael Firman. The temporal opportunist: Self-supervised multi-frame monocular depth. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1164-1174, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[35] Gangwei Xu, Yun Wang, Junda Cheng, Jinhui Tang, and Xin Yang. Accurate and efficient stereo matching via attention concatenation volume. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(4):2461-2474, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.596, + 0.482, + 0.65 + ], + "angle": 0, + "content": "[36] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, Fisher Yu, Dacheng Tao, and Andreas Geiger. Unifying flow, stereo and depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[37] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.721, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[38] Nan Yang, Lukas von Stumberg, Rui Wang, and Daniel Cremers. D3vo: Deep depth, deep pose and deep uncertainty for monocular visual odometry. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1281-1292, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[39] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[40] Zhichao Yin and Jianping Shi. Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1983-1992, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[41] Jason J Yu, Adam W Harley, and Konstantinos G Derpanis. Back to basics: Unsupervised learning of optical flow via brightness constancy and motion smoothness. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part III 14, pages 3-10. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[42] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arXiv:2410.03825, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.248, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[43] Ning Zhang, Francesco Nex, George Vosselman, and Norman Kerle. Lite-mono: A lightweight cnn and transformer architecture for self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18537-18546, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.331, + 0.905, + 0.4 + ], + "angle": 0, + "content": "[44] Wang Zhao, Shaohui Liu, Yezhi Shu, and Yong-Jin Liu. Towards better generalization: Joint depth-posed learning without posenet. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9151–9161, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.402, + 0.905, + 0.444 + ], + "angle": 0, + "content": "[45] Yiran Zhong, Yuchao Dai, and Hongdong Li. Self-supervised learning for stereo matching with self-improving ability. arXiv preprint arXiv:1709.00930, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28934" + } + ] +] \ No newline at end of file diff --git a/2025/Spherical Epipolar Rectification for Deep Two-View Absolute Depth Estimation/580cf2a2-db43-4692-a58d-cadb1e36cc3b_origin.pdf b/2025/Spherical Epipolar Rectification for Deep Two-View Absolute Depth Estimation/580cf2a2-db43-4692-a58d-cadb1e36cc3b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2a93105b7ae0778c6f47693d97c68ad51cd2b444 --- /dev/null +++ b/2025/Spherical Epipolar Rectification for Deep Two-View Absolute Depth Estimation/580cf2a2-db43-4692-a58d-cadb1e36cc3b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:82837a7512e97e2c8b97cd704121b35c5a7deb2231f28a15a9c53d9efc690335 +size 5972734 diff --git a/2025/Spherical Epipolar Rectification for Deep Two-View Absolute Depth Estimation/full.md b/2025/Spherical Epipolar Rectification for Deep Two-View Absolute Depth Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ef9561c8729881d9bba4b4a0734eff2b6ca39fb0 --- /dev/null +++ b/2025/Spherical Epipolar Rectification for Deep Two-View Absolute Depth Estimation/full.md @@ -0,0 +1,402 @@ +# Spherical Epipolar Rectification for Deep Two-View Absolute Depth Estimation + +Pierre-André Brousseau +Université de Montréal + +pierre-andre.brousseau@umontreal.ca + +Sébastien Roy +Université de Montréal +roys@iro.umontreal.ca + +# Abstract + +Absolute depth estimation from single camera sequence of images is a relevant task given that mobile machines increasingly rely on vision to navigate. Deep learning for stereo matching has been demonstrated to improve performance for stereo rectified depth estimation but these methods require straightforward left-right camera setups. This work proposes to introduce deep stereo matching to two views of a monocular image sequence obtained from a camera in motion in a static scene. This paper introduces a novel and principled spherical epipolar rectification model, which handles all camera motions. This rectification model is differentiable and allows self-supervised deep stereo matching algorithms to compute disparity and recover depth, given known camera pose. This paper also introduces a spherical crop operation which limits rectified image size and allows for competitive absolute depth estimation performance. This results in a spherical rectification model that is demonstrated to provide metric depth and compete favorably with current state-of-the-art monocular depth estimators. The code is available at https://gitlab.com/labv3d/spherical-stereo.git. + +# 1. Introduction + +Reconstructing the world is an important task in the computer vision field. Multiple views with a calibrated camera system makes it possible to solve for metric depth passively. As humans, we capture the 3D world with a binocular camera system, our eyes, but most camera systems that capture the world as photos and videos instead rely on a single camera. This work tackles the problem of recovering absolute metric depth from two image frames of a monocular sequence with known camera pose using self-supervised deep neural networks. It will be shown that depth from two views with known pose, known as stereo depth estimation, can be extended to monocular sequences by spherical epipolar rectification. + +The challenge to address with deep networks is to enable stereo depth estimation for devices where only a single cam + +![](images/5f707861d5055841b24c0948b9a2c5483b597ed2a0185242789babd19d6e0a6e.jpg) +Figure 1. Comparing two-view depth estimation from plane sweep and the proposed two-view depth estimation from spherical epipolar rectification. + +era is available. This case occurs in systems where the cost of the second camera is prohibitive but pose can be recovered reliably (GPS, IMU, EKF). In this situation and given two views, it is worth exploring methods that can directly solve for absolute metric depth rather than scaled relative depth. Because of the known pose, an epipolar geometry links both views and reduces the task of optical flow estimation, which has a 2D search space, to a stereo matching task, which has a 1D search space. This 1D search space is expressed to deep neural networks as rectified images where the match for each pixel lies on the same row in the matching image. This process is referred to as rectification whereby images are warped to fit the same row requirement [27]. + +Epipolar rectification for two views of a monocular sequence differs from the standard synchronized two-view stereo camera systems because the rectification cannot be precomputed for image pairs. In the case of two-view depth estimation from a monocular sequence, the epipolar rectification must be computed according to the pose and must handle forward translation. Planar rectification, through homography warping, breaks down when the projection of the second camera center, the epipole, is inside the image. This problem can be addressed by using a rectification with a polar coordinate system, but this category of methods leads to increased deformation of the image structure, i.e. dis + +tortion. Previously proposed methods for polar rectification have not yet been integrated to deep networks because of the constantly changing rectified image dimensions and the large distortion that significantly affects image structure. In this work, we show that when correctly handled, these distorted images lend themselves to deep neural network applications that can learn to solve for stereoscopic correspondences in monocular image pairs. + +This paper will explore spherical epipolar rectification in the case of monocular image sequences for deep neural stereo matching methods. Many devices that have a camera already include an IMU or a GPS for other purposes, and leveraging them for depth estimation adds value at no extra hardware cost. This work proposes a significant step towards metric monocular depth estimation and the contributions of this paper are as follows: + +- We propose a differentiable spherical rectification model that enables deep stereo matching for monocular image pairs. +- We introduce the spherical crop operator which reduces distortion and therefore allows training of neural networks. +- We demonstrate the performance of spherical rectified stereo matching for solving two-view monocular depth estimation. + +To our knowledge, this work is the first to integrate polar rectification into deep stereo matching. + +# 2. Related Work + +Solving depth in a static scene from two images with known camera parameters and pose is a longstanding registration problem. A wide category of methods operate on rectified stereo image pairs to solve for disparity to impressive results. These methods cannot be trivially extended to arbitrary camera geometries. To operate on arbitrary camera geometries, most methods represent the world as 3D world points and solve for depth. This work aims to present the theoretical framework to broaden the concept of stereo matching to two views of a monocular image sequences. + +# 2.1. Polar Rectification for Stereo Matching + +Planar rectification has the drawback of creating rectified images that are potentially infinitely large when the epipoles are inside the images. Cylindrical rectification [25] introduces a new rectification method that guarantees that the rectified images are bounded for any stereo geometry. This rectification is improved by parameterizing the image with polar coordinates [6, 24] which guarantees minimal image size and reduces matching ambiguity to half epipolar lines. Google's Depth from Motion [28] reveals that rectified images can be arbitrarily large and that it should be resized according to an expected maximum disparity. The angular sampling can be improved with a linear transformation [18] + +![](images/47cc29655b7cc97029282d8d1b5ab8e372e72d7f7ad08530a274602c612c9780.jpg) +Figure 2. Spherical Representation. Any point $(x,y,z)$ along a camera ray can be projected on the unit sphere and be expressed by two angles, $\alpha$ and $\beta$ . An epipolar line is selected by $\alpha$ , and $\beta$ is the position along that line. + +and the log-polar [9, 23] coordinate system can be used to reduce distortion. Some works use a variant of cylindrical epipolar rectification in the context of large FOV cameras [22] for stereo matching. Auto-Rectify [2] proposes a deep learning compatible planar epipolar rectification formulation. As it is planar, the rectification cannot handle forward translation. Its experiments therefore deliberately ignore the forward translation because it causes extremely deformed images. + +Our proposed work improves upon these related works by using a generalized spherical transformation operation, by expressing position along an epipolar line as an angle, that holds when cropping to smaller regions. Spherical crop solves the problem of resolving the optimal scaling and angle intervals. This allows for control over image sizes. Our formulation is differentiable and can be readily integrated with deep neural networks. + +# 2.2. Deep Depth from Two Monocular Images + +Depth from Direct Regression. Most methods estimate depth directly with a neural network for a given frame and validate it with other frames using the camera geometry [17, 21]. Feature or image warping provides sufficient constraints to solve for depth [40, 43]. Novel loss functions on the geometry consistency [1] and the epipolar geometry [7] are proposed to isolate general motion from camera motion. Novel frameworks to visual odometry such as a photometric residual [38] or fundamental matrix estimation [44] have also been proposed. Depth Anything V2 [37] solve for depth in image sequences by building a coherent transformer latent space. Most recently, SC-DepthV3 [26] introduces a novel depth ranking loss to distill from a single-image depth estimator. + +Depth from Epipolar Geometry Matching. On the other hand, some approaches use the epipolar geometry to form a cost volume for depth candidates. Differentiable homography [34, 39] warping and plane sweep, such as DeepSfm Revisited [29], can be used to map features from two camera views back to a single 3D cost volume representation linked to the reference view. This depth-discretized epipo + +![](images/72cb1b0521409ba8d16c63061fee2a87d1eef3d6e7b491b8f0782b5806982cb1.jpg) +Figure 3. Spherical rectification on a Spherical Crop. For a given image crop, the ranges $\left[\alpha_{\min}, \alpha_{\max}\right], \left[\beta_{\min}, \beta_{\max}\right]$ can be computed from the pixel coordinates of the corners using spherical epipolar geometry. This can be used to map an image crop (LEFT) of size $H_C \times W_C$ to a spherical crop (RIGHT) $H_S \times W_S$ . This leads to reduced distortion with respect to mapping the full image at once. This rectification process is illustrated in Rectification-Visualization. + +![](images/4e5b66173b1a7c936c2dd49596b6000f088881c16a64f1964d68f1d29e12b161.jpg) + +lar sampling can be used in conjunction with attention [15] to yield cross-attention cost volumes. This has lead to models that unify the formulation for depth and optical flow, GMDepth [36], through a transformer backbone. Most recently, DUSt3R [31] have represented the 3D reconstruction task as an end-to-end 3D pointmap matching problem. This is furthered in MonST3R [42] which integrates semantic segmentation and global pointcloud optimization. + +As shown in Fig. 1, in comparison to the related works, our proposed approach cannot generate depth for a single frame nor does it preemptively discretize depth. It estimates stereoscopic disparity between two frames and computes absolute depth from camera geometry. This disentangles the geometry from the matching process through spherical epipolar rectification. + +# 2.3. Self-Supervised Stereo Matching + +Deep stereo matching has had significant advancements and has lead to powerful and detailed results [8, 20, 35]. In the context of this work, we focus on deep stereo methods that can also be trained without ground truth as it emphasizes the epipolar rectification capabilities. The self-supervised stereo matching task aims to generate a disparity that best explains image displacement without ground truth information. Neural networks can be trained in a self-supervised fashion by minimizing the photometric consistency [45], disparity consistency [13] and smoothing [41]. Novel cost volume representations [4, 19] and epipolar constraints [30] have lead to improved performance by guiding the disparity prediction along epipolar lines. + +Our proposed method does not contribute to stereo matching per se but relies on Permutation Stereo as detailed in [5] and OpenCV SGBM [16]. It applies deep stereo matching to the case of general motion by spherical epipolar rectification. + +# 3. Spherical Rectification Model + +In this work, the proposed spherical rectification model presumes a pair of images taken from a rectilinear camera in motion. This section describes the mathematical basis of spherical epipolar rectification. + +# 3.1. Rectilinear Camera Model + +Let us describe a camera image $I^{i} \in \mathbf{R}^{H_{I} \times W_{I}}$ taken at a moment $i$ . Spherical rectification occurs between images $(I^{i}, I^{j})$ taken by the same camera at two moments in time $(i, j)$ . The image formation model for a point $(x_{w}, y_{w}, z_{w})$ in the 3D world is: + +$$ +I ^ {i} (x, y) = K \cdot \left[ R _ {c} ^ {i} \mid t _ {c} ^ {i} \right] \cdot \left( \begin{array}{l} x _ {w} \\ y _ {w} \\ z _ {w} \\ 1 \end{array} \right) \tag {1} +$$ + +$$ +K = \left[ \begin{array}{c c c} f _ {x} & 0 & p _ {x} \\ 0 & f _ {y} & p _ {y} \\ 0 & 0 & 1 \end{array} \right] +$$ + +The internal parameters matrix $K$ is described by $(f_x, f_y)$ , the focal length of the camera in terms of pixel dimensions in the x-axis and the y-axis respectively, and $(p_x, p_y)$ , the coordinates of the principal point. The pose at instant $i$ is provided by rotation matrix $R_c^i$ and translation vector $t_c^i$ . + +# 3.2. Spherical Geometry from Poses + +Spherical rectification requires that its main axis $z$ be aligned with the direction of translation. This axis represents the epipoles and contains both camera centers. Let us designate the extrinsic parameter of a camera at a moment $i$ as $[R_c^i |t_c^i ]\in \mathbf{R}^{3\times 4}$ . The spherical orientation is described as a rotation matrix $R_{S}$ and is recovered from the orientation + +![](images/8b11457911e07e897ba9466337c9f6948f68ee07616ff2cc218f0ae71df0c62a.jpg) +Figure 4. Spherical Stereo. Crops from image $I^i$ and $I^{i-1}$ are rectified. A shared feature encoder learns stereo features which are then compared with the dot product. Symmetric normalization (N) is applied according to the Permutation Stereo[4] formulation and the estimated disparities are transformed into depths. The reconstructed images and the depths are then derectified back to image space. + +given by the translation of camera centers. + +$$ +r _ {z} = t ^ {i} - t ^ {j} +$$ + +$$ +r _ {y} = (0, 0, 1) \times r _ {z} +$$ + +$$ +r _ {x} = r _ {y} \times r _ {z} \tag {2} +$$ + +$$ +R _ {S} ^ {i, j} = \left(\frac {r _ {x}}{| | r _ {x} | |}, \frac {r _ {y}}{| | r _ {y} | |}, \frac {r _ {z}}{| | r _ {z} | |}\right) ^ {T} +$$ + +# 3.3. Spherical Representation + +The angles $(\alpha, \beta)$ designate a 3D point $(x, y, z)$ on the unit sphere, see Fig. 2, which is aligned with the spherical model principal axis according to the rotation $R_S^{i,j}$ . The coordinate change operation $N_{\alpha \beta \rightarrow xyz}$ is: + +$$ +\begin{array}{l} (x, y, z) ^ {T} = N _ {\alpha \beta \rightarrow x y z} (\alpha , \beta) \\ = \left( \begin{array}{c} \sin (\alpha) \sin (\beta) \\ - \cos (\alpha) \sin (\beta) \\ \cos (\beta) \end{array} \right) \tag {3} \\ \end{array} +$$ + +and the inverse operation $N_{xyz\to \alpha \beta}$ is: + +$$ +\begin{array}{l} (\alpha , \beta) ^ {T} = N _ {x y z \rightarrow \alpha \beta} (x, y, z) \\ = \left( \begin{array}{c} \tan^ {- 1} \left(\frac {x}{- y}\right) \\ \left| \tan^ {- 1} \left(\frac {\operatorname {s g n} (x) \sqrt {x ^ {2} + y ^ {2}}}{z}\right) \right| \end{array} \right) \tag {4} \\ \end{array} +$$ + +# 3.4. Epipolar Rectification + +Because the camera field of view is usually small relative to the full sphere, spherical rectification can yield sparse images. We thus propose to partition the input images into a set of corresponding small regions which are then rectified into shared spherical representations, see Fig. 3. Let us designate a rectified image on the unit sphere, where the pixel intensity is $I_{S}(a,b) \in \mathbf{R}^{H_{S} \times W_{S}}$ . + +The indices of the rows and columns $(a,b)$ of a spherical representation image $I_{S}$ represent uniformly sampled angles $(\alpha ,\beta)$ on the unit sphere such that $(\alpha ,\beta)\in$ $\left([\alpha_{min},\alpha_{max}],[\beta_{min},\beta_{max}]\right)$ in a rectified image size $(H_S,W_S)$ , the spherical image height and width, see Fig. 3 (RIGHT). Conversion operations can be defined $N_{ab\to \alpha \beta}$ between the pair $(a,b)$ and $(\alpha ,\beta)$ . + +$$ +\begin{array}{l} (\alpha , \beta) ^ {T} = N _ {a b \rightarrow \alpha \beta} (a, b) \\ = \left( \begin{array}{c} \frac {a \left(\alpha_ {\text {m a x}} - \alpha_ {\text {m i n}}\right)}{H _ {S}} + \alpha_ {\text {m i n}} \\ \frac {b \left(\beta_ {\text {m a x}} - \beta_ {\text {m i n}}\right)}{W _ {S}} + \beta_ {\text {m i n}} \end{array} \right) \tag {5} \\ \end{array} +$$ + +The spherical epipolar rectification $M_S$ is a rectification function that extracts a pixel at position $(u, v)$ from an unrectified image $I$ given an $(a, b)$ pair. Therefore, pixel intensity in the rectified image is $I_S(a, b) = I(u, v)$ . + +The spherical epipolar rectification $M_S$ from $(u, v)$ coordinates to $(a, b)$ coordinates is: + +$$ +(u, v) = M _ {S} (a, b) +$$ + +$$ +(\alpha , \beta) ^ {T} = N _ {a b \rightarrow \alpha \beta} (a, b) +$$ + +$$ +\left(x _ {p}, y _ {p}, z _ {p}\right) ^ {T} = K \cdot R _ {c} \cdot R _ {s} \cdot N _ {\alpha \beta \rightarrow x y z} (\alpha , \beta) \tag {6} +$$ + +$$ +(u, v) = \left(\frac {y _ {p}}{z _ {p}}, \frac {x _ {p}}{z _ {p}}\right) +$$ + +# 3.5. Epipolar Derectification + +The complementary operator to epipolar rectification is the epipolar derectification. Let us define the conversion operation $N_{\alpha \beta \to ab}$ between the pair $(\alpha, \beta)$ and $(a, b)$ . + +$$ +\begin{array}{l} (a, b) ^ {T} = N _ {\alpha \beta \rightarrow a b} (\alpha , \beta) \\ = \binom {(\alpha - \alpha_ {\min }) H _ {S} / (\alpha_ {\max } - \alpha_ {\min })} {\left. (m o d (\beta , \pi) - \beta_ {\min }\right) W _ {S} / \left(\beta_ {\max } - \beta_ {\min }\right)} \tag {7} \\ \end{array} +$$ + +In the same manner, an unrectified image $I$ with coordinates $(u,v)$ can be generated from a spherical image $I_S$ such that $I(u,v) = I_S(a,b)$ . The spherical epipolar derectification $M_S^{-1}$ operator is an operator that maps a coordinate in the camera coordinate system to a coordinate on the unit sphere. + +$$ +(a, b) = M _ {S} ^ {- 1} (u, v) +$$ + +$$ +\left( \begin{array}{l} x _ {s} \\ y _ {s} \\ z _ {s} \end{array} \right) = R _ {s} ^ {T} \cdot R _ {c} ^ {T} \cdot K ^ {- 1} \cdot \left( \begin{array}{l} u \\ v \\ 1 \end{array} \right) \tag {8} +$$ + +$$ +\left(\alpha , \beta\right) ^ {T} = N _ {x y z \rightarrow \alpha \beta} \left(x _ {s}, y _ {s}, z _ {s}\right) +$$ + +$$ +(a, b) ^ {T} = N _ {\alpha \beta \rightarrow a b} (\alpha , \beta) +$$ + +# 4. Spherical Stereo for Two Monocular Views + +# 4.1. Spherical Crop + +Because spherical epipolar rectification introduces some distortion in the rectified image, we propose a novel spherical crop function illustrated in Fig. 3. This function operates similarly to the random crop data augmentation technique with rectified image dimensions $(H_{S}, W_{S})$ and the selected crop size $(H_{C}, W_{C})$ . During rectification, information loss will be proportional to $\left(\frac{H_{C}}{H_{S}}, \frac{W_{C}}{W_{S}}\right)$ . It is empirically suggested to select a least $(H_{S}, W_{S}) = 1.8(H_{C}, W_{C})$ . + +Illustrated in Fig. 3 (LEFT), given a crop in images $I^i$ and $I^j$ , the corners $c$ are represented as + +$$ +c = \left( \begin{array}{c c c c} u _ {m i n} & u _ {m i n} & u _ {m a x} & u _ {m a x} \\ v _ {m i n} & v _ {m a x} & v _ {m i n} & v _ {m a x} \\ 1 & 1 & 1 & 1 \end{array} \right) \tag {9} +$$ + +and are mapped to rectified image space and then used to compute $(\alpha_{min},\alpha_{max},\beta_{min},\beta_{max})$ .. + +$$ +\begin{array}{l} \left( \begin{array}{c} c _ {S} ^ {i} \\ c _ {S} ^ {j} \end{array} \right) = \left( \begin{array}{c} R _ {s} ^ {T} \cdot (R _ {c} ^ {i}) ^ {T} \cdot K ^ {- 1} \cdot c \\ R _ {s} ^ {T} \cdot (R _ {c} ^ {j}) ^ {T} \cdot K ^ {- 1} \cdot c \end{array} \right) \\ \left( \begin{array}{c} (\alpha^ {i}, \beta^ {i}) ^ {T} \\ (\alpha^ {j}, \beta^ {j}) ^ {T} \end{array} \right) = \left( \begin{array}{c} N _ {x y z \to \alpha \beta} (c _ {S} ^ {i}) \\ N _ {x y z \to \alpha \beta} (c _ {S} ^ {i}) \end{array} \right) \\ \left( \begin{array}{l l} \alpha_ {m i n} & \alpha_ {m a x} \\ \beta_ {m i n} & \beta_ {m a x} \end{array} \right) = \left( \begin{array}{l l} \min \left(\alpha^ {i}, \alpha^ {j}\right) & \max \left(\alpha^ {i}, \alpha^ {j}\right) \\ \min \left(\beta^ {i}, \beta^ {j}\right) & \max \left(\beta^ {i}, \beta^ {j}\right) \end{array} \right) \tag {10} \\ \end{array} +$$ + +Very importantly, it is possible for the alpha range to be around $2\pi$ , i.e. $\alpha_{min} < 2\pi < \alpha_{max}$ . When this happens, the spherical rectification of the crop will not work. To solve for this, an offset of $-\pi$ is added to $\alpha$ . This represents a spherical representation rotated by $\pi$ around its $z$ axis. This is not reflected in the above equations to simplify reading. Similarly to networks trained with cropping data augmentation, the neural network is mapped over images at inference time. + +# 4.2. Overview of Permutation Stereo Matching + +As shown in Fig. 4, the matching process relies on the Permutation Stereo formulation presented in [5]. It is chosen to highlight the contribution of the spherical epipolar rectification. + +$$ +\begin{array}{l} f ^ {i} = F \left(I _ {S} ^ {i}\right), \quad f ^ {j} = F \left(I _ {S} ^ {j}\right) \\ C _ {m, n, o} ^ {i, j} = \sum_ {g} f _ {m, n, g} ^ {i} \cdot f _ {m, n, g} ^ {j}, \text {w h e r e} C \in \mathbb {R} ^ {H _ {S} \times W _ {S} \times W _ {S}} \\ P ^ {i, j} = \operatorname {N o r m} \left(C ^ {i, j}\right) \quad , \text {w h e r e} \quad P \in \mathbb {R} ^ {H _ {S} \times W _ {S} \times W _ {S}} \tag {11} \\ \end{array} +$$ + +The normalization operation Norm is a softmax on rows and columns simultaneously and is detailed in [4]. Disparity is obtained from the permutation volume $P^{i,j}$ for the spherical images $I_S^i$ and $I_S^j$ according to [4], + +$$ +\begin{array}{l} D _ {S} ^ {i, j} = \frac {1}{o} \sum_ {o} P _ {m, n, o} ^ {i, j} * (o - n) \tag {12} \\ D _ {S} ^ {j, i} = \frac {1}{n} \sum_ {n} P _ {m, n, o} ^ {i, j} * (o - n) \\ \end{array} +$$ + +# 4.3. Depth from a Spherical Disparity Map + +In spherical representation, given a disparity map $D_S^{i,j}$ , we solve for depth $Q_S^i$ by triangulation and then detectify it back to the respective image representations $Q^i$ . Notice that while the disparity is defined for an image pair, the depth is defined for the image itself. Here we present the triangulation operator between disparity and depth in spherical representation as $T_{D\to Q}$ + +$$ +\begin{array}{l} Q _ {S} ^ {i} (a, b) = T _ {D \rightarrow Q} (a, b, D _ {S} ^ {i, j} (a, b)) \\ = (R _ {c} ^ {i T} \cdot p _ {s} ^ {i}) | | t ^ {j} - t ^ {i} | | \frac {\sin (\angle (p _ {s} ^ {i} , p _ {s} ^ {j}))}{\sin (\angle (p _ {s} ^ {i} , t ^ {i} - t ^ {j}))} \\ \end{array} +$$ + +where + +$$ +\begin{array}{l} \left( \begin{array}{c} p _ {s} ^ {i} \\ p _ {s} ^ {j} \end{array} \right) = \left( \begin{array}{c} - R _ {s} \cdot N _ {\alpha \beta \to x y z} (\alpha^ {i}, \beta^ {i}) \\ - R _ {s} \cdot N _ {\alpha \beta \to x y z} (\alpha^ {j}, \beta^ {j}) \end{array} \right) \\ \left( \begin{array}{c} (\alpha^ {i}, \beta^ {i}) ^ {T} \\ (\alpha^ {j}, \beta^ {j}) ^ {T} \end{array} \right) = \left( \begin{array}{c} N _ {a b \to \alpha \beta} (a, b) \\ N _ {a b \to \alpha \beta} (a, b + D _ {S} ^ {i, j} (a, b)) \end{array} \right) \\ \end{array} +$$ + +and the vector angle is + +$$ +\angle (g, h) = \cos^ {- 1} \left(\frac {g \cdot h}{| | g | | | | h | |}\right) \tag {13} +$$ + +# 4.4. Training Losses + +Supervised Setting. The supervised method is trained with the L1 Loss between predicted absolute depth $\tilde{d}$ for the first image and ground truth absolute depth $d_{GT}$ . + +$$ +\mathcal {L} = \left| \left| \tilde {d} - d _ {G T} \right| \right| _ {1} \tag {14} +$$ + +
Crop SizeRectified SizeError - Lower is better ↓Accuracy - Higher is better ↑
MethodHC × WCHS × WSAbs RelSq RelRMSERMSE logδ < 1.25δ < 1.252δ < 1.253
No CropSpherical SGM-256 × 2560.6315.224.990.540.510.690.79
Spherical SGM-768 × 7680.398.463.400.420.690.800.86
Spherical Stereo-256 × 2560.5710.094.630.560.470.660.77
Spherical Stereo-768 × 7680.449.053.460.460.650.770.83
CropSpherical SGM120 × 160256 × 2560.254.632.430.220.780.860.91
Spherical Stereo120 × 160256 × 2560.233.322.390.240.750.840.90
+ +Table 1. Spherical Crop ablation study on TartanAir carwelding testing set. Spherical SGM and Spherical Permutation Stereo are used for different image crop $H_{C} \times W_{C}$ and rectified crop $H_{S} \times W_{S}$ sizes. "Crop" indicates the presence of Spherical Crop and "No Crop" indicates an absence of Spherical Crop. Spherical Crop improves performance significantly. + +Self-Supervised Setting. The self-supervised method is trained using photometric losses on the reconstructed images. An image can be reconstructed by warping it with the permutation on its pairing image or on itself: + +$$ +I _ {S} ^ {j ^ {\prime}} = P \cdot I _ {S} ^ {i} \quad I _ {S} ^ {i ^ {\prime}} = P ^ {\top} \cdot I _ {S} ^ {j} \tag {15} +$$ + +$$ +I _ {S} ^ {i ^ {\prime \prime}} = P ^ {\top} \cdot P \cdot I _ {S} ^ {i} \qquad I _ {S} ^ {j ^ {\prime \prime}} = P \cdot P ^ {\top} \cdot I _ {S} ^ {j} +$$ + +The photometric loss between input and target images $\mathcal{L}_{\mathrm{Photo}}(I,I^{*})$ uses the structural similarity (SSIM)[33]. + +$$ +\mathcal {L} _ {\text {P h o t o}} (I, I ^ {*}) = \frac {\alpha}{2} (1 - \operatorname {S S I M} (I, I ^ {*})) + (1 - \alpha) \| I - I ^ {*} \| _ {1} \tag {16} +$$ + +The self-supervised loss function occurs between the images and their reconstructions. This is done in image representation rather than spherical representation. + +$$ +\begin{array}{l} \mathcal {L} = \frac {1}{4} \left(\mathcal {L} _ {\text {P h o t o}} \left(I ^ {i}, I ^ {i ^ {\prime}}\right) + \mathcal {L} _ {\text {P h o t o}} \left(I ^ {i}, I ^ {i ^ {\prime \prime}}\right) \right. \tag {17} \\ \left. + \mathcal {L} _ {\text {P h o t o}} \left(I ^ {j}, I ^ {j ^ {\prime}}\right) + \mathcal {L} _ {\text {P h o t o}} \left(I ^ {j}, I ^ {j ^ {\prime \prime}}\right)\right) \\ \end{array} +$$ + +# 5. Experiments + +Datasets. The presented results evaluate absolute depth estimation on the TartanAir[32] Samples dataset and the KITTI[12] Eigen[10] split. Both datasets provide ground truth pose. The TartanAir Samples is a synthetic dataset that has 6982 images of size $480 \times 640$ pixels split across 16 scenes of various locations from seaside towns to under the ocean with day and night lighting conditions. As a test sequence, 46 images from the carwelding sequence have been selected. The KITTI Eigen dataset is a real world dataset of driving scenes and has approximately $20k$ images for training and 697 for testing with an image size of $375 \times 1242$ pixels. Raw KITTI images were undistorted. Error metrics are computed for absolute depth according to [11] and capped at $50\mathrm{m}$ according to [14]. Implementation details are presented in supplementary. For comparison, a simplified version of SGM [16] is adapted to Spherical Stereo to illustrate independence to Permutation Stereo. The method is called Spherical SGM and is detailed in [5]. + +![](images/b5bdc3dcff895bac65be87611b325326417699188caaa68bf5b80b1e49e46ad6.jpg) +Figure 5. Distortion generated on the rectified image for the different polar rectification methods for a $90^{\circ}$ FOV camera. Planar rectification is the homography based epipolar rectification available among other things in OpenCV [3]. Cylindrical rectification is the method initially proposed in [25] and used in [6, 24, 28]. Spherical rectification increases distortion by adding an additional polar parameter. Spherical Crop applied to spherical rectification limits this weakness and leads to lowest distortion. + +# 5.1. Spherical Crop Ablation Study + +An ablation study is performed for the Spherical Crop as a part of the Spherical Stereo and the results are presented in Tab. 1. With regards to performance, Spherical SGM leads to better accuracy because of its smoothing properties while Spherical Stereo matching has lower error. Irrespective of the method, results illustrate that the Spherical Crop operation leads to significantly improved performances. Spherical Crop not only improves performance, but also makes training the network less challenging. This can be attributed to the fact that the proposed crop significantly reduces distortion. Previous works [24, 25, 28] have identified that the rectified image must be very large to manage distortion for adequate monocular stereo matching. Mapping an image patch to its corresponding $(\alpha, \beta)$ range is demonstrated here to be an appropriate way to handle distortion and makes it possible to integrate polar rectification with deep stereo matching. + +# 5.2. Epipolar Distortion Comparison + +Distortion during spherical rectification is measured and compared against state-of-the-art cylindrical rectification + +
Error - Lower is better ↓Accuracy - Higher is better ↑Nbr. of Params.
MethodAbs RelSq RelRMSERMSE logδ < 1.25δ < 1.252δ < 1.253
Depth Anything V2 [37]Supervised & Scaled10.331.042.630.300.380.800.9897.47M
GMDepth [36]Supervised0.241.202.550.240.600.850.957.32M
Lite-Mono 8M [43]Supervised0.230.611.410.340.700.780.818.78M
DeepSfm Rev [29]Supervised0.150.411.420.140.800.950.9916.64M
MonST3R [42]Supervised & Scaled10.150.491.300.140.790.950.98577.81M
Spherical Stereo (ours)Supervised0.100.410.980.120.920.960.980.73M
Spherical Stereo (ours)Self-Supervised0.233.322.390.240.750.840.900.73M
+ +Table 2. Absolute Metric Depth, Comparative results with state of the art methods. Spherical Stereo is trained with the self-supervised loss and the supervised loss. Spherical Stereo is competitive with state of the art methods. ${}^{1}$ Per-sequence scale & shift with GT depths. + +![](images/ee979d8aeafabf7e3762ed5a91f700a929c65811c61f1e5ed9cf5e5c19403f5f.jpg) + +![](images/a0251621fe2b46941ee32a33ca6dd10ba337633313cb38de795b6d35166acc06.jpg) + +![](images/427a5e56f1c763b2b68a9f9c65380c9f8a770347cea3406e35b6b0aec08d6a74.jpg) + +![](images/1d08d66b27dc0effdcb96f5b8593e9b9640a525ac6d325f931e0a86fc520abb3.jpg) + +![](images/3a9b97b58469f57e8590422638c72d9e7fae99ca05e74e25f469e44115745ff7.jpg) + +![](images/d3bfb3bb36cfbe82b0dac31b74e646ba90325b9caee328d9e91672de2260c5df.jpg) + +![](images/06b4b6df97abd749a0b0ecc9232c26503c172562a0bd366d7683a207f052f802.jpg) +Figure 6. Qualitative results showing the inverse of the predicted depth with error map on a TartanAir carwelding test image. Depths are rounded to reveal depth planes. Additional visualization results are available in Depth-Visualization. + +![](images/0117432f4594e3c7caf5bf17c840ad850fa66ea0a833c73b52a9c40602a4702e.jpg) + +methods [6, 24, 25, 28] in Fig. 5. Distortion can be directly measured as the ratio of area occupied by the biggest and the smallest rectified pixels in the rectified image. Fig. 5 illustrates that planar rectification is not appropriate for general motion and breaks when the epipole is inside the image. As a downside to being differentiable, the proposed spherical rectification model introduces more distortion than the cylindrical rectification model. This is because there is an additional polar parameter which warps the radial direction non-linearly. That said, the proposed Spherical Crop leads to significantly lower distortion with regards to cylindrical rectification especially for geometries where the baseline direction is not completely forward. The forward baseline di + +resection remains the most unfavorable geometry to solve for two-view monocular stereo. This result highlights that the Spherical Crop minimizes distortion during rectification. + +# 5.3. Depth Estimation Comparison + +An evaluation of Spherical Stereo with regards to absolute depth estimation is performed and results are presented in Tab. 2 in juxtaposition to state-of-the-art methods. These methods are selected as the benchmark for the two types of depth estimators presented in section 2.2. Depth Anything V2 and MonST3R are scaled per-sequence according to the best fit curve using the ground truth depths, a scaling function is illustrated in supp. Sec. 3. Other methods have been supervised on the dataset using the same splits and according to their training protocols. This is to allow for comparison on absolute depth estimation. A qualitative example of some approaches is presented in Fig. 6. Spherical Stereo outperforms other methods at absolute depth estimation. In the self-supervised setting, Spherical Stereo performs competitively with many methods that rely on ground truths. This demonstrates that the proposed Spherical Epipolar Rectification is a proper way to extend deep stereo matching to two-view absolute depth estimation. + +# 5.4. Natural Multi-Image Coherence + +Given the binocular nature of stereo matching, an evaluation of the behaviour with regards to the selected images at inference is presented in Tab. 3. Spherical Stereo is trained self-supervised for pairs of images $\{i,i - 1\}$ as a baseline. This means that 2 images, an image and its previous image, are used for stereo matching during training. Results demonstrate that there is no loss in performance when using images $\{i,i + 1\}$ , an image and its following image, at inference. Degradation in performance can be observed when using images that are further away in time, e.g. $\{i,i - 3\}$ , because of more out-of-image pixels. The multi-frame results in Tab. 3 show improvements with respect to the baseline across all selected groups of 3 images with the overall best performance when using 5 images when compared to 2 frames. This shows that using more frames always improves depth estimation performance. Because spherical epipolar rectification properly aligns epipolar lines, this al + +
Error - Lower is better ↓Accuracy - Higher is better ↑
Nb. of FramesFrame TimestepsAbs RelSq RelRMSERMSE logδ < 1.25δ < 1.252δ < 1.253
2{i,i-1}0.233.322.390.240.750.840.90
2{i,i-3}0.407.002.720.290.700.800.86
2{i,i-2}0.304.872.430.250.740.830.89
2{i,i+3}0.325.032.750.280.700.810.87
2{i,i+2}0.243.392.460.250.740.830.89
2{i,i+1}0.223.022.530.250.740.830.89
3{i,i-1,i-2}0.233.162.260.230.750.850.90
3{i,i-2,i+2}0.202.052.030.210.770.870.92
3{i,i-1,i+1}0.192.082.210.210.770.860.91
5{i,i-1,i+1,i-2,i+2}0.171.582.020.200.780.870.92
+ +Table 3. Multi-image coherence evaluation of Spherical Stereo on TartanAir carwelding testing set. Spherical Stereo is trained self-supervised on images $\{i,i - 1\}$ and used for inference on various image pairs, triplets and a quintuplet. Increasing the number of images improves the performance. The method naturally handles using future timesteps when trained on previous timesteps. + +
Error - Lower is better ↓Accuracy - Higher is better ↑
MethodBaselineAbs RelSq RelRMSERMSE logδ < 1.25δ < 1.252δ < 1.253
Spherical SGMAll0.6213.827.270.600.420.640.74
Spherical SGM> 0.30.365.625.190.520.520.760.86
Spherical StereoAll0.496.996.620.650.390.610.73
Spherical Stereo> 0.30.374.475.470.450.470.720.84
+ +Table 4. Real World Navigation results on KITTI Eigen testing set. Spherical SGM and Spherical Stereo are used on frames with and without filtering the pose from the IMU. Pose is filtered by selecting baseline larger than $0.3\mathrm{m}$ . Error decreases when baseline is increased, substantiating Spherical rectification as proper epipolar rectification. + +lows Spherical Stereo to take advantage of added information provided by multiple images. + +# 5.5. Real World Navigation + +Results for absolute depth on the real world data of the KITTI Eigen split are presented in Tab. 4. They show that the proposed Spherical Stereo model solves for absolute depth in driving settings when trained without ground truth depth. KITTI Eigen is selected as a very difficult case given that ground truth pose is noisy and that motion is almost completely forward. Noisy poses prove harder to handle in small baseline cases (Fig. 7) and lead to increased error in the estimated depth. Tab. 4 and Fig. 7 show that performance improves when increasing the baseline. This confirms Spherical Stereo behaves as proper stereo matching. + +# 6. Limitations + +Depth from parallax estimates the depth of objects using the displacement along epipolar lines with a static scene assumption. Objects undergoing independent motion break this assumption and induce unreliable depth estimates. This is a limitation of stereo that still holds for our proposed monocular stereo matching. Our method also does not solve disparity at the focus of expansion (Fig. 6) as it is degenerate. Since the location of this point is known given pose, it could be accounted for. The results are presented in this way because absence of matches is important information. At this point, the user can fill the depth map or not according to the task requirements. It is often more important to have + +![](images/e15668ef20e1b4610183cfe1bbd312c7d86a39022b789889dd20958195f0637c.jpg) +Figure 7. Accuracy $\delta^3$ with respect to the ground truth baseline (the distance between the cameras) for Spherical Stereo trained self-supervised on KITTI Eigen. Average curve is in red. Accuracy improves as a function of baseline which shows that spherical rectification naturally takes advantage of larger disparity range. + +a few accurate and trustworthy matches than have a dense depth map where the uncertainty distribution is unknown. + +# 7. Conclusion + +This paper presents a novel spherical epipolar rectification model for solving absolute depth from two views of a monocular image sequence using deep stereo matching. This differentiable model can be effectively integrated with self-supervised deep stereo matching. The proposed Spherical Crop reduces image distortion compared to previous rectification methods. Spherical Stereo allows for depth images from pairs of views that naturally account for scale and multi-view coherence, key problems of absolute depth estimation. + +# References + +[1] Jiawang Bian, Zhichao Li, Naiyan Wang, Huangying Zhan, Chunhua Shen, Ming-Ming Cheng, and Ian Reid. Unsupervised scale-consistent depth and ego-motion learning from monocular video. Advances in neural information processing systems, 32, 2019. 2 +[2] Jia-Wang Bian, Huangying Zhan, Naiyan Wang, Tat-Jun Chin, Chunhua Shen, and Ian Reid. Auto-rectify network for unsupervised indoor depth estimation. IEEE transactions on pattern analysis and machine intelligence, 44(12):9802-9813, 2021. 2 +[3] Gary Bradski, Adrian Kaehler, et al. Opencv. Dr. Dobb's journal of software tools, 3(2), 2000. 6 +[4] Pierre-André Brousseau and Sébastien Roy. A permutation model for the self-supervised stereo matching problem. In 2022 19th Conference on Robots and Vision (CRV), pages 122–131. IEEE, 2022. 3, 4, 5 +[5] Pierre-André Brousseau and Sébastien Roy. Strong stereo features for self-supervised practical stereo matching. In BMVC, page 89, 2023. 3, 5, 6 +[6] Raub Camaioni, Robert H Luke III, Andrew Buck, and Derek T Anderson. Epidepth: a real-time monocular dense-depth estimation pipeline using generic image rectification. In Geospatial Informatics XII, pages 101-114. SPIE, 2022. 2, 6, 7 +[7] Yuhua Chen, Cordelia Schmid, and Cristian Sminchisescu. Self-supervised learning with geometric constraints in monocular video: Connecting flow, depth, and camera. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7063-7072, 2019. 2 +[8] Xuelian Cheng, Yiran Zhong, Mehrtash Harandi, Yuchao Dai, Xiaojun Chang, Hongdong Li, Tom Drummond, and Zongyuan Ge. Hierarchical neural architecture search for deep stereo matching. Advances in neural information processing systems, 33:22158-22169, 2020. 3 +[9] François Darmon and Pascal Monasse. The polar epipolar rectification. Image processing on line, 11:56-75, 2021. 2 +[10] David Eigen and Rob Fergus. Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. In Proceedings of the IEEE international conference on computer vision, pages 2650-2658, 2015. 6 +[11] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems, 27, 2014. 6 +[12] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 6 +[13] Clément Godard, Oisin Mac Aodha, and Gabriel J Brostow. Unsupervised monocular depth estimation with left-right consistency. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 270-279, 2017. 3 +[14] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular + +depth estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3828-3838, 2019. 6 +[15] Vitor Guizilini, Rares Ambrus, Dian Chen, Sergey Zakharov, and Adrien Gaidon. Multi-frame self-supervised depth with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 160-170, 2022. 3 +[16] Heiko Hirschmuller. Accurate and efficient stereo processing by semi-global matching and mutual information. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), pages 807-814. IEEE, 2005. 3, 6 +[17] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1621, 2021. 2 +[18] Yunseok Lee, Kar-Ann Toh, and Sangyoun Lee. Stereo image rectification based on polar transformation. Optical Engineering, 47(8):087205-087205, 2008. 2 +[19] Ang Li and Zejian Yuan. Occlusion aware stereo matching via cooperative unsupervised learning. In Asian Conference on Computer Vision, pages 197-213. Springer, 2018. 3 +[20] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In 2021 International Conference on 3D Vision (3DV), pages 218-227. IEEE, 2021. 3 +[21] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (ToG), 39(4):71-1, 2020. 2 +[22] Kevin Matzen, Michael F Cohen, Bryce Evans, Johannes Kopf, and Richard Szeliski. Low-cost 360 stereo photography and video capture. ACM Transactions on Graphics (TOG), 36(4):1-12, 2017. 2 +[23] Kimmo Palander and Sami S Brandt. Epipolar geometry and log-polar transform in wide baseline stereo matching. In 2008 19th International Conference on Pattern Recognition, pages 1-4. IEEE, 2008. 2 +[24] Marc Pollefeys, Reinhard Koch, and Luc Van Gool. A simple and efficient rectification method for general motion. In Proceedings of the Seventh IEEE International Conference on Computer Vision, pages 496-501. IEEE, 1999. 2, 6, 7 +[25] Sébastien Roy, Jean Meunier, and Ingemar J Cox. Cylindrical rectification to minimize epipolar distortion. In Proceedings of IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 393-399. IEEE, 1997. 2, 6, 7 +[26] Libo Sun, Jia-Wang Bian, Huangying Zhan, Wei Yin, Ian Reid, and Chunhua Shen. Sc-depthv3: Robust self-supervised monocular depth estimation for dynamic scenes. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2 +[27] Richard Szeliski. Computer vision: algorithms and applications. Springer Nature, 2022. 1 +[28] Julien Valentin, Adarsh Kowdle, Jonathan T Barron, Neal Wadhwa, Max Dzitsiuk, Michael Schoenberg, Vivek Verma, Ambrus Csaszar, Eric Turner, Ivan Dryanovski, et al. Depth + +from motion for smartphone ar. ACM Transactions on Graphics (ToG), 37(6):1-19, 2018. 2, 6, 7 +[29] Jianyuan Wang, Yiran Zhong, Yuchao Dai, Stan Birchfield, Kaihao Zhang, Nikolai Smolyanskiy, and Hongdong Li. Deep two-view structure-from-motion revisited. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 8953-8962, 2021. 2, 7 +[30] Longguang Wang, Yulan Guo, Yingqian Wang, Zhengfa Liang, Zaiping Lin, Jungang Yang, and Wei An. Parallax attention for unsupervised stereo correspondence learning. IEEE transactions on pattern analysis and machine intelligence, 2020. 3 +[31] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 3 +[32] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4909-4916. IEEE, 2020. 6 +[33] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6 +[34] Jamie Watson, Oisin Mac Aodha, Victor Prisacariu, Gabriel Brostow, and Michael Firman. The temporal opportunist: Self-supervised multi-frame monocular depth. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1164-1174, 2021. 2 +[35] Gangwei Xu, Yun Wang, Junda Cheng, Jinhui Tang, and Xin Yang. Accurate and efficient stereo matching via attention concatenation volume. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(4):2461-2474, 2023. 3 +[36] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, Fisher Yu, Dacheng Tao, and Andreas Geiger. Unifying flow, stereo and depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3, 7 +[37] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 2, 7 +[38] Nan Yang, Lukas von Stumberg, Rui Wang, and Daniel Cremers. D3vo: Deep depth, deep pose and deep uncertainty for monocular visual odometry. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1281-1292, 2020. 2 +[39] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 2 +[40] Zhichao Yin and Jianping Shi. Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1983-1992, 2018. 2 + +[41] Jason J Yu, Adam W Harley, and Konstantinos G Derpanis. Back to basics: Unsupervised learning of optical flow via brightness constancy and motion smoothness. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part III 14, pages 3-10. Springer, 2016. 3 +[42] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arXiv:2410.03825, 2024. 3, 7 +[43] Ning Zhang, Francesco Nex, George Vosselman, and Norman Kerle. Lite-mono: A lightweight cnn and transformer architecture for self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18537-18546, 2023. 2, 7 +[44] Wang Zhao, Shaohui Liu, Yezhi Shu, and Yong-Jin Liu. Towards better generalization: Joint depth-posed learning without posenet. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9151–9161, 2020. 2 +[45] Yiran Zhong, Yuchao Dai, and Hongdong Li. Self-supervised learning for stereo matching with self-improving ability. arXiv preprint arXiv:1709.00930, 2017. 3 \ No newline at end of file diff --git a/2025/Spherical Epipolar Rectification for Deep Two-View Absolute Depth Estimation/images.zip b/2025/Spherical Epipolar Rectification for Deep Two-View Absolute Depth Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..93dd085c7f0ea543845f97dd8d3b06764c550e82 --- /dev/null +++ b/2025/Spherical Epipolar Rectification for Deep Two-View Absolute Depth Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2006c207a044b5da0f986d0ff0535523f4b91b93da8bcffeca3c628b4811f382 +size 714056 diff --git a/2025/Spherical Epipolar Rectification for Deep Two-View Absolute Depth Estimation/layout.json b/2025/Spherical Epipolar Rectification for Deep Two-View Absolute Depth Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3b38fc0eecafe60fee63049704badb65a93e7936 --- /dev/null +++ b/2025/Spherical Epipolar Rectification for Deep Two-View Absolute Depth Estimation/layout.json @@ -0,0 +1,9192 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 103, + 550, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 103, + 550, + 121 + ], + "spans": [ + { + "bbox": [ + 59, + 103, + 550, + 121 + ], + "type": "text", + "content": "Spherical Epipolar Rectification for Deep Two-View Absolute Depth Estimation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 169, + 144, + 288, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 144, + 288, + 171 + ], + "spans": [ + { + "bbox": [ + 169, + 144, + 288, + 171 + ], + "type": "text", + "content": "Pierre-André Brousseau \nUniversité de Montréal" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 174, + 324, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 174, + 324, + 185 + ], + "spans": [ + { + "bbox": [ + 133, + 174, + 324, + 185 + ], + "type": "text", + "content": "pierre-andre.brousseau@umontreal.ca" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 360, + 144, + 474, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 144, + 474, + 185 + ], + "spans": [ + { + "bbox": [ + 360, + 144, + 474, + 185 + ], + "type": "text", + "content": "Sébastien Roy \nUniversité de Montréal \nroys@iro.umontreal.ca" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 238, + 297, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 238, + 297, + 490 + ], + "spans": [ + { + "bbox": [ + 54, + 238, + 297, + 490 + ], + "type": "text", + "content": "Absolute depth estimation from single camera sequence of images is a relevant task given that mobile machines increasingly rely on vision to navigate. Deep learning for stereo matching has been demonstrated to improve performance for stereo rectified depth estimation but these methods require straightforward left-right camera setups. This work proposes to introduce deep stereo matching to two views of a monocular image sequence obtained from a camera in motion in a static scene. This paper introduces a novel and principled spherical epipolar rectification model, which handles all camera motions. This rectification model is differentiable and allows self-supervised deep stereo matching algorithms to compute disparity and recover depth, given known camera pose. This paper also introduces a spherical crop operation which limits rectified image size and allows for competitive absolute depth estimation performance. This results in a spherical rectification model that is demonstrated to provide metric depth and compete favorably with current state-of-the-art monocular depth estimators. The code is available at https://gitlab.com/labv3d/spherical-stereo.git." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 514, + 135, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 135, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 135, + 525 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 533, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 688 + ], + "type": "text", + "content": "Reconstructing the world is an important task in the computer vision field. Multiple views with a calibrated camera system makes it possible to solve for metric depth passively. As humans, we capture the 3D world with a binocular camera system, our eyes, but most camera systems that capture the world as photos and videos instead rely on a single camera. This work tackles the problem of recovering absolute metric depth from two image frames of a monocular sequence with known camera pose using self-supervised deep neural networks. It will be shown that depth from two views with known pose, known as stereo depth estimation, can be extended to monocular sequences by spherical epipolar rectification." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "The challenge to address with deep networks is to enable stereo depth estimation for devices where only a single cam" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 319, + 213, + 552, + 342 + ], + "blocks": [ + { + "bbox": [ + 319, + 213, + 552, + 342 + ], + "lines": [ + { + "bbox": [ + 319, + 213, + 552, + 342 + ], + "spans": [ + { + "bbox": [ + 319, + 213, + 552, + 342 + ], + "type": "image", + "image_path": "5f707861d5055841b24c0948b9a2c5483b597ed2a0185242789babd19d6e0a6e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 351, + 555, + 384 + ], + "lines": [ + { + "bbox": [ + 313, + 351, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 555, + 384 + ], + "type": "text", + "content": "Figure 1. Comparing two-view depth estimation from plane sweep and the proposed two-view depth estimation from spherical epipolar rectification." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 401, + 555, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 401, + 555, + 569 + ], + "spans": [ + { + "bbox": [ + 312, + 401, + 555, + 569 + ], + "type": "text", + "content": "era is available. This case occurs in systems where the cost of the second camera is prohibitive but pose can be recovered reliably (GPS, IMU, EKF). In this situation and given two views, it is worth exploring methods that can directly solve for absolute metric depth rather than scaled relative depth. Because of the known pose, an epipolar geometry links both views and reduces the task of optical flow estimation, which has a 2D search space, to a stereo matching task, which has a 1D search space. This 1D search space is expressed to deep neural networks as rectified images where the match for each pixel lies on the same row in the matching image. This process is referred to as rectification whereby images are warped to fit the same row requirement [27]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "Epipolar rectification for two views of a monocular sequence differs from the standard synchronized two-view stereo camera systems because the rectification cannot be precomputed for image pairs. In the case of two-view depth estimation from a monocular sequence, the epipolar rectification must be computed according to the pose and must handle forward translation. Planar rectification, through homography warping, breaks down when the projection of the second camera center, the epipole, is inside the image. This problem can be addressed by using a rectification with a polar coordinate system, but this category of methods leads to increased deformation of the image structure, i.e. dis" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28925" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "tortion. Previously proposed methods for polar rectification have not yet been integrated to deep networks because of the constantly changing rectified image dimensions and the large distortion that significantly affects image structure. In this work, we show that when correctly handled, these distorted images lend themselves to deep neural network applications that can learn to solve for stereoscopic correspondences in monocular image pairs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 168, + 294, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 168, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 294, + 262 + ], + "type": "text", + "content": "This paper will explore spherical epipolar rectification in the case of monocular image sequences for deep neural stereo matching methods. Many devices that have a camera already include an IMU or a GPS for other purposes, and leveraging them for depth estimation adds value at no extra hardware cost. This work proposes a significant step towards metric monocular depth estimation and the contributions of this paper are as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 264, + 294, + 370 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 264, + 294, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 264, + 294, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 294, + 299 + ], + "type": "text", + "content": "- We propose a differentiable spherical rectification model that enables deep stereo matching for monocular image pairs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 300, + 294, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 300, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 56, + 300, + 294, + 334 + ], + "type": "text", + "content": "- We introduce the spherical crop operator which reduces distortion and therefore allows training of neural networks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 335, + 294, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 335, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 335, + 294, + 370 + ], + "type": "text", + "content": "- We demonstrate the performance of spherical rectified stereo matching for solving two-view monocular depth estimation." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 372, + 294, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 294, + 396 + ], + "type": "text", + "content": "To our knowledge, this work is the first to integrate polar rectification into deep stereo matching." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 405, + 142, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 405, + 142, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 142, + 418 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 426, + 295, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 546 + ], + "type": "text", + "content": "Solving depth in a static scene from two images with known camera parameters and pose is a longstanding registration problem. A wide category of methods operate on rectified stereo image pairs to solve for disparity to impressive results. These methods cannot be trivially extended to arbitrary camera geometries. To operate on arbitrary camera geometries, most methods represent the world as 3D world points and solve for depth. This work aims to present the theoretical framework to broaden the concept of stereo matching to two views of a monocular image sequences." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 552, + 262, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 262, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 262, + 565 + ], + "type": "text", + "content": "2.1. Polar Rectification for Stereo Matching" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": "Planar rectification has the drawback of creating rectified images that are potentially infinitely large when the epipoles are inside the images. Cylindrical rectification [25] introduces a new rectification method that guarantees that the rectified images are bounded for any stereo geometry. This rectification is improved by parameterizing the image with polar coordinates [6, 24] which guarantees minimal image size and reduces matching ambiguity to half epipolar lines. Google's Depth from Motion [28] reveals that rectified images can be arbitrarily large and that it should be resized according to an expected maximum disparity. The angular sampling can be improved with a linear transformation [18]" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 358, + 73, + 512, + 166 + ], + "blocks": [ + { + "bbox": [ + 358, + 73, + 512, + 166 + ], + "lines": [ + { + "bbox": [ + 358, + 73, + 512, + 166 + ], + "spans": [ + { + "bbox": [ + 358, + 73, + 512, + 166 + ], + "type": "image", + "image_path": "47cc29655b7cc97029282d8d1b5ab8e372e72d7f7ad08530a274602c612c9780.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 175, + 553, + 220 + ], + "lines": [ + { + "bbox": [ + 313, + 175, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 553, + 220 + ], + "type": "text", + "content": "Figure 2. Spherical Representation. Any point " + }, + { + "bbox": [ + 313, + 175, + 553, + 220 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 313, + 175, + 553, + 220 + ], + "type": "text", + "content": " along a camera ray can be projected on the unit sphere and be expressed by two angles, " + }, + { + "bbox": [ + 313, + 175, + 553, + 220 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 175, + 553, + 220 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 175, + 553, + 220 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 175, + 553, + 220 + ], + "type": "text", + "content": ". An epipolar line is selected by " + }, + { + "bbox": [ + 313, + 175, + 553, + 220 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 175, + 553, + 220 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 175, + 553, + 220 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 175, + 553, + 220 + ], + "type": "text", + "content": " is the position along that line." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 232, + 553, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 553, + 340 + ], + "type": "text", + "content": "and the log-polar [9, 23] coordinate system can be used to reduce distortion. Some works use a variant of cylindrical epipolar rectification in the context of large FOV cameras [22] for stereo matching. Auto-Rectify [2] proposes a deep learning compatible planar epipolar rectification formulation. As it is planar, the rectification cannot handle forward translation. Its experiments therefore deliberately ignore the forward translation because it causes extremely deformed images." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 341, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 554, + 437 + ], + "type": "text", + "content": "Our proposed work improves upon these related works by using a generalized spherical transformation operation, by expressing position along an epipolar line as an angle, that holds when cropping to smaller regions. Spherical crop solves the problem of resolving the optimal scaling and angle intervals. This allows for control over image sizes. Our formulation is differentiable and can be readily integrated with deep neural networks." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 445, + 531, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 531, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 531, + 457 + ], + "type": "text", + "content": "2.2. Deep Depth from Two Monocular Images" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 462, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 554, + 629 + ], + "type": "text", + "content": "Depth from Direct Regression. Most methods estimate depth directly with a neural network for a given frame and validate it with other frames using the camera geometry [17, 21]. Feature or image warping provides sufficient constraints to solve for depth [40, 43]. Novel loss functions on the geometry consistency [1] and the epipolar geometry [7] are proposed to isolate general motion from camera motion. Novel frameworks to visual odometry such as a photometric residual [38] or fundamental matrix estimation [44] have also been proposed. Depth Anything V2 [37] solve for depth in image sequences by building a coherent transformer latent space. Most recently, SC-DepthV3 [26] introduces a novel depth ranking loss to distill from a single-image depth estimator." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "Depth from Epipolar Geometry Matching. On the other hand, some approaches use the epipolar geometry to form a cost volume for depth candidates. Differentiable homography [34, 39] warping and plane sweep, such as DeepSfm Revisited [29], can be used to map features from two camera views back to a single 3D cost volume representation linked to the reference view. This depth-discretized epipo" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28926" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 71, + 386, + 210 + ], + "blocks": [ + { + "bbox": [ + 86, + 71, + 386, + 210 + ], + "lines": [ + { + "bbox": [ + 86, + 71, + 386, + 210 + ], + "spans": [ + { + "bbox": [ + 86, + 71, + 386, + 210 + ], + "type": "image", + "image_path": "72cb1b0521409ba8d16c63061fee2a87d1eef3d6e7b491b8f0782b5806982cb1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 223, + 555, + 268 + ], + "lines": [ + { + "bbox": [ + 55, + 223, + 555, + 268 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 555, + 268 + ], + "type": "text", + "content": "Figure 3. Spherical rectification on a Spherical Crop. For a given image crop, the ranges " + }, + { + "bbox": [ + 55, + 223, + 555, + 268 + ], + "type": "inline_equation", + "content": "\\left[\\alpha_{\\min}, \\alpha_{\\max}\\right], \\left[\\beta_{\\min}, \\beta_{\\max}\\right]" + }, + { + "bbox": [ + 55, + 223, + 555, + 268 + ], + "type": "text", + "content": " can be computed from the pixel coordinates of the corners using spherical epipolar geometry. This can be used to map an image crop (LEFT) of size " + }, + { + "bbox": [ + 55, + 223, + 555, + 268 + ], + "type": "inline_equation", + "content": "H_C \\times W_C" + }, + { + "bbox": [ + 55, + 223, + 555, + 268 + ], + "type": "text", + "content": " to a spherical crop (RIGHT) " + }, + { + "bbox": [ + 55, + 223, + 555, + 268 + ], + "type": "inline_equation", + "content": "H_S \\times W_S" + }, + { + "bbox": [ + 55, + 223, + 555, + 268 + ], + "type": "text", + "content": ". This leads to reduced distortion with respect to mapping the full image at once. This rectification process is illustrated in Rectification-Visualization." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 386, + 71, + 523, + 209 + ], + "blocks": [ + { + "bbox": [ + 386, + 71, + 523, + 209 + ], + "lines": [ + { + "bbox": [ + 386, + 71, + 523, + 209 + ], + "spans": [ + { + "bbox": [ + 386, + 71, + 523, + 209 + ], + "type": "image", + "image_path": "4e5b66173b1a7c936c2dd49596b6000f088881c16a64f1964d68f1d29e12b161.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 280, + 294, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 280, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 54, + 280, + 294, + 376 + ], + "type": "text", + "content": "lar sampling can be used in conjunction with attention [15] to yield cross-attention cost volumes. This has lead to models that unify the formulation for depth and optical flow, GMDepth [36], through a transformer backbone. Most recently, DUSt3R [31] have represented the 3D reconstruction task as an end-to-end 3D pointmap matching problem. This is furthered in MonST3R [42] which integrates semantic segmentation and global pointcloud optimization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 378, + 295, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 378, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 54, + 378, + 295, + 462 + ], + "type": "text", + "content": "As shown in Fig. 1, in comparison to the related works, our proposed approach cannot generate depth for a single frame nor does it preemptively discretize depth. It estimates stereoscopic disparity between two frames and computes absolute depth from camera geometry. This disentangles the geometry from the matching process through spherical epipolar rectification." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 475, + 231, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 231, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 231, + 489 + ], + "type": "text", + "content": "2.3. Self-Supervised Stereo Matching" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 495, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 495, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 54, + 495, + 295, + 652 + ], + "type": "text", + "content": "Deep stereo matching has had significant advancements and has lead to powerful and detailed results [8, 20, 35]. In the context of this work, we focus on deep stereo methods that can also be trained without ground truth as it emphasizes the epipolar rectification capabilities. The self-supervised stereo matching task aims to generate a disparity that best explains image displacement without ground truth information. Neural networks can be trained in a self-supervised fashion by minimizing the photometric consistency [45], disparity consistency [13] and smoothing [41]. Novel cost volume representations [4, 19] and epipolar constraints [30] have lead to improved performance by guiding the disparity prediction along epipolar lines." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 653, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 653, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 54, + 653, + 296, + 712 + ], + "type": "text", + "content": "Our proposed method does not contribute to stereo matching per se but relies on Permutation Stereo as detailed in [5] and OpenCV SGBM [16]. It applies deep stereo matching to the case of general motion by spherical epipolar rectification." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 278, + 481, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 481, + 293 + ], + "type": "text", + "content": "3. Spherical Rectification Model" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 300, + 555, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 350 + ], + "type": "text", + "content": "In this work, the proposed spherical rectification model presumes a pair of images taken from a rectilinear camera in motion. This section describes the mathematical basis of spherical epipolar rectification." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 360, + 462, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 462, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 462, + 372 + ], + "type": "text", + "content": "3.1. Rectilinear Camera Model" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 378, + 555, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 555, + 438 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 555, + 438 + ], + "type": "text", + "content": "Let us describe a camera image " + }, + { + "bbox": [ + 313, + 378, + 555, + 438 + ], + "type": "inline_equation", + "content": "I^{i} \\in \\mathbf{R}^{H_{I} \\times W_{I}}" + }, + { + "bbox": [ + 313, + 378, + 555, + 438 + ], + "type": "text", + "content": " taken at a moment " + }, + { + "bbox": [ + 313, + 378, + 555, + 438 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 378, + 555, + 438 + ], + "type": "text", + "content": ". Spherical rectification occurs between images " + }, + { + "bbox": [ + 313, + 378, + 555, + 438 + ], + "type": "inline_equation", + "content": "(I^{i}, I^{j})" + }, + { + "bbox": [ + 313, + 378, + 555, + 438 + ], + "type": "text", + "content": " taken by the same camera at two moments in time " + }, + { + "bbox": [ + 313, + 378, + 555, + 438 + ], + "type": "inline_equation", + "content": "(i, j)" + }, + { + "bbox": [ + 313, + 378, + 555, + 438 + ], + "type": "text", + "content": ". The image formation model for a point " + }, + { + "bbox": [ + 313, + 378, + 555, + 438 + ], + "type": "inline_equation", + "content": "(x_{w}, y_{w}, z_{w})" + }, + { + "bbox": [ + 313, + 378, + 555, + 438 + ], + "type": "text", + "content": " in the 3D world is:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 369, + 449, + 554, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 449, + 554, + 499 + ], + "spans": [ + { + "bbox": [ + 369, + 449, + 554, + 499 + ], + "type": "interline_equation", + "content": "I ^ {i} (x, y) = K \\cdot \\left[ R _ {c} ^ {i} \\mid t _ {c} ^ {i} \\right] \\cdot \\left( \\begin{array}{l} x _ {w} \\\\ y _ {w} \\\\ z _ {w} \\\\ 1 \\end{array} \\right) \\tag {1}", + "image_path": "12edb2500fbc082dda15cf4fbb129c62d9bc3cc64b8be8271bc24180f97ed03d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 392, + 500, + 479, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 500, + 479, + 540 + ], + "spans": [ + { + "bbox": [ + 392, + 500, + 479, + 540 + ], + "type": "interline_equation", + "content": "K = \\left[ \\begin{array}{c c c} f _ {x} & 0 & p _ {x} \\\\ 0 & f _ {y} & p _ {y} \\\\ 0 & 0 & 1 \\end{array} \\right]", + "image_path": "faa270881ae30774250aac15321b01b37be86e49d539561f8ee51ca1234bf878.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 551, + 554, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 551, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 551, + 554, + 612 + ], + "type": "text", + "content": "The internal parameters matrix " + }, + { + "bbox": [ + 313, + 551, + 554, + 612 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 551, + 554, + 612 + ], + "type": "text", + "content": " is described by " + }, + { + "bbox": [ + 313, + 551, + 554, + 612 + ], + "type": "inline_equation", + "content": "(f_x, f_y)" + }, + { + "bbox": [ + 313, + 551, + 554, + 612 + ], + "type": "text", + "content": ", the focal length of the camera in terms of pixel dimensions in the x-axis and the y-axis respectively, and " + }, + { + "bbox": [ + 313, + 551, + 554, + 612 + ], + "type": "inline_equation", + "content": "(p_x, p_y)" + }, + { + "bbox": [ + 313, + 551, + 554, + 612 + ], + "type": "text", + "content": ", the coordinates of the principal point. The pose at instant " + }, + { + "bbox": [ + 313, + 551, + 554, + 612 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 551, + 554, + 612 + ], + "type": "text", + "content": " is provided by rotation matrix " + }, + { + "bbox": [ + 313, + 551, + 554, + 612 + ], + "type": "inline_equation", + "content": "R_c^i" + }, + { + "bbox": [ + 313, + 551, + 554, + 612 + ], + "type": "text", + "content": " and translation vector " + }, + { + "bbox": [ + 313, + 551, + 554, + 612 + ], + "type": "inline_equation", + "content": "t_c^i" + }, + { + "bbox": [ + 313, + 551, + 554, + 612 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 623, + 485, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 485, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 485, + 635 + ], + "type": "text", + "content": "3.2. Spherical Geometry from Poses" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "content": "Spherical rectification requires that its main axis " + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "content": " be aligned with the direction of translation. This axis represents the epipoles and contains both camera centers. Let us designate the extrinsic parameter of a camera at a moment " + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "inline_equation", + "content": "[R_c^i |t_c^i ]\\in \\mathbf{R}^{3\\times 4}" + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "content": ". The spherical orientation is described as a rotation matrix " + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "inline_equation", + "content": "R_{S}" + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "content": " and is recovered from the orientation" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28927" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 558, + 240 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 558, + 240 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 558, + 240 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 558, + 240 + ], + "type": "image", + "image_path": "8b11457911e07e897ba9466337c9f6948f68ee07616ff2cc218f0ae71df0c62a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 250, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 54, + 250, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 54, + 250, + 555, + 285 + ], + "type": "text", + "content": "Figure 4. Spherical Stereo. Crops from image " + }, + { + "bbox": [ + 54, + 250, + 555, + 285 + ], + "type": "inline_equation", + "content": "I^i" + }, + { + "bbox": [ + 54, + 250, + 555, + 285 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 250, + 555, + 285 + ], + "type": "inline_equation", + "content": "I^{i-1}" + }, + { + "bbox": [ + 54, + 250, + 555, + 285 + ], + "type": "text", + "content": " are rectified. A shared feature encoder learns stereo features which are then compared with the dot product. Symmetric normalization (N) is applied according to the Permutation Stereo[4] formulation and the estimated disparities are transformed into depths. The reconstructed images and the depths are then derectified back to image space." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 297, + 228, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 228, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 228, + 308 + ], + "type": "text", + "content": "given by the translation of camera centers." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 117, + 314, + 170, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 314, + 170, + 327 + ], + "spans": [ + { + "bbox": [ + 117, + 314, + 170, + 327 + ], + "type": "interline_equation", + "content": "r _ {z} = t ^ {i} - t ^ {j}", + "image_path": "2ba33c0cb0e58a4d8f773afc26f9f020f22d1560c29801f8178b9e818ce63451.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 118, + 331, + 195, + 344 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 331, + 195, + 344 + ], + "spans": [ + { + "bbox": [ + 118, + 331, + 195, + 344 + ], + "type": "interline_equation", + "content": "r _ {y} = (0, 0, 1) \\times r _ {z}", + "image_path": "39284311eeb2113ad5a7a5291d3da7fa385d584829b2e7feb9965e1bde4a615b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 117, + 346, + 295, + 358 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 346, + 295, + 358 + ], + "spans": [ + { + "bbox": [ + 117, + 346, + 295, + 358 + ], + "type": "interline_equation", + "content": "r _ {x} = r _ {y} \\times r _ {z} \\tag {2}", + "image_path": "21c7545b90ef12c04848651e71203518c950d37e01191a18e0ed59425a0c024f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 110, + 360, + 241, + 390 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 360, + 241, + 390 + ], + "spans": [ + { + "bbox": [ + 110, + 360, + 241, + 390 + ], + "type": "interline_equation", + "content": "R _ {S} ^ {i, j} = \\left(\\frac {r _ {x}}{| | r _ {x} | |}, \\frac {r _ {y}}{| | r _ {y} | |}, \\frac {r _ {z}}{| | r _ {z} | |}\\right) ^ {T}", + "image_path": "b4b10c76a273bf7f3e9949c7cc2665a2364d5ebf1ea722aeecce4b8759de9e3c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 395, + 197, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 395, + 197, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 395, + 197, + 407 + ], + "type": "text", + "content": "3.3. Spherical Representation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 412, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 296, + 460 + ], + "type": "text", + "content": "The angles " + }, + { + "bbox": [ + 55, + 412, + 296, + 460 + ], + "type": "inline_equation", + "content": "(\\alpha, \\beta)" + }, + { + "bbox": [ + 55, + 412, + 296, + 460 + ], + "type": "text", + "content": " designate a 3D point " + }, + { + "bbox": [ + 55, + 412, + 296, + 460 + ], + "type": "inline_equation", + "content": "(x, y, z)" + }, + { + "bbox": [ + 55, + 412, + 296, + 460 + ], + "type": "text", + "content": " on the unit sphere, see Fig. 2, which is aligned with the spherical model principal axis according to the rotation " + }, + { + "bbox": [ + 55, + 412, + 296, + 460 + ], + "type": "inline_equation", + "content": "R_S^{i,j}" + }, + { + "bbox": [ + 55, + 412, + 296, + 460 + ], + "type": "text", + "content": ". The coordinate change operation " + }, + { + "bbox": [ + 55, + 412, + 296, + 460 + ], + "type": "inline_equation", + "content": "N_{\\alpha \\beta \\rightarrow xyz}" + }, + { + "bbox": [ + 55, + 412, + 296, + 460 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 467, + 295, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 467, + 295, + 522 + ], + "spans": [ + { + "bbox": [ + 107, + 467, + 295, + 522 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} (x, y, z) ^ {T} = N _ {\\alpha \\beta \\rightarrow x y z} (\\alpha , \\beta) \\\\ = \\left( \\begin{array}{c} \\sin (\\alpha) \\sin (\\beta) \\\\ - \\cos (\\alpha) \\sin (\\beta) \\\\ \\cos (\\beta) \\end{array} \\right) \\tag {3} \\\\ \\end{array}", + "image_path": "ee2709019ed65df57ac331b91994bb7965dd25c0c9b7add430361826ea0d4059.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 527, + 215, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 527, + 215, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 527, + 215, + 540 + ], + "type": "text", + "content": "and the inverse operation " + }, + { + "bbox": [ + 55, + 527, + 215, + 540 + ], + "type": "inline_equation", + "content": "N_{xyz\\to \\alpha \\beta}" + }, + { + "bbox": [ + 55, + 527, + 215, + 540 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 90, + 545, + 295, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 545, + 295, + 608 + ], + "spans": [ + { + "bbox": [ + 90, + 545, + 295, + 608 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} (\\alpha , \\beta) ^ {T} = N _ {x y z \\rightarrow \\alpha \\beta} (x, y, z) \\\\ = \\left( \\begin{array}{c} \\tan^ {- 1} \\left(\\frac {x}{- y}\\right) \\\\ \\left| \\tan^ {- 1} \\left(\\frac {\\operatorname {s g n} (x) \\sqrt {x ^ {2} + y ^ {2}}}{z}\\right) \\right| \\end{array} \\right) \\tag {4} \\\\ \\end{array}", + "image_path": "7d73fbf4746df1da88815b8af7461e1906907ae87cb3026cd3dfdc219dfdc707.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 612, + 181, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 181, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 181, + 625 + ], + "type": "text", + "content": "3.4. Epipolar Rectification" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "Because the camera field of view is usually small relative to the full sphere, spherical rectification can yield sparse images. We thus propose to partition the input images into a set of corresponding small regions which are then rectified into shared spherical representations, see Fig. 3. Let us designate a rectified image on the unit sphere, where the pixel intensity is " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "I_{S}(a,b) \\in \\mathbf{R}^{H_{S} \\times W_{S}}" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 297, + 555, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 381 + ], + "type": "text", + "content": "The indices of the rows and columns " + }, + { + "bbox": [ + 313, + 297, + 555, + 381 + ], + "type": "inline_equation", + "content": "(a,b)" + }, + { + "bbox": [ + 313, + 297, + 555, + 381 + ], + "type": "text", + "content": " of a spherical representation image " + }, + { + "bbox": [ + 313, + 297, + 555, + 381 + ], + "type": "inline_equation", + "content": "I_{S}" + }, + { + "bbox": [ + 313, + 297, + 555, + 381 + ], + "type": "text", + "content": " represent uniformly sampled angles " + }, + { + "bbox": [ + 313, + 297, + 555, + 381 + ], + "type": "inline_equation", + "content": "(\\alpha ,\\beta)" + }, + { + "bbox": [ + 313, + 297, + 555, + 381 + ], + "type": "text", + "content": " on the unit sphere such that " + }, + { + "bbox": [ + 313, + 297, + 555, + 381 + ], + "type": "inline_equation", + "content": "(\\alpha ,\\beta)\\in" + }, + { + "bbox": [ + 313, + 297, + 555, + 381 + ], + "type": "inline_equation", + "content": "\\left([\\alpha_{min},\\alpha_{max}],[\\beta_{min},\\beta_{max}]\\right)" + }, + { + "bbox": [ + 313, + 297, + 555, + 381 + ], + "type": "text", + "content": " in a rectified image size " + }, + { + "bbox": [ + 313, + 297, + 555, + 381 + ], + "type": "inline_equation", + "content": "(H_S,W_S)" + }, + { + "bbox": [ + 313, + 297, + 555, + 381 + ], + "type": "text", + "content": ", the spherical image height and width, see Fig. 3 (RIGHT). Conversion operations can be defined " + }, + { + "bbox": [ + 313, + 297, + 555, + 381 + ], + "type": "inline_equation", + "content": "N_{ab\\to \\alpha \\beta}" + }, + { + "bbox": [ + 313, + 297, + 555, + 381 + ], + "type": "text", + "content": " between the pair " + }, + { + "bbox": [ + 313, + 297, + 555, + 381 + ], + "type": "inline_equation", + "content": "(a,b)" + }, + { + "bbox": [ + 313, + 297, + 555, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 297, + 555, + 381 + ], + "type": "inline_equation", + "content": "(\\alpha ,\\beta)" + }, + { + "bbox": [ + 313, + 297, + 555, + 381 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 356, + 387, + 553, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 387, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 356, + 387, + 553, + 437 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} (\\alpha , \\beta) ^ {T} = N _ {a b \\rightarrow \\alpha \\beta} (a, b) \\\\ = \\left( \\begin{array}{c} \\frac {a \\left(\\alpha_ {\\text {m a x}} - \\alpha_ {\\text {m i n}}\\right)}{H _ {S}} + \\alpha_ {\\text {m i n}} \\\\ \\frac {b \\left(\\beta_ {\\text {m a x}} - \\beta_ {\\text {m i n}}\\right)}{W _ {S}} + \\beta_ {\\text {m i n}} \\end{array} \\right) \\tag {5} \\\\ \\end{array}", + "image_path": "ae05a93d630202ce0d21757cd7398f4b1151ee28c195e6eec478567f96ff7bd8.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 441, + 554, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 554, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 554, + 489 + ], + "type": "text", + "content": "The spherical epipolar rectification " + }, + { + "bbox": [ + 313, + 441, + 554, + 489 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 313, + 441, + 554, + 489 + ], + "type": "text", + "content": " is a rectification function that extracts a pixel at position " + }, + { + "bbox": [ + 313, + 441, + 554, + 489 + ], + "type": "inline_equation", + "content": "(u, v)" + }, + { + "bbox": [ + 313, + 441, + 554, + 489 + ], + "type": "text", + "content": " from an unrectified image " + }, + { + "bbox": [ + 313, + 441, + 554, + 489 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 441, + 554, + 489 + ], + "type": "text", + "content": " given an " + }, + { + "bbox": [ + 313, + 441, + 554, + 489 + ], + "type": "inline_equation", + "content": "(a, b)" + }, + { + "bbox": [ + 313, + 441, + 554, + 489 + ], + "type": "text", + "content": " pair. Therefore, pixel intensity in the rectified image is " + }, + { + "bbox": [ + 313, + 441, + 554, + 489 + ], + "type": "inline_equation", + "content": "I_S(a, b) = I(u, v)" + }, + { + "bbox": [ + 313, + 441, + 554, + 489 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 489, + 554, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 554, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 554, + 513 + ], + "type": "text", + "content": "The spherical epipolar rectification " + }, + { + "bbox": [ + 313, + 489, + 554, + 513 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 313, + 489, + 554, + 513 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 489, + 554, + 513 + ], + "type": "inline_equation", + "content": "(u, v)" + }, + { + "bbox": [ + 313, + 489, + 554, + 513 + ], + "type": "text", + "content": " coordinates to " + }, + { + "bbox": [ + 313, + 489, + 554, + 513 + ], + "type": "inline_equation", + "content": "(a, b)" + }, + { + "bbox": [ + 313, + 489, + 554, + 513 + ], + "type": "text", + "content": " coordinates is:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 368, + 520, + 445, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 520, + 445, + 533 + ], + "spans": [ + { + "bbox": [ + 368, + 520, + 445, + 533 + ], + "type": "interline_equation", + "content": "(u, v) = M _ {S} (a, b)", + "image_path": "d88c4fcf1746e7d36110e85cbca07ddbdeb996a842f02e10ae858c775a2e5aae.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 362, + 535, + 466, + 549 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 535, + 466, + 549 + ], + "spans": [ + { + "bbox": [ + 362, + 535, + 466, + 549 + ], + "type": "interline_equation", + "content": "(\\alpha , \\beta) ^ {T} = N _ {a b \\rightarrow \\alpha \\beta} (a, b)", + "image_path": "b185d49c3d79a8ebec22b32f0564483e3974a2cbf01285f184debfd3858d54f2.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 340, + 552, + 553, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 552, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 340, + 552, + 553, + 567 + ], + "type": "interline_equation", + "content": "\\left(x _ {p}, y _ {p}, z _ {p}\\right) ^ {T} = K \\cdot R _ {c} \\cdot R _ {s} \\cdot N _ {\\alpha \\beta \\rightarrow x y z} (\\alpha , \\beta) \\tag {6}", + "image_path": "04ea22f86f4c90a9dde676a915d9fdcfa3541f5eb298a1f92ad70e2996e5e1ee.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 370, + 569, + 451, + 596 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 569, + 451, + 596 + ], + "spans": [ + { + "bbox": [ + 370, + 569, + 451, + 596 + ], + "type": "interline_equation", + "content": "(u, v) = \\left(\\frac {y _ {p}}{z _ {p}}, \\frac {x _ {p}}{z _ {p}}\\right)", + "image_path": "807b073ec05927c506a8f03afcd0e1d08711a848b5164f13344c4cbccc300b0b.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 600, + 450, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 450, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 450, + 613 + ], + "type": "text", + "content": "3.5. Epipolar Derectification" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 618, + 554, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 554, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 554, + 654 + ], + "type": "text", + "content": "The complementary operator to epipolar rectification is the epipolar derectification. Let us define the conversion operation " + }, + { + "bbox": [ + 313, + 618, + 554, + 654 + ], + "type": "inline_equation", + "content": "N_{\\alpha \\beta \\to ab}" + }, + { + "bbox": [ + 313, + 618, + 554, + 654 + ], + "type": "text", + "content": " between the pair " + }, + { + "bbox": [ + 313, + 618, + 554, + 654 + ], + "type": "inline_equation", + "content": "(\\alpha, \\beta)" + }, + { + "bbox": [ + 313, + 618, + 554, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 618, + 554, + 654 + ], + "type": "inline_equation", + "content": "(a, b)" + }, + { + "bbox": [ + 313, + 618, + 554, + 654 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 660, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 660, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 321, + 660, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} (a, b) ^ {T} = N _ {\\alpha \\beta \\rightarrow a b} (\\alpha , \\beta) \\\\ = \\binom {(\\alpha - \\alpha_ {\\min }) H _ {S} / (\\alpha_ {\\max } - \\alpha_ {\\min })} {\\left. (m o d (\\beta , \\pi) - \\beta_ {\\min }\\right) W _ {S} / \\left(\\beta_ {\\max } - \\beta_ {\\min }\\right)} \\tag {7} \\\\ \\end{array}", + "image_path": "c2ac1e8009f8c35c63fa4d270407f26286d54ddbeeab31f497d088d25e615487.jpg" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28928" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "In the same manner, an unrectified image " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " with coordinates " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " can be generated from a spherical image " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "I_S" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "I(u,v) = I_S(a,b)" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": ". The spherical epipolar derectification " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "M_S^{-1}" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " operator is an operator that maps a coordinate in the camera coordinate system to a coordinate on the unit sphere." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 118, + 152, + 197, + 166 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 152, + 197, + 166 + ], + "spans": [ + { + "bbox": [ + 118, + 152, + 197, + 166 + ], + "type": "interline_equation", + "content": "(a, b) = M _ {S} ^ {- 1} (u, v)", + "image_path": "82ea0010538b71562cb66106483795beeabbebb05fbcef839deccd1fa794377c.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 113, + 168, + 294, + 206 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 168, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 113, + 168, + 294, + 206 + ], + "type": "interline_equation", + "content": "\\left( \\begin{array}{l} x _ {s} \\\\ y _ {s} \\\\ z _ {s} \\end{array} \\right) = R _ {s} ^ {T} \\cdot R _ {c} ^ {T} \\cdot K ^ {- 1} \\cdot \\left( \\begin{array}{l} u \\\\ v \\\\ 1 \\end{array} \\right) \\tag {8}", + "image_path": "3340024870702d7bd8e99e91ca984e1eb0ebe5a8910c51068909e6a3367062d5.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 209, + 238, + 222 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 209, + 238, + 222 + ], + "spans": [ + { + "bbox": [ + 107, + 209, + 238, + 222 + ], + "type": "interline_equation", + "content": "\\left(\\alpha , \\beta\\right) ^ {T} = N _ {x y z \\rightarrow \\alpha \\beta} \\left(x _ {s}, y _ {s}, z _ {s}\\right)", + "image_path": "3fc1f84ba50c2852d110703067d7ecf0e60a0a2a4c6756ffce57275d73d4ff71.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 224, + 213, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 224, + 213, + 239 + ], + "spans": [ + { + "bbox": [ + 111, + 224, + 213, + 239 + ], + "type": "interline_equation", + "content": "(a, b) ^ {T} = N _ {\\alpha \\beta \\rightarrow a b} (\\alpha , \\beta)", + "image_path": "323eecf1f2ac0cd77320da6936ef723173ae0d656bbe0218c494d35381a0000e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 250, + 290, + 263 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 250, + 290, + 263 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 290, + 263 + ], + "type": "text", + "content": "4. Spherical Stereo for Two Monocular Views" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 270, + 149, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 149, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 149, + 283 + ], + "type": "text", + "content": "4.1. Spherical Crop" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 288, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 296, + 384 + ], + "type": "text", + "content": "Because spherical epipolar rectification introduces some distortion in the rectified image, we propose a novel spherical crop function illustrated in Fig. 3. This function operates similarly to the random crop data augmentation technique with rectified image dimensions " + }, + { + "bbox": [ + 55, + 288, + 296, + 384 + ], + "type": "inline_equation", + "content": "(H_{S}, W_{S})" + }, + { + "bbox": [ + 55, + 288, + 296, + 384 + ], + "type": "text", + "content": " and the selected crop size " + }, + { + "bbox": [ + 55, + 288, + 296, + 384 + ], + "type": "inline_equation", + "content": "(H_{C}, W_{C})" + }, + { + "bbox": [ + 55, + 288, + 296, + 384 + ], + "type": "text", + "content": ". During rectification, information loss will be proportional to " + }, + { + "bbox": [ + 55, + 288, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\left(\\frac{H_{C}}{H_{S}}, \\frac{W_{C}}{W_{S}}\\right)" + }, + { + "bbox": [ + 55, + 288, + 296, + 384 + ], + "type": "text", + "content": ". It is empirically suggested to select a least " + }, + { + "bbox": [ + 55, + 288, + 296, + 384 + ], + "type": "inline_equation", + "content": "(H_{S}, W_{S}) = 1.8(H_{C}, W_{C})" + }, + { + "bbox": [ + 55, + 288, + 296, + 384 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 384, + 296, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 296, + 409 + ], + "type": "text", + "content": "Illustrated in Fig. 3 (LEFT), given a crop in images " + }, + { + "bbox": [ + 55, + 384, + 296, + 409 + ], + "type": "inline_equation", + "content": "I^i" + }, + { + "bbox": [ + 55, + 384, + 296, + 409 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 384, + 296, + 409 + ], + "type": "inline_equation", + "content": "I^j" + }, + { + "bbox": [ + 55, + 384, + 296, + 409 + ], + "type": "text", + "content": ", the corners " + }, + { + "bbox": [ + 55, + 384, + 296, + 409 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 384, + 296, + 409 + ], + "type": "text", + "content": " are represented as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 96, + 418, + 295, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 418, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 96, + 418, + 295, + 456 + ], + "type": "interline_equation", + "content": "c = \\left( \\begin{array}{c c c c} u _ {m i n} & u _ {m i n} & u _ {m a x} & u _ {m a x} \\\\ v _ {m i n} & v _ {m a x} & v _ {m i n} & v _ {m a x} \\\\ 1 & 1 & 1 & 1 \\end{array} \\right) \\tag {9}", + "image_path": "7868ad94afc7e54c3c879ee1d7da8384459b3979827e518f8efb6e0759797ea3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 466, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 466, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 296, + 491 + ], + "type": "text", + "content": "and are mapped to rectified image space and then used to compute " + }, + { + "bbox": [ + 55, + 466, + 296, + 491 + ], + "type": "inline_equation", + "content": "(\\alpha_{min},\\alpha_{max},\\beta_{min},\\beta_{max})" + }, + { + "bbox": [ + 55, + 466, + 296, + 491 + ], + "type": "text", + "content": " .." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 72, + 501, + 294, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 501, + 294, + 594 + ], + "spans": [ + { + "bbox": [ + 72, + 501, + 294, + 594 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left( \\begin{array}{c} c _ {S} ^ {i} \\\\ c _ {S} ^ {j} \\end{array} \\right) = \\left( \\begin{array}{c} R _ {s} ^ {T} \\cdot (R _ {c} ^ {i}) ^ {T} \\cdot K ^ {- 1} \\cdot c \\\\ R _ {s} ^ {T} \\cdot (R _ {c} ^ {j}) ^ {T} \\cdot K ^ {- 1} \\cdot c \\end{array} \\right) \\\\ \\left( \\begin{array}{c} (\\alpha^ {i}, \\beta^ {i}) ^ {T} \\\\ (\\alpha^ {j}, \\beta^ {j}) ^ {T} \\end{array} \\right) = \\left( \\begin{array}{c} N _ {x y z \\to \\alpha \\beta} (c _ {S} ^ {i}) \\\\ N _ {x y z \\to \\alpha \\beta} (c _ {S} ^ {i}) \\end{array} \\right) \\\\ \\left( \\begin{array}{l l} \\alpha_ {m i n} & \\alpha_ {m a x} \\\\ \\beta_ {m i n} & \\beta_ {m a x} \\end{array} \\right) = \\left( \\begin{array}{l l} \\min \\left(\\alpha^ {i}, \\alpha^ {j}\\right) & \\max \\left(\\alpha^ {i}, \\alpha^ {j}\\right) \\\\ \\min \\left(\\beta^ {i}, \\beta^ {j}\\right) & \\max \\left(\\beta^ {i}, \\beta^ {j}\\right) \\end{array} \\right) \\tag {10} \\\\ \\end{array}", + "image_path": "85be057b4711116a45c31eaef320af7acca000ad1b516a199c7fa52a543c5a27.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 605, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 712 + ], + "type": "text", + "content": "Very importantly, it is possible for the alpha range to be around " + }, + { + "bbox": [ + 55, + 605, + 296, + 712 + ], + "type": "inline_equation", + "content": "2\\pi" + }, + { + "bbox": [ + 55, + 605, + 296, + 712 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 55, + 605, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\alpha_{min} < 2\\pi < \\alpha_{max}" + }, + { + "bbox": [ + 55, + 605, + 296, + 712 + ], + "type": "text", + "content": ". When this happens, the spherical rectification of the crop will not work. To solve for this, an offset of " + }, + { + "bbox": [ + 55, + 605, + 296, + 712 + ], + "type": "inline_equation", + "content": "-\\pi" + }, + { + "bbox": [ + 55, + 605, + 296, + 712 + ], + "type": "text", + "content": " is added to " + }, + { + "bbox": [ + 55, + 605, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 605, + 296, + 712 + ], + "type": "text", + "content": ". This represents a spherical representation rotated by " + }, + { + "bbox": [ + 55, + 605, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 55, + 605, + 296, + 712 + ], + "type": "text", + "content": " around its " + }, + { + "bbox": [ + 55, + 605, + 296, + 712 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 605, + 296, + 712 + ], + "type": "text", + "content": " axis. This is not reflected in the above equations to simplify reading. Similarly to networks trained with cropping data augmentation, the neural network is mapped over images at inference time." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 536, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 536, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 536, + 85 + ], + "type": "text", + "content": "4.2. Overview of Permutation Stereo Matching" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 89, + 554, + 136 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 554, + 136 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 554, + 136 + ], + "type": "text", + "content": "As shown in Fig. 4, the matching process relies on the Permutation Stereo formulation presented in [5]. It is chosen to highlight the contribution of the spherical epipolar rectification." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 144, + 557, + 215 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 144, + 557, + 215 + ], + "spans": [ + { + "bbox": [ + 316, + 144, + 557, + 215 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} f ^ {i} = F \\left(I _ {S} ^ {i}\\right), \\quad f ^ {j} = F \\left(I _ {S} ^ {j}\\right) \\\\ C _ {m, n, o} ^ {i, j} = \\sum_ {g} f _ {m, n, g} ^ {i} \\cdot f _ {m, n, g} ^ {j}, \\text {w h e r e} C \\in \\mathbb {R} ^ {H _ {S} \\times W _ {S} \\times W _ {S}} \\\\ P ^ {i, j} = \\operatorname {N o r m} \\left(C ^ {i, j}\\right) \\quad , \\text {w h e r e} \\quad P \\in \\mathbb {R} ^ {H _ {S} \\times W _ {S} \\times W _ {S}} \\tag {11} \\\\ \\end{array}", + "image_path": "10da4cdd34c93815657a393975b09c89f46578d3984f3cdd3a5efe0263dce0df.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 224, + 554, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 554, + 273 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 554, + 273 + ], + "type": "text", + "content": "The normalization operation Norm is a softmax on rows and columns simultaneously and is detailed in [4]. Disparity is obtained from the permutation volume " + }, + { + "bbox": [ + 313, + 224, + 554, + 273 + ], + "type": "inline_equation", + "content": "P^{i,j}" + }, + { + "bbox": [ + 313, + 224, + 554, + 273 + ], + "type": "text", + "content": " for the spherical images " + }, + { + "bbox": [ + 313, + 224, + 554, + 273 + ], + "type": "inline_equation", + "content": "I_S^i" + }, + { + "bbox": [ + 313, + 224, + 554, + 273 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 224, + 554, + 273 + ], + "type": "inline_equation", + "content": "I_S^j" + }, + { + "bbox": [ + 313, + 224, + 554, + 273 + ], + "type": "text", + "content": " according to [4]," + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 372, + 281, + 553, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 281, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 372, + 281, + 553, + 338 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} D _ {S} ^ {i, j} = \\frac {1}{o} \\sum_ {o} P _ {m, n, o} ^ {i, j} * (o - n) \\tag {12} \\\\ D _ {S} ^ {j, i} = \\frac {1}{n} \\sum_ {n} P _ {m, n, o} ^ {i, j} * (o - n) \\\\ \\end{array}", + "image_path": "249726f0471f98d9fc06fb6bc582730c957aa3b453ba9ec500d50ce1eac554e2.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 347, + 518, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 518, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 518, + 361 + ], + "type": "text", + "content": "4.3. Depth from a Spherical Disparity Map" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 364, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 555, + 449 + ], + "type": "text", + "content": "In spherical representation, given a disparity map " + }, + { + "bbox": [ + 313, + 364, + 555, + 449 + ], + "type": "inline_equation", + "content": "D_S^{i,j}" + }, + { + "bbox": [ + 313, + 364, + 555, + 449 + ], + "type": "text", + "content": ", we solve for depth " + }, + { + "bbox": [ + 313, + 364, + 555, + 449 + ], + "type": "inline_equation", + "content": "Q_S^i" + }, + { + "bbox": [ + 313, + 364, + 555, + 449 + ], + "type": "text", + "content": " by triangulation and then detectify it back to the respective image representations " + }, + { + "bbox": [ + 313, + 364, + 555, + 449 + ], + "type": "inline_equation", + "content": "Q^i" + }, + { + "bbox": [ + 313, + 364, + 555, + 449 + ], + "type": "text", + "content": ". Notice that while the disparity is defined for an image pair, the depth is defined for the image itself. Here we present the triangulation operator between disparity and depth in spherical representation as " + }, + { + "bbox": [ + 313, + 364, + 555, + 449 + ], + "type": "inline_equation", + "content": "T_{D\\to Q}" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 340, + 458, + 547, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 458, + 547, + 502 + ], + "spans": [ + { + "bbox": [ + 340, + 458, + 547, + 502 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} Q _ {S} ^ {i} (a, b) = T _ {D \\rightarrow Q} (a, b, D _ {S} ^ {i, j} (a, b)) \\\\ = (R _ {c} ^ {i T} \\cdot p _ {s} ^ {i}) | | t ^ {j} - t ^ {i} | | \\frac {\\sin (\\angle (p _ {s} ^ {i} , p _ {s} ^ {j}))}{\\sin (\\angle (p _ {s} ^ {i} , t ^ {i} - t ^ {j}))} \\\\ \\end{array}", + "image_path": "746bc76278a5b02412363eeec0ab3aab7994dade89708614e39cad2eb66b4232.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 330, + 505, + 358, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 505, + 358, + 514 + ], + "spans": [ + { + "bbox": [ + 330, + 505, + 358, + 514 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 323, + 518, + 512, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 518, + 512, + 574 + ], + "spans": [ + { + "bbox": [ + 323, + 518, + 512, + 574 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left( \\begin{array}{c} p _ {s} ^ {i} \\\\ p _ {s} ^ {j} \\end{array} \\right) = \\left( \\begin{array}{c} - R _ {s} \\cdot N _ {\\alpha \\beta \\to x y z} (\\alpha^ {i}, \\beta^ {i}) \\\\ - R _ {s} \\cdot N _ {\\alpha \\beta \\to x y z} (\\alpha^ {j}, \\beta^ {j}) \\end{array} \\right) \\\\ \\left( \\begin{array}{c} (\\alpha^ {i}, \\beta^ {i}) ^ {T} \\\\ (\\alpha^ {j}, \\beta^ {j}) ^ {T} \\end{array} \\right) = \\left( \\begin{array}{c} N _ {a b \\to \\alpha \\beta} (a, b) \\\\ N _ {a b \\to \\alpha \\beta} (a, b + D _ {S} ^ {i, j} (a, b)) \\end{array} \\right) \\\\ \\end{array}", + "image_path": "16a7eb876e880e4f56ba7b3f7c84c279ae2d97d546faec3fa4866afa9cd4dc68.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 328, + 577, + 421, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 577, + 421, + 589 + ], + "spans": [ + { + "bbox": [ + 328, + 577, + 421, + 589 + ], + "type": "text", + "content": "and the vector angle is" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 343, + 591, + 553, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 591, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 343, + 591, + 553, + 628 + ], + "type": "interline_equation", + "content": "\\angle (g, h) = \\cos^ {- 1} \\left(\\frac {g \\cdot h}{| | g | | | | h | |}\\right) \\tag {13}", + "image_path": "4985332c0421eefe9f208ccf16a9604ad98ed275cd5d0d09d25aa7db8c1896fb.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 638, + 411, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 638, + 411, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 638, + 411, + 651 + ], + "type": "text", + "content": "4.4. Training Losses" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 656, + 554, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 656, + 554, + 692 + ], + "spans": [ + { + "bbox": [ + 313, + 656, + 554, + 692 + ], + "type": "text", + "content": "Supervised Setting. The supervised method is trained with the L1 Loss between predicted absolute depth " + }, + { + "bbox": [ + 313, + 656, + 554, + 692 + ], + "type": "inline_equation", + "content": "\\tilde{d}" + }, + { + "bbox": [ + 313, + 656, + 554, + 692 + ], + "type": "text", + "content": " for the first image and ground truth absolute depth " + }, + { + "bbox": [ + 313, + 656, + 554, + 692 + ], + "type": "inline_equation", + "content": "d_{GT}" + }, + { + "bbox": [ + 313, + 656, + 554, + 692 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 397, + 700, + 553, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 700, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 397, + 700, + 553, + 714 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\left| \\left| \\tilde {d} - d _ {G T} \\right| \\right| _ {1} \\tag {14}", + "image_path": "c77503f5a5d052ef6fce527b69b602d23719aa33380c6e03961429c07d977da0.jpg" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28929" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 86, + 70, + 525, + 152 + ], + "blocks": [ + { + "bbox": [ + 86, + 70, + 525, + 152 + ], + "lines": [ + { + "bbox": [ + 86, + 70, + 525, + 152 + ], + "spans": [ + { + "bbox": [ + 86, + 70, + 525, + 152 + ], + "type": "table", + "html": "
Crop SizeRectified SizeError - Lower is better ↓Accuracy - Higher is better ↑
MethodHC × WCHS × WSAbs RelSq RelRMSERMSE logδ < 1.25δ < 1.252δ < 1.253
No CropSpherical SGM-256 × 2560.6315.224.990.540.510.690.79
Spherical SGM-768 × 7680.398.463.400.420.690.800.86
Spherical Stereo-256 × 2560.5710.094.630.560.470.660.77
Spherical Stereo-768 × 7680.449.053.460.460.650.770.83
CropSpherical SGM120 × 160256 × 2560.254.632.430.220.780.860.91
Spherical Stereo120 × 160256 × 2560.233.322.390.240.750.840.90
", + "image_path": "9f0460303845f5464228dd502e358d5f141c90339a5587a212e7f8cb96ada41a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 160, + 555, + 195 + ], + "lines": [ + { + "bbox": [ + 55, + 160, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 55, + 160, + 555, + 195 + ], + "type": "text", + "content": "Table 1. Spherical Crop ablation study on TartanAir carwelding testing set. Spherical SGM and Spherical Permutation Stereo are used for different image crop " + }, + { + "bbox": [ + 55, + 160, + 555, + 195 + ], + "type": "inline_equation", + "content": "H_{C} \\times W_{C}" + }, + { + "bbox": [ + 55, + 160, + 555, + 195 + ], + "type": "text", + "content": " and rectified crop " + }, + { + "bbox": [ + 55, + 160, + 555, + 195 + ], + "type": "inline_equation", + "content": "H_{S} \\times W_{S}" + }, + { + "bbox": [ + 55, + 160, + 555, + 195 + ], + "type": "text", + "content": " sizes. \"Crop\" indicates the presence of Spherical Crop and \"No Crop\" indicates an absence of Spherical Crop. Spherical Crop improves performance significantly." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 215, + 296, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 215, + 296, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 296, + 264 + ], + "type": "text", + "content": "Self-Supervised Setting. The self-supervised method is trained using photometric losses on the reconstructed images. An image can be reconstructed by warping it with the permutation on its pairing image or on itself:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 270, + 294, + 292 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 270, + 294, + 292 + ], + "spans": [ + { + "bbox": [ + 91, + 270, + 294, + 292 + ], + "type": "interline_equation", + "content": "I _ {S} ^ {j ^ {\\prime}} = P \\cdot I _ {S} ^ {i} \\quad I _ {S} ^ {i ^ {\\prime}} = P ^ {\\top} \\cdot I _ {S} ^ {j} \\tag {15}", + "image_path": "3cff77a07ac06eebe79dda8f398167c5c2356389f273ab82a101a904af740a65.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 91, + 288, + 261, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 288, + 261, + 304 + ], + "spans": [ + { + "bbox": [ + 91, + 288, + 261, + 304 + ], + "type": "interline_equation", + "content": "I _ {S} ^ {i ^ {\\prime \\prime}} = P ^ {\\top} \\cdot P \\cdot I _ {S} ^ {i} \\qquad I _ {S} ^ {j ^ {\\prime \\prime}} = P \\cdot P ^ {\\top} \\cdot I _ {S} ^ {j}", + "image_path": "01afb8f0f2452f6a4daa8c25ffbe39e650f1dc946a403eb7587dfee4b7c796e7.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 311, + 295, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 311, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 295, + 335 + ], + "type": "text", + "content": "The photometric loss between input and target images " + }, + { + "bbox": [ + 55, + 311, + 295, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Photo}}(I,I^{*})" + }, + { + "bbox": [ + 55, + 311, + 295, + 335 + ], + "type": "text", + "content": " uses the structural similarity (SSIM)[33]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 342, + 294, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 342, + 294, + 372 + ], + "spans": [ + { + "bbox": [ + 57, + 342, + 294, + 372 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {P h o t o}} (I, I ^ {*}) = \\frac {\\alpha}{2} (1 - \\operatorname {S S I M} (I, I ^ {*})) + (1 - \\alpha) \\| I - I ^ {*} \\| _ {1} \\tag {16}", + "image_path": "db9dedab653b9e727f825a6f5a39dcf9b146be96571ea6b78be2ca3d89d01f68.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 373, + 296, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 408 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 408 + ], + "type": "text", + "content": "The self-supervised loss function occurs between the images and their reconstructions. This is done in image representation rather than spherical representation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 79, + 415, + 294, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 415, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 79, + 415, + 294, + 460 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\frac {1}{4} \\left(\\mathcal {L} _ {\\text {P h o t o}} \\left(I ^ {i}, I ^ {i ^ {\\prime}}\\right) + \\mathcal {L} _ {\\text {P h o t o}} \\left(I ^ {i}, I ^ {i ^ {\\prime \\prime}}\\right) \\right. \\tag {17} \\\\ \\left. + \\mathcal {L} _ {\\text {P h o t o}} \\left(I ^ {j}, I ^ {j ^ {\\prime}}\\right) + \\mathcal {L} _ {\\text {P h o t o}} \\left(I ^ {j}, I ^ {j ^ {\\prime \\prime}}\\right)\\right) \\\\ \\end{array}", + "image_path": "fefa9daecbf05c5f0400a0e6960dc6239c4b9ffb39684673b7103e05f4223bfd.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 466, + 137, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 466, + 137, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 137, + 479 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 486, + 296, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 702 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 702 + ], + "type": "text", + "content": "Datasets. The presented results evaluate absolute depth estimation on the TartanAir[32] Samples dataset and the KITTI[12] Eigen[10] split. Both datasets provide ground truth pose. The TartanAir Samples is a synthetic dataset that has 6982 images of size " + }, + { + "bbox": [ + 55, + 486, + 296, + 702 + ], + "type": "inline_equation", + "content": "480 \\times 640" + }, + { + "bbox": [ + 55, + 486, + 296, + 702 + ], + "type": "text", + "content": " pixels split across 16 scenes of various locations from seaside towns to under the ocean with day and night lighting conditions. As a test sequence, 46 images from the carwelding sequence have been selected. The KITTI Eigen dataset is a real world dataset of driving scenes and has approximately " + }, + { + "bbox": [ + 55, + 486, + 296, + 702 + ], + "type": "inline_equation", + "content": "20k" + }, + { + "bbox": [ + 55, + 486, + 296, + 702 + ], + "type": "text", + "content": " images for training and 697 for testing with an image size of " + }, + { + "bbox": [ + 55, + 486, + 296, + 702 + ], + "type": "inline_equation", + "content": "375 \\times 1242" + }, + { + "bbox": [ + 55, + 486, + 296, + 702 + ], + "type": "text", + "content": " pixels. Raw KITTI images were undistorted. Error metrics are computed for absolute depth according to [11] and capped at " + }, + { + "bbox": [ + 55, + 486, + 296, + 702 + ], + "type": "inline_equation", + "content": "50\\mathrm{m}" + }, + { + "bbox": [ + 55, + 486, + 296, + 702 + ], + "type": "text", + "content": " according to [14]. Implementation details are presented in supplementary. For comparison, a simplified version of SGM [16] is adapted to Spherical Stereo to illustrate independence to Permutation Stereo. The method is called Spherical SGM and is detailed in [5]." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 339, + 212, + 531, + 338 + ], + "blocks": [ + { + "bbox": [ + 339, + 212, + 531, + 338 + ], + "lines": [ + { + "bbox": [ + 339, + 212, + 531, + 338 + ], + "spans": [ + { + "bbox": [ + 339, + 212, + 531, + 338 + ], + "type": "image", + "image_path": "b5bdc3dcff895bac65be87611b325326417699188caaa68bf5b80b1e49e46ad6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 346, + 555, + 433 + ], + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 433 + ], + "type": "text", + "content": "Figure 5. Distortion generated on the rectified image for the different polar rectification methods for a " + }, + { + "bbox": [ + 313, + 346, + 555, + 433 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 313, + 346, + 555, + 433 + ], + "type": "text", + "content": " FOV camera. Planar rectification is the homography based epipolar rectification available among other things in OpenCV [3]. Cylindrical rectification is the method initially proposed in [25] and used in [6, 24, 28]. Spherical rectification increases distortion by adding an additional polar parameter. Spherical Crop applied to spherical rectification limits this weakness and leads to lowest distortion." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 441, + 481, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 481, + 454 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 481, + 454 + ], + "type": "text", + "content": "5.1. Spherical Crop Ablation Study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 459, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 555, + 662 + ], + "type": "text", + "content": "An ablation study is performed for the Spherical Crop as a part of the Spherical Stereo and the results are presented in Tab. 1. With regards to performance, Spherical SGM leads to better accuracy because of its smoothing properties while Spherical Stereo matching has lower error. Irrespective of the method, results illustrate that the Spherical Crop operation leads to significantly improved performances. Spherical Crop not only improves performance, but also makes training the network less challenging. This can be attributed to the fact that the proposed crop significantly reduces distortion. Previous works [24, 25, 28] have identified that the rectified image must be very large to manage distortion for adequate monocular stereo matching. Mapping an image patch to its corresponding " + }, + { + "bbox": [ + 313, + 459, + 555, + 662 + ], + "type": "inline_equation", + "content": "(\\alpha, \\beta)" + }, + { + "bbox": [ + 313, + 459, + 555, + 662 + ], + "type": "text", + "content": " range is demonstrated here to be an appropriate way to handle distortion and makes it possible to integrate polar rectification with deep stereo matching." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 671, + 489, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 489, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 489, + 685 + ], + "type": "text", + "content": "5.2. Epipolar Distortion Comparison" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Distortion during spherical rectification is measured and compared against state-of-the-art cylindrical rectification" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28930" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 70, + 539, + 161 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 539, + 161 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 539, + 161 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 539, + 161 + ], + "type": "table", + "html": "
Error - Lower is better ↓Accuracy - Higher is better ↑Nbr. of Params.
MethodAbs RelSq RelRMSERMSE logδ < 1.25δ < 1.252δ < 1.253
Depth Anything V2 [37]Supervised & Scaled10.331.042.630.300.380.800.9897.47M
GMDepth [36]Supervised0.241.202.550.240.600.850.957.32M
Lite-Mono 8M [43]Supervised0.230.611.410.340.700.780.818.78M
DeepSfm Rev [29]Supervised0.150.411.420.140.800.950.9916.64M
MonST3R [42]Supervised & Scaled10.150.491.300.140.790.950.98577.81M
Spherical Stereo (ours)Supervised0.100.410.980.120.920.960.980.73M
Spherical Stereo (ours)Self-Supervised0.233.322.390.240.750.840.900.73M
", + "image_path": "af5adbf353a60bfd8ccf9b37489b87c1bd0eccc4b2c189ed8341143bc5a3b566.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 170, + 555, + 194 + ], + "lines": [ + { + "bbox": [ + 55, + 170, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 170, + 555, + 194 + ], + "type": "text", + "content": "Table 2. Absolute Metric Depth, Comparative results with state of the art methods. Spherical Stereo is trained with the self-supervised loss and the supervised loss. Spherical Stereo is competitive with state of the art methods. " + }, + { + "bbox": [ + 55, + 170, + 555, + 194 + ], + "type": "inline_equation", + "content": "{}^{1}" + }, + { + "bbox": [ + 55, + 170, + 555, + 194 + ], + "type": "text", + "content": " Per-sequence scale & shift with GT depths." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 83, + 203, + 176, + 273 + ], + "blocks": [ + { + "bbox": [ + 83, + 203, + 176, + 273 + ], + "lines": [ + { + "bbox": [ + 83, + 203, + 176, + 273 + ], + "spans": [ + { + "bbox": [ + 83, + 203, + 176, + 273 + ], + "type": "image", + "image_path": "ee979d8aeafabf7e3762ed5a91f700a929c65811c61f1e5ed9cf5e5c19403f5f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 176, + 203, + 269, + 273 + ], + "blocks": [ + { + "bbox": [ + 176, + 203, + 269, + 273 + ], + "lines": [ + { + "bbox": [ + 176, + 203, + 269, + 273 + ], + "spans": [ + { + "bbox": [ + 176, + 203, + 269, + 273 + ], + "type": "image", + "image_path": "a0251621fe2b46941ee32a33ca6dd10ba337633313cb38de795b6d35166acc06.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 83, + 273, + 176, + 343 + ], + "blocks": [ + { + "bbox": [ + 83, + 273, + 176, + 343 + ], + "lines": [ + { + "bbox": [ + 83, + 273, + 176, + 343 + ], + "spans": [ + { + "bbox": [ + 83, + 273, + 176, + 343 + ], + "type": "image", + "image_path": "427a5e56f1c763b2b68a9f9c65380c9f8a770347cea3406e35b6b0aec08d6a74.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 176, + 273, + 269, + 343 + ], + "blocks": [ + { + "bbox": [ + 176, + 273, + 269, + 343 + ], + "lines": [ + { + "bbox": [ + 176, + 273, + 269, + 343 + ], + "spans": [ + { + "bbox": [ + 176, + 273, + 269, + 343 + ], + "type": "image", + "image_path": "1d08d66b27dc0effdcb96f5b8593e9b9640a525ac6d325f931e0a86fc520abb3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 83, + 343, + 176, + 413 + ], + "blocks": [ + { + "bbox": [ + 83, + 343, + 176, + 413 + ], + "lines": [ + { + "bbox": [ + 83, + 343, + 176, + 413 + ], + "spans": [ + { + "bbox": [ + 83, + 343, + 176, + 413 + ], + "type": "image", + "image_path": "3a9b97b58469f57e8590422638c72d9e7fae99ca05e74e25f469e44115745ff7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 176, + 343, + 269, + 412 + ], + "blocks": [ + { + "bbox": [ + 176, + 343, + 269, + 412 + ], + "lines": [ + { + "bbox": [ + 176, + 343, + 269, + 412 + ], + "spans": [ + { + "bbox": [ + 176, + 343, + 269, + 412 + ], + "type": "image", + "image_path": "d3bfb3bb36cfbe82b0dac31b74e646ba90325b9caee328d9e91672de2260c5df.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 83, + 413, + 176, + 483 + ], + "blocks": [ + { + "bbox": [ + 83, + 413, + 176, + 483 + ], + "lines": [ + { + "bbox": [ + 83, + 413, + 176, + 483 + ], + "spans": [ + { + "bbox": [ + 83, + 413, + 176, + 483 + ], + "type": "image", + "image_path": "06b4b6df97abd749a0b0ecc9232c26503c172562a0bd366d7683a207f052f802.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 491, + 297, + 536 + ], + "lines": [ + { + "bbox": [ + 55, + 491, + 297, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 297, + 536 + ], + "type": "text", + "content": "Figure 6. Qualitative results showing the inverse of the predicted depth with error map on a TartanAir carwelding test image. Depths are rounded to reveal depth planes. Additional visualization results are available in Depth-Visualization." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 176, + 413, + 269, + 483 + ], + "blocks": [ + { + "bbox": [ + 176, + 413, + 269, + 483 + ], + "lines": [ + { + "bbox": [ + 176, + 413, + 269, + 483 + ], + "spans": [ + { + "bbox": [ + 176, + 413, + 269, + 483 + ], + "type": "image", + "image_path": "0117432f4594e3c7caf5bf17c840ad850fa66ea0a833c73b52a9c40602a4702e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "text", + "content": "methods [6, 24, 25, 28] in Fig. 5. Distortion can be directly measured as the ratio of area occupied by the biggest and the smallest rectified pixels in the rectified image. Fig. 5 illustrates that planar rectification is not appropriate for general motion and breaks when the epipole is inside the image. As a downside to being differentiable, the proposed spherical rectification model introduces more distortion than the cylindrical rectification model. This is because there is an additional polar parameter which warps the radial direction non-linearly. That said, the proposed Spherical Crop leads to significantly lower distortion with regards to cylindrical rectification especially for geometries where the baseline direction is not completely forward. The forward baseline di" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 205, + 555, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 555, + 242 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 555, + 242 + ], + "type": "text", + "content": "resection remains the most unfavorable geometry to solve for two-view monocular stereo. This result highlights that the Spherical Crop minimizes distortion during rectification." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 250, + 481, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 250, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 250, + 481, + 262 + ], + "type": "text", + "content": "5.3. Depth Estimation Comparison" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 268, + 556, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 556, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 556, + 483 + ], + "type": "text", + "content": "An evaluation of Spherical Stereo with regards to absolute depth estimation is performed and results are presented in Tab. 2 in juxtaposition to state-of-the-art methods. These methods are selected as the benchmark for the two types of depth estimators presented in section 2.2. Depth Anything V2 and MonST3R are scaled per-sequence according to the best fit curve using the ground truth depths, a scaling function is illustrated in supp. Sec. 3. Other methods have been supervised on the dataset using the same splits and according to their training protocols. This is to allow for comparison on absolute depth estimation. A qualitative example of some approaches is presented in Fig. 6. Spherical Stereo outperforms other methods at absolute depth estimation. In the self-supervised setting, Spherical Stereo performs competitively with many methods that rely on ground truths. This demonstrates that the proposed Spherical Epipolar Rectification is a proper way to extend deep stereo matching to two-view absolute depth estimation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 492, + 488, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 492, + 488, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 492, + 488, + 504 + ], + "type": "text", + "content": "5.4. Natural Multi-Image Coherence" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": "Given the binocular nature of stereo matching, an evaluation of the behaviour with regards to the selected images at inference is presented in Tab. 3. Spherical Stereo is trained self-supervised for pairs of images " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\{i,i - 1\\}" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " as a baseline. This means that 2 images, an image and its previous image, are used for stereo matching during training. Results demonstrate that there is no loss in performance when using images " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\{i,i + 1\\}" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": ", an image and its following image, at inference. Degradation in performance can be observed when using images that are further away in time, e.g. " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\{i,i - 3\\}" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": ", because of more out-of-image pixels. The multi-frame results in Tab. 3 show improvements with respect to the baseline across all selected groups of 3 images with the overall best performance when using 5 images when compared to 2 frames. This shows that using more frames always improves depth estimation performance. Because spherical epipolar rectification properly aligns epipolar lines, this al" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28931" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 97, + 70, + 515, + 191 + ], + "blocks": [ + { + "bbox": [ + 97, + 70, + 515, + 191 + ], + "lines": [ + { + "bbox": [ + 97, + 70, + 515, + 191 + ], + "spans": [ + { + "bbox": [ + 97, + 70, + 515, + 191 + ], + "type": "table", + "html": "
Error - Lower is better ↓Accuracy - Higher is better ↑
Nb. of FramesFrame TimestepsAbs RelSq RelRMSERMSE logδ < 1.25δ < 1.252δ < 1.253
2{i,i-1}0.233.322.390.240.750.840.90
2{i,i-3}0.407.002.720.290.700.800.86
2{i,i-2}0.304.872.430.250.740.830.89
2{i,i+3}0.325.032.750.280.700.810.87
2{i,i+2}0.243.392.460.250.740.830.89
2{i,i+1}0.223.022.530.250.740.830.89
3{i,i-1,i-2}0.233.162.260.230.750.850.90
3{i,i-2,i+2}0.202.052.030.210.770.870.92
3{i,i-1,i+1}0.192.082.210.210.770.860.91
5{i,i-1,i+1,i-2,i+2}0.171.582.020.200.780.870.92
", + "image_path": "ae354e18444cca6be1c36c69f12de6d7756300002d1c1834ffb49092aebf8637.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 125, + 243, + 486, + 306 + ], + "blocks": [ + { + "bbox": [ + 54, + 198, + 555, + 234 + ], + "lines": [ + { + "bbox": [ + 54, + 198, + 555, + 234 + ], + "spans": [ + { + "bbox": [ + 54, + 198, + 555, + 234 + ], + "type": "text", + "content": "Table 3. Multi-image coherence evaluation of Spherical Stereo on TartanAir carwelding testing set. Spherical Stereo is trained self-supervised on images " + }, + { + "bbox": [ + 54, + 198, + 555, + 234 + ], + "type": "inline_equation", + "content": "\\{i,i - 1\\}" + }, + { + "bbox": [ + 54, + 198, + 555, + 234 + ], + "type": "text", + "content": " and used for inference on various image pairs, triplets and a quintuplet. Increasing the number of images improves the performance. The method naturally handles using future timesteps when trained on previous timesteps." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 125, + 243, + 486, + 306 + ], + "lines": [ + { + "bbox": [ + 125, + 243, + 486, + 306 + ], + "spans": [ + { + "bbox": [ + 125, + 243, + 486, + 306 + ], + "type": "table", + "html": "
Error - Lower is better ↓Accuracy - Higher is better ↑
MethodBaselineAbs RelSq RelRMSERMSE logδ < 1.25δ < 1.252δ < 1.253
Spherical SGMAll0.6213.827.270.600.420.640.74
Spherical SGM> 0.30.365.625.190.520.520.760.86
Spherical StereoAll0.496.996.620.650.390.610.73
Spherical Stereo> 0.30.374.475.470.450.470.720.84
", + "image_path": "aaf61b7087849fda34b797358fea5ffa202cab5de8334530749edc1bef125f07.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 312, + 555, + 348 + ], + "lines": [ + { + "bbox": [ + 54, + 312, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 54, + 312, + 555, + 348 + ], + "type": "text", + "content": "Table 4. Real World Navigation results on KITTI Eigen testing set. Spherical SGM and Spherical Stereo are used on frames with and without filtering the pose from the IMU. Pose is filtered by selecting baseline larger than " + }, + { + "bbox": [ + 54, + 312, + 555, + 348 + ], + "type": "inline_equation", + "content": "0.3\\mathrm{m}" + }, + { + "bbox": [ + 54, + 312, + 555, + 348 + ], + "type": "text", + "content": ". Error decreases when baseline is increased, substantiating Spherical rectification as proper epipolar rectification." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 358, + 295, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 295, + 384 + ], + "type": "text", + "content": "lows Spherical Stereo to take advantage of added information provided by multiple images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 390, + 186, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 186, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 186, + 403 + ], + "type": "text", + "content": "5.5. Real World Navigation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "text", + "content": "Results for absolute depth on the real world data of the KITTI Eigen split are presented in Tab. 4. They show that the proposed Spherical Stereo model solves for absolute depth in driving settings when trained without ground truth depth. KITTI Eigen is selected as a very difficult case given that ground truth pose is noisy and that motion is almost completely forward. Noisy poses prove harder to handle in small baseline cases (Fig. 7) and lead to increased error in the estimated depth. Tab. 4 and Fig. 7 show that performance improves when increasing the baseline. This confirms Spherical Stereo behaves as proper stereo matching." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 552, + 130, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 130, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 130, + 563 + ], + "type": "text", + "content": "6. Limitations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "Depth from parallax estimates the depth of objects using the displacement along epipolar lines with a static scene assumption. Objects undergoing independent motion break this assumption and induce unreliable depth estimates. This is a limitation of stereo that still holds for our proposed monocular stereo matching. Our method also does not solve disparity at the focus of expansion (Fig. 6) as it is degenerate. Since the location of this point is known given pose, it could be accounted for. The results are presented in this way because absence of matches is important information. At this point, the user can fill the depth map or not according to the task requirements. It is often more important to have" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 351, + 356, + 519, + 467 + ], + "blocks": [ + { + "bbox": [ + 351, + 356, + 519, + 467 + ], + "lines": [ + { + "bbox": [ + 351, + 356, + 519, + 467 + ], + "spans": [ + { + "bbox": [ + 351, + 356, + 519, + 467 + ], + "type": "image", + "image_path": "e15668ef20e1b4610183cfe1bbd312c7d86a39022b789889dd20958195f0637c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 475, + 555, + 533 + ], + "lines": [ + { + "bbox": [ + 313, + 475, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 555, + 533 + ], + "type": "text", + "content": "Figure 7. Accuracy " + }, + { + "bbox": [ + 313, + 475, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\delta^3" + }, + { + "bbox": [ + 313, + 475, + 555, + 533 + ], + "type": "text", + "content": " with respect to the ground truth baseline (the distance between the cameras) for Spherical Stereo trained self-supervised on KITTI Eigen. Average curve is in red. Accuracy improves as a function of baseline which shows that spherical rectification naturally takes advantage of larger disparity range." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 545, + 555, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 570 + ], + "type": "text", + "content": "a few accurate and trustworthy matches than have a dense depth map where the uncertainty distribution is unknown." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 388, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 388, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 388, + 594 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 602, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 602, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 602, + 555, + 713 + ], + "type": "text", + "content": "This paper presents a novel spherical epipolar rectification model for solving absolute depth from two views of a monocular image sequence using deep stereo matching. This differentiable model can be effectively integrated with self-supervised deep stereo matching. The proposed Spherical Crop reduces image distortion compared to previous rectification methods. Spherical Stereo allows for depth images from pairs of views that naturally account for scale and multi-view coherence, key problems of absolute depth estimation." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28932" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Jiawang Bian, Zhichao Li, Naiyan Wang, Huangying Zhan, Chunhua Shen, Ming-Ming Cheng, and Ian Reid. Unsupervised scale-consistent depth and ego-motion learning from monocular video. Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 201 + ], + "type": "text", + "content": "[2] Jia-Wang Bian, Huangying Zhan, Naiyan Wang, Tat-Jun Chin, Chunhua Shen, and Ian Reid. Auto-rectify network for unsupervised indoor depth estimation. IEEE transactions on pattern analysis and machine intelligence, 44(12):9802-9813, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 202, + 295, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 295, + 224 + ], + "type": "text", + "content": "[3] Gary Bradski, Adrian Kaehler, et al. Opencv. Dr. Dobb's journal of software tools, 3(2), 2000. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 225, + 295, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 225, + 295, + 267 + ], + "spans": [ + { + "bbox": [ + 62, + 225, + 295, + 267 + ], + "type": "text", + "content": "[4] Pierre-André Brousseau and Sébastien Roy. A permutation model for the self-supervised stereo matching problem. In 2022 19th Conference on Robots and Vision (CRV), pages 122–131. IEEE, 2022. 3, 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 270, + 295, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 295, + 302 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 295, + 302 + ], + "type": "text", + "content": "[5] Pierre-André Brousseau and Sébastien Roy. Strong stereo features for self-supervised practical stereo matching. In BMVC, page 89, 2023. 3, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 303, + 295, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 303, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 303, + 295, + 357 + ], + "type": "text", + "content": "[6] Raub Camaioni, Robert H Luke III, Andrew Buck, and Derek T Anderson. Epidepth: a real-time monocular dense-depth estimation pipeline using generic image rectification. In Geospatial Informatics XII, pages 101-114. SPIE, 2022. 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 358, + 295, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 358, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 358, + 295, + 413 + ], + "type": "text", + "content": "[7] Yuhua Chen, Cordelia Schmid, and Cristian Sminchisescu. Self-supervised learning with geometric constraints in monocular video: Connecting flow, depth, and camera. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7063-7072, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 414, + 295, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 295, + 468 + ], + "type": "text", + "content": "[8] Xuelian Cheng, Yiran Zhong, Mehrtash Harandi, Yuchao Dai, Xiaojun Chang, Hongdong Li, Tom Drummond, and Zongyuan Ge. Hierarchical neural architecture search for deep stereo matching. Advances in neural information processing systems, 33:22158-22169, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 469, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 469, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 469, + 295, + 491 + ], + "type": "text", + "content": "[9] François Darmon and Pascal Monasse. The polar epipolar rectification. Image processing on line, 11:56-75, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 491, + 295, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 491, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 57, + 491, + 295, + 545 + ], + "type": "text", + "content": "[10] David Eigen and Rob Fergus. Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. In Proceedings of the IEEE international conference on computer vision, pages 2650-2658, 2015. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 547, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 295, + 590 + ], + "type": "text", + "content": "[11] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems, 27, 2014. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 591, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 634 + ], + "type": "text", + "content": "[12] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 636, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 689 + ], + "type": "text", + "content": "[13] Clément Godard, Oisin Mac Aodha, and Gabriel J Brostow. Unsupervised monocular depth estimation with left-right consistency. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 270-279, 2017. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[14] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "text", + "content": "depth estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3828-3838, 2019. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "type": "text", + "content": "[15] Vitor Guizilini, Rares Ambrus, Dian Chen, Sergey Zakharov, and Adrien Gaidon. Multi-frame self-supervised depth with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 160-170, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "text", + "content": "[16] Heiko Hirschmuller. Accurate and efficient stereo processing by semi-global matching and mutual information. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), pages 807-814. IEEE, 2005. 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 219, + 555, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 555, + 262 + ], + "type": "text", + "content": "[17] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1621, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "type": "text", + "content": "[18] Yunseok Lee, Kar-Ann Toh, and Sangyoun Lee. Stereo image rectification based on polar transformation. Optical Engineering, 47(8):087205-087205, 2008. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 298, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 331 + ], + "type": "text", + "content": "[19] Ang Li and Zejian Yuan. Occlusion aware stereo matching via cooperative unsupervised learning. In Asian Conference on Computer Vision, pages 197-213. Springer, 2018. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 332, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 374 + ], + "type": "text", + "content": "[20] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In 2021 International Conference on 3D Vision (3DV), pages 218-227. IEEE, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 376, + 553, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 410 + ], + "type": "text", + "content": "[21] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (ToG), 39(4):71-1, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "type": "text", + "content": "[22] Kevin Matzen, Michael F Cohen, Bryce Evans, Johannes Kopf, and Richard Szeliski. Low-cost 360 stereo photography and video capture. ACM Transactions on Graphics (TOG), 36(4):1-12, 2017. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 456, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 499 + ], + "type": "text", + "content": "[23] Kimmo Palander and Sami S Brandt. Epipolar geometry and log-polar transform in wide baseline stereo matching. In 2008 19th International Conference on Pattern Recognition, pages 1-4. IEEE, 2008. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "type": "text", + "content": "[24] Marc Pollefeys, Reinhard Koch, and Luc Van Gool. A simple and efficient rectification method for general motion. In Proceedings of the Seventh IEEE International Conference on Computer Vision, pages 496-501. IEEE, 1999. 2, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 545, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 599 + ], + "type": "text", + "content": "[25] Sébastien Roy, Jean Meunier, and Ingemar J Cox. Cylindrical rectification to minimize epipolar distortion. In Proceedings of IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 393-399. IEEE, 1997. 2, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 601, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 656 + ], + "type": "text", + "content": "[26] Libo Sun, Jia-Wang Bian, Huangying Zhan, Wei Yin, Ian Reid, and Chunhua Shen. Sc-depthv3: Robust self-supervised monocular depth estimation for dynamic scenes. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 658, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 679 + ], + "type": "text", + "content": "[27] Richard Szeliski. Computer vision: algorithms and applications. Springer Nature, 2022. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[28] Julien Valentin, Adarsh Kowdle, Jonathan T Barron, Neal Wadhwa, Max Dzitsiuk, Michael Schoenberg, Vivek Verma, Ambrus Csaszar, Eric Turner, Ivan Dryanovski, et al. Depth" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28933" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "from motion for smartphone ar. ACM Transactions on Graphics (ToG), 37(6):1-19, 2018. 2, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "text", + "content": "[29] Jianyuan Wang, Yiran Zhong, Yuchao Dai, Stan Birchfield, Kaihao Zhang, Nikolai Smolyanskiy, and Hongdong Li. Deep two-view structure-from-motion revisited. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 8953-8962, 2021. 2, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 294, + 205 + ], + "type": "text", + "content": "[30] Longguang Wang, Yulan Guo, Yingqian Wang, Zhengfa Liang, Zaiping Lin, Jungang Yang, and Wei An. Parallax attention for unsupervised stereo correspondence learning. IEEE transactions on pattern analysis and machine intelligence, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 259 + ], + "type": "text", + "content": "[31] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 325 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 325 + ], + "type": "text", + "content": "[32] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4909-4916. IEEE, 2020. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 369 + ], + "type": "text", + "content": "[33] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 372, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 294, + 426 + ], + "type": "text", + "content": "[34] Jamie Watson, Oisin Mac Aodha, Victor Prisacariu, Gabriel Brostow, and Michael Firman. The temporal opportunist: Self-supervised multi-frame monocular depth. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1164-1174, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "type": "text", + "content": "[35] Gangwei Xu, Yun Wang, Junda Cheng, Jinhui Tang, and Xin Yang. Accurate and efficient stereo matching via attention concatenation volume. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(4):2461-2474, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 472, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 472, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 472, + 294, + 514 + ], + "type": "text", + "content": "[36] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, Fisher Yu, Dacheng Tao, and Andreas Geiger. Unifying flow, stereo and depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 515, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 294, + 569 + ], + "type": "text", + "content": "[37] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 2, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 571, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 571, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 571, + 295, + 624 + ], + "type": "text", + "content": "[38] Nan Yang, Lukas von Stumberg, Rui Wang, and Daniel Cremers. D3vo: Deep depth, deep pose and deep uncertainty for monocular visual odometry. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1281-1292, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 625, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 669 + ], + "type": "text", + "content": "[39] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[40] Zhichao Yin and Jianping Shi. Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1983-1992, 2018. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 351 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "type": "text", + "content": "[41] Jason J Yu, Adam W Harley, and Konstantinos G Derpanis. Back to basics: Unsupervised learning of optical flow via brightness constancy and motion smoothness. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part III 14, pages 3-10. Springer, 2016. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "type": "text", + "content": "[42] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arXiv:2410.03825, 2024. 3, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 196, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 196, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 196, + 553, + 259 + ], + "type": "text", + "content": "[43] Ning Zhang, Francesco Nex, George Vosselman, and Norman Kerle. Lite-mono: A lightweight cnn and transformer architecture for self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18537-18546, 2023. 2, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 262, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 553, + 316 + ], + "type": "text", + "content": "[44] Wang Zhao, Shaohui Liu, Yezhi Shu, and Yong-Jin Liu. Towards better generalization: Joint depth-posed learning without posenet. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9151–9161, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 318, + 553, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 318, + 553, + 351 + ], + "spans": [ + { + "bbox": [ + 316, + 318, + 553, + 351 + ], + "type": "text", + "content": "[45] Yiran Zhong, Yuchao Dai, and Hongdong Li. Self-supervised learning for stereo matching with self-improving ability. arXiv preprint arXiv:1709.00930, 2017. 3" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28934" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SpiLiFormer_ Enhancing Spiking Transformers with Lateral Inhibition/75735d8f-465e-40da-8f9a-a276e0092626_content_list.json b/2025/SpiLiFormer_ Enhancing Spiking Transformers with Lateral Inhibition/75735d8f-465e-40da-8f9a-a276e0092626_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d65446b85b37d6ca27d581acbdd63405740997d6 --- /dev/null +++ b/2025/SpiLiFormer_ Enhancing Spiking Transformers with Lateral Inhibition/75735d8f-465e-40da-8f9a-a276e0092626_content_list.json @@ -0,0 +1,1573 @@ +[ + { + "type": "text", + "text": "SpiLiFormer: Enhancing Spiking Transformers with Lateral Inhibition", + "text_level": 1, + "bbox": [ + 135, + 128, + 859, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zeqi Zheng $^{1,2*}$ , Yanchen Huang $^{2,3*}$ , Yingchao Yu $^{4,2}$ , Zizheng Zhu $^{2,5}$ , Junfeng Tang $^{1,2}$ , Zhaofei Yu $^{6}$ , Yaochu Jin $^{2\\dagger}$", + "bbox": [ + 227, + 178, + 767, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Zhejiang University 2Westlake University 3Nanjing University 4Donghua University 5University of Electronic Science and Technology of China 6Peking University", + "bbox": [ + 133, + 215, + 861, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhengzeqi, huangyanchen, jinyaochu}@westlake.edu.cn", + "bbox": [ + 267, + 253, + 728, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 303, + 326, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Spiking Neural Networks (SNNs) based on Transformers have garnered significant attention due to their superior performance and high energy efficiency. However, the spiking attention modules of most existing Transformer-based SNNs are adapted from those of analog Transformers, failing to fully address the issue of over-allocating attention to irrelevant contexts. To fix this fundamental yet overlooked issue, we propose a Lateral Inhibition-inspired Spiking Transformer (SpiLiFormer). It emulates the brain's lateral inhibition mechanism, guiding the model to enhance attention to relevant tokens while suppressing attention to irrelevant ones. Our model achieves state-of-the-art (SOTA) performance across multiple datasets, including CIFAR-10 (+0.45%), CIFAR-100 (+0.48%), CIFAR10-DVS (+2.70%), N-Caltech101 (+1.94%), and ImageNet-1K (+1.6%)1. Notably, on the ImageNet-1K dataset, SpiLiFormer (69.9M parameters, 4 time steps, 384 resolution) outperforms E-SpikeFormer (173.0M parameters, 8 time steps, 384 resolution), a SOTA spiking Transformer, by 0.46% using only 39% of the parameters and half the time steps. The code and model checkpoints are publicly available at https://github.com/KirinZheng/SpiLiFormer.", + "bbox": [ + 89, + 335, + 485, + 670 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 700, + 220, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Spiking Neural Networks (SNNs), regarded as the third generation of neural networks [18], are seen as a potential alternative to Artificial Neural Networks (ANNs) due to their biological interpretability and high energy efficiency, which stem from their event-driven properties. Transformer [25], originally designed for natural language processing tasks, has now become a dominant neural network architecture, demonstrating remarkable performance across", + "bbox": [ + 89, + 727, + 483, + 849 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "various visual tasks [6, 27, 32, 42]. The success of Transformers has driven the exploration of their integration with SNNs, with the objective of enhancing SNN models' performance across various tasks and progressively bridging the performance gap with ANNs, particularly in image classification [29, 30, 37, 38], object detection [10, 17], action recognition [14, 26], and semantic segmentation [16, 29].", + "bbox": [ + 511, + 305, + 906, + 412 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The attention module plays a crucial role in Transformer-based SNNs, significantly influencing model performance. Recent research advancements in spiking attention modules can be broadly classified into three categories: 1) Self-Spiking Attention (SSA) [36, 38, 39], which directly converts the Query $(Q)$ , Key $(K)$ , and Value $(V)$ into sparse spikes; 2) Spike-Driven Self-Attention (SDSA) [29, 30] and Spiking RWKV (S-RWKV) [40], which replace the dot product in SSA with the Hadamard product, achieving linear complexity while reducing both energy consumption and computational cost; 3) Spatial-Temporal Attention (STA) [12, 28], which refines traditional spatial attention by explicitly modeling the key temporal dependencies intrinsic to spike-based processing.", + "bbox": [ + 511, + 412, + 908, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, we observe that most Transformer-based SNNs exhibit a phenomenon referred to as attention distraction. As illustrated in Fig. 1(a), the model disproportionately assigns high attention weights to background information during decision-making, leading to the neglect of critical object-related features and ultimately impairing classification performance. This issue stems from the fact that spiking attention largely inherits the traditional attention forward propagation paradigm. Previous studies [1, 21, 34] suggest that traditional attention mechanisms process $Q$ , $K$ , and $V$ uniformly in the mapping $softmax(QK^{\\top})V$ , thereby limiting their capacity to regulate contextual sparsity and relevance. Therefore, we posit that constructing additional pathways to process $Q$ , $K$ , and $V$ separately is essential for mitigating attention distraction in Transformer-based SNNs and enhancing model performance.", + "bbox": [ + 511, + 626, + 908, + 882 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "By contrast, the natural lateral inhibition mechanism in", + "bbox": [ + 531, + 885, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. † Corresponding author.", + "bbox": [ + 112, + 862, + 362, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "These results are obtained by comparing with state-of-the-art spiking neural network models that have a similar number of parameters.", + "bbox": [ + 91, + 876, + 482, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24539", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e2b7f11365532398f9418e66da9227cbd6113317d9026495a99ec3d312a8fe34.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 137, + 90, + 625, + 265 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c315987cd2963b7aa7a9e77a289ba5440679566604926badceccb975f8e6d604.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 112, + 872, + 266 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0b579d9f6094e0ca852c1da4d8c5e83e470655fa8ef46df101cf1b81666bd828.jpg", + "image_caption": [ + "Figure 1. Illustration of the severe attention distraction phenomenon in Transformer-based SNNs, the architecture of the Lateral Inhibition-inspired Spiking Transformer (SpiLiFormer), and the detailed processing mechanism of Feedforward-pathway Lateral Differential Inhibition (FF-LiDiff) attention and Feedback-pathway Lateral Difference Inhibition (FB-LiDiff) attention. (a) refers to the attention distraction in mainstream Transformer-based SNNs, where models excessively focus on irrelevant background information, leading to misclassification. (b) denotes the retinal lateral inhibition mechanism, illustrating how horizontal cells regulate neural responses to enhance contrast and suppress noise. (c) presents the architecture of SpiLiFormer. (d) and (e) represent the information processing flows of FF-LiDiff and FB-LiDiff attention, respectively." + ], + "image_footnote": [], + "bbox": [ + 101, + 271, + 893, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the visual system enables the brain to focus on important areas, thereby improving perception and reducing visual overload [4, 19, 22]. Specifically, when neurons within a given area are strongly activated, adjacent neurons are inhibited. This principle is first observed in the retina, where horizontal cells receive input from photoreceptors and inhibit both the stimulated photoreceptors and their neighbors, forming a center-surround receptive field (RF), as shown in Fig. 1(b). In higher visual areas, such as the primary (V1) and secondary (V2) visual cortices, a similar mechanism operates through long-range connections, which contrasts with the short-range inhibition in the retina, helping to refine visual", + "bbox": [ + 89, + 704, + 485, + 886 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "processing by suppressing irrelevant stimuli and selectively enhancing salient features [41, 43].", + "bbox": [ + 511, + 704, + 906, + 736 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the short-range and long-range lateral inhibition mechanisms discussed above, we propose the Lateral Inhibition-inspired Spiking Transformer (SpiLiFormer) to address the attention distraction issue in Transformer-based SNNs, as shown in Fig. 1(c). This model consists of two types of stacked modules: 1) Feedforward-pathway Lateral Differential Inhibition (FF-LiDiff) blocks and 2) Feedback-pathway Lateral Differential Inhibition (FB-LiDiff) blocks. Moreover, it introduces the following innovations: 1) In the shallow blocks (i.e., Stage 1 and Stage 2), it explicitly in-", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24540", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "corporates an inhibition differential attention mechanism by emulating the retinal short-range lateral inhibition process (see Fig. 1(d) for details); 2) In the deep blocks (i.e., Stage 3), it employs feedback-driven processing based on differential attention to simulate the long-range lateral inhibition mechanism observed in the cerebral cortex (see Fig. 1(e) for details). We evaluate our proposed SpiLiFormer on five image classification datasets, including CIFAR-10 [11], CIFAR-100 [11], CIFAR10-DVS [13], N-Caltech101 [20], and ImageNet-1K [5], demonstrating superior performance over current SOTA models. Additionally, adversarial testing and attention heatmap visualizations on the ImageNet-1K test set and noise-augmented datasets (i.e., CIFAR-10-C and ImageNet-1K-C [8]) show that SpiLiFormer effectively suppresses attention on irrelevant areas and background noise. The main contributions of this work are:", + "bbox": [ + 89, + 90, + 480, + 347 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose SpiLiFormer, which incorporates a brain-inspired lateral inhibition mechanism and introduces two novel attention paradigms: FF-LiDiff attention and FB-LiDiff attention. This model effectively mitigates the attention distraction issue common in current Transformer-based SNNs, thereby enhancing image classification performance.", + "We evaluate SpiLiFormer across multiple datasets, achieving new SOTA performance on CIFAR-10, CIFAR100, CIFAR10-DVS, N-Caltech101, and ImageNet-1K. Specifically, on ImageNet-1K, SpiLiFormer (69.9M parameters, 4 time steps) outperforms the optimal SOTA model, E-SpikeFormer (173.0M parameters, 8 time steps), by $0.46\\%$ in top-1 accuracy, while using only 0.39 times the number of parameters and half the time steps.", + "- We comprehensively evaluate the robustness of SpiLiFormer through attention heatmap visualizations and adversarial testing, and find that SpiLiFormer achieves more effective attention allocation and enhanced robustness compared to other baseline spiking models." + ], + "bbox": [ + 89, + 352, + 480, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 670, + 230, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Transformer-based SNNs", + "text_level": 1, + "bbox": [ + 89, + 696, + 320, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although convolution-based SNNs exhibit high energy efficiency, they still demonstrate a significant performance gap compared to ANNs. To address this challenge, researchers have explored integrating Transformer architectures with SNNs. Spikformer [38] was the first to incorporate the Transformer attention into SNNs, cleverly replacing the Softmax operation by leveraging the binary characteristic of spike representations to encode Query, Key, and Value. Building upon this, Spikingformer [36] introduced a pre-activation shortcut to avoid floating-point multiplication while simultaneously reducing the spike firing rate. Inspired by SEW-ResNet [7], Spike-driven Trans", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "former [30] replaced the dot product with a Hadamard product and reshaped the spike residual connection based on the membrane potential to further minimize energy consumption and leverage the event-driven property. Additionally, Spiking RWKV [40] and QKFormer [37] introduced linear attention into SNNs to reduce computational complexity. Meanwhile, STA-Transformer [12] effectively integrated both spatial and temporal dynamics of spike trains to enhance model performance. However, most existing Transformer-based SNNs still adhere to the traditional attention propagation paradigm of ANNs, which results in the severe phenomenon of attention distraction, thereby limiting model performance.", + "bbox": [ + 511, + 90, + 903, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Models with Lateral Inhibition", + "text_level": 1, + "bbox": [ + 511, + 299, + 785, + 314 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Lateral inhibition mechanisms have been widely adopted in various SNN applications. In speech recognition, Spiking-LEAF [23] emulated inner hair cell functionality and leveraged lateral inhibition feedback to enhance the neuron model, significantly improving the efficiency and noise robustness of spike encoding. In the object recognition, LISNN [2] modeled lateral inhibition by establishing spatial neighbor-order relationships between neurons, thereby enhancing the model's ability to focus on important features. In the image classification, Zhang et al. [33] proposed an adaptive self-organizing lateral inhibition strategy, where inhibition strength was adjusted according to the Euclidean distance between neurons. This approach addressed the issue of inefficient feature clustering and reduced redundancy in neuron activation. However, existing research on lateral inhibition has predominantly focused on small-scale datasets for validating model feasibility, with limited exploration of its applicability to larger and more complex datasets. Furthermore, its integration with Transformer-based SNNs remains largely unexplored.", + "bbox": [ + 511, + 321, + 906, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 640, + 604, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section details the proposed SpiLiFormer model, covering the spiking neuron layer, overall architecture, and two lateral inhibition-inspired attention mechanisms. Additionally, we outline the training strategy and loss function.", + "bbox": [ + 511, + 666, + 903, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Spiking Neuron Layer", + "text_level": 1, + "bbox": [ + 511, + 738, + 720, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The spiking neuron layer relies on cumulative firing to integrate spatio-temporal information into the membrane potential, which is then converted into binary spikes to drive computation in the next layer. We adopt the widely used Leaky Integrate-and-Fire (LIF) spiking neuron layer, as established in previous studies [30, 31, 37], with the following dynamics:", + "bbox": [ + 511, + 761, + 903, + 867 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nU [ t ] = H [ t ] (1 - S [ t ]) + U _ {\\text {r e s e t}} S [ t ], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 883, + 903, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24541", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nH [ t ] = U [ t - 1 ] + \\frac {1}{\\tau} (X [ t ] - (U [ t - 1 ] - U _ {\\text {r e s e t}})), \\quad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 114, + 483, + 142 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS [ t ] = \\Theta (H [ t ] - U _ {t h}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 164, + 482, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $U_{\\mathrm{reset}}$ represents the reset potential, and $X[t]$ and $U[t]$ denote the current input and membrane potential at timestep $t$ , respectively. $\\Theta(\\cdot)$ is the Heaviside step function, which outputs 1 if $x \\geq 0$ and 0 otherwise. When $H[t]$ exceeds the firing threshold $U_{th}$ , the spiking neuron fires a spike $S[t]$ , and $U[t]$ is reset to $U_{\\mathrm{reset}}$ at that moment. Additionally, $\\tau$ represents the membrane time constant, which controls the rate of membrane potential leakage.", + "bbox": [ + 89, + 186, + 483, + 308 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Overall Architecture", + "text_level": 1, + "bbox": [ + 89, + 318, + 287, + 332 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall architecture of SpiLiFormer is illustrated in Fig. 1(c), which consists of three stages: Stage 1, Stage 2, and Stage 3, forming a three-level hierarchical structure. The input data is represented as $I \\in \\mathbb{R}^{T \\times N \\times H \\times W}$ , where $N = 3$ for static image data and $N = 2$ for neuromorphic data. In Stage 1, a $4 \\times 4$ patch size is utilized, mapping each patch's input feature dimension $(4 \\times 4 \\times N)$ to a spike-form representation of arbitrary dimension $(C)$ via the Downsampling module. This process reduces the number of tokens to $\\frac{H}{4} \\times \\frac{W}{4}$ . The transformed features are then processed through FF-LiDiff attention and a Spiking MultiLayer Perceptron (SpMLP). FF-LiDiff attention introduces an additional forward pathway and employs a differential approach to mitigate attention distraction in the network's shallower layers (see Sec. 3.3 for details). In Stage 2, the processing mechanism is similar to Stage 1, but with a $2 \\times 2$ patch size, an expanded channel dimension of $2C$ , and the number of tokens reduced to $\\frac{H}{8} \\times \\frac{W}{8}$ . Finally, in Stage 3, we retain the $2 \\times 2$ patch size, expand the channel dimension to $4C$ , and reduce the number of tokens to $\\frac{H}{16} \\times \\frac{W}{16}$ . Additionally, a more complex FB-LiDiff attention module with feedback loops is introduced (see Sec. 3.4 for details) to replace the previous FF-LiDiff attention, further mitigating attention distraction in the model.", + "bbox": [ + 89, + 339, + 483, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The number of spiking blocks (either FF-LiDiff block or FB-LiDiff block) in each stage is denoted as $N_{1}$ , $N_{2}$ , and $N_{3}$ , respectively. Since the Downsampling, SpMLP and classification head modules are based on prior research works [29, 31, 37], the following sections will focus on the detailed description of FF-LiDiff attention and FB-LiDiff attention.", + "bbox": [ + 89, + 702, + 483, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Feedforward-pathway Lateral Differential Inhibition Attention", + "text_level": 1, + "bbox": [ + 89, + 816, + 482, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An overview of FF-LiDiff attention is shown in Figure 1(d). For better mathematical description, in subsequent discussion, we assume $T = 1$ and only use single head attention.", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a spike input $X \\in \\mathbb{R}^{N \\times D}$ ( $N$ is the token number, $D$ is the channel number), Query $(Q)$ and Key $(K)$ can be computed through learnable matrices:", + "bbox": [ + 511, + 90, + 905, + 136 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ = B N \\left(X W _ {Q}\\right), K = B N \\left(X W _ {K}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 148, + 906, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W_{Q}, W_{K} \\in \\mathbb{R}^{D \\times D}$ and $Q, K \\in \\mathbb{R}^{N \\times D}$ .", + "bbox": [ + 511, + 176, + 823, + 193 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To implement the lateral inhibition mechanism, we first split $Q$ along the channel dimension into two parts, obtaining $Q_{1}$ and $Q_{2}$ , before computing the attention matrix. Then, all of $Q_{1}, Q_{2}$ , and $K$ are transformed into spike trains via their respective spiking neuron layers:", + "bbox": [ + 511, + 194, + 905, + 270 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {s 1} = \\mathcal {S N} (Q _ {1}), Q _ {s 2} = \\mathcal {S N} (Q _ {2}), K _ {s} = \\mathcal {S N} (K), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 282, + 906, + 299 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Q_{s1}, Q_{s2} \\in \\mathbb{R}^{N \\times \\frac{D}{2}}$ , $K_{s} \\in \\mathbb{R}^{N \\times D}$ . Next, we sum along the channel dimension to obtain the excitatory attention $A_{e}$ and inhibitory attention $A_{i}$ . Specifically, $A_{e}$ and $A_{i}$ are $N \\times 1$ token attention vectors, representing the importance of different tokens from the perspectives of excitation and inhibition, respectively. Finally, we compute the element-wise difference between $A_{e}$ and $A_{i}$ , and pass the aggregated attention through a spiking neural layer to obtain binary attention $A_{\\text{combined}}$ . This attention is then applied as a token-wise mask to $K_{s}$ via the Hadamard product $\\otimes$ . The above process can be described as:", + "bbox": [ + 511, + 311, + 905, + 479 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA _ {e} = \\sum_ {j} ^ {\\frac {D}{2}} Q _ {s 1} ^ {i, j}, A _ {i} = \\sum_ {j} ^ {\\frac {D}{2}} Q _ {s 2} ^ {i, j}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 491, + 905, + 539 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA _ {\\text {c o m b i n e d}} = \\mathcal {S N} \\left(A _ {e} - A _ {i}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 558, + 903, + 574 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX ^ {\\prime} = A _ {\\text {c o m b i n e d}} \\otimes K _ {s}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 599, + 903, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Feedback-pathway Lateral Differential Inhibition Attention", + "text_level": 1, + "bbox": [ + 511, + 626, + 905, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "FB-LiDiff attention requires two forward propagations as shown in Figure 1(e). Specifically, during the first forward propagation, FB-LiDiff attention takes spike-form $S_{\\mathrm{stage3}}^{Ds}$ from the Downsampling module in Stage 3 or takes spikeform $S_{\\mathrm{stage3}}^i (i \\in [1, N_3])$ from the $i$ -th preceding FB-LiDiff block as input. Given an input $X_1^m$ to the $m$ -th FB-LiDiff attention, the first forward propagation can be described as follows:", + "bbox": [ + 511, + 662, + 905, + 782 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nY _ {s} ^ {1, m} = \\mathcal {S N} \\left(B N \\left(X _ {1} ^ {m} W _ {Y} ^ {m}\\right)\\right), Y \\in \\{Q, K, V \\}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 796, + 906, + 814 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA t t n ^ {1, m} = Q _ {s} ^ {1, m} \\odot \\left(\\left(K _ {s} ^ {1, m}\\right) ^ {\\top} \\odot V _ {s} ^ {1, m}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 840, + 903, + 859 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nO _ {\\text {s t a g e 3}} ^ {m} = \\mathcal {S N} \\left(B N \\left(A t t n ^ {1, m} W _ {a t t n} ^ {m}\\right)\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 883, + 903, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24542", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $X_{1}^{m}\\in \\mathbb{R}^{N\\times D}$ ( $X_{1}^{m}\\in \\{S_{\\mathrm{stage3}}^{Ds},S_{\\mathrm{stage3}}^{i}\\}$ ), and $Y_{s}^{1,m}$ ( $Y\\in \\{Q,K,V\\}$ ) represents the corresponding spiking representation of $Q$ , $K$ and $V$ in the $m$ -th FB-LiDiff attention. $W_{z}^{m}$ ( $z\\in \\{Q,K,V,attn\\}$ ) and $O_{\\mathrm{stage3}}^{m}$ denote the learnable weight matrix and the spike-form output of the $m$ -th FB-LiDiff attention, respectively. $\\odot$ refers to the dot product.", + "bbox": [ + 89, + 89, + 480, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Before the second forward propagation, we use the output $O_{\\mathrm{stage3}}^{N_3}$ from the last block of Stage 3 as feedback information and return it to each FB-LiDiff attention. To filter the feedback information and further mitigate the attention distraction problem, we first introduce learnable prompts $W_{p1}$ and $W_{p2}$ , where $W_{p1}, W_{p2} \\in \\mathbb{R}^{1 \\times D}$ . Next, we compute the dot product between $W_{p1}$ and $O_{\\mathrm{stage3}}^{N_3}$ , and similarly between $W_{p2}$ and $O_{\\mathrm{stage3}}^{N_3}$ . These dot products are then broadcast along the token dimension. The results are scaled to the range [0, 1] using the clamp(\\cdot) function, forming the excitatory feedback attention $A_e^{\\mathrm{FB}}$ and inhibitory feedback attention $A_i^{\\mathrm{FB}}$ . Subsequently, we compute the element-wise difference between $A_e^{\\mathrm{FB}}$ and $A_i^{\\mathrm{FB}}$ to obtain the aggregated attention $A_{\\mathrm{combined}}^{\\mathrm{FB}}$ . Finally, we apply a linear transformation to $A_{\\mathrm{combined}}^{\\mathrm{FB}}$ and pass it through the spiking neuron layer to align it with the feature space of different FB-LiDiff attentions. These processes are formulated as:", + "bbox": [ + 91, + 183, + 483, + 443 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA _ {e} ^ {\\mathrm {F B}} = \\operatorname {c l a m p} \\left(O _ {\\text {s t a g e 3}} ^ {N _ {3}} \\odot W _ {p 1}, 0, 1\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 455, + 482, + 476 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA _ {i} ^ {\\mathrm {F B}} = \\operatorname {c l a m p} \\left(O _ {\\text {s t a g e 3}} ^ {N _ {3}} \\odot W _ {p 2}, 0, 1\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 489, + 482, + 512 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA _ {\\text {c o m b i n e d}} ^ {\\mathrm {F B}} = A _ {e} ^ {\\mathrm {F B}} - A _ {i} ^ {\\mathrm {F B}}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 526, + 480, + 546 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nX _ {\\mathrm {F B}} ^ {m} = \\mathcal {S N} \\left(B N \\left(A _ {\\text {c o m b i n e d}} ^ {\\mathrm {F B}} W _ {\\mathrm {F B}} ^ {m}\\right)\\right). \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 561, + 480, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The second forward propagation begins by taking the processed forward representation $S_{\\mathrm{stage3}}^{Ds}$ and the corresponding feedback information $X_{\\mathrm{FB}}^{1}$ as input to the first FB-LiDiff attention. The objective is to reduce model energy consumption by eliminating redundant computations in the non-feedback loop (i.e., Stage 1 and Stage 2). A detailed energy consumption analysis can be found in Appendix A.3. Additionally, during the second forward propagation, we share all parameter weights across the FB-LiDiff blocks to minimize the number of model parameters. Given the input $X_{2}^{m}$ , the process described above can be formulated as:", + "bbox": [ + 89, + 584, + 482, + 752 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nY _ {s} ^ {2, m} = \\mathcal {S N} \\left(B N \\left(X _ {2} ^ {m} W _ {Y} ^ {m}\\right)\\right), Y \\in \\{Q, K \\}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 773, + 482, + 792 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nV _ {s} ^ {2, m} = \\mathcal {S N} \\left(B N \\left(X _ {2} ^ {m} W _ {V} ^ {m} + X _ {\\mathrm {F B}} ^ {m} W _ {V} ^ {m}\\right)\\right), \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 811, + 482, + 830 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA t t n ^ {2, m} = Q _ {s} ^ {2, m} \\odot \\left(\\left(K _ {s} ^ {2, m}\\right) ^ {\\top} \\odot V _ {s} ^ {2, m}\\right), \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 847, + 482, + 867 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(O _ {\\text {s t a g e} 3} ^ {m}\\right) ^ {\\prime} = \\mathcal {S N} \\left(B N \\left(A t t n ^ {2, m} W _ {a t t n} ^ {m}\\right)\\right). \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 883, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Training Strategy and Loss Function", + "text_level": 1, + "bbox": [ + 513, + 90, + 831, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We employ a surrogate gradient-based method to overcome the non-differentiability of the spiking neuron activation in Eq. (3), enabling direct training of the model from scratch, as demonstrate in previous works [29, 36-38, 40]. In order to make better use of the spiking representations obtained from two forward propagations, we adjust the model's loss function as follows:", + "bbox": [ + 511, + 112, + 903, + 217 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {1} = \\mathcal {L} _ {C E} \\left(C H \\left(O _ {\\text {s t a g e 3}} ^ {N _ {3}}\\right), y\\right), \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 231, + 903, + 251 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {2} = \\mathcal {L} _ {C E} \\left(C H \\left(\\left(O _ {\\text {s t a g e 3}} ^ {N _ {3}}\\right) ^ {\\prime}\\right), y\\right), \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 268, + 903, + 289 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {S p i L i F o r m e r}} = \\alpha \\mathcal {L} _ {1} + (1 - \\alpha) \\mathcal {L} _ {2}, \\tag {22}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 309, + 903, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ is a hyperparameter that dynamically balances the losses from two forward propagations. Unless explicitly specified otherwise, $\\alpha$ is set to 0.5 in the subsequent experiments. The choice of this value is validated through an ablation study, as detailed in Appendix A.4. Additionally, $y$ represents the true label; $O_{\\mathrm{stage3}}^{N_3}$ and $(O_{\\mathrm{stage3}}^{N_3})'$ denote the outputs from the first and second forward propagations of the last FB-LiDiff block, respectively; $CH(\\cdot)$ represents the classification head module, as described in previous studies [29, 30]; and $\\mathcal{L}_{CE}$ denotes the cross-entropy loss function.", + "bbox": [ + 511, + 332, + 906, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 511, + 645, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first evaluate the classification performance of SpiLiFormer model on the large-scale ImageNet-1K dataset [5]. Next, we assess the performance of SpiLiFormer on smaller-scale static datasets, including CIFAR-10 [11] and CIFAR-100 [11]. Additionally, we evaluate SpiLiFormer on two popular neuromorphic datasets: CIFAR10-DVS [13] and N-Caltech101 [20]. Finally, we visualize attention heatmaps and conduct adversarial tests to demonstrate SpiLiFormer's effectiveness in mitigating attention distraction and enhancing robustness.", + "bbox": [ + 511, + 537, + 906, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Results on ImageNet-1K", + "text_level": 1, + "bbox": [ + 511, + 696, + 736, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this experiment, model training is divided into two phases. In the first phase, we use the AdamW optimizer with a base learning rate of $6 \\times 10^{-4}$ . The actual learning rate is calculated by multiplying the base rate by the batch size divided by 256. Additionally, we apply a layer-wise learning rate decay strategy with a decay factor of 1.0 and weight decay of 0.05 to train SpiLiFormer for 200 epochs. In the second phase, we decrease the base learning rate to $2 \\times 10^{-6}$ and continue training for an additional 20 epochs using an epoch-wise learning rate decay strategy. Throughout this entire experiment, we incorporate several data augmentation techniques in alignment with the DeiT approach", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24543", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/134c7e1469e15cf241fc6e8178e29aef71d63b29abe545ebb5b41ba43ee8928c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsTypeArchitectureInput SizeParam(M)Power(mJ)Time StepTop-1 Acc(%)
ViT [6]ANNViT-B/1638486.59254.84177.90
DeiT [24]ANNDeiT-B22486.5980.50181.80
ANNDeiT-B38486.59254.84183.10
Swin [15]ANNSwin Transformer-B22487.7770.84183.50
ANNSwin Transformer-B38487.77216.20184.50
SEW ResNet [7]SNNSEW-ResNet-15222460.1912.89469.26
Spikformer [38]SNNSpikformer-8-76822466.3421.48474.81
Spikingformer [36]SNNSpikingformer-8-76822466.3413.68475.85
S-Transformer [30]SNNS-Transformer-8-51222429.681.13171.68
SNNS-Transformer-8-51222429.684.50474.57
SNNS-Transformer-8-76828866.346.09477.07
QKFormer [37]SNNHST-10-76822464.9638.91484.22
SNNHST-10-768*28864.9664.27485.25
SNNHST-10-76838464.96113.64485.65
E-SpikeFormer [31]SNNE-SpikeFormer224173.035.64*84.7
SNNE-SpikeFormer224173.054.78*85.1
SNNE-SpikeFormer384173.0-8*86.2
SpiLiFormer (Ours)SNNSpiLiFormer-10-76822469.1011.77181.54
SNNSpiLiFormer-10-76822469.1044.17485.82
SNNSpiLiFormer-10-768*28869.1073.52486.62
SNNSpiLiFormer-10-768**38469.10129.45486.66
", + "bbox": [ + 91, + 88, + 893, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Performance comparison on ImageNet-1K. \"SpiLiFormer-L-D\" represents Lateral Inhibition-inspired Transformer with $L$ blocks and a $D$ -dimensional channel. \\* and \\*\\* denote input resolutions of $288^{2}$ and $384^{2}$ during inference, respectively. \\* denotes an integer-valued quantized training approach [17], in which activation values are restricted to integers during training and later expanded into spike trains during inference.", + "bbox": [ + 89, + 481, + 906, + 539 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[24]. These techniques include RandAugment [3], random erasing [35], and stochastic depth [9]. The architecture of SpiLiFormer is composed of 1, 2, and 7 blocks across its three stages, respectively.", + "bbox": [ + 89, + 564, + 482, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 1 demonstrates the superior performance of our proposed SpiLiFormer, significantly outperforming current SNNs models. Specifically, SpiLiFormer (69.10M) achieves $86.66\\%$ top-1 accuracy and $98.116\\%$ top-5 accuracy on ImageNet-1K. To our best knowledge, SpiLiFormer is currently the SOTA model for Transformer-based SNNs with direct training.", + "bbox": [ + 89, + 627, + 482, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare SpiLiFormer against several baseline spiking models under the same image resolution of 224, a time step of 4, and a comparable number of model parameters. SpiLiFormer outperforms Spikformer (66.34M), Spikingformer (66.34M), and QKFormer (64.96M) by $11.01\\%$ , $9.97\\%$ , and $1.6\\%$ , respectively. Moreover, at a higher input resolution of 288, SpiLiFormer outperforms S-Transformer and QKFormer by $9.55\\%$ and $1.37\\%$ , respectively. Similarly, at a resolution of 384, it surpasses QKFormer by $1.01\\%$ . Additionally, we compare SpiLiFormer (69.10M, 4 time steps, 384 resolution) with E-SpikeFormer (173.0M, 8", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "time steps, 384 resolution), the current SOTA Transformer-based SNNs model. It achieves a $0.46\\%$ improvement with only $39\\%$ of parameters and half the time steps.", + "bbox": [ + 511, + 564, + 906, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, compared to the mainstream ANN-based Swin Transformer-B (384 resolution), SpiLiFormer achieves a $2.16\\%$ performance improvement while reducing energy consumption by $40\\%$ and decreasing the number of parameters by 18.67M.", + "bbox": [ + 511, + 609, + 906, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results on CIFAR Datasets", + "text_level": 1, + "bbox": [ + 511, + 696, + 759, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this experiment, we use the AdamW optimizer with a learning rate of $1 \\times 10^{-3}$ and apply a cosine learning rate decay strategy. The weight decay is $6 \\times 10^{-2}$ , the number of time steps is 4, and SpiLiFormer is trained for 400 epochs with a batch size of 64. The network architecture consists of 1, 1, and 2 blocks across its three stages, respectively.", + "bbox": [ + 511, + 719, + 906, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The performance of SpiLiFormer on CIFAR datasets is shown in the upper part of Tab. 2. For CIFAR-10, our proposed model achieves $96.63\\%$ accuracy with 7.04M parameters. Compared to QKFormer, the suboptimal SOTA model, SpiLiFormer improves performance by $0.45\\%$ while increasing the parameters by only $0.04\\%$ . Similarly, for", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24544", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/689d1814a416876ee8ad91f685b255a69158e20a2c229dfd8aba614615813036.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetsMethodsArchitectureParam(M)Time StepTop-1 Acc(%)
CIFAR-10Spikformer [38]Spikformer-4-3849.32495.51
Spikingformer [36]Spikingformer-4-3849.32495.81
S-Transformer [30]S-Transformer-2-51210.28495.60
QKFormer [37]HST-4-3846.74496.18
SpiLiFormer (Ours)SpiLiFormer-4-3847.04496.63
CIFAR-100Spikformer [38]Spikformer-4-3849.32478.21
Spikingformer [36]Spikingformer-4-3849.32479.21
S-Transformer [30]S-Transformer-2-51210.28478.4
QKFormer [37]HST-4-3846.74481.15
SpiLiFormer (Ours)SpiLiFormer-4-3847.04481.63
CIFAR10-DVSSpikformer [38]Spikformer-2-2562.571680.9
Spikingformer [36]Spikingformer-2-2562.571681.3
S-Transformer [30]S-Transformer-2-2562.571680.0
QKFormer [37]HST-2-2561.501684.0
SpiLiFormer (Ours)SpiLiFormer-2-2561.571686.7
N-Caltech101Spikformer [38]Spikformer-2-2562.571683.6
Spikingformer [36]Spikingformer-2-2562.571685.91
S-Transformer [30]S-Transformer-2-2562.571686.3
QKFormer [37]HST-2-2561.501687.24
SpiLiFormer (Ours)SpiLiFormer-2-2561.571689.18
", + "bbox": [ + 114, + 88, + 883, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Performance comparison on CIFAR-10, CIFAR-100, CIFAR10-DVS, and N-Caltech101 datasets.", + "bbox": [ + 181, + 435, + 815, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CIFAR-100, our proposed model achieves $81.63\\%$ accuracy with 7.04M parameters. Specifically, it outperforms QKFormer by $0.48\\%$ while requiring just $0.04\\%$ more parameters.", + "bbox": [ + 89, + 476, + 482, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Results on Neuromorphic Datasets", + "text_level": 1, + "bbox": [ + 89, + 546, + 393, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this experiment, we use a smaller SpiLiFormer architecture, structured with 0, 1, and 1 blocks across its three stages, respectively. Similarly, we adopt the AdamW optimizer with an initial learning rate of $5 \\times 10^{-3}$ , apply a learning rate decay strategy, and set the weight decay to 0.06. The number of time steps is set to 16, and SpiLiFormer is trained for 120 epochs.", + "bbox": [ + 89, + 568, + 482, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The experimental results of the model on the neuromorphic datasets are presented in the lower half of Tab. 2. Compared to the suboptimal model, QKFormer, SpiLi-Former significantly improves performance by $2.7\\%$ on the CIFAR10-DVS dataset and $1.94\\%$ on the N-Caltech101 dataset, with only a $0.13\\%$ increase in parameters. Additionally, we compare the proposed model with other baseline models (i.e., Spikformer, Spikingformer, and S-Transformer). It achieves performance gains of up to $5.8\\%$ on CIFAR10-DVS and $5.55\\%$ on N-Caltech101 while reducing the parameter count by 0.87M.", + "bbox": [ + 89, + 674, + 482, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Visualization and Model Robustness Analysis", + "text_level": 1, + "bbox": [ + 89, + 848, + 475, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We visualize the attention heatmap of SpiLiFormer to assess whether it effectively alleviates the attention distraction is", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sue. Specifically, we examine the attention heatmap of the last FB-LiDiff block in SpiLiFormer Stage 3 under two conditions: standard dataset evaluation and noisy dataset testing.", + "bbox": [ + 511, + 476, + 906, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the standard dataset evaluation, the two sets of images in Fig. 1(a) show examples that are misclassified by the baseline model but correctly classified by SpiLiFormer on the ImageNet-1K test set. We observe that our proposed model effectively eliminates the attention distraction phenomenon, thereby improving classification accuracy. More comparison images are provided in Fig. 4 within the appendix.", + "bbox": [ + 511, + 551, + 906, + 674 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In noisy dataset testing, we evaluate model attention heatmaps using two noisy datasets: CIFAR10-C and ImageNet-1K-C, as shown in Fig. 2. Compared to other baseline models, SpiLiFormer correctly classifies objects or significantly increases confidence in the correct category by focusing on object-relevant information while remaining unaffected by background noise.", + "bbox": [ + 511, + 689, + 908, + 796 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition, we employ white-box adversarial attacks, including FGSM and PGD, to assess the model's robustness across the four datasets mentioned above. As shown in Fig. 3, we find that SpiLiFormer not only enhances model performance but also improves robustness against adversarial and common noise attacks, outperforming QKFormer.", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24545", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3ee79235ecfff21615e4b3a3527c21f7a0fd7f2da9e26794f5f51bc9404e369f.jpg", + "image_caption": [ + "Figure 2. Visualization of corrupted versions of datasets. From left to right, we display the original image, followed by the attention heatmaps obtained from Spike-driven Transformer, QK-Former, and SpiLiFormer (ours), respectively. The information below each image includes the true label, the predicted label for each model, and the confidence score for the predicted category. Red indicates incorrect classification, while green represents correct classification. For more visualizations, refer to Fig. 6 and Fig. 7 in the appendix." + ], + "image_footnote": [], + "bbox": [ + 120, + 89, + 457, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 632, + 243, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct an ablation study on CIFAR-10 (static) and CIFAR10-DVS (neuromorphic) to assess the impact of FF-LiDiff and FB-LiDiff attention on model performance and inference latency. All ablation experiments follow the training details in Sec. 4.2 and Sec. 4.3, unless otherwise specified.", + "bbox": [ + 89, + 655, + 482, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Intuitively, we set up two experimental scenarios: 1) For the case without FF-LiDiff attention (w/o FF-LiDiff attention), we use $Attn = \\mathcal{SN}(\\sum_{j}^{D}Q^{i,j})$ to replace the steps in Eq. (6) and Eq. (7); 2) For the case without FB-LiDiff attention (w/o FB-LiDiff attention), we perform a single forward pass with $Attn = Q\\odot (K^{\\top}\\odot V)$ to calculate the attention, replacing the steps outlined in Eq. (12) to Eq. (19).", + "bbox": [ + 89, + 747, + 482, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Tab. 3, we observe that the model exhibits varying degrees of performance degradation on both datasets under the two aforementioned scenarios. Specifi", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e6ef6abe0370a583876cef3e5dafedede638f8c058874fa4141507144e1a9dc7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 88, + 709, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/46961d053c52fe5fc8dc5698664a551ca706b938c8f8c2c9a6ec4c28a06b33dc.jpg", + "image_caption": [ + "Figure 3. Adversarial robustness comparison between QKFormer and our proposed model on CIFAR-100 (first row) and ImageNet-1K (second row). For PGD, the attack strength is set to 8/255, with a step size of 2/255 per iteration. The comparison results on other datasets are provided in Tab. 6 of within the appendix." + ], + "image_footnote": [], + "bbox": [ + 519, + 186, + 709, + 284 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/df3215a47b0bcca1cd92194ebc47ebec483550843c6ac3f1ad6679b9eddbf938.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 89, + 900, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7f732a35fdc6419c6228c65d0a6763161bb338b6fdedd6c9a15b6eb4c7703c18.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 186, + 900, + 284 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cally, performance exhibits a decrease of $0.31\\%$ and $0.62\\%$ on the CIFAR-10 dataset, and $1.6\\%$ and $3.0\\%$ on the CIFAR10-DVS dataset. These results highlight the effectiveness of FB-LiDiff in mitigating attention distraction and improving model performance. However, its additional forward pass introduces higher inference latency, particularly on static datasets, with a $21 - 23\\%$ increase as shown in Tab. 5 in the appendix.", + "bbox": [ + 511, + 393, + 906, + 513 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/51629e45970592104c0832ebac8b90cb05160b947f60ff971df483b9ef7e2c45.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetsMethodsTop-1 Acc(%)
CIFAR-10w/o FF-LiDiff Attention96.32
w/o FB-LiDiff Attention96.01
SpiLiFormer(ours)96.63
CIFAR10-DVSw/o FF-LiDiff Attention85.1
w/o FB-LiDiff Attention83.7
SpiLiFormer(ours)86.7
", + "bbox": [ + 524, + 525, + 890, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation study on two attention modules", + "bbox": [ + 560, + 650, + 856, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 693, + 633, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we identify attention distraction in mainstream Transformer-based SNNs, a critical issue that limits performance. Inspired by the brain's lateral inhibition mechanism, we propose SpiLiFormer, incorporating FF-LiDiff and FB-LiDiff attention to simulate the inhibitory-excitatory interaction process, guiding the model to focus on the object rather than the irrelevant background. Experimental results show that SpiLiFormer outperforms current Transformer-based SNNs, achieving SOTA performance on five image classification datasets. Additionally, adversarial testing and attention visualization analysis demonstrate its robustness and effectiveness in alleviating attention distraction.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24546", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research was supported in part by Zhejiang Provincial Natural Science Foundation of China under the Grant No. LD25F020006.", + "bbox": [ + 89, + 114, + 483, + 160 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 176, + 187, + 193 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hila Chefer, Idan Schwartz, and Lior Wolf. Optimizing relevance maps of vision transformers improves robustness. Advances in Neural Information Processing Systems, 35: 33618-33632, 2022. 1", + "[2] Xiang Cheng, Yunzhe Hao, Jiaming Xu, and Bo Xu. Lisnn: Improving spiking neural networks with lateral interactions for robust object recognition. In *IJCAI*, pages 1519–1525. Yokohama, 2020. 3", + "[3] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 6", + "[4] Joseph Del Rosario, Stefano Coletta, Soon Ho Kim, Zach Mobile, Kayla Peelman, Brice Williams, Alan J Otsuki, Alejandra Del Castillo Valerio, Kendell Worden, Lou T Blanpain, et al. Lateral inhibition in v1 controls neural and perceptual contrast sensitivity. Nature Neuroscience, pages 1-12, 2025. 2", + "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 3, 5", + "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, G Heigold, S Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 6", + "[7] Wei Fang, Zhaofei Yu, Yanqi Chen, Tiejun Huang, Timothée Masquelier, and Yonghong Tian. Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems, 34:21056-21069, 2021. 3, 6", + "[8] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2019. 3", + "[9] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 646-661. Springer, 2016. 6", + "[10] Seijoon Kim, Seongsik Park, Byunggook Na, and Sungroh Yoon. Spiking-yolo: spiking neural network for energy-efficient object detection. In Proceedings of the AAAI conference on artificial intelligence, pages 11270-11277, 2020. 1", + "[11] Alex Krizhevsky et al. Learning multiple layers of features from tiny images. 2009. 3, 5" + ], + "bbox": [ + 93, + 202, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Donghyun Lee, Yuhang Li, Youngeun Kim, Shiting Xiao, and Priyadarshini Panda. Spiking transformer with spatial-temporal attention. arXiv preprint arXiv:2409.19764, 2024. 1, 3", + "[13] Hongmin Li, Hanchao Liu, Xiangyang Ji, Guoqi Li, and Luping Shi. Cifar10-dvs: an event-stream dataset for object classification. Frontiers in neuroscience, 11:309, 2017. 3,5", + "[14] Qianhui Liu, Dong Xing, Huajin Tang, De Ma, and Gang Pan. Event-based action recognition using motion information and spiking neural networks. In *IJCAI*, pages 1743–1749, 2021. 1", + "[15] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 6", + "[16] Xianlei Long, Xiaxin Zhu, Fangming Guo, Chao Chen, Xiangwei Zhu, Fuqiang Gu, Songyu Yuan, and Chunlong Zhang. Spike-brgnet: Efficient and accurate event-based semantic segmentation with boundary region-guided spiking neural networks. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1", + "[17] Xinhao Luo, Man Yao, Yuhong Chou, Bo Xu, and Guoqi Li. Integer-valued training and spike-driven inference spiking neural network for high-performance and energy-efficient object detection. In European Conference on Computer Vision, pages 253-272. Springer, 2024. 1, 6", + "[18] Wolfgang Maass. Networks of spiking neurons: the third generation of neural network models. Neural networks, 10 (9):1659-1671, 1997. 1", + "[19] James S Magnuson, Anne Marie Crinnion, Sahil Luthra, Phoebe Gaston, and Samantha Grubb. Contra assertions, feedback improves word recognition: How feedback and lateral inhibition sharpen signals over noise. Cognition, 242: 105661, 2024. 2", + "[20] Garrick Orchard, Ajinkya Jayawant, Gregory K Cohen, and Nitish Thakor. Converting static image datasets to spiking neuromorphic datasets using saccades. Frontiers in neuroscience, 9:437, 2015. 3, 5", + "[21] Arda Sahiner, Tolga Ergen, Batu Ozturkler, John Pauly, Morteza Mardani, and Mert Pilanci. Unraveling attention via convex duality: Analysis and interpretations of vision transformers. In International Conference on Machine Learning, pages 19050-19088. PMLR, 2022. 1", + "[22] Stephen R Shaw. Retinal resistance barriers and electrical lateral inhibition. Nature, 255(5508):480-483, 1975. 2", + "[23] Zeyang Song, Jibin Wu, Malu Zhang, Mike Zheng Shou, and Haizhou Li. Spiking-leaf: A learnable auditory front-end for spiking neural networks. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 226-230. IEEE, 2024. 3", + "[24] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 6" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24547", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1", + "[26] Alex Vicente-Sola, Davide L Manna, Paul Kirkland, Gaetano Di Caterina, and Trevor J Bihl. Spiking neural networks for event-based action recognition: A new task to understand their advantage. Neurocomputing, 611:128657, 2025. 1", + "[27] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 568-578, 2021. 1", + "[28] Yuchen Wang, Kexin Shi, Chengzhuo Lu, Yuguo Liu, Malu Zhang, and Hong Qu. Spatial-temporal self-attention for asynchronous spiking neural networks. In IJCAI, pages 3085–3093, 2023. 1", + "[29] Man Yao, JiaKui Hu, Tianxiang Hu, Yifan Xu, Zhaokun Zhou, Yonghong Tian, XU Bo, and Guoqi Li. Spike-driven transformer v2: Meta spiking neural network architecture inspiring the design of next-generation neuromorphic chips. In The Twelfth International Conference on Learning Representations, 2024. 1, 4, 5", + "[30] Man Yao, Jiakui Hu, Zhaokun Zhou, Li Yuan, Yonghong Tian, Bo Xu, and Guoqi Li. Spike-driven transformer. Advances in neural information processing systems, 36, 2024. 1, 3, 5, 6, 7", + "[31] Man Yao, Xuerui Qiu, Tianxiang Hu, Jiakui Hu, Yuhong Chou, Keyu Tian, Jianxing Liao, Luziwei Leng, Bo Xu, and Guoqi Li. Scaling spike-driven transformer with efficient spike firing approximation training. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 3, 4, 6", + "[32] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch on imagenet. In Proceedings of the IEEE/CVF international conference on computer vision, pages 558-567, 2021. 1", + "[33] Geng Zhang, Shuangming Yang, Xuetao Zhang, and Badong Chen. Biologically plausible unsupervised learning for self-organizing spiking neural networks with dendritic computation. Neurocomputing, page 129707, 2025. 3", + "[34] Xuechen Zhang, Xiangyu Chang, Mingchen Li, Amit Roy-Chowdhury, Jiasi Chen, and Samet Oymak. Selective attention: Enhancing transformer through principled context control. Advances in Neural Information Processing Systems, 37:11061-11086, 2025. 1", + "[35] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In Proceedings of the AAAI conference on artificial intelligence, pages 13001-13008, 2020. 6", + "[36] Chenlin Zhou, Liutao Yu, Zhaokun Zhou, Zhengyu Ma, Han Zhang, Huihui Zhou, and Yonghong Tian. Spikingformer: Spike-driven residual learning for transformer-based spiking neural network. arXiv preprint arXiv:2304.11954, 2023. 1, 3, 5, 6, 7" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Chenlin Zhou, Han Zhang, Zhaokun Zhou, Liutao Yu, Liwei Huang, Xiaopeng Fan, Li Yuan, Zhengyu Ma, Huihui Zhou, and Yonghong Tian. Qkformer: Hierarchical spiking transformer using qk attention. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1, 3, 4, 6, 7", + "[38] Zhaokun Zhou, Yuesheng Zhu, Chao He, Yaowei Wang, YAN Shuicheng, Yonghong Tian, and Li Yuan. Spikformer: When spiking neural network meets transformer. In The Eleventh International Conference on Learning Representations, 2023. 1, 3, 5, 6, 7", + "[39] Zhaokun Zhou, Kaiwei Che, Wei Fang, Keyu Tian, Yuesheng Zhu, Shuicheng Yan, Yonghong Tian, and Li Yuan. Spikformer v2: Join the high accuracy club on imagenet with an snn ticket. arXiv preprint arXiv:2401.02020, 2024. 1", + "[40] Rui-Jie Zhu, Qihang Zhao, Guoqi Li, and Jason K Eshraghian. Spikegpt: Generative pre-trained language model with spiking neural networks. arXiv preprint arXiv:2302.13939, 2023. 1, 3, 5", + "[41] Shirui Zhu, Tao Xie, Ziyu Lv, Yan-Bing Leng, Yu-Qi Zhang, Runze Xu, Jingrun Qin, Ye Zhou, Vellaismy AL Roy, and Su-Ting Han. Hierarchies in visual pathway: functions and inspired artificial vision. Advanced Materials, 36(6): 2301986, 2024. 2", + "[42] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2021. 1", + "[43] Petr Znamenskiy, Mean-Hwan Kim, Dylan R Muir, M Florencia Iacaruso, Sonja B Hofer, and Thomas D Mrsic-Flogel. Functional specificity of recurrent inhibition in visual cortex. Neuron, 112(6):991-1000, 2024. 2" + ], + "bbox": [ + 516, + 90, + 903, + 542 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24548", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SpiLiFormer_ Enhancing Spiking Transformers with Lateral Inhibition/75735d8f-465e-40da-8f9a-a276e0092626_model.json b/2025/SpiLiFormer_ Enhancing Spiking Transformers with Lateral Inhibition/75735d8f-465e-40da-8f9a-a276e0092626_model.json new file mode 100644 index 0000000000000000000000000000000000000000..934c99b6a6b348d2247bf5e0cd29f2775b0a72c0 --- /dev/null +++ b/2025/SpiLiFormer_ Enhancing Spiking Transformers with Lateral Inhibition/75735d8f-465e-40da-8f9a-a276e0092626_model.json @@ -0,0 +1,2013 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.137, + 0.13, + 0.861, + 0.154 + ], + "angle": 0, + "content": "SpiLiFormer: Enhancing Spiking Transformers with Lateral Inhibition" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.179, + 0.768, + 0.217 + ], + "angle": 0, + "content": "Zeqi Zheng\\(^{1,2*}\\), Yanchen Huang\\(^{2,3*}\\), Yingchao Yu\\(^{4,2}\\), Zizheng Zhu\\(^{2,5}\\), Junfeng Tang\\(^{1,2}\\), Zhaofei Yu\\(^{6}\\), Yaochu Jin\\(^{2\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.217, + 0.862, + 0.253 + ], + "angle": 0, + "content": "1Zhejiang University 2Westlake University 3Nanjing University 4Donghua University 5University of Electronic Science and Technology of China 6Peking University" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.255, + 0.73, + 0.27 + ], + "angle": 0, + "content": "{zhengzeqi, huangyanchen, jinyaochu}@westlake.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.304, + 0.327, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.337, + 0.486, + 0.671 + ], + "angle": 0, + "content": "Spiking Neural Networks (SNNs) based on Transformers have garnered significant attention due to their superior performance and high energy efficiency. However, the spiking attention modules of most existing Transformer-based SNNs are adapted from those of analog Transformers, failing to fully address the issue of over-allocating attention to irrelevant contexts. To fix this fundamental yet overlooked issue, we propose a Lateral Inhibition-inspired Spiking Transformer (SpiLiFormer). It emulates the brain's lateral inhibition mechanism, guiding the model to enhance attention to relevant tokens while suppressing attention to irrelevant ones. Our model achieves state-of-the-art (SOTA) performance across multiple datasets, including CIFAR-10 (+0.45%), CIFAR-100 (+0.48%), CIFAR10-DVS (+2.70%), N-Caltech101 (+1.94%), and ImageNet-1K (+1.6%)1. Notably, on the ImageNet-1K dataset, SpiLiFormer (69.9M parameters, 4 time steps, 384 resolution) outperforms E-SpikeFormer (173.0M parameters, 8 time steps, 384 resolution), a SOTA spiking Transformer, by 0.46% using only 39% of the parameters and half the time steps. The code and model checkpoints are publicly available at https://github.com/KirinZheng/SpiLiFormer." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.702, + 0.222, + 0.717 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.728, + 0.485, + 0.851 + ], + "angle": 0, + "content": "Spiking Neural Networks (SNNs), regarded as the third generation of neural networks [18], are seen as a potential alternative to Artificial Neural Networks (ANNs) due to their biological interpretability and high energy efficiency, which stem from their event-driven properties. Transformer [25], originally designed for natural language processing tasks, has now become a dominant neural network architecture, demonstrating remarkable performance across" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.306, + 0.907, + 0.413 + ], + "angle": 0, + "content": "various visual tasks [6, 27, 32, 42]. The success of Transformers has driven the exploration of their integration with SNNs, with the objective of enhancing SNN models' performance across various tasks and progressively bridging the performance gap with ANNs, particularly in image classification [29, 30, 37, 38], object detection [10, 17], action recognition [14, 26], and semantic segmentation [16, 29]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.413, + 0.909, + 0.626 + ], + "angle": 0, + "content": "The attention module plays a crucial role in Transformer-based SNNs, significantly influencing model performance. Recent research advancements in spiking attention modules can be broadly classified into three categories: 1) Self-Spiking Attention (SSA) [36, 38, 39], which directly converts the Query \\((Q)\\), Key \\((K)\\), and Value \\((V)\\) into sparse spikes; 2) Spike-Driven Self-Attention (SDSA) [29, 30] and Spiking RWKV (S-RWKV) [40], which replace the dot product in SSA with the Hadamard product, achieving linear complexity while reducing both energy consumption and computational cost; 3) Spatial-Temporal Attention (STA) [12, 28], which refines traditional spatial attention by explicitly modeling the key temporal dependencies intrinsic to spike-based processing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.909, + 0.883 + ], + "angle": 0, + "content": "However, we observe that most Transformer-based SNNs exhibit a phenomenon referred to as attention distraction. As illustrated in Fig. 1(a), the model disproportionately assigns high attention weights to background information during decision-making, leading to the neglect of critical object-related features and ultimately impairing classification performance. This issue stems from the fact that spiking attention largely inherits the traditional attention forward propagation paradigm. Previous studies [1, 21, 34] suggest that traditional attention mechanisms process \\( Q \\), \\( K \\), and \\( V \\) uniformly in the mapping \\( softmax(QK^{\\top})V \\), thereby limiting their capacity to regulate contextual sparsity and relevance. Therefore, we posit that constructing additional pathways to process \\( Q \\), \\( K \\), and \\( V \\) separately is essential for mitigating attention distraction in Transformer-based SNNs and enhancing model performance." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "By contrast, the natural lateral inhibition mechanism in" + }, + { + "type": "page_footnote", + "bbox": [ + 0.113, + 0.863, + 0.363, + 0.876 + ], + "angle": 0, + "content": "* Equal contribution. † Corresponding author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.877, + 0.483, + 0.9 + ], + "angle": 0, + "content": "These results are obtained by comparing with state-of-the-art spiking neural network models that have a similar number of parameters." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.863, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24539" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.138, + 0.092, + 0.626, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.113, + 0.874, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.272, + 0.895, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.581, + 0.907, + 0.679 + ], + "angle": 0, + "content": "Figure 1. Illustration of the severe attention distraction phenomenon in Transformer-based SNNs, the architecture of the Lateral Inhibition-inspired Spiking Transformer (SpiLiFormer), and the detailed processing mechanism of Feedforward-pathway Lateral Differential Inhibition (FF-LiDiff) attention and Feedback-pathway Lateral Difference Inhibition (FB-LiDiff) attention. (a) refers to the attention distraction in mainstream Transformer-based SNNs, where models excessively focus on irrelevant background information, leading to misclassification. (b) denotes the retinal lateral inhibition mechanism, illustrating how horizontal cells regulate neural responses to enhance contrast and suppress noise. (c) presents the architecture of SpiLiFormer. (d) and (e) represent the information processing flows of FF-LiDiff and FB-LiDiff attention, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.486, + 0.887 + ], + "angle": 0, + "content": "the visual system enables the brain to focus on important areas, thereby improving perception and reducing visual overload [4, 19, 22]. Specifically, when neurons within a given area are strongly activated, adjacent neurons are inhibited. This principle is first observed in the retina, where horizontal cells receive input from photoreceptors and inhibit both the stimulated photoreceptors and their neighbors, forming a center-surround receptive field (RF), as shown in Fig. 1(b). In higher visual areas, such as the primary (V1) and secondary (V2) visual cortices, a similar mechanism operates through long-range connections, which contrasts with the short-range inhibition in the retina, helping to refine visual" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.737 + ], + "angle": 0, + "content": "processing by suppressing irrelevant stimuli and selectively enhancing salient features [41, 43]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Based on the short-range and long-range lateral inhibition mechanisms discussed above, we propose the Lateral Inhibition-inspired Spiking Transformer (SpiLiFormer) to address the attention distraction issue in Transformer-based SNNs, as shown in Fig. 1(c). This model consists of two types of stacked modules: 1) Feedforward-pathway Lateral Differential Inhibition (FF-LiDiff) blocks and 2) Feedback-pathway Lateral Differential Inhibition (FB-LiDiff) blocks. Moreover, it introduces the following innovations: 1) In the shallow blocks (i.e., Stage 1 and Stage 2), it explicitly in-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24540" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.348 + ], + "angle": 0, + "content": "corporates an inhibition differential attention mechanism by emulating the retinal short-range lateral inhibition process (see Fig. 1(d) for details); 2) In the deep blocks (i.e., Stage 3), it employs feedback-driven processing based on differential attention to simulate the long-range lateral inhibition mechanism observed in the cerebral cortex (see Fig. 1(e) for details). We evaluate our proposed SpiLiFormer on five image classification datasets, including CIFAR-10 [11], CIFAR-100 [11], CIFAR10-DVS [13], N-Caltech101 [20], and ImageNet-1K [5], demonstrating superior performance over current SOTA models. Additionally, adversarial testing and attention heatmap visualizations on the ImageNet-1K test set and noise-augmented datasets (i.e., CIFAR-10-C and ImageNet-1K-C [8]) show that SpiLiFormer effectively suppresses attention on irrelevant areas and background noise. The main contributions of this work are:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.353, + 0.481, + 0.456 + ], + "angle": 0, + "content": "- We propose SpiLiFormer, which incorporates a brain-inspired lateral inhibition mechanism and introduces two novel attention paradigms: FF-LiDiff attention and FB-LiDiff attention. This model effectively mitigates the attention distraction issue common in current Transformer-based SNNs, thereby enhancing image classification performance." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.458, + 0.48, + 0.579 + ], + "angle": 0, + "content": "We evaluate SpiLiFormer across multiple datasets, achieving new SOTA performance on CIFAR-10, CIFAR100, CIFAR10-DVS, N-Caltech101, and ImageNet-1K. Specifically, on ImageNet-1K, SpiLiFormer (69.9M parameters, 4 time steps) outperforms the optimal SOTA model, E-SpikeFormer (173.0M parameters, 8 time steps), by \\(0.46\\%\\) in top-1 accuracy, while using only 0.39 times the number of parameters and half the time steps." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.58, + 0.48, + 0.655 + ], + "angle": 0, + "content": "- We comprehensively evaluate the robustness of SpiLiFormer through attention heatmap visualizations and adversarial testing, and find that SpiLiFormer achieves more effective attention allocation and enhanced robustness compared to other baseline spiking models." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.353, + 0.481, + 0.655 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.671, + 0.232, + 0.687 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.697, + 0.321, + 0.712 + ], + "angle": 0, + "content": "2.1. Transformer-based SNNs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Although convolution-based SNNs exhibit high energy efficiency, they still demonstrate a significant performance gap compared to ANNs. To address this challenge, researchers have explored integrating Transformer architectures with SNNs. Spikformer [38] was the first to incorporate the Transformer attention into SNNs, cleverly replacing the Softmax operation by leveraging the binary characteristic of spike representations to encode Query, Key, and Value. Building upon this, Spikingformer [36] introduced a pre-activation shortcut to avoid floating-point multiplication while simultaneously reducing the spike firing rate. Inspired by SEW-ResNet [7], Spike-driven Trans" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.289 + ], + "angle": 0, + "content": "former [30] replaced the dot product with a Hadamard product and reshaped the spike residual connection based on the membrane potential to further minimize energy consumption and leverage the event-driven property. Additionally, Spiking RWKV [40] and QKFormer [37] introduced linear attention into SNNs to reduce computational complexity. Meanwhile, STA-Transformer [12] effectively integrated both spatial and temporal dynamics of spike trains to enhance model performance. However, most existing Transformer-based SNNs still adhere to the traditional attention propagation paradigm of ANNs, which results in the severe phenomenon of attention distraction, thereby limiting model performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.3, + 0.787, + 0.315 + ], + "angle": 0, + "content": "2.2. Models with Lateral Inhibition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.322, + 0.907, + 0.626 + ], + "angle": 0, + "content": "Lateral inhibition mechanisms have been widely adopted in various SNN applications. In speech recognition, Spiking-LEAF [23] emulated inner hair cell functionality and leveraged lateral inhibition feedback to enhance the neuron model, significantly improving the efficiency and noise robustness of spike encoding. In the object recognition, LISNN [2] modeled lateral inhibition by establishing spatial neighbor-order relationships between neurons, thereby enhancing the model's ability to focus on important features. In the image classification, Zhang et al. [33] proposed an adaptive self-organizing lateral inhibition strategy, where inhibition strength was adjusted according to the Euclidean distance between neurons. This approach addressed the issue of inefficient feature clustering and reduced redundancy in neuron activation. However, existing research on lateral inhibition has predominantly focused on small-scale datasets for validating model feasibility, with limited exploration of its applicability to larger and more complex datasets. Furthermore, its integration with Transformer-based SNNs remains largely unexplored." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.641, + 0.605, + 0.657 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.667, + 0.905, + 0.728 + ], + "angle": 0, + "content": "This section details the proposed SpiLiFormer model, covering the spiking neuron layer, overall architecture, and two lateral inhibition-inspired attention mechanisms. Additionally, we outline the training strategy and loss function." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.739, + 0.722, + 0.755 + ], + "angle": 0, + "content": "3.1. Spiking Neuron Layer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.762, + 0.905, + 0.868 + ], + "angle": 0, + "content": "The spiking neuron layer relies on cumulative firing to integrate spatio-temporal information into the membrane potential, which is then converted into binary spikes to drive computation in the next layer. We adopt the widely used Leaky Integrate-and-Fire (LIF) spiking neuron layer, as established in previous studies [30, 31, 37], with the following dynamics:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.885, + 0.905, + 0.903 + ], + "angle": 0, + "content": "\\[\nU [ t ] = H [ t ] (1 - S [ t ]) + U _ {\\text {r e s e t}} S [ t ], \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24541" + } + ], + [ + { + "type": "equation", + "bbox": [ + 0.107, + 0.115, + 0.484, + 0.143 + ], + "angle": 0, + "content": "\\[\nH [ t ] = U [ t - 1 ] + \\frac {1}{\\tau} (X [ t ] - (U [ t - 1 ] - U _ {\\text {r e s e t}})), \\quad (2)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.165, + 0.483, + 0.182 + ], + "angle": 0, + "content": "\\[\nS [ t ] = \\Theta (H [ t ] - U _ {t h}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.188, + 0.484, + 0.309 + ], + "angle": 0, + "content": "where \\( U_{\\mathrm{reset}} \\) represents the reset potential, and \\( X[t] \\) and \\( U[t] \\) denote the current input and membrane potential at timestep \\( t \\), respectively. \\( \\Theta(\\cdot) \\) is the Heaviside step function, which outputs 1 if \\( x \\geq 0 \\) and 0 otherwise. When \\( H[t] \\) exceeds the firing threshold \\( U_{th} \\), the spiking neuron fires a spike \\( S[t] \\), and \\( U[t] \\) is reset to \\( U_{\\mathrm{reset}} \\) at that moment. Additionally, \\( \\tau \\) represents the membrane time constant, which controls the rate of membrane potential leakage." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.319, + 0.288, + 0.333 + ], + "angle": 0, + "content": "3.2. Overall Architecture" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.34, + 0.484, + 0.702 + ], + "angle": 0, + "content": "The overall architecture of SpiLiFormer is illustrated in Fig. 1(c), which consists of three stages: Stage 1, Stage 2, and Stage 3, forming a three-level hierarchical structure. The input data is represented as \\( I \\in \\mathbb{R}^{T \\times N \\times H \\times W} \\), where \\( N = 3 \\) for static image data and \\( N = 2 \\) for neuromorphic data. In Stage 1, a \\( 4 \\times 4 \\) patch size is utilized, mapping each patch's input feature dimension \\( (4 \\times 4 \\times N) \\) to a spike-form representation of arbitrary dimension \\( (C) \\) via the Downsampling module. This process reduces the number of tokens to \\( \\frac{H}{4} \\times \\frac{W}{4} \\). The transformed features are then processed through FF-LiDiff attention and a Spiking MultiLayer Perceptron (SpMLP). FF-LiDiff attention introduces an additional forward pathway and employs a differential approach to mitigate attention distraction in the network's shallower layers (see Sec. 3.3 for details). In Stage 2, the processing mechanism is similar to Stage 1, but with a \\( 2 \\times 2 \\) patch size, an expanded channel dimension of \\( 2C \\), and the number of tokens reduced to \\( \\frac{H}{8} \\times \\frac{W}{8} \\). Finally, in Stage 3, we retain the \\( 2 \\times 2 \\) patch size, expand the channel dimension to \\( 4C \\), and reduce the number of tokens to \\( \\frac{H}{16} \\times \\frac{W}{16} \\). Additionally, a more complex FB-LiDiff attention module with feedback loops is introduced (see Sec. 3.4 for details) to replace the previous FF-LiDiff attention, further mitigating attention distraction in the model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.484, + 0.809 + ], + "angle": 0, + "content": "The number of spiking blocks (either FF-LiDiff block or FB-LiDiff block) in each stage is denoted as \\( N_{1} \\), \\( N_{2} \\), and \\( N_{3} \\), respectively. Since the Downsampling, SpMLP and classification head modules are based on prior research works [29, 31, 37], the following sections will focus on the detailed description of FF-LiDiff attention and FB-LiDiff attention." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.483, + 0.849 + ], + "angle": 0, + "content": "3.3. Feedforward-pathway Lateral Differential Inhibition Attention" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "An overview of FF-LiDiff attention is shown in Figure 1(d). For better mathematical description, in subsequent discussion, we assume \\( T = 1 \\) and only use single head attention." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.091, + 0.906, + 0.137 + ], + "angle": 0, + "content": "Given a spike input \\(X \\in \\mathbb{R}^{N \\times D}\\) (\\(N\\) is the token number, \\(D\\) is the channel number), Query \\((Q)\\) and Key \\((K)\\) can be computed through learnable matrices:" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.15, + 0.907, + 0.166 + ], + "angle": 0, + "content": "\\[\nQ = B N \\left(X W _ {Q}\\right), K = B N \\left(X W _ {K}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.177, + 0.824, + 0.194 + ], + "angle": 0, + "content": "where \\(W_{Q}, W_{K} \\in \\mathbb{R}^{D \\times D}\\) and \\(Q, K \\in \\mathbb{R}^{N \\times D}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.195, + 0.906, + 0.271 + ], + "angle": 0, + "content": "To implement the lateral inhibition mechanism, we first split \\(Q\\) along the channel dimension into two parts, obtaining \\(Q_{1}\\) and \\(Q_{2}\\), before computing the attention matrix. Then, all of \\(Q_{1}, Q_{2}\\), and \\(K\\) are transformed into spike trains via their respective spiking neuron layers:" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.283, + 0.907, + 0.3 + ], + "angle": 0, + "content": "\\[\nQ _ {s 1} = \\mathcal {S N} (Q _ {1}), Q _ {s 2} = \\mathcal {S N} (Q _ {2}), K _ {s} = \\mathcal {S N} (K), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.312, + 0.906, + 0.48 + ], + "angle": 0, + "content": "where \\( Q_{s1}, Q_{s2} \\in \\mathbb{R}^{N \\times \\frac{D}{2}} \\), \\( K_{s} \\in \\mathbb{R}^{N \\times D} \\). Next, we sum along the channel dimension to obtain the excitatory attention \\( A_{e} \\) and inhibitory attention \\( A_{i} \\). Specifically, \\( A_{e} \\) and \\( A_{i} \\) are \\( N \\times 1 \\) token attention vectors, representing the importance of different tokens from the perspectives of excitation and inhibition, respectively. Finally, we compute the element-wise difference between \\( A_{e} \\) and \\( A_{i} \\), and pass the aggregated attention through a spiking neural layer to obtain binary attention \\( A_{\\text{combined}} \\). This attention is then applied as a token-wise mask to \\( K_{s} \\) via the Hadamard product \\( \\otimes \\). The above process can be described as:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.492, + 0.906, + 0.54 + ], + "angle": 0, + "content": "\\[\nA _ {e} = \\sum_ {j} ^ {\\frac {D}{2}} Q _ {s 1} ^ {i, j}, A _ {i} = \\sum_ {j} ^ {\\frac {D}{2}} Q _ {s 2} ^ {i, j}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.559, + 0.905, + 0.575 + ], + "angle": 0, + "content": "\\[\nA _ {\\text {c o m b i n e d}} = \\mathcal {S N} \\left(A _ {e} - A _ {i}\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.6, + 0.905, + 0.616 + ], + "angle": 0, + "content": "\\[\nX ^ {\\prime} = A _ {\\text {c o m b i n e d}} \\otimes K _ {s}. \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.627, + 0.906, + 0.656 + ], + "angle": 0, + "content": "3.4. Feedback-pathway Lateral Differential Inhibition Attention" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.664, + 0.906, + 0.784 + ], + "angle": 0, + "content": "FB-LiDiff attention requires two forward propagations as shown in Figure 1(e). Specifically, during the first forward propagation, FB-LiDiff attention takes spike-form \\( S_{\\mathrm{stage3}}^{Ds} \\) from the Downsampling module in Stage 3 or takes spikeform \\( S_{\\mathrm{stage3}}^i (i \\in [1, N_3]) \\) from the \\( i \\)-th preceding FB-LiDiff block as input. Given an input \\( X_1^m \\) to the \\( m \\)-th FB-LiDiff attention, the first forward propagation can be described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.797, + 0.907, + 0.815 + ], + "angle": 0, + "content": "\\[\nY _ {s} ^ {1, m} = \\mathcal {S N} \\left(B N \\left(X _ {1} ^ {m} W _ {Y} ^ {m}\\right)\\right), Y \\in \\{Q, K, V \\}, \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.842, + 0.905, + 0.861 + ], + "angle": 0, + "content": "\\[\nA t t n ^ {1, m} = Q _ {s} ^ {1, m} \\odot \\left(\\left(K _ {s} ^ {1, m}\\right) ^ {\\top} \\odot V _ {s} ^ {1, m}\\right), \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.885, + 0.905, + 0.904 + ], + "angle": 0, + "content": "\\[\nO _ {\\text {s t a g e 3}} ^ {m} = \\mathcal {S N} \\left(B N \\left(A t t n ^ {1, m} W _ {a t t n} ^ {m}\\right)\\right), \\tag {11}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24542" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.09, + 0.482, + 0.182 + ], + "angle": 0, + "content": "where \\(X_{1}^{m}\\in \\mathbb{R}^{N\\times D}\\) (\\(X_{1}^{m}\\in \\{S_{\\mathrm{stage3}}^{Ds},S_{\\mathrm{stage3}}^{i}\\}\\)), and \\(Y_{s}^{1,m}\\) (\\(Y\\in \\{Q,K,V\\}\\)) represents the corresponding spiking representation of \\(Q\\), \\(K\\) and \\(V\\) in the \\(m\\)-th FB-LiDiff attention. \\(W_{z}^{m}\\) (\\(z\\in \\{Q,K,V,attn\\}\\)) and \\(O_{\\mathrm{stage3}}^{m}\\) denote the learnable weight matrix and the spike-form output of the \\(m\\)-th FB-LiDiff attention, respectively. \\(\\odot\\) refers to the dot product." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.184, + 0.484, + 0.444 + ], + "angle": 0, + "content": "Before the second forward propagation, we use the output \\(O_{\\mathrm{stage3}}^{N_3}\\) from the last block of Stage 3 as feedback information and return it to each FB-LiDiff attention. To filter the feedback information and further mitigate the attention distraction problem, we first introduce learnable prompts \\(W_{p1}\\) and \\(W_{p2}\\), where \\(W_{p1}, W_{p2} \\in \\mathbb{R}^{1 \\times D}\\). Next, we compute the dot product between \\(W_{p1}\\) and \\(O_{\\mathrm{stage3}}^{N_3}\\), and similarly between \\(W_{p2}\\) and \\(O_{\\mathrm{stage3}}^{N_3}\\). These dot products are then broadcast along the token dimension. The results are scaled to the range [0, 1] using the clamp(\\cdot) function, forming the excitatory feedback attention \\(A_e^{\\mathrm{FB}}\\) and inhibitory feedback attention \\(A_i^{\\mathrm{FB}}\\). Subsequently, we compute the element-wise difference between \\(A_e^{\\mathrm{FB}}\\) and \\(A_i^{\\mathrm{FB}}\\) to obtain the aggregated attention \\(A_{\\mathrm{combined}}^{\\mathrm{FB}}\\). Finally, we apply a linear transformation to \\(A_{\\mathrm{combined}}^{\\mathrm{FB}}\\) and pass it through the spiking neuron layer to align it with the feature space of different FB-LiDiff attentions. These processes are formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.456, + 0.483, + 0.477 + ], + "angle": 0, + "content": "\\[\nA _ {e} ^ {\\mathrm {F B}} = \\operatorname {c l a m p} \\left(O _ {\\text {s t a g e 3}} ^ {N _ {3}} \\odot W _ {p 1}, 0, 1\\right), \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.491, + 0.483, + 0.513 + ], + "angle": 0, + "content": "\\[\nA _ {i} ^ {\\mathrm {F B}} = \\operatorname {c l a m p} \\left(O _ {\\text {s t a g e 3}} ^ {N _ {3}} \\odot W _ {p 2}, 0, 1\\right), \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.206, + 0.527, + 0.482, + 0.547 + ], + "angle": 0, + "content": "\\[\nA _ {\\text {c o m b i n e d}} ^ {\\mathrm {F B}} = A _ {e} ^ {\\mathrm {F B}} - A _ {i} ^ {\\mathrm {F B}}, \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.563, + 0.482, + 0.582 + ], + "angle": 0, + "content": "\\[\nX _ {\\mathrm {F B}} ^ {m} = \\mathcal {S N} \\left(B N \\left(A _ {\\text {c o m b i n e d}} ^ {\\mathrm {F B}} W _ {\\mathrm {F B}} ^ {m}\\right)\\right). \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.483, + 0.753 + ], + "angle": 0, + "content": "The second forward propagation begins by taking the processed forward representation \\( S_{\\mathrm{stage3}}^{Ds} \\) and the corresponding feedback information \\( X_{\\mathrm{FB}}^{1} \\) as input to the first FB-LiDiff attention. The objective is to reduce model energy consumption by eliminating redundant computations in the non-feedback loop (i.e., Stage 1 and Stage 2). A detailed energy consumption analysis can be found in Appendix A.3. Additionally, during the second forward propagation, we share all parameter weights across the FB-LiDiff blocks to minimize the number of model parameters. Given the input \\( X_{2}^{m} \\), the process described above can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.775, + 0.483, + 0.793 + ], + "angle": 0, + "content": "\\[\nY _ {s} ^ {2, m} = \\mathcal {S N} \\left(B N \\left(X _ {2} ^ {m} W _ {Y} ^ {m}\\right)\\right), Y \\in \\{Q, K \\}, \\tag {16}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.147, + 0.813, + 0.483, + 0.832 + ], + "angle": 0, + "content": "\\[\nV _ {s} ^ {2, m} = \\mathcal {S N} \\left(B N \\left(X _ {2} ^ {m} W _ {V} ^ {m} + X _ {\\mathrm {F B}} ^ {m} W _ {V} ^ {m}\\right)\\right), \\tag {17}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.848, + 0.483, + 0.868 + ], + "angle": 0, + "content": "\\[\nA t t n ^ {2, m} = Q _ {s} ^ {2, m} \\odot \\left(\\left(K _ {s} ^ {2, m}\\right) ^ {\\top} \\odot V _ {s} ^ {2, m}\\right), \\tag {18}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.884, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\left(O _ {\\text {s t a g e} 3} ^ {m}\\right) ^ {\\prime} = \\mathcal {S N} \\left(B N \\left(A t t n ^ {2, m} W _ {a t t n} ^ {m}\\right)\\right). \\tag {19}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.833, + 0.107 + ], + "angle": 0, + "content": "3.5. Training Strategy and Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.905, + 0.218 + ], + "angle": 0, + "content": "We employ a surrogate gradient-based method to overcome the non-differentiability of the spiking neuron activation in Eq. (3), enabling direct training of the model from scratch, as demonstrate in previous works [29, 36-38, 40]. In order to make better use of the spiking representations obtained from two forward propagations, we adjust the model's loss function as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.232, + 0.905, + 0.252 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {1} = \\mathcal {L} _ {C E} \\left(C H \\left(O _ {\\text {s t a g e 3}} ^ {N _ {3}}\\right), y\\right), \\tag {20}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.269, + 0.905, + 0.29 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {2} = \\mathcal {L} _ {C E} \\left(C H \\left(\\left(O _ {\\text {s t a g e 3}} ^ {N _ {3}}\\right) ^ {\\prime}\\right), y\\right), \\tag {21}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.31, + 0.905, + 0.326 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {S p i L i F o r m e r}} = \\alpha \\mathcal {L} _ {1} + (1 - \\alpha) \\mathcal {L} _ {2}, \\tag {22}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.333, + 0.907, + 0.497 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a hyperparameter that dynamically balances the losses from two forward propagations. Unless explicitly specified otherwise, \\(\\alpha\\) is set to 0.5 in the subsequent experiments. The choice of this value is validated through an ablation study, as detailed in Appendix A.4. Additionally, \\(y\\) represents the true label; \\(O_{\\mathrm{stage3}}^{N_3}\\) and \\((O_{\\mathrm{stage3}}^{N_3})'\\) denote the outputs from the first and second forward propagations of the last FB-LiDiff block, respectively; \\(CH(\\cdot)\\) represents the classification head module, as described in previous studies [29, 30]; and \\(\\mathcal{L}_{CE}\\) denotes the cross-entropy loss function." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.512, + 0.646, + 0.529 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.907, + 0.689 + ], + "angle": 0, + "content": "We first evaluate the classification performance of SpiLiFormer model on the large-scale ImageNet-1K dataset [5]. Next, we assess the performance of SpiLiFormer on smaller-scale static datasets, including CIFAR-10 [11] and CIFAR-100 [11]. Additionally, we evaluate SpiLiFormer on two popular neuromorphic datasets: CIFAR10-DVS [13] and N-Caltech101 [20]. Finally, we visualize attention heatmaps and conduct adversarial tests to demonstrate SpiLiFormer's effectiveness in mitigating attention distraction and enhancing robustness." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.698, + 0.738, + 0.714 + ], + "angle": 0, + "content": "4.1. Results on ImageNet-1K" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this experiment, model training is divided into two phases. In the first phase, we use the AdamW optimizer with a base learning rate of \\(6 \\times 10^{-4}\\). The actual learning rate is calculated by multiplying the base rate by the batch size divided by 256. Additionally, we apply a layer-wise learning rate decay strategy with a decay factor of 1.0 and weight decay of 0.05 to train SpiLiFormer for 200 epochs. In the second phase, we decrease the base learning rate to \\(2 \\times 10^{-6}\\) and continue training for an additional 20 epochs using an epoch-wise learning rate decay strategy. Throughout this entire experiment, we incorporate several data augmentation techniques in alignment with the DeiT approach" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24543" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.895, + 0.471 + ], + "angle": 0, + "content": "
MethodsTypeArchitectureInput SizeParam(M)Power(mJ)Time StepTop-1 Acc(%)
ViT [6]ANNViT-B/1638486.59254.84177.90
DeiT [24]ANNDeiT-B22486.5980.50181.80
ANNDeiT-B38486.59254.84183.10
Swin [15]ANNSwin Transformer-B22487.7770.84183.50
ANNSwin Transformer-B38487.77216.20184.50
SEW ResNet [7]SNNSEW-ResNet-15222460.1912.89469.26
Spikformer [38]SNNSpikformer-8-76822466.3421.48474.81
Spikingformer [36]SNNSpikingformer-8-76822466.3413.68475.85
S-Transformer [30]SNNS-Transformer-8-51222429.681.13171.68
SNNS-Transformer-8-51222429.684.50474.57
SNNS-Transformer-8-76828866.346.09477.07
QKFormer [37]SNNHST-10-76822464.9638.91484.22
SNNHST-10-768*28864.9664.27485.25
SNNHST-10-76838464.96113.64485.65
E-SpikeFormer [31]SNNE-SpikeFormer224173.035.64*84.7
SNNE-SpikeFormer224173.054.78*85.1
SNNE-SpikeFormer384173.0-8*86.2
SpiLiFormer (Ours)SNNSpiLiFormer-10-76822469.1011.77181.54
SNNSpiLiFormer-10-76822469.1044.17485.82
SNNSpiLiFormer-10-768*28869.1073.52486.62
SNNSpiLiFormer-10-768**38469.10129.45486.66
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.482, + 0.908, + 0.54 + ], + "angle": 0, + "content": "Table 1. Performance comparison on ImageNet-1K. \"SpiLiFormer-L-D\" represents Lateral Inhibition-inspired Transformer with \\(L\\) blocks and a \\(D\\)-dimensional channel. \\* and \\*\\* denote input resolutions of \\(288^{2}\\) and \\(384^{2}\\) during inference, respectively. \\* denotes an integer-valued quantized training approach [17], in which activation values are restricted to integers during training and later expanded into spike trains during inference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.483, + 0.626 + ], + "angle": 0, + "content": "[24]. These techniques include RandAugment [3], random erasing [35], and stochastic depth [9]. The architecture of SpiLiFormer is composed of 1, 2, and 7 blocks across its three stages, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Tab. 1 demonstrates the superior performance of our proposed SpiLiFormer, significantly outperforming current SNNs models. Specifically, SpiLiFormer (69.10M) achieves \\(86.66\\%\\) top-1 accuracy and \\(98.116\\%\\) top-5 accuracy on ImageNet-1K. To our best knowledge, SpiLiFormer is currently the SOTA model for Transformer-based SNNs with direct training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We compare SpiLiFormer against several baseline spiking models under the same image resolution of 224, a time step of 4, and a comparable number of model parameters. SpiLiFormer outperforms Spikformer (66.34M), Spikingformer (66.34M), and QKFormer (64.96M) by \\(11.01\\%\\), \\(9.97\\%\\), and \\(1.6\\%\\), respectively. Moreover, at a higher input resolution of 288, SpiLiFormer outperforms S-Transformer and QKFormer by \\(9.55\\%\\) and \\(1.37\\%\\), respectively. Similarly, at a resolution of 384, it surpasses QKFormer by \\(1.01\\%\\). Additionally, we compare SpiLiFormer (69.10M, 4 time steps, 384 resolution) with E-SpikeFormer (173.0M, 8" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.907, + 0.611 + ], + "angle": 0, + "content": "time steps, 384 resolution), the current SOTA Transformer-based SNNs model. It achieves a \\(0.46\\%\\) improvement with only \\(39\\%\\) of parameters and half the time steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.907, + 0.687 + ], + "angle": 0, + "content": "Finally, compared to the mainstream ANN-based Swin Transformer-B (384 resolution), SpiLiFormer achieves a \\(2.16\\%\\) performance improvement while reducing energy consumption by \\(40\\%\\) and decreasing the number of parameters by 18.67M." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.697, + 0.761, + 0.713 + ], + "angle": 0, + "content": "4.2. Results on CIFAR Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.81 + ], + "angle": 0, + "content": "In this experiment, we use the AdamW optimizer with a learning rate of \\(1 \\times 10^{-3}\\) and apply a cosine learning rate decay strategy. The weight decay is \\(6 \\times 10^{-2}\\), the number of time steps is 4, and SpiLiFormer is trained for 400 epochs with a batch size of 64. The network architecture consists of 1, 1, and 2 blocks across its three stages, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The performance of SpiLiFormer on CIFAR datasets is shown in the upper part of Tab. 2. For CIFAR-10, our proposed model achieves \\(96.63\\%\\) accuracy with 7.04M parameters. Compared to QKFormer, the suboptimal SOTA model, SpiLiFormer improves performance by \\(0.45\\%\\) while increasing the parameters by only \\(0.04\\%\\). Similarly, for" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24544" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.116, + 0.089, + 0.885, + 0.428 + ], + "angle": 0, + "content": "
DatasetsMethodsArchitectureParam(M)Time StepTop-1 Acc(%)
CIFAR-10Spikformer [38]Spikformer-4-3849.32495.51
Spikingformer [36]Spikingformer-4-3849.32495.81
S-Transformer [30]S-Transformer-2-51210.28495.60
QKFormer [37]HST-4-3846.74496.18
SpiLiFormer (Ours)SpiLiFormer-4-3847.04496.63
CIFAR-100Spikformer [38]Spikformer-4-3849.32478.21
Spikingformer [36]Spikingformer-4-3849.32479.21
S-Transformer [30]S-Transformer-2-51210.28478.4
QKFormer [37]HST-4-3846.74481.15
SpiLiFormer (Ours)SpiLiFormer-4-3847.04481.63
CIFAR10-DVSSpikformer [38]Spikformer-2-2562.571680.9
Spikingformer [36]Spikingformer-2-2562.571681.3
S-Transformer [30]S-Transformer-2-2562.571680.0
QKFormer [37]HST-2-2561.501684.0
SpiLiFormer (Ours)SpiLiFormer-2-2561.571686.7
N-Caltech101Spikformer [38]Spikformer-2-2562.571683.6
Spikingformer [36]Spikingformer-2-2562.571685.91
S-Transformer [30]S-Transformer-2-2562.571686.3
QKFormer [37]HST-2-2561.501687.24
SpiLiFormer (Ours)SpiLiFormer-2-2561.571689.18
" + }, + { + "type": "table_caption", + "bbox": [ + 0.183, + 0.436, + 0.816, + 0.451 + ], + "angle": 0, + "content": "Table 2. Performance comparison on CIFAR-10, CIFAR-100, CIFAR10-DVS, and N-Caltech101 datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.477, + 0.483, + 0.539 + ], + "angle": 0, + "content": "CIFAR-100, our proposed model achieves \\(81.63\\%\\) accuracy with 7.04M parameters. Specifically, it outperforms QKFormer by \\(0.48\\%\\) while requiring just \\(0.04\\%\\) more parameters." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.547, + 0.395, + 0.563 + ], + "angle": 0, + "content": "4.3. Results on Neuromorphic Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.674 + ], + "angle": 0, + "content": "In this experiment, we use a smaller SpiLiFormer architecture, structured with 0, 1, and 1 blocks across its three stages, respectively. Similarly, we adopt the AdamW optimizer with an initial learning rate of \\(5 \\times 10^{-3}\\), apply a learning rate decay strategy, and set the weight decay to 0.06. The number of time steps is set to 16, and SpiLiFormer is trained for 120 epochs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.841 + ], + "angle": 0, + "content": "The experimental results of the model on the neuromorphic datasets are presented in the lower half of Tab. 2. Compared to the suboptimal model, QKFormer, SpiLi-Former significantly improves performance by \\(2.7\\%\\) on the CIFAR10-DVS dataset and \\(1.94\\%\\) on the N-Caltech101 dataset, with only a \\(0.13\\%\\) increase in parameters. Additionally, we compare the proposed model with other baseline models (i.e., Spikformer, Spikingformer, and S-Transformer). It achieves performance gains of up to \\(5.8\\%\\) on CIFAR10-DVS and \\(5.55\\%\\) on N-Caltech101 while reducing the parameter count by 0.87M." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.849, + 0.476, + 0.866 + ], + "angle": 0, + "content": "4.4. Visualization and Model Robustness Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "We visualize the attention heatmap of SpiLiFormer to assess whether it effectively alleviates the attention distraction is" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.907, + 0.539 + ], + "angle": 0, + "content": "sue. Specifically, we examine the attention heatmap of the last FB-LiDiff block in SpiLiFormer Stage 3 under two conditions: standard dataset evaluation and noisy dataset testing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.675 + ], + "angle": 0, + "content": "In the standard dataset evaluation, the two sets of images in Fig. 1(a) show examples that are misclassified by the baseline model but correctly classified by SpiLiFormer on the ImageNet-1K test set. We observe that our proposed model effectively eliminates the attention distraction phenomenon, thereby improving classification accuracy. More comparison images are provided in Fig. 4 within the appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.797 + ], + "angle": 0, + "content": "In noisy dataset testing, we evaluate model attention heatmaps using two noisy datasets: CIFAR10-C and ImageNet-1K-C, as shown in Fig. 2. Compared to other baseline models, SpiLiFormer correctly classifies objects or significantly increases confidence in the correct category by focusing on object-relevant information while remaining unaffected by background noise." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In addition, we employ white-box adversarial attacks, including FGSM and PGD, to assess the model's robustness across the four datasets mentioned above. As shown in Fig. 3, we find that SpiLiFormer not only enhances model performance but also improves robustness against adversarial and common noise attacks, outperforming QKFormer." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24545" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.122, + 0.09, + 0.458, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.478, + 0.484, + 0.604 + ], + "angle": 0, + "content": "Figure 2. Visualization of corrupted versions of datasets. From left to right, we display the original image, followed by the attention heatmaps obtained from Spike-driven Transformer, QK-Former, and SpiLiFormer (ours), respectively. The information below each image includes the true label, the predicted label for each model, and the confidence score for the predicted category. Red indicates incorrect classification, while green represents correct classification. For more visualizations, refer to Fig. 6 and Fig. 7 in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.633, + 0.245, + 0.649 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.656, + 0.483, + 0.746 + ], + "angle": 0, + "content": "We conduct an ablation study on CIFAR-10 (static) and CIFAR10-DVS (neuromorphic) to assess the impact of FF-LiDiff and FB-LiDiff attention on model performance and inference latency. All ablation experiments follow the training details in Sec. 4.2 and Sec. 4.3, unless otherwise specified." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.748, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Intuitively, we set up two experimental scenarios: 1) For the case without FF-LiDiff attention (w/o FF-LiDiff attention), we use \\( Attn = \\mathcal{SN}(\\sum_{j}^{D}Q^{i,j}) \\) to replace the steps in Eq. (6) and Eq. (7); 2) For the case without FB-LiDiff attention (w/o FB-LiDiff attention), we perform a single forward pass with \\( Attn = Q\\odot (K^{\\top}\\odot V) \\) to calculate the attention, replacing the steps outlined in Eq. (12) to Eq. (19)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "As shown in Tab. 3, we observe that the model exhibits varying degrees of performance degradation on both datasets under the two aforementioned scenarios. Specifi" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.089, + 0.71, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.188, + 0.71, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.09, + 0.901, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.188, + 0.901, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.299, + 0.907, + 0.369 + ], + "angle": 0, + "content": "Figure 3. Adversarial robustness comparison between QKFormer and our proposed model on CIFAR-100 (first row) and ImageNet-1K (second row). For PGD, the attack strength is set to 8/255, with a step size of 2/255 per iteration. The comparison results on other datasets are provided in Tab. 6 of within the appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.907, + 0.515 + ], + "angle": 0, + "content": "cally, performance exhibits a decrease of \\(0.31\\%\\) and \\(0.62\\%\\) on the CIFAR-10 dataset, and \\(1.6\\%\\) and \\(3.0\\%\\) on the CIFAR10-DVS dataset. These results highlight the effectiveness of FB-LiDiff in mitigating attention distraction and improving model performance. However, its additional forward pass introduces higher inference latency, particularly on static datasets, with a \\(21 - 23\\%\\) increase as shown in Tab. 5 in the appendix." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.526, + 0.892, + 0.641 + ], + "angle": 0, + "content": "
DatasetsMethodsTop-1 Acc(%)
CIFAR-10w/o FF-LiDiff Attention96.32
w/o FB-LiDiff Attention96.01
SpiLiFormer(ours)96.63
CIFAR10-DVSw/o FF-LiDiff Attention85.1
w/o FB-LiDiff Attention83.7
SpiLiFormer(ours)86.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.561, + 0.651, + 0.857, + 0.665 + ], + "angle": 0, + "content": "Table 3. Ablation study on two attention modules" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.694, + 0.634, + 0.71 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we identify attention distraction in mainstream Transformer-based SNNs, a critical issue that limits performance. Inspired by the brain's lateral inhibition mechanism, we propose SpiLiFormer, incorporating FF-LiDiff and FB-LiDiff attention to simulate the inhibitory-excitatory interaction process, guiding the model to focus on the object rather than the irrelevant background. Experimental results show that SpiLiFormer outperforms current Transformer-based SNNs, achieving SOTA performance on five image classification datasets. Additionally, adversarial testing and attention visualization analysis demonstrate its robustness and effectiveness in alleviating attention distraction." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24546" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.484, + 0.161 + ], + "angle": 0, + "content": "This research was supported in part by Zhejiang Provincial Natural Science Foundation of China under the Grant No. LD25F020006." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.178, + 0.188, + 0.194 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.203, + 0.484, + 0.258 + ], + "angle": 0, + "content": "[1] Hila Chefer, Idan Schwartz, and Lior Wolf. Optimizing relevance maps of vision transformers improves robustness. Advances in Neural Information Processing Systems, 35: 33618-33632, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.261, + 0.484, + 0.315 + ], + "angle": 0, + "content": "[2] Xiang Cheng, Yunzhe Hao, Jiaming Xu, and Bo Xu. Lisnn: Improving spiking neural networks with lateral interactions for robust object recognition. In *IJCAI*, pages 1519–1525. Yokohama, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.318, + 0.484, + 0.387 + ], + "angle": 0, + "content": "[3] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.39, + 0.484, + 0.471 + ], + "angle": 0, + "content": "[4] Joseph Del Rosario, Stefano Coletta, Soon Ho Kim, Zach Mobile, Kayla Peelman, Brice Williams, Alan J Otsuki, Alejandra Del Castillo Valerio, Kendell Worden, Lou T Blanpain, et al. Lateral inhibition in v1 controls neural and perceptual contrast sensitivity. Nature Neuroscience, pages 1-12, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.475, + 0.484, + 0.53 + ], + "angle": 0, + "content": "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.532, + 0.484, + 0.613 + ], + "angle": 0, + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, G Heigold, S Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.617, + 0.484, + 0.671 + ], + "angle": 0, + "content": "[7] Wei Fang, Zhaofei Yu, Yanqi Chen, Tiejun Huang, Timothée Masquelier, and Yonghong Tian. Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems, 34:21056-21069, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.674, + 0.484, + 0.728 + ], + "angle": 0, + "content": "[8] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.732, + 0.484, + 0.801 + ], + "angle": 0, + "content": "[9] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 646-661. Springer, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.484, + 0.87 + ], + "angle": 0, + "content": "[10] Seijoon Kim, Seongsik Park, Byunggook Na, and Sungroh Yoon. Spiking-yolo: spiking neural network for energy-efficient object detection. In Proceedings of the AAAI conference on artificial intelligence, pages 11270-11277, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[11] Alex Krizhevsky et al. Learning multiple layers of features from tiny images. 2009. 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.203, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[12] Donghyun Lee, Yuhang Li, Youngeun Kim, Shiting Xiao, and Priyadarshini Panda. Spiking transformer with spatial-temporal attention. arXiv preprint arXiv:2409.19764, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[13] Hongmin Li, Hanchao Liu, Xiangyang Ji, Guoqi Li, and Luping Shi. Cifar10-dvs: an event-stream dataset for object classification. Frontiers in neuroscience, 11:309, 2017. 3,5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[14] Qianhui Liu, Dong Xing, Huajin Tang, De Ma, and Gang Pan. Event-based action recognition using motion information and spiking neural networks. In *IJCAI*, pages 1743–1749, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.265, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[15] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.336, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[16] Xianlei Long, Xiaxin Zhu, Fangming Guo, Chao Chen, Xiangwei Zhu, Fuqiang Gu, Songyu Yuan, and Chunlong Zhang. Spike-brgnet: Efficient and accurate event-based semantic segmentation with boundary region-guided spiking neural networks. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.421, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[17] Xinhao Luo, Man Yao, Yuhong Chou, Bo Xu, and Guoqi Li. Integer-valued training and spike-driven inference spiking neural network for high-performance and energy-efficient object detection. In European Conference on Computer Vision, pages 253-272. Springer, 2024. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.491, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[18] Wolfgang Maass. Networks of spiking neurons: the third generation of neural network models. Neural networks, 10 (9):1659-1671, 1997. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.534, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[19] James S Magnuson, Anne Marie Crinnion, Sahil Luthra, Phoebe Gaston, and Samantha Grubb. Contra assertions, feedback improves word recognition: How feedback and lateral inhibition sharpen signals over noise. Cognition, 242: 105661, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.605, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[20] Garrick Orchard, Ajinkya Jayawant, Gregory K Cohen, and Nitish Thakor. Converting static image datasets to spiking neuromorphic datasets using saccades. Frontiers in neuroscience, 9:437, 2015. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.662, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[21] Arda Sahiner, Tolga Ergen, Batu Ozturkler, John Pauly, Morteza Mardani, and Mert Pilanci. Unraveling attention via convex duality: Analysis and interpretations of vision transformers. In International Conference on Machine Learning, pages 19050-19088. PMLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[22] Stephen R Shaw. Retinal resistance barriers and electrical lateral inhibition. Nature, 255(5508):480-483, 1975. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.762, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[23] Zeyang Song, Jibin Wu, Malu Zhang, Mike Zheng Shou, and Haizhou Li. Spiking-leaf: A learnable auditory front-end for spiking neural networks. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 226-230. IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.833, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24547" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[25] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.207 + ], + "angle": 0, + "content": "[26] Alex Vicente-Sola, Davide L Manna, Paul Kirkland, Gaetano Di Caterina, and Trevor J Bihl. Spiking neural networks for event-based action recognition: A new task to understand their advantage. Neurocomputing, 611:128657, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[27] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 568-578, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[28] Yuchen Wang, Kexin Shi, Chengzhuo Lu, Yuguo Liu, Malu Zhang, and Hong Qu. Spatial-temporal self-attention for asynchronous spiking neural networks. In IJCAI, pages 3085–3093, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[29] Man Yao, JiaKui Hu, Tianxiang Hu, Yifan Xu, Zhaokun Zhou, Yonghong Tian, XU Bo, and Guoqi Li. Spike-driven transformer v2: Meta spiking neural network architecture inspiring the design of next-generation neuromorphic chips. In The Twelfth International Conference on Learning Representations, 2024. 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[30] Man Yao, Jiakui Hu, Zhaokun Zhou, Li Yuan, Yonghong Tian, Bo Xu, and Guoqi Li. Spike-driven transformer. Advances in neural information processing systems, 36, 2024. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[31] Man Yao, Xuerui Qiu, Tianxiang Hu, Jiakui Hu, Yuhong Chou, Keyu Tian, Jianxing Liao, Luziwei Leng, Bo Xu, and Guoqi Li. Scaling spike-driven transformer with efficient spike firing approximation training. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[32] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch on imagenet. In Proceedings of the IEEE/CVF international conference on computer vision, pages 558-567, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[33] Geng Zhang, Shuangming Yang, Xuetao Zhang, and Badong Chen. Biologically plausible unsupervised learning for self-organizing spiking neural networks with dendritic computation. Neurocomputing, page 129707, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[34] Xuechen Zhang, Xiangyu Chang, Mingchen Li, Amit Roy-Chowdhury, Jiasi Chen, and Samet Oymak. Selective attention: Enhancing transformer through principled context control. Advances in Neural Information Processing Systems, 37:11061-11086, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[35] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In Proceedings of the AAAI conference on artificial intelligence, pages 13001-13008, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[36] Chenlin Zhou, Liutao Yu, Zhaokun Zhou, Zhengyu Ma, Han Zhang, Huihui Zhou, and Yonghong Tian. Spikingformer: Spike-driven residual learning for transformer-based spiking neural network. arXiv preprint arXiv:2304.11954, 2023. 1, 3, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[37] Chenlin Zhou, Han Zhang, Zhaokun Zhou, Liutao Yu, Liwei Huang, Xiaopeng Fan, Li Yuan, Zhengyu Ma, Huihui Zhou, and Yonghong Tian. Qkformer: Hierarchical spiking transformer using qk attention. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[38] Zhaokun Zhou, Yuesheng Zhu, Chao He, Yaowei Wang, YAN Shuicheng, Yonghong Tian, and Li Yuan. Spikformer: When spiking neural network meets transformer. In The Eleventh International Conference on Learning Representations, 2023. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.248, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[39] Zhaokun Zhou, Kaiwei Che, Wei Fang, Keyu Tian, Yuesheng Zhu, Shuicheng Yan, Yonghong Tian, and Li Yuan. Spikformer v2: Join the high accuracy club on imagenet with an snn ticket. arXiv preprint arXiv:2401.02020, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.358 + ], + "angle": 0, + "content": "[40] Rui-Jie Zhu, Qihang Zhao, Guoqi Li, and Jason K Eshraghian. Spikegpt: Generative pre-trained language model with spiking neural networks. arXiv preprint arXiv:2302.13939, 2023. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.905, + 0.429 + ], + "angle": 0, + "content": "[41] Shirui Zhu, Tao Xie, Ziyu Lv, Yan-Bing Leng, Yu-Qi Zhang, Runze Xu, Jingrun Qin, Ye Zhou, Vellaismy AL Roy, and Su-Ting Han. Hierarchies in visual pathway: functions and inspired artificial vision. Advanced Materials, 36(6): 2301986, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.432, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[42] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.488, + 0.905, + 0.543 + ], + "angle": 0, + "content": "[43] Petr Znamenskiy, Mean-Hwan Kim, Dylan R Muir, M Florencia Iacaruso, Sonja B Hofer, and Thomas D Mrsic-Flogel. Functional specificity of recurrent inhibition in visual cortex. Neuron, 112(6):991-1000, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24548" + } + ] +] \ No newline at end of file diff --git a/2025/SpiLiFormer_ Enhancing Spiking Transformers with Lateral Inhibition/75735d8f-465e-40da-8f9a-a276e0092626_origin.pdf b/2025/SpiLiFormer_ Enhancing Spiking Transformers with Lateral Inhibition/75735d8f-465e-40da-8f9a-a276e0092626_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c2105c6f5fe2ddb841cd83252cdd989c37ea1803 --- /dev/null +++ b/2025/SpiLiFormer_ Enhancing Spiking Transformers with Lateral Inhibition/75735d8f-465e-40da-8f9a-a276e0092626_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f544a22faed6f228d89fa38306e32fcf1d8d096d0db4fbf2e0b06ac275c756d6 +size 1201950 diff --git a/2025/SpiLiFormer_ Enhancing Spiking Transformers with Lateral Inhibition/full.md b/2025/SpiLiFormer_ Enhancing Spiking Transformers with Lateral Inhibition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..39a52d66f1a71fd3225dd79fc0f63b53bea64721 --- /dev/null +++ b/2025/SpiLiFormer_ Enhancing Spiking Transformers with Lateral Inhibition/full.md @@ -0,0 +1,321 @@ +# SpiLiFormer: Enhancing Spiking Transformers with Lateral Inhibition + +Zeqi Zheng $^{1,2*}$ , Yanchen Huang $^{2,3*}$ , Yingchao Yu $^{4,2}$ , Zizheng Zhu $^{2,5}$ , Junfeng Tang $^{1,2}$ , Zhaofei Yu $^{6}$ , Yaochu Jin $^{2\dagger}$ + +1Zhejiang University 2Westlake University 3Nanjing University 4Donghua University 5University of Electronic Science and Technology of China 6Peking University + +{zhengzeqi, huangyanchen, jinyaochu}@westlake.edu.cn + +# Abstract + +Spiking Neural Networks (SNNs) based on Transformers have garnered significant attention due to their superior performance and high energy efficiency. However, the spiking attention modules of most existing Transformer-based SNNs are adapted from those of analog Transformers, failing to fully address the issue of over-allocating attention to irrelevant contexts. To fix this fundamental yet overlooked issue, we propose a Lateral Inhibition-inspired Spiking Transformer (SpiLiFormer). It emulates the brain's lateral inhibition mechanism, guiding the model to enhance attention to relevant tokens while suppressing attention to irrelevant ones. Our model achieves state-of-the-art (SOTA) performance across multiple datasets, including CIFAR-10 (+0.45%), CIFAR-100 (+0.48%), CIFAR10-DVS (+2.70%), N-Caltech101 (+1.94%), and ImageNet-1K (+1.6%)1. Notably, on the ImageNet-1K dataset, SpiLiFormer (69.9M parameters, 4 time steps, 384 resolution) outperforms E-SpikeFormer (173.0M parameters, 8 time steps, 384 resolution), a SOTA spiking Transformer, by 0.46% using only 39% of the parameters and half the time steps. The code and model checkpoints are publicly available at https://github.com/KirinZheng/SpiLiFormer. + +# 1. Introduction + +Spiking Neural Networks (SNNs), regarded as the third generation of neural networks [18], are seen as a potential alternative to Artificial Neural Networks (ANNs) due to their biological interpretability and high energy efficiency, which stem from their event-driven properties. Transformer [25], originally designed for natural language processing tasks, has now become a dominant neural network architecture, demonstrating remarkable performance across + +various visual tasks [6, 27, 32, 42]. The success of Transformers has driven the exploration of their integration with SNNs, with the objective of enhancing SNN models' performance across various tasks and progressively bridging the performance gap with ANNs, particularly in image classification [29, 30, 37, 38], object detection [10, 17], action recognition [14, 26], and semantic segmentation [16, 29]. + +The attention module plays a crucial role in Transformer-based SNNs, significantly influencing model performance. Recent research advancements in spiking attention modules can be broadly classified into three categories: 1) Self-Spiking Attention (SSA) [36, 38, 39], which directly converts the Query $(Q)$ , Key $(K)$ , and Value $(V)$ into sparse spikes; 2) Spike-Driven Self-Attention (SDSA) [29, 30] and Spiking RWKV (S-RWKV) [40], which replace the dot product in SSA with the Hadamard product, achieving linear complexity while reducing both energy consumption and computational cost; 3) Spatial-Temporal Attention (STA) [12, 28], which refines traditional spatial attention by explicitly modeling the key temporal dependencies intrinsic to spike-based processing. + +However, we observe that most Transformer-based SNNs exhibit a phenomenon referred to as attention distraction. As illustrated in Fig. 1(a), the model disproportionately assigns high attention weights to background information during decision-making, leading to the neglect of critical object-related features and ultimately impairing classification performance. This issue stems from the fact that spiking attention largely inherits the traditional attention forward propagation paradigm. Previous studies [1, 21, 34] suggest that traditional attention mechanisms process $Q$ , $K$ , and $V$ uniformly in the mapping $softmax(QK^{\top})V$ , thereby limiting their capacity to regulate contextual sparsity and relevance. Therefore, we posit that constructing additional pathways to process $Q$ , $K$ , and $V$ separately is essential for mitigating attention distraction in Transformer-based SNNs and enhancing model performance. + +By contrast, the natural lateral inhibition mechanism in + +![](images/e2b7f11365532398f9418e66da9227cbd6113317d9026495a99ec3d312a8fe34.jpg) + +![](images/c315987cd2963b7aa7a9e77a289ba5440679566604926badceccb975f8e6d604.jpg) + +![](images/0b579d9f6094e0ca852c1da4d8c5e83e470655fa8ef46df101cf1b81666bd828.jpg) +Figure 1. Illustration of the severe attention distraction phenomenon in Transformer-based SNNs, the architecture of the Lateral Inhibition-inspired Spiking Transformer (SpiLiFormer), and the detailed processing mechanism of Feedforward-pathway Lateral Differential Inhibition (FF-LiDiff) attention and Feedback-pathway Lateral Difference Inhibition (FB-LiDiff) attention. (a) refers to the attention distraction in mainstream Transformer-based SNNs, where models excessively focus on irrelevant background information, leading to misclassification. (b) denotes the retinal lateral inhibition mechanism, illustrating how horizontal cells regulate neural responses to enhance contrast and suppress noise. (c) presents the architecture of SpiLiFormer. (d) and (e) represent the information processing flows of FF-LiDiff and FB-LiDiff attention, respectively. + +the visual system enables the brain to focus on important areas, thereby improving perception and reducing visual overload [4, 19, 22]. Specifically, when neurons within a given area are strongly activated, adjacent neurons are inhibited. This principle is first observed in the retina, where horizontal cells receive input from photoreceptors and inhibit both the stimulated photoreceptors and their neighbors, forming a center-surround receptive field (RF), as shown in Fig. 1(b). In higher visual areas, such as the primary (V1) and secondary (V2) visual cortices, a similar mechanism operates through long-range connections, which contrasts with the short-range inhibition in the retina, helping to refine visual + +processing by suppressing irrelevant stimuli and selectively enhancing salient features [41, 43]. + +Based on the short-range and long-range lateral inhibition mechanisms discussed above, we propose the Lateral Inhibition-inspired Spiking Transformer (SpiLiFormer) to address the attention distraction issue in Transformer-based SNNs, as shown in Fig. 1(c). This model consists of two types of stacked modules: 1) Feedforward-pathway Lateral Differential Inhibition (FF-LiDiff) blocks and 2) Feedback-pathway Lateral Differential Inhibition (FB-LiDiff) blocks. Moreover, it introduces the following innovations: 1) In the shallow blocks (i.e., Stage 1 and Stage 2), it explicitly in- + +corporates an inhibition differential attention mechanism by emulating the retinal short-range lateral inhibition process (see Fig. 1(d) for details); 2) In the deep blocks (i.e., Stage 3), it employs feedback-driven processing based on differential attention to simulate the long-range lateral inhibition mechanism observed in the cerebral cortex (see Fig. 1(e) for details). We evaluate our proposed SpiLiFormer on five image classification datasets, including CIFAR-10 [11], CIFAR-100 [11], CIFAR10-DVS [13], N-Caltech101 [20], and ImageNet-1K [5], demonstrating superior performance over current SOTA models. Additionally, adversarial testing and attention heatmap visualizations on the ImageNet-1K test set and noise-augmented datasets (i.e., CIFAR-10-C and ImageNet-1K-C [8]) show that SpiLiFormer effectively suppresses attention on irrelevant areas and background noise. The main contributions of this work are: + +- We propose SpiLiFormer, which incorporates a brain-inspired lateral inhibition mechanism and introduces two novel attention paradigms: FF-LiDiff attention and FB-LiDiff attention. This model effectively mitigates the attention distraction issue common in current Transformer-based SNNs, thereby enhancing image classification performance. +We evaluate SpiLiFormer across multiple datasets, achieving new SOTA performance on CIFAR-10, CIFAR100, CIFAR10-DVS, N-Caltech101, and ImageNet-1K. Specifically, on ImageNet-1K, SpiLiFormer (69.9M parameters, 4 time steps) outperforms the optimal SOTA model, E-SpikeFormer (173.0M parameters, 8 time steps), by $0.46\%$ in top-1 accuracy, while using only 0.39 times the number of parameters and half the time steps. +- We comprehensively evaluate the robustness of SpiLiFormer through attention heatmap visualizations and adversarial testing, and find that SpiLiFormer achieves more effective attention allocation and enhanced robustness compared to other baseline spiking models. + +# 2. Related Work + +# 2.1. Transformer-based SNNs + +Although convolution-based SNNs exhibit high energy efficiency, they still demonstrate a significant performance gap compared to ANNs. To address this challenge, researchers have explored integrating Transformer architectures with SNNs. Spikformer [38] was the first to incorporate the Transformer attention into SNNs, cleverly replacing the Softmax operation by leveraging the binary characteristic of spike representations to encode Query, Key, and Value. Building upon this, Spikingformer [36] introduced a pre-activation shortcut to avoid floating-point multiplication while simultaneously reducing the spike firing rate. Inspired by SEW-ResNet [7], Spike-driven Trans + +former [30] replaced the dot product with a Hadamard product and reshaped the spike residual connection based on the membrane potential to further minimize energy consumption and leverage the event-driven property. Additionally, Spiking RWKV [40] and QKFormer [37] introduced linear attention into SNNs to reduce computational complexity. Meanwhile, STA-Transformer [12] effectively integrated both spatial and temporal dynamics of spike trains to enhance model performance. However, most existing Transformer-based SNNs still adhere to the traditional attention propagation paradigm of ANNs, which results in the severe phenomenon of attention distraction, thereby limiting model performance. + +# 2.2. Models with Lateral Inhibition + +Lateral inhibition mechanisms have been widely adopted in various SNN applications. In speech recognition, Spiking-LEAF [23] emulated inner hair cell functionality and leveraged lateral inhibition feedback to enhance the neuron model, significantly improving the efficiency and noise robustness of spike encoding. In the object recognition, LISNN [2] modeled lateral inhibition by establishing spatial neighbor-order relationships between neurons, thereby enhancing the model's ability to focus on important features. In the image classification, Zhang et al. [33] proposed an adaptive self-organizing lateral inhibition strategy, where inhibition strength was adjusted according to the Euclidean distance between neurons. This approach addressed the issue of inefficient feature clustering and reduced redundancy in neuron activation. However, existing research on lateral inhibition has predominantly focused on small-scale datasets for validating model feasibility, with limited exploration of its applicability to larger and more complex datasets. Furthermore, its integration with Transformer-based SNNs remains largely unexplored. + +# 3. Method + +This section details the proposed SpiLiFormer model, covering the spiking neuron layer, overall architecture, and two lateral inhibition-inspired attention mechanisms. Additionally, we outline the training strategy and loss function. + +# 3.1. Spiking Neuron Layer + +The spiking neuron layer relies on cumulative firing to integrate spatio-temporal information into the membrane potential, which is then converted into binary spikes to drive computation in the next layer. We adopt the widely used Leaky Integrate-and-Fire (LIF) spiking neuron layer, as established in previous studies [30, 31, 37], with the following dynamics: + +$$ +U [ t ] = H [ t ] (1 - S [ t ]) + U _ {\text {r e s e t}} S [ t ], \tag {1} +$$ + +$$ +H [ t ] = U [ t - 1 ] + \frac {1}{\tau} (X [ t ] - (U [ t - 1 ] - U _ {\text {r e s e t}})), \quad (2) +$$ + +$$ +S [ t ] = \Theta (H [ t ] - U _ {t h}), \tag {3} +$$ + +where $U_{\mathrm{reset}}$ represents the reset potential, and $X[t]$ and $U[t]$ denote the current input and membrane potential at timestep $t$ , respectively. $\Theta(\cdot)$ is the Heaviside step function, which outputs 1 if $x \geq 0$ and 0 otherwise. When $H[t]$ exceeds the firing threshold $U_{th}$ , the spiking neuron fires a spike $S[t]$ , and $U[t]$ is reset to $U_{\mathrm{reset}}$ at that moment. Additionally, $\tau$ represents the membrane time constant, which controls the rate of membrane potential leakage. + +# 3.2. Overall Architecture + +The overall architecture of SpiLiFormer is illustrated in Fig. 1(c), which consists of three stages: Stage 1, Stage 2, and Stage 3, forming a three-level hierarchical structure. The input data is represented as $I \in \mathbb{R}^{T \times N \times H \times W}$ , where $N = 3$ for static image data and $N = 2$ for neuromorphic data. In Stage 1, a $4 \times 4$ patch size is utilized, mapping each patch's input feature dimension $(4 \times 4 \times N)$ to a spike-form representation of arbitrary dimension $(C)$ via the Downsampling module. This process reduces the number of tokens to $\frac{H}{4} \times \frac{W}{4}$ . The transformed features are then processed through FF-LiDiff attention and a Spiking MultiLayer Perceptron (SpMLP). FF-LiDiff attention introduces an additional forward pathway and employs a differential approach to mitigate attention distraction in the network's shallower layers (see Sec. 3.3 for details). In Stage 2, the processing mechanism is similar to Stage 1, but with a $2 \times 2$ patch size, an expanded channel dimension of $2C$ , and the number of tokens reduced to $\frac{H}{8} \times \frac{W}{8}$ . Finally, in Stage 3, we retain the $2 \times 2$ patch size, expand the channel dimension to $4C$ , and reduce the number of tokens to $\frac{H}{16} \times \frac{W}{16}$ . Additionally, a more complex FB-LiDiff attention module with feedback loops is introduced (see Sec. 3.4 for details) to replace the previous FF-LiDiff attention, further mitigating attention distraction in the model. + +The number of spiking blocks (either FF-LiDiff block or FB-LiDiff block) in each stage is denoted as $N_{1}$ , $N_{2}$ , and $N_{3}$ , respectively. Since the Downsampling, SpMLP and classification head modules are based on prior research works [29, 31, 37], the following sections will focus on the detailed description of FF-LiDiff attention and FB-LiDiff attention. + +# 3.3. Feedforward-pathway Lateral Differential Inhibition Attention + +An overview of FF-LiDiff attention is shown in Figure 1(d). For better mathematical description, in subsequent discussion, we assume $T = 1$ and only use single head attention. + +Given a spike input $X \in \mathbb{R}^{N \times D}$ ( $N$ is the token number, $D$ is the channel number), Query $(Q)$ and Key $(K)$ can be computed through learnable matrices: + +$$ +Q = B N \left(X W _ {Q}\right), K = B N \left(X W _ {K}\right), \tag {4} +$$ + +where $W_{Q}, W_{K} \in \mathbb{R}^{D \times D}$ and $Q, K \in \mathbb{R}^{N \times D}$ . + +To implement the lateral inhibition mechanism, we first split $Q$ along the channel dimension into two parts, obtaining $Q_{1}$ and $Q_{2}$ , before computing the attention matrix. Then, all of $Q_{1}, Q_{2}$ , and $K$ are transformed into spike trains via their respective spiking neuron layers: + +$$ +Q _ {s 1} = \mathcal {S N} (Q _ {1}), Q _ {s 2} = \mathcal {S N} (Q _ {2}), K _ {s} = \mathcal {S N} (K), \tag {5} +$$ + +where $Q_{s1}, Q_{s2} \in \mathbb{R}^{N \times \frac{D}{2}}$ , $K_{s} \in \mathbb{R}^{N \times D}$ . Next, we sum along the channel dimension to obtain the excitatory attention $A_{e}$ and inhibitory attention $A_{i}$ . Specifically, $A_{e}$ and $A_{i}$ are $N \times 1$ token attention vectors, representing the importance of different tokens from the perspectives of excitation and inhibition, respectively. Finally, we compute the element-wise difference between $A_{e}$ and $A_{i}$ , and pass the aggregated attention through a spiking neural layer to obtain binary attention $A_{\text{combined}}$ . This attention is then applied as a token-wise mask to $K_{s}$ via the Hadamard product $\otimes$ . The above process can be described as: + +$$ +A _ {e} = \sum_ {j} ^ {\frac {D}{2}} Q _ {s 1} ^ {i, j}, A _ {i} = \sum_ {j} ^ {\frac {D}{2}} Q _ {s 2} ^ {i, j}, \tag {6} +$$ + +$$ +A _ {\text {c o m b i n e d}} = \mathcal {S N} \left(A _ {e} - A _ {i}\right), \tag {7} +$$ + +$$ +X ^ {\prime} = A _ {\text {c o m b i n e d}} \otimes K _ {s}. \tag {8} +$$ + +# 3.4. Feedback-pathway Lateral Differential Inhibition Attention + +FB-LiDiff attention requires two forward propagations as shown in Figure 1(e). Specifically, during the first forward propagation, FB-LiDiff attention takes spike-form $S_{\mathrm{stage3}}^{Ds}$ from the Downsampling module in Stage 3 or takes spikeform $S_{\mathrm{stage3}}^i (i \in [1, N_3])$ from the $i$ -th preceding FB-LiDiff block as input. Given an input $X_1^m$ to the $m$ -th FB-LiDiff attention, the first forward propagation can be described as follows: + +$$ +Y _ {s} ^ {1, m} = \mathcal {S N} \left(B N \left(X _ {1} ^ {m} W _ {Y} ^ {m}\right)\right), Y \in \{Q, K, V \}, \tag {9} +$$ + +$$ +A t t n ^ {1, m} = Q _ {s} ^ {1, m} \odot \left(\left(K _ {s} ^ {1, m}\right) ^ {\top} \odot V _ {s} ^ {1, m}\right), \tag {10} +$$ + +$$ +O _ {\text {s t a g e 3}} ^ {m} = \mathcal {S N} \left(B N \left(A t t n ^ {1, m} W _ {a t t n} ^ {m}\right)\right), \tag {11} +$$ + +where $X_{1}^{m}\in \mathbb{R}^{N\times D}$ ( $X_{1}^{m}\in \{S_{\mathrm{stage3}}^{Ds},S_{\mathrm{stage3}}^{i}\}$ ), and $Y_{s}^{1,m}$ ( $Y\in \{Q,K,V\}$ ) represents the corresponding spiking representation of $Q$ , $K$ and $V$ in the $m$ -th FB-LiDiff attention. $W_{z}^{m}$ ( $z\in \{Q,K,V,attn\}$ ) and $O_{\mathrm{stage3}}^{m}$ denote the learnable weight matrix and the spike-form output of the $m$ -th FB-LiDiff attention, respectively. $\odot$ refers to the dot product. + +Before the second forward propagation, we use the output $O_{\mathrm{stage3}}^{N_3}$ from the last block of Stage 3 as feedback information and return it to each FB-LiDiff attention. To filter the feedback information and further mitigate the attention distraction problem, we first introduce learnable prompts $W_{p1}$ and $W_{p2}$ , where $W_{p1}, W_{p2} \in \mathbb{R}^{1 \times D}$ . Next, we compute the dot product between $W_{p1}$ and $O_{\mathrm{stage3}}^{N_3}$ , and similarly between $W_{p2}$ and $O_{\mathrm{stage3}}^{N_3}$ . These dot products are then broadcast along the token dimension. The results are scaled to the range [0, 1] using the clamp(\cdot) function, forming the excitatory feedback attention $A_e^{\mathrm{FB}}$ and inhibitory feedback attention $A_i^{\mathrm{FB}}$ . Subsequently, we compute the element-wise difference between $A_e^{\mathrm{FB}}$ and $A_i^{\mathrm{FB}}$ to obtain the aggregated attention $A_{\mathrm{combined}}^{\mathrm{FB}}$ . Finally, we apply a linear transformation to $A_{\mathrm{combined}}^{\mathrm{FB}}$ and pass it through the spiking neuron layer to align it with the feature space of different FB-LiDiff attentions. These processes are formulated as: + +$$ +A _ {e} ^ {\mathrm {F B}} = \operatorname {c l a m p} \left(O _ {\text {s t a g e 3}} ^ {N _ {3}} \odot W _ {p 1}, 0, 1\right), \tag {12} +$$ + +$$ +A _ {i} ^ {\mathrm {F B}} = \operatorname {c l a m p} \left(O _ {\text {s t a g e 3}} ^ {N _ {3}} \odot W _ {p 2}, 0, 1\right), \tag {13} +$$ + +$$ +A _ {\text {c o m b i n e d}} ^ {\mathrm {F B}} = A _ {e} ^ {\mathrm {F B}} - A _ {i} ^ {\mathrm {F B}}, \tag {14} +$$ + +$$ +X _ {\mathrm {F B}} ^ {m} = \mathcal {S N} \left(B N \left(A _ {\text {c o m b i n e d}} ^ {\mathrm {F B}} W _ {\mathrm {F B}} ^ {m}\right)\right). \tag {15} +$$ + +The second forward propagation begins by taking the processed forward representation $S_{\mathrm{stage3}}^{Ds}$ and the corresponding feedback information $X_{\mathrm{FB}}^{1}$ as input to the first FB-LiDiff attention. The objective is to reduce model energy consumption by eliminating redundant computations in the non-feedback loop (i.e., Stage 1 and Stage 2). A detailed energy consumption analysis can be found in Appendix A.3. Additionally, during the second forward propagation, we share all parameter weights across the FB-LiDiff blocks to minimize the number of model parameters. Given the input $X_{2}^{m}$ , the process described above can be formulated as: + +$$ +Y _ {s} ^ {2, m} = \mathcal {S N} \left(B N \left(X _ {2} ^ {m} W _ {Y} ^ {m}\right)\right), Y \in \{Q, K \}, \tag {16} +$$ + +$$ +V _ {s} ^ {2, m} = \mathcal {S N} \left(B N \left(X _ {2} ^ {m} W _ {V} ^ {m} + X _ {\mathrm {F B}} ^ {m} W _ {V} ^ {m}\right)\right), \tag {17} +$$ + +$$ +A t t n ^ {2, m} = Q _ {s} ^ {2, m} \odot \left(\left(K _ {s} ^ {2, m}\right) ^ {\top} \odot V _ {s} ^ {2, m}\right), \tag {18} +$$ + +$$ +\left(O _ {\text {s t a g e} 3} ^ {m}\right) ^ {\prime} = \mathcal {S N} \left(B N \left(A t t n ^ {2, m} W _ {a t t n} ^ {m}\right)\right). \tag {19} +$$ + +# 3.5. Training Strategy and Loss Function + +We employ a surrogate gradient-based method to overcome the non-differentiability of the spiking neuron activation in Eq. (3), enabling direct training of the model from scratch, as demonstrate in previous works [29, 36-38, 40]. In order to make better use of the spiking representations obtained from two forward propagations, we adjust the model's loss function as follows: + +$$ +\mathcal {L} _ {1} = \mathcal {L} _ {C E} \left(C H \left(O _ {\text {s t a g e 3}} ^ {N _ {3}}\right), y\right), \tag {20} +$$ + +$$ +\mathcal {L} _ {2} = \mathcal {L} _ {C E} \left(C H \left(\left(O _ {\text {s t a g e 3}} ^ {N _ {3}}\right) ^ {\prime}\right), y\right), \tag {21} +$$ + +$$ +\mathcal {L} _ {\text {S p i L i F o r m e r}} = \alpha \mathcal {L} _ {1} + (1 - \alpha) \mathcal {L} _ {2}, \tag {22} +$$ + +where $\alpha$ is a hyperparameter that dynamically balances the losses from two forward propagations. Unless explicitly specified otherwise, $\alpha$ is set to 0.5 in the subsequent experiments. The choice of this value is validated through an ablation study, as detailed in Appendix A.4. Additionally, $y$ represents the true label; $O_{\mathrm{stage3}}^{N_3}$ and $(O_{\mathrm{stage3}}^{N_3})'$ denote the outputs from the first and second forward propagations of the last FB-LiDiff block, respectively; $CH(\cdot)$ represents the classification head module, as described in previous studies [29, 30]; and $\mathcal{L}_{CE}$ denotes the cross-entropy loss function. + +# 4. Experiments + +We first evaluate the classification performance of SpiLiFormer model on the large-scale ImageNet-1K dataset [5]. Next, we assess the performance of SpiLiFormer on smaller-scale static datasets, including CIFAR-10 [11] and CIFAR-100 [11]. Additionally, we evaluate SpiLiFormer on two popular neuromorphic datasets: CIFAR10-DVS [13] and N-Caltech101 [20]. Finally, we visualize attention heatmaps and conduct adversarial tests to demonstrate SpiLiFormer's effectiveness in mitigating attention distraction and enhancing robustness. + +# 4.1. Results on ImageNet-1K + +In this experiment, model training is divided into two phases. In the first phase, we use the AdamW optimizer with a base learning rate of $6 \times 10^{-4}$ . The actual learning rate is calculated by multiplying the base rate by the batch size divided by 256. Additionally, we apply a layer-wise learning rate decay strategy with a decay factor of 1.0 and weight decay of 0.05 to train SpiLiFormer for 200 epochs. In the second phase, we decrease the base learning rate to $2 \times 10^{-6}$ and continue training for an additional 20 epochs using an epoch-wise learning rate decay strategy. Throughout this entire experiment, we incorporate several data augmentation techniques in alignment with the DeiT approach + +
MethodsTypeArchitectureInput SizeParam(M)Power(mJ)Time StepTop-1 Acc(%)
ViT [6]ANNViT-B/1638486.59254.84177.90
DeiT [24]ANNDeiT-B22486.5980.50181.80
ANNDeiT-B38486.59254.84183.10
Swin [15]ANNSwin Transformer-B22487.7770.84183.50
ANNSwin Transformer-B38487.77216.20184.50
SEW ResNet [7]SNNSEW-ResNet-15222460.1912.89469.26
Spikformer [38]SNNSpikformer-8-76822466.3421.48474.81
Spikingformer [36]SNNSpikingformer-8-76822466.3413.68475.85
S-Transformer [30]SNNS-Transformer-8-51222429.681.13171.68
SNNS-Transformer-8-51222429.684.50474.57
SNNS-Transformer-8-76828866.346.09477.07
QKFormer [37]SNNHST-10-76822464.9638.91484.22
SNNHST-10-768*28864.9664.27485.25
SNNHST-10-76838464.96113.64485.65
E-SpikeFormer [31]SNNE-SpikeFormer224173.035.64*84.7
SNNE-SpikeFormer224173.054.78*85.1
SNNE-SpikeFormer384173.0-8*86.2
SpiLiFormer (Ours)SNNSpiLiFormer-10-76822469.1011.77181.54
SNNSpiLiFormer-10-76822469.1044.17485.82
SNNSpiLiFormer-10-768*28869.1073.52486.62
SNNSpiLiFormer-10-768**38469.10129.45486.66
+ +Table 1. Performance comparison on ImageNet-1K. "SpiLiFormer-L-D" represents Lateral Inhibition-inspired Transformer with $L$ blocks and a $D$ -dimensional channel. \* and \*\* denote input resolutions of $288^{2}$ and $384^{2}$ during inference, respectively. \* denotes an integer-valued quantized training approach [17], in which activation values are restricted to integers during training and later expanded into spike trains during inference. + +[24]. These techniques include RandAugment [3], random erasing [35], and stochastic depth [9]. The architecture of SpiLiFormer is composed of 1, 2, and 7 blocks across its three stages, respectively. + +Tab. 1 demonstrates the superior performance of our proposed SpiLiFormer, significantly outperforming current SNNs models. Specifically, SpiLiFormer (69.10M) achieves $86.66\%$ top-1 accuracy and $98.116\%$ top-5 accuracy on ImageNet-1K. To our best knowledge, SpiLiFormer is currently the SOTA model for Transformer-based SNNs with direct training. + +We compare SpiLiFormer against several baseline spiking models under the same image resolution of 224, a time step of 4, and a comparable number of model parameters. SpiLiFormer outperforms Spikformer (66.34M), Spikingformer (66.34M), and QKFormer (64.96M) by $11.01\%$ , $9.97\%$ , and $1.6\%$ , respectively. Moreover, at a higher input resolution of 288, SpiLiFormer outperforms S-Transformer and QKFormer by $9.55\%$ and $1.37\%$ , respectively. Similarly, at a resolution of 384, it surpasses QKFormer by $1.01\%$ . Additionally, we compare SpiLiFormer (69.10M, 4 time steps, 384 resolution) with E-SpikeFormer (173.0M, 8 + +time steps, 384 resolution), the current SOTA Transformer-based SNNs model. It achieves a $0.46\%$ improvement with only $39\%$ of parameters and half the time steps. + +Finally, compared to the mainstream ANN-based Swin Transformer-B (384 resolution), SpiLiFormer achieves a $2.16\%$ performance improvement while reducing energy consumption by $40\%$ and decreasing the number of parameters by 18.67M. + +# 4.2. Results on CIFAR Datasets + +In this experiment, we use the AdamW optimizer with a learning rate of $1 \times 10^{-3}$ and apply a cosine learning rate decay strategy. The weight decay is $6 \times 10^{-2}$ , the number of time steps is 4, and SpiLiFormer is trained for 400 epochs with a batch size of 64. The network architecture consists of 1, 1, and 2 blocks across its three stages, respectively. + +The performance of SpiLiFormer on CIFAR datasets is shown in the upper part of Tab. 2. For CIFAR-10, our proposed model achieves $96.63\%$ accuracy with 7.04M parameters. Compared to QKFormer, the suboptimal SOTA model, SpiLiFormer improves performance by $0.45\%$ while increasing the parameters by only $0.04\%$ . Similarly, for + +
DatasetsMethodsArchitectureParam(M)Time StepTop-1 Acc(%)
CIFAR-10Spikformer [38]Spikformer-4-3849.32495.51
Spikingformer [36]Spikingformer-4-3849.32495.81
S-Transformer [30]S-Transformer-2-51210.28495.60
QKFormer [37]HST-4-3846.74496.18
SpiLiFormer (Ours)SpiLiFormer-4-3847.04496.63
CIFAR-100Spikformer [38]Spikformer-4-3849.32478.21
Spikingformer [36]Spikingformer-4-3849.32479.21
S-Transformer [30]S-Transformer-2-51210.28478.4
QKFormer [37]HST-4-3846.74481.15
SpiLiFormer (Ours)SpiLiFormer-4-3847.04481.63
CIFAR10-DVSSpikformer [38]Spikformer-2-2562.571680.9
Spikingformer [36]Spikingformer-2-2562.571681.3
S-Transformer [30]S-Transformer-2-2562.571680.0
QKFormer [37]HST-2-2561.501684.0
SpiLiFormer (Ours)SpiLiFormer-2-2561.571686.7
N-Caltech101Spikformer [38]Spikformer-2-2562.571683.6
Spikingformer [36]Spikingformer-2-2562.571685.91
S-Transformer [30]S-Transformer-2-2562.571686.3
QKFormer [37]HST-2-2561.501687.24
SpiLiFormer (Ours)SpiLiFormer-2-2561.571689.18
+ +Table 2. Performance comparison on CIFAR-10, CIFAR-100, CIFAR10-DVS, and N-Caltech101 datasets. + +CIFAR-100, our proposed model achieves $81.63\%$ accuracy with 7.04M parameters. Specifically, it outperforms QKFormer by $0.48\%$ while requiring just $0.04\%$ more parameters. + +# 4.3. Results on Neuromorphic Datasets + +In this experiment, we use a smaller SpiLiFormer architecture, structured with 0, 1, and 1 blocks across its three stages, respectively. Similarly, we adopt the AdamW optimizer with an initial learning rate of $5 \times 10^{-3}$ , apply a learning rate decay strategy, and set the weight decay to 0.06. The number of time steps is set to 16, and SpiLiFormer is trained for 120 epochs. + +The experimental results of the model on the neuromorphic datasets are presented in the lower half of Tab. 2. Compared to the suboptimal model, QKFormer, SpiLi-Former significantly improves performance by $2.7\%$ on the CIFAR10-DVS dataset and $1.94\%$ on the N-Caltech101 dataset, with only a $0.13\%$ increase in parameters. Additionally, we compare the proposed model with other baseline models (i.e., Spikformer, Spikingformer, and S-Transformer). It achieves performance gains of up to $5.8\%$ on CIFAR10-DVS and $5.55\%$ on N-Caltech101 while reducing the parameter count by 0.87M. + +# 4.4. Visualization and Model Robustness Analysis + +We visualize the attention heatmap of SpiLiFormer to assess whether it effectively alleviates the attention distraction is + +sue. Specifically, we examine the attention heatmap of the last FB-LiDiff block in SpiLiFormer Stage 3 under two conditions: standard dataset evaluation and noisy dataset testing. + +In the standard dataset evaluation, the two sets of images in Fig. 1(a) show examples that are misclassified by the baseline model but correctly classified by SpiLiFormer on the ImageNet-1K test set. We observe that our proposed model effectively eliminates the attention distraction phenomenon, thereby improving classification accuracy. More comparison images are provided in Fig. 4 within the appendix. + +In noisy dataset testing, we evaluate model attention heatmaps using two noisy datasets: CIFAR10-C and ImageNet-1K-C, as shown in Fig. 2. Compared to other baseline models, SpiLiFormer correctly classifies objects or significantly increases confidence in the correct category by focusing on object-relevant information while remaining unaffected by background noise. + +In addition, we employ white-box adversarial attacks, including FGSM and PGD, to assess the model's robustness across the four datasets mentioned above. As shown in Fig. 3, we find that SpiLiFormer not only enhances model performance but also improves robustness against adversarial and common noise attacks, outperforming QKFormer. + +![](images/3ee79235ecfff21615e4b3a3527c21f7a0fd7f2da9e26794f5f51bc9404e369f.jpg) +Figure 2. Visualization of corrupted versions of datasets. From left to right, we display the original image, followed by the attention heatmaps obtained from Spike-driven Transformer, QK-Former, and SpiLiFormer (ours), respectively. The information below each image includes the true label, the predicted label for each model, and the confidence score for the predicted category. Red indicates incorrect classification, while green represents correct classification. For more visualizations, refer to Fig. 6 and Fig. 7 in the appendix. + +# 4.5. Ablation Study + +We conduct an ablation study on CIFAR-10 (static) and CIFAR10-DVS (neuromorphic) to assess the impact of FF-LiDiff and FB-LiDiff attention on model performance and inference latency. All ablation experiments follow the training details in Sec. 4.2 and Sec. 4.3, unless otherwise specified. + +Intuitively, we set up two experimental scenarios: 1) For the case without FF-LiDiff attention (w/o FF-LiDiff attention), we use $Attn = \mathcal{SN}(\sum_{j}^{D}Q^{i,j})$ to replace the steps in Eq. (6) and Eq. (7); 2) For the case without FB-LiDiff attention (w/o FB-LiDiff attention), we perform a single forward pass with $Attn = Q\odot (K^{\top}\odot V)$ to calculate the attention, replacing the steps outlined in Eq. (12) to Eq. (19). + +As shown in Tab. 3, we observe that the model exhibits varying degrees of performance degradation on both datasets under the two aforementioned scenarios. Specifi + +![](images/e6ef6abe0370a583876cef3e5dafedede638f8c058874fa4141507144e1a9dc7.jpg) + +![](images/46961d053c52fe5fc8dc5698664a551ca706b938c8f8c2c9a6ec4c28a06b33dc.jpg) +Figure 3. Adversarial robustness comparison between QKFormer and our proposed model on CIFAR-100 (first row) and ImageNet-1K (second row). For PGD, the attack strength is set to 8/255, with a step size of 2/255 per iteration. The comparison results on other datasets are provided in Tab. 6 of within the appendix. + +![](images/df3215a47b0bcca1cd92194ebc47ebec483550843c6ac3f1ad6679b9eddbf938.jpg) + +![](images/7f732a35fdc6419c6228c65d0a6763161bb338b6fdedd6c9a15b6eb4c7703c18.jpg) + +cally, performance exhibits a decrease of $0.31\%$ and $0.62\%$ on the CIFAR-10 dataset, and $1.6\%$ and $3.0\%$ on the CIFAR10-DVS dataset. These results highlight the effectiveness of FB-LiDiff in mitigating attention distraction and improving model performance. However, its additional forward pass introduces higher inference latency, particularly on static datasets, with a $21 - 23\%$ increase as shown in Tab. 5 in the appendix. + +
DatasetsMethodsTop-1 Acc(%)
CIFAR-10w/o FF-LiDiff Attention96.32
w/o FB-LiDiff Attention96.01
SpiLiFormer(ours)96.63
CIFAR10-DVSw/o FF-LiDiff Attention85.1
w/o FB-LiDiff Attention83.7
SpiLiFormer(ours)86.7
+ +Table 3. Ablation study on two attention modules + +# 5. Conclusion + +In this paper, we identify attention distraction in mainstream Transformer-based SNNs, a critical issue that limits performance. Inspired by the brain's lateral inhibition mechanism, we propose SpiLiFormer, incorporating FF-LiDiff and FB-LiDiff attention to simulate the inhibitory-excitatory interaction process, guiding the model to focus on the object rather than the irrelevant background. Experimental results show that SpiLiFormer outperforms current Transformer-based SNNs, achieving SOTA performance on five image classification datasets. Additionally, adversarial testing and attention visualization analysis demonstrate its robustness and effectiveness in alleviating attention distraction. + +# 6. Acknowledgments + +This research was supported in part by Zhejiang Provincial Natural Science Foundation of China under the Grant No. LD25F020006. + +# References + +[1] Hila Chefer, Idan Schwartz, and Lior Wolf. Optimizing relevance maps of vision transformers improves robustness. Advances in Neural Information Processing Systems, 35: 33618-33632, 2022. 1 +[2] Xiang Cheng, Yunzhe Hao, Jiaming Xu, and Bo Xu. Lisnn: Improving spiking neural networks with lateral interactions for robust object recognition. In *IJCAI*, pages 1519–1525. Yokohama, 2020. 3 +[3] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 6 +[4] Joseph Del Rosario, Stefano Coletta, Soon Ho Kim, Zach Mobile, Kayla Peelman, Brice Williams, Alan J Otsuki, Alejandra Del Castillo Valerio, Kendell Worden, Lou T Blanpain, et al. Lateral inhibition in v1 controls neural and perceptual contrast sensitivity. Nature Neuroscience, pages 1-12, 2025. 2 +[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 3, 5 +[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, G Heigold, S Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 6 +[7] Wei Fang, Zhaofei Yu, Yanqi Chen, Tiejun Huang, Timothée Masquelier, and Yonghong Tian. Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems, 34:21056-21069, 2021. 3, 6 +[8] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2019. 3 +[9] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 646-661. Springer, 2016. 6 +[10] Seijoon Kim, Seongsik Park, Byunggook Na, and Sungroh Yoon. Spiking-yolo: spiking neural network for energy-efficient object detection. In Proceedings of the AAAI conference on artificial intelligence, pages 11270-11277, 2020. 1 +[11] Alex Krizhevsky et al. Learning multiple layers of features from tiny images. 2009. 3, 5 + +[12] Donghyun Lee, Yuhang Li, Youngeun Kim, Shiting Xiao, and Priyadarshini Panda. Spiking transformer with spatial-temporal attention. arXiv preprint arXiv:2409.19764, 2024. 1, 3 +[13] Hongmin Li, Hanchao Liu, Xiangyang Ji, Guoqi Li, and Luping Shi. Cifar10-dvs: an event-stream dataset for object classification. Frontiers in neuroscience, 11:309, 2017. 3,5 +[14] Qianhui Liu, Dong Xing, Huajin Tang, De Ma, and Gang Pan. Event-based action recognition using motion information and spiking neural networks. In *IJCAI*, pages 1743–1749, 2021. 1 +[15] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 6 +[16] Xianlei Long, Xiaxin Zhu, Fangming Guo, Chao Chen, Xiangwei Zhu, Fuqiang Gu, Songyu Yuan, and Chunlong Zhang. Spike-brgnet: Efficient and accurate event-based semantic segmentation with boundary region-guided spiking neural networks. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1 +[17] Xinhao Luo, Man Yao, Yuhong Chou, Bo Xu, and Guoqi Li. Integer-valued training and spike-driven inference spiking neural network for high-performance and energy-efficient object detection. In European Conference on Computer Vision, pages 253-272. Springer, 2024. 1, 6 +[18] Wolfgang Maass. Networks of spiking neurons: the third generation of neural network models. Neural networks, 10 (9):1659-1671, 1997. 1 +[19] James S Magnuson, Anne Marie Crinnion, Sahil Luthra, Phoebe Gaston, and Samantha Grubb. Contra assertions, feedback improves word recognition: How feedback and lateral inhibition sharpen signals over noise. Cognition, 242: 105661, 2024. 2 +[20] Garrick Orchard, Ajinkya Jayawant, Gregory K Cohen, and Nitish Thakor. Converting static image datasets to spiking neuromorphic datasets using saccades. Frontiers in neuroscience, 9:437, 2015. 3, 5 +[21] Arda Sahiner, Tolga Ergen, Batu Ozturkler, John Pauly, Morteza Mardani, and Mert Pilanci. Unraveling attention via convex duality: Analysis and interpretations of vision transformers. In International Conference on Machine Learning, pages 19050-19088. PMLR, 2022. 1 +[22] Stephen R Shaw. Retinal resistance barriers and electrical lateral inhibition. Nature, 255(5508):480-483, 1975. 2 +[23] Zeyang Song, Jibin Wu, Malu Zhang, Mike Zheng Shou, and Haizhou Li. Spiking-leaf: A learnable auditory front-end for spiking neural networks. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 226-230. IEEE, 2024. 3 +[24] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 6 + +[25] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1 +[26] Alex Vicente-Sola, Davide L Manna, Paul Kirkland, Gaetano Di Caterina, and Trevor J Bihl. Spiking neural networks for event-based action recognition: A new task to understand their advantage. Neurocomputing, 611:128657, 2025. 1 +[27] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 568-578, 2021. 1 +[28] Yuchen Wang, Kexin Shi, Chengzhuo Lu, Yuguo Liu, Malu Zhang, and Hong Qu. Spatial-temporal self-attention for asynchronous spiking neural networks. In IJCAI, pages 3085–3093, 2023. 1 +[29] Man Yao, JiaKui Hu, Tianxiang Hu, Yifan Xu, Zhaokun Zhou, Yonghong Tian, XU Bo, and Guoqi Li. Spike-driven transformer v2: Meta spiking neural network architecture inspiring the design of next-generation neuromorphic chips. In The Twelfth International Conference on Learning Representations, 2024. 1, 4, 5 +[30] Man Yao, Jiakui Hu, Zhaokun Zhou, Li Yuan, Yonghong Tian, Bo Xu, and Guoqi Li. Spike-driven transformer. Advances in neural information processing systems, 36, 2024. 1, 3, 5, 6, 7 +[31] Man Yao, Xuerui Qiu, Tianxiang Hu, Jiakui Hu, Yuhong Chou, Keyu Tian, Jianxing Liao, Luziwei Leng, Bo Xu, and Guoqi Li. Scaling spike-driven transformer with efficient spike firing approximation training. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 3, 4, 6 +[32] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch on imagenet. In Proceedings of the IEEE/CVF international conference on computer vision, pages 558-567, 2021. 1 +[33] Geng Zhang, Shuangming Yang, Xuetao Zhang, and Badong Chen. Biologically plausible unsupervised learning for self-organizing spiking neural networks with dendritic computation. Neurocomputing, page 129707, 2025. 3 +[34] Xuechen Zhang, Xiangyu Chang, Mingchen Li, Amit Roy-Chowdhury, Jiasi Chen, and Samet Oymak. Selective attention: Enhancing transformer through principled context control. Advances in Neural Information Processing Systems, 37:11061-11086, 2025. 1 +[35] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In Proceedings of the AAAI conference on artificial intelligence, pages 13001-13008, 2020. 6 +[36] Chenlin Zhou, Liutao Yu, Zhaokun Zhou, Zhengyu Ma, Han Zhang, Huihui Zhou, and Yonghong Tian. Spikingformer: Spike-driven residual learning for transformer-based spiking neural network. arXiv preprint arXiv:2304.11954, 2023. 1, 3, 5, 6, 7 + +[37] Chenlin Zhou, Han Zhang, Zhaokun Zhou, Liutao Yu, Liwei Huang, Xiaopeng Fan, Li Yuan, Zhengyu Ma, Huihui Zhou, and Yonghong Tian. Qkformer: Hierarchical spiking transformer using qk attention. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1, 3, 4, 6, 7 +[38] Zhaokun Zhou, Yuesheng Zhu, Chao He, Yaowei Wang, YAN Shuicheng, Yonghong Tian, and Li Yuan. Spikformer: When spiking neural network meets transformer. In The Eleventh International Conference on Learning Representations, 2023. 1, 3, 5, 6, 7 +[39] Zhaokun Zhou, Kaiwei Che, Wei Fang, Keyu Tian, Yuesheng Zhu, Shuicheng Yan, Yonghong Tian, and Li Yuan. Spikformer v2: Join the high accuracy club on imagenet with an snn ticket. arXiv preprint arXiv:2401.02020, 2024. 1 +[40] Rui-Jie Zhu, Qihang Zhao, Guoqi Li, and Jason K Eshraghian. Spikegpt: Generative pre-trained language model with spiking neural networks. arXiv preprint arXiv:2302.13939, 2023. 1, 3, 5 +[41] Shirui Zhu, Tao Xie, Ziyu Lv, Yan-Bing Leng, Yu-Qi Zhang, Runze Xu, Jingrun Qin, Ye Zhou, Vellaismy AL Roy, and Su-Ting Han. Hierarchies in visual pathway: functions and inspired artificial vision. Advanced Materials, 36(6): 2301986, 2024. 2 +[42] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2021. 1 +[43] Petr Znamenskiy, Mean-Hwan Kim, Dylan R Muir, M Florencia Iacaruso, Sonja B Hofer, and Thomas D Mrsic-Flogel. Functional specificity of recurrent inhibition in visual cortex. Neuron, 112(6):991-1000, 2024. 2 \ No newline at end of file diff --git a/2025/SpiLiFormer_ Enhancing Spiking Transformers with Lateral Inhibition/images.zip b/2025/SpiLiFormer_ Enhancing Spiking Transformers with Lateral Inhibition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e3c6dcda3268305b4eae87b52d0c02b6df2e59d8 --- /dev/null +++ b/2025/SpiLiFormer_ Enhancing Spiking Transformers with Lateral Inhibition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:378b336eaa210fe7c6a88b968145274a0d9db74ade0eacf61d2eff22eb16775b +size 759932 diff --git a/2025/SpiLiFormer_ Enhancing Spiking Transformers with Lateral Inhibition/layout.json b/2025/SpiLiFormer_ Enhancing Spiking Transformers with Lateral Inhibition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dadbe225521b4cf83fdd2d5ca79f28aebb65e0a9 --- /dev/null +++ b/2025/SpiLiFormer_ Enhancing Spiking Transformers with Lateral Inhibition/layout.json @@ -0,0 +1,9148 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 102, + 526, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 102, + 526, + 121 + ], + "spans": [ + { + "bbox": [ + 83, + 102, + 526, + 121 + ], + "type": "text", + "content": "SpiLiFormer: Enhancing Spiking Transformers with Lateral Inhibition" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 139, + 141, + 470, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 141, + 470, + 171 + ], + "spans": [ + { + "bbox": [ + 139, + 141, + 470, + 171 + ], + "type": "text", + "content": "Zeqi Zheng" + }, + { + "bbox": [ + 139, + 141, + 470, + 171 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 139, + 141, + 470, + 171 + ], + "type": "text", + "content": ", Yanchen Huang" + }, + { + "bbox": [ + 139, + 141, + 470, + 171 + ], + "type": "inline_equation", + "content": "^{2,3*}" + }, + { + "bbox": [ + 139, + 141, + 470, + 171 + ], + "type": "text", + "content": ", Yingchao Yu" + }, + { + "bbox": [ + 139, + 141, + 470, + 171 + ], + "type": "inline_equation", + "content": "^{4,2}" + }, + { + "bbox": [ + 139, + 141, + 470, + 171 + ], + "type": "text", + "content": ", Zizheng Zhu" + }, + { + "bbox": [ + 139, + 141, + 470, + 171 + ], + "type": "inline_equation", + "content": "^{2,5}" + }, + { + "bbox": [ + 139, + 141, + 470, + 171 + ], + "type": "text", + "content": ", Junfeng Tang" + }, + { + "bbox": [ + 139, + 141, + 470, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 139, + 141, + 470, + 171 + ], + "type": "text", + "content": ", Zhaofei Yu" + }, + { + "bbox": [ + 139, + 141, + 470, + 171 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 139, + 141, + 470, + 171 + ], + "type": "text", + "content": ", Yaochu Jin" + }, + { + "bbox": [ + 139, + 141, + 470, + 171 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 171, + 527, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 171, + 527, + 200 + ], + "spans": [ + { + "bbox": [ + 82, + 171, + 527, + 200 + ], + "type": "text", + "content": "1Zhejiang University 2Westlake University 3Nanjing University 4Donghua University 5University of Electronic Science and Technology of China 6Peking University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 164, + 201, + 446, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 201, + 446, + 213 + ], + "spans": [ + { + "bbox": [ + 164, + 201, + 446, + 213 + ], + "type": "text", + "content": "{zhengzeqi, huangyanchen, jinyaochu}@westlake.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 240, + 200, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 240, + 200, + 253 + ], + "spans": [ + { + "bbox": [ + 151, + 240, + 200, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 266, + 297, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 297, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 297, + 531 + ], + "type": "text", + "content": "Spiking Neural Networks (SNNs) based on Transformers have garnered significant attention due to their superior performance and high energy efficiency. However, the spiking attention modules of most existing Transformer-based SNNs are adapted from those of analog Transformers, failing to fully address the issue of over-allocating attention to irrelevant contexts. To fix this fundamental yet overlooked issue, we propose a Lateral Inhibition-inspired Spiking Transformer (SpiLiFormer). It emulates the brain's lateral inhibition mechanism, guiding the model to enhance attention to relevant tokens while suppressing attention to irrelevant ones. Our model achieves state-of-the-art (SOTA) performance across multiple datasets, including CIFAR-10 (+0.45%), CIFAR-100 (+0.48%), CIFAR10-DVS (+2.70%), N-Caltech101 (+1.94%), and ImageNet-1K (+1.6%)1. Notably, on the ImageNet-1K dataset, SpiLiFormer (69.9M parameters, 4 time steps, 384 resolution) outperforms E-SpikeFormer (173.0M parameters, 8 time steps, 384 resolution), a SOTA spiking Transformer, by 0.46% using only 39% of the parameters and half the time steps. The code and model checkpoints are publicly available at https://github.com/KirinZheng/SpiLiFormer." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 555, + 135, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 135, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 135, + 567 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 576, + 296, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 296, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 296, + 673 + ], + "type": "text", + "content": "Spiking Neural Networks (SNNs), regarded as the third generation of neural networks [18], are seen as a potential alternative to Artificial Neural Networks (ANNs) due to their biological interpretability and high energy efficiency, which stem from their event-driven properties. Transformer [25], originally designed for natural language processing tasks, has now become a dominant neural network architecture, demonstrating remarkable performance across" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 242, + 555, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 242, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 555, + 327 + ], + "type": "text", + "content": "various visual tasks [6, 27, 32, 42]. The success of Transformers has driven the exploration of their integration with SNNs, with the objective of enhancing SNN models' performance across various tasks and progressively bridging the performance gap with ANNs, particularly in image classification [29, 30, 37, 38], object detection [10, 17], action recognition [14, 26], and semantic segmentation [16, 29]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 327, + 556, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 556, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 556, + 495 + ], + "type": "text", + "content": "The attention module plays a crucial role in Transformer-based SNNs, significantly influencing model performance. Recent research advancements in spiking attention modules can be broadly classified into three categories: 1) Self-Spiking Attention (SSA) [36, 38, 39], which directly converts the Query " + }, + { + "bbox": [ + 313, + 327, + 556, + 495 + ], + "type": "inline_equation", + "content": "(Q)" + }, + { + "bbox": [ + 313, + 327, + 556, + 495 + ], + "type": "text", + "content": ", Key " + }, + { + "bbox": [ + 313, + 327, + 556, + 495 + ], + "type": "inline_equation", + "content": "(K)" + }, + { + "bbox": [ + 313, + 327, + 556, + 495 + ], + "type": "text", + "content": ", and Value " + }, + { + "bbox": [ + 313, + 327, + 556, + 495 + ], + "type": "inline_equation", + "content": "(V)" + }, + { + "bbox": [ + 313, + 327, + 556, + 495 + ], + "type": "text", + "content": " into sparse spikes; 2) Spike-Driven Self-Attention (SDSA) [29, 30] and Spiking RWKV (S-RWKV) [40], which replace the dot product in SSA with the Hadamard product, achieving linear complexity while reducing both energy consumption and computational cost; 3) Spatial-Temporal Attention (STA) [12, 28], which refines traditional spatial attention by explicitly modeling the key temporal dependencies intrinsic to spike-based processing." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 496, + 556, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 556, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 556, + 699 + ], + "type": "text", + "content": "However, we observe that most Transformer-based SNNs exhibit a phenomenon referred to as attention distraction. As illustrated in Fig. 1(a), the model disproportionately assigns high attention weights to background information during decision-making, leading to the neglect of critical object-related features and ultimately impairing classification performance. This issue stems from the fact that spiking attention largely inherits the traditional attention forward propagation paradigm. Previous studies [1, 21, 34] suggest that traditional attention mechanisms process " + }, + { + "bbox": [ + 313, + 496, + 556, + 699 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 496, + 556, + 699 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 496, + 556, + 699 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 496, + 556, + 699 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 496, + 556, + 699 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 496, + 556, + 699 + ], + "type": "text", + "content": " uniformly in the mapping " + }, + { + "bbox": [ + 313, + 496, + 556, + 699 + ], + "type": "inline_equation", + "content": "softmax(QK^{\\top})V" + }, + { + "bbox": [ + 313, + 496, + 556, + 699 + ], + "type": "text", + "content": ", thereby limiting their capacity to regulate contextual sparsity and relevance. Therefore, we posit that constructing additional pathways to process " + }, + { + "bbox": [ + 313, + 496, + 556, + 699 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 496, + 556, + 699 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 496, + 556, + 699 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 496, + 556, + 699 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 496, + 556, + 699 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 496, + 556, + 699 + ], + "type": "text", + "content": " separately is essential for mitigating attention distraction in Transformer-based SNNs and enhancing model performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "text", + "content": "By contrast, the natural lateral inhibition mechanism in" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 683, + 222, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 683, + 222, + 693 + ], + "spans": [ + { + "bbox": [ + 69, + 683, + 222, + 693 + ], + "type": "text", + "content": "* Equal contribution. † Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 694, + 295, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 694, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 694, + 295, + 712 + ], + "type": "text", + "content": "These results are obtained by comparing with state-of-the-art spiking neural network models that have a similar number of parameters." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24539" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 84, + 72, + 383, + 210 + ], + "blocks": [ + { + "bbox": [ + 84, + 72, + 383, + 210 + ], + "lines": [ + { + "bbox": [ + 84, + 72, + 383, + 210 + ], + "spans": [ + { + "bbox": [ + 84, + 72, + 383, + 210 + ], + "type": "image", + "image_path": "e2b7f11365532398f9418e66da9227cbd6113317d9026495a99ec3d312a8fe34.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 394, + 89, + 534, + 211 + ], + "blocks": [ + { + "bbox": [ + 394, + 89, + 534, + 211 + ], + "lines": [ + { + "bbox": [ + 394, + 89, + 534, + 211 + ], + "spans": [ + { + "bbox": [ + 394, + 89, + 534, + 211 + ], + "type": "image", + "image_path": "c315987cd2963b7aa7a9e77a289ba5440679566604926badceccb975f8e6d604.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 62, + 215, + 547, + 449 + ], + "blocks": [ + { + "bbox": [ + 62, + 215, + 547, + 449 + ], + "lines": [ + { + "bbox": [ + 62, + 215, + 547, + 449 + ], + "spans": [ + { + "bbox": [ + 62, + 215, + 547, + 449 + ], + "type": "image", + "image_path": "0b579d9f6094e0ca852c1da4d8c5e83e470655fa8ef46df101cf1b81666bd828.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 460, + 555, + 537 + ], + "lines": [ + { + "bbox": [ + 55, + 460, + 555, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 555, + 537 + ], + "type": "text", + "content": "Figure 1. Illustration of the severe attention distraction phenomenon in Transformer-based SNNs, the architecture of the Lateral Inhibition-inspired Spiking Transformer (SpiLiFormer), and the detailed processing mechanism of Feedforward-pathway Lateral Differential Inhibition (FF-LiDiff) attention and Feedback-pathway Lateral Difference Inhibition (FB-LiDiff) attention. (a) refers to the attention distraction in mainstream Transformer-based SNNs, where models excessively focus on irrelevant background information, leading to misclassification. (b) denotes the retinal lateral inhibition mechanism, illustrating how horizontal cells regulate neural responses to enhance contrast and suppress noise. (c) presents the architecture of SpiLiFormer. (d) and (e) represent the information processing flows of FF-LiDiff and FB-LiDiff attention, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 558, + 297, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 297, + 702 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 297, + 702 + ], + "type": "text", + "content": "the visual system enables the brain to focus on important areas, thereby improving perception and reducing visual overload [4, 19, 22]. Specifically, when neurons within a given area are strongly activated, adjacent neurons are inhibited. This principle is first observed in the retina, where horizontal cells receive input from photoreceptors and inhibit both the stimulated photoreceptors and their neighbors, forming a center-surround receptive field (RF), as shown in Fig. 1(b). In higher visual areas, such as the primary (V1) and secondary (V2) visual cortices, a similar mechanism operates through long-range connections, which contrasts with the short-range inhibition in the retina, helping to refine visual" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 558, + 555, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 583 + ], + "type": "text", + "content": "processing by suppressing irrelevant stimuli and selectively enhancing salient features [41, 43]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "Based on the short-range and long-range lateral inhibition mechanisms discussed above, we propose the Lateral Inhibition-inspired Spiking Transformer (SpiLiFormer) to address the attention distraction issue in Transformer-based SNNs, as shown in Fig. 1(c). This model consists of two types of stacked modules: 1) Feedforward-pathway Lateral Differential Inhibition (FF-LiDiff) blocks and 2) Feedback-pathway Lateral Differential Inhibition (FB-LiDiff) blocks. Moreover, it introduces the following innovations: 1) In the shallow blocks (i.e., Stage 1 and Stage 2), it explicitly in-" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24540" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "type": "text", + "content": "corporates an inhibition differential attention mechanism by emulating the retinal short-range lateral inhibition process (see Fig. 1(d) for details); 2) In the deep blocks (i.e., Stage 3), it employs feedback-driven processing based on differential attention to simulate the long-range lateral inhibition mechanism observed in the cerebral cortex (see Fig. 1(e) for details). We evaluate our proposed SpiLiFormer on five image classification datasets, including CIFAR-10 [11], CIFAR-100 [11], CIFAR10-DVS [13], N-Caltech101 [20], and ImageNet-1K [5], demonstrating superior performance over current SOTA models. Additionally, adversarial testing and attention heatmap visualizations on the ImageNet-1K test set and noise-augmented datasets (i.e., CIFAR-10-C and ImageNet-1K-C [8]) show that SpiLiFormer effectively suppresses attention on irrelevant areas and background noise. The main contributions of this work are:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 279, + 294, + 518 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 55, + 279, + 294, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 294, + 361 + ], + "type": "text", + "content": "- We propose SpiLiFormer, which incorporates a brain-inspired lateral inhibition mechanism and introduces two novel attention paradigms: FF-LiDiff attention and FB-LiDiff attention. This model effectively mitigates the attention distraction issue common in current Transformer-based SNNs, thereby enhancing image classification performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 362, + 293, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 362, + 293, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 362, + 293, + 458 + ], + "type": "text", + "content": "We evaluate SpiLiFormer across multiple datasets, achieving new SOTA performance on CIFAR-10, CIFAR100, CIFAR10-DVS, N-Caltech101, and ImageNet-1K. Specifically, on ImageNet-1K, SpiLiFormer (69.9M parameters, 4 time steps) outperforms the optimal SOTA model, E-SpikeFormer (173.0M parameters, 8 time steps), by " + }, + { + "bbox": [ + 55, + 362, + 293, + 458 + ], + "type": "inline_equation", + "content": "0.46\\%" + }, + { + "bbox": [ + 55, + 362, + 293, + 458 + ], + "type": "text", + "content": " in top-1 accuracy, while using only 0.39 times the number of parameters and half the time steps." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 459, + 293, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 293, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 293, + 518 + ], + "type": "text", + "content": "- We comprehensively evaluate the robustness of SpiLiFormer through attention heatmap visualizations and adversarial testing, and find that SpiLiFormer achieves more effective attention allocation and enhanced robustness compared to other baseline spiking models." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 531, + 141, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 141, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 141, + 544 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 552, + 196, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 196, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 196, + 563 + ], + "type": "text", + "content": "2.1. Transformer-based SNNs" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "Although convolution-based SNNs exhibit high energy efficiency, they still demonstrate a significant performance gap compared to ANNs. To address this challenge, researchers have explored integrating Transformer architectures with SNNs. Spikformer [38] was the first to incorporate the Transformer attention into SNNs, cleverly replacing the Softmax operation by leveraging the binary characteristic of spike representations to encode Query, Key, and Value. Building upon this, Spikingformer [36] introduced a pre-activation shortcut to avoid floating-point multiplication while simultaneously reducing the spike firing rate. Inspired by SEW-ResNet [7], Spike-driven Trans" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": "former [30] replaced the dot product with a Hadamard product and reshaped the spike residual connection based on the membrane potential to further minimize energy consumption and leverage the event-driven property. Additionally, Spiking RWKV [40] and QKFormer [37] introduced linear attention into SNNs to reduce computational complexity. Meanwhile, STA-Transformer [12] effectively integrated both spatial and temporal dynamics of spike trains to enhance model performance. However, most existing Transformer-based SNNs still adhere to the traditional attention propagation paradigm of ANNs, which results in the severe phenomenon of attention distraction, thereby limiting model performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 237, + 481, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 237, + 481, + 249 + ], + "spans": [ + { + "bbox": [ + 313, + 237, + 481, + 249 + ], + "type": "text", + "content": "2.2. Models with Lateral Inhibition" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 255, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 495 + ], + "type": "text", + "content": "Lateral inhibition mechanisms have been widely adopted in various SNN applications. In speech recognition, Spiking-LEAF [23] emulated inner hair cell functionality and leveraged lateral inhibition feedback to enhance the neuron model, significantly improving the efficiency and noise robustness of spike encoding. In the object recognition, LISNN [2] modeled lateral inhibition by establishing spatial neighbor-order relationships between neurons, thereby enhancing the model's ability to focus on important features. In the image classification, Zhang et al. [33] proposed an adaptive self-organizing lateral inhibition strategy, where inhibition strength was adjusted according to the Euclidean distance between neurons. This approach addressed the issue of inefficient feature clustering and reduced redundancy in neuron activation. However, existing research on lateral inhibition has predominantly focused on small-scale datasets for validating model feasibility, with limited exploration of its applicability to larger and more complex datasets. Furthermore, its integration with Transformer-based SNNs remains largely unexplored." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 507, + 370, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 370, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 370, + 520 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 528, + 553, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 553, + 576 + ], + "type": "text", + "content": "This section details the proposed SpiLiFormer model, covering the spiking neuron layer, overall architecture, and two lateral inhibition-inspired attention mechanisms. Additionally, we outline the training strategy and loss function." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 585, + 441, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 585, + 441, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 585, + 441, + 597 + ], + "type": "text", + "content": "3.1. Spiking Neuron Layer" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 603, + 553, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 603, + 553, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 603, + 553, + 687 + ], + "type": "text", + "content": "The spiking neuron layer relies on cumulative firing to integrate spatio-temporal information into the membrane potential, which is then converted into binary spikes to drive computation in the next layer. We adopt the widely used Leaky Integrate-and-Fire (LIF) spiking neuron layer, as established in previous studies [30, 31, 37], with the following dynamics:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 359, + 700, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 700, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 359, + 700, + 553, + 715 + ], + "type": "interline_equation", + "content": "U [ t ] = H [ t ] (1 - S [ t ]) + U _ {\\text {r e s e t}} S [ t ], \\tag {1}", + "image_path": "b4fc4c7e29bf66be5e3e145d487cb173c0bec0ef5035c608a4e01fd6a0ec562a.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24541" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 65, + 91, + 296, + 113 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 91, + 296, + 113 + ], + "spans": [ + { + "bbox": [ + 65, + 91, + 296, + 113 + ], + "type": "interline_equation", + "content": "H [ t ] = U [ t - 1 ] + \\frac {1}{\\tau} (X [ t ] - (U [ t - 1 ] - U _ {\\text {r e s e t}})), \\quad (2)", + "image_path": "f5a4249e6f6fc484a762e4aba5e1234d26ae52cb2b3be6417c01599631493882.jpg" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 127, + 130, + 295, + 144 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 130, + 295, + 144 + ], + "spans": [ + { + "bbox": [ + 127, + 130, + 295, + 144 + ], + "type": "interline_equation", + "content": "S [ t ] = \\Theta (H [ t ] - U _ {t h}), \\tag {3}", + "image_path": "a6ab940126a79acd08bef54e8ebea04ed2b33bf9c6c1c831a2833cba5957ab18.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 148, + 296, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 148, + 296, + 244 + ], + "spans": [ + { + "bbox": [ + 55, + 148, + 296, + 244 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 148, + 296, + 244 + ], + "type": "inline_equation", + "content": "U_{\\mathrm{reset}}" + }, + { + "bbox": [ + 55, + 148, + 296, + 244 + ], + "type": "text", + "content": " represents the reset potential, and " + }, + { + "bbox": [ + 55, + 148, + 296, + 244 + ], + "type": "inline_equation", + "content": "X[t]" + }, + { + "bbox": [ + 55, + 148, + 296, + 244 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 148, + 296, + 244 + ], + "type": "inline_equation", + "content": "U[t]" + }, + { + "bbox": [ + 55, + 148, + 296, + 244 + ], + "type": "text", + "content": " denote the current input and membrane potential at timestep " + }, + { + "bbox": [ + 55, + 148, + 296, + 244 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 148, + 296, + 244 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 55, + 148, + 296, + 244 + ], + "type": "inline_equation", + "content": "\\Theta(\\cdot)" + }, + { + "bbox": [ + 55, + 148, + 296, + 244 + ], + "type": "text", + "content": " is the Heaviside step function, which outputs 1 if " + }, + { + "bbox": [ + 55, + 148, + 296, + 244 + ], + "type": "inline_equation", + "content": "x \\geq 0" + }, + { + "bbox": [ + 55, + 148, + 296, + 244 + ], + "type": "text", + "content": " and 0 otherwise. When " + }, + { + "bbox": [ + 55, + 148, + 296, + 244 + ], + "type": "inline_equation", + "content": "H[t]" + }, + { + "bbox": [ + 55, + 148, + 296, + 244 + ], + "type": "text", + "content": " exceeds the firing threshold " + }, + { + "bbox": [ + 55, + 148, + 296, + 244 + ], + "type": "inline_equation", + "content": "U_{th}" + }, + { + "bbox": [ + 55, + 148, + 296, + 244 + ], + "type": "text", + "content": ", the spiking neuron fires a spike " + }, + { + "bbox": [ + 55, + 148, + 296, + 244 + ], + "type": "inline_equation", + "content": "S[t]" + }, + { + "bbox": [ + 55, + 148, + 296, + 244 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 148, + 296, + 244 + ], + "type": "inline_equation", + "content": "U[t]" + }, + { + "bbox": [ + 55, + 148, + 296, + 244 + ], + "type": "text", + "content": " is reset to " + }, + { + "bbox": [ + 55, + 148, + 296, + 244 + ], + "type": "inline_equation", + "content": "U_{\\mathrm{reset}}" + }, + { + "bbox": [ + 55, + 148, + 296, + 244 + ], + "type": "text", + "content": " at that moment. Additionally, " + }, + { + "bbox": [ + 55, + 148, + 296, + 244 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 148, + 296, + 244 + ], + "type": "text", + "content": " represents the membrane time constant, which controls the rate of membrane potential leakage." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 252, + 176, + 263 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 176, + 263 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 176, + 263 + ], + "type": "text", + "content": "3.2. Overall Architecture" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 269, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 269, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 296, + 555 + ], + "type": "text", + "content": "The overall architecture of SpiLiFormer is illustrated in Fig. 1(c), which consists of three stages: Stage 1, Stage 2, and Stage 3, forming a three-level hierarchical structure. The input data is represented as " + }, + { + "bbox": [ + 55, + 269, + 296, + 555 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{T \\times N \\times H \\times W}" + }, + { + "bbox": [ + 55, + 269, + 296, + 555 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 269, + 296, + 555 + ], + "type": "inline_equation", + "content": "N = 3" + }, + { + "bbox": [ + 55, + 269, + 296, + 555 + ], + "type": "text", + "content": " for static image data and " + }, + { + "bbox": [ + 55, + 269, + 296, + 555 + ], + "type": "inline_equation", + "content": "N = 2" + }, + { + "bbox": [ + 55, + 269, + 296, + 555 + ], + "type": "text", + "content": " for neuromorphic data. In Stage 1, a " + }, + { + "bbox": [ + 55, + 269, + 296, + 555 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 55, + 269, + 296, + 555 + ], + "type": "text", + "content": " patch size is utilized, mapping each patch's input feature dimension " + }, + { + "bbox": [ + 55, + 269, + 296, + 555 + ], + "type": "inline_equation", + "content": "(4 \\times 4 \\times N)" + }, + { + "bbox": [ + 55, + 269, + 296, + 555 + ], + "type": "text", + "content": " to a spike-form representation of arbitrary dimension " + }, + { + "bbox": [ + 55, + 269, + 296, + 555 + ], + "type": "inline_equation", + "content": "(C)" + }, + { + "bbox": [ + 55, + 269, + 296, + 555 + ], + "type": "text", + "content": " via the Downsampling module. This process reduces the number of tokens to " + }, + { + "bbox": [ + 55, + 269, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\frac{H}{4} \\times \\frac{W}{4}" + }, + { + "bbox": [ + 55, + 269, + 296, + 555 + ], + "type": "text", + "content": ". The transformed features are then processed through FF-LiDiff attention and a Spiking MultiLayer Perceptron (SpMLP). FF-LiDiff attention introduces an additional forward pathway and employs a differential approach to mitigate attention distraction in the network's shallower layers (see Sec. 3.3 for details). In Stage 2, the processing mechanism is similar to Stage 1, but with a " + }, + { + "bbox": [ + 55, + 269, + 296, + 555 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 55, + 269, + 296, + 555 + ], + "type": "text", + "content": " patch size, an expanded channel dimension of " + }, + { + "bbox": [ + 55, + 269, + 296, + 555 + ], + "type": "inline_equation", + "content": "2C" + }, + { + "bbox": [ + 55, + 269, + 296, + 555 + ], + "type": "text", + "content": ", and the number of tokens reduced to " + }, + { + "bbox": [ + 55, + 269, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\frac{H}{8} \\times \\frac{W}{8}" + }, + { + "bbox": [ + 55, + 269, + 296, + 555 + ], + "type": "text", + "content": ". Finally, in Stage 3, we retain the " + }, + { + "bbox": [ + 55, + 269, + 296, + 555 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 55, + 269, + 296, + 555 + ], + "type": "text", + "content": " patch size, expand the channel dimension to " + }, + { + "bbox": [ + 55, + 269, + 296, + 555 + ], + "type": "inline_equation", + "content": "4C" + }, + { + "bbox": [ + 55, + 269, + 296, + 555 + ], + "type": "text", + "content": ", and reduce the number of tokens to " + }, + { + "bbox": [ + 55, + 269, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\frac{H}{16} \\times \\frac{W}{16}" + }, + { + "bbox": [ + 55, + 269, + 296, + 555 + ], + "type": "text", + "content": ". Additionally, a more complex FB-LiDiff attention module with feedback loops is introduced (see Sec. 3.4 for details) to replace the previous FF-LiDiff attention, further mitigating attention distraction in the model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 556, + 296, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 296, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 296, + 640 + ], + "type": "text", + "content": "The number of spiking blocks (either FF-LiDiff block or FB-LiDiff block) in each stage is denoted as " + }, + { + "bbox": [ + 55, + 556, + 296, + 640 + ], + "type": "inline_equation", + "content": "N_{1}" + }, + { + "bbox": [ + 55, + 556, + 296, + 640 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 556, + 296, + 640 + ], + "type": "inline_equation", + "content": "N_{2}" + }, + { + "bbox": [ + 55, + 556, + 296, + 640 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 556, + 296, + 640 + ], + "type": "inline_equation", + "content": "N_{3}" + }, + { + "bbox": [ + 55, + 556, + 296, + 640 + ], + "type": "text", + "content": ", respectively. Since the Downsampling, SpMLP and classification head modules are based on prior research works [29, 31, 37], the following sections will focus on the detailed description of FF-LiDiff attention and FB-LiDiff attention." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 647, + 295, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 295, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 295, + 672 + ], + "type": "text", + "content": "3.3. Feedforward-pathway Lateral Differential Inhibition Attention" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "An overview of FF-LiDiff attention is shown in Figure 1(d). For better mathematical description, in subsequent discussion, we assume " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "T = 1" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " and only use single head attention." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "content": "Given a spike input " + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "content": " is the token number, " + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "content": " is the channel number), Query " + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "inline_equation", + "content": "(Q)" + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "content": " and Key " + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "inline_equation", + "content": "(K)" + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "content": " can be computed through learnable matrices:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 356, + 118, + 555, + 131 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 118, + 555, + 131 + ], + "spans": [ + { + "bbox": [ + 356, + 118, + 555, + 131 + ], + "type": "interline_equation", + "content": "Q = B N \\left(X W _ {Q}\\right), K = B N \\left(X W _ {K}\\right), \\tag {4}", + "image_path": "44ef123bc3f2d8ce465705c98ec782f84229453e0e15f5ef969e1502492708ce.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 140, + 504, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 140, + 504, + 153 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 504, + 153 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 140, + 504, + 153 + ], + "type": "inline_equation", + "content": "W_{Q}, W_{K} \\in \\mathbb{R}^{D \\times D}" + }, + { + "bbox": [ + 313, + 140, + 504, + 153 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 140, + 504, + 153 + ], + "type": "inline_equation", + "content": "Q, K \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 313, + 140, + 504, + 153 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 154, + 554, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 154, + 554, + 214 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 554, + 214 + ], + "type": "text", + "content": "To implement the lateral inhibition mechanism, we first split " + }, + { + "bbox": [ + 313, + 154, + 554, + 214 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 154, + 554, + 214 + ], + "type": "text", + "content": " along the channel dimension into two parts, obtaining " + }, + { + "bbox": [ + 313, + 154, + 554, + 214 + ], + "type": "inline_equation", + "content": "Q_{1}" + }, + { + "bbox": [ + 313, + 154, + 554, + 214 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 154, + 554, + 214 + ], + "type": "inline_equation", + "content": "Q_{2}" + }, + { + "bbox": [ + 313, + 154, + 554, + 214 + ], + "type": "text", + "content": ", before computing the attention matrix. Then, all of " + }, + { + "bbox": [ + 313, + 154, + 554, + 214 + ], + "type": "inline_equation", + "content": "Q_{1}, Q_{2}" + }, + { + "bbox": [ + 313, + 154, + 554, + 214 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 154, + 554, + 214 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 154, + 554, + 214 + ], + "type": "text", + "content": " are transformed into spike trains via their respective spiking neuron layers:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 324, + 224, + 555, + 237 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 224, + 555, + 237 + ], + "spans": [ + { + "bbox": [ + 324, + 224, + 555, + 237 + ], + "type": "interline_equation", + "content": "Q _ {s 1} = \\mathcal {S N} (Q _ {1}), Q _ {s 2} = \\mathcal {S N} (Q _ {2}), K _ {s} = \\mathcal {S N} (K), \\tag {5}", + "image_path": "d491f7c648925298500c4fe66b8ed8c3ef02cc0bed02b359efc8d6c012058f7f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 247, + 554, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 554, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 554, + 380 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 247, + 554, + 380 + ], + "type": "inline_equation", + "content": "Q_{s1}, Q_{s2} \\in \\mathbb{R}^{N \\times \\frac{D}{2}}" + }, + { + "bbox": [ + 313, + 247, + 554, + 380 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 247, + 554, + 380 + ], + "type": "inline_equation", + "content": "K_{s} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 313, + 247, + 554, + 380 + ], + "type": "text", + "content": ". Next, we sum along the channel dimension to obtain the excitatory attention " + }, + { + "bbox": [ + 313, + 247, + 554, + 380 + ], + "type": "inline_equation", + "content": "A_{e}" + }, + { + "bbox": [ + 313, + 247, + 554, + 380 + ], + "type": "text", + "content": " and inhibitory attention " + }, + { + "bbox": [ + 313, + 247, + 554, + 380 + ], + "type": "inline_equation", + "content": "A_{i}" + }, + { + "bbox": [ + 313, + 247, + 554, + 380 + ], + "type": "text", + "content": ". Specifically, " + }, + { + "bbox": [ + 313, + 247, + 554, + 380 + ], + "type": "inline_equation", + "content": "A_{e}" + }, + { + "bbox": [ + 313, + 247, + 554, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 247, + 554, + 380 + ], + "type": "inline_equation", + "content": "A_{i}" + }, + { + "bbox": [ + 313, + 247, + 554, + 380 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 313, + 247, + 554, + 380 + ], + "type": "inline_equation", + "content": "N \\times 1" + }, + { + "bbox": [ + 313, + 247, + 554, + 380 + ], + "type": "text", + "content": " token attention vectors, representing the importance of different tokens from the perspectives of excitation and inhibition, respectively. Finally, we compute the element-wise difference between " + }, + { + "bbox": [ + 313, + 247, + 554, + 380 + ], + "type": "inline_equation", + "content": "A_{e}" + }, + { + "bbox": [ + 313, + 247, + 554, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 247, + 554, + 380 + ], + "type": "inline_equation", + "content": "A_{i}" + }, + { + "bbox": [ + 313, + 247, + 554, + 380 + ], + "type": "text", + "content": ", and pass the aggregated attention through a spiking neural layer to obtain binary attention " + }, + { + "bbox": [ + 313, + 247, + 554, + 380 + ], + "type": "inline_equation", + "content": "A_{\\text{combined}}" + }, + { + "bbox": [ + 313, + 247, + 554, + 380 + ], + "type": "text", + "content": ". This attention is then applied as a token-wise mask to " + }, + { + "bbox": [ + 313, + 247, + 554, + 380 + ], + "type": "inline_equation", + "content": "K_{s}" + }, + { + "bbox": [ + 313, + 247, + 554, + 380 + ], + "type": "text", + "content": " via the Hadamard product " + }, + { + "bbox": [ + 313, + 247, + 554, + 380 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 313, + 247, + 554, + 380 + ], + "type": "text", + "content": ". The above process can be described as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 370, + 389, + 554, + 427 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 389, + 554, + 427 + ], + "spans": [ + { + "bbox": [ + 370, + 389, + 554, + 427 + ], + "type": "interline_equation", + "content": "A _ {e} = \\sum_ {j} ^ {\\frac {D}{2}} Q _ {s 1} ^ {i, j}, A _ {i} = \\sum_ {j} ^ {\\frac {D}{2}} Q _ {s 2} ^ {i, j}, \\tag {6}", + "image_path": "0a69a47755326e36a24645b148e0d4096290f244b98565f3dd55ba5cf37d5aea.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 378, + 442, + 553, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 442, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 378, + 442, + 553, + 455 + ], + "type": "interline_equation", + "content": "A _ {\\text {c o m b i n e d}} = \\mathcal {S N} \\left(A _ {e} - A _ {i}\\right), \\tag {7}", + "image_path": "a8bbee384e4893826083ba4fcae1c6618676863e8707711091cc3a80445e9f3a.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 389, + 475, + 553, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 475, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 389, + 475, + 553, + 487 + ], + "type": "interline_equation", + "content": "X ^ {\\prime} = A _ {\\text {c o m b i n e d}} \\otimes K _ {s}. \\tag {8}", + "image_path": "ff37a51a5f2373b569103dcb9a0882623632dae1bc94c3d1c2ce91bcbf7d9b86.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 496, + 554, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 554, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 554, + 519 + ], + "type": "text", + "content": "3.4. Feedback-pathway Lateral Differential Inhibition Attention" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 525, + 554, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 554, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 554, + 620 + ], + "type": "text", + "content": "FB-LiDiff attention requires two forward propagations as shown in Figure 1(e). Specifically, during the first forward propagation, FB-LiDiff attention takes spike-form " + }, + { + "bbox": [ + 313, + 525, + 554, + 620 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{stage3}}^{Ds}" + }, + { + "bbox": [ + 313, + 525, + 554, + 620 + ], + "type": "text", + "content": " from the Downsampling module in Stage 3 or takes spikeform " + }, + { + "bbox": [ + 313, + 525, + 554, + 620 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{stage3}}^i (i \\in [1, N_3])" + }, + { + "bbox": [ + 313, + 525, + 554, + 620 + ], + "type": "text", + "content": " from the " + }, + { + "bbox": [ + 313, + 525, + 554, + 620 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 525, + 554, + 620 + ], + "type": "text", + "content": "-th preceding FB-LiDiff block as input. Given an input " + }, + { + "bbox": [ + 313, + 525, + 554, + 620 + ], + "type": "inline_equation", + "content": "X_1^m" + }, + { + "bbox": [ + 313, + 525, + 554, + 620 + ], + "type": "text", + "content": " to the " + }, + { + "bbox": [ + 313, + 525, + 554, + 620 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 525, + 554, + 620 + ], + "type": "text", + "content": "-th FB-LiDiff attention, the first forward propagation can be described as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 331, + 631, + 555, + 645 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 631, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 331, + 631, + 555, + 645 + ], + "type": "interline_equation", + "content": "Y _ {s} ^ {1, m} = \\mathcal {S N} \\left(B N \\left(X _ {1} ^ {m} W _ {Y} ^ {m}\\right)\\right), Y \\in \\{Q, K, V \\}, \\tag {9}", + "image_path": "a36a6adbd83b7edca8aae9d2f05c182d5546c80d23dbfc165c4563ed0ba6cb50.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 350, + 666, + 553, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 666, + 553, + 681 + ], + "spans": [ + { + "bbox": [ + 350, + 666, + 553, + 681 + ], + "type": "interline_equation", + "content": "A t t n ^ {1, m} = Q _ {s} ^ {1, m} \\odot \\left(\\left(K _ {s} ^ {1, m}\\right) ^ {\\top} \\odot V _ {s} ^ {1, m}\\right), \\tag {10}", + "image_path": "79edbf1e7d88a94a1d15386364cb635562dfd6ef023701873f133db0bb8b129f.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 357, + 700, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 700, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 357, + 700, + 553, + 715 + ], + "type": "interline_equation", + "content": "O _ {\\text {s t a g e 3}} ^ {m} = \\mathcal {S N} \\left(B N \\left(A t t n ^ {1, m} W _ {a t t n} ^ {m}\\right)\\right), \\tag {11}", + "image_path": "1c639fcaa7c4fee8e9686f83cde29ded81176f26ba9708feeb8b95804df2779b.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24542" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "inline_equation", + "content": "X_{1}^{m}\\in \\mathbb{R}^{N\\times D}" + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "inline_equation", + "content": "X_{1}^{m}\\in \\{S_{\\mathrm{stage3}}^{Ds},S_{\\mathrm{stage3}}^{i}\\}" + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "text", + "content": "), and " + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "inline_equation", + "content": "Y_{s}^{1,m}" + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "inline_equation", + "content": "Y\\in \\{Q,K,V\\}" + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "text", + "content": ") represents the corresponding spiking representation of " + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "text", + "content": "-th FB-LiDiff attention. " + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "inline_equation", + "content": "W_{z}^{m}" + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "inline_equation", + "content": "z\\in \\{Q,K,V,attn\\}" + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "inline_equation", + "content": "O_{\\mathrm{stage3}}^{m}" + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "text", + "content": " denote the learnable weight matrix and the spike-form output of the " + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "text", + "content": "-th FB-LiDiff attention, respectively. " + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "text", + "content": " refers to the dot product." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 145, + 296, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 145, + 296, + 351 + ], + "spans": [ + { + "bbox": [ + 56, + 145, + 296, + 351 + ], + "type": "text", + "content": "Before the second forward propagation, we use the output " + }, + { + "bbox": [ + 56, + 145, + 296, + 351 + ], + "type": "inline_equation", + "content": "O_{\\mathrm{stage3}}^{N_3}" + }, + { + "bbox": [ + 56, + 145, + 296, + 351 + ], + "type": "text", + "content": " from the last block of Stage 3 as feedback information and return it to each FB-LiDiff attention. To filter the feedback information and further mitigate the attention distraction problem, we first introduce learnable prompts " + }, + { + "bbox": [ + 56, + 145, + 296, + 351 + ], + "type": "inline_equation", + "content": "W_{p1}" + }, + { + "bbox": [ + 56, + 145, + 296, + 351 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 145, + 296, + 351 + ], + "type": "inline_equation", + "content": "W_{p2}" + }, + { + "bbox": [ + 56, + 145, + 296, + 351 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 56, + 145, + 296, + 351 + ], + "type": "inline_equation", + "content": "W_{p1}, W_{p2} \\in \\mathbb{R}^{1 \\times D}" + }, + { + "bbox": [ + 56, + 145, + 296, + 351 + ], + "type": "text", + "content": ". Next, we compute the dot product between " + }, + { + "bbox": [ + 56, + 145, + 296, + 351 + ], + "type": "inline_equation", + "content": "W_{p1}" + }, + { + "bbox": [ + 56, + 145, + 296, + 351 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 145, + 296, + 351 + ], + "type": "inline_equation", + "content": "O_{\\mathrm{stage3}}^{N_3}" + }, + { + "bbox": [ + 56, + 145, + 296, + 351 + ], + "type": "text", + "content": ", and similarly between " + }, + { + "bbox": [ + 56, + 145, + 296, + 351 + ], + "type": "inline_equation", + "content": "W_{p2}" + }, + { + "bbox": [ + 56, + 145, + 296, + 351 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 145, + 296, + 351 + ], + "type": "inline_equation", + "content": "O_{\\mathrm{stage3}}^{N_3}" + }, + { + "bbox": [ + 56, + 145, + 296, + 351 + ], + "type": "text", + "content": ". These dot products are then broadcast along the token dimension. The results are scaled to the range [0, 1] using the clamp(\\cdot) function, forming the excitatory feedback attention " + }, + { + "bbox": [ + 56, + 145, + 296, + 351 + ], + "type": "inline_equation", + "content": "A_e^{\\mathrm{FB}}" + }, + { + "bbox": [ + 56, + 145, + 296, + 351 + ], + "type": "text", + "content": " and inhibitory feedback attention " + }, + { + "bbox": [ + 56, + 145, + 296, + 351 + ], + "type": "inline_equation", + "content": "A_i^{\\mathrm{FB}}" + }, + { + "bbox": [ + 56, + 145, + 296, + 351 + ], + "type": "text", + "content": ". Subsequently, we compute the element-wise difference between " + }, + { + "bbox": [ + 56, + 145, + 296, + 351 + ], + "type": "inline_equation", + "content": "A_e^{\\mathrm{FB}}" + }, + { + "bbox": [ + 56, + 145, + 296, + 351 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 145, + 296, + 351 + ], + "type": "inline_equation", + "content": "A_i^{\\mathrm{FB}}" + }, + { + "bbox": [ + 56, + 145, + 296, + 351 + ], + "type": "text", + "content": " to obtain the aggregated attention " + }, + { + "bbox": [ + 56, + 145, + 296, + 351 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{combined}}^{\\mathrm{FB}}" + }, + { + "bbox": [ + 56, + 145, + 296, + 351 + ], + "type": "text", + "content": ". Finally, we apply a linear transformation to " + }, + { + "bbox": [ + 56, + 145, + 296, + 351 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{combined}}^{\\mathrm{FB}}" + }, + { + "bbox": [ + 56, + 145, + 296, + 351 + ], + "type": "text", + "content": " and pass it through the spiking neuron layer to align it with the feature space of different FB-LiDiff attentions. These processes are formulated as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 361, + 295, + 377 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 361, + 295, + 377 + ], + "spans": [ + { + "bbox": [ + 104, + 361, + 295, + 377 + ], + "type": "interline_equation", + "content": "A _ {e} ^ {\\mathrm {F B}} = \\operatorname {c l a m p} \\left(O _ {\\text {s t a g e 3}} ^ {N _ {3}} \\odot W _ {p 1}, 0, 1\\right), \\tag {12}", + "image_path": "021794263d0b1215aab0fa67cd48b2a0c59d2ce427d48238ab73d922b6ad2438.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 388, + 295, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 388, + 295, + 406 + ], + "spans": [ + { + "bbox": [ + 104, + 388, + 295, + 406 + ], + "type": "interline_equation", + "content": "A _ {i} ^ {\\mathrm {F B}} = \\operatorname {c l a m p} \\left(O _ {\\text {s t a g e 3}} ^ {N _ {3}} \\odot W _ {p 2}, 0, 1\\right), \\tag {13}", + "image_path": "45ec9f55230eba8522c4ff1e59ff3d749df482173b4c7eab3ccf429a60f14abb.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 417, + 294, + 433 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 417, + 294, + 433 + ], + "spans": [ + { + "bbox": [ + 126, + 417, + 294, + 433 + ], + "type": "interline_equation", + "content": "A _ {\\text {c o m b i n e d}} ^ {\\mathrm {F B}} = A _ {e} ^ {\\mathrm {F B}} - A _ {i} ^ {\\mathrm {F B}}, \\tag {14}", + "image_path": "499aae704ab03a3fe460fd493522106499634498aadbd320e0102b1f0ba2afca.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 445, + 294, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 445, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 106, + 445, + 294, + 460 + ], + "type": "interline_equation", + "content": "X _ {\\mathrm {F B}} ^ {m} = \\mathcal {S N} \\left(B N \\left(A _ {\\text {c o m b i n e d}} ^ {\\mathrm {F B}} W _ {\\mathrm {F B}} ^ {m}\\right)\\right). \\tag {15}", + "image_path": "f8a02f06202e5277a8d1a8f9243419ab91f8e4ca0615bdb963623cc694a6eaf8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 463, + 295, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 295, + 596 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 295, + 596 + ], + "type": "text", + "content": "The second forward propagation begins by taking the processed forward representation " + }, + { + "bbox": [ + 55, + 463, + 295, + 596 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{stage3}}^{Ds}" + }, + { + "bbox": [ + 55, + 463, + 295, + 596 + ], + "type": "text", + "content": " and the corresponding feedback information " + }, + { + "bbox": [ + 55, + 463, + 295, + 596 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{FB}}^{1}" + }, + { + "bbox": [ + 55, + 463, + 295, + 596 + ], + "type": "text", + "content": " as input to the first FB-LiDiff attention. The objective is to reduce model energy consumption by eliminating redundant computations in the non-feedback loop (i.e., Stage 1 and Stage 2). A detailed energy consumption analysis can be found in Appendix A.3. Additionally, during the second forward propagation, we share all parameter weights across the FB-LiDiff blocks to minimize the number of model parameters. Given the input " + }, + { + "bbox": [ + 55, + 463, + 295, + 596 + ], + "type": "inline_equation", + "content": "X_{2}^{m}" + }, + { + "bbox": [ + 55, + 463, + 295, + 596 + ], + "type": "text", + "content": ", the process described above can be formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 77, + 613, + 295, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 613, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 77, + 613, + 295, + 628 + ], + "type": "interline_equation", + "content": "Y _ {s} ^ {2, m} = \\mathcal {S N} \\left(B N \\left(X _ {2} ^ {m} W _ {Y} ^ {m}\\right)\\right), Y \\in \\{Q, K \\}, \\tag {16}", + "image_path": "5f318888f34525d9092c9260cfd4a8b9b0747459b2b04f78b1466e209c8b90d3.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 89, + 643, + 295, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 643, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 89, + 643, + 295, + 658 + ], + "type": "interline_equation", + "content": "V _ {s} ^ {2, m} = \\mathcal {S N} \\left(B N \\left(X _ {2} ^ {m} W _ {V} ^ {m} + X _ {\\mathrm {F B}} ^ {m} W _ {V} ^ {m}\\right)\\right), \\tag {17}", + "image_path": "614224a682e72904b2aee165c6d1c36f40bd1b8d2defc3cd8196cd1ab8fb35f8.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 92, + 671, + 295, + 687 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 671, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 92, + 671, + 295, + 687 + ], + "type": "interline_equation", + "content": "A t t n ^ {2, m} = Q _ {s} ^ {2, m} \\odot \\left(\\left(K _ {s} ^ {2, m}\\right) ^ {\\top} \\odot V _ {s} ^ {2, m}\\right), \\tag {18}", + "image_path": "4c9aed02c546b514b864c2ff8cc2b94efd37ba4721571fad2a014060b1994358.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 94, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 94, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\left(O _ {\\text {s t a g e} 3} ^ {m}\\right) ^ {\\prime} = \\mathcal {S N} \\left(B N \\left(A t t n ^ {2, m} W _ {a t t n} ^ {m}\\right)\\right). \\tag {19}", + "image_path": "93a7debd6199aeb66440e0dbeb2bfc321e12c3497c2b2407d01f7e8715ae06e8.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 72, + 509, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 509, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 509, + 84 + ], + "type": "text", + "content": "3.5. Training Strategy and Loss Function" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 89, + 553, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 553, + 172 + ], + "type": "text", + "content": "We employ a surrogate gradient-based method to overcome the non-differentiability of the spiking neuron activation in Eq. (3), enabling direct training of the model from scratch, as demonstrate in previous works [29, 36-38, 40]. In order to make better use of the spiking representations obtained from two forward propagations, we adjust the model's loss function as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 375, + 183, + 553, + 199 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 183, + 553, + 199 + ], + "spans": [ + { + "bbox": [ + 375, + 183, + 553, + 199 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {1} = \\mathcal {L} _ {C E} \\left(C H \\left(O _ {\\text {s t a g e 3}} ^ {N _ {3}}\\right), y\\right), \\tag {20}", + "image_path": "56006a3fca3308e3f6546f79177167509d3bd0ec2ace241cee59bdf545d1dcab.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 369, + 213, + 553, + 229 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 213, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 369, + 213, + 553, + 229 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {2} = \\mathcal {L} _ {C E} \\left(C H \\left(\\left(O _ {\\text {s t a g e 3}} ^ {N _ {3}}\\right) ^ {\\prime}\\right), y\\right), \\tag {21}", + "image_path": "5289dccb70fb36ee80e05e60869e1e92564f7889f20e658972017de26bb6a3a3.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 367, + 245, + 553, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 245, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 367, + 245, + 553, + 258 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {S p i L i F o r m e r}} = \\alpha \\mathcal {L} _ {1} + (1 - \\alpha) \\mathcal {L} _ {2}, \\tag {22}", + "image_path": "33fcc983bdc945df89931a68854398385b821954c5bc9edc3d4103f725721c81.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 263, + 555, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 555, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 555, + 393 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 263, + 555, + 393 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 263, + 555, + 393 + ], + "type": "text", + "content": " is a hyperparameter that dynamically balances the losses from two forward propagations. Unless explicitly specified otherwise, " + }, + { + "bbox": [ + 313, + 263, + 555, + 393 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 263, + 555, + 393 + ], + "type": "text", + "content": " is set to 0.5 in the subsequent experiments. The choice of this value is validated through an ablation study, as detailed in Appendix A.4. Additionally, " + }, + { + "bbox": [ + 313, + 263, + 555, + 393 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 263, + 555, + 393 + ], + "type": "text", + "content": " represents the true label; " + }, + { + "bbox": [ + 313, + 263, + 555, + 393 + ], + "type": "inline_equation", + "content": "O_{\\mathrm{stage3}}^{N_3}" + }, + { + "bbox": [ + 313, + 263, + 555, + 393 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 263, + 555, + 393 + ], + "type": "inline_equation", + "content": "(O_{\\mathrm{stage3}}^{N_3})'" + }, + { + "bbox": [ + 313, + 263, + 555, + 393 + ], + "type": "text", + "content": " denote the outputs from the first and second forward propagations of the last FB-LiDiff block, respectively; " + }, + { + "bbox": [ + 313, + 263, + 555, + 393 + ], + "type": "inline_equation", + "content": "CH(\\cdot)" + }, + { + "bbox": [ + 313, + 263, + 555, + 393 + ], + "type": "text", + "content": " represents the classification head module, as described in previous studies [29, 30]; and " + }, + { + "bbox": [ + 313, + 263, + 555, + 393 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}" + }, + { + "bbox": [ + 313, + 263, + 555, + 393 + ], + "type": "text", + "content": " denotes the cross-entropy loss function." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 405, + 395, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 405, + 395, + 418 + ], + "spans": [ + { + "bbox": [ + 314, + 405, + 395, + 418 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 426, + 555, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 545 + ], + "type": "text", + "content": "We first evaluate the classification performance of SpiLiFormer model on the large-scale ImageNet-1K dataset [5]. Next, we assess the performance of SpiLiFormer on smaller-scale static datasets, including CIFAR-10 [11] and CIFAR-100 [11]. Additionally, we evaluate SpiLiFormer on two popular neuromorphic datasets: CIFAR10-DVS [13] and N-Caltech101 [20]. Finally, we visualize attention heatmaps and conduct adversarial tests to demonstrate SpiLiFormer's effectiveness in mitigating attention distraction and enhancing robustness." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 552, + 451, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 451, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 451, + 565 + ], + "type": "text", + "content": "4.1. Results on ImageNet-1K" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "In this experiment, model training is divided into two phases. In the first phase, we use the AdamW optimizer with a base learning rate of " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "6 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": ". The actual learning rate is calculated by multiplying the base rate by the batch size divided by 256. Additionally, we apply a layer-wise learning rate decay strategy with a decay factor of 1.0 and weight decay of 0.05 to train SpiLiFormer for 200 epochs. In the second phase, we decrease the base learning rate to " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-6}" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " and continue training for an additional 20 epochs using an epoch-wise learning rate decay strategy. Throughout this entire experiment, we incorporate several data augmentation techniques in alignment with the DeiT approach" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24543" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 547, + 373 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 547, + 373 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 547, + 373 + ], + "type": "table", + "html": "
MethodsTypeArchitectureInput SizeParam(M)Power(mJ)Time StepTop-1 Acc(%)
ViT [6]ANNViT-B/1638486.59254.84177.90
DeiT [24]ANNDeiT-B22486.5980.50181.80
ANNDeiT-B38486.59254.84183.10
Swin [15]ANNSwin Transformer-B22487.7770.84183.50
ANNSwin Transformer-B38487.77216.20184.50
SEW ResNet [7]SNNSEW-ResNet-15222460.1912.89469.26
Spikformer [38]SNNSpikformer-8-76822466.3421.48474.81
Spikingformer [36]SNNSpikingformer-8-76822466.3413.68475.85
S-Transformer [30]SNNS-Transformer-8-51222429.681.13171.68
SNNS-Transformer-8-51222429.684.50474.57
SNNS-Transformer-8-76828866.346.09477.07
QKFormer [37]SNNHST-10-76822464.9638.91484.22
SNNHST-10-768*28864.9664.27485.25
SNNHST-10-76838464.96113.64485.65
E-SpikeFormer [31]SNNE-SpikeFormer224173.035.64*84.7
SNNE-SpikeFormer224173.054.78*85.1
SNNE-SpikeFormer384173.0-8*86.2
SpiLiFormer (Ours)SNNSpiLiFormer-10-76822469.1011.77181.54
SNNSpiLiFormer-10-76822469.1044.17485.82
SNNSpiLiFormer-10-768*28869.1073.52486.62
SNNSpiLiFormer-10-768**38469.10129.45486.66
", + "image_path": "134c7e1469e15cf241fc6e8178e29aef71d63b29abe545ebb5b41ba43ee8928c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 381, + 555, + 427 + ], + "lines": [ + { + "bbox": [ + 55, + 381, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 55, + 381, + 555, + 427 + ], + "type": "text", + "content": "Table 1. Performance comparison on ImageNet-1K. \"SpiLiFormer-L-D\" represents Lateral Inhibition-inspired Transformer with " + }, + { + "bbox": [ + 55, + 381, + 555, + 427 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 381, + 555, + 427 + ], + "type": "text", + "content": " blocks and a " + }, + { + "bbox": [ + 55, + 381, + 555, + 427 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 381, + 555, + 427 + ], + "type": "text", + "content": "-dimensional channel. \\* and \\*\\* denote input resolutions of " + }, + { + "bbox": [ + 55, + 381, + 555, + 427 + ], + "type": "inline_equation", + "content": "288^{2}" + }, + { + "bbox": [ + 55, + 381, + 555, + 427 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 381, + 555, + 427 + ], + "type": "inline_equation", + "content": "384^{2}" + }, + { + "bbox": [ + 55, + 381, + 555, + 427 + ], + "type": "text", + "content": " during inference, respectively. \\* denotes an integer-valued quantized training approach [17], in which activation values are restricted to integers during training and later expanded into spike trains during inference." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 447, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 295, + 495 + ], + "type": "text", + "content": "[24]. These techniques include RandAugment [3], random erasing [35], and stochastic depth [9]. The architecture of SpiLiFormer is composed of 1, 2, and 7 blocks across its three stages, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 497, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 295, + 581 + ], + "type": "text", + "content": "Tab. 1 demonstrates the superior performance of our proposed SpiLiFormer, significantly outperforming current SNNs models. Specifically, SpiLiFormer (69.10M) achieves " + }, + { + "bbox": [ + 55, + 497, + 295, + 581 + ], + "type": "inline_equation", + "content": "86.66\\%" + }, + { + "bbox": [ + 55, + 497, + 295, + 581 + ], + "type": "text", + "content": " top-1 accuracy and " + }, + { + "bbox": [ + 55, + 497, + 295, + 581 + ], + "type": "inline_equation", + "content": "98.116\\%" + }, + { + "bbox": [ + 55, + 497, + 295, + 581 + ], + "type": "text", + "content": " top-5 accuracy on ImageNet-1K. To our best knowledge, SpiLiFormer is currently the SOTA model for Transformer-based SNNs with direct training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "We compare SpiLiFormer against several baseline spiking models under the same image resolution of 224, a time step of 4, and a comparable number of model parameters. SpiLiFormer outperforms Spikformer (66.34M), Spikingformer (66.34M), and QKFormer (64.96M) by " + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "11.01\\%" + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "9.97\\%" + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": ", respectively. Moreover, at a higher input resolution of 288, SpiLiFormer outperforms S-Transformer and QKFormer by " + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "9.55\\%" + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "1.37\\%" + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": ", respectively. Similarly, at a resolution of 384, it surpasses QKFormer by " + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "1.01\\%" + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": ". Additionally, we compare SpiLiFormer (69.10M, 4 time steps, 384 resolution) with E-SpikeFormer (173.0M, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "type": "text", + "content": "time steps, 384 resolution), the current SOTA Transformer-based SNNs model. It achieves a " + }, + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "type": "inline_equation", + "content": "0.46\\%" + }, + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "type": "text", + "content": " improvement with only " + }, + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "type": "inline_equation", + "content": "39\\%" + }, + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "type": "text", + "content": " of parameters and half the time steps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 483, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 544 + ], + "type": "text", + "content": "Finally, compared to the mainstream ANN-based Swin Transformer-B (384 resolution), SpiLiFormer achieves a " + }, + { + "bbox": [ + 313, + 483, + 555, + 544 + ], + "type": "inline_equation", + "content": "2.16\\%" + }, + { + "bbox": [ + 313, + 483, + 555, + 544 + ], + "type": "text", + "content": " performance improvement while reducing energy consumption by " + }, + { + "bbox": [ + 313, + 483, + 555, + 544 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 483, + 555, + 544 + ], + "type": "text", + "content": " and decreasing the number of parameters by 18.67M." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 552, + 465, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 465, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 465, + 564 + ], + "type": "text", + "content": "4.2. Results on CIFAR Datasets" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 570, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 641 + ], + "type": "text", + "content": "In this experiment, we use the AdamW optimizer with a learning rate of " + }, + { + "bbox": [ + 313, + 570, + 555, + 641 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-3}" + }, + { + "bbox": [ + 313, + 570, + 555, + 641 + ], + "type": "text", + "content": " and apply a cosine learning rate decay strategy. The weight decay is " + }, + { + "bbox": [ + 313, + 570, + 555, + 641 + ], + "type": "inline_equation", + "content": "6 \\times 10^{-2}" + }, + { + "bbox": [ + 313, + 570, + 555, + 641 + ], + "type": "text", + "content": ", the number of time steps is 4, and SpiLiFormer is trained for 400 epochs with a batch size of 64. The network architecture consists of 1, 1, and 2 blocks across its three stages, respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "The performance of SpiLiFormer on CIFAR datasets is shown in the upper part of Tab. 2. For CIFAR-10, our proposed model achieves " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "96.63\\%" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": " accuracy with 7.04M parameters. Compared to QKFormer, the suboptimal SOTA model, SpiLiFormer improves performance by " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "0.45\\%" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": " while increasing the parameters by only " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "0.04\\%" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": ". Similarly, for" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24544" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 541, + 338 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 541, + 338 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 541, + 338 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 541, + 338 + ], + "type": "table", + "html": "
DatasetsMethodsArchitectureParam(M)Time StepTop-1 Acc(%)
CIFAR-10Spikformer [38]Spikformer-4-3849.32495.51
Spikingformer [36]Spikingformer-4-3849.32495.81
S-Transformer [30]S-Transformer-2-51210.28495.60
QKFormer [37]HST-4-3846.74496.18
SpiLiFormer (Ours)SpiLiFormer-4-3847.04496.63
CIFAR-100Spikformer [38]Spikformer-4-3849.32478.21
Spikingformer [36]Spikingformer-4-3849.32479.21
S-Transformer [30]S-Transformer-2-51210.28478.4
QKFormer [37]HST-4-3846.74481.15
SpiLiFormer (Ours)SpiLiFormer-4-3847.04481.63
CIFAR10-DVSSpikformer [38]Spikformer-2-2562.571680.9
Spikingformer [36]Spikingformer-2-2562.571681.3
S-Transformer [30]S-Transformer-2-2562.571680.0
QKFormer [37]HST-2-2561.501684.0
SpiLiFormer (Ours)SpiLiFormer-2-2561.571686.7
N-Caltech101Spikformer [38]Spikformer-2-2562.571683.6
Spikingformer [36]Spikingformer-2-2562.571685.91
S-Transformer [30]S-Transformer-2-2562.571686.3
QKFormer [37]HST-2-2561.501687.24
SpiLiFormer (Ours)SpiLiFormer-2-2561.571689.18
", + "image_path": "689d1814a416876ee8ad91f685b255a69158e20a2c229dfd8aba614615813036.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 111, + 345, + 499, + 357 + ], + "lines": [ + { + "bbox": [ + 111, + 345, + 499, + 357 + ], + "spans": [ + { + "bbox": [ + 111, + 345, + 499, + 357 + ], + "type": "text", + "content": "Table 2. Performance comparison on CIFAR-10, CIFAR-100, CIFAR10-DVS, and N-Caltech101 datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 377, + 295, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 295, + 426 + ], + "type": "text", + "content": "CIFAR-100, our proposed model achieves " + }, + { + "bbox": [ + 55, + 377, + 295, + 426 + ], + "type": "inline_equation", + "content": "81.63\\%" + }, + { + "bbox": [ + 55, + 377, + 295, + 426 + ], + "type": "text", + "content": " accuracy with 7.04M parameters. Specifically, it outperforms QKFormer by " + }, + { + "bbox": [ + 55, + 377, + 295, + 426 + ], + "type": "inline_equation", + "content": "0.48\\%" + }, + { + "bbox": [ + 55, + 377, + 295, + 426 + ], + "type": "text", + "content": " while requiring just " + }, + { + "bbox": [ + 55, + 377, + 295, + 426 + ], + "type": "inline_equation", + "content": "0.04\\%" + }, + { + "bbox": [ + 55, + 377, + 295, + 426 + ], + "type": "text", + "content": " more parameters." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 433, + 241, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 241, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 241, + 445 + ], + "type": "text", + "content": "4.3. Results on Neuromorphic Datasets" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 450, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 533 + ], + "type": "text", + "content": "In this experiment, we use a smaller SpiLiFormer architecture, structured with 0, 1, and 1 blocks across its three stages, respectively. Similarly, we adopt the AdamW optimizer with an initial learning rate of " + }, + { + "bbox": [ + 55, + 450, + 295, + 533 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-3}" + }, + { + "bbox": [ + 55, + 450, + 295, + 533 + ], + "type": "text", + "content": ", apply a learning rate decay strategy, and set the weight decay to 0.06. The number of time steps is set to 16, and SpiLiFormer is trained for 120 epochs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 534, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 666 + ], + "type": "text", + "content": "The experimental results of the model on the neuromorphic datasets are presented in the lower half of Tab. 2. Compared to the suboptimal model, QKFormer, SpiLi-Former significantly improves performance by " + }, + { + "bbox": [ + 55, + 534, + 295, + 666 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 55, + 534, + 295, + 666 + ], + "type": "text", + "content": " on the CIFAR10-DVS dataset and " + }, + { + "bbox": [ + 55, + 534, + 295, + 666 + ], + "type": "inline_equation", + "content": "1.94\\%" + }, + { + "bbox": [ + 55, + 534, + 295, + 666 + ], + "type": "text", + "content": " on the N-Caltech101 dataset, with only a " + }, + { + "bbox": [ + 55, + 534, + 295, + 666 + ], + "type": "inline_equation", + "content": "0.13\\%" + }, + { + "bbox": [ + 55, + 534, + 295, + 666 + ], + "type": "text", + "content": " increase in parameters. Additionally, we compare the proposed model with other baseline models (i.e., Spikformer, Spikingformer, and S-Transformer). It achieves performance gains of up to " + }, + { + "bbox": [ + 55, + 534, + 295, + 666 + ], + "type": "inline_equation", + "content": "5.8\\%" + }, + { + "bbox": [ + 55, + 534, + 295, + 666 + ], + "type": "text", + "content": " on CIFAR10-DVS and " + }, + { + "bbox": [ + 55, + 534, + 295, + 666 + ], + "type": "inline_equation", + "content": "5.55\\%" + }, + { + "bbox": [ + 55, + 534, + 295, + 666 + ], + "type": "text", + "content": " on N-Caltech101 while reducing the parameter count by 0.87M." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 672, + 291, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 291, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 291, + 685 + ], + "type": "text", + "content": "4.4. Visualization and Model Robustness Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "We visualize the attention heatmap of SpiLiFormer to assess whether it effectively alleviates the attention distraction is" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 377, + 555, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 555, + 426 + ], + "type": "text", + "content": "sue. Specifically, we examine the attention heatmap of the last FB-LiDiff block in SpiLiFormer Stage 3 under two conditions: standard dataset evaluation and noisy dataset testing." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 437, + 555, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 534 + ], + "type": "text", + "content": "In the standard dataset evaluation, the two sets of images in Fig. 1(a) show examples that are misclassified by the baseline model but correctly classified by SpiLiFormer on the ImageNet-1K test set. We observe that our proposed model effectively eliminates the attention distraction phenomenon, thereby improving classification accuracy. More comparison images are provided in Fig. 4 within the appendix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 546, + 556, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 631 + ], + "type": "text", + "content": "In noisy dataset testing, we evaluate model attention heatmaps using two noisy datasets: CIFAR10-C and ImageNet-1K-C, as shown in Fig. 2. Compared to other baseline models, SpiLiFormer correctly classifies objects or significantly increases confidence in the correct category by focusing on object-relevant information while remaining unaffected by background noise." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "In addition, we employ white-box adversarial attacks, including FGSM and PGD, to assess the model's robustness across the four datasets mentioned above. As shown in Fig. 3, we find that SpiLiFormer not only enhances model performance but also improves robustness against adversarial and common noise attacks, outperforming QKFormer." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24545" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 74, + 71, + 280, + 371 + ], + "blocks": [ + { + "bbox": [ + 74, + 71, + 280, + 371 + ], + "lines": [ + { + "bbox": [ + 74, + 71, + 280, + 371 + ], + "spans": [ + { + "bbox": [ + 74, + 71, + 280, + 371 + ], + "type": "image", + "image_path": "3ee79235ecfff21615e4b3a3527c21f7a0fd7f2da9e26794f5f51bc9404e369f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 378, + 296, + 478 + ], + "lines": [ + { + "bbox": [ + 55, + 378, + 296, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 296, + 478 + ], + "type": "text", + "content": "Figure 2. Visualization of corrupted versions of datasets. From left to right, we display the original image, followed by the attention heatmaps obtained from Spike-driven Transformer, QK-Former, and SpiLiFormer (ours), respectively. The information below each image includes the true label, the predicted label for each model, and the confidence score for the predicted category. Red indicates incorrect classification, while green represents correct classification. For more visualizations, refer to Fig. 6 and Fig. 7 in the appendix." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 501, + 149, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 149, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 149, + 514 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 519, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 295, + 590 + ], + "type": "text", + "content": "We conduct an ablation study on CIFAR-10 (static) and CIFAR10-DVS (neuromorphic) to assess the impact of FF-LiDiff and FB-LiDiff attention on model performance and inference latency. All ablation experiments follow the training details in Sec. 4.2 and Sec. 4.3, unless otherwise specified." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 592, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 295, + 676 + ], + "type": "text", + "content": "Intuitively, we set up two experimental scenarios: 1) For the case without FF-LiDiff attention (w/o FF-LiDiff attention), we use " + }, + { + "bbox": [ + 55, + 592, + 295, + 676 + ], + "type": "inline_equation", + "content": "Attn = \\mathcal{SN}(\\sum_{j}^{D}Q^{i,j})" + }, + { + "bbox": [ + 55, + 592, + 295, + 676 + ], + "type": "text", + "content": " to replace the steps in Eq. (6) and Eq. (7); 2) For the case without FB-LiDiff attention (w/o FB-LiDiff attention), we perform a single forward pass with " + }, + { + "bbox": [ + 55, + 592, + 295, + 676 + ], + "type": "inline_equation", + "content": "Attn = Q\\odot (K^{\\top}\\odot V)" + }, + { + "bbox": [ + 55, + 592, + 295, + 676 + ], + "type": "text", + "content": " to calculate the attention, replacing the steps outlined in Eq. (12) to Eq. (19)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "As shown in Tab. 3, we observe that the model exhibits varying degrees of performance degradation on both datasets under the two aforementioned scenarios. Specifi" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 318, + 70, + 434, + 148 + ], + "blocks": [ + { + "bbox": [ + 318, + 70, + 434, + 148 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 434, + 148 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 434, + 148 + ], + "type": "image", + "image_path": "e6ef6abe0370a583876cef3e5dafedede638f8c058874fa4141507144e1a9dc7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 318, + 148, + 434, + 225 + ], + "blocks": [ + { + "bbox": [ + 318, + 148, + 434, + 225 + ], + "lines": [ + { + "bbox": [ + 318, + 148, + 434, + 225 + ], + "spans": [ + { + "bbox": [ + 318, + 148, + 434, + 225 + ], + "type": "image", + "image_path": "46961d053c52fe5fc8dc5698664a551ca706b938c8f8c2c9a6ec4c28a06b33dc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 236, + 555, + 292 + ], + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 292 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 292 + ], + "type": "text", + "content": "Figure 3. Adversarial robustness comparison between QKFormer and our proposed model on CIFAR-100 (first row) and ImageNet-1K (second row). For PGD, the attack strength is set to 8/255, with a step size of 2/255 per iteration. The comparison results on other datasets are provided in Tab. 6 of within the appendix." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 435, + 71, + 551, + 148 + ], + "blocks": [ + { + "bbox": [ + 435, + 71, + 551, + 148 + ], + "lines": [ + { + "bbox": [ + 435, + 71, + 551, + 148 + ], + "spans": [ + { + "bbox": [ + 435, + 71, + 551, + 148 + ], + "type": "image", + "image_path": "df3215a47b0bcca1cd92194ebc47ebec483550843c6ac3f1ad6679b9eddbf938.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 435, + 148, + 551, + 225 + ], + "blocks": [ + { + "bbox": [ + 435, + 148, + 551, + 225 + ], + "lines": [ + { + "bbox": [ + 435, + 148, + 551, + 225 + ], + "spans": [ + { + "bbox": [ + 435, + 148, + 551, + 225 + ], + "type": "image", + "image_path": "7f732a35fdc6419c6228c65d0a6763161bb338b6fdedd6c9a15b6eb4c7703c18.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 312, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 407 + ], + "type": "text", + "content": "cally, performance exhibits a decrease of " + }, + { + "bbox": [ + 313, + 312, + 555, + 407 + ], + "type": "inline_equation", + "content": "0.31\\%" + }, + { + "bbox": [ + 313, + 312, + 555, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 312, + 555, + 407 + ], + "type": "inline_equation", + "content": "0.62\\%" + }, + { + "bbox": [ + 313, + 312, + 555, + 407 + ], + "type": "text", + "content": " on the CIFAR-10 dataset, and " + }, + { + "bbox": [ + 313, + 312, + 555, + 407 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 313, + 312, + 555, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 312, + 555, + 407 + ], + "type": "inline_equation", + "content": "3.0\\%" + }, + { + "bbox": [ + 313, + 312, + 555, + 407 + ], + "type": "text", + "content": " on the CIFAR10-DVS dataset. These results highlight the effectiveness of FB-LiDiff in mitigating attention distraction and improving model performance. However, its additional forward pass introduces higher inference latency, particularly on static datasets, with a " + }, + { + "bbox": [ + 313, + 312, + 555, + 407 + ], + "type": "inline_equation", + "content": "21 - 23\\%" + }, + { + "bbox": [ + 313, + 312, + 555, + 407 + ], + "type": "text", + "content": " increase as shown in Tab. 5 in the appendix." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 321, + 416, + 545, + 507 + ], + "blocks": [ + { + "bbox": [ + 321, + 416, + 545, + 507 + ], + "lines": [ + { + "bbox": [ + 321, + 416, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 321, + 416, + 545, + 507 + ], + "type": "table", + "html": "
DatasetsMethodsTop-1 Acc(%)
CIFAR-10w/o FF-LiDiff Attention96.32
w/o FB-LiDiff Attention96.01
SpiLiFormer(ours)96.63
CIFAR10-DVSw/o FF-LiDiff Attention85.1
w/o FB-LiDiff Attention83.7
SpiLiFormer(ours)86.7
", + "image_path": "51629e45970592104c0832ebac8b90cb05160b947f60ff971df483b9ef7e2c45.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 343, + 515, + 524, + 526 + ], + "lines": [ + { + "bbox": [ + 343, + 515, + 524, + 526 + ], + "spans": [ + { + "bbox": [ + 343, + 515, + 524, + 526 + ], + "type": "text", + "content": "Table 3. Ablation study on two attention modules" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 549, + 388, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 549, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 314, + 549, + 388, + 562 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we identify attention distraction in mainstream Transformer-based SNNs, a critical issue that limits performance. Inspired by the brain's lateral inhibition mechanism, we propose SpiLiFormer, incorporating FF-LiDiff and FB-LiDiff attention to simulate the inhibitory-excitatory interaction process, guiding the model to focus on the object rather than the irrelevant background. Experimental results show that SpiLiFormer outperforms current Transformer-based SNNs, achieving SOTA performance on five image classification datasets. Additionally, adversarial testing and attention visualization analysis demonstrate its robustness and effectiveness in alleviating attention distraction." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24546" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "type": "text", + "content": "This research was supported in part by Zhejiang Provincial Natural Science Foundation of China under the Grant No. LD25F020006." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 115, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 115, + 153 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 115, + 153 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 160, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 160, + 296, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 160, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 61, + 160, + 296, + 204 + ], + "type": "text", + "content": "[1] Hila Chefer, Idan Schwartz, and Lior Wolf. Optimizing relevance maps of vision transformers improves robustness. Advances in Neural Information Processing Systems, 35: 33618-33632, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 206, + 296, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 206, + 296, + 249 + ], + "spans": [ + { + "bbox": [ + 61, + 206, + 296, + 249 + ], + "type": "text", + "content": "[2] Xiang Cheng, Yunzhe Hao, Jiaming Xu, and Bo Xu. Lisnn: Improving spiking neural networks with lateral interactions for robust object recognition. In *IJCAI*, pages 1519–1525. Yokohama, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 251, + 296, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 296, + 306 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 296, + 306 + ], + "type": "text", + "content": "[3] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 308, + 296, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 296, + 373 + ], + "type": "text", + "content": "[4] Joseph Del Rosario, Stefano Coletta, Soon Ho Kim, Zach Mobile, Kayla Peelman, Brice Williams, Alan J Otsuki, Alejandra Del Castillo Valerio, Kendell Worden, Lou T Blanpain, et al. Lateral inhibition in v1 controls neural and perceptual contrast sensitivity. Nature Neuroscience, pages 1-12, 2025. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 376, + 296, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 376, + 296, + 419 + ], + "spans": [ + { + "bbox": [ + 62, + 376, + 296, + 419 + ], + "type": "text", + "content": "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 421, + 296, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 421, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 62, + 421, + 296, + 485 + ], + "type": "text", + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, G Heigold, S Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 488, + 296, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 488, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 488, + 296, + 531 + ], + "type": "text", + "content": "[7] Wei Fang, Zhaofei Yu, Yanqi Chen, Tiejun Huang, Timothée Masquelier, and Yonghong Tian. Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems, 34:21056-21069, 2021. 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 533, + 296, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 296, + 576 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 296, + 576 + ], + "type": "text", + "content": "[8] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 579, + 296, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 579, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 62, + 579, + 296, + 634 + ], + "type": "text", + "content": "[9] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 646-661. Springer, 2016. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 296, + 689 + ], + "type": "text", + "content": "[10] Seijoon Kim, Seongsik Park, Byunggook Na, and Sungroh Yoon. Spiking-yolo: spiking neural network for energy-efficient object detection. In Proceedings of the AAAI conference on artificial intelligence, pages 11270-11277, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 692, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 296, + 713 + ], + "type": "text", + "content": "[11] Alex Krizhevsky et al. Learning multiple layers of features from tiny images. 2009. 3, 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[12] Donghyun Lee, Yuhang Li, Youngeun Kim, Shiting Xiao, and Priyadarshini Panda. Spiking transformer with spatial-temporal attention. arXiv preprint arXiv:2409.19764, 2024. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 119, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 161 + ], + "type": "text", + "content": "[13] Hongmin Li, Hanchao Liu, Xiangyang Ji, Guoqi Li, and Luping Shi. Cifar10-dvs: an event-stream dataset for object classification. Frontiers in neuroscience, 11:309, 2017. 3,5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 164, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 553, + 206 + ], + "type": "text", + "content": "[14] Qianhui Liu, Dong Xing, Huajin Tang, De Ma, and Gang Pan. Event-based action recognition using motion information and spiking neural networks. In *IJCAI*, pages 1743–1749, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 209, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 262 + ], + "type": "text", + "content": "[15] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 266, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 266, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 317, + 266, + 553, + 330 + ], + "type": "text", + "content": "[16] Xianlei Long, Xiaxin Zhu, Fangming Guo, Chao Chen, Xiangwei Zhu, Fuqiang Gu, Songyu Yuan, and Chunlong Zhang. Spike-brgnet: Efficient and accurate event-based semantic segmentation with boundary region-guided spiking neural networks. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 333, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 333, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 317, + 333, + 553, + 387 + ], + "type": "text", + "content": "[17] Xinhao Luo, Man Yao, Yuhong Chou, Bo Xu, and Guoqi Li. Integer-valued training and spike-driven inference spiking neural network for high-performance and energy-efficient object detection. In European Conference on Computer Vision, pages 253-272. Springer, 2024. 1, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 388, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 553, + 420 + ], + "type": "text", + "content": "[18] Wolfgang Maass. Networks of spiking neurons: the third generation of neural network models. Neural networks, 10 (9):1659-1671, 1997. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 422, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 553, + 475 + ], + "type": "text", + "content": "[19] James S Magnuson, Anne Marie Crinnion, Sahil Luthra, Phoebe Gaston, and Samantha Grubb. Contra assertions, feedback improves word recognition: How feedback and lateral inhibition sharpen signals over noise. Cognition, 242: 105661, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 479, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 479, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 479, + 553, + 521 + ], + "type": "text", + "content": "[20] Garrick Orchard, Ajinkya Jayawant, Gregory K Cohen, and Nitish Thakor. Converting static image datasets to spiking neuromorphic datasets using saccades. Frontiers in neuroscience, 9:437, 2015. 3, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 524, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 524, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 524, + 553, + 578 + ], + "type": "text", + "content": "[21] Arda Sahiner, Tolga Ergen, Batu Ozturkler, John Pauly, Morteza Mardani, and Mert Pilanci. Unraveling attention via convex duality: Analysis and interpretations of vision transformers. In International Conference on Machine Learning, pages 19050-19088. PMLR, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 580, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 553, + 600 + ], + "type": "text", + "content": "[22] Stephen R Shaw. Retinal resistance barriers and electrical lateral inhibition. Nature, 255(5508):480-483, 1975. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 603, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 603, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 603, + 553, + 657 + ], + "type": "text", + "content": "[23] Zeyang Song, Jibin Wu, Malu Zhang, Mike Zheng Shou, and Haizhou Li. Spiking-leaf: A learnable auditory front-end for spiking neural networks. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 226-230. IEEE, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 659, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 659, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 659, + 553, + 713 + ], + "type": "text", + "content": "[24] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 6" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24547" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[25] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 163 + ], + "type": "text", + "content": "[26] Alex Vicente-Sola, Davide L Manna, Paul Kirkland, Gaetano Di Caterina, and Trevor J Bihl. Spiking neural networks for event-based action recognition: A new task to understand their advantage. Neurocomputing, 611:128657, 2025. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 294, + 229 + ], + "type": "text", + "content": "[27] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 568-578, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "type": "text", + "content": "[28] Yuchen Wang, Kexin Shi, Chengzhuo Lu, Yuguo Liu, Malu Zhang, and Hong Qu. Spatial-temporal self-attention for asynchronous spiking neural networks. In IJCAI, pages 3085–3093, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 341 + ], + "type": "text", + "content": "[29] Man Yao, JiaKui Hu, Tianxiang Hu, Yifan Xu, Zhaokun Zhou, Yonghong Tian, XU Bo, and Guoqi Li. Spike-driven transformer v2: Meta spiking neural network architecture inspiring the design of next-generation neuromorphic chips. In The Twelfth International Conference on Learning Representations, 2024. 1, 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "type": "text", + "content": "[30] Man Yao, Jiakui Hu, Zhaokun Zhou, Li Yuan, Yonghong Tian, Bo Xu, and Guoqi Li. Spike-driven transformer. Advances in neural information processing systems, 36, 2024. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 443 + ], + "type": "text", + "content": "[31] Man Yao, Xuerui Qiu, Tianxiang Hu, Jiakui Hu, Yuhong Chou, Keyu Tian, Jianxing Liao, Luziwei Leng, Bo Xu, and Guoqi Li. Scaling spike-driven transformer with efficient spike firing approximation training. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 3, 4, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 445, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 510 + ], + "type": "text", + "content": "[32] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch on imagenet. In Proceedings of the IEEE/CVF international conference on computer vision, pages 558-567, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 512, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 555 + ], + "type": "text", + "content": "[33] Geng Zhang, Shuangming Yang, Xuetao Zhang, and Badong Chen. Biologically plausible unsupervised learning for self-organizing spiking neural networks with dendritic computation. Neurocomputing, page 129707, 2025. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 557, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 610 + ], + "type": "text", + "content": "[34] Xuechen Zhang, Xiangyu Chang, Mingchen Li, Amit Roy-Chowdhury, Jiasi Chen, and Samet Oymak. Selective attention: Enhancing transformer through principled context control. Advances in Neural Information Processing Systems, 37:11061-11086, 2025. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 613, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 657 + ], + "type": "text", + "content": "[35] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In Proceedings of the AAAI conference on artificial intelligence, pages 13001-13008, 2020. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[36] Chenlin Zhou, Liutao Yu, Zhaokun Zhou, Zhengyu Ma, Han Zhang, Huihui Zhou, and Yonghong Tian. Spikingformer: Spike-driven residual learning for transformer-based spiking neural network. arXiv preprint arXiv:2304.11954, 2023. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 430 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 139 + ], + "type": "text", + "content": "[37] Chenlin Zhou, Han Zhang, Zhaokun Zhou, Liutao Yu, Liwei Huang, Xiaopeng Fan, Li Yuan, Zhengyu Ma, Huihui Zhou, and Yonghong Tian. Qkformer: Hierarchical spiking transformer using qk attention. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1, 3, 4, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "type": "text", + "content": "[38] Zhaokun Zhou, Yuesheng Zhu, Chao He, Yaowei Wang, YAN Shuicheng, Yonghong Tian, and Li Yuan. Spikformer: When spiking neural network meets transformer. In The Eleventh International Conference on Learning Representations, 2023. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 196, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 196, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 196, + 553, + 239 + ], + "type": "text", + "content": "[39] Zhaokun Zhou, Kaiwei Che, Wei Fang, Keyu Tian, Yuesheng Zhu, Shuicheng Yan, Yonghong Tian, and Li Yuan. Spikformer v2: Join the high accuracy club on imagenet with an snn ticket. arXiv preprint arXiv:2401.02020, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 241, + 553, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 283 + ], + "type": "text", + "content": "[40] Rui-Jie Zhu, Qihang Zhao, Guoqi Li, and Jason K Eshraghian. Spikegpt: Generative pre-trained language model with spiking neural networks. arXiv preprint arXiv:2302.13939, 2023. 1, 3, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 285, + 553, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 553, + 339 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 553, + 339 + ], + "type": "text", + "content": "[41] Shirui Zhu, Tao Xie, Ziyu Lv, Yan-Bing Leng, Yu-Qi Zhang, Runze Xu, Jingrun Qin, Ye Zhou, Vellaismy AL Roy, and Su-Ting Han. Hierarchies in visual pathway: functions and inspired artificial vision. Advanced Materials, 36(6): 2301986, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 342, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 385 + ], + "type": "text", + "content": "[42] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 386, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 386, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 386, + 553, + 430 + ], + "type": "text", + "content": "[43] Petr Znamenskiy, Mean-Hwan Kim, Dylan R Muir, M Florencia Iacaruso, Sonja B Hofer, and Thomas D Mrsic-Flogel. Functional specificity of recurrent inhibition in visual cortex. Neuron, 112(6):991-1000, 2024. 2" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24548" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SpikeDiff_ Zero-shot High-Quality Video Reconstruction from Chromatic Spike Camera and Sub-millisecond Spike Streams/3f42c09f-75a5-4dd4-bd4a-8b0a20ea82c7_content_list.json b/2025/SpikeDiff_ Zero-shot High-Quality Video Reconstruction from Chromatic Spike Camera and Sub-millisecond Spike Streams/3f42c09f-75a5-4dd4-bd4a-8b0a20ea82c7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..10e348db84b2347d733b779d0c6578bffb22a158 --- /dev/null +++ b/2025/SpikeDiff_ Zero-shot High-Quality Video Reconstruction from Chromatic Spike Camera and Sub-millisecond Spike Streams/3f42c09f-75a5-4dd4-bd4a-8b0a20ea82c7_content_list.json @@ -0,0 +1,1711 @@ +[ + { + "type": "text", + "text": "SpikeDiff: Zero-shot High-Quality Video Reconstruction from Chromatic Spike Camera and Sub-millisecondskip Spike Streams", + "text_level": 1, + "bbox": [ + 184, + 130, + 818, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Siqi Yang $^{1,2,3}$", + "bbox": [ + 151, + 200, + 261, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jinxiu Liang $^{2,3,4*}$ \nYakun Chang $^{5,6}$", + "bbox": [ + 292, + 202, + 431, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhaojun Huang2,3 Zhaofei Yu1,3", + "bbox": [ + 457, + 203, + 602, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yeliduosi Xiaokaiti2,3", + "bbox": [ + 637, + 203, + 808, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Boxin Shi $^{2,3,1*}$", + "bbox": [ + 596, + 220, + 712, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Institute for Artificial Intelligence, Peking University", + "bbox": [ + 336, + 239, + 661, + 256 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{2}$ State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University", + "$^{3}$ National Engineering Research Center of Visual Technology, School of Computer Science, Peking University" + ], + "bbox": [ + 169, + 257, + 826, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ National Institute of Informatics", + "bbox": [ + 199, + 292, + 403, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{5}$ Institute of Information Science, Beijing Jiaotong University", + "bbox": [ + 428, + 292, + 795, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{6}$ Visual Intelligence +X International Cooperation Joint Laboratory of the Ministry of Education,.cssherryliang@gmail.com, shiboxin@pku.edu.cn", + "bbox": [ + 207, + 310, + 787, + 344 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bd5f75b0432674a69a21f655aa0ecc07b693c85e4cdff72f3ae80e2030a6484c.jpg", + "image_caption": [ + "Figure 1. Spike cameras capture chromatic spikes with high temporal resolution, enabling the observation of scenes with rapid motion. As illustrated in the orange region, the spike camera utilizes a color filter array to capture chromatic information and continuously accumulates photons until reaching a threshold that triggers a spike. To reconstruct high-quality videos from these chromatic spikes, we propose SpikeDiff, a zero-shot framework that leverages principled priors from pretrained diffusion models and integrates physics-based guidance derived from the spike camera mechanism. SpikeDiff generates visually appealing videos without noise or motion blur, significantly outperforming existing chromatic spike reconstruction methods (e.g., SJDD [11] and CSpkNet [12])." + ], + "image_footnote": [], + "bbox": [ + 94, + 349, + 901, + 541 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 641, + 326, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "High-speed video reconstruction from neuromorphic spike cameras offers a promising alternative to traditional frame-based imaging, providing superior temporal resolution and dynamic range with reduced power consumption. Nevertheless, reconstructing high-quality colored videos from spikes captured in ultra-short time intervals (sub-millisecond) remain challenging due to the inherently noisy nature of spikes. While some existing methods extend the temporal capture window to improve reconstruction quality, they inevitably compromise the temporal resolution advantages of spike cameras. In this paper, we introduce SpikeDiff, the first zero-shot framework that leverages pretrained diffusion models to reconstruct high-quality colored videos from sub-millisecond", + "bbox": [ + 89, + 674, + 485, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(0.5ms) chromatic spike streams. By incorporating physics-based guidance into the diffusion sampling process, SpikeDiff bridges the domain gap between chromatic spikes and conventional images, enabling high-fidelity reconstruction without requiring domain-specific training data. Extensive experiments demonstrate that SpikeDiff achieves impressive reconstruction quality while maintaining ultra-high temporal resolution, outperforming existing methods across diverse challenging scenarios in both perceptual quality and structural preservation.", + "bbox": [ + 511, + 642, + 908, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 811, + 645, + 828 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "High-speed motions—from droplet impacts to mechanical dynamics—occur at speeds surpassing human perceptual processing capabilities. Conventional high-speed cameras are capable of capturing these motions but encounter funda-", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding authors.", + "bbox": [ + 109, + 886, + 241, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7905", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mental trade-offs [37, 45, 51]: higher temporal resolution requires shorter exposures, leading to reduced light capture and limited dynamic range, alongside increased power consumption and larger data volumes. These constraints make continuous high-speed capture over extended periods either impractical or impossible.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neuromorphic cameras have emerged as a compelling alternative to conventional high-speed imaging, including event cameras [9, 25, 27] and spike cameras [3, 21, 58]. Drawing inspiration from the fovea in biological visual systems, each pixel in a spike camera continuously accumulates photons, generates spikes upon reaching the predefined threshold, and subsequently proceeds to the next accumulation period. Operating at a sampling rate of $20,000\\mathrm{Hz}$ with $1000\\times 1000$ resolution, spike cameras encode visual information with low data overhead and power consumption, compared to conventional cameras with similar sampling rate and resolution. This enables continuously blur-free capture with high-speed motion and high contrast over several minutes that conventional sensors struggle to handle.", + "bbox": [ + 89, + 183, + 482, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For human perception, the binary representations of spikes are inherently difficult to interpret and require reconstruction into video frames [54, 55, 58]. However, reconstructing high-quality video frames from spikes remains challenging, particularly within ultra-short time intervals where limited spikes make existing methods highly susceptible to noise [57]. This challenge intensifies for color reconstruction, as Color Filter Arrays (CFAs) further reduce the number of available spikes within short time intervals.", + "bbox": [ + 89, + 395, + 482, + 530 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing approaches attempt to mitigate these challenges by extending the observed time intervals to accumulate more spikes [31, 56, 57]. This strategy inadvertently introduces motion blur and ghosting artifacts that compromise the temporal resolution advantage of spike cameras. While recent deep learning-based methods have shown promise in optimizing this trade-off problem, they typically exhibit limited generalization capabilities [10, 11], require extensive training data [12], or demand longer time intervals [48]. Yet, these methods still struggle to completely eliminate the noise [11] (upper right in Figure 1) or avoid introducing motion blur in high-speed motion regions [48] (lower right in Figure 1).", + "bbox": [ + 89, + 532, + 482, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The recent success of diffusion models in image and video restoration [15, 22, 43, 49] suggests a promising direction. These methods have demonstrated remarkable capability in encoding rich priors about natural scene statistics and temporal dynamics. This raises an intriguing question: Can we leverage these powerful priors to reconstruct high-quality frames and videos from extremely noisy chromatic spikes captured within sub-millisecond $^{1}$ time intervals?", + "bbox": [ + 88, + 713, + 482, + 834 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Integrating pretrained diffusion models with spike-to", + "bbox": [ + 109, + 835, + 482, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "video reconstruction presents fundamental challenges: Spike cameras capture visual information through asynchronous voltage accumulation and threshold-based sampling—a process fundamentally different from the fixed-interval capturing mechanism of frame-based cameras. Furthermore, chromatic spikes represent raw sensor measurements without established signal processing pipelines, while pretrained diffusion models are typically trained on processed sRGB frames with optimized color and tone characteristics. These disparities create a significant domain gap that must be addressed for effective reconstruction.", + "bbox": [ + 511, + 90, + 906, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose SpikeDiff, the first zero-shot spike-to-video reconstruction framework that leverages pretrained diffusion models as principled priors for high-quality colored video reconstruction from chromatic spikes. By establishing a physically grounded likelihood model of spike measurements throughout the reverse diffusion process, SpikeDiff approximates the posterior sampling of high-quality videos from sub-milliseconds spike streams. Unlike existing approaches that require extensive domain-specific training data and risk overfitting, our method operates in a zero-shot manner, eliminating both the computational burden of training and the need for large-scale datasets.", + "bbox": [ + 511, + 257, + 908, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our key contributions include:", + "bbox": [ + 532, + 439, + 736, + 452 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- the first framework that leverages pretrained diffusion models as powerful priors for high-quality, zero-shot video reconstruction from chromatic spikes in ultra-short intervals, eliminating the need for domain-specific training data; and", + "- a physics-based and differentiable formulation of the spike generation process, which enables effective integration with pretrained diffusion models, bridging the domain gap between spikes and conventional images." + ], + "bbox": [ + 511, + 453, + 906, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments across diverse challenging scenarios demonstrate that SpikeDiff consistently outperforms prior methods that rely on extended time intervals and compromising temporal resolution, effectively breaking the traditional quality-time resolution trade-off in spike-to-video reconstruction (see our result shown in Figure 1).", + "bbox": [ + 511, + 574, + 906, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 679, + 651, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video reconstruction with monochromatic spikes. Reconstructing high-speed videos from spikes typically employs the imaging model of spike cameras [21]. By accumulating the spikes triggered within a certain time window or estimating the interval between successive spikes, textures can be reconstructed [57, 58]. However, due to the limited number of photons arriving at the pixel in an ultra-short time, such straightforward reconstruction suffers from significant noise. To address this issue, Chang et al. [4] propose enhancing the signal-to-noise ratio (SNR) by combining multi-bit spikes with binary spikes in a rolling-mixed-bit manner. However, this solution requires hardware modifications, limiting its practicality in more general scenes. Zhao et al. [54] present", + "bbox": [ + 511, + 704, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Existing state-of-the-art methods [10-12] require at least a time interval of $40/20, 000 \\mathrm{~Hz} = 2$ ms for high-quality video reconstruction, while we aim at ultra-short time interval of $10/20, 000 \\mathrm{~Hz} = 0.5$ ms.", + "bbox": [ + 89, + 862, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7906", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Spk2ImgNet, a hierarchical architecture that progressively fuses the spikes, offering an alternative approach to improve reconstruction quality. However, its generalization ability is limited since it relies on synthetic data for training. To address the issue of real-world ground truth for training, self-supervised methods [5, 6] are developed to reduce dependence on synthetic datasets. However, it is still challenging to reconstruct color videos from monochromatic spikes. From the perspective of incorporating additional color guidance, Chang et al. [3] develop a hybrid spike-RGB camera system that performs spatial alignment and frame interpolation simultaneously, enabling the recovery of 1000 FPS color video. However, this hybrid camera system necessitates synchronization and optical alignment, while the beam splitter also poses challenges for constructing a compact device.", + "bbox": [ + 89, + 90, + 483, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Demosaicking and denoising for unconventional cameras. Various methods for joint demosaicking and denoising have been proposed to obtain high-resolution and low-noise color images from frame-based RAW data, ranging from traditional model-based techniques [1, 8, 19, 23, 32-34, 36, 50] to more recent data-driven approaches [18, 24, 29, 41, 42, 46, 53]. However, all these methods cannot be applied to binary data captured by the new-generation high-speed cameras, such as event cameras [9, 25, 27], quanta image sensors [16, 17], and spike cameras [21]. In the case of event cameras, Xu et al. [47] and Lu et al. [30] present Transformer-based architectures for demosaicking missing pixel values in RAW domain processing. However, due to the sparsity of events, these approaches still rely on frame-based cameras for texture reconstruction. In the context of quanta image sensors, Elgendy et al. [14] integrate a frequency selection method with a deep neural network based filtering approach, leveraging the luma channels to assist in the denoising of the chroma channels. Ma et al. [31] design a blue-noise pseudo-random RGBW color filter array and successfully reconstruct high-quality color images from mosaicked single-bit frames, even in high-dynamic-range (HDR) scenes with complex and rapid motion. For spike cameras, Dong et al. [11] propose an offset-sharing deformable convolution module to align temporal features of color channels and develop a spike noise estimator to characterize the noise distribution. However, this method is trained on synthetic data, limiting its generalization to real-world data. Yang et al. [48] introduce a self-supervised denoising module trained exclusively on real-world chromatic spikes, achieving 2000 FPS color video reconstruction. However, there remains significant research potential in balancing noise suppression with detail preservation.", + "bbox": [ + 89, + 316, + 486, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 829, + 181, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present the degradation model of chromatic spikes based on the physical formulation and propose SpikeDiff, a zero-shot framework for reconstructing consecutive high-quality", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "frames from chromatic spikes. This approach leverages pretrained diffusion models as principled priors to effectively address the inherent ill-posedness of the problem.", + "bbox": [ + 511, + 90, + 905, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 145, + 653, + 159 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Chromatic spikes. The chromatic spike camera leverages a spike sensor equipped with a Bayer-pattern (RGGB) CFA to capture the scene with chromatic information. As shown in Figure 1, when photons filtered by the CFA reach a pixel, the electrons generated by these photons are continuously accumulated as long as the voltage does not reach the threshold $E_{\\mathrm{th}}$ . Simultaneously, the readout circuit samples the pixel values at a frequency of $20,000\\mathrm{Hz}$ and a signal of 0 is read out at each readout point. Once the accumulated voltage reaches $E_{\\mathrm{th}}$ , a signal of 1 is read out and the voltage of this pixel is reset to continue the next accumulation period. For each pixel $i$ , we denote the accumulated voltage at a readout point $\\tau$ as $E(i,\\tau)$ , the triggered spike at $\\tau$ is:", + "bbox": [ + 511, + 167, + 906, + 364 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nS (i, \\tau) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} E (i, \\tau) \\geq E _ {\\mathrm {t h}}, \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 372, + 906, + 414 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models. Latent diffusion models (LDMs) operate in a compressed latent space learned by an autoencoder pair $(\\mathcal{E},\\mathcal{D})$ , where $X = \\mathcal{D}(Z)$ and $Z = \\mathcal{E}(X)$ . An LDM learns to reverse a diffusion process that gradually adds Gaussian noise to a latent vector $Z_{0}$ . The reverse process, modeled by a stochastic differential equation (SDE), generates a clean latent $Z_{0}$ from pure noise $Z_{T} \\sim \\mathcal{N}(0,I)$ [40]:", + "bbox": [ + 511, + 428, + 908, + 537 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nd \\boldsymbol {Z} = [ - f (\\boldsymbol {Z}, t) - g ^ {2} (t) \\nabla_ {\\boldsymbol {Z} _ {t}} \\log p _ {t} (\\boldsymbol {Z} _ {t}) ] d t + g (t) d w, (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 559, + 906, + 578 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $f(\\cdot, t)$ and $g(t)$ are drift and diffusion coefficients, $w$ represents the standard Brownian motion. The key component is the score function $\\nabla_{\\mathbf{Z}_t} \\log p_t(\\mathbf{Z}_t)$ , which is approximated by a time-conditioned neural network $\\epsilon_{\\theta}(\\mathbf{Z}, t)$ . By iteratively applying the learned score, the reverse SDE can sample from the learned data prior $p(\\mathbf{Z}_0)$ to generate a high-quality image $\\mathbf{X}_0 = \\mathcal{D}(\\mathbf{Z}_0)$ .", + "bbox": [ + 511, + 585, + 908, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Diffusion-based reconstruction from spikes", + "text_level": 1, + "bbox": [ + 511, + 700, + 879, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reconstructing consecutive high-quality frames from real-captured chromatic spike streams is challenging due to the non-negligible noise perturbations, particularly within extremely limited time intervals (e.g., sub-milisecond). To address this challenge, we leverage recent advancements in diffusion models, which incorporate learned priors regarding the distribution of high-quality images, and formulate the reconstruction of video frames from chromatic spikes as a Maximum a Posteriori (MAP) estimation problem:", + "bbox": [ + 511, + 722, + 908, + 859 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {X} ^ {*} = \\operatorname {a r g m a x} _ {\\boldsymbol {X}} p (\\boldsymbol {X} | \\boldsymbol {Y}) \\tag {3} \\\\ = \\operatorname {a r g m a x} _ {\\boldsymbol {X}} p (\\boldsymbol {Y} | \\boldsymbol {X}) p (\\boldsymbol {X}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 867, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7907", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6c3b2d7d61d13e1be9619b110693938f4e7b9a9f2d5aa5f17f87143f192733d2.jpg", + "image_caption": [ + "Figure 2. Pipeline of SpikeDiff. Our diffusion-based posterior sampling procedure follows the iterative denoising process of diffusion models, starting from the initialization $\\mathbf{Z}_{t_s}$ and gradually approaching the desired latent $\\mathbf{Z}_0$ , indicated by the dashed lines. To maximize the likelihood of the reconstructed video given the spike observations $\\mathbf{Y}$ , we enhance each iteration with physics-based spike guidance, which comprises three specially-designed differentiable degradation operators: mosaicking $M$ , color casting $C$ , and quantization $Q$ . Finally, a multiscale loss function bridges the gap between degraded frames $\\mathcal{A}(\\mathbf{X}_{0|t})$ and observed SFR frames $\\mathbf{Y}$ . As shown in the top sequence, the visual quality of generated frames progressively improves until convergence. Note that the internal degraded frames and the SFR frames are all Bayer-patterned mosaic images; we present their corresponding demosaicked results for better visualization." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 898, + 395 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the prior $p(\\mathbf{X})$ encodes knowledge of natural images, and the likelihood $p(\\mathbf{Y}|\\mathbf{X})$ represents the data fidelity to ensure the reconstruction is faithful to the observed chromatic spikes $\\mathbf{Y}$ .", + "bbox": [ + 88, + 503, + 485, + 565 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The degradation model, mapping light intensity to spike signals (illustrated in Figure 1), establishes the connection from RGB frames $X$ — the modality employed in pretrained diffusion models—and chromatic spikes $Y$ , and thereby deriving likelihood $p(\\mathbf{Y}|\\mathbf{X})$ illustrated in the next (Sec. 3.3). This establishes a principled bridge that enables the reconstruction of $X$ from $Y$ via posterior sampling in the MAP framework, as illustrated in the green box in Figure 2, where diffusion models serve as learned priors for $X$ while the likelihood ensures fidelity to $Y$ . This is exactly the key in our method to connect the modality used in pretrained diffusion models, RGB frames, and chromatic spikes.", + "bbox": [ + 88, + 580, + 485, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the significant disparity between chromatic spike streams and conventional images, the data fidelity term $p(\\mathbf{Y}|\\mathbf{X})$ cannot be directly derived from time-continuous chromatic spikes. Instead, we resort to a more compact representation. According to the integrating and firing mechanism of spikes, the light intensity is proportional to the spike firing rate (SFR) [57]. Without considering noise, the SFR at $\\tau'$ can be counted through collecting a number of spikes within", + "bbox": [ + 88, + 779, + 485, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a predefined time interval of size $\\Delta \\tau$", + "bbox": [ + 511, + 503, + 764, + 520 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nY (i, \\tau^ {\\prime}) = \\frac {1}{\\Delta \\tau} \\cdot \\sum_ {\\tau = \\tau^ {\\prime} - \\Delta \\tau / 2} ^ {\\tau^ {\\prime} + \\Delta \\tau / 2} S (i, \\tau). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 527, + 906, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By collecting the SFRs from the pixel array at $\\tau'$ , we obtain mosaicked SFR frames $Y \\in \\mathbb{R}^{T \\times H \\times W}$ , where $T$ denotes the number of frames, and $H, W$ represent the height and width, respectively. The SFR frames, estimated from real-captured chromatic spikes, are sensitive to the inherent noise, especially within limited time intervals, and exhibit misalignment with corresponding consecutive high-quality frames. Consequently, we propose a corresponding degradation model, comprising differentiable operators, as detailed in Sec. 3.3, to bridge the gap between $X$ and $Y$ and resolve the likelihood estimation required for posterior sampling, as illustrated in the orange box in Figure 2.", + "bbox": [ + 511, + 580, + 908, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Posterior sampling. The primary challenge in reconstructing consecutive high-quality frames from sub-millisecond chromatic spikes lies in its inherent ill-posedness, making the prior term $p(\\mathbf{X})$ crucial. We leverage diffusion models as powerful priors. Given the prior distribution of natural images $p(\\mathbf{Z}_t)$ encoded by the pretrained latent diffusion model in Eq. (2) and likelihood defined in the next, we solve the MAP problem in Eq. (3) by sampling from the posterior", + "bbox": [ + 511, + 779, + 910, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7908", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "distribution $p(\\mathbf{X}|\\mathbf{Y})$ . This is achieved by modifying the reverse SDE of the diffusion model to incorporate guidance from our observations. Using Bayes' rule, the score of the posterior distribution $p_t(\\mathbf{Z}_t|\\mathbf{Y})$ can be written as:", + "bbox": [ + 89, + 90, + 483, + 154 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\boldsymbol {Z} _ {t}} \\log p _ {t} (\\boldsymbol {Z} _ {t} | \\boldsymbol {Y}) = \\nabla_ {\\boldsymbol {Z} _ {t}} \\log p (\\boldsymbol {Z} _ {t}) + \\nabla_ {\\boldsymbol {Z} _ {t}} \\log p (\\boldsymbol {Y} | \\boldsymbol {Z} _ {t}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 162, + 483, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the prior term $\\nabla_{\\mathbf{Z}_t}\\log p(\\mathbf{Z}_t)$ is given by the pretrained network $\\epsilon_{\\theta^*}(Z_t,t)$ . The likelihood term $p(\\mathbf{Y}|\\mathbf{Z}_t)$ guides the sampling towards latents $\\mathbf{Z}_t$ that are likely to produce our observed SFR frame $\\mathbf{Y}$ . The full reverse SDE for posterior sampling is:", + "bbox": [ + 89, + 193, + 485, + 270 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} d \\boldsymbol {Z} = [ - f (\\boldsymbol {Z}, t) - g ^ {2} (t) (\\nabla_ {\\boldsymbol {Z} _ {t}} \\log p (\\boldsymbol {Z} _ {t}) \\tag {6} \\\\ + \\nabla_ {\\boldsymbol {Z} _ {t}} \\log p (\\boldsymbol {Y} | \\boldsymbol {Z} _ {t})) ] d t + g (t) d w. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 277, + 485, + 315 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To compute the likelihood score, we must relate the noisy latent $Z_{t}$ at an intermediate timestep $t$ to the observation $Y$ . This involves approximating the conditional probability $p(\\mathbf{Y}|\\mathbf{Z}_t)$ , which can be expressed as:", + "bbox": [ + 89, + 324, + 483, + 387 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} p (\\mathbf {Y} \\mid \\mathbf {Z} _ {t}) = \\int_ {c} p (\\mathbf {Y} \\mid \\mathbf {Z} _ {0}, \\mathbf {Z} _ {t}) p (\\mathbf {Z} _ {0} \\mid \\mathbf {Z} _ {t}) d \\mathbf {Z} _ {0} \\tag {7} \\\\ = \\int p (\\boldsymbol {Y} | \\boldsymbol {Z} _ {0}) p (\\boldsymbol {Z} _ {0} | \\boldsymbol {Z} _ {t}) d \\boldsymbol {Z} _ {0}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 395, + 483, + 459 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This integral is intractable. Instead, we approximate it by first estimating the expected clean latent $\\mathbf{Z}_{0|t} = \\mathbb{E}[Z_0|\\mathbf{Z}_t]$ from $\\mathbf{Z}_t$ . For diffusion models like DDPM [20], the forward process is", + "bbox": [ + 89, + 468, + 483, + 529 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {Z} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {Z} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} z, \\quad z \\sim \\mathcal {N} (0, I), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 539, + 483, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\bar{\\alpha}_t$ is the noise schedule parameter from the diffusion model (e.g., DDPM [20]). Using Tweedie's formula [7], we can estimate the conditional mean:", + "bbox": [ + 89, + 566, + 483, + 612 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {Z} _ {0 | t} = \\mathbb {E} [ \\boldsymbol {Z} _ {0} | \\boldsymbol {Z} _ {t} ] \\\\ = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(\\boldsymbol {Z} _ {t} + \\left(1 - \\bar {\\alpha} _ {t}\\right) \\nabla_ {\\boldsymbol {Z} _ {t}} \\log p _ {t} \\left(\\boldsymbol {Z} _ {t}\\right)\\right). \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 619, + 483, + 670 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, we replace the true score $\\nabla_{Z_t}\\log p_t(Z_t)$ with its neural network approximation $\\epsilon_{\\theta^*}(Z_t,t)$ . With this estimate of the clean latent, $Z_{0|t}$ , following the expression $p(\\mathbf{Y}|\\mathbf{Z}_t) = \\mathbb{E}_{\\mathbf{Z}_0\\sim p(\\mathbf{Z}_0|\\mathbf{Z}_t)}[p(\\mathbf{Y}|\\mathcal{D}(\\mathbf{Z}_0))]$ , we can use a surrogate likelihood approach [7] to make the likelihood tractable:", + "bbox": [ + 89, + 680, + 483, + 771 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np (\\boldsymbol {Y} | \\boldsymbol {Z} _ {t}) \\approx p (\\boldsymbol {Y} | \\mathcal {D} (\\mathbb {E} [ \\boldsymbol {Z} _ {0} | \\boldsymbol {Z} _ {t} ])). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 781, + 483, + 799 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This formulation provides a practical way to compute the likelihood gradient.", + "bbox": [ + 89, + 809, + 483, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Chromatic spikes' likelihood estimation", + "text_level": 1, + "bbox": [ + 89, + 848, + 431, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a set of sequential SFR frames, denoted by $\\mathbf{Y}$ , our objective is to establish a connection between the SFR frames", + "bbox": [ + 89, + 869, + 485, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and the desired target reconstructions $\\mathbf{X}$ , with both $\\mathbf{Y}$ and $\\mathbf{X}$ normalized to the range [0, 1]. Hence, we resolve the likelihood estimation of $p(\\mathbf{Y} | Z_t)$ in Eq. (10) and thereby complete the diffusion-based posterior sampling strategy.", + "bbox": [ + 511, + 90, + 906, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Degradation model of chromatic spikes. The physical process generating SFR frames $\\mathbf{Y}$ can be modeled as a nonlinear inverse problem: $\\mathbf{Y} = \\mathcal{A}(\\mathbf{X}) + \\mathbf{N}$ , where $\\mathcal{A}(\\cdot)$ denotes the known forward operator modeling the degradation process, and $\\mathbf{N}$ represents stochastic noise. This problem is inherently ill-posed due to irreversible information loss during the imaging process and interference from noise in the capture system.", + "bbox": [ + 511, + 170, + 906, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The forward operator $\\mathcal{A}(\\cdot)$ in the imaging pipeline transforms the latent high-quality frame $X$ through a sequence of operations: $\\mathcal{A}(X) = \\mathcal{Q}(CMX)$ , where $\\mathcal{Q}(\\cdot)$ represents pixel-wise quantization, $C$ implements pixel-wise color casting with per-channel scaling, and $M: \\mathbb{R}^{T \\times H \\times W \\times 3} \\to \\mathbb{R}^{T \\times H \\times W}$ denotes the mosaicking operator, modeling the color filter array pattern. This decomposition explicitly models each step of the image formation process, providing a foundation for our reconstruction approach.", + "bbox": [ + 511, + 291, + 906, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Under the assumption of Gaussian noise, we model the likelihood $p(\\mathbf{Y}|\\mathcal{D}(\\mathbf{Z}_{0|t}))$ of observing SFR frames $\\mathbf{Y}$ given the target frame $\\mathbf{X}$ as a Gaussian distribution. Maximizing this likelihood translates to minimizing the following loss:", + "bbox": [ + 511, + 426, + 906, + 488 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} \\left(\\boldsymbol {Z} _ {t}\\right) = \\frac {1}{\\sigma^ {2}} \\| \\boldsymbol {Y} - \\mathcal {A} \\left(\\mathcal {D} \\left(\\boldsymbol {Z} _ {0 | t}\\right)\\right) \\| _ {2} ^ {2}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 498, + 906, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma$ represents the standard deviation of the noise $N$ , and $\\mathcal{A}$ represents the differentiable degradation operators defined later. By substituting the score function into Eq. (5), we obtain:", + "bbox": [ + 511, + 537, + 906, + 595 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\boldsymbol {Z} _ {t}} \\log p _ {t} (\\boldsymbol {Z} _ {t} | \\boldsymbol {Y}) = \\epsilon_ {\\theta^ {*}} (\\boldsymbol {Z} _ {t}, t) - s \\nabla_ {\\boldsymbol {Z} _ {t}} \\mathcal {L} (\\boldsymbol {Z} _ {t}), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 609, + 906, + 627 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $s$ controls the strength of measurement guidance.", + "bbox": [ + 511, + 638, + 880, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Differentiable degradation operators. A key innovation in SpikeDiff is our design of differentiable operators in $\\mathcal{A}$ that model the physical characteristics of chromatic spike cameras, which include color casting $C$ , quantization $\\mathcal{Q}$ , and mosaicking $M$ . Based on the gray-world prior [2] for the target frame, which indicates that each color channel should have a mean value close to one another, we estimate the color casting coefficients $\\alpha_{R}, \\alpha_{G}, \\alpha_{B}$ as $\\alpha_{c} = \\frac{1}{\\mu(\\mathbf{Y}_{c})}$ , where $c \\in \\{\\mathrm{R}, \\mathrm{G}, \\mathrm{B}\\}$ denotes the color channel, $\\mu(\\mathbf{Y}_{c})$ is the mean value of color channel $c$ among $\\mathbf{Y}$ across the temporal axis. Then the channel-wise scaling factors for the color casting operator $C$ can be computed as $\\eta / \\alpha_{c}$ , where parameter $\\eta$ is used to adjust the value range.", + "bbox": [ + 511, + 671, + 906, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SFR frames are inherently quantized and discrete as they are accumulation of binary spike frames, which differ from", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7909", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "8-bit frames $\\mathbf{X}$ . To simulate the quantization effect and maintain the backward differentiability, we design a soft quantization operator $\\mathcal{Q}$ . Given the expected quantization levels determined by SFR estimation, the soft quantization function $q(\\cdot)$ operates on each pixel individually:", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nq (x) = l (x) \\cdot (1 - \\varphi (x)) + u (x) \\cdot \\varphi (x), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 176, + 482, + 194 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\varphi (x) = \\left[ \\tanh \\left(k \\left(x - \\frac {l (x) + u (x)}{2}\\right)\\right) + 1 \\right] / 2, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 196, + 482, + 226 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $l(x)$ and $u(x)$ denote the lower and upper quantization levels of $x$ , $\\varphi(\\cdot)$ quantizes the input value to the range [0, 1] with a smooth curve, and hyperparameter $k$ controls the smoothness, as shown in Figure 3(a).", + "bbox": [ + 89, + 236, + 482, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The naive Bayer-pattern mosaicking operator $M$ converts target frame from dimension $T \\times H \\times W \\times 3$ to $T \\times \\frac{H}{2} \\times \\frac{W}{2} \\times 4$ , resulting in a partially defined backpropagation function due to its gradient discontinuity. To simulate the Bayer-pattern CFA while maintaining a dense gradient flow, we extend the operator's derivative with linear interpolation convolution, similar to demosaicking algorithms.", + "bbox": [ + 89, + 297, + 483, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "By composing these operators, we conclude a differentiable simulation from consecutive frames to SFR frames, enabling optimization in the posterior sampling process.", + "bbox": [ + 89, + 402, + 483, + 448 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multiscale enhancement. To preserve fine spatial details, we incorporate a Laplacian pyramid $\\mathcal{P}(\\boldsymbol{X})$ that guides the diffusion process across multiple scales. The Laplacian pyramid decomposes the generated consecutive frames into a multiscale representation of frequency components, facilitating the preservation of fine details critical for high-quality output. The Laplacian pyramid is calculated by subtracting the upsampled version of the downsampled frames from the original frames:", + "bbox": [ + 89, + 465, + 483, + 602 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} (\\boldsymbol {X}) = \\left\\{\\boldsymbol {X} _ {l} - \\mathbf {U} \\left(\\mathbf {D} \\left(\\boldsymbol {X} _ {l}\\right)\\right) \\right\\} _ {l = 1} ^ {L}, \\boldsymbol {X} _ {l} = \\mathbf {U} \\left(\\mathbf {D} \\left(\\boldsymbol {X} _ {l - 1}\\right)\\right) \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 611, + 482, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{D}$ and $\\mathbf{U}$ are the upsampling operator and downsampling operator, respectively, and $L$ is the number of pyramid levels. This multiscale decomposition effectively guides the diffusion process across different frequency bands, ensuring preservation of high-frequency details in reconstruction.", + "bbox": [ + 89, + 643, + 483, + 719 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 733, + 220, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation details", + "text_level": 1, + "bbox": [ + 89, + 757, + 303, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantization levels. As introduced in Section 3.2, SFR frames are accumulations over predefined time intervals $\\Delta \\tau$ . Given that the spikes are binary, the value range of SFR frames is restricted to $\\left\\{\\frac{u}{\\Delta\\tau}|u\\in \\mathbb{N}\\cap [0,\\Delta \\tau ]\\right\\}$ . For instance, we set $\\Delta \\tau = 10$ in our experiments to achieve sub-milliseconds reconstruction, and the corresponding quantization levels are $0,0.1,0.2,\\ldots ,0.9,1$ , leading to the soft quantization function visualized in Figure 3(a).", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/294efd5fab5154c5cd2f7b37ed10e6f63bedcb3f4471f9f25df13b3f3d8c1a99.jpg", + "image_caption": [ + "(a) Visualization of $q(x)$" + ], + "image_footnote": [], + "bbox": [ + 519, + 89, + 638, + 152 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8a13751b2479c2ec930beb8a8abf065e0e12c6360cbc8ce480ffd837c86f4001.jpg", + "image_caption": [ + "Figure 3. Visualization of the quantization operator. (a) The soft quantization function $q(x)$ for SFR frames with $\\Delta \\tau = 10$ . (b) Detailed demonstration of quantization effectiveness on continuous pixel values." + ], + "image_footnote": [], + "bbox": [ + 648, + 89, + 730, + 152 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b7a0dbe66ff53b0a6e5a6760042b380572b8ce57fe8f7062441cc8b590c11026.jpg", + "image_caption": [ + "(b) Quantization effectiveness" + ], + "image_footnote": [], + "bbox": [ + 732, + 90, + 816, + 152 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/44762d33f751f5cbb9128a1d90d39ceef153c6e51f29f3a6e68fe6543368fd46.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 816, + 90, + 900, + 152 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Color casting estimation. To obtain better color casting coefficients and maintain consistency across video frames reconstruction, we average the estimations of $\\alpha_{c}$ from all SFR frames in the sequence, and apply the same $\\alpha_{c}$ throughout the video reconstruction.", + "bbox": [ + 511, + 234, + 905, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pretrained diffusion model. We implement SpikeDiff upon the pretrained weights from [28], which is adapted from Stable Diffusion v2.1 [38], selected for their strong priors and close alignment with our task. We extend the image-to-image pipeline to spikes-to-video by dividing SFR frames into small chunks, and achieve temporal consistency by leveraging high-fidelity reconstruction from time-continuous chromatic spikes with physics-based guidance.", + "bbox": [ + 511, + 309, + 905, + 430 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluation dataset", + "text_level": 1, + "bbox": [ + 511, + 439, + 692, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As our proposed method leverages pretrained diffusion models, which contain principled priors of the real world images and videos, we qualitatively evaluate the performance of our method on a real-world dataset. To this end, we collect a series of real chromatic spikes within the spike camera. The spike camera is equipped with a Bayer-pattern color filter array and a $1000 \\times 1000$ pixel spike sensor. Limited by the generation ability of pretrained diffusion models and resolution of compared methods [48], we spatially crop the chromatic spikes into a lower resolution of $512 \\times 512$ , maintaining the same Bayer-pattern (RGGB). The collected dataset contains chromatic spikes from various scenes with different levels of motion and color variation. We would like to contribute this dataset to the community for further research.", + "bbox": [ + 511, + 460, + 906, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Despite the real-captured chromatic spikes, we also employ the chromatic spikes simulator described in CSpKNet [12] and SY24 [48] to generate synthetic chromatic spikes from real-world high-frame-rate videos. We collect over 100 dynamic scenarios from X4K1000FPS [39] dataset, and generate the corresponding chromatic spikes with the simulator. The synthetic dataset is used to conduct the quantitative evaluation of our proposed method, which further demonstrates the superiority of SpikeDiff.", + "bbox": [ + 511, + 688, + 906, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Qualitative evaluation", + "text_level": 1, + "bbox": [ + 511, + 833, + 718, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To demonstrate the superiority of our proposed method, we conduct the qualitative evaluation on chromatic spikes captured from the real world, comparing SpikeDiff with exist", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7910", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/35a9e1d5b4299d7c0661ad2935794cf35d3ba972181a0be2c42d6fa62fcfe2ee.jpg", + "image_caption": [ + "(a)", + "Figure 4. Qualitative comparison on real-world captured chromatic spikes (a-c), and synthetic chromatic spikes simulated from real-world high-frame-rate videos (d). Compared methods include TFI [57], TFP [57], 3DRI [10], SJDD [11], CSpkNet [12], and SY24 [48]. Details in red / blue bounding boxes are shown on the top. All results are reconstructed from sub-millisecond spike streams (0.5ms), where most methods suffer from severe noise or extreme degradation while SpikeDiff recovers clean and visually pleasing results." + ], + "image_footnote": [], + "bbox": [ + 120, + 65, + 893, + 601 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing chromatic spike reconstruction methods. Our proposed method can recover high-quality color frames from sub-milisecond time intervals (10 spike frames, corresponding to $0.5\\mathrm{ms}$ ), which none of the existing neural network based method achieve. We adapt and evaluate the existing chromatic spike reconstruction methods (3DRI [10], SJDD [11], CSpkNet [12], and SY24 [48]), combining the SFR estimation methods (TFI [57], TFP [57]) with demosaicking as baselines. We conduct all experiments on sub-millisecond time intervals (0.5ms) to maintain the fairness in comparison. As shown in Figure 4(a), where a fan is fast rotating in front of the checkerboard, SpikeDiff produces the most clean and visually pleasant results. In contrast, other methods, except CSpkNet and SY24, contain significantly noisy pixels as highlighted in the red and blue bounding boxes. The results", + "bbox": [ + 88, + 664, + 486, + 891 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of CSpkNet and SY24 contain blurry or degraded artifacts. As the time intervals are strictly limited, all methods are free of motion blur naturally. As for Figure 4(b), we capture a static scene in front of the lamp with the spike camera slightly shaking, to study the reconstruction quality under hard light. Compared to others, SpikeDiff generates the most clear and sharp details in blue bounding box. And SpikeDiff produces a smoother texture under the direct illumination in the red bounding box, while most of other methods show obvious quantized artifacts. In Figure 4(c), we leverage the spike streams captured by Yang et al. [48], where SpikeDiff outperforms all other methods in both noise suppression and visual quality. We also illustrate a synthetic scene in Figure 4(d), which further demonstrates the superior reconstruction capability of SpikeDiff on chromatic spikes", + "bbox": [ + 511, + 664, + 908, + 892 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7911", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e5e1eae43d8f06579274a840268dde3cf4c47a6760bc13bac8d17e5e9ebc8d0c.jpg", + "table_caption": [ + "Table 1. Quantitative comparison of the proposed method with existing chromatic spike reconstruction methods. The best and second-best results are highlighted in red and blue, respectively." + ], + "table_footnote": [], + "table_body": "
MethodPSNR↑SSIM↑FID↓NIQE↓IL-NIQE↓
SpikeDiff (Ours)18.6940.7502.8805.17338.535
SY24 [48]13.0130.53621.47211.59278.672
SJDD [11]11.3580.32526.00911.27844.711
3DRI [10]18.5120.39016.0368.54141.816
CSpkNet [12]13.3640.6414.6995.81939.948
TFP [57]12.9090.5593.25612.00164.554
TFI [57]14.6400.5527.55812.51855.758
", + "bbox": [ + 107, + 132, + 464, + 241 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/680c40ecc00c5fddcec7a75e0aeaafdd0141ebaa70181b8658c27bf24adc41b2.jpg", + "table_caption": [ + "Table 2. Temporal consistency evaluation, reporting Frame CLIP Score (Frame C.S.), Interpolation Error (Inp. Err.), and Interpolation PSNR (Inp. PSNR) following the metrics from [26]. The best and second-best are highlighted in red and blue, respectively." + ], + "table_footnote": [], + "table_body": "
MethodFrame C.S.↑Inp. Err.↓Inp. PSNR↑
SpikeDiff (Ours)0.9690.05725.45
SY24 [48]0.9630.04128.65
SJDD [11]0.9620.11319.34
3DRI [10]0.9650.10419.87
CSpkNet [12]0.9660.05925.27
", + "bbox": [ + 129, + 306, + 444, + 393 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/01d3436302988abaf90e5cca39c257072453296c5fa7262950208be221aaf4ea.jpg", + "image_caption": [ + "Grad. w/o", + "Mosaicking", + "Figure 5. Analysis of differentiable degradation operators. Generated frames receive sparse gradient from chromatic spikes without mosaicking operator. Similarly, reconstructions exhibit incorrect color distribution without the color casting operator." + ], + "image_footnote": [], + "bbox": [ + 96, + 400, + 191, + 474 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/93bc0d0cc1c89a08dca55893b0206a1e132c70b73e9017b6b578760ae3625407.jpg", + "image_caption": [ + "Grad.w/", + "Mosaicking" + ], + "image_footnote": [], + "bbox": [ + 192, + 401, + 287, + 474 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f66e8b4203c1e4768b364dda0afe21e8ccf79e0a18de712f0ea604271680de63.jpg", + "image_caption": [ + "Result w/o", + "Color casting" + ], + "image_footnote": [], + "bbox": [ + 289, + 401, + 383, + 474 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/39507c09ee4651ee4e0474177738c2f52a4e323ed4bd46a038f3234037f0c856.jpg", + "image_caption": [ + "Result w/", + "Color casting" + ], + "image_footnote": [], + "bbox": [ + 383, + 401, + 475, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "simulated from real-world videos. Note that post-processing (e.g., color casting) cannot eliminate noise or artifacts from other methods' results. Please refer to our supplementary material for more comparison and detailed discussion.", + "bbox": [ + 89, + 561, + 483, + 622 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Quantitative evaluation", + "text_level": 1, + "bbox": [ + 89, + 633, + 307, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Table 1, we quantitatively evaluate the performance of SpikeDiff and existing chromatic spike reconstruction methods on the synthetic dataset. Similar to our qualitative evaluation, we evaluate these methods on submillisecond time intervals (0.5ms). Our method achieves the best performance across all the metrics, including PSNR, SSIM [44], FID [13], NIQE [35], and IL-NIQE [52], which demonstrates the superiority of SpikeDiff, especially in terms of perceptual metrics that reflect visual quality. As shown in Table 2, we also evaluate the temporal consistency metrics of SpikeDiff and existing reconstruction methods, where SpikeDiff achieves the best or comparable performance.", + "bbox": [ + 89, + 656, + 483, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation study", + "text_level": 1, + "bbox": [ + 89, + 848, + 240, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To investigate the effectiveness of our proposed degradation operators, we visualize the internal results after applying", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/07be01eac1443019926966f12191b63307a58fd7a6a61cafa4f908f66e5bcf5e.jpg", + "table_caption": [ + "Table 3. Quantitative ablation study on the contribution of different degradation operators to reconstruction quality." + ], + "table_footnote": [], + "table_body": "
MethodPSNR↑SSIM↑FID↓NIQE↓IL-NIQE↓
SpikeDiff18.6940.7502.8805.17338.535
w/o C16.7520.72912.0646.07244.661
w/o M17.4810.7234.3926.20341.356
w/o Q17.4600.7173.9386.31840.487
", + "bbox": [ + 547, + 117, + 867, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "each operator. The mosaicking operator $M$ converts the generated frames into Bayer-pattern images, which is the beginning step. As shown in Figure 5, without the specific backward formulation for $M$ , the gradient map is not well-defined. The color casting operator $C$ is responsible for color correction, which degrades the normal colored image to a greener one (which is more close to the SFR frames) as shown in Figure 5. For the soft quantization operator $\\mathcal{Q}$ , its effectiveness is demonstrated in Figure 3(b), quantizing the continuous white pixels into discrete levels, similar to the pattern of the SFR frames. As shown in Table 3, we further demonstrate the effectiveness of each degradation operator via quantitative evaluation.", + "bbox": [ + 511, + 199, + 906, + 396 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 410, + 633, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose the first zero-shot chromatic spike reconstruction method, named SpikeDiff, to recover consecutive clean and high-quality colored frames from sub-millisecond spikes. SpikeDiff leverages the deep priors from pretrained diffusion models to address the extreme noise under very limited time intervals and generate visually pleasant results. SpikeDiff integrates the physics-based chromatic spikes' likelihood into the diffusion-based posterior sampling process, which bridges the domain gap between spikes and conventional images. We conduct experiments on both synthetic and real-captured data to demonstrate the superior performance of SpikeDiff over existing reconstruction methods.", + "bbox": [ + 511, + 436, + 906, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and future work. As shown in Figure 4(b), there exist \"overexposure\" cases in the central region under extreme direct illumination. This imperfection is due to the predefined spike firing threshold, resulting in consistent spikes among these pixels. We observed that it is hard to generate ideal details without any explicit guidance. In future research, we plan to explore more degradation model designs and diffusion techniques to handle this issue and further extend the HDR capability of chromatic spike reconstruction.", + "bbox": [ + 511, + 618, + 908, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 513, + 768, + 671, + 786 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by National Natural Science Foundation of China (Grant No. 62302019, 62088102, 62136001), Beijing Natural Science Foundation (Grant No. L233024), and Beijing Municipal Science & Technology Commission, Administrative Commission of Zhongguancun Science Park (Grant No. Z241100003524012). PKU-affiliated authors thank openbayes.com for providing computing resources.", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7912", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Antoni Buades, Bartomeu Coll, Jean-Michel Morel, and Catalina Sbert. Self-similarity driven color demosaicking. IEEE TIP, 18(6):1192-1202, 2009. 3", + "[2] Gershon Buchsbaum. A spatial processor model for object colour perception. Journal of the Franklin institute, 310(1): 1-26, 1980. 5", + "[3] Yakun Chang, Chu Zhou, Yuchen Hong, Liwen Hu, Chao Xu, Tiejun Huang, and Boxin Shi. 1000 FPS HDR video with a Spike-RGB hybrid camera. In CVPR, pages 22180-22190, 2023. 2, 3", + "[4] Yakun Chang, Yeliduosi Xiaokaiti, Yujia Liu, Bin Fan, Zhaojun Huang, Tiejun Huang, and Boxin Shi. Towards HDR and HFR video from rolling-mixed-bit spikings. In CVPR, pages 25117-25127, 2024. 2", + "[5] Shiyan Chen, Chaoteng Duan, Zhaofei Yu, Ruiqin Xiong, and Tiejun Huang. Self-supervised mutual learning for dynamic scene reconstruction of spiking camera. In IJCAI, page 2859-2866, 2022. 3", + "[6] Shiyan Chen, Zhaofei Yu, and Tiejun Huang. Self-supervised joint dynamic scene reconstruction and optical flow estimation for spiking camera. In AAAI, pages 350-358, 2023. 3", + "[7] Hyungjin Chung, Jeongsol Kim, Michael Thompson McCann, Marc Louis Klasky, and Jong Chul Ye. Diffusion Posterior Sampling for General Noisy Inverse Problems. In ICLR, 2023. 5", + "[8] Laurent Condat and Saleh Mosaddegh. Joint demosaicking and denoising by total variation minimization. In ICIP, pages 2781-2784, 2012. 3", + "[9] Tobi Delbruck et al. Frame-free dynamic digital vision. In Proc. of Intl. Symp. on Secure-Life Electronics, Advanced Electronics for Quality Life and Society, pages 21-26. Cite-seer, 2008. 2, 3", + "[10] Yanchen Dong, Jing Zhao, Ruiqin Xiong, and Tiejun Huang. 3D residual interpolation for spike camera demosaicing. In ICIP, pages 1461-1465. IEEE, 2022. 2, 7, 8", + "[11] Yanchen Dong, Ruiqin Xiong, Jing Zhao, Jian Zhang, Xiaopeng Fan, Shuyuan Zhu, and Tiejun Huang. Joint demosaicing and denoising for spike camera. In AAAI, pages 1582-1590, 2024. 1, 2, 3, 7, 8", + "[12] Yanchen Dong, Ruiqin Xiong, Jing Zhao, Jian Zhang, Xiaopeng Fan, Shuyuan Zhu, and Tiejun Huang. Learning a deep demosaicing network for spike camera with color filter array. IEEE TIP, 2024. 1, 2, 6, 7, 8", + "[13] DC Dowson and BV Landau. The frechet distance between multivariate normal distributions. Journal of multivariate analysis, 12(3):450-455, 1982. 8", + "[14] Omar A Elgendy, Abhiram Gnanasambandam, Stanley H Chan, and Jiaju Ma. Low-light demosaicking and denoising for small pixels using learned frequency selection. IEEE TCI, 7:137-150, 2021. 3", + "[15] Berthy T. Feng, Jamie Smith, Michael Rubinstein, Huiwen Chang, Katherine L. Bouman, and William T. Freeman. Score-Based Diffusion Models as Principled Priors for Inverse Imaging. In ICCV, 2023. 2" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Eric R Fossum, Jiaju Ma, and Saleh Masoodian. Quanta image sensor: Concepts and progress. Advanced Photon Counting Techniques X, 9858:985805, 2016. 3", + "[17] Eric R Fossum, Jiaju Ma, Saleh Masoodian, Leo Anzagira, and Rachel Zizza. The quanta image sensor: Every photon counts. Sensors, 16(8):1260, 2016. 3", + "[18] Michael Gharbi, Gaurav Chaurasia, Sylvain Paris, and Frédo Durand. Deep joint demosaicking and denoising. ACM TOG, 35(6):1-12, 2016. 3", + "[19] Felix Heide, Mushfiqur Rouf, Matthias B Hullin, Bjorn Labitzke, Wolfgang Heidrich, and Andreas Kolb. High-quality computational imaging through simple lenses. ACM TOG, 32 (5):1-14, 2013. 3", + "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 5", + "[21] Tiejun Huang, Yajing Zheng, Zhaofei Yu, Rui Chen, Yuan Li, Ruiqin Xiong, Lei Ma, Junwei Zhao, Siwei Dong, Lin Zhu, et al. $1000\\times$ faster camera and machine vision with ordinary devices. Engineering, 25:110-119, 2022. 2, 3", + "[22] Siddhant Jain, Daniel Watson, Eric Tabellion, Aleksander Hołyński, Ben Poole, and Janne Kontkanen. Video Interpolation with Diffusion Models. In CVPR, 2024. 2", + "[23] Daisuke Kiku, Yusuke Monno, Masayuki Tanaka, and Masatoshi Okutomi. Beyond color difference: Residual interpolation for color image demosaicking. IEEE TIP, 25(3): 1288-1300, 2016. 3", + "[24] Filippos Kokkinos and Stamatios Lefkimmiatis. Deep image demosaicking using a cascade of convolutional residual denoising networks. In ECCV, pages 303-319, 2018. 3", + "[25] Juan Antonio Leñero-Bardallo, Teresa Serrano-Gotarredona, and Bernabé Linares-Barranco. A 3.6 $\\mu$ s latency asynchronous frame-free event-driven dynamic-vision-sensor. IEEE Journal of Solid-State Circuits, 46(6):1443-1455, 2011. 2, 3", + "[26] Xinui Li, Chao Ma, Xiaokang Yang, and Ming-Hsuan Yang. VidToMe: Video token merging for zero-shot video editing. In CVPR, pages 7486-7495, 2024. 8", + "[27] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A $128 \\times 128$ 120 dB $15\\mu s$ latency asynchronous temporal contrast vision sensor. IEEE Journal of Solid-State Circuits, 43(2):566-576, 2008. 2, 3", + "[28] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Yu Qiao, Wanli Ouyang, and Chao Dong. Diff-BIR: Toward blind image restoration with generative diffusion prior. In ECCV, pages 430-448. Springer, 2024. 6", + "[29] Lin Liu, Xu Jia, Jianzhuang Liu, and Qi Tian. Joint demosaicing and denoising with self guidance. In CVPR, pages 2240-2249, 2020. 3", + "[30] Yunfan Lu, Yijie Xu, Wenzong Ma, Weiyu Guo, and Hui Xiong. Event camera demosaicing via swim transformer and pixel-focus loss. In CVPR, pages 1095-1105, 2024. 3", + "[31] Sizhuo Ma, Varun Sundar, Paul Mos, Claudio Bruschini, Edoardo Charbon, and Mohit Gupta. Seeing photons in color. ACM TOG, 42(4):1-16, 2023. 2, 3", + "[32] Julien Mairal, Michael Elad, and Guillermo Sapiro. Sparse representation for color image restoration. IEEE TIP, 17(1): 53-69, 2007. 3" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7913", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Julien Mairal, Francis Bach, Jean Ponce, Guillermo Sapiro, and Andrew Zisserman. Non-local sparse models for image restoration. In ICCV, pages 2272-2279, 2009.", + "[34] Henrique S Malvar, Li-wei He, and Ross Cutler. High-quality linear interpolation for demosaicing of bayer-patterned color images. In ICASSP, pages iii-485, 2004. 3", + "[35] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal Processing Letters, 20(3):209-212, 2012. 8", + "[36] Ibrahim Pekkucuksen and Yucel Altunbasak. Gradient based threshold free color filter array interpolation. In ICIP, pages 137-140, 2010. 3", + "[37] Reza Pournaghi and Xiaolin Wu. Coded acquisition of high frame rate video. IEEE TIP, 23(12):5670-5682, 2014. 2", + "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 6", + "[39] Hyeonjun Sim, Jihyong Oh, and Munchurl Kim. XVFI: extreme video frame interpolation. In ICCV, 2021. 6", + "[40] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 3", + "[41] Daniel Stanley Tan, Wei-Yang Chen, and Kai-Lung Hua. DeepDemosaicking: Adaptive image demosaicking via multiple deep fully convolutional networks. IEEE TIP, 27(5): 2408-2419, 2018. 3", + "[42] Runjie Tan, Kai Zhang, Wangmeng Zuo, and Lei Zhang. Color image demosaicking via deep residual learning. In ICME, pages 793-798, 2017. 3", + "[43] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-Shot Image Restoration Using Denoising Diffusion Null-Space Model. In ICLR, 2022. 2", + "[44] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004. 8", + "[45] Bennett Wilburn, Neel Joshi, Vaibhav Vaish, Marc Levoy, and Mark Horowitz. High-speed videography using a dense camera array. CVPR, 2004. 2", + "[46] Wenzhu Xing and Karen Egiazarian. End-to-end learning for joint image demosaicing, denoising and super-resolution. In CVPR, pages 3507-3516, 2021. 3", + "[47] Senyan Xu, Zhijing Sun, Jiaying Zhu, Yurui Zhu, Xueyang Fu, and Zheng-Jun Zha. Demosaicformer: Coarse-to-fine demosaicing network for hybrids camera. In CVPR, pages 1126-1135, 2024. 3", + "[48] Siqi Yang, Zhaojun Huang, Yakun Chang, Bin Fan, Zhaofei Yu, and Boxin Shi. Real-data-driven 2000 FPS color video from mosaicked chromatic spikes. In ECCV, 2024. 2, 3, 6, 7, 8", + "[49] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling Up to Excellence: Practicing Model Scaling for Photo-Realistic Image Restoration In the Wild. In CVPR, 2024. 2" + ], + "bbox": [ + 91, + 92, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Guoshen Yu, Guillermo Sapiro, and Stephane Mallat. Solving inverse problems with piecewise linear estimators: From gaussian mixture models to structured sparsity. IEEE TIP, 21 (5):2481-2499, 2011. 3", + "[51] Jie Zhang, Jonathan Newman, Zeguan Wang, Yong Qian, Pedro Feliciano-Ramos, Wei Guo, Takato Honda, Zhe Sage Chen, Changyang Linghu, Ralph Etienne-Cummings, et al. Pixel-wise programmability enables dynamic high-snr cameras for high-speed microscopy. Nature Communications, 15 (1):4480, 2024. 2", + "[52] Lin Zhang, Lei Zhang, and Alan C Bovik. A feature-enriched completely blind image quality evaluator. IEEE TIP, 24(8): 2579-2591, 2015. 8", + "[53] Zhao Zhang, Huan Zheng, Richang Hong, Mingliang Xu, Shuicheng Yan, and Meng Wang. Deep color consistent network for low-light image enhancement. In CVPR, pages 1899-1908, 2022. 3", + "[54] Jing Zhao, Ruiqin Xiong, Hangfan Liu, Jian Zhang, and Tiejun Huang. Spk2Imgnet: Learning to reconstruct dynamic scene from continuous spike stream. In CVPR, pages 11996-12005, 2021. 2", + "[55] Jing Zhao, Ruiqin Xiong, Jiyu Xie, Boxin Shi, Zhaofei Yu, Wen Gao, and Tiejun Huang. Reconstructing clear image for high-speed motion scene with a retina-inspired spike camera. IEEE TCI, 8:12-27, 2021. 2", + "[56] Yajing Zheng, Lingxiao Zheng, Zhaofei Yu, Boxin Shi, Yonghong Tian, and Tiejun Huang. High-speed image reconstruction through short-term plasticity for spiking cameras. In CVPR, pages 6358-6367, 2021. 2", + "[57] Lin Zhu, Siwei Dong, Tiejun Huang, and Yonghong Tian. A retina-inspired sampling method for visual texture reconstruction. In ICME, pages 1432-1437, 2019. 2, 4, 7, 8", + "[58] Lin Zhu, Siwei Dong, Jianing Li, Tiejun Huang, and Yonghong Tian. Retina-like visual image reconstruction via spiking neural model. In CVPR, pages 1438-1446, 2020. 2" + ], + "bbox": [ + 516, + 92, + 906, + 585 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7914", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SpikeDiff_ Zero-shot High-Quality Video Reconstruction from Chromatic Spike Camera and Sub-millisecond Spike Streams/3f42c09f-75a5-4dd4-bd4a-8b0a20ea82c7_model.json b/2025/SpikeDiff_ Zero-shot High-Quality Video Reconstruction from Chromatic Spike Camera and Sub-millisecond Spike Streams/3f42c09f-75a5-4dd4-bd4a-8b0a20ea82c7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..88e5f01964d31fdfe505ccf23b88221803c1df86 --- /dev/null +++ b/2025/SpikeDiff_ Zero-shot High-Quality Video Reconstruction from Chromatic Spike Camera and Sub-millisecond Spike Streams/3f42c09f-75a5-4dd4-bd4a-8b0a20ea82c7_model.json @@ -0,0 +1,2453 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.185, + 0.131, + 0.819, + 0.178 + ], + "angle": 0, + "content": "SpikeDiff: Zero-shot High-Quality Video Reconstruction from Chromatic Spike Camera and Sub-millisecondskip Spike Streams" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.202, + 0.262, + 0.223 + ], + "angle": 0, + "content": "Siqi Yang\\(^{1,2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.203, + 0.432, + 0.239 + ], + "angle": 0, + "content": "Jinxiu Liang\\(^{2,3,4*}\\) \nYakun Chang\\(^{5,6}\\)" + }, + { + "type": "text", + "bbox": [ + 0.458, + 0.204, + 0.603, + 0.239 + ], + "angle": 0, + "content": "Zhaojun Huang2,3 Zhaofei Yu1,3" + }, + { + "type": "text", + "bbox": [ + 0.638, + 0.204, + 0.81, + 0.222 + ], + "angle": 0, + "content": "Yeliduosi Xiaokaiti2,3" + }, + { + "type": "text", + "bbox": [ + 0.598, + 0.222, + 0.713, + 0.238 + ], + "angle": 0, + "content": "Boxin Shi\\(^{2,3,1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.338, + 0.241, + 0.662, + 0.257 + ], + "angle": 0, + "content": "1 Institute for Artificial Intelligence, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.17, + 0.258, + 0.828, + 0.275 + ], + "angle": 0, + "content": "\\(^{2}\\) State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.17, + 0.276, + 0.827, + 0.292 + ], + "angle": 0, + "content": "\\(^{3}\\) National Engineering Research Center of Visual Technology, School of Computer Science, Peking University" + }, + { + "type": "list", + "bbox": [ + 0.17, + 0.258, + 0.828, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.2, + 0.294, + 0.404, + 0.309 + ], + "angle": 0, + "content": "\\(^{4}\\) National Institute of Informatics" + }, + { + "type": "text", + "bbox": [ + 0.429, + 0.294, + 0.797, + 0.31 + ], + "angle": 0, + "content": "\\(^{5}\\) Institute of Information Science, Beijing Jiaotong University" + }, + { + "type": "text", + "bbox": [ + 0.209, + 0.311, + 0.788, + 0.345 + ], + "angle": 0, + "content": "\\(^{6}\\) Visual Intelligence +X International Cooperation Joint Laboratory of the Ministry of Education,.cssherryliang@gmail.com, shiboxin@pku.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.35, + 0.903, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.545, + 0.907, + 0.63 + ], + "angle": 0, + "content": "Figure 1. Spike cameras capture chromatic spikes with high temporal resolution, enabling the observation of scenes with rapid motion. As illustrated in the orange region, the spike camera utilizes a color filter array to capture chromatic information and continuously accumulates photons until reaching a threshold that triggers a spike. To reconstruct high-quality videos from these chromatic spikes, we propose SpikeDiff, a zero-shot framework that leverages principled priors from pretrained diffusion models and integrates physics-based guidance derived from the spike camera mechanism. SpikeDiff generates visually appealing videos without noise or motion blur, significantly outperforming existing chromatic spike reconstruction methods (e.g., SJDD [11] and CSpkNet [12])." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.642, + 0.327, + 0.658 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.486, + 0.871 + ], + "angle": 0, + "content": "High-speed video reconstruction from neuromorphic spike cameras offers a promising alternative to traditional frame-based imaging, providing superior temporal resolution and dynamic range with reduced power consumption. Nevertheless, reconstructing high-quality colored videos from spikes captured in ultra-short time intervals (sub-millisecond) remain challenging due to the inherently noisy nature of spikes. While some existing methods extend the temporal capture window to improve reconstruction quality, they inevitably compromise the temporal resolution advantages of spike cameras. In this paper, we introduce SpikeDiff, the first zero-shot framework that leverages pretrained diffusion models to reconstruct high-quality colored videos from sub-millisecond" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.909, + 0.794 + ], + "angle": 0, + "content": "(0.5ms) chromatic spike streams. By incorporating physics-based guidance into the diffusion sampling process, SpikeDiff bridges the domain gap between chromatic spikes and conventional images, enabling high-fidelity reconstruction without requiring domain-specific training data. Extensive experiments demonstrate that SpikeDiff achieves impressive reconstruction quality while maintaining ultra-high temporal resolution, outperforming existing methods across diverse challenging scenarios in both perceptual quality and structural preservation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.813, + 0.646, + 0.829 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.91, + 0.903 + ], + "angle": 0, + "content": "High-speed motions—from droplet impacts to mechanical dynamics—occur at speeds surpassing human perceptual processing capabilities. Conventional high-speed cameras are capable of capturing these motions but encounter funda-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.242, + 0.901 + ], + "angle": 0, + "content": "* Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7905" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "mental trade-offs [37, 45, 51]: higher temporal resolution requires shorter exposures, leading to reduced light capture and limited dynamic range, alongside increased power consumption and larger data volumes. These constraints make continuous high-speed capture over extended periods either impractical or impossible." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.184, + 0.483, + 0.394 + ], + "angle": 0, + "content": "Neuromorphic cameras have emerged as a compelling alternative to conventional high-speed imaging, including event cameras [9, 25, 27] and spike cameras [3, 21, 58]. Drawing inspiration from the fovea in biological visual systems, each pixel in a spike camera continuously accumulates photons, generates spikes upon reaching the predefined threshold, and subsequently proceeds to the next accumulation period. Operating at a sampling rate of \\(20,000\\mathrm{Hz}\\) with \\(1000\\times 1000\\) resolution, spike cameras encode visual information with low data overhead and power consumption, compared to conventional cameras with similar sampling rate and resolution. This enables continuously blur-free capture with high-speed motion and high contrast over several minutes that conventional sensors struggle to handle." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.396, + 0.483, + 0.531 + ], + "angle": 0, + "content": "For human perception, the binary representations of spikes are inherently difficult to interpret and require reconstruction into video frames [54, 55, 58]. However, reconstructing high-quality video frames from spikes remains challenging, particularly within ultra-short time intervals where limited spikes make existing methods highly susceptible to noise [57]. This challenge intensifies for color reconstruction, as Color Filter Arrays (CFAs) further reduce the number of available spikes within short time intervals." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.533, + 0.483, + 0.713 + ], + "angle": 0, + "content": "Existing approaches attempt to mitigate these challenges by extending the observed time intervals to accumulate more spikes [31, 56, 57]. This strategy inadvertently introduces motion blur and ghosting artifacts that compromise the temporal resolution advantage of spike cameras. While recent deep learning-based methods have shown promise in optimizing this trade-off problem, they typically exhibit limited generalization capabilities [10, 11], require extensive training data [12], or demand longer time intervals [48]. Yet, these methods still struggle to completely eliminate the noise [11] (upper right in Figure 1) or avoid introducing motion blur in high-speed motion regions [48] (lower right in Figure 1)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.714, + 0.483, + 0.835 + ], + "angle": 0, + "content": "The recent success of diffusion models in image and video restoration [15, 22, 43, 49] suggests a promising direction. These methods have demonstrated remarkable capability in encoding rich priors about natural scene statistics and temporal dynamics. This raises an intriguing question: Can we leverage these powerful priors to reconstruct high-quality frames and videos from extremely noisy chromatic spikes captured within sub-millisecond \\(^{1}\\) time intervals?" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.836, + 0.483, + 0.852 + ], + "angle": 0, + "content": "Integrating pretrained diffusion models with spike-to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.257 + ], + "angle": 0, + "content": "video reconstruction presents fundamental challenges: Spike cameras capture visual information through asynchronous voltage accumulation and threshold-based sampling—a process fundamentally different from the fixed-interval capturing mechanism of frame-based cameras. Furthermore, chromatic spikes represent raw sensor measurements without established signal processing pipelines, while pretrained diffusion models are typically trained on processed sRGB frames with optimized color and tone characteristics. These disparities create a significant domain gap that must be addressed for effective reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.909, + 0.438 + ], + "angle": 0, + "content": "In this paper, we propose SpikeDiff, the first zero-shot spike-to-video reconstruction framework that leverages pretrained diffusion models as principled priors for high-quality colored video reconstruction from chromatic spikes. By establishing a physically grounded likelihood model of spike measurements throughout the reverse diffusion process, SpikeDiff approximates the posterior sampling of high-quality videos from sub-milliseconds spike streams. Unlike existing approaches that require extensive domain-specific training data and risk overfitting, our method operates in a zero-shot manner, eliminating both the computational burden of training and the need for large-scale datasets." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.44, + 0.738, + 0.453 + ], + "angle": 0, + "content": "Our key contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.454, + 0.907, + 0.514 + ], + "angle": 0, + "content": "- the first framework that leverages pretrained diffusion models as powerful priors for high-quality, zero-shot video reconstruction from chromatic spikes in ultra-short intervals, eliminating the need for domain-specific training data; and" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.515, + 0.905, + 0.575 + ], + "angle": 0, + "content": "- a physics-based and differentiable formulation of the spike generation process, which enables effective integration with pretrained diffusion models, bridging the domain gap between spikes and conventional images." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.454, + 0.907, + 0.575 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.575, + 0.908, + 0.665 + ], + "angle": 0, + "content": "Extensive experiments across diverse challenging scenarios demonstrate that SpikeDiff consistently outperforms prior methods that rely on extended time intervals and compromising temporal resolution, effectively breaking the traditional quality-time resolution trade-off in spike-to-video reconstruction (see our result shown in Figure 1)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.68, + 0.652, + 0.695 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Video reconstruction with monochromatic spikes. Reconstructing high-speed videos from spikes typically employs the imaging model of spike cameras [21]. By accumulating the spikes triggered within a certain time window or estimating the interval between successive spikes, textures can be reconstructed [57, 58]. However, due to the limited number of photons arriving at the pixel in an ultra-short time, such straightforward reconstruction suffers from significant noise. To address this issue, Chang et al. [4] propose enhancing the signal-to-noise ratio (SNR) by combining multi-bit spikes with binary spikes in a rolling-mixed-bit manner. However, this solution requires hardware modifications, limiting its practicality in more general scenes. Zhao et al. [54] present" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.483, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\)Existing state-of-the-art methods [10-12] require at least a time interval of \\(40/20, 000 \\mathrm{~Hz} = 2\\) ms for high-quality video reconstruction, while we aim at ultra-short time interval of \\(10/20, 000 \\mathrm{~Hz} = 0.5\\) ms." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7906" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.318 + ], + "angle": 0, + "content": "Spk2ImgNet, a hierarchical architecture that progressively fuses the spikes, offering an alternative approach to improve reconstruction quality. However, its generalization ability is limited since it relies on synthetic data for training. To address the issue of real-world ground truth for training, self-supervised methods [5, 6] are developed to reduce dependence on synthetic datasets. However, it is still challenging to reconstruct color videos from monochromatic spikes. From the perspective of incorporating additional color guidance, Chang et al. [3] develop a hybrid spike-RGB camera system that performs spatial alignment and frame interpolation simultaneously, enabling the recovery of 1000 FPS color video. However, this hybrid camera system necessitates synchronization and optical alignment, while the beam splitter also poses challenges for constructing a compact device." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.487, + 0.818 + ], + "angle": 0, + "content": "Demosaicking and denoising for unconventional cameras. Various methods for joint demosaicking and denoising have been proposed to obtain high-resolution and low-noise color images from frame-based RAW data, ranging from traditional model-based techniques [1, 8, 19, 23, 32-34, 36, 50] to more recent data-driven approaches [18, 24, 29, 41, 42, 46, 53]. However, all these methods cannot be applied to binary data captured by the new-generation high-speed cameras, such as event cameras [9, 25, 27], quanta image sensors [16, 17], and spike cameras [21]. In the case of event cameras, Xu et al. [47] and Lu et al. [30] present Transformer-based architectures for demosaicking missing pixel values in RAW domain processing. However, due to the sparsity of events, these approaches still rely on frame-based cameras for texture reconstruction. In the context of quanta image sensors, Elgendy et al. [14] integrate a frequency selection method with a deep neural network based filtering approach, leveraging the luma channels to assist in the denoising of the chroma channels. Ma et al. [31] design a blue-noise pseudo-random RGBW color filter array and successfully reconstruct high-quality color images from mosaicked single-bit frames, even in high-dynamic-range (HDR) scenes with complex and rapid motion. For spike cameras, Dong et al. [11] propose an offset-sharing deformable convolution module to align temporal features of color channels and develop a spike noise estimator to characterize the noise distribution. However, this method is trained on synthetic data, limiting its generalization to real-world data. Yang et al. [48] introduce a self-supervised denoising module trained exclusively on real-world chromatic spikes, achieving 2000 FPS color video reconstruction. However, there remains significant research potential in balancing noise suppression with detail preservation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.83, + 0.182, + 0.846 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We present the degradation model of chromatic spikes based on the physical formulation and propose SpikeDiff, a zero-shot framework for reconstructing consecutive high-quality" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.138 + ], + "angle": 0, + "content": "frames from chromatic spikes. This approach leverages pretrained diffusion models as principled priors to effectively address the inherent ill-posedness of the problem." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.146, + 0.655, + 0.16 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.907, + 0.365 + ], + "angle": 0, + "content": "Chromatic spikes. The chromatic spike camera leverages a spike sensor equipped with a Bayer-pattern (RGGB) CFA to capture the scene with chromatic information. As shown in Figure 1, when photons filtered by the CFA reach a pixel, the electrons generated by these photons are continuously accumulated as long as the voltage does not reach the threshold \\( E_{\\mathrm{th}} \\). Simultaneously, the readout circuit samples the pixel values at a frequency of \\( 20,000\\mathrm{Hz} \\) and a signal of 0 is read out at each readout point. Once the accumulated voltage reaches \\( E_{\\mathrm{th}} \\), a signal of 1 is read out and the voltage of this pixel is reset to continue the next accumulation period. For each pixel \\( i \\), we denote the accumulated voltage at a readout point \\( \\tau \\) as \\( E(i,\\tau) \\), the triggered spike at \\( \\tau \\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.373, + 0.907, + 0.415 + ], + "angle": 0, + "content": "\\[\nS (i, \\tau) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} E (i, \\tau) \\geq E _ {\\mathrm {t h}}, \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.909, + 0.538 + ], + "angle": 0, + "content": "Diffusion models. Latent diffusion models (LDMs) operate in a compressed latent space learned by an autoencoder pair \\((\\mathcal{E},\\mathcal{D})\\), where \\(X = \\mathcal{D}(Z)\\) and \\(Z = \\mathcal{E}(X)\\). An LDM learns to reverse a diffusion process that gradually adds Gaussian noise to a latent vector \\(Z_{0}\\). The reverse process, modeled by a stochastic differential equation (SDE), generates a clean latent \\(Z_{0}\\) from pure noise \\(Z_{T} \\sim \\mathcal{N}(0,I)\\) [40]:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.56, + 0.907, + 0.579 + ], + "angle": 0, + "content": "\\[\nd \\boldsymbol {Z} = [ - f (\\boldsymbol {Z}, t) - g ^ {2} (t) \\nabla_ {\\boldsymbol {Z} _ {t}} \\log p _ {t} (\\boldsymbol {Z} _ {t}) ] d t + g (t) d w, (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.586, + 0.909, + 0.693 + ], + "angle": 0, + "content": "where \\( f(\\cdot, t) \\) and \\( g(t) \\) are drift and diffusion coefficients, \\( w \\) represents the standard Brownian motion. The key component is the score function \\( \\nabla_{\\mathbf{Z}_t} \\log p_t(\\mathbf{Z}_t) \\), which is approximated by a time-conditioned neural network \\( \\epsilon_{\\theta}(\\mathbf{Z}, t) \\). By iteratively applying the learned score, the reverse SDE can sample from the learned data prior \\( p(\\mathbf{Z}_0) \\) to generate a high-quality image \\( \\mathbf{X}_0 = \\mathcal{D}(\\mathbf{Z}_0) \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.701, + 0.88, + 0.717 + ], + "angle": 0, + "content": "3.2. Diffusion-based reconstruction from spikes" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.723, + 0.909, + 0.86 + ], + "angle": 0, + "content": "Reconstructing consecutive high-quality frames from real-captured chromatic spike streams is challenging due to the non-negligible noise perturbations, particularly within extremely limited time intervals (e.g., sub-milisecond). To address this challenge, we leverage recent advancements in diffusion models, which incorporate learned priors regarding the distribution of high-quality images, and formulate the reconstruction of video frames from chromatic spikes as a Maximum a Posteriori (MAP) estimation problem:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.868, + 0.907, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {X} ^ {*} = \\operatorname {a r g m a x} _ {\\boldsymbol {X}} p (\\boldsymbol {X} | \\boldsymbol {Y}) \\tag {3} \\\\ = \\operatorname {a r g m a x} _ {\\boldsymbol {X}} p (\\boldsymbol {Y} | \\boldsymbol {X}) p (\\boldsymbol {X}), \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7907" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.9, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.399, + 0.907, + 0.497 + ], + "angle": 0, + "content": "Figure 2. Pipeline of SpikeDiff. Our diffusion-based posterior sampling procedure follows the iterative denoising process of diffusion models, starting from the initialization \\( \\mathbf{Z}_{t_s} \\) and gradually approaching the desired latent \\( \\mathbf{Z}_0 \\), indicated by the dashed lines. To maximize the likelihood of the reconstructed video given the spike observations \\( \\mathbf{Y} \\), we enhance each iteration with physics-based spike guidance, which comprises three specially-designed differentiable degradation operators: mosaicking \\( M \\), color casting \\( C \\), and quantization \\( Q \\). Finally, a multiscale loss function bridges the gap between degraded frames \\( \\mathcal{A}(\\mathbf{X}_{0|t}) \\) and observed SFR frames \\( \\mathbf{Y} \\). As shown in the top sequence, the visual quality of generated frames progressively improves until convergence. Note that the internal degraded frames and the SFR frames are all Bayer-patterned mosaic images; we present their corresponding demosaicked results for better visualization." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.504, + 0.486, + 0.566 + ], + "angle": 0, + "content": "where the prior \\(p(\\mathbf{X})\\) encodes knowledge of natural images, and the likelihood \\(p(\\mathbf{Y}|\\mathbf{X})\\) represents the data fidelity to ensure the reconstruction is faithful to the observed chromatic spikes \\(\\mathbf{Y}\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.581, + 0.486, + 0.764 + ], + "angle": 0, + "content": "The degradation model, mapping light intensity to spike signals (illustrated in Figure 1), establishes the connection from RGB frames \\( X \\) — the modality employed in pretrained diffusion models—and chromatic spikes \\( Y \\), and thereby deriving likelihood \\( p(\\mathbf{Y}|\\mathbf{X}) \\) illustrated in the next (Sec. 3.3). This establishes a principled bridge that enables the reconstruction of \\( X \\) from \\( Y \\) via posterior sampling in the MAP framework, as illustrated in the green box in Figure 2, where diffusion models serve as learned priors for \\( X \\) while the likelihood ensures fidelity to \\( Y \\). This is exactly the key in our method to connect the modality used in pretrained diffusion models, RGB frames, and chromatic spikes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.78, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Given the significant disparity between chromatic spike streams and conventional images, the data fidelity term \\( p(\\mathbf{Y}|\\mathbf{X}) \\) cannot be directly derived from time-continuous chromatic spikes. Instead, we resort to a more compact representation. According to the integrating and firing mechanism of spikes, the light intensity is proportional to the spike firing rate (SFR) [57]. Without considering noise, the SFR at \\( \\tau' \\) can be counted through collecting a number of spikes within" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.505, + 0.766, + 0.521 + ], + "angle": 0, + "content": "a predefined time interval of size \\(\\Delta \\tau\\)" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.528, + 0.907, + 0.573 + ], + "angle": 0, + "content": "\\[\nY (i, \\tau^ {\\prime}) = \\frac {1}{\\Delta \\tau} \\cdot \\sum_ {\\tau = \\tau^ {\\prime} - \\Delta \\tau / 2} ^ {\\tau^ {\\prime} + \\Delta \\tau / 2} S (i, \\tau). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.909, + 0.763 + ], + "angle": 0, + "content": "By collecting the SFRs from the pixel array at \\(\\tau'\\), we obtain mosaicked SFR frames \\(Y \\in \\mathbb{R}^{T \\times H \\times W}\\), where \\(T\\) denotes the number of frames, and \\(H, W\\) represent the height and width, respectively. The SFR frames, estimated from real-captured chromatic spikes, are sensitive to the inherent noise, especially within limited time intervals, and exhibit misalignment with corresponding consecutive high-quality frames. Consequently, we propose a corresponding degradation model, comprising differentiable operators, as detailed in Sec. 3.3, to bridge the gap between \\(X\\) and \\(Y\\) and resolve the likelihood estimation required for posterior sampling, as illustrated in the orange box in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Posterior sampling. The primary challenge in reconstructing consecutive high-quality frames from sub-millisecond chromatic spikes lies in its inherent ill-posedness, making the prior term \\( p(\\mathbf{X}) \\) crucial. We leverage diffusion models as powerful priors. Given the prior distribution of natural images \\( p(\\mathbf{Z}_t) \\) encoded by the pretrained latent diffusion model in Eq. (2) and likelihood defined in the next, we solve the MAP problem in Eq. (3) by sampling from the posterior" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7908" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.155 + ], + "angle": 0, + "content": "distribution \\( p(\\mathbf{X}|\\mathbf{Y}) \\). This is achieved by modifying the reverse SDE of the diffusion model to incorporate guidance from our observations. Using Bayes' rule, the score of the posterior distribution \\( p_t(\\mathbf{Z}_t|\\mathbf{Y}) \\) can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.163, + 0.485, + 0.193 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\boldsymbol {Z} _ {t}} \\log p _ {t} (\\boldsymbol {Z} _ {t} | \\boldsymbol {Y}) = \\nabla_ {\\boldsymbol {Z} _ {t}} \\log p (\\boldsymbol {Z} _ {t}) + \\nabla_ {\\boldsymbol {Z} _ {t}} \\log p (\\boldsymbol {Y} | \\boldsymbol {Z} _ {t}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.194, + 0.486, + 0.271 + ], + "angle": 0, + "content": "where the prior term \\(\\nabla_{\\mathbf{Z}_t}\\log p(\\mathbf{Z}_t)\\) is given by the pretrained network \\(\\epsilon_{\\theta^*}(Z_t,t)\\). The likelihood term \\(p(\\mathbf{Y}|\\mathbf{Z}_t)\\) guides the sampling towards latents \\(\\mathbf{Z}_t\\) that are likely to produce our observed SFR frame \\(\\mathbf{Y}\\). The full reverse SDE for posterior sampling is:" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.279, + 0.486, + 0.316 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} d \\boldsymbol {Z} = [ - f (\\boldsymbol {Z}, t) - g ^ {2} (t) (\\nabla_ {\\boldsymbol {Z} _ {t}} \\log p (\\boldsymbol {Z} _ {t}) \\tag {6} \\\\ + \\nabla_ {\\boldsymbol {Z} _ {t}} \\log p (\\boldsymbol {Y} | \\boldsymbol {Z} _ {t})) ] d t + g (t) d w. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.325, + 0.485, + 0.388 + ], + "angle": 0, + "content": "To compute the likelihood score, we must relate the noisy latent \\( Z_{t} \\) at an intermediate timestep \\( t \\) to the observation \\( Y \\). This involves approximating the conditional probability \\( p(\\mathbf{Y}|\\mathbf{Z}_t) \\), which can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.396, + 0.484, + 0.46 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} p (\\mathbf {Y} \\mid \\mathbf {Z} _ {t}) = \\int_ {c} p (\\mathbf {Y} \\mid \\mathbf {Z} _ {0}, \\mathbf {Z} _ {t}) p (\\mathbf {Z} _ {0} \\mid \\mathbf {Z} _ {t}) d \\mathbf {Z} _ {0} \\tag {7} \\\\ = \\int p (\\boldsymbol {Y} | \\boldsymbol {Z} _ {0}) p (\\boldsymbol {Z} _ {0} | \\boldsymbol {Z} _ {t}) d \\boldsymbol {Z} _ {0}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.469, + 0.484, + 0.53 + ], + "angle": 0, + "content": "This integral is intractable. Instead, we approximate it by first estimating the expected clean latent \\( \\mathbf{Z}_{0|t} = \\mathbb{E}[Z_0|\\mathbf{Z}_t] \\) from \\( \\mathbf{Z}_t \\). For diffusion models like DDPM [20], the forward process is" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.54, + 0.484, + 0.557 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {Z} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {Z} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} z, \\quad z \\sim \\mathcal {N} (0, I), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.484, + 0.613 + ], + "angle": 0, + "content": "where \\(\\bar{\\alpha}_t\\) is the noise schedule parameter from the diffusion model (e.g., DDPM [20]). Using Tweedie's formula [7], we can estimate the conditional mean:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.62, + 0.484, + 0.671 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {Z} _ {0 | t} = \\mathbb {E} [ \\boldsymbol {Z} _ {0} | \\boldsymbol {Z} _ {t} ] \\\\ = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(\\boldsymbol {Z} _ {t} + \\left(1 - \\bar {\\alpha} _ {t}\\right) \\nabla_ {\\boldsymbol {Z} _ {t}} \\log p _ {t} \\left(\\boldsymbol {Z} _ {t}\\right)\\right). \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.681, + 0.484, + 0.772 + ], + "angle": 0, + "content": "Here, we replace the true score \\(\\nabla_{Z_t}\\log p_t(Z_t)\\) with its neural network approximation \\(\\epsilon_{\\theta^*}(Z_t,t)\\). With this estimate of the clean latent, \\(Z_{0|t}\\), following the expression \\(p(\\mathbf{Y}|\\mathbf{Z}_t) = \\mathbb{E}_{\\mathbf{Z}_0\\sim p(\\mathbf{Z}_0|\\mathbf{Z}_t)}[p(\\mathbf{Y}|\\mathcal{D}(\\mathbf{Z}_0))]\\), we can use a surrogate likelihood approach [7] to make the likelihood tractable:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.782, + 0.484, + 0.8 + ], + "angle": 0, + "content": "\\[\np (\\boldsymbol {Y} | \\boldsymbol {Z} _ {t}) \\approx p (\\boldsymbol {Y} | \\mathcal {D} (\\mathbb {E} [ \\boldsymbol {Z} _ {0} | \\boldsymbol {Z} _ {t} ])). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.84 + ], + "angle": 0, + "content": "This formulation provides a practical way to compute the likelihood gradient." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.433, + 0.866 + ], + "angle": 0, + "content": "3.3. Chromatic spikes' likelihood estimation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Given a set of sequential SFR frames, denoted by \\( \\mathbf{Y} \\), our objective is to establish a connection between the SFR frames" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.153 + ], + "angle": 0, + "content": "and the desired target reconstructions \\( \\mathbf{X} \\), with both \\( \\mathbf{Y} \\) and \\( \\mathbf{X} \\) normalized to the range [0, 1]. Hence, we resolve the likelihood estimation of \\( p(\\mathbf{Y} | Z_t) \\) in Eq. (10) and thereby complete the diffusion-based posterior sampling strategy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.171, + 0.907, + 0.291 + ], + "angle": 0, + "content": "Degradation model of chromatic spikes. The physical process generating SFR frames \\(\\mathbf{Y}\\) can be modeled as a nonlinear inverse problem: \\(\\mathbf{Y} = \\mathcal{A}(\\mathbf{X}) + \\mathbf{N}\\), where \\(\\mathcal{A}(\\cdot)\\) denotes the known forward operator modeling the degradation process, and \\(\\mathbf{N}\\) represents stochastic noise. This problem is inherently ill-posed due to irreversible information loss during the imaging process and interference from noise in the capture system." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.292, + 0.908, + 0.428 + ], + "angle": 0, + "content": "The forward operator \\(\\mathcal{A}(\\cdot)\\) in the imaging pipeline transforms the latent high-quality frame \\(X\\) through a sequence of operations: \\(\\mathcal{A}(X) = \\mathcal{Q}(CMX)\\), where \\(\\mathcal{Q}(\\cdot)\\) represents pixel-wise quantization, \\(C\\) implements pixel-wise color casting with per-channel scaling, and \\(M: \\mathbb{R}^{T \\times H \\times W \\times 3} \\to \\mathbb{R}^{T \\times H \\times W}\\) denotes the mosaicking operator, modeling the color filter array pattern. This decomposition explicitly models each step of the image formation process, providing a foundation for our reconstruction approach." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.428, + 0.907, + 0.489 + ], + "angle": 0, + "content": "Under the assumption of Gaussian noise, we model the likelihood \\( p(\\mathbf{Y}|\\mathcal{D}(\\mathbf{Z}_{0|t})) \\) of observing SFR frames \\( \\mathbf{Y} \\) given the target frame \\( \\mathbf{X} \\) as a Gaussian distribution. Maximizing this likelihood translates to minimizing the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.499, + 0.907, + 0.528 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} \\left(\\boldsymbol {Z} _ {t}\\right) = \\frac {1}{\\sigma^ {2}} \\| \\boldsymbol {Y} - \\mathcal {A} \\left(\\mathcal {D} \\left(\\boldsymbol {Z} _ {0 | t}\\right)\\right) \\| _ {2} ^ {2}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.908, + 0.597 + ], + "angle": 0, + "content": "where \\(\\sigma\\) represents the standard deviation of the noise \\(N\\), and \\(\\mathcal{A}\\) represents the differentiable degradation operators defined later. By substituting the score function into Eq. (5), we obtain:" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.61, + 0.907, + 0.628 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\boldsymbol {Z} _ {t}} \\log p _ {t} (\\boldsymbol {Z} _ {t} | \\boldsymbol {Y}) = \\epsilon_ {\\theta^ {*}} (\\boldsymbol {Z} _ {t}, t) - s \\nabla_ {\\boldsymbol {Z} _ {t}} \\mathcal {L} (\\boldsymbol {Z} _ {t}), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.639, + 0.882, + 0.654 + ], + "angle": 0, + "content": "where \\(s\\) controls the strength of measurement guidance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.907, + 0.87 + ], + "angle": 0, + "content": "Differentiable degradation operators. A key innovation in SpikeDiff is our design of differentiable operators in \\(\\mathcal{A}\\) that model the physical characteristics of chromatic spike cameras, which include color casting \\(C\\), quantization \\(\\mathcal{Q}\\), and mosaicking \\(M\\). Based on the gray-world prior [2] for the target frame, which indicates that each color channel should have a mean value close to one another, we estimate the color casting coefficients \\(\\alpha_{R}, \\alpha_{G}, \\alpha_{B}\\) as \\(\\alpha_{c} = \\frac{1}{\\mu(\\mathbf{Y}_{c})}\\), where \\(c \\in \\{\\mathrm{R}, \\mathrm{G}, \\mathrm{B}\\}\\) denotes the color channel, \\(\\mu(\\mathbf{Y}_{c})\\) is the mean value of color channel \\(c\\) among \\(\\mathbf{Y}\\) across the temporal axis. Then the channel-wise scaling factors for the color casting operator \\(C\\) can be computed as \\(\\eta / \\alpha_{c}\\), where parameter \\(\\eta\\) is used to adjust the value range." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "SFR frames are inherently quantized and discrete as they are accumulation of binary spike frames, which differ from" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7909" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.168 + ], + "angle": 0, + "content": "8-bit frames \\( \\mathbf{X} \\). To simulate the quantization effect and maintain the backward differentiability, we design a soft quantization operator \\( \\mathcal{Q} \\). Given the expected quantization levels determined by SFR estimation, the soft quantization function \\( q(\\cdot) \\) operates on each pixel individually:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.178, + 0.483, + 0.195 + ], + "angle": 0, + "content": "\\[\nq (x) = l (x) \\cdot (1 - \\varphi (x)) + u (x) \\cdot \\varphi (x), \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.198, + 0.483, + 0.227 + ], + "angle": 0, + "content": "\\[\n\\varphi (x) = \\left[ \\tanh \\left(k \\left(x - \\frac {l (x) + u (x)}{2}\\right)\\right) + 1 \\right] / 2, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.237, + 0.483, + 0.297 + ], + "angle": 0, + "content": "where \\( l(x) \\) and \\( u(x) \\) denote the lower and upper quantization levels of \\( x \\), \\( \\varphi(\\cdot) \\) quantizes the input value to the range [0, 1] with a smooth curve, and hyperparameter \\( k \\) controls the smoothness, as shown in Figure 3(a)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.298, + 0.484, + 0.403 + ], + "angle": 0, + "content": "The naive Bayer-pattern mosaicking operator \\( M \\) converts target frame from dimension \\( T \\times H \\times W \\times 3 \\) to \\( T \\times \\frac{H}{2} \\times \\frac{W}{2} \\times 4 \\), resulting in a partially defined backpropagation function due to its gradient discontinuity. To simulate the Bayer-pattern CFA while maintaining a dense gradient flow, we extend the operator's derivative with linear interpolation convolution, similar to demosaicking algorithms." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.404, + 0.484, + 0.449 + ], + "angle": 0, + "content": "By composing these operators, we conclude a differentiable simulation from consecutive frames to SFR frames, enabling optimization in the posterior sampling process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.467, + 0.484, + 0.603 + ], + "angle": 0, + "content": "Multiscale enhancement. To preserve fine spatial details, we incorporate a Laplacian pyramid \\(\\mathcal{P}(\\boldsymbol{X})\\) that guides the diffusion process across multiple scales. The Laplacian pyramid decomposes the generated consecutive frames into a multiscale representation of frequency components, facilitating the preservation of fine details critical for high-quality output. The Laplacian pyramid is calculated by subtracting the upsampled version of the downsampled frames from the original frames:" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.612, + 0.483, + 0.644 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} (\\boldsymbol {X}) = \\left\\{\\boldsymbol {X} _ {l} - \\mathbf {U} \\left(\\mathbf {D} \\left(\\boldsymbol {X} _ {l}\\right)\\right) \\right\\} _ {l = 1} ^ {L}, \\boldsymbol {X} _ {l} = \\mathbf {U} \\left(\\mathbf {D} \\left(\\boldsymbol {X} _ {l - 1}\\right)\\right) \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.484, + 0.72 + ], + "angle": 0, + "content": "where \\(\\mathbf{D}\\) and \\(\\mathbf{U}\\) are the upsampling operator and downsampling operator, respectively, and \\(L\\) is the number of pyramid levels. This multiscale decomposition effectively guides the diffusion process across different frequency bands, ensuring preservation of high-frequency details in reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.734, + 0.222, + 0.75 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.304, + 0.773 + ], + "angle": 0, + "content": "4.1. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Quantization levels. As introduced in Section 3.2, SFR frames are accumulations over predefined time intervals \\(\\Delta \\tau\\). Given that the spikes are binary, the value range of SFR frames is restricted to \\(\\left\\{\\frac{u}{\\Delta\\tau}|u\\in \\mathbb{N}\\cap [0,\\Delta \\tau ]\\right\\}\\). For instance, we set \\(\\Delta \\tau = 10\\) in our experiments to achieve sub-milliseconds reconstruction, and the corresponding quantization levels are \\(0,0.1,0.2,\\ldots ,0.9,1\\), leading to the soft quantization function visualized in Figure 3(a)." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.09, + 0.639, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.519, + 0.156, + 0.639, + 0.168 + ], + "angle": 0, + "content": "(a) Visualization of \\(q(x)\\)" + }, + { + "type": "image", + "bbox": [ + 0.65, + 0.09, + 0.732, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.733, + 0.091, + 0.817, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.818, + 0.091, + 0.901, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.701, + 0.157, + 0.852, + 0.168 + ], + "angle": 0, + "content": "(b) Quantization effectiveness" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.172, + 0.906, + 0.227 + ], + "angle": 0, + "content": "Figure 3. Visualization of the quantization operator. (a) The soft quantization function \\( q(x) \\) for SFR frames with \\( \\Delta \\tau = 10 \\). (b) Detailed demonstration of quantization effectiveness on continuous pixel values." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.235, + 0.906, + 0.308 + ], + "angle": 0, + "content": "Color casting estimation. To obtain better color casting coefficients and maintain consistency across video frames reconstruction, we average the estimations of \\(\\alpha_{c}\\) from all SFR frames in the sequence, and apply the same \\(\\alpha_{c}\\) throughout the video reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.31, + 0.906, + 0.431 + ], + "angle": 0, + "content": "Pretrained diffusion model. We implement SpikeDiff upon the pretrained weights from [28], which is adapted from Stable Diffusion v2.1 [38], selected for their strong priors and close alignment with our task. We extend the image-to-image pipeline to spikes-to-video by dividing SFR frames into small chunks, and achieve temporal consistency by leveraging high-fidelity reconstruction from time-continuous chromatic spikes with physics-based guidance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.44, + 0.693, + 0.455 + ], + "angle": 0, + "content": "4.2. Evaluation dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.908, + 0.687 + ], + "angle": 0, + "content": "As our proposed method leverages pretrained diffusion models, which contain principled priors of the real world images and videos, we qualitatively evaluate the performance of our method on a real-world dataset. To this end, we collect a series of real chromatic spikes within the spike camera. The spike camera is equipped with a Bayer-pattern color filter array and a \\(1000 \\times 1000\\) pixel spike sensor. Limited by the generation ability of pretrained diffusion models and resolution of compared methods [48], we spatially crop the chromatic spikes into a lower resolution of \\(512 \\times 512\\), maintaining the same Bayer-pattern (RGGB). The collected dataset contains chromatic spikes from various scenes with different levels of motion and color variation. We would like to contribute this dataset to the community for further research." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.908, + 0.825 + ], + "angle": 0, + "content": "Despite the real-captured chromatic spikes, we also employ the chromatic spikes simulator described in CSpKNet [12] and SY24 [48] to generate synthetic chromatic spikes from real-world high-frame-rate videos. We collect over 100 dynamic scenarios from X4K1000FPS [39] dataset, and generate the corresponding chromatic spikes with the simulator. The synthetic dataset is used to conduct the quantitative evaluation of our proposed method, which further demonstrates the superiority of SpikeDiff." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.719, + 0.85 + ], + "angle": 0, + "content": "4.3. Qualitative evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "To demonstrate the superiority of our proposed method, we conduct the qualitative evaluation on chromatic spikes captured from the real world, comparing SpikeDiff with exist" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7910" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.135, + 0.116, + 0.146 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.066, + 0.895, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.602, + 0.907, + 0.659 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison on real-world captured chromatic spikes (a-c), and synthetic chromatic spikes simulated from real-world high-frame-rate videos (d). Compared methods include TFI [57], TFP [57], 3DRI [10], SJDD [11], CSpkNet [12], and SY24 [48]. Details in red / blue bounding boxes are shown on the top. All results are reconstructed from sub-millisecond spike streams (0.5ms), where most methods suffer from severe noise or extreme degradation while SpikeDiff recovers clean and visually pleasing results." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.665, + 0.487, + 0.892 + ], + "angle": 0, + "content": "ing chromatic spike reconstruction methods. Our proposed method can recover high-quality color frames from sub-milisecond time intervals (10 spike frames, corresponding to \\(0.5\\mathrm{ms}\\)), which none of the existing neural network based method achieve. We adapt and evaluate the existing chromatic spike reconstruction methods (3DRI [10], SJDD [11], CSpkNet [12], and SY24 [48]), combining the SFR estimation methods (TFI [57], TFP [57]) with demosaicking as baselines. We conduct all experiments on sub-millisecond time intervals (0.5ms) to maintain the fairness in comparison. As shown in Figure 4(a), where a fan is fast rotating in front of the checkerboard, SpikeDiff produces the most clean and visually pleasant results. In contrast, other methods, except CSpkNet and SY24, contain significantly noisy pixels as highlighted in the red and blue bounding boxes. The results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.665, + 0.909, + 0.893 + ], + "angle": 0, + "content": "of CSpkNet and SY24 contain blurry or degraded artifacts. As the time intervals are strictly limited, all methods are free of motion blur naturally. As for Figure 4(b), we capture a static scene in front of the lamp with the spike camera slightly shaking, to study the reconstruction quality under hard light. Compared to others, SpikeDiff generates the most clear and sharp details in blue bounding box. And SpikeDiff produces a smoother texture under the direct illumination in the red bounding box, while most of other methods show obvious quantized artifacts. In Figure 4(c), we leverage the spike streams captured by Yang et al. [48], where SpikeDiff outperforms all other methods in both noise suppression and visual quality. We also illustrate a synthetic scene in Figure 4(d), which further demonstrates the superior reconstruction capability of SpikeDiff on chromatic spikes" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "7911" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.483, + 0.132 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of the proposed method with existing chromatic spike reconstruction methods. The best and second-best results are highlighted in red and blue, respectively." + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.133, + 0.465, + 0.242 + ], + "angle": 0, + "content": "
MethodPSNR↑SSIM↑FID↓NIQE↓IL-NIQE↓
SpikeDiff (Ours)18.6940.7502.8805.17338.535
SY24 [48]13.0130.53621.47211.59278.672
SJDD [11]11.3580.32526.00911.27844.711
3DRI [10]18.5120.39016.0368.54141.816
CSpkNet [12]13.3640.6414.6995.81939.948
TFP [57]12.9090.5593.25612.00164.554
TFI [57]14.6400.5527.55812.51855.758
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.25, + 0.485, + 0.307 + ], + "angle": 0, + "content": "Table 2. Temporal consistency evaluation, reporting Frame CLIP Score (Frame C.S.), Interpolation Error (Inp. Err.), and Interpolation PSNR (Inp. PSNR) following the metrics from [26]. The best and second-best are highlighted in red and blue, respectively." + }, + { + "type": "table", + "bbox": [ + 0.13, + 0.308, + 0.445, + 0.394 + ], + "angle": 0, + "content": "
MethodFrame C.S.↑Inp. Err.↓Inp. PSNR↑
SpikeDiff (Ours)0.9690.05725.45
SY24 [48]0.9630.04128.65
SJDD [11]0.9620.11319.34
3DRI [10]0.9650.10419.87
CSpkNet [12]0.9660.05925.27
" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.401, + 0.192, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.475, + 0.173, + 0.485 + ], + "angle": 0, + "content": "Grad. w/o" + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.485, + 0.176, + 0.496 + ], + "angle": 0, + "content": "Mosaicking" + }, + { + "type": "image", + "bbox": [ + 0.193, + 0.402, + 0.288, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.475, + 0.264, + 0.484 + ], + "angle": 0, + "content": "Grad.w/" + }, + { + "type": "image_caption", + "bbox": [ + 0.21, + 0.484, + 0.27, + 0.496 + ], + "angle": 0, + "content": "Mosaicking" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.402, + 0.384, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.307, + 0.475, + 0.365, + 0.484 + ], + "angle": 0, + "content": "Result w/o" + }, + { + "type": "image_caption", + "bbox": [ + 0.303, + 0.484, + 0.373, + 0.496 + ], + "angle": 0, + "content": "Color casting" + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.402, + 0.477, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.405, + 0.475, + 0.456, + 0.484 + ], + "angle": 0, + "content": "Result w/" + }, + { + "type": "image_caption", + "bbox": [ + 0.396, + 0.484, + 0.466, + 0.495 + ], + "angle": 0, + "content": "Color casting" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.499, + 0.484, + 0.554 + ], + "angle": 0, + "content": "Figure 5. Analysis of differentiable degradation operators. Generated frames receive sparse gradient from chromatic spikes without mosaicking operator. Similarly, reconstructions exhibit incorrect color distribution without the color casting operator." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.563, + 0.484, + 0.623 + ], + "angle": 0, + "content": "simulated from real-world videos. Note that post-processing (e.g., color casting) cannot eliminate noise or artifacts from other methods' results. Please refer to our supplementary material for more comparison and detailed discussion." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.634, + 0.308, + 0.649 + ], + "angle": 0, + "content": "4.4. Quantitative evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.484, + 0.839 + ], + "angle": 0, + "content": "As shown in Table 1, we quantitatively evaluate the performance of SpikeDiff and existing chromatic spike reconstruction methods on the synthetic dataset. Similar to our qualitative evaluation, we evaluate these methods on submillisecond time intervals (0.5ms). Our method achieves the best performance across all the metrics, including PSNR, SSIM [44], FID [13], NIQE [35], and IL-NIQE [52], which demonstrates the superiority of SpikeDiff, especially in terms of perceptual metrics that reflect visual quality. As shown in Table 2, we also evaluate the temporal consistency metrics of SpikeDiff and existing reconstruction methods, where SpikeDiff achieves the best or comparable performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.241, + 0.865 + ], + "angle": 0, + "content": "4.5. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To investigate the effectiveness of our proposed degradation operators, we visualize the internal results after applying" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.117 + ], + "angle": 0, + "content": "Table 3. Quantitative ablation study on the contribution of different degradation operators to reconstruction quality." + }, + { + "type": "table", + "bbox": [ + 0.549, + 0.118, + 0.869, + 0.193 + ], + "angle": 0, + "content": "
MethodPSNR↑SSIM↑FID↓NIQE↓IL-NIQE↓
SpikeDiff18.6940.7502.8805.17338.535
w/o C16.7520.72912.0646.07244.661
w/o M17.4810.7234.3926.20341.356
w/o Q17.4600.7173.9386.31840.487
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.2, + 0.907, + 0.397 + ], + "angle": 0, + "content": "each operator. The mosaicking operator \\( M \\) converts the generated frames into Bayer-pattern images, which is the beginning step. As shown in Figure 5, without the specific backward formulation for \\( M \\), the gradient map is not well-defined. The color casting operator \\( C \\) is responsible for color correction, which degrades the normal colored image to a greener one (which is more close to the SFR frames) as shown in Figure 5. For the soft quantization operator \\( \\mathcal{Q} \\), its effectiveness is demonstrated in Figure 3(b), quantizing the continuous white pixels into discrete levels, similar to the pattern of the SFR frames. As shown in Table 3, we further demonstrate the effectiveness of each degradation operator via quantitative evaluation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.411, + 0.634, + 0.427 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.438, + 0.907, + 0.619 + ], + "angle": 0, + "content": "We propose the first zero-shot chromatic spike reconstruction method, named SpikeDiff, to recover consecutive clean and high-quality colored frames from sub-millisecond spikes. SpikeDiff leverages the deep priors from pretrained diffusion models to address the extreme noise under very limited time intervals and generate visually pleasant results. SpikeDiff integrates the physics-based chromatic spikes' likelihood into the diffusion-based posterior sampling process, which bridges the domain gap between spikes and conventional images. We conduct experiments on both synthetic and real-captured data to demonstrate the superior performance of SpikeDiff over existing reconstruction methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.619, + 0.909, + 0.756 + ], + "angle": 0, + "content": "Limitations and future work. As shown in Figure 4(b), there exist \"overexposure\" cases in the central region under extreme direct illumination. This imperfection is due to the predefined spike firing threshold, resulting in consistent spikes among these pixels. We observed that it is hard to generate ideal details without any explicit guidance. In future research, we plan to explore more degradation model designs and diffusion techniques to handle this issue and further extend the HDR capability of chromatic spike reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.77, + 0.673, + 0.787 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "This work is supported by National Natural Science Foundation of China (Grant No. 62302019, 62088102, 62136001), Beijing Natural Science Foundation (Grant No. L233024), and Beijing Municipal Science & Technology Commission, Administrative Commission of Zhongguancun Science Park (Grant No. Z241100003524012). PKU-affiliated authors thank openbayes.com for providing computing resources." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7912" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.156 + ], + "angle": 0, + "content": "[1] Antoni Buades, Bartomeu Coll, Jean-Michel Morel, and Catalina Sbert. Self-similarity driven color demosaicking. IEEE TIP, 18(6):1192-1202, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.16, + 0.484, + 0.2 + ], + "angle": 0, + "content": "[2] Gershon Buchsbaum. A spatial processor model for object colour perception. Journal of the Franklin institute, 310(1): 1-26, 1980. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.204, + 0.484, + 0.258 + ], + "angle": 0, + "content": "[3] Yakun Chang, Chu Zhou, Yuchen Hong, Liwen Hu, Chao Xu, Tiejun Huang, and Boxin Shi. 1000 FPS HDR video with a Spike-RGB hybrid camera. In CVPR, pages 22180-22190, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.262, + 0.483, + 0.316 + ], + "angle": 0, + "content": "[4] Yakun Chang, Yeliduosi Xiaokaiti, Yujia Liu, Bin Fan, Zhaojun Huang, Tiejun Huang, and Boxin Shi. Towards HDR and HFR video from rolling-mixed-bit spikings. In CVPR, pages 25117-25127, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.321, + 0.484, + 0.374 + ], + "angle": 0, + "content": "[5] Shiyan Chen, Chaoteng Duan, Zhaofei Yu, Ruiqin Xiong, and Tiejun Huang. Self-supervised mutual learning for dynamic scene reconstruction of spiking camera. In IJCAI, page 2859-2866, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.379, + 0.484, + 0.419 + ], + "angle": 0, + "content": "[6] Shiyan Chen, Zhaofei Yu, and Tiejun Huang. Self-supervised joint dynamic scene reconstruction and optical flow estimation for spiking camera. In AAAI, pages 350-358, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.423, + 0.484, + 0.476 + ], + "angle": 0, + "content": "[7] Hyungjin Chung, Jeongsol Kim, Michael Thompson McCann, Marc Louis Klasky, and Jong Chul Ye. Diffusion Posterior Sampling for General Noisy Inverse Problems. In ICLR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.482, + 0.521 + ], + "angle": 0, + "content": "[8] Laurent Condat and Saleh Mosaddegh. Joint demosaicking and denoising by total variation minimization. In ICIP, pages 2781-2784, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.525, + 0.484, + 0.579 + ], + "angle": 0, + "content": "[9] Tobi Delbruck et al. Frame-free dynamic digital vision. In Proc. of Intl. Symp. on Secure-Life Electronics, Advanced Electronics for Quality Life and Society, pages 21-26. Cite-seer, 2008. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.583, + 0.484, + 0.624 + ], + "angle": 0, + "content": "[10] Yanchen Dong, Jing Zhao, Ruiqin Xiong, and Tiejun Huang. 3D residual interpolation for spike camera demosaicing. In ICIP, pages 1461-1465. IEEE, 2022. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.627, + 0.484, + 0.681 + ], + "angle": 0, + "content": "[11] Yanchen Dong, Ruiqin Xiong, Jing Zhao, Jian Zhang, Xiaopeng Fan, Shuyuan Zhu, and Tiejun Huang. Joint demosaicing and denoising for spike camera. In AAAI, pages 1582-1590, 2024. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.686, + 0.484, + 0.74 + ], + "angle": 0, + "content": "[12] Yanchen Dong, Ruiqin Xiong, Jing Zhao, Jian Zhang, Xiaopeng Fan, Shuyuan Zhu, and Tiejun Huang. Learning a deep demosaicing network for spike camera with color filter array. IEEE TIP, 2024. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.744, + 0.482, + 0.784 + ], + "angle": 0, + "content": "[13] DC Dowson and BV Landau. The frechet distance between multivariate normal distributions. Journal of multivariate analysis, 12(3):450-455, 1982. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.788, + 0.484, + 0.841 + ], + "angle": 0, + "content": "[14] Omar A Elgendy, Abhiram Gnanasambandam, Stanley H Chan, and Jiaju Ma. Low-light demosaicking and denoising for small pixels using learned frequency selection. IEEE TCI, 7:137-150, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[15] Berthy T. Feng, Jamie Smith, Michael Rubinstein, Huiwen Chang, Katherine L. Bouman, and William T. Freeman. Score-Based Diffusion Models as Principled Priors for Inverse Imaging. In ICCV, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[16] Eric R Fossum, Jiaju Ma, and Saleh Masoodian. Quanta image sensor: Concepts and progress. Advanced Photon Counting Techniques X, 9858:985805, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[17] Eric R Fossum, Jiaju Ma, Saleh Masoodian, Leo Anzagira, and Rachel Zizza. The quanta image sensor: Every photon counts. Sensors, 16(8):1260, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[18] Michael Gharbi, Gaurav Chaurasia, Sylvain Paris, and Frédo Durand. Deep joint demosaicking and denoising. ACM TOG, 35(6):1-12, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[19] Felix Heide, Mushfiqur Rouf, Matthias B Hullin, Bjorn Labitzke, Wolfgang Heidrich, and Andreas Kolb. High-quality computational imaging through simple lenses. ACM TOG, 32 (5):1-14, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.313 + ], + "angle": 0, + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[21] Tiejun Huang, Yajing Zheng, Zhaofei Yu, Rui Chen, Yuan Li, Ruiqin Xiong, Lei Ma, Junwei Zhao, Siwei Dong, Lin Zhu, et al. \\(1000\\times\\) faster camera and machine vision with ordinary devices. Engineering, 25:110-119, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[22] Siddhant Jain, Daniel Watson, Eric Tabellion, Aleksander Hołyński, Ben Poole, and Janne Kontkanen. Video Interpolation with Diffusion Models. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.467 + ], + "angle": 0, + "content": "[23] Daisuke Kiku, Yusuke Monno, Masayuki Tanaka, and Masatoshi Okutomi. Beyond color difference: Residual interpolation for color image demosaicking. IEEE TIP, 25(3): 1288-1300, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[24] Filippos Kokkinos and Stamatios Lefkimmiatis. Deep image demosaicking using a cascade of convolutional residual denoising networks. In ECCV, pages 303-319, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.512, + 0.907, + 0.579 + ], + "angle": 0, + "content": "[25] Juan Antonio Leñero-Bardallo, Teresa Serrano-Gotarredona, and Bernabé Linares-Barranco. A 3.6 \\(\\mu\\)s latency asynchronous frame-free event-driven dynamic-vision-sensor. IEEE Journal of Solid-State Circuits, 46(6):1443-1455, 2011. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.582, + 0.907, + 0.621 + ], + "angle": 0, + "content": "[26] Xinui Li, Chao Ma, Xiaokang Yang, and Ming-Hsuan Yang. VidToMe: Video token merging for zero-shot video editing. In CVPR, pages 7486-7495, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.624, + 0.907, + 0.676 + ], + "angle": 0, + "content": "[27] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A \\(128 \\times 128\\) 120 dB \\(15\\mu s\\) latency asynchronous temporal contrast vision sensor. IEEE Journal of Solid-State Circuits, 43(2):566-576, 2008. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.907, + 0.734 + ], + "angle": 0, + "content": "[28] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Yu Qiao, Wanli Ouyang, and Chao Dong. Diff-BIR: Toward blind image restoration with generative diffusion prior. In ECCV, pages 430-448. Springer, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.907, + 0.773 + ], + "angle": 0, + "content": "[29] Lin Liu, Xu Jia, Jianzhuang Liu, and Qi Tian. Joint demosaicing and denoising with self guidance. In CVPR, pages 2240-2249, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[30] Yunfan Lu, Yijie Xu, Wenzong Ma, Weiyu Guo, and Hui Xiong. Event camera demosaicing via swim transformer and pixel-focus loss. In CVPR, pages 1095-1105, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[31] Sizhuo Ma, Varun Sundar, Paul Mos, Claudio Bruschini, Edoardo Charbon, and Mohit Gupta. Seeing photons in color. ACM TOG, 42(4):1-16, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[32] Julien Mairal, Michael Elad, and Guillermo Sapiro. Sparse representation for color image restoration. IEEE TIP, 17(1): 53-69, 2007. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7913" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.135 + ], + "angle": 0, + "content": "[33] Julien Mairal, Francis Bach, Jean Ponce, Guillermo Sapiro, and Andrew Zisserman. Non-local sparse models for image restoration. In ICCV, pages 2272-2279, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.137, + 0.484, + 0.178 + ], + "angle": 0, + "content": "[34] Henrique S Malvar, Li-wei He, and Ross Cutler. High-quality linear interpolation for demosaicing of bayer-patterned color images. In ICASSP, pages iii-485, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.18, + 0.484, + 0.221 + ], + "angle": 0, + "content": "[35] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal Processing Letters, 20(3):209-212, 2012. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.224, + 0.484, + 0.263 + ], + "angle": 0, + "content": "[36] Ibrahim Pekkucuksen and Yucel Altunbasak. Gradient based threshold free color filter array interpolation. In ICIP, pages 137-140, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.267, + 0.484, + 0.295 + ], + "angle": 0, + "content": "[37] Reza Pournaghi and Xiaolin Wu. Coded acquisition of high frame rate video. IEEE TIP, 23(12):5670-5682, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.297, + 0.484, + 0.352 + ], + "angle": 0, + "content": "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.354, + 0.484, + 0.382 + ], + "angle": 0, + "content": "[39] Hyeonjun Sim, Jihyong Oh, and Munchurl Kim. XVFI: extreme video frame interpolation. In ICCV, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.384, + 0.484, + 0.438 + ], + "angle": 0, + "content": "[40] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.484, + 0.496 + ], + "angle": 0, + "content": "[41] Daniel Stanley Tan, Wei-Yang Chen, and Kai-Lung Hua. DeepDemosaicking: Adaptive image demosaicking via multiple deep fully convolutional networks. IEEE TIP, 27(5): 2408-2419, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.499, + 0.484, + 0.54 + ], + "angle": 0, + "content": "[42] Runjie Tan, Kai Zhang, Wangmeng Zuo, and Lei Zhang. Color image demosaicking via deep residual learning. In ICME, pages 793-798, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.543, + 0.484, + 0.582 + ], + "angle": 0, + "content": "[43] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-Shot Image Restoration Using Denoising Diffusion Null-Space Model. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.586, + 0.484, + 0.627 + ], + "angle": 0, + "content": "[44] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.63, + 0.484, + 0.671 + ], + "angle": 0, + "content": "[45] Bennett Wilburn, Neel Joshi, Vaibhav Vaish, Marc Levoy, and Mark Horowitz. High-speed videography using a dense camera array. CVPR, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.674, + 0.484, + 0.714 + ], + "angle": 0, + "content": "[46] Wenzhu Xing and Karen Egiazarian. End-to-end learning for joint image demosaicing, denoising and super-resolution. In CVPR, pages 3507-3516, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.717, + 0.484, + 0.771 + ], + "angle": 0, + "content": "[47] Senyan Xu, Zhijing Sun, Jiaying Zhu, Yurui Zhu, Xueyang Fu, and Zheng-Jun Zha. Demosaicformer: Coarse-to-fine demosaicing network for hybrids camera. In CVPR, pages 1126-1135, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.774, + 0.484, + 0.828 + ], + "angle": 0, + "content": "[48] Siqi Yang, Zhaojun Huang, Yakun Chang, Bin Fan, Zhaofei Yu, and Boxin Shi. Real-data-driven 2000 FPS color video from mosaicked chromatic spikes. In ECCV, 2024. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.484, + 0.899 + ], + "angle": 0, + "content": "[49] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling Up to Excellence: Practicing Model Scaling for Photo-Realistic Image Restoration In the Wild. In CVPR, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.147 + ], + "angle": 0, + "content": "[50] Guoshen Yu, Guillermo Sapiro, and Stephane Mallat. Solving inverse problems with piecewise linear estimators: From gaussian mixture models to structured sparsity. IEEE TIP, 21 (5):2481-2499, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.908, + 0.232 + ], + "angle": 0, + "content": "[51] Jie Zhang, Jonathan Newman, Zeguan Wang, Yong Qian, Pedro Feliciano-Ramos, Wei Guo, Takato Honda, Zhe Sage Chen, Changyang Linghu, Ralph Etienne-Cummings, et al. Pixel-wise programmability enables dynamic high-snr cameras for high-speed microscopy. Nature Communications, 15 (1):4480, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.907, + 0.274 + ], + "angle": 0, + "content": "[52] Lin Zhang, Lei Zhang, and Alan C Bovik. A feature-enriched completely blind image quality evaluator. IEEE TIP, 24(8): 2579-2591, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.907, + 0.33 + ], + "angle": 0, + "content": "[53] Zhao Zhang, Huan Zheng, Richang Hong, Mingliang Xu, Shuicheng Yan, and Meng Wang. Deep color consistent network for low-light image enhancement. In CVPR, pages 1899-1908, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.908, + 0.387 + ], + "angle": 0, + "content": "[54] Jing Zhao, Ruiqin Xiong, Hangfan Liu, Jian Zhang, and Tiejun Huang. Spk2Imgnet: Learning to reconstruct dynamic scene from continuous spike stream. In CVPR, pages 11996-12005, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.908, + 0.444 + ], + "angle": 0, + "content": "[55] Jing Zhao, Ruiqin Xiong, Jiyu Xie, Boxin Shi, Zhaofei Yu, Wen Gao, and Tiejun Huang. Reconstructing clear image for high-speed motion scene with a retina-inspired spike camera. IEEE TCI, 8:12-27, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.446, + 0.908, + 0.501 + ], + "angle": 0, + "content": "[56] Yajing Zheng, Lingxiao Zheng, Zhaofei Yu, Boxin Shi, Yonghong Tian, and Tiejun Huang. High-speed image reconstruction through short-term plasticity for spiking cameras. In CVPR, pages 6358-6367, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.503, + 0.908, + 0.544 + ], + "angle": 0, + "content": "[57] Lin Zhu, Siwei Dong, Tiejun Huang, and Yonghong Tian. A retina-inspired sampling method for visual texture reconstruction. In ICME, pages 1432-1437, 2019. 2, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.546, + 0.907, + 0.587 + ], + "angle": 0, + "content": "[58] Lin Zhu, Siwei Dong, Jianing Li, Tiejun Huang, and Yonghong Tian. Retina-like visual image reconstruction via spiking neural model. In CVPR, pages 1438-1446, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.587 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7914" + } + ] +] \ No newline at end of file diff --git a/2025/SpikeDiff_ Zero-shot High-Quality Video Reconstruction from Chromatic Spike Camera and Sub-millisecond Spike Streams/3f42c09f-75a5-4dd4-bd4a-8b0a20ea82c7_origin.pdf b/2025/SpikeDiff_ Zero-shot High-Quality Video Reconstruction from Chromatic Spike Camera and Sub-millisecond Spike Streams/3f42c09f-75a5-4dd4-bd4a-8b0a20ea82c7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8e478e9658919dab8d03faf988eb7b26fa863eae --- /dev/null +++ b/2025/SpikeDiff_ Zero-shot High-Quality Video Reconstruction from Chromatic Spike Camera and Sub-millisecond Spike Streams/3f42c09f-75a5-4dd4-bd4a-8b0a20ea82c7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c6b5bc179e9d136fce4a07446d8179c7699e7e84387d6616a92bb803414df624 +size 6291534 diff --git a/2025/SpikeDiff_ Zero-shot High-Quality Video Reconstruction from Chromatic Spike Camera and Sub-millisecond Spike Streams/full.md b/2025/SpikeDiff_ Zero-shot High-Quality Video Reconstruction from Chromatic Spike Camera and Sub-millisecond Spike Streams/full.md new file mode 100644 index 0000000000000000000000000000000000000000..19b523a23366bdb87087b9429fc7233e72660979 --- /dev/null +++ b/2025/SpikeDiff_ Zero-shot High-Quality Video Reconstruction from Chromatic Spike Camera and Sub-millisecond Spike Streams/full.md @@ -0,0 +1,360 @@ +# SpikeDiff: Zero-shot High-Quality Video Reconstruction from Chromatic Spike Camera and Sub-millisecondskip Spike Streams + +Siqi Yang $^{1,2,3}$ + +Jinxiu Liang $^{2,3,4*}$ +Yakun Chang $^{5,6}$ + +Zhaojun Huang2,3 Zhaofei Yu1,3 + +Yeliduosi Xiaokaiti2,3 + +Boxin Shi $^{2,3,1*}$ + +1 Institute for Artificial Intelligence, Peking University + +$^{2}$ State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University +$^{3}$ National Engineering Research Center of Visual Technology, School of Computer Science, Peking University + +$^{4}$ National Institute of Informatics + +$^{5}$ Institute of Information Science, Beijing Jiaotong University + +$^{6}$ Visual Intelligence +X International Cooperation Joint Laboratory of the Ministry of Education,.cssherryliang@gmail.com, shiboxin@pku.edu.cn + +![](images/bd5f75b0432674a69a21f655aa0ecc07b693c85e4cdff72f3ae80e2030a6484c.jpg) +Figure 1. Spike cameras capture chromatic spikes with high temporal resolution, enabling the observation of scenes with rapid motion. As illustrated in the orange region, the spike camera utilizes a color filter array to capture chromatic information and continuously accumulates photons until reaching a threshold that triggers a spike. To reconstruct high-quality videos from these chromatic spikes, we propose SpikeDiff, a zero-shot framework that leverages principled priors from pretrained diffusion models and integrates physics-based guidance derived from the spike camera mechanism. SpikeDiff generates visually appealing videos without noise or motion blur, significantly outperforming existing chromatic spike reconstruction methods (e.g., SJDD [11] and CSpkNet [12]). + +# Abstract + +High-speed video reconstruction from neuromorphic spike cameras offers a promising alternative to traditional frame-based imaging, providing superior temporal resolution and dynamic range with reduced power consumption. Nevertheless, reconstructing high-quality colored videos from spikes captured in ultra-short time intervals (sub-millisecond) remain challenging due to the inherently noisy nature of spikes. While some existing methods extend the temporal capture window to improve reconstruction quality, they inevitably compromise the temporal resolution advantages of spike cameras. In this paper, we introduce SpikeDiff, the first zero-shot framework that leverages pretrained diffusion models to reconstruct high-quality colored videos from sub-millisecond + +(0.5ms) chromatic spike streams. By incorporating physics-based guidance into the diffusion sampling process, SpikeDiff bridges the domain gap between chromatic spikes and conventional images, enabling high-fidelity reconstruction without requiring domain-specific training data. Extensive experiments demonstrate that SpikeDiff achieves impressive reconstruction quality while maintaining ultra-high temporal resolution, outperforming existing methods across diverse challenging scenarios in both perceptual quality and structural preservation. + +# 1. Introduction + +High-speed motions—from droplet impacts to mechanical dynamics—occur at speeds surpassing human perceptual processing capabilities. Conventional high-speed cameras are capable of capturing these motions but encounter funda- + +mental trade-offs [37, 45, 51]: higher temporal resolution requires shorter exposures, leading to reduced light capture and limited dynamic range, alongside increased power consumption and larger data volumes. These constraints make continuous high-speed capture over extended periods either impractical or impossible. + +Neuromorphic cameras have emerged as a compelling alternative to conventional high-speed imaging, including event cameras [9, 25, 27] and spike cameras [3, 21, 58]. Drawing inspiration from the fovea in biological visual systems, each pixel in a spike camera continuously accumulates photons, generates spikes upon reaching the predefined threshold, and subsequently proceeds to the next accumulation period. Operating at a sampling rate of $20,000\mathrm{Hz}$ with $1000\times 1000$ resolution, spike cameras encode visual information with low data overhead and power consumption, compared to conventional cameras with similar sampling rate and resolution. This enables continuously blur-free capture with high-speed motion and high contrast over several minutes that conventional sensors struggle to handle. + +For human perception, the binary representations of spikes are inherently difficult to interpret and require reconstruction into video frames [54, 55, 58]. However, reconstructing high-quality video frames from spikes remains challenging, particularly within ultra-short time intervals where limited spikes make existing methods highly susceptible to noise [57]. This challenge intensifies for color reconstruction, as Color Filter Arrays (CFAs) further reduce the number of available spikes within short time intervals. + +Existing approaches attempt to mitigate these challenges by extending the observed time intervals to accumulate more spikes [31, 56, 57]. This strategy inadvertently introduces motion blur and ghosting artifacts that compromise the temporal resolution advantage of spike cameras. While recent deep learning-based methods have shown promise in optimizing this trade-off problem, they typically exhibit limited generalization capabilities [10, 11], require extensive training data [12], or demand longer time intervals [48]. Yet, these methods still struggle to completely eliminate the noise [11] (upper right in Figure 1) or avoid introducing motion blur in high-speed motion regions [48] (lower right in Figure 1). + +The recent success of diffusion models in image and video restoration [15, 22, 43, 49] suggests a promising direction. These methods have demonstrated remarkable capability in encoding rich priors about natural scene statistics and temporal dynamics. This raises an intriguing question: Can we leverage these powerful priors to reconstruct high-quality frames and videos from extremely noisy chromatic spikes captured within sub-millisecond $^{1}$ time intervals? + +Integrating pretrained diffusion models with spike-to + +video reconstruction presents fundamental challenges: Spike cameras capture visual information through asynchronous voltage accumulation and threshold-based sampling—a process fundamentally different from the fixed-interval capturing mechanism of frame-based cameras. Furthermore, chromatic spikes represent raw sensor measurements without established signal processing pipelines, while pretrained diffusion models are typically trained on processed sRGB frames with optimized color and tone characteristics. These disparities create a significant domain gap that must be addressed for effective reconstruction. + +In this paper, we propose SpikeDiff, the first zero-shot spike-to-video reconstruction framework that leverages pretrained diffusion models as principled priors for high-quality colored video reconstruction from chromatic spikes. By establishing a physically grounded likelihood model of spike measurements throughout the reverse diffusion process, SpikeDiff approximates the posterior sampling of high-quality videos from sub-milliseconds spike streams. Unlike existing approaches that require extensive domain-specific training data and risk overfitting, our method operates in a zero-shot manner, eliminating both the computational burden of training and the need for large-scale datasets. + +Our key contributions include: + +- the first framework that leverages pretrained diffusion models as powerful priors for high-quality, zero-shot video reconstruction from chromatic spikes in ultra-short intervals, eliminating the need for domain-specific training data; and +- a physics-based and differentiable formulation of the spike generation process, which enables effective integration with pretrained diffusion models, bridging the domain gap between spikes and conventional images. + +Extensive experiments across diverse challenging scenarios demonstrate that SpikeDiff consistently outperforms prior methods that rely on extended time intervals and compromising temporal resolution, effectively breaking the traditional quality-time resolution trade-off in spike-to-video reconstruction (see our result shown in Figure 1). + +# 2. Related work + +Video reconstruction with monochromatic spikes. Reconstructing high-speed videos from spikes typically employs the imaging model of spike cameras [21]. By accumulating the spikes triggered within a certain time window or estimating the interval between successive spikes, textures can be reconstructed [57, 58]. However, due to the limited number of photons arriving at the pixel in an ultra-short time, such straightforward reconstruction suffers from significant noise. To address this issue, Chang et al. [4] propose enhancing the signal-to-noise ratio (SNR) by combining multi-bit spikes with binary spikes in a rolling-mixed-bit manner. However, this solution requires hardware modifications, limiting its practicality in more general scenes. Zhao et al. [54] present + +Spk2ImgNet, a hierarchical architecture that progressively fuses the spikes, offering an alternative approach to improve reconstruction quality. However, its generalization ability is limited since it relies on synthetic data for training. To address the issue of real-world ground truth for training, self-supervised methods [5, 6] are developed to reduce dependence on synthetic datasets. However, it is still challenging to reconstruct color videos from monochromatic spikes. From the perspective of incorporating additional color guidance, Chang et al. [3] develop a hybrid spike-RGB camera system that performs spatial alignment and frame interpolation simultaneously, enabling the recovery of 1000 FPS color video. However, this hybrid camera system necessitates synchronization and optical alignment, while the beam splitter also poses challenges for constructing a compact device. + +Demosaicking and denoising for unconventional cameras. Various methods for joint demosaicking and denoising have been proposed to obtain high-resolution and low-noise color images from frame-based RAW data, ranging from traditional model-based techniques [1, 8, 19, 23, 32-34, 36, 50] to more recent data-driven approaches [18, 24, 29, 41, 42, 46, 53]. However, all these methods cannot be applied to binary data captured by the new-generation high-speed cameras, such as event cameras [9, 25, 27], quanta image sensors [16, 17], and spike cameras [21]. In the case of event cameras, Xu et al. [47] and Lu et al. [30] present Transformer-based architectures for demosaicking missing pixel values in RAW domain processing. However, due to the sparsity of events, these approaches still rely on frame-based cameras for texture reconstruction. In the context of quanta image sensors, Elgendy et al. [14] integrate a frequency selection method with a deep neural network based filtering approach, leveraging the luma channels to assist in the denoising of the chroma channels. Ma et al. [31] design a blue-noise pseudo-random RGBW color filter array and successfully reconstruct high-quality color images from mosaicked single-bit frames, even in high-dynamic-range (HDR) scenes with complex and rapid motion. For spike cameras, Dong et al. [11] propose an offset-sharing deformable convolution module to align temporal features of color channels and develop a spike noise estimator to characterize the noise distribution. However, this method is trained on synthetic data, limiting its generalization to real-world data. Yang et al. [48] introduce a self-supervised denoising module trained exclusively on real-world chromatic spikes, achieving 2000 FPS color video reconstruction. However, there remains significant research potential in balancing noise suppression with detail preservation. + +# 3. Method + +We present the degradation model of chromatic spikes based on the physical formulation and propose SpikeDiff, a zero-shot framework for reconstructing consecutive high-quality + +frames from chromatic spikes. This approach leverages pretrained diffusion models as principled priors to effectively address the inherent ill-posedness of the problem. + +# 3.1. Preliminaries + +Chromatic spikes. The chromatic spike camera leverages a spike sensor equipped with a Bayer-pattern (RGGB) CFA to capture the scene with chromatic information. As shown in Figure 1, when photons filtered by the CFA reach a pixel, the electrons generated by these photons are continuously accumulated as long as the voltage does not reach the threshold $E_{\mathrm{th}}$ . Simultaneously, the readout circuit samples the pixel values at a frequency of $20,000\mathrm{Hz}$ and a signal of 0 is read out at each readout point. Once the accumulated voltage reaches $E_{\mathrm{th}}$ , a signal of 1 is read out and the voltage of this pixel is reset to continue the next accumulation period. For each pixel $i$ , we denote the accumulated voltage at a readout point $\tau$ as $E(i,\tau)$ , the triggered spike at $\tau$ is: + +$$ +S (i, \tau) = \left\{ \begin{array}{l l} 1, & \text {i f} E (i, \tau) \geq E _ {\mathrm {t h}}, \\ 0, & \text {o t h e r w i s e .} \end{array} \right. \tag {1} +$$ + +Diffusion models. Latent diffusion models (LDMs) operate in a compressed latent space learned by an autoencoder pair $(\mathcal{E},\mathcal{D})$ , where $X = \mathcal{D}(Z)$ and $Z = \mathcal{E}(X)$ . An LDM learns to reverse a diffusion process that gradually adds Gaussian noise to a latent vector $Z_{0}$ . The reverse process, modeled by a stochastic differential equation (SDE), generates a clean latent $Z_{0}$ from pure noise $Z_{T} \sim \mathcal{N}(0,I)$ [40]: + +$$ +d \boldsymbol {Z} = [ - f (\boldsymbol {Z}, t) - g ^ {2} (t) \nabla_ {\boldsymbol {Z} _ {t}} \log p _ {t} (\boldsymbol {Z} _ {t}) ] d t + g (t) d w, (2) +$$ + +where $f(\cdot, t)$ and $g(t)$ are drift and diffusion coefficients, $w$ represents the standard Brownian motion. The key component is the score function $\nabla_{\mathbf{Z}_t} \log p_t(\mathbf{Z}_t)$ , which is approximated by a time-conditioned neural network $\epsilon_{\theta}(\mathbf{Z}, t)$ . By iteratively applying the learned score, the reverse SDE can sample from the learned data prior $p(\mathbf{Z}_0)$ to generate a high-quality image $\mathbf{X}_0 = \mathcal{D}(\mathbf{Z}_0)$ . + +# 3.2. Diffusion-based reconstruction from spikes + +Reconstructing consecutive high-quality frames from real-captured chromatic spike streams is challenging due to the non-negligible noise perturbations, particularly within extremely limited time intervals (e.g., sub-milisecond). To address this challenge, we leverage recent advancements in diffusion models, which incorporate learned priors regarding the distribution of high-quality images, and formulate the reconstruction of video frames from chromatic spikes as a Maximum a Posteriori (MAP) estimation problem: + +$$ +\begin{array}{l} \boldsymbol {X} ^ {*} = \operatorname {a r g m a x} _ {\boldsymbol {X}} p (\boldsymbol {X} | \boldsymbol {Y}) \tag {3} \\ = \operatorname {a r g m a x} _ {\boldsymbol {X}} p (\boldsymbol {Y} | \boldsymbol {X}) p (\boldsymbol {X}), \\ \end{array} +$$ + +![](images/6c3b2d7d61d13e1be9619b110693938f4e7b9a9f2d5aa5f17f87143f192733d2.jpg) +Figure 2. Pipeline of SpikeDiff. Our diffusion-based posterior sampling procedure follows the iterative denoising process of diffusion models, starting from the initialization $\mathbf{Z}_{t_s}$ and gradually approaching the desired latent $\mathbf{Z}_0$ , indicated by the dashed lines. To maximize the likelihood of the reconstructed video given the spike observations $\mathbf{Y}$ , we enhance each iteration with physics-based spike guidance, which comprises three specially-designed differentiable degradation operators: mosaicking $M$ , color casting $C$ , and quantization $Q$ . Finally, a multiscale loss function bridges the gap between degraded frames $\mathcal{A}(\mathbf{X}_{0|t})$ and observed SFR frames $\mathbf{Y}$ . As shown in the top sequence, the visual quality of generated frames progressively improves until convergence. Note that the internal degraded frames and the SFR frames are all Bayer-patterned mosaic images; we present their corresponding demosaicked results for better visualization. + +where the prior $p(\mathbf{X})$ encodes knowledge of natural images, and the likelihood $p(\mathbf{Y}|\mathbf{X})$ represents the data fidelity to ensure the reconstruction is faithful to the observed chromatic spikes $\mathbf{Y}$ . + +The degradation model, mapping light intensity to spike signals (illustrated in Figure 1), establishes the connection from RGB frames $X$ — the modality employed in pretrained diffusion models—and chromatic spikes $Y$ , and thereby deriving likelihood $p(\mathbf{Y}|\mathbf{X})$ illustrated in the next (Sec. 3.3). This establishes a principled bridge that enables the reconstruction of $X$ from $Y$ via posterior sampling in the MAP framework, as illustrated in the green box in Figure 2, where diffusion models serve as learned priors for $X$ while the likelihood ensures fidelity to $Y$ . This is exactly the key in our method to connect the modality used in pretrained diffusion models, RGB frames, and chromatic spikes. + +Given the significant disparity between chromatic spike streams and conventional images, the data fidelity term $p(\mathbf{Y}|\mathbf{X})$ cannot be directly derived from time-continuous chromatic spikes. Instead, we resort to a more compact representation. According to the integrating and firing mechanism of spikes, the light intensity is proportional to the spike firing rate (SFR) [57]. Without considering noise, the SFR at $\tau'$ can be counted through collecting a number of spikes within + +a predefined time interval of size $\Delta \tau$ + +$$ +Y (i, \tau^ {\prime}) = \frac {1}{\Delta \tau} \cdot \sum_ {\tau = \tau^ {\prime} - \Delta \tau / 2} ^ {\tau^ {\prime} + \Delta \tau / 2} S (i, \tau). \tag {4} +$$ + +By collecting the SFRs from the pixel array at $\tau'$ , we obtain mosaicked SFR frames $Y \in \mathbb{R}^{T \times H \times W}$ , where $T$ denotes the number of frames, and $H, W$ represent the height and width, respectively. The SFR frames, estimated from real-captured chromatic spikes, are sensitive to the inherent noise, especially within limited time intervals, and exhibit misalignment with corresponding consecutive high-quality frames. Consequently, we propose a corresponding degradation model, comprising differentiable operators, as detailed in Sec. 3.3, to bridge the gap between $X$ and $Y$ and resolve the likelihood estimation required for posterior sampling, as illustrated in the orange box in Figure 2. + +Posterior sampling. The primary challenge in reconstructing consecutive high-quality frames from sub-millisecond chromatic spikes lies in its inherent ill-posedness, making the prior term $p(\mathbf{X})$ crucial. We leverage diffusion models as powerful priors. Given the prior distribution of natural images $p(\mathbf{Z}_t)$ encoded by the pretrained latent diffusion model in Eq. (2) and likelihood defined in the next, we solve the MAP problem in Eq. (3) by sampling from the posterior + +distribution $p(\mathbf{X}|\mathbf{Y})$ . This is achieved by modifying the reverse SDE of the diffusion model to incorporate guidance from our observations. Using Bayes' rule, the score of the posterior distribution $p_t(\mathbf{Z}_t|\mathbf{Y})$ can be written as: + +$$ +\nabla_ {\boldsymbol {Z} _ {t}} \log p _ {t} (\boldsymbol {Z} _ {t} | \boldsymbol {Y}) = \nabla_ {\boldsymbol {Z} _ {t}} \log p (\boldsymbol {Z} _ {t}) + \nabla_ {\boldsymbol {Z} _ {t}} \log p (\boldsymbol {Y} | \boldsymbol {Z} _ {t}), \tag {5} +$$ + +where the prior term $\nabla_{\mathbf{Z}_t}\log p(\mathbf{Z}_t)$ is given by the pretrained network $\epsilon_{\theta^*}(Z_t,t)$ . The likelihood term $p(\mathbf{Y}|\mathbf{Z}_t)$ guides the sampling towards latents $\mathbf{Z}_t$ that are likely to produce our observed SFR frame $\mathbf{Y}$ . The full reverse SDE for posterior sampling is: + +$$ +\begin{array}{l} d \boldsymbol {Z} = [ - f (\boldsymbol {Z}, t) - g ^ {2} (t) (\nabla_ {\boldsymbol {Z} _ {t}} \log p (\boldsymbol {Z} _ {t}) \tag {6} \\ + \nabla_ {\boldsymbol {Z} _ {t}} \log p (\boldsymbol {Y} | \boldsymbol {Z} _ {t})) ] d t + g (t) d w. \\ \end{array} +$$ + +To compute the likelihood score, we must relate the noisy latent $Z_{t}$ at an intermediate timestep $t$ to the observation $Y$ . This involves approximating the conditional probability $p(\mathbf{Y}|\mathbf{Z}_t)$ , which can be expressed as: + +$$ +\begin{array}{l} p (\mathbf {Y} \mid \mathbf {Z} _ {t}) = \int_ {c} p (\mathbf {Y} \mid \mathbf {Z} _ {0}, \mathbf {Z} _ {t}) p (\mathbf {Z} _ {0} \mid \mathbf {Z} _ {t}) d \mathbf {Z} _ {0} \tag {7} \\ = \int p (\boldsymbol {Y} | \boldsymbol {Z} _ {0}) p (\boldsymbol {Z} _ {0} | \boldsymbol {Z} _ {t}) d \boldsymbol {Z} _ {0}. \\ \end{array} +$$ + +This integral is intractable. Instead, we approximate it by first estimating the expected clean latent $\mathbf{Z}_{0|t} = \mathbb{E}[Z_0|\mathbf{Z}_t]$ from $\mathbf{Z}_t$ . For diffusion models like DDPM [20], the forward process is + +$$ +\boldsymbol {Z} _ {t} = \sqrt {\bar {\alpha} _ {t}} \boldsymbol {Z} _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} z, \quad z \sim \mathcal {N} (0, I), \tag {8} +$$ + +where $\bar{\alpha}_t$ is the noise schedule parameter from the diffusion model (e.g., DDPM [20]). Using Tweedie's formula [7], we can estimate the conditional mean: + +$$ +\begin{array}{l} \boldsymbol {Z} _ {0 | t} = \mathbb {E} [ \boldsymbol {Z} _ {0} | \boldsymbol {Z} _ {t} ] \\ = \frac {1}{\sqrt {\bar {\alpha} _ {t}}} \left(\boldsymbol {Z} _ {t} + \left(1 - \bar {\alpha} _ {t}\right) \nabla_ {\boldsymbol {Z} _ {t}} \log p _ {t} \left(\boldsymbol {Z} _ {t}\right)\right). \tag {9} \\ \end{array} +$$ + +Here, we replace the true score $\nabla_{Z_t}\log p_t(Z_t)$ with its neural network approximation $\epsilon_{\theta^*}(Z_t,t)$ . With this estimate of the clean latent, $Z_{0|t}$ , following the expression $p(\mathbf{Y}|\mathbf{Z}_t) = \mathbb{E}_{\mathbf{Z}_0\sim p(\mathbf{Z}_0|\mathbf{Z}_t)}[p(\mathbf{Y}|\mathcal{D}(\mathbf{Z}_0))]$ , we can use a surrogate likelihood approach [7] to make the likelihood tractable: + +$$ +p (\boldsymbol {Y} | \boldsymbol {Z} _ {t}) \approx p (\boldsymbol {Y} | \mathcal {D} (\mathbb {E} [ \boldsymbol {Z} _ {0} | \boldsymbol {Z} _ {t} ])). \tag {10} +$$ + +This formulation provides a practical way to compute the likelihood gradient. + +# 3.3. Chromatic spikes' likelihood estimation + +Given a set of sequential SFR frames, denoted by $\mathbf{Y}$ , our objective is to establish a connection between the SFR frames + +and the desired target reconstructions $\mathbf{X}$ , with both $\mathbf{Y}$ and $\mathbf{X}$ normalized to the range [0, 1]. Hence, we resolve the likelihood estimation of $p(\mathbf{Y} | Z_t)$ in Eq. (10) and thereby complete the diffusion-based posterior sampling strategy. + +Degradation model of chromatic spikes. The physical process generating SFR frames $\mathbf{Y}$ can be modeled as a nonlinear inverse problem: $\mathbf{Y} = \mathcal{A}(\mathbf{X}) + \mathbf{N}$ , where $\mathcal{A}(\cdot)$ denotes the known forward operator modeling the degradation process, and $\mathbf{N}$ represents stochastic noise. This problem is inherently ill-posed due to irreversible information loss during the imaging process and interference from noise in the capture system. + +The forward operator $\mathcal{A}(\cdot)$ in the imaging pipeline transforms the latent high-quality frame $X$ through a sequence of operations: $\mathcal{A}(X) = \mathcal{Q}(CMX)$ , where $\mathcal{Q}(\cdot)$ represents pixel-wise quantization, $C$ implements pixel-wise color casting with per-channel scaling, and $M: \mathbb{R}^{T \times H \times W \times 3} \to \mathbb{R}^{T \times H \times W}$ denotes the mosaicking operator, modeling the color filter array pattern. This decomposition explicitly models each step of the image formation process, providing a foundation for our reconstruction approach. + +Under the assumption of Gaussian noise, we model the likelihood $p(\mathbf{Y}|\mathcal{D}(\mathbf{Z}_{0|t}))$ of observing SFR frames $\mathbf{Y}$ given the target frame $\mathbf{X}$ as a Gaussian distribution. Maximizing this likelihood translates to minimizing the following loss: + +$$ +\mathcal {L} \left(\boldsymbol {Z} _ {t}\right) = \frac {1}{\sigma^ {2}} \| \boldsymbol {Y} - \mathcal {A} \left(\mathcal {D} \left(\boldsymbol {Z} _ {0 | t}\right)\right) \| _ {2} ^ {2}, \tag {11} +$$ + +where $\sigma$ represents the standard deviation of the noise $N$ , and $\mathcal{A}$ represents the differentiable degradation operators defined later. By substituting the score function into Eq. (5), we obtain: + +$$ +\nabla_ {\boldsymbol {Z} _ {t}} \log p _ {t} (\boldsymbol {Z} _ {t} | \boldsymbol {Y}) = \epsilon_ {\theta^ {*}} (\boldsymbol {Z} _ {t}, t) - s \nabla_ {\boldsymbol {Z} _ {t}} \mathcal {L} (\boldsymbol {Z} _ {t}), \tag {12} +$$ + +where $s$ controls the strength of measurement guidance. + +Differentiable degradation operators. A key innovation in SpikeDiff is our design of differentiable operators in $\mathcal{A}$ that model the physical characteristics of chromatic spike cameras, which include color casting $C$ , quantization $\mathcal{Q}$ , and mosaicking $M$ . Based on the gray-world prior [2] for the target frame, which indicates that each color channel should have a mean value close to one another, we estimate the color casting coefficients $\alpha_{R}, \alpha_{G}, \alpha_{B}$ as $\alpha_{c} = \frac{1}{\mu(\mathbf{Y}_{c})}$ , where $c \in \{\mathrm{R}, \mathrm{G}, \mathrm{B}\}$ denotes the color channel, $\mu(\mathbf{Y}_{c})$ is the mean value of color channel $c$ among $\mathbf{Y}$ across the temporal axis. Then the channel-wise scaling factors for the color casting operator $C$ can be computed as $\eta / \alpha_{c}$ , where parameter $\eta$ is used to adjust the value range. + +SFR frames are inherently quantized and discrete as they are accumulation of binary spike frames, which differ from + +8-bit frames $\mathbf{X}$ . To simulate the quantization effect and maintain the backward differentiability, we design a soft quantization operator $\mathcal{Q}$ . Given the expected quantization levels determined by SFR estimation, the soft quantization function $q(\cdot)$ operates on each pixel individually: + +$$ +q (x) = l (x) \cdot (1 - \varphi (x)) + u (x) \cdot \varphi (x), \tag {13} +$$ + +$$ +\varphi (x) = \left[ \tanh \left(k \left(x - \frac {l (x) + u (x)}{2}\right)\right) + 1 \right] / 2, \tag {14} +$$ + +where $l(x)$ and $u(x)$ denote the lower and upper quantization levels of $x$ , $\varphi(\cdot)$ quantizes the input value to the range [0, 1] with a smooth curve, and hyperparameter $k$ controls the smoothness, as shown in Figure 3(a). + +The naive Bayer-pattern mosaicking operator $M$ converts target frame from dimension $T \times H \times W \times 3$ to $T \times \frac{H}{2} \times \frac{W}{2} \times 4$ , resulting in a partially defined backpropagation function due to its gradient discontinuity. To simulate the Bayer-pattern CFA while maintaining a dense gradient flow, we extend the operator's derivative with linear interpolation convolution, similar to demosaicking algorithms. + +By composing these operators, we conclude a differentiable simulation from consecutive frames to SFR frames, enabling optimization in the posterior sampling process. + +Multiscale enhancement. To preserve fine spatial details, we incorporate a Laplacian pyramid $\mathcal{P}(\boldsymbol{X})$ that guides the diffusion process across multiple scales. The Laplacian pyramid decomposes the generated consecutive frames into a multiscale representation of frequency components, facilitating the preservation of fine details critical for high-quality output. The Laplacian pyramid is calculated by subtracting the upsampled version of the downsampled frames from the original frames: + +$$ +\mathcal {P} (\boldsymbol {X}) = \left\{\boldsymbol {X} _ {l} - \mathbf {U} \left(\mathbf {D} \left(\boldsymbol {X} _ {l}\right)\right) \right\} _ {l = 1} ^ {L}, \boldsymbol {X} _ {l} = \mathbf {U} \left(\mathbf {D} \left(\boldsymbol {X} _ {l - 1}\right)\right) \tag {15} +$$ + +where $\mathbf{D}$ and $\mathbf{U}$ are the upsampling operator and downsampling operator, respectively, and $L$ is the number of pyramid levels. This multiscale decomposition effectively guides the diffusion process across different frequency bands, ensuring preservation of high-frequency details in reconstruction. + +# 4. Experiments + +# 4.1. Implementation details + +Quantization levels. As introduced in Section 3.2, SFR frames are accumulations over predefined time intervals $\Delta \tau$ . Given that the spikes are binary, the value range of SFR frames is restricted to $\left\{\frac{u}{\Delta\tau}|u\in \mathbb{N}\cap [0,\Delta \tau ]\right\}$ . For instance, we set $\Delta \tau = 10$ in our experiments to achieve sub-milliseconds reconstruction, and the corresponding quantization levels are $0,0.1,0.2,\ldots ,0.9,1$ , leading to the soft quantization function visualized in Figure 3(a). + +![](images/294efd5fab5154c5cd2f7b37ed10e6f63bedcb3f4471f9f25df13b3f3d8c1a99.jpg) +(a) Visualization of $q(x)$ + +![](images/8a13751b2479c2ec930beb8a8abf065e0e12c6360cbc8ce480ffd837c86f4001.jpg) +Figure 3. Visualization of the quantization operator. (a) The soft quantization function $q(x)$ for SFR frames with $\Delta \tau = 10$ . (b) Detailed demonstration of quantization effectiveness on continuous pixel values. + +![](images/b7a0dbe66ff53b0a6e5a6760042b380572b8ce57fe8f7062441cc8b590c11026.jpg) +(b) Quantization effectiveness + +![](images/44762d33f751f5cbb9128a1d90d39ceef153c6e51f29f3a6e68fe6543368fd46.jpg) + +Color casting estimation. To obtain better color casting coefficients and maintain consistency across video frames reconstruction, we average the estimations of $\alpha_{c}$ from all SFR frames in the sequence, and apply the same $\alpha_{c}$ throughout the video reconstruction. + +Pretrained diffusion model. We implement SpikeDiff upon the pretrained weights from [28], which is adapted from Stable Diffusion v2.1 [38], selected for their strong priors and close alignment with our task. We extend the image-to-image pipeline to spikes-to-video by dividing SFR frames into small chunks, and achieve temporal consistency by leveraging high-fidelity reconstruction from time-continuous chromatic spikes with physics-based guidance. + +# 4.2. Evaluation dataset + +As our proposed method leverages pretrained diffusion models, which contain principled priors of the real world images and videos, we qualitatively evaluate the performance of our method on a real-world dataset. To this end, we collect a series of real chromatic spikes within the spike camera. The spike camera is equipped with a Bayer-pattern color filter array and a $1000 \times 1000$ pixel spike sensor. Limited by the generation ability of pretrained diffusion models and resolution of compared methods [48], we spatially crop the chromatic spikes into a lower resolution of $512 \times 512$ , maintaining the same Bayer-pattern (RGGB). The collected dataset contains chromatic spikes from various scenes with different levels of motion and color variation. We would like to contribute this dataset to the community for further research. + +Despite the real-captured chromatic spikes, we also employ the chromatic spikes simulator described in CSpKNet [12] and SY24 [48] to generate synthetic chromatic spikes from real-world high-frame-rate videos. We collect over 100 dynamic scenarios from X4K1000FPS [39] dataset, and generate the corresponding chromatic spikes with the simulator. The synthetic dataset is used to conduct the quantitative evaluation of our proposed method, which further demonstrates the superiority of SpikeDiff. + +# 4.3. Qualitative evaluation + +To demonstrate the superiority of our proposed method, we conduct the qualitative evaluation on chromatic spikes captured from the real world, comparing SpikeDiff with exist + +![](images/35a9e1d5b4299d7c0661ad2935794cf35d3ba972181a0be2c42d6fa62fcfe2ee.jpg) +(a) +Figure 4. Qualitative comparison on real-world captured chromatic spikes (a-c), and synthetic chromatic spikes simulated from real-world high-frame-rate videos (d). Compared methods include TFI [57], TFP [57], 3DRI [10], SJDD [11], CSpkNet [12], and SY24 [48]. Details in red / blue bounding boxes are shown on the top. All results are reconstructed from sub-millisecond spike streams (0.5ms), where most methods suffer from severe noise or extreme degradation while SpikeDiff recovers clean and visually pleasing results. + +ing chromatic spike reconstruction methods. Our proposed method can recover high-quality color frames from sub-milisecond time intervals (10 spike frames, corresponding to $0.5\mathrm{ms}$ ), which none of the existing neural network based method achieve. We adapt and evaluate the existing chromatic spike reconstruction methods (3DRI [10], SJDD [11], CSpkNet [12], and SY24 [48]), combining the SFR estimation methods (TFI [57], TFP [57]) with demosaicking as baselines. We conduct all experiments on sub-millisecond time intervals (0.5ms) to maintain the fairness in comparison. As shown in Figure 4(a), where a fan is fast rotating in front of the checkerboard, SpikeDiff produces the most clean and visually pleasant results. In contrast, other methods, except CSpkNet and SY24, contain significantly noisy pixels as highlighted in the red and blue bounding boxes. The results + +of CSpkNet and SY24 contain blurry or degraded artifacts. As the time intervals are strictly limited, all methods are free of motion blur naturally. As for Figure 4(b), we capture a static scene in front of the lamp with the spike camera slightly shaking, to study the reconstruction quality under hard light. Compared to others, SpikeDiff generates the most clear and sharp details in blue bounding box. And SpikeDiff produces a smoother texture under the direct illumination in the red bounding box, while most of other methods show obvious quantized artifacts. In Figure 4(c), we leverage the spike streams captured by Yang et al. [48], where SpikeDiff outperforms all other methods in both noise suppression and visual quality. We also illustrate a synthetic scene in Figure 4(d), which further demonstrates the superior reconstruction capability of SpikeDiff on chromatic spikes + +Table 1. Quantitative comparison of the proposed method with existing chromatic spike reconstruction methods. The best and second-best results are highlighted in red and blue, respectively. + +
MethodPSNR↑SSIM↑FID↓NIQE↓IL-NIQE↓
SpikeDiff (Ours)18.6940.7502.8805.17338.535
SY24 [48]13.0130.53621.47211.59278.672
SJDD [11]11.3580.32526.00911.27844.711
3DRI [10]18.5120.39016.0368.54141.816
CSpkNet [12]13.3640.6414.6995.81939.948
TFP [57]12.9090.5593.25612.00164.554
TFI [57]14.6400.5527.55812.51855.758
+ +Table 2. Temporal consistency evaluation, reporting Frame CLIP Score (Frame C.S.), Interpolation Error (Inp. Err.), and Interpolation PSNR (Inp. PSNR) following the metrics from [26]. The best and second-best are highlighted in red and blue, respectively. + +
MethodFrame C.S.↑Inp. Err.↓Inp. PSNR↑
SpikeDiff (Ours)0.9690.05725.45
SY24 [48]0.9630.04128.65
SJDD [11]0.9620.11319.34
3DRI [10]0.9650.10419.87
CSpkNet [12]0.9660.05925.27
+ +![](images/01d3436302988abaf90e5cca39c257072453296c5fa7262950208be221aaf4ea.jpg) +Grad. w/o +Mosaicking +Figure 5. Analysis of differentiable degradation operators. Generated frames receive sparse gradient from chromatic spikes without mosaicking operator. Similarly, reconstructions exhibit incorrect color distribution without the color casting operator. + +![](images/93bc0d0cc1c89a08dca55893b0206a1e132c70b73e9017b6b578760ae3625407.jpg) +Grad.w/ +Mosaicking + +![](images/f66e8b4203c1e4768b364dda0afe21e8ccf79e0a18de712f0ea604271680de63.jpg) +Result w/o +Color casting + +![](images/39507c09ee4651ee4e0474177738c2f52a4e323ed4bd46a038f3234037f0c856.jpg) +Result w/ +Color casting + +simulated from real-world videos. Note that post-processing (e.g., color casting) cannot eliminate noise or artifacts from other methods' results. Please refer to our supplementary material for more comparison and detailed discussion. + +# 4.4. Quantitative evaluation + +As shown in Table 1, we quantitatively evaluate the performance of SpikeDiff and existing chromatic spike reconstruction methods on the synthetic dataset. Similar to our qualitative evaluation, we evaluate these methods on submillisecond time intervals (0.5ms). Our method achieves the best performance across all the metrics, including PSNR, SSIM [44], FID [13], NIQE [35], and IL-NIQE [52], which demonstrates the superiority of SpikeDiff, especially in terms of perceptual metrics that reflect visual quality. As shown in Table 2, we also evaluate the temporal consistency metrics of SpikeDiff and existing reconstruction methods, where SpikeDiff achieves the best or comparable performance. + +# 4.5. Ablation study + +To investigate the effectiveness of our proposed degradation operators, we visualize the internal results after applying + +Table 3. Quantitative ablation study on the contribution of different degradation operators to reconstruction quality. + +
MethodPSNR↑SSIM↑FID↓NIQE↓IL-NIQE↓
SpikeDiff18.6940.7502.8805.17338.535
w/o C16.7520.72912.0646.07244.661
w/o M17.4810.7234.3926.20341.356
w/o Q17.4600.7173.9386.31840.487
+ +each operator. The mosaicking operator $M$ converts the generated frames into Bayer-pattern images, which is the beginning step. As shown in Figure 5, without the specific backward formulation for $M$ , the gradient map is not well-defined. The color casting operator $C$ is responsible for color correction, which degrades the normal colored image to a greener one (which is more close to the SFR frames) as shown in Figure 5. For the soft quantization operator $\mathcal{Q}$ , its effectiveness is demonstrated in Figure 3(b), quantizing the continuous white pixels into discrete levels, similar to the pattern of the SFR frames. As shown in Table 3, we further demonstrate the effectiveness of each degradation operator via quantitative evaluation. + +# 5. Conclusion + +We propose the first zero-shot chromatic spike reconstruction method, named SpikeDiff, to recover consecutive clean and high-quality colored frames from sub-millisecond spikes. SpikeDiff leverages the deep priors from pretrained diffusion models to address the extreme noise under very limited time intervals and generate visually pleasant results. SpikeDiff integrates the physics-based chromatic spikes' likelihood into the diffusion-based posterior sampling process, which bridges the domain gap between spikes and conventional images. We conduct experiments on both synthetic and real-captured data to demonstrate the superior performance of SpikeDiff over existing reconstruction methods. + +Limitations and future work. As shown in Figure 4(b), there exist "overexposure" cases in the central region under extreme direct illumination. This imperfection is due to the predefined spike firing threshold, resulting in consistent spikes among these pixels. We observed that it is hard to generate ideal details without any explicit guidance. In future research, we plan to explore more degradation model designs and diffusion techniques to handle this issue and further extend the HDR capability of chromatic spike reconstruction. + +# Acknowledgement + +This work is supported by National Natural Science Foundation of China (Grant No. 62302019, 62088102, 62136001), Beijing Natural Science Foundation (Grant No. L233024), and Beijing Municipal Science & Technology Commission, Administrative Commission of Zhongguancun Science Park (Grant No. Z241100003524012). PKU-affiliated authors thank openbayes.com for providing computing resources. + +# References + +[1] Antoni Buades, Bartomeu Coll, Jean-Michel Morel, and Catalina Sbert. Self-similarity driven color demosaicking. IEEE TIP, 18(6):1192-1202, 2009. 3 +[2] Gershon Buchsbaum. A spatial processor model for object colour perception. Journal of the Franklin institute, 310(1): 1-26, 1980. 5 +[3] Yakun Chang, Chu Zhou, Yuchen Hong, Liwen Hu, Chao Xu, Tiejun Huang, and Boxin Shi. 1000 FPS HDR video with a Spike-RGB hybrid camera. In CVPR, pages 22180-22190, 2023. 2, 3 +[4] Yakun Chang, Yeliduosi Xiaokaiti, Yujia Liu, Bin Fan, Zhaojun Huang, Tiejun Huang, and Boxin Shi. Towards HDR and HFR video from rolling-mixed-bit spikings. In CVPR, pages 25117-25127, 2024. 2 +[5] Shiyan Chen, Chaoteng Duan, Zhaofei Yu, Ruiqin Xiong, and Tiejun Huang. Self-supervised mutual learning for dynamic scene reconstruction of spiking camera. In IJCAI, page 2859-2866, 2022. 3 +[6] Shiyan Chen, Zhaofei Yu, and Tiejun Huang. Self-supervised joint dynamic scene reconstruction and optical flow estimation for spiking camera. In AAAI, pages 350-358, 2023. 3 +[7] Hyungjin Chung, Jeongsol Kim, Michael Thompson McCann, Marc Louis Klasky, and Jong Chul Ye. Diffusion Posterior Sampling for General Noisy Inverse Problems. In ICLR, 2023. 5 +[8] Laurent Condat and Saleh Mosaddegh. Joint demosaicking and denoising by total variation minimization. In ICIP, pages 2781-2784, 2012. 3 +[9] Tobi Delbruck et al. Frame-free dynamic digital vision. In Proc. of Intl. Symp. on Secure-Life Electronics, Advanced Electronics for Quality Life and Society, pages 21-26. Cite-seer, 2008. 2, 3 +[10] Yanchen Dong, Jing Zhao, Ruiqin Xiong, and Tiejun Huang. 3D residual interpolation for spike camera demosaicing. In ICIP, pages 1461-1465. IEEE, 2022. 2, 7, 8 +[11] Yanchen Dong, Ruiqin Xiong, Jing Zhao, Jian Zhang, Xiaopeng Fan, Shuyuan Zhu, and Tiejun Huang. Joint demosaicing and denoising for spike camera. In AAAI, pages 1582-1590, 2024. 1, 2, 3, 7, 8 +[12] Yanchen Dong, Ruiqin Xiong, Jing Zhao, Jian Zhang, Xiaopeng Fan, Shuyuan Zhu, and Tiejun Huang. Learning a deep demosaicing network for spike camera with color filter array. IEEE TIP, 2024. 1, 2, 6, 7, 8 +[13] DC Dowson and BV Landau. The frechet distance between multivariate normal distributions. Journal of multivariate analysis, 12(3):450-455, 1982. 8 +[14] Omar A Elgendy, Abhiram Gnanasambandam, Stanley H Chan, and Jiaju Ma. Low-light demosaicking and denoising for small pixels using learned frequency selection. IEEE TCI, 7:137-150, 2021. 3 +[15] Berthy T. Feng, Jamie Smith, Michael Rubinstein, Huiwen Chang, Katherine L. Bouman, and William T. Freeman. Score-Based Diffusion Models as Principled Priors for Inverse Imaging. In ICCV, 2023. 2 + +[16] Eric R Fossum, Jiaju Ma, and Saleh Masoodian. Quanta image sensor: Concepts and progress. Advanced Photon Counting Techniques X, 9858:985805, 2016. 3 +[17] Eric R Fossum, Jiaju Ma, Saleh Masoodian, Leo Anzagira, and Rachel Zizza. The quanta image sensor: Every photon counts. Sensors, 16(8):1260, 2016. 3 +[18] Michael Gharbi, Gaurav Chaurasia, Sylvain Paris, and Frédo Durand. Deep joint demosaicking and denoising. ACM TOG, 35(6):1-12, 2016. 3 +[19] Felix Heide, Mushfiqur Rouf, Matthias B Hullin, Bjorn Labitzke, Wolfgang Heidrich, and Andreas Kolb. High-quality computational imaging through simple lenses. ACM TOG, 32 (5):1-14, 2013. 3 +[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 5 +[21] Tiejun Huang, Yajing Zheng, Zhaofei Yu, Rui Chen, Yuan Li, Ruiqin Xiong, Lei Ma, Junwei Zhao, Siwei Dong, Lin Zhu, et al. $1000\times$ faster camera and machine vision with ordinary devices. Engineering, 25:110-119, 2022. 2, 3 +[22] Siddhant Jain, Daniel Watson, Eric Tabellion, Aleksander Hołyński, Ben Poole, and Janne Kontkanen. Video Interpolation with Diffusion Models. In CVPR, 2024. 2 +[23] Daisuke Kiku, Yusuke Monno, Masayuki Tanaka, and Masatoshi Okutomi. Beyond color difference: Residual interpolation for color image demosaicking. IEEE TIP, 25(3): 1288-1300, 2016. 3 +[24] Filippos Kokkinos and Stamatios Lefkimmiatis. Deep image demosaicking using a cascade of convolutional residual denoising networks. In ECCV, pages 303-319, 2018. 3 +[25] Juan Antonio Leñero-Bardallo, Teresa Serrano-Gotarredona, and Bernabé Linares-Barranco. A 3.6 $\mu$ s latency asynchronous frame-free event-driven dynamic-vision-sensor. IEEE Journal of Solid-State Circuits, 46(6):1443-1455, 2011. 2, 3 +[26] Xinui Li, Chao Ma, Xiaokang Yang, and Ming-Hsuan Yang. VidToMe: Video token merging for zero-shot video editing. In CVPR, pages 7486-7495, 2024. 8 +[27] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A $128 \times 128$ 120 dB $15\mu s$ latency asynchronous temporal contrast vision sensor. IEEE Journal of Solid-State Circuits, 43(2):566-576, 2008. 2, 3 +[28] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Yu Qiao, Wanli Ouyang, and Chao Dong. Diff-BIR: Toward blind image restoration with generative diffusion prior. In ECCV, pages 430-448. Springer, 2024. 6 +[29] Lin Liu, Xu Jia, Jianzhuang Liu, and Qi Tian. Joint demosaicing and denoising with self guidance. In CVPR, pages 2240-2249, 2020. 3 +[30] Yunfan Lu, Yijie Xu, Wenzong Ma, Weiyu Guo, and Hui Xiong. Event camera demosaicing via swim transformer and pixel-focus loss. In CVPR, pages 1095-1105, 2024. 3 +[31] Sizhuo Ma, Varun Sundar, Paul Mos, Claudio Bruschini, Edoardo Charbon, and Mohit Gupta. Seeing photons in color. ACM TOG, 42(4):1-16, 2023. 2, 3 +[32] Julien Mairal, Michael Elad, and Guillermo Sapiro. Sparse representation for color image restoration. IEEE TIP, 17(1): 53-69, 2007. 3 + +[33] Julien Mairal, Francis Bach, Jean Ponce, Guillermo Sapiro, and Andrew Zisserman. Non-local sparse models for image restoration. In ICCV, pages 2272-2279, 2009. +[34] Henrique S Malvar, Li-wei He, and Ross Cutler. High-quality linear interpolation for demosaicing of bayer-patterned color images. In ICASSP, pages iii-485, 2004. 3 +[35] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a "completely blind" image quality analyzer. IEEE Signal Processing Letters, 20(3):209-212, 2012. 8 +[36] Ibrahim Pekkucuksen and Yucel Altunbasak. Gradient based threshold free color filter array interpolation. In ICIP, pages 137-140, 2010. 3 +[37] Reza Pournaghi and Xiaolin Wu. Coded acquisition of high frame rate video. IEEE TIP, 23(12):5670-5682, 2014. 2 +[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 6 +[39] Hyeonjun Sim, Jihyong Oh, and Munchurl Kim. XVFI: extreme video frame interpolation. In ICCV, 2021. 6 +[40] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 3 +[41] Daniel Stanley Tan, Wei-Yang Chen, and Kai-Lung Hua. DeepDemosaicking: Adaptive image demosaicking via multiple deep fully convolutional networks. IEEE TIP, 27(5): 2408-2419, 2018. 3 +[42] Runjie Tan, Kai Zhang, Wangmeng Zuo, and Lei Zhang. Color image demosaicking via deep residual learning. In ICME, pages 793-798, 2017. 3 +[43] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-Shot Image Restoration Using Denoising Diffusion Null-Space Model. In ICLR, 2022. 2 +[44] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004. 8 +[45] Bennett Wilburn, Neel Joshi, Vaibhav Vaish, Marc Levoy, and Mark Horowitz. High-speed videography using a dense camera array. CVPR, 2004. 2 +[46] Wenzhu Xing and Karen Egiazarian. End-to-end learning for joint image demosaicing, denoising and super-resolution. In CVPR, pages 3507-3516, 2021. 3 +[47] Senyan Xu, Zhijing Sun, Jiaying Zhu, Yurui Zhu, Xueyang Fu, and Zheng-Jun Zha. Demosaicformer: Coarse-to-fine demosaicing network for hybrids camera. In CVPR, pages 1126-1135, 2024. 3 +[48] Siqi Yang, Zhaojun Huang, Yakun Chang, Bin Fan, Zhaofei Yu, and Boxin Shi. Real-data-driven 2000 FPS color video from mosaicked chromatic spikes. In ECCV, 2024. 2, 3, 6, 7, 8 +[49] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling Up to Excellence: Practicing Model Scaling for Photo-Realistic Image Restoration In the Wild. In CVPR, 2024. 2 + +[50] Guoshen Yu, Guillermo Sapiro, and Stephane Mallat. Solving inverse problems with piecewise linear estimators: From gaussian mixture models to structured sparsity. IEEE TIP, 21 (5):2481-2499, 2011. 3 +[51] Jie Zhang, Jonathan Newman, Zeguan Wang, Yong Qian, Pedro Feliciano-Ramos, Wei Guo, Takato Honda, Zhe Sage Chen, Changyang Linghu, Ralph Etienne-Cummings, et al. Pixel-wise programmability enables dynamic high-snr cameras for high-speed microscopy. Nature Communications, 15 (1):4480, 2024. 2 +[52] Lin Zhang, Lei Zhang, and Alan C Bovik. A feature-enriched completely blind image quality evaluator. IEEE TIP, 24(8): 2579-2591, 2015. 8 +[53] Zhao Zhang, Huan Zheng, Richang Hong, Mingliang Xu, Shuicheng Yan, and Meng Wang. Deep color consistent network for low-light image enhancement. In CVPR, pages 1899-1908, 2022. 3 +[54] Jing Zhao, Ruiqin Xiong, Hangfan Liu, Jian Zhang, and Tiejun Huang. Spk2Imgnet: Learning to reconstruct dynamic scene from continuous spike stream. In CVPR, pages 11996-12005, 2021. 2 +[55] Jing Zhao, Ruiqin Xiong, Jiyu Xie, Boxin Shi, Zhaofei Yu, Wen Gao, and Tiejun Huang. Reconstructing clear image for high-speed motion scene with a retina-inspired spike camera. IEEE TCI, 8:12-27, 2021. 2 +[56] Yajing Zheng, Lingxiao Zheng, Zhaofei Yu, Boxin Shi, Yonghong Tian, and Tiejun Huang. High-speed image reconstruction through short-term plasticity for spiking cameras. In CVPR, pages 6358-6367, 2021. 2 +[57] Lin Zhu, Siwei Dong, Tiejun Huang, and Yonghong Tian. A retina-inspired sampling method for visual texture reconstruction. In ICME, pages 1432-1437, 2019. 2, 4, 7, 8 +[58] Lin Zhu, Siwei Dong, Jianing Li, Tiejun Huang, and Yonghong Tian. Retina-like visual image reconstruction via spiking neural model. In CVPR, pages 1438-1446, 2020. 2 \ No newline at end of file diff --git a/2025/SpikeDiff_ Zero-shot High-Quality Video Reconstruction from Chromatic Spike Camera and Sub-millisecond Spike Streams/images.zip b/2025/SpikeDiff_ Zero-shot High-Quality Video Reconstruction from Chromatic Spike Camera and Sub-millisecond Spike Streams/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..227041af5eb6b820551644301264934971148637 --- /dev/null +++ b/2025/SpikeDiff_ Zero-shot High-Quality Video Reconstruction from Chromatic Spike Camera and Sub-millisecond Spike Streams/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b197a1c9bbbf38ce41a72df76206c15b286ab78f83eb59c03834dd1106f2963 +size 703057 diff --git a/2025/SpikeDiff_ Zero-shot High-Quality Video Reconstruction from Chromatic Spike Camera and Sub-millisecond Spike Streams/layout.json b/2025/SpikeDiff_ Zero-shot High-Quality Video Reconstruction from Chromatic Spike Camera and Sub-millisecond Spike Streams/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..21031e663090df6ea8469a0c62286ff51196c91b --- /dev/null +++ b/2025/SpikeDiff_ Zero-shot High-Quality Video Reconstruction from Chromatic Spike Camera and Sub-millisecond Spike Streams/layout.json @@ -0,0 +1,10411 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 113, + 103, + 501, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 103, + 501, + 140 + ], + "spans": [ + { + "bbox": [ + 113, + 103, + 501, + 140 + ], + "type": "text", + "content": "SpikeDiff: Zero-shot High-Quality Video Reconstruction from Chromatic Spike Camera and Sub-millisecondskip Spike Streams" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 93, + 159, + 160, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 159, + 160, + 176 + ], + "spans": [ + { + "bbox": [ + 93, + 159, + 160, + 176 + ], + "type": "text", + "content": "Siqi Yang" + }, + { + "bbox": [ + 93, + 159, + 160, + 176 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 179, + 160, + 264, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 160, + 264, + 189 + ], + "spans": [ + { + "bbox": [ + 179, + 160, + 264, + 189 + ], + "type": "text", + "content": "Jinxiu Liang" + }, + { + "bbox": [ + 179, + 160, + 264, + 189 + ], + "type": "inline_equation", + "content": "^{2,3,4*}" + }, + { + "bbox": [ + 179, + 160, + 264, + 189 + ], + "type": "text", + "content": " \nYakun Chang" + }, + { + "bbox": [ + 179, + 160, + 264, + 189 + ], + "type": "inline_equation", + "content": "^{5,6}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 280, + 161, + 369, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 280, + 161, + 369, + 189 + ], + "spans": [ + { + "bbox": [ + 280, + 161, + 369, + 189 + ], + "type": "text", + "content": "Zhaojun Huang2,3 Zhaofei Yu1,3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 390, + 161, + 495, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 161, + 495, + 175 + ], + "spans": [ + { + "bbox": [ + 390, + 161, + 495, + 175 + ], + "type": "text", + "content": "Yeliduosi Xiaokaiti2,3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 365, + 175, + 436, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 175, + 436, + 188 + ], + "spans": [ + { + "bbox": [ + 365, + 175, + 436, + 188 + ], + "type": "text", + "content": "Boxin Shi" + }, + { + "bbox": [ + 365, + 175, + 436, + 188 + ], + "type": "inline_equation", + "content": "^{2,3,1*}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 206, + 190, + 405, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 190, + 405, + 203 + ], + "spans": [ + { + "bbox": [ + 206, + 190, + 405, + 203 + ], + "type": "text", + "content": "1 Institute for Artificial Intelligence, Peking University" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 204, + 506, + 231 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 104, + 204, + 506, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 204, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 104, + 204, + 506, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 104, + 204, + 506, + 217 + ], + "type": "text", + "content": " State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 218, + 506, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 218, + 506, + 231 + ], + "spans": [ + { + "bbox": [ + 104, + 218, + 506, + 231 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 104, + 218, + 506, + 231 + ], + "type": "text", + "content": " National Engineering Research Center of Visual Technology, School of Computer Science, Peking University" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 122, + 232, + 247, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 232, + 247, + 244 + ], + "spans": [ + { + "bbox": [ + 122, + 232, + 247, + 244 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 122, + 232, + 247, + 244 + ], + "type": "text", + "content": " National Institute of Informatics" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 262, + 232, + 487, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 232, + 487, + 245 + ], + "spans": [ + { + "bbox": [ + 262, + 232, + 487, + 245 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 262, + 232, + 487, + 245 + ], + "type": "text", + "content": " Institute of Information Science, Beijing Jiaotong University" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 127, + 246, + 482, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 246, + 482, + 273 + ], + "spans": [ + { + "bbox": [ + 127, + 246, + 482, + 273 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 127, + 246, + 482, + 273 + ], + "type": "text", + "content": " Visual Intelligence +X International Cooperation Joint Laboratory of the Ministry of Education,.cssherryliang@gmail.com, shiboxin@pku.edu.cn" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 58, + 277, + 552, + 429 + ], + "blocks": [ + { + "bbox": [ + 58, + 277, + 552, + 429 + ], + "lines": [ + { + "bbox": [ + 58, + 277, + 552, + 429 + ], + "spans": [ + { + "bbox": [ + 58, + 277, + 552, + 429 + ], + "type": "image", + "image_path": "bd5f75b0432674a69a21f655aa0ecc07b693c85e4cdff72f3ae80e2030a6484c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 431, + 555, + 498 + ], + "lines": [ + { + "bbox": [ + 55, + 431, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 555, + 498 + ], + "type": "text", + "content": "Figure 1. Spike cameras capture chromatic spikes with high temporal resolution, enabling the observation of scenes with rapid motion. As illustrated in the orange region, the spike camera utilizes a color filter array to capture chromatic information and continuously accumulates photons until reaching a threshold that triggers a spike. To reconstruct high-quality videos from these chromatic spikes, we propose SpikeDiff, a zero-shot framework that leverages principled priors from pretrained diffusion models and integrates physics-based guidance derived from the spike camera mechanism. SpikeDiff generates visually appealing videos without noise or motion blur, significantly outperforming existing chromatic spike reconstruction methods (e.g., SJDD [11] and CSpkNet [12])." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 152, + 508, + 200, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 508, + 200, + 521 + ], + "spans": [ + { + "bbox": [ + 152, + 508, + 200, + 521 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 534, + 297, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 297, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 297, + 689 + ], + "type": "text", + "content": "High-speed video reconstruction from neuromorphic spike cameras offers a promising alternative to traditional frame-based imaging, providing superior temporal resolution and dynamic range with reduced power consumption. Nevertheless, reconstructing high-quality colored videos from spikes captured in ultra-short time intervals (sub-millisecond) remain challenging due to the inherently noisy nature of spikes. While some existing methods extend the temporal capture window to improve reconstruction quality, they inevitably compromise the temporal resolution advantages of spike cameras. In this paper, we introduce SpikeDiff, the first zero-shot framework that leverages pretrained diffusion models to reconstruct high-quality colored videos from sub-millisecond" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 509, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 556, + 628 + ], + "type": "text", + "content": "(0.5ms) chromatic spike streams. By incorporating physics-based guidance into the diffusion sampling process, SpikeDiff bridges the domain gap between chromatic spikes and conventional images, enabling high-fidelity reconstruction without requiring domain-specific training data. Extensive experiments demonstrate that SpikeDiff achieves impressive reconstruction quality while maintaining ultra-high temporal resolution, outperforming existing methods across diverse challenging scenarios in both perceptual quality and structural preservation." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 643, + 395, + 656 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 643, + 395, + 656 + ], + "spans": [ + { + "bbox": [ + 314, + 643, + 395, + 656 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "High-speed motions—from droplet impacts to mechanical dynamics—occur at speeds surpassing human perceptual processing capabilities. Conventional high-speed cameras are capable of capturing these motions but encounter funda-" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 148, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 148, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 148, + 713 + ], + "type": "text", + "content": "* Corresponding authors." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7905" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "mental trade-offs [37, 45, 51]: higher temporal resolution requires shorter exposures, leading to reduced light capture and limited dynamic range, alongside increased power consumption and larger data volumes. These constraints make continuous high-speed capture over extended periods either impractical or impossible." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 145, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 145, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 145, + 295, + 312 + ], + "type": "text", + "content": "Neuromorphic cameras have emerged as a compelling alternative to conventional high-speed imaging, including event cameras [9, 25, 27] and spike cameras [3, 21, 58]. Drawing inspiration from the fovea in biological visual systems, each pixel in a spike camera continuously accumulates photons, generates spikes upon reaching the predefined threshold, and subsequently proceeds to the next accumulation period. Operating at a sampling rate of " + }, + { + "bbox": [ + 55, + 145, + 295, + 312 + ], + "type": "inline_equation", + "content": "20,000\\mathrm{Hz}" + }, + { + "bbox": [ + 55, + 145, + 295, + 312 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 145, + 295, + 312 + ], + "type": "inline_equation", + "content": "1000\\times 1000" + }, + { + "bbox": [ + 55, + 145, + 295, + 312 + ], + "type": "text", + "content": " resolution, spike cameras encode visual information with low data overhead and power consumption, compared to conventional cameras with similar sampling rate and resolution. This enables continuously blur-free capture with high-speed motion and high contrast over several minutes that conventional sensors struggle to handle." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 313, + 295, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 313, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 295, + 420 + ], + "type": "text", + "content": "For human perception, the binary representations of spikes are inherently difficult to interpret and require reconstruction into video frames [54, 55, 58]. However, reconstructing high-quality video frames from spikes remains challenging, particularly within ultra-short time intervals where limited spikes make existing methods highly susceptible to noise [57]. This challenge intensifies for color reconstruction, as Color Filter Arrays (CFAs) further reduce the number of available spikes within short time intervals." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 422, + 295, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 295, + 564 + ], + "type": "text", + "content": "Existing approaches attempt to mitigate these challenges by extending the observed time intervals to accumulate more spikes [31, 56, 57]. This strategy inadvertently introduces motion blur and ghosting artifacts that compromise the temporal resolution advantage of spike cameras. While recent deep learning-based methods have shown promise in optimizing this trade-off problem, they typically exhibit limited generalization capabilities [10, 11], require extensive training data [12], or demand longer time intervals [48]. Yet, these methods still struggle to completely eliminate the noise [11] (upper right in Figure 1) or avoid introducing motion blur in high-speed motion regions [48] (lower right in Figure 1)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 565, + 295, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 565, + 295, + 661 + ], + "spans": [ + { + "bbox": [ + 54, + 565, + 295, + 661 + ], + "type": "text", + "content": "The recent success of diffusion models in image and video restoration [15, 22, 43, 49] suggests a promising direction. These methods have demonstrated remarkable capability in encoding rich priors about natural scene statistics and temporal dynamics. This raises an intriguing question: Can we leverage these powerful priors to reconstruct high-quality frames and videos from extremely noisy chromatic spikes captured within sub-millisecond " + }, + { + "bbox": [ + 54, + 565, + 295, + 661 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 54, + 565, + 295, + 661 + ], + "type": "text", + "content": " time intervals?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 662, + 295, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 662, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 67, + 662, + 295, + 674 + ], + "type": "text", + "content": "Integrating pretrained diffusion models with spike-to" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "type": "text", + "content": "video reconstruction presents fundamental challenges: Spike cameras capture visual information through asynchronous voltage accumulation and threshold-based sampling—a process fundamentally different from the fixed-interval capturing mechanism of frame-based cameras. Furthermore, chromatic spikes represent raw sensor measurements without established signal processing pipelines, while pretrained diffusion models are typically trained on processed sRGB frames with optimized color and tone characteristics. These disparities create a significant domain gap that must be addressed for effective reconstruction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 204, + 556, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 556, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 556, + 346 + ], + "type": "text", + "content": "In this paper, we propose SpikeDiff, the first zero-shot spike-to-video reconstruction framework that leverages pretrained diffusion models as principled priors for high-quality colored video reconstruction from chromatic spikes. By establishing a physically grounded likelihood model of spike measurements throughout the reverse diffusion process, SpikeDiff approximates the posterior sampling of high-quality videos from sub-milliseconds spike streams. Unlike existing approaches that require extensive domain-specific training data and risk overfitting, our method operates in a zero-shot manner, eliminating both the computational burden of training and the need for large-scale datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 326, + 348, + 451, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 348, + 451, + 358 + ], + "spans": [ + { + "bbox": [ + 326, + 348, + 451, + 358 + ], + "type": "text", + "content": "Our key contributions include:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 359, + 555, + 455 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 359, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 359, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 314, + 359, + 555, + 407 + ], + "type": "text", + "content": "- the first framework that leverages pretrained diffusion models as powerful priors for high-quality, zero-shot video reconstruction from chromatic spikes in ultra-short intervals, eliminating the need for domain-specific training data; and" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 407, + 553, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 553, + 455 + ], + "type": "text", + "content": "- a physics-based and differentiable formulation of the spike generation process, which enables effective integration with pretrained diffusion models, bridging the domain gap between spikes and conventional images." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 455, + 555, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 555, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 555, + 526 + ], + "type": "text", + "content": "Extensive experiments across diverse challenging scenarios demonstrate that SpikeDiff consistently outperforms prior methods that rely on extended time intervals and compromising temporal resolution, effectively breaking the traditional quality-time resolution trade-off in spike-to-video reconstruction (see our result shown in Figure 1)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 538, + 399, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 538, + 399, + 550 + ], + "spans": [ + { + "bbox": [ + 314, + 538, + 399, + 550 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": "Video reconstruction with monochromatic spikes. Reconstructing high-speed videos from spikes typically employs the imaging model of spike cameras [21]. By accumulating the spikes triggered within a certain time window or estimating the interval between successive spikes, textures can be reconstructed [57, 58]. However, due to the limited number of photons arriving at the pixel in an ultra-short time, such straightforward reconstruction suffers from significant noise. To address this issue, Chang et al. [4] propose enhancing the signal-to-noise ratio (SNR) by combining multi-bit spikes with binary spikes in a rolling-mixed-bit manner. However, this solution requires hardware modifications, limiting its practicality in more general scenes. Zhao et al. [54] present" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": "Existing state-of-the-art methods [10-12] require at least a time interval of " + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "inline_equation", + "content": "40/20, 000 \\mathrm{~Hz} = 2" + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": " ms for high-quality video reconstruction, while we aim at ultra-short time interval of " + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "inline_equation", + "content": "10/20, 000 \\mathrm{~Hz} = 0.5" + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": " ms." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7906" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 251 + ], + "type": "text", + "content": "Spk2ImgNet, a hierarchical architecture that progressively fuses the spikes, offering an alternative approach to improve reconstruction quality. However, its generalization ability is limited since it relies on synthetic data for training. To address the issue of real-world ground truth for training, self-supervised methods [5, 6] are developed to reduce dependence on synthetic datasets. However, it is still challenging to reconstruct color videos from monochromatic spikes. From the perspective of incorporating additional color guidance, Chang et al. [3] develop a hybrid spike-RGB camera system that performs spatial alignment and frame interpolation simultaneously, enabling the recovery of 1000 FPS color video. However, this hybrid camera system necessitates synchronization and optical alignment, while the beam splitter also poses challenges for constructing a compact device." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 251, + 298, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 298, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 298, + 647 + ], + "type": "text", + "content": "Demosaicking and denoising for unconventional cameras. Various methods for joint demosaicking and denoising have been proposed to obtain high-resolution and low-noise color images from frame-based RAW data, ranging from traditional model-based techniques [1, 8, 19, 23, 32-34, 36, 50] to more recent data-driven approaches [18, 24, 29, 41, 42, 46, 53]. However, all these methods cannot be applied to binary data captured by the new-generation high-speed cameras, such as event cameras [9, 25, 27], quanta image sensors [16, 17], and spike cameras [21]. In the case of event cameras, Xu et al. [47] and Lu et al. [30] present Transformer-based architectures for demosaicking missing pixel values in RAW domain processing. However, due to the sparsity of events, these approaches still rely on frame-based cameras for texture reconstruction. In the context of quanta image sensors, Elgendy et al. [14] integrate a frequency selection method with a deep neural network based filtering approach, leveraging the luma channels to assist in the denoising of the chroma channels. Ma et al. [31] design a blue-noise pseudo-random RGBW color filter array and successfully reconstruct high-quality color images from mosaicked single-bit frames, even in high-dynamic-range (HDR) scenes with complex and rapid motion. For spike cameras, Dong et al. [11] propose an offset-sharing deformable convolution module to align temporal features of color channels and develop a spike noise estimator to characterize the noise distribution. However, this method is trained on synthetic data, limiting its generalization to real-world data. Yang et al. [48] introduce a self-supervised denoising module trained exclusively on real-world chromatic spikes, achieving 2000 FPS color video reconstruction. However, there remains significant research potential in balancing noise suppression with detail preservation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 657, + 111, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 111, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 111, + 670 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "We present the degradation model of chromatic spikes based on the physical formulation and propose SpikeDiff, a zero-shot framework for reconstructing consecutive high-quality" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": "frames from chromatic spikes. This approach leverages pretrained diffusion models as principled priors to effectively address the inherent ill-posedness of the problem." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 115, + 400, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 115, + 400, + 126 + ], + "spans": [ + { + "bbox": [ + 313, + 115, + 400, + 126 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 133, + 555, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 289 + ], + "type": "text", + "content": "Chromatic spikes. The chromatic spike camera leverages a spike sensor equipped with a Bayer-pattern (RGGB) CFA to capture the scene with chromatic information. As shown in Figure 1, when photons filtered by the CFA reach a pixel, the electrons generated by these photons are continuously accumulated as long as the voltage does not reach the threshold " + }, + { + "bbox": [ + 313, + 133, + 555, + 289 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{th}}" + }, + { + "bbox": [ + 313, + 133, + 555, + 289 + ], + "type": "text", + "content": ". Simultaneously, the readout circuit samples the pixel values at a frequency of " + }, + { + "bbox": [ + 313, + 133, + 555, + 289 + ], + "type": "inline_equation", + "content": "20,000\\mathrm{Hz}" + }, + { + "bbox": [ + 313, + 133, + 555, + 289 + ], + "type": "text", + "content": " and a signal of 0 is read out at each readout point. Once the accumulated voltage reaches " + }, + { + "bbox": [ + 313, + 133, + 555, + 289 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{th}}" + }, + { + "bbox": [ + 313, + 133, + 555, + 289 + ], + "type": "text", + "content": ", a signal of 1 is read out and the voltage of this pixel is reset to continue the next accumulation period. For each pixel " + }, + { + "bbox": [ + 313, + 133, + 555, + 289 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 133, + 555, + 289 + ], + "type": "text", + "content": ", we denote the accumulated voltage at a readout point " + }, + { + "bbox": [ + 313, + 133, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 133, + 555, + 289 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 133, + 555, + 289 + ], + "type": "inline_equation", + "content": "E(i,\\tau)" + }, + { + "bbox": [ + 313, + 133, + 555, + 289 + ], + "type": "text", + "content": ", the triggered spike at " + }, + { + "bbox": [ + 313, + 133, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 133, + 555, + 289 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 364, + 295, + 555, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 295, + 555, + 328 + ], + "spans": [ + { + "bbox": [ + 364, + 295, + 555, + 328 + ], + "type": "interline_equation", + "content": "S (i, \\tau) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} E (i, \\tau) \\geq E _ {\\mathrm {t h}}, \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {1}", + "image_path": "79c257469956e399161744fb7909cf3d9a3448dfedc5f2fbf7c23dbfb695f101.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 339, + 556, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 556, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 556, + 426 + ], + "type": "text", + "content": "Diffusion models. Latent diffusion models (LDMs) operate in a compressed latent space learned by an autoencoder pair " + }, + { + "bbox": [ + 313, + 339, + 556, + 426 + ], + "type": "inline_equation", + "content": "(\\mathcal{E},\\mathcal{D})" + }, + { + "bbox": [ + 313, + 339, + 556, + 426 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 339, + 556, + 426 + ], + "type": "inline_equation", + "content": "X = \\mathcal{D}(Z)" + }, + { + "bbox": [ + 313, + 339, + 556, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 339, + 556, + 426 + ], + "type": "inline_equation", + "content": "Z = \\mathcal{E}(X)" + }, + { + "bbox": [ + 313, + 339, + 556, + 426 + ], + "type": "text", + "content": ". An LDM learns to reverse a diffusion process that gradually adds Gaussian noise to a latent vector " + }, + { + "bbox": [ + 313, + 339, + 556, + 426 + ], + "type": "inline_equation", + "content": "Z_{0}" + }, + { + "bbox": [ + 313, + 339, + 556, + 426 + ], + "type": "text", + "content": ". The reverse process, modeled by a stochastic differential equation (SDE), generates a clean latent " + }, + { + "bbox": [ + 313, + 339, + 556, + 426 + ], + "type": "inline_equation", + "content": "Z_{0}" + }, + { + "bbox": [ + 313, + 339, + 556, + 426 + ], + "type": "text", + "content": " from pure noise " + }, + { + "bbox": [ + 313, + 339, + 556, + 426 + ], + "type": "inline_equation", + "content": "Z_{T} \\sim \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 313, + 339, + 556, + 426 + ], + "type": "text", + "content": " [40]:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 318, + 443, + 555, + 458 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 443, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 318, + 443, + 555, + 458 + ], + "type": "interline_equation", + "content": "d \\boldsymbol {Z} = [ - f (\\boldsymbol {Z}, t) - g ^ {2} (t) \\nabla_ {\\boldsymbol {Z} _ {t}} \\log p _ {t} (\\boldsymbol {Z} _ {t}) ] d t + g (t) d w, (2)", + "image_path": "d390177fb43b0f1fb1c5fc9e45a7fdca5334b4bf866694be5c4ed62cc9152ec8.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 464, + 556, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 556, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 556, + 548 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 464, + 556, + 548 + ], + "type": "inline_equation", + "content": "f(\\cdot, t)" + }, + { + "bbox": [ + 313, + 464, + 556, + 548 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 464, + 556, + 548 + ], + "type": "inline_equation", + "content": "g(t)" + }, + { + "bbox": [ + 313, + 464, + 556, + 548 + ], + "type": "text", + "content": " are drift and diffusion coefficients, " + }, + { + "bbox": [ + 313, + 464, + 556, + 548 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 464, + 556, + 548 + ], + "type": "text", + "content": " represents the standard Brownian motion. The key component is the score function " + }, + { + "bbox": [ + 313, + 464, + 556, + 548 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{Z}_t} \\log p_t(\\mathbf{Z}_t)" + }, + { + "bbox": [ + 313, + 464, + 556, + 548 + ], + "type": "text", + "content": ", which is approximated by a time-conditioned neural network " + }, + { + "bbox": [ + 313, + 464, + 556, + 548 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{Z}, t)" + }, + { + "bbox": [ + 313, + 464, + 556, + 548 + ], + "type": "text", + "content": ". By iteratively applying the learned score, the reverse SDE can sample from the learned data prior " + }, + { + "bbox": [ + 313, + 464, + 556, + 548 + ], + "type": "inline_equation", + "content": "p(\\mathbf{Z}_0)" + }, + { + "bbox": [ + 313, + 464, + 556, + 548 + ], + "type": "text", + "content": " to generate a high-quality image " + }, + { + "bbox": [ + 313, + 464, + 556, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_0 = \\mathcal{D}(\\mathbf{Z}_0)" + }, + { + "bbox": [ + 313, + 464, + 556, + 548 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 555, + 538, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 538, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 538, + 567 + ], + "type": "text", + "content": "3.2. Diffusion-based reconstruction from spikes" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 572, + 556, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 572, + 556, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 572, + 556, + 681 + ], + "type": "text", + "content": "Reconstructing consecutive high-quality frames from real-captured chromatic spike streams is challenging due to the non-negligible noise perturbations, particularly within extremely limited time intervals (e.g., sub-milisecond). To address this challenge, we leverage recent advancements in diffusion models, which incorporate learned priors regarding the distribution of high-quality images, and formulate the reconstruction of video frames from chromatic spikes as a Maximum a Posteriori (MAP) estimation problem:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 367, + 687, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 687, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 367, + 687, + 555, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {X} ^ {*} = \\operatorname {a r g m a x} _ {\\boldsymbol {X}} p (\\boldsymbol {X} | \\boldsymbol {Y}) \\tag {3} \\\\ = \\operatorname {a r g m a x} _ {\\boldsymbol {X}} p (\\boldsymbol {Y} | \\boldsymbol {X}) p (\\boldsymbol {X}), \\\\ \\end{array}", + "image_path": "762e007412db883830509eaa53335beff853f4a5fb3df400e07847d96e8078ce.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7907" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 550, + 313 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 550, + 313 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 550, + 313 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 550, + 313 + ], + "type": "image", + "image_path": "6c3b2d7d61d13e1be9619b110693938f4e7b9a9f2d5aa5f17f87143f192733d2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 316, + 555, + 393 + ], + "lines": [ + { + "bbox": [ + 54, + 316, + 555, + 393 + ], + "spans": [ + { + "bbox": [ + 54, + 316, + 555, + 393 + ], + "type": "text", + "content": "Figure 2. Pipeline of SpikeDiff. Our diffusion-based posterior sampling procedure follows the iterative denoising process of diffusion models, starting from the initialization " + }, + { + "bbox": [ + 54, + 316, + 555, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{t_s}" + }, + { + "bbox": [ + 54, + 316, + 555, + 393 + ], + "type": "text", + "content": " and gradually approaching the desired latent " + }, + { + "bbox": [ + 54, + 316, + 555, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_0" + }, + { + "bbox": [ + 54, + 316, + 555, + 393 + ], + "type": "text", + "content": ", indicated by the dashed lines. To maximize the likelihood of the reconstructed video given the spike observations " + }, + { + "bbox": [ + 54, + 316, + 555, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 54, + 316, + 555, + 393 + ], + "type": "text", + "content": ", we enhance each iteration with physics-based spike guidance, which comprises three specially-designed differentiable degradation operators: mosaicking " + }, + { + "bbox": [ + 54, + 316, + 555, + 393 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 54, + 316, + 555, + 393 + ], + "type": "text", + "content": ", color casting " + }, + { + "bbox": [ + 54, + 316, + 555, + 393 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 54, + 316, + 555, + 393 + ], + "type": "text", + "content": ", and quantization " + }, + { + "bbox": [ + 54, + 316, + 555, + 393 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 54, + 316, + 555, + 393 + ], + "type": "text", + "content": ". Finally, a multiscale loss function bridges the gap between degraded frames " + }, + { + "bbox": [ + 54, + 316, + 555, + 393 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(\\mathbf{X}_{0|t})" + }, + { + "bbox": [ + 54, + 316, + 555, + 393 + ], + "type": "text", + "content": " and observed SFR frames " + }, + { + "bbox": [ + 54, + 316, + 555, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 54, + 316, + 555, + 393 + ], + "type": "text", + "content": ". As shown in the top sequence, the visual quality of generated frames progressively improves until convergence. Note that the internal degraded frames and the SFR frames are all Bayer-patterned mosaic images; we present their corresponding demosaicked results for better visualization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 399, + 297, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 399, + 297, + 448 + ], + "spans": [ + { + "bbox": [ + 54, + 399, + 297, + 448 + ], + "type": "text", + "content": "where the prior " + }, + { + "bbox": [ + 54, + 399, + 297, + 448 + ], + "type": "inline_equation", + "content": "p(\\mathbf{X})" + }, + { + "bbox": [ + 54, + 399, + 297, + 448 + ], + "type": "text", + "content": " encodes knowledge of natural images, and the likelihood " + }, + { + "bbox": [ + 54, + 399, + 297, + 448 + ], + "type": "inline_equation", + "content": "p(\\mathbf{Y}|\\mathbf{X})" + }, + { + "bbox": [ + 54, + 399, + 297, + 448 + ], + "type": "text", + "content": " represents the data fidelity to ensure the reconstruction is faithful to the observed chromatic spikes " + }, + { + "bbox": [ + 54, + 399, + 297, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 54, + 399, + 297, + 448 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 460, + 297, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 460, + 297, + 605 + ], + "spans": [ + { + "bbox": [ + 54, + 460, + 297, + 605 + ], + "type": "text", + "content": "The degradation model, mapping light intensity to spike signals (illustrated in Figure 1), establishes the connection from RGB frames " + }, + { + "bbox": [ + 54, + 460, + 297, + 605 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 54, + 460, + 297, + 605 + ], + "type": "text", + "content": " — the modality employed in pretrained diffusion models—and chromatic spikes " + }, + { + "bbox": [ + 54, + 460, + 297, + 605 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 54, + 460, + 297, + 605 + ], + "type": "text", + "content": ", and thereby deriving likelihood " + }, + { + "bbox": [ + 54, + 460, + 297, + 605 + ], + "type": "inline_equation", + "content": "p(\\mathbf{Y}|\\mathbf{X})" + }, + { + "bbox": [ + 54, + 460, + 297, + 605 + ], + "type": "text", + "content": " illustrated in the next (Sec. 3.3). This establishes a principled bridge that enables the reconstruction of " + }, + { + "bbox": [ + 54, + 460, + 297, + 605 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 54, + 460, + 297, + 605 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 54, + 460, + 297, + 605 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 54, + 460, + 297, + 605 + ], + "type": "text", + "content": " via posterior sampling in the MAP framework, as illustrated in the green box in Figure 2, where diffusion models serve as learned priors for " + }, + { + "bbox": [ + 54, + 460, + 297, + 605 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 54, + 460, + 297, + 605 + ], + "type": "text", + "content": " while the likelihood ensures fidelity to " + }, + { + "bbox": [ + 54, + 460, + 297, + 605 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 54, + 460, + 297, + 605 + ], + "type": "text", + "content": ". This is exactly the key in our method to connect the modality used in pretrained diffusion models, RGB frames, and chromatic spikes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 617, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 617, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 617, + 297, + 714 + ], + "type": "text", + "content": "Given the significant disparity between chromatic spike streams and conventional images, the data fidelity term " + }, + { + "bbox": [ + 54, + 617, + 297, + 714 + ], + "type": "inline_equation", + "content": "p(\\mathbf{Y}|\\mathbf{X})" + }, + { + "bbox": [ + 54, + 617, + 297, + 714 + ], + "type": "text", + "content": " cannot be directly derived from time-continuous chromatic spikes. Instead, we resort to a more compact representation. According to the integrating and firing mechanism of spikes, the light intensity is proportional to the spike firing rate (SFR) [57]. Without considering noise, the SFR at " + }, + { + "bbox": [ + 54, + 617, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\tau'" + }, + { + "bbox": [ + 54, + 617, + 297, + 714 + ], + "type": "text", + "content": " can be counted through collecting a number of spikes within" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 399, + 468, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 468, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 468, + 412 + ], + "type": "text", + "content": "a predefined time interval of size " + }, + { + "bbox": [ + 313, + 399, + 468, + 412 + ], + "type": "inline_equation", + "content": "\\Delta \\tau" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 359, + 418, + 555, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 418, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 359, + 418, + 555, + 453 + ], + "type": "interline_equation", + "content": "Y (i, \\tau^ {\\prime}) = \\frac {1}{\\Delta \\tau} \\cdot \\sum_ {\\tau = \\tau^ {\\prime} - \\Delta \\tau / 2} ^ {\\tau^ {\\prime} + \\Delta \\tau / 2} S (i, \\tau). \\tag {4}", + "image_path": "7aa6d5c089a341538d4ee9e9803e58bf2a6374c341636255dc3ae3adea645a37.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 460, + 556, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 556, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 556, + 604 + ], + "type": "text", + "content": "By collecting the SFRs from the pixel array at " + }, + { + "bbox": [ + 313, + 460, + 556, + 604 + ], + "type": "inline_equation", + "content": "\\tau'" + }, + { + "bbox": [ + 313, + 460, + 556, + 604 + ], + "type": "text", + "content": ", we obtain mosaicked SFR frames " + }, + { + "bbox": [ + 313, + 460, + 556, + 604 + ], + "type": "inline_equation", + "content": "Y \\in \\mathbb{R}^{T \\times H \\times W}" + }, + { + "bbox": [ + 313, + 460, + 556, + 604 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 460, + 556, + 604 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 460, + 556, + 604 + ], + "type": "text", + "content": " denotes the number of frames, and " + }, + { + "bbox": [ + 313, + 460, + 556, + 604 + ], + "type": "inline_equation", + "content": "H, W" + }, + { + "bbox": [ + 313, + 460, + 556, + 604 + ], + "type": "text", + "content": " represent the height and width, respectively. The SFR frames, estimated from real-captured chromatic spikes, are sensitive to the inherent noise, especially within limited time intervals, and exhibit misalignment with corresponding consecutive high-quality frames. Consequently, we propose a corresponding degradation model, comprising differentiable operators, as detailed in Sec. 3.3, to bridge the gap between " + }, + { + "bbox": [ + 313, + 460, + 556, + 604 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 460, + 556, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 460, + 556, + 604 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 313, + 460, + 556, + 604 + ], + "type": "text", + "content": " and resolve the likelihood estimation required for posterior sampling, as illustrated in the orange box in Figure 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "type": "text", + "content": "Posterior sampling. The primary challenge in reconstructing consecutive high-quality frames from sub-millisecond chromatic spikes lies in its inherent ill-posedness, making the prior term " + }, + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "type": "inline_equation", + "content": "p(\\mathbf{X})" + }, + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "type": "text", + "content": " crucial. We leverage diffusion models as powerful priors. Given the prior distribution of natural images " + }, + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "type": "inline_equation", + "content": "p(\\mathbf{Z}_t)" + }, + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "type": "text", + "content": " encoded by the pretrained latent diffusion model in Eq. (2) and likelihood defined in the next, we solve the MAP problem in Eq. (3) by sampling from the posterior" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7908" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": "distribution " + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "inline_equation", + "content": "p(\\mathbf{X}|\\mathbf{Y})" + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": ". This is achieved by modifying the reverse SDE of the diffusion model to incorporate guidance from our observations. Using Bayes' rule, the score of the posterior distribution " + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "inline_equation", + "content": "p_t(\\mathbf{Z}_t|\\mathbf{Y})" + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": " can be written as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 129, + 296, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 129, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 129, + 296, + 152 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\boldsymbol {Z} _ {t}} \\log p _ {t} (\\boldsymbol {Z} _ {t} | \\boldsymbol {Y}) = \\nabla_ {\\boldsymbol {Z} _ {t}} \\log p (\\boldsymbol {Z} _ {t}) + \\nabla_ {\\boldsymbol {Z} _ {t}} \\log p (\\boldsymbol {Y} | \\boldsymbol {Z} _ {t}), \\tag {5}", + "image_path": "9944ac4457be55929513c1dcf73e03b9155272e877e72a0dde1f243654deaeb9.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 153, + 297, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 153, + 297, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 153, + 297, + 214 + ], + "type": "text", + "content": "where the prior term " + }, + { + "bbox": [ + 55, + 153, + 297, + 214 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{Z}_t}\\log p(\\mathbf{Z}_t)" + }, + { + "bbox": [ + 55, + 153, + 297, + 214 + ], + "type": "text", + "content": " is given by the pretrained network " + }, + { + "bbox": [ + 55, + 153, + 297, + 214 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta^*}(Z_t,t)" + }, + { + "bbox": [ + 55, + 153, + 297, + 214 + ], + "type": "text", + "content": ". The likelihood term " + }, + { + "bbox": [ + 55, + 153, + 297, + 214 + ], + "type": "inline_equation", + "content": "p(\\mathbf{Y}|\\mathbf{Z}_t)" + }, + { + "bbox": [ + 55, + 153, + 297, + 214 + ], + "type": "text", + "content": " guides the sampling towards latents " + }, + { + "bbox": [ + 55, + 153, + 297, + 214 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_t" + }, + { + "bbox": [ + 55, + 153, + 297, + 214 + ], + "type": "text", + "content": " that are likely to produce our observed SFR frame " + }, + { + "bbox": [ + 55, + 153, + 297, + 214 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 55, + 153, + 297, + 214 + ], + "type": "text", + "content": ". The full reverse SDE for posterior sampling is:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 86, + 220, + 297, + 250 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 220, + 297, + 250 + ], + "spans": [ + { + "bbox": [ + 86, + 220, + 297, + 250 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} d \\boldsymbol {Z} = [ - f (\\boldsymbol {Z}, t) - g ^ {2} (t) (\\nabla_ {\\boldsymbol {Z} _ {t}} \\log p (\\boldsymbol {Z} _ {t}) \\tag {6} \\\\ + \\nabla_ {\\boldsymbol {Z} _ {t}} \\log p (\\boldsymbol {Y} | \\boldsymbol {Z} _ {t})) ] d t + g (t) d w. \\\\ \\end{array}", + "image_path": "b7f5af524250890edcaf598cc90d2e9ae3fc708a5ca0121953dd753deffc1547.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 257, + 296, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 296, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 296, + 307 + ], + "type": "text", + "content": "To compute the likelihood score, we must relate the noisy latent " + }, + { + "bbox": [ + 55, + 257, + 296, + 307 + ], + "type": "inline_equation", + "content": "Z_{t}" + }, + { + "bbox": [ + 55, + 257, + 296, + 307 + ], + "type": "text", + "content": " at an intermediate timestep " + }, + { + "bbox": [ + 55, + 257, + 296, + 307 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 257, + 296, + 307 + ], + "type": "text", + "content": " to the observation " + }, + { + "bbox": [ + 55, + 257, + 296, + 307 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 55, + 257, + 296, + 307 + ], + "type": "text", + "content": ". This involves approximating the conditional probability " + }, + { + "bbox": [ + 55, + 257, + 296, + 307 + ], + "type": "inline_equation", + "content": "p(\\mathbf{Y}|\\mathbf{Z}_t)" + }, + { + "bbox": [ + 55, + 257, + 296, + 307 + ], + "type": "text", + "content": ", which can be expressed as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 86, + 313, + 296, + 364 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 313, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 86, + 313, + 296, + 364 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} p (\\mathbf {Y} \\mid \\mathbf {Z} _ {t}) = \\int_ {c} p (\\mathbf {Y} \\mid \\mathbf {Z} _ {0}, \\mathbf {Z} _ {t}) p (\\mathbf {Z} _ {0} \\mid \\mathbf {Z} _ {t}) d \\mathbf {Z} _ {0} \\tag {7} \\\\ = \\int p (\\boldsymbol {Y} | \\boldsymbol {Z} _ {0}) p (\\boldsymbol {Z} _ {0} | \\boldsymbol {Z} _ {t}) d \\boldsymbol {Z} _ {0}. \\\\ \\end{array}", + "image_path": "eee85a33df68c5dce3f40c0c16a13a2d7d1b253d43117dbd59bfab647062939c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 371, + 296, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 296, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 296, + 419 + ], + "type": "text", + "content": "This integral is intractable. Instead, we approximate it by first estimating the expected clean latent " + }, + { + "bbox": [ + 55, + 371, + 296, + 419 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{0|t} = \\mathbb{E}[Z_0|\\mathbf{Z}_t]" + }, + { + "bbox": [ + 55, + 371, + 296, + 419 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 371, + 296, + 419 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_t" + }, + { + "bbox": [ + 55, + 371, + 296, + 419 + ], + "type": "text", + "content": ". For diffusion models like DDPM [20], the forward process is" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 86, + 427, + 296, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 427, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 86, + 427, + 296, + 441 + ], + "type": "interline_equation", + "content": "\\boldsymbol {Z} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {Z} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} z, \\quad z \\sim \\mathcal {N} (0, I), \\tag {8}", + "image_path": "51574eae388a336b322533427b3e1647a50be34d11c98e00b1e8d3c1ca4be589.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 449, + 296, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 485 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 449, + 296, + 485 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t" + }, + { + "bbox": [ + 55, + 449, + 296, + 485 + ], + "type": "text", + "content": " is the noise schedule parameter from the diffusion model (e.g., DDPM [20]). Using Tweedie's formula [7], we can estimate the conditional mean:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 83, + 491, + 296, + 531 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 491, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 83, + 491, + 296, + 531 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {Z} _ {0 | t} = \\mathbb {E} [ \\boldsymbol {Z} _ {0} | \\boldsymbol {Z} _ {t} ] \\\\ = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(\\boldsymbol {Z} _ {t} + \\left(1 - \\bar {\\alpha} _ {t}\\right) \\nabla_ {\\boldsymbol {Z} _ {t}} \\log p _ {t} \\left(\\boldsymbol {Z} _ {t}\\right)\\right). \\tag {9} \\\\ \\end{array}", + "image_path": "d6f10c5208a12c97c5d87ca3c994f926209d6066e7fd9b847b26c89f58f64cc7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 539, + 296, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 296, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 296, + 611 + ], + "type": "text", + "content": "Here, we replace the true score " + }, + { + "bbox": [ + 55, + 539, + 296, + 611 + ], + "type": "inline_equation", + "content": "\\nabla_{Z_t}\\log p_t(Z_t)" + }, + { + "bbox": [ + 55, + 539, + 296, + 611 + ], + "type": "text", + "content": " with its neural network approximation " + }, + { + "bbox": [ + 55, + 539, + 296, + 611 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta^*}(Z_t,t)" + }, + { + "bbox": [ + 55, + 539, + 296, + 611 + ], + "type": "text", + "content": ". With this estimate of the clean latent, " + }, + { + "bbox": [ + 55, + 539, + 296, + 611 + ], + "type": "inline_equation", + "content": "Z_{0|t}" + }, + { + "bbox": [ + 55, + 539, + 296, + 611 + ], + "type": "text", + "content": ", following the expression " + }, + { + "bbox": [ + 55, + 539, + 296, + 611 + ], + "type": "inline_equation", + "content": "p(\\mathbf{Y}|\\mathbf{Z}_t) = \\mathbb{E}_{\\mathbf{Z}_0\\sim p(\\mathbf{Z}_0|\\mathbf{Z}_t)}[p(\\mathbf{Y}|\\mathcal{D}(\\mathbf{Z}_0))]" + }, + { + "bbox": [ + 55, + 539, + 296, + 611 + ], + "type": "text", + "content": ", we can use a surrogate likelihood approach [7] to make the likelihood tractable:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 102, + 619, + 296, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 619, + 296, + 633 + ], + "spans": [ + { + "bbox": [ + 102, + 619, + 296, + 633 + ], + "type": "interline_equation", + "content": "p (\\boldsymbol {Y} | \\boldsymbol {Z} _ {t}) \\approx p (\\boldsymbol {Y} | \\mathcal {D} (\\mathbb {E} [ \\boldsymbol {Z} _ {0} | \\boldsymbol {Z} _ {t} ])). \\tag {10}", + "image_path": "1ce9803ea1489b911c84fb7160ad83c50e83b01c2e0b3794c64bcdd4357e3c18.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 641, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 665 + ], + "type": "text", + "content": "This formulation provides a practical way to compute the likelihood gradient." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 672, + 264, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 264, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 264, + 685 + ], + "type": "text", + "content": "3.3. Chromatic spikes' likelihood estimation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 689, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 297, + 714 + ], + "type": "text", + "content": "Given a set of sequential SFR frames, denoted by " + }, + { + "bbox": [ + 55, + 689, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 55, + 689, + 297, + 714 + ], + "type": "text", + "content": ", our objective is to establish a connection between the SFR frames" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 555, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 121 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 121 + ], + "type": "text", + "content": "and the desired target reconstructions " + }, + { + "bbox": [ + 313, + 72, + 555, + 121 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 72, + 555, + 121 + ], + "type": "text", + "content": ", with both " + }, + { + "bbox": [ + 313, + 72, + 555, + 121 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 313, + 72, + 555, + 121 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 121 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 72, + 555, + 121 + ], + "type": "text", + "content": " normalized to the range [0, 1]. Hence, we resolve the likelihood estimation of " + }, + { + "bbox": [ + 313, + 72, + 555, + 121 + ], + "type": "inline_equation", + "content": "p(\\mathbf{Y} | Z_t)" + }, + { + "bbox": [ + 313, + 72, + 555, + 121 + ], + "type": "text", + "content": " in Eq. (10) and thereby complete the diffusion-based posterior sampling strategy." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 135, + 555, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 135, + 555, + 230 + ], + "spans": [ + { + "bbox": [ + 313, + 135, + 555, + 230 + ], + "type": "text", + "content": "Degradation model of chromatic spikes. The physical process generating SFR frames " + }, + { + "bbox": [ + 313, + 135, + 555, + 230 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 313, + 135, + 555, + 230 + ], + "type": "text", + "content": " can be modeled as a nonlinear inverse problem: " + }, + { + "bbox": [ + 313, + 135, + 555, + 230 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} = \\mathcal{A}(\\mathbf{X}) + \\mathbf{N}" + }, + { + "bbox": [ + 313, + 135, + 555, + 230 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 135, + 555, + 230 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(\\cdot)" + }, + { + "bbox": [ + 313, + 135, + 555, + 230 + ], + "type": "text", + "content": " denotes the known forward operator modeling the degradation process, and " + }, + { + "bbox": [ + 313, + 135, + 555, + 230 + ], + "type": "inline_equation", + "content": "\\mathbf{N}" + }, + { + "bbox": [ + 313, + 135, + 555, + 230 + ], + "type": "text", + "content": " represents stochastic noise. This problem is inherently ill-posed due to irreversible information loss during the imaging process and interference from noise in the capture system." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 231, + 555, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 555, + 338 + ], + "type": "text", + "content": "The forward operator " + }, + { + "bbox": [ + 313, + 231, + 555, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(\\cdot)" + }, + { + "bbox": [ + 313, + 231, + 555, + 338 + ], + "type": "text", + "content": " in the imaging pipeline transforms the latent high-quality frame " + }, + { + "bbox": [ + 313, + 231, + 555, + 338 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 231, + 555, + 338 + ], + "type": "text", + "content": " through a sequence of operations: " + }, + { + "bbox": [ + 313, + 231, + 555, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(X) = \\mathcal{Q}(CMX)" + }, + { + "bbox": [ + 313, + 231, + 555, + 338 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 231, + 555, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}(\\cdot)" + }, + { + "bbox": [ + 313, + 231, + 555, + 338 + ], + "type": "text", + "content": " represents pixel-wise quantization, " + }, + { + "bbox": [ + 313, + 231, + 555, + 338 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 231, + 555, + 338 + ], + "type": "text", + "content": " implements pixel-wise color casting with per-channel scaling, and " + }, + { + "bbox": [ + 313, + 231, + 555, + 338 + ], + "type": "inline_equation", + "content": "M: \\mathbb{R}^{T \\times H \\times W \\times 3} \\to \\mathbb{R}^{T \\times H \\times W}" + }, + { + "bbox": [ + 313, + 231, + 555, + 338 + ], + "type": "text", + "content": " denotes the mosaicking operator, modeling the color filter array pattern. This decomposition explicitly models each step of the image formation process, providing a foundation for our reconstruction approach." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 338, + 555, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 387 + ], + "type": "text", + "content": "Under the assumption of Gaussian noise, we model the likelihood " + }, + { + "bbox": [ + 313, + 338, + 555, + 387 + ], + "type": "inline_equation", + "content": "p(\\mathbf{Y}|\\mathcal{D}(\\mathbf{Z}_{0|t}))" + }, + { + "bbox": [ + 313, + 338, + 555, + 387 + ], + "type": "text", + "content": " of observing SFR frames " + }, + { + "bbox": [ + 313, + 338, + 555, + 387 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 313, + 338, + 555, + 387 + ], + "type": "text", + "content": " given the target frame " + }, + { + "bbox": [ + 313, + 338, + 555, + 387 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 338, + 555, + 387 + ], + "type": "text", + "content": " as a Gaussian distribution. Maximizing this likelihood translates to minimizing the following loss:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 362, + 395, + 555, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 395, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 362, + 395, + 555, + 418 + ], + "type": "interline_equation", + "content": "\\mathcal {L} \\left(\\boldsymbol {Z} _ {t}\\right) = \\frac {1}{\\sigma^ {2}} \\| \\boldsymbol {Y} - \\mathcal {A} \\left(\\mathcal {D} \\left(\\boldsymbol {Z} _ {0 | t}\\right)\\right) \\| _ {2} ^ {2}, \\tag {11}", + "image_path": "642fcc01318b410cceb125348bfd20588d91518cd35010b056d21a454e707bdf.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 426, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 472 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 426, + 555, + 472 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 426, + 555, + 472 + ], + "type": "text", + "content": " represents the standard deviation of the noise " + }, + { + "bbox": [ + 313, + 426, + 555, + 472 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 426, + 555, + 472 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 426, + 555, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 426, + 555, + 472 + ], + "type": "text", + "content": " represents the differentiable degradation operators defined later. By substituting the score function into Eq. (5), we obtain:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 329, + 483, + 555, + 497 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 483, + 555, + 497 + ], + "spans": [ + { + "bbox": [ + 329, + 483, + 555, + 497 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\boldsymbol {Z} _ {t}} \\log p _ {t} (\\boldsymbol {Z} _ {t} | \\boldsymbol {Y}) = \\epsilon_ {\\theta^ {*}} (\\boldsymbol {Z} _ {t}, t) - s \\nabla_ {\\boldsymbol {Z} _ {t}} \\mathcal {L} (\\boldsymbol {Z} _ {t}), \\tag {12}", + "image_path": "be538f3271e844563b1a9dfe126ac0bc2af7c3bff5dc9b4194d6ca1099c7a811.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 506, + 539, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 539, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 539, + 517 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 506, + 539, + 517 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 506, + 539, + 517 + ], + "type": "text", + "content": " controls the strength of measurement guidance." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 532, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 555, + 689 + ], + "type": "text", + "content": "Differentiable degradation operators. A key innovation in SpikeDiff is our design of differentiable operators in " + }, + { + "bbox": [ + 313, + 532, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 532, + 555, + 689 + ], + "type": "text", + "content": " that model the physical characteristics of chromatic spike cameras, which include color casting " + }, + { + "bbox": [ + 313, + 532, + 555, + 689 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 532, + 555, + 689 + ], + "type": "text", + "content": ", quantization " + }, + { + "bbox": [ + 313, + 532, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 313, + 532, + 555, + 689 + ], + "type": "text", + "content": ", and mosaicking " + }, + { + "bbox": [ + 313, + 532, + 555, + 689 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 532, + 555, + 689 + ], + "type": "text", + "content": ". Based on the gray-world prior [2] for the target frame, which indicates that each color channel should have a mean value close to one another, we estimate the color casting coefficients " + }, + { + "bbox": [ + 313, + 532, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\alpha_{R}, \\alpha_{G}, \\alpha_{B}" + }, + { + "bbox": [ + 313, + 532, + 555, + 689 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 532, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\alpha_{c} = \\frac{1}{\\mu(\\mathbf{Y}_{c})}" + }, + { + "bbox": [ + 313, + 532, + 555, + 689 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 532, + 555, + 689 + ], + "type": "inline_equation", + "content": "c \\in \\{\\mathrm{R}, \\mathrm{G}, \\mathrm{B}\\}" + }, + { + "bbox": [ + 313, + 532, + 555, + 689 + ], + "type": "text", + "content": " denotes the color channel, " + }, + { + "bbox": [ + 313, + 532, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\mu(\\mathbf{Y}_{c})" + }, + { + "bbox": [ + 313, + 532, + 555, + 689 + ], + "type": "text", + "content": " is the mean value of color channel " + }, + { + "bbox": [ + 313, + 532, + 555, + 689 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 532, + 555, + 689 + ], + "type": "text", + "content": " among " + }, + { + "bbox": [ + 313, + 532, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 313, + 532, + 555, + 689 + ], + "type": "text", + "content": " across the temporal axis. Then the channel-wise scaling factors for the color casting operator " + }, + { + "bbox": [ + 313, + 532, + 555, + 689 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 532, + 555, + 689 + ], + "type": "text", + "content": " can be computed as " + }, + { + "bbox": [ + 313, + 532, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\eta / \\alpha_{c}" + }, + { + "bbox": [ + 313, + 532, + 555, + 689 + ], + "type": "text", + "content": ", where parameter " + }, + { + "bbox": [ + 313, + 532, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 532, + 555, + 689 + ], + "type": "text", + "content": " is used to adjust the value range." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "SFR frames are inherently quantized and discrete as they are accumulation of binary spike frames, which differ from" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7909" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "8-bit frames " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": ". To simulate the quantization effect and maintain the backward differentiability, we design a soft quantization operator " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": ". Given the expected quantization levels determined by SFR estimation, the soft quantization function " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "q(\\cdot)" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " operates on each pixel individually:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 79, + 140, + 295, + 154 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 140, + 295, + 154 + ], + "spans": [ + { + "bbox": [ + 79, + 140, + 295, + 154 + ], + "type": "interline_equation", + "content": "q (x) = l (x) \\cdot (1 - \\varphi (x)) + u (x) \\cdot \\varphi (x), \\tag {13}", + "image_path": "7fec6bea275187c98cb672920b069559818fb22ec1da9be6eafa772b95e500d2.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 78, + 156, + 295, + 179 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 156, + 295, + 179 + ], + "spans": [ + { + "bbox": [ + 78, + 156, + 295, + 179 + ], + "type": "interline_equation", + "content": "\\varphi (x) = \\left[ \\tanh \\left(k \\left(x - \\frac {l (x) + u (x)}{2}\\right)\\right) + 1 \\right] / 2, \\tag {14}", + "image_path": "55683730b444ed0be55bb5179b67f74a36447e1267e2f6285dda20768c0b88bd.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 187, + 295, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 187, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 55, + 187, + 295, + 235 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 187, + 295, + 235 + ], + "type": "inline_equation", + "content": "l(x)" + }, + { + "bbox": [ + 55, + 187, + 295, + 235 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 187, + 295, + 235 + ], + "type": "inline_equation", + "content": "u(x)" + }, + { + "bbox": [ + 55, + 187, + 295, + 235 + ], + "type": "text", + "content": " denote the lower and upper quantization levels of " + }, + { + "bbox": [ + 55, + 187, + 295, + 235 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 187, + 295, + 235 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 187, + 295, + 235 + ], + "type": "inline_equation", + "content": "\\varphi(\\cdot)" + }, + { + "bbox": [ + 55, + 187, + 295, + 235 + ], + "type": "text", + "content": " quantizes the input value to the range [0, 1] with a smooth curve, and hyperparameter " + }, + { + "bbox": [ + 55, + 187, + 295, + 235 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 187, + 295, + 235 + ], + "type": "text", + "content": " controls the smoothness, as shown in Figure 3(a)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 236, + 296, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 236, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 296, + 319 + ], + "type": "text", + "content": "The naive Bayer-pattern mosaicking operator " + }, + { + "bbox": [ + 55, + 236, + 296, + 319 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 236, + 296, + 319 + ], + "type": "text", + "content": " converts target frame from dimension " + }, + { + "bbox": [ + 55, + 236, + 296, + 319 + ], + "type": "inline_equation", + "content": "T \\times H \\times W \\times 3" + }, + { + "bbox": [ + 55, + 236, + 296, + 319 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 236, + 296, + 319 + ], + "type": "inline_equation", + "content": "T \\times \\frac{H}{2} \\times \\frac{W}{2} \\times 4" + }, + { + "bbox": [ + 55, + 236, + 296, + 319 + ], + "type": "text", + "content": ", resulting in a partially defined backpropagation function due to its gradient discontinuity. To simulate the Bayer-pattern CFA while maintaining a dense gradient flow, we extend the operator's derivative with linear interpolation convolution, similar to demosaicking algorithms." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 319, + 296, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 355 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 355 + ], + "type": "text", + "content": "By composing these operators, we conclude a differentiable simulation from consecutive frames to SFR frames, enabling optimization in the posterior sampling process." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 369, + 296, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 296, + 477 + ], + "type": "text", + "content": "Multiscale enhancement. To preserve fine spatial details, we incorporate a Laplacian pyramid " + }, + { + "bbox": [ + 55, + 369, + 296, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\boldsymbol{X})" + }, + { + "bbox": [ + 55, + 369, + 296, + 477 + ], + "type": "text", + "content": " that guides the diffusion process across multiple scales. The Laplacian pyramid decomposes the generated consecutive frames into a multiscale representation of frequency components, facilitating the preservation of fine details critical for high-quality output. The Laplacian pyramid is calculated by subtracting the upsampled version of the downsampled frames from the original frames:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 63, + 484, + 295, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 484, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 63, + 484, + 295, + 510 + ], + "type": "interline_equation", + "content": "\\mathcal {P} (\\boldsymbol {X}) = \\left\\{\\boldsymbol {X} _ {l} - \\mathbf {U} \\left(\\mathbf {D} \\left(\\boldsymbol {X} _ {l}\\right)\\right) \\right\\} _ {l = 1} ^ {L}, \\boldsymbol {X} _ {l} = \\mathbf {U} \\left(\\mathbf {D} \\left(\\boldsymbol {X} _ {l - 1}\\right)\\right) \\tag {15}", + "image_path": "003dfb6830e6c832c4432ff7dcd01474faf56955ccf38d9003efd737b6556328.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 510, + 296, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 570 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 510, + 296, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 55, + 510, + 296, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 510, + 296, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 55, + 510, + 296, + 570 + ], + "type": "text", + "content": " are the upsampling operator and downsampling operator, respectively, and " + }, + { + "bbox": [ + 55, + 510, + 296, + 570 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 510, + 296, + 570 + ], + "type": "text", + "content": " is the number of pyramid levels. This multiscale decomposition effectively guides the diffusion process across different frequency bands, ensuring preservation of high-frequency details in reconstruction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 581, + 135, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 135, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 135, + 594 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 600, + 186, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 186, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 186, + 612 + ], + "type": "text", + "content": "4.1. Implementation details" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "Quantization levels. As introduced in Section 3.2, SFR frames are accumulations over predefined time intervals " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\Delta \\tau" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": ". Given that the spikes are binary, the value range of SFR frames is restricted to " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\left\\{\\frac{u}{\\Delta\\tau}|u\\in \\mathbb{N}\\cap [0,\\Delta \\tau ]\\right\\}" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": ". For instance, we set " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\Delta \\tau = 10" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": " in our experiments to achieve sub-milliseconds reconstruction, and the corresponding quantization levels are " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "0,0.1,0.2,\\ldots ,0.9,1" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": ", leading to the soft quantization function visualized in Figure 3(a)." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 318, + 71, + 391, + 121 + ], + "blocks": [ + { + "bbox": [ + 318, + 71, + 391, + 121 + ], + "lines": [ + { + "bbox": [ + 318, + 71, + 391, + 121 + ], + "spans": [ + { + "bbox": [ + 318, + 71, + 391, + 121 + ], + "type": "image", + "image_path": "294efd5fab5154c5cd2f7b37ed10e6f63bedcb3f4471f9f25df13b3f3d8c1a99.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 123, + 391, + 133 + ], + "lines": [ + { + "bbox": [ + 317, + 123, + 391, + 133 + ], + "spans": [ + { + "bbox": [ + 317, + 123, + 391, + 133 + ], + "type": "text", + "content": "(a) Visualization of " + }, + { + "bbox": [ + 317, + 123, + 391, + 133 + ], + "type": "inline_equation", + "content": "q(x)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 397, + 71, + 447, + 121 + ], + "blocks": [ + { + "bbox": [ + 397, + 71, + 447, + 121 + ], + "lines": [ + { + "bbox": [ + 397, + 71, + 447, + 121 + ], + "spans": [ + { + "bbox": [ + 397, + 71, + 447, + 121 + ], + "type": "image", + "image_path": "8a13751b2479c2ec930beb8a8abf065e0e12c6360cbc8ce480ffd837c86f4001.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 136, + 554, + 179 + ], + "lines": [ + { + "bbox": [ + 313, + 136, + 554, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 136, + 554, + 179 + ], + "type": "text", + "content": "Figure 3. Visualization of the quantization operator. (a) The soft quantization function " + }, + { + "bbox": [ + 313, + 136, + 554, + 179 + ], + "type": "inline_equation", + "content": "q(x)" + }, + { + "bbox": [ + 313, + 136, + 554, + 179 + ], + "type": "text", + "content": " for SFR frames with " + }, + { + "bbox": [ + 313, + 136, + 554, + 179 + ], + "type": "inline_equation", + "content": "\\Delta \\tau = 10" + }, + { + "bbox": [ + 313, + 136, + 554, + 179 + ], + "type": "text", + "content": ". (b) Detailed demonstration of quantization effectiveness on continuous pixel values." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 448, + 72, + 500, + 121 + ], + "blocks": [ + { + "bbox": [ + 448, + 72, + 500, + 121 + ], + "lines": [ + { + "bbox": [ + 448, + 72, + 500, + 121 + ], + "spans": [ + { + "bbox": [ + 448, + 72, + 500, + 121 + ], + "type": "image", + "image_path": "b7a0dbe66ff53b0a6e5a6760042b380572b8ce57fe8f7062441cc8b590c11026.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 124, + 521, + 133 + ], + "lines": [ + { + "bbox": [ + 429, + 124, + 521, + 133 + ], + "spans": [ + { + "bbox": [ + 429, + 124, + 521, + 133 + ], + "type": "text", + "content": "(b) Quantization effectiveness" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 500, + 72, + 551, + 121 + ], + "blocks": [ + { + "bbox": [ + 500, + 72, + 551, + 121 + ], + "lines": [ + { + "bbox": [ + 500, + 72, + 551, + 121 + ], + "spans": [ + { + "bbox": [ + 500, + 72, + 551, + 121 + ], + "type": "image", + "image_path": "44762d33f751f5cbb9128a1d90d39ceef153c6e51f29f3a6e68fe6543368fd46.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 186, + 554, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 554, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 554, + 243 + ], + "type": "text", + "content": "Color casting estimation. To obtain better color casting coefficients and maintain consistency across video frames reconstruction, we average the estimations of " + }, + { + "bbox": [ + 313, + 186, + 554, + 243 + ], + "type": "inline_equation", + "content": "\\alpha_{c}" + }, + { + "bbox": [ + 313, + 186, + 554, + 243 + ], + "type": "text", + "content": " from all SFR frames in the sequence, and apply the same " + }, + { + "bbox": [ + 313, + 186, + 554, + 243 + ], + "type": "inline_equation", + "content": "\\alpha_{c}" + }, + { + "bbox": [ + 313, + 186, + 554, + 243 + ], + "type": "text", + "content": " throughout the video reconstruction." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 245, + 554, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 554, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 554, + 341 + ], + "type": "text", + "content": "Pretrained diffusion model. We implement SpikeDiff upon the pretrained weights from [28], which is adapted from Stable Diffusion v2.1 [38], selected for their strong priors and close alignment with our task. We extend the image-to-image pipeline to spikes-to-video by dividing SFR frames into small chunks, and achieve temporal consistency by leveraging high-fidelity reconstruction from time-continuous chromatic spikes with physics-based guidance." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 348, + 424, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 424, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 424, + 360 + ], + "type": "text", + "content": "4.2. Evaluation dataset" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 365, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 544 + ], + "type": "text", + "content": "As our proposed method leverages pretrained diffusion models, which contain principled priors of the real world images and videos, we qualitatively evaluate the performance of our method on a real-world dataset. To this end, we collect a series of real chromatic spikes within the spike camera. The spike camera is equipped with a Bayer-pattern color filter array and a " + }, + { + "bbox": [ + 313, + 365, + 555, + 544 + ], + "type": "inline_equation", + "content": "1000 \\times 1000" + }, + { + "bbox": [ + 313, + 365, + 555, + 544 + ], + "type": "text", + "content": " pixel spike sensor. Limited by the generation ability of pretrained diffusion models and resolution of compared methods [48], we spatially crop the chromatic spikes into a lower resolution of " + }, + { + "bbox": [ + 313, + 365, + 555, + 544 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 365, + 555, + 544 + ], + "type": "text", + "content": ", maintaining the same Bayer-pattern (RGGB). The collected dataset contains chromatic spikes from various scenes with different levels of motion and color variation. We would like to contribute this dataset to the community for further research." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 545, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 653 + ], + "type": "text", + "content": "Despite the real-captured chromatic spikes, we also employ the chromatic spikes simulator described in CSpKNet [12] and SY24 [48] to generate synthetic chromatic spikes from real-world high-frame-rate videos. We collect over 100 dynamic scenarios from X4K1000FPS [39] dataset, and generate the corresponding chromatic spikes with the simulator. The synthetic dataset is used to conduct the quantitative evaluation of our proposed method, which further demonstrates the superiority of SpikeDiff." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 660, + 440, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 440, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 440, + 673 + ], + "type": "text", + "content": "4.3. Qualitative evaluation" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "To demonstrate the superiority of our proposed method, we conduct the qualitative evaluation on chromatic spikes captured from the real world, comparing SpikeDiff with exist" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7910" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 74, + 52, + 547, + 476 + ], + "blocks": [ + { + "bbox": [ + 59, + 106, + 70, + 115 + ], + "lines": [ + { + "bbox": [ + 59, + 106, + 70, + 115 + ], + "spans": [ + { + "bbox": [ + 59, + 106, + 70, + 115 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 74, + 52, + 547, + 476 + ], + "lines": [ + { + "bbox": [ + 74, + 52, + 547, + 476 + ], + "spans": [ + { + "bbox": [ + 74, + 52, + 547, + 476 + ], + "type": "image", + "image_path": "35a9e1d5b4299d7c0661ad2935794cf35d3ba972181a0be2c42d6fa62fcfe2ee.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 476, + 555, + 521 + ], + "lines": [ + { + "bbox": [ + 55, + 476, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 476, + 555, + 521 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison on real-world captured chromatic spikes (a-c), and synthetic chromatic spikes simulated from real-world high-frame-rate videos (d). Compared methods include TFI [57], TFP [57], 3DRI [10], SJDD [11], CSpkNet [12], and SY24 [48]. Details in red / blue bounding boxes are shown on the top. All results are reconstructed from sub-millisecond spike streams (0.5ms), where most methods suffer from severe noise or extreme degradation while SpikeDiff recovers clean and visually pleasing results." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 526, + 298, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 526, + 298, + 706 + ], + "spans": [ + { + "bbox": [ + 54, + 526, + 298, + 706 + ], + "type": "text", + "content": "ing chromatic spike reconstruction methods. Our proposed method can recover high-quality color frames from sub-milisecond time intervals (10 spike frames, corresponding to " + }, + { + "bbox": [ + 54, + 526, + 298, + 706 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{ms}" + }, + { + "bbox": [ + 54, + 526, + 298, + 706 + ], + "type": "text", + "content": "), which none of the existing neural network based method achieve. We adapt and evaluate the existing chromatic spike reconstruction methods (3DRI [10], SJDD [11], CSpkNet [12], and SY24 [48]), combining the SFR estimation methods (TFI [57], TFP [57]) with demosaicking as baselines. We conduct all experiments on sub-millisecond time intervals (0.5ms) to maintain the fairness in comparison. As shown in Figure 4(a), where a fan is fast rotating in front of the checkerboard, SpikeDiff produces the most clean and visually pleasant results. In contrast, other methods, except CSpkNet and SY24, contain significantly noisy pixels as highlighted in the red and blue bounding boxes. The results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 526, + 556, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 556, + 707 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 556, + 707 + ], + "type": "text", + "content": "of CSpkNet and SY24 contain blurry or degraded artifacts. As the time intervals are strictly limited, all methods are free of motion blur naturally. As for Figure 4(b), we capture a static scene in front of the lamp with the spike camera slightly shaking, to study the reconstruction quality under hard light. Compared to others, SpikeDiff generates the most clear and sharp details in blue bounding box. And SpikeDiff produces a smoother texture under the direct illumination in the red bounding box, while most of other methods show obvious quantized artifacts. In Figure 4(c), we leverage the spike streams captured by Yang et al. [48], where SpikeDiff outperforms all other methods in both noise suppression and visual quality. We also illustrate a synthetic scene in Figure 4(d), which further demonstrates the superior reconstruction capability of SpikeDiff on chromatic spikes" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7911" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 105, + 284, + 191 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 295, + 104 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 295, + 104 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 295, + 104 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of the proposed method with existing chromatic spike reconstruction methods. The best and second-best results are highlighted in red and blue, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 105, + 284, + 191 + ], + "lines": [ + { + "bbox": [ + 66, + 105, + 284, + 191 + ], + "spans": [ + { + "bbox": [ + 66, + 105, + 284, + 191 + ], + "type": "table", + "html": "
MethodPSNR↑SSIM↑FID↓NIQE↓IL-NIQE↓
SpikeDiff (Ours)18.6940.7502.8805.17338.535
SY24 [48]13.0130.53621.47211.59278.672
SJDD [11]11.3580.32526.00911.27844.711
3DRI [10]18.5120.39016.0368.54141.816
CSpkNet [12]13.3640.6414.6995.81939.948
TFP [57]12.9090.5593.25612.00164.554
TFI [57]14.6400.5527.55812.51855.758
", + "image_path": "e5e1eae43d8f06579274a840268dde3cf4c47a6760bc13bac8d17e5e9ebc8d0c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 79, + 243, + 272, + 312 + ], + "blocks": [ + { + "bbox": [ + 55, + 198, + 296, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 296, + 243 + ], + "type": "text", + "content": "Table 2. Temporal consistency evaluation, reporting Frame CLIP Score (Frame C.S.), Interpolation Error (Inp. Err.), and Interpolation PSNR (Inp. PSNR) following the metrics from [26]. The best and second-best are highlighted in red and blue, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 79, + 243, + 272, + 312 + ], + "lines": [ + { + "bbox": [ + 79, + 243, + 272, + 312 + ], + "spans": [ + { + "bbox": [ + 79, + 243, + 272, + 312 + ], + "type": "table", + "html": "
MethodFrame C.S.↑Inp. Err.↓Inp. PSNR↑
SpikeDiff (Ours)0.9690.05725.45
SY24 [48]0.9630.04128.65
SJDD [11]0.9620.11319.34
3DRI [10]0.9650.10419.87
CSpkNet [12]0.9660.05925.27
", + "image_path": "680c40ecc00c5fddcec7a75e0aeaafdd0141ebaa70181b8658c27bf24adc41b2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 59, + 317, + 117, + 376 + ], + "blocks": [ + { + "bbox": [ + 59, + 317, + 117, + 376 + ], + "lines": [ + { + "bbox": [ + 59, + 317, + 117, + 376 + ], + "spans": [ + { + "bbox": [ + 59, + 317, + 117, + 376 + ], + "type": "image", + "image_path": "01d3436302988abaf90e5cca39c257072453296c5fa7262950208be221aaf4ea.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 376, + 105, + 384 + ], + "lines": [ + { + "bbox": [ + 71, + 376, + 105, + 384 + ], + "spans": [ + { + "bbox": [ + 71, + 376, + 105, + 384 + ], + "type": "text", + "content": "Grad. w/o" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 69, + 384, + 107, + 392 + ], + "lines": [ + { + "bbox": [ + 69, + 384, + 107, + 392 + ], + "spans": [ + { + "bbox": [ + 69, + 384, + 107, + 392 + ], + "type": "text", + "content": "Mosaicking" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 395, + 296, + 438 + ], + "lines": [ + { + "bbox": [ + 55, + 395, + 296, + 438 + ], + "spans": [ + { + "bbox": [ + 55, + 395, + 296, + 438 + ], + "type": "text", + "content": "Figure 5. Analysis of differentiable degradation operators. Generated frames receive sparse gradient from chromatic spikes without mosaicking operator. Similarly, reconstructions exhibit incorrect color distribution without the color casting operator." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 118, + 318, + 176, + 376 + ], + "blocks": [ + { + "bbox": [ + 118, + 318, + 176, + 376 + ], + "lines": [ + { + "bbox": [ + 118, + 318, + 176, + 376 + ], + "spans": [ + { + "bbox": [ + 118, + 318, + 176, + 376 + ], + "type": "image", + "image_path": "93bc0d0cc1c89a08dca55893b0206a1e132c70b73e9017b6b578760ae3625407.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 376, + 161, + 383 + ], + "lines": [ + { + "bbox": [ + 131, + 376, + 161, + 383 + ], + "spans": [ + { + "bbox": [ + 131, + 376, + 161, + 383 + ], + "type": "text", + "content": "Grad.w/" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 128, + 383, + 165, + 392 + ], + "lines": [ + { + "bbox": [ + 128, + 383, + 165, + 392 + ], + "spans": [ + { + "bbox": [ + 128, + 383, + 165, + 392 + ], + "type": "text", + "content": "Mosaicking" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 177, + 318, + 235, + 376 + ], + "blocks": [ + { + "bbox": [ + 177, + 318, + 235, + 376 + ], + "lines": [ + { + "bbox": [ + 177, + 318, + 235, + 376 + ], + "spans": [ + { + "bbox": [ + 177, + 318, + 235, + 376 + ], + "type": "image", + "image_path": "f66e8b4203c1e4768b364dda0afe21e8ccf79e0a18de712f0ea604271680de63.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 376, + 223, + 383 + ], + "lines": [ + { + "bbox": [ + 187, + 376, + 223, + 383 + ], + "spans": [ + { + "bbox": [ + 187, + 376, + 223, + 383 + ], + "type": "text", + "content": "Result w/o" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 185, + 383, + 228, + 392 + ], + "lines": [ + { + "bbox": [ + 185, + 383, + 228, + 392 + ], + "spans": [ + { + "bbox": [ + 185, + 383, + 228, + 392 + ], + "type": "text", + "content": "Color casting" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 235, + 318, + 291, + 376 + ], + "blocks": [ + { + "bbox": [ + 235, + 318, + 291, + 376 + ], + "lines": [ + { + "bbox": [ + 235, + 318, + 291, + 376 + ], + "spans": [ + { + "bbox": [ + 235, + 318, + 291, + 376 + ], + "type": "image", + "image_path": "39507c09ee4651ee4e0474177738c2f52a4e323ed4bd46a038f3234037f0c856.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 376, + 279, + 383 + ], + "lines": [ + { + "bbox": [ + 247, + 376, + 279, + 383 + ], + "spans": [ + { + "bbox": [ + 247, + 376, + 279, + 383 + ], + "type": "text", + "content": "Result w/" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 242, + 383, + 285, + 392 + ], + "lines": [ + { + "bbox": [ + 242, + 383, + 285, + 392 + ], + "spans": [ + { + "bbox": [ + 242, + 383, + 285, + 392 + ], + "type": "text", + "content": "Color casting" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 445, + 296, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 296, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 296, + 493 + ], + "type": "text", + "content": "simulated from real-world videos. Note that post-processing (e.g., color casting) cannot eliminate noise or artifacts from other methods' results. Please refer to our supplementary material for more comparison and detailed discussion." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 502, + 188, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 188, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 188, + 514 + ], + "type": "text", + "content": "4.4. Quantitative evaluation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 520, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 296, + 664 + ], + "type": "text", + "content": "As shown in Table 1, we quantitatively evaluate the performance of SpikeDiff and existing chromatic spike reconstruction methods on the synthetic dataset. Similar to our qualitative evaluation, we evaluate these methods on submillisecond time intervals (0.5ms). Our method achieves the best performance across all the metrics, including PSNR, SSIM [44], FID [13], NIQE [35], and IL-NIQE [52], which demonstrates the superiority of SpikeDiff, especially in terms of perceptual metrics that reflect visual quality. As shown in Table 2, we also evaluate the temporal consistency metrics of SpikeDiff and existing reconstruction methods, where SpikeDiff achieves the best or comparable performance." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 672, + 147, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 147, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 147, + 685 + ], + "type": "text", + "content": "4.5. Ablation study" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "To investigate the effectiveness of our proposed degradation operators, we visualize the internal results after applying" + } + ] + } + ], + "index": 21 + }, + { + "type": "table", + "bbox": [ + 335, + 93, + 531, + 152 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 92 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 92 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 92 + ], + "type": "text", + "content": "Table 3. Quantitative ablation study on the contribution of different degradation operators to reconstruction quality." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 335, + 93, + 531, + 152 + ], + "lines": [ + { + "bbox": [ + 335, + 93, + 531, + 152 + ], + "spans": [ + { + "bbox": [ + 335, + 93, + 531, + 152 + ], + "type": "table", + "html": "
MethodPSNR↑SSIM↑FID↓NIQE↓IL-NIQE↓
SpikeDiff18.6940.7502.8805.17338.535
w/o C16.7520.72912.0646.07244.661
w/o M17.4810.7234.3926.20341.356
w/o Q17.4600.7173.9386.31840.487
", + "image_path": "07be01eac1443019926966f12191b63307a58fd7a6a61cafa4f908f66e5bcf5e.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 158, + 555, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 158, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 158, + 555, + 314 + ], + "type": "text", + "content": "each operator. The mosaicking operator " + }, + { + "bbox": [ + 313, + 158, + 555, + 314 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 158, + 555, + 314 + ], + "type": "text", + "content": " converts the generated frames into Bayer-pattern images, which is the beginning step. As shown in Figure 5, without the specific backward formulation for " + }, + { + "bbox": [ + 313, + 158, + 555, + 314 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 158, + 555, + 314 + ], + "type": "text", + "content": ", the gradient map is not well-defined. The color casting operator " + }, + { + "bbox": [ + 313, + 158, + 555, + 314 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 158, + 555, + 314 + ], + "type": "text", + "content": " is responsible for color correction, which degrades the normal colored image to a greener one (which is more close to the SFR frames) as shown in Figure 5. For the soft quantization operator " + }, + { + "bbox": [ + 313, + 158, + 555, + 314 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 313, + 158, + 555, + 314 + ], + "type": "text", + "content": ", its effectiveness is demonstrated in Figure 3(b), quantizing the continuous white pixels into discrete levels, similar to the pattern of the SFR frames. As shown in Table 3, we further demonstrate the effectiveness of each degradation operator via quantitative evaluation." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 314, + 325, + 388, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 325, + 388, + 338 + ], + "spans": [ + { + "bbox": [ + 314, + 325, + 388, + 338 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 346, + 555, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 490 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 490 + ], + "type": "text", + "content": "We propose the first zero-shot chromatic spike reconstruction method, named SpikeDiff, to recover consecutive clean and high-quality colored frames from sub-millisecond spikes. SpikeDiff leverages the deep priors from pretrained diffusion models to address the extreme noise under very limited time intervals and generate visually pleasant results. SpikeDiff integrates the physics-based chromatic spikes' likelihood into the diffusion-based posterior sampling process, which bridges the domain gap between spikes and conventional images. We conduct experiments on both synthetic and real-captured data to demonstrate the superior performance of SpikeDiff over existing reconstruction methods." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 490, + 556, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 490, + 556, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 490, + 556, + 598 + ], + "type": "text", + "content": "Limitations and future work. As shown in Figure 4(b), there exist \"overexposure\" cases in the central region under extreme direct illumination. This imperfection is due to the predefined spike firing threshold, resulting in consistent spikes among these pixels. We observed that it is hard to generate ideal details without any explicit guidance. In future research, we plan to explore more degradation model designs and diffusion techniques to handle this issue and further extend the HDR capability of chromatic spike reconstruction." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 314, + 609, + 411, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 609, + 411, + 623 + ], + "spans": [ + { + "bbox": [ + 314, + 609, + 411, + 623 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "This work is supported by National Natural Science Foundation of China (Grant No. 62302019, 62088102, 62136001), Beijing Natural Science Foundation (Grant No. L233024), and Beijing Municipal Science & Technology Commission, Administrative Commission of Zhongguancun Science Park (Grant No. Z241100003524012). PKU-affiliated authors thank openbayes.com for providing computing resources." + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7912" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "text", + "content": "[1] Antoni Buades, Bartomeu Coll, Jean-Michel Morel, and Catalina Sbert. Self-similarity driven color demosaicking. IEEE TIP, 18(6):1192-1202, 2009. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 126, + 296, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 126, + 296, + 158 + ], + "spans": [ + { + "bbox": [ + 61, + 126, + 296, + 158 + ], + "type": "text", + "content": "[2] Gershon Buchsbaum. A spatial processor model for object colour perception. Journal of the Franklin institute, 310(1): 1-26, 1980. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 161, + 296, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 161, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 62, + 161, + 296, + 204 + ], + "type": "text", + "content": "[3] Yakun Chang, Chu Zhou, Yuchen Hong, Liwen Hu, Chao Xu, Tiejun Huang, and Boxin Shi. 1000 FPS HDR video with a Spike-RGB hybrid camera. In CVPR, pages 22180-22190, 2023. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 207, + 295, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 207, + 295, + 250 + ], + "spans": [ + { + "bbox": [ + 62, + 207, + 295, + 250 + ], + "type": "text", + "content": "[4] Yakun Chang, Yeliduosi Xiaokaiti, Yujia Liu, Bin Fan, Zhaojun Huang, Tiejun Huang, and Boxin Shi. Towards HDR and HFR video from rolling-mixed-bit spikings. In CVPR, pages 25117-25127, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 254, + 296, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 254, + 296, + 296 + ], + "spans": [ + { + "bbox": [ + 62, + 254, + 296, + 296 + ], + "type": "text", + "content": "[5] Shiyan Chen, Chaoteng Duan, Zhaofei Yu, Ruiqin Xiong, and Tiejun Huang. Self-supervised mutual learning for dynamic scene reconstruction of spiking camera. In IJCAI, page 2859-2866, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 300, + 296, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 300, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 62, + 300, + 296, + 331 + ], + "type": "text", + "content": "[6] Shiyan Chen, Zhaofei Yu, and Tiejun Huang. Self-supervised joint dynamic scene reconstruction and optical flow estimation for spiking camera. In AAAI, pages 350-358, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 335, + 296, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 296, + 376 + ], + "type": "text", + "content": "[7] Hyungjin Chung, Jeongsol Kim, Michael Thompson McCann, Marc Louis Klasky, and Jong Chul Ye. Diffusion Posterior Sampling for General Noisy Inverse Problems. In ICLR, 2023. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 380, + 294, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 294, + 412 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 294, + 412 + ], + "type": "text", + "content": "[8] Laurent Condat and Saleh Mosaddegh. Joint demosaicking and denoising by total variation minimization. In ICIP, pages 2781-2784, 2012. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 415, + 296, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 296, + 458 + ], + "type": "text", + "content": "[9] Tobi Delbruck et al. Frame-free dynamic digital vision. In Proc. of Intl. Symp. on Secure-Life Electronics, Advanced Electronics for Quality Life and Society, pages 21-26. Cite-seer, 2008. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 461, + 296, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 461, + 296, + 494 + ], + "spans": [ + { + "bbox": [ + 57, + 461, + 296, + 494 + ], + "type": "text", + "content": "[10] Yanchen Dong, Jing Zhao, Ruiqin Xiong, and Tiejun Huang. 3D residual interpolation for spike camera demosaicing. In ICIP, pages 1461-1465. IEEE, 2022. 2, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 496, + 296, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 496, + 296, + 539 + ], + "spans": [ + { + "bbox": [ + 57, + 496, + 296, + 539 + ], + "type": "text", + "content": "[11] Yanchen Dong, Ruiqin Xiong, Jing Zhao, Jian Zhang, Xiaopeng Fan, Shuyuan Zhu, and Tiejun Huang. Joint demosaicing and denoising for spike camera. In AAAI, pages 1582-1590, 2024. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 543, + 296, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 543, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 57, + 543, + 296, + 586 + ], + "type": "text", + "content": "[12] Yanchen Dong, Ruiqin Xiong, Jing Zhao, Jian Zhang, Xiaopeng Fan, Shuyuan Zhu, and Tiejun Huang. Learning a deep demosaicing network for spike camera with color filter array. IEEE TIP, 2024. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 589, + 294, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 589, + 294, + 620 + ], + "spans": [ + { + "bbox": [ + 57, + 589, + 294, + 620 + ], + "type": "text", + "content": "[13] DC Dowson and BV Landau. The frechet distance between multivariate normal distributions. Journal of multivariate analysis, 12(3):450-455, 1982. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 624, + 296, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 296, + 666 + ], + "type": "text", + "content": "[14] Omar A Elgendy, Abhiram Gnanasambandam, Stanley H Chan, and Jiaju Ma. Low-light demosaicking and denoising for small pixels using learned frequency selection. IEEE TCI, 7:137-150, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "type": "text", + "content": "[15] Berthy T. Feng, Jamie Smith, Michael Rubinstein, Huiwen Chang, Katherine L. Bouman, and William T. Freeman. Score-Based Diffusion Models as Principled Priors for Inverse Imaging. In ICCV, 2023. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[16] Eric R Fossum, Jiaju Ma, and Saleh Masoodian. Quanta image sensor: Concepts and progress. Advanced Photon Counting Techniques X, 9858:985805, 2016. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 107, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 138 + ], + "type": "text", + "content": "[17] Eric R Fossum, Jiaju Ma, Saleh Masoodian, Leo Anzagira, and Rachel Zizza. The quanta image sensor: Every photon counts. Sensors, 16(8):1260, 2016. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 140, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 171 + ], + "type": "text", + "content": "[18] Michael Gharbi, Gaurav Chaurasia, Sylvain Paris, and Frédo Durand. Deep joint demosaicking and denoising. ACM TOG, 35(6):1-12, 2016. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 173, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 215 + ], + "type": "text", + "content": "[19] Felix Heide, Mushfiqur Rouf, Matthias B Hullin, Bjorn Labitzke, Wolfgang Heidrich, and Andreas Kolb. High-quality computational imaging through simple lenses. ACM TOG, 32 (5):1-14, 2013. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 217, + 553, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 247 + ], + "type": "text", + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 251, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 293 + ], + "type": "text", + "content": "[21] Tiejun Huang, Yajing Zheng, Zhaofei Yu, Rui Chen, Yuan Li, Ruiqin Xiong, Lei Ma, Junwei Zhao, Siwei Dong, Lin Zhu, et al. " + }, + { + "bbox": [ + 316, + 251, + 553, + 293 + ], + "type": "inline_equation", + "content": "1000\\times" + }, + { + "bbox": [ + 316, + 251, + 553, + 293 + ], + "type": "text", + "content": " faster camera and machine vision with ordinary devices. Engineering, 25:110-119, 2022. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 294, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 326 + ], + "type": "text", + "content": "[22] Siddhant Jain, Daniel Watson, Eric Tabellion, Aleksander Hołyński, Ben Poole, and Janne Kontkanen. Video Interpolation with Diffusion Models. In CVPR, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 327, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 369 + ], + "type": "text", + "content": "[23] Daisuke Kiku, Yusuke Monno, Masayuki Tanaka, and Masatoshi Okutomi. Beyond color difference: Residual interpolation for color image demosaicking. IEEE TIP, 25(3): 1288-1300, 2016. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 372, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 403 + ], + "type": "text", + "content": "[24] Filippos Kokkinos and Stamatios Lefkimmiatis. Deep image demosaicking using a cascade of convolutional residual denoising networks. In ECCV, pages 303-319, 2018. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 405, + 555, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 405, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 317, + 405, + 555, + 458 + ], + "type": "text", + "content": "[25] Juan Antonio Leñero-Bardallo, Teresa Serrano-Gotarredona, and Bernabé Linares-Barranco. A 3.6 " + }, + { + "bbox": [ + 317, + 405, + 555, + 458 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 317, + 405, + 555, + 458 + ], + "type": "text", + "content": "s latency asynchronous frame-free event-driven dynamic-vision-sensor. IEEE Journal of Solid-State Circuits, 46(6):1443-1455, 2011. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 460, + 555, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 555, + 491 + ], + "type": "text", + "content": "[26] Xinui Li, Chao Ma, Xiaokang Yang, and Ming-Hsuan Yang. VidToMe: Video token merging for zero-shot video editing. In CVPR, pages 7486-7495, 2024. 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 494, + 555, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 494, + 555, + 535 + ], + "spans": [ + { + "bbox": [ + 317, + 494, + 555, + 535 + ], + "type": "text", + "content": "[27] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A " + }, + { + "bbox": [ + 317, + 494, + 555, + 535 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 317, + 494, + 555, + 535 + ], + "type": "text", + "content": " 120 dB " + }, + { + "bbox": [ + 317, + 494, + 555, + 535 + ], + "type": "inline_equation", + "content": "15\\mu s" + }, + { + "bbox": [ + 317, + 494, + 555, + 535 + ], + "type": "text", + "content": " latency asynchronous temporal contrast vision sensor. IEEE Journal of Solid-State Circuits, 43(2):566-576, 2008. 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 537, + 555, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 555, + 581 + ], + "type": "text", + "content": "[28] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Yu Qiao, Wanli Ouyang, and Chao Dong. Diff-BIR: Toward blind image restoration with generative diffusion prior. In ECCV, pages 430-448. Springer, 2024. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 582, + 555, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 555, + 612 + ], + "type": "text", + "content": "[29] Lin Liu, Xu Jia, Jianzhuang Liu, and Qi Tian. Joint demosaicing and denoising with self guidance. In CVPR, pages 2240-2249, 2020. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 614, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 553, + 646 + ], + "type": "text", + "content": "[30] Yunfan Lu, Yijie Xu, Wenzong Ma, Weiyu Guo, and Hui Xiong. Event camera demosaicing via swim transformer and pixel-focus loss. In CVPR, pages 1095-1105, 2024. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 647, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 679 + ], + "type": "text", + "content": "[31] Sizhuo Ma, Varun Sundar, Paul Mos, Claudio Bruschini, Edoardo Charbon, and Mohit Gupta. Seeing photons in color. ACM TOG, 42(4):1-16, 2023. 2, 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "text", + "content": "[32] Julien Mairal, Michael Elad, and Guillermo Sapiro. Sparse representation for color image restoration. IEEE TIP, 17(1): 53-69, 2007. 3" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7913" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "type": "text", + "content": "[33] Julien Mairal, Francis Bach, Jean Ponce, Guillermo Sapiro, and Andrew Zisserman. Non-local sparse models for image restoration. In ICCV, pages 2272-2279, 2009." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 108, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 108, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 108, + 296, + 140 + ], + "type": "text", + "content": "[34] Henrique S Malvar, Li-wei He, and Ross Cutler. High-quality linear interpolation for demosaicing of bayer-patterned color images. In ICASSP, pages iii-485, 2004. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 142, + 296, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 142, + 296, + 175 + ], + "spans": [ + { + "bbox": [ + 56, + 142, + 296, + 175 + ], + "type": "text", + "content": "[35] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal Processing Letters, 20(3):209-212, 2012. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 177, + 296, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 177, + 296, + 208 + ], + "spans": [ + { + "bbox": [ + 56, + 177, + 296, + 208 + ], + "type": "text", + "content": "[36] Ibrahim Pekkucuksen and Yucel Altunbasak. Gradient based threshold free color filter array interpolation. In ICIP, pages 137-140, 2010. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 211, + 296, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 211, + 296, + 233 + ], + "spans": [ + { + "bbox": [ + 56, + 211, + 296, + 233 + ], + "type": "text", + "content": "[37] Reza Pournaghi and Xiaolin Wu. Coded acquisition of high frame rate video. IEEE TIP, 23(12):5670-5682, 2014. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 235, + 296, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 235, + 296, + 278 + ], + "spans": [ + { + "bbox": [ + 56, + 235, + 296, + 278 + ], + "type": "text", + "content": "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 280, + 296, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 280, + 296, + 302 + ], + "spans": [ + { + "bbox": [ + 56, + 280, + 296, + 302 + ], + "type": "text", + "content": "[39] Hyeonjun Sim, Jihyong Oh, and Munchurl Kim. XVFI: extreme video frame interpolation. In ICCV, 2021. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 304, + 296, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 296, + 346 + ], + "type": "text", + "content": "[40] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 350, + 296, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 296, + 392 + ], + "type": "text", + "content": "[41] Daniel Stanley Tan, Wei-Yang Chen, and Kai-Lung Hua. DeepDemosaicking: Adaptive image demosaicking via multiple deep fully convolutional networks. IEEE TIP, 27(5): 2408-2419, 2018. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 395, + 296, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 395, + 296, + 427 + ], + "spans": [ + { + "bbox": [ + 56, + 395, + 296, + 427 + ], + "type": "text", + "content": "[42] Runjie Tan, Kai Zhang, Wangmeng Zuo, and Lei Zhang. Color image demosaicking via deep residual learning. In ICME, pages 793-798, 2017. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 430, + 296, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 430, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 430, + 296, + 460 + ], + "type": "text", + "content": "[43] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-Shot Image Restoration Using Denoising Diffusion Null-Space Model. In ICLR, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 464, + 296, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 464, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 56, + 464, + 296, + 496 + ], + "type": "text", + "content": "[44] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 498, + 296, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 296, + 531 + ], + "type": "text", + "content": "[45] Bennett Wilburn, Neel Joshi, Vaibhav Vaish, Marc Levoy, and Mark Horowitz. High-speed videography using a dense camera array. CVPR, 2004. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 533, + 296, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 533, + 296, + 565 + ], + "spans": [ + { + "bbox": [ + 56, + 533, + 296, + 565 + ], + "type": "text", + "content": "[46] Wenzhu Xing and Karen Egiazarian. End-to-end learning for joint image demosaicing, denoising and super-resolution. In CVPR, pages 3507-3516, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 567, + 296, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 296, + 610 + ], + "type": "text", + "content": "[47] Senyan Xu, Zhijing Sun, Jiaying Zhu, Yurui Zhu, Xueyang Fu, and Zheng-Jun Zha. Demosaicformer: Coarse-to-fine demosaicing network for hybrids camera. In CVPR, pages 1126-1135, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 613, + 296, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 296, + 655 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 296, + 655 + ], + "type": "text", + "content": "[48] Siqi Yang, Zhaojun Huang, Yakun Chang, Bin Fan, Zhaofei Yu, and Boxin Shi. Real-data-driven 2000 FPS color video from mosaicked chromatic spikes. In ECCV, 2024. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 658, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 296, + 712 + ], + "type": "text", + "content": "[49] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling Up to Excellence: Practicing Model Scaling for Photo-Realistic Image Restoration In the Wild. In CVPR, 2024. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 464 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "text", + "content": "[50] Guoshen Yu, Guillermo Sapiro, and Stephane Mallat. Solving inverse problems with piecewise linear estimators: From gaussian mixture models to structured sparsity. IEEE TIP, 21 (5):2481-2499, 2011. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 118, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 555, + 183 + ], + "type": "text", + "content": "[51] Jie Zhang, Jonathan Newman, Zeguan Wang, Yong Qian, Pedro Feliciano-Ramos, Wei Guo, Takato Honda, Zhe Sage Chen, Changyang Linghu, Ralph Etienne-Cummings, et al. Pixel-wise programmability enables dynamic high-snr cameras for high-speed microscopy. Nature Communications, 15 (1):4480, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 185, + 555, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 555, + 217 + ], + "type": "text", + "content": "[52] Lin Zhang, Lei Zhang, and Alan C Bovik. A feature-enriched completely blind image quality evaluator. IEEE TIP, 24(8): 2579-2591, 2015. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 219, + 555, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 555, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 555, + 261 + ], + "type": "text", + "content": "[53] Zhao Zhang, Huan Zheng, Richang Hong, Mingliang Xu, Shuicheng Yan, and Meng Wang. Deep color consistent network for low-light image enhancement. In CVPR, pages 1899-1908, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 263, + 555, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 555, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 555, + 306 + ], + "type": "text", + "content": "[54] Jing Zhao, Ruiqin Xiong, Hangfan Liu, Jian Zhang, and Tiejun Huang. Spk2Imgnet: Learning to reconstruct dynamic scene from continuous spike stream. In CVPR, pages 11996-12005, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 308, + 555, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 555, + 351 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 555, + 351 + ], + "type": "text", + "content": "[55] Jing Zhao, Ruiqin Xiong, Jiyu Xie, Boxin Shi, Zhaofei Yu, Wen Gao, and Tiejun Huang. Reconstructing clear image for high-speed motion scene with a retina-inspired spike camera. IEEE TCI, 8:12-27, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 353, + 555, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 353, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 353, + 555, + 396 + ], + "type": "text", + "content": "[56] Yajing Zheng, Lingxiao Zheng, Zhaofei Yu, Boxin Shi, Yonghong Tian, and Tiejun Huang. High-speed image reconstruction through short-term plasticity for spiking cameras. In CVPR, pages 6358-6367, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 398, + 555, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 398, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 398, + 555, + 430 + ], + "type": "text", + "content": "[57] Lin Zhu, Siwei Dong, Tiejun Huang, and Yonghong Tian. A retina-inspired sampling method for visual texture reconstruction. In ICME, pages 1432-1437, 2019. 2, 4, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 432, + 555, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 432, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 432, + 555, + 464 + ], + "type": "text", + "content": "[58] Lin Zhu, Siwei Dong, Jianing Li, Tiejun Huang, and Yonghong Tian. Retina-like visual image reconstruction via spiking neural model. In CVPR, pages 1438-1446, 2020. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7914" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SpikePack_ Enhanced Information Flow in Spiking Neural Networks with High Hardware Compatibility/f9abc2f2-d2b5-43b8-9f56-aa36381da767_content_list.json b/2025/SpikePack_ Enhanced Information Flow in Spiking Neural Networks with High Hardware Compatibility/f9abc2f2-d2b5-43b8-9f56-aa36381da767_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a5a136b3a0165074e6126f731872809447873f8b --- /dev/null +++ b/2025/SpikePack_ Enhanced Information Flow in Spiking Neural Networks with High Hardware Compatibility/f9abc2f2-d2b5-43b8-9f56-aa36381da767_content_list.json @@ -0,0 +1,1764 @@ +[ + { + "type": "text", + "text": "SpikePack: Enhanced Information Flow in Spiking Neural Networks with High Hardware Compatibility", + "text_level": 1, + "bbox": [ + 153, + 128, + 843, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Guobin Shen $^{1,2,*}$ ; Jindong Li $^{1,2,*}$ ; Tenglong Li $^{1,2}$ , Dongcheng Zhao $^{2,3}$ , Yi Zeng $^{1,2,3,*}$", + "bbox": [ + 169, + 202, + 833, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ BrainCog Lab, CASIA, $^{2}$ Beijing Institute of AI Safety and Governance, $^{3}$ Center for Long-term Artificial Intelligence $\\{\\text { shenguobin } 2021, \\text { zhaodongcheng } 2016, \\text { litenglong } 2023, \\text { lijindong } 2022, \\text { yi.zeng } \\} @ia.ac.cn$", + "bbox": [ + 102, + 222, + 893, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Spiking Neural Networks (SNNs) hold promise for energy-efficient, biologically inspired computing. We identify substantial information loss during spike transmission, linked to temporal dependencies in traditional Leaky Integrate-and-Fire (LIF) neurons—a key factor potentially limiting SNN performance. Existing SNN architectures also underutilize modern GPUs, constrained by single-bit spike storage and isolated weight-spike operations that restrict computational efficiency. We introduce SpikePack, a neuron model designed to reduce transmission loss while preserving essential features like membrane potential reset and leaky integration. SpikePack achieves constant $\\mathcal{O}(1)$ time and space complexity, enabling efficient parallel processing on GPUs and also supporting serial inference on existing SNN hardware accelerators. Compatible with standard Artificial Neural Network (ANN) architectures, SpikePack facilitates near-lossless ANN-to-SNN conversion across various networks. Experimental results on tasks such as image classification, detection, and segmentation show SpikePack achieves significant gains in accuracy and efficiency for both directly trained and converted SNNs over state-of-the-art models. Tests on FPGA-based platforms further confirm cross-platform flexibility, delivering high performance and enhanced sparsity. By enhancing information flow and rethinking SNN-ANN integration, SpikePack advances efficient SNN deployment across diverse hardware platforms.", + "bbox": [ + 89, + 324, + 485, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 742, + 220, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Spiking Neural Networks (SNNs) [32] have emerged as a promising paradigm for energy-efficient and biologically inspired computing [54]. By emulating the discrete spike-based communication of biological neurons, SNNs offer potential advantages in terms of low-power consumption and event-driven processing, which are particularly ap-", + "bbox": [ + 89, + 768, + 483, + 861 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/038754d1520d9a2966c129331e0df06799fc0ac4cf1a8367476c0802c4bd785a.jpg", + "image_caption": [ + "Figure 1. Performance of SpikePack on Spikeformer-8-512 across different time steps (T), showing error rate $(\\%)$ , time (ep/min), and memory (MB/img). Reference model uses LIF neurons (Ref, T=4)." + ], + "image_footnote": [], + "bbox": [ + 517, + 292, + 903, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pealing for deployment on neuromorphic hardware [22-24, 34, 36].", + "bbox": [ + 511, + 549, + 906, + 579 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite these advantages, SNNs still face significant challenges that impede their widespread application in complex tasks such as image classification [28, 38, 46, 55, 58], object detection [26, 30, 51], and natural language processing [39, 49]. Notably, their performance often lags behind that of Artificial Neural Networks (ANNs). One key reason we have identified is the substantial information loss that occurs during spike transmission, particularly associated with the temporal dependencies inherent in traditional neuron models like the Leaky Integrate-and-Fire (LIF) neuron [3]. This information degradation can limit the network's ability to capture and transmit critical features, thereby hindering overall performance.", + "bbox": [ + 509, + 580, + 908, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Moreover, existing SNN architectures have not fully exploited the capabilities of modern General-Purpose Graphics Processing Units (GPGPUs). The reliance on single-bit spike representations [40] and isolated weight-spike operations [30, 42] leads to inefficient utilization of parallelism, resulting in low training efficiency and slow inference speeds. This inefficiency not only hampers the practicality of SNNs but also complicates their deployment across", + "bbox": [ + 509, + 779, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 875, + 220, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23385", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "diverse hardware platforms [2, 20].", + "bbox": [ + 89, + 90, + 323, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although there has been considerable research into neuromorphic hardware [37] and the development of various SNN accelerators [9, 23], SNNs have not become mainstream. This is partly due to their subpar performance compared to ANNs and their incompatibility with modern ANN architectures [8, 58]. This incompatibility arises not only from the reliance on discrete spikes but also from the temporal dependencies in spiking neurons, which require SNN-specific network designs and training methods. Consequently, adapting ANN models and techniques to SNNs requires complex modifications, limiting SNNs' ability to fully leverage advancements in ANN architectures and optimization.", + "bbox": [ + 89, + 107, + 483, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose SpikePack, a novel neuron model designed to reduce information loss during the transition from pre-synaptic to post-synaptic spikes, minimizing degradation associated with temporal dependencies in traditional neuron models. SpikePack enhances information flow within SNNs while achieving $\\mathcal{O}(1)$ time and space complexity with respect to time steps, enabling efficient time-parallel training and inference on GPUs, as shown in Figure 1. SpikePack also preserves essential biological characteristics, such as membrane potential reset and leaky integration, ensuring biological plausibility.", + "bbox": [ + 89, + 303, + 483, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, SpikePack is compatible with modern ANN architectures, allowing for near-lossless ANN-to-SNN conversion and preserving the inherent sparsity of spike-based computations. This compatibility enables the integration of advanced ANN models within the SNN framework, improving performance across a variety of tasks.", + "bbox": [ + 89, + 484, + 483, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 109, + 577, + 434, + 590 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce SpikePack, a neuron model that minimizes information loss from pre-synaptic to post-synaptic spikes. SpikePack achieves a balance between computational efficiency and biological fidelity in SNNs.", + "- By achieving $\\mathcal{O}(1)$ time and space complexity, SpikePack preserves essential neural dynamics, enabling efficient, biologically relevant behavior. As shown in Figure 2, SpikePack supports direct training, eliminating the need for complex temporal unfolding and enabling more efficient gradient-based optimization. Its compatibility with modern ANN architectures further supports near-lossless ANN-to-SNN conversion while maintaining the inherent sparsity of spike-based computations.", + "- Extensive experiments on image classification, object detection, and segmentation tasks showcase significant improvements over state-of-the-art methods. Additional testing on SNN hardware accelerators further validates the generality and efficiency of our approach." + ], + "bbox": [ + 89, + 595, + 483, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By addressing the fundamental issues of information loss and hardware inefficiency, SpikePack represents a signifi", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cant advancement in the practical deployment of SNNs. It not only enhances the computational capabilities of SNNs but also ensures that these improvements are accessible across various hardware platforms. This work not only brings us closer to realizing the full potential of neuromorphic computing in real-world scenarios but also offers new insights into bridging SNNs and ANNs.", + "bbox": [ + 511, + 90, + 905, + 196 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/04f964ad3c44a3fec9430836307812c9a47f9f42eb5e97c12974b5d93f5a73b6.jpg", + "image_caption": [ + "Figure 2. Forward and backward computation in SpikePack, showing both serial and parallel modes. Parallel computation uses $s_{\\mathrm{zip}}^l$ for efficient global potential $(v_g^l)$ calculation." + ], + "image_footnote": [], + "bbox": [ + 516, + 209, + 903, + 392 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 465, + 653, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our work builds on advancements in three areas of SNNs: information transmission efficiency, efficient training methods, and compatibility with ANNs.", + "bbox": [ + 511, + 491, + 905, + 536 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Information Transmission Efficiency Traditional neuron models like the LIF neuron rely on binary spikes, which limit information capacity. To address this, ternary spike neurons [12] and burst-based LIFB neurons [40] expand spike representations, allowing richer information flow while retaining energy efficiency. Other approaches, such as integer-valued neurons [30] and rectified membrane potentials [11], aim to reduce quantization errors and mitigate degradation in deep SNNs. Although these methods improve information transmission, they often add model complexity. Our SpikePack offers a simpler approach by computing membrane potentials before spikes are generated, effectively enhancing information flow without increasing computational demands.", + "bbox": [ + 511, + 537, + 905, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Efficient Training Methods Training efficiency is a core challenge in SNNs due to the non-differentiability of spike operations, and weight-spike computations, all of which limit hardware utilization. While Backpropagation Through Time (BPTT) [46] has enabled deep SNN training, it incurs high memory and time costs. Approaches like Online Training Through Time (OTTT) [47] and Spatial Learning Through Time (SLTT) [33] reduce memory overhead by prioritizing critical temporal interactions, while Temporal Reversible SNNs (T-RevSNN) [14] leverage reversible ar", + "bbox": [ + 511, + 750, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23386", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "chitectures to lower memory costs. However, these methods primarily optimize training efficiency without addressing the fundamental limitations of binary spikes and underutilized hardware. In contrast, SpikePack inherently supports time-parallel processing, achieving $\\mathcal{O}(1)$ complexity with respect to time steps, thus improving training efficiency and enabling high inference efficiency across diverse hardware.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Compatibility with ANN Architectures Leveraging ANN advancements within the SNN framework is challenging due to fundamental differences in activation dynamics. Traditional ANN-to-SNN conversion approaches rely on rate coding, often requiring many time steps and recalibration [26]. Recently, methods such as Spatio-Temporal Approximation [17] and Expectation Compensation [16] have enabled SNN adaptations of Transformer architectures by approximating non-linear interactions. While effective, these methods are often complex and architecture-specific. By contrast, SpikePack enables near-lossless conversion across various ANN architectures with minimal adjustments, facilitating direct integration with modern ANN models while preserving SNN sparsity and efficiency.", + "bbox": [ + 89, + 198, + 482, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While previous works have addressed aspects of information flow, training efficiency, and ANN compatibility, they often require complex modifications. SpikePack provides a unified, streamlined solution that enhances information flow, supports efficient training, and enables seamless integration with ANN architectures, promoting scalable SNN deployment across diverse hardware.", + "bbox": [ + 89, + 407, + 482, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 527, + 225, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we analyze the limitations of LIF neurons, focusing on their information transmission inefficiencies and computational limitations on GPGPUs. These limitations stem from both information loss during spike transmission and inefficient hardware utilization. To address these challenges, we propose SpikePack, a novel neuron model designed to preserve critical information and support efficient parallel processing on modern hardware. We also provide a theoretical foundation for SpikePack through analysis of information transmission, followed by an explanation of its computational efficiency.", + "bbox": [ + 89, + 551, + 482, + 719 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a47adf05d864ef918462a953ffa9af8c9e7f0cc18e0ddb19e745453d387aad08.jpg", + "image_caption": [ + "Figure 3. Spike sequence compression into integer $s_{\\mathrm{zip}}^l$ for efficient computation of $v_g^l$ without decompression." + ], + "image_footnote": [], + "bbox": [ + 109, + 731, + 465, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Limitations of LIF Neurons", + "text_level": 1, + "bbox": [ + 513, + 90, + 761, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The LIF neuron model is widely used in SNNs due to its simplicity and biological inspiration. However, it suffers from two key limitations that hinder its effectiveness: low information retention during spike transmission and inefficient utilization of modern parallel processing hardware.", + "bbox": [ + 511, + 112, + 903, + 188 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1. Low Information Capacity in Spike Transmission", + "text_level": 1, + "bbox": [ + 511, + 195, + 901, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LIF neurons generate output spikes based on the membrane potential at each discrete time step. However, since the membrane potential is determined by the combined effects of the historical input sequence and the decay factor, the model exhibits certain limitations in fully integrating input information. This means that each spiking decision is based on an incomplete view of the input sequence, causing substantial information loss.", + "bbox": [ + 511, + 214, + 903, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For a given layer $l$ in an SNN, let $\\mathbf{S}^{l - 1}\\in \\{0,1\\}^{N\\times T}$ denote the binary spike matrix from the previous layer, where $N$ is the number of pre-synaptic neurons, and $T$ is the number of time steps. The weight matrix of layer $l$ is represented as $\\mathbf{W}^l\\in \\mathbb{R}^{M\\times N}$ , where $M$ is the number of neurons in the current layer. For our analysis, we focus on a single neuron in this layer, represented by the weight vector $\\mathbf{w}^l = \\mathbf{W}_{i,\\cdot}^l$ . Here, we refer to the output spike sequence for this neuron as $\\mathbf{s}^l\\in \\{0,1\\}^{1\\times T}$ .", + "bbox": [ + 511, + 334, + 905, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The membrane potential $v_{t}$ of a LIF neuron updates according to the rule given in Eq. (1):", + "bbox": [ + 511, + 472, + 903, + 502 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nv _ {t} ^ {l} = \\frac {1}{\\tau} v _ {t - 1} ^ {l} + \\mathbf {w} ^ {l} \\cdot \\mathbf {s} _ {t} ^ {l - 1} - \\theta \\cdot s _ {t - 1} ^ {l}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 512, + 903, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tau$ is the membrane time constant, $\\mathbf{s}_t^{l - 1}$ represents input spikes at time $t$ , and $\\theta$ denotes the firing threshold. A spike is generated when $v_{t}$ exceeds the threshold $\\theta$ , after which the membrane potential is reset to facilitate subsequent spiking dynamics as shown in Eq. (2):", + "bbox": [ + 511, + 551, + 905, + 630 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ns _ {t} ^ {l} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} v _ {t} > \\theta , \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 638, + 903, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since $v_{t}^{l}$ is recursively dependent on $v_{t-1}^{l}$ , each spiking decision is based on partially integrated information from the input sequence. This recursive structure imposes inherent limitations on the mutual information that can be preserved between the input and output, leading to significant information loss during spike transmission. As a result, the ability of the model to fully encode and utilize temporal dependencies in the input signal is constrained.", + "bbox": [ + 511, + 691, + 905, + 814 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.2. Inefficient Hardware Utilization", + "text_level": 1, + "bbox": [ + 511, + 821, + 782, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Another significant limitation of LIF neurons lies in their inefficient utilization of modern parallel processing hardware. The recursive nature of the membrane potential update, as defined in Eq. (1), inherently restricts parallelism, as each", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23387", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "time step must be processed sequentially, resulting in a time complexity of $\\mathcal{O}(T)$ , where $T$ is the number of time steps. Furthermore, the need to store individual spikes for each time step occupies separate integer or floating-point units in memory, leading to inefficient memory utilization and increased space complexity of $\\mathcal{O}(T)$ . These inefficiencies are especially problematic for hardware architectures optimized for parallel computation.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition, the LIF model requires repeated spike-weight operations at every time step. As each input at each time step consists of a single spike value, the effective computation is limited to spike-weight multiplications, thereby underutilizing the General Matrix Multiplication (GeMM) capabilities of modern hardware. During training, the reliance on surrogate gradient approximations [46] to handle the non-differentiable spiking functions further exacerbates computational inefficiency, increasing both the computational overhead and overall complexity.", + "bbox": [ + 89, + 210, + 482, + 363 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. SpikePack", + "text_level": 1, + "bbox": [ + 89, + 371, + 194, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address these challenges, we introduce SpikePack, a neuron model designed to preserve information capacity across the input sequence while supporting efficient parallel computation. SpikePack aggregates information across the entire input sequence into a global membrane potential, denoted as $v_{g}^{l}$ for layer $l$ . This global aggregation enhances information flow and maximizes mutual information between inputs and outputs.", + "bbox": [ + 89, + 393, + 482, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The global membrane potential $v_{g}^{l}$ is computed as:", + "bbox": [ + 109, + 513, + 444, + 531 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nv _ {g} ^ {l} = \\mathbf {w} ^ {l} \\mathbf {S} ^ {l - 1} \\mathbf {q}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 232, + 539, + 480, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{q} = [\\tau^{T - 1},\\tau^{T - 2},\\dots ,\\tau^0 ]^\\top$ applies the influence of the membrane time constant $\\tau$ across time steps. This formulation enables $v_{g}^{l}$ to integrate information from the entire input sequence $\\mathbf{S}^{l - 1}$ , resulting in improved information flow from pre-synaptic to post-synaptic neurons.", + "bbox": [ + 89, + 568, + 482, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After aggregating information into $v_{g}^{l}$ , SpikePack generates the output spike sequence $\\mathbf{S}^{l}$ through a decoding process. The membrane potential $v_{t}^{l}$ is updated as in Eq. (4):", + "bbox": [ + 89, + 647, + 482, + 695 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nv _ {t} ^ {l} = v _ {t - 1} ^ {l} - \\theta_ {t} \\cdot s _ {t - 1} ^ {l}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 702, + 480, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta_{t} = \\frac{\\theta}{\\tau^{t - T}}$ represents a dynamic threshold that adapts over time. The initial membrane potential is set to $v_0^l = v_g^l$ . The spike generation condition is defined as in Eq. (5):", + "bbox": [ + 89, + 731, + 482, + 779 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {t} ^ {l} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} v _ {t} ^ {l} > \\theta_ {t}, \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 209, + 787, + 482, + 828 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Improved Information Capacity in SpikePack", + "text_level": 1, + "bbox": [ + 89, + 835, + 447, + 852 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The initial global membrane potential $v_{g}^{l}$ aggregates information across the entire input sequence, thereby enhancing the mutual information between the input $\\mathbf{S}^{l - 1}$ and the", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "output spike sequence $\\mathbf{s}^l$ . Assuming binary spikes with independent Bernoulli distributions and Gaussian weights, $v_{g}^{l}$ approximates a Gaussian distribution with variance as shown in Eq. (6):", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {v _ {g} ^ {l}} ^ {2} = \\sigma^ {2} N p (1 - p) \\left(\\sum_ {t = 1} ^ {T} q _ {t}\\right) ^ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 162, + 903, + 208 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p$ represents the probability of an input spike, and $q_{t} = \\tau^{t - 1}$", + "bbox": [ + 511, + 213, + 903, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The mutual information between $\\mathbf{S}^{l - 1}$ and $\\mathbf{s}^l$ for SpikePack can thus be approximated as in Eq. (7):", + "bbox": [ + 511, + 242, + 903, + 273 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {\\mathrm {S P}} ^ {l} = \\frac {1}{2} \\log_ {2} \\left(\\frac {1 2 \\sigma_ {v _ {g} ^ {l}} ^ {2}}{\\theta^ {2}}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 284, + 903, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This result, derived theoretically and validated empirically, demonstrates that SpikePack achieves greater information retention across a range of configurations for $N$ and $T$ , affirming its superior transmission capacity.", + "bbox": [ + 511, + 337, + 903, + 397 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For a more detailed theoretical derivation of the mutual information $I(\\mathbf{S}_{\\mathrm{in}}, \\mathbf{s}_{\\mathrm{out}})$ between pre-synaptic and postsynaptic spikes in both the LIF and SpikePack models, please refer to Appendix A. Empirical simulations, as illustrated in Figure 4, confirm that SpikePack consistently achieves higher mutual information across various values of $N$ and $T$ , further substantiating its enhanced information transmission capability.", + "bbox": [ + 511, + 398, + 905, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. Parallel Computation and Hardware Utilization", + "text_level": 1, + "bbox": [ + 511, + 526, + 893, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The SpikePack model leverages the compressed input structure $v_{g}^{l}$ , making its operations hardware-friendly and well-suited for parallel computation. As shown in Figure 3, since", + "bbox": [ + 511, + 545, + 903, + 590 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a4045be74fd289235c8cd33ebc363fe732390e08cc000c6db1555361fcbea53e.jpg", + "image_caption": [ + "Figure 4. Empirical comparison of mutual information between SpikePack and LIF neurons over varying $T$ (time steps) and $N$ (pre-synaptic neurons). SpikePack demonstrates higher information retention across configurations." + ], + "image_footnote": [], + "bbox": [ + 545, + 617, + 874, + 843 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23388", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "all spike information is aggregated into $v_{g}^{l}$ , both time and space complexity are reduced from $\\mathcal{O}(T)$ to $\\mathcal{O}(1)$ , enabling efficient use of parallel processing on modern GPUs.", + "bbox": [ + 89, + 90, + 482, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The packed representation of input spikes enables highly efficient matrix multiplication operations, thereby optimizing computational performance. The global membrane potential $v_{g}^{l}$ is computed as in Eq. (8):", + "bbox": [ + 89, + 136, + 482, + 199 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nv _ {g} ^ {l} = \\mathbf {w} ^ {l} \\mathbf {s} _ {\\mathrm {z i p}} ^ {l - 1}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 238, + 209, + 482, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{s}_{\\mathrm{zip}}^{l - 1} = \\mathbf{S}^{l - 1}\\mathbf{q}$ . This formulation compresses the presynaptic spike matrix into a bitwise representation, facilitating optimized GeMM operations and enhancing computational efficiency.", + "bbox": [ + 89, + 241, + 482, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Utilizing this compressed representation, the total output spike count over $T$ time steps can be derived directly from $v_{g}^{l}$ without iterative updates, as in Eq. (9):", + "bbox": [ + 89, + 303, + 483, + 349 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {\\mathrm {z i p}} ^ {l} = \\left\\lceil \\frac {v _ {g} ^ {l}}{\\theta} \\left. \\right\\rfloor_ {\\tau}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 359, + 482, + 404 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this representation, each bit in $s_{\\mathrm{zip}}^{l}$ indicates whether an output spike $s_t^l$ is generated at time step $t$ , effectively compressing the spike sequence. This compact form aligns with the serial spike generation in Eqs. (4) and (5), preserving spiking behavior without per-step computations and significantly reducing computational and memory demands.", + "bbox": [ + 89, + 414, + 482, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By treating the membrane time constant $\\tau$ as a form of quantization, SpikePack adapts to various hardware configurations, balancing precision and efficiency. For instance, $\\tau = 2$ results in uniform quantization, while $\\tau \\neq 2$ enables non-uniform quantization, allowing the model to adjust its computational footprint based on hardware constraints.", + "bbox": [ + 89, + 508, + 483, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.3. Efficient Gradient Propagation", + "text_level": 1, + "bbox": [ + 89, + 606, + 352, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Figure 2, SpikePack enhances computational efficiency by simplifying gradient propagation, eliminating the need for BPTT. Instead, gradients are computed directly with respect to the compressed input structure $\\mathbf{s}_{\\mathrm{zip}}^{l-1}$ , which encapsulates the entire sequence of input spikes in a single compressed representation. This method significantly reduces memory consumption and computational complexity during training by removing the requirement to unroll across time s", + "bbox": [ + 89, + 625, + 482, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The gradient of the loss $\\mathcal{L}$ with respect to $\\mathbf{s}_{\\mathrm{zip}}^{l - 1}$ is computed as shown in Eq. (10):", + "bbox": [ + 89, + 760, + 482, + 791 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\mathcal {L}}{\\partial \\mathbf {s} _ {\\mathrm {z i p}} ^ {l - 1}} = \\frac {\\partial \\mathcal {L}}{\\partial \\mathbf {s} _ {\\mathrm {z i p}} ^ {l}} \\frac {\\partial \\mathbf {s} _ {\\mathrm {z i p}} ^ {l}}{\\partial v _ {g} ^ {l}} \\frac {\\partial v _ {g} ^ {l}}{\\partial \\mathbf {s} _ {\\mathrm {z i p}} ^ {l - 1}} \\approx \\frac {\\partial \\mathcal {L}}{\\partial \\mathbf {s} _ {\\mathrm {z i p}} ^ {l}} \\frac {\\mathbf {w} ^ {l}}{\\theta}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 801, + 482, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we approximate $\\frac{\\partial[x]_{\\tau}}{\\partial x} \\approx 1$ , effectively bypassing the non-differentiable quantization step. This direct gradient path allows for efficient, memory-saving training", + "bbox": [ + 89, + 852, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "aligned with SpikePack's compressed and parallel computation model.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By leveraging this streamlined gradient computation, SpikePack enables gradient propagation without costly temporal dependencies, as required in traditional SNN models. This design is inherently compatible with hardware architectures that support vectorized and parallel computation, such as SIMD instructions and systolic arrays.", + "bbox": [ + 511, + 122, + 903, + 213 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 511, + 229, + 638, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate SpikePack on tasks including image classification, object detection, and semantic segmentation, comparing its performance with state-of-the-art SNN models, neuron designs, and ANN-to-SNN conversion methods. We also conduct ablation studies to assess the impact of key parameters, demonstrating the versatility and efficiency of SpikePack across various datasets.", + "bbox": [ + 511, + 253, + 906, + 361 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 372, + 702, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct experiments on both static and neuromorphic datasets to thoroughly assess our model. For image classification, we use ImageNet dataset [4]; for object detection, the COCO 2017 dataset [27]; and for semantic segmentation, the ADE20K dataset [57]. To evaluate event-based performance, we use neuromorphic datasets including CIFAR10-DVS [21], DVS-Gesture [1], and N-Caltech101 [35]. Experiments are implemented in PyTorch and run on NVIDIA A100 GPUs, with a default membrane time constant of $\\tau = 2$ . Additional experimental details are provided in Appendix B.", + "bbox": [ + 511, + 395, + 905, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Image and Neuromorphic Data Classification", + "text_level": 1, + "bbox": [ + 511, + 573, + 895, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate SpikePack on the ImageNet dataset and compare its performance with other SNN models and neuron designs.", + "bbox": [ + 511, + 595, + 905, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison with Other Neuron Models We benchmark SpikePack against several neuron models, including LIF, LIFB [40], PSN [10], DSGM [41], and GLIF [52]. For fair evaluation, experiments are conducted using the SEW-ResNet [8] and Spikeformer [59] architectures. Model scales and time steps are adjusted to achieve comparable performance metrics across setups.", + "bbox": [ + 511, + 642, + 905, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 5 illustrates the efficiency of SpikePack, showcasing its balance between accuracy and computational cost, quantified by Synaptic Operations (SOPs). SOPs measure the overall spiking activity and computational workload, defined as the product of the firing rate, operations per time step, and the number of time steps $T$ .", + "bbox": [ + 511, + 748, + 905, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compared to other neuron models, SpikePack consistently achieves higher accuracy at similar or lower computational costs. For instance, within the Spikeformer architecture, SpikePack matches the accuracy of competing neuron", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23389", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/63f92c59178f3be548b0a62d98371046989a2b8bdb1bd591b3ef375a7002a904.jpg", + "image_caption": [ + "Figure 5. Comparison between SpikePack and other neuron models at different time steps on ImageNet 1k. Our method achieves higher accuracy with lower computational cost." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 475, + 287 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/60cd3ea7e76cdd0691aecb5393c163d784237d14feaedd9441149f78f00f02d3.jpg", + "table_caption": [ + "Table 1. Comparison of ANN-to-SNN conversion methods on ImageNet 1k. We report the top-1 accuracy $(\\%)$ at different time steps (T)." + ], + "table_footnote": [], + "table_body": "
ModelMethodANNT=32T=64T=128T=256
ResNet-34RMP [13]----55.65
Opt. [5]70.6433.0159.5267.5470.06
Calib. [25]70.9564.5471.1273.4574.61
SNM [44]75.6655.2862.7265.5369.31
ViT-B/32STA [17]73.3078.7282.3382.5682.79
T=5T=6T=8T=12
ResNet-34SpikePack80.7259.3274.6877.8177.92
ViT-B/32Ours77.9257.2379.2680.6880.72
ViT-L/1488.2785.5988.0088.2288.27
", + "bbox": [ + 94, + 398, + 473, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "models while requiring only $1/10$ of the SOPs. Furthermore, at equivalent SOP levels, SpikePack delivers nearly a $5\\%$ improvement in accuracy, highlighting its superior information transmission and computational efficiency. This advantage persists across varying model sizes and architectures, demonstrating the robustness and scalability of SpikePack.", + "bbox": [ + 89, + 594, + 482, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with Efficient SNN Training Methods We evaluate SpikePack against several efficient SNN training methods, including STBP-tdBN [56], SEW ResNet [8], MS ResNet [15], TEBN [7], TET [6], OTTT [47], SLTT [33], Parallel SNN [10], and T-RevSNN [14].", + "bbox": [ + 89, + 702, + 482, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Table 2, SpikePack outperforms these methods in accuracy while requiring less training time and memory. For instance, using ResNet-34 with 4 time steps, SpikePack achieves $73.4\\%$ accuracy with only 8.1 minutes per epoch and 24.3 MB memory per image, demonstrating reduced training overhead and superior performance. Moreover, experiments with Transformer-based architectures, such as Spikeformer, confirm that SpikePack main", + "bbox": [ + 89, + 779, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tains $\\mathcal{O}(1)$ time and space complexity relative to the number of time steps $T$ , achieving up to $80.1\\%$ accuracy with 8 time steps without any increase in memory or computational load as $T$ grows.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with ANN-to-SNN Conversion Methods We evaluate SpikePack for near-lossless ANN-to-SNN conversion, leveraging its compatibility with various ANN architectures as discussed in Section 3.2.2. Unlike other methods, SpikePack enables efficient conversion without requiring post-conversion training or calibration.", + "bbox": [ + 511, + 151, + 903, + 241 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Table 1, we compare SpikePack with several conversion methods, including RMP [13], Optimal (Opt.) [5], Spike Calibration (Cailb.) [25], SNM [44], and Spatio-Temporal Approximation (STA) [17]. SpikePack achieves high accuracy with as few as 6 time steps and near-lossless performance at 8 time steps—less than $1/10$ of the steps required by other methods. For example, SpikePack achieves $77.92\\%$ accuracy with ResNet-34 and $88.27\\%$ with ViT-L/14, closely matching ANN performance. Figure 6 further highlights SpikePack's ability to maintain high accuracy across different architectures with minimal time steps compared to competing methods.", + "bbox": [ + 511, + 242, + 906, + 422 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ce52f3f709fb0242fea8c3855a240c4a5e34e6a31323f041d2c8c791415275ad.jpg", + "image_caption": [ + "Figure 6. Comparison of ANN-to-SNN conversion methods at different time steps. SpikePack achieves higher accuracy with fewer time steps." + ], + "image_footnote": [], + "bbox": [ + 519, + 430, + 895, + 618 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on Neuromorphic Datasets We evaluate SpikePack on neuromorphic datasets including CIFAR10-DVS [21], DVS-Gesture [1], and N-Caltech101 [35]. Table 3 reports the accuracy at different time steps.", + "bbox": [ + 511, + 674, + 903, + 736 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SpikePack achieves competitive accuracy on neuromorphic datasets, demonstrating adaptability to event-based data. Unlike static datasets, neuromorphic datasets generally require a higher number of time steps $T$ for optimal performance—a trend consistent with prior studies [42]. These results confirm SpikePack's versatility in effectively handling both static and event-driven data.", + "bbox": [ + 511, + 734, + 903, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Object Detection", + "text_level": 1, + "bbox": [ + 511, + 848, + 679, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate SpikePack on the COCO 2017 validation dataset [27], leveraging RTMDet [31] and DINO [53] as", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23390", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5760894bd3472516f0a0b1178cb9d0f4f252fce6ecd9650fdef9312feb8464b1.jpg", + "table_caption": [ + "Table 2. Comparison of SpikePack with state-of-the-art SNN training methods on ImageNet 1k. We report the number of parameters, time steps, training time per epoch, memory usage per image, synaptic operations (SOPs), and top-1 accuracy." + ], + "table_footnote": [], + "table_body": "
MethodsArchitectureParam (M)Time stepsTraining Time (min/ep)Memory (MB/img)SOP (G)Acc (%)
STBP-tdBN [56]ResNet-3421.8629.6186.17.163.7
SEW ResNet [8]SEW-ResNet-3421.845.0224.54.467.0
SEW-ResNet-5025.6410.0596.95.467.8
MS ResNet [15]MS-ResNet-3421.8611.2267.15.769.4
TEBN [7]ResNet-3421.8416.3260.17.164.3
TET [6]SEW-ResNet-3421.8412.5221.04.468.0
Spikformer [59]Spikeformer-8-38416.8414.2580.88.670.2
Spikeformer-8-51229.7416.7767.812.973.4
Spike-driven Transformer [50]Spikeformer-8-38416.8415.4548.94.372.3
Spikeformer-8-51229.7418.8730.05.174.6
OTTT [47]ResNet-3421.8624.284.16.764.2
SLTT [33]ResNet-3421.8618.171.76.766.2
ResNet-5025.6623.4117.38.067.0
Parallel SNN [10]SEW-ResNet-1811.745.8138.7-67.6
SEW-ResNet-3421.848.3179.74.470.5
T-RevSNN [14]ResNet-1815.246.157.51.969.8
ResNet-1829.849.185.73.173.2
SpikePack (Ours)ResNet-1811.145.820.11.870.6
ResNet-3421.848.124.33.773.4
ResNet-5025.6413.553.44.178.7
Spikeformer-8-51229.729.2150.53.973.6
Spikeformer-8-51229.749.2150.57.778.4
Spikeformer-8-51229.769.2150.511.279.2
Spikeformer-8-51229.789.2150.515.180.1
", + "bbox": [ + 178, + 121, + 810, + 506 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5e187278f88469c8e2569e812fa556a82a343c794dce64157628e8ad0b5dcd27.jpg", + "table_caption": [ + "Table 3. Classification accuracy on neuromorphic datasets at different time steps $T$ ." + ], + "table_footnote": [], + "table_body": "
DatasetT=4T=6T=8T=10T=12T=16
CIFAR10-DVS [21]68.176.480.782.483.784.6
DVS-Gesture [1]94.695.596.797.496.797.4
N-Caltech101 [35]76.380.081.782.282.582.6
", + "bbox": [ + 96, + 547, + 468, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "our base architectures. To highlight the effectiveness of our approach, we compare it against other SNN-based object detection models, including Spiking-YOLO [19], Bayesian Optimization [18], Spike Calibration [25], EMSYOLO [43], Meta-SpikeFormer [51], and SpikeYOLO [30].", + "bbox": [ + 89, + 647, + 482, + 723 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4 summarizes the evaluation results. Our approach demonstrates superior mean Average Precision (mAP) while requiring fewer time steps and reducing computational cost. For instance, with DINO-r50 and just 6 time steps, SpikePack achieves an impressive $48.5\\%$ $\\mathrm{mAP@50:95}$ , surpassing previous methods with a substantial reduction in computational overhead.", + "bbox": [ + 89, + 728, + 482, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our method demonstrates competitive performance with significantly fewer parameters and computational cost. For instance, using RTMDet-m with only 67.2G SOPs, we achieve $49.1\\%$ mAP@50:95 at 8 time steps.", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Semantic Segmentation", + "text_level": 1, + "bbox": [ + 511, + 517, + 730, + 534 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5 presents the segmentation results on ADE20K, showcasing SpikePack's strong performance in dense prediction tasks like semantic segmentation. With Segformer-b2 and 10 time steps, SpikePack achieves $45.6\\%$ mIoU, outperforming prior methods while significantly reducing computational cost. These results emphasize the method's efficiency and scalability for challenging benchmarks.", + "bbox": [ + 511, + 540, + 906, + 647 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 659, + 679, + 674 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation studies to analyze the impact of the membrane time constant $\\tau$ on SpikePack's performance, evaluated on ImageNet with ResNet-34 (Figure 7). While $\\tau = 4.0$ achieves the best performance by aligning with activation distributions, $\\tau = 2.0$ is more efficient on GPGPUs, avoiding exponentiation. As a result, $\\tau = 2.0$ has been the preferred choice in prior implementations.", + "bbox": [ + 511, + 681, + 906, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Hardware Compatibility", + "text_level": 1, + "bbox": [ + 511, + 801, + 741, + 818 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Compatibility with Neuromorphic Processors As a spiking neuron, the SpikePack neuron adheres to the binary nature of spiking neurons, generating discrete spike sequences that are fully compatible with existing neuromorphic processors. With higher sparsity, SpikePack achieves superior", + "bbox": [ + 511, + 824, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23391", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a6070ca84de1082fcd3ff35ca06247693e4498068b064108b249022199170b33.jpg", + "table_caption": [ + "Table 4. Performance of object detection on COCO 2017 validation set [27]. We report the number of parameters, computational cost (SOPs), time steps, and mean Average Precision (mAP)." + ], + "table_footnote": [], + "table_body": "
ModelParam (M)SOPs (G)StepmAP@ 50(%)mAP@ 50:95(%)
Spiking-YOLO [19]10.2-3500-25.7
Bayesian Optim [18]10.2-5000-25.9
Spike Calib [25]17.1-51245.4-
EMS-YOLO[43]26.932.2450.130.1
Meta-SpikeFormer (MaskRCNN) [51]34.955.0144.0-
75.0156.4151.2-
Meta-SpikeFormer (YOLO) [51]16.838.7145.0-
16.878.6450.3-
SpikeYOLO [30]23.138.6462.345.5
48.176.1464.647.4
68.893.6466.248.9
SpikePack (Ours) w/ RTMDet-tiny [31]4.88.53655.739.0
4.811.3857.840.9
4.814.11057.941.1
SpikePack (Ours) w/ RTMDet-m [31]24.751.8650.148.5
24.767.2861.749.1
24.786.31061.949.4
SpikePack (Ours) w/ DINO-r50 [53]47.7276666.048.5
47.7359866.750.0
47.74471067.950.1
", + "bbox": [ + 94, + 136, + 480, + 483 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9a0921c05f2cf70e25aea7803573052bf74aa61f7b6b28cbf439e1ee518b560b.jpg", + "table_caption": [ + "Table 5. Performance of semantic segmentation on ADE20K [57]. We report the number of parameters, computational cost (SOPs), simulation time steps, and mIoU (\\%)." + ], + "table_footnote": [], + "table_body": "
ModelParam (M)SOPs (G)StepMIoU(%) 50(%)
Meta-SpikeFormer [51]16.524.6132.3
16.598.2433.6
58.951.7134.8
58.9204.1435.3
SpikePack (Ours) [29] w/ FCN-r5047.2256.8634.1
47.2384.4835.3
47.2476.71035.9
SpikePack (Ours) w/ Segformer-b0 [48]3.7539.3635.3
3.7551.7836.9
3.7563.91037.4
SpikePack (Ours) w/ Segformer-b2 [48]24.883.6642.8
24.8111.5844.1
24.8138.71045.6
", + "bbox": [ + 91, + 547, + 488, + 767 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2131ce71dafd674a9f07b67a018a0284493ae2ac0a25cbb85156aad6d86a6065.jpg", + "table_caption": [ + "Table 6. Latency and Energy Comparison between SpikePack and LIF neuron" + ], + "table_footnote": [], + "table_body": "
SpikePackLIF
ResNet-34ResNet-50ResNet-34ResNet-50
Latency23ms24.1 ms29.1 ms34.7 ms
Energy18.6mJ19.4 mJ23.8 mJ28.4 mJ
", + "bbox": [ + 102, + 803, + 467, + 863 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "speedup and reduced energy consumption compared to tra", + "bbox": [ + 89, + 885, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ccd0c979b02082da71833a14d7f3760142696280bb8d15209265bbb35501be54.jpg", + "image_caption": [ + "Figure 7. Ablation study on the effect of membrane time constant $\\tau$ on ImageNet classification error rate (\\%) using ResNet-34 with SpikePack." + ], + "image_footnote": [], + "bbox": [ + 521, + 93, + 898, + 239 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ditional LIF neurons. To validate this, we conduct hardware experiments comparing SpikePack and LIF neurons under identical conditions. We design a neuromorphic processor that processes binary spike inputs and synaptic weights, filtering zero elements in the spike tensor to ensure only active spikes contribute to computation. The processor includes 64 processing elements for synaptic current accumulation, 16 neuron dynamic units, and a 16-input spike detector that converts spike sequences into addresses for synaptic weight fetching. We implement the processor on xczu3eg FPGA chip running at $300\\mathrm{MHz}$ . The detailed implementation is shown in Appendix C. Using cycle-accurate simulation, we evaluate inference latency for ResNet-34 and ResNet-50 models. The results demonstrate that SpikePack achieves significantly lower latency and power consumption than traditional LIF neurons, highlighting its efficiency on neuromorphic hardware.", + "bbox": [ + 511, + 309, + 906, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Compatibility with Parallel Computing Processors Modern parallel computing processors, such as GPGPUs, NPUs, and SIMD-enabled CPUs, excel at matrix multiplication—a critical operation in neural network acceleration. The SpikePack neuron efficiently utilizes these architectures by maintaining $\\mathcal{O}(1)$ computation and memory usage across time steps. In contrast, traditional LIF neurons require $\\mathcal{O}(T)$ duplication of computational workload and additional storage for membrane potentials, resulting in increased computational overhead.", + "bbox": [ + 511, + 566, + 908, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.7. Discussion", + "text_level": 1, + "bbox": [ + 511, + 727, + 633, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our experimental results demonstrate that SpikePack significantly improves information flow and computational efficiency in SNNs. By minimizing information loss during spike transmission and supporting efficient parallel computation, SpikePack delivers superior performance across diverse tasks and datasets. Additionally, its seamless compatibility with standard ANN architectures enables near-lossless ANN-to-SNN conversion, allowing SNNs to benefit from the latest advancements in ANN models and training techniques.", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23392", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the National Natural Science Foundation of China (Grant No. 62406325).", + "bbox": [ + 89, + 114, + 485, + 146 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 161, + 187, + 176 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Arnon Amir, Brian Taba, David Berg, Timothy Melano, Jeffrey McKinstry, Carmelo Di Nolfo, Tapan Nayak, Alexander Andreopoulos, Guillaume Garreau, Marcela Mendoza, et al. A low power, fully event-based gesture recognition system. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7243-7252, 2017. 5, 6, 7, 3", + "[2] A. Carpegna, A. Savino, and S. D. Carlo. Spiker+: A framework for the generation of efficient spiking neural networks fpga accelerators for inference at the edge. IEEE Transactions on Emerging Topics in Computing, 01(01):1-15, 2024. 2", + "[3] Peter Dayan and Laurence F Abbott. Theoretical neuroscience: computational and mathematical modeling of neural systems. MIT press, 2005. 1", + "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5, 3", + "[5] Shikuang Deng and Shi Gu. Optimal conversion of conventional artificial neural networks to spiking neural networks. In International Conference on Learning Representations, 2021. 6", + "[6] Shikuang Deng, Yuhang Li, Shanghang Zhang, and Shi Gu. Temporal efficient training of spiking neural network via gradient re-weighting. arXiv preprint arXiv:2202.11946, 2022. 6, 7", + "[7] Chaoteng Duan, Jianhao Ding, Shiyan Chen, Zhaofei Yu, and Tiejun Huang. Temporal effective batch normalization in spiking neural networks. Advances in Neural Information Processing Systems, 35:34377-34390, 2022. 6, 7", + "[8] Wei Fang, Zhaofei Yu, Yanqi Chen, Tiejun Huang, Timothée Masquelier, and Yonghong Tian. Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems, 34:21056-21069, 2021. 2, 5, 6, 7", + "[9] Wei Fang, Yanqi Chen, Jianhao Ding, Zhaofei Yu, Timothee Masquelier, Ding Chen, Liwei Huang, Huihui Zhou, Guoqi Li, and Yonghong Tian. Spikingjelly: An open-source machine learning infrastructure platform for spike-based intelligence. Science Advances, 9(40):eadi1480, 2023. 2", + "[10] Wei Fang, Zhaofei Yu, Zhaokun Zhou, Ding Chen, Yanqi Chen, Zhengyu Ma, Timothée Masquelier, and Yonghong Tian. Parallel spiking neurons with high efficiency and ability to learn long-term dependencies. Advances in Neural Information Processing Systems, 36, 2024. 5, 6, 7", + "[11] Yufei Guo, Xinyi Tong, Yuanpei Chen, Liwen Zhang, Xiaode Liu, Zhe Ma, and Xuhui Huang. Recdis-snn: Rectifying membrane potential distribution for directly training spiking neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 326-335, 2022. 2" + ], + "bbox": [ + 93, + 186, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Yufei Guo, Yuanpei Chen, Xiaode Liu, Weihang Peng, Yuhan Zhang, Xuhui Huang, and Zhe Ma. Ternary spike: Learning ternary spikes for spiking neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12244-12252, 2024. 2", + "[13] Bing Han and Kaushik Roy. Deep spiking neural network: Energy efficiency through time based coding. In European conference on computer vision, pages 388-404. Springer, 2020. 6", + "[14] JiaKui Hu, Man Yao, Xuerui Qiu, Yuhong Chou, Yuxuan Cai, Ning Qiao, Yonghong Tian, XU Bo, and Guoqi Li. High-performance temporal reversible spiking neural networks with o(1) training memory and o(1) inference cost. In Forty-first International Conference on Machine Learning, 2024. 2, 6, 7", + "[15] Yifan Hu, Lei Deng, Yujie Wu, Man Yao, and Guoqi Li. Advancing spiking neural networks toward deep residual learning. IEEE Transactions on Neural Networks and Learning Systems, 2024. 6, 7", + "[16] Zihan Huang, Xinyu Shi, Zecheng Hao, Tong Bu, Jianhao Ding, Zhaofei Yu, and Tiejun Huang. Towards high-performance spiking transformers from ann to snn conversion. In ACM Multimedia 2024, 2024. 3", + "[17] Yizhou Jiang, Kunlin Hu, Tianren Zhang, Haichuan Gao, Yuqian Liu, Ying Fang, and Feng Chen. Spatio-temporal approximation: A training-free snn conversion for transformers. In The Twelfth International Conference on Learning Representations, 2024. 3, 6", + "[18] Seijoon Kim, Seongsik Park, Byunggook Na, Jongwan Kim, and Sungroh Yoon. Towards fast and accurate object detection in bio-inspired spiking neural networks through bayesian optimization. IEEE Access, 9:2633-2643, 2020. 7, 8", + "[19] Seijoon Kim, Seongsik Park, Byunggook Na, and Sungroh Yoon. Spiking-yolo: spiking neural network for energy-efficient object detection. In Proceedings of the AAAI conference on artificial intelligence, pages 11270-11277, 2020. 7, 8", + "[20] Takumi Kuwahara, Reon Oshio, Mutsumi Kimura, Renyuan Zhang, and Yasuhiko Nakashima. Fusion synapse by memristor and capacitor for spiking neuromorphic systems. Neurocomputing, 593:127792, 2024. 2", + "[21] Hongmin Li, Hanchao Liu, Xiangyang Ji, Guoqi Li, and Luping Shi. Cifar10-dvs: an event-stream dataset for object classification. Frontiers in neuroscience, 11:309, 2017. 5, 6, 7, 3", + "[22] Jindong Li, Guobin Shen, Dongcheng Zhao, Qian Zhang, and Yi Zeng. Firefly: A high-throughput hardware accelerator for spiking neural networks with efficient tsp and memory optimization. IEEE Transactions on Very Large Scale Integration (VLSI) Systems, 2023. 1", + "[23] Jindong Li, Guobin Shen, Dongcheng Zhao, Qian Zhang, and Yi Zeng. Firefly v2: Advancing hardware support for high-performance spiking neural network with a spatiotemporal fpga accelerator. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2024. 2", + "[24] Tenglong Li, Jindong Li, Guobin Shen, Dongcheng Zhao, Qian Zhang, and Yi Zeng. Firefly-s: Exploiting dual-side" + ], + "bbox": [ + 516, + 92, + 905, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23393", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sparsity for spiking neural networks acceleration with reconfigurable spatial architecture. IEEE Transactions on Circuits and Systems I: Regular Papers, 2024. 1, 4", + "[25] Yang Li and Yi Zeng. Efficient and accurate conversion of spiking neural network with burst spikes. arXiv preprint arXiv:2204.13271, 2022. 6, 7, 8", + "[26] Yang Li, Xiang He, Yiting Dong, Qingqun Kong, and Yi Zeng. Spike calibration: Fast and accurate conversion of spiking neural network for object detection and segmentation. arXiv preprint arXiv:2207.02702, 2022. 1, 3", + "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5, 6, 8, 3", + "[28] Zhuang Liu, Mingjie Sun, Tinghui Zhou, Gao Huang, and Trevor Darrell. Rethinking the value of network pruning. arXiv preprint arXiv:1810.05270, 2018. 1", + "[29] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 8", + "[30] X Luo et al. Integer-valued training and spike-driven inference snn for high-performance and energy-efficient object detection. In ECCV, 2024. 1, 2, 7, 8", + "[31] Chengqi Lyu, Wenwei Zhang, Haian Huang, Yue Zhou, Yudong Wang, Yanyi Liu, Shilong Zhang, and Kai Chen. Rtmdet: An empirical study of designing real-time object detectors. arXiv preprint arXiv:2212.07784, 2022. 6, 8", + "[32] Wolfgang Maass. Networks of spiking neurons: the third generation of neural network models. Neural networks, 10 (9):1659-1671, 1997. 1", + "[33] Qingyan Meng, Mingqing Xiao, Shen Yan, Yisen Wang, Zhouchen Lin, and Zhi-Quan Luo. Towards memory-and time-efficient backpropagation for training spiking neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6166–6176, 2023. 2, 6, 7", + "[34] Paul A Merolla, John V Arthur, Rodrigo Alvarez-Icaza, Andrew S Cassidy, Jun Sawada, Filipp Akopyan, Bryan L Jackson, Nabil Imam, Chen Guo, Yutaka Nakamura, et al. A million spiking-neuron integrated circuit with a scalable communication network and interface. Science, 345(6197):668-673, 2014. 1", + "[35] Garrick Orchard, Ajinkya Jayawant, Gregory K Cohen, and Nitish Thakor. Converting static image datasets to spiking neuromorphic datasets using saccades. Frontiers in neuroscience, 9:437, 2015. 5, 6, 7, 3", + "[36] Sandeep Pande, Fearghal Morgan, Seamus Cawley, Brian McGinley, Snaider Carrillo, Jim Harkin, and Liam McDaid. Embrace-sysc for analysis of noc-based spiking neural network architectures. In 2010 International Symposium on System on Chip, pages 139-145. IEEE, 2010. 1", + "[37] Jing Pei, Lei Deng, Sen Song, Mingguo Zhao, Youhui Zhang, Shuang Wu, Guanrui Wang, Zhe Zou, Zhenzhi Wu, Wei He, et al. Towards artificial general intelligence with" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "hybrid tianjic chip architecture. Nature, 572(7767):106-111, 2019. 2", + "[38] Guobin Shen, Dongcheng Zhao, and Yi Zeng. Backpropagation with biologically plausible spatiotemporal adjustment for training deep spiking neural networks. *Patterns*, 3(6), 2022. 1", + "[39] Guobin Shen, Dongcheng Zhao, Yiting Dong, Yang Li, Jindong Li, Kang Sun, and Yi Zeng. Astrocyte-enabled advancements in spiking neural networks for large language modeling. arXiv preprint arXiv:2312.07625, 2023. 1", + "[40] Guobin Shen, Dongcheng Zhao, and Yi Zeng. Exploiting high performance spiking neural networks with efficient spiking patterns. arXiv preprint arXiv:2301.12356, 2023. 1, 2, 5", + "[41] Guobin Shen, Dongcheng Zhao, and Yi Zeng. Exploiting nonlinear dendritic adaptive computation in training deep spiking neural networks. Neural Networks, 170:190-201, 2024. 5", + "[42] G Shen et al. Are conventional snns really efficient? a perspective from network quantization. In CVPR, 2024. 1, 6", + "[43] Qiaoyi Su, Yuhong Chou, Yifan Hu, Jianing Li, Shijie Mei, Ziyang Zhang, and Guoqi Li. Deep directly-trained spiking neural networks for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6555-6565, 2023. 7, 8", + "[44] Yuchen Wang, Malu Zhang, Yi Chen, and Hong Qu. Signed neuron with memory: Towards simple, accurate and high-efficient ann-snn conversion. In International Joint Conferences on Artificial Intelligence, 2022. 6", + "[45] Bernard Widrow and István Kollár. Quantization noise: roundoff error in digital computation, signal processing, control, and communications. Cambridge University Press, 2008. 1", + "[46] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, and Luping Shi. Spatio-temporal backpropagation for training high-performance spiking neural networks. Frontiers in neuroscience, 12:331, 2018. 1, 2, 4", + "[47] Mingqing Xiao, Qingyan Meng, Zongpeng Zhang, Di He, and Zhouchen Lin. Online training through time for spiking neural networks. Advances in neural information processing systems, 35:20717-20730, 2022. 2, 6, 7", + "[48] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in neural information processing systems, 34: 12077-12090, 2021. 8", + "[49] Xingrun Xing, Boyan Gao, Zheng Zhang, David A Clifton, Shitao Xiao, Li Du, Guoqi Li, and Jiajun Zhang. Spikellm: Scaling up spiking neural network to large language models via saliency-based spiking. arXiv preprint arXiv:2407.04752, 2024. 1", + "[50] Man Yao, Jiakui Hu, Zhaokun Zhou, Li Yuan, Yonghong Tian, Bo Xu, and Guoqi Li. Spike-driven transformer. arXiv preprint arXiv:2307.01694, 2023. 7", + "[51] Man Yao, JiaKui Hu, Tianxiang Hu, Yifan Xu, Zhaokun Zhou, Yonghong Tian, XU Bo, and Guoqi Li. Spike-driven transformer v2: Meta spiking neural network architecture inspiring the design of next-generation neuromorphic chips. In" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23394", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "The Twelfth International Conference on Learning Representations, 2024. 1, 7, 8", + "[52] Xingting Yao, Fanrong Li, Zitao Mo, and Jian Cheng. Glif: A unified gated leaky integrate-and-fire neuron for spiking neural networks. Advances in Neural Information Processing Systems, 35:32160-32171, 2022. 5", + "[53] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 6, 8", + "[54] Youhui Zhang, Peng Qu, Yu Ji, Weihao Zhang, Guangrong Gao, Guanrui Wang, Sen Song, Guoqi Li, Wenguang Chen, Weimin Zheng, et al. A system hierarchy for brain-inspired computing. Nature, 586(7829):378-384, 2020. 1", + "[55] Dongcheng Zhao, Guobin Shen, Yiting Dong, Yang Li, and Yi Zeng. Improving stability and performance of spiking neural networks through enhancing temporal consistency. Pattern Recognition, 159:111094, 2025. 1", + "[56] Hanle Zheng, Yujie Wu, Lei Deng, Yifan Hu, and Guoqi Li. Going deeper with directly-trained larger spiking neural networks. In Proceedings of the AAAI conference on artificial intelligence, pages 11062-11070, 2021. 6, 7", + "[57] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 5, 8, 3", + "[58] Zhaokun Zhou, Yuesheng Zhu, Chao He, Yaowei Wang, Shuicheng Yan, Yonghong Tian, and Li Yuan. Spikformer: When spiking neural network meets transformer. arXiv preprint arXiv:2209.15425, 2022. 1, 2, 3", + "[59] Zhaokun Zhou, Yuesheng Zhu, Chao He, Yaowei Wang, Shuicheng Yan, Yonghong Tian, and Li Yuan. Spikformer: When spiking neural network meets transformer. In The Eleventh International Conference on Learning Representations, 2023. 5, 7" + ], + "bbox": [ + 91, + 90, + 482, + 599 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23395", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SpikePack_ Enhanced Information Flow in Spiking Neural Networks with High Hardware Compatibility/f9abc2f2-d2b5-43b8-9f56-aa36381da767_model.json b/2025/SpikePack_ Enhanced Information Flow in Spiking Neural Networks with High Hardware Compatibility/f9abc2f2-d2b5-43b8-9f56-aa36381da767_model.json new file mode 100644 index 0000000000000000000000000000000000000000..af7160c401cd3f9a665e2e7a906da3ce43f69f3a --- /dev/null +++ b/2025/SpikePack_ Enhanced Information Flow in Spiking Neural Networks with High Hardware Compatibility/f9abc2f2-d2b5-43b8-9f56-aa36381da767_model.json @@ -0,0 +1,2488 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.155, + 0.13, + 0.844, + 0.177 + ], + "angle": 0, + "content": "SpikePack: Enhanced Information Flow in Spiking Neural Networks with High Hardware Compatibility" + }, + { + "type": "text", + "bbox": [ + 0.17, + 0.203, + 0.834, + 0.223 + ], + "angle": 0, + "content": "Guobin Shen\\(^{1,2,*}\\); Jindong Li\\(^{1,2,*}\\); Tenglong Li\\(^{1,2}\\), Dongcheng Zhao\\(^{2,3}\\), Yi Zeng\\(^{1,2,3,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.223, + 0.894, + 0.258 + ], + "angle": 0, + "content": "\\(^{1}\\) BrainCog Lab, CASIA, \\(^{2}\\) Beijing Institute of AI Safety and Governance, \\(^{3}\\) Center for Long-term Artificial Intelligence \\(\\{\\text { shenguobin } 2021, \\text { zhaodongcheng } 2016, \\text { litenglong } 2023, \\text { lijindong } 2022, \\text { yi.zeng } \\} @ia.ac.cn\\)" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.325, + 0.486, + 0.719 + ], + "angle": 0, + "content": "Spiking Neural Networks (SNNs) hold promise for energy-efficient, biologically inspired computing. We identify substantial information loss during spike transmission, linked to temporal dependencies in traditional Leaky Integrate-and-Fire (LIF) neurons—a key factor potentially limiting SNN performance. Existing SNN architectures also underutilize modern GPUs, constrained by single-bit spike storage and isolated weight-spike operations that restrict computational efficiency. We introduce SpikePack, a neuron model designed to reduce transmission loss while preserving essential features like membrane potential reset and leaky integration. SpikePack achieves constant \\(\\mathcal{O}(1)\\) time and space complexity, enabling efficient parallel processing on GPUs and also supporting serial inference on existing SNN hardware accelerators. Compatible with standard Artificial Neural Network (ANN) architectures, SpikePack facilitates near-lossless ANN-to-SNN conversion across various networks. Experimental results on tasks such as image classification, detection, and segmentation show SpikePack achieves significant gains in accuracy and efficiency for both directly trained and converted SNNs over state-of-the-art models. Tests on FPGA-based platforms further confirm cross-platform flexibility, delivering high performance and enhanced sparsity. By enhancing information flow and rethinking SNN-ANN integration, SpikePack advances efficient SNN deployment across diverse hardware platforms." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.743, + 0.222, + 0.759 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.77, + 0.485, + 0.862 + ], + "angle": 0, + "content": "Spiking Neural Networks (SNNs) [32] have emerged as a promising paradigm for energy-efficient and biologically inspired computing [54]. By emulating the discrete spike-based communication of biological neurons, SNNs offer potential advantages in terms of low-power consumption and event-driven processing, which are particularly ap-" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.293, + 0.905, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.486, + 0.907, + 0.542 + ], + "angle": 0, + "content": "Figure 1. Performance of SpikePack on Spikeformer-8-512 across different time steps (T), showing error rate \\((\\%)\\), time (ep/min), and memory (MB/img). Reference model uses LIF neurons (Ref, T=4)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.907, + 0.58 + ], + "angle": 0, + "content": "pealing for deployment on neuromorphic hardware [22-24, 34, 36]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.582, + 0.909, + 0.779 + ], + "angle": 0, + "content": "Despite these advantages, SNNs still face significant challenges that impede their widespread application in complex tasks such as image classification [28, 38, 46, 55, 58], object detection [26, 30, 51], and natural language processing [39, 49]. Notably, their performance often lags behind that of Artificial Neural Networks (ANNs). One key reason we have identified is the substantial information loss that occurs during spike transmission, particularly associated with the temporal dependencies inherent in traditional neuron models like the Leaky Integrate-and-Fire (LIF) neuron [3]. This information degradation can limit the network's ability to capture and transmit critical features, thereby hindering overall performance." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Moreover, existing SNN architectures have not fully exploited the capabilities of modern General-Purpose Graphics Processing Units (GPGPUs). The reliance on single-bit spike representations [40] and isolated weight-spike operations [30, 42] leads to inefficient utilization of parallelism, resulting in low training efficiency and slow inference speeds. This inefficiency not only hampers the practicality of SNNs but also complicates their deployment across" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.221, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.236, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23385" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.325, + 0.106 + ], + "angle": 0, + "content": "diverse hardware platforms [2, 20]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.108, + 0.484, + 0.302 + ], + "angle": 0, + "content": "Although there has been considerable research into neuromorphic hardware [37] and the development of various SNN accelerators [9, 23], SNNs have not become mainstream. This is partly due to their subpar performance compared to ANNs and their incompatibility with modern ANN architectures [8, 58]. This incompatibility arises not only from the reliance on discrete spikes but also from the temporal dependencies in spiking neurons, which require SNN-specific network designs and training methods. Consequently, adapting ANN models and techniques to SNNs requires complex modifications, limiting SNNs' ability to fully leverage advancements in ANN architectures and optimization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.304, + 0.484, + 0.485 + ], + "angle": 0, + "content": "To address these challenges, we propose SpikePack, a novel neuron model designed to reduce information loss during the transition from pre-synaptic to post-synaptic spikes, minimizing degradation associated with temporal dependencies in traditional neuron models. SpikePack enhances information flow within SNNs while achieving \\(\\mathcal{O}(1)\\) time and space complexity with respect to time steps, enabling efficient time-parallel training and inference on GPUs, as shown in Figure 1. SpikePack also preserves essential biological characteristics, such as membrane potential reset and leaky integration, ensuring biological plausibility." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.486, + 0.484, + 0.577 + ], + "angle": 0, + "content": "In addition, SpikePack is compatible with modern ANN architectures, allowing for near-lossless ANN-to-SNN conversion and preserving the inherent sparsity of spike-based computations. This compatibility enables the integration of advanced ANN models within the SNN framework, improving performance across a variety of tasks." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.578, + 0.435, + 0.591 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.596, + 0.483, + 0.655 + ], + "angle": 0, + "content": "- We introduce SpikePack, a neuron model that minimizes information loss from pre-synaptic to post-synaptic spikes. SpikePack achieves a balance between computational efficiency and biological fidelity in SNNs." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.657, + 0.484, + 0.791 + ], + "angle": 0, + "content": "- By achieving \\(\\mathcal{O}(1)\\) time and space complexity, SpikePack preserves essential neural dynamics, enabling efficient, biologically relevant behavior. As shown in Figure 2, SpikePack supports direct training, eliminating the need for complex temporal unfolding and enabling more efficient gradient-based optimization. Its compatibility with modern ANN architectures further supports near-lossless ANN-to-SNN conversion while maintaining the inherent sparsity of spike-based computations." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.792, + 0.484, + 0.868 + ], + "angle": 0, + "content": "- Extensive experiments on image classification, object detection, and segmentation tasks showcase significant improvements over state-of-the-art methods. Additional testing on SNN hardware accelerators further validates the generality and efficiency of our approach." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.596, + 0.484, + 0.868 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "By addressing the fundamental issues of information loss and hardware inefficiency, SpikePack represents a signifi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.198 + ], + "angle": 0, + "content": "cant advancement in the practical deployment of SNNs. It not only enhances the computational capabilities of SNNs but also ensures that these improvements are accessible across various hardware platforms. This work not only brings us closer to realizing the full potential of neuromorphic computing in real-world scenarios but also offers new insights into bridging SNNs and ANNs." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.21, + 0.905, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.404, + 0.907, + 0.449 + ], + "angle": 0, + "content": "Figure 2. Forward and backward computation in SpikePack, showing both serial and parallel modes. Parallel computation uses \\( s_{\\mathrm{zip}}^l \\) for efficient global potential \\( (v_g^l) \\) calculation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.466, + 0.655, + 0.481 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.906, + 0.537 + ], + "angle": 0, + "content": "Our work builds on advancements in three areas of SNNs: information transmission efficiency, efficient training methods, and compatibility with ANNs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.906, + 0.749 + ], + "angle": 0, + "content": "Information Transmission Efficiency Traditional neuron models like the LIF neuron rely on binary spikes, which limit information capacity. To address this, ternary spike neurons [12] and burst-based LIFB neurons [40] expand spike representations, allowing richer information flow while retaining energy efficiency. Other approaches, such as integer-valued neurons [30] and rectified membrane potentials [11], aim to reduce quantization errors and mitigate degradation in deep SNNs. Although these methods improve information transmission, they often add model complexity. Our SpikePack offers a simpler approach by computing membrane potentials before spikes are generated, effectively enhancing information flow without increasing computational demands." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Efficient Training Methods Training efficiency is a core challenge in SNNs due to the non-differentiability of spike operations, and weight-spike computations, all of which limit hardware utilization. While Backpropagation Through Time (BPTT) [46] has enabled deep SNN training, it incurs high memory and time costs. Approaches like Online Training Through Time (OTTT) [47] and Spatial Learning Through Time (SLTT) [33] reduce memory overhead by prioritizing critical temporal interactions, while Temporal Reversible SNNs (T-RevSNN) [14] leverage reversible ar" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23386" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "chitectures to lower memory costs. However, these methods primarily optimize training efficiency without addressing the fundamental limitations of binary spikes and underutilized hardware. In contrast, SpikePack inherently supports time-parallel processing, achieving \\(\\mathcal{O}(1)\\) complexity with respect to time steps, thus improving training efficiency and enabling high inference efficiency across diverse hardware." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.199, + 0.483, + 0.408 + ], + "angle": 0, + "content": "Compatibility with ANN Architectures Leveraging ANN advancements within the SNN framework is challenging due to fundamental differences in activation dynamics. Traditional ANN-to-SNN conversion approaches rely on rate coding, often requiring many time steps and recalibration [26]. Recently, methods such as Spatio-Temporal Approximation [17] and Expectation Compensation [16] have enabled SNN adaptations of Transformer architectures by approximating non-linear interactions. While effective, these methods are often complex and architecture-specific. By contrast, SpikePack enables near-lossless conversion across various ANN architectures with minimal adjustments, facilitating direct integration with modern ANN models while preserving SNN sparsity and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.409, + 0.483, + 0.514 + ], + "angle": 0, + "content": "While previous works have addressed aspects of information flow, training efficiency, and ANN compatibility, they often require complex modifications. SpikePack provides a unified, streamlined solution that enhances information flow, supports efficient training, and enables seamless integration with ANN architectures, promoting scalable SNN deployment across diverse hardware." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.528, + 0.226, + 0.545 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.483, + 0.72 + ], + "angle": 0, + "content": "In this section, we analyze the limitations of LIF neurons, focusing on their information transmission inefficiencies and computational limitations on GPGPUs. These limitations stem from both information loss during spike transmission and inefficient hardware utilization. To address these challenges, we propose SpikePack, a novel neuron model designed to preserve critical information and support efficient parallel processing on modern hardware. We also provide a theoretical foundation for SpikePack through analysis of information transmission, followed by an explanation of its computational efficiency." + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.732, + 0.466, + 0.864 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.873, + 0.483, + 0.904 + ], + "angle": 0, + "content": "Figure 3. Spike sequence compression into integer \\( s_{\\mathrm{zip}}^l \\) for efficient computation of \\( v_g^l \\) without decompression." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.763, + 0.106 + ], + "angle": 0, + "content": "3.1. Limitations of LIF Neurons" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.113, + 0.905, + 0.189 + ], + "angle": 0, + "content": "The LIF neuron model is widely used in SNNs due to its simplicity and biological inspiration. However, it suffers from two key limitations that hinder its effectiveness: low information retention during spike transmission and inefficient utilization of modern parallel processing hardware." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.196, + 0.903, + 0.211 + ], + "angle": 0, + "content": "3.1.1. Low Information Capacity in Spike Transmission" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.215, + 0.905, + 0.334 + ], + "angle": 0, + "content": "LIF neurons generate output spikes based on the membrane potential at each discrete time step. However, since the membrane potential is determined by the combined effects of the historical input sequence and the decay factor, the model exhibits certain limitations in fully integrating input information. This means that each spiking decision is based on an incomplete view of the input sequence, causing substantial information loss." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.335, + 0.906, + 0.472 + ], + "angle": 0, + "content": "For a given layer \\(l\\) in an SNN, let \\(\\mathbf{S}^{l - 1}\\in \\{0,1\\}^{N\\times T}\\) denote the binary spike matrix from the previous layer, where \\(N\\) is the number of pre-synaptic neurons, and \\(T\\) is the number of time steps. The weight matrix of layer \\(l\\) is represented as \\(\\mathbf{W}^l\\in \\mathbb{R}^{M\\times N}\\), where \\(M\\) is the number of neurons in the current layer. For our analysis, we focus on a single neuron in this layer, represented by the weight vector \\(\\mathbf{w}^l = \\mathbf{W}_{i,\\cdot}^l\\). Here, we refer to the output spike sequence for this neuron as \\(\\mathbf{s}^l\\in \\{0,1\\}^{1\\times T}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.473, + 0.905, + 0.503 + ], + "angle": 0, + "content": "The membrane potential \\( v_{t} \\) of a LIF neuron updates according to the rule given in Eq. (1):" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.513, + 0.905, + 0.542 + ], + "angle": 0, + "content": "\\[\nv _ {t} ^ {l} = \\frac {1}{\\tau} v _ {t - 1} ^ {l} + \\mathbf {w} ^ {l} \\cdot \\mathbf {s} _ {t} ^ {l - 1} - \\theta \\cdot s _ {t - 1} ^ {l}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.553, + 0.906, + 0.631 + ], + "angle": 0, + "content": "where \\(\\tau\\) is the membrane time constant, \\(\\mathbf{s}_t^{l - 1}\\) represents input spikes at time \\(t\\), and \\(\\theta\\) denotes the firing threshold. A spike is generated when \\(v_{t}\\) exceeds the threshold \\(\\theta\\), after which the membrane potential is reset to facilitate subsequent spiking dynamics as shown in Eq. (2):" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.64, + 0.905, + 0.681 + ], + "angle": 0, + "content": "\\[\ns _ {t} ^ {l} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} v _ {t} > \\theta , \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.692, + 0.906, + 0.815 + ], + "angle": 0, + "content": "Since \\( v_{t}^{l} \\) is recursively dependent on \\( v_{t-1}^{l} \\), each spiking decision is based on partially integrated information from the input sequence. This recursive structure imposes inherent limitations on the mutual information that can be preserved between the input and output, leading to significant information loss during spike transmission. As a result, the ability of the model to fully encode and utilize temporal dependencies in the input signal is constrained." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.822, + 0.783, + 0.835 + ], + "angle": 0, + "content": "3.1.2. Inefficient Hardware Utilization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Another significant limitation of LIF neurons lies in their inefficient utilization of modern parallel processing hardware. The recursive nature of the membrane potential update, as defined in Eq. (1), inherently restricts parallelism, as each" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23387" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "time step must be processed sequentially, resulting in a time complexity of \\(\\mathcal{O}(T)\\), where \\(T\\) is the number of time steps. Furthermore, the need to store individual spikes for each time step occupies separate integer or floating-point units in memory, leading to inefficient memory utilization and increased space complexity of \\(\\mathcal{O}(T)\\). These inefficiencies are especially problematic for hardware architectures optimized for parallel computation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.212, + 0.483, + 0.364 + ], + "angle": 0, + "content": "In addition, the LIF model requires repeated spike-weight operations at every time step. As each input at each time step consists of a single spike value, the effective computation is limited to spike-weight multiplications, thereby underutilizing the General Matrix Multiplication (GeMM) capabilities of modern hardware. During training, the reliance on surrogate gradient approximations [46] to handle the non-differentiable spiking functions further exacerbates computational inefficiency, increasing both the computational overhead and overall complexity." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.372, + 0.195, + 0.387 + ], + "angle": 0, + "content": "3.2. SpikePack" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.394, + 0.483, + 0.514 + ], + "angle": 0, + "content": "To address these challenges, we introduce SpikePack, a neuron model designed to preserve information capacity across the input sequence while supporting efficient parallel computation. SpikePack aggregates information across the entire input sequence into a global membrane potential, denoted as \\( v_{g}^{l} \\) for layer \\( l \\). This global aggregation enhances information flow and maximizes mutual information between inputs and outputs." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.515, + 0.445, + 0.532 + ], + "angle": 0, + "content": "The global membrane potential \\( v_{g}^{l} \\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.233, + 0.54, + 0.482, + 0.561 + ], + "angle": 0, + "content": "\\[\nv _ {g} ^ {l} = \\mathbf {w} ^ {l} \\mathbf {S} ^ {l - 1} \\mathbf {q}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.648 + ], + "angle": 0, + "content": "where \\(\\mathbf{q} = [\\tau^{T - 1},\\tau^{T - 2},\\dots ,\\tau^0 ]^\\top\\) applies the influence of the membrane time constant \\(\\tau\\) across time steps. This formulation enables \\(v_{g}^{l}\\) to integrate information from the entire input sequence \\(\\mathbf{S}^{l - 1}\\), resulting in improved information flow from pre-synaptic to post-synaptic neurons." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.648, + 0.483, + 0.696 + ], + "angle": 0, + "content": "After aggregating information into \\( v_{g}^{l} \\), SpikePack generates the output spike sequence \\( \\mathbf{S}^{l} \\) through a decoding process. The membrane potential \\( v_{t}^{l} \\) is updated as in Eq. (4):" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.703, + 0.482, + 0.722 + ], + "angle": 0, + "content": "\\[\nv _ {t} ^ {l} = v _ {t - 1} ^ {l} - \\theta_ {t} \\cdot s _ {t - 1} ^ {l}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.732, + 0.483, + 0.78 + ], + "angle": 0, + "content": "where \\(\\theta_{t} = \\frac{\\theta}{\\tau^{t - T}}\\) represents a dynamic threshold that adapts over time. The initial membrane potential is set to \\(v_0^l = v_g^l\\). The spike generation condition is defined as in Eq. (5):" + }, + { + "type": "equation", + "bbox": [ + 0.21, + 0.788, + 0.483, + 0.829 + ], + "angle": 0, + "content": "\\[\ns _ {t} ^ {l} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} v _ {t} ^ {l} > \\theta_ {t}, \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.837, + 0.449, + 0.853 + ], + "angle": 0, + "content": "3.2.1. Improved Information Capacity in SpikePack" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The initial global membrane potential \\( v_{g}^{l} \\) aggregates information across the entire input sequence, thereby enhancing the mutual information between the input \\( \\mathbf{S}^{l - 1} \\) and the" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "output spike sequence \\(\\mathbf{s}^l\\). Assuming binary spikes with independent Bernoulli distributions and Gaussian weights, \\(v_{g}^{l}\\) approximates a Gaussian distribution with variance as shown in Eq. (6):" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.163, + 0.905, + 0.209 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {v _ {g} ^ {l}} ^ {2} = \\sigma^ {2} N p (1 - p) \\left(\\sum_ {t = 1} ^ {T} q _ {t}\\right) ^ {2}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.214, + 0.905, + 0.242 + ], + "angle": 0, + "content": "where \\(p\\) represents the probability of an input spike, and \\(q_{t} = \\tau^{t - 1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.243, + 0.905, + 0.275 + ], + "angle": 0, + "content": "The mutual information between \\(\\mathbf{S}^{l - 1}\\) and \\(\\mathbf{s}^l\\) for SpikePack can thus be approximated as in Eq. (7):" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.285, + 0.905, + 0.328 + ], + "angle": 0, + "content": "\\[\nI _ {\\mathrm {S P}} ^ {l} = \\frac {1}{2} \\log_ {2} \\left(\\frac {1 2 \\sigma_ {v _ {g} ^ {l}} ^ {2}}{\\theta^ {2}}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.338, + 0.905, + 0.398 + ], + "angle": 0, + "content": "This result, derived theoretically and validated empirically, demonstrates that SpikePack achieves greater information retention across a range of configurations for \\( N \\) and \\( T \\), affirming its superior transmission capacity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.906, + 0.52 + ], + "angle": 0, + "content": "For a more detailed theoretical derivation of the mutual information \\( I(\\mathbf{S}_{\\mathrm{in}}, \\mathbf{s}_{\\mathrm{out}}) \\) between pre-synaptic and postsynaptic spikes in both the LIF and SpikePack models, please refer to Appendix A. Empirical simulations, as illustrated in Figure 4, confirm that SpikePack consistently achieves higher mutual information across various values of \\( N \\) and \\( T \\), further substantiating its enhanced information transmission capability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.527, + 0.894, + 0.542 + ], + "angle": 0, + "content": "3.2.2. Parallel Computation and Hardware Utilization" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.546, + 0.905, + 0.592 + ], + "angle": 0, + "content": "The SpikePack model leverages the compressed input structure \\( v_{g}^{l} \\), making its operations hardware-friendly and well-suited for parallel computation. As shown in Figure 3, since" + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.618, + 0.875, + 0.844 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.848, + 0.905, + 0.905 + ], + "angle": 0, + "content": "Figure 4. Empirical comparison of mutual information between SpikePack and LIF neurons over varying \\( T \\) (time steps) and \\( N \\) (pre-synaptic neurons). SpikePack demonstrates higher information retention across configurations." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "23388" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.483, + 0.137 + ], + "angle": 0, + "content": "all spike information is aggregated into \\( v_{g}^{l} \\), both time and space complexity are reduced from \\( \\mathcal{O}(T) \\) to \\( \\mathcal{O}(1) \\), enabling efficient use of parallel processing on modern GPUs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.137, + 0.483, + 0.2 + ], + "angle": 0, + "content": "The packed representation of input spikes enables highly efficient matrix multiplication operations, thereby optimizing computational performance. The global membrane potential \\( v_{g}^{l} \\) is computed as in Eq. (8):" + }, + { + "type": "equation", + "bbox": [ + 0.24, + 0.21, + 0.483, + 0.231 + ], + "angle": 0, + "content": "\\[\nv _ {g} ^ {l} = \\mathbf {w} ^ {l} \\mathbf {s} _ {\\mathrm {z i p}} ^ {l - 1}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.242, + 0.483, + 0.303 + ], + "angle": 0, + "content": "where \\(\\mathbf{s}_{\\mathrm{zip}}^{l - 1} = \\mathbf{S}^{l - 1}\\mathbf{q}\\). This formulation compresses the presynaptic spike matrix into a bitwise representation, facilitating optimized GeMM operations and enhancing computational efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.304, + 0.484, + 0.351 + ], + "angle": 0, + "content": "Utilizing this compressed representation, the total output spike count over \\( T \\) time steps can be derived directly from \\( v_{g}^{l} \\) without iterative updates, as in Eq. (9):" + }, + { + "type": "equation", + "bbox": [ + 0.232, + 0.361, + 0.483, + 0.405 + ], + "angle": 0, + "content": "\\[\ns _ {\\mathrm {z i p}} ^ {l} = \\left\\lceil \\frac {v _ {g} ^ {l}}{\\theta} \\left. \\right\\rfloor_ {\\tau}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.483, + 0.508 + ], + "angle": 0, + "content": "In this representation, each bit in \\( s_{\\mathrm{zip}}^{l} \\) indicates whether an output spike \\( s_t^l \\) is generated at time step \\( t \\), effectively compressing the spike sequence. This compact form aligns with the serial spike generation in Eqs. (4) and (5), preserving spiking behavior without per-step computations and significantly reducing computational and memory demands." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.484, + 0.599 + ], + "angle": 0, + "content": "By treating the membrane time constant \\(\\tau\\) as a form of quantization, SpikePack adapts to various hardware configurations, balancing precision and efficiency. For instance, \\(\\tau = 2\\) results in uniform quantization, while \\(\\tau \\neq 2\\) enables non-uniform quantization, allowing the model to adjust its computational footprint based on hardware constraints." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.607, + 0.354, + 0.621 + ], + "angle": 0, + "content": "3.2.3. Efficient Gradient Propagation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.625, + 0.483, + 0.759 + ], + "angle": 0, + "content": "As shown in Figure 2, SpikePack enhances computational efficiency by simplifying gradient propagation, eliminating the need for BPTT. Instead, gradients are computed directly with respect to the compressed input structure \\( \\mathbf{s}_{\\mathrm{zip}}^{l-1} \\), which encapsulates the entire sequence of input spikes in a single compressed representation. This method significantly reduces memory consumption and computational complexity during training by removing the requirement to unroll across time s" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.761, + 0.483, + 0.792 + ], + "angle": 0, + "content": "The gradient of the loss \\(\\mathcal{L}\\) with respect to \\(\\mathbf{s}_{\\mathrm{zip}}^{l - 1}\\) is computed as shown in Eq. (10):" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.802, + 0.483, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\mathcal {L}}{\\partial \\mathbf {s} _ {\\mathrm {z i p}} ^ {l - 1}} = \\frac {\\partial \\mathcal {L}}{\\partial \\mathbf {s} _ {\\mathrm {z i p}} ^ {l}} \\frac {\\partial \\mathbf {s} _ {\\mathrm {z i p}} ^ {l}}{\\partial v _ {g} ^ {l}} \\frac {\\partial v _ {g} ^ {l}}{\\partial \\mathbf {s} _ {\\mathrm {z i p}} ^ {l - 1}} \\approx \\frac {\\partial \\mathcal {L}}{\\partial \\mathbf {s} _ {\\mathrm {z i p}} ^ {l}} \\frac {\\mathbf {w} ^ {l}}{\\theta}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.853, + 0.484, + 0.903 + ], + "angle": 0, + "content": "where we approximate \\(\\frac{\\partial[x]_{\\tau}}{\\partial x} \\approx 1\\), effectively bypassing the non-differentiable quantization step. This direct gradient path allows for efficient, memory-saving training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "aligned with SpikePack's compressed and parallel computation model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.123, + 0.905, + 0.214 + ], + "angle": 0, + "content": "By leveraging this streamlined gradient computation, SpikePack enables gradient propagation without costly temporal dependencies, as required in traditional SNN models. This design is inherently compatible with hardware architectures that support vectorized and parallel computation, such as SIMD instructions and systolic arrays." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.23, + 0.64, + 0.246 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.255, + 0.907, + 0.362 + ], + "angle": 0, + "content": "We evaluate SpikePack on tasks including image classification, object detection, and semantic segmentation, comparing its performance with state-of-the-art SNN models, neuron designs, and ANN-to-SNN conversion methods. We also conduct ablation studies to assess the impact of key parameters, demonstrating the versatility and efficiency of SpikePack across various datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.373, + 0.704, + 0.389 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.396, + 0.906, + 0.563 + ], + "angle": 0, + "content": "We conduct experiments on both static and neuromorphic datasets to thoroughly assess our model. For image classification, we use ImageNet dataset [4]; for object detection, the COCO 2017 dataset [27]; and for semantic segmentation, the ADE20K dataset [57]. To evaluate event-based performance, we use neuromorphic datasets including CIFAR10-DVS [21], DVS-Gesture [1], and N-Caltech101 [35]. Experiments are implemented in PyTorch and run on NVIDIA A100 GPUs, with a default membrane time constant of \\(\\tau = 2\\). Additional experimental details are provided in Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.574, + 0.897, + 0.59 + ], + "angle": 0, + "content": "4.2. Image and Neuromorphic Data Classification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.906, + 0.642 + ], + "angle": 0, + "content": "We evaluate SpikePack on the ImageNet dataset and compare its performance with other SNN models and neuron designs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.906, + 0.748 + ], + "angle": 0, + "content": "Comparison with Other Neuron Models We benchmark SpikePack against several neuron models, including LIF, LIFB [40], PSN [10], DSGM [41], and GLIF [52]. For fair evaluation, experiments are conducted using the SEW-ResNet [8] and Spikeformer [59] architectures. Model scales and time steps are adjusted to achieve comparable performance metrics across setups." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.749, + 0.906, + 0.84 + ], + "angle": 0, + "content": "Figure 5 illustrates the efficiency of SpikePack, showcasing its balance between accuracy and computational cost, quantified by Synaptic Operations (SOPs). SOPs measure the overall spiking activity and computational workload, defined as the product of the firing rate, operations per time step, and the number of time steps \\( T \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Compared to other neuron models, SpikePack consistently achieves higher accuracy at similar or lower computational costs. For instance, within the Spikeformer architecture, SpikePack matches the accuracy of competing neuron" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23389" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.476, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.302, + 0.483, + 0.344 + ], + "angle": 0, + "content": "Figure 5. Comparison between SpikePack and other neuron models at different time steps on ImageNet 1k. Our method achieves higher accuracy with lower computational cost." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.347, + 0.483, + 0.389 + ], + "angle": 0, + "content": "Table 1. Comparison of ANN-to-SNN conversion methods on ImageNet 1k. We report the top-1 accuracy \\((\\%)\\) at different time steps (T)." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.4, + 0.475, + 0.582 + ], + "angle": 0, + "content": "
ModelMethodANNT=32T=64T=128T=256
ResNet-34RMP [13]----55.65
Opt. [5]70.6433.0159.5267.5470.06
Calib. [25]70.9564.5471.1273.4574.61
SNM [44]75.6655.2862.7265.5369.31
ViT-B/32STA [17]73.3078.7282.3382.5682.79
T=5T=6T=8T=12
ResNet-34SpikePack80.7259.3274.6877.8177.92
ViT-B/32Ours77.9257.2379.2680.6880.72
ViT-L/1488.2785.5988.0088.2288.27
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.595, + 0.483, + 0.701 + ], + "angle": 0, + "content": "models while requiring only \\(1/10\\) of the SOPs. Furthermore, at equivalent SOP levels, SpikePack delivers nearly a \\(5\\%\\) improvement in accuracy, highlighting its superior information transmission and computational efficiency. This advantage persists across varying model sizes and architectures, demonstrating the robustness and scalability of SpikePack." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Comparison with Efficient SNN Training Methods We evaluate SpikePack against several efficient SNN training methods, including STBP-tdBN [56], SEW ResNet [8], MS ResNet [15], TEBN [7], TET [6], OTTT [47], SLTT [33], Parallel SNN [10], and T-RevSNN [14]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "As shown in Table 2, SpikePack outperforms these methods in accuracy while requiring less training time and memory. For instance, using ResNet-34 with 4 time steps, SpikePack achieves \\(73.4\\%\\) accuracy with only 8.1 minutes per epoch and 24.3 MB memory per image, demonstrating reduced training overhead and superior performance. Moreover, experiments with Transformer-based architectures, such as Spikeformer, confirm that SpikePack main" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "tains \\(\\mathcal{O}(1)\\) time and space complexity relative to the number of time steps \\(T\\), achieving up to \\(80.1\\%\\) accuracy with 8 time steps without any increase in memory or computational load as \\(T\\) grows." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.152, + 0.905, + 0.242 + ], + "angle": 0, + "content": "Comparison with ANN-to-SNN Conversion Methods We evaluate SpikePack for near-lossless ANN-to-SNN conversion, leveraging its compatibility with various ANN architectures as discussed in Section 3.2.2. Unlike other methods, SpikePack enables efficient conversion without requiring post-conversion training or calibration." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.243, + 0.907, + 0.424 + ], + "angle": 0, + "content": "As shown in Table 1, we compare SpikePack with several conversion methods, including RMP [13], Optimal (Opt.) [5], Spike Calibration (Cailb.) [25], SNM [44], and Spatio-Temporal Approximation (STA) [17]. SpikePack achieves high accuracy with as few as 6 time steps and near-lossless performance at 8 time steps—less than \\(1/10\\) of the steps required by other methods. For example, SpikePack achieves \\(77.92\\%\\) accuracy with ResNet-34 and \\(88.27\\%\\) with ViT-L/14, closely matching ANN performance. Figure 6 further highlights SpikePack's ability to maintain high accuracy across different architectures with minimal time steps compared to competing methods." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.431, + 0.897, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.634, + 0.905, + 0.675 + ], + "angle": 0, + "content": "Figure 6. Comparison of ANN-to-SNN conversion methods at different time steps. SpikePack achieves higher accuracy with fewer time steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.905, + 0.737 + ], + "angle": 0, + "content": "Results on Neuromorphic Datasets We evaluate SpikePack on neuromorphic datasets including CIFAR10-DVS [21], DVS-Gesture [1], and N-Caltech101 [35]. Table 3 reports the accuracy at different time steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.736, + 0.905, + 0.842 + ], + "angle": 0, + "content": "SpikePack achieves competitive accuracy on neuromorphic datasets, demonstrating adaptability to event-based data. Unlike static datasets, neuromorphic datasets generally require a higher number of time steps \\( T \\) for optimal performance—a trend consistent with prior studies [42]. These results confirm SpikePack's versatility in effectively handling both static and event-driven data." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.681, + 0.866 + ], + "angle": 0, + "content": "4.3. Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "We evaluate SpikePack on the COCO 2017 validation dataset [27], leveraging RTMDet [31] and DINO [53] as" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23390" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 2. Comparison of SpikePack with state-of-the-art SNN training methods on ImageNet 1k. We report the number of parameters, time steps, training time per epoch, memory usage per image, synaptic operations (SOPs), and top-1 accuracy." + }, + { + "type": "table", + "bbox": [ + 0.179, + 0.122, + 0.812, + 0.507 + ], + "angle": 0, + "content": "
MethodsArchitectureParam (M)Time stepsTraining Time (min/ep)Memory (MB/img)SOP (G)Acc (%)
STBP-tdBN [56]ResNet-3421.8629.6186.17.163.7
SEW ResNet [8]SEW-ResNet-3421.845.0224.54.467.0
SEW-ResNet-5025.6410.0596.95.467.8
MS ResNet [15]MS-ResNet-3421.8611.2267.15.769.4
TEBN [7]ResNet-3421.8416.3260.17.164.3
TET [6]SEW-ResNet-3421.8412.5221.04.468.0
Spikformer [59]Spikeformer-8-38416.8414.2580.88.670.2
Spikeformer-8-51229.7416.7767.812.973.4
Spike-driven Transformer [50]Spikeformer-8-38416.8415.4548.94.372.3
Spikeformer-8-51229.7418.8730.05.174.6
OTTT [47]ResNet-3421.8624.284.16.764.2
SLTT [33]ResNet-3421.8618.171.76.766.2
ResNet-5025.6623.4117.38.067.0
Parallel SNN [10]SEW-ResNet-1811.745.8138.7-67.6
SEW-ResNet-3421.848.3179.74.470.5
T-RevSNN [14]ResNet-1815.246.157.51.969.8
ResNet-1829.849.185.73.173.2
SpikePack (Ours)ResNet-1811.145.820.11.870.6
ResNet-3421.848.124.33.773.4
ResNet-5025.6413.553.44.178.7
Spikeformer-8-51229.729.2150.53.973.6
Spikeformer-8-51229.749.2150.57.778.4
Spikeformer-8-51229.769.2150.511.279.2
Spikeformer-8-51229.789.2150.515.180.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.516, + 0.483, + 0.544 + ], + "angle": 0, + "content": "Table 3. Classification accuracy on neuromorphic datasets at different time steps \\( T \\) ." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.549, + 0.47, + 0.626 + ], + "angle": 0, + "content": "
DatasetT=4T=6T=8T=10T=12T=16
CIFAR10-DVS [21]68.176.480.782.483.784.6
DVS-Gesture [1]94.695.596.797.496.797.4
N-Caltech101 [35]76.380.081.782.282.582.6
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.648, + 0.483, + 0.724 + ], + "angle": 0, + "content": "our base architectures. To highlight the effectiveness of our approach, we compare it against other SNN-based object detection models, including Spiking-YOLO [19], Bayesian Optimization [18], Spike Calibration [25], EMSYOLO [43], Meta-SpikeFormer [51], and SpikeYOLO [30]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.729, + 0.483, + 0.835 + ], + "angle": 0, + "content": "Table 4 summarizes the evaluation results. Our approach demonstrates superior mean Average Precision (mAP) while requiring fewer time steps and reducing computational cost. For instance, with DINO-r50 and just 6 time steps, SpikePack achieves an impressive \\(48.5\\%\\) \\(\\mathrm{mAP@50:95}\\), surpassing previous methods with a substantial reduction in computational overhead." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Our method demonstrates competitive performance with significantly fewer parameters and computational cost. For instance, using RTMDet-m with only 67.2G SOPs, we achieve \\(49.1\\%\\) mAP@50:95 at 8 time steps." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.518, + 0.731, + 0.535 + ], + "angle": 0, + "content": "4.4. Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.541, + 0.907, + 0.648 + ], + "angle": 0, + "content": "Table 5 presents the segmentation results on ADE20K, showcasing SpikePack's strong performance in dense prediction tasks like semantic segmentation. With Segformer-b2 and 10 time steps, SpikePack achieves \\(45.6\\%\\) mIoU, outperforming prior methods while significantly reducing computational cost. These results emphasize the method's efficiency and scalability for challenging benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.66, + 0.681, + 0.675 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.683, + 0.907, + 0.79 + ], + "angle": 0, + "content": "We conduct ablation studies to analyze the impact of the membrane time constant \\(\\tau\\) on SpikePack's performance, evaluated on ImageNet with ResNet-34 (Figure 7). While \\(\\tau = 4.0\\) achieves the best performance by aligning with activation distributions, \\(\\tau = 2.0\\) is more efficient on GPGPUs, avoiding exponentiation. As a result, \\(\\tau = 2.0\\) has been the preferred choice in prior implementations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.742, + 0.819 + ], + "angle": 0, + "content": "4.6. Hardware Compatibility" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Compatibility with Neuromorphic Processors As a spiking neuron, the SpikePack neuron adheres to the binary nature of spiking neurons, generating discrete spike sequences that are fully compatible with existing neuromorphic processors. With higher sparsity, SpikePack achieves superior" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23391" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.486, + 0.132 + ], + "angle": 0, + "content": "Table 4. Performance of object detection on COCO 2017 validation set [27]. We report the number of parameters, computational cost (SOPs), time steps, and mean Average Precision (mAP)." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.137, + 0.482, + 0.484 + ], + "angle": 0, + "content": "
ModelParam (M)SOPs (G)StepmAP@ 50(%)mAP@ 50:95(%)
Spiking-YOLO [19]10.2-3500-25.7
Bayesian Optim [18]10.2-5000-25.9
Spike Calib [25]17.1-51245.4-
EMS-YOLO[43]26.932.2450.130.1
Meta-SpikeFormer (MaskRCNN) [51]34.955.0144.0-
75.0156.4151.2-
Meta-SpikeFormer (YOLO) [51]16.838.7145.0-
16.878.6450.3-
SpikeYOLO [30]23.138.6462.345.5
48.176.1464.647.4
68.893.6466.248.9
SpikePack (Ours) w/ RTMDet-tiny [31]4.88.53655.739.0
4.811.3857.840.9
4.814.11057.941.1
SpikePack (Ours) w/ RTMDet-m [31]24.751.8650.148.5
24.767.2861.749.1
24.786.31061.949.4
SpikePack (Ours) w/ DINO-r50 [53]47.7276666.048.5
47.7359866.750.0
47.74471067.950.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.501, + 0.484, + 0.544 + ], + "angle": 0, + "content": "Table 5. Performance of semantic segmentation on ADE20K [57]. We report the number of parameters, computational cost (SOPs), simulation time steps, and mIoU (\\%)." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.548, + 0.489, + 0.768 + ], + "angle": 0, + "content": "
ModelParam (M)SOPs (G)StepMIoU(%) 50(%)
Meta-SpikeFormer [51]16.524.6132.3
16.598.2433.6
58.951.7134.8
58.9204.1435.3
SpikePack (Ours) [29] w/ FCN-r5047.2256.8634.1
47.2384.4835.3
47.2476.71035.9
SpikePack (Ours) w/ Segformer-b0 [48]3.7539.3635.3
3.7551.7836.9
3.7563.91037.4
SpikePack (Ours) w/ Segformer-b2 [48]24.883.6642.8
24.8111.5844.1
24.8138.71045.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.772, + 0.484, + 0.8 + ], + "angle": 0, + "content": "Table 6. Latency and Energy Comparison between SpikePack and LIF neuron" + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.804, + 0.468, + 0.864 + ], + "angle": 0, + "content": "
SpikePackLIF
ResNet-34ResNet-50ResNet-34ResNet-50
Latency23ms24.1 ms29.1 ms34.7 ms
Energy18.6mJ19.4 mJ23.8 mJ28.4 mJ
" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.484, + 0.902 + ], + "angle": 0, + "content": "speedup and reduced energy consumption compared to tra" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.094, + 0.9, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.256, + 0.907, + 0.299 + ], + "angle": 0, + "content": "Figure 7. Ablation study on the effect of membrane time constant \\(\\tau\\) on ImageNet classification error rate (\\%) using ResNet-34 with SpikePack." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.31, + 0.907, + 0.567 + ], + "angle": 0, + "content": "ditional LIF neurons. To validate this, we conduct hardware experiments comparing SpikePack and LIF neurons under identical conditions. We design a neuromorphic processor that processes binary spike inputs and synaptic weights, filtering zero elements in the spike tensor to ensure only active spikes contribute to computation. The processor includes 64 processing elements for synaptic current accumulation, 16 neuron dynamic units, and a 16-input spike detector that converts spike sequences into addresses for synaptic weight fetching. We implement the processor on xczu3eg FPGA chip running at \\(300\\mathrm{MHz}\\). The detailed implementation is shown in Appendix C. Using cycle-accurate simulation, we evaluate inference latency for ResNet-34 and ResNet-50 models. The results demonstrate that SpikePack achieves significantly lower latency and power consumption than traditional LIF neurons, highlighting its efficiency on neuromorphic hardware." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.909, + 0.718 + ], + "angle": 0, + "content": "Compatibility with Parallel Computing Processors Modern parallel computing processors, such as GPGPUs, NPUs, and SIMD-enabled CPUs, excel at matrix multiplication—a critical operation in neural network acceleration. The SpikePack neuron efficiently utilizes these architectures by maintaining \\(\\mathcal{O}(1)\\) computation and memory usage across time steps. In contrast, traditional LIF neurons require \\(\\mathcal{O}(T)\\) duplication of computational workload and additional storage for membrane potentials, resulting in increased computational overhead." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.728, + 0.634, + 0.743 + ], + "angle": 0, + "content": "4.7. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Our experimental results demonstrate that SpikePack significantly improves information flow and computational efficiency in SNNs. By minimizing information loss during spike transmission and supporting efficient parallel computation, SpikePack delivers superior performance across diverse tasks and datasets. Additionally, its seamless compatibility with standard ANN architectures enables near-lossless ANN-to-SNN conversion, allowing SNNs to benefit from the latest advancements in ANN models and training techniques." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23392" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.486, + 0.147 + ], + "angle": 0, + "content": "This work is supported by the National Natural Science Foundation of China (Grant No. 62406325)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.162, + 0.188, + 0.178 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.271 + ], + "angle": 0, + "content": "[1] Arnon Amir, Brian Taba, David Berg, Timothy Melano, Jeffrey McKinstry, Carmelo Di Nolfo, Tapan Nayak, Alexander Andreopoulos, Guillaume Garreau, Marcela Mendoza, et al. A low power, fully event-based gesture recognition system. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7243-7252, 2017. 5, 6, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.273, + 0.484, + 0.342 + ], + "angle": 0, + "content": "[2] A. Carpegna, A. Savino, and S. D. Carlo. Spiker+: A framework for the generation of efficient spiking neural networks fpga accelerators for inference at the edge. IEEE Transactions on Emerging Topics in Computing, 01(01):1-15, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.344, + 0.484, + 0.386 + ], + "angle": 0, + "content": "[3] Peter Dayan and Laurence F Abbott. Theoretical neuroscience: computational and mathematical modeling of neural systems. MIT press, 2005. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.388, + 0.484, + 0.444 + ], + "angle": 0, + "content": "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.446, + 0.484, + 0.5 + ], + "angle": 0, + "content": "[5] Shikuang Deng and Shi Gu. Optimal conversion of conventional artificial neural networks to spiking neural networks. In International Conference on Learning Representations, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.502, + 0.484, + 0.558 + ], + "angle": 0, + "content": "[6] Shikuang Deng, Yuhang Li, Shanghang Zhang, and Shi Gu. Temporal efficient training of spiking neural network via gradient re-weighting. arXiv preprint arXiv:2202.11946, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.56, + 0.484, + 0.616 + ], + "angle": 0, + "content": "[7] Chaoteng Duan, Jianhao Ding, Shiyan Chen, Zhaofei Yu, and Tiejun Huang. Temporal effective batch normalization in spiking neural networks. Advances in Neural Information Processing Systems, 35:34377-34390, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.618, + 0.484, + 0.673 + ], + "angle": 0, + "content": "[8] Wei Fang, Zhaofei Yu, Yanqi Chen, Tiejun Huang, Timothée Masquelier, and Yonghong Tian. Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems, 34:21056-21069, 2021. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.675, + 0.484, + 0.745 + ], + "angle": 0, + "content": "[9] Wei Fang, Yanqi Chen, Jianhao Ding, Zhaofei Yu, Timothee Masquelier, Ding Chen, Liwei Huang, Huihui Zhou, Guoqi Li, and Yonghong Tian. Spikingjelly: An open-source machine learning infrastructure platform for spike-based intelligence. Science Advances, 9(40):eadi1480, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.484, + 0.816 + ], + "angle": 0, + "content": "[10] Wei Fang, Zhaofei Yu, Zhaokun Zhou, Ding Chen, Yanqi Chen, Zhengyu Ma, Timothée Masquelier, and Yonghong Tian. Parallel spiking neurons with high efficiency and ability to learn long-term dependencies. Advances in Neural Information Processing Systems, 36, 2024. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[11] Yufei Guo, Xinyi Tong, Yuanpei Chen, Liwen Zhang, Xiaode Liu, Zhe Ma, and Xuhui Huang. Recdis-snn: Rectifying membrane potential distribution for directly training spiking neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 326-335, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.187, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[12] Yufei Guo, Yuanpei Chen, Xiaode Liu, Weihang Peng, Yuhan Zhang, Xuhui Huang, and Zhe Ma. Ternary spike: Learning ternary spikes for spiking neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12244-12252, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.163, + 0.906, + 0.217 + ], + "angle": 0, + "content": "[13] Bing Han and Kaushik Roy. Deep spiking neural network: Energy efficiency through time based coding. In European conference on computer vision, pages 388-404. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.218, + 0.906, + 0.3 + ], + "angle": 0, + "content": "[14] JiaKui Hu, Man Yao, Xuerui Qiu, Yuhong Chou, Yuxuan Cai, Ning Qiao, Yonghong Tian, XU Bo, and Guoqi Li. High-performance temporal reversible spiking neural networks with o(1) training memory and o(1) inference cost. In Forty-first International Conference on Machine Learning, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.906, + 0.357 + ], + "angle": 0, + "content": "[15] Yifan Hu, Lei Deng, Yujie Wu, Man Yao, and Guoqi Li. Advancing spiking neural networks toward deep residual learning. IEEE Transactions on Neural Networks and Learning Systems, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.906, + 0.412 + ], + "angle": 0, + "content": "[16] Zihan Huang, Xinyu Shi, Zecheng Hao, Tong Bu, Jianhao Ding, Zhaofei Yu, and Tiejun Huang. Towards high-performance spiking transformers from ann to snn conversion. In ACM Multimedia 2024, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.906, + 0.482 + ], + "angle": 0, + "content": "[17] Yizhou Jiang, Kunlin Hu, Tianren Zhang, Haichuan Gao, Yuqian Liu, Ying Fang, and Feng Chen. Spatio-temporal approximation: A training-free snn conversion for transformers. In The Twelfth International Conference on Learning Representations, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.906, + 0.551 + ], + "angle": 0, + "content": "[18] Seijoon Kim, Seongsik Park, Byunggook Na, Jongwan Kim, and Sungroh Yoon. Towards fast and accurate object detection in bio-inspired spiking neural networks through bayesian optimization. IEEE Access, 9:2633-2643, 2020. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.906, + 0.621 + ], + "angle": 0, + "content": "[19] Seijoon Kim, Seongsik Park, Byunggook Na, and Sungroh Yoon. Spiking-yolo: spiking neural network for energy-efficient object detection. In Proceedings of the AAAI conference on artificial intelligence, pages 11270-11277, 2020. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.906, + 0.678 + ], + "angle": 0, + "content": "[20] Takumi Kuwahara, Reon Oshio, Mutsumi Kimura, Renyuan Zhang, and Yasuhiko Nakashima. Fusion synapse by memristor and capacitor for spiking neuromorphic systems. Neurocomputing, 593:127792, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.906, + 0.734 + ], + "angle": 0, + "content": "[21] Hongmin Li, Hanchao Liu, Xiangyang Ji, Guoqi Li, and Luping Shi. Cifar10-dvs: an event-stream dataset for object classification. Frontiers in neuroscience, 11:309, 2017. 5, 6, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.906, + 0.803 + ], + "angle": 0, + "content": "[22] Jindong Li, Guobin Shen, Dongcheng Zhao, Qian Zhang, and Yi Zeng. Firefly: A high-throughput hardware accelerator for spiking neural networks with efficient tsp and memory optimization. IEEE Transactions on Very Large Scale Integration (VLSI) Systems, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.906, + 0.873 + ], + "angle": 0, + "content": "[23] Jindong Li, Guobin Shen, Dongcheng Zhao, Qian Zhang, and Yi Zeng. Firefly v2: Advancing hardware support for high-performance spiking neural network with a spatiotemporal fpga accelerator. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[24] Tenglong Li, Jindong Li, Guobin Shen, Dongcheng Zhao, Qian Zhang, and Yi Zeng. Firefly-s: Exploiting dual-side" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23393" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "sparsity for spiking neural networks acceleration with reconfigurable spatial architecture. IEEE Transactions on Circuits and Systems I: Regular Papers, 2024. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.176 + ], + "angle": 0, + "content": "[25] Yang Li and Yi Zeng. Efficient and accurate conversion of spiking neural network with burst spikes. arXiv preprint arXiv:2204.13271, 2022. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[26] Yang Li, Xiang He, Yiting Dong, Qingqun Kong, and Yi Zeng. Spike calibration: Fast and accurate conversion of spiking neural network for object detection and segmentation. arXiv preprint arXiv:2207.02702, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5, 6, 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.362 + ], + "angle": 0, + "content": "[28] Zhuang Liu, Mingjie Sun, Tinghui Zhou, Gao Huang, and Trevor Darrell. Rethinking the value of network pruning. arXiv preprint arXiv:1810.05270, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.363, + 0.482, + 0.419 + ], + "angle": 0, + "content": "[29] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.461 + ], + "angle": 0, + "content": "[30] X Luo et al. Integer-valued training and spike-driven inference snn for high-performance and energy-efficient object detection. In ECCV, 2024. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.482, + 0.518 + ], + "angle": 0, + "content": "[31] Chengqi Lyu, Wenwei Zhang, Haian Huang, Yue Zhou, Yudong Wang, Yanyi Liu, Shilong Zhang, and Kai Chen. Rtmdet: An empirical study of designing real-time object detectors. arXiv preprint arXiv:2212.07784, 2022. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.52, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[32] Wolfgang Maass. Networks of spiking neurons: the third generation of neural network models. Neural networks, 10 (9):1659-1671, 1997. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.563, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[33] Qingyan Meng, Mingqing Xiao, Shen Yan, Yisen Wang, Zhouchen Lin, and Zhi-Quan Luo. Towards memory-and time-efficient backpropagation for training spiking neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6166–6176, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.648, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[34] Paul A Merolla, John V Arthur, Rodrigo Alvarez-Icaza, Andrew S Cassidy, Jun Sawada, Filipp Akopyan, Bryan L Jackson, Nabil Imam, Chen Guo, Yutaka Nakamura, et al. A million spiking-neuron integrated circuit with a scalable communication network and interface. Science, 345(6197):668-673, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[35] Garrick Orchard, Ajinkya Jayawant, Gregory K Cohen, and Nitish Thakor. Converting static image datasets to spiking neuromorphic datasets using saccades. Frontiers in neuroscience, 9:437, 2015. 5, 6, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[36] Sandeep Pande, Fearghal Morgan, Seamus Cawley, Brian McGinley, Snaider Carrillo, Jim Harkin, and Liam McDaid. Embrace-sysc for analysis of noc-based spiking neural network architectures. In 2010 International Symposium on System on Chip, pages 139-145. IEEE, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] Jing Pei, Lei Deng, Sen Song, Mingguo Zhao, Youhui Zhang, Shuang Wu, Guanrui Wang, Zhe Zou, Zhenzhi Wu, Wei He, et al. Towards artificial general intelligence with" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "hybrid tianjic chip architecture. Nature, 572(7767):106-111, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[38] Guobin Shen, Dongcheng Zhao, and Yi Zeng. Backpropagation with biologically plausible spatiotemporal adjustment for training deep spiking neural networks. *Patterns*, 3(6), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[39] Guobin Shen, Dongcheng Zhao, Yiting Dong, Yang Li, Jindong Li, Kang Sun, and Yi Zeng. Astrocyte-enabled advancements in spiking neural networks for large language modeling. arXiv preprint arXiv:2312.07625, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.286 + ], + "angle": 0, + "content": "[40] Guobin Shen, Dongcheng Zhao, and Yi Zeng. Exploiting high performance spiking neural networks with efficient spiking patterns. arXiv preprint arXiv:2301.12356, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[41] Guobin Shen, Dongcheng Zhao, and Yi Zeng. Exploiting nonlinear dendritic adaptive computation in training deep spiking neural networks. Neural Networks, 170:190-201, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.371 + ], + "angle": 0, + "content": "[42] G Shen et al. Are conventional snns really efficient? a perspective from network quantization. In CVPR, 2024. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[43] Qiaoyi Su, Yuhong Chou, Yifan Hu, Jianing Li, Shijie Mei, Ziyang Zhang, and Guoqi Li. Deep directly-trained spiking neural networks for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6555-6565, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[44] Yuchen Wang, Malu Zhang, Yi Chen, and Hong Qu. Signed neuron with memory: Towards simple, accurate and high-efficient ann-snn conversion. In International Joint Conferences on Artificial Intelligence, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[45] Bernard Widrow and István Kollár. Quantization noise: roundoff error in digital computation, signal processing, control, and communications. Cambridge University Press, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[46] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, and Luping Shi. Spatio-temporal backpropagation for training high-performance spiking neural networks. Frontiers in neuroscience, 12:331, 2018. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[47] Mingqing Xiao, Qingyan Meng, Zongpeng Zhang, Di He, and Zhouchen Lin. Online training through time for spiking neural networks. Advances in neural information processing systems, 35:20717-20730, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[48] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in neural information processing systems, 34: 12077-12090, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[49] Xingrun Xing, Boyan Gao, Zheng Zhang, David A Clifton, Shitao Xiao, Li Du, Guoqi Li, and Jiajun Zhang. Spikellm: Scaling up spiking neural network to large language models via saliency-based spiking. arXiv preprint arXiv:2407.04752, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[50] Man Yao, Jiakui Hu, Zhaokun Zhou, Li Yuan, Yonghong Tian, Bo Xu, and Guoqi Li. Spike-driven transformer. arXiv preprint arXiv:2307.01694, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[51] Man Yao, JiaKui Hu, Tianxiang Hu, Yifan Xu, Zhaokun Zhou, Yonghong Tian, XU Bo, and Guoqi Li. Spike-driven transformer v2: Meta spiking neural network architecture inspiring the design of next-generation neuromorphic chips. In" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23394" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "The Twelfth International Conference on Learning Representations, 2024. 1, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[52] Xingting Yao, Fanrong Li, Zitao Mo, and Jian Cheng. Glif: A unified gated leaky integrate-and-fire neuron for spiking neural networks. Advances in Neural Information Processing Systems, 35:32160-32171, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[53] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[54] Youhui Zhang, Peng Qu, Yu Ji, Weihao Zhang, Guangrong Gao, Guanrui Wang, Sen Song, Guoqi Li, Wenguang Chen, Weimin Zheng, et al. A system hierarchy for brain-inspired computing. Nature, 586(7829):378-384, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[55] Dongcheng Zhao, Guobin Shen, Yiting Dong, Yang Li, and Yi Zeng. Improving stability and performance of spiking neural networks through enhancing temporal consistency. Pattern Recognition, 159:111094, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[56] Hanle Zheng, Yujie Wu, Lei Deng, Yifan Hu, and Guoqi Li. Going deeper with directly-trained larger spiking neural networks. In Proceedings of the AAAI conference on artificial intelligence, pages 11062-11070, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[57] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 5, 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.531 + ], + "angle": 0, + "content": "[58] Zhaokun Zhou, Yuesheng Zhu, Chao He, Yaowei Wang, Shuicheng Yan, Yonghong Tian, and Li Yuan. Spikformer: When spiking neural network meets transformer. arXiv preprint arXiv:2209.15425, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.532, + 0.482, + 0.601 + ], + "angle": 0, + "content": "[59] Zhaokun Zhou, Yuesheng Zhu, Chao He, Yaowei Wang, Shuicheng Yan, Yonghong Tian, and Li Yuan. Spikformer: When spiking neural network meets transformer. In The Eleventh International Conference on Learning Representations, 2023. 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.601 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23395" + } + ] +] \ No newline at end of file diff --git a/2025/SpikePack_ Enhanced Information Flow in Spiking Neural Networks with High Hardware Compatibility/f9abc2f2-d2b5-43b8-9f56-aa36381da767_origin.pdf b/2025/SpikePack_ Enhanced Information Flow in Spiking Neural Networks with High Hardware Compatibility/f9abc2f2-d2b5-43b8-9f56-aa36381da767_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cfbea5cf2ffc089a41e8aa505d8c6855181e424f --- /dev/null +++ b/2025/SpikePack_ Enhanced Information Flow in Spiking Neural Networks with High Hardware Compatibility/f9abc2f2-d2b5-43b8-9f56-aa36381da767_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:397434f8f7df1cf937d2efa3504bc4d68c0303d7a05ac4b21124e68fefc46247 +size 857415 diff --git a/2025/SpikePack_ Enhanced Information Flow in Spiking Neural Networks with High Hardware Compatibility/full.md b/2025/SpikePack_ Enhanced Information Flow in Spiking Neural Networks with High Hardware Compatibility/full.md new file mode 100644 index 0000000000000000000000000000000000000000..76e8de4d234b7b94ce329082aa0a794ff29fe688 --- /dev/null +++ b/2025/SpikePack_ Enhanced Information Flow in Spiking Neural Networks with High Hardware Compatibility/full.md @@ -0,0 +1,357 @@ +# SpikePack: Enhanced Information Flow in Spiking Neural Networks with High Hardware Compatibility + +Guobin Shen $^{1,2,*}$ ; Jindong Li $^{1,2,*}$ ; Tenglong Li $^{1,2}$ , Dongcheng Zhao $^{2,3}$ , Yi Zeng $^{1,2,3,*}$ + +$^{1}$ BrainCog Lab, CASIA, $^{2}$ Beijing Institute of AI Safety and Governance, $^{3}$ Center for Long-term Artificial Intelligence $\{\text { shenguobin } 2021, \text { zhaodongcheng } 2016, \text { litenglong } 2023, \text { lijindong } 2022, \text { yi.zeng } \} @ia.ac.cn$ + +# Abstract + +Spiking Neural Networks (SNNs) hold promise for energy-efficient, biologically inspired computing. We identify substantial information loss during spike transmission, linked to temporal dependencies in traditional Leaky Integrate-and-Fire (LIF) neurons—a key factor potentially limiting SNN performance. Existing SNN architectures also underutilize modern GPUs, constrained by single-bit spike storage and isolated weight-spike operations that restrict computational efficiency. We introduce SpikePack, a neuron model designed to reduce transmission loss while preserving essential features like membrane potential reset and leaky integration. SpikePack achieves constant $\mathcal{O}(1)$ time and space complexity, enabling efficient parallel processing on GPUs and also supporting serial inference on existing SNN hardware accelerators. Compatible with standard Artificial Neural Network (ANN) architectures, SpikePack facilitates near-lossless ANN-to-SNN conversion across various networks. Experimental results on tasks such as image classification, detection, and segmentation show SpikePack achieves significant gains in accuracy and efficiency for both directly trained and converted SNNs over state-of-the-art models. Tests on FPGA-based platforms further confirm cross-platform flexibility, delivering high performance and enhanced sparsity. By enhancing information flow and rethinking SNN-ANN integration, SpikePack advances efficient SNN deployment across diverse hardware platforms. + +# 1. Introduction + +Spiking Neural Networks (SNNs) [32] have emerged as a promising paradigm for energy-efficient and biologically inspired computing [54]. By emulating the discrete spike-based communication of biological neurons, SNNs offer potential advantages in terms of low-power consumption and event-driven processing, which are particularly ap- + +![](images/038754d1520d9a2966c129331e0df06799fc0ac4cf1a8367476c0802c4bd785a.jpg) +Figure 1. Performance of SpikePack on Spikeformer-8-512 across different time steps (T), showing error rate $(\%)$ , time (ep/min), and memory (MB/img). Reference model uses LIF neurons (Ref, T=4). + +pealing for deployment on neuromorphic hardware [22-24, 34, 36]. + +Despite these advantages, SNNs still face significant challenges that impede their widespread application in complex tasks such as image classification [28, 38, 46, 55, 58], object detection [26, 30, 51], and natural language processing [39, 49]. Notably, their performance often lags behind that of Artificial Neural Networks (ANNs). One key reason we have identified is the substantial information loss that occurs during spike transmission, particularly associated with the temporal dependencies inherent in traditional neuron models like the Leaky Integrate-and-Fire (LIF) neuron [3]. This information degradation can limit the network's ability to capture and transmit critical features, thereby hindering overall performance. + +Moreover, existing SNN architectures have not fully exploited the capabilities of modern General-Purpose Graphics Processing Units (GPGPUs). The reliance on single-bit spike representations [40] and isolated weight-spike operations [30, 42] leads to inefficient utilization of parallelism, resulting in low training efficiency and slow inference speeds. This inefficiency not only hampers the practicality of SNNs but also complicates their deployment across + +diverse hardware platforms [2, 20]. + +Although there has been considerable research into neuromorphic hardware [37] and the development of various SNN accelerators [9, 23], SNNs have not become mainstream. This is partly due to their subpar performance compared to ANNs and their incompatibility with modern ANN architectures [8, 58]. This incompatibility arises not only from the reliance on discrete spikes but also from the temporal dependencies in spiking neurons, which require SNN-specific network designs and training methods. Consequently, adapting ANN models and techniques to SNNs requires complex modifications, limiting SNNs' ability to fully leverage advancements in ANN architectures and optimization. + +To address these challenges, we propose SpikePack, a novel neuron model designed to reduce information loss during the transition from pre-synaptic to post-synaptic spikes, minimizing degradation associated with temporal dependencies in traditional neuron models. SpikePack enhances information flow within SNNs while achieving $\mathcal{O}(1)$ time and space complexity with respect to time steps, enabling efficient time-parallel training and inference on GPUs, as shown in Figure 1. SpikePack also preserves essential biological characteristics, such as membrane potential reset and leaky integration, ensuring biological plausibility. + +In addition, SpikePack is compatible with modern ANN architectures, allowing for near-lossless ANN-to-SNN conversion and preserving the inherent sparsity of spike-based computations. This compatibility enables the integration of advanced ANN models within the SNN framework, improving performance across a variety of tasks. + +Our contributions can be summarized as follows: + +- We introduce SpikePack, a neuron model that minimizes information loss from pre-synaptic to post-synaptic spikes. SpikePack achieves a balance between computational efficiency and biological fidelity in SNNs. +- By achieving $\mathcal{O}(1)$ time and space complexity, SpikePack preserves essential neural dynamics, enabling efficient, biologically relevant behavior. As shown in Figure 2, SpikePack supports direct training, eliminating the need for complex temporal unfolding and enabling more efficient gradient-based optimization. Its compatibility with modern ANN architectures further supports near-lossless ANN-to-SNN conversion while maintaining the inherent sparsity of spike-based computations. +- Extensive experiments on image classification, object detection, and segmentation tasks showcase significant improvements over state-of-the-art methods. Additional testing on SNN hardware accelerators further validates the generality and efficiency of our approach. + +By addressing the fundamental issues of information loss and hardware inefficiency, SpikePack represents a signifi + +cant advancement in the practical deployment of SNNs. It not only enhances the computational capabilities of SNNs but also ensures that these improvements are accessible across various hardware platforms. This work not only brings us closer to realizing the full potential of neuromorphic computing in real-world scenarios but also offers new insights into bridging SNNs and ANNs. + +![](images/04f964ad3c44a3fec9430836307812c9a47f9f42eb5e97c12974b5d93f5a73b6.jpg) +Figure 2. Forward and backward computation in SpikePack, showing both serial and parallel modes. Parallel computation uses $s_{\mathrm{zip}}^l$ for efficient global potential $(v_g^l)$ calculation. + +# 2. Related Work + +Our work builds on advancements in three areas of SNNs: information transmission efficiency, efficient training methods, and compatibility with ANNs. + +Information Transmission Efficiency Traditional neuron models like the LIF neuron rely on binary spikes, which limit information capacity. To address this, ternary spike neurons [12] and burst-based LIFB neurons [40] expand spike representations, allowing richer information flow while retaining energy efficiency. Other approaches, such as integer-valued neurons [30] and rectified membrane potentials [11], aim to reduce quantization errors and mitigate degradation in deep SNNs. Although these methods improve information transmission, they often add model complexity. Our SpikePack offers a simpler approach by computing membrane potentials before spikes are generated, effectively enhancing information flow without increasing computational demands. + +Efficient Training Methods Training efficiency is a core challenge in SNNs due to the non-differentiability of spike operations, and weight-spike computations, all of which limit hardware utilization. While Backpropagation Through Time (BPTT) [46] has enabled deep SNN training, it incurs high memory and time costs. Approaches like Online Training Through Time (OTTT) [47] and Spatial Learning Through Time (SLTT) [33] reduce memory overhead by prioritizing critical temporal interactions, while Temporal Reversible SNNs (T-RevSNN) [14] leverage reversible ar + +chitectures to lower memory costs. However, these methods primarily optimize training efficiency without addressing the fundamental limitations of binary spikes and underutilized hardware. In contrast, SpikePack inherently supports time-parallel processing, achieving $\mathcal{O}(1)$ complexity with respect to time steps, thus improving training efficiency and enabling high inference efficiency across diverse hardware. + +Compatibility with ANN Architectures Leveraging ANN advancements within the SNN framework is challenging due to fundamental differences in activation dynamics. Traditional ANN-to-SNN conversion approaches rely on rate coding, often requiring many time steps and recalibration [26]. Recently, methods such as Spatio-Temporal Approximation [17] and Expectation Compensation [16] have enabled SNN adaptations of Transformer architectures by approximating non-linear interactions. While effective, these methods are often complex and architecture-specific. By contrast, SpikePack enables near-lossless conversion across various ANN architectures with minimal adjustments, facilitating direct integration with modern ANN models while preserving SNN sparsity and efficiency. + +While previous works have addressed aspects of information flow, training efficiency, and ANN compatibility, they often require complex modifications. SpikePack provides a unified, streamlined solution that enhances information flow, supports efficient training, and enables seamless integration with ANN architectures, promoting scalable SNN deployment across diverse hardware. + +# 3. Methodology + +In this section, we analyze the limitations of LIF neurons, focusing on their information transmission inefficiencies and computational limitations on GPGPUs. These limitations stem from both information loss during spike transmission and inefficient hardware utilization. To address these challenges, we propose SpikePack, a novel neuron model designed to preserve critical information and support efficient parallel processing on modern hardware. We also provide a theoretical foundation for SpikePack through analysis of information transmission, followed by an explanation of its computational efficiency. + +![](images/a47adf05d864ef918462a953ffa9af8c9e7f0cc18e0ddb19e745453d387aad08.jpg) +Figure 3. Spike sequence compression into integer $s_{\mathrm{zip}}^l$ for efficient computation of $v_g^l$ without decompression. + +# 3.1. Limitations of LIF Neurons + +The LIF neuron model is widely used in SNNs due to its simplicity and biological inspiration. However, it suffers from two key limitations that hinder its effectiveness: low information retention during spike transmission and inefficient utilization of modern parallel processing hardware. + +# 3.1.1. Low Information Capacity in Spike Transmission + +LIF neurons generate output spikes based on the membrane potential at each discrete time step. However, since the membrane potential is determined by the combined effects of the historical input sequence and the decay factor, the model exhibits certain limitations in fully integrating input information. This means that each spiking decision is based on an incomplete view of the input sequence, causing substantial information loss. + +For a given layer $l$ in an SNN, let $\mathbf{S}^{l - 1}\in \{0,1\}^{N\times T}$ denote the binary spike matrix from the previous layer, where $N$ is the number of pre-synaptic neurons, and $T$ is the number of time steps. The weight matrix of layer $l$ is represented as $\mathbf{W}^l\in \mathbb{R}^{M\times N}$ , where $M$ is the number of neurons in the current layer. For our analysis, we focus on a single neuron in this layer, represented by the weight vector $\mathbf{w}^l = \mathbf{W}_{i,\cdot}^l$ . Here, we refer to the output spike sequence for this neuron as $\mathbf{s}^l\in \{0,1\}^{1\times T}$ . + +The membrane potential $v_{t}$ of a LIF neuron updates according to the rule given in Eq. (1): + +$$ +v _ {t} ^ {l} = \frac {1}{\tau} v _ {t - 1} ^ {l} + \mathbf {w} ^ {l} \cdot \mathbf {s} _ {t} ^ {l - 1} - \theta \cdot s _ {t - 1} ^ {l}, \tag {1} +$$ + +where $\tau$ is the membrane time constant, $\mathbf{s}_t^{l - 1}$ represents input spikes at time $t$ , and $\theta$ denotes the firing threshold. A spike is generated when $v_{t}$ exceeds the threshold $\theta$ , after which the membrane potential is reset to facilitate subsequent spiking dynamics as shown in Eq. (2): + +$$ +s _ {t} ^ {l} = \left\{ \begin{array}{l l} 1, & \text {i f} v _ {t} > \theta , \\ 0, & \text {o t h e r w i s e .} \end{array} \right. \tag {2} +$$ + +Since $v_{t}^{l}$ is recursively dependent on $v_{t-1}^{l}$ , each spiking decision is based on partially integrated information from the input sequence. This recursive structure imposes inherent limitations on the mutual information that can be preserved between the input and output, leading to significant information loss during spike transmission. As a result, the ability of the model to fully encode and utilize temporal dependencies in the input signal is constrained. + +# 3.1.2. Inefficient Hardware Utilization + +Another significant limitation of LIF neurons lies in their inefficient utilization of modern parallel processing hardware. The recursive nature of the membrane potential update, as defined in Eq. (1), inherently restricts parallelism, as each + +time step must be processed sequentially, resulting in a time complexity of $\mathcal{O}(T)$ , where $T$ is the number of time steps. Furthermore, the need to store individual spikes for each time step occupies separate integer or floating-point units in memory, leading to inefficient memory utilization and increased space complexity of $\mathcal{O}(T)$ . These inefficiencies are especially problematic for hardware architectures optimized for parallel computation. + +In addition, the LIF model requires repeated spike-weight operations at every time step. As each input at each time step consists of a single spike value, the effective computation is limited to spike-weight multiplications, thereby underutilizing the General Matrix Multiplication (GeMM) capabilities of modern hardware. During training, the reliance on surrogate gradient approximations [46] to handle the non-differentiable spiking functions further exacerbates computational inefficiency, increasing both the computational overhead and overall complexity. + +# 3.2. SpikePack + +To address these challenges, we introduce SpikePack, a neuron model designed to preserve information capacity across the input sequence while supporting efficient parallel computation. SpikePack aggregates information across the entire input sequence into a global membrane potential, denoted as $v_{g}^{l}$ for layer $l$ . This global aggregation enhances information flow and maximizes mutual information between inputs and outputs. + +The global membrane potential $v_{g}^{l}$ is computed as: + +$$ +v _ {g} ^ {l} = \mathbf {w} ^ {l} \mathbf {S} ^ {l - 1} \mathbf {q}, \tag {3} +$$ + +where $\mathbf{q} = [\tau^{T - 1},\tau^{T - 2},\dots ,\tau^0 ]^\top$ applies the influence of the membrane time constant $\tau$ across time steps. This formulation enables $v_{g}^{l}$ to integrate information from the entire input sequence $\mathbf{S}^{l - 1}$ , resulting in improved information flow from pre-synaptic to post-synaptic neurons. + +After aggregating information into $v_{g}^{l}$ , SpikePack generates the output spike sequence $\mathbf{S}^{l}$ through a decoding process. The membrane potential $v_{t}^{l}$ is updated as in Eq. (4): + +$$ +v _ {t} ^ {l} = v _ {t - 1} ^ {l} - \theta_ {t} \cdot s _ {t - 1} ^ {l}, \tag {4} +$$ + +where $\theta_{t} = \frac{\theta}{\tau^{t - T}}$ represents a dynamic threshold that adapts over time. The initial membrane potential is set to $v_0^l = v_g^l$ . The spike generation condition is defined as in Eq. (5): + +$$ +s _ {t} ^ {l} = \left\{ \begin{array}{l l} 1, & \text {i f} v _ {t} ^ {l} > \theta_ {t}, \\ 0, & \text {o t h e r w i s e .} \end{array} \right. \tag {5} +$$ + +# 3.2.1. Improved Information Capacity in SpikePack + +The initial global membrane potential $v_{g}^{l}$ aggregates information across the entire input sequence, thereby enhancing the mutual information between the input $\mathbf{S}^{l - 1}$ and the + +output spike sequence $\mathbf{s}^l$ . Assuming binary spikes with independent Bernoulli distributions and Gaussian weights, $v_{g}^{l}$ approximates a Gaussian distribution with variance as shown in Eq. (6): + +$$ +\sigma_ {v _ {g} ^ {l}} ^ {2} = \sigma^ {2} N p (1 - p) \left(\sum_ {t = 1} ^ {T} q _ {t}\right) ^ {2}, \tag {6} +$$ + +where $p$ represents the probability of an input spike, and $q_{t} = \tau^{t - 1}$ + +The mutual information between $\mathbf{S}^{l - 1}$ and $\mathbf{s}^l$ for SpikePack can thus be approximated as in Eq. (7): + +$$ +I _ {\mathrm {S P}} ^ {l} = \frac {1}{2} \log_ {2} \left(\frac {1 2 \sigma_ {v _ {g} ^ {l}} ^ {2}}{\theta^ {2}}\right). \tag {7} +$$ + +This result, derived theoretically and validated empirically, demonstrates that SpikePack achieves greater information retention across a range of configurations for $N$ and $T$ , affirming its superior transmission capacity. + +For a more detailed theoretical derivation of the mutual information $I(\mathbf{S}_{\mathrm{in}}, \mathbf{s}_{\mathrm{out}})$ between pre-synaptic and postsynaptic spikes in both the LIF and SpikePack models, please refer to Appendix A. Empirical simulations, as illustrated in Figure 4, confirm that SpikePack consistently achieves higher mutual information across various values of $N$ and $T$ , further substantiating its enhanced information transmission capability. + +# 3.2.2. Parallel Computation and Hardware Utilization + +The SpikePack model leverages the compressed input structure $v_{g}^{l}$ , making its operations hardware-friendly and well-suited for parallel computation. As shown in Figure 3, since + +![](images/a4045be74fd289235c8cd33ebc363fe732390e08cc000c6db1555361fcbea53e.jpg) +Figure 4. Empirical comparison of mutual information between SpikePack and LIF neurons over varying $T$ (time steps) and $N$ (pre-synaptic neurons). SpikePack demonstrates higher information retention across configurations. + +all spike information is aggregated into $v_{g}^{l}$ , both time and space complexity are reduced from $\mathcal{O}(T)$ to $\mathcal{O}(1)$ , enabling efficient use of parallel processing on modern GPUs. + +The packed representation of input spikes enables highly efficient matrix multiplication operations, thereby optimizing computational performance. The global membrane potential $v_{g}^{l}$ is computed as in Eq. (8): + +$$ +v _ {g} ^ {l} = \mathbf {w} ^ {l} \mathbf {s} _ {\mathrm {z i p}} ^ {l - 1}, \tag {8} +$$ + +where $\mathbf{s}_{\mathrm{zip}}^{l - 1} = \mathbf{S}^{l - 1}\mathbf{q}$ . This formulation compresses the presynaptic spike matrix into a bitwise representation, facilitating optimized GeMM operations and enhancing computational efficiency. + +Utilizing this compressed representation, the total output spike count over $T$ time steps can be derived directly from $v_{g}^{l}$ without iterative updates, as in Eq. (9): + +$$ +s _ {\mathrm {z i p}} ^ {l} = \left\lceil \frac {v _ {g} ^ {l}}{\theta} \left. \right\rfloor_ {\tau}. \tag {9} +$$ + +In this representation, each bit in $s_{\mathrm{zip}}^{l}$ indicates whether an output spike $s_t^l$ is generated at time step $t$ , effectively compressing the spike sequence. This compact form aligns with the serial spike generation in Eqs. (4) and (5), preserving spiking behavior without per-step computations and significantly reducing computational and memory demands. + +By treating the membrane time constant $\tau$ as a form of quantization, SpikePack adapts to various hardware configurations, balancing precision and efficiency. For instance, $\tau = 2$ results in uniform quantization, while $\tau \neq 2$ enables non-uniform quantization, allowing the model to adjust its computational footprint based on hardware constraints. + +# 3.2.3. Efficient Gradient Propagation + +As shown in Figure 2, SpikePack enhances computational efficiency by simplifying gradient propagation, eliminating the need for BPTT. Instead, gradients are computed directly with respect to the compressed input structure $\mathbf{s}_{\mathrm{zip}}^{l-1}$ , which encapsulates the entire sequence of input spikes in a single compressed representation. This method significantly reduces memory consumption and computational complexity during training by removing the requirement to unroll across time s + +The gradient of the loss $\mathcal{L}$ with respect to $\mathbf{s}_{\mathrm{zip}}^{l - 1}$ is computed as shown in Eq. (10): + +$$ +\frac {\partial \mathcal {L}}{\partial \mathbf {s} _ {\mathrm {z i p}} ^ {l - 1}} = \frac {\partial \mathcal {L}}{\partial \mathbf {s} _ {\mathrm {z i p}} ^ {l}} \frac {\partial \mathbf {s} _ {\mathrm {z i p}} ^ {l}}{\partial v _ {g} ^ {l}} \frac {\partial v _ {g} ^ {l}}{\partial \mathbf {s} _ {\mathrm {z i p}} ^ {l - 1}} \approx \frac {\partial \mathcal {L}}{\partial \mathbf {s} _ {\mathrm {z i p}} ^ {l}} \frac {\mathbf {w} ^ {l}}{\theta}, \tag {10} +$$ + +where we approximate $\frac{\partial[x]_{\tau}}{\partial x} \approx 1$ , effectively bypassing the non-differentiable quantization step. This direct gradient path allows for efficient, memory-saving training + +aligned with SpikePack's compressed and parallel computation model. + +By leveraging this streamlined gradient computation, SpikePack enables gradient propagation without costly temporal dependencies, as required in traditional SNN models. This design is inherently compatible with hardware architectures that support vectorized and parallel computation, such as SIMD instructions and systolic arrays. + +# 4. Experiment + +We evaluate SpikePack on tasks including image classification, object detection, and semantic segmentation, comparing its performance with state-of-the-art SNN models, neuron designs, and ANN-to-SNN conversion methods. We also conduct ablation studies to assess the impact of key parameters, demonstrating the versatility and efficiency of SpikePack across various datasets. + +# 4.1. Experimental Setup + +We conduct experiments on both static and neuromorphic datasets to thoroughly assess our model. For image classification, we use ImageNet dataset [4]; for object detection, the COCO 2017 dataset [27]; and for semantic segmentation, the ADE20K dataset [57]. To evaluate event-based performance, we use neuromorphic datasets including CIFAR10-DVS [21], DVS-Gesture [1], and N-Caltech101 [35]. Experiments are implemented in PyTorch and run on NVIDIA A100 GPUs, with a default membrane time constant of $\tau = 2$ . Additional experimental details are provided in Appendix B. + +# 4.2. Image and Neuromorphic Data Classification + +We evaluate SpikePack on the ImageNet dataset and compare its performance with other SNN models and neuron designs. + +Comparison with Other Neuron Models We benchmark SpikePack against several neuron models, including LIF, LIFB [40], PSN [10], DSGM [41], and GLIF [52]. For fair evaluation, experiments are conducted using the SEW-ResNet [8] and Spikeformer [59] architectures. Model scales and time steps are adjusted to achieve comparable performance metrics across setups. + +Figure 5 illustrates the efficiency of SpikePack, showcasing its balance between accuracy and computational cost, quantified by Synaptic Operations (SOPs). SOPs measure the overall spiking activity and computational workload, defined as the product of the firing rate, operations per time step, and the number of time steps $T$ . + +Compared to other neuron models, SpikePack consistently achieves higher accuracy at similar or lower computational costs. For instance, within the Spikeformer architecture, SpikePack matches the accuracy of competing neuron + +![](images/63f92c59178f3be548b0a62d98371046989a2b8bdb1bd591b3ef375a7002a904.jpg) +Figure 5. Comparison between SpikePack and other neuron models at different time steps on ImageNet 1k. Our method achieves higher accuracy with lower computational cost. + +Table 1. Comparison of ANN-to-SNN conversion methods on ImageNet 1k. We report the top-1 accuracy $(\%)$ at different time steps (T). + +
ModelMethodANNT=32T=64T=128T=256
ResNet-34RMP [13]----55.65
Opt. [5]70.6433.0159.5267.5470.06
Calib. [25]70.9564.5471.1273.4574.61
SNM [44]75.6655.2862.7265.5369.31
ViT-B/32STA [17]73.3078.7282.3382.5682.79
T=5T=6T=8T=12
ResNet-34SpikePack80.7259.3274.6877.8177.92
ViT-B/32Ours77.9257.2379.2680.6880.72
ViT-L/1488.2785.5988.0088.2288.27
+ +models while requiring only $1/10$ of the SOPs. Furthermore, at equivalent SOP levels, SpikePack delivers nearly a $5\%$ improvement in accuracy, highlighting its superior information transmission and computational efficiency. This advantage persists across varying model sizes and architectures, demonstrating the robustness and scalability of SpikePack. + +Comparison with Efficient SNN Training Methods We evaluate SpikePack against several efficient SNN training methods, including STBP-tdBN [56], SEW ResNet [8], MS ResNet [15], TEBN [7], TET [6], OTTT [47], SLTT [33], Parallel SNN [10], and T-RevSNN [14]. + +As shown in Table 2, SpikePack outperforms these methods in accuracy while requiring less training time and memory. For instance, using ResNet-34 with 4 time steps, SpikePack achieves $73.4\%$ accuracy with only 8.1 minutes per epoch and 24.3 MB memory per image, demonstrating reduced training overhead and superior performance. Moreover, experiments with Transformer-based architectures, such as Spikeformer, confirm that SpikePack main + +tains $\mathcal{O}(1)$ time and space complexity relative to the number of time steps $T$ , achieving up to $80.1\%$ accuracy with 8 time steps without any increase in memory or computational load as $T$ grows. + +Comparison with ANN-to-SNN Conversion Methods We evaluate SpikePack for near-lossless ANN-to-SNN conversion, leveraging its compatibility with various ANN architectures as discussed in Section 3.2.2. Unlike other methods, SpikePack enables efficient conversion without requiring post-conversion training or calibration. + +As shown in Table 1, we compare SpikePack with several conversion methods, including RMP [13], Optimal (Opt.) [5], Spike Calibration (Cailb.) [25], SNM [44], and Spatio-Temporal Approximation (STA) [17]. SpikePack achieves high accuracy with as few as 6 time steps and near-lossless performance at 8 time steps—less than $1/10$ of the steps required by other methods. For example, SpikePack achieves $77.92\%$ accuracy with ResNet-34 and $88.27\%$ with ViT-L/14, closely matching ANN performance. Figure 6 further highlights SpikePack's ability to maintain high accuracy across different architectures with minimal time steps compared to competing methods. + +![](images/ce52f3f709fb0242fea8c3855a240c4a5e34e6a31323f041d2c8c791415275ad.jpg) +Figure 6. Comparison of ANN-to-SNN conversion methods at different time steps. SpikePack achieves higher accuracy with fewer time steps. + +Results on Neuromorphic Datasets We evaluate SpikePack on neuromorphic datasets including CIFAR10-DVS [21], DVS-Gesture [1], and N-Caltech101 [35]. Table 3 reports the accuracy at different time steps. + +SpikePack achieves competitive accuracy on neuromorphic datasets, demonstrating adaptability to event-based data. Unlike static datasets, neuromorphic datasets generally require a higher number of time steps $T$ for optimal performance—a trend consistent with prior studies [42]. These results confirm SpikePack's versatility in effectively handling both static and event-driven data. + +# 4.3. Object Detection + +We evaluate SpikePack on the COCO 2017 validation dataset [27], leveraging RTMDet [31] and DINO [53] as + +Table 2. Comparison of SpikePack with state-of-the-art SNN training methods on ImageNet 1k. We report the number of parameters, time steps, training time per epoch, memory usage per image, synaptic operations (SOPs), and top-1 accuracy. + +
MethodsArchitectureParam (M)Time stepsTraining Time (min/ep)Memory (MB/img)SOP (G)Acc (%)
STBP-tdBN [56]ResNet-3421.8629.6186.17.163.7
SEW ResNet [8]SEW-ResNet-3421.845.0224.54.467.0
SEW-ResNet-5025.6410.0596.95.467.8
MS ResNet [15]MS-ResNet-3421.8611.2267.15.769.4
TEBN [7]ResNet-3421.8416.3260.17.164.3
TET [6]SEW-ResNet-3421.8412.5221.04.468.0
Spikformer [59]Spikeformer-8-38416.8414.2580.88.670.2
Spikeformer-8-51229.7416.7767.812.973.4
Spike-driven Transformer [50]Spikeformer-8-38416.8415.4548.94.372.3
Spikeformer-8-51229.7418.8730.05.174.6
OTTT [47]ResNet-3421.8624.284.16.764.2
SLTT [33]ResNet-3421.8618.171.76.766.2
ResNet-5025.6623.4117.38.067.0
Parallel SNN [10]SEW-ResNet-1811.745.8138.7-67.6
SEW-ResNet-3421.848.3179.74.470.5
T-RevSNN [14]ResNet-1815.246.157.51.969.8
ResNet-1829.849.185.73.173.2
SpikePack (Ours)ResNet-1811.145.820.11.870.6
ResNet-3421.848.124.33.773.4
ResNet-5025.6413.553.44.178.7
Spikeformer-8-51229.729.2150.53.973.6
Spikeformer-8-51229.749.2150.57.778.4
Spikeformer-8-51229.769.2150.511.279.2
Spikeformer-8-51229.789.2150.515.180.1
+ +Table 3. Classification accuracy on neuromorphic datasets at different time steps $T$ . + +
DatasetT=4T=6T=8T=10T=12T=16
CIFAR10-DVS [21]68.176.480.782.483.784.6
DVS-Gesture [1]94.695.596.797.496.797.4
N-Caltech101 [35]76.380.081.782.282.582.6
+ +our base architectures. To highlight the effectiveness of our approach, we compare it against other SNN-based object detection models, including Spiking-YOLO [19], Bayesian Optimization [18], Spike Calibration [25], EMSYOLO [43], Meta-SpikeFormer [51], and SpikeYOLO [30]. + +Table 4 summarizes the evaluation results. Our approach demonstrates superior mean Average Precision (mAP) while requiring fewer time steps and reducing computational cost. For instance, with DINO-r50 and just 6 time steps, SpikePack achieves an impressive $48.5\%$ $\mathrm{mAP@50:95}$ , surpassing previous methods with a substantial reduction in computational overhead. + +Our method demonstrates competitive performance with significantly fewer parameters and computational cost. For instance, using RTMDet-m with only 67.2G SOPs, we achieve $49.1\%$ mAP@50:95 at 8 time steps. + +# 4.4. Semantic Segmentation + +Table 5 presents the segmentation results on ADE20K, showcasing SpikePack's strong performance in dense prediction tasks like semantic segmentation. With Segformer-b2 and 10 time steps, SpikePack achieves $45.6\%$ mIoU, outperforming prior methods while significantly reducing computational cost. These results emphasize the method's efficiency and scalability for challenging benchmarks. + +# 4.5. Ablation Studies + +We conduct ablation studies to analyze the impact of the membrane time constant $\tau$ on SpikePack's performance, evaluated on ImageNet with ResNet-34 (Figure 7). While $\tau = 4.0$ achieves the best performance by aligning with activation distributions, $\tau = 2.0$ is more efficient on GPGPUs, avoiding exponentiation. As a result, $\tau = 2.0$ has been the preferred choice in prior implementations. + +# 4.6. Hardware Compatibility + +Compatibility with Neuromorphic Processors As a spiking neuron, the SpikePack neuron adheres to the binary nature of spiking neurons, generating discrete spike sequences that are fully compatible with existing neuromorphic processors. With higher sparsity, SpikePack achieves superior + +Table 4. Performance of object detection on COCO 2017 validation set [27]. We report the number of parameters, computational cost (SOPs), time steps, and mean Average Precision (mAP). + +
ModelParam (M)SOPs (G)StepmAP@ 50(%)mAP@ 50:95(%)
Spiking-YOLO [19]10.2-3500-25.7
Bayesian Optim [18]10.2-5000-25.9
Spike Calib [25]17.1-51245.4-
EMS-YOLO[43]26.932.2450.130.1
Meta-SpikeFormer (MaskRCNN) [51]34.955.0144.0-
75.0156.4151.2-
Meta-SpikeFormer (YOLO) [51]16.838.7145.0-
16.878.6450.3-
SpikeYOLO [30]23.138.6462.345.5
48.176.1464.647.4
68.893.6466.248.9
SpikePack (Ours) w/ RTMDet-tiny [31]4.88.53655.739.0
4.811.3857.840.9
4.814.11057.941.1
SpikePack (Ours) w/ RTMDet-m [31]24.751.8650.148.5
24.767.2861.749.1
24.786.31061.949.4
SpikePack (Ours) w/ DINO-r50 [53]47.7276666.048.5
47.7359866.750.0
47.74471067.950.1
+ +Table 5. Performance of semantic segmentation on ADE20K [57]. We report the number of parameters, computational cost (SOPs), simulation time steps, and mIoU (\%). + +
ModelParam (M)SOPs (G)StepMIoU(%) 50(%)
Meta-SpikeFormer [51]16.524.6132.3
16.598.2433.6
58.951.7134.8
58.9204.1435.3
SpikePack (Ours) [29] w/ FCN-r5047.2256.8634.1
47.2384.4835.3
47.2476.71035.9
SpikePack (Ours) w/ Segformer-b0 [48]3.7539.3635.3
3.7551.7836.9
3.7563.91037.4
SpikePack (Ours) w/ Segformer-b2 [48]24.883.6642.8
24.8111.5844.1
24.8138.71045.6
+ +Table 6. Latency and Energy Comparison between SpikePack and LIF neuron + +
SpikePackLIF
ResNet-34ResNet-50ResNet-34ResNet-50
Latency23ms24.1 ms29.1 ms34.7 ms
Energy18.6mJ19.4 mJ23.8 mJ28.4 mJ
+ +speedup and reduced energy consumption compared to tra + +![](images/ccd0c979b02082da71833a14d7f3760142696280bb8d15209265bbb35501be54.jpg) +Figure 7. Ablation study on the effect of membrane time constant $\tau$ on ImageNet classification error rate (\%) using ResNet-34 with SpikePack. + +ditional LIF neurons. To validate this, we conduct hardware experiments comparing SpikePack and LIF neurons under identical conditions. We design a neuromorphic processor that processes binary spike inputs and synaptic weights, filtering zero elements in the spike tensor to ensure only active spikes contribute to computation. The processor includes 64 processing elements for synaptic current accumulation, 16 neuron dynamic units, and a 16-input spike detector that converts spike sequences into addresses for synaptic weight fetching. We implement the processor on xczu3eg FPGA chip running at $300\mathrm{MHz}$ . The detailed implementation is shown in Appendix C. Using cycle-accurate simulation, we evaluate inference latency for ResNet-34 and ResNet-50 models. The results demonstrate that SpikePack achieves significantly lower latency and power consumption than traditional LIF neurons, highlighting its efficiency on neuromorphic hardware. + +Compatibility with Parallel Computing Processors Modern parallel computing processors, such as GPGPUs, NPUs, and SIMD-enabled CPUs, excel at matrix multiplication—a critical operation in neural network acceleration. The SpikePack neuron efficiently utilizes these architectures by maintaining $\mathcal{O}(1)$ computation and memory usage across time steps. In contrast, traditional LIF neurons require $\mathcal{O}(T)$ duplication of computational workload and additional storage for membrane potentials, resulting in increased computational overhead. + +# 4.7. Discussion + +Our experimental results demonstrate that SpikePack significantly improves information flow and computational efficiency in SNNs. By minimizing information loss during spike transmission and supporting efficient parallel computation, SpikePack delivers superior performance across diverse tasks and datasets. Additionally, its seamless compatibility with standard ANN architectures enables near-lossless ANN-to-SNN conversion, allowing SNNs to benefit from the latest advancements in ANN models and training techniques. + +# Acknowledgements + +This work is supported by the National Natural Science Foundation of China (Grant No. 62406325). + +# References + +[1] Arnon Amir, Brian Taba, David Berg, Timothy Melano, Jeffrey McKinstry, Carmelo Di Nolfo, Tapan Nayak, Alexander Andreopoulos, Guillaume Garreau, Marcela Mendoza, et al. A low power, fully event-based gesture recognition system. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7243-7252, 2017. 5, 6, 7, 3 +[2] A. Carpegna, A. Savino, and S. D. Carlo. Spiker+: A framework for the generation of efficient spiking neural networks fpga accelerators for inference at the edge. IEEE Transactions on Emerging Topics in Computing, 01(01):1-15, 2024. 2 +[3] Peter Dayan and Laurence F Abbott. Theoretical neuroscience: computational and mathematical modeling of neural systems. MIT press, 2005. 1 +[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5, 3 +[5] Shikuang Deng and Shi Gu. Optimal conversion of conventional artificial neural networks to spiking neural networks. In International Conference on Learning Representations, 2021. 6 +[6] Shikuang Deng, Yuhang Li, Shanghang Zhang, and Shi Gu. Temporal efficient training of spiking neural network via gradient re-weighting. arXiv preprint arXiv:2202.11946, 2022. 6, 7 +[7] Chaoteng Duan, Jianhao Ding, Shiyan Chen, Zhaofei Yu, and Tiejun Huang. Temporal effective batch normalization in spiking neural networks. Advances in Neural Information Processing Systems, 35:34377-34390, 2022. 6, 7 +[8] Wei Fang, Zhaofei Yu, Yanqi Chen, Tiejun Huang, Timothée Masquelier, and Yonghong Tian. Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems, 34:21056-21069, 2021. 2, 5, 6, 7 +[9] Wei Fang, Yanqi Chen, Jianhao Ding, Zhaofei Yu, Timothee Masquelier, Ding Chen, Liwei Huang, Huihui Zhou, Guoqi Li, and Yonghong Tian. Spikingjelly: An open-source machine learning infrastructure platform for spike-based intelligence. Science Advances, 9(40):eadi1480, 2023. 2 +[10] Wei Fang, Zhaofei Yu, Zhaokun Zhou, Ding Chen, Yanqi Chen, Zhengyu Ma, Timothée Masquelier, and Yonghong Tian. Parallel spiking neurons with high efficiency and ability to learn long-term dependencies. Advances in Neural Information Processing Systems, 36, 2024. 5, 6, 7 +[11] Yufei Guo, Xinyi Tong, Yuanpei Chen, Liwen Zhang, Xiaode Liu, Zhe Ma, and Xuhui Huang. Recdis-snn: Rectifying membrane potential distribution for directly training spiking neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 326-335, 2022. 2 + +[12] Yufei Guo, Yuanpei Chen, Xiaode Liu, Weihang Peng, Yuhan Zhang, Xuhui Huang, and Zhe Ma. Ternary spike: Learning ternary spikes for spiking neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12244-12252, 2024. 2 +[13] Bing Han and Kaushik Roy. Deep spiking neural network: Energy efficiency through time based coding. In European conference on computer vision, pages 388-404. Springer, 2020. 6 +[14] JiaKui Hu, Man Yao, Xuerui Qiu, Yuhong Chou, Yuxuan Cai, Ning Qiao, Yonghong Tian, XU Bo, and Guoqi Li. High-performance temporal reversible spiking neural networks with o(1) training memory and o(1) inference cost. In Forty-first International Conference on Machine Learning, 2024. 2, 6, 7 +[15] Yifan Hu, Lei Deng, Yujie Wu, Man Yao, and Guoqi Li. Advancing spiking neural networks toward deep residual learning. IEEE Transactions on Neural Networks and Learning Systems, 2024. 6, 7 +[16] Zihan Huang, Xinyu Shi, Zecheng Hao, Tong Bu, Jianhao Ding, Zhaofei Yu, and Tiejun Huang. Towards high-performance spiking transformers from ann to snn conversion. In ACM Multimedia 2024, 2024. 3 +[17] Yizhou Jiang, Kunlin Hu, Tianren Zhang, Haichuan Gao, Yuqian Liu, Ying Fang, and Feng Chen. Spatio-temporal approximation: A training-free snn conversion for transformers. In The Twelfth International Conference on Learning Representations, 2024. 3, 6 +[18] Seijoon Kim, Seongsik Park, Byunggook Na, Jongwan Kim, and Sungroh Yoon. Towards fast and accurate object detection in bio-inspired spiking neural networks through bayesian optimization. IEEE Access, 9:2633-2643, 2020. 7, 8 +[19] Seijoon Kim, Seongsik Park, Byunggook Na, and Sungroh Yoon. Spiking-yolo: spiking neural network for energy-efficient object detection. In Proceedings of the AAAI conference on artificial intelligence, pages 11270-11277, 2020. 7, 8 +[20] Takumi Kuwahara, Reon Oshio, Mutsumi Kimura, Renyuan Zhang, and Yasuhiko Nakashima. Fusion synapse by memristor and capacitor for spiking neuromorphic systems. Neurocomputing, 593:127792, 2024. 2 +[21] Hongmin Li, Hanchao Liu, Xiangyang Ji, Guoqi Li, and Luping Shi. Cifar10-dvs: an event-stream dataset for object classification. Frontiers in neuroscience, 11:309, 2017. 5, 6, 7, 3 +[22] Jindong Li, Guobin Shen, Dongcheng Zhao, Qian Zhang, and Yi Zeng. Firefly: A high-throughput hardware accelerator for spiking neural networks with efficient tsp and memory optimization. IEEE Transactions on Very Large Scale Integration (VLSI) Systems, 2023. 1 +[23] Jindong Li, Guobin Shen, Dongcheng Zhao, Qian Zhang, and Yi Zeng. Firefly v2: Advancing hardware support for high-performance spiking neural network with a spatiotemporal fpga accelerator. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2024. 2 +[24] Tenglong Li, Jindong Li, Guobin Shen, Dongcheng Zhao, Qian Zhang, and Yi Zeng. Firefly-s: Exploiting dual-side + +sparsity for spiking neural networks acceleration with reconfigurable spatial architecture. IEEE Transactions on Circuits and Systems I: Regular Papers, 2024. 1, 4 +[25] Yang Li and Yi Zeng. Efficient and accurate conversion of spiking neural network with burst spikes. arXiv preprint arXiv:2204.13271, 2022. 6, 7, 8 +[26] Yang Li, Xiang He, Yiting Dong, Qingqun Kong, and Yi Zeng. Spike calibration: Fast and accurate conversion of spiking neural network for object detection and segmentation. arXiv preprint arXiv:2207.02702, 2022. 1, 3 +[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5, 6, 8, 3 +[28] Zhuang Liu, Mingjie Sun, Tinghui Zhou, Gao Huang, and Trevor Darrell. Rethinking the value of network pruning. arXiv preprint arXiv:1810.05270, 2018. 1 +[29] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 8 +[30] X Luo et al. Integer-valued training and spike-driven inference snn for high-performance and energy-efficient object detection. In ECCV, 2024. 1, 2, 7, 8 +[31] Chengqi Lyu, Wenwei Zhang, Haian Huang, Yue Zhou, Yudong Wang, Yanyi Liu, Shilong Zhang, and Kai Chen. Rtmdet: An empirical study of designing real-time object detectors. arXiv preprint arXiv:2212.07784, 2022. 6, 8 +[32] Wolfgang Maass. Networks of spiking neurons: the third generation of neural network models. Neural networks, 10 (9):1659-1671, 1997. 1 +[33] Qingyan Meng, Mingqing Xiao, Shen Yan, Yisen Wang, Zhouchen Lin, and Zhi-Quan Luo. Towards memory-and time-efficient backpropagation for training spiking neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6166–6176, 2023. 2, 6, 7 +[34] Paul A Merolla, John V Arthur, Rodrigo Alvarez-Icaza, Andrew S Cassidy, Jun Sawada, Filipp Akopyan, Bryan L Jackson, Nabil Imam, Chen Guo, Yutaka Nakamura, et al. A million spiking-neuron integrated circuit with a scalable communication network and interface. Science, 345(6197):668-673, 2014. 1 +[35] Garrick Orchard, Ajinkya Jayawant, Gregory K Cohen, and Nitish Thakor. Converting static image datasets to spiking neuromorphic datasets using saccades. Frontiers in neuroscience, 9:437, 2015. 5, 6, 7, 3 +[36] Sandeep Pande, Fearghal Morgan, Seamus Cawley, Brian McGinley, Snaider Carrillo, Jim Harkin, and Liam McDaid. Embrace-sysc for analysis of noc-based spiking neural network architectures. In 2010 International Symposium on System on Chip, pages 139-145. IEEE, 2010. 1 +[37] Jing Pei, Lei Deng, Sen Song, Mingguo Zhao, Youhui Zhang, Shuang Wu, Guanrui Wang, Zhe Zou, Zhenzhi Wu, Wei He, et al. Towards artificial general intelligence with + +hybrid tianjic chip architecture. Nature, 572(7767):106-111, 2019. 2 +[38] Guobin Shen, Dongcheng Zhao, and Yi Zeng. Backpropagation with biologically plausible spatiotemporal adjustment for training deep spiking neural networks. *Patterns*, 3(6), 2022. 1 +[39] Guobin Shen, Dongcheng Zhao, Yiting Dong, Yang Li, Jindong Li, Kang Sun, and Yi Zeng. Astrocyte-enabled advancements in spiking neural networks for large language modeling. arXiv preprint arXiv:2312.07625, 2023. 1 +[40] Guobin Shen, Dongcheng Zhao, and Yi Zeng. Exploiting high performance spiking neural networks with efficient spiking patterns. arXiv preprint arXiv:2301.12356, 2023. 1, 2, 5 +[41] Guobin Shen, Dongcheng Zhao, and Yi Zeng. Exploiting nonlinear dendritic adaptive computation in training deep spiking neural networks. Neural Networks, 170:190-201, 2024. 5 +[42] G Shen et al. Are conventional snns really efficient? a perspective from network quantization. In CVPR, 2024. 1, 6 +[43] Qiaoyi Su, Yuhong Chou, Yifan Hu, Jianing Li, Shijie Mei, Ziyang Zhang, and Guoqi Li. Deep directly-trained spiking neural networks for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6555-6565, 2023. 7, 8 +[44] Yuchen Wang, Malu Zhang, Yi Chen, and Hong Qu. Signed neuron with memory: Towards simple, accurate and high-efficient ann-snn conversion. In International Joint Conferences on Artificial Intelligence, 2022. 6 +[45] Bernard Widrow and István Kollár. Quantization noise: roundoff error in digital computation, signal processing, control, and communications. Cambridge University Press, 2008. 1 +[46] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, and Luping Shi. Spatio-temporal backpropagation for training high-performance spiking neural networks. Frontiers in neuroscience, 12:331, 2018. 1, 2, 4 +[47] Mingqing Xiao, Qingyan Meng, Zongpeng Zhang, Di He, and Zhouchen Lin. Online training through time for spiking neural networks. Advances in neural information processing systems, 35:20717-20730, 2022. 2, 6, 7 +[48] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in neural information processing systems, 34: 12077-12090, 2021. 8 +[49] Xingrun Xing, Boyan Gao, Zheng Zhang, David A Clifton, Shitao Xiao, Li Du, Guoqi Li, and Jiajun Zhang. Spikellm: Scaling up spiking neural network to large language models via saliency-based spiking. arXiv preprint arXiv:2407.04752, 2024. 1 +[50] Man Yao, Jiakui Hu, Zhaokun Zhou, Li Yuan, Yonghong Tian, Bo Xu, and Guoqi Li. Spike-driven transformer. arXiv preprint arXiv:2307.01694, 2023. 7 +[51] Man Yao, JiaKui Hu, Tianxiang Hu, Yifan Xu, Zhaokun Zhou, Yonghong Tian, XU Bo, and Guoqi Li. Spike-driven transformer v2: Meta spiking neural network architecture inspiring the design of next-generation neuromorphic chips. In + +The Twelfth International Conference on Learning Representations, 2024. 1, 7, 8 +[52] Xingting Yao, Fanrong Li, Zitao Mo, and Jian Cheng. Glif: A unified gated leaky integrate-and-fire neuron for spiking neural networks. Advances in Neural Information Processing Systems, 35:32160-32171, 2022. 5 +[53] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 6, 8 +[54] Youhui Zhang, Peng Qu, Yu Ji, Weihao Zhang, Guangrong Gao, Guanrui Wang, Sen Song, Guoqi Li, Wenguang Chen, Weimin Zheng, et al. A system hierarchy for brain-inspired computing. Nature, 586(7829):378-384, 2020. 1 +[55] Dongcheng Zhao, Guobin Shen, Yiting Dong, Yang Li, and Yi Zeng. Improving stability and performance of spiking neural networks through enhancing temporal consistency. Pattern Recognition, 159:111094, 2025. 1 +[56] Hanle Zheng, Yujie Wu, Lei Deng, Yifan Hu, and Guoqi Li. Going deeper with directly-trained larger spiking neural networks. In Proceedings of the AAAI conference on artificial intelligence, pages 11062-11070, 2021. 6, 7 +[57] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 5, 8, 3 +[58] Zhaokun Zhou, Yuesheng Zhu, Chao He, Yaowei Wang, Shuicheng Yan, Yonghong Tian, and Li Yuan. Spikformer: When spiking neural network meets transformer. arXiv preprint arXiv:2209.15425, 2022. 1, 2, 3 +[59] Zhaokun Zhou, Yuesheng Zhu, Chao He, Yaowei Wang, Shuicheng Yan, Yonghong Tian, and Li Yuan. Spikformer: When spiking neural network meets transformer. In The Eleventh International Conference on Learning Representations, 2023. 5, 7 \ No newline at end of file diff --git a/2025/SpikePack_ Enhanced Information Flow in Spiking Neural Networks with High Hardware Compatibility/images.zip b/2025/SpikePack_ Enhanced Information Flow in Spiking Neural Networks with High Hardware Compatibility/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4e97a5b7d934b46505895a66cce4c1c0ac507537 --- /dev/null +++ b/2025/SpikePack_ Enhanced Information Flow in Spiking Neural Networks with High Hardware Compatibility/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:04afd9d153723d12b61e2bf1b704c844c58e91b01216156cf9af12331b64a654 +size 654054 diff --git a/2025/SpikePack_ Enhanced Information Flow in Spiking Neural Networks with High Hardware Compatibility/layout.json b/2025/SpikePack_ Enhanced Information Flow in Spiking Neural Networks with High Hardware Compatibility/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5c0a298efef498c2b74b9d51937ff5901c98c74a --- /dev/null +++ b/2025/SpikePack_ Enhanced Information Flow in Spiking Neural Networks with High Hardware Compatibility/layout.json @@ -0,0 +1,9714 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 94, + 102, + 516, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 102, + 516, + 140 + ], + "spans": [ + { + "bbox": [ + 94, + 102, + 516, + 140 + ], + "type": "text", + "content": "SpikePack: Enhanced Information Flow in Spiking Neural Networks with High Hardware Compatibility" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 160, + 510, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 160, + 510, + 176 + ], + "spans": [ + { + "bbox": [ + 104, + 160, + 510, + 176 + ], + "type": "text", + "content": "Guobin Shen" + }, + { + "bbox": [ + 104, + 160, + 510, + 176 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 104, + 160, + 510, + 176 + ], + "type": "text", + "content": "; Jindong Li" + }, + { + "bbox": [ + 104, + 160, + 510, + 176 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 104, + 160, + 510, + 176 + ], + "type": "text", + "content": "; Tenglong Li" + }, + { + "bbox": [ + 104, + 160, + 510, + 176 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 104, + 160, + 510, + 176 + ], + "type": "text", + "content": ", Dongcheng Zhao" + }, + { + "bbox": [ + 104, + 160, + 510, + 176 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 104, + 160, + 510, + 176 + ], + "type": "text", + "content": ", Yi Zeng" + }, + { + "bbox": [ + 104, + 160, + 510, + 176 + ], + "type": "inline_equation", + "content": "^{1,2,3,*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 176, + 547, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 176, + 547, + 204 + ], + "spans": [ + { + "bbox": [ + 63, + 176, + 547, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 63, + 176, + 547, + 204 + ], + "type": "text", + "content": " BrainCog Lab, CASIA, " + }, + { + "bbox": [ + 63, + 176, + 547, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 63, + 176, + 547, + 204 + ], + "type": "text", + "content": " Beijing Institute of AI Safety and Governance, " + }, + { + "bbox": [ + 63, + 176, + 547, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 63, + 176, + 547, + 204 + ], + "type": "text", + "content": " Center for Long-term Artificial Intelligence " + }, + { + "bbox": [ + 63, + 176, + 547, + 204 + ], + "type": "inline_equation", + "content": "\\{\\text { shenguobin } 2021, \\text { zhaodongcheng } 2016, \\text { litenglong } 2023, \\text { lijindong } 2022, \\text { yi.zeng } \\} @ia.ac.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 257, + 297, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 297, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 297, + 569 + ], + "type": "text", + "content": "Spiking Neural Networks (SNNs) hold promise for energy-efficient, biologically inspired computing. We identify substantial information loss during spike transmission, linked to temporal dependencies in traditional Leaky Integrate-and-Fire (LIF) neurons—a key factor potentially limiting SNN performance. Existing SNN architectures also underutilize modern GPUs, constrained by single-bit spike storage and isolated weight-spike operations that restrict computational efficiency. We introduce SpikePack, a neuron model designed to reduce transmission loss while preserving essential features like membrane potential reset and leaky integration. SpikePack achieves constant " + }, + { + "bbox": [ + 55, + 257, + 297, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(1)" + }, + { + "bbox": [ + 55, + 257, + 297, + 569 + ], + "type": "text", + "content": " time and space complexity, enabling efficient parallel processing on GPUs and also supporting serial inference on existing SNN hardware accelerators. Compatible with standard Artificial Neural Network (ANN) architectures, SpikePack facilitates near-lossless ANN-to-SNN conversion across various networks. Experimental results on tasks such as image classification, detection, and segmentation show SpikePack achieves significant gains in accuracy and efficiency for both directly trained and converted SNNs over state-of-the-art models. Tests on FPGA-based platforms further confirm cross-platform flexibility, delivering high performance and enhanced sparsity. By enhancing information flow and rethinking SNN-ANN integration, SpikePack advances efficient SNN deployment across diverse hardware platforms." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 588, + 135, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 588, + 135, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 588, + 135, + 601 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 609, + 296, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 296, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 296, + 682 + ], + "type": "text", + "content": "Spiking Neural Networks (SNNs) [32] have emerged as a promising paradigm for energy-efficient and biologically inspired computing [54]. By emulating the discrete spike-based communication of biological neurons, SNNs offer potential advantages in terms of low-power consumption and event-driven processing, which are particularly ap-" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 232, + 553, + 373 + ], + "blocks": [ + { + "bbox": [ + 317, + 232, + 553, + 373 + ], + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 373 + ], + "type": "image", + "image_path": "038754d1520d9a2966c129331e0df06799fc0ac4cf1a8367476c0802c4bd785a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 384, + 555, + 429 + ], + "lines": [ + { + "bbox": [ + 313, + 384, + 555, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 555, + 429 + ], + "type": "text", + "content": "Figure 1. Performance of SpikePack on Spikeformer-8-512 across different time steps (T), showing error rate " + }, + { + "bbox": [ + 313, + 384, + 555, + 429 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 313, + 384, + 555, + 429 + ], + "type": "text", + "content": ", time (ep/min), and memory (MB/img). Reference model uses LIF neurons (Ref, T=4)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 435, + 555, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 555, + 459 + ], + "type": "text", + "content": "pealing for deployment on neuromorphic hardware [22-24, 34, 36]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 460, + 556, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 460, + 556, + 616 + ], + "spans": [ + { + "bbox": [ + 312, + 460, + 556, + 616 + ], + "type": "text", + "content": "Despite these advantages, SNNs still face significant challenges that impede their widespread application in complex tasks such as image classification [28, 38, 46, 55, 58], object detection [26, 30, 51], and natural language processing [39, 49]. Notably, their performance often lags behind that of Artificial Neural Networks (ANNs). One key reason we have identified is the substantial information loss that occurs during spike transmission, particularly associated with the temporal dependencies inherent in traditional neuron models like the Leaky Integrate-and-Fire (LIF) neuron [3]. This information degradation can limit the network's ability to capture and transmit critical features, thereby hindering overall performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 617, + 556, + 714 + ], + "type": "text", + "content": "Moreover, existing SNN architectures have not fully exploited the capabilities of modern General-Purpose Graphics Processing Units (GPGPUs). The reliance on single-bit spike representations [40] and isolated weight-spike operations [30, 42] leads to inefficient utilization of parallelism, resulting in low training efficiency and slow inference speeds. This inefficiency not only hampers the practicality of SNNs but also complicates their deployment across" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23385" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 198, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 198, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 198, + 83 + ], + "type": "text", + "content": "diverse hardware platforms [2, 20]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 85, + 296, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 85, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 85, + 296, + 239 + ], + "type": "text", + "content": "Although there has been considerable research into neuromorphic hardware [37] and the development of various SNN accelerators [9, 23], SNNs have not become mainstream. This is partly due to their subpar performance compared to ANNs and their incompatibility with modern ANN architectures [8, 58]. This incompatibility arises not only from the reliance on discrete spikes but also from the temporal dependencies in spiking neurons, which require SNN-specific network designs and training methods. Consequently, adapting ANN models and techniques to SNNs requires complex modifications, limiting SNNs' ability to fully leverage advancements in ANN architectures and optimization." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 240, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 296, + 384 + ], + "type": "text", + "content": "To address these challenges, we propose SpikePack, a novel neuron model designed to reduce information loss during the transition from pre-synaptic to post-synaptic spikes, minimizing degradation associated with temporal dependencies in traditional neuron models. SpikePack enhances information flow within SNNs while achieving " + }, + { + "bbox": [ + 55, + 240, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(1)" + }, + { + "bbox": [ + 55, + 240, + 296, + 384 + ], + "type": "text", + "content": " time and space complexity with respect to time steps, enabling efficient time-parallel training and inference on GPUs, as shown in Figure 1. SpikePack also preserves essential biological characteristics, such as membrane potential reset and leaky integration, ensuring biological plausibility." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 384, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 296, + 456 + ], + "type": "text", + "content": "In addition, SpikePack is compatible with modern ANN architectures, allowing for near-lossless ANN-to-SNN conversion and preserving the inherent sparsity of spike-based computations. This compatibility enables the integration of advanced ANN models within the SNN framework, improving performance across a variety of tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 457, + 266, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 457, + 266, + 468 + ], + "spans": [ + { + "bbox": [ + 67, + 457, + 266, + 468 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 472, + 296, + 687 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 472, + 295, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 518 + ], + "type": "text", + "content": "- We introduce SpikePack, a neuron model that minimizes information loss from pre-synaptic to post-synaptic spikes. SpikePack achieves a balance between computational efficiency and biological fidelity in SNNs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 520, + 296, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 296, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 296, + 626 + ], + "type": "text", + "content": "- By achieving " + }, + { + "bbox": [ + 55, + 520, + 296, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(1)" + }, + { + "bbox": [ + 55, + 520, + 296, + 626 + ], + "type": "text", + "content": " time and space complexity, SpikePack preserves essential neural dynamics, enabling efficient, biologically relevant behavior. As shown in Figure 2, SpikePack supports direct training, eliminating the need for complex temporal unfolding and enabling more efficient gradient-based optimization. Its compatibility with modern ANN architectures further supports near-lossless ANN-to-SNN conversion while maintaining the inherent sparsity of spike-based computations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 627, + 296, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 627, + 296, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 627, + 296, + 687 + ], + "type": "text", + "content": "- Extensive experiments on image classification, object detection, and segmentation tasks showcase significant improvements over state-of-the-art methods. Additional testing on SNN hardware accelerators further validates the generality and efficiency of our approach." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "By addressing the fundamental issues of information loss and hardware inefficiency, SpikePack represents a signifi" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": "cant advancement in the practical deployment of SNNs. It not only enhances the computational capabilities of SNNs but also ensures that these improvements are accessible across various hardware platforms. This work not only brings us closer to realizing the full potential of neuromorphic computing in real-world scenarios but also offers new insights into bridging SNNs and ANNs." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 166, + 553, + 311 + ], + "blocks": [ + { + "bbox": [ + 316, + 166, + 553, + 311 + ], + "lines": [ + { + "bbox": [ + 316, + 166, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 316, + 166, + 553, + 311 + ], + "type": "image", + "image_path": "04f964ad3c44a3fec9430836307812c9a47f9f42eb5e97c12974b5d93f5a73b6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 319, + 555, + 355 + ], + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 355 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 355 + ], + "type": "text", + "content": "Figure 2. Forward and backward computation in SpikePack, showing both serial and parallel modes. Parallel computation uses " + }, + { + "bbox": [ + 313, + 319, + 555, + 355 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{zip}}^l" + }, + { + "bbox": [ + 313, + 319, + 555, + 355 + ], + "type": "text", + "content": " for efficient global potential " + }, + { + "bbox": [ + 313, + 319, + 555, + 355 + ], + "type": "inline_equation", + "content": "(v_g^l)" + }, + { + "bbox": [ + 313, + 319, + 555, + 355 + ], + "type": "text", + "content": " calculation." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 369, + 400, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 369, + 400, + 380 + ], + "spans": [ + { + "bbox": [ + 314, + 369, + 400, + 380 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 389, + 554, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 554, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 554, + 425 + ], + "type": "text", + "content": "Our work builds on advancements in three areas of SNNs: information transmission efficiency, efficient training methods, and compatibility with ANNs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 426, + 554, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 554, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 554, + 593 + ], + "type": "text", + "content": "Information Transmission Efficiency Traditional neuron models like the LIF neuron rely on binary spikes, which limit information capacity. To address this, ternary spike neurons [12] and burst-based LIFB neurons [40] expand spike representations, allowing richer information flow while retaining energy efficiency. Other approaches, such as integer-valued neurons [30] and rectified membrane potentials [11], aim to reduce quantization errors and mitigate degradation in deep SNNs. Although these methods improve information transmission, they often add model complexity. Our SpikePack offers a simpler approach by computing membrane potentials before spikes are generated, effectively enhancing information flow without increasing computational demands." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": "Efficient Training Methods Training efficiency is a core challenge in SNNs due to the non-differentiability of spike operations, and weight-spike computations, all of which limit hardware utilization. While Backpropagation Through Time (BPTT) [46] has enabled deep SNN training, it incurs high memory and time costs. Approaches like Online Training Through Time (OTTT) [47] and Spatial Learning Through Time (SLTT) [33] reduce memory overhead by prioritizing critical temporal interactions, while Temporal Reversible SNNs (T-RevSNN) [14] leverage reversible ar" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23386" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "chitectures to lower memory costs. However, these methods primarily optimize training efficiency without addressing the fundamental limitations of binary spikes and underutilized hardware. In contrast, SpikePack inherently supports time-parallel processing, achieving " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(1)" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " complexity with respect to time steps, thus improving training efficiency and enabling high inference efficiency across diverse hardware." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 157, + 295, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 157, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 295, + 323 + ], + "type": "text", + "content": "Compatibility with ANN Architectures Leveraging ANN advancements within the SNN framework is challenging due to fundamental differences in activation dynamics. Traditional ANN-to-SNN conversion approaches rely on rate coding, often requiring many time steps and recalibration [26]. Recently, methods such as Spatio-Temporal Approximation [17] and Expectation Compensation [16] have enabled SNN adaptations of Transformer architectures by approximating non-linear interactions. While effective, these methods are often complex and architecture-specific. By contrast, SpikePack enables near-lossless conversion across various ANN architectures with minimal adjustments, facilitating direct integration with modern ANN models while preserving SNN sparsity and efficiency." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 323, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 295, + 407 + ], + "type": "text", + "content": "While previous works have addressed aspects of information flow, training efficiency, and ANN compatibility, they often require complex modifications. SpikePack provides a unified, streamlined solution that enhances information flow, supports efficient training, and enables seamless integration with ANN architectures, promoting scalable SNN deployment across diverse hardware." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 418, + 138, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 138, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 138, + 431 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 437, + 295, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 570 + ], + "type": "text", + "content": "In this section, we analyze the limitations of LIF neurons, focusing on their information transmission inefficiencies and computational limitations on GPGPUs. These limitations stem from both information loss during spike transmission and inefficient hardware utilization. To address these challenges, we propose SpikePack, a novel neuron model designed to preserve critical information and support efficient parallel processing on modern hardware. We also provide a theoretical foundation for SpikePack through analysis of information transmission, followed by an explanation of its computational efficiency." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 67, + 579, + 285, + 684 + ], + "blocks": [ + { + "bbox": [ + 67, + 579, + 285, + 684 + ], + "lines": [ + { + "bbox": [ + 67, + 579, + 285, + 684 + ], + "spans": [ + { + "bbox": [ + 67, + 579, + 285, + 684 + ], + "type": "image", + "image_path": "a47adf05d864ef918462a953ffa9af8c9e7f0cc18e0ddb19e745453d387aad08.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 691, + 295, + 715 + ], + "lines": [ + { + "bbox": [ + 55, + 691, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 691, + 295, + 715 + ], + "type": "text", + "content": "Figure 3. Spike sequence compression into integer " + }, + { + "bbox": [ + 55, + 691, + 295, + 715 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{zip}}^l" + }, + { + "bbox": [ + 55, + 691, + 295, + 715 + ], + "type": "text", + "content": " for efficient computation of " + }, + { + "bbox": [ + 55, + 691, + 295, + 715 + ], + "type": "inline_equation", + "content": "v_g^l" + }, + { + "bbox": [ + 55, + 691, + 295, + 715 + ], + "type": "text", + "content": " without decompression." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 72, + 466, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 466, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 466, + 83 + ], + "type": "text", + "content": "3.1. Limitations of LIF Neurons" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 89, + 553, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 553, + 149 + ], + "type": "text", + "content": "The LIF neuron model is widely used in SNNs due to its simplicity and biological inspiration. However, it suffers from two key limitations that hinder its effectiveness: low information retention during spike transmission and inefficient utilization of modern parallel processing hardware." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 155, + 552, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 155, + 552, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 155, + 552, + 167 + ], + "type": "text", + "content": "3.1.1. Low Information Capacity in Spike Transmission" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 170, + 553, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 170, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 170, + 553, + 264 + ], + "type": "text", + "content": "LIF neurons generate output spikes based on the membrane potential at each discrete time step. However, since the membrane potential is determined by the combined effects of the historical input sequence and the decay factor, the model exhibits certain limitations in fully integrating input information. This means that each spiking decision is based on an incomplete view of the input sequence, causing substantial information loss." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 265, + 554, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 554, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 554, + 373 + ], + "type": "text", + "content": "For a given layer " + }, + { + "bbox": [ + 313, + 265, + 554, + 373 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 265, + 554, + 373 + ], + "type": "text", + "content": " in an SNN, let " + }, + { + "bbox": [ + 313, + 265, + 554, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^{l - 1}\\in \\{0,1\\}^{N\\times T}" + }, + { + "bbox": [ + 313, + 265, + 554, + 373 + ], + "type": "text", + "content": " denote the binary spike matrix from the previous layer, where " + }, + { + "bbox": [ + 313, + 265, + 554, + 373 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 265, + 554, + 373 + ], + "type": "text", + "content": " is the number of pre-synaptic neurons, and " + }, + { + "bbox": [ + 313, + 265, + 554, + 373 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 265, + 554, + 373 + ], + "type": "text", + "content": " is the number of time steps. The weight matrix of layer " + }, + { + "bbox": [ + 313, + 265, + 554, + 373 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 265, + 554, + 373 + ], + "type": "text", + "content": " is represented as " + }, + { + "bbox": [ + 313, + 265, + 554, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^l\\in \\mathbb{R}^{M\\times N}" + }, + { + "bbox": [ + 313, + 265, + 554, + 373 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 265, + 554, + 373 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 265, + 554, + 373 + ], + "type": "text", + "content": " is the number of neurons in the current layer. For our analysis, we focus on a single neuron in this layer, represented by the weight vector " + }, + { + "bbox": [ + 313, + 265, + 554, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^l = \\mathbf{W}_{i,\\cdot}^l" + }, + { + "bbox": [ + 313, + 265, + 554, + 373 + ], + "type": "text", + "content": ". Here, we refer to the output spike sequence for this neuron as " + }, + { + "bbox": [ + 313, + 265, + 554, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{s}^l\\in \\{0,1\\}^{1\\times T}" + }, + { + "bbox": [ + 313, + 265, + 554, + 373 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 374, + 553, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 553, + 398 + ], + "type": "text", + "content": "The membrane potential " + }, + { + "bbox": [ + 313, + 374, + 553, + 398 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 313, + 374, + 553, + 398 + ], + "type": "text", + "content": " of a LIF neuron updates according to the rule given in Eq. (1):" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 361, + 406, + 553, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 406, + 553, + 429 + ], + "spans": [ + { + "bbox": [ + 361, + 406, + 553, + 429 + ], + "type": "interline_equation", + "content": "v _ {t} ^ {l} = \\frac {1}{\\tau} v _ {t - 1} ^ {l} + \\mathbf {w} ^ {l} \\cdot \\mathbf {s} _ {t} ^ {l - 1} - \\theta \\cdot s _ {t - 1} ^ {l}, \\tag {1}", + "image_path": "771f9ff2927bb924ec157f19b38bbd81ca9c5d983f27684c1a92d3b0c5da58d3.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 437, + 554, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 499 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 437, + 554, + 499 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 437, + 554, + 499 + ], + "type": "text", + "content": " is the membrane time constant, " + }, + { + "bbox": [ + 313, + 437, + 554, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_t^{l - 1}" + }, + { + "bbox": [ + 313, + 437, + 554, + 499 + ], + "type": "text", + "content": " represents input spikes at time " + }, + { + "bbox": [ + 313, + 437, + 554, + 499 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 437, + 554, + 499 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 437, + 554, + 499 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 437, + 554, + 499 + ], + "type": "text", + "content": " denotes the firing threshold. A spike is generated when " + }, + { + "bbox": [ + 313, + 437, + 554, + 499 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 313, + 437, + 554, + 499 + ], + "type": "text", + "content": " exceeds the threshold " + }, + { + "bbox": [ + 313, + 437, + 554, + 499 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 437, + 554, + 499 + ], + "type": "text", + "content": ", after which the membrane potential is reset to facilitate subsequent spiking dynamics as shown in Eq. (2):" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 387, + 506, + 553, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 506, + 553, + 539 + ], + "spans": [ + { + "bbox": [ + 387, + 506, + 553, + 539 + ], + "type": "interline_equation", + "content": "s _ {t} ^ {l} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} v _ {t} > \\theta , \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {2}", + "image_path": "e49d92fc3f5606ba58291af5eff1bba9c373ea8b3eaaeef02a333119e6df26ca.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 548, + 554, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 554, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 554, + 645 + ], + "type": "text", + "content": "Since " + }, + { + "bbox": [ + 313, + 548, + 554, + 645 + ], + "type": "inline_equation", + "content": "v_{t}^{l}" + }, + { + "bbox": [ + 313, + 548, + 554, + 645 + ], + "type": "text", + "content": " is recursively dependent on " + }, + { + "bbox": [ + 313, + 548, + 554, + 645 + ], + "type": "inline_equation", + "content": "v_{t-1}^{l}" + }, + { + "bbox": [ + 313, + 548, + 554, + 645 + ], + "type": "text", + "content": ", each spiking decision is based on partially integrated information from the input sequence. This recursive structure imposes inherent limitations on the mutual information that can be preserved between the input and output, leading to significant information loss during spike transmission. As a result, the ability of the model to fully encode and utilize temporal dependencies in the input signal is constrained." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 651, + 479, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 651, + 479, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 651, + 479, + 661 + ], + "type": "text", + "content": "3.1.2. Inefficient Hardware Utilization" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "Another significant limitation of LIF neurons lies in their inefficient utilization of modern parallel processing hardware. The recursive nature of the membrane potential update, as defined in Eq. (1), inherently restricts parallelism, as each" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23387" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "time step must be processed sequentially, resulting in a time complexity of " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(T)" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " is the number of time steps. Furthermore, the need to store individual spikes for each time step occupies separate integer or floating-point units in memory, leading to inefficient memory utilization and increased space complexity of " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(T)" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": ". These inefficiencies are especially problematic for hardware architectures optimized for parallel computation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 167, + 295, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 167, + 295, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 295, + 288 + ], + "type": "text", + "content": "In addition, the LIF model requires repeated spike-weight operations at every time step. As each input at each time step consists of a single spike value, the effective computation is limited to spike-weight multiplications, thereby underutilizing the General Matrix Multiplication (GeMM) capabilities of modern hardware. During training, the reliance on surrogate gradient approximations [46] to handle the non-differentiable spiking functions further exacerbates computational inefficiency, increasing both the computational overhead and overall complexity." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 294, + 119, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 294, + 119, + 306 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 119, + 306 + ], + "type": "text", + "content": "3.2. SpikePack" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 312, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 295, + 407 + ], + "type": "text", + "content": "To address these challenges, we introduce SpikePack, a neuron model designed to preserve information capacity across the input sequence while supporting efficient parallel computation. SpikePack aggregates information across the entire input sequence into a global membrane potential, denoted as " + }, + { + "bbox": [ + 55, + 312, + 295, + 407 + ], + "type": "inline_equation", + "content": "v_{g}^{l}" + }, + { + "bbox": [ + 55, + 312, + 295, + 407 + ], + "type": "text", + "content": " for layer " + }, + { + "bbox": [ + 55, + 312, + 295, + 407 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 312, + 295, + 407 + ], + "type": "text", + "content": ". This global aggregation enhances information flow and maximizes mutual information between inputs and outputs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 407, + 272, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 407, + 272, + 421 + ], + "spans": [ + { + "bbox": [ + 67, + 407, + 272, + 421 + ], + "type": "text", + "content": "The global membrane potential " + }, + { + "bbox": [ + 67, + 407, + 272, + 421 + ], + "type": "inline_equation", + "content": "v_{g}^{l}" + }, + { + "bbox": [ + 67, + 407, + 272, + 421 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 142, + 427, + 294, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 427, + 294, + 444 + ], + "spans": [ + { + "bbox": [ + 142, + 427, + 294, + 444 + ], + "type": "interline_equation", + "content": "v _ {g} ^ {l} = \\mathbf {w} ^ {l} \\mathbf {S} ^ {l - 1} \\mathbf {q}, \\tag {3}", + "image_path": "c997ed842e7d4a50e8e8d0e7415c49f48929db0a75e089d6286072362b60ff91.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 450, + 295, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 513 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 450, + 295, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{q} = [\\tau^{T - 1},\\tau^{T - 2},\\dots ,\\tau^0 ]^\\top" + }, + { + "bbox": [ + 55, + 450, + 295, + 513 + ], + "type": "text", + "content": " applies the influence of the membrane time constant " + }, + { + "bbox": [ + 55, + 450, + 295, + 513 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 450, + 295, + 513 + ], + "type": "text", + "content": " across time steps. This formulation enables " + }, + { + "bbox": [ + 55, + 450, + 295, + 513 + ], + "type": "inline_equation", + "content": "v_{g}^{l}" + }, + { + "bbox": [ + 55, + 450, + 295, + 513 + ], + "type": "text", + "content": " to integrate information from the entire input sequence " + }, + { + "bbox": [ + 55, + 450, + 295, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^{l - 1}" + }, + { + "bbox": [ + 55, + 450, + 295, + 513 + ], + "type": "text", + "content": ", resulting in improved information flow from pre-synaptic to post-synaptic neurons." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 513, + 295, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 513, + 295, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 513, + 295, + 551 + ], + "type": "text", + "content": "After aggregating information into " + }, + { + "bbox": [ + 55, + 513, + 295, + 551 + ], + "type": "inline_equation", + "content": "v_{g}^{l}" + }, + { + "bbox": [ + 55, + 513, + 295, + 551 + ], + "type": "text", + "content": ", SpikePack generates the output spike sequence " + }, + { + "bbox": [ + 55, + 513, + 295, + 551 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^{l}" + }, + { + "bbox": [ + 55, + 513, + 295, + 551 + ], + "type": "text", + "content": " through a decoding process. The membrane potential " + }, + { + "bbox": [ + 55, + 513, + 295, + 551 + ], + "type": "inline_equation", + "content": "v_{t}^{l}" + }, + { + "bbox": [ + 55, + 513, + 295, + 551 + ], + "type": "text", + "content": " is updated as in Eq. (4):" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 129, + 556, + 294, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 556, + 294, + 571 + ], + "spans": [ + { + "bbox": [ + 129, + 556, + 294, + 571 + ], + "type": "interline_equation", + "content": "v _ {t} ^ {l} = v _ {t - 1} ^ {l} - \\theta_ {t} \\cdot s _ {t - 1} ^ {l}, \\tag {4}", + "image_path": "8348db1933d803cc87db4ec8c601643fbbb0047122e921c145e06b111ffc997a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 579, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 579, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 579, + 295, + 617 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 579, + 295, + 617 + ], + "type": "inline_equation", + "content": "\\theta_{t} = \\frac{\\theta}{\\tau^{t - T}}" + }, + { + "bbox": [ + 55, + 579, + 295, + 617 + ], + "type": "text", + "content": " represents a dynamic threshold that adapts over time. The initial membrane potential is set to " + }, + { + "bbox": [ + 55, + 579, + 295, + 617 + ], + "type": "inline_equation", + "content": "v_0^l = v_g^l" + }, + { + "bbox": [ + 55, + 579, + 295, + 617 + ], + "type": "text", + "content": ". The spike generation condition is defined as in Eq. (5):" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 128, + 624, + 295, + 656 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 624, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 128, + 624, + 295, + 656 + ], + "type": "interline_equation", + "content": "s _ {t} ^ {l} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} v _ {t} ^ {l} > \\theta_ {t}, \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {5}", + "image_path": "2783089ceb17e3b6742075a33353059e75cf1fd5e0286063cdf4ac1949d29cb9.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 662, + 274, + 675 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 662, + 274, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 662, + 274, + 675 + ], + "type": "text", + "content": "3.2.1. Improved Information Capacity in SpikePack" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "The initial global membrane potential " + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "inline_equation", + "content": "v_{g}^{l}" + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": " aggregates information across the entire input sequence, thereby enhancing the mutual information between the input " + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^{l - 1}" + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": " and the" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "output spike sequence " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{s}^l" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": ". Assuming binary spikes with independent Bernoulli distributions and Gaussian weights, " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "v_{g}^{l}" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " approximates a Gaussian distribution with variance as shown in Eq. (6):" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 365, + 129, + 553, + 165 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 129, + 553, + 165 + ], + "spans": [ + { + "bbox": [ + 365, + 129, + 553, + 165 + ], + "type": "interline_equation", + "content": "\\sigma_ {v _ {g} ^ {l}} ^ {2} = \\sigma^ {2} N p (1 - p) \\left(\\sum_ {t = 1} ^ {T} q _ {t}\\right) ^ {2}, \\tag {6}", + "image_path": "bdea9da043b61719e19d3b7013112918200faa3ca9b55297181e810b9a4c434b.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 169, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 169, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 169, + 553, + 191 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 169, + 553, + 191 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 169, + 553, + 191 + ], + "type": "text", + "content": " represents the probability of an input spike, and " + }, + { + "bbox": [ + 313, + 169, + 553, + 191 + ], + "type": "inline_equation", + "content": "q_{t} = \\tau^{t - 1}" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 192, + 553, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 553, + 217 + ], + "type": "text", + "content": "The mutual information between " + }, + { + "bbox": [ + 313, + 192, + 553, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^{l - 1}" + }, + { + "bbox": [ + 313, + 192, + 553, + 217 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 192, + 553, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{s}^l" + }, + { + "bbox": [ + 313, + 192, + 553, + 217 + ], + "type": "text", + "content": " for SpikePack can thus be approximated as in Eq. (7):" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 380, + 225, + 553, + 259 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 225, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 380, + 225, + 553, + 259 + ], + "type": "interline_equation", + "content": "I _ {\\mathrm {S P}} ^ {l} = \\frac {1}{2} \\log_ {2} \\left(\\frac {1 2 \\sigma_ {v _ {g} ^ {l}} ^ {2}}{\\theta^ {2}}\\right). \\tag {7}", + "image_path": "cfb5b692109af4a0b9f11939a683bae6a163d75f346be7b2146b8dedd158e24b.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 267, + 553, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 267, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 267, + 553, + 315 + ], + "type": "text", + "content": "This result, derived theoretically and validated empirically, demonstrates that SpikePack achieves greater information retention across a range of configurations for " + }, + { + "bbox": [ + 313, + 267, + 553, + 315 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 267, + 553, + 315 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 267, + 553, + 315 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 267, + 553, + 315 + ], + "type": "text", + "content": ", affirming its superior transmission capacity." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 316, + 554, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 554, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 554, + 411 + ], + "type": "text", + "content": "For a more detailed theoretical derivation of the mutual information " + }, + { + "bbox": [ + 313, + 316, + 554, + 411 + ], + "type": "inline_equation", + "content": "I(\\mathbf{S}_{\\mathrm{in}}, \\mathbf{s}_{\\mathrm{out}})" + }, + { + "bbox": [ + 313, + 316, + 554, + 411 + ], + "type": "text", + "content": " between pre-synaptic and postsynaptic spikes in both the LIF and SpikePack models, please refer to Appendix A. Empirical simulations, as illustrated in Figure 4, confirm that SpikePack consistently achieves higher mutual information across various values of " + }, + { + "bbox": [ + 313, + 316, + 554, + 411 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 316, + 554, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 316, + 554, + 411 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 316, + 554, + 411 + ], + "type": "text", + "content": ", further substantiating its enhanced information transmission capability." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 417, + 547, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 547, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 547, + 429 + ], + "type": "text", + "content": "3.2.2. Parallel Computation and Hardware Utilization" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 432, + 553, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 553, + 468 + ], + "type": "text", + "content": "The SpikePack model leverages the compressed input structure " + }, + { + "bbox": [ + 313, + 432, + 553, + 468 + ], + "type": "inline_equation", + "content": "v_{g}^{l}" + }, + { + "bbox": [ + 313, + 432, + 553, + 468 + ], + "type": "text", + "content": ", making its operations hardware-friendly and well-suited for parallel computation. As shown in Figure 3, since" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 334, + 489, + 535, + 668 + ], + "blocks": [ + { + "bbox": [ + 334, + 489, + 535, + 668 + ], + "lines": [ + { + "bbox": [ + 334, + 489, + 535, + 668 + ], + "spans": [ + { + "bbox": [ + 334, + 489, + 535, + 668 + ], + "type": "image", + "image_path": "a4045be74fd289235c8cd33ebc363fe732390e08cc000c6db1555361fcbea53e.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 671, + 553, + 716 + ], + "lines": [ + { + "bbox": [ + 313, + 671, + 553, + 716 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 553, + 716 + ], + "type": "text", + "content": "Figure 4. Empirical comparison of mutual information between SpikePack and LIF neurons over varying " + }, + { + "bbox": [ + 313, + 671, + 553, + 716 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 671, + 553, + 716 + ], + "type": "text", + "content": " (time steps) and " + }, + { + "bbox": [ + 313, + 671, + 553, + 716 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 671, + 553, + 716 + ], + "type": "text", + "content": " (pre-synaptic neurons). SpikePack demonstrates higher information retention across configurations." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23388" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": "all spike information is aggregated into " + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "inline_equation", + "content": "v_{g}^{l}" + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": ", both time and space complexity are reduced from " + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(T)" + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(1)" + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": ", enabling efficient use of parallel processing on modern GPUs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 108, + 295, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 295, + 158 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 295, + 158 + ], + "type": "text", + "content": "The packed representation of input spikes enables highly efficient matrix multiplication operations, thereby optimizing computational performance. The global membrane potential " + }, + { + "bbox": [ + 55, + 108, + 295, + 158 + ], + "type": "inline_equation", + "content": "v_{g}^{l}" + }, + { + "bbox": [ + 55, + 108, + 295, + 158 + ], + "type": "text", + "content": " is computed as in Eq. (8):" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 146, + 166, + 295, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 166, + 295, + 182 + ], + "spans": [ + { + "bbox": [ + 146, + 166, + 295, + 182 + ], + "type": "interline_equation", + "content": "v _ {g} ^ {l} = \\mathbf {w} ^ {l} \\mathbf {s} _ {\\mathrm {z i p}} ^ {l - 1}, \\tag {8}", + "image_path": "6daa66dfaa64e5911c4f1fc5b2b0a4da1921870fe3828d0f5435327953fb2baa.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 191, + 295, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 191, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 295, + 239 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 191, + 295, + 239 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{\\mathrm{zip}}^{l - 1} = \\mathbf{S}^{l - 1}\\mathbf{q}" + }, + { + "bbox": [ + 55, + 191, + 295, + 239 + ], + "type": "text", + "content": ". This formulation compresses the presynaptic spike matrix into a bitwise representation, facilitating optimized GeMM operations and enhancing computational efficiency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 240, + 296, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 296, + 277 + ], + "type": "text", + "content": "Utilizing this compressed representation, the total output spike count over " + }, + { + "bbox": [ + 55, + 240, + 296, + 277 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 240, + 296, + 277 + ], + "type": "text", + "content": " time steps can be derived directly from " + }, + { + "bbox": [ + 55, + 240, + 296, + 277 + ], + "type": "inline_equation", + "content": "v_{g}^{l}" + }, + { + "bbox": [ + 55, + 240, + 296, + 277 + ], + "type": "text", + "content": " without iterative updates, as in Eq. (9):" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 285, + 295, + 320 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 285, + 295, + 320 + ], + "spans": [ + { + "bbox": [ + 141, + 285, + 295, + 320 + ], + "type": "interline_equation", + "content": "s _ {\\mathrm {z i p}} ^ {l} = \\left\\lceil \\frac {v _ {g} ^ {l}}{\\theta} \\left. \\right\\rfloor_ {\\tau}. \\tag {9}", + "image_path": "74fcfa0e6931aaa8520ad888cc0f9abbaf21f73f45a0879aca237dbaa7e312d2.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 328, + 295, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 295, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 295, + 402 + ], + "type": "text", + "content": "In this representation, each bit in " + }, + { + "bbox": [ + 55, + 328, + 295, + 402 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{zip}}^{l}" + }, + { + "bbox": [ + 55, + 328, + 295, + 402 + ], + "type": "text", + "content": " indicates whether an output spike " + }, + { + "bbox": [ + 55, + 328, + 295, + 402 + ], + "type": "inline_equation", + "content": "s_t^l" + }, + { + "bbox": [ + 55, + 328, + 295, + 402 + ], + "type": "text", + "content": " is generated at time step " + }, + { + "bbox": [ + 55, + 328, + 295, + 402 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 328, + 295, + 402 + ], + "type": "text", + "content": ", effectively compressing the spike sequence. This compact form aligns with the serial spike generation in Eqs. (4) and (5), preserving spiking behavior without per-step computations and significantly reducing computational and memory demands." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 403, + 296, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 296, + 474 + ], + "type": "text", + "content": "By treating the membrane time constant " + }, + { + "bbox": [ + 55, + 403, + 296, + 474 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 403, + 296, + 474 + ], + "type": "text", + "content": " as a form of quantization, SpikePack adapts to various hardware configurations, balancing precision and efficiency. For instance, " + }, + { + "bbox": [ + 55, + 403, + 296, + 474 + ], + "type": "inline_equation", + "content": "\\tau = 2" + }, + { + "bbox": [ + 55, + 403, + 296, + 474 + ], + "type": "text", + "content": " results in uniform quantization, while " + }, + { + "bbox": [ + 55, + 403, + 296, + 474 + ], + "type": "inline_equation", + "content": "\\tau \\neq 2" + }, + { + "bbox": [ + 55, + 403, + 296, + 474 + ], + "type": "text", + "content": " enables non-uniform quantization, allowing the model to adjust its computational footprint based on hardware constraints." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 480, + 216, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 216, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 216, + 491 + ], + "type": "text", + "content": "3.2.3. Efficient Gradient Propagation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 495, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 295, + 601 + ], + "type": "text", + "content": "As shown in Figure 2, SpikePack enhances computational efficiency by simplifying gradient propagation, eliminating the need for BPTT. Instead, gradients are computed directly with respect to the compressed input structure " + }, + { + "bbox": [ + 55, + 495, + 295, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{\\mathrm{zip}}^{l-1}" + }, + { + "bbox": [ + 55, + 495, + 295, + 601 + ], + "type": "text", + "content": ", which encapsulates the entire sequence of input spikes in a single compressed representation. This method significantly reduces memory consumption and computational complexity during training by removing the requirement to unroll across time s" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 602, + 295, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 295, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 295, + 627 + ], + "type": "text", + "content": "The gradient of the loss " + }, + { + "bbox": [ + 55, + 602, + 295, + 627 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 602, + 295, + 627 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 55, + 602, + 295, + 627 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{\\mathrm{zip}}^{l - 1}" + }, + { + "bbox": [ + 55, + 602, + 295, + 627 + ], + "type": "text", + "content": " is computed as shown in Eq. (10):" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 91, + 635, + 295, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 635, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 91, + 635, + 295, + 666 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\mathcal {L}}{\\partial \\mathbf {s} _ {\\mathrm {z i p}} ^ {l - 1}} = \\frac {\\partial \\mathcal {L}}{\\partial \\mathbf {s} _ {\\mathrm {z i p}} ^ {l}} \\frac {\\partial \\mathbf {s} _ {\\mathrm {z i p}} ^ {l}}{\\partial v _ {g} ^ {l}} \\frac {\\partial v _ {g} ^ {l}}{\\partial \\mathbf {s} _ {\\mathrm {z i p}} ^ {l - 1}} \\approx \\frac {\\partial \\mathcal {L}}{\\partial \\mathbf {s} _ {\\mathrm {z i p}} ^ {l}} \\frac {\\mathbf {w} ^ {l}}{\\theta}, \\tag {10}", + "image_path": "f75fe9897d083d9cfe1ae22a7a652d8e0bf89e98d8bf7c4e0d2ee57d9f170cdb.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 675, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 675, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 675, + 296, + 715 + ], + "type": "text", + "content": "where we approximate " + }, + { + "bbox": [ + 55, + 675, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\frac{\\partial[x]_{\\tau}}{\\partial x} \\approx 1" + }, + { + "bbox": [ + 55, + 675, + 296, + 715 + ], + "type": "text", + "content": ", effectively bypassing the non-differentiable quantization step. This direct gradient path allows for efficient, memory-saving training" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "aligned with SpikePack's compressed and parallel computation model." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 97, + 553, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 97, + 553, + 169 + ], + "spans": [ + { + "bbox": [ + 313, + 97, + 553, + 169 + ], + "type": "text", + "content": "By leveraging this streamlined gradient computation, SpikePack enables gradient propagation without costly temporal dependencies, as required in traditional SNN models. This design is inherently compatible with hardware architectures that support vectorized and parallel computation, such as SIMD instructions and systolic arrays." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 182, + 391, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 391, + 194 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 391, + 194 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 201, + 555, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 555, + 286 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 555, + 286 + ], + "type": "text", + "content": "We evaluate SpikePack on tasks including image classification, object detection, and semantic segmentation, comparing its performance with state-of-the-art SNN models, neuron designs, and ANN-to-SNN conversion methods. We also conduct ablation studies to assess the impact of key parameters, demonstrating the versatility and efficiency of SpikePack across various datasets." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 295, + 430, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 430, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 430, + 308 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 313, + 554, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 313, + 554, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 554, + 445 + ], + "type": "text", + "content": "We conduct experiments on both static and neuromorphic datasets to thoroughly assess our model. For image classification, we use ImageNet dataset [4]; for object detection, the COCO 2017 dataset [27]; and for semantic segmentation, the ADE20K dataset [57]. To evaluate event-based performance, we use neuromorphic datasets including CIFAR10-DVS [21], DVS-Gesture [1], and N-Caltech101 [35]. Experiments are implemented in PyTorch and run on NVIDIA A100 GPUs, with a default membrane time constant of " + }, + { + "bbox": [ + 313, + 313, + 554, + 445 + ], + "type": "inline_equation", + "content": "\\tau = 2" + }, + { + "bbox": [ + 313, + 313, + 554, + 445 + ], + "type": "text", + "content": ". Additional experimental details are provided in Appendix B." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 454, + 548, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 548, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 548, + 467 + ], + "type": "text", + "content": "4.2. Image and Neuromorphic Data Classification" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 472, + 554, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 554, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 554, + 508 + ], + "type": "text", + "content": "We evaluate SpikePack on the ImageNet dataset and compare its performance with other SNN models and neuron designs." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 509, + 554, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 554, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 554, + 592 + ], + "type": "text", + "content": "Comparison with Other Neuron Models We benchmark SpikePack against several neuron models, including LIF, LIFB [40], PSN [10], DSGM [41], and GLIF [52]. For fair evaluation, experiments are conducted using the SEW-ResNet [8] and Spikeformer [59] architectures. Model scales and time steps are adjusted to achieve comparable performance metrics across setups." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 593, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 554, + 665 + ], + "type": "text", + "content": "Figure 5 illustrates the efficiency of SpikePack, showcasing its balance between accuracy and computational cost, quantified by Synaptic Operations (SOPs). SOPs measure the overall spiking activity and computational workload, defined as the product of the firing rate, operations per time step, and the number of time steps " + }, + { + "bbox": [ + 313, + 593, + 554, + 665 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 593, + 554, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "Compared to other neuron models, SpikePack consistently achieves higher accuracy at similar or lower computational costs. For instance, within the Spikeformer architecture, SpikePack matches the accuracy of competing neuron" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23389" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 291, + 228 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 291, + 228 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 291, + 228 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 291, + 228 + ], + "type": "image", + "image_path": "63f92c59178f3be548b0a62d98371046989a2b8bdb1bd591b3ef375a7002a904.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 239, + 295, + 272 + ], + "lines": [ + { + "bbox": [ + 55, + 239, + 295, + 272 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 295, + 272 + ], + "type": "text", + "content": "Figure 5. Comparison between SpikePack and other neuron models at different time steps on ImageNet 1k. Our method achieves higher accuracy with lower computational cost." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 316, + 290, + 460 + ], + "blocks": [ + { + "bbox": [ + 55, + 274, + 295, + 308 + ], + "lines": [ + { + "bbox": [ + 55, + 274, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 295, + 308 + ], + "type": "text", + "content": "Table 1. Comparison of ANN-to-SNN conversion methods on ImageNet 1k. We report the top-1 accuracy " + }, + { + "bbox": [ + 55, + 274, + 295, + 308 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 274, + 295, + 308 + ], + "type": "text", + "content": " at different time steps (T)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 316, + 290, + 460 + ], + "lines": [ + { + "bbox": [ + 58, + 316, + 290, + 460 + ], + "spans": [ + { + "bbox": [ + 58, + 316, + 290, + 460 + ], + "type": "table", + "html": "
ModelMethodANNT=32T=64T=128T=256
ResNet-34RMP [13]----55.65
Opt. [5]70.6433.0159.5267.5470.06
Calib. [25]70.9564.5471.1273.4574.61
SNM [44]75.6655.2862.7265.5369.31
ViT-B/32STA [17]73.3078.7282.3382.5682.79
T=5T=6T=8T=12
ResNet-34SpikePack80.7259.3274.6877.8177.92
ViT-B/32Ours77.9257.2379.2680.6880.72
ViT-L/1488.2785.5988.0088.2288.27
", + "image_path": "60cd3ea7e76cdd0691aecb5393c163d784237d14feaedd9441149f78f00f02d3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 471, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 295, + 555 + ], + "type": "text", + "content": "models while requiring only " + }, + { + "bbox": [ + 55, + 471, + 295, + 555 + ], + "type": "inline_equation", + "content": "1/10" + }, + { + "bbox": [ + 55, + 471, + 295, + 555 + ], + "type": "text", + "content": " of the SOPs. Furthermore, at equivalent SOP levels, SpikePack delivers nearly a " + }, + { + "bbox": [ + 55, + 471, + 295, + 555 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 471, + 295, + 555 + ], + "type": "text", + "content": " improvement in accuracy, highlighting its superior information transmission and computational efficiency. This advantage persists across varying model sizes and architectures, demonstrating the robustness and scalability of SpikePack." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 556, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 295, + 616 + ], + "type": "text", + "content": "Comparison with Efficient SNN Training Methods We evaluate SpikePack against several efficient SNN training methods, including STBP-tdBN [56], SEW ResNet [8], MS ResNet [15], TEBN [7], TET [6], OTTT [47], SLTT [33], Parallel SNN [10], and T-RevSNN [14]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": "As shown in Table 2, SpikePack outperforms these methods in accuracy while requiring less training time and memory. For instance, using ResNet-34 with 4 time steps, SpikePack achieves " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "73.4\\%" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": " accuracy with only 8.1 minutes per epoch and 24.3 MB memory per image, demonstrating reduced training overhead and superior performance. Moreover, experiments with Transformer-based architectures, such as Spikeformer, confirm that SpikePack main" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "tains " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(1)" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " time and space complexity relative to the number of time steps " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": ", achieving up to " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "80.1\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " accuracy with 8 time steps without any increase in memory or computational load as " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " grows." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 120, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 120, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 120, + 553, + 191 + ], + "type": "text", + "content": "Comparison with ANN-to-SNN Conversion Methods We evaluate SpikePack for near-lossless ANN-to-SNN conversion, leveraging its compatibility with various ANN architectures as discussed in Section 3.2.2. Unlike other methods, SpikePack enables efficient conversion without requiring post-conversion training or calibration." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 192, + 555, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 555, + 335 + ], + "type": "text", + "content": "As shown in Table 1, we compare SpikePack with several conversion methods, including RMP [13], Optimal (Opt.) [5], Spike Calibration (Cailb.) [25], SNM [44], and Spatio-Temporal Approximation (STA) [17]. SpikePack achieves high accuracy with as few as 6 time steps and near-lossless performance at 8 time steps—less than " + }, + { + "bbox": [ + 313, + 192, + 555, + 335 + ], + "type": "inline_equation", + "content": "1/10" + }, + { + "bbox": [ + 313, + 192, + 555, + 335 + ], + "type": "text", + "content": " of the steps required by other methods. For example, SpikePack achieves " + }, + { + "bbox": [ + 313, + 192, + 555, + 335 + ], + "type": "inline_equation", + "content": "77.92\\%" + }, + { + "bbox": [ + 313, + 192, + 555, + 335 + ], + "type": "text", + "content": " accuracy with ResNet-34 and " + }, + { + "bbox": [ + 313, + 192, + 555, + 335 + ], + "type": "inline_equation", + "content": "88.27\\%" + }, + { + "bbox": [ + 313, + 192, + 555, + 335 + ], + "type": "text", + "content": " with ViT-L/14, closely matching ANN performance. Figure 6 further highlights SpikePack's ability to maintain high accuracy across different architectures with minimal time steps compared to competing methods." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 341, + 548, + 490 + ], + "blocks": [ + { + "bbox": [ + 318, + 341, + 548, + 490 + ], + "lines": [ + { + "bbox": [ + 318, + 341, + 548, + 490 + ], + "spans": [ + { + "bbox": [ + 318, + 341, + 548, + 490 + ], + "type": "image", + "image_path": "ce52f3f709fb0242fea8c3855a240c4a5e34e6a31323f041d2c8c791415275ad.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 502, + 553, + 534 + ], + "lines": [ + { + "bbox": [ + 313, + 502, + 553, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 553, + 534 + ], + "type": "text", + "content": "Figure 6. Comparison of ANN-to-SNN conversion methods at different time steps. SpikePack achieves higher accuracy with fewer time steps." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 534, + 553, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 553, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 553, + 583 + ], + "type": "text", + "content": "Results on Neuromorphic Datasets We evaluate SpikePack on neuromorphic datasets including CIFAR10-DVS [21], DVS-Gesture [1], and N-Caltech101 [35]. Table 3 reports the accuracy at different time steps." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 553, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 666 + ], + "type": "text", + "content": "SpikePack achieves competitive accuracy on neuromorphic datasets, demonstrating adaptability to event-based data. Unlike static datasets, neuromorphic datasets generally require a higher number of time steps " + }, + { + "bbox": [ + 313, + 582, + 553, + 666 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 582, + 553, + 666 + ], + "type": "text", + "content": " for optimal performance—a trend consistent with prior studies [42]. These results confirm SpikePack's versatility in effectively handling both static and event-driven data." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 672, + 416, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 416, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 416, + 685 + ], + "type": "text", + "content": "4.3. Object Detection" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "We evaluate SpikePack on the COCO 2017 validation dataset [27], leveraging RTMDet [31] and DINO [53] as" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23390" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 109, + 96, + 496, + 401 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 2. Comparison of SpikePack with state-of-the-art SNN training methods on ImageNet 1k. We report the number of parameters, time steps, training time per epoch, memory usage per image, synaptic operations (SOPs), and top-1 accuracy." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 109, + 96, + 496, + 401 + ], + "lines": [ + { + "bbox": [ + 109, + 96, + 496, + 401 + ], + "spans": [ + { + "bbox": [ + 109, + 96, + 496, + 401 + ], + "type": "table", + "html": "
MethodsArchitectureParam (M)Time stepsTraining Time (min/ep)Memory (MB/img)SOP (G)Acc (%)
STBP-tdBN [56]ResNet-3421.8629.6186.17.163.7
SEW ResNet [8]SEW-ResNet-3421.845.0224.54.467.0
SEW-ResNet-5025.6410.0596.95.467.8
MS ResNet [15]MS-ResNet-3421.8611.2267.15.769.4
TEBN [7]ResNet-3421.8416.3260.17.164.3
TET [6]SEW-ResNet-3421.8412.5221.04.468.0
Spikformer [59]Spikeformer-8-38416.8414.2580.88.670.2
Spikeformer-8-51229.7416.7767.812.973.4
Spike-driven Transformer [50]Spikeformer-8-38416.8415.4548.94.372.3
Spikeformer-8-51229.7418.8730.05.174.6
OTTT [47]ResNet-3421.8624.284.16.764.2
SLTT [33]ResNet-3421.8618.171.76.766.2
ResNet-5025.6623.4117.38.067.0
Parallel SNN [10]SEW-ResNet-1811.745.8138.7-67.6
SEW-ResNet-3421.848.3179.74.470.5
T-RevSNN [14]ResNet-1815.246.157.51.969.8
ResNet-1829.849.185.73.173.2
SpikePack (Ours)ResNet-1811.145.820.11.870.6
ResNet-3421.848.124.33.773.4
ResNet-5025.6413.553.44.178.7
Spikeformer-8-51229.729.2150.53.973.6
Spikeformer-8-51229.749.2150.57.778.4
Spikeformer-8-51229.769.2150.511.279.2
Spikeformer-8-51229.789.2150.515.180.1
", + "image_path": "5760894bd3472516f0a0b1178cb9d0f4f252fce6ecd9650fdef9312feb8464b1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 59, + 434, + 287, + 495 + ], + "blocks": [ + { + "bbox": [ + 55, + 408, + 295, + 430 + ], + "lines": [ + { + "bbox": [ + 55, + 408, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 295, + 430 + ], + "type": "text", + "content": "Table 3. Classification accuracy on neuromorphic datasets at different time steps " + }, + { + "bbox": [ + 55, + 408, + 295, + 430 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 408, + 295, + 430 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 434, + 287, + 495 + ], + "lines": [ + { + "bbox": [ + 59, + 434, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 59, + 434, + 287, + 495 + ], + "type": "table", + "html": "
DatasetT=4T=6T=8T=10T=12T=16
CIFAR10-DVS [21]68.176.480.782.483.784.6
DVS-Gesture [1]94.695.596.797.496.797.4
N-Caltech101 [35]76.380.081.782.282.582.6
", + "image_path": "5e187278f88469c8e2569e812fa556a82a343c794dce64157628e8ad0b5dcd27.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 513, + 295, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 513, + 295, + 573 + ], + "spans": [ + { + "bbox": [ + 55, + 513, + 295, + 573 + ], + "type": "text", + "content": "our base architectures. To highlight the effectiveness of our approach, we compare it against other SNN-based object detection models, including Spiking-YOLO [19], Bayesian Optimization [18], Spike Calibration [25], EMSYOLO [43], Meta-SpikeFormer [51], and SpikeYOLO [30]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 577, + 295, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 577, + 295, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 577, + 295, + 661 + ], + "type": "text", + "content": "Table 4 summarizes the evaluation results. Our approach demonstrates superior mean Average Precision (mAP) while requiring fewer time steps and reducing computational cost. For instance, with DINO-r50 and just 6 time steps, SpikePack achieves an impressive " + }, + { + "bbox": [ + 55, + 577, + 295, + 661 + ], + "type": "inline_equation", + "content": "48.5\\%" + }, + { + "bbox": [ + 55, + 577, + 295, + 661 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP@50:95}" + }, + { + "bbox": [ + 55, + 577, + 295, + 661 + ], + "type": "text", + "content": ", surpassing previous methods with a substantial reduction in computational overhead." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Our method demonstrates competitive performance with significantly fewer parameters and computational cost. For instance, using RTMDet-m with only 67.2G SOPs, we achieve " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "49.1\\%" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " mAP@50:95 at 8 time steps." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 410, + 447, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 447, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 447, + 423 + ], + "type": "text", + "content": "4.4. Semantic Segmentation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 428, + 555, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 428, + 555, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 428, + 555, + 513 + ], + "type": "text", + "content": "Table 5 presents the segmentation results on ADE20K, showcasing SpikePack's strong performance in dense prediction tasks like semantic segmentation. With Segformer-b2 and 10 time steps, SpikePack achieves " + }, + { + "bbox": [ + 313, + 428, + 555, + 513 + ], + "type": "inline_equation", + "content": "45.6\\%" + }, + { + "bbox": [ + 313, + 428, + 555, + 513 + ], + "type": "text", + "content": " mIoU, outperforming prior methods while significantly reducing computational cost. These results emphasize the method's efficiency and scalability for challenging benchmarks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 522, + 416, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 416, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 416, + 534 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 540, + 555, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 555, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 555, + 625 + ], + "type": "text", + "content": "We conduct ablation studies to analyze the impact of the membrane time constant " + }, + { + "bbox": [ + 313, + 540, + 555, + 625 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 540, + 555, + 625 + ], + "type": "text", + "content": " on SpikePack's performance, evaluated on ImageNet with ResNet-34 (Figure 7). While " + }, + { + "bbox": [ + 313, + 540, + 555, + 625 + ], + "type": "inline_equation", + "content": "\\tau = 4.0" + }, + { + "bbox": [ + 313, + 540, + 555, + 625 + ], + "type": "text", + "content": " achieves the best performance by aligning with activation distributions, " + }, + { + "bbox": [ + 313, + 540, + 555, + 625 + ], + "type": "inline_equation", + "content": "\\tau = 2.0" + }, + { + "bbox": [ + 313, + 540, + 555, + 625 + ], + "type": "text", + "content": " is more efficient on GPGPUs, avoiding exponentiation. As a result, " + }, + { + "bbox": [ + 313, + 540, + 555, + 625 + ], + "type": "inline_equation", + "content": "\\tau = 2.0" + }, + { + "bbox": [ + 313, + 540, + 555, + 625 + ], + "type": "text", + "content": " has been the preferred choice in prior implementations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 635, + 454, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 454, + 648 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 454, + 648 + ], + "type": "text", + "content": "4.6. Hardware Compatibility" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": "Compatibility with Neuromorphic Processors As a spiking neuron, the SpikePack neuron adheres to the binary nature of spiking neurons, generating discrete spike sequences that are fully compatible with existing neuromorphic processors. With higher sparsity, SpikePack achieves superior" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23391" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 108, + 294, + 383 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 104 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 104 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 104 + ], + "type": "text", + "content": "Table 4. Performance of object detection on COCO 2017 validation set [27]. We report the number of parameters, computational cost (SOPs), time steps, and mean Average Precision (mAP)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 108, + 294, + 383 + ], + "lines": [ + { + "bbox": [ + 58, + 108, + 294, + 383 + ], + "spans": [ + { + "bbox": [ + 58, + 108, + 294, + 383 + ], + "type": "table", + "html": "
ModelParam (M)SOPs (G)StepmAP@ 50(%)mAP@ 50:95(%)
Spiking-YOLO [19]10.2-3500-25.7
Bayesian Optim [18]10.2-5000-25.9
Spike Calib [25]17.1-51245.4-
EMS-YOLO[43]26.932.2450.130.1
Meta-SpikeFormer (MaskRCNN) [51]34.955.0144.0-
75.0156.4151.2-
Meta-SpikeFormer (YOLO) [51]16.838.7145.0-
16.878.6450.3-
SpikeYOLO [30]23.138.6462.345.5
48.176.1464.647.4
68.893.6466.248.9
SpikePack (Ours) w/ RTMDet-tiny [31]4.88.53655.739.0
4.811.3857.840.9
4.814.11057.941.1
SpikePack (Ours) w/ RTMDet-m [31]24.751.8650.148.5
24.767.2861.749.1
24.786.31061.949.4
SpikePack (Ours) w/ DINO-r50 [53]47.7276666.048.5
47.7359866.750.0
47.74471067.950.1
", + "image_path": "a6070ca84de1082fcd3ff35ca06247693e4498068b064108b249022199170b33.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 434, + 299, + 608 + ], + "blocks": [ + { + "bbox": [ + 55, + 396, + 296, + 430 + ], + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 430 + ], + "type": "text", + "content": "Table 5. Performance of semantic segmentation on ADE20K [57]. We report the number of parameters, computational cost (SOPs), simulation time steps, and mIoU (\\%)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 434, + 299, + 608 + ], + "lines": [ + { + "bbox": [ + 56, + 434, + 299, + 608 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 299, + 608 + ], + "type": "table", + "html": "
ModelParam (M)SOPs (G)StepMIoU(%) 50(%)
Meta-SpikeFormer [51]16.524.6132.3
16.598.2433.6
58.951.7134.8
58.9204.1435.3
SpikePack (Ours) [29] w/ FCN-r5047.2256.8634.1
47.2384.4835.3
47.2476.71035.9
SpikePack (Ours) w/ Segformer-b0 [48]3.7539.3635.3
3.7551.7836.9
3.7563.91037.4
SpikePack (Ours) w/ Segformer-b2 [48]24.883.6642.8
24.8111.5844.1
24.8138.71045.6
", + "image_path": "9a0921c05f2cf70e25aea7803573052bf74aa61f7b6b28cbf439e1ee518b560b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 63, + 636, + 286, + 684 + ], + "blocks": [ + { + "bbox": [ + 55, + 611, + 296, + 633 + ], + "lines": [ + { + "bbox": [ + 55, + 611, + 296, + 633 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 296, + 633 + ], + "type": "text", + "content": "Table 6. Latency and Energy Comparison between SpikePack and LIF neuron" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 636, + 286, + 684 + ], + "lines": [ + { + "bbox": [ + 63, + 636, + 286, + 684 + ], + "spans": [ + { + "bbox": [ + 63, + 636, + 286, + 684 + ], + "type": "table", + "html": "
SpikePackLIF
ResNet-34ResNet-50ResNet-34ResNet-50
Latency23ms24.1 ms29.1 ms34.7 ms
Energy18.6mJ19.4 mJ23.8 mJ28.4 mJ
", + "image_path": "2131ce71dafd674a9f07b67a018a0284493ae2ac0a25cbb85156aad6d86a6065.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "content": "speedup and reduced energy consumption compared to tra" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 319, + 74, + 550, + 190 + ], + "blocks": [ + { + "bbox": [ + 319, + 74, + 550, + 190 + ], + "lines": [ + { + "bbox": [ + 319, + 74, + 550, + 190 + ], + "spans": [ + { + "bbox": [ + 319, + 74, + 550, + 190 + ], + "type": "image", + "image_path": "ccd0c979b02082da71833a14d7f3760142696280bb8d15209265bbb35501be54.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 202, + 555, + 236 + ], + "lines": [ + { + "bbox": [ + 313, + 202, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 202, + 555, + 236 + ], + "type": "text", + "content": "Figure 7. Ablation study on the effect of membrane time constant " + }, + { + "bbox": [ + 313, + 202, + 555, + 236 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 202, + 555, + 236 + ], + "type": "text", + "content": " on ImageNet classification error rate (\\%) using ResNet-34 with SpikePack." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 245, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 555, + 449 + ], + "type": "text", + "content": "ditional LIF neurons. To validate this, we conduct hardware experiments comparing SpikePack and LIF neurons under identical conditions. We design a neuromorphic processor that processes binary spike inputs and synaptic weights, filtering zero elements in the spike tensor to ensure only active spikes contribute to computation. The processor includes 64 processing elements for synaptic current accumulation, 16 neuron dynamic units, and a 16-input spike detector that converts spike sequences into addresses for synaptic weight fetching. We implement the processor on xczu3eg FPGA chip running at " + }, + { + "bbox": [ + 313, + 245, + 555, + 449 + ], + "type": "inline_equation", + "content": "300\\mathrm{MHz}" + }, + { + "bbox": [ + 313, + 245, + 555, + 449 + ], + "type": "text", + "content": ". The detailed implementation is shown in Appendix C. Using cycle-accurate simulation, we evaluate inference latency for ResNet-34 and ResNet-50 models. The results demonstrate that SpikePack achieves significantly lower latency and power consumption than traditional LIF neurons, highlighting its efficiency on neuromorphic hardware." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 449, + 556, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 568 + ], + "type": "text", + "content": "Compatibility with Parallel Computing Processors Modern parallel computing processors, such as GPGPUs, NPUs, and SIMD-enabled CPUs, excel at matrix multiplication—a critical operation in neural network acceleration. The SpikePack neuron efficiently utilizes these architectures by maintaining " + }, + { + "bbox": [ + 313, + 449, + 556, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(1)" + }, + { + "bbox": [ + 313, + 449, + 556, + 568 + ], + "type": "text", + "content": " computation and memory usage across time steps. In contrast, traditional LIF neurons require " + }, + { + "bbox": [ + 313, + 449, + 556, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(T)" + }, + { + "bbox": [ + 313, + 449, + 556, + 568 + ], + "type": "text", + "content": " duplication of computational workload and additional storage for membrane potentials, resulting in increased computational overhead." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 576, + 388, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 388, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 388, + 588 + ], + "type": "text", + "content": "4.7. Discussion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "Our experimental results demonstrate that SpikePack significantly improves information flow and computational efficiency in SNNs. By minimizing information loss during spike transmission and supporting efficient parallel computation, SpikePack delivers superior performance across diverse tasks and datasets. Additionally, its seamless compatibility with standard ANN architectures enables near-lossless ANN-to-SNN conversion, allowing SNNs to benefit from the latest advancements in ANN models and training techniques." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23392" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 116 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 116 + ], + "type": "text", + "content": "This work is supported by the National Natural Science Foundation of China (Grant No. 62406325)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 128, + 115, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 115, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 115, + 140 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 148, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 148, + 296, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 214 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 214 + ], + "type": "text", + "content": "[1] Arnon Amir, Brian Taba, David Berg, Timothy Melano, Jeffrey McKinstry, Carmelo Di Nolfo, Tapan Nayak, Alexander Andreopoulos, Guillaume Garreau, Marcela Mendoza, et al. A low power, fully event-based gesture recognition system. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7243-7252, 2017. 5, 6, 7, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 216, + 296, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 216, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 61, + 216, + 296, + 270 + ], + "type": "text", + "content": "[2] A. Carpegna, A. Savino, and S. D. Carlo. Spiker+: A framework for the generation of efficient spiking neural networks fpga accelerators for inference at the edge. IEEE Transactions on Emerging Topics in Computing, 01(01):1-15, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 272, + 296, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 272, + 296, + 305 + ], + "spans": [ + { + "bbox": [ + 61, + 272, + 296, + 305 + ], + "type": "text", + "content": "[3] Peter Dayan and Laurence F Abbott. Theoretical neuroscience: computational and mathematical modeling of neural systems. MIT press, 2005. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 307, + 296, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 307, + 296, + 351 + ], + "spans": [ + { + "bbox": [ + 62, + 307, + 296, + 351 + ], + "type": "text", + "content": "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 353, + 296, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 353, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 353, + 296, + 396 + ], + "type": "text", + "content": "[5] Shikuang Deng and Shi Gu. Optimal conversion of conventional artificial neural networks to spiking neural networks. In International Conference on Learning Representations, 2021. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 397, + 296, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 397, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 397, + 296, + 441 + ], + "type": "text", + "content": "[6] Shikuang Deng, Yuhang Li, Shanghang Zhang, and Shi Gu. Temporal efficient training of spiking neural network via gradient re-weighting. arXiv preprint arXiv:2202.11946, 2022. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 443, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 443, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 443, + 296, + 487 + ], + "type": "text", + "content": "[7] Chaoteng Duan, Jianhao Ding, Shiyan Chen, Zhaofei Yu, and Tiejun Huang. Temporal effective batch normalization in spiking neural networks. Advances in Neural Information Processing Systems, 35:34377-34390, 2022. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 489, + 296, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 489, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 489, + 296, + 533 + ], + "type": "text", + "content": "[8] Wei Fang, Zhaofei Yu, Yanqi Chen, Tiejun Huang, Timothée Masquelier, and Yonghong Tian. Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems, 34:21056-21069, 2021. 2, 5, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 534, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 534, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 534, + 296, + 590 + ], + "type": "text", + "content": "[9] Wei Fang, Yanqi Chen, Jianhao Ding, Zhaofei Yu, Timothee Masquelier, Ding Chen, Liwei Huang, Huihui Zhou, Guoqi Li, and Yonghong Tian. Spikingjelly: An open-source machine learning infrastructure platform for spike-based intelligence. Science Advances, 9(40):eadi1480, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 591, + 296, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 296, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 296, + 646 + ], + "type": "text", + "content": "[10] Wei Fang, Zhaofei Yu, Zhaokun Zhou, Ding Chen, Yanqi Chen, Zhengyu Ma, Timothée Masquelier, and Yonghong Tian. Parallel spiking neurons with high efficiency and ability to learn long-term dependencies. Advances in Neural Information Processing Systems, 36, 2024. 5, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 714 + ], + "type": "text", + "content": "[11] Yufei Guo, Xinyi Tong, Yuanpei Chen, Liwen Zhang, Xiaode Liu, Zhe Ma, and Xuhui Huang. Recdis-snn: Rectifying membrane potential distribution for directly training spiking neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 326-335, 2022. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "text", + "content": "[12] Yufei Guo, Yuanpei Chen, Xiaode Liu, Weihang Peng, Yuhan Zhang, Xuhui Huang, and Zhe Ma. Ternary spike: Learning ternary spikes for spiking neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12244-12252, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 129, + 554, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 554, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 554, + 171 + ], + "type": "text", + "content": "[13] Bing Han and Kaushik Roy. Deep spiking neural network: Energy efficiency through time based coding. In European conference on computer vision, pages 388-404. Springer, 2020. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 172, + 554, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 172, + 554, + 237 + ], + "spans": [ + { + "bbox": [ + 317, + 172, + 554, + 237 + ], + "type": "text", + "content": "[14] JiaKui Hu, Man Yao, Xuerui Qiu, Yuhong Chou, Yuxuan Cai, Ning Qiao, Yonghong Tian, XU Bo, and Guoqi Li. High-performance temporal reversible spiking neural networks with o(1) training memory and o(1) inference cost. In Forty-first International Conference on Machine Learning, 2024. 2, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 239, + 554, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 554, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 554, + 282 + ], + "type": "text", + "content": "[15] Yifan Hu, Lei Deng, Yujie Wu, Man Yao, and Guoqi Li. Advancing spiking neural networks toward deep residual learning. IEEE Transactions on Neural Networks and Learning Systems, 2024. 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 283, + 554, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 554, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 554, + 326 + ], + "type": "text", + "content": "[16] Zihan Huang, Xinyu Shi, Zecheng Hao, Tong Bu, Jianhao Ding, Zhaofei Yu, and Tiejun Huang. Towards high-performance spiking transformers from ann to snn conversion. In ACM Multimedia 2024, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 327, + 554, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 554, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 554, + 381 + ], + "type": "text", + "content": "[17] Yizhou Jiang, Kunlin Hu, Tianren Zhang, Haichuan Gao, Yuqian Liu, Ying Fang, and Feng Chen. Spatio-temporal approximation: A training-free snn conversion for transformers. In The Twelfth International Conference on Learning Representations, 2024. 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 383, + 554, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 554, + 436 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 554, + 436 + ], + "type": "text", + "content": "[18] Seijoon Kim, Seongsik Park, Byunggook Na, Jongwan Kim, and Sungroh Yoon. Towards fast and accurate object detection in bio-inspired spiking neural networks through bayesian optimization. IEEE Access, 9:2633-2643, 2020. 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 437, + 554, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 554, + 491 + ], + "type": "text", + "content": "[19] Seijoon Kim, Seongsik Park, Byunggook Na, and Sungroh Yoon. Spiking-yolo: spiking neural network for energy-efficient object detection. In Proceedings of the AAAI conference on artificial intelligence, pages 11270-11277, 2020. 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 493, + 554, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 554, + 536 + ], + "type": "text", + "content": "[20] Takumi Kuwahara, Reon Oshio, Mutsumi Kimura, Renyuan Zhang, and Yasuhiko Nakashima. Fusion synapse by memristor and capacitor for spiking neuromorphic systems. Neurocomputing, 593:127792, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 537, + 554, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 554, + 581 + ], + "type": "text", + "content": "[21] Hongmin Li, Hanchao Liu, Xiangyang Ji, Guoqi Li, and Luping Shi. Cifar10-dvs: an event-stream dataset for object classification. Frontiers in neuroscience, 11:309, 2017. 5, 6, 7, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 582, + 554, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 554, + 635 + ], + "type": "text", + "content": "[22] Jindong Li, Guobin Shen, Dongcheng Zhao, Qian Zhang, and Yi Zeng. Firefly: A high-throughput hardware accelerator for spiking neural networks with efficient tsp and memory optimization. IEEE Transactions on Very Large Scale Integration (VLSI) Systems, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 636, + 554, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 554, + 691 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 554, + 691 + ], + "type": "text", + "content": "[23] Jindong Li, Guobin Shen, Dongcheng Zhao, Qian Zhang, and Yi Zeng. Firefly v2: Advancing hardware support for high-performance spiking neural network with a spatiotemporal fpga accelerator. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 554, + 714 + ], + "type": "text", + "content": "[24] Tenglong Li, Jindong Li, Guobin Shen, Dongcheng Zhao, Qian Zhang, and Yi Zeng. Firefly-s: Exploiting dual-side" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23393" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 106 + ], + "type": "text", + "content": "sparsity for spiking neural networks acceleration with reconfigurable spatial architecture. IEEE Transactions on Circuits and Systems I: Regular Papers, 2024. 1, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 139 + ], + "type": "text", + "content": "[25] Yang Li and Yi Zeng. Efficient and accurate conversion of spiking neural network with burst spikes. arXiv preprint arXiv:2204.13271, 2022. 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "type": "text", + "content": "[26] Yang Li, Xiang He, Yiting Dong, Qingqun Kong, and Yi Zeng. Spike calibration: Fast and accurate conversion of spiking neural network for object detection and segmentation. arXiv preprint arXiv:2207.02702, 2022. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 252 + ], + "type": "text", + "content": "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5, 6, 8, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 286 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 286 + ], + "type": "text", + "content": "[28] Zhuang Liu, Mingjie Sun, Tinghui Zhou, Gao Huang, and Trevor Darrell. Rethinking the value of network pruning. arXiv preprint arXiv:1810.05270, 2018. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 287, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 294, + 331 + ], + "type": "text", + "content": "[29] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 332, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 365 + ], + "type": "text", + "content": "[30] X Luo et al. Integer-valued training and spike-driven inference snn for high-performance and energy-efficient object detection. In ECCV, 2024. 1, 2, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 366, + 294, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 294, + 410 + ], + "type": "text", + "content": "[31] Chengqi Lyu, Wenwei Zhang, Haian Huang, Yue Zhou, Yudong Wang, Yanyi Liu, Shilong Zhang, and Kai Chen. Rtmdet: An empirical study of designing real-time object detectors. arXiv preprint arXiv:2212.07784, 2022. 6, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 411, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 443 + ], + "type": "text", + "content": "[32] Wolfgang Maass. Networks of spiking neurons: the third generation of neural network models. Neural networks, 10 (9):1659-1671, 1997. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 445, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 510 + ], + "type": "text", + "content": "[33] Qingyan Meng, Mingqing Xiao, Shen Yan, Yisen Wang, Zhouchen Lin, and Zhi-Quan Luo. Towards memory-and time-efficient backpropagation for training spiking neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6166–6176, 2023. 2, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 513, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 513, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 513, + 294, + 578 + ], + "type": "text", + "content": "[34] Paul A Merolla, John V Arthur, Rodrigo Alvarez-Icaza, Andrew S Cassidy, Jun Sawada, Filipp Akopyan, Bryan L Jackson, Nabil Imam, Chen Guo, Yutaka Nakamura, et al. A million spiking-neuron integrated circuit with a scalable communication network and interface. Science, 345(6197):668-673, 2014. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 580, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 294, + 623 + ], + "type": "text", + "content": "[35] Garrick Orchard, Ajinkya Jayawant, Gregory K Cohen, and Nitish Thakor. Converting static image datasets to spiking neuromorphic datasets using saccades. Frontiers in neuroscience, 9:437, 2015. 5, 6, 7, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 679 + ], + "type": "text", + "content": "[36] Sandeep Pande, Fearghal Morgan, Seamus Cawley, Brian McGinley, Snaider Carrillo, Jim Harkin, and Liam McDaid. Embrace-sysc for analysis of noc-based spiking neural network architectures. In 2010 International Symposium on System on Chip, pages 139-145. IEEE, 2010. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[37] Jing Pei, Lei Deng, Sen Song, Mingguo Zhao, Youhui Zhang, Shuang Wu, Guanrui Wang, Zhe Zou, Zhenzhi Wu, Wei He, et al. Towards artificial general intelligence with" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "text", + "content": "hybrid tianjic chip architecture. Nature, 572(7767):106-111, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 95, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 138 + ], + "type": "text", + "content": "[38] Guobin Shen, Dongcheng Zhao, and Yi Zeng. Backpropagation with biologically plausible spatiotemporal adjustment for training deep spiking neural networks. *Patterns*, 3(6), 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "type": "text", + "content": "[39] Guobin Shen, Dongcheng Zhao, Yiting Dong, Yang Li, Jindong Li, Kang Sun, and Yi Zeng. Astrocyte-enabled advancements in spiking neural networks for large language modeling. arXiv preprint arXiv:2312.07625, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 184, + 553, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 226 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 226 + ], + "type": "text", + "content": "[40] Guobin Shen, Dongcheng Zhao, and Yi Zeng. Exploiting high performance spiking neural networks with efficient spiking patterns. arXiv preprint arXiv:2301.12356, 2023. 1, 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 228, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 270 + ], + "type": "text", + "content": "[41] Guobin Shen, Dongcheng Zhao, and Yi Zeng. Exploiting nonlinear dendritic adaptive computation in training deep spiking neural networks. Neural Networks, 170:190-201, 2024. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 272, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 293 + ], + "type": "text", + "content": "[42] G Shen et al. Are conventional snns really efficient? a perspective from network quantization. In CVPR, 2024. 1, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "type": "text", + "content": "[43] Qiaoyi Su, Yuhong Chou, Yifan Hu, Jianing Li, Shijie Mei, Ziyang Zhang, and Guoqi Li. Deep directly-trained spiking neural networks for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6555-6565, 2023. 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "text", + "content": "[44] Yuchen Wang, Malu Zhang, Yi Chen, and Hong Qu. Signed neuron with memory: Towards simple, accurate and high-efficient ann-snn conversion. In International Joint Conferences on Artificial Intelligence, 2022. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 394, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 436 + ], + "type": "text", + "content": "[45] Bernard Widrow and István Kollár. Quantization noise: roundoff error in digital computation, signal processing, control, and communications. Cambridge University Press, 2008. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 437, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 480 + ], + "type": "text", + "content": "[46] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, and Luping Shi. Spatio-temporal backpropagation for training high-performance spiking neural networks. Frontiers in neuroscience, 12:331, 2018. 1, 2, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "type": "text", + "content": "[47] Mingqing Xiao, Qingyan Meng, Zongpeng Zhang, Di He, and Zhouchen Lin. Online training through time for spiking neural networks. Advances in neural information processing systems, 35:20717-20730, 2022. 2, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 526, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 580 + ], + "type": "text", + "content": "[48] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in neural information processing systems, 34: 12077-12090, 2021. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "type": "text", + "content": "[49] Xingrun Xing, Boyan Gao, Zheng Zhang, David A Clifton, Shitao Xiao, Li Du, Guoqi Li, and Jiajun Zhang. Spikellm: Scaling up spiking neural network to large language models via saliency-based spiking. arXiv preprint arXiv:2407.04752, 2024. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "type": "text", + "content": "[50] Man Yao, Jiakui Hu, Zhaokun Zhou, Li Yuan, Yonghong Tian, Bo Xu, and Guoqi Li. Spike-driven transformer. arXiv preprint arXiv:2307.01694, 2023. 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[51] Man Yao, JiaKui Hu, Tianxiang Hu, Yifan Xu, Zhaokun Zhou, Yonghong Tian, XU Bo, and Guoqi Li. Spike-driven transformer v2: Meta spiking neural network architecture inspiring the design of next-generation neuromorphic chips. In" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23394" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 475 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "The Twelfth International Conference on Learning Representations, 2024. 1, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[52] Xingting Yao, Fanrong Li, Zitao Mo, and Jian Cheng. Glif: A unified gated leaky integrate-and-fire neuron for spiking neural networks. Advances in Neural Information Processing Systems, 35:32160-32171, 2022. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "type": "text", + "content": "[53] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 229 + ], + "type": "text", + "content": "[54] Youhui Zhang, Peng Qu, Yu Ji, Weihao Zhang, Guangrong Gao, Guanrui Wang, Sen Song, Guoqi Li, Wenguang Chen, Weimin Zheng, et al. A system hierarchy for brain-inspired computing. Nature, 586(7829):378-384, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "type": "text", + "content": "[55] Dongcheng Zhao, Guobin Shen, Yiting Dong, Yang Li, and Yi Zeng. Improving stability and performance of spiking neural networks through enhancing temporal consistency. Pattern Recognition, 159:111094, 2025. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "text", + "content": "[56] Hanle Zheng, Yujie Wu, Lei Deng, Yifan Hu, and Guoqi Li. Going deeper with directly-trained larger spiking neural networks. In Proceedings of the AAAI conference on artificial intelligence, pages 11062-11070, 2021. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 321, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 374 + ], + "type": "text", + "content": "[57] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 5, 8, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 376, + 294, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 420 + ], + "type": "text", + "content": "[58] Zhaokun Zhou, Yuesheng Zhu, Chao He, Yaowei Wang, Shuicheng Yan, Yonghong Tian, and Li Yuan. Spikformer: When spiking neural network meets transformer. arXiv preprint arXiv:2209.15425, 2022. 1, 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 421, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 294, + 475 + ], + "type": "text", + "content": "[59] Zhaokun Zhou, Yuesheng Zhu, Chao He, Yaowei Wang, Shuicheng Yan, Yonghong Tian, and Li Yuan. Spikformer: When spiking neural network meets transformer. In The Eleventh International Conference on Learning Representations, 2023. 5, 7" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23395" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SpinMeRound_ Consistent Multi-View Identity Generation Using Diffusion Models/d46ce965-2153-45b4-9566-3f0d4d7d97a6_content_list.json b/2025/SpinMeRound_ Consistent Multi-View Identity Generation Using Diffusion Models/d46ce965-2153-45b4-9566-3f0d4d7d97a6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e1e57ff62204ad593ce0bb5f9f715df61365f36b --- /dev/null +++ b/2025/SpinMeRound_ Consistent Multi-View Identity Generation Using Diffusion Models/d46ce965-2153-45b4-9566-3f0d4d7d97a6_content_list.json @@ -0,0 +1,2290 @@ +[ + { + "type": "text", + "text": "SpinMeRound: Consistent Multi-View Identity Generation Using Diffusion Models", + "text_level": 1, + "bbox": [ + 119, + 128, + 875, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Stathis Galanakis", + "bbox": [ + 137, + 202, + 285, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Alexandros Lattas1", + "bbox": [ + 401, + 203, + 553, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bernhard Kainz $^{1,2}$", + "bbox": [ + 316, + 220, + 464, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Stylianos Moschoglou", + "bbox": [ + 671, + 203, + 857, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Stefanos Zafeiriou", + "bbox": [ + 596, + 220, + 751, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Imperial College London, UK", + "bbox": [ + 375, + 255, + 622, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ FAU Erlangen-Nürnberg, Germany", + "bbox": [ + 351, + 273, + 640, + 292 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bb20f158f29f6968ffdbb355f3b07c740b4a6bb02d4d46669aed8360dc346502.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 305, + 279, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0836e9de2f71a06f81d6eb5edef1c613f5e346d0dd77134120eb276784dd2460.jpg", + "image_caption": [ + "Figure 1. SpinMeRound is a multiview diffusion model which generates human portraits from novel viewpoints. Given a single or multiple views, our method produces high-fidelity images along with precise surface normals, ensuring accurate 3D consistency across perspectives." + ], + "image_footnote": [], + "bbox": [ + 99, + 474, + 153, + 523 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ed72d1ea1c1271ea1991e5e982f2596ff60ecac855921160904fd561a1c4f517.jpg", + "image_caption": [ + "Single Image to novel views" + ], + "image_footnote": [], + "bbox": [ + 156, + 474, + 218, + 525 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9cef7845c47785dbb3a477bdd4ac232891b704856525474066cb00a26fb9295e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 474, + 269, + 525 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9797b68e5950dbf682bd41df45d303ba6b4949027585fccffd657e24592ed411.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 306, + 488, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/530921576dfff87d0e844bc7f6510e30a22930cc7e80ec03ff47847db3878c73.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 315, + 474, + 377, + 523 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/32c7851a65ada9ffdcda4b4351029d787bc8fbc58f94fb16213535dee3664c52.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 379, + 474, + 434, + 523 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2b9105ea6dd7efc73c974d913f9068354708c6cf68e0a105dbf3c1f14121679d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 441, + 474, + 504, + 523 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f87676e2202ef8f80c41a804654444a0e024fcacaa3220784c9a46bd8b30ece7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 535, + 306, + 906, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/48dd5b39773110129c5c02a2b84a8f47c3522ad67e436d4a2eb2af4c0c269975.jpg", + "image_caption": [ + "Sparse views to novel views" + ], + "image_footnote": [], + "bbox": [ + 537, + 474, + 584, + 525 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7b096e34531b1e53e1168df01186eb32f2e623a78855597d0f2e71a9412c9c2c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 474, + 653, + 523 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7d06defaff480f16169d3b7f57c2242016f01629842184614866f355640c7317.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 474, + 715, + 523 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a9a96543f3813a273438823097800181f7ffd6a7314e882770b70fa24392ad3e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 474, + 779, + 523 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b449fcc0465bae1bad11a9ad3e3dd469d9ba6428df6cd4fa0da94ff5d8e89b7a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 785, + 474, + 841, + 525 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ad00c0d328cb3b22eea6be8f8e2073676ec03c2bd144379052d0e55c141a8282.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 843, + 474, + 903, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 603, + 326, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite recent progress in diffusion models, generating realistic head portraits from novel viewpoints remains a significant challenge in computer vision. Most current approaches are constrained to limited angular ranges, predominantly focusing on frontal or near-frontal views. Moreover, although the recent emerging large-scale diffusion models have been proven robust in handling 3D scenes, they underperform on facial data, given their complex structure and the uncanny valley pitfalls. In this paper, we propose SpinMeRound, a diffusion-based approach designed to generate consistent and accurate head portraits from novel viewpoints. By leveraging a number of input views alongside an identity embedding, our method effectively synthesizes diverse viewpoints of a subject whilst robustly maintaining its unique identity features. Through experimentation, we showcase our model's generation capabilities in full head synthesis, while beating current state-of-the", + "bbox": [ + 88, + 643, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "art multi-view diffusion models. Project page: spin-meround.github.io.", + "bbox": [ + 511, + 604, + 906, + 635 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 662, + 645, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in deep learning have brought significant progress to fundamental computer vision tasks, notably in image and video generation [6, 56]. The introduction of Diffusion Models [16, 28] has transformed these domains by enabling the generation of high-quality visual content, facilitated by the use of large-scale datasets [5, 56]. Despite these developments, the task of generating consistent and accurate head portraits from a single input image still remains a challenging problem. This difficulty is primarily attributed to the limited availability of comprehensive 3D facial datasets, which constrains the training of models capable of reliably capturing and reconstructing the detailed structure and variations of human faces.", + "bbox": [ + 509, + 688, + 906, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A common practice for modeling human heads from", + "bbox": [ + 532, + 885, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14346", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a single image, incorporates the use of a 3D Morphable Model (3DMM) [4, 7, 8, 37], to represent the facial shape along with an appearance model [17, 21, 34]. However, given typical training data and the difficulty in modeling complex hairstyles, these methods focus solely on the facial region and avoid or miss the full head and hair. The seminal work of Neural Radiance Fields (NeRF) [43] led to an explosion of works on neural rendering of scenes that could not easily be modeled with textured meshes. In addition, pairing such implicit representations with generative models led to a wide variety of approaches [2, 9, 44] that pushed the boundaries on the synthesis of novel facial views, achieving high quality and control. More recently, Panohead [1] first showcased high-quality $360^{\\circ}$ head portrait synthesis. However, because of its adaptive camera training scheme, the back-head synthesis typically contains many artifacts [35], and its inversion on \"in-the-wild\" images is challenging and requires complex fine-tuning [55].", + "bbox": [ + 89, + 90, + 480, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, diffusion models [16] demonstrated superior performance over GAN-based methods in image generation and have achieved great quality in human generation tasks [56]. However, achieving multi-view consistency remains a significant challenge. Despite the lack of accurate 3D datasets, recent advances, such as Score Distillation techniques [49, 73], represent an initial step toward leveraging the 2D generation strengths of diffusion models to construct 3D content without extra training. Nevertheless, these approaches are computationally intensive, require intricate constraints, and do not consistently yield photorealistic results. Meanwhile, multi-view diffusion architectures, employing video diffusion models as a backbone framework [42, 63], still remain very resource-demanding for generating a single novel view, as they rely on comprehensive camera trajectories to generate coherent central objects. Closer to our work, Zero123 [39] introduced a view-conditioned diffusion model that incorporates view features and camera information into the diffusion process. However, the generated images lack strong multi-view consistency and are of low quality, which restricts their performance in photorealistic 3D generation. Although Diff-Portrait3D [25] enables novel view generation through a diffusion process, it still focuses only on near-front views and utilizes a image-conditioned camera mechanism. Other closely related works are Era3D [36] and Morphable Diffusion [10], which generate fixed camera viewpoints, thus limiting their ability to produce full head portraits. The recently proposed Cat3D [20] presents a promising solution by efficiently integrating a number of input views with specified camera poses to achieve a consistent novel viewpoint generation. It integrates 3D attention layers for efficiently sharing common information between all views, along with robust camera pose feature maps. Still, Cat3D is limited by its lack of focus on human generation and is currently", + "bbox": [ + 91, + 369, + 483, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "unavailable as an open-source tool.", + "bbox": [ + 511, + 90, + 746, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present SpinMeRound, a multi-view diffusion model designed to generate novel high-fidelity views of a given human face. In addition to facial images, our model also generates the corresponding normals, which are typically available for human data and, as we show, improve the model's performance and consistency on intricate facial features. Moreover, we show that conditioning the model on an identity embedding and one or more input views of a subject during inference, we can not only sample but also reconstruct multi-view consistent images from \"in-the-wild\" facial images. Our method can accurately synthesize photorealistic head portraits from various angles while preserving essential identity characteristics, which can be used to represent or reconstruct 3D scenes. Given the lack of open-source large-scale multi-view head datasets and the problematic nature of the permissions of such datasets, our method is solely trained on synthetic data acquired using Panohead [1], making this work accessible to experiment and build with. Overall, in this paper:", + "bbox": [ + 511, + 108, + 903, + 393 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce SpinMeRound, a multi-view diffusion model conditioned on identity embeddings and a number of views that generates novel perspectives of an input subject and their respective normals.", + "- We present a novel sampling strategy which, given a single \"in-the-wild\" facial image, generates consistent views encompassing the whole head.", + "- We explore its potential by comparing it with current state-of-the-art multi-view diffusion-based methods and showcase superior results in full-head portrait generation." + ], + "bbox": [ + 513, + 396, + 903, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 564, + 656, + 579 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Face Modeling", + "text_level": 1, + "bbox": [ + 511, + 590, + 663, + 606 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive research has been dedicated to representing 3D faces through a combination of texture maps and 3D meshes, starting with fundamental 3D Morphable Models (3DMM) [4, 7, 8, 37]. These approaches, however, primarily focus on accurately capturing only the frontal region of the head whilst lacking in integrating finer head details such as hair, wrinkles and wearable items. Dealing with this, recent studies have integrated implicit representations for 3D face modeling. Methods such as [23, 45, 72] integrate Signed Distance Functions (SDFs) whereas other [2, 9, 18, 69] use Neural Radiance Field (NeRF) models [43] for generating photorealistic results. Levering the powerful triplane representation [9], RTRF [62] generates nearfrontal views in real-time whilst Panohead [1] introduces an adaptive camera training strategy for enabling full head portrait synthesis. An extension of it is 3DPortraitGAN [67] focusing on all-around upper body generation. As this work focuses only on full-head generation, Panohead can effectively be used to acquire synthetic portrait datasets.", + "bbox": [ + 511, + 613, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14347", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0a9dc190670b992009c40dc82a72a6923cf4261a17c3c3dfe53def1d75fb96fd.jpg", + "image_caption": [ + "Figure 2. Overview of SpinMeRound: Starting with a number of input conditioning views, the identity embedding $\\mathbf{W}$ is extracted via a Face Recognition network (ArcFace [13]). Both the conditioning and target views are then encoded and combined with corresponding ray coordinate maps that represent camera poses. After the sampling step, our method synthesizes photorealistic images from novel angles, along with their associated shape normals $\\mathcal{N}$ ." + ], + "image_footnote": [], + "bbox": [ + 98, + 89, + 893, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Diffusion Models", + "text_level": 1, + "bbox": [ + 89, + 380, + 261, + 396 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Face Generation Diffusion models (DMs) [28] have proven their generative abilities by beating the well-established GANs in image synthesis tasks [16]. The availability of large-scale datasets has revolutionized text-to-image generation [56] and video generation [5, 6] tasks. For face synthesis, a range of methods have emerged to tackle essential tasks, including 3D avatar creation [65, 70], avatar reenactment [14, 15, 33] and texture map generation [19, 46]. We adopt the concatenation strategy outlined in [19, 46] to generate novel viewpoint images and their corresponding shape normals simultaneously. Based on Stable Diffusion [56], Arc2Face [13] and InstantID [64] generate facial images given an input subject. Notably, Arc2Face exhibits strong generalization abilities in facial image generation, leveraging an up-sampled subset of WebFace42M [74]. Also, it introduces a robust identity conditioning mechanism, which is integrated into our method.", + "bbox": [ + 88, + 404, + 485, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Novel View Image Synthesis DreamFusion [49] introduces the use of Score Distillation Sampling (SDS), incorporating a pre-trained text-to-image diffusion model [56] alongside a NeRF model, to synthesize 3D objects from text prompts. In this way, its authors proved that they can efficiently generate 3D objects despite the lack of large-scale datasets; by exploiting the generalization abilities of an image generation network. Although subsequent studies focus on better distillation strategies [51, 53, 73], the aforementioned approaches are time-consuming and require complex balancing and additional constraints [20]. Closer to our work, ID-to-3D [3] and Arc2Avatar [22] use score distillation to generate 3D facial geometry, lacking though photorealistic results. Zero123 [39] introduces a multiview diffusion model conditioned on a reference image", + "bbox": [ + 88, + 672, + 486, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and a camera pose. Following it, methods such as One-2-3-45 [38], SyncDreamer [40], Consistent1-to-3 [66] and Cascade-Zero123 [68] further focus on multi-view consistency by introducing priors during the denoising process. Studies such as Zero $123++$ [59], Era3D [36] and Morphable Diffusion [10] generate fixed viewpoints given the input image, without being able to handle arbitrary views. In Cat3D [20], a general multi-view diffusion model is introduced, using a powerful camera pose conditioning mechanism while using 3D attention layers. Although it has been a robust method for novel view synthesis, it does not focus on facial novel view synthesis, lacks shape normal generation capabilities and is a closed source framework. Diff-Portrait3D [25] showcases novel view capabilities, given an input facial image. Unlike our method, which employs image-free camera conditioning, DiffPortrait3D employs an image-driven approach for the desired output viewpoints and can generate only near-front views. More recent works [50, 61] have explored diffusion-based frameworks for generating controllable and animatable avatars. Concurrently to our approach, Pippo [29] and DiffPortrait360 [26] demonstrated $360^{\\circ}$ avatar generation, without, though, being able to sample novel identities.", + "bbox": [ + 511, + 381, + 908, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Novel View Video Synthesis Recently, video generation models [5, 6, 24] have proven their ability to generate photorealistic models. Methods such as IM-3D [42], V3D [11] and SV3D [63] integrate an off-the-shelf video diffusion model for generating novel viewpoints of a reference object. Still, the video generation step makes these methodologies computationally demanding [20]. Additionally, they are often restricted by the requirement for specific camera trajectories, typically revolving around a central subject. In contrast, our approach focuses on implementing multi-view diffusion networks that handle unordered camera poses.", + "bbox": [ + 511, + 734, + 910, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14348", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a6f86be2864c23ba016d0af10dc59e82a18bd67199a85d0ba5504c46c741d04b.jpg", + "image_caption": [ + "Figure 3. Samples generated by SpinMeRound on \"in-the-wild\" images: Given only the input images (shown in a small box at the top center), our method produces high-fidelity images from novel angles, along with corresponding shape normals $\\mathcal{N}$ ." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 906, + 223 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 273, + 181, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SpinMeRound incorporates a multi-view diffusion model, trained on a 3D facial dataset, capable of generating a full head portrait of an input subject, given their identity embedding and a number of views (Fig. 2). In the following sections, we first describe the proposed diffusion model, its training scheme, and the proposed sampling strategy given a single input view.", + "bbox": [ + 89, + 299, + 483, + 405 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Multi-View Diffusion Model", + "text_level": 1, + "bbox": [ + 89, + 419, + 344, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SpinMeRound employs a latent multi-view diffusion model alongside a powerful identity conditioning mechanism [47], enabling the generation of novel views of an input scene containing a person. Our approach leverages as conditioning inputs $M \\in \\{1,3\\}$ pairs of sparse views of a subject, their respective shape normals and their associated camera poses. Let $\\mathbf{I}_i$ denote a picked conditioned view and $\\bar{\\mathbf{I}}_i$ a cropped and aligned version of $\\mathbf{I}_i$ . We utilize a pre-trained face recognition network (ArcFace [13]) $\\phi$ to extract the identity embedding vector $\\mathbf{w} = \\phi(\\bar{\\mathbf{I}}_i) \\in \\mathbb{R}^{512}$ capable of incorporating crucial identity features. Then, we inject the identity information into the diffusion model as proposed in Arc2Face [13]. In particular, a text prompt of \"a photo of $$ person, is fed to a CLIP text encoder [52] followed by the tokenization step. Followed by padding, the identity embedding vector $\\mathbf{w}$ replaces the corresponding $$ token resulting in $\\mathbf{s} = \\{e_1,e_2,e_3,\\hat{\\mathbf{w}},e_5\\}$ , where $\\hat{\\mathbf{w}} \\in \\mathbb{R}^{768}$ is the padded version of $\\mathbf{w}$ . The vector $\\mathbf{s}$ is fed to the encoder $C$ , and finally we retrieve the corresponding conditioning vector $\\mathbf{c} = C(\\mathbf{s}) \\in \\mathbb{R}^{N \\times 768}$ , where $N$ is the maximum sequence length.", + "bbox": [ + 89, + 441, + 483, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proposed latent multi-view UNet [57] is similar to the one introduced in Cat3D [20]. Our diffusion model is designed to concurrently generate $P = (M + K) = 8$ pairs of facial images and their corresponding normals $\\mathcal{N}$ , given an input conditioning vector $\\mathbf{w}$ , a number of $M$ views and the $P$ camera poses. Let $I_{i}^{cond}$ represent the input human face images, $I_{j}^{tgt}$ the target face images from novel viewpoints, and $\\mathcal{N}_{i}^{cond}, \\mathcal{N}_{j}^{tgt}$ their respective shape normals, where $i \\in \\{1,2,\\dots,M\\}$ , $j \\in \\{1,2,\\dots,N\\}$ with", + "bbox": [ + 89, + 761, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$M$ being the maximum number of the input conditional views and $N$ the number of target viewpoints. Following [19, 20, 46], we employ a pre-trained AutoEncoder of Stable Diffusion 1.5 (SD 1.5) [56], consisting of an encoder $\\mathcal{E}$ and a decoder $\\mathcal{D}$ . For each input conditioning image, we first obtain the corresponding latent feature maps: $\\mathbf{z}_{I_i^{cond}} = \\mathcal{E}(I_i^{cond}) \\in \\mathbb{R}^{4 \\times 64 \\times 64}$ and $\\mathbf{z}_{\\mathcal{N}_i^{cond}} = \\mathcal{E}(\\mathcal{N}_i^{cond}) \\in \\mathbb{R}^{4 \\times 64 \\times 64}$ . These are then concatenated channel-wise. Similarly, we apply the same procedure to extract the latent vectors for the target viewpoints. The corresponding camera pose information is incorporated for both conditioning and target viewpoints using the mechanism proposed in Cat3D [20]. For each of the $P$ views, the latent feature maps $\\mathbf{z}_i$ are concatenated channel-wise with the respective ray representation maps [20, 58] $\\mathbf{r}_i^{cond}, \\mathbf{r}_j^{tgt} \\in \\mathbb{R}^{149 \\times 64 \\times 64}$ , which encode the ray origin and direction. A binary mask $\\mathbf{m} \\in \\{\\mathbf{0}, \\mathbf{1}\\}^{1 \\times 64 \\times 64}$ is then appended to differentiate between conditional and target latent vectors. Finally, we retrieve the conditioning and target latent feature maps $\\bar{\\mathbf{z}}_i^{cond} = \\{\\mathbf{z}_{I_i^{cond}}, \\mathbf{z}_{\\mathcal{N}_i^{cond}}\\mathbf{r}_j^{cond}, \\mathbf{m}_j^{cond}\\}$ and $\\bar{\\mathbf{z}}_i^{tgt} = \\{\\mathbf{z}_{I_i^{tgt}}, \\mathbf{z}_{\\mathcal{N}_i^{tgt}}\\mathbf{r}_j^{tgt}, \\mathbf{m}_j^{tgt}\\}$ .", + "bbox": [ + 511, + 273, + 906, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As proposed in video diffusion models [5], we initialize our method using a pre-trained LDM model (Arc2Face [47]) trained on large-scale datasets while adding additional attention layers, connecting the multiple latent feature maps. As in [20], we integrate 3D attention layers [60] by adding them between the original 2D self-attention layers of the LDM, and we fine-tune all layers of the multi-view UNet for improved multi-view consistency.", + "bbox": [ + 511, + 603, + 908, + 726 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Training Details", + "text_level": 1, + "bbox": [ + 511, + 739, + 671, + 756 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We begin from the publicly available Arc2Face [47], and we adhere to the training schemes introduced in [5, 20]. Specifically, we adapt the EDM framework [30] to Arc2Face, training it for 31,000 iterations on its dedicated dataset. To accommodate the input latent feature maps, we expand the input and output convolutional layers channels, initializing these by copying the existing weights to the shape-normal channels while randomly initializing the camera-pose dimensions. As in [20], we shift the log-to-signal ratio by", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14349", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\log (\\mathrm{N})$ , where $\\mathbf{N}$ is the number of the target images $(\\mathrm{N} = 7)$ We randomly select a conditioning view that includes part of the frontal face during each training iteration, as required for the identity embedding extraction. We then randomly pick the N target images and calculate the relative camera angles. To enhance the dataset, we replace the white background with a random color, in $50\\%$ of the samples. All the training viewpoints are encoded using the encoder $\\mathcal{E}$ ,with noise added only to the target latent vectors, while conditioning vectors remain unchanged. Following the classifier-free guidance (CFG) training scheme [27], with a probability of $\\mathcal{P}_{uncond} = 0.15$ , we randomly replace the identity vector with the empty string, and the conditioning images with zero-ed ones. We first train the model conditioned on a single view for $600\\mathrm{k}$ iterations. Then, for the additional 1M iterations, we vary the conditioning views by randomly choosing O, 1 or 3 conditioning views, corresponding to 8, 7 and 5 target views, each with a probability of $\\mathcal{P}_=1 / 3$", + "bbox": [ + 89, + 90, + 480, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Novel view sampling", + "text_level": 1, + "bbox": [ + 89, + 372, + 285, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SpinMeRound synthesizes novel views for an input subject I, given a number of views. In this section, we introduce a robust sampling strategy designed to produce a large number of consistent views that comprehensively cover a full head, given only a single input image. Achieving consistency across viewpoints requires a carefully structured camera pose selection order. Therefore, we employ a three-step sampling process: a) aligning input views and extracting the corresponding shape normals, b) generating anchor images that provide complete coverage of the $360^{\\circ}$ human head and c) synthesizing intermediate views by leveraging both the input views and the closest anchor images.", + "bbox": [ + 89, + 393, + 482, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Alignment and Shape Normals generation Given the input image $\\mathbf{I}$ , we extract its identity embedding $\\mathbf{w}$ as described in Sec. 3.1. Then, we obtain $\\hat{\\mathbf{I}}$ , a cropped and aligned version of $\\mathbf{I}$ , using the alignment procedure presented in Panohead [1]. To generate the shape normals $\\mathcal{N}$ , we treat this as an in-painting task, thus retrieving the normals through the conditional guidance sampling approach described in Relightify [46]. Specifically, using the aligned image $\\hat{\\mathbf{I}}$ and its identity embedding vector $\\mathbf{w}$ , we employ a binary visibility mask $m$ , marking only the image channels as visible and setting the shape normal channels as nonvisible. By applying the \"channel-wise in-painting\" algorithm, the corresponding shape normals are retrieved. This process uses the EDM sampler presented in [30] alongside the DDPM [28] discretization steps and runs for 50 steps. A more detailed presentation of this approach is presented in the supplemental material.", + "bbox": [ + 89, + 595, + 482, + 851 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Generating anchor and intermediate views SpinMeRound can generate any arbitrary viewpoint, given the input", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "aligned facial image $\\hat{\\mathbf{I}}$ and the corresponding shape normals $\\mathcal{N}$ . However, since it was trained to generate only a limited number of views per sampling process, a carefully designed sampling strategy is essential to produce a wide range of output views. Since the target subject is centered in the scene, we first generate $M = 7$ anchor images $\\mathbf{A}_i$ and corresponding anchor shape normals $\\mathbf{A}_{\\mathcal{N}_i}$ , $i \\in \\{1, \\dots, M\\}$ , covering a $360^\\circ$ angle range of the subject $(\\pm 45^\\circ, \\pm 90^\\circ, \\pm 135^\\circ, 180^\\circ)$ , as proposed in [20]. Using these anchors, any number of intermediate views can then be synthesized by conditioning on image triplets $\\{(\\mathbf{I}, \\mathcal{N}), (\\mathbf{A}_k, \\mathbf{A}_{\\mathcal{N}_k}), (\\mathbf{A}_l, \\mathbf{A}_{\\mathcal{N}_l})\\}$ , where $k, l$ where represent the closest anchor images. This approach ensures that each intermediate view remains consistent with both the input aligned image $\\hat{\\mathbf{I}}$ , and the closest already generated views. For both sampling processes, we use the EDM sampler facilitated by the EDM discretization steps [30], while it runs for 50 steps with a guidance scale set to 3. This method enables the generation of 48, 88, or more novel views for a single input subject $\\mathbf{I}$ , depending on the chosen angle step, thus covering the entire scene.", + "bbox": [ + 511, + 90, + 903, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 422, + 645, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Training Dataset", + "text_level": 1, + "bbox": [ + 511, + 448, + 679, + 464 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training such a model requires a large multi-view dataset containing a great number of subjects $S$ . For each person $S_{i}$ , it is necessary to acquire a set of images $\\mathbf{I}_k^i$ , their corresponding shape normal maps $\\mathcal{N}_k^i$ , camera poses $\\mathcal{C}_k^i$ alongside with their corresponding identity embedding vector $\\mathbf{w}^i$ , where $k \\in \\{1, 2, \\dots, N_i\\}$ and $N_{i}$ is the number of the available views for the $i$ -th scene. Due to the lack of such a dataset, we create a large-scale synthetic dataset using the publicly available Panohead [1]. We first sample $\\sim 10k$ subjects and manually remove instances with artifacts in the back of the head, resulting in $\\sim 7k$ distinct identities. We render images from 125 different viewpoints for each person to cover the entire head. Simultaneously, we obtain each subject's facial shape by extracting their opacity values from the triplane feature maps and then applying the marching cubes [41] algorithm. We render their respective normal maps using the acquired facial shape using Pytorch3d [54]. All the images depicting a frontal head are fed to the ArcFace [13] network to extract their corresponding identity vectors $\\mathbf{w}$ . All in all, we end up with a synthetic dataset containing about 7k individuals, rendered from $N = 125$ different angles, the respective shape normals, camera poses, and their identity embedding vector $\\mathbf{w}$ .", + "bbox": [ + 511, + 470, + 903, + 818 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Qualitative Comparisons", + "text_level": 1, + "bbox": [ + 511, + 828, + 741, + 844 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1. Novel View Synthesis comparisons", + "text_level": 1, + "bbox": [ + 511, + 849, + 799, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we present a qualitative comparison between our model and other state-of-the-art multi-view dif", + "bbox": [ + 511, + 869, + 903, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14350", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d818a6599f3e1d45778cf0e9495b29eab1d9c8dfca098286f33807a469539424.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 85, + 158, + 138 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ae710b886140c2ad071d36e266b11994a06409dae5e52008c7921e9150a8aa6e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 138, + 153, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2ff3e69530abb9e7d0fcd9766d9e34fa4b7973ec676a374f8b63e2dd22a60728.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 165, + 87, + 217, + 138 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 165, + 138, + 218, + 186 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 165, + 188, + 218, + 237 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d01daae13de9efd69a286b34d0b557262e185280f2c2f745c7cd34482464b632.jpg", + "image_caption": [ + "(a) Input" + ], + "image_footnote": [], + "bbox": [ + 94, + 237, + 153, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 165, + 138, + 218, + 186 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 165, + 188, + 218, + 237 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/069db5a8fefd94e18eb7dbf550ca2addf36946ed8b3448354e0e854e76ddb5de.jpg", + "image_caption": [ + "(b) SpinM" + ], + "image_footnote": [], + "bbox": [ + 165, + 237, + 218, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7b34f54acfeb9aa01551968c720205a1307745f58f40ebabe0821aa8b5d946f1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 87, + 272, + 138 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1f7b420fed12bfdff778f94326eee91c6161026a2b8c42194b83c09332b30732.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 138, + 274, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f9455abeb235b7c6e00fea12cb091e21491ba3cbc2f1d88c025240e8d7e38691.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 188, + 274, + 237 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/24197544f59a456f4de064987c0dfaddf121f01757521ccd37f211c76f6af6c5.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison between SpinMeRound, Zero123-XL [12], SV3D [63] and DiffPortrait3D [25] at angles $\\{\\pm 45^{\\circ}, + 180^{\\circ}\\}$ . It is clear that SpinMeRound effectively generates high-quality novel views from \"in-the-wild\" input images, whilst SV3D produces distorted outputs, Zero123-XL generates unnaturally squared avatars and DiffPortrait3D cannot handle such angles." + ], + "image_footnote": [], + "bbox": [ + 220, + 237, + 274, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/004d706813603a4f671352741f48bbe3ec5444279f64a124c91ff445b97ab705.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 87, + 334, + 138 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cd9f16d8849df2ca6ddc258a76cdc428aeb5fc011b1eb3a121e2dd4c0c3c6427.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 138, + 334, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8699bf09479e5df4a791f035e7e857f791d8893ccd5751aa00689989411b851e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 188, + 334, + 237 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/04f10cdcd91e99a63f65548fe9a31d3f6e3b756baacde7ea48d1202118069fe6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 237, + 334, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f9ab34fc4228f91e4b35cc0a9b2c87af9b779635059ffc15675ad92b46121c58.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 349, + 87, + 400, + 138 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7a34fb8a25f158da6ed290436f6758510e6f5e6768135c374d15293a2b891898.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 349, + 138, + 401, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ffaefcfa111f2fd13b2308ada2a267990fb6bef9541358b42d8d263c3736e63a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 344, + 188, + 395, + 237 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/05377c43ce4bfabeaf86a483c9d0f0332b5d49d9fadc46d525aa82a469e43bbc.jpg", + "image_caption": [ + "+45°" + ], + "image_footnote": [], + "bbox": [ + 344, + 237, + 403, + 282 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8f0dc5eebd0e6547c61339c20816af7ddec62b9ba2dbf5ee9e68fe5efabd9b19.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 87, + 465, + 138 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5992d458f8c2f1412672ed6d1860ff4feb1c114500245b79cc729b971ea6bbb3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 138, + 465, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/646d6cea1eedbcf16f53c429e2b0c48ba73f8579eb752e7e74d89d720426a0a4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 188, + 465, + 237 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d68efbfa6f6e8011aeb303eb29e74fcc41597980a617f734d7f2638bfb7225f3.jpg", + "image_caption": [ + "(c) SV3D [63]" + ], + "image_footnote": [], + "bbox": [ + 410, + 237, + 465, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3d4438f889f9eb3241095eee86aac301dbd4393baa704d7d674e803200f893aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 472, + 87, + 524, + 138 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bfc6899dae281503a0e41446a938f0cfcb7370cbfd45cf8c53f3111bc0f8879b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 470, + 138, + 524, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9895b41b7c14eec2b22cc2fdbc3cc210f81a40fba57e442c68c56c371166da1b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 482, + 189, + 522, + 237 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b7289fa4afeb371cd90ea21dfc8336a3553c566058bbfa657ce58ed3a313ca9d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 475, + 238, + 524, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dddd58d01b24ef098c74875ec4b0454845ef662024e5cd971845db10250aabb6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 92, + 584, + 137 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bfd5f75cfee0222a914e12338e13ad309bc70815f18b6c2b2b152efbb3063d55.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 542, + 142, + 581, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fe47f4ec4edcc7094c4f8f417f1a675539611e03bf371db66bb55fc2cae8f3e4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 191, + 580, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/be5f98809405363546d5a098013b232d40c8bcb04f60cd145987d70dc1f89cbf.jpg", + "image_caption": [ + "(d) Zero123-XL [12]" + ], + "image_footnote": [], + "bbox": [ + 537, + 241, + 586, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3333863872599a0617ff4967ca502cc4c2b16de77ff3c3eb5e2057fa876d81e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 92, + 653, + 137 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2104a4404a3747919545ce7b0871b7fb004df43e93f997be76e23cafb7171214.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 143, + 645, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/81e812aa84697c619614684bc481917d3f86153a894794ab2374ad2dfea9301e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 191, + 643, + 233 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1310ecbf5fb84e17db969ff090be4d6d83676c7c78c714b2f91fdebac68b0402.jpg", + "image_caption": [ + "-45°" + ], + "image_footnote": [], + "bbox": [ + 594, + 241, + 653, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8099ab0920fcc49325ca7139a3f2fbc2da4d68e5cf1e6484a0af1915c6bbc485.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 92, + 704, + 137 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/742f95c7c0bb80ab11f81dffbe527c885003f2799149aec4e96049cb9f188fc7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 661, + 140, + 710, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6e16733e0447cc8a3d699a661d34e500328b91624f684fc9c3bb6306fad732d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 668, + 191, + 705, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/81591603fcab7abd2683843363ecdcf0a7b7804876d59bd00345bf6511d77c13.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 237, + 704, + 282 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e92c697f0b1d5b91a372443cdae76534faf8521cf4c86a3baec62704ab63154e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 87, + 841, + 137 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5855816e181b8d0732336cdb3e4f3c34441a0726f086c340fb961a33c26789b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 138, + 841, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7223f9288ef5a6f235757d5b3d3e4ed5e821e1ae664c82ce6270cbe4917dcae2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 188, + 841, + 237 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bc1fd73df9540eeef1b69d20603030eca47620f2edeb6b0d2bbc125b6861268c.jpg", + "image_caption": [ + "(d) DiffPortrait3d [25]" + ], + "image_footnote": [], + "bbox": [ + 715, + 237, + 841, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/50c2ad859c6fbe0ad39011f4255a8d2c4bd1d5203b900d112a38c7e7cfa92ccf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 841, + 87, + 903, + 137 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/62ea2a47c4edb1b9de90a732fb9a30264ebbd33d5adaf490269b50c4749e42f9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 839, + 138, + 903, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6cbbf5a71bfd5097b4d288f6aecccf580bfff7b2228045bf134855d72fed7f50.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 841, + 188, + 903, + 237 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d02a7a584aa561371883b51a416a3a2810dd5f8cd4cb3caaf42f0d1ecd6df08e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 841, + 237, + 901, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "fusion models, SV3D [63] Zero123-XL [12] and DiffPort- trait3D [25], focusing on their output under $\\pm 45^{\\circ}, + 180^{\\circ}$ angles. SV3D, a latent video diffusion model, creates $360^{\\circ}$ videos from a single input image. Zero123-XL, a multiview diffusion model, synthesizes novel perspectives based on the input image and specified camera poses. DiffPort- trait3D [25] generates novel views by conditioning the desired camera on an input image and refining the initial noise through a fine-tuning step to achieve optimal results. Figure 4 demonstrates scenarios where the input \"in-the-wild\" images (Fig. 4a ) are fed to these 4 different networks and their respective generations under $\\pm 45^{\\circ}, + 180^{\\circ}$ viewpoints. As illustrated, SpinMeRound efficiently generates accurate novel viewpoints of the input subject, while SV3D distorts the input image in its generated views, Zero123-XL produces unnatural, square-like facial avatars and DiffPort- trait3D cannot handle those angles at all.", + "bbox": [ + 89, + 378, + 483, + 633 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f3027b182cc163b281867aba12807a0605f15cfd46e26d26986f75e72c96a524.jpg", + "image_caption": [ + "Figure 5. Given the input identities (shown in the small boxes), we present the results of novel view synthesis after applying 3D Gaussian splatting [31] to the views generated by our model." + ], + "image_footnote": [], + "bbox": [ + 91, + 643, + 483, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. 3D Reconstruction", + "text_level": 1, + "bbox": [ + 89, + 801, + 269, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate our model's consistency in generating novel views from an input image, we assess its ability to reconstruct an input identity through Gaussian Splatting (3DGS) [31]. More specifically, given an input identity, we apply the sampling strategy outlined in Section 3.3 to", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "generate 48 novel views. These views are then used to reconstruct the identity through Gaussian Splatting. As in [20], we modify the provided 3DGS code to incorporate the LPIPS [71] loss between the ground-truth and the generated viewpoints. In this way, we can deal with small inconsistencies between nearby viewpoints. We present the novel view synthesis of two input identities (small squares) in Figure 5, which clearly illustrates that our proposed methodology can generate consistent subjects.", + "bbox": [ + 511, + 378, + 906, + 513 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/83c2a33f3fe24124d532df0ebdc66ae96eacebbeebb01ace9391cda9100825ad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 526, + 612, + 603 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5f08c3a2b94be8fe7126488cecb580ad1b64f6df493cf2f251995ca1873a33b4.jpg", + "image_caption": [ + "Figure 6. Samples generated using unconditional sampling. Spin-MeRound can generate novel identities without any prior input." + ], + "image_footnote": [], + "bbox": [ + 517, + 604, + 609, + 640 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f3a6a29d299eb3c8e5d3a7fa25c4338002994d8cc9ce0e91c639b3e3bcb0ec41.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 527, + 661, + 566 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8a6e445ed53d713c46723fd5cb65b9b50c3c8f114be118ad62ae752690387680.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 568, + 661, + 603 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2c9149d69baed028d58782a1fd05123c48e03a0ad97689fbe3d1040e104d68a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 604, + 661, + 640 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6a2c2201223e3b27f0ab06f9ef79c2dbd1480ef321a1eafbe54406c626568ee9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 661, + 527, + 710, + 566 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b02f7a889c81da213224af37d6daa0fad29a0ce0b8dc870cb89587f7f6ad77c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 661, + 568, + 710, + 603 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/be729dd9b6f84d5572990928dc65a6876a66b473a7578ad882ab7d5851eaeb88.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 661, + 604, + 710, + 640 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/302bffb2332ecc37d3a4884bfbbc6d8b3dbb4e6228e4ec66678f84e09878ce1f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 527, + 841, + 566 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1335d900a62c968999befd2ba1c646433ddfeb5cca524dc18712ccb23158c099.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 566, + 841, + 603 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/eb548ecab6db5ee896aeffa79fd5cdb1171897f10b7d295b93fc05e7d3a8a75b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 604, + 841, + 640 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/175221f93b13db844f175279c309b5bb6d39da69577f9e215dee1cec50602c19.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 841, + 527, + 901, + 566 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d19dfe15285c7e9446ade34584deabbb0520e2b031e2ad4a5ce1201a8d51cf0f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 841, + 568, + 901, + 603 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/358f696296781c523fddaba5ffaa22fb87c25e9984665f6f211259046c2a6339.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 841, + 604, + 901, + 640 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Unconditional Sampling", + "text_level": 1, + "bbox": [ + 511, + 708, + 736, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SpinMeRound is trained following the Classifier-Free Guidance training scheme. Thus, our architecture can generate novel multi-view identities without any prior conditioning. By setting the input identity embedding equal to the empty string, our model can generate a novel identity along with different viewpoints of that identity. Figure 6 presents some examples of those identities.", + "bbox": [ + 511, + 729, + 906, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5.Quantitative Comparisons", + "text_level": 1, + "bbox": [ + 511, + 848, + 751, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method's ability to generate novel views with the current state-of-the-art models: a) Eg3D [9] and", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14351", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e83f4079a32afa7ad8ff1c5e0867c4f9b4e3bc144a006a9a40a0d6d3aa1edc1a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodL2 ↓LPIPS↓SSIM↑ID Sim [13] ↑ID Sim [48] ↑
NeRFEg3d [9]0.0250.40.550.310.89
basedPanohead [1]0.0120.320.650.270.88
Zero123 [39]0.1950.5150.550.1690.44
Diffusion basedZero123-XL [12]0.1980.510.5630.1180.442
SV3D [63]0.0870.410.6600.360.881
DiffPortrait3d [25]0.10.50.350.550.887
SpinMeRound (Ours)0.0330.30.730.610.911
", + "bbox": [ + 91, + 88, + 480, + 165 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Panohead [1] which are NeRF-based approaches generating frontal and full-head portraits respectively, b) Zero123 [68], Zero123-XL [12] and DiffPortrai3D [25], which are multiview diffusion models and c) SV3D [63], a multi-view video diffusion model. For SV3D, we selected the SV3Dp variant, as it supports flexible viewing angles, whereas SV3D-u operates only with fixed viewpoints. We validate our approach using the NeRSemble dataset [32], which includes 222 unique identities recorded from 16 angles. We randomly select a timestamp from one of their video sequences for each individual and extract a random subset of views. All selected frames are then centered, and one of these views is used as input for the tested methods. We evaluate their reconstruction performance using L2 distance, LPIPS [71], SSIM, and an Identity Similarity score. To calculate the Identity Similarity score, we feed both the ground-truth and the reconstructed images into the ArcFace [13] and VGGFace [48] networks and measure the cosine similarity between their final feature vectors.", + "bbox": [ + 88, + 250, + 482, + 535 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1 summarizes the reconstruction performance of each network. SpinMeRound, achieves state-of-the-art results across all cases when evaluated on the LPIPS, SSIM, and Identity Similarity metrics. Additionally, it outperforms all diffusion-based approaches in terms of L2 distance. Although our method trails slightly behind NeRF-based approaches, this is expected given the inherent advantages of NeRF-based models in novel view synthesis tasks. Notably, methods like Eg3D and Panohead require a time-intensive fitting process, which can occasionally fail. In contrast, our proposed approach relies solely on an efficient sampling process, eliminating the need for fitting.", + "bbox": [ + 89, + 537, + 482, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 732, + 254, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Component analysis", + "text_level": 1, + "bbox": [ + 89, + 757, + 282, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conduct ablation studies for the importance of the identity mechanism, the shape normals and the input ground truth image. For this reason, we trained the following three models while using the same training data: a) our proposed architecture, without integrating the input conditioning view, b) a Stable Diffusion-based model without integrating the robust identity mechanism, and c) our model without generating any shape normals.", + "bbox": [ + 89, + 780, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/67e69a158bf1d60d300fb33774c413f690ef3bdbf694b4f4b516e498c45e961b.jpg", + "table_caption": [ + "Table 1. Reconstruction performance on the NeRSample [32] dataset shows that SpinMeRound achieves state-of-the-art results in LPIPS, SSIM and ID Sim metrics while performing on par with the leading models in terms of L2 distance." + ], + "table_footnote": [], + "table_body": "
L2 ↓LPIPS ↓SSIM ↑
SpinMeRound (w/o Input Image)0.12460.42990.568
SpinMeRound (w/o identity embedding)0.0280.260.70
SpinMeRound (w/o Normals)0.0560.320.65
SpinMeRound0.0180.220.75
", + "bbox": [ + 550, + 88, + 870, + 155 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation study: We evaluate the performance of Spin-MeRound alongside three variations, demonstrating that the proposed architecture achieves the highest performance across all reconstruction metrics.", + "bbox": [ + 511, + 166, + 903, + 220 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c0db0fb64d4ed31f3ea06fdbde5a275cb7647a6cbb5f1e3f324595d5dc20910a.jpg", + "image_caption": [ + "Figure 7. Ablation Study: We compare the identity similarity between four models: SpinMeRound (Ours), a similar model without the input image, ones without the identity embedding mechanism, and one without generating shape normals $\\mathcal{N}$ . Results show that our proposed architecture achieves the highest identity similarity." + ], + "image_footnote": [], + "bbox": [ + 542, + 261, + 867, + 421 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Use of input image: Starting from a pre-trained Arc2Face model using the EDM framework, we trained a multi-view model to generate novel viewpoints given only an input identity embedding vector. Notably, SpinMeRound has also been trained to be able to generate input identities without providing the conditioning view. In this case, it generates subjects close to the input face, as illustrated in Figure 8.", + "bbox": [ + 511, + 513, + 905, + 621 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Identity embedding mechanism: In this case, we start from a pre-trained Stable diffusion 1.5 architecture trained using the EDM framework. This model gets trained following the same training parameters, as presented in Sec 3.2. Instead of using the proposed identity conditioning mechanism, we feed an empty string in the conditioning attention layers of the denoising UNet.", + "bbox": [ + 511, + 638, + 905, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Use of Normals: We also train a variant of the network similar to SpinMeRound, but without generating shape normals. This version can only generate novel views without having any additional information about the facial shape.", + "bbox": [ + 511, + 763, + 903, + 825 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Aiming to showcase the importance of each component, we conducted a multi-view reconstruction experiment. More specifically, we sample 100 distinct identities using Panohead [1]. For each subject, we sample 10 different views along with an input frontal view. Given the frontal", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14352", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/734d747f27c7e00de1a27859f50087eff27c210f912c3f1ba4ec94ba0215b88c.jpg", + "image_caption": [ + "Figure 8. From the input images on the left, we present samples generated by our method using only the corresponding identity embedding vectors (top row) and using also the input image (bottom row). The resulting subjects closely resemble the input identities." + ], + "image_footnote": [], + "bbox": [ + 99, + 87, + 903, + 414 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/45efca35cf8c8420dbbea15dfa89e5396ae7df4c536affc1b19d2dd76822125e.jpg", + "image_caption": [ + "Figure 9. Identity interpolation between subject pairs, showcasing our method generates smooth transitions between identities." + ], + "image_footnote": [], + "bbox": [ + 91, + 464, + 480, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "view as input to each separate model, we reconstruct the remaining views. We measure the performance of each model by calculating their discrepancy using L2 distance, LPIPS [71], SSIM and ID Similarity. As presented in Table 2 and Figure 7, our proposed methodology performs better regarding all reconstruction metrics while achieving the highest identity similarity score.", + "bbox": [ + 89, + 628, + 482, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Importance of ID features", + "text_level": 1, + "bbox": [ + 89, + 742, + 328, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Choosing to condition the stable diffusion model in identity embeddings is an important design choice for our network. Those forms of representation contain compact information extracted from a FR model (ArcFace [13]), trained in millions of different images and subjects. Those features allow linear interpolation between the facial characteristics of different subjects. Hence, while using only the identity embedding layer, we interpolate between 2 distinct identities and present their linear interpolation results in Figure 9. It is", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "clearly shown that our model generates smooth transitions between the generated identities.", + "bbox": [ + 511, + 467, + 906, + 498 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 521, + 633, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we presented SpinMeRound, a multi-view latent diffusion model, which generates all-around head portraits of an input subject given a number of input views alongside an input identity embedding. Additionally, we introduced a sampling strategy for generating consistent intermediate views, given an unconstrained input \"in-the-wild\" facial image. Being trained sorely on multi-view synthetic data, we showcase our method's abilities by beating the current state-of-the-art multi-view diffusion models in novel view synthesis experiments.", + "bbox": [ + 511, + 550, + 906, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments: S. Zafeiriou and part of the research was funded by the EPSRC Fellowship DEFORM (EP/S010203/1) and EPSRC Project GNOMON (EP/X011364/1). The authors gratefully acknowledge the scientific support and HPC resources provided by the Erlangen National High Performance Computing Center (NHR@FAU) of FAU under the NHR projects b143dc and b180dc. NHR funding is provided by federal and Bavarian state authorities. NHR@FAU hardware is partially funded by the German Research Foundation (DFG) - 440719683. Additional support was also received by the ERC - project MIA-NORMAL 101083647, DFG 513220538, 512819079, and by the state of Bavaria (HTA).", + "bbox": [ + 511, + 704, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14353", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sizhe An, Hongyi Xu, Yichun Shi, Guoxian Song, Umit Ogras, and Linjie Luo. Panohed: Geometry-aware 3d fullhead synthesis in $360^{\\circ}$ , 2023. 2, 5, 7", + "[2] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[3] Francesca Babiloni, Alexandros Lattas, Jiankang Deng, and Stefanos Zafeiriou. Id-to-3d: Expressive id-guided 3d heads via score distillation sampling. In Advances in Neural Information Processing Systems, pages 97167–97183. Curran Associates, Inc., 2024. 3", + "[4] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In SIGGRAPH '99, 1999. 2", + "[5] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets, 2023. 1, 3, 4", + "[6] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 3", + "[7] James Booth, Anastasios Roussos, Stefanos Zafeiriou, Allan Ponniahy, and David Dunaway. A 3D morphable model learnt from 10,000 faces. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 5543-5552, 2016. 2", + "[8] James Booth, Anastasios Roussos, Allan Ponniah, David Dunaway, and Stefanos Zafeiriou. Large scale 3d morphable models. International Journal of Computer Vision, 126(2): 233-254, 2018. 2", + "[9] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J. Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16123-16133, 2022. 2, 6, 7", + "[10] Xiyi Chen, Marko Mihajlovic, Shaofei Wang, Sergey Prokudin, and Siyu Tang. Morphable diffusion: 3d-consistent diffusion for single-image avatar creation. 2024. 2, 3", + "[11] Zilong Chen, Yikai Wang, Feng Wang, Zhengyi Wang, and Huaping Liu. V3d: Video diffusion models are effective 3d generators. arXiv preprint arXiv:2403.06738, 2024. 3", + "[12] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-xl: A universe of $10\\mathrm{m} + 3\\mathrm{d}$ objects. arXiv preprint arXiv:2307.05663, 2023. 6, 7" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 3, 4, 5, 7, 8", + "[14] Yu Deng, Duomin Wang, Xiaohang Ren, Xingyu Chen, and Baoyuan Wang. Portrait4d: Learning one-shot 4d head avatar synthesis using synthetic data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3", + "[15] Yu Deng, Duomin Wang, and Baoyuan Wang. Portrait4d-v2: Pseudo multi-view data creates better 4d head synthesizer. arXiv preprint arXiv:2403.13570, 2024. 3", + "[16] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In Advances in Neural Information Processing Systems, pages 8780-8794. Curran Associates, Inc., 2021. 1, 2, 3", + "[17] Yao Feng, Haiwen Feng, Michael J. Black, and Timo Bolkart. Learning an animatable detailed 3D face model from in-the-wild images. 2021. 2", + "[18] Stathis Galanakis, Baris Gecer, Alexandros Lattas, and Stefanos Zafeiriou. 3dmm-rf: Convolutional radiance fields for 3d face modeling. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 3536-3547, 2023. 2", + "[19] Stathis Galanakis, Alexandros Lattas, Stylianos Moschoglou, and Stefanos Zafeiriou. Fitdiff: Robust monocular 3d facial shape and reflectance estimation using diffusion models, 2024. 3, 4", + "[20] Ruiqi Gao*, Aleksander Holynski*, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul P. Srinivasan, Jonathan T. Barron, and Ben Poole*. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv, 2024. 2, 3, 4, 5, 6", + "[21] Baris Gecer, Stylianos Ploumpis, Irene Kotsia, and Stefanos Zafeiriou. Ganfit: Generative adversarial network fitting for high fidelity 3d face reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[22] Dimitrios Gerogiannis, Foivos Paraperas Papantoniou, Rolandos Alexandros Potamias, Alexandros Lattas, and Stefanos Zafeiriou. Arc2avatar: Generating expressive 3d avatars from a single image via id guidance. arXiv preprint arXiv:2501.05379, 2025. 3", + "[23] Simon Giebenhain, Tobias Kirschstein, Markos Georgopoulos, Martin Rinz, Lourdes Agapito, and Matthias Nießner. Learning neural parametric head models. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[24] Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. Emu video: Factorizing text-to-video generation by explicit image conditioning, 2024. 3", + "[25] Yuming Gu, Hongyi Xu, You Xie, Guoxian Song, Yichun Shi, Di Chang, Jing Yang, and Linjie Luo. Diffportrait3d: Controllable diffusion for zero-shot portrait view synthe" + ], + "bbox": [ + 93, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14354", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10456-10465, 2024. 2, 3, 6, 7", + "[26] Yuming Gu, Phong Tran, Yujuan Zheng, Hongyi Xu, Heyuan Li, Adilbek Karmanov, and Hao Li. Diffportrait360: Consistent portrait diffusion for 360 view synthesis, 2025. 3", + "[27] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5", + "[28] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239, 2020. 1, 3, 5", + "[29] Yash Kant, Ethan Weber, Jin Kyu Kim, Rawal Khirodkar, Su Zhaoen, Julieta Martinez, Igor Gilitschenski, Shunsuke Saito, and Timur Bagautdinov. Pippo: High-resolution multi-view humans from a single image. 2025. 3", + "[30] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Proc. NeurIPS, 2022. 4, 5", + "[31] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 6", + "[32] Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radiance field reconstruction of human heads. ACM Trans. Graph., 42(4), 2023. 7", + "[33] Yushi Lan, Feitong Tan, Di Qiu, Qiangeng Xu, Kyle Genova, Zeng Huang, Sean Fanello, Rohit Pandey, Thomas Funkhouser, Chen Change Loy, and Yinda Zhang. Gaussian3diff: 3d gaussian diffusion for 3d full head synthesis and editing. In ECCV, 2024. 3", + "[34] Alexandros Lattas, Stylianos Moschoglou, Stylianos Ploumpis, Baris Gecer, Jiankang Deng, and Stefanos Zafeiriou. FitMe: Deep photorealistic 3D morphable model avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[35] Heyuan Li, Ce Chen, Tianhao Shi, Yuda Qiu, Sizhe An, Guanying Chen, and Xiaoguang Han. Spherehead: Stable 3d full-head synthesis with spherical tri-plane representation, 2024. 2", + "[36] Peng Li, Yuan Liu, Xiaoxiao Long, Feihu Zhang, Cheng Lin, Mengfei Li, Xingqun Qi, Shanghang Zhang, Wenhan Luo, Ping Tan, et al. Era3d: High-resolution multiview diffusion using efficient row-wise attention. arXiv preprint arXiv:2405.11616, 2024. 2, 3", + "[37] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017. 2", + "[38] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[39] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object, 2023. 2, 3, 7" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 3", + "[41] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. In SIGGRAPH, pages 163-169. ACM, 1987. 5", + "[42] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, Natalia Neverova, Andrea Vedaldi, Oran Gafni, and Filippos Kokkinos. Im-3d: Iterative multiview diffusion and reconstruction for high-quality 3d generation, 2024. 2, 3", + "[43] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 2", + "[44] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13503–13513, 2022. 2", + "[45] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. StyleSDF: High-Resolution 3D-Consistent Image and Geometry Generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13503–13513, 2022. 2", + "[46] Foivos Paraperas Papantoniou, Alexandros Lattas, Stylianos Moschoglou, and Stefanos Zafeiriou. Relightify: Relightable 3d faces from a single image via diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 3, 4, 5", + "[47] Foivos Paraperas Papantoniou, Alexandros Lattas, Stylianos Moschoglou, Jiankang Deng, Bernhard Kainz, and Stefanos Zafeiriou. Arc2face: A foundation model for id-consistent human faces. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 4", + "[48] Omkar M. Parkhi, Andrea Vedaldi, and Andrew Zisserman. Deep face recognition. In *British Machine Vision Conference*, 2015. 7", + "[49] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 2, 3", + "[50] Malte Prinzler, Egor Zakharov, Vanessa Sklyarova, Berna Kabadayi, and Justus Thies. Joker: Conditional 3d head synthesis with extreme facial expressions, 2024. 3", + "[51] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, and Bernard Ghanem. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. In The Twelfth International Conference on Learning Representations (ICLR), 2024. 3", + "[52] Alec Radford, Jong Wook Kim, Chris Hallacy, A. Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14355", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 4", + "[53] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Ben Mildenhall, Nataniel Ruiz, Shiran Zada, Kfir Aberman, Michael Rubenstein, Jonathan Barron, Yuanzhen Li, and Varun Jampani. Dreambooth3d: Subject-driven text-to-3d generation. ICCV, 2023. 3", + "[54] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 5", + "[55] Daniel Roich, Ron Mokady, Amit H. Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images, 2021. 2", + "[56] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2, 3, 4", + "[57] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. CoRR, abs/1505.04597, 2015. 4", + "[58] Mehdi S. M. Sajjadi, Henning Meyer, Etienne Pot, Urs Bergmann, Klaus Greff, Noha Radwan, Suhani Vora, Mario Lucic, Daniel Duckworth, Alexey Dosovitskiy, Jakob Uszkoreit, Thomas Funkhouser, and Andrea Tagliasacchi. Scene Representation Transformer: Geometry-Free Novel View Synthesis Through Set-Latent Scene Representations. CVPR, 2022. 4", + "[59] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model, 2023. 3", + "[60] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv:2308.16512, 2023. 4", + "[61] Felix Taubner, Ruihang Zhang, Mathieu Tuli, and David B. Lindell. Cap4d: Creating animatable 4d portrait avatars with morphable multi-view diffusion models, 2024. 3", + "[62] Alex Trevithick, Matthew Chan, Michael Stengel, Eric R. Chan, Chao Liu, Zhiding Yu, Sameh Khamis, Manmohan Chandraker, Ravi Ramamoorthi, and Koki Nagano. Real-time radiance fields for single-image portrait view synthesis. In ACM Transactions on Graphics (SIGGRAPH), 2023. 2", + "[63] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion, 2024. 2, 3, 6, 7", + "[64] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 3", + "[65] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, and Baining Guo. Rodin: A generative model for sculpting 3d digital avatars using diffusion, 2022. 3" + ], + "bbox": [ + 91, + 90, + 482, + 897 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[66] Haohan Weng, Tianyu Yang, Jianan Wang, Yu Li, Tong Zhang, C. L. Philip Chen, and Lei Zhang. Consistent123: Improve consistency for one image to 3d object synthesis, 2023. 3", + "[67] Yiqian Wu, Hao Xu, Xiangjun Tang, Hongbo Fu, and Xiao-gang Jin. 3dportraitgan: Learning one-quarter headshot 3d gans from a single-view portrait dataset with diverse body poses, 2023. 2", + "[68] Yuyang Huang Taoran Yi Xiaopeng Zhang Lingxi Xie Xinggang Wang Wenrui Dai Hongkai Xiong and Qi Tian Yabo Chen, Jiemin Fang. Cascade-zero123: One image to highly consistent 3d with self-prompted nearby views. arXiv preprint arXiv:2312.04424, 2023. 3, 7", + "[69] Tarun Yenamandra, Ayush Tewari, Florian Bernard, Hans-Peter Seidel, Mohamed Elgharib, Daniel Cremers, and Christian Theobalt. i3dmm: Deep implicit 3d morphable model of human heads. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12803-12813, 2021. 2", + "[70] Bowen Zhang, Yiji Cheng, Chunyu Wang, Ting Zhang, Jiaolong Yang, Yansong Tang, Feng Zhao, Dong Chen, and Baining Guo. Rodinhd: High-fidelity 3d avatar generation with diffusion models. arXiv preprint arXiv:2407.06938, 2024. 3", + "[71] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6, 7, 8", + "[72] Mingwu Zheng, Hongyu Yang, Di Huang, and Liming Chen. Imface: A nonlinear 3d morphable face model with implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20343-20352, 2022. 2", + "[73] Zhenglin Zhou, Fan Ma, Hehe Fan, Zongxin Yang, and Yi Yang. Headstudio: Text to animatable head avatars with 3d gaussian splatting. 2024. 2, 3", + "[74] Zheng Zhu, Guan Huang, Jiankang Deng, Yun Ye, Junjie Huang, Xinze Chen, Jiagang Zhu, Tian Yang, Jiwen Lu, Dalong Du, and Jie Zhou. Webface260m: A benchmark unveiling the power of million-scale deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10492-10502, 2021. 3" + ], + "bbox": [ + 516, + 92, + 903, + 667 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14356", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SpinMeRound_ Consistent Multi-View Identity Generation Using Diffusion Models/d46ce965-2153-45b4-9566-3f0d4d7d97a6_model.json b/2025/SpinMeRound_ Consistent Multi-View Identity Generation Using Diffusion Models/d46ce965-2153-45b4-9566-3f0d4d7d97a6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..027dab6fc69db068cbfa01e15a7c958fdc4ad9ae --- /dev/null +++ b/2025/SpinMeRound_ Consistent Multi-View Identity Generation Using Diffusion Models/d46ce965-2153-45b4-9566-3f0d4d7d97a6_model.json @@ -0,0 +1,3071 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.12, + 0.13, + 0.877, + 0.175 + ], + "angle": 0, + "content": "SpinMeRound: Consistent Multi-View Identity Generation Using Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.138, + 0.203, + 0.286, + 0.221 + ], + "angle": 0, + "content": "Stathis Galanakis" + }, + { + "type": "text", + "bbox": [ + 0.402, + 0.204, + 0.555, + 0.221 + ], + "angle": 0, + "content": "Alexandros Lattas1" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.221, + 0.465, + 0.238 + ], + "angle": 0, + "content": "Bernhard Kainz\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.672, + 0.204, + 0.858, + 0.222 + ], + "angle": 0, + "content": "Stylianos Moschoglou" + }, + { + "type": "text", + "bbox": [ + 0.598, + 0.222, + 0.753, + 0.238 + ], + "angle": 0, + "content": "Stefanos Zafeiriou" + }, + { + "type": "text", + "bbox": [ + 0.377, + 0.256, + 0.623, + 0.275 + ], + "angle": 0, + "content": "1Imperial College London, UK" + }, + { + "type": "text", + "bbox": [ + 0.352, + 0.275, + 0.641, + 0.293 + ], + "angle": 0, + "content": "\\(^{2}\\)FAU Erlangen-Nürnberg, Germany" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.306, + 0.28, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.476, + 0.154, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.157, + 0.476, + 0.219, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.476, + 0.271, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.308, + 0.49, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.316, + 0.476, + 0.378, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.476, + 0.435, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.442, + 0.476, + 0.505, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.202, + 0.538, + 0.383, + 0.552 + ], + "angle": 0, + "content": "Single Image to novel views" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.308, + 0.907, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.476, + 0.586, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.476, + 0.655, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.476, + 0.716, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.476, + 0.78, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.787, + 0.476, + 0.842, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.844, + 0.476, + 0.905, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.628, + 0.538, + 0.808, + 0.551 + ], + "angle": 0, + "content": "Sparse views to novel views" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.563, + 0.908, + 0.593 + ], + "angle": 0, + "content": "Figure 1. SpinMeRound is a multiview diffusion model which generates human portraits from novel viewpoints. Given a single or multiple views, our method produces high-fidelity images along with precise surface normals, ensuring accurate 3D consistency across perspectives." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.604, + 0.327, + 0.62 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.645, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Despite recent progress in diffusion models, generating realistic head portraits from novel viewpoints remains a significant challenge in computer vision. Most current approaches are constrained to limited angular ranges, predominantly focusing on frontal or near-frontal views. Moreover, although the recent emerging large-scale diffusion models have been proven robust in handling 3D scenes, they underperform on facial data, given their complex structure and the uncanny valley pitfalls. In this paper, we propose SpinMeRound, a diffusion-based approach designed to generate consistent and accurate head portraits from novel viewpoints. By leveraging a number of input views alongside an identity embedding, our method effectively synthesizes diverse viewpoints of a subject whilst robustly maintaining its unique identity features. Through experimentation, we showcase our model's generation capabilities in full head synthesis, while beating current state-of-the" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.606, + 0.907, + 0.636 + ], + "angle": 0, + "content": "art multi-view diffusion models. Project page: spin-meround.github.io." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.664, + 0.646, + 0.68 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.689, + 0.907, + 0.885 + ], + "angle": 0, + "content": "Recent advancements in deep learning have brought significant progress to fundamental computer vision tasks, notably in image and video generation [6, 56]. The introduction of Diffusion Models [16, 28] has transformed these domains by enabling the generation of high-quality visual content, facilitated by the use of large-scale datasets [5, 56]. Despite these developments, the task of generating consistent and accurate head portraits from a single input image still remains a challenging problem. This difficulty is primarily attributed to the limited availability of comprehensive 3D facial datasets, which constrains the training of models capable of reliably capturing and reconstructing the detailed structure and variations of human faces." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "A common practice for modeling human heads from" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14346" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.364 + ], + "angle": 0, + "content": "a single image, incorporates the use of a 3D Morphable Model (3DMM) [4, 7, 8, 37], to represent the facial shape along with an appearance model [17, 21, 34]. However, given typical training data and the difficulty in modeling complex hairstyles, these methods focus solely on the facial region and avoid or miss the full head and hair. The seminal work of Neural Radiance Fields (NeRF) [43] led to an explosion of works on neural rendering of scenes that could not easily be modeled with textured meshes. In addition, pairing such implicit representations with generative models led to a wide variety of approaches [2, 9, 44] that pushed the boundaries on the synthesis of novel facial views, achieving high quality and control. More recently, Panohead [1] first showcased high-quality \\(360^{\\circ}\\) head portrait synthesis. However, because of its adaptive camera training scheme, the back-head synthesis typically contains many artifacts [35], and its inversion on \"in-the-wild\" images is challenging and requires complex fine-tuning [55]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.371, + 0.485, + 0.904 + ], + "angle": 0, + "content": "Recently, diffusion models [16] demonstrated superior performance over GAN-based methods in image generation and have achieved great quality in human generation tasks [56]. However, achieving multi-view consistency remains a significant challenge. Despite the lack of accurate 3D datasets, recent advances, such as Score Distillation techniques [49, 73], represent an initial step toward leveraging the 2D generation strengths of diffusion models to construct 3D content without extra training. Nevertheless, these approaches are computationally intensive, require intricate constraints, and do not consistently yield photorealistic results. Meanwhile, multi-view diffusion architectures, employing video diffusion models as a backbone framework [42, 63], still remain very resource-demanding for generating a single novel view, as they rely on comprehensive camera trajectories to generate coherent central objects. Closer to our work, Zero123 [39] introduced a view-conditioned diffusion model that incorporates view features and camera information into the diffusion process. However, the generated images lack strong multi-view consistency and are of low quality, which restricts their performance in photorealistic 3D generation. Although Diff-Portrait3D [25] enables novel view generation through a diffusion process, it still focuses only on near-front views and utilizes a image-conditioned camera mechanism. Other closely related works are Era3D [36] and Morphable Diffusion [10], which generate fixed camera viewpoints, thus limiting their ability to produce full head portraits. The recently proposed Cat3D [20] presents a promising solution by efficiently integrating a number of input views with specified camera poses to achieve a consistent novel viewpoint generation. It integrates 3D attention layers for efficiently sharing common information between all views, along with robust camera pose feature maps. Still, Cat3D is limited by its lack of focus on human generation and is currently" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.747, + 0.107 + ], + "angle": 0, + "content": "unavailable as an open-source tool." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.109, + 0.905, + 0.395 + ], + "angle": 0, + "content": "In this paper, we present SpinMeRound, a multi-view diffusion model designed to generate novel high-fidelity views of a given human face. In addition to facial images, our model also generates the corresponding normals, which are typically available for human data and, as we show, improve the model's performance and consistency on intricate facial features. Moreover, we show that conditioning the model on an identity embedding and one or more input views of a subject during inference, we can not only sample but also reconstruct multi-view consistent images from \"in-the-wild\" facial images. Our method can accurately synthesize photorealistic head portraits from various angles while preserving essential identity characteristics, which can be used to represent or reconstruct 3D scenes. Given the lack of open-source large-scale multi-view head datasets and the problematic nature of the permissions of such datasets, our method is solely trained on synthetic data acquired using Panohead [1], making this work accessible to experiment and build with. Overall, in this paper:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.397, + 0.905, + 0.456 + ], + "angle": 0, + "content": "- We introduce SpinMeRound, a multi-view diffusion model conditioned on identity embeddings and a number of views that generates novel perspectives of an input subject and their respective normals." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.457, + 0.905, + 0.501 + ], + "angle": 0, + "content": "- We present a novel sampling strategy which, given a single \"in-the-wild\" facial image, generates consistent views encompassing the whole head." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.502, + 0.905, + 0.547 + ], + "angle": 0, + "content": "- We explore its potential by comparing it with current state-of-the-art multi-view diffusion-based methods and showcase superior results in full-head portrait generation." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.397, + 0.905, + 0.547 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.565, + 0.657, + 0.58 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.591, + 0.664, + 0.607 + ], + "angle": 0, + "content": "2.1. Face Modeling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Extensive research has been dedicated to representing 3D faces through a combination of texture maps and 3D meshes, starting with fundamental 3D Morphable Models (3DMM) [4, 7, 8, 37]. These approaches, however, primarily focus on accurately capturing only the frontal region of the head whilst lacking in integrating finer head details such as hair, wrinkles and wearable items. Dealing with this, recent studies have integrated implicit representations for 3D face modeling. Methods such as [23, 45, 72] integrate Signed Distance Functions (SDFs) whereas other [2, 9, 18, 69] use Neural Radiance Field (NeRF) models [43] for generating photorealistic results. Levering the powerful triplane representation [9], RTRF [62] generates nearfrontal views in real-time whilst Panohead [1] introduces an adaptive camera training strategy for enabling full head portrait synthesis. An extension of it is 3DPortraitGAN [67] focusing on all-around upper body generation. As this work focuses only on full-head generation, Panohead can effectively be used to acquire synthetic portrait datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14347" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.09, + 0.895, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.312, + 0.908, + 0.37 + ], + "angle": 0, + "content": "Figure 2. Overview of SpinMeRound: Starting with a number of input conditioning views, the identity embedding \\(\\mathbf{W}\\) is extracted via a Face Recognition network (ArcFace [13]). Both the conditioning and target views are then encoded and combined with corresponding ray coordinate maps that represent camera poses. After the sampling step, our method synthesizes photorealistic images from novel angles, along with their associated shape normals \\(\\mathcal{N}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.381, + 0.262, + 0.397 + ], + "angle": 0, + "content": "2.2. Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.405, + 0.486, + 0.664 + ], + "angle": 0, + "content": "Face Generation Diffusion models (DMs) [28] have proven their generative abilities by beating the well-established GANs in image synthesis tasks [16]. The availability of large-scale datasets has revolutionized text-to-image generation [56] and video generation [5, 6] tasks. For face synthesis, a range of methods have emerged to tackle essential tasks, including 3D avatar creation [65, 70], avatar reenactment [14, 15, 33] and texture map generation [19, 46]. We adopt the concatenation strategy outlined in [19, 46] to generate novel viewpoint images and their corresponding shape normals simultaneously. Based on Stable Diffusion [56], Arc2Face [13] and InstantID [64] generate facial images given an input subject. Notably, Arc2Face exhibits strong generalization abilities in facial image generation, leveraging an up-sampled subset of WebFace42M [74]. Also, it introduces a robust identity conditioning mechanism, which is integrated into our method." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.674, + 0.487, + 0.903 + ], + "angle": 0, + "content": "Novel View Image Synthesis DreamFusion [49] introduces the use of Score Distillation Sampling (SDS), incorporating a pre-trained text-to-image diffusion model [56] alongside a NeRF model, to synthesize 3D objects from text prompts. In this way, its authors proved that they can efficiently generate 3D objects despite the lack of large-scale datasets; by exploiting the generalization abilities of an image generation network. Although subsequent studies focus on better distillation strategies [51, 53, 73], the aforementioned approaches are time-consuming and require complex balancing and additional constraints [20]. Closer to our work, ID-to-3D [3] and Arc2Avatar [22] use score distillation to generate 3D facial geometry, lacking though photorealistic results. Zero123 [39] introduces a multiview diffusion model conditioned on a reference image" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.382, + 0.909, + 0.732 + ], + "angle": 0, + "content": "and a camera pose. Following it, methods such as One-2-3-45 [38], SyncDreamer [40], Consistent1-to-3 [66] and Cascade-Zero123 [68] further focus on multi-view consistency by introducing priors during the denoising process. Studies such as Zero \\(123++\\) [59], Era3D [36] and Morphable Diffusion [10] generate fixed viewpoints given the input image, without being able to handle arbitrary views. In Cat3D [20], a general multi-view diffusion model is introduced, using a powerful camera pose conditioning mechanism while using 3D attention layers. Although it has been a robust method for novel view synthesis, it does not focus on facial novel view synthesis, lacks shape normal generation capabilities and is a closed source framework. Diff-Portrait3D [25] showcases novel view capabilities, given an input facial image. Unlike our method, which employs image-free camera conditioning, DiffPortrait3D employs an image-driven approach for the desired output viewpoints and can generate only near-front views. More recent works [50, 61] have explored diffusion-based frameworks for generating controllable and animatable avatars. Concurrently to our approach, Pippo [29] and DiffPortrait360 [26] demonstrated \\(360^{\\circ}\\) avatar generation, without, though, being able to sample novel identities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Novel View Video Synthesis Recently, video generation models [5, 6, 24] have proven their ability to generate photorealistic models. Methods such as IM-3D [42], V3D [11] and SV3D [63] integrate an off-the-shelf video diffusion model for generating novel viewpoints of a reference object. Still, the video generation step makes these methodologies computationally demanding [20]. Additionally, they are often restricted by the requirement for specific camera trajectories, typically revolving around a central subject. In contrast, our approach focuses on implementing multi-view diffusion networks that handle unordered camera poses." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14348" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.087, + 0.907, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.232, + 0.907, + 0.262 + ], + "angle": 0, + "content": "Figure 3. Samples generated by SpinMeRound on \"in-the-wild\" images: Given only the input images (shown in a small box at the top center), our method produces high-fidelity images from novel angles, along with corresponding shape normals \\(\\mathcal{N}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.274, + 0.182, + 0.289 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.3, + 0.484, + 0.406 + ], + "angle": 0, + "content": "SpinMeRound incorporates a multi-view diffusion model, trained on a 3D facial dataset, capable of generating a full head portrait of an input subject, given their identity embedding and a number of views (Fig. 2). In the following sections, we first describe the proposed diffusion model, its training scheme, and the proposed sampling strategy given a single input view." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.42, + 0.345, + 0.435 + ], + "angle": 0, + "content": "3.1. Multi-View Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.443, + 0.484, + 0.76 + ], + "angle": 0, + "content": "SpinMeRound employs a latent multi-view diffusion model alongside a powerful identity conditioning mechanism [47], enabling the generation of novel views of an input scene containing a person. Our approach leverages as conditioning inputs \\( M \\in \\{1,3\\} \\) pairs of sparse views of a subject, their respective shape normals and their associated camera poses. Let \\( \\mathbf{I}_i \\) denote a picked conditioned view and \\( \\bar{\\mathbf{I}}_i \\) a cropped and aligned version of \\( \\mathbf{I}_i \\). We utilize a pre-trained face recognition network (ArcFace [13]) \\( \\phi \\) to extract the identity embedding vector \\( \\mathbf{w} = \\phi(\\bar{\\mathbf{I}}_i) \\in \\mathbb{R}^{512} \\) capable of incorporating crucial identity features. Then, we inject the identity information into the diffusion model as proposed in Arc2Face [13]. In particular, a text prompt of \"a photo of \\( \\) person, is fed to a CLIP text encoder [52] followed by the tokenization step. Followed by padding, the identity embedding vector \\( \\mathbf{w} \\) replaces the corresponding \\( \\) token resulting in \\( \\mathbf{s} = \\{e_1,e_2,e_3,\\hat{\\mathbf{w}},e_5\\} \\), where \\( \\hat{\\mathbf{w}} \\in \\mathbb{R}^{768} \\) is the padded version of \\( \\mathbf{w} \\). The vector \\( \\mathbf{s} \\) is fed to the encoder \\( C \\), and finally we retrieve the corresponding conditioning vector \\( \\mathbf{c} = C(\\mathbf{s}) \\in \\mathbb{R}^{N \\times 768} \\), where \\( N \\) is the maximum sequence length." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.762, + 0.484, + 0.903 + ], + "angle": 0, + "content": "The proposed latent multi-view UNet [57] is similar to the one introduced in Cat3D [20]. Our diffusion model is designed to concurrently generate \\( P = (M + K) = 8 \\) pairs of facial images and their corresponding normals \\( \\mathcal{N} \\), given an input conditioning vector \\( \\mathbf{w} \\), a number of \\( M \\) views and the \\( P \\) camera poses. Let \\( I_{i}^{cond} \\) represent the input human face images, \\( I_{j}^{tgt} \\) the target face images from novel viewpoints, and \\( \\mathcal{N}_{i}^{cond}, \\mathcal{N}_{j}^{tgt} \\) their respective shape normals, where \\( i \\in \\{1,2,\\dots,M\\} \\), \\( j \\in \\{1,2,\\dots,N\\} \\) with" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.275, + 0.907, + 0.603 + ], + "angle": 0, + "content": "\\(M\\) being the maximum number of the input conditional views and \\(N\\) the number of target viewpoints. Following [19, 20, 46], we employ a pre-trained AutoEncoder of Stable Diffusion 1.5 (SD 1.5) [56], consisting of an encoder \\(\\mathcal{E}\\) and a decoder \\(\\mathcal{D}\\). For each input conditioning image, we first obtain the corresponding latent feature maps: \\(\\mathbf{z}_{I_i^{cond}} = \\mathcal{E}(I_i^{cond}) \\in \\mathbb{R}^{4 \\times 64 \\times 64}\\) and \\(\\mathbf{z}_{\\mathcal{N}_i^{cond}} = \\mathcal{E}(\\mathcal{N}_i^{cond}) \\in \\mathbb{R}^{4 \\times 64 \\times 64}\\). These are then concatenated channel-wise. Similarly, we apply the same procedure to extract the latent vectors for the target viewpoints. The corresponding camera pose information is incorporated for both conditioning and target viewpoints using the mechanism proposed in Cat3D [20]. For each of the \\(P\\) views, the latent feature maps \\(\\mathbf{z}_i\\) are concatenated channel-wise with the respective ray representation maps [20, 58] \\(\\mathbf{r}_i^{cond}, \\mathbf{r}_j^{tgt} \\in \\mathbb{R}^{149 \\times 64 \\times 64}\\), which encode the ray origin and direction. A binary mask \\(\\mathbf{m} \\in \\{\\mathbf{0}, \\mathbf{1}\\}^{1 \\times 64 \\times 64}\\) is then appended to differentiate between conditional and target latent vectors. Finally, we retrieve the conditioning and target latent feature maps \\(\\bar{\\mathbf{z}}_i^{cond} = \\{\\mathbf{z}_{I_i^{cond}}, \\mathbf{z}_{\\mathcal{N}_i^{cond}}\\mathbf{r}_j^{cond}, \\mathbf{m}_j^{cond}\\}\\) and \\(\\bar{\\mathbf{z}}_i^{tgt} = \\{\\mathbf{z}_{I_i^{tgt}}, \\mathbf{z}_{\\mathcal{N}_i^{tgt}}\\mathbf{r}_j^{tgt}, \\mathbf{m}_j^{tgt}\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.604, + 0.909, + 0.727 + ], + "angle": 0, + "content": "As proposed in video diffusion models [5], we initialize our method using a pre-trained LDM model (Arc2Face [47]) trained on large-scale datasets while adding additional attention layers, connecting the multiple latent feature maps. As in [20], we integrate 3D attention layers [60] by adding them between the original 2D self-attention layers of the LDM, and we fine-tune all layers of the multi-view UNet for improved multi-view consistency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.74, + 0.673, + 0.757 + ], + "angle": 0, + "content": "3.2. Training Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We begin from the publicly available Arc2Face [47], and we adhere to the training schemes introduced in [5, 20]. Specifically, we adapt the EDM framework [30] to Arc2Face, training it for 31,000 iterations on its dedicated dataset. To accommodate the input latent feature maps, we expand the input and output convolutional layers channels, initializing these by copying the existing weights to the shape-normal channels while randomly initializing the camera-pose dimensions. As in [20], we shift the log-to-signal ratio by" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14349" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.364 + ], + "angle": 0, + "content": "\\(\\log (\\mathrm{N})\\) , where \\(\\mathbf{N}\\) is the number of the target images \\((\\mathrm{N} = 7)\\) We randomly select a conditioning view that includes part of the frontal face during each training iteration, as required for the identity embedding extraction. We then randomly pick the N target images and calculate the relative camera angles. To enhance the dataset, we replace the white background with a random color, in \\(50\\%\\) of the samples. All the training viewpoints are encoded using the encoder \\(\\mathcal{E}\\) ,with noise added only to the target latent vectors, while conditioning vectors remain unchanged. Following the classifier-free guidance (CFG) training scheme [27], with a probability of \\(\\mathcal{P}_{uncond} = 0.15\\) , we randomly replace the identity vector with the empty string, and the conditioning images with zero-ed ones. We first train the model conditioned on a single view for \\(600\\mathrm{k}\\) iterations. Then, for the additional 1M iterations, we vary the conditioning views by randomly choosing O, 1 or 3 conditioning views, corresponding to 8, 7 and 5 target views, each with a probability of \\(\\mathcal{P}_=1 / 3\\)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.373, + 0.286, + 0.389 + ], + "angle": 0, + "content": "3.3. Novel view sampling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.395, + 0.483, + 0.577 + ], + "angle": 0, + "content": "SpinMeRound synthesizes novel views for an input subject I, given a number of views. In this section, we introduce a robust sampling strategy designed to produce a large number of consistent views that comprehensively cover a full head, given only a single input image. Achieving consistency across viewpoints requires a carefully structured camera pose selection order. Therefore, we employ a three-step sampling process: a) aligning input views and extracting the corresponding shape normals, b) generating anchor images that provide complete coverage of the \\(360^{\\circ}\\) human head and c) synthesizing intermediate views by leveraging both the input views and the closest anchor images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.596, + 0.483, + 0.852 + ], + "angle": 0, + "content": "Alignment and Shape Normals generation Given the input image \\(\\mathbf{I}\\), we extract its identity embedding \\(\\mathbf{w}\\) as described in Sec. 3.1. Then, we obtain \\(\\hat{\\mathbf{I}}\\), a cropped and aligned version of \\(\\mathbf{I}\\), using the alignment procedure presented in Panohead [1]. To generate the shape normals \\(\\mathcal{N}\\), we treat this as an in-painting task, thus retrieving the normals through the conditional guidance sampling approach described in Relightify [46]. Specifically, using the aligned image \\(\\hat{\\mathbf{I}}\\) and its identity embedding vector \\(\\mathbf{w}\\), we employ a binary visibility mask \\(m\\), marking only the image channels as visible and setting the shape normal channels as nonvisible. By applying the \"channel-wise in-painting\" algorithm, the corresponding shape normals are retrieved. This process uses the EDM sampler presented in [30] alongside the DDPM [28] discretization steps and runs for 50 steps. A more detailed presentation of this approach is presented in the supplemental material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Generating anchor and intermediate views SpinMeRound can generate any arbitrary viewpoint, given the input" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.905, + 0.41 + ], + "angle": 0, + "content": "aligned facial image \\(\\hat{\\mathbf{I}}\\) and the corresponding shape normals \\(\\mathcal{N}\\). However, since it was trained to generate only a limited number of views per sampling process, a carefully designed sampling strategy is essential to produce a wide range of output views. Since the target subject is centered in the scene, we first generate \\(M = 7\\) anchor images \\(\\mathbf{A}_i\\) and corresponding anchor shape normals \\(\\mathbf{A}_{\\mathcal{N}_i}\\), \\(i \\in \\{1, \\dots, M\\}\\), covering a \\(360^\\circ\\) angle range of the subject \\((\\pm 45^\\circ, \\pm 90^\\circ, \\pm 135^\\circ, 180^\\circ)\\), as proposed in [20]. Using these anchors, any number of intermediate views can then be synthesized by conditioning on image triplets \\(\\{(\\mathbf{I}, \\mathcal{N}), (\\mathbf{A}_k, \\mathbf{A}_{\\mathcal{N}_k}), (\\mathbf{A}_l, \\mathbf{A}_{\\mathcal{N}_l})\\}\\), where \\(k, l\\) where represent the closest anchor images. This approach ensures that each intermediate view remains consistent with both the input aligned image \\(\\hat{\\mathbf{I}}\\), and the closest already generated views. For both sampling processes, we use the EDM sampler facilitated by the EDM discretization steps [30], while it runs for 50 steps with a guidance scale set to 3. This method enables the generation of 48, 88, or more novel views for a single input subject \\(\\mathbf{I}\\), depending on the chosen angle step, thus covering the entire scene." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.424, + 0.646, + 0.44 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.449, + 0.681, + 0.465 + ], + "angle": 0, + "content": "4.1. Training Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.471, + 0.905, + 0.819 + ], + "angle": 0, + "content": "Training such a model requires a large multi-view dataset containing a great number of subjects \\( S \\). For each person \\( S_{i} \\), it is necessary to acquire a set of images \\( \\mathbf{I}_k^i \\), their corresponding shape normal maps \\( \\mathcal{N}_k^i \\), camera poses \\( \\mathcal{C}_k^i \\) alongside with their corresponding identity embedding vector \\( \\mathbf{w}^i \\), where \\( k \\in \\{1, 2, \\dots, N_i\\} \\) and \\( N_{i} \\) is the number of the available views for the \\( i \\)-th scene. Due to the lack of such a dataset, we create a large-scale synthetic dataset using the publicly available Panohead [1]. We first sample \\( \\sim 10k \\) subjects and manually remove instances with artifacts in the back of the head, resulting in \\( \\sim 7k \\) distinct identities. We render images from 125 different viewpoints for each person to cover the entire head. Simultaneously, we obtain each subject's facial shape by extracting their opacity values from the triplane feature maps and then applying the marching cubes [41] algorithm. We render their respective normal maps using the acquired facial shape using Pytorch3d [54]. All the images depicting a frontal head are fed to the ArcFace [13] network to extract their corresponding identity vectors \\( \\mathbf{w} \\). All in all, we end up with a synthetic dataset containing about 7k individuals, rendered from \\( N = 125 \\) different angles, the respective shape normals, camera poses, and their identity embedding vector \\( \\mathbf{w} \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.829, + 0.742, + 0.845 + ], + "angle": 0, + "content": "4.2. Qualitative Comparisons" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.851, + 0.8, + 0.867 + ], + "angle": 0, + "content": "4.2.1. Novel View Synthesis comparisons" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.9 + ], + "angle": 0, + "content": "In this section, we present a qualitative comparison between our model and other state-of-the-art multi-view dif" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14350" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.086, + 0.16, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.14, + 0.155, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.166, + 0.088, + 0.218, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.166, + 0.14, + 0.219, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.166, + 0.189, + 0.219, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.238, + 0.155, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.301, + 0.157, + 0.312 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.166, + 0.14, + 0.219, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.166, + 0.189, + 0.219, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.166, + 0.238, + 0.219, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.088, + 0.274, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.14, + 0.275, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.189, + 0.275, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.238, + 0.275, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.301, + 0.261, + 0.312 + ], + "angle": 0, + "content": "(b) SpinM" + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.088, + 0.336, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.14, + 0.336, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.189, + 0.336, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.238, + 0.335, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.088, + 0.401, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.35, + 0.14, + 0.402, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.189, + 0.397, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.238, + 0.404, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.367, + 0.287, + 0.383, + 0.294 + ], + "angle": 0, + "content": "+45°" + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.088, + 0.467, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.14, + 0.467, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.189, + 0.467, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.238, + 0.467, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.417, + 0.301, + 0.484, + 0.312 + ], + "angle": 0, + "content": "(c) SV3D [63]" + }, + { + "type": "image", + "bbox": [ + 0.473, + 0.088, + 0.525, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.472, + 0.14, + 0.526, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.483, + 0.19, + 0.523, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.477, + 0.239, + 0.525, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.093, + 0.585, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.143, + 0.583, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.192, + 0.581, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.242, + 0.587, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.093, + 0.655, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.144, + 0.646, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.192, + 0.645, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.242, + 0.655, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.58, + 0.301, + 0.678, + 0.312 + ], + "angle": 0, + "content": "(d) Zero123-XL [12]" + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.093, + 0.705, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.663, + 0.141, + 0.712, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.669, + 0.192, + 0.706, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.667, + 0.238, + 0.705, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.67, + 0.288, + 0.696, + 0.294 + ], + "angle": 0, + "content": "-45°" + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.088, + 0.842, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.14, + 0.842, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.189, + 0.842, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.238, + 0.842, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.754, + 0.301, + 0.859, + 0.312 + ], + "angle": 0, + "content": "(d) DiffPortrait3d [25]" + }, + { + "type": "image", + "bbox": [ + 0.842, + 0.088, + 0.905, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.841, + 0.14, + 0.905, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.842, + 0.189, + 0.905, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.842, + 0.238, + 0.902, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.322, + 0.907, + 0.366 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison between SpinMeRound, Zero123-XL [12], SV3D [63] and DiffPortrait3D [25] at angles \\(\\{\\pm 45^{\\circ}, + 180^{\\circ}\\}\\). It is clear that SpinMeRound effectively generates high-quality novel views from \"in-the-wild\" input images, whilst SV3D produces distorted outputs, Zero123-XL generates unnaturally squared avatars and DiffPortrait3D cannot handle such angles." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.379, + 0.485, + 0.635 + ], + "angle": 0, + "content": "fusion models, SV3D [63] Zero123-XL [12] and DiffPort- trait3D [25], focusing on their output under \\(\\pm 45^{\\circ}, + 180^{\\circ}\\) angles. SV3D, a latent video diffusion model, creates \\(360^{\\circ}\\) videos from a single input image. Zero123-XL, a multiview diffusion model, synthesizes novel perspectives based on the input image and specified camera poses. DiffPort- trait3D [25] generates novel views by conditioning the desired camera on an input image and refining the initial noise through a fine-tuning step to achieve optimal results. Figure 4 demonstrates scenarios where the input \"in-the-wild\" images (Fig. 4a ) are fed to these 4 different networks and their respective generations under \\(\\pm 45^{\\circ}, + 180^{\\circ}\\) viewpoints. As illustrated, SpinMeRound efficiently generates accurate novel viewpoints of the input subject, while SV3D distorts the input image in its generated views, Zero123-XL produces unnatural, square-like facial avatars and DiffPort- trait3D cannot handle those angles at all." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.645, + 0.484, + 0.73 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.739, + 0.483, + 0.782 + ], + "angle": 0, + "content": "Figure 5. Given the input identities (shown in the small boxes), we present the results of novel view synthesis after applying 3D Gaussian splatting [31] to the views generated by our model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.803, + 0.271, + 0.818 + ], + "angle": 0, + "content": "4.3. 3D Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "To evaluate our model's consistency in generating novel views from an input image, we assess its ability to reconstruct an input identity through Gaussian Splatting (3DGS) [31]. More specifically, given an input identity, we apply the sampling strategy outlined in Section 3.3 to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.379, + 0.907, + 0.515 + ], + "angle": 0, + "content": "generate 48 novel views. These views are then used to reconstruct the identity through Gaussian Splatting. As in [20], we modify the provided 3DGS code to incorporate the LPIPS [71] loss between the ground-truth and the generated viewpoints. In this way, we can deal with small inconsistencies between nearby viewpoints. We present the novel view synthesis of two input identities (small squares) in Figure 5, which clearly illustrates that our proposed methodology can generate consistent subjects." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.527, + 0.614, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.605, + 0.611, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.528, + 0.663, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.569, + 0.663, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.605, + 0.663, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.663, + 0.528, + 0.712, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.663, + 0.569, + 0.712, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.663, + 0.605, + 0.712, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.528, + 0.842, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.568, + 0.842, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.605, + 0.842, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.842, + 0.528, + 0.902, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.842, + 0.569, + 0.902, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.842, + 0.605, + 0.902, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.654, + 0.906, + 0.682 + ], + "angle": 0, + "content": "Figure 6. Samples generated using unconditional sampling. Spin-MeRound can generate novel identities without any prior input." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.709, + 0.738, + 0.726 + ], + "angle": 0, + "content": "4.4. Unconditional Sampling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.731, + 0.907, + 0.837 + ], + "angle": 0, + "content": "SpinMeRound is trained following the Classifier-Free Guidance training scheme. Thus, our architecture can generate novel multi-view identities without any prior conditioning. By setting the input identity embedding equal to the empty string, our model can generate a novel identity along with different viewpoints of that identity. Figure 6 presents some examples of those identities." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.753, + 0.865 + ], + "angle": 0, + "content": "4.5.Quantitative Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We compare our method's ability to generate novel views with the current state-of-the-art models: a) Eg3D [9] and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14351" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.482, + 0.166 + ], + "angle": 0, + "content": "
MethodL2 ↓LPIPS↓SSIM↑ID Sim [13] ↑ID Sim [48] ↑
NeRFEg3d [9]0.0250.40.550.310.89
basedPanohead [1]0.0120.320.650.270.88
Zero123 [39]0.1950.5150.550.1690.44
Diffusion basedZero123-XL [12]0.1980.510.5630.1180.442
SV3D [63]0.0870.410.6600.360.881
DiffPortrait3d [25]0.10.50.350.550.887
SpinMeRound (Ours)0.0330.30.730.610.911
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.178, + 0.483, + 0.234 + ], + "angle": 0, + "content": "Table 1. Reconstruction performance on the NeRSample [32] dataset shows that SpinMeRound achieves state-of-the-art results in LPIPS, SSIM and ID Sim metrics while performing on par with the leading models in terms of L2 distance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.25, + 0.483, + 0.536 + ], + "angle": 0, + "content": "Panohead [1] which are NeRF-based approaches generating frontal and full-head portraits respectively, b) Zero123 [68], Zero123-XL [12] and DiffPortrai3D [25], which are multiview diffusion models and c) SV3D [63], a multi-view video diffusion model. For SV3D, we selected the SV3Dp variant, as it supports flexible viewing angles, whereas SV3D-u operates only with fixed viewpoints. We validate our approach using the NeRSemble dataset [32], which includes 222 unique identities recorded from 16 angles. We randomly select a timestamp from one of their video sequences for each individual and extract a random subset of views. All selected frames are then centered, and one of these views is used as input for the tested methods. We evaluate their reconstruction performance using L2 distance, LPIPS [71], SSIM, and an Identity Similarity score. To calculate the Identity Similarity score, we feed both the ground-truth and the reconstructed images into the ArcFace [13] and VGGFace [48] networks and measure the cosine similarity between their final feature vectors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.483, + 0.719 + ], + "angle": 0, + "content": "Table 1 summarizes the reconstruction performance of each network. SpinMeRound, achieves state-of-the-art results across all cases when evaluated on the LPIPS, SSIM, and Identity Similarity metrics. Additionally, it outperforms all diffusion-based approaches in terms of L2 distance. Although our method trails slightly behind NeRF-based approaches, this is expected given the inherent advantages of NeRF-based models in novel view synthesis tasks. Notably, methods like Eg3D and Panohead require a time-intensive fitting process, which can occasionally fail. In contrast, our proposed approach relies solely on an efficient sampling process, eliminating the need for fitting." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.733, + 0.255, + 0.749 + ], + "angle": 0, + "content": "5. Ablation Studies" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.283, + 0.774 + ], + "angle": 0, + "content": "5.1. Component analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this section, we conduct ablation studies for the importance of the identity mechanism, the shape normals and the input ground truth image. For this reason, we trained the following three models while using the same training data: a) our proposed architecture, without integrating the input conditioning view, b) a Stable Diffusion-based model without integrating the robust identity mechanism, and c) our model without generating any shape normals." + }, + { + "type": "table", + "bbox": [ + 0.551, + 0.089, + 0.871, + 0.156 + ], + "angle": 0, + "content": "
L2 ↓LPIPS ↓SSIM ↑
SpinMeRound (w/o Input Image)0.12460.42990.568
SpinMeRound (w/o identity embedding)0.0280.260.70
SpinMeRound (w/o Normals)0.0560.320.65
SpinMeRound0.0180.220.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.167, + 0.905, + 0.222 + ], + "angle": 0, + "content": "Table 2. Ablation study: We evaluate the performance of Spin-MeRound alongside three variations, demonstrating that the proposed architecture achieves the highest performance across all reconstruction metrics." + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.262, + 0.868, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.436, + 0.905, + 0.507 + ], + "angle": 0, + "content": "Figure 7. Ablation Study: We compare the identity similarity between four models: SpinMeRound (Ours), a similar model without the input image, ones without the identity embedding mechanism, and one without generating shape normals \\(\\mathcal{N}\\). Results show that our proposed architecture achieves the highest identity similarity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.515, + 0.906, + 0.622 + ], + "angle": 0, + "content": "Use of input image: Starting from a pre-trained Arc2Face model using the EDM framework, we trained a multi-view model to generate novel viewpoints given only an input identity embedding vector. Notably, SpinMeRound has also been trained to be able to generate input identities without providing the conditioning view. In this case, it generates subjects close to the input face, as illustrated in Figure 8." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.906, + 0.746 + ], + "angle": 0, + "content": "Identity embedding mechanism: In this case, we start from a pre-trained Stable diffusion 1.5 architecture trained using the EDM framework. This model gets trained following the same training parameters, as presented in Sec 3.2. Instead of using the proposed identity conditioning mechanism, we feed an empty string in the conditioning attention layers of the denoising UNet." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.905, + 0.826 + ], + "angle": 0, + "content": "Use of Normals: We also train a variant of the network similar to SpinMeRound, but without generating shape normals. This version can only generate novel views without having any additional information about the facial shape." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Aiming to showcase the importance of each component, we conducted a multi-view reconstruction experiment. More specifically, we sample 100 distinct identities using Panohead [1]. For each subject, we sample 10 different views along with an input frontal view. Given the frontal" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14352" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.088, + 0.905, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.424, + 0.908, + 0.455 + ], + "angle": 0, + "content": "Figure 8. From the input images on the left, we present samples generated by our method using only the corresponding identity embedding vectors (top row) and using also the input image (bottom row). The resulting subjects closely resemble the input identities." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.465, + 0.482, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.591, + 0.483, + 0.62 + ], + "angle": 0, + "content": "Figure 9. Identity interpolation between subject pairs, showcasing our method generates smooth transitions between identities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.734 + ], + "angle": 0, + "content": "view as input to each separate model, we reconstruct the remaining views. We measure the performance of each model by calculating their discrepancy using L2 distance, LPIPS [71], SSIM and ID Similarity. As presented in Table 2 and Figure 7, our proposed methodology performs better regarding all reconstruction metrics while achieving the highest identity similarity score." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.329, + 0.759 + ], + "angle": 0, + "content": "5.2. Importance of ID features" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Choosing to condition the stable diffusion model in identity embeddings is an important design choice for our network. Those forms of representation contain compact information extracted from a FR model (ArcFace [13]), trained in millions of different images and subjects. Those features allow linear interpolation between the facial characteristics of different subjects. Hence, while using only the identity embedding layer, we interpolate between 2 distinct identities and present their linear interpolation results in Figure 9. It is" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.468, + 0.907, + 0.499 + ], + "angle": 0, + "content": "clearly shown that our model generates smooth transitions between the generated identities." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.522, + 0.634, + 0.537 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.907, + 0.702 + ], + "angle": 0, + "content": "In this paper, we presented SpinMeRound, a multi-view latent diffusion model, which generates all-around head portraits of an input subject given a number of input views alongside an input identity embedding. Additionally, we introduced a sampling strategy for generating consistent intermediate views, given an unconstrained input \"in-the-wild\" facial image. Being trained sorely on multi-view synthetic data, we showcase our method's abilities by beating the current state-of-the-art multi-view diffusion models in novel view synthesis experiments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments: S. Zafeiriou and part of the research was funded by the EPSRC Fellowship DEFORM (EP/S010203/1) and EPSRC Project GNOMON (EP/X011364/1). The authors gratefully acknowledge the scientific support and HPC resources provided by the Erlangen National High Performance Computing Center (NHR@FAU) of FAU under the NHR projects b143dc and b180dc. NHR funding is provided by federal and Bavarian state authorities. NHR@FAU hardware is partially funded by the German Research Foundation (DFG) - 440719683. Additional support was also received by the ERC - project MIA-NORMAL 101083647, DFG 513220538, 512819079, and by the state of Bavaria (HTA)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14353" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.157 + ], + "angle": 0, + "content": "[1] Sizhe An, Hongyi Xu, Yichun Shi, Guoxian Song, Umit Ogras, and Linjie Luo. Panohed: Geometry-aware 3d fullhead synthesis in \\(360^{\\circ}\\), 2023. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.484, + 0.214 + ], + "angle": 0, + "content": "[2] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.217, + 0.483, + 0.285 + ], + "angle": 0, + "content": "[3] Francesca Babiloni, Alexandros Lattas, Jiankang Deng, and Stefanos Zafeiriou. Id-to-3d: Expressive id-guided 3d heads via score distillation sampling. In Advances in Neural Information Processing Systems, pages 97167–97183. Curran Associates, Inc., 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.288, + 0.483, + 0.316 + ], + "angle": 0, + "content": "[4] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In SIGGRAPH '99, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.318, + 0.483, + 0.387 + ], + "angle": 0, + "content": "[5] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets, 2023. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.389, + 0.483, + 0.459 + ], + "angle": 0, + "content": "[6] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.461, + 0.483, + 0.531 + ], + "angle": 0, + "content": "[7] James Booth, Anastasios Roussos, Stefanos Zafeiriou, Allan Ponniahy, and David Dunaway. A 3D morphable model learnt from 10,000 faces. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 5543-5552, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.532, + 0.483, + 0.587 + ], + "angle": 0, + "content": "[8] James Booth, Anastasios Roussos, Allan Ponniah, David Dunaway, and Stefanos Zafeiriou. Large scale 3d morphable models. International Journal of Computer Vision, 126(2): 233-254, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.59, + 0.483, + 0.699 + ], + "angle": 0, + "content": "[9] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J. Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16123-16133, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.703, + 0.483, + 0.758 + ], + "angle": 0, + "content": "[10] Xiyi Chen, Marko Mihajlovic, Shaofei Wang, Sergey Prokudin, and Siyu Tang. Morphable diffusion: 3d-consistent diffusion for single-image avatar creation. 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[11] Zilong Chen, Yikai Wang, Feng Wang, Zhengyi Wang, and Huaping Liu. V3d: Video diffusion models are effective 3d generators. arXiv preprint arXiv:2403.06738, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[12] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-xl: A universe of \\(10\\mathrm{m} + 3\\mathrm{d}\\) objects. arXiv preprint arXiv:2307.05663, 2023. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[13] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[14] Yu Deng, Duomin Wang, Xiaohang Ren, Xingyu Chen, and Baoyuan Wang. Portrait4d: Learning one-shot 4d head avatar synthesis using synthetic data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.234, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[15] Yu Deng, Duomin Wang, and Baoyuan Wang. Portrait4d-v2: Pseudo multi-view data creates better 4d head synthesizer. arXiv preprint arXiv:2403.13570, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.278, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[16] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In Advances in Neural Information Processing Systems, pages 8780-8794. Curran Associates, Inc., 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.376 + ], + "angle": 0, + "content": "[17] Yao Feng, Haiwen Feng, Michael J. Black, and Timo Bolkart. Learning an animatable detailed 3D face model from in-the-wild images. 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.378, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[18] Stathis Galanakis, Baris Gecer, Alexandros Lattas, and Stefanos Zafeiriou. 3dmm-rf: Convolutional radiance fields for 3d face modeling. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 3536-3547, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.449, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[19] Stathis Galanakis, Alexandros Lattas, Stylianos Moschoglou, and Stefanos Zafeiriou. Fitdiff: Robust monocular 3d facial shape and reflectance estimation using diffusion models, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[20] Ruiqi Gao*, Aleksander Holynski*, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul P. Srinivasan, Jonathan T. Barron, and Ben Poole*. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv, 2024. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.576, + 0.905, + 0.646 + ], + "angle": 0, + "content": "[21] Baris Gecer, Stylianos Ploumpis, Irene Kotsia, and Stefanos Zafeiriou. Ganfit: Generative adversarial network fitting for high fidelity 3d face reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.647, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[22] Dimitrios Gerogiannis, Foivos Paraperas Papantoniou, Rolandos Alexandros Potamias, Alexandros Lattas, and Stefanos Zafeiriou. Arc2avatar: Generating expressive 3d avatars from a single image via id guidance. arXiv preprint arXiv:2501.05379, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[23] Simon Giebenhain, Tobias Kirschstein, Markos Georgopoulos, Martin Rinz, Lourdes Agapito, and Matthias Nießner. Learning neural parametric head models. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.788, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[24] Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. Emu video: Factorizing text-to-video generation by explicit image conditioning, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[25] Yuming Gu, Hongyi Xu, You Xie, Guoxian Song, Yichun Shi, Di Chang, Jing Yang, and Linjie Luo. Diffportrait3d: Controllable diffusion for zero-shot portrait view synthe" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14354" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "sis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10456-10465, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[26] Yuming Gu, Phong Tran, Yujuan Zheng, Hongyi Xu, Heyuan Li, Adilbek Karmanov, and Hao Li. Diffportrait360: Consistent portrait diffusion for 360 view synthesis, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.179, + 0.482, + 0.207 + ], + "angle": 0, + "content": "[27] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.208, + 0.483, + 0.248 + ], + "angle": 0, + "content": "[28] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239, 2020. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.251, + 0.483, + 0.306 + ], + "angle": 0, + "content": "[29] Yash Kant, Ethan Weber, Jin Kyu Kim, Rawal Khirodkar, Su Zhaoen, Julieta Martinez, Igor Gilitschenski, Shunsuke Saito, and Timur Bagautdinov. Pippo: High-resolution multi-view humans from a single image. 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.308, + 0.482, + 0.348 + ], + "angle": 0, + "content": "[30] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Proc. NeurIPS, 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.35, + 0.482, + 0.405 + ], + "angle": 0, + "content": "[31] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.407, + 0.482, + 0.462 + ], + "angle": 0, + "content": "[32] Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radiance field reconstruction of human heads. ACM Trans. Graph., 42(4), 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.464, + 0.482, + 0.533 + ], + "angle": 0, + "content": "[33] Yushi Lan, Feitong Tan, Di Qiu, Qiangeng Xu, Kyle Genova, Zeng Huang, Sean Fanello, Rohit Pandey, Thomas Funkhouser, Chen Change Loy, and Yinda Zhang. Gaussian3diff: 3d gaussian diffusion for 3d full head synthesis and editing. In ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.535, + 0.482, + 0.604 + ], + "angle": 0, + "content": "[34] Alexandros Lattas, Stylianos Moschoglou, Stylianos Ploumpis, Baris Gecer, Jiankang Deng, and Stefanos Zafeiriou. FitMe: Deep photorealistic 3D morphable model avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.605, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[35] Heyuan Li, Ce Chen, Tianhao Shi, Yuda Qiu, Sizhe An, Guanying Chen, and Xiaoguang Han. Spherehead: Stable 3d full-head synthesis with spherical tri-plane representation, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[36] Peng Li, Yuan Liu, Xiaoxiao Long, Feihu Zhang, Cheng Lin, Mengfei Li, Xingqun Qi, Shanghang Zhang, Wenhan Luo, Ping Tan, et al. Era3d: High-resolution multiview diffusion using efficient row-wise attention. arXiv preprint arXiv:2405.11616, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[37] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[38] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[39] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object, 2023. 2, 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[40] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[41] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. In SIGGRAPH, pages 163-169. ACM, 1987. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[42] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, Natalia Neverova, Andrea Vedaldi, Oran Gafni, and Filippos Kokkinos. Im-3d: Iterative multiview diffusion and reconstruction for high-quality 3d generation, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[43] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[44] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13503–13513, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[45] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. StyleSDF: High-Resolution 3D-Consistent Image and Geometry Generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13503–13513, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[46] Foivos Paraperas Papantoniou, Alexandros Lattas, Stylianos Moschoglou, and Stefanos Zafeiriou. Relightify: Relightable 3d faces from a single image via diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[47] Foivos Paraperas Papantoniou, Alexandros Lattas, Stylianos Moschoglou, Jiankang Deng, Bernhard Kainz, and Stefanos Zafeiriou. Arc2face: A foundation model for id-consistent human faces. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.672 + ], + "angle": 0, + "content": "[48] Omkar M. Parkhi, Andrea Vedaldi, and Andrew Zisserman. Deep face recognition. In *British Machine Vision Conference*, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[49] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[50] Malte Prinzler, Egor Zakharov, Vanessa Sklyarova, Berna Kabadayi, and Justus Thies. Joker: Conditional 3d head synthesis with extreme facial expressions, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[51] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, and Bernard Ghanem. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. In The Twelfth International Conference on Learning Representations (ICLR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[52] Alec Radford, Jong Wook Kim, Chris Hallacy, A. Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14355" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.121, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[53] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Ben Mildenhall, Nataniel Ruiz, Shiran Zada, Kfir Aberman, Michael Rubenstein, Jonathan Barron, Yuanzhen Li, and Varun Jampani. Dreambooth3d: Subject-driven text-to-3d generation. ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.19, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[54] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.482, + 0.287 + ], + "angle": 0, + "content": "[55] Daniel Roich, Ron Mokady, Amit H. Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.483, + 0.357 + ], + "angle": 0, + "content": "[56] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.359, + 0.482, + 0.398 + ], + "angle": 0, + "content": "[57] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. CoRR, abs/1505.04597, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.495 + ], + "angle": 0, + "content": "[58] Mehdi S. M. Sajjadi, Henning Meyer, Etienne Pot, Urs Bergmann, Klaus Greff, Noha Radwan, Suhani Vora, Mario Lucic, Daniel Duckworth, Alexey Dosovitskiy, Jakob Uszkoreit, Thomas Funkhouser, and Andrea Tagliasacchi. Scene Representation Transformer: Geometry-Free Novel View Synthesis Through Set-Latent Scene Representations. CVPR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.497, + 0.482, + 0.551 + ], + "angle": 0, + "content": "[59] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.553, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[60] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv:2308.16512, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.595, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[61] Felix Taubner, Ruihang Zhang, Mathieu Tuli, and David B. Lindell. Cap4d: Creating animatable 4d portrait avatars with morphable multi-view diffusion models, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.705 + ], + "angle": 0, + "content": "[62] Alex Trevithick, Matthew Chan, Michael Stengel, Eric R. Chan, Chao Liu, Zhiding Yu, Sameh Khamis, Manmohan Chandraker, Ravi Ramamoorthi, and Koki Nagano. Real-time radiance fields for single-image portrait view synthesis. In ACM Transactions on Graphics (SIGGRAPH), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.774 + ], + "angle": 0, + "content": "[63] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[64] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.482, + 0.898 + ], + "angle": 0, + "content": "[65] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, and Baining Guo. Rodin: A generative model for sculpting 3d digital avatars using diffusion, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.146 + ], + "angle": 0, + "content": "[66] Haohan Weng, Tianyu Yang, Jianan Wang, Yu Li, Tong Zhang, C. L. Philip Chen, and Lei Zhang. Consistent123: Improve consistency for one image to 3d object synthesis, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[67] Yiqian Wu, Hao Xu, Xiangjun Tang, Hongbo Fu, and Xiao-gang Jin. 3dportraitgan: Learning one-quarter headshot 3d gans from a single-view portrait dataset with diverse body poses, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.206, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[68] Yuyang Huang Taoran Yi Xiaopeng Zhang Lingxi Xie Xinggang Wang Wenrui Dai Hongkai Xiong and Qi Tian Yabo Chen, Jiemin Fang. Cascade-zero123: One image to highly consistent 3d with self-prompted nearby views. arXiv preprint arXiv:2312.04424, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[69] Tarun Yenamandra, Ayush Tewari, Florian Bernard, Hans-Peter Seidel, Mohamed Elgharib, Daniel Cremers, and Christian Theobalt. i3dmm: Deep implicit 3d morphable model of human heads. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12803-12813, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.904, + 0.416 + ], + "angle": 0, + "content": "[70] Bowen Zhang, Yiji Cheng, Chunyu Wang, Ting Zhang, Jiaolong Yang, Yansong Tang, Feng Zhao, Dong Chen, and Baining Guo. Rodinhd: High-fidelity 3d avatar generation with diffusion models. arXiv preprint arXiv:2407.06938, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[71] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.529 + ], + "angle": 0, + "content": "[72] Mingwu Zheng, Hongyu Yang, Di Huang, and Liming Chen. Imface: A nonlinear 3d morphable face model with implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20343-20352, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.531, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[73] Zhenglin Zhou, Fan Ma, Hehe Fan, Zongxin Yang, and Yi Yang. Headstudio: Text to animatable head avatars with 3d gaussian splatting. 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.574, + 0.905, + 0.668 + ], + "angle": 0, + "content": "[74] Zheng Zhu, Guan Huang, Jiankang Deng, Yun Ye, Junjie Huang, Xinze Chen, Jiagang Zhu, Tian Yang, Jiwen Lu, Dalong Du, and Jie Zhou. Webface260m: A benchmark unveiling the power of million-scale deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10492-10502, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14356" + } + ] +] \ No newline at end of file diff --git a/2025/SpinMeRound_ Consistent Multi-View Identity Generation Using Diffusion Models/d46ce965-2153-45b4-9566-3f0d4d7d97a6_origin.pdf b/2025/SpinMeRound_ Consistent Multi-View Identity Generation Using Diffusion Models/d46ce965-2153-45b4-9566-3f0d4d7d97a6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c4cd54ac53becf0a66a81e3ec55bb2312cacc7b5 --- /dev/null +++ b/2025/SpinMeRound_ Consistent Multi-View Identity Generation Using Diffusion Models/d46ce965-2153-45b4-9566-3f0d4d7d97a6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db87cfa99adab710866ecaf6891c735cdc17567d38bc89799d813aa82e5efe10 +size 19809237 diff --git a/2025/SpinMeRound_ Consistent Multi-View Identity Generation Using Diffusion Models/full.md b/2025/SpinMeRound_ Consistent Multi-View Identity Generation Using Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dcd4e8fc8c64a49f760edecb5de27ced76a0d73c --- /dev/null +++ b/2025/SpinMeRound_ Consistent Multi-View Identity Generation Using Diffusion Models/full.md @@ -0,0 +1,415 @@ +# SpinMeRound: Consistent Multi-View Identity Generation Using Diffusion Models + +Stathis Galanakis + +Alexandros Lattas1 + +Bernhard Kainz $^{1,2}$ + +Stylianos Moschoglou + +Stefanos Zafeiriou + +1Imperial College London, UK + +$^{2}$ FAU Erlangen-Nürnberg, Germany + +![](images/bb20f158f29f6968ffdbb355f3b07c740b4a6bb02d4d46669aed8360dc346502.jpg) + +![](images/0836e9de2f71a06f81d6eb5edef1c613f5e346d0dd77134120eb276784dd2460.jpg) +Figure 1. SpinMeRound is a multiview diffusion model which generates human portraits from novel viewpoints. Given a single or multiple views, our method produces high-fidelity images along with precise surface normals, ensuring accurate 3D consistency across perspectives. + +![](images/ed72d1ea1c1271ea1991e5e982f2596ff60ecac855921160904fd561a1c4f517.jpg) +Single Image to novel views + +![](images/9cef7845c47785dbb3a477bdd4ac232891b704856525474066cb00a26fb9295e.jpg) + +![](images/9797b68e5950dbf682bd41df45d303ba6b4949027585fccffd657e24592ed411.jpg) + +![](images/530921576dfff87d0e844bc7f6510e30a22930cc7e80ec03ff47847db3878c73.jpg) + +![](images/32c7851a65ada9ffdcda4b4351029d787bc8fbc58f94fb16213535dee3664c52.jpg) + +![](images/2b9105ea6dd7efc73c974d913f9068354708c6cf68e0a105dbf3c1f14121679d.jpg) + +![](images/f87676e2202ef8f80c41a804654444a0e024fcacaa3220784c9a46bd8b30ece7.jpg) + +![](images/48dd5b39773110129c5c02a2b84a8f47c3522ad67e436d4a2eb2af4c0c269975.jpg) +Sparse views to novel views + +![](images/7b096e34531b1e53e1168df01186eb32f2e623a78855597d0f2e71a9412c9c2c.jpg) + +![](images/7d06defaff480f16169d3b7f57c2242016f01629842184614866f355640c7317.jpg) + +![](images/a9a96543f3813a273438823097800181f7ffd6a7314e882770b70fa24392ad3e.jpg) + +![](images/b449fcc0465bae1bad11a9ad3e3dd469d9ba6428df6cd4fa0da94ff5d8e89b7a.jpg) + +![](images/ad00c0d328cb3b22eea6be8f8e2073676ec03c2bd144379052d0e55c141a8282.jpg) + +# Abstract + +Despite recent progress in diffusion models, generating realistic head portraits from novel viewpoints remains a significant challenge in computer vision. Most current approaches are constrained to limited angular ranges, predominantly focusing on frontal or near-frontal views. Moreover, although the recent emerging large-scale diffusion models have been proven robust in handling 3D scenes, they underperform on facial data, given their complex structure and the uncanny valley pitfalls. In this paper, we propose SpinMeRound, a diffusion-based approach designed to generate consistent and accurate head portraits from novel viewpoints. By leveraging a number of input views alongside an identity embedding, our method effectively synthesizes diverse viewpoints of a subject whilst robustly maintaining its unique identity features. Through experimentation, we showcase our model's generation capabilities in full head synthesis, while beating current state-of-the + +art multi-view diffusion models. Project page: spin-meround.github.io. + +# 1. Introduction + +Recent advancements in deep learning have brought significant progress to fundamental computer vision tasks, notably in image and video generation [6, 56]. The introduction of Diffusion Models [16, 28] has transformed these domains by enabling the generation of high-quality visual content, facilitated by the use of large-scale datasets [5, 56]. Despite these developments, the task of generating consistent and accurate head portraits from a single input image still remains a challenging problem. This difficulty is primarily attributed to the limited availability of comprehensive 3D facial datasets, which constrains the training of models capable of reliably capturing and reconstructing the detailed structure and variations of human faces. + +A common practice for modeling human heads from + +a single image, incorporates the use of a 3D Morphable Model (3DMM) [4, 7, 8, 37], to represent the facial shape along with an appearance model [17, 21, 34]. However, given typical training data and the difficulty in modeling complex hairstyles, these methods focus solely on the facial region and avoid or miss the full head and hair. The seminal work of Neural Radiance Fields (NeRF) [43] led to an explosion of works on neural rendering of scenes that could not easily be modeled with textured meshes. In addition, pairing such implicit representations with generative models led to a wide variety of approaches [2, 9, 44] that pushed the boundaries on the synthesis of novel facial views, achieving high quality and control. More recently, Panohead [1] first showcased high-quality $360^{\circ}$ head portrait synthesis. However, because of its adaptive camera training scheme, the back-head synthesis typically contains many artifacts [35], and its inversion on "in-the-wild" images is challenging and requires complex fine-tuning [55]. + +Recently, diffusion models [16] demonstrated superior performance over GAN-based methods in image generation and have achieved great quality in human generation tasks [56]. However, achieving multi-view consistency remains a significant challenge. Despite the lack of accurate 3D datasets, recent advances, such as Score Distillation techniques [49, 73], represent an initial step toward leveraging the 2D generation strengths of diffusion models to construct 3D content without extra training. Nevertheless, these approaches are computationally intensive, require intricate constraints, and do not consistently yield photorealistic results. Meanwhile, multi-view diffusion architectures, employing video diffusion models as a backbone framework [42, 63], still remain very resource-demanding for generating a single novel view, as they rely on comprehensive camera trajectories to generate coherent central objects. Closer to our work, Zero123 [39] introduced a view-conditioned diffusion model that incorporates view features and camera information into the diffusion process. However, the generated images lack strong multi-view consistency and are of low quality, which restricts their performance in photorealistic 3D generation. Although Diff-Portrait3D [25] enables novel view generation through a diffusion process, it still focuses only on near-front views and utilizes a image-conditioned camera mechanism. Other closely related works are Era3D [36] and Morphable Diffusion [10], which generate fixed camera viewpoints, thus limiting their ability to produce full head portraits. The recently proposed Cat3D [20] presents a promising solution by efficiently integrating a number of input views with specified camera poses to achieve a consistent novel viewpoint generation. It integrates 3D attention layers for efficiently sharing common information between all views, along with robust camera pose feature maps. Still, Cat3D is limited by its lack of focus on human generation and is currently + +unavailable as an open-source tool. + +In this paper, we present SpinMeRound, a multi-view diffusion model designed to generate novel high-fidelity views of a given human face. In addition to facial images, our model also generates the corresponding normals, which are typically available for human data and, as we show, improve the model's performance and consistency on intricate facial features. Moreover, we show that conditioning the model on an identity embedding and one or more input views of a subject during inference, we can not only sample but also reconstruct multi-view consistent images from "in-the-wild" facial images. Our method can accurately synthesize photorealistic head portraits from various angles while preserving essential identity characteristics, which can be used to represent or reconstruct 3D scenes. Given the lack of open-source large-scale multi-view head datasets and the problematic nature of the permissions of such datasets, our method is solely trained on synthetic data acquired using Panohead [1], making this work accessible to experiment and build with. Overall, in this paper: + +- We introduce SpinMeRound, a multi-view diffusion model conditioned on identity embeddings and a number of views that generates novel perspectives of an input subject and their respective normals. +- We present a novel sampling strategy which, given a single "in-the-wild" facial image, generates consistent views encompassing the whole head. +- We explore its potential by comparing it with current state-of-the-art multi-view diffusion-based methods and showcase superior results in full-head portrait generation. + +# 2. Related Work + +# 2.1. Face Modeling + +Extensive research has been dedicated to representing 3D faces through a combination of texture maps and 3D meshes, starting with fundamental 3D Morphable Models (3DMM) [4, 7, 8, 37]. These approaches, however, primarily focus on accurately capturing only the frontal region of the head whilst lacking in integrating finer head details such as hair, wrinkles and wearable items. Dealing with this, recent studies have integrated implicit representations for 3D face modeling. Methods such as [23, 45, 72] integrate Signed Distance Functions (SDFs) whereas other [2, 9, 18, 69] use Neural Radiance Field (NeRF) models [43] for generating photorealistic results. Levering the powerful triplane representation [9], RTRF [62] generates nearfrontal views in real-time whilst Panohead [1] introduces an adaptive camera training strategy for enabling full head portrait synthesis. An extension of it is 3DPortraitGAN [67] focusing on all-around upper body generation. As this work focuses only on full-head generation, Panohead can effectively be used to acquire synthetic portrait datasets. + +![](images/0a9dc190670b992009c40dc82a72a6923cf4261a17c3c3dfe53def1d75fb96fd.jpg) +Figure 2. Overview of SpinMeRound: Starting with a number of input conditioning views, the identity embedding $\mathbf{W}$ is extracted via a Face Recognition network (ArcFace [13]). Both the conditioning and target views are then encoded and combined with corresponding ray coordinate maps that represent camera poses. After the sampling step, our method synthesizes photorealistic images from novel angles, along with their associated shape normals $\mathcal{N}$ . + +# 2.2. Diffusion Models + +Face Generation Diffusion models (DMs) [28] have proven their generative abilities by beating the well-established GANs in image synthesis tasks [16]. The availability of large-scale datasets has revolutionized text-to-image generation [56] and video generation [5, 6] tasks. For face synthesis, a range of methods have emerged to tackle essential tasks, including 3D avatar creation [65, 70], avatar reenactment [14, 15, 33] and texture map generation [19, 46]. We adopt the concatenation strategy outlined in [19, 46] to generate novel viewpoint images and their corresponding shape normals simultaneously. Based on Stable Diffusion [56], Arc2Face [13] and InstantID [64] generate facial images given an input subject. Notably, Arc2Face exhibits strong generalization abilities in facial image generation, leveraging an up-sampled subset of WebFace42M [74]. Also, it introduces a robust identity conditioning mechanism, which is integrated into our method. + +Novel View Image Synthesis DreamFusion [49] introduces the use of Score Distillation Sampling (SDS), incorporating a pre-trained text-to-image diffusion model [56] alongside a NeRF model, to synthesize 3D objects from text prompts. In this way, its authors proved that they can efficiently generate 3D objects despite the lack of large-scale datasets; by exploiting the generalization abilities of an image generation network. Although subsequent studies focus on better distillation strategies [51, 53, 73], the aforementioned approaches are time-consuming and require complex balancing and additional constraints [20]. Closer to our work, ID-to-3D [3] and Arc2Avatar [22] use score distillation to generate 3D facial geometry, lacking though photorealistic results. Zero123 [39] introduces a multiview diffusion model conditioned on a reference image + +and a camera pose. Following it, methods such as One-2-3-45 [38], SyncDreamer [40], Consistent1-to-3 [66] and Cascade-Zero123 [68] further focus on multi-view consistency by introducing priors during the denoising process. Studies such as Zero $123++$ [59], Era3D [36] and Morphable Diffusion [10] generate fixed viewpoints given the input image, without being able to handle arbitrary views. In Cat3D [20], a general multi-view diffusion model is introduced, using a powerful camera pose conditioning mechanism while using 3D attention layers. Although it has been a robust method for novel view synthesis, it does not focus on facial novel view synthesis, lacks shape normal generation capabilities and is a closed source framework. Diff-Portrait3D [25] showcases novel view capabilities, given an input facial image. Unlike our method, which employs image-free camera conditioning, DiffPortrait3D employs an image-driven approach for the desired output viewpoints and can generate only near-front views. More recent works [50, 61] have explored diffusion-based frameworks for generating controllable and animatable avatars. Concurrently to our approach, Pippo [29] and DiffPortrait360 [26] demonstrated $360^{\circ}$ avatar generation, without, though, being able to sample novel identities. + +Novel View Video Synthesis Recently, video generation models [5, 6, 24] have proven their ability to generate photorealistic models. Methods such as IM-3D [42], V3D [11] and SV3D [63] integrate an off-the-shelf video diffusion model for generating novel viewpoints of a reference object. Still, the video generation step makes these methodologies computationally demanding [20]. Additionally, they are often restricted by the requirement for specific camera trajectories, typically revolving around a central subject. In contrast, our approach focuses on implementing multi-view diffusion networks that handle unordered camera poses. + +![](images/a6f86be2864c23ba016d0af10dc59e82a18bd67199a85d0ba5504c46c741d04b.jpg) +Figure 3. Samples generated by SpinMeRound on "in-the-wild" images: Given only the input images (shown in a small box at the top center), our method produces high-fidelity images from novel angles, along with corresponding shape normals $\mathcal{N}$ . + +# 3. Method + +SpinMeRound incorporates a multi-view diffusion model, trained on a 3D facial dataset, capable of generating a full head portrait of an input subject, given their identity embedding and a number of views (Fig. 2). In the following sections, we first describe the proposed diffusion model, its training scheme, and the proposed sampling strategy given a single input view. + +# 3.1. Multi-View Diffusion Model + +SpinMeRound employs a latent multi-view diffusion model alongside a powerful identity conditioning mechanism [47], enabling the generation of novel views of an input scene containing a person. Our approach leverages as conditioning inputs $M \in \{1,3\}$ pairs of sparse views of a subject, their respective shape normals and their associated camera poses. Let $\mathbf{I}_i$ denote a picked conditioned view and $\bar{\mathbf{I}}_i$ a cropped and aligned version of $\mathbf{I}_i$ . We utilize a pre-trained face recognition network (ArcFace [13]) $\phi$ to extract the identity embedding vector $\mathbf{w} = \phi(\bar{\mathbf{I}}_i) \in \mathbb{R}^{512}$ capable of incorporating crucial identity features. Then, we inject the identity information into the diffusion model as proposed in Arc2Face [13]. In particular, a text prompt of "a photo of $$ person, is fed to a CLIP text encoder [52] followed by the tokenization step. Followed by padding, the identity embedding vector $\mathbf{w}$ replaces the corresponding $$ token resulting in $\mathbf{s} = \{e_1,e_2,e_3,\hat{\mathbf{w}},e_5\}$ , where $\hat{\mathbf{w}} \in \mathbb{R}^{768}$ is the padded version of $\mathbf{w}$ . The vector $\mathbf{s}$ is fed to the encoder $C$ , and finally we retrieve the corresponding conditioning vector $\mathbf{c} = C(\mathbf{s}) \in \mathbb{R}^{N \times 768}$ , where $N$ is the maximum sequence length. + +The proposed latent multi-view UNet [57] is similar to the one introduced in Cat3D [20]. Our diffusion model is designed to concurrently generate $P = (M + K) = 8$ pairs of facial images and their corresponding normals $\mathcal{N}$ , given an input conditioning vector $\mathbf{w}$ , a number of $M$ views and the $P$ camera poses. Let $I_{i}^{cond}$ represent the input human face images, $I_{j}^{tgt}$ the target face images from novel viewpoints, and $\mathcal{N}_{i}^{cond}, \mathcal{N}_{j}^{tgt}$ their respective shape normals, where $i \in \{1,2,\dots,M\}$ , $j \in \{1,2,\dots,N\}$ with + +$M$ being the maximum number of the input conditional views and $N$ the number of target viewpoints. Following [19, 20, 46], we employ a pre-trained AutoEncoder of Stable Diffusion 1.5 (SD 1.5) [56], consisting of an encoder $\mathcal{E}$ and a decoder $\mathcal{D}$ . For each input conditioning image, we first obtain the corresponding latent feature maps: $\mathbf{z}_{I_i^{cond}} = \mathcal{E}(I_i^{cond}) \in \mathbb{R}^{4 \times 64 \times 64}$ and $\mathbf{z}_{\mathcal{N}_i^{cond}} = \mathcal{E}(\mathcal{N}_i^{cond}) \in \mathbb{R}^{4 \times 64 \times 64}$ . These are then concatenated channel-wise. Similarly, we apply the same procedure to extract the latent vectors for the target viewpoints. The corresponding camera pose information is incorporated for both conditioning and target viewpoints using the mechanism proposed in Cat3D [20]. For each of the $P$ views, the latent feature maps $\mathbf{z}_i$ are concatenated channel-wise with the respective ray representation maps [20, 58] $\mathbf{r}_i^{cond}, \mathbf{r}_j^{tgt} \in \mathbb{R}^{149 \times 64 \times 64}$ , which encode the ray origin and direction. A binary mask $\mathbf{m} \in \{\mathbf{0}, \mathbf{1}\}^{1 \times 64 \times 64}$ is then appended to differentiate between conditional and target latent vectors. Finally, we retrieve the conditioning and target latent feature maps $\bar{\mathbf{z}}_i^{cond} = \{\mathbf{z}_{I_i^{cond}}, \mathbf{z}_{\mathcal{N}_i^{cond}}\mathbf{r}_j^{cond}, \mathbf{m}_j^{cond}\}$ and $\bar{\mathbf{z}}_i^{tgt} = \{\mathbf{z}_{I_i^{tgt}}, \mathbf{z}_{\mathcal{N}_i^{tgt}}\mathbf{r}_j^{tgt}, \mathbf{m}_j^{tgt}\}$ . + +As proposed in video diffusion models [5], we initialize our method using a pre-trained LDM model (Arc2Face [47]) trained on large-scale datasets while adding additional attention layers, connecting the multiple latent feature maps. As in [20], we integrate 3D attention layers [60] by adding them between the original 2D self-attention layers of the LDM, and we fine-tune all layers of the multi-view UNet for improved multi-view consistency. + +# 3.2. Training Details + +We begin from the publicly available Arc2Face [47], and we adhere to the training schemes introduced in [5, 20]. Specifically, we adapt the EDM framework [30] to Arc2Face, training it for 31,000 iterations on its dedicated dataset. To accommodate the input latent feature maps, we expand the input and output convolutional layers channels, initializing these by copying the existing weights to the shape-normal channels while randomly initializing the camera-pose dimensions. As in [20], we shift the log-to-signal ratio by + +$\log (\mathrm{N})$ , where $\mathbf{N}$ is the number of the target images $(\mathrm{N} = 7)$ We randomly select a conditioning view that includes part of the frontal face during each training iteration, as required for the identity embedding extraction. We then randomly pick the N target images and calculate the relative camera angles. To enhance the dataset, we replace the white background with a random color, in $50\%$ of the samples. All the training viewpoints are encoded using the encoder $\mathcal{E}$ ,with noise added only to the target latent vectors, while conditioning vectors remain unchanged. Following the classifier-free guidance (CFG) training scheme [27], with a probability of $\mathcal{P}_{uncond} = 0.15$ , we randomly replace the identity vector with the empty string, and the conditioning images with zero-ed ones. We first train the model conditioned on a single view for $600\mathrm{k}$ iterations. Then, for the additional 1M iterations, we vary the conditioning views by randomly choosing O, 1 or 3 conditioning views, corresponding to 8, 7 and 5 target views, each with a probability of $\mathcal{P}_=1 / 3$ + +# 3.3. Novel view sampling + +SpinMeRound synthesizes novel views for an input subject I, given a number of views. In this section, we introduce a robust sampling strategy designed to produce a large number of consistent views that comprehensively cover a full head, given only a single input image. Achieving consistency across viewpoints requires a carefully structured camera pose selection order. Therefore, we employ a three-step sampling process: a) aligning input views and extracting the corresponding shape normals, b) generating anchor images that provide complete coverage of the $360^{\circ}$ human head and c) synthesizing intermediate views by leveraging both the input views and the closest anchor images. + +Alignment and Shape Normals generation Given the input image $\mathbf{I}$ , we extract its identity embedding $\mathbf{w}$ as described in Sec. 3.1. Then, we obtain $\hat{\mathbf{I}}$ , a cropped and aligned version of $\mathbf{I}$ , using the alignment procedure presented in Panohead [1]. To generate the shape normals $\mathcal{N}$ , we treat this as an in-painting task, thus retrieving the normals through the conditional guidance sampling approach described in Relightify [46]. Specifically, using the aligned image $\hat{\mathbf{I}}$ and its identity embedding vector $\mathbf{w}$ , we employ a binary visibility mask $m$ , marking only the image channels as visible and setting the shape normal channels as nonvisible. By applying the "channel-wise in-painting" algorithm, the corresponding shape normals are retrieved. This process uses the EDM sampler presented in [30] alongside the DDPM [28] discretization steps and runs for 50 steps. A more detailed presentation of this approach is presented in the supplemental material. + +Generating anchor and intermediate views SpinMeRound can generate any arbitrary viewpoint, given the input + +aligned facial image $\hat{\mathbf{I}}$ and the corresponding shape normals $\mathcal{N}$ . However, since it was trained to generate only a limited number of views per sampling process, a carefully designed sampling strategy is essential to produce a wide range of output views. Since the target subject is centered in the scene, we first generate $M = 7$ anchor images $\mathbf{A}_i$ and corresponding anchor shape normals $\mathbf{A}_{\mathcal{N}_i}$ , $i \in \{1, \dots, M\}$ , covering a $360^\circ$ angle range of the subject $(\pm 45^\circ, \pm 90^\circ, \pm 135^\circ, 180^\circ)$ , as proposed in [20]. Using these anchors, any number of intermediate views can then be synthesized by conditioning on image triplets $\{(\mathbf{I}, \mathcal{N}), (\mathbf{A}_k, \mathbf{A}_{\mathcal{N}_k}), (\mathbf{A}_l, \mathbf{A}_{\mathcal{N}_l})\}$ , where $k, l$ where represent the closest anchor images. This approach ensures that each intermediate view remains consistent with both the input aligned image $\hat{\mathbf{I}}$ , and the closest already generated views. For both sampling processes, we use the EDM sampler facilitated by the EDM discretization steps [30], while it runs for 50 steps with a guidance scale set to 3. This method enables the generation of 48, 88, or more novel views for a single input subject $\mathbf{I}$ , depending on the chosen angle step, thus covering the entire scene. + +# 4. Experiments + +# 4.1. Training Dataset + +Training such a model requires a large multi-view dataset containing a great number of subjects $S$ . For each person $S_{i}$ , it is necessary to acquire a set of images $\mathbf{I}_k^i$ , their corresponding shape normal maps $\mathcal{N}_k^i$ , camera poses $\mathcal{C}_k^i$ alongside with their corresponding identity embedding vector $\mathbf{w}^i$ , where $k \in \{1, 2, \dots, N_i\}$ and $N_{i}$ is the number of the available views for the $i$ -th scene. Due to the lack of such a dataset, we create a large-scale synthetic dataset using the publicly available Panohead [1]. We first sample $\sim 10k$ subjects and manually remove instances with artifacts in the back of the head, resulting in $\sim 7k$ distinct identities. We render images from 125 different viewpoints for each person to cover the entire head. Simultaneously, we obtain each subject's facial shape by extracting their opacity values from the triplane feature maps and then applying the marching cubes [41] algorithm. We render their respective normal maps using the acquired facial shape using Pytorch3d [54]. All the images depicting a frontal head are fed to the ArcFace [13] network to extract their corresponding identity vectors $\mathbf{w}$ . All in all, we end up with a synthetic dataset containing about 7k individuals, rendered from $N = 125$ different angles, the respective shape normals, camera poses, and their identity embedding vector $\mathbf{w}$ . + +# 4.2. Qualitative Comparisons + +# 4.2.1. Novel View Synthesis comparisons + +In this section, we present a qualitative comparison between our model and other state-of-the-art multi-view dif + +![](images/d818a6599f3e1d45778cf0e9495b29eab1d9c8dfca098286f33807a469539424.jpg) + +![](images/ae710b886140c2ad071d36e266b11994a06409dae5e52008c7921e9150a8aa6e.jpg) + +![](images/2ff3e69530abb9e7d0fcd9766d9e34fa4b7973ec676a374f8b63e2dd22a60728.jpg) + +![](images/d01daae13de9efd69a286b34d0b557262e185280f2c2f745c7cd34482464b632.jpg) +(a) Input + +![](images/069db5a8fefd94e18eb7dbf550ca2addf36946ed8b3448354e0e854e76ddb5de.jpg) +(b) SpinM + +![](images/7b34f54acfeb9aa01551968c720205a1307745f58f40ebabe0821aa8b5d946f1.jpg) + +![](images/1f7b420fed12bfdff778f94326eee91c6161026a2b8c42194b83c09332b30732.jpg) + +![](images/f9455abeb235b7c6e00fea12cb091e21491ba3cbc2f1d88c025240e8d7e38691.jpg) + +![](images/24197544f59a456f4de064987c0dfaddf121f01757521ccd37f211c76f6af6c5.jpg) +Figure 4. Qualitative comparison between SpinMeRound, Zero123-XL [12], SV3D [63] and DiffPortrait3D [25] at angles $\{\pm 45^{\circ}, + 180^{\circ}\}$ . It is clear that SpinMeRound effectively generates high-quality novel views from "in-the-wild" input images, whilst SV3D produces distorted outputs, Zero123-XL generates unnaturally squared avatars and DiffPortrait3D cannot handle such angles. + +![](images/004d706813603a4f671352741f48bbe3ec5444279f64a124c91ff445b97ab705.jpg) + +![](images/cd9f16d8849df2ca6ddc258a76cdc428aeb5fc011b1eb3a121e2dd4c0c3c6427.jpg) + +![](images/8699bf09479e5df4a791f035e7e857f791d8893ccd5751aa00689989411b851e.jpg) + +![](images/04f10cdcd91e99a63f65548fe9a31d3f6e3b756baacde7ea48d1202118069fe6.jpg) + +![](images/f9ab34fc4228f91e4b35cc0a9b2c87af9b779635059ffc15675ad92b46121c58.jpg) + +![](images/7a34fb8a25f158da6ed290436f6758510e6f5e6768135c374d15293a2b891898.jpg) + +![](images/ffaefcfa111f2fd13b2308ada2a267990fb6bef9541358b42d8d263c3736e63a.jpg) + +![](images/05377c43ce4bfabeaf86a483c9d0f0332b5d49d9fadc46d525aa82a469e43bbc.jpg) ++45° + +![](images/8f0dc5eebd0e6547c61339c20816af7ddec62b9ba2dbf5ee9e68fe5efabd9b19.jpg) + +![](images/5992d458f8c2f1412672ed6d1860ff4feb1c114500245b79cc729b971ea6bbb3.jpg) + +![](images/646d6cea1eedbcf16f53c429e2b0c48ba73f8579eb752e7e74d89d720426a0a4.jpg) + +![](images/d68efbfa6f6e8011aeb303eb29e74fcc41597980a617f734d7f2638bfb7225f3.jpg) +(c) SV3D [63] + +![](images/3d4438f889f9eb3241095eee86aac301dbd4393baa704d7d674e803200f893aa.jpg) + +![](images/bfc6899dae281503a0e41446a938f0cfcb7370cbfd45cf8c53f3111bc0f8879b.jpg) + +![](images/9895b41b7c14eec2b22cc2fdbc3cc210f81a40fba57e442c68c56c371166da1b.jpg) + +![](images/b7289fa4afeb371cd90ea21dfc8336a3553c566058bbfa657ce58ed3a313ca9d.jpg) + +![](images/dddd58d01b24ef098c74875ec4b0454845ef662024e5cd971845db10250aabb6.jpg) + +![](images/bfd5f75cfee0222a914e12338e13ad309bc70815f18b6c2b2b152efbb3063d55.jpg) + +![](images/fe47f4ec4edcc7094c4f8f417f1a675539611e03bf371db66bb55fc2cae8f3e4.jpg) + +![](images/be5f98809405363546d5a098013b232d40c8bcb04f60cd145987d70dc1f89cbf.jpg) +(d) Zero123-XL [12] + +![](images/3333863872599a0617ff4967ca502cc4c2b16de77ff3c3eb5e2057fa876d81e9.jpg) + +![](images/2104a4404a3747919545ce7b0871b7fb004df43e93f997be76e23cafb7171214.jpg) + +![](images/81e812aa84697c619614684bc481917d3f86153a894794ab2374ad2dfea9301e.jpg) + +![](images/1310ecbf5fb84e17db969ff090be4d6d83676c7c78c714b2f91fdebac68b0402.jpg) +-45° + +![](images/8099ab0920fcc49325ca7139a3f2fbc2da4d68e5cf1e6484a0af1915c6bbc485.jpg) + +![](images/742f95c7c0bb80ab11f81dffbe527c885003f2799149aec4e96049cb9f188fc7.jpg) + +![](images/6e16733e0447cc8a3d699a661d34e500328b91624f684fc9c3bb6306fad732d0.jpg) + +![](images/81591603fcab7abd2683843363ecdcf0a7b7804876d59bd00345bf6511d77c13.jpg) + +![](images/e92c697f0b1d5b91a372443cdae76534faf8521cf4c86a3baec62704ab63154e.jpg) + +![](images/5855816e181b8d0732336cdb3e4f3c34441a0726f086c340fb961a33c26789b7.jpg) + +![](images/7223f9288ef5a6f235757d5b3d3e4ed5e821e1ae664c82ce6270cbe4917dcae2.jpg) + +![](images/bc1fd73df9540eeef1b69d20603030eca47620f2edeb6b0d2bbc125b6861268c.jpg) +(d) DiffPortrait3d [25] + +![](images/50c2ad859c6fbe0ad39011f4255a8d2c4bd1d5203b900d112a38c7e7cfa92ccf.jpg) + +![](images/62ea2a47c4edb1b9de90a732fb9a30264ebbd33d5adaf490269b50c4749e42f9.jpg) + +![](images/6cbbf5a71bfd5097b4d288f6aecccf580bfff7b2228045bf134855d72fed7f50.jpg) + +![](images/d02a7a584aa561371883b51a416a3a2810dd5f8cd4cb3caaf42f0d1ecd6df08e.jpg) + +fusion models, SV3D [63] Zero123-XL [12] and DiffPort- trait3D [25], focusing on their output under $\pm 45^{\circ}, + 180^{\circ}$ angles. SV3D, a latent video diffusion model, creates $360^{\circ}$ videos from a single input image. Zero123-XL, a multiview diffusion model, synthesizes novel perspectives based on the input image and specified camera poses. DiffPort- trait3D [25] generates novel views by conditioning the desired camera on an input image and refining the initial noise through a fine-tuning step to achieve optimal results. Figure 4 demonstrates scenarios where the input "in-the-wild" images (Fig. 4a ) are fed to these 4 different networks and their respective generations under $\pm 45^{\circ}, + 180^{\circ}$ viewpoints. As illustrated, SpinMeRound efficiently generates accurate novel viewpoints of the input subject, while SV3D distorts the input image in its generated views, Zero123-XL produces unnatural, square-like facial avatars and DiffPort- trait3D cannot handle those angles at all. + +![](images/f3027b182cc163b281867aba12807a0605f15cfd46e26d26986f75e72c96a524.jpg) +Figure 5. Given the input identities (shown in the small boxes), we present the results of novel view synthesis after applying 3D Gaussian splatting [31] to the views generated by our model. + +# 4.3. 3D Reconstruction + +To evaluate our model's consistency in generating novel views from an input image, we assess its ability to reconstruct an input identity through Gaussian Splatting (3DGS) [31]. More specifically, given an input identity, we apply the sampling strategy outlined in Section 3.3 to + +generate 48 novel views. These views are then used to reconstruct the identity through Gaussian Splatting. As in [20], we modify the provided 3DGS code to incorporate the LPIPS [71] loss between the ground-truth and the generated viewpoints. In this way, we can deal with small inconsistencies between nearby viewpoints. We present the novel view synthesis of two input identities (small squares) in Figure 5, which clearly illustrates that our proposed methodology can generate consistent subjects. + +![](images/83c2a33f3fe24124d532df0ebdc66ae96eacebbeebb01ace9391cda9100825ad.jpg) + +![](images/5f08c3a2b94be8fe7126488cecb580ad1b64f6df493cf2f251995ca1873a33b4.jpg) +Figure 6. Samples generated using unconditional sampling. Spin-MeRound can generate novel identities without any prior input. + +![](images/f3a6a29d299eb3c8e5d3a7fa25c4338002994d8cc9ce0e91c639b3e3bcb0ec41.jpg) + +![](images/8a6e445ed53d713c46723fd5cb65b9b50c3c8f114be118ad62ae752690387680.jpg) + +![](images/2c9149d69baed028d58782a1fd05123c48e03a0ad97689fbe3d1040e104d68a0.jpg) + +![](images/6a2c2201223e3b27f0ab06f9ef79c2dbd1480ef321a1eafbe54406c626568ee9.jpg) + +![](images/b02f7a889c81da213224af37d6daa0fad29a0ce0b8dc870cb89587f7f6ad77c8.jpg) + +![](images/be729dd9b6f84d5572990928dc65a6876a66b473a7578ad882ab7d5851eaeb88.jpg) + +![](images/302bffb2332ecc37d3a4884bfbbc6d8b3dbb4e6228e4ec66678f84e09878ce1f.jpg) + +![](images/1335d900a62c968999befd2ba1c646433ddfeb5cca524dc18712ccb23158c099.jpg) + +![](images/eb548ecab6db5ee896aeffa79fd5cdb1171897f10b7d295b93fc05e7d3a8a75b.jpg) + +![](images/175221f93b13db844f175279c309b5bb6d39da69577f9e215dee1cec50602c19.jpg) + +![](images/d19dfe15285c7e9446ade34584deabbb0520e2b031e2ad4a5ce1201a8d51cf0f.jpg) + +![](images/358f696296781c523fddaba5ffaa22fb87c25e9984665f6f211259046c2a6339.jpg) + +# 4.4. Unconditional Sampling + +SpinMeRound is trained following the Classifier-Free Guidance training scheme. Thus, our architecture can generate novel multi-view identities without any prior conditioning. By setting the input identity embedding equal to the empty string, our model can generate a novel identity along with different viewpoints of that identity. Figure 6 presents some examples of those identities. + +# 4.5.Quantitative Comparisons + +We compare our method's ability to generate novel views with the current state-of-the-art models: a) Eg3D [9] and + +
MethodL2 ↓LPIPS↓SSIM↑ID Sim [13] ↑ID Sim [48] ↑
NeRFEg3d [9]0.0250.40.550.310.89
basedPanohead [1]0.0120.320.650.270.88
Zero123 [39]0.1950.5150.550.1690.44
Diffusion basedZero123-XL [12]0.1980.510.5630.1180.442
SV3D [63]0.0870.410.6600.360.881
DiffPortrait3d [25]0.10.50.350.550.887
SpinMeRound (Ours)0.0330.30.730.610.911
+ +Panohead [1] which are NeRF-based approaches generating frontal and full-head portraits respectively, b) Zero123 [68], Zero123-XL [12] and DiffPortrai3D [25], which are multiview diffusion models and c) SV3D [63], a multi-view video diffusion model. For SV3D, we selected the SV3Dp variant, as it supports flexible viewing angles, whereas SV3D-u operates only with fixed viewpoints. We validate our approach using the NeRSemble dataset [32], which includes 222 unique identities recorded from 16 angles. We randomly select a timestamp from one of their video sequences for each individual and extract a random subset of views. All selected frames are then centered, and one of these views is used as input for the tested methods. We evaluate their reconstruction performance using L2 distance, LPIPS [71], SSIM, and an Identity Similarity score. To calculate the Identity Similarity score, we feed both the ground-truth and the reconstructed images into the ArcFace [13] and VGGFace [48] networks and measure the cosine similarity between their final feature vectors. + +Table 1 summarizes the reconstruction performance of each network. SpinMeRound, achieves state-of-the-art results across all cases when evaluated on the LPIPS, SSIM, and Identity Similarity metrics. Additionally, it outperforms all diffusion-based approaches in terms of L2 distance. Although our method trails slightly behind NeRF-based approaches, this is expected given the inherent advantages of NeRF-based models in novel view synthesis tasks. Notably, methods like Eg3D and Panohead require a time-intensive fitting process, which can occasionally fail. In contrast, our proposed approach relies solely on an efficient sampling process, eliminating the need for fitting. + +# 5. Ablation Studies + +# 5.1. Component analysis + +In this section, we conduct ablation studies for the importance of the identity mechanism, the shape normals and the input ground truth image. For this reason, we trained the following three models while using the same training data: a) our proposed architecture, without integrating the input conditioning view, b) a Stable Diffusion-based model without integrating the robust identity mechanism, and c) our model without generating any shape normals. + +Table 1. Reconstruction performance on the NeRSample [32] dataset shows that SpinMeRound achieves state-of-the-art results in LPIPS, SSIM and ID Sim metrics while performing on par with the leading models in terms of L2 distance. + +
L2 ↓LPIPS ↓SSIM ↑
SpinMeRound (w/o Input Image)0.12460.42990.568
SpinMeRound (w/o identity embedding)0.0280.260.70
SpinMeRound (w/o Normals)0.0560.320.65
SpinMeRound0.0180.220.75
+ +Table 2. Ablation study: We evaluate the performance of Spin-MeRound alongside three variations, demonstrating that the proposed architecture achieves the highest performance across all reconstruction metrics. + +![](images/c0db0fb64d4ed31f3ea06fdbde5a275cb7647a6cbb5f1e3f324595d5dc20910a.jpg) +Figure 7. Ablation Study: We compare the identity similarity between four models: SpinMeRound (Ours), a similar model without the input image, ones without the identity embedding mechanism, and one without generating shape normals $\mathcal{N}$ . Results show that our proposed architecture achieves the highest identity similarity. + +Use of input image: Starting from a pre-trained Arc2Face model using the EDM framework, we trained a multi-view model to generate novel viewpoints given only an input identity embedding vector. Notably, SpinMeRound has also been trained to be able to generate input identities without providing the conditioning view. In this case, it generates subjects close to the input face, as illustrated in Figure 8. + +Identity embedding mechanism: In this case, we start from a pre-trained Stable diffusion 1.5 architecture trained using the EDM framework. This model gets trained following the same training parameters, as presented in Sec 3.2. Instead of using the proposed identity conditioning mechanism, we feed an empty string in the conditioning attention layers of the denoising UNet. + +Use of Normals: We also train a variant of the network similar to SpinMeRound, but without generating shape normals. This version can only generate novel views without having any additional information about the facial shape. + +Aiming to showcase the importance of each component, we conducted a multi-view reconstruction experiment. More specifically, we sample 100 distinct identities using Panohead [1]. For each subject, we sample 10 different views along with an input frontal view. Given the frontal + +![](images/734d747f27c7e00de1a27859f50087eff27c210f912c3f1ba4ec94ba0215b88c.jpg) +Figure 8. From the input images on the left, we present samples generated by our method using only the corresponding identity embedding vectors (top row) and using also the input image (bottom row). The resulting subjects closely resemble the input identities. + +![](images/45efca35cf8c8420dbbea15dfa89e5396ae7df4c536affc1b19d2dd76822125e.jpg) +Figure 9. Identity interpolation between subject pairs, showcasing our method generates smooth transitions between identities. + +view as input to each separate model, we reconstruct the remaining views. We measure the performance of each model by calculating their discrepancy using L2 distance, LPIPS [71], SSIM and ID Similarity. As presented in Table 2 and Figure 7, our proposed methodology performs better regarding all reconstruction metrics while achieving the highest identity similarity score. + +# 5.2. Importance of ID features + +Choosing to condition the stable diffusion model in identity embeddings is an important design choice for our network. Those forms of representation contain compact information extracted from a FR model (ArcFace [13]), trained in millions of different images and subjects. Those features allow linear interpolation between the facial characteristics of different subjects. Hence, while using only the identity embedding layer, we interpolate between 2 distinct identities and present their linear interpolation results in Figure 9. It is + +clearly shown that our model generates smooth transitions between the generated identities. + +# 6. Conclusion + +In this paper, we presented SpinMeRound, a multi-view latent diffusion model, which generates all-around head portraits of an input subject given a number of input views alongside an input identity embedding. Additionally, we introduced a sampling strategy for generating consistent intermediate views, given an unconstrained input "in-the-wild" facial image. Being trained sorely on multi-view synthetic data, we showcase our method's abilities by beating the current state-of-the-art multi-view diffusion models in novel view synthesis experiments. + +Acknowledgments: S. Zafeiriou and part of the research was funded by the EPSRC Fellowship DEFORM (EP/S010203/1) and EPSRC Project GNOMON (EP/X011364/1). The authors gratefully acknowledge the scientific support and HPC resources provided by the Erlangen National High Performance Computing Center (NHR@FAU) of FAU under the NHR projects b143dc and b180dc. NHR funding is provided by federal and Bavarian state authorities. NHR@FAU hardware is partially funded by the German Research Foundation (DFG) - 440719683. Additional support was also received by the ERC - project MIA-NORMAL 101083647, DFG 513220538, 512819079, and by the state of Bavaria (HTA). + +# References + +[1] Sizhe An, Hongyi Xu, Yichun Shi, Guoxian Song, Umit Ogras, and Linjie Luo. Panohed: Geometry-aware 3d fullhead synthesis in $360^{\circ}$ , 2023. 2, 5, 7 +[2] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[3] Francesca Babiloni, Alexandros Lattas, Jiankang Deng, and Stefanos Zafeiriou. Id-to-3d: Expressive id-guided 3d heads via score distillation sampling. In Advances in Neural Information Processing Systems, pages 97167–97183. Curran Associates, Inc., 2024. 3 +[4] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In SIGGRAPH '99, 1999. 2 +[5] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets, 2023. 1, 3, 4 +[6] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 3 +[7] James Booth, Anastasios Roussos, Stefanos Zafeiriou, Allan Ponniahy, and David Dunaway. A 3D morphable model learnt from 10,000 faces. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 5543-5552, 2016. 2 +[8] James Booth, Anastasios Roussos, Allan Ponniah, David Dunaway, and Stefanos Zafeiriou. Large scale 3d morphable models. International Journal of Computer Vision, 126(2): 233-254, 2018. 2 +[9] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J. Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16123-16133, 2022. 2, 6, 7 +[10] Xiyi Chen, Marko Mihajlovic, Shaofei Wang, Sergey Prokudin, and Siyu Tang. Morphable diffusion: 3d-consistent diffusion for single-image avatar creation. 2024. 2, 3 +[11] Zilong Chen, Yikai Wang, Feng Wang, Zhengyi Wang, and Huaping Liu. V3d: Video diffusion models are effective 3d generators. arXiv preprint arXiv:2403.06738, 2024. 3 +[12] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-xl: A universe of $10\mathrm{m} + 3\mathrm{d}$ objects. arXiv preprint arXiv:2307.05663, 2023. 6, 7 + +[13] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 3, 4, 5, 7, 8 +[14] Yu Deng, Duomin Wang, Xiaohang Ren, Xingyu Chen, and Baoyuan Wang. Portrait4d: Learning one-shot 4d head avatar synthesis using synthetic data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3 +[15] Yu Deng, Duomin Wang, and Baoyuan Wang. Portrait4d-v2: Pseudo multi-view data creates better 4d head synthesizer. arXiv preprint arXiv:2403.13570, 2024. 3 +[16] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In Advances in Neural Information Processing Systems, pages 8780-8794. Curran Associates, Inc., 2021. 1, 2, 3 +[17] Yao Feng, Haiwen Feng, Michael J. Black, and Timo Bolkart. Learning an animatable detailed 3D face model from in-the-wild images. 2021. 2 +[18] Stathis Galanakis, Baris Gecer, Alexandros Lattas, and Stefanos Zafeiriou. 3dmm-rf: Convolutional radiance fields for 3d face modeling. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 3536-3547, 2023. 2 +[19] Stathis Galanakis, Alexandros Lattas, Stylianos Moschoglou, and Stefanos Zafeiriou. Fitdiff: Robust monocular 3d facial shape and reflectance estimation using diffusion models, 2024. 3, 4 +[20] Ruiqi Gao*, Aleksander Holynski*, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul P. Srinivasan, Jonathan T. Barron, and Ben Poole*. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv, 2024. 2, 3, 4, 5, 6 +[21] Baris Gecer, Stylianos Ploumpis, Irene Kotsia, and Stefanos Zafeiriou. Ganfit: Generative adversarial network fitting for high fidelity 3d face reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[22] Dimitrios Gerogiannis, Foivos Paraperas Papantoniou, Rolandos Alexandros Potamias, Alexandros Lattas, and Stefanos Zafeiriou. Arc2avatar: Generating expressive 3d avatars from a single image via id guidance. arXiv preprint arXiv:2501.05379, 2025. 3 +[23] Simon Giebenhain, Tobias Kirschstein, Markos Georgopoulos, Martin Rinz, Lourdes Agapito, and Matthias Nießner. Learning neural parametric head models. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[24] Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. Emu video: Factorizing text-to-video generation by explicit image conditioning, 2024. 3 +[25] Yuming Gu, Hongyi Xu, You Xie, Guoxian Song, Yichun Shi, Di Chang, Jing Yang, and Linjie Luo. Diffportrait3d: Controllable diffusion for zero-shot portrait view synthe + +sis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10456-10465, 2024. 2, 3, 6, 7 +[26] Yuming Gu, Phong Tran, Yujuan Zheng, Hongyi Xu, Heyuan Li, Adilbek Karmanov, and Hao Li. Diffportrait360: Consistent portrait diffusion for 360 view synthesis, 2025. 3 +[27] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5 +[28] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239, 2020. 1, 3, 5 +[29] Yash Kant, Ethan Weber, Jin Kyu Kim, Rawal Khirodkar, Su Zhaoen, Julieta Martinez, Igor Gilitschenski, Shunsuke Saito, and Timur Bagautdinov. Pippo: High-resolution multi-view humans from a single image. 2025. 3 +[30] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Proc. NeurIPS, 2022. 4, 5 +[31] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 6 +[32] Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radiance field reconstruction of human heads. ACM Trans. Graph., 42(4), 2023. 7 +[33] Yushi Lan, Feitong Tan, Di Qiu, Qiangeng Xu, Kyle Genova, Zeng Huang, Sean Fanello, Rohit Pandey, Thomas Funkhouser, Chen Change Loy, and Yinda Zhang. Gaussian3diff: 3d gaussian diffusion for 3d full head synthesis and editing. In ECCV, 2024. 3 +[34] Alexandros Lattas, Stylianos Moschoglou, Stylianos Ploumpis, Baris Gecer, Jiankang Deng, and Stefanos Zafeiriou. FitMe: Deep photorealistic 3D morphable model avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[35] Heyuan Li, Ce Chen, Tianhao Shi, Yuda Qiu, Sizhe An, Guanying Chen, and Xiaoguang Han. Spherehead: Stable 3d full-head synthesis with spherical tri-plane representation, 2024. 2 +[36] Peng Li, Yuan Liu, Xiaoxiao Long, Feihu Zhang, Cheng Lin, Mengfei Li, Xingqun Qi, Shanghang Zhang, Wenhan Luo, Ping Tan, et al. Era3d: High-resolution multiview diffusion using efficient row-wise attention. arXiv preprint arXiv:2405.11616, 2024. 2, 3 +[37] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017. 2 +[38] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems, 36, 2024. 3 +[39] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object, 2023. 2, 3, 7 + +[40] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 3 +[41] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. In SIGGRAPH, pages 163-169. ACM, 1987. 5 +[42] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, Natalia Neverova, Andrea Vedaldi, Oran Gafni, and Filippos Kokkinos. Im-3d: Iterative multiview diffusion and reconstruction for high-quality 3d generation, 2024. 2, 3 +[43] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 2 +[44] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13503–13513, 2022. 2 +[45] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. StyleSDF: High-Resolution 3D-Consistent Image and Geometry Generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13503–13513, 2022. 2 +[46] Foivos Paraperas Papantoniou, Alexandros Lattas, Stylianos Moschoglou, and Stefanos Zafeiriou. Relightify: Relightable 3d faces from a single image via diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 3, 4, 5 +[47] Foivos Paraperas Papantoniou, Alexandros Lattas, Stylianos Moschoglou, Jiankang Deng, Bernhard Kainz, and Stefanos Zafeiriou. Arc2face: A foundation model for id-consistent human faces. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 4 +[48] Omkar M. Parkhi, Andrea Vedaldi, and Andrew Zisserman. Deep face recognition. In *British Machine Vision Conference*, 2015. 7 +[49] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 2, 3 +[50] Malte Prinzler, Egor Zakharov, Vanessa Sklyarova, Berna Kabadayi, and Justus Thies. Joker: Conditional 3d head synthesis with extreme facial expressions, 2024. 3 +[51] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, and Bernard Ghanem. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. In The Twelfth International Conference on Learning Representations (ICLR), 2024. 3 +[52] Alec Radford, Jong Wook Kim, Chris Hallacy, A. Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and + +Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 4 +[53] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Ben Mildenhall, Nataniel Ruiz, Shiran Zada, Kfir Aberman, Michael Rubenstein, Jonathan Barron, Yuanzhen Li, and Varun Jampani. Dreambooth3d: Subject-driven text-to-3d generation. ICCV, 2023. 3 +[54] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 5 +[55] Daniel Roich, Ron Mokady, Amit H. Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images, 2021. 2 +[56] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2, 3, 4 +[57] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. CoRR, abs/1505.04597, 2015. 4 +[58] Mehdi S. M. Sajjadi, Henning Meyer, Etienne Pot, Urs Bergmann, Klaus Greff, Noha Radwan, Suhani Vora, Mario Lucic, Daniel Duckworth, Alexey Dosovitskiy, Jakob Uszkoreit, Thomas Funkhouser, and Andrea Tagliasacchi. Scene Representation Transformer: Geometry-Free Novel View Synthesis Through Set-Latent Scene Representations. CVPR, 2022. 4 +[59] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model, 2023. 3 +[60] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv:2308.16512, 2023. 4 +[61] Felix Taubner, Ruihang Zhang, Mathieu Tuli, and David B. Lindell. Cap4d: Creating animatable 4d portrait avatars with morphable multi-view diffusion models, 2024. 3 +[62] Alex Trevithick, Matthew Chan, Michael Stengel, Eric R. Chan, Chao Liu, Zhiding Yu, Sameh Khamis, Manmohan Chandraker, Ravi Ramamoorthi, and Koki Nagano. Real-time radiance fields for single-image portrait view synthesis. In ACM Transactions on Graphics (SIGGRAPH), 2023. 2 +[63] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion, 2024. 2, 3, 6, 7 +[64] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 3 +[65] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, and Baining Guo. Rodin: A generative model for sculpting 3d digital avatars using diffusion, 2022. 3 + +[66] Haohan Weng, Tianyu Yang, Jianan Wang, Yu Li, Tong Zhang, C. L. Philip Chen, and Lei Zhang. Consistent123: Improve consistency for one image to 3d object synthesis, 2023. 3 +[67] Yiqian Wu, Hao Xu, Xiangjun Tang, Hongbo Fu, and Xiao-gang Jin. 3dportraitgan: Learning one-quarter headshot 3d gans from a single-view portrait dataset with diverse body poses, 2023. 2 +[68] Yuyang Huang Taoran Yi Xiaopeng Zhang Lingxi Xie Xinggang Wang Wenrui Dai Hongkai Xiong and Qi Tian Yabo Chen, Jiemin Fang. Cascade-zero123: One image to highly consistent 3d with self-prompted nearby views. arXiv preprint arXiv:2312.04424, 2023. 3, 7 +[69] Tarun Yenamandra, Ayush Tewari, Florian Bernard, Hans-Peter Seidel, Mohamed Elgharib, Daniel Cremers, and Christian Theobalt. i3dmm: Deep implicit 3d morphable model of human heads. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12803-12813, 2021. 2 +[70] Bowen Zhang, Yiji Cheng, Chunyu Wang, Ting Zhang, Jiaolong Yang, Yansong Tang, Feng Zhao, Dong Chen, and Baining Guo. Rodinhd: High-fidelity 3d avatar generation with diffusion models. arXiv preprint arXiv:2407.06938, 2024. 3 +[71] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6, 7, 8 +[72] Mingwu Zheng, Hongyu Yang, Di Huang, and Liming Chen. Imface: A nonlinear 3d morphable face model with implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20343-20352, 2022. 2 +[73] Zhenglin Zhou, Fan Ma, Hehe Fan, Zongxin Yang, and Yi Yang. Headstudio: Text to animatable head avatars with 3d gaussian splatting. 2024. 2, 3 +[74] Zheng Zhu, Guan Huang, Jiankang Deng, Yun Ye, Junjie Huang, Xinze Chen, Jiagang Zhu, Tian Yang, Jiwen Lu, Dalong Du, and Jie Zhou. Webface260m: A benchmark unveiling the power of million-scale deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10492-10502, 2021. 3 \ No newline at end of file diff --git a/2025/SpinMeRound_ Consistent Multi-View Identity Generation Using Diffusion Models/images.zip b/2025/SpinMeRound_ Consistent Multi-View Identity Generation Using Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2509226cf965ae6f2a37ae8cc4f1d0818442ef61 --- /dev/null +++ b/2025/SpinMeRound_ Consistent Multi-View Identity Generation Using Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c3c3f6f3a15eacdd75eb8fdae6a02127c987d6bb7d0a393e1deff040633fd606 +size 661945 diff --git a/2025/SpinMeRound_ Consistent Multi-View Identity Generation Using Diffusion Models/layout.json b/2025/SpinMeRound_ Consistent Multi-View Identity Generation Using Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fde92311b055f11e003f8425eacedb1611aa5fff --- /dev/null +++ b/2025/SpinMeRound_ Consistent Multi-View Identity Generation Using Diffusion Models/layout.json @@ -0,0 +1,11991 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 73, + 102, + 536, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 102, + 536, + 138 + ], + "spans": [ + { + "bbox": [ + 73, + 102, + 536, + 138 + ], + "type": "text", + "content": "SpinMeRound: Consistent Multi-View Identity Generation Using Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 84, + 160, + 175, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 160, + 175, + 175 + ], + "spans": [ + { + "bbox": [ + 84, + 160, + 175, + 175 + ], + "type": "text", + "content": "Stathis Galanakis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 246, + 161, + 339, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 161, + 339, + 175 + ], + "spans": [ + { + "bbox": [ + 246, + 161, + 339, + 175 + ], + "type": "text", + "content": "Alexandros Lattas1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 194, + 175, + 284, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 175, + 284, + 188 + ], + "spans": [ + { + "bbox": [ + 194, + 175, + 284, + 188 + ], + "type": "text", + "content": "Bernhard Kainz" + }, + { + "bbox": [ + 194, + 175, + 284, + 188 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 411, + 161, + 525, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 161, + 525, + 175 + ], + "spans": [ + { + "bbox": [ + 411, + 161, + 525, + 175 + ], + "type": "text", + "content": "Stylianos Moschoglou" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 365, + 175, + 460, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 175, + 460, + 188 + ], + "spans": [ + { + "bbox": [ + 365, + 175, + 460, + 188 + ], + "type": "text", + "content": "Stefanos Zafeiriou" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 230, + 202, + 381, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 202, + 381, + 217 + ], + "spans": [ + { + "bbox": [ + 230, + 202, + 381, + 217 + ], + "type": "text", + "content": "1Imperial College London, UK" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 215, + 217, + 392, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 217, + 392, + 232 + ], + "spans": [ + { + "bbox": [ + 215, + 217, + 392, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 215, + 217, + 392, + 232 + ], + "type": "text", + "content": "FAU Erlangen-Nürnberg, Germany" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 61, + 242, + 171, + 376 + ], + "blocks": [ + { + "bbox": [ + 61, + 242, + 171, + 376 + ], + "lines": [ + { + "bbox": [ + 61, + 242, + 171, + 376 + ], + "spans": [ + { + "bbox": [ + 61, + 242, + 171, + 376 + ], + "type": "image", + "image_path": "bb20f158f29f6968ffdbb355f3b07c740b4a6bb02d4d46669aed8360dc346502.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 61, + 376, + 94, + 415 + ], + "blocks": [ + { + "bbox": [ + 61, + 376, + 94, + 415 + ], + "lines": [ + { + "bbox": [ + 61, + 376, + 94, + 415 + ], + "spans": [ + { + "bbox": [ + 61, + 376, + 94, + 415 + ], + "type": "image", + "image_path": "0836e9de2f71a06f81d6eb5edef1c613f5e346d0dd77134120eb276784dd2460.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 445, + 555, + 469 + ], + "lines": [ + { + "bbox": [ + 55, + 445, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 555, + 469 + ], + "type": "text", + "content": "Figure 1. SpinMeRound is a multiview diffusion model which generates human portraits from novel viewpoints. Given a single or multiple views, our method produces high-fidelity images along with precise surface normals, ensuring accurate 3D consistency across perspectives." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 96, + 376, + 134, + 416 + ], + "blocks": [ + { + "bbox": [ + 96, + 376, + 134, + 416 + ], + "lines": [ + { + "bbox": [ + 96, + 376, + 134, + 416 + ], + "spans": [ + { + "bbox": [ + 96, + 376, + 134, + 416 + ], + "type": "image", + "image_path": "ed72d1ea1c1271ea1991e5e982f2596ff60ecac855921160904fd561a1c4f517.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 123, + 426, + 234, + 437 + ], + "lines": [ + { + "bbox": [ + 123, + 426, + 234, + 437 + ], + "spans": [ + { + "bbox": [ + 123, + 426, + 234, + 437 + ], + "type": "text", + "content": "Single Image to novel views" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 134, + 376, + 165, + 416 + ], + "blocks": [ + { + "bbox": [ + 134, + 376, + 165, + 416 + ], + "lines": [ + { + "bbox": [ + 134, + 376, + 165, + 416 + ], + "spans": [ + { + "bbox": [ + 134, + 376, + 165, + 416 + ], + "type": "image", + "image_path": "9cef7845c47785dbb3a477bdd4ac232891b704856525474066cb00a26fb9295e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 206, + 243, + 299, + 376 + ], + "blocks": [ + { + "bbox": [ + 206, + 243, + 299, + 376 + ], + "lines": [ + { + "bbox": [ + 206, + 243, + 299, + 376 + ], + "spans": [ + { + "bbox": [ + 206, + 243, + 299, + 376 + ], + "type": "image", + "image_path": "9797b68e5950dbf682bd41df45d303ba6b4949027585fccffd657e24592ed411.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 193, + 376, + 231, + 415 + ], + "blocks": [ + { + "bbox": [ + 193, + 376, + 231, + 415 + ], + "lines": [ + { + "bbox": [ + 193, + 376, + 231, + 415 + ], + "spans": [ + { + "bbox": [ + 193, + 376, + 231, + 415 + ], + "type": "image", + "image_path": "530921576dfff87d0e844bc7f6510e30a22930cc7e80ec03ff47847db3878c73.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 232, + 376, + 266, + 415 + ], + "blocks": [ + { + "bbox": [ + 232, + 376, + 266, + 415 + ], + "lines": [ + { + "bbox": [ + 232, + 376, + 266, + 415 + ], + "spans": [ + { + "bbox": [ + 232, + 376, + 266, + 415 + ], + "type": "image", + "image_path": "32c7851a65ada9ffdcda4b4351029d787bc8fbc58f94fb16213535dee3664c52.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 270, + 376, + 309, + 415 + ], + "blocks": [ + { + "bbox": [ + 270, + 376, + 309, + 415 + ], + "lines": [ + { + "bbox": [ + 270, + 376, + 309, + 415 + ], + "spans": [ + { + "bbox": [ + 270, + 376, + 309, + 415 + ], + "type": "image", + "image_path": "2b9105ea6dd7efc73c974d913f9068354708c6cf68e0a105dbf3c1f14121679d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 328, + 243, + 555, + 376 + ], + "blocks": [ + { + "bbox": [ + 328, + 243, + 555, + 376 + ], + "lines": [ + { + "bbox": [ + 328, + 243, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 328, + 243, + 555, + 376 + ], + "type": "image", + "image_path": "f87676e2202ef8f80c41a804654444a0e024fcacaa3220784c9a46bd8b30ece7.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 329, + 376, + 358, + 416 + ], + "blocks": [ + { + "bbox": [ + 329, + 376, + 358, + 416 + ], + "lines": [ + { + "bbox": [ + 329, + 376, + 358, + 416 + ], + "spans": [ + { + "bbox": [ + 329, + 376, + 358, + 416 + ], + "type": "image", + "image_path": "48dd5b39773110129c5c02a2b84a8f47c3522ad67e436d4a2eb2af4c0c269975.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 426, + 494, + 436 + ], + "lines": [ + { + "bbox": [ + 384, + 426, + 494, + 436 + ], + "spans": [ + { + "bbox": [ + 384, + 426, + 494, + 436 + ], + "type": "text", + "content": "Sparse views to novel views" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 361, + 376, + 400, + 415 + ], + "blocks": [ + { + "bbox": [ + 361, + 376, + 400, + 415 + ], + "lines": [ + { + "bbox": [ + 361, + 376, + 400, + 415 + ], + "spans": [ + { + "bbox": [ + 361, + 376, + 400, + 415 + ], + "type": "image", + "image_path": "7b096e34531b1e53e1168df01186eb32f2e623a78855597d0f2e71a9412c9c2c.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 406, + 376, + 438, + 415 + ], + "blocks": [ + { + "bbox": [ + 406, + 376, + 438, + 415 + ], + "lines": [ + { + "bbox": [ + 406, + 376, + 438, + 415 + ], + "spans": [ + { + "bbox": [ + 406, + 376, + 438, + 415 + ], + "type": "image", + "image_path": "7d06defaff480f16169d3b7f57c2242016f01629842184614866f355640c7317.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 439, + 376, + 477, + 415 + ], + "blocks": [ + { + "bbox": [ + 439, + 376, + 477, + 415 + ], + "lines": [ + { + "bbox": [ + 439, + 376, + 477, + 415 + ], + "spans": [ + { + "bbox": [ + 439, + 376, + 477, + 415 + ], + "type": "image", + "image_path": "a9a96543f3813a273438823097800181f7ffd6a7314e882770b70fa24392ad3e.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 481, + 376, + 515, + 416 + ], + "blocks": [ + { + "bbox": [ + 481, + 376, + 515, + 416 + ], + "lines": [ + { + "bbox": [ + 481, + 376, + 515, + 416 + ], + "spans": [ + { + "bbox": [ + 481, + 376, + 515, + 416 + ], + "type": "image", + "image_path": "b449fcc0465bae1bad11a9ad3e3dd469d9ba6428df6cd4fa0da94ff5d8e89b7a.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 516, + 376, + 553, + 415 + ], + "blocks": [ + { + "bbox": [ + 516, + 376, + 553, + 415 + ], + "lines": [ + { + "bbox": [ + 516, + 376, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 516, + 376, + 553, + 415 + ], + "type": "image", + "image_path": "ad00c0d328cb3b22eea6be8f8e2073676ec03c2bd144379052d0e55c141a8282.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 152, + 478, + 200, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 478, + 200, + 491 + ], + "spans": [ + { + "bbox": [ + 152, + 478, + 200, + 491 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 54, + 510, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 510, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 510, + 297, + 714 + ], + "type": "text", + "content": "Despite recent progress in diffusion models, generating realistic head portraits from novel viewpoints remains a significant challenge in computer vision. Most current approaches are constrained to limited angular ranges, predominantly focusing on frontal or near-frontal views. Moreover, although the recent emerging large-scale diffusion models have been proven robust in handling 3D scenes, they underperform on facial data, given their complex structure and the uncanny valley pitfalls. In this paper, we propose SpinMeRound, a diffusion-based approach designed to generate consistent and accurate head portraits from novel viewpoints. By leveraging a number of input views alongside an identity embedding, our method effectively synthesizes diverse viewpoints of a subject whilst robustly maintaining its unique identity features. Through experimentation, we showcase our model's generation capabilities in full head synthesis, while beating current state-of-the" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 479, + 555, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 555, + 503 + ], + "type": "text", + "content": "art multi-view diffusion models. Project page: spin-meround.github.io." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 314, + 525, + 395, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 525, + 395, + 538 + ], + "spans": [ + { + "bbox": [ + 314, + 525, + 395, + 538 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 312, + 545, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 545, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 312, + 545, + 555, + 700 + ], + "type": "text", + "content": "Recent advancements in deep learning have brought significant progress to fundamental computer vision tasks, notably in image and video generation [6, 56]. The introduction of Diffusion Models [16, 28] has transformed these domains by enabling the generation of high-quality visual content, facilitated by the use of large-scale datasets [5, 56]. Despite these developments, the task of generating consistent and accurate head portraits from a single input image still remains a challenging problem. This difficulty is primarily attributed to the limited availability of comprehensive 3D facial datasets, which constrains the training of models capable of reliably capturing and reconstructing the detailed structure and variations of human faces." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "content": "A common practice for modeling human heads from" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14346" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "content": "a single image, incorporates the use of a 3D Morphable Model (3DMM) [4, 7, 8, 37], to represent the facial shape along with an appearance model [17, 21, 34]. However, given typical training data and the difficulty in modeling complex hairstyles, these methods focus solely on the facial region and avoid or miss the full head and hair. The seminal work of Neural Radiance Fields (NeRF) [43] led to an explosion of works on neural rendering of scenes that could not easily be modeled with textured meshes. In addition, pairing such implicit representations with generative models led to a wide variety of approaches [2, 9, 44] that pushed the boundaries on the synthesis of novel facial views, achieving high quality and control. More recently, Panohead [1] first showcased high-quality " + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "content": " head portrait synthesis. However, because of its adaptive camera training scheme, the back-head synthesis typically contains many artifacts [35], and its inversion on \"in-the-wild\" images is challenging and requires complex fine-tuning [55]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 293, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 296, + 715 + ], + "type": "text", + "content": "Recently, diffusion models [16] demonstrated superior performance over GAN-based methods in image generation and have achieved great quality in human generation tasks [56]. However, achieving multi-view consistency remains a significant challenge. Despite the lack of accurate 3D datasets, recent advances, such as Score Distillation techniques [49, 73], represent an initial step toward leveraging the 2D generation strengths of diffusion models to construct 3D content without extra training. Nevertheless, these approaches are computationally intensive, require intricate constraints, and do not consistently yield photorealistic results. Meanwhile, multi-view diffusion architectures, employing video diffusion models as a backbone framework [42, 63], still remain very resource-demanding for generating a single novel view, as they rely on comprehensive camera trajectories to generate coherent central objects. Closer to our work, Zero123 [39] introduced a view-conditioned diffusion model that incorporates view features and camera information into the diffusion process. However, the generated images lack strong multi-view consistency and are of low quality, which restricts their performance in photorealistic 3D generation. Although Diff-Portrait3D [25] enables novel view generation through a diffusion process, it still focuses only on near-front views and utilizes a image-conditioned camera mechanism. Other closely related works are Era3D [36] and Morphable Diffusion [10], which generate fixed camera viewpoints, thus limiting their ability to produce full head portraits. The recently proposed Cat3D [20] presents a promising solution by efficiently integrating a number of input views with specified camera poses to achieve a consistent novel viewpoint generation. It integrates 3D attention layers for efficiently sharing common information between all views, along with robust camera pose feature maps. Still, Cat3D is limited by its lack of focus on human generation and is currently" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 313, + 72, + 457, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 457, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 457, + 84 + ], + "type": "text", + "content": "unavailable as an open-source tool." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 86, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 86, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 86, + 553, + 312 + ], + "type": "text", + "content": "In this paper, we present SpinMeRound, a multi-view diffusion model designed to generate novel high-fidelity views of a given human face. In addition to facial images, our model also generates the corresponding normals, which are typically available for human data and, as we show, improve the model's performance and consistency on intricate facial features. Moreover, we show that conditioning the model on an identity embedding and one or more input views of a subject during inference, we can not only sample but also reconstruct multi-view consistent images from \"in-the-wild\" facial images. Our method can accurately synthesize photorealistic head portraits from various angles while preserving essential identity characteristics, which can be used to represent or reconstruct 3D scenes. Given the lack of open-source large-scale multi-view head datasets and the problematic nature of the permissions of such datasets, our method is solely trained on synthetic data acquired using Panohead [1], making this work accessible to experiment and build with. Overall, in this paper:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 314, + 553, + 433 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 314, + 314, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 314, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 314, + 314, + 553, + 361 + ], + "type": "text", + "content": "- We introduce SpinMeRound, a multi-view diffusion model conditioned on identity embeddings and a number of views that generates novel perspectives of an input subject and their respective normals." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 361, + 553, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 361, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 314, + 361, + 553, + 396 + ], + "type": "text", + "content": "- We present a novel sampling strategy which, given a single \"in-the-wild\" facial image, generates consistent views encompassing the whole head." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 397, + 553, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 397, + 553, + 433 + ], + "spans": [ + { + "bbox": [ + 314, + 397, + 553, + 433 + ], + "type": "text", + "content": "- We explore its potential by comparing it with current state-of-the-art multi-view diffusion-based methods and showcase superior results in full-head portrait generation." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 447, + 402, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 402, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 402, + 459 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 468, + 406, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 406, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 406, + 480 + ], + "type": "text", + "content": "2.1. Face Modeling" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": "Extensive research has been dedicated to representing 3D faces through a combination of texture maps and 3D meshes, starting with fundamental 3D Morphable Models (3DMM) [4, 7, 8, 37]. These approaches, however, primarily focus on accurately capturing only the frontal region of the head whilst lacking in integrating finer head details such as hair, wrinkles and wearable items. Dealing with this, recent studies have integrated implicit representations for 3D face modeling. Methods such as [23, 45, 72] integrate Signed Distance Functions (SDFs) whereas other [2, 9, 18, 69] use Neural Radiance Field (NeRF) models [43] for generating photorealistic results. Levering the powerful triplane representation [9], RTRF [62] generates nearfrontal views in real-time whilst Panohead [1] introduces an adaptive camera training strategy for enabling full head portrait synthesis. An extension of it is 3DPortraitGAN [67] focusing on all-around upper body generation. As this work focuses only on full-head generation, Panohead can effectively be used to acquire synthetic portrait datasets." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14347" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 71, + 547, + 240 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 547, + 240 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 547, + 240 + ], + "type": "image", + "image_path": "0a9dc190670b992009c40dc82a72a6923cf4261a17c3c3dfe53def1d75fb96fd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 247, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 54, + 247, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 54, + 247, + 555, + 293 + ], + "type": "text", + "content": "Figure 2. Overview of SpinMeRound: Starting with a number of input conditioning views, the identity embedding " + }, + { + "bbox": [ + 54, + 247, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 54, + 247, + 555, + 293 + ], + "type": "text", + "content": " is extracted via a Face Recognition network (ArcFace [13]). Both the conditioning and target views are then encoded and combined with corresponding ray coordinate maps that represent camera poses. After the sampling step, our method synthesizes photorealistic images from novel angles, along with their associated shape normals " + }, + { + "bbox": [ + 54, + 247, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 54, + 247, + 555, + 293 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 301, + 160, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 160, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 160, + 314 + ], + "type": "text", + "content": "2.2. Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 320, + 297, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 320, + 297, + 525 + ], + "spans": [ + { + "bbox": [ + 54, + 320, + 297, + 525 + ], + "type": "text", + "content": "Face Generation Diffusion models (DMs) [28] have proven their generative abilities by beating the well-established GANs in image synthesis tasks [16]. The availability of large-scale datasets has revolutionized text-to-image generation [56] and video generation [5, 6] tasks. For face synthesis, a range of methods have emerged to tackle essential tasks, including 3D avatar creation [65, 70], avatar reenactment [14, 15, 33] and texture map generation [19, 46]. We adopt the concatenation strategy outlined in [19, 46] to generate novel viewpoint images and their corresponding shape normals simultaneously. Based on Stable Diffusion [56], Arc2Face [13] and InstantID [64] generate facial images given an input subject. Notably, Arc2Face exhibits strong generalization abilities in facial image generation, leveraging an up-sampled subset of WebFace42M [74]. Also, it introduces a robust identity conditioning mechanism, which is integrated into our method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 533, + 298, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 533, + 298, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 533, + 298, + 715 + ], + "type": "text", + "content": "Novel View Image Synthesis DreamFusion [49] introduces the use of Score Distillation Sampling (SDS), incorporating a pre-trained text-to-image diffusion model [56] alongside a NeRF model, to synthesize 3D objects from text prompts. In this way, its authors proved that they can efficiently generate 3D objects despite the lack of large-scale datasets; by exploiting the generalization abilities of an image generation network. Although subsequent studies focus on better distillation strategies [51, 53, 73], the aforementioned approaches are time-consuming and require complex balancing and additional constraints [20]. Closer to our work, ID-to-3D [3] and Arc2Avatar [22] use score distillation to generate 3D facial geometry, lacking though photorealistic results. Zero123 [39] introduces a multiview diffusion model conditioned on a reference image" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 302, + 556, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 302, + 556, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 556, + 579 + ], + "type": "text", + "content": "and a camera pose. Following it, methods such as One-2-3-45 [38], SyncDreamer [40], Consistent1-to-3 [66] and Cascade-Zero123 [68] further focus on multi-view consistency by introducing priors during the denoising process. Studies such as Zero " + }, + { + "bbox": [ + 313, + 302, + 556, + 579 + ], + "type": "inline_equation", + "content": "123++" + }, + { + "bbox": [ + 313, + 302, + 556, + 579 + ], + "type": "text", + "content": " [59], Era3D [36] and Morphable Diffusion [10] generate fixed viewpoints given the input image, without being able to handle arbitrary views. In Cat3D [20], a general multi-view diffusion model is introduced, using a powerful camera pose conditioning mechanism while using 3D attention layers. Although it has been a robust method for novel view synthesis, it does not focus on facial novel view synthesis, lacks shape normal generation capabilities and is a closed source framework. Diff-Portrait3D [25] showcases novel view capabilities, given an input facial image. Unlike our method, which employs image-free camera conditioning, DiffPortrait3D employs an image-driven approach for the desired output viewpoints and can generate only near-front views. More recent works [50, 61] have explored diffusion-based frameworks for generating controllable and animatable avatars. Concurrently to our approach, Pippo [29] and DiffPortrait360 [26] demonstrated " + }, + { + "bbox": [ + 313, + 302, + 556, + 579 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 302, + 556, + 579 + ], + "type": "text", + "content": " avatar generation, without, though, being able to sample novel identities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "content": "Novel View Video Synthesis Recently, video generation models [5, 6, 24] have proven their ability to generate photorealistic models. Methods such as IM-3D [42], V3D [11] and SV3D [63] integrate an off-the-shelf video diffusion model for generating novel viewpoints of a reference object. Still, the video generation step makes these methodologies computationally demanding [20]. Additionally, they are often restricted by the requirement for specific camera trajectories, typically revolving around a central subject. In contrast, our approach focuses on implementing multi-view diffusion networks that handle unordered camera poses." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14348" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 555, + 177 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 555, + 177 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 555, + 177 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 555, + 177 + ], + "type": "image", + "image_path": "a6f86be2864c23ba016d0af10dc59e82a18bd67199a85d0ba5504c46c741d04b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 183, + 555, + 207 + ], + "lines": [ + { + "bbox": [ + 54, + 183, + 555, + 207 + ], + "spans": [ + { + "bbox": [ + 54, + 183, + 555, + 207 + ], + "type": "text", + "content": "Figure 3. Samples generated by SpinMeRound on \"in-the-wild\" images: Given only the input images (shown in a small box at the top center), our method produces high-fidelity images from novel angles, along with corresponding shape normals " + }, + { + "bbox": [ + 54, + 183, + 555, + 207 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 54, + 183, + 555, + 207 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 217, + 111, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 111, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 111, + 228 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 237, + 296, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 237, + 296, + 321 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 296, + 321 + ], + "type": "text", + "content": "SpinMeRound incorporates a multi-view diffusion model, trained on a 3D facial dataset, capable of generating a full head portrait of an input subject, given their identity embedding and a number of views (Fig. 2). In the following sections, we first describe the proposed diffusion model, its training scheme, and the proposed sampling strategy given a single input view." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 332, + 211, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 332, + 211, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 332, + 211, + 344 + ], + "type": "text", + "content": "3.1. Multi-View Diffusion Model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "type": "text", + "content": "SpinMeRound employs a latent multi-view diffusion model alongside a powerful identity conditioning mechanism [47], enabling the generation of novel views of an input scene containing a person. Our approach leverages as conditioning inputs " + }, + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "type": "inline_equation", + "content": "M \\in \\{1,3\\}" + }, + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "type": "text", + "content": " pairs of sparse views of a subject, their respective shape normals and their associated camera poses. Let " + }, + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "type": "text", + "content": " denote a picked conditioned view and " + }, + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{I}}_i" + }, + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "type": "text", + "content": " a cropped and aligned version of " + }, + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "type": "text", + "content": ". We utilize a pre-trained face recognition network (ArcFace [13]) " + }, + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "type": "text", + "content": " to extract the identity embedding vector " + }, + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{w} = \\phi(\\bar{\\mathbf{I}}_i) \\in \\mathbb{R}^{512}" + }, + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "type": "text", + "content": " capable of incorporating crucial identity features. Then, we inject the identity information into the diffusion model as proposed in Arc2Face [13]. In particular, a text prompt of \"a photo of " + }, + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "type": "inline_equation", + "content": "" + }, + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "type": "text", + "content": " person, is fed to a CLIP text encoder [52] followed by the tokenization step. Followed by padding, the identity embedding vector " + }, + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "type": "text", + "content": " replaces the corresponding " + }, + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "type": "inline_equation", + "content": "" + }, + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "type": "text", + "content": " token resulting in " + }, + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{s} = \\{e_1,e_2,e_3,\\hat{\\mathbf{w}},e_5\\}" + }, + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{w}} \\in \\mathbb{R}^{768}" + }, + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "type": "text", + "content": " is the padded version of " + }, + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "type": "text", + "content": ". The vector " + }, + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "type": "text", + "content": " is fed to the encoder " + }, + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "type": "text", + "content": ", and finally we retrieve the corresponding conditioning vector " + }, + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{c} = C(\\mathbf{s}) \\in \\mathbb{R}^{N \\times 768}" + }, + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 350, + 296, + 601 + ], + "type": "text", + "content": " is the maximum sequence length." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "text", + "content": "The proposed latent multi-view UNet [57] is similar to the one introduced in Cat3D [20]. Our diffusion model is designed to concurrently generate " + }, + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "inline_equation", + "content": "P = (M + K) = 8" + }, + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "text", + "content": " pairs of facial images and their corresponding normals " + }, + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "text", + "content": ", given an input conditioning vector " + }, + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "text", + "content": ", a number of " + }, + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "text", + "content": " views and the " + }, + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "text", + "content": " camera poses. Let " + }, + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "inline_equation", + "content": "I_{i}^{cond}" + }, + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "text", + "content": " represent the input human face images, " + }, + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "inline_equation", + "content": "I_{j}^{tgt}" + }, + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "text", + "content": " the target face images from novel viewpoints, and " + }, + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{i}^{cond}, \\mathcal{N}_{j}^{tgt}" + }, + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "text", + "content": " their respective shape normals, where " + }, + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "inline_equation", + "content": "i \\in \\{1,2,\\dots,M\\}" + }, + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "inline_equation", + "content": "j \\in \\{1,2,\\dots,N\\}" + }, + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "text", + "content": " with" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 217, + 555, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 477 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 217, + 555, + 477 + ], + "type": "text", + "content": " being the maximum number of the input conditional views and " + }, + { + "bbox": [ + 313, + 217, + 555, + 477 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 217, + 555, + 477 + ], + "type": "text", + "content": " the number of target viewpoints. Following [19, 20, 46], we employ a pre-trained AutoEncoder of Stable Diffusion 1.5 (SD 1.5) [56], consisting of an encoder " + }, + { + "bbox": [ + 313, + 217, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 217, + 555, + 477 + ], + "type": "text", + "content": " and a decoder " + }, + { + "bbox": [ + 313, + 217, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 217, + 555, + 477 + ], + "type": "text", + "content": ". For each input conditioning image, we first obtain the corresponding latent feature maps: " + }, + { + "bbox": [ + 313, + 217, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{I_i^{cond}} = \\mathcal{E}(I_i^{cond}) \\in \\mathbb{R}^{4 \\times 64 \\times 64}" + }, + { + "bbox": [ + 313, + 217, + 555, + 477 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 217, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathcal{N}_i^{cond}} = \\mathcal{E}(\\mathcal{N}_i^{cond}) \\in \\mathbb{R}^{4 \\times 64 \\times 64}" + }, + { + "bbox": [ + 313, + 217, + 555, + 477 + ], + "type": "text", + "content": ". These are then concatenated channel-wise. Similarly, we apply the same procedure to extract the latent vectors for the target viewpoints. The corresponding camera pose information is incorporated for both conditioning and target viewpoints using the mechanism proposed in Cat3D [20]. For each of the " + }, + { + "bbox": [ + 313, + 217, + 555, + 477 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 217, + 555, + 477 + ], + "type": "text", + "content": " views, the latent feature maps " + }, + { + "bbox": [ + 313, + 217, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 313, + 217, + 555, + 477 + ], + "type": "text", + "content": " are concatenated channel-wise with the respective ray representation maps [20, 58] " + }, + { + "bbox": [ + 313, + 217, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_i^{cond}, \\mathbf{r}_j^{tgt} \\in \\mathbb{R}^{149 \\times 64 \\times 64}" + }, + { + "bbox": [ + 313, + 217, + 555, + 477 + ], + "type": "text", + "content": ", which encode the ray origin and direction. A binary mask " + }, + { + "bbox": [ + 313, + 217, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\mathbf{m} \\in \\{\\mathbf{0}, \\mathbf{1}\\}^{1 \\times 64 \\times 64}" + }, + { + "bbox": [ + 313, + 217, + 555, + 477 + ], + "type": "text", + "content": " is then appended to differentiate between conditional and target latent vectors. Finally, we retrieve the conditioning and target latent feature maps " + }, + { + "bbox": [ + 313, + 217, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{z}}_i^{cond} = \\{\\mathbf{z}_{I_i^{cond}}, \\mathbf{z}_{\\mathcal{N}_i^{cond}}\\mathbf{r}_j^{cond}, \\mathbf{m}_j^{cond}\\}" + }, + { + "bbox": [ + 313, + 217, + 555, + 477 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 217, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{z}}_i^{tgt} = \\{\\mathbf{z}_{I_i^{tgt}}, \\mathbf{z}_{\\mathcal{N}_i^{tgt}}\\mathbf{r}_j^{tgt}, \\mathbf{m}_j^{tgt}\\}" + }, + { + "bbox": [ + 313, + 217, + 555, + 477 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 478, + 556, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 556, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 556, + 575 + ], + "type": "text", + "content": "As proposed in video diffusion models [5], we initialize our method using a pre-trained LDM model (Arc2Face [47]) trained on large-scale datasets while adding additional attention layers, connecting the multiple latent feature maps. As in [20], we integrate 3D attention layers [60] by adding them between the original 2D self-attention layers of the LDM, and we fine-tune all layers of the multi-view UNet for improved multi-view consistency." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 586, + 411, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 411, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 411, + 599 + ], + "type": "text", + "content": "3.2. Training Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "We begin from the publicly available Arc2Face [47], and we adhere to the training schemes introduced in [5, 20]. Specifically, we adapt the EDM framework [30] to Arc2Face, training it for 31,000 iterations on its dedicated dataset. To accommodate the input latent feature maps, we expand the input and output convolutional layers channels, initializing these by copying the existing weights to the shape-normal channels while randomly initializing the camera-pose dimensions. As in [20], we shift the log-to-signal ratio by" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14349" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "inline_equation", + "content": "\\log (\\mathrm{N})" + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "inline_equation", + "content": "\\mathbf{N}" + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "content": " is the number of the target images " + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "inline_equation", + "content": "(\\mathrm{N} = 7)" + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "content": " We randomly select a conditioning view that includes part of the frontal face during each training iteration, as required for the identity embedding extraction. We then randomly pick the N target images and calculate the relative camera angles. To enhance the dataset, we replace the white background with a random color, in " + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "content": " of the samples. All the training viewpoints are encoded using the encoder " + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "content": " ,with noise added only to the target latent vectors, while conditioning vectors remain unchanged. Following the classifier-free guidance (CFG) training scheme [27], with a probability of " + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{uncond} = 0.15" + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "content": " , we randomly replace the identity vector with the empty string, and the conditioning images with zero-ed ones. We first train the model conditioned on a single view for " + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "inline_equation", + "content": "600\\mathrm{k}" + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "content": " iterations. Then, for the additional 1M iterations, we vary the conditioning views by randomly choosing O, 1 or 3 conditioning views, corresponding to 8, 7 and 5 target views, each with a probability of " + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_=1 / 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 295, + 175, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 175, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 175, + 308 + ], + "type": "text", + "content": "3.3. Novel view sampling" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 312, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 295, + 456 + ], + "type": "text", + "content": "SpinMeRound synthesizes novel views for an input subject I, given a number of views. In this section, we introduce a robust sampling strategy designed to produce a large number of consistent views that comprehensively cover a full head, given only a single input image. Achieving consistency across viewpoints requires a carefully structured camera pose selection order. Therefore, we employ a three-step sampling process: a) aligning input views and extracting the corresponding shape normals, b) generating anchor images that provide complete coverage of the " + }, + { + "bbox": [ + 55, + 312, + 295, + 456 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 55, + 312, + 295, + 456 + ], + "type": "text", + "content": " human head and c) synthesizing intermediate views by leveraging both the input views and the closest anchor images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 472, + 295, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 674 + ], + "type": "text", + "content": "Alignment and Shape Normals generation Given the input image " + }, + { + "bbox": [ + 55, + 472, + 295, + 674 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 55, + 472, + 295, + 674 + ], + "type": "text", + "content": ", we extract its identity embedding " + }, + { + "bbox": [ + 55, + 472, + 295, + 674 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 55, + 472, + 295, + 674 + ], + "type": "text", + "content": " as described in Sec. 3.1. Then, we obtain " + }, + { + "bbox": [ + 55, + 472, + 295, + 674 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}" + }, + { + "bbox": [ + 55, + 472, + 295, + 674 + ], + "type": "text", + "content": ", a cropped and aligned version of " + }, + { + "bbox": [ + 55, + 472, + 295, + 674 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 55, + 472, + 295, + 674 + ], + "type": "text", + "content": ", using the alignment procedure presented in Panohead [1]. To generate the shape normals " + }, + { + "bbox": [ + 55, + 472, + 295, + 674 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 55, + 472, + 295, + 674 + ], + "type": "text", + "content": ", we treat this as an in-painting task, thus retrieving the normals through the conditional guidance sampling approach described in Relightify [46]. Specifically, using the aligned image " + }, + { + "bbox": [ + 55, + 472, + 295, + 674 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}" + }, + { + "bbox": [ + 55, + 472, + 295, + 674 + ], + "type": "text", + "content": " and its identity embedding vector " + }, + { + "bbox": [ + 55, + 472, + 295, + 674 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 55, + 472, + 295, + 674 + ], + "type": "text", + "content": ", we employ a binary visibility mask " + }, + { + "bbox": [ + 55, + 472, + 295, + 674 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 472, + 295, + 674 + ], + "type": "text", + "content": ", marking only the image channels as visible and setting the shape normal channels as nonvisible. By applying the \"channel-wise in-painting\" algorithm, the corresponding shape normals are retrieved. This process uses the EDM sampler presented in [30] alongside the DDPM [28] discretization steps and runs for 50 steps. A more detailed presentation of this approach is presented in the supplemental material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "Generating anchor and intermediate views SpinMeRound can generate any arbitrary viewpoint, given the input" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 324 + ], + "type": "text", + "content": "aligned facial image " + }, + { + "bbox": [ + 313, + 72, + 553, + 324 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 324 + ], + "type": "text", + "content": " and the corresponding shape normals " + }, + { + "bbox": [ + 313, + 72, + 553, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 313, + 72, + 553, + 324 + ], + "type": "text", + "content": ". However, since it was trained to generate only a limited number of views per sampling process, a carefully designed sampling strategy is essential to produce a wide range of output views. Since the target subject is centered in the scene, we first generate " + }, + { + "bbox": [ + 313, + 72, + 553, + 324 + ], + "type": "inline_equation", + "content": "M = 7" + }, + { + "bbox": [ + 313, + 72, + 553, + 324 + ], + "type": "text", + "content": " anchor images " + }, + { + "bbox": [ + 313, + 72, + 553, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i" + }, + { + "bbox": [ + 313, + 72, + 553, + 324 + ], + "type": "text", + "content": " and corresponding anchor shape normals " + }, + { + "bbox": [ + 313, + 72, + 553, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{\\mathcal{N}_i}" + }, + { + "bbox": [ + 313, + 72, + 553, + 324 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 553, + 324 + ], + "type": "inline_equation", + "content": "i \\in \\{1, \\dots, M\\}" + }, + { + "bbox": [ + 313, + 72, + 553, + 324 + ], + "type": "text", + "content": ", covering a " + }, + { + "bbox": [ + 313, + 72, + 553, + 324 + ], + "type": "inline_equation", + "content": "360^\\circ" + }, + { + "bbox": [ + 313, + 72, + 553, + 324 + ], + "type": "text", + "content": " angle range of the subject " + }, + { + "bbox": [ + 313, + 72, + 553, + 324 + ], + "type": "inline_equation", + "content": "(\\pm 45^\\circ, \\pm 90^\\circ, \\pm 135^\\circ, 180^\\circ)" + }, + { + "bbox": [ + 313, + 72, + 553, + 324 + ], + "type": "text", + "content": ", as proposed in [20]. Using these anchors, any number of intermediate views can then be synthesized by conditioning on image triplets " + }, + { + "bbox": [ + 313, + 72, + 553, + 324 + ], + "type": "inline_equation", + "content": "\\{(\\mathbf{I}, \\mathcal{N}), (\\mathbf{A}_k, \\mathbf{A}_{\\mathcal{N}_k}), (\\mathbf{A}_l, \\mathbf{A}_{\\mathcal{N}_l})\\}" + }, + { + "bbox": [ + 313, + 72, + 553, + 324 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 553, + 324 + ], + "type": "inline_equation", + "content": "k, l" + }, + { + "bbox": [ + 313, + 72, + 553, + 324 + ], + "type": "text", + "content": " where represent the closest anchor images. This approach ensures that each intermediate view remains consistent with both the input aligned image " + }, + { + "bbox": [ + 313, + 72, + 553, + 324 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 324 + ], + "type": "text", + "content": ", and the closest already generated views. For both sampling processes, we use the EDM sampler facilitated by the EDM discretization steps [30], while it runs for 50 steps with a guidance scale set to 3. This method enables the generation of 48, 88, or more novel views for a single input subject " + }, + { + "bbox": [ + 313, + 72, + 553, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 313, + 72, + 553, + 324 + ], + "type": "text", + "content": ", depending on the chosen angle step, thus covering the entire scene." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 335, + 395, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 395, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 395, + 348 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 355, + 416, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 416, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 416, + 368 + ], + "type": "text", + "content": "4.1. Training Dataset" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 373, + 553, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 553, + 648 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 553, + 648 + ], + "type": "text", + "content": "Training such a model requires a large multi-view dataset containing a great number of subjects " + }, + { + "bbox": [ + 313, + 373, + 553, + 648 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 373, + 553, + 648 + ], + "type": "text", + "content": ". For each person " + }, + { + "bbox": [ + 313, + 373, + 553, + 648 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 313, + 373, + 553, + 648 + ], + "type": "text", + "content": ", it is necessary to acquire a set of images " + }, + { + "bbox": [ + 313, + 373, + 553, + 648 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_k^i" + }, + { + "bbox": [ + 313, + 373, + 553, + 648 + ], + "type": "text", + "content": ", their corresponding shape normal maps " + }, + { + "bbox": [ + 313, + 373, + 553, + 648 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_k^i" + }, + { + "bbox": [ + 313, + 373, + 553, + 648 + ], + "type": "text", + "content": ", camera poses " + }, + { + "bbox": [ + 313, + 373, + 553, + 648 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_k^i" + }, + { + "bbox": [ + 313, + 373, + 553, + 648 + ], + "type": "text", + "content": " alongside with their corresponding identity embedding vector " + }, + { + "bbox": [ + 313, + 373, + 553, + 648 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^i" + }, + { + "bbox": [ + 313, + 373, + 553, + 648 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 373, + 553, + 648 + ], + "type": "inline_equation", + "content": "k \\in \\{1, 2, \\dots, N_i\\}" + }, + { + "bbox": [ + 313, + 373, + 553, + 648 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 373, + 553, + 648 + ], + "type": "inline_equation", + "content": "N_{i}" + }, + { + "bbox": [ + 313, + 373, + 553, + 648 + ], + "type": "text", + "content": " is the number of the available views for the " + }, + { + "bbox": [ + 313, + 373, + 553, + 648 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 373, + 553, + 648 + ], + "type": "text", + "content": "-th scene. Due to the lack of such a dataset, we create a large-scale synthetic dataset using the publicly available Panohead [1]. We first sample " + }, + { + "bbox": [ + 313, + 373, + 553, + 648 + ], + "type": "inline_equation", + "content": "\\sim 10k" + }, + { + "bbox": [ + 313, + 373, + 553, + 648 + ], + "type": "text", + "content": " subjects and manually remove instances with artifacts in the back of the head, resulting in " + }, + { + "bbox": [ + 313, + 373, + 553, + 648 + ], + "type": "inline_equation", + "content": "\\sim 7k" + }, + { + "bbox": [ + 313, + 373, + 553, + 648 + ], + "type": "text", + "content": " distinct identities. We render images from 125 different viewpoints for each person to cover the entire head. Simultaneously, we obtain each subject's facial shape by extracting their opacity values from the triplane feature maps and then applying the marching cubes [41] algorithm. We render their respective normal maps using the acquired facial shape using Pytorch3d [54]. All the images depicting a frontal head are fed to the ArcFace [13] network to extract their corresponding identity vectors " + }, + { + "bbox": [ + 313, + 373, + 553, + 648 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 313, + 373, + 553, + 648 + ], + "type": "text", + "content": ". All in all, we end up with a synthetic dataset containing about 7k individuals, rendered from " + }, + { + "bbox": [ + 313, + 373, + 553, + 648 + ], + "type": "inline_equation", + "content": "N = 125" + }, + { + "bbox": [ + 313, + 373, + 553, + 648 + ], + "type": "text", + "content": " different angles, the respective shape normals, camera poses, and their identity embedding vector " + }, + { + "bbox": [ + 313, + 373, + 553, + 648 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 313, + 373, + 553, + 648 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 656, + 454, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 656, + 454, + 669 + ], + "spans": [ + { + "bbox": [ + 313, + 656, + 454, + 669 + ], + "type": "text", + "content": "4.2. Qualitative Comparisons" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 673, + 489, + 686 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 673, + 489, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 673, + 489, + 686 + ], + "type": "text", + "content": "4.2.1. Novel View Synthesis comparisons" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "type": "text", + "content": "In this section, we present a qualitative comparison between our model and other state-of-the-art multi-view dif" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14350" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 68, + 97, + 110 + ], + "blocks": [ + { + "bbox": [ + 58, + 68, + 97, + 110 + ], + "lines": [ + { + "bbox": [ + 58, + 68, + 97, + 110 + ], + "spans": [ + { + "bbox": [ + 58, + 68, + 97, + 110 + ], + "type": "image", + "image_path": "d818a6599f3e1d45778cf0e9495b29eab1d9c8dfca098286f33807a469539424.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 60, + 110, + 94, + 148 + ], + "blocks": [ + { + "bbox": [ + 60, + 110, + 94, + 148 + ], + "lines": [ + { + "bbox": [ + 60, + 110, + 94, + 148 + ], + "spans": [ + { + "bbox": [ + 60, + 110, + 94, + 148 + ], + "type": "image", + "image_path": "ae710b886140c2ad071d36e266b11994a06409dae5e52008c7921e9150a8aa6e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 101, + 69, + 133, + 110 + ], + "blocks": [ + { + "bbox": [ + 101, + 69, + 133, + 110 + ], + "lines": [ + { + "bbox": [ + 101, + 69, + 133, + 110 + ], + "spans": [ + { + "bbox": [ + 101, + 69, + 133, + 110 + ], + "type": "image", + "image_path": "2ff3e69530abb9e7d0fcd9766d9e34fa4b7973ec676a374f8b63e2dd22a60728.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 101, + 110, + 134, + 148 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 110, + 134, + 148 + ], + "spans": [ + { + "bbox": [ + 101, + 110, + 134, + 148 + ], + "type": "image", + "image_path": "2d59d04a8782fe73dae75a8876e0423b868bbb8cdd332b35c2c5fd71fbb7dbe8.jpg" + } + ] + } + ], + "index": 3, + "type": "text" + }, + { + "bbox": [ + 101, + 149, + 134, + 188 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 149, + 134, + 188 + ], + "spans": [ + { + "bbox": [ + 101, + 149, + 134, + 188 + ], + "type": "image", + "image_path": "8f4fd68612c9acd4395000061a42ec028f12eda18ae84c23263b4fa22bb1e65b.jpg" + } + ] + } + ], + "index": 4, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 58, + 188, + 94, + 232 + ], + "blocks": [ + { + "bbox": [ + 58, + 188, + 94, + 232 + ], + "lines": [ + { + "bbox": [ + 58, + 188, + 94, + 232 + ], + "spans": [ + { + "bbox": [ + 58, + 188, + 94, + 232 + ], + "type": "image", + "image_path": "d01daae13de9efd69a286b34d0b557262e185280f2c2f745c7cd34482464b632.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 238, + 96, + 247 + ], + "lines": [ + { + "bbox": [ + 69, + 238, + 96, + 247 + ], + "spans": [ + { + "bbox": [ + 69, + 238, + 96, + 247 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 101, + 110, + 134, + 148 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 110, + 134, + 148 + ], + "spans": [ + { + "bbox": [ + 101, + 110, + 134, + 148 + ], + "type": "image", + "image_path": "2d59d04a8782fe73dae75a8876e0423b868bbb8cdd332b35c2c5fd71fbb7dbe8.jpg" + } + ] + } + ], + "index": 7, + "type": "text" + }, + { + "bbox": [ + 101, + 149, + 134, + 188 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 149, + 134, + 188 + ], + "spans": [ + { + "bbox": [ + 101, + 149, + 134, + 188 + ], + "type": "image", + "image_path": "8f4fd68612c9acd4395000061a42ec028f12eda18ae84c23263b4fa22bb1e65b.jpg" + } + ] + } + ], + "index": 8, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 101, + 188, + 134, + 232 + ], + "blocks": [ + { + "bbox": [ + 101, + 188, + 134, + 232 + ], + "lines": [ + { + "bbox": [ + 101, + 188, + 134, + 232 + ], + "spans": [ + { + "bbox": [ + 101, + 188, + 134, + 232 + ], + "type": "image", + "image_path": "069db5a8fefd94e18eb7dbf550ca2addf36946ed8b3448354e0e854e76ddb5de.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 238, + 159, + 247 + ], + "lines": [ + { + "bbox": [ + 130, + 238, + 159, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 238, + 159, + 247 + ], + "type": "text", + "content": "(b) SpinM" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 135, + 69, + 167, + 110 + ], + "blocks": [ + { + "bbox": [ + 135, + 69, + 167, + 110 + ], + "lines": [ + { + "bbox": [ + 135, + 69, + 167, + 110 + ], + "spans": [ + { + "bbox": [ + 135, + 69, + 167, + 110 + ], + "type": "image", + "image_path": "7b34f54acfeb9aa01551968c720205a1307745f58f40ebabe0821aa8b5d946f1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 135, + 110, + 168, + 148 + ], + "blocks": [ + { + "bbox": [ + 135, + 110, + 168, + 148 + ], + "lines": [ + { + "bbox": [ + 135, + 110, + 168, + 148 + ], + "spans": [ + { + "bbox": [ + 135, + 110, + 168, + 148 + ], + "type": "image", + "image_path": "1f7b420fed12bfdff778f94326eee91c6161026a2b8c42194b83c09332b30732.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 135, + 149, + 168, + 188 + ], + "blocks": [ + { + "bbox": [ + 135, + 149, + 168, + 188 + ], + "lines": [ + { + "bbox": [ + 135, + 149, + 168, + 188 + ], + "spans": [ + { + "bbox": [ + 135, + 149, + 168, + 188 + ], + "type": "image", + "image_path": "f9455abeb235b7c6e00fea12cb091e21491ba3cbc2f1d88c025240e8d7e38691.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 135, + 188, + 168, + 232 + ], + "blocks": [ + { + "bbox": [ + 135, + 188, + 168, + 232 + ], + "lines": [ + { + "bbox": [ + 135, + 188, + 168, + 232 + ], + "spans": [ + { + "bbox": [ + 135, + 188, + 168, + 232 + ], + "type": "image", + "image_path": "24197544f59a456f4de064987c0dfaddf121f01757521ccd37f211c76f6af6c5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 255, + 555, + 289 + ], + "lines": [ + { + "bbox": [ + 55, + 255, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 555, + 289 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison between SpinMeRound, Zero123-XL [12], SV3D [63] and DiffPortrait3D [25] at angles " + }, + { + "bbox": [ + 55, + 255, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\{\\pm 45^{\\circ}, + 180^{\\circ}\\}" + }, + { + "bbox": [ + 55, + 255, + 555, + 289 + ], + "type": "text", + "content": ". It is clear that SpinMeRound effectively generates high-quality novel views from \"in-the-wild\" input images, whilst SV3D produces distorted outputs, Zero123-XL generates unnaturally squared avatars and DiffPortrait3D cannot handle such angles." + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 170, + 69, + 205, + 110 + ], + "blocks": [ + { + "bbox": [ + 170, + 69, + 205, + 110 + ], + "lines": [ + { + "bbox": [ + 170, + 69, + 205, + 110 + ], + "spans": [ + { + "bbox": [ + 170, + 69, + 205, + 110 + ], + "type": "image", + "image_path": "004d706813603a4f671352741f48bbe3ec5444279f64a124c91ff445b97ab705.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 170, + 110, + 205, + 148 + ], + "blocks": [ + { + "bbox": [ + 170, + 110, + 205, + 148 + ], + "lines": [ + { + "bbox": [ + 170, + 110, + 205, + 148 + ], + "spans": [ + { + "bbox": [ + 170, + 110, + 205, + 148 + ], + "type": "image", + "image_path": "cd9f16d8849df2ca6ddc258a76cdc428aeb5fc011b1eb3a121e2dd4c0c3c6427.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 170, + 149, + 205, + 188 + ], + "blocks": [ + { + "bbox": [ + 170, + 149, + 205, + 188 + ], + "lines": [ + { + "bbox": [ + 170, + 149, + 205, + 188 + ], + "spans": [ + { + "bbox": [ + 170, + 149, + 205, + 188 + ], + "type": "image", + "image_path": "8699bf09479e5df4a791f035e7e857f791d8893ccd5751aa00689989411b851e.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 170, + 188, + 205, + 232 + ], + "blocks": [ + { + "bbox": [ + 170, + 188, + 205, + 232 + ], + "lines": [ + { + "bbox": [ + 170, + 188, + 205, + 232 + ], + "spans": [ + { + "bbox": [ + 170, + 188, + 205, + 232 + ], + "type": "image", + "image_path": "04f10cdcd91e99a63f65548fe9a31d3f6e3b756baacde7ea48d1202118069fe6.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 214, + 69, + 245, + 110 + ], + "blocks": [ + { + "bbox": [ + 214, + 69, + 245, + 110 + ], + "lines": [ + { + "bbox": [ + 214, + 69, + 245, + 110 + ], + "spans": [ + { + "bbox": [ + 214, + 69, + 245, + 110 + ], + "type": "image", + "image_path": "f9ab34fc4228f91e4b35cc0a9b2c87af9b779635059ffc15675ad92b46121c58.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 214, + 110, + 246, + 148 + ], + "blocks": [ + { + "bbox": [ + 214, + 110, + 246, + 148 + ], + "lines": [ + { + "bbox": [ + 214, + 110, + 246, + 148 + ], + "spans": [ + { + "bbox": [ + 214, + 110, + 246, + 148 + ], + "type": "image", + "image_path": "7a34fb8a25f158da6ed290436f6758510e6f5e6768135c374d15293a2b891898.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 211, + 149, + 242, + 188 + ], + "blocks": [ + { + "bbox": [ + 211, + 149, + 242, + 188 + ], + "lines": [ + { + "bbox": [ + 211, + 149, + 242, + 188 + ], + "spans": [ + { + "bbox": [ + 211, + 149, + 242, + 188 + ], + "type": "image", + "image_path": "ffaefcfa111f2fd13b2308ada2a267990fb6bef9541358b42d8d263c3736e63a.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 211, + 188, + 247, + 224 + ], + "blocks": [ + { + "bbox": [ + 211, + 188, + 247, + 224 + ], + "lines": [ + { + "bbox": [ + 211, + 188, + 247, + 224 + ], + "spans": [ + { + "bbox": [ + 211, + 188, + 247, + 224 + ], + "type": "image", + "image_path": "05377c43ce4bfabeaf86a483c9d0f0332b5d49d9fadc46d525aa82a469e43bbc.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 224, + 227, + 234, + 232 + ], + "lines": [ + { + "bbox": [ + 224, + 227, + 234, + 232 + ], + "spans": [ + { + "bbox": [ + 224, + 227, + 234, + 232 + ], + "type": "text", + "content": "+45°" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 252, + 69, + 285, + 110 + ], + "blocks": [ + { + "bbox": [ + 252, + 69, + 285, + 110 + ], + "lines": [ + { + "bbox": [ + 252, + 69, + 285, + 110 + ], + "spans": [ + { + "bbox": [ + 252, + 69, + 285, + 110 + ], + "type": "image", + "image_path": "8f0dc5eebd0e6547c61339c20816af7ddec62b9ba2dbf5ee9e68fe5efabd9b19.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 252, + 110, + 285, + 148 + ], + "blocks": [ + { + "bbox": [ + 252, + 110, + 285, + 148 + ], + "lines": [ + { + "bbox": [ + 252, + 110, + 285, + 148 + ], + "spans": [ + { + "bbox": [ + 252, + 110, + 285, + 148 + ], + "type": "image", + "image_path": "5992d458f8c2f1412672ed6d1860ff4feb1c114500245b79cc729b971ea6bbb3.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 252, + 149, + 285, + 188 + ], + "blocks": [ + { + "bbox": [ + 252, + 149, + 285, + 188 + ], + "lines": [ + { + "bbox": [ + 252, + 149, + 285, + 188 + ], + "spans": [ + { + "bbox": [ + 252, + 149, + 285, + 188 + ], + "type": "image", + "image_path": "646d6cea1eedbcf16f53c429e2b0c48ba73f8579eb752e7e74d89d720426a0a4.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 251, + 188, + 285, + 232 + ], + "blocks": [ + { + "bbox": [ + 251, + 188, + 285, + 232 + ], + "lines": [ + { + "bbox": [ + 251, + 188, + 285, + 232 + ], + "spans": [ + { + "bbox": [ + 251, + 188, + 285, + 232 + ], + "type": "image", + "image_path": "d68efbfa6f6e8011aeb303eb29e74fcc41597980a617f734d7f2638bfb7225f3.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 238, + 296, + 247 + ], + "lines": [ + { + "bbox": [ + 255, + 238, + 296, + 247 + ], + "spans": [ + { + "bbox": [ + 255, + 238, + 296, + 247 + ], + "type": "text", + "content": "(c) SV3D [63]" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 289, + 69, + 321, + 110 + ], + "blocks": [ + { + "bbox": [ + 289, + 69, + 321, + 110 + ], + "lines": [ + { + "bbox": [ + 289, + 69, + 321, + 110 + ], + "spans": [ + { + "bbox": [ + 289, + 69, + 321, + 110 + ], + "type": "image", + "image_path": "3d4438f889f9eb3241095eee86aac301dbd4393baa704d7d674e803200f893aa.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 288, + 110, + 321, + 148 + ], + "blocks": [ + { + "bbox": [ + 288, + 110, + 321, + 148 + ], + "lines": [ + { + "bbox": [ + 288, + 110, + 321, + 148 + ], + "spans": [ + { + "bbox": [ + 288, + 110, + 321, + 148 + ], + "type": "image", + "image_path": "bfc6899dae281503a0e41446a938f0cfcb7370cbfd45cf8c53f3111bc0f8879b.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 295, + 150, + 320, + 188 + ], + "blocks": [ + { + "bbox": [ + 295, + 150, + 320, + 188 + ], + "lines": [ + { + "bbox": [ + 295, + 150, + 320, + 188 + ], + "spans": [ + { + "bbox": [ + 295, + 150, + 320, + 188 + ], + "type": "image", + "image_path": "9895b41b7c14eec2b22cc2fdbc3cc210f81a40fba57e442c68c56c371166da1b.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 291, + 189, + 321, + 232 + ], + "blocks": [ + { + "bbox": [ + 291, + 189, + 321, + 232 + ], + "lines": [ + { + "bbox": [ + 291, + 189, + 321, + 232 + ], + "spans": [ + { + "bbox": [ + 291, + 189, + 321, + 232 + ], + "type": "image", + "image_path": "b7289fa4afeb371cd90ea21dfc8336a3553c566058bbfa657ce58ed3a313ca9d.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 331, + 73, + 358, + 109 + ], + "blocks": [ + { + "bbox": [ + 331, + 73, + 358, + 109 + ], + "lines": [ + { + "bbox": [ + 331, + 73, + 358, + 109 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 358, + 109 + ], + "type": "image", + "image_path": "dddd58d01b24ef098c74875ec4b0454845ef662024e5cd971845db10250aabb6.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 332, + 113, + 356, + 148 + ], + "blocks": [ + { + "bbox": [ + 332, + 113, + 356, + 148 + ], + "lines": [ + { + "bbox": [ + 332, + 113, + 356, + 148 + ], + "spans": [ + { + "bbox": [ + 332, + 113, + 356, + 148 + ], + "type": "image", + "image_path": "bfd5f75cfee0222a914e12338e13ad309bc70815f18b6c2b2b152efbb3063d55.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 329, + 152, + 355, + 186 + ], + "blocks": [ + { + "bbox": [ + 329, + 152, + 355, + 186 + ], + "lines": [ + { + "bbox": [ + 329, + 152, + 355, + 186 + ], + "spans": [ + { + "bbox": [ + 329, + 152, + 355, + 186 + ], + "type": "image", + "image_path": "fe47f4ec4edcc7094c4f8f417f1a675539611e03bf371db66bb55fc2cae8f3e4.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 329, + 191, + 359, + 232 + ], + "blocks": [ + { + "bbox": [ + 329, + 191, + 359, + 232 + ], + "lines": [ + { + "bbox": [ + 329, + 191, + 359, + 232 + ], + "spans": [ + { + "bbox": [ + 329, + 191, + 359, + 232 + ], + "type": "image", + "image_path": "be5f98809405363546d5a098013b232d40c8bcb04f60cd145987d70dc1f89cbf.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 238, + 414, + 247 + ], + "lines": [ + { + "bbox": [ + 354, + 238, + 414, + 247 + ], + "spans": [ + { + "bbox": [ + 354, + 238, + 414, + 247 + ], + "type": "text", + "content": "(d) Zero123-XL [12]" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 354, + 73, + 400, + 109 + ], + "blocks": [ + { + "bbox": [ + 354, + 73, + 400, + 109 + ], + "lines": [ + { + "bbox": [ + 354, + 73, + 400, + 109 + ], + "spans": [ + { + "bbox": [ + 354, + 73, + 400, + 109 + ], + "type": "image", + "image_path": "3333863872599a0617ff4967ca502cc4c2b16de77ff3c3eb5e2057fa876d81e9.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 354, + 114, + 395, + 148 + ], + "blocks": [ + { + "bbox": [ + 354, + 114, + 395, + 148 + ], + "lines": [ + { + "bbox": [ + 354, + 114, + 395, + 148 + ], + "spans": [ + { + "bbox": [ + 354, + 114, + 395, + 148 + ], + "type": "image", + "image_path": "2104a4404a3747919545ce7b0871b7fb004df43e93f997be76e23cafb7171214.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 369, + 152, + 394, + 185 + ], + "blocks": [ + { + "bbox": [ + 369, + 152, + 394, + 185 + ], + "lines": [ + { + "bbox": [ + 369, + 152, + 394, + 185 + ], + "spans": [ + { + "bbox": [ + 369, + 152, + 394, + 185 + ], + "type": "image", + "image_path": "81e812aa84697c619614684bc481917d3f86153a894794ab2374ad2dfea9301e.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 364, + 191, + 400, + 232 + ], + "blocks": [ + { + "bbox": [ + 364, + 191, + 400, + 232 + ], + "lines": [ + { + "bbox": [ + 364, + 191, + 400, + 232 + ], + "spans": [ + { + "bbox": [ + 364, + 191, + 400, + 232 + ], + "type": "image", + "image_path": "1310ecbf5fb84e17db969ff090be4d6d83676c7c78c714b2f91fdebac68b0402.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 228, + 425, + 232 + ], + "lines": [ + { + "bbox": [ + 410, + 228, + 425, + 232 + ], + "spans": [ + { + "bbox": [ + 410, + 228, + 425, + 232 + ], + "type": "text", + "content": "-45°" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 403, + 73, + 431, + 109 + ], + "blocks": [ + { + "bbox": [ + 403, + 73, + 431, + 109 + ], + "lines": [ + { + "bbox": [ + 403, + 73, + 431, + 109 + ], + "spans": [ + { + "bbox": [ + 403, + 73, + 431, + 109 + ], + "type": "image", + "image_path": "8099ab0920fcc49325ca7139a3f2fbc2da4d68e5cf1e6484a0af1915c6bbc485.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 405, + 111, + 435, + 148 + ], + "blocks": [ + { + "bbox": [ + 405, + 111, + 435, + 148 + ], + "lines": [ + { + "bbox": [ + 405, + 111, + 435, + 148 + ], + "spans": [ + { + "bbox": [ + 405, + 111, + 435, + 148 + ], + "type": "image", + "image_path": "742f95c7c0bb80ab11f81dffbe527c885003f2799149aec4e96049cb9f188fc7.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 409, + 152, + 432, + 186 + ], + "blocks": [ + { + "bbox": [ + 409, + 152, + 432, + 186 + ], + "lines": [ + { + "bbox": [ + 409, + 152, + 432, + 186 + ], + "spans": [ + { + "bbox": [ + 409, + 152, + 432, + 186 + ], + "type": "image", + "image_path": "6e16733e0447cc8a3d699a661d34e500328b91624f684fc9c3bb6306fad732d0.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 408, + 188, + 431, + 224 + ], + "blocks": [ + { + "bbox": [ + 408, + 188, + 431, + 224 + ], + "lines": [ + { + "bbox": [ + 408, + 188, + 431, + 224 + ], + "spans": [ + { + "bbox": [ + 408, + 188, + 431, + 224 + ], + "type": "image", + "image_path": "81591603fcab7abd2683843363ecdcf0a7b7804876d59bd00345bf6511d77c13.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 438, + 69, + 515, + 109 + ], + "blocks": [ + { + "bbox": [ + 438, + 69, + 515, + 109 + ], + "lines": [ + { + "bbox": [ + 438, + 69, + 515, + 109 + ], + "spans": [ + { + "bbox": [ + 438, + 69, + 515, + 109 + ], + "type": "image", + "image_path": "e92c697f0b1d5b91a372443cdae76534faf8521cf4c86a3baec62704ab63154e.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 438, + 110, + 515, + 148 + ], + "blocks": [ + { + "bbox": [ + 438, + 110, + 515, + 148 + ], + "lines": [ + { + "bbox": [ + 438, + 110, + 515, + 148 + ], + "spans": [ + { + "bbox": [ + 438, + 110, + 515, + 148 + ], + "type": "image", + "image_path": "5855816e181b8d0732336cdb3e4f3c34441a0726f086c340fb961a33c26789b7.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 438, + 149, + 515, + 188 + ], + "blocks": [ + { + "bbox": [ + 438, + 149, + 515, + 188 + ], + "lines": [ + { + "bbox": [ + 438, + 149, + 515, + 188 + ], + "spans": [ + { + "bbox": [ + 438, + 149, + 515, + 188 + ], + "type": "image", + "image_path": "7223f9288ef5a6f235757d5b3d3e4ed5e821e1ae664c82ce6270cbe4917dcae2.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 438, + 188, + 515, + 232 + ], + "blocks": [ + { + "bbox": [ + 438, + 188, + 515, + 232 + ], + "lines": [ + { + "bbox": [ + 438, + 188, + 515, + 232 + ], + "spans": [ + { + "bbox": [ + 438, + 188, + 515, + 232 + ], + "type": "image", + "image_path": "bc1fd73df9540eeef1b69d20603030eca47620f2edeb6b0d2bbc125b6861268c.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 461, + 238, + 525, + 247 + ], + "lines": [ + { + "bbox": [ + 461, + 238, + 525, + 247 + ], + "spans": [ + { + "bbox": [ + 461, + 238, + 525, + 247 + ], + "type": "text", + "content": "(d) DiffPortrait3d [25]" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 515, + 69, + 553, + 109 + ], + "blocks": [ + { + "bbox": [ + 515, + 69, + 553, + 109 + ], + "lines": [ + { + "bbox": [ + 515, + 69, + 553, + 109 + ], + "spans": [ + { + "bbox": [ + 515, + 69, + 553, + 109 + ], + "type": "image", + "image_path": "50c2ad859c6fbe0ad39011f4255a8d2c4bd1d5203b900d112a38c7e7cfa92ccf.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 514, + 110, + 553, + 148 + ], + "blocks": [ + { + "bbox": [ + 514, + 110, + 553, + 148 + ], + "lines": [ + { + "bbox": [ + 514, + 110, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 514, + 110, + 553, + 148 + ], + "type": "image", + "image_path": "62ea2a47c4edb1b9de90a732fb9a30264ebbd33d5adaf490269b50c4749e42f9.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 515, + 149, + 553, + 188 + ], + "blocks": [ + { + "bbox": [ + 515, + 149, + 553, + 188 + ], + "lines": [ + { + "bbox": [ + 515, + 149, + 553, + 188 + ], + "spans": [ + { + "bbox": [ + 515, + 149, + 553, + 188 + ], + "type": "image", + "image_path": "6cbbf5a71bfd5097b4d288f6aecccf580bfff7b2228045bf134855d72fed7f50.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 515, + 188, + 552, + 232 + ], + "blocks": [ + { + "bbox": [ + 515, + 188, + 552, + 232 + ], + "lines": [ + { + "bbox": [ + 515, + 188, + 552, + 232 + ], + "spans": [ + { + "bbox": [ + 515, + 188, + 552, + 232 + ], + "type": "image", + "image_path": "d02a7a584aa561371883b51a416a3a2810dd5f8cd4cb3caaf42f0d1ecd6df08e.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + } + ], + "index": 55 + }, + { + "bbox": [ + 55, + 300, + 296, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 502 + ], + "type": "text", + "content": "fusion models, SV3D [63] Zero123-XL [12] and DiffPort- trait3D [25], focusing on their output under " + }, + { + "bbox": [ + 55, + 300, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\pm 45^{\\circ}, + 180^{\\circ}" + }, + { + "bbox": [ + 55, + 300, + 296, + 502 + ], + "type": "text", + "content": " angles. SV3D, a latent video diffusion model, creates " + }, + { + "bbox": [ + 55, + 300, + 296, + 502 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 55, + 300, + 296, + 502 + ], + "type": "text", + "content": " videos from a single input image. Zero123-XL, a multiview diffusion model, synthesizes novel perspectives based on the input image and specified camera poses. DiffPort- trait3D [25] generates novel views by conditioning the desired camera on an input image and refining the initial noise through a fine-tuning step to achieve optimal results. Figure 4 demonstrates scenarios where the input \"in-the-wild\" images (Fig. 4a ) are fed to these 4 different networks and their respective generations under " + }, + { + "bbox": [ + 55, + 300, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\pm 45^{\\circ}, + 180^{\\circ}" + }, + { + "bbox": [ + 55, + 300, + 296, + 502 + ], + "type": "text", + "content": " viewpoints. As illustrated, SpinMeRound efficiently generates accurate novel viewpoints of the input subject, while SV3D distorts the input image in its generated views, Zero123-XL produces unnatural, square-like facial avatars and DiffPort- trait3D cannot handle those angles at all." + } + ] + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 56, + 510, + 296, + 578 + ], + "blocks": [ + { + "bbox": [ + 56, + 510, + 296, + 578 + ], + "lines": [ + { + "bbox": [ + 56, + 510, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 296, + 578 + ], + "type": "image", + "image_path": "f3027b182cc163b281867aba12807a0605f15cfd46e26d26986f75e72c96a524.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 585, + 295, + 619 + ], + "lines": [ + { + "bbox": [ + 55, + 585, + 295, + 619 + ], + "spans": [ + { + "bbox": [ + 55, + 585, + 295, + 619 + ], + "type": "text", + "content": "Figure 5. Given the input identities (shown in the small boxes), we present the results of novel view synthesis after applying 3D Gaussian splatting [31] to the views generated by our model." + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_caption" + } + ], + "index": 58 + }, + { + "bbox": [ + 55, + 635, + 165, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 165, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 165, + 647 + ], + "type": "text", + "content": "4.3. 3D Reconstruction" + } + ] + } + ], + "index": 60 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "To evaluate our model's consistency in generating novel views from an input image, we assess its ability to reconstruct an input identity through Gaussian Splatting (3DGS) [31]. More specifically, given an input identity, we apply the sampling strategy outlined in Section 3.3 to" + } + ] + } + ], + "index": 61 + }, + { + "bbox": [ + 313, + 300, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 407 + ], + "type": "text", + "content": "generate 48 novel views. These views are then used to reconstruct the identity through Gaussian Splatting. As in [20], we modify the provided 3DGS code to incorporate the LPIPS [71] loss between the ground-truth and the generated viewpoints. In this way, we can deal with small inconsistencies between nearby viewpoints. We present the novel view synthesis of two input identities (small squares) in Figure 5, which clearly illustrates that our proposed methodology can generate consistent subjects." + } + ] + } + ], + "index": 62 + }, + { + "type": "image", + "bbox": [ + 316, + 417, + 375, + 478 + ], + "blocks": [ + { + "bbox": [ + 316, + 417, + 375, + 478 + ], + "lines": [ + { + "bbox": [ + 316, + 417, + 375, + 478 + ], + "spans": [ + { + "bbox": [ + 316, + 417, + 375, + 478 + ], + "type": "image", + "image_path": "83c2a33f3fe24124d532df0ebdc66ae96eacebbeebb01ace9391cda9100825ad.jpg" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_body" + } + ], + "index": 63 + }, + { + "type": "image", + "bbox": [ + 317, + 479, + 373, + 507 + ], + "blocks": [ + { + "bbox": [ + 317, + 479, + 373, + 507 + ], + "lines": [ + { + "bbox": [ + 317, + 479, + 373, + 507 + ], + "spans": [ + { + "bbox": [ + 317, + 479, + 373, + 507 + ], + "type": "image", + "image_path": "5f08c3a2b94be8fe7126488cecb580ad1b64f6df493cf2f251995ca1873a33b4.jpg" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 517, + 554, + 540 + ], + "lines": [ + { + "bbox": [ + 313, + 517, + 554, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 554, + 540 + ], + "type": "text", + "content": "Figure 6. Samples generated using unconditional sampling. Spin-MeRound can generate novel identities without any prior input." + } + ] + } + ], + "index": 77, + "angle": 0, + "type": "image_caption" + } + ], + "index": 64 + }, + { + "type": "image", + "bbox": [ + 378, + 418, + 405, + 449 + ], + "blocks": [ + { + "bbox": [ + 378, + 418, + 405, + 449 + ], + "lines": [ + { + "bbox": [ + 378, + 418, + 405, + 449 + ], + "spans": [ + { + "bbox": [ + 378, + 418, + 405, + 449 + ], + "type": "image", + "image_path": "f3a6a29d299eb3c8e5d3a7fa25c4338002994d8cc9ce0e91c639b3e3bcb0ec41.jpg" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_body" + } + ], + "index": 65 + }, + { + "type": "image", + "bbox": [ + 378, + 450, + 405, + 478 + ], + "blocks": [ + { + "bbox": [ + 378, + 450, + 405, + 478 + ], + "lines": [ + { + "bbox": [ + 378, + 450, + 405, + 478 + ], + "spans": [ + { + "bbox": [ + 378, + 450, + 405, + 478 + ], + "type": "image", + "image_path": "8a6e445ed53d713c46723fd5cb65b9b50c3c8f114be118ad62ae752690387680.jpg" + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "image_body" + } + ], + "index": 66 + }, + { + "type": "image", + "bbox": [ + 378, + 479, + 405, + 507 + ], + "blocks": [ + { + "bbox": [ + 378, + 479, + 405, + 507 + ], + "lines": [ + { + "bbox": [ + 378, + 479, + 405, + 507 + ], + "spans": [ + { + "bbox": [ + 378, + 479, + 405, + 507 + ], + "type": "image", + "image_path": "2c9149d69baed028d58782a1fd05123c48e03a0ad97689fbe3d1040e104d68a0.jpg" + } + ] + } + ], + "index": 67, + "angle": 0, + "type": "image_body" + } + ], + "index": 67 + }, + { + "type": "image", + "bbox": [ + 405, + 418, + 435, + 449 + ], + "blocks": [ + { + "bbox": [ + 405, + 418, + 435, + 449 + ], + "lines": [ + { + "bbox": [ + 405, + 418, + 435, + 449 + ], + "spans": [ + { + "bbox": [ + 405, + 418, + 435, + 449 + ], + "type": "image", + "image_path": "6a2c2201223e3b27f0ab06f9ef79c2dbd1480ef321a1eafbe54406c626568ee9.jpg" + } + ] + } + ], + "index": 68, + "angle": 0, + "type": "image_body" + } + ], + "index": 68 + }, + { + "type": "image", + "bbox": [ + 405, + 450, + 435, + 478 + ], + "blocks": [ + { + "bbox": [ + 405, + 450, + 435, + 478 + ], + "lines": [ + { + "bbox": [ + 405, + 450, + 435, + 478 + ], + "spans": [ + { + "bbox": [ + 405, + 450, + 435, + 478 + ], + "type": "image", + "image_path": "b02f7a889c81da213224af37d6daa0fad29a0ce0b8dc870cb89587f7f6ad77c8.jpg" + } + ] + } + ], + "index": 69, + "angle": 0, + "type": "image_body" + } + ], + "index": 69 + }, + { + "type": "image", + "bbox": [ + 405, + 479, + 435, + 507 + ], + "blocks": [ + { + "bbox": [ + 405, + 479, + 435, + 507 + ], + "lines": [ + { + "bbox": [ + 405, + 479, + 435, + 507 + ], + "spans": [ + { + "bbox": [ + 405, + 479, + 435, + 507 + ], + "type": "image", + "image_path": "be729dd9b6f84d5572990928dc65a6876a66b473a7578ad882ab7d5851eaeb88.jpg" + } + ] + } + ], + "index": 70, + "angle": 0, + "type": "image_body" + } + ], + "index": 70 + }, + { + "type": "image", + "bbox": [ + 437, + 418, + 515, + 449 + ], + "blocks": [ + { + "bbox": [ + 437, + 418, + 515, + 449 + ], + "lines": [ + { + "bbox": [ + 437, + 418, + 515, + 449 + ], + "spans": [ + { + "bbox": [ + 437, + 418, + 515, + 449 + ], + "type": "image", + "image_path": "302bffb2332ecc37d3a4884bfbbc6d8b3dbb4e6228e4ec66678f84e09878ce1f.jpg" + } + ] + } + ], + "index": 71, + "angle": 0, + "type": "image_body" + } + ], + "index": 71 + }, + { + "type": "image", + "bbox": [ + 437, + 449, + 515, + 478 + ], + "blocks": [ + { + "bbox": [ + 437, + 449, + 515, + 478 + ], + "lines": [ + { + "bbox": [ + 437, + 449, + 515, + 478 + ], + "spans": [ + { + "bbox": [ + 437, + 449, + 515, + 478 + ], + "type": "image", + "image_path": "1335d900a62c968999befd2ba1c646433ddfeb5cca524dc18712ccb23158c099.jpg" + } + ] + } + ], + "index": 72, + "angle": 0, + "type": "image_body" + } + ], + "index": 72 + }, + { + "type": "image", + "bbox": [ + 437, + 479, + 515, + 507 + ], + "blocks": [ + { + "bbox": [ + 437, + 479, + 515, + 507 + ], + "lines": [ + { + "bbox": [ + 437, + 479, + 515, + 507 + ], + "spans": [ + { + "bbox": [ + 437, + 479, + 515, + 507 + ], + "type": "image", + "image_path": "eb548ecab6db5ee896aeffa79fd5cdb1171897f10b7d295b93fc05e7d3a8a75b.jpg" + } + ] + } + ], + "index": 73, + "angle": 0, + "type": "image_body" + } + ], + "index": 73 + }, + { + "type": "image", + "bbox": [ + 515, + 418, + 552, + 449 + ], + "blocks": [ + { + "bbox": [ + 515, + 418, + 552, + 449 + ], + "lines": [ + { + "bbox": [ + 515, + 418, + 552, + 449 + ], + "spans": [ + { + "bbox": [ + 515, + 418, + 552, + 449 + ], + "type": "image", + "image_path": "175221f93b13db844f175279c309b5bb6d39da69577f9e215dee1cec50602c19.jpg" + } + ] + } + ], + "index": 74, + "angle": 0, + "type": "image_body" + } + ], + "index": 74 + }, + { + "type": "image", + "bbox": [ + 515, + 450, + 552, + 478 + ], + "blocks": [ + { + "bbox": [ + 515, + 450, + 552, + 478 + ], + "lines": [ + { + "bbox": [ + 515, + 450, + 552, + 478 + ], + "spans": [ + { + "bbox": [ + 515, + 450, + 552, + 478 + ], + "type": "image", + "image_path": "d19dfe15285c7e9446ade34584deabbb0520e2b031e2ad4a5ce1201a8d51cf0f.jpg" + } + ] + } + ], + "index": 75, + "angle": 0, + "type": "image_body" + } + ], + "index": 75 + }, + { + "type": "image", + "bbox": [ + 515, + 479, + 552, + 507 + ], + "blocks": [ + { + "bbox": [ + 515, + 479, + 552, + 507 + ], + "lines": [ + { + "bbox": [ + 515, + 479, + 552, + 507 + ], + "spans": [ + { + "bbox": [ + 515, + 479, + 552, + 507 + ], + "type": "image", + "image_path": "358f696296781c523fddaba5ffaa22fb87c25e9984665f6f211259046c2a6339.jpg" + } + ] + } + ], + "index": 76, + "angle": 0, + "type": "image_body" + } + ], + "index": 76 + }, + { + "bbox": [ + 313, + 561, + 451, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 451, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 451, + 574 + ], + "type": "text", + "content": "4.4. Unconditional Sampling" + } + ] + } + ], + "index": 78 + }, + { + "bbox": [ + 313, + 578, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 555, + 662 + ], + "type": "text", + "content": "SpinMeRound is trained following the Classifier-Free Guidance training scheme. Thus, our architecture can generate novel multi-view identities without any prior conditioning. By setting the input identity embedding equal to the empty string, our model can generate a novel identity along with different viewpoints of that identity. Figure 6 presents some examples of those identities." + } + ] + } + ], + "index": 79 + }, + { + "bbox": [ + 313, + 672, + 460, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 460, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 460, + 685 + ], + "type": "text", + "content": "4.5.Quantitative Comparisons" + } + ] + } + ], + "index": 80 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "We compare our method's ability to generate novel views with the current state-of-the-art models: a) Eg3D [9] and" + } + ] + } + ], + "index": 81 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14351" + } + ] + } + ], + "index": 82 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 294, + 131 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 294, + 131 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 294, + 131 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 294, + 131 + ], + "type": "table", + "html": "
MethodL2 ↓LPIPS↓SSIM↑ID Sim [13] ↑ID Sim [48] ↑
NeRFEg3d [9]0.0250.40.550.310.89
basedPanohead [1]0.0120.320.650.270.88
Zero123 [39]0.1950.5150.550.1690.44
Diffusion basedZero123-XL [12]0.1980.510.5630.1180.442
SV3D [63]0.0870.410.6600.360.881
DiffPortrait3d [25]0.10.50.350.550.887
SpinMeRound (Ours)0.0330.30.730.610.911
", + "image_path": "e83f4079a32afa7ad8ff1c5e0867c4f9b4e3bc144a006a9a40a0d6d3aa1edc1a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 198, + 295, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 198, + 295, + 424 + ], + "spans": [ + { + "bbox": [ + 54, + 198, + 295, + 424 + ], + "type": "text", + "content": "Panohead [1] which are NeRF-based approaches generating frontal and full-head portraits respectively, b) Zero123 [68], Zero123-XL [12] and DiffPortrai3D [25], which are multiview diffusion models and c) SV3D [63], a multi-view video diffusion model. For SV3D, we selected the SV3Dp variant, as it supports flexible viewing angles, whereas SV3D-u operates only with fixed viewpoints. We validate our approach using the NeRSemble dataset [32], which includes 222 unique identities recorded from 16 angles. We randomly select a timestamp from one of their video sequences for each individual and extract a random subset of views. All selected frames are then centered, and one of these views is used as input for the tested methods. We evaluate their reconstruction performance using L2 distance, LPIPS [71], SSIM, and an Identity Similarity score. To calculate the Identity Similarity score, we feed both the ground-truth and the reconstructed images into the ArcFace [13] and VGGFace [48] networks and measure the cosine similarity between their final feature vectors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 426, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 569 + ], + "type": "text", + "content": "Table 1 summarizes the reconstruction performance of each network. SpinMeRound, achieves state-of-the-art results across all cases when evaluated on the LPIPS, SSIM, and Identity Similarity metrics. Additionally, it outperforms all diffusion-based approaches in terms of L2 distance. Although our method trails slightly behind NeRF-based approaches, this is expected given the inherent advantages of NeRF-based models in novel view synthesis tasks. Notably, methods like Eg3D and Panohead require a time-intensive fitting process, which can occasionally fail. In contrast, our proposed approach relies solely on an efficient sampling process, eliminating the need for fitting." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 580, + 156, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 156, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 156, + 593 + ], + "type": "text", + "content": "5. Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 600, + 173, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 173, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 173, + 613 + ], + "type": "text", + "content": "5.1. Component analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": "In this section, we conduct ablation studies for the importance of the identity mechanism, the shape normals and the input ground truth image. For this reason, we trained the following three models while using the same training data: a) our proposed architecture, without integrating the input conditioning view, b) a Stable Diffusion-based model without integrating the robust identity mechanism, and c) our model without generating any shape normals." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 337, + 70, + 533, + 123 + ], + "blocks": [ + { + "bbox": [ + 55, + 140, + 295, + 185 + ], + "lines": [ + { + "bbox": [ + 55, + 140, + 295, + 185 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 295, + 185 + ], + "type": "text", + "content": "Table 1. Reconstruction performance on the NeRSample [32] dataset shows that SpinMeRound achieves state-of-the-art results in LPIPS, SSIM and ID Sim metrics while performing on par with the leading models in terms of L2 distance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 337, + 70, + 533, + 123 + ], + "lines": [ + { + "bbox": [ + 337, + 70, + 533, + 123 + ], + "spans": [ + { + "bbox": [ + 337, + 70, + 533, + 123 + ], + "type": "table", + "html": "
L2 ↓LPIPS ↓SSIM ↑
SpinMeRound (w/o Input Image)0.12460.42990.568
SpinMeRound (w/o identity embedding)0.0280.260.70
SpinMeRound (w/o Normals)0.0560.320.65
SpinMeRound0.0180.220.75
", + "image_path": "67e69a158bf1d60d300fb33774c413f690ef3bdbf694b4f4b516e498c45e961b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 132, + 553, + 175 + ], + "lines": [ + { + "bbox": [ + 313, + 132, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 313, + 132, + 553, + 175 + ], + "type": "text", + "content": "Table 2. Ablation study: We evaluate the performance of Spin-MeRound alongside three variations, demonstrating that the proposed architecture achieves the highest performance across all reconstruction metrics." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 332, + 207, + 531, + 334 + ], + "blocks": [ + { + "bbox": [ + 332, + 207, + 531, + 334 + ], + "lines": [ + { + "bbox": [ + 332, + 207, + 531, + 334 + ], + "spans": [ + { + "bbox": [ + 332, + 207, + 531, + 334 + ], + "type": "image", + "image_path": "c0db0fb64d4ed31f3ea06fdbde5a275cb7647a6cbb5f1e3f324595d5dc20910a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 345, + 553, + 401 + ], + "lines": [ + { + "bbox": [ + 313, + 345, + 553, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 553, + 401 + ], + "type": "text", + "content": "Figure 7. Ablation Study: We compare the identity similarity between four models: SpinMeRound (Ours), a similar model without the input image, ones without the identity embedding mechanism, and one without generating shape normals " + }, + { + "bbox": [ + 313, + 345, + 553, + 401 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 313, + 345, + 553, + 401 + ], + "type": "text", + "content": ". Results show that our proposed architecture achieves the highest identity similarity." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 407, + 554, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 554, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 554, + 492 + ], + "type": "text", + "content": "Use of input image: Starting from a pre-trained Arc2Face model using the EDM framework, we trained a multi-view model to generate novel viewpoints given only an input identity embedding vector. Notably, SpinMeRound has also been trained to be able to generate input identities without providing the conditioning view. In this case, it generates subjects close to the input face, as illustrated in Figure 8." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 506, + 554, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 554, + 590 + ], + "type": "text", + "content": "Identity embedding mechanism: In this case, we start from a pre-trained Stable diffusion 1.5 architecture trained using the EDM framework. This model gets trained following the same training parameters, as presented in Sec 3.2. Instead of using the proposed identity conditioning mechanism, we feed an empty string in the conditioning attention layers of the denoising UNet." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 553, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 654 + ], + "type": "text", + "content": "Use of Normals: We also train a variant of the network similar to SpinMeRound, but without generating shape normals. This version can only generate novel views without having any additional information about the facial shape." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "Aiming to showcase the importance of each component, we conducted a multi-view reconstruction experiment. More specifically, we sample 100 distinct identities using Panohead [1]. For each subject, we sample 10 different views along with an input frontal view. Given the frontal" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14352" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 69, + 553, + 328 + ], + "blocks": [ + { + "bbox": [ + 61, + 69, + 553, + 328 + ], + "lines": [ + { + "bbox": [ + 61, + 69, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 61, + 69, + 553, + 328 + ], + "type": "image", + "image_path": "734d747f27c7e00de1a27859f50087eff27c210f912c3f1ba4ec94ba0215b88c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 335, + 555, + 360 + ], + "lines": [ + { + "bbox": [ + 55, + 335, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 555, + 360 + ], + "type": "text", + "content": "Figure 8. From the input images on the left, we present samples generated by our method using only the corresponding identity embedding vectors (top row) and using also the input image (bottom row). The resulting subjects closely resemble the input identities." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 368, + 294, + 460 + ], + "blocks": [ + { + "bbox": [ + 56, + 368, + 294, + 460 + ], + "lines": [ + { + "bbox": [ + 56, + 368, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 368, + 294, + 460 + ], + "type": "image", + "image_path": "45efca35cf8c8420dbbea15dfa89e5396ae7df4c536affc1b19d2dd76822125e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 468, + 295, + 491 + ], + "lines": [ + { + "bbox": [ + 55, + 468, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 295, + 491 + ], + "type": "text", + "content": "Figure 9. Identity interpolation between subject pairs, showcasing our method generates smooth transitions between identities." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 498, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 581 + ], + "type": "text", + "content": "view as input to each separate model, we reconstruct the remaining views. We measure the performance of each model by calculating their discrepancy using L2 distance, LPIPS [71], SSIM and ID Similarity. As presented in Table 2 and Figure 7, our proposed methodology performs better regarding all reconstruction metrics while achieving the highest identity similarity score." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 588, + 201, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 201, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 201, + 601 + ], + "type": "text", + "content": "5.2. Importance of ID features" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "Choosing to condition the stable diffusion model in identity embeddings is an important design choice for our network. Those forms of representation contain compact information extracted from a FR model (ArcFace [13]), trained in millions of different images and subjects. Those features allow linear interpolation between the facial characteristics of different subjects. Hence, while using only the identity embedding layer, we interpolate between 2 distinct identities and present their linear interpolation results in Figure 9. It is" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 370, + 555, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 555, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 555, + 395 + ], + "type": "text", + "content": "clearly shown that our model generates smooth transitions between the generated identities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 413, + 388, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 413, + 388, + 425 + ], + "spans": [ + { + "bbox": [ + 314, + 413, + 388, + 425 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 436, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 555, + 555 + ], + "type": "text", + "content": "In this paper, we presented SpinMeRound, a multi-view latent diffusion model, which generates all-around head portraits of an input subject given a number of input views alongside an input identity embedding. Additionally, we introduced a sampling strategy for generating consistent intermediate views, given an unconstrained input \"in-the-wild\" facial image. Being trained sorely on multi-view synthetic data, we showcase our method's abilities by beating the current state-of-the-art multi-view diffusion models in novel view synthesis experiments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": "Acknowledgments: S. Zafeiriou and part of the research was funded by the EPSRC Fellowship DEFORM (EP/S010203/1) and EPSRC Project GNOMON (EP/X011364/1). The authors gratefully acknowledge the scientific support and HPC resources provided by the Erlangen National High Performance Computing Center (NHR@FAU) of FAU under the NHR projects b143dc and b180dc. NHR funding is provided by federal and Bavarian state authorities. NHR@FAU hardware is partially funded by the German Research Foundation (DFG) - 440719683. Additional support was also received by the ERC - project MIA-NORMAL 101083647, DFG 513220538, 512819079, and by the state of Bavaria (HTA)." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14353" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "type": "text", + "content": "[1] Sizhe An, Hongyi Xu, Yichun Shi, Guoxian Song, Umit Ogras, and Linjie Luo. Panohed: Geometry-aware 3d fullhead synthesis in " + }, + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "type": "text", + "content": ", 2023. 2, 5, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 169 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 169 + ], + "type": "text", + "content": "[2] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 171, + 295, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 171, + 295, + 225 + ], + "spans": [ + { + "bbox": [ + 62, + 171, + 295, + 225 + ], + "type": "text", + "content": "[3] Francesca Babiloni, Alexandros Lattas, Jiankang Deng, and Stefanos Zafeiriou. Id-to-3d: Expressive id-guided 3d heads via score distillation sampling. In Advances in Neural Information Processing Systems, pages 97167–97183. Curran Associates, Inc., 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 228, + 295, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 228, + 295, + 250 + ], + "spans": [ + { + "bbox": [ + 62, + 228, + 295, + 250 + ], + "type": "text", + "content": "[4] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In SIGGRAPH '99, 1999. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 251, + 295, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 295, + 306 + ], + "type": "text", + "content": "[5] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets, 2023. 1, 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 308, + 295, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 295, + 363 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 295, + 363 + ], + "type": "text", + "content": "[6] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 365, + 295, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 365, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 62, + 365, + 295, + 420 + ], + "type": "text", + "content": "[7] James Booth, Anastasios Roussos, Stefanos Zafeiriou, Allan Ponniahy, and David Dunaway. A 3D morphable model learnt from 10,000 faces. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 5543-5552, 2016. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 421, + 295, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 421, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 62, + 421, + 295, + 464 + ], + "type": "text", + "content": "[8] James Booth, Anastasios Roussos, Allan Ponniah, David Dunaway, and Stefanos Zafeiriou. Large scale 3d morphable models. International Journal of Computer Vision, 126(2): 233-254, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 467, + 295, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 467, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 62, + 467, + 295, + 553 + ], + "type": "text", + "content": "[9] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J. Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16123-16133, 2022. 2, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 556, + 295, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 556, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 57, + 556, + 295, + 600 + ], + "type": "text", + "content": "[10] Xiyi Chen, Marko Mihajlovic, Shaofei Wang, Sergey Prokudin, and Siyu Tang. Morphable diffusion: 3d-consistent diffusion for single-image avatar creation. 2024. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 602, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 635 + ], + "type": "text", + "content": "[11] Zilong Chen, Yikai Wang, Feng Wang, Zhengyi Wang, and Huaping Liu. V3d: Video diffusion models are effective 3d generators. arXiv preprint arXiv:2403.06738, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 636, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 712 + ], + "type": "text", + "content": "[12] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-xl: A universe of " + }, + { + "bbox": [ + 57, + 636, + 295, + 712 + ], + "type": "inline_equation", + "content": "10\\mathrm{m} + 3\\mathrm{d}" + }, + { + "bbox": [ + 57, + 636, + 295, + 712 + ], + "type": "text", + "content": " objects. arXiv preprint arXiv:2307.05663, 2023. 6, 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[13] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 129, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 183 + ], + "type": "text", + "content": "[14] Yu Deng, Duomin Wang, Xiaohang Ren, Xingyu Chen, and Baoyuan Wang. Portrait4d: Learning one-shot 4d head avatar synthesis using synthetic data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 185, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 553, + 218 + ], + "type": "text", + "content": "[15] Yu Deng, Duomin Wang, and Baoyuan Wang. Portrait4d-v2: Pseudo multi-view data creates better 4d head synthesizer. arXiv preprint arXiv:2403.13570, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 220, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 220, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 220, + 553, + 262 + ], + "type": "text", + "content": "[16] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In Advances in Neural Information Processing Systems, pages 8780-8794. Curran Associates, Inc., 2021. 1, 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "type": "text", + "content": "[17] Yao Feng, Haiwen Feng, Michael J. Black, and Timo Bolkart. Learning an animatable detailed 3D face model from in-the-wild images. 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 299, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 299, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 317, + 299, + 553, + 354 + ], + "type": "text", + "content": "[18] Stathis Galanakis, Baris Gecer, Alexandros Lattas, and Stefanos Zafeiriou. 3dmm-rf: Convolutional radiance fields for 3d face modeling. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 3536-3547, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 355, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 355, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 317, + 355, + 553, + 398 + ], + "type": "text", + "content": "[19] Stathis Galanakis, Alexandros Lattas, Stylianos Moschoglou, and Stefanos Zafeiriou. Fitdiff: Robust monocular 3d facial shape and reflectance estimation using diffusion models, 2024. 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 399, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 454 + ], + "type": "text", + "content": "[20] Ruiqi Gao*, Aleksander Holynski*, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul P. Srinivasan, Jonathan T. Barron, and Ben Poole*. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv, 2024. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 456, + 553, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 553, + 511 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 553, + 511 + ], + "type": "text", + "content": "[21] Baris Gecer, Stylianos Ploumpis, Irene Kotsia, and Stefanos Zafeiriou. Ganfit: Generative adversarial network fitting for high fidelity 3d face reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 512, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 512, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 512, + 553, + 567 + ], + "type": "text", + "content": "[22] Dimitrios Gerogiannis, Foivos Paraperas Papantoniou, Rolandos Alexandros Potamias, Alexandros Lattas, and Stefanos Zafeiriou. Arc2avatar: Generating expressive 3d avatars from a single image via id guidance. arXiv preprint arXiv:2501.05379, 2025. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 568, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 553, + 622 + ], + "type": "text", + "content": "[23] Simon Giebenhain, Tobias Kirschstein, Markos Georgopoulos, Martin Rinz, Lourdes Agapito, and Matthias Nießner. Learning neural parametric head models. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 553, + 678 + ], + "type": "text", + "content": "[24] Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. Emu video: Factorizing text-to-video generation by explicit image conditioning, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "type": "text", + "content": "[25] Yuming Gu, Hongyi Xu, You Xie, Guoxian Song, Yichun Shi, Di Chang, Jing Yang, and Linjie Luo. Diffportrait3d: Controllable diffusion for zero-shot portrait view synthe" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14354" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 106 + ], + "type": "text", + "content": "sis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10456-10465, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "text", + "content": "[26] Yuming Gu, Phong Tran, Yujuan Zheng, Hongyi Xu, Heyuan Li, Adilbek Karmanov, and Hao Li. Diffportrait360: Consistent portrait diffusion for 360 view synthesis, 2025. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 141, + 294, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 141, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 57, + 141, + 294, + 163 + ], + "type": "text", + "content": "[27] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 164, + 295, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 164, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 57, + 164, + 295, + 196 + ], + "type": "text", + "content": "[28] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239, 2020. 1, 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 198, + 295, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 198, + 295, + 242 + ], + "spans": [ + { + "bbox": [ + 57, + 198, + 295, + 242 + ], + "type": "text", + "content": "[29] Yash Kant, Ethan Weber, Jin Kyu Kim, Rawal Khirodkar, Su Zhaoen, Julieta Martinez, Igor Gilitschenski, Shunsuke Saito, and Timur Bagautdinov. Pippo: High-resolution multi-view humans from a single image. 2025. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 243, + 294, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 243, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 57, + 243, + 294, + 275 + ], + "type": "text", + "content": "[30] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Proc. NeurIPS, 2022. 4, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 277, + 294, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 277, + 294, + 320 + ], + "spans": [ + { + "bbox": [ + 56, + 277, + 294, + 320 + ], + "type": "text", + "content": "[31] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 322, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 322, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 322, + 294, + 365 + ], + "type": "text", + "content": "[32] Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radiance field reconstruction of human heads. ACM Trans. Graph., 42(4), 2023. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 367, + 294, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 367, + 294, + 422 + ], + "spans": [ + { + "bbox": [ + 56, + 367, + 294, + 422 + ], + "type": "text", + "content": "[33] Yushi Lan, Feitong Tan, Di Qiu, Qiangeng Xu, Kyle Genova, Zeng Huang, Sean Fanello, Rohit Pandey, Thomas Funkhouser, Chen Change Loy, and Yinda Zhang. Gaussian3diff: 3d gaussian diffusion for 3d full head synthesis and editing. In ECCV, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 423, + 294, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 423, + 294, + 478 + ], + "spans": [ + { + "bbox": [ + 56, + 423, + 294, + 478 + ], + "type": "text", + "content": "[34] Alexandros Lattas, Stylianos Moschoglou, Stylianos Ploumpis, Baris Gecer, Jiankang Deng, and Stefanos Zafeiriou. FitMe: Deep photorealistic 3D morphable model avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 479, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 521 + ], + "type": "text", + "content": "[35] Heyuan Li, Ce Chen, Tianhao Shi, Yuda Qiu, Sizhe An, Guanying Chen, and Xiaoguang Han. Spherehead: Stable 3d full-head synthesis with spherical tri-plane representation, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "type": "text", + "content": "[36] Peng Li, Yuan Liu, Xiaoxiao Long, Feihu Zhang, Cheng Lin, Mengfei Li, Xingqun Qi, Shanghang Zhang, Wenhan Luo, Ping Tan, et al. Era3d: High-resolution multiview diffusion using efficient row-wise attention. arXiv preprint arXiv:2405.11616, 2024. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 579, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 623 + ], + "type": "text", + "content": "[37] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "type": "text", + "content": "[38] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[39] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object, 2023. 2, 3, 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[40] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 119, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 151 + ], + "type": "text", + "content": "[41] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. In SIGGRAPH, pages 163-169. ACM, 1987. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 152, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 196 + ], + "type": "text", + "content": "[42] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, Natalia Neverova, Andrea Vedaldi, Oran Gafni, and Filippos Kokkinos. Im-3d: Iterative multiview diffusion and reconstruction for high-quality 3d generation, 2024. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 198, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 252 + ], + "type": "text", + "content": "[43] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 254, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 319 + ], + "type": "text", + "content": "[44] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13503–13513, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 320, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 387 + ], + "type": "text", + "content": "[45] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. StyleSDF: High-Resolution 3D-Consistent Image and Geometry Generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13503–13513, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "type": "text", + "content": "[46] Foivos Paraperas Papantoniou, Alexandros Lattas, Stylianos Moschoglou, and Stefanos Zafeiriou. Relightify: Relightable 3d faces from a single image via diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 3, 4, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 444, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 498 + ], + "type": "text", + "content": "[47] Foivos Paraperas Papantoniou, Alexandros Lattas, Stylianos Moschoglou, Jiankang Deng, Bernhard Kainz, and Stefanos Zafeiriou. Arc2face: A foundation model for id-consistent human faces. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 500, + 553, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 532 + ], + "type": "text", + "content": "[48] Omkar M. Parkhi, Andrea Vedaldi, and Andrew Zisserman. Deep face recognition. In *British Machine Vision Conference*, 2015. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 534, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 553, + 567 + ], + "type": "text", + "content": "[49] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 568, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 601 + ], + "type": "text", + "content": "[50] Malte Prinzler, Egor Zakharov, Vanessa Sklyarova, Berna Kabadayi, and Justus Thies. Joker: Conditional 3d head synthesis with extreme facial expressions, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 602, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 678 + ], + "type": "text", + "content": "[51] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, and Bernard Ghanem. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. In The Twelfth International Conference on Learning Representations (ICLR), 2024. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[52] Alec Radford, Jong Wook Kim, Chris Hallacy, A. Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14355" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 711 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "text", + "content": "Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 95, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 95, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 57, + 95, + 295, + 149 + ], + "type": "text", + "content": "[53] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Ben Mildenhall, Nataniel Ruiz, Shiran Zada, Kfir Aberman, Michael Rubenstein, Jonathan Barron, Yuanzhen Li, and Varun Jampani. Dreambooth3d: Subject-driven text-to-3d generation. ICCV, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 150, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 150, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 57, + 150, + 294, + 194 + ], + "type": "text", + "content": "[54] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 194, + 294, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 294, + 227 + ], + "type": "text", + "content": "[55] Daniel Roich, Ron Mokady, Amit H. Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 228, + 295, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 295, + 282 + ], + "type": "text", + "content": "[56] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2, 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 284, + 294, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 284, + 294, + 315 + ], + "spans": [ + { + "bbox": [ + 57, + 284, + 294, + 315 + ], + "type": "text", + "content": "[57] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. CoRR, abs/1505.04597, 2015. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 316, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 392 + ], + "type": "text", + "content": "[58] Mehdi S. M. Sajjadi, Henning Meyer, Etienne Pot, Urs Bergmann, Klaus Greff, Noha Radwan, Suhani Vora, Mario Lucic, Daniel Duckworth, Alexey Dosovitskiy, Jakob Uszkoreit, Thomas Funkhouser, and Andrea Tagliasacchi. Scene Representation Transformer: Geometry-Free Novel View Synthesis Through Set-Latent Scene Representations. CVPR, 2022. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 393, + 294, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 393, + 294, + 436 + ], + "spans": [ + { + "bbox": [ + 57, + 393, + 294, + 436 + ], + "type": "text", + "content": "[59] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 437, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 437, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 57, + 437, + 294, + 470 + ], + "type": "text", + "content": "[60] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv:2308.16512, 2023. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 471, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 471, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 57, + 471, + 294, + 503 + ], + "type": "text", + "content": "[61] Felix Taubner, Ruihang Zhang, Mathieu Tuli, and David B. Lindell. Cap4d: Creating animatable 4d portrait avatars with morphable multi-view diffusion models, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 504, + 294, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 558 + ], + "type": "text", + "content": "[62] Alex Trevithick, Matthew Chan, Michael Stengel, Eric R. Chan, Chao Liu, Zhiding Yu, Sameh Khamis, Manmohan Chandraker, Ravi Ramamoorthi, and Koki Nagano. Real-time radiance fields for single-image portrait view synthesis. In ACM Transactions on Graphics (SIGGRAPH), 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 559, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 613 + ], + "type": "text", + "content": "[63] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 614, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 294, + 656 + ], + "type": "text", + "content": "[64] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 294, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 711 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 711 + ], + "type": "text", + "content": "[65] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, and Baining Guo. Rodin: A generative model for sculpting 3d digital avatars using diffusion, 2022. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 529 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "type": "text", + "content": "[66] Haohan Weng, Tianyu Yang, Jianan Wang, Yu Li, Tong Zhang, C. L. Philip Chen, and Lei Zhang. Consistent123: Improve consistency for one image to 3d object synthesis, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "type": "text", + "content": "[67] Yiqian Wu, Hao Xu, Xiangjun Tang, Hongbo Fu, and Xiao-gang Jin. 3dportraitgan: Learning one-quarter headshot 3d gans from a single-view portrait dataset with diverse body poses, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 217 + ], + "type": "text", + "content": "[68] Yuyang Huang Taoran Yi Xiaopeng Zhang Lingxi Xie Xinggang Wang Wenrui Dai Hongkai Xiong and Qi Tian Yabo Chen, Jiemin Fang. Cascade-zero123: One image to highly consistent 3d with self-prompted nearby views. arXiv preprint arXiv:2312.04424, 2023. 3, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 219, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 284 + ], + "type": "text", + "content": "[69] Tarun Yenamandra, Ayush Tewari, Florian Bernard, Hans-Peter Seidel, Mohamed Elgharib, Daniel Cremers, and Christian Theobalt. i3dmm: Deep implicit 3d morphable model of human heads. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12803-12813, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 285, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 553, + 329 + ], + "type": "text", + "content": "[70] Bowen Zhang, Yiji Cheng, Chunyu Wang, Ting Zhang, Jiaolong Yang, Yansong Tang, Feng Zhao, Dong Chen, and Baining Guo. Rodinhd: High-fidelity 3d avatar generation with diffusion models. arXiv preprint arXiv:2407.06938, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 331, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 363 + ], + "type": "text", + "content": "[71] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 365, + 553, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 418 + ], + "type": "text", + "content": "[72] Mingwu Zheng, Hongyu Yang, Di Huang, and Liming Chen. Imface: A nonlinear 3d morphable face model with implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20343-20352, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 420, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 420, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 420, + 553, + 453 + ], + "type": "text", + "content": "[73] Zhenglin Zhou, Fan Ma, Hehe Fan, Zongxin Yang, and Yi Yang. Headstudio: Text to animatable head avatars with 3d gaussian splatting. 2024. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 454, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 454, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 454, + 553, + 529 + ], + "type": "text", + "content": "[74] Zheng Zhu, Guan Huang, Jiankang Deng, Yun Ye, Junjie Huang, Xinze Chen, Jiagang Zhu, Tian Yang, Jiwen Lu, Dalong Du, and Jie Zhou. Webface260m: A benchmark unveiling the power of million-scale deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10492-10502, 2021. 3" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "14356" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SplArt_ Articulation Estimation and Part-Level Reconstruction with 3D Gaussian Splatting/99f05b1c-e917-486f-913a-4812f9c2d857_content_list.json b/2025/SplArt_ Articulation Estimation and Part-Level Reconstruction with 3D Gaussian Splatting/99f05b1c-e917-486f-913a-4812f9c2d857_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..730cb38e6bbeafefbeec7dcd1855e717cd0cf82d --- /dev/null +++ b/2025/SplArt_ Articulation Estimation and Part-Level Reconstruction with 3D Gaussian Splatting/99f05b1c-e917-486f-913a-4812f9c2d857_content_list.json @@ -0,0 +1,1929 @@ +[ + { + "type": "text", + "text": "SPLART: Articulation Estimation and Part-Level Reconstruction with 3D Gaussian Splatting", + "text_level": 1, + "bbox": [ + 169, + 130, + 828, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shengjie Lin†, Jiading Fang†, Muhammad Zubair Irshad‡, Vitor Campagnolo Guizilini‡, Rares Andrei Ambrus‡, Greg Shakhnarovich†‡, Matthew R. Walter†", + "bbox": [ + 166, + 202, + 861, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "† Toyota Technological Institute at Chicago, ‡ Toyota Research Institute", + "bbox": [ + 230, + 238, + 797, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{slin, fjd, greg, mwalter}@ttic.edu, {zubair.irshad, vitor.quizilini, rares.ambrus}@tri.global", + "bbox": [ + 99, + 258, + 926, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 309, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reconstructing articulated objects prevalent in daily environments is crucial for applications in augmented/virtual reality and robotics. However, existing methods face scalability limitations (requiring 3D supervision or costly annotations), robustness issues (being susceptible to local optima), and rendering shortcomings (lacking speed or photorealism). We introduce SPLART, a self-supervised, category-agnostic framework that uses 3D Gaussian Splattig (3DGS) to reconstruct and infer the kinematics of articulated objects from two sets of posed RGB images captured at different articulation states, enabling real-time photorealistic rendering for novel viewpoints and articulations. SPLART augments 3DGS with a differentiable mobility parameter per Gaussian, achieving refined part segmentation. A multi-stage optimization strategy is employed to progressively handle reconstruction, part segmentation, and articulation estimation, significantly enhancing robustness and accuracy. SPLART exploits geometric self-supervision, effectively addressing challenging scenarios without requiring 3D annotations or category-specific priors. Evaluations on established and newly proposed benchmarks, along with applications to real-world scenarios using a handheld RGB camera, demonstrate SPLART's state-of-the-art performance and real-world practicality. Code is publicly available at https://github.com/ripl/splart.", + "bbox": [ + 88, + 340, + 483, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 739, + 220, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Articulated objects, such as drawers, doors, and scissors, are ubiquitous in our daily lives, yet their dynamic nature poses significant challenges for 3D reconstruction—a critical task for applications in augmented/virtual reality [36, 51], robotics [1, 5, 16, 24, 50, 57], and computer vision [20, 45]. Existing methods for reconstructing articulated objects are hindered by several key limitations: they often require labor-intensive supervision (e.g., part-level segmentation or articulation annotations) [23, 29,", + "bbox": [ + 89, + 763, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8c14f10e462f496c2e116707619e37f0dff2c7fbbf757212e58adff3a0acdffe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 306, + 696, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/739155a868204e504885cf6d17dc0b8e5a8bd4aaa5adfd8248302b293390ad17.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 308, + 898, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4e480fa9bf958ffd84a582c3e8ba48fd1f638390542b6b1e11a139f0c8135f82.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 422, + 606, + 511 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/edfef345eb7328e4432c033b9531328821d4ce0cfe56dd7ff2df97a62174cb15.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 436, + 763, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "SPLART", + "bbox": [ + 779, + 460, + 880, + 486 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5899ad4519229e08ade1d7230a0f3479cfcc6013fc039e1cd7190c5aaec7a93d.jpg", + "image_caption": [ + "Articulation" + ], + "image_footnote": [], + "bbox": [ + 517, + 518, + 606, + 599 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a469f80e92abc52db462b98744b78fbfd58d798534210babb4fdfc7f01304873.jpg", + "image_caption": [ + "Figure 1. Given (top) RGB images of an object at two articulation states, (middle) SPLART uses 3DGS to simultaneously reconstruct its static and dynamic parts and estimate the kinematic articulation model. SPLART is then able to (bottom) render high-fidelity 3D reconstructions of the object along with part-level segmentations for novel articulation states, allowing for novel view synthesis." + ], + "image_footnote": [], + "bbox": [ + 611, + 537, + 700, + 599 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f36a1e71cdc3e76fc452d65ea457f42324dafc504f926d76c06b5eed854bf1ea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 537, + 794, + 599 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8ffa9581e67e0231c89d4649935deddf7c23c714f3b806973cd9ca3d7fcbe7aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 795, + 539, + 888, + 598 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "40, 54, 56], they depend on 3D supervision that restricts practical use [23, 29, 33, 40, 60], they produce category-specific models that limit scalability [29, 40, 54, 59], or they are not capable of real-time, photorealistic rendering [6, 13, 23, 29, 30, 40, 54, 56, 59, 60]. To address these challenges, we introduce SPLART, a novel self-supervised and category-agnostic framework that leverages 3D Gaussian Splitting (3DGS) [26] to reconstruct articulated objects from minimal input—two sets of posed RGB images at distinct articulation states. SPLART reconstructs the object's parts and infers its kinematics, enabling real-time, photorealistic rendering for novel views and articulation states.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8841", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Central to SPLART is the augmentation of 3DGS [26] to include a differentiable mobility parameter for each Gaussian, which enables a more refined segmentation of static and mobile parts through gradient-based optimization. This results in enhanced reconstruction quality, while preserving the real-time, photorealistic rendering capabilities of 3DGS—offering a speedup of more than $100\\times$ over methods [6, 30] based on neural radiance fields [38].", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To enhance robustness, SPLART employs a multi-stage optimization strategy that decouples the part-level reconstruction and articulation estimation processes. Unlike end-to-end approaches prone to local optima [30], SPLART first independently reconstructs each articulation state, then estimates each Gaussian's mobility parameter for part segmentation, and finally refines both the articulation and mobility estimates jointly. This structured approach ensures stable and accurate convergence, avoiding the stringent initialization requirements of existing methods, thereby providing a practical solution for challenging articulated structures.", + "bbox": [ + 89, + 212, + 480, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on this foundation, SPLART leverages geometric self-supervision to eliminate the need for manual annotations or 3D supervision. By enforcing geometric consistency between reconstructions through complementary loss formulations, SPLART robustly estimates articulation parameters across diverse scenarios. This self-supervised strategy enhances scalability, enabling SPLART to reconstruct a wide range of articulated objects without relying on prior structural or categorical knowledge.", + "bbox": [ + 89, + 380, + 480, + 515 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive evaluations on both established and newly introduced benchmarks demonstrate SPLART's superior articulation accuracy and reconstruction quality, surpassing state-of-the-art methods without requiring 3D supervision. Real-world experiments further validate its practicality, showcasing successful reconstructions of diverse articulated objects using only a handheld RGB camera.", + "bbox": [ + 89, + 517, + 480, + 622 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, this work contributes:", + "bbox": [ + 109, + 623, + 341, + 637 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. An extension of 3DGS with a differentiable mobility value per Gaussian that enables precise part segmentation using gradient-based optimization.", + "2. A multi-stage optimization strategy that decouples reconstruction and articulation estimation, enhancing robustness and accuracy.", + "3. Complementary formulations of geometric self-supervision for articulation estimation, eliminating the need for 3D supervision or laborious annotations.", + "4. A challenging dataset and new metrics for comprehensive evaluation of articulated object reconstruction." + ], + "bbox": [ + 89, + 638, + 480, + 804 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 821, + 232, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Data-Driven Articulation Learning", + "text_level": 1, + "bbox": [ + 89, + 847, + 395, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Estimating the pose and joint properties of articulated objects is crucial for robot manipulation and interac", + "bbox": [ + 89, + 869, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion [10, 11, 19, 31]. Recent learning-based methods [9, 13, 15, 22, 29, 32, 56, 64] infer articulation properties from point clouds via end-to-end training. For instance, Shape2Motion [56] analyzes motion parts from a single point cloud in a supervised setting, while ANCSH [29] performs category-level pose estimation, but requires class-specific models. RPM-Net [62] enhances generalization across categories for part segmentation and kinematic prediction, and DITTO [23] predicts motion and geometry from 3D point cloud pairs without labels. However, these methods depend on costly 3D supervision and annotations. In contrast, our approach reconstructs accurate 3D geometry and detailed appearance, capturing articulation without 3D supervision or priors.", + "bbox": [ + 511, + 90, + 903, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Representations for Object Reconstruction", + "text_level": 1, + "bbox": [ + 511, + 313, + 877, + 329 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early 3D object reconstruction methods predicted point clouds, voxels, or meshes from partial observations [3, 8, 17]. Recent advances in implicit scene representations [18, 26, 37, 38, 44, 65] enable detailed geometry and appearance reconstruction via differentiable rendering [12, 43, 52, 55]. While neural fields suffer from slow rendering, 3D Gaussian Splatting (3DGS) [26] overcomes this by using explicit 3D Gaussians. We leverage 3DGS for self-supervised articulated object reconstruction from posed RGB images, achieving fast, realistic synthesis of novel views and articulations in real time.", + "bbox": [ + 511, + 335, + 903, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Articulated Object Reconstruction", + "text_level": 1, + "bbox": [ + 511, + 512, + 815, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent methods leverage differentiable 3D representations [26, 38, 44] to jointly reconstruct articulated objects and infer articulation parameters. Training-based approaches use synthetic 3D data to predict joint parameters and segment parts [13, 14, 23, 25, 40, 42, 59]. Self-supervised methods optimize shape, appearance, and articulation per scene without extensive training [6, 27, 30, 32, 33, 49, 60]. Some handle objects with more than two parts, but require the number of parts to be known and the mobile parts to be directly connected to the static part [6, 33, 60]. Other works enhance articulation estimation using large language or vision-language models [28, 35]. In contrast, our self-supervised method reconstructs two-part articulated objects from RGB images across articulation states using 3DGS [26]. As the first to apply 3DGS to this task without 3D supervision or pre-trained priors, it robustly handles challenging cases and achieves real-time performance.", + "bbox": [ + 511, + 535, + 903, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 806, + 648, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 511, + 832, + 624, + 847 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Consider an arbitrary object composed of two rigid parts: a static parent part, and a child part that can move relative to its parent through either a revolute or prismatic articulation.", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8842", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/949f48fd63c7627d87b3ddba7965c91d5267f78a243ccf93e30db8e0b339e1e6.jpg", + "image_caption": [ + "(a) Stage 1: Per-state reconstruction." + ], + "image_footnote": [], + "bbox": [ + 91, + 128, + 181, + 193 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/65d52b5e2530a042a40745b877f08db5d0627f3ad8648ed93ec38e1365793423.jpg", + "image_caption": [ + "(Section 3.2.1)" + ], + "image_footnote": [], + "bbox": [ + 189, + 128, + 269, + 193 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a682f366330860f5adc56c9e1de27e762e3ab7efe3a607caeaba740e2653b232.jpg", + "image_caption": [ + "Figure 2. Methodology overview of SPLART. It consists of three decoupled stages for the purpose of stable optimization. Stage 1 constructs a 3DGS for each state from posed RGB images using the photometric loss. Stage 2 proposes cross-static formulation for mobility estimation. Intuitively, it combines static parts of both states and the mobile part of the desired state for the target Gaussians. Stage 3 proposes cross-mobile formulation for articulation estimation. Intuitively, it combines static parts of both states, the mobile part of the desired state and the transformed mobile part of the other state together for the target Gaussians." + ], + "image_footnote": [], + "bbox": [ + 94, + 195, + 173, + 271 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/89dd5b274ef6c9ec9cb37e73fac5119891da0a47e96c665fd4043b933b2cf2a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 187, + 195, + 264, + 271 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/56152ec1f68d402376abfd4ece5c2250a479c0577d0df1d3c58e428f593016d6.jpg", + "image_caption": [ + "(b) Stage 2: Cross-static formulation for mobility estimation." + ], + "image_footnote": [], + "bbox": [ + 277, + 117, + 573, + 271 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f9ba36a2f893efaaab5e87cb625da3b3743ac339695c80e7b75b3b6cd0c08780.jpg", + "image_caption": [ + "(c) Stage 3: Cross-mobile formulation for articulation estimation and mobility refinement." + ], + "image_footnote": [], + "bbox": [ + 581, + 117, + 901, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our objective is twofold: (1) to reconstruct the articulated object at the part level; and (2) to estimate its articulated motion. Assuming a known articulation type (i.e., either revolute or prismatic), the input to our method consists of two sets of posed RGB images (i.e., images with known camera intrinsics and extrinsics), each capturing the articulated object at one end state of the motion.", + "bbox": [ + 88, + 364, + 482, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formally, let $l$ denote the articulation state label, where $l = 0$ and $l = 1$ correspond to the two end states of the observed articulation. For reconstruction, SPLART uses observations $\\mathcal{O}_l = \\{(I_l^i,P_l^i,K_l^i)\\}_{i = 1}^{N_l}, l\\in \\{0,1\\}$ , where $I_{l}^{i}$ is the $i$ -th observed RGB image of the articulated object at state $l$ , $P_{l}^{i}$ and $K_{l}^{i}$ denote its camera extrinsic and intrinsic, respectively, and $N_{l}$ represents the number of data samples for state $l$ . Note that $P_{l}^{i}$ is specified in a common world frame for both states, while the articulated motion involves only one moving part w.r.t. the world space. SPLART models a revolute articulation with its rotation axis $\\mathbf{a}$ ( $\\| \\mathbf{a}\\| = 1$ ), pivot $\\mathbf{p}$ , and rotation angle $\\theta$ , such that a point $\\mathbf{x}$ (in the world frame) on the mobile part at state $l = 0$ will move to", + "bbox": [ + 89, + 470, + 483, + 667 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n{ } ^ { 1 } T _ { 0 } \\mathbf { x } = R _ { \\mathbf { a } , \\theta } ( \\mathbf { x } - \\mathbf { p } ) + \\mathbf { p } \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 676, + 482, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "at state $l = 1$ , where $R_{\\mathbf{a},\\theta}$ is the rotation induced by the axis-angle notation. A prismatic articulation is instead modeled by its translation axis $\\mathbf{a}$ ( $\\| \\mathbf{a} \\| = 1$ ) and distance $d$ . The goal is to reconstruct the articulated object at the part level using a chosen representation while estimating the articulated motion $^1 T_0$ , ensuring that the renderings at each articulation state are with the observations.", + "bbox": [ + 89, + 703, + 482, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SPLART extends the 3DGS representation for articulated objects, decoupling the goals of part-level reconstruction and articulation estimation across three stages: (1) separate reconstructions for each articulation state (Sec. 3.2.1), (2) mobility estimation using the cross-static formulation (Sec. 3.2.2), and (3) articulation estimation and mobility re", + "bbox": [ + 89, + 810, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "finement using the cross-mobile formulation (Sec. 3.2.3). To facilitate the application of SPLART to real-world objects, we leverage modern structure-from-motion and image segmentation techniques, developing a framework that enables general users to reconstruct articulated objects in their surroundings using only images captured by a handheld camera device (Sec. 3.3).", + "bbox": [ + 511, + 364, + 906, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Decoupled Multi-Stage Optimization", + "text_level": 1, + "bbox": [ + 511, + 479, + 831, + 496 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Jointly performing part-level reconstruction and articulation estimation on an articulated object is prone to local minima [6, 30], making it preferable to strategically decouple the problem into multiple stages.", + "bbox": [ + 511, + 502, + 905, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1. Separate per-state reconstruction", + "text_level": 1, + "bbox": [ + 511, + 571, + 792, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Stage 1, SPLART separately optimizes two 3DGS models following the standard procedure, one for each end state of the articulation. In addition to the attributes that are standard to 3DGS, each Gaussian is initialized with a persistent binary state label $l$ , equally drawn from $\\{0,1\\}$ . We denote the set of Gaussians representing state $l$ as $\\mathcal{G}_l^{\\mathrm{ref}}$ , where $ref$ emphasizes that $\\mathcal{G}_l^{\\mathrm{ref}}$ is the reference reconstruction unaffected by the other state. Given a data sample observed at state $l$ , the optimization follows as:", + "bbox": [ + 511, + 589, + 908, + 726 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathcal {G} _ {l} ^ {\\text {r e f}}} \\Delta_ {I} \\left(\\hat {I} _ {l} ^ {\\text {r e f}, i}, I _ {l} ^ {i}\\right), \\tag {2a}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 736, + 903, + 763 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {I} _ {l} ^ {\\text {r e f}, i} = R \\left(\\mathcal {G} _ {l} ^ {\\text {r e f}}, P _ {l} ^ {i}, K _ {l} ^ {i}\\right), l \\in \\{0, 1 \\}, \\tag {2b}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 767, + 903, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $R$ is the 3DGS rendering function, and $\\Delta_I$ represents the photometric loss. For simplicity, the view index $i$ and camera parameters $P_l^i, K_l^i$ will be omitted from now on.", + "bbox": [ + 511, + 797, + 905, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.2. Cross-static formulation for mobility estimation", + "text_level": 1, + "bbox": [ + 511, + 851, + 890, + 866 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The focus of Stage 2 is mobility estimation for each Gaussian. To keep $\\mathcal{G}^{\\mathrm{ref}}$ dedicated to single-state reconstruction,", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8843", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "we first duplicate $\\mathcal{G}_l^{\\mathrm{ref}}$ as $\\mathcal{G}_l^{\\mathrm{tgt}}$ for both states (i.e., $l\\in \\{0,1\\}$ ), which is intended as the target representation that will fulfill the goals of part-level reconstruction and articulation estimation. By design, $\\mathcal{G}^{\\mathrm{tgt}}$ shares neither data storage nor gradient flow with $\\mathcal{G}^{\\mathrm{ref}}$ after its creation.", + "bbox": [ + 89, + 89, + 480, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each Gaussian in $\\mathcal{G}^{\\mathrm{tgt}}$ , we further extend its set of attributes with a mobility value $m \\in [0,1]$ , initialized at 0.5. By design, $m$ enables the break-down of a Gaussian to its static and mobile components, where each component inherits all the original Gaussian attributes except for the opacity $\\sigma$ . The static component has its opacity modified to $\\sigma \\cdot (1 - m)$ , and the mobile component to $\\sigma \\cdot m$ . For simplicity, letting $\\mathcal{M}$ be the set of mobilities for $\\mathcal{G}$ , we use the element-wise product $\\mathcal{G} \\cdot (1 - \\mathcal{M})$ to denote the static component of $\\mathcal{G}$ , and $\\mathcal{G} \\cdot \\mathcal{M}$ for the mobile component.", + "bbox": [ + 89, + 167, + 482, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To estimate the mobilities $\\mathcal{M}$ , we employ the intuition that the static components from both states should constitute the static part of the articulated object. Formally, we introduce the cross-static (cs) formulation, where the static part of the articulated object is jointly represented as", + "bbox": [ + 89, + 316, + 480, + 393 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} ^ {\\mathrm {s}} = \\mathcal {G} _ {l} ^ {\\mathrm {t g t}} \\cdot (1 - \\mathcal {M} _ {l}) \\oplus \\mathcal {G} _ {1 - l} ^ {\\mathrm {t g t}} \\cdot (1 - \\mathcal {M} _ {1 - l}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 401, + 480, + 420 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\oplus$ denotes concatenation. For state $l$ , the target representation thus becomes", + "bbox": [ + 89, + 428, + 482, + 455 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {l} ^ {\\mathrm {t g t}, \\mathrm {c s}} = \\mathcal {G} ^ {\\mathrm {s}} \\oplus \\mathcal {G} _ {l} ^ {\\mathrm {t g t}} \\cdot \\mathcal {M} _ {l}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 465, + 480, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With this formulation, Stage 2 is further divided into two sub-stages as described below.", + "bbox": [ + 89, + 493, + 480, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stage 2(a): Coarse mobility estimation via cross-static geometric consistency. To measure the geometric distance between two Gaussian sets, we design a weighted version of the Chamfer distance. Specifically, letting $X = \\{(x_{i},w_{x_{i}})\\}_{i = 1}^{M}$ and $Y = \\{(y_{j},w_{y_{j}})\\}_{j = 1}^{N}$ be two sets of point-weight pairs, the weighted Chamfer distance is then:", + "bbox": [ + 89, + 526, + 483, + 617 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname {Chamfer}(X,Y) = \\sum_{\\left(x_{i},w_{x_{i}}\\right)\\in X}\\tilde{w}_{x_{i}}\\min_{\\left(y_{j},w_{y_{j}}\\right)\\in Y}\\| x_{i} - y_{j}\\|^{2} \\\\ + \\sum_ {\\left(y _ {j}, w _ {y _ {j}}\\right) \\in Y} \\tilde {w} _ {y _ {j}} \\min _ {\\left(x _ {i}, w _ {x _ {i}}\\right) \\in X} \\| x _ {i} - y _ {j} \\| ^ {2}, \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 628, + 482, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tilde{w}_{x_i} = w_{x_i} / \\sum_{(x_i, w_{x_i}) \\in X} w_{x_i}$ and $\\tilde{w}_{y_j}$ are normalized weights. For each Gaussian, we use its mobility-modified opacity $\\sigma'$ as the weight, essentially treating it as $\\sigma'$ points overlapped at its mean position. The mobilities are then optimized by encouraging the geometric consistency formulated as follows:", + "bbox": [ + 89, + 718, + 482, + 808 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathcal {M}} \\mathrm {C D} _ {0} ^ {\\mathrm {c s}} + \\mathrm {C D} _ {1} ^ {\\mathrm {c s}} + \\lambda_ {m} ^ {\\text {g e o m}} \\| \\mathcal {M} \\|, \\tag {6a}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 818, + 480, + 840 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {C D} _ {l} ^ {\\mathrm {c s}} = \\operatorname {C h a m f e r} \\left(\\mathcal {G} _ {l} ^ {\\mathrm {t g t , c s}}, \\mathcal {G} _ {l} ^ {\\mathrm {r e f}}\\right), \\tag {6b}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 842, + 480, + 862 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{Chamfer}(\\cdot)$ denotes the weighted Chamfer distance, and $\\lambda_{m}^{\\mathrm{geom}}\\| \\mathcal{M}\\|$ is a regularization term that encourages", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "smaller mobilities. Note how $\\mathcal{M}$ affects the weighted Chamfer distance by modifying the opacities, and that $\\mathcal{M} = 1$ is a trivial solution without the regularization. Without photometric supervision, the mobilities obtained from Eqn. 6a are generally noisy. However, being relatively fast (taking only tens of seconds), they still serve as a good initialization for the next sub-stage.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stage 2(b): Joint mobility and Gaussian optimization via cross-static rendering. To more accurately estimate the mobilities while jointly refining the full Gaussian attributes, we perform cross-static rendering as follows:", + "bbox": [ + 511, + 198, + 903, + 255 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathcal {G} ^ {\\mathrm {t g t}}, \\mathcal {M} _ {1 - l}} \\Delta_ {I} \\left(\\hat {I} _ {l} ^ {\\mathrm {c s}}, I _ {l}\\right) + \\lambda_ {m} ^ {\\text {p h o t o}} \\| \\mathcal {M} _ {1 - l} \\|, \\tag {7a}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 261, + 903, + 284 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {I} _ {l} ^ {\\mathrm {c s}} = R \\left(\\mathcal {G} _ {l} ^ {\\mathrm {t g t}, \\mathrm {c s}}\\right), \\tag {7b}\n$$\n", + "text_format": "latex", + "bbox": [ + 658, + 287, + 903, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda_{m}^{\\mathrm{photo}}\\| \\mathcal{M}_{1 - l}\\|$ is a mobility regularization term similar to that in Eqn. 6a.", + "bbox": [ + 511, + 314, + 903, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.3. Cross-mobile formulation for articulation estimation and mobility refinement", + "text_level": 1, + "bbox": [ + 511, + 353, + 903, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The focus of Stage 3 is to estimate the articulation parameters shared by the mobile components of all Gaussians. To this end, we employ the intuition that the mobile components from the two end states are related through the articulated motion. Formally, we introduce the cross-mobile (cm) formulation, where the mobile part of the articulated object at state $l$ is jointly represented as", + "bbox": [ + 511, + 387, + 903, + 493 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {l} ^ {\\mathrm {m}} = \\mathcal {G} _ {l} ^ {\\mathrm {t g t}} \\cdot \\mathcal {M} _ {l} \\oplus {} ^ {l} T _ {1 - l} \\left(\\mathcal {G} _ {1 - l} ^ {\\mathrm {t g t}} \\cdot \\mathcal {M} _ {1 - l}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 501, + 903, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where ${}^l T_{1 - l}(\\mathcal{G})$ denotes the transformation of Gaussians $\\mathcal{G}$ according to the articulated motion from state $1 - l$ to $l$ . For state $l$ , the target representation then becomes", + "bbox": [ + 511, + 530, + 903, + 575 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {l} ^ {\\mathrm {t g t}, \\mathrm {c m}} = \\mathcal {G} ^ {\\mathrm {s}} \\oplus \\mathcal {G} _ {l} ^ {\\mathrm {m}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 585, + 903, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{G}^{\\mathrm{s}}$ is defined in Eqn. 3. With this formulation, Stage 3 is further divided into three sub-stages as below.", + "bbox": [ + 511, + 612, + 903, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stage 3(a): Coarse articulation estimation via geometric consistency. Similar to Eqn. 6a, weighted Chamfer distance is used for computing the cross-mobile geometric consistency, which can be optimized over both mobilities and articulation parameters as follows:", + "bbox": [ + 511, + 645, + 903, + 720 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathcal {M}, T} \\mathrm {C D} _ {0} ^ {\\mathrm {c m}} + \\mathrm {C D} _ {1} ^ {\\mathrm {c m}}, \\tag {10a}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 729, + 903, + 753 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {C D} _ {l} ^ {\\mathrm {c m}} = \\operatorname {C h a m f e r} \\left(\\mathcal {G} _ {l} ^ {\\mathrm {t g t , c m}}, \\mathcal {G} _ {l} ^ {\\mathrm {r e f}}\\right). \\tag {10b}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 756, + 903, + 775 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, we notice that this formulation is still susceptible to local minima if the mobile part is too geometrically insignificant. We qualitatively show one such failure in Fig. 3a. To remedy this, we further propose mobile-only geometric consistency, which considers only the mobile components of both states as follows:", + "bbox": [ + 511, + 784, + 903, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {C D} ^ {\\mathrm {m}} = \\operatorname {C h a m f e r} \\left(\\mathcal {G} _ {l} ^ {\\mathrm {t g t}} \\cdot \\mathcal {M} _ {l}, ^ {l} T _ {1 - l} \\left(\\mathcal {G} _ {1 - l} ^ {\\mathrm {t g t}} \\cdot \\mathcal {M} _ {1 - l}\\right)\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 883, + 903, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8844", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/dc57c7c2627f33a42e1bccd935173f966d2e15276d8c88445328b5773ef508e3.jpg", + "image_caption": [ + "(a) Failure with cross-mobile geometric consistency.", + "(b) Success with mobile-only geometric consistency." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 467, + 161 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While successfully addressing these scenarios, mobile-only geometric consistency still falls short in other circumstances, especially when the mobile components in the two states exhibit large discrepancies. We qualitatively show one such failure in Fig. 4a. To take advantage of both formulations and to facilitate robustness in the inherently nonconvex optimization of articulation parameters, we propose the following practical strategy:", + "bbox": [ + 89, + 303, + 483, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "1. With $K^{\\mathrm{m}}$ randomized tries: articulation estimation via", + "bbox": [ + 91, + 424, + 470, + 436 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nT ^ {m} = \\arg \\min _ {T} \\mathrm {C D} ^ {\\mathrm {m}}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 450, + 480, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2. With $K^{\\mathrm{cm}}$ randomized tries, plus another initialized with $T^{\\mathrm{m}}$ : articulation estimation via", + "bbox": [ + 89, + 481, + 482, + 510 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nT ^ {\\mathrm {c m}} = \\arg \\min _ {T} \\mathrm {C D} _ {0} ^ {\\mathrm {c m}} + \\mathrm {C D} _ {1} ^ {\\mathrm {c m}}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 522, + 480, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3. Final run initialized with $T^{\\mathrm{cm}}$ : joint articulation estimation and mobility refinement via Eqn. 10a.", + "bbox": [ + 89, + 554, + 482, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stage 3(b): Joint articulation, mobility, and Gaussian optimization via cross-mobile rendering. As in Stage 2(b), we perform a full joint optimization over articulation parameters, mobilities, and Gaussians utilizing photometric supervision via cross-mobile rendering as follows:", + "bbox": [ + 89, + 587, + 483, + 662 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathcal {G} ^ {\\mathrm {i g t}}, \\mathcal {M} _ {1 - l}, ^ {l} T _ {1 - l}} \\Delta_ {I} \\left(\\hat {I} _ {l} ^ {\\mathrm {c m}}, I _ {l}\\right), \\tag {14a}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 672, + 480, + 698 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {I} _ {l} ^ {\\mathrm {c m}} = R \\left(\\mathcal {G} _ {l} ^ {\\mathrm {t g t , c m}}\\right), l \\in \\{0, 1 \\}. \\tag {14b}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 702, + 480, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stage 3(c): Mobility correction via cross-mobile geometric consistency. Stage 3(b) solely relies on photometric supervision, which is limited to the training views. On rare occasions, a Gaussian may be mistakenly estimated as mobile if the articulated motion moves it out of sight from most views, effectively losing supervision. On the other hand, geometric consistency is not affected by visibility limitations, which we leverage for mobility correction as follows:", + "bbox": [ + 89, + 731, + 483, + 852 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathcal {M}} \\mathrm {C D} _ {0} ^ {\\mathrm {c m}} + \\mathrm {C D} _ {1} ^ {\\mathrm {c m}}. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 863, + 482, + 885 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e05432a5de3057a0ce9daff2e3a02637f34a98379f58e3a003239a47c4a616da.jpg", + "image_caption": [ + "Figure 3. 103549-Toaster has a flat slider as its mobile part, similar in geometric curvature to the static component. This similarity makes the nearest-neighbor data association error-prone when computing the Chamfer distance for cross-mobile geometric consistency. On the other hand, mobile-only geometric consistency focuses only on the mobile part and successfully handles this case.", + "(a) Failure with mobile-only geometric consistency.", + "(b) Success with cross-mobile geometric consistency.", + "Figure 4. 101713-Pen has both ends of a pen as the mobile part, with only one end visible from any given observation. Thus, the reconstructed mobile components from the two states essentially represent distinct parts, violating the intuition for mobile-only geometric consistency. On the other hand, cross-mobile geometric consistency considers the joint representation from both states as a whole, successfully estimating the articulation." + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 900, + 183 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Real-World Application", + "text_level": 1, + "bbox": [ + 511, + 325, + 733, + 343 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For an articulated object to be reconstructed in the real world, we first collect two sets of RGB images, one for each articulation state, by imaging the object from a surrounding hemisphere. We then preprocess the data to determine the posed images that serve as the input to SPLART. This involves using SAM 2 [46] to perform foreground-background segmentation, which also removes dynamic contents of the scene. We then perform structure from motion (SfM) to determine camera poses using COLMAP [48] with SuperPoint [7] descriptors and SuperGlue [47] matching on the segmented backgrounds from both sets of images to construct a joint coordinate frame for the object in both states. Once we obtain the joint coordinate frame and the foreground target object is localized, we run SPLART to reconstruct the articulated objects.", + "bbox": [ + 511, + 349, + 906, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 590, + 643, + 607 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 511, + 616, + 614, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "PARIS PartNet-Mobility Subset. Our evaluation includes PARIS [30] as a baseline, which selects 10 instances (8 revolute, 2 prismatic) from PartNet-Mobility [2, 39, 61], a large-scale dataset of simulatable 3D objects with part-level mobility. We refer to this subset as PARIS-PMS. For each articulation state, the dataset provides 100 calibrated object-centric views for training and 50 for testing, sampled from the upper hemisphere. However, the released dataset lacks ground-truth depth and part segmentation maps. To enable their evaluations, we augment PARIS-PMS with the necessary ground-truth data by following its data generation procedure. The dataset does not provide test views of intermediate articulation states, which limits the quantitative evaluation of novel articulation synthesis.", + "bbox": [ + 511, + 638, + 906, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SPLART PartNet-Mobility Subset. We curate an additional articulated object dataset from PartNet-Mobility, dubbed SPLART-PMS, to enable a more comprehensive", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8845", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0f83d2a657f7430bae79d483a2a850dadeb59ec6b75e71ca720e998daf835123.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TypeMethodSuccess Rate %↑erra(x10-2DEG)↓errp(x10-3)↓errr(x10-2DEG)↓errt(x10-3)↓CDs(x10-3)↓CDm(x10-3)↓CDw(x10-3)↓
RevolutePARIS40.0121.90 ± 68.455.27 ± 4.55127.80 ± 55.12N/A3.48 ± 0.3968.44 ± 13.0313.11 ± 1.16
DTA†98.812.32 ± 4.161.86 ± 1.6919.41 ± 16.28N/A2.20 ± 0.491.27 ± 1.871.73 ± 0.03
SPLART100.03.70 ± 0.380.40 ± 0.074.85 ± 0.33N/A4.08 ± 0.321.06 ± 0.783.59 ± 0.21
PrismaticPARIS50.027.97 ± 13.09N/AN/A4.28 ± 3.029.21 ± 1.94151.78 ± 35.007.99 ± 0.49
DTA†100.016.26 ± 3.54N/AN/A1.15 ± 0.142.69 ± 0.0415.74 ± 0.282.22 ± 0.03
SPLART95.02.36 ± 0.44N/AN/A0.33 ± 0.046.02 ± 0.2016.77 ± 1.143.69 ± 0.14
OverallPARIS42.0103.12 ± 57.385.27 ± 4.55127.80 ± 55.124.28 ± 3.024.63 ± 0.7085.11 ± 17.4212.09 ± 1.03
DTA†99.013.11 ± 4.041.86 ± 1.6919.41 ± 16.281.15 ± 0.142.30 ± 0.404.16 ± 1.551.83 ± 0.03
SPLART99.03.44 ± 0.390.40 ± 0.074.85 ± 0.330.33 ± 0.044.47 ± 0.304.20 ± 0.853.61 ± 0.19
", + "bbox": [ + 94, + 88, + 903, + 233 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/97a57adb50ed9030444d0d0a0d999e1878348d4d404bf2f013e059411f635bd8.jpg", + "table_caption": [ + "Table 1. PARIS-PMS Articulation and Mesh Reconstruction Metrics. Best results in bold. ${}^{ \\dagger }$ DTA requires ground-truth depth." + ], + "table_footnote": [], + "table_body": "
TypeMethodPSNR ↑Depth MAE ↓mIoU ↑
RevolutePARIS32.210.0930.955
DTA†N/A0.0310.941
SPLART43.530.0390.974
PrismaticPARIS33.750.1080.902
DTA†N/A0.0660.844
SPLART44.380.0250.890
OverallPARIS32.520.0960.942
DTA†N/A0.0380.922
SPLART43.700.0360.957
", + "bbox": [ + 94, + 263, + 480, + 433 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. PARIS-PMS Novel View Synthesis Metrics, averaged over successful runs (see Tab. 1 for success rates). Best results in bold. ${}^{ \\dagger }$ DTA requires ground-truth depth.", + "bbox": [ + 89, + 436, + 482, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "evaluation including the quantitative assessment of novel articulation synthesis. Our dataset consists of 22 object instances from categories not included in PARIS-PMS, with 12 revolute and 10 prismatic objects. For the test split, we also generate ground-truth depth and part segmentation maps to facilitate relevant evaluations. Crucially, while PARIS-PMS restricts articulation states to binary values, SPLART-PMS uniformly samples states from $[-0.1, 1.1]$ , demanding accurate estimation of articulation, part segmentation, and object reconstruction for correct view synthesis. Our experiments reveal that SPLART-PMS poses a greater challenge than PARIS-PMS—the latter being almost perfectly solved by our method.", + "bbox": [ + 88, + 489, + 482, + 685 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Real-World Dataset. To show the efficacy of SPLART on real-world data, we collect a dataset that consists of 7 real-world objects with 9 articulations in total, including common objects like a monitor, cabinet, glasses, etc. All images are captured by a hand-held phone camera. Camera parameters are calibrated following strategies detailed in Sec. 3.3. We qualitatively evaluate SPLART on this dataset.", + "bbox": [ + 89, + 686, + 482, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 89, + 803, + 274, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Articulation Estimation. We evaluate the accuracy of articulation estimation depending on its type.", + "bbox": [ + 89, + 825, + 482, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- A revolute articulation describes a rotation around some line in 3D space, parameterized by the pivot point $\\mathbf{p}$ on the line, the line's directional axis $\\mathbf{a}$ , and the angle of", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ed97e3e78886aff688c7f20c9ab3d22384ec4fa1f464cfe9a66b9e987b33960e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 513, + 263, + 906, + 479 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b06e4b21ac42b7874d2a098070d890b030fe3091729386227ace9bba6f167263.jpg", + "table_caption": [ + "Figure 5. Real-World Articulated Object Reconstructions." + ], + "table_footnote": [], + "table_body": "
TypeMethodSuccess Rate %errn(×10-2DEG) ↓errp(×10-3) ↓errr(×10-2DEG) ↓errt(×10-3) ↓
RevolutePARIS18.3167.34 ± 42.479.36 ± 3.73270.37 ± 91.19N/A
DTA†33.332.76 ± 2.381.28 ± 0.3940.80 ± 4.66N/A
SPLART-2a78.52.91 ± 0.552.53 ± 2.1852.06 ± 32.00N/A
SPLART-2b67.04.62 ± 1.693.92 ± 3.07131.99 ± 64.45N/A
SPLART-3a0.8296.61 ± 0.002.83 ± 0.00427.10 ± 0.00N/A
SPLART-3b76.784.13 ± 19.135.11 ± 3.03206.98 ± 56.30N/A
SPLART77.520.76 ± 0.813.93 ± 2.4178.31 ± 31.06N/A
PrismaticPARIS39.049.49 ± 18.92N/AN/A14.96 ± 5.51
DTA†90.0102.77 ± 12.48N/AN/A9.25 ± 0.73
SPLART-2a90.025.88 ± 23.25N/AN/A1.50 ± 1.63
SPLART-2b85.715.63 ± 20.40N/AN/A0.73 ± 0.58
SPLART-3a13.015.17 ± 0.59N/AN/A2.26 ± 0.04
SPLART-3b86.061.00 ± 19.76N/AN/A9.43 ± 2.37
SPLART95.025.72 ± 17.63N/AN/A1.04 ± 0.32
OverallPARIS27.7103.88 ± 29.799.36 ± 3.73270.37 ± 91.1914.96 ± 5.51
DTA†59.177.76 ± 8.871.28 ± 0.3940.80 ± 4.669.25 ± 0.73
SPLART-2a83.713.85 ± 11.362.53 ± 2.1852.06 ± 32.001.50 ± 1.63
SPLART-2b74.79.34 ± 9.713.92 ± 3.07131.99 ± 64.450.73 ± 0.58
SPLART-3a6.485.53 ± 0.442.83 ± 0.00427.10 ± 0.002.26 ± 0.04
SPLART-3b80.973.62 ± 19.425.11 ± 3.03206.98 ± 56.309.43 ± 2.37
SPLART85.523.02 ± 8.463.93 ± 2.4178.31 ± 31.061.04 ± 0.32
", + "bbox": [ + 517, + 515, + 903, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. SPLART-PMS Articulation Metrics. Best results in bold. ${}^{ \\dagger }$ DTA requires ground-truth depth.", + "bbox": [ + 513, + 738, + 903, + 767 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "rotation $\\theta$ . We report the angular error $\\mathrm{err}_{\\mathbf{a}}$ (in $\\times 10^{-2}$ DEG) between the predicted and ground-truth axes, the geodesic distance $\\mathrm{err}_{\\mathrm{r}}$ (in $\\times 10^{-2}$ DEG) between the predicted and ground-truth rotations induced by the axis-angle ( $\\mathbf{a} - \\theta$ ) pair, and the pivot point error $\\mathrm{err}_{\\mathbf{p}}$ . Since the pivot can move arbitrarily along the axis, $\\mathrm{err}_{\\mathbf{p}}$ is computed as the closest distance between the predicted and ground-truth lines induced by the axis-pivot ( $\\mathbf{a} - \\mathbf{p}$ ) pair.", + "bbox": [ + 524, + 779, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8846", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8ee028dbb6801161217d0a59f87602dcd9013e20416dad3198293dcd28fb2e46.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TypeMethodPSNR ↑Depth MAE ↓mIoU ↑
RevolutePARIS32.750.1250.897
DTA†N/A0.0530.886
SPLART36.380.0320.897
PrismaticPARIS32.300.1240.902
DTA†N/A0.0600.857
SPLART37.770.0330.938
OverallPARIS32.510.1250.900
DTA†N/A0.0580.867
SPLART37.010.0320.916
", + "bbox": [ + 96, + 88, + 478, + 253 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. SPLART-PMS Novel View and Articulation Synthesis Metrics, averaged over successful runs (see Tab. 3 for success rates). Best results in bold. $\\dagger$ DTA requires ground-truth depth.", + "bbox": [ + 89, + 260, + 482, + 305 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- A prismatic articulation describes a translation along a specific direction, parameterized by the axis $\\mathbf{a}$ of the translation direction and the translation distance $d$ . We report the axis error $\\mathrm{err}_{\\mathbf{a}}$ , as in the revolute case, and the translation error $\\mathrm{err}_{\\mathrm{t}}$ , which is the distance between the predicted and ground-truth translations induced by the axis-distance $(\\mathbf{a} - d)$ pair.", + "bbox": [ + 89, + 313, + 483, + 420 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Part-Level Reconstruction. The evaluation of part-level reconstruction accuracy is threefold: photometric rendering quality, geometric accuracy, and part segmentation accuracy. For all these aspects, we use novel view synthesis as the surrogate task. We perform volume rendering for each view in the test split, generating outputs that include the RGB image, depth map, and part segmentation map. Photometric rendering quality is evaluated by reporting image metrics PSNR. Geometric accuracy is assessed by reporting the mean absolute error of the depth map (i.e., Depth MAE). Part segmentation accuracy is evaluated using the intersection-over-union (IoU) ratio for three categories: static part $(\\mathrm{IoU_s})$ , mobile part $(\\mathrm{IoU_m})$ , and background $(\\mathrm{IoU_{bg}})$ . The mean IoU (mIoU) is then reported as the average IoU across these categories. In addition, we also evaluate geometric accuracy through mesh reconstruction. To extract a mesh from 3DGS, we: (1) render depth images from uniformly sampled spherical viewpoints, (2) fuse rendered depths into a TSDF representation [4, 41], and (3) extract the mesh using the Marching Cubes algorithm [34]. For each reconstruction, we separately extract the meshes corresponding to the static, mobile, and whole parts. For evaluation, we follow the procedure used in previous works [30, 60]: 10 000 points are uniformly sampled from both the reconstructed and the ground-truth meshes, and the Chamfer distance is computed for each category—static $(\\mathrm{CD_s})$ , mobile $(\\mathrm{CD_m})$ , and whole $(\\mathrm{CD_w})$ .", + "bbox": [ + 91, + 426, + 483, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of Randomness. Randomness can play a significant role in the joint optimization of articulation and part reconstruction [6, 30]. In order to effectively demonstrate the robustness of a method, we conduct ten trials for each", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2fc06219444e5914e71f5ae360c285e9c2734e3be1238f7c20584afe6651661b.jpg", + "image_caption": [ + "Figure 6. State interpolation from the estimated articulation." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 905, + 210 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "scene, each with a different fixed random seed. We report results averaged over the ten runs. Additionally, we notice that for a failed run converging to a local minimum, the evaluation metrics, such as $\\mathrm{err}_{\\mathbf{r}}$ , are typically orders-of-magnitude worse than those of a successful run. To better reflect a method's performance, we report the number of successful runs out of the total number (i.e., ten), where a run is deemed successful if and only if each of the following criteria are met: (1) $\\mathrm{err}_{\\mathbf{a}} < 5$ DEG; (2) $\\mathrm{err}_{\\mathbf{p}} < 0.05$ for revolute articulation; (3) $\\mathrm{err}_{\\mathbf{r}} < 10$ DEG for revolute articulation; and (4) $\\mathrm{err}_{\\mathbf{t}} < 0.05$ for prismatic articulation. We then proceed to report the evaluation metrics detailed previously, averaging only over the successful runs.", + "bbox": [ + 511, + 244, + 906, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Baselines", + "text_level": 1, + "bbox": [ + 513, + 450, + 622, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "PARIS [30] addresses the same challenge we tackle: reconstructing the part-level shape, appearance, and motion of an articulated object from multi-view RGB images captured at two articulation states. It employs a NeRF-based representation, modeling the object with static and mobile fields for each part, alongside a transformation to capture state transitions. These components are jointly optimized end-to-end, primarily through an image-rendering loss.", + "bbox": [ + 511, + 473, + 905, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DTA [60] adopts a similar setup: reconstructing digital twins of articulated objects from multi-view RGB-D observations at two articulation states. It operates in two stages: Stage 1 optimizes neural object fields for each state using RGB-D images and extracts meshes; Stage 2 optimizes a part segmentation field and per-part motions using consistency, matching, and collision losses to determine point correspondences between states. Note: Unlike SPLART, DTA requires ground-truth depth as input and produces plain object meshes unsuitable for photorealistic rendering.", + "bbox": [ + 511, + 597, + 905, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Experiment Results", + "text_level": 1, + "bbox": [ + 511, + 757, + 702, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Articulation Estimation. We quantitatively evaluate SPLART against baselines for articulation estimation accuracy, reporting scene-averaged results per articulation type on the PARIS-PMS dataset (Tab. 1) and the SPLART-PMS dataset (Tab. 3). SPLART consistently outperforms the baselines across both datasets. Additionally, we qualitatively compare SPLART with PARIS on selected scenes from these datasets in Fig. 7. In the \"color\" column, each", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8847", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bae55b184d543a19622de379ebc3e97a12b08962128574169594e36aba0b8bb2.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison on part-level reconstruction and articulation estimation." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 906, + 286 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "image visualizes the ground-truth articulation (green arrow) and the estimated articulation (red arrow), except in the ground-truth column. When the estimate aligns perfectly with the ground truth, only the red arrow appears; if significantly misaligned, only the green arrow is visible.", + "bbox": [ + 89, + 315, + 483, + 391 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Part-Level Reconstruction. We assess SPLART's part-level reconstruction accuracy against baselines, using novel view synthesis as a surrogate task. Scene-averaged results per articulation type are reported on the PARIS-PMS dataset (Tab. 2) and the SPLART-PMS dataset (Tab. 4). We also evaluate geometric accuracy through mesh reconstruction, presenting scene-averaged results per articulation type on the PARIS-PMS dataset (Tab. 1). Without depth supervision, which is required for the DTA baseline, SPLART matches DTA's performance, both substantially exceeding PARIS. Qualitative comparisons with PARIS, including RGB renderings and part segmentation maps, are provided for selected scenes from both datasets in Fig. 7. Additional qualitative results on articulation synthesis appear in Fig. 6.", + "bbox": [ + 89, + 404, + 483, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Real-World Reconstructions. We qualitatively evaluate SPLART on the aforementioned real-world dataset, with results shown in Fig. 5. Each image overlays the estimated articulation, visualized with a red arrow.", + "bbox": [ + 89, + 628, + 483, + 690 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Studies. We perform extensive ablation studies to demonstrate the contribution of each stage. Specifically, we perform the following four ablations: (1) No geometric supervision for mobility estimation, i.e., skipping Stage 2(a), denoted SPLART-2a; (2) No photometric supervision for mobility estimation, i.e., skipping Stage 2(b), denoted SPLART-2b; (3) No geometric supervision for articulation estimation, i.e., skipping Stage 3(a), denoted SPLART-3a; (4) No photometric supervision for articulation estimation, i.e., skipping Stage 3(b), denoted SPLART-3b. Like the main experiments, we conduct 10 runs for each scene, but only on the more challenging dataset of SPLART-PMS. We report the articulation estimation results in Tab. 3.", + "bbox": [ + 89, + 704, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0b517baf1144dfd9df63c40504710b4fc005b4854d23a0fd7254f1f8780bfc26.jpg", + "image_caption": [ + "Figure 8. Failure cases due to inherent ambiguities in articulation." + ], + "image_footnote": [], + "bbox": [ + 514, + 311, + 906, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 426, + 633, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present SPLART, the first self-supervised method that utilizes 3D Gaussian Splatting for articulated object reconstruction from two-state RGB observations, without 3D supervision or category-specific priors. SPLART delivers robust optimization and effectively handles challenging scenarios—outperforming prior approaches while eliminating the need for 3D supervision, articulation annotations, or semantic labels. Extensive evaluations on synthetic and real-world datasets show SPLART's articulation estimation accuracy and view synthesis quality that surpass existing methods. Real-world experiments further validate its practicality, showcasing successful reconstructions of diverse articulated objects using only a handheld RGB camera. Our framework empowers non-expert users to effortlessly create high-fidelity, articulated digital twins, supporting applications in augmented/virtual reality and robotics. However, while SPLART effectively addresses two-part articulations, it lacks direct support for objects with more articulated parts, though iteratively applying it to incremental state pairs may help. Some existing methods [6, 33, 60] handle such objects, but require known part counts and specific articulation structures. Future research will focus on a comprehensive adaptation to multi-level articulated objects. Additionally, inherent ambiguities may arise when multiple articulations explain state differences, as shown in Fig. 8 with examples like a clock's hand rotating on an unexpected axis or a globe's surface spinning independently. Overall, SPLART represents a scalable, robust, and practical solution that pushes the boundaries of articulated object reconstruction, laying a strong foundation for future work.", + "bbox": [ + 511, + 446, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8848", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ben Abbatematteo, Stefanie TELlex, and George Konidaris. Learning to generalize kinematic models to novel objects. In Proceedings of the Conference on Robot Learning (CoRL), pages 1289-1299, 2019. 1", + "[2] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. ShapeNet: An information-rich 3D model repository. arXiv preprint arXiv:1512.03012, 2015. 5", + "[3] Christopher B. Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3D-R2N2: A unified approach for single and multi-view 3D object reconstruction. In Proceedings of the European Conference on Computer Vision (ECCV), pages 628-644, 2016. 2", + "[4] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), pages 303-312, 1996. 7", + "[5] Andrea F. Daniele, Thomas M. Howard, and Matthew R. Walter. A multiview approach to learning articulated motion models. In Proceedings of the International Symposium of Robotics Research (ISRR), pages 371-386, 2017. 1", + "[6] Jianning Deng, Kartic Subr, and Hakan Bilen. Articulate your NeRF: Unsupervised articulated object modeling via conditional view synthesis. In Advances in Neural Information Processing Systems (NeurIPS), pages 119717-119741, 2025. 1, 2, 3, 7, 8", + "[7] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 224-236, 2018. 5", + "[8] Haoqiang Fan, Hao Su, and Leonidas J. Guibas. A point set generation network for 3D object reconstruction from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2463-2471, 2017. 2", + "[9] Samir Yitzhak Gadre, Kiana Ehsani, and Shuran Song. Act the part: Learning interaction strategies for articulated object part discovery. In Proceedings of the International Conference on Computer Vision (ICCV), pages 15732-15741, 2021. 2", + "[10] Haoran Geng, Ziming Li, Yiran Geng, Jiayi Chen, Hao Dong, and He Wang. PartManip: Learning cross-category generalizable part manipulation policy from point cloud observations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2978-2988, 2023. 2", + "[11] Haoran Geng, Helin Xu, Chengyang Zhao, Chao Xu, Li Yi, Siyuan Huang, and He Wang. GAPartNet: Cross-category domain-generalizable object perception and manipulation via generalizable and actionable parts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7081-7091, 2023. 2" + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Antoine Guédon and Vincent Lepetit. SuGaR: Surface-aligned Gaussian splattering for efficient 3D mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5354-5363, 2024. 2", + "[13] Nick Heppert, Muhammad Zubair Irshad, Sergey Zakharov, Katherine Liu, Rares Andrei Ambrus, Jeannette Bohg, Abhinav Valada, and Thomas Kollar. CARTO: Category and joint agnostic reconstruction of articulated objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21201-21210, 2023. 1, 2", + "[14] Cheng-Chun Hsu, Zhenyu Jiang, and Yuke Zhu. Ditto in the house: Building articulation models of indoor scenes through interactive perception. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 3933-3939, 2023. 2", + "[15] Ruizhen Hu, Wenchao Li, Oliver Van Kaick, Ariel Shamir, Hao Zhang, and Hui Huang. Learning to predict part mobility from a single static snapshot. ACM Transactions On Graphics, 36(6):1-13, 2017. 2", + "[16] Xiaoxia Huang, Ian Walker, and Stan Birchfield. Occlusion-aware reconstruction and manipulation of 3D articulated objects. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 1365-1371, 2012. 1", + "[17] Muhammad Zubair Irshad, Thomas Kollar, Michael Laskey, Kevin Stone, and Zsolt Kira. CenterSnap: Single-shot multi-object 3D shape reconstruction and categorical 6D pose and size estimation. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 10632-10640, 2022. 2", + "[18] Muhammad Zubair Irshad, Sergey Zakharov, Rares Ambrus, Thomas Kollar, Zsolt Kira, and Adrien Gaidon. ShAPO: Implicit representations for multi-object shape, appearance, and pose optimization. In Proceedings of the European Conference on Computer Vision (ECCV), pages 275–292, 2022. 2", + "[19] Muhammad Zubair Irshad, Mauro Comi, Yen-Chen Lin, Nick Heppert, Abhinav Valada, Rares Ambrus, Zsolt Kira, and Jonathan Tremblay. Neural fields in robotics: A survey, 2024. 2", + "[20] Ajinkya Jain, Rudolf Lioutikov, Caleb Chuck, and Scott Niekum. ScrewNet: Category-independent articulation model estimation from depth images using screw theory. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 13670-13677, 2021. 1", + "[21] Anil K. Jain. Fundamentals of Digital Image Processing. Prentice-Hall, Inc., 1989. 2", + "[22] Hanxiao Jiang, Yongsen Mao, Manolis Savva, and Angel X. Chang. OPD: Single-view 3D openable part detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 410-426, 2022. 2", + "[23] Zhenyu Jiang, Cheng-Chun Hsu, and Yuke Zhu. Ditto: Building digital twins of articulated objects from interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5606-5616, 2022. 1, 2" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8849", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Dov Katz, Moslem Kazemi, J. Andrew Bagnell, and Anthony Stentz. Interactive segmentation, tracking, and kinematic modeling of unknown 3D articulated objects. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 5003-5010, 2013. 1", + "[25] Yuki Kawana, Yusuke Mukuta, and Tatsuya Harada. Unsupervised pose-aware part decomposition for man-made articulated objects. In Proceedings of the European Conference on Computer Vision (ECCV), pages 558-575, 2022. 2", + "[26] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4):1-14, 2023. 1, 2", + "[27] Justin Kerr, Chung Min Kim, Mingxuan Wu, Brent Yi, Qianqian Wang, Ken Goldberg, and Angjoo Kanazawa. Robot see robot do: Imitating articulated object manipulation with monocular 4d reconstruction. In Proceedings of the Conference on Robot Learning (CoRL), pages 587-603, 2024. 2", + "[28] Long Le, Jason Xie, William Liang, Hung-Ju Wang, Yue Yang, Yecheng Jason Ma, Kyle Vedder, Arjun Krishna, Dinesh Jayaraman, and Eric Eaton. Articulate-anything: Automatic modeling of articulated objects via a vision-language foundation model. In Proceedings of the International Conference on Learning Representations (ICLR), 2025. 2", + "[29] Xiaolong Li, He Wang, Li Yi, Leonidas J. Guibas, A. Lynn Abbott, and Shuran Song. Category-level articulated object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3703-3712, 2020. 1, 2", + "[30] Jiayi Liu, Ali Mahdavi-Amiri, and Manolis Savva. PARIS: Part-level reconstruction and motion analysis for articulated objects. In Proceedings of the International Conference on Computer Vision (ICCV), pages 352-363, 2023. 1, 2, 3, 5, 7", + "[31] Liu Liu, Wenqiang Xu, Haoyuan Fu, Sucheng Qian, Qiao-jun Yu, Yang Han, and Cewu Lu. A real-world articulated object knowledge base. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14789-14798, 2022. 2", + "[32] Shaowei Liu, Saurabh Gupta, and Shenlong Wang. Building rearticulable models for arbitrary 3D objects from 4D point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21138-21147, 2023. 2", + "[33] Yu Liu, Baoxiong Jia, Ruijie Lu, Junfeng Ni, Song-Chun Zhu, and Siyuan Huang. Building interactable replicas of complex articulated objects via gaussian splatting. In Proceedings of the International Conference on Learning Representations (ICLR), 2025. 1, 2, 8", + "[34] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3D surface construction algorithm. In *Seminal Graphics: Pioneering Efforts That Shaped the Field*, pages 347–353, 1998. 7", + "[35] Zhao Mandi, Yijia Weng, Dominik Bauer, and Shuran Song. Real2code: Reconstruct articulated objects via code generation. In Proceedings of the International Conference on Learning Representations (ICLR), 2025. 2", + "[36] Eric Marchand, Hideaki Uchiyama, and Fabien Spindler. Pose estimation for augmented reality: A hands-on survey." + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE Transactions on Visualization and Computer Graphics, 22(12):2633-2651, 2016. 1", + "[37] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3D reconstruction in function space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4460-4470, 2019. 2", + "[38] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 1", + "[39] Kaichun Mo, Shilin Zhu, Angel X. Chang, Li Yi, Subarna Tripathi, Leonidas J. Guibas, and Hao Su. PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 909-918, 2019. 5", + "[40] Jiteng Mu, Weichao Qiu, Adam Kortylewski, Alan Yuille, Nuno Vasconcelos, and Xiaolong Wang. A-SDF: Learning disentangled signed distance functions for articulated shape representation. In Proceedings of the International Conference on Computer Vision (ICCV), pages 12981-12991, 2021. 1, 2", + "[41] Richard A. Newcombe, Andrew Fitzgibbon, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J. Davison, Pushmeet Kohi, Jamie Shotton, and Steve Hodges. KinectFusion: Real-time dense surface mapping and tracking. In Proceedings of the IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 127-136, 2011. 7", + "[42] Neil Nie, Samir Yitzhak Gadre, Kiana Ehsani, and Shuran Song. Structure from action: Learning interactions for articulated object 3D structure discovery, 2022. 2", + "[43] Michael Oechsle, Songyou Peng, and Andreas Geiger. UNISURF: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the International Conference on Computer Vision (ICCV), pages 5589-5599, 2021. 2", + "[44] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 165-174, 2019. 2", + "[45] Sudeep Pillai, Matthew R. Walter, and Seth Teller. Learning articulated motions from visual demonstrations. In Proceedings of Robotics: Science and Systems (RSS), 2014. 1", + "[46] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross B. Girshick, Piotr Dólar, and Christoph Feichtenhofer. SAM 2: Segment anything in images and videos. In Proceedings of the International Conference on Learning Representations (ICLR), 2025. 5", + "[47] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8850", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "matching with graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4937-4946, 2020. 5", + "[48] Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4104-4113, 2016. 5", + "[49] Chaoyue Song, Jiacheng Wei, Chuan Sheng Foo, Guosheng Lin, and Fayao Liu. REACTO: Reconstructing articulated objects from a single video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5384-5395, 2024. 2", + "[50] Jürgen Sturm, Cyril Stachniss, and Wolfram Burgard. A probabilistic framework for learning kinematic models of articulated objects. Journal of Artificial Intelligence Research, 41:477-526, 2011. 1", + "[51] Ryo Suzuki, Adnan Karim, Tian Xia, Hooman Hedayati, and Nicolai Marquardt. Augmented reality and robotics: A survey and taxonomy for ar-enhanced human-robot interaction and robotic interfaces. In Proceedings of the Chi Conference on Human Factors in Computing Systems, pages 1–33, 2022. 1", + "[52] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3D shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11353-11362, 2021. 2", + "[53] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Justin Kerr, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, David McAllister, and Angjoo Kanazawa. Nerfstudio: A modular framework for neural radiance field development. In Proceedings of the International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), pages 1-12, 2023. 1", + "[54] Wei-Cheng Tseng, Hung-Ju Liao, Lin Yen-Chen, and Min Sun. CLA-NeRF: Category-level articulated neural radiance field. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 8454–8460, 2022. 1", + "[55] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In Advances in Neural Information Processing Systems (NeurIPS), pages 27171-27183, 2021. 2", + "[56] Xiaogang Wang, Bin Zhou, Yahao Shi, Xiaowu Chen, Qinping Zhao, and Kai Xu. Shape2Motion: Joint analysis of motion parts and attributes from 3D shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8868-8876, 2019. 1, 2" + ], + "bbox": [ + 91, + 92, + 483, + 811 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Xian Wang, Luyao Shen, and Lik-Hang Lee. A systematic review of XR-based remote human-robot interaction systems. ACM Computing Surveys, 57(11):1-37, 2025. 1", + "[58] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 2", + "[59] Fangyin Wei, Rohan Chabra, Lingni Ma, Christoph Lassner, Michael Zollhöfer, Szymon Rusinkiewicz, Chris Sweeney, Richard Newcombe, and Mira Slavcheva. Self-supervised neural articulated shape and appearance models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15795-15805, 2022. 1, 2", + "[60] Yijia Weng, Bowen Wen, Jonathan Tremblay, Valts Blukis, Dieter Fox, Leonidas Guibas, and Stan Birchfield. Neural implicit representation for building digital twins of unknown articulated objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3141-3150, 2024. 1, 2, 7, 8", + "[61] Fanbo Xiang, Yuzhe Qin, Kaichun Mo, Yikuan Xia, Hao Zhu, Fangchen Liu, Minghua Liu, Hanxiao Jiang, Yifu Yuan, He Wang, Li Yi, Angel X. Chang, Leonidas J. Guibas, and Hao Su. SAPIEN: A simulated part-based interactive environment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11094–11104, 2020. 5", + "[62] Zihao Yan, Ruizhen Hu, Xingguang Yan, Luanmin Chen, Oliver van Kaick, Hao Zhang, and Hui Huang. RPM-net: Recurrent prediction of motion and parts from point cloud. ACM Transactions On Graphics, 38(6):240:1-240:15, 2019. 2", + "[63] Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jeffrey Hu, Matthew Tancik, and Angjoo Kanazawa. Gsplat: An open-source library for Gaussian splattering, 2024. 1", + "[64] Li Yi, Haibin Huang, Difan Liu, Evangelos Kalogerakis, Hao Su, and Leonidas Guibas. Deep part induction from articulated object pairs. ACM Transactions on Graphics, 37(6): 1-15, 2018. 2", + "[65] Sergey Zakharov, Katherine Liu, Adrien Gaidon, and Rares Ambrus. ReFiNe: Recursive field networks for cross-modal multi-scene representation. In Proceedings of the International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), pages 1-9, 2024. 2", + "[66] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 2" + ], + "bbox": [ + 516, + 92, + 906, + 811 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8851", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SplArt_ Articulation Estimation and Part-Level Reconstruction with 3D Gaussian Splatting/99f05b1c-e917-486f-913a-4812f9c2d857_model.json b/2025/SplArt_ Articulation Estimation and Part-Level Reconstruction with 3D Gaussian Splatting/99f05b1c-e917-486f-913a-4812f9c2d857_model.json new file mode 100644 index 0000000000000000000000000000000000000000..306fdaa1e0aa8da50b894bc334576de76195cc0c --- /dev/null +++ b/2025/SplArt_ Articulation Estimation and Part-Level Reconstruction with 3D Gaussian Splatting/99f05b1c-e917-486f-913a-4812f9c2d857_model.json @@ -0,0 +1,2730 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.17, + 0.131, + 0.83, + 0.178 + ], + "angle": 0, + "content": "SPLART: Articulation Estimation and Part-Level Reconstruction with 3D Gaussian Splatting" + }, + { + "type": "text", + "bbox": [ + 0.168, + 0.203, + 0.862, + 0.24 + ], + "angle": 0, + "content": "Shengjie Lin†, Jiading Fang†, Muhammad Zubair Irshad‡, Vitor Campagnolo Guizilini‡, Rares Andrei Ambrus‡, Greg Shakhnarovich†‡, Matthew R. Walter†" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.239, + 0.799, + 0.258 + ], + "angle": 0, + "content": "† Toyota Technological Institute at Chicago, ‡ Toyota Research Institute" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.26, + 0.928, + 0.276 + ], + "angle": 0, + "content": "{slin, fjd, greg, mwalter}@ttic.edu, {zubair.irshad, vitor.quizilini, rares.ambrus}@tri.global" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.31, + 0.327, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.342, + 0.485, + 0.721 + ], + "angle": 0, + "content": "Reconstructing articulated objects prevalent in daily environments is crucial for applications in augmented/virtual reality and robotics. However, existing methods face scalability limitations (requiring 3D supervision or costly annotations), robustness issues (being susceptible to local optima), and rendering shortcomings (lacking speed or photorealism). We introduce SPLART, a self-supervised, category-agnostic framework that uses 3D Gaussian Splattig (3DGS) to reconstruct and infer the kinematics of articulated objects from two sets of posed RGB images captured at different articulation states, enabling real-time photorealistic rendering for novel viewpoints and articulations. SPLART augments 3DGS with a differentiable mobility parameter per Gaussian, achieving refined part segmentation. A multi-stage optimization strategy is employed to progressively handle reconstruction, part segmentation, and articulation estimation, significantly enhancing robustness and accuracy. SPLART exploits geometric self-supervision, effectively addressing challenging scenarios without requiring 3D annotations or category-specific priors. Evaluations on established and newly proposed benchmarks, along with applications to real-world scenarios using a handheld RGB camera, demonstrate SPLART's state-of-the-art performance and real-world practicality. Code is publicly available at https://github.com/ripl/splart." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.74, + 0.222, + 0.755 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Articulated objects, such as drawers, doors, and scissors, are ubiquitous in our daily lives, yet their dynamic nature poses significant challenges for 3D reconstruction—a critical task for applications in augmented/virtual reality [36, 51], robotics [1, 5, 16, 24, 50, 57], and computer vision [20, 45]. Existing methods for reconstructing articulated objects are hindered by several key limitations: they often require labor-intensive supervision (e.g., part-level segmentation or articulation annotations) [23, 29," + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.308, + 0.697, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.309, + 0.899, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.424, + 0.607, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.438, + 0.764, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.78, + 0.462, + 0.882, + 0.487 + ], + "angle": 0, + "content": "SPLART" + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.52, + 0.607, + 0.6 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.603, + 0.598, + 0.613 + ], + "angle": 0, + "content": "Articulation" + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.538, + 0.701, + 0.6 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.539, + 0.795, + 0.6 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.796, + 0.54, + 0.89, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.626, + 0.907, + 0.71 + ], + "angle": 0, + "content": "Figure 1. Given (top) RGB images of an object at two articulation states, (middle) SPLART uses 3DGS to simultaneously reconstruct its static and dynamic parts and estimate the kinematic articulation model. SPLART is then able to (bottom) render high-fidelity 3D reconstructions of the object along with part-level segmentations for novel articulation states, allowing for novel view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "40, 54, 56], they depend on 3D supervision that restricts practical use [23, 29, 33, 40, 60], they produce category-specific models that limit scalability [29, 40, 54, 59], or they are not capable of real-time, photorealistic rendering [6, 13, 23, 29, 30, 40, 54, 56, 59, 60]. To address these challenges, we introduce SPLART, a novel self-supervised and category-agnostic framework that leverages 3D Gaussian Splitting (3DGS) [26] to reconstruct articulated objects from minimal input—two sets of posed RGB images at distinct articulation states. SPLART reconstructs the object's parts and infers its kinematics, enabling real-time, photorealistic rendering for novel views and articulation states." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8841" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "Central to SPLART is the augmentation of 3DGS [26] to include a differentiable mobility parameter for each Gaussian, which enables a more refined segmentation of static and mobile parts through gradient-based optimization. This results in enhanced reconstruction quality, while preserving the real-time, photorealistic rendering capabilities of 3DGS—offering a speedup of more than \\(100\\times\\) over methods [6, 30] based on neural radiance fields [38]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.213, + 0.482, + 0.379 + ], + "angle": 0, + "content": "To enhance robustness, SPLART employs a multi-stage optimization strategy that decouples the part-level reconstruction and articulation estimation processes. Unlike end-to-end approaches prone to local optima [30], SPLART first independently reconstructs each articulation state, then estimates each Gaussian's mobility parameter for part segmentation, and finally refines both the articulation and mobility estimates jointly. This structured approach ensures stable and accurate convergence, avoiding the stringent initialization requirements of existing methods, thereby providing a practical solution for challenging articulated structures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.381, + 0.482, + 0.516 + ], + "angle": 0, + "content": "Building on this foundation, SPLART leverages geometric self-supervision to eliminate the need for manual annotations or 3D supervision. By enforcing geometric consistency between reconstructions through complementary loss formulations, SPLART robustly estimates articulation parameters across diverse scenarios. This self-supervised strategy enhances scalability, enabling SPLART to reconstruct a wide range of articulated objects without relying on prior structural or categorical knowledge." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.482, + 0.623 + ], + "angle": 0, + "content": "Extensive evaluations on both established and newly introduced benchmarks demonstrate SPLART's superior articulation accuracy and reconstruction quality, surpassing state-of-the-art methods without requiring 3D supervision. Real-world experiments further validate its practicality, showcasing successful reconstructions of diverse articulated objects using only a handheld RGB camera." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.624, + 0.343, + 0.638 + ], + "angle": 0, + "content": "In summary, this work contributes:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.64, + 0.481, + 0.685 + ], + "angle": 0, + "content": "1. An extension of 3DGS with a differentiable mobility value per Gaussian that enables precise part segmentation using gradient-based optimization." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.686, + 0.481, + 0.729 + ], + "angle": 0, + "content": "2. A multi-stage optimization strategy that decouples reconstruction and articulation estimation, enhancing robustness and accuracy." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.731, + 0.481, + 0.774 + ], + "angle": 0, + "content": "3. Complementary formulations of geometric self-supervision for articulation estimation, eliminating the need for 3D supervision or laborious annotations." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.776, + 0.481, + 0.805 + ], + "angle": 0, + "content": "4. A challenging dataset and new metrics for comprehensive evaluation of articulated object reconstruction." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.64, + 0.481, + 0.805 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.822, + 0.233, + 0.837 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.396, + 0.864 + ], + "angle": 0, + "content": "2.1. Data-Driven Articulation Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Estimating the pose and joint properties of articulated objects is crucial for robot manipulation and interac" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.304 + ], + "angle": 0, + "content": "tion [10, 11, 19, 31]. Recent learning-based methods [9, 13, 15, 22, 29, 32, 56, 64] infer articulation properties from point clouds via end-to-end training. For instance, Shape2Motion [56] analyzes motion parts from a single point cloud in a supervised setting, while ANCSH [29] performs category-level pose estimation, but requires class-specific models. RPM-Net [62] enhances generalization across categories for part segmentation and kinematic prediction, and DITTO [23] predicts motion and geometry from 3D point cloud pairs without labels. However, these methods depend on costly 3D supervision and annotations. In contrast, our approach reconstructs accurate 3D geometry and detailed appearance, capturing articulation without 3D supervision or priors." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.314, + 0.878, + 0.33 + ], + "angle": 0, + "content": "2.2. Representations for Object Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.336, + 0.905, + 0.502 + ], + "angle": 0, + "content": "Early 3D object reconstruction methods predicted point clouds, voxels, or meshes from partial observations [3, 8, 17]. Recent advances in implicit scene representations [18, 26, 37, 38, 44, 65] enable detailed geometry and appearance reconstruction via differentiable rendering [12, 43, 52, 55]. While neural fields suffer from slow rendering, 3D Gaussian Splatting (3DGS) [26] overcomes this by using explicit 3D Gaussians. We leverage 3DGS for self-supervised articulated object reconstruction from posed RGB images, achieving fast, realistic synthesis of novel views and articulations in real time." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.513, + 0.816, + 0.529 + ], + "angle": 0, + "content": "2.3. Articulated Object Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.536, + 0.905, + 0.793 + ], + "angle": 0, + "content": "Recent methods leverage differentiable 3D representations [26, 38, 44] to jointly reconstruct articulated objects and infer articulation parameters. Training-based approaches use synthetic 3D data to predict joint parameters and segment parts [13, 14, 23, 25, 40, 42, 59]. Self-supervised methods optimize shape, appearance, and articulation per scene without extensive training [6, 27, 30, 32, 33, 49, 60]. Some handle objects with more than two parts, but require the number of parts to be known and the mobile parts to be directly connected to the static part [6, 33, 60]. Other works enhance articulation estimation using large language or vision-language models [28, 35]. In contrast, our self-supervised method reconstructs two-part articulated objects from RGB images across articulation states using 3DGS [26]. As the first to apply 3DGS to this task without 3D supervision or pre-trained priors, it robustly handles challenging cases and achieves real-time performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.808, + 0.649, + 0.825 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.625, + 0.848 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Consider an arbitrary object composed of two rigid parts: a static parent part, and a child part that can move relative to its parent through either a revolute or prismatic articulation." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8842" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.122, + 0.093, + 0.241, + 0.116 + ], + "angle": 0, + "content": "(a) Stage 1: Per-state reconstruction." + }, + { + "type": "image_caption", + "bbox": [ + 0.142, + 0.117, + 0.221, + 0.128 + ], + "angle": 0, + "content": "(Section 3.2.1)" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.13, + 0.182, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.19, + 0.13, + 0.271, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.196, + 0.174, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.188, + 0.196, + 0.266, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.093, + 0.529, + 0.117 + ], + "angle": 0, + "content": "(b) Stage 2: Cross-static formulation for mobility estimation." + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.118, + 0.575, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.605, + 0.093, + 0.877, + 0.117 + ], + "angle": 0, + "content": "(c) Stage 3: Cross-mobile formulation for articulation estimation and mobility refinement." + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.118, + 0.902, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.286, + 0.907, + 0.356 + ], + "angle": 0, + "content": "Figure 2. Methodology overview of SPLART. It consists of three decoupled stages for the purpose of stable optimization. Stage 1 constructs a 3DGS for each state from posed RGB images using the photometric loss. Stage 2 proposes cross-static formulation for mobility estimation. Intuitively, it combines static parts of both states and the mobile part of the desired state for the target Gaussians. Stage 3 proposes cross-mobile formulation for articulation estimation. Intuitively, it combines static parts of both states, the mobile part of the desired state and the transformed mobile part of the other state together for the target Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.365, + 0.483, + 0.47 + ], + "angle": 0, + "content": "Our objective is twofold: (1) to reconstruct the articulated object at the part level; and (2) to estimate its articulated motion. Assuming a known articulation type (i.e., either revolute or prismatic), the input to our method consists of two sets of posed RGB images (i.e., images with known camera intrinsics and extrinsics), each capturing the articulated object at one end state of the motion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.471, + 0.484, + 0.668 + ], + "angle": 0, + "content": "Formally, let \\( l \\) denote the articulation state label, where \\( l = 0 \\) and \\( l = 1 \\) correspond to the two end states of the observed articulation. For reconstruction, SPLART uses observations \\( \\mathcal{O}_l = \\{(I_l^i,P_l^i,K_l^i)\\}_{i = 1}^{N_l}, l\\in \\{0,1\\} \\), where \\( I_{l}^{i} \\) is the \\( i \\)-th observed RGB image of the articulated object at state \\( l \\), \\( P_{l}^{i} \\) and \\( K_{l}^{i} \\) denote its camera extrinsic and intrinsic, respectively, and \\( N_{l} \\) represents the number of data samples for state \\( l \\). Note that \\( P_{l}^{i} \\) is specified in a common world frame for both states, while the articulated motion involves only one moving part w.r.t. the world space. SPLART models a revolute articulation with its rotation axis \\( \\mathbf{a} \\) (\\( \\| \\mathbf{a}\\| = 1 \\)), pivot \\( \\mathbf{p} \\), and rotation angle \\( \\theta \\), such that a point \\( \\mathbf{x} \\) (in the world frame) on the mobile part at state \\( l = 0 \\) will move to" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.677, + 0.483, + 0.695 + ], + "angle": 0, + "content": "\\[\n{ } ^ { 1 } T _ { 0 } \\mathbf { x } = R _ { \\mathbf { a } , \\theta } ( \\mathbf { x } - \\mathbf { p } ) + \\mathbf { p } \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.809 + ], + "angle": 0, + "content": "at state \\( l = 1 \\), where \\( R_{\\mathbf{a},\\theta} \\) is the rotation induced by the axis-angle notation. A prismatic articulation is instead modeled by its translation axis \\( \\mathbf{a} \\) (\\( \\| \\mathbf{a} \\| = 1 \\)) and distance \\( d \\). The goal is to reconstruct the articulated object at the part level using a chosen representation while estimating the articulated motion \\( ^1 T_0 \\), ensuring that the renderings at each articulation state are with the observations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.901 + ], + "angle": 0, + "content": "SPLART extends the 3DGS representation for articulated objects, decoupling the goals of part-level reconstruction and articulation estimation across three stages: (1) separate reconstructions for each articulation state (Sec. 3.2.1), (2) mobility estimation using the cross-static formulation (Sec. 3.2.2), and (3) articulation estimation and mobility re" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.365, + 0.907, + 0.472 + ], + "angle": 0, + "content": "finement using the cross-mobile formulation (Sec. 3.2.3). To facilitate the application of SPLART to real-world objects, we leverage modern structure-from-motion and image segmentation techniques, developing a framework that enables general users to reconstruct articulated objects in their surroundings using only images captured by a handheld camera device (Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.481, + 0.833, + 0.497 + ], + "angle": 0, + "content": "3.2. Decoupled Multi-Stage Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.503, + 0.906, + 0.565 + ], + "angle": 0, + "content": "Jointly performing part-level reconstruction and articulation estimation on an articulated object is prone to local minima [6, 30], making it preferable to strategically decouple the problem into multiple stages." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.572, + 0.794, + 0.587 + ], + "angle": 0, + "content": "3.2.1. Separate per-state reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.59, + 0.909, + 0.727 + ], + "angle": 0, + "content": "In Stage 1, SPLART separately optimizes two 3DGS models following the standard procedure, one for each end state of the articulation. In addition to the attributes that are standard to 3DGS, each Gaussian is initialized with a persistent binary state label \\( l \\), equally drawn from \\( \\{0,1\\} \\). We denote the set of Gaussians representing state \\( l \\) as \\( \\mathcal{G}_l^{\\mathrm{ref}} \\), where \\( ref \\) emphasizes that \\( \\mathcal{G}_l^{\\mathrm{ref}} \\) is the reference reconstruction unaffected by the other state. Given a data sample observed at state \\( l \\), the optimization follows as:" + }, + { + "type": "equation", + "bbox": [ + 0.645, + 0.737, + 0.905, + 0.765 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathcal {G} _ {l} ^ {\\text {r e f}}} \\Delta_ {I} \\left(\\hat {I} _ {l} ^ {\\text {r e f}, i}, I _ {l} ^ {i}\\right), \\tag {2a}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.768, + 0.905, + 0.787 + ], + "angle": 0, + "content": "\\[\n\\hat {I} _ {l} ^ {\\text {r e f}, i} = R \\left(\\mathcal {G} _ {l} ^ {\\text {r e f}}, P _ {l} ^ {i}, K _ {l} ^ {i}\\right), l \\in \\{0, 1 \\}, \\tag {2b}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.798, + 0.906, + 0.845 + ], + "angle": 0, + "content": "where \\( R \\) is the 3DGS rendering function, and \\( \\Delta_I \\) represents the photometric loss. For simplicity, the view index \\( i \\) and camera parameters \\( P_l^i, K_l^i \\) will be omitted from now on." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.852, + 0.892, + 0.867 + ], + "angle": 0, + "content": "3.2.2. Cross-static formulation for mobility estimation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "The focus of Stage 2 is mobility estimation for each Gaussian. To keep \\(\\mathcal{G}^{\\mathrm{ref}}\\) dedicated to single-state reconstruction," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8843" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.09, + 0.482, + 0.166 + ], + "angle": 0, + "content": "we first duplicate \\(\\mathcal{G}_l^{\\mathrm{ref}}\\) as \\(\\mathcal{G}_l^{\\mathrm{tgt}}\\) for both states (i.e., \\(l\\in \\{0,1\\}\\)), which is intended as the target representation that will fulfill the goals of part-level reconstruction and articulation estimation. By design, \\(\\mathcal{G}^{\\mathrm{tgt}}\\) shares neither data storage nor gradient flow with \\(\\mathcal{G}^{\\mathrm{ref}}\\) after its creation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.483, + 0.318 + ], + "angle": 0, + "content": "For each Gaussian in \\(\\mathcal{G}^{\\mathrm{tgt}}\\), we further extend its set of attributes with a mobility value \\(m \\in [0,1]\\), initialized at 0.5. By design, \\(m\\) enables the break-down of a Gaussian to its static and mobile components, where each component inherits all the original Gaussian attributes except for the opacity \\(\\sigma\\). The static component has its opacity modified to \\(\\sigma \\cdot (1 - m)\\), and the mobile component to \\(\\sigma \\cdot m\\). For simplicity, letting \\(\\mathcal{M}\\) be the set of mobilities for \\(\\mathcal{G}\\), we use the element-wise product \\(\\mathcal{G} \\cdot (1 - \\mathcal{M})\\) to denote the static component of \\(\\mathcal{G}\\), and \\(\\mathcal{G} \\cdot \\mathcal{M}\\) for the mobile component." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.482, + 0.395 + ], + "angle": 0, + "content": "To estimate the mobilities \\(\\mathcal{M}\\), we employ the intuition that the static components from both states should constitute the static part of the articulated object. Formally, we introduce the cross-static (cs) formulation, where the static part of the articulated object is jointly represented as" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.402, + 0.482, + 0.421 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} ^ {\\mathrm {s}} = \\mathcal {G} _ {l} ^ {\\mathrm {t g t}} \\cdot (1 - \\mathcal {M} _ {l}) \\oplus \\mathcal {G} _ {1 - l} ^ {\\mathrm {t g t}} \\cdot (1 - \\mathcal {M} _ {1 - l}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.429, + 0.483, + 0.457 + ], + "angle": 0, + "content": "where \\(\\oplus\\) denotes concatenation. For state \\(l\\), the target representation thus becomes" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.467, + 0.482, + 0.486 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {l} ^ {\\mathrm {t g t}, \\mathrm {c s}} = \\mathcal {G} ^ {\\mathrm {s}} \\oplus \\mathcal {G} _ {l} ^ {\\mathrm {t g t}} \\cdot \\mathcal {M} _ {l}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.494, + 0.482, + 0.525 + ], + "angle": 0, + "content": "With this formulation, Stage 2 is further divided into two sub-stages as described below." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.527, + 0.484, + 0.618 + ], + "angle": 0, + "content": "Stage 2(a): Coarse mobility estimation via cross-static geometric consistency. To measure the geometric distance between two Gaussian sets, we design a weighted version of the Chamfer distance. Specifically, letting \\( X = \\{(x_{i},w_{x_{i}})\\}_{i = 1}^{M} \\) and \\( Y = \\{(y_{j},w_{y_{j}})\\}_{j = 1}^{N} \\) be two sets of point-weight pairs, the weighted Chamfer distance is then:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.63, + 0.483, + 0.71 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname {Chamfer}(X,Y) = \\sum_{\\left(x_{i},w_{x_{i}}\\right)\\in X}\\tilde{w}_{x_{i}}\\min_{\\left(y_{j},w_{y_{j}}\\right)\\in Y}\\| x_{i} - y_{j}\\|^{2} \\\\ + \\sum_ {\\left(y _ {j}, w _ {y _ {j}}\\right) \\in Y} \\tilde {w} _ {y _ {j}} \\min _ {\\left(x _ {i}, w _ {x _ {i}}\\right) \\in X} \\| x _ {i} - y _ {j} \\| ^ {2}, \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.809 + ], + "angle": 0, + "content": "where \\(\\tilde{w}_{x_i} = w_{x_i} / \\sum_{(x_i, w_{x_i}) \\in X} w_{x_i}\\) and \\(\\tilde{w}_{y_j}\\) are normalized weights. For each Gaussian, we use its mobility-modified opacity \\(\\sigma'\\) as the weight, essentially treating it as \\(\\sigma'\\) points overlapped at its mean position. The mobilities are then optimized by encouraging the geometric consistency formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.819, + 0.482, + 0.841 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathcal {M}} \\mathrm {C D} _ {0} ^ {\\mathrm {c s}} + \\mathrm {C D} _ {1} ^ {\\mathrm {c s}} + \\lambda_ {m} ^ {\\text {g e o m}} \\| \\mathcal {M} \\|, \\tag {6a}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.843, + 0.482, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\mathrm {C D} _ {l} ^ {\\mathrm {c s}} = \\operatorname {C h a m f e r} \\left(\\mathcal {G} _ {l} ^ {\\mathrm {t g t , c s}}, \\mathcal {G} _ {l} ^ {\\mathrm {r e f}}\\right), \\tag {6b}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathrm{Chamfer}(\\cdot)\\) denotes the weighted Chamfer distance, and \\(\\lambda_{m}^{\\mathrm{geom}}\\| \\mathcal{M}\\|\\) is a regularization term that encourages" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "smaller mobilities. Note how \\(\\mathcal{M}\\) affects the weighted Chamfer distance by modifying the opacities, and that \\(\\mathcal{M} = 1\\) is a trivial solution without the regularization. Without photometric supervision, the mobilities obtained from Eqn. 6a are generally noisy. However, being relatively fast (taking only tens of seconds), they still serve as a good initialization for the next sub-stage." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.905, + 0.256 + ], + "angle": 0, + "content": "Stage 2(b): Joint mobility and Gaussian optimization via cross-static rendering. To more accurately estimate the mobilities while jointly refining the full Gaussian attributes, we perform cross-static rendering as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.262, + 0.905, + 0.285 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathcal {G} ^ {\\mathrm {t g t}}, \\mathcal {M} _ {1 - l}} \\Delta_ {I} \\left(\\hat {I} _ {l} ^ {\\mathrm {c s}}, I _ {l}\\right) + \\lambda_ {m} ^ {\\text {p h o t o}} \\| \\mathcal {M} _ {1 - l} \\|, \\tag {7a}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.66, + 0.289, + 0.905, + 0.307 + ], + "angle": 0, + "content": "\\[\n\\hat {I} _ {l} ^ {\\mathrm {c s}} = R \\left(\\mathcal {G} _ {l} ^ {\\mathrm {t g t}, \\mathrm {c s}}\\right), \\tag {7b}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.315, + 0.904, + 0.347 + ], + "angle": 0, + "content": "where \\(\\lambda_{m}^{\\mathrm{photo}}\\| \\mathcal{M}_{1 - l}\\|\\) is a mobility regularization term similar to that in Eqn. 6a." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.354, + 0.905, + 0.384 + ], + "angle": 0, + "content": "3.2.3. Cross-mobile formulation for articulation estimation and mobility refinement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.905, + 0.494 + ], + "angle": 0, + "content": "The focus of Stage 3 is to estimate the articulation parameters shared by the mobile components of all Gaussians. To this end, we employ the intuition that the mobile components from the two end states are related through the articulated motion. Formally, we introduce the cross-mobile (cm) formulation, where the mobile part of the articulated object at state \\( l \\) is jointly represented as" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.502, + 0.905, + 0.521 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {l} ^ {\\mathrm {m}} = \\mathcal {G} _ {l} ^ {\\mathrm {t g t}} \\cdot \\mathcal {M} _ {l} \\oplus {} ^ {l} T _ {1 - l} \\left(\\mathcal {G} _ {1 - l} ^ {\\mathrm {t g t}} \\cdot \\mathcal {M} _ {1 - l}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.531, + 0.905, + 0.577 + ], + "angle": 0, + "content": "where \\( {}^l T_{1 - l}(\\mathcal{G}) \\) denotes the transformation of Gaussians \\( \\mathcal{G} \\) according to the articulated motion from state \\( 1 - l \\) to \\( l \\). For state \\( l \\), the target representation then becomes" + }, + { + "type": "equation", + "bbox": [ + 0.644, + 0.586, + 0.905, + 0.605 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {l} ^ {\\mathrm {t g t}, \\mathrm {c m}} = \\mathcal {G} ^ {\\mathrm {s}} \\oplus \\mathcal {G} _ {l} ^ {\\mathrm {m}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.904, + 0.643 + ], + "angle": 0, + "content": "where \\(\\mathcal{G}^{\\mathrm{s}}\\) is defined in Eqn. 3. With this formulation, Stage 3 is further divided into three sub-stages as below." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.646, + 0.905, + 0.721 + ], + "angle": 0, + "content": "Stage 3(a): Coarse articulation estimation via geometric consistency. Similar to Eqn. 6a, weighted Chamfer distance is used for computing the cross-mobile geometric consistency, which can be optimized over both mobilities and articulation parameters as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.731, + 0.905, + 0.754 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathcal {M}, T} \\mathrm {C D} _ {0} ^ {\\mathrm {c m}} + \\mathrm {C D} _ {1} ^ {\\mathrm {c m}}, \\tag {10a}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.757, + 0.905, + 0.776 + ], + "angle": 0, + "content": "\\[\n\\mathrm {C D} _ {l} ^ {\\mathrm {c m}} = \\operatorname {C h a m f e r} \\left(\\mathcal {G} _ {l} ^ {\\mathrm {t g t , c m}}, \\mathcal {G} _ {l} ^ {\\mathrm {r e f}}\\right). \\tag {10b}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.785, + 0.905, + 0.876 + ], + "angle": 0, + "content": "However, we notice that this formulation is still susceptible to local minima if the mobile part is too geometrically insignificant. We qualitatively show one such failure in Fig. 3a. To remedy this, we further propose mobile-only geometric consistency, which considers only the mobile components of both states as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.884, + 0.905, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathrm {C D} ^ {\\mathrm {m}} = \\operatorname {C h a m f e r} \\left(\\mathcal {G} _ {l} ^ {\\mathrm {t g t}} \\cdot \\mathcal {M} _ {l}, ^ {l} T _ {1 - l} \\left(\\mathcal {G} _ {1 - l} ^ {\\mathrm {t g t}} \\cdot \\mathcal {M} _ {1 - l}\\right)\\right). \\tag {11}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8844" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.468, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.165, + 0.282, + 0.189 + ], + "angle": 0, + "content": "(a) Failure with cross-mobile geometric consistency." + }, + { + "type": "image_caption", + "bbox": [ + 0.289, + 0.165, + 0.478, + 0.189 + ], + "angle": 0, + "content": "(b) Success with mobile-only geometric consistency." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.195, + 0.483, + 0.281 + ], + "angle": 0, + "content": "Figure 3. 103549-Toaster has a flat slider as its mobile part, similar in geometric curvature to the static component. This similarity makes the nearest-neighbor data association error-prone when computing the Chamfer distance for cross-mobile geometric consistency. On the other hand, mobile-only geometric consistency focuses only on the mobile part and successfully handles this case." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.304, + 0.484, + 0.424 + ], + "angle": 0, + "content": "While successfully addressing these scenarios, mobile-only geometric consistency still falls short in other circumstances, especially when the mobile components in the two states exhibit large discrepancies. We qualitatively show one such failure in Fig. 4a. To take advantage of both formulations and to facilitate robustness in the inherently nonconvex optimization of articulation parameters, we propose the following practical strategy:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.425, + 0.472, + 0.438 + ], + "angle": 0, + "content": "1. With \\( K^{\\mathrm{m}} \\) randomized tries: articulation estimation via" + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.451, + 0.482, + 0.473 + ], + "angle": 0, + "content": "\\[\nT ^ {m} = \\arg \\min _ {T} \\mathrm {C D} ^ {\\mathrm {m}}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.482, + 0.483, + 0.511 + ], + "angle": 0, + "content": "2. With \\( K^{\\mathrm{cm}} \\) randomized tries, plus another initialized with \\( T^{\\mathrm{m}} \\): articulation estimation via" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.523, + 0.482, + 0.545 + ], + "angle": 0, + "content": "\\[\nT ^ {\\mathrm {c m}} = \\arg \\min _ {T} \\mathrm {C D} _ {0} ^ {\\mathrm {c m}} + \\mathrm {C D} _ {1} ^ {\\mathrm {c m}}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.555, + 0.483, + 0.586 + ], + "angle": 0, + "content": "3. Final run initialized with \\( T^{\\mathrm{cm}} \\): joint articulation estimation and mobility refinement via Eqn. 10a." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.588, + 0.484, + 0.664 + ], + "angle": 0, + "content": "Stage 3(b): Joint articulation, mobility, and Gaussian optimization via cross-mobile rendering. As in Stage 2(b), we perform a full joint optimization over articulation parameters, mobilities, and Gaussians utilizing photometric supervision via cross-mobile rendering as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.674, + 0.482, + 0.699 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathcal {G} ^ {\\mathrm {i g t}}, \\mathcal {M} _ {1 - l}, ^ {l} T _ {1 - l}} \\Delta_ {I} \\left(\\hat {I} _ {l} ^ {\\mathrm {c m}}, I _ {l}\\right), \\tag {14a}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.703, + 0.482, + 0.722 + ], + "angle": 0, + "content": "\\[\n\\hat {I} _ {l} ^ {\\mathrm {c m}} = R \\left(\\mathcal {G} _ {l} ^ {\\mathrm {t g t , c m}}\\right), l \\in \\{0, 1 \\}. \\tag {14b}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.732, + 0.484, + 0.853 + ], + "angle": 0, + "content": "Stage 3(c): Mobility correction via cross-mobile geometric consistency. Stage 3(b) solely relies on photometric supervision, which is limited to the training views. On rare occasions, a Gaussian may be mistakenly estimated as mobile if the articulated motion moves it out of sight from most views, effectively losing supervision. On the other hand, geometric consistency is not affected by visibility limitations, which we leverage for mobility correction as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.864, + 0.483, + 0.886 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathcal {M}} \\mathrm {C D} _ {0} ^ {\\mathrm {c m}} + \\mathrm {C D} _ {1} ^ {\\mathrm {c m}}. \\tag {15}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.901, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.188, + 0.704, + 0.213 + ], + "angle": 0, + "content": "(a) Failure with mobile-only geometric consistency." + }, + { + "type": "image_caption", + "bbox": [ + 0.712, + 0.189, + 0.901, + 0.213 + ], + "angle": 0, + "content": "(b) Success with cross-mobile geometric consistency." + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.22, + 0.905, + 0.317 + ], + "angle": 0, + "content": "Figure 4. 101713-Pen has both ends of a pen as the mobile part, with only one end visible from any given observation. Thus, the reconstructed mobile components from the two states essentially represent distinct parts, violating the intuition for mobile-only geometric consistency. On the other hand, cross-mobile geometric consistency considers the joint representation from both states as a whole, successfully estimating the articulation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.327, + 0.734, + 0.344 + ], + "angle": 0, + "content": "3.3. Real-World Application" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.35, + 0.907, + 0.577 + ], + "angle": 0, + "content": "For an articulated object to be reconstructed in the real world, we first collect two sets of RGB images, one for each articulation state, by imaging the object from a surrounding hemisphere. We then preprocess the data to determine the posed images that serve as the input to SPLART. This involves using SAM 2 [46] to perform foreground-background segmentation, which also removes dynamic contents of the scene. We then perform structure from motion (SfM) to determine camera poses using COLMAP [48] with SuperPoint [7] descriptors and SuperGlue [47] matching on the segmented backgrounds from both sets of images to construct a joint coordinate frame for the object in both states. Once we obtain the joint coordinate frame and the foreground target object is localized, we run SPLART to reconstruct the articulated objects." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.591, + 0.645, + 0.608 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.617, + 0.615, + 0.631 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.639, + 0.907, + 0.85 + ], + "angle": 0, + "content": "PARIS PartNet-Mobility Subset. Our evaluation includes PARIS [30] as a baseline, which selects 10 instances (8 revolute, 2 prismatic) from PartNet-Mobility [2, 39, 61], a large-scale dataset of simulatable 3D objects with part-level mobility. We refer to this subset as PARIS-PMS. For each articulation state, the dataset provides 100 calibrated object-centric views for training and 50 for testing, sampled from the upper hemisphere. However, the released dataset lacks ground-truth depth and part segmentation maps. To enable their evaluations, we augment PARIS-PMS with the necessary ground-truth data by following its data generation procedure. The dataset does not provide test views of intermediate articulation states, which limits the quantitative evaluation of novel articulation synthesis." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "SPLART PartNet-Mobility Subset. We curate an additional articulated object dataset from PartNet-Mobility, dubbed SPLART-PMS, to enable a more comprehensive" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8845" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.905, + 0.234 + ], + "angle": 0, + "content": "
TypeMethodSuccess Rate %↑erra(x10-2DEG)↓errp(x10-3)↓errr(x10-2DEG)↓errt(x10-3)↓CDs(x10-3)↓CDm(x10-3)↓CDw(x10-3)↓
RevolutePARIS40.0121.90 ± 68.455.27 ± 4.55127.80 ± 55.12N/A3.48 ± 0.3968.44 ± 13.0313.11 ± 1.16
DTA†98.812.32 ± 4.161.86 ± 1.6919.41 ± 16.28N/A2.20 ± 0.491.27 ± 1.871.73 ± 0.03
SPLART100.03.70 ± 0.380.40 ± 0.074.85 ± 0.33N/A4.08 ± 0.321.06 ± 0.783.59 ± 0.21
PrismaticPARIS50.027.97 ± 13.09N/AN/A4.28 ± 3.029.21 ± 1.94151.78 ± 35.007.99 ± 0.49
DTA†100.016.26 ± 3.54N/AN/A1.15 ± 0.142.69 ± 0.0415.74 ± 0.282.22 ± 0.03
SPLART95.02.36 ± 0.44N/AN/A0.33 ± 0.046.02 ± 0.2016.77 ± 1.143.69 ± 0.14
OverallPARIS42.0103.12 ± 57.385.27 ± 4.55127.80 ± 55.124.28 ± 3.024.63 ± 0.7085.11 ± 17.4212.09 ± 1.03
DTA†99.013.11 ± 4.041.86 ± 1.6919.41 ± 16.281.15 ± 0.142.30 ± 0.404.16 ± 1.551.83 ± 0.03
SPLART99.03.44 ± 0.390.40 ± 0.074.85 ± 0.330.33 ± 0.044.47 ± 0.304.20 ± 0.853.61 ± 0.19
" + }, + { + "type": "table_caption", + "bbox": [ + 0.128, + 0.238, + 0.868, + 0.253 + ], + "angle": 0, + "content": "Table 1. PARIS-PMS Articulation and Mesh Reconstruction Metrics. Best results in bold. \\( {}^{ \\dagger } \\) DTA requires ground-truth depth." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.265, + 0.481, + 0.434 + ], + "angle": 0, + "content": "
TypeMethodPSNR ↑Depth MAE ↓mIoU ↑
RevolutePARIS32.210.0930.955
DTA†N/A0.0310.941
SPLART43.530.0390.974
PrismaticPARIS33.750.1080.902
DTA†N/A0.0660.844
SPLART44.380.0250.890
OverallPARIS32.520.0960.942
DTA†N/A0.0380.922
SPLART43.700.0360.957
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.438, + 0.483, + 0.482 + ], + "angle": 0, + "content": "Table 2. PARIS-PMS Novel View Synthesis Metrics, averaged over successful runs (see Tab. 1 for success rates). Best results in bold. \\( {}^{ \\dagger } \\) DTA requires ground-truth depth." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.49, + 0.483, + 0.686 + ], + "angle": 0, + "content": "evaluation including the quantitative assessment of novel articulation synthesis. Our dataset consists of 22 object instances from categories not included in PARIS-PMS, with 12 revolute and 10 prismatic objects. For the test split, we also generate ground-truth depth and part segmentation maps to facilitate relevant evaluations. Crucially, while PARIS-PMS restricts articulation states to binary values, SPLART-PMS uniformly samples states from \\([-0.1, 1.1]\\), demanding accurate estimation of articulation, part segmentation, and object reconstruction for correct view synthesis. Our experiments reveal that SPLART-PMS poses a greater challenge than PARIS-PMS—the latter being almost perfectly solved by our method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.794 + ], + "angle": 0, + "content": "Real-World Dataset. To show the efficacy of SPLART on real-world data, we collect a dataset that consists of 7 real-world objects with 9 articulations in total, including common objects like a monitor, cabinet, glasses, etc. All images are captured by a hand-held phone camera. Camera parameters are calibrated following strategies detailed in Sec. 3.3. We qualitatively evaluate SPLART on this dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.804, + 0.275, + 0.818 + ], + "angle": 0, + "content": "4.2. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Articulation Estimation. We evaluate the accuracy of articulation estimation depending on its type." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "- A revolute articulation describes a rotation around some line in 3D space, parameterized by the pivot point \\(\\mathbf{p}\\) on the line, the line's directional axis \\(\\mathbf{a}\\), and the angle of" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.265, + 0.907, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.537, + 0.49, + 0.882, + 0.504 + ], + "angle": 0, + "content": "Figure 5. Real-World Articulated Object Reconstructions." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.516, + 0.905, + 0.735 + ], + "angle": 0, + "content": "
TypeMethodSuccess Rate %errn(×10-2DEG) ↓errp(×10-3) ↓errr(×10-2DEG) ↓errt(×10-3) ↓
RevolutePARIS18.3167.34 ± 42.479.36 ± 3.73270.37 ± 91.19N/A
DTA†33.332.76 ± 2.381.28 ± 0.3940.80 ± 4.66N/A
SPLART-2a78.52.91 ± 0.552.53 ± 2.1852.06 ± 32.00N/A
SPLART-2b67.04.62 ± 1.693.92 ± 3.07131.99 ± 64.45N/A
SPLART-3a0.8296.61 ± 0.002.83 ± 0.00427.10 ± 0.00N/A
SPLART-3b76.784.13 ± 19.135.11 ± 3.03206.98 ± 56.30N/A
SPLART77.520.76 ± 0.813.93 ± 2.4178.31 ± 31.06N/A
PrismaticPARIS39.049.49 ± 18.92N/AN/A14.96 ± 5.51
DTA†90.0102.77 ± 12.48N/AN/A9.25 ± 0.73
SPLART-2a90.025.88 ± 23.25N/AN/A1.50 ± 1.63
SPLART-2b85.715.63 ± 20.40N/AN/A0.73 ± 0.58
SPLART-3a13.015.17 ± 0.59N/AN/A2.26 ± 0.04
SPLART-3b86.061.00 ± 19.76N/AN/A9.43 ± 2.37
SPLART95.025.72 ± 17.63N/AN/A1.04 ± 0.32
OverallPARIS27.7103.88 ± 29.799.36 ± 3.73270.37 ± 91.1914.96 ± 5.51
DTA†59.177.76 ± 8.871.28 ± 0.3940.80 ± 4.669.25 ± 0.73
SPLART-2a83.713.85 ± 11.362.53 ± 2.1852.06 ± 32.001.50 ± 1.63
SPLART-2b74.79.34 ± 9.713.92 ± 3.07131.99 ± 64.450.73 ± 0.58
SPLART-3a6.485.53 ± 0.442.83 ± 0.00427.10 ± 0.002.26 ± 0.04
SPLART-3b80.973.62 ± 19.425.11 ± 3.03206.98 ± 56.309.43 ± 2.37
SPLART85.523.02 ± 8.463.93 ± 2.4178.31 ± 31.061.04 ± 0.32
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.739, + 0.905, + 0.768 + ], + "angle": 0, + "content": "Table 3. SPLART-PMS Articulation Metrics. Best results in bold. \\( {}^{ \\dagger } \\) DTA requires ground-truth depth." + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.78, + 0.907, + 0.903 + ], + "angle": 0, + "content": "rotation \\(\\theta\\). We report the angular error \\(\\mathrm{err}_{\\mathbf{a}}\\) (in \\(\\times 10^{-2}\\) DEG) between the predicted and ground-truth axes, the geodesic distance \\(\\mathrm{err}_{\\mathrm{r}}\\) (in \\(\\times 10^{-2}\\) DEG) between the predicted and ground-truth rotations induced by the axis-angle (\\(\\mathbf{a} - \\theta\\)) pair, and the pivot point error \\(\\mathrm{err}_{\\mathbf{p}}\\). Since the pivot can move arbitrarily along the axis, \\(\\mathrm{err}_{\\mathbf{p}}\\) is computed as the closest distance between the predicted and ground-truth lines induced by the axis-pivot (\\(\\mathbf{a} - \\mathbf{p}\\)) pair." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8846" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.48, + 0.255 + ], + "angle": 0, + "content": "
TypeMethodPSNR ↑Depth MAE ↓mIoU ↑
RevolutePARIS32.750.1250.897
DTA†N/A0.0530.886
SPLART36.380.0320.897
PrismaticPARIS32.300.1240.902
DTA†N/A0.0600.857
SPLART37.770.0330.938
OverallPARIS32.510.1250.900
DTA†N/A0.0580.867
SPLART37.010.0320.916
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.261, + 0.483, + 0.306 + ], + "angle": 0, + "content": "Table 4. SPLART-PMS Novel View and Articulation Synthesis Metrics, averaged over successful runs (see Tab. 3 for success rates). Best results in bold. \\(\\dagger\\)DTA requires ground-truth depth." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.314, + 0.484, + 0.421 + ], + "angle": 0, + "content": "- A prismatic articulation describes a translation along a specific direction, parameterized by the axis \\( \\mathbf{a} \\) of the translation direction and the translation distance \\( d \\). We report the axis error \\( \\mathrm{err}_{\\mathbf{a}} \\), as in the revolute case, and the translation error \\( \\mathrm{err}_{\\mathrm{t}} \\), which is the distance between the predicted and ground-truth translations induced by the axis-distance \\( (\\mathbf{a} - d) \\) pair." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.427, + 0.484, + 0.835 + ], + "angle": 0, + "content": "Part-Level Reconstruction. The evaluation of part-level reconstruction accuracy is threefold: photometric rendering quality, geometric accuracy, and part segmentation accuracy. For all these aspects, we use novel view synthesis as the surrogate task. We perform volume rendering for each view in the test split, generating outputs that include the RGB image, depth map, and part segmentation map. Photometric rendering quality is evaluated by reporting image metrics PSNR. Geometric accuracy is assessed by reporting the mean absolute error of the depth map (i.e., Depth MAE). Part segmentation accuracy is evaluated using the intersection-over-union (IoU) ratio for three categories: static part \\((\\mathrm{IoU_s})\\), mobile part \\((\\mathrm{IoU_m})\\), and background \\((\\mathrm{IoU_{bg}})\\). The mean IoU (mIoU) is then reported as the average IoU across these categories. In addition, we also evaluate geometric accuracy through mesh reconstruction. To extract a mesh from 3DGS, we: (1) render depth images from uniformly sampled spherical viewpoints, (2) fuse rendered depths into a TSDF representation [4, 41], and (3) extract the mesh using the Marching Cubes algorithm [34]. For each reconstruction, we separately extract the meshes corresponding to the static, mobile, and whole parts. For evaluation, we follow the procedure used in previous works [30, 60]: 10 000 points are uniformly sampled from both the reconstructed and the ground-truth meshes, and the Chamfer distance is computed for each category—static \\((\\mathrm{CD_s})\\), mobile \\((\\mathrm{CD_m})\\), and whole \\((\\mathrm{CD_w})\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Impact of Randomness. Randomness can play a significant role in the joint optimization of articulation and part reconstruction [6, 30]. In order to effectively demonstrate the robustness of a method, we conduct ten trials for each" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.906, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.53, + 0.222, + 0.889, + 0.236 + ], + "angle": 0, + "content": "Figure 6. State interpolation from the estimated articulation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.246, + 0.907, + 0.443 + ], + "angle": 0, + "content": "scene, each with a different fixed random seed. We report results averaged over the ten runs. Additionally, we notice that for a failed run converging to a local minimum, the evaluation metrics, such as \\(\\mathrm{err}_{\\mathbf{r}}\\), are typically orders-of-magnitude worse than those of a successful run. To better reflect a method's performance, we report the number of successful runs out of the total number (i.e., ten), where a run is deemed successful if and only if each of the following criteria are met: (1) \\(\\mathrm{err}_{\\mathbf{a}} < 5\\) DEG; (2) \\(\\mathrm{err}_{\\mathbf{p}} < 0.05\\) for revolute articulation; (3) \\(\\mathrm{err}_{\\mathbf{r}} < 10\\) DEG for revolute articulation; and (4) \\(\\mathrm{err}_{\\mathbf{t}} < 0.05\\) for prismatic articulation. We then proceed to report the evaluation metrics detailed previously, averaging only over the successful runs." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.452, + 0.623, + 0.467 + ], + "angle": 0, + "content": "4.3. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.474, + 0.906, + 0.595 + ], + "angle": 0, + "content": "PARIS [30] addresses the same challenge we tackle: reconstructing the part-level shape, appearance, and motion of an articulated object from multi-view RGB images captured at two articulation states. It employs a NeRF-based representation, modeling the object with static and mobile fields for each part, alongside a transformation to capture state transitions. These components are jointly optimized end-to-end, primarily through an image-rendering loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.906, + 0.75 + ], + "angle": 0, + "content": "DTA [60] adopts a similar setup: reconstructing digital twins of articulated objects from multi-view RGB-D observations at two articulation states. It operates in two stages: Stage 1 optimizes neural object fields for each state using RGB-D images and extracts meshes; Stage 2 optimizes a part segmentation field and per-part motions using consistency, matching, and collision losses to determine point correspondences between states. Note: Unlike SPLART, DTA requires ground-truth depth as input and produces plain object meshes unsuitable for photorealistic rendering." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.704, + 0.774 + ], + "angle": 0, + "content": "4.4. Experiment Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Articulation Estimation. We quantitatively evaluate SPLART against baselines for articulation estimation accuracy, reporting scene-averaged results per articulation type on the PARIS-PMS dataset (Tab. 1) and the SPLART-PMS dataset (Tab. 3). SPLART consistently outperforms the baselines across both datasets. Additionally, we qualitatively compare SPLART with PARIS on selected scenes from these datasets in Fig. 7. In the \"color\" column, each" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8847" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.907, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.233, + 0.292, + 0.761, + 0.307 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison on part-level reconstruction and articulation estimation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.316, + 0.484, + 0.392 + ], + "angle": 0, + "content": "image visualizes the ground-truth articulation (green arrow) and the estimated articulation (red arrow), except in the ground-truth column. When the estimate aligns perfectly with the ground truth, only the red arrow appears; if significantly misaligned, only the green arrow is visible." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.405, + 0.484, + 0.618 + ], + "angle": 0, + "content": "Part-Level Reconstruction. We assess SPLART's part-level reconstruction accuracy against baselines, using novel view synthesis as a surrogate task. Scene-averaged results per articulation type are reported on the PARIS-PMS dataset (Tab. 2) and the SPLART-PMS dataset (Tab. 4). We also evaluate geometric accuracy through mesh reconstruction, presenting scene-averaged results per articulation type on the PARIS-PMS dataset (Tab. 1). Without depth supervision, which is required for the DTA baseline, SPLART matches DTA's performance, both substantially exceeding PARIS. Qualitative comparisons with PARIS, including RGB renderings and part segmentation maps, are provided for selected scenes from both datasets in Fig. 7. Additional qualitative results on articulation synthesis appear in Fig. 6." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.484, + 0.691 + ], + "angle": 0, + "content": "Real-World Reconstructions. We qualitatively evaluate SPLART on the aforementioned real-world dataset, with results shown in Fig. 5. Each image overlays the estimated articulation, visualized with a red arrow." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Ablation Studies. We perform extensive ablation studies to demonstrate the contribution of each stage. Specifically, we perform the following four ablations: (1) No geometric supervision for mobility estimation, i.e., skipping Stage 2(a), denoted SPLART-2a; (2) No photometric supervision for mobility estimation, i.e., skipping Stage 2(b), denoted SPLART-2b; (3) No geometric supervision for articulation estimation, i.e., skipping Stage 3(a), denoted SPLART-3a; (4) No photometric supervision for articulation estimation, i.e., skipping Stage 3(b), denoted SPLART-3b. Like the main experiments, we conduct 10 runs for each scene, but only on the more challenging dataset of SPLART-PMS. We report the articulation estimation results in Tab. 3." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.313, + 0.907, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.408, + 0.904, + 0.422 + ], + "angle": 0, + "content": "Figure 8. Failure cases due to inherent ambiguities in articulation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.427, + 0.634, + 0.442 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We present SPLART, the first self-supervised method that utilizes 3D Gaussian Splatting for articulated object reconstruction from two-state RGB observations, without 3D supervision or category-specific priors. SPLART delivers robust optimization and effectively handles challenging scenarios—outperforming prior approaches while eliminating the need for 3D supervision, articulation annotations, or semantic labels. Extensive evaluations on synthetic and real-world datasets show SPLART's articulation estimation accuracy and view synthesis quality that surpass existing methods. Real-world experiments further validate its practicality, showcasing successful reconstructions of diverse articulated objects using only a handheld RGB camera. Our framework empowers non-expert users to effortlessly create high-fidelity, articulated digital twins, supporting applications in augmented/virtual reality and robotics. However, while SPLART effectively addresses two-part articulations, it lacks direct support for objects with more articulated parts, though iteratively applying it to incremental state pairs may help. Some existing methods [6, 33, 60] handle such objects, but require known part counts and specific articulation structures. Future research will focus on a comprehensive adaptation to multi-level articulated objects. Additionally, inherent ambiguities may arise when multiple articulations explain state differences, as shown in Fig. 8 with examples like a clock's hand rotating on an unexpected axis or a globe's surface spinning independently. Overall, SPLART represents a scalable, robust, and practical solution that pushes the boundaries of articulated object reconstruction, laying a strong foundation for future work." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8848" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Ben Abbatematteo, Stefanie TELlex, and George Konidaris. Learning to generalize kinematic models to novel objects. In Proceedings of the Conference on Robot Learning (CoRL), pages 1289-1299, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.243 + ], + "angle": 0, + "content": "[2] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. ShapeNet: An information-rich 3D model repository. arXiv preprint arXiv:1512.03012, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.484, + 0.315 + ], + "angle": 0, + "content": "[3] Christopher B. Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3D-R2N2: A unified approach for single and multi-view 3D object reconstruction. In Proceedings of the European Conference on Computer Vision (ECCV), pages 628-644, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.317, + 0.484, + 0.384 + ], + "angle": 0, + "content": "[4] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), pages 303-312, 1996. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.387, + 0.484, + 0.444 + ], + "angle": 0, + "content": "[5] Andrea F. Daniele, Thomas M. Howard, and Matthew R. Walter. A multiview approach to learning articulated motion models. In Proceedings of the International Symposium of Robotics Research (ISRR), pages 371-386, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.446, + 0.484, + 0.515 + ], + "angle": 0, + "content": "[6] Jianning Deng, Kartic Subr, and Hakan Bilen. Articulate your NeRF: Unsupervised articulated object modeling via conditional view synthesis. In Advances in Neural Information Processing Systems (NeurIPS), pages 119717-119741, 2025. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.484, + 0.587 + ], + "angle": 0, + "content": "[7] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 224-236, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.589, + 0.484, + 0.658 + ], + "angle": 0, + "content": "[8] Haoqiang Fan, Hao Su, and Leonidas J. Guibas. A point set generation network for 3D object reconstruction from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2463-2471, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.66, + 0.484, + 0.73 + ], + "angle": 0, + "content": "[9] Samir Yitzhak Gadre, Kiana Ehsani, and Shuran Song. Act the part: Learning interaction strategies for articulated object part discovery. In Proceedings of the International Conference on Computer Vision (ICCV), pages 15732-15741, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.733, + 0.484, + 0.815 + ], + "angle": 0, + "content": "[10] Haoran Geng, Ziming Li, Yiran Geng, Jiayi Chen, Hao Dong, and He Wang. PartManip: Learning cross-category generalizable part manipulation policy from point cloud observations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2978-2988, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[11] Haoran Geng, Helin Xu, Chengyang Zhao, Chao Xu, Li Yi, Siyuan Huang, and He Wang. GAPartNet: Cross-category domain-generalizable object perception and manipulation via generalizable and actionable parts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7081-7091, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.162 + ], + "angle": 0, + "content": "[12] Antoine Guédon and Vincent Lepetit. SuGaR: Surface-aligned Gaussian splattering for efficient 3D mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5354-5363, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.163, + 0.907, + 0.258 + ], + "angle": 0, + "content": "[13] Nick Heppert, Muhammad Zubair Irshad, Sergey Zakharov, Katherine Liu, Rares Andrei Ambrus, Jeannette Bohg, Abhinav Valada, and Thomas Kollar. CARTO: Category and joint agnostic reconstruction of articulated objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21201-21210, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.26, + 0.906, + 0.328 + ], + "angle": 0, + "content": "[14] Cheng-Chun Hsu, Zhenyu Jiang, and Yuke Zhu. Ditto in the house: Building articulation models of indoor scenes through interactive perception. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 3933-3939, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.329, + 0.906, + 0.385 + ], + "angle": 0, + "content": "[15] Ruizhen Hu, Wenchao Li, Oliver Van Kaick, Ariel Shamir, Hao Zhang, and Hui Huang. Learning to predict part mobility from a single static snapshot. ACM Transactions On Graphics, 36(6):1-13, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.906, + 0.453 + ], + "angle": 0, + "content": "[16] Xiaoxia Huang, Ian Walker, and Stan Birchfield. Occlusion-aware reconstruction and manipulation of 3D articulated objects. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 1365-1371, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.455, + 0.906, + 0.538 + ], + "angle": 0, + "content": "[17] Muhammad Zubair Irshad, Thomas Kollar, Michael Laskey, Kevin Stone, and Zsolt Kira. CenterSnap: Single-shot multi-object 3D shape reconstruction and categorical 6D pose and size estimation. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 10632-10640, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.906, + 0.607 + ], + "angle": 0, + "content": "[18] Muhammad Zubair Irshad, Sergey Zakharov, Rares Ambrus, Thomas Kollar, Zsolt Kira, and Adrien Gaidon. ShAPO: Implicit representations for multi-object shape, appearance, and pose optimization. In Proceedings of the European Conference on Computer Vision (ECCV), pages 275–292, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.608, + 0.906, + 0.662 + ], + "angle": 0, + "content": "[19] Muhammad Zubair Irshad, Mauro Comi, Yen-Chen Lin, Nick Heppert, Abhinav Valada, Rares Ambrus, Zsolt Kira, and Jonathan Tremblay. Neural fields in robotics: A survey, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.664, + 0.906, + 0.746 + ], + "angle": 0, + "content": "[20] Ajinkya Jain, Rudolf Lioutikov, Caleb Chuck, and Scott Niekum. ScrewNet: Category-independent articulation model estimation from depth images using screw theory. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 13670-13677, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[21] Anil K. Jain. Fundamentals of Digital Image Processing. Prentice-Hall, Inc., 1989. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.906, + 0.831 + ], + "angle": 0, + "content": "[22] Hanxiao Jiang, Yongsen Mao, Manolis Savva, and Angel X. Chang. OPD: Single-view 3D openable part detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 410-426, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[23] Zhenyu Jiang, Cheng-Chun Hsu, and Yuke Zhu. Ditto: Building digital twins of articulated objects from interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5606-5616, 2022. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8849" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[24] Dov Katz, Moslem Kazemi, J. Andrew Bagnell, and Anthony Stentz. Interactive segmentation, tracking, and kinematic modeling of unknown 3D articulated objects. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 5003-5010, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[25] Yuki Kawana, Yusuke Mukuta, and Tatsuya Harada. Unsupervised pose-aware part decomposition for man-made articulated objects. In Proceedings of the European Conference on Computer Vision (ECCV), pages 558-575, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.272 + ], + "angle": 0, + "content": "[26] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4):1-14, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.343 + ], + "angle": 0, + "content": "[27] Justin Kerr, Chung Min Kim, Mingxuan Wu, Brent Yi, Qianqian Wang, Ken Goldberg, and Angjoo Kanazawa. Robot see robot do: Imitating articulated object manipulation with monocular 4d reconstruction. In Proceedings of the Conference on Robot Learning (CoRL), pages 587-603, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[28] Long Le, Jason Xie, William Liang, Hung-Ju Wang, Yue Yang, Yecheng Jason Ma, Kyle Vedder, Arjun Krishna, Dinesh Jayaraman, and Eric Eaton. Articulate-anything: Automatic modeling of articulated objects via a vision-language foundation model. In Proceedings of the International Conference on Learning Representations (ICLR), 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.427, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[29] Xiaolong Li, He Wang, Li Yi, Leonidas J. Guibas, A. Lynn Abbott, and Shuran Song. Category-level articulated object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3703-3712, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[30] Jiayi Liu, Ali Mahdavi-Amiri, and Manolis Savva. PARIS: Part-level reconstruction and motion analysis for articulated objects. In Proceedings of the International Conference on Computer Vision (ICCV), pages 352-363, 2023. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[31] Liu Liu, Wenqiang Xu, Haoyuan Fu, Sucheng Qian, Qiao-jun Yu, Yang Han, and Cewu Lu. A real-world articulated object knowledge base. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14789-14798, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[32] Shaowei Liu, Saurabh Gupta, and Shenlong Wang. Building rearticulable models for arbitrary 3D objects from 4D point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21138-21147, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[33] Yu Liu, Baoxiong Jia, Ruijie Lu, Junfeng Ni, Song-Chun Zhu, and Siyuan Huang. Building interactable replicas of complex articulated objects via gaussian splatting. In Proceedings of the International Conference on Learning Representations (ICLR), 2025. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[34] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3D surface construction algorithm. In *Seminal Graphics: Pioneering Efforts That Shaped the Field*, pages 347–353, 1998. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[35] Zhao Mandi, Yijia Weng, Dominik Bauer, and Shuran Song. Real2code: Reconstruct articulated objects via code generation. In Proceedings of the International Conference on Learning Representations (ICLR), 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[36] Eric Marchand, Hideaki Uchiyama, and Fabien Spindler. Pose estimation for augmented reality: A hands-on survey." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "IEEE Transactions on Visualization and Computer Graphics, 22(12):2633-2651, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[37] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3D reconstruction in function space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4460-4470, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.19, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[38] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.26, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[39] Kaichun Mo, Shilin Zhu, Angel X. Chang, Li Yi, Subarna Tripathi, Leonidas J. Guibas, and Hao Su. PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 909-918, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.425 + ], + "angle": 0, + "content": "[40] Jiteng Mu, Weichao Qiu, Adam Kortylewski, Alan Yuille, Nuno Vasconcelos, and Xiaolong Wang. A-SDF: Learning disentangled signed distance functions for articulated shape representation. In Proceedings of the International Conference on Computer Vision (ICCV), pages 12981-12991, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.427, + 0.905, + 0.522 + ], + "angle": 0, + "content": "[41] Richard A. Newcombe, Andrew Fitzgibbon, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J. Davison, Pushmeet Kohi, Jamie Shotton, and Steve Hodges. KinectFusion: Real-time dense surface mapping and tracking. In Proceedings of the IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 127-136, 2011. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.524, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[42] Neil Nie, Samir Yitzhak Gadre, Kiana Ehsani, and Shuran Song. Structure from action: Learning interactions for articulated object 3D structure discovery, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.634 + ], + "angle": 0, + "content": "[43] Michael Oechsle, Songyou Peng, and Andreas Geiger. UNISURF: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the International Conference on Computer Vision (ICCV), pages 5589-5599, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.636, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[44] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 165-174, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.719, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[45] Sudeep Pillai, Matthew R. Walter, and Seth Teller. Learning articulated motions from visual demonstrations. In Proceedings of Robotics: Science and Systems (RSS), 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[46] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross B. Girshick, Piotr Dólar, and Christoph Feichtenhofer. SAM 2: Segment anything in images and videos. In Proceedings of the International Conference on Learning Representations (ICLR), 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[47] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8850" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.484, + 0.134 + ], + "angle": 0, + "content": "matching with graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4937-4946, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.135, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[48] Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4104-4113, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[49] Chaoyue Song, Jiacheng Wei, Chuan Sheng Foo, Guosheng Lin, and Fayao Liu. REACTO: Reconstructing articulated objects from a single video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5384-5395, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.482, + 0.313 + ], + "angle": 0, + "content": "[50] Jürgen Sturm, Cyril Stachniss, and Wolfram Burgard. A probabilistic framework for learning kinematic models of articulated objects. Journal of Artificial Intelligence Research, 41:477-526, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.315, + 0.482, + 0.395 + ], + "angle": 0, + "content": "[51] Ryo Suzuki, Adnan Karim, Tian Xia, Hooman Hedayati, and Nicolai Marquardt. Augmented reality and robotics: A survey and taxonomy for ar-enhanced human-robot interaction and robotic interfaces. In Proceedings of the Chi Conference on Human Factors in Computing Systems, pages 1–33, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.398, + 0.482, + 0.492 + ], + "angle": 0, + "content": "[52] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3D shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11353-11362, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.494, + 0.482, + 0.604 + ], + "angle": 0, + "content": "[53] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Justin Kerr, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, David McAllister, and Angjoo Kanazawa. Nerfstudio: A modular framework for neural radiance field development. In Proceedings of the International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), pages 1-12, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.606, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[54] Wei-Cheng Tseng, Hung-Ju Liao, Lin Yen-Chen, and Min Sun. CLA-NeRF: Category-level articulated neural radiance field. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 8454–8460, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[55] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In Advances in Neural Information Processing Systems (NeurIPS), pages 27171-27183, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.744, + 0.482, + 0.813 + ], + "angle": 0, + "content": "[56] Xiaogang Wang, Bin Zhou, Yahao Shi, Xiaowu Chen, Qinping Zhao, and Kai Xu. Shape2Motion: Joint analysis of motion parts and attributes from 3D shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8868-8876, 2019. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.813 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[57] Xian Wang, Luyao Shen, and Lik-Hang Lee. A systematic review of XR-based remote human-robot interaction systems. ACM Computing Surveys, 57(11):1-37, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.193 + ], + "angle": 0, + "content": "[58] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.196, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[59] Fangyin Wei, Rohan Chabra, Lingni Ma, Christoph Lassner, Michael Zollhöfer, Szymon Rusinkiewicz, Chris Sweeney, Richard Newcombe, and Mira Slavcheva. Self-supervised neural articulated shape and appearance models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15795-15805, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.296, + 0.905, + 0.379 + ], + "angle": 0, + "content": "[60] Yijia Weng, Bowen Wen, Jonathan Tremblay, Valts Blukis, Dieter Fox, Leonidas Guibas, and Stan Birchfield. Neural implicit representation for building digital twins of unknown articulated objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3141-3150, 2024. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.382, + 0.905, + 0.478 + ], + "angle": 0, + "content": "[61] Fanbo Xiang, Yuzhe Qin, Kaichun Mo, Yikuan Xia, Hao Zhu, Fangchen Liu, Minghua Liu, Hanxiao Jiang, Yifu Yuan, He Wang, Li Yi, Angel X. Chang, Leonidas J. Guibas, and Hao Su. SAPIEN: A simulated part-based interactive environment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11094–11104, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.482, + 0.905, + 0.55 + ], + "angle": 0, + "content": "[62] Zihao Yan, Ruizhen Hu, Xingguang Yan, Luanmin Chen, Oliver van Kaick, Hao Zhang, and Hui Huang. RPM-net: Recurrent prediction of motion and parts from point cloud. ACM Transactions On Graphics, 38(6):240:1-240:15, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.61 + ], + "angle": 0, + "content": "[63] Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jeffrey Hu, Matthew Tancik, and Angjoo Kanazawa. Gsplat: An open-source library for Gaussian splattering, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.613, + 0.905, + 0.667 + ], + "angle": 0, + "content": "[64] Li Yi, Haibin Huang, Difan Liu, Evangelos Kalogerakis, Hao Su, and Leonidas Guibas. Deep part induction from articulated object pairs. ACM Transactions on Graphics, 37(6): 1-15, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.671, + 0.905, + 0.74 + ], + "angle": 0, + "content": "[65] Sergey Zakharov, Katherine Liu, Adrien Gaidon, and Rares Ambrus. ReFiNe: Recursive field networks for cross-modal multi-scene representation. In Proceedings of the International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), pages 1-9, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.743, + 0.907, + 0.813 + ], + "angle": 0, + "content": "[66] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.813 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "8851" + } + ] +] \ No newline at end of file diff --git a/2025/SplArt_ Articulation Estimation and Part-Level Reconstruction with 3D Gaussian Splatting/99f05b1c-e917-486f-913a-4812f9c2d857_origin.pdf b/2025/SplArt_ Articulation Estimation and Part-Level Reconstruction with 3D Gaussian Splatting/99f05b1c-e917-486f-913a-4812f9c2d857_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e09ec8a37c949375dd83f41685bb3ca0979a9600 --- /dev/null +++ b/2025/SplArt_ Articulation Estimation and Part-Level Reconstruction with 3D Gaussian Splatting/99f05b1c-e917-486f-913a-4812f9c2d857_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8a80c981c8662d8758b54fde1f31a0ae094d98b528e564f333f1a2331ca7320 +size 10410022 diff --git a/2025/SplArt_ Articulation Estimation and Part-Level Reconstruction with 3D Gaussian Splatting/full.md b/2025/SplArt_ Articulation Estimation and Part-Level Reconstruction with 3D Gaussian Splatting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c4d0aea3bf5716f1ea245a81b966823e0d1f450a --- /dev/null +++ b/2025/SplArt_ Articulation Estimation and Part-Level Reconstruction with 3D Gaussian Splatting/full.md @@ -0,0 +1,410 @@ +# SPLART: Articulation Estimation and Part-Level Reconstruction with 3D Gaussian Splatting + +Shengjie Lin†, Jiading Fang†, Muhammad Zubair Irshad‡, Vitor Campagnolo Guizilini‡, Rares Andrei Ambrus‡, Greg Shakhnarovich†‡, Matthew R. Walter† + +† Toyota Technological Institute at Chicago, ‡ Toyota Research Institute + +{slin, fjd, greg, mwalter}@ttic.edu, {zubair.irshad, vitor.quizilini, rares.ambrus}@tri.global + +# Abstract + +Reconstructing articulated objects prevalent in daily environments is crucial for applications in augmented/virtual reality and robotics. However, existing methods face scalability limitations (requiring 3D supervision or costly annotations), robustness issues (being susceptible to local optima), and rendering shortcomings (lacking speed or photorealism). We introduce SPLART, a self-supervised, category-agnostic framework that uses 3D Gaussian Splattig (3DGS) to reconstruct and infer the kinematics of articulated objects from two sets of posed RGB images captured at different articulation states, enabling real-time photorealistic rendering for novel viewpoints and articulations. SPLART augments 3DGS with a differentiable mobility parameter per Gaussian, achieving refined part segmentation. A multi-stage optimization strategy is employed to progressively handle reconstruction, part segmentation, and articulation estimation, significantly enhancing robustness and accuracy. SPLART exploits geometric self-supervision, effectively addressing challenging scenarios without requiring 3D annotations or category-specific priors. Evaluations on established and newly proposed benchmarks, along with applications to real-world scenarios using a handheld RGB camera, demonstrate SPLART's state-of-the-art performance and real-world practicality. Code is publicly available at https://github.com/ripl/splart. + +# 1. Introduction + +Articulated objects, such as drawers, doors, and scissors, are ubiquitous in our daily lives, yet their dynamic nature poses significant challenges for 3D reconstruction—a critical task for applications in augmented/virtual reality [36, 51], robotics [1, 5, 16, 24, 50, 57], and computer vision [20, 45]. Existing methods for reconstructing articulated objects are hindered by several key limitations: they often require labor-intensive supervision (e.g., part-level segmentation or articulation annotations) [23, 29, + +![](images/8c14f10e462f496c2e116707619e37f0dff2c7fbbf757212e58adff3a0acdffe.jpg) + +![](images/739155a868204e504885cf6d17dc0b8e5a8bd4aaa5adfd8248302b293390ad17.jpg) + +![](images/4e480fa9bf958ffd84a582c3e8ba48fd1f638390542b6b1e11a139f0c8135f82.jpg) + +![](images/edfef345eb7328e4432c033b9531328821d4ce0cfe56dd7ff2df97a62174cb15.jpg) + +SPLART + +![](images/5899ad4519229e08ade1d7230a0f3479cfcc6013fc039e1cd7190c5aaec7a93d.jpg) +Articulation + +![](images/a469f80e92abc52db462b98744b78fbfd58d798534210babb4fdfc7f01304873.jpg) +Figure 1. Given (top) RGB images of an object at two articulation states, (middle) SPLART uses 3DGS to simultaneously reconstruct its static and dynamic parts and estimate the kinematic articulation model. SPLART is then able to (bottom) render high-fidelity 3D reconstructions of the object along with part-level segmentations for novel articulation states, allowing for novel view synthesis. + +![](images/f36a1e71cdc3e76fc452d65ea457f42324dafc504f926d76c06b5eed854bf1ea.jpg) + +![](images/8ffa9581e67e0231c89d4649935deddf7c23c714f3b806973cd9ca3d7fcbe7aa.jpg) + +40, 54, 56], they depend on 3D supervision that restricts practical use [23, 29, 33, 40, 60], they produce category-specific models that limit scalability [29, 40, 54, 59], or they are not capable of real-time, photorealistic rendering [6, 13, 23, 29, 30, 40, 54, 56, 59, 60]. To address these challenges, we introduce SPLART, a novel self-supervised and category-agnostic framework that leverages 3D Gaussian Splitting (3DGS) [26] to reconstruct articulated objects from minimal input—two sets of posed RGB images at distinct articulation states. SPLART reconstructs the object's parts and infers its kinematics, enabling real-time, photorealistic rendering for novel views and articulation states. + +Central to SPLART is the augmentation of 3DGS [26] to include a differentiable mobility parameter for each Gaussian, which enables a more refined segmentation of static and mobile parts through gradient-based optimization. This results in enhanced reconstruction quality, while preserving the real-time, photorealistic rendering capabilities of 3DGS—offering a speedup of more than $100\times$ over methods [6, 30] based on neural radiance fields [38]. + +To enhance robustness, SPLART employs a multi-stage optimization strategy that decouples the part-level reconstruction and articulation estimation processes. Unlike end-to-end approaches prone to local optima [30], SPLART first independently reconstructs each articulation state, then estimates each Gaussian's mobility parameter for part segmentation, and finally refines both the articulation and mobility estimates jointly. This structured approach ensures stable and accurate convergence, avoiding the stringent initialization requirements of existing methods, thereby providing a practical solution for challenging articulated structures. + +Building on this foundation, SPLART leverages geometric self-supervision to eliminate the need for manual annotations or 3D supervision. By enforcing geometric consistency between reconstructions through complementary loss formulations, SPLART robustly estimates articulation parameters across diverse scenarios. This self-supervised strategy enhances scalability, enabling SPLART to reconstruct a wide range of articulated objects without relying on prior structural or categorical knowledge. + +Extensive evaluations on both established and newly introduced benchmarks demonstrate SPLART's superior articulation accuracy and reconstruction quality, surpassing state-of-the-art methods without requiring 3D supervision. Real-world experiments further validate its practicality, showcasing successful reconstructions of diverse articulated objects using only a handheld RGB camera. + +In summary, this work contributes: + +1. An extension of 3DGS with a differentiable mobility value per Gaussian that enables precise part segmentation using gradient-based optimization. +2. A multi-stage optimization strategy that decouples reconstruction and articulation estimation, enhancing robustness and accuracy. +3. Complementary formulations of geometric self-supervision for articulation estimation, eliminating the need for 3D supervision or laborious annotations. +4. A challenging dataset and new metrics for comprehensive evaluation of articulated object reconstruction. + +# 2. Related Work + +# 2.1. Data-Driven Articulation Learning + +Estimating the pose and joint properties of articulated objects is crucial for robot manipulation and interac + +tion [10, 11, 19, 31]. Recent learning-based methods [9, 13, 15, 22, 29, 32, 56, 64] infer articulation properties from point clouds via end-to-end training. For instance, Shape2Motion [56] analyzes motion parts from a single point cloud in a supervised setting, while ANCSH [29] performs category-level pose estimation, but requires class-specific models. RPM-Net [62] enhances generalization across categories for part segmentation and kinematic prediction, and DITTO [23] predicts motion and geometry from 3D point cloud pairs without labels. However, these methods depend on costly 3D supervision and annotations. In contrast, our approach reconstructs accurate 3D geometry and detailed appearance, capturing articulation without 3D supervision or priors. + +# 2.2. Representations for Object Reconstruction + +Early 3D object reconstruction methods predicted point clouds, voxels, or meshes from partial observations [3, 8, 17]. Recent advances in implicit scene representations [18, 26, 37, 38, 44, 65] enable detailed geometry and appearance reconstruction via differentiable rendering [12, 43, 52, 55]. While neural fields suffer from slow rendering, 3D Gaussian Splatting (3DGS) [26] overcomes this by using explicit 3D Gaussians. We leverage 3DGS for self-supervised articulated object reconstruction from posed RGB images, achieving fast, realistic synthesis of novel views and articulations in real time. + +# 2.3. Articulated Object Reconstruction + +Recent methods leverage differentiable 3D representations [26, 38, 44] to jointly reconstruct articulated objects and infer articulation parameters. Training-based approaches use synthetic 3D data to predict joint parameters and segment parts [13, 14, 23, 25, 40, 42, 59]. Self-supervised methods optimize shape, appearance, and articulation per scene without extensive training [6, 27, 30, 32, 33, 49, 60]. Some handle objects with more than two parts, but require the number of parts to be known and the mobile parts to be directly connected to the static part [6, 33, 60]. Other works enhance articulation estimation using large language or vision-language models [28, 35]. In contrast, our self-supervised method reconstructs two-part articulated objects from RGB images across articulation states using 3DGS [26]. As the first to apply 3DGS to this task without 3D supervision or pre-trained priors, it robustly handles challenging cases and achieves real-time performance. + +# 3. Methodology + +# 3.1. Overview + +Consider an arbitrary object composed of two rigid parts: a static parent part, and a child part that can move relative to its parent through either a revolute or prismatic articulation. + +![](images/949f48fd63c7627d87b3ddba7965c91d5267f78a243ccf93e30db8e0b339e1e6.jpg) +(a) Stage 1: Per-state reconstruction. + +![](images/65d52b5e2530a042a40745b877f08db5d0627f3ad8648ed93ec38e1365793423.jpg) +(Section 3.2.1) + +![](images/a682f366330860f5adc56c9e1de27e762e3ab7efe3a607caeaba740e2653b232.jpg) +Figure 2. Methodology overview of SPLART. It consists of three decoupled stages for the purpose of stable optimization. Stage 1 constructs a 3DGS for each state from posed RGB images using the photometric loss. Stage 2 proposes cross-static formulation for mobility estimation. Intuitively, it combines static parts of both states and the mobile part of the desired state for the target Gaussians. Stage 3 proposes cross-mobile formulation for articulation estimation. Intuitively, it combines static parts of both states, the mobile part of the desired state and the transformed mobile part of the other state together for the target Gaussians. + +![](images/89dd5b274ef6c9ec9cb37e73fac5119891da0a47e96c665fd4043b933b2cf2a3.jpg) + +![](images/56152ec1f68d402376abfd4ece5c2250a479c0577d0df1d3c58e428f593016d6.jpg) +(b) Stage 2: Cross-static formulation for mobility estimation. + +![](images/f9ba36a2f893efaaab5e87cb625da3b3743ac339695c80e7b75b3b6cd0c08780.jpg) +(c) Stage 3: Cross-mobile formulation for articulation estimation and mobility refinement. + +Our objective is twofold: (1) to reconstruct the articulated object at the part level; and (2) to estimate its articulated motion. Assuming a known articulation type (i.e., either revolute or prismatic), the input to our method consists of two sets of posed RGB images (i.e., images with known camera intrinsics and extrinsics), each capturing the articulated object at one end state of the motion. + +Formally, let $l$ denote the articulation state label, where $l = 0$ and $l = 1$ correspond to the two end states of the observed articulation. For reconstruction, SPLART uses observations $\mathcal{O}_l = \{(I_l^i,P_l^i,K_l^i)\}_{i = 1}^{N_l}, l\in \{0,1\}$ , where $I_{l}^{i}$ is the $i$ -th observed RGB image of the articulated object at state $l$ , $P_{l}^{i}$ and $K_{l}^{i}$ denote its camera extrinsic and intrinsic, respectively, and $N_{l}$ represents the number of data samples for state $l$ . Note that $P_{l}^{i}$ is specified in a common world frame for both states, while the articulated motion involves only one moving part w.r.t. the world space. SPLART models a revolute articulation with its rotation axis $\mathbf{a}$ ( $\| \mathbf{a}\| = 1$ ), pivot $\mathbf{p}$ , and rotation angle $\theta$ , such that a point $\mathbf{x}$ (in the world frame) on the mobile part at state $l = 0$ will move to + +$$ +{ } ^ { 1 } T _ { 0 } \mathbf { x } = R _ { \mathbf { a } , \theta } ( \mathbf { x } - \mathbf { p } ) + \mathbf { p } \tag {1} +$$ + +at state $l = 1$ , where $R_{\mathbf{a},\theta}$ is the rotation induced by the axis-angle notation. A prismatic articulation is instead modeled by its translation axis $\mathbf{a}$ ( $\| \mathbf{a} \| = 1$ ) and distance $d$ . The goal is to reconstruct the articulated object at the part level using a chosen representation while estimating the articulated motion $^1 T_0$ , ensuring that the renderings at each articulation state are with the observations. + +SPLART extends the 3DGS representation for articulated objects, decoupling the goals of part-level reconstruction and articulation estimation across three stages: (1) separate reconstructions for each articulation state (Sec. 3.2.1), (2) mobility estimation using the cross-static formulation (Sec. 3.2.2), and (3) articulation estimation and mobility re + +finement using the cross-mobile formulation (Sec. 3.2.3). To facilitate the application of SPLART to real-world objects, we leverage modern structure-from-motion and image segmentation techniques, developing a framework that enables general users to reconstruct articulated objects in their surroundings using only images captured by a handheld camera device (Sec. 3.3). + +# 3.2. Decoupled Multi-Stage Optimization + +Jointly performing part-level reconstruction and articulation estimation on an articulated object is prone to local minima [6, 30], making it preferable to strategically decouple the problem into multiple stages. + +# 3.2.1. Separate per-state reconstruction + +In Stage 1, SPLART separately optimizes two 3DGS models following the standard procedure, one for each end state of the articulation. In addition to the attributes that are standard to 3DGS, each Gaussian is initialized with a persistent binary state label $l$ , equally drawn from $\{0,1\}$ . We denote the set of Gaussians representing state $l$ as $\mathcal{G}_l^{\mathrm{ref}}$ , where $ref$ emphasizes that $\mathcal{G}_l^{\mathrm{ref}}$ is the reference reconstruction unaffected by the other state. Given a data sample observed at state $l$ , the optimization follows as: + +$$ +\min _ {\mathcal {G} _ {l} ^ {\text {r e f}}} \Delta_ {I} \left(\hat {I} _ {l} ^ {\text {r e f}, i}, I _ {l} ^ {i}\right), \tag {2a} +$$ + +$$ +\hat {I} _ {l} ^ {\text {r e f}, i} = R \left(\mathcal {G} _ {l} ^ {\text {r e f}}, P _ {l} ^ {i}, K _ {l} ^ {i}\right), l \in \{0, 1 \}, \tag {2b} +$$ + +where $R$ is the 3DGS rendering function, and $\Delta_I$ represents the photometric loss. For simplicity, the view index $i$ and camera parameters $P_l^i, K_l^i$ will be omitted from now on. + +# 3.2.2. Cross-static formulation for mobility estimation + +The focus of Stage 2 is mobility estimation for each Gaussian. To keep $\mathcal{G}^{\mathrm{ref}}$ dedicated to single-state reconstruction, + +we first duplicate $\mathcal{G}_l^{\mathrm{ref}}$ as $\mathcal{G}_l^{\mathrm{tgt}}$ for both states (i.e., $l\in \{0,1\}$ ), which is intended as the target representation that will fulfill the goals of part-level reconstruction and articulation estimation. By design, $\mathcal{G}^{\mathrm{tgt}}$ shares neither data storage nor gradient flow with $\mathcal{G}^{\mathrm{ref}}$ after its creation. + +For each Gaussian in $\mathcal{G}^{\mathrm{tgt}}$ , we further extend its set of attributes with a mobility value $m \in [0,1]$ , initialized at 0.5. By design, $m$ enables the break-down of a Gaussian to its static and mobile components, where each component inherits all the original Gaussian attributes except for the opacity $\sigma$ . The static component has its opacity modified to $\sigma \cdot (1 - m)$ , and the mobile component to $\sigma \cdot m$ . For simplicity, letting $\mathcal{M}$ be the set of mobilities for $\mathcal{G}$ , we use the element-wise product $\mathcal{G} \cdot (1 - \mathcal{M})$ to denote the static component of $\mathcal{G}$ , and $\mathcal{G} \cdot \mathcal{M}$ for the mobile component. + +To estimate the mobilities $\mathcal{M}$ , we employ the intuition that the static components from both states should constitute the static part of the articulated object. Formally, we introduce the cross-static (cs) formulation, where the static part of the articulated object is jointly represented as + +$$ +\mathcal {G} ^ {\mathrm {s}} = \mathcal {G} _ {l} ^ {\mathrm {t g t}} \cdot (1 - \mathcal {M} _ {l}) \oplus \mathcal {G} _ {1 - l} ^ {\mathrm {t g t}} \cdot (1 - \mathcal {M} _ {1 - l}), \tag {3} +$$ + +where $\oplus$ denotes concatenation. For state $l$ , the target representation thus becomes + +$$ +\mathcal {G} _ {l} ^ {\mathrm {t g t}, \mathrm {c s}} = \mathcal {G} ^ {\mathrm {s}} \oplus \mathcal {G} _ {l} ^ {\mathrm {t g t}} \cdot \mathcal {M} _ {l}. \tag {4} +$$ + +With this formulation, Stage 2 is further divided into two sub-stages as described below. + +Stage 2(a): Coarse mobility estimation via cross-static geometric consistency. To measure the geometric distance between two Gaussian sets, we design a weighted version of the Chamfer distance. Specifically, letting $X = \{(x_{i},w_{x_{i}})\}_{i = 1}^{M}$ and $Y = \{(y_{j},w_{y_{j}})\}_{j = 1}^{N}$ be two sets of point-weight pairs, the weighted Chamfer distance is then: + +$$ +\begin{array}{l} \operatorname {Chamfer}(X,Y) = \sum_{\left(x_{i},w_{x_{i}}\right)\in X}\tilde{w}_{x_{i}}\min_{\left(y_{j},w_{y_{j}}\right)\in Y}\| x_{i} - y_{j}\|^{2} \\ + \sum_ {\left(y _ {j}, w _ {y _ {j}}\right) \in Y} \tilde {w} _ {y _ {j}} \min _ {\left(x _ {i}, w _ {x _ {i}}\right) \in X} \| x _ {i} - y _ {j} \| ^ {2}, \tag {5} \\ \end{array} +$$ + +where $\tilde{w}_{x_i} = w_{x_i} / \sum_{(x_i, w_{x_i}) \in X} w_{x_i}$ and $\tilde{w}_{y_j}$ are normalized weights. For each Gaussian, we use its mobility-modified opacity $\sigma'$ as the weight, essentially treating it as $\sigma'$ points overlapped at its mean position. The mobilities are then optimized by encouraging the geometric consistency formulated as follows: + +$$ +\min _ {\mathcal {M}} \mathrm {C D} _ {0} ^ {\mathrm {c s}} + \mathrm {C D} _ {1} ^ {\mathrm {c s}} + \lambda_ {m} ^ {\text {g e o m}} \| \mathcal {M} \|, \tag {6a} +$$ + +$$ +\mathrm {C D} _ {l} ^ {\mathrm {c s}} = \operatorname {C h a m f e r} \left(\mathcal {G} _ {l} ^ {\mathrm {t g t , c s}}, \mathcal {G} _ {l} ^ {\mathrm {r e f}}\right), \tag {6b} +$$ + +where $\mathrm{Chamfer}(\cdot)$ denotes the weighted Chamfer distance, and $\lambda_{m}^{\mathrm{geom}}\| \mathcal{M}\|$ is a regularization term that encourages + +smaller mobilities. Note how $\mathcal{M}$ affects the weighted Chamfer distance by modifying the opacities, and that $\mathcal{M} = 1$ is a trivial solution without the regularization. Without photometric supervision, the mobilities obtained from Eqn. 6a are generally noisy. However, being relatively fast (taking only tens of seconds), they still serve as a good initialization for the next sub-stage. + +Stage 2(b): Joint mobility and Gaussian optimization via cross-static rendering. To more accurately estimate the mobilities while jointly refining the full Gaussian attributes, we perform cross-static rendering as follows: + +$$ +\min _ {\mathcal {G} ^ {\mathrm {t g t}}, \mathcal {M} _ {1 - l}} \Delta_ {I} \left(\hat {I} _ {l} ^ {\mathrm {c s}}, I _ {l}\right) + \lambda_ {m} ^ {\text {p h o t o}} \| \mathcal {M} _ {1 - l} \|, \tag {7a} +$$ + +$$ +\hat {I} _ {l} ^ {\mathrm {c s}} = R \left(\mathcal {G} _ {l} ^ {\mathrm {t g t}, \mathrm {c s}}\right), \tag {7b} +$$ + +where $\lambda_{m}^{\mathrm{photo}}\| \mathcal{M}_{1 - l}\|$ is a mobility regularization term similar to that in Eqn. 6a. + +# 3.2.3. Cross-mobile formulation for articulation estimation and mobility refinement + +The focus of Stage 3 is to estimate the articulation parameters shared by the mobile components of all Gaussians. To this end, we employ the intuition that the mobile components from the two end states are related through the articulated motion. Formally, we introduce the cross-mobile (cm) formulation, where the mobile part of the articulated object at state $l$ is jointly represented as + +$$ +\mathcal {G} _ {l} ^ {\mathrm {m}} = \mathcal {G} _ {l} ^ {\mathrm {t g t}} \cdot \mathcal {M} _ {l} \oplus {} ^ {l} T _ {1 - l} \left(\mathcal {G} _ {1 - l} ^ {\mathrm {t g t}} \cdot \mathcal {M} _ {1 - l}\right), \tag {8} +$$ + +where ${}^l T_{1 - l}(\mathcal{G})$ denotes the transformation of Gaussians $\mathcal{G}$ according to the articulated motion from state $1 - l$ to $l$ . For state $l$ , the target representation then becomes + +$$ +\mathcal {G} _ {l} ^ {\mathrm {t g t}, \mathrm {c m}} = \mathcal {G} ^ {\mathrm {s}} \oplus \mathcal {G} _ {l} ^ {\mathrm {m}}, \tag {9} +$$ + +where $\mathcal{G}^{\mathrm{s}}$ is defined in Eqn. 3. With this formulation, Stage 3 is further divided into three sub-stages as below. + +Stage 3(a): Coarse articulation estimation via geometric consistency. Similar to Eqn. 6a, weighted Chamfer distance is used for computing the cross-mobile geometric consistency, which can be optimized over both mobilities and articulation parameters as follows: + +$$ +\min _ {\mathcal {M}, T} \mathrm {C D} _ {0} ^ {\mathrm {c m}} + \mathrm {C D} _ {1} ^ {\mathrm {c m}}, \tag {10a} +$$ + +$$ +\mathrm {C D} _ {l} ^ {\mathrm {c m}} = \operatorname {C h a m f e r} \left(\mathcal {G} _ {l} ^ {\mathrm {t g t , c m}}, \mathcal {G} _ {l} ^ {\mathrm {r e f}}\right). \tag {10b} +$$ + +However, we notice that this formulation is still susceptible to local minima if the mobile part is too geometrically insignificant. We qualitatively show one such failure in Fig. 3a. To remedy this, we further propose mobile-only geometric consistency, which considers only the mobile components of both states as follows: + +$$ +\mathrm {C D} ^ {\mathrm {m}} = \operatorname {C h a m f e r} \left(\mathcal {G} _ {l} ^ {\mathrm {t g t}} \cdot \mathcal {M} _ {l}, ^ {l} T _ {1 - l} \left(\mathcal {G} _ {1 - l} ^ {\mathrm {t g t}} \cdot \mathcal {M} _ {1 - l}\right)\right). \tag {11} +$$ + +![](images/dc57c7c2627f33a42e1bccd935173f966d2e15276d8c88445328b5773ef508e3.jpg) +(a) Failure with cross-mobile geometric consistency. +(b) Success with mobile-only geometric consistency. + +While successfully addressing these scenarios, mobile-only geometric consistency still falls short in other circumstances, especially when the mobile components in the two states exhibit large discrepancies. We qualitatively show one such failure in Fig. 4a. To take advantage of both formulations and to facilitate robustness in the inherently nonconvex optimization of articulation parameters, we propose the following practical strategy: + +1. With $K^{\mathrm{m}}$ randomized tries: articulation estimation via + +$$ +T ^ {m} = \arg \min _ {T} \mathrm {C D} ^ {\mathrm {m}}. \tag {12} +$$ + +2. With $K^{\mathrm{cm}}$ randomized tries, plus another initialized with $T^{\mathrm{m}}$ : articulation estimation via + +$$ +T ^ {\mathrm {c m}} = \arg \min _ {T} \mathrm {C D} _ {0} ^ {\mathrm {c m}} + \mathrm {C D} _ {1} ^ {\mathrm {c m}}. \tag {13} +$$ + +3. Final run initialized with $T^{\mathrm{cm}}$ : joint articulation estimation and mobility refinement via Eqn. 10a. + +Stage 3(b): Joint articulation, mobility, and Gaussian optimization via cross-mobile rendering. As in Stage 2(b), we perform a full joint optimization over articulation parameters, mobilities, and Gaussians utilizing photometric supervision via cross-mobile rendering as follows: + +$$ +\min _ {\mathcal {G} ^ {\mathrm {i g t}}, \mathcal {M} _ {1 - l}, ^ {l} T _ {1 - l}} \Delta_ {I} \left(\hat {I} _ {l} ^ {\mathrm {c m}}, I _ {l}\right), \tag {14a} +$$ + +$$ +\hat {I} _ {l} ^ {\mathrm {c m}} = R \left(\mathcal {G} _ {l} ^ {\mathrm {t g t , c m}}\right), l \in \{0, 1 \}. \tag {14b} +$$ + +Stage 3(c): Mobility correction via cross-mobile geometric consistency. Stage 3(b) solely relies on photometric supervision, which is limited to the training views. On rare occasions, a Gaussian may be mistakenly estimated as mobile if the articulated motion moves it out of sight from most views, effectively losing supervision. On the other hand, geometric consistency is not affected by visibility limitations, which we leverage for mobility correction as follows: + +$$ +\min _ {\mathcal {M}} \mathrm {C D} _ {0} ^ {\mathrm {c m}} + \mathrm {C D} _ {1} ^ {\mathrm {c m}}. \tag {15} +$$ + +![](images/e05432a5de3057a0ce9daff2e3a02637f34a98379f58e3a003239a47c4a616da.jpg) +Figure 3. 103549-Toaster has a flat slider as its mobile part, similar in geometric curvature to the static component. This similarity makes the nearest-neighbor data association error-prone when computing the Chamfer distance for cross-mobile geometric consistency. On the other hand, mobile-only geometric consistency focuses only on the mobile part and successfully handles this case. +(a) Failure with mobile-only geometric consistency. +(b) Success with cross-mobile geometric consistency. +Figure 4. 101713-Pen has both ends of a pen as the mobile part, with only one end visible from any given observation. Thus, the reconstructed mobile components from the two states essentially represent distinct parts, violating the intuition for mobile-only geometric consistency. On the other hand, cross-mobile geometric consistency considers the joint representation from both states as a whole, successfully estimating the articulation. + +# 3.3. Real-World Application + +For an articulated object to be reconstructed in the real world, we first collect two sets of RGB images, one for each articulation state, by imaging the object from a surrounding hemisphere. We then preprocess the data to determine the posed images that serve as the input to SPLART. This involves using SAM 2 [46] to perform foreground-background segmentation, which also removes dynamic contents of the scene. We then perform structure from motion (SfM) to determine camera poses using COLMAP [48] with SuperPoint [7] descriptors and SuperGlue [47] matching on the segmented backgrounds from both sets of images to construct a joint coordinate frame for the object in both states. Once we obtain the joint coordinate frame and the foreground target object is localized, we run SPLART to reconstruct the articulated objects. + +# 4. Experiments + +# 4.1. Datasets + +PARIS PartNet-Mobility Subset. Our evaluation includes PARIS [30] as a baseline, which selects 10 instances (8 revolute, 2 prismatic) from PartNet-Mobility [2, 39, 61], a large-scale dataset of simulatable 3D objects with part-level mobility. We refer to this subset as PARIS-PMS. For each articulation state, the dataset provides 100 calibrated object-centric views for training and 50 for testing, sampled from the upper hemisphere. However, the released dataset lacks ground-truth depth and part segmentation maps. To enable their evaluations, we augment PARIS-PMS with the necessary ground-truth data by following its data generation procedure. The dataset does not provide test views of intermediate articulation states, which limits the quantitative evaluation of novel articulation synthesis. + +SPLART PartNet-Mobility Subset. We curate an additional articulated object dataset from PartNet-Mobility, dubbed SPLART-PMS, to enable a more comprehensive + +
TypeMethodSuccess Rate %↑erra(x10-2DEG)↓errp(x10-3)↓errr(x10-2DEG)↓errt(x10-3)↓CDs(x10-3)↓CDm(x10-3)↓CDw(x10-3)↓
RevolutePARIS40.0121.90 ± 68.455.27 ± 4.55127.80 ± 55.12N/A3.48 ± 0.3968.44 ± 13.0313.11 ± 1.16
DTA†98.812.32 ± 4.161.86 ± 1.6919.41 ± 16.28N/A2.20 ± 0.491.27 ± 1.871.73 ± 0.03
SPLART100.03.70 ± 0.380.40 ± 0.074.85 ± 0.33N/A4.08 ± 0.321.06 ± 0.783.59 ± 0.21
PrismaticPARIS50.027.97 ± 13.09N/AN/A4.28 ± 3.029.21 ± 1.94151.78 ± 35.007.99 ± 0.49
DTA†100.016.26 ± 3.54N/AN/A1.15 ± 0.142.69 ± 0.0415.74 ± 0.282.22 ± 0.03
SPLART95.02.36 ± 0.44N/AN/A0.33 ± 0.046.02 ± 0.2016.77 ± 1.143.69 ± 0.14
OverallPARIS42.0103.12 ± 57.385.27 ± 4.55127.80 ± 55.124.28 ± 3.024.63 ± 0.7085.11 ± 17.4212.09 ± 1.03
DTA†99.013.11 ± 4.041.86 ± 1.6919.41 ± 16.281.15 ± 0.142.30 ± 0.404.16 ± 1.551.83 ± 0.03
SPLART99.03.44 ± 0.390.40 ± 0.074.85 ± 0.330.33 ± 0.044.47 ± 0.304.20 ± 0.853.61 ± 0.19
+ +Table 1. PARIS-PMS Articulation and Mesh Reconstruction Metrics. Best results in bold. ${}^{ \dagger }$ DTA requires ground-truth depth. + +
TypeMethodPSNR ↑Depth MAE ↓mIoU ↑
RevolutePARIS32.210.0930.955
DTA†N/A0.0310.941
SPLART43.530.0390.974
PrismaticPARIS33.750.1080.902
DTA†N/A0.0660.844
SPLART44.380.0250.890
OverallPARIS32.520.0960.942
DTA†N/A0.0380.922
SPLART43.700.0360.957
+ +Table 2. PARIS-PMS Novel View Synthesis Metrics, averaged over successful runs (see Tab. 1 for success rates). Best results in bold. ${}^{ \dagger }$ DTA requires ground-truth depth. + +evaluation including the quantitative assessment of novel articulation synthesis. Our dataset consists of 22 object instances from categories not included in PARIS-PMS, with 12 revolute and 10 prismatic objects. For the test split, we also generate ground-truth depth and part segmentation maps to facilitate relevant evaluations. Crucially, while PARIS-PMS restricts articulation states to binary values, SPLART-PMS uniformly samples states from $[-0.1, 1.1]$ , demanding accurate estimation of articulation, part segmentation, and object reconstruction for correct view synthesis. Our experiments reveal that SPLART-PMS poses a greater challenge than PARIS-PMS—the latter being almost perfectly solved by our method. + +Real-World Dataset. To show the efficacy of SPLART on real-world data, we collect a dataset that consists of 7 real-world objects with 9 articulations in total, including common objects like a monitor, cabinet, glasses, etc. All images are captured by a hand-held phone camera. Camera parameters are calibrated following strategies detailed in Sec. 3.3. We qualitatively evaluate SPLART on this dataset. + +# 4.2. Evaluation Metrics + +Articulation Estimation. We evaluate the accuracy of articulation estimation depending on its type. + +- A revolute articulation describes a rotation around some line in 3D space, parameterized by the pivot point $\mathbf{p}$ on the line, the line's directional axis $\mathbf{a}$ , and the angle of + +![](images/ed97e3e78886aff688c7f20c9ab3d22384ec4fa1f464cfe9a66b9e987b33960e.jpg) + +Figure 5. Real-World Articulated Object Reconstructions. + +
TypeMethodSuccess Rate %errn(×10-2DEG) ↓errp(×10-3) ↓errr(×10-2DEG) ↓errt(×10-3) ↓
RevolutePARIS18.3167.34 ± 42.479.36 ± 3.73270.37 ± 91.19N/A
DTA†33.332.76 ± 2.381.28 ± 0.3940.80 ± 4.66N/A
SPLART-2a78.52.91 ± 0.552.53 ± 2.1852.06 ± 32.00N/A
SPLART-2b67.04.62 ± 1.693.92 ± 3.07131.99 ± 64.45N/A
SPLART-3a0.8296.61 ± 0.002.83 ± 0.00427.10 ± 0.00N/A
SPLART-3b76.784.13 ± 19.135.11 ± 3.03206.98 ± 56.30N/A
SPLART77.520.76 ± 0.813.93 ± 2.4178.31 ± 31.06N/A
PrismaticPARIS39.049.49 ± 18.92N/AN/A14.96 ± 5.51
DTA†90.0102.77 ± 12.48N/AN/A9.25 ± 0.73
SPLART-2a90.025.88 ± 23.25N/AN/A1.50 ± 1.63
SPLART-2b85.715.63 ± 20.40N/AN/A0.73 ± 0.58
SPLART-3a13.015.17 ± 0.59N/AN/A2.26 ± 0.04
SPLART-3b86.061.00 ± 19.76N/AN/A9.43 ± 2.37
SPLART95.025.72 ± 17.63N/AN/A1.04 ± 0.32
OverallPARIS27.7103.88 ± 29.799.36 ± 3.73270.37 ± 91.1914.96 ± 5.51
DTA†59.177.76 ± 8.871.28 ± 0.3940.80 ± 4.669.25 ± 0.73
SPLART-2a83.713.85 ± 11.362.53 ± 2.1852.06 ± 32.001.50 ± 1.63
SPLART-2b74.79.34 ± 9.713.92 ± 3.07131.99 ± 64.450.73 ± 0.58
SPLART-3a6.485.53 ± 0.442.83 ± 0.00427.10 ± 0.002.26 ± 0.04
SPLART-3b80.973.62 ± 19.425.11 ± 3.03206.98 ± 56.309.43 ± 2.37
SPLART85.523.02 ± 8.463.93 ± 2.4178.31 ± 31.061.04 ± 0.32
+ +Table 3. SPLART-PMS Articulation Metrics. Best results in bold. ${}^{ \dagger }$ DTA requires ground-truth depth. + +rotation $\theta$ . We report the angular error $\mathrm{err}_{\mathbf{a}}$ (in $\times 10^{-2}$ DEG) between the predicted and ground-truth axes, the geodesic distance $\mathrm{err}_{\mathrm{r}}$ (in $\times 10^{-2}$ DEG) between the predicted and ground-truth rotations induced by the axis-angle ( $\mathbf{a} - \theta$ ) pair, and the pivot point error $\mathrm{err}_{\mathbf{p}}$ . Since the pivot can move arbitrarily along the axis, $\mathrm{err}_{\mathbf{p}}$ is computed as the closest distance between the predicted and ground-truth lines induced by the axis-pivot ( $\mathbf{a} - \mathbf{p}$ ) pair. + +
TypeMethodPSNR ↑Depth MAE ↓mIoU ↑
RevolutePARIS32.750.1250.897
DTA†N/A0.0530.886
SPLART36.380.0320.897
PrismaticPARIS32.300.1240.902
DTA†N/A0.0600.857
SPLART37.770.0330.938
OverallPARIS32.510.1250.900
DTA†N/A0.0580.867
SPLART37.010.0320.916
+ +Table 4. SPLART-PMS Novel View and Articulation Synthesis Metrics, averaged over successful runs (see Tab. 3 for success rates). Best results in bold. $\dagger$ DTA requires ground-truth depth. + +- A prismatic articulation describes a translation along a specific direction, parameterized by the axis $\mathbf{a}$ of the translation direction and the translation distance $d$ . We report the axis error $\mathrm{err}_{\mathbf{a}}$ , as in the revolute case, and the translation error $\mathrm{err}_{\mathrm{t}}$ , which is the distance between the predicted and ground-truth translations induced by the axis-distance $(\mathbf{a} - d)$ pair. + +Part-Level Reconstruction. The evaluation of part-level reconstruction accuracy is threefold: photometric rendering quality, geometric accuracy, and part segmentation accuracy. For all these aspects, we use novel view synthesis as the surrogate task. We perform volume rendering for each view in the test split, generating outputs that include the RGB image, depth map, and part segmentation map. Photometric rendering quality is evaluated by reporting image metrics PSNR. Geometric accuracy is assessed by reporting the mean absolute error of the depth map (i.e., Depth MAE). Part segmentation accuracy is evaluated using the intersection-over-union (IoU) ratio for three categories: static part $(\mathrm{IoU_s})$ , mobile part $(\mathrm{IoU_m})$ , and background $(\mathrm{IoU_{bg}})$ . The mean IoU (mIoU) is then reported as the average IoU across these categories. In addition, we also evaluate geometric accuracy through mesh reconstruction. To extract a mesh from 3DGS, we: (1) render depth images from uniformly sampled spherical viewpoints, (2) fuse rendered depths into a TSDF representation [4, 41], and (3) extract the mesh using the Marching Cubes algorithm [34]. For each reconstruction, we separately extract the meshes corresponding to the static, mobile, and whole parts. For evaluation, we follow the procedure used in previous works [30, 60]: 10 000 points are uniformly sampled from both the reconstructed and the ground-truth meshes, and the Chamfer distance is computed for each category—static $(\mathrm{CD_s})$ , mobile $(\mathrm{CD_m})$ , and whole $(\mathrm{CD_w})$ . + +Impact of Randomness. Randomness can play a significant role in the joint optimization of articulation and part reconstruction [6, 30]. In order to effectively demonstrate the robustness of a method, we conduct ten trials for each + +![](images/2fc06219444e5914e71f5ae360c285e9c2734e3be1238f7c20584afe6651661b.jpg) +Figure 6. State interpolation from the estimated articulation. + +scene, each with a different fixed random seed. We report results averaged over the ten runs. Additionally, we notice that for a failed run converging to a local minimum, the evaluation metrics, such as $\mathrm{err}_{\mathbf{r}}$ , are typically orders-of-magnitude worse than those of a successful run. To better reflect a method's performance, we report the number of successful runs out of the total number (i.e., ten), where a run is deemed successful if and only if each of the following criteria are met: (1) $\mathrm{err}_{\mathbf{a}} < 5$ DEG; (2) $\mathrm{err}_{\mathbf{p}} < 0.05$ for revolute articulation; (3) $\mathrm{err}_{\mathbf{r}} < 10$ DEG for revolute articulation; and (4) $\mathrm{err}_{\mathbf{t}} < 0.05$ for prismatic articulation. We then proceed to report the evaluation metrics detailed previously, averaging only over the successful runs. + +# 4.3. Baselines + +PARIS [30] addresses the same challenge we tackle: reconstructing the part-level shape, appearance, and motion of an articulated object from multi-view RGB images captured at two articulation states. It employs a NeRF-based representation, modeling the object with static and mobile fields for each part, alongside a transformation to capture state transitions. These components are jointly optimized end-to-end, primarily through an image-rendering loss. + +DTA [60] adopts a similar setup: reconstructing digital twins of articulated objects from multi-view RGB-D observations at two articulation states. It operates in two stages: Stage 1 optimizes neural object fields for each state using RGB-D images and extracts meshes; Stage 2 optimizes a part segmentation field and per-part motions using consistency, matching, and collision losses to determine point correspondences between states. Note: Unlike SPLART, DTA requires ground-truth depth as input and produces plain object meshes unsuitable for photorealistic rendering. + +# 4.4. Experiment Results + +Articulation Estimation. We quantitatively evaluate SPLART against baselines for articulation estimation accuracy, reporting scene-averaged results per articulation type on the PARIS-PMS dataset (Tab. 1) and the SPLART-PMS dataset (Tab. 3). SPLART consistently outperforms the baselines across both datasets. Additionally, we qualitatively compare SPLART with PARIS on selected scenes from these datasets in Fig. 7. In the "color" column, each + +![](images/bae55b184d543a19622de379ebc3e97a12b08962128574169594e36aba0b8bb2.jpg) +Figure 7. Qualitative comparison on part-level reconstruction and articulation estimation. + +image visualizes the ground-truth articulation (green arrow) and the estimated articulation (red arrow), except in the ground-truth column. When the estimate aligns perfectly with the ground truth, only the red arrow appears; if significantly misaligned, only the green arrow is visible. + +Part-Level Reconstruction. We assess SPLART's part-level reconstruction accuracy against baselines, using novel view synthesis as a surrogate task. Scene-averaged results per articulation type are reported on the PARIS-PMS dataset (Tab. 2) and the SPLART-PMS dataset (Tab. 4). We also evaluate geometric accuracy through mesh reconstruction, presenting scene-averaged results per articulation type on the PARIS-PMS dataset (Tab. 1). Without depth supervision, which is required for the DTA baseline, SPLART matches DTA's performance, both substantially exceeding PARIS. Qualitative comparisons with PARIS, including RGB renderings and part segmentation maps, are provided for selected scenes from both datasets in Fig. 7. Additional qualitative results on articulation synthesis appear in Fig. 6. + +Real-World Reconstructions. We qualitatively evaluate SPLART on the aforementioned real-world dataset, with results shown in Fig. 5. Each image overlays the estimated articulation, visualized with a red arrow. + +Ablation Studies. We perform extensive ablation studies to demonstrate the contribution of each stage. Specifically, we perform the following four ablations: (1) No geometric supervision for mobility estimation, i.e., skipping Stage 2(a), denoted SPLART-2a; (2) No photometric supervision for mobility estimation, i.e., skipping Stage 2(b), denoted SPLART-2b; (3) No geometric supervision for articulation estimation, i.e., skipping Stage 3(a), denoted SPLART-3a; (4) No photometric supervision for articulation estimation, i.e., skipping Stage 3(b), denoted SPLART-3b. Like the main experiments, we conduct 10 runs for each scene, but only on the more challenging dataset of SPLART-PMS. We report the articulation estimation results in Tab. 3. + +![](images/0b517baf1144dfd9df63c40504710b4fc005b4854d23a0fd7254f1f8780bfc26.jpg) +Figure 8. Failure cases due to inherent ambiguities in articulation. + +# 5. Conclusion + +We present SPLART, the first self-supervised method that utilizes 3D Gaussian Splatting for articulated object reconstruction from two-state RGB observations, without 3D supervision or category-specific priors. SPLART delivers robust optimization and effectively handles challenging scenarios—outperforming prior approaches while eliminating the need for 3D supervision, articulation annotations, or semantic labels. Extensive evaluations on synthetic and real-world datasets show SPLART's articulation estimation accuracy and view synthesis quality that surpass existing methods. Real-world experiments further validate its practicality, showcasing successful reconstructions of diverse articulated objects using only a handheld RGB camera. Our framework empowers non-expert users to effortlessly create high-fidelity, articulated digital twins, supporting applications in augmented/virtual reality and robotics. However, while SPLART effectively addresses two-part articulations, it lacks direct support for objects with more articulated parts, though iteratively applying it to incremental state pairs may help. Some existing methods [6, 33, 60] handle such objects, but require known part counts and specific articulation structures. Future research will focus on a comprehensive adaptation to multi-level articulated objects. Additionally, inherent ambiguities may arise when multiple articulations explain state differences, as shown in Fig. 8 with examples like a clock's hand rotating on an unexpected axis or a globe's surface spinning independently. Overall, SPLART represents a scalable, robust, and practical solution that pushes the boundaries of articulated object reconstruction, laying a strong foundation for future work. + +# References + +[1] Ben Abbatematteo, Stefanie TELlex, and George Konidaris. Learning to generalize kinematic models to novel objects. In Proceedings of the Conference on Robot Learning (CoRL), pages 1289-1299, 2019. 1 +[2] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. ShapeNet: An information-rich 3D model repository. arXiv preprint arXiv:1512.03012, 2015. 5 +[3] Christopher B. Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3D-R2N2: A unified approach for single and multi-view 3D object reconstruction. In Proceedings of the European Conference on Computer Vision (ECCV), pages 628-644, 2016. 2 +[4] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), pages 303-312, 1996. 7 +[5] Andrea F. Daniele, Thomas M. Howard, and Matthew R. Walter. A multiview approach to learning articulated motion models. In Proceedings of the International Symposium of Robotics Research (ISRR), pages 371-386, 2017. 1 +[6] Jianning Deng, Kartic Subr, and Hakan Bilen. Articulate your NeRF: Unsupervised articulated object modeling via conditional view synthesis. In Advances in Neural Information Processing Systems (NeurIPS), pages 119717-119741, 2025. 1, 2, 3, 7, 8 +[7] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 224-236, 2018. 5 +[8] Haoqiang Fan, Hao Su, and Leonidas J. Guibas. A point set generation network for 3D object reconstruction from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2463-2471, 2017. 2 +[9] Samir Yitzhak Gadre, Kiana Ehsani, and Shuran Song. Act the part: Learning interaction strategies for articulated object part discovery. In Proceedings of the International Conference on Computer Vision (ICCV), pages 15732-15741, 2021. 2 +[10] Haoran Geng, Ziming Li, Yiran Geng, Jiayi Chen, Hao Dong, and He Wang. PartManip: Learning cross-category generalizable part manipulation policy from point cloud observations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2978-2988, 2023. 2 +[11] Haoran Geng, Helin Xu, Chengyang Zhao, Chao Xu, Li Yi, Siyuan Huang, and He Wang. GAPartNet: Cross-category domain-generalizable object perception and manipulation via generalizable and actionable parts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7081-7091, 2023. 2 + +[12] Antoine Guédon and Vincent Lepetit. SuGaR: Surface-aligned Gaussian splattering for efficient 3D mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5354-5363, 2024. 2 +[13] Nick Heppert, Muhammad Zubair Irshad, Sergey Zakharov, Katherine Liu, Rares Andrei Ambrus, Jeannette Bohg, Abhinav Valada, and Thomas Kollar. CARTO: Category and joint agnostic reconstruction of articulated objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21201-21210, 2023. 1, 2 +[14] Cheng-Chun Hsu, Zhenyu Jiang, and Yuke Zhu. Ditto in the house: Building articulation models of indoor scenes through interactive perception. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 3933-3939, 2023. 2 +[15] Ruizhen Hu, Wenchao Li, Oliver Van Kaick, Ariel Shamir, Hao Zhang, and Hui Huang. Learning to predict part mobility from a single static snapshot. ACM Transactions On Graphics, 36(6):1-13, 2017. 2 +[16] Xiaoxia Huang, Ian Walker, and Stan Birchfield. Occlusion-aware reconstruction and manipulation of 3D articulated objects. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 1365-1371, 2012. 1 +[17] Muhammad Zubair Irshad, Thomas Kollar, Michael Laskey, Kevin Stone, and Zsolt Kira. CenterSnap: Single-shot multi-object 3D shape reconstruction and categorical 6D pose and size estimation. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 10632-10640, 2022. 2 +[18] Muhammad Zubair Irshad, Sergey Zakharov, Rares Ambrus, Thomas Kollar, Zsolt Kira, and Adrien Gaidon. ShAPO: Implicit representations for multi-object shape, appearance, and pose optimization. In Proceedings of the European Conference on Computer Vision (ECCV), pages 275–292, 2022. 2 +[19] Muhammad Zubair Irshad, Mauro Comi, Yen-Chen Lin, Nick Heppert, Abhinav Valada, Rares Ambrus, Zsolt Kira, and Jonathan Tremblay. Neural fields in robotics: A survey, 2024. 2 +[20] Ajinkya Jain, Rudolf Lioutikov, Caleb Chuck, and Scott Niekum. ScrewNet: Category-independent articulation model estimation from depth images using screw theory. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 13670-13677, 2021. 1 +[21] Anil K. Jain. Fundamentals of Digital Image Processing. Prentice-Hall, Inc., 1989. 2 +[22] Hanxiao Jiang, Yongsen Mao, Manolis Savva, and Angel X. Chang. OPD: Single-view 3D openable part detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 410-426, 2022. 2 +[23] Zhenyu Jiang, Cheng-Chun Hsu, and Yuke Zhu. Ditto: Building digital twins of articulated objects from interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5606-5616, 2022. 1, 2 + +[24] Dov Katz, Moslem Kazemi, J. Andrew Bagnell, and Anthony Stentz. Interactive segmentation, tracking, and kinematic modeling of unknown 3D articulated objects. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 5003-5010, 2013. 1 +[25] Yuki Kawana, Yusuke Mukuta, and Tatsuya Harada. Unsupervised pose-aware part decomposition for man-made articulated objects. In Proceedings of the European Conference on Computer Vision (ECCV), pages 558-575, 2022. 2 +[26] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4):1-14, 2023. 1, 2 +[27] Justin Kerr, Chung Min Kim, Mingxuan Wu, Brent Yi, Qianqian Wang, Ken Goldberg, and Angjoo Kanazawa. Robot see robot do: Imitating articulated object manipulation with monocular 4d reconstruction. In Proceedings of the Conference on Robot Learning (CoRL), pages 587-603, 2024. 2 +[28] Long Le, Jason Xie, William Liang, Hung-Ju Wang, Yue Yang, Yecheng Jason Ma, Kyle Vedder, Arjun Krishna, Dinesh Jayaraman, and Eric Eaton. Articulate-anything: Automatic modeling of articulated objects via a vision-language foundation model. In Proceedings of the International Conference on Learning Representations (ICLR), 2025. 2 +[29] Xiaolong Li, He Wang, Li Yi, Leonidas J. Guibas, A. Lynn Abbott, and Shuran Song. Category-level articulated object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3703-3712, 2020. 1, 2 +[30] Jiayi Liu, Ali Mahdavi-Amiri, and Manolis Savva. PARIS: Part-level reconstruction and motion analysis for articulated objects. In Proceedings of the International Conference on Computer Vision (ICCV), pages 352-363, 2023. 1, 2, 3, 5, 7 +[31] Liu Liu, Wenqiang Xu, Haoyuan Fu, Sucheng Qian, Qiao-jun Yu, Yang Han, and Cewu Lu. A real-world articulated object knowledge base. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14789-14798, 2022. 2 +[32] Shaowei Liu, Saurabh Gupta, and Shenlong Wang. Building rearticulable models for arbitrary 3D objects from 4D point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21138-21147, 2023. 2 +[33] Yu Liu, Baoxiong Jia, Ruijie Lu, Junfeng Ni, Song-Chun Zhu, and Siyuan Huang. Building interactable replicas of complex articulated objects via gaussian splatting. In Proceedings of the International Conference on Learning Representations (ICLR), 2025. 1, 2, 8 +[34] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3D surface construction algorithm. In *Seminal Graphics: Pioneering Efforts That Shaped the Field*, pages 347–353, 1998. 7 +[35] Zhao Mandi, Yijia Weng, Dominik Bauer, and Shuran Song. Real2code: Reconstruct articulated objects via code generation. In Proceedings of the International Conference on Learning Representations (ICLR), 2025. 2 +[36] Eric Marchand, Hideaki Uchiyama, and Fabien Spindler. Pose estimation for augmented reality: A hands-on survey. + +IEEE Transactions on Visualization and Computer Graphics, 22(12):2633-2651, 2016. 1 +[37] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3D reconstruction in function space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4460-4470, 2019. 2 +[38] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 1 +[39] Kaichun Mo, Shilin Zhu, Angel X. Chang, Li Yi, Subarna Tripathi, Leonidas J. Guibas, and Hao Su. PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 909-918, 2019. 5 +[40] Jiteng Mu, Weichao Qiu, Adam Kortylewski, Alan Yuille, Nuno Vasconcelos, and Xiaolong Wang. A-SDF: Learning disentangled signed distance functions for articulated shape representation. In Proceedings of the International Conference on Computer Vision (ICCV), pages 12981-12991, 2021. 1, 2 +[41] Richard A. Newcombe, Andrew Fitzgibbon, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J. Davison, Pushmeet Kohi, Jamie Shotton, and Steve Hodges. KinectFusion: Real-time dense surface mapping and tracking. In Proceedings of the IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 127-136, 2011. 7 +[42] Neil Nie, Samir Yitzhak Gadre, Kiana Ehsani, and Shuran Song. Structure from action: Learning interactions for articulated object 3D structure discovery, 2022. 2 +[43] Michael Oechsle, Songyou Peng, and Andreas Geiger. UNISURF: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the International Conference on Computer Vision (ICCV), pages 5589-5599, 2021. 2 +[44] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 165-174, 2019. 2 +[45] Sudeep Pillai, Matthew R. Walter, and Seth Teller. Learning articulated motions from visual demonstrations. In Proceedings of Robotics: Science and Systems (RSS), 2014. 1 +[46] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross B. Girshick, Piotr Dólar, and Christoph Feichtenhofer. SAM 2: Segment anything in images and videos. In Proceedings of the International Conference on Learning Representations (ICLR), 2025. 5 +[47] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature + +matching with graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4937-4946, 2020. 5 +[48] Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4104-4113, 2016. 5 +[49] Chaoyue Song, Jiacheng Wei, Chuan Sheng Foo, Guosheng Lin, and Fayao Liu. REACTO: Reconstructing articulated objects from a single video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5384-5395, 2024. 2 +[50] Jürgen Sturm, Cyril Stachniss, and Wolfram Burgard. A probabilistic framework for learning kinematic models of articulated objects. Journal of Artificial Intelligence Research, 41:477-526, 2011. 1 +[51] Ryo Suzuki, Adnan Karim, Tian Xia, Hooman Hedayati, and Nicolai Marquardt. Augmented reality and robotics: A survey and taxonomy for ar-enhanced human-robot interaction and robotic interfaces. In Proceedings of the Chi Conference on Human Factors in Computing Systems, pages 1–33, 2022. 1 +[52] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3D shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11353-11362, 2021. 2 +[53] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Justin Kerr, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, David McAllister, and Angjoo Kanazawa. Nerfstudio: A modular framework for neural radiance field development. In Proceedings of the International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), pages 1-12, 2023. 1 +[54] Wei-Cheng Tseng, Hung-Ju Liao, Lin Yen-Chen, and Min Sun. CLA-NeRF: Category-level articulated neural radiance field. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 8454–8460, 2022. 1 +[55] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In Advances in Neural Information Processing Systems (NeurIPS), pages 27171-27183, 2021. 2 +[56] Xiaogang Wang, Bin Zhou, Yahao Shi, Xiaowu Chen, Qinping Zhao, and Kai Xu. Shape2Motion: Joint analysis of motion parts and attributes from 3D shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8868-8876, 2019. 1, 2 + +[57] Xian Wang, Luyao Shen, and Lik-Hang Lee. A systematic review of XR-based remote human-robot interaction systems. ACM Computing Surveys, 57(11):1-37, 2025. 1 +[58] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 2 +[59] Fangyin Wei, Rohan Chabra, Lingni Ma, Christoph Lassner, Michael Zollhöfer, Szymon Rusinkiewicz, Chris Sweeney, Richard Newcombe, and Mira Slavcheva. Self-supervised neural articulated shape and appearance models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15795-15805, 2022. 1, 2 +[60] Yijia Weng, Bowen Wen, Jonathan Tremblay, Valts Blukis, Dieter Fox, Leonidas Guibas, and Stan Birchfield. Neural implicit representation for building digital twins of unknown articulated objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3141-3150, 2024. 1, 2, 7, 8 +[61] Fanbo Xiang, Yuzhe Qin, Kaichun Mo, Yikuan Xia, Hao Zhu, Fangchen Liu, Minghua Liu, Hanxiao Jiang, Yifu Yuan, He Wang, Li Yi, Angel X. Chang, Leonidas J. Guibas, and Hao Su. SAPIEN: A simulated part-based interactive environment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11094–11104, 2020. 5 +[62] Zihao Yan, Ruizhen Hu, Xingguang Yan, Luanmin Chen, Oliver van Kaick, Hao Zhang, and Hui Huang. RPM-net: Recurrent prediction of motion and parts from point cloud. ACM Transactions On Graphics, 38(6):240:1-240:15, 2019. 2 +[63] Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jeffrey Hu, Matthew Tancik, and Angjoo Kanazawa. Gsplat: An open-source library for Gaussian splattering, 2024. 1 +[64] Li Yi, Haibin Huang, Difan Liu, Evangelos Kalogerakis, Hao Su, and Leonidas Guibas. Deep part induction from articulated object pairs. ACM Transactions on Graphics, 37(6): 1-15, 2018. 2 +[65] Sergey Zakharov, Katherine Liu, Adrien Gaidon, and Rares Ambrus. ReFiNe: Recursive field networks for cross-modal multi-scene representation. In Proceedings of the International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), pages 1-9, 2024. 2 +[66] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 2 \ No newline at end of file diff --git a/2025/SplArt_ Articulation Estimation and Part-Level Reconstruction with 3D Gaussian Splatting/images.zip b/2025/SplArt_ Articulation Estimation and Part-Level Reconstruction with 3D Gaussian Splatting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f77b7301f828f87c763db2c4eec50844444368be --- /dev/null +++ b/2025/SplArt_ Articulation Estimation and Part-Level Reconstruction with 3D Gaussian Splatting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:47265cce66199aabec39fb17bd399e73dbd9214f890761a0f52f29359e5ebafa +size 618461 diff --git a/2025/SplArt_ Articulation Estimation and Part-Level Reconstruction with 3D Gaussian Splatting/layout.json b/2025/SplArt_ Articulation Estimation and Part-Level Reconstruction with 3D Gaussian Splatting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fccffb1738152382dcb1fc4511c1d03900da7aaa --- /dev/null +++ b/2025/SplArt_ Articulation Estimation and Part-Level Reconstruction with 3D Gaussian Splatting/layout.json @@ -0,0 +1,10624 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 104, + 103, + 507, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 103, + 507, + 140 + ], + "spans": [ + { + "bbox": [ + 104, + 103, + 507, + 140 + ], + "type": "text", + "content": "SPLART: Articulation Estimation and Part-Level Reconstruction with 3D Gaussian Splatting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 160, + 527, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 160, + 527, + 190 + ], + "spans": [ + { + "bbox": [ + 102, + 160, + 527, + 190 + ], + "type": "text", + "content": "Shengjie Lin†, Jiading Fang†, Muhammad Zubair Irshad‡, Vitor Campagnolo Guizilini‡, Rares Andrei Ambrus‡, Greg Shakhnarovich†‡, Matthew R. Walter†" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 141, + 189, + 488, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 189, + 488, + 204 + ], + "spans": [ + { + "bbox": [ + 141, + 189, + 488, + 204 + ], + "type": "text", + "content": "† Toyota Technological Institute at Chicago, ‡ Toyota Research Institute" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 205, + 567, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 205, + 567, + 218 + ], + "spans": [ + { + "bbox": [ + 61, + 205, + 567, + 218 + ], + "type": "text", + "content": "{slin, fjd, greg, mwalter}@ttic.edu, {zubair.irshad, vitor.quizilini, rares.ambrus}@tri.global" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 245, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 245, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 245, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 270, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 296, + 571 + ], + "type": "text", + "content": "Reconstructing articulated objects prevalent in daily environments is crucial for applications in augmented/virtual reality and robotics. However, existing methods face scalability limitations (requiring 3D supervision or costly annotations), robustness issues (being susceptible to local optima), and rendering shortcomings (lacking speed or photorealism). We introduce SPLART, a self-supervised, category-agnostic framework that uses 3D Gaussian Splattig (3DGS) to reconstruct and infer the kinematics of articulated objects from two sets of posed RGB images captured at different articulation states, enabling real-time photorealistic rendering for novel viewpoints and articulations. SPLART augments 3DGS with a differentiable mobility parameter per Gaussian, achieving refined part segmentation. A multi-stage optimization strategy is employed to progressively handle reconstruction, part segmentation, and articulation estimation, significantly enhancing robustness and accuracy. SPLART exploits geometric self-supervision, effectively addressing challenging scenarios without requiring 3D annotations or category-specific priors. Evaluations on established and newly proposed benchmarks, along with applications to real-world scenarios using a handheld RGB camera, demonstrate SPLART's state-of-the-art performance and real-world practicality. Code is publicly available at https://github.com/ripl/splart." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 586, + 135, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 135, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 135, + 597 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": "Articulated objects, such as drawers, doors, and scissors, are ubiquitous in our daily lives, yet their dynamic nature poses significant challenges for 3D reconstruction—a critical task for applications in augmented/virtual reality [36, 51], robotics [1, 5, 16, 24, 50, 57], and computer vision [20, 45]. Existing methods for reconstructing articulated objects are hindered by several key limitations: they often require labor-intensive supervision (e.g., part-level segmentation or articulation annotations) [23, 29," + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 243, + 426, + 331 + ], + "blocks": [ + { + "bbox": [ + 317, + 243, + 426, + 331 + ], + "lines": [ + { + "bbox": [ + 317, + 243, + 426, + 331 + ], + "spans": [ + { + "bbox": [ + 317, + 243, + 426, + 331 + ], + "type": "image", + "image_path": "8c14f10e462f496c2e116707619e37f0dff2c7fbbf757212e58adff3a0acdffe.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 432, + 244, + 550, + 331 + ], + "blocks": [ + { + "bbox": [ + 432, + 244, + 550, + 331 + ], + "lines": [ + { + "bbox": [ + 432, + 244, + 550, + 331 + ], + "spans": [ + { + "bbox": [ + 432, + 244, + 550, + 331 + ], + "type": "image", + "image_path": "739155a868204e504885cf6d17dc0b8e5a8bd4aaa5adfd8248302b293390ad17.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 315, + 335, + 371, + 405 + ], + "blocks": [ + { + "bbox": [ + 315, + 335, + 371, + 405 + ], + "lines": [ + { + "bbox": [ + 315, + 335, + 371, + 405 + ], + "spans": [ + { + "bbox": [ + 315, + 335, + 371, + 405 + ], + "type": "image", + "image_path": "4e480fa9bf958ffd84a582c3e8ba48fd1f638390542b6b1e11a139f0c8135f82.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 373, + 346, + 467, + 403 + ], + "blocks": [ + { + "bbox": [ + 373, + 346, + 467, + 403 + ], + "lines": [ + { + "bbox": [ + 373, + 346, + 467, + 403 + ], + "spans": [ + { + "bbox": [ + 373, + 346, + 467, + 403 + ], + "type": "image", + "image_path": "edfef345eb7328e4432c033b9531328821d4ce0cfe56dd7ff2df97a62174cb15.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 477, + 365, + 539, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 477, + 365, + 539, + 385 + ], + "spans": [ + { + "bbox": [ + 477, + 365, + 539, + 385 + ], + "type": "text", + "content": "SPLART" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 317, + 411, + 371, + 475 + ], + "blocks": [ + { + "bbox": [ + 317, + 411, + 371, + 475 + ], + "lines": [ + { + "bbox": [ + 317, + 411, + 371, + 475 + ], + "spans": [ + { + "bbox": [ + 317, + 411, + 371, + 475 + ], + "type": "image", + "image_path": "5899ad4519229e08ade1d7230a0f3479cfcc6013fc039e1cd7190c5aaec7a93d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 477, + 365, + 485 + ], + "lines": [ + { + "bbox": [ + 325, + 477, + 365, + 485 + ], + "spans": [ + { + "bbox": [ + 325, + 477, + 365, + 485 + ], + "type": "text", + "content": "Articulation" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 374, + 426, + 429, + 475 + ], + "blocks": [ + { + "bbox": [ + 374, + 426, + 429, + 475 + ], + "lines": [ + { + "bbox": [ + 374, + 426, + 429, + 475 + ], + "spans": [ + { + "bbox": [ + 374, + 426, + 429, + 475 + ], + "type": "image", + "image_path": "a469f80e92abc52db462b98744b78fbfd58d798534210babb4fdfc7f01304873.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 495, + 555, + 562 + ], + "lines": [ + { + "bbox": [ + 313, + 495, + 555, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 555, + 562 + ], + "type": "text", + "content": "Figure 1. Given (top) RGB images of an object at two articulation states, (middle) SPLART uses 3DGS to simultaneously reconstruct its static and dynamic parts and estimate the kinematic articulation model. SPLART is then able to (bottom) render high-fidelity 3D reconstructions of the object along with part-level segmentations for novel articulation states, allowing for novel view synthesis." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 430, + 426, + 486, + 475 + ], + "blocks": [ + { + "bbox": [ + 430, + 426, + 486, + 475 + ], + "lines": [ + { + "bbox": [ + 430, + 426, + 486, + 475 + ], + "spans": [ + { + "bbox": [ + 430, + 426, + 486, + 475 + ], + "type": "image", + "image_path": "f36a1e71cdc3e76fc452d65ea457f42324dafc504f926d76c06b5eed854bf1ea.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 487, + 427, + 544, + 474 + ], + "blocks": [ + { + "bbox": [ + 487, + 427, + 544, + 474 + ], + "lines": [ + { + "bbox": [ + 487, + 427, + 544, + 474 + ], + "spans": [ + { + "bbox": [ + 487, + 427, + 544, + 474 + ], + "type": "image", + "image_path": "8ffa9581e67e0231c89d4649935deddf7c23c714f3b806973cd9ca3d7fcbe7aa.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "40, 54, 56], they depend on 3D supervision that restricts practical use [23, 29, 33, 40, 60], they produce category-specific models that limit scalability [29, 40, 54, 59], or they are not capable of real-time, photorealistic rendering [6, 13, 23, 29, 30, 40, 54, 56, 59, 60]. To address these challenges, we introduce SPLART, a novel self-supervised and category-agnostic framework that leverages 3D Gaussian Splitting (3DGS) [26] to reconstruct articulated objects from minimal input—two sets of posed RGB images at distinct articulation states. SPLART reconstructs the object's parts and infers its kinematics, enabling real-time, photorealistic rendering for novel views and articulation states." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8841" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "Central to SPLART is the augmentation of 3DGS [26] to include a differentiable mobility parameter for each Gaussian, which enables a more refined segmentation of static and mobile parts through gradient-based optimization. This results in enhanced reconstruction quality, while preserving the real-time, photorealistic rendering capabilities of 3DGS—offering a speedup of more than " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "100\\times" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " over methods [6, 30] based on neural radiance fields [38]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 168, + 294, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 168, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 294, + 300 + ], + "type": "text", + "content": "To enhance robustness, SPLART employs a multi-stage optimization strategy that decouples the part-level reconstruction and articulation estimation processes. Unlike end-to-end approaches prone to local optima [30], SPLART first independently reconstructs each articulation state, then estimates each Gaussian's mobility parameter for part segmentation, and finally refines both the articulation and mobility estimates jointly. This structured approach ensures stable and accurate convergence, avoiding the stringent initialization requirements of existing methods, thereby providing a practical solution for challenging articulated structures." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 301, + 294, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 294, + 408 + ], + "type": "text", + "content": "Building on this foundation, SPLART leverages geometric self-supervision to eliminate the need for manual annotations or 3D supervision. By enforcing geometric consistency between reconstructions through complementary loss formulations, SPLART robustly estimates articulation parameters across diverse scenarios. This self-supervised strategy enhances scalability, enabling SPLART to reconstruct a wide range of articulated objects without relying on prior structural or categorical knowledge." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 410, + 294, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 294, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 294, + 493 + ], + "type": "text", + "content": "Extensive evaluations on both established and newly introduced benchmarks demonstrate SPLART's superior articulation accuracy and reconstruction quality, surpassing state-of-the-art methods without requiring 3D supervision. Real-world experiments further validate its practicality, showcasing successful reconstructions of diverse articulated objects using only a handheld RGB camera." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 494, + 209, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 494, + 209, + 505 + ], + "spans": [ + { + "bbox": [ + 67, + 494, + 209, + 505 + ], + "type": "text", + "content": "In summary, this work contributes:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 506, + 294, + 637 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 56, + 506, + 294, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 506, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 506, + 294, + 542 + ], + "type": "text", + "content": "1. An extension of 3DGS with a differentiable mobility value per Gaussian that enables precise part segmentation using gradient-based optimization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 543, + 294, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 294, + 577 + ], + "type": "text", + "content": "2. A multi-stage optimization strategy that decouples reconstruction and articulation estimation, enhancing robustness and accuracy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 578, + 294, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 294, + 613 + ], + "type": "text", + "content": "3. Complementary formulations of geometric self-supervision for articulation estimation, eliminating the need for 3D supervision or laborious annotations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 614, + 294, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 614, + 294, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 614, + 294, + 637 + ], + "type": "text", + "content": "4. A challenging dataset and new metrics for comprehensive evaluation of articulated object reconstruction." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 651, + 142, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 142, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 142, + 662 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 671, + 242, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 242, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 242, + 684 + ], + "type": "text", + "content": "2.1. Data-Driven Articulation Learning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "type": "text", + "content": "Estimating the pose and joint properties of articulated objects is crucial for robot manipulation and interac" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "content": "tion [10, 11, 19, 31]. Recent learning-based methods [9, 13, 15, 22, 29, 32, 56, 64] infer articulation properties from point clouds via end-to-end training. For instance, Shape2Motion [56] analyzes motion parts from a single point cloud in a supervised setting, while ANCSH [29] performs category-level pose estimation, but requires class-specific models. RPM-Net [62] enhances generalization across categories for part segmentation and kinematic prediction, and DITTO [23] predicts motion and geometry from 3D point cloud pairs without labels. However, these methods depend on costly 3D supervision and annotations. In contrast, our approach reconstructs accurate 3D geometry and detailed appearance, capturing articulation without 3D supervision or priors." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 248, + 537, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 248, + 537, + 261 + ], + "spans": [ + { + "bbox": [ + 313, + 248, + 537, + 261 + ], + "type": "text", + "content": "2.2. Representations for Object Reconstruction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 266, + 553, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 553, + 397 + ], + "type": "text", + "content": "Early 3D object reconstruction methods predicted point clouds, voxels, or meshes from partial observations [3, 8, 17]. Recent advances in implicit scene representations [18, 26, 37, 38, 44, 65] enable detailed geometry and appearance reconstruction via differentiable rendering [12, 43, 52, 55]. While neural fields suffer from slow rendering, 3D Gaussian Splatting (3DGS) [26] overcomes this by using explicit 3D Gaussians. We leverage 3DGS for self-supervised articulated object reconstruction from posed RGB images, achieving fast, realistic synthesis of novel views and articulations in real time." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 406, + 499, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 499, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 499, + 418 + ], + "type": "text", + "content": "2.3. Articulated Object Reconstruction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 424, + 553, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 553, + 628 + ], + "type": "text", + "content": "Recent methods leverage differentiable 3D representations [26, 38, 44] to jointly reconstruct articulated objects and infer articulation parameters. Training-based approaches use synthetic 3D data to predict joint parameters and segment parts [13, 14, 23, 25, 40, 42, 59]. Self-supervised methods optimize shape, appearance, and articulation per scene without extensive training [6, 27, 30, 32, 33, 49, 60]. Some handle objects with more than two parts, but require the number of parts to be known and the mobile parts to be directly connected to the static part [6, 33, 60]. Other works enhance articulation estimation using large language or vision-language models [28, 35]. In contrast, our self-supervised method reconstructs two-part articulated objects from RGB images across articulation states using 3DGS [26]. As the first to apply 3DGS to this task without 3D supervision or pre-trained priors, it robustly handles challenging cases and achieves real-time performance." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 639, + 397, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 397, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 397, + 653 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 659, + 382, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 382, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 382, + 671 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Consider an arbitrary object composed of two rigid parts: a static parent part, and a child part that can move relative to its parent through either a revolute or prismatic articulation." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8842" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 102, + 111, + 153 + ], + "blocks": [ + { + "bbox": [ + 74, + 73, + 147, + 91 + ], + "lines": [ + { + "bbox": [ + 74, + 73, + 147, + 91 + ], + "spans": [ + { + "bbox": [ + 74, + 73, + 147, + 91 + ], + "type": "text", + "content": "(a) Stage 1: Per-state reconstruction." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 102, + 111, + 153 + ], + "lines": [ + { + "bbox": [ + 56, + 102, + 111, + 153 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 111, + 153 + ], + "type": "image", + "image_path": "949f48fd63c7627d87b3ddba7965c91d5267f78a243ccf93e30db8e0b339e1e6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 116, + 102, + 165, + 153 + ], + "blocks": [ + { + "bbox": [ + 86, + 92, + 135, + 101 + ], + "lines": [ + { + "bbox": [ + 86, + 92, + 135, + 101 + ], + "spans": [ + { + "bbox": [ + 86, + 92, + 135, + 101 + ], + "type": "text", + "content": "(Section 3.2.1)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 116, + 102, + 165, + 153 + ], + "lines": [ + { + "bbox": [ + 116, + 102, + 165, + 153 + ], + "spans": [ + { + "bbox": [ + 116, + 102, + 165, + 153 + ], + "type": "image", + "image_path": "65d52b5e2530a042a40745b877f08db5d0627f3ad8648ed93ec38e1365793423.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 155, + 106, + 215 + ], + "blocks": [ + { + "bbox": [ + 58, + 155, + 106, + 215 + ], + "lines": [ + { + "bbox": [ + 58, + 155, + 106, + 215 + ], + "spans": [ + { + "bbox": [ + 58, + 155, + 106, + 215 + ], + "type": "image", + "image_path": "a682f366330860f5adc56c9e1de27e762e3ab7efe3a607caeaba740e2653b232.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 226, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 55, + 226, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 555, + 281 + ], + "type": "text", + "content": "Figure 2. Methodology overview of SPLART. It consists of three decoupled stages for the purpose of stable optimization. Stage 1 constructs a 3DGS for each state from posed RGB images using the photometric loss. Stage 2 proposes cross-static formulation for mobility estimation. Intuitively, it combines static parts of both states and the mobile part of the desired state for the target Gaussians. Stage 3 proposes cross-mobile formulation for articulation estimation. Intuitively, it combines static parts of both states, the mobile part of the desired state and the transformed mobile part of the other state together for the target Gaussians." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 115, + 155, + 162, + 215 + ], + "blocks": [ + { + "bbox": [ + 115, + 155, + 162, + 215 + ], + "lines": [ + { + "bbox": [ + 115, + 155, + 162, + 215 + ], + "spans": [ + { + "bbox": [ + 115, + 155, + 162, + 215 + ], + "type": "image", + "image_path": "89dd5b274ef6c9ec9cb37e73fac5119891da0a47e96c665fd4043b933b2cf2a3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 170, + 93, + 351, + 215 + ], + "blocks": [ + { + "bbox": [ + 198, + 73, + 323, + 92 + ], + "lines": [ + { + "bbox": [ + 198, + 73, + 323, + 92 + ], + "spans": [ + { + "bbox": [ + 198, + 73, + 323, + 92 + ], + "type": "text", + "content": "(b) Stage 2: Cross-static formulation for mobility estimation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 170, + 93, + 351, + 215 + ], + "lines": [ + { + "bbox": [ + 170, + 93, + 351, + 215 + ], + "spans": [ + { + "bbox": [ + 170, + 93, + 351, + 215 + ], + "type": "image", + "image_path": "56152ec1f68d402376abfd4ece5c2250a479c0577d0df1d3c58e428f593016d6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 356, + 93, + 552, + 215 + ], + "blocks": [ + { + "bbox": [ + 370, + 73, + 536, + 92 + ], + "lines": [ + { + "bbox": [ + 370, + 73, + 536, + 92 + ], + "spans": [ + { + "bbox": [ + 370, + 73, + 536, + 92 + ], + "type": "text", + "content": "(c) Stage 3: Cross-mobile formulation for articulation estimation and mobility refinement." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 356, + 93, + 552, + 215 + ], + "lines": [ + { + "bbox": [ + 356, + 93, + 552, + 215 + ], + "spans": [ + { + "bbox": [ + 356, + 93, + 552, + 215 + ], + "type": "image", + "image_path": "f9ba36a2f893efaaab5e87cb625da3b3743ac339695c80e7b75b3b6cd0c08780.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 289, + 295, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 289, + 295, + 372 + ], + "spans": [ + { + "bbox": [ + 54, + 289, + 295, + 372 + ], + "type": "text", + "content": "Our objective is twofold: (1) to reconstruct the articulated object at the part level; and (2) to estimate its articulated motion. Assuming a known articulation type (i.e., either revolute or prismatic), the input to our method consists of two sets of posed RGB images (i.e., images with known camera intrinsics and extrinsics), each capturing the articulated object at one end state of the motion." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "text", + "content": "Formally, let " + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "text", + "content": " denote the articulation state label, where " + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "inline_equation", + "content": "l = 0" + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "inline_equation", + "content": "l = 1" + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "text", + "content": " correspond to the two end states of the observed articulation. For reconstruction, SPLART uses observations " + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_l = \\{(I_l^i,P_l^i,K_l^i)\\}_{i = 1}^{N_l}, l\\in \\{0,1\\}" + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "inline_equation", + "content": "I_{l}^{i}" + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "text", + "content": "-th observed RGB image of the articulated object at state " + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "inline_equation", + "content": "P_{l}^{i}" + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "inline_equation", + "content": "K_{l}^{i}" + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "text", + "content": " denote its camera extrinsic and intrinsic, respectively, and " + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "inline_equation", + "content": "N_{l}" + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "text", + "content": " represents the number of data samples for state " + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "inline_equation", + "content": "P_{l}^{i}" + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "text", + "content": " is specified in a common world frame for both states, while the articulated motion involves only one moving part w.r.t. the world space. SPLART models a revolute articulation with its rotation axis " + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{a}\\| = 1" + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "text", + "content": "), pivot " + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "text", + "content": ", and rotation angle " + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "text", + "content": ", such that a point " + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "text", + "content": " (in the world frame) on the mobile part at state " + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "inline_equation", + "content": "l = 0" + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "text", + "content": " will move to" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 122, + 536, + 295, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 536, + 295, + 550 + ], + "spans": [ + { + "bbox": [ + 122, + 536, + 295, + 550 + ], + "type": "interline_equation", + "content": "{ } ^ { 1 } T _ { 0 } \\mathbf { x } = R _ { \\mathbf { a } , \\theta } ( \\mathbf { x } - \\mathbf { p } ) + \\mathbf { p } \\tag {1}", + "image_path": "90f204bee88e653f150bacf3d6aaeb0bd7f696cd1660219c7183b259762138a4.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 557, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 640 + ], + "type": "text", + "content": "at state " + }, + { + "bbox": [ + 55, + 557, + 295, + 640 + ], + "type": "inline_equation", + "content": "l = 1" + }, + { + "bbox": [ + 55, + 557, + 295, + 640 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 557, + 295, + 640 + ], + "type": "inline_equation", + "content": "R_{\\mathbf{a},\\theta}" + }, + { + "bbox": [ + 55, + 557, + 295, + 640 + ], + "type": "text", + "content": " is the rotation induced by the axis-angle notation. A prismatic articulation is instead modeled by its translation axis " + }, + { + "bbox": [ + 55, + 557, + 295, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 55, + 557, + 295, + 640 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 557, + 295, + 640 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{a} \\| = 1" + }, + { + "bbox": [ + 55, + 557, + 295, + 640 + ], + "type": "text", + "content": ") and distance " + }, + { + "bbox": [ + 55, + 557, + 295, + 640 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 557, + 295, + 640 + ], + "type": "text", + "content": ". The goal is to reconstruct the articulated object at the part level using a chosen representation while estimating the articulated motion " + }, + { + "bbox": [ + 55, + 557, + 295, + 640 + ], + "type": "inline_equation", + "content": "^1 T_0" + }, + { + "bbox": [ + 55, + 557, + 295, + 640 + ], + "type": "text", + "content": ", ensuring that the renderings at each articulation state are with the observations." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": "SPLART extends the 3DGS representation for articulated objects, decoupling the goals of part-level reconstruction and articulation estimation across three stages: (1) separate reconstructions for each articulation state (Sec. 3.2.1), (2) mobility estimation using the cross-static formulation (Sec. 3.2.2), and (3) articulation estimation and mobility re" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 289, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 555, + 373 + ], + "type": "text", + "content": "finement using the cross-mobile formulation (Sec. 3.2.3). To facilitate the application of SPLART to real-world objects, we leverage modern structure-from-motion and image segmentation techniques, developing a framework that enables general users to reconstruct articulated objects in their surroundings using only images captured by a handheld camera device (Sec. 3.3)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 380, + 509, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 509, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 509, + 393 + ], + "type": "text", + "content": "3.2. Decoupled Multi-Stage Optimization" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 398, + 554, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 554, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 554, + 447 + ], + "type": "text", + "content": "Jointly performing part-level reconstruction and articulation estimation on an articulated object is prone to local minima [6, 30], making it preferable to strategically decouple the problem into multiple stages." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 453, + 485, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 485, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 485, + 464 + ], + "type": "text", + "content": "3.2.1. Separate per-state reconstruction" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 467, + 556, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 556, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 556, + 575 + ], + "type": "text", + "content": "In Stage 1, SPLART separately optimizes two 3DGS models following the standard procedure, one for each end state of the articulation. In addition to the attributes that are standard to 3DGS, each Gaussian is initialized with a persistent binary state label " + }, + { + "bbox": [ + 313, + 467, + 556, + 575 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 467, + 556, + 575 + ], + "type": "text", + "content": ", equally drawn from " + }, + { + "bbox": [ + 313, + 467, + 556, + 575 + ], + "type": "inline_equation", + "content": "\\{0,1\\}" + }, + { + "bbox": [ + 313, + 467, + 556, + 575 + ], + "type": "text", + "content": ". We denote the set of Gaussians representing state " + }, + { + "bbox": [ + 313, + 467, + 556, + 575 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 467, + 556, + 575 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 467, + 556, + 575 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_l^{\\mathrm{ref}}" + }, + { + "bbox": [ + 313, + 467, + 556, + 575 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 467, + 556, + 575 + ], + "type": "inline_equation", + "content": "ref" + }, + { + "bbox": [ + 313, + 467, + 556, + 575 + ], + "type": "text", + "content": " emphasizes that " + }, + { + "bbox": [ + 313, + 467, + 556, + 575 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_l^{\\mathrm{ref}}" + }, + { + "bbox": [ + 313, + 467, + 556, + 575 + ], + "type": "text", + "content": " is the reference reconstruction unaffected by the other state. Given a data sample observed at state " + }, + { + "bbox": [ + 313, + 467, + 556, + 575 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 467, + 556, + 575 + ], + "type": "text", + "content": ", the optimization follows as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 394, + 583, + 553, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 583, + 553, + 605 + ], + "spans": [ + { + "bbox": [ + 394, + 583, + 553, + 605 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathcal {G} _ {l} ^ {\\text {r e f}}} \\Delta_ {I} \\left(\\hat {I} _ {l} ^ {\\text {r e f}, i}, I _ {l} ^ {i}\\right), \\tag {2a}", + "image_path": "f37be80b268db17f256f344beca7c182b35850162d362cafa558b775c4f315e9.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 361, + 608, + 553, + 623 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 608, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 361, + 608, + 553, + 623 + ], + "type": "interline_equation", + "content": "\\hat {I} _ {l} ^ {\\text {r e f}, i} = R \\left(\\mathcal {G} _ {l} ^ {\\text {r e f}}, P _ {l} ^ {i}, K _ {l} ^ {i}\\right), l \\in \\{0, 1 \\}, \\tag {2b}", + "image_path": "6839030f1c50a3cf07e1602fb2b2afbed0d29195fcda0ac8d67e186cc1240f38.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 632, + 554, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 632, + 554, + 669 + ], + "spans": [ + { + "bbox": [ + 313, + 632, + 554, + 669 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 632, + 554, + 669 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 632, + 554, + 669 + ], + "type": "text", + "content": " is the 3DGS rendering function, and " + }, + { + "bbox": [ + 313, + 632, + 554, + 669 + ], + "type": "inline_equation", + "content": "\\Delta_I" + }, + { + "bbox": [ + 313, + 632, + 554, + 669 + ], + "type": "text", + "content": " represents the photometric loss. For simplicity, the view index " + }, + { + "bbox": [ + 313, + 632, + 554, + 669 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 632, + 554, + 669 + ], + "type": "text", + "content": " and camera parameters " + }, + { + "bbox": [ + 313, + 632, + 554, + 669 + ], + "type": "inline_equation", + "content": "P_l^i, K_l^i" + }, + { + "bbox": [ + 313, + 632, + 554, + 669 + ], + "type": "text", + "content": " will be omitted from now on." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 674, + 545, + 686 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 674, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 674, + 545, + 686 + ], + "type": "text", + "content": "3.2.2. Cross-static formulation for mobility estimation" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "The focus of Stage 2 is mobility estimation for each Gaussian. To keep " + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{\\mathrm{ref}}" + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": " dedicated to single-state reconstruction," + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8843" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 294, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 294, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 294, + 131 + ], + "type": "text", + "content": "we first duplicate " + }, + { + "bbox": [ + 55, + 71, + 294, + 131 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_l^{\\mathrm{ref}}" + }, + { + "bbox": [ + 55, + 71, + 294, + 131 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 71, + 294, + 131 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_l^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 55, + 71, + 294, + 131 + ], + "type": "text", + "content": " for both states (i.e., " + }, + { + "bbox": [ + 55, + 71, + 294, + 131 + ], + "type": "inline_equation", + "content": "l\\in \\{0,1\\}" + }, + { + "bbox": [ + 55, + 71, + 294, + 131 + ], + "type": "text", + "content": "), which is intended as the target representation that will fulfill the goals of part-level reconstruction and articulation estimation. By design, " + }, + { + "bbox": [ + 55, + 71, + 294, + 131 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 55, + 71, + 294, + 131 + ], + "type": "text", + "content": " shares neither data storage nor gradient flow with " + }, + { + "bbox": [ + 55, + 71, + 294, + 131 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{\\mathrm{ref}}" + }, + { + "bbox": [ + 55, + 71, + 294, + 131 + ], + "type": "text", + "content": " after its creation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 295, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 251 + ], + "type": "text", + "content": "For each Gaussian in " + }, + { + "bbox": [ + 55, + 133, + 295, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 55, + 133, + 295, + 251 + ], + "type": "text", + "content": ", we further extend its set of attributes with a mobility value " + }, + { + "bbox": [ + 55, + 133, + 295, + 251 + ], + "type": "inline_equation", + "content": "m \\in [0,1]" + }, + { + "bbox": [ + 55, + 133, + 295, + 251 + ], + "type": "text", + "content": ", initialized at 0.5. By design, " + }, + { + "bbox": [ + 55, + 133, + 295, + 251 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 133, + 295, + 251 + ], + "type": "text", + "content": " enables the break-down of a Gaussian to its static and mobile components, where each component inherits all the original Gaussian attributes except for the opacity " + }, + { + "bbox": [ + 55, + 133, + 295, + 251 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 133, + 295, + 251 + ], + "type": "text", + "content": ". The static component has its opacity modified to " + }, + { + "bbox": [ + 55, + 133, + 295, + 251 + ], + "type": "inline_equation", + "content": "\\sigma \\cdot (1 - m)" + }, + { + "bbox": [ + 55, + 133, + 295, + 251 + ], + "type": "text", + "content": ", and the mobile component to " + }, + { + "bbox": [ + 55, + 133, + 295, + 251 + ], + "type": "inline_equation", + "content": "\\sigma \\cdot m" + }, + { + "bbox": [ + 55, + 133, + 295, + 251 + ], + "type": "text", + "content": ". For simplicity, letting " + }, + { + "bbox": [ + 55, + 133, + 295, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 133, + 295, + 251 + ], + "type": "text", + "content": " be the set of mobilities for " + }, + { + "bbox": [ + 55, + 133, + 295, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 133, + 295, + 251 + ], + "type": "text", + "content": ", we use the element-wise product " + }, + { + "bbox": [ + 55, + 133, + 295, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{G} \\cdot (1 - \\mathcal{M})" + }, + { + "bbox": [ + 55, + 133, + 295, + 251 + ], + "type": "text", + "content": " to denote the static component of " + }, + { + "bbox": [ + 55, + 133, + 295, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 133, + 295, + 251 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 133, + 295, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{G} \\cdot \\mathcal{M}" + }, + { + "bbox": [ + 55, + 133, + 295, + 251 + ], + "type": "text", + "content": " for the mobile component." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 251, + 294, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 294, + 312 + ], + "type": "text", + "content": "To estimate the mobilities " + }, + { + "bbox": [ + 55, + 251, + 294, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 251, + 294, + 312 + ], + "type": "text", + "content": ", we employ the intuition that the static components from both states should constitute the static part of the articulated object. Formally, we introduce the cross-static (cs) formulation, where the static part of the articulated object is jointly represented as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 318, + 294, + 333 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 318, + 294, + 333 + ], + "spans": [ + { + "bbox": [ + 85, + 318, + 294, + 333 + ], + "type": "interline_equation", + "content": "\\mathcal {G} ^ {\\mathrm {s}} = \\mathcal {G} _ {l} ^ {\\mathrm {t g t}} \\cdot (1 - \\mathcal {M} _ {l}) \\oplus \\mathcal {G} _ {1 - l} ^ {\\mathrm {t g t}} \\cdot (1 - \\mathcal {M} _ {1 - l}), \\tag {3}", + "image_path": "d73054887139cfa98520be0e7ca9a9d6aefc671690bae9501512d784c42149c0.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 339, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 295, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 339, + 295, + 361 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 55, + 339, + 295, + 361 + ], + "type": "text", + "content": " denotes concatenation. For state " + }, + { + "bbox": [ + 55, + 339, + 295, + 361 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 339, + 295, + 361 + ], + "type": "text", + "content": ", the target representation thus becomes" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 125, + 369, + 294, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 369, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 125, + 369, + 294, + 384 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {l} ^ {\\mathrm {t g t}, \\mathrm {c s}} = \\mathcal {G} ^ {\\mathrm {s}} \\oplus \\mathcal {G} _ {l} ^ {\\mathrm {t g t}} \\cdot \\mathcal {M} _ {l}. \\tag {4}", + "image_path": "a8814cae84b1cb3bda8bc56a46e9873653f25c2d4d0b75ee179c5302763a03f9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 391, + 294, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 294, + 415 + ], + "type": "text", + "content": "With this formulation, Stage 2 is further divided into two sub-stages as described below." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 417, + 296, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 417, + 296, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 296, + 489 + ], + "type": "text", + "content": "Stage 2(a): Coarse mobility estimation via cross-static geometric consistency. To measure the geometric distance between two Gaussian sets, we design a weighted version of the Chamfer distance. Specifically, letting " + }, + { + "bbox": [ + 55, + 417, + 296, + 489 + ], + "type": "inline_equation", + "content": "X = \\{(x_{i},w_{x_{i}})\\}_{i = 1}^{M}" + }, + { + "bbox": [ + 55, + 417, + 296, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 417, + 296, + 489 + ], + "type": "inline_equation", + "content": "Y = \\{(y_{j},w_{y_{j}})\\}_{j = 1}^{N}" + }, + { + "bbox": [ + 55, + 417, + 296, + 489 + ], + "type": "text", + "content": " be two sets of point-weight pairs, the weighted Chamfer distance is then:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 498, + 295, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 498, + 295, + 562 + ], + "spans": [ + { + "bbox": [ + 66, + 498, + 295, + 562 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname {Chamfer}(X,Y) = \\sum_{\\left(x_{i},w_{x_{i}}\\right)\\in X}\\tilde{w}_{x_{i}}\\min_{\\left(y_{j},w_{y_{j}}\\right)\\in Y}\\| x_{i} - y_{j}\\|^{2} \\\\ + \\sum_ {\\left(y _ {j}, w _ {y _ {j}}\\right) \\in Y} \\tilde {w} _ {y _ {j}} \\min _ {\\left(x _ {i}, w _ {x _ {i}}\\right) \\in X} \\| x _ {i} - y _ {j} \\| ^ {2}, \\tag {5} \\\\ \\end{array}", + "image_path": "00b4e67a4f96ee66220297f1d3df1b1b184f0121dfe7bcba1a3068ecbc3ba85a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 569, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 569, + 295, + 640 + ], + "type": "inline_equation", + "content": "\\tilde{w}_{x_i} = w_{x_i} / \\sum_{(x_i, w_{x_i}) \\in X} w_{x_i}" + }, + { + "bbox": [ + 55, + 569, + 295, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 569, + 295, + 640 + ], + "type": "inline_equation", + "content": "\\tilde{w}_{y_j}" + }, + { + "bbox": [ + 55, + 569, + 295, + 640 + ], + "type": "text", + "content": " are normalized weights. For each Gaussian, we use its mobility-modified opacity " + }, + { + "bbox": [ + 55, + 569, + 295, + 640 + ], + "type": "inline_equation", + "content": "\\sigma'" + }, + { + "bbox": [ + 55, + 569, + 295, + 640 + ], + "type": "text", + "content": " as the weight, essentially treating it as " + }, + { + "bbox": [ + 55, + 569, + 295, + 640 + ], + "type": "inline_equation", + "content": "\\sigma'" + }, + { + "bbox": [ + 55, + 569, + 295, + 640 + ], + "type": "text", + "content": " points overlapped at its mean position. The mobilities are then optimized by encouraging the geometric consistency formulated as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 648, + 294, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 648, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 648, + 294, + 666 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathcal {M}} \\mathrm {C D} _ {0} ^ {\\mathrm {c s}} + \\mathrm {C D} _ {1} ^ {\\mathrm {c s}} + \\lambda_ {m} ^ {\\text {g e o m}} \\| \\mathcal {M} \\|, \\tag {6a}", + "image_path": "0cbbd6cd19d879e8b3d558dc8884241633bf9830d56657c8e0df23379c7c6c15.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 112, + 667, + 294, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 667, + 294, + 683 + ], + "spans": [ + { + "bbox": [ + 112, + 667, + 294, + 683 + ], + "type": "interline_equation", + "content": "\\mathrm {C D} _ {l} ^ {\\mathrm {c s}} = \\operatorname {C h a m f e r} \\left(\\mathcal {G} _ {l} ^ {\\mathrm {t g t , c s}}, \\mathcal {G} _ {l} ^ {\\mathrm {r e f}}\\right), \\tag {6b}", + "image_path": "b56bfa8f4c6cb28a0daf8707b3e19c32b9131958b226cb2b3dce7569b287591d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{Chamfer}(\\cdot)" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " denotes the weighted Chamfer distance, and " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{m}^{\\mathrm{geom}}\\| \\mathcal{M}\\|" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " is a regularization term that encourages" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "smaller mobilities. Note how " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " affects the weighted Chamfer distance by modifying the opacities, and that " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = 1" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " is a trivial solution without the regularization. Without photometric supervision, the mobilities obtained from Eqn. 6a are generally noisy. However, being relatively fast (taking only tens of seconds), they still serve as a good initialization for the next sub-stage." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 157, + 553, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 553, + 202 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 553, + 202 + ], + "type": "text", + "content": "Stage 2(b): Joint mobility and Gaussian optimization via cross-static rendering. To more accurately estimate the mobilities while jointly refining the full Gaussian attributes, we perform cross-static rendering as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 361, + 207, + 553, + 225 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 207, + 553, + 225 + ], + "spans": [ + { + "bbox": [ + 361, + 207, + 553, + 225 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathcal {G} ^ {\\mathrm {t g t}}, \\mathcal {M} _ {1 - l}} \\Delta_ {I} \\left(\\hat {I} _ {l} ^ {\\mathrm {c s}}, I _ {l}\\right) + \\lambda_ {m} ^ {\\text {p h o t o}} \\| \\mathcal {M} _ {1 - l} \\|, \\tag {7a}", + "image_path": "44129f2a64b086e07726215ed0e72a78ea3875bf13ce91ed1e349dd128086a3f.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 403, + 228, + 553, + 243 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 228, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 403, + 228, + 553, + 243 + ], + "type": "interline_equation", + "content": "\\hat {I} _ {l} ^ {\\mathrm {c s}} = R \\left(\\mathcal {G} _ {l} ^ {\\mathrm {t g t}, \\mathrm {c s}}\\right), \\tag {7b}", + "image_path": "cadc2a5b64f70c2c93df7049ed0bedc083e83ca9276654de99f06b7708dc5957.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 249, + 553, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 249, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 553, + 274 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 249, + 553, + 274 + ], + "type": "inline_equation", + "content": "\\lambda_{m}^{\\mathrm{photo}}\\| \\mathcal{M}_{1 - l}\\|" + }, + { + "bbox": [ + 313, + 249, + 553, + 274 + ], + "type": "text", + "content": " is a mobility regularization term similar to that in Eqn. 6a." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 280, + 553, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 280, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 553, + 304 + ], + "type": "text", + "content": "3.2.3. Cross-mobile formulation for articulation estimation and mobility refinement" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 307, + 553, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 553, + 391 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 553, + 391 + ], + "type": "text", + "content": "The focus of Stage 3 is to estimate the articulation parameters shared by the mobile components of all Gaussians. To this end, we employ the intuition that the mobile components from the two end states are related through the articulated motion. Formally, we introduce the cross-mobile (cm) formulation, where the mobile part of the articulated object at state " + }, + { + "bbox": [ + 313, + 307, + 553, + 391 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 307, + 553, + 391 + ], + "type": "text", + "content": " is jointly represented as" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 352, + 397, + 553, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 397, + 553, + 412 + ], + "spans": [ + { + "bbox": [ + 352, + 397, + 553, + 412 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {l} ^ {\\mathrm {m}} = \\mathcal {G} _ {l} ^ {\\mathrm {t g t}} \\cdot \\mathcal {M} _ {l} \\oplus {} ^ {l} T _ {1 - l} \\left(\\mathcal {G} _ {1 - l} ^ {\\mathrm {t g t}} \\cdot \\mathcal {M} _ {1 - l}\\right), \\tag {8}", + "image_path": "bfa9c3f8fcc6de9f46e740cc3d79d89ef4007edcd2579c7cc085bfc38bc9338a.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 420, + 553, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 553, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 553, + 456 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 420, + 553, + 456 + ], + "type": "inline_equation", + "content": "{}^l T_{1 - l}(\\mathcal{G})" + }, + { + "bbox": [ + 313, + 420, + 553, + 456 + ], + "type": "text", + "content": " denotes the transformation of Gaussians " + }, + { + "bbox": [ + 313, + 420, + 553, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 420, + 553, + 456 + ], + "type": "text", + "content": " according to the articulated motion from state " + }, + { + "bbox": [ + 313, + 420, + 553, + 456 + ], + "type": "inline_equation", + "content": "1 - l" + }, + { + "bbox": [ + 313, + 420, + 553, + 456 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 420, + 553, + 456 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 420, + 553, + 456 + ], + "type": "text", + "content": ". For state " + }, + { + "bbox": [ + 313, + 420, + 553, + 456 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 420, + 553, + 456 + ], + "type": "text", + "content": ", the target representation then becomes" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 394, + 464, + 553, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 464, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 394, + 464, + 553, + 479 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {l} ^ {\\mathrm {t g t}, \\mathrm {c m}} = \\mathcal {G} ^ {\\mathrm {s}} \\oplus \\mathcal {G} _ {l} ^ {\\mathrm {m}}, \\tag {9}", + "image_path": "a0b9db369c7380566e75c818c6b5f09b6d16f7722f6c09fb83a25de067b3500f.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 485, + 553, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 553, + 509 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 485, + 553, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{\\mathrm{s}}" + }, + { + "bbox": [ + 313, + 485, + 553, + 509 + ], + "type": "text", + "content": " is defined in Eqn. 3. With this formulation, Stage 3 is further divided into three sub-stages as below." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 511, + 553, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 511, + 553, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 553, + 571 + ], + "type": "text", + "content": "Stage 3(a): Coarse articulation estimation via geometric consistency. Similar to Eqn. 6a, weighted Chamfer distance is used for computing the cross-mobile geometric consistency, which can be optimized over both mobilities and articulation parameters as follows:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 390, + 578, + 553, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 578, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 390, + 578, + 553, + 597 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathcal {M}, T} \\mathrm {C D} _ {0} ^ {\\mathrm {c m}} + \\mathrm {C D} _ {1} ^ {\\mathrm {c m}}, \\tag {10a}", + "image_path": "c3da55ce0b6e966dd7e684abc3e940afc70c3584af2be3b24d4a97bf1564d896.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 368, + 599, + 553, + 614 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 599, + 553, + 614 + ], + "spans": [ + { + "bbox": [ + 368, + 599, + 553, + 614 + ], + "type": "interline_equation", + "content": "\\mathrm {C D} _ {l} ^ {\\mathrm {c m}} = \\operatorname {C h a m f e r} \\left(\\mathcal {G} _ {l} ^ {\\mathrm {t g t , c m}}, \\mathcal {G} _ {l} ^ {\\mathrm {r e f}}\\right). \\tag {10b}", + "image_path": "56725bb406ad3b80aea2a185e79305e086351f44c6261982a0088087508a4405.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 621, + 553, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 553, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 553, + 693 + ], + "type": "text", + "content": "However, we notice that this formulation is still susceptible to local minima if the mobile part is too geometrically insignificant. We qualitatively show one such failure in Fig. 3a. To remedy this, we further propose mobile-only geometric consistency, which considers only the mobile components of both states as follows:" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 321, + 700, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 700, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 321, + 700, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\mathrm {C D} ^ {\\mathrm {m}} = \\operatorname {C h a m f e r} \\left(\\mathcal {G} _ {l} ^ {\\mathrm {t g t}} \\cdot \\mathcal {M} _ {l}, ^ {l} T _ {1 - l} \\left(\\mathcal {G} _ {1 - l} ^ {\\mathrm {t g t}} \\cdot \\mathcal {M} _ {1 - l}\\right)\\right). \\tag {11}", + "image_path": "c2b928ce999d1bae87d0575b5dad6950a38a5864a0b0861531221112ce5d08a9.jpg" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8844" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 286, + 128 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 286, + 128 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 286, + 128 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 286, + 128 + ], + "type": "image", + "image_path": "dc57c7c2627f33a42e1bccd935173f966d2e15276d8c88445328b5773ef508e3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 130, + 172, + 149 + ], + "lines": [ + { + "bbox": [ + 55, + 130, + 172, + 149 + ], + "spans": [ + { + "bbox": [ + 55, + 130, + 172, + 149 + ], + "type": "text", + "content": "(a) Failure with cross-mobile geometric consistency." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 176, + 130, + 292, + 149 + ], + "lines": [ + { + "bbox": [ + 176, + 130, + 292, + 149 + ], + "spans": [ + { + "bbox": [ + 176, + 130, + 292, + 149 + ], + "type": "text", + "content": "(b) Success with mobile-only geometric consistency." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 240, + 296, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 296, + 335 + ], + "type": "text", + "content": "While successfully addressing these scenarios, mobile-only geometric consistency still falls short in other circumstances, especially when the mobile components in the two states exhibit large discrepancies. We qualitatively show one such failure in Fig. 4a. To take advantage of both formulations and to facilitate robustness in the inherently nonconvex optimization of articulation parameters, we propose the following practical strategy:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 336, + 288, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 336, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 56, + 336, + 288, + 346 + ], + "type": "text", + "content": "1. With " + }, + { + "bbox": [ + 56, + 336, + 288, + 346 + ], + "type": "inline_equation", + "content": "K^{\\mathrm{m}}" + }, + { + "bbox": [ + 56, + 336, + 288, + 346 + ], + "type": "text", + "content": " randomized tries: articulation estimation via" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 357, + 294, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 357, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 136, + 357, + 294, + 374 + ], + "type": "interline_equation", + "content": "T ^ {m} = \\arg \\min _ {T} \\mathrm {C D} ^ {\\mathrm {m}}. \\tag {12}", + "image_path": "b1135220334f116b7a765be58b028f75f19925818bc7b9da897be3b5d8bbcbc0.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 381, + 295, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 381, + 295, + 404 + ], + "spans": [ + { + "bbox": [ + 55, + 381, + 295, + 404 + ], + "type": "text", + "content": "2. With " + }, + { + "bbox": [ + 55, + 381, + 295, + 404 + ], + "type": "inline_equation", + "content": "K^{\\mathrm{cm}}" + }, + { + "bbox": [ + 55, + 381, + 295, + 404 + ], + "type": "text", + "content": " randomized tries, plus another initialized with " + }, + { + "bbox": [ + 55, + 381, + 295, + 404 + ], + "type": "inline_equation", + "content": "T^{\\mathrm{m}}" + }, + { + "bbox": [ + 55, + 381, + 295, + 404 + ], + "type": "text", + "content": ": articulation estimation via" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 414, + 294, + 431 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 414, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 115, + 414, + 294, + 431 + ], + "type": "interline_equation", + "content": "T ^ {\\mathrm {c m}} = \\arg \\min _ {T} \\mathrm {C D} _ {0} ^ {\\mathrm {c m}} + \\mathrm {C D} _ {1} ^ {\\mathrm {c m}}. \\tag {13}", + "image_path": "6ebc45790687e28c864786608be18fc7e5bd91b926ea78c1e8ae2740aac5e89f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 439, + 295, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 439, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 295, + 464 + ], + "type": "text", + "content": "3. Final run initialized with " + }, + { + "bbox": [ + 55, + 439, + 295, + 464 + ], + "type": "inline_equation", + "content": "T^{\\mathrm{cm}}" + }, + { + "bbox": [ + 55, + 439, + 295, + 464 + ], + "type": "text", + "content": ": joint articulation estimation and mobility refinement via Eqn. 10a." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 465, + 296, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 465, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 465, + 296, + 525 + ], + "type": "text", + "content": "Stage 3(b): Joint articulation, mobility, and Gaussian optimization via cross-mobile rendering. As in Stage 2(b), we perform a full joint optimization over articulation parameters, mobilities, and Gaussians utilizing photometric supervision via cross-mobile rendering as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 118, + 533, + 294, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 533, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 118, + 533, + 294, + 553 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathcal {G} ^ {\\mathrm {i g t}}, \\mathcal {M} _ {1 - l}, ^ {l} T _ {1 - l}} \\Delta_ {I} \\left(\\hat {I} _ {l} ^ {\\mathrm {c m}}, I _ {l}\\right), \\tag {14a}", + "image_path": "a608b384cfb74949b9fa31e4a47f7b6854e79984952fff3fb2adf9f75bc10ced.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 556, + 294, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 556, + 294, + 571 + ], + "spans": [ + { + "bbox": [ + 116, + 556, + 294, + 571 + ], + "type": "interline_equation", + "content": "\\hat {I} _ {l} ^ {\\mathrm {c m}} = R \\left(\\mathcal {G} _ {l} ^ {\\mathrm {t g t , c m}}\\right), l \\in \\{0, 1 \\}. \\tag {14b}", + "image_path": "4d724dfd2d0228ab65613b3936652eb39c3a119039462d1a2ec910685870ceb8.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 579, + 296, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 579, + 296, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 579, + 296, + 675 + ], + "type": "text", + "content": "Stage 3(c): Mobility correction via cross-mobile geometric consistency. Stage 3(b) solely relies on photometric supervision, which is limited to the training views. On rare occasions, a Gaussian may be mistakenly estimated as mobile if the articulated motion moves it out of sight from most views, effectively losing supervision. On the other hand, geometric consistency is not affected by visibility limitations, which we leverage for mobility correction as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 684, + 295, + 701 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 684, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 132, + 684, + 295, + 701 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathcal {M}} \\mathrm {C D} _ {0} ^ {\\mathrm {c m}} + \\mathrm {C D} _ {1} ^ {\\mathrm {c m}}. \\tag {15}", + "image_path": "cbb747b9d6edf346ddf259af9e7fdaf2d55738a7e03d944aa4d1cf24cabe40e6.jpg" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 551, + 145 + ], + "blocks": [ + { + "bbox": [ + 55, + 154, + 295, + 222 + ], + "lines": [ + { + "bbox": [ + 55, + 154, + 295, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 154, + 295, + 222 + ], + "type": "text", + "content": "Figure 3. 103549-Toaster has a flat slider as its mobile part, similar in geometric curvature to the static component. This similarity makes the nearest-neighbor data association error-prone when computing the Chamfer distance for cross-mobile geometric consistency. On the other hand, mobile-only geometric consistency focuses only on the mobile part and successfully handles this case." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 70, + 551, + 145 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 551, + 145 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 551, + 145 + ], + "type": "image", + "image_path": "e05432a5de3057a0ce9daff2e3a02637f34a98379f58e3a003239a47c4a616da.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 148, + 430, + 168 + ], + "lines": [ + { + "bbox": [ + 314, + 148, + 430, + 168 + ], + "spans": [ + { + "bbox": [ + 314, + 148, + 430, + 168 + ], + "type": "text", + "content": "(a) Failure with mobile-only geometric consistency." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 435, + 149, + 551, + 168 + ], + "lines": [ + { + "bbox": [ + 435, + 149, + 551, + 168 + ], + "spans": [ + { + "bbox": [ + 435, + 149, + 551, + 168 + ], + "type": "text", + "content": "(b) Success with cross-mobile geometric consistency." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 174, + 553, + 251 + ], + "lines": [ + { + "bbox": [ + 313, + 174, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 553, + 251 + ], + "type": "text", + "content": "Figure 4. 101713-Pen has both ends of a pen as the mobile part, with only one end visible from any given observation. Thus, the reconstructed mobile components from the two states essentially represent distinct parts, violating the intuition for mobile-only geometric consistency. On the other hand, cross-mobile geometric consistency considers the joint representation from both states as a whole, successfully estimating the articulation." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 258, + 449, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 449, + 272 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 449, + 272 + ], + "type": "text", + "content": "3.3. Real-World Application" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 277, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 456 + ], + "type": "text", + "content": "For an articulated object to be reconstructed in the real world, we first collect two sets of RGB images, one for each articulation state, by imaging the object from a surrounding hemisphere. We then preprocess the data to determine the posed images that serve as the input to SPLART. This involves using SAM 2 [46] to perform foreground-background segmentation, which also removes dynamic contents of the scene. We then perform structure from motion (SfM) to determine camera poses using COLMAP [48] with SuperPoint [7] descriptors and SuperGlue [47] matching on the segmented backgrounds from both sets of images to construct a joint coordinate frame for the object in both states. Once we obtain the joint coordinate frame and the foreground target object is localized, we run SPLART to reconstruct the articulated objects." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 468, + 394, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 394, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 394, + 481 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 488, + 376, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 488, + 376, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 488, + 376, + 499 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 506, + 555, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 673 + ], + "type": "text", + "content": "PARIS PartNet-Mobility Subset. Our evaluation includes PARIS [30] as a baseline, which selects 10 instances (8 revolute, 2 prismatic) from PartNet-Mobility [2, 39, 61], a large-scale dataset of simulatable 3D objects with part-level mobility. We refer to this subset as PARIS-PMS. For each articulation state, the dataset provides 100 calibrated object-centric views for training and 50 for testing, sampled from the upper hemisphere. However, the released dataset lacks ground-truth depth and part segmentation maps. To enable their evaluations, we augment PARIS-PMS with the necessary ground-truth data by following its data generation procedure. The dataset does not provide test views of intermediate articulation states, which limits the quantitative evaluation of novel articulation synthesis." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "SPLART PartNet-Mobility Subset. We curate an additional articulated object dataset from PartNet-Mobility, dubbed SPLART-PMS, to enable a more comprehensive" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8845" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 185 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 185 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 185 + ], + "type": "table", + "html": "
TypeMethodSuccess Rate %↑erra(x10-2DEG)↓errp(x10-3)↓errr(x10-2DEG)↓errt(x10-3)↓CDs(x10-3)↓CDm(x10-3)↓CDw(x10-3)↓
RevolutePARIS40.0121.90 ± 68.455.27 ± 4.55127.80 ± 55.12N/A3.48 ± 0.3968.44 ± 13.0313.11 ± 1.16
DTA†98.812.32 ± 4.161.86 ± 1.6919.41 ± 16.28N/A2.20 ± 0.491.27 ± 1.871.73 ± 0.03
SPLART100.03.70 ± 0.380.40 ± 0.074.85 ± 0.33N/A4.08 ± 0.321.06 ± 0.783.59 ± 0.21
PrismaticPARIS50.027.97 ± 13.09N/AN/A4.28 ± 3.029.21 ± 1.94151.78 ± 35.007.99 ± 0.49
DTA†100.016.26 ± 3.54N/AN/A1.15 ± 0.142.69 ± 0.0415.74 ± 0.282.22 ± 0.03
SPLART95.02.36 ± 0.44N/AN/A0.33 ± 0.046.02 ± 0.2016.77 ± 1.143.69 ± 0.14
OverallPARIS42.0103.12 ± 57.385.27 ± 4.55127.80 ± 55.124.28 ± 3.024.63 ± 0.7085.11 ± 17.4212.09 ± 1.03
DTA†99.013.11 ± 4.041.86 ± 1.6919.41 ± 16.281.15 ± 0.142.30 ± 0.404.16 ± 1.551.83 ± 0.03
SPLART99.03.44 ± 0.390.40 ± 0.074.85 ± 0.330.33 ± 0.044.47 ± 0.304.20 ± 0.853.61 ± 0.19
", + "image_path": "0f83d2a657f7430bae79d483a2a850dadeb59ec6b75e71ca720e998daf835123.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 209, + 294, + 343 + ], + "blocks": [ + { + "bbox": [ + 78, + 188, + 531, + 200 + ], + "lines": [ + { + "bbox": [ + 78, + 188, + 531, + 200 + ], + "spans": [ + { + "bbox": [ + 78, + 188, + 531, + 200 + ], + "type": "text", + "content": "Table 1. PARIS-PMS Articulation and Mesh Reconstruction Metrics. Best results in bold. " + }, + { + "bbox": [ + 78, + 188, + 531, + 200 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 78, + 188, + 531, + 200 + ], + "type": "text", + "content": " DTA requires ground-truth depth." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 209, + 294, + 343 + ], + "lines": [ + { + "bbox": [ + 58, + 209, + 294, + 343 + ], + "spans": [ + { + "bbox": [ + 58, + 209, + 294, + 343 + ], + "type": "table", + "html": "
TypeMethodPSNR ↑Depth MAE ↓mIoU ↑
RevolutePARIS32.210.0930.955
DTA†N/A0.0310.941
SPLART43.530.0390.974
PrismaticPARIS33.750.1080.902
DTA†N/A0.0660.844
SPLART44.380.0250.890
OverallPARIS32.520.0960.942
DTA†N/A0.0380.922
SPLART43.700.0360.957
", + "image_path": "97a57adb50ed9030444d0d0a0d999e1878348d4d404bf2f013e059411f635bd8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 346, + 295, + 381 + ], + "lines": [ + { + "bbox": [ + 55, + 346, + 295, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 295, + 381 + ], + "type": "text", + "content": "Table 2. PARIS-PMS Novel View Synthesis Metrics, averaged over successful runs (see Tab. 1 for success rates). Best results in bold. " + }, + { + "bbox": [ + 55, + 346, + 295, + 381 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 55, + 346, + 295, + 381 + ], + "type": "text", + "content": " DTA requires ground-truth depth." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 388, + 295, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 388, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 54, + 388, + 295, + 543 + ], + "type": "text", + "content": "evaluation including the quantitative assessment of novel articulation synthesis. Our dataset consists of 22 object instances from categories not included in PARIS-PMS, with 12 revolute and 10 prismatic objects. For the test split, we also generate ground-truth depth and part segmentation maps to facilitate relevant evaluations. Crucially, while PARIS-PMS restricts articulation states to binary values, SPLART-PMS uniformly samples states from " + }, + { + "bbox": [ + 54, + 388, + 295, + 543 + ], + "type": "inline_equation", + "content": "[-0.1, 1.1]" + }, + { + "bbox": [ + 54, + 388, + 295, + 543 + ], + "type": "text", + "content": ", demanding accurate estimation of articulation, part segmentation, and object reconstruction for correct view synthesis. Our experiments reveal that SPLART-PMS poses a greater challenge than PARIS-PMS—the latter being almost perfectly solved by our method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 544, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 628 + ], + "type": "text", + "content": "Real-World Dataset. To show the efficacy of SPLART on real-world data, we collect a dataset that consists of 7 real-world objects with 9 articulations in total, including common objects like a monitor, cabinet, glasses, etc. All images are captured by a hand-held phone camera. Camera parameters are calibrated following strategies detailed in Sec. 3.3. We qualitatively evaluate SPLART on this dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 636, + 168, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 168, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 168, + 647 + ], + "type": "text", + "content": "4.2. Evaluation Metrics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 677 + ], + "type": "text", + "content": "Articulation Estimation. We evaluate the accuracy of articulation estimation depending on its type." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "- A revolute articulation describes a rotation around some line in 3D space, parameterized by the pivot point " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " on the line, the line's directional axis " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": ", and the angle of" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 314, + 209, + 555, + 380 + ], + "blocks": [ + { + "bbox": [ + 314, + 209, + 555, + 380 + ], + "lines": [ + { + "bbox": [ + 314, + 209, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 314, + 209, + 555, + 380 + ], + "type": "image", + "image_path": "ed97e3e78886aff688c7f20c9ab3d22384ec4fa1f464cfe9a66b9e987b33960e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 317, + 408, + 553, + 582 + ], + "blocks": [ + { + "bbox": [ + 328, + 388, + 539, + 399 + ], + "lines": [ + { + "bbox": [ + 328, + 388, + 539, + 399 + ], + "spans": [ + { + "bbox": [ + 328, + 388, + 539, + 399 + ], + "type": "text", + "content": "Figure 5. Real-World Articulated Object Reconstructions." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 408, + 553, + 582 + ], + "lines": [ + { + "bbox": [ + 317, + 408, + 553, + 582 + ], + "spans": [ + { + "bbox": [ + 317, + 408, + 553, + 582 + ], + "type": "table", + "html": "
TypeMethodSuccess Rate %errn(×10-2DEG) ↓errp(×10-3) ↓errr(×10-2DEG) ↓errt(×10-3) ↓
RevolutePARIS18.3167.34 ± 42.479.36 ± 3.73270.37 ± 91.19N/A
DTA†33.332.76 ± 2.381.28 ± 0.3940.80 ± 4.66N/A
SPLART-2a78.52.91 ± 0.552.53 ± 2.1852.06 ± 32.00N/A
SPLART-2b67.04.62 ± 1.693.92 ± 3.07131.99 ± 64.45N/A
SPLART-3a0.8296.61 ± 0.002.83 ± 0.00427.10 ± 0.00N/A
SPLART-3b76.784.13 ± 19.135.11 ± 3.03206.98 ± 56.30N/A
SPLART77.520.76 ± 0.813.93 ± 2.4178.31 ± 31.06N/A
PrismaticPARIS39.049.49 ± 18.92N/AN/A14.96 ± 5.51
DTA†90.0102.77 ± 12.48N/AN/A9.25 ± 0.73
SPLART-2a90.025.88 ± 23.25N/AN/A1.50 ± 1.63
SPLART-2b85.715.63 ± 20.40N/AN/A0.73 ± 0.58
SPLART-3a13.015.17 ± 0.59N/AN/A2.26 ± 0.04
SPLART-3b86.061.00 ± 19.76N/AN/A9.43 ± 2.37
SPLART95.025.72 ± 17.63N/AN/A1.04 ± 0.32
OverallPARIS27.7103.88 ± 29.799.36 ± 3.73270.37 ± 91.1914.96 ± 5.51
DTA†59.177.76 ± 8.871.28 ± 0.3940.80 ± 4.669.25 ± 0.73
SPLART-2a83.713.85 ± 11.362.53 ± 2.1852.06 ± 32.001.50 ± 1.63
SPLART-2b74.79.34 ± 9.713.92 ± 3.07131.99 ± 64.450.73 ± 0.58
SPLART-3a6.485.53 ± 0.442.83 ± 0.00427.10 ± 0.002.26 ± 0.04
SPLART-3b80.973.62 ± 19.425.11 ± 3.03206.98 ± 56.309.43 ± 2.37
SPLART85.523.02 ± 8.463.93 ± 2.4178.31 ± 31.061.04 ± 0.32
", + "image_path": "b06e4b21ac42b7874d2a098070d890b030fe3091729386227ace9bba6f167263.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 585, + 553, + 608 + ], + "lines": [ + { + "bbox": [ + 314, + 585, + 553, + 608 + ], + "spans": [ + { + "bbox": [ + 314, + 585, + 553, + 608 + ], + "type": "text", + "content": "Table 3. SPLART-PMS Articulation Metrics. Best results in bold. " + }, + { + "bbox": [ + 314, + 585, + 553, + 608 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 314, + 585, + 553, + 608 + ], + "type": "text", + "content": " DTA requires ground-truth depth." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 321, + 617, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 617, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 321, + 617, + 555, + 715 + ], + "type": "text", + "content": "rotation " + }, + { + "bbox": [ + 321, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 321, + 617, + 555, + 715 + ], + "type": "text", + "content": ". We report the angular error " + }, + { + "bbox": [ + 321, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{err}_{\\mathbf{a}}" + }, + { + "bbox": [ + 321, + 617, + 555, + 715 + ], + "type": "text", + "content": " (in " + }, + { + "bbox": [ + 321, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\times 10^{-2}" + }, + { + "bbox": [ + 321, + 617, + 555, + 715 + ], + "type": "text", + "content": " DEG) between the predicted and ground-truth axes, the geodesic distance " + }, + { + "bbox": [ + 321, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{err}_{\\mathrm{r}}" + }, + { + "bbox": [ + 321, + 617, + 555, + 715 + ], + "type": "text", + "content": " (in " + }, + { + "bbox": [ + 321, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\times 10^{-2}" + }, + { + "bbox": [ + 321, + 617, + 555, + 715 + ], + "type": "text", + "content": " DEG) between the predicted and ground-truth rotations induced by the axis-angle (" + }, + { + "bbox": [ + 321, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{a} - \\theta" + }, + { + "bbox": [ + 321, + 617, + 555, + 715 + ], + "type": "text", + "content": ") pair, and the pivot point error " + }, + { + "bbox": [ + 321, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{err}_{\\mathbf{p}}" + }, + { + "bbox": [ + 321, + 617, + 555, + 715 + ], + "type": "text", + "content": ". Since the pivot can move arbitrarily along the axis, " + }, + { + "bbox": [ + 321, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{err}_{\\mathbf{p}}" + }, + { + "bbox": [ + 321, + 617, + 555, + 715 + ], + "type": "text", + "content": " is computed as the closest distance between the predicted and ground-truth lines induced by the axis-pivot (" + }, + { + "bbox": [ + 321, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{a} - \\mathbf{p}" + }, + { + "bbox": [ + 321, + 617, + 555, + 715 + ], + "type": "text", + "content": ") pair." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8846" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 293, + 201 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 293, + 201 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 293, + 201 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 293, + 201 + ], + "type": "table", + "html": "
TypeMethodPSNR ↑Depth MAE ↓mIoU ↑
RevolutePARIS32.750.1250.897
DTA†N/A0.0530.886
SPLART36.380.0320.897
PrismaticPARIS32.300.1240.902
DTA†N/A0.0600.857
SPLART37.770.0330.938
OverallPARIS32.510.1250.900
DTA†N/A0.0580.867
SPLART37.010.0320.916
", + "image_path": "8ee028dbb6801161217d0a59f87602dcd9013e20416dad3198293dcd28fb2e46.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 206, + 295, + 242 + ], + "lines": [ + { + "bbox": [ + 55, + 206, + 295, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 295, + 242 + ], + "type": "text", + "content": "Table 4. SPLART-PMS Novel View and Articulation Synthesis Metrics, averaged over successful runs (see Tab. 3 for success rates). Best results in bold. " + }, + { + "bbox": [ + 55, + 206, + 295, + 242 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 206, + 295, + 242 + ], + "type": "text", + "content": "DTA requires ground-truth depth." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 248, + 296, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 248, + 296, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 296, + 333 + ], + "type": "text", + "content": "- A prismatic articulation describes a translation along a specific direction, parameterized by the axis " + }, + { + "bbox": [ + 55, + 248, + 296, + 333 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 55, + 248, + 296, + 333 + ], + "type": "text", + "content": " of the translation direction and the translation distance " + }, + { + "bbox": [ + 55, + 248, + 296, + 333 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 248, + 296, + 333 + ], + "type": "text", + "content": ". We report the axis error " + }, + { + "bbox": [ + 55, + 248, + 296, + 333 + ], + "type": "inline_equation", + "content": "\\mathrm{err}_{\\mathbf{a}}" + }, + { + "bbox": [ + 55, + 248, + 296, + 333 + ], + "type": "text", + "content": ", as in the revolute case, and the translation error " + }, + { + "bbox": [ + 55, + 248, + 296, + 333 + ], + "type": "inline_equation", + "content": "\\mathrm{err}_{\\mathrm{t}}" + }, + { + "bbox": [ + 55, + 248, + 296, + 333 + ], + "type": "text", + "content": ", which is the distance between the predicted and ground-truth translations induced by the axis-distance " + }, + { + "bbox": [ + 55, + 248, + 296, + 333 + ], + "type": "inline_equation", + "content": "(\\mathbf{a} - d)" + }, + { + "bbox": [ + 55, + 248, + 296, + 333 + ], + "type": "text", + "content": " pair." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 338, + 296, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 296, + 661 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 296, + 661 + ], + "type": "text", + "content": "Part-Level Reconstruction. The evaluation of part-level reconstruction accuracy is threefold: photometric rendering quality, geometric accuracy, and part segmentation accuracy. For all these aspects, we use novel view synthesis as the surrogate task. We perform volume rendering for each view in the test split, generating outputs that include the RGB image, depth map, and part segmentation map. Photometric rendering quality is evaluated by reporting image metrics PSNR. Geometric accuracy is assessed by reporting the mean absolute error of the depth map (i.e., Depth MAE). Part segmentation accuracy is evaluated using the intersection-over-union (IoU) ratio for three categories: static part " + }, + { + "bbox": [ + 56, + 338, + 296, + 661 + ], + "type": "inline_equation", + "content": "(\\mathrm{IoU_s})" + }, + { + "bbox": [ + 56, + 338, + 296, + 661 + ], + "type": "text", + "content": ", mobile part " + }, + { + "bbox": [ + 56, + 338, + 296, + 661 + ], + "type": "inline_equation", + "content": "(\\mathrm{IoU_m})" + }, + { + "bbox": [ + 56, + 338, + 296, + 661 + ], + "type": "text", + "content": ", and background " + }, + { + "bbox": [ + 56, + 338, + 296, + 661 + ], + "type": "inline_equation", + "content": "(\\mathrm{IoU_{bg}})" + }, + { + "bbox": [ + 56, + 338, + 296, + 661 + ], + "type": "text", + "content": ". The mean IoU (mIoU) is then reported as the average IoU across these categories. In addition, we also evaluate geometric accuracy through mesh reconstruction. To extract a mesh from 3DGS, we: (1) render depth images from uniformly sampled spherical viewpoints, (2) fuse rendered depths into a TSDF representation [4, 41], and (3) extract the mesh using the Marching Cubes algorithm [34]. For each reconstruction, we separately extract the meshes corresponding to the static, mobile, and whole parts. For evaluation, we follow the procedure used in previous works [30, 60]: 10 000 points are uniformly sampled from both the reconstructed and the ground-truth meshes, and the Chamfer distance is computed for each category—static " + }, + { + "bbox": [ + 56, + 338, + 296, + 661 + ], + "type": "inline_equation", + "content": "(\\mathrm{CD_s})" + }, + { + "bbox": [ + 56, + 338, + 296, + 661 + ], + "type": "text", + "content": ", mobile " + }, + { + "bbox": [ + 56, + 338, + 296, + 661 + ], + "type": "inline_equation", + "content": "(\\mathrm{CD_m})" + }, + { + "bbox": [ + 56, + 338, + 296, + 661 + ], + "type": "text", + "content": ", and whole " + }, + { + "bbox": [ + 56, + 338, + 296, + 661 + ], + "type": "inline_equation", + "content": "(\\mathrm{CD_w})" + }, + { + "bbox": [ + 56, + 338, + 296, + 661 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Impact of Randomness. Randomness can play a significant role in the joint optimization of articulation and part reconstruction [6, 30]. In order to effectively demonstrate the robustness of a method, we conduct ten trials for each" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 554, + 167 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 554, + 167 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 554, + 167 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 554, + 167 + ], + "type": "image", + "image_path": "2fc06219444e5914e71f5ae360c285e9c2734e3be1238f7c20584afe6651661b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 175, + 544, + 186 + ], + "lines": [ + { + "bbox": [ + 324, + 175, + 544, + 186 + ], + "spans": [ + { + "bbox": [ + 324, + 175, + 544, + 186 + ], + "type": "text", + "content": "Figure 6. State interpolation from the estimated articulation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 194, + 555, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 555, + 350 + ], + "type": "text", + "content": "scene, each with a different fixed random seed. We report results averaged over the ten runs. Additionally, we notice that for a failed run converging to a local minimum, the evaluation metrics, such as " + }, + { + "bbox": [ + 313, + 194, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\mathrm{err}_{\\mathbf{r}}" + }, + { + "bbox": [ + 313, + 194, + 555, + 350 + ], + "type": "text", + "content": ", are typically orders-of-magnitude worse than those of a successful run. To better reflect a method's performance, we report the number of successful runs out of the total number (i.e., ten), where a run is deemed successful if and only if each of the following criteria are met: (1) " + }, + { + "bbox": [ + 313, + 194, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\mathrm{err}_{\\mathbf{a}} < 5" + }, + { + "bbox": [ + 313, + 194, + 555, + 350 + ], + "type": "text", + "content": " DEG; (2) " + }, + { + "bbox": [ + 313, + 194, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\mathrm{err}_{\\mathbf{p}} < 0.05" + }, + { + "bbox": [ + 313, + 194, + 555, + 350 + ], + "type": "text", + "content": " for revolute articulation; (3) " + }, + { + "bbox": [ + 313, + 194, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\mathrm{err}_{\\mathbf{r}} < 10" + }, + { + "bbox": [ + 313, + 194, + 555, + 350 + ], + "type": "text", + "content": " DEG for revolute articulation; and (4) " + }, + { + "bbox": [ + 313, + 194, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\mathrm{err}_{\\mathbf{t}} < 0.05" + }, + { + "bbox": [ + 313, + 194, + 555, + 350 + ], + "type": "text", + "content": " for prismatic articulation. We then proceed to report the evaluation metrics detailed previously, averaging only over the successful runs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 357, + 381, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 357, + 381, + 369 + ], + "spans": [ + { + "bbox": [ + 314, + 357, + 381, + 369 + ], + "type": "text", + "content": "4.3. Baselines" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 375, + 554, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 554, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 554, + 471 + ], + "type": "text", + "content": "PARIS [30] addresses the same challenge we tackle: reconstructing the part-level shape, appearance, and motion of an articulated object from multi-view RGB images captured at two articulation states. It employs a NeRF-based representation, modeling the object with static and mobile fields for each part, alongside a transformation to capture state transitions. These components are jointly optimized end-to-end, primarily through an image-rendering loss." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 473, + 554, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 554, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 554, + 594 + ], + "type": "text", + "content": "DTA [60] adopts a similar setup: reconstructing digital twins of articulated objects from multi-view RGB-D observations at two articulation states. It operates in two stages: Stage 1 optimizes neural object fields for each state using RGB-D images and extracts meshes; Stage 2 optimizes a part segmentation field and per-part motions using consistency, matching, and collision losses to determine point correspondences between states. Note: Unlike SPLART, DTA requires ground-truth depth as input and produces plain object meshes unsuitable for photorealistic rendering." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 600, + 430, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 430, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 430, + 613 + ], + "type": "text", + "content": "4.4. Experiment Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "Articulation Estimation. We quantitatively evaluate SPLART against baselines for articulation estimation accuracy, reporting scene-averaged results per articulation type on the PARIS-PMS dataset (Tab. 1) and the SPLART-PMS dataset (Tab. 3). SPLART consistently outperforms the baselines across both datasets. Additionally, we qualitatively compare SPLART with PARIS on selected scenes from these datasets in Fig. 7. In the \"color\" column, each" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8847" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 555, + 227 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 227 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 227 + ], + "type": "image", + "image_path": "bae55b184d543a19622de379ebc3e97a12b08962128574169594e36aba0b8bb2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 142, + 231, + 465, + 243 + ], + "lines": [ + { + "bbox": [ + 142, + 231, + 465, + 243 + ], + "spans": [ + { + "bbox": [ + 142, + 231, + 465, + 243 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison on part-level reconstruction and articulation estimation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 250, + 296, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 250, + 296, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 296, + 310 + ], + "type": "text", + "content": "image visualizes the ground-truth articulation (green arrow) and the estimated articulation (red arrow), except in the ground-truth column. When the estimate aligns perfectly with the ground truth, only the red arrow appears; if significantly misaligned, only the green arrow is visible." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 320, + 296, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 320, + 296, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 320, + 296, + 489 + ], + "type": "text", + "content": "Part-Level Reconstruction. We assess SPLART's part-level reconstruction accuracy against baselines, using novel view synthesis as a surrogate task. Scene-averaged results per articulation type are reported on the PARIS-PMS dataset (Tab. 2) and the SPLART-PMS dataset (Tab. 4). We also evaluate geometric accuracy through mesh reconstruction, presenting scene-averaged results per articulation type on the PARIS-PMS dataset (Tab. 1). Without depth supervision, which is required for the DTA baseline, SPLART matches DTA's performance, both substantially exceeding PARIS. Qualitative comparisons with PARIS, including RGB renderings and part segmentation maps, are provided for selected scenes from both datasets in Fig. 7. Additional qualitative results on articulation synthesis appear in Fig. 6." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 498, + 296, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 547 + ], + "type": "text", + "content": "Real-World Reconstructions. We qualitatively evaluate SPLART on the aforementioned real-world dataset, with results shown in Fig. 5. Each image overlays the estimated articulation, visualized with a red arrow." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": "Ablation Studies. We perform extensive ablation studies to demonstrate the contribution of each stage. Specifically, we perform the following four ablations: (1) No geometric supervision for mobility estimation, i.e., skipping Stage 2(a), denoted SPLART-2a; (2) No photometric supervision for mobility estimation, i.e., skipping Stage 2(b), denoted SPLART-2b; (3) No geometric supervision for articulation estimation, i.e., skipping Stage 3(a), denoted SPLART-3a; (4) No photometric supervision for articulation estimation, i.e., skipping Stage 3(b), denoted SPLART-3b. Like the main experiments, we conduct 10 runs for each scene, but only on the more challenging dataset of SPLART-PMS. We report the articulation estimation results in Tab. 3." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 315, + 247, + 555, + 319 + ], + "blocks": [ + { + "bbox": [ + 315, + 247, + 555, + 319 + ], + "lines": [ + { + "bbox": [ + 315, + 247, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 315, + 247, + 555, + 319 + ], + "type": "image", + "image_path": "0b517baf1144dfd9df63c40504710b4fc005b4854d23a0fd7254f1f8780bfc26.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 323, + 553, + 334 + ], + "lines": [ + { + "bbox": [ + 313, + 323, + 553, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 553, + 334 + ], + "type": "text", + "content": "Figure 8. Failure cases due to inherent ambiguities in articulation." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 338, + 388, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 388, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 388, + 350 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 354, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 713 + ], + "type": "text", + "content": "We present SPLART, the first self-supervised method that utilizes 3D Gaussian Splatting for articulated object reconstruction from two-state RGB observations, without 3D supervision or category-specific priors. SPLART delivers robust optimization and effectively handles challenging scenarios—outperforming prior approaches while eliminating the need for 3D supervision, articulation annotations, or semantic labels. Extensive evaluations on synthetic and real-world datasets show SPLART's articulation estimation accuracy and view synthesis quality that surpass existing methods. Real-world experiments further validate its practicality, showcasing successful reconstructions of diverse articulated objects using only a handheld RGB camera. Our framework empowers non-expert users to effortlessly create high-fidelity, articulated digital twins, supporting applications in augmented/virtual reality and robotics. However, while SPLART effectively addresses two-part articulations, it lacks direct support for objects with more articulated parts, though iteratively applying it to incremental state pairs may help. Some existing methods [6, 33, 60] handle such objects, but require known part counts and specific articulation structures. Future research will focus on a comprehensive adaptation to multi-level articulated objects. Additionally, inherent ambiguities may arise when multiple articulations explain state differences, as shown in Fig. 8 with examples like a clock's hand rotating on an unexpected axis or a globe's surface spinning independently. Overall, SPLART represents a scalable, robust, and practical solution that pushes the boundaries of articulated object reconstruction, laying a strong foundation for future work." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8848" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Ben Abbatematteo, Stefanie TELlex, and George Konidaris. Learning to generalize kinematic models to novel objects. In Proceedings of the Conference on Robot Learning (CoRL), pages 1289-1299, 2019. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "type": "text", + "content": "[2] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. ShapeNet: An information-rich 3D model repository. arXiv preprint arXiv:1512.03012, 2015. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 296, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 296, + 249 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 296, + 249 + ], + "type": "text", + "content": "[3] Christopher B. Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3D-R2N2: A unified approach for single and multi-view 3D object reconstruction. In Proceedings of the European Conference on Computer Vision (ECCV), pages 628-644, 2016. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 251, + 296, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 296, + 304 + ], + "type": "text", + "content": "[4] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), pages 303-312, 1996. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 306, + 296, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 306, + 296, + 351 + ], + "spans": [ + { + "bbox": [ + 62, + 306, + 296, + 351 + ], + "type": "text", + "content": "[5] Andrea F. Daniele, Thomas M. Howard, and Matthew R. Walter. A multiview approach to learning articulated motion models. In Proceedings of the International Symposium of Robotics Research (ISRR), pages 371-386, 2017. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 353, + 296, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 353, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 353, + 296, + 407 + ], + "type": "text", + "content": "[6] Jianning Deng, Kartic Subr, and Hakan Bilen. Articulate your NeRF: Unsupervised articulated object modeling via conditional view synthesis. In Advances in Neural Information Processing Systems (NeurIPS), pages 119717-119741, 2025. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 409, + 296, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 296, + 464 + ], + "type": "text", + "content": "[7] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 224-236, 2018. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 466, + 296, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 466, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 466, + 296, + 521 + ], + "type": "text", + "content": "[8] Haoqiang Fan, Hao Su, and Leonidas J. Guibas. A point set generation network for 3D object reconstruction from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2463-2471, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 522, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 522, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 522, + 296, + 578 + ], + "type": "text", + "content": "[9] Samir Yitzhak Gadre, Kiana Ehsani, and Shuran Song. Act the part: Learning interaction strategies for articulated object part discovery. In Proceedings of the International Conference on Computer Vision (ICCV), pages 15732-15741, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 580, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 580, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 580, + 296, + 645 + ], + "type": "text", + "content": "[10] Haoran Geng, Ziming Li, Yiran Geng, Jiayi Chen, Hao Dong, and He Wang. PartManip: Learning cross-category generalizable part manipulation policy from point cloud observations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2978-2988, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 647, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 714 + ], + "type": "text", + "content": "[11] Haoran Geng, Helin Xu, Chengyang Zhao, Chao Xu, Li Yi, Siyuan Huang, and He Wang. GAPartNet: Cross-category domain-generalizable object perception and manipulation via generalizable and actionable parts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7081-7091, 2023. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "type": "text", + "content": "[12] Antoine Guédon and Vincent Lepetit. SuGaR: Surface-aligned Gaussian splattering for efficient 3D mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5354-5363, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 129, + 555, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 555, + 204 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 555, + 204 + ], + "type": "text", + "content": "[13] Nick Heppert, Muhammad Zubair Irshad, Sergey Zakharov, Katherine Liu, Rares Andrei Ambrus, Jeannette Bohg, Abhinav Valada, and Thomas Kollar. CARTO: Category and joint agnostic reconstruction of articulated objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21201-21210, 2023. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 205, + 554, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 205, + 554, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 205, + 554, + 259 + ], + "type": "text", + "content": "[14] Cheng-Chun Hsu, Zhenyu Jiang, and Yuke Zhu. Ditto in the house: Building articulation models of indoor scenes through interactive perception. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 3933-3939, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 260, + 554, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 260, + 554, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 260, + 554, + 304 + ], + "type": "text", + "content": "[15] Ruizhen Hu, Wenchao Li, Oliver Van Kaick, Ariel Shamir, Hao Zhang, and Hui Huang. Learning to predict part mobility from a single static snapshot. ACM Transactions On Graphics, 36(6):1-13, 2017. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 304, + 554, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 554, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 554, + 358 + ], + "type": "text", + "content": "[16] Xiaoxia Huang, Ian Walker, and Stan Birchfield. Occlusion-aware reconstruction and manipulation of 3D articulated objects. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 1365-1371, 2012. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 360, + 554, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 360, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 317, + 360, + 554, + 426 + ], + "type": "text", + "content": "[17] Muhammad Zubair Irshad, Thomas Kollar, Michael Laskey, Kevin Stone, and Zsolt Kira. CenterSnap: Single-shot multi-object 3D shape reconstruction and categorical 6D pose and size estimation. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 10632-10640, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 426, + 554, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 554, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 554, + 480 + ], + "type": "text", + "content": "[18] Muhammad Zubair Irshad, Sergey Zakharov, Rares Ambrus, Thomas Kollar, Zsolt Kira, and Adrien Gaidon. ShAPO: Implicit representations for multi-object shape, appearance, and pose optimization. In Proceedings of the European Conference on Computer Vision (ECCV), pages 275–292, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 481, + 554, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 481, + 554, + 524 + ], + "spans": [ + { + "bbox": [ + 317, + 481, + 554, + 524 + ], + "type": "text", + "content": "[19] Muhammad Zubair Irshad, Mauro Comi, Yen-Chen Lin, Nick Heppert, Abhinav Valada, Rares Ambrus, Zsolt Kira, and Jonathan Tremblay. Neural fields in robotics: A survey, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 525, + 554, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 525, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 317, + 525, + 554, + 590 + ], + "type": "text", + "content": "[20] Ajinkya Jain, Rudolf Lioutikov, Caleb Chuck, and Scott Niekum. ScrewNet: Category-independent articulation model estimation from depth images using screw theory. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 13670-13677, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 592, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 553, + 613 + ], + "type": "text", + "content": "[21] Anil K. Jain. Fundamentals of Digital Image Processing. Prentice-Hall, Inc., 1989. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 614, + 554, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 554, + 658 + ], + "type": "text", + "content": "[22] Hanxiao Jiang, Yongsen Mao, Manolis Savva, and Angel X. Chang. OPD: Single-view 3D openable part detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 410-426, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "type": "text", + "content": "[23] Zhenyu Jiang, Cheng-Chun Hsu, and Yuke Zhu. Ditto: Building digital twins of articulated objects from interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5606-5616, 2022. 1, 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8849" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[24] Dov Katz, Moslem Kazemi, J. Andrew Bagnell, and Anthony Stentz. Interactive segmentation, tracking, and kinematic modeling of unknown 3D articulated objects. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 5003-5010, 2013. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "type": "text", + "content": "[25] Yuki Kawana, Yusuke Mukuta, and Tatsuya Harada. Unsupervised pose-aware part decomposition for man-made articulated objects. In Proceedings of the European Conference on Computer Vision (ECCV), pages 558-575, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 294, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 215 + ], + "type": "text", + "content": "[26] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4):1-14, 2023. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 271 + ], + "type": "text", + "content": "[27] Justin Kerr, Chung Min Kim, Mingxuan Wu, Brent Yi, Qianqian Wang, Ken Goldberg, and Angjoo Kanazawa. Robot see robot do: Imitating articulated object manipulation with monocular 4d reconstruction. In Proceedings of the Conference on Robot Learning (CoRL), pages 587-603, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 337 + ], + "type": "text", + "content": "[28] Long Le, Jason Xie, William Liang, Hung-Ju Wang, Yue Yang, Yecheng Jason Ma, Kyle Vedder, Arjun Krishna, Dinesh Jayaraman, and Eric Eaton. Articulate-anything: Automatic modeling of articulated objects via a vision-language foundation model. In Proceedings of the International Conference on Learning Representations (ICLR), 2025. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "type": "text", + "content": "[29] Xiaolong Li, He Wang, Li Yi, Leonidas J. Guibas, A. Lynn Abbott, and Shuran Song. Category-level articulated object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3703-3712, 2020. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "type": "text", + "content": "[30] Jiayi Liu, Ali Mahdavi-Amiri, and Manolis Savva. PARIS: Part-level reconstruction and motion analysis for articulated objects. In Proceedings of the International Conference on Computer Vision (ICCV), pages 352-363, 2023. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "type": "text", + "content": "[31] Liu Liu, Wenqiang Xu, Haoyuan Fu, Sucheng Qian, Qiao-jun Yu, Yang Han, and Cewu Lu. A real-world articulated object knowledge base. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14789-14798, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 492, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 294, + 546 + ], + "type": "text", + "content": "[32] Shaowei Liu, Saurabh Gupta, and Shenlong Wang. Building rearticulable models for arbitrary 3D objects from 4D point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21138-21147, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "type": "text", + "content": "[33] Yu Liu, Baoxiong Jia, Ruijie Lu, Junfeng Ni, Song-Chun Zhu, and Siyuan Huang. Building interactable replicas of complex articulated objects via gaussian splatting. In Proceedings of the International Conference on Learning Representations (ICLR), 2025. 1, 2, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 603, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 646 + ], + "type": "text", + "content": "[34] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3D surface construction algorithm. In *Seminal Graphics: Pioneering Efforts That Shaped the Field*, pages 347–353, 1998. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "type": "text", + "content": "[35] Zhao Mandi, Yijia Weng, Dominik Bauer, and Shuran Song. Real2code: Reconstruct articulated objects via code generation. In Proceedings of the International Conference on Learning Representations (ICLR), 2025. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[36] Eric Marchand, Hideaki Uchiyama, and Fabien Spindler. Pose estimation for augmented reality: A hands-on survey." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "IEEE Transactions on Visualization and Computer Graphics, 22(12):2633-2651, 2016. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 95, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 149 + ], + "type": "text", + "content": "[37] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3D reconstruction in function space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4460-4470, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 150, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 150, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 553, + 205 + ], + "type": "text", + "content": "[38] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 205, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 205, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 205, + 553, + 271 + ], + "type": "text", + "content": "[39] Kaichun Mo, Shilin Zhu, Angel X. Chang, Li Yi, Subarna Tripathi, Leonidas J. Guibas, and Hao Su. PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 909-918, 2019. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 272, + 553, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 336 + ], + "type": "text", + "content": "[40] Jiteng Mu, Weichao Qiu, Adam Kortylewski, Alan Yuille, Nuno Vasconcelos, and Xiaolong Wang. A-SDF: Learning disentangled signed distance functions for articulated shape representation. In Proceedings of the International Conference on Computer Vision (ICCV), pages 12981-12991, 2021. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 338, + 553, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 413 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 413 + ], + "type": "text", + "content": "[41] Richard A. Newcombe, Andrew Fitzgibbon, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J. Davison, Pushmeet Kohi, Jamie Shotton, and Steve Hodges. KinectFusion: Real-time dense surface mapping and tracking. In Proceedings of the IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 127-136, 2011. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 415, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 448 + ], + "type": "text", + "content": "[42] Neil Nie, Samir Yitzhak Gadre, Kiana Ehsani, and Shuran Song. Structure from action: Learning interactions for articulated object 3D structure discovery, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "type": "text", + "content": "[43] Michael Oechsle, Songyou Peng, and Andreas Geiger. UNISURF: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the International Conference on Computer Vision (ICCV), pages 5589-5599, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 503, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 503, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 503, + 553, + 568 + ], + "type": "text", + "content": "[44] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 165-174, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 569, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 569, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 569, + 553, + 602 + ], + "type": "text", + "content": "[45] Sudeep Pillai, Matthew R. Walter, and Seth Teller. Learning articulated motions from visual demonstrations. In Proceedings of Robotics: Science and Systems (RSS), 2014. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 603, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 690 + ], + "type": "text", + "content": "[46] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross B. Girshick, Piotr Dólar, and Christoph Feichtenhofer. SAM 2: Segment anything in images and videos. In Proceedings of the International Conference on Learning Representations (ICLR), 2025. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[47] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8850" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 643 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 76, + 73, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 296, + 106 + ], + "type": "text", + "content": "matching with graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4937-4946, 2020. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 149 + ], + "type": "text", + "content": "[48] Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4104-4113, 2016. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "type": "text", + "content": "[49] Chaoyue Song, Jiacheng Wei, Chuan Sheng Foo, Guosheng Lin, and Fayao Liu. REACTO: Reconstructing articulated objects from a single video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5384-5395, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 205, + 294, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 294, + 247 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 294, + 247 + ], + "type": "text", + "content": "[50] Jürgen Sturm, Cyril Stachniss, and Wolfram Burgard. A probabilistic framework for learning kinematic models of articulated objects. Journal of Artificial Intelligence Research, 41:477-526, 2011. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 249, + 294, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 249, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 56, + 249, + 294, + 312 + ], + "type": "text", + "content": "[51] Ryo Suzuki, Adnan Karim, Tian Xia, Hooman Hedayati, and Nicolai Marquardt. Augmented reality and robotics: A survey and taxonomy for ar-enhanced human-robot interaction and robotic interfaces. In Proceedings of the Chi Conference on Human Factors in Computing Systems, pages 1–33, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 315, + 294, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 315, + 294, + 389 + ], + "spans": [ + { + "bbox": [ + 56, + 315, + 294, + 389 + ], + "type": "text", + "content": "[52] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3D shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11353-11362, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 391, + 294, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 391, + 294, + 478 + ], + "spans": [ + { + "bbox": [ + 56, + 391, + 294, + 478 + ], + "type": "text", + "content": "[53] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Justin Kerr, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, David McAllister, and Angjoo Kanazawa. Nerfstudio: A modular framework for neural radiance field development. In Proceedings of the International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), pages 1-12, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 479, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 533 + ], + "type": "text", + "content": "[54] Wei-Cheng Tseng, Hung-Ju Liao, Lin Yen-Chen, and Min Sun. CLA-NeRF: Category-level articulated neural radiance field. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 8454–8460, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "type": "text", + "content": "[55] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In Advances in Neural Information Processing Systems (NeurIPS), pages 27171-27183, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 589, + 294, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 589, + 294, + 643 + ], + "spans": [ + { + "bbox": [ + 56, + 589, + 294, + 643 + ], + "type": "text", + "content": "[56] Xiaogang Wang, Bin Zhou, Yahao Shi, Xiaowu Chen, Qinping Zhao, and Kai Xu. Shape2Motion: Joint analysis of motion parts and attributes from 3D shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8868-8876, 2019. 1, 2" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 643 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[57] Xian Wang, Luyao Shen, and Lik-Hang Lee. A systematic review of XR-based remote human-robot interaction systems. ACM Computing Surveys, 57(11):1-37, 2025. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "type": "text", + "content": "[58] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 155, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 155, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 155, + 553, + 230 + ], + "type": "text", + "content": "[59] Fangyin Wei, Rohan Chabra, Lingni Ma, Christoph Lassner, Michael Zollhöfer, Szymon Rusinkiewicz, Chris Sweeney, Richard Newcombe, and Mira Slavcheva. Self-supervised neural articulated shape and appearance models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15795-15805, 2022. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 234, + 553, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 234, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 316, + 234, + 553, + 300 + ], + "type": "text", + "content": "[60] Yijia Weng, Bowen Wen, Jonathan Tremblay, Valts Blukis, Dieter Fox, Leonidas Guibas, and Stan Birchfield. Neural implicit representation for building digital twins of unknown articulated objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3141-3150, 2024. 1, 2, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 302, + 553, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 302, + 553, + 378 + ], + "spans": [ + { + "bbox": [ + 316, + 302, + 553, + 378 + ], + "type": "text", + "content": "[61] Fanbo Xiang, Yuzhe Qin, Kaichun Mo, Yikuan Xia, Hao Zhu, Fangchen Liu, Minghua Liu, Hanxiao Jiang, Yifu Yuan, He Wang, Li Yi, Angel X. Chang, Leonidas J. Guibas, and Hao Su. SAPIEN: A simulated part-based interactive environment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11094–11104, 2020. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 381, + 553, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 381, + 553, + 435 + ], + "spans": [ + { + "bbox": [ + 316, + 381, + 553, + 435 + ], + "type": "text", + "content": "[62] Zihao Yan, Ruizhen Hu, Xingguang Yan, Luanmin Chen, Oliver van Kaick, Hao Zhang, and Hui Huang. RPM-net: Recurrent prediction of motion and parts from point cloud. ACM Transactions On Graphics, 38(6):240:1-240:15, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 438, + 553, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 483 + ], + "type": "text", + "content": "[63] Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jeffrey Hu, Matthew Tancik, and Angjoo Kanazawa. Gsplat: An open-source library for Gaussian splattering, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 485, + 553, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 485, + 553, + 528 + ], + "spans": [ + { + "bbox": [ + 316, + 485, + 553, + 528 + ], + "type": "text", + "content": "[64] Li Yi, Haibin Huang, Difan Liu, Evangelos Kalogerakis, Hao Su, and Leonidas Guibas. Deep part induction from articulated object pairs. ACM Transactions on Graphics, 37(6): 1-15, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 531, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 531, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 316, + 531, + 553, + 586 + ], + "type": "text", + "content": "[65] Sergey Zakharov, Katherine Liu, Adrien Gaidon, and Rares Ambrus. ReFiNe: Recursive field networks for cross-modal multi-scene representation. In Proceedings of the International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), pages 1-9, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 588, + 555, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 588, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 316, + 588, + 555, + 643 + ], + "type": "text", + "content": "[66] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 2" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8851" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Splat-LOAM_ Gaussian Splatting LiDAR Odometry and Mapping/21f438cd-a199-46e8-b668-60b759526f59_content_list.json b/2025/Splat-LOAM_ Gaussian Splatting LiDAR Odometry and Mapping/21f438cd-a199-46e8-b668-60b759526f59_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cd762cd2e45dd707810387deeb21ffa4a2800e38 --- /dev/null +++ b/2025/Splat-LOAM_ Gaussian Splatting LiDAR Odometry and Mapping/21f438cd-a199-46e8-b668-60b759526f59_content_list.json @@ -0,0 +1,1918 @@ +[ + { + "type": "text", + "text": "Splat-LOAM: Gaussian Splitting LiDAR Odometry and Mapping", + "text_level": 1, + "bbox": [ + 163, + 130, + 833, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Emanuele Giacomini1", + "bbox": [ + 112, + 179, + 290, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Luca Di Giammarino1", + "bbox": [ + 326, + 180, + 504, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lorenzo De Rebotti1", + "bbox": [ + 542, + 180, + 710, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Giorgio Grisetti1", + "bbox": [ + 746, + 180, + 880, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Martin R. Oswald", + "bbox": [ + 421, + 198, + 575, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Sapienza University of Rome, 2University of Amsterdam, Netherlands", + "bbox": [ + 215, + 215, + 782, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "LiDARs provide accurate geometric measurements, making them valuable for ego-motion estimation and reconstruction tasks. Although its success, managing an accurate and lightweight representation of the environment still poses challenges. Both classic and NeRF-based solutions have to trade off accuracy over memory and processing times. In this work, we build on recent advancements in Gaussian Splatting methods to develop a novel LiDAR odometry and mapping pipeline that exclusively relies on Gaussian primitives for its scene representation. Leveraging spherical projection, we drive the refinement of the primitives uniquely from LiDAR measurements. Experiments show that our approach matches the current registration performance, while achieving SOTA results for mapping tasks with minimal GPU requirements. This efficiency makes it a strong candidate for further exploration and potential adoption in real-time robotics estimation tasks.", + "bbox": [ + 88, + 300, + 485, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 587, + 220, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "LiDAR sensors provide accurate spatial measurements of the environment, making them valuable for ego-motion estimation and reconstruction tasks. Since the measurements already capture the 3d structure, many LiDAR Simultaneous Localization And Mapping (SLAM) pipelines do not explicitly refine the underlying point representation [2, 3, 7, 9, 12, 37]. Moreover, a global map can be obtained by directly stacking the measurements together. However, this typically leads to extremely large point clouds that cannot be explicitly used for online applications. Several approaches attempted to use surfels [1, 31] and meshes [32]. Although these approaches manage to simultaneously estimate the sensor's ego-motion while optimizing the map representation, they result in a trade-off between accuracy, memory usage, and runtime.", + "bbox": [ + 89, + 612, + 483, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The recent advent of NeRF [26] sparked fresh interest in Novel View Synthesis (NVS) tasks. Specifically, given a set of input views and triangulated points (i.e. obtainable via COLMAP [35, 36]), NeRF learns a continuous", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d6a2cc84b189922b1b99fae9b5c855a2f6e1ee4a85e162f99fcbe63b6fb47ea6.jpg", + "image_caption": [ + "Figure 1. Performance overview of Splat-LOAM. F1 score to Active Memory (MB) and Runtime (s). The plots provide a quantitative comparison between state-of-the-art mapping pipelines, while PIN-SLAM and ours also perform online odometry." + ], + "image_footnote": [], + "bbox": [ + 524, + 265, + 705, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/970daaf16d5d54c429942fa3602d52c4af5362dc77d943901894dee7fdbd14ea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 266, + 890, + 431 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "volumetric scene function. Color and density information is propagated throughout the radiance field by ray-casting pixels from the view cameras and, through a Multi-Layer Perceptron (MLP), the method learns a representation that ensures multi-view consistency. Concurrently, NeRFs inspired the computer vision community to tackle the problem of dense visual SLAM through implicit methods. The pioneering work in this context was iMAP [40], which stores the global appearance and geometry of the scene through a single MLP. Despite its success in driving a new research direction, the method suffered from issues related to the limited capacity of the model, leading to low reconstruction quality and catastrophic forgetting during the exploration of larger areas. These issues were later handled by shifting the paradigm and by moving some of the appearance and geometric information over a hierarchical feature grid [54] or neural point clouds [22, 33]. Similarly, these techniques were employed on LiDAR measurements to provide more accurate and lighter explicit dense representations [8, 16, 30, 39, 52]. However, these methods still require tailored sampling techniques to estimate the underlying Signed Distance Function (SDF) accurately. This bottleneck still poses problems for online execution.", + "bbox": [ + 511, + 521, + 906, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A recent, explicit alternative to NeRFs is 3D Gaussian Splitting (3DGS) [20]. This approach leverages 3D", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27630", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Gaussian-shaped primitives and a differentiable, tile-based rasterizer to generate an appearance-accurate representation. Furthermore, having no need to model empty areas and no neural components, 3DGS earned a remarkable result in accuracy and training speeds. Additionally, several approaches further enhanced the reconstruction capabilities of this representation [6, 13, 15]. Being fast and accurate, this representation is now sparking interest in dense visual SLAM. Recently, 3D Gaussians were employed in several works, yielding superior results over implicit solutions [24, 47, 53].", + "bbox": [ + 89, + 90, + 480, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One issue with Gaussian Splitting relates to the primitive initialization. In regions where few or no points are provided by SFM, adaptive densification tends to fail, often yielding under-reconstructed regions. LiDAR priors are particularly useful in solving this issue due to explicit spatial measurements that can be used to initialize the local representation [14, 46].", + "bbox": [ + 89, + 258, + 480, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, to our knowledge, no attempt has been made to evaluate the this representation for LiDAR-only scenarios. This technique could prove interesting for visual NVs initialization and LiDAR mapping as it could produce a lightweight, dense, and consistent representation. These insights led us to the development of Splat-LOAM, the first LiDAR Odometry and Mapping pipeline that only leverages Gaussian primitives as its surface representation. Our system demonstrates results on par with other SOTA pipelines at a fraction of the computational demands, proving as an additional research direction for real-time perception in autonomous systems.", + "bbox": [ + 89, + 364, + 480, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 561, + 232, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Classic LiDAR Odometry and Mapping. Feature-based methods that leverage specific points or groups of points to perform incremental registration. For instance, [38, 49] track feature points on sharp edges and planar surface patches, enabling high-frequency odometry estimation. On the other hand, Direct methods leverage the whole cloud to perform registration. Specifically, these methods can be categorized based on the subjects of the alignment. Scanto-Scan methods matches subsequent clouds, either explicitly [2, 3, 37] or through neural methods [21, 44], while Scan-to-Model methods match clouds with either a local or a global map. Typically, the map is represented using points [7, 12], surfels [1, 31], meshes [32]. Another explored solution project the measurements onto a spherical projection plane to leverage visual techniques for egomotion estimation [9, 10, 28, 51].", + "bbox": [ + 89, + 592, + 482, + 834 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Implicit Methods. Concerning mapping only methods, Zhong et al. [52] proposed the first method for LiDAR implicit mapping that, given a set of point clouds and the corresponding sensor poses, used hierarchical feature grids to", + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "estimate the SDF of the scene. Their results demonstrated once again the advantages of such representation for surface reconstruction in terms of accuracy and memory footprint. A similar approach from Song et al. [39] improves the mapping accuracy by introducing SDF normal guided sampling and a hierarchical, voxel-guided sampling strategy for local optimization. Building on these advancements, several LiDAR odometry and mapping techniques were proposed. Deng et al. [8] presented the first implicit LiDAR LOAM system using an octree-based feature representation to encode the scene's SDF, used both for tracking and mapping. Similarly, Isaacson et al. [16] proposes a hierarchical feature grid to store the SDF information while using point-to-plane ICP to register new clouds. Pan et al. [30] leverages a neural point cloud representation to ensure a globally consistent estimate. The new clouds are registered using a correspondence-free point-to-implicit model approach. These methods prove that implicit representation can offer SOTA results in accuracy at the cost of potentially high execution times or memory requirements. Although targeting a different problem setting, related are also a series of visual neural SLAM methods with RGB or RGBD input [19, 22, 23, 33, 34, 47, 48, 54], see [42] for a survey.", + "bbox": [ + 511, + 90, + 903, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Gaussian Splatting. Few works tackle the use of LiDAR within the context of Gaussian Splatting. Wu et al. [46] propose a multi-modal fusion system for SLAM. Specifically, by knowing the LiDAR to camera rigid pose, the initial pose estimate is obtained through point cloud registration and further refined via photometric error minimization. In this framework, the LiDAR points are leveraged to initialize the new 3D Gaussians. In a related approach, Hong et al. [14] proposes a LiDAR-Inertial-Visual SLAM system. The initial estimate is computed through a LiDAR-Inertial odometry. Points are partitioned using size-adaptive voxels to initialize 3D Gaussians using per-voxel covariances. Both the primitives and the trajectory are further refined via photometric error minimization. While these methods introduce LiDAR measurement, 3D Gaussians are still inherently processed by cameras. Recently, Chen et al. [5] applies 3D Gaussians for the task of LiDAR NVS for re-simulation. The authors propose the use of Periodic Vibrating 3D Gaussian primitives to account for dynamic objects present in the scene. The primitives are initialized using a lightweight MLP, and the rasterization is carried out in a spherical frame by computing a per-primitive plane orthogonal to the ray that connects the primitive's mean to the LiDAR frame, thus removing any distortion in the projection process. Focusing on a different formulation, Jiang et al. [17] propose a method of LiDAR NVS that leverages Periodic Vibrating 2D Gaussian primitives. The primitives are initialized by randomly sampling points and, further optimized using the losses described in [15], along with a Chamfer loss is introduced to constrain the 3D structures of the synthesized point", + "bbox": [ + 511, + 446, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27631", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8983ddc263f458559d96d6758114177ef7ef8023e5854e9900551b79d10cb212.jpg", + "image_caption": [ + "Figure 2. Splat-LOAM Overview. Given a LiDAR point cloud, we leverage the spherical projection to generate an image-like representation. Moreover, using an ad-hoc differentiable rasterizer, we guide the optimization for structural parameters of 2D Gaussians. The underlying representation is concurrently used to incrementally register new measurements." + ], + "image_footnote": [], + "bbox": [ + 99, + 85, + 906, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "clouds, and an additional ray-drop term to account for phenomena like non-returning laser pulses. This term is further refined through a U-Net that considers other factors, such as the distance of the surface from the sensor. Compared to this work, we provide a thorough methodology to render 2D Gaussians on spherical frames while accounting for coordinates singularity and a cloud registration technique to allow for simultaneous odometry and mapping. To our knowledge, Splat-LOAM is the first pure LiDAR Odometry and Mapping pipeline that leverages Gaussian primitives both for mapping and tracking. In sum, our contributions are", + "bbox": [ + 88, + 371, + 480, + 537 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A differentiable, tile-based rasterizer for 2D Gaussians over spherical manifolds.", + "- A mapping pipeline that allows the integration of LiDAR measurements in a 2D Gaussian representation.", + "- A tracking schema that leverages both 3D and 2D representations to register new measurements and estimate the sensor ego-motion." + ], + "bbox": [ + 89, + 537, + 483, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 659, + 181, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section introduces our novel LiDAR odometry and mapping method based on 2D Gaussian primitives. We detail a mapping strategy for initializing, refining, and integrating these primitives alongside a registration method that leverages geometric and photometric cues from the continuous local model for ego-motion estimation. Additional details can be found in the Supplementary material.", + "bbox": [ + 89, + 685, + 483, + 791 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Spherical Projection Model", + "text_level": 1, + "bbox": [ + 89, + 801, + 336, + 818 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While original Gaussian Splatting leverages pinhole-camera projection to render or refine 3D primitives, Li-DAR input provides $360^{\\circ}$ panoramic input. To this end, we employ spherical projection to encode LiDAR measurements into an image-like representation that we use to guide", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the Gaussians optimization. A projection is a mapping $\\phi : \\mathbb{R}^3 \\to \\Gamma \\subset \\mathbb{R}^2$ from a world point $\\mathbf{p} = (x, y, z)^T$ to image coordinates $\\mathbf{u} = (u, v)^T$ . Knowing the range $d = \\| \\mathbf{p} \\|$ of an image point $\\mathbf{u}$ , we can calculate the inverse mapping $\\phi^{-1} : \\Gamma \\times \\mathbb{R} \\to \\mathbb{R}^3$ , more explicitly $\\mathbf{p} = \\phi^{-1}(\\mathbf{u}, d)$ . We refer to this operation as back-projection. To ease the clarity of this work, it is worth mentioning that, compared to the classical pinhole camera, the optical reference frame is rearranged; the $x$ -axis points forward, $y$ -axis points to the left, and $z$ -axis points upwards. Let $\\mathbf{K}$ be a camera intrinsic matrix that can be computed directly from the point cloud (see Supplementary Material), with function $\\psi$ mapping a 3D point to azimuth and elevation. Thus, the spherical projection of a point is given by", + "bbox": [ + 511, + 371, + 906, + 583 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\phi (\\mathbf {p}) = \\mathbf {K} \\psi (\\mathbf {p}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 595, + 903, + 612 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\psi (\\mathbf {v}) = \\left[ \\begin{array}{c} \\operatorname {a t a n 2} \\left(v _ {y}, v _ {x}\\right) \\\\ \\operatorname {a t a n 2} \\left(v _ {z}, \\sqrt {v _ {x} ^ {2} + v _ {y} ^ {2}}\\right) \\\\ 1 \\end{array} \\right]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 614, + 903, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We used spherical projection to obtain a range map $\\hat{\\mathbf{D}}$ of the LiDAR point cloud $\\{\\mathbf{q}_p\\}_{p=1}^Q$ of size $Q$ .", + "bbox": [ + 511, + 683, + 906, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. 2D Gaussian Splitting", + "text_level": 1, + "bbox": [ + 511, + 724, + 723, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our scene representation is based on 2D Gaussians due to their affinity to surface reconstruction [6, 15]. A 2D Gaussian is defined by its opacity $o \\in [0,1]$ , centroid $\\pmb{\\mu} \\in \\mathbb{R}^3$ , origin plane defined by two tangential vectors $\\mathbf{t}_{\\alpha} \\in \\mathbb{R}^3$ and $\\mathbf{t}_{\\beta} \\in \\mathbb{R}^3$ , and scaling vector $\\mathbf{s} = (s_{\\alpha}, s_{\\beta}) \\in \\mathbb{R}^2$ that controls the variance of the Gaussian kernel [15]. Points on the splan's plane can be described in world space using:", + "bbox": [ + 511, + 748, + 906, + 854 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} = \\left( \\begin{array}{c c c c} s _ {\\alpha} \\mathbf {t} _ {\\alpha} & s _ {\\beta} \\mathbf {t} _ {\\beta} & \\mathbf {0} & \\boldsymbol {\\mu} \\\\ 0 & 0 & 0 & 1 \\end{array} \\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 864, + 906, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27632", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "while the Gaussian kernel can be evaluated on a point $(\\alpha, \\beta)$ in splat space using:", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} (\\alpha , \\beta) = \\exp \\left(- \\frac {\\alpha^ {2} + \\beta^ {2}}{2}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 133, + 482, + 169 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Rasterization", + "text_level": 1, + "bbox": [ + 89, + 180, + 230, + 194 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We render the Gaussians via $\\alpha$ -blending as in [15, 20]. Specifically, for each pixel $\\mathbf{u}$ , we integrate Gaussians from front to back to obtain the range $d$ , normal $\\mathbf{n}$ , and opacity $o$ as follows:", + "bbox": [ + 89, + 199, + 483, + 258 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd = \\sum_ {i = 1} ^ {T} o _ {i} \\mathcal {G} _ {i} d _ {i} \\prod_ {j = 1} ^ {i - 1} \\left(1 - o _ {j} \\mathcal {G} _ {j}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 270, + 482, + 314 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {n} = \\sum_ {i = 1} ^ {T} o _ {i} \\mathcal {G} _ {i} \\mathbf {t} _ {n _ {i}} \\prod_ {j = 1} ^ {i - 1} \\left(1 - o _ {j} \\mathcal {G} _ {j}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 316, + 482, + 359 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\no = \\sum_ {i = 1} ^ {T} o _ {i} \\mathcal {G} _ {i} \\prod_ {j = 1} ^ {i - 1} \\left(1 - o _ {j} \\mathcal {G} _ {j}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 362, + 482, + 405 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Different from the pinhole model, the spherical projection model described in Sec. 3.1 is non-affine, resulting in strong distortion artifacts on Gaussians that cover a significant extent in the image plane, if rasterizing using the local affine approximation method proposed in [20].", + "bbox": [ + 89, + 417, + 483, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To this extent, as in [15], our rasterizer leverages exact ray-splat intersections to render 2D Gaussians. This approach is particularly suited for our projection model as it explicitly relies on per-pixel directions to evaluate the Gaussian kernels, thus allowing us to account for the structure of our projection manifold surface.", + "bbox": [ + 89, + 494, + 483, + 584 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The ray-splat intersection is based on [45], and differently from [15], we reformulate the ray parametrization to cover the whole azimuthal space around the sensor. Additionally, we propose a method to compute the image-space extent of each Gaussian that is robust to the azimuthal coordinate singularity, ensuring accurate rasterization for any Gaussian configuration.", + "bbox": [ + 89, + 585, + 483, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. Ray-splat Intersection", + "text_level": 1, + "bbox": [ + 89, + 700, + 290, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $\\mathbf{v} = \\phi^{-1}\\left(\\mathbf{K}^{-1}\\mathbf{u}\\right)$ be the normalized direction of pixel $\\mathbf{u}$ , parametrized as the intersection of two orthogonal planes", + "bbox": [ + 89, + 719, + 483, + 752 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} _ {x} = \\frac {\\mathbf {v} \\times \\mathbf {u} _ {z}}{\\left\\| \\mathbf {v} \\times \\mathbf {u} _ {z} \\right\\|}, \\quad \\mathbf {h} _ {y} = \\mathbf {h} _ {x} \\times \\mathbf {v}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 763, + 482, + 796 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{u}_z$ is the unit $z$ -direction vector. This formulation is valid when $\\mathbf{v} \\neq \\mathbf{u}_z$ which is a fair assumption in LiDAR measurements. Moreover, the planes are described in splat's space as", + "bbox": [ + 89, + 806, + 483, + 867 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} _ {\\alpha} = \\left(\\mathbf {T} _ {w} ^ {c} \\mathbf {H}\\right) ^ {T} \\mathbf {h} _ {x}, \\quad \\mathbf {h} _ {\\beta} = \\left(\\mathbf {T} _ {w} ^ {c} \\mathbf {H}\\right) ^ {T} \\mathbf {h} _ {y}, \\quad (9)\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 877, + 483, + 898 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c949944ef033cc573c706a2bbb36afb53fc94b0fbfdd94b3d4e2b17e168fec96.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 87, + 668, + 242 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cb55ea3b6bbc2d40e6e25a29a56faede3fab92dd57de1c208429805fe24bf5f9.jpg", + "image_caption": [ + "Figure 3. Bounding box computation for near-singularity splats. (a) shows the 3D configuration of a splat that approximately lies behind the camera. (b) shows the corresponding spherical image with the projected bounding box vertices. The distortion is removed by shifting the vertices along the horizontal direction to align the projected center to the image center. (c) being far from the coordinate singularity, we compute the maximum extent of the splat. (d) We revert the shift and propagate to the corresponding tiles via an offset from the central vertex, matching with the tiles highlighted on (a)." + ], + "image_footnote": [], + "bbox": [ + 676, + 88, + 906, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{T}_w^k\\in \\mathbb{S}\\mathbb{E}(3)$ describes the world frame in the $k$ -th sensor reference frame. We express the intersection of the two planes bundle as:", + "bbox": [ + 511, + 431, + 905, + 477 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} _ {\\alpha} (\\alpha , \\beta , 1, 1) ^ {T} = \\mathbf {h} _ {\\beta} (\\alpha , \\beta , 1, 1) ^ {T} = 0. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 486, + 905, + 507 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The solution $\\alpha$ and $\\beta$ is then computed by solving the homogeneous linear system:", + "bbox": [ + 511, + 516, + 905, + 545 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha \\left(\\mathbf {u} _ {s}\\right) = \\frac {\\mathbf {h} _ {\\alpha} ^ {2} \\mathbf {h} _ {\\beta} ^ {4} - \\mathbf {h} _ {\\alpha} ^ {4} \\mathbf {h} _ {\\beta} ^ {2}}{\\mathbf {h} _ {\\alpha} ^ {1} \\mathbf {h} _ {\\beta} ^ {2} - \\mathbf {h} _ {\\alpha} ^ {2} \\mathbf {h} _ {\\beta} ^ {1}}, \\quad \\beta \\left(\\mathbf {u} _ {s}\\right) = \\frac {\\mathbf {h} _ {\\alpha} ^ {4} \\mathbf {h} _ {\\beta} ^ {1} - \\mathbf {h} _ {\\alpha} ^ {1} \\mathbf {h} _ {\\beta} ^ {4}}{\\mathbf {h} _ {\\alpha} ^ {1} \\mathbf {h} _ {\\beta} ^ {2} - \\mathbf {h} _ {\\alpha} ^ {2} \\mathbf {h} _ {\\beta} ^ {1}}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 551, + 905, + 588 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.3. Bounding Box Computation", + "text_level": 1, + "bbox": [ + 511, + 595, + 756, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To efficiently rasterize Gaussians, each tile is informed of which Gaussians it should process. This is achieved by computing the per-Gaussian image-space extent from its projected central point and using it to inform each covered tile to rasterize the Gaussian. In our work, we compute the per-Gaussian extent by projecting the vertices of the $3\\sigma$ splat-space bounding box on the spherical frame. Due to the horizontal periodic nature of the spherical frame, for Gaussians whose center is projected near the horizontal image boundaries and has a significant image-space size, one or more vertices of the bounding box may be projected on the opposite side of the image. If not handled, this situation can degrade the rasterization since not all the corresponding tiles may be informed, thus skipping rendering for some pixels and slowing the rendering process since the computed extent might almost cover the whole image plane.", + "bbox": [ + 511, + 613, + 906, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We solve this by computing the Gaussian extent and tile propagation in a condition where the coordinate singularity cannot be harmful. As shown in Fig. 3, for each Gaussian,", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27633", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we project its bounding box and shift its projected points towards the center of the image plane while accounting for coordinate singularity. Similarly, after computing the Gaussian radius, the information is propagated to neighboring tiles while accounting for periodicity, ensuring correct rasterization and reducing the number of selected tiles.", + "bbox": [ + 89, + 90, + 483, + 183 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Odometry And Mapping", + "text_level": 1, + "bbox": [ + 89, + 189, + 318, + 207 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following the modern literature of RGB-D SLAM [22, 34, 47, 53], we rely on keyframing to optimize local maps. We choose this approach for the following advantages: first, continuous integration over the same model can have adverse effects on artifacts and, more importantly, on runtime. Instead of decreasing the local density, we reset to a new local model if certain conditions are met, while also restricting the number of frames joining the optimization stage to allow effective online processing. Based on this, we define each local map as an individual Gaussian model $\\mathbf{P}^s$ as", + "bbox": [ + 89, + 210, + 483, + 363 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} ^ {s} = \\left\\{\\mathcal {G} (\\boldsymbol {\\mu}, \\boldsymbol {\\Sigma}, o) | i = 1, \\dots , N \\right\\}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 373, + 482, + 392 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.1. Local model initialization", + "text_level": 1, + "bbox": [ + 89, + 398, + 316, + 414 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We initialize a new local model using the input LiDAR point cloud whenever necessary, such as at system startup or when visibility conditions require it. As a first step, we generate a valid pixel mask via the indicator function $\\mathbb{1}[\\cdot ]$ as", + "bbox": [ + 89, + 417, + 483, + 481 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {v} = \\mathbb {1} [ \\hat {\\mathbf {D}} > 0 ], \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 489, + 482, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and compute the range gradients $\\nabla_{\\mathbf{D}}$ to construct a weight map over the range image. We use a weighted sampling of $n_d$ points to prioritize complex regions. Splat positions are computed by back-projecting the range image, while their orientations are initialized by directing surface normals toward the sensor center to enhance initial visibility.", + "bbox": [ + 89, + 516, + 483, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2. Local Model Refinement", + "text_level": 1, + "bbox": [ + 89, + 614, + 308, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We perform a limited number of refinement iterations $n_o$ on the most recent keyframes. Unlike [47, 53], we employ a geometric distribution-based sampling scheme that guarantees at least $40\\%$ probability of selecting the most recent keyframe and progressively decreases the likelihood of choosing the older ones. To filter out artifacts caused by ray-drop phenomena in the LiDAR measurements, we only consider valid pixels to refine the local model parameters $\\mathbf{x}$ . We start by applying a densification strategy similar to the one described in Sec. 3.3.1, with two extra terms. The first one, $\\mathbf{M}_n = \\mathbb{1}[\\mathbf{O}_i \\leq \\lambda_{d,o}]$ , target newly discovered areas while the second, $\\mathbf{M}_e = \\mathbb{1}[|\\mathbf{D}_i - \\hat{\\mathbf{D}}_i| \\geq \\lambda_{d,e}]$ , target under-reconstructed regions.", + "bbox": [ + 89, + 633, + 482, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To optimize the geometric consistency of the local model, we employ a loss term that minimizes the $L_{1}$ error:", + "bbox": [ + 89, + 830, + 482, + 862 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d} = \\sum_ {\\mathbf {u} \\in \\mathbf {M} _ {v}} \\rho_ {d} \\| \\mathbf {D} (\\mathbf {u}, \\mathbf {x}) - \\hat {\\mathbf {D}} (\\mathbf {u}) \\|, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 869, + 482, + 904 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\rho_{d}$ is a weight function dependent on the measurement's range. In addition, we employ a self-regularization term to align the splat's normals to the surface normals $\\mathbf{N}$ estimated by the gradients of the range map $\\mathbf{D}$ [15]:", + "bbox": [ + 511, + 90, + 905, + 152 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {n} = \\sum_ {\\mathbf {u} \\in \\mathbf {M} _ {v}} 1 - \\mathbf {n} ^ {T} \\mathbf {N} (\\mathbf {D} (\\mathbf {u}, \\mathbf {x})) \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 160, + 906, + 193 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Furthermore, to promote the expansion of splats over uniform surfaces, we introduce an additional term that operates on the opacity channel of the rasterized images. Specifically, we drive the splats to cover the areas of the image containing valid measurements by correlating the opacity image $\\mathbf{O}$ with the valid mask $\\mathbf{M}_v$ .", + "bbox": [ + 511, + 202, + 905, + 292 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {o} = \\sum_ {\\mathbf {u} \\in \\mathbf {M} _ {v}} - \\log (\\mathbf {O} (\\mathbf {u}, \\mathbf {x})). \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 301, + 906, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While $\\mathcal{L}_o$ allows the splats to grow, it can also cause some splats to become extremely large in unobserved areas. We employ an additional regularization term on the scaling parameter of all primitives $N$ to compensate for this effect. We propose a novel regularization that allows the splat to extend up to a certain value $\\tau_s$ before penalizing its growth:", + "bbox": [ + 511, + 343, + 905, + 429 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} = \\left\\{ \\begin{array}{c l} \\tau_ {s} - \\max \\left(s _ {\\alpha}, s _ {\\beta}\\right) & \\text {i f} \\max \\left(s _ {\\alpha}, s _ {\\beta}\\right) > \\tau_ {s} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 435, + 906, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We found this term to be more effective rather than directly minimizing the deviation from the average [24, 47, 53] as it allows for anisotropic splats that are particularly useful for mapping details and edges, and provides more control on the density and structure of the model.", + "bbox": [ + 511, + 472, + 905, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final mapping objective function is defined as:", + "bbox": [ + 532, + 547, + 867, + 564 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a p}} = \\mathcal {L} _ {d} + \\lambda_ {o} \\mathcal {L} _ {o} + \\lambda_ {n} \\mathcal {L} _ {n} + \\lambda_ {s} \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {s _ {i}}, \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 573, + 905, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with $\\lambda_{o}, \\lambda_{n}, \\lambda_{s} \\in \\mathbb{R}$ being loss weights. This update strategy allows our method to maintain a geometrically accurate scene representation, while selectively densifying under-reconstructed regions that require more resolution, using a limited number of iterations. To avoid catastrophic forgetting, we do not perform any opacity reset steps. Moreover, we prune Gaussians that are either transparent or very small, as they typically cannot be observed during rendering.", + "bbox": [ + 511, + 622, + 905, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.3. Frame-To-Model Registration", + "text_level": 1, + "bbox": [ + 511, + 750, + 769, + 766 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Each time a new keyframe is selected, we sample the local model by back-projecting the rasterized range image $\\mathbf{D}$ onto a point cloud $\\{\\mathbf{p}_q\\}_{q=1}^{W \\times H}$ at its estimated pose. We design an ad-hoc tracking schema to benefit from both geometric and photometric consistencies provided by the Li-DAR and the rendered local model. Hence, the total odometry loss is composed of the sum of both residuals:", + "bbox": [ + 511, + 768, + 905, + 875 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {o d o m}} = \\mathcal {L} _ {\\mathrm {g e o}} + \\mathcal {L} _ {\\mathrm {p h o t o}}. \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 885, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27634", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Geometric Registration. To associate geometric entities, we employ a PCA-based kd-tree [12]. The kd-tree is built on the back-projected rendered range map $\\{\\mathbf{p}_q\\}_{q=1}^{W \\times H}$ and segmented into tree leaves, corresponding to planar patches. Each leaf corresponds to $l = \\langle \\mathbf{p}_l, \\mathbf{n}_l \\rangle$ , that is, the mean point $\\mathbf{p}_l \\in \\mathbb{R}^3$ and the surface normal $\\mathbf{n}_l \\in \\mathbb{R}^3$ . The geometric loss $\\mathcal{L}_{\\mathrm{geo}}$ represents the sum of the point-to-plane distance between the mean leaf $\\mathbf{p}_l$ and the point of the current measurement point cloud $k$ with $\\{\\mathbf{q}_p\\}_{p=1}^Q$ , along the normal $\\mathbf{n}_l$ expressed in the local reference frame $\\mathbf{T}_w^k$ . Specifically, we have:", + "bbox": [ + 89, + 90, + 483, + 258 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {g e o}} = \\sum_ {p, q \\in \\{a \\}} \\rho \\left(\\left(\\mathbf {T} _ {w} ^ {k} \\mathbf {n} _ {l _ {q}}\\right) ^ {T} \\left(\\mathbf {T} _ {w} ^ {k} \\mathbf {q} _ {p} - \\mathbf {p} _ {q _ {i}}\\right)\\right), \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 266, + 483, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $w$ is the global frame, $k$ is the local frame, $\\{a\\}$ is the set of leaf associations with the point from the measurement point cloud, and $\\rho$ is the Huber robust loss function.", + "bbox": [ + 89, + 306, + 483, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Photometric Registration. Leveraging the rendered range map $\\mathbf{D}$ , we employ photometric registration for subpixel refinement, minimizing the photometric distance between the rendered and the spherical projected query point cloud $\\hat{\\mathbf{D}}$ . The photometric loss is formulated as:", + "bbox": [ + 89, + 372, + 483, + 445 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {p h o t o}} = \\sum_ {\\mathbf {u}} \\left\\| \\rho \\left(\\mathbf {D} (\\mathbf {u}) - \\hat {\\mathbf {D}} \\underbrace {\\left(\\phi \\left(\\mathbf {T} _ {w} ^ {k} \\phi^ {- 1} (\\mathbf {u} , \\hat {d})\\right)\\right)} _ {\\mathbf {u} ^ {\\prime}}\\right) \\right\\| ^ {2}. \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 450, + 482, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The evaluation point $\\mathbf{u}'$ in the query image $\\hat{\\mathbf{D}}$ is computed by first back-projecting the pixel $\\mathbf{u}$ , applying the transform $\\mathbf{T}_w^k$ , and then projecting it back. Pose updates $\\delta$ are parameterized as local updates in the Lie algebra $\\mathfrak{se}(3)$ . Therefore, the transformation $\\mathbf{T}_k^w$ of the local reference frame, expressed in the global reference frame, is updated as:", + "bbox": [ + 89, + 498, + 483, + 589 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} _ {k} ^ {w} \\leftarrow \\mathbf {T} _ {k} ^ {w} \\exp (\\delta). \\tag {22}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 597, + 482, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This update is carried out using a second-order GaussNewton method. Local updates ensure that rotation updates are well-defined [11].", + "bbox": [ + 89, + 622, + 483, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 679, + 223, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we report the results of our method on several publicly available datasets. We evaluate our pipeline on tracking and mapping. We recall that, to our knowledge, this is the first Gaussian Splatting LiDAR Odometry and Mapping pipeline, and no direct competitors are available. We compared our approach with other well-known geometric and neural implicit methods. The experiments were executed on a PC with an Intel Core i9-14900K @ 3.20Ghz, 64GB of RAM, and an NVIDIA RTX 4090 GPU with 24 GB of VRAM.", + "bbox": [ + 89, + 704, + 482, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For odometry evaluation, we include several baselines. The first one is a basic point-to-plane ICP odometry, followed by SLAMesh, that simultaneously estimates a mesh", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "representation of the scene via Gaussian Processes and performs registration onto it [32]. Moreover, MAD-ICP leverages a forest of PCA-based KD-Trees to perform accurate registration. Furthermore, we include PIN-SLAM as SOTA baseline for implicit LiDAR SLAM [30]. It relies on neural point primitives and interleaves an incremental learning of the model's SDF and a correspondence-free, point-to-implicit registration schema.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For mapping evaluation, we include OpenVDB [27] and VoxBlox [29] as \"classic\" baselines. OpenVDB provides a robust volumetric data structure to handle 3D Points, while VoxBlox combines adaptive weights and grouped ray-casting for an efficient and accurate Truncated Signed Distance Function (TSDF) integration. Additionally, we include as neural-implicit approaches, $N^3$ -Mapping [39] and PIN-SLAM [30]. $N^3$ -Mapping is a neural-implicit non-projective SDF mapping approach that leverages normal guidance to produce more accurate SDFs, leading to SOTA results for offline LiDAR mapping. Moreover, PIN-SLAM is included as meshes can be extracted directly from the implicit representation via marching cubes. Below, we report the datasets and the evaluation metrics employed. We do not include SLAMesh [32] since it could not be run with ground-truth poses.", + "bbox": [ + 511, + 212, + 906, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 513, + 467, + 614, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We used the following four publicly available datasets:", + "bbox": [ + 513, + 489, + 877, + 503 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Newer College Dataset (NC) [50]: Collected with a handheld Ouster OS0-128 LiDAR in structured and vegetated areas. Ground truth was generated using the Leica BLK360 scanner, achieving centimeter-level accuracy over poses and points in the map.", + "- A Vision Benchmark in Rome (VBR) [4]: Recorded in Rome using OS1-64 (car-mounted) and OS0-128 (hand-held) LiDARs, capturing large-scale urban scenarios with narrow streets and dynamic objects. Ground truth trajectories were obtained by fusing LiDAR, IMU, and RTK" + ], + "bbox": [ + 514, + 505, + 905, + 661 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fca29fcdfcbbabff88fd491a1ed51f2d78802e34d1760e18bd3389b1aa213bc2.jpg", + "image_caption": [ + "Figure 4. RPE evaluation. Number of successful sequences across RPE thresholds. It includes the sequences of Newer College [49], VBR [4], Oxford Spires [41] and Mai City [43]." + ], + "image_footnote": [], + "bbox": [ + 519, + 689, + 898, + 845 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27635", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/817655b460b2639ad0f1bc9a8fe87a4850f7c9929e1d4b333d56c6fa8d689f46.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetNewer College[50]Oxford Spires[41]Mai-City[43]
quad-easymath-easykeble-02bodleian-02observatory-01mai-01mai-02
ApproachC-11↓F-score↑C-11↓F-score↑C-11↓F-score↑C-11↓F-score↑C-11↓F-score↑C-11↓F-score↑C-11↓F-score↑
OpenVDB[27]7.9288.8511.6984.517.1991.747.5189.689.5986.163.3397.293.2697.37
VoxBlox[29]16.564.6311.9380.5115.0371.6315.2458.7715.1270.457.0493.815.8095.51
N3-Mapping[39]8.0494.54failfail7.0193.477.8990.369.3587.942.6499.062.7699.04
PIN-SLAM[30]12.8988.0514.6973.8211.8379.6510.7482.7114.4972.314.9193.844.8294.11
Ours5.3796.749.1590.027.4394.417.6090.0910.5683.043.6497.275.0595.31
", + "bbox": [ + 135, + 88, + 864, + 212 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Reconstruction evaluation. The pipelines were run with ground-truth poses. Voxel size is ${20}\\mathrm{\\;{cm}}$ and F-score is computed with ${20}\\mathrm{\\;{cm}}$ threshold. Splat-LOAM yield competitive mapping performance on both the Newer College[50] and Oxford Spires[41] dataset.", + "bbox": [ + 89, + 217, + 906, + 246 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "GNSS data, ensuring centimeter-level accuracy over the poses. This dataset is not used for mapping evaluation as no ground-truth maps are provided.", + "bbox": [ + 102, + 272, + 482, + 316 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Oxford Spires [41]: Recorded with a hand-held Hesai QT64 LiDAR featuring a $360^{\\circ}$ horizontal FoV, $104^{\\circ}$ vertical FoV, 64 vertical channels, and 60 meters of range detection. Similar to [49], each sequence includes a prior map obtained via a survey-grade 3D imaging laser scanner, used for ground-truth trajectory estimation and mapping evaluation. Specifically, we choose the Keble College, Bodleian Library, and Radcliffe Observatory sequences to include both indoor and outdoor scenarios with different levels of detail.", + "- Mai City [43]: A synthetic dataset captured using a car-like simulated LiDAR with 120 meters of range detection. The measurements are generated via ray-casting on an underlying mesh, providing error-free, motion-undistorted data. We select the $01$ and $02$ sequences, which capture similar scenarios with different vertical resolutions." + ], + "bbox": [ + 89, + 321, + 482, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Evaluation", + "text_level": 1, + "bbox": [ + 89, + 577, + 210, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use Relative Pose Error (RPE) computed with progressively increasing delta steps to evaluate the odometry accuracy. Specifically, we adapt the deltas to the trajectory length to provide a more meaningful result [4].", + "bbox": [ + 89, + 599, + 482, + 659 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Differently, to evaluate mapping quality, we use several metrics [25]: Accuracy (Acc) measures the mean distance of points on the estimated mesh with their nearest neighbors on the reference cloud. Completeness (Comp) measures the opposite distance, and Chamfer-11 (C-11) describes the mean of the two. Additionally, we use the F-score computed with $20\\mathrm{cm}$ error threshold. Table 1 reports compact results for C-11 and F1-score, additional tables are available in the Supplementary Material.", + "bbox": [ + 89, + 660, + 482, + 795 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 803, + 243, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our approach employs Gaussian primitives for LiDAR odometry estimation and mapping, yielding results comparable to state-of-the-art methods while significantly enhancing computational efficiency. In this section, we evaluate some key aspects of our pipeline and evaluate their", + "bbox": [ + 89, + 825, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "contributions.", + "bbox": [ + 513, + 273, + 607, + 285 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Memory and Runtime Analysis. In Fig. 5, we report how the increment of active primitives affects the active GPU memory requirements and the per-iteration mapping frequency. It shows an experiment run over the large-scale campus sequence [4] where we set a maximum of 100 keyframes per local map and sample, at most, $50\\%$ of points for the incoming point cloud. It's possible to notice that the mapping FPS remains stable between 200k and 300k primitives. This is most likely related to the saturation of the rasterizer.", + "bbox": [ + 511, + 306, + 906, + 457 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6f1dd620c0146ede4fd189a3037f4501e87349e6880d43a9e754bfece47c38e4.jpg", + "image_caption": [ + "Figure 5. Memory and Runtime Analysis. The plot relates the used GPU Memory and the mapping iteration frequency with the number of active primitives. The measurements were obtained over the longest sequence we reported: campus [4]." + ], + "image_footnote": [], + "bbox": [ + 542, + 473, + 877, + 623 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Odometry. Fig. 7 shows the results for different tracking methods over our scene representation. Using both geometric and photometric components, we achieve a better result than using point-to-point or point-to-plane. The last three bars show an ablation of geometric and photometric loss components. The results are best for the joint use of both terms which support our design choice.", + "bbox": [ + 511, + 714, + 905, + 820 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mesh generation. We can sample the Gaussian scene to produce a mesh representation. Similar to [13], the meshing process involves sampling $n_m$ points from each keyframe rendered range and normal maps, and running Poisson reconstruction [18]. As shown in Tab. 2, we found", + "bbox": [ + 511, + 825, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27636", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a47a6ccec38933cc3ff53dd986481cd0c763ed9714a02013cba541374b10ec3f.jpg", + "image_caption": [ + "PIN-SLAM" + ], + "image_footnote": [], + "bbox": [ + 93, + 104, + 292, + 193 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1f8e480cb538ed36666c8be4e68f4247029873dbaba7e73832a1020c7f064854.jpg", + "image_caption": [ + "$\\mathbf{N}^3$ -Mapping" + ], + "image_footnote": [], + "bbox": [ + 295, + 104, + 496, + 193 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c3e7e666ea73a21b76ea9b7c989bce04d8c19abf7ae21be458852c99ade76367.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 501, + 104, + 700, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/93fefcc024af56629cf061fb2c87a54030a581e8c2e3fcb1be70387a71eb2862.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 704, + 104, + 903, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/06cff9a1acd60fe66d5720707149a8fbaf1393ce209f470ef8707be1d508de2d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 194, + 292, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4b1389376e896f0a53bf765ed5595b392257f0be201800b3a3fc4e56c94aa456.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 194, + 496, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c0ecd9bf88726ed22e365266e2089e86efd5dfe471faa64b095a5a0d911497f1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 194, + 700, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/19bd142339bf312ce8748194350edeabee2683bdf20337e59f28a84c62cddbf0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 194, + 903, + 281 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0c39cbab10d5fb386e1d75c9028f56ad31a424c01935b8d4eb0be133c52cbae9.jpg", + "image_caption": [ + "Figure 6. Comparison of Mesh Reconstruction. The figure shows reconstruction results for quad-easy sequence from the newer college dataset. Our method recovers a geometry with much higher data fidelity. PIN-SLAM lacks many details and exhibits a large level of noise. $N^3$ -Mapping performs more similar to ours, but oversmoothes fine geometric details." + ], + "image_footnote": [], + "bbox": [ + 91, + 284, + 292, + 371 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/205ad13a24b0e70972c290a0bc262b26c492a9c4d6330a034615c5b4c15cfbe1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 282, + 496, + 371 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d9d7c9816b7d2a45a1125af51c52fa18af51828882f975071df7da6d812f965a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 282, + 700, + 371 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e1bf2831635340691d4a5b09610f706800ef7ba227b4d3d2ad5ed628513cbd10.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 282, + 903, + 371 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ed8fbcb19e771faa56bf43822a3a2a2e891feab34d32f99e001066088a5137de.jpg", + "image_caption": [ + "Figure 7. Ablation on registration methods. The plot reports the RPE (\\%) of several tested registration methods on the quad-easy sequence [50]. Enabling both geometric and photometric factors in sequence, provides a more robust estimate." + ], + "image_footnote": [], + "bbox": [ + 99, + 446, + 475, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "that this method produces consistently better results than directly running TSDF integration and marching cubes over the rendered frames, or by running Poisson reconstruction directly on the Gaussian centers. Specifically, these method produces worse results in the regions that overlap multiple local models.", + "bbox": [ + 89, + 705, + 483, + 796 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 813, + 209, + 829 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present the first LiDAR Odometry and Mapping pipeline that leverages 2D Gaussian primitives as the sole scene representation. Through an ad-hoc tile-based Gaussian rasterizer for spherical images, we leverage LiDAR", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b2c2d044c03949fb10cf59629b9a4a44912cb0cad8a52ab25ceb27707c352d69.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Extraction MethodAcc↓Com↓C-11↓F-score↑
Marching Cubes [27]16.765.5311.1476.76
Poisson (centers)10.156.708.4392.33
Ours6.644.095.3796.74
", + "bbox": [ + 517, + 441, + 901, + 506 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Ablation on Meshing Methods. We report mapping results with varying meshing methods on the quad-easy sequence [50].", + "bbox": [ + 511, + 510, + 906, + 554 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "measurements to optimize the local model. Furthermore, we demonstrate the effectiveness of combining a geometric and photometric tracker to register new LiDAR point clouds over the Gaussian local model. The experiments show that our pipeline obtains tracking and mapping scores that are on par with the current SOTA at a fraction of the computational demands.", + "bbox": [ + 511, + 594, + 906, + 699 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Future Work. We plan on improving Splat-LOAM by simultaneously estimating the sensor pose and velocity to compensate the rolling-shutter distortions caused by motion while acquisition, typical of the LiDAR measurements. Moreover, we plan on including Loop Closure (LC) to improve the pose estimates, along with the mapping accuracy.", + "bbox": [ + 511, + 710, + 905, + 804 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 513, + 837, + 671, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work has been supported by PNRR MUR project PE0000013-FAIR.", + "bbox": [ + 511, + 869, + 872, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27637", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jens Behley and Cyril Stachniss. Efficient Surfel-Based SLAM using 3D Laser Range Data in Urban Environments. In Proc. of Robotics: Science and Systems (RSS). Robotics: Science and Systems Foundation, 2018. 1, 2", + "[2] P.J. Besl and Neil D. McKay. A method for registration of 3-D shapes. IEEE TPAMI, 14(2):239-256, 1992. 1, 2", + "[3] Dorit Borrmann, Jan Elseberg, Kai Lingemann, Andreas Nüchter, and Joachim Hertzberg. Globally consistent 3D mapping with scan matching. Journal on Robotics and Autonomous Systems (RAS), 56(2):130-142, 2008. 1, 2", + "[4] Leonardo Brizi, Emanuele Giacomini, Luca Di Giammarino, Simone Ferrari, Omar Salem, Lorenzo De Rebotti, and Giorgio Grisetti. VBR: A Vision Benchmark in Rome. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), pages 15868-15874, 2024. 6, 7", + "[5] Qifeng Chen, Sheng Yang, Sicong Du, Tao Tang, Peng Chen, and Yuchi Huo. LiDAR-GS:Real-time LiDAR Re-Simulation using Gaussian Splitting, 2024. 2", + "[6] Pinxuan Dai, Jiamin Xu, Wenxiang Xie, Xinguo Liu, Huamin Wang, and Weiwei Xu. High-quality Surface Reconstruction using Gaussian Surfels. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, New York, NY, USA, 2024. Association for Computing Machinery. 2, 3", + "[7] Pierre Dellenbach, Jean-Emmanuel Deschaud, Bastien Jacquet, and François Goulette. CT-ICP: Real-time Elastic LiDAR Odometry with Loop Closure. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), pages 5580–5586, 2022. 1, 2", + "[8] Junyuan Deng, Qi Wu, Xieyuanli Chen, Songpengcheng Xia, Zhen Sun, Guoqing Liu, Wenxian Yu, and Ling Pei. NeRF-LOAM: Neural Implicit Representation for Large-Scale Incremental LiDAR Odometry and Mapping. In ICCV, pages 8184–8193, 2023. 1, 2", + "[9] Luca Di Giammarino, Leonardo Brizi, Tiziano Guadagnino, Cyrill Stachniss, and Giorgio Grisetti. MD-SLAM: Multicue Direct SLAM. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), pages 11047-11054, 2022. 1, 2", + "[10] Luca Di Giammarino, Emanuele Giacomini, Leonardo Brizi, Omar Salem, and Giorgio Grisetti. Photometric LiDAR and RGB-D Bundle Adjustment. IEEE Robotics and Automation Letters (RA-L), 8(7):4362-4369, 2023. 2", + "[11] Chris Engels, Henrik Stewénius, and David Nister. Bundle Adjustment Rules. Photogrammetric computer vision, 2(32), 2006. 6", + "[12] Simone Ferrari, Luca Di Giammarino, Leonardo Brizi, and Giorgio Grisetti. MAD-ICP: It is All About Matching Data – Robust and Informed LiDAR Odometry. IEEE Robotics and Automation Letters (RA-L), 9(11):9175–9182, 2024. 1, 2, 6", + "[13] Antoine Guédon and Vincent Lepetit. SuGaR: Surface-Aligned Gaussian Splating for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering. In CVPR, pages 5354-5363, 2024. 2, 7", + "[14] Sheng Hong, Junjie He, Xinhu Zheng, and Chunran Zheng. LIV-GaussMap: LiDAR-Inertial-Visual Fusion for Real" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Time 3D Radiance Field Map Rendering. IEEE Robotics and Automation Letters (RA-L), 9(11):9765-9772, 2024. 2", + "[15] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2D Gaussian Splatting for Geometrically Accurate Radiance Fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, New York, NY, USA, 2024. Association for Computing Machinery. 2, 3, 4, 5", + "[16] Seth Isaacson, Pou-Chun Kung, Mani Ramanagopal, Ram Vasudevan, and Katherine A. Skinner. LONER: LiDAR Only Neural Representations for Real-Time SLAM. IEEE Robotics and Automation Letters (RA-L), 8(12):8042-8049, 2023. 1, 2", + "[17] Junzhe Jiang, Chun Gu, Yurui Chen, and Li Zhang. GS-LiDAR: Generating Realistic LiDAR Point Clouds with Panoramic Gaussian Splatting. In ICLR, 2025. 2", + "[18] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the Fourth Eurographics Symposium on Geometry Processing, pages 61-70, Goslar, DEU, 2006. Eurographics Association. 7", + "[19] Nikhil Keetha, Jay Karhade, Krishna Murthy Jatavallabhula, Gengshan Yang, Sebastian Scherer, Deva Ramanan, and Jonathon Luiten. Splatam: Splat, track and map 3d gaussians for dense rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2", + "[20] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM TOG, 42(4):139:1-139:14, 2023. 1, 4", + "[21] Qing Li, Shaoyang Chen, Cheng Wang, Xin Li, Chenglu Wen, Ming Cheng, and Jonathan Li. LO-Net: Deep Real-Time Lidar Odometry. In CVPR, pages 8465–8474, 2019. 2", + "[22] Lorenzo Liso, Erik Sandström, Vladimir Yugay, Luc Van Gool, and Martin R. Oswald. Loopy-SLAM: Dense Neural SLAM with Loop Closures. In CVPR, pages 20363-20373, 2024. 1, 2, 5", + "[23] Hidenobu Matsuki, Riku Murai, Paul HJ Kelly, and Andrew J Davison. Gaussian splatting slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18039-18048, 2024. 2", + "[24] Hidenobu Matsuki, Riku Murai, Paul H. J. Kelly, and Andrew J. Davison. Gaussian Splitting SLAM. In CVPR, pages 18039-18048, 2024. 2, 5", + "[25] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In CVPR, 2019. 7", + "[26] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In ECCV, pages 405-421, Cham, 2020. Springer International Publishing. 1", + "[27] Ken Museth. VDB: High-resolution sparse volumes with dynamic topology. ACM TOG, 32(3):27:1-27:22, 2013. 6, 7, 8", + "[28] Austin Nicolai, Ryan Steele, Christopher Eriksen, and Geoffrey A. Hollinger. Deep Learning for Laser Based Odometry" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27638", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Estimation. In RSS workshop Limits and Potentials of Deep Learning in Robotics, page 1, 2016. 2", + "[29] Helen Oleynikova, Zachary Taylor, Marius Fehr, Roland Siegwart, and Juan Nieto. Voxblox: Incremental 3D Euclidean Signed Distance Fields for on-board MAV planning. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), pages 1366-1373, 2017. 6, 7", + "[30] Yue Pan, Xingguang Zhong, Louis Wiesmann, Thorbjörn Posewsky, Jens Behley, and Cyril Stachniss. PIN-SLAM: LiDAR SLAM Using a Point-Based Implicit Neural Representation for Achieving Global Map Consistency. IEEE Trans. on Robotics, 40:4045-4064, 2024. 1, 2, 6, 7", + "[31] Jan Quenzel and Sven Behnke. Real-time Multi-Adaptive-Resolution-Surfel 6D LiDAR Odometry using Continuous-time Trajectory Optimization. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), pages 5499–5506, 2021. 1, 2", + "[32] Jianyuan Ruan, Bo Li, Yibo Wang, and Yuxiang Sun. SLAMesh: Real-time LiDAR Simultaneous Localization and Meshing. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), pages 3546-3552, 2023. 1, 2, 6, 4", + "[33] Erik Sandström, Yue Li, Luc Van Gool, and Martin R. Oswald. Point-SLAM: Dense Neural Point Cloud-based SLAM. In ICCV, pages 18387-18398, 2023. 1, 2", + "[34] Erik Sandström, Keisuke Tateno, Michael Oechsle, Michael Niemeyer, Luc Van Gool, Martin R. Oswald, and Federico Tombari. Splat-SLAM: Globally Optimized RGB-only SLAM with 3D Gaussians, 2024. 2, 5", + "[35] Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-Motion Revisited. In CVPR, pages 4104-4113, 2016. 1", + "[36] Johannes L. Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise View Selection for Unstructured Multi-View Stereo. In ECCV, pages 501–518, Cham, 2016. Springer International Publishing. 1", + "[37] A. Segal, D. Haehnel, and S. Thrun. Generalized-ICP. In Proc. of Robotics: Science and Systems (RSS). Robotics: Science and Systems Foundation, 2009. 1, 2", + "[38] Tixiao Shan and Brendan Englot. LeGO-LOAM: Lightweight and Ground-Optimized Lidar Odometry and Mapping on Variable Terrain. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), pages 4758-4765, 2018. 2", + "[39] Shuangfu Song, Junqiao Zhao, Kai Huang, Jiaye Lin, Chen Ye, and Tiantian Feng. $\\mathbf{N}^3$ -Mapping: Normal Guided Neural Non-Projective Signed Distance Fields for Large-Scale 3D Mapping. IEEE Robotics and Automation Letters (RA-L), 9 (6):5935-5942, 2024. 1, 2, 6, 7", + "[40] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J. Davison. iMAP: Implicit Mapping and Positioning in Real-Time. In ICCV, pages 6209-6218, 2021. 1", + "[41] Yifu Tao, Miguel Ángel Muñoz-Bañón, Lintong Zhang, Jiāhao Wang, Lanke Frank Tarimo Fu, and Maurice Fallon. The Oxford Spires Dataset: Benchmarking Large-Scale LiDAR-Visual Localisation, Reconstruction and Radiance Field Methods, 2024. 6, 7, 3, 4, 5", + "[42] Fabio Tosi, Youmin Zhang, Ziren Gong, Erik Sandström, Stefano Mattoccia, Martin R. Oswald, and Matteo Poggi." + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "How nerfs and 3d gaussian splatting are reshaping slam: a survey, 2024. 2", + "[43] Ignacio Vizzo, Xieyuanli Chen, Nived Chebrolu, Jens Behley, and Cyril Stachniss. Poisson Surface Reconstruction for LiDAR Odometry and Mapping. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), pages 5624-5630, 2021. 6, 7, 3", + "[44] Guangming Wang, Xinrui Wu, Zhe Liu, and Hesheng Wang. PWCLO-Net: Deep LiDAR Odometry in 3D Point Clouds Using Hierarchical Embedding Mask Optimization. In CVPR, pages 15905–15914, 2021. 2", + "[45] Tim Weyrich, Simon Heinzle, Timo Aila, Daniel B. Fasnacht, Stephan Oetiker, Mario Botsch, Cyril Flaig, Simon Mall, Kaspar Rohrer, Norbert Felber, Hubert Kaeslin, and Markus Gross. A hardware architecture for surface splattering. ACM TOG, 26(3):90-es, 2007. 4", + "[46] Chenyang Wu, Yifan Duan, Xinran Zhang, Yu Sheng, Jianmin Ji, and Yanyong Zhang. MM-Gaussian: 3D Gaussian-based Multi-modal Fusion for Localization and Reconstruction in Unbounded Scenes. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), pages 12287-12293, 2024. 2", + "[47] Vladimir Yugay, Yue Li, Theo Gevers, and Martin R. Oswald. Gaussian-SLAM: Photo-realistic Dense SLAM with Gaussian Splatting, 2024. 2, 5", + "[48] Ganlin Zhang, Erik Sandström, Youmin Zhang, Manthan Patel, Luc Van Gool, and Martin R Oswald. Glorie-slam: Globally optimized rgb-only implicit encoding point cloud slam. arXiv preprint arXiv:2403.19549, 2024. 2", + "[49] Ji Zhang and Sanjiv Singh. LOAM: Lidar Odometry and Mapping in Real-time. In Proc. of Robotics: Science and Systems (RSS). Robotics: Science and Systems Foundation, 2014. 2, 6, 7", + "[50] Lintong Zhang, Marco Camurri, David Wisth, and Maurice Fallon. Multi-Camera LiDAR Inertial Extension to the Newer College Dataset, 2021. 6, 7, 8, 3", + "[51] Xin Zheng and Jianke Zhu. Efficient LiDAR Odometry for Autonomous Driving. IEEE Robotics and Automation Letters (RA-L), 6(4):8458-8465, 2021. 2", + "[52] Xingguang Zhong, Yue Pan, Jens Behley, and Cyril Stachniss. SHINE-Mapping: Large-Scale 3D Mapping Using Sparse Hierarchical Implicit Neural Representations. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), pages 8371-8377, 2023. 1, 2", + "[53] Liyuan Zhu, Yue Li, Erik Sandstrom, Shengyu Huang, Konrad Schindler, and Iro Armeni. LoopSplat: Loop Closure by Registering 3D Gaussian Splats. In Proc. of the International Conference on 3D Vision (3DV), 2025. 2, 5", + "[54] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Pollefeys. NICE-SLAM: Neural Implicit Scalable Encoding for SLAM. In CVPR, pages 12776-12786, 2022. 1, 2" + ], + "bbox": [ + 516, + 92, + 903, + 825 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27639", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Splat-LOAM_ Gaussian Splatting LiDAR Odometry and Mapping/21f438cd-a199-46e8-b668-60b759526f59_model.json b/2025/Splat-LOAM_ Gaussian Splatting LiDAR Odometry and Mapping/21f438cd-a199-46e8-b668-60b759526f59_model.json new file mode 100644 index 0000000000000000000000000000000000000000..31b73c458b4a7fbba9b8fb57761f408412d56d4f --- /dev/null +++ b/2025/Splat-LOAM_ Gaussian Splatting LiDAR Odometry and Mapping/21f438cd-a199-46e8-b668-60b759526f59_model.json @@ -0,0 +1,2574 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.165, + 0.131, + 0.834, + 0.154 + ], + "angle": 0, + "content": "Splat-LOAM: Gaussian Splitting LiDAR Odometry and Mapping" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.18, + 0.291, + 0.198 + ], + "angle": 0, + "content": "Emanuele Giacomini1" + }, + { + "type": "text", + "bbox": [ + 0.328, + 0.181, + 0.506, + 0.198 + ], + "angle": 0, + "content": "Luca Di Giammarino1" + }, + { + "type": "text", + "bbox": [ + 0.544, + 0.181, + 0.711, + 0.198 + ], + "angle": 0, + "content": "Lorenzo De Rebotti1" + }, + { + "type": "text", + "bbox": [ + 0.748, + 0.181, + 0.882, + 0.199 + ], + "angle": 0, + "content": "Giorgio Grisetti1" + }, + { + "type": "text", + "bbox": [ + 0.423, + 0.199, + 0.576, + 0.216 + ], + "angle": 0, + "content": "Martin R. Oswald" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.216, + 0.783, + 0.235 + ], + "angle": 0, + "content": "1Sapienza University of Rome, 2University of Amsterdam, Netherlands" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.301, + 0.486, + 0.558 + ], + "angle": 0, + "content": "LiDARs provide accurate geometric measurements, making them valuable for ego-motion estimation and reconstruction tasks. Although its success, managing an accurate and lightweight representation of the environment still poses challenges. Both classic and NeRF-based solutions have to trade off accuracy over memory and processing times. In this work, we build on recent advancements in Gaussian Splatting methods to develop a novel LiDAR odometry and mapping pipeline that exclusively relies on Gaussian primitives for its scene representation. Leveraging spherical projection, we drive the refinement of the primitives uniquely from LiDAR measurements. Experiments show that our approach matches the current registration performance, while achieving SOTA results for mapping tasks with minimal GPU requirements. This efficiency makes it a strong candidate for further exploration and potential adoption in real-time robotics estimation tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.588, + 0.222, + 0.604 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.484, + 0.84 + ], + "angle": 0, + "content": "LiDAR sensors provide accurate spatial measurements of the environment, making them valuable for ego-motion estimation and reconstruction tasks. Since the measurements already capture the 3d structure, many LiDAR Simultaneous Localization And Mapping (SLAM) pipelines do not explicitly refine the underlying point representation [2, 3, 7, 9, 12, 37]. Moreover, a global map can be obtained by directly stacking the measurements together. However, this typically leads to extremely large point clouds that cannot be explicitly used for online applications. Several approaches attempted to use surfels [1, 31] and meshes [32]. Although these approaches manage to simultaneously estimate the sensor's ego-motion while optimizing the map representation, they result in a trade-off between accuracy, memory usage, and runtime." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The recent advent of NeRF [26] sparked fresh interest in Novel View Synthesis (NVS) tasks. Specifically, given a set of input views and triangulated points (i.e. obtainable via COLMAP [35, 36]), NeRF learns a continuous" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.266, + 0.707, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.267, + 0.891, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.435, + 0.908, + 0.492 + ], + "angle": 0, + "content": "Figure 1. Performance overview of Splat-LOAM. F1 score to Active Memory (MB) and Runtime (s). The plots provide a quantitative comparison between state-of-the-art mapping pipelines, while PIN-SLAM and ours also perform online odometry." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.907, + 0.869 + ], + "angle": 0, + "content": "volumetric scene function. Color and density information is propagated throughout the radiance field by ray-casting pixels from the view cameras and, through a Multi-Layer Perceptron (MLP), the method learns a representation that ensures multi-view consistency. Concurrently, NeRFs inspired the computer vision community to tackle the problem of dense visual SLAM through implicit methods. The pioneering work in this context was iMAP [40], which stores the global appearance and geometry of the scene through a single MLP. Despite its success in driving a new research direction, the method suffered from issues related to the limited capacity of the model, leading to low reconstruction quality and catastrophic forgetting during the exploration of larger areas. These issues were later handled by shifting the paradigm and by moving some of the appearance and geometric information over a hierarchical feature grid [54] or neural point clouds [22, 33]. Similarly, these techniques were employed on LiDAR measurements to provide more accurate and lighter explicit dense representations [8, 16, 30, 39, 52]. However, these methods still require tailored sampling techniques to estimate the underlying Signed Distance Function (SDF) accurately. This bottleneck still poses problems for online execution." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "A recent, explicit alternative to NeRFs is 3D Gaussian Splitting (3DGS) [20]. This approach leverages 3D" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27630" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.257 + ], + "angle": 0, + "content": "Gaussian-shaped primitives and a differentiable, tile-based rasterizer to generate an appearance-accurate representation. Furthermore, having no need to model empty areas and no neural components, 3DGS earned a remarkable result in accuracy and training speeds. Additionally, several approaches further enhanced the reconstruction capabilities of this representation [6, 13, 15]. Being fast and accurate, this representation is now sparking interest in dense visual SLAM. Recently, 3D Gaussians were employed in several works, yielding superior results over implicit solutions [24, 47, 53]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.259, + 0.482, + 0.363 + ], + "angle": 0, + "content": "One issue with Gaussian Splitting relates to the primitive initialization. In regions where few or no points are provided by SFM, adaptive densification tends to fail, often yielding under-reconstructed regions. LiDAR priors are particularly useful in solving this issue due to explicit spatial measurements that can be used to initialize the local representation [14, 46]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.365, + 0.482, + 0.545 + ], + "angle": 0, + "content": "However, to our knowledge, no attempt has been made to evaluate the this representation for LiDAR-only scenarios. This technique could prove interesting for visual NVs initialization and LiDAR mapping as it could produce a lightweight, dense, and consistent representation. These insights led us to the development of Splat-LOAM, the first LiDAR Odometry and Mapping pipeline that only leverages Gaussian primitives as its surface representation. Our system demonstrates results on par with other SOTA pipelines at a fraction of the computational demands, proving as an additional research direction for real-time perception in autonomous systems." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.562, + 0.233, + 0.577 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.593, + 0.483, + 0.835 + ], + "angle": 0, + "content": "Classic LiDAR Odometry and Mapping. Feature-based methods that leverage specific points or groups of points to perform incremental registration. For instance, [38, 49] track feature points on sharp edges and planar surface patches, enabling high-frequency odometry estimation. On the other hand, Direct methods leverage the whole cloud to perform registration. Specifically, these methods can be categorized based on the subjects of the alignment. Scanto-Scan methods matches subsequent clouds, either explicitly [2, 3, 37] or through neural methods [21, 44], while Scan-to-Model methods match clouds with either a local or a global map. Typically, the map is represented using points [7, 12], surfels [1, 31], meshes [32]. Another explored solution project the measurements onto a spherical projection plane to leverage visual techniques for egomotion estimation [9, 10, 28, 51]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Implicit Methods. Concerning mapping only methods, Zhong et al. [52] proposed the first method for LiDAR implicit mapping that, given a set of point clouds and the corresponding sensor poses, used hierarchical feature grids to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.44 + ], + "angle": 0, + "content": "estimate the SDF of the scene. Their results demonstrated once again the advantages of such representation for surface reconstruction in terms of accuracy and memory footprint. A similar approach from Song et al. [39] improves the mapping accuracy by introducing SDF normal guided sampling and a hierarchical, voxel-guided sampling strategy for local optimization. Building on these advancements, several LiDAR odometry and mapping techniques were proposed. Deng et al. [8] presented the first implicit LiDAR LOAM system using an octree-based feature representation to encode the scene's SDF, used both for tracking and mapping. Similarly, Isaacson et al. [16] proposes a hierarchical feature grid to store the SDF information while using point-to-plane ICP to register new clouds. Pan et al. [30] leverages a neural point cloud representation to ensure a globally consistent estimate. The new clouds are registered using a correspondence-free point-to-implicit model approach. These methods prove that implicit representation can offer SOTA results in accuracy at the cost of potentially high execution times or memory requirements. Although targeting a different problem setting, related are also a series of visual neural SLAM methods with RGB or RGBD input [19, 22, 23, 33, 34, 47, 48, 54], see [42] for a survey." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Gaussian Splatting. Few works tackle the use of LiDAR within the context of Gaussian Splatting. Wu et al. [46] propose a multi-modal fusion system for SLAM. Specifically, by knowing the LiDAR to camera rigid pose, the initial pose estimate is obtained through point cloud registration and further refined via photometric error minimization. In this framework, the LiDAR points are leveraged to initialize the new 3D Gaussians. In a related approach, Hong et al. [14] proposes a LiDAR-Inertial-Visual SLAM system. The initial estimate is computed through a LiDAR-Inertial odometry. Points are partitioned using size-adaptive voxels to initialize 3D Gaussians using per-voxel covariances. Both the primitives and the trajectory are further refined via photometric error minimization. While these methods introduce LiDAR measurement, 3D Gaussians are still inherently processed by cameras. Recently, Chen et al. [5] applies 3D Gaussians for the task of LiDAR NVS for re-simulation. The authors propose the use of Periodic Vibrating 3D Gaussian primitives to account for dynamic objects present in the scene. The primitives are initialized using a lightweight MLP, and the rasterization is carried out in a spherical frame by computing a per-primitive plane orthogonal to the ray that connects the primitive's mean to the LiDAR frame, thus removing any distortion in the projection process. Focusing on a different formulation, Jiang et al. [17] propose a method of LiDAR NVS that leverages Periodic Vibrating 2D Gaussian primitives. The primitives are initialized by randomly sampling points and, further optimized using the losses described in [15], along with a Chamfer loss is introduced to constrain the 3D structures of the synthesized point" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27631" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.087, + 0.907, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.303, + 0.908, + 0.348 + ], + "angle": 0, + "content": "Figure 2. Splat-LOAM Overview. Given a LiDAR point cloud, we leverage the spherical projection to generate an image-like representation. Moreover, using an ad-hoc differentiable rasterizer, we guide the optimization for structural parameters of 2D Gaussians. The underlying representation is concurrently used to incrementally register new measurements." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.372, + 0.482, + 0.538 + ], + "angle": 0, + "content": "clouds, and an additional ray-drop term to account for phenomena like non-returning laser pulses. This term is further refined through a U-Net that considers other factors, such as the distance of the surface from the sensor. Compared to this work, we provide a thorough methodology to render 2D Gaussians on spherical frames while accounting for coordinates singularity and a cloud registration technique to allow for simultaneous odometry and mapping. To our knowledge, Splat-LOAM is the first pure LiDAR Odometry and Mapping pipeline that leverages Gaussian primitives both for mapping and tracking. In sum, our contributions are" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.539, + 0.483, + 0.569 + ], + "angle": 0, + "content": "- A differentiable, tile-based rasterizer for 2D Gaussians over spherical manifolds." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.57, + 0.483, + 0.599 + ], + "angle": 0, + "content": "- A mapping pipeline that allows the integration of LiDAR measurements in a 2D Gaussian representation." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.6, + 0.484, + 0.645 + ], + "angle": 0, + "content": "- A tracking schema that leverages both 3D and 2D representations to register new measurements and estimate the sensor ego-motion." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.539, + 0.484, + 0.645 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.66, + 0.182, + 0.675 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.686, + 0.484, + 0.792 + ], + "angle": 0, + "content": "This section introduces our novel LiDAR odometry and mapping method based on 2D Gaussian primitives. We detail a mapping strategy for initializing, refining, and integrating these primitives alongside a registration method that leverages geometric and photometric cues from the continuous local model for ego-motion estimation. Additional details can be found in the Supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.803, + 0.338, + 0.819 + ], + "angle": 0, + "content": "3.1. Spherical Projection Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "While original Gaussian Splatting leverages pinhole-camera projection to render or refine 3D primitives, Li-DAR input provides \\(360^{\\circ}\\) panoramic input. To this end, we employ spherical projection to encode LiDAR measurements into an image-like representation that we use to guide" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.907, + 0.584 + ], + "angle": 0, + "content": "the Gaussians optimization. A projection is a mapping \\(\\phi : \\mathbb{R}^3 \\to \\Gamma \\subset \\mathbb{R}^2\\) from a world point \\(\\mathbf{p} = (x, y, z)^T\\) to image coordinates \\(\\mathbf{u} = (u, v)^T\\). Knowing the range \\(d = \\| \\mathbf{p} \\|\\) of an image point \\(\\mathbf{u}\\), we can calculate the inverse mapping \\(\\phi^{-1} : \\Gamma \\times \\mathbb{R} \\to \\mathbb{R}^3\\), more explicitly \\(\\mathbf{p} = \\phi^{-1}(\\mathbf{u}, d)\\). We refer to this operation as back-projection. To ease the clarity of this work, it is worth mentioning that, compared to the classical pinhole camera, the optical reference frame is rearranged; the \\(x\\)-axis points forward, \\(y\\)-axis points to the left, and \\(z\\)-axis points upwards. Let \\(\\mathbf{K}\\) be a camera intrinsic matrix that can be computed directly from the point cloud (see Supplementary Material), with function \\(\\psi\\) mapping a 3D point to azimuth and elevation. Thus, the spherical projection of a point is given by" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.596, + 0.905, + 0.613 + ], + "angle": 0, + "content": "\\[\n\\phi (\\mathbf {p}) = \\mathbf {K} \\psi (\\mathbf {p}), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.615, + 0.905, + 0.672 + ], + "angle": 0, + "content": "\\[\n\\psi (\\mathbf {v}) = \\left[ \\begin{array}{c} \\operatorname {a t a n 2} \\left(v _ {y}, v _ {x}\\right) \\\\ \\operatorname {a t a n 2} \\left(v _ {z}, \\sqrt {v _ {x} ^ {2} + v _ {y} ^ {2}}\\right) \\\\ 1 \\end{array} \\right]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.684, + 0.907, + 0.719 + ], + "angle": 0, + "content": "We used spherical projection to obtain a range map \\(\\hat{\\mathbf{D}}\\) of the LiDAR point cloud \\(\\{\\mathbf{q}_p\\}_{p=1}^Q\\) of size \\(Q\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.726, + 0.724, + 0.743 + ], + "angle": 0, + "content": "3.2. 2D Gaussian Splitting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.749, + 0.907, + 0.855 + ], + "angle": 0, + "content": "Our scene representation is based on 2D Gaussians due to their affinity to surface reconstruction [6, 15]. A 2D Gaussian is defined by its opacity \\(o \\in [0,1]\\), centroid \\(\\pmb{\\mu} \\in \\mathbb{R}^3\\), origin plane defined by two tangential vectors \\(\\mathbf{t}_{\\alpha} \\in \\mathbb{R}^3\\) and \\(\\mathbf{t}_{\\beta} \\in \\mathbb{R}^3\\), and scaling vector \\(\\mathbf{s} = (s_{\\alpha}, s_{\\beta}) \\in \\mathbb{R}^2\\) that controls the variance of the Gaussian kernel [15]. Points on the splan's plane can be described in world space using:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.866, + 0.907, + 0.9 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} = \\left( \\begin{array}{c c c c} s _ {\\alpha} \\mathbf {t} _ {\\alpha} & s _ {\\beta} \\mathbf {t} _ {\\beta} & \\mathbf {0} & \\boldsymbol {\\mu} \\\\ 0 & 0 & 0 & 1 \\end{array} \\right), \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27632" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "while the Gaussian kernel can be evaluated on a point \\((\\alpha, \\beta)\\) in splat space using:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.134, + 0.483, + 0.17 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} (\\alpha , \\beta) = \\exp \\left(- \\frac {\\alpha^ {2} + \\beta^ {2}}{2}\\right). \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.181, + 0.232, + 0.195 + ], + "angle": 0, + "content": "3.2.1. Rasterization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.2, + 0.484, + 0.26 + ], + "angle": 0, + "content": "We render the Gaussians via \\(\\alpha\\)-blending as in [15, 20]. Specifically, for each pixel \\(\\mathbf{u}\\), we integrate Gaussians from front to back to obtain the range \\(d\\), normal \\(\\mathbf{n}\\), and opacity \\(o\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.271, + 0.483, + 0.315 + ], + "angle": 0, + "content": "\\[\nd = \\sum_ {i = 1} ^ {T} o _ {i} \\mathcal {G} _ {i} d _ {i} \\prod_ {j = 1} ^ {i - 1} \\left(1 - o _ {j} \\mathcal {G} _ {j}\\right) \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.317, + 0.483, + 0.36 + ], + "angle": 0, + "content": "\\[\n\\mathbf {n} = \\sum_ {i = 1} ^ {T} o _ {i} \\mathcal {G} _ {i} \\mathbf {t} _ {n _ {i}} \\prod_ {j = 1} ^ {i - 1} \\left(1 - o _ {j} \\mathcal {G} _ {j}\\right) \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.363, + 0.483, + 0.406 + ], + "angle": 0, + "content": "\\[\no = \\sum_ {i = 1} ^ {T} o _ {i} \\mathcal {G} _ {i} \\prod_ {j = 1} ^ {i - 1} \\left(1 - o _ {j} \\mathcal {G} _ {j}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.418, + 0.484, + 0.493 + ], + "angle": 0, + "content": "Different from the pinhole model, the spherical projection model described in Sec. 3.1 is non-affine, resulting in strong distortion artifacts on Gaussians that cover a significant extent in the image plane, if rasterizing using the local affine approximation method proposed in [20]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.495, + 0.484, + 0.585 + ], + "angle": 0, + "content": "To this extent, as in [15], our rasterizer leverages exact ray-splat intersections to render 2D Gaussians. This approach is particularly suited for our projection model as it explicitly relies on per-pixel directions to evaluate the Gaussian kernels, thus allowing us to account for the structure of our projection manifold surface." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.586, + 0.484, + 0.692 + ], + "angle": 0, + "content": "The ray-splat intersection is based on [45], and differently from [15], we reformulate the ray parametrization to cover the whole azimuthal space around the sensor. Additionally, we propose a method to compute the image-space extent of each Gaussian that is robust to the azimuthal coordinate singularity, ensuring accurate rasterization for any Gaussian configuration." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.702, + 0.292, + 0.717 + ], + "angle": 0, + "content": "3.2.2. Ray-splat Intersection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.753 + ], + "angle": 0, + "content": "Let \\(\\mathbf{v} = \\phi^{-1}\\left(\\mathbf{K}^{-1}\\mathbf{u}\\right)\\) be the normalized direction of pixel \\(\\mathbf{u}\\), parametrized as the intersection of two orthogonal planes" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.765, + 0.483, + 0.797 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} _ {x} = \\frac {\\mathbf {v} \\times \\mathbf {u} _ {z}}{\\left\\| \\mathbf {v} \\times \\mathbf {u} _ {z} \\right\\|}, \\quad \\mathbf {h} _ {y} = \\mathbf {h} _ {x} \\times \\mathbf {v}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.807, + 0.484, + 0.868 + ], + "angle": 0, + "content": "where \\(\\mathbf{u}_z\\) is the unit \\(z\\)-direction vector. This formulation is valid when \\(\\mathbf{v} \\neq \\mathbf{u}_z\\) which is a fair assumption in LiDAR measurements. Moreover, the planes are described in splat's space as" + }, + { + "type": "equation", + "bbox": [ + 0.129, + 0.878, + 0.484, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} _ {\\alpha} = \\left(\\mathbf {T} _ {w} ^ {c} \\mathbf {H}\\right) ^ {T} \\mathbf {h} _ {x}, \\quad \\mathbf {h} _ {\\beta} = \\left(\\mathbf {T} _ {w} ^ {c} \\mathbf {H}\\right) ^ {T} \\mathbf {h} _ {y}, \\quad (9)\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.088, + 0.669, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.677, + 0.089, + 0.907, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.268, + 0.907, + 0.407 + ], + "angle": 0, + "content": "Figure 3. Bounding box computation for near-singularity splats. (a) shows the 3D configuration of a splat that approximately lies behind the camera. (b) shows the corresponding spherical image with the projected bounding box vertices. The distortion is removed by shifting the vertices along the horizontal direction to align the projected center to the image center. (c) being far from the coordinate singularity, we compute the maximum extent of the splat. (d) We revert the shift and propagate to the corresponding tiles via an offset from the central vertex, matching with the tiles highlighted on (a)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.906, + 0.478 + ], + "angle": 0, + "content": "where \\(\\mathbf{T}_w^k\\in \\mathbb{S}\\mathbb{E}(3)\\) describes the world frame in the \\(k\\) -th sensor reference frame. We express the intersection of the two planes bundle as:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.487, + 0.906, + 0.508 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} _ {\\alpha} (\\alpha , \\beta , 1, 1) ^ {T} = \\mathbf {h} _ {\\beta} (\\alpha , \\beta , 1, 1) ^ {T} = 0. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.517, + 0.906, + 0.546 + ], + "angle": 0, + "content": "The solution \\(\\alpha\\) and \\(\\beta\\) is then computed by solving the homogeneous linear system:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.553, + 0.906, + 0.589 + ], + "angle": 0, + "content": "\\[\n\\alpha \\left(\\mathbf {u} _ {s}\\right) = \\frac {\\mathbf {h} _ {\\alpha} ^ {2} \\mathbf {h} _ {\\beta} ^ {4} - \\mathbf {h} _ {\\alpha} ^ {4} \\mathbf {h} _ {\\beta} ^ {2}}{\\mathbf {h} _ {\\alpha} ^ {1} \\mathbf {h} _ {\\beta} ^ {2} - \\mathbf {h} _ {\\alpha} ^ {2} \\mathbf {h} _ {\\beta} ^ {1}}, \\quad \\beta \\left(\\mathbf {u} _ {s}\\right) = \\frac {\\mathbf {h} _ {\\alpha} ^ {4} \\mathbf {h} _ {\\beta} ^ {1} - \\mathbf {h} _ {\\alpha} ^ {1} \\mathbf {h} _ {\\beta} ^ {4}}{\\mathbf {h} _ {\\alpha} ^ {1} \\mathbf {h} _ {\\beta} ^ {2} - \\mathbf {h} _ {\\alpha} ^ {2} \\mathbf {h} _ {\\beta} ^ {1}}. \\tag {11}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.596, + 0.757, + 0.611 + ], + "angle": 0, + "content": "3.2.3. Bounding Box Computation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.856 + ], + "angle": 0, + "content": "To efficiently rasterize Gaussians, each tile is informed of which Gaussians it should process. This is achieved by computing the per-Gaussian image-space extent from its projected central point and using it to inform each covered tile to rasterize the Gaussian. In our work, we compute the per-Gaussian extent by projecting the vertices of the \\(3\\sigma\\) splat-space bounding box on the spherical frame. Due to the horizontal periodic nature of the spherical frame, for Gaussians whose center is projected near the horizontal image boundaries and has a significant image-space size, one or more vertices of the bounding box may be projected on the opposite side of the image. If not handled, this situation can degrade the rasterization since not all the corresponding tiles may be informed, thus skipping rendering for some pixels and slowing the rendering process since the computed extent might almost cover the whole image plane." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We solve this by computing the Gaussian extent and tile propagation in a condition where the coordinate singularity cannot be harmful. As shown in Fig. 3, for each Gaussian," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27633" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.184 + ], + "angle": 0, + "content": "we project its bounding box and shift its projected points towards the center of the image plane while accounting for coordinate singularity. Similarly, after computing the Gaussian radius, the information is propagated to neighboring tiles while accounting for periodicity, ensuring correct rasterization and reducing the number of selected tiles." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.19, + 0.32, + 0.208 + ], + "angle": 0, + "content": "3.3. Odometry And Mapping" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.212, + 0.485, + 0.364 + ], + "angle": 0, + "content": "Following the modern literature of RGB-D SLAM [22, 34, 47, 53], we rely on keyframing to optimize local maps. We choose this approach for the following advantages: first, continuous integration over the same model can have adverse effects on artifacts and, more importantly, on runtime. Instead of decreasing the local density, we reset to a new local model if certain conditions are met, while also restricting the number of frames joining the optimization stage to allow effective online processing. Based on this, we define each local map as an individual Gaussian model \\(\\mathbf{P}^s\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.374, + 0.483, + 0.393 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} ^ {s} = \\left\\{\\mathcal {G} (\\boldsymbol {\\mu}, \\boldsymbol {\\Sigma}, o) | i = 1, \\dots , N \\right\\}. \\tag {12}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.4, + 0.317, + 0.415 + ], + "angle": 0, + "content": "3.3.1. Local model initialization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.419, + 0.485, + 0.482 + ], + "angle": 0, + "content": "We initialize a new local model using the input LiDAR point cloud whenever necessary, such as at system startup or when visibility conditions require it. As a first step, we generate a valid pixel mask via the indicator function \\(\\mathbb{1}[\\cdot ]\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.227, + 0.491, + 0.483, + 0.509 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {v} = \\mathbb {1} [ \\hat {\\mathbf {D}} > 0 ], \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.517, + 0.485, + 0.609 + ], + "angle": 0, + "content": "and compute the range gradients \\(\\nabla_{\\mathbf{D}}\\) to construct a weight map over the range image. We use a weighted sampling of \\(n_d\\) points to prioritize complex regions. Splat positions are computed by back-projecting the range image, while their orientations are initialized by directing surface normals toward the sensor center to enhance initial visibility." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.615, + 0.31, + 0.629 + ], + "angle": 0, + "content": "3.3.2. Local Model Refinement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.634, + 0.483, + 0.831 + ], + "angle": 0, + "content": "We perform a limited number of refinement iterations \\( n_o \\) on the most recent keyframes. Unlike [47, 53], we employ a geometric distribution-based sampling scheme that guarantees at least \\( 40\\% \\) probability of selecting the most recent keyframe and progressively decreases the likelihood of choosing the older ones. To filter out artifacts caused by ray-drop phenomena in the LiDAR measurements, we only consider valid pixels to refine the local model parameters \\( \\mathbf{x} \\). We start by applying a densification strategy similar to the one described in Sec. 3.3.1, with two extra terms. The first one, \\( \\mathbf{M}_n = \\mathbb{1}[\\mathbf{O}_i \\leq \\lambda_{d,o}] \\), target newly discovered areas while the second, \\( \\mathbf{M}_e = \\mathbb{1}[|\\mathbf{D}_i - \\hat{\\mathbf{D}}_i| \\geq \\lambda_{d,e}] \\), target under-reconstructed regions." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.832, + 0.483, + 0.863 + ], + "angle": 0, + "content": "To optimize the geometric consistency of the local model, we employ a loss term that minimizes the \\(L_{1}\\) error:" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.871, + 0.483, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d} = \\sum_ {\\mathbf {u} \\in \\mathbf {M} _ {v}} \\rho_ {d} \\| \\mathbf {D} (\\mathbf {u}, \\mathbf {x}) - \\hat {\\mathbf {D}} (\\mathbf {u}) \\|, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.153 + ], + "angle": 0, + "content": "where \\(\\rho_{d}\\) is a weight function dependent on the measurement's range. In addition, we employ a self-regularization term to align the splat's normals to the surface normals \\(\\mathbf{N}\\) estimated by the gradients of the range map \\(\\mathbf{D}\\) [15]:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.161, + 0.907, + 0.194 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {n} = \\sum_ {\\mathbf {u} \\in \\mathbf {M} _ {v}} 1 - \\mathbf {n} ^ {T} \\mathbf {N} (\\mathbf {D} (\\mathbf {u}, \\mathbf {x})) \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.203, + 0.906, + 0.294 + ], + "angle": 0, + "content": "Furthermore, to promote the expansion of splats over uniform surfaces, we introduce an additional term that operates on the opacity channel of the rasterized images. Specifically, we drive the splats to cover the areas of the image containing valid measurements by correlating the opacity image \\(\\mathbf{O}\\) with the valid mask \\(\\mathbf{M}_v\\)." + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.303, + 0.907, + 0.337 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {o} = \\sum_ {\\mathbf {u} \\in \\mathbf {M} _ {v}} - \\log (\\mathbf {O} (\\mathbf {u}, \\mathbf {x})). \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.344, + 0.906, + 0.43 + ], + "angle": 0, + "content": "While \\(\\mathcal{L}_o\\) allows the splats to grow, it can also cause some splats to become extremely large in unobserved areas. We employ an additional regularization term on the scaling parameter of all primitives \\(N\\) to compensate for this effect. We propose a novel regularization that allows the splat to extend up to a certain value \\(\\tau_s\\) before penalizing its growth:" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.436, + 0.907, + 0.467 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} = \\left\\{ \\begin{array}{c l} \\tau_ {s} - \\max \\left(s _ {\\alpha}, s _ {\\beta}\\right) & \\text {i f} \\max \\left(s _ {\\alpha}, s _ {\\beta}\\right) > \\tau_ {s} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.906, + 0.548 + ], + "angle": 0, + "content": "We found this term to be more effective rather than directly minimizing the deviation from the average [24, 47, 53] as it allows for anisotropic splats that are particularly useful for mapping details and edges, and provides more control on the density and structure of the model." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.549, + 0.869, + 0.565 + ], + "angle": 0, + "content": "The final mapping objective function is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.574, + 0.906, + 0.615 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a p}} = \\mathcal {L} _ {d} + \\lambda_ {o} \\mathcal {L} _ {o} + \\lambda_ {n} \\mathcal {L} _ {n} + \\lambda_ {s} \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {s _ {i}}, \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.623, + 0.906, + 0.745 + ], + "angle": 0, + "content": "with \\(\\lambda_{o}, \\lambda_{n}, \\lambda_{s} \\in \\mathbb{R}\\) being loss weights. This update strategy allows our method to maintain a geometrically accurate scene representation, while selectively densifying under-reconstructed regions that require more resolution, using a limited number of iterations. To avoid catastrophic forgetting, we do not perform any opacity reset steps. Moreover, we prune Gaussians that are either transparent or very small, as they typically cannot be observed during rendering." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.751, + 0.771, + 0.767 + ], + "angle": 0, + "content": "3.3.3. Frame-To-Model Registration" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.77, + 0.906, + 0.875 + ], + "angle": 0, + "content": "Each time a new keyframe is selected, we sample the local model by back-projecting the rasterized range image \\(\\mathbf{D}\\) onto a point cloud \\(\\{\\mathbf{p}_q\\}_{q=1}^{W \\times H}\\) at its estimated pose. We design an ad-hoc tracking schema to benefit from both geometric and photometric consistencies provided by the Li-DAR and the rendered local model. Hence, the total odometry loss is composed of the sum of both residuals:" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.886, + 0.907, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {o d o m}} = \\mathcal {L} _ {\\mathrm {g e o}} + \\mathcal {L} _ {\\mathrm {p h o t o}}. \\tag {19}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27634" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.26 + ], + "angle": 0, + "content": "Geometric Registration. To associate geometric entities, we employ a PCA-based kd-tree [12]. The kd-tree is built on the back-projected rendered range map \\(\\{\\mathbf{p}_q\\}_{q=1}^{W \\times H}\\) and segmented into tree leaves, corresponding to planar patches. Each leaf corresponds to \\(l = \\langle \\mathbf{p}_l, \\mathbf{n}_l \\rangle\\), that is, the mean point \\(\\mathbf{p}_l \\in \\mathbb{R}^3\\) and the surface normal \\(\\mathbf{n}_l \\in \\mathbb{R}^3\\). The geometric loss \\(\\mathcal{L}_{\\mathrm{geo}}\\) represents the sum of the point-to-plane distance between the mean leaf \\(\\mathbf{p}_l\\) and the point of the current measurement point cloud \\(k\\) with \\(\\{\\mathbf{q}_p\\}_{p=1}^Q\\), along the normal \\(\\mathbf{n}_l\\) expressed in the local reference frame \\(\\mathbf{T}_w^k\\). Specifically, we have:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.267, + 0.484, + 0.302 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {g e o}} = \\sum_ {p, q \\in \\{a \\}} \\rho \\left(\\left(\\mathbf {T} _ {w} ^ {k} \\mathbf {n} _ {l _ {q}}\\right) ^ {T} \\left(\\mathbf {T} _ {w} ^ {k} \\mathbf {q} _ {p} - \\mathbf {p} _ {q _ {i}}\\right)\\right), \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.308, + 0.484, + 0.356 + ], + "angle": 0, + "content": "where \\( w \\) is the global frame, \\( k \\) is the local frame, \\( \\{a\\} \\) is the set of leaf associations with the point from the measurement point cloud, and \\( \\rho \\) is the Huber robust loss function." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.373, + 0.484, + 0.446 + ], + "angle": 0, + "content": "Photometric Registration. Leveraging the rendered range map \\(\\mathbf{D}\\), we employ photometric registration for subpixel refinement, minimizing the photometric distance between the rendered and the spherical projected query point cloud \\(\\hat{\\mathbf{D}}\\). The photometric loss is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.451, + 0.483, + 0.491 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {p h o t o}} = \\sum_ {\\mathbf {u}} \\left\\| \\rho \\left(\\mathbf {D} (\\mathbf {u}) - \\hat {\\mathbf {D}} \\underbrace {\\left(\\phi \\left(\\mathbf {T} _ {w} ^ {k} \\phi^ {- 1} (\\mathbf {u} , \\hat {d})\\right)\\right)} _ {\\mathbf {u} ^ {\\prime}}\\right) \\right\\| ^ {2}. \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.499, + 0.484, + 0.59 + ], + "angle": 0, + "content": "The evaluation point \\(\\mathbf{u}'\\) in the query image \\(\\hat{\\mathbf{D}}\\) is computed by first back-projecting the pixel \\(\\mathbf{u}\\), applying the transform \\(\\mathbf{T}_w^k\\), and then projecting it back. Pose updates \\(\\delta\\) are parameterized as local updates in the Lie algebra \\(\\mathfrak{se}(3)\\). Therefore, the transformation \\(\\mathbf{T}_k^w\\) of the local reference frame, expressed in the global reference frame, is updated as:" + }, + { + "type": "equation", + "bbox": [ + 0.221, + 0.598, + 0.483, + 0.615 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} _ {k} ^ {w} \\leftarrow \\mathbf {T} _ {k} ^ {w} \\exp (\\delta). \\tag {22}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.623, + 0.484, + 0.668 + ], + "angle": 0, + "content": "This update is carried out using a second-order GaussNewton method. Local updates ensure that rotation updates are well-defined [11]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.68, + 0.224, + 0.696 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.854 + ], + "angle": 0, + "content": "In this section, we report the results of our method on several publicly available datasets. We evaluate our pipeline on tracking and mapping. We recall that, to our knowledge, this is the first Gaussian Splatting LiDAR Odometry and Mapping pipeline, and no direct competitors are available. We compared our approach with other well-known geometric and neural implicit methods. The experiments were executed on a PC with an Intel Core i9-14900K @ 3.20Ghz, 64GB of RAM, and an NVIDIA RTX 4090 GPU with 24 GB of VRAM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "For odometry evaluation, we include several baselines. The first one is a basic point-to-plane ICP odometry, followed by SLAMesh, that simultaneously estimates a mesh" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "representation of the scene via Gaussian Processes and performs registration onto it [32]. Moreover, MAD-ICP leverages a forest of PCA-based KD-Trees to perform accurate registration. Furthermore, we include PIN-SLAM as SOTA baseline for implicit LiDAR SLAM [30]. It relies on neural point primitives and interleaves an incremental learning of the model's SDF and a correspondence-free, point-to-implicit registration schema." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.907, + 0.456 + ], + "angle": 0, + "content": "For mapping evaluation, we include OpenVDB [27] and VoxBlox [29] as \"classic\" baselines. OpenVDB provides a robust volumetric data structure to handle 3D Points, while VoxBlox combines adaptive weights and grouped ray-casting for an efficient and accurate Truncated Signed Distance Function (TSDF) integration. Additionally, we include as neural-implicit approaches, \\( N^3 \\)-Mapping [39] and PIN-SLAM [30]. \\( N^3 \\)-Mapping is a neural-implicit non-projective SDF mapping approach that leverages normal guidance to produce more accurate SDFs, leading to SOTA results for offline LiDAR mapping. Moreover, PIN-SLAM is included as meshes can be extracted directly from the implicit representation via marching cubes. Below, we report the datasets and the evaluation metrics employed. We do not include SLAMesh [32] since it could not be run with ground-truth poses." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.468, + 0.616, + 0.482 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.49, + 0.878, + 0.505 + ], + "angle": 0, + "content": "We used the following four publicly available datasets:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.506, + 0.906, + 0.582 + ], + "angle": 0, + "content": "- Newer College Dataset (NC) [50]: Collected with a handheld Ouster OS0-128 LiDAR in structured and vegetated areas. Ground truth was generated using the Leica BLK360 scanner, achieving centimeter-level accuracy over poses and points in the map." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.586, + 0.906, + 0.662 + ], + "angle": 0, + "content": "- A Vision Benchmark in Rome (VBR) [4]: Recorded in Rome using OS1-64 (car-mounted) and OS0-128 (hand-held) LiDARs, capturing large-scale urban scenarios with narrow streets and dynamic objects. Ground truth trajectories were obtained by fusing LiDAR, IMU, and RTK" + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.506, + 0.906, + 0.662 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.69, + 0.9, + 0.846 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.855, + 0.906, + 0.898 + ], + "angle": 0, + "content": "Figure 4. RPE evaluation. Number of successful sequences across RPE thresholds. It includes the sequences of Newer College [49], VBR [4], Oxford Spires [41] and Mai City [43]." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27635" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.137, + 0.089, + 0.865, + 0.213 + ], + "angle": 0, + "content": "
DatasetNewer College[50]Oxford Spires[41]Mai-City[43]
quad-easymath-easykeble-02bodleian-02observatory-01mai-01mai-02
ApproachC-11↓F-score↑C-11↓F-score↑C-11↓F-score↑C-11↓F-score↑C-11↓F-score↑C-11↓F-score↑C-11↓F-score↑
OpenVDB[27]7.9288.8511.6984.517.1991.747.5189.689.5986.163.3397.293.2697.37
VoxBlox[29]16.564.6311.9380.5115.0371.6315.2458.7715.1270.457.0493.815.8095.51
N3-Mapping[39]8.0494.54failfail7.0193.477.8990.369.3587.942.6499.062.7699.04
PIN-SLAM[30]12.8988.0514.6973.8211.8379.6510.7482.7114.4972.314.9193.844.8294.11
Ours5.3796.749.1590.027.4394.417.6090.0910.5683.043.6497.275.0595.31
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.218, + 0.907, + 0.247 + ], + "angle": 0, + "content": "Table 1. Reconstruction evaluation. The pipelines were run with ground-truth poses. Voxel size is \\( {20}\\mathrm{\\;{cm}} \\) and F-score is computed with \\( {20}\\mathrm{\\;{cm}} \\) threshold. Splat-LOAM yield competitive mapping performance on both the Newer College[50] and Oxford Spires[41] dataset." + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.273, + 0.483, + 0.318 + ], + "angle": 0, + "content": "GNSS data, ensuring centimeter-level accuracy over the poses. This dataset is not used for mapping evaluation as no ground-truth maps are provided." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.323, + 0.483, + 0.473 + ], + "angle": 0, + "content": "- Oxford Spires [41]: Recorded with a hand-held Hesai QT64 LiDAR featuring a \\(360^{\\circ}\\) horizontal FoV, \\(104^{\\circ}\\) vertical FoV, 64 vertical channels, and 60 meters of range detection. Similar to [49], each sequence includes a prior map obtained via a survey-grade 3D imaging laser scanner, used for ground-truth trajectory estimation and mapping evaluation. Specifically, we choose the Keble College, Bodleian Library, and Radcliffe Observatory sequences to include both indoor and outdoor scenarios with different levels of detail." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.479, + 0.483, + 0.569 + ], + "angle": 0, + "content": "- Mai City [43]: A synthetic dataset captured using a car-like simulated LiDAR with 120 meters of range detection. The measurements are generated via ray-casting on an underlying mesh, providing error-free, motion-undistorted data. We select the \\(01\\) and \\(02\\) sequences, which capture similar scenarios with different vertical resolutions." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.323, + 0.483, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.578, + 0.212, + 0.592 + ], + "angle": 0, + "content": "4.2. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.483, + 0.66 + ], + "angle": 0, + "content": "We use Relative Pose Error (RPE) computed with progressively increasing delta steps to evaluate the odometry accuracy. Specifically, we adapt the deltas to the trajectory length to provide a more meaningful result [4]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.661, + 0.483, + 0.796 + ], + "angle": 0, + "content": "Differently, to evaluate mapping quality, we use several metrics [25]: Accuracy (Acc) measures the mean distance of points on the estimated mesh with their nearest neighbors on the reference cloud. Completeness (Comp) measures the opposite distance, and Chamfer-11 (C-11) describes the mean of the two. Additionally, we use the F-score computed with \\(20\\mathrm{cm}\\) error threshold. Table 1 reports compact results for C-11 and F1-score, additional tables are available in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.804, + 0.245, + 0.82 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Our approach employs Gaussian primitives for LiDAR odometry estimation and mapping, yielding results comparable to state-of-the-art methods while significantly enhancing computational efficiency. In this section, we evaluate some key aspects of our pipeline and evaluate their" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.274, + 0.608, + 0.286 + ], + "angle": 0, + "content": "contributions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.308, + 0.907, + 0.458 + ], + "angle": 0, + "content": "Memory and Runtime Analysis. In Fig. 5, we report how the increment of active primitives affects the active GPU memory requirements and the per-iteration mapping frequency. It shows an experiment run over the large-scale campus sequence [4] where we set a maximum of 100 keyframes per local map and sample, at most, \\(50\\%\\) of points for the incoming point cloud. It's possible to notice that the mapping FPS remains stable between 200k and 300k primitives. This is most likely related to the saturation of the rasterizer." + }, + { + "type": "image", + "bbox": [ + 0.544, + 0.474, + 0.878, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.636, + 0.907, + 0.693 + ], + "angle": 0, + "content": "Figure 5. Memory and Runtime Analysis. The plot relates the used GPU Memory and the mapping iteration frequency with the number of active primitives. The measurements were obtained over the longest sequence we reported: campus [4]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.715, + 0.906, + 0.821 + ], + "angle": 0, + "content": "Odometry. Fig. 7 shows the results for different tracking methods over our scene representation. Using both geometric and photometric components, we achieve a better result than using point-to-point or point-to-plane. The last three bars show an ablation of geometric and photometric loss components. The results are best for the joint use of both terms which support our design choice." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Mesh generation. We can sample the Gaussian scene to produce a mesh representation. Similar to [13], the meshing process involves sampling \\( n_m \\) points from each keyframe rendered range and normal maps, and running Poisson reconstruction [18]. As shown in Tab. 2, we found" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27636" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.151, + 0.089, + 0.236, + 0.102 + ], + "angle": 0, + "content": "PIN-SLAM" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.105, + 0.294, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.35, + 0.089, + 0.444, + 0.105 + ], + "angle": 0, + "content": "\\(\\mathbf{N}^3\\) -Mapping" + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.106, + 0.498, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.582, + 0.09, + 0.621, + 0.103 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.106, + 0.702, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.791, + 0.089, + 0.818, + 0.102 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.106, + 0.905, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.195, + 0.294, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.195, + 0.498, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.195, + 0.702, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.195, + 0.905, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.285, + 0.294, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.284, + 0.498, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.284, + 0.702, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.284, + 0.905, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.377, + 0.908, + 0.419 + ], + "angle": 0, + "content": "Figure 6. Comparison of Mesh Reconstruction. The figure shows reconstruction results for quad-easy sequence from the newer college dataset. Our method recovers a geometry with much higher data fidelity. PIN-SLAM lacks many details and exhibits a large level of noise. \\(N^3\\)-Mapping performs more similar to ours, but oversmoothes fine geometric details." + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.447, + 0.476, + 0.618 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.621, + 0.483, + 0.678 + ], + "angle": 0, + "content": "Figure 7. Ablation on registration methods. The plot reports the RPE (\\%) of several tested registration methods on the quad-easy sequence [50]. Enabling both geometric and photometric factors in sequence, provides a more robust estimate." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.706, + 0.484, + 0.797 + ], + "angle": 0, + "content": "that this method produces consistently better results than directly running TSDF integration and marching cubes over the rendered frames, or by running Poisson reconstruction directly on the Gaussian centers. Specifically, these method produces worse results in the regions that overlap multiple local models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.814, + 0.21, + 0.83 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We present the first LiDAR Odometry and Mapping pipeline that leverages 2D Gaussian primitives as the sole scene representation. Through an ad-hoc tile-based Gaussian rasterizer for spherical images, we leverage LiDAR" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.442, + 0.903, + 0.507 + ], + "angle": 0, + "content": "
Extraction MethodAcc↓Com↓C-11↓F-score↑
Marching Cubes [27]16.765.5311.1476.76
Poisson (centers)10.156.708.4392.33
Ours6.644.095.3796.74
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.511, + 0.907, + 0.555 + ], + "angle": 0, + "content": "Table 2. Ablation on Meshing Methods. We report mapping results with varying meshing methods on the quad-easy sequence [50]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.907, + 0.7 + ], + "angle": 0, + "content": "measurements to optimize the local model. Furthermore, we demonstrate the effectiveness of combining a geometric and photometric tracker to register new LiDAR point clouds over the Gaussian local model. The experiments show that our pipeline obtains tracking and mapping scores that are on par with the current SOTA at a fraction of the computational demands." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.712, + 0.906, + 0.805 + ], + "angle": 0, + "content": "Future Work. We plan on improving Splat-LOAM by simultaneously estimating the sensor pose and velocity to compensate the rolling-shutter distortions caused by motion while acquisition, typical of the LiDAR measurements. Moreover, we plan on including Loop Closure (LC) to improve the pose estimates, along with the mapping accuracy." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.838, + 0.673, + 0.856 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.873, + 0.9 + ], + "angle": 0, + "content": "This work has been supported by PNRR MUR project PE0000013-FAIR." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27637" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Jens Behley and Cyril Stachniss. Efficient Surfel-Based SLAM using 3D Laser Range Data in Urban Environments. In Proc. of Robotics: Science and Systems (RSS). Robotics: Science and Systems Foundation, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.2 + ], + "angle": 0, + "content": "[2] P.J. Besl and Neil D. McKay. A method for registration of 3-D shapes. IEEE TPAMI, 14(2):239-256, 1992. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.202, + 0.483, + 0.257 + ], + "angle": 0, + "content": "[3] Dorit Borrmann, Jan Elseberg, Kai Lingemann, Andreas Nüchter, and Joachim Hertzberg. Globally consistent 3D mapping with scan matching. Journal on Robotics and Autonomous Systems (RAS), 56(2):130-142, 2008. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.26, + 0.484, + 0.328 + ], + "angle": 0, + "content": "[4] Leonardo Brizi, Emanuele Giacomini, Luca Di Giammarino, Simone Ferrari, Omar Salem, Lorenzo De Rebotti, and Giorgio Grisetti. VBR: A Vision Benchmark in Rome. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), pages 15868-15874, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.331, + 0.483, + 0.372 + ], + "angle": 0, + "content": "[5] Qifeng Chen, Sheng Yang, Sicong Du, Tao Tang, Peng Chen, and Yuchi Huo. LiDAR-GS:Real-time LiDAR Re-Simulation using Gaussian Splitting, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.374, + 0.483, + 0.442 + ], + "angle": 0, + "content": "[6] Pinxuan Dai, Jiamin Xu, Wenxiang Xie, Xinguo Liu, Huamin Wang, and Weiwei Xu. High-quality Surface Reconstruction using Gaussian Surfels. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, New York, NY, USA, 2024. Association for Computing Machinery. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.445, + 0.483, + 0.513 + ], + "angle": 0, + "content": "[7] Pierre Dellenbach, Jean-Emmanuel Deschaud, Bastien Jacquet, and François Goulette. CT-ICP: Real-time Elastic LiDAR Odometry with Loop Closure. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), pages 5580–5586, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.516, + 0.483, + 0.584 + ], + "angle": 0, + "content": "[8] Junyuan Deng, Qi Wu, Xieyuanli Chen, Songpengcheng Xia, Zhen Sun, Guoqing Liu, Wenxian Yu, and Ling Pei. NeRF-LOAM: Neural Implicit Representation for Large-Scale Incremental LiDAR Odometry and Mapping. In ICCV, pages 8184–8193, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.587, + 0.483, + 0.655 + ], + "angle": 0, + "content": "[9] Luca Di Giammarino, Leonardo Brizi, Tiziano Guadagnino, Cyrill Stachniss, and Giorgio Grisetti. MD-SLAM: Multicue Direct SLAM. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), pages 11047-11054, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.659, + 0.483, + 0.713 + ], + "angle": 0, + "content": "[10] Luca Di Giammarino, Emanuele Giacomini, Leonardo Brizi, Omar Salem, and Giorgio Grisetti. Photometric LiDAR and RGB-D Bundle Adjustment. IEEE Robotics and Automation Letters (RA-L), 8(7):4362-4369, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.716, + 0.483, + 0.755 + ], + "angle": 0, + "content": "[11] Chris Engels, Henrik Stewénius, and David Nister. Bundle Adjustment Rules. Photogrammetric computer vision, 2(32), 2006. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.759, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[12] Simone Ferrari, Luca Di Giammarino, Leonardo Brizi, and Giorgio Grisetti. MAD-ICP: It is All About Matching Data – Robust and Informed LiDAR Odometry. IEEE Robotics and Automation Letters (RA-L), 9(11):9175–9182, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[13] Antoine Guédon and Vincent Lepetit. SuGaR: Surface-Aligned Gaussian Splating for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering. In CVPR, pages 5354-5363, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Sheng Hong, Junjie He, Xinhu Zheng, and Chunran Zheng. LIV-GaussMap: LiDAR-Inertial-Visual Fusion for Real" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Time 3D Radiance Field Map Rendering. IEEE Robotics and Automation Letters (RA-L), 9(11):9765-9772, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[15] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2D Gaussian Splatting for Geometrically Accurate Radiance Fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, New York, NY, USA, 2024. Association for Computing Machinery. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.905, + 0.258 + ], + "angle": 0, + "content": "[16] Seth Isaacson, Pou-Chun Kung, Mani Ramanagopal, Ram Vasudevan, and Katherine A. Skinner. LONER: LiDAR Only Neural Representations for Real-Time SLAM. IEEE Robotics and Automation Letters (RA-L), 8(12):8042-8049, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.904, + 0.3 + ], + "angle": 0, + "content": "[17] Junzhe Jiang, Chun Gu, Yurui Chen, and Li Zhang. GS-LiDAR: Generating Realistic LiDAR Point Clouds with Panoramic Gaussian Splatting. In ICLR, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[18] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the Fourth Eurographics Symposium on Geometry Processing, pages 61-70, Goslar, DEU, 2006. Eurographics Association. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.905, + 0.438 + ], + "angle": 0, + "content": "[19] Nikhil Keetha, Jay Karhade, Krishna Murthy Jatavallabhula, Gengshan Yang, Sebastian Scherer, Deva Ramanan, and Jonathon Luiten. Splatam: Splat, track and map 3d gaussians for dense rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.495 + ], + "angle": 0, + "content": "[20] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM TOG, 42(4):139:1-139:14, 2023. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.55 + ], + "angle": 0, + "content": "[21] Qing Li, Shaoyang Chen, Cheng Wang, Xin Li, Chenglu Wen, Ming Cheng, and Jonathan Li. LO-Net: Deep Real-Time Lidar Odometry. In CVPR, pages 8465–8474, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[22] Lorenzo Liso, Erik Sandström, Vladimir Yugay, Luc Van Gool, and Martin R. Oswald. Loopy-SLAM: Dense Neural SLAM with Loop Closures. In CVPR, pages 20363-20373, 2024. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[23] Hidenobu Matsuki, Riku Murai, Paul HJ Kelly, and Andrew J Davison. Gaussian splatting slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18039-18048, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[24] Hidenobu Matsuki, Riku Murai, Paul H. J. Kelly, and Andrew J. Davison. Gaussian Splitting SLAM. In CVPR, pages 18039-18048, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.76 + ], + "angle": 0, + "content": "[25] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In CVPR, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[26] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In ECCV, pages 405-421, Cham, 2020. Springer International Publishing. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[27] Ken Museth. VDB: High-resolution sparse volumes with dynamic topology. ACM TOG, 32(3):27:1-27:22, 2013. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[28] Austin Nicolai, Ryan Steele, Christopher Eriksen, and Geoffrey A. Hollinger. Deep Learning for Laser Based Odometry" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27638" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Estimation. In RSS workshop Limits and Potentials of Deep Learning in Robotics, page 1, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[29] Helen Oleynikova, Zachary Taylor, Marius Fehr, Roland Siegwart, and Juan Nieto. Voxblox: Incremental 3D Euclidean Signed Distance Fields for on-board MAV planning. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), pages 1366-1373, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[30] Yue Pan, Xingguang Zhong, Louis Wiesmann, Thorbjörn Posewsky, Jens Behley, and Cyril Stachniss. PIN-SLAM: LiDAR SLAM Using a Point-Based Implicit Neural Representation for Achieving Global Map Consistency. IEEE Trans. on Robotics, 40:4045-4064, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[31] Jan Quenzel and Sven Behnke. Real-time Multi-Adaptive-Resolution-Surfel 6D LiDAR Odometry using Continuous-time Trajectory Optimization. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), pages 5499–5506, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.384 + ], + "angle": 0, + "content": "[32] Jianyuan Ruan, Bo Li, Yibo Wang, and Yuxiang Sun. SLAMesh: Real-time LiDAR Simultaneous Localization and Meshing. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), pages 3546-3552, 2023. 1, 2, 6, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[33] Erik Sandström, Yue Li, Luc Van Gool, and Martin R. Oswald. Point-SLAM: Dense Neural Point Cloud-based SLAM. In ICCV, pages 18387-18398, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.481 + ], + "angle": 0, + "content": "[34] Erik Sandström, Keisuke Tateno, Michael Oechsle, Michael Niemeyer, Luc Van Gool, Martin R. Oswald, and Federico Tombari. Splat-SLAM: Globally Optimized RGB-only SLAM with 3D Gaussians, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.482, + 0.522 + ], + "angle": 0, + "content": "[35] Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-Motion Revisited. In CVPR, pages 4104-4113, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[36] Johannes L. Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise View Selection for Unstructured Multi-View Stereo. In ECCV, pages 501–518, Cham, 2016. Springer International Publishing. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.622 + ], + "angle": 0, + "content": "[37] A. Segal, D. Haehnel, and S. Thrun. Generalized-ICP. In Proc. of Robotics: Science and Systems (RSS). Robotics: Science and Systems Foundation, 2009. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[38] Tixiao Shan and Brendan Englot. LeGO-LOAM: Lightweight and Ground-Optimized Lidar Odometry and Mapping on Variable Terrain. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), pages 4758-4765, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[39] Shuangfu Song, Junqiao Zhao, Kai Huang, Jiaye Lin, Chen Ye, and Tiantian Feng. \\(\\mathbf{N}^3\\)-Mapping: Normal Guided Neural Non-Projective Signed Distance Fields for Large-Scale 3D Mapping. IEEE Robotics and Automation Letters (RA-L), 9 (6):5935-5942, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[40] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J. Davison. iMAP: Implicit Mapping and Positioning in Real-Time. In ICCV, pages 6209-6218, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[41] Yifu Tao, Miguel Ángel Muñoz-Bañón, Lintong Zhang, Jiāhao Wang, Lanke Frank Tarimo Fu, and Maurice Fallon. The Oxford Spires Dataset: Benchmarking Large-Scale LiDAR-Visual Localisation, Reconstruction and Radiance Field Methods, 2024. 6, 7, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[42] Fabio Tosi, Youmin Zhang, Ziren Gong, Erik Sandström, Stefano Mattoccia, Martin R. Oswald, and Matteo Poggi." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "How nerfs and 3d gaussian splatting are reshaping slam: a survey, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[43] Ignacio Vizzo, Xieyuanli Chen, Nived Chebrolu, Jens Behley, and Cyril Stachniss. Poisson Surface Reconstruction for LiDAR Odometry and Mapping. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), pages 5624-5630, 2021. 6, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[44] Guangming Wang, Xinrui Wu, Zhe Liu, and Hesheng Wang. PWCLO-Net: Deep LiDAR Odometry in 3D Point Clouds Using Hierarchical Embedding Mask Optimization. In CVPR, pages 15905–15914, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[45] Tim Weyrich, Simon Heinzle, Timo Aila, Daniel B. Fasnacht, Stephan Oetiker, Mario Botsch, Cyril Flaig, Simon Mall, Kaspar Rohrer, Norbert Felber, Hubert Kaeslin, and Markus Gross. A hardware architecture for surface splattering. ACM TOG, 26(3):90-es, 2007. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[46] Chenyang Wu, Yifan Duan, Xinran Zhang, Yu Sheng, Jianmin Ji, and Yanyong Zhang. MM-Gaussian: 3D Gaussian-based Multi-modal Fusion for Localization and Reconstruction in Unbounded Scenes. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), pages 12287-12293, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.403, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[47] Vladimir Yugay, Yue Li, Theo Gevers, and Martin R. Oswald. Gaussian-SLAM: Photo-realistic Dense SLAM with Gaussian Splatting, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.501 + ], + "angle": 0, + "content": "[48] Ganlin Zhang, Erik Sandström, Youmin Zhang, Manthan Patel, Luc Van Gool, and Martin R Oswald. Glorie-slam: Globally optimized rgb-only implicit encoding point cloud slam. arXiv preprint arXiv:2403.19549, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.503, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[49] Ji Zhang and Sanjiv Singh. LOAM: Lidar Odometry and Mapping in Real-time. In Proc. of Robotics: Science and Systems (RSS). Robotics: Science and Systems Foundation, 2014. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.56, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[50] Lintong Zhang, Marco Camurri, David Wisth, and Maurice Fallon. Multi-Camera LiDAR Inertial Extension to the Newer College Dataset, 2021. 6, 7, 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.602, + 0.905, + 0.643 + ], + "angle": 0, + "content": "[51] Xin Zheng and Jianke Zhu. Efficient LiDAR Odometry for Autonomous Driving. IEEE Robotics and Automation Letters (RA-L), 6(4):8458-8465, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.645, + 0.905, + 0.713 + ], + "angle": 0, + "content": "[52] Xingguang Zhong, Yue Pan, Jens Behley, and Cyril Stachniss. SHINE-Mapping: Large-Scale 3D Mapping Using Sparse Hierarchical Implicit Neural Representations. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), pages 8371-8377, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.715, + 0.905, + 0.77 + ], + "angle": 0, + "content": "[53] Liyuan Zhu, Yue Li, Erik Sandstrom, Shengyu Huang, Konrad Schindler, and Iro Armeni. LoopSplat: Loop Closure by Registering 3D Gaussian Splats. In Proc. of the International Conference on 3D Vision (3DV), 2025. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.772, + 0.905, + 0.827 + ], + "angle": 0, + "content": "[54] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Pollefeys. NICE-SLAM: Neural Implicit Scalable Encoding for SLAM. In CVPR, pages 12776-12786, 2022. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.827 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27639" + } + ] +] \ No newline at end of file diff --git a/2025/Splat-LOAM_ Gaussian Splatting LiDAR Odometry and Mapping/21f438cd-a199-46e8-b668-60b759526f59_origin.pdf b/2025/Splat-LOAM_ Gaussian Splatting LiDAR Odometry and Mapping/21f438cd-a199-46e8-b668-60b759526f59_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c0bdc062539a120cc2f3c000711cee1b02f79ac5 --- /dev/null +++ b/2025/Splat-LOAM_ Gaussian Splatting LiDAR Odometry and Mapping/21f438cd-a199-46e8-b668-60b759526f59_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:740e532e6ecdc4d48f116e469664fe4cea3cdf368d3f39b27a2accd6176a1bc4 +size 10246331 diff --git a/2025/Splat-LOAM_ Gaussian Splatting LiDAR Odometry and Mapping/full.md b/2025/Splat-LOAM_ Gaussian Splatting LiDAR Odometry and Mapping/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9bd3c96d1fd4359bc10f054d4b6f68b2876269c3 --- /dev/null +++ b/2025/Splat-LOAM_ Gaussian Splatting LiDAR Odometry and Mapping/full.md @@ -0,0 +1,401 @@ +# Splat-LOAM: Gaussian Splitting LiDAR Odometry and Mapping + +Emanuele Giacomini1 + +Luca Di Giammarino1 + +Lorenzo De Rebotti1 + +Giorgio Grisetti1 + +Martin R. Oswald + +1Sapienza University of Rome, 2University of Amsterdam, Netherlands + +# Abstract + +LiDARs provide accurate geometric measurements, making them valuable for ego-motion estimation and reconstruction tasks. Although its success, managing an accurate and lightweight representation of the environment still poses challenges. Both classic and NeRF-based solutions have to trade off accuracy over memory and processing times. In this work, we build on recent advancements in Gaussian Splatting methods to develop a novel LiDAR odometry and mapping pipeline that exclusively relies on Gaussian primitives for its scene representation. Leveraging spherical projection, we drive the refinement of the primitives uniquely from LiDAR measurements. Experiments show that our approach matches the current registration performance, while achieving SOTA results for mapping tasks with minimal GPU requirements. This efficiency makes it a strong candidate for further exploration and potential adoption in real-time robotics estimation tasks. + +# 1. Introduction + +LiDAR sensors provide accurate spatial measurements of the environment, making them valuable for ego-motion estimation and reconstruction tasks. Since the measurements already capture the 3d structure, many LiDAR Simultaneous Localization And Mapping (SLAM) pipelines do not explicitly refine the underlying point representation [2, 3, 7, 9, 12, 37]. Moreover, a global map can be obtained by directly stacking the measurements together. However, this typically leads to extremely large point clouds that cannot be explicitly used for online applications. Several approaches attempted to use surfels [1, 31] and meshes [32]. Although these approaches manage to simultaneously estimate the sensor's ego-motion while optimizing the map representation, they result in a trade-off between accuracy, memory usage, and runtime. + +The recent advent of NeRF [26] sparked fresh interest in Novel View Synthesis (NVS) tasks. Specifically, given a set of input views and triangulated points (i.e. obtainable via COLMAP [35, 36]), NeRF learns a continuous + +![](images/d6a2cc84b189922b1b99fae9b5c855a2f6e1ee4a85e162f99fcbe63b6fb47ea6.jpg) +Figure 1. Performance overview of Splat-LOAM. F1 score to Active Memory (MB) and Runtime (s). The plots provide a quantitative comparison between state-of-the-art mapping pipelines, while PIN-SLAM and ours also perform online odometry. + +![](images/970daaf16d5d54c429942fa3602d52c4af5362dc77d943901894dee7fdbd14ea.jpg) + +volumetric scene function. Color and density information is propagated throughout the radiance field by ray-casting pixels from the view cameras and, through a Multi-Layer Perceptron (MLP), the method learns a representation that ensures multi-view consistency. Concurrently, NeRFs inspired the computer vision community to tackle the problem of dense visual SLAM through implicit methods. The pioneering work in this context was iMAP [40], which stores the global appearance and geometry of the scene through a single MLP. Despite its success in driving a new research direction, the method suffered from issues related to the limited capacity of the model, leading to low reconstruction quality and catastrophic forgetting during the exploration of larger areas. These issues were later handled by shifting the paradigm and by moving some of the appearance and geometric information over a hierarchical feature grid [54] or neural point clouds [22, 33]. Similarly, these techniques were employed on LiDAR measurements to provide more accurate and lighter explicit dense representations [8, 16, 30, 39, 52]. However, these methods still require tailored sampling techniques to estimate the underlying Signed Distance Function (SDF) accurately. This bottleneck still poses problems for online execution. + +A recent, explicit alternative to NeRFs is 3D Gaussian Splitting (3DGS) [20]. This approach leverages 3D + +Gaussian-shaped primitives and a differentiable, tile-based rasterizer to generate an appearance-accurate representation. Furthermore, having no need to model empty areas and no neural components, 3DGS earned a remarkable result in accuracy and training speeds. Additionally, several approaches further enhanced the reconstruction capabilities of this representation [6, 13, 15]. Being fast and accurate, this representation is now sparking interest in dense visual SLAM. Recently, 3D Gaussians were employed in several works, yielding superior results over implicit solutions [24, 47, 53]. + +One issue with Gaussian Splitting relates to the primitive initialization. In regions where few or no points are provided by SFM, adaptive densification tends to fail, often yielding under-reconstructed regions. LiDAR priors are particularly useful in solving this issue due to explicit spatial measurements that can be used to initialize the local representation [14, 46]. + +However, to our knowledge, no attempt has been made to evaluate the this representation for LiDAR-only scenarios. This technique could prove interesting for visual NVs initialization and LiDAR mapping as it could produce a lightweight, dense, and consistent representation. These insights led us to the development of Splat-LOAM, the first LiDAR Odometry and Mapping pipeline that only leverages Gaussian primitives as its surface representation. Our system demonstrates results on par with other SOTA pipelines at a fraction of the computational demands, proving as an additional research direction for real-time perception in autonomous systems. + +# 2. Related Work + +Classic LiDAR Odometry and Mapping. Feature-based methods that leverage specific points or groups of points to perform incremental registration. For instance, [38, 49] track feature points on sharp edges and planar surface patches, enabling high-frequency odometry estimation. On the other hand, Direct methods leverage the whole cloud to perform registration. Specifically, these methods can be categorized based on the subjects of the alignment. Scanto-Scan methods matches subsequent clouds, either explicitly [2, 3, 37] or through neural methods [21, 44], while Scan-to-Model methods match clouds with either a local or a global map. Typically, the map is represented using points [7, 12], surfels [1, 31], meshes [32]. Another explored solution project the measurements onto a spherical projection plane to leverage visual techniques for egomotion estimation [9, 10, 28, 51]. + +Implicit Methods. Concerning mapping only methods, Zhong et al. [52] proposed the first method for LiDAR implicit mapping that, given a set of point clouds and the corresponding sensor poses, used hierarchical feature grids to + +estimate the SDF of the scene. Their results demonstrated once again the advantages of such representation for surface reconstruction in terms of accuracy and memory footprint. A similar approach from Song et al. [39] improves the mapping accuracy by introducing SDF normal guided sampling and a hierarchical, voxel-guided sampling strategy for local optimization. Building on these advancements, several LiDAR odometry and mapping techniques were proposed. Deng et al. [8] presented the first implicit LiDAR LOAM system using an octree-based feature representation to encode the scene's SDF, used both for tracking and mapping. Similarly, Isaacson et al. [16] proposes a hierarchical feature grid to store the SDF information while using point-to-plane ICP to register new clouds. Pan et al. [30] leverages a neural point cloud representation to ensure a globally consistent estimate. The new clouds are registered using a correspondence-free point-to-implicit model approach. These methods prove that implicit representation can offer SOTA results in accuracy at the cost of potentially high execution times or memory requirements. Although targeting a different problem setting, related are also a series of visual neural SLAM methods with RGB or RGBD input [19, 22, 23, 33, 34, 47, 48, 54], see [42] for a survey. + +Gaussian Splatting. Few works tackle the use of LiDAR within the context of Gaussian Splatting. Wu et al. [46] propose a multi-modal fusion system for SLAM. Specifically, by knowing the LiDAR to camera rigid pose, the initial pose estimate is obtained through point cloud registration and further refined via photometric error minimization. In this framework, the LiDAR points are leveraged to initialize the new 3D Gaussians. In a related approach, Hong et al. [14] proposes a LiDAR-Inertial-Visual SLAM system. The initial estimate is computed through a LiDAR-Inertial odometry. Points are partitioned using size-adaptive voxels to initialize 3D Gaussians using per-voxel covariances. Both the primitives and the trajectory are further refined via photometric error minimization. While these methods introduce LiDAR measurement, 3D Gaussians are still inherently processed by cameras. Recently, Chen et al. [5] applies 3D Gaussians for the task of LiDAR NVS for re-simulation. The authors propose the use of Periodic Vibrating 3D Gaussian primitives to account for dynamic objects present in the scene. The primitives are initialized using a lightweight MLP, and the rasterization is carried out in a spherical frame by computing a per-primitive plane orthogonal to the ray that connects the primitive's mean to the LiDAR frame, thus removing any distortion in the projection process. Focusing on a different formulation, Jiang et al. [17] propose a method of LiDAR NVS that leverages Periodic Vibrating 2D Gaussian primitives. The primitives are initialized by randomly sampling points and, further optimized using the losses described in [15], along with a Chamfer loss is introduced to constrain the 3D structures of the synthesized point + +![](images/8983ddc263f458559d96d6758114177ef7ef8023e5854e9900551b79d10cb212.jpg) +Figure 2. Splat-LOAM Overview. Given a LiDAR point cloud, we leverage the spherical projection to generate an image-like representation. Moreover, using an ad-hoc differentiable rasterizer, we guide the optimization for structural parameters of 2D Gaussians. The underlying representation is concurrently used to incrementally register new measurements. + +clouds, and an additional ray-drop term to account for phenomena like non-returning laser pulses. This term is further refined through a U-Net that considers other factors, such as the distance of the surface from the sensor. Compared to this work, we provide a thorough methodology to render 2D Gaussians on spherical frames while accounting for coordinates singularity and a cloud registration technique to allow for simultaneous odometry and mapping. To our knowledge, Splat-LOAM is the first pure LiDAR Odometry and Mapping pipeline that leverages Gaussian primitives both for mapping and tracking. In sum, our contributions are + +- A differentiable, tile-based rasterizer for 2D Gaussians over spherical manifolds. +- A mapping pipeline that allows the integration of LiDAR measurements in a 2D Gaussian representation. +- A tracking schema that leverages both 3D and 2D representations to register new measurements and estimate the sensor ego-motion. + +# 3. Method + +This section introduces our novel LiDAR odometry and mapping method based on 2D Gaussian primitives. We detail a mapping strategy for initializing, refining, and integrating these primitives alongside a registration method that leverages geometric and photometric cues from the continuous local model for ego-motion estimation. Additional details can be found in the Supplementary material. + +# 3.1. Spherical Projection Model + +While original Gaussian Splatting leverages pinhole-camera projection to render or refine 3D primitives, Li-DAR input provides $360^{\circ}$ panoramic input. To this end, we employ spherical projection to encode LiDAR measurements into an image-like representation that we use to guide + +the Gaussians optimization. A projection is a mapping $\phi : \mathbb{R}^3 \to \Gamma \subset \mathbb{R}^2$ from a world point $\mathbf{p} = (x, y, z)^T$ to image coordinates $\mathbf{u} = (u, v)^T$ . Knowing the range $d = \| \mathbf{p} \|$ of an image point $\mathbf{u}$ , we can calculate the inverse mapping $\phi^{-1} : \Gamma \times \mathbb{R} \to \mathbb{R}^3$ , more explicitly $\mathbf{p} = \phi^{-1}(\mathbf{u}, d)$ . We refer to this operation as back-projection. To ease the clarity of this work, it is worth mentioning that, compared to the classical pinhole camera, the optical reference frame is rearranged; the $x$ -axis points forward, $y$ -axis points to the left, and $z$ -axis points upwards. Let $\mathbf{K}$ be a camera intrinsic matrix that can be computed directly from the point cloud (see Supplementary Material), with function $\psi$ mapping a 3D point to azimuth and elevation. Thus, the spherical projection of a point is given by + +$$ +\phi (\mathbf {p}) = \mathbf {K} \psi (\mathbf {p}), \tag {1} +$$ + +$$ +\psi (\mathbf {v}) = \left[ \begin{array}{c} \operatorname {a t a n 2} \left(v _ {y}, v _ {x}\right) \\ \operatorname {a t a n 2} \left(v _ {z}, \sqrt {v _ {x} ^ {2} + v _ {y} ^ {2}}\right) \\ 1 \end{array} \right]. \tag {2} +$$ + +We used spherical projection to obtain a range map $\hat{\mathbf{D}}$ of the LiDAR point cloud $\{\mathbf{q}_p\}_{p=1}^Q$ of size $Q$ . + +# 3.2. 2D Gaussian Splitting + +Our scene representation is based on 2D Gaussians due to their affinity to surface reconstruction [6, 15]. A 2D Gaussian is defined by its opacity $o \in [0,1]$ , centroid $\pmb{\mu} \in \mathbb{R}^3$ , origin plane defined by two tangential vectors $\mathbf{t}_{\alpha} \in \mathbb{R}^3$ and $\mathbf{t}_{\beta} \in \mathbb{R}^3$ , and scaling vector $\mathbf{s} = (s_{\alpha}, s_{\beta}) \in \mathbb{R}^2$ that controls the variance of the Gaussian kernel [15]. Points on the splan's plane can be described in world space using: + +$$ +\mathbf {H} = \left( \begin{array}{c c c c} s _ {\alpha} \mathbf {t} _ {\alpha} & s _ {\beta} \mathbf {t} _ {\beta} & \mathbf {0} & \boldsymbol {\mu} \\ 0 & 0 & 0 & 1 \end{array} \right), \tag {3} +$$ + +while the Gaussian kernel can be evaluated on a point $(\alpha, \beta)$ in splat space using: + +$$ +\mathcal {G} (\alpha , \beta) = \exp \left(- \frac {\alpha^ {2} + \beta^ {2}}{2}\right). \tag {4} +$$ + +# 3.2.1. Rasterization + +We render the Gaussians via $\alpha$ -blending as in [15, 20]. Specifically, for each pixel $\mathbf{u}$ , we integrate Gaussians from front to back to obtain the range $d$ , normal $\mathbf{n}$ , and opacity $o$ as follows: + +$$ +d = \sum_ {i = 1} ^ {T} o _ {i} \mathcal {G} _ {i} d _ {i} \prod_ {j = 1} ^ {i - 1} \left(1 - o _ {j} \mathcal {G} _ {j}\right) \tag {5} +$$ + +$$ +\mathbf {n} = \sum_ {i = 1} ^ {T} o _ {i} \mathcal {G} _ {i} \mathbf {t} _ {n _ {i}} \prod_ {j = 1} ^ {i - 1} \left(1 - o _ {j} \mathcal {G} _ {j}\right) \tag {6} +$$ + +$$ +o = \sum_ {i = 1} ^ {T} o _ {i} \mathcal {G} _ {i} \prod_ {j = 1} ^ {i - 1} \left(1 - o _ {j} \mathcal {G} _ {j}\right) \tag {7} +$$ + +Different from the pinhole model, the spherical projection model described in Sec. 3.1 is non-affine, resulting in strong distortion artifacts on Gaussians that cover a significant extent in the image plane, if rasterizing using the local affine approximation method proposed in [20]. + +To this extent, as in [15], our rasterizer leverages exact ray-splat intersections to render 2D Gaussians. This approach is particularly suited for our projection model as it explicitly relies on per-pixel directions to evaluate the Gaussian kernels, thus allowing us to account for the structure of our projection manifold surface. + +The ray-splat intersection is based on [45], and differently from [15], we reformulate the ray parametrization to cover the whole azimuthal space around the sensor. Additionally, we propose a method to compute the image-space extent of each Gaussian that is robust to the azimuthal coordinate singularity, ensuring accurate rasterization for any Gaussian configuration. + +# 3.2.2. Ray-splat Intersection + +Let $\mathbf{v} = \phi^{-1}\left(\mathbf{K}^{-1}\mathbf{u}\right)$ be the normalized direction of pixel $\mathbf{u}$ , parametrized as the intersection of two orthogonal planes + +$$ +\mathbf {h} _ {x} = \frac {\mathbf {v} \times \mathbf {u} _ {z}}{\left\| \mathbf {v} \times \mathbf {u} _ {z} \right\|}, \quad \mathbf {h} _ {y} = \mathbf {h} _ {x} \times \mathbf {v}, \tag {8} +$$ + +where $\mathbf{u}_z$ is the unit $z$ -direction vector. This formulation is valid when $\mathbf{v} \neq \mathbf{u}_z$ which is a fair assumption in LiDAR measurements. Moreover, the planes are described in splat's space as + +$$ +\mathbf {h} _ {\alpha} = \left(\mathbf {T} _ {w} ^ {c} \mathbf {H}\right) ^ {T} \mathbf {h} _ {x}, \quad \mathbf {h} _ {\beta} = \left(\mathbf {T} _ {w} ^ {c} \mathbf {H}\right) ^ {T} \mathbf {h} _ {y}, \quad (9) +$$ + +![](images/c949944ef033cc573c706a2bbb36afb53fc94b0fbfdd94b3d4e2b17e168fec96.jpg) + +![](images/cb55ea3b6bbc2d40e6e25a29a56faede3fab92dd57de1c208429805fe24bf5f9.jpg) +Figure 3. Bounding box computation for near-singularity splats. (a) shows the 3D configuration of a splat that approximately lies behind the camera. (b) shows the corresponding spherical image with the projected bounding box vertices. The distortion is removed by shifting the vertices along the horizontal direction to align the projected center to the image center. (c) being far from the coordinate singularity, we compute the maximum extent of the splat. (d) We revert the shift and propagate to the corresponding tiles via an offset from the central vertex, matching with the tiles highlighted on (a). + +where $\mathbf{T}_w^k\in \mathbb{S}\mathbb{E}(3)$ describes the world frame in the $k$ -th sensor reference frame. We express the intersection of the two planes bundle as: + +$$ +\mathbf {h} _ {\alpha} (\alpha , \beta , 1, 1) ^ {T} = \mathbf {h} _ {\beta} (\alpha , \beta , 1, 1) ^ {T} = 0. \tag {10} +$$ + +The solution $\alpha$ and $\beta$ is then computed by solving the homogeneous linear system: + +$$ +\alpha \left(\mathbf {u} _ {s}\right) = \frac {\mathbf {h} _ {\alpha} ^ {2} \mathbf {h} _ {\beta} ^ {4} - \mathbf {h} _ {\alpha} ^ {4} \mathbf {h} _ {\beta} ^ {2}}{\mathbf {h} _ {\alpha} ^ {1} \mathbf {h} _ {\beta} ^ {2} - \mathbf {h} _ {\alpha} ^ {2} \mathbf {h} _ {\beta} ^ {1}}, \quad \beta \left(\mathbf {u} _ {s}\right) = \frac {\mathbf {h} _ {\alpha} ^ {4} \mathbf {h} _ {\beta} ^ {1} - \mathbf {h} _ {\alpha} ^ {1} \mathbf {h} _ {\beta} ^ {4}}{\mathbf {h} _ {\alpha} ^ {1} \mathbf {h} _ {\beta} ^ {2} - \mathbf {h} _ {\alpha} ^ {2} \mathbf {h} _ {\beta} ^ {1}}. \tag {11} +$$ + +# 3.2.3. Bounding Box Computation + +To efficiently rasterize Gaussians, each tile is informed of which Gaussians it should process. This is achieved by computing the per-Gaussian image-space extent from its projected central point and using it to inform each covered tile to rasterize the Gaussian. In our work, we compute the per-Gaussian extent by projecting the vertices of the $3\sigma$ splat-space bounding box on the spherical frame. Due to the horizontal periodic nature of the spherical frame, for Gaussians whose center is projected near the horizontal image boundaries and has a significant image-space size, one or more vertices of the bounding box may be projected on the opposite side of the image. If not handled, this situation can degrade the rasterization since not all the corresponding tiles may be informed, thus skipping rendering for some pixels and slowing the rendering process since the computed extent might almost cover the whole image plane. + +We solve this by computing the Gaussian extent and tile propagation in a condition where the coordinate singularity cannot be harmful. As shown in Fig. 3, for each Gaussian, + +we project its bounding box and shift its projected points towards the center of the image plane while accounting for coordinate singularity. Similarly, after computing the Gaussian radius, the information is propagated to neighboring tiles while accounting for periodicity, ensuring correct rasterization and reducing the number of selected tiles. + +# 3.3. Odometry And Mapping + +Following the modern literature of RGB-D SLAM [22, 34, 47, 53], we rely on keyframing to optimize local maps. We choose this approach for the following advantages: first, continuous integration over the same model can have adverse effects on artifacts and, more importantly, on runtime. Instead of decreasing the local density, we reset to a new local model if certain conditions are met, while also restricting the number of frames joining the optimization stage to allow effective online processing. Based on this, we define each local map as an individual Gaussian model $\mathbf{P}^s$ as + +$$ +\mathbf {P} ^ {s} = \left\{\mathcal {G} (\boldsymbol {\mu}, \boldsymbol {\Sigma}, o) | i = 1, \dots , N \right\}. \tag {12} +$$ + +# 3.3.1. Local model initialization + +We initialize a new local model using the input LiDAR point cloud whenever necessary, such as at system startup or when visibility conditions require it. As a first step, we generate a valid pixel mask via the indicator function $\mathbb{1}[\cdot ]$ as + +$$ +\mathbf {M} _ {v} = \mathbb {1} [ \hat {\mathbf {D}} > 0 ], \tag {13} +$$ + +and compute the range gradients $\nabla_{\mathbf{D}}$ to construct a weight map over the range image. We use a weighted sampling of $n_d$ points to prioritize complex regions. Splat positions are computed by back-projecting the range image, while their orientations are initialized by directing surface normals toward the sensor center to enhance initial visibility. + +# 3.3.2. Local Model Refinement + +We perform a limited number of refinement iterations $n_o$ on the most recent keyframes. Unlike [47, 53], we employ a geometric distribution-based sampling scheme that guarantees at least $40\%$ probability of selecting the most recent keyframe and progressively decreases the likelihood of choosing the older ones. To filter out artifacts caused by ray-drop phenomena in the LiDAR measurements, we only consider valid pixels to refine the local model parameters $\mathbf{x}$ . We start by applying a densification strategy similar to the one described in Sec. 3.3.1, with two extra terms. The first one, $\mathbf{M}_n = \mathbb{1}[\mathbf{O}_i \leq \lambda_{d,o}]$ , target newly discovered areas while the second, $\mathbf{M}_e = \mathbb{1}[|\mathbf{D}_i - \hat{\mathbf{D}}_i| \geq \lambda_{d,e}]$ , target under-reconstructed regions. + +To optimize the geometric consistency of the local model, we employ a loss term that minimizes the $L_{1}$ error: + +$$ +\mathcal {L} _ {d} = \sum_ {\mathbf {u} \in \mathbf {M} _ {v}} \rho_ {d} \| \mathbf {D} (\mathbf {u}, \mathbf {x}) - \hat {\mathbf {D}} (\mathbf {u}) \|, \tag {14} +$$ + +where $\rho_{d}$ is a weight function dependent on the measurement's range. In addition, we employ a self-regularization term to align the splat's normals to the surface normals $\mathbf{N}$ estimated by the gradients of the range map $\mathbf{D}$ [15]: + +$$ +\mathcal {L} _ {n} = \sum_ {\mathbf {u} \in \mathbf {M} _ {v}} 1 - \mathbf {n} ^ {T} \mathbf {N} (\mathbf {D} (\mathbf {u}, \mathbf {x})) \tag {15} +$$ + +Furthermore, to promote the expansion of splats over uniform surfaces, we introduce an additional term that operates on the opacity channel of the rasterized images. Specifically, we drive the splats to cover the areas of the image containing valid measurements by correlating the opacity image $\mathbf{O}$ with the valid mask $\mathbf{M}_v$ . + +$$ +\mathcal {L} _ {o} = \sum_ {\mathbf {u} \in \mathbf {M} _ {v}} - \log (\mathbf {O} (\mathbf {u}, \mathbf {x})). \tag {16} +$$ + +While $\mathcal{L}_o$ allows the splats to grow, it can also cause some splats to become extremely large in unobserved areas. We employ an additional regularization term on the scaling parameter of all primitives $N$ to compensate for this effect. We propose a novel regularization that allows the splat to extend up to a certain value $\tau_s$ before penalizing its growth: + +$$ +\mathcal {L} _ {s} = \left\{ \begin{array}{c l} \tau_ {s} - \max \left(s _ {\alpha}, s _ {\beta}\right) & \text {i f} \max \left(s _ {\alpha}, s _ {\beta}\right) > \tau_ {s} \\ 0 & \text {o t h e r w i s e} \end{array} \right. \tag {17} +$$ + +We found this term to be more effective rather than directly minimizing the deviation from the average [24, 47, 53] as it allows for anisotropic splats that are particularly useful for mapping details and edges, and provides more control on the density and structure of the model. + +The final mapping objective function is defined as: + +$$ +\mathcal {L} _ {\text {m a p}} = \mathcal {L} _ {d} + \lambda_ {o} \mathcal {L} _ {o} + \lambda_ {n} \mathcal {L} _ {n} + \lambda_ {s} \sum_ {i = 1} ^ {N} \mathcal {L} _ {s _ {i}}, \tag {18} +$$ + +with $\lambda_{o}, \lambda_{n}, \lambda_{s} \in \mathbb{R}$ being loss weights. This update strategy allows our method to maintain a geometrically accurate scene representation, while selectively densifying under-reconstructed regions that require more resolution, using a limited number of iterations. To avoid catastrophic forgetting, we do not perform any opacity reset steps. Moreover, we prune Gaussians that are either transparent or very small, as they typically cannot be observed during rendering. + +# 3.3.3. Frame-To-Model Registration + +Each time a new keyframe is selected, we sample the local model by back-projecting the rasterized range image $\mathbf{D}$ onto a point cloud $\{\mathbf{p}_q\}_{q=1}^{W \times H}$ at its estimated pose. We design an ad-hoc tracking schema to benefit from both geometric and photometric consistencies provided by the Li-DAR and the rendered local model. Hence, the total odometry loss is composed of the sum of both residuals: + +$$ +\mathcal {L} _ {\mathrm {o d o m}} = \mathcal {L} _ {\mathrm {g e o}} + \mathcal {L} _ {\mathrm {p h o t o}}. \tag {19} +$$ + +Geometric Registration. To associate geometric entities, we employ a PCA-based kd-tree [12]. The kd-tree is built on the back-projected rendered range map $\{\mathbf{p}_q\}_{q=1}^{W \times H}$ and segmented into tree leaves, corresponding to planar patches. Each leaf corresponds to $l = \langle \mathbf{p}_l, \mathbf{n}_l \rangle$ , that is, the mean point $\mathbf{p}_l \in \mathbb{R}^3$ and the surface normal $\mathbf{n}_l \in \mathbb{R}^3$ . The geometric loss $\mathcal{L}_{\mathrm{geo}}$ represents the sum of the point-to-plane distance between the mean leaf $\mathbf{p}_l$ and the point of the current measurement point cloud $k$ with $\{\mathbf{q}_p\}_{p=1}^Q$ , along the normal $\mathbf{n}_l$ expressed in the local reference frame $\mathbf{T}_w^k$ . Specifically, we have: + +$$ +\mathcal {L} _ {\mathrm {g e o}} = \sum_ {p, q \in \{a \}} \rho \left(\left(\mathbf {T} _ {w} ^ {k} \mathbf {n} _ {l _ {q}}\right) ^ {T} \left(\mathbf {T} _ {w} ^ {k} \mathbf {q} _ {p} - \mathbf {p} _ {q _ {i}}\right)\right), \tag {20} +$$ + +where $w$ is the global frame, $k$ is the local frame, $\{a\}$ is the set of leaf associations with the point from the measurement point cloud, and $\rho$ is the Huber robust loss function. + +Photometric Registration. Leveraging the rendered range map $\mathbf{D}$ , we employ photometric registration for subpixel refinement, minimizing the photometric distance between the rendered and the spherical projected query point cloud $\hat{\mathbf{D}}$ . The photometric loss is formulated as: + +$$ +\mathcal {L} _ {\mathrm {p h o t o}} = \sum_ {\mathbf {u}} \left\| \rho \left(\mathbf {D} (\mathbf {u}) - \hat {\mathbf {D}} \underbrace {\left(\phi \left(\mathbf {T} _ {w} ^ {k} \phi^ {- 1} (\mathbf {u} , \hat {d})\right)\right)} _ {\mathbf {u} ^ {\prime}}\right) \right\| ^ {2}. \tag {21} +$$ + +The evaluation point $\mathbf{u}'$ in the query image $\hat{\mathbf{D}}$ is computed by first back-projecting the pixel $\mathbf{u}$ , applying the transform $\mathbf{T}_w^k$ , and then projecting it back. Pose updates $\delta$ are parameterized as local updates in the Lie algebra $\mathfrak{se}(3)$ . Therefore, the transformation $\mathbf{T}_k^w$ of the local reference frame, expressed in the global reference frame, is updated as: + +$$ +\mathbf {T} _ {k} ^ {w} \leftarrow \mathbf {T} _ {k} ^ {w} \exp (\delta). \tag {22} +$$ + +This update is carried out using a second-order GaussNewton method. Local updates ensure that rotation updates are well-defined [11]. + +# 4. Experiments + +In this section, we report the results of our method on several publicly available datasets. We evaluate our pipeline on tracking and mapping. We recall that, to our knowledge, this is the first Gaussian Splatting LiDAR Odometry and Mapping pipeline, and no direct competitors are available. We compared our approach with other well-known geometric and neural implicit methods. The experiments were executed on a PC with an Intel Core i9-14900K @ 3.20Ghz, 64GB of RAM, and an NVIDIA RTX 4090 GPU with 24 GB of VRAM. + +For odometry evaluation, we include several baselines. The first one is a basic point-to-plane ICP odometry, followed by SLAMesh, that simultaneously estimates a mesh + +representation of the scene via Gaussian Processes and performs registration onto it [32]. Moreover, MAD-ICP leverages a forest of PCA-based KD-Trees to perform accurate registration. Furthermore, we include PIN-SLAM as SOTA baseline for implicit LiDAR SLAM [30]. It relies on neural point primitives and interleaves an incremental learning of the model's SDF and a correspondence-free, point-to-implicit registration schema. + +For mapping evaluation, we include OpenVDB [27] and VoxBlox [29] as "classic" baselines. OpenVDB provides a robust volumetric data structure to handle 3D Points, while VoxBlox combines adaptive weights and grouped ray-casting for an efficient and accurate Truncated Signed Distance Function (TSDF) integration. Additionally, we include as neural-implicit approaches, $N^3$ -Mapping [39] and PIN-SLAM [30]. $N^3$ -Mapping is a neural-implicit non-projective SDF mapping approach that leverages normal guidance to produce more accurate SDFs, leading to SOTA results for offline LiDAR mapping. Moreover, PIN-SLAM is included as meshes can be extracted directly from the implicit representation via marching cubes. Below, we report the datasets and the evaluation metrics employed. We do not include SLAMesh [32] since it could not be run with ground-truth poses. + +# 4.1. Datasets + +We used the following four publicly available datasets: + +- Newer College Dataset (NC) [50]: Collected with a handheld Ouster OS0-128 LiDAR in structured and vegetated areas. Ground truth was generated using the Leica BLK360 scanner, achieving centimeter-level accuracy over poses and points in the map. +- A Vision Benchmark in Rome (VBR) [4]: Recorded in Rome using OS1-64 (car-mounted) and OS0-128 (hand-held) LiDARs, capturing large-scale urban scenarios with narrow streets and dynamic objects. Ground truth trajectories were obtained by fusing LiDAR, IMU, and RTK + +![](images/fca29fcdfcbbabff88fd491a1ed51f2d78802e34d1760e18bd3389b1aa213bc2.jpg) +Figure 4. RPE evaluation. Number of successful sequences across RPE thresholds. It includes the sequences of Newer College [49], VBR [4], Oxford Spires [41] and Mai City [43]. + +
DatasetNewer College[50]Oxford Spires[41]Mai-City[43]
quad-easymath-easykeble-02bodleian-02observatory-01mai-01mai-02
ApproachC-11↓F-score↑C-11↓F-score↑C-11↓F-score↑C-11↓F-score↑C-11↓F-score↑C-11↓F-score↑C-11↓F-score↑
OpenVDB[27]7.9288.8511.6984.517.1991.747.5189.689.5986.163.3397.293.2697.37
VoxBlox[29]16.564.6311.9380.5115.0371.6315.2458.7715.1270.457.0493.815.8095.51
N3-Mapping[39]8.0494.54failfail7.0193.477.8990.369.3587.942.6499.062.7699.04
PIN-SLAM[30]12.8988.0514.6973.8211.8379.6510.7482.7114.4972.314.9193.844.8294.11
Ours5.3796.749.1590.027.4394.417.6090.0910.5683.043.6497.275.0595.31
+ +Table 1. Reconstruction evaluation. The pipelines were run with ground-truth poses. Voxel size is ${20}\mathrm{\;{cm}}$ and F-score is computed with ${20}\mathrm{\;{cm}}$ threshold. Splat-LOAM yield competitive mapping performance on both the Newer College[50] and Oxford Spires[41] dataset. + +GNSS data, ensuring centimeter-level accuracy over the poses. This dataset is not used for mapping evaluation as no ground-truth maps are provided. + +- Oxford Spires [41]: Recorded with a hand-held Hesai QT64 LiDAR featuring a $360^{\circ}$ horizontal FoV, $104^{\circ}$ vertical FoV, 64 vertical channels, and 60 meters of range detection. Similar to [49], each sequence includes a prior map obtained via a survey-grade 3D imaging laser scanner, used for ground-truth trajectory estimation and mapping evaluation. Specifically, we choose the Keble College, Bodleian Library, and Radcliffe Observatory sequences to include both indoor and outdoor scenarios with different levels of detail. +- Mai City [43]: A synthetic dataset captured using a car-like simulated LiDAR with 120 meters of range detection. The measurements are generated via ray-casting on an underlying mesh, providing error-free, motion-undistorted data. We select the $01$ and $02$ sequences, which capture similar scenarios with different vertical resolutions. + +# 4.2. Evaluation + +We use Relative Pose Error (RPE) computed with progressively increasing delta steps to evaluate the odometry accuracy. Specifically, we adapt the deltas to the trajectory length to provide a more meaningful result [4]. + +Differently, to evaluate mapping quality, we use several metrics [25]: Accuracy (Acc) measures the mean distance of points on the estimated mesh with their nearest neighbors on the reference cloud. Completeness (Comp) measures the opposite distance, and Chamfer-11 (C-11) describes the mean of the two. Additionally, we use the F-score computed with $20\mathrm{cm}$ error threshold. Table 1 reports compact results for C-11 and F1-score, additional tables are available in the Supplementary Material. + +# 4.3. Ablation Study + +Our approach employs Gaussian primitives for LiDAR odometry estimation and mapping, yielding results comparable to state-of-the-art methods while significantly enhancing computational efficiency. In this section, we evaluate some key aspects of our pipeline and evaluate their + +contributions. + +Memory and Runtime Analysis. In Fig. 5, we report how the increment of active primitives affects the active GPU memory requirements and the per-iteration mapping frequency. It shows an experiment run over the large-scale campus sequence [4] where we set a maximum of 100 keyframes per local map and sample, at most, $50\%$ of points for the incoming point cloud. It's possible to notice that the mapping FPS remains stable between 200k and 300k primitives. This is most likely related to the saturation of the rasterizer. + +![](images/6f1dd620c0146ede4fd189a3037f4501e87349e6880d43a9e754bfece47c38e4.jpg) +Figure 5. Memory and Runtime Analysis. The plot relates the used GPU Memory and the mapping iteration frequency with the number of active primitives. The measurements were obtained over the longest sequence we reported: campus [4]. + +Odometry. Fig. 7 shows the results for different tracking methods over our scene representation. Using both geometric and photometric components, we achieve a better result than using point-to-point or point-to-plane. The last three bars show an ablation of geometric and photometric loss components. The results are best for the joint use of both terms which support our design choice. + +Mesh generation. We can sample the Gaussian scene to produce a mesh representation. Similar to [13], the meshing process involves sampling $n_m$ points from each keyframe rendered range and normal maps, and running Poisson reconstruction [18]. As shown in Tab. 2, we found + +![](images/a47a6ccec38933cc3ff53dd986481cd0c763ed9714a02013cba541374b10ec3f.jpg) +PIN-SLAM + +![](images/1f8e480cb538ed36666c8be4e68f4247029873dbaba7e73832a1020c7f064854.jpg) +$\mathbf{N}^3$ -Mapping + +![](images/c3e7e666ea73a21b76ea9b7c989bce04d8c19abf7ae21be458852c99ade76367.jpg) +Ours + +![](images/93fefcc024af56629cf061fb2c87a54030a581e8c2e3fcb1be70387a71eb2862.jpg) +GT + +![](images/06cff9a1acd60fe66d5720707149a8fbaf1393ce209f470ef8707be1d508de2d.jpg) + +![](images/4b1389376e896f0a53bf765ed5595b392257f0be201800b3a3fc4e56c94aa456.jpg) + +![](images/c0ecd9bf88726ed22e365266e2089e86efd5dfe471faa64b095a5a0d911497f1.jpg) + +![](images/19bd142339bf312ce8748194350edeabee2683bdf20337e59f28a84c62cddbf0.jpg) + +![](images/0c39cbab10d5fb386e1d75c9028f56ad31a424c01935b8d4eb0be133c52cbae9.jpg) +Figure 6. Comparison of Mesh Reconstruction. The figure shows reconstruction results for quad-easy sequence from the newer college dataset. Our method recovers a geometry with much higher data fidelity. PIN-SLAM lacks many details and exhibits a large level of noise. $N^3$ -Mapping performs more similar to ours, but oversmoothes fine geometric details. + +![](images/205ad13a24b0e70972c290a0bc262b26c492a9c4d6330a034615c5b4c15cfbe1.jpg) + +![](images/d9d7c9816b7d2a45a1125af51c52fa18af51828882f975071df7da6d812f965a.jpg) + +![](images/e1bf2831635340691d4a5b09610f706800ef7ba227b4d3d2ad5ed628513cbd10.jpg) + +![](images/ed8fbcb19e771faa56bf43822a3a2a2e891feab34d32f99e001066088a5137de.jpg) +Figure 7. Ablation on registration methods. The plot reports the RPE (\%) of several tested registration methods on the quad-easy sequence [50]. Enabling both geometric and photometric factors in sequence, provides a more robust estimate. + +that this method produces consistently better results than directly running TSDF integration and marching cubes over the rendered frames, or by running Poisson reconstruction directly on the Gaussian centers. Specifically, these method produces worse results in the regions that overlap multiple local models. + +# 5. Conclusion + +We present the first LiDAR Odometry and Mapping pipeline that leverages 2D Gaussian primitives as the sole scene representation. Through an ad-hoc tile-based Gaussian rasterizer for spherical images, we leverage LiDAR + +
Extraction MethodAcc↓Com↓C-11↓F-score↑
Marching Cubes [27]16.765.5311.1476.76
Poisson (centers)10.156.708.4392.33
Ours6.644.095.3796.74
+ +Table 2. Ablation on Meshing Methods. We report mapping results with varying meshing methods on the quad-easy sequence [50]. + +measurements to optimize the local model. Furthermore, we demonstrate the effectiveness of combining a geometric and photometric tracker to register new LiDAR point clouds over the Gaussian local model. The experiments show that our pipeline obtains tracking and mapping scores that are on par with the current SOTA at a fraction of the computational demands. + +Future Work. We plan on improving Splat-LOAM by simultaneously estimating the sensor pose and velocity to compensate the rolling-shutter distortions caused by motion while acquisition, typical of the LiDAR measurements. Moreover, we plan on including Loop Closure (LC) to improve the pose estimates, along with the mapping accuracy. + +# Acknowledgments + +This work has been supported by PNRR MUR project PE0000013-FAIR. + +# References + +[1] Jens Behley and Cyril Stachniss. Efficient Surfel-Based SLAM using 3D Laser Range Data in Urban Environments. In Proc. of Robotics: Science and Systems (RSS). Robotics: Science and Systems Foundation, 2018. 1, 2 +[2] P.J. Besl and Neil D. McKay. A method for registration of 3-D shapes. IEEE TPAMI, 14(2):239-256, 1992. 1, 2 +[3] Dorit Borrmann, Jan Elseberg, Kai Lingemann, Andreas Nüchter, and Joachim Hertzberg. Globally consistent 3D mapping with scan matching. Journal on Robotics and Autonomous Systems (RAS), 56(2):130-142, 2008. 1, 2 +[4] Leonardo Brizi, Emanuele Giacomini, Luca Di Giammarino, Simone Ferrari, Omar Salem, Lorenzo De Rebotti, and Giorgio Grisetti. VBR: A Vision Benchmark in Rome. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), pages 15868-15874, 2024. 6, 7 +[5] Qifeng Chen, Sheng Yang, Sicong Du, Tao Tang, Peng Chen, and Yuchi Huo. LiDAR-GS:Real-time LiDAR Re-Simulation using Gaussian Splitting, 2024. 2 +[6] Pinxuan Dai, Jiamin Xu, Wenxiang Xie, Xinguo Liu, Huamin Wang, and Weiwei Xu. High-quality Surface Reconstruction using Gaussian Surfels. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, New York, NY, USA, 2024. Association for Computing Machinery. 2, 3 +[7] Pierre Dellenbach, Jean-Emmanuel Deschaud, Bastien Jacquet, and François Goulette. CT-ICP: Real-time Elastic LiDAR Odometry with Loop Closure. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), pages 5580–5586, 2022. 1, 2 +[8] Junyuan Deng, Qi Wu, Xieyuanli Chen, Songpengcheng Xia, Zhen Sun, Guoqing Liu, Wenxian Yu, and Ling Pei. NeRF-LOAM: Neural Implicit Representation for Large-Scale Incremental LiDAR Odometry and Mapping. In ICCV, pages 8184–8193, 2023. 1, 2 +[9] Luca Di Giammarino, Leonardo Brizi, Tiziano Guadagnino, Cyrill Stachniss, and Giorgio Grisetti. MD-SLAM: Multicue Direct SLAM. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), pages 11047-11054, 2022. 1, 2 +[10] Luca Di Giammarino, Emanuele Giacomini, Leonardo Brizi, Omar Salem, and Giorgio Grisetti. Photometric LiDAR and RGB-D Bundle Adjustment. IEEE Robotics and Automation Letters (RA-L), 8(7):4362-4369, 2023. 2 +[11] Chris Engels, Henrik Stewénius, and David Nister. Bundle Adjustment Rules. Photogrammetric computer vision, 2(32), 2006. 6 +[12] Simone Ferrari, Luca Di Giammarino, Leonardo Brizi, and Giorgio Grisetti. MAD-ICP: It is All About Matching Data – Robust and Informed LiDAR Odometry. IEEE Robotics and Automation Letters (RA-L), 9(11):9175–9182, 2024. 1, 2, 6 +[13] Antoine Guédon and Vincent Lepetit. SuGaR: Surface-Aligned Gaussian Splating for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering. In CVPR, pages 5354-5363, 2024. 2, 7 +[14] Sheng Hong, Junjie He, Xinhu Zheng, and Chunran Zheng. LIV-GaussMap: LiDAR-Inertial-Visual Fusion for Real + +Time 3D Radiance Field Map Rendering. IEEE Robotics and Automation Letters (RA-L), 9(11):9765-9772, 2024. 2 +[15] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2D Gaussian Splatting for Geometrically Accurate Radiance Fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, New York, NY, USA, 2024. Association for Computing Machinery. 2, 3, 4, 5 +[16] Seth Isaacson, Pou-Chun Kung, Mani Ramanagopal, Ram Vasudevan, and Katherine A. Skinner. LONER: LiDAR Only Neural Representations for Real-Time SLAM. IEEE Robotics and Automation Letters (RA-L), 8(12):8042-8049, 2023. 1, 2 +[17] Junzhe Jiang, Chun Gu, Yurui Chen, and Li Zhang. GS-LiDAR: Generating Realistic LiDAR Point Clouds with Panoramic Gaussian Splatting. In ICLR, 2025. 2 +[18] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the Fourth Eurographics Symposium on Geometry Processing, pages 61-70, Goslar, DEU, 2006. Eurographics Association. 7 +[19] Nikhil Keetha, Jay Karhade, Krishna Murthy Jatavallabhula, Gengshan Yang, Sebastian Scherer, Deva Ramanan, and Jonathon Luiten. Splatam: Splat, track and map 3d gaussians for dense rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2 +[20] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM TOG, 42(4):139:1-139:14, 2023. 1, 4 +[21] Qing Li, Shaoyang Chen, Cheng Wang, Xin Li, Chenglu Wen, Ming Cheng, and Jonathan Li. LO-Net: Deep Real-Time Lidar Odometry. In CVPR, pages 8465–8474, 2019. 2 +[22] Lorenzo Liso, Erik Sandström, Vladimir Yugay, Luc Van Gool, and Martin R. Oswald. Loopy-SLAM: Dense Neural SLAM with Loop Closures. In CVPR, pages 20363-20373, 2024. 1, 2, 5 +[23] Hidenobu Matsuki, Riku Murai, Paul HJ Kelly, and Andrew J Davison. Gaussian splatting slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18039-18048, 2024. 2 +[24] Hidenobu Matsuki, Riku Murai, Paul H. J. Kelly, and Andrew J. Davison. Gaussian Splitting SLAM. In CVPR, pages 18039-18048, 2024. 2, 5 +[25] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In CVPR, 2019. 7 +[26] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In ECCV, pages 405-421, Cham, 2020. Springer International Publishing. 1 +[27] Ken Museth. VDB: High-resolution sparse volumes with dynamic topology. ACM TOG, 32(3):27:1-27:22, 2013. 6, 7, 8 +[28] Austin Nicolai, Ryan Steele, Christopher Eriksen, and Geoffrey A. Hollinger. Deep Learning for Laser Based Odometry + +Estimation. In RSS workshop Limits and Potentials of Deep Learning in Robotics, page 1, 2016. 2 +[29] Helen Oleynikova, Zachary Taylor, Marius Fehr, Roland Siegwart, and Juan Nieto. Voxblox: Incremental 3D Euclidean Signed Distance Fields for on-board MAV planning. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), pages 1366-1373, 2017. 6, 7 +[30] Yue Pan, Xingguang Zhong, Louis Wiesmann, Thorbjörn Posewsky, Jens Behley, and Cyril Stachniss. PIN-SLAM: LiDAR SLAM Using a Point-Based Implicit Neural Representation for Achieving Global Map Consistency. IEEE Trans. on Robotics, 40:4045-4064, 2024. 1, 2, 6, 7 +[31] Jan Quenzel and Sven Behnke. Real-time Multi-Adaptive-Resolution-Surfel 6D LiDAR Odometry using Continuous-time Trajectory Optimization. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), pages 5499–5506, 2021. 1, 2 +[32] Jianyuan Ruan, Bo Li, Yibo Wang, and Yuxiang Sun. SLAMesh: Real-time LiDAR Simultaneous Localization and Meshing. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), pages 3546-3552, 2023. 1, 2, 6, 4 +[33] Erik Sandström, Yue Li, Luc Van Gool, and Martin R. Oswald. Point-SLAM: Dense Neural Point Cloud-based SLAM. In ICCV, pages 18387-18398, 2023. 1, 2 +[34] Erik Sandström, Keisuke Tateno, Michael Oechsle, Michael Niemeyer, Luc Van Gool, Martin R. Oswald, and Federico Tombari. Splat-SLAM: Globally Optimized RGB-only SLAM with 3D Gaussians, 2024. 2, 5 +[35] Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-Motion Revisited. In CVPR, pages 4104-4113, 2016. 1 +[36] Johannes L. Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise View Selection for Unstructured Multi-View Stereo. In ECCV, pages 501–518, Cham, 2016. Springer International Publishing. 1 +[37] A. Segal, D. Haehnel, and S. Thrun. Generalized-ICP. In Proc. of Robotics: Science and Systems (RSS). Robotics: Science and Systems Foundation, 2009. 1, 2 +[38] Tixiao Shan and Brendan Englot. LeGO-LOAM: Lightweight and Ground-Optimized Lidar Odometry and Mapping on Variable Terrain. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), pages 4758-4765, 2018. 2 +[39] Shuangfu Song, Junqiao Zhao, Kai Huang, Jiaye Lin, Chen Ye, and Tiantian Feng. $\mathbf{N}^3$ -Mapping: Normal Guided Neural Non-Projective Signed Distance Fields for Large-Scale 3D Mapping. IEEE Robotics and Automation Letters (RA-L), 9 (6):5935-5942, 2024. 1, 2, 6, 7 +[40] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J. Davison. iMAP: Implicit Mapping and Positioning in Real-Time. In ICCV, pages 6209-6218, 2021. 1 +[41] Yifu Tao, Miguel Ángel Muñoz-Bañón, Lintong Zhang, Jiāhao Wang, Lanke Frank Tarimo Fu, and Maurice Fallon. The Oxford Spires Dataset: Benchmarking Large-Scale LiDAR-Visual Localisation, Reconstruction and Radiance Field Methods, 2024. 6, 7, 3, 4, 5 +[42] Fabio Tosi, Youmin Zhang, Ziren Gong, Erik Sandström, Stefano Mattoccia, Martin R. Oswald, and Matteo Poggi. + +How nerfs and 3d gaussian splatting are reshaping slam: a survey, 2024. 2 +[43] Ignacio Vizzo, Xieyuanli Chen, Nived Chebrolu, Jens Behley, and Cyril Stachniss. Poisson Surface Reconstruction for LiDAR Odometry and Mapping. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), pages 5624-5630, 2021. 6, 7, 3 +[44] Guangming Wang, Xinrui Wu, Zhe Liu, and Hesheng Wang. PWCLO-Net: Deep LiDAR Odometry in 3D Point Clouds Using Hierarchical Embedding Mask Optimization. In CVPR, pages 15905–15914, 2021. 2 +[45] Tim Weyrich, Simon Heinzle, Timo Aila, Daniel B. Fasnacht, Stephan Oetiker, Mario Botsch, Cyril Flaig, Simon Mall, Kaspar Rohrer, Norbert Felber, Hubert Kaeslin, and Markus Gross. A hardware architecture for surface splattering. ACM TOG, 26(3):90-es, 2007. 4 +[46] Chenyang Wu, Yifan Duan, Xinran Zhang, Yu Sheng, Jianmin Ji, and Yanyong Zhang. MM-Gaussian: 3D Gaussian-based Multi-modal Fusion for Localization and Reconstruction in Unbounded Scenes. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), pages 12287-12293, 2024. 2 +[47] Vladimir Yugay, Yue Li, Theo Gevers, and Martin R. Oswald. Gaussian-SLAM: Photo-realistic Dense SLAM with Gaussian Splatting, 2024. 2, 5 +[48] Ganlin Zhang, Erik Sandström, Youmin Zhang, Manthan Patel, Luc Van Gool, and Martin R Oswald. Glorie-slam: Globally optimized rgb-only implicit encoding point cloud slam. arXiv preprint arXiv:2403.19549, 2024. 2 +[49] Ji Zhang and Sanjiv Singh. LOAM: Lidar Odometry and Mapping in Real-time. In Proc. of Robotics: Science and Systems (RSS). Robotics: Science and Systems Foundation, 2014. 2, 6, 7 +[50] Lintong Zhang, Marco Camurri, David Wisth, and Maurice Fallon. Multi-Camera LiDAR Inertial Extension to the Newer College Dataset, 2021. 6, 7, 8, 3 +[51] Xin Zheng and Jianke Zhu. Efficient LiDAR Odometry for Autonomous Driving. IEEE Robotics and Automation Letters (RA-L), 6(4):8458-8465, 2021. 2 +[52] Xingguang Zhong, Yue Pan, Jens Behley, and Cyril Stachniss. SHINE-Mapping: Large-Scale 3D Mapping Using Sparse Hierarchical Implicit Neural Representations. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), pages 8371-8377, 2023. 1, 2 +[53] Liyuan Zhu, Yue Li, Erik Sandstrom, Shengyu Huang, Konrad Schindler, and Iro Armeni. LoopSplat: Loop Closure by Registering 3D Gaussian Splats. In Proc. of the International Conference on 3D Vision (3DV), 2025. 2, 5 +[54] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Pollefeys. NICE-SLAM: Neural Implicit Scalable Encoding for SLAM. In CVPR, pages 12776-12786, 2022. 1, 2 \ No newline at end of file diff --git a/2025/Splat-LOAM_ Gaussian Splatting LiDAR Odometry and Mapping/images.zip b/2025/Splat-LOAM_ Gaussian Splatting LiDAR Odometry and Mapping/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6a7b4771be450649c5c9bbcbbdfc3e76c760f82e --- /dev/null +++ b/2025/Splat-LOAM_ Gaussian Splatting LiDAR Odometry and Mapping/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:73ae5f119852e8947dc89e58d67bb8103330bb524037f2ff7a9cd1e387d8cb12 +size 587894 diff --git a/2025/Splat-LOAM_ Gaussian Splatting LiDAR Odometry and Mapping/layout.json b/2025/Splat-LOAM_ Gaussian Splatting LiDAR Odometry and Mapping/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..71670cb660806ad6d29da8996172ab60435293fe --- /dev/null +++ b/2025/Splat-LOAM_ Gaussian Splatting LiDAR Odometry and Mapping/layout.json @@ -0,0 +1,9817 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 100, + 103, + 510, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 103, + 510, + 121 + ], + "spans": [ + { + "bbox": [ + 100, + 103, + 510, + 121 + ], + "type": "text", + "content": "Splat-LOAM: Gaussian Splitting LiDAR Odometry and Mapping" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 142, + 178, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 142, + 178, + 156 + ], + "spans": [ + { + "bbox": [ + 69, + 142, + 178, + 156 + ], + "type": "text", + "content": "Emanuele Giacomini1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 200, + 143, + 309, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 143, + 309, + 156 + ], + "spans": [ + { + "bbox": [ + 200, + 143, + 309, + 156 + ], + "type": "text", + "content": "Luca Di Giammarino1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 332, + 143, + 435, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 143, + 435, + 156 + ], + "spans": [ + { + "bbox": [ + 332, + 143, + 435, + 156 + ], + "type": "text", + "content": "Lorenzo De Rebotti1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 457, + 143, + 539, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 457, + 143, + 539, + 157 + ], + "spans": [ + { + "bbox": [ + 457, + 143, + 539, + 157 + ], + "type": "text", + "content": "Giorgio Grisetti1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 258, + 157, + 352, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 157, + 352, + 171 + ], + "spans": [ + { + "bbox": [ + 258, + 157, + 352, + 171 + ], + "type": "text", + "content": "Martin R. Oswald" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 171, + 479, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 479, + 186 + ], + "type": "text", + "content": "1Sapienza University of Rome, 2University of Amsterdam, Netherlands" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 238, + 297, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 238, + 297, + 441 + ], + "spans": [ + { + "bbox": [ + 54, + 238, + 297, + 441 + ], + "type": "text", + "content": "LiDARs provide accurate geometric measurements, making them valuable for ego-motion estimation and reconstruction tasks. Although its success, managing an accurate and lightweight representation of the environment still poses challenges. Both classic and NeRF-based solutions have to trade off accuracy over memory and processing times. In this work, we build on recent advancements in Gaussian Splatting methods to develop a novel LiDAR odometry and mapping pipeline that exclusively relies on Gaussian primitives for its scene representation. Leveraging spherical projection, we drive the refinement of the primitives uniquely from LiDAR measurements. Experiments show that our approach matches the current registration performance, while achieving SOTA results for mapping tasks with minimal GPU requirements. This efficiency makes it a strong candidate for further exploration and potential adoption in real-time robotics estimation tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 465, + 135, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 465, + 135, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 465, + 135, + 478 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 485, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 296, + 665 + ], + "type": "text", + "content": "LiDAR sensors provide accurate spatial measurements of the environment, making them valuable for ego-motion estimation and reconstruction tasks. Since the measurements already capture the 3d structure, many LiDAR Simultaneous Localization And Mapping (SLAM) pipelines do not explicitly refine the underlying point representation [2, 3, 7, 9, 12, 37]. Moreover, a global map can be obtained by directly stacking the measurements together. However, this typically leads to extremely large point clouds that cannot be explicitly used for online applications. Several approaches attempted to use surfels [1, 31] and meshes [32]. Although these approaches manage to simultaneously estimate the sensor's ego-motion while optimizing the map representation, they result in a trade-off between accuracy, memory usage, and runtime." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "The recent advent of NeRF [26] sparked fresh interest in Novel View Synthesis (NVS) tasks. Specifically, given a set of input views and triangulated points (i.e. obtainable via COLMAP [35, 36]), NeRF learns a continuous" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 321, + 210, + 432, + 342 + ], + "blocks": [ + { + "bbox": [ + 321, + 210, + 432, + 342 + ], + "lines": [ + { + "bbox": [ + 321, + 210, + 432, + 342 + ], + "spans": [ + { + "bbox": [ + 321, + 210, + 432, + 342 + ], + "type": "image", + "image_path": "d6a2cc84b189922b1b99fae9b5c855a2f6e1ee4a85e162f99fcbe63b6fb47ea6.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 344, + 555, + 389 + ], + "lines": [ + { + "bbox": [ + 313, + 344, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 344, + 555, + 389 + ], + "type": "text", + "content": "Figure 1. Performance overview of Splat-LOAM. F1 score to Active Memory (MB) and Runtime (s). The plots provide a quantitative comparison between state-of-the-art mapping pipelines, while PIN-SLAM and ours also perform online odometry." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 434, + 211, + 545, + 342 + ], + "blocks": [ + { + "bbox": [ + 434, + 211, + 545, + 342 + ], + "lines": [ + { + "bbox": [ + 434, + 211, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 434, + 211, + 545, + 342 + ], + "type": "image", + "image_path": "970daaf16d5d54c429942fa3602d52c4af5362dc77d943901894dee7fdbd14ea.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 413, + 555, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 555, + 688 + ], + "type": "text", + "content": "volumetric scene function. Color and density information is propagated throughout the radiance field by ray-casting pixels from the view cameras and, through a Multi-Layer Perceptron (MLP), the method learns a representation that ensures multi-view consistency. Concurrently, NeRFs inspired the computer vision community to tackle the problem of dense visual SLAM through implicit methods. The pioneering work in this context was iMAP [40], which stores the global appearance and geometry of the scene through a single MLP. Despite its success in driving a new research direction, the method suffered from issues related to the limited capacity of the model, leading to low reconstruction quality and catastrophic forgetting during the exploration of larger areas. These issues were later handled by shifting the paradigm and by moving some of the appearance and geometric information over a hierarchical feature grid [54] or neural point clouds [22, 33]. Similarly, these techniques were employed on LiDAR measurements to provide more accurate and lighter explicit dense representations [8, 16, 30, 39, 52]. However, these methods still require tailored sampling techniques to estimate the underlying Signed Distance Function (SDF) accurately. This bottleneck still poses problems for online execution." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "A recent, explicit alternative to NeRFs is 3D Gaussian Splitting (3DGS) [20]. This approach leverages 3D" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27630" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "content": "Gaussian-shaped primitives and a differentiable, tile-based rasterizer to generate an appearance-accurate representation. Furthermore, having no need to model empty areas and no neural components, 3DGS earned a remarkable result in accuracy and training speeds. Additionally, several approaches further enhanced the reconstruction capabilities of this representation [6, 13, 15]. Being fast and accurate, this representation is now sparking interest in dense visual SLAM. Recently, 3D Gaussians were employed in several works, yielding superior results over implicit solutions [24, 47, 53]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 205, + 294, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 294, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 294, + 287 + ], + "type": "text", + "content": "One issue with Gaussian Splitting relates to the primitive initialization. In regions where few or no points are provided by SFM, adaptive densification tends to fail, often yielding under-reconstructed regions. LiDAR priors are particularly useful in solving this issue due to explicit spatial measurements that can be used to initialize the local representation [14, 46]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 289, + 294, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 294, + 431 + ], + "type": "text", + "content": "However, to our knowledge, no attempt has been made to evaluate the this representation for LiDAR-only scenarios. This technique could prove interesting for visual NVs initialization and LiDAR mapping as it could produce a lightweight, dense, and consistent representation. These insights led us to the development of Splat-LOAM, the first LiDAR Odometry and Mapping pipeline that only leverages Gaussian primitives as its surface representation. Our system demonstrates results on par with other SOTA pipelines at a fraction of the computational demands, proving as an additional research direction for real-time perception in autonomous systems." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 445, + 142, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 142, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 142, + 456 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 469, + 295, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 469, + 295, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 295, + 661 + ], + "type": "text", + "content": "Classic LiDAR Odometry and Mapping. Feature-based methods that leverage specific points or groups of points to perform incremental registration. For instance, [38, 49] track feature points on sharp edges and planar surface patches, enabling high-frequency odometry estimation. On the other hand, Direct methods leverage the whole cloud to perform registration. Specifically, these methods can be categorized based on the subjects of the alignment. Scanto-Scan methods matches subsequent clouds, either explicitly [2, 3, 37] or through neural methods [21, 44], while Scan-to-Model methods match clouds with either a local or a global map. Typically, the map is represented using points [7, 12], surfels [1, 31], meshes [32]. Another explored solution project the measurements onto a spherical projection plane to leverage visual techniques for egomotion estimation [9, 10, 28, 51]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": "Implicit Methods. Concerning mapping only methods, Zhong et al. [52] proposed the first method for LiDAR implicit mapping that, given a set of point clouds and the corresponding sensor poses, used hierarchical feature grids to" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 348 + ], + "type": "text", + "content": "estimate the SDF of the scene. Their results demonstrated once again the advantages of such representation for surface reconstruction in terms of accuracy and memory footprint. A similar approach from Song et al. [39] improves the mapping accuracy by introducing SDF normal guided sampling and a hierarchical, voxel-guided sampling strategy for local optimization. Building on these advancements, several LiDAR odometry and mapping techniques were proposed. Deng et al. [8] presented the first implicit LiDAR LOAM system using an octree-based feature representation to encode the scene's SDF, used both for tracking and mapping. Similarly, Isaacson et al. [16] proposes a hierarchical feature grid to store the SDF information while using point-to-plane ICP to register new clouds. Pan et al. [30] leverages a neural point cloud representation to ensure a globally consistent estimate. The new clouds are registered using a correspondence-free point-to-implicit model approach. These methods prove that implicit representation can offer SOTA results in accuracy at the cost of potentially high execution times or memory requirements. Although targeting a different problem setting, related are also a series of visual neural SLAM methods with RGB or RGBD input [19, 22, 23, 33, 34, 47, 48, 54], see [42] for a survey." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 354, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 553, + 713 + ], + "type": "text", + "content": "Gaussian Splatting. Few works tackle the use of LiDAR within the context of Gaussian Splatting. Wu et al. [46] propose a multi-modal fusion system for SLAM. Specifically, by knowing the LiDAR to camera rigid pose, the initial pose estimate is obtained through point cloud registration and further refined via photometric error minimization. In this framework, the LiDAR points are leveraged to initialize the new 3D Gaussians. In a related approach, Hong et al. [14] proposes a LiDAR-Inertial-Visual SLAM system. The initial estimate is computed through a LiDAR-Inertial odometry. Points are partitioned using size-adaptive voxels to initialize 3D Gaussians using per-voxel covariances. Both the primitives and the trajectory are further refined via photometric error minimization. While these methods introduce LiDAR measurement, 3D Gaussians are still inherently processed by cameras. Recently, Chen et al. [5] applies 3D Gaussians for the task of LiDAR NVS for re-simulation. The authors propose the use of Periodic Vibrating 3D Gaussian primitives to account for dynamic objects present in the scene. The primitives are initialized using a lightweight MLP, and the rasterization is carried out in a spherical frame by computing a per-primitive plane orthogonal to the ray that connects the primitive's mean to the LiDAR frame, thus removing any distortion in the projection process. Focusing on a different formulation, Jiang et al. [17] propose a method of LiDAR NVS that leverages Periodic Vibrating 2D Gaussian primitives. The primitives are initialized by randomly sampling points and, further optimized using the losses described in [15], along with a Chamfer loss is introduced to constrain the 3D structures of the synthesized point" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "27631" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 68, + 555, + 236 + ], + "blocks": [ + { + "bbox": [ + 61, + 68, + 555, + 236 + ], + "lines": [ + { + "bbox": [ + 61, + 68, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 61, + 68, + 555, + 236 + ], + "type": "image", + "image_path": "8983ddc263f458559d96d6758114177ef7ef8023e5854e9900551b79d10cb212.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 239, + 555, + 275 + ], + "lines": [ + { + "bbox": [ + 54, + 239, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 54, + 239, + 555, + 275 + ], + "type": "text", + "content": "Figure 2. Splat-LOAM Overview. Given a LiDAR point cloud, we leverage the spherical projection to generate an image-like representation. Moreover, using an ad-hoc differentiable rasterizer, we guide the optimization for structural parameters of 2D Gaussians. The underlying representation is concurrently used to incrementally register new measurements." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 294, + 294, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 294, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 54, + 294, + 294, + 426 + ], + "type": "text", + "content": "clouds, and an additional ray-drop term to account for phenomena like non-returning laser pulses. This term is further refined through a U-Net that considers other factors, such as the distance of the surface from the sensor. Compared to this work, we provide a thorough methodology to render 2D Gaussians on spherical frames while accounting for coordinates singularity and a cloud registration technique to allow for simultaneous odometry and mapping. To our knowledge, Splat-LOAM is the first pure LiDAR Odometry and Mapping pipeline that leverages Gaussian primitives both for mapping and tracking. In sum, our contributions are" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 426, + 296, + 510 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 426, + 295, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 450 + ], + "type": "text", + "content": "- A differentiable, tile-based rasterizer for 2D Gaussians over spherical manifolds." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 451, + 295, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 295, + 474 + ], + "type": "text", + "content": "- A mapping pipeline that allows the integration of LiDAR measurements in a 2D Gaussian representation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 475, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 296, + 510 + ], + "type": "text", + "content": "- A tracking schema that leverages both 3D and 2D representations to register new measurements and estimate the sensor ego-motion." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 522, + 111, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 111, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 111, + 534 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 543, + 296, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 296, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 296, + 627 + ], + "type": "text", + "content": "This section introduces our novel LiDAR odometry and mapping method based on 2D Gaussian primitives. We detail a mapping strategy for initializing, refining, and integrating these primitives alongside a registration method that leverages geometric and photometric cues from the continuous local model for ego-motion estimation. Additional details can be found in the Supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 635, + 206, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 206, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 206, + 648 + ], + "type": "text", + "content": "3.1. Spherical Projection Model" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "While original Gaussian Splatting leverages pinhole-camera projection to render or refine 3D primitives, Li-DAR input provides " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " panoramic input. To this end, we employ spherical projection to encode LiDAR measurements into an image-like representation that we use to guide" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 294, + 555, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 555, + 462 + ], + "type": "text", + "content": "the Gaussians optimization. A projection is a mapping " + }, + { + "bbox": [ + 313, + 294, + 555, + 462 + ], + "type": "inline_equation", + "content": "\\phi : \\mathbb{R}^3 \\to \\Gamma \\subset \\mathbb{R}^2" + }, + { + "bbox": [ + 313, + 294, + 555, + 462 + ], + "type": "text", + "content": " from a world point " + }, + { + "bbox": [ + 313, + 294, + 555, + 462 + ], + "type": "inline_equation", + "content": "\\mathbf{p} = (x, y, z)^T" + }, + { + "bbox": [ + 313, + 294, + 555, + 462 + ], + "type": "text", + "content": " to image coordinates " + }, + { + "bbox": [ + 313, + 294, + 555, + 462 + ], + "type": "inline_equation", + "content": "\\mathbf{u} = (u, v)^T" + }, + { + "bbox": [ + 313, + 294, + 555, + 462 + ], + "type": "text", + "content": ". Knowing the range " + }, + { + "bbox": [ + 313, + 294, + 555, + 462 + ], + "type": "inline_equation", + "content": "d = \\| \\mathbf{p} \\|" + }, + { + "bbox": [ + 313, + 294, + 555, + 462 + ], + "type": "text", + "content": " of an image point " + }, + { + "bbox": [ + 313, + 294, + 555, + 462 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 313, + 294, + 555, + 462 + ], + "type": "text", + "content": ", we can calculate the inverse mapping " + }, + { + "bbox": [ + 313, + 294, + 555, + 462 + ], + "type": "inline_equation", + "content": "\\phi^{-1} : \\Gamma \\times \\mathbb{R} \\to \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 294, + 555, + 462 + ], + "type": "text", + "content": ", more explicitly " + }, + { + "bbox": [ + 313, + 294, + 555, + 462 + ], + "type": "inline_equation", + "content": "\\mathbf{p} = \\phi^{-1}(\\mathbf{u}, d)" + }, + { + "bbox": [ + 313, + 294, + 555, + 462 + ], + "type": "text", + "content": ". We refer to this operation as back-projection. To ease the clarity of this work, it is worth mentioning that, compared to the classical pinhole camera, the optical reference frame is rearranged; the " + }, + { + "bbox": [ + 313, + 294, + 555, + 462 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 294, + 555, + 462 + ], + "type": "text", + "content": "-axis points forward, " + }, + { + "bbox": [ + 313, + 294, + 555, + 462 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 294, + 555, + 462 + ], + "type": "text", + "content": "-axis points to the left, and " + }, + { + "bbox": [ + 313, + 294, + 555, + 462 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 294, + 555, + 462 + ], + "type": "text", + "content": "-axis points upwards. Let " + }, + { + "bbox": [ + 313, + 294, + 555, + 462 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 313, + 294, + 555, + 462 + ], + "type": "text", + "content": " be a camera intrinsic matrix that can be computed directly from the point cloud (see Supplementary Material), with function " + }, + { + "bbox": [ + 313, + 294, + 555, + 462 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 313, + 294, + 555, + 462 + ], + "type": "text", + "content": " mapping a 3D point to azimuth and elevation. Thus, the spherical projection of a point is given by" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 359, + 472, + 553, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 472, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 359, + 472, + 553, + 485 + ], + "type": "interline_equation", + "content": "\\phi (\\mathbf {p}) = \\mathbf {K} \\psi (\\mathbf {p}), \\tag {1}", + "image_path": "0d87d06a58d132983f4e72d242242c10b5bef4ab4f12da2ac0adca09af4e2b5f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 359, + 487, + 553, + 532 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 487, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 359, + 487, + 553, + 532 + ], + "type": "interline_equation", + "content": "\\psi (\\mathbf {v}) = \\left[ \\begin{array}{c} \\operatorname {a t a n 2} \\left(v _ {y}, v _ {x}\\right) \\\\ \\operatorname {a t a n 2} \\left(v _ {z}, \\sqrt {v _ {x} ^ {2} + v _ {y} ^ {2}}\\right) \\\\ 1 \\end{array} \\right]. \\tag {2}", + "image_path": "262ab81c289daf34e883725c216325f4338a0be6d0f07571247eea7e928ef7d1.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 541, + 555, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 541, + 555, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 541, + 555, + 569 + ], + "type": "text", + "content": "We used spherical projection to obtain a range map " + }, + { + "bbox": [ + 313, + 541, + 555, + 569 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{D}}" + }, + { + "bbox": [ + 313, + 541, + 555, + 569 + ], + "type": "text", + "content": " of the LiDAR point cloud " + }, + { + "bbox": [ + 313, + 541, + 555, + 569 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{q}_p\\}_{p=1}^Q" + }, + { + "bbox": [ + 313, + 541, + 555, + 569 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 313, + 541, + 555, + 569 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 541, + 555, + 569 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 574, + 443, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 443, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 443, + 588 + ], + "type": "text", + "content": "3.2. 2D Gaussian Splitting" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 593, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 555, + 677 + ], + "type": "text", + "content": "Our scene representation is based on 2D Gaussians due to their affinity to surface reconstruction [6, 15]. A 2D Gaussian is defined by its opacity " + }, + { + "bbox": [ + 313, + 593, + 555, + 677 + ], + "type": "inline_equation", + "content": "o \\in [0,1]" + }, + { + "bbox": [ + 313, + 593, + 555, + 677 + ], + "type": "text", + "content": ", centroid " + }, + { + "bbox": [ + 313, + 593, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 593, + 555, + 677 + ], + "type": "text", + "content": ", origin plane defined by two tangential vectors " + }, + { + "bbox": [ + 313, + 593, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{\\alpha} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 593, + 555, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 593, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{\\beta} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 593, + 555, + 677 + ], + "type": "text", + "content": ", and scaling vector " + }, + { + "bbox": [ + 313, + 593, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{s} = (s_{\\alpha}, s_{\\beta}) \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 313, + 593, + 555, + 677 + ], + "type": "text", + "content": " that controls the variance of the Gaussian kernel [15]. Points on the splan's plane can be described in world space using:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 369, + 685, + 555, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 685, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 369, + 685, + 555, + 712 + ], + "type": "interline_equation", + "content": "\\mathbf {H} = \\left( \\begin{array}{c c c c} s _ {\\alpha} \\mathbf {t} _ {\\alpha} & s _ {\\beta} \\mathbf {t} _ {\\beta} & \\mathbf {0} & \\boldsymbol {\\mu} \\\\ 0 & 0 & 0 & 1 \\end{array} \\right), \\tag {3}", + "image_path": "719360ce340f9d986450528cf7816d2d6328f628101014397399d6fd7b1b3c08.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27632" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "while the Gaussian kernel can be evaluated on a point " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "(\\alpha, \\beta)" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": " in splat space using:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 111, + 106, + 295, + 134 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 106, + 295, + 134 + ], + "spans": [ + { + "bbox": [ + 111, + 106, + 295, + 134 + ], + "type": "interline_equation", + "content": "\\mathcal {G} (\\alpha , \\beta) = \\exp \\left(- \\frac {\\alpha^ {2} + \\beta^ {2}}{2}\\right). \\tag {4}", + "image_path": "92b78add4a7442ba8a33a94d93635eace89613c04c22065e8d7984c24858f54c.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 143, + 141, + 154 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 143, + 141, + 154 + ], + "spans": [ + { + "bbox": [ + 55, + 143, + 141, + 154 + ], + "type": "text", + "content": "3.2.1. Rasterization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 158, + 296, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 158, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 296, + 205 + ], + "type": "text", + "content": "We render the Gaussians via " + }, + { + "bbox": [ + 55, + 158, + 296, + 205 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 158, + 296, + 205 + ], + "type": "text", + "content": "-blending as in [15, 20]. Specifically, for each pixel " + }, + { + "bbox": [ + 55, + 158, + 296, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 55, + 158, + 296, + 205 + ], + "type": "text", + "content": ", we integrate Gaussians from front to back to obtain the range " + }, + { + "bbox": [ + 55, + 158, + 296, + 205 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 158, + 296, + 205 + ], + "type": "text", + "content": ", normal " + }, + { + "bbox": [ + 55, + 158, + 296, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 55, + 158, + 296, + 205 + ], + "type": "text", + "content": ", and opacity " + }, + { + "bbox": [ + 55, + 158, + 296, + 205 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 55, + 158, + 296, + 205 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 110, + 214, + 295, + 249 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 214, + 295, + 249 + ], + "spans": [ + { + "bbox": [ + 110, + 214, + 295, + 249 + ], + "type": "interline_equation", + "content": "d = \\sum_ {i = 1} ^ {T} o _ {i} \\mathcal {G} _ {i} d _ {i} \\prod_ {j = 1} ^ {i - 1} \\left(1 - o _ {j} \\mathcal {G} _ {j}\\right) \\tag {5}", + "image_path": "546ae03bfafee5ac7c4678148a8b194af7209fed2c8d91dc577553b9233b0085.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 110, + 251, + 295, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 251, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 110, + 251, + 295, + 285 + ], + "type": "interline_equation", + "content": "\\mathbf {n} = \\sum_ {i = 1} ^ {T} o _ {i} \\mathcal {G} _ {i} \\mathbf {t} _ {n _ {i}} \\prod_ {j = 1} ^ {i - 1} \\left(1 - o _ {j} \\mathcal {G} _ {j}\\right) \\tag {6}", + "image_path": "a520510f92bea44486701ed564f4572fa61d5dab331226bee70adf3b70b22624.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 110, + 287, + 295, + 321 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 287, + 295, + 321 + ], + "spans": [ + { + "bbox": [ + 110, + 287, + 295, + 321 + ], + "type": "interline_equation", + "content": "o = \\sum_ {i = 1} ^ {T} o _ {i} \\mathcal {G} _ {i} \\prod_ {j = 1} ^ {i - 1} \\left(1 - o _ {j} \\mathcal {G} _ {j}\\right) \\tag {7}", + "image_path": "a88e75c7f94592d43a9d464ce3cea5984a5689df1205106088fbc9d16b259d3c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 331, + 296, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 296, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 296, + 390 + ], + "type": "text", + "content": "Different from the pinhole model, the spherical projection model described in Sec. 3.1 is non-affine, resulting in strong distortion artifacts on Gaussians that cover a significant extent in the image plane, if rasterizing using the local affine approximation method proposed in [20]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 392, + 296, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 296, + 463 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 296, + 463 + ], + "type": "text", + "content": "To this extent, as in [15], our rasterizer leverages exact ray-splat intersections to render 2D Gaussians. This approach is particularly suited for our projection model as it explicitly relies on per-pixel directions to evaluate the Gaussian kernels, thus allowing us to account for the structure of our projection manifold surface." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 464, + 296, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 296, + 548 + ], + "type": "text", + "content": "The ray-splat intersection is based on [45], and differently from [15], we reformulate the ray parametrization to cover the whole azimuthal space around the sensor. Additionally, we propose a method to compute the image-space extent of each Gaussian that is robust to the azimuthal coordinate singularity, ensuring accurate rasterization for any Gaussian configuration." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 555, + 178, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 178, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 178, + 567 + ], + "type": "text", + "content": "3.2.2. Ray-splat Intersection" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 570, + 296, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 596 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 596 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 570, + 296, + 596 + ], + "type": "inline_equation", + "content": "\\mathbf{v} = \\phi^{-1}\\left(\\mathbf{K}^{-1}\\mathbf{u}\\right)" + }, + { + "bbox": [ + 55, + 570, + 296, + 596 + ], + "type": "text", + "content": " be the normalized direction of pixel " + }, + { + "bbox": [ + 55, + 570, + 296, + 596 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 55, + 570, + 296, + 596 + ], + "type": "text", + "content": ", parametrized as the intersection of two orthogonal planes" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 88, + 605, + 295, + 631 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 605, + 295, + 631 + ], + "spans": [ + { + "bbox": [ + 88, + 605, + 295, + 631 + ], + "type": "interline_equation", + "content": "\\mathbf {h} _ {x} = \\frac {\\mathbf {v} \\times \\mathbf {u} _ {z}}{\\left\\| \\mathbf {v} \\times \\mathbf {u} _ {z} \\right\\|}, \\quad \\mathbf {h} _ {y} = \\mathbf {h} _ {x} \\times \\mathbf {v}, \\tag {8}", + "image_path": "3b372f39a849a1869d4401cc6bd07efcf5da70aad5bfb9370d62f614b5775f9b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 639, + 296, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 296, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 296, + 687 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 639, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_z" + }, + { + "bbox": [ + 55, + 639, + 296, + 687 + ], + "type": "text", + "content": " is the unit " + }, + { + "bbox": [ + 55, + 639, + 296, + 687 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 639, + 296, + 687 + ], + "type": "text", + "content": "-direction vector. This formulation is valid when " + }, + { + "bbox": [ + 55, + 639, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{v} \\neq \\mathbf{u}_z" + }, + { + "bbox": [ + 55, + 639, + 296, + 687 + ], + "type": "text", + "content": " which is a fair assumption in LiDAR measurements. Moreover, the planes are described in splat's space as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 78, + 695, + 296, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 695, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 78, + 695, + 296, + 712 + ], + "type": "interline_equation", + "content": "\\mathbf {h} _ {\\alpha} = \\left(\\mathbf {T} _ {w} ^ {c} \\mathbf {H}\\right) ^ {T} \\mathbf {h} _ {x}, \\quad \\mathbf {h} _ {\\beta} = \\left(\\mathbf {T} _ {w} ^ {c} \\mathbf {H}\\right) ^ {T} \\mathbf {h} _ {y}, \\quad (9)", + "image_path": "1ecdf8107a2d5fc989ef3c80c6e3e294e5b4867228162588b4db9dde368642c0.jpg" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 315, + 69, + 409, + 192 + ], + "blocks": [ + { + "bbox": [ + 315, + 69, + 409, + 192 + ], + "lines": [ + { + "bbox": [ + 315, + 69, + 409, + 192 + ], + "spans": [ + { + "bbox": [ + 315, + 69, + 409, + 192 + ], + "type": "image", + "image_path": "c949944ef033cc573c706a2bbb36afb53fc94b0fbfdd94b3d4e2b17e168fec96.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 414, + 70, + 555, + 209 + ], + "blocks": [ + { + "bbox": [ + 414, + 70, + 555, + 209 + ], + "lines": [ + { + "bbox": [ + 414, + 70, + 555, + 209 + ], + "spans": [ + { + "bbox": [ + 414, + 70, + 555, + 209 + ], + "type": "image", + "image_path": "cb55ea3b6bbc2d40e6e25a29a56faede3fab92dd57de1c208429805fe24bf5f9.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 212, + 555, + 322 + ], + "lines": [ + { + "bbox": [ + 313, + 212, + 555, + 322 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 555, + 322 + ], + "type": "text", + "content": "Figure 3. Bounding box computation for near-singularity splats. (a) shows the 3D configuration of a splat that approximately lies behind the camera. (b) shows the corresponding spherical image with the projected bounding box vertices. The distortion is removed by shifting the vertices along the horizontal direction to align the projected center to the image center. (c) being far from the coordinate singularity, we compute the maximum extent of the splat. (d) We revert the shift and propagate to the corresponding tiles via an offset from the central vertex, matching with the tiles highlighted on (a)." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 342, + 554, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 554, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 554, + 378 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 342, + 554, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_w^k\\in \\mathbb{S}\\mathbb{E}(3)" + }, + { + "bbox": [ + 313, + 342, + 554, + 378 + ], + "type": "text", + "content": " describes the world frame in the " + }, + { + "bbox": [ + 313, + 342, + 554, + 378 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 342, + 554, + 378 + ], + "type": "text", + "content": " -th sensor reference frame. We express the intersection of the two planes bundle as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 349, + 385, + 554, + 402 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 385, + 554, + 402 + ], + "spans": [ + { + "bbox": [ + 349, + 385, + 554, + 402 + ], + "type": "interline_equation", + "content": "\\mathbf {h} _ {\\alpha} (\\alpha , \\beta , 1, 1) ^ {T} = \\mathbf {h} _ {\\beta} (\\alpha , \\beta , 1, 1) ^ {T} = 0. \\tag {10}", + "image_path": "d2d333fa339c239bf343ad9f521dea779228ebb73316f1ae5ec400fe9ba7f496.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 409, + 554, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 554, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 554, + 432 + ], + "type": "text", + "content": "The solution " + }, + { + "bbox": [ + 313, + 409, + 554, + 432 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 409, + 554, + 432 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 409, + 554, + 432 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 409, + 554, + 432 + ], + "type": "text", + "content": " is then computed by solving the homogeneous linear system:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 319, + 437, + 554, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 437, + 554, + 466 + ], + "spans": [ + { + "bbox": [ + 319, + 437, + 554, + 466 + ], + "type": "interline_equation", + "content": "\\alpha \\left(\\mathbf {u} _ {s}\\right) = \\frac {\\mathbf {h} _ {\\alpha} ^ {2} \\mathbf {h} _ {\\beta} ^ {4} - \\mathbf {h} _ {\\alpha} ^ {4} \\mathbf {h} _ {\\beta} ^ {2}}{\\mathbf {h} _ {\\alpha} ^ {1} \\mathbf {h} _ {\\beta} ^ {2} - \\mathbf {h} _ {\\alpha} ^ {2} \\mathbf {h} _ {\\beta} ^ {1}}, \\quad \\beta \\left(\\mathbf {u} _ {s}\\right) = \\frac {\\mathbf {h} _ {\\alpha} ^ {4} \\mathbf {h} _ {\\beta} ^ {1} - \\mathbf {h} _ {\\alpha} ^ {1} \\mathbf {h} _ {\\beta} ^ {4}}{\\mathbf {h} _ {\\alpha} ^ {1} \\mathbf {h} _ {\\beta} ^ {2} - \\mathbf {h} _ {\\alpha} ^ {2} \\mathbf {h} _ {\\beta} ^ {1}}. \\tag {11}", + "image_path": "4a023a7c9550e137bd3b92f627a4d14e7e988744a6a813f97e74d2989a4cfbeb.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 472, + 463, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 463, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 463, + 483 + ], + "type": "text", + "content": "3.2.3. Bounding Box Computation" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 486, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 677 + ], + "type": "text", + "content": "To efficiently rasterize Gaussians, each tile is informed of which Gaussians it should process. This is achieved by computing the per-Gaussian image-space extent from its projected central point and using it to inform each covered tile to rasterize the Gaussian. In our work, we compute the per-Gaussian extent by projecting the vertices of the " + }, + { + "bbox": [ + 313, + 486, + 555, + 677 + ], + "type": "inline_equation", + "content": "3\\sigma" + }, + { + "bbox": [ + 313, + 486, + 555, + 677 + ], + "type": "text", + "content": " splat-space bounding box on the spherical frame. Due to the horizontal periodic nature of the spherical frame, for Gaussians whose center is projected near the horizontal image boundaries and has a significant image-space size, one or more vertices of the bounding box may be projected on the opposite side of the image. If not handled, this situation can degrade the rasterization since not all the corresponding tiles may be informed, thus skipping rendering for some pixels and slowing the rendering process since the computed extent might almost cover the whole image plane." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "We solve this by computing the Gaussian extent and tile propagation in a condition where the coordinate singularity cannot be harmful. As shown in Fig. 3, for each Gaussian," + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27633" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": "we project its bounding box and shift its projected points towards the center of the image plane while accounting for coordinate singularity. Similarly, after computing the Gaussian radius, the information is propagated to neighboring tiles while accounting for periodicity, ensuring correct rasterization and reducing the number of selected tiles." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 150, + 195, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 150, + 195, + 164 + ], + "spans": [ + { + "bbox": [ + 55, + 150, + 195, + 164 + ], + "type": "text", + "content": "3.3. Odometry And Mapping" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 167, + 296, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 167, + 296, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 296, + 288 + ], + "type": "text", + "content": "Following the modern literature of RGB-D SLAM [22, 34, 47, 53], we rely on keyframing to optimize local maps. We choose this approach for the following advantages: first, continuous integration over the same model can have adverse effects on artifacts and, more importantly, on runtime. Instead of decreasing the local density, we reset to a new local model if certain conditions are met, while also restricting the number of frames joining the optimization stage to allow effective online processing. Based on this, we define each local map as an individual Gaussian model " + }, + { + "bbox": [ + 55, + 167, + 296, + 288 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^s" + }, + { + "bbox": [ + 55, + 167, + 296, + 288 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 296, + 295, + 311 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 296, + 295, + 311 + ], + "spans": [ + { + "bbox": [ + 104, + 296, + 295, + 311 + ], + "type": "interline_equation", + "content": "\\mathbf {P} ^ {s} = \\left\\{\\mathcal {G} (\\boldsymbol {\\mu}, \\boldsymbol {\\Sigma}, o) | i = 1, \\dots , N \\right\\}. \\tag {12}", + "image_path": "b0ba12f643b724e7b07989e2261d5f95e59be25a849aadb54c0d28df9ec86e9d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 316, + 194, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 194, + 328 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 194, + 328 + ], + "type": "text", + "content": "3.3.1. Local model initialization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 331, + 296, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 296, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 296, + 381 + ], + "type": "text", + "content": "We initialize a new local model using the input LiDAR point cloud whenever necessary, such as at system startup or when visibility conditions require it. As a first step, we generate a valid pixel mask via the indicator function " + }, + { + "bbox": [ + 55, + 331, + 296, + 381 + ], + "type": "inline_equation", + "content": "\\mathbb{1}[\\cdot ]" + }, + { + "bbox": [ + 55, + 331, + 296, + 381 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 388, + 295, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 388, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 388, + 295, + 403 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {v} = \\mathbb {1} [ \\hat {\\mathbf {D}} > 0 ], \\tag {13}", + "image_path": "90103ebe6d70227e40b31c9a369396595a35975a5095c08fd5667799be6f1d78.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 409, + 296, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 296, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 296, + 482 + ], + "type": "text", + "content": "and compute the range gradients " + }, + { + "bbox": [ + 55, + 409, + 296, + 482 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{D}}" + }, + { + "bbox": [ + 55, + 409, + 296, + 482 + ], + "type": "text", + "content": " to construct a weight map over the range image. We use a weighted sampling of " + }, + { + "bbox": [ + 55, + 409, + 296, + 482 + ], + "type": "inline_equation", + "content": "n_d" + }, + { + "bbox": [ + 55, + 409, + 296, + 482 + ], + "type": "text", + "content": " points to prioritize complex regions. Splat positions are computed by back-projecting the range image, while their orientations are initialized by directing surface normals toward the sensor center to enhance initial visibility." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 487, + 189, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 189, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 189, + 498 + ], + "type": "text", + "content": "3.3.2. Local Model Refinement" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 502, + 295, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 295, + 658 + ], + "type": "text", + "content": "We perform a limited number of refinement iterations " + }, + { + "bbox": [ + 55, + 502, + 295, + 658 + ], + "type": "inline_equation", + "content": "n_o" + }, + { + "bbox": [ + 55, + 502, + 295, + 658 + ], + "type": "text", + "content": " on the most recent keyframes. Unlike [47, 53], we employ a geometric distribution-based sampling scheme that guarantees at least " + }, + { + "bbox": [ + 55, + 502, + 295, + 658 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 55, + 502, + 295, + 658 + ], + "type": "text", + "content": " probability of selecting the most recent keyframe and progressively decreases the likelihood of choosing the older ones. To filter out artifacts caused by ray-drop phenomena in the LiDAR measurements, we only consider valid pixels to refine the local model parameters " + }, + { + "bbox": [ + 55, + 502, + 295, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 502, + 295, + 658 + ], + "type": "text", + "content": ". We start by applying a densification strategy similar to the one described in Sec. 3.3.1, with two extra terms. The first one, " + }, + { + "bbox": [ + 55, + 502, + 295, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_n = \\mathbb{1}[\\mathbf{O}_i \\leq \\lambda_{d,o}]" + }, + { + "bbox": [ + 55, + 502, + 295, + 658 + ], + "type": "text", + "content": ", target newly discovered areas while the second, " + }, + { + "bbox": [ + 55, + 502, + 295, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_e = \\mathbb{1}[|\\mathbf{D}_i - \\hat{\\mathbf{D}}_i| \\geq \\lambda_{d,e}]" + }, + { + "bbox": [ + 55, + 502, + 295, + 658 + ], + "type": "text", + "content": ", target under-reconstructed regions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 658, + 295, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 295, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 295, + 683 + ], + "type": "text", + "content": "To optimize the geometric consistency of the local model, we employ a loss term that minimizes the " + }, + { + "bbox": [ + 55, + 658, + 295, + 683 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 55, + 658, + 295, + 683 + ], + "type": "text", + "content": " error:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 103, + 689, + 295, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 689, + 295, + 716 + ], + "spans": [ + { + "bbox": [ + 103, + 689, + 295, + 716 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d} = \\sum_ {\\mathbf {u} \\in \\mathbf {M} _ {v}} \\rho_ {d} \\| \\mathbf {D} (\\mathbf {u}, \\mathbf {x}) - \\hat {\\mathbf {D}} (\\mathbf {u}) \\|, \\tag {14}", + "image_path": "1a8871d4418476451ea530470c2d86de8a45c1371a6fedb27e83724ca6af901a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "inline_equation", + "content": "\\rho_{d}" + }, + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "text", + "content": " is a weight function dependent on the measurement's range. In addition, we employ a self-regularization term to align the splat's normals to the surface normals " + }, + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "inline_equation", + "content": "\\mathbf{N}" + }, + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "text", + "content": " estimated by the gradients of the range map " + }, + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "text", + "content": " [15]:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 366, + 127, + 555, + 153 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 127, + 555, + 153 + ], + "spans": [ + { + "bbox": [ + 366, + 127, + 555, + 153 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {n} = \\sum_ {\\mathbf {u} \\in \\mathbf {M} _ {v}} 1 - \\mathbf {n} ^ {T} \\mathbf {N} (\\mathbf {D} (\\mathbf {u}, \\mathbf {x})) \\tag {15}", + "image_path": "0498cc7f08c3fc79da43714586eca030ea0132e960da01ca471bf27e98cfa36d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 160, + 554, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 160, + 554, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 160, + 554, + 232 + ], + "type": "text", + "content": "Furthermore, to promote the expansion of splats over uniform surfaces, we introduce an additional term that operates on the opacity channel of the rasterized images. Specifically, we drive the splats to cover the areas of the image containing valid measurements by correlating the opacity image " + }, + { + "bbox": [ + 313, + 160, + 554, + 232 + ], + "type": "inline_equation", + "content": "\\mathbf{O}" + }, + { + "bbox": [ + 313, + 160, + 554, + 232 + ], + "type": "text", + "content": " with the valid mask " + }, + { + "bbox": [ + 313, + 160, + 554, + 232 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_v" + }, + { + "bbox": [ + 313, + 160, + 554, + 232 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 373, + 239, + 555, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 239, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 373, + 239, + 555, + 266 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {o} = \\sum_ {\\mathbf {u} \\in \\mathbf {M} _ {v}} - \\log (\\mathbf {O} (\\mathbf {u}, \\mathbf {x})). \\tag {16}", + "image_path": "8deef6883a42e28694647a35fc511915e77bb13e73cd0e3b26308e3ead227708.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 272, + 554, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 272, + 554, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 272, + 554, + 340 + ], + "type": "text", + "content": "While " + }, + { + "bbox": [ + 313, + 272, + 554, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_o" + }, + { + "bbox": [ + 313, + 272, + 554, + 340 + ], + "type": "text", + "content": " allows the splats to grow, it can also cause some splats to become extremely large in unobserved areas. We employ an additional regularization term on the scaling parameter of all primitives " + }, + { + "bbox": [ + 313, + 272, + 554, + 340 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 272, + 554, + 340 + ], + "type": "text", + "content": " to compensate for this effect. We propose a novel regularization that allows the splat to extend up to a certain value " + }, + { + "bbox": [ + 313, + 272, + 554, + 340 + ], + "type": "inline_equation", + "content": "\\tau_s" + }, + { + "bbox": [ + 313, + 272, + 554, + 340 + ], + "type": "text", + "content": " before penalizing its growth:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 326, + 345, + 555, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 345, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 326, + 345, + 555, + 369 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} = \\left\\{ \\begin{array}{c l} \\tau_ {s} - \\max \\left(s _ {\\alpha}, s _ {\\beta}\\right) & \\text {i f} \\max \\left(s _ {\\alpha}, s _ {\\beta}\\right) > \\tau_ {s} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {17}", + "image_path": "a58e3018d8fae14f27538d12aa20bddaef05c66f2c9acb3c1037589d3602ac96.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 374, + 554, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 554, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 554, + 434 + ], + "type": "text", + "content": "We found this term to be more effective rather than directly minimizing the deviation from the average [24, 47, 53] as it allows for anisotropic splats that are particularly useful for mapping details and edges, and provides more control on the density and structure of the model." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 326, + 434, + 531, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 434, + 531, + 447 + ], + "spans": [ + { + "bbox": [ + 326, + 434, + 531, + 447 + ], + "type": "text", + "content": "The final mapping objective function is defined as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 335, + 454, + 554, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 454, + 554, + 487 + ], + "spans": [ + { + "bbox": [ + 335, + 454, + 554, + 487 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a p}} = \\mathcal {L} _ {d} + \\lambda_ {o} \\mathcal {L} _ {o} + \\lambda_ {n} \\mathcal {L} _ {n} + \\lambda_ {s} \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {s _ {i}}, \\tag {18}", + "image_path": "26161039fb58ff9c0882a35a02b455529d54aad831f91f40fc707f6bdf979a3a.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 493, + 554, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 554, + 590 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 313, + 493, + 554, + 590 + ], + "type": "inline_equation", + "content": "\\lambda_{o}, \\lambda_{n}, \\lambda_{s} \\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 493, + 554, + 590 + ], + "type": "text", + "content": " being loss weights. This update strategy allows our method to maintain a geometrically accurate scene representation, while selectively densifying under-reconstructed regions that require more resolution, using a limited number of iterations. To avoid catastrophic forgetting, we do not perform any opacity reset steps. Moreover, we prune Gaussians that are either transparent or very small, as they typically cannot be observed during rendering." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 594, + 471, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 471, + 607 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 471, + 607 + ], + "type": "text", + "content": "3.3.3. Frame-To-Model Registration" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 609, + 554, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 554, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 554, + 693 + ], + "type": "text", + "content": "Each time a new keyframe is selected, we sample the local model by back-projecting the rasterized range image " + }, + { + "bbox": [ + 313, + 609, + 554, + 693 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 313, + 609, + 554, + 693 + ], + "type": "text", + "content": " onto a point cloud " + }, + { + "bbox": [ + 313, + 609, + 554, + 693 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{p}_q\\}_{q=1}^{W \\times H}" + }, + { + "bbox": [ + 313, + 609, + 554, + 693 + ], + "type": "text", + "content": " at its estimated pose. We design an ad-hoc tracking schema to benefit from both geometric and photometric consistencies provided by the Li-DAR and the rendered local model. Hence, the total odometry loss is composed of the sum of both residuals:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 382, + 701, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 701, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 382, + 701, + 555, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {o d o m}} = \\mathcal {L} _ {\\mathrm {g e o}} + \\mathcal {L} _ {\\mathrm {p h o t o}}. \\tag {19}", + "image_path": "e34fa3b8f4c92abfc056976daf7bb595d64bd868093de8a3fa5f9ef8acaf0fe1.jpg" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27634" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "text", + "content": "Geometric Registration. To associate geometric entities, we employ a PCA-based kd-tree [12]. The kd-tree is built on the back-projected rendered range map " + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{p}_q\\}_{q=1}^{W \\times H}" + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "text", + "content": " and segmented into tree leaves, corresponding to planar patches. Each leaf corresponds to " + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "inline_equation", + "content": "l = \\langle \\mathbf{p}_l, \\mathbf{n}_l \\rangle" + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "text", + "content": ", that is, the mean point " + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_l \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "text", + "content": " and the surface normal " + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_l \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "text", + "content": ". The geometric loss " + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{geo}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "text", + "content": " represents the sum of the point-to-plane distance between the mean leaf " + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_l" + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "text", + "content": " and the point of the current measurement point cloud " + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{q}_p\\}_{p=1}^Q" + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "text", + "content": ", along the normal " + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_l" + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "text", + "content": " expressed in the local reference frame " + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_w^k" + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "text", + "content": ". Specifically, we have:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 73, + 211, + 296, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 211, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 73, + 211, + 296, + 239 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {g e o}} = \\sum_ {p, q \\in \\{a \\}} \\rho \\left(\\left(\\mathbf {T} _ {w} ^ {k} \\mathbf {n} _ {l _ {q}}\\right) ^ {T} \\left(\\mathbf {T} _ {w} ^ {k} \\mathbf {q} _ {p} - \\mathbf {p} _ {q _ {i}}\\right)\\right), \\tag {20}", + "image_path": "7fb973e65ed0a6f33e74e066197d4c66952f20f8acf45bb6f0153a541cfb685c.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 243, + 296, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 296, + 281 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 243, + 296, + 281 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 243, + 296, + 281 + ], + "type": "text", + "content": " is the global frame, " + }, + { + "bbox": [ + 55, + 243, + 296, + 281 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 243, + 296, + 281 + ], + "type": "text", + "content": " is the local frame, " + }, + { + "bbox": [ + 55, + 243, + 296, + 281 + ], + "type": "inline_equation", + "content": "\\{a\\}" + }, + { + "bbox": [ + 55, + 243, + 296, + 281 + ], + "type": "text", + "content": " is the set of leaf associations with the point from the measurement point cloud, and " + }, + { + "bbox": [ + 55, + 243, + 296, + 281 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 55, + 243, + 296, + 281 + ], + "type": "text", + "content": " is the Huber robust loss function." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 295, + 296, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 296, + 353 + ], + "type": "text", + "content": "Photometric Registration. Leveraging the rendered range map " + }, + { + "bbox": [ + 55, + 295, + 296, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 55, + 295, + 296, + 353 + ], + "type": "text", + "content": ", we employ photometric registration for subpixel refinement, minimizing the photometric distance between the rendered and the spherical projected query point cloud " + }, + { + "bbox": [ + 55, + 295, + 296, + 353 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{D}}" + }, + { + "bbox": [ + 55, + 295, + 296, + 353 + ], + "type": "text", + "content": ". The photometric loss is formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 65, + 357, + 295, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 357, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 65, + 357, + 295, + 388 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {p h o t o}} = \\sum_ {\\mathbf {u}} \\left\\| \\rho \\left(\\mathbf {D} (\\mathbf {u}) - \\hat {\\mathbf {D}} \\underbrace {\\left(\\phi \\left(\\mathbf {T} _ {w} ^ {k} \\phi^ {- 1} (\\mathbf {u} , \\hat {d})\\right)\\right)} _ {\\mathbf {u} ^ {\\prime}}\\right) \\right\\| ^ {2}. \\tag {21}", + "image_path": "c8728b744cec7d23500b305748ba290db766495a4878c91e79725f68a6d1ae73.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 395, + 296, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 395, + 296, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 395, + 296, + 467 + ], + "type": "text", + "content": "The evaluation point " + }, + { + "bbox": [ + 55, + 395, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{u}'" + }, + { + "bbox": [ + 55, + 395, + 296, + 467 + ], + "type": "text", + "content": " in the query image " + }, + { + "bbox": [ + 55, + 395, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{D}}" + }, + { + "bbox": [ + 55, + 395, + 296, + 467 + ], + "type": "text", + "content": " is computed by first back-projecting the pixel " + }, + { + "bbox": [ + 55, + 395, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 55, + 395, + 296, + 467 + ], + "type": "text", + "content": ", applying the transform " + }, + { + "bbox": [ + 55, + 395, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_w^k" + }, + { + "bbox": [ + 55, + 395, + 296, + 467 + ], + "type": "text", + "content": ", and then projecting it back. Pose updates " + }, + { + "bbox": [ + 55, + 395, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 395, + 296, + 467 + ], + "type": "text", + "content": " are parameterized as local updates in the Lie algebra " + }, + { + "bbox": [ + 55, + 395, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\mathfrak{se}(3)" + }, + { + "bbox": [ + 55, + 395, + 296, + 467 + ], + "type": "text", + "content": ". Therefore, the transformation " + }, + { + "bbox": [ + 55, + 395, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_k^w" + }, + { + "bbox": [ + 55, + 395, + 296, + 467 + ], + "type": "text", + "content": " of the local reference frame, expressed in the global reference frame, is updated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 135, + 473, + 295, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 473, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 135, + 473, + 295, + 487 + ], + "type": "interline_equation", + "content": "\\mathbf {T} _ {k} ^ {w} \\leftarrow \\mathbf {T} _ {k} ^ {w} \\exp (\\delta). \\tag {22}", + "image_path": "e94239ea0f1f79ace2a6b10ba420f259d6a6c7a8c88ba8355340d66a6d24ee73.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 493, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 296, + 529 + ], + "type": "text", + "content": "This update is carried out using a second-order GaussNewton method. Local updates ensure that rotation updates are well-defined [11]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 538, + 137, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 137, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 137, + 551 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 558, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 676 + ], + "type": "text", + "content": "In this section, we report the results of our method on several publicly available datasets. We evaluate our pipeline on tracking and mapping. We recall that, to our knowledge, this is the first Gaussian Splatting LiDAR Odometry and Mapping pipeline, and no direct competitors are available. We compared our approach with other well-known geometric and neural implicit methods. The experiments were executed on a PC with an Intel Core i9-14900K @ 3.20Ghz, 64GB of RAM, and an NVIDIA RTX 4090 GPU with 24 GB of VRAM." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "For odometry evaluation, we include several baselines. The first one is a basic point-to-plane ICP odometry, followed by SLAMesh, that simultaneously estimates a mesh" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "representation of the scene via Gaussian Processes and performs registration onto it [32]. Moreover, MAD-ICP leverages a forest of PCA-based KD-Trees to perform accurate registration. Furthermore, we include PIN-SLAM as SOTA baseline for implicit LiDAR SLAM [30]. It relies on neural point primitives and interleaves an incremental learning of the model's SDF and a correspondence-free, point-to-implicit registration schema." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 168, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 555, + 361 + ], + "type": "text", + "content": "For mapping evaluation, we include OpenVDB [27] and VoxBlox [29] as \"classic\" baselines. OpenVDB provides a robust volumetric data structure to handle 3D Points, while VoxBlox combines adaptive weights and grouped ray-casting for an efficient and accurate Truncated Signed Distance Function (TSDF) integration. Additionally, we include as neural-implicit approaches, " + }, + { + "bbox": [ + 313, + 168, + 555, + 361 + ], + "type": "inline_equation", + "content": "N^3" + }, + { + "bbox": [ + 313, + 168, + 555, + 361 + ], + "type": "text", + "content": "-Mapping [39] and PIN-SLAM [30]. " + }, + { + "bbox": [ + 313, + 168, + 555, + 361 + ], + "type": "inline_equation", + "content": "N^3" + }, + { + "bbox": [ + 313, + 168, + 555, + 361 + ], + "type": "text", + "content": "-Mapping is a neural-implicit non-projective SDF mapping approach that leverages normal guidance to produce more accurate SDFs, leading to SOTA results for offline LiDAR mapping. Moreover, PIN-SLAM is included as meshes can be extracted directly from the implicit representation via marching cubes. Below, we report the datasets and the evaluation metrics employed. We do not include SLAMesh [32] since it could not be run with ground-truth poses." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 370, + 376, + 381 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 370, + 376, + 381 + ], + "spans": [ + { + "bbox": [ + 314, + 370, + 376, + 381 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 388, + 537, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 388, + 537, + 399 + ], + "spans": [ + { + "bbox": [ + 314, + 388, + 537, + 399 + ], + "type": "text", + "content": "We used the following four publicly available datasets:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 400, + 554, + 524 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 315, + 400, + 554, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 400, + 554, + 460 + ], + "spans": [ + { + "bbox": [ + 315, + 400, + 554, + 460 + ], + "type": "text", + "content": "- Newer College Dataset (NC) [50]: Collected with a handheld Ouster OS0-128 LiDAR in structured and vegetated areas. Ground truth was generated using the Leica BLK360 scanner, achieving centimeter-level accuracy over poses and points in the map." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 464, + 554, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 464, + 554, + 524 + ], + "spans": [ + { + "bbox": [ + 315, + 464, + 554, + 524 + ], + "type": "text", + "content": "- A Vision Benchmark in Rome (VBR) [4]: Recorded in Rome using OS1-64 (car-mounted) and OS0-128 (hand-held) LiDARs, capturing large-scale urban scenarios with narrow streets and dynamic objects. Ground truth trajectories were obtained by fusing LiDAR, IMU, and RTK" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 318, + 546, + 550, + 670 + ], + "blocks": [ + { + "bbox": [ + 318, + 546, + 550, + 670 + ], + "lines": [ + { + "bbox": [ + 318, + 546, + 550, + 670 + ], + "spans": [ + { + "bbox": [ + 318, + 546, + 550, + 670 + ], + "type": "image", + "image_path": "fca29fcdfcbbabff88fd491a1ed51f2d78802e34d1760e18bd3389b1aa213bc2.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 677, + 554, + 711 + ], + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 711 + ], + "type": "text", + "content": "Figure 4. RPE evaluation. Number of successful sequences across RPE thresholds. It includes the sequences of Newer College [49], VBR [4], Oxford Spires [41] and Mai City [43]." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27635" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 70, + 529, + 168 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 529, + 168 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 529, + 168 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 529, + 168 + ], + "type": "table", + "html": "
DatasetNewer College[50]Oxford Spires[41]Mai-City[43]
quad-easymath-easykeble-02bodleian-02observatory-01mai-01mai-02
ApproachC-11↓F-score↑C-11↓F-score↑C-11↓F-score↑C-11↓F-score↑C-11↓F-score↑C-11↓F-score↑C-11↓F-score↑
OpenVDB[27]7.9288.8511.6984.517.1991.747.5189.689.5986.163.3397.293.2697.37
VoxBlox[29]16.564.6311.9380.5115.0371.6315.2458.7715.1270.457.0493.815.8095.51
N3-Mapping[39]8.0494.54failfail7.0193.477.8990.369.3587.942.6499.062.7699.04
PIN-SLAM[30]12.8988.0514.6973.8211.8379.6510.7482.7114.4972.314.9193.844.8294.11
Ours5.3796.749.1590.027.4394.417.6090.0910.5683.043.6497.275.0595.31
", + "image_path": "817655b460b2639ad0f1bc9a8fe87a4850f7c9929e1d4b333d56c6fa8d689f46.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 172, + 555, + 195 + ], + "lines": [ + { + "bbox": [ + 55, + 172, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 55, + 172, + 555, + 195 + ], + "type": "text", + "content": "Table 1. Reconstruction evaluation. The pipelines were run with ground-truth poses. Voxel size is " + }, + { + "bbox": [ + 55, + 172, + 555, + 195 + ], + "type": "inline_equation", + "content": "{20}\\mathrm{\\;{cm}}" + }, + { + "bbox": [ + 55, + 172, + 555, + 195 + ], + "type": "text", + "content": " and F-score is computed with " + }, + { + "bbox": [ + 55, + 172, + 555, + 195 + ], + "type": "inline_equation", + "content": "{20}\\mathrm{\\;{cm}}" + }, + { + "bbox": [ + 55, + 172, + 555, + 195 + ], + "type": "text", + "content": " threshold. Splat-LOAM yield competitive mapping performance on both the Newer College[50] and Oxford Spires[41] dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 63, + 216, + 295, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 216, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 63, + 216, + 295, + 251 + ], + "type": "text", + "content": "GNSS data, ensuring centimeter-level accuracy over the poses. This dataset is not used for mapping evaluation as no ground-truth maps are provided." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 255, + 295, + 450 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 255, + 295, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 295, + 374 + ], + "type": "text", + "content": "- Oxford Spires [41]: Recorded with a hand-held Hesai QT64 LiDAR featuring a " + }, + { + "bbox": [ + 55, + 255, + 295, + 374 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 55, + 255, + 295, + 374 + ], + "type": "text", + "content": " horizontal FoV, " + }, + { + "bbox": [ + 55, + 255, + 295, + 374 + ], + "type": "inline_equation", + "content": "104^{\\circ}" + }, + { + "bbox": [ + 55, + 255, + 295, + 374 + ], + "type": "text", + "content": " vertical FoV, 64 vertical channels, and 60 meters of range detection. Similar to [49], each sequence includes a prior map obtained via a survey-grade 3D imaging laser scanner, used for ground-truth trajectory estimation and mapping evaluation. Specifically, we choose the Keble College, Bodleian Library, and Radcliffe Observatory sequences to include both indoor and outdoor scenarios with different levels of detail." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 379, + 295, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 295, + 450 + ], + "type": "text", + "content": "- Mai City [43]: A synthetic dataset captured using a car-like simulated LiDAR with 120 meters of range detection. The measurements are generated via ray-casting on an underlying mesh, providing error-free, motion-undistorted data. We select the " + }, + { + "bbox": [ + 55, + 379, + 295, + 450 + ], + "type": "inline_equation", + "content": "01" + }, + { + "bbox": [ + 55, + 379, + 295, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 379, + 295, + 450 + ], + "type": "inline_equation", + "content": "02" + }, + { + "bbox": [ + 55, + 379, + 295, + 450 + ], + "type": "text", + "content": " sequences, which capture similar scenarios with different vertical resolutions." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 457, + 129, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 457, + 129, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 129, + 468 + ], + "type": "text", + "content": "4.2. Evaluation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 475, + 295, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 295, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 295, + 522 + ], + "type": "text", + "content": "We use Relative Pose Error (RPE) computed with progressively increasing delta steps to evaluate the odometry accuracy. Specifically, we adapt the deltas to the trajectory length to provide a more meaningful result [4]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 523, + 295, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 523, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 523, + 295, + 630 + ], + "type": "text", + "content": "Differently, to evaluate mapping quality, we use several metrics [25]: Accuracy (Acc) measures the mean distance of points on the estimated mesh with their nearest neighbors on the reference cloud. Completeness (Comp) measures the opposite distance, and Chamfer-11 (C-11) describes the mean of the two. Additionally, we use the F-score computed with " + }, + { + "bbox": [ + 55, + 523, + 295, + 630 + ], + "type": "inline_equation", + "content": "20\\mathrm{cm}" + }, + { + "bbox": [ + 55, + 523, + 295, + 630 + ], + "type": "text", + "content": " error threshold. Table 1 reports compact results for C-11 and F1-score, additional tables are available in the Supplementary Material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 636, + 149, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 149, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 149, + 649 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "content": "Our approach employs Gaussian primitives for LiDAR odometry estimation and mapping, yielding results comparable to state-of-the-art methods while significantly enhancing computational efficiency. In this section, we evaluate some key aspects of our pipeline and evaluate their" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 217, + 372, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 217, + 372, + 226 + ], + "spans": [ + { + "bbox": [ + 314, + 217, + 372, + 226 + ], + "type": "text", + "content": "contributions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 243, + 555, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 362 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 362 + ], + "type": "text", + "content": "Memory and Runtime Analysis. In Fig. 5, we report how the increment of active primitives affects the active GPU memory requirements and the per-iteration mapping frequency. It shows an experiment run over the large-scale campus sequence [4] where we set a maximum of 100 keyframes per local map and sample, at most, " + }, + { + "bbox": [ + 313, + 243, + 555, + 362 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 243, + 555, + 362 + ], + "type": "text", + "content": " of points for the incoming point cloud. It's possible to notice that the mapping FPS remains stable between 200k and 300k primitives. This is most likely related to the saturation of the rasterizer." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 332, + 375, + 537, + 494 + ], + "blocks": [ + { + "bbox": [ + 332, + 375, + 537, + 494 + ], + "lines": [ + { + "bbox": [ + 332, + 375, + 537, + 494 + ], + "spans": [ + { + "bbox": [ + 332, + 375, + 537, + 494 + ], + "type": "image", + "image_path": "6f1dd620c0146ede4fd189a3037f4501e87349e6880d43a9e754bfece47c38e4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 503, + 555, + 548 + ], + "lines": [ + { + "bbox": [ + 313, + 503, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 555, + 548 + ], + "type": "text", + "content": "Figure 5. Memory and Runtime Analysis. The plot relates the used GPU Memory and the mapping iteration frequency with the number of active primitives. The measurements were obtained over the longest sequence we reported: campus [4]." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 566, + 554, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 566, + 554, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 566, + 554, + 650 + ], + "type": "text", + "content": "Odometry. Fig. 7 shows the results for different tracking methods over our scene representation. Using both geometric and photometric components, we achieve a better result than using point-to-point or point-to-plane. The last three bars show an ablation of geometric and photometric loss components. The results are best for the joint use of both terms which support our design choice." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": "Mesh generation. We can sample the Gaussian scene to produce a mesh representation. Similar to [13], the meshing process involves sampling " + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "inline_equation", + "content": "n_m" + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": " points from each keyframe rendered range and normal maps, and running Poisson reconstruction [18]. As shown in Tab. 2, we found" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "27636" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 83, + 179, + 153 + ], + "blocks": [ + { + "bbox": [ + 92, + 70, + 144, + 80 + ], + "lines": [ + { + "bbox": [ + 92, + 70, + 144, + 80 + ], + "spans": [ + { + "bbox": [ + 92, + 70, + 144, + 80 + ], + "type": "text", + "content": "PIN-SLAM" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 83, + 179, + 153 + ], + "lines": [ + { + "bbox": [ + 57, + 83, + 179, + 153 + ], + "spans": [ + { + "bbox": [ + 57, + 83, + 179, + 153 + ], + "type": "image", + "image_path": "a47a6ccec38933cc3ff53dd986481cd0c763ed9714a02013cba541374b10ec3f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 181, + 83, + 304, + 153 + ], + "blocks": [ + { + "bbox": [ + 214, + 70, + 271, + 83 + ], + "lines": [ + { + "bbox": [ + 214, + 70, + 271, + 83 + ], + "spans": [ + { + "bbox": [ + 214, + 70, + 271, + 83 + ], + "type": "inline_equation", + "content": "\\mathbf{N}^3" + }, + { + "bbox": [ + 214, + 70, + 271, + 83 + ], + "type": "text", + "content": " -Mapping" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 181, + 83, + 304, + 153 + ], + "lines": [ + { + "bbox": [ + 181, + 83, + 304, + 153 + ], + "spans": [ + { + "bbox": [ + 181, + 83, + 304, + 153 + ], + "type": "image", + "image_path": "1f8e480cb538ed36666c8be4e68f4247029873dbaba7e73832a1020c7f064854.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 307, + 83, + 429, + 152 + ], + "blocks": [ + { + "bbox": [ + 356, + 71, + 380, + 81 + ], + "lines": [ + { + "bbox": [ + 356, + 71, + 380, + 81 + ], + "spans": [ + { + "bbox": [ + 356, + 71, + 380, + 81 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 83, + 429, + 152 + ], + "lines": [ + { + "bbox": [ + 307, + 83, + 429, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 83, + 429, + 152 + ], + "type": "image", + "image_path": "c3e7e666ea73a21b76ea9b7c989bce04d8c19abf7ae21be458852c99ade76367.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 431, + 83, + 553, + 152 + ], + "blocks": [ + { + "bbox": [ + 484, + 70, + 500, + 80 + ], + "lines": [ + { + "bbox": [ + 484, + 70, + 500, + 80 + ], + "spans": [ + { + "bbox": [ + 484, + 70, + 500, + 80 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 431, + 83, + 553, + 152 + ], + "lines": [ + { + "bbox": [ + 431, + 83, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 431, + 83, + 553, + 152 + ], + "type": "image", + "image_path": "93fefcc024af56629cf061fb2c87a54030a581e8c2e3fcb1be70387a71eb2862.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 57, + 154, + 179, + 224 + ], + "blocks": [ + { + "bbox": [ + 57, + 154, + 179, + 224 + ], + "lines": [ + { + "bbox": [ + 57, + 154, + 179, + 224 + ], + "spans": [ + { + "bbox": [ + 57, + 154, + 179, + 224 + ], + "type": "image", + "image_path": "06cff9a1acd60fe66d5720707149a8fbaf1393ce209f470ef8707be1d508de2d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 181, + 154, + 304, + 224 + ], + "blocks": [ + { + "bbox": [ + 181, + 154, + 304, + 224 + ], + "lines": [ + { + "bbox": [ + 181, + 154, + 304, + 224 + ], + "spans": [ + { + "bbox": [ + 181, + 154, + 304, + 224 + ], + "type": "image", + "image_path": "4b1389376e896f0a53bf765ed5595b392257f0be201800b3a3fc4e56c94aa456.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 154, + 429, + 224 + ], + "blocks": [ + { + "bbox": [ + 307, + 154, + 429, + 224 + ], + "lines": [ + { + "bbox": [ + 307, + 154, + 429, + 224 + ], + "spans": [ + { + "bbox": [ + 307, + 154, + 429, + 224 + ], + "type": "image", + "image_path": "c0ecd9bf88726ed22e365266e2089e86efd5dfe471faa64b095a5a0d911497f1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 431, + 154, + 553, + 223 + ], + "blocks": [ + { + "bbox": [ + 431, + 154, + 553, + 223 + ], + "lines": [ + { + "bbox": [ + 431, + 154, + 553, + 223 + ], + "spans": [ + { + "bbox": [ + 431, + 154, + 553, + 223 + ], + "type": "image", + "image_path": "19bd142339bf312ce8748194350edeabee2683bdf20337e59f28a84c62cddbf0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 56, + 225, + 179, + 294 + ], + "blocks": [ + { + "bbox": [ + 56, + 225, + 179, + 294 + ], + "lines": [ + { + "bbox": [ + 56, + 225, + 179, + 294 + ], + "spans": [ + { + "bbox": [ + 56, + 225, + 179, + 294 + ], + "type": "image", + "image_path": "0c39cbab10d5fb386e1d75c9028f56ad31a424c01935b8d4eb0be133c52cbae9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 298, + 555, + 331 + ], + "lines": [ + { + "bbox": [ + 55, + 298, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 555, + 331 + ], + "type": "text", + "content": "Figure 6. Comparison of Mesh Reconstruction. The figure shows reconstruction results for quad-easy sequence from the newer college dataset. Our method recovers a geometry with much higher data fidelity. PIN-SLAM lacks many details and exhibits a large level of noise. " + }, + { + "bbox": [ + 55, + 298, + 555, + 331 + ], + "type": "inline_equation", + "content": "N^3" + }, + { + "bbox": [ + 55, + 298, + 555, + 331 + ], + "type": "text", + "content": "-Mapping performs more similar to ours, but oversmoothes fine geometric details." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 181, + 224, + 304, + 294 + ], + "blocks": [ + { + "bbox": [ + 181, + 224, + 304, + 294 + ], + "lines": [ + { + "bbox": [ + 181, + 224, + 304, + 294 + ], + "spans": [ + { + "bbox": [ + 181, + 224, + 304, + 294 + ], + "type": "image", + "image_path": "205ad13a24b0e70972c290a0bc262b26c492a9c4d6330a034615c5b4c15cfbe1.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 307, + 224, + 429, + 294 + ], + "blocks": [ + { + "bbox": [ + 307, + 224, + 429, + 294 + ], + "lines": [ + { + "bbox": [ + 307, + 224, + 429, + 294 + ], + "spans": [ + { + "bbox": [ + 307, + 224, + 429, + 294 + ], + "type": "image", + "image_path": "d9d7c9816b7d2a45a1125af51c52fa18af51828882f975071df7da6d812f965a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 431, + 224, + 553, + 294 + ], + "blocks": [ + { + "bbox": [ + 431, + 224, + 553, + 294 + ], + "lines": [ + { + "bbox": [ + 431, + 224, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 431, + 224, + 553, + 294 + ], + "type": "image", + "image_path": "e1bf2831635340691d4a5b09610f706800ef7ba227b4d3d2ad5ed628513cbd10.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 61, + 354, + 291, + 489 + ], + "blocks": [ + { + "bbox": [ + 61, + 354, + 291, + 489 + ], + "lines": [ + { + "bbox": [ + 61, + 354, + 291, + 489 + ], + "spans": [ + { + "bbox": [ + 61, + 354, + 291, + 489 + ], + "type": "image", + "image_path": "ed8fbcb19e771faa56bf43822a3a2a2e891feab34d32f99e001066088a5137de.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 491, + 295, + 536 + ], + "lines": [ + { + "bbox": [ + 55, + 491, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 295, + 536 + ], + "type": "text", + "content": "Figure 7. Ablation on registration methods. The plot reports the RPE (\\%) of several tested registration methods on the quad-easy sequence [50]. Enabling both geometric and photometric factors in sequence, provides a more robust estimate." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 559, + 296, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 296, + 631 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 296, + 631 + ], + "type": "text", + "content": "that this method produces consistently better results than directly running TSDF integration and marching cubes over the rendered frames, or by running Poisson reconstruction directly on the Gaussian centers. Specifically, these method produces worse results in the regions that overlap multiple local models." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 644, + 128, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 644, + 128, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 644, + 128, + 657 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "We present the first LiDAR Odometry and Mapping pipeline that leverages 2D Gaussian primitives as the sole scene representation. Through an ad-hoc tile-based Gaussian rasterizer for spherical images, we leverage LiDAR" + } + ] + } + ], + "index": 21 + }, + { + "type": "table", + "bbox": [ + 317, + 350, + 552, + 401 + ], + "blocks": [ + { + "bbox": [ + 317, + 350, + 552, + 401 + ], + "lines": [ + { + "bbox": [ + 317, + 350, + 552, + 401 + ], + "spans": [ + { + "bbox": [ + 317, + 350, + 552, + 401 + ], + "type": "table", + "html": "
Extraction MethodAcc↓Com↓C-11↓F-score↑
Marching Cubes [27]16.765.5311.1476.76
Poisson (centers)10.156.708.4392.33
Ours6.644.095.3796.74
", + "image_path": "b2c2d044c03949fb10cf59629b9a4a44912cb0cad8a52ab25ceb27707c352d69.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 404, + 555, + 439 + ], + "lines": [ + { + "bbox": [ + 313, + 404, + 555, + 439 + ], + "spans": [ + { + "bbox": [ + 313, + 404, + 555, + 439 + ], + "type": "text", + "content": "Table 2. Ablation on Meshing Methods. We report mapping results with varying meshing methods on the quad-easy sequence [50]." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 471, + 555, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 555, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 555, + 554 + ], + "type": "text", + "content": "measurements to optimize the local model. Furthermore, we demonstrate the effectiveness of combining a geometric and photometric tracker to register new LiDAR point clouds over the Gaussian local model. The experiments show that our pipeline obtains tracking and mapping scores that are on par with the current SOTA at a fraction of the computational demands." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 563, + 554, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 554, + 637 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 554, + 637 + ], + "type": "text", + "content": "Future Work. We plan on improving Splat-LOAM by simultaneously estimating the sensor pose and velocity to compensate the rolling-shutter distortions caused by motion while acquisition, typical of the LiDAR measurements. Moreover, we plan on including Loop Closure (LC) to improve the pose estimates, along with the mapping accuracy." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 314, + 663, + 411, + 677 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 663, + 411, + 677 + ], + "spans": [ + { + "bbox": [ + 314, + 663, + 411, + 677 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 689, + 534, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 534, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 534, + 712 + ], + "type": "text", + "content": "This work has been supported by PNRR MUR project PE0000013-FAIR." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27637" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Jens Behley and Cyril Stachniss. Efficient Surfel-Based SLAM using 3D Laser Range Data in Urban Environments. In Proc. of Robotics: Science and Systems (RSS). Robotics: Science and Systems Foundation, 2018. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 158 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 158 + ], + "type": "text", + "content": "[2] P.J. Besl and Neil D. McKay. A method for registration of 3-D shapes. IEEE TPAMI, 14(2):239-256, 1992. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 159, + 295, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 159, + 295, + 203 + ], + "spans": [ + { + "bbox": [ + 62, + 159, + 295, + 203 + ], + "type": "text", + "content": "[3] Dorit Borrmann, Jan Elseberg, Kai Lingemann, Andreas Nüchter, and Joachim Hertzberg. Globally consistent 3D mapping with scan matching. Journal on Robotics and Autonomous Systems (RAS), 56(2):130-142, 2008. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 205, + 296, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 296, + 259 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 296, + 259 + ], + "type": "text", + "content": "[4] Leonardo Brizi, Emanuele Giacomini, Luca Di Giammarino, Simone Ferrari, Omar Salem, Lorenzo De Rebotti, and Giorgio Grisetti. VBR: A Vision Benchmark in Rome. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), pages 15868-15874, 2024. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 262, + 295, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 262, + 295, + 294 + ], + "spans": [ + { + "bbox": [ + 62, + 262, + 295, + 294 + ], + "type": "text", + "content": "[5] Qifeng Chen, Sheng Yang, Sicong Du, Tao Tang, Peng Chen, and Yuchi Huo. LiDAR-GS:Real-time LiDAR Re-Simulation using Gaussian Splitting, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 296, + 295, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 296, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 296, + 295, + 350 + ], + "type": "text", + "content": "[6] Pinxuan Dai, Jiamin Xu, Wenxiang Xie, Xinguo Liu, Huamin Wang, and Weiwei Xu. High-quality Surface Reconstruction using Gaussian Surfels. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, New York, NY, USA, 2024. Association for Computing Machinery. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 352, + 295, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 352, + 295, + 406 + ], + "spans": [ + { + "bbox": [ + 62, + 352, + 295, + 406 + ], + "type": "text", + "content": "[7] Pierre Dellenbach, Jean-Emmanuel Deschaud, Bastien Jacquet, and François Goulette. CT-ICP: Real-time Elastic LiDAR Odometry with Loop Closure. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), pages 5580–5586, 2022. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 408, + 295, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 408, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 62, + 408, + 295, + 462 + ], + "type": "text", + "content": "[8] Junyuan Deng, Qi Wu, Xieyuanli Chen, Songpengcheng Xia, Zhen Sun, Guoqing Liu, Wenxian Yu, and Ling Pei. NeRF-LOAM: Neural Implicit Representation for Large-Scale Incremental LiDAR Odometry and Mapping. In ICCV, pages 8184–8193, 2023. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 464, + 295, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 464, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 62, + 464, + 295, + 518 + ], + "type": "text", + "content": "[9] Luca Di Giammarino, Leonardo Brizi, Tiziano Guadagnino, Cyrill Stachniss, and Giorgio Grisetti. MD-SLAM: Multicue Direct SLAM. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), pages 11047-11054, 2022. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 521, + 295, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 521, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 57, + 521, + 295, + 564 + ], + "type": "text", + "content": "[10] Luca Di Giammarino, Emanuele Giacomini, Leonardo Brizi, Omar Salem, and Giorgio Grisetti. Photometric LiDAR and RGB-D Bundle Adjustment. IEEE Robotics and Automation Letters (RA-L), 8(7):4362-4369, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 567, + 295, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 295, + 597 + ], + "type": "text", + "content": "[11] Chris Engels, Henrik Stewénius, and David Nister. Bundle Adjustment Rules. Photogrammetric computer vision, 2(32), 2006. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 601, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 295, + 644 + ], + "type": "text", + "content": "[12] Simone Ferrari, Luca Di Giammarino, Leonardo Brizi, and Giorgio Grisetti. MAD-ICP: It is All About Matching Data – Robust and Informed LiDAR Odometry. IEEE Robotics and Automation Letters (RA-L), 9(11):9175–9182, 2024. 1, 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "text", + "content": "[13] Antoine Guédon and Vincent Lepetit. SuGaR: Surface-Aligned Gaussian Splating for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering. In CVPR, pages 5354-5363, 2024. 2, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "text", + "content": "[14] Sheng Hong, Junjie He, Xinhu Zheng, and Chunran Zheng. LIV-GaussMap: LiDAR-Inertial-Visual Fusion for Real" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "Time 3D Radiance Field Map Rendering. IEEE Robotics and Automation Letters (RA-L), 9(11):9765-9772, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 96, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 149 + ], + "type": "text", + "content": "[15] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2D Gaussian Splatting for Geometrically Accurate Radiance Fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, New York, NY, USA, 2024. Association for Computing Machinery. 2, 3, 4, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 151, + 553, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 553, + 204 + ], + "type": "text", + "content": "[16] Seth Isaacson, Pou-Chun Kung, Mani Ramanagopal, Ram Vasudevan, and Katherine A. Skinner. LONER: LiDAR Only Neural Representations for Real-Time SLAM. IEEE Robotics and Automation Letters (RA-L), 8(12):8042-8049, 2023. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 206, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 237 + ], + "type": "text", + "content": "[17] Junzhe Jiang, Chun Gu, Yurui Chen, and Li Zhang. GS-LiDAR: Generating Realistic LiDAR Point Clouds with Panoramic Gaussian Splatting. In ICLR, 2025. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "text", + "content": "[18] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the Fourth Eurographics Symposium on Geometry Processing, pages 61-70, Goslar, DEU, 2006. Eurographics Association. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 283, + 553, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 346 + ], + "type": "text", + "content": "[19] Nikhil Keetha, Jay Karhade, Krishna Murthy Jatavallabhula, Gengshan Yang, Sebastian Scherer, Deva Ramanan, and Jonathon Luiten. Splatam: Splat, track and map 3d gaussians for dense rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "text", + "content": "[20] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM TOG, 42(4):139:1-139:14, 2023. 1, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 393, + 553, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 435 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 435 + ], + "type": "text", + "content": "[21] Qing Li, Shaoyang Chen, Cheng Wang, Xin Li, Chenglu Wen, Ming Cheng, and Jonathan Li. LO-Net: Deep Real-Time Lidar Odometry. In CVPR, pages 8465–8474, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 437, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 480 + ], + "type": "text", + "content": "[22] Lorenzo Liso, Erik Sandström, Vladimir Yugay, Luc Van Gool, and Martin R. Oswald. Loopy-SLAM: Dense Neural SLAM with Loop Closures. In CVPR, pages 20363-20373, 2024. 1, 2, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "type": "text", + "content": "[23] Hidenobu Matsuki, Riku Murai, Paul HJ Kelly, and Andrew J Davison. Gaussian splatting slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18039-18048, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 526, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 557 + ], + "type": "text", + "content": "[24] Hidenobu Matsuki, Riku Murai, Paul H. J. Kelly, and Andrew J. Davison. Gaussian Splitting SLAM. In CVPR, pages 18039-18048, 2024. 2, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 559, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 601 + ], + "type": "text", + "content": "[25] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In CVPR, 2019. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 603, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 658 + ], + "type": "text", + "content": "[26] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In ECCV, pages 405-421, Cham, 2020. Springer International Publishing. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "text", + "content": "[27] Ken Museth. VDB: High-resolution sparse volumes with dynamic topology. ACM TOG, 32(3):27:1-27:22, 2013. 6, 7, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[28] Austin Nicolai, Ryan Steele, Christopher Eriksen, and Geoffrey A. Hollinger. Deep Learning for Laser Based Odometry" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27638" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "Estimation. In RSS workshop Limits and Potentials of Deep Learning in Robotics, page 1, 2016. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "type": "text", + "content": "[29] Helen Oleynikova, Zachary Taylor, Marius Fehr, Roland Siegwart, and Juan Nieto. Voxblox: Incremental 3D Euclidean Signed Distance Fields for on-board MAV planning. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), pages 1366-1373, 2017. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "type": "text", + "content": "[30] Yue Pan, Xingguang Zhong, Louis Wiesmann, Thorbjörn Posewsky, Jens Behley, and Cyril Stachniss. PIN-SLAM: LiDAR SLAM Using a Point-Based Implicit Neural Representation for Achieving Global Map Consistency. IEEE Trans. on Robotics, 40:4045-4064, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 259 + ], + "type": "text", + "content": "[31] Jan Quenzel and Sven Behnke. Real-time Multi-Adaptive-Resolution-Surfel 6D LiDAR Odometry using Continuous-time Trajectory Optimization. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), pages 5499–5506, 2021. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "type": "text", + "content": "[32] Jianyuan Ruan, Bo Li, Yibo Wang, and Yuxiang Sun. SLAMesh: Real-time LiDAR Simultaneous Localization and Meshing. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), pages 3546-3552, 2023. 1, 2, 6, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 304, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 294, + 337 + ], + "type": "text", + "content": "[33] Erik Sandström, Yue Li, Luc Van Gool, and Martin R. Oswald. Point-SLAM: Dense Neural Point Cloud-based SLAM. In ICCV, pages 18387-18398, 2023. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 380 + ], + "type": "text", + "content": "[34] Erik Sandström, Keisuke Tateno, Michael Oechsle, Michael Niemeyer, Luc Van Gool, Martin R. Oswald, and Federico Tombari. Splat-SLAM: Globally Optimized RGB-only SLAM with 3D Gaussians, 2024. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 382, + 294, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 294, + 413 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 294, + 413 + ], + "type": "text", + "content": "[35] Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-Motion Revisited. In CVPR, pages 4104-4113, 2016. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 415, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 459 + ], + "type": "text", + "content": "[36] Johannes L. Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise View Selection for Unstructured Multi-View Stereo. In ECCV, pages 501–518, Cham, 2016. Springer International Publishing. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 460, + 294, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 492 + ], + "type": "text", + "content": "[37] A. Segal, D. Haehnel, and S. Thrun. Generalized-ICP. In Proc. of Robotics: Science and Systems (RSS). Robotics: Science and Systems Foundation, 2009. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 493, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 546 + ], + "type": "text", + "content": "[38] Tixiao Shan and Brendan Englot. LeGO-LOAM: Lightweight and Ground-Optimized Lidar Odometry and Mapping on Variable Terrain. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), pages 4758-4765, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "type": "text", + "content": "[39] Shuangfu Song, Junqiao Zhao, Kai Huang, Jiaye Lin, Chen Ye, and Tiantian Feng. " + }, + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{N}^3" + }, + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "type": "text", + "content": "-Mapping: Normal Guided Neural Non-Projective Signed Distance Fields for Large-Scale 3D Mapping. IEEE Robotics and Automation Letters (RA-L), 9 (6):5935-5942, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 603, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 635 + ], + "type": "text", + "content": "[40] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J. Davison. iMAP: Implicit Mapping and Positioning in Real-Time. In ICCV, pages 6209-6218, 2021. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "type": "text", + "content": "[41] Yifu Tao, Miguel Ángel Muñoz-Bañón, Lintong Zhang, Jiāhao Wang, Lanke Frank Tarimo Fu, and Maurice Fallon. The Oxford Spires Dataset: Benchmarking Large-Scale LiDAR-Visual Localisation, Reconstruction and Radiance Field Methods, 2024. 6, 7, 3, 4, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[42] Fabio Tosi, Youmin Zhang, Ziren Gong, Erik Sandström, Stefano Mattoccia, Martin R. Oswald, and Matteo Poggi." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 654 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "How nerfs and 3d gaussian splatting are reshaping slam: a survey, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "text", + "content": "[43] Ignacio Vizzo, Xieyuanli Chen, Nived Chebrolu, Jens Behley, and Cyril Stachniss. Poisson Surface Reconstruction for LiDAR Odometry and Mapping. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), pages 5624-5630, 2021. 6, 7, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[44] Guangming Wang, Xinrui Wu, Zhe Liu, and Hesheng Wang. PWCLO-Net: Deep LiDAR Odometry in 3D Point Clouds Using Hierarchical Embedding Mask Optimization. In CVPR, pages 15905–15914, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "type": "text", + "content": "[45] Tim Weyrich, Simon Heinzle, Timo Aila, Daniel B. Fasnacht, Stephan Oetiker, Mario Botsch, Cyril Flaig, Simon Mall, Kaspar Rohrer, Norbert Felber, Hubert Kaeslin, and Markus Gross. A hardware architecture for surface splattering. ACM TOG, 26(3):90-es, 2007. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 253, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 317 + ], + "type": "text", + "content": "[46] Chenyang Wu, Yifan Duan, Xinran Zhang, Yu Sheng, Jianmin Ji, and Yanyong Zhang. MM-Gaussian: 3D Gaussian-based Multi-modal Fusion for Localization and Reconstruction in Unbounded Scenes. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), pages 12287-12293, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 319, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 352 + ], + "type": "text", + "content": "[47] Vladimir Yugay, Yue Li, Theo Gevers, and Martin R. Oswald. Gaussian-SLAM: Photo-realistic Dense SLAM with Gaussian Splatting, 2024. 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "type": "text", + "content": "[48] Ganlin Zhang, Erik Sandström, Youmin Zhang, Manthan Patel, Luc Van Gool, and Martin R Oswald. Glorie-slam: Globally optimized rgb-only implicit encoding point cloud slam. arXiv preprint arXiv:2403.19549, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 398, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 398, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 398, + 553, + 441 + ], + "type": "text", + "content": "[49] Ji Zhang and Sanjiv Singh. LOAM: Lidar Odometry and Mapping in Real-time. In Proc. of Robotics: Science and Systems (RSS). Robotics: Science and Systems Foundation, 2014. 2, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 443, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 443, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 443, + 553, + 475 + ], + "type": "text", + "content": "[50] Lintong Zhang, Marco Camurri, David Wisth, and Maurice Fallon. Multi-Camera LiDAR Inertial Extension to the Newer College Dataset, 2021. 6, 7, 8, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 476, + 553, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 476, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 316, + 476, + 553, + 509 + ], + "type": "text", + "content": "[51] Xin Zheng and Jianke Zhu. Efficient LiDAR Odometry for Autonomous Driving. IEEE Robotics and Automation Letters (RA-L), 6(4):8458-8465, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 510, + 553, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 553, + 564 + ], + "type": "text", + "content": "[52] Xingguang Zhong, Yue Pan, Jens Behley, and Cyril Stachniss. SHINE-Mapping: Large-Scale 3D Mapping Using Sparse Hierarchical Implicit Neural Representations. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), pages 8371-8377, 2023. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 566, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 566, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 566, + 553, + 609 + ], + "type": "text", + "content": "[53] Liyuan Zhu, Yue Li, Erik Sandstrom, Shengyu Huang, Konrad Schindler, and Iro Armeni. LoopSplat: Loop Closure by Registering 3D Gaussian Splats. In Proc. of the International Conference on 3D Vision (3DV), 2025. 2, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 611, + 553, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 611, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 316, + 611, + 553, + 654 + ], + "type": "text", + "content": "[54] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Pollefeys. NICE-SLAM: Neural Implicit Scalable Encoding for SLAM. In CVPR, pages 12776-12786, 2022. 1, 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27639" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Splat-based 3D Scene Reconstruction with Extreme Motion-blur/73a5b991-b304-4700-8590-8334fbf2bdee_content_list.json b/2025/Splat-based 3D Scene Reconstruction with Extreme Motion-blur/73a5b991-b304-4700-8590-8334fbf2bdee_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0c111f71f7e6d57943d7b3979bc3353aec387214 --- /dev/null +++ b/2025/Splat-based 3D Scene Reconstruction with Extreme Motion-blur/73a5b991-b304-4700-8590-8334fbf2bdee_content_list.json @@ -0,0 +1,1703 @@ +[ + { + "type": "text", + "text": "Splat-based 3D Scene Reconstruction with Extreme Motion-blur", + "text_level": 1, + "bbox": [ + 173, + 130, + 823, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hyeonjoong Jang†*", + "bbox": [ + 119, + 178, + 276, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dongyoung Choi† \n† KAIST", + "bbox": [ + 297, + 180, + 442, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Donggun Kim $^{\\dagger}$ Woohyun Kang $^{\\dagger}$ \n* HYPERGRAM", + "bbox": [ + 464, + 180, + 738, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Min H. Kim†*", + "bbox": [ + 761, + 180, + 874, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{hjjang; dychoi; dgkim; whkang; minhkim}@vclab.kaist.ac.kr", + "bbox": [ + 240, + 219, + 751, + 234 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8148e79216c51fa76cfd8cd39f2083c8c2e187dfed867ff6e83d7798d0aeb7a1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 257, + 254, + 327 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3364c044bf5b03f3d65916a7cfcaa64f0813a1a389d52f3f42e8489bc2aae960.jpg", + "image_caption": [ + "Input", + "out" + ], + "image_footnote": [], + "bbox": [ + 98, + 328, + 176, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2028d66bd913728fe7c1763ca132e1251ce9c1f26dd9dfad91350bbe78d58980.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 259, + 257, + 416, + 328 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0cd06fcab1244798ddc1c096785dbe6f73dc615a4a6e5a401b04bf88573d6e35.jpg", + "image_caption": [ + "ElasticFusion", + "Figure 1. We propose a robust 3D scene reconstruction method from RGB-D input that effectively addresses extreme motion blur. Our approach achieves accurate camera poses and dense point clouds, producing clearer, deblurred scenes compared to existing methods." + ], + "image_footnote": [], + "bbox": [ + 259, + 328, + 336, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2dd837bf8b329554cdfcb60a391ce455e5829ef9b8d9d9ecfe14f723b1e732be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 338, + 328, + 416, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0736a2bd1733fe9528015726d26bd4aadf52fe4862261af60a90203859754934.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 257, + 576, + 328 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6d12761ac801f1e68e66ce23de016c5453509695c66e3030268357845f31825a.jpg", + "image_caption": [ + "DROID-SLAM" + ], + "image_footnote": [], + "bbox": [ + 419, + 328, + 576, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dab89590977b7d2f0ffab195557d5456ff3b9e05450a21199942c3d1bd7837c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 580, + 257, + 736, + 328 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/98007c7640395b4a7ea0fd334de03eee117587749d63b69d2917e0865e8b0982.jpg", + "image_caption": [ + "MonoGS" + ], + "image_footnote": [], + "bbox": [ + 580, + 328, + 684, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d227d0808ee008285d54552f1ad14a28d5b7af7f569d4cfff1f5f9b5366ae8fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 740, + 257, + 898, + 328 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eeb28c3f72a0685f986a2b85bf64539fd8f7ebb4ad6fd0ceec92291c16c1356f.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 740, + 328, + 818, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9bd8095922966743884d2b6d3357fcf3d227469ec8b33f13417f011982375a8b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 820, + 328, + 898, + 388 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 441, + 326, + 458 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose a splat-based 3D scene reconstruction method from RGB-D input that effectively handles extreme motion blur, a frequent challenge in low-light environments. Under dim illumination, RGB frames often suffer from severe motion blur due to extended exposure times, causing traditional camera pose estimation methods, such as COLMAP, to fail. This results in inaccurate camera pose and blurry color input, compromising the quality of 3D reconstructions. Although recent 3D reconstruction techniques like Neural Radiance Fields and Gaussian Splatting have demonstrated impressive results, they rely on accurate camera trajectory estimation, which becomes challenging under fast motion or poor lighting conditions. Furthermore, rapid camera movement and the limited field of view of depth sensors reduce point cloud overlap, limiting the effectiveness of pose estimation with the ICP algorithm. To address these issues, we introduce a method that combines camera pose estimation and image deblurring using a Gaussian Splatting framework, leveraging both 3D Gaussian splats and depth inputs for enhanced scene representation. Our method first aligns consecutive RGB-D frames through optical flow and ICP, then refines camera poses and 3D geometry by adjusting Gaussian positions for optimal depth alignment. To handle motion blur, we model camera movement during exposure and deblur images by comparing the input with a series of sharp, rendered frames. Experiments on a new RGB-D dataset with extreme motion blur show that our method outperforms existing approaches,", + "bbox": [ + 88, + 477, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "enabling high-quality reconstructions even in challenging conditions. This approach has broad implications for 3D mapping applications in robotics, autonomous navigation, and augmented reality. Both code and dataset are publicly available on https://github.com/KAIST-VCLAB/gs-extreme-motion-blur.", + "bbox": [ + 511, + 444, + 908, + 536 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 569, + 643, + 584 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "High-quality 3D scene reconstruction is one of the most important and challenging applications in computer vision. The accuracy of 3D reconstruction hinges on the quality of essential components, such as camera poses, RGB images, and depth maps. These elements are interconnected and influence each other; a failure in one component can adversely impact the others. For example, blurry RGB images or noisy depth maps can significantly impair camera pose estimation, which, in turn, degrades overall reconstruction quality. This challenge is particularly pronounced when input data is captured in low-light conditions or during rapid camera movement—common scenarios in everyday, casual capture. Such conditions result in degraded color and depth frames, leading to poorly estimated camera poses and reconstructions with blurry textures and distorted or smoothed geometry. Therefore, achieving high-quality 3D reconstruction requires clear RGB images and accurate depth maps.", + "bbox": [ + 509, + 595, + 906, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The interdependence of these input components suggests an opportunity for compensating or restoring degraded elements. However, this process presents a chicken-and-egg", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26425", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "problem, complicating the reconstruction further. For instance, accurate camera pose estimation requires sharp images, while deblurring motion-blurred images depend on reliable camera pose information. Although numerous methods [4, 41] exist for recovering sharp images from motion-blurred ones, they often rely on supervised learning models trained on specific datasets. Consequently, their deblurring performance can degrade significantly when faced with new cameras or motion blur scenarios that differ from the training data.", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent works [35, 44, 45, 49] have made advances in addressing these challenges by optimizing camera trajectories during exposure while simultaneously learning sharp RGB colors. This enables simultaneous camera pose estimation and image deblurring. However, these methods require an initial camera pose and, when using Gaussian Splatting [14], also need a sparse point cloud as input. This requirement limits their applicability, especially when severe motion blur prevents structure-from-motion (SfM) methods like COLMAP [31] from functioning effectively, making it impossible to start the optimization process.", + "bbox": [ + 89, + 244, + 480, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome these limitations, we introduce a robust approach for 3D scene reconstruction that effectively compensates for severe motion blur and addresses the challenges of camera pose estimation without relying on a precise initial pose. Our method leverages RGB-D inputs and incorporates both optical flow and depth information to align camera poses accurately, even in the presence of challenging motion blur and lighting conditions. We first perform a global alignment between consecutive frames using optical flow and the ICP algorithm, which enables effective local alignment of point clouds generated from the depth maps.", + "bbox": [ + 89, + 412, + 480, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "After this initial alignment, we refine the camera poses and 3D geometry by integrating them into a Gaussian Splating pipeline. This approach allows us to initialize dense 3D Gaussians from depth maps, which we scale to ensure a detailed representation of the scene geometry. Our refinement process iteratively adjusts both camera poses and the positions of 3D Gaussians, achieving global consistency by minimizing a depth alignment loss that compares rendered depth maps with input depth measurements. Through this adjustment, we eliminate loop-closure artifacts and reduce the drift that often accumulates in traditional SLAM-based methods.", + "bbox": [ + 89, + 580, + 480, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For scenes with significant motion blur, we further optimize the deblurring process by modeling the camera poses at the start and end of each frame's exposure. This temporal modeling allows us to simulate the effects of motion during the exposure period, which we incorporate into our Gaussian Splatting framework. We minimize the difference between the observed motion-blurred image and the average of a set of sharp images rendered from multiple viewpoints along the exposure trajectory, resulting in more accurate,", + "bbox": [ + 89, + 763, + 480, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "geometrically consistent deblurring. Our deblur loss function combines image alignment, structural similarity, and depth consistency, ensuring that the final reconstructions retain sharpness and fidelity to the original scene structure.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Through extensive evaluation on a newly constructed RGB-D dataset featuring extreme motion blur, we demonstrate that our approach significantly outperforms existing methods in both accuracy and robustness. By addressing the key challenges of pose estimation, depth alignment, and image deblurring, our method provides a versatile and effective solution for high-quality 3D reconstruction under real-world capture conditions, such as low light and fast motion. The proposed method holds promising applications in areas requiring reliable 3D mapping and reconstruction, including robotics, autonomous navigation, and augmented reality. We will make both our dataset and implementation publicly available to foster further research and development in this field.", + "bbox": [ + 511, + 152, + 903, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 383, + 653, + 398 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "RGB-D scene reconstruction. A popular approach to 3D geometry reconstruction defines a 3D voxel grid and constructs a signed distance function (SDF) volume by accumulating depth estimates captured by sensors. Kinect-Fusion [27] uses a Kinect camera to capture RGB-D images, then estimates relative camera poses between adjacent frames with the ICP algorithm. Next, it constructs and updates a truncated signed distance function (TSDF) volume, enabling real-time 3D reconstruction of objects. Niessner et al. [28] improve memory efficiency by combining hashing algorithms to store and update SDF values only where needed, thus reducing memory requirements.", + "bbox": [ + 511, + 412, + 903, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several methods [8, 11, 20, 22] adopt memory-efficient, hierarchical data structures for depth fusion. BundleFusion [6] simultaneously refines camera poses and geometry using bundle adjustment, improving reconstruction accuracy and addressing the loop closure problem caused by accumulated camera pose and depth estimation errors. ElasticFusion [47] builds a map with surfels—point primitives containing position, normal, and radius—and updates the map through a deformation graph to minimize error and perform loop closure. Other methods [13, 33, 38] directly fuse depth estimates into point clouds, optimizing memory use. However, these methods generally require slow camera motion to capture sharp images and accurately estimate camera transformations.", + "bbox": [ + 511, + 597, + 903, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Simultaneous localization and mapping (SLAM) methods [3, 15, 25, 32, 46] track camera motion and construct a map in real-time by extracting visual features from input images to establish 3D point correspondences. CodeSLAM [2] and DeepTAM [50] use neural networks to combine depth maps with color images, while Azinović", + "bbox": [ + 511, + 809, + 903, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26426", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c45ed40390f47755f0dc6308204674dc878c1ccd30a154ffa616bf7d1259fb64.jpg", + "image_caption": [ + "Figure 2. Overview of our method pipeline. We begin with global alignment by comparing optical flow and reprojected pixel differences between adjacent RGB-D image pairs, followed by local estimation of relative camera poses using ICP. Each depth map and camera pose is then used to initialize 3D Gaussians. By fixing the scale of the Gaussians and optimizing with depth and opacity losses, we achieve global refinement of camera poses and geometry. Finally, we deblur the scene texture by minimizing the deblur loss between the input image and averaged virtual images." + ], + "image_footnote": [], + "bbox": [ + 153, + 65, + 843, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "et al. [1] train multi-layer perceptrons (MLP) to learn SDF and color values for scene reconstruction. Their approach also optimizes camera transformation correction variables during training to refine poses. However, these methods struggle with input images that contain motion blur. Nice-SLAM [51], Point-SLAM [30], and DROID-SLAM [42] use neural networks to track camera poses. SplaTAM [12] and MonoGS [21] introduced SLAM solutions based on 3D Gaussian Splatting. However, these methods are not designed to handle motion blur.", + "bbox": [ + 88, + 311, + 485, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image deblurring. Motion blur occurs when camera motion during the exposure time changes the projected pixel coordinates of rays on the camera sensor. Early studies [7, 16, 36] tackle this issue by developing kernels to restore blurry images. Convolutional neural networks (CNNs) demonstrate powerful performance in learning deblur kernels [34, 40], and deep learning techniques such as ResNet with skip connections and multi-scale networks further improved restoration quality [26, 39]. SRN-DeblurNet [41] adopt recurrent networks [10, 37] and combine them with an encoder-decoder network structure, while NAFNet [4] simplify the network structure, extracting only essential components and proposing a nonlinear activation-free network. However, since these methods rely on curated datasets like the GoPro dataset [26], they struggle to effectively deblur images with extreme motion blur or those captured on different cameras.", + "bbox": [ + 91, + 472, + 482, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There are NeRF-based approaches [24] that tackle motion blur in input images. Deblur-NeRF [19] and DP-NeRF [17] use a set of motion-blurred images along with camera poses estimated by COLMAP as input to restore sharp images through NeRF optimization. BAD-NeRF [44] enhances this by jointly optimizing virtual camera poses and radiance fields during exposure. The latest approaches utilize Gaussian Splatting [14], which offers faster training times than NeRF by rendering a set of 3D Gaussians with a dedicated rasterizer rather than optimizing a neural network. Gaussian Splatting-based deblurring meth", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ods [35, 45, 49] optimize the virtual camera trajectory and minimize differences between input blurry images and the average of rendered images at each virtual camera position.", + "bbox": [ + 511, + 311, + 903, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "All of these methods rely on initial camera poses, typically obtained from COLMAP, and for Gaussian Splatting-based methods, a sparse point cloud as well. However, structure-from-motion methods like COLMAP often fail with blurry inputs—such as images captured by fast-moving cameras or in low-light conditions that require long exposure times. Even with dense depth inputs, ICP requires acceptable global alignment, and recent point cloud matching methods [9, 18] cannot accurately determine relative camera transformations between two point clouds if they consist mainly of planar structures.", + "bbox": [ + 511, + 358, + 905, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address these issues, we propose an effective camera pose estimation algorithm tailored for RGB-D image sequences that jointly refines both the camera trajectory and the reconstructed 3D point cloud, resulting in improved sharpness in 3D reconstruction.", + "bbox": [ + 511, + 525, + 905, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 614, + 604, + 630 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. In datasets with significant motion blur, using COLMAP to estimate camera poses becomes impractical. Additionally, rapid camera motion and the limited field of view of depth sensors further complicate this task, making ICP registration less effective. To address these challenges, we first perform global alignment using both RGB and depth images, which establishes sufficient initial alignment for the ICP algorithm to operate effectively. After achieving this global alignment, we apply ICP again to refine and accurately estimate the camera poses. However, estimating camera poses by analyzing only two consecutive frames in a sequence can result in accumulated errors, leading to drift over time. To resolve this issue, it is crucial to adjust the 3D geometry and align the camera poses across the entire sequence so that depth maps and camera views from all perspectives remain well-aligned. We accomplish this through bundle adjustment, optimizing the positions of", + "bbox": [ + 511, + 643, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26427", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/528a4657a2fdc0e9e33f5463e6b9ae45155ec1baac2556b41a81a71da06a5775.jpg", + "image_caption": [ + "Figure 3. The illustration of our pose initialization loss. When $\\mathbf{p}_i$ is reprojected to $\\mathbf{p}_j$ using the depth value, it is considered geometrically consistent if $\\overline{\\mathbf{P}_i\\mathbf{H}_i}$ equals $\\overline{\\mathbf{P}_j\\mathbf{H}_j}$ ." + ], + "image_footnote": [], + "bbox": [ + 140, + 78, + 433, + 207 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D Gaussians and camera poses within the Gaussian Splating pipeline. Finally, to account for motion blur, we approximate the camera trajectory over time and optimize the movement during exposure, enabling effective deblurring. See Figure 2 for an overview.", + "bbox": [ + 89, + 267, + 483, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Local Pose Estimation", + "text_level": 1, + "bbox": [ + 89, + 357, + 299, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To estimate relative camera transformations, we select consecutive frames from the input RGB-D image sequence. Our approach begins with global alignment using optical flow, followed by local alignment of the two-point clouds using the ICP algorithm. For each consecutive RGB image and depth map pair at times $t$ and $t + 1$ , we optimize the transformation $\\xi \\in \\mathfrak{se}(3)$ between the frames. Here, $\\xi_{t\\rightarrow t + 1} = [\\mathbf{x}^{\\intercal},\\boldsymbol{\\omega}^{\\intercal}]^{\\intercal}$ , where $\\mathbf{x}$ represents the translation vector between the two camera origins, and $\\boldsymbol{\\omega}$ is the rotation element of $\\mathfrak{se}(3)$ . We ultimately calculate a camera-to-world transformation matrix $\\xi_t$ for each timestamp $t$ .", + "bbox": [ + 89, + 381, + 483, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To achieve this, we first compute the difference between the estimated optical flow and the reprojected pixel from the camera at time $t$ to $t + 1$ using the depth estimate for each pixel in the depth map $D_{t}$ and the intrinsic matrix $\\mathbf{K}$ . A pixel $\\mathbf{p}_i$ in the depth map is backprojected to 3D space as: $\\mathbf{P}_i = \\pi^{-1}(\\mathbf{K},\\mathbf{p}_i,D_t(\\mathbf{p}_i))$ . We then project this 3D point to the camera at $t + 1$ : $\\mathbf{p_j} = \\pi (\\mathbf{K},\\exp (\\xi_{t\\to t + 1}^{\\wedge})\\mathbf{P}_i)$ , where $\\pi$ and $\\pi^{-1}$ represent the projection and backprojection operators, respectively. The pixel movement from $\\mathbf{p}_i$ to $\\mathbf{p}_j$ is then compared with the estimated optical flow $F_{t\\rightarrow t + 1}(\\mathbf{p}_i)$ to compute the optical flow loss:", + "bbox": [ + 89, + 549, + 483, + 717 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {F} = \\sum_ {i} M _ {t} (\\mathbf {p} _ {i}) M _ {t + 1} (\\mathbf {p} _ {j}) \\| F _ {t \\to t + 1} (\\mathbf {p} _ {i}) - (\\mathbf {p} _ {j} - \\mathbf {p} _ {i}) \\| _ {2}.\n$$\n", + "text_format": "latex", + "bbox": [ + 93, + 729, + 480, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this formulation, $M_{t}$ is a mask indicating valid depth values in $D_{t}$ at pixel $\\mathbf{p}_{i}$ . The mask $M_{t}$ dynamically updates during optimization, as $\\mathbf{p}_{j}$ depends on the transformation $\\xi$ : $M_{t}(\\mathbf{p}_{i}) = D_{t}(\\mathbf{p}_{i}) > 0$ .", + "bbox": [ + 89, + 777, + 483, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We employ a state-of-the-art pretrained optical flow estimation model [48] designed hierarchically to produce reliable optical flow even from two blurry input images. Next, we calculate the photometric loss between color values at", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "pixel $\\mathbf{p}_i$ in image $I_{t}$ and pixel $\\mathbf{p}_j$ in image $I_{t + 1}$ :", + "bbox": [ + 511, + 90, + 836, + 107 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {I} = \\sum_ {i} M _ {t} (\\mathbf {p} _ {i}) M _ {t + 1} (\\mathbf {p} _ {j}) \\| I _ {t} (\\mathbf {p} _ {i}) - I _ {t + 1} (\\mathbf {p} _ {j}) \\| _ {1}.\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 118, + 879, + 150 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To ensure geometric consistency, we introduce a depth consistency loss by using the two depth maps captured by the depth camera to optimize relative poses. Specifically, as illustrated in Figure 3, we compute and compare the vertical components $\\overline{\\mathbf{P}_i\\mathbf{H}_i}$ and $\\overline{\\mathbf{P}_j\\mathbf{H}_j}$ based on the baseline vector $\\overline{\\mathbf{O}_t\\mathbf{O}_{t + 1}}$ between the two cameras. We calculate $\\overline{\\mathbf{P}_i\\mathbf{H}_i}$ as follows:", + "bbox": [ + 511, + 159, + 906, + 265 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{\\mathbf {P} _ {i} \\mathbf {H} _ {i}}} = \\| \\mathbf {P} _ {i} \\| _ {2} \\sin \\left(\\cos^ {- 1} \\left(\\frac {\\mathbf {P} _ {i} \\cdot \\mathbf {x}}{\\| \\mathbf {P} _ {i} \\| _ {2} \\| \\mathbf {x} \\| _ {2}}\\right)\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 272, + 864, + 309 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarly, we compute $\\overline{\\mathbf{P}_j\\mathbf{H}_j}$ as:", + "bbox": [ + 511, + 318, + 735, + 335 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{\\mathbf {P} _ {j} \\mathbf {H} _ {j}}} = \\| \\mathbf {P} _ {j} \\| _ {2} \\sin \\left(\\cos^ {- 1} \\left(\\frac {\\mathbf {P} _ {j} \\cdot (- \\exp (\\boldsymbol {\\omega}) ^ {\\intercal} \\mathbf {x})}{\\| \\mathbf {P} _ {j} \\| _ {2} \\| - \\exp (\\boldsymbol {\\omega}) ^ {\\intercal} \\mathbf {x} \\| _ {2}}\\right)\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 345, + 895, + 377 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This depth consistency loss is calculated as follows:", + "bbox": [ + 511, + 383, + 859, + 398 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C} = \\sum_ {i} M _ {t} (\\mathbf {p} _ {i}) M _ {t + 1} (\\mathbf {p} _ {j}) \\| \\overline {{\\mathbf {P} _ {i} \\mathbf {H} _ {i}}} - \\overline {{\\mathbf {P} _ {j} \\mathbf {H} _ {j}}} \\|, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 409, + 906, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\exp()$ is the exponential mapping operator in Lie algebra, which helps maintain alignment. Our final loss function combines these terms with weighted coefficients:", + "bbox": [ + 511, + 452, + 905, + 497 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p o s e}} = \\mathcal {L} _ {F} + \\lambda_ {I} \\mathcal {L} _ {I} + \\lambda_ {C} \\mathcal {L} _ {C}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 508, + 906, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We start optimization with $\\lambda_I = 0$ and $\\lambda_C = 0$ , allowing only the flow loss $\\mathcal{L}_F$ to guide the relative pose alignment with global optical flow. After initial convergence, we increase weights to $\\lambda_I = 75$ and $\\lambda_C = 25$ and continue optimization to a second convergence.", + "bbox": [ + 511, + 537, + 905, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we apply a point-to-plane ICP algorithm on the two globally aligned point clouds to estimate camera poses accurately. Since optical flow estimation between two blurry images introduces some errors, this final camera pose estimation relies exclusively on depth maps.", + "bbox": [ + 511, + 613, + 905, + 689 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Global Pose Estimation/Geometry Refinement", + "text_level": 1, + "bbox": [ + 511, + 696, + 903, + 713 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Starting from the initial camera poses and point clouds obtained in the previous step, we merge them into a global point cloud within the world coordinate system (Figure 4). To manage memory efficiently, we downsample the point cloud with a specified voxel size $s$ . However, due to accumulated errors in the initial camera poses, these poses may not align well with the global point cloud, leading to seam artifacts at image boundaries and a drifting phenomenon that creates loop-closure issues. To resolve this, we refine both the camera poses and 3D point cloud positions simultaneously within the Gaussian Splatting pipeline, ensuring global and geometric consistency.", + "bbox": [ + 511, + 719, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26428", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/240642e034bbd39f7e39698d52f74ef4049cb02e47960e7012347066845330f2.jpg", + "image_caption": [ + "Input color image" + ], + "image_footnote": [], + "bbox": [ + 94, + 78, + 287, + 164 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/eaa9e2f5231815953c24155779cac99f8f10632e06766e37a17deffd12748646.jpg", + "image_caption": [ + "Input depth map" + ], + "image_footnote": [], + "bbox": [ + 287, + 78, + 478, + 164 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ff9a09ca1da2b85b480d6029603136c185d8538391ba56c461d748063bea0af6.jpg", + "image_caption": [ + "DROID-SLAM" + ], + "image_footnote": [], + "bbox": [ + 98, + 176, + 218, + 236 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7cb32aaf4b47cb9da9c1a0a64fae1c50678584e4607a84f0760ee431aec63553.jpg", + "image_caption": [ + "MonoGS", + "Figure 4. Reconstructed initial point clouds from the RGB-D image sequence. Our method achieves accurate camera poses and generates a dense, high-quality point cloud." + ], + "image_footnote": [], + "bbox": [ + 222, + 176, + 348, + 239 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8bd656b686752171a2fd64d2b71b711c74e1be9d51d7b89072b949957a7913ac.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 348, + 178, + 478, + 238 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We initialize the Gaussian Splatting pipeline by loading the point cloud along with the estimated camera poses. To capture every depth value per pixel from the dense depth maps, we initialize 3D Gaussians, fixing their sizes and setting their scale to $s / 2$ . This dense arrangement of 3D Gaussians provides a detailed geometric representation. We then train only the camera poses, 3D positions, and opacities of the Gaussians by comparing a rendered depth map $\\widetilde{D}$ with the input depth map $D$ .", + "bbox": [ + 89, + 310, + 483, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training, we randomly select camera viewpoints from among the trainable camera pose variables to render each depth map. For a depth value $\\widetilde{D}_t(\\mathbf{p}_i)$ at pixel $\\mathbf{p}_i$ in the rendered map, we calculate the depth loss as:", + "bbox": [ + 89, + 446, + 483, + 506 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D} = \\sum_ {i} M _ {t} (\\mathbf {p} _ {i}) \\left\\| D _ {t} (\\mathbf {p} _ {i}) - \\widetilde {D} _ {t} (\\mathbf {p} _ {i}) \\right\\| _ {1}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 511, + 482, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also initialize each 3D Gaussian's opacity $o_j$ to 0.5, optimizing them to values of 0 or 1:", + "bbox": [ + 89, + 550, + 482, + 580 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {o p a c i t y}} = \\sum_ {j} o _ {j} ^ {2} \\left(1 - o _ {j}\\right) ^ {2}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 585, + 482, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This optimization aims to retain only Gaussians with an opacity of 1, pruning others to reduce ambiguity for the subsequent deblurring stage. We prune Gaussians with an opacity less than 0.8 and reset opacity with 0.5 again every certain amount of iteration. As iterations progress, the rendered depth maps increasingly align with the input depth maps, indicating that all 3D Gaussian positions and camera poses are globally well-aligned (Figure 5).", + "bbox": [ + 89, + 625, + 482, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final refinement loss is a weighted combination of the depth and opacity losses:", + "bbox": [ + 89, + 744, + 483, + 775 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e f i n e m e n t}} = \\lambda_ {D} \\mathcal {L} _ {D} + \\lambda_ {o} \\mathcal {L} _ {\\text {o p a c i t y}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 782, + 482, + 800 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Image Deblur", + "text_level": 1, + "bbox": [ + 89, + 804, + 235, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Motion blur caused by camera movement can generally be approximated as:", + "bbox": [ + 89, + 825, + 482, + 856 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI (t) \\approx B (t) = \\frac {1}{k} \\sum_ {i = 0} ^ {k} C _ {i} (t) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 863, + 482, + 904 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/164be333210da6b4a6d4068eee8e3c6b1251d2805000ba79443ba1baef62c0db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 79, + 645, + 152 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dd21ab2a4f4b577f34e8fa12b5459d4dbefcefca81e5a90d71de5492a2a8cad8.jpg", + "image_caption": [ + "Iteration 100", + "Figure 5. The refinement process corrects misaligned 3D Gaussians by updating camera poses and geometry, with effectiveness demonstrated in Table 3." + ], + "image_footnote": [], + "bbox": [ + 517, + 152, + 643, + 200 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b7a7242cd27897475e6ae65ffea3624d168713cc42fc97f33558d7801b02828e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 79, + 772, + 152 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/426e721ff4e3e49c7249f59700eac77120e5c0c75a0f93d0992916877512c6fa.jpg", + "image_caption": [ + "Iteration 700" + ], + "image_footnote": [], + "bbox": [ + 645, + 152, + 772, + 200 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9cf0d2fc9dba3009e263cf7fb0af56a6050dc15af3b345732d813391a06a9d61.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 79, + 901, + 152 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fa0ef4ca7b74b4f3581200f478bf0ff41b19d82859438d3b1586740577121153.jpg", + "image_caption": [ + "Iteration 20000" + ], + "image_footnote": [], + "bbox": [ + 774, + 152, + 901, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $B(t)$ represents the motion-blurred image, which we model as the average of $k$ sharp images $C_i(t)$ rendered from the 3D Gaussians at $k$ virtual camera poses interpolated between the camera poses at the start and end of the exposure time period $\\pmb{\\xi}_t^s$ and $\\pmb{\\xi}_t^e$ . This approximation allows $B(t)$ to closely match the input image $I(t)$ .", + "bbox": [ + 511, + 273, + 903, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Using the refined camera poses and point cloud from the previous step, we incorporate depth information to achieve geometrically accurate deblurring, following the approach in recent work [49]. Specifically, we minimize the difference between the input image $I(t)$ and the averaged rendered image $B(t)$ , as shown below:", + "bbox": [ + 511, + 364, + 905, + 455 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {B} = \\sum_ {t} \\sum_ {i} \\| I _ {t} (\\mathbf {p} _ {i}) - B _ {t} (\\mathbf {p} _ {i}) \\| _ {1}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 464, + 903, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final deblur loss function integrates this alignment loss with additional terms for structural similarity and depth consistency:", + "bbox": [ + 511, + 503, + 905, + 550 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d e b l u r}} = \\left(1 - \\lambda_ {B}\\right) \\mathcal {L} _ {B} + \\lambda_ {B} \\mathcal {L} _ {\\mathrm {D} - \\mathrm {S S I M}} + \\lambda_ {D} \\mathcal {L} _ {D}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 560, + 903, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This combined loss helps ensure that the deblurred output aligns well with both the input image and depth structure, leading to sharper and more geometrically consistent results.", + "bbox": [ + 511, + 585, + 905, + 646 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Results", + "text_level": 1, + "bbox": [ + 511, + 660, + 599, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset preparation. We prepare both synthetic and real-world datasets to evaluate our method's performance. In the absence of an RGB-D dataset with extreme motion blur, we generate a synthetic reference dataset by rendering directly from 3D models. We create RGB-D image sequences and record per-frame camera trajectories to evaluate our pose estimation and deblurring accuracy. The sequences, ranging from 50 to 150 frames, are captured along simulated trajectories that include rapid translations, 360-degree rotations, and human movement paths (using Blender's Walk Navigation feature). The RGB and depth images have a resolution of $640 \\times 480$ , and motion blur is applied by enabling Blender's motion blur setting. The 3D models are sourced from the McGuire archive [23] and the Blender archive [43].", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26429", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2cf99cda93bb7ddcff2d8a58587e6a4118a5d9f87ab4ea640829dab6d1959041.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison of deblurring performance from a novel view. Our method excels in restoring high-frequency details, whereas several methods, including COLMAP [31], fail under severe motion blur conditions (see Section 4.1). The corresponding quantitative results are provided in Table 2." + ], + "image_footnote": [], + "bbox": [ + 91, + 83, + 897, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For real scene evaluation, we used an Azure Kinect RGB-D camera with a 33.3ms fixed exposure time, 5000K fixed white balance, NFOV-2 × 2-binned depth capture mode, and $1280 \\times 720$ RGB resolution. We used the provided intrinsic camera parameters to undistort the images and aligned depth maps to color images using the SDK's transformation functions. Depth maps were eroded three times to remove outliers, which commonly occur along edges.", + "bbox": [ + 89, + 659, + 482, + 796 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. Our code is based on 3D Gaussian Splatting [14], with data processing partially adapted from CF-3DGS [14]. For initial camera pose estimation, we use the Unimatch [48] optical flow network and the ICP [5] algorithm from Open3D as part of the optimization process. For an input image sequence of $N$ frames,", + "bbox": [ + 89, + 809, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "we prune Gaussians every $100N$ iterations, reset opacity to 0.5 every $200N$ iterations, and run a total of $600N$ iterations (Section 3.2). The number of virtual views for deblurring is set to 15 for synthetic dataset evaluation and 17 for real-scene comparison. All experiments were conducted on an NVIDIA RTX A6000 GPU (48GB), with initial camera pose optimization taking 11 seconds per frame, 0.1 seconds per iteration for pose and geometry refinement, and 0.26 seconds per iteration for deblurring at $640 \\times 480$ resolution.", + "bbox": [ + 511, + 659, + 906, + 796 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 511, + 801, + 735, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the camera pose accuracy and deblurring performance on three synthetic scenes from our dataset, each rendered with extreme motion blur. For pose accuracy, we compare our method against a variety of approaches capable of generating a 3D point cloud and camera trajectory:", + "bbox": [ + 511, + 825, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26430", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/63db02ccec4b5616198a2333aa1148c6ba4dfb8a09d5fcfe25fee7f35c63a572.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison of deblurring performance from novel views on real scenes captured with an Azure Kinect camera. Our method effectively restores high-frequency details, outperforming other methods even when processing RGB-D sequences with large spatial gaps between frames due to rapid camera motion and severe motion blur." + ], + "image_footnote": [], + "bbox": [ + 96, + 83, + 897, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "COLMAP [31], ICP with FPFH feature-based global alignment [29], ORB-SLAM3 [3], ElasticFusion [47], and several RGB-D SLAM methods, including NICE-SLAM [51], Point-SLAM [30], and DROID-SLAM [42]. Additionally, we include recent Gaussian Splatting-based SLAM approaches such as MonoGS [21] and SplaTAM [12]. For each method, we calculate absolute trajectory error (ATE) and relative pose error (RPE) in meter and degree units, using the ground truth camera trajectories as references.", + "bbox": [ + 88, + 686, + 480, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate deblurring accuracy, we modify a state-of-the-art Gaussian Splitting-based image restoration method [49] to operate in a depth-aware manner using undistorted color and depth images. We select one out of every three images as a test view and render the correspond", + "bbox": [ + 88, + 824, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing novel view for comparison with the ground truth. For quantitative metrics, we calculate PSNR, SSIM, and LPIPS on the rendered images, as well as RMSE between the rendered and ground truth depth maps in the inverse depth domain $(m^{-1})$ .", + "bbox": [ + 511, + 686, + 906, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 772, + 666, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To validate the contributions of each component in our method, we perform an ablation study evaluating camera pose and deblurring accuracy by selectively omitting key elements. Specifically, we conduct experiments without fixing the scale of the Gaussians (Section 3.2), without the pose and geometry refinement process (Equation 5), and without depth loss (Equation 8).", + "bbox": [ + 511, + 794, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26431", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/62f4a7e3811fa0f2c54a8a9981e7eed931cdb98a5b27672cbb7427697a8e675d.jpg", + "table_caption": [ + "Table 1. Quantitative comparison of camera pose accuracy against ground truth poses. We report scores only for methods that successfully provide both camera poses and a point cloud (see Section 4.1). Our method achieves the best scores on RPE metrics and shows ATE accuracy comparable to DROID-SLAM [42]. Each color highlights the best and second best results." + ], + "table_footnote": [], + "table_body": "
BedroomLivingroomOffice
ATE ↓RPE (trans) ↓RPE (rot) ↓ATE ↓RPE (trans) ↓RPE (rot) ↓ATE ↓RPE (trans) ↓RPE (rot) ↓
ElasticFusion [47]1.3030.0792.2181.0140.0562.7770.4790.0380.480
FPFH-ICP [29]2.7600.1182.3102.0280.0921.2130.1390.0090.167
DROID-SLAM [42]0.0590.0441.2120.0500.0140.3170.0110.0110.245
MonoGS [21]0.5060.0410.8120.1350.0090.1240.0450.0110.180
Ours0.1020.0060.0920.0050.0020.0320.0410.0030.046
", + "bbox": [ + 91, + 133, + 906, + 228 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/af31e8321d226fbe0eb064773042dc376a8df05fb0858346202b0262cc711bcc.jpg", + "table_caption": [ + "Table 2. Quantitative comparison of deblurring performance from novel views. Our method demonstrates high accuracy across all metrics for both color and depth images. RMSE is calculated by comparing the rendered depth with GT depth in the inverse depth domain $(m^{-1})$ ." + ], + "table_footnote": [], + "table_body": "
BedroomLivingroomOffice
PSNR ↑SSIM ↑LPIPS ↓RMSE ↓PSNR ↑SSIM ↑LPIPS ↓RMSE ↓PSNR ↑SSIM ↑LPIPS ↓RMSE ↓
ElasticFusion [47]18.7660.5860.4661.24913.7990.4480.5540.14619.0520.5330.4010.079
FPFH-ICP [29]20.6230.6170.4180.45018.1720.5980.4220.10222.2120.6840.2410.033
DROID-SLAM [42]21.3750.6660.3020.01321.2220.7350.2500.03525.6800.7690.1440.013
MonoGS [21]23.7280.7170.2850.00924.4140.8220.1880.02724.1940.7250.1840.184
Ours26.7450.8240.2060.00527.6500.9000.1500.01025.6490.7910.1600.014
", + "bbox": [ + 91, + 270, + 906, + 357 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/273dbbba05fb83080d825dc6f01171b82d8d82dbede7a8a8215a2e1cb1d9dea5.jpg", + "table_caption": [ + "Table 3. Ablation study on the effects of the global refinement process and fixed scale of Gaussians. Our refinement significantly enhances deblurring and depth accuracy while fixing the scale of the Gaussians leads to more accurate camera pose estimation." + ], + "table_footnote": [], + "table_body": "
PSNRSSIMLPIPSRMSEATERPE(trans)RPE(rot)
w/o scale fix21.6560.7030.3080.0270.1550.0531.488
w/o refinement22.2230.7340.2380.0270.1230.0070.047
w/o depth loss25.5180.8070.1700.0140.1030.0120.195
Scale fix + refinement26.6810.8380.1720.0100.1030.0120.195
", + "bbox": [ + 91, + 430, + 480, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We assess the effectiveness of these components by performing deblurring based on results from Section 3.1—omitting refinement of camera poses and geometry—and by allowing the scale of the 3D Gaussians to be optimized freely. Table 3 demonstrates that both the global refinement process and fixed Gaussian scale are essential for achieving high accuracy. Notably, during pose refinement, the algorithm reduces absolute pose error even at the expense of a slight increase in relative pose error, underscoring the importance of our refinement approach for stable, high-quality reconstruction.", + "bbox": [ + 88, + 491, + 482, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method achieves superior RPE scores across all scenes, as shown in Table 1, outperforming other approaches in pose and geometry estimation under severe motion blur. While our ATE scores are comparable to those of DROID-SLAM, which uses a bundle adjustment (BA) module, ATE alone does not correlate strongly with deblurring quality. Our method consistently produces more accurate and denser point clouds, which directly contributes to improved deblurring performance, as evidenced by the metrics in Table 2. These results confirm that our integrated approach to pose estimation and deblurring is more resilient to challenging conditions than other methods.", + "bbox": [ + 88, + 657, + 482, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 89, + 848, + 302, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figures 6 and 7 present qualitative comparisons of our method's performance in synthetic and real scenes. The", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "densification capability of Gaussian Splitting provides detailed scene representation even from a relatively sparse point cloud; however, our approach, with its initially dense and accurate point cloud generation, achieves notably superior deblurring performance, especially in restoring high-frequency details. Our method consistently demonstrates a significant advantage in visual fidelity and detail restoration across diverse conditions.", + "bbox": [ + 511, + 376, + 906, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 503, + 633, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have introduced a robust method for 3D scene reconstruction from RGB-D image sequences that effectively addresses the challenges of extreme motion blur and low-light conditions. Our approach leverages optical flow from color images and a carefully designed geometric loss to achieve accurate global alignment, followed by local refinement of camera poses and geometry using ICP. By integrating these steps into a Gaussian Splatting pipeline, we further refine camera poses and 3D geometry by minimizing depth and opacity losses. Additionally, fixing the scale of the 3D Gaussians ensures that the depth information from the RGB-D input is fully utilized, allowing for precise camera pose estimation and improved deblurring performance. Our method demonstrates superior performance on real and synthetic RGB-D scenes with significant motion blur, outperforming existing approaches.", + "bbox": [ + 511, + 527, + 906, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 511, + 780, + 681, + 797 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Min H. Kim acknowledges the Samsung Research Funding & Incubation Center (SRFC-IT2402-02), the Korea NRF grant (RS-2024-00357548), the MSIT/IITP of Korea (RS2022-00155620, RS-2024-00398830, 2022-0-00058, and 2017-0-00072), Microsoft Research Asia, and Samsung Electronics.", + "bbox": [ + 511, + 800, + 905, + 891 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26432", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Dejan Azinović, Ricardo Martin-Brualla, Dan B Goldman, Matthias Nießner, and Justus Thies. Neural rgb-d surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6290-6301, 2022. 3", + "[2] Michael Bloesch, Jan Czarnowski, Ronald Clark, Stefan Leutenegger, and Andrew J Davison. Codeslam—learning a compact, optimisable representation for dense visual slam. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2560–2568, 2018. 2", + "[3] Carlos Campos, Richard Elvira, Juan J Gómez Rodríguez, José MM Montiel, and Juan D Tardós. Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics, 37(6):1874-1890, 2021. 2, 7", + "[4] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. arXiv preprint arXiv:2204.04676, 2022. 2, 3", + "[5] Yang Chen and Gérard Medioni. Object modelling by registration of multiple range images. Image and vision computing, 10(3):145-155, 1992. 6", + "[6] Angela Dai, Matthias Nießner, Michael Zollofer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface re-integration. ACM Transactions on Graphics 2017 (TOG), 2017. 2", + "[7] Rob Fergus, Barun Singh, Aaron Hertzmann, Sam T Roweis, and William T Freeman. Removing camera shake from a single photograph. In Acm Siggraph 2006 Papers, pages 787-794. 2006. 3", + "[8] Kyle Genova, Forrester Cole, Avneesh Sud, Aaron Sarna, and Thomas Funkhouser. Deep structured implicit functions. arXiv preprint arXiv:1912.06126, 2:2, 2019. 2", + "[9] Zan Gojcic, Caifa Zhou, Jan D Wegner, and Andreas Wieser. The perfect match: 3d point cloud matching with smoothed densities. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5545-5554, 2019. 3", + "[10] Alex Graves and Alex Graves. Long short-term memory. Supervised sequence labelling with recurrent neural networks, pages 37-45, 2012. 3", + "[11] Olaf Kähler, Victor Prisacariu, Julien Valentin, and David Murray. Hierarchical voxel block hashing for efficient integration of depth images. IEEE Robotics and Automation Letters, 1(1):192-197, 2015. 2", + "[12] Nikhil Keetha, Jay Karhade, Krishna Murthy Jatavallabhula, Gengshan Yang, Sebastian Scherer, Deva Ramanan, and Jonathon Luiten. Splatam: Splat track & map 3d gaussians for dense rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21357-21366, 2024. 3, 7", + "[13] Maik Keller, Damien Lefloch, Martin Lambers, Shahram Izadi, Tim Weyrich, and Andreas Kolb. Real-time 3d reconstruction in dynamic scenes using point-based fusion. In 2013 International Conference on 3D Vision-3DV 2013, pages 1-8. IEEE, 2013. 2" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 3, 6", + "[15] Christian Kerl, Jürgen Sturm, and Daniel Cremers. Dense visual slam for rgb-d cameras. In 2013 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 2100-2106. IEEE, 2013. 2", + "[16] Dilip Krishnan and Rob Fergus. Fast image deconvolution using hyper-laplacian priors. Advances in neural information processing systems, 22, 2009. 3", + "[17] Dogyoon Lee, Minhyeok Lee, Chajin Shin, and Sangyoun Lee. Dp-nerf: Deblurred neural radiance field with physical scene priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12386–12396, 2023. 3", + "[18] Yang Li and Tatsuya Harada. Lepard: Learning partial point cloud matching in rigid and deformable scenes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5554-5564, 2022. 3", + "[19] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-nerf: Neural radiance fields from blurry images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12861–12870, 2022. 3", + "[20] Nico Marniok, Ole Johannsen, and Bastian Goldluecke. An efficient octree design for local variational range image fusion. In Pattern Recognition: 39th German Conference, GCPR 2017, Basel, Switzerland, September 12-15, 2017, Proceedings 39, pages 401-412. Springer, 2017. 2", + "[21] Hidenobu Matsuki, Riku Murai, Paul HJ Kelly, and Andrew J Davison. Gaussian splatting slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18039-18048, 2024. 3, 7, 8", + "[22] John McCormac, Ankur Handa, Andrew Davison, and Stefan Leutenegger. Semantic fusion: Dense 3d semantic mapping with convolutional neural networks. In 2017 IEEE International Conference on Robotics and automation (ICRA), pages 4628-4635. IEEE, 2017. 2", + "[23] Morgan McGuire. Computer graphics archive, 2017. 5", + "[24] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3", + "[25] Raul Mur-Artal and Juan D Tardós. Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras. IEEE transactions on robotics, 33(5):1255-1262, 2017. 2", + "[26] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3883-3891, 2017. 3", + "[27] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon." + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26433", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Kinectfusion: Real-time dense surface mapping and tracking. In 2011 10th IEEE international symposium on mixed and augmented reality, pages 127-136. IEEE, 2011. 2", + "[28] Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Marc Stamminger. Real-time 3d reconstruction at scale using voxel hashing. ACM Transactions on Graphics (ToG), 32 (6):1-11, 2013. 2", + "[29] Radu Bogdan Rusu, Nico Blodow, and Michael Beetz. Fast point feature histograms (fpfh) for 3d registration. In 2009 IEEE international conference on robotics and automation, pages 3212-3217. IEEE, 2009. 7, 8", + "[30] Erik Sandström, Yue Li, Luc Van Gool, and Martin R Oswald. Point-slam: Dense neural point cloud-based slam. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18433-18444, 2023. 3, 7", + "[31] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 6, 7", + "[32] Thomas Schops, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 134-144, 2019. 2", + "[33] Thomas Schöps, Torsten Sattler, and Marc Pollefeys. Surfelmeshing: Online surfel-based mesh reconstruction. IEEE transactions on pattern analysis and machine intelligence, 42(10):2494-2507, 2019. 2", + "[34] Christian J Schuler, Michael Hirsch, Stefan Harmeling, and Bernhard Scholkopf. Learning to deblur. IEEE transactions on pattern analysis and machine intelligence, 38(7):1439-1451, 2015. 3", + "[35] Otto Seiskari, Jerry Yilammi, Valtteri Kaatrasalo, Pekka Rantalankila, Matias Turkulainen, Juho Kannala, and Arno Solin. Gaussian splattering on the move: Blur and rolling shutter compensation for natural camera motion, 2024. 2, 3", + "[36] Qi Shan, Jiaya Jia, and Aseem Agarwala. High-quality motion deblurring from a single image. Acm transactions on graphics (tog), 27(3):1-10, 2008. 3", + "[37] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. Advances in neural information processing systems, 28, 2015. 3", + "[38] Jörg Stuckler and Sven Behnke. Multi-resolution surfel maps for efficient dense 3d modeling and tracking. Journal of Visual Communication and Image Representation, 25(1):137-147, 2014. 2", + "[39] Shuochen Su, Maurizio Delbracio, Jue Wang, Guillermo Sapiro, Wolfgang Heidrich, and Oliver Wang. Deep video deblurring for hand-held cameras. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1279-1288, 2017. 3", + "[40] Jian Sun, Wenfei Cao, Zongben Xu, and Jean Ponce. Learning a convolutional neural network for non-uniform motion blur removal. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 769-777, 2015. 3" + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 3", + "[42] Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. Advances in neural information processing systems, 2021. 3, 7, 8", + "[43] Flavio Della Tommasa. Blender demo files, 2024. 5", + "[44] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. BAD-NeRF: Bundle Adjusted Deblur Neural Radiance Fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4170–4179, 2023. 2, 3", + "[45] Chen Wenbo and Liu Ligang. Deblur-gs: 3d gaussian splattering from camera motion blurred images. Proc. ACM Comput. Graph. Interact. Tech. (Proceedings of I3D 2024), 7(1), 2024. 2, 3", + "[46] Thomas Whelan, Michael Kaess, Hordur Johannsson, Maurice Fallon, John J Leonard, and John McDonald. Real-time large-scale densergb-d slam with volumetric fusion. The International Journal of Robotics Research, 34(4-5):598-626, 2015. 2", + "[47] Thomas Whelan, Renato F Salas-Moreno, Ben Glocker, Andrew J Davison, and Stefan Leutenegger. Elasticfusion: Real-time dense slam and light source estimation. The International Journal of Robotics Research, 35(14):1697-1716, 2016. 2, 7, 8", + "[48] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, Fisher Yu, Dacheng Tao, and Andreas Geiger. Unifying flow, stereo and depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 4, 6", + "[49] Lingzhe Zhao, Peng Wang, and Peidong Liu. BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splatting. 2024. 2, 3, 5, 7", + "[50] Huizhong Zhou, Benjamin Ummenhofer, and Thomas Brox. Deeptam: Deep tracking and mapping. In Proceedings of the European conference on computer vision (ECCV), pages 822-838, 2018. 2", + "[51] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12786-12796, 2022. 3, 7" + ], + "bbox": [ + 514, + 90, + 906, + 724 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26434", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Splat-based 3D Scene Reconstruction with Extreme Motion-blur/73a5b991-b304-4700-8590-8334fbf2bdee_model.json b/2025/Splat-based 3D Scene Reconstruction with Extreme Motion-blur/73a5b991-b304-4700-8590-8334fbf2bdee_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a9d63189ef78b9139467be4c2f403b306eb793c0 --- /dev/null +++ b/2025/Splat-based 3D Scene Reconstruction with Extreme Motion-blur/73a5b991-b304-4700-8590-8334fbf2bdee_model.json @@ -0,0 +1,2365 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.174, + 0.131, + 0.825, + 0.153 + ], + "angle": 0, + "content": "Splat-based 3D Scene Reconstruction with Extreme Motion-blur" + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.179, + 0.277, + 0.201 + ], + "angle": 0, + "content": "Hyeonjoong Jang†*" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.181, + 0.443, + 0.215 + ], + "angle": 0, + "content": "Dongyoung Choi† \n† KAIST" + }, + { + "type": "text", + "bbox": [ + 0.465, + 0.181, + 0.74, + 0.216 + ], + "angle": 0, + "content": "Donggun Kim\\(^{\\dagger}\\) Woohyun Kang\\(^{\\dagger}\\) \n* HYPERGRAM" + }, + { + "type": "text", + "bbox": [ + 0.763, + 0.181, + 0.875, + 0.198 + ], + "angle": 0, + "content": "Min H. Kim†*" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.22, + 0.753, + 0.235 + ], + "angle": 0, + "content": "{hjjang; dychoi; dgkim; whkang; minhkim}@vclab.kaist.ac.kr" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.258, + 0.256, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.329, + 0.177, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.16, + 0.391, + 0.195, + 0.404 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.389, + 0.195, + 0.404 + ], + "angle": 0, + "content": "out" + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.258, + 0.417, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.329, + 0.338, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.298, + 0.39, + 0.38, + 0.402 + ], + "angle": 0, + "content": "ElasticFusion" + }, + { + "type": "image", + "bbox": [ + 0.339, + 0.329, + 0.418, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.421, + 0.258, + 0.578, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.421, + 0.329, + 0.577, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.453, + 0.39, + 0.545, + 0.401 + ], + "angle": 0, + "content": "DROID-SLAM" + }, + { + "type": "image", + "bbox": [ + 0.581, + 0.258, + 0.738, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.581, + 0.329, + 0.686, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.633, + 0.39, + 0.688, + 0.401 + ], + "angle": 0, + "content": "MonoGS" + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.258, + 0.899, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.329, + 0.82, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.821, + 0.329, + 0.899, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.805, + 0.391, + 0.836, + 0.403 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.405, + 0.907, + 0.434 + ], + "angle": 0, + "content": "Figure 1. We propose a robust 3D scene reconstruction method from RGB-D input that effectively addresses extreme motion blur. Our approach achieves accurate camera poses and dense point clouds, producing clearer, deblurred scenes compared to existing methods." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.443, + 0.327, + 0.459 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.478, + 0.486, + 0.902 + ], + "angle": 0, + "content": "We propose a splat-based 3D scene reconstruction method from RGB-D input that effectively handles extreme motion blur, a frequent challenge in low-light environments. Under dim illumination, RGB frames often suffer from severe motion blur due to extended exposure times, causing traditional camera pose estimation methods, such as COLMAP, to fail. This results in inaccurate camera pose and blurry color input, compromising the quality of 3D reconstructions. Although recent 3D reconstruction techniques like Neural Radiance Fields and Gaussian Splatting have demonstrated impressive results, they rely on accurate camera trajectory estimation, which becomes challenging under fast motion or poor lighting conditions. Furthermore, rapid camera movement and the limited field of view of depth sensors reduce point cloud overlap, limiting the effectiveness of pose estimation with the ICP algorithm. To address these issues, we introduce a method that combines camera pose estimation and image deblurring using a Gaussian Splatting framework, leveraging both 3D Gaussian splats and depth inputs for enhanced scene representation. Our method first aligns consecutive RGB-D frames through optical flow and ICP, then refines camera poses and 3D geometry by adjusting Gaussian positions for optimal depth alignment. To handle motion blur, we model camera movement during exposure and deblur images by comparing the input with a series of sharp, rendered frames. Experiments on a new RGB-D dataset with extreme motion blur show that our method outperforms existing approaches," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.445, + 0.909, + 0.537 + ], + "angle": 0, + "content": "enabling high-quality reconstructions even in challenging conditions. This approach has broad implications for 3D mapping applications in robotics, autonomous navigation, and augmented reality. Both code and dataset are publicly available on https://github.com/KAIST-VCLAB/gs-extreme-motion-blur." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.57, + 0.645, + 0.585 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.597, + 0.907, + 0.854 + ], + "angle": 0, + "content": "High-quality 3D scene reconstruction is one of the most important and challenging applications in computer vision. The accuracy of 3D reconstruction hinges on the quality of essential components, such as camera poses, RGB images, and depth maps. These elements are interconnected and influence each other; a failure in one component can adversely impact the others. For example, blurry RGB images or noisy depth maps can significantly impair camera pose estimation, which, in turn, degrades overall reconstruction quality. This challenge is particularly pronounced when input data is captured in low-light conditions or during rapid camera movement—common scenarios in everyday, casual capture. Such conditions result in degraded color and depth frames, leading to poorly estimated camera poses and reconstructions with blurry textures and distorted or smoothed geometry. Therefore, achieving high-quality 3D reconstruction requires clear RGB images and accurate depth maps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The interdependence of these input components suggests an opportunity for compensating or restoring degraded elements. However, this process presents a chicken-and-egg" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "26425" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "problem, complicating the reconstruction further. For instance, accurate camera pose estimation requires sharp images, while deblurring motion-blurred images depend on reliable camera pose information. Although numerous methods [4, 41] exist for recovering sharp images from motion-blurred ones, they often rely on supervised learning models trained on specific datasets. Consequently, their deblurring performance can degrade significantly when faced with new cameras or motion blur scenarios that differ from the training data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.245, + 0.482, + 0.412 + ], + "angle": 0, + "content": "Recent works [35, 44, 45, 49] have made advances in addressing these challenges by optimizing camera trajectories during exposure while simultaneously learning sharp RGB colors. This enables simultaneous camera pose estimation and image deblurring. However, these methods require an initial camera pose and, when using Gaussian Splatting [14], also need a sparse point cloud as input. This requirement limits their applicability, especially when severe motion blur prevents structure-from-motion (SfM) methods like COLMAP [31] from functioning effectively, making it impossible to start the optimization process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.482, + 0.581 + ], + "angle": 0, + "content": "To overcome these limitations, we introduce a robust approach for 3D scene reconstruction that effectively compensates for severe motion blur and addresses the challenges of camera pose estimation without relying on a precise initial pose. Our method leverages RGB-D inputs and incorporates both optical flow and depth information to align camera poses accurately, even in the presence of challenging motion blur and lighting conditions. We first perform a global alignment between consecutive frames using optical flow and the ICP algorithm, which enables effective local alignment of point clouds generated from the depth maps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.482, + 0.763 + ], + "angle": 0, + "content": "After this initial alignment, we refine the camera poses and 3D geometry by integrating them into a Gaussian Splating pipeline. This approach allows us to initialize dense 3D Gaussians from depth maps, which we scale to ensure a detailed representation of the scene geometry. Our refinement process iteratively adjusts both camera poses and the positions of 3D Gaussians, achieving global consistency by minimizing a depth alignment loss that compares rendered depth maps with input depth measurements. Through this adjustment, we eliminate loop-closure artifacts and reduce the drift that often accumulates in traditional SLAM-based methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.482, + 0.903 + ], + "angle": 0, + "content": "For scenes with significant motion blur, we further optimize the deblurring process by modeling the camera poses at the start and end of each frame's exposure. This temporal modeling allows us to simulate the effects of motion during the exposure period, which we incorporate into our Gaussian Splatting framework. We minimize the difference between the observed motion-blurred image and the average of a set of sharp images rendered from multiple viewpoints along the exposure trajectory, resulting in more accurate," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "geometrically consistent deblurring. Our deblur loss function combines image alignment, structural similarity, and depth consistency, ensuring that the final reconstructions retain sharpness and fidelity to the original scene structure." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.153, + 0.905, + 0.365 + ], + "angle": 0, + "content": "Through extensive evaluation on a newly constructed RGB-D dataset featuring extreme motion blur, we demonstrate that our approach significantly outperforms existing methods in both accuracy and robustness. By addressing the key challenges of pose estimation, depth alignment, and image deblurring, our method provides a versatile and effective solution for high-quality 3D reconstruction under real-world capture conditions, such as low light and fast motion. The proposed method holds promising applications in areas requiring reliable 3D mapping and reconstruction, including robotics, autonomous navigation, and augmented reality. We will make both our dataset and implementation publicly available to foster further research and development in this field." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.384, + 0.655, + 0.4 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.414, + 0.905, + 0.596 + ], + "angle": 0, + "content": "RGB-D scene reconstruction. A popular approach to 3D geometry reconstruction defines a 3D voxel grid and constructs a signed distance function (SDF) volume by accumulating depth estimates captured by sensors. Kinect-Fusion [27] uses a Kinect camera to capture RGB-D images, then estimates relative camera poses between adjacent frames with the ICP algorithm. Next, it constructs and updates a truncated signed distance function (TSDF) volume, enabling real-time 3D reconstruction of objects. Niessner et al. [28] improve memory efficiency by combining hashing algorithms to store and update SDF values only where needed, thus reducing memory requirements." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.598, + 0.905, + 0.808 + ], + "angle": 0, + "content": "Several methods [8, 11, 20, 22] adopt memory-efficient, hierarchical data structures for depth fusion. BundleFusion [6] simultaneously refines camera poses and geometry using bundle adjustment, improving reconstruction accuracy and addressing the loop closure problem caused by accumulated camera pose and depth estimation errors. ElasticFusion [47] builds a map with surfels—point primitives containing position, normal, and radius—and updates the map through a deformation graph to minimize error and perform loop closure. Other methods [13, 33, 38] directly fuse depth estimates into point clouds, optimizing memory use. However, these methods generally require slow camera motion to capture sharp images and accurately estimate camera transformations." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.81, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Simultaneous localization and mapping (SLAM) methods [3, 15, 25, 32, 46] track camera motion and construct a map in real-time by extracting visual features from input images to establish 3D point correspondences. CodeSLAM [2] and DeepTAM [50] use neural networks to combine depth maps with color images, while Azinović" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26426" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.154, + 0.066, + 0.844, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.227, + 0.908, + 0.298 + ], + "angle": 0, + "content": "Figure 2. Overview of our method pipeline. We begin with global alignment by comparing optical flow and reprojected pixel differences between adjacent RGB-D image pairs, followed by local estimation of relative camera poses using ICP. Each depth map and camera pose is then used to initialize 3D Gaussians. By fixing the scale of the Gaussians and optimizing with depth and opacity losses, we achieve global refinement of camera poses and geometry. Finally, we deblur the scene texture by minimizing the deblur loss between the input image and averaged virtual images." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.313, + 0.486, + 0.465 + ], + "angle": 0, + "content": "et al. [1] train multi-layer perceptrons (MLP) to learn SDF and color values for scene reconstruction. Their approach also optimizes camera transformation correction variables during training to refine poses. However, these methods struggle with input images that contain motion blur. Nice-SLAM [51], Point-SLAM [30], and DROID-SLAM [42] use neural networks to track camera poses. SplaTAM [12] and MonoGS [21] introduced SLAM solutions based on 3D Gaussian Splatting. However, these methods are not designed to handle motion blur." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.473, + 0.483, + 0.729 + ], + "angle": 0, + "content": "Image deblurring. Motion blur occurs when camera motion during the exposure time changes the projected pixel coordinates of rays on the camera sensor. Early studies [7, 16, 36] tackle this issue by developing kernels to restore blurry images. Convolutional neural networks (CNNs) demonstrate powerful performance in learning deblur kernels [34, 40], and deep learning techniques such as ResNet with skip connections and multi-scale networks further improved restoration quality [26, 39]. SRN-DeblurNet [41] adopt recurrent networks [10, 37] and combine them with an encoder-decoder network structure, while NAFNet [4] simplify the network structure, extracting only essential components and proposing a nonlinear activation-free network. However, since these methods rely on curated datasets like the GoPro dataset [26], they struggle to effectively deblur images with extreme motion blur or those captured on different cameras." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "There are NeRF-based approaches [24] that tackle motion blur in input images. Deblur-NeRF [19] and DP-NeRF [17] use a set of motion-blurred images along with camera poses estimated by COLMAP as input to restore sharp images through NeRF optimization. BAD-NeRF [44] enhances this by jointly optimizing virtual camera poses and radiance fields during exposure. The latest approaches utilize Gaussian Splatting [14], which offers faster training times than NeRF by rendering a set of 3D Gaussians with a dedicated rasterizer rather than optimizing a neural network. Gaussian Splatting-based deblurring meth" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.313, + 0.905, + 0.358 + ], + "angle": 0, + "content": "ods [35, 45, 49] optimize the virtual camera trajectory and minimize differences between input blurry images and the average of rendered images at each virtual camera position." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.359, + 0.906, + 0.525 + ], + "angle": 0, + "content": "All of these methods rely on initial camera poses, typically obtained from COLMAP, and for Gaussian Splatting-based methods, a sparse point cloud as well. However, structure-from-motion methods like COLMAP often fail with blurry inputs—such as images captured by fast-moving cameras or in low-light conditions that require long exposure times. Even with dense depth inputs, ICP requires acceptable global alignment, and recent point cloud matching methods [9, 18] cannot accurately determine relative camera transformations between two point clouds if they consist mainly of planar structures." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.526, + 0.906, + 0.601 + ], + "angle": 0, + "content": "To address these issues, we propose an effective camera pose estimation algorithm tailored for RGB-D image sequences that jointly refines both the camera trajectory and the reconstructed 3D point cloud, resulting in improved sharpness in 3D reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.615, + 0.605, + 0.631 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Overview. In datasets with significant motion blur, using COLMAP to estimate camera poses becomes impractical. Additionally, rapid camera motion and the limited field of view of depth sensors further complicate this task, making ICP registration less effective. To address these challenges, we first perform global alignment using both RGB and depth images, which establishes sufficient initial alignment for the ICP algorithm to operate effectively. After achieving this global alignment, we apply ICP again to refine and accurately estimate the camera poses. However, estimating camera poses by analyzing only two consecutive frames in a sequence can result in accumulated errors, leading to drift over time. To resolve this issue, it is crucial to adjust the 3D geometry and align the camera poses across the entire sequence so that depth maps and camera views from all perspectives remain well-aligned. We accomplish this through bundle adjustment, optimizing the positions of" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26427" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.142, + 0.079, + 0.434, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.21, + 0.486, + 0.255 + ], + "angle": 0, + "content": "Figure 3. The illustration of our pose initialization loss. When \\(\\mathbf{p}_i\\) is reprojected to \\(\\mathbf{p}_j\\) using the depth value, it is considered geometrically consistent if \\(\\overline{\\mathbf{P}_i\\mathbf{H}_i}\\) equals \\(\\overline{\\mathbf{P}_j\\mathbf{H}_j}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.268, + 0.485, + 0.345 + ], + "angle": 0, + "content": "3D Gaussians and camera poses within the Gaussian Splating pipeline. Finally, to account for motion blur, we approximate the camera trajectory over time and optimize the movement during exposure, enabling effective deblurring. See Figure 2 for an overview." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.358, + 0.3, + 0.373 + ], + "angle": 0, + "content": "3.1. Local Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.382, + 0.484, + 0.548 + ], + "angle": 0, + "content": "To estimate relative camera transformations, we select consecutive frames from the input RGB-D image sequence. Our approach begins with global alignment using optical flow, followed by local alignment of the two-point clouds using the ICP algorithm. For each consecutive RGB image and depth map pair at times \\( t \\) and \\( t + 1 \\), we optimize the transformation \\( \\xi \\in \\mathfrak{se}(3) \\) between the frames. Here, \\( \\xi_{t\\rightarrow t + 1} = [\\mathbf{x}^{\\intercal},\\boldsymbol{\\omega}^{\\intercal}]^{\\intercal} \\), where \\( \\mathbf{x} \\) represents the translation vector between the two camera origins, and \\( \\boldsymbol{\\omega} \\) is the rotation element of \\( \\mathfrak{se}(3) \\). We ultimately calculate a camera-to-world transformation matrix \\( \\xi_t \\) for each timestamp \\( t \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.55, + 0.484, + 0.718 + ], + "angle": 0, + "content": "To achieve this, we first compute the difference between the estimated optical flow and the reprojected pixel from the camera at time \\(t\\) to \\(t + 1\\) using the depth estimate for each pixel in the depth map \\(D_{t}\\) and the intrinsic matrix \\(\\mathbf{K}\\). A pixel \\(\\mathbf{p}_i\\) in the depth map is backprojected to 3D space as: \\(\\mathbf{P}_i = \\pi^{-1}(\\mathbf{K},\\mathbf{p}_i,D_t(\\mathbf{p}_i))\\). We then project this 3D point to the camera at \\(t + 1\\): \\(\\mathbf{p_j} = \\pi (\\mathbf{K},\\exp (\\xi_{t\\to t + 1}^{\\wedge})\\mathbf{P}_i)\\), where \\(\\pi\\) and \\(\\pi^{-1}\\) represent the projection and backprojection operators, respectively. The pixel movement from \\(\\mathbf{p}_i\\) to \\(\\mathbf{p}_j\\) is then compared with the estimated optical flow \\(F_{t\\rightarrow t + 1}(\\mathbf{p}_i)\\) to compute the optical flow loss:" + }, + { + "type": "equation", + "bbox": [ + 0.094, + 0.731, + 0.482, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {F} = \\sum_ {i} M _ {t} (\\mathbf {p} _ {i}) M _ {t + 1} (\\mathbf {p} _ {j}) \\| F _ {t \\to t + 1} (\\mathbf {p} _ {i}) - (\\mathbf {p} _ {j} - \\mathbf {p} _ {i}) \\| _ {2}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.779, + 0.485, + 0.84 + ], + "angle": 0, + "content": "In this formulation, \\( M_{t} \\) is a mask indicating valid depth values in \\( D_{t} \\) at pixel \\( \\mathbf{p}_{i} \\). The mask \\( M_{t} \\) dynamically updates during optimization, as \\( \\mathbf{p}_{j} \\) depends on the transformation \\( \\xi \\): \\( M_{t}(\\mathbf{p}_{i}) = D_{t}(\\mathbf{p}_{i}) > 0 \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.902 + ], + "angle": 0, + "content": "We employ a state-of-the-art pretrained optical flow estimation model [48] designed hierarchically to produce reliable optical flow even from two blurry input images. Next, we calculate the photometric loss between color values at" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.837, + 0.108 + ], + "angle": 0, + "content": "pixel \\(\\mathbf{p}_i\\) in image \\(I_{t}\\) and pixel \\(\\mathbf{p}_j\\) in image \\(I_{t + 1}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.119, + 0.88, + 0.151 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {I} = \\sum_ {i} M _ {t} (\\mathbf {p} _ {i}) M _ {t + 1} (\\mathbf {p} _ {j}) \\| I _ {t} (\\mathbf {p} _ {i}) - I _ {t + 1} (\\mathbf {p} _ {j}) \\| _ {1}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.16, + 0.907, + 0.266 + ], + "angle": 0, + "content": "To ensure geometric consistency, we introduce a depth consistency loss by using the two depth maps captured by the depth camera to optimize relative poses. Specifically, as illustrated in Figure 3, we compute and compare the vertical components \\(\\overline{\\mathbf{P}_i\\mathbf{H}_i}\\) and \\(\\overline{\\mathbf{P}_j\\mathbf{H}_j}\\) based on the baseline vector \\(\\overline{\\mathbf{O}_t\\mathbf{O}_{t + 1}}\\) between the two cameras. We calculate \\(\\overline{\\mathbf{P}_i\\mathbf{H}_i}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.273, + 0.866, + 0.31 + ], + "angle": 0, + "content": "\\[\n\\overline {{\\mathbf {P} _ {i} \\mathbf {H} _ {i}}} = \\| \\mathbf {P} _ {i} \\| _ {2} \\sin \\left(\\cos^ {- 1} \\left(\\frac {\\mathbf {P} _ {i} \\cdot \\mathbf {x}}{\\| \\mathbf {P} _ {i} \\| _ {2} \\| \\mathbf {x} \\| _ {2}}\\right)\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.319, + 0.736, + 0.337 + ], + "angle": 0, + "content": "Similarly, we compute \\(\\overline{\\mathbf{P}_j\\mathbf{H}_j}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.346, + 0.897, + 0.378 + ], + "angle": 0, + "content": "\\[\n\\overline {{\\mathbf {P} _ {j} \\mathbf {H} _ {j}}} = \\| \\mathbf {P} _ {j} \\| _ {2} \\sin \\left(\\cos^ {- 1} \\left(\\frac {\\mathbf {P} _ {j} \\cdot (- \\exp (\\boldsymbol {\\omega}) ^ {\\intercal} \\mathbf {x})}{\\| \\mathbf {P} _ {j} \\| _ {2} \\| - \\exp (\\boldsymbol {\\omega}) ^ {\\intercal} \\mathbf {x} \\| _ {2}}\\right)\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.384, + 0.86, + 0.4 + ], + "angle": 0, + "content": "This depth consistency loss is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.41, + 0.907, + 0.442 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C} = \\sum_ {i} M _ {t} (\\mathbf {p} _ {i}) M _ {t + 1} (\\mathbf {p} _ {j}) \\| \\overline {{\\mathbf {P} _ {i} \\mathbf {H} _ {i}}} - \\overline {{\\mathbf {P} _ {j} \\mathbf {H} _ {j}}} \\|, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.453, + 0.906, + 0.498 + ], + "angle": 0, + "content": "where \\(\\exp()\\) is the exponential mapping operator in Lie algebra, which helps maintain alignment. Our final loss function combines these terms with weighted coefficients:" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.51, + 0.907, + 0.528 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p o s e}} = \\mathcal {L} _ {F} + \\lambda_ {I} \\mathcal {L} _ {I} + \\lambda_ {C} \\mathcal {L} _ {C}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.906, + 0.613 + ], + "angle": 0, + "content": "We start optimization with \\(\\lambda_I = 0\\) and \\(\\lambda_C = 0\\), allowing only the flow loss \\(\\mathcal{L}_F\\) to guide the relative pose alignment with global optical flow. After initial convergence, we increase weights to \\(\\lambda_I = 75\\) and \\(\\lambda_C = 25\\) and continue optimization to a second convergence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.906, + 0.69 + ], + "angle": 0, + "content": "Finally, we apply a point-to-plane ICP algorithm on the two globally aligned point clouds to estimate camera poses accurately. Since optical flow estimation between two blurry images introduces some errors, this final camera pose estimation relies exclusively on depth maps." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.698, + 0.904, + 0.714 + ], + "angle": 0, + "content": "3.2. Global Pose Estimation/Geometry Refinement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Starting from the initial camera poses and point clouds obtained in the previous step, we merge them into a global point cloud within the world coordinate system (Figure 4). To manage memory efficiently, we downsample the point cloud with a specified voxel size \\( s \\). However, due to accumulated errors in the initial camera poses, these poses may not align well with the global point cloud, leading to seam artifacts at image boundaries and a drifting phenomenon that creates loop-closure issues. To resolve this, we refine both the camera poses and 3D point cloud positions simultaneously within the Gaussian Splatting pipeline, ensuring global and geometric consistency." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26428" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.079, + 0.288, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.165, + 0.245, + 0.176 + ], + "angle": 0, + "content": "Input color image" + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.079, + 0.48, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.334, + 0.165, + 0.419, + 0.177 + ], + "angle": 0, + "content": "Input depth map" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.178, + 0.22, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.241, + 0.2, + 0.252 + ], + "angle": 0, + "content": "DROID-SLAM" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.178, + 0.349, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.242, + 0.306, + 0.252 + ], + "angle": 0, + "content": "MonoGS" + }, + { + "type": "image", + "bbox": [ + 0.349, + 0.179, + 0.48, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.398, + 0.242, + 0.425, + 0.252 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.257, + 0.483, + 0.3 + ], + "angle": 0, + "content": "Figure 4. Reconstructed initial point clouds from the RGB-D image sequence. Our method achieves accurate camera poses and generates a dense, high-quality point cloud." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.311, + 0.484, + 0.446 + ], + "angle": 0, + "content": "We initialize the Gaussian Splatting pipeline by loading the point cloud along with the estimated camera poses. To capture every depth value per pixel from the dense depth maps, we initialize 3D Gaussians, fixing their sizes and setting their scale to \\( s / 2 \\). This dense arrangement of 3D Gaussians provides a detailed geometric representation. We then train only the camera poses, 3D positions, and opacities of the Gaussians by comparing a rendered depth map \\( \\widetilde{D} \\) with the input depth map \\( D \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.484, + 0.507 + ], + "angle": 0, + "content": "During training, we randomly select camera viewpoints from among the trainable camera pose variables to render each depth map. For a depth value \\(\\widetilde{D}_t(\\mathbf{p}_i)\\) at pixel \\(\\mathbf{p}_i\\) in the rendered map, we calculate the depth loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.512, + 0.483, + 0.545 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D} = \\sum_ {i} M _ {t} (\\mathbf {p} _ {i}) \\left\\| D _ {t} (\\mathbf {p} _ {i}) - \\widetilde {D} _ {t} (\\mathbf {p} _ {i}) \\right\\| _ {1}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.483, + 0.581 + ], + "angle": 0, + "content": "We also initialize each 3D Gaussian's opacity \\(o_j\\) to 0.5, optimizing them to values of 0 or 1:" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.586, + 0.483, + 0.62 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {o p a c i t y}} = \\sum_ {j} o _ {j} ^ {2} \\left(1 - o _ {j}\\right) ^ {2}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.625, + 0.483, + 0.746 + ], + "angle": 0, + "content": "This optimization aims to retain only Gaussians with an opacity of 1, pruning others to reduce ambiguity for the subsequent deblurring stage. We prune Gaussians with an opacity less than 0.8 and reset opacity with 0.5 again every certain amount of iteration. As iterations progress, the rendered depth maps increasingly align with the input depth maps, indicating that all 3D Gaussian positions and camera poses are globally well-aligned (Figure 5)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.746, + 0.484, + 0.776 + ], + "angle": 0, + "content": "The final refinement loss is a weighted combination of the depth and opacity losses:" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.783, + 0.483, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e f i n e m e n t}} = \\lambda_ {D} \\mathcal {L} _ {D} + \\lambda_ {o} \\mathcal {L} _ {\\text {o p a c i t y}}. \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.805, + 0.236, + 0.82 + ], + "angle": 0, + "content": "3.3. Image Deblur" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.827, + 0.483, + 0.857 + ], + "angle": 0, + "content": "Motion blur caused by camera movement can generally be approximated as:" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.864, + 0.483, + 0.905 + ], + "angle": 0, + "content": "\\[\nI (t) \\approx B (t) = \\frac {1}{k} \\sum_ {i = 0} ^ {k} C _ {i} (t) \\tag {6}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.08, + 0.646, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.153, + 0.645, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.203, + 0.613, + 0.213 + ], + "angle": 0, + "content": "Iteration 100" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.08, + 0.773, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.153, + 0.774, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.203, + 0.742, + 0.213 + ], + "angle": 0, + "content": "Iteration 700" + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.08, + 0.903, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.153, + 0.903, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.801, + 0.203, + 0.878, + 0.213 + ], + "angle": 0, + "content": "Iteration 20000" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.218, + 0.905, + 0.258 + ], + "angle": 0, + "content": "Figure 5. The refinement process corrects misaligned 3D Gaussians by updating camera poses and geometry, with effectiveness demonstrated in Table 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.274, + 0.905, + 0.365 + ], + "angle": 0, + "content": "Here, \\(B(t)\\) represents the motion-blurred image, which we model as the average of \\(k\\) sharp images \\(C_i(t)\\) rendered from the 3D Gaussians at \\(k\\) virtual camera poses interpolated between the camera poses at the start and end of the exposure time period \\(\\pmb{\\xi}_t^s\\) and \\(\\pmb{\\xi}_t^e\\). This approximation allows \\(B(t)\\) to closely match the input image \\(I(t)\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.365, + 0.906, + 0.456 + ], + "angle": 0, + "content": "Using the refined camera poses and point cloud from the previous step, we incorporate depth information to achieve geometrically accurate deblurring, following the approach in recent work [49]. Specifically, we minimize the difference between the input image \\(I(t)\\) and the averaged rendered image \\(B(t)\\), as shown below:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.465, + 0.905, + 0.497 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {B} = \\sum_ {t} \\sum_ {i} \\| I _ {t} (\\mathbf {p} _ {i}) - B _ {t} (\\mathbf {p} _ {i}) \\| _ {1}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.505, + 0.906, + 0.551 + ], + "angle": 0, + "content": "The final deblur loss function integrates this alignment loss with additional terms for structural similarity and depth consistency:" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.561, + 0.905, + 0.579 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d e b l u r}} = \\left(1 - \\lambda_ {B}\\right) \\mathcal {L} _ {B} + \\lambda_ {B} \\mathcal {L} _ {\\mathrm {D} - \\mathrm {S S I M}} + \\lambda_ {D} \\mathcal {L} _ {D}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.587, + 0.906, + 0.647 + ], + "angle": 0, + "content": "This combined loss helps ensure that the deblurred output aligns well with both the input image and depth structure, leading to sharper and more geometrically consistent results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.661, + 0.6, + 0.677 + ], + "angle": 0, + "content": "4. Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Dataset preparation. We prepare both synthetic and real-world datasets to evaluate our method's performance. In the absence of an RGB-D dataset with extreme motion blur, we generate a synthetic reference dataset by rendering directly from 3D models. We create RGB-D image sequences and record per-frame camera trajectories to evaluate our pose estimation and deblurring accuracy. The sequences, ranging from 50 to 150 frames, are captured along simulated trajectories that include rapid translations, 360-degree rotations, and human movement paths (using Blender's Walk Navigation feature). The RGB and depth images have a resolution of \\(640 \\times 480\\), and motion blur is applied by enabling Blender's motion blur setting. The 3D models are sourced from the McGuire archive [23] and the Blender archive [43]." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26429" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.084, + 0.898, + 0.593 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.599, + 0.908, + 0.641 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison of deblurring performance from a novel view. Our method excels in restoring high-frequency details, whereas several methods, including COLMAP [31], fail under severe motion blur conditions (see Section 4.1). The corresponding quantitative results are provided in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.797 + ], + "angle": 0, + "content": "For real scene evaluation, we used an Azure Kinect RGB-D camera with a 33.3ms fixed exposure time, 5000K fixed white balance, NFOV-2 × 2-binned depth capture mode, and \\(1280 \\times 720\\) RGB resolution. We used the provided intrinsic camera parameters to undistort the images and aligned depth maps to color images using the SDK's transformation functions. Depth maps were eroded three times to remove outliers, which commonly occur along edges." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Implementation details. Our code is based on 3D Gaussian Splatting [14], with data processing partially adapted from CF-3DGS [14]. For initial camera pose estimation, we use the Unimatch [48] optical flow network and the ICP [5] algorithm from Open3D as part of the optimization process. For an input image sequence of \\(N\\) frames," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.907, + 0.797 + ], + "angle": 0, + "content": "we prune Gaussians every \\(100N\\) iterations, reset opacity to 0.5 every \\(200N\\) iterations, and run a total of \\(600N\\) iterations (Section 3.2). The number of virtual views for deblurring is set to 15 for synthetic dataset evaluation and 17 for real-scene comparison. All experiments were conducted on an NVIDIA RTX A6000 GPU (48GB), with initial camera pose optimization taking 11 seconds per frame, 0.1 seconds per iteration for pose and geometry refinement, and 0.26 seconds per iteration for deblurring at \\(640 \\times 480\\) resolution." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.736, + 0.82 + ], + "angle": 0, + "content": "4.1.Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We evaluate the camera pose accuracy and deblurring performance on three synthetic scenes from our dataset, each rendered with extreme motion blur. For pose accuracy, we compare our method against a variety of approaches capable of generating a 3D point cloud and camera trajectory:" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26430" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.084, + 0.898, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.625, + 0.908, + 0.669 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison of deblurring performance from novel views on real scenes captured with an Azure Kinect camera. Our method effectively restores high-frequency details, outperforming other methods even when processing RGB-D sequences with large spatial gaps between frames due to rapid camera motion and severe motion blur." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.687, + 0.482, + 0.822 + ], + "angle": 0, + "content": "COLMAP [31], ICP with FPFH feature-based global alignment [29], ORB-SLAM3 [3], ElasticFusion [47], and several RGB-D SLAM methods, including NICE-SLAM [51], Point-SLAM [30], and DROID-SLAM [42]. Additionally, we include recent Gaussian Splatting-based SLAM approaches such as MonoGS [21] and SplaTAM [12]. For each method, we calculate absolute trajectory error (ATE) and relative pose error (RPE) in meter and degree units, using the ground truth camera trajectories as references." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.825, + 0.483, + 0.903 + ], + "angle": 0, + "content": "To evaluate deblurring accuracy, we modify a state-of-the-art Gaussian Splitting-based image restoration method [49] to operate in a depth-aware manner using undistorted color and depth images. We select one out of every three images as a test view and render the correspond" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.907, + 0.763 + ], + "angle": 0, + "content": "ing novel view for comparison with the ground truth. For quantitative metrics, we calculate PSNR, SSIM, and LPIPS on the rendered images, as well as RMSE between the rendered and ground truth depth maps in the inverse depth domain \\((m^{-1})\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.667, + 0.79 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.903 + ], + "angle": 0, + "content": "To validate the contributions of each component in our method, we perform an ablation study evaluating camera pose and deblurring accuracy by selectively omitting key elements. Specifically, we conduct experiments without fixing the scale of the Gaussians (Section 3.2), without the pose and geometry refinement process (Equation 5), and without depth loss (Equation 8)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26431" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of camera pose accuracy against ground truth poses. We report scores only for methods that successfully provide both camera poses and a point cloud (see Section 4.1). Our method achieves the best scores on RPE metrics and shows ATE accuracy comparable to DROID-SLAM [42]. Each color highlights the best and second best results." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.134, + 0.907, + 0.229 + ], + "angle": 0, + "content": "
BedroomLivingroomOffice
ATE ↓RPE (trans) ↓RPE (rot) ↓ATE ↓RPE (trans) ↓RPE (rot) ↓ATE ↓RPE (trans) ↓RPE (rot) ↓
ElasticFusion [47]1.3030.0792.2181.0140.0562.7770.4790.0380.480
FPFH-ICP [29]2.7600.1182.3102.0280.0921.2130.1390.0090.167
DROID-SLAM [42]0.0590.0441.2120.0500.0140.3170.0110.0110.245
MonoGS [21]0.5060.0410.8120.1350.0090.1240.0450.0110.180
Ours0.1020.0060.0920.0050.0020.0320.0410.0030.046
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.241, + 0.907, + 0.27 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison of deblurring performance from novel views. Our method demonstrates high accuracy across all metrics for both color and depth images. RMSE is calculated by comparing the rendered depth with GT depth in the inverse depth domain \\((m^{-1})\\)." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.271, + 0.907, + 0.358 + ], + "angle": 0, + "content": "
BedroomLivingroomOffice
PSNR ↑SSIM ↑LPIPS ↓RMSE ↓PSNR ↑SSIM ↑LPIPS ↓RMSE ↓PSNR ↑SSIM ↑LPIPS ↓RMSE ↓
ElasticFusion [47]18.7660.5860.4661.24913.7990.4480.5540.14619.0520.5330.4010.079
FPFH-ICP [29]20.6230.6170.4180.45018.1720.5980.4220.10222.2120.6840.2410.033
DROID-SLAM [42]21.3750.6660.3020.01321.2220.7350.2500.03525.6800.7690.1440.013
MonoGS [21]23.7280.7170.2850.00924.4140.8220.1880.02724.1940.7250.1840.184
Ours26.7450.8240.2060.00527.6500.9000.1500.01025.6490.7910.1600.014
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.375, + 0.483, + 0.43 + ], + "angle": 0, + "content": "Table 3. Ablation study on the effects of the global refinement process and fixed scale of Gaussians. Our refinement significantly enhances deblurring and depth accuracy while fixing the scale of the Gaussians leads to more accurate camera pose estimation." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.431, + 0.481, + 0.48 + ], + "angle": 0, + "content": "
PSNRSSIMLPIPSRMSEATERPE(trans)RPE(rot)
w/o scale fix21.6560.7030.3080.0270.1550.0531.488
w/o refinement22.2230.7340.2380.0270.1230.0070.047
w/o depth loss25.5180.8070.1700.0140.1030.0120.195
Scale fix + refinement26.6810.8380.1720.0100.1030.0120.195
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.492, + 0.483, + 0.658 + ], + "angle": 0, + "content": "We assess the effectiveness of these components by performing deblurring based on results from Section 3.1—omitting refinement of camera poses and geometry—and by allowing the scale of the 3D Gaussians to be optimized freely. Table 3 demonstrates that both the global refinement process and fixed Gaussian scale are essential for achieving high accuracy. Notably, during pose refinement, the algorithm reduces absolute pose error even at the expense of a slight increase in relative pose error, underscoring the importance of our refinement approach for stable, high-quality reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.658, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Our method achieves superior RPE scores across all scenes, as shown in Table 1, outperforming other approaches in pose and geometry estimation under severe motion blur. While our ATE scores are comparable to those of DROID-SLAM, which uses a bundle adjustment (BA) module, ATE alone does not correlate strongly with deblurring quality. Our method consistently produces more accurate and denser point clouds, which directly contributes to improved deblurring performance, as evidenced by the metrics in Table 2. These results confirm that our integrated approach to pose estimation and deblurring is more resilient to challenging conditions than other methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.303, + 0.866 + ], + "angle": 0, + "content": "4.3. Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Figures 6 and 7 present qualitative comparisons of our method's performance in synthetic and real scenes. The" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.377, + 0.907, + 0.497 + ], + "angle": 0, + "content": "densification capability of Gaussian Splitting provides detailed scene representation even from a relatively sparse point cloud; however, our approach, with its initially dense and accurate point cloud generation, achieves notably superior deblurring performance, especially in restoring high-frequency details. Our method consistently demonstrates a significant advantage in visual fidelity and detail restoration across diverse conditions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.504, + 0.634, + 0.52 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.529, + 0.907, + 0.772 + ], + "angle": 0, + "content": "We have introduced a robust method for 3D scene reconstruction from RGB-D image sequences that effectively addresses the challenges of extreme motion blur and low-light conditions. Our approach leverages optical flow from color images and a carefully designed geometric loss to achieve accurate global alignment, followed by local refinement of camera poses and geometry using ICP. By integrating these steps into a Gaussian Splatting pipeline, we further refine camera poses and 3D geometry by minimizing depth and opacity losses. Additionally, fixing the scale of the 3D Gaussians ensures that the depth information from the RGB-D input is fully utilized, allowing for precise camera pose estimation and improved deblurring performance. Our method demonstrates superior performance on real and synthetic RGB-D scenes with significant motion blur, outperforming existing approaches." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.781, + 0.682, + 0.798 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.801, + 0.906, + 0.892 + ], + "angle": 0, + "content": "Min H. Kim acknowledges the Samsung Research Funding & Incubation Center (SRFC-IT2402-02), the Korea NRF grant (RS-2024-00357548), the MSIT/IITP of Korea (RS2022-00155620, RS-2024-00398830, 2022-0-00058, and 2017-0-00072), Microsoft Research Asia, and Samsung Electronics." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26432" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Dejan Azinović, Ricardo Martin-Brualla, Dan B Goldman, Matthias Nießner, and Justus Thies. Neural rgb-d surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6290-6301, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[2] Michael Bloesch, Jan Czarnowski, Ronald Clark, Stefan Leutenegger, and Andrew J Davison. Codeslam—learning a compact, optimisable representation for dense visual slam. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2560–2568, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.483, + 0.324 + ], + "angle": 0, + "content": "[3] Carlos Campos, Richard Elvira, Juan J Gómez Rodríguez, José MM Montiel, and Juan D Tardós. Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics, 37(6):1874-1890, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.326, + 0.483, + 0.367 + ], + "angle": 0, + "content": "[4] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. arXiv preprint arXiv:2204.04676, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.482, + 0.409 + ], + "angle": 0, + "content": "[5] Yang Chen and Gérard Medioni. Object modelling by registration of multiple range images. Image and vision computing, 10(3):145-155, 1992. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.41, + 0.483, + 0.478 + ], + "angle": 0, + "content": "[6] Angela Dai, Matthias Nießner, Michael Zollofer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface re-integration. ACM Transactions on Graphics 2017 (TOG), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.483, + 0.535 + ], + "angle": 0, + "content": "[7] Rob Fergus, Barun Singh, Aaron Hertzmann, Sam T Roweis, and William T Freeman. Removing camera shake from a single photograph. In Acm Siggraph 2006 Papers, pages 787-794. 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.537, + 0.482, + 0.578 + ], + "angle": 0, + "content": "[8] Kyle Genova, Forrester Cole, Avneesh Sud, Aaron Sarna, and Thomas Funkhouser. Deep structured implicit functions. arXiv preprint arXiv:1912.06126, 2:2, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.58, + 0.483, + 0.647 + ], + "angle": 0, + "content": "[9] Zan Gojcic, Caifa Zhou, Jan D Wegner, and Andreas Wieser. The perfect match: 3d point cloud matching with smoothed densities. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5545-5554, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.649, + 0.483, + 0.691 + ], + "angle": 0, + "content": "[10] Alex Graves and Alex Graves. Long short-term memory. Supervised sequence labelling with recurrent neural networks, pages 37-45, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.692, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[11] Olaf Kähler, Victor Prisacariu, Julien Valentin, and David Murray. Hierarchical voxel block hashing for efficient integration of depth images. IEEE Robotics and Automation Letters, 1(1):192-197, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.748, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[12] Nikhil Keetha, Jay Karhade, Krishna Murthy Jatavallabhula, Gengshan Yang, Sebastian Scherer, Deva Ramanan, and Jonathon Luiten. Splatam: Splat track & map 3d gaussians for dense rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21357-21366, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Maik Keller, Damien Lefloch, Martin Lambers, Shahram Izadi, Tim Weyrich, and Andreas Kolb. Real-time 3d reconstruction in dynamic scenes using point-based fusion. In 2013 International Conference on 3D Vision-3DV 2013, pages 1-8. IEEE, 2013. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[14] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[15] Christian Kerl, Jürgen Sturm, and Daniel Cremers. Dense visual slam for rgb-d cameras. In 2013 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 2100-2106. IEEE, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[16] Dilip Krishnan and Rob Fergus. Fast image deconvolution using hyper-laplacian priors. Advances in neural information processing systems, 22, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[17] Dogyoon Lee, Minhyeok Lee, Chajin Shin, and Sangyoun Lee. Dp-nerf: Deblurred neural radiance field with physical scene priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12386–12396, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[18] Yang Li and Tatsuya Harada. Lepard: Learning partial point cloud matching in rigid and deformable scenes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5554-5564, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.378, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[19] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-nerf: Neural radiance fields from blurry images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12861–12870, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[20] Nico Marniok, Ole Johannsen, and Bastian Goldluecke. An efficient octree design for local variational range image fusion. In Pattern Recognition: 39th German Conference, GCPR 2017, Basel, Switzerland, September 12-15, 2017, Proceedings 39, pages 401-412. Springer, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[21] Hidenobu Matsuki, Riku Murai, Paul HJ Kelly, and Andrew J Davison. Gaussian splatting slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18039-18048, 2024. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.576, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[22] John McCormac, Ankur Handa, Andrew Davison, and Stefan Leutenegger. Semantic fusion: Dense 3d semantic mapping with convolutional neural networks. In 2017 IEEE International Conference on Robotics and automation (ICRA), pages 4628-4635. IEEE, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.646, + 0.874, + 0.659 + ], + "angle": 0, + "content": "[23] Morgan McGuire. Computer graphics archive, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.662, + 0.905, + 0.728 + ], + "angle": 0, + "content": "[24] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[25] Raul Mur-Artal and Juan D Tardós. Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras. IEEE transactions on robotics, 33(5):1255-1262, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.856 + ], + "angle": 0, + "content": "[26] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3883-3891, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26433" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "Kinectfusion: Real-time dense surface mapping and tracking. In 2011 10th IEEE international symposium on mixed and augmented reality, pages 127-136. IEEE, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.19 + ], + "angle": 0, + "content": "[28] Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Marc Stamminger. Real-time 3d reconstruction at scale using voxel hashing. ACM Transactions on Graphics (ToG), 32 (6):1-11, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.484, + 0.248 + ], + "angle": 0, + "content": "[29] Radu Bogdan Rusu, Nico Blodow, and Michael Beetz. Fast point feature histograms (fpfh) for 3d registration. In 2009 IEEE international conference on robotics and automation, pages 3212-3217. IEEE, 2009. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.483, + 0.305 + ], + "angle": 0, + "content": "[30] Erik Sandström, Yue Li, Luc Van Gool, and Martin R Oswald. Point-slam: Dense neural point cloud-based slam. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18433-18444, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.483, + 0.359 + ], + "angle": 0, + "content": "[31] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.363, + 0.484, + 0.419 + ], + "angle": 0, + "content": "[32] Thomas Schops, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 134-144, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.421, + 0.483, + 0.475 + ], + "angle": 0, + "content": "[33] Thomas Schöps, Torsten Sattler, and Marc Pollefeys. Surfelmeshing: Online surfel-based mesh reconstruction. IEEE transactions on pattern analysis and machine intelligence, 42(10):2494-2507, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.483, + 0.532 + ], + "angle": 0, + "content": "[34] Christian J Schuler, Michael Hirsch, Stefan Harmeling, and Bernhard Scholkopf. Learning to deblur. IEEE transactions on pattern analysis and machine intelligence, 38(7):1439-1451, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.483, + 0.589 + ], + "angle": 0, + "content": "[35] Otto Seiskari, Jerry Yilammi, Valtteri Kaatrasalo, Pekka Rantalankila, Matias Turkulainen, Juho Kannala, and Arno Solin. Gaussian splattering on the move: Blur and rolling shutter compensation for natural camera motion, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.483, + 0.632 + ], + "angle": 0, + "content": "[36] Qi Shan, Jiaya Jia, and Aseem Agarwala. High-quality motion deblurring from a single image. Acm transactions on graphics (tog), 27(3):1-10, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.634, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[37] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. Advances in neural information processing systems, 28, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.483, + 0.759 + ], + "angle": 0, + "content": "[38] Jörg Stuckler and Sven Behnke. Multi-resolution surfel maps for efficient dense 3d modeling and tracking. Journal of Visual Communication and Image Representation, 25(1):137-147, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.483, + 0.831 + ], + "angle": 0, + "content": "[39] Shuochen Su, Maurizio Delbracio, Jue Wang, Guillermo Sapiro, Wolfgang Heidrich, and Oliver Wang. Deep video deblurring for hand-held cameras. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1279-1288, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[40] Jian Sun, Wenfei Cao, Zongben Xu, and Jean Ponce. Learning a convolutional neural network for non-uniform motion blur removal. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 769-777, 2015. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[41] Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.149, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[42] Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. Advances in neural information processing systems, 2021. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.206, + 0.857, + 0.22 + ], + "angle": 0, + "content": "[43] Flavio Della Tommasa. Blender demo files, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.222, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[44] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. BAD-NeRF: Bundle Adjusted Deblur Neural Radiance Fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4170–4179, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.292, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[45] Chen Wenbo and Liu Ligang. Deblur-gs: 3d gaussian splattering from camera motion blurred images. Proc. ACM Comput. Graph. Interact. Tech. (Proceedings of I3D 2024), 7(1), 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.348, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[46] Thomas Whelan, Michael Kaess, Hordur Johannsson, Maurice Fallon, John J Leonard, and John McDonald. Real-time large-scale densergb-d slam with volumetric fusion. The International Journal of Robotics Research, 34(4-5):598-626, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.418, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[47] Thomas Whelan, Renato F Salas-Moreno, Ben Glocker, Andrew J Davison, and Stefan Leutenegger. Elasticfusion: Real-time dense slam and light source estimation. The International Journal of Robotics Research, 35(14):1697-1716, 2016. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.489, + 0.905, + 0.544 + ], + "angle": 0, + "content": "[48] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, Fisher Yu, Dacheng Tao, and Andreas Geiger. Unifying flow, stereo and depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.546, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[49] Lingzhe Zhao, Peng Wang, and Peidong Liu. BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splatting. 2024. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.589, + 0.907, + 0.642 + ], + "angle": 0, + "content": "[50] Huizhong Zhou, Benjamin Ummenhofer, and Thomas Brox. Deeptam: Deep tracking and mapping. In Proceedings of the European conference on computer vision (ECCV), pages 822-838, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.645, + 0.905, + 0.726 + ], + "angle": 0, + "content": "[51] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12786-12796, 2022. 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.092, + 0.907, + 0.726 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26434" + } + ] +] \ No newline at end of file diff --git a/2025/Splat-based 3D Scene Reconstruction with Extreme Motion-blur/73a5b991-b304-4700-8590-8334fbf2bdee_origin.pdf b/2025/Splat-based 3D Scene Reconstruction with Extreme Motion-blur/73a5b991-b304-4700-8590-8334fbf2bdee_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7f891952cd2cedc1965b01be5440d8dde6614a93 --- /dev/null +++ b/2025/Splat-based 3D Scene Reconstruction with Extreme Motion-blur/73a5b991-b304-4700-8590-8334fbf2bdee_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:511c0a7ba632953cfdcffe0d742682f343430614ed32a9890e6dc228caf6ea58 +size 3342971 diff --git a/2025/Splat-based 3D Scene Reconstruction with Extreme Motion-blur/full.md b/2025/Splat-based 3D Scene Reconstruction with Extreme Motion-blur/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ee835131d0fc9818d2ffce422fe5d3b30a669f11 --- /dev/null +++ b/2025/Splat-based 3D Scene Reconstruction with Extreme Motion-blur/full.md @@ -0,0 +1,350 @@ +# Splat-based 3D Scene Reconstruction with Extreme Motion-blur + +Hyeonjoong Jang†* + +Dongyoung Choi† +† KAIST + +Donggun Kim $^{\dagger}$ Woohyun Kang $^{\dagger}$ +* HYPERGRAM + +Min H. Kim†* + +{hjjang; dychoi; dgkim; whkang; minhkim}@vclab.kaist.ac.kr + +![](images/8148e79216c51fa76cfd8cd39f2083c8c2e187dfed867ff6e83d7798d0aeb7a1.jpg) + +![](images/3364c044bf5b03f3d65916a7cfcaa64f0813a1a389d52f3f42e8489bc2aae960.jpg) +Input +out + +![](images/2028d66bd913728fe7c1763ca132e1251ce9c1f26dd9dfad91350bbe78d58980.jpg) + +![](images/0cd06fcab1244798ddc1c096785dbe6f73dc615a4a6e5a401b04bf88573d6e35.jpg) +ElasticFusion +Figure 1. We propose a robust 3D scene reconstruction method from RGB-D input that effectively addresses extreme motion blur. Our approach achieves accurate camera poses and dense point clouds, producing clearer, deblurred scenes compared to existing methods. + +![](images/2dd837bf8b329554cdfcb60a391ce455e5829ef9b8d9d9ecfe14f723b1e732be.jpg) + +![](images/0736a2bd1733fe9528015726d26bd4aadf52fe4862261af60a90203859754934.jpg) + +![](images/6d12761ac801f1e68e66ce23de016c5453509695c66e3030268357845f31825a.jpg) +DROID-SLAM + +![](images/dab89590977b7d2f0ffab195557d5456ff3b9e05450a21199942c3d1bd7837c5.jpg) + +![](images/98007c7640395b4a7ea0fd334de03eee117587749d63b69d2917e0865e8b0982.jpg) +MonoGS + +![](images/d227d0808ee008285d54552f1ad14a28d5b7af7f569d4cfff1f5f9b5366ae8fe.jpg) + +![](images/eeb28c3f72a0685f986a2b85bf64539fd8f7ebb4ad6fd0ceec92291c16c1356f.jpg) +Ours + +![](images/9bd8095922966743884d2b6d3357fcf3d227469ec8b33f13417f011982375a8b.jpg) + +# Abstract + +We propose a splat-based 3D scene reconstruction method from RGB-D input that effectively handles extreme motion blur, a frequent challenge in low-light environments. Under dim illumination, RGB frames often suffer from severe motion blur due to extended exposure times, causing traditional camera pose estimation methods, such as COLMAP, to fail. This results in inaccurate camera pose and blurry color input, compromising the quality of 3D reconstructions. Although recent 3D reconstruction techniques like Neural Radiance Fields and Gaussian Splatting have demonstrated impressive results, they rely on accurate camera trajectory estimation, which becomes challenging under fast motion or poor lighting conditions. Furthermore, rapid camera movement and the limited field of view of depth sensors reduce point cloud overlap, limiting the effectiveness of pose estimation with the ICP algorithm. To address these issues, we introduce a method that combines camera pose estimation and image deblurring using a Gaussian Splatting framework, leveraging both 3D Gaussian splats and depth inputs for enhanced scene representation. Our method first aligns consecutive RGB-D frames through optical flow and ICP, then refines camera poses and 3D geometry by adjusting Gaussian positions for optimal depth alignment. To handle motion blur, we model camera movement during exposure and deblur images by comparing the input with a series of sharp, rendered frames. Experiments on a new RGB-D dataset with extreme motion blur show that our method outperforms existing approaches, + +enabling high-quality reconstructions even in challenging conditions. This approach has broad implications for 3D mapping applications in robotics, autonomous navigation, and augmented reality. Both code and dataset are publicly available on https://github.com/KAIST-VCLAB/gs-extreme-motion-blur. + +# 1. Introduction + +High-quality 3D scene reconstruction is one of the most important and challenging applications in computer vision. The accuracy of 3D reconstruction hinges on the quality of essential components, such as camera poses, RGB images, and depth maps. These elements are interconnected and influence each other; a failure in one component can adversely impact the others. For example, blurry RGB images or noisy depth maps can significantly impair camera pose estimation, which, in turn, degrades overall reconstruction quality. This challenge is particularly pronounced when input data is captured in low-light conditions or during rapid camera movement—common scenarios in everyday, casual capture. Such conditions result in degraded color and depth frames, leading to poorly estimated camera poses and reconstructions with blurry textures and distorted or smoothed geometry. Therefore, achieving high-quality 3D reconstruction requires clear RGB images and accurate depth maps. + +The interdependence of these input components suggests an opportunity for compensating or restoring degraded elements. However, this process presents a chicken-and-egg + +problem, complicating the reconstruction further. For instance, accurate camera pose estimation requires sharp images, while deblurring motion-blurred images depend on reliable camera pose information. Although numerous methods [4, 41] exist for recovering sharp images from motion-blurred ones, they often rely on supervised learning models trained on specific datasets. Consequently, their deblurring performance can degrade significantly when faced with new cameras or motion blur scenarios that differ from the training data. + +Recent works [35, 44, 45, 49] have made advances in addressing these challenges by optimizing camera trajectories during exposure while simultaneously learning sharp RGB colors. This enables simultaneous camera pose estimation and image deblurring. However, these methods require an initial camera pose and, when using Gaussian Splatting [14], also need a sparse point cloud as input. This requirement limits their applicability, especially when severe motion blur prevents structure-from-motion (SfM) methods like COLMAP [31] from functioning effectively, making it impossible to start the optimization process. + +To overcome these limitations, we introduce a robust approach for 3D scene reconstruction that effectively compensates for severe motion blur and addresses the challenges of camera pose estimation without relying on a precise initial pose. Our method leverages RGB-D inputs and incorporates both optical flow and depth information to align camera poses accurately, even in the presence of challenging motion blur and lighting conditions. We first perform a global alignment between consecutive frames using optical flow and the ICP algorithm, which enables effective local alignment of point clouds generated from the depth maps. + +After this initial alignment, we refine the camera poses and 3D geometry by integrating them into a Gaussian Splating pipeline. This approach allows us to initialize dense 3D Gaussians from depth maps, which we scale to ensure a detailed representation of the scene geometry. Our refinement process iteratively adjusts both camera poses and the positions of 3D Gaussians, achieving global consistency by minimizing a depth alignment loss that compares rendered depth maps with input depth measurements. Through this adjustment, we eliminate loop-closure artifacts and reduce the drift that often accumulates in traditional SLAM-based methods. + +For scenes with significant motion blur, we further optimize the deblurring process by modeling the camera poses at the start and end of each frame's exposure. This temporal modeling allows us to simulate the effects of motion during the exposure period, which we incorporate into our Gaussian Splatting framework. We minimize the difference between the observed motion-blurred image and the average of a set of sharp images rendered from multiple viewpoints along the exposure trajectory, resulting in more accurate, + +geometrically consistent deblurring. Our deblur loss function combines image alignment, structural similarity, and depth consistency, ensuring that the final reconstructions retain sharpness and fidelity to the original scene structure. + +Through extensive evaluation on a newly constructed RGB-D dataset featuring extreme motion blur, we demonstrate that our approach significantly outperforms existing methods in both accuracy and robustness. By addressing the key challenges of pose estimation, depth alignment, and image deblurring, our method provides a versatile and effective solution for high-quality 3D reconstruction under real-world capture conditions, such as low light and fast motion. The proposed method holds promising applications in areas requiring reliable 3D mapping and reconstruction, including robotics, autonomous navigation, and augmented reality. We will make both our dataset and implementation publicly available to foster further research and development in this field. + +# 2. Related Work + +RGB-D scene reconstruction. A popular approach to 3D geometry reconstruction defines a 3D voxel grid and constructs a signed distance function (SDF) volume by accumulating depth estimates captured by sensors. Kinect-Fusion [27] uses a Kinect camera to capture RGB-D images, then estimates relative camera poses between adjacent frames with the ICP algorithm. Next, it constructs and updates a truncated signed distance function (TSDF) volume, enabling real-time 3D reconstruction of objects. Niessner et al. [28] improve memory efficiency by combining hashing algorithms to store and update SDF values only where needed, thus reducing memory requirements. + +Several methods [8, 11, 20, 22] adopt memory-efficient, hierarchical data structures for depth fusion. BundleFusion [6] simultaneously refines camera poses and geometry using bundle adjustment, improving reconstruction accuracy and addressing the loop closure problem caused by accumulated camera pose and depth estimation errors. ElasticFusion [47] builds a map with surfels—point primitives containing position, normal, and radius—and updates the map through a deformation graph to minimize error and perform loop closure. Other methods [13, 33, 38] directly fuse depth estimates into point clouds, optimizing memory use. However, these methods generally require slow camera motion to capture sharp images and accurately estimate camera transformations. + +Simultaneous localization and mapping (SLAM) methods [3, 15, 25, 32, 46] track camera motion and construct a map in real-time by extracting visual features from input images to establish 3D point correspondences. CodeSLAM [2] and DeepTAM [50] use neural networks to combine depth maps with color images, while Azinović + +![](images/c45ed40390f47755f0dc6308204674dc878c1ccd30a154ffa616bf7d1259fb64.jpg) +Figure 2. Overview of our method pipeline. We begin with global alignment by comparing optical flow and reprojected pixel differences between adjacent RGB-D image pairs, followed by local estimation of relative camera poses using ICP. Each depth map and camera pose is then used to initialize 3D Gaussians. By fixing the scale of the Gaussians and optimizing with depth and opacity losses, we achieve global refinement of camera poses and geometry. Finally, we deblur the scene texture by minimizing the deblur loss between the input image and averaged virtual images. + +et al. [1] train multi-layer perceptrons (MLP) to learn SDF and color values for scene reconstruction. Their approach also optimizes camera transformation correction variables during training to refine poses. However, these methods struggle with input images that contain motion blur. Nice-SLAM [51], Point-SLAM [30], and DROID-SLAM [42] use neural networks to track camera poses. SplaTAM [12] and MonoGS [21] introduced SLAM solutions based on 3D Gaussian Splatting. However, these methods are not designed to handle motion blur. + +Image deblurring. Motion blur occurs when camera motion during the exposure time changes the projected pixel coordinates of rays on the camera sensor. Early studies [7, 16, 36] tackle this issue by developing kernels to restore blurry images. Convolutional neural networks (CNNs) demonstrate powerful performance in learning deblur kernels [34, 40], and deep learning techniques such as ResNet with skip connections and multi-scale networks further improved restoration quality [26, 39]. SRN-DeblurNet [41] adopt recurrent networks [10, 37] and combine them with an encoder-decoder network structure, while NAFNet [4] simplify the network structure, extracting only essential components and proposing a nonlinear activation-free network. However, since these methods rely on curated datasets like the GoPro dataset [26], they struggle to effectively deblur images with extreme motion blur or those captured on different cameras. + +There are NeRF-based approaches [24] that tackle motion blur in input images. Deblur-NeRF [19] and DP-NeRF [17] use a set of motion-blurred images along with camera poses estimated by COLMAP as input to restore sharp images through NeRF optimization. BAD-NeRF [44] enhances this by jointly optimizing virtual camera poses and radiance fields during exposure. The latest approaches utilize Gaussian Splatting [14], which offers faster training times than NeRF by rendering a set of 3D Gaussians with a dedicated rasterizer rather than optimizing a neural network. Gaussian Splatting-based deblurring meth + +ods [35, 45, 49] optimize the virtual camera trajectory and minimize differences between input blurry images and the average of rendered images at each virtual camera position. + +All of these methods rely on initial camera poses, typically obtained from COLMAP, and for Gaussian Splatting-based methods, a sparse point cloud as well. However, structure-from-motion methods like COLMAP often fail with blurry inputs—such as images captured by fast-moving cameras or in low-light conditions that require long exposure times. Even with dense depth inputs, ICP requires acceptable global alignment, and recent point cloud matching methods [9, 18] cannot accurately determine relative camera transformations between two point clouds if they consist mainly of planar structures. + +To address these issues, we propose an effective camera pose estimation algorithm tailored for RGB-D image sequences that jointly refines both the camera trajectory and the reconstructed 3D point cloud, resulting in improved sharpness in 3D reconstruction. + +# 3. Method + +Overview. In datasets with significant motion blur, using COLMAP to estimate camera poses becomes impractical. Additionally, rapid camera motion and the limited field of view of depth sensors further complicate this task, making ICP registration less effective. To address these challenges, we first perform global alignment using both RGB and depth images, which establishes sufficient initial alignment for the ICP algorithm to operate effectively. After achieving this global alignment, we apply ICP again to refine and accurately estimate the camera poses. However, estimating camera poses by analyzing only two consecutive frames in a sequence can result in accumulated errors, leading to drift over time. To resolve this issue, it is crucial to adjust the 3D geometry and align the camera poses across the entire sequence so that depth maps and camera views from all perspectives remain well-aligned. We accomplish this through bundle adjustment, optimizing the positions of + +![](images/528a4657a2fdc0e9e33f5463e6b9ae45155ec1baac2556b41a81a71da06a5775.jpg) +Figure 3. The illustration of our pose initialization loss. When $\mathbf{p}_i$ is reprojected to $\mathbf{p}_j$ using the depth value, it is considered geometrically consistent if $\overline{\mathbf{P}_i\mathbf{H}_i}$ equals $\overline{\mathbf{P}_j\mathbf{H}_j}$ . + +3D Gaussians and camera poses within the Gaussian Splating pipeline. Finally, to account for motion blur, we approximate the camera trajectory over time and optimize the movement during exposure, enabling effective deblurring. See Figure 2 for an overview. + +# 3.1. Local Pose Estimation + +To estimate relative camera transformations, we select consecutive frames from the input RGB-D image sequence. Our approach begins with global alignment using optical flow, followed by local alignment of the two-point clouds using the ICP algorithm. For each consecutive RGB image and depth map pair at times $t$ and $t + 1$ , we optimize the transformation $\xi \in \mathfrak{se}(3)$ between the frames. Here, $\xi_{t\rightarrow t + 1} = [\mathbf{x}^{\intercal},\boldsymbol{\omega}^{\intercal}]^{\intercal}$ , where $\mathbf{x}$ represents the translation vector between the two camera origins, and $\boldsymbol{\omega}$ is the rotation element of $\mathfrak{se}(3)$ . We ultimately calculate a camera-to-world transformation matrix $\xi_t$ for each timestamp $t$ . + +To achieve this, we first compute the difference between the estimated optical flow and the reprojected pixel from the camera at time $t$ to $t + 1$ using the depth estimate for each pixel in the depth map $D_{t}$ and the intrinsic matrix $\mathbf{K}$ . A pixel $\mathbf{p}_i$ in the depth map is backprojected to 3D space as: $\mathbf{P}_i = \pi^{-1}(\mathbf{K},\mathbf{p}_i,D_t(\mathbf{p}_i))$ . We then project this 3D point to the camera at $t + 1$ : $\mathbf{p_j} = \pi (\mathbf{K},\exp (\xi_{t\to t + 1}^{\wedge})\mathbf{P}_i)$ , where $\pi$ and $\pi^{-1}$ represent the projection and backprojection operators, respectively. The pixel movement from $\mathbf{p}_i$ to $\mathbf{p}_j$ is then compared with the estimated optical flow $F_{t\rightarrow t + 1}(\mathbf{p}_i)$ to compute the optical flow loss: + +$$ +\mathcal {L} _ {F} = \sum_ {i} M _ {t} (\mathbf {p} _ {i}) M _ {t + 1} (\mathbf {p} _ {j}) \| F _ {t \to t + 1} (\mathbf {p} _ {i}) - (\mathbf {p} _ {j} - \mathbf {p} _ {i}) \| _ {2}. +$$ + +In this formulation, $M_{t}$ is a mask indicating valid depth values in $D_{t}$ at pixel $\mathbf{p}_{i}$ . The mask $M_{t}$ dynamically updates during optimization, as $\mathbf{p}_{j}$ depends on the transformation $\xi$ : $M_{t}(\mathbf{p}_{i}) = D_{t}(\mathbf{p}_{i}) > 0$ . + +We employ a state-of-the-art pretrained optical flow estimation model [48] designed hierarchically to produce reliable optical flow even from two blurry input images. Next, we calculate the photometric loss between color values at + +pixel $\mathbf{p}_i$ in image $I_{t}$ and pixel $\mathbf{p}_j$ in image $I_{t + 1}$ : + +$$ +\mathcal {L} _ {I} = \sum_ {i} M _ {t} (\mathbf {p} _ {i}) M _ {t + 1} (\mathbf {p} _ {j}) \| I _ {t} (\mathbf {p} _ {i}) - I _ {t + 1} (\mathbf {p} _ {j}) \| _ {1}. +$$ + +To ensure geometric consistency, we introduce a depth consistency loss by using the two depth maps captured by the depth camera to optimize relative poses. Specifically, as illustrated in Figure 3, we compute and compare the vertical components $\overline{\mathbf{P}_i\mathbf{H}_i}$ and $\overline{\mathbf{P}_j\mathbf{H}_j}$ based on the baseline vector $\overline{\mathbf{O}_t\mathbf{O}_{t + 1}}$ between the two cameras. We calculate $\overline{\mathbf{P}_i\mathbf{H}_i}$ as follows: + +$$ +\overline {{\mathbf {P} _ {i} \mathbf {H} _ {i}}} = \| \mathbf {P} _ {i} \| _ {2} \sin \left(\cos^ {- 1} \left(\frac {\mathbf {P} _ {i} \cdot \mathbf {x}}{\| \mathbf {P} _ {i} \| _ {2} \| \mathbf {x} \| _ {2}}\right)\right). +$$ + +Similarly, we compute $\overline{\mathbf{P}_j\mathbf{H}_j}$ as: + +$$ +\overline {{\mathbf {P} _ {j} \mathbf {H} _ {j}}} = \| \mathbf {P} _ {j} \| _ {2} \sin \left(\cos^ {- 1} \left(\frac {\mathbf {P} _ {j} \cdot (- \exp (\boldsymbol {\omega}) ^ {\intercal} \mathbf {x})}{\| \mathbf {P} _ {j} \| _ {2} \| - \exp (\boldsymbol {\omega}) ^ {\intercal} \mathbf {x} \| _ {2}}\right)\right). +$$ + +This depth consistency loss is calculated as follows: + +$$ +\mathcal {L} _ {C} = \sum_ {i} M _ {t} (\mathbf {p} _ {i}) M _ {t + 1} (\mathbf {p} _ {j}) \| \overline {{\mathbf {P} _ {i} \mathbf {H} _ {i}}} - \overline {{\mathbf {P} _ {j} \mathbf {H} _ {j}}} \|, \tag {1} +$$ + +where $\exp()$ is the exponential mapping operator in Lie algebra, which helps maintain alignment. Our final loss function combines these terms with weighted coefficients: + +$$ +\mathcal {L} _ {\text {p o s e}} = \mathcal {L} _ {F} + \lambda_ {I} \mathcal {L} _ {I} + \lambda_ {C} \mathcal {L} _ {C}. \tag {2} +$$ + +We start optimization with $\lambda_I = 0$ and $\lambda_C = 0$ , allowing only the flow loss $\mathcal{L}_F$ to guide the relative pose alignment with global optical flow. After initial convergence, we increase weights to $\lambda_I = 75$ and $\lambda_C = 25$ and continue optimization to a second convergence. + +Finally, we apply a point-to-plane ICP algorithm on the two globally aligned point clouds to estimate camera poses accurately. Since optical flow estimation between two blurry images introduces some errors, this final camera pose estimation relies exclusively on depth maps. + +# 3.2. Global Pose Estimation/Geometry Refinement + +Starting from the initial camera poses and point clouds obtained in the previous step, we merge them into a global point cloud within the world coordinate system (Figure 4). To manage memory efficiently, we downsample the point cloud with a specified voxel size $s$ . However, due to accumulated errors in the initial camera poses, these poses may not align well with the global point cloud, leading to seam artifacts at image boundaries and a drifting phenomenon that creates loop-closure issues. To resolve this, we refine both the camera poses and 3D point cloud positions simultaneously within the Gaussian Splatting pipeline, ensuring global and geometric consistency. + +![](images/240642e034bbd39f7e39698d52f74ef4049cb02e47960e7012347066845330f2.jpg) +Input color image + +![](images/eaa9e2f5231815953c24155779cac99f8f10632e06766e37a17deffd12748646.jpg) +Input depth map + +![](images/ff9a09ca1da2b85b480d6029603136c185d8538391ba56c461d748063bea0af6.jpg) +DROID-SLAM + +![](images/7cb32aaf4b47cb9da9c1a0a64fae1c50678584e4607a84f0760ee431aec63553.jpg) +MonoGS +Figure 4. Reconstructed initial point clouds from the RGB-D image sequence. Our method achieves accurate camera poses and generates a dense, high-quality point cloud. + +![](images/8bd656b686752171a2fd64d2b71b711c74e1be9d51d7b89072b949957a7913ac.jpg) +Ours + +We initialize the Gaussian Splatting pipeline by loading the point cloud along with the estimated camera poses. To capture every depth value per pixel from the dense depth maps, we initialize 3D Gaussians, fixing their sizes and setting their scale to $s / 2$ . This dense arrangement of 3D Gaussians provides a detailed geometric representation. We then train only the camera poses, 3D positions, and opacities of the Gaussians by comparing a rendered depth map $\widetilde{D}$ with the input depth map $D$ . + +During training, we randomly select camera viewpoints from among the trainable camera pose variables to render each depth map. For a depth value $\widetilde{D}_t(\mathbf{p}_i)$ at pixel $\mathbf{p}_i$ in the rendered map, we calculate the depth loss as: + +$$ +\mathcal {L} _ {D} = \sum_ {i} M _ {t} (\mathbf {p} _ {i}) \left\| D _ {t} (\mathbf {p} _ {i}) - \widetilde {D} _ {t} (\mathbf {p} _ {i}) \right\| _ {1}. \tag {3} +$$ + +We also initialize each 3D Gaussian's opacity $o_j$ to 0.5, optimizing them to values of 0 or 1: + +$$ +\mathcal {L} _ {\text {o p a c i t y}} = \sum_ {j} o _ {j} ^ {2} \left(1 - o _ {j}\right) ^ {2}. \tag {4} +$$ + +This optimization aims to retain only Gaussians with an opacity of 1, pruning others to reduce ambiguity for the subsequent deblurring stage. We prune Gaussians with an opacity less than 0.8 and reset opacity with 0.5 again every certain amount of iteration. As iterations progress, the rendered depth maps increasingly align with the input depth maps, indicating that all 3D Gaussian positions and camera poses are globally well-aligned (Figure 5). + +The final refinement loss is a weighted combination of the depth and opacity losses: + +$$ +\mathcal {L} _ {\text {r e f i n e m e n t}} = \lambda_ {D} \mathcal {L} _ {D} + \lambda_ {o} \mathcal {L} _ {\text {o p a c i t y}}. \tag {5} +$$ + +# 3.3. Image Deblur + +Motion blur caused by camera movement can generally be approximated as: + +$$ +I (t) \approx B (t) = \frac {1}{k} \sum_ {i = 0} ^ {k} C _ {i} (t) \tag {6} +$$ + +![](images/164be333210da6b4a6d4068eee8e3c6b1251d2805000ba79443ba1baef62c0db.jpg) + +![](images/dd21ab2a4f4b577f34e8fa12b5459d4dbefcefca81e5a90d71de5492a2a8cad8.jpg) +Iteration 100 +Figure 5. The refinement process corrects misaligned 3D Gaussians by updating camera poses and geometry, with effectiveness demonstrated in Table 3. + +![](images/b7a7242cd27897475e6ae65ffea3624d168713cc42fc97f33558d7801b02828e.jpg) + +![](images/426e721ff4e3e49c7249f59700eac77120e5c0c75a0f93d0992916877512c6fa.jpg) +Iteration 700 + +![](images/9cf0d2fc9dba3009e263cf7fb0af56a6050dc15af3b345732d813391a06a9d61.jpg) + +![](images/fa0ef4ca7b74b4f3581200f478bf0ff41b19d82859438d3b1586740577121153.jpg) +Iteration 20000 + +Here, $B(t)$ represents the motion-blurred image, which we model as the average of $k$ sharp images $C_i(t)$ rendered from the 3D Gaussians at $k$ virtual camera poses interpolated between the camera poses at the start and end of the exposure time period $\pmb{\xi}_t^s$ and $\pmb{\xi}_t^e$ . This approximation allows $B(t)$ to closely match the input image $I(t)$ . + +Using the refined camera poses and point cloud from the previous step, we incorporate depth information to achieve geometrically accurate deblurring, following the approach in recent work [49]. Specifically, we minimize the difference between the input image $I(t)$ and the averaged rendered image $B(t)$ , as shown below: + +$$ +\mathcal {L} _ {B} = \sum_ {t} \sum_ {i} \| I _ {t} (\mathbf {p} _ {i}) - B _ {t} (\mathbf {p} _ {i}) \| _ {1}. \tag {7} +$$ + +The final deblur loss function integrates this alignment loss with additional terms for structural similarity and depth consistency: + +$$ +\mathcal {L} _ {\text {d e b l u r}} = \left(1 - \lambda_ {B}\right) \mathcal {L} _ {B} + \lambda_ {B} \mathcal {L} _ {\mathrm {D} - \mathrm {S S I M}} + \lambda_ {D} \mathcal {L} _ {D}. \tag {8} +$$ + +This combined loss helps ensure that the deblurred output aligns well with both the input image and depth structure, leading to sharper and more geometrically consistent results. + +# 4. Results + +Dataset preparation. We prepare both synthetic and real-world datasets to evaluate our method's performance. In the absence of an RGB-D dataset with extreme motion blur, we generate a synthetic reference dataset by rendering directly from 3D models. We create RGB-D image sequences and record per-frame camera trajectories to evaluate our pose estimation and deblurring accuracy. The sequences, ranging from 50 to 150 frames, are captured along simulated trajectories that include rapid translations, 360-degree rotations, and human movement paths (using Blender's Walk Navigation feature). The RGB and depth images have a resolution of $640 \times 480$ , and motion blur is applied by enabling Blender's motion blur setting. The 3D models are sourced from the McGuire archive [23] and the Blender archive [43]. + +![](images/2cf99cda93bb7ddcff2d8a58587e6a4118a5d9f87ab4ea640829dab6d1959041.jpg) +Figure 6. Qualitative comparison of deblurring performance from a novel view. Our method excels in restoring high-frequency details, whereas several methods, including COLMAP [31], fail under severe motion blur conditions (see Section 4.1). The corresponding quantitative results are provided in Table 2. + +For real scene evaluation, we used an Azure Kinect RGB-D camera with a 33.3ms fixed exposure time, 5000K fixed white balance, NFOV-2 × 2-binned depth capture mode, and $1280 \times 720$ RGB resolution. We used the provided intrinsic camera parameters to undistort the images and aligned depth maps to color images using the SDK's transformation functions. Depth maps were eroded three times to remove outliers, which commonly occur along edges. + +Implementation details. Our code is based on 3D Gaussian Splatting [14], with data processing partially adapted from CF-3DGS [14]. For initial camera pose estimation, we use the Unimatch [48] optical flow network and the ICP [5] algorithm from Open3D as part of the optimization process. For an input image sequence of $N$ frames, + +we prune Gaussians every $100N$ iterations, reset opacity to 0.5 every $200N$ iterations, and run a total of $600N$ iterations (Section 3.2). The number of virtual views for deblurring is set to 15 for synthetic dataset evaluation and 17 for real-scene comparison. All experiments were conducted on an NVIDIA RTX A6000 GPU (48GB), with initial camera pose optimization taking 11 seconds per frame, 0.1 seconds per iteration for pose and geometry refinement, and 0.26 seconds per iteration for deblurring at $640 \times 480$ resolution. + +# 4.1.Quantitative Evaluation + +We evaluate the camera pose accuracy and deblurring performance on three synthetic scenes from our dataset, each rendered with extreme motion blur. For pose accuracy, we compare our method against a variety of approaches capable of generating a 3D point cloud and camera trajectory: + +![](images/63db02ccec4b5616198a2333aa1148c6ba4dfb8a09d5fcfe25fee7f35c63a572.jpg) +Figure 7. Qualitative comparison of deblurring performance from novel views on real scenes captured with an Azure Kinect camera. Our method effectively restores high-frequency details, outperforming other methods even when processing RGB-D sequences with large spatial gaps between frames due to rapid camera motion and severe motion blur. + +COLMAP [31], ICP with FPFH feature-based global alignment [29], ORB-SLAM3 [3], ElasticFusion [47], and several RGB-D SLAM methods, including NICE-SLAM [51], Point-SLAM [30], and DROID-SLAM [42]. Additionally, we include recent Gaussian Splatting-based SLAM approaches such as MonoGS [21] and SplaTAM [12]. For each method, we calculate absolute trajectory error (ATE) and relative pose error (RPE) in meter and degree units, using the ground truth camera trajectories as references. + +To evaluate deblurring accuracy, we modify a state-of-the-art Gaussian Splitting-based image restoration method [49] to operate in a depth-aware manner using undistorted color and depth images. We select one out of every three images as a test view and render the correspond + +ing novel view for comparison with the ground truth. For quantitative metrics, we calculate PSNR, SSIM, and LPIPS on the rendered images, as well as RMSE between the rendered and ground truth depth maps in the inverse depth domain $(m^{-1})$ . + +# 4.2. Ablation Study + +To validate the contributions of each component in our method, we perform an ablation study evaluating camera pose and deblurring accuracy by selectively omitting key elements. Specifically, we conduct experiments without fixing the scale of the Gaussians (Section 3.2), without the pose and geometry refinement process (Equation 5), and without depth loss (Equation 8). + +Table 1. Quantitative comparison of camera pose accuracy against ground truth poses. We report scores only for methods that successfully provide both camera poses and a point cloud (see Section 4.1). Our method achieves the best scores on RPE metrics and shows ATE accuracy comparable to DROID-SLAM [42]. Each color highlights the best and second best results. + +
BedroomLivingroomOffice
ATE ↓RPE (trans) ↓RPE (rot) ↓ATE ↓RPE (trans) ↓RPE (rot) ↓ATE ↓RPE (trans) ↓RPE (rot) ↓
ElasticFusion [47]1.3030.0792.2181.0140.0562.7770.4790.0380.480
FPFH-ICP [29]2.7600.1182.3102.0280.0921.2130.1390.0090.167
DROID-SLAM [42]0.0590.0441.2120.0500.0140.3170.0110.0110.245
MonoGS [21]0.5060.0410.8120.1350.0090.1240.0450.0110.180
Ours0.1020.0060.0920.0050.0020.0320.0410.0030.046
+ +Table 2. Quantitative comparison of deblurring performance from novel views. Our method demonstrates high accuracy across all metrics for both color and depth images. RMSE is calculated by comparing the rendered depth with GT depth in the inverse depth domain $(m^{-1})$ . + +
BedroomLivingroomOffice
PSNR ↑SSIM ↑LPIPS ↓RMSE ↓PSNR ↑SSIM ↑LPIPS ↓RMSE ↓PSNR ↑SSIM ↑LPIPS ↓RMSE ↓
ElasticFusion [47]18.7660.5860.4661.24913.7990.4480.5540.14619.0520.5330.4010.079
FPFH-ICP [29]20.6230.6170.4180.45018.1720.5980.4220.10222.2120.6840.2410.033
DROID-SLAM [42]21.3750.6660.3020.01321.2220.7350.2500.03525.6800.7690.1440.013
MonoGS [21]23.7280.7170.2850.00924.4140.8220.1880.02724.1940.7250.1840.184
Ours26.7450.8240.2060.00527.6500.9000.1500.01025.6490.7910.1600.014
+ +Table 3. Ablation study on the effects of the global refinement process and fixed scale of Gaussians. Our refinement significantly enhances deblurring and depth accuracy while fixing the scale of the Gaussians leads to more accurate camera pose estimation. + +
PSNRSSIMLPIPSRMSEATERPE(trans)RPE(rot)
w/o scale fix21.6560.7030.3080.0270.1550.0531.488
w/o refinement22.2230.7340.2380.0270.1230.0070.047
w/o depth loss25.5180.8070.1700.0140.1030.0120.195
Scale fix + refinement26.6810.8380.1720.0100.1030.0120.195
+ +We assess the effectiveness of these components by performing deblurring based on results from Section 3.1—omitting refinement of camera poses and geometry—and by allowing the scale of the 3D Gaussians to be optimized freely. Table 3 demonstrates that both the global refinement process and fixed Gaussian scale are essential for achieving high accuracy. Notably, during pose refinement, the algorithm reduces absolute pose error even at the expense of a slight increase in relative pose error, underscoring the importance of our refinement approach for stable, high-quality reconstruction. + +Our method achieves superior RPE scores across all scenes, as shown in Table 1, outperforming other approaches in pose and geometry estimation under severe motion blur. While our ATE scores are comparable to those of DROID-SLAM, which uses a bundle adjustment (BA) module, ATE alone does not correlate strongly with deblurring quality. Our method consistently produces more accurate and denser point clouds, which directly contributes to improved deblurring performance, as evidenced by the metrics in Table 2. These results confirm that our integrated approach to pose estimation and deblurring is more resilient to challenging conditions than other methods. + +# 4.3. Qualitative Evaluation + +Figures 6 and 7 present qualitative comparisons of our method's performance in synthetic and real scenes. The + +densification capability of Gaussian Splitting provides detailed scene representation even from a relatively sparse point cloud; however, our approach, with its initially dense and accurate point cloud generation, achieves notably superior deblurring performance, especially in restoring high-frequency details. Our method consistently demonstrates a significant advantage in visual fidelity and detail restoration across diverse conditions. + +# 5. Conclusion + +We have introduced a robust method for 3D scene reconstruction from RGB-D image sequences that effectively addresses the challenges of extreme motion blur and low-light conditions. Our approach leverages optical flow from color images and a carefully designed geometric loss to achieve accurate global alignment, followed by local refinement of camera poses and geometry using ICP. By integrating these steps into a Gaussian Splatting pipeline, we further refine camera poses and 3D geometry by minimizing depth and opacity losses. Additionally, fixing the scale of the 3D Gaussians ensures that the depth information from the RGB-D input is fully utilized, allowing for precise camera pose estimation and improved deblurring performance. Our method demonstrates superior performance on real and synthetic RGB-D scenes with significant motion blur, outperforming existing approaches. + +# Acknowledgements + +Min H. Kim acknowledges the Samsung Research Funding & Incubation Center (SRFC-IT2402-02), the Korea NRF grant (RS-2024-00357548), the MSIT/IITP of Korea (RS2022-00155620, RS-2024-00398830, 2022-0-00058, and 2017-0-00072), Microsoft Research Asia, and Samsung Electronics. + +# References + +[1] Dejan Azinović, Ricardo Martin-Brualla, Dan B Goldman, Matthias Nießner, and Justus Thies. Neural rgb-d surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6290-6301, 2022. 3 +[2] Michael Bloesch, Jan Czarnowski, Ronald Clark, Stefan Leutenegger, and Andrew J Davison. Codeslam—learning a compact, optimisable representation for dense visual slam. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2560–2568, 2018. 2 +[3] Carlos Campos, Richard Elvira, Juan J Gómez Rodríguez, José MM Montiel, and Juan D Tardós. Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics, 37(6):1874-1890, 2021. 2, 7 +[4] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. arXiv preprint arXiv:2204.04676, 2022. 2, 3 +[5] Yang Chen and Gérard Medioni. Object modelling by registration of multiple range images. Image and vision computing, 10(3):145-155, 1992. 6 +[6] Angela Dai, Matthias Nießner, Michael Zollofer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface re-integration. ACM Transactions on Graphics 2017 (TOG), 2017. 2 +[7] Rob Fergus, Barun Singh, Aaron Hertzmann, Sam T Roweis, and William T Freeman. Removing camera shake from a single photograph. In Acm Siggraph 2006 Papers, pages 787-794. 2006. 3 +[8] Kyle Genova, Forrester Cole, Avneesh Sud, Aaron Sarna, and Thomas Funkhouser. Deep structured implicit functions. arXiv preprint arXiv:1912.06126, 2:2, 2019. 2 +[9] Zan Gojcic, Caifa Zhou, Jan D Wegner, and Andreas Wieser. The perfect match: 3d point cloud matching with smoothed densities. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5545-5554, 2019. 3 +[10] Alex Graves and Alex Graves. Long short-term memory. Supervised sequence labelling with recurrent neural networks, pages 37-45, 2012. 3 +[11] Olaf Kähler, Victor Prisacariu, Julien Valentin, and David Murray. Hierarchical voxel block hashing for efficient integration of depth images. IEEE Robotics and Automation Letters, 1(1):192-197, 2015. 2 +[12] Nikhil Keetha, Jay Karhade, Krishna Murthy Jatavallabhula, Gengshan Yang, Sebastian Scherer, Deva Ramanan, and Jonathon Luiten. Splatam: Splat track & map 3d gaussians for dense rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21357-21366, 2024. 3, 7 +[13] Maik Keller, Damien Lefloch, Martin Lambers, Shahram Izadi, Tim Weyrich, and Andreas Kolb. Real-time 3d reconstruction in dynamic scenes using point-based fusion. In 2013 International Conference on 3D Vision-3DV 2013, pages 1-8. IEEE, 2013. 2 + +[14] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 3, 6 +[15] Christian Kerl, Jürgen Sturm, and Daniel Cremers. Dense visual slam for rgb-d cameras. In 2013 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 2100-2106. IEEE, 2013. 2 +[16] Dilip Krishnan and Rob Fergus. Fast image deconvolution using hyper-laplacian priors. Advances in neural information processing systems, 22, 2009. 3 +[17] Dogyoon Lee, Minhyeok Lee, Chajin Shin, and Sangyoun Lee. Dp-nerf: Deblurred neural radiance field with physical scene priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12386–12396, 2023. 3 +[18] Yang Li and Tatsuya Harada. Lepard: Learning partial point cloud matching in rigid and deformable scenes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5554-5564, 2022. 3 +[19] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-nerf: Neural radiance fields from blurry images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12861–12870, 2022. 3 +[20] Nico Marniok, Ole Johannsen, and Bastian Goldluecke. An efficient octree design for local variational range image fusion. In Pattern Recognition: 39th German Conference, GCPR 2017, Basel, Switzerland, September 12-15, 2017, Proceedings 39, pages 401-412. Springer, 2017. 2 +[21] Hidenobu Matsuki, Riku Murai, Paul HJ Kelly, and Andrew J Davison. Gaussian splatting slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18039-18048, 2024. 3, 7, 8 +[22] John McCormac, Ankur Handa, Andrew Davison, and Stefan Leutenegger. Semantic fusion: Dense 3d semantic mapping with convolutional neural networks. In 2017 IEEE International Conference on Robotics and automation (ICRA), pages 4628-4635. IEEE, 2017. 2 +[23] Morgan McGuire. Computer graphics archive, 2017. 5 +[24] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3 +[25] Raul Mur-Artal and Juan D Tardós. Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras. IEEE transactions on robotics, 33(5):1255-1262, 2017. 2 +[26] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3883-3891, 2017. 3 +[27] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. + +Kinectfusion: Real-time dense surface mapping and tracking. In 2011 10th IEEE international symposium on mixed and augmented reality, pages 127-136. IEEE, 2011. 2 +[28] Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Marc Stamminger. Real-time 3d reconstruction at scale using voxel hashing. ACM Transactions on Graphics (ToG), 32 (6):1-11, 2013. 2 +[29] Radu Bogdan Rusu, Nico Blodow, and Michael Beetz. Fast point feature histograms (fpfh) for 3d registration. In 2009 IEEE international conference on robotics and automation, pages 3212-3217. IEEE, 2009. 7, 8 +[30] Erik Sandström, Yue Li, Luc Van Gool, and Martin R Oswald. Point-slam: Dense neural point cloud-based slam. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18433-18444, 2023. 3, 7 +[31] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 6, 7 +[32] Thomas Schops, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 134-144, 2019. 2 +[33] Thomas Schöps, Torsten Sattler, and Marc Pollefeys. Surfelmeshing: Online surfel-based mesh reconstruction. IEEE transactions on pattern analysis and machine intelligence, 42(10):2494-2507, 2019. 2 +[34] Christian J Schuler, Michael Hirsch, Stefan Harmeling, and Bernhard Scholkopf. Learning to deblur. IEEE transactions on pattern analysis and machine intelligence, 38(7):1439-1451, 2015. 3 +[35] Otto Seiskari, Jerry Yilammi, Valtteri Kaatrasalo, Pekka Rantalankila, Matias Turkulainen, Juho Kannala, and Arno Solin. Gaussian splattering on the move: Blur and rolling shutter compensation for natural camera motion, 2024. 2, 3 +[36] Qi Shan, Jiaya Jia, and Aseem Agarwala. High-quality motion deblurring from a single image. Acm transactions on graphics (tog), 27(3):1-10, 2008. 3 +[37] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. Advances in neural information processing systems, 28, 2015. 3 +[38] Jörg Stuckler and Sven Behnke. Multi-resolution surfel maps for efficient dense 3d modeling and tracking. Journal of Visual Communication and Image Representation, 25(1):137-147, 2014. 2 +[39] Shuochen Su, Maurizio Delbracio, Jue Wang, Guillermo Sapiro, Wolfgang Heidrich, and Oliver Wang. Deep video deblurring for hand-held cameras. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1279-1288, 2017. 3 +[40] Jian Sun, Wenfei Cao, Zongben Xu, and Jean Ponce. Learning a convolutional neural network for non-uniform motion blur removal. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 769-777, 2015. 3 + +[41] Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 3 +[42] Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. Advances in neural information processing systems, 2021. 3, 7, 8 +[43] Flavio Della Tommasa. Blender demo files, 2024. 5 +[44] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. BAD-NeRF: Bundle Adjusted Deblur Neural Radiance Fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4170–4179, 2023. 2, 3 +[45] Chen Wenbo and Liu Ligang. Deblur-gs: 3d gaussian splattering from camera motion blurred images. Proc. ACM Comput. Graph. Interact. Tech. (Proceedings of I3D 2024), 7(1), 2024. 2, 3 +[46] Thomas Whelan, Michael Kaess, Hordur Johannsson, Maurice Fallon, John J Leonard, and John McDonald. Real-time large-scale densergb-d slam with volumetric fusion. The International Journal of Robotics Research, 34(4-5):598-626, 2015. 2 +[47] Thomas Whelan, Renato F Salas-Moreno, Ben Glocker, Andrew J Davison, and Stefan Leutenegger. Elasticfusion: Real-time dense slam and light source estimation. The International Journal of Robotics Research, 35(14):1697-1716, 2016. 2, 7, 8 +[48] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, Fisher Yu, Dacheng Tao, and Andreas Geiger. Unifying flow, stereo and depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 4, 6 +[49] Lingzhe Zhao, Peng Wang, and Peidong Liu. BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splatting. 2024. 2, 3, 5, 7 +[50] Huizhong Zhou, Benjamin Ummenhofer, and Thomas Brox. Deeptam: Deep tracking and mapping. In Proceedings of the European conference on computer vision (ECCV), pages 822-838, 2018. 2 +[51] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12786-12796, 2022. 3, 7 \ No newline at end of file diff --git a/2025/Splat-based 3D Scene Reconstruction with Extreme Motion-blur/images.zip b/2025/Splat-based 3D Scene Reconstruction with Extreme Motion-blur/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..afe169d84288bb82bbfef9253e88597c72d2e8ae --- /dev/null +++ b/2025/Splat-based 3D Scene Reconstruction with Extreme Motion-blur/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7731cf13c86149ff9fb1640dc950fcf16fad4e77db2796b4846f187a0a2e794d +size 868564 diff --git a/2025/Splat-based 3D Scene Reconstruction with Extreme Motion-blur/layout.json b/2025/Splat-based 3D Scene Reconstruction with Extreme Motion-blur/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..847394d1c73c4a06157dfc8fb15ef219e32611a6 --- /dev/null +++ b/2025/Splat-based 3D Scene Reconstruction with Extreme Motion-blur/layout.json @@ -0,0 +1,8964 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 106, + 103, + 504, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 103, + 504, + 121 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 504, + 121 + ], + "type": "text", + "content": "Splat-based 3D Scene Reconstruction with Extreme Motion-blur" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 73, + 141, + 169, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 141, + 169, + 159 + ], + "spans": [ + { + "bbox": [ + 73, + 141, + 169, + 159 + ], + "type": "text", + "content": "Hyeonjoong Jang†*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 182, + 143, + 271, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 143, + 271, + 170 + ], + "spans": [ + { + "bbox": [ + 182, + 143, + 271, + 170 + ], + "type": "text", + "content": "Dongyoung Choi† \n† KAIST" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 284, + 143, + 452, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 143, + 452, + 171 + ], + "spans": [ + { + "bbox": [ + 284, + 143, + 452, + 171 + ], + "type": "text", + "content": "Donggun Kim" + }, + { + "bbox": [ + 284, + 143, + 452, + 171 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 284, + 143, + 452, + 171 + ], + "type": "text", + "content": " Woohyun Kang" + }, + { + "bbox": [ + 284, + 143, + 452, + 171 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 284, + 143, + 452, + 171 + ], + "type": "text", + "content": " \n* HYPERGRAM" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 466, + 143, + 535, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 466, + 143, + 535, + 156 + ], + "spans": [ + { + "bbox": [ + 466, + 143, + 535, + 156 + ], + "type": "text", + "content": "Min H. Kim†*" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 147, + 174, + 460, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 174, + 460, + 186 + ], + "spans": [ + { + "bbox": [ + 147, + 174, + 460, + 186 + ], + "type": "text", + "content": "{hjjang; dychoi; dgkim; whkang; minhkim}@vclab.kaist.ac.kr" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 60, + 204, + 156, + 259 + ], + "blocks": [ + { + "bbox": [ + 60, + 204, + 156, + 259 + ], + "lines": [ + { + "bbox": [ + 60, + 204, + 156, + 259 + ], + "spans": [ + { + "bbox": [ + 60, + 204, + 156, + 259 + ], + "type": "image", + "image_path": "8148e79216c51fa76cfd8cd39f2083c8c2e187dfed867ff6e83d7798d0aeb7a1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 60, + 260, + 108, + 308 + ], + "blocks": [ + { + "bbox": [ + 60, + 260, + 108, + 308 + ], + "lines": [ + { + "bbox": [ + 60, + 260, + 108, + 308 + ], + "spans": [ + { + "bbox": [ + 60, + 260, + 108, + 308 + ], + "type": "image", + "image_path": "3364c044bf5b03f3d65916a7cfcaa64f0813a1a389d52f3f42e8489bc2aae960.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 309, + 119, + 319 + ], + "lines": [ + { + "bbox": [ + 97, + 309, + 119, + 319 + ], + "spans": [ + { + "bbox": [ + 97, + 309, + 119, + 319 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 110, + 308, + 119, + 319 + ], + "lines": [ + { + "bbox": [ + 110, + 308, + 119, + 319 + ], + "spans": [ + { + "bbox": [ + 110, + 308, + 119, + 319 + ], + "type": "text", + "content": "out" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 159, + 204, + 255, + 260 + ], + "blocks": [ + { + "bbox": [ + 159, + 204, + 255, + 260 + ], + "lines": [ + { + "bbox": [ + 159, + 204, + 255, + 260 + ], + "spans": [ + { + "bbox": [ + 159, + 204, + 255, + 260 + ], + "type": "image", + "image_path": "2028d66bd913728fe7c1763ca132e1251ce9c1f26dd9dfad91350bbe78d58980.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 159, + 260, + 206, + 308 + ], + "blocks": [ + { + "bbox": [ + 159, + 260, + 206, + 308 + ], + "lines": [ + { + "bbox": [ + 159, + 260, + 206, + 308 + ], + "spans": [ + { + "bbox": [ + 159, + 260, + 206, + 308 + ], + "type": "image", + "image_path": "0cd06fcab1244798ddc1c096785dbe6f73dc615a4a6e5a401b04bf88573d6e35.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 308, + 232, + 318 + ], + "lines": [ + { + "bbox": [ + 182, + 308, + 232, + 318 + ], + "spans": [ + { + "bbox": [ + 182, + 308, + 232, + 318 + ], + "type": "text", + "content": "ElasticFusion" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 320, + 555, + 343 + ], + "lines": [ + { + "bbox": [ + 55, + 320, + 555, + 343 + ], + "spans": [ + { + "bbox": [ + 55, + 320, + 555, + 343 + ], + "type": "text", + "content": "Figure 1. We propose a robust 3D scene reconstruction method from RGB-D input that effectively addresses extreme motion blur. Our approach achieves accurate camera poses and dense point clouds, producing clearer, deblurred scenes compared to existing methods." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 207, + 260, + 255, + 308 + ], + "blocks": [ + { + "bbox": [ + 207, + 260, + 255, + 308 + ], + "lines": [ + { + "bbox": [ + 207, + 260, + 255, + 308 + ], + "spans": [ + { + "bbox": [ + 207, + 260, + 255, + 308 + ], + "type": "image", + "image_path": "2dd837bf8b329554cdfcb60a391ce455e5829ef9b8d9d9ecfe14f723b1e732be.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 257, + 204, + 353, + 260 + ], + "blocks": [ + { + "bbox": [ + 257, + 204, + 353, + 260 + ], + "lines": [ + { + "bbox": [ + 257, + 204, + 353, + 260 + ], + "spans": [ + { + "bbox": [ + 257, + 204, + 353, + 260 + ], + "type": "image", + "image_path": "0736a2bd1733fe9528015726d26bd4aadf52fe4862261af60a90203859754934.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 257, + 260, + 353, + 308 + ], + "blocks": [ + { + "bbox": [ + 257, + 260, + 353, + 308 + ], + "lines": [ + { + "bbox": [ + 257, + 260, + 353, + 308 + ], + "spans": [ + { + "bbox": [ + 257, + 260, + 353, + 308 + ], + "type": "image", + "image_path": "6d12761ac801f1e68e66ce23de016c5453509695c66e3030268357845f31825a.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 308, + 333, + 317 + ], + "lines": [ + { + "bbox": [ + 277, + 308, + 333, + 317 + ], + "spans": [ + { + "bbox": [ + 277, + 308, + 333, + 317 + ], + "type": "text", + "content": "DROID-SLAM" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 355, + 204, + 451, + 260 + ], + "blocks": [ + { + "bbox": [ + 355, + 204, + 451, + 260 + ], + "lines": [ + { + "bbox": [ + 355, + 204, + 451, + 260 + ], + "spans": [ + { + "bbox": [ + 355, + 204, + 451, + 260 + ], + "type": "image", + "image_path": "dab89590977b7d2f0ffab195557d5456ff3b9e05450a21199942c3d1bd7837c5.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 355, + 260, + 419, + 308 + ], + "blocks": [ + { + "bbox": [ + 355, + 260, + 419, + 308 + ], + "lines": [ + { + "bbox": [ + 355, + 260, + 419, + 308 + ], + "spans": [ + { + "bbox": [ + 355, + 260, + 419, + 308 + ], + "type": "image", + "image_path": "98007c7640395b4a7ea0fd334de03eee117587749d63b69d2917e0865e8b0982.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 387, + 308, + 421, + 317 + ], + "lines": [ + { + "bbox": [ + 387, + 308, + 421, + 317 + ], + "spans": [ + { + "bbox": [ + 387, + 308, + 421, + 317 + ], + "type": "text", + "content": "MonoGS" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 453, + 204, + 550, + 260 + ], + "blocks": [ + { + "bbox": [ + 453, + 204, + 550, + 260 + ], + "lines": [ + { + "bbox": [ + 453, + 204, + 550, + 260 + ], + "spans": [ + { + "bbox": [ + 453, + 204, + 550, + 260 + ], + "type": "image", + "image_path": "d227d0808ee008285d54552f1ad14a28d5b7af7f569d4cfff1f5f9b5366ae8fe.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 453, + 260, + 501, + 308 + ], + "blocks": [ + { + "bbox": [ + 453, + 260, + 501, + 308 + ], + "lines": [ + { + "bbox": [ + 453, + 260, + 501, + 308 + ], + "spans": [ + { + "bbox": [ + 453, + 260, + 501, + 308 + ], + "type": "image", + "image_path": "eeb28c3f72a0685f986a2b85bf64539fd8f7ebb4ad6fd0ceec92291c16c1356f.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 492, + 309, + 511, + 319 + ], + "lines": [ + { + "bbox": [ + 492, + 309, + 511, + 319 + ], + "spans": [ + { + "bbox": [ + 492, + 309, + 511, + 319 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 502, + 260, + 550, + 308 + ], + "blocks": [ + { + "bbox": [ + 502, + 260, + 550, + 308 + ], + "lines": [ + { + "bbox": [ + 502, + 260, + 550, + 308 + ], + "spans": [ + { + "bbox": [ + 502, + 260, + 550, + 308 + ], + "type": "image", + "image_path": "9bd8095922966743884d2b6d3357fcf3d227469ec8b33f13417f011982375a8b.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 152, + 350, + 200, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 350, + 200, + 363 + ], + "spans": [ + { + "bbox": [ + 152, + 350, + 200, + 363 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 54, + 378, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 378, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 378, + 297, + 714 + ], + "type": "text", + "content": "We propose a splat-based 3D scene reconstruction method from RGB-D input that effectively handles extreme motion blur, a frequent challenge in low-light environments. Under dim illumination, RGB frames often suffer from severe motion blur due to extended exposure times, causing traditional camera pose estimation methods, such as COLMAP, to fail. This results in inaccurate camera pose and blurry color input, compromising the quality of 3D reconstructions. Although recent 3D reconstruction techniques like Neural Radiance Fields and Gaussian Splatting have demonstrated impressive results, they rely on accurate camera trajectory estimation, which becomes challenging under fast motion or poor lighting conditions. Furthermore, rapid camera movement and the limited field of view of depth sensors reduce point cloud overlap, limiting the effectiveness of pose estimation with the ICP algorithm. To address these issues, we introduce a method that combines camera pose estimation and image deblurring using a Gaussian Splatting framework, leveraging both 3D Gaussian splats and depth inputs for enhanced scene representation. Our method first aligns consecutive RGB-D frames through optical flow and ICP, then refines camera poses and 3D geometry by adjusting Gaussian positions for optimal depth alignment. To handle motion blur, we model camera movement during exposure and deblur images by comparing the input with a series of sharp, rendered frames. Experiments on a new RGB-D dataset with extreme motion blur show that our method outperforms existing approaches," + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 352, + 556, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 352, + 556, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 556, + 425 + ], + "type": "text", + "content": "enabling high-quality reconstructions even in challenging conditions. This approach has broad implications for 3D mapping applications in robotics, autonomous navigation, and augmented reality. Both code and dataset are publicly available on https://github.com/KAIST-VCLAB/gs-extreme-motion-blur." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 314, + 451, + 394, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 451, + 394, + 463 + ], + "spans": [ + { + "bbox": [ + 314, + 451, + 394, + 463 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 472, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 472, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 312, + 472, + 555, + 676 + ], + "type": "text", + "content": "High-quality 3D scene reconstruction is one of the most important and challenging applications in computer vision. The accuracy of 3D reconstruction hinges on the quality of essential components, such as camera poses, RGB images, and depth maps. These elements are interconnected and influence each other; a failure in one component can adversely impact the others. For example, blurry RGB images or noisy depth maps can significantly impair camera pose estimation, which, in turn, degrades overall reconstruction quality. This challenge is particularly pronounced when input data is captured in low-light conditions or during rapid camera movement—common scenarios in everyday, casual capture. Such conditions result in degraded color and depth frames, leading to poorly estimated camera poses and reconstructions with blurry textures and distorted or smoothed geometry. Therefore, achieving high-quality 3D reconstruction requires clear RGB images and accurate depth maps." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "The interdependence of these input components suggests an opportunity for compensating or restoring degraded elements. However, this process presents a chicken-and-egg" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "26425" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "problem, complicating the reconstruction further. For instance, accurate camera pose estimation requires sharp images, while deblurring motion-blurred images depend on reliable camera pose information. Although numerous methods [4, 41] exist for recovering sharp images from motion-blurred ones, they often rely on supervised learning models trained on specific datasets. Consequently, their deblurring performance can degrade significantly when faced with new cameras or motion blur scenarios that differ from the training data." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 194, + 294, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 294, + 326 + ], + "type": "text", + "content": "Recent works [35, 44, 45, 49] have made advances in addressing these challenges by optimizing camera trajectories during exposure while simultaneously learning sharp RGB colors. This enables simultaneous camera pose estimation and image deblurring. However, these methods require an initial camera pose and, when using Gaussian Splatting [14], also need a sparse point cloud as input. This requirement limits their applicability, especially when severe motion blur prevents structure-from-motion (SfM) methods like COLMAP [31] from functioning effectively, making it impossible to start the optimization process." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 327, + 294, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 294, + 460 + ], + "type": "text", + "content": "To overcome these limitations, we introduce a robust approach for 3D scene reconstruction that effectively compensates for severe motion blur and addresses the challenges of camera pose estimation without relying on a precise initial pose. Our method leverages RGB-D inputs and incorporates both optical flow and depth information to align camera poses accurately, even in the presence of challenging motion blur and lighting conditions. We first perform a global alignment between consecutive frames using optical flow and the ICP algorithm, which enables effective local alignment of point clouds generated from the depth maps." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 460, + 294, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 294, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 294, + 604 + ], + "type": "text", + "content": "After this initial alignment, we refine the camera poses and 3D geometry by integrating them into a Gaussian Splating pipeline. This approach allows us to initialize dense 3D Gaussians from depth maps, which we scale to ensure a detailed representation of the scene geometry. Our refinement process iteratively adjusts both camera poses and the positions of 3D Gaussians, achieving global consistency by minimizing a depth alignment loss that compares rendered depth maps with input depth measurements. Through this adjustment, we eliminate loop-closure artifacts and reduce the drift that often accumulates in traditional SLAM-based methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 605, + 294, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 294, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 294, + 715 + ], + "type": "text", + "content": "For scenes with significant motion blur, we further optimize the deblurring process by modeling the camera poses at the start and end of each frame's exposure. This temporal modeling allows us to simulate the effects of motion during the exposure period, which we incorporate into our Gaussian Splatting framework. We minimize the difference between the observed motion-blurred image and the average of a set of sharp images rendered from multiple viewpoints along the exposure trajectory, resulting in more accurate," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "geometrically consistent deblurring. Our deblur loss function combines image alignment, structural similarity, and depth consistency, ensuring that the final reconstructions retain sharpness and fidelity to the original scene structure." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 121, + 553, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 553, + 289 + ], + "type": "text", + "content": "Through extensive evaluation on a newly constructed RGB-D dataset featuring extreme motion blur, we demonstrate that our approach significantly outperforms existing methods in both accuracy and robustness. By addressing the key challenges of pose estimation, depth alignment, and image deblurring, our method provides a versatile and effective solution for high-quality 3D reconstruction under real-world capture conditions, such as low light and fast motion. The proposed method holds promising applications in areas requiring reliable 3D mapping and reconstruction, including robotics, autonomous navigation, and augmented reality. We will make both our dataset and implementation publicly available to foster further research and development in this field." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 304, + 400, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 304, + 400, + 316 + ], + "spans": [ + { + "bbox": [ + 314, + 304, + 400, + 316 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 327, + 553, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 553, + 472 + ], + "type": "text", + "content": "RGB-D scene reconstruction. A popular approach to 3D geometry reconstruction defines a 3D voxel grid and constructs a signed distance function (SDF) volume by accumulating depth estimates captured by sensors. Kinect-Fusion [27] uses a Kinect camera to capture RGB-D images, then estimates relative camera poses between adjacent frames with the ICP algorithm. Next, it constructs and updates a truncated signed distance function (TSDF) volume, enabling real-time 3D reconstruction of objects. Niessner et al. [28] improve memory efficiency by combining hashing algorithms to store and update SDF values only where needed, thus reducing memory requirements." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 473, + 553, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 553, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 553, + 639 + ], + "type": "text", + "content": "Several methods [8, 11, 20, 22] adopt memory-efficient, hierarchical data structures for depth fusion. BundleFusion [6] simultaneously refines camera poses and geometry using bundle adjustment, improving reconstruction accuracy and addressing the loop closure problem caused by accumulated camera pose and depth estimation errors. ElasticFusion [47] builds a map with surfels—point primitives containing position, normal, and radius—and updates the map through a deformation graph to minimize error and perform loop closure. Other methods [13, 33, 38] directly fuse depth estimates into point clouds, optimizing memory use. However, these methods generally require slow camera motion to capture sharp images and accurately estimate camera transformations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 641, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 553, + 714 + ], + "type": "text", + "content": "Simultaneous localization and mapping (SLAM) methods [3, 15, 25, 32, 46] track camera motion and construct a map in real-time by extracting visual features from input images to establish 3D point correspondences. CodeSLAM [2] and DeepTAM [50] use neural networks to combine depth maps with color images, while Azinović" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26426" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 94, + 52, + 516, + 175 + ], + "blocks": [ + { + "bbox": [ + 94, + 52, + 516, + 175 + ], + "lines": [ + { + "bbox": [ + 94, + 52, + 516, + 175 + ], + "spans": [ + { + "bbox": [ + 94, + 52, + 516, + 175 + ], + "type": "image", + "image_path": "c45ed40390f47755f0dc6308204674dc878c1ccd30a154ffa616bf7d1259fb64.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 179, + 555, + 236 + ], + "lines": [ + { + "bbox": [ + 54, + 179, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 54, + 179, + 555, + 236 + ], + "type": "text", + "content": "Figure 2. Overview of our method pipeline. We begin with global alignment by comparing optical flow and reprojected pixel differences between adjacent RGB-D image pairs, followed by local estimation of relative camera poses using ICP. Each depth map and camera pose is then used to initialize 3D Gaussians. By fixing the scale of the Gaussians and optimizing with depth and opacity losses, we achieve global refinement of camera poses and geometry. Finally, we deblur the scene texture by minimizing the deblur loss between the input image and averaged virtual images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 247, + 297, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 247, + 297, + 368 + ], + "spans": [ + { + "bbox": [ + 54, + 247, + 297, + 368 + ], + "type": "text", + "content": "et al. [1] train multi-layer perceptrons (MLP) to learn SDF and color values for scene reconstruction. Their approach also optimizes camera transformation correction variables during training to refine poses. However, these methods struggle with input images that contain motion blur. Nice-SLAM [51], Point-SLAM [30], and DROID-SLAM [42] use neural networks to track camera poses. SplaTAM [12] and MonoGS [21] introduced SLAM solutions based on 3D Gaussian Splatting. However, these methods are not designed to handle motion blur." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 374, + 295, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 374, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 374, + 295, + 577 + ], + "type": "text", + "content": "Image deblurring. Motion blur occurs when camera motion during the exposure time changes the projected pixel coordinates of rays on the camera sensor. Early studies [7, 16, 36] tackle this issue by developing kernels to restore blurry images. Convolutional neural networks (CNNs) demonstrate powerful performance in learning deblur kernels [34, 40], and deep learning techniques such as ResNet with skip connections and multi-scale networks further improved restoration quality [26, 39]. SRN-DeblurNet [41] adopt recurrent networks [10, 37] and combine them with an encoder-decoder network structure, while NAFNet [4] simplify the network structure, extracting only essential components and proposing a nonlinear activation-free network. However, since these methods rely on curated datasets like the GoPro dataset [26], they struggle to effectively deblur images with extreme motion blur or those captured on different cameras." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "There are NeRF-based approaches [24] that tackle motion blur in input images. Deblur-NeRF [19] and DP-NeRF [17] use a set of motion-blurred images along with camera poses estimated by COLMAP as input to restore sharp images through NeRF optimization. BAD-NeRF [44] enhances this by jointly optimizing virtual camera poses and radiance fields during exposure. The latest approaches utilize Gaussian Splatting [14], which offers faster training times than NeRF by rendering a set of 3D Gaussians with a dedicated rasterizer rather than optimizing a neural network. Gaussian Splatting-based deblurring meth" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 247, + 553, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 553, + 283 + ], + "type": "text", + "content": "ods [35, 45, 49] optimize the virtual camera trajectory and minimize differences between input blurry images and the average of rendered images at each virtual camera position." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 284, + 554, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 284, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 554, + 415 + ], + "type": "text", + "content": "All of these methods rely on initial camera poses, typically obtained from COLMAP, and for Gaussian Splatting-based methods, a sparse point cloud as well. However, structure-from-motion methods like COLMAP often fail with blurry inputs—such as images captured by fast-moving cameras or in low-light conditions that require long exposure times. Even with dense depth inputs, ICP requires acceptable global alignment, and recent point cloud matching methods [9, 18] cannot accurately determine relative camera transformations between two point clouds if they consist mainly of planar structures." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 416, + 554, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 416, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 554, + 475 + ], + "type": "text", + "content": "To address these issues, we propose an effective camera pose estimation algorithm tailored for RGB-D image sequences that jointly refines both the camera trajectory and the reconstructed 3D point cloud, resulting in improved sharpness in 3D reconstruction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 487, + 370, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 370, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 370, + 499 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "text", + "content": "Overview. In datasets with significant motion blur, using COLMAP to estimate camera poses becomes impractical. Additionally, rapid camera motion and the limited field of view of depth sensors further complicate this task, making ICP registration less effective. To address these challenges, we first perform global alignment using both RGB and depth images, which establishes sufficient initial alignment for the ICP algorithm to operate effectively. After achieving this global alignment, we apply ICP again to refine and accurately estimate the camera poses. However, estimating camera poses by analyzing only two consecutive frames in a sequence can result in accumulated errors, leading to drift over time. To resolve this issue, it is crucial to adjust the 3D geometry and align the camera poses across the entire sequence so that depth maps and camera views from all perspectives remain well-aligned. We accomplish this through bundle adjustment, optimizing the positions of" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26427" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 62, + 265, + 164 + ], + "blocks": [ + { + "bbox": [ + 86, + 62, + 265, + 164 + ], + "lines": [ + { + "bbox": [ + 86, + 62, + 265, + 164 + ], + "spans": [ + { + "bbox": [ + 86, + 62, + 265, + 164 + ], + "type": "image", + "image_path": "528a4657a2fdc0e9e33f5463e6b9ae45155ec1baac2556b41a81a71da06a5775.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 166, + 297, + 201 + ], + "lines": [ + { + "bbox": [ + 55, + 166, + 297, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 166, + 297, + 201 + ], + "type": "text", + "content": "Figure 3. The illustration of our pose initialization loss. When " + }, + { + "bbox": [ + 55, + 166, + 297, + 201 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 55, + 166, + 297, + 201 + ], + "type": "text", + "content": " is reprojected to " + }, + { + "bbox": [ + 55, + 166, + 297, + 201 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_j" + }, + { + "bbox": [ + 55, + 166, + 297, + 201 + ], + "type": "text", + "content": " using the depth value, it is considered geometrically consistent if " + }, + { + "bbox": [ + 55, + 166, + 297, + 201 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{P}_i\\mathbf{H}_i}" + }, + { + "bbox": [ + 55, + 166, + 297, + 201 + ], + "type": "text", + "content": " equals " + }, + { + "bbox": [ + 55, + 166, + 297, + 201 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{P}_j\\mathbf{H}_j}" + }, + { + "bbox": [ + 55, + 166, + 297, + 201 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 212, + 296, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 212, + 296, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 296, + 273 + ], + "type": "text", + "content": "3D Gaussians and camera poses within the Gaussian Splating pipeline. Finally, to account for motion blur, we approximate the camera trajectory over time and optimize the movement during exposure, enabling effective deblurring. See Figure 2 for an overview." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 283, + 183, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 183, + 295 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 183, + 295 + ], + "type": "text", + "content": "3.1. Local Pose Estimation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 302, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 302, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 302, + 296, + 434 + ], + "type": "text", + "content": "To estimate relative camera transformations, we select consecutive frames from the input RGB-D image sequence. Our approach begins with global alignment using optical flow, followed by local alignment of the two-point clouds using the ICP algorithm. For each consecutive RGB image and depth map pair at times " + }, + { + "bbox": [ + 55, + 302, + 296, + 434 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 302, + 296, + 434 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 302, + 296, + 434 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 55, + 302, + 296, + 434 + ], + "type": "text", + "content": ", we optimize the transformation " + }, + { + "bbox": [ + 55, + 302, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\xi \\in \\mathfrak{se}(3)" + }, + { + "bbox": [ + 55, + 302, + 296, + 434 + ], + "type": "text", + "content": " between the frames. Here, " + }, + { + "bbox": [ + 55, + 302, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\xi_{t\\rightarrow t + 1} = [\\mathbf{x}^{\\intercal},\\boldsymbol{\\omega}^{\\intercal}]^{\\intercal}" + }, + { + "bbox": [ + 55, + 302, + 296, + 434 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 302, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 302, + 296, + 434 + ], + "type": "text", + "content": " represents the translation vector between the two camera origins, and " + }, + { + "bbox": [ + 55, + 302, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\omega}" + }, + { + "bbox": [ + 55, + 302, + 296, + 434 + ], + "type": "text", + "content": " is the rotation element of " + }, + { + "bbox": [ + 55, + 302, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\mathfrak{se}(3)" + }, + { + "bbox": [ + 55, + 302, + 296, + 434 + ], + "type": "text", + "content": ". We ultimately calculate a camera-to-world transformation matrix " + }, + { + "bbox": [ + 55, + 302, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\xi_t" + }, + { + "bbox": [ + 55, + 302, + 296, + 434 + ], + "type": "text", + "content": " for each timestamp " + }, + { + "bbox": [ + 55, + 302, + 296, + 434 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 302, + 296, + 434 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 435, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 296, + 568 + ], + "type": "text", + "content": "To achieve this, we first compute the difference between the estimated optical flow and the reprojected pixel from the camera at time " + }, + { + "bbox": [ + 55, + 435, + 296, + 568 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 435, + 296, + 568 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 435, + 296, + 568 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 55, + 435, + 296, + 568 + ], + "type": "text", + "content": " using the depth estimate for each pixel in the depth map " + }, + { + "bbox": [ + 55, + 435, + 296, + 568 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 55, + 435, + 296, + 568 + ], + "type": "text", + "content": " and the intrinsic matrix " + }, + { + "bbox": [ + 55, + 435, + 296, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 55, + 435, + 296, + 568 + ], + "type": "text", + "content": ". A pixel " + }, + { + "bbox": [ + 55, + 435, + 296, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 55, + 435, + 296, + 568 + ], + "type": "text", + "content": " in the depth map is backprojected to 3D space as: " + }, + { + "bbox": [ + 55, + 435, + 296, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i = \\pi^{-1}(\\mathbf{K},\\mathbf{p}_i,D_t(\\mathbf{p}_i))" + }, + { + "bbox": [ + 55, + 435, + 296, + 568 + ], + "type": "text", + "content": ". We then project this 3D point to the camera at " + }, + { + "bbox": [ + 55, + 435, + 296, + 568 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 55, + 435, + 296, + 568 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 55, + 435, + 296, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{p_j} = \\pi (\\mathbf{K},\\exp (\\xi_{t\\to t + 1}^{\\wedge})\\mathbf{P}_i)" + }, + { + "bbox": [ + 55, + 435, + 296, + 568 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 435, + 296, + 568 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 55, + 435, + 296, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 435, + 296, + 568 + ], + "type": "inline_equation", + "content": "\\pi^{-1}" + }, + { + "bbox": [ + 55, + 435, + 296, + 568 + ], + "type": "text", + "content": " represent the projection and backprojection operators, respectively. The pixel movement from " + }, + { + "bbox": [ + 55, + 435, + 296, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 55, + 435, + 296, + 568 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 435, + 296, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_j" + }, + { + "bbox": [ + 55, + 435, + 296, + 568 + ], + "type": "text", + "content": " is then compared with the estimated optical flow " + }, + { + "bbox": [ + 55, + 435, + 296, + 568 + ], + "type": "inline_equation", + "content": "F_{t\\rightarrow t + 1}(\\mathbf{p}_i)" + }, + { + "bbox": [ + 55, + 435, + 296, + 568 + ], + "type": "text", + "content": " to compute the optical flow loss:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 578, + 294, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 294, + 604 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 294, + 604 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {F} = \\sum_ {i} M _ {t} (\\mathbf {p} _ {i}) M _ {t + 1} (\\mathbf {p} _ {j}) \\| F _ {t \\to t + 1} (\\mathbf {p} _ {i}) - (\\mathbf {p} _ {j} - \\mathbf {p} _ {i}) \\| _ {2}.", + "image_path": "6a68ac42baccfe9f2c88fd51886f339c6243a3fd59667e8b666f8ba2d9a0930d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 616, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 296, + 665 + ], + "type": "text", + "content": "In this formulation, " + }, + { + "bbox": [ + 55, + 616, + 296, + 665 + ], + "type": "inline_equation", + "content": "M_{t}" + }, + { + "bbox": [ + 55, + 616, + 296, + 665 + ], + "type": "text", + "content": " is a mask indicating valid depth values in " + }, + { + "bbox": [ + 55, + 616, + 296, + 665 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 55, + 616, + 296, + 665 + ], + "type": "text", + "content": " at pixel " + }, + { + "bbox": [ + 55, + 616, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{i}" + }, + { + "bbox": [ + 55, + 616, + 296, + 665 + ], + "type": "text", + "content": ". The mask " + }, + { + "bbox": [ + 55, + 616, + 296, + 665 + ], + "type": "inline_equation", + "content": "M_{t}" + }, + { + "bbox": [ + 55, + 616, + 296, + 665 + ], + "type": "text", + "content": " dynamically updates during optimization, as " + }, + { + "bbox": [ + 55, + 616, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{j}" + }, + { + "bbox": [ + 55, + 616, + 296, + 665 + ], + "type": "text", + "content": " depends on the transformation " + }, + { + "bbox": [ + 55, + 616, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 55, + 616, + 296, + 665 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 55, + 616, + 296, + 665 + ], + "type": "inline_equation", + "content": "M_{t}(\\mathbf{p}_{i}) = D_{t}(\\mathbf{p}_{i}) > 0" + }, + { + "bbox": [ + 55, + 616, + 296, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "We employ a state-of-the-art pretrained optical flow estimation model [48] designed hierarchically to produce reliable optical flow even from two blurry input images. Next, we calculate the photometric loss between color values at" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 512, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 512, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 512, + 85 + ], + "type": "text", + "content": "pixel " + }, + { + "bbox": [ + 313, + 72, + 512, + 85 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 313, + 72, + 512, + 85 + ], + "type": "text", + "content": " in image " + }, + { + "bbox": [ + 313, + 72, + 512, + 85 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 313, + 72, + 512, + 85 + ], + "type": "text", + "content": " and pixel " + }, + { + "bbox": [ + 313, + 72, + 512, + 85 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_j" + }, + { + "bbox": [ + 313, + 72, + 512, + 85 + ], + "type": "text", + "content": " in image " + }, + { + "bbox": [ + 313, + 72, + 512, + 85 + ], + "type": "inline_equation", + "content": "I_{t + 1}" + }, + { + "bbox": [ + 313, + 72, + 512, + 85 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 330, + 94, + 538, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 94, + 538, + 119 + ], + "spans": [ + { + "bbox": [ + 330, + 94, + 538, + 119 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {I} = \\sum_ {i} M _ {t} (\\mathbf {p} _ {i}) M _ {t + 1} (\\mathbf {p} _ {j}) \\| I _ {t} (\\mathbf {p} _ {i}) - I _ {t + 1} (\\mathbf {p} _ {j}) \\| _ {1}.", + "image_path": "85c63c954977dc9628c0343a4ce5192c3d5476ca29b687448cb77257eeebab58.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 126, + 555, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 126, + 555, + 210 + ], + "spans": [ + { + "bbox": [ + 313, + 126, + 555, + 210 + ], + "type": "text", + "content": "To ensure geometric consistency, we introduce a depth consistency loss by using the two depth maps captured by the depth camera to optimize relative poses. Specifically, as illustrated in Figure 3, we compute and compare the vertical components " + }, + { + "bbox": [ + 313, + 126, + 555, + 210 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{P}_i\\mathbf{H}_i}" + }, + { + "bbox": [ + 313, + 126, + 555, + 210 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 126, + 555, + 210 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{P}_j\\mathbf{H}_j}" + }, + { + "bbox": [ + 313, + 126, + 555, + 210 + ], + "type": "text", + "content": " based on the baseline vector " + }, + { + "bbox": [ + 313, + 126, + 555, + 210 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{O}_t\\mathbf{O}_{t + 1}}" + }, + { + "bbox": [ + 313, + 126, + 555, + 210 + ], + "type": "text", + "content": " between the two cameras. We calculate " + }, + { + "bbox": [ + 313, + 126, + 555, + 210 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{P}_i\\mathbf{H}_i}" + }, + { + "bbox": [ + 313, + 126, + 555, + 210 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 339, + 216, + 529, + 245 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 216, + 529, + 245 + ], + "spans": [ + { + "bbox": [ + 339, + 216, + 529, + 245 + ], + "type": "interline_equation", + "content": "\\overline {{\\mathbf {P} _ {i} \\mathbf {H} _ {i}}} = \\| \\mathbf {P} _ {i} \\| _ {2} \\sin \\left(\\cos^ {- 1} \\left(\\frac {\\mathbf {P} _ {i} \\cdot \\mathbf {x}}{\\| \\mathbf {P} _ {i} \\| _ {2} \\| \\mathbf {x} \\| _ {2}}\\right)\\right).", + "image_path": "a3c5251d12a087a131655410b82f9e8c76b252fdfe4357fb2ec9599f9b5a1396.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 252, + 450, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 252, + 450, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 252, + 450, + 266 + ], + "type": "text", + "content": "Similarly, we compute " + }, + { + "bbox": [ + 313, + 252, + 450, + 266 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{P}_j\\mathbf{H}_j}" + }, + { + "bbox": [ + 313, + 252, + 450, + 266 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 274, + 548, + 299 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 548, + 299 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 548, + 299 + ], + "type": "interline_equation", + "content": "\\overline {{\\mathbf {P} _ {j} \\mathbf {H} _ {j}}} = \\| \\mathbf {P} _ {j} \\| _ {2} \\sin \\left(\\cos^ {- 1} \\left(\\frac {\\mathbf {P} _ {j} \\cdot (- \\exp (\\boldsymbol {\\omega}) ^ {\\intercal} \\mathbf {x})}{\\| \\mathbf {P} _ {j} \\| _ {2} \\| - \\exp (\\boldsymbol {\\omega}) ^ {\\intercal} \\mathbf {x} \\| _ {2}}\\right)\\right).", + "image_path": "e578fb540cc5100099b21b4e3385b0872e331689782d8ee28b46a16105a42965.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 304, + 526, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 526, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 526, + 316 + ], + "type": "text", + "content": "This depth consistency loss is calculated as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 337, + 324, + 555, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 324, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 337, + 324, + 555, + 350 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C} = \\sum_ {i} M _ {t} (\\mathbf {p} _ {i}) M _ {t + 1} (\\mathbf {p} _ {j}) \\| \\overline {{\\mathbf {P} _ {i} \\mathbf {H} _ {i}}} - \\overline {{\\mathbf {P} _ {j} \\mathbf {H} _ {j}}} \\|, \\tag {1}", + "image_path": "b2805e05edcdb301398021bcff3228ce9b60b90a25b3364c73a21a56ebcbedc9.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 358, + 554, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 554, + 394 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 554, + 394 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 358, + 554, + 394 + ], + "type": "inline_equation", + "content": "\\exp()" + }, + { + "bbox": [ + 313, + 358, + 554, + 394 + ], + "type": "text", + "content": " is the exponential mapping operator in Lie algebra, which helps maintain alignment. Our final loss function combines these terms with weighted coefficients:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 371, + 403, + 555, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 403, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 371, + 403, + 555, + 418 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p o s e}} = \\mathcal {L} _ {F} + \\lambda_ {I} \\mathcal {L} _ {I} + \\lambda_ {C} \\mathcal {L} _ {C}. \\tag {2}", + "image_path": "eee8b9acbf0bac96db45624f8497b09df024e762205863077198c8c1739c2cc8.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 426, + 554, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 554, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 554, + 485 + ], + "type": "text", + "content": "We start optimization with " + }, + { + "bbox": [ + 313, + 426, + 554, + 485 + ], + "type": "inline_equation", + "content": "\\lambda_I = 0" + }, + { + "bbox": [ + 313, + 426, + 554, + 485 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 426, + 554, + 485 + ], + "type": "inline_equation", + "content": "\\lambda_C = 0" + }, + { + "bbox": [ + 313, + 426, + 554, + 485 + ], + "type": "text", + "content": ", allowing only the flow loss " + }, + { + "bbox": [ + 313, + 426, + 554, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_F" + }, + { + "bbox": [ + 313, + 426, + 554, + 485 + ], + "type": "text", + "content": " to guide the relative pose alignment with global optical flow. After initial convergence, we increase weights to " + }, + { + "bbox": [ + 313, + 426, + 554, + 485 + ], + "type": "inline_equation", + "content": "\\lambda_I = 75" + }, + { + "bbox": [ + 313, + 426, + 554, + 485 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 426, + 554, + 485 + ], + "type": "inline_equation", + "content": "\\lambda_C = 25" + }, + { + "bbox": [ + 313, + 426, + 554, + 485 + ], + "type": "text", + "content": " and continue optimization to a second convergence." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 486, + 554, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 554, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 554, + 546 + ], + "type": "text", + "content": "Finally, we apply a point-to-plane ICP algorithm on the two globally aligned point clouds to estimate camera poses accurately. Since optical flow estimation between two blurry images introduces some errors, this final camera pose estimation relies exclusively on depth maps." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 552, + 553, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 553, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 553, + 565 + ], + "type": "text", + "content": "3.2. Global Pose Estimation/Geometry Refinement" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": "Starting from the initial camera poses and point clouds obtained in the previous step, we merge them into a global point cloud within the world coordinate system (Figure 4). To manage memory efficiently, we downsample the point cloud with a specified voxel size " + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": ". However, due to accumulated errors in the initial camera poses, these poses may not align well with the global point cloud, leading to seam artifacts at image boundaries and a drifting phenomenon that creates loop-closure issues. To resolve this, we refine both the camera poses and 3D point cloud positions simultaneously within the Gaussian Splatting pipeline, ensuring global and geometric consistency." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26428" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 62, + 176, + 130 + ], + "blocks": [ + { + "bbox": [ + 58, + 62, + 176, + 130 + ], + "lines": [ + { + "bbox": [ + 58, + 62, + 176, + 130 + ], + "spans": [ + { + "bbox": [ + 58, + 62, + 176, + 130 + ], + "type": "image", + "image_path": "240642e034bbd39f7e39698d52f74ef4049cb02e47960e7012347066845330f2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 93, + 130, + 149, + 139 + ], + "lines": [ + { + "bbox": [ + 93, + 130, + 149, + 139 + ], + "spans": [ + { + "bbox": [ + 93, + 130, + 149, + 139 + ], + "type": "text", + "content": "Input color image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 176, + 62, + 293, + 130 + ], + "blocks": [ + { + "bbox": [ + 176, + 62, + 293, + 130 + ], + "lines": [ + { + "bbox": [ + 176, + 62, + 293, + 130 + ], + "spans": [ + { + "bbox": [ + 176, + 62, + 293, + 130 + ], + "type": "image", + "image_path": "eaa9e2f5231815953c24155779cac99f8f10632e06766e37a17deffd12748646.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 204, + 130, + 256, + 140 + ], + "lines": [ + { + "bbox": [ + 204, + 130, + 256, + 140 + ], + "spans": [ + { + "bbox": [ + 204, + 130, + 256, + 140 + ], + "type": "text", + "content": "Input depth map" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 60, + 140, + 134, + 187 + ], + "blocks": [ + { + "bbox": [ + 60, + 140, + 134, + 187 + ], + "lines": [ + { + "bbox": [ + 60, + 140, + 134, + 187 + ], + "spans": [ + { + "bbox": [ + 60, + 140, + 134, + 187 + ], + "type": "image", + "image_path": "ff9a09ca1da2b85b480d6029603136c185d8538391ba56c461d748063bea0af6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 190, + 122, + 199 + ], + "lines": [ + { + "bbox": [ + 72, + 190, + 122, + 199 + ], + "spans": [ + { + "bbox": [ + 72, + 190, + 122, + 199 + ], + "type": "text", + "content": "DROID-SLAM" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 136, + 140, + 213, + 190 + ], + "blocks": [ + { + "bbox": [ + 136, + 140, + 213, + 190 + ], + "lines": [ + { + "bbox": [ + 136, + 140, + 213, + 190 + ], + "spans": [ + { + "bbox": [ + 136, + 140, + 213, + 190 + ], + "type": "image", + "image_path": "7cb32aaf4b47cb9da9c1a0a64fae1c50678584e4607a84f0760ee431aec63553.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 191, + 187, + 199 + ], + "lines": [ + { + "bbox": [ + 157, + 191, + 187, + 199 + ], + "spans": [ + { + "bbox": [ + 157, + 191, + 187, + 199 + ], + "type": "text", + "content": "MonoGS" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 203, + 295, + 237 + ], + "lines": [ + { + "bbox": [ + 55, + 203, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 203, + 295, + 237 + ], + "type": "text", + "content": "Figure 4. Reconstructed initial point clouds from the RGB-D image sequence. Our method achieves accurate camera poses and generates a dense, high-quality point cloud." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 213, + 141, + 293, + 189 + ], + "blocks": [ + { + "bbox": [ + 213, + 141, + 293, + 189 + ], + "lines": [ + { + "bbox": [ + 213, + 141, + 293, + 189 + ], + "spans": [ + { + "bbox": [ + 213, + 141, + 293, + 189 + ], + "type": "image", + "image_path": "8bd656b686752171a2fd64d2b71b711c74e1be9d51d7b89072b949957a7913ac.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 243, + 191, + 260, + 199 + ], + "lines": [ + { + "bbox": [ + 243, + 191, + 260, + 199 + ], + "spans": [ + { + "bbox": [ + 243, + 191, + 260, + 199 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 246, + 296, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 246, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 246, + 296, + 353 + ], + "type": "text", + "content": "We initialize the Gaussian Splatting pipeline by loading the point cloud along with the estimated camera poses. To capture every depth value per pixel from the dense depth maps, we initialize 3D Gaussians, fixing their sizes and setting their scale to " + }, + { + "bbox": [ + 55, + 246, + 296, + 353 + ], + "type": "inline_equation", + "content": "s / 2" + }, + { + "bbox": [ + 55, + 246, + 296, + 353 + ], + "type": "text", + "content": ". This dense arrangement of 3D Gaussians provides a detailed geometric representation. We then train only the camera poses, 3D positions, and opacities of the Gaussians by comparing a rendered depth map " + }, + { + "bbox": [ + 55, + 246, + 296, + 353 + ], + "type": "inline_equation", + "content": "\\widetilde{D}" + }, + { + "bbox": [ + 55, + 246, + 296, + 353 + ], + "type": "text", + "content": " with the input depth map " + }, + { + "bbox": [ + 55, + 246, + 296, + 353 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 246, + 296, + 353 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 354, + 296, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 401 + ], + "type": "text", + "content": "During training, we randomly select camera viewpoints from among the trainable camera pose variables to render each depth map. For a depth value " + }, + { + "bbox": [ + 55, + 354, + 296, + 401 + ], + "type": "inline_equation", + "content": "\\widetilde{D}_t(\\mathbf{p}_i)" + }, + { + "bbox": [ + 55, + 354, + 296, + 401 + ], + "type": "text", + "content": " at pixel " + }, + { + "bbox": [ + 55, + 354, + 296, + 401 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 55, + 354, + 296, + 401 + ], + "type": "text", + "content": " in the rendered map, we calculate the depth loss as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 92, + 405, + 295, + 431 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 405, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 92, + 405, + 295, + 431 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D} = \\sum_ {i} M _ {t} (\\mathbf {p} _ {i}) \\left\\| D _ {t} (\\mathbf {p} _ {i}) - \\widetilde {D} _ {t} (\\mathbf {p} _ {i}) \\right\\| _ {1}. \\tag {3}", + "image_path": "2c0a5f8679e4c6f1b0bf5e094fd362b00454f830971f2792b401080156ed8cbe.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 436, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 295, + 460 + ], + "type": "text", + "content": "We also initialize each 3D Gaussian's opacity " + }, + { + "bbox": [ + 55, + 436, + 295, + 460 + ], + "type": "inline_equation", + "content": "o_j" + }, + { + "bbox": [ + 55, + 436, + 295, + 460 + ], + "type": "text", + "content": " to 0.5, optimizing them to values of 0 or 1:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 119, + 464, + 295, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 464, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 119, + 464, + 295, + 491 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {o p a c i t y}} = \\sum_ {j} o _ {j} ^ {2} \\left(1 - o _ {j}\\right) ^ {2}. \\tag {4}", + "image_path": "cf988779de9de42a13675502f27b227f708b09a639e35f02c0390bc5e05a5760.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 495, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 295, + 590 + ], + "type": "text", + "content": "This optimization aims to retain only Gaussians with an opacity of 1, pruning others to reduce ambiguity for the subsequent deblurring stage. We prune Gaussians with an opacity less than 0.8 and reset opacity with 0.5 again every certain amount of iteration. As iterations progress, the rendered depth maps increasingly align with the input depth maps, indicating that all 3D Gaussian positions and camera poses are globally well-aligned (Figure 5)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 590, + 296, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 296, + 614 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 296, + 614 + ], + "type": "text", + "content": "The final refinement loss is a weighted combination of the depth and opacity losses:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 107, + 620, + 295, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 620, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 107, + 620, + 295, + 634 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e f i n e m e n t}} = \\lambda_ {D} \\mathcal {L} _ {D} + \\lambda_ {o} \\mathcal {L} _ {\\text {o p a c i t y}}. \\tag {5}", + "image_path": "3c4967fd5b683deec67b00b3d59a392cef05ae8315c6216b7137b75a639154fa.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 637, + 144, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 637, + 144, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 637, + 144, + 649 + ], + "type": "text", + "content": "3.3. Image Deblur" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 654, + 295, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 678 + ], + "type": "text", + "content": "Motion blur caused by camera movement can generally be approximated as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 118, + 684, + 295, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 684, + 295, + 716 + ], + "spans": [ + { + "bbox": [ + 118, + 684, + 295, + 716 + ], + "type": "interline_equation", + "content": "I (t) \\approx B (t) = \\frac {1}{k} \\sum_ {i = 0} ^ {k} C _ {i} (t) \\tag {6}", + "image_path": "6d1b18caf3bcfa88f1fbefeced740a3d2535e4b138e1aab168d0146fc0c696e5.jpg" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 317, + 63, + 395, + 121 + ], + "blocks": [ + { + "bbox": [ + 317, + 63, + 395, + 121 + ], + "lines": [ + { + "bbox": [ + 317, + 63, + 395, + 121 + ], + "spans": [ + { + "bbox": [ + 317, + 63, + 395, + 121 + ], + "type": "image", + "image_path": "164be333210da6b4a6d4068eee8e3c6b1251d2805000ba79443ba1baef62c0db.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 317, + 121, + 394, + 159 + ], + "blocks": [ + { + "bbox": [ + 317, + 121, + 394, + 159 + ], + "lines": [ + { + "bbox": [ + 317, + 121, + 394, + 159 + ], + "spans": [ + { + "bbox": [ + 317, + 121, + 394, + 159 + ], + "type": "image", + "image_path": "dd21ab2a4f4b577f34e8fa12b5459d4dbefcefca81e5a90d71de5492a2a8cad8.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 160, + 375, + 168 + ], + "lines": [ + { + "bbox": [ + 335, + 160, + 375, + 168 + ], + "spans": [ + { + "bbox": [ + 335, + 160, + 375, + 168 + ], + "type": "text", + "content": "Iteration 100" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 172, + 553, + 204 + ], + "lines": [ + { + "bbox": [ + 313, + 172, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 172, + 553, + 204 + ], + "type": "text", + "content": "Figure 5. The refinement process corrects misaligned 3D Gaussians by updating camera poses and geometry, with effectiveness demonstrated in Table 3." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 395, + 63, + 473, + 121 + ], + "blocks": [ + { + "bbox": [ + 395, + 63, + 473, + 121 + ], + "lines": [ + { + "bbox": [ + 395, + 63, + 473, + 121 + ], + "spans": [ + { + "bbox": [ + 395, + 63, + 473, + 121 + ], + "type": "image", + "image_path": "b7a7242cd27897475e6ae65ffea3624d168713cc42fc97f33558d7801b02828e.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 395, + 121, + 473, + 159 + ], + "blocks": [ + { + "bbox": [ + 395, + 121, + 473, + 159 + ], + "lines": [ + { + "bbox": [ + 395, + 121, + 473, + 159 + ], + "spans": [ + { + "bbox": [ + 395, + 121, + 473, + 159 + ], + "type": "image", + "image_path": "426e721ff4e3e49c7249f59700eac77120e5c0c75a0f93d0992916877512c6fa.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 160, + 454, + 168 + ], + "lines": [ + { + "bbox": [ + 414, + 160, + 454, + 168 + ], + "spans": [ + { + "bbox": [ + 414, + 160, + 454, + 168 + ], + "type": "text", + "content": "Iteration 700" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 474, + 63, + 552, + 121 + ], + "blocks": [ + { + "bbox": [ + 474, + 63, + 552, + 121 + ], + "lines": [ + { + "bbox": [ + 474, + 63, + 552, + 121 + ], + "spans": [ + { + "bbox": [ + 474, + 63, + 552, + 121 + ], + "type": "image", + "image_path": "9cf0d2fc9dba3009e263cf7fb0af56a6050dc15af3b345732d813391a06a9d61.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 474, + 121, + 552, + 159 + ], + "blocks": [ + { + "bbox": [ + 474, + 121, + 552, + 159 + ], + "lines": [ + { + "bbox": [ + 474, + 121, + 552, + 159 + ], + "spans": [ + { + "bbox": [ + 474, + 121, + 552, + 159 + ], + "type": "image", + "image_path": "fa0ef4ca7b74b4f3581200f478bf0ff41b19d82859438d3b1586740577121153.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 490, + 160, + 537, + 168 + ], + "lines": [ + { + "bbox": [ + 490, + 160, + 537, + 168 + ], + "spans": [ + { + "bbox": [ + 490, + 160, + 537, + 168 + ], + "type": "text", + "content": "Iteration 20000" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 217, + 553, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 553, + 289 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 217, + 553, + 289 + ], + "type": "inline_equation", + "content": "B(t)" + }, + { + "bbox": [ + 313, + 217, + 553, + 289 + ], + "type": "text", + "content": " represents the motion-blurred image, which we model as the average of " + }, + { + "bbox": [ + 313, + 217, + 553, + 289 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 217, + 553, + 289 + ], + "type": "text", + "content": " sharp images " + }, + { + "bbox": [ + 313, + 217, + 553, + 289 + ], + "type": "inline_equation", + "content": "C_i(t)" + }, + { + "bbox": [ + 313, + 217, + 553, + 289 + ], + "type": "text", + "content": " rendered from the 3D Gaussians at " + }, + { + "bbox": [ + 313, + 217, + 553, + 289 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 217, + 553, + 289 + ], + "type": "text", + "content": " virtual camera poses interpolated between the camera poses at the start and end of the exposure time period " + }, + { + "bbox": [ + 313, + 217, + 553, + 289 + ], + "type": "inline_equation", + "content": "\\pmb{\\xi}_t^s" + }, + { + "bbox": [ + 313, + 217, + 553, + 289 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 217, + 553, + 289 + ], + "type": "inline_equation", + "content": "\\pmb{\\xi}_t^e" + }, + { + "bbox": [ + 313, + 217, + 553, + 289 + ], + "type": "text", + "content": ". This approximation allows " + }, + { + "bbox": [ + 313, + 217, + 553, + 289 + ], + "type": "inline_equation", + "content": "B(t)" + }, + { + "bbox": [ + 313, + 217, + 553, + 289 + ], + "type": "text", + "content": " to closely match the input image " + }, + { + "bbox": [ + 313, + 217, + 553, + 289 + ], + "type": "inline_equation", + "content": "I(t)" + }, + { + "bbox": [ + 313, + 217, + 553, + 289 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 289, + 554, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 554, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 554, + 361 + ], + "type": "text", + "content": "Using the refined camera poses and point cloud from the previous step, we incorporate depth information to achieve geometrically accurate deblurring, following the approach in recent work [49]. Specifically, we minimize the difference between the input image " + }, + { + "bbox": [ + 313, + 289, + 554, + 361 + ], + "type": "inline_equation", + "content": "I(t)" + }, + { + "bbox": [ + 313, + 289, + 554, + 361 + ], + "type": "text", + "content": " and the averaged rendered image " + }, + { + "bbox": [ + 313, + 289, + 554, + 361 + ], + "type": "inline_equation", + "content": "B(t)" + }, + { + "bbox": [ + 313, + 289, + 554, + 361 + ], + "type": "text", + "content": ", as shown below:" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 361, + 368, + 553, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 368, + 553, + 393 + ], + "spans": [ + { + "bbox": [ + 361, + 368, + 553, + 393 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {B} = \\sum_ {t} \\sum_ {i} \\| I _ {t} (\\mathbf {p} _ {i}) - B _ {t} (\\mathbf {p} _ {i}) \\| _ {1}. \\tag {7}", + "image_path": "bf2bb2921c0a4906e421e46958ef344cb9edfb2ea5de83f307abba52b345af53.jpg" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 399, + 554, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 554, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 554, + 436 + ], + "type": "text", + "content": "The final deblur loss function integrates this alignment loss with additional terms for structural similarity and depth consistency:" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 337, + 444, + 553, + 458 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 444, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 337, + 444, + 553, + 458 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d e b l u r}} = \\left(1 - \\lambda_ {B}\\right) \\mathcal {L} _ {B} + \\lambda_ {B} \\mathcal {L} _ {\\mathrm {D} - \\mathrm {S S I M}} + \\lambda_ {D} \\mathcal {L} _ {D}. \\tag {8}", + "image_path": "c47d305edb56468cb65a5afee936e86db54aa6f999fe328a8ca42f4b6de263e0.jpg" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 464, + 554, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 554, + 512 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 554, + 512 + ], + "type": "text", + "content": "This combined loss helps ensure that the deblurred output aligns well with both the input image and depth structure, leading to sharper and more geometrically consistent results." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 523, + 367, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 523, + 367, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 523, + 367, + 536 + ], + "type": "text", + "content": "4. Results" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "Dataset preparation. We prepare both synthetic and real-world datasets to evaluate our method's performance. In the absence of an RGB-D dataset with extreme motion blur, we generate a synthetic reference dataset by rendering directly from 3D models. We create RGB-D image sequences and record per-frame camera trajectories to evaluate our pose estimation and deblurring accuracy. The sequences, ranging from 50 to 150 frames, are captured along simulated trajectories that include rapid translations, 360-degree rotations, and human movement paths (using Blender's Walk Navigation feature). The RGB and depth images have a resolution of " + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "inline_equation", + "content": "640 \\times 480" + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": ", and motion blur is applied by enabling Blender's motion blur setting. The 3D models are sourced from the McGuire archive [23] and the Blender archive [43]." + } + ] + } + ], + "index": 39 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26429" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 66, + 549, + 469 + ], + "blocks": [ + { + "bbox": [ + 56, + 66, + 549, + 469 + ], + "lines": [ + { + "bbox": [ + 56, + 66, + 549, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 66, + 549, + 469 + ], + "type": "image", + "image_path": "2cf99cda93bb7ddcff2d8a58587e6a4118a5d9f87ab4ea640829dab6d1959041.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 474, + 555, + 507 + ], + "lines": [ + { + "bbox": [ + 55, + 474, + 555, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 555, + 507 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison of deblurring performance from a novel view. Our method excels in restoring high-frequency details, whereas several methods, including COLMAP [31], fail under severe motion blur conditions (see Section 4.1). The corresponding quantitative results are provided in Table 2." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 522, + 295, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 631 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 631 + ], + "type": "text", + "content": "For real scene evaluation, we used an Azure Kinect RGB-D camera with a 33.3ms fixed exposure time, 5000K fixed white balance, NFOV-2 × 2-binned depth capture mode, and " + }, + { + "bbox": [ + 55, + 522, + 295, + 631 + ], + "type": "inline_equation", + "content": "1280 \\times 720" + }, + { + "bbox": [ + 55, + 522, + 295, + 631 + ], + "type": "text", + "content": " RGB resolution. We used the provided intrinsic camera parameters to undistort the images and aligned depth maps to color images using the SDK's transformation functions. Depth maps were eroded three times to remove outliers, which commonly occur along edges." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "content": "Implementation details. Our code is based on 3D Gaussian Splatting [14], with data processing partially adapted from CF-3DGS [14]. For initial camera pose estimation, we use the Unimatch [48] optical flow network and the ICP [5] algorithm from Open3D as part of the optimization process. For an input image sequence of " + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "content": " frames," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 522, + 555, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 555, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 555, + 631 + ], + "type": "text", + "content": "we prune Gaussians every " + }, + { + "bbox": [ + 313, + 522, + 555, + 631 + ], + "type": "inline_equation", + "content": "100N" + }, + { + "bbox": [ + 313, + 522, + 555, + 631 + ], + "type": "text", + "content": " iterations, reset opacity to 0.5 every " + }, + { + "bbox": [ + 313, + 522, + 555, + 631 + ], + "type": "inline_equation", + "content": "200N" + }, + { + "bbox": [ + 313, + 522, + 555, + 631 + ], + "type": "text", + "content": " iterations, and run a total of " + }, + { + "bbox": [ + 313, + 522, + 555, + 631 + ], + "type": "inline_equation", + "content": "600N" + }, + { + "bbox": [ + 313, + 522, + 555, + 631 + ], + "type": "text", + "content": " iterations (Section 3.2). The number of virtual views for deblurring is set to 15 for synthetic dataset evaluation and 17 for real-scene comparison. All experiments were conducted on an NVIDIA RTX A6000 GPU (48GB), with initial camera pose optimization taking 11 seconds per frame, 0.1 seconds per iteration for pose and geometry refinement, and 0.26 seconds per iteration for deblurring at " + }, + { + "bbox": [ + 313, + 522, + 555, + 631 + ], + "type": "inline_equation", + "content": "640 \\times 480" + }, + { + "bbox": [ + 313, + 522, + 555, + 631 + ], + "type": "text", + "content": " resolution." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 635, + 450, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 450, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 450, + 649 + ], + "type": "text", + "content": "4.1.Quantitative Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": "We evaluate the camera pose accuracy and deblurring performance on three synthetic scenes from our dataset, each rendered with extreme motion blur. For pose accuracy, we compare our method against a variety of approaches capable of generating a 3D point cloud and camera trajectory:" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26430" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 66, + 549, + 487 + ], + "blocks": [ + { + "bbox": [ + 59, + 66, + 549, + 487 + ], + "lines": [ + { + "bbox": [ + 59, + 66, + 549, + 487 + ], + "spans": [ + { + "bbox": [ + 59, + 66, + 549, + 487 + ], + "type": "image", + "image_path": "63db02ccec4b5616198a2333aa1148c6ba4dfb8a09d5fcfe25fee7f35c63a572.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 495, + 555, + 529 + ], + "lines": [ + { + "bbox": [ + 55, + 495, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 555, + 529 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison of deblurring performance from novel views on real scenes captured with an Azure Kinect camera. Our method effectively restores high-frequency details, outperforming other methods even when processing RGB-D sequences with large spatial gaps between frames due to rapid camera motion and severe motion blur." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 544, + 294, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 544, + 294, + 651 + ], + "spans": [ + { + "bbox": [ + 54, + 544, + 294, + 651 + ], + "type": "text", + "content": "COLMAP [31], ICP with FPFH feature-based global alignment [29], ORB-SLAM3 [3], ElasticFusion [47], and several RGB-D SLAM methods, including NICE-SLAM [51], Point-SLAM [30], and DROID-SLAM [42]. Additionally, we include recent Gaussian Splatting-based SLAM approaches such as MonoGS [21] and SplaTAM [12]. For each method, we calculate absolute trajectory error (ATE) and relative pose error (RPE) in meter and degree units, using the ground truth camera trajectories as references." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 653, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 653, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 653, + 295, + 715 + ], + "type": "text", + "content": "To evaluate deblurring accuracy, we modify a state-of-the-art Gaussian Splitting-based image restoration method [49] to operate in a depth-aware manner using undistorted color and depth images. We select one out of every three images as a test view and render the correspond" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 544, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 555, + 604 + ], + "type": "text", + "content": "ing novel view for comparison with the ground truth. For quantitative metrics, we calculate PSNR, SSIM, and LPIPS on the rendered images, as well as RMSE between the rendered and ground truth depth maps in the inverse depth domain " + }, + { + "bbox": [ + 313, + 544, + 555, + 604 + ], + "type": "inline_equation", + "content": "(m^{-1})" + }, + { + "bbox": [ + 313, + 544, + 555, + 604 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 612, + 408, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 408, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 408, + 625 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": "To validate the contributions of each component in our method, we perform an ablation study evaluating camera pose and deblurring accuracy by selectively omitting key elements. Specifically, we conduct experiments without fixing the scale of the Gaussians (Section 3.2), without the pose and geometry refinement process (Equation 5), and without depth loss (Equation 8)." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26431" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 106, + 555, + 181 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of camera pose accuracy against ground truth poses. We report scores only for methods that successfully provide both camera poses and a point cloud (see Section 4.1). Our method achieves the best scores on RPE metrics and shows ATE accuracy comparable to DROID-SLAM [42]. Each color highlights the best and second best results." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 106, + 555, + 181 + ], + "lines": [ + { + "bbox": [ + 56, + 106, + 555, + 181 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 555, + 181 + ], + "type": "table", + "html": "
BedroomLivingroomOffice
ATE ↓RPE (trans) ↓RPE (rot) ↓ATE ↓RPE (trans) ↓RPE (rot) ↓ATE ↓RPE (trans) ↓RPE (rot) ↓
ElasticFusion [47]1.3030.0792.2181.0140.0562.7770.4790.0380.480
FPFH-ICP [29]2.7600.1182.3102.0280.0921.2130.1390.0090.167
DROID-SLAM [42]0.0590.0441.2120.0500.0140.3170.0110.0110.245
MonoGS [21]0.5060.0410.8120.1350.0090.1240.0450.0110.180
Ours0.1020.0060.0920.0050.0020.0320.0410.0030.046
", + "image_path": "62f4a7e3811fa0f2c54a8a9981e7eed931cdb98a5b27672cbb7427697a8e675d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 214, + 555, + 283 + ], + "blocks": [ + { + "bbox": [ + 55, + 190, + 555, + 213 + ], + "lines": [ + { + "bbox": [ + 55, + 190, + 555, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 555, + 213 + ], + "type": "text", + "content": "Table 2. Quantitative comparison of deblurring performance from novel views. Our method demonstrates high accuracy across all metrics for both color and depth images. RMSE is calculated by comparing the rendered depth with GT depth in the inverse depth domain " + }, + { + "bbox": [ + 55, + 190, + 555, + 213 + ], + "type": "inline_equation", + "content": "(m^{-1})" + }, + { + "bbox": [ + 55, + 190, + 555, + 213 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 214, + 555, + 283 + ], + "lines": [ + { + "bbox": [ + 56, + 214, + 555, + 283 + ], + "spans": [ + { + "bbox": [ + 56, + 214, + 555, + 283 + ], + "type": "table", + "html": "
BedroomLivingroomOffice
PSNR ↑SSIM ↑LPIPS ↓RMSE ↓PSNR ↑SSIM ↑LPIPS ↓RMSE ↓PSNR ↑SSIM ↑LPIPS ↓RMSE ↓
ElasticFusion [47]18.7660.5860.4661.24913.7990.4480.5540.14619.0520.5330.4010.079
FPFH-ICP [29]20.6230.6170.4180.45018.1720.5980.4220.10222.2120.6840.2410.033
DROID-SLAM [42]21.3750.6660.3020.01321.2220.7350.2500.03525.6800.7690.1440.013
MonoGS [21]23.7280.7170.2850.00924.4140.8220.1880.02724.1940.7250.1840.184
Ours26.7450.8240.2060.00527.6500.9000.1500.01025.6490.7910.1600.014
", + "image_path": "af31e8321d226fbe0eb064773042dc376a8df05fb0858346202b0262cc711bcc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 56, + 341, + 294, + 380 + ], + "blocks": [ + { + "bbox": [ + 55, + 297, + 295, + 340 + ], + "lines": [ + { + "bbox": [ + 55, + 297, + 295, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 295, + 340 + ], + "type": "text", + "content": "Table 3. Ablation study on the effects of the global refinement process and fixed scale of Gaussians. Our refinement significantly enhances deblurring and depth accuracy while fixing the scale of the Gaussians leads to more accurate camera pose estimation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 341, + 294, + 380 + ], + "lines": [ + { + "bbox": [ + 56, + 341, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 341, + 294, + 380 + ], + "type": "table", + "html": "
PSNRSSIMLPIPSRMSEATERPE(trans)RPE(rot)
w/o scale fix21.6560.7030.3080.0270.1550.0531.488
w/o refinement22.2230.7340.2380.0270.1230.0070.047
w/o depth loss25.5180.8070.1700.0140.1030.0120.195
Scale fix + refinement26.6810.8380.1720.0100.1030.0120.195
", + "image_path": "273dbbba05fb83080d825dc6f01171b82d8d82dbede7a8a8215a2e1cb1d9dea5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 389, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 389, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 54, + 389, + 295, + 521 + ], + "type": "text", + "content": "We assess the effectiveness of these components by performing deblurring based on results from Section 3.1—omitting refinement of camera poses and geometry—and by allowing the scale of the 3D Gaussians to be optimized freely. Table 3 demonstrates that both the global refinement process and fixed Gaussian scale are essential for achieving high accuracy. Notably, during pose refinement, the algorithm reduces absolute pose error even at the expense of a slight increase in relative pose error, underscoring the importance of our refinement approach for stable, high-quality reconstruction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 521, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 521, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 54, + 521, + 295, + 664 + ], + "type": "text", + "content": "Our method achieves superior RPE scores across all scenes, as shown in Table 1, outperforming other approaches in pose and geometry estimation under severe motion blur. While our ATE scores are comparable to those of DROID-SLAM, which uses a bundle adjustment (BA) module, ATE alone does not correlate strongly with deblurring quality. Our method consistently produces more accurate and denser point clouds, which directly contributes to improved deblurring performance, as evidenced by the metrics in Table 2. These results confirm that our integrated approach to pose estimation and deblurring is more resilient to challenging conditions than other methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 672, + 185, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 185, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 185, + 685 + ], + "type": "text", + "content": "4.3. Qualitative Evaluation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "Figures 6 and 7 present qualitative comparisons of our method's performance in synthetic and real scenes. The" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 298, + 555, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 298, + 555, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 298, + 555, + 393 + ], + "type": "text", + "content": "densification capability of Gaussian Splitting provides detailed scene representation even from a relatively sparse point cloud; however, our approach, with its initially dense and accurate point cloud generation, achieves notably superior deblurring performance, especially in restoring high-frequency details. Our method consistently demonstrates a significant advantage in visual fidelity and detail restoration across diverse conditions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 399, + 388, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 388, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 388, + 411 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 418, + 555, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 555, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 555, + 611 + ], + "type": "text", + "content": "We have introduced a robust method for 3D scene reconstruction from RGB-D image sequences that effectively addresses the challenges of extreme motion blur and low-light conditions. Our approach leverages optical flow from color images and a carefully designed geometric loss to achieve accurate global alignment, followed by local refinement of camera poses and geometry using ICP. By integrating these steps into a Gaussian Splatting pipeline, we further refine camera poses and 3D geometry by minimizing depth and opacity losses. Additionally, fixing the scale of the 3D Gaussians ensures that the depth information from the RGB-D input is fully utilized, allowing for precise camera pose estimation and improved deblurring performance. Our method demonstrates superior performance on real and synthetic RGB-D scenes with significant motion blur, outperforming existing approaches." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 618, + 417, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 417, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 417, + 632 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 634, + 554, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 634, + 554, + 706 + ], + "spans": [ + { + "bbox": [ + 313, + 634, + 554, + 706 + ], + "type": "text", + "content": "Min H. Kim acknowledges the Samsung Research Funding & Incubation Center (SRFC-IT2402-02), the Korea NRF grant (RS-2024-00357548), the MSIT/IITP of Korea (RS2022-00155620, RS-2024-00398830, 2022-0-00058, and 2017-0-00072), Microsoft Research Asia, and Samsung Electronics." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26432" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Dejan Azinović, Ricardo Martin-Brualla, Dan B Goldman, Matthias Nießner, and Justus Thies. Neural rgb-d surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6290-6301, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 201 + ], + "type": "text", + "content": "[2] Michael Bloesch, Jan Czarnowski, Ronald Clark, Stefan Leutenegger, and Andrew J Davison. Codeslam—learning a compact, optimisable representation for dense visual slam. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2560–2568, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 202, + 295, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 295, + 256 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 295, + 256 + ], + "type": "text", + "content": "[3] Carlos Campos, Richard Elvira, Juan J Gómez Rodríguez, José MM Montiel, and Juan D Tardós. Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics, 37(6):1874-1890, 2021. 2, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 258, + 295, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 290 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 290 + ], + "type": "text", + "content": "[4] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. arXiv preprint arXiv:2204.04676, 2022. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 291, + 294, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 294, + 323 + ], + "type": "text", + "content": "[5] Yang Chen and Gérard Medioni. Object modelling by registration of multiple range images. Image and vision computing, 10(3):145-155, 1992. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 324, + 295, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 324, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 324, + 295, + 378 + ], + "type": "text", + "content": "[6] Angela Dai, Matthias Nießner, Michael Zollofer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface re-integration. ACM Transactions on Graphics 2017 (TOG), 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "type": "text", + "content": "[7] Rob Fergus, Barun Singh, Aaron Hertzmann, Sam T Roweis, and William T Freeman. Removing camera shake from a single photograph. In Acm Siggraph 2006 Papers, pages 787-794. 2006. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 425, + 294, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 425, + 294, + 457 + ], + "spans": [ + { + "bbox": [ + 62, + 425, + 294, + 457 + ], + "type": "text", + "content": "[8] Kyle Genova, Forrester Cole, Avneesh Sud, Aaron Sarna, and Thomas Funkhouser. Deep structured implicit functions. arXiv preprint arXiv:1912.06126, 2:2, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 459, + 295, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 459, + 295, + 512 + ], + "spans": [ + { + "bbox": [ + 62, + 459, + 295, + 512 + ], + "type": "text", + "content": "[9] Zan Gojcic, Caifa Zhou, Jan D Wegner, and Andreas Wieser. The perfect match: 3d point cloud matching with smoothed densities. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5545-5554, 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 514, + 295, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 295, + 547 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 295, + 547 + ], + "type": "text", + "content": "[10] Alex Graves and Alex Graves. Long short-term memory. Supervised sequence labelling with recurrent neural networks, pages 37-45, 2012. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 548, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 548, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 548, + 295, + 590 + ], + "type": "text", + "content": "[11] Olaf Kähler, Victor Prisacariu, Julien Valentin, and David Murray. Hierarchical voxel block hashing for efficient integration of depth images. IEEE Robotics and Automation Letters, 1(1):192-197, 2015. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 592, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 592, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 57, + 592, + 295, + 657 + ], + "type": "text", + "content": "[12] Nikhil Keetha, Jay Karhade, Krishna Murthy Jatavallabhula, Gengshan Yang, Sebastian Scherer, Deva Ramanan, and Jonathon Luiten. Splatam: Splat track & map 3d gaussians for dense rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21357-21366, 2024. 3, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "text", + "content": "[13] Maik Keller, Damien Lefloch, Martin Lambers, Shahram Izadi, Tim Weyrich, and Andreas Kolb. Real-time 3d reconstruction in dynamic scenes using point-based fusion. In 2013 International Conference on 3D Vision-3DV 2013, pages 1-8. IEEE, 2013. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[14] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "type": "text", + "content": "[15] Christian Kerl, Jürgen Sturm, and Daniel Cremers. Dense visual slam for rgb-d cameras. In 2013 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 2100-2106. IEEE, 2013. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 196 + ], + "type": "text", + "content": "[16] Dilip Krishnan and Rob Fergus. Fast image deconvolution using hyper-laplacian priors. Advances in neural information processing systems, 22, 2009. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "type": "text", + "content": "[17] Dogyoon Lee, Minhyeok Lee, Chajin Shin, and Sangyoun Lee. Dp-nerf: Deblurred neural radiance field with physical scene priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12386–12396, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "type": "text", + "content": "[18] Yang Li and Tatsuya Harada. Lepard: Learning partial point cloud matching in rigid and deformable scenes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5554-5564, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 299, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 299, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 553, + 353 + ], + "type": "text", + "content": "[19] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-nerf: Neural radiance fields from blurry images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12861–12870, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 354, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 409 + ], + "type": "text", + "content": "[20] Nico Marniok, Ole Johannsen, and Bastian Goldluecke. An efficient octree design for local variational range image fusion. In Pattern Recognition: 39th German Conference, GCPR 2017, Basel, Switzerland, September 12-15, 2017, Proceedings 39, pages 401-412. Springer, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "type": "text", + "content": "[21] Hidenobu Matsuki, Riku Murai, Paul HJ Kelly, and Andrew J Davison. Gaussian splatting slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18039-18048, 2024. 3, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 456, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 553, + 510 + ], + "type": "text", + "content": "[22] John McCormac, Ankur Handa, Andrew Davison, and Stefan Leutenegger. Semantic fusion: Dense 3d semantic mapping with convolutional neural networks. In 2017 IEEE International Conference on Robotics and automation (ICRA), pages 4628-4635. IEEE, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 511, + 534, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 511, + 534, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 511, + 534, + 521 + ], + "type": "text", + "content": "[23] Morgan McGuire. Computer graphics archive, 2017. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 524, + 553, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 524, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 317, + 524, + 553, + 576 + ], + "type": "text", + "content": "[24] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 579, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 621 + ], + "type": "text", + "content": "[25] Raul Mur-Artal and Juan D Tardós. Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras. IEEE transactions on robotics, 33(5):1255-1262, 2017. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 624, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 677 + ], + "type": "text", + "content": "[26] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3883-3891, 2017. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[27] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26433" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "Kinectfusion: Real-time dense surface mapping and tracking. In 2011 10th IEEE international symposium on mixed and augmented reality, pages 127-136. IEEE, 2011. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "type": "text", + "content": "[28] Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Marc Stamminger. Real-time 3d reconstruction at scale using voxel hashing. ACM Transactions on Graphics (ToG), 32 (6):1-11, 2013. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 196 + ], + "type": "text", + "content": "[29] Radu Bogdan Rusu, Nico Blodow, and Michael Beetz. Fast point feature histograms (fpfh) for 3d registration. In 2009 IEEE international conference on robotics and automation, pages 3212-3217. IEEE, 2009. 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 295, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 295, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 295, + 241 + ], + "type": "text", + "content": "[30] Erik Sandström, Yue Li, Luc Van Gool, and Martin R Oswald. Point-slam: Dense neural point cloud-based slam. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18433-18444, 2023. 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 295, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 295, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 295, + 284 + ], + "type": "text", + "content": "[31] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 287, + 296, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 296, + 331 + ], + "type": "text", + "content": "[32] Thomas Schops, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 134-144, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 333, + 295, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 295, + 376 + ], + "type": "text", + "content": "[33] Thomas Schöps, Torsten Sattler, and Marc Pollefeys. Surfelmeshing: Online surfel-based mesh reconstruction. IEEE transactions on pattern analysis and machine intelligence, 42(10):2494-2507, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 377, + 295, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 295, + 421 + ], + "type": "text", + "content": "[34] Christian J Schuler, Michael Hirsch, Stefan Harmeling, and Bernhard Scholkopf. Learning to deblur. IEEE transactions on pattern analysis and machine intelligence, 38(7):1439-1451, 2015. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 422, + 295, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 295, + 466 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 295, + 466 + ], + "type": "text", + "content": "[35] Otto Seiskari, Jerry Yilammi, Valtteri Kaatrasalo, Pekka Rantalankila, Matias Turkulainen, Juho Kannala, and Arno Solin. Gaussian splattering on the move: Blur and rolling shutter compensation for natural camera motion, 2024. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 468, + 295, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 295, + 500 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 295, + 500 + ], + "type": "text", + "content": "[36] Qi Shan, Jiaya Jia, and Aseem Agarwala. High-quality motion deblurring from a single image. Acm transactions on graphics (tog), 27(3):1-10, 2008. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 502, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 502, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 502, + 295, + 555 + ], + "type": "text", + "content": "[37] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. Advances in neural information processing systems, 28, 2015. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 557, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 295, + 601 + ], + "type": "text", + "content": "[38] Jörg Stuckler and Sven Behnke. Multi-resolution surfel maps for efficient dense 3d modeling and tracking. Journal of Visual Communication and Image Representation, 25(1):137-147, 2014. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 603, + 295, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 295, + 658 + ], + "type": "text", + "content": "[39] Shuochen Su, Maurizio Delbracio, Jue Wang, Guillermo Sapiro, Wolfgang Heidrich, and Oliver Wang. Deep video deblurring for hand-held cameras. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1279-1288, 2017. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "text", + "content": "[40] Jian Sun, Wenfei Cao, Zongben Xu, and Jean Ponce. Learning a convolutional neural network for non-uniform motion blur removal. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 769-777, 2015. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 555, + 574 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "type": "text", + "content": "[41] Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 553, + 161 + ], + "type": "text", + "content": "[42] Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. Advances in neural information processing systems, 2021. 3, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 163, + 524, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 163, + 524, + 174 + ], + "spans": [ + { + "bbox": [ + 315, + 163, + 524, + 174 + ], + "type": "text", + "content": "[43] Flavio Della Tommasa. Blender demo files, 2024. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 175, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 175, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 315, + 175, + 553, + 229 + ], + "type": "text", + "content": "[44] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. BAD-NeRF: Bundle Adjusted Deblur Neural Radiance Fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4170–4179, 2023. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 231, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 231, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 315, + 231, + 553, + 274 + ], + "type": "text", + "content": "[45] Chen Wenbo and Liu Ligang. Deblur-gs: 3d gaussian splattering from camera motion blurred images. Proc. ACM Comput. Graph. Interact. Tech. (Proceedings of I3D 2024), 7(1), 2024. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 275, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 275, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 315, + 275, + 553, + 329 + ], + "type": "text", + "content": "[46] Thomas Whelan, Michael Kaess, Hordur Johannsson, Maurice Fallon, John J Leonard, and John McDonald. Real-time large-scale densergb-d slam with volumetric fusion. The International Journal of Robotics Research, 34(4-5):598-626, 2015. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 331, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 331, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 315, + 331, + 553, + 385 + ], + "type": "text", + "content": "[47] Thomas Whelan, Renato F Salas-Moreno, Ben Glocker, Andrew J Davison, and Stefan Leutenegger. Elasticfusion: Real-time dense slam and light source estimation. The International Journal of Robotics Research, 35(14):1697-1716, 2016. 2, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 387, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 387, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 315, + 387, + 553, + 430 + ], + "type": "text", + "content": "[48] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, Fisher Yu, Dacheng Tao, and Andreas Geiger. Unifying flow, stereo and depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 4, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 432, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 432, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 315, + 432, + 553, + 464 + ], + "type": "text", + "content": "[49] Lingzhe Zhao, Peng Wang, and Peidong Liu. BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splatting. 2024. 2, 3, 5, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 466, + 555, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 466, + 555, + 508 + ], + "spans": [ + { + "bbox": [ + 315, + 466, + 555, + 508 + ], + "type": "text", + "content": "[50] Huizhong Zhou, Benjamin Ummenhofer, and Thomas Brox. Deeptam: Deep tracking and mapping. In Proceedings of the European conference on computer vision (ECCV), pages 822-838, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 510, + 553, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 510, + 553, + 574 + ], + "spans": [ + { + "bbox": [ + 315, + 510, + 553, + 574 + ], + "type": "text", + "content": "[51] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12786-12796, 2022. 3, 7" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26434" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SplatTalk_ 3D VQA with Gaussian Splatting/71c4eb8c-9be8-49e8-bab3-c4048f8e7d2a_content_list.json b/2025/SplatTalk_ 3D VQA with Gaussian Splatting/71c4eb8c-9be8-49e8-bab3-c4048f8e7d2a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5c946bcef0b7f16663b6995aa7822b905c56d7f2 --- /dev/null +++ b/2025/SplatTalk_ 3D VQA with Gaussian Splatting/71c4eb8c-9be8-49e8-bab3-c4048f8e7d2a_content_list.json @@ -0,0 +1,1318 @@ +[ + { + "type": "text", + "text": "SplatTalk: 3D VQA with Gaussian Splatting", + "text_level": 1, + "bbox": [ + 272, + 130, + 723, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anh Thai $^{1,2*}$", + "bbox": [ + 127, + 179, + 230, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Songyou Peng²", + "bbox": [ + 251, + 179, + 377, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kyle Genova2", + "bbox": [ + 398, + 180, + 513, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Leonidas Guibas2", + "bbox": [ + 532, + 180, + 676, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Thomas Funkhouser", + "bbox": [ + 697, + 180, + 867, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Georgia Institute of Technology", + "bbox": [ + 269, + 204, + 529, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Google DeepMind", + "bbox": [ + 570, + 205, + 727, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "*Work done as a student researcher at Google DeepMind.", + "bbox": [ + 326, + 224, + 665, + 239 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/10c8644c27e3ab27ac7da88a7566145a44b407cd2f94be94301b677c673aae01.jpg", + "image_caption": [ + "Figure 1. SplatTalk. We propose a self-supervised 3D-Language Gaussian Splitting model trained from multi-view RGB images. First, images are encoded using a pretrained 2D Vision-Language Model (VLM) and projected into visual-language feature maps via a multimodal projector. These feature maps are then learned within a feed-forward 3D-Language Gaussian Splitting model, producing a 3D-Language Gaussian Field that encodes spatial and semantic information in 3D space. During inference, the 3D Gaussian features are directly queried by a Large Language Model (LLM) to perform 3D question-answering (3D VQA) tasks." + ], + "image_footnote": [], + "bbox": [ + 101, + 276, + 897, + 448 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 537, + 326, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Language-guided 3D scene understanding is important for advancing applications in robotics, AR/VR, and human-computer interaction, enabling models to comprehend and interact with 3D environments through natural language. While 2D vision-language models (VLMs) have achieved remarkable success in 2D VQA tasks, progress in the 3D domain has been significantly slower due to the complexity of 3D data and the high cost of manual annotations. In this work, we introduce SplatTalk, a novel method that uses a generalizable 3D Gaussian Splatting (3DGS) framework to produce 3D tokens suitable for direct input into a pretrained LLM, enabling effective zero-shot 3D visual question answering (3D VQA) for scenes with only posed images. During experiments on multiple benchmarks, our approach outperforms both 3D models trained specifically for the task and previous 2D-LMM-based models utilizing only images (our setting), while achieving competitive performance with state-of-the-art 3D LMMs that additionally utilize 3D inputs.", + "bbox": [ + 88, + 595, + 485, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Project website: https://splat-talk.github.io/", + "bbox": [ + 89, + 868, + 383, + 882 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 537, + 643, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The rise of Large Multimodal Models (LMMs), following the success of Large Language Models (LLMs), has become increasingly significant in a world where visual understanding is crucial. While 2D LMMs have seen rapid improvements, enabling models to analyze images and answer complex questions [21, 28, 37, 40], progress in 3D LMMs has been slower. This gap is largely due to the scarcity of 3D data compared to 2D datasets. Further, the availability of language-annotated 3D data is even more limited, as annotating 3D scenes since it is significantly more complex and expensive than labeling 2D images. These challenges have hindered the development of language-guided 3D reasoning, making 3D LMMs an important yet underexplored area in multimodal research. In this work, we focus on 3D Visual Question Answering (3D VQA) as it provides a strong benchmark for evaluating a model's ability to understand and reason about 3D scenes.", + "bbox": [ + 511, + 565, + 908, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, significant progress has been made in 3D multimodal learning, particularly for indoor scenes [15, 17, 27, 48]. Several approaches rely on posed images and 3D inputs, like meshes [41], point clouds [9, 17, 34] and voxels [13], to associate 3D geometry with language features. As such,", + "bbox": [ + 511, + 825, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4712", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "they have limited applicability in scenarios where a full 3D surface reconstruction is not available.", + "bbox": [ + 89, + 90, + 480, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous works have explored integrating semantic features into NeRF [19] or 3D Gaussian Splatting (3DGS) [35], which rely only upon RGB image inputs. However, research on effectively associating 3D geometry with language features remains limited. Some methods align individual 3D locations with language features via implicit functions optimized individually for every scene [10, 19]. Others employ SAM [20] to first segment objects and object parts, followed by CLIP [36] to extract object-centric features from the generated masks. These approaches allow models to query information about individual points and/or objects. However, it disregards spatial relationships between objects, making it unsuitable for 3D VQA spatial reasoning tasks.", + "bbox": [ + 88, + 123, + 482, + 319 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More recent work [21] has demonstrated the power of LLMs to answer VQA questions using only a sampling of tokens extracted from images, without any underlying 3D representation. While this approach is inspiring, it relies upon the LLM to make sense of a set of tokens that are arranged according to images rather than elements in the 3D scene. As a result, it struggles to answer queries about specific 3D locations (e.g., \"What is here?\") and questions requiring a holistic understanding of the entire 3D scene, where the answer depends on establishing correspondences between objects across multiple images (e.g., \"What is opposite the door at the other end of the room\"?). In this work, we aim to address these issues by consolidating tokens in 3D before providing them to the LLM.", + "bbox": [ + 91, + 321, + 485, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach is called SplatTalk. Given a set of posed RGB images, we integrate language features extracted from the RGB images to build a 3D Gaussian Splitting representation, from which 3D tokens can be extracted for direct input into a pretrained LLM to answer challenging 3D VQA queries (see Fig. 1). The main research contributions are the new methods for 1) integrating language features into the Gaussian representation, and 2) extracting tokens that can be provided directly to an LLM. These methods avoid the issues of concurrent work, ChatSplat [6], which produces tokens that are not directly compatible with an LLM, making language-based reasoning more difficult.", + "bbox": [ + 89, + 536, + 482, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overall, SplatTalk has a unique combination of three advantages. First, it does not require any 3D surface representation as input. Second, it leverages the full power of a pre-trained LLM to answer difficult 3D VQA questions. Third, it produces a concise 3D representation that can be queried and/or edited in 3D (e.g., cropped to a 3D region, rendered from a novel view, etc.). During experiments with 3D VQA benchmarks, we find that it outperforms 3D models trained only for the benchmark tasks, demonstrating the value of utilizing pre-trained LLMs. SplatTalk further outperforms prior image-only methods using LLMs, showing the value of building a 3D scene representation. Finally,", + "bbox": [ + 89, + 719, + 485, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "it even achieves competitive performance with SOTA 3D LMMs that utilize 3D scene representation inputs.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our key contributions are:", + "bbox": [ + 532, + 122, + 790, + 136 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce the first self-supervised 3D Gaussian-based method for large-scale zero-shot 3D VQA, SplatTalk, requiring only multi-view images without explicit depth, point clouds or 3D-language annotated supervision.", + "- We show that mean feature extraction from 3D Gaussians encodes holistic scene concepts for 3D VQA, supported by theoretical justification.", + "- We introduce entropy-adaptive token sampling, leading to better 3D VQA performance without additional training.", + "- We demonstrate that SplatTalk outperforms 2D LMMs and achieves performance on par with 3D models using point clouds on various 3D VQA benchmarks, demonstrating the effectiveness of our 3D-aware representations." + ], + "bbox": [ + 511, + 137, + 906, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 345, + 653, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2D LMMs. With advancements in large language models (LLMs), the ability of models to comprehend images and respond to language prompts related to their content has improved significantly. Notably, SOTA models [1, 11, 21, 30, 32] have demonstrated remarkable capabilities in processing multiple images of a 3D scene and answering questions about the shared content across these images. Although multiple RGB images inherently contain 3D information, they do not explicitly model 3D understanding, potentially leading to sub-optimal performance on tasks that require explicit 3D reasoning. In this work, we build upon the strong performance of 2D vision-language models (2D-VLMs), particularly LLaVA-OneVision [21], and integrate them with the 3D priors provided by the 3DGS framework to improve 3D spatial understanding capabilities.", + "bbox": [ + 511, + 371, + 906, + 598 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D LMMs. Recent advancements in 3D indoor scene understanding have explored integrating 3D priors into LLMs using inputs such as 3D point clouds [7, 17] or multi-view RGB images combined with depth or 3D point clouds [48]. These approaches employ various strategies to embed 3D structural information and language features into a shared representation space. While most methods [17, 48] fine-tune pretrained visual encoders (both 2D and 3D) alongside LoRA parameters or even the full set of LLM parameters, others [18, 24] opt for a more lightweight approach, training only a projection layer to map visual features into the language feature space. In contrast, our method does not require any explicit depth or 3D inputs. Instead, we leverage the 3DGS framework, which has shown strong capabilities in representing 3D scenes using only multi-view RGB images. Generalizable 3D Gaussian Splatting. This body of work explores generalizable 3DGS as an alternative to the perscene optimization paradigm used in traditional 3DGS methods, aiming to improve novel view synthesis across unseen scenes. Our method builds on FreeSplat [43], which is", + "bbox": [ + 511, + 598, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4713", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/84c7612776e61cc762c577c9f35bae46a16175dfc583846ddcb7ca4e1203f7c8.jpg", + "image_caption": [ + "Figure 2. Overview. Left: During the self-supervised 3D-language Gaussian Splitting training phase, multiple RGB input views are first encoded into Gaussian latent features (Gaussian triplets). These latent features are then decoded into Gaussian parameters for rendering, along with a low-dimensional visual-language feature $f$ . To ensure proper supervision of this low-dimensional feature, we train an autoencoder that maps the high-dimensional, unbounded features obtained from LLaVA-OV, specifically, the visual tokens serving as direct inputs to the LLM, onto a low-dimensional hypersphere space. Right: During 3D VQA inference, visual-language features are directly extracted from the 3D Gaussians. These features are then mapped back to the original high-dimensional space using the pretrained decoder and subsequently used as direct visual token inputs to the LLM. LoRA fine-tuning of the LLM is optional." + ], + "image_footnote": [], + "bbox": [ + 98, + 89, + 669, + 299 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/85c094a9cea333eefbe69806bb71b0124ee4c9f8bfaf92e316236e14d9993baf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 89, + 901, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "designed for sparse-view synthesis in indoor scenes. The method begins by extracting multi-scale features from input RGB images using CNN backbones. It then constructs adaptive cost volumes from neighboring views to predict depth maps. These predicted depth maps are used to backproject 2D CNN-encoded features into 3D Gaussian triplets $(\\mu, \\omega, f)$ where $\\mu$ denotes the Gaussian's center, $\\omega$ the weight, and $f$ the latent feature. To integrate multi-view information, FreeSplat introduces a Pixel-wise Triplet Fusion (PTF) module that uses a lightweight GRU to align and fuse overlapping 3D Gaussians across views efficiently. Finally, the fused triplets are passed through an MLP to decode rendering parameters, $\\Sigma, \\alpha, s$ . Our approach extends the FreeSplat framework to enable the learning of generalizable 3D Gaussian language fields.", + "bbox": [ + 88, + 415, + 483, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similar to FreeSplat [43], [5, 8] leverage multi-view stereo (MVS) cues to refine depth estimation and improve geometric consistency in 3D Gaussian representations. These methods primarily focus on RGB novel view synthesis from sparse inputs (2-10 views). Our work builds on the SOTA model, FreeSplat [43], but diverges significantly in both objectives and methodology. First, while previous methods focus on photorealistic RGB reconstruction, we aim at semantic 3D scene understanding. Second, instead of reconstructing only a localized portion of the scene as shown in these works, our goal is to represent entire 3D environments, capturing both spatial and semantic relationships holistically. Language-guided 3D Gaussian Splatting. LangSplat [35] is one of the first approaches that integrates semantic features into 3DGS. However, this method requires training on individual scenes, which is significantly expensive, and lacks generalization to unseen scenes. Similarly, follow-up", + "bbox": [ + 89, + 642, + 485, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "works [23, 45] explore 3DGS for open-vocabulary object segmentation but do not focus on reasoning tasks like 3D VQA. Further, they primarily rely on scene-specific training, limiting their scalability. In contrast, we propose a fully self-supervised framework that learns 3D-aware visual representations without requiring per-scene training. While some works [12, 42] explores a generalizable 3DGS and semantic modeling approach, they primarily focus on open-vocabulary segmentation, without addressing the broader challenges of free-form language reasoning required for 3D VQA.", + "bbox": [ + 511, + 415, + 906, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ChatSplat [6] is a concurrent work that explores natural language conversation within the 3DGS framework. However, it optimizes on a per-scene basis and requires rendering 3D features into 2D feature maps during inference. Additionally, its training pipeline is not fully end-to-end, as it first trains the RGB Gaussians separately before incorporating language features. In contrast, our approach leverages scene latent features that encapsulate both RGB and semantic information, enabling the joint optimization of both modalities in a fully end-to-end manner. Our approach extends the FreeSplat framework to enable the learning of generalizable 3D Gaussian language fields.", + "bbox": [ + 511, + 566, + 908, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 766, + 604, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The problem we address is formulated as follows: Given a sequence of $N$ posed RGB images captured from a 3D indoor scene, along with language prompts related to the scene's content, the goal is to generate the corresponding language responses. That is, we address 3D VQA tasks with solely posed RGB images as inputs.", + "bbox": [ + 511, + 792, + 905, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We provide an overview of our approach in Fig. 2, which", + "bbox": [ + 532, + 885, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4714", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/396b4405b322ed870007efe67c8c51355a6dbc783b7af71262143472d6623b90.jpg", + "image_caption": [ + "Figure 3. Qualitative results on ScanQA, scene0030_00 and scene0084_00. We compare responses from LLaVA-OV, our model (Ours), and the ground truth (GT) for spatial reasoning questions in 3D VQA. Each scene highlights the referenced objects with red circles, and key relative objects are marked in blue. The answers from each model are displayed in color-coded speech bubbles: LLaVA-OV (brown), Ours (blue), and GT (green). With its 3D-aware representation, our model exhibits improved spatial reasoning capabilities, accurately identifying relationships between objects that are far apart and may never co-occur in the same image (e.g., door and window). More examples in Supp." + ], + "image_footnote": [], + "bbox": [ + 99, + 92, + 488, + 324 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cd69f61788950ca6a3f9c919091ec419ab45ec136019317853b7eaa878c619d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 90, + 893, + 324 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "consists of three main steps: 1) Training the feature autoencoder, 2) Self-supervised 3D-Language Gaussian Splitting Training and 3) 3D VQA Inference. First, we extract the high-dimensional language features from LLaVA-OV [21] for each RGB image by using the pretrained image encoder and multimodal projector. Since these features are high-dimensional, unbounded and sparse, we train an autoencoder to project them into a more compact, low-dimensional space. These low-dimensional features then serve as pseudo groundtruth for training the language representations in SplatTalk. SplatTalk architecture is a feed-forward network consisting of a Gaussian encoder and decoder that predicts Gaussian parameters for rendering. During 3D VQA inference stage, we extract the trained language features from the 3D Gaussians and directly feed them into the LLM, along with the language prompt, to generate free-form responses.", + "bbox": [ + 88, + 417, + 485, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Integrating Language Features", + "text_level": 1, + "bbox": [ + 89, + 667, + 366, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An important step of our computational pipeline is the self-supervised method to integrate language features into a generalizable 3DGS framework. We identify four key factors that shape our approach for this step: (1) Leveraging features that align with the LLM's input space for direct language reasoning, (2) Efficient feature dimensionality reduction for effective CUDA-based rendering, and (3) Joint training of RGB and semantic features for coherent multimodal representation learning.", + "bbox": [ + 89, + 688, + 485, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Visual Tokens for 2D Pseudo Ground-truth Features. We extract the visual tokens used as inputs to the LLM for training our 3DGS framework. Our empirical findings indicate that extracting features directly from the visual encoder for training 3DGS requires retraining the multimodal projector", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "layer which aligns the visual with language spaces. This observation is further corroborated by [6], which demonstrates that features extracted before the multimodal projector cannot be seamlessly integrated into the language model. A possible explanation is that the image encoder's native feature space (before the projector) is high-dimensional and unbounded, leading to sparser features with greater variance across different dimensions. As a result, the projector becomes more sensitive to subtle variations in the input features, making it difficult to directly map them into a token representation that aligns with the LLM's understanding. In contrast, the visual tokens generated after the projector are already well-aligned with the language model's latent space, allowing the LLM to interpret and reason with them directly.", + "bbox": [ + 511, + 417, + 908, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Feature Dimensionality Reduction. Building on the previous factor, directly optimizing these language embeddings within a differentiable rendering pipeline is challenging because they are extremely high-dimensional (e.g., 3584 dimensions), sparse, and contain unbounded values, making learning both unstable and impractical. Therefore, we learn a generalizable autoencoder that projects these sparse, high-dimensional visual tokens into a compact unit hypersphere with much lower dimension (e.g., 256 dimensions), ensuring feature smoothness. Note that the language representations from LLaVA, which support reasoning and free-form language comprehension, are significantly more complex than those from CLIP or segmentation-based models like SAM. Different from prior works that compress semantic features into extremely low-dimensional vectors (e.g., 3-16 dimensions [35, 45]), which simplifies rendering but results in significant information loss, our approach prior", + "bbox": [ + 511, + 643, + 910, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4715", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/2d9625c50e510f125b8027ac5a4e4b090d7e7089e298e11ac1a3613e46dd76da.jpg", + "table_caption": [ + "Table 1. Evaluation of 3D VQA on ScanQA [2] and SQA3D [31] datasets. We compare SplatTalk against various baselines, including 3D LMMs, which take point clouds (PC) or both point clouds and images $(\\mathrm{PC} + \\mathrm{I})$ as visual inputs, and 2D LMM-based models, which rely solely on images (I). As a 2D LMM-based approach, SplatTalk outperforms other 2D LMM-based models, surpasses some 3D LMMs, and achieves competitive performance against the strongest 3D LMM baselines. Our models are highlighted in blue, bold indicates best performance among the 2D-LMM methods." + ], + "table_footnote": [], + "table_body": "
MethodModalityScanQA (val)SQA3D (test)
CIDErMETEORROUGEEM@1EM@1-REM@1EM@1-R
Specialist Models
ScanQA [2]PC64.913.133.121.1-47.2-
ScanRefer+MCan [46]PC55.411.530.018.6---
3D-VisTA [49]PC69.613.935.722.4-48.5-
SQA3D [31]PC-----46.6-
Generalist 3D LMMs
Chat-Scene [16]PC+I87.718.041.621.6-54.657.5
Scene-LLM [13]PC+I80.015.835.9--53.6-
Grounded 3D-LLM [9]PC72.7------
Chat-3D-v2 [44]PC77.116.140.1--54.7-
Finetuned 3D LMMs
3D-LLM [15]PC+I69.414.535.720.5---
LEO [17]PC+I101.420.049.224.547.650.052.4
Scene-LLM [13]PC+I80.016.640.027.2-54.2-
LL3DA [7]PC76.815.937.3----
Chat-3D [44]PC53.211.928.5----
2D LMM-Based
VideoChat2 [25]I49.29.528.219.2-37.3-
GPT-4V [32]I59.613.533.4----
Claude [1]I57.710.029.3----
LLaVA-NeXT-Video [22]I46.29.127.8--34.2-
LLaVA-OV [21]I50.013.029.415.632.525.435.8
SplatTalkI61.714.232.717.132.226.141.3
SplatTalk-ScanQA-FTI77.115.438.722.338.342.044.7
SplatTalk-3DVQA-FTI77.515.638.522.438.347.649.4
", + "bbox": [ + 189, + 170, + 803, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "itizes preserving rich language features by maintaining a higher-dimensional representation (e.g., 256). To efficiently render high-dimensional feature maps on CUDA, we draw inspiration from [47] and implement a parallel CUDA rasterizer pipeline, where semantic features and RGB are rendered simultaneously, using shared Gaussian parameters.", + "bbox": [ + 89, + 625, + 482, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike previous methods [6, 35] that train a separate autoencoder for each individual scene, we train a single autoencoder across all training scenes, each consisting of multiple frames. This approach improves the autoencoder's generalizability, reducing computational overhead by eliminating the need for retraining to compress features from unseen scenes. Training the autoencoder separately from the 3DGS model ensures that it is dedicated solely to compressing high-dimensional feature vectors while preserving the essential scene information without being influenced by the complexities of the rendering pipeline. Our empirical findings further indicate that jointly training a feature de", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "coder to project low-dimensional embeddings back to the high-dimensional space introduces significant instabilities in the training process.", + "bbox": [ + 511, + 625, + 906, + 670 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Joint-training RGB and Language. In contrast to previous approaches that first train RGB Gaussians and then freeze them before integrating language features [6, 35], we adopt a joint training strategy, where RGB and language features are learned simultaneously. This is due to the fact that visual RGB and semantic features play a complementary role in effectively representing 3D scenes [33]. In feedforward models like FreeSplat, where the scene is first encoded into a Gaussian latent space before being decoded into specific Gaussian parameters for RGB rendering, these Gaussian latent features inherently capture holistic scene information, including both appearance (RGB) and semantics. Based on this observation, we introduce a new Gaussian parameter decoder head that predicts semantic features for each 3D Gaussian alongside rendering parameters, enabling joint op-", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4716", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1d005e0cfa789f2fe6e39717f63eb987e351c52c66073cc60f401c7807de748e.jpg", + "table_caption": [ + "Table 2. Evaluation of 3D VQA on MSR3D [27]. We compare SplatTalk against various baselines, all evaluated in a zero-shot setting. T denotes models that rely solely on text inputs without interleaving images, PC indicates models that use point cloud as the visual input, and I represents models that process only image-based inputs. Our approach, highlighted in blue, outperforms all baselines." + ], + "table_footnote": [], + "table_body": "
ModelInputSceneSettingCountingExistenceAttributesSpatialNavigationOthersOverall
LEO [17]TPCzero-shot0.8015.511.87.32.315.37.8
LLaVA-OV [21]TIzero-shot18.631.224.819.516.736.324.0
SplatTalkTIzero-shot19.660.344.035.835.561.841.8
SplatTalk-ScanQA-FTTIzero-shot28.966.043.228.333.860.041.5
SplatTalk-3DVQA-FTTIzero-shot24.357.036.17.816.435.026.5
", + "bbox": [ + 153, + 142, + 841, + 247 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tization. Unlike approaches that explicitly model these interactions [33, 39] using extra cross-attention or projection layers, SplatTalk implicitly integrates semantic and visual information within the Gaussian representation. This paradigm naturally facilitates multimodal interactions between RGB and language features.", + "bbox": [ + 88, + 258, + 480, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To incorporate more views and comprehensively cover the scene, we reduce the input resolution of the RGB images. This adjustment aligns well with the inherently low spatial resolution of the visual tokens extracted from LLaVA-OV, and also prevents the model from disproportionately allocating resources to modeling RGB features, thus encouraging a balanced representation of semantic and RGB information.", + "bbox": [ + 88, + 352, + 482, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "These design choices establish our approach as a fundamentally novel method for integrating language features into 3DGS, grounded in first-principles observations rather than heuristic adaptations. By addressing the core challenges of multimodal alignment, feature compression, and joint optimization, our approach is not merely a trivial extension of existing works, but a carefully designed framework that ensures both effectiveness and stability in learning rich, spatially-aware semantic representations.", + "bbox": [ + 88, + 459, + 482, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2. Extracting 3D Language Features", + "text_level": 1, + "bbox": [ + 89, + 611, + 385, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During inference, we directly extract language features at the mean position of each 3D Gaussian. Although the rendering process incorporates each Gaussian's covariance and opacity when projecting 3D features to 2D, we demonstrate that scene semantics can be effectively captured by the mean feature of each Gaussian, $\\pmb{f}_i$ , alone, without relying on these additional factors. This is supported by the conceptual similarity between our training optimization process and the Expectation-Maximization (EM) algorithm. For simplification, let $\\mathcal{L} = \\sum_{t=1}^{T} \\| \\pmb{F}_t(x) - \\pmb{F}_t^{gt}(x) \\|^2$ be the loss, where $\\pmb{F}_t(x)$ is the predicted rendered feature map of frame $t$ at pixel $x$ and $\\pmb{F}^{gt}$ is the pseudo ground-truth feature map obtained from LLaVA-OV. Replacing $\\pmb{F}_t(x)$ by the rendering equation and solving for the optimal $\\pmb{f}_i^*$ , we obtain $\\pmb{f}_i^* = \\frac{\\sum_{t,x} R_i(t,x) \\pmb{F}_t^{gt}(x)}{\\sum_{t,x} R_i(t,x)}$ where $R_i(t,x) = \\frac{\\alpha_i \\mathcal{N}(x; \\pmb{\\mu}_i, \\pmb{\\Sigma}_i)}{\\sum_j \\alpha_j \\mathcal{N}(x; \\pmb{\\mu}_j, \\pmb{\\Sigma}_j)}$ is the Gaussian's contribution to pixel $x$ at viewpoint $t$ . This mirrors the EM procedure: estimating $R$ by rendering (E-step)", + "bbox": [ + 91, + 633, + 485, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and updating $f_{i}$ by minimizing reconstruction loss (M-step). Consequently, the optimal 3D feature $\\pmb{f}_{i}^{*}$ for each Gaussian is a weighted sum of the 2D feature maps, and the set of 3D Gaussian features $\\{\\pmb{f}_{i}\\}_{i=1}^{N}$ encodes the scene's semantic holistically from $T$ 2D inputs. These features are then projected back into the original language embedding space using the pretrained decoder and directly provided as inputs to the LLM for downstream 3D-language tasks.", + "bbox": [ + 511, + 258, + 906, + 380 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In open-vocabulary segmentation, querying Gaussian centers is effective [4, 35, 45] because each Gaussian is assumed to represent a distinct, meaningful object or region. These methods use object-centric embeddings (e.g., SAM-masked CLIP or LSeg) and retrieve objects via cosine similarity to text queries. In contrast, 3D reasoning tasks lack this one-to-one correspondence. Each Gaussian encodes parts of a scene-level representation, where meaning stems from spatial and relational structure. Querying single Gaussians is thus insufficient for context-dependent reasoning. While our method also queries at Gaussian means, this is not trivial; our EM-based formulation offers a principled justification.", + "bbox": [ + 511, + 381, + 908, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3.3D VQA Inference", + "text_level": 1, + "bbox": [ + 511, + 574, + 692, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Entropy Adaptive Gaussian Sampling. SplatTalk leverages triplet fusion adopted in FreeSplat [43], which reduces the number of Gaussians. However, complex scenes still require significantly more Gaussians than the LLM's visual token limit, and not all are equally informative. To address this, we propose entropy-based sampling: selecting the top $k$ Gaussians with the highest language feature entropy, where $k$ matches the LLM's visual token capacity.", + "bbox": [ + 511, + 597, + 906, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D Tokens as Visual Inputs. After 3DGS training, each 3D Gaussian encodes a feature that is a weighted aggregation of 2D semantic information across all input views (Sec. 3.2). These features are also spatially grounded in their 3D coordinates. Because each token corresponds to a specific location in 3D space, its position is already implicitly encoded in the feature through the rendering and fusion process. Unlike 2D patches in a fixed grid, these 3D tokens form an unordered set, similar to a point cloud where spatial relationships are captured by the features themselves, not by sequence position. Therefore, we do not require explicit positional encodings or a fixed input order. The 3D tokens,", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4717", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "extracted by sampling features at each Gaussian's mean, are directly fed into the LLM.", + "bbox": [ + 89, + 90, + 480, + 122 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4. Optional - Finetuning on 3D VQA Datasets", + "text_level": 1, + "bbox": [ + 89, + 131, + 457, + 148 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While SplatTalk is capable of performing the 3D VQA task in a zero-shot manner, it also allows for flexible fine-tuning on 3D VQA datasets to enhance alignment and reasoning capabilities. Rather than training the model end-to-end, which is both computationally expensive and prone to overfitting on limited datasets, we optimize only the LoRA parameters applied to the LLM during fine-tuning.", + "bbox": [ + 89, + 152, + 482, + 260 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 273, + 222, + 291 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 299, + 307, + 315 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We train SplatTalk with 500 scenes from the ScanQA training set. The initial self-supervised stage focuses purely on learning the 3D scene representation, without leveraging any text-annotated information. For each scene, we uniformly sample 100 views around the scene and extract their per-patch features by passing them through the visual encoder of LLaVA-OV followed by the multi-modal projector. The subsequent SplatTalk model is trained on a single NVIDIA H100 (80GB) GPU. Both the LoRA fine-tuning and inference are also conducted on a single H100. More details can be found in Supp. Material.", + "bbox": [ + 89, + 321, + 483, + 487 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Datasets and Metrics", + "text_level": 1, + "bbox": [ + 89, + 497, + 287, + 512 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Dataset. We evaluate our method on ScanNet-based QA datasets: ScanQA [2], SQA3D [31], and MSR3D [27]. ScanQA is a benchmark designed for question answering in 3D indoor scenes, containing detailed spatial and semantic information extracted from ScanNet. SQA3D and MSR3D further extend this by introducing scenarios that describe an embodied agent's 3D position and state.", + "bbox": [ + 89, + 520, + 483, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics. For ScanQA and SQA3D, we evaluate performance using standard n-gram-based metrics, specifically CIDEr [38], METEOR [3], ROUGE [26], EM@1 [14], and EM@1-Refined for ScanQA, and EM@1 and EM@1-Refined for SQA3D. In contrast, for MSR3D, we utilize the LLM-Match score [29] as recommended by the authors.", + "bbox": [ + 89, + 626, + 483, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Performance on 3D Question Answering Tasks", + "text_level": 1, + "bbox": [ + 89, + 726, + 482, + 743 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ScanQA. We evaluate SplatTalk's performance against several model families, each representing different levels of specialization and generalization. \"Specialist Models\" refer to models specifically trained to solve ScanQA without exposure to other tasks. \"Generalist 3D LMMs\" consist of models trained across a diverse range of 3D vision-language tasks, enabling broader generalization. \"Finetuned 3D LMMs\" are models that have been further fine-tuned on the ScanQA training set to optimize performance for this particular dataset. Lastly, \"2D LMM-Based Models\" include", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "models that process only image inputs without leveraging depth or explicit 3D information. SplatTalk and its fine-tuned variants fall within this category.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate three different variants of our model on this task. The base model (SplatTalk) is trained solely to encode the semantic representation of the scene from multiple RGB images using 3DGS, without exposure to question-answering tasks. Note that this model is trained entirely in a self-supervised manner, without any access to language supervision or textual annotations during training. To further adapt it to downstream 3D VQA tasks, we fine-tune this base model on the ScanQA training set, referring to it as SplatTalk-ScanQA-FT, and on the combined training sets of ScanQA and SQA3D, denoting as SplatTalk-3DVQA-FT. We present results in Table 1.", + "bbox": [ + 511, + 137, + 906, + 318 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Compared to LLaVA-OV, whose 2D language feature maps are used to train SplatTalk, our base model, which benefits from a more 3D-aware representation due to the 3D Gaussian training process, achieves significantly higher performance (e.g. 61.7 vs 50.0 on CIDEr). Our language finetuned model variants, SplatTalk-ScanQA-FT and SplatTalk-3DVQA-FT, improve performance further. Notably, despite not being trained with depth supervision or explicit 3D information, our models achieve comparable results to both Generalist 3D LMMs and Finetuned 3D LMMs, while outperforming Specialist Models. Additionally, among 2D LMM-Based approaches, our base model without being fine-tuned with 3D VQA datasets outperforms other methods, with significantly higher performance on some metrics, including both open-source and closed-source models. We show qualitative results in Fig. 3.", + "bbox": [ + 511, + 319, + 908, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SQA3D. Similar to our evaluation on ScanQA, we compare our model against both 3D LMMs and 2D LMMs. Compared to 2D LLaVA-OV features, our base model with 3D-aware features achieves a significant performance gain, highlighting the advantage of incorporating 3D information via 3DGS which aligns with our findings on ScanQA. Further finetuning on ScanQA alone and on the combined ScanQA + SQA3D dataset leads to substantial improvements. Notably, SplatTalk-3DVQA-FT surpasses specialist models that leverage 3D point cloud inputs, demonstrating the strength of our approach in 3D scene understanding.", + "bbox": [ + 511, + 561, + 908, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "MSR3D. We evaluated our model's performance against LLaVA-OV and $\\mathsf{LEO}^1$ [17] on the Scannet split of MSR3D. Importantly, all comparisons are conducted in a zero-shot setting, as none of the models have been trained on MSR3D. This is an interleaved dataset where inputs consist of both text and images. However, since all models are tested in a zero-shot manner, they rely solely on text inputs, where the agent's state descriptions and the question are combined into", + "bbox": [ + 511, + 729, + 908, + 849 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "1The reported number reflects performance across all splits of MSR3D. However, since LEO was trained on ScanNet and 3RScan, its performance remains largely consistent across splits.", + "bbox": [ + 511, + 862, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4718", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3f658b393c12ad03731f4020caaa97273d25d391926e1323495a2696c5003985.jpg", + "table_caption": [ + "Table 3. Ablation Study on Gaussian Sampling Strategies. Our entropy-based sampling strategy consistently outperforms across multiple metrics." + ], + "table_footnote": [], + "table_body": "
Sampling MethodEM@1METEORROUGESPICEAvg
Random17.013.862.114.026.73
Point Density16.214.059.714.126.00
FPS17.713.862.013.626.78
Entropy17.114.661.714.226.90
", + "bbox": [ + 165, + 125, + 426, + 189 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "a single text prompt. On the visual processing side, LEO operates on 3D point cloud inputs, whereas both LLaVA-OV and our model rely only on RGB images.", + "bbox": [ + 89, + 208, + 482, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results, presented in Table 2, demonstrate that our base model outperforms LLaVA-OV across all question types. Notably, aside from the \"counting\" category, our base model achieves approximately twice the performance of LLaVA-OV on other question types. Both models significantly outperform LEO, which struggles due to the absence of fine-tuning on this dataset.", + "bbox": [ + 89, + 255, + 483, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Interestingly, we observe a slight performance drop when comparing SplatTalk with SplatTalk-ScanQA-FT. However, this drop becomes much more pronounced when compared with SplatTalk-3DVQA-FT. A plausible explanation is that fine-tuning on other QA datasets introduces overfitting, causing the model to adapt too specifically to those datasets. Furthermore, the observed performance drop suggests that MSR3D differs significantly from ScanQA and SQA3D, making direct knowledge transfer less effective. This distinction may also explain why LEO struggles to generalize in a zero-shot setting. Despite being trained on multiple QA datasets, none of them closely resemble MSR3D, limiting its ability to adapt without fine-tuning.", + "bbox": [ + 89, + 362, + 483, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 571, + 254, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Gaussian Sampling Strategies. We explore various Gaussian sampling strategies for selecting tokens as inputs to the LLM, including: (1) Random Sampling—Gaussians are selected randomly across the scene. (2) Point Density Adaptive Sampling—Gaussians are sampled proportionally to the density of 3D points. (3) Furthest Point Sampling (FPS)—Gaussians are iteratively selected at locations maximally distant from previously chosen points. (4) Entropy Adaptive Sampling—Gaussians are prioritized based on language feature entropy, selecting points with the highest uncertainty.", + "bbox": [ + 89, + 595, + 483, + 746 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct experiments on the ScanQA validation set. The results, presented in Table 3, demonstrate that entropy-adaptive sampling outperforms other sampling strategies, followed by FPS. This finding highlights entropy-adaptive sampling as the most effective approach for selecting informative tokens, making it the optimal strategy for downstream 3D VQA tasks.", + "bbox": [ + 89, + 748, + 483, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Length of Visual Tokens. We evaluate the performance of SplatTalk on ScanQA, SQA3D, and MSR3D under two different settings: (1) using only 1 image-worth of visual", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a22b545a6a0f8e7679ec8fc4158821a00592774e6d8c8dbaf430218b51695621.jpg", + "table_caption": [ + "Table 4. Ablation Study on Visual Input Length. We compare model performance using 729 tokens (equivalent to a single image) versus 32,076 tokens (corresponding to 44 images), analyzing the impact of visual context on reasoning." + ], + "table_footnote": [], + "table_body": "
Visual LengthScanQA (val)SQA3D (test)MSR3D (test)
EM@1EM@1-REM@1EM@1-REM@1EM@1-R
729 tokens16.230.725.838.07.917.7
32,076 tokens17.132.226.141.314.123.3
", + "bbox": [ + 537, + 132, + 890, + 194 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tokens (729 tokens) and (2) using the full context length corresponding to 44 images-worth of visual tokens (32,076 tokens). The results, presented in Table 4, evaluate model performance using EM@1 and EM@1-Refined, as these metrics offer a reliable measure of how precisely model predictions align with ground truth answers, both in literal accuracy and permissible variations in phrasing. We employ entropy-adaptive sampling (see Sec. 3.3), which prioritizes high-information tokens and remains deterministic given a fixed set of visual tokens.", + "bbox": [ + 511, + 208, + 903, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A key observation is that performance improved across all datasets as the number of input tokens increased, with MSR3D showing the largest gain, nearly doubling in EM@1. In contrast, ScanQA and SQA3D saw only marginal improvements, even with significantly more tokens. This suggests that additional scene context may not substantially improve performance, possibly due to: (1) Dataset limitations where the tasks may not require rich spatial context; or (2) Model limitations where the model may be unable to effectively leverage extra visual tokens for spatial reasoning.", + "bbox": [ + 511, + 359, + 905, + 511 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 527, + 633, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduce the first method to integrate free-form language features into a 3D Gaussian Splitting framework, enabling zero-shot generalization for 3D Visual Question Answering (3D VQA) without requiring 3D point clouds or depth maps. Our approach surpasses methods that rely solely on 2D language feature maps and achieves performance on par with 3D LMMs that explicitly utilize 3D structural information across multiple 3D VQA benchmarks. We hope this work inspires further research into leveraging multi-view RGB inputs to encapsulate rich semantic representations of 3D scenes, ultimately enhancing 3D VQA capabilities without the reliance on explicit geometric priors.", + "bbox": [ + 511, + 555, + 906, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 511, + 768, + 665, + 786 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We thank the creators of the datasets and codebases used in our work, ScanQA, SQA3D, MSR3D, ScanNet, LLaVA-OV, and FreeSplat, for making their resources publicly available. We are also grateful to the anonymous reviewers for their constructive feedback during the review process. Special thanks to Xiang Li and James M. Rehg for the insightful discussions.", + "bbox": [ + 511, + 794, + 905, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4719", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anthropic. Claude 3. https://www.anthropic.com/index/claude, 2024. Accessed: 2024-02-28. 2, 5", + "[2] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19129-19139, 2022. 5, 7", + "[3] Satanjeev Banerjee and Alon Lavie. METEOR: An Automatic Metric for MT Evaluation with Improved Correlation with Human Judgments. In Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization, pages 65-72, 2005. 7", + "[4] Yash Bhalgat, Iro Laina, Joao F Henriques, Andrew Zisserman, and Andrea Vedaldi. N2f2: Hierarchical scene understanding with nested neural feature fields. In European Conference on Computer Vision, pages 197-214. Springer, 2024. 6", + "[5] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19457-19467, 2024. 3", + "[6] Hanlin Chen, Fangyin Wei, and Gim Hee Lee. Chat-splat: 3d conversational gaussian splatting. arXiv preprint arXiv:2412.00734, 2024. 2, 3, 4, 5", + "[7] Sijin Chen, Xin Chen, Chi Zhang, Mingsheng Li, Gang Yu, Hao Fei, Hongyuan Zhu, Jiayuan Fan, and Tao Chen. L13da: Visual interactive instruction tuning for omni-3d understanding reasoning and planning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26428-26438, 2024. 2, 5", + "[8] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In European Conference on Computer Vision, pages 370-386. Springer, 2024. 3", + "[9] Yilun Chen, Shuai Yang, Haifeng Huang, Tai Wang, Runsen Xu, Ruiyuan Lyu, Dahua Lin, and Jiangmiao Pang. Grounded 3d-llm with referent tokens. arXiv preprint arXiv:2405.10370, 2024. 1, 5", + "[10] Francis Engelmann, Fabian Manhardt, Michael Niemeyer, Keisuke Tateno, Marc Pollefeys, and Federico Tombari. Opennerf: open set 3d neural scene segmentation with pixelwise features and rendered novel views. arXiv preprint arXiv:2404.03650, 2024. 2", + "[11] Gemini Team et al. Gemini: A family of highly capable multimodal models, 2024. 2", + "[12] Zhiwen Fan, Jian Zhang, Wenyan Cong, Peihao Wang, Renjie Li, Kairun Wen, Shijie Zhou, Achuta Kadambi, Zhangyang Wang, Danfei Xu, et al. Large spatial model: End-to-end unposed images to semantic 3d. Advances in Neural Information Processing Systems, 37:40212-40229, 2025. 3", + "[13] Rao Fu, Jingyu Liu, Xilun Chen, Yixin Nie, and Wenhan Xiong. Scene-llm: Extending language model for 3d visual understanding and reasoning. arXiv preprint arXiv:2403.11401, 2024. 1, 5" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the V in VQA Matter: Elevating the Role of Image Understanding in Visual Question Answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6904-6913, 2017. 7", + "[15] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems, 36:20482-20494, 2023. 1, 5", + "[16] Haifeng Huang, Yilun Chen, Zehan Wang, Rongjie Huang, Runsen Xu, Tai Wang, Luping Liu, Xize Cheng, Yang Zhao, Jiangmiao Pang, et al. Chat-scene: Bridging 3d scene and large language models with object identifiers. arXiv preprint arXiv:2312.08168, 2023.5", + "[17] Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. An embodied generalist agent in 3d world. arXiv preprint arXiv:2311.12871, 2023. 1, 2, 5, 6, 7", + "[18] Weitai Kang, Haifeng Huang, Yuzhang Shang, Mubarak Shah, and Yan Yan. Robin3d: Improving 3d large language model via robust instruction tuning, 2025. 2", + "[19] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. 2", + "[20] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 2", + "[21] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 2, 4, 5, 6", + "[22] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava-last-interleave: Tackling multi-image, video, and 3d in large multimodal models, 2024. 5", + "[23] Hajjie Li, Yanmin Wu, Jiarui Meng, Qiankun Gao, Zhiyao Zhang, Ronggang Wang, and Jian Zhang. Instancegaussian: Appearance-semantic joint gaussian representation for 3d instance-level perception. arXiv preprint arXiv:2411.19235, 2024. 3", + "[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 2", + "[25] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 5" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4720", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Chin-Yew Lin. ROUGE: A Package for Automatic Evaluation of Summaries. In Proceedings of the Workshop on Text Summarization Branches Out, pages 74-81, 2004. 7", + "[27] Xiongkun Linghu, Jiangyong Huang, Xuesong Niu, Xiaojian Shawn Ma, Baoxiong Jia, and Siyuan Huang. Multimodal situated reasoning in 3d scenes. Advances in Neural Information Processing Systems, 37:140903-140936, 2025. 1, 6, 7", + "[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36:34892-34916, 2023. 1", + "[29] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. MMBench: Is your multi-modal model an all-around player? In Proceedings of the European Conference on Computer Vision (ECCV), 2024. Accepted as Oral Presentation. 7", + "[30] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, Yaofeng Sun, Chengqi Deng, Hanwei Xu, Zhenda Xie, and Chong Ruan. Deepseek-vl: Towards real-world vision-language understanding, 2024. 2", + "[31] Xiaojian Ma, Silong Yong, Zilong Zheng, Qing Li, Yitao Liang, Song-Chun Zhu, and Siyuan Huang. Sqa3d: Situated question answering in 3d scenes. arXiv preprint arXiv:2210.07474, 2022. 5, 7", + "[32] OpenAI. Gpt-4 with vision (gpt-4v). https://openai.com/research/gpt-4, 2023. Accessed: 2024-02-28. 2, 5", + "[33] Qucheng Peng, Benjamin Planche, Zhongpai Gao, Meng Zheng, Anwesa Choudhuri, Terrence Chen, Chen Chen, and Ziyan Wu. 3d vision-language gaussian splatting. arXiv preprint arXiv:2410.07577, 2024. 5, 6", + "[34] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 815-824, 2023. 1", + "[35] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20051-20060, 2024. 2, 3, 4, 5, 6", + "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 2", + "[37] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13009-13018, 2024. 1", + "[38] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. CIDEr: Consensus-based Image Description Evaluation. In Proceedings of the IEEE Conference on Computer" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Vision and Pattern Recognition (CVPR), pages 4566-4575, 2015. 7", + "[39] Peihao Wang, Zhiwen Fan, Zhangyang Wang, Hao Su, Ravi Ramamoorthi, et al. Lift3d: Zero-shot lifting of any 2d vision model to 3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21367-21377, 2024. 6", + "[40] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Song XiXuan, et al. Cogvlm: Visual expert for pretrained language models. Advances in Neural Information Processing Systems, 37:121475-121499, 2025. 1", + "[41] Xingrui Wang, Wufei Ma, Zhuowan Li, Adam Kortyewski, and Alan L Yuille. 3d-aware visual question answering about parts, poses and occlusions. Advances in Neural Information Processing Systems, 36:58717-58735, 2023. 1", + "[42] Xingrui Wang, Cuiling Lan, Hanxin Zhu, Zhibo Chen, and Yan Lu. Gsemsplat: Generalizable semantic 3d gaussian splatting from uncalibrated image pairs, 2024. 3", + "[43] Yunsong Wang, Tianxin Huang, Hanlin Chen, and Gim Hee Lee. Freesplat: Generalizable 3d gaussian splatting towards free view synthesis of indoor scenes. Advances in Neural Information Processing Systems, 37:107326-107349, 2025. 2, 3, 6", + "[44] Zehan Wang, Haifeng Huang, Yang Zhao, Ziang Zhang, and Zhou Zhao. Chat-3d: Data-efficiently tuning large language model for universal dialogue of 3d scenes. arXiv preprint arXiv:2308.08769, 2023. 5", + "[45] Yanmin Wu, Jiarui Meng, Hajjie Li, Chenming Wu, Yahao Shi, Xinhua Cheng, Chen Zhao, Haocheng Feng, Errui Ding, Jingdong Wang, et al. Opengaussian: Towards point-level 3d gaussian-based open vocabulary understanding. Advances in Neural Information Processing Systems, 37:19114-19138, 2025. 3, 4, 6", + "[46] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6281–6290, 2019. 5", + "[47] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splattering to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676-21685, 2024. 5", + "[48] Chenming Zhu, Tai Wang, Wenwei Zhang, Jiangmiao Pang, and Xihui Liu. Llava-3d: A simple yet effective pathway to empowering lmm's with 3d-awareness. arXiv preprint arXiv:2409.18125, 2024. 1, 2", + "[49] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2911-2921, 2023. 5" + ], + "bbox": [ + 516, + 90, + 906, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4721", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SplatTalk_ 3D VQA with Gaussian Splatting/71c4eb8c-9be8-49e8-bab3-c4048f8e7d2a_model.json b/2025/SplatTalk_ 3D VQA with Gaussian Splatting/71c4eb8c-9be8-49e8-bab3-c4048f8e7d2a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..514230614bdfb783cab02c1a04adac438e6b169c --- /dev/null +++ b/2025/SplatTalk_ 3D VQA with Gaussian Splatting/71c4eb8c-9be8-49e8-bab3-c4048f8e7d2a_model.json @@ -0,0 +1,1892 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.273, + 0.131, + 0.725, + 0.154 + ], + "angle": 0, + "content": "SplatTalk: 3D VQA with Gaussian Splatting" + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.18, + 0.232, + 0.198 + ], + "angle": 0, + "content": "Anh Thai\\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.18, + 0.379, + 0.199 + ], + "angle": 0, + "content": "Songyou Peng²" + }, + { + "type": "text", + "bbox": [ + 0.399, + 0.181, + 0.514, + 0.199 + ], + "angle": 0, + "content": "Kyle Genova2" + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.181, + 0.678, + 0.198 + ], + "angle": 0, + "content": "Leonidas Guibas2" + }, + { + "type": "text", + "bbox": [ + 0.698, + 0.181, + 0.869, + 0.198 + ], + "angle": 0, + "content": "Thomas Funkhouser" + }, + { + "type": "text", + "bbox": [ + 0.27, + 0.205, + 0.531, + 0.224 + ], + "angle": 0, + "content": "1Georgia Institute of Technology" + }, + { + "type": "text", + "bbox": [ + 0.571, + 0.206, + 0.728, + 0.224 + ], + "angle": 0, + "content": "\\(^{2}\\)Google DeepMind" + }, + { + "type": "text", + "bbox": [ + 0.327, + 0.226, + 0.666, + 0.241 + ], + "angle": 0, + "content": "*Work done as a student researcher at Google DeepMind." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.277, + 0.898, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.455, + 0.908, + 0.526 + ], + "angle": 0, + "content": "Figure 1. SplatTalk. We propose a self-supervised 3D-Language Gaussian Splitting model trained from multi-view RGB images. First, images are encoded using a pretrained 2D Vision-Language Model (VLM) and projected into visual-language feature maps via a multimodal projector. These feature maps are then learned within a feed-forward 3D-Language Gaussian Splitting model, producing a 3D-Language Gaussian Field that encodes spatial and semantic information in 3D space. During inference, the 3D Gaussian features are directly queried by a Large Language Model (LLM) to perform 3D question-answering (3D VQA) tasks." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.538, + 0.327, + 0.553 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.597, + 0.486, + 0.868 + ], + "angle": 0, + "content": "Language-guided 3D scene understanding is important for advancing applications in robotics, AR/VR, and human-computer interaction, enabling models to comprehend and interact with 3D environments through natural language. While 2D vision-language models (VLMs) have achieved remarkable success in 2D VQA tasks, progress in the 3D domain has been significantly slower due to the complexity of 3D data and the high cost of manual annotations. In this work, we introduce SplatTalk, a novel method that uses a generalizable 3D Gaussian Splatting (3DGS) framework to produce 3D tokens suitable for direct input into a pretrained LLM, enabling effective zero-shot 3D visual question answering (3D VQA) for scenes with only posed images. During experiments on multiple benchmarks, our approach outperforms both 3D models trained specifically for the task and previous 2D-LMM-based models utilizing only images (our setting), while achieving competitive performance with state-of-the-art 3D LMMs that additionally utilize 3D inputs." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.869, + 0.385, + 0.883 + ], + "angle": 0, + "content": "Project website: https://splat-talk.github.io/" + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.538, + 0.645, + 0.553 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.566, + 0.909, + 0.822 + ], + "angle": 0, + "content": "The rise of Large Multimodal Models (LMMs), following the success of Large Language Models (LLMs), has become increasingly significant in a world where visual understanding is crucial. While 2D LMMs have seen rapid improvements, enabling models to analyze images and answer complex questions [21, 28, 37, 40], progress in 3D LMMs has been slower. This gap is largely due to the scarcity of 3D data compared to 2D datasets. Further, the availability of language-annotated 3D data is even more limited, as annotating 3D scenes since it is significantly more complex and expensive than labeling 2D images. These challenges have hindered the development of language-guided 3D reasoning, making 3D LMMs an important yet underexplored area in multimodal research. In this work, we focus on 3D Visual Question Answering (3D VQA) as it provides a strong benchmark for evaluating a model's ability to understand and reason about 3D scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Recently, significant progress has been made in 3D multimodal learning, particularly for indoor scenes [15, 17, 27, 48]. Several approaches rely on posed images and 3D inputs, like meshes [41], point clouds [9, 17, 34] and voxels [13], to associate 3D geometry with language features. As such," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4712" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "they have limited applicability in scenarios where a full 3D surface reconstruction is not available." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.124, + 0.483, + 0.32 + ], + "angle": 0, + "content": "Previous works have explored integrating semantic features into NeRF [19] or 3D Gaussian Splatting (3DGS) [35], which rely only upon RGB image inputs. However, research on effectively associating 3D geometry with language features remains limited. Some methods align individual 3D locations with language features via implicit functions optimized individually for every scene [10, 19]. Others employ SAM [20] to first segment objects and object parts, followed by CLIP [36] to extract object-centric features from the generated masks. These approaches allow models to query information about individual points and/or objects. However, it disregards spatial relationships between objects, making it unsuitable for 3D VQA spatial reasoning tasks." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.323, + 0.486, + 0.534 + ], + "angle": 0, + "content": "More recent work [21] has demonstrated the power of LLMs to answer VQA questions using only a sampling of tokens extracted from images, without any underlying 3D representation. While this approach is inspiring, it relies upon the LLM to make sense of a set of tokens that are arranged according to images rather than elements in the 3D scene. As a result, it struggles to answer queries about specific 3D locations (e.g., \"What is here?\") and questions requiring a holistic understanding of the entire 3D scene, where the answer depends on establishing correspondences between objects across multiple images (e.g., \"What is opposite the door at the other end of the room\"?). In this work, we aim to address these issues by consolidating tokens in 3D before providing them to the LLM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.537, + 0.483, + 0.717 + ], + "angle": 0, + "content": "Our approach is called SplatTalk. Given a set of posed RGB images, we integrate language features extracted from the RGB images to build a 3D Gaussian Splitting representation, from which 3D tokens can be extracted for direct input into a pretrained LLM to answer challenging 3D VQA queries (see Fig. 1). The main research contributions are the new methods for 1) integrating language features into the Gaussian representation, and 2) extracting tokens that can be provided directly to an LLM. These methods avoid the issues of concurrent work, ChatSplat [6], which produces tokens that are not directly compatible with an LLM, making language-based reasoning more difficult." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Overall, SplatTalk has a unique combination of three advantages. First, it does not require any 3D surface representation as input. Second, it leverages the full power of a pre-trained LLM to answer difficult 3D VQA questions. Third, it produces a concise 3D representation that can be queried and/or edited in 3D (e.g., cropped to a 3D region, rendered from a novel view, etc.). During experiments with 3D VQA benchmarks, we find that it outperforms 3D models trained only for the benchmark tasks, demonstrating the value of utilizing pre-trained LLMs. SplatTalk further outperforms prior image-only methods using LLMs, showing the value of building a 3D scene representation. Finally," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "it even achieves competitive performance with SOTA 3D LMMs that utilize 3D scene representation inputs." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.123, + 0.792, + 0.137 + ], + "angle": 0, + "content": "In summary, our key contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.138, + 0.907, + 0.197 + ], + "angle": 0, + "content": "- We introduce the first self-supervised 3D Gaussian-based method for large-scale zero-shot 3D VQA, SplatTalk, requiring only multi-view images without explicit depth, point clouds or 3D-language annotated supervision." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.198, + 0.905, + 0.242 + ], + "angle": 0, + "content": "- We show that mean feature extraction from 3D Gaussians encodes holistic scene concepts for 3D VQA, supported by theoretical justification." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.243, + 0.905, + 0.272 + ], + "angle": 0, + "content": "- We introduce entropy-adaptive token sampling, leading to better 3D VQA performance without additional training." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.273, + 0.905, + 0.333 + ], + "angle": 0, + "content": "- We demonstrate that SplatTalk outperforms 2D LMMs and achieves performance on par with 3D models using point clouds on various 3D VQA benchmarks, demonstrating the effectiveness of our 3D-aware representations." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.138, + 0.907, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.347, + 0.655, + 0.362 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.907, + 0.599 + ], + "angle": 0, + "content": "2D LMMs. With advancements in large language models (LLMs), the ability of models to comprehend images and respond to language prompts related to their content has improved significantly. Notably, SOTA models [1, 11, 21, 30, 32] have demonstrated remarkable capabilities in processing multiple images of a 3D scene and answering questions about the shared content across these images. Although multiple RGB images inherently contain 3D information, they do not explicitly model 3D understanding, potentially leading to sub-optimal performance on tasks that require explicit 3D reasoning. In this work, we build upon the strong performance of 2D vision-language models (2D-VLMs), particularly LLaVA-OneVision [21], and integrate them with the 3D priors provided by the 3DGS framework to improve 3D spatial understanding capabilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.909, + 0.901 + ], + "angle": 0, + "content": "3D LMMs. Recent advancements in 3D indoor scene understanding have explored integrating 3D priors into LLMs using inputs such as 3D point clouds [7, 17] or multi-view RGB images combined with depth or 3D point clouds [48]. These approaches employ various strategies to embed 3D structural information and language features into a shared representation space. While most methods [17, 48] fine-tune pretrained visual encoders (both 2D and 3D) alongside LoRA parameters or even the full set of LLM parameters, others [18, 24] opt for a more lightweight approach, training only a projection layer to map visual features into the language feature space. In contrast, our method does not require any explicit depth or 3D inputs. Instead, we leverage the 3DGS framework, which has shown strong capabilities in representing 3D scenes using only multi-view RGB images. Generalizable 3D Gaussian Splatting. This body of work explores generalizable 3DGS as an alternative to the perscene optimization paradigm used in traditional 3DGS methods, aiming to improve novel view synthesis across unseen scenes. Our method builds on FreeSplat [43], which is" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4713" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.09, + 0.67, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.09, + 0.903, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.31, + 0.907, + 0.409 + ], + "angle": 0, + "content": "Figure 2. Overview. Left: During the self-supervised 3D-language Gaussian Splitting training phase, multiple RGB input views are first encoded into Gaussian latent features (Gaussian triplets). These latent features are then decoded into Gaussian parameters for rendering, along with a low-dimensional visual-language feature \\( f \\). To ensure proper supervision of this low-dimensional feature, we train an autoencoder that maps the high-dimensional, unbounded features obtained from LLaVA-OV, specifically, the visual tokens serving as direct inputs to the LLM, onto a low-dimensional hypersphere space. Right: During 3D VQA inference, visual-language features are directly extracted from the 3D Gaussians. These features are then mapped back to the original high-dimensional space using the pretrained decoder and subsequently used as direct visual token inputs to the LLM. LoRA fine-tuning of the LLM is optional." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.416, + 0.484, + 0.642 + ], + "angle": 0, + "content": "designed for sparse-view synthesis in indoor scenes. The method begins by extracting multi-scale features from input RGB images using CNN backbones. It then constructs adaptive cost volumes from neighboring views to predict depth maps. These predicted depth maps are used to backproject 2D CNN-encoded features into 3D Gaussian triplets \\((\\mu, \\omega, f)\\) where \\(\\mu\\) denotes the Gaussian's center, \\(\\omega\\) the weight, and \\(f\\) the latent feature. To integrate multi-view information, FreeSplat introduces a Pixel-wise Triplet Fusion (PTF) module that uses a lightweight GRU to align and fuse overlapping 3D Gaussians across views efficiently. Finally, the fused triplets are passed through an MLP to decode rendering parameters, \\(\\Sigma, \\alpha, s\\). Our approach extends the FreeSplat framework to enable the learning of generalizable 3D Gaussian language fields." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Similar to FreeSplat [43], [5, 8] leverage multi-view stereo (MVS) cues to refine depth estimation and improve geometric consistency in 3D Gaussian representations. These methods primarily focus on RGB novel view synthesis from sparse inputs (2-10 views). Our work builds on the SOTA model, FreeSplat [43], but diverges significantly in both objectives and methodology. First, while previous methods focus on photorealistic RGB reconstruction, we aim at semantic 3D scene understanding. Second, instead of reconstructing only a localized portion of the scene as shown in these works, our goal is to represent entire 3D environments, capturing both spatial and semantic relationships holistically. Language-guided 3D Gaussian Splatting. LangSplat [35] is one of the first approaches that integrates semantic features into 3DGS. However, this method requires training on individual scenes, which is significantly expensive, and lacks generalization to unseen scenes. Similarly, follow-up" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.907, + 0.567 + ], + "angle": 0, + "content": "works [23, 45] explore 3DGS for open-vocabulary object segmentation but do not focus on reasoning tasks like 3D VQA. Further, they primarily rely on scene-specific training, limiting their scalability. In contrast, we propose a fully self-supervised framework that learns 3D-aware visual representations without requiring per-scene training. While some works [12, 42] explores a generalizable 3DGS and semantic modeling approach, they primarily focus on open-vocabulary segmentation, without addressing the broader challenges of free-form language reasoning required for 3D VQA." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.909, + 0.751 + ], + "angle": 0, + "content": "ChatSplat [6] is a concurrent work that explores natural language conversation within the 3DGS framework. However, it optimizes on a per-scene basis and requires rendering 3D features into 2D feature maps during inference. Additionally, its training pipeline is not fully end-to-end, as it first trains the RGB Gaussians separately before incorporating language features. In contrast, our approach leverages scene latent features that encapsulate both RGB and semantic information, enabling the joint optimization of both modalities in a fully end-to-end manner. Our approach extends the FreeSplat framework to enable the learning of generalizable 3D Gaussian language fields." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.767, + 0.605, + 0.783 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.794, + 0.906, + 0.884 + ], + "angle": 0, + "content": "The problem we address is formulated as follows: Given a sequence of \\( N \\) posed RGB images captured from a 3D indoor scene, along with language prompts related to the scene's content, the goal is to generate the corresponding language responses. That is, we address 3D VQA tasks with solely posed RGB images as inputs." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.906, + 0.902 + ], + "angle": 0, + "content": "We provide an overview of our approach in Fig. 2, which" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4714" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.093, + 0.49, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.092, + 0.895, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.34, + 0.908, + 0.412 + ], + "angle": 0, + "content": "Figure 3. Qualitative results on ScanQA, scene0030_00 and scene0084_00. We compare responses from LLaVA-OV, our model (Ours), and the ground truth (GT) for spatial reasoning questions in 3D VQA. Each scene highlights the referenced objects with red circles, and key relative objects are marked in blue. The answers from each model are displayed in color-coded speech bubbles: LLaVA-OV (brown), Ours (blue), and GT (green). With its 3D-aware representation, our model exhibits improved spatial reasoning capabilities, accurately identifying relationships between objects that are far apart and may never co-occur in the same image (e.g., door and window). More examples in Supp." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.418, + 0.486, + 0.66 + ], + "angle": 0, + "content": "consists of three main steps: 1) Training the feature autoencoder, 2) Self-supervised 3D-Language Gaussian Splitting Training and 3) 3D VQA Inference. First, we extract the high-dimensional language features from LLaVA-OV [21] for each RGB image by using the pretrained image encoder and multimodal projector. Since these features are high-dimensional, unbounded and sparse, we train an autoencoder to project them into a more compact, low-dimensional space. These low-dimensional features then serve as pseudo groundtruth for training the language representations in SplatTalk. SplatTalk architecture is a feed-forward network consisting of a Gaussian encoder and decoder that predicts Gaussian parameters for rendering. During 3D VQA inference stage, we extract the trained language features from the 3D Gaussians and directly feed them into the LLM, along with the language prompt, to generate free-form responses." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.668, + 0.367, + 0.684 + ], + "angle": 0, + "content": "3.1. Integrating Language Features" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.486, + 0.825 + ], + "angle": 0, + "content": "An important step of our computational pipeline is the self-supervised method to integrate language features into a generalizable 3DGS framework. We identify four key factors that shape our approach for this step: (1) Leveraging features that align with the LLM's input space for direct language reasoning, (2) Efficient feature dimensionality reduction for effective CUDA-based rendering, and (3) Joint training of RGB and semantic features for coherent multimodal representation learning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Visual Tokens for 2D Pseudo Ground-truth Features. We extract the visual tokens used as inputs to the LLM for training our 3DGS framework. Our empirical findings indicate that extracting features directly from the visual encoder for training 3DGS requires retraining the multimodal projector" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.909, + 0.63 + ], + "angle": 0, + "content": "layer which aligns the visual with language spaces. This observation is further corroborated by [6], which demonstrates that features extracted before the multimodal projector cannot be seamlessly integrated into the language model. A possible explanation is that the image encoder's native feature space (before the projector) is high-dimensional and unbounded, leading to sparser features with greater variance across different dimensions. As a result, the projector becomes more sensitive to subtle variations in the input features, making it difficult to directly map them into a token representation that aligns with the LLM's understanding. In contrast, the visual tokens generated after the projector are already well-aligned with the language model's latent space, allowing the LLM to interpret and reason with them directly." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Feature Dimensionality Reduction. Building on the previous factor, directly optimizing these language embeddings within a differentiable rendering pipeline is challenging because they are extremely high-dimensional (e.g., 3584 dimensions), sparse, and contain unbounded values, making learning both unstable and impractical. Therefore, we learn a generalizable autoencoder that projects these sparse, high-dimensional visual tokens into a compact unit hypersphere with much lower dimension (e.g., 256 dimensions), ensuring feature smoothness. Note that the language representations from LLaVA, which support reasoning and free-form language comprehension, are significantly more complex than those from CLIP or segmentation-based models like SAM. Different from prior works that compress semantic features into extremely low-dimensional vectors (e.g., 3-16 dimensions [35, 45]), which simplifies rendering but results in significant information loss, our approach prior" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4715" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.16 + ], + "angle": 0, + "content": "Table 1. Evaluation of 3D VQA on ScanQA [2] and SQA3D [31] datasets. We compare SplatTalk against various baselines, including 3D LMMs, which take point clouds (PC) or both point clouds and images \\((\\mathrm{PC} + \\mathrm{I})\\) as visual inputs, and 2D LMM-based models, which rely solely on images (I). As a 2D LMM-based approach, SplatTalk outperforms other 2D LMM-based models, surpasses some 3D LMMs, and achieves competitive performance against the strongest 3D LMM baselines. Our models are highlighted in blue, bold indicates best performance among the 2D-LMM methods." + }, + { + "type": "table", + "bbox": [ + 0.191, + 0.171, + 0.805, + 0.613 + ], + "angle": 0, + "content": "
MethodModalityScanQA (val)SQA3D (test)
CIDErMETEORROUGEEM@1EM@1-REM@1EM@1-R
Specialist Models
ScanQA [2]PC64.913.133.121.1-47.2-
ScanRefer+MCan [46]PC55.411.530.018.6---
3D-VisTA [49]PC69.613.935.722.4-48.5-
SQA3D [31]PC-----46.6-
Generalist 3D LMMs
Chat-Scene [16]PC+I87.718.041.621.6-54.657.5
Scene-LLM [13]PC+I80.015.835.9--53.6-
Grounded 3D-LLM [9]PC72.7------
Chat-3D-v2 [44]PC77.116.140.1--54.7-
Finetuned 3D LMMs
3D-LLM [15]PC+I69.414.535.720.5---
LEO [17]PC+I101.420.049.224.547.650.052.4
Scene-LLM [13]PC+I80.016.640.027.2-54.2-
LL3DA [7]PC76.815.937.3----
Chat-3D [44]PC53.211.928.5----
2D LMM-Based
VideoChat2 [25]I49.29.528.219.2-37.3-
GPT-4V [32]I59.613.533.4----
Claude [1]I57.710.029.3----
LLaVA-NeXT-Video [22]I46.29.127.8--34.2-
LLaVA-OV [21]I50.013.029.415.632.525.435.8
SplatTalkI61.714.232.717.132.226.141.3
SplatTalk-ScanQA-FTI77.115.438.722.338.342.044.7
SplatTalk-3DVQA-FTI77.515.638.522.438.347.649.4
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.625, + 0.483, + 0.716 + ], + "angle": 0, + "content": "itizes preserving rich language features by maintaining a higher-dimensional representation (e.g., 256). To efficiently render high-dimensional feature maps on CUDA, we draw inspiration from [47] and implement a parallel CUDA rasterizer pipeline, where semantic features and RGB are rendered simultaneously, using shared Gaussian parameters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Unlike previous methods [6, 35] that train a separate autoencoder for each individual scene, we train a single autoencoder across all training scenes, each consisting of multiple frames. This approach improves the autoencoder's generalizability, reducing computational overhead by eliminating the need for retraining to compress features from unseen scenes. Training the autoencoder separately from the 3DGS model ensures that it is dedicated solely to compressing high-dimensional feature vectors while preserving the essential scene information without being influenced by the complexities of the rendering pipeline. Our empirical findings further indicate that jointly training a feature de" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.625, + 0.907, + 0.671 + ], + "angle": 0, + "content": "coder to project low-dimensional embeddings back to the high-dimensional space introduces significant instabilities in the training process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Joint-training RGB and Language. In contrast to previous approaches that first train RGB Gaussians and then freeze them before integrating language features [6, 35], we adopt a joint training strategy, where RGB and language features are learned simultaneously. This is due to the fact that visual RGB and semantic features play a complementary role in effectively representing 3D scenes [33]. In feedforward models like FreeSplat, where the scene is first encoded into a Gaussian latent space before being decoded into specific Gaussian parameters for RGB rendering, these Gaussian latent features inherently capture holistic scene information, including both appearance (RGB) and semantics. Based on this observation, we introduce a new Gaussian parameter decoder head that predicts semantic features for each 3D Gaussian alongside rendering parameters, enabling joint op-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4716" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 2. Evaluation of 3D VQA on MSR3D [27]. We compare SplatTalk against various baselines, all evaluated in a zero-shot setting. T denotes models that rely solely on text inputs without interleaving images, PC indicates models that use point cloud as the visual input, and I represents models that process only image-based inputs. Our approach, highlighted in blue, outperforms all baselines." + }, + { + "type": "table", + "bbox": [ + 0.154, + 0.143, + 0.842, + 0.248 + ], + "angle": 0, + "content": "
ModelInputSceneSettingCountingExistenceAttributesSpatialNavigationOthersOverall
LEO [17]TPCzero-shot0.8015.511.87.32.315.37.8
LLaVA-OV [21]TIzero-shot18.631.224.819.516.736.324.0
SplatTalkTIzero-shot19.660.344.035.835.561.841.8
SplatTalk-ScanQA-FTTIzero-shot28.966.043.228.333.860.041.5
SplatTalk-3DVQA-FTTIzero-shot24.357.036.17.816.435.026.5
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.259, + 0.482, + 0.351 + ], + "angle": 0, + "content": "tization. Unlike approaches that explicitly model these interactions [33, 39] using extra cross-attention or projection layers, SplatTalk implicitly integrates semantic and visual information within the Gaussian representation. This paradigm naturally facilitates multimodal interactions between RGB and language features." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.353, + 0.483, + 0.457 + ], + "angle": 0, + "content": "To incorporate more views and comprehensively cover the scene, we reduce the input resolution of the RGB images. This adjustment aligns well with the inherently low spatial resolution of the visual tokens extracted from LLaVA-OV, and also prevents the model from disproportionately allocating resources to modeling RGB features, thus encouraging a balanced representation of semantic and RGB information." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.46, + 0.483, + 0.596 + ], + "angle": 0, + "content": "These design choices establish our approach as a fundamentally novel method for integrating language features into 3DGS, grounded in first-principles observations rather than heuristic adaptations. By addressing the core challenges of multimodal alignment, feature compression, and joint optimization, our approach is not merely a trivial extension of existing works, but a carefully designed framework that ensures both effectiveness and stability in learning rich, spatially-aware semantic representations." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.612, + 0.387, + 0.629 + ], + "angle": 0, + "content": "3.2. Extracting 3D Language Features" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.635, + 0.486, + 0.902 + ], + "angle": 0, + "content": "During inference, we directly extract language features at the mean position of each 3D Gaussian. Although the rendering process incorporates each Gaussian's covariance and opacity when projecting 3D features to 2D, we demonstrate that scene semantics can be effectively captured by the mean feature of each Gaussian, \\( \\pmb{f}_i \\), alone, without relying on these additional factors. This is supported by the conceptual similarity between our training optimization process and the Expectation-Maximization (EM) algorithm. For simplification, let \\( \\mathcal{L} = \\sum_{t=1}^{T} \\| \\pmb{F}_t(x) - \\pmb{F}_t^{gt}(x) \\|^2 \\) be the loss, where \\( \\pmb{F}_t(x) \\) is the predicted rendered feature map of frame \\( t \\) at pixel \\( x \\) and \\( \\pmb{F}^{gt} \\) is the pseudo ground-truth feature map obtained from LLaVA-OV. Replacing \\( \\pmb{F}_t(x) \\) by the rendering equation and solving for the optimal \\( \\pmb{f}_i^* \\), we obtain \\( \\pmb{f}_i^* = \\frac{\\sum_{t,x} R_i(t,x) \\pmb{F}_t^{gt}(x)}{\\sum_{t,x} R_i(t,x)} \\) where \\( R_i(t,x) = \\frac{\\alpha_i \\mathcal{N}(x; \\pmb{\\mu}_i, \\pmb{\\Sigma}_i)}{\\sum_j \\alpha_j \\mathcal{N}(x; \\pmb{\\mu}_j, \\pmb{\\Sigma}_j)} \\) is the Gaussian's contribution to pixel \\( x \\) at viewpoint \\( t \\). This mirrors the EM procedure: estimating \\( R \\) by rendering (E-step)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.26, + 0.907, + 0.381 + ], + "angle": 0, + "content": "and updating \\( f_{i} \\) by minimizing reconstruction loss (M-step). Consequently, the optimal 3D feature \\( \\pmb{f}_{i}^{*} \\) for each Gaussian is a weighted sum of the 2D feature maps, and the set of 3D Gaussian features \\( \\{\\pmb{f}_{i}\\}_{i=1}^{N} \\) encodes the scene's semantic holistically from \\( T \\) 2D inputs. These features are then projected back into the original language embedding space using the pretrained decoder and directly provided as inputs to the LLM for downstream 3D-language tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.382, + 0.909, + 0.563 + ], + "angle": 0, + "content": "In open-vocabulary segmentation, querying Gaussian centers is effective [4, 35, 45] because each Gaussian is assumed to represent a distinct, meaningful object or region. These methods use object-centric embeddings (e.g., SAM-masked CLIP or LSeg) and retrieve objects via cosine similarity to text queries. In contrast, 3D reasoning tasks lack this one-to-one correspondence. Each Gaussian encodes parts of a scene-level representation, where meaning stems from spatial and relational structure. Querying single Gaussians is thus insufficient for context-dependent reasoning. While our method also queries at Gaussian means, this is not trivial; our EM-based formulation offers a principled justification." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.575, + 0.694, + 0.591 + ], + "angle": 0, + "content": "3.3.3D VQA Inference" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.907, + 0.718 + ], + "angle": 0, + "content": "Entropy Adaptive Gaussian Sampling. SplatTalk leverages triplet fusion adopted in FreeSplat [43], which reduces the number of Gaussians. However, complex scenes still require significantly more Gaussians than the LLM's visual token limit, and not all are equally informative. To address this, we propose entropy-based sampling: selecting the top \\( k \\) Gaussians with the highest language feature entropy, where \\( k \\) matches the LLM's visual token capacity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "3D Tokens as Visual Inputs. After 3DGS training, each 3D Gaussian encodes a feature that is a weighted aggregation of 2D semantic information across all input views (Sec. 3.2). These features are also spatially grounded in their 3D coordinates. Because each token corresponds to a specific location in 3D space, its position is already implicitly encoded in the feature through the rendering and fusion process. Unlike 2D patches in a fixed grid, these 3D tokens form an unordered set, similar to a point cloud where spatial relationships are captured by the features themselves, not by sequence position. Therefore, we do not require explicit positional encodings or a fixed input order. The 3D tokens," + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4717" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.123 + ], + "angle": 0, + "content": "extracted by sampling features at each Gaussian's mean, are directly fed into the LLM." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.132, + 0.458, + 0.149 + ], + "angle": 0, + "content": "3.4. Optional - Finetuning on 3D VQA Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.154, + 0.483, + 0.261 + ], + "angle": 0, + "content": "While SplatTalk is capable of performing the 3D VQA task in a zero-shot manner, it also allows for flexible fine-tuning on 3D VQA datasets to enhance alignment and reasoning capabilities. Rather than training the model end-to-end, which is both computationally expensive and prone to overfitting on limited datasets, we optimize only the LoRA parameters applied to the LLM during fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.275, + 0.223, + 0.292 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.3, + 0.308, + 0.316 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.484, + 0.488 + ], + "angle": 0, + "content": "We train SplatTalk with 500 scenes from the ScanQA training set. The initial self-supervised stage focuses purely on learning the 3D scene representation, without leveraging any text-annotated information. For each scene, we uniformly sample 100 views around the scene and extract their per-patch features by passing them through the visual encoder of LLaVA-OV followed by the multi-modal projector. The subsequent SplatTalk model is trained on a single NVIDIA H100 (80GB) GPU. Both the LoRA fine-tuning and inference are also conducted on a single H100. More details can be found in Supp. Material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.498, + 0.289, + 0.513 + ], + "angle": 0, + "content": "4.2. Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.521, + 0.484, + 0.626 + ], + "angle": 0, + "content": "Evaluation Dataset. We evaluate our method on ScanNet-based QA datasets: ScanQA [2], SQA3D [31], and MSR3D [27]. ScanQA is a benchmark designed for question answering in 3D indoor scenes, containing detailed spatial and semantic information extracted from ScanNet. SQA3D and MSR3D further extend this by introducing scenarios that describe an embodied agent's 3D position and state." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.484, + 0.718 + ], + "angle": 0, + "content": "Metrics. For ScanQA and SQA3D, we evaluate performance using standard n-gram-based metrics, specifically CIDEr [38], METEOR [3], ROUGE [26], EM@1 [14], and EM@1-Refined for ScanQA, and EM@1 and EM@1-Refined for SQA3D. In contrast, for MSR3D, we utilize the LLM-Match score [29] as recommended by the authors." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.483, + 0.744 + ], + "angle": 0, + "content": "4.3. Performance on 3D Question Answering Tasks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "ScanQA. We evaluate SplatTalk's performance against several model families, each representing different levels of specialization and generalization. \"Specialist Models\" refer to models specifically trained to solve ScanQA without exposure to other tasks. \"Generalist 3D LMMs\" consist of models trained across a diverse range of 3D vision-language tasks, enabling broader generalization. \"Finetuned 3D LMMs\" are models that have been further fine-tuned on the ScanQA training set to optimize performance for this particular dataset. Lastly, \"2D LMM-Based Models\" include" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "models that process only image inputs without leveraging depth or explicit 3D information. SplatTalk and its fine-tuned variants fall within this category." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.138, + 0.907, + 0.319 + ], + "angle": 0, + "content": "We evaluate three different variants of our model on this task. The base model (SplatTalk) is trained solely to encode the semantic representation of the scene from multiple RGB images using 3DGS, without exposure to question-answering tasks. Note that this model is trained entirely in a self-supervised manner, without any access to language supervision or textual annotations during training. To further adapt it to downstream 3D VQA tasks, we fine-tune this base model on the ScanQA training set, referring to it as SplatTalk-ScanQA-FT, and on the combined training sets of ScanQA and SQA3D, denoting as SplatTalk-3DVQA-FT. We present results in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.32, + 0.909, + 0.562 + ], + "angle": 0, + "content": "Compared to LLaVA-OV, whose 2D language feature maps are used to train SplatTalk, our base model, which benefits from a more 3D-aware representation due to the 3D Gaussian training process, achieves significantly higher performance (e.g. 61.7 vs 50.0 on CIDEr). Our language finetuned model variants, SplatTalk-ScanQA-FT and SplatTalk-3DVQA-FT, improve performance further. Notably, despite not being trained with depth supervision or explicit 3D information, our models achieve comparable results to both Generalist 3D LMMs and Finetuned 3D LMMs, while outperforming Specialist Models. Additionally, among 2D LMM-Based approaches, our base model without being fine-tuned with 3D VQA datasets outperforms other methods, with significantly higher performance on some metrics, including both open-source and closed-source models. We show qualitative results in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.563, + 0.909, + 0.729 + ], + "angle": 0, + "content": "SQA3D. Similar to our evaluation on ScanQA, we compare our model against both 3D LMMs and 2D LMMs. Compared to 2D LLaVA-OV features, our base model with 3D-aware features achieves a significant performance gain, highlighting the advantage of incorporating 3D information via 3DGS which aligns with our findings on ScanQA. Further finetuning on ScanQA alone and on the combined ScanQA + SQA3D dataset leads to substantial improvements. Notably, SplatTalk-3DVQA-FT surpasses specialist models that leverage 3D point cloud inputs, demonstrating the strength of our approach in 3D scene understanding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.73, + 0.909, + 0.851 + ], + "angle": 0, + "content": "MSR3D. We evaluated our model's performance against LLaVA-OV and \\(\\mathsf{LEO}^1\\) [17] on the Scannet split of MSR3D. Importantly, all comparisons are conducted in a zero-shot setting, as none of the models have been trained on MSR3D. This is an interleaved dataset where inputs consist of both text and images. However, since all models are tested in a zero-shot manner, they rely solely on text inputs, where the agent's state descriptions and the question are combined into" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.863, + 0.908, + 0.901 + ], + "angle": 0, + "content": "1The reported number reflects performance across all splits of MSR3D. However, since LEO was trained on ScanNet and 3RScan, its performance remains largely consistent across splits." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4718" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.082, + 0.503, + 0.125 + ], + "angle": 0, + "content": "Table 3. Ablation Study on Gaussian Sampling Strategies. Our entropy-based sampling strategy consistently outperforms across multiple metrics." + }, + { + "type": "table", + "bbox": [ + 0.166, + 0.125, + 0.427, + 0.19 + ], + "angle": 0, + "content": "
Sampling MethodEM@1METEORROUGESPICEAvg
Random17.013.862.114.026.73
Point Density16.214.059.714.126.00
FPS17.713.862.013.626.78
Entropy17.114.661.714.226.90
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.209, + 0.483, + 0.255 + ], + "angle": 0, + "content": "a single text prompt. On the visual processing side, LEO operates on 3D point cloud inputs, whereas both LLaVA-OV and our model rely only on RGB images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.256, + 0.484, + 0.36 + ], + "angle": 0, + "content": "The results, presented in Table 2, demonstrate that our base model outperforms LLaVA-OV across all question types. Notably, aside from the \"counting\" category, our base model achieves approximately twice the performance of LLaVA-OV on other question types. Both models significantly outperform LEO, which struggles due to the absence of fine-tuning on this dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.363, + 0.484, + 0.558 + ], + "angle": 0, + "content": "Interestingly, we observe a slight performance drop when comparing SplatTalk with SplatTalk-ScanQA-FT. However, this drop becomes much more pronounced when compared with SplatTalk-3DVQA-FT. A plausible explanation is that fine-tuning on other QA datasets introduces overfitting, causing the model to adapt too specifically to those datasets. Furthermore, the observed performance drop suggests that MSR3D differs significantly from ScanQA and SQA3D, making direct knowledge transfer less effective. This distinction may also explain why LEO struggles to generalize in a zero-shot setting. Despite being trained on multiple QA datasets, none of them closely resemble MSR3D, limiting its ability to adapt without fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.573, + 0.255, + 0.587 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.596, + 0.484, + 0.747 + ], + "angle": 0, + "content": "Gaussian Sampling Strategies. We explore various Gaussian sampling strategies for selecting tokens as inputs to the LLM, including: (1) Random Sampling—Gaussians are selected randomly across the scene. (2) Point Density Adaptive Sampling—Gaussians are sampled proportionally to the density of 3D points. (3) Furthest Point Sampling (FPS)—Gaussians are iteratively selected at locations maximally distant from previously chosen points. (4) Entropy Adaptive Sampling—Gaussians are prioritized based on language feature entropy, selecting points with the highest uncertainty." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.749, + 0.484, + 0.854 + ], + "angle": 0, + "content": "We conduct experiments on the ScanQA validation set. The results, presented in Table 3, demonstrate that entropy-adaptive sampling outperforms other sampling strategies, followed by FPS. This finding highlights entropy-adaptive sampling as the most effective approach for selecting informative tokens, making it the optimal strategy for downstream 3D VQA tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Length of Visual Tokens. We evaluate the performance of SplatTalk on ScanQA, SQA3D, and MSR3D under two different settings: (1) using only 1 image-worth of visual" + }, + { + "type": "table_caption", + "bbox": [ + 0.509, + 0.077, + 0.921, + 0.132 + ], + "angle": 0, + "content": "Table 4. Ablation Study on Visual Input Length. We compare model performance using 729 tokens (equivalent to a single image) versus 32,076 tokens (corresponding to 44 images), analyzing the impact of visual context on reasoning." + }, + { + "type": "table", + "bbox": [ + 0.538, + 0.133, + 0.892, + 0.195 + ], + "angle": 0, + "content": "
Visual LengthScanQA (val)SQA3D (test)MSR3D (test)
EM@1EM@1-REM@1EM@1-REM@1EM@1-R
729 tokens16.230.725.838.07.917.7
32,076 tokens17.132.226.141.314.123.3
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.209, + 0.905, + 0.358 + ], + "angle": 0, + "content": "tokens (729 tokens) and (2) using the full context length corresponding to 44 images-worth of visual tokens (32,076 tokens). The results, presented in Table 4, evaluate model performance using EM@1 and EM@1-Refined, as these metrics offer a reliable measure of how precisely model predictions align with ground truth answers, both in literal accuracy and permissible variations in phrasing. We employ entropy-adaptive sampling (see Sec. 3.3), which prioritizes high-information tokens and remains deterministic given a fixed set of visual tokens." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.361, + 0.906, + 0.512 + ], + "angle": 0, + "content": "A key observation is that performance improved across all datasets as the number of input tokens increased, with MSR3D showing the largest gain, nearly doubling in EM@1. In contrast, ScanQA and SQA3D saw only marginal improvements, even with significantly more tokens. This suggests that additional scene context may not substantially improve performance, possibly due to: (1) Dataset limitations where the tasks may not require rich spatial context; or (2) Model limitations where the model may be unable to effectively leverage extra visual tokens for spatial reasoning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.529, + 0.634, + 0.545 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.556, + 0.907, + 0.737 + ], + "angle": 0, + "content": "In this work, we introduce the first method to integrate free-form language features into a 3D Gaussian Splitting framework, enabling zero-shot generalization for 3D Visual Question Answering (3D VQA) without requiring 3D point clouds or depth maps. Our approach surpasses methods that rely solely on 2D language feature maps and achieves performance on par with 3D LMMs that explicitly utilize 3D structural information across multiple 3D VQA benchmarks. We hope this work inspires further research into leveraging multi-view RGB inputs to encapsulate rich semantic representations of 3D scenes, ultimately enhancing 3D VQA capabilities without the reliance on explicit geometric priors." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.769, + 0.666, + 0.787 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.9 + ], + "angle": 0, + "content": "We thank the creators of the datasets and codebases used in our work, ScanQA, SQA3D, MSR3D, ScanNet, LLaVA-OV, and FreeSplat, for making their resources publicly available. We are also grateful to the anonymous reviewers for their constructive feedback during the review process. Special thanks to Xiang Li and James M. Rehg for the insightful discussions." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4719" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.142 + ], + "angle": 0, + "content": "[1] Anthropic. Claude 3. https://www.anthropic.com/index/claude, 2024. Accessed: 2024-02-28. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.145, + 0.484, + 0.212 + ], + "angle": 0, + "content": "[2] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19129-19139, 2022. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.215, + 0.483, + 0.284 + ], + "angle": 0, + "content": "[3] Satanjeev Banerjee and Alon Lavie. METEOR: An Automatic Metric for MT Evaluation with Improved Correlation with Human Judgments. In Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization, pages 65-72, 2005. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.285, + 0.484, + 0.352 + ], + "angle": 0, + "content": "[4] Yash Bhalgat, Iro Laina, Joao F Henriques, Andrew Zisserman, and Andrea Vedaldi. N2f2: Hierarchical scene understanding with nested neural feature fields. In European Conference on Computer Vision, pages 197-214. Springer, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.355, + 0.484, + 0.424 + ], + "angle": 0, + "content": "[5] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19457-19467, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.425, + 0.484, + 0.465 + ], + "angle": 0, + "content": "[6] Hanlin Chen, Fangyin Wei, and Gim Hee Lee. Chat-splat: 3d conversational gaussian splatting. arXiv preprint arXiv:2412.00734, 2024. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.467, + 0.484, + 0.55 + ], + "angle": 0, + "content": "[7] Sijin Chen, Xin Chen, Chi Zhang, Mingsheng Li, Gang Yu, Hao Fei, Hongyuan Zhu, Jiayuan Fan, and Tao Chen. L13da: Visual interactive instruction tuning for omni-3d understanding reasoning and planning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26428-26438, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.552, + 0.484, + 0.62 + ], + "angle": 0, + "content": "[8] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In European Conference on Computer Vision, pages 370-386. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.622, + 0.484, + 0.675 + ], + "angle": 0, + "content": "[9] Yilun Chen, Shuai Yang, Haifeng Huang, Tai Wang, Runsen Xu, Ruiyuan Lyu, Dahua Lin, and Jiangmiao Pang. Grounded 3d-llm with referent tokens. arXiv preprint arXiv:2405.10370, 2024. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.678, + 0.484, + 0.745 + ], + "angle": 0, + "content": "[10] Francis Engelmann, Fabian Manhardt, Michael Niemeyer, Keisuke Tateno, Marc Pollefeys, and Federico Tombari. Opennerf: open set 3d neural scene segmentation with pixelwise features and rendered novel views. arXiv preprint arXiv:2404.03650, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.774 + ], + "angle": 0, + "content": "[11] Gemini Team et al. Gemini: A family of highly capable multimodal models, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[12] Zhiwen Fan, Jian Zhang, Wenyan Cong, Peihao Wang, Renjie Li, Kairun Wen, Shijie Zhou, Achuta Kadambi, Zhangyang Wang, Danfei Xu, et al. Large spatial model: End-to-end unposed images to semantic 3d. Advances in Neural Information Processing Systems, 37:40212-40229, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[13] Rao Fu, Jingyu Liu, Xilun Chen, Yixin Nie, and Wenhan Xiong. Scene-llm: Extending language model for 3d visual understanding and reasoning. arXiv preprint arXiv:2403.11401, 2024. 1, 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.175 + ], + "angle": 0, + "content": "[14] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the V in VQA Matter: Elevating the Role of Image Understanding in Visual Question Answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6904-6913, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.907, + 0.234 + ], + "angle": 0, + "content": "[15] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems, 36:20482-20494, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.907, + 0.303 + ], + "angle": 0, + "content": "[16] Haifeng Huang, Yilun Chen, Zehan Wang, Rongjie Huang, Runsen Xu, Tai Wang, Luping Liu, Xize Cheng, Yang Zhao, Jiangmiao Pang, et al. Chat-scene: Bridging 3d scene and large language models with object identifiers. arXiv preprint arXiv:2312.08168, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.305, + 0.907, + 0.373 + ], + "angle": 0, + "content": "[17] Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. An embodied generalist agent in 3d world. arXiv preprint arXiv:2311.12871, 2023. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.906, + 0.418 + ], + "angle": 0, + "content": "[18] Weitai Kang, Haifeng Huang, Yuzhang Shang, Mubarak Shah, and Yan Yan. Robin3d: Improving 3d large language model via robust instruction tuning, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.42, + 0.907, + 0.488 + ], + "angle": 0, + "content": "[19] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.49, + 0.907, + 0.56 + ], + "angle": 0, + "content": "[20] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.907, + 0.617 + ], + "angle": 0, + "content": "[21] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.619, + 0.907, + 0.673 + ], + "angle": 0, + "content": "[22] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava-last-interleave: Tackling multi-image, video, and 3d in large multimodal models, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.907, + 0.744 + ], + "angle": 0, + "content": "[23] Hajjie Li, Yanmin Wu, Jiarui Meng, Qiankun Gao, Zhiyao Zhang, Ronggang Wang, and Jian Zhang. Instancegaussian: Appearance-semantic joint gaussian representation for 3d instance-level perception. arXiv preprint arXiv:2411.19235, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.747, + 0.907, + 0.815 + ], + "angle": 0, + "content": "[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[25] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4720" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.135 + ], + "angle": 0, + "content": "[26] Chin-Yew Lin. ROUGE: A Package for Automatic Evaluation of Summaries. In Proceedings of the Workshop on Text Summarization Branches Out, pages 74-81, 2004. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.485, + 0.205 + ], + "angle": 0, + "content": "[27] Xiongkun Linghu, Jiangyong Huang, Xuesong Niu, Xiaojian Shawn Ma, Baoxiong Jia, and Siyuan Huang. Multimodal situated reasoning in 3d scenes. Advances in Neural Information Processing Systems, 37:140903-140936, 2025. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.484, + 0.248 + ], + "angle": 0, + "content": "[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36:34892-34916, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.484, + 0.332 + ], + "angle": 0, + "content": "[29] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. MMBench: Is your multi-modal model an all-around player? In Proceedings of the European Conference on Computer Vision (ECCV), 2024. Accepted as Oral Presentation. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.484, + 0.404 + ], + "angle": 0, + "content": "[30] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, Yaofeng Sun, Chengqi Deng, Hanwei Xu, Zhenda Xie, and Chong Ruan. Deepseek-vl: Towards real-world vision-language understanding, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.484, + 0.461 + ], + "angle": 0, + "content": "[31] Xiaojian Ma, Silong Yong, Zilong Zheng, Qing Li, Yitao Liang, Song-Chun Zhu, and Siyuan Huang. Sqa3d: Situated question answering in 3d scenes. arXiv preprint arXiv:2210.07474, 2022. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.484, + 0.49 + ], + "angle": 0, + "content": "[32] OpenAI. Gpt-4 with vision (gpt-4v). https://openai.com/research/gpt-4, 2023. Accessed: 2024-02-28. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.492, + 0.484, + 0.547 + ], + "angle": 0, + "content": "[33] Qucheng Peng, Benjamin Planche, Zhongpai Gao, Meng Zheng, Anwesa Choudhuri, Terrence Chen, Chen Chen, and Ziyan Wu. 3d vision-language gaussian splatting. arXiv preprint arXiv:2410.07577, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.549, + 0.484, + 0.619 + ], + "angle": 0, + "content": "[34] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 815-824, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.62, + 0.484, + 0.688 + ], + "angle": 0, + "content": "[35] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20051-20060, 2024. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.69, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[37] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13009-13018, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[38] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. CIDEr: Consensus-based Image Description Evaluation. In Proceedings of the IEEE Conference on Computer" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.908, + 0.12 + ], + "angle": 0, + "content": "Vision and Pattern Recognition (CVPR), pages 4566-4575, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.908, + 0.19 + ], + "angle": 0, + "content": "[39] Peihao Wang, Zhiwen Fan, Zhangyang Wang, Hao Su, Ravi Ramamoorthi, et al. Lift3d: Zero-shot lifting of any 2d vision model to 3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21367-21377, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.907, + 0.261 + ], + "angle": 0, + "content": "[40] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Song XiXuan, et al. Cogvlm: Visual expert for pretrained language models. Advances in Neural Information Processing Systems, 37:121475-121499, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.907, + 0.318 + ], + "angle": 0, + "content": "[41] Xingrui Wang, Wufei Ma, Zhuowan Li, Adam Kortyewski, and Alan L Yuille. 3d-aware visual question answering about parts, poses and occlusions. Advances in Neural Information Processing Systems, 36:58717-58735, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.906, + 0.361 + ], + "angle": 0, + "content": "[42] Xingrui Wang, Cuiling Lan, Hanxin Zhu, Zhibo Chen, and Yan Lu. Gsemsplat: Generalizable semantic 3d gaussian splatting from uncalibrated image pairs, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.907, + 0.431 + ], + "angle": 0, + "content": "[43] Yunsong Wang, Tianxin Huang, Hanlin Chen, and Gim Hee Lee. Freesplat: Generalizable 3d gaussian splatting towards free view synthesis of indoor scenes. Advances in Neural Information Processing Systems, 37:107326-107349, 2025. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.906, + 0.488 + ], + "angle": 0, + "content": "[44] Zehan Wang, Haifeng Huang, Yang Zhao, Ziang Zhang, and Zhou Zhao. Chat-3d: Data-efficiently tuning large language model for universal dialogue of 3d scenes. arXiv preprint arXiv:2308.08769, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.489, + 0.907, + 0.572 + ], + "angle": 0, + "content": "[45] Yanmin Wu, Jiarui Meng, Hajjie Li, Chenming Wu, Yahao Shi, Xinhua Cheng, Chen Zhao, Haocheng Feng, Errui Ding, Jingdong Wang, et al. Opengaussian: Towards point-level 3d gaussian-based open vocabulary understanding. Advances in Neural Information Processing Systems, 37:19114-19138, 2025. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.574, + 0.907, + 0.629 + ], + "angle": 0, + "content": "[46] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6281–6290, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.63, + 0.908, + 0.713 + ], + "angle": 0, + "content": "[47] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splattering to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676-21685, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.715, + 0.908, + 0.769 + ], + "angle": 0, + "content": "[48] Chenming Zhu, Tai Wang, Wenwei Zhang, Jiangmiao Pang, and Xihui Liu. Llava-3d: A simple yet effective pathway to empowering lmm's with 3d-awareness. arXiv preprint arXiv:2409.18125, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.771, + 0.908, + 0.84 + ], + "angle": 0, + "content": "[49] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2911-2921, 2023. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.84 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4721" + } + ] +] \ No newline at end of file diff --git a/2025/SplatTalk_ 3D VQA with Gaussian Splatting/71c4eb8c-9be8-49e8-bab3-c4048f8e7d2a_origin.pdf b/2025/SplatTalk_ 3D VQA with Gaussian Splatting/71c4eb8c-9be8-49e8-bab3-c4048f8e7d2a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..97e3ef94c8d7bfff49b7910c3ce61fc5f5b14e03 --- /dev/null +++ b/2025/SplatTalk_ 3D VQA with Gaussian Splatting/71c4eb8c-9be8-49e8-bab3-c4048f8e7d2a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a8f25ba7454b543130ad56e06be31d2ad5d1aa3df6b925481c4a07842d861dcd +size 791068 diff --git a/2025/SplatTalk_ 3D VQA with Gaussian Splatting/full.md b/2025/SplatTalk_ 3D VQA with Gaussian Splatting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..432fe38afe2e1891016367ede42631b103e41d91 --- /dev/null +++ b/2025/SplatTalk_ 3D VQA with Gaussian Splatting/full.md @@ -0,0 +1,251 @@ +# SplatTalk: 3D VQA with Gaussian Splatting + +Anh Thai $^{1,2*}$ + +Songyou Peng² + +Kyle Genova2 + +Leonidas Guibas2 + +Thomas Funkhouser + +1Georgia Institute of Technology + +$^{2}$ Google DeepMind + +*Work done as a student researcher at Google DeepMind. + +![](images/10c8644c27e3ab27ac7da88a7566145a44b407cd2f94be94301b677c673aae01.jpg) +Figure 1. SplatTalk. We propose a self-supervised 3D-Language Gaussian Splitting model trained from multi-view RGB images. First, images are encoded using a pretrained 2D Vision-Language Model (VLM) and projected into visual-language feature maps via a multimodal projector. These feature maps are then learned within a feed-forward 3D-Language Gaussian Splitting model, producing a 3D-Language Gaussian Field that encodes spatial and semantic information in 3D space. During inference, the 3D Gaussian features are directly queried by a Large Language Model (LLM) to perform 3D question-answering (3D VQA) tasks. + +# Abstract + +Language-guided 3D scene understanding is important for advancing applications in robotics, AR/VR, and human-computer interaction, enabling models to comprehend and interact with 3D environments through natural language. While 2D vision-language models (VLMs) have achieved remarkable success in 2D VQA tasks, progress in the 3D domain has been significantly slower due to the complexity of 3D data and the high cost of manual annotations. In this work, we introduce SplatTalk, a novel method that uses a generalizable 3D Gaussian Splatting (3DGS) framework to produce 3D tokens suitable for direct input into a pretrained LLM, enabling effective zero-shot 3D visual question answering (3D VQA) for scenes with only posed images. During experiments on multiple benchmarks, our approach outperforms both 3D models trained specifically for the task and previous 2D-LMM-based models utilizing only images (our setting), while achieving competitive performance with state-of-the-art 3D LMMs that additionally utilize 3D inputs. + +Project website: https://splat-talk.github.io/ + +# 1. Introduction + +The rise of Large Multimodal Models (LMMs), following the success of Large Language Models (LLMs), has become increasingly significant in a world where visual understanding is crucial. While 2D LMMs have seen rapid improvements, enabling models to analyze images and answer complex questions [21, 28, 37, 40], progress in 3D LMMs has been slower. This gap is largely due to the scarcity of 3D data compared to 2D datasets. Further, the availability of language-annotated 3D data is even more limited, as annotating 3D scenes since it is significantly more complex and expensive than labeling 2D images. These challenges have hindered the development of language-guided 3D reasoning, making 3D LMMs an important yet underexplored area in multimodal research. In this work, we focus on 3D Visual Question Answering (3D VQA) as it provides a strong benchmark for evaluating a model's ability to understand and reason about 3D scenes. + +Recently, significant progress has been made in 3D multimodal learning, particularly for indoor scenes [15, 17, 27, 48]. Several approaches rely on posed images and 3D inputs, like meshes [41], point clouds [9, 17, 34] and voxels [13], to associate 3D geometry with language features. As such, + +they have limited applicability in scenarios where a full 3D surface reconstruction is not available. + +Previous works have explored integrating semantic features into NeRF [19] or 3D Gaussian Splatting (3DGS) [35], which rely only upon RGB image inputs. However, research on effectively associating 3D geometry with language features remains limited. Some methods align individual 3D locations with language features via implicit functions optimized individually for every scene [10, 19]. Others employ SAM [20] to first segment objects and object parts, followed by CLIP [36] to extract object-centric features from the generated masks. These approaches allow models to query information about individual points and/or objects. However, it disregards spatial relationships between objects, making it unsuitable for 3D VQA spatial reasoning tasks. + +More recent work [21] has demonstrated the power of LLMs to answer VQA questions using only a sampling of tokens extracted from images, without any underlying 3D representation. While this approach is inspiring, it relies upon the LLM to make sense of a set of tokens that are arranged according to images rather than elements in the 3D scene. As a result, it struggles to answer queries about specific 3D locations (e.g., "What is here?") and questions requiring a holistic understanding of the entire 3D scene, where the answer depends on establishing correspondences between objects across multiple images (e.g., "What is opposite the door at the other end of the room"?). In this work, we aim to address these issues by consolidating tokens in 3D before providing them to the LLM. + +Our approach is called SplatTalk. Given a set of posed RGB images, we integrate language features extracted from the RGB images to build a 3D Gaussian Splitting representation, from which 3D tokens can be extracted for direct input into a pretrained LLM to answer challenging 3D VQA queries (see Fig. 1). The main research contributions are the new methods for 1) integrating language features into the Gaussian representation, and 2) extracting tokens that can be provided directly to an LLM. These methods avoid the issues of concurrent work, ChatSplat [6], which produces tokens that are not directly compatible with an LLM, making language-based reasoning more difficult. + +Overall, SplatTalk has a unique combination of three advantages. First, it does not require any 3D surface representation as input. Second, it leverages the full power of a pre-trained LLM to answer difficult 3D VQA questions. Third, it produces a concise 3D representation that can be queried and/or edited in 3D (e.g., cropped to a 3D region, rendered from a novel view, etc.). During experiments with 3D VQA benchmarks, we find that it outperforms 3D models trained only for the benchmark tasks, demonstrating the value of utilizing pre-trained LLMs. SplatTalk further outperforms prior image-only methods using LLMs, showing the value of building a 3D scene representation. Finally, + +it even achieves competitive performance with SOTA 3D LMMs that utilize 3D scene representation inputs. + +In summary, our key contributions are: + +- We introduce the first self-supervised 3D Gaussian-based method for large-scale zero-shot 3D VQA, SplatTalk, requiring only multi-view images without explicit depth, point clouds or 3D-language annotated supervision. +- We show that mean feature extraction from 3D Gaussians encodes holistic scene concepts for 3D VQA, supported by theoretical justification. +- We introduce entropy-adaptive token sampling, leading to better 3D VQA performance without additional training. +- We demonstrate that SplatTalk outperforms 2D LMMs and achieves performance on par with 3D models using point clouds on various 3D VQA benchmarks, demonstrating the effectiveness of our 3D-aware representations. + +# 2. Related Work + +2D LMMs. With advancements in large language models (LLMs), the ability of models to comprehend images and respond to language prompts related to their content has improved significantly. Notably, SOTA models [1, 11, 21, 30, 32] have demonstrated remarkable capabilities in processing multiple images of a 3D scene and answering questions about the shared content across these images. Although multiple RGB images inherently contain 3D information, they do not explicitly model 3D understanding, potentially leading to sub-optimal performance on tasks that require explicit 3D reasoning. In this work, we build upon the strong performance of 2D vision-language models (2D-VLMs), particularly LLaVA-OneVision [21], and integrate them with the 3D priors provided by the 3DGS framework to improve 3D spatial understanding capabilities. + +3D LMMs. Recent advancements in 3D indoor scene understanding have explored integrating 3D priors into LLMs using inputs such as 3D point clouds [7, 17] or multi-view RGB images combined with depth or 3D point clouds [48]. These approaches employ various strategies to embed 3D structural information and language features into a shared representation space. While most methods [17, 48] fine-tune pretrained visual encoders (both 2D and 3D) alongside LoRA parameters or even the full set of LLM parameters, others [18, 24] opt for a more lightweight approach, training only a projection layer to map visual features into the language feature space. In contrast, our method does not require any explicit depth or 3D inputs. Instead, we leverage the 3DGS framework, which has shown strong capabilities in representing 3D scenes using only multi-view RGB images. Generalizable 3D Gaussian Splatting. This body of work explores generalizable 3DGS as an alternative to the perscene optimization paradigm used in traditional 3DGS methods, aiming to improve novel view synthesis across unseen scenes. Our method builds on FreeSplat [43], which is + +![](images/84c7612776e61cc762c577c9f35bae46a16175dfc583846ddcb7ca4e1203f7c8.jpg) +Figure 2. Overview. Left: During the self-supervised 3D-language Gaussian Splitting training phase, multiple RGB input views are first encoded into Gaussian latent features (Gaussian triplets). These latent features are then decoded into Gaussian parameters for rendering, along with a low-dimensional visual-language feature $f$ . To ensure proper supervision of this low-dimensional feature, we train an autoencoder that maps the high-dimensional, unbounded features obtained from LLaVA-OV, specifically, the visual tokens serving as direct inputs to the LLM, onto a low-dimensional hypersphere space. Right: During 3D VQA inference, visual-language features are directly extracted from the 3D Gaussians. These features are then mapped back to the original high-dimensional space using the pretrained decoder and subsequently used as direct visual token inputs to the LLM. LoRA fine-tuning of the LLM is optional. + +![](images/85c094a9cea333eefbe69806bb71b0124ee4c9f8bfaf92e316236e14d9993baf.jpg) + +designed for sparse-view synthesis in indoor scenes. The method begins by extracting multi-scale features from input RGB images using CNN backbones. It then constructs adaptive cost volumes from neighboring views to predict depth maps. These predicted depth maps are used to backproject 2D CNN-encoded features into 3D Gaussian triplets $(\mu, \omega, f)$ where $\mu$ denotes the Gaussian's center, $\omega$ the weight, and $f$ the latent feature. To integrate multi-view information, FreeSplat introduces a Pixel-wise Triplet Fusion (PTF) module that uses a lightweight GRU to align and fuse overlapping 3D Gaussians across views efficiently. Finally, the fused triplets are passed through an MLP to decode rendering parameters, $\Sigma, \alpha, s$ . Our approach extends the FreeSplat framework to enable the learning of generalizable 3D Gaussian language fields. + +Similar to FreeSplat [43], [5, 8] leverage multi-view stereo (MVS) cues to refine depth estimation and improve geometric consistency in 3D Gaussian representations. These methods primarily focus on RGB novel view synthesis from sparse inputs (2-10 views). Our work builds on the SOTA model, FreeSplat [43], but diverges significantly in both objectives and methodology. First, while previous methods focus on photorealistic RGB reconstruction, we aim at semantic 3D scene understanding. Second, instead of reconstructing only a localized portion of the scene as shown in these works, our goal is to represent entire 3D environments, capturing both spatial and semantic relationships holistically. Language-guided 3D Gaussian Splatting. LangSplat [35] is one of the first approaches that integrates semantic features into 3DGS. However, this method requires training on individual scenes, which is significantly expensive, and lacks generalization to unseen scenes. Similarly, follow-up + +works [23, 45] explore 3DGS for open-vocabulary object segmentation but do not focus on reasoning tasks like 3D VQA. Further, they primarily rely on scene-specific training, limiting their scalability. In contrast, we propose a fully self-supervised framework that learns 3D-aware visual representations without requiring per-scene training. While some works [12, 42] explores a generalizable 3DGS and semantic modeling approach, they primarily focus on open-vocabulary segmentation, without addressing the broader challenges of free-form language reasoning required for 3D VQA. + +ChatSplat [6] is a concurrent work that explores natural language conversation within the 3DGS framework. However, it optimizes on a per-scene basis and requires rendering 3D features into 2D feature maps during inference. Additionally, its training pipeline is not fully end-to-end, as it first trains the RGB Gaussians separately before incorporating language features. In contrast, our approach leverages scene latent features that encapsulate both RGB and semantic information, enabling the joint optimization of both modalities in a fully end-to-end manner. Our approach extends the FreeSplat framework to enable the learning of generalizable 3D Gaussian language fields. + +# 3. Method + +The problem we address is formulated as follows: Given a sequence of $N$ posed RGB images captured from a 3D indoor scene, along with language prompts related to the scene's content, the goal is to generate the corresponding language responses. That is, we address 3D VQA tasks with solely posed RGB images as inputs. + +We provide an overview of our approach in Fig. 2, which + +![](images/396b4405b322ed870007efe67c8c51355a6dbc783b7af71262143472d6623b90.jpg) +Figure 3. Qualitative results on ScanQA, scene0030_00 and scene0084_00. We compare responses from LLaVA-OV, our model (Ours), and the ground truth (GT) for spatial reasoning questions in 3D VQA. Each scene highlights the referenced objects with red circles, and key relative objects are marked in blue. The answers from each model are displayed in color-coded speech bubbles: LLaVA-OV (brown), Ours (blue), and GT (green). With its 3D-aware representation, our model exhibits improved spatial reasoning capabilities, accurately identifying relationships between objects that are far apart and may never co-occur in the same image (e.g., door and window). More examples in Supp. + +![](images/cd69f61788950ca6a3f9c919091ec419ab45ec136019317853b7eaa878c619d4.jpg) + +consists of three main steps: 1) Training the feature autoencoder, 2) Self-supervised 3D-Language Gaussian Splitting Training and 3) 3D VQA Inference. First, we extract the high-dimensional language features from LLaVA-OV [21] for each RGB image by using the pretrained image encoder and multimodal projector. Since these features are high-dimensional, unbounded and sparse, we train an autoencoder to project them into a more compact, low-dimensional space. These low-dimensional features then serve as pseudo groundtruth for training the language representations in SplatTalk. SplatTalk architecture is a feed-forward network consisting of a Gaussian encoder and decoder that predicts Gaussian parameters for rendering. During 3D VQA inference stage, we extract the trained language features from the 3D Gaussians and directly feed them into the LLM, along with the language prompt, to generate free-form responses. + +# 3.1. Integrating Language Features + +An important step of our computational pipeline is the self-supervised method to integrate language features into a generalizable 3DGS framework. We identify four key factors that shape our approach for this step: (1) Leveraging features that align with the LLM's input space for direct language reasoning, (2) Efficient feature dimensionality reduction for effective CUDA-based rendering, and (3) Joint training of RGB and semantic features for coherent multimodal representation learning. + +Visual Tokens for 2D Pseudo Ground-truth Features. We extract the visual tokens used as inputs to the LLM for training our 3DGS framework. Our empirical findings indicate that extracting features directly from the visual encoder for training 3DGS requires retraining the multimodal projector + +layer which aligns the visual with language spaces. This observation is further corroborated by [6], which demonstrates that features extracted before the multimodal projector cannot be seamlessly integrated into the language model. A possible explanation is that the image encoder's native feature space (before the projector) is high-dimensional and unbounded, leading to sparser features with greater variance across different dimensions. As a result, the projector becomes more sensitive to subtle variations in the input features, making it difficult to directly map them into a token representation that aligns with the LLM's understanding. In contrast, the visual tokens generated after the projector are already well-aligned with the language model's latent space, allowing the LLM to interpret and reason with them directly. + +Feature Dimensionality Reduction. Building on the previous factor, directly optimizing these language embeddings within a differentiable rendering pipeline is challenging because they are extremely high-dimensional (e.g., 3584 dimensions), sparse, and contain unbounded values, making learning both unstable and impractical. Therefore, we learn a generalizable autoencoder that projects these sparse, high-dimensional visual tokens into a compact unit hypersphere with much lower dimension (e.g., 256 dimensions), ensuring feature smoothness. Note that the language representations from LLaVA, which support reasoning and free-form language comprehension, are significantly more complex than those from CLIP or segmentation-based models like SAM. Different from prior works that compress semantic features into extremely low-dimensional vectors (e.g., 3-16 dimensions [35, 45]), which simplifies rendering but results in significant information loss, our approach prior + +Table 1. Evaluation of 3D VQA on ScanQA [2] and SQA3D [31] datasets. We compare SplatTalk against various baselines, including 3D LMMs, which take point clouds (PC) or both point clouds and images $(\mathrm{PC} + \mathrm{I})$ as visual inputs, and 2D LMM-based models, which rely solely on images (I). As a 2D LMM-based approach, SplatTalk outperforms other 2D LMM-based models, surpasses some 3D LMMs, and achieves competitive performance against the strongest 3D LMM baselines. Our models are highlighted in blue, bold indicates best performance among the 2D-LMM methods. + +
MethodModalityScanQA (val)SQA3D (test)
CIDErMETEORROUGEEM@1EM@1-REM@1EM@1-R
Specialist Models
ScanQA [2]PC64.913.133.121.1-47.2-
ScanRefer+MCan [46]PC55.411.530.018.6---
3D-VisTA [49]PC69.613.935.722.4-48.5-
SQA3D [31]PC-----46.6-
Generalist 3D LMMs
Chat-Scene [16]PC+I87.718.041.621.6-54.657.5
Scene-LLM [13]PC+I80.015.835.9--53.6-
Grounded 3D-LLM [9]PC72.7------
Chat-3D-v2 [44]PC77.116.140.1--54.7-
Finetuned 3D LMMs
3D-LLM [15]PC+I69.414.535.720.5---
LEO [17]PC+I101.420.049.224.547.650.052.4
Scene-LLM [13]PC+I80.016.640.027.2-54.2-
LL3DA [7]PC76.815.937.3----
Chat-3D [44]PC53.211.928.5----
2D LMM-Based
VideoChat2 [25]I49.29.528.219.2-37.3-
GPT-4V [32]I59.613.533.4----
Claude [1]I57.710.029.3----
LLaVA-NeXT-Video [22]I46.29.127.8--34.2-
LLaVA-OV [21]I50.013.029.415.632.525.435.8
SplatTalkI61.714.232.717.132.226.141.3
SplatTalk-ScanQA-FTI77.115.438.722.338.342.044.7
SplatTalk-3DVQA-FTI77.515.638.522.438.347.649.4
+ +itizes preserving rich language features by maintaining a higher-dimensional representation (e.g., 256). To efficiently render high-dimensional feature maps on CUDA, we draw inspiration from [47] and implement a parallel CUDA rasterizer pipeline, where semantic features and RGB are rendered simultaneously, using shared Gaussian parameters. + +Unlike previous methods [6, 35] that train a separate autoencoder for each individual scene, we train a single autoencoder across all training scenes, each consisting of multiple frames. This approach improves the autoencoder's generalizability, reducing computational overhead by eliminating the need for retraining to compress features from unseen scenes. Training the autoencoder separately from the 3DGS model ensures that it is dedicated solely to compressing high-dimensional feature vectors while preserving the essential scene information without being influenced by the complexities of the rendering pipeline. Our empirical findings further indicate that jointly training a feature de + +coder to project low-dimensional embeddings back to the high-dimensional space introduces significant instabilities in the training process. + +Joint-training RGB and Language. In contrast to previous approaches that first train RGB Gaussians and then freeze them before integrating language features [6, 35], we adopt a joint training strategy, where RGB and language features are learned simultaneously. This is due to the fact that visual RGB and semantic features play a complementary role in effectively representing 3D scenes [33]. In feedforward models like FreeSplat, where the scene is first encoded into a Gaussian latent space before being decoded into specific Gaussian parameters for RGB rendering, these Gaussian latent features inherently capture holistic scene information, including both appearance (RGB) and semantics. Based on this observation, we introduce a new Gaussian parameter decoder head that predicts semantic features for each 3D Gaussian alongside rendering parameters, enabling joint op- + +Table 2. Evaluation of 3D VQA on MSR3D [27]. We compare SplatTalk against various baselines, all evaluated in a zero-shot setting. T denotes models that rely solely on text inputs without interleaving images, PC indicates models that use point cloud as the visual input, and I represents models that process only image-based inputs. Our approach, highlighted in blue, outperforms all baselines. + +
ModelInputSceneSettingCountingExistenceAttributesSpatialNavigationOthersOverall
LEO [17]TPCzero-shot0.8015.511.87.32.315.37.8
LLaVA-OV [21]TIzero-shot18.631.224.819.516.736.324.0
SplatTalkTIzero-shot19.660.344.035.835.561.841.8
SplatTalk-ScanQA-FTTIzero-shot28.966.043.228.333.860.041.5
SplatTalk-3DVQA-FTTIzero-shot24.357.036.17.816.435.026.5
+ +tization. Unlike approaches that explicitly model these interactions [33, 39] using extra cross-attention or projection layers, SplatTalk implicitly integrates semantic and visual information within the Gaussian representation. This paradigm naturally facilitates multimodal interactions between RGB and language features. + +To incorporate more views and comprehensively cover the scene, we reduce the input resolution of the RGB images. This adjustment aligns well with the inherently low spatial resolution of the visual tokens extracted from LLaVA-OV, and also prevents the model from disproportionately allocating resources to modeling RGB features, thus encouraging a balanced representation of semantic and RGB information. + +These design choices establish our approach as a fundamentally novel method for integrating language features into 3DGS, grounded in first-principles observations rather than heuristic adaptations. By addressing the core challenges of multimodal alignment, feature compression, and joint optimization, our approach is not merely a trivial extension of existing works, but a carefully designed framework that ensures both effectiveness and stability in learning rich, spatially-aware semantic representations. + +# 3.2. Extracting 3D Language Features + +During inference, we directly extract language features at the mean position of each 3D Gaussian. Although the rendering process incorporates each Gaussian's covariance and opacity when projecting 3D features to 2D, we demonstrate that scene semantics can be effectively captured by the mean feature of each Gaussian, $\pmb{f}_i$ , alone, without relying on these additional factors. This is supported by the conceptual similarity between our training optimization process and the Expectation-Maximization (EM) algorithm. For simplification, let $\mathcal{L} = \sum_{t=1}^{T} \| \pmb{F}_t(x) - \pmb{F}_t^{gt}(x) \|^2$ be the loss, where $\pmb{F}_t(x)$ is the predicted rendered feature map of frame $t$ at pixel $x$ and $\pmb{F}^{gt}$ is the pseudo ground-truth feature map obtained from LLaVA-OV. Replacing $\pmb{F}_t(x)$ by the rendering equation and solving for the optimal $\pmb{f}_i^*$ , we obtain $\pmb{f}_i^* = \frac{\sum_{t,x} R_i(t,x) \pmb{F}_t^{gt}(x)}{\sum_{t,x} R_i(t,x)}$ where $R_i(t,x) = \frac{\alpha_i \mathcal{N}(x; \pmb{\mu}_i, \pmb{\Sigma}_i)}{\sum_j \alpha_j \mathcal{N}(x; \pmb{\mu}_j, \pmb{\Sigma}_j)}$ is the Gaussian's contribution to pixel $x$ at viewpoint $t$ . This mirrors the EM procedure: estimating $R$ by rendering (E-step) + +and updating $f_{i}$ by minimizing reconstruction loss (M-step). Consequently, the optimal 3D feature $\pmb{f}_{i}^{*}$ for each Gaussian is a weighted sum of the 2D feature maps, and the set of 3D Gaussian features $\{\pmb{f}_{i}\}_{i=1}^{N}$ encodes the scene's semantic holistically from $T$ 2D inputs. These features are then projected back into the original language embedding space using the pretrained decoder and directly provided as inputs to the LLM for downstream 3D-language tasks. + +In open-vocabulary segmentation, querying Gaussian centers is effective [4, 35, 45] because each Gaussian is assumed to represent a distinct, meaningful object or region. These methods use object-centric embeddings (e.g., SAM-masked CLIP or LSeg) and retrieve objects via cosine similarity to text queries. In contrast, 3D reasoning tasks lack this one-to-one correspondence. Each Gaussian encodes parts of a scene-level representation, where meaning stems from spatial and relational structure. Querying single Gaussians is thus insufficient for context-dependent reasoning. While our method also queries at Gaussian means, this is not trivial; our EM-based formulation offers a principled justification. + +# 3.3.3D VQA Inference + +Entropy Adaptive Gaussian Sampling. SplatTalk leverages triplet fusion adopted in FreeSplat [43], which reduces the number of Gaussians. However, complex scenes still require significantly more Gaussians than the LLM's visual token limit, and not all are equally informative. To address this, we propose entropy-based sampling: selecting the top $k$ Gaussians with the highest language feature entropy, where $k$ matches the LLM's visual token capacity. + +3D Tokens as Visual Inputs. After 3DGS training, each 3D Gaussian encodes a feature that is a weighted aggregation of 2D semantic information across all input views (Sec. 3.2). These features are also spatially grounded in their 3D coordinates. Because each token corresponds to a specific location in 3D space, its position is already implicitly encoded in the feature through the rendering and fusion process. Unlike 2D patches in a fixed grid, these 3D tokens form an unordered set, similar to a point cloud where spatial relationships are captured by the features themselves, not by sequence position. Therefore, we do not require explicit positional encodings or a fixed input order. The 3D tokens, + +extracted by sampling features at each Gaussian's mean, are directly fed into the LLM. + +# 3.4. Optional - Finetuning on 3D VQA Datasets + +While SplatTalk is capable of performing the 3D VQA task in a zero-shot manner, it also allows for flexible fine-tuning on 3D VQA datasets to enhance alignment and reasoning capabilities. Rather than training the model end-to-end, which is both computationally expensive and prone to overfitting on limited datasets, we optimize only the LoRA parameters applied to the LLM during fine-tuning. + +# 4. Experiments + +# 4.1. Implementation Details + +We train SplatTalk with 500 scenes from the ScanQA training set. The initial self-supervised stage focuses purely on learning the 3D scene representation, without leveraging any text-annotated information. For each scene, we uniformly sample 100 views around the scene and extract their per-patch features by passing them through the visual encoder of LLaVA-OV followed by the multi-modal projector. The subsequent SplatTalk model is trained on a single NVIDIA H100 (80GB) GPU. Both the LoRA fine-tuning and inference are also conducted on a single H100. More details can be found in Supp. Material. + +# 4.2. Datasets and Metrics + +Evaluation Dataset. We evaluate our method on ScanNet-based QA datasets: ScanQA [2], SQA3D [31], and MSR3D [27]. ScanQA is a benchmark designed for question answering in 3D indoor scenes, containing detailed spatial and semantic information extracted from ScanNet. SQA3D and MSR3D further extend this by introducing scenarios that describe an embodied agent's 3D position and state. + +Metrics. For ScanQA and SQA3D, we evaluate performance using standard n-gram-based metrics, specifically CIDEr [38], METEOR [3], ROUGE [26], EM@1 [14], and EM@1-Refined for ScanQA, and EM@1 and EM@1-Refined for SQA3D. In contrast, for MSR3D, we utilize the LLM-Match score [29] as recommended by the authors. + +# 4.3. Performance on 3D Question Answering Tasks + +ScanQA. We evaluate SplatTalk's performance against several model families, each representing different levels of specialization and generalization. "Specialist Models" refer to models specifically trained to solve ScanQA without exposure to other tasks. "Generalist 3D LMMs" consist of models trained across a diverse range of 3D vision-language tasks, enabling broader generalization. "Finetuned 3D LMMs" are models that have been further fine-tuned on the ScanQA training set to optimize performance for this particular dataset. Lastly, "2D LMM-Based Models" include + +models that process only image inputs without leveraging depth or explicit 3D information. SplatTalk and its fine-tuned variants fall within this category. + +We evaluate three different variants of our model on this task. The base model (SplatTalk) is trained solely to encode the semantic representation of the scene from multiple RGB images using 3DGS, without exposure to question-answering tasks. Note that this model is trained entirely in a self-supervised manner, without any access to language supervision or textual annotations during training. To further adapt it to downstream 3D VQA tasks, we fine-tune this base model on the ScanQA training set, referring to it as SplatTalk-ScanQA-FT, and on the combined training sets of ScanQA and SQA3D, denoting as SplatTalk-3DVQA-FT. We present results in Table 1. + +Compared to LLaVA-OV, whose 2D language feature maps are used to train SplatTalk, our base model, which benefits from a more 3D-aware representation due to the 3D Gaussian training process, achieves significantly higher performance (e.g. 61.7 vs 50.0 on CIDEr). Our language finetuned model variants, SplatTalk-ScanQA-FT and SplatTalk-3DVQA-FT, improve performance further. Notably, despite not being trained with depth supervision or explicit 3D information, our models achieve comparable results to both Generalist 3D LMMs and Finetuned 3D LMMs, while outperforming Specialist Models. Additionally, among 2D LMM-Based approaches, our base model without being fine-tuned with 3D VQA datasets outperforms other methods, with significantly higher performance on some metrics, including both open-source and closed-source models. We show qualitative results in Fig. 3. + +SQA3D. Similar to our evaluation on ScanQA, we compare our model against both 3D LMMs and 2D LMMs. Compared to 2D LLaVA-OV features, our base model with 3D-aware features achieves a significant performance gain, highlighting the advantage of incorporating 3D information via 3DGS which aligns with our findings on ScanQA. Further finetuning on ScanQA alone and on the combined ScanQA + SQA3D dataset leads to substantial improvements. Notably, SplatTalk-3DVQA-FT surpasses specialist models that leverage 3D point cloud inputs, demonstrating the strength of our approach in 3D scene understanding. + +MSR3D. We evaluated our model's performance against LLaVA-OV and $\mathsf{LEO}^1$ [17] on the Scannet split of MSR3D. Importantly, all comparisons are conducted in a zero-shot setting, as none of the models have been trained on MSR3D. This is an interleaved dataset where inputs consist of both text and images. However, since all models are tested in a zero-shot manner, they rely solely on text inputs, where the agent's state descriptions and the question are combined into + +Table 3. Ablation Study on Gaussian Sampling Strategies. Our entropy-based sampling strategy consistently outperforms across multiple metrics. + +
Sampling MethodEM@1METEORROUGESPICEAvg
Random17.013.862.114.026.73
Point Density16.214.059.714.126.00
FPS17.713.862.013.626.78
Entropy17.114.661.714.226.90
+ +a single text prompt. On the visual processing side, LEO operates on 3D point cloud inputs, whereas both LLaVA-OV and our model rely only on RGB images. + +The results, presented in Table 2, demonstrate that our base model outperforms LLaVA-OV across all question types. Notably, aside from the "counting" category, our base model achieves approximately twice the performance of LLaVA-OV on other question types. Both models significantly outperform LEO, which struggles due to the absence of fine-tuning on this dataset. + +Interestingly, we observe a slight performance drop when comparing SplatTalk with SplatTalk-ScanQA-FT. However, this drop becomes much more pronounced when compared with SplatTalk-3DVQA-FT. A plausible explanation is that fine-tuning on other QA datasets introduces overfitting, causing the model to adapt too specifically to those datasets. Furthermore, the observed performance drop suggests that MSR3D differs significantly from ScanQA and SQA3D, making direct knowledge transfer less effective. This distinction may also explain why LEO struggles to generalize in a zero-shot setting. Despite being trained on multiple QA datasets, none of them closely resemble MSR3D, limiting its ability to adapt without fine-tuning. + +# 4.4. Ablation Studies + +Gaussian Sampling Strategies. We explore various Gaussian sampling strategies for selecting tokens as inputs to the LLM, including: (1) Random Sampling—Gaussians are selected randomly across the scene. (2) Point Density Adaptive Sampling—Gaussians are sampled proportionally to the density of 3D points. (3) Furthest Point Sampling (FPS)—Gaussians are iteratively selected at locations maximally distant from previously chosen points. (4) Entropy Adaptive Sampling—Gaussians are prioritized based on language feature entropy, selecting points with the highest uncertainty. + +We conduct experiments on the ScanQA validation set. The results, presented in Table 3, demonstrate that entropy-adaptive sampling outperforms other sampling strategies, followed by FPS. This finding highlights entropy-adaptive sampling as the most effective approach for selecting informative tokens, making it the optimal strategy for downstream 3D VQA tasks. + +Length of Visual Tokens. We evaluate the performance of SplatTalk on ScanQA, SQA3D, and MSR3D under two different settings: (1) using only 1 image-worth of visual + +Table 4. Ablation Study on Visual Input Length. We compare model performance using 729 tokens (equivalent to a single image) versus 32,076 tokens (corresponding to 44 images), analyzing the impact of visual context on reasoning. + +
Visual LengthScanQA (val)SQA3D (test)MSR3D (test)
EM@1EM@1-REM@1EM@1-REM@1EM@1-R
729 tokens16.230.725.838.07.917.7
32,076 tokens17.132.226.141.314.123.3
+ +tokens (729 tokens) and (2) using the full context length corresponding to 44 images-worth of visual tokens (32,076 tokens). The results, presented in Table 4, evaluate model performance using EM@1 and EM@1-Refined, as these metrics offer a reliable measure of how precisely model predictions align with ground truth answers, both in literal accuracy and permissible variations in phrasing. We employ entropy-adaptive sampling (see Sec. 3.3), which prioritizes high-information tokens and remains deterministic given a fixed set of visual tokens. + +A key observation is that performance improved across all datasets as the number of input tokens increased, with MSR3D showing the largest gain, nearly doubling in EM@1. In contrast, ScanQA and SQA3D saw only marginal improvements, even with significantly more tokens. This suggests that additional scene context may not substantially improve performance, possibly due to: (1) Dataset limitations where the tasks may not require rich spatial context; or (2) Model limitations where the model may be unable to effectively leverage extra visual tokens for spatial reasoning. + +# 5. Conclusion + +In this work, we introduce the first method to integrate free-form language features into a 3D Gaussian Splitting framework, enabling zero-shot generalization for 3D Visual Question Answering (3D VQA) without requiring 3D point clouds or depth maps. Our approach surpasses methods that rely solely on 2D language feature maps and achieves performance on par with 3D LMMs that explicitly utilize 3D structural information across multiple 3D VQA benchmarks. We hope this work inspires further research into leveraging multi-view RGB inputs to encapsulate rich semantic representations of 3D scenes, ultimately enhancing 3D VQA capabilities without the reliance on explicit geometric priors. + +# Acknowledgment + +We thank the creators of the datasets and codebases used in our work, ScanQA, SQA3D, MSR3D, ScanNet, LLaVA-OV, and FreeSplat, for making their resources publicly available. We are also grateful to the anonymous reviewers for their constructive feedback during the review process. Special thanks to Xiang Li and James M. Rehg for the insightful discussions. + +# References + +[1] Anthropic. Claude 3. https://www.anthropic.com/index/claude, 2024. Accessed: 2024-02-28. 2, 5 +[2] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19129-19139, 2022. 5, 7 +[3] Satanjeev Banerjee and Alon Lavie. METEOR: An Automatic Metric for MT Evaluation with Improved Correlation with Human Judgments. In Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization, pages 65-72, 2005. 7 +[4] Yash Bhalgat, Iro Laina, Joao F Henriques, Andrew Zisserman, and Andrea Vedaldi. N2f2: Hierarchical scene understanding with nested neural feature fields. In European Conference on Computer Vision, pages 197-214. Springer, 2024. 6 +[5] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19457-19467, 2024. 3 +[6] Hanlin Chen, Fangyin Wei, and Gim Hee Lee. Chat-splat: 3d conversational gaussian splatting. arXiv preprint arXiv:2412.00734, 2024. 2, 3, 4, 5 +[7] Sijin Chen, Xin Chen, Chi Zhang, Mingsheng Li, Gang Yu, Hao Fei, Hongyuan Zhu, Jiayuan Fan, and Tao Chen. L13da: Visual interactive instruction tuning for omni-3d understanding reasoning and planning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26428-26438, 2024. 2, 5 +[8] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In European Conference on Computer Vision, pages 370-386. Springer, 2024. 3 +[9] Yilun Chen, Shuai Yang, Haifeng Huang, Tai Wang, Runsen Xu, Ruiyuan Lyu, Dahua Lin, and Jiangmiao Pang. Grounded 3d-llm with referent tokens. arXiv preprint arXiv:2405.10370, 2024. 1, 5 +[10] Francis Engelmann, Fabian Manhardt, Michael Niemeyer, Keisuke Tateno, Marc Pollefeys, and Federico Tombari. Opennerf: open set 3d neural scene segmentation with pixelwise features and rendered novel views. arXiv preprint arXiv:2404.03650, 2024. 2 +[11] Gemini Team et al. Gemini: A family of highly capable multimodal models, 2024. 2 +[12] Zhiwen Fan, Jian Zhang, Wenyan Cong, Peihao Wang, Renjie Li, Kairun Wen, Shijie Zhou, Achuta Kadambi, Zhangyang Wang, Danfei Xu, et al. Large spatial model: End-to-end unposed images to semantic 3d. Advances in Neural Information Processing Systems, 37:40212-40229, 2025. 3 +[13] Rao Fu, Jingyu Liu, Xilun Chen, Yixin Nie, and Wenhan Xiong. Scene-llm: Extending language model for 3d visual understanding and reasoning. arXiv preprint arXiv:2403.11401, 2024. 1, 5 + +[14] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the V in VQA Matter: Elevating the Role of Image Understanding in Visual Question Answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6904-6913, 2017. 7 +[15] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems, 36:20482-20494, 2023. 1, 5 +[16] Haifeng Huang, Yilun Chen, Zehan Wang, Rongjie Huang, Runsen Xu, Tai Wang, Luping Liu, Xize Cheng, Yang Zhao, Jiangmiao Pang, et al. Chat-scene: Bridging 3d scene and large language models with object identifiers. arXiv preprint arXiv:2312.08168, 2023.5 +[17] Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. An embodied generalist agent in 3d world. arXiv preprint arXiv:2311.12871, 2023. 1, 2, 5, 6, 7 +[18] Weitai Kang, Haifeng Huang, Yuzhang Shang, Mubarak Shah, and Yan Yan. Robin3d: Improving 3d large language model via robust instruction tuning, 2025. 2 +[19] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. 2 +[20] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 2 +[21] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 2, 4, 5, 6 +[22] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava-last-interleave: Tackling multi-image, video, and 3d in large multimodal models, 2024. 5 +[23] Hajjie Li, Yanmin Wu, Jiarui Meng, Qiankun Gao, Zhiyao Zhang, Ronggang Wang, and Jian Zhang. Instancegaussian: Appearance-semantic joint gaussian representation for 3d instance-level perception. arXiv preprint arXiv:2411.19235, 2024. 3 +[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 2 +[25] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 5 + +[26] Chin-Yew Lin. ROUGE: A Package for Automatic Evaluation of Summaries. In Proceedings of the Workshop on Text Summarization Branches Out, pages 74-81, 2004. 7 +[27] Xiongkun Linghu, Jiangyong Huang, Xuesong Niu, Xiaojian Shawn Ma, Baoxiong Jia, and Siyuan Huang. Multimodal situated reasoning in 3d scenes. Advances in Neural Information Processing Systems, 37:140903-140936, 2025. 1, 6, 7 +[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36:34892-34916, 2023. 1 +[29] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. MMBench: Is your multi-modal model an all-around player? In Proceedings of the European Conference on Computer Vision (ECCV), 2024. Accepted as Oral Presentation. 7 +[30] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, Yaofeng Sun, Chengqi Deng, Hanwei Xu, Zhenda Xie, and Chong Ruan. Deepseek-vl: Towards real-world vision-language understanding, 2024. 2 +[31] Xiaojian Ma, Silong Yong, Zilong Zheng, Qing Li, Yitao Liang, Song-Chun Zhu, and Siyuan Huang. Sqa3d: Situated question answering in 3d scenes. arXiv preprint arXiv:2210.07474, 2022. 5, 7 +[32] OpenAI. Gpt-4 with vision (gpt-4v). https://openai.com/research/gpt-4, 2023. Accessed: 2024-02-28. 2, 5 +[33] Qucheng Peng, Benjamin Planche, Zhongpai Gao, Meng Zheng, Anwesa Choudhuri, Terrence Chen, Chen Chen, and Ziyan Wu. 3d vision-language gaussian splatting. arXiv preprint arXiv:2410.07577, 2024. 5, 6 +[34] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 815-824, 2023. 1 +[35] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20051-20060, 2024. 2, 3, 4, 5, 6 +[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 2 +[37] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13009-13018, 2024. 1 +[38] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. CIDEr: Consensus-based Image Description Evaluation. In Proceedings of the IEEE Conference on Computer + +Vision and Pattern Recognition (CVPR), pages 4566-4575, 2015. 7 +[39] Peihao Wang, Zhiwen Fan, Zhangyang Wang, Hao Su, Ravi Ramamoorthi, et al. Lift3d: Zero-shot lifting of any 2d vision model to 3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21367-21377, 2024. 6 +[40] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Song XiXuan, et al. Cogvlm: Visual expert for pretrained language models. Advances in Neural Information Processing Systems, 37:121475-121499, 2025. 1 +[41] Xingrui Wang, Wufei Ma, Zhuowan Li, Adam Kortyewski, and Alan L Yuille. 3d-aware visual question answering about parts, poses and occlusions. Advances in Neural Information Processing Systems, 36:58717-58735, 2023. 1 +[42] Xingrui Wang, Cuiling Lan, Hanxin Zhu, Zhibo Chen, and Yan Lu. Gsemsplat: Generalizable semantic 3d gaussian splatting from uncalibrated image pairs, 2024. 3 +[43] Yunsong Wang, Tianxin Huang, Hanlin Chen, and Gim Hee Lee. Freesplat: Generalizable 3d gaussian splatting towards free view synthesis of indoor scenes. Advances in Neural Information Processing Systems, 37:107326-107349, 2025. 2, 3, 6 +[44] Zehan Wang, Haifeng Huang, Yang Zhao, Ziang Zhang, and Zhou Zhao. Chat-3d: Data-efficiently tuning large language model for universal dialogue of 3d scenes. arXiv preprint arXiv:2308.08769, 2023. 5 +[45] Yanmin Wu, Jiarui Meng, Hajjie Li, Chenming Wu, Yahao Shi, Xinhua Cheng, Chen Zhao, Haocheng Feng, Errui Ding, Jingdong Wang, et al. Opengaussian: Towards point-level 3d gaussian-based open vocabulary understanding. Advances in Neural Information Processing Systems, 37:19114-19138, 2025. 3, 4, 6 +[46] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6281–6290, 2019. 5 +[47] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splattering to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676-21685, 2024. 5 +[48] Chenming Zhu, Tai Wang, Wenwei Zhang, Jiangmiao Pang, and Xihui Liu. Llava-3d: A simple yet effective pathway to empowering lmm's with 3d-awareness. arXiv preprint arXiv:2409.18125, 2024. 1, 2 +[49] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2911-2921, 2023. 5 \ No newline at end of file diff --git a/2025/SplatTalk_ 3D VQA with Gaussian Splatting/images.zip b/2025/SplatTalk_ 3D VQA with Gaussian Splatting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3ee4cbcce693ef4fdd20e0431036c7a20d5e7f1c --- /dev/null +++ b/2025/SplatTalk_ 3D VQA with Gaussian Splatting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:433a30a0117772a9a32a243fcae004dc80827f7fc693d5ebc05ef787a1d9456e +size 397854 diff --git a/2025/SplatTalk_ 3D VQA with Gaussian Splatting/layout.json b/2025/SplatTalk_ 3D VQA with Gaussian Splatting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..646890acd61f1eea606c26e22ceaf8ae2bc4185e --- /dev/null +++ b/2025/SplatTalk_ 3D VQA with Gaussian Splatting/layout.json @@ -0,0 +1,6323 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 167, + 103, + 443, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 103, + 443, + 121 + ], + "spans": [ + { + "bbox": [ + 167, + 103, + 443, + 121 + ], + "type": "text", + "content": "SplatTalk: 3D VQA with Gaussian Splatting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 142, + 141, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 142, + 141, + 156 + ], + "spans": [ + { + "bbox": [ + 78, + 142, + 141, + 156 + ], + "type": "text", + "content": "Anh Thai" + }, + { + "bbox": [ + 78, + 142, + 141, + 156 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 154, + 142, + 231, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 142, + 231, + 157 + ], + "spans": [ + { + "bbox": [ + 154, + 142, + 231, + 157 + ], + "type": "text", + "content": "Songyou Peng²" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 244, + 143, + 314, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 143, + 314, + 157 + ], + "spans": [ + { + "bbox": [ + 244, + 143, + 314, + 157 + ], + "type": "text", + "content": "Kyle Genova2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 326, + 143, + 414, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 143, + 414, + 156 + ], + "spans": [ + { + "bbox": [ + 326, + 143, + 414, + 156 + ], + "type": "text", + "content": "Leonidas Guibas2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 427, + 143, + 531, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 427, + 143, + 531, + 156 + ], + "spans": [ + { + "bbox": [ + 427, + 143, + 531, + 156 + ], + "type": "text", + "content": "Thomas Funkhouser" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 165, + 162, + 324, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 162, + 324, + 177 + ], + "spans": [ + { + "bbox": [ + 165, + 162, + 324, + 177 + ], + "type": "text", + "content": "1Georgia Institute of Technology" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 349, + 163, + 445, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 163, + 445, + 177 + ], + "spans": [ + { + "bbox": [ + 349, + 163, + 445, + 177 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 349, + 163, + 445, + 177 + ], + "type": "text", + "content": "Google DeepMind" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 200, + 178, + 407, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 178, + 407, + 190 + ], + "spans": [ + { + "bbox": [ + 200, + 178, + 407, + 190 + ], + "type": "text", + "content": "*Work done as a student researcher at Google DeepMind." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 62, + 219, + 549, + 355 + ], + "blocks": [ + { + "bbox": [ + 62, + 219, + 549, + 355 + ], + "lines": [ + { + "bbox": [ + 62, + 219, + 549, + 355 + ], + "spans": [ + { + "bbox": [ + 62, + 219, + 549, + 355 + ], + "type": "image", + "image_path": "10c8644c27e3ab27ac7da88a7566145a44b407cd2f94be94301b677c673aae01.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 360, + 555, + 416 + ], + "lines": [ + { + "bbox": [ + 54, + 360, + 555, + 416 + ], + "spans": [ + { + "bbox": [ + 54, + 360, + 555, + 416 + ], + "type": "text", + "content": "Figure 1. SplatTalk. We propose a self-supervised 3D-Language Gaussian Splitting model trained from multi-view RGB images. First, images are encoded using a pretrained 2D Vision-Language Model (VLM) and projected into visual-language feature maps via a multimodal projector. These feature maps are then learned within a feed-forward 3D-Language Gaussian Splitting model, producing a 3D-Language Gaussian Field that encodes spatial and semantic information in 3D space. During inference, the 3D Gaussian features are directly queried by a Large Language Model (LLM) to perform 3D question-answering (3D VQA) tasks." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 151, + 426, + 200, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 426, + 200, + 437 + ], + "spans": [ + { + "bbox": [ + 151, + 426, + 200, + 437 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 54, + 472, + 297, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 472, + 297, + 687 + ], + "spans": [ + { + "bbox": [ + 54, + 472, + 297, + 687 + ], + "type": "text", + "content": "Language-guided 3D scene understanding is important for advancing applications in robotics, AR/VR, and human-computer interaction, enabling models to comprehend and interact with 3D environments through natural language. While 2D vision-language models (VLMs) have achieved remarkable success in 2D VQA tasks, progress in the 3D domain has been significantly slower due to the complexity of 3D data and the high cost of manual annotations. In this work, we introduce SplatTalk, a novel method that uses a generalizable 3D Gaussian Splatting (3DGS) framework to produce 3D tokens suitable for direct input into a pretrained LLM, enabling effective zero-shot 3D visual question answering (3D VQA) for scenes with only posed images. During experiments on multiple benchmarks, our approach outperforms both 3D models trained specifically for the task and previous 2D-LMM-based models utilizing only images (our setting), while achieving competitive performance with state-of-the-art 3D LMMs that additionally utilize 3D inputs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 688, + 235, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 688, + 235, + 699 + ], + "spans": [ + { + "bbox": [ + 55, + 688, + 235, + 699 + ], + "type": "text", + "content": "Project website: https://splat-talk.github.io/" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 426, + 394, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 426, + 394, + 437 + ], + "spans": [ + { + "bbox": [ + 315, + 426, + 394, + 437 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 448, + 556, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 556, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 556, + 651 + ], + "type": "text", + "content": "The rise of Large Multimodal Models (LMMs), following the success of Large Language Models (LLMs), has become increasingly significant in a world where visual understanding is crucial. While 2D LMMs have seen rapid improvements, enabling models to analyze images and answer complex questions [21, 28, 37, 40], progress in 3D LMMs has been slower. This gap is largely due to the scarcity of 3D data compared to 2D datasets. Further, the availability of language-annotated 3D data is even more limited, as annotating 3D scenes since it is significantly more complex and expensive than labeling 2D images. These challenges have hindered the development of language-guided 3D reasoning, making 3D LMMs an important yet underexplored area in multimodal research. In this work, we focus on 3D Visual Question Answering (3D VQA) as it provides a strong benchmark for evaluating a model's ability to understand and reason about 3D scenes." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 654, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 557, + 715 + ], + "type": "text", + "content": "Recently, significant progress has been made in 3D multimodal learning, particularly for indoor scenes [15, 17, 27, 48]. Several approaches rely on posed images and 3D inputs, like meshes [41], point clouds [9, 17, 34] and voxels [13], to associate 3D geometry with language features. As such," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4712" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": "they have limited applicability in scenarios where a full 3D surface reconstruction is not available." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 98, + 295, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 98, + 295, + 253 + ], + "spans": [ + { + "bbox": [ + 54, + 98, + 295, + 253 + ], + "type": "text", + "content": "Previous works have explored integrating semantic features into NeRF [19] or 3D Gaussian Splatting (3DGS) [35], which rely only upon RGB image inputs. However, research on effectively associating 3D geometry with language features remains limited. Some methods align individual 3D locations with language features via implicit functions optimized individually for every scene [10, 19]. Others employ SAM [20] to first segment objects and object parts, followed by CLIP [36] to extract object-centric features from the generated masks. These approaches allow models to query information about individual points and/or objects. However, it disregards spatial relationships between objects, making it unsuitable for 3D VQA spatial reasoning tasks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 255, + 297, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 297, + 422 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 297, + 422 + ], + "type": "text", + "content": "More recent work [21] has demonstrated the power of LLMs to answer VQA questions using only a sampling of tokens extracted from images, without any underlying 3D representation. While this approach is inspiring, it relies upon the LLM to make sense of a set of tokens that are arranged according to images rather than elements in the 3D scene. As a result, it struggles to answer queries about specific 3D locations (e.g., \"What is here?\") and questions requiring a holistic understanding of the entire 3D scene, where the answer depends on establishing correspondences between objects across multiple images (e.g., \"What is opposite the door at the other end of the room\"?). In this work, we aim to address these issues by consolidating tokens in 3D before providing them to the LLM." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 425, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 295, + 567 + ], + "type": "text", + "content": "Our approach is called SplatTalk. Given a set of posed RGB images, we integrate language features extracted from the RGB images to build a 3D Gaussian Splitting representation, from which 3D tokens can be extracted for direct input into a pretrained LLM to answer challenging 3D VQA queries (see Fig. 1). The main research contributions are the new methods for 1) integrating language features into the Gaussian representation, and 2) extracting tokens that can be provided directly to an LLM. These methods avoid the issues of concurrent work, ChatSplat [6], which produces tokens that are not directly compatible with an LLM, making language-based reasoning more difficult." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": "Overall, SplatTalk has a unique combination of three advantages. First, it does not require any 3D surface representation as input. Second, it leverages the full power of a pre-trained LLM to answer difficult 3D VQA questions. Third, it produces a concise 3D representation that can be queried and/or edited in 3D (e.g., cropped to a 3D region, rendered from a novel view, etc.). During experiments with 3D VQA benchmarks, we find that it outperforms 3D models trained only for the benchmark tasks, demonstrating the value of utilizing pre-trained LLMs. SplatTalk further outperforms prior image-only methods using LLMs, showing the value of building a 3D scene representation. Finally," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "it even achieves competitive performance with SOTA 3D LMMs that utilize 3D scene representation inputs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 326, + 97, + 484, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 97, + 484, + 108 + ], + "spans": [ + { + "bbox": [ + 326, + 97, + 484, + 108 + ], + "type": "text", + "content": "In summary, our key contributions are:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 109, + 555, + 263 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 313, + 109, + 555, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 555, + 156 + ], + "type": "text", + "content": "- We introduce the first self-supervised 3D Gaussian-based method for large-scale zero-shot 3D VQA, SplatTalk, requiring only multi-view images without explicit depth, point clouds or 3D-language annotated supervision." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 156, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 553, + 191 + ], + "type": "text", + "content": "- We show that mean feature extraction from 3D Gaussians encodes holistic scene concepts for 3D VQA, supported by theoretical justification." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 192, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 553, + 215 + ], + "type": "text", + "content": "- We introduce entropy-adaptive token sampling, leading to better 3D VQA performance without additional training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 216, + 553, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 553, + 263 + ], + "type": "text", + "content": "- We demonstrate that SplatTalk outperforms 2D LMMs and achieves performance on par with 3D models using point clouds on various 3D VQA benchmarks, demonstrating the effectiveness of our 3D-aware representations." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 274, + 400, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 274, + 400, + 286 + ], + "spans": [ + { + "bbox": [ + 314, + 274, + 400, + 286 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 294, + 555, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 555, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 555, + 474 + ], + "type": "text", + "content": "2D LMMs. With advancements in large language models (LLMs), the ability of models to comprehend images and respond to language prompts related to their content has improved significantly. Notably, SOTA models [1, 11, 21, 30, 32] have demonstrated remarkable capabilities in processing multiple images of a 3D scene and answering questions about the shared content across these images. Although multiple RGB images inherently contain 3D information, they do not explicitly model 3D understanding, potentially leading to sub-optimal performance on tasks that require explicit 3D reasoning. In this work, we build upon the strong performance of 2D vision-language models (2D-VLMs), particularly LLaVA-OneVision [21], and integrate them with the 3D priors provided by the 3DGS framework to improve 3D spatial understanding capabilities." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 474, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 713 + ], + "type": "text", + "content": "3D LMMs. Recent advancements in 3D indoor scene understanding have explored integrating 3D priors into LLMs using inputs such as 3D point clouds [7, 17] or multi-view RGB images combined with depth or 3D point clouds [48]. These approaches employ various strategies to embed 3D structural information and language features into a shared representation space. While most methods [17, 48] fine-tune pretrained visual encoders (both 2D and 3D) alongside LoRA parameters or even the full set of LLM parameters, others [18, 24] opt for a more lightweight approach, training only a projection layer to map visual features into the language feature space. In contrast, our method does not require any explicit depth or 3D inputs. Instead, we leverage the 3DGS framework, which has shown strong capabilities in representing 3D scenes using only multi-view RGB images. Generalizable 3D Gaussian Splatting. This body of work explores generalizable 3DGS as an alternative to the perscene optimization paradigm used in traditional 3DGS methods, aiming to improve novel view synthesis across unseen scenes. Our method builds on FreeSplat [43], which is" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4713" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 71, + 410, + 237 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 410, + 237 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 410, + 237 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 410, + 237 + ], + "type": "image", + "image_path": "84c7612776e61cc762c577c9f35bae46a16175dfc583846ddcb7ca4e1203f7c8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 245, + 555, + 323 + ], + "lines": [ + { + "bbox": [ + 54, + 245, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 54, + 245, + 555, + 323 + ], + "type": "text", + "content": "Figure 2. Overview. Left: During the self-supervised 3D-language Gaussian Splitting training phase, multiple RGB input views are first encoded into Gaussian latent features (Gaussian triplets). These latent features are then decoded into Gaussian parameters for rendering, along with a low-dimensional visual-language feature " + }, + { + "bbox": [ + 54, + 245, + 555, + 323 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 54, + 245, + 555, + 323 + ], + "type": "text", + "content": ". To ensure proper supervision of this low-dimensional feature, we train an autoencoder that maps the high-dimensional, unbounded features obtained from LLaVA-OV, specifically, the visual tokens serving as direct inputs to the LLM, onto a low-dimensional hypersphere space. Right: During 3D VQA inference, visual-language features are directly extracted from the 3D Gaussians. These features are then mapped back to the original high-dimensional space using the pretrained decoder and subsequently used as direct visual token inputs to the LLM. LoRA fine-tuning of the LLM is optional." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 411, + 71, + 552, + 235 + ], + "blocks": [ + { + "bbox": [ + 411, + 71, + 552, + 235 + ], + "lines": [ + { + "bbox": [ + 411, + 71, + 552, + 235 + ], + "spans": [ + { + "bbox": [ + 411, + 71, + 552, + 235 + ], + "type": "image", + "image_path": "85c094a9cea333eefbe69806bb71b0124ee4c9f8bfaf92e316236e14d9993baf.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 329, + 296, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 329, + 296, + 508 + ], + "spans": [ + { + "bbox": [ + 54, + 329, + 296, + 508 + ], + "type": "text", + "content": "designed for sparse-view synthesis in indoor scenes. The method begins by extracting multi-scale features from input RGB images using CNN backbones. It then constructs adaptive cost volumes from neighboring views to predict depth maps. These predicted depth maps are used to backproject 2D CNN-encoded features into 3D Gaussian triplets " + }, + { + "bbox": [ + 54, + 329, + 296, + 508 + ], + "type": "inline_equation", + "content": "(\\mu, \\omega, f)" + }, + { + "bbox": [ + 54, + 329, + 296, + 508 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 54, + 329, + 296, + 508 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 54, + 329, + 296, + 508 + ], + "type": "text", + "content": " denotes the Gaussian's center, " + }, + { + "bbox": [ + 54, + 329, + 296, + 508 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 54, + 329, + 296, + 508 + ], + "type": "text", + "content": " the weight, and " + }, + { + "bbox": [ + 54, + 329, + 296, + 508 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 54, + 329, + 296, + 508 + ], + "type": "text", + "content": " the latent feature. To integrate multi-view information, FreeSplat introduces a Pixel-wise Triplet Fusion (PTF) module that uses a lightweight GRU to align and fuse overlapping 3D Gaussians across views efficiently. Finally, the fused triplets are passed through an MLP to decode rendering parameters, " + }, + { + "bbox": [ + 54, + 329, + 296, + 508 + ], + "type": "inline_equation", + "content": "\\Sigma, \\alpha, s" + }, + { + "bbox": [ + 54, + 329, + 296, + 508 + ], + "type": "text", + "content": ". Our approach extends the FreeSplat framework to enable the learning of generalizable 3D Gaussian language fields." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 509, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 297, + 714 + ], + "type": "text", + "content": "Similar to FreeSplat [43], [5, 8] leverage multi-view stereo (MVS) cues to refine depth estimation and improve geometric consistency in 3D Gaussian representations. These methods primarily focus on RGB novel view synthesis from sparse inputs (2-10 views). Our work builds on the SOTA model, FreeSplat [43], but diverges significantly in both objectives and methodology. First, while previous methods focus on photorealistic RGB reconstruction, we aim at semantic 3D scene understanding. Second, instead of reconstructing only a localized portion of the scene as shown in these works, our goal is to represent entire 3D environments, capturing both spatial and semantic relationships holistically. Language-guided 3D Gaussian Splatting. LangSplat [35] is one of the first approaches that integrates semantic features into 3DGS. However, this method requires training on individual scenes, which is significantly expensive, and lacks generalization to unseen scenes. Similarly, follow-up" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 329, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 555, + 449 + ], + "type": "text", + "content": "works [23, 45] explore 3DGS for open-vocabulary object segmentation but do not focus on reasoning tasks like 3D VQA. Further, they primarily rely on scene-specific training, limiting their scalability. In contrast, we propose a fully self-supervised framework that learns 3D-aware visual representations without requiring per-scene training. While some works [12, 42] explores a generalizable 3DGS and semantic modeling approach, they primarily focus on open-vocabulary segmentation, without addressing the broader challenges of free-form language reasoning required for 3D VQA." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 449, + 556, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 594 + ], + "type": "text", + "content": "ChatSplat [6] is a concurrent work that explores natural language conversation within the 3DGS framework. However, it optimizes on a per-scene basis and requires rendering 3D features into 2D feature maps during inference. Additionally, its training pipeline is not fully end-to-end, as it first trains the RGB Gaussians separately before incorporating language features. In contrast, our approach leverages scene latent features that encapsulate both RGB and semantic information, enabling the joint optimization of both modalities in a fully end-to-end manner. Our approach extends the FreeSplat framework to enable the learning of generalizable 3D Gaussian language fields." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 607, + 370, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 607, + 370, + 620 + ], + "spans": [ + { + "bbox": [ + 314, + 607, + 370, + 620 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 628, + 554, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 554, + 700 + ], + "type": "text", + "content": "The problem we address is formulated as follows: Given a sequence of " + }, + { + "bbox": [ + 313, + 628, + 554, + 700 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 628, + 554, + 700 + ], + "type": "text", + "content": " posed RGB images captured from a 3D indoor scene, along with language prompts related to the scene's content, the goal is to generate the corresponding language responses. That is, we address 3D VQA tasks with solely posed RGB images as inputs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 326, + 701, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 554, + 714 + ], + "type": "text", + "content": "We provide an overview of our approach in Fig. 2, which" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4714" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 73, + 299, + 257 + ], + "blocks": [ + { + "bbox": [ + 61, + 73, + 299, + 257 + ], + "lines": [ + { + "bbox": [ + 61, + 73, + 299, + 257 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 299, + 257 + ], + "type": "image", + "image_path": "396b4405b322ed870007efe67c8c51355a6dbc783b7af71262143472d6623b90.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 269, + 555, + 326 + ], + "lines": [ + { + "bbox": [ + 54, + 269, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 54, + 269, + 555, + 326 + ], + "type": "text", + "content": "Figure 3. Qualitative results on ScanQA, scene0030_00 and scene0084_00. We compare responses from LLaVA-OV, our model (Ours), and the ground truth (GT) for spatial reasoning questions in 3D VQA. Each scene highlights the referenced objects with red circles, and key relative objects are marked in blue. The answers from each model are displayed in color-coded speech bubbles: LLaVA-OV (brown), Ours (blue), and GT (green). With its 3D-aware representation, our model exhibits improved spatial reasoning capabilities, accurately identifying relationships between objects that are far apart and may never co-occur in the same image (e.g., door and window). More examples in Supp." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 547, + 257 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 547, + 257 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 547, + 257 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 547, + 257 + ], + "type": "image", + "image_path": "cd69f61788950ca6a3f9c919091ec419ab45ec136019317853b7eaa878c619d4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 331, + 297, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 331, + 297, + 522 + ], + "spans": [ + { + "bbox": [ + 54, + 331, + 297, + 522 + ], + "type": "text", + "content": "consists of three main steps: 1) Training the feature autoencoder, 2) Self-supervised 3D-Language Gaussian Splitting Training and 3) 3D VQA Inference. First, we extract the high-dimensional language features from LLaVA-OV [21] for each RGB image by using the pretrained image encoder and multimodal projector. Since these features are high-dimensional, unbounded and sparse, we train an autoencoder to project them into a more compact, low-dimensional space. These low-dimensional features then serve as pseudo groundtruth for training the language representations in SplatTalk. SplatTalk architecture is a feed-forward network consisting of a Gaussian encoder and decoder that predicts Gaussian parameters for rendering. During 3D VQA inference stage, we extract the trained language features from the 3D Gaussians and directly feed them into the LLM, along with the language prompt, to generate free-form responses." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 529, + 224, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 224, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 224, + 541 + ], + "type": "text", + "content": "3.1. Integrating Language Features" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 545, + 297, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 297, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 297, + 653 + ], + "type": "text", + "content": "An important step of our computational pipeline is the self-supervised method to integrate language features into a generalizable 3DGS framework. We identify four key factors that shape our approach for this step: (1) Leveraging features that align with the LLM's input space for direct language reasoning, (2) Efficient feature dimensionality reduction for effective CUDA-based rendering, and (3) Joint training of RGB and semantic features for coherent multimodal representation learning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Visual Tokens for 2D Pseudo Ground-truth Features. We extract the visual tokens used as inputs to the LLM for training our 3DGS framework. Our empirical findings indicate that extracting features directly from the visual encoder for training 3DGS requires retraining the multimodal projector" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 331, + 556, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 556, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 556, + 498 + ], + "type": "text", + "content": "layer which aligns the visual with language spaces. This observation is further corroborated by [6], which demonstrates that features extracted before the multimodal projector cannot be seamlessly integrated into the language model. A possible explanation is that the image encoder's native feature space (before the projector) is high-dimensional and unbounded, leading to sparser features with greater variance across different dimensions. As a result, the projector becomes more sensitive to subtle variations in the input features, making it difficult to directly map them into a token representation that aligns with the LLM's understanding. In contrast, the visual tokens generated after the projector are already well-aligned with the language model's latent space, allowing the LLM to interpret and reason with them directly." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 510, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 557, + 714 + ], + "type": "text", + "content": "Feature Dimensionality Reduction. Building on the previous factor, directly optimizing these language embeddings within a differentiable rendering pipeline is challenging because they are extremely high-dimensional (e.g., 3584 dimensions), sparse, and contain unbounded values, making learning both unstable and impractical. Therefore, we learn a generalizable autoencoder that projects these sparse, high-dimensional visual tokens into a compact unit hypersphere with much lower dimension (e.g., 256 dimensions), ensuring feature smoothness. Note that the language representations from LLaVA, which support reasoning and free-form language comprehension, are significantly more complex than those from CLIP or segmentation-based models like SAM. Different from prior works that compress semantic features into extremely low-dimensional vectors (e.g., 3-16 dimensions [35, 45]), which simplifies rendering but results in significant information loss, our approach prior" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4715" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 116, + 135, + 492, + 485 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "text", + "content": "Table 1. Evaluation of 3D VQA on ScanQA [2] and SQA3D [31] datasets. We compare SplatTalk against various baselines, including 3D LMMs, which take point clouds (PC) or both point clouds and images " + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "inline_equation", + "content": "(\\mathrm{PC} + \\mathrm{I})" + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "text", + "content": " as visual inputs, and 2D LMM-based models, which rely solely on images (I). As a 2D LMM-based approach, SplatTalk outperforms other 2D LMM-based models, surpasses some 3D LMMs, and achieves competitive performance against the strongest 3D LMM baselines. Our models are highlighted in blue, bold indicates best performance among the 2D-LMM methods." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 116, + 135, + 492, + 485 + ], + "lines": [ + { + "bbox": [ + 116, + 135, + 492, + 485 + ], + "spans": [ + { + "bbox": [ + 116, + 135, + 492, + 485 + ], + "type": "table", + "html": "
MethodModalityScanQA (val)SQA3D (test)
CIDErMETEORROUGEEM@1EM@1-REM@1EM@1-R
Specialist Models
ScanQA [2]PC64.913.133.121.1-47.2-
ScanRefer+MCan [46]PC55.411.530.018.6---
3D-VisTA [49]PC69.613.935.722.4-48.5-
SQA3D [31]PC-----46.6-
Generalist 3D LMMs
Chat-Scene [16]PC+I87.718.041.621.6-54.657.5
Scene-LLM [13]PC+I80.015.835.9--53.6-
Grounded 3D-LLM [9]PC72.7------
Chat-3D-v2 [44]PC77.116.140.1--54.7-
Finetuned 3D LMMs
3D-LLM [15]PC+I69.414.535.720.5---
LEO [17]PC+I101.420.049.224.547.650.052.4
Scene-LLM [13]PC+I80.016.640.027.2-54.2-
LL3DA [7]PC76.815.937.3----
Chat-3D [44]PC53.211.928.5----
2D LMM-Based
VideoChat2 [25]I49.29.528.219.2-37.3-
GPT-4V [32]I59.613.533.4----
Claude [1]I57.710.029.3----
LLaVA-NeXT-Video [22]I46.29.127.8--34.2-
LLaVA-OV [21]I50.013.029.415.632.525.435.8
SplatTalkI61.714.232.717.132.226.141.3
SplatTalk-ScanQA-FTI77.115.438.722.338.342.044.7
SplatTalk-3DVQA-FTI77.515.638.522.438.347.649.4
", + "image_path": "2d9625c50e510f125b8027ac5a4e4b090d7e7089e298e11ac1a3613e46dd76da.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 495, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 295, + 567 + ], + "type": "text", + "content": "itizes preserving rich language features by maintaining a higher-dimensional representation (e.g., 256). To efficiently render high-dimensional feature maps on CUDA, we draw inspiration from [47] and implement a parallel CUDA rasterizer pipeline, where semantic features and RGB are rendered simultaneously, using shared Gaussian parameters." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "Unlike previous methods [6, 35] that train a separate autoencoder for each individual scene, we train a single autoencoder across all training scenes, each consisting of multiple frames. This approach improves the autoencoder's generalizability, reducing computational overhead by eliminating the need for retraining to compress features from unseen scenes. Training the autoencoder separately from the 3DGS model ensures that it is dedicated solely to compressing high-dimensional feature vectors while preserving the essential scene information without being influenced by the complexities of the rendering pipeline. Our empirical findings further indicate that jointly training a feature de" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 495, + 555, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 555, + 531 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 555, + 531 + ], + "type": "text", + "content": "coder to project low-dimensional embeddings back to the high-dimensional space introduces significant instabilities in the training process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "Joint-training RGB and Language. In contrast to previous approaches that first train RGB Gaussians and then freeze them before integrating language features [6, 35], we adopt a joint training strategy, where RGB and language features are learned simultaneously. This is due to the fact that visual RGB and semantic features play a complementary role in effectively representing 3D scenes [33]. In feedforward models like FreeSplat, where the scene is first encoded into a Gaussian latent space before being decoded into specific Gaussian parameters for RGB rendering, these Gaussian latent features inherently capture holistic scene information, including both appearance (RGB) and semantics. Based on this observation, we introduce a new Gaussian parameter decoder head that predicts semantic features for each 3D Gaussian alongside rendering parameters, enabling joint op-" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4716" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 94, + 113, + 515, + 196 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 2. Evaluation of 3D VQA on MSR3D [27]. We compare SplatTalk against various baselines, all evaluated in a zero-shot setting. T denotes models that rely solely on text inputs without interleaving images, PC indicates models that use point cloud as the visual input, and I represents models that process only image-based inputs. Our approach, highlighted in blue, outperforms all baselines." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 94, + 113, + 515, + 196 + ], + "lines": [ + { + "bbox": [ + 94, + 113, + 515, + 196 + ], + "spans": [ + { + "bbox": [ + 94, + 113, + 515, + 196 + ], + "type": "table", + "html": "
ModelInputSceneSettingCountingExistenceAttributesSpatialNavigationOthersOverall
LEO [17]TPCzero-shot0.8015.511.87.32.315.37.8
LLaVA-OV [21]TIzero-shot18.631.224.819.516.736.324.0
SplatTalkTIzero-shot19.660.344.035.835.561.841.8
SplatTalk-ScanQA-FTTIzero-shot28.966.043.228.333.860.041.5
SplatTalk-3DVQA-FTTIzero-shot24.357.036.17.816.435.026.5
", + "image_path": "1d005e0cfa789f2fe6e39717f63eb987e351c52c66073cc60f401c7807de748e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 205, + 294, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 205, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 54, + 205, + 294, + 277 + ], + "type": "text", + "content": "tization. Unlike approaches that explicitly model these interactions [33, 39] using extra cross-attention or projection layers, SplatTalk implicitly integrates semantic and visual information within the Gaussian representation. This paradigm naturally facilitates multimodal interactions between RGB and language features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 279, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 279, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 54, + 279, + 295, + 361 + ], + "type": "text", + "content": "To incorporate more views and comprehensively cover the scene, we reduce the input resolution of the RGB images. This adjustment aligns well with the inherently low spatial resolution of the visual tokens extracted from LLaVA-OV, and also prevents the model from disproportionately allocating resources to modeling RGB features, thus encouraging a balanced representation of semantic and RGB information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 364, + 295, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 364, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 54, + 364, + 295, + 472 + ], + "type": "text", + "content": "These design choices establish our approach as a fundamentally novel method for integrating language features into 3DGS, grounded in first-principles observations rather than heuristic adaptations. By addressing the core challenges of multimodal alignment, feature compression, and joint optimization, our approach is not merely a trivial extension of existing works, but a carefully designed framework that ensures both effectiveness and stability in learning rich, spatially-aware semantic representations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 484, + 236, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 236, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 236, + 498 + ], + "type": "text", + "content": "3.2. Extracting 3D Language Features" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 502, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 502, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 502, + 297, + 714 + ], + "type": "text", + "content": "During inference, we directly extract language features at the mean position of each 3D Gaussian. Although the rendering process incorporates each Gaussian's covariance and opacity when projecting 3D features to 2D, we demonstrate that scene semantics can be effectively captured by the mean feature of each Gaussian, " + }, + { + "bbox": [ + 56, + 502, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i" + }, + { + "bbox": [ + 56, + 502, + 297, + 714 + ], + "type": "text", + "content": ", alone, without relying on these additional factors. This is supported by the conceptual similarity between our training optimization process and the Expectation-Maximization (EM) algorithm. For simplification, let " + }, + { + "bbox": [ + 56, + 502, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\sum_{t=1}^{T} \\| \\pmb{F}_t(x) - \\pmb{F}_t^{gt}(x) \\|^2" + }, + { + "bbox": [ + 56, + 502, + 297, + 714 + ], + "type": "text", + "content": " be the loss, where " + }, + { + "bbox": [ + 56, + 502, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{F}_t(x)" + }, + { + "bbox": [ + 56, + 502, + 297, + 714 + ], + "type": "text", + "content": " is the predicted rendered feature map of frame " + }, + { + "bbox": [ + 56, + 502, + 297, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 56, + 502, + 297, + 714 + ], + "type": "text", + "content": " at pixel " + }, + { + "bbox": [ + 56, + 502, + 297, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 56, + 502, + 297, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 502, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{F}^{gt}" + }, + { + "bbox": [ + 56, + 502, + 297, + 714 + ], + "type": "text", + "content": " is the pseudo ground-truth feature map obtained from LLaVA-OV. Replacing " + }, + { + "bbox": [ + 56, + 502, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{F}_t(x)" + }, + { + "bbox": [ + 56, + 502, + 297, + 714 + ], + "type": "text", + "content": " by the rendering equation and solving for the optimal " + }, + { + "bbox": [ + 56, + 502, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i^*" + }, + { + "bbox": [ + 56, + 502, + 297, + 714 + ], + "type": "text", + "content": ", we obtain " + }, + { + "bbox": [ + 56, + 502, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i^* = \\frac{\\sum_{t,x} R_i(t,x) \\pmb{F}_t^{gt}(x)}{\\sum_{t,x} R_i(t,x)}" + }, + { + "bbox": [ + 56, + 502, + 297, + 714 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 56, + 502, + 297, + 714 + ], + "type": "inline_equation", + "content": "R_i(t,x) = \\frac{\\alpha_i \\mathcal{N}(x; \\pmb{\\mu}_i, \\pmb{\\Sigma}_i)}{\\sum_j \\alpha_j \\mathcal{N}(x; \\pmb{\\mu}_j, \\pmb{\\Sigma}_j)}" + }, + { + "bbox": [ + 56, + 502, + 297, + 714 + ], + "type": "text", + "content": " is the Gaussian's contribution to pixel " + }, + { + "bbox": [ + 56, + 502, + 297, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 56, + 502, + 297, + 714 + ], + "type": "text", + "content": " at viewpoint " + }, + { + "bbox": [ + 56, + 502, + 297, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 56, + 502, + 297, + 714 + ], + "type": "text", + "content": ". This mirrors the EM procedure: estimating " + }, + { + "bbox": [ + 56, + 502, + 297, + 714 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 56, + 502, + 297, + 714 + ], + "type": "text", + "content": " by rendering (E-step)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 205, + 555, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 555, + 301 + ], + "type": "text", + "content": "and updating " + }, + { + "bbox": [ + 313, + 205, + 555, + 301 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 313, + 205, + 555, + 301 + ], + "type": "text", + "content": " by minimizing reconstruction loss (M-step). Consequently, the optimal 3D feature " + }, + { + "bbox": [ + 313, + 205, + 555, + 301 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{i}^{*}" + }, + { + "bbox": [ + 313, + 205, + 555, + 301 + ], + "type": "text", + "content": " for each Gaussian is a weighted sum of the 2D feature maps, and the set of 3D Gaussian features " + }, + { + "bbox": [ + 313, + 205, + 555, + 301 + ], + "type": "inline_equation", + "content": "\\{\\pmb{f}_{i}\\}_{i=1}^{N}" + }, + { + "bbox": [ + 313, + 205, + 555, + 301 + ], + "type": "text", + "content": " encodes the scene's semantic holistically from " + }, + { + "bbox": [ + 313, + 205, + 555, + 301 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 205, + 555, + 301 + ], + "type": "text", + "content": " 2D inputs. These features are then projected back into the original language embedding space using the pretrained decoder and directly provided as inputs to the LLM for downstream 3D-language tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 302, + 556, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 302, + 556, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 556, + 445 + ], + "type": "text", + "content": "In open-vocabulary segmentation, querying Gaussian centers is effective [4, 35, 45] because each Gaussian is assumed to represent a distinct, meaningful object or region. These methods use object-centric embeddings (e.g., SAM-masked CLIP or LSeg) and retrieve objects via cosine similarity to text queries. In contrast, 3D reasoning tasks lack this one-to-one correspondence. Each Gaussian encodes parts of a scene-level representation, where meaning stems from spatial and relational structure. Querying single Gaussians is thus insufficient for context-dependent reasoning. While our method also queries at Gaussian means, this is not trivial; our EM-based formulation offers a principled justification." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 455, + 424, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 424, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 424, + 468 + ], + "type": "text", + "content": "3.3.3D VQA Inference" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "type": "text", + "content": "Entropy Adaptive Gaussian Sampling. SplatTalk leverages triplet fusion adopted in FreeSplat [43], which reduces the number of Gaussians. However, complex scenes still require significantly more Gaussians than the LLM's visual token limit, and not all are equally informative. To address this, we propose entropy-based sampling: selecting the top " + }, + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "type": "text", + "content": " Gaussians with the highest language feature entropy, where " + }, + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "type": "text", + "content": " matches the LLM's visual token capacity." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "3D Tokens as Visual Inputs. After 3DGS training, each 3D Gaussian encodes a feature that is a weighted aggregation of 2D semantic information across all input views (Sec. 3.2). These features are also spatially grounded in their 3D coordinates. Because each token corresponds to a specific location in 3D space, its position is already implicitly encoded in the feature through the rendering and fusion process. Unlike 2D patches in a fixed grid, these 3D tokens form an unordered set, similar to a point cloud where spatial relationships are captured by the features themselves, not by sequence position. Therefore, we do not require explicit positional encodings or a fixed input order. The 3D tokens," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4717" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 97 + ], + "type": "text", + "content": "extracted by sampling features at each Gaussian's mean, are directly fed into the LLM." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 104, + 280, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 104, + 280, + 118 + ], + "spans": [ + { + "bbox": [ + 55, + 104, + 280, + 118 + ], + "type": "text", + "content": "3.4. Optional - Finetuning on 3D VQA Datasets" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 121, + 295, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 295, + 206 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 295, + 206 + ], + "type": "text", + "content": "While SplatTalk is capable of performing the 3D VQA task in a zero-shot manner, it also allows for flexible fine-tuning on 3D VQA datasets to enhance alignment and reasoning capabilities. Rather than training the model end-to-end, which is both computationally expensive and prone to overfitting on limited datasets, we optimize only the LoRA parameters applied to the LLM during fine-tuning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 217, + 136, + 231 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 136, + 231 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 136, + 231 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 237, + 188, + 250 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 237, + 188, + 250 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 188, + 250 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 255, + 296, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 386 + ], + "type": "text", + "content": "We train SplatTalk with 500 scenes from the ScanQA training set. The initial self-supervised stage focuses purely on learning the 3D scene representation, without leveraging any text-annotated information. For each scene, we uniformly sample 100 views around the scene and extract their per-patch features by passing them through the visual encoder of LLaVA-OV followed by the multi-modal projector. The subsequent SplatTalk model is trained on a single NVIDIA H100 (80GB) GPU. Both the LoRA fine-tuning and inference are also conducted on a single H100. More details can be found in Supp. Material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 394, + 176, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 394, + 176, + 406 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 176, + 406 + ], + "type": "text", + "content": "4.2. Datasets and Metrics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 412, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 296, + 495 + ], + "type": "text", + "content": "Evaluation Dataset. We evaluate our method on ScanNet-based QA datasets: ScanQA [2], SQA3D [31], and MSR3D [27]. ScanQA is a benchmark designed for question answering in 3D indoor scenes, containing detailed spatial and semantic information extracted from ScanNet. SQA3D and MSR3D further extend this by introducing scenarios that describe an embodied agent's 3D position and state." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 496, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 296, + 568 + ], + "type": "text", + "content": "Metrics. For ScanQA and SQA3D, we evaluate performance using standard n-gram-based metrics, specifically CIDEr [38], METEOR [3], ROUGE [26], EM@1 [14], and EM@1-Refined for ScanQA, and EM@1 and EM@1-Refined for SQA3D. In contrast, for MSR3D, we utilize the LLM-Match score [29] as recommended by the authors." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 575, + 295, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 295, + 589 + ], + "type": "text", + "content": "4.3. Performance on 3D Question Answering Tasks" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "ScanQA. We evaluate SplatTalk's performance against several model families, each representing different levels of specialization and generalization. \"Specialist Models\" refer to models specifically trained to solve ScanQA without exposure to other tasks. \"Generalist 3D LMMs\" consist of models trained across a diverse range of 3D vision-language tasks, enabling broader generalization. \"Finetuned 3D LMMs\" are models that have been further fine-tuned on the ScanQA training set to optimize performance for this particular dataset. Lastly, \"2D LMM-Based Models\" include" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "models that process only image inputs without leveraging depth or explicit 3D information. SplatTalk and its fine-tuned variants fall within this category." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 109, + 555, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 555, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 555, + 252 + ], + "type": "text", + "content": "We evaluate three different variants of our model on this task. The base model (SplatTalk) is trained solely to encode the semantic representation of the scene from multiple RGB images using 3DGS, without exposure to question-answering tasks. Note that this model is trained entirely in a self-supervised manner, without any access to language supervision or textual annotations during training. To further adapt it to downstream 3D VQA tasks, we fine-tune this base model on the ScanQA training set, referring to it as SplatTalk-ScanQA-FT, and on the combined training sets of ScanQA and SQA3D, denoting as SplatTalk-3DVQA-FT. We present results in Table 1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 253, + 556, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 556, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 556, + 445 + ], + "type": "text", + "content": "Compared to LLaVA-OV, whose 2D language feature maps are used to train SplatTalk, our base model, which benefits from a more 3D-aware representation due to the 3D Gaussian training process, achieves significantly higher performance (e.g. 61.7 vs 50.0 on CIDEr). Our language finetuned model variants, SplatTalk-ScanQA-FT and SplatTalk-3DVQA-FT, improve performance further. Notably, despite not being trained with depth supervision or explicit 3D information, our models achieve comparable results to both Generalist 3D LMMs and Finetuned 3D LMMs, while outperforming Specialist Models. Additionally, among 2D LMM-Based approaches, our base model without being fine-tuned with 3D VQA datasets outperforms other methods, with significantly higher performance on some metrics, including both open-source and closed-source models. We show qualitative results in Fig. 3." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 445, + 556, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 556, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 556, + 577 + ], + "type": "text", + "content": "SQA3D. Similar to our evaluation on ScanQA, we compare our model against both 3D LMMs and 2D LMMs. Compared to 2D LLaVA-OV features, our base model with 3D-aware features achieves a significant performance gain, highlighting the advantage of incorporating 3D information via 3DGS which aligns with our findings on ScanQA. Further finetuning on ScanQA alone and on the combined ScanQA + SQA3D dataset leads to substantial improvements. Notably, SplatTalk-3DVQA-FT surpasses specialist models that leverage 3D point cloud inputs, demonstrating the strength of our approach in 3D scene understanding." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 578, + 556, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 556, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 556, + 673 + ], + "type": "text", + "content": "MSR3D. We evaluated our model's performance against LLaVA-OV and " + }, + { + "bbox": [ + 313, + 578, + 556, + 673 + ], + "type": "inline_equation", + "content": "\\mathsf{LEO}^1" + }, + { + "bbox": [ + 313, + 578, + 556, + 673 + ], + "type": "text", + "content": " [17] on the Scannet split of MSR3D. Importantly, all comparisons are conducted in a zero-shot setting, as none of the models have been trained on MSR3D. This is an interleaved dataset where inputs consist of both text and images. However, since all models are tested in a zero-shot manner, they rely solely on text inputs, where the agent's state descriptions and the question are combined into" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "text", + "content": "1The reported number reflects performance across all splits of MSR3D. However, since LEO was trained on ScanNet and 3RScan, its performance remains largely consistent across splits." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4718" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 101, + 99, + 261, + 150 + ], + "blocks": [ + { + "bbox": [ + 55, + 64, + 307, + 99 + ], + "lines": [ + { + "bbox": [ + 55, + 64, + 307, + 99 + ], + "spans": [ + { + "bbox": [ + 55, + 64, + 307, + 99 + ], + "type": "text", + "content": "Table 3. Ablation Study on Gaussian Sampling Strategies. Our entropy-based sampling strategy consistently outperforms across multiple metrics." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 101, + 99, + 261, + 150 + ], + "lines": [ + { + "bbox": [ + 101, + 99, + 261, + 150 + ], + "spans": [ + { + "bbox": [ + 101, + 99, + 261, + 150 + ], + "type": "table", + "html": "
Sampling MethodEM@1METEORROUGESPICEAvg
Random17.013.862.114.026.73
Point Density16.214.059.714.126.00
FPS17.713.862.013.626.78
Entropy17.114.661.714.226.90
", + "image_path": "3f658b393c12ad03731f4020caaa97273d25d391926e1323495a2696c5003985.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 165, + 295, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 165, + 295, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 165, + 295, + 201 + ], + "type": "text", + "content": "a single text prompt. On the visual processing side, LEO operates on 3D point cloud inputs, whereas both LLaVA-OV and our model rely only on RGB images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 202, + 296, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 202, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 296, + 285 + ], + "type": "text", + "content": "The results, presented in Table 2, demonstrate that our base model outperforms LLaVA-OV across all question types. Notably, aside from the \"counting\" category, our base model achieves approximately twice the performance of LLaVA-OV on other question types. Both models significantly outperform LEO, which struggles due to the absence of fine-tuning on this dataset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 287, + 296, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 287, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 296, + 441 + ], + "type": "text", + "content": "Interestingly, we observe a slight performance drop when comparing SplatTalk with SplatTalk-ScanQA-FT. However, this drop becomes much more pronounced when compared with SplatTalk-3DVQA-FT. A plausible explanation is that fine-tuning on other QA datasets introduces overfitting, causing the model to adapt too specifically to those datasets. Furthermore, the observed performance drop suggests that MSR3D differs significantly from ScanQA and SQA3D, making direct knowledge transfer less effective. This distinction may also explain why LEO struggles to generalize in a zero-shot setting. Despite being trained on multiple QA datasets, none of them closely resemble MSR3D, limiting its ability to adapt without fine-tuning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 453, + 156, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 156, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 156, + 464 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 472, + 296, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 591 + ], + "type": "text", + "content": "Gaussian Sampling Strategies. We explore various Gaussian sampling strategies for selecting tokens as inputs to the LLM, including: (1) Random Sampling—Gaussians are selected randomly across the scene. (2) Point Density Adaptive Sampling—Gaussians are sampled proportionally to the density of 3D points. (3) Furthest Point Sampling (FPS)—Gaussians are iteratively selected at locations maximally distant from previously chosen points. (4) Entropy Adaptive Sampling—Gaussians are prioritized based on language feature entropy, selecting points with the highest uncertainty." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 593, + 296, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 296, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 296, + 676 + ], + "type": "text", + "content": "We conduct experiments on the ScanQA validation set. The results, presented in Table 3, demonstrate that entropy-adaptive sampling outperforms other sampling strategies, followed by FPS. This finding highlights entropy-adaptive sampling as the most effective approach for selecting informative tokens, making it the optimal strategy for downstream 3D VQA tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Length of Visual Tokens. We evaluate the performance of SplatTalk on ScanQA, SQA3D, and MSR3D under two different settings: (1) using only 1 image-worth of visual" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 329, + 105, + 545, + 154 + ], + "blocks": [ + { + "bbox": [ + 311, + 60, + 563, + 104 + ], + "lines": [ + { + "bbox": [ + 311, + 60, + 563, + 104 + ], + "spans": [ + { + "bbox": [ + 311, + 60, + 563, + 104 + ], + "type": "text", + "content": "Table 4. Ablation Study on Visual Input Length. We compare model performance using 729 tokens (equivalent to a single image) versus 32,076 tokens (corresponding to 44 images), analyzing the impact of visual context on reasoning." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 329, + 105, + 545, + 154 + ], + "lines": [ + { + "bbox": [ + 329, + 105, + 545, + 154 + ], + "spans": [ + { + "bbox": [ + 329, + 105, + 545, + 154 + ], + "type": "table", + "html": "
Visual LengthScanQA (val)SQA3D (test)MSR3D (test)
EM@1EM@1-REM@1EM@1-REM@1EM@1-R
729 tokens16.230.725.838.07.917.7
32,076 tokens17.132.226.141.314.123.3
", + "image_path": "a22b545a6a0f8e7679ec8fc4158821a00592774e6d8c8dbaf430218b51695621.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 165, + 553, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 165, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 313, + 165, + 553, + 283 + ], + "type": "text", + "content": "tokens (729 tokens) and (2) using the full context length corresponding to 44 images-worth of visual tokens (32,076 tokens). The results, presented in Table 4, evaluate model performance using EM@1 and EM@1-Refined, as these metrics offer a reliable measure of how precisely model predictions align with ground truth answers, both in literal accuracy and permissible variations in phrasing. We employ entropy-adaptive sampling (see Sec. 3.3), which prioritizes high-information tokens and remains deterministic given a fixed set of visual tokens." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 285, + 554, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 554, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 554, + 405 + ], + "type": "text", + "content": "A key observation is that performance improved across all datasets as the number of input tokens increased, with MSR3D showing the largest gain, nearly doubling in EM@1. In contrast, ScanQA and SQA3D saw only marginal improvements, even with significantly more tokens. This suggests that additional scene context may not substantially improve performance, possibly due to: (1) Dataset limitations where the tasks may not require rich spatial context; or (2) Model limitations where the model may be unable to effectively leverage extra visual tokens for spatial reasoning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 418, + 388, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 388, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 388, + 431 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 440, + 555, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 440, + 555, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 440, + 555, + 583 + ], + "type": "text", + "content": "In this work, we introduce the first method to integrate free-form language features into a 3D Gaussian Splitting framework, enabling zero-shot generalization for 3D Visual Question Answering (3D VQA) without requiring 3D point clouds or depth maps. Our approach surpasses methods that rely solely on 2D language feature maps and achieves performance on par with 3D LMMs that explicitly utilize 3D structural information across multiple 3D VQA benchmarks. We hope this work inspires further research into leveraging multi-view RGB inputs to encapsulate rich semantic representations of 3D scenes, ultimately enhancing 3D VQA capabilities without the reliance on explicit geometric priors." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 609, + 407, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 407, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 407, + 623 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 629, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 712 + ], + "type": "text", + "content": "We thank the creators of the datasets and codebases used in our work, ScanQA, SQA3D, MSR3D, ScanNet, LLaVA-OV, and FreeSplat, for making their resources publicly available. We are also grateful to the anonymous reviewers for their constructive feedback during the review process. Special thanks to Xiang Li and James M. Rehg for the insightful discussions." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4719" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 112 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 112 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 112 + ], + "type": "text", + "content": "[1] Anthropic. Claude 3. https://www.anthropic.com/index/claude, 2024. Accessed: 2024-02-28. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 114, + 296, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 114, + 296, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 114, + 296, + 167 + ], + "type": "text", + "content": "[2] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19129-19139, 2022. 5, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 170, + 295, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 170, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 62, + 170, + 295, + 224 + ], + "type": "text", + "content": "[3] Satanjeev Banerjee and Alon Lavie. METEOR: An Automatic Metric for MT Evaluation with Improved Correlation with Human Judgments. In Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization, pages 65-72, 2005. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 225, + 296, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 225, + 296, + 278 + ], + "spans": [ + { + "bbox": [ + 62, + 225, + 296, + 278 + ], + "type": "text", + "content": "[4] Yash Bhalgat, Iro Laina, Joao F Henriques, Andrew Zisserman, and Andrea Vedaldi. N2f2: Hierarchical scene understanding with nested neural feature fields. In European Conference on Computer Vision, pages 197-214. Springer, 2024. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 281, + 296, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 281, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 62, + 281, + 296, + 335 + ], + "type": "text", + "content": "[5] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19457-19467, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 336, + 296, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 336, + 296, + 368 + ], + "spans": [ + { + "bbox": [ + 62, + 336, + 296, + 368 + ], + "type": "text", + "content": "[6] Hanlin Chen, Fangyin Wei, and Gim Hee Lee. Chat-splat: 3d conversational gaussian splatting. arXiv preprint arXiv:2412.00734, 2024. 2, 3, 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 369, + 296, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 296, + 435 + ], + "type": "text", + "content": "[7] Sijin Chen, Xin Chen, Chi Zhang, Mingsheng Li, Gang Yu, Hao Fei, Hongyuan Zhu, Jiayuan Fan, and Tao Chen. L13da: Visual interactive instruction tuning for omni-3d understanding reasoning and planning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26428-26438, 2024. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 437, + 296, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 437, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 437, + 296, + 491 + ], + "type": "text", + "content": "[8] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In European Conference on Computer Vision, pages 370-386. Springer, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 492, + 296, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 492, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 62, + 492, + 296, + 534 + ], + "type": "text", + "content": "[9] Yilun Chen, Shuai Yang, Haifeng Huang, Tai Wang, Runsen Xu, Ruiyuan Lyu, Dahua Lin, and Jiangmiao Pang. Grounded 3d-llm with referent tokens. arXiv preprint arXiv:2405.10370, 2024. 1, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 536, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 296, + 590 + ], + "type": "text", + "content": "[10] Francis Engelmann, Fabian Manhardt, Michael Niemeyer, Keisuke Tateno, Marc Pollefeys, and Federico Tombari. Opennerf: open set 3d neural scene segmentation with pixelwise features and rendered novel views. arXiv preprint arXiv:2404.03650, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 591, + 295, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 613 + ], + "type": "text", + "content": "[11] Gemini Team et al. Gemini: A family of highly capable multimodal models, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 614, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 295, + 668 + ], + "type": "text", + "content": "[12] Zhiwen Fan, Jian Zhang, Wenyan Cong, Peihao Wang, Renjie Li, Kairun Wen, Shijie Zhou, Achuta Kadambi, Zhangyang Wang, Danfei Xu, et al. Large spatial model: End-to-end unposed images to semantic 3d. Advances in Neural Information Processing Systems, 37:40212-40229, 2025. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[13] Rao Fu, Jingyu Liu, Xilun Chen, Yixin Nie, and Wenhan Xiong. Scene-llm: Extending language model for 3d visual understanding and reasoning. arXiv preprint arXiv:2403.11401, 2024. 1, 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 138 + ], + "type": "text", + "content": "[14] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the V in VQA Matter: Elevating the Role of Image Understanding in Visual Question Answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6904-6913, 2017. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 140, + 555, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 555, + 185 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 555, + 185 + ], + "type": "text", + "content": "[15] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems, 36:20482-20494, 2023. 1, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 186, + 555, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 555, + 239 + ], + "type": "text", + "content": "[16] Haifeng Huang, Yilun Chen, Zehan Wang, Rongjie Huang, Runsen Xu, Tai Wang, Luping Liu, Xize Cheng, Yang Zhao, Jiangmiao Pang, et al. Chat-scene: Bridging 3d scene and large language models with object identifiers. arXiv preprint arXiv:2312.08168, 2023.5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 241, + 555, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 241, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 317, + 241, + 555, + 295 + ], + "type": "text", + "content": "[17] Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. An embodied generalist agent in 3d world. arXiv preprint arXiv:2311.12871, 2023. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 297, + 554, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 554, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 554, + 331 + ], + "type": "text", + "content": "[18] Weitai Kang, Haifeng Huang, Yuzhang Shang, Mubarak Shah, and Yan Yan. Robin3d: Improving 3d large language model via robust instruction tuning, 2025. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 332, + 555, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 332, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 317, + 332, + 555, + 386 + ], + "type": "text", + "content": "[19] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 388, + 555, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 555, + 443 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 555, + 443 + ], + "type": "text", + "content": "[20] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 445, + 555, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 555, + 488 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 555, + 488 + ], + "type": "text", + "content": "[21] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 490, + 555, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 490, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 490, + 555, + 533 + ], + "type": "text", + "content": "[22] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava-last-interleave: Tackling multi-image, video, and 3d in large multimodal models, 2024. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 535, + 555, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 555, + 589 + ], + "type": "text", + "content": "[23] Hajjie Li, Yanmin Wu, Jiarui Meng, Qiankun Gao, Zhiyao Zhang, Ronggang Wang, and Jian Zhang. Instancegaussian: Appearance-semantic joint gaussian representation for 3d instance-level perception. arXiv preprint arXiv:2411.19235, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 591, + 555, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 591, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 591, + 555, + 645 + ], + "type": "text", + "content": "[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 647, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 555, + 712 + ], + "type": "text", + "content": "[25] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 5" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4720" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 106 + ], + "type": "text", + "content": "[26] Chin-Yew Lin. ROUGE: A Package for Automatic Evaluation of Summaries. In Proceedings of the Workshop on Text Summarization Branches Out, pages 74-81, 2004. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "type": "text", + "content": "[27] Xiongkun Linghu, Jiangyong Huang, Xuesong Niu, Xiaojian Shawn Ma, Baoxiong Jia, and Siyuan Huang. Multimodal situated reasoning in 3d scenes. Advances in Neural Information Processing Systems, 37:140903-140936, 2025. 1, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 196 + ], + "type": "text", + "content": "[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36:34892-34916, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 296, + 262 + ], + "type": "text", + "content": "[29] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. MMBench: Is your multi-modal model an all-around player? In Proceedings of the European Conference on Computer Vision (ECCV), 2024. Accepted as Oral Presentation. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 296, + 319 + ], + "type": "text", + "content": "[30] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, Yaofeng Sun, Chengqi Deng, Hanwei Xu, Zhenda Xie, and Chong Ruan. Deepseek-vl: Towards real-world vision-language understanding, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 296, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 296, + 365 + ], + "type": "text", + "content": "[31] Xiaojian Ma, Silong Yong, Zilong Zheng, Qing Li, Yitao Liang, Song-Chun Zhu, and Siyuan Huang. Sqa3d: Situated question answering in 3d scenes. arXiv preprint arXiv:2210.07474, 2022. 5, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 366, + 296, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 296, + 388 + ], + "type": "text", + "content": "[32] OpenAI. Gpt-4 with vision (gpt-4v). https://openai.com/research/gpt-4, 2023. Accessed: 2024-02-28. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 389, + 296, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 389, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 56, + 389, + 296, + 433 + ], + "type": "text", + "content": "[33] Qucheng Peng, Benjamin Planche, Zhongpai Gao, Meng Zheng, Anwesa Choudhuri, Terrence Chen, Chen Chen, and Ziyan Wu. 3d vision-language gaussian splatting. arXiv preprint arXiv:2410.07577, 2024. 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 296, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 296, + 490 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 296, + 490 + ], + "type": "text", + "content": "[34] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 815-824, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 491, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 491, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 491, + 296, + 544 + ], + "type": "text", + "content": "[35] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20051-20060, 2024. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 546, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 546, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 296, + 612 + ], + "type": "text", + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 613, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 296, + 679 + ], + "type": "text", + "content": "[37] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13009-13018, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 681, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 296, + 713 + ], + "type": "text", + "content": "[38] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. CIDEr: Consensus-based Image Description Evaluation. In Proceedings of the IEEE Conference on Computer" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 665 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 72, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 555, + 95 + ], + "type": "text", + "content": "Vision and Pattern Recognition (CVPR), pages 4566-4575, 2015. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "type": "text", + "content": "[39] Peihao Wang, Zhiwen Fan, Zhangyang Wang, Hao Su, Ravi Ramamoorthi, et al. Lift3d: Zero-shot lifting of any 2d vision model to 3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21367-21377, 2024. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 555, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 206 + ], + "type": "text", + "content": "[40] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Song XiXuan, et al. Cogvlm: Visual expert for pretrained language models. Advances in Neural Information Processing Systems, 37:121475-121499, 2025. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 208, + 555, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 555, + 251 + ], + "type": "text", + "content": "[41] Xingrui Wang, Wufei Ma, Zhuowan Li, Adam Kortyewski, and Alan L Yuille. 3d-aware visual question answering about parts, poses and occlusions. Advances in Neural Information Processing Systems, 36:58717-58735, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 253, + 554, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 554, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 554, + 285 + ], + "type": "text", + "content": "[42] Xingrui Wang, Cuiling Lan, Hanxin Zhu, Zhibo Chen, and Yan Lu. Gsemsplat: Generalizable semantic 3d gaussian splatting from uncalibrated image pairs, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 286, + 555, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 555, + 341 + ], + "type": "text", + "content": "[43] Yunsong Wang, Tianxin Huang, Hanlin Chen, and Gim Hee Lee. Freesplat: Generalizable 3d gaussian splatting towards free view synthesis of indoor scenes. Advances in Neural Information Processing Systems, 37:107326-107349, 2025. 2, 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 342, + 554, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 554, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 554, + 386 + ], + "type": "text", + "content": "[44] Zehan Wang, Haifeng Huang, Yang Zhao, Ziang Zhang, and Zhou Zhao. Chat-3d: Data-efficiently tuning large language model for universal dialogue of 3d scenes. arXiv preprint arXiv:2308.08769, 2023. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 387, + 555, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 387, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 387, + 555, + 453 + ], + "type": "text", + "content": "[45] Yanmin Wu, Jiarui Meng, Hajjie Li, Chenming Wu, Yahao Shi, Xinhua Cheng, Chen Zhao, Haocheng Feng, Errui Ding, Jingdong Wang, et al. Opengaussian: Towards point-level 3d gaussian-based open vocabulary understanding. Advances in Neural Information Processing Systems, 37:19114-19138, 2025. 3, 4, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 454, + 555, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 454, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 454, + 555, + 498 + ], + "type": "text", + "content": "[46] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6281–6290, 2019. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 498, + 555, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 555, + 564 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 555, + 564 + ], + "type": "text", + "content": "[47] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splattering to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676-21685, 2024. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 566, + 555, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 566, + 555, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 566, + 555, + 609 + ], + "type": "text", + "content": "[48] Chenming Zhu, Tai Wang, Wenwei Zhang, Jiangmiao Pang, and Xihui Liu. Llava-3d: A simple yet effective pathway to empowering lmm's with 3d-awareness. arXiv preprint arXiv:2409.18125, 2024. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 610, + 555, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 610, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 316, + 610, + 555, + 665 + ], + "type": "text", + "content": "[49] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2911-2921, 2023. 5" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "4721" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Split-and-Combine_ Enhancing Style Augmentation for Single Domain Generalization/3714ebba-fcba-4212-aaf3-03100ed83e52_content_list.json b/2025/Split-and-Combine_ Enhancing Style Augmentation for Single Domain Generalization/3714ebba-fcba-4212-aaf3-03100ed83e52_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..94676e688ad6cb3f7fe9215317055a41d29e977a --- /dev/null +++ b/2025/Split-and-Combine_ Enhancing Style Augmentation for Single Domain Generalization/3714ebba-fcba-4212-aaf3-03100ed83e52_content_list.json @@ -0,0 +1,1606 @@ +[ + { + "type": "text", + "text": "Split-and-Combine: Enhancing Style Augmentation for Single Domain Generalization", + "text_level": 1, + "bbox": [ + 142, + 128, + 854, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhen Zhang $^{1}$ , Shuai Yang $^{1\\dagger}$ , Qianlong Dang $^{2}$ , Zhize Wu $^{3}$ , Lichuan Gu $^{1\\dagger}$ $^{1}$ Anhui Agricultural University $^{2}$ Northwest A & F University $^{3}$ Hefei University", + "bbox": [ + 184, + 202, + 810, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhangz@stu.ahau.edu.cn, yangs@ahau.edu.cn, dangqianlong@nwafu.edu.cn,", + "bbox": [ + 191, + 241, + 800, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "wuzz@hfuu.edu.cn, glc@ahau.edu.cn", + "bbox": [ + 346, + 258, + 643, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Single domain generalization aims to learn a model with good generalization ability from a single source domain. Recent advances in this field have focused on increasing the diversity of the training data through style (e.g., color and texture) augmentation. However, most existing methods apply uniform perturbations to the entire image, failing to simulate complex images with multiple distinct stylistic regions. To address this, we propose a \"Split-And-Combine\" (SAC) strategy to enhance style diversity. Specifically, SAC first performs patch-aware augmentation, which splits an image into multiple patches and applies style augmentation independently to each patch, enabling distinct color variations across regions. Then, SAC combines these patches to reconstruct a complete image and applies adaptive random convolutions, which utilizes a deformable convolution layer with random and Gaussian filters to enhance texture diversity while preserving object integrity. Notably, SAC leverages entropy as a risk assessment criterion to adaptively determine whether a sample should undergo augmentation within the iterative process of random convolutions, preventing excessive augmentation. Furthermore, SAC introduces an energy-based distribution discrepancy score to quantify out-of-distribution likelihood, systematically expanding the augmented data's distribution. SAC can serve as a plug-and-play component to improve the performance of recent methods. Extensive experiments on four datasets demonstrate the effectiveness of SAC.", + "bbox": [ + 88, + 340, + 485, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 777, + 220, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Modern machine learning algorithms [11, 16, 30] have achieved remarkable performance under the assumption that the source (training) and target (test) domains share the same distribution. However, in reality, such an assumption is often violated due to factors such as variations in", + "bbox": [ + 89, + 803, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "illumination, background, or weather, which are known as the domain shift problem [41]. Unfortunately, these traditional algorithms often face significant performance degradation when there are distributional discrepancies between the source and target domains [24, 25, 38, 39, 47].", + "bbox": [ + 511, + 310, + 906, + 385 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To tackle this problem, multi-source domain generalization [19, 22, 28, 46] has been presented, aiming to learn a robust model by using knowledge from multiple different but related source domains without accessing target domain data. The mainstream works depend heavily on domain identifiers as supervision signals to capture domain invariant representations. However, the collection and annotation of data from multiple domains can be prohibitively expensive, making multi-source domain generalization less feasible in practical applications. To alleviate this issue, a more challenging and realistic setting known as single domain generalization [31, 33, 43] has emerged and garnered significant attention recently, which only leverages a single source domain to learn a model with good generalizability.", + "bbox": [ + 511, + 386, + 908, + 598 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The difficulty in single domain generalization arises from mitigating distributional discrepancies between the source and unknown target domains. In reality, it is observed that content features, such as object features, remain relatively stable despite environmental changes, whereas style features, such as backgrounds, exhibit considerable variability across different domains [1, 2]. For example, in animal image classification, fundamental elements like eyes, ears, and face shape constitute the content, whereas style elements include background and lighting conditions. Therefore, the key to effective single domain generalization is to eliminate spurious correlations between style features and labels. The current mainstream research on single domain generalization primarily focuses on increasing the capacity of the training data through the generation of novel domains, thereby reducing the model's reliance on style features due to the increased diversity of styles within the training data. One category of methods [3, 32, 45] perturb the original data using the back-propagated gradient to obtain augmented data. However, gradient-based perturba", + "bbox": [ + 511, + 599, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 107, + 886, + 236, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15616", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/13392f7c48309c7ff72b485abe5e59cec212dfcda5d5150947dbb820a1f30104.jpg", + "image_caption": [ + "(a) UDP" + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 222, + 152 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9066d5d6e3caf5b114b20fdbb3ab6b0fae3a724d0ffde415e917ab5933ee559b.jpg", + "image_caption": [ + "(b) PSDG" + ], + "image_footnote": [], + "bbox": [ + 230, + 88, + 343, + 152 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e032456b63610bdfffbf40b8b7a7a0746f90c1912bad734b944bb3f4bef2f114.jpg", + "image_caption": [ + "(c) RandConv" + ], + "image_footnote": [], + "bbox": [ + 351, + 88, + 462, + 152 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/41423f839c1cfd1aa6654a908c3e8755979cecea770708fd958ed47358e053b4.jpg", + "image_caption": [ + "(d) Pro-RandConv" + ], + "image_footnote": [], + "bbox": [ + 112, + 162, + 222, + 226 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9cc13c06a9c6ce5afd3758a3db6698f48c03d8fc5990d850f5ddab83772e9882.jpg", + "image_caption": [ + "(e) Patch", + "Figure 1. Examples of images generated by different methods." + ], + "image_footnote": [], + "bbox": [ + 232, + 162, + 343, + 227 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fe05c9a4da7a52b5bc8b5282f6f495a9ffaf44f35fc3c8815b7c026c2960b8ad.jpg", + "image_caption": [ + "(f) SAC" + ], + "image_footnote": [], + "bbox": [ + 351, + 162, + 462, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tions are visually imperceptible, rendering these methods inadequate for simulating real-world domain shifts. Another mainstream category trains a generative model to synthesize novel data with style (e.g., color and texture) shifts [4, 10, 20], resulting in visually distinct samples.", + "bbox": [ + 89, + 294, + 482, + 369 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent methods leveraging generative models for data augmentation have achieved remarkable success by enhancing color and texture diversity. For instance, several approaches [10, 29, 40] introduce Gaussian noise to perturb images, effectively enriching the color distribution within a single generated image. Alternatively, other methods [4, 36] utilize a convolution layer with random filters to generate images with diverse local textures. However, these methods enhance style diversity from only a single aspect, either color or texture, leading to inherent limitations in the other aspect. For example, Gaussian noise-based methods, such as UDP [10] and PSDG [40], only generate simplistic textures (see Fig. 1 (a) and (b)), whereas methods leveraging random convolutions, such as RandConv [36] and ProRandConv [4], tend to produce images dominated by a single color (see Fig. 1 (c) and (d)). In contrast, real-world scenarios typically involve a wide range of colors and complex textures, with different regions of an image (e.g., a person's clothing, grass, trees, and the sky within a park), each exhibiting unique texture and color characteristics. Therefore, improving style diversity solely from a single perspective cannot simulate images from complex real-world scenarios. This raises a crucial challenge: Can we combine the advantages of Gaussian noise and random convolutions to generate images with both rich colors and complex textures?", + "bbox": [ + 91, + 371, + 482, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a \"Split-And-Combine\" (SAC) strategy to improve style diversity, which first splits images into a series of patches to enhance the color diversity, then combines the augmented patches into the complete image and uses adaptive random convolutions to improve the texture diversity. In the split step, we propose patch-aware augmentation (PAA) that introduces rich color variations. Specifically, we divide an image into several patches and each patch undergoes independent style generation within Gaussian noise, yielding generated images", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with different local colors. While PAA effectively generates diverse styles for each patch, it leads to texture inconsistencies in a single object, resulting in the model misinterpreting the object as a collection of fragmented parts rather than a unified entity (see Fig. 1 (e)). For instance, a tree trunk may exhibit a smooth texture on one side and a rough, uneven texture on the other, leading the model to perceive it as two distinct components instead of a coherent whole. To address this challenge, we propose adaptive random convolutions (ARC) to enhance texture complexity and restore object integrity in the combine step. Specifically, we first initialize a deformable convolution layer with both random filters and Gaussian filters. Subsequently, we iteratively apply this layer and affine transformation to generate random textures, progressively increasing texture complexity. Moreover, we use entropy values to adaptively control this progressive process, preventing excessive application of random convolutions that could distort semantic information. Furthermore, we use an energy-based distribution discrepancy score to amplify the distribution differences between the original and generated data, enabling the synthesis of out-of-distribution data. We conduct experiments on four well-established benchmarks, where SAC consistently achieves state-of-the-art performance in various scenarios. Overall, our contributions can be summarized as follows:", + "bbox": [ + 511, + 90, + 903, + 467 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a \"Split-And-Combine\" strategy that first splits images into patches for color augmentation and then combines the patches to undergo texture augmentation.", + "- We design the patch-aware augmentation that divides an image into a series of patches for separate style transformation to promote color diversity for generated images.", + "- We develop the adaptive random convolution that utilizes entropy values to adaptively control the progressive process of the random deformable convolution to improve the texture diversity.", + "- We present an energy-based perspective to maximize the distribution discrepancy between the original and augmented data, generating out-of-distribution data.", + "- We conduct experiments on four established datasets. The results demonstrate that the proposed SAC achieves consistent state-of-the-art performance." + ], + "bbox": [ + 513, + 470, + 903, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 724, + 653, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, since only requiring a single training domain, single domain generalization has attracted substantial interest, which can be roughly categorized into two paradigms.", + "bbox": [ + 511, + 750, + 903, + 795 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Gradient-based augmentation methods. This paradigm of work attempts to generate data using backpropagated gradients. One approach involves directly adding gradients to the original data, thereby synthesizing augmented data. For example, ADA [32] employs backpropagated gradients to synthesize data. Subsequently, M-ADA [27] enhances ADA with a Wasserstein Auto", + "bbox": [ + 511, + 795, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15617", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Encoder to increase the distribution differences between the augmented and original data. Similarly, ME-ADA [44] introduces a maximum-entropy regularizer to amplify distribution shifts in augmented data by minimizing mutual information between augmented and original data. Later on, NCDG [31] presents neuron coverage to improve generalization performance by maximizing gradient similarity between the original and augmented data. CAA [3] relies on angular distance to push generated data away from class centroids. AdvST [45] uses back-propagated gradients as guidance to fully exploit the effectiveness of several standard data augmentation methods. However, these methods fail in target domains with large domain shifts due to visually imperceptible style variations.", + "bbox": [ + 89, + 90, + 480, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Style-based augmentation methods. This paradigm of work trains a generative model to produce data with different styles (e.g., contrast, colors, and textures). The most well-established method is PDEN [20], which progressively trains multiple generative models to synthesize different augmented domains and uses cycle consistency loss to preserve semantic information. Additionally, L2D [34] generates data and captures invariant features through a max-min game of mutual information. ATS [37] introduces a teacher-student network to learn domain-generalized knowledge by minimizing the discrepancy between the normalized features of the original and augmented data. CSDG [23] simultaneously changes the textures and intensities of the original data to generate new data. However, due to the limited styles provided by AdaIN [14], these methods only synthesize data with simple colors and textures, which limits their applicability in real-world scenarios. To enrich the styles, Pro-RandConv [4] introduces a deformable convolution block to improve diversification by applying it along with affine transformation to generate data with new styles. UDP [10] adds extra Gaussian noise to perturb the data for more complex styles and enhances the shifts in generated data by maximizing the uncertainty between the generated and original data from an information theory perspective. Subsequently, PDG [29] enhances UDP by modifying the structure of the generative model, using $1 \\times 1$ convolutions to preserve the input shape while employing random and larger kernel sizes to alter the local texture. PSDG [40] was recently proposed to improve generalizability by optimizing the model during both training and testing. AdvFrequency [12] improves frequency discrepancies between the generated and original data through the use of adversarial learning, thereby reducing the risk of the particular frequency.", + "bbox": [ + 91, + 306, + 483, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previous methods typically enhance style diversity from a single perspective, focusing on either color or texture, and fail to simulate the complexity of real-world images. In contrast, SAC improves both color and texture diversity simultaneously through PAA and ARC, thereby enhancing the model's ability to capture invariant representations.", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 89, + 604, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Setup. Consider single domain generalization defined on source domain $D_{s} = \\{\\mathbf{x}_{i},\\mathbf{y}_{i}\\}_{i = 1}^{n}$ , where $\\mathbf{x}_i$ is the $i$ -th sample, $\\mathbf{y}_i$ is the corresponding label of $\\mathbf{x}_i$ , and $n$ is the number of samples. The task is to train a model with good generalizability in the unknown target domain $D_{t}$ .", + "bbox": [ + 511, + 114, + 903, + 189 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overall Framework. We propose the \"Split-And-Combine\" strategy to synthesize images with rich colors and complex textures for single domain generalization tasks. SAC comprises two key components: a style generation module, which synthesizes data with novel styles, and a representation learning module, which learns domain invariant features. The style generation module consists of an Encoder, a generative model $\\mathcal{T}(\\cdot)$ , and a Decoder. To enhance color and texture diversity, we introduce patch-aware augmentation and adaptive random convolutions, enabling the generation of more complex styles. Additionally, we design a score to expand domain distribution through energy discrepancies. The representation learning module, which consists of a feature extractor $F$ and a classifier $P$ , combines both original and augmented data to learn invariant representations. The framework of SAC is shown in Fig. 2.", + "bbox": [ + 511, + 190, + 906, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Style Generation Module", + "text_level": 1, + "bbox": [ + 513, + 440, + 743, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As we proposed before, the key to solving single domain generalization is to synthesize data with rich colors and complex textures. However, the augmented data generated by previous methods contains simple textures or colors, failing to simulate images in real-world scenarios. Unlike previous methods, we independently perform style augmentation on each patch to enrich colors and use adaptive random deformable convolutions to enhance texture complexity.", + "bbox": [ + 511, + 463, + 905, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Patch-Aware Augmentation. To improve the color diversity, we propose patch-aware augmentation to ensure that the images have distinct local colors. Given an image $\\mathbf{x} \\in \\mathbb{R}^{H \\times W}$ , we first use an encoder to project the image into the latent space to obtain rich feature representations:", + "bbox": [ + 511, + 584, + 905, + 659 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {r} = \\operatorname {E n c o d e r} (\\mathbf {x}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 648, + 670, + 903, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{r}$ is from the encoder, which consists of learnable deep layers instantiated as two convolutional layers and each with a ReLU activation function behind. Afterward, we divide $\\mathbf{r}$ into $N\\times N$ patches. Each patch $p_{i,j}\\in \\mathbb{R}^{\\frac{H}{N}\\times \\frac{W}{N}}$ can be written as:", + "bbox": [ + 511, + 696, + 905, + 771 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} p _ {i, j} = \\mathbf {r} [ h _ {i}: h _ {i + 1}, w _ {j}: w _ {j + 1} ], \\quad i, j = 0, \\dots , N - 1 \\\\ h _ {i} = \\frac {H}{N} \\times i, \\quad w _ {j} = \\frac {W}{N} \\times j, \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 779, + 903, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $h_i, w_j$ indicate the locations in the image. Next, following the convention in generative models [20, 40], we employ two fully connected layers $f_1(\\cdot)$ and $f_2(\\cdot)$ to learn a randomly initialized Gaussian noise, which can simulate an", + "bbox": [ + 511, + 840, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15618", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bc0ca0cbca69ea061010a9fdf17d765b80997e9a5d9e5763e21d246d27a5da2f.jpg", + "image_caption": [ + "Figure 2. Framework of SAC, where \"PAA\" is patch-aware augmentation, \"ARC\" is adaptive random convolutions, \"En\" is the Encoder, \"De\" is the Decoder, $F$ is the feature extractor, and $P$ is the classifier." + ], + "image_footnote": [], + "bbox": [ + 196, + 85, + 803, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "unknown style. Notably, as styles generated in this process are simple, we go beyond previous methods and present to separately apply style augmentation to each patch rather than the entire image:", + "bbox": [ + 89, + 443, + 483, + 503 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {p} _ {i, j} = \\mathcal {T} (p _ {i, j}), \\quad i, j = 0, \\ldots , N - 1\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 510, + 411, + 529 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} \\left(p _ {i, j}\\right) = \\left(1 + f _ {1} \\left(\\varepsilon_ {i, j}\\right)\\right) \\left(\\frac {p _ {i , j} - \\mu \\left(p _ {i , j}\\right)}{\\sigma \\left(p _ {i , j}\\right)} + \\eta_ {i, j}\\right) + f _ {2} \\left(\\varepsilon_ {i, j}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 531, + 480, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mu (\\cdot)$ and $\\sigma (\\cdot)$ are mean and standard variance respectively, and $\\varepsilon_{i,j},\\eta_{i,j}\\sim \\mathcal{N}(0,1)$ . Here, we adopt $\\eta_{i,j}$ to perturb the semantic information and a constant 1 to prevent large semantic shifts. Notably, the generative model $\\mathcal{T}(\\cdot)$ is flexible, which can be replaced by AdaIN [14], AdaINnoise [10], or StyIN [40], etc. This flexibility allows SAC to serve as a plug-and-play component for other methods. Subsequently, we combine each patch in its original order:", + "bbox": [ + 89, + 575, + 483, + 698 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde {\\mathbf {r}} = \\operatorname {C o n c a t} \\left(\\left[ \\tilde {p} _ {i, j} \\right] _ {i, j = 0, \\dots , N - 1}\\right) \\\\ = \\left[ \\begin{array}{c c c c} \\tilde {p} _ {0, 0} & \\tilde {p} _ {0, 1} & \\dots & \\tilde {p} _ {0, N - 1} \\\\ \\tilde {p} _ {1, 0} & \\tilde {p} _ {1, 1} & \\dots & \\tilde {p} _ {1, N - 1} \\\\ \\vdots & \\vdots & \\ddots & \\vdots \\\\ \\tilde {p} _ {N - 1, 0} & \\tilde {p} _ {N - 1, 1} & \\dots & \\tilde {p} _ {N - 1, N - 1} \\end{array} \\right]. \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 703, + 482, + 804 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By projecting back to the pixel space using the decoder, we can obtain an augmented image $\\tilde{\\mathbf{x}}\\in \\mathbb{R}^{H\\times W}$", + "bbox": [ + 89, + 809, + 482, + 839 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {x}} = \\operatorname {D e c o d e r} (\\tilde {\\mathbf {r}}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 845, + 482, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tilde{\\mathbf{x}}$ is from the decoder, which is learnable deep layers instantiated as two convolutional layers, followed by ReLU", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and Sigmoid activation function respectively. In summary, our patch-aware augmentation is:", + "bbox": [ + 511, + 443, + 906, + 473 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {x}} = \\operatorname {P A A} (\\mathbf {x}, \\mathcal {Q}),\n$$\n", + "text_format": "latex", + "bbox": [ + 647, + 484, + 764, + 500 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w h e r e} \\quad \\mathcal {Q} = \\left\\{\\left\\{\\varepsilon_ {i, j} \\right\\} _ {i, j = 0} ^ {N - 1}, \\left\\{\\eta_ {i, j} \\right\\} _ {i, j = 0} ^ {N - 1} \\right\\}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 500, + 903, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By applying distinct Gaussian noise to different patches, each patch acquires a unique color, resulting in an image with different local colors. However, $\\tilde{\\mathbf{x}}$ cannot be directly used to capture domain invariant representations, since the introduction of illogical styles during patch-aware augmentation. Specifically, nearby pixels with similar textures can be grouped into primitive shapes that represent parts of objects in images [36]. For instance, in animal classification, an animal's eyes as a whole should present a similar texture. However, this consistency can be disrupted by patch-aware augmentation, which relies on the uniform grid division. When an object spans multiple patches, independent style transformations can lead to conflicting texture regions, causing the model to misinterpret the object as fragmented components rather than a unified whole.", + "bbox": [ + 511, + 537, + 906, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adaptive Random Convolutions. To preserve object integrity, we introduce adaptive random convolutions. Inspired by [4, 36], we use the deformable convolution layer with random filters to project random and coherent texture into images. Furthermore, we multiply the convolution kernel element-wise with a Gaussian filter to further smooth the texture in an object spanning multiple patches. Additionally, we choose the deformable convolution rather than the conventional convolution since the deformable convo", + "bbox": [ + 511, + 763, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15619", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "lution can better sample object contours instead of being confined to patch boundaries [7], helping smooth the texture. (see Tab. 7 for ablation studies on convolution types). Given an augmented sample $\\tilde{\\mathbf{x}}$ , we first initialize a $3 \\times 3$ kernel with dilation 1, which defines a regular grid $\\mathcal{R}$ as $\\{(-1, -1), (-1, 0), \\dots, (0, 1), (1, 1)\\}$ . Here, $\\mathcal{R}$ represents the fixed sampling locations of a standard $3 \\times 3$ convolution kernel, where each coordinate pair $(i_m, j_m)$ denotes the fixed offsets relative to the center pixel $(i_0, j_0)$ . In deformable convolution, these offsets are perturbed by random values $(\\Delta i_m, \\Delta j_m)$ , enabling the sampling locations to deviate from the regular grid. We apply this deformable convolution to $\\tilde{\\mathbf{x}}$ :", + "bbox": [ + 89, + 90, + 483, + 287 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde{\\mathbf{x}}^{\\prime}[i_{0},j_{0}] = \\sum_{(i_{m},j_{m})\\in \\mathcal{R}}\\mathbf{w}[i_{m},j_{m}]\\cdot \\tilde{\\mathbf{x}} [i_{0} + i_{m} + \\Delta i_{m},j_{0} + j_{m} + \\Delta j_{m}], \\\\ \\text {w h e r e} \\quad \\mathbf {w} [ i _ {m}, j _ {m} ] = \\mathbf {g} [ i _ {m}, j _ {m} ] \\cdot w [ i _ {m}, j _ {m} ], \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 296, + 506, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $i_m, j_m$ are the locations in $\\mathcal{R}$ , $\\mathbf{w}$ indicates the weights of the convolution kernel that are randomized, $\\mathbf{g}[i_m, j_m] = \\exp(-\\frac{i_m^2 + j_m^2}{2\\sigma_g^2})$ is a Gaussian filter, $w[i_m, j_m]$ is sampled from a Gaussian distribution $\\mathcal{N}(0, \\frac{1}{k^2C_{in}})$ , $k$ is the size of kernel size, $C_{in}$ is the number of input channels, and $\\Delta i_m, \\Delta j_m$ denote the random offsets of deformable convolution. Each location $(i_0, j_0)$ on $\\tilde{\\mathbf{x}}'$ is transformed by the weighted summation of $\\mathbf{w}$ and the pixel values at irregular locations $(i_0 + i_m + \\Delta i_m, j_0 + j_m + \\Delta j_m)$ of $\\tilde{\\mathbf{x}}$ .", + "bbox": [ + 89, + 363, + 482, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Afterward, following the convention of random convolution [4, 36], we employ affine transformation with two parameters $\\gamma$ and $\\beta$ to further enhance the texture diversity. Given an input image $\\tilde{\\mathbf{x}}'$ , this process can be written as:", + "bbox": [ + 89, + 508, + 483, + 568 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A R C} (\\tilde {\\mathbf {x}}, \\gamma , \\beta) = \\tanh (\\operatorname {A d a I N} (\\tilde {\\mathbf {x}} ^ {\\prime}, \\gamma , \\beta)),\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 578, + 423, + 595 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A d a I N} \\left(\\tilde {\\mathbf {x}} ^ {\\prime}, \\gamma , \\beta\\right) = \\gamma \\frac {\\tilde {\\mathbf {x}} ^ {\\prime} - \\mu \\left(\\tilde {\\mathbf {x}} ^ {\\prime}\\right)}{\\sqrt {\\sigma^ {2} \\left(\\tilde {\\mathbf {x}} ^ {\\prime}\\right) + \\epsilon}} + \\beta , \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 598, + 482, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mu (\\cdot)$ and $\\sigma (\\cdot)$ represent the mean and variance, $\\epsilon$ is a positive number, and $\\gamma ,\\beta \\sim \\mathcal{N}(0,1)$ . Then, the tanh is applied to scale the augmented images. We aim to iterate Eq. (7) and Eq. (8) $T$ times to progressively improve the texture diversity. However, this iterative process presents potential risks. Specifically, excessive iteration can distort semantic information and hinder the learning of invariant representations, while insufficient iterations may fail to adequately enhance texture diversity. Moreover, applying a uniform iteration count across all samples in a batch without considering individual discrepancies can lead to semantic distortions in some samples, degrading model performance. To this end, we introduce an adaptive strategy using the entropy to assess whether a sample requires augmentation during the iterative process. Specifically, for each batch of data, we iterate up to 10 times. In the first iteration, all samples undergo augmentation. In subsequent iterations, we", + "bbox": [ + 89, + 643, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "compute the entropy of each sample and apply augmentation only to those with entropy values below a predefined threshold, while high-entropy samples remain unchanged. Entropy serves as a measure of uncertainty, where high entropy indicates a high likelihood of distribution shifts. Our strategy adaptively controls augmentation based on entropy, ensuring that low-entropy samples undergo augmentation to adequately enhance texture complexity, while high-entropy samples remain unchanged to prevent distortion of semantic information. Our strategy can be summarized as follows:", + "bbox": [ + 511, + 90, + 906, + 243 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {t + 1} ^ {+} = \\left\\{ \\begin{array}{c l} \\operatorname {A R C} (\\mathbf {x} _ {t} ^ {+}, \\gamma , \\beta), & \\text {i f} E (\\mathbf {x} _ {t} ^ {+}) < E _ {0}, \\\\ \\mathbf {x} _ {t} ^ {+}, & \\text {i f} E (\\mathbf {x} _ {t} ^ {+}) \\geq E _ {0}, \\end{array} \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 253, + 906, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_t^+$ denotes the sample after the $t^{th}$ iteration ( $t \\in [1,9]$ ), $E_0$ is the threshold defined as $e \\times \\log(C)$ (where $C$ represents the number of classes in the dataset and $e$ is set to 0.3) and $E(\\cdot)$ denotes the entropy loss.", + "bbox": [ + 511, + 299, + 906, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Energy-based Distribution Discrepancy Score. To enhance the ability of the generative model $\\mathcal{T}(\\cdot)$ in the feature-level, we propose an energy-based distribution discrepancy score to facilitate the expansion of generated domain shifts using a trained classifier. To assess the energy, we use logits from the classifier for energy representations, which is widely used to evaluate Helmholtz free energy [9]. Given the input data $\\mathbf{x}$ , and the trained $c$ -class classifier $P$ , where $c$ is the number of classes, we can compute the Helmholtz free energy for $\\mathbf{x}$ :", + "bbox": [ + 511, + 361, + 906, + 512 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE n (\\mathbf {x}) = - \\log \\sum_ {i} ^ {c} e ^ {P _ {i} (\\mathbf {x})}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 523, + 905, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $P_{i}(\\cdot)$ denotes the probability of the class $i$ . It is known that the energy exhibits the uncertainty of the input data with respect to the classifier, where high energy indicates the high probability of being out-of-distribution [21]. Thus, the energy discrepancy can serve as the score to measure the distribution discrepancy between two data under the same view. Given input data $\\mathbf{x}_1$ , $\\mathbf{x}_2$ , and the trained $c$ -class classifier $P$ , we obtain the distribution discrepancy score by:", + "bbox": [ + 511, + 573, + 906, + 709 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS (\\mathbf {x} _ {1}, \\mathbf {x} _ {2}) = \\frac {1}{\\left\\| E n (\\mathbf {x} _ {1}) - E n (\\mathbf {x} _ {2}) \\right\\| _ {1}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 720, + 905, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\| \\cdot \\|_1$ is the mean absolute error [35]. Here, $S(\\cdot, \\cdot)$ represents the approximate distribution distance between two data. During training, we leverage this score to enhance the distribution discrepancy between the original data and generated data, as well as among different generated data, improving the expansion of the training data distribution.", + "bbox": [ + 511, + 763, + 905, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Objective. Since the parameters in ARC are randomly initialized, our loss term in the style generation module is only used to constrain PAA. To prevent semantic", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15620", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "information from distortion, we use the cross-entropy to ensure that the augmented data can be correctly predicted by the model. Specifically, since the original and augmented data correspond to different styles but share the same semantic information, the model should assign them the same label if it classifies based on semantic information.", + "bbox": [ + 89, + 90, + 483, + 181 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c e} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(\\ell \\left(P (F (\\tilde {\\mathbf {x}} _ {i, 1})), \\mathbf {y} _ {i}\\right) + \\ell \\left(P (F (\\tilde {\\mathbf {x}} _ {i, 2})), \\mathbf {y} _ {i}\\right)\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 188, + 478, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\tilde{\\mathbf{x}}_{i,1} = \\mathrm{PAA}(\\mathbf{x}_i,\\mathcal{Q}_1),\\quad \\tilde{\\mathbf{x}}_{i,2} = \\mathrm{PAA}(\\mathbf{x}_i,\\mathcal{Q}_2),$ (12)", + "bbox": [ + 96, + 229, + 482, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_i$ indicates input original data, $\\ell$ is the cross-entropy loss, $\\mathbf{y}_i$ is the label for $\\mathbf{x}_i$ , and $n$ is the number of available data. The augmented data $\\tilde{\\mathbf{x}}_{i,1}$ and $\\tilde{\\mathbf{x}}_{i,2}$ are synthesized using different parameter sets $\\mathcal{Q}_1$ and $\\mathcal{Q}_2$ , respectively. Next, we maximize distribution discrepancies among the original data $\\mathbf{x}$ and different augmented data $\\tilde{\\mathbf{x}}_1$ , $\\tilde{\\mathbf{x}}_2$ , encouraging the generative model to consistently generate data that exhibit large domain shifts relative to the original data and previously generated data:", + "bbox": [ + 89, + 261, + 483, + 397 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(s \\left(\\tilde {\\mathbf {x}} _ {i, 1}, \\mathbf {x} _ {i}\\right) + s \\left(\\tilde {\\mathbf {x}} _ {i, 2}, \\mathbf {x} _ {i}\\right) + s \\left(\\tilde {\\mathbf {x}} _ {i, 1}, \\tilde {\\mathbf {x}} _ {i, 2}\\right)\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 406, + 482, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, we use mean absolute error to compel the generative model to consistently generate data with pixel-level discrepancies:", + "bbox": [ + 89, + 470, + 483, + 516 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\| \\tilde {\\mathbf {x}} _ {i, 1} - \\tilde {\\mathbf {x}} _ {i, 2} \\| _ {1}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 525, + 482, + 564 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In summary, the overall objective is as follows:", + "bbox": [ + 89, + 573, + 403, + 588 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S} = \\mathcal {L} _ {c e} + \\lambda_ {d} \\mathcal {L} _ {d} - \\lambda_ {p} \\mathcal {L} _ {p}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 599, + 482, + 616 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda_{d}$ and $\\lambda_{p}$ are the hyper-parameters.", + "bbox": [ + 89, + 625, + 374, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2. Representation Learning Module", + "text_level": 1, + "bbox": [ + 89, + 648, + 382, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Benefiting from our style generation module, the representation learning module can be supervised by a simple cross-entropy loss function and learn domain invariant representations. The loss for each sample pair is:", + "bbox": [ + 89, + 671, + 483, + 733 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {R} = \\frac {1}{2 n} \\sum_ {i = 1} ^ {n} \\left(\\ell \\left(P \\left(F \\left(\\mathbf {x} _ {i} ^ {+}\\right)\\right), \\mathbf {y} _ {i}\\right) + \\ell \\left(P \\left(F \\left(\\mathbf {x} _ {i} ^ {+ +}\\right)\\right), \\mathbf {y} _ {i}\\right)\\right) \\\\ + \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\ell \\left(P (F (\\mathbf {x} _ {i})), \\mathbf {y} _ {i}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 741, + 468, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_i^+ = \\mathrm{ARC}(\\tilde{\\mathbf{x}}_{i,1},\\gamma_1,\\beta_1),\\mathbf{x}_i^{+ + } = \\mathrm{ARC}(\\tilde{\\mathbf{x}}_{i,2},\\gamma_2,\\beta_2)$ (16)", + "bbox": [ + 89, + 824, + 488, + 858 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_i^+$ and $\\mathbf{x}_i^{+ + }$ are synthesized from $\\mathbf{x}_i$ through distinct parameters.", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a49b7426624b19a21972703ca3292d30b44ed5f2dceae94678e3ffb6097e5e01.jpg", + "table_caption": [ + "Table 1. Accuracy (\\%) of different methods on Digits. Each column title indicates the target domain." + ], + "table_footnote": [], + "table_body": "
MethodsSVHNMNIST-MSYNUSPSAvg.
ERM [15]27.852.739.776.949.3
ADA [32]NeurIPS'1835.560.445.377.354.6
M-ADA [27]CVPR'2042.667.949.078.559.5
ME-ADA [44]NeurIPS'2042.663.350.481.059.3
PDEN [20]CVPR'2162.282.269.485.374.8
L2D [34]ICCV'2162.987.363.784.074.5
NCDG [31]TPAMI'2359.777.463.892.673.4
SADA [42]AAAI'2368.575.672.989.376.6
Pro-RandConv [4]CVPR'2369.782.379.893.781.4
UDP [10]MM'2372.479.781.796.382.5
AdvST [45]AAAI'2467.579.878.195.480.1
PSDG [40]KDD'2473.186.783.495.784.7
AdvFrequency [12]NeurIPS'2473.682.779.989.381.4
SAC77.791.087.996.488.3
", + "bbox": [ + 517, + 127, + 903, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Training Pipeline", + "text_level": 1, + "bbox": [ + 513, + 325, + 683, + 343 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In real-world scenarios, since the styles of target domains exhibit complexity, using a single generated domain may not produce data with sufficient styles. To address this limitation, we employ a progressively adversarial max-min game to optimize the style generation and representation learning module. We use empirical risk minimization (ERM) to pretrain the representation learning module with original data. During training, we first fix the weights of the representation learning module and train the style generation module $S_{1}$ using PAA and ARC to synthesize the augmented data. Next, we fix the weights of the style generation module and utilize the original and augmented data to optimize the representation learning module using Eq. (16). This adversarial process is repeated several times until we obtain $K$ style generation modules $\\{S_{1}, S_{2}, \\dots, S_{K}\\}$ and a trained representation learning module with good generalization ability. Finally, we use the representation learning model to predict test data labels.", + "bbox": [ + 511, + 349, + 906, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 642, + 645, + 659 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and Implementations", + "text_level": 1, + "bbox": [ + 511, + 669, + 782, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To show the superiority of SAC for single domain generalization, we conduct experiments on four datasets: Digits [32], PACS [18], CIFAR10-C [13] and DomainNet [26]. Please see the supplementary material for more implementation details.", + "bbox": [ + 511, + 694, + 905, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison Methods", + "text_level": 1, + "bbox": [ + 511, + 785, + 717, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare SAC with 13 competitive methods, including ERM [15], ADA [32], M-ADA [27], ME-ADA [44], PDEN [20], L2D [34], NCDG [31], SADA [42], Pro-RandConv [4], UDP [10], AdvST [45], PSDG [40], and AdvFrequency [12]. Note that all baselines are compared using the same structures and * indicates our implementation.", + "bbox": [ + 511, + 809, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15621", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7e9293ef2a085105fd0b509da2a3ca7f6e378a14b0064f956372ff89a8194df1.jpg", + "table_caption": [ + "Table 2. Accuracy (\\%) of different methods on PACS. Each column title indicates the source domain." + ], + "table_footnote": [], + "table_body": "
MethodsPhotoArtCartoonSketchAvg.
ERM [15]42.070.976.553.160.7
ADA [32]NeurIPS'1841.768.869.335.153.7
M-ADA [27]CVPR'2043.168.071.933.954.2
ME-ADA [44]NeurIPS'2042.169.270.636.854.7
PDEN [20]CVPR'2158.377.974.657.967.2
L2D [34]ICCV'2152.376.977.953.765.2
NCDG [31]TPAMI'2349.076.676.453.163.8
SADA [42]AAAI'2351.277.876.957.866.2
Pro-RandConv [4]CVPR'2362.977.078.557.168.9
UDP [10]MM'2356.777.177.657.567.2
AdvST [45]AAAI'2464.177.976.057.168.8
PSDG [40]KDD'2462.178.278.363.070.4
AdvFrequency [12]NeurIPS'2458.475.778.658.967.9
SAC61.779.279.264.671.2
", + "bbox": [ + 93, + 127, + 480, + 303 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Experimental Results", + "text_level": 1, + "bbox": [ + 89, + 330, + 294, + 348 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparisons on Digits. We evaluate the effectiveness of SAC on the Digits dataset, as shown in Tab. 1. Compared to the state-of-the-art baseline PSDG, SAC achieves an average accuracy of $88.3\\%$ , representing a $3.6\\%$ improvement. This result demonstrates that SAC generates augmented data with more diverse styles and captures more domain invariant representations than previous approaches. Notably, SAC achieves the highest accuracy across all tasks, underscoring the effectiveness of our proposed PAA and ARC in synthesizing complex styles, thereby facilitating the model generalizing to the unknown domains. These remarked results indicate that SAC captures more generalizable information than previous approaches, further highlighting the advantage of both improving color and texture diversity.", + "bbox": [ + 88, + 354, + 482, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparisons on PACS. Tab. 2 shows the experimental results on PACS dataset, demonstrating the superior performance of SAC compared to competing methods. We first observe that SAC achieves a higher average accuracy than the previous state-of-the-art method PSDG. Specifically, SAC outperforms PSDG by an average of $0.8\\%$ , which clearly demonstrates its effectiveness. Notably, SAC achieves impressive accuracy on the most challenging Sketch task, which uses simplistic images without backgrounds as the training data. The impressive accuracy in this task highlights the superiority of SAC. In addition, although SAC performs a $2.4\\%$ lower accuracy in the Photo task compared to AdvST, it achieves an overall average improvement of $2.4\\%$ , further demonstrating its enhanced generalizability relative to AdvST.", + "bbox": [ + 89, + 566, + 482, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparisons on CIFAR10-C. As shown in Tab. 3, SAC achieves state-of-the-art (SOTA) performance, significantly outperforming competitors in accuracy. SAC demonstrates notable improvements, achieving a high average accuracy of $81.7\\%$ , which is a $1.6\\%$ increase over the previous SOTA method, UDP. Furthermore, compared to UDP, our method achieves substantial gains, with improvements of $3.9\\%$ for", + "bbox": [ + 89, + 794, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/70145bb730414a4c7e31694c8941403bffa833a76ec351d822694b5a44547e0c.jpg", + "table_caption": [ + "Table 3. Accuracy (\\%) of different methods on CIFAR10-C under different corruption categories at the level 5. Each column title indicates the target domain." + ], + "table_footnote": [], + "table_body": "
MethodsWeatherBlurNoiseDigitAvg.
ERM [15]67.356.730.062.354.1
ADA [32]NeurIPS'1872.767.040.066.661.6
M-ADA [27]CVPR'2075.536.754.265.164.7
ME-ADA [44]NeurIPS'2074.471.466.570.870.8
PDEN [20]CVPR'2178.676.778.673.176.8
L2D [34]ICCV'2176.070.273.372.072.9
NCDG [31]TPAMI'2382.282.870.976.278.0
SADA [42]AAAI'2379.182.471.478.477.8
Pro-RandConv* [4]CVPR'2377.878.676.880.778.5
UDP [10]MM'2380.783.679.976.180.1
AdvST [45]AAAI'2481.578.476.579.479.0
PSDG [40]KDD'2482.881.779.176.179.9
AdvFrequency [12]NeurIPS'2480.481.773.777.578.3
SAC80.182.683.880.481.7
", + "bbox": [ + 517, + 141, + 903, + 321 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/35074a2b586fca91a945404153d29c3c6963edb0c5404e0a1c97274504edd73b.jpg", + "table_caption": [ + "Table 4. Accuracy (\\%) of different methods on DomainNet. Each column title indicates the target domain." + ], + "table_footnote": [], + "table_body": "
MethodsPaintingInfographClipartSketchQuickdrawAvg.
ERM* [15]38.314.035.322.92.122.5
ME-ADA [44]NeurIPS/2039.014.041.025.34.324.7
PDEN* [20]CVPR/2139.013.138.929.74.725.1
L2D [34]ICCV/2138.412.038.730.16.525.2
Pro-RandConv* [4]CVPR/2342.314.235.433.25.927.4
UDP* [10]MM/2338.613.039.130.14.725.1
AdvST [45]AAAI/2442.314.841.530.85.927.1
PSDG* [40]KDD/2439.013.539.530.54.325.4
AdvFrequency [12]NeurIPS/2442.815.142.731.26.927.7
SAC42.513.943.334.97.128.4
", + "bbox": [ + 516, + 375, + 903, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Noise and $4.3\\%$ for Digit. These impressive results highlight the advantages of SAC.", + "bbox": [ + 511, + 503, + 903, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparisons on DomainNet. To further evaluate the superiority of SAC, we compare it with the recent methods on the large-scale DomainNet benchmark. The results are presented in Tab. 4. We can observe that SAC achieves the highest average accuracy among all baselines. Additionally, the methods that solely introduce Gaussian noise to improve color diversity, such as UDP and PSDG, exhibit relatively poor performance. We conjecture that generating images with simplistic textures fails to simulate the complexity of real-world images. While Pro-RandConv achieves competitive accuracy by enhancing texture diversity, SAC still outperforms it in terms of average accuracy. These results validate the effectiveness of SAC, which simultaneously enhances both color and texture complexity.", + "bbox": [ + 509, + 535, + 906, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Analysis Experiment", + "text_level": 1, + "bbox": [ + 511, + 757, + 710, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Study. We provide the ablation study on theDigits dataset to evaluate the effectiveness of key components in SAC, as shown in Tab. 5. Specifically, we introduce some variants to analyze the contributions of SAC, ARC, the Gaussian filter $\\mathbf{g}$ , $\\mathcal{L}_d$ , and $\\mathcal{L}_p$ . When we remove a component, all others remain default settings. We first delete ARC (SAC W/O ARC) to assess the effectiveness of preserving object integrity. We observe that the performance", + "bbox": [ + 511, + 779, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15622", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7ac64e5fab025f954fddccd7d5be94d25ee5082085f508c5325e546fad3e7da4.jpg", + "table_caption": [ + "Table 5. Ablation study on Digits. Each column title indicates the target domain." + ], + "table_footnote": [], + "table_body": "
MethodsSVHNMNIST-MSYNUSPSAvg.
SAC W/O ARC64.179.175.192.777.7
SAC W/O PAA74.985.386.096.185.3
SAC W/O g77.688.886.594.286.8
SAC W/O Ld75.590.387.096.787.4
SAC W/O Lp74.989.786.796.787.0
SAC W/O Ada76.390.585.595.486.9
SAC W/O Ada max73.989.885.495.386.1
SAC77.791.087.996.488.3
", + "bbox": [ + 93, + 127, + 478, + 247 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "drops seriously, which indicates the harm of illogical textures. Next, to investigate the influence of PAA, we only employ ARC (SAC W/O PAA) to generate images. The lower accuracy compared to SAC suggests that focusing solely on texture diversity is insufficient to enhance overall style diversity. We then remove the Gaussian filter $\\mathbf{g}$ (SAC W/O $\\mathbf{g}$ ) to evaluate its effectiveness in preserving object integrity. The degraded performance demonstrates that $\\mathbf{g}$ can help maintain object integrity and prevent texture inconsistencies. Moreover, we see that eliminating $\\mathcal{L}_d$ (SAC W/O $\\mathcal{L}_d$ ) or $\\mathcal{L}_p$ (SAC W/O $\\mathcal{L}_p$ ) would lead to inferior performance compared to SAC, highlighting the importance of expanding the discrepancy at both the feature and pixel levels. Finally, we remove Eq. (9) (SAC W/O Ada) and follow the settings in [4] by iterating the random convolutions $T$ times, where $T$ is sampled from the uniform distribution $U$ [1, 10]. The inferior accuracy indicates that our proposed adaptive strategy effectively improves texture diversity while preventing semantic distortion. To further demonstrate its effectiveness in preserving semantic information, we conduct an experiment in which $T$ is fixed at 10 (SAC W/O $\\mathrm{Ada}_{\\mathrm{max}}$ ). The approximately $2\\%$ performance drop confirms the importance of Eq. (9).", + "bbox": [ + 93, + 277, + 482, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Complementarity with Other Approaches. SAC can serve as a plug-and-play component to enhance the generalizability of other methods. Specifically, we apply PAA and ARC to the approaches of PDEN, L2D, RandConv, ProRandConv, and PSDG, respectively. The experimental results on Digits are shown in Tab. 6. These results indicate that ARC can serve as a general enhancer, improving the generalizability of other methods. Notably, when we apply ARC to PDEN, we find that \"PDEN+ARC\" achieves an $8.1\\%$ improvement of average accuracy under default settings compared to PDEN.", + "bbox": [ + 89, + 626, + 482, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparative Analysis of Different Convolutions. In Tab. 7, we evaluate the effectiveness of various convolution types on the Digits dataset, including conventional convolution [17] (SAC-Conv), depthwise separable convolution [5] (SAC-DSCov), wavelet convolution [8] (SAC-WTConv), and deformable convolution [6] (SAC). SAC and its variants achieve outstanding performance, highlighting the ef", + "bbox": [ + 89, + 794, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/81f348d402e7f1dde6ca52128028c1d016b4ad76218c7abbffa9606598a8433c.jpg", + "table_caption": [ + "Table 6. Accuracy (\\%) on Digits when SAC is applied to recent methods as a plug-and-play component. Each column title indicates the target domain." + ], + "table_footnote": [], + "table_body": "
MethodsSVHNMNIST-MSYNUSPSAvg.
PDEN [20]62.282.269.485.374.8
PDEN+SAC73.886.781.590.183.0
L2D [34]62.987.363.784.074.5
L2D+SAC59.882.270.693.176.4
PSDG [40]73.186.783.495.784.7
PSDG+SAC75.890.885.096.387.0
RandConv [36]62.187.963.984.474.6
RandConv+ARC68.286.871.085.878.0
Pro-RandConv [4]69.782.379.893.781.4
Pro-RandConv+ARC70.585.982.795.383.6
", + "bbox": [ + 517, + 141, + 903, + 282 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fd7865a8caffbfd67107c5824d0b65cb06cc81fe2e7e262f899fb078ea271d68.jpg", + "table_caption": [ + "Table 7. Performance comparison of different convolution on Digits. Each column title indicates the target domain." + ], + "table_footnote": [], + "table_body": "
MethodsSVHNMNIST-MSYNUSPSAvg.
SAC-Conv [36]78.387.387.492.586.4
SAC-DSConv [5]78.789.187.492.486.9
SAC-WTConv [8]74.387.883.895.985.5
SAC77.791.087.996.488.3
", + "bbox": [ + 517, + 340, + 901, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "fectiveness of our proposed \"Split-And-Combine\" strategy. Specifically, the conventional convolution and depthwise separable convolution remain constrained to uniform grids, limiting their ability to accurately capture object contours and preserve object integrity. While the wavelet convolution expands the receptive field through wavelet transformation, it still extracts features within a uniform grid, making it inherently constrained by grid boundaries. In contrast, the deformable convolution overcomes these limitations by allowing sampling points to deviate from a fixed uniform grid, thereby capturing object shapes and structural details more effectively and contributing to texture smoothing.", + "bbox": [ + 511, + 441, + 906, + 623 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 645, + 633, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a novel \"Split-and-Combine\" method for single domain generalization. Instead of applying style augmentation to entire images, our method first performs patch-aware augmentation, which conducts separate style generation to each patch of an image, enriching colors. Then, we introduce adaptive random convolutions to enhance texture complexity while preserving object integrity. Through these synergistic components, our model can synthesize augmented data with different local colors and textures. Additionally, we design an energy-based distribution discrepancy score to assess distribution differences between the original and augmented data, which can encourage SAC to generate novel data with large distribution shifts. SAC consistently achieves state-of-the-art performance across four datasets, demonstrating the effectiveness of our method.", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15623", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China under Grants 62306008, 32472007, 62406095, and 62402004, the Key Project of Anhui Province's Science and Technology Innovation Tackle Plan under Grant 202423k09020040, the China Postdoctoral Science Foundation under Grant 2024M762632, the Natural Science Foundation of Shaanxi under Grant 2024JC-YBQN-0687, the Social Science Foundation of Shaanxi under Grant 2024R027, and the Hefei Key Technology R&D \"Champion-Based Selection\" Project under Grant 2023SGJ011.", + "bbox": [ + 89, + 114, + 485, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 295, + 187, + 311 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kartik Ahuja, Karthikeyan Shanmugam, Kush R. Varshney, and Amit Dhurandhar. Invariant risk minimization games. In International Conference on Machine Learning, Virtual Event, July 13-18, 2020. 1", + "[2] Kartik Ahuja, Ethan Caballero, Dinghuai Zhang, Jean-Christophe Gagnon-Audet, Yoshua Bengio, Ioannis Mitliagkas, and Irina Rish. Invariance principle meets information bottleneck for out-of-distribution generalization. In Neural Information Processing Systems, virtual, December 6-14, 2021. 1", + "[3] Tianle Chen, Mahsa Baktashmotlagh, Zijian Wang, and Mathieu Salzmann. Center-aware adversarial augmentation for single domain generalization. In Winter Conference on Applications of Computer Vision, Waikoloa, HI, USA, January 2-7, 2023. 1, 3", + "[4] Seokeon Choi, Debasmit Das, Sungha Choi, Seunghan Yang, Hyunsin Park, and Sungrack Yun. Progressive random convolutions for single domain generalization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, Vancouver, BC, Canada, June 17-24, pages 10312-10322, 2023. 2, 3, 4, 5, 6, 7, 8", + "[5] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Conference on Computer Vision and Pattern Recognition, Honolulu, HI, USA, July 21-26, 2017. 8", + "[6] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In International Conference on Computer Vision, Venice, Italy, October 22-29, 2017. 8", + "[7] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In International Conference on Computer Vision, Venice, Italy, October 22-29, 2017. 5", + "[8] Shahaf E. Finder, Roy Amoyal, Eran Treister, and Oren Freifeld. Wavelet convolutions for large receptive fields. In European Conference on Computer Vision, MiCo Milano, September 29 - October 4, 2024. 8", + "[9] Will Grathwohl, Kuan-Chieh Wang, Jorn-Henrik Jacobsen, David Duvenaud, Mohammad Norouzi, and Kevin Swersky. Your classifier is secretly an energy based model and you should treat it like one. In International Conference on" + ], + "bbox": [ + 99, + 320, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Learning Representations, Addis Ababa, Ethiopia, April 26-30, 2020. 5", + "[10] Kehua Guo, Rui Ding, Tian Qiu, Xiangyuan Zhu, Zheng Wu, Liwei Wang, and Hui Fang. Single domain generalization via unsupervised diversity probe. In ACM International Conference on Multimedia, Ottawa, ON, Canada, October 29-November 3, pages 2101-2111, 2023. 2, 3, 4, 6, 7", + "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Conference on Computer Vision and Pattern Recognition, Las Vegas, NV, USA, June 27-30, pages 770-778, 2016. 1", + "[12] Xilin He, Jingyu Hu, Qinliang Lin, Cheng Luo, Weicheng Xie, Siyang Song, Muhammad Haris Khan, and Linlin Shen. Towards combating frequency simplicity-biased learning for domain generalization. In Conference on Neural Information Processing Systems, Vancouver, BC, Canada, December 10 - 15, 2024. 3, 6, 7", + "[13] Dan Hendrycks and Thomas G. Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, New Orleans, LA, USA, May 6-9, 2019. 6", + "[14] Xun Huang and Serge J. Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In International Conference on Computer Vision, Venice, Italy, October 22-29, pages 1510-1519, 2017. 3, 4", + "[15] Vladimir Koltchinskii. Oracle inequalities in empirical risk minimization and sparse recovery problems: École D'Été de Probabilités de Saint-Flour XXXVIII-2008. Springer Science & Business Media, 2011. 6, 7", + "[16] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In Conference on Neural Information Processing Systems, Lake Tahoe, Nevada, United States, December 3-6, 2012. 1", + "[17] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 1998. 8", + "[18] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M. Hospedales. Deeper, broader and artier domain generalization. In IEEE International Conference on Computer Vision, Venice, Italy, October 22-29, pages 5542-5550, 2017. 6", + "[19] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M. Hospedales. Learning to generalize: Meta-learning for domain generalization. In Proceedings of Conference on Artificial Intelligence, New Orleans, Louisiana, USA, February 2-7, 2018. 1", + "[20] Lei Li, Ke Gao, Juan Cao, Ziyao Huang, Yepeng Weng, Xiaoyue Mi, Zhengze Yu, Xiaoya Li, and Boyang Xia. Progressive domain expansion network for single domain generalization. In Conference on Computer Vision and Pattern Recognition, virtual, June 19-25, pages 224-233, 2021. 2, 3, 6, 7, 8", + "[21] Weitang Liu, Xiaoyun Wang, John D. Owens, and Yixuan Li. Energy-based out-of-distribution detection. In Conference on Neural Information Processing Systems, virtual, December 6-12, 2020. 5", + "[22] Krikamol Muandet, David Balduzzi, and Bernhard Scholkopf. Domain generalization via invariant feature" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15624", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "representation. In Proceedings of International Conference on Machine Learning Atlanta, GA, USA, June 16-21, pages 10-18, 2013. 1", + "[23] Cheng Ouyang, Chen Chen, Surui Li, Zeju Li, Chen Qin, Wenjia Bai, and Daniel Rueckert. Causality-inspired single-source domain generalization for medical image segmentation. IEEE Transactions on Medical Imaging, 2023. 3", + "[24] Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. IEEE Transactions on Knowledge and Data Engineering, 22(10):1345-1359, 2010. 1", + "[25] Sinno Jialin Pan, Ivor W. Tsang, James T. Kwok, and Qiang Yang. Domain adaptation via transfer component analysis. In International Joint Conference on Artificial Intelligence, Pasadena, California, USA, July 11-17, 2009. 1", + "[26] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In International Conference on Computer Vision, Seoul, Korea (South), October 27 - November 2, pages 1406-1415. IEEE, 2019. 6", + "[27] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In Conference on Computer Vision and Pattern Recognition, Seattle, WA, USA, June 13-19, pages 12553-12562, 2020. 2, 6, 7", + "[28] Alexander Robey, George J. Pappas, and Hamed Hassani. Model-based domain generalization. In Conference on Neural Information Processing Systems, virtual, December 6-14, 2021. 1", + "[29] Ding Rui, Kehua Guo, Xiangyuan Zhu, Zheng Wu, and Hui Fang and. Progressive diversity generation for single domain generalization. IEEE Transactions on Multimedia, 2024. 2, 3", + "[30] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott E. Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Conference on Computer Vision and Pattern Recognition, Boston, MA, USA, June 7-12, 2015. 1", + "[31] Chris Xing Tian, Haoliang Li, Xiaofei Xie, Yang Liu, and Shiqi Wang. Neuron coverage-guided domain generalization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):1302-1311, 2023. 1, 3, 6, 7", + "[32] Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C. Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. In Conference on Neural Information Processing Systems, Montreal, Canada, December 3-8, pages 5339-5349, 2018. 1, 2, 6, 7", + "[33] Chaoqun Wan, Xu Shen, Yonggang Zhang, Zhiheng Yin, Xinmei Tian, Feng Gao, Jianqiang Huang, and Xian-Sheng Hua. Meta convolutional neural networks for single domain generalization. In Conference on Computer Vision and Pattern Recognition, New Orleans, LA, USA, June 18-24, pages 4672-4681, 2022. 1", + "[34] Zijian Wang, Yadan Luo, Ruihong Qiu, Zi Huang, and Mahsa Baktashmotlagh. Learning to diversify for single domain generalization. In International Conference on Computer Vision, Montreal, QC, Canada, October 10-17, pages 814-823, 2021. 3, 6, 7, 8" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Cort J Willmott and Kenji Matsuura. Advantages of the mean absolute error (mae) over the root mean square error (rmse) in assessing average model performance. Climate research, 2005. 5", + "[36] Zhenlin Xu, Deyi Liu, Junlin Yang, Colin Raffel, and Marc Niethammer. Robust and generalizable visual representation learning via random convolutions. In International Conference on Learning Representations, virtual, Austria, May 3-7, 2021. 2, 4, 5, 8", + "[37] Fu-En Yang, Yuan-Chia Cheng, Zu-Yun Shiau, and Yu-Chiang Frank Wang. Adversarial teacher-student representation learning for domain generalization. In Conference on Neural Information Processing Systems, December 6-14, virtual, 2021. 3", + "[38] Shuai Yang, Kui Yu, Fuyuan Cao, Hao Wang, and Xindong Wu. Dual-representation-based autoencoder for domain adaptation. IEEE Transactions on Cybernetics, 52(8): 7464-7477, 2022. 1", + "[39] Shuai Yang, Kui Yu, Fuyuan Cao, Lin Liu, Hao Wang, and Jiuyong Li. Learning causal representations for robust domain adaptation. IEEE Transactions on Knowledge and Data Engineering, 35(3):2750-2764, 2023. 1", + "[40] Shuai Yang, Zhen Zhang, and Lichuan Gu. Practical single domain generalization via training-time and test-time learning. In ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Barcelona, Spain, August 25-29, 2024. 2, 3, 4, 6, 7, 8", + "[41] Haotian Ye, Chuanlong Xie, Tianle Cai, Ruichen Li, Zhenguo Li, and Liwei Wang. Towards a theoretical framework of out-of-distribution generalization. In Neural Information Processing Systems, virtual, December 6-14, 2021. 1", + "[42] Jiajin Zhang, Hanqing Chao, Amit Dhurandhar, Pin-Yu Chen, Ali Tajer, Yangyang Xu, and Pingkun Yan. When neural networks fail to generalize? A model sensitivity perspective. In AAAI Conference on Artificial Intelligence, Washington, DC, USA, February 7-14, 2023. 6, 7", + "[43] Zuyu Zhang, Yan Li, and Byung-Seok Shin. Embracing domain gradient conflicts: Domain generalization using domain gradient equilibrium. In International Conference on Multimedia, Melbourne, VIC, Australia, October 28 - November 1, 2024. 1", + "[44] Long Zhao, Ting Liu, Xi Peng, and Dimitris N. Metaxas. Maximum-entropy adversarial data augmentation for improved generalization and robustness. In Advances in Neural Information Processing Systems, virtual, December 6-12, pages 1-13, 2020. 3, 6, 7", + "[45] Guangtao Zheng, Mengdi Huai, and Aidong Zhang. Advst: Revisiting data augmentations for single domain generalization. In Thirty-Eighth AAAI Conference on Artificial Intelligence, February 20-27, pages 1-9, 2024. 1, 3, 6, 7", + "[46] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4396-4415, 2023. 1", + "[47] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A comprehensive survey on transfer learning. Proceedings of the IEEE, 109(1):43-76, 2021. 1" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15625", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Split-and-Combine_ Enhancing Style Augmentation for Single Domain Generalization/3714ebba-fcba-4212-aaf3-03100ed83e52_model.json b/2025/Split-and-Combine_ Enhancing Style Augmentation for Single Domain Generalization/3714ebba-fcba-4212-aaf3-03100ed83e52_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9b57494ab0e5709aa8eeb941a43712cee715f3c6 --- /dev/null +++ b/2025/Split-and-Combine_ Enhancing Style Augmentation for Single Domain Generalization/3714ebba-fcba-4212-aaf3-03100ed83e52_model.json @@ -0,0 +1,2222 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.143, + 0.13, + 0.856, + 0.175 + ], + "angle": 0, + "content": "Split-and-Combine: Enhancing Style Augmentation for Single Domain Generalization" + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.203, + 0.812, + 0.24 + ], + "angle": 0, + "content": "Zhen Zhang\\(^{1}\\), Shuai Yang\\(^{1\\dagger}\\), Qianlong Dang\\(^{2}\\), Zhize Wu\\(^{3}\\), Lichuan Gu\\(^{1\\dagger}\\) \n\\(^{1}\\)Anhui Agricultural University \\(^{2}\\)Northwest A & F University \\(^{3}\\)Hefei University" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.242, + 0.802, + 0.257 + ], + "angle": 0, + "content": "zhangz@stu.ahau.edu.cn, yangs@ahau.edu.cn, dangqianlong@nwafu.edu.cn," + }, + { + "type": "text", + "bbox": [ + 0.348, + 0.26, + 0.644, + 0.274 + ], + "angle": 0, + "content": "wuzz@hfuu.edu.cn, glc@ahau.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.341, + 0.486, + 0.751 + ], + "angle": 0, + "content": "Single domain generalization aims to learn a model with good generalization ability from a single source domain. Recent advances in this field have focused on increasing the diversity of the training data through style (e.g., color and texture) augmentation. However, most existing methods apply uniform perturbations to the entire image, failing to simulate complex images with multiple distinct stylistic regions. To address this, we propose a \"Split-And-Combine\" (SAC) strategy to enhance style diversity. Specifically, SAC first performs patch-aware augmentation, which splits an image into multiple patches and applies style augmentation independently to each patch, enabling distinct color variations across regions. Then, SAC combines these patches to reconstruct a complete image and applies adaptive random convolutions, which utilizes a deformable convolution layer with random and Gaussian filters to enhance texture diversity while preserving object integrity. Notably, SAC leverages entropy as a risk assessment criterion to adaptively determine whether a sample should undergo augmentation within the iterative process of random convolutions, preventing excessive augmentation. Furthermore, SAC introduces an energy-based distribution discrepancy score to quantify out-of-distribution likelihood, systematically expanding the augmented data's distribution. SAC can serve as a plug-and-play component to improve the performance of recent methods. Extensive experiments on four datasets demonstrate the effectiveness of SAC." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.778, + 0.222, + 0.794 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.804, + 0.485, + 0.88 + ], + "angle": 0, + "content": "Modern machine learning algorithms [11, 16, 30] have achieved remarkable performance under the assumption that the source (training) and target (test) domains share the same distribution. However, in reality, such an assumption is often violated due to factors such as variations in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.311, + 0.907, + 0.386 + ], + "angle": 0, + "content": "illumination, background, or weather, which are known as the domain shift problem [41]. Unfortunately, these traditional algorithms often face significant performance degradation when there are distributional discrepancies between the source and target domains [24, 25, 38, 39, 47]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.909, + 0.599 + ], + "angle": 0, + "content": "To tackle this problem, multi-source domain generalization [19, 22, 28, 46] has been presented, aiming to learn a robust model by using knowledge from multiple different but related source domains without accessing target domain data. The mainstream works depend heavily on domain identifiers as supervision signals to capture domain invariant representations. However, the collection and annotation of data from multiple domains can be prohibitively expensive, making multi-source domain generalization less feasible in practical applications. To alleviate this issue, a more challenging and realistic setting known as single domain generalization [31, 33, 43] has emerged and garnered significant attention recently, which only leverages a single source domain to learn a model with good generalizability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.909, + 0.902 + ], + "angle": 0, + "content": "The difficulty in single domain generalization arises from mitigating distributional discrepancies between the source and unknown target domains. In reality, it is observed that content features, such as object features, remain relatively stable despite environmental changes, whereas style features, such as backgrounds, exhibit considerable variability across different domains [1, 2]. For example, in animal image classification, fundamental elements like eyes, ears, and face shape constitute the content, whereas style elements include background and lighting conditions. Therefore, the key to effective single domain generalization is to eliminate spurious correlations between style features and labels. The current mainstream research on single domain generalization primarily focuses on increasing the capacity of the training data through the generation of novel domains, thereby reducing the model's reliance on style features due to the increased diversity of styles within the training data. One category of methods [3, 32, 45] perturb the original data using the back-propagated gradient to obtain augmented data. However, gradient-based perturba" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.238, + 0.901 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "15616" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.089, + 0.223, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.144, + 0.154, + 0.192, + 0.164 + ], + "angle": 0, + "content": "(a) UDP" + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.089, + 0.344, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.259, + 0.154, + 0.316, + 0.164 + ], + "angle": 0, + "content": "(b) PSDG" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.089, + 0.463, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.154, + 0.444, + 0.164 + ], + "angle": 0, + "content": "(c) RandConv" + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.164, + 0.223, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.228, + 0.219, + 0.239 + ], + "angle": 0, + "content": "(d) Pro-RandConv" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.164, + 0.344, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.263, + 0.229, + 0.313, + 0.24 + ], + "angle": 0, + "content": "(e) Patch" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.164, + 0.463, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.385, + 0.229, + 0.431, + 0.24 + ], + "angle": 0, + "content": "(f) SAC" + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.251, + 0.471, + 0.266 + ], + "angle": 0, + "content": "Figure 1. Examples of images generated by different methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.295, + 0.483, + 0.37 + ], + "angle": 0, + "content": "tions are visually imperceptible, rendering these methods inadequate for simulating real-world domain shifts. Another mainstream category trains a generative model to synthesize novel data with style (e.g., color and texture) shifts [4, 10, 20], resulting in visually distinct samples." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.372, + 0.483, + 0.749 + ], + "angle": 0, + "content": "Recent methods leveraging generative models for data augmentation have achieved remarkable success by enhancing color and texture diversity. For instance, several approaches [10, 29, 40] introduce Gaussian noise to perturb images, effectively enriching the color distribution within a single generated image. Alternatively, other methods [4, 36] utilize a convolution layer with random filters to generate images with diverse local textures. However, these methods enhance style diversity from only a single aspect, either color or texture, leading to inherent limitations in the other aspect. For example, Gaussian noise-based methods, such as UDP [10] and PSDG [40], only generate simplistic textures (see Fig. 1 (a) and (b)), whereas methods leveraging random convolutions, such as RandConv [36] and ProRandConv [4], tend to produce images dominated by a single color (see Fig. 1 (c) and (d)). In contrast, real-world scenarios typically involve a wide range of colors and complex textures, with different regions of an image (e.g., a person's clothing, grass, trees, and the sky within a park), each exhibiting unique texture and color characteristics. Therefore, improving style diversity solely from a single perspective cannot simulate images from complex real-world scenarios. This raises a crucial challenge: Can we combine the advantages of Gaussian noise and random convolutions to generate images with both rich colors and complex textures?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this paper, we propose a \"Split-And-Combine\" (SAC) strategy to improve style diversity, which first splits images into a series of patches to enhance the color diversity, then combines the augmented patches into the complete image and uses adaptive random convolutions to improve the texture diversity. In the split step, we propose patch-aware augmentation (PAA) that introduces rich color variations. Specifically, we divide an image into several patches and each patch undergoes independent style generation within Gaussian noise, yielding generated images" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.468 + ], + "angle": 0, + "content": "with different local colors. While PAA effectively generates diverse styles for each patch, it leads to texture inconsistencies in a single object, resulting in the model misinterpreting the object as a collection of fragmented parts rather than a unified entity (see Fig. 1 (e)). For instance, a tree trunk may exhibit a smooth texture on one side and a rough, uneven texture on the other, leading the model to perceive it as two distinct components instead of a coherent whole. To address this challenge, we propose adaptive random convolutions (ARC) to enhance texture complexity and restore object integrity in the combine step. Specifically, we first initialize a deformable convolution layer with both random filters and Gaussian filters. Subsequently, we iteratively apply this layer and affine transformation to generate random textures, progressively increasing texture complexity. Moreover, we use entropy values to adaptively control this progressive process, preventing excessive application of random convolutions that could distort semantic information. Furthermore, we use an energy-based distribution discrepancy score to amplify the distribution differences between the original and generated data, enabling the synthesis of out-of-distribution data. We conduct experiments on four well-established benchmarks, where SAC consistently achieves state-of-the-art performance in various scenarios. Overall, our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.471, + 0.905, + 0.516 + ], + "angle": 0, + "content": "- We propose a \"Split-And-Combine\" strategy that first splits images into patches for color augmentation and then combines the patches to undergo texture augmentation." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.517, + 0.905, + 0.561 + ], + "angle": 0, + "content": "- We design the patch-aware augmentation that divides an image into a series of patches for separate style transformation to promote color diversity for generated images." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.562, + 0.905, + 0.621 + ], + "angle": 0, + "content": "- We develop the adaptive random convolution that utilizes entropy values to adaptively control the progressive process of the random deformable convolution to improve the texture diversity." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.622, + 0.905, + 0.667 + ], + "angle": 0, + "content": "- We present an energy-based perspective to maximize the distribution discrepancy between the original and augmented data, generating out-of-distribution data." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.668, + 0.905, + 0.713 + ], + "angle": 0, + "content": "- We conduct experiments on four established datasets. The results demonstrate that the proposed SAC achieves consistent state-of-the-art performance." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.471, + 0.905, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.725, + 0.655, + 0.74 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.905, + 0.796 + ], + "angle": 0, + "content": "Recently, since only requiring a single training domain, single domain generalization has attracted substantial interest, which can be roughly categorized into two paradigms." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Gradient-based augmentation methods. This paradigm of work attempts to generate data using backpropagated gradients. One approach involves directly adding gradients to the original data, thereby synthesizing augmented data. For example, ADA [32] employs backpropagated gradients to synthesize data. Subsequently, M-ADA [27] enhances ADA with a Wasserstein Auto" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15617" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.304 + ], + "angle": 0, + "content": "Encoder to increase the distribution differences between the augmented and original data. Similarly, ME-ADA [44] introduces a maximum-entropy regularizer to amplify distribution shifts in augmented data by minimizing mutual information between augmented and original data. Later on, NCDG [31] presents neuron coverage to improve generalization performance by maximizing gradient similarity between the original and augmented data. CAA [3] relies on angular distance to push generated data away from class centroids. AdvST [45] uses back-propagated gradients as guidance to fully exploit the effectiveness of several standard data augmentation methods. However, these methods fail in target domains with large domain shifts due to visually imperceptible style variations." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.308, + 0.485, + 0.807 + ], + "angle": 0, + "content": "Style-based augmentation methods. This paradigm of work trains a generative model to produce data with different styles (e.g., contrast, colors, and textures). The most well-established method is PDEN [20], which progressively trains multiple generative models to synthesize different augmented domains and uses cycle consistency loss to preserve semantic information. Additionally, L2D [34] generates data and captures invariant features through a max-min game of mutual information. ATS [37] introduces a teacher-student network to learn domain-generalized knowledge by minimizing the discrepancy between the normalized features of the original and augmented data. CSDG [23] simultaneously changes the textures and intensities of the original data to generate new data. However, due to the limited styles provided by AdaIN [14], these methods only synthesize data with simple colors and textures, which limits their applicability in real-world scenarios. To enrich the styles, Pro-RandConv [4] introduces a deformable convolution block to improve diversification by applying it along with affine transformation to generate data with new styles. UDP [10] adds extra Gaussian noise to perturb the data for more complex styles and enhances the shifts in generated data by maximizing the uncertainty between the generated and original data from an information theory perspective. Subsequently, PDG [29] enhances UDP by modifying the structure of the generative model, using \\( 1 \\times 1 \\) convolutions to preserve the input shape while employing random and larger kernel sizes to alter the local texture. PSDG [40] was recently proposed to improve generalizability by optimizing the model during both training and testing. AdvFrequency [12] improves frequency discrepancies between the generated and original data through the use of adversarial learning, thereby reducing the risk of the particular frequency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Previous methods typically enhance style diversity from a single perspective, focusing on either color or texture, and fail to simulate the complexity of real-world images. In contrast, SAC improves both color and texture diversity simultaneously through PAA and ARC, thereby enhancing the model's ability to capture invariant representations." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.605, + 0.107 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.115, + 0.905, + 0.19 + ], + "angle": 0, + "content": "Problem Setup. Consider single domain generalization defined on source domain \\(D_{s} = \\{\\mathbf{x}_{i},\\mathbf{y}_{i}\\}_{i = 1}^{n}\\), where \\(\\mathbf{x}_i\\) is the \\(i\\)-th sample, \\(\\mathbf{y}_i\\) is the corresponding label of \\(\\mathbf{x}_i\\), and \\(n\\) is the number of samples. The task is to train a model with good generalizability in the unknown target domain \\(D_{t}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.191, + 0.907, + 0.433 + ], + "angle": 0, + "content": "Overall Framework. We propose the \"Split-And-Combine\" strategy to synthesize images with rich colors and complex textures for single domain generalization tasks. SAC comprises two key components: a style generation module, which synthesizes data with novel styles, and a representation learning module, which learns domain invariant features. The style generation module consists of an Encoder, a generative model \\(\\mathcal{T}(\\cdot)\\), and a Decoder. To enhance color and texture diversity, we introduce patch-aware augmentation and adaptive random convolutions, enabling the generation of more complex styles. Additionally, we design a score to expand domain distribution through energy discrepancies. The representation learning module, which consists of a feature extractor \\(F\\) and a classifier \\(P\\), combines both original and augmented data to learn invariant representations. The framework of SAC is shown in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.441, + 0.744, + 0.457 + ], + "angle": 0, + "content": "3.1. Style Generation Module" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.464, + 0.906, + 0.584 + ], + "angle": 0, + "content": "As we proposed before, the key to solving single domain generalization is to synthesize data with rich colors and complex textures. However, the augmented data generated by previous methods contains simple textures or colors, failing to simulate images in real-world scenarios. Unlike previous methods, we independently perform style augmentation on each patch to enrich colors and use adaptive random deformable convolutions to enhance texture complexity." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.585, + 0.906, + 0.66 + ], + "angle": 0, + "content": "Patch-Aware Augmentation. To improve the color diversity, we propose patch-aware augmentation to ensure that the images have distinct local colors. Given an image \\(\\mathbf{x} \\in \\mathbb{R}^{H \\times W}\\), we first use an encoder to project the image into the latent space to obtain rich feature representations:" + }, + { + "type": "equation", + "bbox": [ + 0.649, + 0.671, + 0.905, + 0.687 + ], + "angle": 0, + "content": "\\[\n\\mathbf {r} = \\operatorname {E n c o d e r} (\\mathbf {x}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.697, + 0.906, + 0.772 + ], + "angle": 0, + "content": "where \\(\\mathbf{r}\\) is from the encoder, which consists of learnable deep layers instantiated as two convolutional layers and each with a ReLU activation function behind. Afterward, we divide \\(\\mathbf{r}\\) into \\(N\\times N\\) patches. Each patch \\(p_{i,j}\\in \\mathbb{R}^{\\frac{H}{N}\\times \\frac{W}{N}}\\) can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.78, + 0.905, + 0.841 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} p _ {i, j} = \\mathbf {r} [ h _ {i}: h _ {i + 1}, w _ {j}: w _ {j + 1} ], \\quad i, j = 0, \\dots , N - 1 \\\\ h _ {i} = \\frac {H}{N} \\times i, \\quad w _ {j} = \\frac {W}{N} \\times j, \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.841, + 0.906, + 0.901 + ], + "angle": 0, + "content": "where \\( h_i, w_j \\) indicate the locations in the image. Next, following the convention in generative models [20, 40], we employ two fully connected layers \\( f_1(\\cdot) \\) and \\( f_2(\\cdot) \\) to learn a randomly initialized Gaussian noise, which can simulate an" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15618" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.197, + 0.087, + 0.805, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.389, + 0.908, + 0.418 + ], + "angle": 0, + "content": "Figure 2. Framework of SAC, where \"PAA\" is patch-aware augmentation, \"ARC\" is adaptive random convolutions, \"En\" is the Encoder, \"De\" is the Decoder, \\(F\\) is the feature extractor, and \\(P\\) is the classifier." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.444, + 0.484, + 0.505 + ], + "angle": 0, + "content": "unknown style. Notably, as styles generated in this process are simple, we go beyond previous methods and present to separately apply style augmentation to each patch rather than the entire image:" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.511, + 0.413, + 0.53 + ], + "angle": 0, + "content": "\\[\n\\tilde {p} _ {i, j} = \\mathcal {T} (p _ {i, j}), \\quad i, j = 0, \\ldots , N - 1\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.532, + 0.482, + 0.576 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} \\left(p _ {i, j}\\right) = \\left(1 + f _ {1} \\left(\\varepsilon_ {i, j}\\right)\\right) \\left(\\frac {p _ {i , j} - \\mu \\left(p _ {i , j}\\right)}{\\sigma \\left(p _ {i , j}\\right)} + \\eta_ {i, j}\\right) + f _ {2} \\left(\\varepsilon_ {i, j}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.577, + 0.484, + 0.699 + ], + "angle": 0, + "content": "where \\(\\mu (\\cdot)\\) and \\(\\sigma (\\cdot)\\) are mean and standard variance respectively, and \\(\\varepsilon_{i,j},\\eta_{i,j}\\sim \\mathcal{N}(0,1)\\). Here, we adopt \\(\\eta_{i,j}\\) to perturb the semantic information and a constant 1 to prevent large semantic shifts. Notably, the generative model \\(\\mathcal{T}(\\cdot)\\) is flexible, which can be replaced by AdaIN [14], AdaINnoise [10], or StyIN [40], etc. This flexibility allows SAC to serve as a plug-and-play component for other methods. Subsequently, we combine each patch in its original order:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.704, + 0.483, + 0.805 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tilde {\\mathbf {r}} = \\operatorname {C o n c a t} \\left(\\left[ \\tilde {p} _ {i, j} \\right] _ {i, j = 0, \\dots , N - 1}\\right) \\\\ = \\left[ \\begin{array}{c c c c} \\tilde {p} _ {0, 0} & \\tilde {p} _ {0, 1} & \\dots & \\tilde {p} _ {0, N - 1} \\\\ \\tilde {p} _ {1, 0} & \\tilde {p} _ {1, 1} & \\dots & \\tilde {p} _ {1, N - 1} \\\\ \\vdots & \\vdots & \\ddots & \\vdots \\\\ \\tilde {p} _ {N - 1, 0} & \\tilde {p} _ {N - 1, 1} & \\dots & \\tilde {p} _ {N - 1, N - 1} \\end{array} \\right]. \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.81, + 0.483, + 0.84 + ], + "angle": 0, + "content": "By projecting back to the pixel space using the decoder, we can obtain an augmented image \\(\\tilde{\\mathbf{x}}\\in \\mathbb{R}^{H\\times W}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.226, + 0.847, + 0.483, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {x}} = \\operatorname {D e c o d e r} (\\tilde {\\mathbf {r}}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\tilde{\\mathbf{x}}\\) is from the decoder, which is learnable deep layers instantiated as two convolutional layers, followed by ReLU" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.444, + 0.907, + 0.474 + ], + "angle": 0, + "content": "and Sigmoid activation function respectively. In summary, our patch-aware augmentation is:" + }, + { + "type": "equation", + "bbox": [ + 0.648, + 0.485, + 0.766, + 0.501 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {x}} = \\operatorname {P A A} (\\mathbf {x}, \\mathcal {Q}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.501, + 0.905, + 0.531 + ], + "angle": 0, + "content": "\\[\n\\text {w h e r e} \\quad \\mathcal {Q} = \\left\\{\\left\\{\\varepsilon_ {i, j} \\right\\} _ {i, j = 0} ^ {N - 1}, \\left\\{\\eta_ {i, j} \\right\\} _ {i, j = 0} ^ {N - 1} \\right\\}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.764 + ], + "angle": 0, + "content": "By applying distinct Gaussian noise to different patches, each patch acquires a unique color, resulting in an image with different local colors. However, \\(\\tilde{\\mathbf{x}}\\) cannot be directly used to capture domain invariant representations, since the introduction of illogical styles during patch-aware augmentation. Specifically, nearby pixels with similar textures can be grouped into primitive shapes that represent parts of objects in images [36]. For instance, in animal classification, an animal's eyes as a whole should present a similar texture. However, this consistency can be disrupted by patch-aware augmentation, which relies on the uniform grid division. When an object spans multiple patches, independent style transformations can lead to conflicting texture regions, causing the model to misinterpret the object as fragmented components rather than a unified whole." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Adaptive Random Convolutions. To preserve object integrity, we introduce adaptive random convolutions. Inspired by [4, 36], we use the deformable convolution layer with random filters to project random and coherent texture into images. Furthermore, we multiply the convolution kernel element-wise with a Gaussian filter to further smooth the texture in an object spanning multiple patches. Additionally, we choose the deformable convolution rather than the conventional convolution since the deformable convo" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15619" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.288 + ], + "angle": 0, + "content": "lution can better sample object contours instead of being confined to patch boundaries [7], helping smooth the texture. (see Tab. 7 for ablation studies on convolution types). Given an augmented sample \\(\\tilde{\\mathbf{x}}\\), we first initialize a \\(3 \\times 3\\) kernel with dilation 1, which defines a regular grid \\(\\mathcal{R}\\) as \\(\\{(-1, -1), (-1, 0), \\dots, (0, 1), (1, 1)\\}\\). Here, \\(\\mathcal{R}\\) represents the fixed sampling locations of a standard \\(3 \\times 3\\) convolution kernel, where each coordinate pair \\((i_m, j_m)\\) denotes the fixed offsets relative to the center pixel \\((i_0, j_0)\\). In deformable convolution, these offsets are perturbed by random values \\((\\Delta i_m, \\Delta j_m)\\), enabling the sampling locations to deviate from the regular grid. We apply this deformable convolution to \\(\\tilde{\\mathbf{x}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.297, + 0.508, + 0.363 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tilde{\\mathbf{x}}^{\\prime}[i_{0},j_{0}] = \\sum_{(i_{m},j_{m})\\in \\mathcal{R}}\\mathbf{w}[i_{m},j_{m}]\\cdot \\tilde{\\mathbf{x}} [i_{0} + i_{m} + \\Delta i_{m},j_{0} + j_{m} + \\Delta j_{m}], \\\\ \\text {w h e r e} \\quad \\mathbf {w} [ i _ {m}, j _ {m} ] = \\mathbf {g} [ i _ {m}, j _ {m} ] \\cdot w [ i _ {m}, j _ {m} ], \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.483, + 0.509 + ], + "angle": 0, + "content": "where \\(i_m, j_m\\) are the locations in \\(\\mathcal{R}\\), \\(\\mathbf{w}\\) indicates the weights of the convolution kernel that are randomized, \\(\\mathbf{g}[i_m, j_m] = \\exp(-\\frac{i_m^2 + j_m^2}{2\\sigma_g^2})\\) is a Gaussian filter, \\(w[i_m, j_m]\\) is sampled from a Gaussian distribution \\(\\mathcal{N}(0, \\frac{1}{k^2C_{in}})\\), \\(k\\) is the size of kernel size, \\(C_{in}\\) is the number of input channels, and \\(\\Delta i_m, \\Delta j_m\\) denote the random offsets of deformable convolution. Each location \\((i_0, j_0)\\) on \\(\\tilde{\\mathbf{x}}'\\) is transformed by the weighted summation of \\(\\mathbf{w}\\) and the pixel values at irregular locations \\((i_0 + i_m + \\Delta i_m, j_0 + j_m + \\Delta j_m)\\) of \\(\\tilde{\\mathbf{x}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.484, + 0.569 + ], + "angle": 0, + "content": "Afterward, following the convention of random convolution [4, 36], we employ affine transformation with two parameters \\(\\gamma\\) and \\(\\beta\\) to further enhance the texture diversity. Given an input image \\(\\tilde{\\mathbf{x}}'\\), this process can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.579, + 0.424, + 0.596 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A R C} (\\tilde {\\mathbf {x}}, \\gamma , \\beta) = \\tanh (\\operatorname {A d a I N} (\\tilde {\\mathbf {x}} ^ {\\prime}, \\gamma , \\beta)),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.599, + 0.483, + 0.635 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A d a I N} \\left(\\tilde {\\mathbf {x}} ^ {\\prime}, \\gamma , \\beta\\right) = \\gamma \\frac {\\tilde {\\mathbf {x}} ^ {\\prime} - \\mu \\left(\\tilde {\\mathbf {x}} ^ {\\prime}\\right)}{\\sqrt {\\sigma^ {2} \\left(\\tilde {\\mathbf {x}} ^ {\\prime}\\right) + \\epsilon}} + \\beta , \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mu (\\cdot)\\) and \\(\\sigma (\\cdot)\\) represent the mean and variance, \\(\\epsilon\\) is a positive number, and \\(\\gamma ,\\beta \\sim \\mathcal{N}(0,1)\\). Then, the tanh is applied to scale the augmented images. We aim to iterate Eq. (7) and Eq. (8) \\(T\\) times to progressively improve the texture diversity. However, this iterative process presents potential risks. Specifically, excessive iteration can distort semantic information and hinder the learning of invariant representations, while insufficient iterations may fail to adequately enhance texture diversity. Moreover, applying a uniform iteration count across all samples in a batch without considering individual discrepancies can lead to semantic distortions in some samples, degrading model performance. To this end, we introduce an adaptive strategy using the entropy to assess whether a sample requires augmentation during the iterative process. Specifically, for each batch of data, we iterate up to 10 times. In the first iteration, all samples undergo augmentation. In subsequent iterations, we" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.244 + ], + "angle": 0, + "content": "compute the entropy of each sample and apply augmentation only to those with entropy values below a predefined threshold, while high-entropy samples remain unchanged. Entropy serves as a measure of uncertainty, where high entropy indicates a high likelihood of distribution shifts. Our strategy adaptively controls augmentation based on entropy, ensuring that low-entropy samples undergo augmentation to adequately enhance texture complexity, while high-entropy samples remain unchanged to prevent distortion of semantic information. Our strategy can be summarized as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.254, + 0.907, + 0.289 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {t + 1} ^ {+} = \\left\\{ \\begin{array}{c l} \\operatorname {A R C} (\\mathbf {x} _ {t} ^ {+}, \\gamma , \\beta), & \\text {i f} E (\\mathbf {x} _ {t} ^ {+}) < E _ {0}, \\\\ \\mathbf {x} _ {t} ^ {+}, & \\text {i f} E (\\mathbf {x} _ {t} ^ {+}) \\geq E _ {0}, \\end{array} \\right. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.3, + 0.907, + 0.36 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_t^+\\) denotes the sample after the \\(t^{th}\\) iteration (\\(t \\in [1,9]\\)), \\(E_0\\) is the threshold defined as \\(e \\times \\log(C)\\) (where \\(C\\) represents the number of classes in the dataset and \\(e\\) is set to 0.3) and \\(E(\\cdot)\\) denotes the entropy loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.362, + 0.907, + 0.513 + ], + "angle": 0, + "content": "Energy-based Distribution Discrepancy Score. To enhance the ability of the generative model \\(\\mathcal{T}(\\cdot)\\) in the feature-level, we propose an energy-based distribution discrepancy score to facilitate the expansion of generated domain shifts using a trained classifier. To assess the energy, we use logits from the classifier for energy representations, which is widely used to evaluate Helmholtz free energy [9]. Given the input data \\(\\mathbf{x}\\), and the trained \\(c\\)-class classifier \\(P\\), where \\(c\\) is the number of classes, we can compute the Helmholtz free energy for \\(\\mathbf{x}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.524, + 0.906, + 0.562 + ], + "angle": 0, + "content": "\\[\nE n (\\mathbf {x}) = - \\log \\sum_ {i} ^ {c} e ^ {P _ {i} (\\mathbf {x})}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.574, + 0.907, + 0.71 + ], + "angle": 0, + "content": "where \\(P_{i}(\\cdot)\\) denotes the probability of the class \\(i\\). It is known that the energy exhibits the uncertainty of the input data with respect to the classifier, where high energy indicates the high probability of being out-of-distribution [21]. Thus, the energy discrepancy can serve as the score to measure the distribution discrepancy between two data under the same view. Given input data \\(\\mathbf{x}_1\\), \\(\\mathbf{x}_2\\), and the trained \\(c\\)-class classifier \\(P\\), we obtain the distribution discrepancy score by:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.721, + 0.906, + 0.756 + ], + "angle": 0, + "content": "\\[\nS (\\mathbf {x} _ {1}, \\mathbf {x} _ {2}) = \\frac {1}{\\left\\| E n (\\mathbf {x} _ {1}) - E n (\\mathbf {x} _ {2}) \\right\\| _ {1}}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.856 + ], + "angle": 0, + "content": "where \\( \\| \\cdot \\|_1 \\) is the mean absolute error [35]. Here, \\( S(\\cdot, \\cdot) \\) represents the approximate distribution distance between two data. During training, we leverage this score to enhance the distribution discrepancy between the original data and generated data, as well as among different generated data, improving the expansion of the training data distribution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Training Objective. Since the parameters in ARC are randomly initialized, our loss term in the style generation module is only used to constrain PAA. To prevent semantic" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15620" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.182 + ], + "angle": 0, + "content": "information from distortion, we use the cross-entropy to ensure that the augmented data can be correctly predicted by the model. Specifically, since the original and augmented data correspond to different styles but share the same semantic information, the model should assign them the same label if it classifies based on semantic information." + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.189, + 0.479, + 0.228 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c e} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(\\ell \\left(P (F (\\tilde {\\mathbf {x}} _ {i, 1})), \\mathbf {y} _ {i}\\right) + \\ell \\left(P (F (\\tilde {\\mathbf {x}} _ {i, 2})), \\mathbf {y} _ {i}\\right)\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.231, + 0.483, + 0.26 + ], + "angle": 0, + "content": "where \\(\\tilde{\\mathbf{x}}_{i,1} = \\mathrm{PAA}(\\mathbf{x}_i,\\mathcal{Q}_1),\\quad \\tilde{\\mathbf{x}}_{i,2} = \\mathrm{PAA}(\\mathbf{x}_i,\\mathcal{Q}_2),\\) (12)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.262, + 0.484, + 0.398 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_i\\) indicates input original data, \\(\\ell\\) is the cross-entropy loss, \\(\\mathbf{y}_i\\) is the label for \\(\\mathbf{x}_i\\), and \\(n\\) is the number of available data. The augmented data \\(\\tilde{\\mathbf{x}}_{i,1}\\) and \\(\\tilde{\\mathbf{x}}_{i,2}\\) are synthesized using different parameter sets \\(\\mathcal{Q}_1\\) and \\(\\mathcal{Q}_2\\), respectively. Next, we maximize distribution discrepancies among the original data \\(\\mathbf{x}\\) and different augmented data \\(\\tilde{\\mathbf{x}}_1\\), \\(\\tilde{\\mathbf{x}}_2\\), encouraging the generative model to consistently generate data that exhibit large domain shifts relative to the original data and previously generated data:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.407, + 0.483, + 0.461 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(s \\left(\\tilde {\\mathbf {x}} _ {i, 1}, \\mathbf {x} _ {i}\\right) + s \\left(\\tilde {\\mathbf {x}} _ {i, 2}, \\mathbf {x} _ {i}\\right) + s \\left(\\tilde {\\mathbf {x}} _ {i, 1}, \\tilde {\\mathbf {x}} _ {i, 2}\\right)\\right). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.471, + 0.484, + 0.517 + ], + "angle": 0, + "content": "Additionally, we use mean absolute error to compel the generative model to consistently generate data with pixel-level discrepancies:" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.526, + 0.483, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\| \\tilde {\\mathbf {x}} _ {i, 1} - \\tilde {\\mathbf {x}} _ {i, 2} \\| _ {1}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.574, + 0.405, + 0.589 + ], + "angle": 0, + "content": "In summary, the overall objective is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.6, + 0.483, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S} = \\mathcal {L} _ {c e} + \\lambda_ {d} \\mathcal {L} _ {d} - \\lambda_ {p} \\mathcal {L} _ {p}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.626, + 0.375, + 0.643 + ], + "angle": 0, + "content": "where \\(\\lambda_{d}\\) and \\(\\lambda_{p}\\) are the hyper-parameters." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.65, + 0.383, + 0.667 + ], + "angle": 0, + "content": "3.2. Representation Learning Module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.484, + 0.734 + ], + "angle": 0, + "content": "Benefiting from our style generation module, the representation learning module can be supervised by a simple cross-entropy loss function and learn domain invariant representations. The loss for each sample pair is:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.742, + 0.47, + 0.823 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {R} = \\frac {1}{2 n} \\sum_ {i = 1} ^ {n} \\left(\\ell \\left(P \\left(F \\left(\\mathbf {x} _ {i} ^ {+}\\right)\\right), \\mathbf {y} _ {i}\\right) + \\ell \\left(P \\left(F \\left(\\mathbf {x} _ {i} ^ {+ +}\\right)\\right), \\mathbf {y} _ {i}\\right)\\right) \\\\ + \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\ell \\left(P (F (\\mathbf {x} _ {i})), \\mathbf {y} _ {i}\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.825, + 0.49, + 0.859 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_i^+ = \\mathrm{ARC}(\\tilde{\\mathbf{x}}_{i,1},\\gamma_1,\\beta_1),\\mathbf{x}_i^{+ + } = \\mathrm{ARC}(\\tilde{\\mathbf{x}}_{i,2},\\gamma_2,\\beta_2)\\) (16)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.87, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_i^+\\) and \\(\\mathbf{x}_i^{+ + }\\) are synthesized from \\(\\mathbf{x}_i\\) through distinct parameters." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.906, + 0.118 + ], + "angle": 0, + "content": "Table 1. Accuracy (\\%) of different methods on Digits. Each column title indicates the target domain." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.128, + 0.905, + 0.297 + ], + "angle": 0, + "content": "
MethodsSVHNMNIST-MSYNUSPSAvg.
ERM [15]27.852.739.776.949.3
ADA [32]NeurIPS'1835.560.445.377.354.6
M-ADA [27]CVPR'2042.667.949.078.559.5
ME-ADA [44]NeurIPS'2042.663.350.481.059.3
PDEN [20]CVPR'2162.282.269.485.374.8
L2D [34]ICCV'2162.987.363.784.074.5
NCDG [31]TPAMI'2359.777.463.892.673.4
SADA [42]AAAI'2368.575.672.989.376.6
Pro-RandConv [4]CVPR'2369.782.379.893.781.4
UDP [10]MM'2372.479.781.796.382.5
AdvST [45]AAAI'2467.579.878.195.480.1
PSDG [40]KDD'2473.186.783.495.784.7
AdvFrequency [12]NeurIPS'2473.682.779.989.381.4
SAC77.791.087.996.488.3
" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.327, + 0.684, + 0.344 + ], + "angle": 0, + "content": "3.3. Training Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.351, + 0.907, + 0.624 + ], + "angle": 0, + "content": "In real-world scenarios, since the styles of target domains exhibit complexity, using a single generated domain may not produce data with sufficient styles. To address this limitation, we employ a progressively adversarial max-min game to optimize the style generation and representation learning module. We use empirical risk minimization (ERM) to pretrain the representation learning module with original data. During training, we first fix the weights of the representation learning module and train the style generation module \\( S_{1} \\) using PAA and ARC to synthesize the augmented data. Next, we fix the weights of the style generation module and utilize the original and augmented data to optimize the representation learning module using Eq. (16). This adversarial process is repeated several times until we obtain \\( K \\) style generation modules \\( \\{S_{1}, S_{2}, \\dots, S_{K}\\} \\) and a trained representation learning module with good generalization ability. Finally, we use the representation learning model to predict test data labels." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.643, + 0.646, + 0.66 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.67, + 0.783, + 0.687 + ], + "angle": 0, + "content": "4.1. Datasets and Implementations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.695, + 0.906, + 0.77 + ], + "angle": 0, + "content": "To show the superiority of SAC for single domain generalization, we conduct experiments on four datasets: Digits [32], PACS [18], CIFAR10-C [13] and DomainNet [26]. Please see the supplementary material for more implementation details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.786, + 0.718, + 0.802 + ], + "angle": 0, + "content": "4.2. Comparison Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.902 + ], + "angle": 0, + "content": "We compare SAC with 13 competitive methods, including ERM [15], ADA [32], M-ADA [27], ME-ADA [44], PDEN [20], L2D [34], NCDG [31], SADA [42], Pro-RandConv [4], UDP [10], AdvST [45], PSDG [40], and AdvFrequency [12]. Note that all baselines are compared using the same structures and * indicates our implementation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15621" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.483, + 0.119 + ], + "angle": 0, + "content": "Table 2. Accuracy (\\%) of different methods on PACS. Each column title indicates the source domain." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.128, + 0.481, + 0.304 + ], + "angle": 0, + "content": "
MethodsPhotoArtCartoonSketchAvg.
ERM [15]42.070.976.553.160.7
ADA [32]NeurIPS'1841.768.869.335.153.7
M-ADA [27]CVPR'2043.168.071.933.954.2
ME-ADA [44]NeurIPS'2042.169.270.636.854.7
PDEN [20]CVPR'2158.377.974.657.967.2
L2D [34]ICCV'2152.376.977.953.765.2
NCDG [31]TPAMI'2349.076.676.453.163.8
SADA [42]AAAI'2351.277.876.957.866.2
Pro-RandConv [4]CVPR'2362.977.078.557.168.9
UDP [10]MM'2356.777.177.657.567.2
AdvST [45]AAAI'2464.177.976.057.168.8
PSDG [40]KDD'2462.178.278.363.070.4
AdvFrequency [12]NeurIPS'2458.475.778.658.967.9
SAC61.779.279.264.671.2
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.332, + 0.295, + 0.349 + ], + "angle": 0, + "content": "4.3. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.355, + 0.483, + 0.567 + ], + "angle": 0, + "content": "Comparisons on Digits. We evaluate the effectiveness of SAC on the Digits dataset, as shown in Tab. 1. Compared to the state-of-the-art baseline PSDG, SAC achieves an average accuracy of \\(88.3\\%\\), representing a \\(3.6\\%\\) improvement. This result demonstrates that SAC generates augmented data with more diverse styles and captures more domain invariant representations than previous approaches. Notably, SAC achieves the highest accuracy across all tasks, underscoring the effectiveness of our proposed PAA and ARC in synthesizing complex styles, thereby facilitating the model generalizing to the unknown domains. These remarked results indicate that SAC captures more generalizable information than previous approaches, further highlighting the advantage of both improving color and texture diversity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.483, + 0.794 + ], + "angle": 0, + "content": "Comparisons on PACS. Tab. 2 shows the experimental results on PACS dataset, demonstrating the superior performance of SAC compared to competing methods. We first observe that SAC achieves a higher average accuracy than the previous state-of-the-art method PSDG. Specifically, SAC outperforms PSDG by an average of \\(0.8\\%\\), which clearly demonstrates its effectiveness. Notably, SAC achieves impressive accuracy on the most challenging Sketch task, which uses simplistic images without backgrounds as the training data. The impressive accuracy in this task highlights the superiority of SAC. In addition, although SAC performs a \\(2.4\\%\\) lower accuracy in the Photo task compared to AdvST, it achieves an overall average improvement of \\(2.4\\%\\), further demonstrating its enhanced generalizability relative to AdvST." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Comparisons on CIFAR10-C. As shown in Tab. 3, SAC achieves state-of-the-art (SOTA) performance, significantly outperforming competitors in accuracy. SAC demonstrates notable improvements, achieving a high average accuracy of \\(81.7\\%\\), which is a \\(1.6\\%\\) increase over the previous SOTA method, UDP. Furthermore, compared to UDP, our method achieves substantial gains, with improvements of \\(3.9\\%\\) for" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 3. Accuracy (\\%) of different methods on CIFAR10-C under different corruption categories at the level 5. Each column title indicates the target domain." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.142, + 0.905, + 0.322 + ], + "angle": 0, + "content": "
MethodsWeatherBlurNoiseDigitAvg.
ERM [15]67.356.730.062.354.1
ADA [32]NeurIPS'1872.767.040.066.661.6
M-ADA [27]CVPR'2075.536.754.265.164.7
ME-ADA [44]NeurIPS'2074.471.466.570.870.8
PDEN [20]CVPR'2178.676.778.673.176.8
L2D [34]ICCV'2176.070.273.372.072.9
NCDG [31]TPAMI'2382.282.870.976.278.0
SADA [42]AAAI'2379.182.471.478.477.8
Pro-RandConv* [4]CVPR'2377.878.676.880.778.5
UDP [10]MM'2380.783.679.976.180.1
AdvST [45]AAAI'2481.578.476.579.479.0
PSDG [40]KDD'2482.881.779.176.179.9
AdvFrequency [12]NeurIPS'2480.481.773.777.578.3
SAC80.182.683.880.481.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.336, + 0.907, + 0.365 + ], + "angle": 0, + "content": "Table 4. Accuracy (\\%) of different methods on DomainNet. Each column title indicates the target domain." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.375, + 0.905, + 0.478 + ], + "angle": 0, + "content": "
MethodsPaintingInfographClipartSketchQuickdrawAvg.
ERM* [15]38.314.035.322.92.122.5
ME-ADA [44]NeurIPS/2039.014.041.025.34.324.7
PDEN* [20]CVPR/2139.013.138.929.74.725.1
L2D [34]ICCV/2138.412.038.730.16.525.2
Pro-RandConv* [4]CVPR/2342.314.235.433.25.927.4
UDP* [10]MM/2338.613.039.130.14.725.1
AdvST [45]AAAI/2442.314.841.530.85.927.1
PSDG* [40]KDD/2439.013.539.530.54.325.4
AdvFrequency [12]NeurIPS/2442.815.142.731.26.927.7
SAC42.513.943.334.97.128.4
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.505, + 0.905, + 0.534 + ], + "angle": 0, + "content": "Noise and \\(4.3\\%\\) for Digit. These impressive results highlight the advantages of SAC." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.536, + 0.907, + 0.747 + ], + "angle": 0, + "content": "Comparisons on DomainNet. To further evaluate the superiority of SAC, we compare it with the recent methods on the large-scale DomainNet benchmark. The results are presented in Tab. 4. We can observe that SAC achieves the highest average accuracy among all baselines. Additionally, the methods that solely introduce Gaussian noise to improve color diversity, such as UDP and PSDG, exhibit relatively poor performance. We conjecture that generating images with simplistic textures fails to simulate the complexity of real-world images. While Pro-RandConv achieves competitive accuracy by enhancing texture diversity, SAC still outperforms it in terms of average accuracy. These results validate the effectiveness of SAC, which simultaneously enhances both color and texture complexity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.712, + 0.774 + ], + "angle": 0, + "content": "4.4. Analysis Experiment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Ablation Study. We provide the ablation study on theDigits dataset to evaluate the effectiveness of key components in SAC, as shown in Tab. 5. Specifically, we introduce some variants to analyze the contributions of SAC, ARC, the Gaussian filter \\(\\mathbf{g}\\), \\(\\mathcal{L}_d\\), and \\(\\mathcal{L}_p\\). When we remove a component, all others remain default settings. We first delete ARC (SAC W/O ARC) to assess the effectiveness of preserving object integrity. We observe that the performance" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15622" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.089, + 0.483, + 0.119 + ], + "angle": 0, + "content": "Table 5. Ablation study on Digits. Each column title indicates the target domain." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.128, + 0.48, + 0.248 + ], + "angle": 0, + "content": "
MethodsSVHNMNIST-MSYNUSPSAvg.
SAC W/O ARC64.179.175.192.777.7
SAC W/O PAA74.985.386.096.185.3
SAC W/O g77.688.886.594.286.8
SAC W/O Ld75.590.387.096.787.4
SAC W/O Lp74.989.786.796.787.0
SAC W/O Ada76.390.585.595.486.9
SAC W/O Ada max73.989.885.495.386.1
SAC77.791.087.996.488.3
" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.279, + 0.483, + 0.625 + ], + "angle": 0, + "content": "drops seriously, which indicates the harm of illogical textures. Next, to investigate the influence of PAA, we only employ ARC (SAC W/O PAA) to generate images. The lower accuracy compared to SAC suggests that focusing solely on texture diversity is insufficient to enhance overall style diversity. We then remove the Gaussian filter \\(\\mathbf{g}\\) (SAC W/O \\(\\mathbf{g}\\)) to evaluate its effectiveness in preserving object integrity. The degraded performance demonstrates that \\(\\mathbf{g}\\) can help maintain object integrity and prevent texture inconsistencies. Moreover, we see that eliminating \\(\\mathcal{L}_d\\) (SAC W/O \\(\\mathcal{L}_d\\)) or \\(\\mathcal{L}_p\\) (SAC W/O \\(\\mathcal{L}_p\\)) would lead to inferior performance compared to SAC, highlighting the importance of expanding the discrepancy at both the feature and pixel levels. Finally, we remove Eq. (9) (SAC W/O Ada) and follow the settings in [4] by iterating the random convolutions \\(T\\) times, where \\(T\\) is sampled from the uniform distribution \\(U\\) [1, 10]. The inferior accuracy indicates that our proposed adaptive strategy effectively improves texture diversity while preventing semantic distortion. To further demonstrate its effectiveness in preserving semantic information, we conduct an experiment in which \\(T\\) is fixed at 10 (SAC W/O \\(\\mathrm{Ada}_{\\mathrm{max}}\\)). The approximately \\(2\\%\\) performance drop confirms the importance of Eq. (9)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.483, + 0.793 + ], + "angle": 0, + "content": "Complementarity with Other Approaches. SAC can serve as a plug-and-play component to enhance the generalizability of other methods. Specifically, we apply PAA and ARC to the approaches of PDEN, L2D, RandConv, ProRandConv, and PSDG, respectively. The experimental results on Digits are shown in Tab. 6. These results indicate that ARC can serve as a general enhancer, improving the generalizability of other methods. Notably, when we apply ARC to PDEN, we find that \"PDEN+ARC\" achieves an \\(8.1\\%\\) improvement of average accuracy under default settings compared to PDEN." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Comparative Analysis of Different Convolutions. In Tab. 7, we evaluate the effectiveness of various convolution types on the Digits dataset, including conventional convolution [17] (SAC-Conv), depthwise separable convolution [5] (SAC-DSCov), wavelet convolution [8] (SAC-WTConv), and deformable convolution [6] (SAC). SAC and its variants achieve outstanding performance, highlighting the ef" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.132 + ], + "angle": 0, + "content": "Table 6. Accuracy (\\%) on Digits when SAC is applied to recent methods as a plug-and-play component. Each column title indicates the target domain." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.142, + 0.905, + 0.283 + ], + "angle": 0, + "content": "
MethodsSVHNMNIST-MSYNUSPSAvg.
PDEN [20]62.282.269.485.374.8
PDEN+SAC73.886.781.590.183.0
L2D [34]62.987.363.784.074.5
L2D+SAC59.882.270.693.176.4
PSDG [40]73.186.783.495.784.7
PSDG+SAC75.890.885.096.387.0
RandConv [36]62.187.963.984.474.6
RandConv+ARC68.286.871.085.878.0
Pro-RandConv [4]69.782.379.893.781.4
Pro-RandConv+ARC70.585.982.795.383.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.302, + 0.905, + 0.332 + ], + "angle": 0, + "content": "Table 7. Performance comparison of different convolution on Digits. Each column title indicates the target domain." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.341, + 0.902, + 0.41 + ], + "angle": 0, + "content": "
MethodsSVHNMNIST-MSYNUSPSAvg.
SAC-Conv [36]78.387.387.492.586.4
SAC-DSConv [5]78.789.187.492.486.9
SAC-WTConv [8]74.387.883.895.985.5
SAC77.791.087.996.488.3
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.442, + 0.907, + 0.624 + ], + "angle": 0, + "content": "fectiveness of our proposed \"Split-And-Combine\" strategy. Specifically, the conventional convolution and depthwise separable convolution remain constrained to uniform grids, limiting their ability to accurately capture object contours and preserve object integrity. While the wavelet convolution expands the receptive field through wavelet transformation, it still extracts features within a uniform grid, making it inherently constrained by grid boundaries. In contrast, the deformable convolution overcomes these limitations by allowing sampling points to deviate from a fixed uniform grid, thereby capturing object shapes and structural details more effectively and contributing to texture smoothing." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.646, + 0.634, + 0.663 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We present a novel \"Split-and-Combine\" method for single domain generalization. Instead of applying style augmentation to entire images, our method first performs patch-aware augmentation, which conducts separate style generation to each patch of an image, enriching colors. Then, we introduce adaptive random convolutions to enhance texture complexity while preserving object integrity. Through these synergistic components, our model can synthesize augmented data with different local colors and textures. Additionally, we design an energy-based distribution discrepancy score to assess distribution differences between the original and augmented data, which can encourage SAC to generate novel data with large distribution shifts. SAC consistently achieves state-of-the-art performance across four datasets, demonstrating the effectiveness of our method." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15623" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.486, + 0.283 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China under Grants 62306008, 32472007, 62406095, and 62402004, the Key Project of Anhui Province's Science and Technology Innovation Tackle Plan under Grant 202423k09020040, the China Postdoctoral Science Foundation under Grant 2024M762632, the Natural Science Foundation of Shaanxi under Grant 2024JC-YBQN-0687, the Social Science Foundation of Shaanxi under Grant 2024R027, and the Hefei Key Technology R&D \"Champion-Based Selection\" Project under Grant 2023SGJ011." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.296, + 0.188, + 0.312 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.321, + 0.484, + 0.376 + ], + "angle": 0, + "content": "[1] Kartik Ahuja, Karthikeyan Shanmugam, Kush R. Varshney, and Amit Dhurandhar. Invariant risk minimization games. In International Conference on Machine Learning, Virtual Event, July 13-18, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.379, + 0.484, + 0.46 + ], + "angle": 0, + "content": "[2] Kartik Ahuja, Ethan Caballero, Dinghuai Zhang, Jean-Christophe Gagnon-Audet, Yoshua Bengio, Ioannis Mitliagkas, and Irina Rish. Invariance principle meets information bottleneck for out-of-distribution generalization. In Neural Information Processing Systems, virtual, December 6-14, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.463, + 0.484, + 0.532 + ], + "angle": 0, + "content": "[3] Tianle Chen, Mahsa Baktashmotlagh, Zijian Wang, and Mathieu Salzmann. Center-aware adversarial augmentation for single domain generalization. In Winter Conference on Applications of Computer Vision, Waikoloa, HI, USA, January 2-7, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.534, + 0.484, + 0.616 + ], + "angle": 0, + "content": "[4] Seokeon Choi, Debasmit Das, Sungha Choi, Seunghan Yang, Hyunsin Park, and Sungrack Yun. Progressive random convolutions for single domain generalization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, Vancouver, BC, Canada, June 17-24, pages 10312-10322, 2023. 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.618, + 0.484, + 0.672 + ], + "angle": 0, + "content": "[5] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Conference on Computer Vision and Pattern Recognition, Honolulu, HI, USA, July 21-26, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.675, + 0.484, + 0.731 + ], + "angle": 0, + "content": "[6] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In International Conference on Computer Vision, Venice, Italy, October 22-29, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.733, + 0.484, + 0.787 + ], + "angle": 0, + "content": "[7] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In International Conference on Computer Vision, Venice, Italy, October 22-29, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.789, + 0.484, + 0.844 + ], + "angle": 0, + "content": "[8] Shahaf E. Finder, Roy Amoyal, Eran Treister, and Oren Freifeld. Wavelet convolutions for large receptive fields. In European Conference on Computer Vision, MiCo Milano, September 29 - October 4, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[9] Will Grathwohl, Kuan-Chieh Wang, Jorn-Henrik Jacobsen, David Duvenaud, Mohammad Norouzi, and Kevin Swersky. Your classifier is secretly an energy based model and you should treat it like one. In International Conference on" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.321, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "Learning Representations, Addis Ababa, Ethiopia, April 26-30, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[10] Kehua Guo, Rui Ding, Tian Qiu, Xiangyuan Zhu, Zheng Wu, Liwei Wang, and Hui Fang. Single domain generalization via unsupervised diversity probe. In ACM International Conference on Multimedia, Ottawa, ON, Canada, October 29-November 3, pages 2101-2111, 2023. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Conference on Computer Vision and Pattern Recognition, Las Vegas, NV, USA, June 27-30, pages 770-778, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[12] Xilin He, Jingyu Hu, Qinliang Lin, Cheng Luo, Weicheng Xie, Siyang Song, Muhammad Haris Khan, and Linlin Shen. Towards combating frequency simplicity-biased learning for domain generalization. In Conference on Neural Information Processing Systems, Vancouver, BC, Canada, December 10 - 15, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[13] Dan Hendrycks and Thomas G. Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, New Orleans, LA, USA, May 6-9, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.387, + 0.905, + 0.441 + ], + "angle": 0, + "content": "[14] Xun Huang and Serge J. Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In International Conference on Computer Vision, Venice, Italy, October 22-29, pages 1510-1519, 2017. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.443, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[15] Vladimir Koltchinskii. Oracle inequalities in empirical risk minimization and sparse recovery problems: École D'Été de Probabilités de Saint-Flour XXXVIII-2008. Springer Science & Business Media, 2011. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[16] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In Conference on Neural Information Processing Systems, Lake Tahoe, Nevada, United States, December 3-6, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.905, + 0.608 + ], + "angle": 0, + "content": "[17] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 1998. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.61, + 0.905, + 0.664 + ], + "angle": 0, + "content": "[18] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M. Hospedales. Deeper, broader and artier domain generalization. In IEEE International Conference on Computer Vision, Venice, Italy, October 22-29, pages 5542-5550, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[19] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M. Hospedales. Learning to generalize: Meta-learning for domain generalization. In Proceedings of Conference on Artificial Intelligence, New Orleans, Louisiana, USA, February 2-7, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[20] Lei Li, Ke Gao, Juan Cao, Ziyao Huang, Yepeng Weng, Xiaoyue Mi, Zhengze Yu, Xiaoya Li, and Boyang Xia. Progressive domain expansion network for single domain generalization. In Conference on Computer Vision and Pattern Recognition, virtual, June 19-25, pages 224-233, 2021. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[21] Weitang Liu, Xiaoyun Wang, John D. Owens, and Yixuan Li. Energy-based out-of-distribution detection. In Conference on Neural Information Processing Systems, virtual, December 6-12, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[22] Krikamol Muandet, David Balduzzi, and Bernhard Scholkopf. Domain generalization via invariant feature" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15624" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "representation. In Proceedings of International Conference on Machine Learning Atlanta, GA, USA, June 16-21, pages 10-18, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[23] Cheng Ouyang, Chen Chen, Surui Li, Zeju Li, Chen Qin, Wenjia Bai, and Daniel Rueckert. Causality-inspired single-source domain generalization for medical image segmentation. IEEE Transactions on Medical Imaging, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[24] Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. IEEE Transactions on Knowledge and Data Engineering, 22(10):1345-1359, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.482, + 0.291 + ], + "angle": 0, + "content": "[25] Sinno Jialin Pan, Ivor W. Tsang, James T. Kwok, and Qiang Yang. Domain adaptation via transfer component analysis. In International Joint Conference on Artificial Intelligence, Pasadena, California, USA, July 11-17, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.482, + 0.361 + ], + "angle": 0, + "content": "[26] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In International Conference on Computer Vision, Seoul, Korea (South), October 27 - November 2, pages 1406-1415. IEEE, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.364, + 0.482, + 0.419 + ], + "angle": 0, + "content": "[27] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In Conference on Computer Vision and Pattern Recognition, Seattle, WA, USA, June 13-19, pages 12553-12562, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.421, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[28] Alexander Robey, George J. Pappas, and Hamed Hassani. Model-based domain generalization. In Conference on Neural Information Processing Systems, virtual, December 6-14, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.478, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[29] Ding Rui, Kehua Guo, Xiangyuan Zhu, Zheng Wu, and Hui Fang and. Progressive diversity generation for single domain generalization. IEEE Transactions on Multimedia, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.535, + 0.482, + 0.604 + ], + "angle": 0, + "content": "[30] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott E. Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Conference on Computer Vision and Pattern Recognition, Boston, MA, USA, June 7-12, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.606, + 0.482, + 0.66 + ], + "angle": 0, + "content": "[31] Chris Xing Tian, Haoliang Li, Xiaofei Xie, Yang Liu, and Shiqi Wang. Neuron coverage-guided domain generalization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):1302-1311, 2023. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.663, + 0.482, + 0.745 + ], + "angle": 0, + "content": "[32] Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C. Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. In Conference on Neural Information Processing Systems, Montreal, Canada, December 3-8, pages 5339-5349, 2018. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[33] Chaoqun Wan, Xu Shen, Yonggang Zhang, Zhiheng Yin, Xinmei Tian, Feng Gao, Jianqiang Huang, and Xian-Sheng Hua. Meta convolutional neural networks for single domain generalization. In Conference on Computer Vision and Pattern Recognition, New Orleans, LA, USA, June 18-24, pages 4672-4681, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[34] Zijian Wang, Yadan Luo, Ruihong Qiu, Zi Huang, and Mahsa Baktashmotlagh. Learning to diversify for single domain generalization. In International Conference on Computer Vision, Montreal, QC, Canada, October 10-17, pages 814-823, 2021. 3, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.146 + ], + "angle": 0, + "content": "[35] Cort J Willmott and Kenji Matsuura. Advantages of the mean absolute error (mae) over the root mean square error (rmse) in assessing average model performance. Climate research, 2005. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.216 + ], + "angle": 0, + "content": "[36] Zhenlin Xu, Deyi Liu, Junlin Yang, Colin Raffel, and Marc Niethammer. Robust and generalizable visual representation learning via random convolutions. In International Conference on Learning Representations, virtual, Austria, May 3-7, 2021. 2, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.286 + ], + "angle": 0, + "content": "[37] Fu-En Yang, Yuan-Chia Cheng, Zu-Yun Shiau, and Yu-Chiang Frank Wang. Adversarial teacher-student representation learning for domain generalization. In Conference on Neural Information Processing Systems, December 6-14, virtual, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[38] Shuai Yang, Kui Yu, Fuyuan Cao, Hao Wang, and Xindong Wu. Dual-representation-based autoencoder for domain adaptation. IEEE Transactions on Cybernetics, 52(8): 7464-7477, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[39] Shuai Yang, Kui Yu, Fuyuan Cao, Lin Liu, Hao Wang, and Jiuyong Li. Learning causal representations for robust domain adaptation. IEEE Transactions on Knowledge and Data Engineering, 35(3):2750-2764, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.467 + ], + "angle": 0, + "content": "[40] Shuai Yang, Zhen Zhang, and Lichuan Gu. Practical single domain generalization via training-time and test-time learning. In ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Barcelona, Spain, August 25-29, 2024. 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[41] Haotian Ye, Chuanlong Xie, Tianle Cai, Ruichen Li, Zhenguo Li, and Liwei Wang. Towards a theoretical framework of out-of-distribution generalization. In Neural Information Processing Systems, virtual, December 6-14, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[42] Jiajin Zhang, Hanqing Chao, Amit Dhurandhar, Pin-Yu Chen, Ali Tajer, Yangyang Xu, and Pingkun Yan. When neural networks fail to generalize? A model sensitivity perspective. In AAAI Conference on Artificial Intelligence, Washington, DC, USA, February 7-14, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.662 + ], + "angle": 0, + "content": "[43] Zuyu Zhang, Yan Li, and Byung-Seok Shin. Embracing domain gradient conflicts: Domain generalization using domain gradient equilibrium. In International Conference on Multimedia, Melbourne, VIC, Australia, October 28 - November 1, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[44] Long Zhao, Ting Liu, Xi Peng, and Dimitris N. Metaxas. Maximum-entropy adversarial data augmentation for improved generalization and robustness. In Advances in Neural Information Processing Systems, virtual, December 6-12, pages 1-13, 2020. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[45] Guangtao Zheng, Mengdi Huai, and Aidong Zhang. Advst: Revisiting data augmentations for single domain generalization. In Thirty-Eighth AAAI Conference on Artificial Intelligence, February 20-27, pages 1-9, 2024. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[46] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4396-4415, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[47] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A comprehensive survey on transfer learning. Proceedings of the IEEE, 109(1):43-76, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15625" + } + ] +] \ No newline at end of file diff --git a/2025/Split-and-Combine_ Enhancing Style Augmentation for Single Domain Generalization/3714ebba-fcba-4212-aaf3-03100ed83e52_origin.pdf b/2025/Split-and-Combine_ Enhancing Style Augmentation for Single Domain Generalization/3714ebba-fcba-4212-aaf3-03100ed83e52_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..564c256a1a2a3b12c085701a8247162193f6a876 --- /dev/null +++ b/2025/Split-and-Combine_ Enhancing Style Augmentation for Single Domain Generalization/3714ebba-fcba-4212-aaf3-03100ed83e52_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1eecad84c7c5abda2f323828f583dab1f80ef9571e8eb0e78ce1df59a57dd253 +size 10386157 diff --git a/2025/Split-and-Combine_ Enhancing Style Augmentation for Single Domain Generalization/full.md b/2025/Split-and-Combine_ Enhancing Style Augmentation for Single Domain Generalization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..91bb636f10458bec894083016fdbbde0c5b2a99f --- /dev/null +++ b/2025/Split-and-Combine_ Enhancing Style Augmentation for Single Domain Generalization/full.md @@ -0,0 +1,344 @@ +# Split-and-Combine: Enhancing Style Augmentation for Single Domain Generalization + +Zhen Zhang $^{1}$ , Shuai Yang $^{1\dagger}$ , Qianlong Dang $^{2}$ , Zhize Wu $^{3}$ , Lichuan Gu $^{1\dagger}$ $^{1}$ Anhui Agricultural University $^{2}$ Northwest A & F University $^{3}$ Hefei University + +zhangz@stu.ahau.edu.cn, yangs@ahau.edu.cn, dangqianlong@nwafu.edu.cn, + +wuzz@hfuu.edu.cn, glc@ahau.edu.cn + +# Abstract + +Single domain generalization aims to learn a model with good generalization ability from a single source domain. Recent advances in this field have focused on increasing the diversity of the training data through style (e.g., color and texture) augmentation. However, most existing methods apply uniform perturbations to the entire image, failing to simulate complex images with multiple distinct stylistic regions. To address this, we propose a "Split-And-Combine" (SAC) strategy to enhance style diversity. Specifically, SAC first performs patch-aware augmentation, which splits an image into multiple patches and applies style augmentation independently to each patch, enabling distinct color variations across regions. Then, SAC combines these patches to reconstruct a complete image and applies adaptive random convolutions, which utilizes a deformable convolution layer with random and Gaussian filters to enhance texture diversity while preserving object integrity. Notably, SAC leverages entropy as a risk assessment criterion to adaptively determine whether a sample should undergo augmentation within the iterative process of random convolutions, preventing excessive augmentation. Furthermore, SAC introduces an energy-based distribution discrepancy score to quantify out-of-distribution likelihood, systematically expanding the augmented data's distribution. SAC can serve as a plug-and-play component to improve the performance of recent methods. Extensive experiments on four datasets demonstrate the effectiveness of SAC. + +# 1. Introduction + +Modern machine learning algorithms [11, 16, 30] have achieved remarkable performance under the assumption that the source (training) and target (test) domains share the same distribution. However, in reality, such an assumption is often violated due to factors such as variations in + +illumination, background, or weather, which are known as the domain shift problem [41]. Unfortunately, these traditional algorithms often face significant performance degradation when there are distributional discrepancies between the source and target domains [24, 25, 38, 39, 47]. + +To tackle this problem, multi-source domain generalization [19, 22, 28, 46] has been presented, aiming to learn a robust model by using knowledge from multiple different but related source domains without accessing target domain data. The mainstream works depend heavily on domain identifiers as supervision signals to capture domain invariant representations. However, the collection and annotation of data from multiple domains can be prohibitively expensive, making multi-source domain generalization less feasible in practical applications. To alleviate this issue, a more challenging and realistic setting known as single domain generalization [31, 33, 43] has emerged and garnered significant attention recently, which only leverages a single source domain to learn a model with good generalizability. + +The difficulty in single domain generalization arises from mitigating distributional discrepancies between the source and unknown target domains. In reality, it is observed that content features, such as object features, remain relatively stable despite environmental changes, whereas style features, such as backgrounds, exhibit considerable variability across different domains [1, 2]. For example, in animal image classification, fundamental elements like eyes, ears, and face shape constitute the content, whereas style elements include background and lighting conditions. Therefore, the key to effective single domain generalization is to eliminate spurious correlations between style features and labels. The current mainstream research on single domain generalization primarily focuses on increasing the capacity of the training data through the generation of novel domains, thereby reducing the model's reliance on style features due to the increased diversity of styles within the training data. One category of methods [3, 32, 45] perturb the original data using the back-propagated gradient to obtain augmented data. However, gradient-based perturba + +![](images/13392f7c48309c7ff72b485abe5e59cec212dfcda5d5150947dbb820a1f30104.jpg) +(a) UDP + +![](images/9066d5d6e3caf5b114b20fdbb3ab6b0fae3a724d0ffde415e917ab5933ee559b.jpg) +(b) PSDG + +![](images/e032456b63610bdfffbf40b8b7a7a0746f90c1912bad734b944bb3f4bef2f114.jpg) +(c) RandConv + +![](images/41423f839c1cfd1aa6654a908c3e8755979cecea770708fd958ed47358e053b4.jpg) +(d) Pro-RandConv + +![](images/9cc13c06a9c6ce5afd3758a3db6698f48c03d8fc5990d850f5ddab83772e9882.jpg) +(e) Patch +Figure 1. Examples of images generated by different methods. + +![](images/fe05c9a4da7a52b5bc8b5282f6f495a9ffaf44f35fc3c8815b7c026c2960b8ad.jpg) +(f) SAC + +tions are visually imperceptible, rendering these methods inadequate for simulating real-world domain shifts. Another mainstream category trains a generative model to synthesize novel data with style (e.g., color and texture) shifts [4, 10, 20], resulting in visually distinct samples. + +Recent methods leveraging generative models for data augmentation have achieved remarkable success by enhancing color and texture diversity. For instance, several approaches [10, 29, 40] introduce Gaussian noise to perturb images, effectively enriching the color distribution within a single generated image. Alternatively, other methods [4, 36] utilize a convolution layer with random filters to generate images with diverse local textures. However, these methods enhance style diversity from only a single aspect, either color or texture, leading to inherent limitations in the other aspect. For example, Gaussian noise-based methods, such as UDP [10] and PSDG [40], only generate simplistic textures (see Fig. 1 (a) and (b)), whereas methods leveraging random convolutions, such as RandConv [36] and ProRandConv [4], tend to produce images dominated by a single color (see Fig. 1 (c) and (d)). In contrast, real-world scenarios typically involve a wide range of colors and complex textures, with different regions of an image (e.g., a person's clothing, grass, trees, and the sky within a park), each exhibiting unique texture and color characteristics. Therefore, improving style diversity solely from a single perspective cannot simulate images from complex real-world scenarios. This raises a crucial challenge: Can we combine the advantages of Gaussian noise and random convolutions to generate images with both rich colors and complex textures? + +In this paper, we propose a "Split-And-Combine" (SAC) strategy to improve style diversity, which first splits images into a series of patches to enhance the color diversity, then combines the augmented patches into the complete image and uses adaptive random convolutions to improve the texture diversity. In the split step, we propose patch-aware augmentation (PAA) that introduces rich color variations. Specifically, we divide an image into several patches and each patch undergoes independent style generation within Gaussian noise, yielding generated images + +with different local colors. While PAA effectively generates diverse styles for each patch, it leads to texture inconsistencies in a single object, resulting in the model misinterpreting the object as a collection of fragmented parts rather than a unified entity (see Fig. 1 (e)). For instance, a tree trunk may exhibit a smooth texture on one side and a rough, uneven texture on the other, leading the model to perceive it as two distinct components instead of a coherent whole. To address this challenge, we propose adaptive random convolutions (ARC) to enhance texture complexity and restore object integrity in the combine step. Specifically, we first initialize a deformable convolution layer with both random filters and Gaussian filters. Subsequently, we iteratively apply this layer and affine transformation to generate random textures, progressively increasing texture complexity. Moreover, we use entropy values to adaptively control this progressive process, preventing excessive application of random convolutions that could distort semantic information. Furthermore, we use an energy-based distribution discrepancy score to amplify the distribution differences between the original and generated data, enabling the synthesis of out-of-distribution data. We conduct experiments on four well-established benchmarks, where SAC consistently achieves state-of-the-art performance in various scenarios. Overall, our contributions can be summarized as follows: + +- We propose a "Split-And-Combine" strategy that first splits images into patches for color augmentation and then combines the patches to undergo texture augmentation. +- We design the patch-aware augmentation that divides an image into a series of patches for separate style transformation to promote color diversity for generated images. +- We develop the adaptive random convolution that utilizes entropy values to adaptively control the progressive process of the random deformable convolution to improve the texture diversity. +- We present an energy-based perspective to maximize the distribution discrepancy between the original and augmented data, generating out-of-distribution data. +- We conduct experiments on four established datasets. The results demonstrate that the proposed SAC achieves consistent state-of-the-art performance. + +# 2. Related Work + +Recently, since only requiring a single training domain, single domain generalization has attracted substantial interest, which can be roughly categorized into two paradigms. + +Gradient-based augmentation methods. This paradigm of work attempts to generate data using backpropagated gradients. One approach involves directly adding gradients to the original data, thereby synthesizing augmented data. For example, ADA [32] employs backpropagated gradients to synthesize data. Subsequently, M-ADA [27] enhances ADA with a Wasserstein Auto + +Encoder to increase the distribution differences between the augmented and original data. Similarly, ME-ADA [44] introduces a maximum-entropy regularizer to amplify distribution shifts in augmented data by minimizing mutual information between augmented and original data. Later on, NCDG [31] presents neuron coverage to improve generalization performance by maximizing gradient similarity between the original and augmented data. CAA [3] relies on angular distance to push generated data away from class centroids. AdvST [45] uses back-propagated gradients as guidance to fully exploit the effectiveness of several standard data augmentation methods. However, these methods fail in target domains with large domain shifts due to visually imperceptible style variations. + +Style-based augmentation methods. This paradigm of work trains a generative model to produce data with different styles (e.g., contrast, colors, and textures). The most well-established method is PDEN [20], which progressively trains multiple generative models to synthesize different augmented domains and uses cycle consistency loss to preserve semantic information. Additionally, L2D [34] generates data and captures invariant features through a max-min game of mutual information. ATS [37] introduces a teacher-student network to learn domain-generalized knowledge by minimizing the discrepancy between the normalized features of the original and augmented data. CSDG [23] simultaneously changes the textures and intensities of the original data to generate new data. However, due to the limited styles provided by AdaIN [14], these methods only synthesize data with simple colors and textures, which limits their applicability in real-world scenarios. To enrich the styles, Pro-RandConv [4] introduces a deformable convolution block to improve diversification by applying it along with affine transformation to generate data with new styles. UDP [10] adds extra Gaussian noise to perturb the data for more complex styles and enhances the shifts in generated data by maximizing the uncertainty between the generated and original data from an information theory perspective. Subsequently, PDG [29] enhances UDP by modifying the structure of the generative model, using $1 \times 1$ convolutions to preserve the input shape while employing random and larger kernel sizes to alter the local texture. PSDG [40] was recently proposed to improve generalizability by optimizing the model during both training and testing. AdvFrequency [12] improves frequency discrepancies between the generated and original data through the use of adversarial learning, thereby reducing the risk of the particular frequency. + +Previous methods typically enhance style diversity from a single perspective, focusing on either color or texture, and fail to simulate the complexity of real-world images. In contrast, SAC improves both color and texture diversity simultaneously through PAA and ARC, thereby enhancing the model's ability to capture invariant representations. + +# 3. Method + +Problem Setup. Consider single domain generalization defined on source domain $D_{s} = \{\mathbf{x}_{i},\mathbf{y}_{i}\}_{i = 1}^{n}$ , where $\mathbf{x}_i$ is the $i$ -th sample, $\mathbf{y}_i$ is the corresponding label of $\mathbf{x}_i$ , and $n$ is the number of samples. The task is to train a model with good generalizability in the unknown target domain $D_{t}$ . + +Overall Framework. We propose the "Split-And-Combine" strategy to synthesize images with rich colors and complex textures for single domain generalization tasks. SAC comprises two key components: a style generation module, which synthesizes data with novel styles, and a representation learning module, which learns domain invariant features. The style generation module consists of an Encoder, a generative model $\mathcal{T}(\cdot)$ , and a Decoder. To enhance color and texture diversity, we introduce patch-aware augmentation and adaptive random convolutions, enabling the generation of more complex styles. Additionally, we design a score to expand domain distribution through energy discrepancies. The representation learning module, which consists of a feature extractor $F$ and a classifier $P$ , combines both original and augmented data to learn invariant representations. The framework of SAC is shown in Fig. 2. + +# 3.1. Style Generation Module + +As we proposed before, the key to solving single domain generalization is to synthesize data with rich colors and complex textures. However, the augmented data generated by previous methods contains simple textures or colors, failing to simulate images in real-world scenarios. Unlike previous methods, we independently perform style augmentation on each patch to enrich colors and use adaptive random deformable convolutions to enhance texture complexity. + +Patch-Aware Augmentation. To improve the color diversity, we propose patch-aware augmentation to ensure that the images have distinct local colors. Given an image $\mathbf{x} \in \mathbb{R}^{H \times W}$ , we first use an encoder to project the image into the latent space to obtain rich feature representations: + +$$ +\mathbf {r} = \operatorname {E n c o d e r} (\mathbf {x}), \tag {1} +$$ + +where $\mathbf{r}$ is from the encoder, which consists of learnable deep layers instantiated as two convolutional layers and each with a ReLU activation function behind. Afterward, we divide $\mathbf{r}$ into $N\times N$ patches. Each patch $p_{i,j}\in \mathbb{R}^{\frac{H}{N}\times \frac{W}{N}}$ can be written as: + +$$ +\begin{array}{l} p _ {i, j} = \mathbf {r} [ h _ {i}: h _ {i + 1}, w _ {j}: w _ {j + 1} ], \quad i, j = 0, \dots , N - 1 \\ h _ {i} = \frac {H}{N} \times i, \quad w _ {j} = \frac {W}{N} \times j, \tag {2} \\ \end{array} +$$ + +where $h_i, w_j$ indicate the locations in the image. Next, following the convention in generative models [20, 40], we employ two fully connected layers $f_1(\cdot)$ and $f_2(\cdot)$ to learn a randomly initialized Gaussian noise, which can simulate an + +![](images/bc0ca0cbca69ea061010a9fdf17d765b80997e9a5d9e5763e21d246d27a5da2f.jpg) +Figure 2. Framework of SAC, where "PAA" is patch-aware augmentation, "ARC" is adaptive random convolutions, "En" is the Encoder, "De" is the Decoder, $F$ is the feature extractor, and $P$ is the classifier. + +unknown style. Notably, as styles generated in this process are simple, we go beyond previous methods and present to separately apply style augmentation to each patch rather than the entire image: + +$$ +\tilde {p} _ {i, j} = \mathcal {T} (p _ {i, j}), \quad i, j = 0, \ldots , N - 1 +$$ + +$$ +\mathcal {T} \left(p _ {i, j}\right) = \left(1 + f _ {1} \left(\varepsilon_ {i, j}\right)\right) \left(\frac {p _ {i , j} - \mu \left(p _ {i , j}\right)}{\sigma \left(p _ {i , j}\right)} + \eta_ {i, j}\right) + f _ {2} \left(\varepsilon_ {i, j}\right), \tag {3} +$$ + +where $\mu (\cdot)$ and $\sigma (\cdot)$ are mean and standard variance respectively, and $\varepsilon_{i,j},\eta_{i,j}\sim \mathcal{N}(0,1)$ . Here, we adopt $\eta_{i,j}$ to perturb the semantic information and a constant 1 to prevent large semantic shifts. Notably, the generative model $\mathcal{T}(\cdot)$ is flexible, which can be replaced by AdaIN [14], AdaINnoise [10], or StyIN [40], etc. This flexibility allows SAC to serve as a plug-and-play component for other methods. Subsequently, we combine each patch in its original order: + +$$ +\begin{array}{l} \tilde {\mathbf {r}} = \operatorname {C o n c a t} \left(\left[ \tilde {p} _ {i, j} \right] _ {i, j = 0, \dots , N - 1}\right) \\ = \left[ \begin{array}{c c c c} \tilde {p} _ {0, 0} & \tilde {p} _ {0, 1} & \dots & \tilde {p} _ {0, N - 1} \\ \tilde {p} _ {1, 0} & \tilde {p} _ {1, 1} & \dots & \tilde {p} _ {1, N - 1} \\ \vdots & \vdots & \ddots & \vdots \\ \tilde {p} _ {N - 1, 0} & \tilde {p} _ {N - 1, 1} & \dots & \tilde {p} _ {N - 1, N - 1} \end{array} \right]. \tag {4} \\ \end{array} +$$ + +By projecting back to the pixel space using the decoder, we can obtain an augmented image $\tilde{\mathbf{x}}\in \mathbb{R}^{H\times W}$ + +$$ +\tilde {\mathbf {x}} = \operatorname {D e c o d e r} (\tilde {\mathbf {r}}), \tag {5} +$$ + +where $\tilde{\mathbf{x}}$ is from the decoder, which is learnable deep layers instantiated as two convolutional layers, followed by ReLU + +and Sigmoid activation function respectively. In summary, our patch-aware augmentation is: + +$$ +\tilde {\mathbf {x}} = \operatorname {P A A} (\mathbf {x}, \mathcal {Q}), +$$ + +$$ +\text {w h e r e} \quad \mathcal {Q} = \left\{\left\{\varepsilon_ {i, j} \right\} _ {i, j = 0} ^ {N - 1}, \left\{\eta_ {i, j} \right\} _ {i, j = 0} ^ {N - 1} \right\}. \tag {6} +$$ + +By applying distinct Gaussian noise to different patches, each patch acquires a unique color, resulting in an image with different local colors. However, $\tilde{\mathbf{x}}$ cannot be directly used to capture domain invariant representations, since the introduction of illogical styles during patch-aware augmentation. Specifically, nearby pixels with similar textures can be grouped into primitive shapes that represent parts of objects in images [36]. For instance, in animal classification, an animal's eyes as a whole should present a similar texture. However, this consistency can be disrupted by patch-aware augmentation, which relies on the uniform grid division. When an object spans multiple patches, independent style transformations can lead to conflicting texture regions, causing the model to misinterpret the object as fragmented components rather than a unified whole. + +Adaptive Random Convolutions. To preserve object integrity, we introduce adaptive random convolutions. Inspired by [4, 36], we use the deformable convolution layer with random filters to project random and coherent texture into images. Furthermore, we multiply the convolution kernel element-wise with a Gaussian filter to further smooth the texture in an object spanning multiple patches. Additionally, we choose the deformable convolution rather than the conventional convolution since the deformable convo + +lution can better sample object contours instead of being confined to patch boundaries [7], helping smooth the texture. (see Tab. 7 for ablation studies on convolution types). Given an augmented sample $\tilde{\mathbf{x}}$ , we first initialize a $3 \times 3$ kernel with dilation 1, which defines a regular grid $\mathcal{R}$ as $\{(-1, -1), (-1, 0), \dots, (0, 1), (1, 1)\}$ . Here, $\mathcal{R}$ represents the fixed sampling locations of a standard $3 \times 3$ convolution kernel, where each coordinate pair $(i_m, j_m)$ denotes the fixed offsets relative to the center pixel $(i_0, j_0)$ . In deformable convolution, these offsets are perturbed by random values $(\Delta i_m, \Delta j_m)$ , enabling the sampling locations to deviate from the regular grid. We apply this deformable convolution to $\tilde{\mathbf{x}}$ : + +$$ +\begin{array}{l} \tilde{\mathbf{x}}^{\prime}[i_{0},j_{0}] = \sum_{(i_{m},j_{m})\in \mathcal{R}}\mathbf{w}[i_{m},j_{m}]\cdot \tilde{\mathbf{x}} [i_{0} + i_{m} + \Delta i_{m},j_{0} + j_{m} + \Delta j_{m}], \\ \text {w h e r e} \quad \mathbf {w} [ i _ {m}, j _ {m} ] = \mathbf {g} [ i _ {m}, j _ {m} ] \cdot w [ i _ {m}, j _ {m} ], \tag {7} \\ \end{array} +$$ + +where $i_m, j_m$ are the locations in $\mathcal{R}$ , $\mathbf{w}$ indicates the weights of the convolution kernel that are randomized, $\mathbf{g}[i_m, j_m] = \exp(-\frac{i_m^2 + j_m^2}{2\sigma_g^2})$ is a Gaussian filter, $w[i_m, j_m]$ is sampled from a Gaussian distribution $\mathcal{N}(0, \frac{1}{k^2C_{in}})$ , $k$ is the size of kernel size, $C_{in}$ is the number of input channels, and $\Delta i_m, \Delta j_m$ denote the random offsets of deformable convolution. Each location $(i_0, j_0)$ on $\tilde{\mathbf{x}}'$ is transformed by the weighted summation of $\mathbf{w}$ and the pixel values at irregular locations $(i_0 + i_m + \Delta i_m, j_0 + j_m + \Delta j_m)$ of $\tilde{\mathbf{x}}$ . + +Afterward, following the convention of random convolution [4, 36], we employ affine transformation with two parameters $\gamma$ and $\beta$ to further enhance the texture diversity. Given an input image $\tilde{\mathbf{x}}'$ , this process can be written as: + +$$ +\operatorname {A R C} (\tilde {\mathbf {x}}, \gamma , \beta) = \tanh (\operatorname {A d a I N} (\tilde {\mathbf {x}} ^ {\prime}, \gamma , \beta)), +$$ + +$$ +\operatorname {A d a I N} \left(\tilde {\mathbf {x}} ^ {\prime}, \gamma , \beta\right) = \gamma \frac {\tilde {\mathbf {x}} ^ {\prime} - \mu \left(\tilde {\mathbf {x}} ^ {\prime}\right)}{\sqrt {\sigma^ {2} \left(\tilde {\mathbf {x}} ^ {\prime}\right) + \epsilon}} + \beta , \tag {8} +$$ + +where $\mu (\cdot)$ and $\sigma (\cdot)$ represent the mean and variance, $\epsilon$ is a positive number, and $\gamma ,\beta \sim \mathcal{N}(0,1)$ . Then, the tanh is applied to scale the augmented images. We aim to iterate Eq. (7) and Eq. (8) $T$ times to progressively improve the texture diversity. However, this iterative process presents potential risks. Specifically, excessive iteration can distort semantic information and hinder the learning of invariant representations, while insufficient iterations may fail to adequately enhance texture diversity. Moreover, applying a uniform iteration count across all samples in a batch without considering individual discrepancies can lead to semantic distortions in some samples, degrading model performance. To this end, we introduce an adaptive strategy using the entropy to assess whether a sample requires augmentation during the iterative process. Specifically, for each batch of data, we iterate up to 10 times. In the first iteration, all samples undergo augmentation. In subsequent iterations, we + +compute the entropy of each sample and apply augmentation only to those with entropy values below a predefined threshold, while high-entropy samples remain unchanged. Entropy serves as a measure of uncertainty, where high entropy indicates a high likelihood of distribution shifts. Our strategy adaptively controls augmentation based on entropy, ensuring that low-entropy samples undergo augmentation to adequately enhance texture complexity, while high-entropy samples remain unchanged to prevent distortion of semantic information. Our strategy can be summarized as follows: + +$$ +\mathbf {x} _ {t + 1} ^ {+} = \left\{ \begin{array}{c l} \operatorname {A R C} (\mathbf {x} _ {t} ^ {+}, \gamma , \beta), & \text {i f} E (\mathbf {x} _ {t} ^ {+}) < E _ {0}, \\ \mathbf {x} _ {t} ^ {+}, & \text {i f} E (\mathbf {x} _ {t} ^ {+}) \geq E _ {0}, \end{array} \right. \tag {9} +$$ + +where $\mathbf{x}_t^+$ denotes the sample after the $t^{th}$ iteration ( $t \in [1,9]$ ), $E_0$ is the threshold defined as $e \times \log(C)$ (where $C$ represents the number of classes in the dataset and $e$ is set to 0.3) and $E(\cdot)$ denotes the entropy loss. + +Energy-based Distribution Discrepancy Score. To enhance the ability of the generative model $\mathcal{T}(\cdot)$ in the feature-level, we propose an energy-based distribution discrepancy score to facilitate the expansion of generated domain shifts using a trained classifier. To assess the energy, we use logits from the classifier for energy representations, which is widely used to evaluate Helmholtz free energy [9]. Given the input data $\mathbf{x}$ , and the trained $c$ -class classifier $P$ , where $c$ is the number of classes, we can compute the Helmholtz free energy for $\mathbf{x}$ : + +$$ +E n (\mathbf {x}) = - \log \sum_ {i} ^ {c} e ^ {P _ {i} (\mathbf {x})}, \tag {10} +$$ + +where $P_{i}(\cdot)$ denotes the probability of the class $i$ . It is known that the energy exhibits the uncertainty of the input data with respect to the classifier, where high energy indicates the high probability of being out-of-distribution [21]. Thus, the energy discrepancy can serve as the score to measure the distribution discrepancy between two data under the same view. Given input data $\mathbf{x}_1$ , $\mathbf{x}_2$ , and the trained $c$ -class classifier $P$ , we obtain the distribution discrepancy score by: + +$$ +S (\mathbf {x} _ {1}, \mathbf {x} _ {2}) = \frac {1}{\left\| E n (\mathbf {x} _ {1}) - E n (\mathbf {x} _ {2}) \right\| _ {1}}, \tag {11} +$$ + +where $\| \cdot \|_1$ is the mean absolute error [35]. Here, $S(\cdot, \cdot)$ represents the approximate distribution distance between two data. During training, we leverage this score to enhance the distribution discrepancy between the original data and generated data, as well as among different generated data, improving the expansion of the training data distribution. + +Training Objective. Since the parameters in ARC are randomly initialized, our loss term in the style generation module is only used to constrain PAA. To prevent semantic + +information from distortion, we use the cross-entropy to ensure that the augmented data can be correctly predicted by the model. Specifically, since the original and augmented data correspond to different styles but share the same semantic information, the model should assign them the same label if it classifies based on semantic information. + +$$ +\mathcal {L} _ {c e} = \frac {1}{n} \sum_ {i = 1} ^ {n} \left(\ell \left(P (F (\tilde {\mathbf {x}} _ {i, 1})), \mathbf {y} _ {i}\right) + \ell \left(P (F (\tilde {\mathbf {x}} _ {i, 2})), \mathbf {y} _ {i}\right)\right), +$$ + +where $\tilde{\mathbf{x}}_{i,1} = \mathrm{PAA}(\mathbf{x}_i,\mathcal{Q}_1),\quad \tilde{\mathbf{x}}_{i,2} = \mathrm{PAA}(\mathbf{x}_i,\mathcal{Q}_2),$ (12) + +where $\mathbf{x}_i$ indicates input original data, $\ell$ is the cross-entropy loss, $\mathbf{y}_i$ is the label for $\mathbf{x}_i$ , and $n$ is the number of available data. The augmented data $\tilde{\mathbf{x}}_{i,1}$ and $\tilde{\mathbf{x}}_{i,2}$ are synthesized using different parameter sets $\mathcal{Q}_1$ and $\mathcal{Q}_2$ , respectively. Next, we maximize distribution discrepancies among the original data $\mathbf{x}$ and different augmented data $\tilde{\mathbf{x}}_1$ , $\tilde{\mathbf{x}}_2$ , encouraging the generative model to consistently generate data that exhibit large domain shifts relative to the original data and previously generated data: + +$$ +\mathcal {L} _ {d} = \frac {1}{n} \sum_ {i = 1} ^ {n} \left(s \left(\tilde {\mathbf {x}} _ {i, 1}, \mathbf {x} _ {i}\right) + s \left(\tilde {\mathbf {x}} _ {i, 2}, \mathbf {x} _ {i}\right) + s \left(\tilde {\mathbf {x}} _ {i, 1}, \tilde {\mathbf {x}} _ {i, 2}\right)\right). \tag {13} +$$ + +Additionally, we use mean absolute error to compel the generative model to consistently generate data with pixel-level discrepancies: + +$$ +\mathcal {L} _ {p} = \frac {1}{n} \sum_ {i = 1} ^ {n} \| \tilde {\mathbf {x}} _ {i, 1} - \tilde {\mathbf {x}} _ {i, 2} \| _ {1}. \tag {14} +$$ + +In summary, the overall objective is as follows: + +$$ +\mathcal {L} _ {S} = \mathcal {L} _ {c e} + \lambda_ {d} \mathcal {L} _ {d} - \lambda_ {p} \mathcal {L} _ {p}, \tag {15} +$$ + +where $\lambda_{d}$ and $\lambda_{p}$ are the hyper-parameters. + +# 3.2. Representation Learning Module + +Benefiting from our style generation module, the representation learning module can be supervised by a simple cross-entropy loss function and learn domain invariant representations. The loss for each sample pair is: + +$$ +\begin{array}{l} \mathcal {L} _ {R} = \frac {1}{2 n} \sum_ {i = 1} ^ {n} \left(\ell \left(P \left(F \left(\mathbf {x} _ {i} ^ {+}\right)\right), \mathbf {y} _ {i}\right) + \ell \left(P \left(F \left(\mathbf {x} _ {i} ^ {+ +}\right)\right), \mathbf {y} _ {i}\right)\right) \\ + \frac {1}{n} \sum_ {i = 1} ^ {n} \ell \left(P (F (\mathbf {x} _ {i})), \mathbf {y} _ {i}\right), \\ \end{array} +$$ + +where $\mathbf{x}_i^+ = \mathrm{ARC}(\tilde{\mathbf{x}}_{i,1},\gamma_1,\beta_1),\mathbf{x}_i^{+ + } = \mathrm{ARC}(\tilde{\mathbf{x}}_{i,2},\gamma_2,\beta_2)$ (16) + +where $\mathbf{x}_i^+$ and $\mathbf{x}_i^{+ + }$ are synthesized from $\mathbf{x}_i$ through distinct parameters. + +Table 1. Accuracy (\%) of different methods on Digits. Each column title indicates the target domain. + +
MethodsSVHNMNIST-MSYNUSPSAvg.
ERM [15]27.852.739.776.949.3
ADA [32]NeurIPS'1835.560.445.377.354.6
M-ADA [27]CVPR'2042.667.949.078.559.5
ME-ADA [44]NeurIPS'2042.663.350.481.059.3
PDEN [20]CVPR'2162.282.269.485.374.8
L2D [34]ICCV'2162.987.363.784.074.5
NCDG [31]TPAMI'2359.777.463.892.673.4
SADA [42]AAAI'2368.575.672.989.376.6
Pro-RandConv [4]CVPR'2369.782.379.893.781.4
UDP [10]MM'2372.479.781.796.382.5
AdvST [45]AAAI'2467.579.878.195.480.1
PSDG [40]KDD'2473.186.783.495.784.7
AdvFrequency [12]NeurIPS'2473.682.779.989.381.4
SAC77.791.087.996.488.3
+ +# 3.3. Training Pipeline + +In real-world scenarios, since the styles of target domains exhibit complexity, using a single generated domain may not produce data with sufficient styles. To address this limitation, we employ a progressively adversarial max-min game to optimize the style generation and representation learning module. We use empirical risk minimization (ERM) to pretrain the representation learning module with original data. During training, we first fix the weights of the representation learning module and train the style generation module $S_{1}$ using PAA and ARC to synthesize the augmented data. Next, we fix the weights of the style generation module and utilize the original and augmented data to optimize the representation learning module using Eq. (16). This adversarial process is repeated several times until we obtain $K$ style generation modules $\{S_{1}, S_{2}, \dots, S_{K}\}$ and a trained representation learning module with good generalization ability. Finally, we use the representation learning model to predict test data labels. + +# 4. Experiments + +# 4.1. Datasets and Implementations + +To show the superiority of SAC for single domain generalization, we conduct experiments on four datasets: Digits [32], PACS [18], CIFAR10-C [13] and DomainNet [26]. Please see the supplementary material for more implementation details. + +# 4.2. Comparison Methods + +We compare SAC with 13 competitive methods, including ERM [15], ADA [32], M-ADA [27], ME-ADA [44], PDEN [20], L2D [34], NCDG [31], SADA [42], Pro-RandConv [4], UDP [10], AdvST [45], PSDG [40], and AdvFrequency [12]. Note that all baselines are compared using the same structures and * indicates our implementation. + +Table 2. Accuracy (\%) of different methods on PACS. Each column title indicates the source domain. + +
MethodsPhotoArtCartoonSketchAvg.
ERM [15]42.070.976.553.160.7
ADA [32]NeurIPS'1841.768.869.335.153.7
M-ADA [27]CVPR'2043.168.071.933.954.2
ME-ADA [44]NeurIPS'2042.169.270.636.854.7
PDEN [20]CVPR'2158.377.974.657.967.2
L2D [34]ICCV'2152.376.977.953.765.2
NCDG [31]TPAMI'2349.076.676.453.163.8
SADA [42]AAAI'2351.277.876.957.866.2
Pro-RandConv [4]CVPR'2362.977.078.557.168.9
UDP [10]MM'2356.777.177.657.567.2
AdvST [45]AAAI'2464.177.976.057.168.8
PSDG [40]KDD'2462.178.278.363.070.4
AdvFrequency [12]NeurIPS'2458.475.778.658.967.9
SAC61.779.279.264.671.2
+ +# 4.3. Experimental Results + +Comparisons on Digits. We evaluate the effectiveness of SAC on the Digits dataset, as shown in Tab. 1. Compared to the state-of-the-art baseline PSDG, SAC achieves an average accuracy of $88.3\%$ , representing a $3.6\%$ improvement. This result demonstrates that SAC generates augmented data with more diverse styles and captures more domain invariant representations than previous approaches. Notably, SAC achieves the highest accuracy across all tasks, underscoring the effectiveness of our proposed PAA and ARC in synthesizing complex styles, thereby facilitating the model generalizing to the unknown domains. These remarked results indicate that SAC captures more generalizable information than previous approaches, further highlighting the advantage of both improving color and texture diversity. + +Comparisons on PACS. Tab. 2 shows the experimental results on PACS dataset, demonstrating the superior performance of SAC compared to competing methods. We first observe that SAC achieves a higher average accuracy than the previous state-of-the-art method PSDG. Specifically, SAC outperforms PSDG by an average of $0.8\%$ , which clearly demonstrates its effectiveness. Notably, SAC achieves impressive accuracy on the most challenging Sketch task, which uses simplistic images without backgrounds as the training data. The impressive accuracy in this task highlights the superiority of SAC. In addition, although SAC performs a $2.4\%$ lower accuracy in the Photo task compared to AdvST, it achieves an overall average improvement of $2.4\%$ , further demonstrating its enhanced generalizability relative to AdvST. + +Comparisons on CIFAR10-C. As shown in Tab. 3, SAC achieves state-of-the-art (SOTA) performance, significantly outperforming competitors in accuracy. SAC demonstrates notable improvements, achieving a high average accuracy of $81.7\%$ , which is a $1.6\%$ increase over the previous SOTA method, UDP. Furthermore, compared to UDP, our method achieves substantial gains, with improvements of $3.9\%$ for + +Table 3. Accuracy (\%) of different methods on CIFAR10-C under different corruption categories at the level 5. Each column title indicates the target domain. + +
MethodsWeatherBlurNoiseDigitAvg.
ERM [15]67.356.730.062.354.1
ADA [32]NeurIPS'1872.767.040.066.661.6
M-ADA [27]CVPR'2075.536.754.265.164.7
ME-ADA [44]NeurIPS'2074.471.466.570.870.8
PDEN [20]CVPR'2178.676.778.673.176.8
L2D [34]ICCV'2176.070.273.372.072.9
NCDG [31]TPAMI'2382.282.870.976.278.0
SADA [42]AAAI'2379.182.471.478.477.8
Pro-RandConv* [4]CVPR'2377.878.676.880.778.5
UDP [10]MM'2380.783.679.976.180.1
AdvST [45]AAAI'2481.578.476.579.479.0
PSDG [40]KDD'2482.881.779.176.179.9
AdvFrequency [12]NeurIPS'2480.481.773.777.578.3
SAC80.182.683.880.481.7
+ +Table 4. Accuracy (\%) of different methods on DomainNet. Each column title indicates the target domain. + +
MethodsPaintingInfographClipartSketchQuickdrawAvg.
ERM* [15]38.314.035.322.92.122.5
ME-ADA [44]NeurIPS/2039.014.041.025.34.324.7
PDEN* [20]CVPR/2139.013.138.929.74.725.1
L2D [34]ICCV/2138.412.038.730.16.525.2
Pro-RandConv* [4]CVPR/2342.314.235.433.25.927.4
UDP* [10]MM/2338.613.039.130.14.725.1
AdvST [45]AAAI/2442.314.841.530.85.927.1
PSDG* [40]KDD/2439.013.539.530.54.325.4
AdvFrequency [12]NeurIPS/2442.815.142.731.26.927.7
SAC42.513.943.334.97.128.4
+ +Noise and $4.3\%$ for Digit. These impressive results highlight the advantages of SAC. + +Comparisons on DomainNet. To further evaluate the superiority of SAC, we compare it with the recent methods on the large-scale DomainNet benchmark. The results are presented in Tab. 4. We can observe that SAC achieves the highest average accuracy among all baselines. Additionally, the methods that solely introduce Gaussian noise to improve color diversity, such as UDP and PSDG, exhibit relatively poor performance. We conjecture that generating images with simplistic textures fails to simulate the complexity of real-world images. While Pro-RandConv achieves competitive accuracy by enhancing texture diversity, SAC still outperforms it in terms of average accuracy. These results validate the effectiveness of SAC, which simultaneously enhances both color and texture complexity. + +# 4.4. Analysis Experiment + +Ablation Study. We provide the ablation study on theDigits dataset to evaluate the effectiveness of key components in SAC, as shown in Tab. 5. Specifically, we introduce some variants to analyze the contributions of SAC, ARC, the Gaussian filter $\mathbf{g}$ , $\mathcal{L}_d$ , and $\mathcal{L}_p$ . When we remove a component, all others remain default settings. We first delete ARC (SAC W/O ARC) to assess the effectiveness of preserving object integrity. We observe that the performance + +Table 5. Ablation study on Digits. Each column title indicates the target domain. + +
MethodsSVHNMNIST-MSYNUSPSAvg.
SAC W/O ARC64.179.175.192.777.7
SAC W/O PAA74.985.386.096.185.3
SAC W/O g77.688.886.594.286.8
SAC W/O Ld75.590.387.096.787.4
SAC W/O Lp74.989.786.796.787.0
SAC W/O Ada76.390.585.595.486.9
SAC W/O Ada max73.989.885.495.386.1
SAC77.791.087.996.488.3
+ +drops seriously, which indicates the harm of illogical textures. Next, to investigate the influence of PAA, we only employ ARC (SAC W/O PAA) to generate images. The lower accuracy compared to SAC suggests that focusing solely on texture diversity is insufficient to enhance overall style diversity. We then remove the Gaussian filter $\mathbf{g}$ (SAC W/O $\mathbf{g}$ ) to evaluate its effectiveness in preserving object integrity. The degraded performance demonstrates that $\mathbf{g}$ can help maintain object integrity and prevent texture inconsistencies. Moreover, we see that eliminating $\mathcal{L}_d$ (SAC W/O $\mathcal{L}_d$ ) or $\mathcal{L}_p$ (SAC W/O $\mathcal{L}_p$ ) would lead to inferior performance compared to SAC, highlighting the importance of expanding the discrepancy at both the feature and pixel levels. Finally, we remove Eq. (9) (SAC W/O Ada) and follow the settings in [4] by iterating the random convolutions $T$ times, where $T$ is sampled from the uniform distribution $U$ [1, 10]. The inferior accuracy indicates that our proposed adaptive strategy effectively improves texture diversity while preventing semantic distortion. To further demonstrate its effectiveness in preserving semantic information, we conduct an experiment in which $T$ is fixed at 10 (SAC W/O $\mathrm{Ada}_{\mathrm{max}}$ ). The approximately $2\%$ performance drop confirms the importance of Eq. (9). + +Complementarity with Other Approaches. SAC can serve as a plug-and-play component to enhance the generalizability of other methods. Specifically, we apply PAA and ARC to the approaches of PDEN, L2D, RandConv, ProRandConv, and PSDG, respectively. The experimental results on Digits are shown in Tab. 6. These results indicate that ARC can serve as a general enhancer, improving the generalizability of other methods. Notably, when we apply ARC to PDEN, we find that "PDEN+ARC" achieves an $8.1\%$ improvement of average accuracy under default settings compared to PDEN. + +Comparative Analysis of Different Convolutions. In Tab. 7, we evaluate the effectiveness of various convolution types on the Digits dataset, including conventional convolution [17] (SAC-Conv), depthwise separable convolution [5] (SAC-DSCov), wavelet convolution [8] (SAC-WTConv), and deformable convolution [6] (SAC). SAC and its variants achieve outstanding performance, highlighting the ef + +Table 6. Accuracy (\%) on Digits when SAC is applied to recent methods as a plug-and-play component. Each column title indicates the target domain. + +
MethodsSVHNMNIST-MSYNUSPSAvg.
PDEN [20]62.282.269.485.374.8
PDEN+SAC73.886.781.590.183.0
L2D [34]62.987.363.784.074.5
L2D+SAC59.882.270.693.176.4
PSDG [40]73.186.783.495.784.7
PSDG+SAC75.890.885.096.387.0
RandConv [36]62.187.963.984.474.6
RandConv+ARC68.286.871.085.878.0
Pro-RandConv [4]69.782.379.893.781.4
Pro-RandConv+ARC70.585.982.795.383.6
+ +Table 7. Performance comparison of different convolution on Digits. Each column title indicates the target domain. + +
MethodsSVHNMNIST-MSYNUSPSAvg.
SAC-Conv [36]78.387.387.492.586.4
SAC-DSConv [5]78.789.187.492.486.9
SAC-WTConv [8]74.387.883.895.985.5
SAC77.791.087.996.488.3
+ +fectiveness of our proposed "Split-And-Combine" strategy. Specifically, the conventional convolution and depthwise separable convolution remain constrained to uniform grids, limiting their ability to accurately capture object contours and preserve object integrity. While the wavelet convolution expands the receptive field through wavelet transformation, it still extracts features within a uniform grid, making it inherently constrained by grid boundaries. In contrast, the deformable convolution overcomes these limitations by allowing sampling points to deviate from a fixed uniform grid, thereby capturing object shapes and structural details more effectively and contributing to texture smoothing. + +# 5. Conclusion + +We present a novel "Split-and-Combine" method for single domain generalization. Instead of applying style augmentation to entire images, our method first performs patch-aware augmentation, which conducts separate style generation to each patch of an image, enriching colors. Then, we introduce adaptive random convolutions to enhance texture complexity while preserving object integrity. Through these synergistic components, our model can synthesize augmented data with different local colors and textures. Additionally, we design an energy-based distribution discrepancy score to assess distribution differences between the original and augmented data, which can encourage SAC to generate novel data with large distribution shifts. SAC consistently achieves state-of-the-art performance across four datasets, demonstrating the effectiveness of our method. + +# Acknowledgments + +This work was supported in part by the National Natural Science Foundation of China under Grants 62306008, 32472007, 62406095, and 62402004, the Key Project of Anhui Province's Science and Technology Innovation Tackle Plan under Grant 202423k09020040, the China Postdoctoral Science Foundation under Grant 2024M762632, the Natural Science Foundation of Shaanxi under Grant 2024JC-YBQN-0687, the Social Science Foundation of Shaanxi under Grant 2024R027, and the Hefei Key Technology R&D "Champion-Based Selection" Project under Grant 2023SGJ011. + +# References + +[1] Kartik Ahuja, Karthikeyan Shanmugam, Kush R. Varshney, and Amit Dhurandhar. Invariant risk minimization games. In International Conference on Machine Learning, Virtual Event, July 13-18, 2020. 1 +[2] Kartik Ahuja, Ethan Caballero, Dinghuai Zhang, Jean-Christophe Gagnon-Audet, Yoshua Bengio, Ioannis Mitliagkas, and Irina Rish. Invariance principle meets information bottleneck for out-of-distribution generalization. In Neural Information Processing Systems, virtual, December 6-14, 2021. 1 +[3] Tianle Chen, Mahsa Baktashmotlagh, Zijian Wang, and Mathieu Salzmann. Center-aware adversarial augmentation for single domain generalization. In Winter Conference on Applications of Computer Vision, Waikoloa, HI, USA, January 2-7, 2023. 1, 3 +[4] Seokeon Choi, Debasmit Das, Sungha Choi, Seunghan Yang, Hyunsin Park, and Sungrack Yun. Progressive random convolutions for single domain generalization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, Vancouver, BC, Canada, June 17-24, pages 10312-10322, 2023. 2, 3, 4, 5, 6, 7, 8 +[5] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Conference on Computer Vision and Pattern Recognition, Honolulu, HI, USA, July 21-26, 2017. 8 +[6] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In International Conference on Computer Vision, Venice, Italy, October 22-29, 2017. 8 +[7] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In International Conference on Computer Vision, Venice, Italy, October 22-29, 2017. 5 +[8] Shahaf E. Finder, Roy Amoyal, Eran Treister, and Oren Freifeld. Wavelet convolutions for large receptive fields. In European Conference on Computer Vision, MiCo Milano, September 29 - October 4, 2024. 8 +[9] Will Grathwohl, Kuan-Chieh Wang, Jorn-Henrik Jacobsen, David Duvenaud, Mohammad Norouzi, and Kevin Swersky. Your classifier is secretly an energy based model and you should treat it like one. In International Conference on + +Learning Representations, Addis Ababa, Ethiopia, April 26-30, 2020. 5 +[10] Kehua Guo, Rui Ding, Tian Qiu, Xiangyuan Zhu, Zheng Wu, Liwei Wang, and Hui Fang. Single domain generalization via unsupervised diversity probe. In ACM International Conference on Multimedia, Ottawa, ON, Canada, October 29-November 3, pages 2101-2111, 2023. 2, 3, 4, 6, 7 +[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Conference on Computer Vision and Pattern Recognition, Las Vegas, NV, USA, June 27-30, pages 770-778, 2016. 1 +[12] Xilin He, Jingyu Hu, Qinliang Lin, Cheng Luo, Weicheng Xie, Siyang Song, Muhammad Haris Khan, and Linlin Shen. Towards combating frequency simplicity-biased learning for domain generalization. In Conference on Neural Information Processing Systems, Vancouver, BC, Canada, December 10 - 15, 2024. 3, 6, 7 +[13] Dan Hendrycks and Thomas G. Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, New Orleans, LA, USA, May 6-9, 2019. 6 +[14] Xun Huang and Serge J. Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In International Conference on Computer Vision, Venice, Italy, October 22-29, pages 1510-1519, 2017. 3, 4 +[15] Vladimir Koltchinskii. Oracle inequalities in empirical risk minimization and sparse recovery problems: École D'Été de Probabilités de Saint-Flour XXXVIII-2008. Springer Science & Business Media, 2011. 6, 7 +[16] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In Conference on Neural Information Processing Systems, Lake Tahoe, Nevada, United States, December 3-6, 2012. 1 +[17] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 1998. 8 +[18] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M. Hospedales. Deeper, broader and artier domain generalization. In IEEE International Conference on Computer Vision, Venice, Italy, October 22-29, pages 5542-5550, 2017. 6 +[19] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M. Hospedales. Learning to generalize: Meta-learning for domain generalization. In Proceedings of Conference on Artificial Intelligence, New Orleans, Louisiana, USA, February 2-7, 2018. 1 +[20] Lei Li, Ke Gao, Juan Cao, Ziyao Huang, Yepeng Weng, Xiaoyue Mi, Zhengze Yu, Xiaoya Li, and Boyang Xia. Progressive domain expansion network for single domain generalization. In Conference on Computer Vision and Pattern Recognition, virtual, June 19-25, pages 224-233, 2021. 2, 3, 6, 7, 8 +[21] Weitang Liu, Xiaoyun Wang, John D. Owens, and Yixuan Li. Energy-based out-of-distribution detection. In Conference on Neural Information Processing Systems, virtual, December 6-12, 2020. 5 +[22] Krikamol Muandet, David Balduzzi, and Bernhard Scholkopf. Domain generalization via invariant feature + +representation. In Proceedings of International Conference on Machine Learning Atlanta, GA, USA, June 16-21, pages 10-18, 2013. 1 +[23] Cheng Ouyang, Chen Chen, Surui Li, Zeju Li, Chen Qin, Wenjia Bai, and Daniel Rueckert. Causality-inspired single-source domain generalization for medical image segmentation. IEEE Transactions on Medical Imaging, 2023. 3 +[24] Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. IEEE Transactions on Knowledge and Data Engineering, 22(10):1345-1359, 2010. 1 +[25] Sinno Jialin Pan, Ivor W. Tsang, James T. Kwok, and Qiang Yang. Domain adaptation via transfer component analysis. In International Joint Conference on Artificial Intelligence, Pasadena, California, USA, July 11-17, 2009. 1 +[26] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In International Conference on Computer Vision, Seoul, Korea (South), October 27 - November 2, pages 1406-1415. IEEE, 2019. 6 +[27] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In Conference on Computer Vision and Pattern Recognition, Seattle, WA, USA, June 13-19, pages 12553-12562, 2020. 2, 6, 7 +[28] Alexander Robey, George J. Pappas, and Hamed Hassani. Model-based domain generalization. In Conference on Neural Information Processing Systems, virtual, December 6-14, 2021. 1 +[29] Ding Rui, Kehua Guo, Xiangyuan Zhu, Zheng Wu, and Hui Fang and. Progressive diversity generation for single domain generalization. IEEE Transactions on Multimedia, 2024. 2, 3 +[30] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott E. Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Conference on Computer Vision and Pattern Recognition, Boston, MA, USA, June 7-12, 2015. 1 +[31] Chris Xing Tian, Haoliang Li, Xiaofei Xie, Yang Liu, and Shiqi Wang. Neuron coverage-guided domain generalization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):1302-1311, 2023. 1, 3, 6, 7 +[32] Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C. Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. In Conference on Neural Information Processing Systems, Montreal, Canada, December 3-8, pages 5339-5349, 2018. 1, 2, 6, 7 +[33] Chaoqun Wan, Xu Shen, Yonggang Zhang, Zhiheng Yin, Xinmei Tian, Feng Gao, Jianqiang Huang, and Xian-Sheng Hua. Meta convolutional neural networks for single domain generalization. In Conference on Computer Vision and Pattern Recognition, New Orleans, LA, USA, June 18-24, pages 4672-4681, 2022. 1 +[34] Zijian Wang, Yadan Luo, Ruihong Qiu, Zi Huang, and Mahsa Baktashmotlagh. Learning to diversify for single domain generalization. In International Conference on Computer Vision, Montreal, QC, Canada, October 10-17, pages 814-823, 2021. 3, 6, 7, 8 + +[35] Cort J Willmott and Kenji Matsuura. Advantages of the mean absolute error (mae) over the root mean square error (rmse) in assessing average model performance. Climate research, 2005. 5 +[36] Zhenlin Xu, Deyi Liu, Junlin Yang, Colin Raffel, and Marc Niethammer. Robust and generalizable visual representation learning via random convolutions. In International Conference on Learning Representations, virtual, Austria, May 3-7, 2021. 2, 4, 5, 8 +[37] Fu-En Yang, Yuan-Chia Cheng, Zu-Yun Shiau, and Yu-Chiang Frank Wang. Adversarial teacher-student representation learning for domain generalization. In Conference on Neural Information Processing Systems, December 6-14, virtual, 2021. 3 +[38] Shuai Yang, Kui Yu, Fuyuan Cao, Hao Wang, and Xindong Wu. Dual-representation-based autoencoder for domain adaptation. IEEE Transactions on Cybernetics, 52(8): 7464-7477, 2022. 1 +[39] Shuai Yang, Kui Yu, Fuyuan Cao, Lin Liu, Hao Wang, and Jiuyong Li. Learning causal representations for robust domain adaptation. IEEE Transactions on Knowledge and Data Engineering, 35(3):2750-2764, 2023. 1 +[40] Shuai Yang, Zhen Zhang, and Lichuan Gu. Practical single domain generalization via training-time and test-time learning. In ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Barcelona, Spain, August 25-29, 2024. 2, 3, 4, 6, 7, 8 +[41] Haotian Ye, Chuanlong Xie, Tianle Cai, Ruichen Li, Zhenguo Li, and Liwei Wang. Towards a theoretical framework of out-of-distribution generalization. In Neural Information Processing Systems, virtual, December 6-14, 2021. 1 +[42] Jiajin Zhang, Hanqing Chao, Amit Dhurandhar, Pin-Yu Chen, Ali Tajer, Yangyang Xu, and Pingkun Yan. When neural networks fail to generalize? A model sensitivity perspective. In AAAI Conference on Artificial Intelligence, Washington, DC, USA, February 7-14, 2023. 6, 7 +[43] Zuyu Zhang, Yan Li, and Byung-Seok Shin. Embracing domain gradient conflicts: Domain generalization using domain gradient equilibrium. In International Conference on Multimedia, Melbourne, VIC, Australia, October 28 - November 1, 2024. 1 +[44] Long Zhao, Ting Liu, Xi Peng, and Dimitris N. Metaxas. Maximum-entropy adversarial data augmentation for improved generalization and robustness. In Advances in Neural Information Processing Systems, virtual, December 6-12, pages 1-13, 2020. 3, 6, 7 +[45] Guangtao Zheng, Mengdi Huai, and Aidong Zhang. Advst: Revisiting data augmentations for single domain generalization. In Thirty-Eighth AAAI Conference on Artificial Intelligence, February 20-27, pages 1-9, 2024. 1, 3, 6, 7 +[46] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4396-4415, 2023. 1 +[47] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A comprehensive survey on transfer learning. Proceedings of the IEEE, 109(1):43-76, 2021. 1 \ No newline at end of file diff --git a/2025/Split-and-Combine_ Enhancing Style Augmentation for Single Domain Generalization/images.zip b/2025/Split-and-Combine_ Enhancing Style Augmentation for Single Domain Generalization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..13e8c4ac6ab0854221716c11b5245f80dd7db5dd --- /dev/null +++ b/2025/Split-and-Combine_ Enhancing Style Augmentation for Single Domain Generalization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:55c7b3f96c978e272a202799747bf2e13e44a89fa938b32d6fd2e51fd2cad575 +size 554214 diff --git a/2025/Split-and-Combine_ Enhancing Style Augmentation for Single Domain Generalization/layout.json b/2025/Split-and-Combine_ Enhancing Style Augmentation for Single Domain Generalization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b659fd8673f43b9fd7de191f03a09f4734a6a1e3 --- /dev/null +++ b/2025/Split-and-Combine_ Enhancing Style Augmentation for Single Domain Generalization/layout.json @@ -0,0 +1,9575 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 87, + 102, + 523, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 102, + 523, + 138 + ], + "spans": [ + { + "bbox": [ + 87, + 102, + 523, + 138 + ], + "type": "text", + "content": "Split-and-Combine: Enhancing Style Augmentation for Single Domain Generalization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "spans": [ + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "text", + "content": "Zhen Zhang" + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "text", + "content": ", Shuai Yang" + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "text", + "content": ", Qianlong Dang" + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "text", + "content": ", Zhize Wu" + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "text", + "content": ", Lichuan Gu" + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "text", + "content": "Anhui Agricultural University " + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "text", + "content": "Northwest A & F University " + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "text", + "content": "Hefei University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 117, + 191, + 490, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 191, + 490, + 203 + ], + "spans": [ + { + "bbox": [ + 117, + 191, + 490, + 203 + ], + "type": "text", + "content": "zhangz@stu.ahau.edu.cn, yangs@ahau.edu.cn, dangqianlong@nwafu.edu.cn," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 212, + 205, + 394, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 205, + 394, + 217 + ], + "spans": [ + { + "bbox": [ + 212, + 205, + 394, + 217 + ], + "type": "text", + "content": "wuzz@hfuu.edu.cn, glc@ahau.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 270, + 297, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 297, + 594 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 297, + 594 + ], + "type": "text", + "content": "Single domain generalization aims to learn a model with good generalization ability from a single source domain. Recent advances in this field have focused on increasing the diversity of the training data through style (e.g., color and texture) augmentation. However, most existing methods apply uniform perturbations to the entire image, failing to simulate complex images with multiple distinct stylistic regions. To address this, we propose a \"Split-And-Combine\" (SAC) strategy to enhance style diversity. Specifically, SAC first performs patch-aware augmentation, which splits an image into multiple patches and applies style augmentation independently to each patch, enabling distinct color variations across regions. Then, SAC combines these patches to reconstruct a complete image and applies adaptive random convolutions, which utilizes a deformable convolution layer with random and Gaussian filters to enhance texture diversity while preserving object integrity. Notably, SAC leverages entropy as a risk assessment criterion to adaptively determine whether a sample should undergo augmentation within the iterative process of random convolutions, preventing excessive augmentation. Furthermore, SAC introduces an energy-based distribution discrepancy score to quantify out-of-distribution likelihood, systematically expanding the augmented data's distribution. SAC can serve as a plug-and-play component to improve the performance of recent methods. Extensive experiments on four datasets demonstrate the effectiveness of SAC." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 616, + 135, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 135, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 135, + 628 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 636, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 296, + 696 + ], + "type": "text", + "content": "Modern machine learning algorithms [11, 16, 30] have achieved remarkable performance under the assumption that the source (training) and target (test) domains share the same distribution. However, in reality, such an assumption is often violated due to factors such as variations in" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 246, + 555, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 555, + 305 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 555, + 305 + ], + "type": "text", + "content": "illumination, background, or weather, which are known as the domain shift problem [41]. Unfortunately, these traditional algorithms often face significant performance degradation when there are distributional discrepancies between the source and target domains [24, 25, 38, 39, 47]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 306, + 556, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 556, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 556, + 474 + ], + "type": "text", + "content": "To tackle this problem, multi-source domain generalization [19, 22, 28, 46] has been presented, aiming to learn a robust model by using knowledge from multiple different but related source domains without accessing target domain data. The mainstream works depend heavily on domain identifiers as supervision signals to capture domain invariant representations. However, the collection and annotation of data from multiple domains can be prohibitively expensive, making multi-source domain generalization less feasible in practical applications. To alleviate this issue, a more challenging and realistic setting known as single domain generalization [31, 33, 43] has emerged and garnered significant attention recently, which only leverages a single source domain to learn a model with good generalizability." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 475, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 556, + 714 + ], + "type": "text", + "content": "The difficulty in single domain generalization arises from mitigating distributional discrepancies between the source and unknown target domains. In reality, it is observed that content features, such as object features, remain relatively stable despite environmental changes, whereas style features, such as backgrounds, exhibit considerable variability across different domains [1, 2]. For example, in animal image classification, fundamental elements like eyes, ears, and face shape constitute the content, whereas style elements include background and lighting conditions. Therefore, the key to effective single domain generalization is to eliminate spurious correlations between style features and labels. The current mainstream research on single domain generalization primarily focuses on increasing the capacity of the training data through the generation of novel domains, thereby reducing the model's reliance on style features due to the increased diversity of styles within the training data. One category of methods [3, 32, 45] perturb the original data using the back-propagated gradient to obtain augmented data. However, gradient-based perturba" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 145, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 145, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 145, + 713 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15616" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 136, + 121 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 136, + 121 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 136, + 121 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 136, + 121 + ], + "type": "image", + "image_path": "13392f7c48309c7ff72b485abe5e59cec212dfcda5d5150947dbb820a1f30104.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 121, + 117, + 129 + ], + "lines": [ + { + "bbox": [ + 88, + 121, + 117, + 129 + ], + "spans": [ + { + "bbox": [ + 88, + 121, + 117, + 129 + ], + "type": "text", + "content": "(a) UDP" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 141, + 70, + 210, + 121 + ], + "blocks": [ + { + "bbox": [ + 141, + 70, + 210, + 121 + ], + "lines": [ + { + "bbox": [ + 141, + 70, + 210, + 121 + ], + "spans": [ + { + "bbox": [ + 141, + 70, + 210, + 121 + ], + "type": "image", + "image_path": "9066d5d6e3caf5b114b20fdbb3ab6b0fae3a724d0ffde415e917ab5933ee559b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 158, + 121, + 193, + 129 + ], + "lines": [ + { + "bbox": [ + 158, + 121, + 193, + 129 + ], + "spans": [ + { + "bbox": [ + 158, + 121, + 193, + 129 + ], + "type": "text", + "content": "(b) PSDG" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 215, + 70, + 283, + 121 + ], + "blocks": [ + { + "bbox": [ + 215, + 70, + 283, + 121 + ], + "lines": [ + { + "bbox": [ + 215, + 70, + 283, + 121 + ], + "spans": [ + { + "bbox": [ + 215, + 70, + 283, + 121 + ], + "type": "image", + "image_path": "e032456b63610bdfffbf40b8b7a7a0746f90c1912bad734b944bb3f4bef2f114.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 121, + 271, + 129 + ], + "lines": [ + { + "bbox": [ + 223, + 121, + 271, + 129 + ], + "spans": [ + { + "bbox": [ + 223, + 121, + 271, + 129 + ], + "type": "text", + "content": "(c) RandConv" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 69, + 129, + 136, + 179 + ], + "blocks": [ + { + "bbox": [ + 69, + 129, + 136, + 179 + ], + "lines": [ + { + "bbox": [ + 69, + 129, + 136, + 179 + ], + "spans": [ + { + "bbox": [ + 69, + 129, + 136, + 179 + ], + "type": "image", + "image_path": "41423f839c1cfd1aa6654a908c3e8755979cecea770708fd958ed47358e053b4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 180, + 134, + 189 + ], + "lines": [ + { + "bbox": [ + 71, + 180, + 134, + 189 + ], + "spans": [ + { + "bbox": [ + 71, + 180, + 134, + 189 + ], + "type": "text", + "content": "(d) Pro-RandConv" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 142, + 129, + 210, + 180 + ], + "blocks": [ + { + "bbox": [ + 142, + 129, + 210, + 180 + ], + "lines": [ + { + "bbox": [ + 142, + 129, + 210, + 180 + ], + "spans": [ + { + "bbox": [ + 142, + 129, + 210, + 180 + ], + "type": "image", + "image_path": "9cc13c06a9c6ce5afd3758a3db6698f48c03d8fc5990d850f5ddab83772e9882.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 181, + 191, + 190 + ], + "lines": [ + { + "bbox": [ + 160, + 181, + 191, + 190 + ], + "spans": [ + { + "bbox": [ + 160, + 181, + 191, + 190 + ], + "type": "text", + "content": "(e) Patch" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 198, + 288, + 210 + ], + "lines": [ + { + "bbox": [ + 61, + 198, + 288, + 210 + ], + "spans": [ + { + "bbox": [ + 61, + 198, + 288, + 210 + ], + "type": "text", + "content": "Figure 1. Examples of images generated by different methods." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 215, + 129, + 283, + 180 + ], + "blocks": [ + { + "bbox": [ + 215, + 129, + 283, + 180 + ], + "lines": [ + { + "bbox": [ + 215, + 129, + 283, + 180 + ], + "spans": [ + { + "bbox": [ + 215, + 129, + 283, + 180 + ], + "type": "image", + "image_path": "fe05c9a4da7a52b5bc8b5282f6f495a9ffaf44f35fc3c8815b7c026c2960b8ad.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 181, + 263, + 190 + ], + "lines": [ + { + "bbox": [ + 235, + 181, + 263, + 190 + ], + "spans": [ + { + "bbox": [ + 235, + 181, + 263, + 190 + ], + "type": "text", + "content": "(f) SAC" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 233, + 295, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 233, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 233, + 295, + 293 + ], + "type": "text", + "content": "tions are visually imperceptible, rendering these methods inadequate for simulating real-world domain shifts. Another mainstream category trains a generative model to synthesize novel data with style (e.g., color and texture) shifts [4, 10, 20], resulting in visually distinct samples." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 294, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 295, + 593 + ], + "type": "text", + "content": "Recent methods leveraging generative models for data augmentation have achieved remarkable success by enhancing color and texture diversity. For instance, several approaches [10, 29, 40] introduce Gaussian noise to perturb images, effectively enriching the color distribution within a single generated image. Alternatively, other methods [4, 36] utilize a convolution layer with random filters to generate images with diverse local textures. However, these methods enhance style diversity from only a single aspect, either color or texture, leading to inherent limitations in the other aspect. For example, Gaussian noise-based methods, such as UDP [10] and PSDG [40], only generate simplistic textures (see Fig. 1 (a) and (b)), whereas methods leveraging random convolutions, such as RandConv [36] and ProRandConv [4], tend to produce images dominated by a single color (see Fig. 1 (c) and (d)). In contrast, real-world scenarios typically involve a wide range of colors and complex textures, with different regions of an image (e.g., a person's clothing, grass, trees, and the sky within a park), each exhibiting unique texture and color characteristics. Therefore, improving style diversity solely from a single perspective cannot simulate images from complex real-world scenarios. This raises a crucial challenge: Can we combine the advantages of Gaussian noise and random convolutions to generate images with both rich colors and complex textures?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "In this paper, we propose a \"Split-And-Combine\" (SAC) strategy to improve style diversity, which first splits images into a series of patches to enhance the color diversity, then combines the augmented patches into the complete image and uses adaptive random convolutions to improve the texture diversity. In the split step, we propose patch-aware augmentation (PAA) that introduces rich color variations. Specifically, we divide an image into several patches and each patch undergoes independent style generation within Gaussian noise, yielding generated images" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 72, + 553, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 370 + ], + "type": "text", + "content": "with different local colors. While PAA effectively generates diverse styles for each patch, it leads to texture inconsistencies in a single object, resulting in the model misinterpreting the object as a collection of fragmented parts rather than a unified entity (see Fig. 1 (e)). For instance, a tree trunk may exhibit a smooth texture on one side and a rough, uneven texture on the other, leading the model to perceive it as two distinct components instead of a coherent whole. To address this challenge, we propose adaptive random convolutions (ARC) to enhance texture complexity and restore object integrity in the combine step. Specifically, we first initialize a deformable convolution layer with both random filters and Gaussian filters. Subsequently, we iteratively apply this layer and affine transformation to generate random textures, progressively increasing texture complexity. Moreover, we use entropy values to adaptively control this progressive process, preventing excessive application of random convolutions that could distort semantic information. Furthermore, we use an energy-based distribution discrepancy score to amplify the distribution differences between the original and generated data, enabling the synthesis of out-of-distribution data. We conduct experiments on four well-established benchmarks, where SAC consistently achieves state-of-the-art performance in various scenarios. Overall, our contributions can be summarized as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 373, + 553, + 564 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 314, + 373, + 553, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 373, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 314, + 373, + 553, + 408 + ], + "type": "text", + "content": "- We propose a \"Split-And-Combine\" strategy that first splits images into patches for color augmentation and then combines the patches to undergo texture augmentation." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 409, + 553, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 409, + 553, + 444 + ], + "spans": [ + { + "bbox": [ + 314, + 409, + 553, + 444 + ], + "type": "text", + "content": "- We design the patch-aware augmentation that divides an image into a series of patches for separate style transformation to promote color diversity for generated images." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 445, + 553, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 445, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 314, + 445, + 553, + 491 + ], + "type": "text", + "content": "- We develop the adaptive random convolution that utilizes entropy values to adaptively control the progressive process of the random deformable convolution to improve the texture diversity." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 492, + 553, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 492, + 553, + 528 + ], + "spans": [ + { + "bbox": [ + 314, + 492, + 553, + 528 + ], + "type": "text", + "content": "- We present an energy-based perspective to maximize the distribution discrepancy between the original and augmented data, generating out-of-distribution data." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 529, + 553, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 529, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 314, + 529, + 553, + 564 + ], + "type": "text", + "content": "- We conduct experiments on four established datasets. The results demonstrate that the proposed SAC achieves consistent state-of-the-art performance." + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 574, + 400, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 574, + 400, + 586 + ], + "spans": [ + { + "bbox": [ + 314, + 574, + 400, + 586 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 594, + 553, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 630 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 630 + ], + "type": "text", + "content": "Recently, since only requiring a single training domain, single domain generalization has attracted substantial interest, which can be roughly categorized into two paradigms." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": "Gradient-based augmentation methods. This paradigm of work attempts to generate data using backpropagated gradients. One approach involves directly adding gradients to the original data, thereby synthesizing augmented data. For example, ADA [32] employs backpropagated gradients to synthesize data. Subsequently, M-ADA [27] enhances ADA with a Wasserstein Auto" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15617" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "content": "Encoder to increase the distribution differences between the augmented and original data. Similarly, ME-ADA [44] introduces a maximum-entropy regularizer to amplify distribution shifts in augmented data by minimizing mutual information between augmented and original data. Later on, NCDG [31] presents neuron coverage to improve generalization performance by maximizing gradient similarity between the original and augmented data. CAA [3] relies on angular distance to push generated data away from class centroids. AdvST [45] uses back-propagated gradients as guidance to fully exploit the effectiveness of several standard data augmentation methods. However, these methods fail in target domains with large domain shifts due to visually imperceptible style variations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 243, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 296, + 639 + ], + "type": "text", + "content": "Style-based augmentation methods. This paradigm of work trains a generative model to produce data with different styles (e.g., contrast, colors, and textures). The most well-established method is PDEN [20], which progressively trains multiple generative models to synthesize different augmented domains and uses cycle consistency loss to preserve semantic information. Additionally, L2D [34] generates data and captures invariant features through a max-min game of mutual information. ATS [37] introduces a teacher-student network to learn domain-generalized knowledge by minimizing the discrepancy between the normalized features of the original and augmented data. CSDG [23] simultaneously changes the textures and intensities of the original data to generate new data. However, due to the limited styles provided by AdaIN [14], these methods only synthesize data with simple colors and textures, which limits their applicability in real-world scenarios. To enrich the styles, Pro-RandConv [4] introduces a deformable convolution block to improve diversification by applying it along with affine transformation to generate data with new styles. UDP [10] adds extra Gaussian noise to perturb the data for more complex styles and enhances the shifts in generated data by maximizing the uncertainty between the generated and original data from an information theory perspective. Subsequently, PDG [29] enhances UDP by modifying the structure of the generative model, using " + }, + { + "bbox": [ + 56, + 243, + 296, + 639 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 56, + 243, + 296, + 639 + ], + "type": "text", + "content": " convolutions to preserve the input shape while employing random and larger kernel sizes to alter the local texture. PSDG [40] was recently proposed to improve generalizability by optimizing the model during both training and testing. AdvFrequency [12] improves frequency discrepancies between the generated and original data through the use of adversarial learning, thereby reducing the risk of the particular frequency." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Previous methods typically enhance style diversity from a single perspective, focusing on either color or texture, and fail to simulate the complexity of real-world images. In contrast, SAC improves both color and texture diversity simultaneously through PAA and ARC, thereby enhancing the model's ability to capture invariant representations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 314, + 71, + 370, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 370, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 370, + 84 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 91, + 553, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 150 + ], + "type": "text", + "content": "Problem Setup. Consider single domain generalization defined on source domain " + }, + { + "bbox": [ + 313, + 91, + 553, + 150 + ], + "type": "inline_equation", + "content": "D_{s} = \\{\\mathbf{x}_{i},\\mathbf{y}_{i}\\}_{i = 1}^{n}" + }, + { + "bbox": [ + 313, + 91, + 553, + 150 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 91, + 553, + 150 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 313, + 91, + 553, + 150 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 91, + 553, + 150 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 91, + 553, + 150 + ], + "type": "text", + "content": "-th sample, " + }, + { + "bbox": [ + 313, + 91, + 553, + 150 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i" + }, + { + "bbox": [ + 313, + 91, + 553, + 150 + ], + "type": "text", + "content": " is the corresponding label of " + }, + { + "bbox": [ + 313, + 91, + 553, + 150 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 313, + 91, + 553, + 150 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 91, + 553, + 150 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 91, + 553, + 150 + ], + "type": "text", + "content": " is the number of samples. The task is to train a model with good generalizability in the unknown target domain " + }, + { + "bbox": [ + 313, + 91, + 553, + 150 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 313, + 91, + 553, + 150 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 151, + 555, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 151, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 151, + 555, + 342 + ], + "type": "text", + "content": "Overall Framework. We propose the \"Split-And-Combine\" strategy to synthesize images with rich colors and complex textures for single domain generalization tasks. SAC comprises two key components: a style generation module, which synthesizes data with novel styles, and a representation learning module, which learns domain invariant features. The style generation module consists of an Encoder, a generative model " + }, + { + "bbox": [ + 313, + 151, + 555, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\cdot)" + }, + { + "bbox": [ + 313, + 151, + 555, + 342 + ], + "type": "text", + "content": ", and a Decoder. To enhance color and texture diversity, we introduce patch-aware augmentation and adaptive random convolutions, enabling the generation of more complex styles. Additionally, we design a score to expand domain distribution through energy discrepancies. The representation learning module, which consists of a feature extractor " + }, + { + "bbox": [ + 313, + 151, + 555, + 342 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 151, + 555, + 342 + ], + "type": "text", + "content": " and a classifier " + }, + { + "bbox": [ + 313, + 151, + 555, + 342 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 151, + 555, + 342 + ], + "type": "text", + "content": ", combines both original and augmented data to learn invariant representations. The framework of SAC is shown in Fig. 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 349, + 455, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 349, + 455, + 361 + ], + "spans": [ + { + "bbox": [ + 314, + 349, + 455, + 361 + ], + "type": "text", + "content": "3.1. Style Generation Module" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 367, + 554, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 554, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 554, + 462 + ], + "type": "text", + "content": "As we proposed before, the key to solving single domain generalization is to synthesize data with rich colors and complex textures. However, the augmented data generated by previous methods contains simple textures or colors, failing to simulate images in real-world scenarios. Unlike previous methods, we independently perform style augmentation on each patch to enrich colors and use adaptive random deformable convolutions to enhance texture complexity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 463, + 554, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 554, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 554, + 522 + ], + "type": "text", + "content": "Patch-Aware Augmentation. To improve the color diversity, we propose patch-aware augmentation to ensure that the images have distinct local colors. Given an image " + }, + { + "bbox": [ + 313, + 463, + 554, + 522 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 313, + 463, + 554, + 522 + ], + "type": "text", + "content": ", we first use an encoder to project the image into the latent space to obtain rich feature representations:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 397, + 531, + 553, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 531, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 397, + 531, + 553, + 544 + ], + "type": "interline_equation", + "content": "\\mathbf {r} = \\operatorname {E n c o d e r} (\\mathbf {x}), \\tag {1}", + "image_path": "df14513e744eda1df957c70e65cd0c8b481e3ba70c32b6e6bd257ee19d97f1a8.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 552, + 554, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 554, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 554, + 611 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 552, + 554, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 313, + 552, + 554, + 611 + ], + "type": "text", + "content": " is from the encoder, which consists of learnable deep layers instantiated as two convolutional layers and each with a ReLU activation function behind. Afterward, we divide " + }, + { + "bbox": [ + 313, + 552, + 554, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 313, + 552, + 554, + 611 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 313, + 552, + 554, + 611 + ], + "type": "inline_equation", + "content": "N\\times N" + }, + { + "bbox": [ + 313, + 552, + 554, + 611 + ], + "type": "text", + "content": " patches. Each patch " + }, + { + "bbox": [ + 313, + 552, + 554, + 611 + ], + "type": "inline_equation", + "content": "p_{i,j}\\in \\mathbb{R}^{\\frac{H}{N}\\times \\frac{W}{N}}" + }, + { + "bbox": [ + 313, + 552, + 554, + 611 + ], + "type": "text", + "content": " can be written as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 324, + 617, + 553, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 617, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 324, + 617, + 553, + 666 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} p _ {i, j} = \\mathbf {r} [ h _ {i}: h _ {i + 1}, w _ {j}: w _ {j + 1} ], \\quad i, j = 0, \\dots , N - 1 \\\\ h _ {i} = \\frac {H}{N} \\times i, \\quad w _ {j} = \\frac {W}{N} \\times j, \\tag {2} \\\\ \\end{array}", + "image_path": "0647fd467847d16b85a9824844721099e29a89fd6b3859b0c5a2815a72e750c4.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "inline_equation", + "content": "h_i, w_j" + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": " indicate the locations in the image. Next, following the convention in generative models [20, 40], we employ two fully connected layers " + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "inline_equation", + "content": "f_1(\\cdot)" + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "inline_equation", + "content": "f_2(\\cdot)" + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": " to learn a randomly initialized Gaussian noise, which can simulate an" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15618" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 120, + 68, + 492, + 297 + ], + "blocks": [ + { + "bbox": [ + 120, + 68, + 492, + 297 + ], + "lines": [ + { + "bbox": [ + 120, + 68, + 492, + 297 + ], + "spans": [ + { + "bbox": [ + 120, + 68, + 492, + 297 + ], + "type": "image", + "image_path": "bc0ca0cbca69ea061010a9fdf17d765b80997e9a5d9e5763e21d246d27a5da2f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 308, + 555, + 331 + ], + "lines": [ + { + "bbox": [ + 55, + 308, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 555, + 331 + ], + "type": "text", + "content": "Figure 2. Framework of SAC, where \"PAA\" is patch-aware augmentation, \"ARC\" is adaptive random convolutions, \"En\" is the Encoder, \"De\" is the Decoder, " + }, + { + "bbox": [ + 55, + 308, + 555, + 331 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 308, + 555, + 331 + ], + "type": "text", + "content": " is the feature extractor, and " + }, + { + "bbox": [ + 55, + 308, + 555, + 331 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 308, + 555, + 331 + ], + "type": "text", + "content": " is the classifier." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 351, + 296, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 296, + 399 + ], + "type": "text", + "content": "unknown style. Notably, as styles generated in this process are simple, we go beyond previous methods and present to separately apply style augmentation to each patch rather than the entire image:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 404, + 252, + 419 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 404, + 252, + 419 + ], + "spans": [ + { + "bbox": [ + 96, + 404, + 252, + 419 + ], + "type": "interline_equation", + "content": "\\tilde {p} _ {i, j} = \\mathcal {T} (p _ {i, j}), \\quad i, j = 0, \\ldots , N - 1", + "image_path": "07008c2f5a555a3ba3c5be1cfb5b1dc7e2146e16fcdf5f79370f76ce2b6e36da.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 421, + 294, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 294, + 456 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 294, + 456 + ], + "type": "interline_equation", + "content": "\\mathcal {T} \\left(p _ {i, j}\\right) = \\left(1 + f _ {1} \\left(\\varepsilon_ {i, j}\\right)\\right) \\left(\\frac {p _ {i , j} - \\mu \\left(p _ {i , j}\\right)}{\\sigma \\left(p _ {i , j}\\right)} + \\eta_ {i, j}\\right) + f _ {2} \\left(\\varepsilon_ {i, j}\\right), \\tag {3}", + "image_path": "115a1d52e206e2905879ee5b76434f669e6b41a8af4e29530e03d5ee71a9bdbd.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 456, + 296, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 296, + 553 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 456, + 296, + 553 + ], + "type": "inline_equation", + "content": "\\mu (\\cdot)" + }, + { + "bbox": [ + 55, + 456, + 296, + 553 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 456, + 296, + 553 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 55, + 456, + 296, + 553 + ], + "type": "text", + "content": " are mean and standard variance respectively, and " + }, + { + "bbox": [ + 55, + 456, + 296, + 553 + ], + "type": "inline_equation", + "content": "\\varepsilon_{i,j},\\eta_{i,j}\\sim \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 55, + 456, + 296, + 553 + ], + "type": "text", + "content": ". Here, we adopt " + }, + { + "bbox": [ + 55, + 456, + 296, + 553 + ], + "type": "inline_equation", + "content": "\\eta_{i,j}" + }, + { + "bbox": [ + 55, + 456, + 296, + 553 + ], + "type": "text", + "content": " to perturb the semantic information and a constant 1 to prevent large semantic shifts. Notably, the generative model " + }, + { + "bbox": [ + 55, + 456, + 296, + 553 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\cdot)" + }, + { + "bbox": [ + 55, + 456, + 296, + 553 + ], + "type": "text", + "content": " is flexible, which can be replaced by AdaIN [14], AdaINnoise [10], or StyIN [40], etc. This flexibility allows SAC to serve as a plug-and-play component for other methods. Subsequently, we combine each patch in its original order:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 83, + 557, + 295, + 637 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 557, + 295, + 637 + ], + "spans": [ + { + "bbox": [ + 83, + 557, + 295, + 637 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tilde {\\mathbf {r}} = \\operatorname {C o n c a t} \\left(\\left[ \\tilde {p} _ {i, j} \\right] _ {i, j = 0, \\dots , N - 1}\\right) \\\\ = \\left[ \\begin{array}{c c c c} \\tilde {p} _ {0, 0} & \\tilde {p} _ {0, 1} & \\dots & \\tilde {p} _ {0, N - 1} \\\\ \\tilde {p} _ {1, 0} & \\tilde {p} _ {1, 1} & \\dots & \\tilde {p} _ {1, N - 1} \\\\ \\vdots & \\vdots & \\ddots & \\vdots \\\\ \\tilde {p} _ {N - 1, 0} & \\tilde {p} _ {N - 1, 1} & \\dots & \\tilde {p} _ {N - 1, N - 1} \\end{array} \\right]. \\tag {4} \\\\ \\end{array}", + "image_path": "0f5cd8aa9b3982fa5bf1e848bbaae582605dd7f755bd17b842a5307884d4b374.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 641, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 665 + ], + "type": "text", + "content": "By projecting back to the pixel space using the decoder, we can obtain an augmented image " + }, + { + "bbox": [ + 55, + 641, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}\\in \\mathbb{R}^{H\\times W}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 670, + 295, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 670, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 138, + 670, + 295, + 685 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {x}} = \\operatorname {D e c o d e r} (\\tilde {\\mathbf {r}}), \\tag {5}", + "image_path": "26e263c046ec575f330551a8321ecbb0ef2ca08c997e66894d2f92ada0431eab.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is from the decoder, which is learnable deep layers instantiated as two convolutional layers, followed by ReLU" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 351, + 555, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 555, + 375 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 555, + 375 + ], + "type": "text", + "content": "and Sigmoid activation function respectively. In summary, our patch-aware augmentation is:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 396, + 384, + 468, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 384, + 468, + 396 + ], + "spans": [ + { + "bbox": [ + 396, + 384, + 468, + 396 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {x}} = \\operatorname {P A A} (\\mathbf {x}, \\mathcal {Q}),", + "image_path": "994245ccf309cb2682d9543e36b3d063cb02aede68893513841b88cf195e4f88.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 350, + 396, + 553, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 396, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 350, + 396, + 553, + 420 + ], + "type": "interline_equation", + "content": "\\text {w h e r e} \\quad \\mathcal {Q} = \\left\\{\\left\\{\\varepsilon_ {i, j} \\right\\} _ {i, j = 0} ^ {N - 1}, \\left\\{\\eta_ {i, j} \\right\\} _ {i, j = 0} ^ {N - 1} \\right\\}. \\tag {6}", + "image_path": "2840e14c9587c16d749d7e0908a8494ba4bdd87c8b8e11c97581dbf7b5e8a713.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 426, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 605 + ], + "type": "text", + "content": "By applying distinct Gaussian noise to different patches, each patch acquires a unique color, resulting in an image with different local colors. However, " + }, + { + "bbox": [ + 313, + 426, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}" + }, + { + "bbox": [ + 313, + 426, + 555, + 605 + ], + "type": "text", + "content": " cannot be directly used to capture domain invariant representations, since the introduction of illogical styles during patch-aware augmentation. Specifically, nearby pixels with similar textures can be grouped into primitive shapes that represent parts of objects in images [36]. For instance, in animal classification, an animal's eyes as a whole should present a similar texture. However, this consistency can be disrupted by patch-aware augmentation, which relies on the uniform grid division. When an object spans multiple patches, independent style transformations can lead to conflicting texture regions, causing the model to misinterpret the object as fragmented components rather than a unified whole." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": "Adaptive Random Convolutions. To preserve object integrity, we introduce adaptive random convolutions. Inspired by [4, 36], we use the deformable convolution layer with random filters to project random and coherent texture into images. Furthermore, we multiply the convolution kernel element-wise with a Gaussian filter to further smooth the texture in an object spanning multiple patches. Additionally, we choose the deformable convolution rather than the conventional convolution since the deformable convo" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15619" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": "lution can better sample object contours instead of being confined to patch boundaries [7], helping smooth the texture. (see Tab. 7 for ablation studies on convolution types). Given an augmented sample " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": ", we first initialize a " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": " kernel with dilation 1, which defines a regular grid " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\{(-1, -1), (-1, 0), \\dots, (0, 1), (1, 1)\\}" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": " represents the fixed sampling locations of a standard " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": " convolution kernel, where each coordinate pair " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "(i_m, j_m)" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": " denotes the fixed offsets relative to the center pixel " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "(i_0, j_0)" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": ". In deformable convolution, these offsets are perturbed by random values " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "(\\Delta i_m, \\Delta j_m)" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": ", enabling the sampling locations to deviate from the regular grid. We apply this deformable convolution to " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 235, + 310, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 235, + 310, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 235, + 310, + 287 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tilde{\\mathbf{x}}^{\\prime}[i_{0},j_{0}] = \\sum_{(i_{m},j_{m})\\in \\mathcal{R}}\\mathbf{w}[i_{m},j_{m}]\\cdot \\tilde{\\mathbf{x}} [i_{0} + i_{m} + \\Delta i_{m},j_{0} + j_{m} + \\Delta j_{m}], \\\\ \\text {w h e r e} \\quad \\mathbf {w} [ i _ {m}, j _ {m} ] = \\mathbf {g} [ i _ {m}, j _ {m} ] \\cdot w [ i _ {m}, j _ {m} ], \\tag {7} \\\\ \\end{array}", + "image_path": "74d2105582ccc5569c16ab50f90c687a5bb8cf4df0c819ed3c4f396534fa8a03.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 288, + 295, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 295, + 403 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 288, + 295, + 403 + ], + "type": "inline_equation", + "content": "i_m, j_m" + }, + { + "bbox": [ + 55, + 288, + 295, + 403 + ], + "type": "text", + "content": " are the locations in " + }, + { + "bbox": [ + 55, + 288, + 295, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 55, + 288, + 295, + 403 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 288, + 295, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 55, + 288, + 295, + 403 + ], + "type": "text", + "content": " indicates the weights of the convolution kernel that are randomized, " + }, + { + "bbox": [ + 55, + 288, + 295, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{g}[i_m, j_m] = \\exp(-\\frac{i_m^2 + j_m^2}{2\\sigma_g^2})" + }, + { + "bbox": [ + 55, + 288, + 295, + 403 + ], + "type": "text", + "content": " is a Gaussian filter, " + }, + { + "bbox": [ + 55, + 288, + 295, + 403 + ], + "type": "inline_equation", + "content": "w[i_m, j_m]" + }, + { + "bbox": [ + 55, + 288, + 295, + 403 + ], + "type": "text", + "content": " is sampled from a Gaussian distribution " + }, + { + "bbox": [ + 55, + 288, + 295, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0, \\frac{1}{k^2C_{in}})" + }, + { + "bbox": [ + 55, + 288, + 295, + 403 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 288, + 295, + 403 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 288, + 295, + 403 + ], + "type": "text", + "content": " is the size of kernel size, " + }, + { + "bbox": [ + 55, + 288, + 295, + 403 + ], + "type": "inline_equation", + "content": "C_{in}" + }, + { + "bbox": [ + 55, + 288, + 295, + 403 + ], + "type": "text", + "content": " is the number of input channels, and " + }, + { + "bbox": [ + 55, + 288, + 295, + 403 + ], + "type": "inline_equation", + "content": "\\Delta i_m, \\Delta j_m" + }, + { + "bbox": [ + 55, + 288, + 295, + 403 + ], + "type": "text", + "content": " denote the random offsets of deformable convolution. Each location " + }, + { + "bbox": [ + 55, + 288, + 295, + 403 + ], + "type": "inline_equation", + "content": "(i_0, j_0)" + }, + { + "bbox": [ + 55, + 288, + 295, + 403 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 55, + 288, + 295, + 403 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}'" + }, + { + "bbox": [ + 55, + 288, + 295, + 403 + ], + "type": "text", + "content": " is transformed by the weighted summation of " + }, + { + "bbox": [ + 55, + 288, + 295, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 55, + 288, + 295, + 403 + ], + "type": "text", + "content": " and the pixel values at irregular locations " + }, + { + "bbox": [ + 55, + 288, + 295, + 403 + ], + "type": "inline_equation", + "content": "(i_0 + i_m + \\Delta i_m, j_0 + j_m + \\Delta j_m)" + }, + { + "bbox": [ + 55, + 288, + 295, + 403 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 288, + 295, + 403 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}" + }, + { + "bbox": [ + 55, + 288, + 295, + 403 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 403, + 296, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 296, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 296, + 450 + ], + "type": "text", + "content": "Afterward, following the convention of random convolution [4, 36], we employ affine transformation with two parameters " + }, + { + "bbox": [ + 55, + 403, + 296, + 450 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 403, + 296, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 403, + 296, + 450 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 403, + 296, + 450 + ], + "type": "text", + "content": " to further enhance the texture diversity. Given an input image " + }, + { + "bbox": [ + 55, + 403, + 296, + 450 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}'" + }, + { + "bbox": [ + 55, + 403, + 296, + 450 + ], + "type": "text", + "content": ", this process can be written as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 91, + 458, + 259, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 458, + 259, + 472 + ], + "spans": [ + { + "bbox": [ + 91, + 458, + 259, + 472 + ], + "type": "interline_equation", + "content": "\\operatorname {A R C} (\\tilde {\\mathbf {x}}, \\gamma , \\beta) = \\tanh (\\operatorname {A d a I N} (\\tilde {\\mathbf {x}} ^ {\\prime}, \\gamma , \\beta)),", + "image_path": "821d021ce40f6e9f39232df6d698f183c1ff7dd52470b697f814ba84f6ebdc61.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 474, + 295, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 474, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 91, + 474, + 295, + 502 + ], + "type": "interline_equation", + "content": "\\operatorname {A d a I N} \\left(\\tilde {\\mathbf {x}} ^ {\\prime}, \\gamma , \\beta\\right) = \\gamma \\frac {\\tilde {\\mathbf {x}} ^ {\\prime} - \\mu \\left(\\tilde {\\mathbf {x}} ^ {\\prime}\\right)}{\\sqrt {\\sigma^ {2} \\left(\\tilde {\\mathbf {x}} ^ {\\prime}\\right) + \\epsilon}} + \\beta , \\tag {8}", + "image_path": "d3d3547a903bfbb244ae87c280d1579d4c1921d4aee8e742334a09976e2ecfd7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mu (\\cdot)" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " represent the mean and variance, " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " is a positive number, and " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\gamma ,\\beta \\sim \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": ". Then, the tanh is applied to scale the augmented images. We aim to iterate Eq. (7) and Eq. (8) " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " times to progressively improve the texture diversity. However, this iterative process presents potential risks. Specifically, excessive iteration can distort semantic information and hinder the learning of invariant representations, while insufficient iterations may fail to adequately enhance texture diversity. Moreover, applying a uniform iteration count across all samples in a batch without considering individual discrepancies can lead to semantic distortions in some samples, degrading model performance. To this end, we introduce an adaptive strategy using the entropy to assess whether a sample requires augmentation during the iterative process. Specifically, for each batch of data, we iterate up to 10 times. In the first iteration, all samples undergo augmentation. In subsequent iterations, we" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": "compute the entropy of each sample and apply augmentation only to those with entropy values below a predefined threshold, while high-entropy samples remain unchanged. Entropy serves as a measure of uncertainty, where high entropy indicates a high likelihood of distribution shifts. Our strategy adaptively controls augmentation based on entropy, ensuring that low-entropy samples undergo augmentation to adequately enhance texture complexity, while high-entropy samples remain unchanged to prevent distortion of semantic information. Our strategy can be summarized as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 328, + 201, + 555, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 201, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 328, + 201, + 555, + 228 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {t + 1} ^ {+} = \\left\\{ \\begin{array}{c l} \\operatorname {A R C} (\\mathbf {x} _ {t} ^ {+}, \\gamma , \\beta), & \\text {i f} E (\\mathbf {x} _ {t} ^ {+}) < E _ {0}, \\\\ \\mathbf {x} _ {t} ^ {+}, & \\text {i f} E (\\mathbf {x} _ {t} ^ {+}) \\geq E _ {0}, \\end{array} \\right. \\tag {9}", + "image_path": "9d2b79a4940c6cb28e67e0534c711eb37b8c5ed0fba17fa61a298db0dc723a3e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 237, + 555, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 237, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 237, + 555, + 285 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 237, + 555, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t^+" + }, + { + "bbox": [ + 313, + 237, + 555, + 285 + ], + "type": "text", + "content": " denotes the sample after the " + }, + { + "bbox": [ + 313, + 237, + 555, + 285 + ], + "type": "inline_equation", + "content": "t^{th}" + }, + { + "bbox": [ + 313, + 237, + 555, + 285 + ], + "type": "text", + "content": " iteration (" + }, + { + "bbox": [ + 313, + 237, + 555, + 285 + ], + "type": "inline_equation", + "content": "t \\in [1,9]" + }, + { + "bbox": [ + 313, + 237, + 555, + 285 + ], + "type": "text", + "content": "), " + }, + { + "bbox": [ + 313, + 237, + 555, + 285 + ], + "type": "inline_equation", + "content": "E_0" + }, + { + "bbox": [ + 313, + 237, + 555, + 285 + ], + "type": "text", + "content": " is the threshold defined as " + }, + { + "bbox": [ + 313, + 237, + 555, + 285 + ], + "type": "inline_equation", + "content": "e \\times \\log(C)" + }, + { + "bbox": [ + 313, + 237, + 555, + 285 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 313, + 237, + 555, + 285 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 237, + 555, + 285 + ], + "type": "text", + "content": " represents the number of classes in the dataset and " + }, + { + "bbox": [ + 313, + 237, + 555, + 285 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 313, + 237, + 555, + 285 + ], + "type": "text", + "content": " is set to 0.3) and " + }, + { + "bbox": [ + 313, + 237, + 555, + 285 + ], + "type": "inline_equation", + "content": "E(\\cdot)" + }, + { + "bbox": [ + 313, + 237, + 555, + 285 + ], + "type": "text", + "content": " denotes the entropy loss." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 286, + 555, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 286, + 555, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 286, + 555, + 406 + ], + "type": "text", + "content": "Energy-based Distribution Discrepancy Score. To enhance the ability of the generative model " + }, + { + "bbox": [ + 313, + 286, + 555, + 406 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\cdot)" + }, + { + "bbox": [ + 313, + 286, + 555, + 406 + ], + "type": "text", + "content": " in the feature-level, we propose an energy-based distribution discrepancy score to facilitate the expansion of generated domain shifts using a trained classifier. To assess the energy, we use logits from the classifier for energy representations, which is widely used to evaluate Helmholtz free energy [9]. Given the input data " + }, + { + "bbox": [ + 313, + 286, + 555, + 406 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 286, + 555, + 406 + ], + "type": "text", + "content": ", and the trained " + }, + { + "bbox": [ + 313, + 286, + 555, + 406 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 286, + 555, + 406 + ], + "type": "text", + "content": "-class classifier " + }, + { + "bbox": [ + 313, + 286, + 555, + 406 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 286, + 555, + 406 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 286, + 555, + 406 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 286, + 555, + 406 + ], + "type": "text", + "content": " is the number of classes, we can compute the Helmholtz free energy for " + }, + { + "bbox": [ + 313, + 286, + 555, + 406 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 286, + 555, + 406 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 378, + 415, + 554, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 415, + 554, + 445 + ], + "spans": [ + { + "bbox": [ + 378, + 415, + 554, + 445 + ], + "type": "interline_equation", + "content": "E n (\\mathbf {x}) = - \\log \\sum_ {i} ^ {c} e ^ {P _ {i} (\\mathbf {x})}, \\tag {10}", + "image_path": "4dddaa55084ee51d94ff09e51111bfadf9a8ba4f12ea18c362f8574950281dcd.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 454, + 555, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 555, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 555, + 562 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 454, + 555, + 562 + ], + "type": "inline_equation", + "content": "P_{i}(\\cdot)" + }, + { + "bbox": [ + 313, + 454, + 555, + 562 + ], + "type": "text", + "content": " denotes the probability of the class " + }, + { + "bbox": [ + 313, + 454, + 555, + 562 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 454, + 555, + 562 + ], + "type": "text", + "content": ". It is known that the energy exhibits the uncertainty of the input data with respect to the classifier, where high energy indicates the high probability of being out-of-distribution [21]. Thus, the energy discrepancy can serve as the score to measure the distribution discrepancy between two data under the same view. Given input data " + }, + { + "bbox": [ + 313, + 454, + 555, + 562 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_1" + }, + { + "bbox": [ + 313, + 454, + 555, + 562 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 454, + 555, + 562 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_2" + }, + { + "bbox": [ + 313, + 454, + 555, + 562 + ], + "type": "text", + "content": ", and the trained " + }, + { + "bbox": [ + 313, + 454, + 555, + 562 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 454, + 555, + 562 + ], + "type": "text", + "content": "-class classifier " + }, + { + "bbox": [ + 313, + 454, + 555, + 562 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 454, + 555, + 562 + ], + "type": "text", + "content": ", we obtain the distribution discrepancy score by:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 358, + 571, + 554, + 598 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 571, + 554, + 598 + ], + "spans": [ + { + "bbox": [ + 358, + 571, + 554, + 598 + ], + "type": "interline_equation", + "content": "S (\\mathbf {x} _ {1}, \\mathbf {x} _ {2}) = \\frac {1}{\\left\\| E n (\\mathbf {x} _ {1}) - E n (\\mathbf {x} _ {2}) \\right\\| _ {1}}, \\tag {11}", + "image_path": "7478ad499d696192a8b5fd5244993ab3a665cdcc6ab6cf0404e2079edcf1256b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_1" + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "text", + "content": " is the mean absolute error [35]. Here, " + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "inline_equation", + "content": "S(\\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "text", + "content": " represents the approximate distribution distance between two data. During training, we leverage this score to enhance the distribution discrepancy between the original data and generated data, as well as among different generated data, improving the expansion of the training data distribution." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Training Objective. Since the parameters in ARC are randomly initialized, our loss term in the style generation module is only used to constrain PAA. To prevent semantic" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15620" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": "information from distortion, we use the cross-entropy to ensure that the augmented data can be correctly predicted by the model. Specifically, since the original and augmented data correspond to different styles but share the same semantic information, the model should assign them the same label if it classifies based on semantic information." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 149, + 293, + 180 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 149, + 293, + 180 + ], + "spans": [ + { + "bbox": [ + 59, + 149, + 293, + 180 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c e} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(\\ell \\left(P (F (\\tilde {\\mathbf {x}} _ {i, 1})), \\mathbf {y} _ {i}\\right) + \\ell \\left(P (F (\\tilde {\\mathbf {x}} _ {i, 2})), \\mathbf {y} _ {i}\\right)\\right),", + "image_path": "2dc28095a361b99f20487deaab434867aaa302b9d2d1cee43386325f629e7ffc.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 182, + 295, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 182, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 59, + 182, + 295, + 205 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 59, + 182, + 295, + 205 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_{i,1} = \\mathrm{PAA}(\\mathbf{x}_i,\\mathcal{Q}_1),\\quad \\tilde{\\mathbf{x}}_{i,2} = \\mathrm{PAA}(\\mathbf{x}_i,\\mathcal{Q}_2)," + }, + { + "bbox": [ + 59, + 182, + 295, + 205 + ], + "type": "text", + "content": " (12)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 207, + 296, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 207, + 296, + 315 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 296, + 315 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 207, + 296, + 315 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 55, + 207, + 296, + 315 + ], + "type": "text", + "content": " indicates input original data, " + }, + { + "bbox": [ + 55, + 207, + 296, + 315 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 55, + 207, + 296, + 315 + ], + "type": "text", + "content": " is the cross-entropy loss, " + }, + { + "bbox": [ + 55, + 207, + 296, + 315 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i" + }, + { + "bbox": [ + 55, + 207, + 296, + 315 + ], + "type": "text", + "content": " is the label for " + }, + { + "bbox": [ + 55, + 207, + 296, + 315 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 55, + 207, + 296, + 315 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 207, + 296, + 315 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 207, + 296, + 315 + ], + "type": "text", + "content": " is the number of available data. The augmented data " + }, + { + "bbox": [ + 55, + 207, + 296, + 315 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_{i,1}" + }, + { + "bbox": [ + 55, + 207, + 296, + 315 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 207, + 296, + 315 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_{i,2}" + }, + { + "bbox": [ + 55, + 207, + 296, + 315 + ], + "type": "text", + "content": " are synthesized using different parameter sets " + }, + { + "bbox": [ + 55, + 207, + 296, + 315 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_1" + }, + { + "bbox": [ + 55, + 207, + 296, + 315 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 207, + 296, + 315 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_2" + }, + { + "bbox": [ + 55, + 207, + 296, + 315 + ], + "type": "text", + "content": ", respectively. Next, we maximize distribution discrepancies among the original data " + }, + { + "bbox": [ + 55, + 207, + 296, + 315 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 207, + 296, + 315 + ], + "type": "text", + "content": " and different augmented data " + }, + { + "bbox": [ + 55, + 207, + 296, + 315 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_1" + }, + { + "bbox": [ + 55, + 207, + 296, + 315 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 207, + 296, + 315 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_2" + }, + { + "bbox": [ + 55, + 207, + 296, + 315 + ], + "type": "text", + "content": ", encouraging the generative model to consistently generate data that exhibit large domain shifts relative to the original data and previously generated data:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 322, + 295, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 322, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 62, + 322, + 295, + 365 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(s \\left(\\tilde {\\mathbf {x}} _ {i, 1}, \\mathbf {x} _ {i}\\right) + s \\left(\\tilde {\\mathbf {x}} _ {i, 2}, \\mathbf {x} _ {i}\\right) + s \\left(\\tilde {\\mathbf {x}} _ {i, 1}, \\tilde {\\mathbf {x}} _ {i, 2}\\right)\\right). \\tag {13}", + "image_path": "e9fced1da7d9a372d1e30f3388bbd7e4b3edc819ca0bdbc6fe32acf199b0c870.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 373, + 296, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 409 + ], + "type": "text", + "content": "Additionally, we use mean absolute error to compel the generative model to consistently generate data with pixel-level discrepancies:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 416, + 295, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 416, + 295, + 447 + ], + "spans": [ + { + "bbox": [ + 116, + 416, + 295, + 447 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\| \\tilde {\\mathbf {x}} _ {i, 1} - \\tilde {\\mathbf {x}} _ {i, 2} \\| _ {1}. \\tag {14}", + "image_path": "ffc54c14b9b832ea7eacdf39b1ae3f26d4653ef9b38c6a03063ad749b48887e6.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 454, + 247, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 247, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 247, + 466 + ], + "type": "text", + "content": "In summary, the overall objective is as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 117, + 475, + 295, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 475, + 295, + 488 + ], + "spans": [ + { + "bbox": [ + 117, + 475, + 295, + 488 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S} = \\mathcal {L} _ {c e} + \\lambda_ {d} \\mathcal {L} _ {d} - \\lambda_ {p} \\mathcal {L} _ {p}, \\tag {15}", + "image_path": "99d201058a285eda3b9c1390141a8028ade1c88bc3f0a4376180b00d88ba252e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 495, + 229, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 229, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 229, + 509 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 495, + 229, + 509 + ], + "type": "inline_equation", + "content": "\\lambda_{d}" + }, + { + "bbox": [ + 55, + 495, + 229, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 495, + 229, + 509 + ], + "type": "inline_equation", + "content": "\\lambda_{p}" + }, + { + "bbox": [ + 55, + 495, + 229, + 509 + ], + "type": "text", + "content": " are the hyper-parameters." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 514, + 234, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 234, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 234, + 528 + ], + "type": "text", + "content": "3.2. Representation Learning Module" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 532, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 296, + 581 + ], + "type": "text", + "content": "Benefiting from our style generation module, the representation learning module can be supervised by a simple cross-entropy loss function and learn domain invariant representations. The loss for each sample pair is:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 58, + 587, + 287, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 587, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 58, + 587, + 287, + 651 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {R} = \\frac {1}{2 n} \\sum_ {i = 1} ^ {n} \\left(\\ell \\left(P \\left(F \\left(\\mathbf {x} _ {i} ^ {+}\\right)\\right), \\mathbf {y} _ {i}\\right) + \\ell \\left(P \\left(F \\left(\\mathbf {x} _ {i} ^ {+ +}\\right)\\right), \\mathbf {y} _ {i}\\right)\\right) \\\\ + \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\ell \\left(P (F (\\mathbf {x} _ {i})), \\mathbf {y} _ {i}\\right), \\\\ \\end{array}", + "image_path": "3fb7afbfaccebfca962258fb9f4122b5d52461ea4a2ff3e947366015d5cd9dd9.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 653, + 299, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 299, + 680 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 299, + 680 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 653, + 299, + 680 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^+ = \\mathrm{ARC}(\\tilde{\\mathbf{x}}_{i,1},\\gamma_1,\\beta_1),\\mathbf{x}_i^{+ + } = \\mathrm{ARC}(\\tilde{\\mathbf{x}}_{i,2},\\gamma_2,\\beta_2)" + }, + { + "bbox": [ + 55, + 653, + 299, + 680 + ], + "type": "text", + "content": " (16)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^+" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^{+ + }" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " are synthesized from " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " through distinct parameters." + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 317, + 101, + 553, + 235 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 554, + 93 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 554, + 93 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 554, + 93 + ], + "type": "text", + "content": "Table 1. Accuracy (\\%) of different methods on Digits. Each column title indicates the target domain." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 101, + 553, + 235 + ], + "lines": [ + { + "bbox": [ + 317, + 101, + 553, + 235 + ], + "spans": [ + { + "bbox": [ + 317, + 101, + 553, + 235 + ], + "type": "table", + "html": "
MethodsSVHNMNIST-MSYNUSPSAvg.
ERM [15]27.852.739.776.949.3
ADA [32]NeurIPS'1835.560.445.377.354.6
M-ADA [27]CVPR'2042.667.949.078.559.5
ME-ADA [44]NeurIPS'2042.663.350.481.059.3
PDEN [20]CVPR'2162.282.269.485.374.8
L2D [34]ICCV'2162.987.363.784.074.5
NCDG [31]TPAMI'2359.777.463.892.673.4
SADA [42]AAAI'2368.575.672.989.376.6
Pro-RandConv [4]CVPR'2369.782.379.893.781.4
UDP [10]MM'2372.479.781.796.382.5
AdvST [45]AAAI'2467.579.878.195.480.1
PSDG [40]KDD'2473.186.783.495.784.7
AdvFrequency [12]NeurIPS'2473.682.779.989.381.4
SAC77.791.087.996.488.3
", + "image_path": "a49b7426624b19a21972703ca3292d30b44ed5f2dceae94678e3ffb6097e5e01.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 258, + 418, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 258, + 418, + 272 + ], + "spans": [ + { + "bbox": [ + 314, + 258, + 418, + 272 + ], + "type": "text", + "content": "3.3. Training Pipeline" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 277, + 555, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 494 + ], + "type": "text", + "content": "In real-world scenarios, since the styles of target domains exhibit complexity, using a single generated domain may not produce data with sufficient styles. To address this limitation, we employ a progressively adversarial max-min game to optimize the style generation and representation learning module. We use empirical risk minimization (ERM) to pretrain the representation learning module with original data. During training, we first fix the weights of the representation learning module and train the style generation module " + }, + { + "bbox": [ + 313, + 277, + 555, + 494 + ], + "type": "inline_equation", + "content": "S_{1}" + }, + { + "bbox": [ + 313, + 277, + 555, + 494 + ], + "type": "text", + "content": " using PAA and ARC to synthesize the augmented data. Next, we fix the weights of the style generation module and utilize the original and augmented data to optimize the representation learning module using Eq. (16). This adversarial process is repeated several times until we obtain " + }, + { + "bbox": [ + 313, + 277, + 555, + 494 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 277, + 555, + 494 + ], + "type": "text", + "content": " style generation modules " + }, + { + "bbox": [ + 313, + 277, + 555, + 494 + ], + "type": "inline_equation", + "content": "\\{S_{1}, S_{2}, \\dots, S_{K}\\}" + }, + { + "bbox": [ + 313, + 277, + 555, + 494 + ], + "type": "text", + "content": " and a trained representation learning module with good generalization ability. Finally, we use the representation learning model to predict test data labels." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 509, + 395, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 509, + 395, + 522 + ], + "spans": [ + { + "bbox": [ + 314, + 509, + 395, + 522 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 530, + 479, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 479, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 479, + 544 + ], + "type": "text", + "content": "4.1. Datasets and Implementations" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 550, + 554, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 554, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 554, + 609 + ], + "type": "text", + "content": "To show the superiority of SAC for single domain generalization, we conduct experiments on four datasets: Digits [32], PACS [18], CIFAR10-C [13] and DomainNet [26]. Please see the supplementary material for more implementation details." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 622, + 439, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 622, + 439, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 622, + 439, + 635 + ], + "type": "text", + "content": "4.2. Comparison Methods" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": "We compare SAC with 13 competitive methods, including ERM [15], ADA [32], M-ADA [27], ME-ADA [44], PDEN [20], L2D [34], NCDG [31], SADA [42], Pro-RandConv [4], UDP [10], AdvST [45], PSDG [40], and AdvFrequency [12]. Note that all baselines are compared using the same structures and * indicates our implementation." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15621" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 101, + 294, + 240 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 295, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 295, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 295, + 94 + ], + "type": "text", + "content": "Table 2. Accuracy (\\%) of different methods on PACS. Each column title indicates the source domain." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 101, + 294, + 240 + ], + "lines": [ + { + "bbox": [ + 57, + 101, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 57, + 101, + 294, + 240 + ], + "type": "table", + "html": "
MethodsPhotoArtCartoonSketchAvg.
ERM [15]42.070.976.553.160.7
ADA [32]NeurIPS'1841.768.869.335.153.7
M-ADA [27]CVPR'2043.168.071.933.954.2
ME-ADA [44]NeurIPS'2042.169.270.636.854.7
PDEN [20]CVPR'2158.377.974.657.967.2
L2D [34]ICCV'2152.376.977.953.765.2
NCDG [31]TPAMI'2349.076.676.453.163.8
SADA [42]AAAI'2351.277.876.957.866.2
Pro-RandConv [4]CVPR'2362.977.078.557.168.9
UDP [10]MM'2356.777.177.657.567.2
AdvST [45]AAAI'2464.177.976.057.168.8
PSDG [40]KDD'2462.178.278.363.070.4
AdvFrequency [12]NeurIPS'2458.475.778.658.967.9
SAC61.779.279.264.671.2
", + "image_path": "7e9293ef2a085105fd0b509da2a3ca7f6e378a14b0064f956372ff89a8194df1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 262, + 180, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 180, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 180, + 276 + ], + "type": "text", + "content": "4.3. Experimental Results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 281, + 295, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 281, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 54, + 281, + 295, + 449 + ], + "type": "text", + "content": "Comparisons on Digits. We evaluate the effectiveness of SAC on the Digits dataset, as shown in Tab. 1. Compared to the state-of-the-art baseline PSDG, SAC achieves an average accuracy of " + }, + { + "bbox": [ + 54, + 281, + 295, + 449 + ], + "type": "inline_equation", + "content": "88.3\\%" + }, + { + "bbox": [ + 54, + 281, + 295, + 449 + ], + "type": "text", + "content": ", representing a " + }, + { + "bbox": [ + 54, + 281, + 295, + 449 + ], + "type": "inline_equation", + "content": "3.6\\%" + }, + { + "bbox": [ + 54, + 281, + 295, + 449 + ], + "type": "text", + "content": " improvement. This result demonstrates that SAC generates augmented data with more diverse styles and captures more domain invariant representations than previous approaches. Notably, SAC achieves the highest accuracy across all tasks, underscoring the effectiveness of our proposed PAA and ARC in synthesizing complex styles, thereby facilitating the model generalizing to the unknown domains. These remarked results indicate that SAC captures more generalizable information than previous approaches, further highlighting the advantage of both improving color and texture diversity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 449, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 628 + ], + "type": "text", + "content": "Comparisons on PACS. Tab. 2 shows the experimental results on PACS dataset, demonstrating the superior performance of SAC compared to competing methods. We first observe that SAC achieves a higher average accuracy than the previous state-of-the-art method PSDG. Specifically, SAC outperforms PSDG by an average of " + }, + { + "bbox": [ + 55, + 449, + 295, + 628 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 55, + 449, + 295, + 628 + ], + "type": "text", + "content": ", which clearly demonstrates its effectiveness. Notably, SAC achieves impressive accuracy on the most challenging Sketch task, which uses simplistic images without backgrounds as the training data. The impressive accuracy in this task highlights the superiority of SAC. In addition, although SAC performs a " + }, + { + "bbox": [ + 55, + 449, + 295, + 628 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 55, + 449, + 295, + 628 + ], + "type": "text", + "content": " lower accuracy in the Photo task compared to AdvST, it achieves an overall average improvement of " + }, + { + "bbox": [ + 55, + 449, + 295, + 628 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 55, + 449, + 295, + 628 + ], + "type": "text", + "content": ", further demonstrating its enhanced generalizability relative to AdvST." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": "Comparisons on CIFAR10-C. As shown in Tab. 3, SAC achieves state-of-the-art (SOTA) performance, significantly outperforming competitors in accuracy. SAC demonstrates notable improvements, achieving a high average accuracy of " + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "inline_equation", + "content": "81.7\\%" + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": ", which is a " + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": " increase over the previous SOTA method, UDP. Furthermore, compared to UDP, our method achieves substantial gains, with improvements of " + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "inline_equation", + "content": "3.9\\%" + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": " for" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 112, + 553, + 255 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 3. Accuracy (\\%) of different methods on CIFAR10-C under different corruption categories at the level 5. Each column title indicates the target domain." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 112, + 553, + 255 + ], + "lines": [ + { + "bbox": [ + 317, + 112, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 317, + 112, + 553, + 255 + ], + "type": "table", + "html": "
MethodsWeatherBlurNoiseDigitAvg.
ERM [15]67.356.730.062.354.1
ADA [32]NeurIPS'1872.767.040.066.661.6
M-ADA [27]CVPR'2075.536.754.265.164.7
ME-ADA [44]NeurIPS'2074.471.466.570.870.8
PDEN [20]CVPR'2178.676.778.673.176.8
L2D [34]ICCV'2176.070.273.372.072.9
NCDG [31]TPAMI'2382.282.870.976.278.0
SADA [42]AAAI'2379.182.471.478.477.8
Pro-RandConv* [4]CVPR'2377.878.676.880.778.5
UDP [10]MM'2380.783.679.976.180.1
AdvST [45]AAAI'2481.578.476.579.479.0
PSDG [40]KDD'2482.881.779.176.179.9
AdvFrequency [12]NeurIPS'2480.481.773.777.578.3
SAC80.182.683.880.481.7
", + "image_path": "70145bb730414a4c7e31694c8941403bffa833a76ec351d822694b5a44547e0c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 316, + 297, + 553, + 378 + ], + "blocks": [ + { + "bbox": [ + 313, + 266, + 555, + 289 + ], + "lines": [ + { + "bbox": [ + 313, + 266, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 555, + 289 + ], + "type": "text", + "content": "Table 4. Accuracy (\\%) of different methods on DomainNet. Each column title indicates the target domain." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 297, + 553, + 378 + ], + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 378 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 378 + ], + "type": "table", + "html": "
MethodsPaintingInfographClipartSketchQuickdrawAvg.
ERM* [15]38.314.035.322.92.122.5
ME-ADA [44]NeurIPS/2039.014.041.025.34.324.7
PDEN* [20]CVPR/2139.013.138.929.74.725.1
L2D [34]ICCV/2138.412.038.730.16.525.2
Pro-RandConv* [4]CVPR/2342.314.235.433.25.927.4
UDP* [10]MM/2338.613.039.130.14.725.1
AdvST [45]AAAI/2442.314.841.530.85.927.1
PSDG* [40]KDD/2439.013.539.530.54.325.4
AdvFrequency [12]NeurIPS/2442.815.142.731.26.927.7
SAC42.513.943.334.97.128.4
", + "image_path": "35074a2b586fca91a945404153d29c3c6963edb0c5404e0a1c97274504edd73b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 399, + 553, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 553, + 422 + ], + "type": "text", + "content": "Noise and " + }, + { + "bbox": [ + 313, + 399, + 553, + 422 + ], + "type": "inline_equation", + "content": "4.3\\%" + }, + { + "bbox": [ + 313, + 399, + 553, + 422 + ], + "type": "text", + "content": " for Digit. These impressive results highlight the advantages of SAC." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 424, + 555, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 424, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 312, + 424, + 555, + 591 + ], + "type": "text", + "content": "Comparisons on DomainNet. To further evaluate the superiority of SAC, we compare it with the recent methods on the large-scale DomainNet benchmark. The results are presented in Tab. 4. We can observe that SAC achieves the highest average accuracy among all baselines. Additionally, the methods that solely introduce Gaussian noise to improve color diversity, such as UDP and PSDG, exhibit relatively poor performance. We conjecture that generating images with simplistic textures fails to simulate the complexity of real-world images. While Pro-RandConv achieves competitive accuracy by enhancing texture diversity, SAC still outperforms it in terms of average accuracy. These results validate the effectiveness of SAC, which simultaneously enhances both color and texture complexity." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 600, + 435, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 435, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 435, + 613 + ], + "type": "text", + "content": "4.4. Analysis Experiment" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": "Ablation Study. We provide the ablation study on theDigits dataset to evaluate the effectiveness of key components in SAC, as shown in Tab. 5. Specifically, we introduce some variants to analyze the contributions of SAC, ARC, the Gaussian filter " + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{g}" + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": ". When we remove a component, all others remain default settings. We first delete ARC (SAC W/O ARC) to assess the effectiveness of preserving object integrity. We observe that the performance" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15622" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 101, + 293, + 196 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 295, + 94 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 295, + 94 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 295, + 94 + ], + "type": "text", + "content": "Table 5. Ablation study on Digits. Each column title indicates the target domain." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 101, + 293, + 196 + ], + "lines": [ + { + "bbox": [ + 57, + 101, + 293, + 196 + ], + "spans": [ + { + "bbox": [ + 57, + 101, + 293, + 196 + ], + "type": "table", + "html": "
MethodsSVHNMNIST-MSYNUSPSAvg.
SAC W/O ARC64.179.175.192.777.7
SAC W/O PAA74.985.386.096.185.3
SAC W/O g77.688.886.594.286.8
SAC W/O Ld75.590.387.096.787.4
SAC W/O Lp74.989.786.796.787.0
SAC W/O Ada76.390.585.595.486.9
SAC W/O Ada max73.989.885.495.386.1
SAC77.791.087.996.488.3
", + "image_path": "7ac64e5fab025f954fddccd7d5be94d25ee5082085f508c5325e546fad3e7da4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 220, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 220, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 57, + 220, + 295, + 495 + ], + "type": "text", + "content": "drops seriously, which indicates the harm of illogical textures. Next, to investigate the influence of PAA, we only employ ARC (SAC W/O PAA) to generate images. The lower accuracy compared to SAC suggests that focusing solely on texture diversity is insufficient to enhance overall style diversity. We then remove the Gaussian filter " + }, + { + "bbox": [ + 57, + 220, + 295, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{g}" + }, + { + "bbox": [ + 57, + 220, + 295, + 495 + ], + "type": "text", + "content": " (SAC W/O " + }, + { + "bbox": [ + 57, + 220, + 295, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{g}" + }, + { + "bbox": [ + 57, + 220, + 295, + 495 + ], + "type": "text", + "content": ") to evaluate its effectiveness in preserving object integrity. The degraded performance demonstrates that " + }, + { + "bbox": [ + 57, + 220, + 295, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{g}" + }, + { + "bbox": [ + 57, + 220, + 295, + 495 + ], + "type": "text", + "content": " can help maintain object integrity and prevent texture inconsistencies. Moreover, we see that eliminating " + }, + { + "bbox": [ + 57, + 220, + 295, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 57, + 220, + 295, + 495 + ], + "type": "text", + "content": " (SAC W/O " + }, + { + "bbox": [ + 57, + 220, + 295, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 57, + 220, + 295, + 495 + ], + "type": "text", + "content": ") or " + }, + { + "bbox": [ + 57, + 220, + 295, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 57, + 220, + 295, + 495 + ], + "type": "text", + "content": " (SAC W/O " + }, + { + "bbox": [ + 57, + 220, + 295, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 57, + 220, + 295, + 495 + ], + "type": "text", + "content": ") would lead to inferior performance compared to SAC, highlighting the importance of expanding the discrepancy at both the feature and pixel levels. Finally, we remove Eq. (9) (SAC W/O Ada) and follow the settings in [4] by iterating the random convolutions " + }, + { + "bbox": [ + 57, + 220, + 295, + 495 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 57, + 220, + 295, + 495 + ], + "type": "text", + "content": " times, where " + }, + { + "bbox": [ + 57, + 220, + 295, + 495 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 57, + 220, + 295, + 495 + ], + "type": "text", + "content": " is sampled from the uniform distribution " + }, + { + "bbox": [ + 57, + 220, + 295, + 495 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 57, + 220, + 295, + 495 + ], + "type": "text", + "content": " [1, 10]. The inferior accuracy indicates that our proposed adaptive strategy effectively improves texture diversity while preventing semantic distortion. To further demonstrate its effectiveness in preserving semantic information, we conduct an experiment in which " + }, + { + "bbox": [ + 57, + 220, + 295, + 495 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 57, + 220, + 295, + 495 + ], + "type": "text", + "content": " is fixed at 10 (SAC W/O " + }, + { + "bbox": [ + 57, + 220, + 295, + 495 + ], + "type": "inline_equation", + "content": "\\mathrm{Ada}_{\\mathrm{max}}" + }, + { + "bbox": [ + 57, + 220, + 295, + 495 + ], + "type": "text", + "content": "). The approximately " + }, + { + "bbox": [ + 57, + 220, + 295, + 495 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 57, + 220, + 295, + 495 + ], + "type": "text", + "content": " performance drop confirms the importance of Eq. (9)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 496, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 295, + 628 + ], + "type": "text", + "content": "Complementarity with Other Approaches. SAC can serve as a plug-and-play component to enhance the generalizability of other methods. Specifically, we apply PAA and ARC to the approaches of PDEN, L2D, RandConv, ProRandConv, and PSDG, respectively. The experimental results on Digits are shown in Tab. 6. These results indicate that ARC can serve as a general enhancer, improving the generalizability of other methods. Notably, when we apply ARC to PDEN, we find that \"PDEN+ARC\" achieves an " + }, + { + "bbox": [ + 55, + 496, + 295, + 628 + ], + "type": "inline_equation", + "content": "8.1\\%" + }, + { + "bbox": [ + 55, + 496, + 295, + 628 + ], + "type": "text", + "content": " improvement of average accuracy under default settings compared to PDEN." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": "Comparative Analysis of Different Convolutions. In Tab. 7, we evaluate the effectiveness of various convolution types on the Digits dataset, including conventional convolution [17] (SAC-Conv), depthwise separable convolution [5] (SAC-DSCov), wavelet convolution [8] (SAC-WTConv), and deformable convolution [6] (SAC). SAC and its variants achieve outstanding performance, highlighting the ef" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 317, + 112, + 553, + 224 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 104 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 104 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 104 + ], + "type": "text", + "content": "Table 6. Accuracy (\\%) on Digits when SAC is applied to recent methods as a plug-and-play component. Each column title indicates the target domain." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 112, + 553, + 224 + ], + "lines": [ + { + "bbox": [ + 317, + 112, + 553, + 224 + ], + "spans": [ + { + "bbox": [ + 317, + 112, + 553, + 224 + ], + "type": "table", + "html": "
MethodsSVHNMNIST-MSYNUSPSAvg.
PDEN [20]62.282.269.485.374.8
PDEN+SAC73.886.781.590.183.0
L2D [34]62.987.363.784.074.5
L2D+SAC59.882.270.693.176.4
PSDG [40]73.186.783.495.784.7
PSDG+SAC75.890.885.096.387.0
RandConv [36]62.187.963.984.474.6
RandConv+ARC68.286.871.085.878.0
Pro-RandConv [4]69.782.379.893.781.4
Pro-RandConv+ARC70.585.982.795.383.6
", + "image_path": "81f348d402e7f1dde6ca52128028c1d016b4ad76218c7abbffa9606598a8433c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 270, + 552, + 324 + ], + "blocks": [ + { + "bbox": [ + 313, + 239, + 553, + 262 + ], + "lines": [ + { + "bbox": [ + 313, + 239, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 553, + 262 + ], + "type": "text", + "content": "Table 7. Performance comparison of different convolution on Digits. Each column title indicates the target domain." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 270, + 552, + 324 + ], + "lines": [ + { + "bbox": [ + 317, + 270, + 552, + 324 + ], + "spans": [ + { + "bbox": [ + 317, + 270, + 552, + 324 + ], + "type": "table", + "html": "
MethodsSVHNMNIST-MSYNUSPSAvg.
SAC-Conv [36]78.387.387.492.586.4
SAC-DSConv [5]78.789.187.492.486.9
SAC-WTConv [8]74.387.883.895.985.5
SAC77.791.087.996.488.3
", + "image_path": "fd7865a8caffbfd67107c5824d0b65cb06cc81fe2e7e262f899fb078ea271d68.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 350, + 555, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 555, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 555, + 494 + ], + "type": "text", + "content": "fectiveness of our proposed \"Split-And-Combine\" strategy. Specifically, the conventional convolution and depthwise separable convolution remain constrained to uniform grids, limiting their ability to accurately capture object contours and preserve object integrity. While the wavelet convolution expands the receptive field through wavelet transformation, it still extracts features within a uniform grid, making it inherently constrained by grid boundaries. In contrast, the deformable convolution overcomes these limitations by allowing sampling points to deviate from a fixed uniform grid, thereby capturing object shapes and structural details more effectively and contributing to texture smoothing." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 511, + 388, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 511, + 388, + 525 + ], + "spans": [ + { + "bbox": [ + 314, + 511, + 388, + 525 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "We present a novel \"Split-and-Combine\" method for single domain generalization. Instead of applying style augmentation to entire images, our method first performs patch-aware augmentation, which conducts separate style generation to each patch of an image, enriching colors. Then, we introduce adaptive random convolutions to enhance texture complexity while preserving object integrity. Through these synergistic components, our model can synthesize augmented data with different local colors and textures. Additionally, we design an energy-based distribution discrepancy score to assess distribution differences between the original and augmented data, which can encourage SAC to generate novel data with large distribution shifts. SAC consistently achieves state-of-the-art performance across four datasets, demonstrating the effectiveness of our method." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15623" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 224 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China under Grants 62306008, 32472007, 62406095, and 62402004, the Key Project of Anhui Province's Science and Technology Innovation Tackle Plan under Grant 202423k09020040, the China Postdoctoral Science Foundation under Grant 2024M762632, the Natural Science Foundation of Shaanxi under Grant 2024JC-YBQN-0687, the Social Science Foundation of Shaanxi under Grant 2024R027, and the Hefei Key Technology R&D \"Champion-Based Selection\" Project under Grant 2023SGJ011." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 234, + 115, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 234, + 115, + 247 + ], + "spans": [ + { + "bbox": [ + 56, + 234, + 115, + 247 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 254, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 254, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 254, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 61, + 254, + 296, + 297 + ], + "type": "text", + "content": "[1] Kartik Ahuja, Karthikeyan Shanmugam, Kush R. Varshney, and Amit Dhurandhar. Invariant risk minimization games. In International Conference on Machine Learning, Virtual Event, July 13-18, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 300, + 296, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 300, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 61, + 300, + 296, + 364 + ], + "type": "text", + "content": "[2] Kartik Ahuja, Ethan Caballero, Dinghuai Zhang, Jean-Christophe Gagnon-Audet, Yoshua Bengio, Ioannis Mitliagkas, and Irina Rish. Invariance principle meets information bottleneck for out-of-distribution generalization. In Neural Information Processing Systems, virtual, December 6-14, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 366, + 296, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 366, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 61, + 366, + 296, + 421 + ], + "type": "text", + "content": "[3] Tianle Chen, Mahsa Baktashmotlagh, Zijian Wang, and Mathieu Salzmann. Center-aware adversarial augmentation for single domain generalization. In Winter Conference on Applications of Computer Vision, Waikoloa, HI, USA, January 2-7, 2023. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 422, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 422, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 422, + 296, + 487 + ], + "type": "text", + "content": "[4] Seokeon Choi, Debasmit Das, Sungha Choi, Seunghan Yang, Hyunsin Park, and Sungrack Yun. Progressive random convolutions for single domain generalization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, Vancouver, BC, Canada, June 17-24, pages 10312-10322, 2023. 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 489, + 296, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 489, + 296, + 532 + ], + "spans": [ + { + "bbox": [ + 62, + 489, + 296, + 532 + ], + "type": "text", + "content": "[5] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Conference on Computer Vision and Pattern Recognition, Honolulu, HI, USA, July 21-26, 2017. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 534, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 534, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 534, + 296, + 578 + ], + "type": "text", + "content": "[6] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In International Conference on Computer Vision, Venice, Italy, October 22-29, 2017. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 580, + 296, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 580, + 296, + 623 + ], + "spans": [ + { + "bbox": [ + 62, + 580, + 296, + 623 + ], + "type": "text", + "content": "[7] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In International Conference on Computer Vision, Venice, Italy, October 22-29, 2017. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 624, + 296, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 296, + 668 + ], + "type": "text", + "content": "[8] Shahaf E. Finder, Roy Amoyal, Eran Treister, and Oren Freifeld. Wavelet convolutions for large receptive fields. In European Conference on Computer Vision, MiCo Milano, September 29 - October 4, 2024. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 296, + 713 + ], + "type": "text", + "content": "[9] Will Grathwohl, Kuan-Chieh Wang, Jorn-Henrik Jacobsen, David Duvenaud, Mohammad Norouzi, and Kevin Swersky. Your classifier is secretly an energy based model and you should treat it like one. In International Conference on" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "text", + "content": "Learning Representations, Addis Ababa, Ethiopia, April 26-30, 2020. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[10] Kehua Guo, Rui Ding, Tian Qiu, Xiangyuan Zhu, Zheng Wu, Liwei Wang, and Hui Fang. Single domain generalization via unsupervised diversity probe. In ACM International Conference on Multimedia, Ottawa, ON, Canada, October 29-November 3, pages 2101-2111, 2023. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 194 + ], + "type": "text", + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Conference on Computer Vision and Pattern Recognition, Las Vegas, NV, USA, June 27-30, pages 770-778, 2016. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 195, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 259 + ], + "type": "text", + "content": "[12] Xilin He, Jingyu Hu, Qinliang Lin, Cheng Luo, Weicheng Xie, Siyang Song, Muhammad Haris Khan, and Linlin Shen. Towards combating frequency simplicity-biased learning for domain generalization. In Conference on Neural Information Processing Systems, Vancouver, BC, Canada, December 10 - 15, 2024. 3, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "type": "text", + "content": "[13] Dan Hendrycks and Thomas G. Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, New Orleans, LA, USA, May 6-9, 2019. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 306, + 553, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 306, + 553, + 349 + ], + "spans": [ + { + "bbox": [ + 316, + 306, + 553, + 349 + ], + "type": "text", + "content": "[14] Xun Huang and Serge J. Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In International Conference on Computer Vision, Venice, Italy, October 22-29, pages 1510-1519, 2017. 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "text", + "content": "[15] Vladimir Koltchinskii. Oracle inequalities in empirical risk minimization and sparse recovery problems: École D'Été de Probabilités de Saint-Flour XXXVIII-2008. Springer Science & Business Media, 2011. 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 394, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 448 + ], + "type": "text", + "content": "[16] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In Conference on Neural Information Processing Systems, Lake Tahoe, Nevada, United States, December 3-6, 2012. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 449, + 553, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 481 + ], + "type": "text", + "content": "[17] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 1998. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 483, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 483, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 483, + 553, + 525 + ], + "type": "text", + "content": "[18] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M. Hospedales. Deeper, broader and artier domain generalization. In IEEE International Conference on Computer Vision, Venice, Italy, October 22-29, pages 5542-5550, 2017. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 526, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 580 + ], + "type": "text", + "content": "[19] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M. Hospedales. Learning to generalize: Meta-learning for domain generalization. In Proceedings of Conference on Artificial Intelligence, New Orleans, Louisiana, USA, February 2-7, 2018. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 582, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 553, + 645 + ], + "type": "text", + "content": "[20] Lei Li, Ke Gao, Juan Cao, Ziyao Huang, Yepeng Weng, Xiaoyue Mi, Zhengze Yu, Xiaoya Li, and Boyang Xia. Progressive domain expansion network for single domain generalization. In Conference on Computer Vision and Pattern Recognition, virtual, June 19-25, pages 224-233, 2021. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "text", + "content": "[21] Weitang Liu, Xiaoyun Wang, John D. Owens, and Yixuan Li. Energy-based out-of-distribution detection. In Conference on Neural Information Processing Systems, virtual, December 6-12, 2020. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[22] Krikamol Muandet, David Balduzzi, and Bernhard Scholkopf. Domain generalization via invariant feature" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15624" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 105 + ], + "type": "text", + "content": "representation. In Proceedings of International Conference on Machine Learning Atlanta, GA, USA, June 16-21, pages 10-18, 2013. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 151 + ], + "type": "text", + "content": "[23] Cheng Ouyang, Chen Chen, Surui Li, Zeju Li, Chen Qin, Wenjia Bai, and Daniel Rueckert. Causality-inspired single-source domain generalization for medical image segmentation. IEEE Transactions on Medical Imaging, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "text", + "content": "[24] Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. IEEE Transactions on Knowledge and Data Engineering, 22(10):1345-1359, 2010. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 230 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 230 + ], + "type": "text", + "content": "[25] Sinno Jialin Pan, Ivor W. Tsang, James T. Kwok, and Qiang Yang. Domain adaptation via transfer component analysis. In International Joint Conference on Artificial Intelligence, Pasadena, California, USA, July 11-17, 2009. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 232, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 294, + 285 + ], + "type": "text", + "content": "[26] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In International Conference on Computer Vision, Seoul, Korea (South), October 27 - November 2, pages 1406-1415. IEEE, 2019. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 288, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 288, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 288, + 294, + 331 + ], + "type": "text", + "content": "[27] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In Conference on Computer Vision and Pattern Recognition, Seattle, WA, USA, June 13-19, pages 12553-12562, 2020. 2, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 333, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 294, + 376 + ], + "type": "text", + "content": "[28] Alexander Robey, George J. Pappas, and Hamed Hassani. Model-based domain generalization. In Conference on Neural Information Processing Systems, virtual, December 6-14, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 378, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 378, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 378, + 294, + 421 + ], + "type": "text", + "content": "[29] Ding Rui, Kehua Guo, Xiangyuan Zhu, Zheng Wu, and Hui Fang and. Progressive diversity generation for single domain generalization. IEEE Transactions on Multimedia, 2024. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 423, + 294, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 423, + 294, + 478 + ], + "spans": [ + { + "bbox": [ + 56, + 423, + 294, + 478 + ], + "type": "text", + "content": "[30] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott E. Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Conference on Computer Vision and Pattern Recognition, Boston, MA, USA, June 7-12, 2015. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 479, + 294, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 522 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 522 + ], + "type": "text", + "content": "[31] Chris Xing Tian, Haoliang Li, Xiaofei Xie, Yang Liu, and Shiqi Wang. Neuron coverage-guided domain generalization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):1302-1311, 2023. 1, 3, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 525, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 294, + 590 + ], + "type": "text", + "content": "[32] Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C. Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. In Conference on Neural Information Processing Systems, Montreal, Canada, December 3-8, pages 5339-5349, 2018. 1, 2, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 592, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 656 + ], + "type": "text", + "content": "[33] Chaoqun Wan, Xu Shen, Yonggang Zhang, Zhiheng Yin, Xinmei Tian, Feng Gao, Jianqiang Huang, and Xian-Sheng Hua. Meta convolutional neural networks for single domain generalization. In Conference on Computer Vision and Pattern Recognition, New Orleans, LA, USA, June 18-24, pages 4672-4681, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[34] Zijian Wang, Yadan Luo, Ruihong Qiu, Zi Huang, and Mahsa Baktashmotlagh. Learning to diversify for single domain generalization. In International Conference on Computer Vision, Montreal, QC, Canada, October 10-17, pages 814-823, 2021. 3, 6, 7, 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "type": "text", + "content": "[35] Cort J Willmott and Kenji Matsuura. Advantages of the mean absolute error (mae) over the root mean square error (rmse) in assessing average model performance. Climate research, 2005. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "text", + "content": "[36] Zhenlin Xu, Deyi Liu, Junlin Yang, Colin Raffel, and Marc Niethammer. Robust and generalizable visual representation learning via random convolutions. In International Conference on Learning Representations, virtual, Austria, May 3-7, 2021. 2, 4, 5, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 173, + 553, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 226 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 226 + ], + "type": "text", + "content": "[37] Fu-En Yang, Yuan-Chia Cheng, Zu-Yun Shiau, and Yu-Chiang Frank Wang. Adversarial teacher-student representation learning for domain generalization. In Conference on Neural Information Processing Systems, December 6-14, virtual, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 228, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 270 + ], + "type": "text", + "content": "[38] Shuai Yang, Kui Yu, Fuyuan Cao, Hao Wang, and Xindong Wu. Dual-representation-based autoencoder for domain adaptation. IEEE Transactions on Cybernetics, 52(8): 7464-7477, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 272, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 316 + ], + "type": "text", + "content": "[39] Shuai Yang, Kui Yu, Fuyuan Cao, Lin Liu, Hao Wang, and Jiuyong Li. Learning causal representations for robust domain adaptation. IEEE Transactions on Knowledge and Data Engineering, 35(3):2750-2764, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 316, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 369 + ], + "type": "text", + "content": "[40] Shuai Yang, Zhen Zhang, and Lichuan Gu. Practical single domain generalization via training-time and test-time learning. In ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Barcelona, Spain, August 25-29, 2024. 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 371, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 415 + ], + "type": "text", + "content": "[41] Haotian Ye, Chuanlong Xie, Tianle Cai, Ruichen Li, Zhenguo Li, and Liwei Wang. Towards a theoretical framework of out-of-distribution generalization. In Neural Information Processing Systems, virtual, December 6-14, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 415, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 469 + ], + "type": "text", + "content": "[42] Jiajin Zhang, Hanqing Chao, Amit Dhurandhar, Pin-Yu Chen, Ali Tajer, Yangyang Xu, and Pingkun Yan. When neural networks fail to generalize? A model sensitivity perspective. In AAAI Conference on Artificial Intelligence, Washington, DC, USA, February 7-14, 2023. 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 471, + 553, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 524 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 524 + ], + "type": "text", + "content": "[43] Zuyu Zhang, Yan Li, and Byung-Seok Shin. Embracing domain gradient conflicts: Domain generalization using domain gradient equilibrium. In International Conference on Multimedia, Melbourne, VIC, Australia, October 28 - November 1, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 526, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 580 + ], + "type": "text", + "content": "[44] Long Zhao, Ting Liu, Xi Peng, and Dimitris N. Metaxas. Maximum-entropy adversarial data augmentation for improved generalization and robustness. In Advances in Neural Information Processing Systems, virtual, December 6-12, pages 1-13, 2020. 3, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "type": "text", + "content": "[45] Guangtao Zheng, Mengdi Huai, and Aidong Zhang. Advst: Revisiting data augmentations for single domain generalization. In Thirty-Eighth AAAI Conference on Artificial Intelligence, February 20-27, pages 1-9, 2024. 1, 3, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "type": "text", + "content": "[46] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4396-4415, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[47] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A comprehensive survey on transfer learning. Proceedings of the IEEE, 109(1):43-76, 2021. 1" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15625" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/St4RTrack_ Simultaneous 4D Reconstruction and Tracking in the World/8996ce18-d088-44bb-a34b-53f80b972348_content_list.json b/2025/St4RTrack_ Simultaneous 4D Reconstruction and Tracking in the World/8996ce18-d088-44bb-a34b-53f80b972348_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e3ed87309c3ba2b5b172a39c1e589c911f711692 --- /dev/null +++ b/2025/St4RTrack_ Simultaneous 4D Reconstruction and Tracking in the World/8996ce18-d088-44bb-a34b-53f80b972348_content_list.json @@ -0,0 +1,1771 @@ +[ + { + "type": "text", + "text": "St4RTrack: Simultaneous 4D Reconstruction and Tracking in the World", + "text_level": 1, + "bbox": [ + 133, + 84, + 864, + 107 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haiwen Feng $^{1,2}$ * Junyi Zhang $^{1}$ * Qianqian Wang $^{1}$ Yufei Ye $^{3}$ Pengcheng Yu $^{2}$ Michael J. Black $^{2}$ Trevor Darrell $^{1}$ Angjoo Kanazawa $^{1}$", + "bbox": [ + 156, + 133, + 841, + 172 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ UC Berkeley $^{2}$ Max Planck Institute for Intelligent Systems $^{3}$ Stanford University", + "bbox": [ + 158, + 179, + 841, + 200 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e4cfba1126e92c9e355bab8c7db34ddf1f91ab4871227ed3126373ba26c8f337.jpg", + "image_caption": [ + "Figure 1. St4RTrack: Given an RGB video capturing dynamic scenes, St4RTrack simultaneously tracks the points from the initial frame (visualized in purple) and reconstructs the geometry of the subsequent frames (in orange) in a consistent world coordinate frame. St4RTrack is a feed-forward framework that takes a pair of images as input and outputs two pointmaps in the world frame, as the visualization shown in the middle. By iteratively processing the first frame paired with each subsequent frame, St4RTrack achieves simultaneous tracking (right bottom) and reconstruction (right top) for the entire video. Interactive results on our webpage: https://St4RTrack.github.io/." + ], + "image_footnote": [], + "bbox": [ + 91, + 215, + 906, + 426 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 536, + 328, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dynamic 3D reconstruction and point tracking in videos are typically treated as separate tasks, despite their deep connection. We propose St4RTrack, a feed-forward framework that simultaneously reconstructs and tracks dynamic video content in a world coordinate frame from RGB inputs. This is achieved by predicting two appropriately defined pointmaps for a pair of frames captured at different moments. Specifically, we predict both pointmaps at the same moment, in the same world, capturing both static and dynamic scene geometry while maintaining 3D correspondences. Chaining these predictions through the video sequence with respect to a reference frame naturally computes long-range correspondences, effectively combining 3D reconstruction with 3D tracking. Unlike prior methods that rely heavily on 4D ground truth supervision, we employ a novel adaptation scheme based on a reprojection loss. We establish a new extensive benchmark for world-frame reconstruction and tracking, demonstrating the effectiveness and efficiency of our unified, data-driven framework. Our code, model, and benchmark will be released.", + "bbox": [ + 88, + 565, + 485, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 536, + 645, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "When asked about the three most important problems in computer vision, Takeo Kanade replied, \"Correspondence, Correspondence, Correspondence!\" This remark is especially pertinent to multi-view 3D reconstruction, where 3D geometry and 2D correspondence are two sides of the same coin; that is, a 3D point in the physical world naturally brings 2D correspondence across its projections in different views, and conversely, corresponded 2D points across views reconstruct the same 3D point after triangulation. This synergy between 3D geometry and 2D correspondence has long formed the foundation of multi-view geometry [17]. However, when the scene becomes dynamic, this synergy appears to vanish, as existing methods—particularly the recent data-driven ones—tend to treat dynamic reconstruction [22, 58, 67] and correspondence [24, 38, 51] as separate, disconnected tasks. We argue that this is a missed opportunity; the synergy between 3D reconstruction and 2D correspondence is not lost in dynamic scenes—it simply requires an additional element: understanding how the scene content evolves over time. This evolution is captured by 3D motion estimation (i.e., dense 3D point tracking), which, when computed across a sequence, effectively ex", + "bbox": [ + 511, + 568, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution, listed alphabetically.", + "bbox": [ + 114, + 886, + 336, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8503", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "plains scene motion. Once tracking accounts for scene dynamics, the problem effectively reduces to the rigid case, where the natural interplay between reconstruction and correspondence can once again be leveraged.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose St4RTrack, a learning framework that unifies reconstruction and tracking directly from RGB video frames. St4RTrack simultaneously reconstructs and tracks dynamic video content in a single consistent world, achieving world-frame 3D tracking, as demonstrated in Fig. 1. Tracking in the world frame decouples the scene and the camera motion, essential for domains where both the camera and content are in motion. Our approach also reconstructs the 3D geometry of observed image content for both static and dynamic portions of the scene. St4RTrack directly predicts their reconstruction and tracking in the world without requiring an additional alignment stage.", + "bbox": [ + 88, + 154, + 482, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our key insight stems from the observation that a feedforward static 3D reconstruction method, DUSt3R [61], can be adapted to dynamic scenes simply by changing the pointmaps' annotation [67]. Building on this, we reexamine the pointmaps definition in the 4D scenario and opt to simply redefine its geometric interpretation for both reconstruction and tracking, as illustrated in Fig. 2. Concretely, we achieve it by predicting two pointmaps at the same timestamp and in the same world from a pair of image frames depicting two different timestamps. More specific, given images $(\\mathbf{I}_i,\\mathbf{I}_j)$ , both pointmaps are predicted in the coordinate frame of $\\mathbf{I}_i$ , but at the time specified by $\\mathbf{I}_j$ . Our method is realized through a feed-forward network comprising of two branches: the reconstruction branch reconstructs the content of $\\mathbf{I}_j$ in the $\\mathbf{I}_i$ coordinate frame; and the tracking branch, which reconstructs the content of $\\mathbf{I}_i$ in the $\\mathbf{I}_i$ (its own) coordinate frame, but at the time indicated by $\\mathbf{I}_j$ . Essentially, the tracking branch predicts how the scene content in $\\mathbf{I}_i$ evolves to match the moment captured in $\\mathbf{I}_j$ . This is enabled through a DUSt3R-like dual cross-attention mechanism, where the tracking branch relies on the reconstruction branch to decide how to move points. This minimal change proves sufficient for unifying both dynamic reconstruction and 3D point tracking in the world coordinate system.", + "bbox": [ + 91, + 339, + 482, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, unlike existing methods [58, 61, 67] that rely solely on 4D supervision, our approach unlocks 4D reconstruction training on in-the-wild videos via reprojection loss without 4D supervision. This is possible because St4RTrack simultaneously establishes camera parameters, 3D geometry and motion. Specifically, based on the outputs of the reconstruction branch, the camera parameters for $\\mathbf{I}_j$ can be differentiably computed via PnP. Using these cameras, the pointmap of $\\mathbf{I}_i$ is projected into the $j$ -th frame, enabling training with a reprojection loss that leverages 2D correspondences and monocular depth predictions from off-the-shelf approaches [23, 59]. Consequently, the monocular supervisions facilitate effective test-time adaptation of", + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "St4RTrack to in-the-wild videos, which can differ substantially from the synthetic data used during pretraining.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While prior 3D point tracking benchmarks focus on camera coordinate frames [64], our approach enables worldframe 3D tracking. To evaluate this capability, we establish a novel benchmark, WorldTrack, for both tracking and reconstruction in the world coordinate system. We find that our unified method outperforms the strong baselines that combine several pieces on each individual task. Furthermore, we show that our feedforward results can be improved via test-time adaptation. We believe this is a step towards a unified task-agnostic 4D perception system that can be trained on a large-scale video. Our code, model, and benchmark will be released.", + "bbox": [ + 511, + 123, + 903, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 321, + 663, + 338 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Camera Estimation and Scene Reconstruction. Jointly estimating camera motion and scene geometry has been studied for decades, often in the context of Structure from Motion (SfM) [1, 47, 48, 56] or Simultaneous Localization and Mapping (SLAM) [7, 10, 37, 52, 60]. However, these methods are primarily designed for static scenes and typically do not model dynamic scene content. Recent advances in learning-based monocular and video depth estimation methods [2, 42, 44, 65] have opened new opportunities to reconstruct dynamic scenes. Notably, R-CVD [25], CasualSAM [68] and MegaSAM [33] jointly optimize camera parameters and per-frame dense depth maps leveraging monocular depth priors, producing consistent depth estimates for dynamic objects and accurate camera parameters even in challenging cases with minimal camera parallax. Another notable recent method, DUSt3R [61], introduces a two-pointmap representation that enables joint estimation of camera motion and scene geometry of a pair of images. While DUSt3R itself primarily focuses on reconstructing static scenes, follow-up effort such as MonST3R [67], demonstrate that this formulation can also effectively handle dynamic scenes with minimal modification on supervisions. Despite these advances, none of the aforementioned methods explicitly estimate 3D scene motion, meaning they do not track the movement of individual 3D points over time. In contrast, our method simultaneously performs joint reconstruction and tracking for dynamic scenes.", + "bbox": [ + 511, + 349, + 906, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2D/3D Point Tracking. Tracking pixel motion over time is a fundamental problem in computer vision. Optical/Scene flow methods [3, 18, 19, 35, 50, 51, 53, 55] produce dense 2D/3D motion vectors but are inherently short-ranged, struggling with large displacements and occlusions. While long-range point tracking [45, 46] has been studied for decades, it has recently been revitalized via supervised learning [8, 9, 16, 23, 24], enabling more robust tracking over extended time periods and overcoming these limita-", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8504", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/41aa5171dca6774fdbdd14db755bf094b7fc028fe61fc026cd8563dd1c1ac99d.jpg", + "image_caption": [ + "Figure 2. Pointmap Comparison of MonST3R and St4RTrack. Given two input frames, MonST3R handles dynamic scenes by reconstructing both pointmaps in their own timestamp. St4RTrack predicts where the points in the first frame move in the second frame, and reconstructs the geometry of the second frame. More details of the representation definition are introduced in Sec. 3.1." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 186, + 220 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f1e32e9c6ab572e2e4767e4091e3f9e9987ef8eb4f791ef2a30f977db9cd9e45.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 205, + 90, + 331, + 220 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f645ec427d3a3c046c2467d50953e0b115cd37cb432f1f2b22444b492e6249ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 90, + 475, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tions. However, these methods still produce only 2D pixel trajectories. More recently, several works [38, 64] achieve 3D tracking by lifting points into 3D space using monocular depth priors and performing tracking in 3D. While closely related to our approach, these methods still operate in the camera frame space, meaning they lack camera motion estimation and do not explicitly separate scene motion from camera motion. In contrast, our method jointly estimates disentangled camera and scene motion, enabling world-space tracking for a more complete understanding of 3D scene dynamics.", + "bbox": [ + 89, + 349, + 483, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Joint Dynamic Reconstruction and Tracking. While traditional non-rigid SfMs have studied joint tracking and reconstruction [4, 6, 12, 39, 70] from 2D correspondences, jointly optimizing them from raw videos is highly challenging and typically requires multi-view synchronized videos as input [13, 30, 36, 41, 63]. With recent advances in neural rendering and data-driven geometric priors, reconstructing and rendering dynamic scenes from monocular videos became possible. However, as Gao et al. [14] point out, many methods [43, 66] focus on \"teleporting\" input data, which are effectively multi-view and not representative of real-world videos. In addition, since the main focus is view synthesis, motion estimation serves a secondary role in facilitating information fusion between neighboring frames [31, 32]. More recently, several works [27, 34, 57] focus on jointly recovering camera parameters, persistent scene geometry, and long-range 3D tracks from single, causally captured videos. However, these methods take off-the-shelf priors as given and design per-video optimization techniques that optimize a representation from scratch. Most recently, Stereo4D [20]—a concurrent effort to our work—proposed a pipeline for crafting a real-world 4D tracking dataset using internet stereo videos, enabling the supervised regression of 3D trajectories and geometries between frames. In contrast, we propose a feed-forward", + "bbox": [ + 89, + 523, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "method that simultaneously performs reconstruction and tracking, while the same architecture also supports test-time adaptation on unlabeled videos to approach the high quality of optimization-based methods.", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Simultaneous Reconstruction and Tracking", + "text_level": 1, + "bbox": [ + 511, + 167, + 901, + 185 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present a framework that simultaneously reconstructs and tracks dynamic video content in 3D within a single consistent world coordinate frame. The core idea is simple yet powerful: reconstructing and tracking can both be achieved by predicting two appropriately defined pointmaps, where both pointmaps reconstruct the scene content observed in each image at the same timestamp in a consistent coordinate system. This enables simultaneous reconstruction of both dynamic and static contents, while tracking across a sequence of image frames in a video. Since all geometry, camera, and motion (i.e. 3D correspondence over time) can be derived from the representation, it can be adapted to videos without any explicit 4D supervision. Below, we discuss the main insight, how St4RTrack compares to prior works. Then, we discuss the details of the model and how it can be trained and adapted to videos.", + "bbox": [ + 511, + 193, + 906, + 435 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Unified 4D Representation of St4RTrack", + "text_level": 1, + "bbox": [ + 511, + 446, + 859, + 463 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given two images $\\mathbf{I}_i, \\mathbf{I}_j$ with dynamic content (see Figure 2), how can one devise a single feedforward approach that simultaneously performs reconstruction and tracking? We argue that the underlying representation must (1) capture camera motion to establish a world coordinate frame, (2) reconstruct the 3D geometry of all observed points, and (3) estimate 3D motion that maintain explicit correspondence over time. In this work, we show that just two properly defined pointmaps suffice to fulfill these requirements.", + "bbox": [ + 511, + 469, + 905, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Time-Dependent Pointmap. A pointmap representation assumes that each pixel in an image $\\mathbf{I}$ of shape $H\\times W$ is associated with a corresponding 3D point, forming a pointmap $\\mathbf{X}\\in \\mathbb{R}^{H\\times W\\times 3}$ . For the case of static scenes, DUSt3R [61] only considers two factors of pointmaps—(1) the source frame of the content and (2) the camera coordinate system in which the points are expressed. However, this definition is insufficient for modeling the dynamic scenario of monocular video. To address this, we introduce a previously overlooked yet decisive factor: time.", + "bbox": [ + 511, + 612, + 905, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, we define a time-dependent pointmap that encodes the 3D positions of the scene points in a chosen (world) coordinate system at a specific timestamp. For clarity, we denote this representation as ${}^a\\mathbf{X}_t^b$ , which denotes the 3D pointmap of physical content from frame $b$ , at time $t$ , expressed in the coordinate system established by frame $a$ . For example, ${}^i\\mathbf{X}_j^i$ represents the geometry originally seen in frame $i$ in frame $i$ 's own coordinate system, but described at timestamp $j$ . The time-dependency is achieved by construc", + "bbox": [ + 511, + 763, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8505", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/838ff1e8eeefd1c74338805156461b5e20fed63826feab1939d84f794bab6951.jpg", + "image_caption": [ + "Figure 3. Overview of St4RTrack. Given frame 1 and frame $j$ as input, the tracking branch outputs ${}^{1}\\mathbf{X}_{j}^{1}$ , the pointmap that corresponds to observed content of the first frame at timestep $j$ in its own camera coordinate (i.e. world coordinate); the reconstruction branch outputs ${}^{1}\\mathbf{X}_{j}^{j}$ , the pointmap of the content in frame $j$ at its own timestamp in the world coordinate. To adapt to new videos without any 4D labels, the camera is computed via differentiable PnP from the pointmap, enabling reprojected supervision signals (e.g., 2D trajectories and monocular depth). We finetune both branches during training (Sec. 3.2) with synthetic data, and when adapting to a new video (Sec. 3.3), only the tracking branch is fine-tuned using these reprojected supervision signals." + ], + "image_footnote": [], + "bbox": [ + 106, + 66, + 883, + 290 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tion without explicit timestamp conditioning, as described in the next section.", + "bbox": [ + 89, + 402, + 482, + 433 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unified 4D Modeling. St4RTrack learns a function $f$ that maps two images $\\mathbf{I}_i, \\mathbf{I}_j$ , captured at timestamp $i$ and $j$ , into two pointmaps:", + "bbox": [ + 89, + 440, + 483, + 484 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\theta} \\left(\\mathbf {I} _ {i}, \\mathbf {I} _ {j}\\right) = ^ {i} \\mathbf {X} _ {j} ^ {i}, ^ {i} \\mathbf {X} _ {j} ^ {j}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 209, + 497, + 482, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The second image $\\mathbf{I}_j$ is reconstructed as the pointmap ${}^i\\mathbf{X}_j^j$ in the first image $\\mathbf{I}_i$ 's coordinate frame. Meanwhile, it predicts ${}^i\\mathbf{X}_j^i$ , representing the 3D motion of how the content from the first image $\\mathbf{I}_i$ moves at timestamp $j$ . Thus, both geometry and motion (tracking) are estimated from this unified prediction.", + "bbox": [ + 89, + 530, + 483, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To handle a full video consisting of $T$ frames, we perform tracking and reconstruction by always selecting the first frame as the anchor frame $\\mathbf{I}_i$ . Each subsequent frame $\\mathbf{I}_j$ is then paired with this initial frame, ensuring that every new frame is consistently aligned to the coordinate system of the first frame. Specifically, $\\{f(\\mathbf{I}_1,\\mathbf{I}_1),f(\\mathbf{I}_1,\\mathbf{I}_2),\\ldots ,f(\\mathbf{I}_1,\\mathbf{I}_T)\\}$ are computed in the same reference, $\\mathbf{I}_1$ , which naturally serves as the world coordinate frame. Thus, world-frame 3D tracking is achieved by explicitly following how points observed in $\\mathbf{I}_1$ are placed throughout the sequence, $\\{\\mathbf{^1X_1^1},\\mathbf{^1X_2^1},\\dots,\\mathbf{^1X_T^1}\\}$ , while the world frame dynamic reconstruction is obtained by the paired geometry estimation per-frame, $\\{\\mathbf{^1X_1^1},\\mathbf{^1X_2^2},\\dots,\\mathbf{^1X_T^T}\\}$ .", + "bbox": [ + 89, + 622, + 483, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Relation to prior works. Our formulation of the 4D modeling generalizes prior works in a unified framework. DUSt3R reconstructs and establishes correspondences but is limited to rigid scenes, as correspondence and recon", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "struction are dual tasks for static scenes. With this perspective, one can see that if there is no dynamic component (i.e. frozen moment in time or rigid scenes), our formulation is equivalent to DUSt3R, where both images share the same timestamp $t = i$ :", + "bbox": [ + 511, + 402, + 906, + 479 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\theta} (\\mathbf {I} _ {i}, \\mathbf {I} _ {j}) = ^ {i} \\mathbf {X} _ {i} ^ {i}, ^ {i} \\mathbf {X} _ {i} ^ {j}. \\qquad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 491, + 906, + 511 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In a static world, 3D reconstruction from two pointmaps inherently yields the correspondences between them, allowing synergy to arise naturally. However, when objects or the scene are in motion, the dynamic component appears differently in different frames, it becomes crucial to account for 3D scene motion to preserve this synergy. St4RTrack addresses this challenge by predicting the 3D content from the first image at future timestamps.", + "bbox": [ + 511, + 523, + 905, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the same framework, we see that MonST3R, the dynamic follow-up of DUSt3R can be expressed as such:", + "bbox": [ + 511, + 645, + 905, + 675 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\theta} (\\mathbf {I} _ {i}, \\mathbf {I} _ {j}) = ^ {i} \\mathbf {X} _ {i} ^ {i}, ^ {i} \\mathbf {X} _ {j} ^ {j}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 686, + 906, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where each image's 3D geometry is reconstructed in its timestamp, such that the dynamic contents separately align with their frame inputs. While it's sufficient for obtaining dynamic scene geometry, there is no temporal correspondence being established, as illustrated in Fig. 2. Furthermore, both DUSt3R and MonST3R compute the pairwise graphs and perform global alignment.", + "bbox": [ + 511, + 718, + 905, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since we always designate the first frame as the reference for tracking, the world coordinates are consistently established by the first frame. For simplicity, we omit the explicit notation of the world coordinate in subsequent equations and paragraphs, i.e., $\\mathbf{X}_j^i \\coloneqq {}^i\\mathbf{X}_j^i$", + "bbox": [ + 511, + 825, + 905, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8506", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Joint Learning of Tracking and Reconstruction", + "text_level": 1, + "bbox": [ + 89, + 90, + 480, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we describe how our framework implements equation 1 within a pair-wise framework as DUSt3R. For each pair of frames, $\\mathbf{I}_1$ and $\\mathbf{I}_j$ , we first encode them into token representations using a ViT encoder, then process these tokens through a siamese transformer decoder. The decoder sequentially applies self-attention (allowing tokens within each frame to interact), followed by cross-attention (enabling tokens from one frame to attend to tokens in the other), and finally passes the tokens through an MLP. This continuous information flow between the two branches is crucial for generating spatial-aligned 3D pointmaps in a shared coordinate system, as illustrated in Fig. 3.", + "bbox": [ + 89, + 113, + 483, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our siamese architecture processes two input views concurrently and generates two 3D pointmaps that are expressed in a common reference frame established by the first view. Although the two branches share the same architectural structure, they serve distinct purposes:", + "bbox": [ + 89, + 296, + 483, + 371 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Tracking branch predicts the pointmap $\\mathbf{X}_j^1$ , which represents the geometry of the first frame at timestamp $j$ in the first frame's coordinates (i.e., the world coordinates).", + "- Reconstruction branch predicts the pointmap $\\mathbf{X}_j^j$ , which represents the geometry of frame $j$ at its own timestamp, also expressed in the first frame's camera coordinates." + ], + "bbox": [ + 89, + 372, + 482, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since this architecture is exactly the same as proposed by DUSt3R and subsequently adopted by MonST3R, with the only difference being the output paired pointmaps (Eq.1-3), our network can be initialized with pretrained 3D knowledge from either DUSt3R or MonST3R.", + "bbox": [ + 89, + 464, + 483, + 540 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pretraining with 4D Synthetic Data. Our proposed representation requires specialized supervision for the Tracking Branch—namely, ensuring that the pointmap from the first frame is correctly positioned in the world across all frames. Achieving this necessitates complete 4D information of the dynamic scene. Therefore, we leverage existing 4D synthetic datasets [22, 69] that provide both the 3D geometry and motion of the rendered content. Specifically, for each dataset, we use the scene mesh vertices (expressed in world coordinates) to provide sparse, masked supervision for the Tracking Branch pointmap representation, and employ per-frame depth maps and camera ground-truth to supervise the Reconstruction Branch. For this fully supervised training process, we use the objectives from DUSt3R. We initialize our dual-branch transformer with weights from MAST3R [29], a DUSt3R variant that has been adapted for 2D correspondence learning. Additional details regarding the 4D synthetic datasets are provided in Sec. 4.1.", + "bbox": [ + 89, + 547, + 483, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Adapt to Any Video without 4D Label", + "text_level": 1, + "bbox": [ + 89, + 832, + 418, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While the synthetic datasets are small-scale and unrealistic, they are sufficient for our network to learn the newly proposed representations. However, fully supervised training", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "on these datasets presents two key limitations: 1) The 4D synthetic data is limited in scale and does not encompass the full range of motion and geometry present in real-world dynamic scenes; 2) Our proposed pointmap representation requires the capability to freely move the pointmap within the world coordinates—a departure from conventional pixel-aligned geometry predictions, making small-scale training insufficient for achieving fine-grained predictions. These limitations motivate us to further leverage the 3D geometry and motion inherent in the St4RTrack framework to perform domain adaptation on any video without 4D labels. Specifically, we first show how we can derive camera parameters differentially, and with which we can design reprojected 2D trajectory loss and monocular depth loss to supervise the network.", + "bbox": [ + 511, + 90, + 903, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Solving Camera Parameters. The intrinsic matrix $\\mathbf{K}$ is first estimated from the tracking branch's first-frame pointmap prediction, following DUSt3R [61]. In this process, the principal point is assumed to be centered, and pixels are treated as square. The focal length is assumed static across frames and estimated using a fast iterative solver based on the Weiszfeld algorithm [62]. Next, the extrinsic parameters $\\mathbf{P}^j = [\\mathbf{R}^j | \\mathbf{T}^j]$ for each frame $j$ are derived using the \"reconstruction\" pointmap $\\mathbf{X}_j^j$ . Specifically, each pixel $\\mathbf{x}^{j,n}$ in frame $j$ is associated with a 3D coordinate $\\mathbf{X}_j^{j,n}$ in the shared world coordinate system (established by the first camera), thus forming 2D-to-3D correspondences. We could then solve for $\\mathbf{R}^j$ and $\\mathbf{T}^j$ via a Perspective- $n$ -Points (PnP) [28] solver with RANSAC [11] for outlier rejection:", + "bbox": [ + 511, + 324, + 906, + 551 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {R} ^ {j}, \\mathbf {T} ^ {j} = \\underset {\\mathbf {R}, \\mathbf {T}} {\\operatorname {a r g m i n}} \\sum_ {n \\in \\mathcal {I} _ {j}} \\left\\| \\mathbf {x} ^ {j, n} - \\pi \\left(\\mathbf {K} \\left(\\mathbf {R} \\mathbf {X} _ {j} ^ {j, n} + \\mathbf {T}\\right)\\right) \\right\\| ^ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 565, + 903, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pi (\\cdot)$ is the projection $(x,y,z)\\to (x / z,y / z)$ .", + "bbox": [ + 511, + 616, + 856, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For differentiability, we adopt a derivative-based Gauss-Newton solver following [5], ensuring that gradients from the reprojection loss can adjust both the camera pose and the 3D pointmaps. Further details are provided in ??", + "bbox": [ + 511, + 633, + 903, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reprojection Loss. With the camera pose of frame $j$ derived, the tracking pointmap $\\mathbf{X}_j^1$ and reconstruction pointmap $\\mathbf{X}_j^j$ can be transformed from the world coordinate system into the camera coordinate system of frame $j$ . This transformation enables self-supervised training by enforcing two types of consistency: (1) 2D correspondence consistency, which aligns the projected 2D tracks (from $\\mathbf{X}_j^1$ ) with the pseudo-ground truth tracking from CoTracker [23, 24], and (2) geometric consistency, which aligns the scale-invariant depth (from $\\mathbf{X}_j^j$ ) with the pseudoground truth monocular depth from MoGe [59].", + "bbox": [ + 511, + 700, + 905, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "More specifically, given the estimated camera pose $(\\mathbf{R}^j,\\mathbf{T}^j)$ from frame $j$ and the tracking pointmap $\\mathbf{X}_j^1$ , we", + "bbox": [ + 511, + 869, + 903, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8507", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "reproject these 3D points into the image plane of frame $j$ :", + "bbox": [ + 89, + 90, + 472, + 107 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} ^ {j, n} = \\pi \\left(\\mathbf {K} \\left(\\mathbf {R} ^ {j} \\mathbf {X} _ {j} ^ {1, n} + \\mathbf {T} ^ {j}\\right)\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 114, + 482, + 135 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "These reprojected points serve as the predicted 2D tracks and are compared with the pseudo-ground truth tracking points $\\mathbf{x}_{\\mathrm{trk}}^{j,n}$ from CoTracker3 [23].", + "bbox": [ + 89, + 143, + 482, + 188 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To mitigate minor focal inaccuracies that may induce scaling shifts, the reprojection loss is computed in a scale-invariant manner. Let $\\mathbf{p}_n = \\hat{\\mathbf{x}}^{j,n}$ and $\\mathbf{g}_n = \\mathbf{x}_{\\mathrm{trk}}^{j,n}$ for $n = 1,\\dots ,N$ , and denote the image center by $c$ . The scale factor and adjusted predictions are computed together as:", + "bbox": [ + 89, + 188, + 483, + 265 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {p}} _ {n} = \\left(\\mathbf {p} _ {n} - c\\right) s + c, \\quad s = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\frac {\\left\\| \\mathbf {g} _ {n} - c \\right\\| _ {2}}{\\left\\| \\mathbf {p} _ {n} - c \\right\\| _ {2}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 273, + 483, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Then, the scale-invariant 2D reprojection loss is defined as", + "bbox": [ + 89, + 324, + 478, + 339 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {t r a j}} = \\frac {1}{N} \\sum_ {n \\in \\mathcal {I} _ {j}} \\| \\hat {\\mathbf {p}} _ {n} - \\mathbf {g} _ {n} \\| ^ {2}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 348, + 482, + 386 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Similarly, to enforce geometric consistency with the mono-depth predictions from MoGe [59], we use the reconstruction pointmap $\\mathbf{X}_j^j$ that transformed to frame $j$ 's camera coordinate. The depth of each transformed 3D point is", + "bbox": [ + 89, + 395, + 483, + 455 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nz _ {\\text {p r o j}} ^ {j, n} = \\left(\\mathbf {R} ^ {j} \\mathbf {X} _ {j} ^ {j, n} + \\mathbf {T} ^ {j}\\right) _ {z}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 465, + 482, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $(\\cdot)_z$ denotes the third (depth) component. Denote the corresponding mono-depth pseudo ground-truth by $z_{\\mathrm{mono}}^{j,n}$ . After solving for an optimal scaling factor $\\alpha^*$ to align the two depth maps, the scale-invariant mono-depth loss is defined as", + "bbox": [ + 89, + 500, + 483, + 575 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {d e p t h}} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\left(\\alpha^ {*} z _ {\\text {p r o j}} ^ {j, n} - z _ {\\text {m o n o}} ^ {j, n}\\right) ^ {2}, \\quad \\alpha^ {*} = \\frac {\\sum_ {i} z _ {\\text {p r o j}} ^ {j , n} z _ {\\text {m o n o}} ^ {j , n}}{\\sum_ {i} \\left(z _ {\\text {p r o j}} ^ {j , n}\\right) ^ {2}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 583, + 483, + 645 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D Self-Consistency. Beyond the 2D reprojection losses, we introduce a 3D self-consistency term that aligns the tracking pointmap $\\mathbf{X}_j^1$ with the reconstruction pointmap $\\mathbf{X}_j^j$ . Let $\\mathcal{I}_1^\\prime$ be the set of points in frame 1 that remain visible in frame $j$ , and for each point $n \\in \\mathcal{I}_1^\\prime$ , denote its corresponding point (provided by CoTracker) in frame $j$ by $n^\\prime$ . We then penalize the distance between their predicted 3D positions:", + "bbox": [ + 89, + 651, + 483, + 775 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l i g n}} = \\sum_ {n \\in \\mathcal {I} _ {1} ^ {\\prime}} \\left\\| \\mathbf {X} _ {j} ^ {1, n} - \\mathbf {X} _ {j} ^ {j, n ^ {\\prime}} \\right\\| ^ {2}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 784, + 483, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Minimizing $\\mathcal{L}_{\\mathrm{align}}$ ensures that both branches produce consistent geometry in the same timestamp.", + "bbox": [ + 89, + 829, + 482, + 859 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The overall self-supervision loss is given by:", + "bbox": [ + 109, + 859, + 406, + 875 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e p r o j}} = \\mathcal {L} _ {\\text {t r a j}} + \\lambda_ {1} \\mathcal {L} _ {\\text {d e p t h}} + \\lambda_ {2} \\mathcal {L} _ {\\text {a l i g n}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 885, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with $\\lambda_{1}$ and $\\lambda_{2}$ being the weighting factors. Minimizing $\\mathcal{L}_{\\mathrm{proj}}$ aligns the projected 3D structure with the 2D tracking and monocular depth cues and their 3D self-consistency, enabling unsupervised, target-specific refinement of the 3D geometry and point tracking.", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Test-Time Adaptation. To address the gap between synthetic pretraining and real-world data, we incorporate reprojection-based losses to enable test-time adaptation in St4RTrack. Our framework supports two adaptation paradigms: (1) Instance-level adaptation. During testing, we update St4RTrack on new sequences using only the aforementioned reprojected losses while freezing the reconstruction branch. We freeze these weights because both the 2D trajectories and depth are computed under a purely monocular setting, which does not provide view-alignment supervision. This approach preserves the view-alignment capability captured during pretraining. Moreover, since the pretrained network already encodes strong task-relevant representations, this sequence-specific optimization converges rapidly compared to test-time optimization methods that start from scratch. (2) Domain-level adaptation. Unlike tabula-rasa approaches such as [27, 57], which require full re-optimization for each new sequence, St4RTrack is an end-to-end learning framework that enables test-time adaptation to align the model from its pretraining data distribution to the target video domain. After adapting to a sparse set of target-domain samples, St4RTrack can directly perform simultaneous reconstruction and tracking on new sequences from the adapted domain without additional optimization.", + "bbox": [ + 511, + 172, + 906, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 564, + 643, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "St4RTrack performs both dense 3D point tracking and dynamic reconstruction in a unified world coordinate system, all within a single inference. In the following section, we first evaluate our method on 3D tracking and dynamic reconstruction separately, and then present the joint results. We also introduce a new benchmark, WorldTrack, for 3D tracking in world coordinates, which is not directly covered by previous methods.", + "bbox": [ + 511, + 589, + 905, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Details", + "text_level": 1, + "bbox": [ + 511, + 720, + 712, + 736 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. For fully supervised training, we use three synthetic datasets: Point Odyssey (PO) [69], Dynamic Replica (DR) [22], and Kubric [15]. All three datasets contain scene and camera motion and provide mesh vertex positions as ground-truth 3D point trajectories. We randomly sample 24 frames with a stride of $1 \\sim 6$ for each sample sequence. We also filter out less semantically meaningful sequences in PO, resulting in a total of 9.8k sequences for PO, 8.5k for DR, and 5.7k for Kubric dataset.", + "bbox": [ + 511, + 743, + 905, + 878 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training and Inference. During training, we sample 600", + "bbox": [ + 513, + 885, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8508", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/de09abb813879a15dd6b5734fd3003f82d934644f16cd34c0c7f2d13dd0867bd.jpg", + "table_caption": [ + "Table 1. World Coordinate 3D Point Tracking. We report the performance of average points under distance $(\\mathrm{APD}_{3\\mathrm{D}})$ after global median alignment. We evaluate the accuracy of both all points and dynamic points. The best results are bold. See ?? for more results." + ], + "table_footnote": [], + "table_body": "
CategoryMethodsAll PointsDynamic Points
PODRADTPStudioPODRADTPStudio
CombinationalSpaTracker+RANSAC-Procrustes44.0355.0150.8752.0553.7758.5866.4952.05
SpaTracker+MonST3R47.6555.4951.9550.1658.6159.2169.9450.16
Feed-forwardMonST3R33.4758.0674.3551.3239.3651.8667.9251.32
SpaTracker38.5454.8545.6562.5951.2058.6567.6562.59
St4RTrack (Ours)67.9573.7476.0169.6768.7268.1375.3469.67
", + "bbox": [ + 93, + 95, + 906, + 250 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6b75972565ee627880afcc4e04852f0e4737671303d93a324e52606a4dbf6902.jpg", + "table_caption": [ + "Table 2. World Coordinate 3D Reconstruction. We report performance on both Point Odyssey (PO) and TUM-Dynamics after global median scaling. The best results are in bold." + ], + "table_footnote": [], + "table_body": "
CategoryMethodsPoint OdysseyTUM-Dynamics
EPE↓APD↑EPE↓APD↑
w/ Global Align.DUSt3R+GA0.608843.900.314770.49
MASt3R+GA0.403060.440.518668.38
MonST3R+GA0.262972.310.342963.87
Feed-forwardDUSt3R0.638645.790.289172.26
MASt3R0.464456.900.551066.22
MonST3R0.304468.250.364661.38
St4RTrack (Ours)0.240678.730.185483.42
", + "bbox": [ + 93, + 306, + 483, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sequences from each dataset per epoch. We use the AdamW optimizer with a learning rate of $5 \\times 10^{-5}$ and a mini-batch size of 1 per GPU. The model is trained for 50 epochs on 4 A100 GPUs, which takes about one day. For test-time adaptation, we run 500 optimization steps on a single sequence, taking approximately 5 minutes on 4 A100 GPUs. At inference time, the model runs at 30 FPS on an RTX 4090. Although the model is trained on sequences of 24 frames, our pair-wise approach allows it to operate on arbitrarily long videos during inference. Refer to ?? for more details regarding test-time adaptation.", + "bbox": [ + 88, + 491, + 482, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. 3D Tracking in World Coordinates", + "text_level": 1, + "bbox": [ + 89, + 666, + 393, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets. 3D tracking in world coordinates is a critical aspect that has been largely overlooked by previous benchmarks [26], which are limited to camera coordinate systems. To address this limitation, we propose a new benchmark for 3D tracking in world coordinates. Our benchmark leverages two real-world datasets—Aerial Digital Twin (ADT) [40] and Panoptic Studio [21]—by converting the TAPVid-3D [26] sequences to world coordinates using paired extrinsic parameters. However, it is noteworthy that the limitations of these datasets: the ADT sequences exhibit minimal scene motion, while the Panoptic Studio lacks camera motion. To overcome these shortcomings, we include two additional synthetic test sets from Point Odyssey and Dynamic Replica, which have both scene and", + "bbox": [ + 88, + 688, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "camera motion. In total, our benchmark comprises four datasets, each containing 50 sequences of 64 frames. Refer to ?? for more details and visual examples of the benchmark.", + "bbox": [ + 511, + 262, + 906, + 321 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Metrics. We follow the TAPVid-3D protocol and use the Average percent of Points within Delta (APD) metric for evaluation. Specifically, we first align the predicted 3D point trajectories with the ground truth by normalizing them with their global median. We then compute the prediction error and measure the percentage of points whose error falls below a given threshold $\\delta_{3D}$ (with $\\delta_{3D} \\in \\{0.1\\mathrm{m}, 0.3\\mathrm{m}, 0.5\\mathrm{m}, 1.0\\mathrm{m}\\}$ ) over the first 64 frames. Let $\\hat{\\mathbf{P}}_t^i$ denote the $i$ -th predicted point at time $t$ and $\\mathbf{P}_t^i$ denote its corresponding ground-truth location. The resulting $\\mathrm{APD}_{3D}$ is then computed as follows:", + "bbox": [ + 511, + 328, + 908, + 496 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {A P D} _ {\\mathrm {3 D}} \\equiv \\sum_ {i, t} \\mathbb {1} \\left(\\left\\| \\hat {\\mathbf {P}} _ {t} ^ {i} - \\mathbf {P} _ {t} ^ {i} \\right\\| < \\delta_ {3 D}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 507, + 906, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbb{1}(\\cdot)$ is the indicator function and $\\| \\cdot \\|$ denotes the Euclidean norm.", + "bbox": [ + 511, + 554, + 905, + 584 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. Since no existing work explicitly performs 3D tracking in world coordinates, for our feedforward baselines, we compare against the camera coordinate 3D tracking method SpatialTracker [64] and a dynamic 3D reconstruction method MonST3R [67] (as a non-tracking baseline). In addition, we implement two combinational baselines for world coordinate 3D tracking. The first baseline applies Procrustes alignment [54] and RANSAC [11] to the camera coordinate 3D tracks predicted by SpatialTracker to offset the camera motion. The second baseline leverages the camera poses predicted by the dynamic SLAM method MonST3R to compensate for camera motion.", + "bbox": [ + 511, + 590, + 908, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. As shown in Tab. 1, we achieve state-of-the-art performance, with test-time adaptation proving particularly beneficial for dynamic points. Notably, on the Panoptic Studio dataset, which is captured with a fixed camera and can be considered a fair benchmark for camera coordinate tracking methods, our approach still outperforms SpatialTracker [64]. It is worth noting that although our model is trained on sequences of 24 frames, it generalizes well to", + "bbox": [ + 511, + 779, + 910, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8509", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8326b61c09a351b48947af958d39883c51ff9ad97e5979afb1933fd2311f71ab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 112, + 74, + 204, + 114 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eaa581cf9c9b9f9077f85026ace1801dae1ca7a68f8b5552801c1cd46f5f914c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 112, + 116, + 204, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8c8d8982a2cb365c4df587b6df14e7e89c3425d4a2ac59d4c3654fce0ec6fe35.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 63, + 433, + 170 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4ee0e24f3babbf21a88a367b57e343135347e3c2bcd5e96b72600f1b1efdc5d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 63, + 666, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e78831837f016db6990ccc530b742ff54534d0ac2c0011f1a9b2ae972cd2b883.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 61, + 882, + 165 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ff00bfcb118553f05b5cb325eb8bcb24a0aff8700cdb55007cd24367e7adf729.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 112, + 181, + 204, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6e9c87060f4f471cb1074682ba5e4abef0d8897e338fc7a9f0a49ab24e35dcec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 112, + 224, + 202, + 266 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/341e505abe930d474ac9ab960e4251b8b8880a59e27ed2d1db7e0fbcc349de07.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 186, + 433, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d0e2c17e8160286cd13eeedac026722f00639e0c7adc81536454e65a6bf51d62.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 189, + 661, + 268 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a4c80a539f678cad3b17e98f21d41035787064689b9d55147a082748c55c7913.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 684, + 188, + 877, + 268 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9558911e2e4ae72a9b195696679f6efd7f7793f6a72e8897a953cf24e820a480.jpg", + "image_caption": [ + "Figure 4. Qualitative Results. From left to right, we show our results in feed-forward inference: 1) the input video, 2) two pointmaps at frame $j$ overlayed together, 3) the accumulated reconstruction branch result, and 4) the accumulated tracking branch result. The accumulated reconstruction demonstrates a stable reconstruction of the dynamic scene geometry, while the accumulated tracking illustrates long-term, dense tracking of scene motion. More qualitative results in???" + ], + "image_footnote": [], + "bbox": [ + 112, + 284, + 204, + 325 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/edd1f728eb60a1d24bc90afd7ce3fb97e7cd473f6b7e8d7a3ac6f0fd4bd28031.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 112, + 325, + 202, + 369 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/26807ecb796ea967a5522700d8ca0a5497d7f124c2f92f1f547f3020f3a144f1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 212, + 287, + 429, + 364 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4da768f12644a4c491ff80cb1b51537ec4500180b907397941bfc048994c6644.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 447, + 290, + 658, + 367 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/79936bed42e3066c8b42123becebbc5451dc9b376eddc23da7f97c28f2d50a0b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 285, + 882, + 376 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Input Video", + "bbox": [ + 122, + 383, + 204, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Pair Output", + "bbox": [ + 281, + 383, + 364, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Accumulated Reconstruction", + "bbox": [ + 447, + 383, + 647, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Accumulated Tracking", + "bbox": [ + 694, + 383, + 852, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "longer sequences, including 64-frame videos. Refer to ?? for more results.", + "bbox": [ + 89, + 474, + 482, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Dynamic 3D Reconstruction", + "text_level": 1, + "bbox": [ + 89, + 512, + 344, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets. We evaluate on both synthetic and real-world data. For the synthetic data, we use Point Odyssey [69]. For real-world evaluation, we employ TUM-Dynamics [49], a subset of a SLAM dataset featuring moving people, dense depth maps, and accurate camera poses.", + "bbox": [ + 89, + 534, + 483, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation Metrics. Unlike prior works [58, 67] that separately evaluate video depth and camera pose estimation, we directly compare the reconstructed 3D point clouds to the ground truth using the Average percent of Points within Distance (APD) and End-Point Error (EPE) metrics. We filter out ambiguous floating points in the ground truth data and align the point clouds for each sequence using the median scale before evaluation.", + "bbox": [ + 89, + 616, + 483, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Baselines. We compare our method against MonST3R, MASt3R and DUSt3R, both with global alignment (w/ GA) and in feedforward mode. For the feedforward baselines, we construct image pairs of a video in the form of St4RTrack, that align all frames to a common anchor frame.", + "bbox": [ + 89, + 743, + 483, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results. As in Tab. 2, we also achieve state-of-the-art performance on the task of 3D reconstruction in the world coordinate. Although MonST3R is designed for 3D reconstruction for dynamic scenes, it still underperforms St4RTrack even with global alignment. This further highlights the ben", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "efit of jointly tracking and reconstruction. Since we freeze the reconstruction head to preserve the 3D prior, 3D reconstruction results are similar with test-time adaptation.", + "bbox": [ + 511, + 474, + 906, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Joint Tracking and Reconstruction in the World", + "text_level": 1, + "bbox": [ + 511, + 527, + 906, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method simultaneously predicts 3D point trajectories and point clouds in a single feed-forward pass, evaluated separately in prior sections. Here, we present qualitative results visualizing both raw 3D point trajectories and point clouds within the same world coordinates, as shown in Fig. 4. The \"pair output\" demonstrates that outputs from the tracking and reconstruction branches align well at the same time step. Additionally, the accumulated reconstruction indicates consistency in static regions, while accumulated tracking shows that our method estimates accurate and smooth 3D tracks over time.", + "bbox": [ + 511, + 550, + 906, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 724, + 633, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce St4RTrack, a feed-forward framework that simultaneously achieves 3D point tracking and dynamic reconstruction in the world coordinate from monocular videos using a unified representation. Alongside, we present a novel benchmark, WorldTrack, for systematically evaluating dynamic 3D scene geometry and motion estimation in a global reference frame. Our method achieves state-of-the-art performance on both synthetic and real-world datasets, while also extending beyond fully supervised paradigms by enabling test-time adaptation.", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8510", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 89, + 90, + 277, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank Aleksander Holynski, Yifei Zhang, Chung Min Kim, Brent Yi, and Zehao Yu for helpful discussions. We especially thank Aleksander Holynski for his guidance and feedback. This project is supported in part by NSF CNS-2235013, Bakar Fellows, and BAIR Commons Industry Alliance Sponsors.", + "bbox": [ + 89, + 114, + 483, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 220, + 187, + 237 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M Seitz, and Richard Szeliski. Building rome in a day. Communications of the ACM, 54 (10):105-112, 2011. 2", + "[2] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 2", + "[3] Michael J Black and Padmanabhan Anandan. A framework for the robust estimation of optical flow. In 1993 (4th) International Conference on Computer Vision, pages 231-236. IEEE, 1993. 2", + "[4] Christoph Bregler, Aaron Hertzmann, and Henning Biermann. Recovering non-rigid 3d shape from image streams. In Proceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No. PR00662), pages 690-696. IEEE, 2000. 3", + "[5] Hansheng Chen, Wei Tian, Pichao Wang, Fan Wang, Lu Xiong, and Hao Li. EPro-PnP: Generalized end-to-end probabilistic perspective-n-points for monocular object pose estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 5", + "[6] Yuchao Dai, Hongdong Li, and Mingyi He. A simple prior-free method for non-rigid structure-from-motion factorization. International Journal of Computer Vision, 107:101-122, 2014. 3", + "[7] Andrew J Davison, Ian D Reid, Nicholas D Molton, and Olivier Stasse. Monoslam: Real-time single camera slam. IEEE transactions on pattern analysis and machine intelligence, 29(6):1052-1067, 2007. 2", + "[8] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. TAP-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 35:13610-13626, 2022. 2", + "[9] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. TAPIR: Tracking any point with per-frame initialization and temporal refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10061-10072, 2023. 2", + "[10] Hugh Durrant-Whyte and Tim Bailey. Simultaneous localization and mapping: part i. IEEE robotics & automation magazine, 13(2):99–110, 2006. 2", + "[11] Martin A Fischler and Robert C. Bolles. Random sample consensus: a paradigm for model fitting with applications to" + ], + "bbox": [ + 93, + 246, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 5, 7", + "[12] Katerina Fragkiadaki, Marta Salas, Pablo Arbelaez, and Jitendra Malik. Grouping-based low-rank trajectory completion and 3d reconstruction. advances in neural information processing systems, 27, 2014. 3", + "[13] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 3", + "[14] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. Advances in Neural Information Processing Systems, 35:33768-33780, 2022. 3", + "[15] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3749-3761, 2022. 6", + "[16] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In European Conference on Computer Vision, pages 59-75. Springer, 2022. 2", + "[17] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 1", + "[18] Berthold KP Horn and Brian G Schunck. Determining optical flow. Artificial intelligence, 17(1-3):185–203, 1981. 2", + "[19] Junhwa Hur and Stefan Roth. Self-supervised monocular scene flow estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7396-7405, 2020. 2", + "[20] Linyi Jin, Richard Tucker, Zhengqi Li, David Fouhey, Noah Snavely, and Aleksander Holynski. Stereo4d: Learning how things move in 3d from internet stereo videos. arXiv preprint arXiv:2412.09621, 2024. 3", + "[21] Hanbyul Joo, Tomas Simon, Xulong Li, Hao Liu, Lei Tan, Lin Gui, Sean Banerjee, Timothy Godisart, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social interaction capture, 2016. 7", + "[22] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Dynamic stereo: Consistent dynamic depth from stereo videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13229-13239, 2023. 1, 5, 6", + "[23] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker3: Simpler and better point tracking by pseudolabelling real videos. 2024. 2, 5, 6", + "[24] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. In Proc. ECCV, 2024. 1, 2, 5" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8511", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1621, 2021. 2", + "[26] Skanda Koppula, Ignacio Rocco, Yi Yang, Joe Heyward, João Carreira, Andrew Zisserman, Gabriel Brostow, and Carl Doersch. Tapvid-3d: A benchmark for tracking any point in 3d, 2024. 7", + "[27] Jiahui Lei, Yijia Weng, Adam W. Harley, Leonidas Guibas, and Kostas Daniilidis. MoSca: Dynamic gaussian fusion from casual videos via 4d motion scaffolds. arXiv preprint arXiv:2405.17421, 2024. 3, 6", + "[28] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. EPnP: An accurate O(n) solution to the PnP problem. IJCV, 81:155-166, 2009. 5", + "[29] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. In European Conference on Computer Vision, pages 71-91. Springer, 2024. 5", + "[30] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5521-5531, 2022. 3", + "[31] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 3", + "[32] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4273-4284, 2023. 3", + "[33] Zhengqi Li, Richard Tucker, Forrester Cole, Qianqian Wang, Linyi Jin, Vickie Ye, Angjoo Kanazawa, Aleksander Holynski, and Noah Snavelly. MegaSaM: Accurate, fast, and robust structure and motion from casual dynamic videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 2", + "[34] Qingming Liu, Yuan Liu, Jiepeng Wang, Xianqiang Lyv, Peng Wang, Wenping Wang, and Junhui Hou. Modgs: Dynamic gaussian splattering from casually-captured monocular videos, 2024. 3", + "[35] Bruce D Lucas and Takeo Kanade. An iterative image registration technique with an application to stereo vision. In IJCAI'81: 7th international joint conference on Artificial intelligence, pages 674-679, 1981. 2", + "[36] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis, 2023. 3", + "[37] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: A versatile and accurate monocular slam system. IEEE transactions on robotics, 31(5):1147-1163, 2015. 2", + "[38] Tuan Duc Ngo, Peiye Zhuang, Chuang Gan, Evangelos Kalogerakis, Sergey Tulyakov, Hsin-Ying Lee, and" + ], + "bbox": [ + 91, + 90, + 485, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Chaoyang Wang. DELTA: Dense efficient long-range 3d tracking for any video. In International Conference on Learning Representations (ICLR), 2025. 1, 3", + "[39] David Novotny, Nikhila Ravi, Benjamin Graham, Natalia Neverova, and Andrea Vedaldi. C3dpo: Canonical 3d pose networks for non-rigid structure from motion. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7688-7697, 2019. 3", + "[40] Xiaqing Pan, Nicholas Charron, Yongqian Yang, Scott Peters, Thomas Whelan, Chen Kong, Omkar Parkhi, Richard Newcombe, and Carl Yuheng Ren. Aria digital twin: A new benchmark dataset for egocentric 3d machine perception, 2023. 7", + "[41] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), 2021. 3", + "[42] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Undepth: Universal monocular metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10106-10116, 2024. 2", + "[43] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10318–10327, 2021. 3", + "[44] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. 2", + "[45] Michael Rubinstein, Ce Liu, and William T Freeman. Towards longer long-range motion trajectories. 2012. 2", + "[46] P. Sand and S. Teller. Particle video: Long-range motion estimation using point trajectories. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), pages 2195-2202, 2006. 2", + "[47] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 2", + "[48] Steven M Seitz, Brian Curless, James Diebel, Daniel Scharstein, and Richard Szeliski. A comparison and evaluation of multi-view stereo reconstruction algorithms. In 2006 IEEE computer society conference on computer vision and pattern recognition (CVPR'06), pages 519-528. IEEE, 2006. 2", + "[49] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of RGB-D SLAM systems. pages 573-580, 2012. 8", + "[50] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8934-8943, 2018. 2", + "[51] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow, 2020. 1, 2" + ], + "bbox": [ + 516, + 90, + 906, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8512", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 34:16558-16569, 2021. 2", + "[53] Zachary Teed and Jia Deng. Raft-3d: Scene flow using rigid-motion embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8375–8384, 2021. 2", + "[54] Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. IEEE Transactions on Pattern Analysis & Machine Intelligence, 13(04): 376-380, 1991. 7", + "[55] Sundar Vedula, Simon Baker, Peter Rander, Robert Collins, and Takeo Kanade. Three-dimensional scene flow. In Proceedings of the Seventh IEEE International Conference on Computer Vision, pages 722-729. IEEE, 1999. 2", + "[56] Jianyuan Wang, Nikita Karaev, Christian Rupprecht, and David Novotny. Vggsfm: Visual geometry grounded deep structure from motion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21686-21697, 2024. 2", + "[57] Qianqian Wang, Vickie Ye, Hang Gao, Weijia Zeng, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. 3, 6", + "[58] Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A. Efros, and Angjoo Kanazawa. Continuous 3d perception model with persistent state. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 1, 2, 8", + "[59] Ruicheng Wang, Sicheng Xu, Cassie Dai, Jianfeng Xiang, Yu Deng, Xin Tong, and Jiaolong Yang. Moge: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision, 2024. 2, 5, 6", + "[60] Sen Wang, Ronald Clark, Hongkai Wen, and Niki Trigoni. Deepvo: Towards end-to-end visual odometry with deep recurrent convolutional neural networks. In 2017 IEEE international conference on robotics and automation (ICRA), pages 2043-2050. IEEE, 2017. 2", + "[61] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jérôme Revaud. Dust3r: Geometric 3d vision made easy. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20697-20709, 2023. 2, 3, 5", + "[62] Endre Weiszfeld. Sur le point pour lequel la somme des distances de n points données est minimum. Tohoku Mathematical Journal, First Series, 43:355-386, 1937. 5", + "[63] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20310-20320, 2024. 3", + "[64] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. Spatialtracker: Tracking any 2d pixels in 3d space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 7" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 2", + "[66] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20331-20341, 2024. 3", + "[67] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. MonST3R: A simple approach for estimating geometry in the presence of motion. In International Conference on Learning Representations (ICLR), 2025. 1, 2, 7, 8", + "[68] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T. Freeman. Structure and motion from casual videos. In Proceedings of the European Conference on Computer Vision (ECCV), pages 20-37, 2022. 2", + "[69] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19855-19865, 2023. 5, 6, 8", + "[70] Xiaowei Zhou, Menglong Zhu, Spyridon Leonardos, Konstantinos G Derpanis, and Kostas Daniilidis. Sparseness meets deepness: 3d human pose estimation from monocular video. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4966-4975, 2016. 3" + ], + "bbox": [ + 516, + 90, + 906, + 527 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8513", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/St4RTrack_ Simultaneous 4D Reconstruction and Tracking in the World/8996ce18-d088-44bb-a34b-53f80b972348_model.json b/2025/St4RTrack_ Simultaneous 4D Reconstruction and Tracking in the World/8996ce18-d088-44bb-a34b-53f80b972348_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b55f7733d86dfe2e9d40695fe54037584670bdad --- /dev/null +++ b/2025/St4RTrack_ Simultaneous 4D Reconstruction and Tracking in the World/8996ce18-d088-44bb-a34b-53f80b972348_model.json @@ -0,0 +1,2499 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.134, + 0.085, + 0.865, + 0.108 + ], + "angle": 0, + "content": "St4RTrack: Simultaneous 4D Reconstruction and Tracking in the World" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.134, + 0.842, + 0.173 + ], + "angle": 0, + "content": "Haiwen Feng\\(^{1,2}\\) * Junyi Zhang\\(^{1}\\) * Qianqian Wang\\(^{1}\\) Yufei Ye\\(^{3}\\) Pengcheng Yu\\(^{2}\\) Michael J. Black\\(^{2}\\) Trevor Darrell\\(^{1}\\) Angjoo Kanazawa\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.159, + 0.18, + 0.842, + 0.201 + ], + "angle": 0, + "content": "\\(^{1}\\)UC Berkeley \\(^{2}\\)Max Planck Institute for Intelligent Systems \\(^{3}\\)Stanford University" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.217, + 0.907, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.441, + 0.908, + 0.514 + ], + "angle": 0, + "content": "Figure 1. St4RTrack: Given an RGB video capturing dynamic scenes, St4RTrack simultaneously tracks the points from the initial frame (visualized in purple) and reconstructs the geometry of the subsequent frames (in orange) in a consistent world coordinate frame. St4RTrack is a feed-forward framework that takes a pair of images as input and outputs two pointmaps in the world frame, as the visualization shown in the middle. By iteratively processing the first frame paired with each subsequent frame, St4RTrack achieves simultaneous tracking (right bottom) and reconstruction (right top) for the entire video. Interactive results on our webpage: https://St4RTrack.github.io/." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.537, + 0.33, + 0.553 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.566, + 0.486, + 0.869 + ], + "angle": 0, + "content": "Dynamic 3D reconstruction and point tracking in videos are typically treated as separate tasks, despite their deep connection. We propose St4RTrack, a feed-forward framework that simultaneously reconstructs and tracks dynamic video content in a world coordinate frame from RGB inputs. This is achieved by predicting two appropriately defined pointmaps for a pair of frames captured at different moments. Specifically, we predict both pointmaps at the same moment, in the same world, capturing both static and dynamic scene geometry while maintaining 3D correspondences. Chaining these predictions through the video sequence with respect to a reference frame naturally computes long-range correspondences, effectively combining 3D reconstruction with 3D tracking. Unlike prior methods that rely heavily on 4D ground truth supervision, we employ a novel adaptation scheme based on a reprojection loss. We establish a new extensive benchmark for world-frame reconstruction and tracking, demonstrating the effectiveness and efficiency of our unified, data-driven framework. Our code, model, and benchmark will be released." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.537, + 0.646, + 0.552 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.909, + 0.903 + ], + "angle": 0, + "content": "When asked about the three most important problems in computer vision, Takeo Kanade replied, \"Correspondence, Correspondence, Correspondence!\" This remark is especially pertinent to multi-view 3D reconstruction, where 3D geometry and 2D correspondence are two sides of the same coin; that is, a 3D point in the physical world naturally brings 2D correspondence across its projections in different views, and conversely, corresponded 2D points across views reconstruct the same 3D point after triangulation. This synergy between 3D geometry and 2D correspondence has long formed the foundation of multi-view geometry [17]. However, when the scene becomes dynamic, this synergy appears to vanish, as existing methods—particularly the recent data-driven ones—tend to treat dynamic reconstruction [22, 58, 67] and correspondence [24, 38, 51] as separate, disconnected tasks. We argue that this is a missed opportunity; the synergy between 3D reconstruction and 2D correspondence is not lost in dynamic scenes—it simply requires an additional element: understanding how the scene content evolves over time. This evolution is captured by 3D motion estimation (i.e., dense 3D point tracking), which, when computed across a sequence, effectively ex" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.338, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution, listed alphabetically." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8503" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "plains scene motion. Once tracking accounts for scene dynamics, the problem effectively reduces to the rigid case, where the natural interplay between reconstruction and correspondence can once again be leveraged." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.155, + 0.483, + 0.337 + ], + "angle": 0, + "content": "We propose St4RTrack, a learning framework that unifies reconstruction and tracking directly from RGB video frames. St4RTrack simultaneously reconstructs and tracks dynamic video content in a single consistent world, achieving world-frame 3D tracking, as demonstrated in Fig. 1. Tracking in the world frame decouples the scene and the camera motion, essential for domains where both the camera and content are in motion. Our approach also reconstructs the 3D geometry of observed image content for both static and dynamic portions of the scene. St4RTrack directly predicts their reconstruction and tracking in the world without requiring an additional alignment stage." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.34, + 0.483, + 0.702 + ], + "angle": 0, + "content": "Our key insight stems from the observation that a feedforward static 3D reconstruction method, DUSt3R [61], can be adapted to dynamic scenes simply by changing the pointmaps' annotation [67]. Building on this, we reexamine the pointmaps definition in the 4D scenario and opt to simply redefine its geometric interpretation for both reconstruction and tracking, as illustrated in Fig. 2. Concretely, we achieve it by predicting two pointmaps at the same timestamp and in the same world from a pair of image frames depicting two different timestamps. More specific, given images \\((\\mathbf{I}_i,\\mathbf{I}_j)\\), both pointmaps are predicted in the coordinate frame of \\(\\mathbf{I}_i\\), but at the time specified by \\(\\mathbf{I}_j\\). Our method is realized through a feed-forward network comprising of two branches: the reconstruction branch reconstructs the content of \\(\\mathbf{I}_j\\) in the \\(\\mathbf{I}_i\\) coordinate frame; and the tracking branch, which reconstructs the content of \\(\\mathbf{I}_i\\) in the \\(\\mathbf{I}_i\\) (its own) coordinate frame, but at the time indicated by \\(\\mathbf{I}_j\\). Essentially, the tracking branch predicts how the scene content in \\(\\mathbf{I}_i\\) evolves to match the moment captured in \\(\\mathbf{I}_j\\). This is enabled through a DUSt3R-like dual cross-attention mechanism, where the tracking branch relies on the reconstruction branch to decide how to move points. This minimal change proves sufficient for unifying both dynamic reconstruction and 3D point tracking in the world coordinate system." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Furthermore, unlike existing methods [58, 61, 67] that rely solely on 4D supervision, our approach unlocks 4D reconstruction training on in-the-wild videos via reprojection loss without 4D supervision. This is possible because St4RTrack simultaneously establishes camera parameters, 3D geometry and motion. Specifically, based on the outputs of the reconstruction branch, the camera parameters for \\(\\mathbf{I}_j\\) can be differentiably computed via PnP. Using these cameras, the pointmap of \\(\\mathbf{I}_i\\) is projected into the \\(j\\)-th frame, enabling training with a reprojection loss that leverages 2D correspondences and monocular depth predictions from off-the-shelf approaches [23, 59]. Consequently, the monocular supervisions facilitate effective test-time adaptation of" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "St4RTrack to in-the-wild videos, which can differ substantially from the synthetic data used during pretraining." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.124, + 0.905, + 0.304 + ], + "angle": 0, + "content": "While prior 3D point tracking benchmarks focus on camera coordinate frames [64], our approach enables worldframe 3D tracking. To evaluate this capability, we establish a novel benchmark, WorldTrack, for both tracking and reconstruction in the world coordinate system. We find that our unified method outperforms the strong baselines that combine several pieces on each individual task. Furthermore, we show that our feedforward results can be improved via test-time adaptation. We believe this is a step towards a unified task-agnostic 4D perception system that can be trained on a large-scale video. Our code, model, and benchmark will be released." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.323, + 0.665, + 0.339 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.35, + 0.908, + 0.759 + ], + "angle": 0, + "content": "Camera Estimation and Scene Reconstruction. Jointly estimating camera motion and scene geometry has been studied for decades, often in the context of Structure from Motion (SfM) [1, 47, 48, 56] or Simultaneous Localization and Mapping (SLAM) [7, 10, 37, 52, 60]. However, these methods are primarily designed for static scenes and typically do not model dynamic scene content. Recent advances in learning-based monocular and video depth estimation methods [2, 42, 44, 65] have opened new opportunities to reconstruct dynamic scenes. Notably, R-CVD [25], CasualSAM [68] and MegaSAM [33] jointly optimize camera parameters and per-frame dense depth maps leveraging monocular depth priors, producing consistent depth estimates for dynamic objects and accurate camera parameters even in challenging cases with minimal camera parallax. Another notable recent method, DUSt3R [61], introduces a two-pointmap representation that enables joint estimation of camera motion and scene geometry of a pair of images. While DUSt3R itself primarily focuses on reconstructing static scenes, follow-up effort such as MonST3R [67], demonstrate that this formulation can also effectively handle dynamic scenes with minimal modification on supervisions. Despite these advances, none of the aforementioned methods explicitly estimate 3D scene motion, meaning they do not track the movement of individual 3D points over time. In contrast, our method simultaneously performs joint reconstruction and tracking for dynamic scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "2D/3D Point Tracking. Tracking pixel motion over time is a fundamental problem in computer vision. Optical/Scene flow methods [3, 18, 19, 35, 50, 51, 53, 55] produce dense 2D/3D motion vectors but are inherently short-ranged, struggling with large displacements and occlusions. While long-range point tracking [45, 46] has been studied for decades, it has recently been revitalized via supervised learning [8, 9, 16, 23, 24], enabling more robust tracking over extended time periods and overcoming these limita-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8504" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.09, + 0.187, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.207, + 0.092, + 0.332, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.092, + 0.477, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.237, + 0.483, + 0.321 + ], + "angle": 0, + "content": "Figure 2. Pointmap Comparison of MonST3R and St4RTrack. Given two input frames, MonST3R handles dynamic scenes by reconstructing both pointmaps in their own timestamp. St4RTrack predicts where the points in the first frame move in the second frame, and reconstructs the geometry of the second frame. More details of the representation definition are introduced in Sec. 3.1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.35, + 0.484, + 0.516 + ], + "angle": 0, + "content": "tions. However, these methods still produce only 2D pixel trajectories. More recently, several works [38, 64] achieve 3D tracking by lifting points into 3D space using monocular depth priors and performing tracking in 3D. While closely related to our approach, these methods still operate in the camera frame space, meaning they lack camera motion estimation and do not explicitly separate scene motion from camera motion. In contrast, our method jointly estimates disentangled camera and scene motion, enabling world-space tracking for a more complete understanding of 3D scene dynamics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Joint Dynamic Reconstruction and Tracking. While traditional non-rigid SfMs have studied joint tracking and reconstruction [4, 6, 12, 39, 70] from 2D correspondences, jointly optimizing them from raw videos is highly challenging and typically requires multi-view synchronized videos as input [13, 30, 36, 41, 63]. With recent advances in neural rendering and data-driven geometric priors, reconstructing and rendering dynamic scenes from monocular videos became possible. However, as Gao et al. [14] point out, many methods [43, 66] focus on \"teleporting\" input data, which are effectively multi-view and not representative of real-world videos. In addition, since the main focus is view synthesis, motion estimation serves a secondary role in facilitating information fusion between neighboring frames [31, 32]. More recently, several works [27, 34, 57] focus on jointly recovering camera parameters, persistent scene geometry, and long-range 3D tracks from single, causally captured videos. However, these methods take off-the-shelf priors as given and design per-video optimization techniques that optimize a representation from scratch. Most recently, Stereo4D [20]—a concurrent effort to our work—proposed a pipeline for crafting a real-world 4D tracking dataset using internet stereo videos, enabling the supervised regression of 3D trajectories and geometries between frames. In contrast, we propose a feed-forward" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "method that simultaneously performs reconstruction and tracking, while the same architecture also supports test-time adaptation on unlabeled videos to approach the high quality of optimization-based methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.168, + 0.902, + 0.186 + ], + "angle": 0, + "content": "3. Simultaneous Reconstruction and Tracking" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.194, + 0.907, + 0.436 + ], + "angle": 0, + "content": "We present a framework that simultaneously reconstructs and tracks dynamic video content in 3D within a single consistent world coordinate frame. The core idea is simple yet powerful: reconstructing and tracking can both be achieved by predicting two appropriately defined pointmaps, where both pointmaps reconstruct the scene content observed in each image at the same timestamp in a consistent coordinate system. This enables simultaneous reconstruction of both dynamic and static contents, while tracking across a sequence of image frames in a video. Since all geometry, camera, and motion (i.e. 3D correspondence over time) can be derived from the representation, it can be adapted to videos without any explicit 4D supervision. Below, we discuss the main insight, how St4RTrack compares to prior works. Then, we discuss the details of the model and how it can be trained and adapted to videos." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.448, + 0.86, + 0.464 + ], + "angle": 0, + "content": "3.1. Unified 4D Representation of St4RTrack" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.906, + 0.607 + ], + "angle": 0, + "content": "Given two images \\(\\mathbf{I}_i, \\mathbf{I}_j\\) with dynamic content (see Figure 2), how can one devise a single feedforward approach that simultaneously performs reconstruction and tracking? We argue that the underlying representation must (1) capture camera motion to establish a world coordinate frame, (2) reconstruct the 3D geometry of all observed points, and (3) estimate 3D motion that maintain explicit correspondence over time. In this work, we show that just two properly defined pointmaps suffice to fulfill these requirements." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.906, + 0.764 + ], + "angle": 0, + "content": "Time-Dependent Pointmap. A pointmap representation assumes that each pixel in an image \\(\\mathbf{I}\\) of shape \\(H\\times W\\) is associated with a corresponding 3D point, forming a pointmap \\(\\mathbf{X}\\in \\mathbb{R}^{H\\times W\\times 3}\\). For the case of static scenes, DUSt3R [61] only considers two factors of pointmaps—(1) the source frame of the content and (2) the camera coordinate system in which the points are expressed. However, this definition is insufficient for modeling the dynamic scenario of monocular video. To address this, we introduce a previously overlooked yet decisive factor: time." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Specifically, we define a time-dependent pointmap that encodes the 3D positions of the scene points in a chosen (world) coordinate system at a specific timestamp. For clarity, we denote this representation as \\( {}^a\\mathbf{X}_t^b \\), which denotes the 3D pointmap of physical content from frame \\( b \\), at time \\( t \\), expressed in the coordinate system established by frame \\( a \\). For example, \\( {}^i\\mathbf{X}_j^i \\) represents the geometry originally seen in frame \\( i \\) in frame \\( i \\)'s own coordinate system, but described at timestamp \\( j \\). The time-dependency is achieved by construc" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8505" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.068, + 0.885, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.306, + 0.908, + 0.39 + ], + "angle": 0, + "content": "Figure 3. Overview of St4RTrack. Given frame 1 and frame \\(j\\) as input, the tracking branch outputs \\({}^{1}\\mathbf{X}_{j}^{1}\\), the pointmap that corresponds to observed content of the first frame at timestep \\(j\\) in its own camera coordinate (i.e. world coordinate); the reconstruction branch outputs \\({}^{1}\\mathbf{X}_{j}^{j}\\), the pointmap of the content in frame \\(j\\) at its own timestamp in the world coordinate. To adapt to new videos without any 4D labels, the camera is computed via differentiable PnP from the pointmap, enabling reprojected supervision signals (e.g., 2D trajectories and monocular depth). We finetune both branches during training (Sec. 3.2) with synthetic data, and when adapting to a new video (Sec. 3.3), only the tracking branch is fine-tuned using these reprojected supervision signals." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.404, + 0.483, + 0.434 + ], + "angle": 0, + "content": "tion without explicit timestamp conditioning, as described in the next section." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.441, + 0.484, + 0.486 + ], + "angle": 0, + "content": "Unified 4D Modeling. St4RTrack learns a function \\( f \\) that maps two images \\( \\mathbf{I}_i, \\mathbf{I}_j \\), captured at timestamp \\( i \\) and \\( j \\), into two pointmaps:" + }, + { + "type": "equation", + "bbox": [ + 0.21, + 0.498, + 0.483, + 0.518 + ], + "angle": 0, + "content": "\\[\nf _ {\\theta} \\left(\\mathbf {I} _ {i}, \\mathbf {I} _ {j}\\right) = ^ {i} \\mathbf {X} _ {j} ^ {i}, ^ {i} \\mathbf {X} _ {j} ^ {j}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.531, + 0.484, + 0.621 + ], + "angle": 0, + "content": "The second image \\(\\mathbf{I}_j\\) is reconstructed as the pointmap \\({}^i\\mathbf{X}_j^j\\) in the first image \\(\\mathbf{I}_i\\)'s coordinate frame. Meanwhile, it predicts \\({}^i\\mathbf{X}_j^i\\), representing the 3D motion of how the content from the first image \\(\\mathbf{I}_i\\) moves at timestamp \\(j\\). Thus, both geometry and motion (tracking) are estimated from this unified prediction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.623, + 0.484, + 0.835 + ], + "angle": 0, + "content": "To handle a full video consisting of \\(T\\) frames, we perform tracking and reconstruction by always selecting the first frame as the anchor frame \\(\\mathbf{I}_i\\). Each subsequent frame \\(\\mathbf{I}_j\\) is then paired with this initial frame, ensuring that every new frame is consistently aligned to the coordinate system of the first frame. Specifically, \\(\\{f(\\mathbf{I}_1,\\mathbf{I}_1),f(\\mathbf{I}_1,\\mathbf{I}_2),\\ldots ,f(\\mathbf{I}_1,\\mathbf{I}_T)\\}\\) are computed in the same reference, \\(\\mathbf{I}_1\\), which naturally serves as the world coordinate frame. Thus, world-frame 3D tracking is achieved by explicitly following how points observed in \\(\\mathbf{I}_1\\) are placed throughout the sequence, \\(\\{\\mathbf{^1X_1^1},\\mathbf{^1X_2^1},\\dots,\\mathbf{^1X_T^1}\\}\\), while the world frame dynamic reconstruction is obtained by the paired geometry estimation per-frame, \\(\\{\\mathbf{^1X_1^1},\\mathbf{^1X_2^2},\\dots,\\mathbf{^1X_T^T}\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Relation to prior works. Our formulation of the 4D modeling generalizes prior works in a unified framework. DUSt3R reconstructs and establishes correspondences but is limited to rigid scenes, as correspondence and recon" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.404, + 0.907, + 0.48 + ], + "angle": 0, + "content": "struction are dual tasks for static scenes. With this perspective, one can see that if there is no dynamic component (i.e. frozen moment in time or rigid scenes), our formulation is equivalent to DUSt3R, where both images share the same timestamp \\( t = i \\):" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.492, + 0.907, + 0.512 + ], + "angle": 0, + "content": "\\[\nf _ {\\theta} (\\mathbf {I} _ {i}, \\mathbf {I} _ {j}) = ^ {i} \\mathbf {X} _ {i} ^ {i}, ^ {i} \\mathbf {X} _ {i} ^ {j}. \\qquad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.906, + 0.645 + ], + "angle": 0, + "content": "In a static world, 3D reconstruction from two pointmaps inherently yields the correspondences between them, allowing synergy to arise naturally. However, when objects or the scene are in motion, the dynamic component appears differently in different frames, it becomes crucial to account for 3D scene motion to preserve this synergy. St4RTrack addresses this challenge by predicting the 3D content from the first image at future timestamps." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.646, + 0.906, + 0.676 + ], + "angle": 0, + "content": "In the same framework, we see that MonST3R, the dynamic follow-up of DUSt3R can be expressed as such:" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.688, + 0.907, + 0.709 + ], + "angle": 0, + "content": "\\[\nf _ {\\theta} (\\mathbf {I} _ {i}, \\mathbf {I} _ {j}) = ^ {i} \\mathbf {X} _ {i} ^ {i}, ^ {i} \\mathbf {X} _ {j} ^ {j}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.906, + 0.825 + ], + "angle": 0, + "content": "where each image's 3D geometry is reconstructed in its timestamp, such that the dynamic contents separately align with their frame inputs. While it's sufficient for obtaining dynamic scene geometry, there is no temporal correspondence being established, as illustrated in Fig. 2. Furthermore, both DUSt3R and MonST3R compute the pairwise graphs and perform global alignment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.904 + ], + "angle": 0, + "content": "Since we always designate the first frame as the reference for tracking, the world coordinates are consistently established by the first frame. For simplicity, we omit the explicit notation of the world coordinate in subsequent equations and paragraphs, i.e., \\(\\mathbf{X}_j^i \\coloneqq {}^i\\mathbf{X}_j^i\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8506" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.482, + 0.108 + ], + "angle": 0, + "content": "3.2. Joint Learning of Tracking and Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.295 + ], + "angle": 0, + "content": "In this section, we describe how our framework implements equation 1 within a pair-wise framework as DUSt3R. For each pair of frames, \\(\\mathbf{I}_1\\) and \\(\\mathbf{I}_j\\), we first encode them into token representations using a ViT encoder, then process these tokens through a siamese transformer decoder. The decoder sequentially applies self-attention (allowing tokens within each frame to interact), followed by cross-attention (enabling tokens from one frame to attend to tokens in the other), and finally passes the tokens through an MLP. This continuous information flow between the two branches is crucial for generating spatial-aligned 3D pointmaps in a shared coordinate system, as illustrated in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.297, + 0.484, + 0.372 + ], + "angle": 0, + "content": "Our siamese architecture processes two input views concurrently and generates two 3D pointmaps that are expressed in a common reference frame established by the first view. Although the two branches share the same architectural structure, they serve distinct purposes:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.373, + 0.483, + 0.418 + ], + "angle": 0, + "content": "- Tracking branch predicts the pointmap \\(\\mathbf{X}_j^1\\), which represents the geometry of the first frame at timestamp \\(j\\) in the first frame's coordinates (i.e., the world coordinates)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.419, + 0.482, + 0.463 + ], + "angle": 0, + "content": "- Reconstruction branch predicts the pointmap \\(\\mathbf{X}_j^j\\), which represents the geometry of frame \\(j\\) at its own timestamp, also expressed in the first frame's camera coordinates." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.373, + 0.483, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.465, + 0.484, + 0.541 + ], + "angle": 0, + "content": "Since this architecture is exactly the same as proposed by DUSt3R and subsequently adopted by MonST3R, with the only difference being the output paired pointmaps (Eq.1-3), our network can be initialized with pretrained 3D knowledge from either DUSt3R or MonST3R." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.548, + 0.484, + 0.82 + ], + "angle": 0, + "content": "Pretraining with 4D Synthetic Data. Our proposed representation requires specialized supervision for the Tracking Branch—namely, ensuring that the pointmap from the first frame is correctly positioned in the world across all frames. Achieving this necessitates complete 4D information of the dynamic scene. Therefore, we leverage existing 4D synthetic datasets [22, 69] that provide both the 3D geometry and motion of the rendered content. Specifically, for each dataset, we use the scene mesh vertices (expressed in world coordinates) to provide sparse, masked supervision for the Tracking Branch pointmap representation, and employ per-frame depth maps and camera ground-truth to supervise the Reconstruction Branch. For this fully supervised training process, we use the objectives from DUSt3R. We initialize our dual-branch transformer with weights from MAST3R [29], a DUSt3R variant that has been adapted for 2D correspondence learning. Additional details regarding the 4D synthetic datasets are provided in Sec. 4.1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.419, + 0.849 + ], + "angle": 0, + "content": "3.3. Adapt to Any Video without 4D Label" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "While the synthetic datasets are small-scale and unrealistic, they are sufficient for our network to learn the newly proposed representations. However, fully supervised training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.318 + ], + "angle": 0, + "content": "on these datasets presents two key limitations: 1) The 4D synthetic data is limited in scale and does not encompass the full range of motion and geometry present in real-world dynamic scenes; 2) Our proposed pointmap representation requires the capability to freely move the pointmap within the world coordinates—a departure from conventional pixel-aligned geometry predictions, making small-scale training insufficient for achieving fine-grained predictions. These limitations motivate us to further leverage the 3D geometry and motion inherent in the St4RTrack framework to perform domain adaptation on any video without 4D labels. Specifically, we first show how we can derive camera parameters differentially, and with which we can design reprojected 2D trajectory loss and monocular depth loss to supervise the network." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.325, + 0.907, + 0.553 + ], + "angle": 0, + "content": "Solving Camera Parameters. The intrinsic matrix \\(\\mathbf{K}\\) is first estimated from the tracking branch's first-frame pointmap prediction, following DUSt3R [61]. In this process, the principal point is assumed to be centered, and pixels are treated as square. The focal length is assumed static across frames and estimated using a fast iterative solver based on the Weiszfeld algorithm [62]. Next, the extrinsic parameters \\(\\mathbf{P}^j = [\\mathbf{R}^j | \\mathbf{T}^j]\\) for each frame \\(j\\) are derived using the \"reconstruction\" pointmap \\(\\mathbf{X}_j^j\\). Specifically, each pixel \\(\\mathbf{x}^{j,n}\\) in frame \\(j\\) is associated with a 3D coordinate \\(\\mathbf{X}_j^{j,n}\\) in the shared world coordinate system (established by the first camera), thus forming 2D-to-3D correspondences. We could then solve for \\(\\mathbf{R}^j\\) and \\(\\mathbf{T}^j\\) via a Perspective-\\(n\\)-Points (PnP) [28] solver with RANSAC [11] for outlier rejection:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.566, + 0.905, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\mathbf {R} ^ {j}, \\mathbf {T} ^ {j} = \\underset {\\mathbf {R}, \\mathbf {T}} {\\operatorname {a r g m i n}} \\sum_ {n \\in \\mathcal {I} _ {j}} \\left\\| \\mathbf {x} ^ {j, n} - \\pi \\left(\\mathbf {K} \\left(\\mathbf {R} \\mathbf {X} _ {j} ^ {j, n} + \\mathbf {T}\\right)\\right) \\right\\| ^ {2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.617, + 0.857, + 0.633 + ], + "angle": 0, + "content": "where \\(\\pi (\\cdot)\\) is the projection \\((x,y,z)\\to (x / z,y / z)\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.634, + 0.905, + 0.695 + ], + "angle": 0, + "content": "For differentiability, we adopt a derivative-based Gauss-Newton solver following [5], ensuring that gradients from the reprojection loss can adjust both the camera pose and the 3D pointmaps. Further details are provided in ??" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.701, + 0.906, + 0.87 + ], + "angle": 0, + "content": "Reprojection Loss. With the camera pose of frame \\(j\\) derived, the tracking pointmap \\(\\mathbf{X}_j^1\\) and reconstruction pointmap \\(\\mathbf{X}_j^j\\) can be transformed from the world coordinate system into the camera coordinate system of frame \\(j\\). This transformation enables self-supervised training by enforcing two types of consistency: (1) 2D correspondence consistency, which aligns the projected 2D tracks (from \\(\\mathbf{X}_j^1\\)) with the pseudo-ground truth tracking from CoTracker [23, 24], and (2) geometric consistency, which aligns the scale-invariant depth (from \\(\\mathbf{X}_j^j\\)) with the pseudoground truth monocular depth from MoGe [59]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.903 + ], + "angle": 0, + "content": "More specifically, given the estimated camera pose \\((\\mathbf{R}^j,\\mathbf{T}^j)\\) from frame \\(j\\) and the tracking pointmap \\(\\mathbf{X}_j^1\\), we" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8507" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.473, + 0.108 + ], + "angle": 0, + "content": "reproject these 3D points into the image plane of frame \\(j\\):" + }, + { + "type": "equation", + "bbox": [ + 0.181, + 0.115, + 0.483, + 0.136 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} ^ {j, n} = \\pi \\left(\\mathbf {K} \\left(\\mathbf {R} ^ {j} \\mathbf {X} _ {j} ^ {1, n} + \\mathbf {T} ^ {j}\\right)\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.144, + 0.483, + 0.189 + ], + "angle": 0, + "content": "These reprojected points serve as the predicted 2D tracks and are compared with the pseudo-ground truth tracking points \\(\\mathbf{x}_{\\mathrm{trk}}^{j,n}\\) from CoTracker3 [23]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.189, + 0.484, + 0.266 + ], + "angle": 0, + "content": "To mitigate minor focal inaccuracies that may induce scaling shifts, the reprojection loss is computed in a scale-invariant manner. Let \\(\\mathbf{p}_n = \\hat{\\mathbf{x}}^{j,n}\\) and \\(\\mathbf{g}_n = \\mathbf{x}_{\\mathrm{trk}}^{j,n}\\) for \\(n = 1,\\dots ,N\\) , and denote the image center by \\(c\\) . The scale factor and adjusted predictions are computed together as:" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.275, + 0.484, + 0.317 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {p}} _ {n} = \\left(\\mathbf {p} _ {n} - c\\right) s + c, \\quad s = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\frac {\\left\\| \\mathbf {g} _ {n} - c \\right\\| _ {2}}{\\left\\| \\mathbf {p} _ {n} - c \\right\\| _ {2}}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.325, + 0.479, + 0.34 + ], + "angle": 0, + "content": "Then, the scale-invariant 2D reprojection loss is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.349, + 0.483, + 0.387 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {t r a j}} = \\frac {1}{N} \\sum_ {n \\in \\mathcal {I} _ {j}} \\| \\hat {\\mathbf {p}} _ {n} - \\mathbf {g} _ {n} \\| ^ {2}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.396, + 0.484, + 0.456 + ], + "angle": 0, + "content": "Similarly, to enforce geometric consistency with the mono-depth predictions from MoGe [59], we use the reconstruction pointmap \\(\\mathbf{X}_j^j\\) that transformed to frame \\(j\\)'s camera coordinate. The depth of each transformed 3D point is" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.466, + 0.483, + 0.492 + ], + "angle": 0, + "content": "\\[\nz _ {\\text {p r o j}} ^ {j, n} = \\left(\\mathbf {R} ^ {j} \\mathbf {X} _ {j} ^ {j, n} + \\mathbf {T} ^ {j}\\right) _ {z}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.501, + 0.484, + 0.576 + ], + "angle": 0, + "content": "where \\((\\cdot)_z\\) denotes the third (depth) component. Denote the corresponding mono-depth pseudo ground-truth by \\(z_{\\mathrm{mono}}^{j,n}\\). After solving for an optimal scaling factor \\(\\alpha^*\\) to align the two depth maps, the scale-invariant mono-depth loss is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.584, + 0.484, + 0.646 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {d e p t h}} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\left(\\alpha^ {*} z _ {\\text {p r o j}} ^ {j, n} - z _ {\\text {m o n o}} ^ {j, n}\\right) ^ {2}, \\quad \\alpha^ {*} = \\frac {\\sum_ {i} z _ {\\text {p r o j}} ^ {j , n} z _ {\\text {m o n o}} ^ {j , n}}{\\sum_ {i} \\left(z _ {\\text {p r o j}} ^ {j , n}\\right) ^ {2}}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.652, + 0.484, + 0.776 + ], + "angle": 0, + "content": "3D Self-Consistency. Beyond the 2D reprojection losses, we introduce a 3D self-consistency term that aligns the tracking pointmap \\(\\mathbf{X}_j^1\\) with the reconstruction pointmap \\(\\mathbf{X}_j^j\\). Let \\(\\mathcal{I}_1^\\prime\\) be the set of points in frame 1 that remain visible in frame \\(j\\), and for each point \\(n \\in \\mathcal{I}_1^\\prime\\), denote its corresponding point (provided by CoTracker) in frame \\(j\\) by \\(n^\\prime\\). We then penalize the distance between their predicted 3D positions:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.785, + 0.484, + 0.82 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l i g n}} = \\sum_ {n \\in \\mathcal {I} _ {1} ^ {\\prime}} \\left\\| \\mathbf {X} _ {j} ^ {1, n} - \\mathbf {X} _ {j} ^ {j, n ^ {\\prime}} \\right\\| ^ {2}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.83, + 0.483, + 0.86 + ], + "angle": 0, + "content": "Minimizing \\(\\mathcal{L}_{\\mathrm{align}}\\) ensures that both branches produce consistent geometry in the same timestamp." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.86, + 0.408, + 0.875 + ], + "angle": 0, + "content": "The overall self-supervision loss is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.886, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e p r o j}} = \\mathcal {L} _ {\\text {t r a j}} + \\lambda_ {1} \\mathcal {L} _ {\\text {d e p t h}} + \\lambda_ {2} \\mathcal {L} _ {\\text {a l i g n}}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "with \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) being the weighting factors. Minimizing \\(\\mathcal{L}_{\\mathrm{proj}}\\) aligns the projected 3D structure with the 2D tracking and monocular depth cues and their 3D self-consistency, enabling unsupervised, target-specific refinement of the 3D geometry and point tracking." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.173, + 0.907, + 0.552 + ], + "angle": 0, + "content": "Test-Time Adaptation. To address the gap between synthetic pretraining and real-world data, we incorporate reprojection-based losses to enable test-time adaptation in St4RTrack. Our framework supports two adaptation paradigms: (1) Instance-level adaptation. During testing, we update St4RTrack on new sequences using only the aforementioned reprojected losses while freezing the reconstruction branch. We freeze these weights because both the 2D trajectories and depth are computed under a purely monocular setting, which does not provide view-alignment supervision. This approach preserves the view-alignment capability captured during pretraining. Moreover, since the pretrained network already encodes strong task-relevant representations, this sequence-specific optimization converges rapidly compared to test-time optimization methods that start from scratch. (2) Domain-level adaptation. Unlike tabula-rasa approaches such as [27, 57], which require full re-optimization for each new sequence, St4RTrack is an end-to-end learning framework that enables test-time adaptation to align the model from its pretraining data distribution to the target video domain. After adapting to a sparse set of target-domain samples, St4RTrack can directly perform simultaneous reconstruction and tracking on new sequences from the adapted domain without additional optimization." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.565, + 0.645, + 0.582 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.59, + 0.906, + 0.711 + ], + "angle": 0, + "content": "St4RTrack performs both dense 3D point tracking and dynamic reconstruction in a unified world coordinate system, all within a single inference. In the following section, we first evaluate our method on 3D tracking and dynamic reconstruction separately, and then present the joint results. We also introduce a new benchmark, WorldTrack, for 3D tracking in world coordinates, which is not directly covered by previous methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.721, + 0.714, + 0.737 + ], + "angle": 0, + "content": "4.1. Experimental Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.744, + 0.906, + 0.879 + ], + "angle": 0, + "content": "Datasets. For fully supervised training, we use three synthetic datasets: Point Odyssey (PO) [69], Dynamic Replica (DR) [22], and Kubric [15]. All three datasets contain scene and camera motion and provide mesh vertex positions as ground-truth 3D point trajectories. We randomly sample 24 frames with a stride of \\(1 \\sim 6\\) for each sample sequence. We also filter out less semantically meaningful sequences in PO, resulting in a total of 9.8k sequences for PO, 8.5k for DR, and 5.7k for Kubric dataset." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.886, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Training and Inference. During training, we sample 600" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8508" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.064, + 0.907, + 0.093 + ], + "angle": 0, + "content": "Table 1. World Coordinate 3D Point Tracking. We report the performance of average points under distance \\((\\mathrm{APD}_{3\\mathrm{D}})\\) after global median alignment. We evaluate the accuracy of both all points and dynamic points. The best results are bold. See ?? for more results." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.097, + 0.907, + 0.25 + ], + "angle": 0, + "content": "
CategoryMethodsAll PointsDynamic Points
PODRADTPStudioPODRADTPStudio
CombinationalSpaTracker+RANSAC-Procrustes44.0355.0150.8752.0553.7758.5866.4952.05
SpaTracker+MonST3R47.6555.4951.9550.1658.6159.2169.9450.16
Feed-forwardMonST3R33.4758.0674.3551.3239.3651.8667.9251.32
SpaTracker38.5454.8545.6562.5951.2058.6567.6562.59
St4RTrack (Ours)67.9573.7476.0169.6768.7268.1375.3469.67
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.26, + 0.483, + 0.304 + ], + "angle": 0, + "content": "Table 2. World Coordinate 3D Reconstruction. We report performance on both Point Odyssey (PO) and TUM-Dynamics after global median scaling. The best results are in bold." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.308, + 0.485, + 0.476 + ], + "angle": 0, + "content": "
CategoryMethodsPoint OdysseyTUM-Dynamics
EPE↓APD↑EPE↓APD↑
w/ Global Align.DUSt3R+GA0.608843.900.314770.49
MASt3R+GA0.403060.440.518668.38
MonST3R+GA0.262972.310.342963.87
Feed-forwardDUSt3R0.638645.790.289172.26
MASt3R0.464456.900.551066.22
MonST3R0.304468.250.364661.38
St4RTrack (Ours)0.240678.730.185483.42
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.492, + 0.483, + 0.658 + ], + "angle": 0, + "content": "sequences from each dataset per epoch. We use the AdamW optimizer with a learning rate of \\(5 \\times 10^{-5}\\) and a mini-batch size of 1 per GPU. The model is trained for 50 epochs on 4 A100 GPUs, which takes about one day. For test-time adaptation, we run 500 optimization steps on a single sequence, taking approximately 5 minutes on 4 A100 GPUs. At inference time, the model runs at 30 FPS on an RTX 4090. Although the model is trained on sequences of 24 frames, our pair-wise approach allows it to operate on arbitrarily long videos during inference. Refer to ?? for more details regarding test-time adaptation." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.667, + 0.395, + 0.684 + ], + "angle": 0, + "content": "4.2. 3D Tracking in World Coordinates" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Datasets. 3D tracking in world coordinates is a critical aspect that has been largely overlooked by previous benchmarks [26], which are limited to camera coordinate systems. To address this limitation, we propose a new benchmark for 3D tracking in world coordinates. Our benchmark leverages two real-world datasets—Aerial Digital Twin (ADT) [40] and Panoptic Studio [21]—by converting the TAPVid-3D [26] sequences to world coordinates using paired extrinsic parameters. However, it is noteworthy that the limitations of these datasets: the ADT sequences exhibit minimal scene motion, while the Panoptic Studio lacks camera motion. To overcome these shortcomings, we include two additional synthetic test sets from Point Odyssey and Dynamic Replica, which have both scene and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.263, + 0.907, + 0.322 + ], + "angle": 0, + "content": "camera motion. In total, our benchmark comprises four datasets, each containing 50 sequences of 64 frames. Refer to ?? for more details and visual examples of the benchmark." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.329, + 0.909, + 0.497 + ], + "angle": 0, + "content": "Evaluation Metrics. We follow the TAPVid-3D protocol and use the Average percent of Points within Delta (APD) metric for evaluation. Specifically, we first align the predicted 3D point trajectories with the ground truth by normalizing them with their global median. We then compute the prediction error and measure the percentage of points whose error falls below a given threshold \\(\\delta_{3D}\\) (with \\(\\delta_{3D} \\in \\{0.1\\mathrm{m}, 0.3\\mathrm{m}, 0.5\\mathrm{m}, 1.0\\mathrm{m}\\}\\)) over the first 64 frames. Let \\(\\hat{\\mathbf{P}}_t^i\\) denote the \\(i\\)-th predicted point at time \\(t\\) and \\(\\mathbf{P}_t^i\\) denote its corresponding ground-truth location. The resulting \\(\\mathrm{APD}_{3D}\\) is then computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.508, + 0.907, + 0.543 + ], + "angle": 0, + "content": "\\[\n\\mathrm {A P D} _ {\\mathrm {3 D}} \\equiv \\sum_ {i, t} \\mathbb {1} \\left(\\left\\| \\hat {\\mathbf {P}} _ {t} ^ {i} - \\mathbf {P} _ {t} ^ {i} \\right\\| < \\delta_ {3 D}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.555, + 0.906, + 0.585 + ], + "angle": 0, + "content": "where \\(\\mathbb{1}(\\cdot)\\) is the indicator function and \\(\\| \\cdot \\|\\) denotes the Euclidean norm." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.592, + 0.909, + 0.775 + ], + "angle": 0, + "content": "Baselines. Since no existing work explicitly performs 3D tracking in world coordinates, for our feedforward baselines, we compare against the camera coordinate 3D tracking method SpatialTracker [64] and a dynamic 3D reconstruction method MonST3R [67] (as a non-tracking baseline). In addition, we implement two combinational baselines for world coordinate 3D tracking. The first baseline applies Procrustes alignment [54] and RANSAC [11] to the camera coordinate 3D tracks predicted by SpatialTracker to offset the camera motion. The second baseline leverages the camera poses predicted by the dynamic SLAM method MonST3R to compensate for camera motion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Results. As shown in Tab. 1, we achieve state-of-the-art performance, with test-time adaptation proving particularly beneficial for dynamic points. Notably, on the Panoptic Studio dataset, which is captured with a fixed camera and can be considered a fair benchmark for camera coordinate tracking methods, our approach still outperforms SpatialTracker [64]. It is worth noting that although our model is trained on sequences of 24 frames, it generalizes well to" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8509" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.075, + 0.205, + 0.116 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.117, + 0.205, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.064, + 0.434, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.064, + 0.667, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.67, + 0.063, + 0.883, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.183, + 0.205, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.225, + 0.204, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.188, + 0.434, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.19, + 0.662, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.189, + 0.878, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.285, + 0.205, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.327, + 0.204, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.288, + 0.43, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.448, + 0.291, + 0.659, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.67, + 0.286, + 0.883, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.384, + 0.205, + 0.398 + ], + "angle": 0, + "content": "Input Video" + }, + { + "type": "text", + "bbox": [ + 0.282, + 0.384, + 0.366, + 0.399 + ], + "angle": 0, + "content": "Pair Output" + }, + { + "type": "text", + "bbox": [ + 0.448, + 0.384, + 0.648, + 0.398 + ], + "angle": 0, + "content": "Accumulated Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.696, + 0.384, + 0.853, + 0.399 + ], + "angle": 0, + "content": "Accumulated Tracking" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.405, + 0.907, + 0.46 + ], + "angle": 0, + "content": "Figure 4. Qualitative Results. From left to right, we show our results in feed-forward inference: 1) the input video, 2) two pointmaps at frame \\( j \\) overlayed together, 3) the accumulated reconstruction branch result, and 4) the accumulated tracking branch result. The accumulated reconstruction demonstrates a stable reconstruction of the dynamic scene geometry, while the accumulated tracking illustrates long-term, dense tracking of scene motion. More qualitative results in???" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.475, + 0.483, + 0.504 + ], + "angle": 0, + "content": "longer sequences, including 64-frame videos. Refer to ?? for more results." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.513, + 0.345, + 0.529 + ], + "angle": 0, + "content": "4.3. Dynamic 3D Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.484, + 0.611 + ], + "angle": 0, + "content": "Datasets. We evaluate on both synthetic and real-world data. For the synthetic data, we use Point Odyssey [69]. For real-world evaluation, we employ TUM-Dynamics [49], a subset of a SLAM dataset featuring moving people, dense depth maps, and accurate camera poses." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.617, + 0.484, + 0.737 + ], + "angle": 0, + "content": "Evaluation Metrics. Unlike prior works [58, 67] that separately evaluate video depth and camera pose estimation, we directly compare the reconstructed 3D point clouds to the ground truth using the Average percent of Points within Distance (APD) and End-Point Error (EPE) metrics. We filter out ambiguous floating points in the ground truth data and align the point clouds for each sequence using the median scale before evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.744, + 0.484, + 0.82 + ], + "angle": 0, + "content": "Baselines. We compare our method against MonST3R, MASt3R and DUSt3R, both with global alignment (w/ GA) and in feedforward mode. For the feedforward baselines, we construct image pairs of a video in the form of St4RTrack, that align all frames to a common anchor frame." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Results. As in Tab. 2, we also achieve state-of-the-art performance on the task of 3D reconstruction in the world coordinate. Although MonST3R is designed for 3D reconstruction for dynamic scenes, it still underperforms St4RTrack even with global alignment. This further highlights the ben" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.475, + 0.907, + 0.52 + ], + "angle": 0, + "content": "efit of jointly tracking and reconstruction. Since we freeze the reconstruction head to preserve the 3D prior, 3D reconstruction results are similar with test-time adaptation." + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.529, + 0.907, + 0.545 + ], + "angle": 0, + "content": "4.4. Joint Tracking and Reconstruction in the World" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.907, + 0.716 + ], + "angle": 0, + "content": "Our method simultaneously predicts 3D point trajectories and point clouds in a single feed-forward pass, evaluated separately in prior sections. Here, we present qualitative results visualizing both raw 3D point trajectories and point clouds within the same world coordinates, as shown in Fig. 4. The \"pair output\" demonstrates that outputs from the tracking and reconstruction branches align well at the same time step. Additionally, the accumulated reconstruction indicates consistency in static regions, while accumulated tracking shows that our method estimates accurate and smooth 3D tracks over time." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.725, + 0.634, + 0.74 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We introduce St4RTrack, a feed-forward framework that simultaneously achieves 3D point tracking and dynamic reconstruction in the world coordinate from monocular videos using a unified representation. Alongside, we present a novel benchmark, WorldTrack, for systematically evaluating dynamic 3D scene geometry and motion estimation in a global reference frame. Our method achieves state-of-the-art performance on both synthetic and real-world datasets, while also extending beyond fully supervised paradigms by enabling test-time adaptation." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8510" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.279, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.207 + ], + "angle": 0, + "content": "We thank Aleksander Holynski, Yifei Zhang, Chung Min Kim, Brent Yi, and Zehao Yu for helpful discussions. We especially thank Aleksander Holynski for his guidance and feedback. This project is supported in part by NSF CNS-2235013, Bakar Fellows, and BAIR Commons Industry Alliance Sponsors." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.222, + 0.188, + 0.238 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.247, + 0.484, + 0.302 + ], + "angle": 0, + "content": "[1] Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M Seitz, and Richard Szeliski. Building rome in a day. Communications of the ACM, 54 (10):105-112, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.304, + 0.484, + 0.358 + ], + "angle": 0, + "content": "[2] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.361, + 0.483, + 0.415 + ], + "angle": 0, + "content": "[3] Michael J Black and Padmanabhan Anandan. A framework for the robust estimation of optical flow. In 1993 (4th) International Conference on Computer Vision, pages 231-236. IEEE, 1993. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.418, + 0.484, + 0.487 + ], + "angle": 0, + "content": "[4] Christoph Bregler, Aaron Hertzmann, and Henning Biermann. Recovering non-rigid 3d shape from image streams. In Proceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No. PR00662), pages 690-696. IEEE, 2000. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.489, + 0.484, + 0.558 + ], + "angle": 0, + "content": "[5] Hansheng Chen, Wei Tian, Pichao Wang, Fan Wang, Lu Xiong, and Hao Li. EPro-PnP: Generalized end-to-end probabilistic perspective-n-points for monocular object pose estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.561, + 0.484, + 0.614 + ], + "angle": 0, + "content": "[6] Yuchao Dai, Hongdong Li, and Mingyi He. A simple prior-free method for non-rigid structure-from-motion factorization. International Journal of Computer Vision, 107:101-122, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.617, + 0.484, + 0.672 + ], + "angle": 0, + "content": "[7] Andrew J Davison, Ian D Reid, Nicholas D Molton, and Olivier Stasse. Monoslam: Real-time single camera slam. IEEE transactions on pattern analysis and machine intelligence, 29(6):1052-1067, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.675, + 0.484, + 0.743 + ], + "angle": 0, + "content": "[8] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. TAP-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 35:13610-13626, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.746, + 0.484, + 0.827 + ], + "angle": 0, + "content": "[9] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. TAPIR: Tracking any point with per-frame initialization and temporal refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10061-10072, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.83, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[10] Hugh Durrant-Whyte and Tim Bailey. Simultaneous localization and mapping: part i. IEEE robotics & automation magazine, 13(2):99–110, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[11] Martin A Fischler and Robert C. Bolles. Random sample consensus: a paradigm for model fitting with applications to" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.247, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[12] Katerina Fragkiadaki, Marta Salas, Pablo Arbelaez, and Jitendra Malik. Grouping-based low-rank trajectory completion and 3d reconstruction. advances in neural information processing systems, 27, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[13] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[14] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. Advances in Neural Information Processing Systems, 35:33768-33780, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[15] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3749-3761, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[16] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In European Conference on Computer Vision, pages 59-75. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[17] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.518 + ], + "angle": 0, + "content": "[18] Berthold KP Horn and Brian G Schunck. Determining optical flow. Artificial intelligence, 17(1-3):185–203, 1981. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.52, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[19] Junhwa Hur and Stefan Roth. Self-supervised monocular scene flow estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7396-7405, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.577, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[20] Linyi Jin, Richard Tucker, Zhengqi Li, David Fouhey, Noah Snavely, and Aleksander Holynski. Stereo4d: Learning how things move in 3d from internet stereo videos. arXiv preprint arXiv:2412.09621, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.634, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[21] Hanbyul Joo, Tomas Simon, Xulong Li, Hao Liu, Lei Tan, Lin Gui, Sean Banerjee, Timothy Godisart, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social interaction capture, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[22] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Dynamic stereo: Consistent dynamic depth from stereo videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13229-13239, 2023. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[23] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker3: Simpler and better point tracking by pseudolabelling real videos. 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[24] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. In Proc. ECCV, 2024. 1, 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8511" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.148 + ], + "angle": 0, + "content": "[25] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1621, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[26] Skanda Koppula, Ignacio Rocco, Yi Yang, Joe Heyward, João Carreira, Andrew Zisserman, Gabriel Brostow, and Carl Doersch. Tapvid-3d: A benchmark for tracking any point in 3d, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.483, + 0.262 + ], + "angle": 0, + "content": "[27] Jiahui Lei, Yijia Weng, Adam W. Harley, Leonidas Guibas, and Kostas Daniilidis. MoSca: Dynamic gaussian fusion from casual videos via 4d motion scaffolds. arXiv preprint arXiv:2405.17421, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.483, + 0.304 + ], + "angle": 0, + "content": "[28] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. EPnP: An accurate O(n) solution to the PnP problem. IJCV, 81:155-166, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.483, + 0.348 + ], + "angle": 0, + "content": "[29] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. In European Conference on Computer Vision, pages 71-91. Springer, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.483, + 0.433 + ], + "angle": 0, + "content": "[30] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5521-5531, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.483, + 0.503 + ], + "angle": 0, + "content": "[31] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.483, + 0.573 + ], + "angle": 0, + "content": "[32] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4273-4284, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.483, + 0.658 + ], + "angle": 0, + "content": "[33] Zhengqi Li, Richard Tucker, Forrester Cole, Qianqian Wang, Linyi Jin, Vickie Ye, Angjoo Kanazawa, Aleksander Holynski, and Noah Snavelly. MegaSaM: Accurate, fast, and robust structure and motion from casual dynamic videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.66, + 0.483, + 0.715 + ], + "angle": 0, + "content": "[34] Qingming Liu, Yuan Liu, Jiepeng Wang, Xianqiang Lyv, Peng Wang, Wenping Wang, and Junhui Hou. Modgs: Dynamic gaussian splattering from casually-captured monocular videos, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.716, + 0.483, + 0.772 + ], + "angle": 0, + "content": "[35] Bruce D Lucas and Takeo Kanade. An iterative image registration technique with an application to stereo vision. In IJCAI'81: 7th international joint conference on Artificial intelligence, pages 674-679, 1981. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.773, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[36] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.816, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[37] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: A versatile and accurate monocular slam system. IEEE transactions on robotics, 31(5):1147-1163, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.872, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[38] Tuan Duc Ngo, Peiye Zhuang, Chuang Gan, Evangelos Kalogerakis, Sergey Tulyakov, Hsin-Ying Lee, and" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Chaoyang Wang. DELTA: Dense efficient long-range 3d tracking for any video. In International Conference on Learning Representations (ICLR), 2025. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[39] David Novotny, Nikhila Ravi, Benjamin Graham, Natalia Neverova, and Andrea Vedaldi. C3dpo: Canonical 3d pose networks for non-rigid structure from motion. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7688-7697, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[40] Xiaqing Pan, Nicholas Charron, Yongqian Yang, Scott Peters, Thomas Whelan, Chen Kong, Omkar Parkhi, Richard Newcombe, and Carl Yuheng Ren. Aria digital twin: A new benchmark dataset for egocentric 3d machine perception, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[41] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[42] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Undepth: Universal monocular metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10106-10116, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[43] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10318–10327, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.483, + 0.907, + 0.538 + ], + "angle": 0, + "content": "[44] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[45] Michael Rubinstein, Ce Liu, and William T Freeman. Towards longer long-range motion trajectories. 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.622 + ], + "angle": 0, + "content": "[46] P. Sand and S. Teller. Particle video: Long-range motion estimation using point trajectories. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), pages 2195-2202, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[47] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[48] Steven M Seitz, Brian Curless, James Diebel, Daniel Scharstein, and Richard Szeliski. A comparison and evaluation of multi-view stereo reconstruction algorithms. In 2006 IEEE computer society conference on computer vision and pattern recognition (CVPR'06), pages 519-528. IEEE, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[49] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of RGB-D SLAM systems. pages 573-580, 2012. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[50] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8934-8943, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.872, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[51] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow, 2020. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8512" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[52] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 34:16558-16569, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[53] Zachary Teed and Jia Deng. Raft-3d: Scene flow using rigid-motion embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8375–8384, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[54] Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. IEEE Transactions on Pattern Analysis & Machine Intelligence, 13(04): 376-380, 1991. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[55] Sundar Vedula, Simon Baker, Peter Rander, Robert Collins, and Takeo Kanade. Three-dimensional scene flow. In Proceedings of the Seventh IEEE International Conference on Computer Vision, pages 722-729. IEEE, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[56] Jianyuan Wang, Nikita Karaev, Christian Rupprecht, and David Novotny. Vggsfm: Visual geometry grounded deep structure from motion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21686-21697, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[57] Qianqian Wang, Vickie Ye, Hang Gao, Weijia Zeng, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[58] Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A. Efros, and Angjoo Kanazawa. Continuous 3d perception model with persistent state. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.574 + ], + "angle": 0, + "content": "[59] Ruicheng Wang, Sicheng Xu, Cassie Dai, Jianfeng Xiang, Yu Deng, Xin Tong, and Jiaolong Yang. Moge: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[60] Sen Wang, Ronald Clark, Hongkai Wen, and Niki Trigoni. Deepvo: Towards end-to-end visual odometry with deep recurrent convolutional neural networks. In 2017 IEEE international conference on robotics and automation (ICRA), pages 2043-2050. IEEE, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[61] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jérôme Revaud. Dust3r: Geometric 3d vision made easy. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20697-20709, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[62] Endre Weiszfeld. Sur le point pour lequel la somme des distances de n points données est minimum. Tohoku Mathematical Journal, First Series, 43:355-386, 1937. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[63] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20310-20320, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[64] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. Spatialtracker: Tracking any 2d pixels in 3d space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.161 + ], + "angle": 0, + "content": "[65] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.906, + 0.232 + ], + "angle": 0, + "content": "[66] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20331-20341, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.906, + 0.315 + ], + "angle": 0, + "content": "[67] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. MonST3R: A simple approach for estimating geometry in the presence of motion. In International Conference on Learning Representations (ICLR), 2025. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.318, + 0.906, + 0.386 + ], + "angle": 0, + "content": "[68] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T. Freeman. Structure and motion from casual videos. In Proceedings of the European Conference on Computer Vision (ECCV), pages 20-37, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.388, + 0.906, + 0.457 + ], + "angle": 0, + "content": "[69] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19855-19865, 2023. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.459, + 0.906, + 0.528 + ], + "angle": 0, + "content": "[70] Xiaowei Zhou, Menglong Zhu, Spyridon Leonardos, Konstantinos G Derpanis, and Kostas Daniilidis. Sparseness meets deepness: 3d human pose estimation from monocular video. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4966-4975, 2016. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8513" + } + ] +] \ No newline at end of file diff --git a/2025/St4RTrack_ Simultaneous 4D Reconstruction and Tracking in the World/8996ce18-d088-44bb-a34b-53f80b972348_origin.pdf b/2025/St4RTrack_ Simultaneous 4D Reconstruction and Tracking in the World/8996ce18-d088-44bb-a34b-53f80b972348_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f81e7e54e371a7799f3cbd421b6856ae43a912b7 --- /dev/null +++ b/2025/St4RTrack_ Simultaneous 4D Reconstruction and Tracking in the World/8996ce18-d088-44bb-a34b-53f80b972348_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:97331f7c6f884be443ce95898062df28e7bbf56e11545ab8597bc2fd2c49371c +size 4342862 diff --git a/2025/St4RTrack_ Simultaneous 4D Reconstruction and Tracking in the World/full.md b/2025/St4RTrack_ Simultaneous 4D Reconstruction and Tracking in the World/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ac7af46402ecaf34cc5fae44bb490d479fa128f2 --- /dev/null +++ b/2025/St4RTrack_ Simultaneous 4D Reconstruction and Tracking in the World/full.md @@ -0,0 +1,360 @@ +# St4RTrack: Simultaneous 4D Reconstruction and Tracking in the World + +Haiwen Feng $^{1,2}$ * Junyi Zhang $^{1}$ * Qianqian Wang $^{1}$ Yufei Ye $^{3}$ Pengcheng Yu $^{2}$ Michael J. Black $^{2}$ Trevor Darrell $^{1}$ Angjoo Kanazawa $^{1}$ + +$^{1}$ UC Berkeley $^{2}$ Max Planck Institute for Intelligent Systems $^{3}$ Stanford University + +![](images/e4cfba1126e92c9e355bab8c7db34ddf1f91ab4871227ed3126373ba26c8f337.jpg) +Figure 1. St4RTrack: Given an RGB video capturing dynamic scenes, St4RTrack simultaneously tracks the points from the initial frame (visualized in purple) and reconstructs the geometry of the subsequent frames (in orange) in a consistent world coordinate frame. St4RTrack is a feed-forward framework that takes a pair of images as input and outputs two pointmaps in the world frame, as the visualization shown in the middle. By iteratively processing the first frame paired with each subsequent frame, St4RTrack achieves simultaneous tracking (right bottom) and reconstruction (right top) for the entire video. Interactive results on our webpage: https://St4RTrack.github.io/. + +# Abstract + +Dynamic 3D reconstruction and point tracking in videos are typically treated as separate tasks, despite their deep connection. We propose St4RTrack, a feed-forward framework that simultaneously reconstructs and tracks dynamic video content in a world coordinate frame from RGB inputs. This is achieved by predicting two appropriately defined pointmaps for a pair of frames captured at different moments. Specifically, we predict both pointmaps at the same moment, in the same world, capturing both static and dynamic scene geometry while maintaining 3D correspondences. Chaining these predictions through the video sequence with respect to a reference frame naturally computes long-range correspondences, effectively combining 3D reconstruction with 3D tracking. Unlike prior methods that rely heavily on 4D ground truth supervision, we employ a novel adaptation scheme based on a reprojection loss. We establish a new extensive benchmark for world-frame reconstruction and tracking, demonstrating the effectiveness and efficiency of our unified, data-driven framework. Our code, model, and benchmark will be released. + +# 1. Introduction + +When asked about the three most important problems in computer vision, Takeo Kanade replied, "Correspondence, Correspondence, Correspondence!" This remark is especially pertinent to multi-view 3D reconstruction, where 3D geometry and 2D correspondence are two sides of the same coin; that is, a 3D point in the physical world naturally brings 2D correspondence across its projections in different views, and conversely, corresponded 2D points across views reconstruct the same 3D point after triangulation. This synergy between 3D geometry and 2D correspondence has long formed the foundation of multi-view geometry [17]. However, when the scene becomes dynamic, this synergy appears to vanish, as existing methods—particularly the recent data-driven ones—tend to treat dynamic reconstruction [22, 58, 67] and correspondence [24, 38, 51] as separate, disconnected tasks. We argue that this is a missed opportunity; the synergy between 3D reconstruction and 2D correspondence is not lost in dynamic scenes—it simply requires an additional element: understanding how the scene content evolves over time. This evolution is captured by 3D motion estimation (i.e., dense 3D point tracking), which, when computed across a sequence, effectively ex + +plains scene motion. Once tracking accounts for scene dynamics, the problem effectively reduces to the rigid case, where the natural interplay between reconstruction and correspondence can once again be leveraged. + +We propose St4RTrack, a learning framework that unifies reconstruction and tracking directly from RGB video frames. St4RTrack simultaneously reconstructs and tracks dynamic video content in a single consistent world, achieving world-frame 3D tracking, as demonstrated in Fig. 1. Tracking in the world frame decouples the scene and the camera motion, essential for domains where both the camera and content are in motion. Our approach also reconstructs the 3D geometry of observed image content for both static and dynamic portions of the scene. St4RTrack directly predicts their reconstruction and tracking in the world without requiring an additional alignment stage. + +Our key insight stems from the observation that a feedforward static 3D reconstruction method, DUSt3R [61], can be adapted to dynamic scenes simply by changing the pointmaps' annotation [67]. Building on this, we reexamine the pointmaps definition in the 4D scenario and opt to simply redefine its geometric interpretation for both reconstruction and tracking, as illustrated in Fig. 2. Concretely, we achieve it by predicting two pointmaps at the same timestamp and in the same world from a pair of image frames depicting two different timestamps. More specific, given images $(\mathbf{I}_i,\mathbf{I}_j)$ , both pointmaps are predicted in the coordinate frame of $\mathbf{I}_i$ , but at the time specified by $\mathbf{I}_j$ . Our method is realized through a feed-forward network comprising of two branches: the reconstruction branch reconstructs the content of $\mathbf{I}_j$ in the $\mathbf{I}_i$ coordinate frame; and the tracking branch, which reconstructs the content of $\mathbf{I}_i$ in the $\mathbf{I}_i$ (its own) coordinate frame, but at the time indicated by $\mathbf{I}_j$ . Essentially, the tracking branch predicts how the scene content in $\mathbf{I}_i$ evolves to match the moment captured in $\mathbf{I}_j$ . This is enabled through a DUSt3R-like dual cross-attention mechanism, where the tracking branch relies on the reconstruction branch to decide how to move points. This minimal change proves sufficient for unifying both dynamic reconstruction and 3D point tracking in the world coordinate system. + +Furthermore, unlike existing methods [58, 61, 67] that rely solely on 4D supervision, our approach unlocks 4D reconstruction training on in-the-wild videos via reprojection loss without 4D supervision. This is possible because St4RTrack simultaneously establishes camera parameters, 3D geometry and motion. Specifically, based on the outputs of the reconstruction branch, the camera parameters for $\mathbf{I}_j$ can be differentiably computed via PnP. Using these cameras, the pointmap of $\mathbf{I}_i$ is projected into the $j$ -th frame, enabling training with a reprojection loss that leverages 2D correspondences and monocular depth predictions from off-the-shelf approaches [23, 59]. Consequently, the monocular supervisions facilitate effective test-time adaptation of + +St4RTrack to in-the-wild videos, which can differ substantially from the synthetic data used during pretraining. + +While prior 3D point tracking benchmarks focus on camera coordinate frames [64], our approach enables worldframe 3D tracking. To evaluate this capability, we establish a novel benchmark, WorldTrack, for both tracking and reconstruction in the world coordinate system. We find that our unified method outperforms the strong baselines that combine several pieces on each individual task. Furthermore, we show that our feedforward results can be improved via test-time adaptation. We believe this is a step towards a unified task-agnostic 4D perception system that can be trained on a large-scale video. Our code, model, and benchmark will be released. + +# 2. Related Works + +Camera Estimation and Scene Reconstruction. Jointly estimating camera motion and scene geometry has been studied for decades, often in the context of Structure from Motion (SfM) [1, 47, 48, 56] or Simultaneous Localization and Mapping (SLAM) [7, 10, 37, 52, 60]. However, these methods are primarily designed for static scenes and typically do not model dynamic scene content. Recent advances in learning-based monocular and video depth estimation methods [2, 42, 44, 65] have opened new opportunities to reconstruct dynamic scenes. Notably, R-CVD [25], CasualSAM [68] and MegaSAM [33] jointly optimize camera parameters and per-frame dense depth maps leveraging monocular depth priors, producing consistent depth estimates for dynamic objects and accurate camera parameters even in challenging cases with minimal camera parallax. Another notable recent method, DUSt3R [61], introduces a two-pointmap representation that enables joint estimation of camera motion and scene geometry of a pair of images. While DUSt3R itself primarily focuses on reconstructing static scenes, follow-up effort such as MonST3R [67], demonstrate that this formulation can also effectively handle dynamic scenes with minimal modification on supervisions. Despite these advances, none of the aforementioned methods explicitly estimate 3D scene motion, meaning they do not track the movement of individual 3D points over time. In contrast, our method simultaneously performs joint reconstruction and tracking for dynamic scenes. + +2D/3D Point Tracking. Tracking pixel motion over time is a fundamental problem in computer vision. Optical/Scene flow methods [3, 18, 19, 35, 50, 51, 53, 55] produce dense 2D/3D motion vectors but are inherently short-ranged, struggling with large displacements and occlusions. While long-range point tracking [45, 46] has been studied for decades, it has recently been revitalized via supervised learning [8, 9, 16, 23, 24], enabling more robust tracking over extended time periods and overcoming these limita- + +![](images/41aa5171dca6774fdbdd14db755bf094b7fc028fe61fc026cd8563dd1c1ac99d.jpg) +Figure 2. Pointmap Comparison of MonST3R and St4RTrack. Given two input frames, MonST3R handles dynamic scenes by reconstructing both pointmaps in their own timestamp. St4RTrack predicts where the points in the first frame move in the second frame, and reconstructs the geometry of the second frame. More details of the representation definition are introduced in Sec. 3.1. + +![](images/f1e32e9c6ab572e2e4767e4091e3f9e9987ef8eb4f791ef2a30f977db9cd9e45.jpg) + +![](images/f645ec427d3a3c046c2467d50953e0b115cd37cb432f1f2b22444b492e6249ca.jpg) + +tions. However, these methods still produce only 2D pixel trajectories. More recently, several works [38, 64] achieve 3D tracking by lifting points into 3D space using monocular depth priors and performing tracking in 3D. While closely related to our approach, these methods still operate in the camera frame space, meaning they lack camera motion estimation and do not explicitly separate scene motion from camera motion. In contrast, our method jointly estimates disentangled camera and scene motion, enabling world-space tracking for a more complete understanding of 3D scene dynamics. + +Joint Dynamic Reconstruction and Tracking. While traditional non-rigid SfMs have studied joint tracking and reconstruction [4, 6, 12, 39, 70] from 2D correspondences, jointly optimizing them from raw videos is highly challenging and typically requires multi-view synchronized videos as input [13, 30, 36, 41, 63]. With recent advances in neural rendering and data-driven geometric priors, reconstructing and rendering dynamic scenes from monocular videos became possible. However, as Gao et al. [14] point out, many methods [43, 66] focus on "teleporting" input data, which are effectively multi-view and not representative of real-world videos. In addition, since the main focus is view synthesis, motion estimation serves a secondary role in facilitating information fusion between neighboring frames [31, 32]. More recently, several works [27, 34, 57] focus on jointly recovering camera parameters, persistent scene geometry, and long-range 3D tracks from single, causally captured videos. However, these methods take off-the-shelf priors as given and design per-video optimization techniques that optimize a representation from scratch. Most recently, Stereo4D [20]—a concurrent effort to our work—proposed a pipeline for crafting a real-world 4D tracking dataset using internet stereo videos, enabling the supervised regression of 3D trajectories and geometries between frames. In contrast, we propose a feed-forward + +method that simultaneously performs reconstruction and tracking, while the same architecture also supports test-time adaptation on unlabeled videos to approach the high quality of optimization-based methods. + +# 3. Simultaneous Reconstruction and Tracking + +We present a framework that simultaneously reconstructs and tracks dynamic video content in 3D within a single consistent world coordinate frame. The core idea is simple yet powerful: reconstructing and tracking can both be achieved by predicting two appropriately defined pointmaps, where both pointmaps reconstruct the scene content observed in each image at the same timestamp in a consistent coordinate system. This enables simultaneous reconstruction of both dynamic and static contents, while tracking across a sequence of image frames in a video. Since all geometry, camera, and motion (i.e. 3D correspondence over time) can be derived from the representation, it can be adapted to videos without any explicit 4D supervision. Below, we discuss the main insight, how St4RTrack compares to prior works. Then, we discuss the details of the model and how it can be trained and adapted to videos. + +# 3.1. Unified 4D Representation of St4RTrack + +Given two images $\mathbf{I}_i, \mathbf{I}_j$ with dynamic content (see Figure 2), how can one devise a single feedforward approach that simultaneously performs reconstruction and tracking? We argue that the underlying representation must (1) capture camera motion to establish a world coordinate frame, (2) reconstruct the 3D geometry of all observed points, and (3) estimate 3D motion that maintain explicit correspondence over time. In this work, we show that just two properly defined pointmaps suffice to fulfill these requirements. + +Time-Dependent Pointmap. A pointmap representation assumes that each pixel in an image $\mathbf{I}$ of shape $H\times W$ is associated with a corresponding 3D point, forming a pointmap $\mathbf{X}\in \mathbb{R}^{H\times W\times 3}$ . For the case of static scenes, DUSt3R [61] only considers two factors of pointmaps—(1) the source frame of the content and (2) the camera coordinate system in which the points are expressed. However, this definition is insufficient for modeling the dynamic scenario of monocular video. To address this, we introduce a previously overlooked yet decisive factor: time. + +Specifically, we define a time-dependent pointmap that encodes the 3D positions of the scene points in a chosen (world) coordinate system at a specific timestamp. For clarity, we denote this representation as ${}^a\mathbf{X}_t^b$ , which denotes the 3D pointmap of physical content from frame $b$ , at time $t$ , expressed in the coordinate system established by frame $a$ . For example, ${}^i\mathbf{X}_j^i$ represents the geometry originally seen in frame $i$ in frame $i$ 's own coordinate system, but described at timestamp $j$ . The time-dependency is achieved by construc + +![](images/838ff1e8eeefd1c74338805156461b5e20fed63826feab1939d84f794bab6951.jpg) +Figure 3. Overview of St4RTrack. Given frame 1 and frame $j$ as input, the tracking branch outputs ${}^{1}\mathbf{X}_{j}^{1}$ , the pointmap that corresponds to observed content of the first frame at timestep $j$ in its own camera coordinate (i.e. world coordinate); the reconstruction branch outputs ${}^{1}\mathbf{X}_{j}^{j}$ , the pointmap of the content in frame $j$ at its own timestamp in the world coordinate. To adapt to new videos without any 4D labels, the camera is computed via differentiable PnP from the pointmap, enabling reprojected supervision signals (e.g., 2D trajectories and monocular depth). We finetune both branches during training (Sec. 3.2) with synthetic data, and when adapting to a new video (Sec. 3.3), only the tracking branch is fine-tuned using these reprojected supervision signals. + +tion without explicit timestamp conditioning, as described in the next section. + +Unified 4D Modeling. St4RTrack learns a function $f$ that maps two images $\mathbf{I}_i, \mathbf{I}_j$ , captured at timestamp $i$ and $j$ , into two pointmaps: + +$$ +f _ {\theta} \left(\mathbf {I} _ {i}, \mathbf {I} _ {j}\right) = ^ {i} \mathbf {X} _ {j} ^ {i}, ^ {i} \mathbf {X} _ {j} ^ {j}. \tag {1} +$$ + +The second image $\mathbf{I}_j$ is reconstructed as the pointmap ${}^i\mathbf{X}_j^j$ in the first image $\mathbf{I}_i$ 's coordinate frame. Meanwhile, it predicts ${}^i\mathbf{X}_j^i$ , representing the 3D motion of how the content from the first image $\mathbf{I}_i$ moves at timestamp $j$ . Thus, both geometry and motion (tracking) are estimated from this unified prediction. + +To handle a full video consisting of $T$ frames, we perform tracking and reconstruction by always selecting the first frame as the anchor frame $\mathbf{I}_i$ . Each subsequent frame $\mathbf{I}_j$ is then paired with this initial frame, ensuring that every new frame is consistently aligned to the coordinate system of the first frame. Specifically, $\{f(\mathbf{I}_1,\mathbf{I}_1),f(\mathbf{I}_1,\mathbf{I}_2),\ldots ,f(\mathbf{I}_1,\mathbf{I}_T)\}$ are computed in the same reference, $\mathbf{I}_1$ , which naturally serves as the world coordinate frame. Thus, world-frame 3D tracking is achieved by explicitly following how points observed in $\mathbf{I}_1$ are placed throughout the sequence, $\{\mathbf{^1X_1^1},\mathbf{^1X_2^1},\dots,\mathbf{^1X_T^1}\}$ , while the world frame dynamic reconstruction is obtained by the paired geometry estimation per-frame, $\{\mathbf{^1X_1^1},\mathbf{^1X_2^2},\dots,\mathbf{^1X_T^T}\}$ . + +Relation to prior works. Our formulation of the 4D modeling generalizes prior works in a unified framework. DUSt3R reconstructs and establishes correspondences but is limited to rigid scenes, as correspondence and recon + +struction are dual tasks for static scenes. With this perspective, one can see that if there is no dynamic component (i.e. frozen moment in time or rigid scenes), our formulation is equivalent to DUSt3R, where both images share the same timestamp $t = i$ : + +$$ +f _ {\theta} (\mathbf {I} _ {i}, \mathbf {I} _ {j}) = ^ {i} \mathbf {X} _ {i} ^ {i}, ^ {i} \mathbf {X} _ {i} ^ {j}. \qquad (2) +$$ + +In a static world, 3D reconstruction from two pointmaps inherently yields the correspondences between them, allowing synergy to arise naturally. However, when objects or the scene are in motion, the dynamic component appears differently in different frames, it becomes crucial to account for 3D scene motion to preserve this synergy. St4RTrack addresses this challenge by predicting the 3D content from the first image at future timestamps. + +In the same framework, we see that MonST3R, the dynamic follow-up of DUSt3R can be expressed as such: + +$$ +f _ {\theta} (\mathbf {I} _ {i}, \mathbf {I} _ {j}) = ^ {i} \mathbf {X} _ {i} ^ {i}, ^ {i} \mathbf {X} _ {j} ^ {j}, \tag {3} +$$ + +where each image's 3D geometry is reconstructed in its timestamp, such that the dynamic contents separately align with their frame inputs. While it's sufficient for obtaining dynamic scene geometry, there is no temporal correspondence being established, as illustrated in Fig. 2. Furthermore, both DUSt3R and MonST3R compute the pairwise graphs and perform global alignment. + +Since we always designate the first frame as the reference for tracking, the world coordinates are consistently established by the first frame. For simplicity, we omit the explicit notation of the world coordinate in subsequent equations and paragraphs, i.e., $\mathbf{X}_j^i \coloneqq {}^i\mathbf{X}_j^i$ + +# 3.2. Joint Learning of Tracking and Reconstruction + +In this section, we describe how our framework implements equation 1 within a pair-wise framework as DUSt3R. For each pair of frames, $\mathbf{I}_1$ and $\mathbf{I}_j$ , we first encode them into token representations using a ViT encoder, then process these tokens through a siamese transformer decoder. The decoder sequentially applies self-attention (allowing tokens within each frame to interact), followed by cross-attention (enabling tokens from one frame to attend to tokens in the other), and finally passes the tokens through an MLP. This continuous information flow between the two branches is crucial for generating spatial-aligned 3D pointmaps in a shared coordinate system, as illustrated in Fig. 3. + +Our siamese architecture processes two input views concurrently and generates two 3D pointmaps that are expressed in a common reference frame established by the first view. Although the two branches share the same architectural structure, they serve distinct purposes: + +- Tracking branch predicts the pointmap $\mathbf{X}_j^1$ , which represents the geometry of the first frame at timestamp $j$ in the first frame's coordinates (i.e., the world coordinates). +- Reconstruction branch predicts the pointmap $\mathbf{X}_j^j$ , which represents the geometry of frame $j$ at its own timestamp, also expressed in the first frame's camera coordinates. + +Since this architecture is exactly the same as proposed by DUSt3R and subsequently adopted by MonST3R, with the only difference being the output paired pointmaps (Eq.1-3), our network can be initialized with pretrained 3D knowledge from either DUSt3R or MonST3R. + +Pretraining with 4D Synthetic Data. Our proposed representation requires specialized supervision for the Tracking Branch—namely, ensuring that the pointmap from the first frame is correctly positioned in the world across all frames. Achieving this necessitates complete 4D information of the dynamic scene. Therefore, we leverage existing 4D synthetic datasets [22, 69] that provide both the 3D geometry and motion of the rendered content. Specifically, for each dataset, we use the scene mesh vertices (expressed in world coordinates) to provide sparse, masked supervision for the Tracking Branch pointmap representation, and employ per-frame depth maps and camera ground-truth to supervise the Reconstruction Branch. For this fully supervised training process, we use the objectives from DUSt3R. We initialize our dual-branch transformer with weights from MAST3R [29], a DUSt3R variant that has been adapted for 2D correspondence learning. Additional details regarding the 4D synthetic datasets are provided in Sec. 4.1. + +# 3.3. Adapt to Any Video without 4D Label + +While the synthetic datasets are small-scale and unrealistic, they are sufficient for our network to learn the newly proposed representations. However, fully supervised training + +on these datasets presents two key limitations: 1) The 4D synthetic data is limited in scale and does not encompass the full range of motion and geometry present in real-world dynamic scenes; 2) Our proposed pointmap representation requires the capability to freely move the pointmap within the world coordinates—a departure from conventional pixel-aligned geometry predictions, making small-scale training insufficient for achieving fine-grained predictions. These limitations motivate us to further leverage the 3D geometry and motion inherent in the St4RTrack framework to perform domain adaptation on any video without 4D labels. Specifically, we first show how we can derive camera parameters differentially, and with which we can design reprojected 2D trajectory loss and monocular depth loss to supervise the network. + +Solving Camera Parameters. The intrinsic matrix $\mathbf{K}$ is first estimated from the tracking branch's first-frame pointmap prediction, following DUSt3R [61]. In this process, the principal point is assumed to be centered, and pixels are treated as square. The focal length is assumed static across frames and estimated using a fast iterative solver based on the Weiszfeld algorithm [62]. Next, the extrinsic parameters $\mathbf{P}^j = [\mathbf{R}^j | \mathbf{T}^j]$ for each frame $j$ are derived using the "reconstruction" pointmap $\mathbf{X}_j^j$ . Specifically, each pixel $\mathbf{x}^{j,n}$ in frame $j$ is associated with a 3D coordinate $\mathbf{X}_j^{j,n}$ in the shared world coordinate system (established by the first camera), thus forming 2D-to-3D correspondences. We could then solve for $\mathbf{R}^j$ and $\mathbf{T}^j$ via a Perspective- $n$ -Points (PnP) [28] solver with RANSAC [11] for outlier rejection: + +$$ +\mathbf {R} ^ {j}, \mathbf {T} ^ {j} = \underset {\mathbf {R}, \mathbf {T}} {\operatorname {a r g m i n}} \sum_ {n \in \mathcal {I} _ {j}} \left\| \mathbf {x} ^ {j, n} - \pi \left(\mathbf {K} \left(\mathbf {R} \mathbf {X} _ {j} ^ {j, n} + \mathbf {T}\right)\right) \right\| ^ {2}, \tag {4} +$$ + +where $\pi (\cdot)$ is the projection $(x,y,z)\to (x / z,y / z)$ . + +For differentiability, we adopt a derivative-based Gauss-Newton solver following [5], ensuring that gradients from the reprojection loss can adjust both the camera pose and the 3D pointmaps. Further details are provided in ?? + +Reprojection Loss. With the camera pose of frame $j$ derived, the tracking pointmap $\mathbf{X}_j^1$ and reconstruction pointmap $\mathbf{X}_j^j$ can be transformed from the world coordinate system into the camera coordinate system of frame $j$ . This transformation enables self-supervised training by enforcing two types of consistency: (1) 2D correspondence consistency, which aligns the projected 2D tracks (from $\mathbf{X}_j^1$ ) with the pseudo-ground truth tracking from CoTracker [23, 24], and (2) geometric consistency, which aligns the scale-invariant depth (from $\mathbf{X}_j^j$ ) with the pseudoground truth monocular depth from MoGe [59]. + +More specifically, given the estimated camera pose $(\mathbf{R}^j,\mathbf{T}^j)$ from frame $j$ and the tracking pointmap $\mathbf{X}_j^1$ , we + +reproject these 3D points into the image plane of frame $j$ : + +$$ +\hat {\mathbf {x}} ^ {j, n} = \pi \left(\mathbf {K} \left(\mathbf {R} ^ {j} \mathbf {X} _ {j} ^ {1, n} + \mathbf {T} ^ {j}\right)\right). \tag {5} +$$ + +These reprojected points serve as the predicted 2D tracks and are compared with the pseudo-ground truth tracking points $\mathbf{x}_{\mathrm{trk}}^{j,n}$ from CoTracker3 [23]. + +To mitigate minor focal inaccuracies that may induce scaling shifts, the reprojection loss is computed in a scale-invariant manner. Let $\mathbf{p}_n = \hat{\mathbf{x}}^{j,n}$ and $\mathbf{g}_n = \mathbf{x}_{\mathrm{trk}}^{j,n}$ for $n = 1,\dots ,N$ , and denote the image center by $c$ . The scale factor and adjusted predictions are computed together as: + +$$ +\hat {\mathbf {p}} _ {n} = \left(\mathbf {p} _ {n} - c\right) s + c, \quad s = \frac {1}{N} \sum_ {n = 1} ^ {N} \frac {\left\| \mathbf {g} _ {n} - c \right\| _ {2}}{\left\| \mathbf {p} _ {n} - c \right\| _ {2}}. \tag {6} +$$ + +Then, the scale-invariant 2D reprojection loss is defined as + +$$ +\mathcal {L} _ {\mathrm {t r a j}} = \frac {1}{N} \sum_ {n \in \mathcal {I} _ {j}} \| \hat {\mathbf {p}} _ {n} - \mathbf {g} _ {n} \| ^ {2}. \tag {7} +$$ + +Similarly, to enforce geometric consistency with the mono-depth predictions from MoGe [59], we use the reconstruction pointmap $\mathbf{X}_j^j$ that transformed to frame $j$ 's camera coordinate. The depth of each transformed 3D point is + +$$ +z _ {\text {p r o j}} ^ {j, n} = \left(\mathbf {R} ^ {j} \mathbf {X} _ {j} ^ {j, n} + \mathbf {T} ^ {j}\right) _ {z}, \tag {8} +$$ + +where $(\cdot)_z$ denotes the third (depth) component. Denote the corresponding mono-depth pseudo ground-truth by $z_{\mathrm{mono}}^{j,n}$ . After solving for an optimal scaling factor $\alpha^*$ to align the two depth maps, the scale-invariant mono-depth loss is defined as + +$$ +L _ {\text {d e p t h}} = \frac {1}{N} \sum_ {n = 1} ^ {N} \left(\alpha^ {*} z _ {\text {p r o j}} ^ {j, n} - z _ {\text {m o n o}} ^ {j, n}\right) ^ {2}, \quad \alpha^ {*} = \frac {\sum_ {i} z _ {\text {p r o j}} ^ {j , n} z _ {\text {m o n o}} ^ {j , n}}{\sum_ {i} \left(z _ {\text {p r o j}} ^ {j , n}\right) ^ {2}}. \tag {9} +$$ + +3D Self-Consistency. Beyond the 2D reprojection losses, we introduce a 3D self-consistency term that aligns the tracking pointmap $\mathbf{X}_j^1$ with the reconstruction pointmap $\mathbf{X}_j^j$ . Let $\mathcal{I}_1^\prime$ be the set of points in frame 1 that remain visible in frame $j$ , and for each point $n \in \mathcal{I}_1^\prime$ , denote its corresponding point (provided by CoTracker) in frame $j$ by $n^\prime$ . We then penalize the distance between their predicted 3D positions: + +$$ +\mathcal {L} _ {\text {a l i g n}} = \sum_ {n \in \mathcal {I} _ {1} ^ {\prime}} \left\| \mathbf {X} _ {j} ^ {1, n} - \mathbf {X} _ {j} ^ {j, n ^ {\prime}} \right\| ^ {2}. \tag {10} +$$ + +Minimizing $\mathcal{L}_{\mathrm{align}}$ ensures that both branches produce consistent geometry in the same timestamp. + +The overall self-supervision loss is given by: + +$$ +\mathcal {L} _ {\text {r e p r o j}} = \mathcal {L} _ {\text {t r a j}} + \lambda_ {1} \mathcal {L} _ {\text {d e p t h}} + \lambda_ {2} \mathcal {L} _ {\text {a l i g n}}, \tag {11} +$$ + +with $\lambda_{1}$ and $\lambda_{2}$ being the weighting factors. Minimizing $\mathcal{L}_{\mathrm{proj}}$ aligns the projected 3D structure with the 2D tracking and monocular depth cues and their 3D self-consistency, enabling unsupervised, target-specific refinement of the 3D geometry and point tracking. + +Test-Time Adaptation. To address the gap between synthetic pretraining and real-world data, we incorporate reprojection-based losses to enable test-time adaptation in St4RTrack. Our framework supports two adaptation paradigms: (1) Instance-level adaptation. During testing, we update St4RTrack on new sequences using only the aforementioned reprojected losses while freezing the reconstruction branch. We freeze these weights because both the 2D trajectories and depth are computed under a purely monocular setting, which does not provide view-alignment supervision. This approach preserves the view-alignment capability captured during pretraining. Moreover, since the pretrained network already encodes strong task-relevant representations, this sequence-specific optimization converges rapidly compared to test-time optimization methods that start from scratch. (2) Domain-level adaptation. Unlike tabula-rasa approaches such as [27, 57], which require full re-optimization for each new sequence, St4RTrack is an end-to-end learning framework that enables test-time adaptation to align the model from its pretraining data distribution to the target video domain. After adapting to a sparse set of target-domain samples, St4RTrack can directly perform simultaneous reconstruction and tracking on new sequences from the adapted domain without additional optimization. + +# 4. Experiments + +St4RTrack performs both dense 3D point tracking and dynamic reconstruction in a unified world coordinate system, all within a single inference. In the following section, we first evaluate our method on 3D tracking and dynamic reconstruction separately, and then present the joint results. We also introduce a new benchmark, WorldTrack, for 3D tracking in world coordinates, which is not directly covered by previous methods. + +# 4.1. Experimental Details + +Datasets. For fully supervised training, we use three synthetic datasets: Point Odyssey (PO) [69], Dynamic Replica (DR) [22], and Kubric [15]. All three datasets contain scene and camera motion and provide mesh vertex positions as ground-truth 3D point trajectories. We randomly sample 24 frames with a stride of $1 \sim 6$ for each sample sequence. We also filter out less semantically meaningful sequences in PO, resulting in a total of 9.8k sequences for PO, 8.5k for DR, and 5.7k for Kubric dataset. + +Training and Inference. During training, we sample 600 + +Table 1. World Coordinate 3D Point Tracking. We report the performance of average points under distance $(\mathrm{APD}_{3\mathrm{D}})$ after global median alignment. We evaluate the accuracy of both all points and dynamic points. The best results are bold. See ?? for more results. + +
CategoryMethodsAll PointsDynamic Points
PODRADTPStudioPODRADTPStudio
CombinationalSpaTracker+RANSAC-Procrustes44.0355.0150.8752.0553.7758.5866.4952.05
SpaTracker+MonST3R47.6555.4951.9550.1658.6159.2169.9450.16
Feed-forwardMonST3R33.4758.0674.3551.3239.3651.8667.9251.32
SpaTracker38.5454.8545.6562.5951.2058.6567.6562.59
St4RTrack (Ours)67.9573.7476.0169.6768.7268.1375.3469.67
+ +Table 2. World Coordinate 3D Reconstruction. We report performance on both Point Odyssey (PO) and TUM-Dynamics after global median scaling. The best results are in bold. + +
CategoryMethodsPoint OdysseyTUM-Dynamics
EPE↓APD↑EPE↓APD↑
w/ Global Align.DUSt3R+GA0.608843.900.314770.49
MASt3R+GA0.403060.440.518668.38
MonST3R+GA0.262972.310.342963.87
Feed-forwardDUSt3R0.638645.790.289172.26
MASt3R0.464456.900.551066.22
MonST3R0.304468.250.364661.38
St4RTrack (Ours)0.240678.730.185483.42
+ +sequences from each dataset per epoch. We use the AdamW optimizer with a learning rate of $5 \times 10^{-5}$ and a mini-batch size of 1 per GPU. The model is trained for 50 epochs on 4 A100 GPUs, which takes about one day. For test-time adaptation, we run 500 optimization steps on a single sequence, taking approximately 5 minutes on 4 A100 GPUs. At inference time, the model runs at 30 FPS on an RTX 4090. Although the model is trained on sequences of 24 frames, our pair-wise approach allows it to operate on arbitrarily long videos during inference. Refer to ?? for more details regarding test-time adaptation. + +# 4.2. 3D Tracking in World Coordinates + +Datasets. 3D tracking in world coordinates is a critical aspect that has been largely overlooked by previous benchmarks [26], which are limited to camera coordinate systems. To address this limitation, we propose a new benchmark for 3D tracking in world coordinates. Our benchmark leverages two real-world datasets—Aerial Digital Twin (ADT) [40] and Panoptic Studio [21]—by converting the TAPVid-3D [26] sequences to world coordinates using paired extrinsic parameters. However, it is noteworthy that the limitations of these datasets: the ADT sequences exhibit minimal scene motion, while the Panoptic Studio lacks camera motion. To overcome these shortcomings, we include two additional synthetic test sets from Point Odyssey and Dynamic Replica, which have both scene and + +camera motion. In total, our benchmark comprises four datasets, each containing 50 sequences of 64 frames. Refer to ?? for more details and visual examples of the benchmark. + +Evaluation Metrics. We follow the TAPVid-3D protocol and use the Average percent of Points within Delta (APD) metric for evaluation. Specifically, we first align the predicted 3D point trajectories with the ground truth by normalizing them with their global median. We then compute the prediction error and measure the percentage of points whose error falls below a given threshold $\delta_{3D}$ (with $\delta_{3D} \in \{0.1\mathrm{m}, 0.3\mathrm{m}, 0.5\mathrm{m}, 1.0\mathrm{m}\}$ ) over the first 64 frames. Let $\hat{\mathbf{P}}_t^i$ denote the $i$ -th predicted point at time $t$ and $\mathbf{P}_t^i$ denote its corresponding ground-truth location. The resulting $\mathrm{APD}_{3D}$ is then computed as follows: + +$$ +\mathrm {A P D} _ {\mathrm {3 D}} \equiv \sum_ {i, t} \mathbb {1} \left(\left\| \hat {\mathbf {P}} _ {t} ^ {i} - \mathbf {P} _ {t} ^ {i} \right\| < \delta_ {3 D}\right), \tag {12} +$$ + +where $\mathbb{1}(\cdot)$ is the indicator function and $\| \cdot \|$ denotes the Euclidean norm. + +Baselines. Since no existing work explicitly performs 3D tracking in world coordinates, for our feedforward baselines, we compare against the camera coordinate 3D tracking method SpatialTracker [64] and a dynamic 3D reconstruction method MonST3R [67] (as a non-tracking baseline). In addition, we implement two combinational baselines for world coordinate 3D tracking. The first baseline applies Procrustes alignment [54] and RANSAC [11] to the camera coordinate 3D tracks predicted by SpatialTracker to offset the camera motion. The second baseline leverages the camera poses predicted by the dynamic SLAM method MonST3R to compensate for camera motion. + +Results. As shown in Tab. 1, we achieve state-of-the-art performance, with test-time adaptation proving particularly beneficial for dynamic points. Notably, on the Panoptic Studio dataset, which is captured with a fixed camera and can be considered a fair benchmark for camera coordinate tracking methods, our approach still outperforms SpatialTracker [64]. It is worth noting that although our model is trained on sequences of 24 frames, it generalizes well to + +![](images/8326b61c09a351b48947af958d39883c51ff9ad97e5979afb1933fd2311f71ab.jpg) + +![](images/eaa581cf9c9b9f9077f85026ace1801dae1ca7a68f8b5552801c1cd46f5f914c.jpg) + +![](images/8c8d8982a2cb365c4df587b6df14e7e89c3425d4a2ac59d4c3654fce0ec6fe35.jpg) + +![](images/4ee0e24f3babbf21a88a367b57e343135347e3c2bcd5e96b72600f1b1efdc5d4.jpg) + +![](images/e78831837f016db6990ccc530b742ff54534d0ac2c0011f1a9b2ae972cd2b883.jpg) + +![](images/ff00bfcb118553f05b5cb325eb8bcb24a0aff8700cdb55007cd24367e7adf729.jpg) + +![](images/6e9c87060f4f471cb1074682ba5e4abef0d8897e338fc7a9f0a49ab24e35dcec.jpg) + +![](images/341e505abe930d474ac9ab960e4251b8b8880a59e27ed2d1db7e0fbcc349de07.jpg) + +![](images/d0e2c17e8160286cd13eeedac026722f00639e0c7adc81536454e65a6bf51d62.jpg) + +![](images/a4c80a539f678cad3b17e98f21d41035787064689b9d55147a082748c55c7913.jpg) + +![](images/9558911e2e4ae72a9b195696679f6efd7f7793f6a72e8897a953cf24e820a480.jpg) +Figure 4. Qualitative Results. From left to right, we show our results in feed-forward inference: 1) the input video, 2) two pointmaps at frame $j$ overlayed together, 3) the accumulated reconstruction branch result, and 4) the accumulated tracking branch result. The accumulated reconstruction demonstrates a stable reconstruction of the dynamic scene geometry, while the accumulated tracking illustrates long-term, dense tracking of scene motion. More qualitative results in??? + +![](images/edd1f728eb60a1d24bc90afd7ce3fb97e7cd473f6b7e8d7a3ac6f0fd4bd28031.jpg) + +![](images/26807ecb796ea967a5522700d8ca0a5497d7f124c2f92f1f547f3020f3a144f1.jpg) + +![](images/4da768f12644a4c491ff80cb1b51537ec4500180b907397941bfc048994c6644.jpg) + +![](images/79936bed42e3066c8b42123becebbc5451dc9b376eddc23da7f97c28f2d50a0b.jpg) + +Input Video + +Pair Output + +Accumulated Reconstruction + +Accumulated Tracking + +longer sequences, including 64-frame videos. Refer to ?? for more results. + +# 4.3. Dynamic 3D Reconstruction + +Datasets. We evaluate on both synthetic and real-world data. For the synthetic data, we use Point Odyssey [69]. For real-world evaluation, we employ TUM-Dynamics [49], a subset of a SLAM dataset featuring moving people, dense depth maps, and accurate camera poses. + +Evaluation Metrics. Unlike prior works [58, 67] that separately evaluate video depth and camera pose estimation, we directly compare the reconstructed 3D point clouds to the ground truth using the Average percent of Points within Distance (APD) and End-Point Error (EPE) metrics. We filter out ambiguous floating points in the ground truth data and align the point clouds for each sequence using the median scale before evaluation. + +Baselines. We compare our method against MonST3R, MASt3R and DUSt3R, both with global alignment (w/ GA) and in feedforward mode. For the feedforward baselines, we construct image pairs of a video in the form of St4RTrack, that align all frames to a common anchor frame. + +Results. As in Tab. 2, we also achieve state-of-the-art performance on the task of 3D reconstruction in the world coordinate. Although MonST3R is designed for 3D reconstruction for dynamic scenes, it still underperforms St4RTrack even with global alignment. This further highlights the ben + +efit of jointly tracking and reconstruction. Since we freeze the reconstruction head to preserve the 3D prior, 3D reconstruction results are similar with test-time adaptation. + +# 4.4. Joint Tracking and Reconstruction in the World + +Our method simultaneously predicts 3D point trajectories and point clouds in a single feed-forward pass, evaluated separately in prior sections. Here, we present qualitative results visualizing both raw 3D point trajectories and point clouds within the same world coordinates, as shown in Fig. 4. The "pair output" demonstrates that outputs from the tracking and reconstruction branches align well at the same time step. Additionally, the accumulated reconstruction indicates consistency in static regions, while accumulated tracking shows that our method estimates accurate and smooth 3D tracks over time. + +# 5. Conclusion + +We introduce St4RTrack, a feed-forward framework that simultaneously achieves 3D point tracking and dynamic reconstruction in the world coordinate from monocular videos using a unified representation. Alongside, we present a novel benchmark, WorldTrack, for systematically evaluating dynamic 3D scene geometry and motion estimation in a global reference frame. Our method achieves state-of-the-art performance on both synthetic and real-world datasets, while also extending beyond fully supervised paradigms by enabling test-time adaptation. + +# 6. Acknowledgements + +We thank Aleksander Holynski, Yifei Zhang, Chung Min Kim, Brent Yi, and Zehao Yu for helpful discussions. We especially thank Aleksander Holynski for his guidance and feedback. This project is supported in part by NSF CNS-2235013, Bakar Fellows, and BAIR Commons Industry Alliance Sponsors. + +# References + +[1] Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M Seitz, and Richard Szeliski. Building rome in a day. Communications of the ACM, 54 (10):105-112, 2011. 2 +[2] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 2 +[3] Michael J Black and Padmanabhan Anandan. A framework for the robust estimation of optical flow. In 1993 (4th) International Conference on Computer Vision, pages 231-236. IEEE, 1993. 2 +[4] Christoph Bregler, Aaron Hertzmann, and Henning Biermann. Recovering non-rigid 3d shape from image streams. In Proceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No. PR00662), pages 690-696. IEEE, 2000. 3 +[5] Hansheng Chen, Wei Tian, Pichao Wang, Fan Wang, Lu Xiong, and Hao Li. EPro-PnP: Generalized end-to-end probabilistic perspective-n-points for monocular object pose estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 5 +[6] Yuchao Dai, Hongdong Li, and Mingyi He. A simple prior-free method for non-rigid structure-from-motion factorization. International Journal of Computer Vision, 107:101-122, 2014. 3 +[7] Andrew J Davison, Ian D Reid, Nicholas D Molton, and Olivier Stasse. Monoslam: Real-time single camera slam. IEEE transactions on pattern analysis and machine intelligence, 29(6):1052-1067, 2007. 2 +[8] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. TAP-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 35:13610-13626, 2022. 2 +[9] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. TAPIR: Tracking any point with per-frame initialization and temporal refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10061-10072, 2023. 2 +[10] Hugh Durrant-Whyte and Tim Bailey. Simultaneous localization and mapping: part i. IEEE robotics & automation magazine, 13(2):99–110, 2006. 2 +[11] Martin A Fischler and Robert C. Bolles. Random sample consensus: a paradigm for model fitting with applications to + +image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 5, 7 +[12] Katerina Fragkiadaki, Marta Salas, Pablo Arbelaez, and Jitendra Malik. Grouping-based low-rank trajectory completion and 3d reconstruction. advances in neural information processing systems, 27, 2014. 3 +[13] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 3 +[14] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. Advances in Neural Information Processing Systems, 35:33768-33780, 2022. 3 +[15] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3749-3761, 2022. 6 +[16] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In European Conference on Computer Vision, pages 59-75. Springer, 2022. 2 +[17] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 1 +[18] Berthold KP Horn and Brian G Schunck. Determining optical flow. Artificial intelligence, 17(1-3):185–203, 1981. 2 +[19] Junhwa Hur and Stefan Roth. Self-supervised monocular scene flow estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7396-7405, 2020. 2 +[20] Linyi Jin, Richard Tucker, Zhengqi Li, David Fouhey, Noah Snavely, and Aleksander Holynski. Stereo4d: Learning how things move in 3d from internet stereo videos. arXiv preprint arXiv:2412.09621, 2024. 3 +[21] Hanbyul Joo, Tomas Simon, Xulong Li, Hao Liu, Lei Tan, Lin Gui, Sean Banerjee, Timothy Godisart, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social interaction capture, 2016. 7 +[22] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Dynamic stereo: Consistent dynamic depth from stereo videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13229-13239, 2023. 1, 5, 6 +[23] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker3: Simpler and better point tracking by pseudolabelling real videos. 2024. 2, 5, 6 +[24] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. In Proc. ECCV, 2024. 1, 2, 5 + +[25] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1621, 2021. 2 +[26] Skanda Koppula, Ignacio Rocco, Yi Yang, Joe Heyward, João Carreira, Andrew Zisserman, Gabriel Brostow, and Carl Doersch. Tapvid-3d: A benchmark for tracking any point in 3d, 2024. 7 +[27] Jiahui Lei, Yijia Weng, Adam W. Harley, Leonidas Guibas, and Kostas Daniilidis. MoSca: Dynamic gaussian fusion from casual videos via 4d motion scaffolds. arXiv preprint arXiv:2405.17421, 2024. 3, 6 +[28] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. EPnP: An accurate O(n) solution to the PnP problem. IJCV, 81:155-166, 2009. 5 +[29] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. In European Conference on Computer Vision, pages 71-91. Springer, 2024. 5 +[30] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5521-5531, 2022. 3 +[31] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 3 +[32] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4273-4284, 2023. 3 +[33] Zhengqi Li, Richard Tucker, Forrester Cole, Qianqian Wang, Linyi Jin, Vickie Ye, Angjoo Kanazawa, Aleksander Holynski, and Noah Snavelly. MegaSaM: Accurate, fast, and robust structure and motion from casual dynamic videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 2 +[34] Qingming Liu, Yuan Liu, Jiepeng Wang, Xianqiang Lyv, Peng Wang, Wenping Wang, and Junhui Hou. Modgs: Dynamic gaussian splattering from casually-captured monocular videos, 2024. 3 +[35] Bruce D Lucas and Takeo Kanade. An iterative image registration technique with an application to stereo vision. In IJCAI'81: 7th international joint conference on Artificial intelligence, pages 674-679, 1981. 2 +[36] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis, 2023. 3 +[37] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: A versatile and accurate monocular slam system. IEEE transactions on robotics, 31(5):1147-1163, 2015. 2 +[38] Tuan Duc Ngo, Peiye Zhuang, Chuang Gan, Evangelos Kalogerakis, Sergey Tulyakov, Hsin-Ying Lee, and + +Chaoyang Wang. DELTA: Dense efficient long-range 3d tracking for any video. In International Conference on Learning Representations (ICLR), 2025. 1, 3 +[39] David Novotny, Nikhila Ravi, Benjamin Graham, Natalia Neverova, and Andrea Vedaldi. C3dpo: Canonical 3d pose networks for non-rigid structure from motion. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7688-7697, 2019. 3 +[40] Xiaqing Pan, Nicholas Charron, Yongqian Yang, Scott Peters, Thomas Whelan, Chen Kong, Omkar Parkhi, Richard Newcombe, and Carl Yuheng Ren. Aria digital twin: A new benchmark dataset for egocentric 3d machine perception, 2023. 7 +[41] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), 2021. 3 +[42] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Undepth: Universal monocular metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10106-10116, 2024. 2 +[43] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10318–10327, 2021. 3 +[44] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. 2 +[45] Michael Rubinstein, Ce Liu, and William T Freeman. Towards longer long-range motion trajectories. 2012. 2 +[46] P. Sand and S. Teller. Particle video: Long-range motion estimation using point trajectories. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), pages 2195-2202, 2006. 2 +[47] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 2 +[48] Steven M Seitz, Brian Curless, James Diebel, Daniel Scharstein, and Richard Szeliski. A comparison and evaluation of multi-view stereo reconstruction algorithms. In 2006 IEEE computer society conference on computer vision and pattern recognition (CVPR'06), pages 519-528. IEEE, 2006. 2 +[49] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of RGB-D SLAM systems. pages 573-580, 2012. 8 +[50] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8934-8943, 2018. 2 +[51] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow, 2020. 1, 2 + +[52] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 34:16558-16569, 2021. 2 +[53] Zachary Teed and Jia Deng. Raft-3d: Scene flow using rigid-motion embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8375–8384, 2021. 2 +[54] Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. IEEE Transactions on Pattern Analysis & Machine Intelligence, 13(04): 376-380, 1991. 7 +[55] Sundar Vedula, Simon Baker, Peter Rander, Robert Collins, and Takeo Kanade. Three-dimensional scene flow. In Proceedings of the Seventh IEEE International Conference on Computer Vision, pages 722-729. IEEE, 1999. 2 +[56] Jianyuan Wang, Nikita Karaev, Christian Rupprecht, and David Novotny. Vggsfm: Visual geometry grounded deep structure from motion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21686-21697, 2024. 2 +[57] Qianqian Wang, Vickie Ye, Hang Gao, Weijia Zeng, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. 3, 6 +[58] Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A. Efros, and Angjoo Kanazawa. Continuous 3d perception model with persistent state. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 1, 2, 8 +[59] Ruicheng Wang, Sicheng Xu, Cassie Dai, Jianfeng Xiang, Yu Deng, Xin Tong, and Jiaolong Yang. Moge: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision, 2024. 2, 5, 6 +[60] Sen Wang, Ronald Clark, Hongkai Wen, and Niki Trigoni. Deepvo: Towards end-to-end visual odometry with deep recurrent convolutional neural networks. In 2017 IEEE international conference on robotics and automation (ICRA), pages 2043-2050. IEEE, 2017. 2 +[61] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jérôme Revaud. Dust3r: Geometric 3d vision made easy. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20697-20709, 2023. 2, 3, 5 +[62] Endre Weiszfeld. Sur le point pour lequel la somme des distances de n points données est minimum. Tohoku Mathematical Journal, First Series, 43:355-386, 1937. 5 +[63] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20310-20320, 2024. 3 +[64] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. Spatialtracker: Tracking any 2d pixels in 3d space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 7 + +[65] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 2 +[66] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20331-20341, 2024. 3 +[67] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. MonST3R: A simple approach for estimating geometry in the presence of motion. In International Conference on Learning Representations (ICLR), 2025. 1, 2, 7, 8 +[68] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T. Freeman. Structure and motion from casual videos. In Proceedings of the European Conference on Computer Vision (ECCV), pages 20-37, 2022. 2 +[69] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19855-19865, 2023. 5, 6, 8 +[70] Xiaowei Zhou, Menglong Zhu, Spyridon Leonardos, Konstantinos G Derpanis, and Kostas Daniilidis. Sparseness meets deepness: 3d human pose estimation from monocular video. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4966-4975, 2016. 3 \ No newline at end of file diff --git a/2025/St4RTrack_ Simultaneous 4D Reconstruction and Tracking in the World/images.zip b/2025/St4RTrack_ Simultaneous 4D Reconstruction and Tracking in the World/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..284d65d6d9230f4ff6dd207202e19817b4ccb161 --- /dev/null +++ b/2025/St4RTrack_ Simultaneous 4D Reconstruction and Tracking in the World/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a05f645e4f2be6e832112860c3e8e093cc6405043c425e9a94e52b4d93dd69f +size 556127 diff --git a/2025/St4RTrack_ Simultaneous 4D Reconstruction and Tracking in the World/layout.json b/2025/St4RTrack_ Simultaneous 4D Reconstruction and Tracking in the World/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..05e61258ccbbcd5c7090f81331fc7f5757e5bbd9 --- /dev/null +++ b/2025/St4RTrack_ Simultaneous 4D Reconstruction and Tracking in the World/layout.json @@ -0,0 +1,10127 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 82, + 67, + 529, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 67, + 529, + 85 + ], + "spans": [ + { + "bbox": [ + 82, + 67, + 529, + 85 + ], + "type": "text", + "content": "St4RTrack: Simultaneous 4D Reconstruction and Tracking in the World" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 106, + 515, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 106, + 515, + 137 + ], + "spans": [ + { + "bbox": [ + 96, + 106, + 515, + 137 + ], + "type": "text", + "content": "Haiwen Feng" + }, + { + "bbox": [ + 96, + 106, + 515, + 137 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 96, + 106, + 515, + 137 + ], + "type": "text", + "content": " * Junyi Zhang" + }, + { + "bbox": [ + 96, + 106, + 515, + 137 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 96, + 106, + 515, + 137 + ], + "type": "text", + "content": " * Qianqian Wang" + }, + { + "bbox": [ + 96, + 106, + 515, + 137 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 96, + 106, + 515, + 137 + ], + "type": "text", + "content": " Yufei Ye" + }, + { + "bbox": [ + 96, + 106, + 515, + 137 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 96, + 106, + 515, + 137 + ], + "type": "text", + "content": " Pengcheng Yu" + }, + { + "bbox": [ + 96, + 106, + 515, + 137 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 106, + 515, + 137 + ], + "type": "text", + "content": " Michael J. Black" + }, + { + "bbox": [ + 96, + 106, + 515, + 137 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 106, + 515, + 137 + ], + "type": "text", + "content": " Trevor Darrell" + }, + { + "bbox": [ + 96, + 106, + 515, + 137 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 96, + 106, + 515, + 137 + ], + "type": "text", + "content": " Angjoo Kanazawa" + }, + { + "bbox": [ + 96, + 106, + 515, + 137 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 97, + 142, + 515, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 142, + 515, + 159 + ], + "spans": [ + { + "bbox": [ + 97, + 142, + 515, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 142, + 515, + 159 + ], + "type": "text", + "content": "UC Berkeley " + }, + { + "bbox": [ + 97, + 142, + 515, + 159 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 142, + 515, + 159 + ], + "type": "text", + "content": "Max Planck Institute for Intelligent Systems " + }, + { + "bbox": [ + 97, + 142, + 515, + 159 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 97, + 142, + 515, + 159 + ], + "type": "text", + "content": "Stanford University" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 56, + 171, + 555, + 338 + ], + "blocks": [ + { + "bbox": [ + 56, + 171, + 555, + 338 + ], + "lines": [ + { + "bbox": [ + 56, + 171, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 171, + 555, + 338 + ], + "type": "image", + "image_path": "e4cfba1126e92c9e355bab8c7db34ddf1f91ab4871227ed3126373ba26c8f337.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 349, + 555, + 407 + ], + "lines": [ + { + "bbox": [ + 54, + 349, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 54, + 349, + 555, + 407 + ], + "type": "text", + "content": "Figure 1. St4RTrack: Given an RGB video capturing dynamic scenes, St4RTrack simultaneously tracks the points from the initial frame (visualized in purple) and reconstructs the geometry of the subsequent frames (in orange) in a consistent world coordinate frame. St4RTrack is a feed-forward framework that takes a pair of images as input and outputs two pointmaps in the world frame, as the visualization shown in the middle. By iteratively processing the first frame paired with each subsequent frame, St4RTrack achieves simultaneous tracking (right bottom) and reconstruction (right top) for the entire video. Interactive results on our webpage: https://St4RTrack.github.io/." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 425, + 201, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 425, + 201, + 437 + ], + "spans": [ + { + "bbox": [ + 151, + 425, + 201, + 437 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 448, + 297, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 448, + 297, + 688 + ], + "spans": [ + { + "bbox": [ + 54, + 448, + 297, + 688 + ], + "type": "text", + "content": "Dynamic 3D reconstruction and point tracking in videos are typically treated as separate tasks, despite their deep connection. We propose St4RTrack, a feed-forward framework that simultaneously reconstructs and tracks dynamic video content in a world coordinate frame from RGB inputs. This is achieved by predicting two appropriately defined pointmaps for a pair of frames captured at different moments. Specifically, we predict both pointmaps at the same moment, in the same world, capturing both static and dynamic scene geometry while maintaining 3D correspondences. Chaining these predictions through the video sequence with respect to a reference frame naturally computes long-range correspondences, effectively combining 3D reconstruction with 3D tracking. Unlike prior methods that rely heavily on 4D ground truth supervision, we employ a novel adaptation scheme based on a reprojection loss. We establish a new extensive benchmark for world-frame reconstruction and tracking, demonstrating the effectiveness and efficiency of our unified, data-driven framework. Our code, model, and benchmark will be released." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 425, + 395, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 425, + 395, + 437 + ], + "spans": [ + { + "bbox": [ + 315, + 425, + 395, + 437 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "content": "When asked about the three most important problems in computer vision, Takeo Kanade replied, \"Correspondence, Correspondence, Correspondence!\" This remark is especially pertinent to multi-view 3D reconstruction, where 3D geometry and 2D correspondence are two sides of the same coin; that is, a 3D point in the physical world naturally brings 2D correspondence across its projections in different views, and conversely, corresponded 2D points across views reconstruct the same 3D point after triangulation. This synergy between 3D geometry and 2D correspondence has long formed the foundation of multi-view geometry [17]. However, when the scene becomes dynamic, this synergy appears to vanish, as existing methods—particularly the recent data-driven ones—tend to treat dynamic reconstruction [22, 58, 67] and correspondence [24, 38, 51] as separate, disconnected tasks. We argue that this is a missed opportunity; the synergy between 3D reconstruction and 2D correspondence is not lost in dynamic scenes—it simply requires an additional element: understanding how the scene content evolves over time. This evolution is captured by 3D motion estimation (i.e., dense 3D point tracking), which, when computed across a sequence, effectively ex" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 206, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 206, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 206, + 713 + ], + "type": "text", + "content": "*Equal contribution, listed alphabetically." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8503" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "plains scene motion. Once tracking accounts for scene dynamics, the problem effectively reduces to the rigid case, where the natural interplay between reconstruction and correspondence can once again be leveraged." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 122, + 295, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 122, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 54, + 122, + 295, + 266 + ], + "type": "text", + "content": "We propose St4RTrack, a learning framework that unifies reconstruction and tracking directly from RGB video frames. St4RTrack simultaneously reconstructs and tracks dynamic video content in a single consistent world, achieving world-frame 3D tracking, as demonstrated in Fig. 1. Tracking in the world frame decouples the scene and the camera motion, essential for domains where both the camera and content are in motion. Our approach also reconstructs the 3D geometry of observed image content for both static and dynamic portions of the scene. St4RTrack directly predicts their reconstruction and tracking in the world without requiring an additional alignment stage." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 269, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 269, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 269, + 295, + 555 + ], + "type": "text", + "content": "Our key insight stems from the observation that a feedforward static 3D reconstruction method, DUSt3R [61], can be adapted to dynamic scenes simply by changing the pointmaps' annotation [67]. Building on this, we reexamine the pointmaps definition in the 4D scenario and opt to simply redefine its geometric interpretation for both reconstruction and tracking, as illustrated in Fig. 2. Concretely, we achieve it by predicting two pointmaps at the same timestamp and in the same world from a pair of image frames depicting two different timestamps. More specific, given images " + }, + { + "bbox": [ + 56, + 269, + 295, + 555 + ], + "type": "inline_equation", + "content": "(\\mathbf{I}_i,\\mathbf{I}_j)" + }, + { + "bbox": [ + 56, + 269, + 295, + 555 + ], + "type": "text", + "content": ", both pointmaps are predicted in the coordinate frame of " + }, + { + "bbox": [ + 56, + 269, + 295, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 56, + 269, + 295, + 555 + ], + "type": "text", + "content": ", but at the time specified by " + }, + { + "bbox": [ + 56, + 269, + 295, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_j" + }, + { + "bbox": [ + 56, + 269, + 295, + 555 + ], + "type": "text", + "content": ". Our method is realized through a feed-forward network comprising of two branches: the reconstruction branch reconstructs the content of " + }, + { + "bbox": [ + 56, + 269, + 295, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_j" + }, + { + "bbox": [ + 56, + 269, + 295, + 555 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 56, + 269, + 295, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 56, + 269, + 295, + 555 + ], + "type": "text", + "content": " coordinate frame; and the tracking branch, which reconstructs the content of " + }, + { + "bbox": [ + 56, + 269, + 295, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 56, + 269, + 295, + 555 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 56, + 269, + 295, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 56, + 269, + 295, + 555 + ], + "type": "text", + "content": " (its own) coordinate frame, but at the time indicated by " + }, + { + "bbox": [ + 56, + 269, + 295, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_j" + }, + { + "bbox": [ + 56, + 269, + 295, + 555 + ], + "type": "text", + "content": ". Essentially, the tracking branch predicts how the scene content in " + }, + { + "bbox": [ + 56, + 269, + 295, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 56, + 269, + 295, + 555 + ], + "type": "text", + "content": " evolves to match the moment captured in " + }, + { + "bbox": [ + 56, + 269, + 295, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_j" + }, + { + "bbox": [ + 56, + 269, + 295, + 555 + ], + "type": "text", + "content": ". This is enabled through a DUSt3R-like dual cross-attention mechanism, where the tracking branch relies on the reconstruction branch to decide how to move points. This minimal change proves sufficient for unifying both dynamic reconstruction and 3D point tracking in the world coordinate system." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": "Furthermore, unlike existing methods [58, 61, 67] that rely solely on 4D supervision, our approach unlocks 4D reconstruction training on in-the-wild videos via reprojection loss without 4D supervision. This is possible because St4RTrack simultaneously establishes camera parameters, 3D geometry and motion. Specifically, based on the outputs of the reconstruction branch, the camera parameters for " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_j" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": " can be differentiably computed via PnP. Using these cameras, the pointmap of " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": " is projected into the " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": "-th frame, enabling training with a reprojection loss that leverages 2D correspondences and monocular depth predictions from off-the-shelf approaches [23, 59]. Consequently, the monocular supervisions facilitate effective test-time adaptation of" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "St4RTrack to in-the-wild videos, which can differ substantially from the synthetic data used during pretraining." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 98, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 98, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 98, + 553, + 240 + ], + "type": "text", + "content": "While prior 3D point tracking benchmarks focus on camera coordinate frames [64], our approach enables worldframe 3D tracking. To evaluate this capability, we establish a novel benchmark, WorldTrack, for both tracking and reconstruction in the world coordinate system. We find that our unified method outperforms the strong baselines that combine several pieces on each individual task. Furthermore, we show that our feedforward results can be improved via test-time adaptation. We believe this is a step towards a unified task-agnostic 4D perception system that can be trained on a large-scale video. Our code, model, and benchmark will be released." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 255, + 406, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 255, + 406, + 268 + ], + "spans": [ + { + "bbox": [ + 314, + 255, + 406, + 268 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 277, + 555, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 601 + ], + "type": "text", + "content": "Camera Estimation and Scene Reconstruction. Jointly estimating camera motion and scene geometry has been studied for decades, often in the context of Structure from Motion (SfM) [1, 47, 48, 56] or Simultaneous Localization and Mapping (SLAM) [7, 10, 37, 52, 60]. However, these methods are primarily designed for static scenes and typically do not model dynamic scene content. Recent advances in learning-based monocular and video depth estimation methods [2, 42, 44, 65] have opened new opportunities to reconstruct dynamic scenes. Notably, R-CVD [25], CasualSAM [68] and MegaSAM [33] jointly optimize camera parameters and per-frame dense depth maps leveraging monocular depth priors, producing consistent depth estimates for dynamic objects and accurate camera parameters even in challenging cases with minimal camera parallax. Another notable recent method, DUSt3R [61], introduces a two-pointmap representation that enables joint estimation of camera motion and scene geometry of a pair of images. While DUSt3R itself primarily focuses on reconstructing static scenes, follow-up effort such as MonST3R [67], demonstrate that this formulation can also effectively handle dynamic scenes with minimal modification on supervisions. Despite these advances, none of the aforementioned methods explicitly estimate 3D scene motion, meaning they do not track the movement of individual 3D points over time. In contrast, our method simultaneously performs joint reconstruction and tracking for dynamic scenes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "2D/3D Point Tracking. Tracking pixel motion over time is a fundamental problem in computer vision. Optical/Scene flow methods [3, 18, 19, 35, 50, 51, 53, 55] produce dense 2D/3D motion vectors but are inherently short-ranged, struggling with large displacements and occlusions. While long-range point tracking [45, 46] has been studied for decades, it has recently been revitalized via supervised learning [8, 9, 16, 23, 24], enabling more robust tracking over extended time periods and overcoming these limita-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8504" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 114, + 175 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 114, + 175 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 114, + 175 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 114, + 175 + ], + "type": "image", + "image_path": "41aa5171dca6774fdbdd14db755bf094b7fc028fe61fc026cd8563dd1c1ac99d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 187, + 295, + 254 + ], + "lines": [ + { + "bbox": [ + 55, + 187, + 295, + 254 + ], + "spans": [ + { + "bbox": [ + 55, + 187, + 295, + 254 + ], + "type": "text", + "content": "Figure 2. Pointmap Comparison of MonST3R and St4RTrack. Given two input frames, MonST3R handles dynamic scenes by reconstructing both pointmaps in their own timestamp. St4RTrack predicts where the points in the first frame move in the second frame, and reconstructs the geometry of the second frame. More details of the representation definition are introduced in Sec. 3.1." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 126, + 72, + 203, + 175 + ], + "blocks": [ + { + "bbox": [ + 126, + 72, + 203, + 175 + ], + "lines": [ + { + "bbox": [ + 126, + 72, + 203, + 175 + ], + "spans": [ + { + "bbox": [ + 126, + 72, + 203, + 175 + ], + "type": "image", + "image_path": "f1e32e9c6ab572e2e4767e4091e3f9e9987ef8eb4f791ef2a30f977db9cd9e45.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 215, + 72, + 291, + 175 + ], + "blocks": [ + { + "bbox": [ + 215, + 72, + 291, + 175 + ], + "lines": [ + { + "bbox": [ + 215, + 72, + 291, + 175 + ], + "spans": [ + { + "bbox": [ + 215, + 72, + 291, + 175 + ], + "type": "image", + "image_path": "f645ec427d3a3c046c2467d50953e0b115cd37cb432f1f2b22444b492e6249ca.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 277, + 296, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 296, + 408 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 296, + 408 + ], + "type": "text", + "content": "tions. However, these methods still produce only 2D pixel trajectories. More recently, several works [38, 64] achieve 3D tracking by lifting points into 3D space using monocular depth priors and performing tracking in 3D. While closely related to our approach, these methods still operate in the camera frame space, meaning they lack camera motion estimation and do not explicitly separate scene motion from camera motion. In contrast, our method jointly estimates disentangled camera and scene motion, enabling world-space tracking for a more complete understanding of 3D scene dynamics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 415, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 714 + ], + "type": "text", + "content": "Joint Dynamic Reconstruction and Tracking. While traditional non-rigid SfMs have studied joint tracking and reconstruction [4, 6, 12, 39, 70] from 2D correspondences, jointly optimizing them from raw videos is highly challenging and typically requires multi-view synchronized videos as input [13, 30, 36, 41, 63]. With recent advances in neural rendering and data-driven geometric priors, reconstructing and rendering dynamic scenes from monocular videos became possible. However, as Gao et al. [14] point out, many methods [43, 66] focus on \"teleporting\" input data, which are effectively multi-view and not representative of real-world videos. In addition, since the main focus is view synthesis, motion estimation serves a secondary role in facilitating information fusion between neighboring frames [31, 32]. More recently, several works [27, 34, 57] focus on jointly recovering camera parameters, persistent scene geometry, and long-range 3D tracks from single, causally captured videos. However, these methods take off-the-shelf priors as given and design per-video optimization techniques that optimize a representation from scratch. Most recently, Stereo4D [20]—a concurrent effort to our work—proposed a pipeline for crafting a real-world 4D tracking dataset using internet stereo videos, enabling the supervised regression of 3D trajectories and geometries between frames. In contrast, we propose a feed-forward" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "method that simultaneously performs reconstruction and tracking, while the same architecture also supports test-time adaptation on unlabeled videos to approach the high quality of optimization-based methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 133, + 552, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 552, + 147 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 552, + 147 + ], + "type": "text", + "content": "3. Simultaneous Reconstruction and Tracking" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 153, + 555, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 153, + 555, + 345 + ], + "spans": [ + { + "bbox": [ + 313, + 153, + 555, + 345 + ], + "type": "text", + "content": "We present a framework that simultaneously reconstructs and tracks dynamic video content in 3D within a single consistent world coordinate frame. The core idea is simple yet powerful: reconstructing and tracking can both be achieved by predicting two appropriately defined pointmaps, where both pointmaps reconstruct the scene content observed in each image at the same timestamp in a consistent coordinate system. This enables simultaneous reconstruction of both dynamic and static contents, while tracking across a sequence of image frames in a video. Since all geometry, camera, and motion (i.e. 3D correspondence over time) can be derived from the representation, it can be adapted to videos without any explicit 4D supervision. Below, we discuss the main insight, how St4RTrack compares to prior works. Then, we discuss the details of the model and how it can be trained and adapted to videos." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 354, + 526, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 526, + 367 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 526, + 367 + ], + "type": "text", + "content": "3.1. Unified 4D Representation of St4RTrack" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 372, + 554, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 554, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 554, + 480 + ], + "type": "text", + "content": "Given two images " + }, + { + "bbox": [ + 313, + 372, + 554, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i, \\mathbf{I}_j" + }, + { + "bbox": [ + 313, + 372, + 554, + 480 + ], + "type": "text", + "content": " with dynamic content (see Figure 2), how can one devise a single feedforward approach that simultaneously performs reconstruction and tracking? We argue that the underlying representation must (1) capture camera motion to establish a world coordinate frame, (2) reconstruct the 3D geometry of all observed points, and (3) estimate 3D motion that maintain explicit correspondence over time. In this work, we show that just two properly defined pointmaps suffice to fulfill these requirements." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 485, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 554, + 605 + ], + "type": "text", + "content": "Time-Dependent Pointmap. A pointmap representation assumes that each pixel in an image " + }, + { + "bbox": [ + 313, + 485, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 313, + 485, + 554, + 605 + ], + "type": "text", + "content": " of shape " + }, + { + "bbox": [ + 313, + 485, + 554, + 605 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 313, + 485, + 554, + 605 + ], + "type": "text", + "content": " is associated with a corresponding 3D point, forming a pointmap " + }, + { + "bbox": [ + 313, + 485, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{X}\\in \\mathbb{R}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 313, + 485, + 554, + 605 + ], + "type": "text", + "content": ". For the case of static scenes, DUSt3R [61] only considers two factors of pointmaps—(1) the source frame of the content and (2) the camera coordinate system in which the points are expressed. However, this definition is insufficient for modeling the dynamic scenario of monocular video. To address this, we introduce a previously overlooked yet decisive factor: time." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "content": "Specifically, we define a time-dependent pointmap that encodes the 3D positions of the scene points in a chosen (world) coordinate system at a specific timestamp. For clarity, we denote this representation as " + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "inline_equation", + "content": "{}^a\\mathbf{X}_t^b" + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "content": ", which denotes the 3D pointmap of physical content from frame " + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "content": ", at time " + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "content": ", expressed in the coordinate system established by frame " + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "content": ". For example, " + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "inline_equation", + "content": "{}^i\\mathbf{X}_j^i" + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "content": " represents the geometry originally seen in frame " + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "content": " in frame " + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "content": "'s own coordinate system, but described at timestamp " + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "content": ". The time-dependency is achieved by construc" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8505" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 53, + 541, + 230 + ], + "blocks": [ + { + "bbox": [ + 65, + 53, + 541, + 230 + ], + "lines": [ + { + "bbox": [ + 65, + 53, + 541, + 230 + ], + "spans": [ + { + "bbox": [ + 65, + 53, + 541, + 230 + ], + "type": "image", + "image_path": "838ff1e8eeefd1c74338805156461b5e20fed63826feab1939d84f794bab6951.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 242, + 555, + 308 + ], + "lines": [ + { + "bbox": [ + 54, + 242, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 54, + 242, + 555, + 308 + ], + "type": "text", + "content": "Figure 3. Overview of St4RTrack. Given frame 1 and frame " + }, + { + "bbox": [ + 54, + 242, + 555, + 308 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 54, + 242, + 555, + 308 + ], + "type": "text", + "content": " as input, the tracking branch outputs " + }, + { + "bbox": [ + 54, + 242, + 555, + 308 + ], + "type": "inline_equation", + "content": "{}^{1}\\mathbf{X}_{j}^{1}" + }, + { + "bbox": [ + 54, + 242, + 555, + 308 + ], + "type": "text", + "content": ", the pointmap that corresponds to observed content of the first frame at timestep " + }, + { + "bbox": [ + 54, + 242, + 555, + 308 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 54, + 242, + 555, + 308 + ], + "type": "text", + "content": " in its own camera coordinate (i.e. world coordinate); the reconstruction branch outputs " + }, + { + "bbox": [ + 54, + 242, + 555, + 308 + ], + "type": "inline_equation", + "content": "{}^{1}\\mathbf{X}_{j}^{j}" + }, + { + "bbox": [ + 54, + 242, + 555, + 308 + ], + "type": "text", + "content": ", the pointmap of the content in frame " + }, + { + "bbox": [ + 54, + 242, + 555, + 308 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 54, + 242, + 555, + 308 + ], + "type": "text", + "content": " at its own timestamp in the world coordinate. To adapt to new videos without any 4D labels, the camera is computed via differentiable PnP from the pointmap, enabling reprojected supervision signals (e.g., 2D trajectories and monocular depth). We finetune both branches during training (Sec. 3.2) with synthetic data, and when adapting to a new video (Sec. 3.3), only the tracking branch is fine-tuned using these reprojected supervision signals." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 319, + 295, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 295, + 343 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 295, + 343 + ], + "type": "text", + "content": "tion without explicit timestamp conditioning, as described in the next section." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 349, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 296, + 384 + ], + "type": "text", + "content": "Unified 4D Modeling. St4RTrack learns a function " + }, + { + "bbox": [ + 55, + 349, + 296, + 384 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 349, + 296, + 384 + ], + "type": "text", + "content": " that maps two images " + }, + { + "bbox": [ + 55, + 349, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i, \\mathbf{I}_j" + }, + { + "bbox": [ + 55, + 349, + 296, + 384 + ], + "type": "text", + "content": ", captured at timestamp " + }, + { + "bbox": [ + 55, + 349, + 296, + 384 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 349, + 296, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 349, + 296, + 384 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 349, + 296, + 384 + ], + "type": "text", + "content": ", into two pointmaps:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 128, + 394, + 295, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 394, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 128, + 394, + 295, + 410 + ], + "type": "interline_equation", + "content": "f _ {\\theta} \\left(\\mathbf {I} _ {i}, \\mathbf {I} _ {j}\\right) = ^ {i} \\mathbf {X} _ {j} ^ {i}, ^ {i} \\mathbf {X} _ {j} ^ {j}. \\tag {1}", + "image_path": "5287f11bd6ad76209d5118f731e12982793112fd4e2e71b37267a9e26a56880e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 420, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 296, + 491 + ], + "type": "text", + "content": "The second image " + }, + { + "bbox": [ + 55, + 420, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_j" + }, + { + "bbox": [ + 55, + 420, + 296, + 491 + ], + "type": "text", + "content": " is reconstructed as the pointmap " + }, + { + "bbox": [ + 55, + 420, + 296, + 491 + ], + "type": "inline_equation", + "content": "{}^i\\mathbf{X}_j^j" + }, + { + "bbox": [ + 55, + 420, + 296, + 491 + ], + "type": "text", + "content": " in the first image " + }, + { + "bbox": [ + 55, + 420, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 55, + 420, + 296, + 491 + ], + "type": "text", + "content": "'s coordinate frame. Meanwhile, it predicts " + }, + { + "bbox": [ + 55, + 420, + 296, + 491 + ], + "type": "inline_equation", + "content": "{}^i\\mathbf{X}_j^i" + }, + { + "bbox": [ + 55, + 420, + 296, + 491 + ], + "type": "text", + "content": ", representing the 3D motion of how the content from the first image " + }, + { + "bbox": [ + 55, + 420, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 55, + 420, + 296, + 491 + ], + "type": "text", + "content": " moves at timestamp " + }, + { + "bbox": [ + 55, + 420, + 296, + 491 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 420, + 296, + 491 + ], + "type": "text", + "content": ". Thus, both geometry and motion (tracking) are estimated from this unified prediction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 493, + 296, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 296, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 296, + 661 + ], + "type": "text", + "content": "To handle a full video consisting of " + }, + { + "bbox": [ + 55, + 493, + 296, + 661 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 493, + 296, + 661 + ], + "type": "text", + "content": " frames, we perform tracking and reconstruction by always selecting the first frame as the anchor frame " + }, + { + "bbox": [ + 55, + 493, + 296, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 55, + 493, + 296, + 661 + ], + "type": "text", + "content": ". Each subsequent frame " + }, + { + "bbox": [ + 55, + 493, + 296, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_j" + }, + { + "bbox": [ + 55, + 493, + 296, + 661 + ], + "type": "text", + "content": " is then paired with this initial frame, ensuring that every new frame is consistently aligned to the coordinate system of the first frame. Specifically, " + }, + { + "bbox": [ + 55, + 493, + 296, + 661 + ], + "type": "inline_equation", + "content": "\\{f(\\mathbf{I}_1,\\mathbf{I}_1),f(\\mathbf{I}_1,\\mathbf{I}_2),\\ldots ,f(\\mathbf{I}_1,\\mathbf{I}_T)\\}" + }, + { + "bbox": [ + 55, + 493, + 296, + 661 + ], + "type": "text", + "content": " are computed in the same reference, " + }, + { + "bbox": [ + 55, + 493, + 296, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_1" + }, + { + "bbox": [ + 55, + 493, + 296, + 661 + ], + "type": "text", + "content": ", which naturally serves as the world coordinate frame. Thus, world-frame 3D tracking is achieved by explicitly following how points observed in " + }, + { + "bbox": [ + 55, + 493, + 296, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_1" + }, + { + "bbox": [ + 55, + 493, + 296, + 661 + ], + "type": "text", + "content": " are placed throughout the sequence, " + }, + { + "bbox": [ + 55, + 493, + 296, + 661 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{^1X_1^1},\\mathbf{^1X_2^1},\\dots,\\mathbf{^1X_T^1}\\}" + }, + { + "bbox": [ + 55, + 493, + 296, + 661 + ], + "type": "text", + "content": ", while the world frame dynamic reconstruction is obtained by the paired geometry estimation per-frame, " + }, + { + "bbox": [ + 55, + 493, + 296, + 661 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{^1X_1^1},\\mathbf{^1X_2^2},\\dots,\\mathbf{^1X_T^T}\\}" + }, + { + "bbox": [ + 55, + 493, + 296, + 661 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Relation to prior works. Our formulation of the 4D modeling generalizes prior works in a unified framework. DUSt3R reconstructs and establishes correspondences but is limited to rigid scenes, as correspondence and recon" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 319, + 555, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 380 + ], + "type": "text", + "content": "struction are dual tasks for static scenes. With this perspective, one can see that if there is no dynamic component (i.e. frozen moment in time or rigid scenes), our formulation is equivalent to DUSt3R, where both images share the same timestamp " + }, + { + "bbox": [ + 313, + 319, + 555, + 380 + ], + "type": "inline_equation", + "content": "t = i" + }, + { + "bbox": [ + 313, + 319, + 555, + 380 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 388, + 389, + 555, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 389, + 555, + 405 + ], + "spans": [ + { + "bbox": [ + 388, + 389, + 555, + 405 + ], + "type": "interline_equation", + "content": "f _ {\\theta} (\\mathbf {I} _ {i}, \\mathbf {I} _ {j}) = ^ {i} \\mathbf {X} _ {i} ^ {i}, ^ {i} \\mathbf {X} _ {i} ^ {j}. \\qquad (2)", + "image_path": "61e18eb8e5ce5f01b6b2149883e6e5c72ba1826ac46f3f75766d3dc1007a8181.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 415, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 510 + ], + "type": "text", + "content": "In a static world, 3D reconstruction from two pointmaps inherently yields the correspondences between them, allowing synergy to arise naturally. However, when objects or the scene are in motion, the dynamic component appears differently in different frames, it becomes crucial to account for 3D scene motion to preserve this synergy. St4RTrack addresses this challenge by predicting the 3D content from the first image at future timestamps." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 511, + 554, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 511, + 554, + 535 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 554, + 535 + ], + "type": "text", + "content": "In the same framework, we see that MonST3R, the dynamic follow-up of DUSt3R can be expressed as such:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 388, + 544, + 555, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 544, + 555, + 561 + ], + "spans": [ + { + "bbox": [ + 388, + 544, + 555, + 561 + ], + "type": "interline_equation", + "content": "f _ {\\theta} (\\mathbf {I} _ {i}, \\mathbf {I} _ {j}) = ^ {i} \\mathbf {X} _ {i} ^ {i}, ^ {i} \\mathbf {X} _ {j} ^ {j}, \\tag {3}", + "image_path": "78be9d929f029d395b432cc82d261622fdc4ef076ed9095e6fa9a10b249ef40c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 569, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 554, + 653 + ], + "type": "text", + "content": "where each image's 3D geometry is reconstructed in its timestamp, such that the dynamic contents separately align with their frame inputs. While it's sufficient for obtaining dynamic scene geometry, there is no temporal correspondence being established, as illustrated in Fig. 2. Furthermore, both DUSt3R and MonST3R compute the pairwise graphs and perform global alignment." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "content": "Since we always designate the first frame as the reference for tracking, the world coordinates are consistently established by the first frame. For simplicity, we omit the explicit notation of the world coordinate in subsequent equations and paragraphs, i.e., " + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_j^i \\coloneqq {}^i\\mathbf{X}_j^i" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8506" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 85 + ], + "type": "text", + "content": "3.2. Joint Learning of Tracking and Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 233 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 233 + ], + "type": "text", + "content": "In this section, we describe how our framework implements equation 1 within a pair-wise framework as DUSt3R. For each pair of frames, " + }, + { + "bbox": [ + 55, + 90, + 296, + 233 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_1" + }, + { + "bbox": [ + 55, + 90, + 296, + 233 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 90, + 296, + 233 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_j" + }, + { + "bbox": [ + 55, + 90, + 296, + 233 + ], + "type": "text", + "content": ", we first encode them into token representations using a ViT encoder, then process these tokens through a siamese transformer decoder. The decoder sequentially applies self-attention (allowing tokens within each frame to interact), followed by cross-attention (enabling tokens from one frame to attend to tokens in the other), and finally passes the tokens through an MLP. This continuous information flow between the two branches is crucial for generating spatial-aligned 3D pointmaps in a shared coordinate system, as illustrated in Fig. 3." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 235, + 296, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 235, + 296, + 294 + ], + "spans": [ + { + "bbox": [ + 55, + 235, + 296, + 294 + ], + "type": "text", + "content": "Our siamese architecture processes two input views concurrently and generates two 3D pointmaps that are expressed in a common reference frame established by the first view. Although the two branches share the same architectural structure, they serve distinct purposes:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 295, + 295, + 366 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 295, + 295, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 295, + 331 + ], + "type": "text", + "content": "- Tracking branch predicts the pointmap " + }, + { + "bbox": [ + 55, + 295, + 295, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_j^1" + }, + { + "bbox": [ + 55, + 295, + 295, + 331 + ], + "type": "text", + "content": ", which represents the geometry of the first frame at timestamp " + }, + { + "bbox": [ + 55, + 295, + 295, + 331 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 295, + 295, + 331 + ], + "type": "text", + "content": " in the first frame's coordinates (i.e., the world coordinates)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 331, + 294, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 294, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 294, + 366 + ], + "type": "text", + "content": "- Reconstruction branch predicts the pointmap " + }, + { + "bbox": [ + 55, + 331, + 294, + 366 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_j^j" + }, + { + "bbox": [ + 55, + 331, + 294, + 366 + ], + "type": "text", + "content": ", which represents the geometry of frame " + }, + { + "bbox": [ + 55, + 331, + 294, + 366 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 331, + 294, + 366 + ], + "type": "text", + "content": " at its own timestamp, also expressed in the first frame's camera coordinates." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 368, + 296, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 368, + 296, + 428 + ], + "spans": [ + { + "bbox": [ + 55, + 368, + 296, + 428 + ], + "type": "text", + "content": "Since this architecture is exactly the same as proposed by DUSt3R and subsequently adopted by MonST3R, with the only difference being the output paired pointmaps (Eq.1-3), our network can be initialized with pretrained 3D knowledge from either DUSt3R or MonST3R." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 434, + 296, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 296, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 296, + 649 + ], + "type": "text", + "content": "Pretraining with 4D Synthetic Data. Our proposed representation requires specialized supervision for the Tracking Branch—namely, ensuring that the pointmap from the first frame is correctly positioned in the world across all frames. Achieving this necessitates complete 4D information of the dynamic scene. Therefore, we leverage existing 4D synthetic datasets [22, 69] that provide both the 3D geometry and motion of the rendered content. Specifically, for each dataset, we use the scene mesh vertices (expressed in world coordinates) to provide sparse, masked supervision for the Tracking Branch pointmap representation, and employ per-frame depth maps and camera ground-truth to supervise the Reconstruction Branch. For this fully supervised training process, we use the objectives from DUSt3R. We initialize our dual-branch transformer with weights from MAST3R [29], a DUSt3R variant that has been adapted for 2D correspondence learning. Additional details regarding the 4D synthetic datasets are provided in Sec. 4.1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 659, + 256, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 256, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 256, + 672 + ], + "type": "text", + "content": "3.3. Adapt to Any Video without 4D Label" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "While the synthetic datasets are small-scale and unrealistic, they are sufficient for our network to learn the newly proposed representations. However, fully supervised training" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": "on these datasets presents two key limitations: 1) The 4D synthetic data is limited in scale and does not encompass the full range of motion and geometry present in real-world dynamic scenes; 2) Our proposed pointmap representation requires the capability to freely move the pointmap within the world coordinates—a departure from conventional pixel-aligned geometry predictions, making small-scale training insufficient for achieving fine-grained predictions. These limitations motivate us to further leverage the 3D geometry and motion inherent in the St4RTrack framework to perform domain adaptation on any video without 4D labels. Specifically, we first show how we can derive camera parameters differentially, and with which we can design reprojected 2D trajectory loss and monocular depth loss to supervise the network." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 257, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 257, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 555, + 437 + ], + "type": "text", + "content": "Solving Camera Parameters. The intrinsic matrix " + }, + { + "bbox": [ + 313, + 257, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 313, + 257, + 555, + 437 + ], + "type": "text", + "content": " is first estimated from the tracking branch's first-frame pointmap prediction, following DUSt3R [61]. In this process, the principal point is assumed to be centered, and pixels are treated as square. The focal length is assumed static across frames and estimated using a fast iterative solver based on the Weiszfeld algorithm [62]. Next, the extrinsic parameters " + }, + { + "bbox": [ + 313, + 257, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^j = [\\mathbf{R}^j | \\mathbf{T}^j]" + }, + { + "bbox": [ + 313, + 257, + 555, + 437 + ], + "type": "text", + "content": " for each frame " + }, + { + "bbox": [ + 313, + 257, + 555, + 437 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 257, + 555, + 437 + ], + "type": "text", + "content": " are derived using the \"reconstruction\" pointmap " + }, + { + "bbox": [ + 313, + 257, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_j^j" + }, + { + "bbox": [ + 313, + 257, + 555, + 437 + ], + "type": "text", + "content": ". Specifically, each pixel " + }, + { + "bbox": [ + 313, + 257, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{j,n}" + }, + { + "bbox": [ + 313, + 257, + 555, + 437 + ], + "type": "text", + "content": " in frame " + }, + { + "bbox": [ + 313, + 257, + 555, + 437 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 257, + 555, + 437 + ], + "type": "text", + "content": " is associated with a 3D coordinate " + }, + { + "bbox": [ + 313, + 257, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_j^{j,n}" + }, + { + "bbox": [ + 313, + 257, + 555, + 437 + ], + "type": "text", + "content": " in the shared world coordinate system (established by the first camera), thus forming 2D-to-3D correspondences. We could then solve for " + }, + { + "bbox": [ + 313, + 257, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^j" + }, + { + "bbox": [ + 313, + 257, + 555, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 257, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{T}^j" + }, + { + "bbox": [ + 313, + 257, + 555, + 437 + ], + "type": "text", + "content": " via a Perspective-" + }, + { + "bbox": [ + 313, + 257, + 555, + 437 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 257, + 555, + 437 + ], + "type": "text", + "content": "-Points (PnP) [28] solver with RANSAC [11] for outlier rejection:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 319, + 448, + 553, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 448, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 319, + 448, + 553, + 488 + ], + "type": "interline_equation", + "content": "\\mathbf {R} ^ {j}, \\mathbf {T} ^ {j} = \\underset {\\mathbf {R}, \\mathbf {T}} {\\operatorname {a r g m i n}} \\sum_ {n \\in \\mathcal {I} _ {j}} \\left\\| \\mathbf {x} ^ {j, n} - \\pi \\left(\\mathbf {K} \\left(\\mathbf {R} \\mathbf {X} _ {j} ^ {j, n} + \\mathbf {T}\\right)\\right) \\right\\| ^ {2}, \\tag {4}", + "image_path": "551d5255c6e38571ebb999e6a9cedd3bc0faaf7270998016f6a6667e014aac63.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 488, + 524, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 488, + 524, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 488, + 524, + 501 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 488, + 524, + 501 + ], + "type": "inline_equation", + "content": "\\pi (\\cdot)" + }, + { + "bbox": [ + 313, + 488, + 524, + 501 + ], + "type": "text", + "content": " is the projection " + }, + { + "bbox": [ + 313, + 488, + 524, + 501 + ], + "type": "inline_equation", + "content": "(x,y,z)\\to (x / z,y / z)" + }, + { + "bbox": [ + 313, + 488, + 524, + 501 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 502, + 553, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 553, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 553, + 550 + ], + "type": "text", + "content": "For differentiability, we adopt a derivative-based Gauss-Newton solver following [5], ensuring that gradients from the reprojection loss can adjust both the camera pose and the 3D pointmaps. Further details are provided in ??" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 555, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 554, + 689 + ], + "type": "text", + "content": "Reprojection Loss. With the camera pose of frame " + }, + { + "bbox": [ + 313, + 555, + 554, + 689 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 555, + 554, + 689 + ], + "type": "text", + "content": " derived, the tracking pointmap " + }, + { + "bbox": [ + 313, + 555, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_j^1" + }, + { + "bbox": [ + 313, + 555, + 554, + 689 + ], + "type": "text", + "content": " and reconstruction pointmap " + }, + { + "bbox": [ + 313, + 555, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_j^j" + }, + { + "bbox": [ + 313, + 555, + 554, + 689 + ], + "type": "text", + "content": " can be transformed from the world coordinate system into the camera coordinate system of frame " + }, + { + "bbox": [ + 313, + 555, + 554, + 689 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 555, + 554, + 689 + ], + "type": "text", + "content": ". This transformation enables self-supervised training by enforcing two types of consistency: (1) 2D correspondence consistency, which aligns the projected 2D tracks (from " + }, + { + "bbox": [ + 313, + 555, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_j^1" + }, + { + "bbox": [ + 313, + 555, + 554, + 689 + ], + "type": "text", + "content": ") with the pseudo-ground truth tracking from CoTracker [23, 24], and (2) geometric consistency, which aligns the scale-invariant depth (from " + }, + { + "bbox": [ + 313, + 555, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_j^j" + }, + { + "bbox": [ + 313, + 555, + 554, + 689 + ], + "type": "text", + "content": ") with the pseudoground truth monocular depth from MoGe [59]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 715 + ], + "type": "text", + "content": "More specifically, given the estimated camera pose " + }, + { + "bbox": [ + 313, + 689, + 553, + 715 + ], + "type": "inline_equation", + "content": "(\\mathbf{R}^j,\\mathbf{T}^j)" + }, + { + "bbox": [ + 313, + 689, + 553, + 715 + ], + "type": "text", + "content": " from frame " + }, + { + "bbox": [ + 313, + 689, + 553, + 715 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 689, + 553, + 715 + ], + "type": "text", + "content": " and the tracking pointmap " + }, + { + "bbox": [ + 313, + 689, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_j^1" + }, + { + "bbox": [ + 313, + 689, + 553, + 715 + ], + "type": "text", + "content": ", we" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8507" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 289, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 289, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 289, + 85 + ], + "type": "text", + "content": "reproject these 3D points into the image plane of frame " + }, + { + "bbox": [ + 55, + 72, + 289, + 85 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 72, + 289, + 85 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 110, + 91, + 295, + 107 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 91, + 295, + 107 + ], + "spans": [ + { + "bbox": [ + 110, + 91, + 295, + 107 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} ^ {j, n} = \\pi \\left(\\mathbf {K} \\left(\\mathbf {R} ^ {j} \\mathbf {X} _ {j} ^ {1, n} + \\mathbf {T} ^ {j}\\right)\\right). \\tag {5}", + "image_path": "e054f10775faa83fced971670153788b5396075d9003f98b1b5ab7898284342b.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 114, + 295, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 114, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 55, + 114, + 295, + 149 + ], + "type": "text", + "content": "These reprojected points serve as the predicted 2D tracks and are compared with the pseudo-ground truth tracking points " + }, + { + "bbox": [ + 55, + 114, + 295, + 149 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{trk}}^{j,n}" + }, + { + "bbox": [ + 55, + 114, + 295, + 149 + ], + "type": "text", + "content": " from CoTracker3 [23]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 149, + 296, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 149, + 296, + 210 + ], + "spans": [ + { + "bbox": [ + 55, + 149, + 296, + 210 + ], + "type": "text", + "content": "To mitigate minor focal inaccuracies that may induce scaling shifts, the reprojection loss is computed in a scale-invariant manner. Let " + }, + { + "bbox": [ + 55, + 149, + 296, + 210 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_n = \\hat{\\mathbf{x}}^{j,n}" + }, + { + "bbox": [ + 55, + 149, + 296, + 210 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 149, + 296, + 210 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_n = \\mathbf{x}_{\\mathrm{trk}}^{j,n}" + }, + { + "bbox": [ + 55, + 149, + 296, + 210 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 149, + 296, + 210 + ], + "type": "inline_equation", + "content": "n = 1,\\dots ,N" + }, + { + "bbox": [ + 55, + 149, + 296, + 210 + ], + "type": "text", + "content": " , and denote the image center by " + }, + { + "bbox": [ + 55, + 149, + 296, + 210 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 149, + 296, + 210 + ], + "type": "text", + "content": " . The scale factor and adjusted predictions are computed together as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 217, + 296, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 217, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 70, + 217, + 296, + 251 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {p}} _ {n} = \\left(\\mathbf {p} _ {n} - c\\right) s + c, \\quad s = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\frac {\\left\\| \\mathbf {g} _ {n} - c \\right\\| _ {2}}{\\left\\| \\mathbf {p} _ {n} - c \\right\\| _ {2}}. \\tag {6}", + "image_path": "4c1cbdd45d9047548032afa8b79ae23af1ee43d9a8cb109a7e56f46399a804e9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 257, + 293, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 293, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 293, + 269 + ], + "type": "text", + "content": "Then, the scale-invariant 2D reprojection loss is defined as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 276, + 295, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 276, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 116, + 276, + 295, + 306 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {t r a j}} = \\frac {1}{N} \\sum_ {n \\in \\mathcal {I} _ {j}} \\| \\hat {\\mathbf {p}} _ {n} - \\mathbf {g} _ {n} \\| ^ {2}. \\tag {7}", + "image_path": "2f7955d1c778202620b0710acff49d767f45f67cccfeaa51fed8b7f062adf6e6.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 313, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 313, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 296, + 361 + ], + "type": "text", + "content": "Similarly, to enforce geometric consistency with the mono-depth predictions from MoGe [59], we use the reconstruction pointmap " + }, + { + "bbox": [ + 55, + 313, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_j^j" + }, + { + "bbox": [ + 55, + 313, + 296, + 361 + ], + "type": "text", + "content": " that transformed to frame " + }, + { + "bbox": [ + 55, + 313, + 296, + 361 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 313, + 296, + 361 + ], + "type": "text", + "content": "'s camera coordinate. The depth of each transformed 3D point is" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 120, + 369, + 295, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 369, + 295, + 389 + ], + "spans": [ + { + "bbox": [ + 120, + 369, + 295, + 389 + ], + "type": "interline_equation", + "content": "z _ {\\text {p r o j}} ^ {j, n} = \\left(\\mathbf {R} ^ {j} \\mathbf {X} _ {j} ^ {j, n} + \\mathbf {T} ^ {j}\\right) _ {z}, \\tag {8}", + "image_path": "4c9a8e6416d42e5a00d3cc58f0a2bcce15f41eacb93f349c4ebd926b83249958.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 396, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 456 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 396, + 296, + 456 + ], + "type": "inline_equation", + "content": "(\\cdot)_z" + }, + { + "bbox": [ + 55, + 396, + 296, + 456 + ], + "type": "text", + "content": " denotes the third (depth) component. Denote the corresponding mono-depth pseudo ground-truth by " + }, + { + "bbox": [ + 55, + 396, + 296, + 456 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{mono}}^{j,n}" + }, + { + "bbox": [ + 55, + 396, + 296, + 456 + ], + "type": "text", + "content": ". After solving for an optimal scaling factor " + }, + { + "bbox": [ + 55, + 396, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\alpha^*" + }, + { + "bbox": [ + 55, + 396, + 296, + 456 + ], + "type": "text", + "content": " to align the two depth maps, the scale-invariant mono-depth loss is defined as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 462, + 296, + 511 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 296, + 511 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 296, + 511 + ], + "type": "interline_equation", + "content": "L _ {\\text {d e p t h}} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\left(\\alpha^ {*} z _ {\\text {p r o j}} ^ {j, n} - z _ {\\text {m o n o}} ^ {j, n}\\right) ^ {2}, \\quad \\alpha^ {*} = \\frac {\\sum_ {i} z _ {\\text {p r o j}} ^ {j , n} z _ {\\text {m o n o}} ^ {j , n}}{\\sum_ {i} \\left(z _ {\\text {p r o j}} ^ {j , n}\\right) ^ {2}}. \\tag {9}", + "image_path": "ce177e8478c0e3274a617d0e2dc6b219a96b920ad9eb0078bf006a18200a815e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 516, + 296, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 296, + 614 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 296, + 614 + ], + "type": "text", + "content": "3D Self-Consistency. Beyond the 2D reprojection losses, we introduce a 3D self-consistency term that aligns the tracking pointmap " + }, + { + "bbox": [ + 55, + 516, + 296, + 614 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_j^1" + }, + { + "bbox": [ + 55, + 516, + 296, + 614 + ], + "type": "text", + "content": " with the reconstruction pointmap " + }, + { + "bbox": [ + 55, + 516, + 296, + 614 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_j^j" + }, + { + "bbox": [ + 55, + 516, + 296, + 614 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 55, + 516, + 296, + 614 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_1^\\prime" + }, + { + "bbox": [ + 55, + 516, + 296, + 614 + ], + "type": "text", + "content": " be the set of points in frame 1 that remain visible in frame " + }, + { + "bbox": [ + 55, + 516, + 296, + 614 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 516, + 296, + 614 + ], + "type": "text", + "content": ", and for each point " + }, + { + "bbox": [ + 55, + 516, + 296, + 614 + ], + "type": "inline_equation", + "content": "n \\in \\mathcal{I}_1^\\prime" + }, + { + "bbox": [ + 55, + 516, + 296, + 614 + ], + "type": "text", + "content": ", denote its corresponding point (provided by CoTracker) in frame " + }, + { + "bbox": [ + 55, + 516, + 296, + 614 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 516, + 296, + 614 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 516, + 296, + 614 + ], + "type": "inline_equation", + "content": "n^\\prime" + }, + { + "bbox": [ + 55, + 516, + 296, + 614 + ], + "type": "text", + "content": ". We then penalize the distance between their predicted 3D positions:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 110, + 621, + 296, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 621, + 296, + 649 + ], + "spans": [ + { + "bbox": [ + 110, + 621, + 296, + 649 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l i g n}} = \\sum_ {n \\in \\mathcal {I} _ {1} ^ {\\prime}} \\left\\| \\mathbf {X} _ {j} ^ {1, n} - \\mathbf {X} _ {j} ^ {j, n ^ {\\prime}} \\right\\| ^ {2}. \\tag {10}", + "image_path": "ba0b4cc20e4b805381aed7cbe097c2560af8f3176ccb773f264f3010f003aa27.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 657, + 295, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 295, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 295, + 681 + ], + "type": "text", + "content": "Minimizing " + }, + { + "bbox": [ + 55, + 657, + 295, + 681 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{align}}" + }, + { + "bbox": [ + 55, + 657, + 295, + 681 + ], + "type": "text", + "content": " ensures that both branches produce consistent geometry in the same timestamp." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 67, + 681, + 249, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 681, + 249, + 693 + ], + "spans": [ + { + "bbox": [ + 67, + 681, + 249, + 693 + ], + "type": "text", + "content": "The overall self-supervision loss is given by:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 99, + 701, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 701, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 99, + 701, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e p r o j}} = \\mathcal {L} _ {\\text {t r a j}} + \\lambda_ {1} \\mathcal {L} _ {\\text {d e p t h}} + \\lambda_ {2} \\mathcal {L} _ {\\text {a l i g n}}, \\tag {11}", + "image_path": "a5fb78825ca5168d2a2b8af4b602815fb4fb2b9e1fcfcb55c59daa876ab46305.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " being the weighting factors. Minimizing " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{proj}}" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " aligns the projected 3D structure with the 2D tracking and monocular depth cues and their 3D self-consistency, enabling unsupervised, target-specific refinement of the 3D geometry and point tracking." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 137, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 137, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 555, + 437 + ], + "type": "text", + "content": "Test-Time Adaptation. To address the gap between synthetic pretraining and real-world data, we incorporate reprojection-based losses to enable test-time adaptation in St4RTrack. Our framework supports two adaptation paradigms: (1) Instance-level adaptation. During testing, we update St4RTrack on new sequences using only the aforementioned reprojected losses while freezing the reconstruction branch. We freeze these weights because both the 2D trajectories and depth are computed under a purely monocular setting, which does not provide view-alignment supervision. This approach preserves the view-alignment capability captured during pretraining. Moreover, since the pretrained network already encodes strong task-relevant representations, this sequence-specific optimization converges rapidly compared to test-time optimization methods that start from scratch. (2) Domain-level adaptation. Unlike tabula-rasa approaches such as [27, 57], which require full re-optimization for each new sequence, St4RTrack is an end-to-end learning framework that enables test-time adaptation to align the model from its pretraining data distribution to the target video domain. After adapting to a sparse set of target-domain samples, St4RTrack can directly perform simultaneous reconstruction and tracking on new sequences from the adapted domain without additional optimization." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 447, + 394, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 447, + 394, + 460 + ], + "spans": [ + { + "bbox": [ + 314, + 447, + 394, + 460 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 467, + 554, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 554, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 554, + 563 + ], + "type": "text", + "content": "St4RTrack performs both dense 3D point tracking and dynamic reconstruction in a unified world coordinate system, all within a single inference. In the following section, we first evaluate our method on 3D tracking and dynamic reconstruction separately, and then present the joint results. We also introduce a new benchmark, WorldTrack, for 3D tracking in world coordinates, which is not directly covered by previous methods." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 571, + 436, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 571, + 436, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 571, + 436, + 583 + ], + "type": "text", + "content": "4.1. Experimental Details" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 589, + 554, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 589, + 554, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 589, + 554, + 696 + ], + "type": "text", + "content": "Datasets. For fully supervised training, we use three synthetic datasets: Point Odyssey (PO) [69], Dynamic Replica (DR) [22], and Kubric [15]. All three datasets contain scene and camera motion and provide mesh vertex positions as ground-truth 3D point trajectories. We randomly sample 24 frames with a stride of " + }, + { + "bbox": [ + 313, + 589, + 554, + 696 + ], + "type": "inline_equation", + "content": "1 \\sim 6" + }, + { + "bbox": [ + 313, + 589, + 554, + 696 + ], + "type": "text", + "content": " for each sample sequence. We also filter out less semantically meaningful sequences in PO, resulting in a total of 9.8k sequences for PO, 8.5k for DR, and 5.7k for Kubric dataset." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 314, + 701, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 701, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 314, + 701, + 554, + 714 + ], + "type": "text", + "content": "Training and Inference. During training, we sample 600" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8508" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 76, + 555, + 198 + ], + "blocks": [ + { + "bbox": [ + 55, + 50, + 555, + 73 + ], + "lines": [ + { + "bbox": [ + 55, + 50, + 555, + 73 + ], + "spans": [ + { + "bbox": [ + 55, + 50, + 555, + 73 + ], + "type": "text", + "content": "Table 1. World Coordinate 3D Point Tracking. We report the performance of average points under distance " + }, + { + "bbox": [ + 55, + 50, + 555, + 73 + ], + "type": "inline_equation", + "content": "(\\mathrm{APD}_{3\\mathrm{D}})" + }, + { + "bbox": [ + 55, + 50, + 555, + 73 + ], + "type": "text", + "content": " after global median alignment. We evaluate the accuracy of both all points and dynamic points. The best results are bold. See ?? for more results." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 76, + 555, + 198 + ], + "lines": [ + { + "bbox": [ + 57, + 76, + 555, + 198 + ], + "spans": [ + { + "bbox": [ + 57, + 76, + 555, + 198 + ], + "type": "table", + "html": "
CategoryMethodsAll PointsDynamic Points
PODRADTPStudioPODRADTPStudio
CombinationalSpaTracker+RANSAC-Procrustes44.0355.0150.8752.0553.7758.5866.4952.05
SpaTracker+MonST3R47.6555.4951.9550.1658.6159.2169.9450.16
Feed-forwardMonST3R33.4758.0674.3551.3239.3651.8667.9251.32
SpaTracker38.5454.8545.6562.5951.2058.6567.6562.59
St4RTrack (Ours)67.9573.7476.0169.6768.7268.1375.3469.67
", + "image_path": "de09abb813879a15dd6b5734fd3003f82d934644f16cd34c0c7f2d13dd0867bd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 57, + 243, + 296, + 376 + ], + "blocks": [ + { + "bbox": [ + 55, + 205, + 295, + 240 + ], + "lines": [ + { + "bbox": [ + 55, + 205, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 295, + 240 + ], + "type": "text", + "content": "Table 2. World Coordinate 3D Reconstruction. We report performance on both Point Odyssey (PO) and TUM-Dynamics after global median scaling. The best results are in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 243, + 296, + 376 + ], + "lines": [ + { + "bbox": [ + 57, + 243, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 57, + 243, + 296, + 376 + ], + "type": "table", + "html": "
CategoryMethodsPoint OdysseyTUM-Dynamics
EPE↓APD↑EPE↓APD↑
w/ Global Align.DUSt3R+GA0.608843.900.314770.49
MASt3R+GA0.403060.440.518668.38
MonST3R+GA0.262972.310.342963.87
Feed-forwardDUSt3R0.638645.790.289172.26
MASt3R0.464456.900.551066.22
MonST3R0.304468.250.364661.38
St4RTrack (Ours)0.240678.730.185483.42
", + "image_path": "6b75972565ee627880afcc4e04852f0e4737671303d93a324e52606a4dbf6902.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 389, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 389, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 54, + 389, + 295, + 521 + ], + "type": "text", + "content": "sequences from each dataset per epoch. We use the AdamW optimizer with a learning rate of " + }, + { + "bbox": [ + 54, + 389, + 295, + 521 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-5}" + }, + { + "bbox": [ + 54, + 389, + 295, + 521 + ], + "type": "text", + "content": " and a mini-batch size of 1 per GPU. The model is trained for 50 epochs on 4 A100 GPUs, which takes about one day. For test-time adaptation, we run 500 optimization steps on a single sequence, taking approximately 5 minutes on 4 A100 GPUs. At inference time, the model runs at 30 FPS on an RTX 4090. Although the model is trained on sequences of 24 frames, our pair-wise approach allows it to operate on arbitrarily long videos during inference. Refer to ?? for more details regarding test-time adaptation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 528, + 241, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 241, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 241, + 541 + ], + "type": "text", + "content": "4.2. 3D Tracking in World Coordinates" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 545, + 295, + 713 + ], + "type": "text", + "content": "Datasets. 3D tracking in world coordinates is a critical aspect that has been largely overlooked by previous benchmarks [26], which are limited to camera coordinate systems. To address this limitation, we propose a new benchmark for 3D tracking in world coordinates. Our benchmark leverages two real-world datasets—Aerial Digital Twin (ADT) [40] and Panoptic Studio [21]—by converting the TAPVid-3D [26] sequences to world coordinates using paired extrinsic parameters. However, it is noteworthy that the limitations of these datasets: the ADT sequences exhibit minimal scene motion, while the Panoptic Studio lacks camera motion. To overcome these shortcomings, we include two additional synthetic test sets from Point Odyssey and Dynamic Replica, which have both scene and" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 208, + 555, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 208, + 555, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 208, + 555, + 255 + ], + "type": "text", + "content": "camera motion. In total, our benchmark comprises four datasets, each containing 50 sequences of 64 frames. Refer to ?? for more details and visual examples of the benchmark." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 260, + 556, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 556, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 556, + 393 + ], + "type": "text", + "content": "Evaluation Metrics. We follow the TAPVid-3D protocol and use the Average percent of Points within Delta (APD) metric for evaluation. Specifically, we first align the predicted 3D point trajectories with the ground truth by normalizing them with their global median. We then compute the prediction error and measure the percentage of points whose error falls below a given threshold " + }, + { + "bbox": [ + 313, + 260, + 556, + 393 + ], + "type": "inline_equation", + "content": "\\delta_{3D}" + }, + { + "bbox": [ + 313, + 260, + 556, + 393 + ], + "type": "text", + "content": " (with " + }, + { + "bbox": [ + 313, + 260, + 556, + 393 + ], + "type": "inline_equation", + "content": "\\delta_{3D} \\in \\{0.1\\mathrm{m}, 0.3\\mathrm{m}, 0.5\\mathrm{m}, 1.0\\mathrm{m}\\}" + }, + { + "bbox": [ + 313, + 260, + 556, + 393 + ], + "type": "text", + "content": ") over the first 64 frames. Let " + }, + { + "bbox": [ + 313, + 260, + 556, + 393 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}}_t^i" + }, + { + "bbox": [ + 313, + 260, + 556, + 393 + ], + "type": "text", + "content": " denote the " + }, + { + "bbox": [ + 313, + 260, + 556, + 393 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 260, + 556, + 393 + ], + "type": "text", + "content": "-th predicted point at time " + }, + { + "bbox": [ + 313, + 260, + 556, + 393 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 260, + 556, + 393 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 260, + 556, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_t^i" + }, + { + "bbox": [ + 313, + 260, + 556, + 393 + ], + "type": "text", + "content": " denote its corresponding ground-truth location. The resulting " + }, + { + "bbox": [ + 313, + 260, + 556, + 393 + ], + "type": "inline_equation", + "content": "\\mathrm{APD}_{3D}" + }, + { + "bbox": [ + 313, + 260, + 556, + 393 + ], + "type": "text", + "content": " is then computed as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 355, + 402, + 555, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 402, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 355, + 402, + 555, + 430 + ], + "type": "interline_equation", + "content": "\\mathrm {A P D} _ {\\mathrm {3 D}} \\equiv \\sum_ {i, t} \\mathbb {1} \\left(\\left\\| \\hat {\\mathbf {P}} _ {t} ^ {i} - \\mathbf {P} _ {t} ^ {i} \\right\\| < \\delta_ {3 D}\\right), \\tag {12}", + "image_path": "b7c94efd188ef2eab65f3c2260c04def5db011cf56804412246862f3ceccc4e3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 439, + 554, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 439, + 554, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 554, + 463 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 439, + 554, + 463 + ], + "type": "inline_equation", + "content": "\\mathbb{1}(\\cdot)" + }, + { + "bbox": [ + 313, + 439, + 554, + 463 + ], + "type": "text", + "content": " is the indicator function and " + }, + { + "bbox": [ + 313, + 439, + 554, + 463 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|" + }, + { + "bbox": [ + 313, + 439, + 554, + 463 + ], + "type": "text", + "content": " denotes the Euclidean norm." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 468, + 556, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 556, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 556, + 613 + ], + "type": "text", + "content": "Baselines. Since no existing work explicitly performs 3D tracking in world coordinates, for our feedforward baselines, we compare against the camera coordinate 3D tracking method SpatialTracker [64] and a dynamic 3D reconstruction method MonST3R [67] (as a non-tracking baseline). In addition, we implement two combinational baselines for world coordinate 3D tracking. The first baseline applies Procrustes alignment [54] and RANSAC [11] to the camera coordinate 3D tracks predicted by SpatialTracker to offset the camera motion. The second baseline leverages the camera poses predicted by the dynamic SLAM method MonST3R to compensate for camera motion." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "type": "text", + "content": "Results. As shown in Tab. 1, we achieve state-of-the-art performance, with test-time adaptation proving particularly beneficial for dynamic points. Notably, on the Panoptic Studio dataset, which is captured with a fixed camera and can be considered a fair benchmark for camera coordinate tracking methods, our approach still outperforms SpatialTracker [64]. It is worth noting that although our model is trained on sequences of 24 frames, it generalizes well to" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8509" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 59, + 125, + 91 + ], + "blocks": [ + { + "bbox": [ + 69, + 59, + 125, + 91 + ], + "lines": [ + { + "bbox": [ + 69, + 59, + 125, + 91 + ], + "spans": [ + { + "bbox": [ + 69, + 59, + 125, + 91 + ], + "type": "image", + "image_path": "8326b61c09a351b48947af958d39883c51ff9ad97e5979afb1933fd2311f71ab.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 69, + 92, + 125, + 125 + ], + "blocks": [ + { + "bbox": [ + 69, + 92, + 125, + 125 + ], + "lines": [ + { + "bbox": [ + 69, + 92, + 125, + 125 + ], + "spans": [ + { + "bbox": [ + 69, + 92, + 125, + 125 + ], + "type": "image", + "image_path": "eaa581cf9c9b9f9077f85026ace1801dae1ca7a68f8b5552801c1cd46f5f914c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 137, + 50, + 265, + 135 + ], + "blocks": [ + { + "bbox": [ + 137, + 50, + 265, + 135 + ], + "lines": [ + { + "bbox": [ + 137, + 50, + 265, + 135 + ], + "spans": [ + { + "bbox": [ + 137, + 50, + 265, + 135 + ], + "type": "image", + "image_path": "8c8d8982a2cb365c4df587b6df14e7e89c3425d4a2ac59d4c3654fce0ec6fe35.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 271, + 50, + 408, + 140 + ], + "blocks": [ + { + "bbox": [ + 271, + 50, + 408, + 140 + ], + "lines": [ + { + "bbox": [ + 271, + 50, + 408, + 140 + ], + "spans": [ + { + "bbox": [ + 271, + 50, + 408, + 140 + ], + "type": "image", + "image_path": "4ee0e24f3babbf21a88a367b57e343135347e3c2bcd5e96b72600f1b1efdc5d4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 410, + 49, + 540, + 131 + ], + "blocks": [ + { + "bbox": [ + 410, + 49, + 540, + 131 + ], + "lines": [ + { + "bbox": [ + 410, + 49, + 540, + 131 + ], + "spans": [ + { + "bbox": [ + 410, + 49, + 540, + 131 + ], + "type": "image", + "image_path": "e78831837f016db6990ccc530b742ff54534d0ac2c0011f1a9b2ae972cd2b883.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 69, + 144, + 125, + 178 + ], + "blocks": [ + { + "bbox": [ + 69, + 144, + 125, + 178 + ], + "lines": [ + { + "bbox": [ + 69, + 144, + 125, + 178 + ], + "spans": [ + { + "bbox": [ + 69, + 144, + 125, + 178 + ], + "type": "image", + "image_path": "ff00bfcb118553f05b5cb325eb8bcb24a0aff8700cdb55007cd24367e7adf729.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 69, + 178, + 124, + 211 + ], + "blocks": [ + { + "bbox": [ + 69, + 178, + 124, + 211 + ], + "lines": [ + { + "bbox": [ + 69, + 178, + 124, + 211 + ], + "spans": [ + { + "bbox": [ + 69, + 178, + 124, + 211 + ], + "type": "image", + "image_path": "6e9c87060f4f471cb1074682ba5e4abef0d8897e338fc7a9f0a49ab24e35dcec.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 132, + 148, + 265, + 220 + ], + "blocks": [ + { + "bbox": [ + 132, + 148, + 265, + 220 + ], + "lines": [ + { + "bbox": [ + 132, + 148, + 265, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 148, + 265, + 220 + ], + "type": "image", + "image_path": "341e505abe930d474ac9ab960e4251b8b8880a59e27ed2d1db7e0fbcc349de07.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 271, + 150, + 405, + 213 + ], + "blocks": [ + { + "bbox": [ + 271, + 150, + 405, + 213 + ], + "lines": [ + { + "bbox": [ + 271, + 150, + 405, + 213 + ], + "spans": [ + { + "bbox": [ + 271, + 150, + 405, + 213 + ], + "type": "image", + "image_path": "d0e2c17e8160286cd13eeedac026722f00639e0c7adc81536454e65a6bf51d62.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 419, + 149, + 537, + 213 + ], + "blocks": [ + { + "bbox": [ + 419, + 149, + 537, + 213 + ], + "lines": [ + { + "bbox": [ + 419, + 149, + 537, + 213 + ], + "spans": [ + { + "bbox": [ + 419, + 149, + 537, + 213 + ], + "type": "image", + "image_path": "a4c80a539f678cad3b17e98f21d41035787064689b9d55147a082748c55c7913.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 69, + 225, + 125, + 258 + ], + "blocks": [ + { + "bbox": [ + 69, + 225, + 125, + 258 + ], + "lines": [ + { + "bbox": [ + 69, + 225, + 125, + 258 + ], + "spans": [ + { + "bbox": [ + 69, + 225, + 125, + 258 + ], + "type": "image", + "image_path": "9558911e2e4ae72a9b195696679f6efd7f7793f6a72e8897a953cf24e820a480.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 320, + 555, + 364 + ], + "lines": [ + { + "bbox": [ + 55, + 320, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 55, + 320, + 555, + 364 + ], + "type": "text", + "content": "Figure 4. Qualitative Results. From left to right, we show our results in feed-forward inference: 1) the input video, 2) two pointmaps at frame " + }, + { + "bbox": [ + 55, + 320, + 555, + 364 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 320, + 555, + 364 + ], + "type": "text", + "content": " overlayed together, 3) the accumulated reconstruction branch result, and 4) the accumulated tracking branch result. The accumulated reconstruction demonstrates a stable reconstruction of the dynamic scene geometry, while the accumulated tracking illustrates long-term, dense tracking of scene motion. More qualitative results in???" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 69, + 258, + 124, + 293 + ], + "blocks": [ + { + "bbox": [ + 69, + 258, + 124, + 293 + ], + "lines": [ + { + "bbox": [ + 69, + 258, + 124, + 293 + ], + "spans": [ + { + "bbox": [ + 69, + 258, + 124, + 293 + ], + "type": "image", + "image_path": "edd1f728eb60a1d24bc90afd7ce3fb97e7cd473f6b7e8d7a3ac6f0fd4bd28031.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 130, + 228, + 263, + 289 + ], + "blocks": [ + { + "bbox": [ + 130, + 228, + 263, + 289 + ], + "lines": [ + { + "bbox": [ + 130, + 228, + 263, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 263, + 289 + ], + "type": "image", + "image_path": "26807ecb796ea967a5522700d8ca0a5497d7f124c2f92f1f547f3020f3a144f1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 274, + 230, + 403, + 291 + ], + "blocks": [ + { + "bbox": [ + 274, + 230, + 403, + 291 + ], + "lines": [ + { + "bbox": [ + 274, + 230, + 403, + 291 + ], + "spans": [ + { + "bbox": [ + 274, + 230, + 403, + 291 + ], + "type": "image", + "image_path": "4da768f12644a4c491ff80cb1b51537ec4500180b907397941bfc048994c6644.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 410, + 226, + 540, + 298 + ], + "blocks": [ + { + "bbox": [ + 410, + 226, + 540, + 298 + ], + "lines": [ + { + "bbox": [ + 410, + 226, + 540, + 298 + ], + "spans": [ + { + "bbox": [ + 410, + 226, + 540, + 298 + ], + "type": "image", + "image_path": "79936bed42e3066c8b42123becebbc5451dc9b376eddc23da7f97c28f2d50a0b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 75, + 304, + 125, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 304, + 125, + 315 + ], + "spans": [ + { + "bbox": [ + 75, + 304, + 125, + 315 + ], + "type": "text", + "content": "Input Video" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 172, + 304, + 223, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 304, + 223, + 316 + ], + "spans": [ + { + "bbox": [ + 172, + 304, + 223, + 316 + ], + "type": "text", + "content": "Pair Output" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 274, + 304, + 396, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 304, + 396, + 315 + ], + "spans": [ + { + "bbox": [ + 274, + 304, + 396, + 315 + ], + "type": "text", + "content": "Accumulated Reconstruction" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 425, + 304, + 522, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 304, + 522, + 316 + ], + "spans": [ + { + "bbox": [ + 425, + 304, + 522, + 316 + ], + "type": "text", + "content": "Accumulated Tracking" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 376, + 295, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 399 + ], + "type": "text", + "content": "longer sequences, including 64-frame videos. Refer to ?? for more results." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 406, + 211, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 406, + 211, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 211, + 418 + ], + "type": "text", + "content": "4.3. Dynamic 3D Reconstruction" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 423, + 296, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 296, + 483 + ], + "type": "text", + "content": "Datasets. We evaluate on both synthetic and real-world data. For the synthetic data, we use Point Odyssey [69]. For real-world evaluation, we employ TUM-Dynamics [49], a subset of a SLAM dataset featuring moving people, dense depth maps, and accurate camera poses." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 488, + 296, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 488, + 296, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 488, + 296, + 583 + ], + "type": "text", + "content": "Evaluation Metrics. Unlike prior works [58, 67] that separately evaluate video depth and camera pose estimation, we directly compare the reconstructed 3D point clouds to the ground truth using the Average percent of Points within Distance (APD) and End-Point Error (EPE) metrics. We filter out ambiguous floating points in the ground truth data and align the point clouds for each sequence using the median scale before evaluation." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 589, + 296, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 589, + 296, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 589, + 296, + 649 + ], + "type": "text", + "content": "Baselines. We compare our method against MonST3R, MASt3R and DUSt3R, both with global alignment (w/ GA) and in feedforward mode. For the feedforward baselines, we construct image pairs of a video in the form of St4RTrack, that align all frames to a common anchor frame." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Results. As in Tab. 2, we also achieve state-of-the-art performance on the task of 3D reconstruction in the world coordinate. Although MonST3R is designed for 3D reconstruction for dynamic scenes, it still underperforms St4RTrack even with global alignment. This further highlights the ben" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 376, + 555, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 411 + ], + "type": "text", + "content": "efit of jointly tracking and reconstruction. Since we freeze the reconstruction head to preserve the 3D prior, 3D reconstruction results are similar with test-time adaptation." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 418, + 555, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 555, + 431 + ], + "type": "text", + "content": "4.4. Joint Tracking and Reconstruction in the World" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 436, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 555, + 567 + ], + "type": "text", + "content": "Our method simultaneously predicts 3D point trajectories and point clouds in a single feed-forward pass, evaluated separately in prior sections. Here, we present qualitative results visualizing both raw 3D point trajectories and point clouds within the same world coordinates, as shown in Fig. 4. The \"pair output\" demonstrates that outputs from the tracking and reconstruction branches align well at the same time step. Additionally, the accumulated reconstruction indicates consistency in static regions, while accumulated tracking shows that our method estimates accurate and smooth 3D tracks over time." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 574, + 388, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 388, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 388, + 586 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "We introduce St4RTrack, a feed-forward framework that simultaneously achieves 3D point tracking and dynamic reconstruction in the world coordinate from monocular videos using a unified representation. Alongside, we present a novel benchmark, WorldTrack, for systematically evaluating dynamic 3D scene geometry and motion estimation in a global reference frame. Our method achieves state-of-the-art performance on both synthetic and real-world datasets, while also extending beyond fully supervised paradigms by enabling test-time adaptation." + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8510" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "We thank Aleksander Holynski, Yifei Zhang, Chung Min Kim, Brent Yi, and Zehao Yu for helpful discussions. We especially thank Aleksander Holynski for his guidance and feedback. This project is supported in part by NSF CNS-2235013, Bakar Fellows, and BAIR Commons Industry Alliance Sponsors." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 115, + 188 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 115, + 188 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 115, + 188 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 195, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 195, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 195, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 61, + 195, + 296, + 239 + ], + "type": "text", + "content": "[1] Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M Seitz, and Richard Szeliski. Building rome in a day. Communications of the ACM, 54 (10):105-112, 2011. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 240, + 296, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 240, + 296, + 283 + ], + "spans": [ + { + "bbox": [ + 61, + 240, + 296, + 283 + ], + "type": "text", + "content": "[2] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 285, + 295, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 285, + 295, + 328 + ], + "spans": [ + { + "bbox": [ + 61, + 285, + 295, + 328 + ], + "type": "text", + "content": "[3] Michael J Black and Padmanabhan Anandan. A framework for the robust estimation of optical flow. In 1993 (4th) International Conference on Computer Vision, pages 231-236. IEEE, 1993. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 331, + 296, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 331, + 296, + 385 + ], + "spans": [ + { + "bbox": [ + 61, + 331, + 296, + 385 + ], + "type": "text", + "content": "[4] Christoph Bregler, Aaron Hertzmann, and Henning Biermann. Recovering non-rigid 3d shape from image streams. In Proceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No. PR00662), pages 690-696. IEEE, 2000. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 387, + 296, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 387, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 387, + 296, + 441 + ], + "type": "text", + "content": "[5] Hansheng Chen, Wei Tian, Pichao Wang, Fan Wang, Lu Xiong, and Hao Li. EPro-PnP: Generalized end-to-end probabilistic perspective-n-points for monocular object pose estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 444, + 296, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 444, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 62, + 444, + 296, + 486 + ], + "type": "text", + "content": "[6] Yuchao Dai, Hongdong Li, and Mingyi He. A simple prior-free method for non-rigid structure-from-motion factorization. International Journal of Computer Vision, 107:101-122, 2014. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 488, + 296, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 488, + 296, + 532 + ], + "spans": [ + { + "bbox": [ + 62, + 488, + 296, + 532 + ], + "type": "text", + "content": "[7] Andrew J Davison, Ian D Reid, Nicholas D Molton, and Olivier Stasse. Monoslam: Real-time single camera slam. IEEE transactions on pattern analysis and machine intelligence, 29(6):1052-1067, 2007. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 534, + 296, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 534, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 62, + 534, + 296, + 588 + ], + "type": "text", + "content": "[8] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. TAP-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 35:13610-13626, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 590, + 296, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 590, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 62, + 590, + 296, + 654 + ], + "type": "text", + "content": "[9] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. TAPIR: Tracking any point with per-frame initialization and temporal refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10061-10072, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 657, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 657, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 657, + 295, + 689 + ], + "type": "text", + "content": "[10] Hugh Durrant-Whyte and Tim Bailey. Simultaneous localization and mapping: part i. IEEE robotics & automation magazine, 13(2):99–110, 2006. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "text", + "content": "[11] Martin A Fischler and Robert C. Bolles. Random sample consensus: a paradigm for model fitting with applications to" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 5, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[12] Katerina Fragkiadaki, Marta Salas, Pablo Arbelaez, and Jitendra Malik. Grouping-based low-rank trajectory completion and 3d reconstruction. advances in neural information processing systems, 27, 2014. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "type": "text", + "content": "[13] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 198, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 240 + ], + "type": "text", + "content": "[14] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. Advances in Neural Information Processing Systems, 35:33768-33780, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 243, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 308 + ], + "type": "text", + "content": "[15] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3749-3761, 2022. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 309, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 354 + ], + "type": "text", + "content": "[16] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In European Conference on Computer Vision, pages 59-75. Springer, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 354, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 386 + ], + "type": "text", + "content": "[17] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 388, + 553, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 410 + ], + "type": "text", + "content": "[18] Berthold KP Horn and Brian G Schunck. Determining optical flow. Artificial intelligence, 17(1-3):185–203, 1981. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "type": "text", + "content": "[19] Junhwa Hur and Stefan Roth. Self-supervised monocular scene flow estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7396-7405, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 456, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 499 + ], + "type": "text", + "content": "[20] Linyi Jin, Richard Tucker, Zhengqi Li, David Fouhey, Noah Snavely, and Aleksander Holynski. Stereo4d: Learning how things move in 3d from internet stereo videos. arXiv preprint arXiv:2412.09621, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 502, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 502, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 502, + 553, + 555 + ], + "type": "text", + "content": "[21] Hanbyul Joo, Tomas Simon, Xulong Li, Hao Liu, Lei Tan, Lin Gui, Sean Banerjee, Timothy Godisart, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social interaction capture, 2016. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 557, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 623 + ], + "type": "text", + "content": "[22] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Dynamic stereo: Consistent dynamic depth from stereo videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13229-13239, 2023. 1, 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 624, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 668 + ], + "type": "text", + "content": "[23] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker3: Simpler and better point tracking by pseudolabelling real videos. 2024. 2, 5, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[24] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. In Proc. ECCV, 2024. 1, 2, 5" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8511" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 117 + ], + "type": "text", + "content": "[25] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1621, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "type": "text", + "content": "[26] Skanda Koppula, Ignacio Rocco, Yi Yang, Joe Heyward, João Carreira, Andrew Zisserman, Gabriel Brostow, and Carl Doersch. Tapvid-3d: A benchmark for tracking any point in 3d, 2024. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 207 + ], + "type": "text", + "content": "[27] Jiahui Lei, Yijia Weng, Adam W. Harley, Leonidas Guibas, and Kostas Daniilidis. MoSca: Dynamic gaussian fusion from casual videos via 4d motion scaffolds. arXiv preprint arXiv:2405.17421, 2024. 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 295, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 295, + 240 + ], + "type": "text", + "content": "[28] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. EPnP: An accurate O(n) solution to the PnP problem. IJCV, 81:155-166, 2009. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 295, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 295, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 295, + 275 + ], + "type": "text", + "content": "[29] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. In European Conference on Computer Vision, pages 71-91. Springer, 2024. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 295, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 295, + 342 + ], + "type": "text", + "content": "[30] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5521-5531, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 295, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 295, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 295, + 398 + ], + "type": "text", + "content": "[31] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 295, + 453 + ], + "type": "text", + "content": "[32] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4273-4284, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 455, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 295, + 521 + ], + "type": "text", + "content": "[33] Zhengqi Li, Richard Tucker, Forrester Cole, Qianqian Wang, Linyi Jin, Vickie Ye, Angjoo Kanazawa, Aleksander Holynski, and Noah Snavelly. MegaSaM: Accurate, fast, and robust structure and motion from casual dynamic videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 522, + 295, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 522, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 522, + 295, + 566 + ], + "type": "text", + "content": "[34] Qingming Liu, Yuan Liu, Jiepeng Wang, Xianqiang Lyv, Peng Wang, Wenping Wang, and Junhui Hou. Modgs: Dynamic gaussian splattering from casually-captured monocular videos, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 567, + 295, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 295, + 611 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 295, + 611 + ], + "type": "text", + "content": "[35] Bruce D Lucas and Takeo Kanade. An iterative image registration technique with an application to stereo vision. In IJCAI'81: 7th international joint conference on Artificial intelligence, pages 674-679, 1981. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 612, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 612, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 612, + 295, + 645 + ], + "type": "text", + "content": "[36] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 646, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 646, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 646, + 295, + 689 + ], + "type": "text", + "content": "[37] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: A versatile and accurate monocular slam system. IEEE transactions on robotics, 31(5):1147-1163, 2015. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 690, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 690, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 690, + 295, + 714 + ], + "type": "text", + "content": "[38] Tuan Duc Ngo, Peiye Zhuang, Chuang Gan, Evangelos Kalogerakis, Sergey Tulyakov, Hsin-Ying Lee, and" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "type": "text", + "content": "Chaoyang Wang. DELTA: Dense efficient long-range 3d tracking for any video. In International Conference on Learning Representations (ICLR), 2025. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 106, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 553, + 161 + ], + "type": "text", + "content": "[39] David Novotny, Nikhila Ravi, Benjamin Graham, Natalia Neverova, and Andrea Vedaldi. C3dpo: Canonical 3d pose networks for non-rigid structure from motion. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7688-7697, 2019. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 162, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 215 + ], + "type": "text", + "content": "[40] Xiaqing Pan, Nicholas Charron, Yongqian Yang, Scott Peters, Thomas Whelan, Chen Kong, Omkar Parkhi, Richard Newcombe, and Carl Yuheng Ren. Aria digital twin: A new benchmark dataset for egocentric 3d machine perception, 2023. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "type": "text", + "content": "[41] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 272, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 327 + ], + "type": "text", + "content": "[42] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Undepth: Universal monocular metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10106-10116, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "type": "text", + "content": "[43] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10318–10327, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 382, + 555, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 382, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 382, + 555, + 426 + ], + "type": "text", + "content": "[44] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 426, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 448 + ], + "type": "text", + "content": "[45] Michael Rubinstein, Ce Liu, and William T Freeman. Towards longer long-range motion trajectories. 2012. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "type": "text", + "content": "[46] P. Sand and S. Teller. Particle video: Long-range motion estimation using point trajectories. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), pages 2195-2202, 2006. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 493, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 536 + ], + "type": "text", + "content": "[47] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 537, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 601 + ], + "type": "text", + "content": "[48] Steven M Seitz, Brian Curless, James Diebel, Daniel Scharstein, and Richard Szeliski. A comparison and evaluation of multi-view stereo reconstruction algorithms. In 2006 IEEE computer society conference on computer vision and pattern recognition (CVPR'06), pages 519-528. IEEE, 2006. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 602, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 635 + ], + "type": "text", + "content": "[49] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of RGB-D SLAM systems. pages 573-580, 2012. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 636, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 689 + ], + "type": "text", + "content": "[50] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8934-8943, 2018. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 690, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 690, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 690, + 553, + 714 + ], + "type": "text", + "content": "[51] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow, 2020. 1, 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8512" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "text", + "content": "[52] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 34:16558-16569, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "text", + "content": "[53] Zachary Teed and Jia Deng. Raft-3d: Scene flow using rigid-motion embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8375–8384, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "text", + "content": "[54] Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. IEEE Transactions on Pattern Analysis & Machine Intelligence, 13(04): 376-380, 1991. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "text", + "content": "[55] Sundar Vedula, Simon Baker, Peter Rander, Robert Collins, and Takeo Kanade. Three-dimensional scene flow. In Proceedings of the Seventh IEEE International Conference on Computer Vision, pages 722-729. IEEE, 1999. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "type": "text", + "content": "[56] Jianyuan Wang, Nikita Karaev, Christian Rupprecht, and David Novotny. Vggsfm: Visual geometry grounded deep structure from motion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21686-21697, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "type": "text", + "content": "[57] Qianqian Wang, Vickie Ye, Hang Gao, Weijia Zeng, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 409 + ], + "type": "text", + "content": "[58] Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A. Efros, and Angjoo Kanazawa. Continuous 3d perception model with persistent state. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 1, 2, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 454 + ], + "type": "text", + "content": "[59] Ruicheng Wang, Sicheng Xu, Cassie Dai, Jianfeng Xiang, Yu Deng, Xin Tong, and Jiaolong Yang. Moge: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision, 2024. 2, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "type": "text", + "content": "[60] Sen Wang, Ronald Clark, Hongkai Wen, and Niki Trigoni. Deepvo: Towards end-to-end visual odometry with deep recurrent convolutional neural networks. In 2017 IEEE international conference on robotics and automation (ICRA), pages 2043-2050. IEEE, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 566 + ], + "type": "text", + "content": "[61] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jérôme Revaud. Dust3r: Geometric 3d vision made easy. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20697-20709, 2023. 2, 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 601 + ], + "type": "text", + "content": "[62] Endre Weiszfeld. Sur le point pour lequel la somme des distances de n points données est minimum. Tohoku Mathematical Journal, First Series, 43:355-386, 1937. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 602, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 657 + ], + "type": "text", + "content": "[63] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20310-20320, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[64] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. Spatialtracker: Tracking any 2d pixels in 3d space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 418 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 127 + ], + "type": "text", + "content": "[65] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 554, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 183 + ], + "type": "text", + "content": "[66] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20331-20341, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 185, + 554, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 554, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 554, + 249 + ], + "type": "text", + "content": "[67] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. MonST3R: A simple approach for estimating geometry in the presence of motion. In International Conference on Learning Representations (ICLR), 2025. 1, 2, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 251, + 554, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 554, + 305 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 554, + 305 + ], + "type": "text", + "content": "[68] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T. Freeman. Structure and motion from casual videos. In Proceedings of the European Conference on Computer Vision (ECCV), pages 20-37, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 307, + 554, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 307, + 554, + 361 + ], + "spans": [ + { + "bbox": [ + 316, + 307, + 554, + 361 + ], + "type": "text", + "content": "[69] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19855-19865, 2023. 5, 6, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 363, + 554, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 363, + 554, + 418 + ], + "spans": [ + { + "bbox": [ + 316, + 363, + 554, + 418 + ], + "type": "text", + "content": "[70] Xiaowei Zhou, Menglong Zhu, Spyridon Leonardos, Konstantinos G Derpanis, and Kostas Daniilidis. Sparseness meets deepness: 3d human pose estimation from monocular video. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4966-4975, 2016. 3" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8513" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Stable Diffusion Models are Secretly Good at Visual In-Context Learning/94d38526-8561-492c-b6db-5bcfedbe9083_content_list.json b/2025/Stable Diffusion Models are Secretly Good at Visual In-Context Learning/94d38526-8561-492c-b6db-5bcfedbe9083_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4536aaf4dcb0160f231a05a26c90d9d63655c8f3 --- /dev/null +++ b/2025/Stable Diffusion Models are Secretly Good at Visual In-Context Learning/94d38526-8561-492c-b6db-5bcfedbe9083_content_list.json @@ -0,0 +1,1467 @@ +[ + { + "type": "text", + "text": "Stable Diffusion Models are Secretly Good at Visual In-Context Learning", + "text_level": 1, + "bbox": [ + 129, + 130, + 867, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Trevine Oorloff $^{1,2*}$ Ali Shafahi $^{1}$ Vishwanath Sindagi $^{1}$ Wele Gedara Chaminda Bandara $^{1}$ Amin Ghiasi $^{1}$ Charan Prakash $^{1}$ Reza Ardekani $^{1}$ Apple $^{2}$ University of Maryland - College Park", + "bbox": [ + 169, + 178, + 825, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 279, + 326, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large language models (LLM) in natural language processing (NLP) have demonstrated great potential for in-context learning (ICL) — the ability to leverage a few sets of example prompts to adapt to various tasks without having to explicitly update the model weights. ICL has recently been explored for computer vision tasks with promising early outcomes. These approaches involve specialized training and/or additional data that complicate the process and limit its generalizability. In this work, we show that off-the-shelf Stable Diffusion models can be repurposed for visual in-context learning (V-ICL). Specifically, we formulate an in-place attention re-computation within the self-attention layers of the Stable Diffusion architecture that explicitly incorporates context between the query and example prompts. Without any additional fine-tuning, we show that this repurposed Stable Diffusion model is able to adapt to six different tasks: foreground segmentation, single object detection, semantic segmentation, keypoint detection, edge detection, and colorization. For example, the proposed approach improves the mean intersection over union (mIoU) for the foreground segmentation task on Pascal-5i dataset by $8.9\\%$ and $3.2\\%$ over recent methods such as Visual Prompting and IMProv, respectively. Additionally, we show that the proposed method is able to effectively leverage multiple prompts through ensembling to infer the task better and further improve the performance.", + "bbox": [ + 88, + 311, + 483, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 732, + 220, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In-context learning (ICL) refers to the paradigm where a foundation model leverages exemplar source-target pair(s), known as prompts, to infer the task and perform the inferred task on an input, known as query. ICL is an emergent property of large language models (LLMs) and has been widely explored in the natural language processing (NLP) domain [4, 8, 25, 32]. ICL facilitates a model to adapt to novel or out-of-domain tasks without the need for fine-tuning which", + "bbox": [ + 89, + 757, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6928d690523c365b98323bbd57e7fa4d0b84754a197a8c52a34011fe72cb79b6.jpg", + "image_caption": [ + "Figure 1. Existing V-ICL methods require specialized training of foundation models on curated/uncurated data before using them for novel downstream tasks. In contrast, our approach repurposes an off-the-shelf Stable Diffusion model without any additional training or data." + ], + "image_footnote": [], + "bbox": [ + 526, + 277, + 895, + 407 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "not only eliminates task-specific training but also reduces dependency on task-specific annotated datasets. Recent research, such as [3, 16, 28-31, 33], have made promising attempts to harness the potential of in-context learning for computer vision tasks. These works can be broadly divided into two categories based on the datasets that are used for training: (1) uncurated datasets (e.g. CVF [3], S2CV [33]) and (2) curated/annotated task-related datasets (e.g. COCO [15], NYUDv2 [23]).", + "bbox": [ + 511, + 502, + 906, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The first category of approaches such as Visual Prompting [3] and Improv [33] enable foundation models to perform in-context learning by training with an inpainting loss on uncurated datasets (e.g. CVF, S2CV). These datasets consist of images extracted from computer vision papers and consist of grid-like structures of source-target pairs. The combination of this structured input and random masking during training allows the model to implicitly learn the relationships between the source and target images. During inference, when the model is presented with a grid-like canvas of the query image and the example prompt(s) (source-target pairs) along with the location of the output for the query image masked-out (see Fig. 2a), the model is expected to inpaint the prediction.", + "bbox": [ + 511, + 641, + 908, + 851 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The second category of approaches [16, 28-31] enable a foundation model for in-context learning by allowing the model to train on curated datasets (e.g. COCO, NYUDv2)", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was completed during internship at Apple", + "bbox": [ + 107, + 886, + 393, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23604", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d9003c96a323cc4558ba96ed193a7b6170f24899366c22243c83ef07998cb917.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 150, + 87, + 426, + 166 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/85bf4bf65636944260f9bc181941388b88bf311f0f7469e668053ea2944e1b82.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 148, + 181, + 263, + 267 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/11528fe0005272a20065ffa39da83d3287a2d8451192160a885406149400ca71.jpg", + "image_caption": [ + "(c)", + "Figure 2. (a) Existing approaches like Visual Prompting [3] train a foundation model with an inpainting objective on uncurated data. At inference, a grid of query and example prompt are input to the model. (b) Visual Prompting [3] struggles to predict accurately even when the prompt is the same as the query. (c) The proposed approach, when presented with the same prompt as the query, is able to fully leverage the prompt to make accurate predictions." + ], + "image_footnote": [], + "bbox": [ + 312, + 181, + 426, + 267 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "related to the out-of-domain tasks that the model is expected to adapt to. While these methods do not employ task-specific loss functions, they still rely on task-representative datasets. For example, to enable the Painter [28] model to adapt to open-vocabulary segmentation (e.g. FSS-1000 [14]), the model is trained on a related dataset like the COCO segmentation dataset. In other words, as long as the model has encountered a task during training, it can perform that task on query images from unseen categories. However, this implies that the model needs to access large annotated datasets during training thereby undermining the key advantage of V-ICL. Furthermore, the generalization of such models is limited to tasks related to those seen during training. These methods show promising early outcomes of vision foundation models on unseen tasks, however, achieving similar success as LLMs in NLP domain with off-the-shelf vision foundation models (e.g. Stable Diffusion [20], ViT [7]) remains an area yet to be fully explored.", + "bbox": [ + 91, + 401, + 483, + 674 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although these existing V-ICL approaches show promising outcomes, they involve the use of additional training steps and/or data from out-of-domain tasks to which the model is being used to adapt. Different from existing approaches, in this work, we focus on an approach that is more in-line with the ICL methods from the NLP community. Specifically, we address the question — Can we use off-the-shelf foundation models for visual in-context learning without any additional training steps or data? To this end, we propose the Stable Diffusion based visual in-context learning (SD-VICL) pipeline where we show that an off-the-shelf Stable Diffusion model can be repurposed, without any fine-tuning, to adapt to several new out-of-domain", + "bbox": [ + 89, + 675, + 483, + 872 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tasks (see Fig. 1). To achieve this, we formulate an in-place attention re-computation within the self-attention layers of the Stable Diffusion architecture that explicitly incorporates the context between the query and example prompts. This formulation is based on the drawbacks (Fig. 2) of existing approaches like Visual Prompting [3] where these models are unable to fully leverage the example prompt.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, we note that existing approaches such as Bar et al. [3] and Xu et al. [33] demonstrate the benefits of using multiple example prompts to boost the prediction abilities of the V-ICL foundation models. However, their approach of ensembling in the image space by creating a composite of multiple examples comes at the cost of reduced the resolution per prompt — limiting the potential of prompt ensembling. To address this issue, inspired by feature ensembling in SegGPT [29], we perform prompt ensembling in the latent space for in-context learning. Note that the feature ensembling in SegGPT involves averaging of query features after each attention layer which results in uniform weighting of all prompts. In contrast, we perform implicitly-weighted prompt ensembling within each attention layer thereby enabling the model to better infer the information from multiple prompts.", + "bbox": [ + 511, + 198, + 906, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct extensive evaluations to demonstrate the effectiveness of the proposed approach. Specifically, we show that a pre-trained Stable Diffusion model has the ability to adapt to various tasks like foreground segmentation, single object detection, semantic segmentation, keypoint detection, edge detection, and colorization.", + "bbox": [ + 511, + 440, + 905, + 530 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, the main contributions of this work are as follows,", + "bbox": [ + 511, + 532, + 905, + 561 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The first training-free method to enable visual in-context learning properties within a foundation model, setting a new direction for visual in-context learning research.", + "- A novel pipeline that explicitly incorporates the context between the query image and prompts by introducing an in-place attention re-computation within the self-attention layers of an off-the-shelf Stable Diffusion model.", + "- Extensive evaluations of the proposed approach to demonstrate its ability to generalize to multiple out-of-domain tasks.", + "- Implicitly-weighted prompt ensembling, which enables effective use of multiple prompts for V-ICL." + ], + "bbox": [ + 513, + 563, + 905, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Method", + "text_level": 1, + "bbox": [ + 511, + 761, + 604, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Motivation and Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 787, + 774, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In ICL, a foundation model is expected to infer the task based on a few source-target example prompts and predict the appropriate output for the query image. More importantly, as demonstrated in the NLP community, in-context learning is supposed to be an emergent property of a foundation model where the model does not require addi", + "bbox": [ + 511, + 809, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1Refer Sec. B of supp. for a thorough discussion on related work.", + "bbox": [ + 107, + 886, + 455, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23605", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tional training. However, existing visual in-context learning approaches involve training and/or use of additional task-related data. Further, when we investigated V-ICL approaches based on uncurated datasets $^2$ (e.g. Bar et al. [3]) through multiple experiments, we observe the following, (1) the model fails to reconstruct the input prompt, despite it being unmasked in the input, as shown in the top-left image of the foundation model's output in Fig. 2a, (2) the model struggles in the presence of multiple objects in the image as shown in Fig. 2a where it incorrectly predicts the bird feeder as the foreground, and (3) when provided with the query and its groundtruth as the example prompt, the model struggles to make correct predictions despite having access to the groundtruth for the query (see Fig. 2b). These observations indicate that the model is unable to fully leverage the prompts. We attribute this inability of the model to lack of appropriate context interpretation (1) between the source and target of the example prompt which is essential for task inference, and (2) between the query and the source of the example prompt which is essential for accurate predictions.", + "bbox": [ + 88, + 90, + 482, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To alleviate these limitations, we propose a novel inference-only Stable Diffusion based visual in-context learning pipeline (SD-VICL), that unlike existing approaches, requires no additional training. Stable Diffusion [20] is a latent diffusion model that generates high-quality images by iteratively refining random noise through a denoising process. The denoising process at a given time step $t$ employs a denoising U-Net which comprises of multiple self-attention layers operating at resolutions $16 \\times 16$ , $32 \\times 32$ , and $64 \\times 64$ . At each self-attention layer ( $l$ ), the input features of the intermediate noise latent, $\\phi(z_{t})$ , are transformed to Query ( $Q$ ), Key ( $K$ ), and Value ( $V$ ) vectors using linear layers. The $Q$ and $K$ vectors are used to compute the self-attention map using:", + "bbox": [ + 89, + 393, + 483, + 604 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha^ {(t)} = \\operatorname {s o f t m a x} \\left(\\frac {Q ^ {(t)} \\cdot K ^ {(t) ^ {T}}}{\\sqrt {d}}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 609, + 482, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which captures the correspondences within the image. The parameter $d$ denotes the feature dimension of the $Q$ vector. The self-attention map $(\\alpha)$ is used to update the intermediate feature map using the feature update, $\\Delta \\phi$ , which is computed as,", + "bbox": [ + 89, + 652, + 483, + 726 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\phi^ {(t)} = \\alpha^ {(t)} \\cdot V ^ {(t)}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 726, + 482, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This updated set of features, along with the updates from the cross-attention layers are used to compute the intermediate noise predictions. In this paper, we focus only on the self-attention computations. Hence, for simplicity, we write the iterative denoising process as", + "bbox": [ + 89, + 746, + 483, + 821 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz ^ {(t - 1)} = z ^ {(t)} - \\epsilon \\left(Q ^ {(t)}, K ^ {(t)}, V ^ {(t)}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 825, + 482, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the predicted noise, $\\epsilon$ , is a function of the $Q$ , $K$ , and $V$ vectors of the self-attention layers.", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Repurposing Stable Diffusion for Visual In-Context Learning (SD-VICL)", + "text_level": 1, + "bbox": [ + 511, + 130, + 906, + 162 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As discussed previously, for successful in-context learning, a foundation model needs to appropriately infer (1) the task: the relationship between the prompt image $(A)$ and prompt groundtruth $(B)$ , and (2) the context: the relationship between the query image $(C)$ and the prompt image $(A)$ . In an attempt to ensure that these relationships are effectively inferred, we formulate a novel attention computation in place of the traditional self-attention computation in the upsample layers of the denoising U-Net. This formulation is inspired by a few recent research [1, 5, 18, 26] that repurpose the self-attention layers of Stable Diffusion for tasks like image editing and style transfer.", + "bbox": [ + 509, + 167, + 906, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "First, each of the input images: prompt image $(A)$ , prompt groundtruth $(B)$ , and query image $(C)$ are inverted to the Stable Diffusion's noise space to obtain, $z_A^{(T)}$ , $z_B^{(T)}$ , and $z_C^{(T)}$ respectively, using an off-the-shelf inversion model [13]. Then, each of these images are iteratively denoised in parallel without any changes to the default denoising pipeline as follows:", + "bbox": [ + 511, + 349, + 905, + 460 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {p} ^ {(t - 1)} = z _ {p} ^ {(t)} - \\epsilon \\left(Q _ {p} ^ {(t)}, K _ {p} ^ {(t)}, V _ {p} ^ {(t)}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 468, + 906, + 487 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $p\\in \\{A,B,C\\}$ $t\\in [T,1],t\\in \\mathbb{Z}^{+}$", + "bbox": [ + 511, + 496, + 787, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since the output image is expected to be structurally similar to the query image for most vision tasks, we initialize the noise space in the prediction pipeline $D$ using the noise space of the query image (i.e., $z_{D}^{(T)} \\gets z_{C}^{(T)}$ ).", + "bbox": [ + 511, + 512, + 905, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Different from Eq. (4), the query prediction is denoised using,", + "bbox": [ + 511, + 574, + 905, + 603 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {D} ^ {(t - 1)} = z _ {D} ^ {(t)} - \\epsilon \\left(Q _ {C} ^ {(t)}, K _ {A} ^ {(t)}, V _ {B} ^ {(t)}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 603, + 903, + 621 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $t\\in [T,1],t\\in \\mathbb{Z}^{+}$", + "bbox": [ + 511, + 625, + 683, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Expanding the feature update formulation for $D$ we get,", + "bbox": [ + 531, + 638, + 901, + 655 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\phi_ {D} ^ {(t)} = \\alpha_ {D} ^ {(t)} \\cdot V _ {B} ^ {(t)}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 662, + 903, + 683 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\phi_ {D} ^ {(t)} = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {C} ^ {(t)} \\cdot K _ {A} ^ {(t) ^ {T}}}{\\tau \\cdot \\sqrt {d}}\\right) \\cdot V _ {B} ^ {(t)}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 684, + 903, + 722 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast to Eqs. (1) and (2), the attention map is formulated using the Query vector of the query image $(C)$ and the Key vector of the prompt image $(A)$ . The $Q_{C}$ vector comprises of the semantics of each spatial location in $C$ , and the $K_{A}$ vector offers the context within $A$ that each query can attend to. Hence, each element in the attention map, i.e., $\\alpha_{D(i,j)}^{t}$ , captures how the $i^{th}$ patch of the query image correlates with the $j^{th}$ patch of the prompt image. This formulation explicitly enforces the context between the query image and the prompt image. The additional temperature hyperparameter $(\\tau)$ (in Eq. (7)) for the softmax computation", + "bbox": [ + 511, + 731, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2We use approaches based on uncurated datasets as the basis for our analysis/formulations since they do not explicitly use images from out-of-domain tasks for training and are much closer to in-context learning in the NLP community as compared to the approaches that use curated datasets.", + "bbox": [ + 89, + 849, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23606", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/630f5816d67c371b5e728424dbd09b019ddde6f498acbdbeb9752e51478cee34.jpg", + "image_caption": [ + "Figure 3. Latent representations of the query and prompt images from the Stable Diffusion model's VAE encoder are transformed to the noise space using [13]. Each noisy latent is iteratively denoised using the denoising U-Net. Prediction $(D)$ path is initialized with the noise space of $C$ , and at each denoising step, the computations within the self-attention layers are modified to enhance visual ICL capabilities. With the Query (from prompt image- $A$ ), the Key (from prompt groundtruth- $B$ ), and the Value (from query image- $C$ ) vectors, we reformulate the intermediate feature update to infuse context between the example prompts and the query image into the foundation model. This process is iteratively performed for $T$ denoising steps to obtain the denoised prediction latent, which is then fed to the VAE decoder to obtain the final prediction $D$ ." + ], + "image_footnote": [], + "bbox": [ + 148, + 85, + 851, + 329 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "controls the sharpness of the correlation. By adjusting $\\tau$ , we can modulate the model's focus on attention/correlation between $C$ and $A$ . A lower $\\tau$ sharpens the attention, emphasizing stronger correlation, while a higher $\\tau$ smoothens the attention, allowing for a broader distribution of focus across multiple correlations. Furthermore, we use the Value vector from the prompt groundtruth $(B)$ to compute $\\Delta \\phi_D^{(t)}$ . This use of value vector from a different latent, in contrast to the standard implementation of cross-attention [27], where both Key and Value vectors typically come from the same set of intermediate latents, is to facilitate the prediction to be in the same domain as that of the prompt groundtruth $B$ . This formulation, which is in-place of the computation within the standard self-attention layer of Stable Diffusion, facilitates the explicit infusion of the context between the query image and the prompt in addition to improved task inference. However, the modification of the default self-attention computation and the use of $Q$ , $K$ , $V$ vectors that are extracted from three distinct images create a domain gap that needs to be addressed to enhance the quality of the prediction. Hence, we adapt the attention map contrasting, swap-guidance, and AdaIN mechanisms employed in [1].", + "bbox": [ + 91, + 448, + 483, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Attention map contrasting supplements the function of the temperature hyperparameter $(\\tau)$ , where it enhances the focus on the relevant regions (i.e. semantically similar regions) between $C$ and $A$ , while attenuating the irrelevant (i.e. color similarity between unrelated objects). Additionally, this operation adjusts the scale of the values to suit the pre-trained Stable Diffusion pipeline. This contrasting op", + "bbox": [ + 89, + 795, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "eration can be expressed as,", + "bbox": [ + 511, + 449, + 699, + 462 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {D} ^ {(t)} = \\mu \\left(\\alpha_ {D} ^ {(t)}\\right) + \\beta \\cdot \\left(\\alpha_ {D} ^ {(t)} - \\mu \\left(\\alpha_ {D} ^ {(t)}\\right)\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 468, + 906, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mu$ is the mean operation and $\\beta$ is a hyperparameter that controls the scale.", + "bbox": [ + 511, + 496, + 906, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Swap-guidance is derived from classifier-free diffusion guidance implementation proposed by Ho and Salimans [11]. Given the noise prediction from the default self-attention formulation $(\\eta_{default}^{(t)})$ and the noise prediction from our modified formulation $(\\eta_{modified}^{(t)})$ , the final noise prediction $(\\eta^{(t)})$ can be expressed as,", + "bbox": [ + 511, + 525, + 905, + 626 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\eta^ {(t)} = \\eta_ {\\text {d e f a u l t}} ^ {(t)} + \\frac {\\gamma \\cdot (T - t)}{T} \\cdot \\left(\\eta_ {\\text {m o d i f i e d}} ^ {(t)} - \\eta_ {\\text {d e f a u l t}} ^ {(t)}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 631, + 906, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\gamma$ is a hyperparameter controlling the scale. This mechanism gradually incorporates the modified noise prediction as denoising progresses, which directs the denoising process through denser regions of the Stable Diffusion's generative pipeline, thus alleviating unwarranted artifacts.", + "bbox": [ + 511, + 662, + 905, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adaptive instance normalization (AdaIN), which was proposed by Huang and Belongie [12] is used to align the color distribution between the prediction $(D)$ , which is initialized using the noise space of the query image $(C)$ , and the final groundtruth/task color space (i.e. color space of $B$ ). AdaIN is implemented as follows,", + "bbox": [ + 511, + 737, + 905, + 820 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {D} ^ {(t)} \\leftarrow \\operatorname {A d a I N} \\left(z _ {D} ^ {(t)}, z _ {B} ^ {(t)}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 825, + 903, + 843 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {D} ^ {(t)} \\leftarrow \\frac {z _ {D} ^ {(t)} - \\mu \\left(z _ {D} ^ {(t)}\\right)}{\\sigma \\left(z _ {D} ^ {(t)}\\right)} \\cdot \\sigma \\left(z _ {B} ^ {(t)}\\right) + \\mu \\left(z _ {B} ^ {(t)}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 845, + 903, + 881 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mu$ and $\\sigma$ are mean and standard deviation operations.", + "bbox": [ + 511, + 886, + 903, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23607", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3. Implicitly-Weighted Prompt Ensembling (IWPE)", + "text_level": 1, + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Bar et al. [3] and Xu et al. [33] demonstrate that providing the model with multiple source-target example prompts improves the prediction performance of visual in-context learning. These methods ensemble multiple prompts at the input level by stitching together prompt images and corresponding groundtruths to form a grid-like structure in the form of a composite image. This composite image is then input to the model, which is expected to leverage information from multiple prompts to make a better prediction compared to the single prompt scenario. However, for a fixed-resolution foundation model, as the number of source-target examples increases, the effective size of each image in the grid is reduced. This leads to performance deterioration, mainly because of the loss of details in the input [33].", + "bbox": [ + 89, + 127, + 483, + 338 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address this issue, instead of ensembling prompts in the image space through a composite input, SegGPT [29] proposed a feature space ensembling method. In this method, multiple prompts are processed in parallel and aggregated by averaging the features at the end of each attention layer. This approach assumes that all the example prompts are equally informative and they are uniformly weighted during the averaging process. However, such uniform weighting of prompts is sub-optimal since it prevents the model from benefiting from more relevant prompts. To this end, we propose a simple but effective ensembling method by integrating the ensembling into the attention computation, allowing the prompt patches to be implicitly weighted based on their correspondences with each query patch. Specifically, we concatenate each of the $K$ and $V$ vectors from each of the prompts prior to computing $\\Delta \\phi_D^{(t)}$ . Specifically, for $n$ number of prompts, we re-write Eq. (7):", + "bbox": [ + 89, + 339, + 483, + 598 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\phi_ {D} ^ {(t)} = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {C} ^ {(t)} \\cdot \\left(\\bigoplus_ {i = 1} ^ {n} K _ {A _ {i}} ^ {(t)}\\right) ^ {T}}{\\tau \\cdot \\sqrt {d}}\\right) \\cdot \\left(\\bigoplus_ {i = 1} ^ {n} V _ {B _ {i}} ^ {(t)}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 602, + 482, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\oplus$ is the concatenation operation and $i$ is the index for each prompt.", + "bbox": [ + 89, + 650, + 482, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3. Experiments and Evaluations", + "text_level": 1, + "bbox": [ + 89, + 694, + 362, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To demonstrate the ability of the proposed approach to generalize across different tasks, we evaluate it on six downstream tasks. For all these tasks, we use the unsupervised prompt retrieval [35] to select the candidates for the prompt images. Specifically, this method chooses the nearest neighbors of the query image as prompt candidates. The nearest neighbor retrieval is based on the cosine similarity of CLIP's vision encoder [19] embeddings between the query image and the example prompt images.", + "bbox": [ + 89, + 719, + 482, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For a fair comparison, we use Visual Prompting [3] and IMProv [33] as baselines. Although these approaches involve a training step, unlike [28, 31], they do not rely on", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "curated or annotated data that come from related out-of-domain tasks. Additionally, since IMProv supports supplementary text guidance, we include results with and without text guidance in our comparisons. Further, we also evaluate and compare the effectiveness of multiple prompts. Please refer to the supplementary for details on datasets, evaluation metrics, and sensitivity analysis for different hyperparameters. For the results in the following subsections, we choose one set of hyperparameters that provided optimal performance for all the tasks.", + "bbox": [ + 511, + 90, + 905, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1. Results and Analysis", + "text_level": 1, + "bbox": [ + 511, + 251, + 709, + 267 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantitative metrics for all the six tasks are reported in Tab. 1. The proposed approach consistently outperforms existing uncurated based approaches in all the tasks.", + "bbox": [ + 511, + 273, + 905, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Foreground segmentation: With a single prompt, the proposed method achieves absolute improvements of $8.9\\%$ and $3.2\\%$ in mean intersection over union (mIoU) on an average across all splits3 in the Pascal-5i dataset [22], compared to Visual Prompting and IMProv, respectively. The proposed ensembling method with five prompts using the repurposed Stable Diffusion model further improves the mIoU by $11.6\\%$ (in absolute terms) as compared to the single prompt case.", + "bbox": [ + 511, + 319, + 905, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Single object detection: Similar to foreground segmentation, with a single prompt, the proposed method improves the mIoU by $5.3\\%$ and $7.1\\%$ (in absolute terms) on average across all splits3 in the Pascal-5i dataset [22] as compared to Visual Prompting and IMProv, respectively. The use of multiple prompts through implicitly-weighted prompt assembling results in an additional $5.6\\%$ of improvement as compared to the single prompt case.", + "bbox": [ + 511, + 455, + 905, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Semantic segmentation: On the Cityscapes dataset [6], our method with a single prompt results in an absolute improvement of $0.8\\%$ in mIoU and $3.4\\%$ increase in accuracy as compared to Visual Prompting. Similarly, we observe an absolute gain of $6.1\\%$ in mIoU and a $6.5\\%$ in accuracy as compared to IMProv. With IWPE, the mIoU and accuracy further improve by $4.9\\%$ and $6.8\\%$ (in absolute terms) as compared to the single prompt based approach. The proposed approach with a single prompt in the case of semantic segmentation achieves lower performance improvements over existing methods as compared to other tasks like foreground segmentation and keypoint detection. However, with additional prompts, the proposed method is able to achieve much higher gains similar to what is observed in other tasks. This performance boost can be attributed to two key factors: (1) multiple prompts help resolve conflicts between similar or overlapping classes by mitigating inter-class attention ambiguities, and (2) a single prompt is unlikely to contain all classes that are potentially present in the query image of a multi-class task and in such cases, ac", + "bbox": [ + 511, + 575, + 905, + 878 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "See supplementary for the performance on each split.", + "bbox": [ + 529, + 886, + 820, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23608", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/190352a69e33d040911ffed063d4f2259ca8211c588d9edcef0c0533b5d0cdfe.jpg", + "image_caption": [ + "Figure 4. Qualitative evaluation, where we compare the performance of Visual Prompting [3] and IMProv [33] with our proposed method on six different tasks. It can be seen that our method produces visually superior results as compared to the baselines." + ], + "image_footnote": [], + "bbox": [ + 166, + 88, + 831, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "cess to multiple prompts increases the probability that all relevant classes are available to the model for making accurate predictions.", + "bbox": [ + 89, + 522, + 482, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Keypoint detection: When evaluated on the DeepFashion dataset [17], our method demonstrates a substantial improvement, reducing the MSE by $6 \\times$ while achieving a $7 \\times$ gain in percentage of correct keypoints (PCK), compared to the best-performing baseline. Furthermore, the use of five prompts enhances these metrics, yielding additional relative gains of $18.5\\%$ for MSE and $6.5\\%$ for PCK.", + "bbox": [ + 89, + 566, + 482, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Edge detection: On the NYUDv2 dataset [23], our approach reduces the mean squared error (MSE) by $72.0\\%$ and LPIPS [34] by $60.6\\%$ as compared to the best performing existing approach. Note that these metrics improve an additional $24.5\\%$ and $21.4\\%$ , respectively, with our prompt ensembling method.", + "bbox": [ + 89, + 672, + 482, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Colorization: Evaluating on the ImageNet dataset [21], our model reduces LPIPS by $27.9\\%$ and FID [10] by $49.0\\%$ , relative to IMProv, which yields the next best performance, and shows further gains of $19.1\\%$ and $16.0\\%$ respectively, with five prompts.", + "bbox": [ + 89, + 763, + 482, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overall, our method outperforms both Visual Prompting and IMProv across all tasks by considerable margins. Additionally, the integration of multiple prompts using IWPE consistently yields performance improvements over", + "bbox": [ + 89, + 840, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the single-prompt scenario.", + "bbox": [ + 511, + 522, + 694, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Supplementing the quantitative results, we also demonstrate visual comparisons of the outputs for each task in Fig. 4 $^{4}$ , further highlighting the superior performance of our method over baselines. We observe that while both Visual Prompting and IMProv perform reasonably in single object detection and foreground segmentation, they show a weaker performance in tasks such as colorization, edge detection, and semantic segmentation, where capturing fine details and preserving structural information of the query is crucial. Moreover, both baselines notably struggle with keypoint detection, whereas the proposed approach demonstrates far superior results, indicating its robustness to different out-of-domain tasks.", + "bbox": [ + 511, + 537, + 906, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2. Additional Evaluations", + "text_level": 1, + "bbox": [ + 511, + 746, + 725, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with V-ICL models trained on task-related data: To provide a comprehensive evaluation, we compare our training-free approach against prior methods trained explicitly on task-related data, including Painter [28], LVM [2], and Prompt Diffusion [31]. Quantitative results are presented in Tab. 2, and qualitative comparisons are shown in Fig. 7 (supplementary). Despite the lack of any ex", + "bbox": [ + 511, + 768, + 905, + 875 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "4See supplementary for additional qualitative results for each task.", + "bbox": [ + 529, + 886, + 880, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23609", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/51c8a9189460b5be061cf6092d15756f7404c608bfc8dc0c2e6a28d7c4ccc39f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelForeground Seg. (Pascal-5i) mIoU ↑Object Det. (Pascal-5i) mIoU ↑Semantic Segmentation (Cityscapes) mIoU ↑Keypoint Detection (DeepFashion) Acc. ↑MSE ↓PCK ↑Edge Detection (NYUDv2) MSE ↓LPIPS ↓Colorization (ImageNet) LPIPS ↓FID ↓
Number of Example Prompts: 1
Visual Prompting [3]35.0446.2921.3071.5235.8310.870.10060.39250.4166111.06
IMProv (w/o text) [33]39.5043.8215.9768.4431.3018.380.10590.42780.3895104.74
IMProv (w/ text) [33]40.6944.4215.8067.8844.672.810.11250.51460.3899105.90
SD-VICL (ours)43.9251.5522.0874.935.3677.190.02820.15480.281053.40
Number of Example Prompts: 5
Visual Prompting36.3748.0921.7070.9536.5710.320.10000.39000.4156112.91
SD-VICL (ours)55.4957.1027.0181.754.3782.240.02130.12160.227244.84
", + "bbox": [ + 112, + 88, + 883, + 234 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/83fffe75b23d41e1c446e6671b0851ee9a8b42534f280bdb70566b61f6be467e.jpg", + "table_caption": [ + "Table 1. Quantitative performance comparison of the proposed approach with recent approaches on foreground segmentation, single object detection, semantic segmentation, keypoint detection, edge detection, and colorization." + ], + "table_footnote": [], + "table_body": "
ModelFG Seg. mIoU ↑Obj. Det. mIoU ↑Edge Det.Colorization
MSE ↓LPIPS ↓LPIPS ↓FID ↓
Painter [28]55.0954.280.09260.72940.347464.16
LVM [2]50.9852.670.04990.42590.314256.40
Prompt Diffusion [31]17.5326.860.04300.33110.699798.91
SD-VICL (ours)55.4957.100.02130.12160.227244.84
", + "bbox": [ + 93, + 282, + 531, + 361 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "plicit training, our method consistently outperforms all three baselines across multiple vision tasks. Analyzing the results, we observe that all three models tend to overfit to their training tasks, leading to poor generalization when presented with novel tasks. For instance, Painter incorrectly outputs a depth map instead of edges, despite the prompt pair defining the task has an edge map as the target (Fig. 7 in supplementary). Further, while Painter and LVM perform well on tasks, foreground segmentation and object detection, when the query image consists of only a single class, they struggle in multi-class scenarios, often segmenting all classes instead of the intended one (Fig. 8a in supplementary). Additionally, LVM exhibits inconsistencies in its outputs, where for a given task, despite the format/domain of the inputs remaining unchanged, we observe that the generated output belongs to diverse domains (Fig. 8b in supplementary). Moreover, Prompt Diffusion performs notably poorly on most tasks, often generating incorrect outputs, with the exception of edge detection — a task it was exposed to during training. Ideally, a V-ICL model should infer context and task from the input, but all three models fall short in this aspect. In contrast, our training-free method demonstrates superior generalization and effective context and task inference as intended by V-ICL, underscoring the advantages of uncovering V-ICL properties without additional training. For a detailed discussion and visual examples, we refer the reader to Sec. E in the supplementary.", + "bbox": [ + 88, + 411, + 485, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with task-specific models based on Stable Diffusion: Here, we compare our method against approaches that adapt Stable Diffusion to facilitate specific tasks, such as semantic segmentation and keypoint detection. DAS [24] is one such work that proposes a pipeline", + "bbox": [ + 89, + 825, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/10a6e0a4003c886f96d5844506c12cee0667110da712dd1cafe3c4137382336c.jpg", + "table_caption": [ + "Table 2. Extended quantitative evaluations against V-ICL models that train on task-related data." + ], + "table_footnote": [], + "table_body": "
ModelSem.Seg.Keypoint Det.
mIoU ↑Acc. ↑MSE ↓PCK ↑
DAS [24]21.276.0--
StableKeypoints [9]--6.4670.0
SD-VICL (ours)31.5282.724.3782.24
", + "bbox": [ + 576, + 286, + 898, + 357 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Comparison with task-specific models based on Stable Diffusion.", + "bbox": [ + 571, + 367, + 901, + 395 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "that aggregates the self-attention layers of Stable Diffusion iteratively to predict segmentation masks for each class facilitating semantic segmentation. Similarly, StableKeypoints [9] optimizes on the text embeddings such that cross-attention maps condense to landmarks in a given image facilitating keypoint detection. We present the performance comparisons against these task-specific models in Tab. 3, where our approach outperforms both [24] and [9]. For semantic segmentation, we follow the evaluation protocol of DAS (i.e. Hungarian matching) to ensure a fair comparison and achieve absolute improvements of $10.3\\%$ in mIoU and $6.7\\%$ in accuracy. In keypoint detection, we surpass StableKeypoints with a $32.4\\%$ reduction in MSE and $12.2\\%$ improvement in PCK. While DAS and StableKeypoints achieve strong results in their respective tasks, they remain constrained by their task-specific nature. In contrast, our method not only delivers superior performance but also generalizes across diverse tasks without requiring any modification in the inference. Furthermore, unlike these methods, which derive predictions from intermediate feature or attention maps, our model directly utilizes the generative process to produce outputs, making it inherently more flexible and effective.", + "bbox": [ + 511, + 411, + 906, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 771, + 666, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we present the details of various ablations conducted to demonstrate the significance of key design choices. Specifically, we evaluate the effectiveness of implicitly-weighted prompt ensembling and examine the impact of number of prompts on performance. For additional ablations, including alternative attention formulations and sensitivity analyses on hyperparameters such as", + "bbox": [ + 511, + 794, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23610", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8d11171e4926762ce0d3d66f945f8b2c6840a47e1598d27c673b6e5953d122fa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
# PromptsEnsemblingmIoU ↑Acc. ↑
1N/A22.0874.93
5FE22.9778.84
5IWPE27.0181.75
", + "bbox": [ + 155, + 88, + 415, + 151 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation of prompt-ensembling evaluated on semantic segmentation.", + "bbox": [ + 89, + 155, + 482, + 184 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "attention temperature, self-attention resolution, contrastive strength, swap-guidance scale, and AdaIN, we refer the reader to Sec. F in the supplementary.", + "bbox": [ + 89, + 200, + 482, + 246 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of implicitly-weighted prompt ensembling (IWPE): Here, we compare our proposed ensembling method, IWPE, to SegGPT's feature ensembling (FE) approach. For all tasks except semantic segmentation, we observe that both methods yield similar gains over the single prompt scenario. In the case of semantic segmentation, IWPE demonstrates significantly better performance. As tabulated in Tab. 4, while FE yields an absolute improvement of merely $0.9\\%$ in mIoU and $3.9\\%$ in accuracy, IWPE achieves substantial gains, improving mIoU by $4.9\\%$ and accuracy by $6.8\\%$ over the single prompt case. This notable improvement is primarily due to the capability of our ensembling method to dynamically weigh each prompt based on its correspondence with the query image. In contrast, as discussed previously, the feature ensembling approach assumes that all prompts are equally informative, which is sub-optimal, especially in the case of tasks involving multiple classes like semantic segmentation. For instance, one prompt may provide a stronger correlation with a specific class, while another prompt, would have a weaker association with that class. Our method by implicitly weighing the influence of different prompts based on their relative correspondence to the query, better captures these nuances, resulting in a significant boost in performance compared to SegGPT's uniformly weighted feature ensembling. Therefore, while both FE and IWPE perform at par in simpler tasks where the prompts are equally informative, our method is more beneficial for tasks that involve complex multi-class structures.", + "bbox": [ + 91, + 247, + 483, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of number of prompts: Here, we investigate the effect of the number of prompts on the performance and the inference speed. As illustrated in Fig. 5a, the mIoU scores improve with an increasing number of prompts. This result can be attributed to the model's ability to effectively leverage multiple prompts to better infer the task and establish the correspondence between the query and the prompts. The additional prompts help reduce ambiguities that may arise from using a single prompt. However, these gains come at the cost of reduced inference speed; specifically using five prompts cuts the inference speed to approximately half that of a single prompt. Besides the prompts, the number of denoising steps in the Stable Diffusion model is a key factor impacting the inference speed. This led us to the question", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/95a3b0d67bc9aea035eea773115a6a9bf36187125884c0e058a51fd27b512476.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 522, + 90, + 705, + 198 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9ae8dd8c3a21a68e05c6107bf3ecaf5e8e4601bb5f2d51018995c631547dbf85.jpg", + "image_caption": [ + "(b)", + "Figure 5. Effect of the number of prompts on the performance and inference speed5. While (a) depicts the variation in mIoU and the inference speed, with the number of prompts, (b) illustrates the variation in mIoU with the number of denoising steps for both single and five prompt cases." + ], + "image_footnote": [], + "bbox": [ + 709, + 87, + 911, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "— With the availability of multiple prompts, does the model require as many denoising steps as in the single prompt case? In other words, we explored if we can trade-off a few denoising steps without impacting the performance when additional context is provided by multiple prompts. To this end, we conducted experiments where we evaluated the SD-VICL approach with different denoising steps. As shown in Fig. 5b, the mIoU for the five-prompt scenario at 30 denoising steps surpasses the maximum performance of the single-prompt scenario. Moreover, the maximum performance for the single-prompt case, achieved with 70 denoising steps, requires 38.4 seconds of inference time. In contrast, the five-prompt case achieves higher performance in just 35.9 seconds with 30 denoising steps5.", + "bbox": [ + 511, + 305, + 908, + 517 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 537, + 633, + 553 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel pipeline for visual in-context learning that explicitly incorporates the context between the query image and the prompt. Unlike existing visual in-context learning methods which rely on training or fine-tuning foundation models on curated or uncurated data to enable in-context learning, our approach is entirely training-free. Specifically, we introduce an in-place attention re-computation within the self-attention layers of an off-the-shelf Stable Diffusion model. Additionally, the proposed implicitly-weighted prompt ensembling technique facilitates effective integration of context through multiple prompts by implicitly weighing the prompts based on the relative correspondences. The versatility of the proposed method is demonstrated by its successful generalization across six different tasks, where it outperforms both V-ICL models trained on uncurated and task-related data as well as task-specific models based on Stable Diffusion. Crucially, our method achieves these results without requiring any additional training or annotated task-specific datasets, highlighting its ability to uncover and leverage the V-ICL capabilities of Stable Diffusion.", + "bbox": [ + 511, + 564, + 908, + 869 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "5The inference times are based on evaluations on a single A100 GPU.", + "bbox": [ + 529, + 886, + 901, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23611", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer. In ACM SIGGRAPH 2024 Conference Papers, pages 1–12, 2024. 3, 4", + "[2] Yutong Bai, Xinyang Geng, Karttikeya Mangalam, Amir Bar, Alan L Yuille, Trevor Darrell, Jitendra Malik, and Alexei A Efros. Sequential modeling enables scalable learning for large vision models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22861-22872, 2024. 6, 7", + "[3] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei Efros. Visual prompting via image inpainting. Advances in Neural Information Processing Systems, 35:25005-25017, 2022. 1, 2, 3, 5, 6, 7", + "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1", + "[5] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560-22570, 2023. 3", + "[6] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5", + "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2", + "[8] Yaru Hao, Haoyu Song, Li Dong, Shaohan Huang, Zewen Chi, Wenhui Wang, Shuming Ma, and Furu Wei. Language models are general-purpose interfaces. arXiv preprint arXiv:2206.06336, 2022. 1", + "[9] Eric Hedlin, Gopal Sharma, Shweta Mahajan, Xingzhe He, Hossam Isack, Abhishek Kar, Helge Rhodin, Andrea Tagliasacchi, and Kwang Moo Yi. Unsupervised keypoints from pretrained diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22820-22830, 2024. 7", + "[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6", + "[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4", + "[12] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceed-" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 4", + "[13] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12469-12478, 2024. 3, 4", + "[14] Xiang Li, Tianhan Wei, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. Fss-1000: A 1000-class dataset for few-shot segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2869–2878, 2020. 2", + "[15] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 1", + "[16] Weihuang Liu, Xi Shen, Chi-Man Pun, and Xiaodong Cun. Explicit visual prompting for low-level structure segmentations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19434-19445, 2023. 1", + "[17] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1096-1104, 2016. 6", + "[18] Or Patashnik, Daniel Garibi, Idan Azuri, Hadar Averbuch-Elor, and Daniel Cohen-Or. Localizing object-level shape variations with text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23051–23061, 2023. 3", + "[19] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5", + "[20] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3", + "[21] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 6", + "[22] Amirreza Shaban, Shray Bansal, Zhen Liu, Irfan Essa, and Byron Boots. One-shot learning for semantic segmentation. arXiv preprint arXiv:1709.03410, 2017. 5", + "[23] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 1, 6" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23612", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Junjiao Tian, Lavisha Aggarwal, Andrea Colaco, Zsolt Kira, and Mar Gonzalez-Franco. Diffuse attend and segment: Unsupervised zero-shot segmentation using stable diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3554-3563, 2024. 7", + "[25] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1", + "[26] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 3", + "[27] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 4", + "[28] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6830-6839, 2023. 1, 2, 5, 6, 7", + "[29] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023. 2,5", + "[30] Xinshun Wang, Zhongbin Fang, Xia Li, Xiangtai Li, Chen Chen, and Mengyuan Liu. Skeleton-in-context: Unified skeleton sequence modeling with in-context learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2436-2446, 2024.", + "[31] Zhendong Wang, Yifan Jiang, Yadong Lu, Pengcheng He, Weizhu Chen, Zhangyang Wang, Mingyuan Zhou, et al. Incontext learning unlocked for diffusion models. Advances in Neural Information Processing Systems, 36:8542-8562, 2023. 1, 5, 6, 7", + "[32] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. Emergent abilities of large language models. Transactions on Machine Learning Research, 2022. 1", + "[33] Jiarui Xu, Yossi Gandelsman, Amir Bar, Jianwei Yang, Jianfeng Gao, Trevor Darrell, and Xiaolong Wang. Improv: Inpainting-based multimodal prompting for computer vision tasks. arXiv preprint arXiv:2312.01771, 2023. 1, 2, 5, 6, 7", + "[34] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6", + "[35] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. What makes good examples for visual in-context learning? Advances in Neural Information Processing Systems, 36:17773-17794, 2023. 5" + ], + "bbox": [ + 91, + 92, + 480, + 824 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23613", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Stable Diffusion Models are Secretly Good at Visual In-Context Learning/94d38526-8561-492c-b6db-5bcfedbe9083_model.json b/2025/Stable Diffusion Models are Secretly Good at Visual In-Context Learning/94d38526-8561-492c-b6db-5bcfedbe9083_model.json new file mode 100644 index 0000000000000000000000000000000000000000..006ccea3200e5ecdfb15b46888eeb72cd64872bf --- /dev/null +++ b/2025/Stable Diffusion Models are Secretly Good at Visual In-Context Learning/94d38526-8561-492c-b6db-5bcfedbe9083_model.json @@ -0,0 +1,1936 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.13, + 0.131, + 0.868, + 0.154 + ], + "angle": 0, + "content": "Stable Diffusion Models are Secretly Good at Visual In-Context Learning" + }, + { + "type": "text", + "bbox": [ + 0.17, + 0.179, + 0.826, + 0.248 + ], + "angle": 0, + "content": "Trevine Oorloff\\(^{1,2*}\\) Ali Shafahi\\(^{1}\\) Vishwanath Sindagi\\(^{1}\\) Wele Gedara Chaminda Bandara\\(^{1}\\) Amin Ghiasi\\(^{1}\\) Charan Prakash\\(^{1}\\) Reza Ardekani\\(^{1}\\) Apple \\(^{2}\\)University of Maryland - College Park" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.28, + 0.327, + 0.296 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.312, + 0.485, + 0.705 + ], + "angle": 0, + "content": "Large language models (LLM) in natural language processing (NLP) have demonstrated great potential for in-context learning (ICL) — the ability to leverage a few sets of example prompts to adapt to various tasks without having to explicitly update the model weights. ICL has recently been explored for computer vision tasks with promising early outcomes. These approaches involve specialized training and/or additional data that complicate the process and limit its generalizability. In this work, we show that off-the-shelf Stable Diffusion models can be repurposed for visual in-context learning (V-ICL). Specifically, we formulate an in-place attention re-computation within the self-attention layers of the Stable Diffusion architecture that explicitly incorporates context between the query and example prompts. Without any additional fine-tuning, we show that this repurposed Stable Diffusion model is able to adapt to six different tasks: foreground segmentation, single object detection, semantic segmentation, keypoint detection, edge detection, and colorization. For example, the proposed approach improves the mean intersection over union (mIoU) for the foreground segmentation task on Pascal-5i dataset by \\(8.9\\%\\) and \\(3.2\\%\\) over recent methods such as Visual Prompting and IMProv, respectively. Additionally, we show that the proposed method is able to effectively leverage multiple prompts through ensembling to infer the task better and further improve the performance." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.733, + 0.222, + 0.749 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.758, + 0.483, + 0.879 + ], + "angle": 0, + "content": "In-context learning (ICL) refers to the paradigm where a foundation model leverages exemplar source-target pair(s), known as prompts, to infer the task and perform the inferred task on an input, known as query. ICL is an emergent property of large language models (LLMs) and has been widely explored in the natural language processing (NLP) domain [4, 8, 25, 32]. ICL facilitates a model to adapt to novel or out-of-domain tasks without the need for fine-tuning which" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.279, + 0.896, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.419, + 0.908, + 0.49 + ], + "angle": 0, + "content": "Figure 1. Existing V-ICL methods require specialized training of foundation models on curated/uncurated data before using them for novel downstream tasks. In contrast, our approach repurposes an off-the-shelf Stable Diffusion model without any additional training or data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.503, + 0.907, + 0.639 + ], + "angle": 0, + "content": "not only eliminates task-specific training but also reduces dependency on task-specific annotated datasets. Recent research, such as [3, 16, 28-31, 33], have made promising attempts to harness the potential of in-context learning for computer vision tasks. These works can be broadly divided into two categories based on the datasets that are used for training: (1) uncurated datasets (e.g. CVF [3], S2CV [33]) and (2) curated/annotated task-related datasets (e.g. COCO [15], NYUDv2 [23])." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.909, + 0.852 + ], + "angle": 0, + "content": "The first category of approaches such as Visual Prompting [3] and Improv [33] enable foundation models to perform in-context learning by training with an inpainting loss on uncurated datasets (e.g. CVF, S2CV). These datasets consist of images extracted from computer vision papers and consist of grid-like structures of source-target pairs. The combination of this structured input and random masking during training allows the model to implicitly learn the relationships between the source and target images. During inference, when the model is presented with a grid-like canvas of the query image and the example prompt(s) (source-target pairs) along with the location of the output for the query image masked-out (see Fig. 2a), the model is expected to inpaint the prediction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "The second category of approaches [16, 28-31] enable a foundation model for in-context learning by allowing the model to train on curated datasets (e.g. COCO, NYUDv2)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.395, + 0.901 + ], + "angle": 0, + "content": "*This work was completed during internship at Apple" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23604" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.151, + 0.088, + 0.428, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.279, + 0.17, + 0.296, + 0.181 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.149, + 0.182, + 0.264, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.199, + 0.269, + 0.214, + 0.278 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.313, + 0.182, + 0.427, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.362, + 0.269, + 0.376, + 0.278 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.29, + 0.484, + 0.388 + ], + "angle": 0, + "content": "Figure 2. (a) Existing approaches like Visual Prompting [3] train a foundation model with an inpainting objective on uncurated data. At inference, a grid of query and example prompt are input to the model. (b) Visual Prompting [3] struggles to predict accurately even when the prompt is the same as the query. (c) The proposed approach, when presented with the same prompt as the query, is able to fully leverage the prompt to make accurate predictions." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.402, + 0.485, + 0.675 + ], + "angle": 0, + "content": "related to the out-of-domain tasks that the model is expected to adapt to. While these methods do not employ task-specific loss functions, they still rely on task-representative datasets. For example, to enable the Painter [28] model to adapt to open-vocabulary segmentation (e.g. FSS-1000 [14]), the model is trained on a related dataset like the COCO segmentation dataset. In other words, as long as the model has encountered a task during training, it can perform that task on query images from unseen categories. However, this implies that the model needs to access large annotated datasets during training thereby undermining the key advantage of V-ICL. Furthermore, the generalization of such models is limited to tasks related to those seen during training. These methods show promising early outcomes of vision foundation models on unseen tasks, however, achieving similar success as LLMs in NLP domain with off-the-shelf vision foundation models (e.g. Stable Diffusion [20], ViT [7]) remains an area yet to be fully explored." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.676, + 0.484, + 0.873 + ], + "angle": 0, + "content": "Although these existing V-ICL approaches show promising outcomes, they involve the use of additional training steps and/or data from out-of-domain tasks to which the model is being used to adapt. Different from existing approaches, in this work, we focus on an approach that is more in-line with the ICL methods from the NLP community. Specifically, we address the question — Can we use off-the-shelf foundation models for visual in-context learning without any additional training steps or data? To this end, we propose the Stable Diffusion based visual in-context learning (SD-VICL) pipeline where we show that an off-the-shelf Stable Diffusion model can be repurposed, without any fine-tuning, to adapt to several new out-of-domain" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "tasks (see Fig. 1). To achieve this, we formulate an in-place attention re-computation within the self-attention layers of the Stable Diffusion architecture that explicitly incorporates the context between the query and example prompts. This formulation is based on the drawbacks (Fig. 2) of existing approaches like Visual Prompting [3] where these models are unable to fully leverage the example prompt." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.907, + 0.44 + ], + "angle": 0, + "content": "Furthermore, we note that existing approaches such as Bar et al. [3] and Xu et al. [33] demonstrate the benefits of using multiple example prompts to boost the prediction abilities of the V-ICL foundation models. However, their approach of ensembling in the image space by creating a composite of multiple examples comes at the cost of reduced the resolution per prompt — limiting the potential of prompt ensembling. To address this issue, inspired by feature ensembling in SegGPT [29], we perform prompt ensembling in the latent space for in-context learning. Note that the feature ensembling in SegGPT involves averaging of query features after each attention layer which results in uniform weighting of all prompts. In contrast, we perform implicitly-weighted prompt ensembling within each attention layer thereby enabling the model to better infer the information from multiple prompts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.906, + 0.531 + ], + "angle": 0, + "content": "We conduct extensive evaluations to demonstrate the effectiveness of the proposed approach. Specifically, we show that a pre-trained Stable Diffusion model has the ability to adapt to various tasks like foreground segmentation, single object detection, semantic segmentation, keypoint detection, edge detection, and colorization." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.533, + 0.906, + 0.562 + ], + "angle": 0, + "content": "To summarize, the main contributions of this work are as follows," + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.564, + 0.906, + 0.609 + ], + "angle": 0, + "content": "- The first training-free method to enable visual in-context learning properties within a foundation model, setting a new direction for visual in-context learning research." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.61, + 0.906, + 0.669 + ], + "angle": 0, + "content": "- A novel pipeline that explicitly incorporates the context between the query image and prompts by introducing an in-place attention re-computation within the self-attention layers of an off-the-shelf Stable Diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.67, + 0.905, + 0.713 + ], + "angle": 0, + "content": "- Extensive evaluations of the proposed approach to demonstrate its ability to generalize to multiple out-of-domain tasks." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.715, + 0.905, + 0.745 + ], + "angle": 0, + "content": "- Implicitly-weighted prompt ensembling, which enables effective use of multiple prompts for V-ICL." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.564, + 0.906, + 0.745 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.762, + 0.605, + 0.777 + ], + "angle": 0, + "content": "2. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.776, + 0.803 + ], + "angle": 0, + "content": "2.1. Motivation and Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In ICL, a foundation model is expected to infer the task based on a few source-target example prompts and predict the appropriate output for the query image. More importantly, as demonstrated in the NLP community, in-context learning is supposed to be an emergent property of a foundation model where the model does not require addi" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.456, + 0.901 + ], + "angle": 0, + "content": "1Refer Sec. B of supp. for a thorough discussion on related work." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23605" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.483, + 0.394 + ], + "angle": 0, + "content": "tional training. However, existing visual in-context learning approaches involve training and/or use of additional task-related data. Further, when we investigated V-ICL approaches based on uncurated datasets\\(^2\\) (e.g. Bar et al. [3]) through multiple experiments, we observe the following, (1) the model fails to reconstruct the input prompt, despite it being unmasked in the input, as shown in the top-left image of the foundation model's output in Fig. 2a, (2) the model struggles in the presence of multiple objects in the image as shown in Fig. 2a where it incorrectly predicts the bird feeder as the foreground, and (3) when provided with the query and its groundtruth as the example prompt, the model struggles to make correct predictions despite having access to the groundtruth for the query (see Fig. 2b). These observations indicate that the model is unable to fully leverage the prompts. We attribute this inability of the model to lack of appropriate context interpretation (1) between the source and target of the example prompt which is essential for task inference, and (2) between the query and the source of the example prompt which is essential for accurate predictions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.394, + 0.484, + 0.606 + ], + "angle": 0, + "content": "To alleviate these limitations, we propose a novel inference-only Stable Diffusion based visual in-context learning pipeline (SD-VICL), that unlike existing approaches, requires no additional training. Stable Diffusion [20] is a latent diffusion model that generates high-quality images by iteratively refining random noise through a denoising process. The denoising process at a given time step \\(t\\) employs a denoising U-Net which comprises of multiple self-attention layers operating at resolutions \\(16 \\times 16\\), \\(32 \\times 32\\), and \\(64 \\times 64\\). At each self-attention layer (\\(l\\)), the input features of the intermediate noise latent, \\(\\phi(z_{t})\\), are transformed to Query (\\(Q\\)), Key (\\(K\\)), and Value (\\(V\\)) vectors using linear layers. The \\(Q\\) and \\(K\\) vectors are used to compute the self-attention map using:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.611, + 0.483, + 0.649 + ], + "angle": 0, + "content": "\\[\n\\alpha^ {(t)} = \\operatorname {s o f t m a x} \\left(\\frac {Q ^ {(t)} \\cdot K ^ {(t) ^ {T}}}{\\sqrt {d}}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.654, + 0.484, + 0.727 + ], + "angle": 0, + "content": "which captures the correspondences within the image. The parameter \\(d\\) denotes the feature dimension of the \\(Q\\) vector. The self-attention map \\((\\alpha)\\) is used to update the intermediate feature map using the feature update, \\(\\Delta \\phi\\), which is computed as," + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.727, + 0.483, + 0.743 + ], + "angle": 0, + "content": "\\[\n\\Delta \\phi^ {(t)} = \\alpha^ {(t)} \\cdot V ^ {(t)}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.747, + 0.484, + 0.822 + ], + "angle": 0, + "content": "This updated set of features, along with the updates from the cross-attention layers are used to compute the intermediate noise predictions. In this paper, we focus only on the self-attention computations. Hence, for simplicity, we write the iterative denoising process as" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.826, + 0.483, + 0.845 + ], + "angle": 0, + "content": "\\[\nz ^ {(t - 1)} = z ^ {(t)} - \\epsilon \\left(Q ^ {(t)}, K ^ {(t)}, V ^ {(t)}\\right), \\tag {3}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.851, + 0.484, + 0.901 + ], + "angle": 0, + "content": "2We use approaches based on uncurated datasets as the basis for our analysis/formulations since they do not explicitly use images from out-of-domain tasks for training and are much closer to in-context learning in the NLP community as compared to the approaches that use curated datasets." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "where the predicted noise, \\(\\epsilon\\), is a function of the \\(Q\\), \\(K\\), and \\(V\\) vectors of the self-attention layers." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.131, + 0.907, + 0.163 + ], + "angle": 0, + "content": "2.2. Repurposing Stable Diffusion for Visual In-Context Learning (SD-VICL)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.169, + 0.907, + 0.35 + ], + "angle": 0, + "content": "As discussed previously, for successful in-context learning, a foundation model needs to appropriately infer (1) the task: the relationship between the prompt image \\((A)\\) and prompt groundtruth \\((B)\\), and (2) the context: the relationship between the query image \\((C)\\) and the prompt image \\((A)\\). In an attempt to ensure that these relationships are effectively inferred, we formulate a novel attention computation in place of the traditional self-attention computation in the upsample layers of the denoising U-Net. This formulation is inspired by a few recent research [1, 5, 18, 26] that repurpose the self-attention layers of Stable Diffusion for tasks like image editing and style transfer." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.351, + 0.906, + 0.461 + ], + "angle": 0, + "content": "First, each of the input images: prompt image \\((A)\\), prompt groundtruth \\((B)\\), and query image \\((C)\\) are inverted to the Stable Diffusion's noise space to obtain, \\(z_A^{(T)}\\), \\(z_B^{(T)}\\), and \\(z_C^{(T)}\\) respectively, using an off-the-shelf inversion model [13]. Then, each of these images are iteratively denoised in parallel without any changes to the default denoising pipeline as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.469, + 0.907, + 0.488 + ], + "angle": 0, + "content": "\\[\nz _ {p} ^ {(t - 1)} = z _ {p} ^ {(t)} - \\epsilon \\left(Q _ {p} ^ {(t)}, K _ {p} ^ {(t)}, V _ {p} ^ {(t)}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.497, + 0.789, + 0.513 + ], + "angle": 0, + "content": "where \\(p\\in \\{A,B,C\\}\\) \\(t\\in [T,1],t\\in \\mathbb{Z}^{+}\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.513, + 0.906, + 0.576 + ], + "angle": 0, + "content": "Since the output image is expected to be structurally similar to the query image for most vision tasks, we initialize the noise space in the prediction pipeline \\( D \\) using the noise space of the query image (i.e., \\( z_{D}^{(T)} \\gets z_{C}^{(T)} \\))." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.575, + 0.906, + 0.604 + ], + "angle": 0, + "content": "Different from Eq. (4), the query prediction is denoised using," + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.604, + 0.905, + 0.622 + ], + "angle": 0, + "content": "\\[\nz _ {D} ^ {(t - 1)} = z _ {D} ^ {(t)} - \\epsilon \\left(Q _ {C} ^ {(t)}, K _ {A} ^ {(t)}, V _ {B} ^ {(t)}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.626, + 0.684, + 0.64 + ], + "angle": 0, + "content": "where \\(t\\in [T,1],t\\in \\mathbb{Z}^{+}\\)" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.64, + 0.903, + 0.656 + ], + "angle": 0, + "content": "Expanding the feature update formulation for \\(D\\) we get," + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.664, + 0.905, + 0.684 + ], + "angle": 0, + "content": "\\[\n\\Delta \\phi_ {D} ^ {(t)} = \\alpha_ {D} ^ {(t)} \\cdot V _ {B} ^ {(t)}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.685, + 0.905, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\Delta \\phi_ {D} ^ {(t)} = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {C} ^ {(t)} \\cdot K _ {A} ^ {(t) ^ {T}}}{\\tau \\cdot \\sqrt {d}}\\right) \\cdot V _ {B} ^ {(t)}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.732, + 0.907, + 0.902 + ], + "angle": 0, + "content": "In contrast to Eqs. (1) and (2), the attention map is formulated using the Query vector of the query image \\((C)\\) and the Key vector of the prompt image \\((A)\\). The \\(Q_{C}\\) vector comprises of the semantics of each spatial location in \\(C\\), and the \\(K_{A}\\) vector offers the context within \\(A\\) that each query can attend to. Hence, each element in the attention map, i.e., \\(\\alpha_{D(i,j)}^{t}\\), captures how the \\(i^{th}\\) patch of the query image correlates with the \\(j^{th}\\) patch of the prompt image. This formulation explicitly enforces the context between the query image and the prompt image. The additional temperature hyperparameter \\((\\tau)\\) (in Eq. (7)) for the softmax computation" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23606" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.15, + 0.087, + 0.852, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.339, + 0.908, + 0.437 + ], + "angle": 0, + "content": "Figure 3. Latent representations of the query and prompt images from the Stable Diffusion model's VAE encoder are transformed to the noise space using [13]. Each noisy latent is iteratively denoised using the denoising U-Net. Prediction \\((D)\\) path is initialized with the noise space of \\(C\\), and at each denoising step, the computations within the self-attention layers are modified to enhance visual ICL capabilities. With the Query (from prompt image-\\(A\\)), the Key (from prompt groundtruth-\\(B\\)), and the Value (from query image-\\(C\\)) vectors, we reformulate the intermediate feature update to infuse context between the example prompts and the query image into the foundation model. This process is iteratively performed for \\(T\\) denoising steps to obtain the denoised prediction latent, which is then fed to the VAE decoder to obtain the final prediction \\(D\\)." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.449, + 0.485, + 0.784 + ], + "angle": 0, + "content": "controls the sharpness of the correlation. By adjusting \\(\\tau\\), we can modulate the model's focus on attention/correlation between \\(C\\) and \\(A\\). A lower \\(\\tau\\) sharpens the attention, emphasizing stronger correlation, while a higher \\(\\tau\\) smoothens the attention, allowing for a broader distribution of focus across multiple correlations. Furthermore, we use the Value vector from the prompt groundtruth \\((B)\\) to compute \\(\\Delta \\phi_D^{(t)}\\). This use of value vector from a different latent, in contrast to the standard implementation of cross-attention [27], where both Key and Value vectors typically come from the same set of intermediate latents, is to facilitate the prediction to be in the same domain as that of the prompt groundtruth \\(B\\). This formulation, which is in-place of the computation within the standard self-attention layer of Stable Diffusion, facilitates the explicit infusion of the context between the query image and the prompt in addition to improved task inference. However, the modification of the default self-attention computation and the use of \\(Q\\), \\(K\\), \\(V\\) vectors that are extracted from three distinct images create a domain gap that needs to be addressed to enhance the quality of the prediction. Hence, we adapt the attention map contrasting, swap-guidance, and AdaIN mechanisms employed in [1]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Attention map contrasting supplements the function of the temperature hyperparameter \\((\\tau)\\), where it enhances the focus on the relevant regions (i.e. semantically similar regions) between \\(C\\) and \\(A\\), while attenuating the irrelevant (i.e. color similarity between unrelated objects). Additionally, this operation adjusts the scale of the values to suit the pre-trained Stable Diffusion pipeline. This contrasting op" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.45, + 0.7, + 0.463 + ], + "angle": 0, + "content": "eration can be expressed as," + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.469, + 0.907, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {D} ^ {(t)} = \\mu \\left(\\alpha_ {D} ^ {(t)}\\right) + \\beta \\cdot \\left(\\alpha_ {D} ^ {(t)} - \\mu \\left(\\alpha_ {D} ^ {(t)}\\right)\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.497, + 0.907, + 0.525 + ], + "angle": 0, + "content": "where \\(\\mu\\) is the mean operation and \\(\\beta\\) is a hyperparameter that controls the scale." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.526, + 0.906, + 0.627 + ], + "angle": 0, + "content": "Swap-guidance is derived from classifier-free diffusion guidance implementation proposed by Ho and Salimans [11]. Given the noise prediction from the default self-attention formulation \\((\\eta_{default}^{(t)})\\) and the noise prediction from our modified formulation \\((\\eta_{modified}^{(t)})\\), the final noise prediction \\((\\eta^{(t)})\\) can be expressed as," + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.632, + 0.907, + 0.658 + ], + "angle": 0, + "content": "\\[\n\\eta^ {(t)} = \\eta_ {\\text {d e f a u l t}} ^ {(t)} + \\frac {\\gamma \\cdot (T - t)}{T} \\cdot \\left(\\eta_ {\\text {m o d i f i e d}} ^ {(t)} - \\eta_ {\\text {d e f a u l t}} ^ {(t)}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.663, + 0.906, + 0.737 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is a hyperparameter controlling the scale. This mechanism gradually incorporates the modified noise prediction as denoising progresses, which directs the denoising process through denser regions of the Stable Diffusion's generative pipeline, thus alleviating unwarranted artifacts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.738, + 0.906, + 0.821 + ], + "angle": 0, + "content": "Adaptive instance normalization (AdaIN), which was proposed by Huang and Belongie [12] is used to align the color distribution between the prediction \\((D)\\), which is initialized using the noise space of the query image \\((C)\\), and the final groundtruth/task color space (i.e. color space of \\(B\\)). AdaIN is implemented as follows," + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.826, + 0.905, + 0.844 + ], + "angle": 0, + "content": "\\[\nz _ {D} ^ {(t)} \\leftarrow \\operatorname {A d a I N} \\left(z _ {D} ^ {(t)}, z _ {B} ^ {(t)}\\right), \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.847, + 0.905, + 0.882 + ], + "angle": 0, + "content": "\\[\nz _ {D} ^ {(t)} \\leftarrow \\frac {z _ {D} ^ {(t)} - \\mu \\left(z _ {D} ^ {(t)}\\right)}{\\sigma \\left(z _ {D} ^ {(t)}\\right)} \\cdot \\sigma \\left(z _ {B} ^ {(t)}\\right) + \\mu \\left(z _ {B} ^ {(t)}\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.887, + 0.905, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mu\\) and \\(\\sigma\\) are mean and standard deviation operations." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23607" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.483, + 0.122 + ], + "angle": 0, + "content": "2.3. Implicitly-Weighted Prompt Ensembling (IWPE)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.128, + 0.484, + 0.339 + ], + "angle": 0, + "content": "Bar et al. [3] and Xu et al. [33] demonstrate that providing the model with multiple source-target example prompts improves the prediction performance of visual in-context learning. These methods ensemble multiple prompts at the input level by stitching together prompt images and corresponding groundtruths to form a grid-like structure in the form of a composite image. This composite image is then input to the model, which is expected to leverage information from multiple prompts to make a better prediction compared to the single prompt scenario. However, for a fixed-resolution foundation model, as the number of source-target examples increases, the effective size of each image in the grid is reduced. This leads to performance deterioration, mainly because of the loss of details in the input [33]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.34, + 0.484, + 0.599 + ], + "angle": 0, + "content": "To address this issue, instead of ensembling prompts in the image space through a composite input, SegGPT [29] proposed a feature space ensembling method. In this method, multiple prompts are processed in parallel and aggregated by averaging the features at the end of each attention layer. This approach assumes that all the example prompts are equally informative and they are uniformly weighted during the averaging process. However, such uniform weighting of prompts is sub-optimal since it prevents the model from benefiting from more relevant prompts. To this end, we propose a simple but effective ensembling method by integrating the ensembling into the attention computation, allowing the prompt patches to be implicitly weighted based on their correspondences with each query patch. Specifically, we concatenate each of the \\(K\\) and \\(V\\) vectors from each of the prompts prior to computing \\(\\Delta \\phi_D^{(t)}\\). Specifically, for \\(n\\) number of prompts, we re-write Eq. (7):" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.603, + 0.483, + 0.648 + ], + "angle": 0, + "content": "\\[\n\\Delta \\phi_ {D} ^ {(t)} = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {C} ^ {(t)} \\cdot \\left(\\bigoplus_ {i = 1} ^ {n} K _ {A _ {i}} ^ {(t)}\\right) ^ {T}}{\\tau \\cdot \\sqrt {d}}\\right) \\cdot \\left(\\bigoplus_ {i = 1} ^ {n} V _ {B _ {i}} ^ {(t)}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.651, + 0.483, + 0.682 + ], + "angle": 0, + "content": "where \\(\\oplus\\) is the concatenation operation and \\(i\\) is the index for each prompt." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.695, + 0.364, + 0.712 + ], + "angle": 0, + "content": "3. Experiments and Evaluations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.856 + ], + "angle": 0, + "content": "To demonstrate the ability of the proposed approach to generalize across different tasks, we evaluate it on six downstream tasks. For all these tasks, we use the unsupervised prompt retrieval [35] to select the candidates for the prompt images. Specifically, this method chooses the nearest neighbors of the query image as prompt candidates. The nearest neighbor retrieval is based on the cosine similarity of CLIP's vision encoder [19] embeddings between the query image and the example prompt images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "For a fair comparison, we use Visual Prompting [3] and IMProv [33] as baselines. Although these approaches involve a training step, unlike [28, 31], they do not rely on" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.243 + ], + "angle": 0, + "content": "curated or annotated data that come from related out-of-domain tasks. Additionally, since IMProv supports supplementary text guidance, we include results with and without text guidance in our comparisons. Further, we also evaluate and compare the effectiveness of multiple prompts. Please refer to the supplementary for details on datasets, evaluation metrics, and sensitivity analysis for different hyperparameters. For the results in the following subsections, we choose one set of hyperparameters that provided optimal performance for all the tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.252, + 0.71, + 0.268 + ], + "angle": 0, + "content": "3.1. Results and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.274, + 0.906, + 0.319 + ], + "angle": 0, + "content": "Quantitative metrics for all the six tasks are reported in Tab. 1. The proposed approach consistently outperforms existing uncurated based approaches in all the tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.32, + 0.906, + 0.456 + ], + "angle": 0, + "content": "Foreground segmentation: With a single prompt, the proposed method achieves absolute improvements of \\(8.9\\%\\) and \\(3.2\\%\\) in mean intersection over union (mIoU) on an average across all splits3 in the Pascal-5i dataset [22], compared to Visual Prompting and IMProv, respectively. The proposed ensembling method with five prompts using the repurposed Stable Diffusion model further improves the mIoU by \\(11.6\\%\\) (in absolute terms) as compared to the single prompt case." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.906, + 0.576 + ], + "angle": 0, + "content": "Single object detection: Similar to foreground segmentation, with a single prompt, the proposed method improves the mIoU by \\(5.3\\%\\) and \\(7.1\\%\\) (in absolute terms) on average across all splits3 in the Pascal-5i dataset [22] as compared to Visual Prompting and IMProv, respectively. The use of multiple prompts through implicitly-weighted prompt assembling results in an additional \\(5.6\\%\\) of improvement as compared to the single prompt case." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.577, + 0.906, + 0.879 + ], + "angle": 0, + "content": "Semantic segmentation: On the Cityscapes dataset [6], our method with a single prompt results in an absolute improvement of \\(0.8\\%\\) in mIoU and \\(3.4\\%\\) increase in accuracy as compared to Visual Prompting. Similarly, we observe an absolute gain of \\(6.1\\%\\) in mIoU and a \\(6.5\\%\\) in accuracy as compared to IMProv. With IWPE, the mIoU and accuracy further improve by \\(4.9\\%\\) and \\(6.8\\%\\) (in absolute terms) as compared to the single prompt based approach. The proposed approach with a single prompt in the case of semantic segmentation achieves lower performance improvements over existing methods as compared to other tasks like foreground segmentation and keypoint detection. However, with additional prompts, the proposed method is able to achieve much higher gains similar to what is observed in other tasks. This performance boost can be attributed to two key factors: (1) multiple prompts help resolve conflicts between similar or overlapping classes by mitigating inter-class attention ambiguities, and (2) a single prompt is unlikely to contain all classes that are potentially present in the query image of a multi-class task and in such cases, ac" + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.887, + 0.821, + 0.901 + ], + "angle": 0, + "content": "See supplementary for the performance on each split." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23608" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.167, + 0.089, + 0.833, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.481, + 0.908, + 0.511 + ], + "angle": 0, + "content": "Figure 4. Qualitative evaluation, where we compare the performance of Visual Prompting [3] and IMProv [33] with our proposed method on six different tasks. It can be seen that our method produces visually superior results as compared to the baselines." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.523, + 0.483, + 0.568 + ], + "angle": 0, + "content": "cess to multiple prompts increases the probability that all relevant classes are available to the model for making accurate predictions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.483, + 0.674 + ], + "angle": 0, + "content": "Keypoint detection: When evaluated on the DeepFashion dataset [17], our method demonstrates a substantial improvement, reducing the MSE by \\(6 \\times\\) while achieving a \\(7 \\times\\) gain in percentage of correct keypoints (PCK), compared to the best-performing baseline. Furthermore, the use of five prompts enhances these metrics, yielding additional relative gains of \\(18.5\\%\\) for MSE and \\(6.5\\%\\) for PCK." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.765 + ], + "angle": 0, + "content": "Edge detection: On the NYUDv2 dataset [23], our approach reduces the mean squared error (MSE) by \\(72.0\\%\\) and LPIPS [34] by \\(60.6\\%\\) as compared to the best performing existing approach. Note that these metrics improve an additional \\(24.5\\%\\) and \\(21.4\\%\\), respectively, with our prompt ensembling method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.842 + ], + "angle": 0, + "content": "Colorization: Evaluating on the ImageNet dataset [21], our model reduces LPIPS by \\(27.9\\%\\) and FID [10] by \\(49.0\\%\\), relative to IMProv, which yields the next best performance, and shows further gains of \\(19.1\\%\\) and \\(16.0\\%\\) respectively, with five prompts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Overall, our method outperforms both Visual Prompting and IMProv across all tasks by considerable margins. Additionally, the integration of multiple prompts using IWPE consistently yields performance improvements over" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.523, + 0.696, + 0.538 + ], + "angle": 0, + "content": "the single-prompt scenario." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.735 + ], + "angle": 0, + "content": "Supplementing the quantitative results, we also demonstrate visual comparisons of the outputs for each task in Fig. 4\\(^{4}\\), further highlighting the superior performance of our method over baselines. We observe that while both Visual Prompting and IMProv perform reasonably in single object detection and foreground segmentation, they show a weaker performance in tasks such as colorization, edge detection, and semantic segmentation, where capturing fine details and preserving structural information of the query is crucial. Moreover, both baselines notably struggle with keypoint detection, whereas the proposed approach demonstrates far superior results, indicating its robustness to different out-of-domain tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.747, + 0.727, + 0.761 + ], + "angle": 0, + "content": "3.2. Additional Evaluations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.769, + 0.906, + 0.876 + ], + "angle": 0, + "content": "Comparison with V-ICL models trained on task-related data: To provide a comprehensive evaluation, we compare our training-free approach against prior methods trained explicitly on task-related data, including Painter [28], LVM [2], and Prompt Diffusion [31]. Quantitative results are presented in Tab. 2, and qualitative comparisons are shown in Fig. 7 (supplementary). Despite the lack of any ex" + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.887, + 0.882, + 0.901 + ], + "angle": 0, + "content": "4See supplementary for additional qualitative results for each task." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23609" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.114, + 0.089, + 0.885, + 0.236 + ], + "angle": 0, + "content": "
ModelForeground Seg. (Pascal-5i) mIoU ↑Object Det. (Pascal-5i) mIoU ↑Semantic Segmentation (Cityscapes) mIoU ↑Keypoint Detection (DeepFashion) Acc. ↑MSE ↓PCK ↑Edge Detection (NYUDv2) MSE ↓LPIPS ↓Colorization (ImageNet) LPIPS ↓FID ↓
Number of Example Prompts: 1
Visual Prompting [3]35.0446.2921.3071.5235.8310.870.10060.39250.4166111.06
IMProv (w/o text) [33]39.5043.8215.9768.4431.3018.380.10590.42780.3895104.74
IMProv (w/ text) [33]40.6944.4215.8067.8844.672.810.11250.51460.3899105.90
SD-VICL (ours)43.9251.5522.0874.935.3677.190.02820.15480.281053.40
Number of Example Prompts: 5
Visual Prompting36.3748.0921.7070.9536.5710.320.10000.39000.4156112.91
SD-VICL (ours)55.4957.1027.0181.754.3782.240.02130.12160.227244.84
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.241, + 0.907, + 0.27 + ], + "angle": 0, + "content": "Table 1. Quantitative performance comparison of the proposed approach with recent approaches on foreground segmentation, single object detection, semantic segmentation, keypoint detection, edge detection, and colorization." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.283, + 0.532, + 0.362 + ], + "angle": 0, + "content": "
ModelFG Seg. mIoU ↑Obj. Det. mIoU ↑Edge Det.Colorization
MSE ↓LPIPS ↓LPIPS ↓FID ↓
Painter [28]55.0954.280.09260.72940.347464.16
LVM [2]50.9852.670.04990.42590.314256.40
Prompt Diffusion [31]17.5326.860.04300.33110.699798.91
SD-VICL (ours)55.4957.100.02130.12160.227244.84
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.372, + 0.534, + 0.4 + ], + "angle": 0, + "content": "Table 2. Extended quantitative evaluations against V-ICL models that train on task-related data." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.412, + 0.486, + 0.82 + ], + "angle": 0, + "content": "plicit training, our method consistently outperforms all three baselines across multiple vision tasks. Analyzing the results, we observe that all three models tend to overfit to their training tasks, leading to poor generalization when presented with novel tasks. For instance, Painter incorrectly outputs a depth map instead of edges, despite the prompt pair defining the task has an edge map as the target (Fig. 7 in supplementary). Further, while Painter and LVM perform well on tasks, foreground segmentation and object detection, when the query image consists of only a single class, they struggle in multi-class scenarios, often segmenting all classes instead of the intended one (Fig. 8a in supplementary). Additionally, LVM exhibits inconsistencies in its outputs, where for a given task, despite the format/domain of the inputs remaining unchanged, we observe that the generated output belongs to diverse domains (Fig. 8b in supplementary). Moreover, Prompt Diffusion performs notably poorly on most tasks, often generating incorrect outputs, with the exception of edge detection — a task it was exposed to during training. Ideally, a V-ICL model should infer context and task from the input, but all three models fall short in this aspect. In contrast, our training-free method demonstrates superior generalization and effective context and task inference as intended by V-ICL, underscoring the advantages of uncovering V-ICL properties without additional training. For a detailed discussion and visual examples, we refer the reader to Sec. E in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Comparison with task-specific models based on Stable Diffusion: Here, we compare our method against approaches that adapt Stable Diffusion to facilitate specific tasks, such as semantic segmentation and keypoint detection. DAS [24] is one such work that proposes a pipeline" + }, + { + "type": "table", + "bbox": [ + 0.577, + 0.287, + 0.9, + 0.358 + ], + "angle": 0, + "content": "
ModelSem.Seg.Keypoint Det.
mIoU ↑Acc. ↑MSE ↓PCK ↑
DAS [24]21.276.0--
StableKeypoints [9]--6.4670.0
SD-VICL (ours)31.5282.724.3782.24
" + }, + { + "type": "table_caption", + "bbox": [ + 0.572, + 0.368, + 0.903, + 0.396 + ], + "angle": 0, + "content": "Table 3. Comparison with task-specific models based on Stable Diffusion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.412, + 0.907, + 0.759 + ], + "angle": 0, + "content": "that aggregates the self-attention layers of Stable Diffusion iteratively to predict segmentation masks for each class facilitating semantic segmentation. Similarly, StableKeypoints [9] optimizes on the text embeddings such that cross-attention maps condense to landmarks in a given image facilitating keypoint detection. We present the performance comparisons against these task-specific models in Tab. 3, where our approach outperforms both [24] and [9]. For semantic segmentation, we follow the evaluation protocol of DAS (i.e. Hungarian matching) to ensure a fair comparison and achieve absolute improvements of \\(10.3\\%\\) in mIoU and \\(6.7\\%\\) in accuracy. In keypoint detection, we surpass StableKeypoints with a \\(32.4\\%\\) reduction in MSE and \\(12.2\\%\\) improvement in PCK. While DAS and StableKeypoints achieve strong results in their respective tasks, they remain constrained by their task-specific nature. In contrast, our method not only delivers superior performance but also generalizes across diverse tasks without requiring any modification in the inference. Furthermore, unlike these methods, which derive predictions from intermediate feature or attention maps, our model directly utilizes the generative process to produce outputs, making it inherently more flexible and effective." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.667, + 0.789 + ], + "angle": 0, + "content": "3.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this section, we present the details of various ablations conducted to demonstrate the significance of key design choices. Specifically, we evaluate the effectiveness of implicitly-weighted prompt ensembling and examine the impact of number of prompts on performance. For additional ablations, including alternative attention formulations and sensitivity analyses on hyperparameters such as" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23610" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.156, + 0.089, + 0.416, + 0.152 + ], + "angle": 0, + "content": "
# PromptsEnsemblingmIoU ↑Acc. ↑
1N/A22.0874.93
5FE22.9778.84
5IWPE27.0181.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.156, + 0.483, + 0.185 + ], + "angle": 0, + "content": "Table 4. Ablation of prompt-ensembling evaluated on semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.201, + 0.483, + 0.247 + ], + "angle": 0, + "content": "attention temperature, self-attention resolution, contrastive strength, swap-guidance scale, and AdaIN, we refer the reader to Sec. F in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.248, + 0.485, + 0.687 + ], + "angle": 0, + "content": "Effectiveness of implicitly-weighted prompt ensembling (IWPE): Here, we compare our proposed ensembling method, IWPE, to SegGPT's feature ensembling (FE) approach. For all tasks except semantic segmentation, we observe that both methods yield similar gains over the single prompt scenario. In the case of semantic segmentation, IWPE demonstrates significantly better performance. As tabulated in Tab. 4, while FE yields an absolute improvement of merely \\(0.9\\%\\) in mIoU and \\(3.9\\%\\) in accuracy, IWPE achieves substantial gains, improving mIoU by \\(4.9\\%\\) and accuracy by \\(6.8\\%\\) over the single prompt case. This notable improvement is primarily due to the capability of our ensembling method to dynamically weigh each prompt based on its correspondence with the query image. In contrast, as discussed previously, the feature ensembling approach assumes that all prompts are equally informative, which is sub-optimal, especially in the case of tasks involving multiple classes like semantic segmentation. For instance, one prompt may provide a stronger correlation with a specific class, while another prompt, would have a weaker association with that class. Our method by implicitly weighing the influence of different prompts based on their relative correspondence to the query, better captures these nuances, resulting in a significant boost in performance compared to SegGPT's uniformly weighted feature ensembling. Therefore, while both FE and IWPE perform at par in simpler tasks where the prompts are equally informative, our method is more beneficial for tasks that involve complex multi-class structures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Impact of number of prompts: Here, we investigate the effect of the number of prompts on the performance and the inference speed. As illustrated in Fig. 5a, the mIoU scores improve with an increasing number of prompts. This result can be attributed to the model's ability to effectively leverage multiple prompts to better infer the task and establish the correspondence between the query and the prompts. The additional prompts help reduce ambiguities that may arise from using a single prompt. However, these gains come at the cost of reduced inference speed; specifically using five prompts cuts the inference speed to approximately half that of a single prompt. Besides the prompts, the number of denoising steps in the Stable Diffusion model is a key factor impacting the inference speed. This led us to the question" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.091, + 0.707, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.204, + 0.625, + 0.216 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.088, + 0.912, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.803, + 0.204, + 0.82, + 0.215 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.22, + 0.908, + 0.29 + ], + "angle": 0, + "content": "Figure 5. Effect of the number of prompts on the performance and inference speed5. While (a) depicts the variation in mIoU and the inference speed, with the number of prompts, (b) illustrates the variation in mIoU with the number of denoising steps for both single and five prompt cases." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.306, + 0.909, + 0.518 + ], + "angle": 0, + "content": "— With the availability of multiple prompts, does the model require as many denoising steps as in the single prompt case? In other words, we explored if we can trade-off a few denoising steps without impacting the performance when additional context is provided by multiple prompts. To this end, we conducted experiments where we evaluated the SD-VICL approach with different denoising steps. As shown in Fig. 5b, the mIoU for the five-prompt scenario at 30 denoising steps surpasses the maximum performance of the single-prompt scenario. Moreover, the maximum performance for the single-prompt case, achieved with 70 denoising steps, requires 38.4 seconds of inference time. In contrast, the five-prompt case achieves higher performance in just 35.9 seconds with 30 denoising steps5." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.538, + 0.634, + 0.554 + ], + "angle": 0, + "content": "4. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.909, + 0.87 + ], + "angle": 0, + "content": "In this paper, we propose a novel pipeline for visual in-context learning that explicitly incorporates the context between the query image and the prompt. Unlike existing visual in-context learning methods which rely on training or fine-tuning foundation models on curated or uncurated data to enable in-context learning, our approach is entirely training-free. Specifically, we introduce an in-place attention re-computation within the self-attention layers of an off-the-shelf Stable Diffusion model. Additionally, the proposed implicitly-weighted prompt ensembling technique facilitates effective integration of context through multiple prompts by implicitly weighing the prompts based on the relative correspondences. The versatility of the proposed method is demonstrated by its successful generalization across six different tasks, where it outperforms both V-ICL models trained on uncurated and task-related data as well as task-specific models based on Stable Diffusion. Crucially, our method achieves these results without requiring any additional training or annotated task-specific datasets, highlighting its ability to uncover and leverage the V-ICL capabilities of Stable Diffusion." + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.887, + 0.902, + 0.9 + ], + "angle": 0, + "content": "5The inference times are based on evaluations on a single A100 GPU." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23611" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer. In ACM SIGGRAPH 2024 Conference Papers, pages 1–12, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[2] Yutong Bai, Xinyang Geng, Karttikeya Mangalam, Amir Bar, Alan L Yuille, Trevor Darrell, Jitendra Malik, and Alexei A Efros. Sequential modeling enables scalable learning for large vision models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22861-22872, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.483, + 0.31 + ], + "angle": 0, + "content": "[3] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei Efros. Visual prompting via image inpainting. Advances in Neural Information Processing Systems, 35:25005-25017, 2022. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.483, + 0.382 + ], + "angle": 0, + "content": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.483, + 0.451 + ], + "angle": 0, + "content": "[5] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560-22570, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.453, + 0.483, + 0.535 + ], + "angle": 0, + "content": "[6] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.537, + 0.483, + 0.632 + ], + "angle": 0, + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.634, + 0.483, + 0.689 + ], + "angle": 0, + "content": "[8] Yaru Hao, Haoyu Song, Li Dong, Shaohan Huang, Zewen Chi, Wenhui Wang, Shuming Ma, and Furu Wei. Language models are general-purpose interfaces. arXiv preprint arXiv:2206.06336, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.69, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[9] Eric Hedlin, Gopal Sharma, Shweta Mahajan, Xingzhe He, Hossam Isack, Abhishek Kar, Helge Rhodin, Andrea Tagliasacchi, and Kwang Moo Yi. Unsupervised keypoints from pretrained diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22820-22830, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.774, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.845, + 0.483, + 0.873 + ], + "angle": 0, + "content": "[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceed-" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "ings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[13] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12469-12478, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.19, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[14] Xiang Li, Tianhan Wei, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. Fss-1000: A 1000-class dataset for few-shot segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2869–2878, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.26, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[15] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[16] Weihuang Liu, Xi Shen, Chi-Man Pun, and Xiaodong Cun. Explicit visual prompting for low-level structure segmentations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19434-19445, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.413, + 0.905, + 0.483 + ], + "angle": 0, + "content": "[17] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1096-1104, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[18] Or Patashnik, Daniel Garibi, Idan Azuri, Hadar Averbuch-Elor, and Daniel Cohen-Or. Localizing object-level shape variations with text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23051–23061, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[19] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.636, + 0.905, + 0.706 + ], + "angle": 0, + "content": "[20] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[21] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[22] Amirreza Shaban, Shray Bansal, Zhen Liu, Irfan Essa, and Byron Boots. One-shot learning for semantic segmentation. arXiv preprint arXiv:1709.03410, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 1, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23612" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[24] Junjiao Tian, Lavisha Aggarwal, Andrea Colaco, Zsolt Kira, and Mar Gonzalez-Franco. Diffuse attend and segment: Unsupervised zero-shot segmentation using stable diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3554-3563, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[25] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[26] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.304, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[27] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.402 + ], + "angle": 0, + "content": "[28] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6830-6839, 2023. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.404, + 0.482, + 0.457 + ], + "angle": 0, + "content": "[29] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023. 2,5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.46, + 0.482, + 0.529 + ], + "angle": 0, + "content": "[30] Xinshun Wang, Zhongbin Fang, Xia Li, Xiangtai Li, Chen Chen, and Mengyuan Liu. Skeleton-in-context: Unified skeleton sequence modeling with in-context learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2436-2446, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.531, + 0.482, + 0.599 + ], + "angle": 0, + "content": "[31] Zhendong Wang, Yifan Jiang, Yadong Lu, Pengcheng He, Weizhu Chen, Zhangyang Wang, Mingyuan Zhou, et al. Incontext learning unlocked for diffusion models. Advances in Neural Information Processing Systems, 36:8542-8562, 2023. 1, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.601, + 0.482, + 0.669 + ], + "angle": 0, + "content": "[32] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. Emergent abilities of large language models. Transactions on Machine Learning Research, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.671, + 0.482, + 0.727 + ], + "angle": 0, + "content": "[33] Jiarui Xu, Yossi Gandelsman, Amir Bar, Jianwei Yang, Jianfeng Gao, Trevor Darrell, and Xiaolong Wang. Improv: Inpainting-based multimodal prompting for computer vision tasks. arXiv preprint arXiv:2312.01771, 2023. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.728, + 0.482, + 0.769 + ], + "angle": 0, + "content": "[34] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.771, + 0.482, + 0.825 + ], + "angle": 0, + "content": "[35] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. What makes good examples for visual in-context learning? Advances in Neural Information Processing Systems, 36:17773-17794, 2023. 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.825 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23613" + } + ] +] \ No newline at end of file diff --git a/2025/Stable Diffusion Models are Secretly Good at Visual In-Context Learning/94d38526-8561-492c-b6db-5bcfedbe9083_origin.pdf b/2025/Stable Diffusion Models are Secretly Good at Visual In-Context Learning/94d38526-8561-492c-b6db-5bcfedbe9083_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f58149f4f61de61ef929da661dab947231b88ff2 --- /dev/null +++ b/2025/Stable Diffusion Models are Secretly Good at Visual In-Context Learning/94d38526-8561-492c-b6db-5bcfedbe9083_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:968b5dc3ec3fe62eaa1075671e079b3e563e100c43e5553cc89d2c8a3ceb09ef +size 10287815 diff --git a/2025/Stable Diffusion Models are Secretly Good at Visual In-Context Learning/full.md b/2025/Stable Diffusion Models are Secretly Good at Visual In-Context Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..22d4edf3df22cb1b7d20766b1716fa1df3a58cc7 --- /dev/null +++ b/2025/Stable Diffusion Models are Secretly Good at Visual In-Context Learning/full.md @@ -0,0 +1,281 @@ +# Stable Diffusion Models are Secretly Good at Visual In-Context Learning + +Trevine Oorloff $^{1,2*}$ Ali Shafahi $^{1}$ Vishwanath Sindagi $^{1}$ Wele Gedara Chaminda Bandara $^{1}$ Amin Ghiasi $^{1}$ Charan Prakash $^{1}$ Reza Ardekani $^{1}$ Apple $^{2}$ University of Maryland - College Park + +# Abstract + +Large language models (LLM) in natural language processing (NLP) have demonstrated great potential for in-context learning (ICL) — the ability to leverage a few sets of example prompts to adapt to various tasks without having to explicitly update the model weights. ICL has recently been explored for computer vision tasks with promising early outcomes. These approaches involve specialized training and/or additional data that complicate the process and limit its generalizability. In this work, we show that off-the-shelf Stable Diffusion models can be repurposed for visual in-context learning (V-ICL). Specifically, we formulate an in-place attention re-computation within the self-attention layers of the Stable Diffusion architecture that explicitly incorporates context between the query and example prompts. Without any additional fine-tuning, we show that this repurposed Stable Diffusion model is able to adapt to six different tasks: foreground segmentation, single object detection, semantic segmentation, keypoint detection, edge detection, and colorization. For example, the proposed approach improves the mean intersection over union (mIoU) for the foreground segmentation task on Pascal-5i dataset by $8.9\%$ and $3.2\%$ over recent methods such as Visual Prompting and IMProv, respectively. Additionally, we show that the proposed method is able to effectively leverage multiple prompts through ensembling to infer the task better and further improve the performance. + +# 1. Introduction + +In-context learning (ICL) refers to the paradigm where a foundation model leverages exemplar source-target pair(s), known as prompts, to infer the task and perform the inferred task on an input, known as query. ICL is an emergent property of large language models (LLMs) and has been widely explored in the natural language processing (NLP) domain [4, 8, 25, 32]. ICL facilitates a model to adapt to novel or out-of-domain tasks without the need for fine-tuning which + +![](images/6928d690523c365b98323bbd57e7fa4d0b84754a197a8c52a34011fe72cb79b6.jpg) +Figure 1. Existing V-ICL methods require specialized training of foundation models on curated/uncurated data before using them for novel downstream tasks. In contrast, our approach repurposes an off-the-shelf Stable Diffusion model without any additional training or data. + +not only eliminates task-specific training but also reduces dependency on task-specific annotated datasets. Recent research, such as [3, 16, 28-31, 33], have made promising attempts to harness the potential of in-context learning for computer vision tasks. These works can be broadly divided into two categories based on the datasets that are used for training: (1) uncurated datasets (e.g. CVF [3], S2CV [33]) and (2) curated/annotated task-related datasets (e.g. COCO [15], NYUDv2 [23]). + +The first category of approaches such as Visual Prompting [3] and Improv [33] enable foundation models to perform in-context learning by training with an inpainting loss on uncurated datasets (e.g. CVF, S2CV). These datasets consist of images extracted from computer vision papers and consist of grid-like structures of source-target pairs. The combination of this structured input and random masking during training allows the model to implicitly learn the relationships between the source and target images. During inference, when the model is presented with a grid-like canvas of the query image and the example prompt(s) (source-target pairs) along with the location of the output for the query image masked-out (see Fig. 2a), the model is expected to inpaint the prediction. + +The second category of approaches [16, 28-31] enable a foundation model for in-context learning by allowing the model to train on curated datasets (e.g. COCO, NYUDv2) + +![](images/d9003c96a323cc4558ba96ed193a7b6170f24899366c22243c83ef07998cb917.jpg) +(a) + +![](images/85bf4bf65636944260f9bc181941388b88bf311f0f7469e668053ea2944e1b82.jpg) +(b) + +![](images/11528fe0005272a20065ffa39da83d3287a2d8451192160a885406149400ca71.jpg) +(c) +Figure 2. (a) Existing approaches like Visual Prompting [3] train a foundation model with an inpainting objective on uncurated data. At inference, a grid of query and example prompt are input to the model. (b) Visual Prompting [3] struggles to predict accurately even when the prompt is the same as the query. (c) The proposed approach, when presented with the same prompt as the query, is able to fully leverage the prompt to make accurate predictions. + +related to the out-of-domain tasks that the model is expected to adapt to. While these methods do not employ task-specific loss functions, they still rely on task-representative datasets. For example, to enable the Painter [28] model to adapt to open-vocabulary segmentation (e.g. FSS-1000 [14]), the model is trained on a related dataset like the COCO segmentation dataset. In other words, as long as the model has encountered a task during training, it can perform that task on query images from unseen categories. However, this implies that the model needs to access large annotated datasets during training thereby undermining the key advantage of V-ICL. Furthermore, the generalization of such models is limited to tasks related to those seen during training. These methods show promising early outcomes of vision foundation models on unseen tasks, however, achieving similar success as LLMs in NLP domain with off-the-shelf vision foundation models (e.g. Stable Diffusion [20], ViT [7]) remains an area yet to be fully explored. + +Although these existing V-ICL approaches show promising outcomes, they involve the use of additional training steps and/or data from out-of-domain tasks to which the model is being used to adapt. Different from existing approaches, in this work, we focus on an approach that is more in-line with the ICL methods from the NLP community. Specifically, we address the question — Can we use off-the-shelf foundation models for visual in-context learning without any additional training steps or data? To this end, we propose the Stable Diffusion based visual in-context learning (SD-VICL) pipeline where we show that an off-the-shelf Stable Diffusion model can be repurposed, without any fine-tuning, to adapt to several new out-of-domain + +tasks (see Fig. 1). To achieve this, we formulate an in-place attention re-computation within the self-attention layers of the Stable Diffusion architecture that explicitly incorporates the context between the query and example prompts. This formulation is based on the drawbacks (Fig. 2) of existing approaches like Visual Prompting [3] where these models are unable to fully leverage the example prompt. + +Furthermore, we note that existing approaches such as Bar et al. [3] and Xu et al. [33] demonstrate the benefits of using multiple example prompts to boost the prediction abilities of the V-ICL foundation models. However, their approach of ensembling in the image space by creating a composite of multiple examples comes at the cost of reduced the resolution per prompt — limiting the potential of prompt ensembling. To address this issue, inspired by feature ensembling in SegGPT [29], we perform prompt ensembling in the latent space for in-context learning. Note that the feature ensembling in SegGPT involves averaging of query features after each attention layer which results in uniform weighting of all prompts. In contrast, we perform implicitly-weighted prompt ensembling within each attention layer thereby enabling the model to better infer the information from multiple prompts. + +We conduct extensive evaluations to demonstrate the effectiveness of the proposed approach. Specifically, we show that a pre-trained Stable Diffusion model has the ability to adapt to various tasks like foreground segmentation, single object detection, semantic segmentation, keypoint detection, edge detection, and colorization. + +To summarize, the main contributions of this work are as follows, + +- The first training-free method to enable visual in-context learning properties within a foundation model, setting a new direction for visual in-context learning research. +- A novel pipeline that explicitly incorporates the context between the query image and prompts by introducing an in-place attention re-computation within the self-attention layers of an off-the-shelf Stable Diffusion model. +- Extensive evaluations of the proposed approach to demonstrate its ability to generalize to multiple out-of-domain tasks. +- Implicitly-weighted prompt ensembling, which enables effective use of multiple prompts for V-ICL. + +# 2. Method + +# 2.1. Motivation and Preliminaries + +In ICL, a foundation model is expected to infer the task based on a few source-target example prompts and predict the appropriate output for the query image. More importantly, as demonstrated in the NLP community, in-context learning is supposed to be an emergent property of a foundation model where the model does not require addi + +tional training. However, existing visual in-context learning approaches involve training and/or use of additional task-related data. Further, when we investigated V-ICL approaches based on uncurated datasets $^2$ (e.g. Bar et al. [3]) through multiple experiments, we observe the following, (1) the model fails to reconstruct the input prompt, despite it being unmasked in the input, as shown in the top-left image of the foundation model's output in Fig. 2a, (2) the model struggles in the presence of multiple objects in the image as shown in Fig. 2a where it incorrectly predicts the bird feeder as the foreground, and (3) when provided with the query and its groundtruth as the example prompt, the model struggles to make correct predictions despite having access to the groundtruth for the query (see Fig. 2b). These observations indicate that the model is unable to fully leverage the prompts. We attribute this inability of the model to lack of appropriate context interpretation (1) between the source and target of the example prompt which is essential for task inference, and (2) between the query and the source of the example prompt which is essential for accurate predictions. + +To alleviate these limitations, we propose a novel inference-only Stable Diffusion based visual in-context learning pipeline (SD-VICL), that unlike existing approaches, requires no additional training. Stable Diffusion [20] is a latent diffusion model that generates high-quality images by iteratively refining random noise through a denoising process. The denoising process at a given time step $t$ employs a denoising U-Net which comprises of multiple self-attention layers operating at resolutions $16 \times 16$ , $32 \times 32$ , and $64 \times 64$ . At each self-attention layer ( $l$ ), the input features of the intermediate noise latent, $\phi(z_{t})$ , are transformed to Query ( $Q$ ), Key ( $K$ ), and Value ( $V$ ) vectors using linear layers. The $Q$ and $K$ vectors are used to compute the self-attention map using: + +$$ +\alpha^ {(t)} = \operatorname {s o f t m a x} \left(\frac {Q ^ {(t)} \cdot K ^ {(t) ^ {T}}}{\sqrt {d}}\right), \tag {1} +$$ + +which captures the correspondences within the image. The parameter $d$ denotes the feature dimension of the $Q$ vector. The self-attention map $(\alpha)$ is used to update the intermediate feature map using the feature update, $\Delta \phi$ , which is computed as, + +$$ +\Delta \phi^ {(t)} = \alpha^ {(t)} \cdot V ^ {(t)}. \tag {2} +$$ + +This updated set of features, along with the updates from the cross-attention layers are used to compute the intermediate noise predictions. In this paper, we focus only on the self-attention computations. Hence, for simplicity, we write the iterative denoising process as + +$$ +z ^ {(t - 1)} = z ^ {(t)} - \epsilon \left(Q ^ {(t)}, K ^ {(t)}, V ^ {(t)}\right), \tag {3} +$$ + +where the predicted noise, $\epsilon$ , is a function of the $Q$ , $K$ , and $V$ vectors of the self-attention layers. + +# 2.2. Repurposing Stable Diffusion for Visual In-Context Learning (SD-VICL) + +As discussed previously, for successful in-context learning, a foundation model needs to appropriately infer (1) the task: the relationship between the prompt image $(A)$ and prompt groundtruth $(B)$ , and (2) the context: the relationship between the query image $(C)$ and the prompt image $(A)$ . In an attempt to ensure that these relationships are effectively inferred, we formulate a novel attention computation in place of the traditional self-attention computation in the upsample layers of the denoising U-Net. This formulation is inspired by a few recent research [1, 5, 18, 26] that repurpose the self-attention layers of Stable Diffusion for tasks like image editing and style transfer. + +First, each of the input images: prompt image $(A)$ , prompt groundtruth $(B)$ , and query image $(C)$ are inverted to the Stable Diffusion's noise space to obtain, $z_A^{(T)}$ , $z_B^{(T)}$ , and $z_C^{(T)}$ respectively, using an off-the-shelf inversion model [13]. Then, each of these images are iteratively denoised in parallel without any changes to the default denoising pipeline as follows: + +$$ +z _ {p} ^ {(t - 1)} = z _ {p} ^ {(t)} - \epsilon \left(Q _ {p} ^ {(t)}, K _ {p} ^ {(t)}, V _ {p} ^ {(t)}\right), \tag {4} +$$ + +where $p\in \{A,B,C\}$ $t\in [T,1],t\in \mathbb{Z}^{+}$ + +Since the output image is expected to be structurally similar to the query image for most vision tasks, we initialize the noise space in the prediction pipeline $D$ using the noise space of the query image (i.e., $z_{D}^{(T)} \gets z_{C}^{(T)}$ ). + +Different from Eq. (4), the query prediction is denoised using, + +$$ +z _ {D} ^ {(t - 1)} = z _ {D} ^ {(t)} - \epsilon \left(Q _ {C} ^ {(t)}, K _ {A} ^ {(t)}, V _ {B} ^ {(t)}\right), \tag {5} +$$ + +where $t\in [T,1],t\in \mathbb{Z}^{+}$ + +Expanding the feature update formulation for $D$ we get, + +$$ +\Delta \phi_ {D} ^ {(t)} = \alpha_ {D} ^ {(t)} \cdot V _ {B} ^ {(t)}, \tag {6} +$$ + +$$ +\Delta \phi_ {D} ^ {(t)} = \operatorname {s o f t m a x} \left(\frac {Q _ {C} ^ {(t)} \cdot K _ {A} ^ {(t) ^ {T}}}{\tau \cdot \sqrt {d}}\right) \cdot V _ {B} ^ {(t)}. \tag {7} +$$ + +In contrast to Eqs. (1) and (2), the attention map is formulated using the Query vector of the query image $(C)$ and the Key vector of the prompt image $(A)$ . The $Q_{C}$ vector comprises of the semantics of each spatial location in $C$ , and the $K_{A}$ vector offers the context within $A$ that each query can attend to. Hence, each element in the attention map, i.e., $\alpha_{D(i,j)}^{t}$ , captures how the $i^{th}$ patch of the query image correlates with the $j^{th}$ patch of the prompt image. This formulation explicitly enforces the context between the query image and the prompt image. The additional temperature hyperparameter $(\tau)$ (in Eq. (7)) for the softmax computation + +![](images/630f5816d67c371b5e728424dbd09b019ddde6f498acbdbeb9752e51478cee34.jpg) +Figure 3. Latent representations of the query and prompt images from the Stable Diffusion model's VAE encoder are transformed to the noise space using [13]. Each noisy latent is iteratively denoised using the denoising U-Net. Prediction $(D)$ path is initialized with the noise space of $C$ , and at each denoising step, the computations within the self-attention layers are modified to enhance visual ICL capabilities. With the Query (from prompt image- $A$ ), the Key (from prompt groundtruth- $B$ ), and the Value (from query image- $C$ ) vectors, we reformulate the intermediate feature update to infuse context between the example prompts and the query image into the foundation model. This process is iteratively performed for $T$ denoising steps to obtain the denoised prediction latent, which is then fed to the VAE decoder to obtain the final prediction $D$ . + +controls the sharpness of the correlation. By adjusting $\tau$ , we can modulate the model's focus on attention/correlation between $C$ and $A$ . A lower $\tau$ sharpens the attention, emphasizing stronger correlation, while a higher $\tau$ smoothens the attention, allowing for a broader distribution of focus across multiple correlations. Furthermore, we use the Value vector from the prompt groundtruth $(B)$ to compute $\Delta \phi_D^{(t)}$ . This use of value vector from a different latent, in contrast to the standard implementation of cross-attention [27], where both Key and Value vectors typically come from the same set of intermediate latents, is to facilitate the prediction to be in the same domain as that of the prompt groundtruth $B$ . This formulation, which is in-place of the computation within the standard self-attention layer of Stable Diffusion, facilitates the explicit infusion of the context between the query image and the prompt in addition to improved task inference. However, the modification of the default self-attention computation and the use of $Q$ , $K$ , $V$ vectors that are extracted from three distinct images create a domain gap that needs to be addressed to enhance the quality of the prediction. Hence, we adapt the attention map contrasting, swap-guidance, and AdaIN mechanisms employed in [1]. + +Attention map contrasting supplements the function of the temperature hyperparameter $(\tau)$ , where it enhances the focus on the relevant regions (i.e. semantically similar regions) between $C$ and $A$ , while attenuating the irrelevant (i.e. color similarity between unrelated objects). Additionally, this operation adjusts the scale of the values to suit the pre-trained Stable Diffusion pipeline. This contrasting op + +eration can be expressed as, + +$$ +\alpha_ {D} ^ {(t)} = \mu \left(\alpha_ {D} ^ {(t)}\right) + \beta \cdot \left(\alpha_ {D} ^ {(t)} - \mu \left(\alpha_ {D} ^ {(t)}\right)\right), \tag {8} +$$ + +where $\mu$ is the mean operation and $\beta$ is a hyperparameter that controls the scale. + +Swap-guidance is derived from classifier-free diffusion guidance implementation proposed by Ho and Salimans [11]. Given the noise prediction from the default self-attention formulation $(\eta_{default}^{(t)})$ and the noise prediction from our modified formulation $(\eta_{modified}^{(t)})$ , the final noise prediction $(\eta^{(t)})$ can be expressed as, + +$$ +\eta^ {(t)} = \eta_ {\text {d e f a u l t}} ^ {(t)} + \frac {\gamma \cdot (T - t)}{T} \cdot \left(\eta_ {\text {m o d i f i e d}} ^ {(t)} - \eta_ {\text {d e f a u l t}} ^ {(t)}\right), \tag {9} +$$ + +where $\gamma$ is a hyperparameter controlling the scale. This mechanism gradually incorporates the modified noise prediction as denoising progresses, which directs the denoising process through denser regions of the Stable Diffusion's generative pipeline, thus alleviating unwarranted artifacts. + +Adaptive instance normalization (AdaIN), which was proposed by Huang and Belongie [12] is used to align the color distribution between the prediction $(D)$ , which is initialized using the noise space of the query image $(C)$ , and the final groundtruth/task color space (i.e. color space of $B$ ). AdaIN is implemented as follows, + +$$ +z _ {D} ^ {(t)} \leftarrow \operatorname {A d a I N} \left(z _ {D} ^ {(t)}, z _ {B} ^ {(t)}\right), \tag {10} +$$ + +$$ +z _ {D} ^ {(t)} \leftarrow \frac {z _ {D} ^ {(t)} - \mu \left(z _ {D} ^ {(t)}\right)}{\sigma \left(z _ {D} ^ {(t)}\right)} \cdot \sigma \left(z _ {B} ^ {(t)}\right) + \mu \left(z _ {B} ^ {(t)}\right), \tag {11} +$$ + +where $\mu$ and $\sigma$ are mean and standard deviation operations. + +# 2.3. Implicitly-Weighted Prompt Ensembling (IWPE) + +Bar et al. [3] and Xu et al. [33] demonstrate that providing the model with multiple source-target example prompts improves the prediction performance of visual in-context learning. These methods ensemble multiple prompts at the input level by stitching together prompt images and corresponding groundtruths to form a grid-like structure in the form of a composite image. This composite image is then input to the model, which is expected to leverage information from multiple prompts to make a better prediction compared to the single prompt scenario. However, for a fixed-resolution foundation model, as the number of source-target examples increases, the effective size of each image in the grid is reduced. This leads to performance deterioration, mainly because of the loss of details in the input [33]. + +To address this issue, instead of ensembling prompts in the image space through a composite input, SegGPT [29] proposed a feature space ensembling method. In this method, multiple prompts are processed in parallel and aggregated by averaging the features at the end of each attention layer. This approach assumes that all the example prompts are equally informative and they are uniformly weighted during the averaging process. However, such uniform weighting of prompts is sub-optimal since it prevents the model from benefiting from more relevant prompts. To this end, we propose a simple but effective ensembling method by integrating the ensembling into the attention computation, allowing the prompt patches to be implicitly weighted based on their correspondences with each query patch. Specifically, we concatenate each of the $K$ and $V$ vectors from each of the prompts prior to computing $\Delta \phi_D^{(t)}$ . Specifically, for $n$ number of prompts, we re-write Eq. (7): + +$$ +\Delta \phi_ {D} ^ {(t)} = \operatorname {s o f t m a x} \left(\frac {Q _ {C} ^ {(t)} \cdot \left(\bigoplus_ {i = 1} ^ {n} K _ {A _ {i}} ^ {(t)}\right) ^ {T}}{\tau \cdot \sqrt {d}}\right) \cdot \left(\bigoplus_ {i = 1} ^ {n} V _ {B _ {i}} ^ {(t)}\right), \tag {12} +$$ + +where $\oplus$ is the concatenation operation and $i$ is the index for each prompt. + +# 3. Experiments and Evaluations + +To demonstrate the ability of the proposed approach to generalize across different tasks, we evaluate it on six downstream tasks. For all these tasks, we use the unsupervised prompt retrieval [35] to select the candidates for the prompt images. Specifically, this method chooses the nearest neighbors of the query image as prompt candidates. The nearest neighbor retrieval is based on the cosine similarity of CLIP's vision encoder [19] embeddings between the query image and the example prompt images. + +For a fair comparison, we use Visual Prompting [3] and IMProv [33] as baselines. Although these approaches involve a training step, unlike [28, 31], they do not rely on + +curated or annotated data that come from related out-of-domain tasks. Additionally, since IMProv supports supplementary text guidance, we include results with and without text guidance in our comparisons. Further, we also evaluate and compare the effectiveness of multiple prompts. Please refer to the supplementary for details on datasets, evaluation metrics, and sensitivity analysis for different hyperparameters. For the results in the following subsections, we choose one set of hyperparameters that provided optimal performance for all the tasks. + +# 3.1. Results and Analysis + +Quantitative metrics for all the six tasks are reported in Tab. 1. The proposed approach consistently outperforms existing uncurated based approaches in all the tasks. + +Foreground segmentation: With a single prompt, the proposed method achieves absolute improvements of $8.9\%$ and $3.2\%$ in mean intersection over union (mIoU) on an average across all splits3 in the Pascal-5i dataset [22], compared to Visual Prompting and IMProv, respectively. The proposed ensembling method with five prompts using the repurposed Stable Diffusion model further improves the mIoU by $11.6\%$ (in absolute terms) as compared to the single prompt case. + +Single object detection: Similar to foreground segmentation, with a single prompt, the proposed method improves the mIoU by $5.3\%$ and $7.1\%$ (in absolute terms) on average across all splits3 in the Pascal-5i dataset [22] as compared to Visual Prompting and IMProv, respectively. The use of multiple prompts through implicitly-weighted prompt assembling results in an additional $5.6\%$ of improvement as compared to the single prompt case. + +Semantic segmentation: On the Cityscapes dataset [6], our method with a single prompt results in an absolute improvement of $0.8\%$ in mIoU and $3.4\%$ increase in accuracy as compared to Visual Prompting. Similarly, we observe an absolute gain of $6.1\%$ in mIoU and a $6.5\%$ in accuracy as compared to IMProv. With IWPE, the mIoU and accuracy further improve by $4.9\%$ and $6.8\%$ (in absolute terms) as compared to the single prompt based approach. The proposed approach with a single prompt in the case of semantic segmentation achieves lower performance improvements over existing methods as compared to other tasks like foreground segmentation and keypoint detection. However, with additional prompts, the proposed method is able to achieve much higher gains similar to what is observed in other tasks. This performance boost can be attributed to two key factors: (1) multiple prompts help resolve conflicts between similar or overlapping classes by mitigating inter-class attention ambiguities, and (2) a single prompt is unlikely to contain all classes that are potentially present in the query image of a multi-class task and in such cases, ac + +![](images/190352a69e33d040911ffed063d4f2259ca8211c588d9edcef0c0533b5d0cdfe.jpg) +Figure 4. Qualitative evaluation, where we compare the performance of Visual Prompting [3] and IMProv [33] with our proposed method on six different tasks. It can be seen that our method produces visually superior results as compared to the baselines. + +cess to multiple prompts increases the probability that all relevant classes are available to the model for making accurate predictions. + +Keypoint detection: When evaluated on the DeepFashion dataset [17], our method demonstrates a substantial improvement, reducing the MSE by $6 \times$ while achieving a $7 \times$ gain in percentage of correct keypoints (PCK), compared to the best-performing baseline. Furthermore, the use of five prompts enhances these metrics, yielding additional relative gains of $18.5\%$ for MSE and $6.5\%$ for PCK. + +Edge detection: On the NYUDv2 dataset [23], our approach reduces the mean squared error (MSE) by $72.0\%$ and LPIPS [34] by $60.6\%$ as compared to the best performing existing approach. Note that these metrics improve an additional $24.5\%$ and $21.4\%$ , respectively, with our prompt ensembling method. + +Colorization: Evaluating on the ImageNet dataset [21], our model reduces LPIPS by $27.9\%$ and FID [10] by $49.0\%$ , relative to IMProv, which yields the next best performance, and shows further gains of $19.1\%$ and $16.0\%$ respectively, with five prompts. + +Overall, our method outperforms both Visual Prompting and IMProv across all tasks by considerable margins. Additionally, the integration of multiple prompts using IWPE consistently yields performance improvements over + +the single-prompt scenario. + +Supplementing the quantitative results, we also demonstrate visual comparisons of the outputs for each task in Fig. 4 $^{4}$ , further highlighting the superior performance of our method over baselines. We observe that while both Visual Prompting and IMProv perform reasonably in single object detection and foreground segmentation, they show a weaker performance in tasks such as colorization, edge detection, and semantic segmentation, where capturing fine details and preserving structural information of the query is crucial. Moreover, both baselines notably struggle with keypoint detection, whereas the proposed approach demonstrates far superior results, indicating its robustness to different out-of-domain tasks. + +# 3.2. Additional Evaluations + +Comparison with V-ICL models trained on task-related data: To provide a comprehensive evaluation, we compare our training-free approach against prior methods trained explicitly on task-related data, including Painter [28], LVM [2], and Prompt Diffusion [31]. Quantitative results are presented in Tab. 2, and qualitative comparisons are shown in Fig. 7 (supplementary). Despite the lack of any ex + +
ModelForeground Seg. (Pascal-5i) mIoU ↑Object Det. (Pascal-5i) mIoU ↑Semantic Segmentation (Cityscapes) mIoU ↑Keypoint Detection (DeepFashion) Acc. ↑MSE ↓PCK ↑Edge Detection (NYUDv2) MSE ↓LPIPS ↓Colorization (ImageNet) LPIPS ↓FID ↓
Number of Example Prompts: 1
Visual Prompting [3]35.0446.2921.3071.5235.8310.870.10060.39250.4166111.06
IMProv (w/o text) [33]39.5043.8215.9768.4431.3018.380.10590.42780.3895104.74
IMProv (w/ text) [33]40.6944.4215.8067.8844.672.810.11250.51460.3899105.90
SD-VICL (ours)43.9251.5522.0874.935.3677.190.02820.15480.281053.40
Number of Example Prompts: 5
Visual Prompting36.3748.0921.7070.9536.5710.320.10000.39000.4156112.91
SD-VICL (ours)55.4957.1027.0181.754.3782.240.02130.12160.227244.84
+ +Table 1. Quantitative performance comparison of the proposed approach with recent approaches on foreground segmentation, single object detection, semantic segmentation, keypoint detection, edge detection, and colorization. + +
ModelFG Seg. mIoU ↑Obj. Det. mIoU ↑Edge Det.Colorization
MSE ↓LPIPS ↓LPIPS ↓FID ↓
Painter [28]55.0954.280.09260.72940.347464.16
LVM [2]50.9852.670.04990.42590.314256.40
Prompt Diffusion [31]17.5326.860.04300.33110.699798.91
SD-VICL (ours)55.4957.100.02130.12160.227244.84
+ +plicit training, our method consistently outperforms all three baselines across multiple vision tasks. Analyzing the results, we observe that all three models tend to overfit to their training tasks, leading to poor generalization when presented with novel tasks. For instance, Painter incorrectly outputs a depth map instead of edges, despite the prompt pair defining the task has an edge map as the target (Fig. 7 in supplementary). Further, while Painter and LVM perform well on tasks, foreground segmentation and object detection, when the query image consists of only a single class, they struggle in multi-class scenarios, often segmenting all classes instead of the intended one (Fig. 8a in supplementary). Additionally, LVM exhibits inconsistencies in its outputs, where for a given task, despite the format/domain of the inputs remaining unchanged, we observe that the generated output belongs to diverse domains (Fig. 8b in supplementary). Moreover, Prompt Diffusion performs notably poorly on most tasks, often generating incorrect outputs, with the exception of edge detection — a task it was exposed to during training. Ideally, a V-ICL model should infer context and task from the input, but all three models fall short in this aspect. In contrast, our training-free method demonstrates superior generalization and effective context and task inference as intended by V-ICL, underscoring the advantages of uncovering V-ICL properties without additional training. For a detailed discussion and visual examples, we refer the reader to Sec. E in the supplementary. + +Comparison with task-specific models based on Stable Diffusion: Here, we compare our method against approaches that adapt Stable Diffusion to facilitate specific tasks, such as semantic segmentation and keypoint detection. DAS [24] is one such work that proposes a pipeline + +Table 2. Extended quantitative evaluations against V-ICL models that train on task-related data. + +
ModelSem.Seg.Keypoint Det.
mIoU ↑Acc. ↑MSE ↓PCK ↑
DAS [24]21.276.0--
StableKeypoints [9]--6.4670.0
SD-VICL (ours)31.5282.724.3782.24
+ +Table 3. Comparison with task-specific models based on Stable Diffusion. + +that aggregates the self-attention layers of Stable Diffusion iteratively to predict segmentation masks for each class facilitating semantic segmentation. Similarly, StableKeypoints [9] optimizes on the text embeddings such that cross-attention maps condense to landmarks in a given image facilitating keypoint detection. We present the performance comparisons against these task-specific models in Tab. 3, where our approach outperforms both [24] and [9]. For semantic segmentation, we follow the evaluation protocol of DAS (i.e. Hungarian matching) to ensure a fair comparison and achieve absolute improvements of $10.3\%$ in mIoU and $6.7\%$ in accuracy. In keypoint detection, we surpass StableKeypoints with a $32.4\%$ reduction in MSE and $12.2\%$ improvement in PCK. While DAS and StableKeypoints achieve strong results in their respective tasks, they remain constrained by their task-specific nature. In contrast, our method not only delivers superior performance but also generalizes across diverse tasks without requiring any modification in the inference. Furthermore, unlike these methods, which derive predictions from intermediate feature or attention maps, our model directly utilizes the generative process to produce outputs, making it inherently more flexible and effective. + +# 3.3. Ablation Study + +In this section, we present the details of various ablations conducted to demonstrate the significance of key design choices. Specifically, we evaluate the effectiveness of implicitly-weighted prompt ensembling and examine the impact of number of prompts on performance. For additional ablations, including alternative attention formulations and sensitivity analyses on hyperparameters such as + +
# PromptsEnsemblingmIoU ↑Acc. ↑
1N/A22.0874.93
5FE22.9778.84
5IWPE27.0181.75
+ +Table 4. Ablation of prompt-ensembling evaluated on semantic segmentation. + +attention temperature, self-attention resolution, contrastive strength, swap-guidance scale, and AdaIN, we refer the reader to Sec. F in the supplementary. + +Effectiveness of implicitly-weighted prompt ensembling (IWPE): Here, we compare our proposed ensembling method, IWPE, to SegGPT's feature ensembling (FE) approach. For all tasks except semantic segmentation, we observe that both methods yield similar gains over the single prompt scenario. In the case of semantic segmentation, IWPE demonstrates significantly better performance. As tabulated in Tab. 4, while FE yields an absolute improvement of merely $0.9\%$ in mIoU and $3.9\%$ in accuracy, IWPE achieves substantial gains, improving mIoU by $4.9\%$ and accuracy by $6.8\%$ over the single prompt case. This notable improvement is primarily due to the capability of our ensembling method to dynamically weigh each prompt based on its correspondence with the query image. In contrast, as discussed previously, the feature ensembling approach assumes that all prompts are equally informative, which is sub-optimal, especially in the case of tasks involving multiple classes like semantic segmentation. For instance, one prompt may provide a stronger correlation with a specific class, while another prompt, would have a weaker association with that class. Our method by implicitly weighing the influence of different prompts based on their relative correspondence to the query, better captures these nuances, resulting in a significant boost in performance compared to SegGPT's uniformly weighted feature ensembling. Therefore, while both FE and IWPE perform at par in simpler tasks where the prompts are equally informative, our method is more beneficial for tasks that involve complex multi-class structures. + +Impact of number of prompts: Here, we investigate the effect of the number of prompts on the performance and the inference speed. As illustrated in Fig. 5a, the mIoU scores improve with an increasing number of prompts. This result can be attributed to the model's ability to effectively leverage multiple prompts to better infer the task and establish the correspondence between the query and the prompts. The additional prompts help reduce ambiguities that may arise from using a single prompt. However, these gains come at the cost of reduced inference speed; specifically using five prompts cuts the inference speed to approximately half that of a single prompt. Besides the prompts, the number of denoising steps in the Stable Diffusion model is a key factor impacting the inference speed. This led us to the question + +![](images/95a3b0d67bc9aea035eea773115a6a9bf36187125884c0e058a51fd27b512476.jpg) +(a) + +![](images/9ae8dd8c3a21a68e05c6107bf3ecaf5e8e4601bb5f2d51018995c631547dbf85.jpg) +(b) +Figure 5. Effect of the number of prompts on the performance and inference speed5. While (a) depicts the variation in mIoU and the inference speed, with the number of prompts, (b) illustrates the variation in mIoU with the number of denoising steps for both single and five prompt cases. + +— With the availability of multiple prompts, does the model require as many denoising steps as in the single prompt case? In other words, we explored if we can trade-off a few denoising steps without impacting the performance when additional context is provided by multiple prompts. To this end, we conducted experiments where we evaluated the SD-VICL approach with different denoising steps. As shown in Fig. 5b, the mIoU for the five-prompt scenario at 30 denoising steps surpasses the maximum performance of the single-prompt scenario. Moreover, the maximum performance for the single-prompt case, achieved with 70 denoising steps, requires 38.4 seconds of inference time. In contrast, the five-prompt case achieves higher performance in just 35.9 seconds with 30 denoising steps5. + +# 4. Conclusion + +In this paper, we propose a novel pipeline for visual in-context learning that explicitly incorporates the context between the query image and the prompt. Unlike existing visual in-context learning methods which rely on training or fine-tuning foundation models on curated or uncurated data to enable in-context learning, our approach is entirely training-free. Specifically, we introduce an in-place attention re-computation within the self-attention layers of an off-the-shelf Stable Diffusion model. Additionally, the proposed implicitly-weighted prompt ensembling technique facilitates effective integration of context through multiple prompts by implicitly weighing the prompts based on the relative correspondences. The versatility of the proposed method is demonstrated by its successful generalization across six different tasks, where it outperforms both V-ICL models trained on uncurated and task-related data as well as task-specific models based on Stable Diffusion. Crucially, our method achieves these results without requiring any additional training or annotated task-specific datasets, highlighting its ability to uncover and leverage the V-ICL capabilities of Stable Diffusion. + +# References + +[1] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer. In ACM SIGGRAPH 2024 Conference Papers, pages 1–12, 2024. 3, 4 +[2] Yutong Bai, Xinyang Geng, Karttikeya Mangalam, Amir Bar, Alan L Yuille, Trevor Darrell, Jitendra Malik, and Alexei A Efros. Sequential modeling enables scalable learning for large vision models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22861-22872, 2024. 6, 7 +[3] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei Efros. Visual prompting via image inpainting. Advances in Neural Information Processing Systems, 35:25005-25017, 2022. 1, 2, 3, 5, 6, 7 +[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1 +[5] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560-22570, 2023. 3 +[6] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5 +[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2 +[8] Yaru Hao, Haoyu Song, Li Dong, Shaohan Huang, Zewen Chi, Wenhui Wang, Shuming Ma, and Furu Wei. Language models are general-purpose interfaces. arXiv preprint arXiv:2206.06336, 2022. 1 +[9] Eric Hedlin, Gopal Sharma, Shweta Mahajan, Xingzhe He, Hossam Isack, Abhishek Kar, Helge Rhodin, Andrea Tagliasacchi, and Kwang Moo Yi. Unsupervised keypoints from pretrained diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22820-22830, 2024. 7 +[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4 +[12] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceed- + +ings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 4 +[13] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12469-12478, 2024. 3, 4 +[14] Xiang Li, Tianhan Wei, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. Fss-1000: A 1000-class dataset for few-shot segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2869–2878, 2020. 2 +[15] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 1 +[16] Weihuang Liu, Xi Shen, Chi-Man Pun, and Xiaodong Cun. Explicit visual prompting for low-level structure segmentations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19434-19445, 2023. 1 +[17] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1096-1104, 2016. 6 +[18] Or Patashnik, Daniel Garibi, Idan Azuri, Hadar Averbuch-Elor, and Daniel Cohen-Or. Localizing object-level shape variations with text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23051–23061, 2023. 3 +[19] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5 +[20] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3 +[21] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 6 +[22] Amirreza Shaban, Shray Bansal, Zhen Liu, Irfan Essa, and Byron Boots. One-shot learning for semantic segmentation. arXiv preprint arXiv:1709.03410, 2017. 5 +[23] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 1, 6 + +[24] Junjiao Tian, Lavisha Aggarwal, Andrea Colaco, Zsolt Kira, and Mar Gonzalez-Franco. Diffuse attend and segment: Unsupervised zero-shot segmentation using stable diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3554-3563, 2024. 7 +[25] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1 +[26] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 3 +[27] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 4 +[28] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6830-6839, 2023. 1, 2, 5, 6, 7 +[29] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023. 2,5 +[30] Xinshun Wang, Zhongbin Fang, Xia Li, Xiangtai Li, Chen Chen, and Mengyuan Liu. Skeleton-in-context: Unified skeleton sequence modeling with in-context learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2436-2446, 2024. +[31] Zhendong Wang, Yifan Jiang, Yadong Lu, Pengcheng He, Weizhu Chen, Zhangyang Wang, Mingyuan Zhou, et al. Incontext learning unlocked for diffusion models. Advances in Neural Information Processing Systems, 36:8542-8562, 2023. 1, 5, 6, 7 +[32] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. Emergent abilities of large language models. Transactions on Machine Learning Research, 2022. 1 +[33] Jiarui Xu, Yossi Gandelsman, Amir Bar, Jianwei Yang, Jianfeng Gao, Trevor Darrell, and Xiaolong Wang. Improv: Inpainting-based multimodal prompting for computer vision tasks. arXiv preprint arXiv:2312.01771, 2023. 1, 2, 5, 6, 7 +[34] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6 +[35] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. What makes good examples for visual in-context learning? Advances in Neural Information Processing Systems, 36:17773-17794, 2023. 5 \ No newline at end of file diff --git a/2025/Stable Diffusion Models are Secretly Good at Visual In-Context Learning/images.zip b/2025/Stable Diffusion Models are Secretly Good at Visual In-Context Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3c3c1a46348c8488fe3bc5a06def78be36d666ba --- /dev/null +++ b/2025/Stable Diffusion Models are Secretly Good at Visual In-Context Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e80c42fcff7ba37d07e9dc6b591b6aa93f22c6dee4112ddb7a1e2b893ed8bd13 +size 478923 diff --git a/2025/Stable Diffusion Models are Secretly Good at Visual In-Context Learning/layout.json b/2025/Stable Diffusion Models are Secretly Good at Visual In-Context Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fbd8ab7ea21188eb1343d41ad136e65822a790e1 --- /dev/null +++ b/2025/Stable Diffusion Models are Secretly Good at Visual In-Context Learning/layout.json @@ -0,0 +1,8452 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 79, + 103, + 531, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 103, + 531, + 121 + ], + "spans": [ + { + "bbox": [ + 79, + 103, + 531, + 121 + ], + "type": "text", + "content": "Stable Diffusion Models are Secretly Good at Visual In-Context Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 141, + 505, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 141, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 104, + 141, + 505, + 196 + ], + "type": "text", + "content": "Trevine Oorloff" + }, + { + "bbox": [ + 104, + 141, + 505, + 196 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 104, + 141, + 505, + 196 + ], + "type": "text", + "content": " Ali Shafahi" + }, + { + "bbox": [ + 104, + 141, + 505, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 104, + 141, + 505, + 196 + ], + "type": "text", + "content": " Vishwanath Sindagi" + }, + { + "bbox": [ + 104, + 141, + 505, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 104, + 141, + 505, + 196 + ], + "type": "text", + "content": " Wele Gedara Chaminda Bandara" + }, + { + "bbox": [ + 104, + 141, + 505, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 104, + 141, + 505, + 196 + ], + "type": "text", + "content": " Amin Ghiasi" + }, + { + "bbox": [ + 104, + 141, + 505, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 104, + 141, + 505, + 196 + ], + "type": "text", + "content": " Charan Prakash" + }, + { + "bbox": [ + 104, + 141, + 505, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 104, + 141, + 505, + 196 + ], + "type": "text", + "content": " Reza Ardekani" + }, + { + "bbox": [ + 104, + 141, + 505, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 104, + 141, + 505, + 196 + ], + "type": "text", + "content": " Apple " + }, + { + "bbox": [ + 104, + 141, + 505, + 196 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 104, + 141, + 505, + 196 + ], + "type": "text", + "content": "University of Maryland - College Park" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 221, + 200, + 234 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 221, + 200, + 234 + ], + "spans": [ + { + "bbox": [ + 151, + 221, + 200, + 234 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 247, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 247, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 54, + 247, + 296, + 558 + ], + "type": "text", + "content": "Large language models (LLM) in natural language processing (NLP) have demonstrated great potential for in-context learning (ICL) — the ability to leverage a few sets of example prompts to adapt to various tasks without having to explicitly update the model weights. ICL has recently been explored for computer vision tasks with promising early outcomes. These approaches involve specialized training and/or additional data that complicate the process and limit its generalizability. In this work, we show that off-the-shelf Stable Diffusion models can be repurposed for visual in-context learning (V-ICL). Specifically, we formulate an in-place attention re-computation within the self-attention layers of the Stable Diffusion architecture that explicitly incorporates context between the query and example prompts. Without any additional fine-tuning, we show that this repurposed Stable Diffusion model is able to adapt to six different tasks: foreground segmentation, single object detection, semantic segmentation, keypoint detection, edge detection, and colorization. For example, the proposed approach improves the mean intersection over union (mIoU) for the foreground segmentation task on Pascal-5i dataset by " + }, + { + "bbox": [ + 54, + 247, + 296, + 558 + ], + "type": "inline_equation", + "content": "8.9\\%" + }, + { + "bbox": [ + 54, + 247, + 296, + 558 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 247, + 296, + 558 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 54, + 247, + 296, + 558 + ], + "type": "text", + "content": " over recent methods such as Visual Prompting and IMProv, respectively. Additionally, we show that the proposed method is able to effectively leverage multiple prompts through ensembling to infer the task better and further improve the performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 580, + 135, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 135, + 593 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 135, + 593 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 600, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 295, + 696 + ], + "type": "text", + "content": "In-context learning (ICL) refers to the paradigm where a foundation model leverages exemplar source-target pair(s), known as prompts, to infer the task and perform the inferred task on an input, known as query. ICL is an emergent property of large language models (LLMs) and has been widely explored in the natural language processing (NLP) domain [4, 8, 25, 32]. ICL facilitates a model to adapt to novel or out-of-domain tasks without the need for fine-tuning which" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 322, + 220, + 548, + 323 + ], + "blocks": [ + { + "bbox": [ + 322, + 220, + 548, + 323 + ], + "lines": [ + { + "bbox": [ + 322, + 220, + 548, + 323 + ], + "spans": [ + { + "bbox": [ + 322, + 220, + 548, + 323 + ], + "type": "image", + "image_path": "6928d690523c365b98323bbd57e7fa4d0b84754a197a8c52a34011fe72cb79b6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 331, + 555, + 388 + ], + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 388 + ], + "type": "text", + "content": "Figure 1. Existing V-ICL methods require specialized training of foundation models on curated/uncurated data before using them for novel downstream tasks. In contrast, our approach repurposes an off-the-shelf Stable Diffusion model without any additional training or data." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 398, + 555, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 555, + 506 + ], + "type": "text", + "content": "not only eliminates task-specific training but also reduces dependency on task-specific annotated datasets. Recent research, such as [3, 16, 28-31, 33], have made promising attempts to harness the potential of in-context learning for computer vision tasks. These works can be broadly divided into two categories based on the datasets that are used for training: (1) uncurated datasets (e.g. CVF [3], S2CV [33]) and (2) curated/annotated task-related datasets (e.g. COCO [15], NYUDv2 [23])." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 508, + 556, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 556, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 556, + 674 + ], + "type": "text", + "content": "The first category of approaches such as Visual Prompting [3] and Improv [33] enable foundation models to perform in-context learning by training with an inpainting loss on uncurated datasets (e.g. CVF, S2CV). These datasets consist of images extracted from computer vision papers and consist of grid-like structures of source-target pairs. The combination of this structured input and random masking during training allows the model to implicitly learn the relationships between the source and target images. During inference, when the model is presented with a grid-like canvas of the query image and the example prompt(s) (source-target pairs) along with the location of the output for the query image masked-out (see Fig. 2a), the model is expected to inpaint the prediction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "The second category of approaches [16, 28-31] enable a foundation model for in-context learning by allowing the model to train on curated datasets (e.g. COCO, NYUDv2)" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 241, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 241, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 241, + 713 + ], + "type": "text", + "content": "*This work was completed during internship at Apple" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23604" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 92, + 69, + 261, + 132 + ], + "blocks": [ + { + "bbox": [ + 92, + 69, + 261, + 132 + ], + "lines": [ + { + "bbox": [ + 92, + 69, + 261, + 132 + ], + "spans": [ + { + "bbox": [ + 92, + 69, + 261, + 132 + ], + "type": "image", + "image_path": "d9003c96a323cc4558ba96ed193a7b6170f24899366c22243c83ef07998cb917.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 134, + 181, + 143 + ], + "lines": [ + { + "bbox": [ + 170, + 134, + 181, + 143 + ], + "spans": [ + { + "bbox": [ + 170, + 134, + 181, + 143 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 91, + 144, + 161, + 212 + ], + "blocks": [ + { + "bbox": [ + 91, + 144, + 161, + 212 + ], + "lines": [ + { + "bbox": [ + 91, + 144, + 161, + 212 + ], + "spans": [ + { + "bbox": [ + 91, + 144, + 161, + 212 + ], + "type": "image", + "image_path": "85bf4bf65636944260f9bc181941388b88bf311f0f7469e668053ea2944e1b82.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 121, + 213, + 130, + 220 + ], + "lines": [ + { + "bbox": [ + 121, + 213, + 130, + 220 + ], + "spans": [ + { + "bbox": [ + 121, + 213, + 130, + 220 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 191, + 144, + 261, + 212 + ], + "blocks": [ + { + "bbox": [ + 191, + 144, + 261, + 212 + ], + "lines": [ + { + "bbox": [ + 191, + 144, + 261, + 212 + ], + "spans": [ + { + "bbox": [ + 191, + 144, + 261, + 212 + ], + "type": "image", + "image_path": "11528fe0005272a20065ffa39da83d3287a2d8451192160a885406149400ca71.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 221, + 213, + 230, + 220 + ], + "lines": [ + { + "bbox": [ + 221, + 213, + 230, + 220 + ], + "spans": [ + { + "bbox": [ + 221, + 213, + 230, + 220 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 229, + 296, + 307 + ], + "lines": [ + { + "bbox": [ + 55, + 229, + 296, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 296, + 307 + ], + "type": "text", + "content": "Figure 2. (a) Existing approaches like Visual Prompting [3] train a foundation model with an inpainting objective on uncurated data. At inference, a grid of query and example prompt are input to the model. (b) Visual Prompting [3] struggles to predict accurately even when the prompt is the same as the query. (c) The proposed approach, when presented with the same prompt as the query, is able to fully leverage the prompt to make accurate predictions." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 318, + 296, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 318, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 56, + 318, + 296, + 534 + ], + "type": "text", + "content": "related to the out-of-domain tasks that the model is expected to adapt to. While these methods do not employ task-specific loss functions, they still rely on task-representative datasets. For example, to enable the Painter [28] model to adapt to open-vocabulary segmentation (e.g. FSS-1000 [14]), the model is trained on a related dataset like the COCO segmentation dataset. In other words, as long as the model has encountered a task during training, it can perform that task on query images from unseen categories. However, this implies that the model needs to access large annotated datasets during training thereby undermining the key advantage of V-ICL. Furthermore, the generalization of such models is limited to tasks related to those seen during training. These methods show promising early outcomes of vision foundation models on unseen tasks, however, achieving similar success as LLMs in NLP domain with off-the-shelf vision foundation models (e.g. Stable Diffusion [20], ViT [7]) remains an area yet to be fully explored." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 535, + 296, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 535, + 296, + 691 + ], + "spans": [ + { + "bbox": [ + 55, + 535, + 296, + 691 + ], + "type": "text", + "content": "Although these existing V-ICL approaches show promising outcomes, they involve the use of additional training steps and/or data from out-of-domain tasks to which the model is being used to adapt. Different from existing approaches, in this work, we focus on an approach that is more in-line with the ICL methods from the NLP community. Specifically, we address the question — Can we use off-the-shelf foundation models for visual in-context learning without any additional training steps or data? To this end, we propose the Stable Diffusion based visual in-context learning (SD-VICL) pipeline where we show that an off-the-shelf Stable Diffusion model can be repurposed, without any fine-tuning, to adapt to several new out-of-domain" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "tasks (see Fig. 1). To achieve this, we formulate an in-place attention re-computation within the self-attention layers of the Stable Diffusion architecture that explicitly incorporates the context between the query and example prompts. This formulation is based on the drawbacks (Fig. 2) of existing approaches like Visual Prompting [3] where these models are unable to fully leverage the example prompt." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 157, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 555, + 348 + ], + "type": "text", + "content": "Furthermore, we note that existing approaches such as Bar et al. [3] and Xu et al. [33] demonstrate the benefits of using multiple example prompts to boost the prediction abilities of the V-ICL foundation models. However, their approach of ensembling in the image space by creating a composite of multiple examples comes at the cost of reduced the resolution per prompt — limiting the potential of prompt ensembling. To address this issue, inspired by feature ensembling in SegGPT [29], we perform prompt ensembling in the latent space for in-context learning. Note that the feature ensembling in SegGPT involves averaging of query features after each attention layer which results in uniform weighting of all prompts. In contrast, we perform implicitly-weighted prompt ensembling within each attention layer thereby enabling the model to better infer the information from multiple prompts." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 349, + 554, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 554, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 554, + 420 + ], + "type": "text", + "content": "We conduct extensive evaluations to demonstrate the effectiveness of the proposed approach. Specifically, we show that a pre-trained Stable Diffusion model has the ability to adapt to various tasks like foreground segmentation, single object detection, semantic segmentation, keypoint detection, edge detection, and colorization." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 422, + 554, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 554, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 554, + 445 + ], + "type": "text", + "content": "To summarize, the main contributions of this work are as follows," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 446, + 554, + 590 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 314, + 446, + 554, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 446, + 554, + 482 + ], + "spans": [ + { + "bbox": [ + 314, + 446, + 554, + 482 + ], + "type": "text", + "content": "- The first training-free method to enable visual in-context learning properties within a foundation model, setting a new direction for visual in-context learning research." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 483, + 554, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 483, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 315, + 483, + 554, + 529 + ], + "type": "text", + "content": "- A novel pipeline that explicitly incorporates the context between the query image and prompts by introducing an in-place attention re-computation within the self-attention layers of an off-the-shelf Stable Diffusion model." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 530, + 553, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 530, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 315, + 530, + 553, + 564 + ], + "type": "text", + "content": "- Extensive evaluations of the proposed approach to demonstrate its ability to generalize to multiple out-of-domain tasks." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 566, + 553, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 566, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 314, + 566, + 553, + 590 + ], + "type": "text", + "content": "- Implicitly-weighted prompt ensembling, which enables effective use of multiple prompts for V-ICL." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 603, + 370, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 603, + 370, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 603, + 370, + 615 + ], + "type": "text", + "content": "2. Method" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 624, + 474, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 474, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 474, + 635 + ], + "type": "text", + "content": "2.1. Motivation and Preliminaries" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "content": "In ICL, a foundation model is expected to infer the task based on a few source-target example prompts and predict the appropriate output for the query image. More importantly, as demonstrated in the NLP community, in-context learning is supposed to be an emergent property of a foundation model where the model does not require addi" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 279, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 279, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 279, + 713 + ], + "type": "text", + "content": "1Refer Sec. B of supp. for a thorough discussion on related work." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23605" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 295, + 312 + ], + "type": "text", + "content": "tional training. However, existing visual in-context learning approaches involve training and/or use of additional task-related data. Further, when we investigated V-ICL approaches based on uncurated datasets" + }, + { + "bbox": [ + 54, + 72, + 295, + 312 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 54, + 72, + 295, + 312 + ], + "type": "text", + "content": " (e.g. Bar et al. [3]) through multiple experiments, we observe the following, (1) the model fails to reconstruct the input prompt, despite it being unmasked in the input, as shown in the top-left image of the foundation model's output in Fig. 2a, (2) the model struggles in the presence of multiple objects in the image as shown in Fig. 2a where it incorrectly predicts the bird feeder as the foreground, and (3) when provided with the query and its groundtruth as the example prompt, the model struggles to make correct predictions despite having access to the groundtruth for the query (see Fig. 2b). These observations indicate that the model is unable to fully leverage the prompts. We attribute this inability of the model to lack of appropriate context interpretation (1) between the source and target of the example prompt which is essential for task inference, and (2) between the query and the source of the example prompt which is essential for accurate predictions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "text", + "content": "To alleviate these limitations, we propose a novel inference-only Stable Diffusion based visual in-context learning pipeline (SD-VICL), that unlike existing approaches, requires no additional training. Stable Diffusion [20] is a latent diffusion model that generates high-quality images by iteratively refining random noise through a denoising process. The denoising process at a given time step " + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "text", + "content": " employs a denoising U-Net which comprises of multiple self-attention layers operating at resolutions " + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "text", + "content": ". At each self-attention layer (" + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "text", + "content": "), the input features of the intermediate noise latent, " + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\phi(z_{t})" + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "text", + "content": ", are transformed to Query (" + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "text", + "content": "), Key (" + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "text", + "content": "), and Value (" + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "text", + "content": ") vectors using linear layers. The " + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "text", + "content": " vectors are used to compute the self-attention map using:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 110, + 483, + 295, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 483, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 110, + 483, + 295, + 514 + ], + "type": "interline_equation", + "content": "\\alpha^ {(t)} = \\operatorname {s o f t m a x} \\left(\\frac {Q ^ {(t)} \\cdot K ^ {(t) ^ {T}}}{\\sqrt {d}}\\right), \\tag {1}", + "image_path": "ace88b3ad27ba47634312fbacdf6c9add4bf54e2e48bfac7de3553a133ed3116.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 517, + 296, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 296, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 296, + 575 + ], + "type": "text", + "content": "which captures the correspondences within the image. The parameter " + }, + { + "bbox": [ + 55, + 517, + 296, + 575 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 517, + 296, + 575 + ], + "type": "text", + "content": " denotes the feature dimension of the " + }, + { + "bbox": [ + 55, + 517, + 296, + 575 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 517, + 296, + 575 + ], + "type": "text", + "content": " vector. The self-attention map " + }, + { + "bbox": [ + 55, + 517, + 296, + 575 + ], + "type": "inline_equation", + "content": "(\\alpha)" + }, + { + "bbox": [ + 55, + 517, + 296, + 575 + ], + "type": "text", + "content": " is used to update the intermediate feature map using the feature update, " + }, + { + "bbox": [ + 55, + 517, + 296, + 575 + ], + "type": "inline_equation", + "content": "\\Delta \\phi" + }, + { + "bbox": [ + 55, + 517, + 296, + 575 + ], + "type": "text", + "content": ", which is computed as," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 575, + 295, + 588 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 575, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 136, + 575, + 295, + 588 + ], + "type": "interline_equation", + "content": "\\Delta \\phi^ {(t)} = \\alpha^ {(t)} \\cdot V ^ {(t)}. \\tag {2}", + "image_path": "b687e4922f57f7534aa0c445d8327a1fce36e8c2c28fd7fddb3690520b09c179.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 591, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 296, + 651 + ], + "type": "text", + "content": "This updated set of features, along with the updates from the cross-attention layers are used to compute the intermediate noise predictions. In this paper, we focus only on the self-attention computations. Hence, for simplicity, we write the iterative denoising process as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 654, + 295, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 654, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 107, + 654, + 295, + 669 + ], + "type": "interline_equation", + "content": "z ^ {(t - 1)} = z ^ {(t)} - \\epsilon \\left(Q ^ {(t)}, K ^ {(t)}, V ^ {(t)}\\right), \\tag {3}", + "image_path": "2b53b0483f297a5e06bd5242f13fd0532b0e94c6a41050e061dc9c185b79f63c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "where the predicted noise, " + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": ", is a function of the " + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": " vectors of the self-attention layers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 103, + 555, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 103, + 555, + 129 + ], + "spans": [ + { + "bbox": [ + 313, + 103, + 555, + 129 + ], + "type": "text", + "content": "2.2. Repurposing Stable Diffusion for Visual In-Context Learning (SD-VICL)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 133, + 555, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 133, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 312, + 133, + 555, + 277 + ], + "type": "text", + "content": "As discussed previously, for successful in-context learning, a foundation model needs to appropriately infer (1) the task: the relationship between the prompt image " + }, + { + "bbox": [ + 312, + 133, + 555, + 277 + ], + "type": "inline_equation", + "content": "(A)" + }, + { + "bbox": [ + 312, + 133, + 555, + 277 + ], + "type": "text", + "content": " and prompt groundtruth " + }, + { + "bbox": [ + 312, + 133, + 555, + 277 + ], + "type": "inline_equation", + "content": "(B)" + }, + { + "bbox": [ + 312, + 133, + 555, + 277 + ], + "type": "text", + "content": ", and (2) the context: the relationship between the query image " + }, + { + "bbox": [ + 312, + 133, + 555, + 277 + ], + "type": "inline_equation", + "content": "(C)" + }, + { + "bbox": [ + 312, + 133, + 555, + 277 + ], + "type": "text", + "content": " and the prompt image " + }, + { + "bbox": [ + 312, + 133, + 555, + 277 + ], + "type": "inline_equation", + "content": "(A)" + }, + { + "bbox": [ + 312, + 133, + 555, + 277 + ], + "type": "text", + "content": ". In an attempt to ensure that these relationships are effectively inferred, we formulate a novel attention computation in place of the traditional self-attention computation in the upsample layers of the denoising U-Net. This formulation is inspired by a few recent research [1, 5, 18, 26] that repurpose the self-attention layers of Stable Diffusion for tasks like image editing and style transfer." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 277, + 554, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 554, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 554, + 365 + ], + "type": "text", + "content": "First, each of the input images: prompt image " + }, + { + "bbox": [ + 313, + 277, + 554, + 365 + ], + "type": "inline_equation", + "content": "(A)" + }, + { + "bbox": [ + 313, + 277, + 554, + 365 + ], + "type": "text", + "content": ", prompt groundtruth " + }, + { + "bbox": [ + 313, + 277, + 554, + 365 + ], + "type": "inline_equation", + "content": "(B)" + }, + { + "bbox": [ + 313, + 277, + 554, + 365 + ], + "type": "text", + "content": ", and query image " + }, + { + "bbox": [ + 313, + 277, + 554, + 365 + ], + "type": "inline_equation", + "content": "(C)" + }, + { + "bbox": [ + 313, + 277, + 554, + 365 + ], + "type": "text", + "content": " are inverted to the Stable Diffusion's noise space to obtain, " + }, + { + "bbox": [ + 313, + 277, + 554, + 365 + ], + "type": "inline_equation", + "content": "z_A^{(T)}" + }, + { + "bbox": [ + 313, + 277, + 554, + 365 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 277, + 554, + 365 + ], + "type": "inline_equation", + "content": "z_B^{(T)}" + }, + { + "bbox": [ + 313, + 277, + 554, + 365 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 277, + 554, + 365 + ], + "type": "inline_equation", + "content": "z_C^{(T)}" + }, + { + "bbox": [ + 313, + 277, + 554, + 365 + ], + "type": "text", + "content": " respectively, using an off-the-shelf inversion model [13]. Then, each of these images are iteratively denoised in parallel without any changes to the default denoising pipeline as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 366, + 371, + 555, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 371, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 366, + 371, + 555, + 386 + ], + "type": "interline_equation", + "content": "z _ {p} ^ {(t - 1)} = z _ {p} ^ {(t)} - \\epsilon \\left(Q _ {p} ^ {(t)}, K _ {p} ^ {(t)}, V _ {p} ^ {(t)}\\right), \\tag {4}", + "image_path": "1987672f0f79fac2afbc1950c4ed24850442a7f129588e601b3cdabbf113036b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 393, + 482, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 482, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 482, + 406 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 393, + 482, + 406 + ], + "type": "inline_equation", + "content": "p\\in \\{A,B,C\\}" + }, + { + "bbox": [ + 313, + 393, + 482, + 406 + ], + "type": "inline_equation", + "content": "t\\in [T,1],t\\in \\mathbb{Z}^{+}" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 406, + 554, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 554, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 554, + 456 + ], + "type": "text", + "content": "Since the output image is expected to be structurally similar to the query image for most vision tasks, we initialize the noise space in the prediction pipeline " + }, + { + "bbox": [ + 313, + 406, + 554, + 456 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 406, + 554, + 456 + ], + "type": "text", + "content": " using the noise space of the query image (i.e., " + }, + { + "bbox": [ + 313, + 406, + 554, + 456 + ], + "type": "inline_equation", + "content": "z_{D}^{(T)} \\gets z_{C}^{(T)}" + }, + { + "bbox": [ + 313, + 406, + 554, + 456 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 455, + 554, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 554, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 554, + 478 + ], + "type": "text", + "content": "Different from Eq. (4), the query prediction is denoised using," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 366, + 478, + 553, + 492 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 478, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 366, + 478, + 553, + 492 + ], + "type": "interline_equation", + "content": "z _ {D} ^ {(t - 1)} = z _ {D} ^ {(t)} - \\epsilon \\left(Q _ {C} ^ {(t)}, K _ {A} ^ {(t)}, V _ {B} ^ {(t)}\\right), \\tag {5}", + "image_path": "f60d4fb0a822246120db76a1a5a73c64c13b87834ef871c7dce66d0dcb0b6ef4.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 495, + 418, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 418, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 418, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 495, + 418, + 506 + ], + "type": "inline_equation", + "content": "t\\in [T,1],t\\in \\mathbb{Z}^{+}" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 325, + 506, + 552, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 506, + 552, + 519 + ], + "spans": [ + { + "bbox": [ + 325, + 506, + 552, + 519 + ], + "type": "text", + "content": "Expanding the feature update formulation for " + }, + { + "bbox": [ + 325, + 506, + 552, + 519 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 325, + 506, + 552, + 519 + ], + "type": "text", + "content": " we get," + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 353, + 525, + 553, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 525, + 553, + 541 + ], + "spans": [ + { + "bbox": [ + 353, + 525, + 553, + 541 + ], + "type": "interline_equation", + "content": "\\Delta \\phi_ {D} ^ {(t)} = \\alpha_ {D} ^ {(t)} \\cdot V _ {B} ^ {(t)}, \\tag {6}", + "image_path": "e001baf71a710a038b5fe4030b308591b6085a0dbb024dad9b51bac4681b7e4b.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 354, + 542, + 553, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 542, + 553, + 572 + ], + "spans": [ + { + "bbox": [ + 354, + 542, + 553, + 572 + ], + "type": "interline_equation", + "content": "\\Delta \\phi_ {D} ^ {(t)} = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {C} ^ {(t)} \\cdot K _ {A} ^ {(t) ^ {T}}}{\\tau \\cdot \\sqrt {d}}\\right) \\cdot V _ {B} ^ {(t)}. \\tag {7}", + "image_path": "4a56b8f423f3648469351e32daedb17c6d85cc201cc34cb50baf80447dd59643.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "type": "text", + "content": "In contrast to Eqs. (1) and (2), the attention map is formulated using the Query vector of the query image " + }, + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "type": "inline_equation", + "content": "(C)" + }, + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "type": "text", + "content": " and the Key vector of the prompt image " + }, + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "type": "inline_equation", + "content": "(A)" + }, + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "type": "inline_equation", + "content": "Q_{C}" + }, + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "type": "text", + "content": " vector comprises of the semantics of each spatial location in " + }, + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "type": "text", + "content": ", and the " + }, + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "type": "inline_equation", + "content": "K_{A}" + }, + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "type": "text", + "content": " vector offers the context within " + }, + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "type": "text", + "content": " that each query can attend to. Hence, each element in the attention map, i.e., " + }, + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\alpha_{D(i,j)}^{t}" + }, + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "type": "text", + "content": ", captures how the " + }, + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "type": "text", + "content": " patch of the query image correlates with the " + }, + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "type": "text", + "content": " patch of the prompt image. This formulation explicitly enforces the context between the query image and the prompt image. The additional temperature hyperparameter " + }, + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "type": "inline_equation", + "content": "(\\tau)" + }, + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "type": "text", + "content": " (in Eq. (7)) for the softmax computation" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 673, + 296, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 673, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 673, + 296, + 713 + ], + "type": "text", + "content": "2We use approaches based on uncurated datasets as the basis for our analysis/formulations since they do not explicitly use images from out-of-domain tasks for training and are much closer to in-context learning in the NLP community as compared to the approaches that use curated datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23606" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 91, + 68, + 521, + 261 + ], + "blocks": [ + { + "bbox": [ + 91, + 68, + 521, + 261 + ], + "lines": [ + { + "bbox": [ + 91, + 68, + 521, + 261 + ], + "spans": [ + { + "bbox": [ + 91, + 68, + 521, + 261 + ], + "type": "image", + "image_path": "630f5816d67c371b5e728424dbd09b019ddde6f498acbdbeb9752e51478cee34.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 268, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 54, + 268, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 54, + 268, + 555, + 346 + ], + "type": "text", + "content": "Figure 3. Latent representations of the query and prompt images from the Stable Diffusion model's VAE encoder are transformed to the noise space using [13]. Each noisy latent is iteratively denoised using the denoising U-Net. Prediction " + }, + { + "bbox": [ + 54, + 268, + 555, + 346 + ], + "type": "inline_equation", + "content": "(D)" + }, + { + "bbox": [ + 54, + 268, + 555, + 346 + ], + "type": "text", + "content": " path is initialized with the noise space of " + }, + { + "bbox": [ + 54, + 268, + 555, + 346 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 54, + 268, + 555, + 346 + ], + "type": "text", + "content": ", and at each denoising step, the computations within the self-attention layers are modified to enhance visual ICL capabilities. With the Query (from prompt image-" + }, + { + "bbox": [ + 54, + 268, + 555, + 346 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 54, + 268, + 555, + 346 + ], + "type": "text", + "content": "), the Key (from prompt groundtruth-" + }, + { + "bbox": [ + 54, + 268, + 555, + 346 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 54, + 268, + 555, + 346 + ], + "type": "text", + "content": "), and the Value (from query image-" + }, + { + "bbox": [ + 54, + 268, + 555, + 346 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 54, + 268, + 555, + 346 + ], + "type": "text", + "content": ") vectors, we reformulate the intermediate feature update to infuse context between the example prompts and the query image into the foundation model. This process is iteratively performed for " + }, + { + "bbox": [ + 54, + 268, + 555, + 346 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 54, + 268, + 555, + 346 + ], + "type": "text", + "content": " denoising steps to obtain the denoised prediction latent, which is then fed to the VAE decoder to obtain the final prediction " + }, + { + "bbox": [ + 54, + 268, + 555, + 346 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 54, + 268, + 555, + 346 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 355, + 296, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 296, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 296, + 620 + ], + "type": "text", + "content": "controls the sharpness of the correlation. By adjusting " + }, + { + "bbox": [ + 56, + 355, + 296, + 620 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 56, + 355, + 296, + 620 + ], + "type": "text", + "content": ", we can modulate the model's focus on attention/correlation between " + }, + { + "bbox": [ + 56, + 355, + 296, + 620 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 56, + 355, + 296, + 620 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 355, + 296, + 620 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 56, + 355, + 296, + 620 + ], + "type": "text", + "content": ". A lower " + }, + { + "bbox": [ + 56, + 355, + 296, + 620 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 56, + 355, + 296, + 620 + ], + "type": "text", + "content": " sharpens the attention, emphasizing stronger correlation, while a higher " + }, + { + "bbox": [ + 56, + 355, + 296, + 620 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 56, + 355, + 296, + 620 + ], + "type": "text", + "content": " smoothens the attention, allowing for a broader distribution of focus across multiple correlations. Furthermore, we use the Value vector from the prompt groundtruth " + }, + { + "bbox": [ + 56, + 355, + 296, + 620 + ], + "type": "inline_equation", + "content": "(B)" + }, + { + "bbox": [ + 56, + 355, + 296, + 620 + ], + "type": "text", + "content": " to compute " + }, + { + "bbox": [ + 56, + 355, + 296, + 620 + ], + "type": "inline_equation", + "content": "\\Delta \\phi_D^{(t)}" + }, + { + "bbox": [ + 56, + 355, + 296, + 620 + ], + "type": "text", + "content": ". This use of value vector from a different latent, in contrast to the standard implementation of cross-attention [27], where both Key and Value vectors typically come from the same set of intermediate latents, is to facilitate the prediction to be in the same domain as that of the prompt groundtruth " + }, + { + "bbox": [ + 56, + 355, + 296, + 620 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 56, + 355, + 296, + 620 + ], + "type": "text", + "content": ". This formulation, which is in-place of the computation within the standard self-attention layer of Stable Diffusion, facilitates the explicit infusion of the context between the query image and the prompt in addition to improved task inference. However, the modification of the default self-attention computation and the use of " + }, + { + "bbox": [ + 56, + 355, + 296, + 620 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 56, + 355, + 296, + 620 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 355, + 296, + 620 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 56, + 355, + 296, + 620 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 355, + 296, + 620 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 56, + 355, + 296, + 620 + ], + "type": "text", + "content": " vectors that are extracted from three distinct images create a domain gap that needs to be addressed to enhance the quality of the prediction. Hence, we adapt the attention map contrasting, swap-guidance, and AdaIN mechanisms employed in [1]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": "Attention map contrasting supplements the function of the temperature hyperparameter " + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "(\\tau)" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": ", where it enhances the focus on the relevant regions (i.e. semantically similar regions) between " + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": ", while attenuating the irrelevant (i.e. color similarity between unrelated objects). Additionally, this operation adjusts the scale of the values to suit the pre-trained Stable Diffusion pipeline. This contrasting op" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 356, + 428, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 428, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 428, + 366 + ], + "type": "text", + "content": "eration can be expressed as," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 358, + 371, + 555, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 371, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 358, + 371, + 555, + 389 + ], + "type": "interline_equation", + "content": "\\alpha_ {D} ^ {(t)} = \\mu \\left(\\alpha_ {D} ^ {(t)}\\right) + \\beta \\cdot \\left(\\alpha_ {D} ^ {(t)} - \\mu \\left(\\alpha_ {D} ^ {(t)}\\right)\\right), \\tag {8}", + "image_path": "9e460e123568f2dd2946d652767e01b4c1d5272168fbc97beb222ab6b2133fab.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 393, + 555, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 555, + 415 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 393, + 555, + 415 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 393, + 555, + 415 + ], + "type": "text", + "content": " is the mean operation and " + }, + { + "bbox": [ + 313, + 393, + 555, + 415 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 393, + 555, + 415 + ], + "type": "text", + "content": " is a hyperparameter that controls the scale." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 416, + 554, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 416, + 554, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 554, + 496 + ], + "type": "text", + "content": "Swap-guidance is derived from classifier-free diffusion guidance implementation proposed by Ho and Salimans [11]. Given the noise prediction from the default self-attention formulation " + }, + { + "bbox": [ + 313, + 416, + 554, + 496 + ], + "type": "inline_equation", + "content": "(\\eta_{default}^{(t)})" + }, + { + "bbox": [ + 313, + 416, + 554, + 496 + ], + "type": "text", + "content": " and the noise prediction from our modified formulation " + }, + { + "bbox": [ + 313, + 416, + 554, + 496 + ], + "type": "inline_equation", + "content": "(\\eta_{modified}^{(t)})" + }, + { + "bbox": [ + 313, + 416, + 554, + 496 + ], + "type": "text", + "content": ", the final noise prediction " + }, + { + "bbox": [ + 313, + 416, + 554, + 496 + ], + "type": "inline_equation", + "content": "(\\eta^{(t)})" + }, + { + "bbox": [ + 313, + 416, + 554, + 496 + ], + "type": "text", + "content": " can be expressed as," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 321, + 500, + 555, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 500, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 321, + 500, + 555, + 521 + ], + "type": "interline_equation", + "content": "\\eta^ {(t)} = \\eta_ {\\text {d e f a u l t}} ^ {(t)} + \\frac {\\gamma \\cdot (T - t)}{T} \\cdot \\left(\\eta_ {\\text {m o d i f i e d}} ^ {(t)} - \\eta_ {\\text {d e f a u l t}} ^ {(t)}\\right), \\tag {9}", + "image_path": "46abd446bbda90bfcc3806454379cc16c011a42dcf2a9d24f6a226fff9137dfb.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 525, + 554, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 554, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 554, + 583 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 525, + 554, + 583 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 525, + 554, + 583 + ], + "type": "text", + "content": " is a hyperparameter controlling the scale. This mechanism gradually incorporates the modified noise prediction as denoising progresses, which directs the denoising process through denser regions of the Stable Diffusion's generative pipeline, thus alleviating unwarranted artifacts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 584, + 554, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 584, + 554, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 584, + 554, + 650 + ], + "type": "text", + "content": "Adaptive instance normalization (AdaIN), which was proposed by Huang and Belongie [12] is used to align the color distribution between the prediction " + }, + { + "bbox": [ + 313, + 584, + 554, + 650 + ], + "type": "inline_equation", + "content": "(D)" + }, + { + "bbox": [ + 313, + 584, + 554, + 650 + ], + "type": "text", + "content": ", which is initialized using the noise space of the query image " + }, + { + "bbox": [ + 313, + 584, + 554, + 650 + ], + "type": "inline_equation", + "content": "(C)" + }, + { + "bbox": [ + 313, + 584, + 554, + 650 + ], + "type": "text", + "content": ", and the final groundtruth/task color space (i.e. color space of " + }, + { + "bbox": [ + 313, + 584, + 554, + 650 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 584, + 554, + 650 + ], + "type": "text", + "content": "). AdaIN is implemented as follows," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 354, + 654, + 553, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 654, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 354, + 654, + 553, + 668 + ], + "type": "interline_equation", + "content": "z _ {D} ^ {(t)} \\leftarrow \\operatorname {A d a I N} \\left(z _ {D} ^ {(t)}, z _ {B} ^ {(t)}\\right), \\tag {10}", + "image_path": "092a664f407ab758ba53e1b7c605997a62ff1a836881f81f93b93ab393c73254.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 354, + 670, + 553, + 698 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 670, + 553, + 698 + ], + "spans": [ + { + "bbox": [ + 354, + 670, + 553, + 698 + ], + "type": "interline_equation", + "content": "z _ {D} ^ {(t)} \\leftarrow \\frac {z _ {D} ^ {(t)} - \\mu \\left(z _ {D} ^ {(t)}\\right)}{\\sigma \\left(z _ {D} ^ {(t)}\\right)} \\cdot \\sigma \\left(z _ {B} ^ {(t)}\\right) + \\mu \\left(z _ {B} ^ {(t)}\\right), \\tag {11}", + "image_path": "34775926934f01b721f4eab3c9e34cce3611e3f5b2cb95becba3fac0195c2f42.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 702, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 702, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 702, + 553, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 702, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 702, + 553, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 702, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 702, + 553, + 714 + ], + "type": "text", + "content": " are mean and standard deviation operations." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23607" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "2.3. Implicitly-Weighted Prompt Ensembling (IWPE)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 101, + 296, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 101, + 296, + 268 + ], + "spans": [ + { + "bbox": [ + 55, + 101, + 296, + 268 + ], + "type": "text", + "content": "Bar et al. [3] and Xu et al. [33] demonstrate that providing the model with multiple source-target example prompts improves the prediction performance of visual in-context learning. These methods ensemble multiple prompts at the input level by stitching together prompt images and corresponding groundtruths to form a grid-like structure in the form of a composite image. This composite image is then input to the model, which is expected to leverage information from multiple prompts to make a better prediction compared to the single prompt scenario. However, for a fixed-resolution foundation model, as the number of source-target examples increases, the effective size of each image in the grid is reduced. This leads to performance deterioration, mainly because of the loss of details in the input [33]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 269, + 296, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 269, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 296, + 474 + ], + "type": "text", + "content": "To address this issue, instead of ensembling prompts in the image space through a composite input, SegGPT [29] proposed a feature space ensembling method. In this method, multiple prompts are processed in parallel and aggregated by averaging the features at the end of each attention layer. This approach assumes that all the example prompts are equally informative and they are uniformly weighted during the averaging process. However, such uniform weighting of prompts is sub-optimal since it prevents the model from benefiting from more relevant prompts. To this end, we propose a simple but effective ensembling method by integrating the ensembling into the attention computation, allowing the prompt patches to be implicitly weighted based on their correspondences with each query patch. Specifically, we concatenate each of the " + }, + { + "bbox": [ + 55, + 269, + 296, + 474 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 269, + 296, + 474 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 269, + 296, + 474 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 269, + 296, + 474 + ], + "type": "text", + "content": " vectors from each of the prompts prior to computing " + }, + { + "bbox": [ + 55, + 269, + 296, + 474 + ], + "type": "inline_equation", + "content": "\\Delta \\phi_D^{(t)}" + }, + { + "bbox": [ + 55, + 269, + 296, + 474 + ], + "type": "text", + "content": ". Specifically, for " + }, + { + "bbox": [ + 55, + 269, + 296, + 474 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 269, + 296, + 474 + ], + "type": "text", + "content": " number of prompts, we re-write Eq. (7):" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 477, + 295, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 477, + 295, + 513 + ], + "spans": [ + { + "bbox": [ + 63, + 477, + 295, + 513 + ], + "type": "interline_equation", + "content": "\\Delta \\phi_ {D} ^ {(t)} = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {C} ^ {(t)} \\cdot \\left(\\bigoplus_ {i = 1} ^ {n} K _ {A _ {i}} ^ {(t)}\\right) ^ {T}}{\\tau \\cdot \\sqrt {d}}\\right) \\cdot \\left(\\bigoplus_ {i = 1} ^ {n} V _ {B _ {i}} ^ {(t)}\\right), \\tag {12}", + "image_path": "d5b4cf4cafd8d7d463ef36a4cf22153262ecb59461bc5743b5b011138ffac53a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 515, + 295, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 295, + 540 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 515, + 295, + 540 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 55, + 515, + 295, + 540 + ], + "type": "text", + "content": " is the concatenation operation and " + }, + { + "bbox": [ + 55, + 515, + 295, + 540 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 515, + 295, + 540 + ], + "type": "text", + "content": " is the index for each prompt." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 550, + 222, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 222, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 222, + 563 + ], + "type": "text", + "content": "3. Experiments and Evaluations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "text", + "content": "To demonstrate the ability of the proposed approach to generalize across different tasks, we evaluate it on six downstream tasks. For all these tasks, we use the unsupervised prompt retrieval [35] to select the candidates for the prompt images. Specifically, this method chooses the nearest neighbors of the query image as prompt candidates. The nearest neighbor retrieval is based on the cosine similarity of CLIP's vision encoder [19] embeddings between the query image and the example prompt images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "For a fair comparison, we use Visual Prompting [3] and IMProv [33] as baselines. Although these approaches involve a training step, unlike [28, 31], they do not rely on" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 554, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 192 + ], + "type": "text", + "content": "curated or annotated data that come from related out-of-domain tasks. Additionally, since IMProv supports supplementary text guidance, we include results with and without text guidance in our comparisons. Further, we also evaluate and compare the effectiveness of multiple prompts. Please refer to the supplementary for details on datasets, evaluation metrics, and sensitivity analysis for different hyperparameters. For the results in the following subsections, we choose one set of hyperparameters that provided optimal performance for all the tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 199, + 434, + 212 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 199, + 434, + 212 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 434, + 212 + ], + "type": "text", + "content": "3.1. Results and Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 217, + 554, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 554, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 554, + 252 + ], + "type": "text", + "content": "Quantitative metrics for all the six tasks are reported in Tab. 1. The proposed approach consistently outperforms existing uncurated based approaches in all the tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 253, + 554, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 554, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 554, + 361 + ], + "type": "text", + "content": "Foreground segmentation: With a single prompt, the proposed method achieves absolute improvements of " + }, + { + "bbox": [ + 313, + 253, + 554, + 361 + ], + "type": "inline_equation", + "content": "8.9\\%" + }, + { + "bbox": [ + 313, + 253, + 554, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 253, + 554, + 361 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 313, + 253, + 554, + 361 + ], + "type": "text", + "content": " in mean intersection over union (mIoU) on an average across all splits3 in the Pascal-5i dataset [22], compared to Visual Prompting and IMProv, respectively. The proposed ensembling method with five prompts using the repurposed Stable Diffusion model further improves the mIoU by " + }, + { + "bbox": [ + 313, + 253, + 554, + 361 + ], + "type": "inline_equation", + "content": "11.6\\%" + }, + { + "bbox": [ + 313, + 253, + 554, + 361 + ], + "type": "text", + "content": " (in absolute terms) as compared to the single prompt case." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 361, + 554, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 554, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 554, + 456 + ], + "type": "text", + "content": "Single object detection: Similar to foreground segmentation, with a single prompt, the proposed method improves the mIoU by " + }, + { + "bbox": [ + 313, + 361, + 554, + 456 + ], + "type": "inline_equation", + "content": "5.3\\%" + }, + { + "bbox": [ + 313, + 361, + 554, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 361, + 554, + 456 + ], + "type": "inline_equation", + "content": "7.1\\%" + }, + { + "bbox": [ + 313, + 361, + 554, + 456 + ], + "type": "text", + "content": " (in absolute terms) on average across all splits3 in the Pascal-5i dataset [22] as compared to Visual Prompting and IMProv, respectively. The use of multiple prompts through implicitly-weighted prompt assembling results in an additional " + }, + { + "bbox": [ + 313, + 361, + 554, + 456 + ], + "type": "inline_equation", + "content": "5.6\\%" + }, + { + "bbox": [ + 313, + 361, + 554, + 456 + ], + "type": "text", + "content": " of improvement as compared to the single prompt case." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 456, + 554, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 554, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 554, + 696 + ], + "type": "text", + "content": "Semantic segmentation: On the Cityscapes dataset [6], our method with a single prompt results in an absolute improvement of " + }, + { + "bbox": [ + 313, + 456, + 554, + 696 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 313, + 456, + 554, + 696 + ], + "type": "text", + "content": " in mIoU and " + }, + { + "bbox": [ + 313, + 456, + 554, + 696 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 313, + 456, + 554, + 696 + ], + "type": "text", + "content": " increase in accuracy as compared to Visual Prompting. Similarly, we observe an absolute gain of " + }, + { + "bbox": [ + 313, + 456, + 554, + 696 + ], + "type": "inline_equation", + "content": "6.1\\%" + }, + { + "bbox": [ + 313, + 456, + 554, + 696 + ], + "type": "text", + "content": " in mIoU and a " + }, + { + "bbox": [ + 313, + 456, + 554, + 696 + ], + "type": "inline_equation", + "content": "6.5\\%" + }, + { + "bbox": [ + 313, + 456, + 554, + 696 + ], + "type": "text", + "content": " in accuracy as compared to IMProv. With IWPE, the mIoU and accuracy further improve by " + }, + { + "bbox": [ + 313, + 456, + 554, + 696 + ], + "type": "inline_equation", + "content": "4.9\\%" + }, + { + "bbox": [ + 313, + 456, + 554, + 696 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 456, + 554, + 696 + ], + "type": "inline_equation", + "content": "6.8\\%" + }, + { + "bbox": [ + 313, + 456, + 554, + 696 + ], + "type": "text", + "content": " (in absolute terms) as compared to the single prompt based approach. The proposed approach with a single prompt in the case of semantic segmentation achieves lower performance improvements over existing methods as compared to other tasks like foreground segmentation and keypoint detection. However, with additional prompts, the proposed method is able to achieve much higher gains similar to what is observed in other tasks. This performance boost can be attributed to two key factors: (1) multiple prompts help resolve conflicts between similar or overlapping classes by mitigating inter-class attention ambiguities, and (2) a single prompt is unlikely to contain all classes that are potentially present in the query image of a multi-class task and in such cases, ac" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 702, + 502, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 702, + 502, + 713 + ], + "spans": [ + { + "bbox": [ + 324, + 702, + 502, + 713 + ], + "type": "text", + "content": "See supplementary for the performance on each split." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23608" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 102, + 70, + 509, + 373 + ], + "blocks": [ + { + "bbox": [ + 102, + 70, + 509, + 373 + ], + "lines": [ + { + "bbox": [ + 102, + 70, + 509, + 373 + ], + "spans": [ + { + "bbox": [ + 102, + 70, + 509, + 373 + ], + "type": "image", + "image_path": "190352a69e33d040911ffed063d4f2259ca8211c588d9edcef0c0533b5d0cdfe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 380, + 555, + 404 + ], + "lines": [ + { + "bbox": [ + 55, + 380, + 555, + 404 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 555, + 404 + ], + "type": "text", + "content": "Figure 4. Qualitative evaluation, where we compare the performance of Visual Prompting [3] and IMProv [33] with our proposed method on six different tasks. It can be seen that our method produces visually superior results as compared to the baselines." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 414, + 295, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 414, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 414, + 295, + 449 + ], + "type": "text", + "content": "cess to multiple prompts increases the probability that all relevant classes are available to the model for making accurate predictions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 449, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 533 + ], + "type": "text", + "content": "Keypoint detection: When evaluated on the DeepFashion dataset [17], our method demonstrates a substantial improvement, reducing the MSE by " + }, + { + "bbox": [ + 55, + 449, + 295, + 533 + ], + "type": "inline_equation", + "content": "6 \\times" + }, + { + "bbox": [ + 55, + 449, + 295, + 533 + ], + "type": "text", + "content": " while achieving a " + }, + { + "bbox": [ + 55, + 449, + 295, + 533 + ], + "type": "inline_equation", + "content": "7 \\times" + }, + { + "bbox": [ + 55, + 449, + 295, + 533 + ], + "type": "text", + "content": " gain in percentage of correct keypoints (PCK), compared to the best-performing baseline. Furthermore, the use of five prompts enhances these metrics, yielding additional relative gains of " + }, + { + "bbox": [ + 55, + 449, + 295, + 533 + ], + "type": "inline_equation", + "content": "18.5\\%" + }, + { + "bbox": [ + 55, + 449, + 295, + 533 + ], + "type": "text", + "content": " for MSE and " + }, + { + "bbox": [ + 55, + 449, + 295, + 533 + ], + "type": "inline_equation", + "content": "6.5\\%" + }, + { + "bbox": [ + 55, + 449, + 295, + 533 + ], + "type": "text", + "content": " for PCK." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 533, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 605 + ], + "type": "text", + "content": "Edge detection: On the NYUDv2 dataset [23], our approach reduces the mean squared error (MSE) by " + }, + { + "bbox": [ + 55, + 533, + 295, + 605 + ], + "type": "inline_equation", + "content": "72.0\\%" + }, + { + "bbox": [ + 55, + 533, + 295, + 605 + ], + "type": "text", + "content": " and LPIPS [34] by " + }, + { + "bbox": [ + 55, + 533, + 295, + 605 + ], + "type": "inline_equation", + "content": "60.6\\%" + }, + { + "bbox": [ + 55, + 533, + 295, + 605 + ], + "type": "text", + "content": " as compared to the best performing existing approach. Note that these metrics improve an additional " + }, + { + "bbox": [ + 55, + 533, + 295, + 605 + ], + "type": "inline_equation", + "content": "24.5\\%" + }, + { + "bbox": [ + 55, + 533, + 295, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 533, + 295, + 605 + ], + "type": "inline_equation", + "content": "21.4\\%" + }, + { + "bbox": [ + 55, + 533, + 295, + 605 + ], + "type": "text", + "content": ", respectively, with our prompt ensembling method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "text", + "content": "Colorization: Evaluating on the ImageNet dataset [21], our model reduces LPIPS by " + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "inline_equation", + "content": "27.9\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "text", + "content": " and FID [10] by " + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "inline_equation", + "content": "49.0\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "text", + "content": ", relative to IMProv, which yields the next best performance, and shows further gains of " + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "inline_equation", + "content": "19.1\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "inline_equation", + "content": "16.0\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "text", + "content": " respectively, with five prompts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "type": "text", + "content": "Overall, our method outperforms both Visual Prompting and IMProv across all tasks by considerable margins. Additionally, the integration of multiple prompts using IWPE consistently yields performance improvements over" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 414, + 425, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 425, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 425, + 426 + ], + "type": "text", + "content": "the single-prompt scenario." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 426, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 582 + ], + "type": "text", + "content": "Supplementing the quantitative results, we also demonstrate visual comparisons of the outputs for each task in Fig. 4" + }, + { + "bbox": [ + 313, + 426, + 555, + 582 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 313, + 426, + 555, + 582 + ], + "type": "text", + "content": ", further highlighting the superior performance of our method over baselines. We observe that while both Visual Prompting and IMProv perform reasonably in single object detection and foreground segmentation, they show a weaker performance in tasks such as colorization, edge detection, and semantic segmentation, where capturing fine details and preserving structural information of the query is crucial. Moreover, both baselines notably struggle with keypoint detection, whereas the proposed approach demonstrates far superior results, indicating its robustness to different out-of-domain tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 591, + 444, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 591, + 444, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 591, + 444, + 602 + ], + "type": "text", + "content": "3.2. Additional Evaluations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 609, + 554, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 554, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 554, + 693 + ], + "type": "text", + "content": "Comparison with V-ICL models trained on task-related data: To provide a comprehensive evaluation, we compare our training-free approach against prior methods trained explicitly on task-related data, including Painter [28], LVM [2], and Prompt Diffusion [31]. Quantitative results are presented in Tab. 2, and qualitative comparisons are shown in Fig. 7 (supplementary). Despite the lack of any ex" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 702, + 539, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 702, + 539, + 713 + ], + "spans": [ + { + "bbox": [ + 324, + 702, + 539, + 713 + ], + "type": "text", + "content": "4See supplementary for additional qualitative results for each task." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23609" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 70, + 541, + 186 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 541, + 186 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 541, + 186 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 541, + 186 + ], + "type": "table", + "html": "
ModelForeground Seg. (Pascal-5i) mIoU ↑Object Det. (Pascal-5i) mIoU ↑Semantic Segmentation (Cityscapes) mIoU ↑Keypoint Detection (DeepFashion) Acc. ↑MSE ↓PCK ↑Edge Detection (NYUDv2) MSE ↓LPIPS ↓Colorization (ImageNet) LPIPS ↓FID ↓
Number of Example Prompts: 1
Visual Prompting [3]35.0446.2921.3071.5235.8310.870.10060.39250.4166111.06
IMProv (w/o text) [33]39.5043.8215.9768.4431.3018.380.10590.42780.3895104.74
IMProv (w/ text) [33]40.6944.4215.8067.8844.672.810.11250.51460.3899105.90
SD-VICL (ours)43.9251.5522.0874.935.3677.190.02820.15480.281053.40
Number of Example Prompts: 5
Visual Prompting36.3748.0921.7070.9536.5710.320.10000.39000.4156112.91
SD-VICL (ours)55.4957.1027.0181.754.3782.240.02130.12160.227244.84
", + "image_path": "51c8a9189460b5be061cf6092d15756f7404c608bfc8dc0c2e6a28d7c4ccc39f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 224, + 325, + 286 + ], + "blocks": [ + { + "bbox": [ + 55, + 190, + 555, + 213 + ], + "lines": [ + { + "bbox": [ + 55, + 190, + 555, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 555, + 213 + ], + "type": "text", + "content": "Table 1. Quantitative performance comparison of the proposed approach with recent approaches on foreground segmentation, single object detection, semantic segmentation, keypoint detection, edge detection, and colorization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 224, + 325, + 286 + ], + "lines": [ + { + "bbox": [ + 57, + 224, + 325, + 286 + ], + "spans": [ + { + "bbox": [ + 57, + 224, + 325, + 286 + ], + "type": "table", + "html": "
ModelFG Seg. mIoU ↑Obj. Det. mIoU ↑Edge Det.Colorization
MSE ↓LPIPS ↓LPIPS ↓FID ↓
Painter [28]55.0954.280.09260.72940.347464.16
LVM [2]50.9852.670.04990.42590.314256.40
Prompt Diffusion [31]17.5326.860.04300.33110.699798.91
SD-VICL (ours)55.4957.100.02130.12160.227244.84
", + "image_path": "83fffe75b23d41e1c446e6671b0851ee9a8b42534f280bdb70566b61f6be467e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 326, + 297, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 326, + 297, + 649 + ], + "spans": [ + { + "bbox": [ + 54, + 326, + 297, + 649 + ], + "type": "text", + "content": "plicit training, our method consistently outperforms all three baselines across multiple vision tasks. Analyzing the results, we observe that all three models tend to overfit to their training tasks, leading to poor generalization when presented with novel tasks. For instance, Painter incorrectly outputs a depth map instead of edges, despite the prompt pair defining the task has an edge map as the target (Fig. 7 in supplementary). Further, while Painter and LVM perform well on tasks, foreground segmentation and object detection, when the query image consists of only a single class, they struggle in multi-class scenarios, often segmenting all classes instead of the intended one (Fig. 8a in supplementary). Additionally, LVM exhibits inconsistencies in its outputs, where for a given task, despite the format/domain of the inputs remaining unchanged, we observe that the generated output belongs to diverse domains (Fig. 8b in supplementary). Moreover, Prompt Diffusion performs notably poorly on most tasks, often generating incorrect outputs, with the exception of edge detection — a task it was exposed to during training. Ideally, a V-ICL model should infer context and task from the input, but all three models fall short in this aspect. In contrast, our training-free method demonstrates superior generalization and effective context and task inference as intended by V-ICL, underscoring the advantages of uncovering V-ICL properties without additional training. For a detailed discussion and visual examples, we refer the reader to Sec. E in the supplementary." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": "Comparison with task-specific models based on Stable Diffusion: Here, we compare our method against approaches that adapt Stable Diffusion to facilitate specific tasks, such as semantic segmentation and keypoint detection. DAS [24] is one such work that proposes a pipeline" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 353, + 227, + 550, + 283 + ], + "blocks": [ + { + "bbox": [ + 55, + 294, + 326, + 316 + ], + "lines": [ + { + "bbox": [ + 55, + 294, + 326, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 326, + 316 + ], + "type": "text", + "content": "Table 2. Extended quantitative evaluations against V-ICL models that train on task-related data." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 353, + 227, + 550, + 283 + ], + "lines": [ + { + "bbox": [ + 353, + 227, + 550, + 283 + ], + "spans": [ + { + "bbox": [ + 353, + 227, + 550, + 283 + ], + "type": "table", + "html": "
ModelSem.Seg.Keypoint Det.
mIoU ↑Acc. ↑MSE ↓PCK ↑
DAS [24]21.276.0--
StableKeypoints [9]--6.4670.0
SD-VICL (ours)31.5282.724.3782.24
", + "image_path": "10a6e0a4003c886f96d5844506c12cee0667110da712dd1cafe3c4137382336c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 350, + 291, + 552, + 313 + ], + "lines": [ + { + "bbox": [ + 350, + 291, + 552, + 313 + ], + "spans": [ + { + "bbox": [ + 350, + 291, + 552, + 313 + ], + "type": "text", + "content": "Table 3. Comparison with task-specific models based on Stable Diffusion." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 326, + 555, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 555, + 601 + ], + "type": "text", + "content": "that aggregates the self-attention layers of Stable Diffusion iteratively to predict segmentation masks for each class facilitating semantic segmentation. Similarly, StableKeypoints [9] optimizes on the text embeddings such that cross-attention maps condense to landmarks in a given image facilitating keypoint detection. We present the performance comparisons against these task-specific models in Tab. 3, where our approach outperforms both [24] and [9]. For semantic segmentation, we follow the evaluation protocol of DAS (i.e. Hungarian matching) to ensure a fair comparison and achieve absolute improvements of " + }, + { + "bbox": [ + 313, + 326, + 555, + 601 + ], + "type": "inline_equation", + "content": "10.3\\%" + }, + { + "bbox": [ + 313, + 326, + 555, + 601 + ], + "type": "text", + "content": " in mIoU and " + }, + { + "bbox": [ + 313, + 326, + 555, + 601 + ], + "type": "inline_equation", + "content": "6.7\\%" + }, + { + "bbox": [ + 313, + 326, + 555, + 601 + ], + "type": "text", + "content": " in accuracy. In keypoint detection, we surpass StableKeypoints with a " + }, + { + "bbox": [ + 313, + 326, + 555, + 601 + ], + "type": "inline_equation", + "content": "32.4\\%" + }, + { + "bbox": [ + 313, + 326, + 555, + 601 + ], + "type": "text", + "content": " reduction in MSE and " + }, + { + "bbox": [ + 313, + 326, + 555, + 601 + ], + "type": "inline_equation", + "content": "12.2\\%" + }, + { + "bbox": [ + 313, + 326, + 555, + 601 + ], + "type": "text", + "content": " improvement in PCK. While DAS and StableKeypoints achieve strong results in their respective tasks, they remain constrained by their task-specific nature. In contrast, our method not only delivers superior performance but also generalizes across diverse tasks without requiring any modification in the inference. Furthermore, unlike these methods, which derive predictions from intermediate feature or attention maps, our model directly utilizes the generative process to produce outputs, making it inherently more flexible and effective." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 611, + 408, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 408, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 408, + 624 + ], + "type": "text", + "content": "3.3. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": "In this section, we present the details of various ablations conducted to demonstrate the significance of key design choices. Specifically, we evaluate the effectiveness of implicitly-weighted prompt ensembling and examine the impact of number of prompts on performance. For additional ablations, including alternative attention formulations and sensitivity analyses on hyperparameters such as" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23610" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 95, + 70, + 254, + 120 + ], + "blocks": [ + { + "bbox": [ + 95, + 70, + 254, + 120 + ], + "lines": [ + { + "bbox": [ + 95, + 70, + 254, + 120 + ], + "spans": [ + { + "bbox": [ + 95, + 70, + 254, + 120 + ], + "type": "table", + "html": "
# PromptsEnsemblingmIoU ↑Acc. ↑
1N/A22.0874.93
5FE22.9778.84
5IWPE27.0181.75
", + "image_path": "8d11171e4926762ce0d3d66f945f8b2c6840a47e1598d27c673b6e5953d122fa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 123, + 295, + 146 + ], + "lines": [ + { + "bbox": [ + 55, + 123, + 295, + 146 + ], + "spans": [ + { + "bbox": [ + 55, + 123, + 295, + 146 + ], + "type": "text", + "content": "Table 4. Ablation of prompt-ensembling evaluated on semantic segmentation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 159, + 295, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 159, + 295, + 195 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 295, + 195 + ], + "type": "text", + "content": "attention temperature, self-attention resolution, contrastive strength, swap-guidance scale, and AdaIN, we refer the reader to Sec. F in the supplementary." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 196, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 196, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 196, + 296, + 544 + ], + "type": "text", + "content": "Effectiveness of implicitly-weighted prompt ensembling (IWPE): Here, we compare our proposed ensembling method, IWPE, to SegGPT's feature ensembling (FE) approach. For all tasks except semantic segmentation, we observe that both methods yield similar gains over the single prompt scenario. In the case of semantic segmentation, IWPE demonstrates significantly better performance. As tabulated in Tab. 4, while FE yields an absolute improvement of merely " + }, + { + "bbox": [ + 56, + 196, + 296, + 544 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 56, + 196, + 296, + 544 + ], + "type": "text", + "content": " in mIoU and " + }, + { + "bbox": [ + 56, + 196, + 296, + 544 + ], + "type": "inline_equation", + "content": "3.9\\%" + }, + { + "bbox": [ + 56, + 196, + 296, + 544 + ], + "type": "text", + "content": " in accuracy, IWPE achieves substantial gains, improving mIoU by " + }, + { + "bbox": [ + 56, + 196, + 296, + 544 + ], + "type": "inline_equation", + "content": "4.9\\%" + }, + { + "bbox": [ + 56, + 196, + 296, + 544 + ], + "type": "text", + "content": " and accuracy by " + }, + { + "bbox": [ + 56, + 196, + 296, + 544 + ], + "type": "inline_equation", + "content": "6.8\\%" + }, + { + "bbox": [ + 56, + 196, + 296, + 544 + ], + "type": "text", + "content": " over the single prompt case. This notable improvement is primarily due to the capability of our ensembling method to dynamically weigh each prompt based on its correspondence with the query image. In contrast, as discussed previously, the feature ensembling approach assumes that all prompts are equally informative, which is sub-optimal, especially in the case of tasks involving multiple classes like semantic segmentation. For instance, one prompt may provide a stronger correlation with a specific class, while another prompt, would have a weaker association with that class. Our method by implicitly weighing the influence of different prompts based on their relative correspondence to the query, better captures these nuances, resulting in a significant boost in performance compared to SegGPT's uniformly weighted feature ensembling. Therefore, while both FE and IWPE perform at par in simpler tasks where the prompts are equally informative, our method is more beneficial for tasks that involve complex multi-class structures." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "Impact of number of prompts: Here, we investigate the effect of the number of prompts on the performance and the inference speed. As illustrated in Fig. 5a, the mIoU scores improve with an increasing number of prompts. This result can be attributed to the model's ability to effectively leverage multiple prompts to better infer the task and establish the correspondence between the query and the prompts. The additional prompts help reduce ambiguities that may arise from using a single prompt. However, these gains come at the cost of reduced inference speed; specifically using five prompts cuts the inference speed to approximately half that of a single prompt. Besides the prompts, the number of denoising steps in the Stable Diffusion model is a key factor impacting the inference speed. This led us to the question" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 320, + 72, + 432, + 157 + ], + "blocks": [ + { + "bbox": [ + 320, + 72, + 432, + 157 + ], + "lines": [ + { + "bbox": [ + 320, + 72, + 432, + 157 + ], + "spans": [ + { + "bbox": [ + 320, + 72, + 432, + 157 + ], + "type": "image", + "image_path": "95a3b0d67bc9aea035eea773115a6a9bf36187125884c0e058a51fd27b512476.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 161, + 382, + 171 + ], + "lines": [ + { + "bbox": [ + 371, + 161, + 382, + 171 + ], + "spans": [ + { + "bbox": [ + 371, + 161, + 382, + 171 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 434, + 69, + 558, + 159 + ], + "blocks": [ + { + "bbox": [ + 434, + 69, + 558, + 159 + ], + "lines": [ + { + "bbox": [ + 434, + 69, + 558, + 159 + ], + "spans": [ + { + "bbox": [ + 434, + 69, + 558, + 159 + ], + "type": "image", + "image_path": "9ae8dd8c3a21a68e05c6107bf3ecaf5e8e4601bb5f2d51018995c631547dbf85.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 491, + 161, + 501, + 170 + ], + "lines": [ + { + "bbox": [ + 491, + 161, + 501, + 170 + ], + "spans": [ + { + "bbox": [ + 491, + 161, + 501, + 170 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 174, + 555, + 229 + ], + "lines": [ + { + "bbox": [ + 313, + 174, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 555, + 229 + ], + "type": "text", + "content": "Figure 5. Effect of the number of prompts on the performance and inference speed5. While (a) depicts the variation in mIoU and the inference speed, with the number of prompts, (b) illustrates the variation in mIoU with the number of denoising steps for both single and five prompt cases." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 242, + 556, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 242, + 556, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 556, + 410 + ], + "type": "text", + "content": "— With the availability of multiple prompts, does the model require as many denoising steps as in the single prompt case? In other words, we explored if we can trade-off a few denoising steps without impacting the performance when additional context is provided by multiple prompts. To this end, we conducted experiments where we evaluated the SD-VICL approach with different denoising steps. As shown in Fig. 5b, the mIoU for the five-prompt scenario at 30 denoising steps surpasses the maximum performance of the single-prompt scenario. Moreover, the maximum performance for the single-prompt case, achieved with 70 denoising steps, requires 38.4 seconds of inference time. In contrast, the five-prompt case achieves higher performance in just 35.9 seconds with 30 denoising steps5." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 426, + 388, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 388, + 438 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 388, + 438 + ], + "type": "text", + "content": "4. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 447, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 556, + 689 + ], + "type": "text", + "content": "In this paper, we propose a novel pipeline for visual in-context learning that explicitly incorporates the context between the query image and the prompt. Unlike existing visual in-context learning methods which rely on training or fine-tuning foundation models on curated or uncurated data to enable in-context learning, our approach is entirely training-free. Specifically, we introduce an in-place attention re-computation within the self-attention layers of an off-the-shelf Stable Diffusion model. Additionally, the proposed implicitly-weighted prompt ensembling technique facilitates effective integration of context through multiple prompts by implicitly weighing the prompts based on the relative correspondences. The versatility of the proposed method is demonstrated by its successful generalization across six different tasks, where it outperforms both V-ICL models trained on uncurated and task-related data as well as task-specific models based on Stable Diffusion. Crucially, our method achieves these results without requiring any additional training or annotated task-specific datasets, highlighting its ability to uncover and leverage the V-ICL capabilities of Stable Diffusion." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 702, + 552, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 702, + 552, + 712 + ], + "spans": [ + { + "bbox": [ + 324, + 702, + 552, + 712 + ], + "type": "text", + "content": "5The inference times are based on evaluations on a single A100 GPU." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23611" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer. In ACM SIGGRAPH 2024 Conference Papers, pages 1–12, 2024. 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 201 + ], + "type": "text", + "content": "[2] Yutong Bai, Xinyang Geng, Karttikeya Mangalam, Amir Bar, Alan L Yuille, Trevor Darrell, Jitendra Malik, and Alexei A Efros. Sequential modeling enables scalable learning for large vision models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22861-22872, 2024. 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 202, + 295, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 295, + 245 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 295, + 245 + ], + "type": "text", + "content": "[3] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei Efros. Visual prompting via image inpainting. Advances in Neural Information Processing Systems, 35:25005-25017, 2022. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 247, + 295, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 295, + 302 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 295, + 302 + ], + "type": "text", + "content": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 302, + 295, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 295, + 357 + ], + "type": "text", + "content": "[5] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560-22570, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 358, + 295, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 358, + 295, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 358, + 295, + 423 + ], + "type": "text", + "content": "[6] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 425, + 295, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 425, + 295, + 500 + ], + "spans": [ + { + "bbox": [ + 62, + 425, + 295, + 500 + ], + "type": "text", + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 502, + 295, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 295, + 545 + ], + "type": "text", + "content": "[8] Yaru Hao, Haoyu Song, Li Dong, Shaohan Huang, Zewen Chi, Wenhui Wang, Shuming Ma, and Furu Wei. Language models are general-purpose interfaces. arXiv preprint arXiv:2206.06336, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 546, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 546, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 546, + 295, + 612 + ], + "type": "text", + "content": "[9] Eric Hedlin, Gopal Sharma, Shweta Mahajan, Xingzhe He, Hossam Isack, Abhishek Kar, Helge Rhodin, Andrea Tagliasacchi, and Kwang Moo Yi. Unsupervised keypoints from pretrained diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22820-22830, 2024. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 613, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 667 + ], + "type": "text", + "content": "[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 669, + 295, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 669, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 57, + 669, + 295, + 691 + ], + "type": "text", + "content": "[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[12] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceed-" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "ings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "text", + "content": "[13] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12469-12478, 2024. 3, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 150, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 150, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 553, + 205 + ], + "type": "text", + "content": "[14] Xiang Li, Tianhan Wei, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. Fss-1000: A 1000-class dataset for few-shot segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2869–2878, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 205, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 205, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 205, + 553, + 271 + ], + "type": "text", + "content": "[15] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 272, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 326 + ], + "type": "text", + "content": "[16] Weihuang Liu, Xi Shen, Chi-Man Pun, and Xiaodong Cun. Explicit visual prompting for low-level structure segmentations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19434-19445, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 327, + 553, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 382 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 382 + ], + "type": "text", + "content": "[17] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1096-1104, 2016. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 383, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 437 + ], + "type": "text", + "content": "[18] Or Patashnik, Daniel Garibi, Idan Azuri, Hadar Averbuch-Elor, and Daniel Cohen-Or. Localizing object-level shape variations with text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23051–23061, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 437, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 502 + ], + "type": "text", + "content": "[19] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 503, + 553, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 503, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 316, + 503, + 553, + 559 + ], + "type": "text", + "content": "[20] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 559, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 613 + ], + "type": "text", + "content": "[21] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 614, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 647 + ], + "type": "text", + "content": "[22] Amirreza Shaban, Shray Bansal, Zhen Liu, Irfan Essa, and Byron Boots. One-shot learning for semantic segmentation. arXiv preprint arXiv:1709.03410, 2017. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 713 + ], + "type": "text", + "content": "[23] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 1, 6" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23612" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 653 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[24] Junjiao Tian, Lavisha Aggarwal, Andrea Colaco, Zsolt Kira, and Mar Gonzalez-Franco. Diffuse attend and segment: Unsupervised zero-shot segmentation using stable diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3554-3563, 2024. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "text", + "content": "[25] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "type": "text", + "content": "[26] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 240, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 240, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 294, + 262 + ], + "type": "text", + "content": "[27] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 294, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 318 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 318 + ], + "type": "text", + "content": "[28] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6830-6839, 2023. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 294, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 361 + ], + "type": "text", + "content": "[29] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023. 2,5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 364, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 364, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 364, + 294, + 418 + ], + "type": "text", + "content": "[30] Xinshun Wang, Zhongbin Fang, Xia Li, Xiangtai Li, Chen Chen, and Mengyuan Liu. Skeleton-in-context: Unified skeleton sequence modeling with in-context learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2436-2446, 2024." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 420, + 294, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 420, + 294, + 474 + ], + "spans": [ + { + "bbox": [ + 56, + 420, + 294, + 474 + ], + "type": "text", + "content": "[31] Zhendong Wang, Yifan Jiang, Yadong Lu, Pengcheng He, Weizhu Chen, Zhangyang Wang, Mingyuan Zhou, et al. Incontext learning unlocked for diffusion models. Advances in Neural Information Processing Systems, 36:8542-8562, 2023. 1, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 475, + 294, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 294, + 529 + ], + "type": "text", + "content": "[32] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. Emergent abilities of large language models. Transactions on Machine Learning Research, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 531, + 294, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 531, + 294, + 575 + ], + "spans": [ + { + "bbox": [ + 56, + 531, + 294, + 575 + ], + "type": "text", + "content": "[33] Jiarui Xu, Yossi Gandelsman, Amir Bar, Jianwei Yang, Jianfeng Gao, Trevor Darrell, and Xiaolong Wang. Improv: Inpainting-based multimodal prompting for computer vision tasks. arXiv preprint arXiv:2312.01771, 2023. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 576, + 294, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 576, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 576, + 294, + 609 + ], + "type": "text", + "content": "[34] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 610, + 294, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 610, + 294, + 653 + ], + "spans": [ + { + "bbox": [ + 56, + 610, + 294, + 653 + ], + "type": "text", + "content": "[35] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. What makes good examples for visual in-context learning? Advances in Neural Information Processing Systems, 36:17773-17794, 2023. 5" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23613" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Stable Score Distillation/62930789-4f50-45e0-b2ce-8d1494eced86_content_list.json b/2025/Stable Score Distillation/62930789-4f50-45e0-b2ce-8d1494eced86_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a492347e11418d73b300bd394ce8a5aa59d5a42e --- /dev/null +++ b/2025/Stable Score Distillation/62930789-4f50-45e0-b2ce-8d1494eced86_content_list.json @@ -0,0 +1,2353 @@ +[ + { + "type": "text", + "text": "Stable Score Distillation", + "text_level": 1, + "bbox": [ + 374, + 130, + 622, + 150 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haiming Zhu $^{1}$ , Yangyang Xu $^{2*}$ , Chenshu Xu $^{1}$ , Tingrui Shen $^{3}$ , Wenxi Liu $^{4}$ , Yong Du $^{5}$ , Jun Yu $^{2}$ , Shengfeng He $^{1*}$", + "bbox": [ + 256, + 179, + 738, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Singapore Management University $^{2}$ Harbin Institute of Technology (Shenzhen)", + "bbox": [ + 220, + 217, + 776, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ South China University of Technology $^{4}$ Fuzhou University $^{5}$ Ocean University of China", + "bbox": [ + 178, + 234, + 813, + 250 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/215e2790b3793d17d15859ef9997fc328f48e5c57b8796648d55ed6cb98b5ec8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 253, + 243, + 340 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/94379b3d762a60744f9d4afc109879d218b487c1a02403c497c68dad6352031b.jpg", + "image_caption": [ + "(a) Original Views" + ], + "image_footnote": [], + "bbox": [ + 94, + 340, + 238, + 425 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cc09d0cc8b5fac1f29b5496fddb5464f1e45bcdda3d2c9bca51cdbe82300ee06.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 245, + 255, + 442, + 340 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f83feac7a9b6dbc14a1b6005bc55d6e14b6e47965ba8113917b44c124f3e85fb.jpg", + "image_caption": [ + "(b) CSD[52]" + ], + "image_footnote": [], + "bbox": [ + 245, + 340, + 455, + 424 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b69aca6b82b03c70b0636af851357aafc4d4644a62e8db6343325dc46726ffbd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 460, + 255, + 656, + 340 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/12492b653667c289aac1cee6e4554b4d08a8b414bbfc37ca5e0cd529fd795f9c.jpg", + "image_caption": [ + "(c) DDS[10]", + "Figure 1. We propose Stable Score Distillation (SSD), a method that improves text-guided editing by preserving original content structure and enhancing realism in edited results. The first column shows the original views, while the remaining columns display comparative results with existing score distillation methods on the SD model. SSD demonstrates superior performance, maintaining the integrity of the original content and producing more realistic edits." + ], + "image_footnote": [], + "bbox": [ + 460, + 340, + 656, + 424 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/58b1d4ad15639ae1d16011af1c25387c59da9c38aa0ab9a43a7f9145b9405fa9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 255, + 897, + 340 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c9e7d6162ad4fe9abefbc17693f4d165e0c6bd0bb674c6968388be0193d2ddc9.jpg", + "image_caption": [ + "(d) Ours" + ], + "image_footnote": [], + "bbox": [ + 676, + 340, + 895, + 424 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 518, + 326, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-guided image and 3D editing have advanced with diffusion-based models, yet methods like Delta Denoising Score often struggle with stability, spatial control, and editing strength. These limitations stem from reliance on complex auxiliary structures, which introduce conflicting optimization signals and restrict precise, localized edits. We introduce Stable Score Distillation (SSD), a streamlined framework that enhances stability and alignment in the editing process by anchoring a single classifier to the source prompt. Specifically, SSD utilizes Classifier-Free Guidance (CFG) equation to achieve cross-prompt alignment, and introduces a constant term null-text branch to stabilize the optimization process. This approach preserves the original content's structure and ensures that editing trajectories are closely aligned with the source prompt, enabling smooth, prompt-specific modifications while maintaining coherence in surrounding regions. Additionally, SSD incorporates a prompt enhancement branch to boost editing strength, particularly for style transformations. Our method achieves state-of-the-art results in 2D and 3D editing tasks, including NeRF and text-driven style edits, with faster conver", + "bbox": [ + 86, + 549, + 485, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "gence and reduced complexity, providing a robust and efficient solution for text-guided editing. Code is available at: https://github.com/Alex-Zhu1/SSD.", + "bbox": [ + 511, + 520, + 908, + 565 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 587, + 645, + 602 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-based image generation has achieved remarkable progress, particularly with the advent of diffusion models [13, 28, 40, 41, 43, 56]. These models leverage strong priors to produce high-quality images, facilitating significant advances in text-to-3D generation [37, 47, 58]. Moreover, text-guided 3D editing has enabled intricate modifications to shape and texture [26, 31], supporting flexible and precise 3D scene manipulation.", + "bbox": [ + 509, + 612, + 906, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unlike generation tasks that create new content, editing tasks aim to modify specific elements within an image while preserving surrounding areas [51]. However, directly applying methods like Score Distillation Sampling (SDS) to editing tasks can yield undesired effects, such as blurring across the image. This arises because SDS optimizes globally to the prompt, affecting regions beyond the targeted area [10, 20]. DDS [10] addresses this by introducing a dual-branch architecture, pairing the source image with its description to leverage the model's inherent bias and isolate specific prompt changes. Further, CSD [52] achieved", + "bbox": [ + 509, + 734, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors: Yangyang Xu (xuyangyang@hit.edu.cn) and Shengfeng He (shengfenghe@smu.edu.sg).", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16597", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/113e4cf3086202d6189ea23ab0e6c94a7fef3b378152c624fb16ea51a380b2b7.jpg", + "image_caption": [ + "Figure 2. Illustration of three distillation-based approaches. Note that we assume a 2-step optimization process for illustration, where the subscript $t$ represents the iteration number. DDS utilizes the source branch to obtain initial latent $Z_0^*$ , while CSD employs two classifiers to derive $Z_1'$ and $Z_1''$ for cross-prompt editing. Our SSD method designs a CFG classifier to determine the cross-prompt editing, introduces the null-text branch as the initial latent $Z_0^*$ , and further constructs the cross-trajectory term (see Sec. 4.1) for stable optimization." + ], + "image_footnote": [], + "bbox": [ + 106, + 93, + 321, + 213 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c5589cd515662891cef249906174257aa7b3542a2aa1bf446dcc8e922b46d715.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 328, + 93, + 545, + 214 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1b693ad852678cb0b00bdc1dd1b74f854c74679f7328eae1298787e3f4cc59b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 93, + 769, + 214 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dad6abc620176c768dfa99228a5f05c5ab73f7b654a77aa876f396f92fc6ba21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 93, + 890, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/479da4509d074ccb14bce7ce3df49b68e806012d1bbbe6aecabfaea49663ba2a.jpg", + "image_caption": [ + "Figure 3. The optimization process of DDS and our SSD. SSD preserves the source structure effectively during optimization iterations, while DDS cannot preserve it effectively." + ], + "image_footnote": [], + "bbox": [ + 89, + 287, + 490, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "scene editing by incorporating a classifier component within Classifier-Free Guidance (CFG) [12] to refine the prediction score by applying the classifier to both the source and target prompts. NFSD [20] further decomposes the CFG score, highlighting the classifier as the primary driver of prompt direction.", + "bbox": [ + 89, + 517, + 483, + 607 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite their success, we argue that current distillation-based approaches face inherent limitations, such as low editing quality and loss of source content. As shown in Fig. 2, DDS [10] relies on the source branch to remove model bias, but lacks the explicit guidance to preserve the source content [24] during optimization. As shown in Fig. 1c, DDS changes man's clothes during editing his faces. Additionally, although introducing source prompt components is intended to improve prompt specificity [19], it can amplify noise and introduce overlapping objectives that hinder stable convergence. This results in artifacts or unintended variations, especially in the unedited regions. Correspondingly, CSD [52] utilizes dual classifiers to refine the prompt editing direction (see Fig. 2). However, it lacks the explicit source preservation to restrict edits precisely to the target areas. As shown in Fig. 1b, this causes the structure deformation and annoying artifacts in the edited regions.", + "bbox": [ + 89, + 609, + 483, + 866 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our insights into these limitations lead to two key observations: (1) Cross-prompt: a single classifier, providing", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the editing direction from source prompt to target, and (2) Cross-trajectory: stability in the editing process can be achieved by aligning the editing direction closely with the structure of the source content.", + "bbox": [ + 511, + 292, + 906, + 351 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose Stable Score Distillation (SSD), a streamlined approach for stable and precise text-guided editing. To achieve a smooth editing direction, we employ the CFG equation for both the source and target prompts, ensuring a gradual transition of the original contextual texture as the model adapts to the specified changes. This approach contrasts with DDS [10], as it eliminates the need for a auxiliary source branch, enabling our method to focus editing gradients precisely within target regions while ensuring a stable transition, as illustrated in Fig. 3. Moreover, for aligning the editing direction with the source prompt, facilitating smoother and more controlled progression toward the target prompt, we design an cross-trajectory strategy to ensure that edits respect the original structure, supporting subtle and stable transformations within designated areas. While NFSD [20] utilizes negative-branch and DDS utilizes source branch to enhance output clarity, as shown in SSD in Fig. 2, we introduce a null-text branch aligned with the \"no-edit\" direction to integrate a \"reconstruction\" term to explicitly enforce source content preservation, which enhances consistency and produces reliable edits across diverse tasks. Based on above designs, our framework remains streamlined and efficient, achieving both precision and stability without the complexity of additional components.", + "bbox": [ + 511, + 354, + 908, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our framework integrates seamlessly into existing DDS-based editing pipelines and applications, such as text-driven NeRF editing [22, 24, 36] and 2D image editing [35]. Notably, our approach's \"clear\" editing direction preserves source content, making a carefully designed identity regularization [22] unnecessary. Moreover, standard DDS methods often lack sufficient editing strength, resulting in minimal or negligible changes in output, particularly in style editing [23]. Our approach, with its streamlined and stable framework, allows for the seamless integration of a prompt enhancement branch to amplify editing capability.", + "bbox": [ + 511, + 717, + 908, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With these improvements, our method achieves faster and", + "bbox": [ + 532, + 885, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16598", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "more effective edits during optimization, remains compatible with the Stable Diffusion Model [40] without requiring LoRA [14] or fine-tuning, and integrates effectively with Instructpix2pix [2]. Additionally, by incorporating nonincreasing timestep sampling [15], we accelerate convergence, reducing the required iterations to approximately 3,000 for NeRF [33] and 1,500 for Gaussian splatting [21]. In summary, our contributions are as follows:", + "bbox": [ + 89, + 90, + 485, + 210 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a novel editing framework, Stable Score Distillation, that leverages a single, anchored classifier to achieve targeted and stable edits in 3D scene editing.", + "- We introduce a prompt enhancement strategy, effectively improve the prompt-alignment, especially style editing in 2D-image editing.", + "- We demonstrate the effectiveness of our approach across NeRF-editing and image-editing tasks, achieving state-of-the-art results with a streamlined and efficient framework." + ], + "bbox": [ + 89, + 213, + 480, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 364, + 230, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Diffusion Models", + "text_level": 1, + "bbox": [ + 89, + 388, + 258, + 404 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models [13, 39, 43, 53] have made significant advancements in generating diverse and high-fidelity images. Starting from a gaussian noise, diffusion models can predict the noise-less sample at each time step, until finally obtaining clear samples. Commonly, the denoising process can utilize U-net model to predict the noise. Some works [13, 44] have observed that is proportional to the predicted score function [17] of the smoothed density. Thus, intuitively, taking steps in the direction of the score function gradually moves the sample towards the data distribution.", + "bbox": [ + 89, + 412, + 483, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To generate images aligned with a target prompt, guidance is typically introduced to explicitly control the weight assigned to the conditioning information. The popular guidance methods include Classifier Guidance [8] and Class-free Guidance (CFG) [12]. While the former relies on a separately learned classifier, the latter directly introduces null-text samples to the model. CFG modifies the score function to steer the process towards regions with a higher ratio of conditional density to the unconditional one. However, it has been observed that CFG trades sample fidelity for diversity [12]. Based on the insights gained from the decomposition of the CFG equation, we propose a novel Stable Score Distillation (SSD) method to guide the SDS optimization process.", + "bbox": [ + 89, + 565, + 483, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Score Distillation Sampling (SDS)", + "text_level": 1, + "bbox": [ + 89, + 771, + 383, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Benefit from the data scale-law, diffusion model [38, 40, 41] achieve high-quality image generation and text-to-image generation. Specifically, Score Distillation Sampling (SDS) [37] leveraging the priors of pre-trained text-to-image models to facilitate text-conditioned generation in 3D content generation. Specifically, SDS is an optimization approach that updates the rendering parameter towards the image distribu", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion of diffusion models by enforcing the noise prediction on noisy rendered images to match sampled noise. While SDS provides an elegant mechanism for leveraging pretrained text-to-image models, SDS-generated results often suffer from oversaturation and lack of fine realistic details. VSD [47] proposed a particle-based optimization framework that treats the 3D parameter as a random variable of target distribution. Furthermore, by regarding SDS as a reverse diffusion process, decreasing timesteps sampling [15, 58] to imitate the diffusion reverse sampling, which can improve the quality of the generated 3D assets.", + "bbox": [ + 511, + 90, + 903, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In image editing, Delta Denoising Score (DDS) [10] found that Score Distillation Sampling (SDS) introduces noticeable artifacts and over-smoothing in edited images due to inherent bias. To mitigate this bias, DDS employs a subtraction of two SDS scores of the source and target images to obtain a delta score, which is then used to guide the optimization process.", + "bbox": [ + 511, + 262, + 903, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Text-Driven 3D-Scene Editing", + "text_level": 1, + "bbox": [ + 511, + 393, + 779, + 410 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-driven 3D scene editing has been a popular research topic [18, 29, 56]. IN2N [9] proposed a Iterative Dataset Update method that can edit 3D scenes from text descriptions. By leveraging advancements in 2D diffusion editing techniques, notably InstructP2P [2] and ControlNet [54], GaussianEditor[6] and GaussCtrl [49] utilize edit multi-view images latent to optimize the 3D scene. We consider utilize score distillation to guide the 3D scene editing, which is more flexible and efficient for the text-driven 3D scene editing.", + "bbox": [ + 511, + 421, + 906, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Building upon the foundational SDS loss introduced by DreamFusion [37], some work has explored SDS loss in the text-driven 3D scene editing. RePaint-NeRF [57] has advanced the application of SDS in 3D editing by integrating a semantic mask to guide and constrain modifications within the background elements. CSD [52] utilize two classifiers to achieve editing. In a similar vein, ED-NeRF [36] has introduced an enhanced loss function specifically designed for 3D editing tasks. PDS [24] proposed a posterior distillation sampling to match stochastic latent [16]. Piva [25] fine-tuned the model while introducing a regularization term to preserve identity. Unfortunately, these methods are still limited to the long-time diffusion reverse sampling process, which is not suboptimal for the text-driven 3D scene editing. DreamCatalyst [22] extends the PDS optimization processing to ID-preserving and edit-ability based on decreasing timesteps sampling.", + "bbox": [ + 511, + 578, + 908, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Different from the above methods, ours firstly improve the DDS optimization process by introducing a single classifier, and further introduce a null-text branch to achieve a more stable and precise editing process.", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16599", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1c65990c7e27298d4db995adee328213b5e3471bf049410d212daf798bf05600.jpg", + "image_caption": [ + "Figure 4. The overview of SSD. Given the parameter 3D-model or image, SSD provides effective editing gradient to guide the optimization process. We utilize CFG equation between the predicted target noise $\\epsilon_{\\phi}(z_t,y,t)$ and source noise $\\epsilon_{\\phi}(z_t,\\hat{y},t)$ , which generate the gradual editing direction. Furthermore, we introduce a null-text branch $\\epsilon_{\\phi}(\\hat{z}_t,\\emptyset ,t)$ to regularize the optimization process and achieve stable optimization. We further analyzing and decompose ours design term into three parts: cross-prompt, cross-trajectory, and prompt-enhance." + ], + "image_footnote": [], + "bbox": [ + 116, + 89, + 880, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 414, + 217, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we first discuss existing optimization-based approaches to handle parametric images. Then, we will introduce our novel parametric image editing method in Section 4.", + "bbox": [ + 89, + 439, + 483, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Score Distillation Sampling", + "text_level": 1, + "bbox": [ + 89, + 508, + 336, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Score Distillation Sampling (SDS) [37] is proposed to generate parametric images by leveraging the 2D prior of pretrained text-to-image diffusion models. Specifically, given a pretrained diffusion model $\\epsilon_{\\phi}$ , SDS optimizes a set of parameters $\\theta$ of a differentiable parametric image generator $g$ , using the gradient of the loss $L_{\\mathrm{SDS}}$ with respect to $\\theta$ :", + "bbox": [ + 89, + 531, + 485, + 625 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\theta} L _ {\\mathrm {S D S}} = w (t) \\left(\\epsilon_ {\\phi} \\left(z _ {t} (x); y, t\\right) - \\epsilon\\right) \\frac {\\partial x}{\\partial \\theta}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 632, + 483, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x = g(\\theta)$ is an image rendered by $\\theta$ , $z_{t}(x)$ is obtained by adding a Gaussian noise $\\epsilon$ to $x$ corresponding to the $t$ -th timestep of the diffusion process, and $y$ is a condition to the diffusion model. As Noise-Free Score Distillation (NFSD) [20] has shown, the score $\\epsilon_{\\phi}(z_{t}(x);y,t)$ provides the direction in which this noised version of $x$ should be moved towards a denser region in the distribution of real images.", + "bbox": [ + 89, + 672, + 485, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Delta Distillation Sampling", + "text_level": 1, + "bbox": [ + 89, + 801, + 336, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although SDS get excellent generation ability, for editing task, an undesired component from the pretrained model, $\\delta_{\\mathrm{bias}}$ , interferes with the process and causes the image to become smooth and blurry in some parts [10]. Based on the observations that a matched source prompt $\\hat{y}$ and source", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "latent $\\hat{z}_t$ can estimate the noisy direction $\\delta_{\\mathrm{bias}}$ , thus, the DDS method aims to remove the $\\delta_{\\mathrm{bias}}$ by introducing source branch, as shown in Eq. 2:", + "bbox": [ + 511, + 415, + 906, + 460 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\theta} L _ {\\mathrm {D D S}} = \\left(\\epsilon_ {\\phi} ^ {c} \\left(z _ {t}, y, t\\right) - \\epsilon_ {\\phi} ^ {c} \\left(\\hat {z} _ {t}, \\hat {y}, t\\right)\\right) \\frac {\\partial z}{\\partial \\theta}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 465, + 906, + 496 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\phi}^{c}(z_{t},y,t)$ and $\\epsilon_{\\phi}^{c}(\\hat{z}_{t},\\hat{y},t)$ are pretrained model predictions $\\epsilon$ , with the superscript $c$ indicating the CFG results. Thus, DDS pushes the optimized image into the direction of the target prompt without the interference of the noise component, namely, $\\nabla_{\\theta}L_{\\mathrm{DDS}}\\approx \\delta_{\\mathrm{ext}}$ . Obviously, $\\nabla_{\\delta_{\\mathrm{ext}}}$ is contingent on classifier part from $\\epsilon_{\\phi}^{c}(z_{t},y,t)$ as discussed in CSD [52] and NFSD [20]. Note that in the following manuscript, we decompose the CFG results without the superscript $c$ and omit the timestep $t$ for simplicity.", + "bbox": [ + 511, + 500, + 908, + 637 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Further exploring prompt editing direction, CSD [52] method proposed a dual-classifier to refine the editing score and achieve more precise editing, as shown in Eq. 3:", + "bbox": [ + 511, + 637, + 906, + 681 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {\\theta} L _ {\\mathrm {C S D}} = \\left(w _ {a} \\left(\\epsilon_ {\\phi} \\left(z _ {t}, y\\right) - \\epsilon_ {\\phi} \\left(z _ {t}, \\varnothing\\right)\\right) \\right. \\\\ \\left. - w _ {b} \\left(\\epsilon_ {\\phi} \\left(z _ {t}, \\hat {y}\\right) - \\epsilon_ {\\phi} \\left(z _ {t}, \\emptyset\\right)\\right)\\right) \\frac {\\partial z}{\\partial \\theta}, \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 688, + 900, + 738 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "while the $\\epsilon_{\\phi}(z_t,y,t)$ and $\\epsilon_{\\phi}(z_t,\\hat{y},t)$ are current latent $z_{t}$ predictions for the target prompt $y$ and source prompt $\\hat{y}$ , respectively. $w_{a}$ and $w_{b}$ are weights of classifiers. Simply put, CSD aims to refine the prompt editing direction by determining the difference between two classifiers, which can be regarded as a cross-prompt term.", + "bbox": [ + 511, + 742, + 908, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 511, + 844, + 604, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In 3D scene editing process, which requires consideration of both the target prompt and the original source content,", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16600", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we consider two key aspects: (1) smooth editing direction towards the target prompt (2) and editing results respect the original structure. Based on these, in this section, we introduce our novel editing framework Stable Score Distillation.", + "bbox": [ + 89, + 90, + 485, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Stable Score Distillation", + "text_level": 1, + "bbox": [ + 89, + 162, + 312, + 178 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Firstly, we introduce the design of a cross-prompt editing direction. As discussed about CSD method in Sec. 3.2, the key role in cross-prompt editing is to provide a smooth transition from the source prompt to the target. As the CFG guidance [12] steers the process towards regions with a higher ratio of conditional density to the unconditional one, accordingly, we can modify the SDS score function, as shown below:", + "bbox": [ + 89, + 186, + 483, + 306 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nG r a d = \\epsilon_ {\\phi} (z _ {t}, \\hat {y}) + s (\\epsilon_ {\\phi} (z _ {t}, y) - \\epsilon_ {\\phi} (z _ {t}, \\hat {y})), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 321, + 483, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\phi}(z_t,y)$ and $\\epsilon_{\\phi}(z_t,\\hat{y})$ are pretrained model predictions. The scale factor $s$ is equal to control weight.", + "bbox": [ + 89, + 352, + 483, + 382 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although the cross-prompt term provides a smooth texture transition in the edited region, we observed that the optimization process leads to abrupt structural changes, often resulting in artifacts and unappealing outcomes, similar to CSD in Fig. 1b. To address this, we introduce an additional regularization term to constrain the structural transition. Interestingly, as shown in Fig. 1c, DDS achieves better results than CSD by incorporating a source branch. However, DDS still lacks a mechanism to ensure the original structure remains intact, leading to modification on unedited regions. To address this, we introduce a null-text branch $\\epsilon_{\\phi}(\\hat{z}_t,\\emptyset)$ to regularize the optimization process, as shown in Eq. 5:", + "bbox": [ + 89, + 383, + 483, + 565 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {s s d}} = \\epsilon_ {\\phi} (z _ {t}, \\hat {y}) + s \\left(\\epsilon_ {\\phi} (z _ {t}, y) - \\epsilon_ {\\phi} (z _ {t}, \\hat {y})\\right) - \\epsilon_ {\\phi} (\\hat {z} _ {t}, \\varnothing). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 578, + 483, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Eq. 5 is ours Stable Score Distillation, and we can further decompose above equation into two parts, and the latter is regarded as a cross-trajectory term.", + "bbox": [ + 89, + 609, + 483, + 655 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {s s d}} = \\underbrace {w _ {p} \\left(\\epsilon_ {\\phi} \\left(z _ {t} , y\\right) - \\epsilon_ {\\phi} \\left(z _ {t} , \\hat {y}\\right)\\right)} _ {\\text {c r o s s - p r o m p t}} + \\underbrace {w _ {t} \\left(\\epsilon_ {\\phi} \\left(z _ {t} , \\hat {y}\\right) - \\epsilon_ {\\phi} \\left(\\hat {z} _ {t} , \\emptyset\\right)\\right)} _ {\\text {c r o s s - t r a j e c t o r y}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 667, + 504, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the $w_{t}$ and $w_{p}$ control the strength of the cross-trajectory and cross-prompt, respectively.", + "bbox": [ + 89, + 719, + 483, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The cross-trajectory term can be interpreted as the distance between the transitions of two latents, ensuring that the original structure remains smooth and does not change abruptly (more details are provided in the supplementary material). In Fig. 5, we can see that the cross-trajectory term can provide a strong structure constraint ability, guiding the optimization process to preserve the source image structure. Specifically, when set $w_{t} = 0$ , the optimization process behaves similarly to the CSD[52] method, which fails to retain the original image structure.", + "bbox": [ + 89, + 750, + 485, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2d730c9f07c9dab167114f14a75e98f13ac70892f2ebebb8bf73cce3055a8358.jpg", + "image_caption": [ + "Figure 5. The effect of increasing the strength of the prompt enhancement term $(w_{e})$ and cross-trajectory term $(w_{t})$ , with the cross-prompt term fixed at 7.5. Both terms contribute to prompt-aligned results, while setting $(w_{t} = 0)$ leads to saturation and discard source content." + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 903, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Improving Prompt Alignment", + "text_level": 1, + "bbox": [ + 511, + 416, + 782, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although Eq. 5 can achieve gradual editing results, we found Eq. 5 have similar limitation with DDS [7], which have insufficient editing strength. The editing results neither get successful editing nor retain the source image structure, often leads to little or no change in the final. Benefit from the cross-prompt editing design as Eq. 5, we can add a target prompt enhancement branch to guide the optimization process. The target prompt alignment branch will provide the direction of the target prompt, as shown in Eq. 7:", + "bbox": [ + 511, + 438, + 906, + 575 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {a l i g n}} = w _ {e} \\left(\\epsilon_ {\\phi} \\left(z _ {t}, y\\right) - \\epsilon_ {\\phi} \\left(z _ {t}, \\varnothing\\right)\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 588, + 584, + 906, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w_{e}$ is the prompt enhancement scale. As shown in Fig. 5, the synchronous scaling of both the cross-trajectory and prompt-enhancement terms results in effective visual editing outcomes.", + "bbox": [ + 511, + 609, + 905, + 669 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Source Latent Regularization", + "text_level": 1, + "bbox": [ + 511, + 678, + 777, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Empirically, we found that directly using latent-space loss rather than pixel-level loss can lead to optimization difficulties in local regions of 3DGS. For example, the bright spots appearing in Fig. 5. To suppress the steep gradients in these areas, we incorporate ID regularization to guide the stable optimization process. Differ with PDS [24] use source latent $\\hat{x}_0$ , we can use the noisy latent $\\hat{x}_t$ to avoid partial exploding gradient, as shown in Eq. 8:", + "bbox": [ + 511, + 700, + 906, + 820 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {I D}} = w (t) \\cdot \\left(x _ {t} - \\hat {x} _ {t}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 629, + 830, + 906, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the $w(t)$ is the iteration-dependent strength, designed as a decreasing function of $t$ . Notably, the $w(t)$ is not necessary to well-designed in our design. Our final loss function", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16601", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "as shown in Eq. 9:", + "bbox": [ + 89, + 90, + 217, + 104 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {f i n a l}} = L _ {\\text {s s d}} + L _ {\\text {a l i g n}} + L _ {\\text {I D}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 119, + 482, + 135 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Based on the above design, we achieve a more prompt-aligned editing method, which integrates seamlessly into the Stable Diffusion Model [40] without requiring LoRA [14] or fine-tuning. Moreover, we will further introduce our method's connection with InstructPix2Pix [2].", + "bbox": [ + 89, + 148, + 483, + 224 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Connecting with IP2P", + "text_level": 1, + "bbox": [ + 89, + 236, + 297, + 253 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The final design of our method is shown in Eq 9. We found that ours edit gard provide new angle to understand about InstructP2P [2] one-step reverse sampling.", + "bbox": [ + 89, + 258, + 483, + 305 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\epsilon_ {\\theta} \\left(z _ {t}, c _ {I}, c _ {T}\\right) = \\epsilon_ {\\theta} \\left(z _ {t}, \\varnothing , \\varnothing\\right) \\\\ + s _ {I} \\left(\\epsilon_ {\\theta} \\left(z _ {t}, c _ {I}, \\varnothing\\right) - \\epsilon_ {\\theta} \\left(z _ {t}, \\varnothing , \\varnothing\\right)\\right) \\\\ + s _ {T} \\left(\\epsilon_ {\\theta} \\left(z _ {t}, c _ {I}, c _ {T}\\right) - \\epsilon_ {\\theta} \\left(z _ {t}, c _ {I}, \\varnothing\\right)\\right), \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 316, + 478, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $c_{I}$ and $c_{T}$ are input-image and instruction prompt separately, $s_{I}$ and $s_{T}$ are the source image control and instruction prompt control strength. The Eq. 10 is the InstructP2P one-step reverse sampling, which can provide the direction of the target prompt. We can see that the InstructP2P is the simple version of our method, the middle term of Eq. 10 is cross-trajectory regularization, and the last term of Eq. 10 is ours cross-prompt term. Simply put, as analyzing the Eq. 5, subtracting the constant correction term $\\epsilon_{\\theta}(\\hat{z}_t; \\varnothing; \\varnothing)$ is edit grad. Ours method reveals that apply DDS loss in the InstructP2P model can only editing branch, and don't have to provide the source branch.", + "bbox": [ + 89, + 402, + 483, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 598, + 223, + 616 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we conduct editing experiments across two types of parameterized images. Section 5.1 evaluates the effectiveness of our method on 3D Scenes Editing, and Section 5.2 evaluates the effectiveness of our method on 2D Image Editing. We also conduct ablation studies to analyze the effectiveness of ours components in Section 5.3.", + "bbox": [ + 89, + 625, + 483, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1.3D Scenes Editing", + "text_level": 1, + "bbox": [ + 89, + 726, + 266, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. To evaluate the effectiveness of our method, we conduct experiments on the scenes from IN2N [9] and other real-world datasets, including LLFF [32] and Mip-Nerf360 [1].", + "bbox": [ + 89, + 750, + 483, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our method with several state-of-the-art inversion methods. We use 3DGS [21] as the 3D representation, and compare our method with InstructNerf2Nerf [9], DDS [10], GS-Edit [6], and DGE [5]. For fairness, we implement the DDS version based on the official GS-Edit code. PDS [24] is designed for addition of objects to unspecified", + "bbox": [ + 89, + 810, + 485, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/65f60e7560a5debb561683f4d96607bafb4cde437da3657fcef787a51a42b0fc.jpg", + "table_caption": [ + "Table 1. Quantitative evaluations under 3D editing scenes." + ], + "table_footnote": [], + "table_body": "
MethodCLIP Sim.↑Sim Dir.↑User Study ↑
IN2N0.16760.070714.54%
DDS0.17800.04015.45%
GS-Editor0.17580.042914.54%
DGE0.17580.056323.63%
Ours0.18460.077341.81%
", + "bbox": [ + 553, + 104, + 859, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "regions, we will provide the comparison results in supplementary material.", + "bbox": [ + 511, + 196, + 906, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We follow common practice [5, 6, 9] to evaluate the effectiveness of our method. CLIP Similarity is to evaluate the alignment between the render images and the target prompts, i.e., the cosine similarity between the text and image embeddings encoded by CLIP. Specifically, follow DGE [5], randomly sample 20 camera poses to evaluate. CLIP Directional Similarity is to measure the editing effect, i.e., the cosine similarity between the image and text editing directions (target embeddings minus source embeddings). We evaluate all methods on 6 different scenes and 10 different prompts.", + "bbox": [ + 511, + 228, + 906, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results. We begin by evaluating our method, starting with a qualitative assessment. In Fig. 6, we present a comparison of results with competing methods. Our approach generates more visually appealing images that are better aligned with the editing instructions. In contrast, methods based on the Iterative Dataset Update (IDU) strategy, such as IN2N [9] and GS-Editor [6], fail to produce the desired editing outcomes, resulting in blurrier or lower-fidelity reconstructions and noticeable artifacts. For example, in the scene of \"Spider-Man with a mask\", IN2N generates a mask with reduced fidelity, while GS-Editor produces a low-detail mask. In the multi-view consistency setup, DGE [5] performs well on common attributes but is constrained to \"rainbow\" editing and tends to generate artifacts outside the segmentation mask. Our method works seamlessly with masks, producing results with rich details.", + "bbox": [ + 511, + 395, + 908, + 635 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Tab. 1, We present a quantitative comparison. Our method outperforms the baselines in terms of CLIP Similarity and CLIP Directional Similarity. Notably, Dire Sim is not sensitive with the editing quality, much focus on the instruction attributes. We conducted a user study with a survey of 55 participants to evaluate the editing quality. The results show that our method received the most popular votes.", + "bbox": [ + 511, + 638, + 908, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. 2D Image Editing", + "text_level": 1, + "bbox": [ + 511, + 755, + 687, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. To evaluate the effectiveness of our method, we conduct experiments on the PIE-Bench dataset proposed by PNPIv [19], which consists of 700 images with 9 editing types. Each image is annotated with the source and target prompts.", + "bbox": [ + 511, + 777, + 906, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our method with several classical editing methods based on DDIM [43] inversion, including P2P [11], PNP [46] and MasaCtrl [3]. For optimization-", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16602", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/13ce1d72ac22be6ba8aaad7589261cefbec4998271c9b205f0bb383fe9db2855.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 104, + 250, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f2805ad1f9a08b27f8a067ef9505fff4aad4faacd96bb493599b24da55750bc8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 256, + 103, + 403, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/35a59f56e4c89d5ff0a5cd0b0f38718f9b1613950998792f6b07be31b594739c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 103, + 560, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f9cedc028a8c2f06f38b8fd7bb2eee568a7f3d437ab0c979f75559f4afed319b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 570, + 103, + 715, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/878a0b066fa3d1b0f5600b04680a85667a95f9fce37bf619173f7e6750a5472b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 723, + 104, + 872, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0bd1a105b87c004a30efa34371eeaf0d431a6e913e790f2dc689c63a78c01e1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 218, + 251, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6e56458c83a5e5663d8cba8b7d62b6ceca5b4909a68f04cb7c5ee27c331b792c.jpg", + "image_caption": [ + "A man wearing T-shirt with a pineapple pattern" + ], + "image_footnote": [], + "bbox": [ + 251, + 218, + 408, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/07bfd5534b16378060d25b3b2e02843f14eeb98b8e1f139bfab069accea1a362.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 218, + 563, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d846415dd336729eeee72a7b26c7afd54e6ded75b8f11cc135cb1e7c7410e36f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 565, + 218, + 720, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8da663ffc801f6d2c6aa9c769369662d88bb36f0a503105e0c7923847be8f956.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 218, + 877, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a1f6047a6c410edb2977679a0dd1d46fc3c84fe44b9194420f848cb970991b8e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 340, + 251, + 431 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e20c01a14c58cf70b717178cb863b1b7e64edf64dec5eb4e5eb6d28ef4aca5f8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 251, + 340, + 408, + 431 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f7b88149ae0d3f96d9139d4b733219f7502d881a0888483373573daef3935c90.jpg", + "image_caption": [ + "Rainbow horn fossil" + ], + "image_footnote": [], + "bbox": [ + 408, + 340, + 563, + 431 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c7cb61146101236bca3aac01285297c694df1bdcb1cf9f604bfc7713e409c9a2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 565, + 340, + 720, + 431 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/25227a4b1e6df7d5e6be676a438e4bd74903c1c649de2c99b23dafe131223ea8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 340, + 877, + 431 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/18b19580bde98279c991ce488677c952c1566e7b4919d77ef6e2daa542bc7d2a.jpg", + "image_caption": [ + "Input Views", + "Figure 6. Qualitative comparisons with related works. SDS demonstrates outstanding performance in effectively preserve source structure in the modified region." + ], + "image_footnote": [], + "bbox": [ + 96, + 446, + 251, + 527 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2d130333d4f3face37b0455d5e73cf8928624983748be7a9bc200e89ec888418.jpg", + "image_caption": [ + "A tree stump with some leaves on fire", + "InstructN2N" + ], + "image_footnote": [], + "bbox": [ + 251, + 446, + 408, + 527 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f40b944de1aed3a3a1b649e778e674236769dec39541cb27e4306f0ac65e3458.jpg", + "image_caption": [ + "GaussianEditor" + ], + "image_footnote": [], + "bbox": [ + 408, + 446, + 563, + 527 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/903d458fc60da3419c0cae26b709da119a308a7789f4201121dac0279e03ad19.jpg", + "image_caption": [ + "DGE" + ], + "image_footnote": [], + "bbox": [ + 565, + 446, + 720, + 527 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/935d95a40c02d21f1d8cdfcfc5ce9531ed9d7c5608c90f6ab0547459fea4b6aa.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 720, + 446, + 877, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "based editing method, we compare with NT [34] and StyleD [27]. Besides, we report the comparison with DT [19]. Further, we compare with DDS [10] and its extended method CDS [35].", + "bbox": [ + 89, + 579, + 482, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Metrics. We follow DT [19] which uses several metrics to evaluate our method. We use the Structure Distance assessed by DINO score [4] to evaluate the structure distance between original and edited images. We also introduce several metrics to evaluating the background preservation, which includes LPIPS [55] and MSE. Besides, we introduce CLIP Similarity [48] to evaluate the text-image consistency between edited images and corresponding target editing text prompts.", + "bbox": [ + 89, + 641, + 483, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. We present a qualitative comparison of our method with competitors in Fig. 7. Our method generates images that are more aligned with the target prompts and preserve the source structure. In \"blue butterfly\", ours successfully changes the color of the butterfly to blue, while DDS [10] and CDS [35] generate similar color from source. Especially, ours method successfully changes the style of the image and generates appealing results, which is challenging for DDS-", + "bbox": [ + 89, + 779, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a2097ad2573be0bb9a7dc491a9722a9c285d2ca961f2adcac4f0b6d2686c2428.jpg", + "table_caption": [ + "Table 2. Quantitative evaluation in PIE-Bench dataset." + ], + "table_footnote": [], + "table_body": "
MethodDistance×103↓LPIPS×103↓MSE×104↓CLIP↑
DDIM + P2P69.43208.80219.8825.01
DDIM + PNP28.22113.4683.6425.41
DDIM + MasaCtrl28.38106.6286.9723.96
NT + P2P13.4460.6735.8624.75
StyleD + P2P11.6566.1038.6324.78
DT + P2P11.6554.5532.8625.02
DDS14.7450.5845.0925.86
DDS + CDS7.1533.1425.2924.96
Ours28.1382.4386.6426.94
Ours + CDS6.9032.1524.2125.12
", + "bbox": [ + 513, + 590, + 919, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "based methods. Compared to optimization-based methods, NT [34] preserves the general source structure during the inversion process, but tends to discard some content as evident in the distortion of the girl's fingers in Fig. 7. Additionally, due to limitations in the editing methods, the editing results are unsuccessful.", + "bbox": [ + 511, + 761, + 906, + 851 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Tab. 2, we present a quantitative comparison. Our method strikes a balance between structure distance and editability. Notably, we observe that the distance is much", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "A photo of a man $\\rightarrow$ Spider man with a mask", + "bbox": [ + 362, + 87, + 633, + 101 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16603", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A blue and white bird $\\rightarrow$ butterfly sits on a branch", + "bbox": [ + 151, + 90, + 419, + 102 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cf61a5a6f51cf7b19f3024fd9ff691faa65d02314a8a45379864b4b8f6716f58.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 104, + 168, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/00c0fde9295dd7240d285eefe4980fad3185dc361640f4013af1364939c2fb0b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 169, + 104, + 246, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/02fdae951c4a701eb43ef5c93133d3db311c1268b309d76794f8f3f1bfe3617a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 250, + 104, + 325, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5954546facdaca4fea820fb04998d791af9fedee2c1c451396c01012d381bced.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 326, + 104, + 401, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e36d09ce7a7403e49e4865b73ecd12453b00bfbd0c912fbb787a148ada4e0f45.jpg", + "image_caption": [ + "Figure 8. Effect of source latent regularization. In most experiments, the source ID term helps prevent partial gradient explosion. In the left image, the yellow arrow highlights an irregular color. As the weight of the ID term increases, the color becomes more regular, however, the spider on the character's chest is affected." + ], + "image_footnote": [], + "bbox": [ + 403, + 104, + 480, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Kids crayon drawing of a man with a long beard and a long sword", + "bbox": [ + 109, + 164, + 464, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6844614c16f7a3153254083073477af46f2a162e8cd2b1871f19b12700d2f463.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 176, + 169, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/51f3b4407f6e9ff81716cc3792649a66a201a15cd0fc834b0f805d0538b9c5b3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 171, + 176, + 248, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bfa1737ca5a4b77528bf1ff09b1cf5a87757c8b1df837ba69eaa24f68b197eed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 250, + 176, + 325, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2b4433e62d7577275a90ed28b5e200c6c80dcaa7a94d5e709bc2a5f2d131d9d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 326, + 176, + 401, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0e712de57c892101d9095f9a7bcd6c60c57744ef59ecdab4bae094786d536f5f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 403, + 176, + 480, + 234 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Black and white sketch of a young girl with painted hands and face", + "bbox": [ + 109, + 237, + 464, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9359ff7cb352ff71f1503aa5d09b4a88e6581ee50d98942dd30bd524fd6d949f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 252, + 169, + 309 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/76395dace40a6df6c10cbf3fd0eee41fb5cd2353c620d1a0c34f7ee8645ec28d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 171, + 252, + 248, + 309 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c32718a6a8bcbb0e8d88f7671a3fd7d559862444a6a1321639464486c1371461.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 250, + 252, + 325, + 309 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/48e72db346cae2dde2118d0dc124f8cae46706aa7ff7f730f05f372ca883328e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 326, + 252, + 401, + 309 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e5450e97c47ce50415b03b58ab28db0b3632bdc842d45b5524c55ebae1f10cf5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 403, + 252, + 480, + 309 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A monkey $\\rightarrow$ man wearing colorful goggles and a colorful scarf", + "bbox": [ + 114, + 311, + 455, + 323 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fa1e77d8489ed52e472799bfeb95b24b43cddf11fcbe46268f65c13e4f205dfa.jpg", + "image_caption": [ + "Figure 7. Comparison of different editing methods on various objects and styles." + ], + "image_footnote": [], + "bbox": [ + 94, + 324, + 169, + 382 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cfd0cb6fd2309746225d4176b5723e1b12697f932a48a12cbea1fd8ba17b0d9b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 171, + 324, + 248, + 382 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3a4b9116d403bc2969adba0180a410d118bd10aab311aee5251702f5d76f5aca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 250, + 324, + 325, + 382 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a6ad37aba4a9dfe6cf5eea7ce806d6c952033df2bedfc2258e2f35a8a893b5ee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 326, + 323, + 401, + 382 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7e7f55139e27b356337557a9444fac7e08ba0dee13c0e8baf828bfc193734545.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 403, + 323, + 480, + 382 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Source", + "bbox": [ + 114, + 383, + 161, + 395 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DDS", + "bbox": [ + 197, + 383, + 230, + 395 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "CDS", + "bbox": [ + 266, + 383, + 297, + 395 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "NT+P2P", + "bbox": [ + 334, + 383, + 388, + 395 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ours", + "bbox": [ + 426, + 383, + 457, + 395 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "lower when no editing occurs, which is particularly visible in DDS-based methods applied to style editing. Our methods achieving better editing score but with a slightly higher structure distance. In terms of precise structure preservation, when combined with CDS [35], it achieves good preservation of the un-edited areas. Our full model achieves the best performance in the CLIP Similarity metric, demonstrating the effectiveness of our prompt enhancement branch. While CDS excels in preserving unedited regions, it suffers from the inferior editability of DDS-based methods. Optimization-based methods [27, 34] refine the inversion process, achieving excellent performance in structure preservation. However, they struggle with editing methods (like P2P), resulting in limited editability.", + "bbox": [ + 89, + 431, + 483, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 652, + 254, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we conduct an ablation experiment to analyze different choices in our SSD. Due to space limitations, we first present a qualitative evaluation in the main text. Please refer to the Supp. for quantitative evaluation.", + "bbox": [ + 89, + 674, + 482, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The effectiveness of cross-trajectory. In Sec. 4.1, we have analyzed the necessity of cross-trajectory. This term makes the optimization process more stable and provides the source content regularization in ours design, which is also the key difference between ours and Classifier Score Distillation(CSD) [52]. In Fig. 1 and Fig. 5, we present the comparison of the results with and without cross-trajectory. The results show that the cross-trajectory term can provide the direction of generating high-quality images. Please refer to the supplement for more details.", + "bbox": [ + 89, + 734, + 482, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The effectiveness of prompt-enhancement. The enhance-", + "bbox": [ + 89, + 885, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ac2d77153e3c555c622e7de06d3621d0ff7940ddb8aa9aa6a7b0d6c56a6848f4.jpg", + "image_caption": [ + "w/o ID Regular" + ], + "image_footnote": [], + "bbox": [ + 519, + 88, + 642, + 159 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/10d4cd7612166efab2a8900afe039440be70e8834f230d6440cc0ea132aeb552.jpg", + "image_caption": [ + "ID Regular $\\times 1.0$" + ], + "image_footnote": [], + "bbox": [ + 645, + 88, + 769, + 159 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a63e838151873b6b242f4712f0ede5ce0fc27a1a5c7d6387d9a6eb7bcb3c9b6a.jpg", + "image_caption": [ + "ID Regular $\\times 2.0$" + ], + "image_footnote": [], + "bbox": [ + 774, + 88, + 897, + 159 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ment of the target prompt branch is another key component in our method, which is designed to improve the editability aligned with the target prompt in 2D-image task. In Fig. 7, we observe a clear distinction from DDS in style editing. The results show that the prompt-enhancement term effectively overcomes the challenging from style editing.", + "bbox": [ + 511, + 257, + 906, + 348 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The effectiveness of ID regularization. ID regularization is designed to ensure stable optimization in 3DGS. In Fig. 8, we compare results with and without ID regularization. The area marked by the yellow arrow highlights its effect in 3D scene editing. However, excessive ID regularization may constrain editing quality by limiting certain attributes, presenting a trade-off in our design.", + "bbox": [ + 511, + 348, + 906, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions, Limitations, and Future Work", + "text_level": 1, + "bbox": [ + 511, + 467, + 901, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a novel method for text-guided image editing, capable of handling both 3D scenes and 2D images. Our approach is built on a score distillation framework that leverages the powerful priors of diffusion models. For editing tasks, we design an effective optimization strategy that produces high-quality results aligned with target prompts while ensuring stable and consistent optimization.", + "bbox": [ + 511, + 493, + 906, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method achieves state-of-the-art performance in both 3D scene and 2D image editing, delivering realistic edits with excellent preservation of the original content. It demonstrates strong adaptability to various editing tasks and target prompts, making it a robust solution for complex scenarios. However, while effective, the optimization process is relatively time-intensive compared to recent one-step methods [50] or few-step approaches [7]. Future work could explore integrating advanced techniques such as LCM [30] or SD-Turbo [42], which show potential for accelerating the optimization process [45].", + "bbox": [ + 511, + 599, + 908, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment: This project is supported by the National Natural Science Foundation of China (62125201, U24B20174, 62102381), the Guangdong Natural Science Funds for Distinguished Young Scholars (Grant 2023B1515020097), the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No.: AISG3-GV-2023-011), the Singapore Ministry of Education AcRF Tier 1 Grant (Grant No.: MSS25C004), the Lee Kong Chian Fellowships.", + "bbox": [ + 511, + 765, + 908, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16604", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, pages 5470–5479, 2022. 6", + "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, pages 18392-18402, 2023. 3, 6", + "[3] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In ICCV, pages 22560-22570, 2023. 6", + "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 7", + "[5] Minghao Chen, Iro Laina, and Andrea Vedaldi. Dge: Direct gaussian 3d editing by consistent multi-view editing. In ECCV, pages 74-92, 2024. 6", + "[6] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. In CVPR, pages 21476-21485, 2024. 3, 6", + "[7] Gilad Deutch, Rinon Gal, Daniel Garibi, Or Patashnik, and Daniel Cohen-Or. Turboedit: Text-based image editing using few-step diffusion models. In SIGGRAPHA, 2024. 5, 8", + "[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 34:8780-8794, 2021. 3", + "[9] Ayaan Haque, Matthew Tancik, Alexei Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. In ICCV, 2023. 3, 6", + "[10] Amir Hertz, Kfir Aberman, and Daniel Cohen-Or. Delta denoising score. In ICCV, pages 2328-2337, 2023. 1, 2, 3, 4, 6, 7", + "[11] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In ICLR, 2023. 6", + "[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshop, 2021. 2, 3, 5", + "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 1, 3", + "[14] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 3, 6", + "[15] Yukun Huang, Jianan Wang, Yukai Shi, Boshi Tang, Xianbiao Qi, and Lei Zhang. Dreamtime: An improved optimization strategy for diffusion-guided 3d generation. In ICLR, 2023. 3", + "[16] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. In CVPR, pages 12469-12478, 2024. 3", + "[17] Aapo Hyvarinen and Peter Dayan. Estimation of non-normalized statistical models by score matching. JMLR, 6(4), 2005. 3" + ], + "bbox": [ + 93, + 114, + 485, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Yutao Jiang, Yang Zhou, Yuan Liang, Wenxi Liu, Jianbo Jiao, Yuhui Quan, and Shengfeng He. Diffuse3d: Wide-angle 3d photography via bilateral diffusion. In ICCV, pages 8998-9008, 2023. 3", + "[19] Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. Pnp inversion: Boosting diffusion-based editing with 3 lines of code. In ICLR, 2024. 2, 6, 7", + "[20] Oren Katzir, Or Patashnik, Daniel Cohen-Or, and Dani Lischinski. Noise-free score distillation. In ICLR, 2024. 1, 2, 4", + "[21] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splattering for real-time radiance field rendering. ACM TOG, 42(4):1-14, 2023. 3, 6", + "[22] Jiwook Kim, Seonho Lee, Jaeyo Shin, Jiho Choi, and Hyunjung Shim. Dreamcatalyst: Fast and high-quality 3d editing via controlling editability and identity preservation. In ICLR, 2025. 2, 3", + "[23] Hubert Kompanowski and Binh-Son Hua. Dream-in-style: Text-to-3d generation using stylized score distillation. In 3DV, 2025. 2", + "[24] Juil Koo, Chanho Park, and Minhyuk Sung. Posterior distillation sampling. In CVPR, pages 13352-13361, 2024. 2, 3, 5, 6", + "[25] Duong H Le, Tuan Pham, Aniruddha Kembhavi, Stephan Mandt, Wei-Chiu Ma, and Jiasen Lu. Preserving identity with variational score for general-purpose 3d editing. arXiv preprint arXiv:2406.08953, 2024. 3", + "[26] Kehan Li, Yanbo Fan, Yang Wu, Zhongqian Sun, Wei Yang, Xiangyang Ji, Li Yuan, and Jie Chen. Learning pseudo 3d guidance for view-consistent texturing with 2d diffusion. In ECCV, pages 18-34, 2025. 1", + "[27] Senmao Li, Joost Van De Weijer, Taihang Hu, Fahad Shahbaz Khan, Qibin Hou, Yaxing Wang, and Jian Yang. Stylediffusion: Prompt-embedding inversion for text-based editing. arXiv preprint arXiv:2303.15649, 2023. 7, 8", + "[28] Haofeng Liu, Chenshu Xu, Yifei Yang, Lihua Zeng, and Shengfeng He. Drag your noise: Interactive point-based editing via diffusion semantic propagation. In CVPR, pages 6743-6752, 2024. 1", + "[29] Yuqin Lu, Bailin Deng, Zhixuan Zhong, Tianle Zhang, Yuhui Quan, Hongmin Cai, and Shengfeng He. 3d snapshot: Invertible embedding of 3d neural representations in a single image. 46(12):11524-11531, 2024. 3", + "[30] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 8", + "[31] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In CVPR, pages 12663–12673, 2023. 1", + "[32] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM TOG, 38(4): 1-14, 2019. 6" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16605", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, pages 405-421, 2020. 3", + "[34] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, pages 6038–6047, 2023. 7, 8", + "[35] Hyelin Nam, Gihyun Kwon, Geon Yeong Park, and Jong Chul Ye. Contrastive denoising score for text-guided latent diffusion image editing. In CVPR, pages 9192-9201, 2024. 2, 7, 8", + "[36] JangHo Park, Gihyun Kwon, and Jong Chul Ye. ED-NeRF: Efficient text-guided editing of 3d scene with latent space nef. In ICLR, 2024. 2, 3", + "[37] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 1, 3, 4", + "[38] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2): 3, 2022. 3", + "[39] Jingjing Ren, Wenbo Li, Haoyu Chen, Renjing Pei, Bin Shao, Yong Guo, Long Peng, Fenglong Song, and Lei Zhu. Ultrapixel: Advancing ultra high-resolution image synthesis to new peaks. In NeurIPS, pages 111131-111171, 2024. 3", + "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 3, 6", + "[41] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, pages 36479-36494, 2022. 1, 3", + "[42] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In ECCV, pages 87-103, 2025. 8", + "[43] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 1, 3, 6", + "[44] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 3", + "[45] Feng Tian, Yixuan Li, Yichao Yan, Shanyan Guan, Yanhao Ge, and Xiaokang Yang. Postedit: Posterior sampling for efficient zero-shot image editing. In ICLR, 2025. 8", + "[46] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In CVPR, pages 1921-1930, 2023. 6", + "[47] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In NeurIPS, 2024. 1, 3", + "[48] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva:" + ], + "bbox": [ + 91, + 92, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 7", + "[49] Jing Wu, Jia-Wang Bian, Xinghui Li, Guangrun Wang, Ian Reid, Philip Torr, and Victor Adrian Prisacariu. Gaussctrl: Multi-view consistent text-driven 3d gaussian splatting editing. In ECCV, pages 55-71, 2024. 3", + "[50] Sihan Xu, Yidong Huang, Jiayi Pan, Ziqiao Ma, and Joyce Chai. Inversion-free image editing with natural language. In CVPR, 2024. 8", + "[51] Yangyang Xu, Shengfeng He, Kwan-Yee K Wong, and Ping Luo. Rigid: Recurrent gan inversion and editing of real face videos and beyond. IJCV, 133(6):3437-3455, 2025. 1", + "[52] Xin Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Song-Hai Zhang, and Xiaojuan Qi. Text-to-3d with classifier score distillation. In ICLR, 2024. 1, 2, 3, 4, 5, 8", + "[53] Yuyang Yu, Bangzhen Liu, Chenxi Zheng, Xuemiao Xu, Huaidong Zhang, and Shengfeng He. Beyond textual constraints: Learning novel diffusion conditions with fewer examples. In CVPR, pages 7109-7118, 2024. 3", + "[54] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836-3847, 2023. 3", + "[55] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7", + "[56] Chenxi Zheng, Yihong Lin, Bangzhen Liu, Xuemiao Xu, Yongwei Nie, and Shengfeng He. Recdreamer: Consistent text-to-3d generation via uniform score distillation. In ICLR, 2025. 1, 3", + "[57] Xingchen Zhou, Ying He, F Richard Yu, Jianqiang Li, and You Li. RePaint-NeRF: Nerf editing via semantic masks and diffusion models. In IJCAI, 2023. 3", + "[58] Junzhe Zhu, Peiye Zhuang, and Sanmi Koyejo. HIFA: High-fidelity text-to-3d generation with advanced diffusion guidance. In ICLR, 2024. 1, 3" + ], + "bbox": [ + 516, + 92, + 906, + 587 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16606", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Stable Score Distillation/62930789-4f50-45e0-b2ce-8d1494eced86_model.json b/2025/Stable Score Distillation/62930789-4f50-45e0-b2ce-8d1494eced86_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bd609421ccfdc55d370723a7892028f9d56057a0 --- /dev/null +++ b/2025/Stable Score Distillation/62930789-4f50-45e0-b2ce-8d1494eced86_model.json @@ -0,0 +1,3058 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.375, + 0.131, + 0.624, + 0.151 + ], + "angle": 0, + "content": "Stable Score Distillation" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.18, + 0.74, + 0.217 + ], + "angle": 0, + "content": "Haiming Zhu\\(^{1}\\), Yangyang Xu\\(^{2*}\\), Chenshu Xu\\(^{1}\\), Tingrui Shen\\(^{3}\\), Wenxi Liu\\(^{4}\\), Yong Du\\(^{5}\\), Jun Yu\\(^{2}\\), Shengfeng He\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.218, + 0.777, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\)Singapore Management University \\(^{2}\\)Harbin Institute of Technology (Shenzhen)" + }, + { + "type": "text", + "bbox": [ + 0.179, + 0.236, + 0.815, + 0.251 + ], + "angle": 0, + "content": "\\(^{3}\\)South China University of Technology \\(^{4}\\)Fuzhou University \\(^{5}\\)Ocean University of China" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.255, + 0.244, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.341, + 0.24, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.428, + 0.22, + 0.44 + ], + "angle": 0, + "content": "(a) Original Views" + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.256, + 0.444, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.341, + 0.457, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.32, + 0.428, + 0.387, + 0.44 + ], + "angle": 0, + "content": "(b) CSD[52]" + }, + { + "type": "image", + "bbox": [ + 0.461, + 0.256, + 0.658, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.461, + 0.341, + 0.658, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.428, + 0.604, + 0.44 + ], + "angle": 0, + "content": "(c) DDS[10]" + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.256, + 0.898, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.677, + 0.342, + 0.896, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.763, + 0.428, + 0.807, + 0.44 + ], + "angle": 0, + "content": "(d) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.446, + 0.908, + 0.503 + ], + "angle": 0, + "content": "Figure 1. We propose Stable Score Distillation (SSD), a method that improves text-guided editing by preserving original content structure and enhancing realism in edited results. The first column shows the original views, while the remaining columns display comparative results with existing score distillation methods on the SD model. SSD demonstrates superior performance, maintaining the integrity of the original content and producing more realistic edits." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.519, + 0.327, + 0.535 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.55, + 0.486, + 0.868 + ], + "angle": 0, + "content": "Text-guided image and 3D editing have advanced with diffusion-based models, yet methods like Delta Denoising Score often struggle with stability, spatial control, and editing strength. These limitations stem from reliance on complex auxiliary structures, which introduce conflicting optimization signals and restrict precise, localized edits. We introduce Stable Score Distillation (SSD), a streamlined framework that enhances stability and alignment in the editing process by anchoring a single classifier to the source prompt. Specifically, SSD utilizes Classifier-Free Guidance (CFG) equation to achieve cross-prompt alignment, and introduces a constant term null-text branch to stabilize the optimization process. This approach preserves the original content's structure and ensures that editing trajectories are closely aligned with the source prompt, enabling smooth, prompt-specific modifications while maintaining coherence in surrounding regions. Additionally, SSD incorporates a prompt enhancement branch to boost editing strength, particularly for style transformations. Our method achieves state-of-the-art results in 2D and 3D editing tasks, including NeRF and text-driven style edits, with faster conver" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.521, + 0.909, + 0.566 + ], + "angle": 0, + "content": "gence and reduced complexity, providing a robust and efficient solution for text-guided editing. Code is available at: https://github.com/Alex-Zhu1/SSD." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.588, + 0.646, + 0.603 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.613, + 0.907, + 0.734 + ], + "angle": 0, + "content": "Text-based image generation has achieved remarkable progress, particularly with the advent of diffusion models [13, 28, 40, 41, 43, 56]. These models leverage strong priors to produce high-quality images, facilitating significant advances in text-to-3D generation [37, 47, 58]. Moreover, text-guided 3D editing has enabled intricate modifications to shape and texture [26, 31], supporting flexible and precise 3D scene manipulation." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.735, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Unlike generation tasks that create new content, editing tasks aim to modify specific elements within an image while preserving surrounding areas [51]. However, directly applying methods like Score Distillation Sampling (SDS) to editing tasks can yield undesired effects, such as blurring across the image. This arises because SDS optimizes globally to the prompt, affecting regions beyond the targeted area [10, 20]. DDS [10] addresses this by introducing a dual-branch architecture, pairing the source image with its description to leverage the model's inherent bias and isolate specific prompt changes. Further, CSD [52] achieved" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "*Corresponding authors: Yangyang Xu (xuyangyang@hit.edu.cn) and Shengfeng He (shengfenghe@smu.edu.sg)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16597" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.094, + 0.323, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.094, + 0.547, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.094, + 0.77, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.094, + 0.892, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.221, + 0.908, + 0.278 + ], + "angle": 0, + "content": "Figure 2. Illustration of three distillation-based approaches. Note that we assume a 2-step optimization process for illustration, where the subscript \\( t \\) represents the iteration number. DDS utilizes the source branch to obtain initial latent \\( Z_0^* \\), while CSD employs two classifiers to derive \\( Z_1' \\) and \\( Z_1'' \\) for cross-prompt editing. Our SSD method designs a CFG classifier to determine the cross-prompt editing, introduces the null-text branch as the initial latent \\( Z_0^* \\), and further constructs the cross-trajectory term (see Sec. 4.1) for stable optimization." + }, + { + "type": "image", + "bbox": [ + 0.09, + 0.289, + 0.491, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.462, + 0.486, + 0.504 + ], + "angle": 0, + "content": "Figure 3. The optimization process of DDS and our SSD. SSD preserves the source structure effectively during optimization iterations, while DDS cannot preserve it effectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.484, + 0.608 + ], + "angle": 0, + "content": "scene editing by incorporating a classifier component within Classifier-Free Guidance (CFG) [12] to refine the prediction score by applying the classifier to both the source and target prompts. NFSD [20] further decomposes the CFG score, highlighting the classifier as the primary driver of prompt direction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.484, + 0.867 + ], + "angle": 0, + "content": "Despite their success, we argue that current distillation-based approaches face inherent limitations, such as low editing quality and loss of source content. As shown in Fig. 2, DDS [10] relies on the source branch to remove model bias, but lacks the explicit guidance to preserve the source content [24] during optimization. As shown in Fig. 1c, DDS changes man's clothes during editing his faces. Additionally, although introducing source prompt components is intended to improve prompt specificity [19], it can amplify noise and introduce overlapping objectives that hinder stable convergence. This results in artifacts or unintended variations, especially in the unedited regions. Correspondingly, CSD [52] utilizes dual classifiers to refine the prompt editing direction (see Fig. 2). However, it lacks the explicit source preservation to restrict edits precisely to the target areas. As shown in Fig. 1b, this causes the structure deformation and annoying artifacts in the edited regions." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Our insights into these limitations lead to two key observations: (1) Cross-prompt: a single classifier, providing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.907, + 0.352 + ], + "angle": 0, + "content": "the editing direction from source prompt to target, and (2) Cross-trajectory: stability in the editing process can be achieved by aligning the editing direction closely with the structure of the source content." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.355, + 0.909, + 0.717 + ], + "angle": 0, + "content": "In this paper, we propose Stable Score Distillation (SSD), a streamlined approach for stable and precise text-guided editing. To achieve a smooth editing direction, we employ the CFG equation for both the source and target prompts, ensuring a gradual transition of the original contextual texture as the model adapts to the specified changes. This approach contrasts with DDS [10], as it eliminates the need for a auxiliary source branch, enabling our method to focus editing gradients precisely within target regions while ensuring a stable transition, as illustrated in Fig. 3. Moreover, for aligning the editing direction with the source prompt, facilitating smoother and more controlled progression toward the target prompt, we design an cross-trajectory strategy to ensure that edits respect the original structure, supporting subtle and stable transformations within designated areas. While NFSD [20] utilizes negative-branch and DDS utilizes source branch to enhance output clarity, as shown in SSD in Fig. 2, we introduce a null-text branch aligned with the \"no-edit\" direction to integrate a \"reconstruction\" term to explicitly enforce source content preservation, which enhances consistency and produces reliable edits across diverse tasks. Based on above designs, our framework remains streamlined and efficient, achieving both precision and stability without the complexity of additional components." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.909, + 0.884 + ], + "angle": 0, + "content": "Our framework integrates seamlessly into existing DDS-based editing pipelines and applications, such as text-driven NeRF editing [22, 24, 36] and 2D image editing [35]. Notably, our approach's \"clear\" editing direction preserves source content, making a carefully designed identity regularization [22] unnecessary. Moreover, standard DDS methods often lack sufficient editing strength, resulting in minimal or negligible changes in output, particularly in style editing [23]. Our approach, with its streamlined and stable framework, allows for the seamless integration of a prompt enhancement branch to amplify editing capability." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.907, + 0.901 + ], + "angle": 0, + "content": "With these improvements, our method achieves faster and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16598" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.212 + ], + "angle": 0, + "content": "more effective edits during optimization, remains compatible with the Stable Diffusion Model [40] without requiring LoRA [14] or fine-tuning, and integrates effectively with Instructpix2pix [2]. Additionally, by incorporating nonincreasing timestep sampling [15], we accelerate convergence, reducing the required iterations to approximately 3,000 for NeRF [33] and 1,500 for Gaussian splatting [21]. In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.214, + 0.482, + 0.258 + ], + "angle": 0, + "content": "- We introduce a novel editing framework, Stable Score Distillation, that leverages a single, anchored classifier to achieve targeted and stable edits in 3D scene editing." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.26, + 0.482, + 0.304 + ], + "angle": 0, + "content": "- We introduce a prompt enhancement strategy, effectively improve the prompt-alignment, especially style editing in 2D-image editing." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.305, + 0.482, + 0.348 + ], + "angle": 0, + "content": "- We demonstrate the effectiveness of our approach across NeRF-editing and image-editing tasks, achieving state-of-the-art results with a streamlined and efficient framework." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.214, + 0.482, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.365, + 0.232, + 0.381 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.39, + 0.259, + 0.405 + ], + "angle": 0, + "content": "2.1. Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.484, + 0.563 + ], + "angle": 0, + "content": "Diffusion models [13, 39, 43, 53] have made significant advancements in generating diverse and high-fidelity images. Starting from a gaussian noise, diffusion models can predict the noise-less sample at each time step, until finally obtaining clear samples. Commonly, the denoising process can utilize U-net model to predict the noise. Some works [13, 44] have observed that is proportional to the predicted score function [17] of the smoothed density. Thus, intuitively, taking steps in the direction of the score function gradually moves the sample towards the data distribution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.566, + 0.484, + 0.761 + ], + "angle": 0, + "content": "To generate images aligned with a target prompt, guidance is typically introduced to explicitly control the weight assigned to the conditioning information. The popular guidance methods include Classifier Guidance [8] and Class-free Guidance (CFG) [12]. While the former relies on a separately learned classifier, the latter directly introduces null-text samples to the model. CFG modifies the score function to steer the process towards regions with a higher ratio of conditional density to the unconditional one. However, it has been observed that CFG trades sample fidelity for diversity [12]. Based on the insights gained from the decomposition of the CFG equation, we propose a novel Stable Score Distillation (SSD) method to guide the SDS optimization process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.772, + 0.385, + 0.789 + ], + "angle": 0, + "content": "2.2. Score Distillation Sampling (SDS)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Benefit from the data scale-law, diffusion model [38, 40, 41] achieve high-quality image generation and text-to-image generation. Specifically, Score Distillation Sampling (SDS) [37] leveraging the priors of pre-trained text-to-image models to facilitate text-conditioned generation in 3D content generation. Specifically, SDS is an optimization approach that updates the rendering parameter towards the image distribu" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.258 + ], + "angle": 0, + "content": "tion of diffusion models by enforcing the noise prediction on noisy rendered images to match sampled noise. While SDS provides an elegant mechanism for leveraging pretrained text-to-image models, SDS-generated results often suffer from oversaturation and lack of fine realistic details. VSD [47] proposed a particle-based optimization framework that treats the 3D parameter as a random variable of target distribution. Furthermore, by regarding SDS as a reverse diffusion process, decreasing timesteps sampling [15, 58] to imitate the diffusion reverse sampling, which can improve the quality of the generated 3D assets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.263, + 0.905, + 0.368 + ], + "angle": 0, + "content": "In image editing, Delta Denoising Score (DDS) [10] found that Score Distillation Sampling (SDS) introduces noticeable artifacts and over-smoothing in edited images due to inherent bias. To mitigate this bias, DDS employs a subtraction of two SDS scores of the source and target images to obtain a delta score, which is then used to guide the optimization process." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.395, + 0.78, + 0.411 + ], + "angle": 0, + "content": "2.3. Text-Driven 3D-Scene Editing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.422, + 0.907, + 0.573 + ], + "angle": 0, + "content": "Text-driven 3D scene editing has been a popular research topic [18, 29, 56]. IN2N [9] proposed a Iterative Dataset Update method that can edit 3D scenes from text descriptions. By leveraging advancements in 2D diffusion editing techniques, notably InstructP2P [2] and ControlNet [54], GaussianEditor[6] and GaussCtrl [49] utilize edit multi-view images latent to optimize the 3D scene. We consider utilize score distillation to guide the 3D scene editing, which is more flexible and efficient for the text-driven 3D scene editing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.579, + 0.909, + 0.836 + ], + "angle": 0, + "content": "Building upon the foundational SDS loss introduced by DreamFusion [37], some work has explored SDS loss in the text-driven 3D scene editing. RePaint-NeRF [57] has advanced the application of SDS in 3D editing by integrating a semantic mask to guide and constrain modifications within the background elements. CSD [52] utilize two classifiers to achieve editing. In a similar vein, ED-NeRF [36] has introduced an enhanced loss function specifically designed for 3D editing tasks. PDS [24] proposed a posterior distillation sampling to match stochastic latent [16]. Piva [25] fine-tuned the model while introducing a regularization term to preserve identity. Unfortunately, these methods are still limited to the long-time diffusion reverse sampling process, which is not suboptimal for the text-driven 3D scene editing. DreamCatalyst [22] extends the PDS optimization processing to ID-preserving and edit-ability based on decreasing timesteps sampling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Different from the above methods, ours firstly improve the DDS optimization process by introducing a single classifier, and further introduce a null-text branch to achieve a more stable and precise editing process." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16599" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.09, + 0.881, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.348, + 0.908, + 0.406 + ], + "angle": 0, + "content": "Figure 4. The overview of SSD. Given the parameter 3D-model or image, SSD provides effective editing gradient to guide the optimization process. We utilize CFG equation between the predicted target noise \\(\\epsilon_{\\phi}(z_t,y,t)\\) and source noise \\(\\epsilon_{\\phi}(z_t,\\hat{y},t)\\), which generate the gradual editing direction. Furthermore, we introduce a null-text branch \\(\\epsilon_{\\phi}(\\hat{z}_t,\\emptyset ,t)\\) to regularize the optimization process and achieve stable optimization. We further analyzing and decompose ours design term into three parts: cross-prompt, cross-trajectory, and prompt-enhance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.415, + 0.218, + 0.433 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.44, + 0.484, + 0.501 + ], + "angle": 0, + "content": "In this section, we first discuss existing optimization-based approaches to handle parametric images. Then, we will introduce our novel parametric image editing method in Section 4." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.51, + 0.338, + 0.527 + ], + "angle": 0, + "content": "3.1. Score Distillation Sampling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.532, + 0.486, + 0.625 + ], + "angle": 0, + "content": "Score Distillation Sampling (SDS) [37] is proposed to generate parametric images by leveraging the 2D prior of pretrained text-to-image diffusion models. Specifically, given a pretrained diffusion model \\(\\epsilon_{\\phi}\\), SDS optimizes a set of parameters \\(\\theta\\) of a differentiable parametric image generator \\(g\\), using the gradient of the loss \\(L_{\\mathrm{SDS}}\\) with respect to \\(\\theta\\):" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.633, + 0.484, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\theta} L _ {\\mathrm {S D S}} = w (t) \\left(\\epsilon_ {\\phi} \\left(z _ {t} (x); y, t\\right) - \\epsilon\\right) \\frac {\\partial x}{\\partial \\theta}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.486, + 0.795 + ], + "angle": 0, + "content": "where \\(x = g(\\theta)\\) is an image rendered by \\(\\theta\\), \\(z_{t}(x)\\) is obtained by adding a Gaussian noise \\(\\epsilon\\) to \\(x\\) corresponding to the \\(t\\)-th timestep of the diffusion process, and \\(y\\) is a condition to the diffusion model. As Noise-Free Score Distillation (NFSD) [20] has shown, the score \\(\\epsilon_{\\phi}(z_{t}(x);y,t)\\) provides the direction in which this noised version of \\(x\\) should be moved towards a denser region in the distribution of real images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.803, + 0.337, + 0.82 + ], + "angle": 0, + "content": "3.2. Delta Distillation Sampling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Although SDS get excellent generation ability, for editing task, an undesired component from the pretrained model, \\(\\delta_{\\mathrm{bias}}\\), interferes with the process and causes the image to become smooth and blurry in some parts [10]. Based on the observations that a matched source prompt \\(\\hat{y}\\) and source" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.907, + 0.462 + ], + "angle": 0, + "content": "latent \\(\\hat{z}_t\\) can estimate the noisy direction \\(\\delta_{\\mathrm{bias}}\\), thus, the DDS method aims to remove the \\(\\delta_{\\mathrm{bias}}\\) by introducing source branch, as shown in Eq. 2:" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.467, + 0.907, + 0.497 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\theta} L _ {\\mathrm {D D S}} = \\left(\\epsilon_ {\\phi} ^ {c} \\left(z _ {t}, y, t\\right) - \\epsilon_ {\\phi} ^ {c} \\left(\\hat {z} _ {t}, \\hat {y}, t\\right)\\right) \\frac {\\partial z}{\\partial \\theta}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.909, + 0.638 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\phi}^{c}(z_{t},y,t)\\) and \\(\\epsilon_{\\phi}^{c}(\\hat{z}_{t},\\hat{y},t)\\) are pretrained model predictions \\(\\epsilon\\), with the superscript \\(c\\) indicating the CFG results. Thus, DDS pushes the optimized image into the direction of the target prompt without the interference of the noise component, namely, \\(\\nabla_{\\theta}L_{\\mathrm{DDS}}\\approx \\delta_{\\mathrm{ext}}\\). Obviously, \\(\\nabla_{\\delta_{\\mathrm{ext}}}\\) is contingent on classifier part from \\(\\epsilon_{\\phi}^{c}(z_{t},y,t)\\) as discussed in CSD [52] and NFSD [20]. Note that in the following manuscript, we decompose the CFG results without the superscript \\(c\\) and omit the timestep \\(t\\) for simplicity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.638, + 0.908, + 0.683 + ], + "angle": 0, + "content": "Further exploring prompt editing direction, CSD [52] method proposed a dual-classifier to refine the editing score and achieve more precise editing, as shown in Eq. 3:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.689, + 0.901, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {\\theta} L _ {\\mathrm {C S D}} = \\left(w _ {a} \\left(\\epsilon_ {\\phi} \\left(z _ {t}, y\\right) - \\epsilon_ {\\phi} \\left(z _ {t}, \\varnothing\\right)\\right) \\right. \\\\ \\left. - w _ {b} \\left(\\epsilon_ {\\phi} \\left(z _ {t}, \\hat {y}\\right) - \\epsilon_ {\\phi} \\left(z _ {t}, \\emptyset\\right)\\right)\\right) \\frac {\\partial z}{\\partial \\theta}, \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.743, + 0.909, + 0.835 + ], + "angle": 0, + "content": "while the \\(\\epsilon_{\\phi}(z_t,y,t)\\) and \\(\\epsilon_{\\phi}(z_t,\\hat{y},t)\\) are current latent \\(z_{t}\\) predictions for the target prompt \\(y\\) and source prompt \\(\\hat{y}\\), respectively. \\(w_{a}\\) and \\(w_{b}\\) are weights of classifiers. Simply put, CSD aims to refine the prompt editing direction by determining the difference between two classifiers, which can be regarded as a cross-prompt term." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.845, + 0.605, + 0.86 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In 3D scene editing process, which requires consideration of both the target prompt and the original source content," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16600" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.152 + ], + "angle": 0, + "content": "we consider two key aspects: (1) smooth editing direction towards the target prompt (2) and editing results respect the original structure. Based on these, in this section, we introduce our novel editing framework Stable Score Distillation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.164, + 0.313, + 0.179 + ], + "angle": 0, + "content": "4.1. Stable Score Distillation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.188, + 0.485, + 0.307 + ], + "angle": 0, + "content": "Firstly, we introduce the design of a cross-prompt editing direction. As discussed about CSD method in Sec. 3.2, the key role in cross-prompt editing is to provide a smooth transition from the source prompt to the target. As the CFG guidance [12] steers the process towards regions with a higher ratio of conditional density to the unconditional one, accordingly, we can modify the SDS score function, as shown below:" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.322, + 0.484, + 0.34 + ], + "angle": 0, + "content": "\\[\nG r a d = \\epsilon_ {\\phi} (z _ {t}, \\hat {y}) + s (\\epsilon_ {\\phi} (z _ {t}, y) - \\epsilon_ {\\phi} (z _ {t}, \\hat {y})), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.353, + 0.484, + 0.383 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\phi}(z_t,y)\\) and \\(\\epsilon_{\\phi}(z_t,\\hat{y})\\) are pretrained model predictions. The scale factor \\(s\\) is equal to control weight." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.384, + 0.484, + 0.566 + ], + "angle": 0, + "content": "Although the cross-prompt term provides a smooth texture transition in the edited region, we observed that the optimization process leads to abrupt structural changes, often resulting in artifacts and unappealing outcomes, similar to CSD in Fig. 1b. To address this, we introduce an additional regularization term to constrain the structural transition. Interestingly, as shown in Fig. 1c, DDS achieves better results than CSD by incorporating a source branch. However, DDS still lacks a mechanism to ensure the original structure remains intact, leading to modification on unedited regions. To address this, we introduce a null-text branch \\(\\epsilon_{\\phi}(\\hat{z}_t,\\emptyset)\\) to regularize the optimization process, as shown in Eq. 5:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.579, + 0.484, + 0.596 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {s s d}} = \\epsilon_ {\\phi} (z _ {t}, \\hat {y}) + s \\left(\\epsilon_ {\\phi} (z _ {t}, y) - \\epsilon_ {\\phi} (z _ {t}, \\hat {y})\\right) - \\epsilon_ {\\phi} (\\hat {z} _ {t}, \\varnothing). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.61, + 0.484, + 0.656 + ], + "angle": 0, + "content": "Eq. 5 is ours Stable Score Distillation, and we can further decompose above equation into two parts, and the latter is regarded as a cross-trajectory term." + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.669, + 0.506, + 0.718 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {s s d}} = \\underbrace {w _ {p} \\left(\\epsilon_ {\\phi} \\left(z _ {t} , y\\right) - \\epsilon_ {\\phi} \\left(z _ {t} , \\hat {y}\\right)\\right)} _ {\\text {c r o s s - p r o m p t}} + \\underbrace {w _ {t} \\left(\\epsilon_ {\\phi} \\left(z _ {t} , \\hat {y}\\right) - \\epsilon_ {\\phi} \\left(\\hat {z} _ {t} , \\emptyset\\right)\\right)} _ {\\text {c r o s s - t r a j e c t o r y}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.749 + ], + "angle": 0, + "content": "where the \\(w_{t}\\) and \\(w_{p}\\) control the strength of the cross-trajectory and cross-prompt, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.486, + 0.901 + ], + "angle": 0, + "content": "The cross-trajectory term can be interpreted as the distance between the transitions of two latents, ensuring that the original structure remains smooth and does not change abruptly (more details are provided in the supplementary material). In Fig. 5, we can see that the cross-trajectory term can provide a strong structure constraint ability, guiding the optimization process to preserve the source image structure. Specifically, when set \\( w_{t} = 0 \\), the optimization process behaves similarly to the CSD[52] method, which fails to retain the original image structure." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.346, + 0.908, + 0.414 + ], + "angle": 0, + "content": "Figure 5. The effect of increasing the strength of the prompt enhancement term \\((w_{e})\\) and cross-trajectory term \\((w_{t})\\), with the cross-prompt term fixed at 7.5. Both terms contribute to prompt-aligned results, while setting \\((w_{t} = 0)\\) leads to saturation and discard source content." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.417, + 0.783, + 0.433 + ], + "angle": 0, + "content": "4.2. Improving Prompt Alignment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.439, + 0.907, + 0.576 + ], + "angle": 0, + "content": "Although Eq. 5 can achieve gradual editing results, we found Eq. 5 have similar limitation with DDS [7], which have insufficient editing strength. The editing results neither get successful editing nor retain the source image structure, often leads to little or no change in the final. Benefit from the cross-prompt editing design as Eq. 5, we can add a target prompt enhancement branch to guide the optimization process. The target prompt alignment branch will provide the direction of the target prompt, as shown in Eq. 7:" + }, + { + "type": "equation", + "bbox": [ + 0.589, + 0.585, + 0.907, + 0.601 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {a l i g n}} = w _ {e} \\left(\\epsilon_ {\\phi} \\left(z _ {t}, y\\right) - \\epsilon_ {\\phi} \\left(z _ {t}, \\varnothing\\right)\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.906, + 0.67 + ], + "angle": 0, + "content": "where \\(w_{e}\\) is the prompt enhancement scale. As shown in Fig. 5, the synchronous scaling of both the cross-trajectory and prompt-enhancement terms results in effective visual editing outcomes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.679, + 0.779, + 0.695 + ], + "angle": 0, + "content": "4.3. Source Latent Regularization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.701, + 0.907, + 0.821 + ], + "angle": 0, + "content": "Empirically, we found that directly using latent-space loss rather than pixel-level loss can lead to optimization difficulties in local regions of 3DGS. For example, the bright spots appearing in Fig. 5. To suppress the steep gradients in these areas, we incorporate ID regularization to guide the stable optimization process. Differ with PDS [24] use source latent \\(\\hat{x}_0\\), we can use the noisy latent \\(\\hat{x}_t\\) to avoid partial exploding gradient, as shown in Eq. 8:" + }, + { + "type": "equation", + "bbox": [ + 0.63, + 0.831, + 0.907, + 0.847 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {I D}} = w (t) \\cdot \\left(x _ {t} - \\hat {x} _ {t}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "where the \\( w(t) \\) is the iteration-dependent strength, designed as a decreasing function of \\( t \\). Notably, the \\( w(t) \\) is not necessary to well-designed in our design. Our final loss function" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16601" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.218, + 0.106 + ], + "angle": 0, + "content": "as shown in Eq. 9:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.12, + 0.483, + 0.136 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {f i n a l}} = L _ {\\text {s s d}} + L _ {\\text {a l i g n}} + L _ {\\text {I D}}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.15, + 0.485, + 0.226 + ], + "angle": 0, + "content": "Based on the above design, we achieve a more prompt-aligned editing method, which integrates seamlessly into the Stable Diffusion Model [40] without requiring LoRA [14] or fine-tuning. Moreover, we will further introduce our method's connection with InstructPix2Pix [2]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.237, + 0.298, + 0.254 + ], + "angle": 0, + "content": "4.4. Connecting with IP2P" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.259, + 0.484, + 0.306 + ], + "angle": 0, + "content": "The final design of our method is shown in Eq 9. We found that ours edit gard provide new angle to understand about InstructP2P [2] one-step reverse sampling." + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.317, + 0.48, + 0.39 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\epsilon_ {\\theta} \\left(z _ {t}, c _ {I}, c _ {T}\\right) = \\epsilon_ {\\theta} \\left(z _ {t}, \\varnothing , \\varnothing\\right) \\\\ + s _ {I} \\left(\\epsilon_ {\\theta} \\left(z _ {t}, c _ {I}, \\varnothing\\right) - \\epsilon_ {\\theta} \\left(z _ {t}, \\varnothing , \\varnothing\\right)\\right) \\\\ + s _ {T} \\left(\\epsilon_ {\\theta} \\left(z _ {t}, c _ {I}, c _ {T}\\right) - \\epsilon_ {\\theta} \\left(z _ {t}, c _ {I}, \\varnothing\\right)\\right), \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.484, + 0.584 + ], + "angle": 0, + "content": "where \\( c_{I} \\) and \\( c_{T} \\) are input-image and instruction prompt separately, \\( s_{I} \\) and \\( s_{T} \\) are the source image control and instruction prompt control strength. The Eq. 10 is the InstructP2P one-step reverse sampling, which can provide the direction of the target prompt. We can see that the InstructP2P is the simple version of our method, the middle term of Eq. 10 is cross-trajectory regularization, and the last term of Eq. 10 is ours cross-prompt term. Simply put, as analyzing the Eq. 5, subtracting the constant correction term \\( \\epsilon_{\\theta}(\\hat{z}_t; \\varnothing; \\varnothing) \\) is edit grad. Ours method reveals that apply DDS loss in the InstructP2P model can only editing branch, and don't have to provide the source branch." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.599, + 0.224, + 0.617 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.625, + 0.484, + 0.716 + ], + "angle": 0, + "content": "In this section, we conduct editing experiments across two types of parameterized images. Section 5.1 evaluates the effectiveness of our method on 3D Scenes Editing, and Section 5.2 evaluates the effectiveness of our method on 2D Image Editing. We also conduct ablation studies to analyze the effectiveness of ours components in Section 5.3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.267, + 0.743 + ], + "angle": 0, + "content": "5.1.3D Scenes Editing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.485, + 0.809 + ], + "angle": 0, + "content": "Dataset. To evaluate the effectiveness of our method, we conduct experiments on the scenes from IN2N [9] and other real-world datasets, including LLFF [32] and Mip-Nerf360 [1]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Baselines. We compare our method with several state-of-the-art inversion methods. We use 3DGS [21] as the 3D representation, and compare our method with InstructNerf2Nerf [9], DDS [10], GS-Edit [6], and DGE [5]. For fairness, we implement the DDS version based on the official GS-Edit code. PDS [24] is designed for addition of objects to unspecified" + }, + { + "type": "table_caption", + "bbox": [ + 0.536, + 0.09, + 0.882, + 0.104 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluations under 3D editing scenes." + }, + { + "type": "table", + "bbox": [ + 0.554, + 0.105, + 0.86, + 0.192 + ], + "angle": 0, + "content": "
MethodCLIP Sim.↑Sim Dir.↑User Study ↑
IN2N0.16760.070714.54%
DDS0.17800.04015.45%
GS-Editor0.17580.042914.54%
DGE0.17580.056323.63%
Ours0.18460.077341.81%
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.198, + 0.908, + 0.228 + ], + "angle": 0, + "content": "regions, we will provide the comparison results in supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.229, + 0.907, + 0.395 + ], + "angle": 0, + "content": "Evaluation Metrics. We follow common practice [5, 6, 9] to evaluate the effectiveness of our method. CLIP Similarity is to evaluate the alignment between the render images and the target prompts, i.e., the cosine similarity between the text and image embeddings encoded by CLIP. Specifically, follow DGE [5], randomly sample 20 camera poses to evaluate. CLIP Directional Similarity is to measure the editing effect, i.e., the cosine similarity between the image and text editing directions (target embeddings minus source embeddings). We evaluate all methods on 6 different scenes and 10 different prompts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.396, + 0.909, + 0.636 + ], + "angle": 0, + "content": "Results. We begin by evaluating our method, starting with a qualitative assessment. In Fig. 6, we present a comparison of results with competing methods. Our approach generates more visually appealing images that are better aligned with the editing instructions. In contrast, methods based on the Iterative Dataset Update (IDU) strategy, such as IN2N [9] and GS-Editor [6], fail to produce the desired editing outcomes, resulting in blurrier or lower-fidelity reconstructions and noticeable artifacts. For example, in the scene of \"Spider-Man with a mask\", IN2N generates a mask with reduced fidelity, while GS-Editor produces a low-detail mask. In the multi-view consistency setup, DGE [5] performs well on common attributes but is constrained to \"rainbow\" editing and tends to generate artifacts outside the segmentation mask. Our method works seamlessly with masks, producing results with rich details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.639, + 0.909, + 0.745 + ], + "angle": 0, + "content": "In Tab. 1, We present a quantitative comparison. Our method outperforms the baselines in terms of CLIP Similarity and CLIP Directional Similarity. Notably, Dire Sim is not sensitive with the editing quality, much focus on the instruction attributes. We conducted a user study with a survey of 55 participants to evaluate the editing quality. The results show that our method received the most popular votes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.756, + 0.688, + 0.772 + ], + "angle": 0, + "content": "5.2. 2D Image Editing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.779, + 0.907, + 0.854 + ], + "angle": 0, + "content": "Dataset. To evaluate the effectiveness of our method, we conduct experiments on the PIE-Bench dataset proposed by PNPIv [19], which consists of 700 images with 9 editing types. Each image is annotated with the source and target prompts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Baselines. We compare our method with several classical editing methods based on DDIM [43] inversion, including P2P [11], PNP [46] and MasaCtrl [3]. For optimization-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16602" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.363, + 0.088, + 0.634, + 0.102 + ], + "angle": 0, + "content": "A photo of a man \\(\\rightarrow\\) Spider man with a mask" + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.105, + 0.25, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.104, + 0.404, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.104, + 0.561, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.104, + 0.717, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.105, + 0.874, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.362, + 0.206, + 0.646, + 0.219 + ], + "angle": 0, + "content": "A man wearing T-shirt with a pineapple pattern" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.219, + 0.252, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.219, + 0.409, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.219, + 0.565, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.566, + 0.219, + 0.721, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.722, + 0.219, + 0.879, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.439, + 0.328, + 0.56, + 0.341 + ], + "angle": 0, + "content": "Rainbow horn fossil" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.342, + 0.252, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.342, + 0.409, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.342, + 0.565, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.566, + 0.342, + 0.721, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.722, + 0.342, + 0.878, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.387, + 0.435, + 0.61, + 0.447 + ], + "angle": 0, + "content": "A tree stump with some leaves on fire" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.448, + 0.252, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.158, + 0.53, + 0.232, + 0.542 + ], + "angle": 0, + "content": "Input Views" + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.448, + 0.409, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.308, + 0.531, + 0.385, + 0.542 + ], + "angle": 0, + "content": "InstructN2N" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.448, + 0.565, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.452, + 0.531, + 0.546, + 0.543 + ], + "angle": 0, + "content": "GaussianEditor" + }, + { + "type": "image", + "bbox": [ + 0.566, + 0.448, + 0.721, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.635, + 0.53, + 0.668, + 0.542 + ], + "angle": 0, + "content": "DGE" + }, + { + "type": "image", + "bbox": [ + 0.722, + 0.448, + 0.878, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.788, + 0.531, + 0.819, + 0.542 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.543, + 0.907, + 0.571 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparisons with related works. SDS demonstrates outstanding performance in effectively preserve source structure in the modified region." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.483, + 0.64 + ], + "angle": 0, + "content": "based editing method, we compare with NT [34] and StyleD [27]. Besides, we report the comparison with DT [19]. Further, we compare with DDS [10] and its extended method CDS [35]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.484, + 0.779 + ], + "angle": 0, + "content": "Evaluation Metrics. We follow DT [19] which uses several metrics to evaluate our method. We use the Structure Distance assessed by DINO score [4] to evaluate the structure distance between original and edited images. We also introduce several metrics to evaluating the background preservation, which includes LPIPS [55] and MSE. Besides, we introduce CLIP Similarity [48] to evaluate the text-image consistency between edited images and corresponding target editing text prompts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Results. We present a qualitative comparison of our method with competitors in Fig. 7. Our method generates images that are more aligned with the target prompts and preserve the source structure. In \"blue butterfly\", ours successfully changes the color of the butterfly to blue, while DDS [10] and CDS [35] generate similar color from source. Especially, ours method successfully changes the style of the image and generates appealing results, which is challenging for DDS-" + }, + { + "type": "table_caption", + "bbox": [ + 0.547, + 0.578, + 0.871, + 0.591 + ], + "angle": 0, + "content": "Table 2. Quantitative evaluation in PIE-Bench dataset." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.592, + 0.92, + 0.751 + ], + "angle": 0, + "content": "
MethodDistance×103↓LPIPS×103↓MSE×104↓CLIP↑
DDIM + P2P69.43208.80219.8825.01
DDIM + PNP28.22113.4683.6425.41
DDIM + MasaCtrl28.38106.6286.9723.96
NT + P2P13.4460.6735.8624.75
StyleD + P2P11.6566.1038.6324.78
DT + P2P11.6554.5532.8625.02
DDS14.7450.5845.0925.86
DDS + CDS7.1533.1425.2924.96
Ours28.1382.4386.6426.94
Ours + CDS6.9032.1524.2125.12
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.762, + 0.907, + 0.852 + ], + "angle": 0, + "content": "based methods. Compared to optimization-based methods, NT [34] preserves the general source structure during the inversion process, but tends to discard some content as evident in the distortion of the girl's fingers in Fig. 7. Additionally, due to limitations in the editing methods, the editing results are unsuccessful." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In Tab. 2, we present a quantitative comparison. Our method strikes a balance between structure distance and editability. Notably, we observe that the distance is much" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16603" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.153, + 0.092, + 0.421, + 0.103 + ], + "angle": 0, + "content": "A blue and white bird \\(\\rightarrow\\) butterfly sits on a branch" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.105, + 0.169, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.171, + 0.105, + 0.248, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.25, + 0.105, + 0.326, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.105, + 0.403, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.404, + 0.105, + 0.481, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.165, + 0.465, + 0.177 + ], + "angle": 0, + "content": "Kids crayon drawing of a man with a long beard and a long sword" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.178, + 0.171, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.172, + 0.178, + 0.249, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.178, + 0.326, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.178, + 0.403, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.404, + 0.178, + 0.481, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.238, + 0.465, + 0.251 + ], + "angle": 0, + "content": "Black and white sketch of a young girl with painted hands and face" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.253, + 0.171, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.172, + 0.253, + 0.249, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.253, + 0.326, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.253, + 0.403, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.404, + 0.253, + 0.481, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.312, + 0.457, + 0.324 + ], + "angle": 0, + "content": "A monkey \\(\\rightarrow\\) man wearing colorful goggles and a colorful scarf" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.325, + 0.171, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.172, + 0.325, + 0.249, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.325, + 0.326, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.324, + 0.403, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.404, + 0.324, + 0.481, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.385, + 0.162, + 0.396 + ], + "angle": 0, + "content": "Source" + }, + { + "type": "text", + "bbox": [ + 0.198, + 0.385, + 0.231, + 0.396 + ], + "angle": 0, + "content": "DDS" + }, + { + "type": "text", + "bbox": [ + 0.267, + 0.385, + 0.299, + 0.396 + ], + "angle": 0, + "content": "CDS" + }, + { + "type": "text", + "bbox": [ + 0.335, + 0.385, + 0.39, + 0.396 + ], + "angle": 0, + "content": "NT+P2P" + }, + { + "type": "text", + "bbox": [ + 0.427, + 0.385, + 0.458, + 0.396 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.397, + 0.483, + 0.424 + ], + "angle": 0, + "content": "Figure 7. Comparison of different editing methods on various objects and styles." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.432, + 0.485, + 0.644 + ], + "angle": 0, + "content": "lower when no editing occurs, which is particularly visible in DDS-based methods applied to style editing. Our methods achieving better editing score but with a slightly higher structure distance. In terms of precise structure preservation, when combined with CDS [35], it achieves good preservation of the un-edited areas. Our full model achieves the best performance in the CLIP Similarity metric, demonstrating the effectiveness of our prompt enhancement branch. While CDS excels in preserving unedited regions, it suffers from the inferior editability of DDS-based methods. Optimization-based methods [27, 34] refine the inversion process, achieving excellent performance in structure preservation. However, they struggle with editing methods (like P2P), resulting in limited editability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.653, + 0.255, + 0.667 + ], + "angle": 0, + "content": "5.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.735 + ], + "angle": 0, + "content": "In this section, we conduct an ablation experiment to analyze different choices in our SSD. Due to space limitations, we first present a qualitative evaluation in the main text. Please refer to the Supp. for quantitative evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.885 + ], + "angle": 0, + "content": "The effectiveness of cross-trajectory. In Sec. 4.1, we have analyzed the necessity of cross-trajectory. This term makes the optimization process more stable and provides the source content regularization in ours design, which is also the key difference between ours and Classifier Score Distillation(CSD) [52]. In Fig. 1 and Fig. 5, we present the comparison of the results with and without cross-trajectory. The results show that the cross-trajectory term can provide the direction of generating high-quality images. Please refer to the supplement for more details." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The effectiveness of prompt-enhancement. The enhance-" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.089, + 0.643, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.165, + 0.629, + 0.178 + ], + "angle": 0, + "content": "w/o ID Regular" + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.089, + 0.771, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.656, + 0.165, + 0.762, + 0.178 + ], + "angle": 0, + "content": "ID Regular \\(\\times 1.0\\)" + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.089, + 0.898, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.783, + 0.165, + 0.889, + 0.178 + ], + "angle": 0, + "content": "ID Regular \\(\\times 2.0\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.18, + 0.907, + 0.249 + ], + "angle": 0, + "content": "Figure 8. Effect of source latent regularization. In most experiments, the source ID term helps prevent partial gradient explosion. In the left image, the yellow arrow highlights an irregular color. As the weight of the ID term increases, the color becomes more regular, however, the spider on the character's chest is affected." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.907, + 0.349 + ], + "angle": 0, + "content": "ment of the target prompt branch is another key component in our method, which is designed to improve the editability aligned with the target prompt in 2D-image task. In Fig. 7, we observe a clear distinction from DDS in style editing. The results show that the prompt-enhancement term effectively overcomes the challenging from style editing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.349, + 0.907, + 0.456 + ], + "angle": 0, + "content": "The effectiveness of ID regularization. ID regularization is designed to ensure stable optimization in 3DGS. In Fig. 8, we compare results with and without ID regularization. The area marked by the yellow arrow highlights its effect in 3D scene editing. However, excessive ID regularization may constrain editing quality by limiting certain attributes, presenting a trade-off in our design." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.468, + 0.903, + 0.485 + ], + "angle": 0, + "content": "6. Conclusions, Limitations, and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.907, + 0.6 + ], + "angle": 0, + "content": "In this work, we propose a novel method for text-guided image editing, capable of handling both 3D scenes and 2D images. Our approach is built on a score distillation framework that leverages the powerful priors of diffusion models. For editing tasks, we design an effective optimization strategy that produces high-quality results aligned with target prompts while ensuring stable and consistent optimization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.909, + 0.765 + ], + "angle": 0, + "content": "Our method achieves state-of-the-art performance in both 3D scene and 2D image editing, delivering realistic edits with excellent preservation of the original content. It demonstrates strong adaptability to various editing tasks and target prompts, making it a robust solution for complex scenarios. However, while effective, the optimization process is relatively time-intensive compared to recent one-step methods [50] or few-step approaches [7]. Future work could explore integrating advanced techniques such as LCM [30] or SD-Turbo [42], which show potential for accelerating the optimization process [45]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Acknowledgment: This project is supported by the National Natural Science Foundation of China (62125201, U24B20174, 62102381), the Guangdong Natural Science Funds for Distinguished Young Scholars (Grant 2023B1515020097), the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No.: AISG3-GV-2023-011), the Singapore Ministry of Education AcRF Tier 1 Grant (Grant No.: MSS25C004), the Lee Kong Chian Fellowships." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16604" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.486, + 0.17 + ], + "angle": 0, + "content": "[1] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, pages 5470–5479, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.171, + 0.485, + 0.213 + ], + "angle": 0, + "content": "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, pages 18392-18402, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.213, + 0.484, + 0.269 + ], + "angle": 0, + "content": "[3] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In ICCV, pages 22560-22570, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.27, + 0.484, + 0.323 + ], + "angle": 0, + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.326, + 0.484, + 0.367 + ], + "angle": 0, + "content": "[5] Minghao Chen, Iro Laina, and Andrea Vedaldi. Dge: Direct gaussian 3d editing by consistent multi-view editing. In ECCV, pages 74-92, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.484, + 0.436 + ], + "angle": 0, + "content": "[6] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. In CVPR, pages 21476-21485, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.484, + 0.478 + ], + "angle": 0, + "content": "[7] Gilad Deutch, Rinon Gal, Daniel Garibi, Or Patashnik, and Daniel Cohen-Or. Turboedit: Text-based image editing using few-step diffusion models. In SIGGRAPHA, 2024. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.48, + 0.484, + 0.519 + ], + "angle": 0, + "content": "[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 34:8780-8794, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.522, + 0.484, + 0.563 + ], + "angle": 0, + "content": "[9] Ayaan Haque, Matthew Tancik, Alexei Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. In ICCV, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.565, + 0.484, + 0.605 + ], + "angle": 0, + "content": "[10] Amir Hertz, Kfir Aberman, and Daniel Cohen-Or. Delta denoising score. In ICCV, pages 2328-2337, 2023. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.607, + 0.484, + 0.647 + ], + "angle": 0, + "content": "[11] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In ICLR, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.649, + 0.482, + 0.676 + ], + "angle": 0, + "content": "[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshop, 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.677, + 0.484, + 0.717 + ], + "angle": 0, + "content": "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.719, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[14] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[15] Yukun Huang, Jianan Wang, Yukai Shi, Boshi Tang, Xianbiao Qi, and Lei Zhang. Dreamtime: An improved optimization strategy for diffusion-guided 3d generation. In ICLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.482, + 0.859 + ], + "angle": 0, + "content": "[16] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. In CVPR, pages 12469-12478, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.484, + 0.899 + ], + "angle": 0, + "content": "[17] Aapo Hyvarinen and Peter Dayan. Estimation of non-normalized statistical models by score matching. JMLR, 6(4), 2005. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.486, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.147 + ], + "angle": 0, + "content": "[18] Yutao Jiang, Yang Zhou, Yuan Liang, Wenxi Liu, Jianbo Jiao, Yuhui Quan, and Shengfeng He. Diffuse3d: Wide-angle 3d photography via bilateral diffusion. In ICCV, pages 8998-9008, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.906, + 0.191 + ], + "angle": 0, + "content": "[19] Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. Pnp inversion: Boosting diffusion-based editing with 3 lines of code. In ICLR, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.908, + 0.235 + ], + "angle": 0, + "content": "[20] Oren Katzir, Or Patashnik, Daniel Cohen-Or, and Dani Lischinski. Noise-free score distillation. In ICLR, 2024. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.238, + 0.906, + 0.279 + ], + "angle": 0, + "content": "[21] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splattering for real-time radiance field rendering. ACM TOG, 42(4):1-14, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.282, + 0.908, + 0.336 + ], + "angle": 0, + "content": "[22] Jiwook Kim, Seonho Lee, Jaeyo Shin, Jiho Choi, and Hyunjung Shim. Dreamcatalyst: Fast and high-quality 3d editing via controlling editability and identity preservation. In ICLR, 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.34, + 0.908, + 0.38 + ], + "angle": 0, + "content": "[23] Hubert Kompanowski and Binh-Son Hua. Dream-in-style: Text-to-3d generation using stylized score distillation. In 3DV, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.384, + 0.908, + 0.423 + ], + "angle": 0, + "content": "[24] Juil Koo, Chanho Park, and Minhyuk Sung. Posterior distillation sampling. In CVPR, pages 13352-13361, 2024. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.427, + 0.906, + 0.482 + ], + "angle": 0, + "content": "[25] Duong H Le, Tuan Pham, Aniruddha Kembhavi, Stephan Mandt, Wei-Chiu Ma, and Jiasen Lu. Preserving identity with variational score for general-purpose 3d editing. arXiv preprint arXiv:2406.08953, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.485, + 0.906, + 0.54 + ], + "angle": 0, + "content": "[26] Kehan Li, Yanbo Fan, Yang Wu, Zhongqian Sun, Wei Yang, Xiangyang Ji, Li Yuan, and Jie Chen. Learning pseudo 3d guidance for view-consistent texturing with 2d diffusion. In ECCV, pages 18-34, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.543, + 0.908, + 0.598 + ], + "angle": 0, + "content": "[27] Senmao Li, Joost Van De Weijer, Taihang Hu, Fahad Shahbaz Khan, Qibin Hou, Yaxing Wang, and Jian Yang. Stylediffusion: Prompt-embedding inversion for text-based editing. arXiv preprint arXiv:2303.15649, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.601, + 0.906, + 0.655 + ], + "angle": 0, + "content": "[28] Haofeng Liu, Chenshu Xu, Yifei Yang, Lihua Zeng, and Shengfeng He. Drag your noise: Interactive point-based editing via diffusion semantic propagation. In CVPR, pages 6743-6752, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.658, + 0.908, + 0.713 + ], + "angle": 0, + "content": "[29] Yuqin Lu, Bailin Deng, Zhixuan Zhong, Tianle Zhang, Yuhui Quan, Hongmin Cai, and Shengfeng He. 3d snapshot: Invertible embedding of 3d neural representations in a single image. 46(12):11524-11531, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.716, + 0.908, + 0.77 + ], + "angle": 0, + "content": "[30] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.908, + 0.828 + ], + "angle": 0, + "content": "[31] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In CVPR, pages 12663–12673, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.908, + 0.899 + ], + "angle": 0, + "content": "[32] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM TOG, 38(4): 1-14, 2019. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16605" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[33] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, pages 405-421, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.204 + ], + "angle": 0, + "content": "[34] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, pages 6038–6047, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.206, + 0.484, + 0.26 + ], + "angle": 0, + "content": "[35] Hyelin Nam, Gihyun Kwon, Geon Yeong Park, and Jong Chul Ye. Contrastive denoising score for text-guided latent diffusion image editing. In CVPR, pages 9192-9201, 2024. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.262, + 0.483, + 0.303 + ], + "angle": 0, + "content": "[36] JangHo Park, Gihyun Kwon, and Jong Chul Ye. ED-NeRF: Efficient text-guided editing of 3d scene with latent space nef. In ICLR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.484, + 0.346 + ], + "angle": 0, + "content": "[37] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.484, + 0.403 + ], + "angle": 0, + "content": "[38] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2): 3, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.484, + 0.46 + ], + "angle": 0, + "content": "[39] Jingjing Ren, Wenbo Li, Haoyu Chen, Renjing Pei, Bin Shao, Yong Guo, Long Peng, Fenglong Song, and Lei Zhu. Ultrapixel: Advancing ultra high-resolution image synthesis to new peaks. In NeurIPS, pages 111131-111171, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.484, + 0.516 + ], + "angle": 0, + "content": "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.518, + 0.484, + 0.599 + ], + "angle": 0, + "content": "[41] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, pages 36479-36494, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.602, + 0.484, + 0.643 + ], + "angle": 0, + "content": "[42] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In ECCV, pages 87-103, 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.645, + 0.483, + 0.673 + ], + "angle": 0, + "content": "[43] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.484, + 0.729 + ], + "angle": 0, + "content": "[44] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.731, + 0.483, + 0.772 + ], + "angle": 0, + "content": "[45] Feng Tian, Yixuan Li, Yichao Yan, Shanyan Guan, Yanhao Ge, and Xiaokang Yang. Postedit: Posterior sampling for efficient zero-shot image editing. In ICLR, 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.774, + 0.484, + 0.815 + ], + "angle": 0, + "content": "[46] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In CVPR, pages 1921-1930, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.484, + 0.87 + ], + "angle": 0, + "content": "[47] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In NeurIPS, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[48] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva:" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.12 + ], + "angle": 0, + "content": "Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.908, + 0.177 + ], + "angle": 0, + "content": "[49] Jing Wu, Jia-Wang Bian, Xinghui Li, Guangrun Wang, Ian Reid, Philip Torr, and Victor Adrian Prisacariu. Gaussctrl: Multi-view consistent text-driven 3d gaussian splatting editing. In ECCV, pages 55-71, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.907, + 0.219 + ], + "angle": 0, + "content": "[50] Sihan Xu, Yidong Huang, Jiayi Pan, Ziqiao Ma, and Joyce Chai. Inversion-free image editing with natural language. In CVPR, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.906, + 0.262 + ], + "angle": 0, + "content": "[51] Yangyang Xu, Shengfeng He, Kwan-Yee K Wong, and Ping Luo. Rigid: Recurrent gan inversion and editing of real face videos and beyond. IJCV, 133(6):3437-3455, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.907, + 0.304 + ], + "angle": 0, + "content": "[52] Xin Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Song-Hai Zhang, and Xiaojuan Qi. Text-to-3d with classifier score distillation. In ICLR, 2024. 1, 2, 3, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.908, + 0.361 + ], + "angle": 0, + "content": "[53] Yuyang Yu, Bangzhen Liu, Chenxi Zheng, Xuemiao Xu, Huaidong Zhang, and Shengfeng He. Beyond textual constraints: Learning novel diffusion conditions with fewer examples. In CVPR, pages 7109-7118, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.906, + 0.405 + ], + "angle": 0, + "content": "[54] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836-3847, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.407, + 0.907, + 0.447 + ], + "angle": 0, + "content": "[55] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.907, + 0.503 + ], + "angle": 0, + "content": "[56] Chenxi Zheng, Yihong Lin, Bangzhen Liu, Xuemiao Xu, Yongwei Nie, and Shengfeng He. Recdreamer: Consistent text-to-3d generation via uniform score distillation. In ICLR, 2025. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.906, + 0.545 + ], + "angle": 0, + "content": "[57] Xingchen Zhou, Ying He, F Richard Yu, Jianqiang Li, and You Li. RePaint-NeRF: Nerf editing via semantic masks and diffusion models. In IJCAI, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.908, + 0.588 + ], + "angle": 0, + "content": "[58] Junzhe Zhu, Peiye Zhuang, and Sanmi Koyejo. HIFA: High-fidelity text-to-3d generation with advanced diffusion guidance. In ICLR, 2024. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.588 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16606" + } + ] +] \ No newline at end of file diff --git a/2025/Stable Score Distillation/62930789-4f50-45e0-b2ce-8d1494eced86_origin.pdf b/2025/Stable Score Distillation/62930789-4f50-45e0-b2ce-8d1494eced86_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..85af6bfdbded0710e82577f8a78c203ddfbf476b --- /dev/null +++ b/2025/Stable Score Distillation/62930789-4f50-45e0-b2ce-8d1494eced86_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:09d84146bc5e0d8917702f0e9dda06cc8b99e61a69fcbe2f46da7157f6bfef49 +size 12627813 diff --git a/2025/Stable Score Distillation/full.md b/2025/Stable Score Distillation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..09c1ceaf93f218b47f3f9d16d44065ee3561ce15 --- /dev/null +++ b/2025/Stable Score Distillation/full.md @@ -0,0 +1,453 @@ +# Stable Score Distillation + +Haiming Zhu $^{1}$ , Yangyang Xu $^{2*}$ , Chenshu Xu $^{1}$ , Tingrui Shen $^{3}$ , Wenxi Liu $^{4}$ , Yong Du $^{5}$ , Jun Yu $^{2}$ , Shengfeng He $^{1*}$ + +$^{1}$ Singapore Management University $^{2}$ Harbin Institute of Technology (Shenzhen) + +$^{3}$ South China University of Technology $^{4}$ Fuzhou University $^{5}$ Ocean University of China + +![](images/215e2790b3793d17d15859ef9997fc328f48e5c57b8796648d55ed6cb98b5ec8.jpg) + +![](images/94379b3d762a60744f9d4afc109879d218b487c1a02403c497c68dad6352031b.jpg) +(a) Original Views + +![](images/cc09d0cc8b5fac1f29b5496fddb5464f1e45bcdda3d2c9bca51cdbe82300ee06.jpg) + +![](images/f83feac7a9b6dbc14a1b6005bc55d6e14b6e47965ba8113917b44c124f3e85fb.jpg) +(b) CSD[52] + +![](images/b69aca6b82b03c70b0636af851357aafc4d4644a62e8db6343325dc46726ffbd.jpg) + +![](images/12492b653667c289aac1cee6e4554b4d08a8b414bbfc37ca5e0cd529fd795f9c.jpg) +(c) DDS[10] +Figure 1. We propose Stable Score Distillation (SSD), a method that improves text-guided editing by preserving original content structure and enhancing realism in edited results. The first column shows the original views, while the remaining columns display comparative results with existing score distillation methods on the SD model. SSD demonstrates superior performance, maintaining the integrity of the original content and producing more realistic edits. + +![](images/58b1d4ad15639ae1d16011af1c25387c59da9c38aa0ab9a43a7f9145b9405fa9.jpg) + +![](images/c9e7d6162ad4fe9abefbc17693f4d165e0c6bd0bb674c6968388be0193d2ddc9.jpg) +(d) Ours + +# Abstract + +Text-guided image and 3D editing have advanced with diffusion-based models, yet methods like Delta Denoising Score often struggle with stability, spatial control, and editing strength. These limitations stem from reliance on complex auxiliary structures, which introduce conflicting optimization signals and restrict precise, localized edits. We introduce Stable Score Distillation (SSD), a streamlined framework that enhances stability and alignment in the editing process by anchoring a single classifier to the source prompt. Specifically, SSD utilizes Classifier-Free Guidance (CFG) equation to achieve cross-prompt alignment, and introduces a constant term null-text branch to stabilize the optimization process. This approach preserves the original content's structure and ensures that editing trajectories are closely aligned with the source prompt, enabling smooth, prompt-specific modifications while maintaining coherence in surrounding regions. Additionally, SSD incorporates a prompt enhancement branch to boost editing strength, particularly for style transformations. Our method achieves state-of-the-art results in 2D and 3D editing tasks, including NeRF and text-driven style edits, with faster conver + +gence and reduced complexity, providing a robust and efficient solution for text-guided editing. Code is available at: https://github.com/Alex-Zhu1/SSD. + +# 1. Introduction + +Text-based image generation has achieved remarkable progress, particularly with the advent of diffusion models [13, 28, 40, 41, 43, 56]. These models leverage strong priors to produce high-quality images, facilitating significant advances in text-to-3D generation [37, 47, 58]. Moreover, text-guided 3D editing has enabled intricate modifications to shape and texture [26, 31], supporting flexible and precise 3D scene manipulation. + +Unlike generation tasks that create new content, editing tasks aim to modify specific elements within an image while preserving surrounding areas [51]. However, directly applying methods like Score Distillation Sampling (SDS) to editing tasks can yield undesired effects, such as blurring across the image. This arises because SDS optimizes globally to the prompt, affecting regions beyond the targeted area [10, 20]. DDS [10] addresses this by introducing a dual-branch architecture, pairing the source image with its description to leverage the model's inherent bias and isolate specific prompt changes. Further, CSD [52] achieved + +![](images/113e4cf3086202d6189ea23ab0e6c94a7fef3b378152c624fb16ea51a380b2b7.jpg) +Figure 2. Illustration of three distillation-based approaches. Note that we assume a 2-step optimization process for illustration, where the subscript $t$ represents the iteration number. DDS utilizes the source branch to obtain initial latent $Z_0^*$ , while CSD employs two classifiers to derive $Z_1'$ and $Z_1''$ for cross-prompt editing. Our SSD method designs a CFG classifier to determine the cross-prompt editing, introduces the null-text branch as the initial latent $Z_0^*$ , and further constructs the cross-trajectory term (see Sec. 4.1) for stable optimization. + +![](images/c5589cd515662891cef249906174257aa7b3542a2aa1bf446dcc8e922b46d715.jpg) + +![](images/1b693ad852678cb0b00bdc1dd1b74f854c74679f7328eae1298787e3f4cc59b8.jpg) + +![](images/dad6abc620176c768dfa99228a5f05c5ab73f7b654a77aa876f396f92fc6ba21.jpg) + +![](images/479da4509d074ccb14bce7ce3df49b68e806012d1bbbe6aecabfaea49663ba2a.jpg) +Figure 3. The optimization process of DDS and our SSD. SSD preserves the source structure effectively during optimization iterations, while DDS cannot preserve it effectively. + +scene editing by incorporating a classifier component within Classifier-Free Guidance (CFG) [12] to refine the prediction score by applying the classifier to both the source and target prompts. NFSD [20] further decomposes the CFG score, highlighting the classifier as the primary driver of prompt direction. + +Despite their success, we argue that current distillation-based approaches face inherent limitations, such as low editing quality and loss of source content. As shown in Fig. 2, DDS [10] relies on the source branch to remove model bias, but lacks the explicit guidance to preserve the source content [24] during optimization. As shown in Fig. 1c, DDS changes man's clothes during editing his faces. Additionally, although introducing source prompt components is intended to improve prompt specificity [19], it can amplify noise and introduce overlapping objectives that hinder stable convergence. This results in artifacts or unintended variations, especially in the unedited regions. Correspondingly, CSD [52] utilizes dual classifiers to refine the prompt editing direction (see Fig. 2). However, it lacks the explicit source preservation to restrict edits precisely to the target areas. As shown in Fig. 1b, this causes the structure deformation and annoying artifacts in the edited regions. + +Our insights into these limitations lead to two key observations: (1) Cross-prompt: a single classifier, providing + +the editing direction from source prompt to target, and (2) Cross-trajectory: stability in the editing process can be achieved by aligning the editing direction closely with the structure of the source content. + +In this paper, we propose Stable Score Distillation (SSD), a streamlined approach for stable and precise text-guided editing. To achieve a smooth editing direction, we employ the CFG equation for both the source and target prompts, ensuring a gradual transition of the original contextual texture as the model adapts to the specified changes. This approach contrasts with DDS [10], as it eliminates the need for a auxiliary source branch, enabling our method to focus editing gradients precisely within target regions while ensuring a stable transition, as illustrated in Fig. 3. Moreover, for aligning the editing direction with the source prompt, facilitating smoother and more controlled progression toward the target prompt, we design an cross-trajectory strategy to ensure that edits respect the original structure, supporting subtle and stable transformations within designated areas. While NFSD [20] utilizes negative-branch and DDS utilizes source branch to enhance output clarity, as shown in SSD in Fig. 2, we introduce a null-text branch aligned with the "no-edit" direction to integrate a "reconstruction" term to explicitly enforce source content preservation, which enhances consistency and produces reliable edits across diverse tasks. Based on above designs, our framework remains streamlined and efficient, achieving both precision and stability without the complexity of additional components. + +Our framework integrates seamlessly into existing DDS-based editing pipelines and applications, such as text-driven NeRF editing [22, 24, 36] and 2D image editing [35]. Notably, our approach's "clear" editing direction preserves source content, making a carefully designed identity regularization [22] unnecessary. Moreover, standard DDS methods often lack sufficient editing strength, resulting in minimal or negligible changes in output, particularly in style editing [23]. Our approach, with its streamlined and stable framework, allows for the seamless integration of a prompt enhancement branch to amplify editing capability. + +With these improvements, our method achieves faster and + +more effective edits during optimization, remains compatible with the Stable Diffusion Model [40] without requiring LoRA [14] or fine-tuning, and integrates effectively with Instructpix2pix [2]. Additionally, by incorporating nonincreasing timestep sampling [15], we accelerate convergence, reducing the required iterations to approximately 3,000 for NeRF [33] and 1,500 for Gaussian splatting [21]. In summary, our contributions are as follows: + +- We introduce a novel editing framework, Stable Score Distillation, that leverages a single, anchored classifier to achieve targeted and stable edits in 3D scene editing. +- We introduce a prompt enhancement strategy, effectively improve the prompt-alignment, especially style editing in 2D-image editing. +- We demonstrate the effectiveness of our approach across NeRF-editing and image-editing tasks, achieving state-of-the-art results with a streamlined and efficient framework. + +# 2. Related Work + +# 2.1. Diffusion Models + +Diffusion models [13, 39, 43, 53] have made significant advancements in generating diverse and high-fidelity images. Starting from a gaussian noise, diffusion models can predict the noise-less sample at each time step, until finally obtaining clear samples. Commonly, the denoising process can utilize U-net model to predict the noise. Some works [13, 44] have observed that is proportional to the predicted score function [17] of the smoothed density. Thus, intuitively, taking steps in the direction of the score function gradually moves the sample towards the data distribution. + +To generate images aligned with a target prompt, guidance is typically introduced to explicitly control the weight assigned to the conditioning information. The popular guidance methods include Classifier Guidance [8] and Class-free Guidance (CFG) [12]. While the former relies on a separately learned classifier, the latter directly introduces null-text samples to the model. CFG modifies the score function to steer the process towards regions with a higher ratio of conditional density to the unconditional one. However, it has been observed that CFG trades sample fidelity for diversity [12]. Based on the insights gained from the decomposition of the CFG equation, we propose a novel Stable Score Distillation (SSD) method to guide the SDS optimization process. + +# 2.2. Score Distillation Sampling (SDS) + +Benefit from the data scale-law, diffusion model [38, 40, 41] achieve high-quality image generation and text-to-image generation. Specifically, Score Distillation Sampling (SDS) [37] leveraging the priors of pre-trained text-to-image models to facilitate text-conditioned generation in 3D content generation. Specifically, SDS is an optimization approach that updates the rendering parameter towards the image distribu + +tion of diffusion models by enforcing the noise prediction on noisy rendered images to match sampled noise. While SDS provides an elegant mechanism for leveraging pretrained text-to-image models, SDS-generated results often suffer from oversaturation and lack of fine realistic details. VSD [47] proposed a particle-based optimization framework that treats the 3D parameter as a random variable of target distribution. Furthermore, by regarding SDS as a reverse diffusion process, decreasing timesteps sampling [15, 58] to imitate the diffusion reverse sampling, which can improve the quality of the generated 3D assets. + +In image editing, Delta Denoising Score (DDS) [10] found that Score Distillation Sampling (SDS) introduces noticeable artifacts and over-smoothing in edited images due to inherent bias. To mitigate this bias, DDS employs a subtraction of two SDS scores of the source and target images to obtain a delta score, which is then used to guide the optimization process. + +# 2.3. Text-Driven 3D-Scene Editing + +Text-driven 3D scene editing has been a popular research topic [18, 29, 56]. IN2N [9] proposed a Iterative Dataset Update method that can edit 3D scenes from text descriptions. By leveraging advancements in 2D diffusion editing techniques, notably InstructP2P [2] and ControlNet [54], GaussianEditor[6] and GaussCtrl [49] utilize edit multi-view images latent to optimize the 3D scene. We consider utilize score distillation to guide the 3D scene editing, which is more flexible and efficient for the text-driven 3D scene editing. + +Building upon the foundational SDS loss introduced by DreamFusion [37], some work has explored SDS loss in the text-driven 3D scene editing. RePaint-NeRF [57] has advanced the application of SDS in 3D editing by integrating a semantic mask to guide and constrain modifications within the background elements. CSD [52] utilize two classifiers to achieve editing. In a similar vein, ED-NeRF [36] has introduced an enhanced loss function specifically designed for 3D editing tasks. PDS [24] proposed a posterior distillation sampling to match stochastic latent [16]. Piva [25] fine-tuned the model while introducing a regularization term to preserve identity. Unfortunately, these methods are still limited to the long-time diffusion reverse sampling process, which is not suboptimal for the text-driven 3D scene editing. DreamCatalyst [22] extends the PDS optimization processing to ID-preserving and edit-ability based on decreasing timesteps sampling. + +Different from the above methods, ours firstly improve the DDS optimization process by introducing a single classifier, and further introduce a null-text branch to achieve a more stable and precise editing process. + +![](images/1c65990c7e27298d4db995adee328213b5e3471bf049410d212daf798bf05600.jpg) +Figure 4. The overview of SSD. Given the parameter 3D-model or image, SSD provides effective editing gradient to guide the optimization process. We utilize CFG equation between the predicted target noise $\epsilon_{\phi}(z_t,y,t)$ and source noise $\epsilon_{\phi}(z_t,\hat{y},t)$ , which generate the gradual editing direction. Furthermore, we introduce a null-text branch $\epsilon_{\phi}(\hat{z}_t,\emptyset ,t)$ to regularize the optimization process and achieve stable optimization. We further analyzing and decompose ours design term into three parts: cross-prompt, cross-trajectory, and prompt-enhance. + +# 3. Preliminary + +In this section, we first discuss existing optimization-based approaches to handle parametric images. Then, we will introduce our novel parametric image editing method in Section 4. + +# 3.1. Score Distillation Sampling + +Score Distillation Sampling (SDS) [37] is proposed to generate parametric images by leveraging the 2D prior of pretrained text-to-image diffusion models. Specifically, given a pretrained diffusion model $\epsilon_{\phi}$ , SDS optimizes a set of parameters $\theta$ of a differentiable parametric image generator $g$ , using the gradient of the loss $L_{\mathrm{SDS}}$ with respect to $\theta$ : + +$$ +\nabla_ {\theta} L _ {\mathrm {S D S}} = w (t) \left(\epsilon_ {\phi} \left(z _ {t} (x); y, t\right) - \epsilon\right) \frac {\partial x}{\partial \theta}, \tag {1} +$$ + +where $x = g(\theta)$ is an image rendered by $\theta$ , $z_{t}(x)$ is obtained by adding a Gaussian noise $\epsilon$ to $x$ corresponding to the $t$ -th timestep of the diffusion process, and $y$ is a condition to the diffusion model. As Noise-Free Score Distillation (NFSD) [20] has shown, the score $\epsilon_{\phi}(z_{t}(x);y,t)$ provides the direction in which this noised version of $x$ should be moved towards a denser region in the distribution of real images. + +# 3.2. Delta Distillation Sampling + +Although SDS get excellent generation ability, for editing task, an undesired component from the pretrained model, $\delta_{\mathrm{bias}}$ , interferes with the process and causes the image to become smooth and blurry in some parts [10]. Based on the observations that a matched source prompt $\hat{y}$ and source + +latent $\hat{z}_t$ can estimate the noisy direction $\delta_{\mathrm{bias}}$ , thus, the DDS method aims to remove the $\delta_{\mathrm{bias}}$ by introducing source branch, as shown in Eq. 2: + +$$ +\nabla_ {\theta} L _ {\mathrm {D D S}} = \left(\epsilon_ {\phi} ^ {c} \left(z _ {t}, y, t\right) - \epsilon_ {\phi} ^ {c} \left(\hat {z} _ {t}, \hat {y}, t\right)\right) \frac {\partial z}{\partial \theta}, \tag {2} +$$ + +where $\epsilon_{\phi}^{c}(z_{t},y,t)$ and $\epsilon_{\phi}^{c}(\hat{z}_{t},\hat{y},t)$ are pretrained model predictions $\epsilon$ , with the superscript $c$ indicating the CFG results. Thus, DDS pushes the optimized image into the direction of the target prompt without the interference of the noise component, namely, $\nabla_{\theta}L_{\mathrm{DDS}}\approx \delta_{\mathrm{ext}}$ . Obviously, $\nabla_{\delta_{\mathrm{ext}}}$ is contingent on classifier part from $\epsilon_{\phi}^{c}(z_{t},y,t)$ as discussed in CSD [52] and NFSD [20]. Note that in the following manuscript, we decompose the CFG results without the superscript $c$ and omit the timestep $t$ for simplicity. + +Further exploring prompt editing direction, CSD [52] method proposed a dual-classifier to refine the editing score and achieve more precise editing, as shown in Eq. 3: + +$$ +\begin{array}{l} \nabla_ {\theta} L _ {\mathrm {C S D}} = \left(w _ {a} \left(\epsilon_ {\phi} \left(z _ {t}, y\right) - \epsilon_ {\phi} \left(z _ {t}, \varnothing\right)\right) \right. \\ \left. - w _ {b} \left(\epsilon_ {\phi} \left(z _ {t}, \hat {y}\right) - \epsilon_ {\phi} \left(z _ {t}, \emptyset\right)\right)\right) \frac {\partial z}{\partial \theta}, \tag {3} \\ \end{array} +$$ + +while the $\epsilon_{\phi}(z_t,y,t)$ and $\epsilon_{\phi}(z_t,\hat{y},t)$ are current latent $z_{t}$ predictions for the target prompt $y$ and source prompt $\hat{y}$ , respectively. $w_{a}$ and $w_{b}$ are weights of classifiers. Simply put, CSD aims to refine the prompt editing direction by determining the difference between two classifiers, which can be regarded as a cross-prompt term. + +# 4. Method + +In 3D scene editing process, which requires consideration of both the target prompt and the original source content, + +we consider two key aspects: (1) smooth editing direction towards the target prompt (2) and editing results respect the original structure. Based on these, in this section, we introduce our novel editing framework Stable Score Distillation. + +# 4.1. Stable Score Distillation + +Firstly, we introduce the design of a cross-prompt editing direction. As discussed about CSD method in Sec. 3.2, the key role in cross-prompt editing is to provide a smooth transition from the source prompt to the target. As the CFG guidance [12] steers the process towards regions with a higher ratio of conditional density to the unconditional one, accordingly, we can modify the SDS score function, as shown below: + +$$ +G r a d = \epsilon_ {\phi} (z _ {t}, \hat {y}) + s (\epsilon_ {\phi} (z _ {t}, y) - \epsilon_ {\phi} (z _ {t}, \hat {y})), \tag {4} +$$ + +where $\epsilon_{\phi}(z_t,y)$ and $\epsilon_{\phi}(z_t,\hat{y})$ are pretrained model predictions. The scale factor $s$ is equal to control weight. + +Although the cross-prompt term provides a smooth texture transition in the edited region, we observed that the optimization process leads to abrupt structural changes, often resulting in artifacts and unappealing outcomes, similar to CSD in Fig. 1b. To address this, we introduce an additional regularization term to constrain the structural transition. Interestingly, as shown in Fig. 1c, DDS achieves better results than CSD by incorporating a source branch. However, DDS still lacks a mechanism to ensure the original structure remains intact, leading to modification on unedited regions. To address this, we introduce a null-text branch $\epsilon_{\phi}(\hat{z}_t,\emptyset)$ to regularize the optimization process, as shown in Eq. 5: + +$$ +L _ {\mathrm {s s d}} = \epsilon_ {\phi} (z _ {t}, \hat {y}) + s \left(\epsilon_ {\phi} (z _ {t}, y) - \epsilon_ {\phi} (z _ {t}, \hat {y})\right) - \epsilon_ {\phi} (\hat {z} _ {t}, \varnothing). \tag {5} +$$ + +Eq. 5 is ours Stable Score Distillation, and we can further decompose above equation into two parts, and the latter is regarded as a cross-trajectory term. + +$$ +L _ {\mathrm {s s d}} = \underbrace {w _ {p} \left(\epsilon_ {\phi} \left(z _ {t} , y\right) - \epsilon_ {\phi} \left(z _ {t} , \hat {y}\right)\right)} _ {\text {c r o s s - p r o m p t}} + \underbrace {w _ {t} \left(\epsilon_ {\phi} \left(z _ {t} , \hat {y}\right) - \epsilon_ {\phi} \left(\hat {z} _ {t} , \emptyset\right)\right)} _ {\text {c r o s s - t r a j e c t o r y}}, \tag {6} +$$ + +where the $w_{t}$ and $w_{p}$ control the strength of the cross-trajectory and cross-prompt, respectively. + +The cross-trajectory term can be interpreted as the distance between the transitions of two latents, ensuring that the original structure remains smooth and does not change abruptly (more details are provided in the supplementary material). In Fig. 5, we can see that the cross-trajectory term can provide a strong structure constraint ability, guiding the optimization process to preserve the source image structure. Specifically, when set $w_{t} = 0$ , the optimization process behaves similarly to the CSD[52] method, which fails to retain the original image structure. + +![](images/2d730c9f07c9dab167114f14a75e98f13ac70892f2ebebb8bf73cce3055a8358.jpg) +Figure 5. The effect of increasing the strength of the prompt enhancement term $(w_{e})$ and cross-trajectory term $(w_{t})$ , with the cross-prompt term fixed at 7.5. Both terms contribute to prompt-aligned results, while setting $(w_{t} = 0)$ leads to saturation and discard source content. + +# 4.2. Improving Prompt Alignment + +Although Eq. 5 can achieve gradual editing results, we found Eq. 5 have similar limitation with DDS [7], which have insufficient editing strength. The editing results neither get successful editing nor retain the source image structure, often leads to little or no change in the final. Benefit from the cross-prompt editing design as Eq. 5, we can add a target prompt enhancement branch to guide the optimization process. The target prompt alignment branch will provide the direction of the target prompt, as shown in Eq. 7: + +$$ +L _ {\text {a l i g n}} = w _ {e} \left(\epsilon_ {\phi} \left(z _ {t}, y\right) - \epsilon_ {\phi} \left(z _ {t}, \varnothing\right)\right), \tag {7} +$$ + +where $w_{e}$ is the prompt enhancement scale. As shown in Fig. 5, the synchronous scaling of both the cross-trajectory and prompt-enhancement terms results in effective visual editing outcomes. + +# 4.3. Source Latent Regularization + +Empirically, we found that directly using latent-space loss rather than pixel-level loss can lead to optimization difficulties in local regions of 3DGS. For example, the bright spots appearing in Fig. 5. To suppress the steep gradients in these areas, we incorporate ID regularization to guide the stable optimization process. Differ with PDS [24] use source latent $\hat{x}_0$ , we can use the noisy latent $\hat{x}_t$ to avoid partial exploding gradient, as shown in Eq. 8: + +$$ +L _ {\mathrm {I D}} = w (t) \cdot \left(x _ {t} - \hat {x} _ {t}\right), \tag {8} +$$ + +where the $w(t)$ is the iteration-dependent strength, designed as a decreasing function of $t$ . Notably, the $w(t)$ is not necessary to well-designed in our design. Our final loss function + +as shown in Eq. 9: + +$$ +L _ {\text {f i n a l}} = L _ {\text {s s d}} + L _ {\text {a l i g n}} + L _ {\text {I D}}. \tag {9} +$$ + +Based on the above design, we achieve a more prompt-aligned editing method, which integrates seamlessly into the Stable Diffusion Model [40] without requiring LoRA [14] or fine-tuning. Moreover, we will further introduce our method's connection with InstructPix2Pix [2]. + +# 4.4. Connecting with IP2P + +The final design of our method is shown in Eq 9. We found that ours edit gard provide new angle to understand about InstructP2P [2] one-step reverse sampling. + +$$ +\begin{array}{l} \epsilon_ {\theta} \left(z _ {t}, c _ {I}, c _ {T}\right) = \epsilon_ {\theta} \left(z _ {t}, \varnothing , \varnothing\right) \\ + s _ {I} \left(\epsilon_ {\theta} \left(z _ {t}, c _ {I}, \varnothing\right) - \epsilon_ {\theta} \left(z _ {t}, \varnothing , \varnothing\right)\right) \\ + s _ {T} \left(\epsilon_ {\theta} \left(z _ {t}, c _ {I}, c _ {T}\right) - \epsilon_ {\theta} \left(z _ {t}, c _ {I}, \varnothing\right)\right), \tag {10} \\ \end{array} +$$ + +where $c_{I}$ and $c_{T}$ are input-image and instruction prompt separately, $s_{I}$ and $s_{T}$ are the source image control and instruction prompt control strength. The Eq. 10 is the InstructP2P one-step reverse sampling, which can provide the direction of the target prompt. We can see that the InstructP2P is the simple version of our method, the middle term of Eq. 10 is cross-trajectory regularization, and the last term of Eq. 10 is ours cross-prompt term. Simply put, as analyzing the Eq. 5, subtracting the constant correction term $\epsilon_{\theta}(\hat{z}_t; \varnothing; \varnothing)$ is edit grad. Ours method reveals that apply DDS loss in the InstructP2P model can only editing branch, and don't have to provide the source branch. + +# 5. Experiments + +In this section, we conduct editing experiments across two types of parameterized images. Section 5.1 evaluates the effectiveness of our method on 3D Scenes Editing, and Section 5.2 evaluates the effectiveness of our method on 2D Image Editing. We also conduct ablation studies to analyze the effectiveness of ours components in Section 5.3. + +# 5.1.3D Scenes Editing + +Dataset. To evaluate the effectiveness of our method, we conduct experiments on the scenes from IN2N [9] and other real-world datasets, including LLFF [32] and Mip-Nerf360 [1]. + +Baselines. We compare our method with several state-of-the-art inversion methods. We use 3DGS [21] as the 3D representation, and compare our method with InstructNerf2Nerf [9], DDS [10], GS-Edit [6], and DGE [5]. For fairness, we implement the DDS version based on the official GS-Edit code. PDS [24] is designed for addition of objects to unspecified + +Table 1. Quantitative evaluations under 3D editing scenes. + +
MethodCLIP Sim.↑Sim Dir.↑User Study ↑
IN2N0.16760.070714.54%
DDS0.17800.04015.45%
GS-Editor0.17580.042914.54%
DGE0.17580.056323.63%
Ours0.18460.077341.81%
+ +regions, we will provide the comparison results in supplementary material. + +Evaluation Metrics. We follow common practice [5, 6, 9] to evaluate the effectiveness of our method. CLIP Similarity is to evaluate the alignment between the render images and the target prompts, i.e., the cosine similarity between the text and image embeddings encoded by CLIP. Specifically, follow DGE [5], randomly sample 20 camera poses to evaluate. CLIP Directional Similarity is to measure the editing effect, i.e., the cosine similarity between the image and text editing directions (target embeddings minus source embeddings). We evaluate all methods on 6 different scenes and 10 different prompts. + +Results. We begin by evaluating our method, starting with a qualitative assessment. In Fig. 6, we present a comparison of results with competing methods. Our approach generates more visually appealing images that are better aligned with the editing instructions. In contrast, methods based on the Iterative Dataset Update (IDU) strategy, such as IN2N [9] and GS-Editor [6], fail to produce the desired editing outcomes, resulting in blurrier or lower-fidelity reconstructions and noticeable artifacts. For example, in the scene of "Spider-Man with a mask", IN2N generates a mask with reduced fidelity, while GS-Editor produces a low-detail mask. In the multi-view consistency setup, DGE [5] performs well on common attributes but is constrained to "rainbow" editing and tends to generate artifacts outside the segmentation mask. Our method works seamlessly with masks, producing results with rich details. + +In Tab. 1, We present a quantitative comparison. Our method outperforms the baselines in terms of CLIP Similarity and CLIP Directional Similarity. Notably, Dire Sim is not sensitive with the editing quality, much focus on the instruction attributes. We conducted a user study with a survey of 55 participants to evaluate the editing quality. The results show that our method received the most popular votes. + +# 5.2. 2D Image Editing + +Dataset. To evaluate the effectiveness of our method, we conduct experiments on the PIE-Bench dataset proposed by PNPIv [19], which consists of 700 images with 9 editing types. Each image is annotated with the source and target prompts. + +Baselines. We compare our method with several classical editing methods based on DDIM [43] inversion, including P2P [11], PNP [46] and MasaCtrl [3]. For optimization- + +![](images/13ce1d72ac22be6ba8aaad7589261cefbec4998271c9b205f0bb383fe9db2855.jpg) + +![](images/f2805ad1f9a08b27f8a067ef9505fff4aad4faacd96bb493599b24da55750bc8.jpg) + +![](images/35a59f56e4c89d5ff0a5cd0b0f38718f9b1613950998792f6b07be31b594739c.jpg) + +![](images/f9cedc028a8c2f06f38b8fd7bb2eee568a7f3d437ab0c979f75559f4afed319b.jpg) + +![](images/878a0b066fa3d1b0f5600b04680a85667a95f9fce37bf619173f7e6750a5472b.jpg) + +![](images/0bd1a105b87c004a30efa34371eeaf0d431a6e913e790f2dc689c63a78c01e1e.jpg) + +![](images/6e56458c83a5e5663d8cba8b7d62b6ceca5b4909a68f04cb7c5ee27c331b792c.jpg) +A man wearing T-shirt with a pineapple pattern + +![](images/07bfd5534b16378060d25b3b2e02843f14eeb98b8e1f139bfab069accea1a362.jpg) + +![](images/d846415dd336729eeee72a7b26c7afd54e6ded75b8f11cc135cb1e7c7410e36f.jpg) + +![](images/8da663ffc801f6d2c6aa9c769369662d88bb36f0a503105e0c7923847be8f956.jpg) + +![](images/a1f6047a6c410edb2977679a0dd1d46fc3c84fe44b9194420f848cb970991b8e.jpg) + +![](images/e20c01a14c58cf70b717178cb863b1b7e64edf64dec5eb4e5eb6d28ef4aca5f8.jpg) + +![](images/f7b88149ae0d3f96d9139d4b733219f7502d881a0888483373573daef3935c90.jpg) +Rainbow horn fossil + +![](images/c7cb61146101236bca3aac01285297c694df1bdcb1cf9f604bfc7713e409c9a2.jpg) + +![](images/25227a4b1e6df7d5e6be676a438e4bd74903c1c649de2c99b23dafe131223ea8.jpg) + +![](images/18b19580bde98279c991ce488677c952c1566e7b4919d77ef6e2daa542bc7d2a.jpg) +Input Views +Figure 6. Qualitative comparisons with related works. SDS demonstrates outstanding performance in effectively preserve source structure in the modified region. + +![](images/2d130333d4f3face37b0455d5e73cf8928624983748be7a9bc200e89ec888418.jpg) +A tree stump with some leaves on fire +InstructN2N + +![](images/f40b944de1aed3a3a1b649e778e674236769dec39541cb27e4306f0ac65e3458.jpg) +GaussianEditor + +![](images/903d458fc60da3419c0cae26b709da119a308a7789f4201121dac0279e03ad19.jpg) +DGE + +![](images/935d95a40c02d21f1d8cdfcfc5ce9531ed9d7c5608c90f6ab0547459fea4b6aa.jpg) +Ours + +based editing method, we compare with NT [34] and StyleD [27]. Besides, we report the comparison with DT [19]. Further, we compare with DDS [10] and its extended method CDS [35]. + +Evaluation Metrics. We follow DT [19] which uses several metrics to evaluate our method. We use the Structure Distance assessed by DINO score [4] to evaluate the structure distance between original and edited images. We also introduce several metrics to evaluating the background preservation, which includes LPIPS [55] and MSE. Besides, we introduce CLIP Similarity [48] to evaluate the text-image consistency between edited images and corresponding target editing text prompts. + +Results. We present a qualitative comparison of our method with competitors in Fig. 7. Our method generates images that are more aligned with the target prompts and preserve the source structure. In "blue butterfly", ours successfully changes the color of the butterfly to blue, while DDS [10] and CDS [35] generate similar color from source. Especially, ours method successfully changes the style of the image and generates appealing results, which is challenging for DDS- + +Table 2. Quantitative evaluation in PIE-Bench dataset. + +
MethodDistance×103↓LPIPS×103↓MSE×104↓CLIP↑
DDIM + P2P69.43208.80219.8825.01
DDIM + PNP28.22113.4683.6425.41
DDIM + MasaCtrl28.38106.6286.9723.96
NT + P2P13.4460.6735.8624.75
StyleD + P2P11.6566.1038.6324.78
DT + P2P11.6554.5532.8625.02
DDS14.7450.5845.0925.86
DDS + CDS7.1533.1425.2924.96
Ours28.1382.4386.6426.94
Ours + CDS6.9032.1524.2125.12
+ +based methods. Compared to optimization-based methods, NT [34] preserves the general source structure during the inversion process, but tends to discard some content as evident in the distortion of the girl's fingers in Fig. 7. Additionally, due to limitations in the editing methods, the editing results are unsuccessful. + +In Tab. 2, we present a quantitative comparison. Our method strikes a balance between structure distance and editability. Notably, we observe that the distance is much + +A blue and white bird $\rightarrow$ butterfly sits on a branch + +![](images/cf61a5a6f51cf7b19f3024fd9ff691faa65d02314a8a45379864b4b8f6716f58.jpg) + +![](images/00c0fde9295dd7240d285eefe4980fad3185dc361640f4013af1364939c2fb0b.jpg) + +![](images/02fdae951c4a701eb43ef5c93133d3db311c1268b309d76794f8f3f1bfe3617a.jpg) + +![](images/5954546facdaca4fea820fb04998d791af9fedee2c1c451396c01012d381bced.jpg) + +![](images/e36d09ce7a7403e49e4865b73ecd12453b00bfbd0c912fbb787a148ada4e0f45.jpg) +Figure 8. Effect of source latent regularization. In most experiments, the source ID term helps prevent partial gradient explosion. In the left image, the yellow arrow highlights an irregular color. As the weight of the ID term increases, the color becomes more regular, however, the spider on the character's chest is affected. + +Kids crayon drawing of a man with a long beard and a long sword + +![](images/6844614c16f7a3153254083073477af46f2a162e8cd2b1871f19b12700d2f463.jpg) + +![](images/51f3b4407f6e9ff81716cc3792649a66a201a15cd0fc834b0f805d0538b9c5b3.jpg) + +![](images/bfa1737ca5a4b77528bf1ff09b1cf5a87757c8b1df837ba69eaa24f68b197eed.jpg) + +![](images/2b4433e62d7577275a90ed28b5e200c6c80dcaa7a94d5e709bc2a5f2d131d9d8.jpg) + +![](images/0e712de57c892101d9095f9a7bcd6c60c57744ef59ecdab4bae094786d536f5f.jpg) + +Black and white sketch of a young girl with painted hands and face + +![](images/9359ff7cb352ff71f1503aa5d09b4a88e6581ee50d98942dd30bd524fd6d949f.jpg) + +![](images/76395dace40a6df6c10cbf3fd0eee41fb5cd2353c620d1a0c34f7ee8645ec28d.jpg) + +![](images/c32718a6a8bcbb0e8d88f7671a3fd7d559862444a6a1321639464486c1371461.jpg) + +![](images/48e72db346cae2dde2118d0dc124f8cae46706aa7ff7f730f05f372ca883328e.jpg) + +![](images/e5450e97c47ce50415b03b58ab28db0b3632bdc842d45b5524c55ebae1f10cf5.jpg) + +A monkey $\rightarrow$ man wearing colorful goggles and a colorful scarf + +![](images/fa1e77d8489ed52e472799bfeb95b24b43cddf11fcbe46268f65c13e4f205dfa.jpg) +Figure 7. Comparison of different editing methods on various objects and styles. + +![](images/cfd0cb6fd2309746225d4176b5723e1b12697f932a48a12cbea1fd8ba17b0d9b.jpg) + +![](images/3a4b9116d403bc2969adba0180a410d118bd10aab311aee5251702f5d76f5aca.jpg) + +![](images/a6ad37aba4a9dfe6cf5eea7ce806d6c952033df2bedfc2258e2f35a8a893b5ee.jpg) + +![](images/7e7f55139e27b356337557a9444fac7e08ba0dee13c0e8baf828bfc193734545.jpg) + +Source + +DDS + +CDS + +NT+P2P + +Ours + +lower when no editing occurs, which is particularly visible in DDS-based methods applied to style editing. Our methods achieving better editing score but with a slightly higher structure distance. In terms of precise structure preservation, when combined with CDS [35], it achieves good preservation of the un-edited areas. Our full model achieves the best performance in the CLIP Similarity metric, demonstrating the effectiveness of our prompt enhancement branch. While CDS excels in preserving unedited regions, it suffers from the inferior editability of DDS-based methods. Optimization-based methods [27, 34] refine the inversion process, achieving excellent performance in structure preservation. However, they struggle with editing methods (like P2P), resulting in limited editability. + +# 5.3. Ablation Studies + +In this section, we conduct an ablation experiment to analyze different choices in our SSD. Due to space limitations, we first present a qualitative evaluation in the main text. Please refer to the Supp. for quantitative evaluation. + +The effectiveness of cross-trajectory. In Sec. 4.1, we have analyzed the necessity of cross-trajectory. This term makes the optimization process more stable and provides the source content regularization in ours design, which is also the key difference between ours and Classifier Score Distillation(CSD) [52]. In Fig. 1 and Fig. 5, we present the comparison of the results with and without cross-trajectory. The results show that the cross-trajectory term can provide the direction of generating high-quality images. Please refer to the supplement for more details. + +The effectiveness of prompt-enhancement. The enhance- + +![](images/ac2d77153e3c555c622e7de06d3621d0ff7940ddb8aa9aa6a7b0d6c56a6848f4.jpg) +w/o ID Regular + +![](images/10d4cd7612166efab2a8900afe039440be70e8834f230d6440cc0ea132aeb552.jpg) +ID Regular $\times 1.0$ + +![](images/a63e838151873b6b242f4712f0ede5ce0fc27a1a5c7d6387d9a6eb7bcb3c9b6a.jpg) +ID Regular $\times 2.0$ + +ment of the target prompt branch is another key component in our method, which is designed to improve the editability aligned with the target prompt in 2D-image task. In Fig. 7, we observe a clear distinction from DDS in style editing. The results show that the prompt-enhancement term effectively overcomes the challenging from style editing. + +The effectiveness of ID regularization. ID regularization is designed to ensure stable optimization in 3DGS. In Fig. 8, we compare results with and without ID regularization. The area marked by the yellow arrow highlights its effect in 3D scene editing. However, excessive ID regularization may constrain editing quality by limiting certain attributes, presenting a trade-off in our design. + +# 6. Conclusions, Limitations, and Future Work + +In this work, we propose a novel method for text-guided image editing, capable of handling both 3D scenes and 2D images. Our approach is built on a score distillation framework that leverages the powerful priors of diffusion models. For editing tasks, we design an effective optimization strategy that produces high-quality results aligned with target prompts while ensuring stable and consistent optimization. + +Our method achieves state-of-the-art performance in both 3D scene and 2D image editing, delivering realistic edits with excellent preservation of the original content. It demonstrates strong adaptability to various editing tasks and target prompts, making it a robust solution for complex scenarios. However, while effective, the optimization process is relatively time-intensive compared to recent one-step methods [50] or few-step approaches [7]. Future work could explore integrating advanced techniques such as LCM [30] or SD-Turbo [42], which show potential for accelerating the optimization process [45]. + +Acknowledgment: This project is supported by the National Natural Science Foundation of China (62125201, U24B20174, 62102381), the Guangdong Natural Science Funds for Distinguished Young Scholars (Grant 2023B1515020097), the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No.: AISG3-GV-2023-011), the Singapore Ministry of Education AcRF Tier 1 Grant (Grant No.: MSS25C004), the Lee Kong Chian Fellowships. + +# References + +[1] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, pages 5470–5479, 2022. 6 +[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, pages 18392-18402, 2023. 3, 6 +[3] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In ICCV, pages 22560-22570, 2023. 6 +[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 7 +[5] Minghao Chen, Iro Laina, and Andrea Vedaldi. Dge: Direct gaussian 3d editing by consistent multi-view editing. In ECCV, pages 74-92, 2024. 6 +[6] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. In CVPR, pages 21476-21485, 2024. 3, 6 +[7] Gilad Deutch, Rinon Gal, Daniel Garibi, Or Patashnik, and Daniel Cohen-Or. Turboedit: Text-based image editing using few-step diffusion models. In SIGGRAPHA, 2024. 5, 8 +[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 34:8780-8794, 2021. 3 +[9] Ayaan Haque, Matthew Tancik, Alexei Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. In ICCV, 2023. 3, 6 +[10] Amir Hertz, Kfir Aberman, and Daniel Cohen-Or. Delta denoising score. In ICCV, pages 2328-2337, 2023. 1, 2, 3, 4, 6, 7 +[11] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In ICLR, 2023. 6 +[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshop, 2021. 2, 3, 5 +[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 1, 3 +[14] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 3, 6 +[15] Yukun Huang, Jianan Wang, Yukai Shi, Boshi Tang, Xianbiao Qi, and Lei Zhang. Dreamtime: An improved optimization strategy for diffusion-guided 3d generation. In ICLR, 2023. 3 +[16] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. In CVPR, pages 12469-12478, 2024. 3 +[17] Aapo Hyvarinen and Peter Dayan. Estimation of non-normalized statistical models by score matching. JMLR, 6(4), 2005. 3 + +[18] Yutao Jiang, Yang Zhou, Yuan Liang, Wenxi Liu, Jianbo Jiao, Yuhui Quan, and Shengfeng He. Diffuse3d: Wide-angle 3d photography via bilateral diffusion. In ICCV, pages 8998-9008, 2023. 3 +[19] Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. Pnp inversion: Boosting diffusion-based editing with 3 lines of code. In ICLR, 2024. 2, 6, 7 +[20] Oren Katzir, Or Patashnik, Daniel Cohen-Or, and Dani Lischinski. Noise-free score distillation. In ICLR, 2024. 1, 2, 4 +[21] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splattering for real-time radiance field rendering. ACM TOG, 42(4):1-14, 2023. 3, 6 +[22] Jiwook Kim, Seonho Lee, Jaeyo Shin, Jiho Choi, and Hyunjung Shim. Dreamcatalyst: Fast and high-quality 3d editing via controlling editability and identity preservation. In ICLR, 2025. 2, 3 +[23] Hubert Kompanowski and Binh-Son Hua. Dream-in-style: Text-to-3d generation using stylized score distillation. In 3DV, 2025. 2 +[24] Juil Koo, Chanho Park, and Minhyuk Sung. Posterior distillation sampling. In CVPR, pages 13352-13361, 2024. 2, 3, 5, 6 +[25] Duong H Le, Tuan Pham, Aniruddha Kembhavi, Stephan Mandt, Wei-Chiu Ma, and Jiasen Lu. Preserving identity with variational score for general-purpose 3d editing. arXiv preprint arXiv:2406.08953, 2024. 3 +[26] Kehan Li, Yanbo Fan, Yang Wu, Zhongqian Sun, Wei Yang, Xiangyang Ji, Li Yuan, and Jie Chen. Learning pseudo 3d guidance for view-consistent texturing with 2d diffusion. In ECCV, pages 18-34, 2025. 1 +[27] Senmao Li, Joost Van De Weijer, Taihang Hu, Fahad Shahbaz Khan, Qibin Hou, Yaxing Wang, and Jian Yang. Stylediffusion: Prompt-embedding inversion for text-based editing. arXiv preprint arXiv:2303.15649, 2023. 7, 8 +[28] Haofeng Liu, Chenshu Xu, Yifei Yang, Lihua Zeng, and Shengfeng He. Drag your noise: Interactive point-based editing via diffusion semantic propagation. In CVPR, pages 6743-6752, 2024. 1 +[29] Yuqin Lu, Bailin Deng, Zhixuan Zhong, Tianle Zhang, Yuhui Quan, Hongmin Cai, and Shengfeng He. 3d snapshot: Invertible embedding of 3d neural representations in a single image. 46(12):11524-11531, 2024. 3 +[30] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 8 +[31] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In CVPR, pages 12663–12673, 2023. 1 +[32] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM TOG, 38(4): 1-14, 2019. 6 + +[33] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, pages 405-421, 2020. 3 +[34] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, pages 6038–6047, 2023. 7, 8 +[35] Hyelin Nam, Gihyun Kwon, Geon Yeong Park, and Jong Chul Ye. Contrastive denoising score for text-guided latent diffusion image editing. In CVPR, pages 9192-9201, 2024. 2, 7, 8 +[36] JangHo Park, Gihyun Kwon, and Jong Chul Ye. ED-NeRF: Efficient text-guided editing of 3d scene with latent space nef. In ICLR, 2024. 2, 3 +[37] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 1, 3, 4 +[38] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2): 3, 2022. 3 +[39] Jingjing Ren, Wenbo Li, Haoyu Chen, Renjing Pei, Bin Shao, Yong Guo, Long Peng, Fenglong Song, and Lei Zhu. Ultrapixel: Advancing ultra high-resolution image synthesis to new peaks. In NeurIPS, pages 111131-111171, 2024. 3 +[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 3, 6 +[41] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, pages 36479-36494, 2022. 1, 3 +[42] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In ECCV, pages 87-103, 2025. 8 +[43] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 1, 3, 6 +[44] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 3 +[45] Feng Tian, Yixuan Li, Yichao Yan, Shanyan Guan, Yanhao Ge, and Xiaokang Yang. Postedit: Posterior sampling for efficient zero-shot image editing. In ICLR, 2025. 8 +[46] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In CVPR, pages 1921-1930, 2023. 6 +[47] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In NeurIPS, 2024. 1, 3 +[48] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: + +Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 7 +[49] Jing Wu, Jia-Wang Bian, Xinghui Li, Guangrun Wang, Ian Reid, Philip Torr, and Victor Adrian Prisacariu. Gaussctrl: Multi-view consistent text-driven 3d gaussian splatting editing. In ECCV, pages 55-71, 2024. 3 +[50] Sihan Xu, Yidong Huang, Jiayi Pan, Ziqiao Ma, and Joyce Chai. Inversion-free image editing with natural language. In CVPR, 2024. 8 +[51] Yangyang Xu, Shengfeng He, Kwan-Yee K Wong, and Ping Luo. Rigid: Recurrent gan inversion and editing of real face videos and beyond. IJCV, 133(6):3437-3455, 2025. 1 +[52] Xin Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Song-Hai Zhang, and Xiaojuan Qi. Text-to-3d with classifier score distillation. In ICLR, 2024. 1, 2, 3, 4, 5, 8 +[53] Yuyang Yu, Bangzhen Liu, Chenxi Zheng, Xuemiao Xu, Huaidong Zhang, and Shengfeng He. Beyond textual constraints: Learning novel diffusion conditions with fewer examples. In CVPR, pages 7109-7118, 2024. 3 +[54] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836-3847, 2023. 3 +[55] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7 +[56] Chenxi Zheng, Yihong Lin, Bangzhen Liu, Xuemiao Xu, Yongwei Nie, and Shengfeng He. Recdreamer: Consistent text-to-3d generation via uniform score distillation. In ICLR, 2025. 1, 3 +[57] Xingchen Zhou, Ying He, F Richard Yu, Jianqiang Li, and You Li. RePaint-NeRF: Nerf editing via semantic masks and diffusion models. In IJCAI, 2023. 3 +[58] Junzhe Zhu, Peiye Zhuang, and Sanmi Koyejo. HIFA: High-fidelity text-to-3d generation with advanced diffusion guidance. In ICLR, 2024. 1, 3 \ No newline at end of file diff --git a/2025/Stable Score Distillation/images.zip b/2025/Stable Score Distillation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..76ba52c37d948796db70bce86dc808f7099a8f74 --- /dev/null +++ b/2025/Stable Score Distillation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a1c81b45609baf1c705ab3ff89efd3feb958f336863fed6bc9e031c28ed01ffd +size 854168 diff --git a/2025/Stable Score Distillation/layout.json b/2025/Stable Score Distillation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..74bc0ad494420fc658951988b0e3126eb9ba1122 --- /dev/null +++ b/2025/Stable Score Distillation/layout.json @@ -0,0 +1,11451 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 229, + 103, + 381, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 103, + 381, + 119 + ], + "spans": [ + { + "bbox": [ + 229, + 103, + 381, + 119 + ], + "type": "text", + "content": "Stable Score Distillation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 157, + 142, + 452, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 142, + 452, + 171 + ], + "spans": [ + { + "bbox": [ + 157, + 142, + 452, + 171 + ], + "type": "text", + "content": "Haiming Zhu" + }, + { + "bbox": [ + 157, + 142, + 452, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 157, + 142, + 452, + 171 + ], + "type": "text", + "content": ", Yangyang Xu" + }, + { + "bbox": [ + 157, + 142, + 452, + 171 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 157, + 142, + 452, + 171 + ], + "type": "text", + "content": ", Chenshu Xu" + }, + { + "bbox": [ + 157, + 142, + 452, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 157, + 142, + 452, + 171 + ], + "type": "text", + "content": ", Tingrui Shen" + }, + { + "bbox": [ + 157, + 142, + 452, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 157, + 142, + 452, + 171 + ], + "type": "text", + "content": ", Wenxi Liu" + }, + { + "bbox": [ + 157, + 142, + 452, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 157, + 142, + 452, + 171 + ], + "type": "text", + "content": ", Yong Du" + }, + { + "bbox": [ + 157, + 142, + 452, + 171 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 157, + 142, + 452, + 171 + ], + "type": "text", + "content": ", Jun Yu" + }, + { + "bbox": [ + 157, + 142, + 452, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 157, + 142, + 452, + 171 + ], + "type": "text", + "content": ", Shengfeng He" + }, + { + "bbox": [ + 157, + 142, + 452, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 172, + 475, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 172, + 475, + 185 + ], + "spans": [ + { + "bbox": [ + 135, + 172, + 475, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 172, + 475, + 185 + ], + "type": "text", + "content": "Singapore Management University " + }, + { + "bbox": [ + 135, + 172, + 475, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 135, + 172, + 475, + 185 + ], + "type": "text", + "content": "Harbin Institute of Technology (Shenzhen)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 109, + 186, + 498, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 186, + 498, + 198 + ], + "spans": [ + { + "bbox": [ + 109, + 186, + 498, + 198 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 109, + 186, + 498, + 198 + ], + "type": "text", + "content": "South China University of Technology " + }, + { + "bbox": [ + 109, + 186, + 498, + 198 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 109, + 186, + 498, + 198 + ], + "type": "text", + "content": "Fuzhou University " + }, + { + "bbox": [ + 109, + 186, + 498, + 198 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 109, + 186, + 498, + 198 + ], + "type": "text", + "content": "Ocean University of China" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 58, + 201, + 149, + 270 + ], + "blocks": [ + { + "bbox": [ + 58, + 201, + 149, + 270 + ], + "lines": [ + { + "bbox": [ + 58, + 201, + 149, + 270 + ], + "spans": [ + { + "bbox": [ + 58, + 201, + 149, + 270 + ], + "type": "image", + "image_path": "215e2790b3793d17d15859ef9997fc328f48e5c57b8796648d55ed6cb98b5ec8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 58, + 270, + 146, + 337 + ], + "blocks": [ + { + "bbox": [ + 58, + 270, + 146, + 337 + ], + "lines": [ + { + "bbox": [ + 58, + 270, + 146, + 337 + ], + "spans": [ + { + "bbox": [ + 58, + 270, + 146, + 337 + ], + "type": "image", + "image_path": "94379b3d762a60744f9d4afc109879d218b487c1a02403c497c68dad6352031b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 73, + 338, + 134, + 348 + ], + "lines": [ + { + "bbox": [ + 73, + 338, + 134, + 348 + ], + "spans": [ + { + "bbox": [ + 73, + 338, + 134, + 348 + ], + "type": "text", + "content": "(a) Original Views" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 150, + 202, + 271, + 270 + ], + "blocks": [ + { + "bbox": [ + 150, + 202, + 271, + 270 + ], + "lines": [ + { + "bbox": [ + 150, + 202, + 271, + 270 + ], + "spans": [ + { + "bbox": [ + 150, + 202, + 271, + 270 + ], + "type": "image", + "image_path": "cc09d0cc8b5fac1f29b5496fddb5464f1e45bcdda3d2c9bca51cdbe82300ee06.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 150, + 270, + 279, + 336 + ], + "blocks": [ + { + "bbox": [ + 150, + 270, + 279, + 336 + ], + "lines": [ + { + "bbox": [ + 150, + 270, + 279, + 336 + ], + "spans": [ + { + "bbox": [ + 150, + 270, + 279, + 336 + ], + "type": "image", + "image_path": "f83feac7a9b6dbc14a1b6005bc55d6e14b6e47965ba8113917b44c124f3e85fb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 338, + 236, + 348 + ], + "lines": [ + { + "bbox": [ + 195, + 338, + 236, + 348 + ], + "spans": [ + { + "bbox": [ + 195, + 338, + 236, + 348 + ], + "type": "text", + "content": "(b) CSD[52]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 282, + 202, + 402, + 270 + ], + "blocks": [ + { + "bbox": [ + 282, + 202, + 402, + 270 + ], + "lines": [ + { + "bbox": [ + 282, + 202, + 402, + 270 + ], + "spans": [ + { + "bbox": [ + 282, + 202, + 402, + 270 + ], + "type": "image", + "image_path": "b69aca6b82b03c70b0636af851357aafc4d4644a62e8db6343325dc46726ffbd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 282, + 270, + 402, + 336 + ], + "blocks": [ + { + "bbox": [ + 282, + 270, + 402, + 336 + ], + "lines": [ + { + "bbox": [ + 282, + 270, + 402, + 336 + ], + "spans": [ + { + "bbox": [ + 282, + 270, + 402, + 336 + ], + "type": "image", + "image_path": "12492b653667c289aac1cee6e4554b4d08a8b414bbfc37ca5e0cd529fd795f9c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 338, + 369, + 348 + ], + "lines": [ + { + "bbox": [ + 328, + 338, + 369, + 348 + ], + "spans": [ + { + "bbox": [ + 328, + 338, + 369, + 348 + ], + "type": "text", + "content": "(c) DDS[10]" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 353, + 555, + 398 + ], + "lines": [ + { + "bbox": [ + 54, + 353, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 54, + 353, + 555, + 398 + ], + "type": "text", + "content": "Figure 1. We propose Stable Score Distillation (SSD), a method that improves text-guided editing by preserving original content structure and enhancing realism in edited results. The first column shows the original views, while the remaining columns display comparative results with existing score distillation methods on the SD model. SSD demonstrates superior performance, maintaining the integrity of the original content and producing more realistic edits." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 413, + 202, + 549, + 270 + ], + "blocks": [ + { + "bbox": [ + 413, + 202, + 549, + 270 + ], + "lines": [ + { + "bbox": [ + 413, + 202, + 549, + 270 + ], + "spans": [ + { + "bbox": [ + 413, + 202, + 549, + 270 + ], + "type": "image", + "image_path": "58b1d4ad15639ae1d16011af1c25387c59da9c38aa0ab9a43a7f9145b9405fa9.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 414, + 270, + 548, + 336 + ], + "blocks": [ + { + "bbox": [ + 414, + 270, + 548, + 336 + ], + "lines": [ + { + "bbox": [ + 414, + 270, + 548, + 336 + ], + "spans": [ + { + "bbox": [ + 414, + 270, + 548, + 336 + ], + "type": "image", + "image_path": "c9e7d6162ad4fe9abefbc17693f4d165e0c6bd0bb674c6968388be0193d2ddc9.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 466, + 338, + 493, + 348 + ], + "lines": [ + { + "bbox": [ + 466, + 338, + 493, + 348 + ], + "spans": [ + { + "bbox": [ + 466, + 338, + 493, + 348 + ], + "type": "text", + "content": "(d) Ours" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 151, + 411, + 200, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 411, + 200, + 423 + ], + "spans": [ + { + "bbox": [ + 151, + 411, + 200, + 423 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 53, + 435, + 297, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 435, + 297, + 687 + ], + "spans": [ + { + "bbox": [ + 53, + 435, + 297, + 687 + ], + "type": "text", + "content": "Text-guided image and 3D editing have advanced with diffusion-based models, yet methods like Delta Denoising Score often struggle with stability, spatial control, and editing strength. These limitations stem from reliance on complex auxiliary structures, which introduce conflicting optimization signals and restrict precise, localized edits. We introduce Stable Score Distillation (SSD), a streamlined framework that enhances stability and alignment in the editing process by anchoring a single classifier to the source prompt. Specifically, SSD utilizes Classifier-Free Guidance (CFG) equation to achieve cross-prompt alignment, and introduces a constant term null-text branch to stabilize the optimization process. This approach preserves the original content's structure and ensures that editing trajectories are closely aligned with the source prompt, enabling smooth, prompt-specific modifications while maintaining coherence in surrounding regions. Additionally, SSD incorporates a prompt enhancement branch to boost editing strength, particularly for style transformations. Our method achieves state-of-the-art results in 2D and 3D editing tasks, including NeRF and text-driven style edits, with faster conver" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 412, + 556, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 556, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 556, + 448 + ], + "type": "text", + "content": "gence and reduced complexity, providing a robust and efficient solution for text-guided editing. Code is available at: https://github.com/Alex-Zhu1/SSD." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 314, + 465, + 395, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 465, + 395, + 477 + ], + "spans": [ + { + "bbox": [ + 314, + 465, + 395, + 477 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 485, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 485, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 312, + 485, + 555, + 581 + ], + "type": "text", + "content": "Text-based image generation has achieved remarkable progress, particularly with the advent of diffusion models [13, 28, 40, 41, 43, 56]. These models leverage strong priors to produce high-quality images, facilitating significant advances in text-to-3D generation [37, 47, 58]. Moreover, text-guided 3D editing has enabled intricate modifications to shape and texture [26, 31], supporting flexible and precise 3D scene manipulation." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 582, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 582, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 582, + 556, + 714 + ], + "type": "text", + "content": "Unlike generation tasks that create new content, editing tasks aim to modify specific elements within an image while preserving surrounding areas [51]. However, directly applying methods like Score Distillation Sampling (SDS) to editing tasks can yield undesired effects, such as blurring across the image. This arises because SDS optimizes globally to the prompt, affecting regions beyond the targeted area [10, 20]. DDS [10] addresses this by introducing a dual-branch architecture, pairing the source image with its description to leverage the model's inherent bias and isolate specific prompt changes. Further, CSD [52] achieved" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "*Corresponding authors: Yangyang Xu (xuyangyang@hit.edu.cn) and Shengfeng He (shengfenghe@smu.edu.sg)." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16597" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 74, + 197, + 169 + ], + "blocks": [ + { + "bbox": [ + 65, + 74, + 197, + 169 + ], + "lines": [ + { + "bbox": [ + 65, + 74, + 197, + 169 + ], + "spans": [ + { + "bbox": [ + 65, + 74, + 197, + 169 + ], + "type": "image", + "image_path": "113e4cf3086202d6189ea23ab0e6c94a7fef3b378152c624fb16ea51a380b2b7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 175, + 555, + 220 + ], + "lines": [ + { + "bbox": [ + 55, + 175, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 555, + 220 + ], + "type": "text", + "content": "Figure 2. Illustration of three distillation-based approaches. Note that we assume a 2-step optimization process for illustration, where the subscript " + }, + { + "bbox": [ + 55, + 175, + 555, + 220 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 175, + 555, + 220 + ], + "type": "text", + "content": " represents the iteration number. DDS utilizes the source branch to obtain initial latent " + }, + { + "bbox": [ + 55, + 175, + 555, + 220 + ], + "type": "inline_equation", + "content": "Z_0^*" + }, + { + "bbox": [ + 55, + 175, + 555, + 220 + ], + "type": "text", + "content": ", while CSD employs two classifiers to derive " + }, + { + "bbox": [ + 55, + 175, + 555, + 220 + ], + "type": "inline_equation", + "content": "Z_1'" + }, + { + "bbox": [ + 55, + 175, + 555, + 220 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 175, + 555, + 220 + ], + "type": "inline_equation", + "content": "Z_1''" + }, + { + "bbox": [ + 55, + 175, + 555, + 220 + ], + "type": "text", + "content": " for cross-prompt editing. Our SSD method designs a CFG classifier to determine the cross-prompt editing, introduces the null-text branch as the initial latent " + }, + { + "bbox": [ + 55, + 175, + 555, + 220 + ], + "type": "inline_equation", + "content": "Z_0^*" + }, + { + "bbox": [ + 55, + 175, + 555, + 220 + ], + "type": "text", + "content": ", and further constructs the cross-trajectory term (see Sec. 4.1) for stable optimization." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 201, + 74, + 334, + 170 + ], + "blocks": [ + { + "bbox": [ + 201, + 74, + 334, + 170 + ], + "lines": [ + { + "bbox": [ + 201, + 74, + 334, + 170 + ], + "spans": [ + { + "bbox": [ + 201, + 74, + 334, + 170 + ], + "type": "image", + "image_path": "c5589cd515662891cef249906174257aa7b3542a2aa1bf446dcc8e922b46d715.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 337, + 74, + 471, + 170 + ], + "blocks": [ + { + "bbox": [ + 337, + 74, + 471, + 170 + ], + "lines": [ + { + "bbox": [ + 337, + 74, + 471, + 170 + ], + "spans": [ + { + "bbox": [ + 337, + 74, + 471, + 170 + ], + "type": "image", + "image_path": "1b693ad852678cb0b00bdc1dd1b74f854c74679f7328eae1298787e3f4cc59b8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 474, + 74, + 545, + 171 + ], + "blocks": [ + { + "bbox": [ + 474, + 74, + 545, + 171 + ], + "lines": [ + { + "bbox": [ + 474, + 74, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 474, + 74, + 545, + 171 + ], + "type": "image", + "image_path": "dad6abc620176c768dfa99228a5f05c5ab73f7b654a77aa876f396f92fc6ba21.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 55, + 228, + 300, + 365 + ], + "blocks": [ + { + "bbox": [ + 55, + 228, + 300, + 365 + ], + "lines": [ + { + "bbox": [ + 55, + 228, + 300, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 300, + 365 + ], + "type": "image", + "image_path": "479da4509d074ccb14bce7ce3df49b68e806012d1bbbe6aecabfaea49663ba2a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 365, + 297, + 399 + ], + "lines": [ + { + "bbox": [ + 55, + 365, + 297, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 297, + 399 + ], + "type": "text", + "content": "Figure 3. The optimization process of DDS and our SSD. SSD preserves the source structure effectively during optimization iterations, while DDS cannot preserve it effectively." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 410, + 296, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 296, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 296, + 481 + ], + "type": "text", + "content": "scene editing by incorporating a classifier component within Classifier-Free Guidance (CFG) [12] to refine the prediction score by applying the classifier to both the source and target prompts. NFSD [20] further decomposes the CFG score, highlighting the classifier as the primary driver of prompt direction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 483, + 296, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 686 + ], + "type": "text", + "content": "Despite their success, we argue that current distillation-based approaches face inherent limitations, such as low editing quality and loss of source content. As shown in Fig. 2, DDS [10] relies on the source branch to remove model bias, but lacks the explicit guidance to preserve the source content [24] during optimization. As shown in Fig. 1c, DDS changes man's clothes during editing his faces. Additionally, although introducing source prompt components is intended to improve prompt specificity [19], it can amplify noise and introduce overlapping objectives that hinder stable convergence. This results in artifacts or unintended variations, especially in the unedited regions. Correspondingly, CSD [52] utilizes dual classifiers to refine the prompt editing direction (see Fig. 2). However, it lacks the explicit source preservation to restrict edits precisely to the target areas. As shown in Fig. 1b, this causes the structure deformation and annoying artifacts in the edited regions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Our insights into these limitations lead to two key observations: (1) Cross-prompt: a single classifier, providing" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 232, + 555, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 278 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 278 + ], + "type": "text", + "content": "the editing direction from source prompt to target, and (2) Cross-trajectory: stability in the editing process can be achieved by aligning the editing direction closely with the structure of the source content." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 281, + 556, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 556, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 556, + 567 + ], + "type": "text", + "content": "In this paper, we propose Stable Score Distillation (SSD), a streamlined approach for stable and precise text-guided editing. To achieve a smooth editing direction, we employ the CFG equation for both the source and target prompts, ensuring a gradual transition of the original contextual texture as the model adapts to the specified changes. This approach contrasts with DDS [10], as it eliminates the need for a auxiliary source branch, enabling our method to focus editing gradients precisely within target regions while ensuring a stable transition, as illustrated in Fig. 3. Moreover, for aligning the editing direction with the source prompt, facilitating smoother and more controlled progression toward the target prompt, we design an cross-trajectory strategy to ensure that edits respect the original structure, supporting subtle and stable transformations within designated areas. While NFSD [20] utilizes negative-branch and DDS utilizes source branch to enhance output clarity, as shown in SSD in Fig. 2, we introduce a null-text branch aligned with the \"no-edit\" direction to integrate a \"reconstruction\" term to explicitly enforce source content preservation, which enhances consistency and produces reliable edits across diverse tasks. Based on above designs, our framework remains streamlined and efficient, achieving both precision and stability without the complexity of additional components." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 568, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 556, + 700 + ], + "type": "text", + "content": "Our framework integrates seamlessly into existing DDS-based editing pipelines and applications, such as text-driven NeRF editing [22, 24, 36] and 2D image editing [35]. Notably, our approach's \"clear\" editing direction preserves source content, making a carefully designed identity regularization [22] unnecessary. Moreover, standard DDS methods often lack sufficient editing strength, resulting in minimal or negligible changes in output, particularly in style editing [23]. Our approach, with its streamlined and stable framework, allows for the seamless integration of a prompt enhancement branch to amplify editing capability." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 326, + 701, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 713 + ], + "type": "text", + "content": "With these improvements, our method achieves faster and" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16598" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 167 + ], + "type": "text", + "content": "more effective edits during optimization, remains compatible with the Stable Diffusion Model [40] without requiring LoRA [14] or fine-tuning, and integrates effectively with Instructpix2pix [2]. Additionally, by incorporating nonincreasing timestep sampling [15], we accelerate convergence, reducing the required iterations to approximately 3,000 for NeRF [33] and 1,500 for Gaussian splatting [21]. In summary, our contributions are as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 169, + 294, + 275 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 55, + 169, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 169, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 169, + 294, + 204 + ], + "type": "text", + "content": "- We introduce a novel editing framework, Stable Score Distillation, that leverages a single, anchored classifier to achieve targeted and stable edits in 3D scene editing." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 205, + 294, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 294, + 240 + ], + "type": "text", + "content": "- We introduce a prompt enhancement strategy, effectively improve the prompt-alignment, especially style editing in 2D-image editing." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 241, + 294, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 275 + ], + "type": "text", + "content": "- We demonstrate the effectiveness of our approach across NeRF-editing and image-editing tasks, achieving state-of-the-art results with a streamlined and efficient framework." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 289, + 141, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 141, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 141, + 301 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 308, + 158, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 158, + 320 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 158, + 320 + ], + "type": "text", + "content": "2.1. Diffusion Models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 327, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 445 + ], + "type": "text", + "content": "Diffusion models [13, 39, 43, 53] have made significant advancements in generating diverse and high-fidelity images. Starting from a gaussian noise, diffusion models can predict the noise-less sample at each time step, until finally obtaining clear samples. Commonly, the denoising process can utilize U-net model to predict the noise. Some works [13, 44] have observed that is proportional to the predicted score function [17] of the smoothed density. Thus, intuitively, taking steps in the direction of the score function gradually moves the sample towards the data distribution." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 448, + 296, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 296, + 602 + ], + "type": "text", + "content": "To generate images aligned with a target prompt, guidance is typically introduced to explicitly control the weight assigned to the conditioning information. The popular guidance methods include Classifier Guidance [8] and Class-free Guidance (CFG) [12]. While the former relies on a separately learned classifier, the latter directly introduces null-text samples to the model. CFG modifies the score function to steer the process towards regions with a higher ratio of conditional density to the unconditional one. However, it has been observed that CFG trades sample fidelity for diversity [12]. Based on the insights gained from the decomposition of the CFG equation, we propose a novel Stable Score Distillation (SSD) method to guide the SDS optimization process." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 611, + 235, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 235, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 235, + 624 + ], + "type": "text", + "content": "2.2. Score Distillation Sampling (SDS)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "Benefit from the data scale-law, diffusion model [38, 40, 41] achieve high-quality image generation and text-to-image generation. Specifically, Score Distillation Sampling (SDS) [37] leveraging the priors of pre-trained text-to-image models to facilitate text-conditioned generation in 3D content generation. Specifically, SDS is an optimization approach that updates the rendering parameter towards the image distribu" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": "tion of diffusion models by enforcing the noise prediction on noisy rendered images to match sampled noise. While SDS provides an elegant mechanism for leveraging pretrained text-to-image models, SDS-generated results often suffer from oversaturation and lack of fine realistic details. VSD [47] proposed a particle-based optimization framework that treats the 3D parameter as a random variable of target distribution. Furthermore, by regarding SDS as a reverse diffusion process, decreasing timesteps sampling [15, 58] to imitate the diffusion reverse sampling, which can improve the quality of the generated 3D assets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 208, + 553, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 208, + 553, + 291 + ], + "spans": [ + { + "bbox": [ + 313, + 208, + 553, + 291 + ], + "type": "text", + "content": "In image editing, Delta Denoising Score (DDS) [10] found that Score Distillation Sampling (SDS) introduces noticeable artifacts and over-smoothing in edited images due to inherent bias. To mitigate this bias, DDS employs a subtraction of two SDS scores of the source and target images to obtain a delta score, which is then used to guide the optimization process." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 312, + 477, + 325 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 477, + 325 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 477, + 325 + ], + "type": "text", + "content": "2.3. Text-Driven 3D-Scene Editing" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 334, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 555, + 453 + ], + "type": "text", + "content": "Text-driven 3D scene editing has been a popular research topic [18, 29, 56]. IN2N [9] proposed a Iterative Dataset Update method that can edit 3D scenes from text descriptions. By leveraging advancements in 2D diffusion editing techniques, notably InstructP2P [2] and ControlNet [54], GaussianEditor[6] and GaussCtrl [49] utilize edit multi-view images latent to optimize the 3D scene. We consider utilize score distillation to guide the 3D scene editing, which is more flexible and efficient for the text-driven 3D scene editing." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 458, + 556, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 458, + 556, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 458, + 556, + 662 + ], + "type": "text", + "content": "Building upon the foundational SDS loss introduced by DreamFusion [37], some work has explored SDS loss in the text-driven 3D scene editing. RePaint-NeRF [57] has advanced the application of SDS in 3D editing by integrating a semantic mask to guide and constrain modifications within the background elements. CSD [52] utilize two classifiers to achieve editing. In a similar vein, ED-NeRF [36] has introduced an enhanced loss function specifically designed for 3D editing tasks. PDS [24] proposed a posterior distillation sampling to match stochastic latent [16]. Piva [25] fine-tuned the model while introducing a regularization term to preserve identity. Unfortunately, these methods are still limited to the long-time diffusion reverse sampling process, which is not suboptimal for the text-driven 3D scene editing. DreamCatalyst [22] extends the PDS optimization processing to ID-preserving and edit-ability based on decreasing timesteps sampling." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "Different from the above methods, ours firstly improve the DDS optimization process by introducing a single classifier, and further introduce a null-text branch to achieve a more stable and precise editing process." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16599" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 71, + 539, + 274 + ], + "blocks": [ + { + "bbox": [ + 71, + 71, + 539, + 274 + ], + "lines": [ + { + "bbox": [ + 71, + 71, + 539, + 274 + ], + "spans": [ + { + "bbox": [ + 71, + 71, + 539, + 274 + ], + "type": "image", + "image_path": "1c65990c7e27298d4db995adee328213b5e3471bf049410d212daf798bf05600.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 275, + 555, + 321 + ], + "lines": [ + { + "bbox": [ + 54, + 275, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 54, + 275, + 555, + 321 + ], + "type": "text", + "content": "Figure 4. The overview of SSD. Given the parameter 3D-model or image, SSD provides effective editing gradient to guide the optimization process. We utilize CFG equation between the predicted target noise " + }, + { + "bbox": [ + 54, + 275, + 555, + 321 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}(z_t,y,t)" + }, + { + "bbox": [ + 54, + 275, + 555, + 321 + ], + "type": "text", + "content": " and source noise " + }, + { + "bbox": [ + 54, + 275, + 555, + 321 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}(z_t,\\hat{y},t)" + }, + { + "bbox": [ + 54, + 275, + 555, + 321 + ], + "type": "text", + "content": ", which generate the gradual editing direction. Furthermore, we introduce a null-text branch " + }, + { + "bbox": [ + 54, + 275, + 555, + 321 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}(\\hat{z}_t,\\emptyset ,t)" + }, + { + "bbox": [ + 54, + 275, + 555, + 321 + ], + "type": "text", + "content": " to regularize the optimization process and achieve stable optimization. We further analyzing and decompose ours design term into three parts: cross-prompt, cross-trajectory, and prompt-enhance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 328, + 133, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 133, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 133, + 342 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 348, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 296, + 396 + ], + "type": "text", + "content": "In this section, we first discuss existing optimization-based approaches to handle parametric images. Then, we will introduce our novel parametric image editing method in Section 4." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 403, + 206, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 206, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 206, + 417 + ], + "type": "text", + "content": "3.1. Score Distillation Sampling" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 421, + 297, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 297, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 297, + 495 + ], + "type": "text", + "content": "Score Distillation Sampling (SDS) [37] is proposed to generate parametric images by leveraging the 2D prior of pretrained text-to-image diffusion models. Specifically, given a pretrained diffusion model " + }, + { + "bbox": [ + 55, + 421, + 297, + 495 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}" + }, + { + "bbox": [ + 55, + 421, + 297, + 495 + ], + "type": "text", + "content": ", SDS optimizes a set of parameters " + }, + { + "bbox": [ + 55, + 421, + 297, + 495 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 421, + 297, + 495 + ], + "type": "text", + "content": " of a differentiable parametric image generator " + }, + { + "bbox": [ + 55, + 421, + 297, + 495 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 421, + 297, + 495 + ], + "type": "text", + "content": ", using the gradient of the loss " + }, + { + "bbox": [ + 55, + 421, + 297, + 495 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{SDS}}" + }, + { + "bbox": [ + 55, + 421, + 297, + 495 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 55, + 421, + 297, + 495 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 421, + 297, + 495 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 91, + 501, + 296, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 501, + 296, + 526 + ], + "spans": [ + { + "bbox": [ + 91, + 501, + 296, + 526 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\theta} L _ {\\mathrm {S D S}} = w (t) \\left(\\epsilon_ {\\phi} \\left(z _ {t} (x); y, t\\right) - \\epsilon\\right) \\frac {\\partial x}{\\partial \\theta}, \\tag {1}", + "image_path": "86ae501dc2d7415289ac0b27a187366726b5ccafd630455363be9a402e0cc636.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 533, + 297, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 297, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 297, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 533, + 297, + 629 + ], + "type": "inline_equation", + "content": "x = g(\\theta)" + }, + { + "bbox": [ + 55, + 533, + 297, + 629 + ], + "type": "text", + "content": " is an image rendered by " + }, + { + "bbox": [ + 55, + 533, + 297, + 629 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 533, + 297, + 629 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 533, + 297, + 629 + ], + "type": "inline_equation", + "content": "z_{t}(x)" + }, + { + "bbox": [ + 55, + 533, + 297, + 629 + ], + "type": "text", + "content": " is obtained by adding a Gaussian noise " + }, + { + "bbox": [ + 55, + 533, + 297, + 629 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 533, + 297, + 629 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 533, + 297, + 629 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 533, + 297, + 629 + ], + "type": "text", + "content": " corresponding to the " + }, + { + "bbox": [ + 55, + 533, + 297, + 629 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 533, + 297, + 629 + ], + "type": "text", + "content": "-th timestep of the diffusion process, and " + }, + { + "bbox": [ + 55, + 533, + 297, + 629 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 533, + 297, + 629 + ], + "type": "text", + "content": " is a condition to the diffusion model. As Noise-Free Score Distillation (NFSD) [20] has shown, the score " + }, + { + "bbox": [ + 55, + 533, + 297, + 629 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}(z_{t}(x);y,t)" + }, + { + "bbox": [ + 55, + 533, + 297, + 629 + ], + "type": "text", + "content": " provides the direction in which this noised version of " + }, + { + "bbox": [ + 55, + 533, + 297, + 629 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 533, + 297, + 629 + ], + "type": "text", + "content": " should be moved towards a denser region in the distribution of real images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 635, + 206, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 206, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 206, + 649 + ], + "type": "text", + "content": "3.2. Delta Distillation Sampling" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Although SDS get excellent generation ability, for editing task, an undesired component from the pretrained model, " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{bias}}" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": ", interferes with the process and causes the image to become smooth and blurry in some parts [10]. Based on the observations that a matched source prompt " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " and source" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 329, + 555, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 555, + 365 + ], + "type": "text", + "content": "latent " + }, + { + "bbox": [ + 313, + 329, + 555, + 365 + ], + "type": "inline_equation", + "content": "\\hat{z}_t" + }, + { + "bbox": [ + 313, + 329, + 555, + 365 + ], + "type": "text", + "content": " can estimate the noisy direction " + }, + { + "bbox": [ + 313, + 329, + 555, + 365 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{bias}}" + }, + { + "bbox": [ + 313, + 329, + 555, + 365 + ], + "type": "text", + "content": ", thus, the DDS method aims to remove the " + }, + { + "bbox": [ + 313, + 329, + 555, + 365 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{bias}}" + }, + { + "bbox": [ + 313, + 329, + 555, + 365 + ], + "type": "text", + "content": " by introducing source branch, as shown in Eq. 2:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 346, + 369, + 555, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 369, + 555, + 393 + ], + "spans": [ + { + "bbox": [ + 346, + 369, + 555, + 393 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\theta} L _ {\\mathrm {D D S}} = \\left(\\epsilon_ {\\phi} ^ {c} \\left(z _ {t}, y, t\\right) - \\epsilon_ {\\phi} ^ {c} \\left(\\hat {z} _ {t}, \\hat {y}, t\\right)\\right) \\frac {\\partial z}{\\partial \\theta}, \\tag {2}", + "image_path": "03693eb0302f7a8cbc62f7ad054e40691cffbc3bd6c75bb0857598b67adb1820.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 396, + 556, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 556, + 505 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 556, + 505 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 396, + 556, + 505 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}^{c}(z_{t},y,t)" + }, + { + "bbox": [ + 313, + 396, + 556, + 505 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 396, + 556, + 505 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}^{c}(\\hat{z}_{t},\\hat{y},t)" + }, + { + "bbox": [ + 313, + 396, + 556, + 505 + ], + "type": "text", + "content": " are pretrained model predictions " + }, + { + "bbox": [ + 313, + 396, + 556, + 505 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 396, + 556, + 505 + ], + "type": "text", + "content": ", with the superscript " + }, + { + "bbox": [ + 313, + 396, + 556, + 505 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 396, + 556, + 505 + ], + "type": "text", + "content": " indicating the CFG results. Thus, DDS pushes the optimized image into the direction of the target prompt without the interference of the noise component, namely, " + }, + { + "bbox": [ + 313, + 396, + 556, + 505 + ], + "type": "inline_equation", + "content": "\\nabla_{\\theta}L_{\\mathrm{DDS}}\\approx \\delta_{\\mathrm{ext}}" + }, + { + "bbox": [ + 313, + 396, + 556, + 505 + ], + "type": "text", + "content": ". Obviously, " + }, + { + "bbox": [ + 313, + 396, + 556, + 505 + ], + "type": "inline_equation", + "content": "\\nabla_{\\delta_{\\mathrm{ext}}}" + }, + { + "bbox": [ + 313, + 396, + 556, + 505 + ], + "type": "text", + "content": " is contingent on classifier part from " + }, + { + "bbox": [ + 313, + 396, + 556, + 505 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}^{c}(z_{t},y,t)" + }, + { + "bbox": [ + 313, + 396, + 556, + 505 + ], + "type": "text", + "content": " as discussed in CSD [52] and NFSD [20]. Note that in the following manuscript, we decompose the CFG results without the superscript " + }, + { + "bbox": [ + 313, + 396, + 556, + 505 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 396, + 556, + 505 + ], + "type": "text", + "content": " and omit the timestep " + }, + { + "bbox": [ + 313, + 396, + 556, + 505 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 396, + 556, + 505 + ], + "type": "text", + "content": " for simplicity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 505, + 555, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 555, + 540 + ], + "type": "text", + "content": "Further exploring prompt editing direction, CSD [52] method proposed a dual-classifier to refine the editing score and achieve more precise editing, as shown in Eq. 3:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 341, + 545, + 551, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 545, + 551, + 585 + ], + "spans": [ + { + "bbox": [ + 341, + 545, + 551, + 585 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {\\theta} L _ {\\mathrm {C S D}} = \\left(w _ {a} \\left(\\epsilon_ {\\phi} \\left(z _ {t}, y\\right) - \\epsilon_ {\\phi} \\left(z _ {t}, \\varnothing\\right)\\right) \\right. \\\\ \\left. - w _ {b} \\left(\\epsilon_ {\\phi} \\left(z _ {t}, \\hat {y}\\right) - \\epsilon_ {\\phi} \\left(z _ {t}, \\emptyset\\right)\\right)\\right) \\frac {\\partial z}{\\partial \\theta}, \\tag {3} \\\\ \\end{array}", + "image_path": "91c0b33da937d6138cd06b8c4f6701d0e6a7fc8678e2ac2b90a08abf3cbb0259.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 588, + 556, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 556, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 556, + 661 + ], + "type": "text", + "content": "while the " + }, + { + "bbox": [ + 313, + 588, + 556, + 661 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}(z_t,y,t)" + }, + { + "bbox": [ + 313, + 588, + 556, + 661 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 588, + 556, + 661 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}(z_t,\\hat{y},t)" + }, + { + "bbox": [ + 313, + 588, + 556, + 661 + ], + "type": "text", + "content": " are current latent " + }, + { + "bbox": [ + 313, + 588, + 556, + 661 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 313, + 588, + 556, + 661 + ], + "type": "text", + "content": " predictions for the target prompt " + }, + { + "bbox": [ + 313, + 588, + 556, + 661 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 588, + 556, + 661 + ], + "type": "text", + "content": " and source prompt " + }, + { + "bbox": [ + 313, + 588, + 556, + 661 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 313, + 588, + 556, + 661 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 313, + 588, + 556, + 661 + ], + "type": "inline_equation", + "content": "w_{a}" + }, + { + "bbox": [ + 313, + 588, + 556, + 661 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 588, + 556, + 661 + ], + "type": "inline_equation", + "content": "w_{b}" + }, + { + "bbox": [ + 313, + 588, + 556, + 661 + ], + "type": "text", + "content": " are weights of classifiers. Simply put, CSD aims to refine the prompt editing direction by determining the difference between two classifiers, which can be regarded as a cross-prompt term." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 669, + 370, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 669, + 370, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 669, + 370, + 681 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "In 3D scene editing process, which requires consideration of both the target prompt and the original source content," + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16600" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "content": "we consider two key aspects: (1) smooth editing direction towards the target prompt (2) and editing results respect the original structure. Based on these, in this section, we introduce our novel editing framework Stable Score Distillation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 129, + 191, + 141 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 129, + 191, + 141 + ], + "spans": [ + { + "bbox": [ + 55, + 129, + 191, + 141 + ], + "type": "text", + "content": "4.1. Stable Score Distillation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 148, + 296, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 148, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 148, + 296, + 243 + ], + "type": "text", + "content": "Firstly, we introduce the design of a cross-prompt editing direction. As discussed about CSD method in Sec. 3.2, the key role in cross-prompt editing is to provide a smooth transition from the source prompt to the target. As the CFG guidance [12] steers the process towards regions with a higher ratio of conditional density to the unconditional one, accordingly, we can modify the SDS score function, as shown below:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 255, + 296, + 269 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 255, + 296, + 269 + ], + "spans": [ + { + "bbox": [ + 82, + 255, + 296, + 269 + ], + "type": "interline_equation", + "content": "G r a d = \\epsilon_ {\\phi} (z _ {t}, \\hat {y}) + s (\\epsilon_ {\\phi} (z _ {t}, y) - \\epsilon_ {\\phi} (z _ {t}, \\hat {y})), \\tag {4}", + "image_path": "87fc3c38b278291aef51141267aa024c1c91f9e56061652fdf1978ed438bcf38.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 279, + 296, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 296, + 303 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 296, + 303 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 279, + 296, + 303 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}(z_t,y)" + }, + { + "bbox": [ + 55, + 279, + 296, + 303 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 279, + 296, + 303 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}(z_t,\\hat{y})" + }, + { + "bbox": [ + 55, + 279, + 296, + 303 + ], + "type": "text", + "content": " are pretrained model predictions. The scale factor " + }, + { + "bbox": [ + 55, + 279, + 296, + 303 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 279, + 296, + 303 + ], + "type": "text", + "content": " is equal to control weight." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 304, + 296, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 296, + 448 + ], + "type": "text", + "content": "Although the cross-prompt term provides a smooth texture transition in the edited region, we observed that the optimization process leads to abrupt structural changes, often resulting in artifacts and unappealing outcomes, similar to CSD in Fig. 1b. To address this, we introduce an additional regularization term to constrain the structural transition. Interestingly, as shown in Fig. 1c, DDS achieves better results than CSD by incorporating a source branch. However, DDS still lacks a mechanism to ensure the original structure remains intact, leading to modification on unedited regions. To address this, we introduce a null-text branch " + }, + { + "bbox": [ + 55, + 304, + 296, + 448 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}(\\hat{z}_t,\\emptyset)" + }, + { + "bbox": [ + 55, + 304, + 296, + 448 + ], + "type": "text", + "content": " to regularize the optimization process, as shown in Eq. 5:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 458, + 296, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 458, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 61, + 458, + 296, + 472 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {s s d}} = \\epsilon_ {\\phi} (z _ {t}, \\hat {y}) + s \\left(\\epsilon_ {\\phi} (z _ {t}, y) - \\epsilon_ {\\phi} (z _ {t}, \\hat {y})\\right) - \\epsilon_ {\\phi} (\\hat {z} _ {t}, \\varnothing). \\tag {5}", + "image_path": "acf284ddca3d1a8354461a37a8d0ff2e476a6f1b96f70de2845a754ae59e954e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 483, + 296, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 519 + ], + "type": "text", + "content": "Eq. 5 is ours Stable Score Distillation, and we can further decompose above equation into two parts, and the latter is regarded as a cross-trajectory term." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 529, + 309, + 568 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 309, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 309, + 568 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {s s d}} = \\underbrace {w _ {p} \\left(\\epsilon_ {\\phi} \\left(z _ {t} , y\\right) - \\epsilon_ {\\phi} \\left(z _ {t} , \\hat {y}\\right)\\right)} _ {\\text {c r o s s - p r o m p t}} + \\underbrace {w _ {t} \\left(\\epsilon_ {\\phi} \\left(z _ {t} , \\hat {y}\\right) - \\epsilon_ {\\phi} \\left(\\hat {z} _ {t} , \\emptyset\\right)\\right)} _ {\\text {c r o s s - t r a j e c t o r y}}, \\tag {6}", + "image_path": "72b32a12f549dcf1289d56ecfccf534f31097c9ff23c6321bc0f16f9c4c03b22.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 570, + 296, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 593 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 55, + 570, + 296, + 593 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 55, + 570, + 296, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 296, + 593 + ], + "type": "inline_equation", + "content": "w_{p}" + }, + { + "bbox": [ + 55, + 570, + 296, + 593 + ], + "type": "text", + "content": " control the strength of the cross-trajectory and cross-prompt, respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 594, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 297, + 713 + ], + "type": "text", + "content": "The cross-trajectory term can be interpreted as the distance between the transitions of two latents, ensuring that the original structure remains smooth and does not change abruptly (more details are provided in the supplementary material). In Fig. 5, we can see that the cross-trajectory term can provide a strong structure constraint ability, guiding the optimization process to preserve the source image structure. Specifically, when set " + }, + { + "bbox": [ + 55, + 594, + 297, + 713 + ], + "type": "inline_equation", + "content": "w_{t} = 0" + }, + { + "bbox": [ + 55, + 594, + 297, + 713 + ], + "type": "text", + "content": ", the optimization process behaves similarly to the CSD[52] method, which fails to retain the original image structure." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 553, + 273 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 553, + 273 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 273 + ], + "type": "image", + "image_path": "2d730c9f07c9dab167114f14a75e98f13ac70892f2ebebb8bf73cce3055a8358.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 274, + 555, + 327 + ], + "lines": [ + { + "bbox": [ + 313, + 274, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 555, + 327 + ], + "type": "text", + "content": "Figure 5. The effect of increasing the strength of the prompt enhancement term " + }, + { + "bbox": [ + 313, + 274, + 555, + 327 + ], + "type": "inline_equation", + "content": "(w_{e})" + }, + { + "bbox": [ + 313, + 274, + 555, + 327 + ], + "type": "text", + "content": " and cross-trajectory term " + }, + { + "bbox": [ + 313, + 274, + 555, + 327 + ], + "type": "inline_equation", + "content": "(w_{t})" + }, + { + "bbox": [ + 313, + 274, + 555, + 327 + ], + "type": "text", + "content": ", with the cross-prompt term fixed at 7.5. Both terms contribute to prompt-aligned results, while setting " + }, + { + "bbox": [ + 313, + 274, + 555, + 327 + ], + "type": "inline_equation", + "content": "(w_{t} = 0)" + }, + { + "bbox": [ + 313, + 274, + 555, + 327 + ], + "type": "text", + "content": " leads to saturation and discard source content." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 330, + 479, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 479, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 479, + 342 + ], + "type": "text", + "content": "4.2. Improving Prompt Alignment" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 347, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 555, + 456 + ], + "type": "text", + "content": "Although Eq. 5 can achieve gradual editing results, we found Eq. 5 have similar limitation with DDS [7], which have insufficient editing strength. The editing results neither get successful editing nor retain the source image structure, often leads to little or no change in the final. Benefit from the cross-prompt editing design as Eq. 5, we can add a target prompt enhancement branch to guide the optimization process. The target prompt alignment branch will provide the direction of the target prompt, as shown in Eq. 7:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 360, + 463, + 555, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 463, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 360, + 463, + 555, + 475 + ], + "type": "interline_equation", + "content": "L _ {\\text {a l i g n}} = w _ {e} \\left(\\epsilon_ {\\phi} \\left(z _ {t}, y\\right) - \\epsilon_ {\\phi} \\left(z _ {t}, \\varnothing\\right)\\right), \\tag {7}", + "image_path": "898b9236d40bb3b3047b8a63a74f959869c10c6f7f5a70267b5de817341acda0.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 483, + 554, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 554, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 554, + 530 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 483, + 554, + 530 + ], + "type": "inline_equation", + "content": "w_{e}" + }, + { + "bbox": [ + 313, + 483, + 554, + 530 + ], + "type": "text", + "content": " is the prompt enhancement scale. As shown in Fig. 5, the synchronous scaling of both the cross-trajectory and prompt-enhancement terms results in effective visual editing outcomes." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 537, + 476, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 476, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 476, + 550 + ], + "type": "text", + "content": "4.3. Source Latent Regularization" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 555, + 555, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 555, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 555, + 650 + ], + "type": "text", + "content": "Empirically, we found that directly using latent-space loss rather than pixel-level loss can lead to optimization difficulties in local regions of 3DGS. For example, the bright spots appearing in Fig. 5. To suppress the steep gradients in these areas, we incorporate ID regularization to guide the stable optimization process. Differ with PDS [24] use source latent " + }, + { + "bbox": [ + 313, + 555, + 555, + 650 + ], + "type": "inline_equation", + "content": "\\hat{x}_0" + }, + { + "bbox": [ + 313, + 555, + 555, + 650 + ], + "type": "text", + "content": ", we can use the noisy latent " + }, + { + "bbox": [ + 313, + 555, + 555, + 650 + ], + "type": "inline_equation", + "content": "\\hat{x}_t" + }, + { + "bbox": [ + 313, + 555, + 555, + 650 + ], + "type": "text", + "content": " to avoid partial exploding gradient, as shown in Eq. 8:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 385, + 658, + 555, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 658, + 555, + 670 + ], + "spans": [ + { + "bbox": [ + 385, + 658, + 555, + 670 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {I D}} = w (t) \\cdot \\left(x _ {t} - \\hat {x} _ {t}\\right), \\tag {8}", + "image_path": "58f4dc6ae4c3a64bb8b1fcdb4fd5e4962d2e24680e4a1d2ac6a2de32e428248e.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": " is the iteration-dependent strength, designed as a decreasing function of " + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": ". Notably, the " + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": " is not necessary to well-designed in our design. Our final loss function" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16601" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 133, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 133, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 133, + 83 + ], + "type": "text", + "content": "as shown in Eq. 9:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 117, + 95, + 295, + 107 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 95, + 295, + 107 + ], + "spans": [ + { + "bbox": [ + 117, + 95, + 295, + 107 + ], + "type": "interline_equation", + "content": "L _ {\\text {f i n a l}} = L _ {\\text {s s d}} + L _ {\\text {a l i g n}} + L _ {\\text {I D}}. \\tag {9}", + "image_path": "6e9ad4a59c6bba8df3a17082607ac9bfc1188978eba2a11d75c9f97b244e7414.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 118, + 296, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 118, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 296, + 178 + ], + "type": "text", + "content": "Based on the above design, we achieve a more prompt-aligned editing method, which integrates seamlessly into the Stable Diffusion Model [40] without requiring LoRA [14] or fine-tuning. Moreover, we will further introduce our method's connection with InstructPix2Pix [2]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 187, + 182, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 187, + 182, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 187, + 182, + 201 + ], + "type": "text", + "content": "4.4. Connecting with IP2P" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 205, + 296, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 296, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 296, + 242 + ], + "type": "text", + "content": "The final design of our method is shown in Eq 9. We found that ours edit gard provide new angle to understand about InstructP2P [2] one-step reverse sampling." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 251, + 293, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 293, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 293, + 308 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\epsilon_ {\\theta} \\left(z _ {t}, c _ {I}, c _ {T}\\right) = \\epsilon_ {\\theta} \\left(z _ {t}, \\varnothing , \\varnothing\\right) \\\\ + s _ {I} \\left(\\epsilon_ {\\theta} \\left(z _ {t}, c _ {I}, \\varnothing\\right) - \\epsilon_ {\\theta} \\left(z _ {t}, \\varnothing , \\varnothing\\right)\\right) \\\\ + s _ {T} \\left(\\epsilon_ {\\theta} \\left(z _ {t}, c _ {I}, c _ {T}\\right) - \\epsilon_ {\\theta} \\left(z _ {t}, c _ {I}, \\varnothing\\right)\\right), \\tag {10} \\\\ \\end{array}", + "image_path": "0d85d3f12231718ef24f0f4ef635555992cfa2d4540dd4440c2d2c5cf2c98c36.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 319, + 296, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 462 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 319, + 296, + 462 + ], + "type": "inline_equation", + "content": "c_{I}" + }, + { + "bbox": [ + 55, + 319, + 296, + 462 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 319, + 296, + 462 + ], + "type": "inline_equation", + "content": "c_{T}" + }, + { + "bbox": [ + 55, + 319, + 296, + 462 + ], + "type": "text", + "content": " are input-image and instruction prompt separately, " + }, + { + "bbox": [ + 55, + 319, + 296, + 462 + ], + "type": "inline_equation", + "content": "s_{I}" + }, + { + "bbox": [ + 55, + 319, + 296, + 462 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 319, + 296, + 462 + ], + "type": "inline_equation", + "content": "s_{T}" + }, + { + "bbox": [ + 55, + 319, + 296, + 462 + ], + "type": "text", + "content": " are the source image control and instruction prompt control strength. The Eq. 10 is the InstructP2P one-step reverse sampling, which can provide the direction of the target prompt. We can see that the InstructP2P is the simple version of our method, the middle term of Eq. 10 is cross-trajectory regularization, and the last term of Eq. 10 is ours cross-prompt term. Simply put, as analyzing the Eq. 5, subtracting the constant correction term " + }, + { + "bbox": [ + 55, + 319, + 296, + 462 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\hat{z}_t; \\varnothing; \\varnothing)" + }, + { + "bbox": [ + 55, + 319, + 296, + 462 + ], + "type": "text", + "content": " is edit grad. Ours method reveals that apply DDS loss in the InstructP2P model can only editing branch, and don't have to provide the source branch." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 474, + 137, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 137, + 488 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 137, + 488 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 495, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 296, + 567 + ], + "type": "text", + "content": "In this section, we conduct editing experiments across two types of parameterized images. Section 5.1 evaluates the effectiveness of our method on 3D Scenes Editing, and Section 5.2 evaluates the effectiveness of our method on 2D Image Editing. We also conduct ablation studies to analyze the effectiveness of ours components in Section 5.3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 575, + 163, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 163, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 163, + 588 + ], + "type": "text", + "content": "5.1.3D Scenes Editing" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 594, + 296, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 640 + ], + "type": "text", + "content": "Dataset. To evaluate the effectiveness of our method, we conduct experiments on the scenes from IN2N [9] and other real-world datasets, including LLFF [32] and Mip-Nerf360 [1]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 642, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 297, + 714 + ], + "type": "text", + "content": "Baselines. We compare our method with several state-of-the-art inversion methods. We use 3DGS [21] as the 3D representation, and compare our method with InstructNerf2Nerf [9], DDS [10], GS-Edit [6], and DGE [5]. For fairness, we implement the DDS version based on the official GS-Edit code. PDS [24] is designed for addition of objects to unspecified" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 339, + 83, + 526, + 152 + ], + "blocks": [ + { + "bbox": [ + 328, + 71, + 539, + 82 + ], + "lines": [ + { + "bbox": [ + 328, + 71, + 539, + 82 + ], + "spans": [ + { + "bbox": [ + 328, + 71, + 539, + 82 + ], + "type": "text", + "content": "Table 1. Quantitative evaluations under 3D editing scenes." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 339, + 83, + 526, + 152 + ], + "lines": [ + { + "bbox": [ + 339, + 83, + 526, + 152 + ], + "spans": [ + { + "bbox": [ + 339, + 83, + 526, + 152 + ], + "type": "table", + "html": "
MethodCLIP Sim.↑Sim Dir.↑User Study ↑
IN2N0.16760.070714.54%
DDS0.17800.04015.45%
GS-Editor0.17580.042914.54%
DGE0.17580.056323.63%
Ours0.18460.077341.81%
", + "image_path": "65f60e7560a5debb561683f4d96607bafb4cde437da3657fcef787a51a42b0fc.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 156, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 555, + 180 + ], + "type": "text", + "content": "regions, we will provide the comparison results in supplementary material." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 181, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 181, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 181, + 555, + 312 + ], + "type": "text", + "content": "Evaluation Metrics. We follow common practice [5, 6, 9] to evaluate the effectiveness of our method. CLIP Similarity is to evaluate the alignment between the render images and the target prompts, i.e., the cosine similarity between the text and image embeddings encoded by CLIP. Specifically, follow DGE [5], randomly sample 20 camera poses to evaluate. CLIP Directional Similarity is to measure the editing effect, i.e., the cosine similarity between the image and text editing directions (target embeddings minus source embeddings). We evaluate all methods on 6 different scenes and 10 different prompts." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 313, + 556, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 313, + 556, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 556, + 503 + ], + "type": "text", + "content": "Results. We begin by evaluating our method, starting with a qualitative assessment. In Fig. 6, we present a comparison of results with competing methods. Our approach generates more visually appealing images that are better aligned with the editing instructions. In contrast, methods based on the Iterative Dataset Update (IDU) strategy, such as IN2N [9] and GS-Editor [6], fail to produce the desired editing outcomes, resulting in blurrier or lower-fidelity reconstructions and noticeable artifacts. For example, in the scene of \"Spider-Man with a mask\", IN2N generates a mask with reduced fidelity, while GS-Editor produces a low-detail mask. In the multi-view consistency setup, DGE [5] performs well on common attributes but is constrained to \"rainbow\" editing and tends to generate artifacts outside the segmentation mask. Our method works seamlessly with masks, producing results with rich details." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 506, + 556, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 556, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 556, + 590 + ], + "type": "text", + "content": "In Tab. 1, We present a quantitative comparison. Our method outperforms the baselines in terms of CLIP Similarity and CLIP Directional Similarity. Notably, Dire Sim is not sensitive with the editing quality, much focus on the instruction attributes. We conducted a user study with a survey of 55 participants to evaluate the editing quality. The results show that our method received the most popular votes." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 598, + 421, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 598, + 421, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 598, + 421, + 611 + ], + "type": "text", + "content": "5.2. 2D Image Editing" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 616, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 555, + 676 + ], + "type": "text", + "content": "Dataset. To evaluate the effectiveness of our method, we conduct experiments on the PIE-Bench dataset proposed by PNPIv [19], which consists of 700 images with 9 editing types. Each image is annotated with the source and target prompts." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "Baselines. We compare our method with several classical editing methods based on DDIM [43] inversion, including P2P [11], PNP [46] and MasaCtrl [3]. For optimization-" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16602" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 83, + 153, + 158 + ], + "blocks": [ + { + "bbox": [ + 61, + 83, + 153, + 158 + ], + "lines": [ + { + "bbox": [ + 61, + 83, + 153, + 158 + ], + "spans": [ + { + "bbox": [ + 61, + 83, + 153, + 158 + ], + "type": "image", + "image_path": "13ce1d72ac22be6ba8aaad7589261cefbec4998271c9b205f0bb383fe9db2855.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 157, + 82, + 247, + 158 + ], + "blocks": [ + { + "bbox": [ + 157, + 82, + 247, + 158 + ], + "lines": [ + { + "bbox": [ + 157, + 82, + 247, + 158 + ], + "spans": [ + { + "bbox": [ + 157, + 82, + 247, + 158 + ], + "type": "image", + "image_path": "f2805ad1f9a08b27f8a067ef9505fff4aad4faacd96bb493599b24da55750bc8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 253, + 82, + 343, + 158 + ], + "blocks": [ + { + "bbox": [ + 253, + 82, + 343, + 158 + ], + "lines": [ + { + "bbox": [ + 253, + 82, + 343, + 158 + ], + "spans": [ + { + "bbox": [ + 253, + 82, + 343, + 158 + ], + "type": "image", + "image_path": "35a59f56e4c89d5ff0a5cd0b0f38718f9b1613950998792f6b07be31b594739c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 349, + 82, + 438, + 158 + ], + "blocks": [ + { + "bbox": [ + 349, + 82, + 438, + 158 + ], + "lines": [ + { + "bbox": [ + 349, + 82, + 438, + 158 + ], + "spans": [ + { + "bbox": [ + 349, + 82, + 438, + 158 + ], + "type": "image", + "image_path": "f9cedc028a8c2f06f38b8fd7bb2eee568a7f3d437ab0c979f75559f4afed319b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 443, + 83, + 534, + 158 + ], + "blocks": [ + { + "bbox": [ + 443, + 83, + 534, + 158 + ], + "lines": [ + { + "bbox": [ + 443, + 83, + 534, + 158 + ], + "spans": [ + { + "bbox": [ + 443, + 83, + 534, + 158 + ], + "type": "image", + "image_path": "878a0b066fa3d1b0f5600b04680a85667a95f9fce37bf619173f7e6750a5472b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 59, + 173, + 154, + 258 + ], + "blocks": [ + { + "bbox": [ + 59, + 173, + 154, + 258 + ], + "lines": [ + { + "bbox": [ + 59, + 173, + 154, + 258 + ], + "spans": [ + { + "bbox": [ + 59, + 173, + 154, + 258 + ], + "type": "image", + "image_path": "0bd1a105b87c004a30efa34371eeaf0d431a6e913e790f2dc689c63a78c01e1e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 154, + 173, + 250, + 258 + ], + "blocks": [ + { + "bbox": [ + 221, + 163, + 395, + 173 + ], + "lines": [ + { + "bbox": [ + 221, + 163, + 395, + 173 + ], + "spans": [ + { + "bbox": [ + 221, + 163, + 395, + 173 + ], + "type": "text", + "content": "A man wearing T-shirt with a pineapple pattern" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 154, + 173, + 250, + 258 + ], + "lines": [ + { + "bbox": [ + 154, + 173, + 250, + 258 + ], + "spans": [ + { + "bbox": [ + 154, + 173, + 250, + 258 + ], + "type": "image", + "image_path": "6e56458c83a5e5663d8cba8b7d62b6ceca5b4909a68f04cb7c5ee27c331b792c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 250, + 173, + 345, + 258 + ], + "blocks": [ + { + "bbox": [ + 250, + 173, + 345, + 258 + ], + "lines": [ + { + "bbox": [ + 250, + 173, + 345, + 258 + ], + "spans": [ + { + "bbox": [ + 250, + 173, + 345, + 258 + ], + "type": "image", + "image_path": "07bfd5534b16378060d25b3b2e02843f14eeb98b8e1f139bfab069accea1a362.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 346, + 173, + 441, + 258 + ], + "blocks": [ + { + "bbox": [ + 346, + 173, + 441, + 258 + ], + "lines": [ + { + "bbox": [ + 346, + 173, + 441, + 258 + ], + "spans": [ + { + "bbox": [ + 346, + 173, + 441, + 258 + ], + "type": "image", + "image_path": "d846415dd336729eeee72a7b26c7afd54e6ded75b8f11cc135cb1e7c7410e36f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 441, + 173, + 537, + 258 + ], + "blocks": [ + { + "bbox": [ + 441, + 173, + 537, + 258 + ], + "lines": [ + { + "bbox": [ + 441, + 173, + 537, + 258 + ], + "spans": [ + { + "bbox": [ + 441, + 173, + 537, + 258 + ], + "type": "image", + "image_path": "8da663ffc801f6d2c6aa9c769369662d88bb36f0a503105e0c7923847be8f956.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 59, + 270, + 154, + 342 + ], + "blocks": [ + { + "bbox": [ + 59, + 270, + 154, + 342 + ], + "lines": [ + { + "bbox": [ + 59, + 270, + 154, + 342 + ], + "spans": [ + { + "bbox": [ + 59, + 270, + 154, + 342 + ], + "type": "image", + "image_path": "a1f6047a6c410edb2977679a0dd1d46fc3c84fe44b9194420f848cb970991b8e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 154, + 270, + 250, + 342 + ], + "blocks": [ + { + "bbox": [ + 154, + 270, + 250, + 342 + ], + "lines": [ + { + "bbox": [ + 154, + 270, + 250, + 342 + ], + "spans": [ + { + "bbox": [ + 154, + 270, + 250, + 342 + ], + "type": "image", + "image_path": "e20c01a14c58cf70b717178cb863b1b7e64edf64dec5eb4e5eb6d28ef4aca5f8.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 250, + 270, + 345, + 342 + ], + "blocks": [ + { + "bbox": [ + 268, + 259, + 342, + 270 + ], + "lines": [ + { + "bbox": [ + 268, + 259, + 342, + 270 + ], + "spans": [ + { + "bbox": [ + 268, + 259, + 342, + 270 + ], + "type": "text", + "content": "Rainbow horn fossil" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 250, + 270, + 345, + 342 + ], + "lines": [ + { + "bbox": [ + 250, + 270, + 345, + 342 + ], + "spans": [ + { + "bbox": [ + 250, + 270, + 345, + 342 + ], + "type": "image", + "image_path": "f7b88149ae0d3f96d9139d4b733219f7502d881a0888483373573daef3935c90.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 346, + 270, + 441, + 342 + ], + "blocks": [ + { + "bbox": [ + 346, + 270, + 441, + 342 + ], + "lines": [ + { + "bbox": [ + 346, + 270, + 441, + 342 + ], + "spans": [ + { + "bbox": [ + 346, + 270, + 441, + 342 + ], + "type": "image", + "image_path": "c7cb61146101236bca3aac01285297c694df1bdcb1cf9f604bfc7713e409c9a2.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 441, + 270, + 537, + 342 + ], + "blocks": [ + { + "bbox": [ + 441, + 270, + 537, + 342 + ], + "lines": [ + { + "bbox": [ + 441, + 270, + 537, + 342 + ], + "spans": [ + { + "bbox": [ + 441, + 270, + 537, + 342 + ], + "type": "image", + "image_path": "25227a4b1e6df7d5e6be676a438e4bd74903c1c649de2c99b23dafe131223ea8.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 59, + 354, + 154, + 418 + ], + "blocks": [ + { + "bbox": [ + 59, + 354, + 154, + 418 + ], + "lines": [ + { + "bbox": [ + 59, + 354, + 154, + 418 + ], + "spans": [ + { + "bbox": [ + 59, + 354, + 154, + 418 + ], + "type": "image", + "image_path": "18b19580bde98279c991ce488677c952c1566e7b4919d77ef6e2daa542bc7d2a.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 96, + 419, + 141, + 429 + ], + "lines": [ + { + "bbox": [ + 96, + 419, + 141, + 429 + ], + "spans": [ + { + "bbox": [ + 96, + 419, + 141, + 429 + ], + "type": "text", + "content": "Input Views" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 430, + 555, + 452 + ], + "lines": [ + { + "bbox": [ + 55, + 430, + 555, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 555, + 452 + ], + "type": "text", + "content": "Figure 6. Qualitative comparisons with related works. SDS demonstrates outstanding performance in effectively preserve source structure in the modified region." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 154, + 354, + 250, + 418 + ], + "blocks": [ + { + "bbox": [ + 236, + 344, + 373, + 354 + ], + "lines": [ + { + "bbox": [ + 236, + 344, + 373, + 354 + ], + "spans": [ + { + "bbox": [ + 236, + 344, + 373, + 354 + ], + "type": "text", + "content": "A tree stump with some leaves on fire" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 154, + 354, + 250, + 418 + ], + "lines": [ + { + "bbox": [ + 154, + 354, + 250, + 418 + ], + "spans": [ + { + "bbox": [ + 154, + 354, + 250, + 418 + ], + "type": "image", + "image_path": "2d130333d4f3face37b0455d5e73cf8928624983748be7a9bc200e89ec888418.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 188, + 420, + 235, + 429 + ], + "lines": [ + { + "bbox": [ + 188, + 420, + 235, + 429 + ], + "spans": [ + { + "bbox": [ + 188, + 420, + 235, + 429 + ], + "type": "text", + "content": "InstructN2N" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 250, + 354, + 345, + 418 + ], + "blocks": [ + { + "bbox": [ + 250, + 354, + 345, + 418 + ], + "lines": [ + { + "bbox": [ + 250, + 354, + 345, + 418 + ], + "spans": [ + { + "bbox": [ + 250, + 354, + 345, + 418 + ], + "type": "image", + "image_path": "f40b944de1aed3a3a1b649e778e674236769dec39541cb27e4306f0ac65e3458.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 276, + 420, + 334, + 430 + ], + "lines": [ + { + "bbox": [ + 276, + 420, + 334, + 430 + ], + "spans": [ + { + "bbox": [ + 276, + 420, + 334, + 430 + ], + "type": "text", + "content": "GaussianEditor" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 346, + 354, + 441, + 418 + ], + "blocks": [ + { + "bbox": [ + 346, + 354, + 441, + 418 + ], + "lines": [ + { + "bbox": [ + 346, + 354, + 441, + 418 + ], + "spans": [ + { + "bbox": [ + 346, + 354, + 441, + 418 + ], + "type": "image", + "image_path": "903d458fc60da3419c0cae26b709da119a308a7789f4201121dac0279e03ad19.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 419, + 408, + 429 + ], + "lines": [ + { + "bbox": [ + 388, + 419, + 408, + 429 + ], + "spans": [ + { + "bbox": [ + 388, + 419, + 408, + 429 + ], + "type": "text", + "content": "DGE" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 441, + 354, + 537, + 418 + ], + "blocks": [ + { + "bbox": [ + 441, + 354, + 537, + 418 + ], + "lines": [ + { + "bbox": [ + 441, + 354, + 537, + 418 + ], + "spans": [ + { + "bbox": [ + 441, + 354, + 537, + 418 + ], + "type": "image", + "image_path": "935d95a40c02d21f1d8cdfcfc5ce9531ed9d7c5608c90f6ab0547459fea4b6aa.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 482, + 420, + 501, + 429 + ], + "lines": [ + { + "bbox": [ + 482, + 420, + 501, + 429 + ], + "spans": [ + { + "bbox": [ + 482, + 420, + 501, + 429 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 459, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 295, + 506 + ], + "type": "text", + "content": "based editing method, we compare with NT [34] and StyleD [27]. Besides, we report the comparison with DT [19]. Further, we compare with DDS [10] and its extended method CDS [35]." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 55, + 508, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 296, + 616 + ], + "type": "text", + "content": "Evaluation Metrics. We follow DT [19] which uses several metrics to evaluate our method. We use the Structure Distance assessed by DINO score [4] to evaluate the structure distance between original and edited images. We also introduce several metrics to evaluating the background preservation, which includes LPIPS [55] and MSE. Besides, we introduce CLIP Similarity [48] to evaluate the text-image consistency between edited images and corresponding target editing text prompts." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": "Results. We present a qualitative comparison of our method with competitors in Fig. 7. Our method generates images that are more aligned with the target prompts and preserve the source structure. In \"blue butterfly\", ours successfully changes the color of the butterfly to blue, while DDS [10] and CDS [35] generate similar color from source. Especially, ours method successfully changes the style of the image and generates appealing results, which is challenging for DDS-" + } + ] + } + ], + "index": 32 + }, + { + "type": "table", + "bbox": [ + 314, + 468, + 563, + 594 + ], + "blocks": [ + { + "bbox": [ + 334, + 457, + 533, + 468 + ], + "lines": [ + { + "bbox": [ + 334, + 457, + 533, + 468 + ], + "spans": [ + { + "bbox": [ + 334, + 457, + 533, + 468 + ], + "type": "text", + "content": "Table 2. Quantitative evaluation in PIE-Bench dataset." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 468, + 563, + 594 + ], + "lines": [ + { + "bbox": [ + 314, + 468, + 563, + 594 + ], + "spans": [ + { + "bbox": [ + 314, + 468, + 563, + 594 + ], + "type": "table", + "html": "
MethodDistance×103↓LPIPS×103↓MSE×104↓CLIP↑
DDIM + P2P69.43208.80219.8825.01
DDIM + PNP28.22113.4683.6425.41
DDIM + MasaCtrl28.38106.6286.9723.96
NT + P2P13.4460.6735.8624.75
StyleD + P2P11.6566.1038.6324.78
DT + P2P11.6554.5532.8625.02
DDS14.7450.5845.0925.86
DDS + CDS7.1533.1425.2924.96
Ours28.1382.4386.6426.94
Ours + CDS6.9032.1524.2125.12
", + "image_path": "a2097ad2573be0bb9a7dc491a9722a9c285d2ca961f2adcac4f0b6d2686c2428.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "table_body" + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 603, + 555, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 603, + 555, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 603, + 555, + 674 + ], + "type": "text", + "content": "based methods. Compared to optimization-based methods, NT [34] preserves the general source structure during the inversion process, but tends to discard some content as evident in the distortion of the girl's fingers in Fig. 7. Additionally, due to limitations in the editing methods, the editing results are unsuccessful." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "In Tab. 2, we present a quantitative comparison. Our method strikes a balance between structure distance and editability. Notably, we observe that the distance is much" + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 222, + 69, + 388, + 80 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 69, + 388, + 80 + ], + "spans": [ + { + "bbox": [ + 222, + 69, + 388, + 80 + ], + "type": "text", + "content": "A photo of a man " + }, + { + "bbox": [ + 222, + 69, + 388, + 80 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 222, + 69, + 388, + 80 + ], + "type": "text", + "content": " Spider man with a mask" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16603" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 93, + 72, + 257, + 81 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 72, + 257, + 81 + ], + "spans": [ + { + "bbox": [ + 93, + 72, + 257, + 81 + ], + "type": "text", + "content": "A blue and white bird " + }, + { + "bbox": [ + 93, + 72, + 257, + 81 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 93, + 72, + 257, + 81 + ], + "type": "text", + "content": " butterfly sits on a branch" + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 83, + 103, + 129 + ], + "blocks": [ + { + "bbox": [ + 58, + 83, + 103, + 129 + ], + "lines": [ + { + "bbox": [ + 58, + 83, + 103, + 129 + ], + "spans": [ + { + "bbox": [ + 58, + 83, + 103, + 129 + ], + "type": "image", + "image_path": "cf61a5a6f51cf7b19f3024fd9ff691faa65d02314a8a45379864b4b8f6716f58.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 104, + 83, + 151, + 129 + ], + "blocks": [ + { + "bbox": [ + 104, + 83, + 151, + 129 + ], + "lines": [ + { + "bbox": [ + 104, + 83, + 151, + 129 + ], + "spans": [ + { + "bbox": [ + 104, + 83, + 151, + 129 + ], + "type": "image", + "image_path": "00c0fde9295dd7240d285eefe4980fad3185dc361640f4013af1364939c2fb0b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 153, + 83, + 199, + 129 + ], + "blocks": [ + { + "bbox": [ + 153, + 83, + 199, + 129 + ], + "lines": [ + { + "bbox": [ + 153, + 83, + 199, + 129 + ], + "spans": [ + { + "bbox": [ + 153, + 83, + 199, + 129 + ], + "type": "image", + "image_path": "02fdae951c4a701eb43ef5c93133d3db311c1268b309d76794f8f3f1bfe3617a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 200, + 83, + 246, + 129 + ], + "blocks": [ + { + "bbox": [ + 200, + 83, + 246, + 129 + ], + "lines": [ + { + "bbox": [ + 200, + 83, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 200, + 83, + 246, + 129 + ], + "type": "image", + "image_path": "5954546facdaca4fea820fb04998d791af9fedee2c1c451396c01012d381bced.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 247, + 83, + 294, + 129 + ], + "blocks": [ + { + "bbox": [ + 247, + 83, + 294, + 129 + ], + "lines": [ + { + "bbox": [ + 247, + 83, + 294, + 129 + ], + "spans": [ + { + "bbox": [ + 247, + 83, + 294, + 129 + ], + "type": "image", + "image_path": "e36d09ce7a7403e49e4865b73ecd12453b00bfbd0c912fbb787a148ada4e0f45.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 142, + 555, + 197 + ], + "lines": [ + { + "bbox": [ + 313, + 142, + 555, + 197 + ], + "spans": [ + { + "bbox": [ + 313, + 142, + 555, + 197 + ], + "type": "text", + "content": "Figure 8. Effect of source latent regularization. In most experiments, the source ID term helps prevent partial gradient explosion. In the left image, the yellow arrow highlights an irregular color. As the weight of the ID term increases, the color becomes more regular, however, the spider on the character's chest is affected." + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 130, + 284, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 130, + 284, + 140 + ], + "spans": [ + { + "bbox": [ + 67, + 130, + 284, + 140 + ], + "type": "text", + "content": "Kids crayon drawing of a man with a long beard and a long sword" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 58, + 140, + 104, + 186 + ], + "blocks": [ + { + "bbox": [ + 58, + 140, + 104, + 186 + ], + "lines": [ + { + "bbox": [ + 58, + 140, + 104, + 186 + ], + "spans": [ + { + "bbox": [ + 58, + 140, + 104, + 186 + ], + "type": "image", + "image_path": "6844614c16f7a3153254083073477af46f2a162e8cd2b1871f19b12700d2f463.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 105, + 140, + 152, + 186 + ], + "blocks": [ + { + "bbox": [ + 105, + 140, + 152, + 186 + ], + "lines": [ + { + "bbox": [ + 105, + 140, + 152, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 140, + 152, + 186 + ], + "type": "image", + "image_path": "51f3b4407f6e9ff81716cc3792649a66a201a15cd0fc834b0f805d0538b9c5b3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 153, + 140, + 199, + 186 + ], + "blocks": [ + { + "bbox": [ + 153, + 140, + 199, + 186 + ], + "lines": [ + { + "bbox": [ + 153, + 140, + 199, + 186 + ], + "spans": [ + { + "bbox": [ + 153, + 140, + 199, + 186 + ], + "type": "image", + "image_path": "bfa1737ca5a4b77528bf1ff09b1cf5a87757c8b1df837ba69eaa24f68b197eed.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 200, + 140, + 246, + 186 + ], + "blocks": [ + { + "bbox": [ + 200, + 140, + 246, + 186 + ], + "lines": [ + { + "bbox": [ + 200, + 140, + 246, + 186 + ], + "spans": [ + { + "bbox": [ + 200, + 140, + 246, + 186 + ], + "type": "image", + "image_path": "2b4433e62d7577275a90ed28b5e200c6c80dcaa7a94d5e709bc2a5f2d131d9d8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 247, + 140, + 294, + 186 + ], + "blocks": [ + { + "bbox": [ + 247, + 140, + 294, + 186 + ], + "lines": [ + { + "bbox": [ + 247, + 140, + 294, + 186 + ], + "spans": [ + { + "bbox": [ + 247, + 140, + 294, + 186 + ], + "type": "image", + "image_path": "0e712de57c892101d9095f9a7bcd6c60c57744ef59ecdab4bae094786d536f5f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 188, + 284, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 188, + 284, + 198 + ], + "spans": [ + { + "bbox": [ + 67, + 188, + 284, + 198 + ], + "type": "text", + "content": "Black and white sketch of a young girl with painted hands and face" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 58, + 200, + 104, + 245 + ], + "blocks": [ + { + "bbox": [ + 58, + 200, + 104, + 245 + ], + "lines": [ + { + "bbox": [ + 58, + 200, + 104, + 245 + ], + "spans": [ + { + "bbox": [ + 58, + 200, + 104, + 245 + ], + "type": "image", + "image_path": "9359ff7cb352ff71f1503aa5d09b4a88e6581ee50d98942dd30bd524fd6d949f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 105, + 200, + 152, + 245 + ], + "blocks": [ + { + "bbox": [ + 105, + 200, + 152, + 245 + ], + "lines": [ + { + "bbox": [ + 105, + 200, + 152, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 152, + 245 + ], + "type": "image", + "image_path": "76395dace40a6df6c10cbf3fd0eee41fb5cd2353c620d1a0c34f7ee8645ec28d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 153, + 200, + 199, + 245 + ], + "blocks": [ + { + "bbox": [ + 153, + 200, + 199, + 245 + ], + "lines": [ + { + "bbox": [ + 153, + 200, + 199, + 245 + ], + "spans": [ + { + "bbox": [ + 153, + 200, + 199, + 245 + ], + "type": "image", + "image_path": "c32718a6a8bcbb0e8d88f7671a3fd7d559862444a6a1321639464486c1371461.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 200, + 200, + 246, + 245 + ], + "blocks": [ + { + "bbox": [ + 200, + 200, + 246, + 245 + ], + "lines": [ + { + "bbox": [ + 200, + 200, + 246, + 245 + ], + "spans": [ + { + "bbox": [ + 200, + 200, + 246, + 245 + ], + "type": "image", + "image_path": "48e72db346cae2dde2118d0dc124f8cae46706aa7ff7f730f05f372ca883328e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 247, + 200, + 294, + 245 + ], + "blocks": [ + { + "bbox": [ + 247, + 200, + 294, + 245 + ], + "lines": [ + { + "bbox": [ + 247, + 200, + 294, + 245 + ], + "spans": [ + { + "bbox": [ + 247, + 200, + 294, + 245 + ], + "type": "image", + "image_path": "e5450e97c47ce50415b03b58ab28db0b3632bdc842d45b5524c55ebae1f10cf5.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 70, + 247, + 279, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 247, + 279, + 256 + ], + "spans": [ + { + "bbox": [ + 70, + 247, + 279, + 256 + ], + "type": "text", + "content": "A monkey " + }, + { + "bbox": [ + 70, + 247, + 279, + 256 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 70, + 247, + 279, + 256 + ], + "type": "text", + "content": " man wearing colorful goggles and a colorful scarf" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 58, + 257, + 104, + 303 + ], + "blocks": [ + { + "bbox": [ + 58, + 257, + 104, + 303 + ], + "lines": [ + { + "bbox": [ + 58, + 257, + 104, + 303 + ], + "spans": [ + { + "bbox": [ + 58, + 257, + 104, + 303 + ], + "type": "image", + "image_path": "fa1e77d8489ed52e472799bfeb95b24b43cddf11fcbe46268f65c13e4f205dfa.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 314, + 295, + 335 + ], + "lines": [ + { + "bbox": [ + 55, + 314, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 314, + 295, + 335 + ], + "type": "text", + "content": "Figure 7. Comparison of different editing methods on various objects and styles." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 105, + 257, + 152, + 303 + ], + "blocks": [ + { + "bbox": [ + 105, + 257, + 152, + 303 + ], + "lines": [ + { + "bbox": [ + 105, + 257, + 152, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 152, + 303 + ], + "type": "image", + "image_path": "cfd0cb6fd2309746225d4176b5723e1b12697f932a48a12cbea1fd8ba17b0d9b.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 153, + 257, + 199, + 303 + ], + "blocks": [ + { + "bbox": [ + 153, + 257, + 199, + 303 + ], + "lines": [ + { + "bbox": [ + 153, + 257, + 199, + 303 + ], + "spans": [ + { + "bbox": [ + 153, + 257, + 199, + 303 + ], + "type": "image", + "image_path": "3a4b9116d403bc2969adba0180a410d118bd10aab311aee5251702f5d76f5aca.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 200, + 256, + 246, + 303 + ], + "blocks": [ + { + "bbox": [ + 200, + 256, + 246, + 303 + ], + "lines": [ + { + "bbox": [ + 200, + 256, + 246, + 303 + ], + "spans": [ + { + "bbox": [ + 200, + 256, + 246, + 303 + ], + "type": "image", + "image_path": "a6ad37aba4a9dfe6cf5eea7ce806d6c952033df2bedfc2258e2f35a8a893b5ee.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 247, + 256, + 294, + 303 + ], + "blocks": [ + { + "bbox": [ + 247, + 256, + 294, + 303 + ], + "lines": [ + { + "bbox": [ + 247, + 256, + 294, + 303 + ], + "spans": [ + { + "bbox": [ + 247, + 256, + 294, + 303 + ], + "type": "image", + "image_path": "7e7f55139e27b356337557a9444fac7e08ba0dee13c0e8baf828bfc193734545.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 70, + 304, + 99, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 304, + 99, + 313 + ], + "spans": [ + { + "bbox": [ + 70, + 304, + 99, + 313 + ], + "type": "text", + "content": "Source" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 121, + 304, + 141, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 304, + 141, + 313 + ], + "spans": [ + { + "bbox": [ + 121, + 304, + 141, + 313 + ], + "type": "text", + "content": "DDS" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 163, + 304, + 182, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 163, + 304, + 182, + 313 + ], + "spans": [ + { + "bbox": [ + 163, + 304, + 182, + 313 + ], + "type": "text", + "content": "CDS" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 205, + 304, + 238, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 304, + 238, + 313 + ], + "spans": [ + { + "bbox": [ + 205, + 304, + 238, + 313 + ], + "type": "text", + "content": "NT+P2P" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 261, + 304, + 280, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 304, + 280, + 313 + ], + "spans": [ + { + "bbox": [ + 261, + 304, + 280, + 313 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 55, + 342, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 510 + ], + "type": "text", + "content": "lower when no editing occurs, which is particularly visible in DDS-based methods applied to style editing. Our methods achieving better editing score but with a slightly higher structure distance. In terms of precise structure preservation, when combined with CDS [35], it achieves good preservation of the un-edited areas. Our full model achieves the best performance in the CLIP Similarity metric, demonstrating the effectiveness of our prompt enhancement branch. While CDS excels in preserving unedited regions, it suffers from the inferior editability of DDS-based methods. Optimization-based methods [27, 34] refine the inversion process, achieving excellent performance in structure preservation. However, they struggle with editing methods (like P2P), resulting in limited editability." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 55, + 517, + 156, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 156, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 156, + 528 + ], + "type": "text", + "content": "5.3. Ablation Studies" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 55, + 534, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 582 + ], + "type": "text", + "content": "In this section, we conduct an ablation experiment to analyze different choices in our SSD. Due to space limitations, we first present a qualitative evaluation in the main text. Please refer to the Supp. for quantitative evaluation." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 55, + 582, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 700 + ], + "type": "text", + "content": "The effectiveness of cross-trajectory. In Sec. 4.1, we have analyzed the necessity of cross-trajectory. This term makes the optimization process more stable and provides the source content regularization in ours design, which is also the key difference between ours and Classifier Score Distillation(CSD) [52]. In Fig. 1 and Fig. 5, we present the comparison of the results with and without cross-trajectory. The results show that the cross-trajectory term can provide the direction of generating high-quality images. Please refer to the supplement for more details." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "type": "text", + "content": "The effectiveness of prompt-enhancement. The enhance-" + } + ] + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 318, + 70, + 393, + 126 + ], + "blocks": [ + { + "bbox": [ + 318, + 70, + 393, + 126 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 393, + 126 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 393, + 126 + ], + "type": "image", + "image_path": "ac2d77153e3c555c622e7de06d3621d0ff7940ddb8aa9aa6a7b0d6c56a6848f4.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 130, + 384, + 140 + ], + "lines": [ + { + "bbox": [ + 326, + 130, + 384, + 140 + ], + "spans": [ + { + "bbox": [ + 326, + 130, + 384, + 140 + ], + "type": "text", + "content": "w/o ID Regular" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 395, + 70, + 471, + 126 + ], + "blocks": [ + { + "bbox": [ + 395, + 70, + 471, + 126 + ], + "lines": [ + { + "bbox": [ + 395, + 70, + 471, + 126 + ], + "spans": [ + { + "bbox": [ + 395, + 70, + 471, + 126 + ], + "type": "image", + "image_path": "10d4cd7612166efab2a8900afe039440be70e8834f230d6440cc0ea132aeb552.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 401, + 130, + 466, + 140 + ], + "lines": [ + { + "bbox": [ + 401, + 130, + 466, + 140 + ], + "spans": [ + { + "bbox": [ + 401, + 130, + 466, + 140 + ], + "type": "text", + "content": "ID Regular " + }, + { + "bbox": [ + 401, + 130, + 466, + 140 + ], + "type": "inline_equation", + "content": "\\times 1.0" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 474, + 70, + 549, + 126 + ], + "blocks": [ + { + "bbox": [ + 474, + 70, + 549, + 126 + ], + "lines": [ + { + "bbox": [ + 474, + 70, + 549, + 126 + ], + "spans": [ + { + "bbox": [ + 474, + 70, + 549, + 126 + ], + "type": "image", + "image_path": "a63e838151873b6b242f4712f0ede5ce0fc27a1a5c7d6387d9a6eb7bcb3c9b6a.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 479, + 130, + 544, + 140 + ], + "lines": [ + { + "bbox": [ + 479, + 130, + 544, + 140 + ], + "spans": [ + { + "bbox": [ + 479, + 130, + 544, + 140 + ], + "type": "text", + "content": "ID Regular " + }, + { + "bbox": [ + 479, + 130, + 544, + 140 + ], + "type": "inline_equation", + "content": "\\times 2.0" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 204, + 555, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 555, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 555, + 276 + ], + "type": "text", + "content": "ment of the target prompt branch is another key component in our method, which is designed to improve the editability aligned with the target prompt in 2D-image task. In Fig. 7, we observe a clear distinction from DDS in style editing. The results show that the prompt-enhancement term effectively overcomes the challenging from style editing." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 313, + 276, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 555, + 361 + ], + "type": "text", + "content": "The effectiveness of ID regularization. ID regularization is designed to ensure stable optimization in 3DGS. In Fig. 8, we compare results with and without ID regularization. The area marked by the yellow arrow highlights its effect in 3D scene editing. However, excessive ID regularization may constrain editing quality by limiting certain attributes, presenting a trade-off in our design." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 313, + 370, + 552, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 552, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 552, + 384 + ], + "type": "text", + "content": "6. Conclusions, Limitations, and Future Work" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 313, + 391, + 555, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 555, + 475 + ], + "type": "text", + "content": "In this work, we propose a novel method for text-guided image editing, capable of handling both 3D scenes and 2D images. Our approach is built on a score distillation framework that leverages the powerful priors of diffusion models. For editing tasks, we design an effective optimization strategy that produces high-quality results aligned with target prompts while ensuring stable and consistent optimization." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 313, + 475, + 556, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 556, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 556, + 605 + ], + "type": "text", + "content": "Our method achieves state-of-the-art performance in both 3D scene and 2D image editing, delivering realistic edits with excellent preservation of the original content. It demonstrates strong adaptability to various editing tasks and target prompts, making it a robust solution for complex scenarios. However, while effective, the optimization process is relatively time-intensive compared to recent one-step methods [50] or few-step approaches [7]. Future work could explore integrating advanced techniques such as LCM [30] or SD-Turbo [42], which show potential for accelerating the optimization process [45]." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "text", + "content": "Acknowledgment: This project is supported by the National Natural Science Foundation of China (62125201, U24B20174, 62102381), the Guangdong Natural Science Funds for Distinguished Young Scholars (Grant 2023B1515020097), the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No.: AISG3-GV-2023-011), the Singapore Ministry of Education AcRF Tier 1 Grant (Grant No.: MSS25C004), the Lee Kong Chian Fellowships." + } + ] + } + ], + "index": 47 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16604" + } + ] + } + ], + "index": 48 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 134 + ], + "type": "text", + "content": "[1] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, pages 5470–5479, 2022. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 135, + 296, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 135, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 61, + 135, + 296, + 168 + ], + "type": "text", + "content": "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, pages 18392-18402, 2023. 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 168, + 296, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 168, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 168, + 296, + 213 + ], + "type": "text", + "content": "[3] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In ICCV, pages 22560-22570, 2023. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 213, + 296, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 213, + 296, + 255 + ], + "spans": [ + { + "bbox": [ + 62, + 213, + 296, + 255 + ], + "type": "text", + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 258, + 296, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 296, + 290 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 296, + 290 + ], + "type": "text", + "content": "[5] Minghao Chen, Iro Laina, and Andrea Vedaldi. Dge: Direct gaussian 3d editing by consistent multi-view editing. In ECCV, pages 74-92, 2024. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 291, + 296, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 296, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 296, + 345 + ], + "type": "text", + "content": "[6] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. In CVPR, pages 21476-21485, 2024. 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 346, + 296, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 296, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 296, + 378 + ], + "type": "text", + "content": "[7] Gilad Deutch, Rinon Gal, Daniel Garibi, Or Patashnik, and Daniel Cohen-Or. Turboedit: Text-based image editing using few-step diffusion models. In SIGGRAPHA, 2024. 5, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 380, + 296, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 296, + 411 + ], + "type": "text", + "content": "[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 34:8780-8794, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 413, + 296, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 413, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 413, + 296, + 445 + ], + "type": "text", + "content": "[9] Ayaan Haque, Matthew Tancik, Alexei Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. In ICCV, 2023. 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 447, + 296, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 447, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 57, + 447, + 296, + 479 + ], + "type": "text", + "content": "[10] Amir Hertz, Kfir Aberman, and Daniel Cohen-Or. Delta denoising score. In ICCV, pages 2328-2337, 2023. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 480, + 296, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 480, + 296, + 512 + ], + "spans": [ + { + "bbox": [ + 57, + 480, + 296, + 512 + ], + "type": "text", + "content": "[11] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In ICLR, 2023. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 514, + 294, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 294, + 535 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 294, + 535 + ], + "type": "text", + "content": "[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshop, 2021. 2, 3, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 536, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 296, + 567 + ], + "type": "text", + "content": "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 569, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 569, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 569, + 296, + 612 + ], + "type": "text", + "content": "[14] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 614, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 294, + 646 + ], + "type": "text", + "content": "[15] Yukun Huang, Jianan Wang, Yukai Shi, Boshi Tang, Xianbiao Qi, and Lei Zhang. Dreamtime: An improved optimization strategy for diffusion-guided 3d generation. In ICLR, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 647, + 294, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 680 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 680 + ], + "type": "text", + "content": "[16] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. In CVPR, pages 12469-12478, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 681, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 296, + 712 + ], + "type": "text", + "content": "[17] Aapo Hyvarinen and Peter Dayan. Estimation of non-normalized statistical models by score matching. JMLR, 6(4), 2005. 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "text", + "content": "[18] Yutao Jiang, Yang Zhou, Yuan Liang, Wenxi Liu, Jianbo Jiao, Yuhui Quan, and Shengfeng He. Diffuse3d: Wide-angle 3d photography via bilateral diffusion. In ICCV, pages 8998-9008, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 119, + 554, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 554, + 151 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 554, + 151 + ], + "type": "text", + "content": "[19] Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. Pnp inversion: Boosting diffusion-based editing with 3 lines of code. In ICLR, 2024. 2, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 153, + 555, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 555, + 186 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 555, + 186 + ], + "type": "text", + "content": "[20] Oren Katzir, Or Patashnik, Daniel Cohen-Or, and Dani Lischinski. Noise-free score distillation. In ICLR, 2024. 1, 2, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 188, + 554, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 188, + 554, + 220 + ], + "spans": [ + { + "bbox": [ + 317, + 188, + 554, + 220 + ], + "type": "text", + "content": "[21] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splattering for real-time radiance field rendering. ACM TOG, 42(4):1-14, 2023. 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 223, + 555, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 223, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 223, + 555, + 266 + ], + "type": "text", + "content": "[22] Jiwook Kim, Seonho Lee, Jaeyo Shin, Jiho Choi, and Hyunjung Shim. Dreamcatalyst: Fast and high-quality 3d editing via controlling editability and identity preservation. In ICLR, 2025. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 269, + 555, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 269, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 316, + 269, + 555, + 300 + ], + "type": "text", + "content": "[23] Hubert Kompanowski and Binh-Son Hua. Dream-in-style: Text-to-3d generation using stylized score distillation. In 3DV, 2025. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 304, + 555, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 555, + 335 + ], + "type": "text", + "content": "[24] Juil Koo, Chanho Park, and Minhyuk Sung. Posterior distillation sampling. In CVPR, pages 13352-13361, 2024. 2, 3, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 338, + 554, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 554, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 554, + 381 + ], + "type": "text", + "content": "[25] Duong H Le, Tuan Pham, Aniruddha Kembhavi, Stephan Mandt, Wei-Chiu Ma, and Jiasen Lu. Preserving identity with variational score for general-purpose 3d editing. arXiv preprint arXiv:2406.08953, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 384, + 554, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 384, + 554, + 427 + ], + "spans": [ + { + "bbox": [ + 316, + 384, + 554, + 427 + ], + "type": "text", + "content": "[26] Kehan Li, Yanbo Fan, Yang Wu, Zhongqian Sun, Wei Yang, Xiangyang Ji, Li Yuan, and Jie Chen. Learning pseudo 3d guidance for view-consistent texturing with 2d diffusion. In ECCV, pages 18-34, 2025. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 430, + 555, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 430, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 316, + 430, + 555, + 473 + ], + "type": "text", + "content": "[27] Senmao Li, Joost Van De Weijer, Taihang Hu, Fahad Shahbaz Khan, Qibin Hou, Yaxing Wang, and Jian Yang. Stylediffusion: Prompt-embedding inversion for text-based editing. arXiv preprint arXiv:2303.15649, 2023. 7, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 475, + 554, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 554, + 518 + ], + "type": "text", + "content": "[28] Haofeng Liu, Chenshu Xu, Yifei Yang, Lihua Zeng, and Shengfeng He. Drag your noise: Interactive point-based editing via diffusion semantic propagation. In CVPR, pages 6743-6752, 2024. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 521, + 555, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 521, + 555, + 564 + ], + "spans": [ + { + "bbox": [ + 316, + 521, + 555, + 564 + ], + "type": "text", + "content": "[29] Yuqin Lu, Bailin Deng, Zhixuan Zhong, Tianle Zhang, Yuhui Quan, Hongmin Cai, and Shengfeng He. 3d snapshot: Invertible embedding of 3d neural representations in a single image. 46(12):11524-11531, 2024. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 567, + 555, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 555, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 555, + 609 + ], + "type": "text", + "content": "[30] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 8" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 613, + 555, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 555, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 555, + 655 + ], + "type": "text", + "content": "[31] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In CVPR, pages 12663–12673, 2023. 1" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "type": "text", + "content": "[32] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM TOG, 38(4): 1-14, 2019. 6" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16605" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 116 + ], + "type": "text", + "content": "[33] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, pages 405-421, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 161 + ], + "type": "text", + "content": "[34] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, pages 6038–6047, 2023. 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 205 + ], + "type": "text", + "content": "[35] Hyelin Nam, Gihyun Kwon, Geon Yeong Park, and Jong Chul Ye. Contrastive denoising score for text-guided latent diffusion image editing. In CVPR, pages 9192-9201, 2024. 2, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 207, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 207, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 207, + 295, + 239 + ], + "type": "text", + "content": "[36] JangHo Park, Gihyun Kwon, and Jong Chul Ye. ED-NeRF: Efficient text-guided editing of 3d scene with latent space nef. In ICLR, 2024. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 296, + 274 + ], + "type": "text", + "content": "[37] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 1, 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 275, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 296, + 319 + ], + "type": "text", + "content": "[38] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2): 3, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 320, + 296, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 296, + 364 + ], + "type": "text", + "content": "[39] Jingjing Ren, Wenbo Li, Haoyu Chen, Renjing Pei, Bin Shao, Yong Guo, Long Peng, Fenglong Song, and Lei Zhu. Ultrapixel: Advancing ultra high-resolution image synthesis to new peaks. In NeurIPS, pages 111131-111171, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 365, + 296, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 296, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 296, + 408 + ], + "type": "text", + "content": "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 410, + 296, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 410, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 296, + 474 + ], + "type": "text", + "content": "[41] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, pages 36479-36494, 2022. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 476, + 296, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 476, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 56, + 476, + 296, + 509 + ], + "type": "text", + "content": "[42] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In ECCV, pages 87-103, 2025. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 510, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 295, + 533 + ], + "type": "text", + "content": "[43] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 1, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 534, + 296, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 296, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 296, + 577 + ], + "type": "text", + "content": "[44] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 578, + 295, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 295, + 611 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 295, + 611 + ], + "type": "text", + "content": "[45] Feng Tian, Yixuan Li, Yichao Yan, Shanyan Guan, Yanhao Ge, and Xiaokang Yang. Postedit: Posterior sampling for efficient zero-shot image editing. In ICLR, 2025. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 613, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 296, + 645 + ], + "type": "text", + "content": "[46] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In CVPR, pages 1921-1930, 2023. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 647, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 296, + 689 + ], + "type": "text", + "content": "[47] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In NeurIPS, 2024. 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "text", + "content": "[48] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva:" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 465 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 96, + 555, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 140 + ], + "type": "text", + "content": "[49] Jing Wu, Jia-Wang Bian, Xinghui Li, Guangrun Wang, Ian Reid, Philip Torr, and Victor Adrian Prisacariu. Gaussctrl: Multi-view consistent text-driven 3d gaussian splatting editing. In ECCV, pages 55-71, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 141, + 555, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 555, + 173 + ], + "type": "text", + "content": "[50] Sihan Xu, Yidong Huang, Jiayi Pan, Ziqiao Ma, and Joyce Chai. Inversion-free image editing with natural language. In CVPR, 2024. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 175, + 554, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 554, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 554, + 207 + ], + "type": "text", + "content": "[51] Yangyang Xu, Shengfeng He, Kwan-Yee K Wong, and Ping Luo. Rigid: Recurrent gan inversion and editing of real face videos and beyond. IJCV, 133(6):3437-3455, 2025. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 209, + 555, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 555, + 240 + ], + "type": "text", + "content": "[52] Xin Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Song-Hai Zhang, and Xiaojuan Qi. Text-to-3d with classifier score distillation. In ICLR, 2024. 1, 2, 3, 4, 5, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 243, + 555, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 555, + 285 + ], + "type": "text", + "content": "[53] Yuyang Yu, Bangzhen Liu, Chenxi Zheng, Xuemiao Xu, Huaidong Zhang, and Shengfeng He. Beyond textual constraints: Learning novel diffusion conditions with fewer examples. In CVPR, pages 7109-7118, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 287, + 554, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 554, + 320 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 554, + 320 + ], + "type": "text", + "content": "[54] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836-3847, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 322, + 555, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 322, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 322, + 555, + 354 + ], + "type": "text", + "content": "[55] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 355, + 555, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 555, + 398 + ], + "type": "text", + "content": "[56] Chenxi Zheng, Yihong Lin, Bangzhen Liu, Xuemiao Xu, Yongwei Nie, and Shengfeng He. Recdreamer: Consistent text-to-3d generation via uniform score distillation. In ICLR, 2025. 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 399, + 554, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 554, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 554, + 431 + ], + "type": "text", + "content": "[57] Xingchen Zhou, Ying He, F Richard Yu, Jianqiang Li, and You Li. RePaint-NeRF: Nerf editing via semantic masks and diffusion models. In IJCAI, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 434, + 555, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 555, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 555, + 465 + ], + "type": "text", + "content": "[58] Junzhe Zhu, Peiye Zhuang, and Sanmi Koyejo. HIFA: High-fidelity text-to-3d generation with advanced diffusion guidance. In ICLR, 2024. 1, 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16606" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Stable Virtual Camera_ Generative View Synthesis with Diffusion Models/8df1f571-dcd1-4fc3-a74d-5f6f5bf1e485_content_list.json b/2025/Stable Virtual Camera_ Generative View Synthesis with Diffusion Models/8df1f571-dcd1-4fc3-a74d-5f6f5bf1e485_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c38a07696be77f097bfdfb1b8a496b785e9d2f0d --- /dev/null +++ b/2025/Stable Virtual Camera_ Generative View Synthesis with Diffusion Models/8df1f571-dcd1-4fc3-a74d-5f6f5bf1e485_content_list.json @@ -0,0 +1,1470 @@ +[ + { + "type": "text", + "text": "STABLE VIRTUAL CAMERA: Generative View Synthesis with Diffusion Models", + "text_level": 1, + "bbox": [ + 96, + 130, + 895, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jensen (Jinghao) Zhou $^{1,2,*,\\dagger}$ Hang Gao $^{1,3,*,\\dagger}$ \nVikram Voleti $^{1}$ Aaryaman Vasishta $^{1}$ Chun-Han Yao $^{1}$ Mark Boss $^{1}$ Philip Torr $^{2}$ Christian Rupprecht $^{2}$ Varun Jampani $^{1}$", + "bbox": [ + 210, + 161, + 782, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Stability AI $^{2}$ University of Oxford $^{3}$ University of California, Berkeley", + "bbox": [ + 197, + 224, + 797, + 244 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7972dd8723f41d3ffe2accb551f716bcfa11ff6f527fc96112d36828f2b1c496.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 252, + 617, + 271 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present STABLE VIRTUAL CAMERA (SEVA), a generalist diffusion model that creates novel views of a scene, given any number of input views and target cameras. Existing works struggle to generate either large viewpoint changes or temporally smooth samples, while relying on specific task configurations. Our approach overcomes these limitations through simple model design, optimized training recipe, and flexible sampling strategy that generalize across view synthesis tasks at test time. As a result, our samples maintain high consistency without requiring additional 3D representation-based distillation, thus streamlining view synthesis in the wild. Furthermore, we show that our method can generate high-quality videos lasting up to half a minute with seamless loop closure. Extensive benchmarking demonstrates that SEVA outperforms existing methods across different datasets and settings.", + "bbox": [ + 89, + 319, + 483, + 563 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 584, + 220, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Novel view synthesis (NVS) aims to generate realistic, 3D-consistent images of a scene from arbitrary camera viewpoints given any number of camera-posed input views. Traditional methods, which rely on dense input views, treat NVS as a 3D reconstruction and rendering problem [1-3], but this approach fails with sparse inputs. Generative view synthesis addresses this limitation by leveraging modern deep network priors [4, 5], enabling immersive 3D interactions in uncontrolled environments without the need to capture large image sets per scene. In this work, we focus on generative view synthesis and, unless otherwise specified, refer to it simply as NVS for clarity.", + "bbox": [ + 89, + 611, + 483, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite recent progress [6-12], NVS in the wild remains limited due to two key challenges: First, existing methods struggle to generate both large viewpoint changes [9, 10] and temporally smooth samples [6-8, 13] while being constrained by rigid task configurations, such as a fixed number of input and target views [7, 9, 11, 12], reviewed in Tab. 1. Second, their sampling consistency is often insufficient, ne", + "bbox": [ + 89, + 795, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fb92f49ac881fc46a4d8f22a85a7508e73995c0848ced9210adf4443401260a8.jpg", + "image_caption": [ + "Figure 1. Generative view synthesis. STABLE VIRTUAL CAMERA generates novel views from any number of input views and target cameras, which the user can specify anywhere. We show three examples: single view with simple orbit camera trajectory (top); two views with long camera trajectory (middle); and nine views with large spatial range (bottom). Please refer to the supplementary webpage and video for the video results." + ], + "image_footnote": [], + "bbox": [ + 514, + 284, + 906, + 492 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cessitating additional NeRF distillation to fuse inconsistent results into a coherent representation [7, 8, 13]. These limitations hinder their applicability across diverse NVS tasks, which we address in this work.", + "bbox": [ + 511, + 604, + 905, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present _STABLE_VIRTUAL CAMERA1 (SEVA), a diffusion-based NVS model that generalizes across a spectrum of view synthesis tasks without requiring NeRF distillation. With a single network, SEVA generates high-quality novel views that strike both large viewpoint changes and temporal smoothness while supporting any number of input and target views. Our approach simplifies the NVS pipeline without requiring distillation from a 3D representation, thus streamlining it for real-world applications. For the first time, we demonstrate high-quality videos lasting up to half a minute with precise camera control and seamless loop closure in 3D. We highlight these results in Fig. 1 and", + "bbox": [ + 511, + 666, + 908, + 849 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1We are naming this model in tribute to the Virtual Camera [14] cinematography technology, a pre-visualization technique to simulate real-world camera movements.", + "bbox": [ + 511, + 862, + 906, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12405", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "showcase more examples of camera control in Fig. 2.", + "bbox": [ + 89, + 90, + 442, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve this, we carefully design our pipeline in three key aspects: model design, training recipe, and sampling method at inference. First, SEVA avoids explicit 3D representations within the network, allowing the model to inherit strong priors from pre-trained 2D models. Second, during training, we carefully craft our view selection strategy to cover both small and large viewpoint changes, ensuring strong generalization to diverse NVS tasks. Third, at inference, we introduce a two-pass procedural sampling approach that supports flexible input-target configurations. Together, these design choices create a versatile 3D \"virtual camera simulation system\" capable of synthesizing novel views along arbitrary camera trajectories with any number of input and target views, without using a 3D representation.", + "bbox": [ + 88, + 106, + 480, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conducted a unified benchmark across 10 datasets and a variety of experimental settings, including both open-source and proprietary models. It reflects the diversity of real-world NVS tasks across the board and systematically evaluates existing methods beyond their comfort zones. We find that SEVA consistently outperforms previous works, achieving $+1.5$ dB PSNR over the state of the art CAT3D [8]. Our method generalizes well to in-the-wild user captures, with input views ranging from 1 to 32.", + "bbox": [ + 88, + 316, + 480, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our key contributions with the SEVA model include: (1) a training strategy for jointly modeling large viewpoint changes and temporal smoothness, (2) a two-pass procedural sampling method for smooth video generation along arbitrary long camera trajectories, (3) a comprehensive benchmark that evaluates NVS methods across different datasets and settings, and (4) an open-source release of model weights to support future research.", + "bbox": [ + 88, + 453, + 482, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background", + "text_level": 1, + "bbox": [ + 89, + 587, + 218, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We consider the evaluation of an NVS model across three key criteria: (1) generation capacity—the ability to synthesize missing regions for large viewpoint changes; (2) interpolation smoothness—the ability to produce seamless transitions between views; and (3) input flexibility—the ability to handle a variable number of input and target views. We review existing NVS models based on these criteria in Tab. 1, including the types of training data.", + "bbox": [ + 88, + 612, + 482, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Types of NVS Tasks", + "text_level": 1, + "bbox": [ + 89, + 742, + 279, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given $M$ input view images $\\mathbf{I}^{\\mathrm{inp}}\\in \\mathbb{R}^{M\\times H\\times W\\times 3}$ of $H\\times W$ resolution, along with their corresponding cameras $\\pi^{\\mathrm{inp}}$ , NVS involves predicting $N$ targets views $\\mathbf{I}^{\\mathrm{tgt}}\\in \\mathbb{R}^{N\\times H\\times W\\times 3}$ , specified by their respective cameras $\\pi^{\\mathrm{tgt}}$ . For each camera, we assume we know both intrinsics and extrinsics. Based on the number of input views, we define the \"sparse-view regime\" as having up to 8 input views, and the \"semi-dense-view regime\" as an intermediate state bridging the sparse-view regime and dense captures, which", + "bbox": [ + 89, + 762, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/127edfaca72ede73b9e414e914fa8866af67936bbd80405eb52b0938d618241b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
modeltraining datageneration capacityinterpolation smoothnessinput flexibility
Regression-based
pixelNeRF [4]▲X✓sparse (1)
pixelSplat [15]▲X✓sparse (2)
MVSplat [16]▲X✓sparse (2)
Long-LRM [17]▲X✓semi-dense (\\{16, 32\\})
LVSM [11]/▲X✓sparse (\\{2, 4\\})
Diffusion-based: image models
Zero123 [6]▲✓Xsparse (1)
ZeroNVS [13]▲✓Xsparse (1)
ReconFusion [7]▲✓Xsparse (3)
CAT3D [8]▲✓Xsparse ([1, 9])
Diffusion-based: video models
SV3D [18]▲X✓sparse (1)
MotionCtrl [10]▲X✓sparse (1)
ViewCrafter [9]▲X✓sparse (2)
4DiM [12]▲✓✓sparse (\\{1, 2, 8\\})
SEVA▲✓✓sparse ([1, 8]), semi-dense ([9, 32*])
", + "bbox": [ + 516, + 88, + 903, + 368 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Comparison of existing NVS models based on the source of training data and key attributes. SEVA is trained on both object-level (●) and scene-level (▲) data, offering flexible input conditioning, strong generation capacity, and smooth view interpolation. We define generation capacity and interpolation smoothness of each work based on their evaluation setting and our benchmark results. *This upper-bound can be up to hundreds for dense captures, we test our model up to 32 views in practice.", + "bbox": [ + 511, + 375, + 906, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "typically involve hundreds of views. Based on the nature of their target views, we bucket a broad range of NVs tasks into \"set NVS\" and \"trajectory NVS\", as shown in Fig. 3.", + "bbox": [ + 511, + 497, + 906, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Set NVs considers a set of target views in arbitrary order, usually across a large spatial range. The order of views is often not helpful here, and a good NVS model requires great generation capacity to excel at this task. We note that some works address only this task (e.g. ReconFusion [7]).", + "bbox": [ + 511, + 553, + 905, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Trajectory NVs regards target views along a smooth camera trajectory, such that they form a video sequence. However, they are often sampled within a small spatial range in a shorter video. To solve this task, a good NVS model requires great interpolation smoothness to produce consistent and non-flickering results. We note that some existing works address only this task (e.g. ViewCrafter [9]).", + "bbox": [ + 511, + 638, + 905, + 746 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Existing Models", + "text_level": 1, + "bbox": [ + 511, + 756, + 671, + 772 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We group existing approaches into regression- and diffusion-based models based on their high-level design choices. A more detailed discussion of related works can be found in the appendix.", + "bbox": [ + 511, + 779, + 905, + 840 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Regression-based models learn a deterministic mapping:", + "bbox": [ + 511, + 851, + 903, + 867 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\theta} \\left(\\mathbf {I} ^ {\\text {i n p}}, \\boldsymbol {\\pi} ^ {\\text {i n p}}, \\boldsymbol {\\pi} ^ {\\text {t g t}}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 878, + 774, + 904 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12406", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d5b834cf8c302efb60c3cc213a069ca1a0f3c04730ee044608b07528d09ade78.jpg", + "image_caption": [ + "Figure 2. Diverse camera control. SEVA generates photorealistic novel views following diverse camera trajectories. This includes orbit, spiral, zoom out, dolly zooms, and any user-specified trajectories. Please visit our website for more visual results." + ], + "image_footnote": [], + "bbox": [ + 94, + 99, + 903, + 506 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to directly generate $\\mathbf{I}^{\\mathrm{tgt}}$ deterministically from $\\mathbf{I}^{\\mathrm{inp}}, \\boldsymbol{\\pi}^{\\mathrm{inp}}, \\boldsymbol{\\pi}^{\\mathrm{tgt}}$ . $f_{\\theta}$ can be either an end-to-end network parameterized by $\\theta$ , or a composition of a feed-forward prediction of an intermediate 3D representation and then a neural renderer (e.g., NeRF [19] or 3DGS [3]). For the latter case, set NVs and trajectory NVs are solved in the same way since there exists a persistent 3D representation.", + "bbox": [ + 89, + 547, + 482, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion-based models capture the conditional distribution:", + "bbox": [ + 89, + 661, + 482, + 689 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(\\mathbf {I} ^ {\\mathrm {t g t}} \\mid \\mathbf {I} ^ {\\mathrm {i n p}}, \\boldsymbol {\\pi} ^ {\\mathrm {i n p}}, \\boldsymbol {\\pi} ^ {\\mathrm {t g t}}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 686, + 370, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "from which $\\mathbf{I}^{\\mathrm{tgt}}$ are sampled [20] iteratively. We highlight two types of models within this scope: Image and Video models. Image models are trained on unordered image sets, such that $(\\mathbf{I}^{\\mathrm{inp}},\\mathbf{I}^{\\mathrm{tgt}})\\sim \\mathcal{I}$ , where $\\mathcal{I} = \\{\\mathbf{I}_{\\rho (1)},\\mathbf{I}_{\\rho (2)},\\dots ,\\mathbf{I}_{\\rho (M + N)}\\}$ is an image batch, and $\\rho (\\cdot)$ is a random permutation function (camera parameters are omitted for simplicity). Image models usually thrive at set NVs but struggle at trajectory NVS since they are designed to generate images and not videos. Moreover, the unordered nature of all views solicits flexible input conditioning. Video models are instead trained on ordered views, such that $(\\mathbf{I}^{\\mathrm{inp}},\\mathbf{I}^{\\mathrm{tgt}})\\sim \\mathcal{V}$ , where $\\mathcal{V} = \\{\\mathbf{I}_1,\\mathbf{I}_2,\\dots ,\\mathbf{I}_{M + N}\\}$", + "bbox": [ + 89, + 719, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4518ec28f6c75a7e52b1e70835d8bce7d8bba039f0fbb7a69b5b8a76f6374b69.jpg", + "image_caption": [ + "SetNVS" + ], + "image_footnote": [], + "bbox": [ + 517, + 545, + 687, + 614 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d3b77916ee65256010f7877cfd0d879c34e6fca09a92b63d881f6cc0b5880459.jpg", + "image_caption": [ + "Trajectory NVS", + "Figure 3. Set NVS versus trajectory NVS. Set NVS generates target views as an image set, whereas trajectory NVS produces them as a trajectory video." + ], + "image_footnote": [], + "bbox": [ + 692, + 545, + 859, + 614 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ce0250ea6326add0094e34b1c625f739992a3a56a17efd99669ca3f462d11661.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 861, + 545, + 903, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "is a set of video frames with preserved ordering. Additional temporal operators may also be used to improve temporal smoothness, such as temporal positional encoding and temporal attention. In contrast with image models, video models thrive at trajectory NVS but struggle at set NVS. Moreover, all existing video models require both input and target views to be ordered (input views followed by target ones), constraining their input flexibility [10, 18, 21, 22].", + "bbox": [ + 511, + 680, + 906, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 815, + 604, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It is challenging to develop a unified model that works for different NVS tasks while being flexible in input and output domains. Our design choices enable SEVA to achieve high generation capacity, smooth view interpolation, and flexi", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12407", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1bf2f170e5012697b960197a8454e34e52415c909a54c43e646dc0e6d9483f57.jpg", + "image_caption": [ + "Figure 4. Method. SEVA is trained with fixed sequence length as a “ $M$ -in $N$ -out” multi-view diffusion model with standard architecture. During sampling, SEVA can be cast as a generative “ $P$ -in $Q$ -out” renderer that works with variable sequence length, where $P$ and $Q$ need not be equal to $M$ and $N$ . To enhance temporal and 3D consistency across generated views, especially when generating along a trajectory, we present procedural two-pass sampling as a general strategy." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 370, + 244 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/801909587f252e64b3dd1adecbc53ed19c318cae5441d2991497717a92097bac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 87, + 906, + 246 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ble input conditioning (Tab. 1) in both set NVs as well as trajectory NVs. A system overview is provided in Fig. 4.", + "bbox": [ + 89, + 311, + 482, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Model Design and Training", + "text_level": 1, + "bbox": [ + 89, + 349, + 338, + 366 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We consider an \" $M$ -in $N$ -out\" multi-view diffusion model $p_{\\theta}$ , as noted in Sec. 2.2. We formulate this learning problem as a standard diffusion process [20] without any change.", + "bbox": [ + 89, + 372, + 483, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Architecture. Our model is based on the publicly available SD 2.1 [23], which consists of an auto-encoder and a latent denoising U-Net. Following [8], we inflate the 2D self-attention of each low-resolution residual block into 3D self-attention [24] within the U-Net. To improve model capacity, we add 1D self-attention along the view axis after each self-attention block via skip connection [25, 26], bumping the model parameters from 870M to 1.3B. Optionally, we further tame this model into a video model by introducing 3D convolutions after each residual block via skip connection, similar to [27, 28], yielding 1.5B total parameters. The temporal pathway can be enabled during inference when frames within one forward pass are known to be spatially ordered, enhancing output's smoothness. For the conditioning of our model, please refer to the appendix.", + "bbox": [ + 89, + 425, + 483, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training. Let us define the training context window length $T = |\\mathbf{I}^{\\mathrm{inp}}| + |\\mathbf{I}^{\\mathrm{tgt}}| = M + N$ . One natural goal is to support large $T$ such that we can generate larger set of frames. However, we find that naive training is prone to divergence, and we thus employ a two-stage training curriculum. During the first stage, we train our model with $T = 8$ with a batch size of 1472 for $100K$ iterations. In the second stage, we train our model with $T = 21$ with a batch size of 512 for $600K$ iterations. Given a training video sequence, we randomly sample the number of input frames $M \\in [1,T - 1]$ and the frames $(\\mathbf{I}^{\\mathrm{inp}},\\mathbf{I}^{\\mathrm{gt}})$ across the entire sequence to encourage diversity. We find it important to jointly sample with a smaller subsampling stride to ensure sufficient temporal granularity and avoid missing critical transitions with a small probability (0.2 is used in practice). In the optional video training stage, we only train", + "bbox": [ + 89, + 659, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "temporal weights with data sampled with a small subsampling stride and a batch size of 512 for $200K$ iterations. We shift the signal-to-noise ratio (SNR) in all stages as more noise is necessary to remove the input information, corroborating findings from [8, 29, 30]. The model is trained with squared images with $H = W = 576$ . The training data is a superset of data used in [8] and derives from all public academic datasets, consisting of around $340K$ objects and $80K$ scenes.", + "bbox": [ + 511, + 311, + 906, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Sampling Novel Views", + "text_level": 1, + "bbox": [ + 511, + 459, + 720, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Once the diffusion model is trained, we can sample it for a wide range of NVS tasks during test time. Formally, let us consider a \" $P$ -in $Q$ -out\" NVS task during testing, where we are given $P = |\\mathbf{I}^{\\mathrm{inp}}|$ input frames and aim to produce $Q = |\\mathbf{I}^{\\mathrm{tgt}}|$ target frames. Our goal is to design a generic sampling strategy that works for all $P$ and $Q$ configurations, where $P$ and $Q$ need not be equal to $M$ and $N$ .", + "bbox": [ + 511, + 482, + 905, + 588 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We make two key observations: First, within a single forward pass, predictions are 3D consistent, provided the model is well-trained. Second, when $P + Q > T$ , $\\mathbf{I}^{\\mathrm{gt}}$ must be split into smaller chunks of size $Q_{i}$ such that $P + Q_{i} \\leqslant T$ for the $i^{\\mathrm{th}}$ forward pass. We term this practice one-pass sampling. However, predictions across these forward passes would be inconsistent unless they share common frames to maintain local consistency within a spatial neighborhood. Building on these observations, we summarize our sampling process under two scenarios: $P + Q \\leqslant T$ and $P + Q > T$ .", + "bbox": [ + 511, + 589, + 905, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$P + Q \\leqslant T$ . We fit the task within one forward pass for simplicity and consistency. As shown in the appendix, we find it works better to pad the forward pass to have exactly $T$ frames by repeating the first input image, compared to changing the context window $T$ zero-shot.", + "bbox": [ + 511, + 766, + 905, + 843 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$P + Q > T$ . We propose procedural two-pass sampling: In the first pass, we generate anchor frames $\\mathbf{I}^{\\mathrm{acr}}$ using all input frames $\\mathbf{I}^{\\mathrm{inp}}$ . In the second pass, we divide $\\mathbf{I}^{\\mathrm{tgt}}$ into", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12408", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "chunks and generate them using $\\mathbf{I}^{\\mathrm{acr}}$ (and optionally $\\mathbf{I}^{\\mathrm{inp}}$ ) according to the spatial distribution of $\\mathbf{I}^{\\mathrm{acr}}$ and $\\mathbf{I}^{\\mathrm{gt}}$ .", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the distinct nature of the two tasks of interest—set NVS and trajectory NVS—e.g., differences in the availability of views' ordering, we design tailored chunking strategies for each task.", + "bbox": [ + 89, + 121, + 482, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For set NVS, we consider nearest procedural sampling. We first generate $\\mathbf{I}^{\\mathrm{acr}}$ based on pre-defined trajectory priors, similar to [8], e.g., 360 trajectories for object-centric scenes or spiral trajectories for forward-facing scenes. We then divide $\\mathbf{I}^{\\mathrm{gt}}$ into chunks w.r.t. $\\mathbf{I}^{\\mathrm{acr}}$ using nearest neighbor. Specifically, the $i^{\\mathrm{th}}$ forward pass involves:", + "bbox": [ + 89, + 181, + 483, + 272 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {n e a r e s t}: \\left\\{\\mathbf {I} _ {i} ^ {\\mathrm {a c r}} \\right\\} \\cup \\left\\{\\mathbf {I} _ {j} ^ {\\mathrm {t g t}} \\mid \\mathrm {N N} (\\mathbf {I} _ {j} ^ {\\mathrm {t g t}}, \\mathbf {I} ^ {\\mathrm {a c r}}) = \\mathbf {I} _ {i} ^ {\\mathrm {a c r}} \\right\\}.\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 282, + 442, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We considered two strategies of procedural sampling: nearest as described above, and $\\mathrm{gt} +$ nearest strategy by appending $\\mathbf{I}^{\\mathrm{inp}}$ into each forward pass. We find that the $\\mathrm{gt} +$ nearest strategy performs better than nearest, and thus default to it instead. In the absence of trajectory priors, we revert to one-pass sampling. In practice, employing nearest anchors enhances qualitative consistency, albeit on a limited scale.", + "bbox": [ + 89, + 310, + 483, + 430 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For trajectory NVS, we consider interp procedural sampling. We first generate a subset of target frames as $\\mathbf{I}^{\\mathrm{acr}}$ by uniformly spanning the target camera path with a stride $\\Delta = \\left\\lfloor \\frac{Q}{T - 2}\\right\\rfloor$ . We then generate the rest of $\\mathbf{I}^{\\mathrm{tgt}}$ as segments between those anchors:", + "bbox": [ + 89, + 431, + 483, + 506 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {i n t e r p}: \\{\\mathbf {I} _ {i} ^ {\\mathrm {a c r}}, \\mathbf {I} _ {i \\cdot \\Delta + 1} ^ {\\mathrm {t g t}}, \\dots , \\mathbf {I} _ {(i + 1) \\cdot \\Delta - 1} ^ {\\mathrm {t g t}}, \\mathbf {I} _ {i + 1} ^ {\\mathrm {a c r}} \\}.\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 516, + 437, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the input to the model is ordered, we can leverage temporal weights to further improve smoothness (Sec. 4.3). Similarly, $\\mathrm{gt} + \\mathrm{interp}$ is possible by appending $\\mathbf{I}^{\\mathrm{inp}}$ with $\\Delta = \\left\\lfloor \\frac{Q}{T - P - 2} \\right\\rfloor$ . We find that interp is sufficiently robust, and choose it as the default option. The interp strategy drastically outperforms its counterparts (e.g., one-pass, or $\\mathrm{gt} + \\mathrm{nearest}$ procedural sampling) in terms of temporal smoothness.", + "bbox": [ + 89, + 545, + 483, + 665 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Scaling Sampling for Large $P$ and $Q$", + "text_level": 1, + "bbox": [ + 89, + 674, + 413, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we examine two special cases when $P + Q > T$ : $P > T$ and $Q \\gg T$ . Here, we make a tailored design for anchor generation in the first pass, while keeping target generation in the second pass unchanged.", + "bbox": [ + 89, + 696, + 483, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$P > T$ . In the semi-dense-view regime (e.g., $P = 32$ ), we extend the context window length $T$ zero-shot to accommodate all $P$ input views and anchor views in one pass during anchor generation. Empirically, $T$ can even be extended up to hundreds without severe degradation in photorealism in the generated outputs. We find that the diffusion model generalizes well in this case as long as the input views cover the majority of the scene, shifting the task from generation to primarily interpolation. In the sparse-view", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a43cdb8665181349f6de4fcef3de81638119f70716dba8864ecbf09cb5e37967.jpg", + "image_caption": [ + "Figure 5. Anchor generation when $Q \\gg T$ . We introduce a memory bank composed of previously-generated anchor views and their corresponding camera poses. The lookup of spatial neighbors helps improve long-term 3D consistency." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 903, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "regime $(P \\leq 8)$ , we observe similar performance degradation caused by zero-shot extension of $T$ compared to what we have found when $P + Q \\leqslant T$ . Refer to the appendix for a detailed discussion.", + "bbox": [ + 511, + 289, + 906, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$Q \\gg T$ . When the number of target views $Q$ is large, e.g., in large-set NVS or long-trajectory NVS, even anchors will be chunked into different forwards in the first pass, leading to the inconsistency of anchors. To this end, we maintain a memory bank of anchor views previously generated, as shown in Fig. 5. We generate anchors autoregressively by retrieving their spatially nearest ones from the memory bank, similar to the nearest strategy introduced above for the second pass. In Sec. 4.4, we show that this strategy drastically outperforms the standard practice of reusing temporally nearest anchors previously generated in long video literature [22], in terms of long-range 3D consistency, especially for hard trajectories.", + "bbox": [ + 511, + 373, + 908, + 570 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 582, + 645, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We employ a single model for a spectrum of settings and find that SEVA model generalizes well under the three criteria (Tab. 1). We cover different NVS tasks (set NVS and trajectory NVS) and examine one special task of interest—long trajectory NVS. We also cover different input regimes (single-view, sparse-view, and semi-dense-view). A discussion about key properties is presented in the appendix.", + "bbox": [ + 511, + 608, + 906, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Benchmark", + "text_level": 1, + "bbox": [ + 511, + 726, + 640, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets, splits, and the number of input views. We collect 10 commonly used datasets to benchmark NVS, encompassing a diverse range of scene distributions and complexities. We consider (1) object datasets [32, 33]; (2) object-centric scene datasets [34-38]; and (3) scene datasets [39-41]. To establish a comprehensive and rigorous comparison with baselines, we consider different dataset splits utilized in prior works with the same input-view configuration, unless specified as our split (O). These include splits used in 4DiM [12] (D), ViewCrafter [9] (V),", + "bbox": [ + 511, + 750, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12409", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a4b8491584ad3d6abe80aa4aa7086eb632e6d09a48470af878ef5e46c8e001c0.jpg", + "image_caption": [ + "Figure 6. SOTA comparison on set NVs (top) and trajectory NVs (bottom) across varying numbers of input views. From left to right and top to bottom, inputs views are (top, i.e. set NVS) 1, 2, 3, 32; (bottom, i.e. trajectory NVS) 1, 2, 1, 1, 3, 3. For cases with multiple input views, the view closest to the target is shown over the rest. We compare with open-source approaches—ViewCrafter [9] (VC) and DepthSplat [31] (DS)—as well as proprietary ones including LVSM [11], Long-LRM [17] (LLRM), 4DiM [12], and CAT3D [8]." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 906, + 425 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "pixelSplat [15] (P), ReconFusion [7] (R), SV3D [18] (S), and Long-LRM [17] (L).", + "bbox": [ + 89, + 491, + 482, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Small-viewpoint versus large-viewpoint NVS. Sweeping across all datasets, splits, and input-view configurations reveals a diverse benchmark of setups. To better evaluate models' generation capacity and interpolation smoothness (Sec. 2.1), we propose to categorize this setup into two groups—small-viewpoint NVS and large-viewpoint NVS—depending on the disparity between $\\mathbf{I}^{\\mathrm{tgt}}$ and $\\mathbf{I}^{\\mathrm{inp}}$ .", + "bbox": [ + 89, + 542, + 482, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We consider a range of proprietary models, including ReconFusion [7], CAT3D [8], 4DiM [12], LVSM [11], and Long-LRM [17]. We also consider various open-source models, including SV3D [18], MVSplat [16], depthSplat [31], MotionCtrl [42], and ViewCrafter [9]. As outlined in Sec. 2.2, these baselines encompass both regression-based and diffusion-based approaches, providing a comprehensive framework for comparison. Please refer to the appendix for more details of the benchmark.", + "bbox": [ + 89, + 670, + 482, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Set NVs", + "text_level": 1, + "bbox": [ + 89, + 816, + 192, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative comparison. The input and target views are chosen following the splits used in previous methods. The order of target views is not preserved. We use standard metrics of peak signal-to-noise ratio (PSNR), structural simi", + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "larity index measure [44] (SSIM), and learned perceptual image patch similarity [45] (LPIPS). Only PSNR is showcased here due to space limits, with the rest deferred to the appendix. Empirically, our method shows a greater performance improvement on LPIPS, reflecting the photorealism of our results.", + "bbox": [ + 511, + 491, + 906, + 582 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For small-viewpoint set NVs, Tab. 2 shows that SEVA sets state-of-the-art results in the majority of splits. In the sparse-view regime (i.e., $P \\leqslant 8$ ), SEVA excels across different datasets when $P > 1$ . For example, a performance gain of +6.0 dB PSNR is achieved on LLFF with $P = 3$ . In the semi-dense-view regime (e.g., $P = 32$ ), SEVA surprisingly performs favorably against the specialized model [17], despite not being specifically designed for this setup. For example, SEVA lags behind the state-of-the-art method [17] by only 1.7 dB on T&T. A detailed discussion on the RE10K dataset can be found in the appendix. On object datasets OO3D and GSO, SEVA achieves a significantly higher state-of-the-art PSNR compared to all other methods.", + "bbox": [ + 511, + 583, + 906, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For large-viewpoint set NVs, Tab. 3 shows that SEVA's quantitative advantages are even more prominent here, revealing clear benefits of SEVA in terms of generation capacity when the camera spans a large spatial range. On Mip360 with $P = 3$ , SEVA improves over previous state-of-the-art method CAT3D [8] by $+0.6$ dB PSNR. On harder scenes like DL3DV and T&T with different input-view configurations, SEVA obtains a clear performance lead.", + "bbox": [ + 511, + 779, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12410", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/48ef1fdff6384acc2831050f6090b14d45fbf5a3b56359e23f650fa4eb8d9db0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methoddataset split POO3D OGSO ORE10KLLFF R [7] 1 3DTU R [7] 1 3CO3D V [9] R [7] 1 3WRGBD Oe 3 6Mip360 O L [17] 6 32DL3DV T & T V [9] L [17] 1 32
O 3O 3D [12] 1V [9] 1P [15] 2R [7] 13
Regression-based models
Long-LRM [17]-------------23.86 - 18.20
MVSplat [16]14.78 15.6715.21 16.5220.42 20.9020.32 19.2426.39 27.4421.56 21.8725.64 22.5411.23 12.0712.50 12.6213.87 15.52 14.1513.52 13.52 13.2314.56 12.54 15.9313.56 14.24 14.2313.24 13.22 14.35
DepthSplat [31]
LVSM [11]----29.67---------
Diffusion-based models
MotionCtrl [42]--12.74 17.0816.29 ------15.46 ----13.29 -
4DiM [12]--17.08-----------
ViewCrafter [9]14.64 30.3015.93 31.5320.43 17.9922.04 18.5621.42 25.6620.88 18.1122.81 27.5710.53 14.0313.52 14.4712.66 16.40 18.4718.96 14.72 18.4016.42 12.66 19.7514.59 13.78 16.70- 18.07 - 15.16
SEVA
", + "bbox": [ + 101, + 88, + 893, + 277 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fd88ef218fbbf8ca821867ff3dec15305457e250dea210e8fc17864eeb982b71.jpg", + "table_caption": [ + "Table 2. PSNR† on small-viewpoint set NVS. $P$ denotes the number of input views. For all results with $P = 1$ , we sweep the unit length for camera normalization due to the model's scale ambiguity. $\\mathrm{O}_e$ and $\\mathrm{O}_h$ denote the easy and hard split of our split, respectively. Underlined numbers are run by us using the official released code." + ], + "table_footnote": [], + "table_body": "
Methoddataset split POO3D S [18] 1GSO S [18] 1CO3D R [7] 1WRGBD Oh 1Mip360 R [7] 3DL3DV O 1T&T O 31369
SV3D [18]19.2820.38----------
DepthSplat [31]11.5612.3210.429.3513.5310.4912.549.6312.528.639.7810.12
CAT3D [8]------15.15-----
ViewCrafter [9]10.5611.4210.119.1213.459.7910.348.9711.509.239.8810.32
SEVA19.2520.6515.3014.3717.2812.9315.7813.0115.9511.2812.6513.80
", + "bbox": [ + 94, + 330, + 602, + 455 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/508a3fc5dbfb74647a83806a2d668e2fd53b711dbb112cf0fc88f955522aefb8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methodsmall-viewpointlarge-viewpoint
RE10KLLFFDTUCO3DMip360
ZipNeRF [43]20.7717.239.1814.3412.77
ZeroNVS [13]19.1115.9116.7117.1314.44
ReconFusion [7]25.8421.3420.7419.5915.50
CAT3D [8]26.7821.5822.0220.5716.62
SEVA27.9521.8822.6821.8817.82
", + "bbox": [ + 607, + 333, + 901, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. PSNR† on large-viewpoint set NVs. For all results with $P = 1$ , we sweep the Table 4. PSNR† on 3DGS renderings for set unit length for camera normalization due to the model's scale ambiguity. Underlined NVS. Results are reported on the ReconFusion [7] numbers are run by us using the officially released code. split with $P = 3$ .", + "bbox": [ + 89, + 465, + 903, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison of 3D reconstruction. To enable a direct quantitative comparison with prior works [7, 8], we adopt the few-view 3D reconstruction pipeline described in [8]. We optimize the camera parameters and apply LPIPS loss [46] during the distillation. Finally, we render the distilled 3D model on the test views and report the performance in Tab. 4. SEVA shows a consistent performance lead.", + "bbox": [ + 88, + 516, + 482, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative comparison. Fig. 6 top panel shows a qualitative comparison with diverse baselines. For small-viewpoint set NVs, the output from SEVA with the best scale exhibits desirable alignment with the ground truth while being more photorealistic in details. Compared with LVSM [11] on the P split of RE10K, SEVA produces sharper images, also corroborating that lower PSNR arises from scale ambiguity rather than interpolation quality. Similar trends hold when compared to Long-LRM [17] on DL3DV with $P = 32$ . For large-viewpoint set NVs, we compare with DepthSplat [31] on DL3DV with $P = 3$ . DepthSplat fails to produce reasonable results when the viewpoint change is too large and falls short in overall visual quality.", + "bbox": [ + 88, + 628, + 482, + 825 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Trajectory NVS", + "text_level": 1, + "bbox": [ + 89, + 833, + 250, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative results. Fig. 2 presents qualitative results, illustrating trajectories of varying complexities with different numbers of input views across diverse types.", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the single-view regime (i.e., $P = 1$ ), we manually craft a set of common camera movements/effects, e.g., look-at 360, spiral, panning, zoom-in, zoom-out, dolly zoom, etc. We observe that SEVA generalizes to a wide range of images and demonstrates accurate camera-following capacity. An extensive sweeping of camera movements on 4 types of images is provided in the appendix.", + "bbox": [ + 511, + 516, + 906, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the sparse-view regime with few input views (i.e., $1 < P \\leqslant 8$ ), we observe that SEVA demonstrates strong generalization to in-the-wild real-world images and versatility in adapting to different numbers of input views. The output forms a smooth trajectory video with subtle temporal flickering, revealing its capacity to interpolate between views smoothly. In the last row of Fig. 2, our model generates plausible results at the end of the trajectory—an area unseen in the input observations—demonstrating its strong generation capacity. In the semi-dense-view regime (i.e., $P > 9$ ), we similarly find that SEVA is surprisingly able to produce a smooth trajectory video with minimal artifacts. Please check the supplementary website for video results.", + "bbox": [ + 511, + 623, + 908, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative comparison. Fig. 6 bottom panel presents a qualitative comparison with diverse baselines. In the single-view regime (i.e., $P = 1$ ), we compare to 4DiM [12] and CAT3D [8]. We observe more photo-realistic and sharper", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12411", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e132f50a1e0a1699e7008268e344305c60b18995ea4027079a8baf39bc357971.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methodsplit datasetsmall-viewpointlarge-viewpoint
V [9]O
RECO3DT&TREDTUWRDLT&T
MotionCtrl [42]16.2915.4613.29------
DepthSplat [31]19.2413.2314.3525.2314.6812.4511.329.11
ViewCrafter [9]22.0418.9618.0726.5418.9913.4411.459.68
SEVA18.5618.4015.1627.3419.9917.7915.7611.92
SEVA (+ temp.)18.6218.4315.1327.3620.1917.9315.7811.99
", + "bbox": [ + 96, + 89, + 477, + 218 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. PSNR $\\uparrow$ on trajectory NVS. temp. denotes optional temporal pathway. RE, WR, and DL denotes RE10K, WRGBD, and DL3DV, respectively. For the V [9] split, $P = 1$ with unit length swept; for the O split, $P = 3$ . Underlined numbers are run by us using the officially released code.", + "bbox": [ + 89, + 229, + 480, + 299 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7429eaa8fd38800a41e78b4526e383df46bbff51fb6a7508b0cdc85b3a1575dd.jpg", + "image_caption": [ + "Figure 7. Temporal quality. Vertical slices of a rendered novel camera path on the BONSAI scene from Mip-NeRF360 [38] illustrate the temporal quality across adjacent viewpoints. One-pass or gt + nearest procedural sampling results in notable flickering, whereas interp procedural sampling ensures temporally smooth rendering." + ], + "image_footnote": [], + "bbox": [ + 93, + 306, + 480, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "output from our model, especially in the background area for object-centric scenes. 4DiM outputs tend to be cartoonish and over-simplistic, given that the model is only trained on RE10K. In the sparse-view regime with few input views (i.e., $P = 3$ ), we compare with CAT3D and observe that our model demonstrates more photo-realistic textures, especially in the background.", + "bbox": [ + 89, + 592, + 483, + 699 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quantitative comparison. We use the same input views as in set NVS for each split. We use all frames from each scene as target views such that they form a smoothly transitioning trajectory video, i.e., $\\mathbf{I}^{\\mathrm{gt}}\\sim \\mathcal{V}$ . For small-viewpoint trajectory NVS, Tab. 5 compares SEVA with baselines on PSNR. SEVA performs favorably against other methods in V split with $P = 1$ . The performance lead of ViewCrafter is mainly attributed to its training on high-resolution images. For large-viewpoint trajectory NVS with $P = 3$ , SEVA consistency sets new state-of-the-art results. Applying the temporal pathway further boosts performance and improves smoothness, indicating the benefits of the gated", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1116035c33588665ed1055470f28fc4ffbe5e890116e24c52949be06a628600d.jpg", + "image_caption": [ + "Figure 8. Long-range 3D consistency. We visualize samples following a camera path looping three times around the TELEPHONE-BOOTH scene. Lookup using spatial neighbors from the memory bank (ours) notably improves view consistency and reduces artifacts in recurring locations across different loops, compared to lookup using temporal neighbors (baseline)." + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 906, + 290 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "architecture.", + "bbox": [ + 513, + 390, + 599, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on two-pass procedural sampling. Fig. 7 shows that one-pass sampling introduces visible temporal flickering and abrupt scene changes. In contrast, interp procedural sampling produces the smoothest transitions, outperforming the alternative $\\mathrm{gt} +$ nearest strategy and mitigating noticeable flickering. Additional quantitative ablation is deferred to the appendix, including 3DGS distillation and video quality metrics.", + "bbox": [ + 511, + 422, + 906, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Long-Trajectory NVS", + "text_level": 1, + "bbox": [ + 511, + 551, + 718, + 569 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 8 presents a qualitative demonstration of NVS over a long trajectory of up to 1000 frames. With the memory bank maintaining previously generated anchors, SEVA achieves robust 3D consistency for long-trajectory NVS, e.g., the building in front of and the plantation after the booth. Comparing it to using temporal nearest anchors previously generated, using spatially nearest ones demonstrates a clear advantage in terms of long-term consistency.", + "bbox": [ + 511, + 575, + 905, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 709, + 633, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce SEVA, a generalist diffusion model for novel view synthesis that balances large viewpoint changes and smooth interpolation while supporting flexible input and target configurations. By designing a diffusion-based architecture without 3D representation, a structured training strategy, and a two-pass procedural sampling approach, SEVA achieves 3D consistent rendering across diverse NVs tasks. Extensive benchmarking demonstrates its superiority over existing methods, with strong generalization to real-world scenes. For broader impact and limitations, please refer to the appendix.", + "bbox": [ + 511, + 734, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12412", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, pages 405-421, 2020. 1", + "[2] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 2022.", + "[3] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4), July 2023. 1, 3", + "[4] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In CVPR, 2021. 1, 2", + "[5] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3D using 2D diffusion. arXiv, 2022. 1", + "[6] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3D object. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 2", + "[7] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21551-21561, 2024. 1, 2, 6, 7", + "[8] Ruiqi Gao*, Aleksander Holynski*, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul P. Srinivasan, Jonathan T. Barron, and Ben Poole*. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv, 2024. 1, 2, 4, 5, 6, 7", + "[9] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. ViewCrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 1, 2, 5, 6, 7, 8", + "[10] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1, 2, 3", + "[11] Haian Jin, Hanwen Jiang, Hao Tan, Kai Zhang, Sai Bi, Tianyuan Zhang, Fujun Luan, Noah Snavely, and Zexiang Xu. Lvsm: A large view synthesis model with minimal 3d inductive bias, 2024. 1, 2, 6, 7", + "[12] Daniel Watson, Saurabh Saxena, Lala Li, Andrea Tagliasacchi, and David J Fleet. Controlling space and time with diffusion models. arXiv preprint arXiv:2407.07860, 2024. 1, 2, 5, 6, 7", + "[13] Kyle Sargent, Zizhang Li, Tanmay Shah, Charles Herrmann, Hong-Xing Yu, Yunzhi Zhang, Eric Ryan Chan, Dmitry Lagun, Li Fei-Fei, Deqing Sun, et al. Zeronvs: Zero-shot" + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "360-degree view synthesis from a single real image. arXiv preprint arXiv:2310.17994, 2023. 1, 2, 7", + "[14] Virtual camera system. https://en.wikipedia.org/wiki/Virtual_camera_system. Accessed: 5-Mar-2025. 1", + "[15] David Charatan, Sizhe Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In arXiv, 2023. 2, 6, 7", + "[16] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In European Conference on Computer Vision, pages 370-386. Springer, 2025. 2, 6, 7", + "[17] Chen Ziwen, Hao Tan, Kai Zhang, Sai Bi, Fujun Luan, Yicong Hong, Li Fuxin, and Zexiang Xu. Long-lrm: Long-sequence large reconstruction model for wide-coverage gaussian splats. arXiv preprint arXiv:2410.12781, 2024. 2, 6, 7", + "[18] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitrii Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. SV3D: Novel multi-view synthesis and 3D generation from a single image using latent video diffusion. In European Conference on Computer Vision, 2024. 2, 3, 6, 7", + "[19] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 3", + "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3, 4", + "[21] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrl: Enabling camera control for text-to-video generation. arXiv preprint arXiv:2404.02101, 2024. 3", + "[22] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. arXiv, 2022. 3, 5", + "[23] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, June 2022. 4", + "[24] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. MVDream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 4", + "[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4", + "[26] P Goyal. Accurate, large minibatch sg d: training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 4", + "[27] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Content" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12413", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 4", + "[28] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 4", + "[29] Emiel Hoogeboom, Jonathan Heek, and Tim Salimans. simple diffusion: End-to-end diffusion for high resolution images. arXiv preprint arXiv:2301.11093, 2023. 4", + "[30] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICLR, 2024. 4", + "[31] Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. Depthsplat: Connecting gaussian splatting and depth. arXiv preprint arXiv:2410.13862, 2024. 6, 7, 8", + "[32] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, et al. Omniobject3D: Large-vocabulary 3D object dataset for realistic perception, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 803-814, 2023. 5", + "[33] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google Scanned Objects: A high-quality dataset of 3D scanned household items. In 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560. IEEE, 2022. 5", + "[34] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (ToG), 38(4):1-14, 2019. 5", + "[35] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 406-413, 2014.", + "[36] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In International Conference on Computer Vision, 2021.", + "[37] Hongchi Xia, Yang Fu, Sifei Liu, and Xiaolong Wang. RGBd objects in the wild: Scaling real-world 3d object learning from rgb-d videos, 2024.", + "[38] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-NeRF: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 5, 8", + "[39] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018. 5" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024.", + "[41] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 5", + "[42] Tsai-Shien Chen, Chieh Hubert Lin, Hung-Yu Tseng, Tsung-Yi Lin, and Ming-Hsuan Yang. Motion-conditioned diffusion model for controllable video synthesis. arXiv preprint arXiv:2304.14404, 2023. 6, 7, 8", + "[43] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 7", + "[44] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6", + "[45] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6", + "[46] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric, 2018. 7" + ], + "bbox": [ + 516, + 92, + 903, + 529 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12414", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Stable Virtual Camera_ Generative View Synthesis with Diffusion Models/8df1f571-dcd1-4fc3-a74d-5f6f5bf1e485_model.json b/2025/Stable Virtual Camera_ Generative View Synthesis with Diffusion Models/8df1f571-dcd1-4fc3-a74d-5f6f5bf1e485_model.json new file mode 100644 index 0000000000000000000000000000000000000000..406d6fe40168a3932ba7c5f7f3bef2f784fd120a --- /dev/null +++ b/2025/Stable Virtual Camera_ Generative View Synthesis with Diffusion Models/8df1f571-dcd1-4fc3-a74d-5f6f5bf1e485_model.json @@ -0,0 +1,2002 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.097, + 0.131, + 0.896, + 0.152 + ], + "angle": 0, + "content": "STABLE VIRTUAL CAMERA: Generative View Synthesis with Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.162, + 0.784, + 0.22 + ], + "angle": 0, + "content": "Jensen (Jinghao) Zhou\\(^{1,2,*,\\dagger}\\) Hang Gao\\(^{1,3,*,\\dagger}\\) \nVikram Voleti\\(^{1}\\) Aaryaman Vasishta\\(^{1}\\) Chun-Han Yao\\(^{1}\\) Mark Boss\\(^{1}\\) Philip Torr\\(^{2}\\) Christian Rupprecht\\(^{2}\\) Varun Jampani\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.225, + 0.798, + 0.245 + ], + "angle": 0, + "content": "\\(^{1}\\)Stability AI \\(^{2}\\)University of Oxford \\(^{3}\\)University of California, Berkeley" + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.253, + 0.618, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.32, + 0.485, + 0.564 + ], + "angle": 0, + "content": "We present STABLE VIRTUAL CAMERA (SEVA), a generalist diffusion model that creates novel views of a scene, given any number of input views and target cameras. Existing works struggle to generate either large viewpoint changes or temporally smooth samples, while relying on specific task configurations. Our approach overcomes these limitations through simple model design, optimized training recipe, and flexible sampling strategy that generalize across view synthesis tasks at test time. As a result, our samples maintain high consistency without requiring additional 3D representation-based distillation, thus streamlining view synthesis in the wild. Furthermore, we show that our method can generate high-quality videos lasting up to half a minute with seamless loop closure. Extensive benchmarking demonstrates that SEVA outperforms existing methods across different datasets and settings." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.585, + 0.222, + 0.601 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.484, + 0.794 + ], + "angle": 0, + "content": "Novel view synthesis (NVS) aims to generate realistic, 3D-consistent images of a scene from arbitrary camera viewpoints given any number of camera-posed input views. Traditional methods, which rely on dense input views, treat NVS as a 3D reconstruction and rendering problem [1-3], but this approach fails with sparse inputs. Generative view synthesis addresses this limitation by leveraging modern deep network priors [4, 5], enabling immersive 3D interactions in uncontrolled environments without the need to capture large image sets per scene. In this work, we focus on generative view synthesis and, unless otherwise specified, refer to it simply as NVS for clarity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Despite recent progress [6-12], NVS in the wild remains limited due to two key challenges: First, existing methods struggle to generate both large viewpoint changes [9, 10] and temporally smooth samples [6-8, 13] while being constrained by rigid task configurations, such as a fixed number of input and target views [7, 9, 11, 12], reviewed in Tab. 1. Second, their sampling consistency is often insufficient, ne" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.285, + 0.907, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.5, + 0.907, + 0.597 + ], + "angle": 0, + "content": "Figure 1. Generative view synthesis. STABLE VIRTUAL CAMERA generates novel views from any number of input views and target cameras, which the user can specify anywhere. We show three examples: single view with simple orbit camera trajectory (top); two views with long camera trajectory (middle); and nine views with large spatial range (bottom). Please refer to the supplementary webpage and video for the video results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.605, + 0.906, + 0.664 + ], + "angle": 0, + "content": "cessitating additional NeRF distillation to fuse inconsistent results into a coherent representation [7, 8, 13]. These limitations hinder their applicability across diverse NVS tasks, which we address in this work." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.667, + 0.909, + 0.85 + ], + "angle": 0, + "content": "We present _STABLE_VIRTUAL CAMERA1 (SEVA), a diffusion-based NVS model that generalizes across a spectrum of view synthesis tasks without requiring NeRF distillation. With a single network, SEVA generates high-quality novel views that strike both large viewpoint changes and temporal smoothness while supporting any number of input and target views. Our approach simplifies the NVS pipeline without requiring distillation from a 3D representation, thus streamlining it for real-world applications. For the first time, we demonstrate high-quality videos lasting up to half a minute with precise camera control and seamless loop closure in 3D. We highlight these results in Fig. 1 and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.863, + 0.907, + 0.9 + ], + "angle": 0, + "content": "1We are naming this model in tribute to the Virtual Camera [14] cinematography technology, a pre-visualization technique to simulate real-world camera movements." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12405" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.443, + 0.106 + ], + "angle": 0, + "content": "showcase more examples of camera control in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.107, + 0.482, + 0.318 + ], + "angle": 0, + "content": "To achieve this, we carefully design our pipeline in three key aspects: model design, training recipe, and sampling method at inference. First, SEVA avoids explicit 3D representations within the network, allowing the model to inherit strong priors from pre-trained 2D models. Second, during training, we carefully craft our view selection strategy to cover both small and large viewpoint changes, ensuring strong generalization to diverse NVS tasks. Third, at inference, we introduce a two-pass procedural sampling approach that supports flexible input-target configurations. Together, these design choices create a versatile 3D \"virtual camera simulation system\" capable of synthesizing novel views along arbitrary camera trajectories with any number of input and target views, without using a 3D representation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.318, + 0.482, + 0.453 + ], + "angle": 0, + "content": "We conducted a unified benchmark across 10 datasets and a variety of experimental settings, including both open-source and proprietary models. It reflects the diversity of real-world NVS tasks across the board and systematically evaluates existing methods beyond their comfort zones. We find that SEVA consistently outperforms previous works, achieving \\(+1.5\\) dB PSNR over the state of the art CAT3D [8]. Our method generalizes well to in-the-wild user captures, with input views ranging from 1 to 32." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.454, + 0.483, + 0.575 + ], + "angle": 0, + "content": "In summary, our key contributions with the SEVA model include: (1) a training strategy for jointly modeling large viewpoint changes and temporal smoothness, (2) a two-pass procedural sampling method for smooth video generation along arbitrary long camera trajectories, (3) a comprehensive benchmark that evaluates NVS methods across different datasets and settings, and (4) an open-source release of model weights to support future research." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.588, + 0.22, + 0.605 + ], + "angle": 0, + "content": "2. Background" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.613, + 0.483, + 0.734 + ], + "angle": 0, + "content": "We consider the evaluation of an NVS model across three key criteria: (1) generation capacity—the ability to synthesize missing regions for large viewpoint changes; (2) interpolation smoothness—the ability to produce seamless transitions between views; and (3) input flexibility—the ability to handle a variable number of input and target views. We review existing NVS models based on these criteria in Tab. 1, including the types of training data." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.281, + 0.759 + ], + "angle": 0, + "content": "2.1. Types of NVS Tasks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.763, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Given \\(M\\) input view images \\(\\mathbf{I}^{\\mathrm{inp}}\\in \\mathbb{R}^{M\\times H\\times W\\times 3}\\) of \\(H\\times W\\) resolution, along with their corresponding cameras \\(\\pi^{\\mathrm{inp}}\\), NVS involves predicting \\(N\\) targets views \\(\\mathbf{I}^{\\mathrm{tgt}}\\in \\mathbb{R}^{N\\times H\\times W\\times 3}\\), specified by their respective cameras \\(\\pi^{\\mathrm{tgt}}\\). For each camera, we assume we know both intrinsics and extrinsics. Based on the number of input views, we define the \"sparse-view regime\" as having up to 8 input views, and the \"semi-dense-view regime\" as an intermediate state bridging the sparse-view regime and dense captures, which" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.089, + 0.905, + 0.369 + ], + "angle": 0, + "content": "
modeltraining datageneration capacityinterpolation smoothnessinput flexibility
Regression-based
pixelNeRF [4]▲X✓sparse (1)
pixelSplat [15]▲X✓sparse (2)
MVSplat [16]▲X✓sparse (2)
Long-LRM [17]▲X✓semi-dense (\\{16, 32\\})
LVSM [11]/▲X✓sparse (\\{2, 4\\})
Diffusion-based: image models
Zero123 [6]▲✓Xsparse (1)
ZeroNVS [13]▲✓Xsparse (1)
ReconFusion [7]▲✓Xsparse (3)
CAT3D [8]▲✓Xsparse ([1, 9])
Diffusion-based: video models
SV3D [18]▲X✓sparse (1)
MotionCtrl [10]▲X✓sparse (1)
ViewCrafter [9]▲X✓sparse (2)
4DiM [12]▲✓✓sparse (\\{1, 2, 8\\})
SEVA▲✓✓sparse ([1, 8]), semi-dense ([9, 32*])
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.376, + 0.907, + 0.488 + ], + "angle": 0, + "content": "Table 1. Comparison of existing NVS models based on the source of training data and key attributes. SEVA is trained on both object-level (●) and scene-level (▲) data, offering flexible input conditioning, strong generation capacity, and smooth view interpolation. We define generation capacity and interpolation smoothness of each work based on their evaluation setting and our benchmark results. *This upper-bound can be up to hundreds for dense captures, we test our model up to 32 views in practice." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.498, + 0.907, + 0.543 + ], + "angle": 0, + "content": "typically involve hundreds of views. Based on the nature of their target views, we bucket a broad range of NVs tasks into \"set NVS\" and \"trajectory NVS\", as shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.906, + 0.63 + ], + "angle": 0, + "content": "Set NVs considers a set of target views in arbitrary order, usually across a large spatial range. The order of views is often not helpful here, and a good NVS model requires great generation capacity to excel at this task. We note that some works address only this task (e.g. ReconFusion [7])." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.906, + 0.747 + ], + "angle": 0, + "content": "Trajectory NVs regards target views along a smooth camera trajectory, such that they form a video sequence. However, they are often sampled within a small spatial range in a shorter video. To solve this task, a good NVS model requires great interpolation smoothness to produce consistent and non-flickering results. We note that some existing works address only this task (e.g. ViewCrafter [9])." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.757, + 0.673, + 0.773 + ], + "angle": 0, + "content": "2.2. Existing Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.841 + ], + "angle": 0, + "content": "We group existing approaches into regression- and diffusion-based models based on their high-level design choices. A more detailed discussion of related works can be found in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.852, + 0.905, + 0.868 + ], + "angle": 0, + "content": "Regression-based models learn a deterministic mapping:" + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.879, + 0.776, + 0.905 + ], + "angle": 0, + "content": "\\[\nf _ {\\theta} \\left(\\mathbf {I} ^ {\\text {i n p}}, \\boldsymbol {\\pi} ^ {\\text {i n p}}, \\boldsymbol {\\pi} ^ {\\text {t g t}}\\right),\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12406" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.1, + 0.904, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.511, + 0.907, + 0.54 + ], + "angle": 0, + "content": "Figure 2. Diverse camera control. SEVA generates photorealistic novel views following diverse camera trajectories. This includes orbit, spiral, zoom out, dolly zooms, and any user-specified trajectories. Please visit our website for more visual results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.548, + 0.483, + 0.653 + ], + "angle": 0, + "content": "to directly generate \\(\\mathbf{I}^{\\mathrm{tgt}}\\) deterministically from \\(\\mathbf{I}^{\\mathrm{inp}}, \\boldsymbol{\\pi}^{\\mathrm{inp}}, \\boldsymbol{\\pi}^{\\mathrm{tgt}}\\). \\(f_{\\theta}\\) can be either an end-to-end network parameterized by \\(\\theta\\), or a composition of a feed-forward prediction of an intermediate 3D representation and then a neural renderer (e.g., NeRF [19] or 3DGS [3]). For the latter case, set NVs and trajectory NVs are solved in the same way since there exists a persistent 3D representation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.662, + 0.483, + 0.69 + ], + "angle": 0, + "content": "Diffusion-based models capture the conditional distribution:" + }, + { + "type": "equation", + "bbox": [ + 0.203, + 0.688, + 0.372, + 0.713 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(\\mathbf {I} ^ {\\mathrm {t g t}} \\mid \\mathbf {I} ^ {\\mathrm {i n p}}, \\boldsymbol {\\pi} ^ {\\mathrm {i n p}}, \\boldsymbol {\\pi} ^ {\\mathrm {t g t}}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.903 + ], + "angle": 0, + "content": "from which \\(\\mathbf{I}^{\\mathrm{tgt}}\\) are sampled [20] iteratively. We highlight two types of models within this scope: Image and Video models. Image models are trained on unordered image sets, such that \\((\\mathbf{I}^{\\mathrm{inp}},\\mathbf{I}^{\\mathrm{tgt}})\\sim \\mathcal{I}\\), where \\(\\mathcal{I} = \\{\\mathbf{I}_{\\rho (1)},\\mathbf{I}_{\\rho (2)},\\dots ,\\mathbf{I}_{\\rho (M + N)}\\}\\) is an image batch, and \\(\\rho (\\cdot)\\) is a random permutation function (camera parameters are omitted for simplicity). Image models usually thrive at set NVs but struggle at trajectory NVS since they are designed to generate images and not videos. Moreover, the unordered nature of all views solicits flexible input conditioning. Video models are instead trained on ordered views, such that \\((\\mathbf{I}^{\\mathrm{inp}},\\mathbf{I}^{\\mathrm{tgt}})\\sim \\mathcal{V}\\), where \\(\\mathcal{V} = \\{\\mathbf{I}_1,\\mathbf{I}_2,\\dots ,\\mathbf{I}_{M + N}\\}\\)" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.546, + 0.688, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.578, + 0.619, + 0.627, + 0.629 + ], + "angle": 0, + "content": "SetNVS" + }, + { + "type": "image", + "bbox": [ + 0.694, + 0.546, + 0.86, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.732, + 0.618, + 0.819, + 0.631 + ], + "angle": 0, + "content": "Trajectory NVS" + }, + { + "type": "image", + "bbox": [ + 0.862, + 0.546, + 0.905, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.635, + 0.907, + 0.678 + ], + "angle": 0, + "content": "Figure 3. Set NVS versus trajectory NVS. Set NVS generates target views as an image set, whereas trajectory NVS produces them as a trajectory video." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.681, + 0.907, + 0.803 + ], + "angle": 0, + "content": "is a set of video frames with preserved ordering. Additional temporal operators may also be used to improve temporal smoothness, such as temporal positional encoding and temporal attention. In contrast with image models, video models thrive at trajectory NVS but struggle at set NVS. Moreover, all existing video models require both input and target views to be ordered (input views followed by target ones), constraining their input flexibility [10, 18, 21, 22]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.816, + 0.605, + 0.831 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.903 + ], + "angle": 0, + "content": "It is challenging to develop a unified model that works for different NVS tasks while being flexible in input and output domains. Our design choices enable SEVA to achieve high generation capacity, smooth view interpolation, and flexi" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12407" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.371, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.088, + 0.908, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.248, + 0.908, + 0.304 + ], + "angle": 0, + "content": "Figure 4. Method. SEVA is trained with fixed sequence length as a “\\(M\\)-in \\(N\\)-out” multi-view diffusion model with standard architecture. During sampling, SEVA can be cast as a generative “\\(P\\)-in \\(Q\\)-out” renderer that works with variable sequence length, where \\(P\\) and \\(Q\\) need not be equal to \\(M\\) and \\(N\\). To enhance temporal and 3D consistency across generated views, especially when generating along a trajectory, we present procedural two-pass sampling as a general strategy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.312, + 0.483, + 0.344 + ], + "angle": 0, + "content": "ble input conditioning (Tab. 1) in both set NVs as well as trajectory NVs. A system overview is provided in Fig. 4." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.351, + 0.339, + 0.367 + ], + "angle": 0, + "content": "3.1. Model Design and Training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.373, + 0.484, + 0.419 + ], + "angle": 0, + "content": "We consider an \" \\(M\\)-in \\(N\\)-out\" multi-view diffusion model \\(p_{\\theta}\\), as noted in Sec. 2.2. We formulate this learning problem as a standard diffusion process [20] without any change." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.426, + 0.484, + 0.653 + ], + "angle": 0, + "content": "Architecture. Our model is based on the publicly available SD 2.1 [23], which consists of an auto-encoder and a latent denoising U-Net. Following [8], we inflate the 2D self-attention of each low-resolution residual block into 3D self-attention [24] within the U-Net. To improve model capacity, we add 1D self-attention along the view axis after each self-attention block via skip connection [25, 26], bumping the model parameters from 870M to 1.3B. Optionally, we further tame this model into a video model by introducing 3D convolutions after each residual block via skip connection, similar to [27, 28], yielding 1.5B total parameters. The temporal pathway can be enabled during inference when frames within one forward pass are known to be spatially ordered, enhancing output's smoothness. For the conditioning of our model, please refer to the appendix." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Training. Let us define the training context window length \\( T = |\\mathbf{I}^{\\mathrm{inp}}| + |\\mathbf{I}^{\\mathrm{tgt}}| = M + N \\). One natural goal is to support large \\( T \\) such that we can generate larger set of frames. However, we find that naive training is prone to divergence, and we thus employ a two-stage training curriculum. During the first stage, we train our model with \\( T = 8 \\) with a batch size of 1472 for \\( 100K \\) iterations. In the second stage, we train our model with \\( T = 21 \\) with a batch size of 512 for \\( 600K \\) iterations. Given a training video sequence, we randomly sample the number of input frames \\( M \\in [1,T - 1] \\) and the frames \\( (\\mathbf{I}^{\\mathrm{inp}},\\mathbf{I}^{\\mathrm{gt}}) \\) across the entire sequence to encourage diversity. We find it important to jointly sample with a smaller subsampling stride to ensure sufficient temporal granularity and avoid missing critical transitions with a small probability (0.2 is used in practice). In the optional video training stage, we only train" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.312, + 0.907, + 0.448 + ], + "angle": 0, + "content": "temporal weights with data sampled with a small subsampling stride and a batch size of 512 for \\(200K\\) iterations. We shift the signal-to-noise ratio (SNR) in all stages as more noise is necessary to remove the input information, corroborating findings from [8, 29, 30]. The model is trained with squared images with \\(H = W = 576\\). The training data is a superset of data used in [8] and derives from all public academic datasets, consisting of around \\(340K\\) objects and \\(80K\\) scenes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.46, + 0.722, + 0.476 + ], + "angle": 0, + "content": "3.2. Sampling Novel Views" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.483, + 0.906, + 0.589 + ], + "angle": 0, + "content": "Once the diffusion model is trained, we can sample it for a wide range of NVS tasks during test time. Formally, let us consider a \" \\(P\\)-in \\(Q\\)-out\" NVS task during testing, where we are given \\(P = |\\mathbf{I}^{\\mathrm{inp}}|\\) input frames and aim to produce \\(Q = |\\mathbf{I}^{\\mathrm{tgt}}|\\) target frames. Our goal is to design a generic sampling strategy that works for all \\(P\\) and \\(Q\\) configurations, where \\(P\\) and \\(Q\\) need not be equal to \\(M\\) and \\(N\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.59, + 0.906, + 0.756 + ], + "angle": 0, + "content": "We make two key observations: First, within a single forward pass, predictions are 3D consistent, provided the model is well-trained. Second, when \\( P + Q > T \\), \\( \\mathbf{I}^{\\mathrm{gt}} \\) must be split into smaller chunks of size \\( Q_{i} \\) such that \\( P + Q_{i} \\leqslant T \\) for the \\( i^{\\mathrm{th}} \\) forward pass. We term this practice one-pass sampling. However, predictions across these forward passes would be inconsistent unless they share common frames to maintain local consistency within a spatial neighborhood. Building on these observations, we summarize our sampling process under two scenarios: \\( P + Q \\leqslant T \\) and \\( P + Q > T \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.767, + 0.906, + 0.844 + ], + "angle": 0, + "content": "\\(P + Q \\leqslant T\\). We fit the task within one forward pass for simplicity and consistency. As shown in the appendix, we find it works better to pad the forward pass to have exactly \\(T\\) frames by repeating the first input image, compared to changing the context window \\(T\\) zero-shot." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "\\(P + Q > T\\). We propose procedural two-pass sampling: In the first pass, we generate anchor frames \\(\\mathbf{I}^{\\mathrm{acr}}\\) using all input frames \\(\\mathbf{I}^{\\mathrm{inp}}\\). In the second pass, we divide \\(\\mathbf{I}^{\\mathrm{tgt}}\\) into" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12408" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.483, + 0.121 + ], + "angle": 0, + "content": "chunks and generate them using \\(\\mathbf{I}^{\\mathrm{acr}}\\) (and optionally \\(\\mathbf{I}^{\\mathrm{inp}}\\)) according to the spatial distribution of \\(\\mathbf{I}^{\\mathrm{acr}}\\) and \\(\\mathbf{I}^{\\mathrm{gt}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.483, + 0.182 + ], + "angle": 0, + "content": "Given the distinct nature of the two tasks of interest—set NVS and trajectory NVS—e.g., differences in the availability of views' ordering, we design tailored chunking strategies for each task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.484, + 0.273 + ], + "angle": 0, + "content": "For set NVS, we consider nearest procedural sampling. We first generate \\(\\mathbf{I}^{\\mathrm{acr}}\\) based on pre-defined trajectory priors, similar to [8], e.g., 360 trajectories for object-centric scenes or spiral trajectories for forward-facing scenes. We then divide \\(\\mathbf{I}^{\\mathrm{gt}}\\) into chunks w.r.t. \\(\\mathbf{I}^{\\mathrm{acr}}\\) using nearest neighbor. Specifically, the \\(i^{\\mathrm{th}}\\) forward pass involves:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.283, + 0.443, + 0.302 + ], + "angle": 0, + "content": "\\[\n\\operatorname {n e a r e s t}: \\left\\{\\mathbf {I} _ {i} ^ {\\mathrm {a c r}} \\right\\} \\cup \\left\\{\\mathbf {I} _ {j} ^ {\\mathrm {t g t}} \\mid \\mathrm {N N} (\\mathbf {I} _ {j} ^ {\\mathrm {t g t}}, \\mathbf {I} ^ {\\mathrm {a c r}}) = \\mathbf {I} _ {i} ^ {\\mathrm {a c r}} \\right\\}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.311, + 0.484, + 0.431 + ], + "angle": 0, + "content": "We considered two strategies of procedural sampling: nearest as described above, and \\(\\mathrm{gt} +\\) nearest strategy by appending \\(\\mathbf{I}^{\\mathrm{inp}}\\) into each forward pass. We find that the \\(\\mathrm{gt} +\\) nearest strategy performs better than nearest, and thus default to it instead. In the absence of trajectory priors, we revert to one-pass sampling. In practice, employing nearest anchors enhances qualitative consistency, albeit on a limited scale." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.432, + 0.484, + 0.507 + ], + "angle": 0, + "content": "For trajectory NVS, we consider interp procedural sampling. We first generate a subset of target frames as \\(\\mathbf{I}^{\\mathrm{acr}}\\) by uniformly spanning the target camera path with a stride \\(\\Delta = \\left\\lfloor \\frac{Q}{T - 2}\\right\\rfloor\\). We then generate the rest of \\(\\mathbf{I}^{\\mathrm{tgt}}\\) as segments between those anchors:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.517, + 0.438, + 0.538 + ], + "angle": 0, + "content": "\\[\n\\mathrm {i n t e r p}: \\{\\mathbf {I} _ {i} ^ {\\mathrm {a c r}}, \\mathbf {I} _ {i \\cdot \\Delta + 1} ^ {\\mathrm {t g t}}, \\dots , \\mathbf {I} _ {(i + 1) \\cdot \\Delta - 1} ^ {\\mathrm {t g t}}, \\mathbf {I} _ {i + 1} ^ {\\mathrm {a c r}} \\}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.546, + 0.484, + 0.666 + ], + "angle": 0, + "content": "Since the input to the model is ordered, we can leverage temporal weights to further improve smoothness (Sec. 4.3). Similarly, \\(\\mathrm{gt} + \\mathrm{interp}\\) is possible by appending \\(\\mathbf{I}^{\\mathrm{inp}}\\) with \\(\\Delta = \\left\\lfloor \\frac{Q}{T - P - 2} \\right\\rfloor\\). We find that interp is sufficiently robust, and choose it as the default option. The interp strategy drastically outperforms its counterparts (e.g., one-pass, or \\(\\mathrm{gt} + \\mathrm{nearest}\\) procedural sampling) in terms of temporal smoothness." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.675, + 0.414, + 0.692 + ], + "angle": 0, + "content": "3.3. Scaling Sampling for Large \\(P\\) and \\(Q\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.697, + 0.484, + 0.759 + ], + "angle": 0, + "content": "Next, we examine two special cases when \\( P + Q > T \\): \\( P > T \\) and \\( Q \\gg T \\). Here, we make a tailored design for anchor generation in the first pass, while keeping target generation in the second pass unchanged." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "\\(P > T\\). In the semi-dense-view regime (e.g., \\(P = 32\\)), we extend the context window length \\(T\\) zero-shot to accommodate all \\(P\\) input views and anchor views in one pass during anchor generation. Empirically, \\(T\\) can even be extended up to hundreds without severe degradation in photorealism in the generated outputs. We find that the diffusion model generalizes well in this case as long as the input views cover the majority of the scene, shifting the task from generation to primarily interpolation. In the sparse-view" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.905, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.224, + 0.907, + 0.28 + ], + "angle": 0, + "content": "Figure 5. Anchor generation when \\( Q \\gg T \\). We introduce a memory bank composed of previously-generated anchor views and their corresponding camera poses. The lookup of spatial neighbors helps improve long-term 3D consistency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.29, + 0.907, + 0.351 + ], + "angle": 0, + "content": "regime \\((P \\leq 8)\\), we observe similar performance degradation caused by zero-shot extension of \\(T\\) compared to what we have found when \\(P + Q \\leqslant T\\). Refer to the appendix for a detailed discussion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.374, + 0.909, + 0.571 + ], + "angle": 0, + "content": "\\(Q \\gg T\\). When the number of target views \\(Q\\) is large, e.g., in large-set NVS or long-trajectory NVS, even anchors will be chunked into different forwards in the first pass, leading to the inconsistency of anchors. To this end, we maintain a memory bank of anchor views previously generated, as shown in Fig. 5. We generate anchors autoregressively by retrieving their spatially nearest ones from the memory bank, similar to the nearest strategy introduced above for the second pass. In Sec. 4.4, we show that this strategy drastically outperforms the standard practice of reusing temporally nearest anchors previously generated in long video literature [22], in terms of long-range 3D consistency, especially for hard trajectories." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.583, + 0.646, + 0.601 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.907, + 0.716 + ], + "angle": 0, + "content": "We employ a single model for a spectrum of settings and find that SEVA model generalizes well under the three criteria (Tab. 1). We cover different NVS tasks (set NVS and trajectory NVS) and examine one special task of interest—long trajectory NVS. We also cover different input regimes (single-view, sparse-view, and semi-dense-view). A discussion about key properties is presented in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.727, + 0.641, + 0.743 + ], + "angle": 0, + "content": "4.1. Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Datasets, splits, and the number of input views. We collect 10 commonly used datasets to benchmark NVS, encompassing a diverse range of scene distributions and complexities. We consider (1) object datasets [32, 33]; (2) object-centric scene datasets [34-38]; and (3) scene datasets [39-41]. To establish a comprehensive and rigorous comparison with baselines, we consider different dataset splits utilized in prior works with the same input-view configuration, unless specified as our split (O). These include splits used in 4DiM [12] (D), ViewCrafter [9] (V)," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12409" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.907, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.428, + 0.907, + 0.486 + ], + "angle": 0, + "content": "Figure 6. SOTA comparison on set NVs (top) and trajectory NVs (bottom) across varying numbers of input views. From left to right and top to bottom, inputs views are (top, i.e. set NVS) 1, 2, 3, 32; (bottom, i.e. trajectory NVS) 1, 2, 1, 1, 3, 3. For cases with multiple input views, the view closest to the target is shown over the rest. We compare with open-source approaches—ViewCrafter [9] (VC) and DepthSplat [31] (DS)—as well as proprietary ones including LVSM [11], Long-LRM [17] (LLRM), 4DiM [12], and CAT3D [8]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.483, + 0.525 + ], + "angle": 0, + "content": "pixelSplat [15] (P), ReconFusion [7] (R), SV3D [18] (S), and Long-LRM [17] (L)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.544, + 0.483, + 0.651 + ], + "angle": 0, + "content": "Small-viewpoint versus large-viewpoint NVS. Sweeping across all datasets, splits, and input-view configurations reveals a diverse benchmark of setups. To better evaluate models' generation capacity and interpolation smoothness (Sec. 2.1), we propose to categorize this setup into two groups—small-viewpoint NVS and large-viewpoint NVS—depending on the disparity between \\(\\mathbf{I}^{\\mathrm{tgt}}\\) and \\(\\mathbf{I}^{\\mathrm{inp}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.671, + 0.483, + 0.807 + ], + "angle": 0, + "content": "Baselines. We consider a range of proprietary models, including ReconFusion [7], CAT3D [8], 4DiM [12], LVSM [11], and Long-LRM [17]. We also consider various open-source models, including SV3D [18], MVSplat [16], depthSplat [31], MotionCtrl [42], and ViewCrafter [9]. As outlined in Sec. 2.2, these baselines encompass both regression-based and diffusion-based approaches, providing a comprehensive framework for comparison. Please refer to the appendix for more details of the benchmark." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.194, + 0.833 + ], + "angle": 0, + "content": "4.2. Set NVs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Quantitative comparison. The input and target views are chosen following the splits used in previous methods. The order of target views is not preserved. We use standard metrics of peak signal-to-noise ratio (PSNR), structural simi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.907, + 0.583 + ], + "angle": 0, + "content": "larity index measure [44] (SSIM), and learned perceptual image patch similarity [45] (LPIPS). Only PSNR is showcased here due to space limits, with the rest deferred to the appendix. Empirically, our method shows a greater performance improvement on LPIPS, reflecting the photorealism of our results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.78 + ], + "angle": 0, + "content": "For small-viewpoint set NVs, Tab. 2 shows that SEVA sets state-of-the-art results in the majority of splits. In the sparse-view regime (i.e., \\( P \\leqslant 8 \\)), SEVA excels across different datasets when \\( P > 1 \\). For example, a performance gain of +6.0 dB PSNR is achieved on LLFF with \\( P = 3 \\). In the semi-dense-view regime (e.g., \\( P = 32 \\)), SEVA surprisingly performs favorably against the specialized model [17], despite not being specifically designed for this setup. For example, SEVA lags behind the state-of-the-art method [17] by only 1.7 dB on T&T. A detailed discussion on the RE10K dataset can be found in the appendix. On object datasets OO3D and GSO, SEVA achieves a significantly higher state-of-the-art PSNR compared to all other methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.901 + ], + "angle": 0, + "content": "For large-viewpoint set NVs, Tab. 3 shows that SEVA's quantitative advantages are even more prominent here, revealing clear benefits of SEVA in terms of generation capacity when the camera spans a large spatial range. On Mip360 with \\( P = 3 \\), SEVA improves over previous state-of-the-art method CAT3D [8] by \\( +0.6 \\) dB PSNR. On harder scenes like DL3DV and T&T with different input-view configurations, SEVA obtains a clear performance lead." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12410" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.089, + 0.895, + 0.278 + ], + "angle": 0, + "content": "
Methoddataset split POO3D OGSO ORE10KLLFF R [7] 1 3DTU R [7] 1 3CO3D V [9] R [7] 1 3WRGBD Oe 3 6Mip360 O L [17] 6 32DL3DV T & T V [9] L [17] 1 32
O 3O 3D [12] 1V [9] 1P [15] 2R [7] 13
Regression-based models
Long-LRM [17]-------------23.86 - 18.20
MVSplat [16]14.78 15.6715.21 16.5220.42 20.9020.32 19.2426.39 27.4421.56 21.8725.64 22.5411.23 12.0712.50 12.6213.87 15.52 14.1513.52 13.52 13.2314.56 12.54 15.9313.56 14.24 14.2313.24 13.22 14.35
DepthSplat [31]
LVSM [11]----29.67---------
Diffusion-based models
MotionCtrl [42]--12.74 17.0816.29 ------15.46 ----13.29 -
4DiM [12]--17.08-----------
ViewCrafter [9]14.64 30.3015.93 31.5320.43 17.9922.04 18.5621.42 25.6620.88 18.1122.81 27.5710.53 14.0313.52 14.4712.66 16.40 18.4718.96 14.72 18.4016.42 12.66 19.7514.59 13.78 16.70- 18.07 - 15.16
SEVA
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.282, + 0.908, + 0.325 + ], + "angle": 0, + "content": "Table 2. PSNR† on small-viewpoint set NVS. \\( P \\) denotes the number of input views. For all results with \\( P = 1 \\), we sweep the unit length for camera normalization due to the model's scale ambiguity. \\( \\mathrm{O}_e \\) and \\( \\mathrm{O}_h \\) denote the easy and hard split of our split, respectively. Underlined numbers are run by us using the official released code." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.332, + 0.603, + 0.457 + ], + "angle": 0, + "content": "
Methoddataset split POO3D S [18] 1GSO S [18] 1CO3D R [7] 1WRGBD Oh 1Mip360 R [7] 3DL3DV O 1T&T O 31369
SV3D [18]19.2820.38----------
DepthSplat [31]11.5612.3210.429.3513.5310.4912.549.6312.528.639.7810.12
CAT3D [8]------15.15-----
ViewCrafter [9]10.5611.4210.119.1213.459.7910.348.9711.509.239.8810.32
SEVA19.2520.6515.3014.3717.2812.9315.7813.0115.9511.2812.6513.80
" + }, + { + "type": "table", + "bbox": [ + 0.608, + 0.334, + 0.902, + 0.453 + ], + "angle": 0, + "content": "
Methodsmall-viewpointlarge-viewpoint
RE10KLLFFDTUCO3DMip360
ZipNeRF [43]20.7717.239.1814.3412.77
ZeroNVS [13]19.1115.9116.7117.1314.44
ReconFusion [7]25.8421.3420.7419.5915.50
CAT3D [8]26.7821.5822.0220.5716.62
SEVA27.9521.8822.6821.8817.82
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.466, + 0.904, + 0.509 + ], + "angle": 0, + "content": "Table 3. PSNR† on large-viewpoint set NVs. For all results with \\( P = 1 \\), we sweep the Table 4. PSNR† on 3DGS renderings for set unit length for camera normalization due to the model's scale ambiguity. Underlined NVS. Results are reported on the ReconFusion [7] numbers are run by us using the officially released code. split with \\( P = 3 \\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.517, + 0.483, + 0.625 + ], + "angle": 0, + "content": "Comparison of 3D reconstruction. To enable a direct quantitative comparison with prior works [7, 8], we adopt the few-view 3D reconstruction pipeline described in [8]. We optimize the camera parameters and apply LPIPS loss [46] during the distillation. Finally, we render the distilled 3D model on the test views and report the performance in Tab. 4. SEVA shows a consistent performance lead." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.63, + 0.483, + 0.827 + ], + "angle": 0, + "content": "Qualitative comparison. Fig. 6 top panel shows a qualitative comparison with diverse baselines. For small-viewpoint set NVs, the output from SEVA with the best scale exhibits desirable alignment with the ground truth while being more photorealistic in details. Compared with LVSM [11] on the P split of RE10K, SEVA produces sharper images, also corroborating that lower PSNR arises from scale ambiguity rather than interpolation quality. Similar trends hold when compared to Long-LRM [17] on DL3DV with \\(P = 32\\). For large-viewpoint set NVs, we compare with DepthSplat [31] on DL3DV with \\(P = 3\\). DepthSplat fails to produce reasonable results when the viewpoint change is too large and falls short in overall visual quality." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.251, + 0.851 + ], + "angle": 0, + "content": "4.3. Trajectory NVS" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Qualitative results. Fig. 2 presents qualitative results, illustrating trajectories of varying complexities with different numbers of input views across diverse types." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.517, + 0.907, + 0.623 + ], + "angle": 0, + "content": "In the single-view regime (i.e., \\( P = 1 \\)), we manually craft a set of common camera movements/effects, e.g., look-at 360, spiral, panning, zoom-in, zoom-out, dolly zoom, etc. We observe that SEVA generalizes to a wide range of images and demonstrates accurate camera-following capacity. An extensive sweeping of camera movements on 4 types of images is provided in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.624, + 0.909, + 0.82 + ], + "angle": 0, + "content": "In the sparse-view regime with few input views (i.e., \\(1 < P \\leqslant 8\\)), we observe that SEVA demonstrates strong generalization to in-the-wild real-world images and versatility in adapting to different numbers of input views. The output forms a smooth trajectory video with subtle temporal flickering, revealing its capacity to interpolate between views smoothly. In the last row of Fig. 2, our model generates plausible results at the end of the trajectory—an area unseen in the input observations—demonstrating its strong generation capacity. In the semi-dense-view regime (i.e., \\(P > 9\\)), we similarly find that SEVA is surprisingly able to produce a smooth trajectory video with minimal artifacts. Please check the supplementary website for video results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Qualitative comparison. Fig. 6 bottom panel presents a qualitative comparison with diverse baselines. In the single-view regime (i.e., \\( P = 1 \\)), we compare to 4DiM [12] and CAT3D [8]. We observe more photo-realistic and sharper" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12411" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.098, + 0.09, + 0.478, + 0.219 + ], + "angle": 0, + "content": "
Methodsplit datasetsmall-viewpointlarge-viewpoint
V [9]O
RECO3DT&TREDTUWRDLT&T
MotionCtrl [42]16.2915.4613.29------
DepthSplat [31]19.2413.2314.3525.2314.6812.4511.329.11
ViewCrafter [9]22.0418.9618.0726.5418.9913.4411.459.68
SEVA18.5618.4015.1627.3419.9917.7915.7611.92
SEVA (+ temp.)18.6218.4315.1327.3620.1917.9315.7811.99
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.23, + 0.482, + 0.3 + ], + "angle": 0, + "content": "Table 5. PSNR\\(\\uparrow\\) on trajectory NVS. temp. denotes optional temporal pathway. RE, WR, and DL denotes RE10K, WRGBD, and DL3DV, respectively. For the V [9] split, \\(P = 1\\) with unit length swept; for the O split, \\(P = 3\\). Underlined numbers are run by us using the officially released code." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.308, + 0.482, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.499, + 0.483, + 0.584 + ], + "angle": 0, + "content": "Figure 7. Temporal quality. Vertical slices of a rendered novel camera path on the BONSAI scene from Mip-NeRF360 [38] illustrate the temporal quality across adjacent viewpoints. One-pass or gt + nearest procedural sampling results in notable flickering, whereas interp procedural sampling ensures temporally smooth rendering." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.593, + 0.484, + 0.7 + ], + "angle": 0, + "content": "output from our model, especially in the background area for object-centric scenes. 4DiM outputs tend to be cartoonish and over-simplistic, given that the model is only trained on RE10K. In the sparse-view regime with few input views (i.e., \\( P = 3 \\)), we compare with CAT3D and observe that our model demonstrates more photo-realistic textures, especially in the background." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Quantitative comparison. We use the same input views as in set NVS for each split. We use all frames from each scene as target views such that they form a smoothly transitioning trajectory video, i.e., \\(\\mathbf{I}^{\\mathrm{gt}}\\sim \\mathcal{V}\\). For small-viewpoint trajectory NVS, Tab. 5 compares SEVA with baselines on PSNR. SEVA performs favorably against other methods in V split with \\(P = 1\\). The performance lead of ViewCrafter is mainly attributed to its training on high-resolution images. For large-viewpoint trajectory NVS with \\(P = 3\\), SEVA consistency sets new state-of-the-art results. Applying the temporal pathway further boosts performance and improves smoothness, indicating the benefits of the gated" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.907, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.299, + 0.907, + 0.385 + ], + "angle": 0, + "content": "Figure 8. Long-range 3D consistency. We visualize samples following a camera path looping three times around the TELEPHONE-BOOTH scene. Lookup using spatial neighbors from the memory bank (ours) notably improves view consistency and reduces artifacts in recurring locations across different loops, compared to lookup using temporal neighbors (baseline)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.391, + 0.6, + 0.404 + ], + "angle": 0, + "content": "architecture." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.907, + 0.545 + ], + "angle": 0, + "content": "Ablation on two-pass procedural sampling. Fig. 7 shows that one-pass sampling introduces visible temporal flickering and abrupt scene changes. In contrast, interp procedural sampling produces the smoothest transitions, outperforming the alternative \\(\\mathrm{gt} +\\) nearest strategy and mitigating noticeable flickering. Additional quantitative ablation is deferred to the appendix, including 3DGS distillation and video quality metrics." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.553, + 0.719, + 0.57 + ], + "angle": 0, + "content": "4.4. Long-Trajectory NVS" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.576, + 0.906, + 0.697 + ], + "angle": 0, + "content": "Fig. 8 presents a qualitative demonstration of NVS over a long trajectory of up to 1000 frames. With the memory bank maintaining previously generated anchors, SEVA achieves robust 3D consistency for long-trajectory NVS, e.g., the building in front of and the plantation after the booth. Comparing it to using temporal nearest anchors previously generated, using spatially nearest ones demonstrates a clear advantage in terms of long-term consistency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.71, + 0.634, + 0.725 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We introduce SEVA, a generalist diffusion model for novel view synthesis that balances large viewpoint changes and smooth interpolation while supporting flexible input and target configurations. By designing a diffusion-based architecture without 3D representation, a structured training strategy, and a two-pass procedural sampling approach, SEVA achieves 3D consistent rendering across diverse NVs tasks. Extensive benchmarking demonstrates its superiority over existing methods, with strong generalization to real-world scenes. For broader impact and limitations, please refer to the appendix." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12412" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, pages 405-421, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.228 + ], + "angle": 0, + "content": "[2] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.231, + 0.483, + 0.285 + ], + "angle": 0, + "content": "[3] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4), July 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.288, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[4] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In CVPR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.331, + 0.482, + 0.372 + ], + "angle": 0, + "content": "[5] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3D using 2D diffusion. arXiv, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.375, + 0.483, + 0.443 + ], + "angle": 0, + "content": "[6] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3D object. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.446, + 0.484, + 0.529 + ], + "angle": 0, + "content": "[7] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21551-21561, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.483, + 0.599 + ], + "angle": 0, + "content": "[8] Ruiqi Gao*, Aleksander Holynski*, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul P. Srinivasan, Jonathan T. Barron, and Ben Poole*. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv, 2024. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.602, + 0.483, + 0.671 + ], + "angle": 0, + "content": "[9] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. ViewCrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.673, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[10] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.745, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[11] Haian Jin, Hanwen Jiang, Hao Tan, Kai Zhang, Sai Bi, Tianyuan Zhang, Fujun Luan, Noah Snavely, and Zexiang Xu. Lvsm: A large view synthesis model with minimal 3d inductive bias, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[12] Daniel Watson, Saurabh Saxena, Lala Li, Andrea Tagliasacchi, and David J Fleet. Controlling space and time with diffusion models. arXiv preprint arXiv:2407.07860, 2024. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[13] Kyle Sargent, Zizhang Li, Tanmay Shah, Charles Herrmann, Hong-Xing Yu, Yunzhi Zhang, Eric Ryan Chan, Dmitry Lagun, Li Fei-Fei, Deqing Sun, et al. Zeronvs: Zero-shot" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "360-degree view synthesis from a single real image. arXiv preprint arXiv:2310.17994, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[14] Virtual camera system. https://en.wikipedia.org/wiki/Virtual_camera_system. Accessed: 5-Mar-2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[15] David Charatan, Sizhe Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In arXiv, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[16] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In European Conference on Computer Vision, pages 370-386. Springer, 2025. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.293, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[17] Chen Ziwen, Hao Tan, Kai Zhang, Sai Bi, Fujun Luan, Yicong Hong, Li Fuxin, and Zexiang Xu. Long-lrm: Long-sequence large reconstruction model for wide-coverage gaussian splats. arXiv preprint arXiv:2410.12781, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[18] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitrii Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. SV3D: Novel multi-view synthesis and 3D generation from a single image using latent video diffusion. In European Conference on Computer Vision, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[19] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[21] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrl: Enabling camera control for text-to-video generation. arXiv preprint arXiv:2404.02101, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[22] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. arXiv, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.647, + 0.907, + 0.715 + ], + "angle": 0, + "content": "[23] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, June 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.717, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[24] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. MVDream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[26] P Goyal. Accurate, large minibatch sg d: training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Content" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12413" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.093, + 0.482, + 0.119 + ], + "angle": 0, + "content": "ference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[28] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[29] Emiel Hoogeboom, Jonathan Heek, and Tim Salimans. simple diffusion: End-to-end diffusion for high resolution images. arXiv preprint arXiv:2301.11093, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[30] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICLR, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[31] Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. Depthsplat: Connecting gaussian splatting and depth. arXiv preprint arXiv:2410.13862, 2024. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.359, + 0.482, + 0.44 + ], + "angle": 0, + "content": "[32] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, et al. Omniobject3D: Large-vocabulary 3D object dataset for realistic perception, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 803-814, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[33] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google Scanned Objects: A high-quality dataset of 3D scanned household items. In 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560. IEEE, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[34] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (ToG), 38(4):1-14, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.596, + 0.482, + 0.65 + ], + "angle": 0, + "content": "[35] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 406-413, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[36] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In International Conference on Computer Vision, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.721, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[37] Hongchi Xia, Yang Fu, Sifei Liu, and Xiaolong Wang. RGBd objects in the wild: Scaling real-world 3d object learning from rgb-d videos, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.763, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[38] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-NeRF: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[39] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018. 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[40] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[41] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[42] Tsai-Shien Chen, Chieh Hubert Lin, Hung-Yu Tseng, Tsung-Yi Lin, and Ming-Hsuan Yang. Motion-conditioned diffusion model for controllable video synthesis. arXiv preprint arXiv:2304.14404, 2023. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[43] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.905, + 0.415 + ], + "angle": 0, + "content": "[44] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.486 + ], + "angle": 0, + "content": "[45] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.488, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[46] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric, 2018. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12414" + } + ] +] \ No newline at end of file diff --git a/2025/Stable Virtual Camera_ Generative View Synthesis with Diffusion Models/8df1f571-dcd1-4fc3-a74d-5f6f5bf1e485_origin.pdf b/2025/Stable Virtual Camera_ Generative View Synthesis with Diffusion Models/8df1f571-dcd1-4fc3-a74d-5f6f5bf1e485_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..611eaf3042163791623a5b825b7dfcd01e2a406f --- /dev/null +++ b/2025/Stable Virtual Camera_ Generative View Synthesis with Diffusion Models/8df1f571-dcd1-4fc3-a74d-5f6f5bf1e485_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c1fd044fe5517aa9f1a69609e44fdfb4551bfff2206098085c8f31bdb5085711 +size 8896472 diff --git a/2025/Stable Virtual Camera_ Generative View Synthesis with Diffusion Models/full.md b/2025/Stable Virtual Camera_ Generative View Synthesis with Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..88194123f803c4c44dd884035dc7c8ce8c206c54 --- /dev/null +++ b/2025/Stable Virtual Camera_ Generative View Synthesis with Diffusion Models/full.md @@ -0,0 +1,280 @@ +# STABLE VIRTUAL CAMERA: Generative View Synthesis with Diffusion Models + +Jensen (Jinghao) Zhou $^{1,2,*,\dagger}$ Hang Gao $^{1,3,*,\dagger}$ +Vikram Voleti $^{1}$ Aaryaman Vasishta $^{1}$ Chun-Han Yao $^{1}$ Mark Boss $^{1}$ Philip Torr $^{2}$ Christian Rupprecht $^{2}$ Varun Jampani $^{1}$ + +$^{1}$ Stability AI $^{2}$ University of Oxford $^{3}$ University of California, Berkeley + +![](images/7972dd8723f41d3ffe2accb551f716bcfa11ff6f527fc96112d36828f2b1c496.jpg) + +# Abstract + +We present STABLE VIRTUAL CAMERA (SEVA), a generalist diffusion model that creates novel views of a scene, given any number of input views and target cameras. Existing works struggle to generate either large viewpoint changes or temporally smooth samples, while relying on specific task configurations. Our approach overcomes these limitations through simple model design, optimized training recipe, and flexible sampling strategy that generalize across view synthesis tasks at test time. As a result, our samples maintain high consistency without requiring additional 3D representation-based distillation, thus streamlining view synthesis in the wild. Furthermore, we show that our method can generate high-quality videos lasting up to half a minute with seamless loop closure. Extensive benchmarking demonstrates that SEVA outperforms existing methods across different datasets and settings. + +# 1. Introduction + +Novel view synthesis (NVS) aims to generate realistic, 3D-consistent images of a scene from arbitrary camera viewpoints given any number of camera-posed input views. Traditional methods, which rely on dense input views, treat NVS as a 3D reconstruction and rendering problem [1-3], but this approach fails with sparse inputs. Generative view synthesis addresses this limitation by leveraging modern deep network priors [4, 5], enabling immersive 3D interactions in uncontrolled environments without the need to capture large image sets per scene. In this work, we focus on generative view synthesis and, unless otherwise specified, refer to it simply as NVS for clarity. + +Despite recent progress [6-12], NVS in the wild remains limited due to two key challenges: First, existing methods struggle to generate both large viewpoint changes [9, 10] and temporally smooth samples [6-8, 13] while being constrained by rigid task configurations, such as a fixed number of input and target views [7, 9, 11, 12], reviewed in Tab. 1. Second, their sampling consistency is often insufficient, ne + +![](images/fb92f49ac881fc46a4d8f22a85a7508e73995c0848ced9210adf4443401260a8.jpg) +Figure 1. Generative view synthesis. STABLE VIRTUAL CAMERA generates novel views from any number of input views and target cameras, which the user can specify anywhere. We show three examples: single view with simple orbit camera trajectory (top); two views with long camera trajectory (middle); and nine views with large spatial range (bottom). Please refer to the supplementary webpage and video for the video results. + +cessitating additional NeRF distillation to fuse inconsistent results into a coherent representation [7, 8, 13]. These limitations hinder their applicability across diverse NVS tasks, which we address in this work. + +We present _STABLE_VIRTUAL CAMERA1 (SEVA), a diffusion-based NVS model that generalizes across a spectrum of view synthesis tasks without requiring NeRF distillation. With a single network, SEVA generates high-quality novel views that strike both large viewpoint changes and temporal smoothness while supporting any number of input and target views. Our approach simplifies the NVS pipeline without requiring distillation from a 3D representation, thus streamlining it for real-world applications. For the first time, we demonstrate high-quality videos lasting up to half a minute with precise camera control and seamless loop closure in 3D. We highlight these results in Fig. 1 and + +showcase more examples of camera control in Fig. 2. + +To achieve this, we carefully design our pipeline in three key aspects: model design, training recipe, and sampling method at inference. First, SEVA avoids explicit 3D representations within the network, allowing the model to inherit strong priors from pre-trained 2D models. Second, during training, we carefully craft our view selection strategy to cover both small and large viewpoint changes, ensuring strong generalization to diverse NVS tasks. Third, at inference, we introduce a two-pass procedural sampling approach that supports flexible input-target configurations. Together, these design choices create a versatile 3D "virtual camera simulation system" capable of synthesizing novel views along arbitrary camera trajectories with any number of input and target views, without using a 3D representation. + +We conducted a unified benchmark across 10 datasets and a variety of experimental settings, including both open-source and proprietary models. It reflects the diversity of real-world NVS tasks across the board and systematically evaluates existing methods beyond their comfort zones. We find that SEVA consistently outperforms previous works, achieving $+1.5$ dB PSNR over the state of the art CAT3D [8]. Our method generalizes well to in-the-wild user captures, with input views ranging from 1 to 32. + +In summary, our key contributions with the SEVA model include: (1) a training strategy for jointly modeling large viewpoint changes and temporal smoothness, (2) a two-pass procedural sampling method for smooth video generation along arbitrary long camera trajectories, (3) a comprehensive benchmark that evaluates NVS methods across different datasets and settings, and (4) an open-source release of model weights to support future research. + +# 2. Background + +We consider the evaluation of an NVS model across three key criteria: (1) generation capacity—the ability to synthesize missing regions for large viewpoint changes; (2) interpolation smoothness—the ability to produce seamless transitions between views; and (3) input flexibility—the ability to handle a variable number of input and target views. We review existing NVS models based on these criteria in Tab. 1, including the types of training data. + +# 2.1. Types of NVS Tasks + +Given $M$ input view images $\mathbf{I}^{\mathrm{inp}}\in \mathbb{R}^{M\times H\times W\times 3}$ of $H\times W$ resolution, along with their corresponding cameras $\pi^{\mathrm{inp}}$ , NVS involves predicting $N$ targets views $\mathbf{I}^{\mathrm{tgt}}\in \mathbb{R}^{N\times H\times W\times 3}$ , specified by their respective cameras $\pi^{\mathrm{tgt}}$ . For each camera, we assume we know both intrinsics and extrinsics. Based on the number of input views, we define the "sparse-view regime" as having up to 8 input views, and the "semi-dense-view regime" as an intermediate state bridging the sparse-view regime and dense captures, which + +
modeltraining datageneration capacityinterpolation smoothnessinput flexibility
Regression-based
pixelNeRF [4]▲X✓sparse (1)
pixelSplat [15]▲X✓sparse (2)
MVSplat [16]▲X✓sparse (2)
Long-LRM [17]▲X✓semi-dense (\{16, 32\})
LVSM [11]/▲X✓sparse (\{2, 4\})
Diffusion-based: image models
Zero123 [6]▲✓Xsparse (1)
ZeroNVS [13]▲✓Xsparse (1)
ReconFusion [7]▲✓Xsparse (3)
CAT3D [8]▲✓Xsparse ([1, 9])
Diffusion-based: video models
SV3D [18]▲X✓sparse (1)
MotionCtrl [10]▲X✓sparse (1)
ViewCrafter [9]▲X✓sparse (2)
4DiM [12]▲✓✓sparse (\{1, 2, 8\})
SEVA▲✓✓sparse ([1, 8]), semi-dense ([9, 32*])
+ +Table 1. Comparison of existing NVS models based on the source of training data and key attributes. SEVA is trained on both object-level (●) and scene-level (▲) data, offering flexible input conditioning, strong generation capacity, and smooth view interpolation. We define generation capacity and interpolation smoothness of each work based on their evaluation setting and our benchmark results. *This upper-bound can be up to hundreds for dense captures, we test our model up to 32 views in practice. + +typically involve hundreds of views. Based on the nature of their target views, we bucket a broad range of NVs tasks into "set NVS" and "trajectory NVS", as shown in Fig. 3. + +Set NVs considers a set of target views in arbitrary order, usually across a large spatial range. The order of views is often not helpful here, and a good NVS model requires great generation capacity to excel at this task. We note that some works address only this task (e.g. ReconFusion [7]). + +Trajectory NVs regards target views along a smooth camera trajectory, such that they form a video sequence. However, they are often sampled within a small spatial range in a shorter video. To solve this task, a good NVS model requires great interpolation smoothness to produce consistent and non-flickering results. We note that some existing works address only this task (e.g. ViewCrafter [9]). + +# 2.2. Existing Models + +We group existing approaches into regression- and diffusion-based models based on their high-level design choices. A more detailed discussion of related works can be found in the appendix. + +Regression-based models learn a deterministic mapping: + +$$ +f _ {\theta} \left(\mathbf {I} ^ {\text {i n p}}, \boldsymbol {\pi} ^ {\text {i n p}}, \boldsymbol {\pi} ^ {\text {t g t}}\right), +$$ + +![](images/d5b834cf8c302efb60c3cc213a069ca1a0f3c04730ee044608b07528d09ade78.jpg) +Figure 2. Diverse camera control. SEVA generates photorealistic novel views following diverse camera trajectories. This includes orbit, spiral, zoom out, dolly zooms, and any user-specified trajectories. Please visit our website for more visual results. + +to directly generate $\mathbf{I}^{\mathrm{tgt}}$ deterministically from $\mathbf{I}^{\mathrm{inp}}, \boldsymbol{\pi}^{\mathrm{inp}}, \boldsymbol{\pi}^{\mathrm{tgt}}$ . $f_{\theta}$ can be either an end-to-end network parameterized by $\theta$ , or a composition of a feed-forward prediction of an intermediate 3D representation and then a neural renderer (e.g., NeRF [19] or 3DGS [3]). For the latter case, set NVs and trajectory NVs are solved in the same way since there exists a persistent 3D representation. + +Diffusion-based models capture the conditional distribution: + +$$ +p _ {\theta} \left(\mathbf {I} ^ {\mathrm {t g t}} \mid \mathbf {I} ^ {\mathrm {i n p}}, \boldsymbol {\pi} ^ {\mathrm {i n p}}, \boldsymbol {\pi} ^ {\mathrm {t g t}}\right), +$$ + +from which $\mathbf{I}^{\mathrm{tgt}}$ are sampled [20] iteratively. We highlight two types of models within this scope: Image and Video models. Image models are trained on unordered image sets, such that $(\mathbf{I}^{\mathrm{inp}},\mathbf{I}^{\mathrm{tgt}})\sim \mathcal{I}$ , where $\mathcal{I} = \{\mathbf{I}_{\rho (1)},\mathbf{I}_{\rho (2)},\dots ,\mathbf{I}_{\rho (M + N)}\}$ is an image batch, and $\rho (\cdot)$ is a random permutation function (camera parameters are omitted for simplicity). Image models usually thrive at set NVs but struggle at trajectory NVS since they are designed to generate images and not videos. Moreover, the unordered nature of all views solicits flexible input conditioning. Video models are instead trained on ordered views, such that $(\mathbf{I}^{\mathrm{inp}},\mathbf{I}^{\mathrm{tgt}})\sim \mathcal{V}$ , where $\mathcal{V} = \{\mathbf{I}_1,\mathbf{I}_2,\dots ,\mathbf{I}_{M + N}\}$ + +![](images/4518ec28f6c75a7e52b1e70835d8bce7d8bba039f0fbb7a69b5b8a76f6374b69.jpg) +SetNVS + +![](images/d3b77916ee65256010f7877cfd0d879c34e6fca09a92b63d881f6cc0b5880459.jpg) +Trajectory NVS +Figure 3. Set NVS versus trajectory NVS. Set NVS generates target views as an image set, whereas trajectory NVS produces them as a trajectory video. + +![](images/ce0250ea6326add0094e34b1c625f739992a3a56a17efd99669ca3f462d11661.jpg) + +is a set of video frames with preserved ordering. Additional temporal operators may also be used to improve temporal smoothness, such as temporal positional encoding and temporal attention. In contrast with image models, video models thrive at trajectory NVS but struggle at set NVS. Moreover, all existing video models require both input and target views to be ordered (input views followed by target ones), constraining their input flexibility [10, 18, 21, 22]. + +# 3. Method + +It is challenging to develop a unified model that works for different NVS tasks while being flexible in input and output domains. Our design choices enable SEVA to achieve high generation capacity, smooth view interpolation, and flexi + +![](images/1bf2f170e5012697b960197a8454e34e52415c909a54c43e646dc0e6d9483f57.jpg) +Figure 4. Method. SEVA is trained with fixed sequence length as a “ $M$ -in $N$ -out” multi-view diffusion model with standard architecture. During sampling, SEVA can be cast as a generative “ $P$ -in $Q$ -out” renderer that works with variable sequence length, where $P$ and $Q$ need not be equal to $M$ and $N$ . To enhance temporal and 3D consistency across generated views, especially when generating along a trajectory, we present procedural two-pass sampling as a general strategy. + +![](images/801909587f252e64b3dd1adecbc53ed19c318cae5441d2991497717a92097bac.jpg) + +ble input conditioning (Tab. 1) in both set NVs as well as trajectory NVs. A system overview is provided in Fig. 4. + +# 3.1. Model Design and Training + +We consider an " $M$ -in $N$ -out" multi-view diffusion model $p_{\theta}$ , as noted in Sec. 2.2. We formulate this learning problem as a standard diffusion process [20] without any change. + +Architecture. Our model is based on the publicly available SD 2.1 [23], which consists of an auto-encoder and a latent denoising U-Net. Following [8], we inflate the 2D self-attention of each low-resolution residual block into 3D self-attention [24] within the U-Net. To improve model capacity, we add 1D self-attention along the view axis after each self-attention block via skip connection [25, 26], bumping the model parameters from 870M to 1.3B. Optionally, we further tame this model into a video model by introducing 3D convolutions after each residual block via skip connection, similar to [27, 28], yielding 1.5B total parameters. The temporal pathway can be enabled during inference when frames within one forward pass are known to be spatially ordered, enhancing output's smoothness. For the conditioning of our model, please refer to the appendix. + +Training. Let us define the training context window length $T = |\mathbf{I}^{\mathrm{inp}}| + |\mathbf{I}^{\mathrm{tgt}}| = M + N$ . One natural goal is to support large $T$ such that we can generate larger set of frames. However, we find that naive training is prone to divergence, and we thus employ a two-stage training curriculum. During the first stage, we train our model with $T = 8$ with a batch size of 1472 for $100K$ iterations. In the second stage, we train our model with $T = 21$ with a batch size of 512 for $600K$ iterations. Given a training video sequence, we randomly sample the number of input frames $M \in [1,T - 1]$ and the frames $(\mathbf{I}^{\mathrm{inp}},\mathbf{I}^{\mathrm{gt}})$ across the entire sequence to encourage diversity. We find it important to jointly sample with a smaller subsampling stride to ensure sufficient temporal granularity and avoid missing critical transitions with a small probability (0.2 is used in practice). In the optional video training stage, we only train + +temporal weights with data sampled with a small subsampling stride and a batch size of 512 for $200K$ iterations. We shift the signal-to-noise ratio (SNR) in all stages as more noise is necessary to remove the input information, corroborating findings from [8, 29, 30]. The model is trained with squared images with $H = W = 576$ . The training data is a superset of data used in [8] and derives from all public academic datasets, consisting of around $340K$ objects and $80K$ scenes. + +# 3.2. Sampling Novel Views + +Once the diffusion model is trained, we can sample it for a wide range of NVS tasks during test time. Formally, let us consider a " $P$ -in $Q$ -out" NVS task during testing, where we are given $P = |\mathbf{I}^{\mathrm{inp}}|$ input frames and aim to produce $Q = |\mathbf{I}^{\mathrm{tgt}}|$ target frames. Our goal is to design a generic sampling strategy that works for all $P$ and $Q$ configurations, where $P$ and $Q$ need not be equal to $M$ and $N$ . + +We make two key observations: First, within a single forward pass, predictions are 3D consistent, provided the model is well-trained. Second, when $P + Q > T$ , $\mathbf{I}^{\mathrm{gt}}$ must be split into smaller chunks of size $Q_{i}$ such that $P + Q_{i} \leqslant T$ for the $i^{\mathrm{th}}$ forward pass. We term this practice one-pass sampling. However, predictions across these forward passes would be inconsistent unless they share common frames to maintain local consistency within a spatial neighborhood. Building on these observations, we summarize our sampling process under two scenarios: $P + Q \leqslant T$ and $P + Q > T$ . + +$P + Q \leqslant T$ . We fit the task within one forward pass for simplicity and consistency. As shown in the appendix, we find it works better to pad the forward pass to have exactly $T$ frames by repeating the first input image, compared to changing the context window $T$ zero-shot. + +$P + Q > T$ . We propose procedural two-pass sampling: In the first pass, we generate anchor frames $\mathbf{I}^{\mathrm{acr}}$ using all input frames $\mathbf{I}^{\mathrm{inp}}$ . In the second pass, we divide $\mathbf{I}^{\mathrm{tgt}}$ into + +chunks and generate them using $\mathbf{I}^{\mathrm{acr}}$ (and optionally $\mathbf{I}^{\mathrm{inp}}$ ) according to the spatial distribution of $\mathbf{I}^{\mathrm{acr}}$ and $\mathbf{I}^{\mathrm{gt}}$ . + +Given the distinct nature of the two tasks of interest—set NVS and trajectory NVS—e.g., differences in the availability of views' ordering, we design tailored chunking strategies for each task. + +For set NVS, we consider nearest procedural sampling. We first generate $\mathbf{I}^{\mathrm{acr}}$ based on pre-defined trajectory priors, similar to [8], e.g., 360 trajectories for object-centric scenes or spiral trajectories for forward-facing scenes. We then divide $\mathbf{I}^{\mathrm{gt}}$ into chunks w.r.t. $\mathbf{I}^{\mathrm{acr}}$ using nearest neighbor. Specifically, the $i^{\mathrm{th}}$ forward pass involves: + +$$ +\operatorname {n e a r e s t}: \left\{\mathbf {I} _ {i} ^ {\mathrm {a c r}} \right\} \cup \left\{\mathbf {I} _ {j} ^ {\mathrm {t g t}} \mid \mathrm {N N} (\mathbf {I} _ {j} ^ {\mathrm {t g t}}, \mathbf {I} ^ {\mathrm {a c r}}) = \mathbf {I} _ {i} ^ {\mathrm {a c r}} \right\}. +$$ + +We considered two strategies of procedural sampling: nearest as described above, and $\mathrm{gt} +$ nearest strategy by appending $\mathbf{I}^{\mathrm{inp}}$ into each forward pass. We find that the $\mathrm{gt} +$ nearest strategy performs better than nearest, and thus default to it instead. In the absence of trajectory priors, we revert to one-pass sampling. In practice, employing nearest anchors enhances qualitative consistency, albeit on a limited scale. + +For trajectory NVS, we consider interp procedural sampling. We first generate a subset of target frames as $\mathbf{I}^{\mathrm{acr}}$ by uniformly spanning the target camera path with a stride $\Delta = \left\lfloor \frac{Q}{T - 2}\right\rfloor$ . We then generate the rest of $\mathbf{I}^{\mathrm{tgt}}$ as segments between those anchors: + +$$ +\mathrm {i n t e r p}: \{\mathbf {I} _ {i} ^ {\mathrm {a c r}}, \mathbf {I} _ {i \cdot \Delta + 1} ^ {\mathrm {t g t}}, \dots , \mathbf {I} _ {(i + 1) \cdot \Delta - 1} ^ {\mathrm {t g t}}, \mathbf {I} _ {i + 1} ^ {\mathrm {a c r}} \}. +$$ + +Since the input to the model is ordered, we can leverage temporal weights to further improve smoothness (Sec. 4.3). Similarly, $\mathrm{gt} + \mathrm{interp}$ is possible by appending $\mathbf{I}^{\mathrm{inp}}$ with $\Delta = \left\lfloor \frac{Q}{T - P - 2} \right\rfloor$ . We find that interp is sufficiently robust, and choose it as the default option. The interp strategy drastically outperforms its counterparts (e.g., one-pass, or $\mathrm{gt} + \mathrm{nearest}$ procedural sampling) in terms of temporal smoothness. + +# 3.3. Scaling Sampling for Large $P$ and $Q$ + +Next, we examine two special cases when $P + Q > T$ : $P > T$ and $Q \gg T$ . Here, we make a tailored design for anchor generation in the first pass, while keeping target generation in the second pass unchanged. + +$P > T$ . In the semi-dense-view regime (e.g., $P = 32$ ), we extend the context window length $T$ zero-shot to accommodate all $P$ input views and anchor views in one pass during anchor generation. Empirically, $T$ can even be extended up to hundreds without severe degradation in photorealism in the generated outputs. We find that the diffusion model generalizes well in this case as long as the input views cover the majority of the scene, shifting the task from generation to primarily interpolation. In the sparse-view + +![](images/a43cdb8665181349f6de4fcef3de81638119f70716dba8864ecbf09cb5e37967.jpg) +Figure 5. Anchor generation when $Q \gg T$ . We introduce a memory bank composed of previously-generated anchor views and their corresponding camera poses. The lookup of spatial neighbors helps improve long-term 3D consistency. + +regime $(P \leq 8)$ , we observe similar performance degradation caused by zero-shot extension of $T$ compared to what we have found when $P + Q \leqslant T$ . Refer to the appendix for a detailed discussion. + +$Q \gg T$ . When the number of target views $Q$ is large, e.g., in large-set NVS or long-trajectory NVS, even anchors will be chunked into different forwards in the first pass, leading to the inconsistency of anchors. To this end, we maintain a memory bank of anchor views previously generated, as shown in Fig. 5. We generate anchors autoregressively by retrieving their spatially nearest ones from the memory bank, similar to the nearest strategy introduced above for the second pass. In Sec. 4.4, we show that this strategy drastically outperforms the standard practice of reusing temporally nearest anchors previously generated in long video literature [22], in terms of long-range 3D consistency, especially for hard trajectories. + +# 4. Experiments + +We employ a single model for a spectrum of settings and find that SEVA model generalizes well under the three criteria (Tab. 1). We cover different NVS tasks (set NVS and trajectory NVS) and examine one special task of interest—long trajectory NVS. We also cover different input regimes (single-view, sparse-view, and semi-dense-view). A discussion about key properties is presented in the appendix. + +# 4.1. Benchmark + +Datasets, splits, and the number of input views. We collect 10 commonly used datasets to benchmark NVS, encompassing a diverse range of scene distributions and complexities. We consider (1) object datasets [32, 33]; (2) object-centric scene datasets [34-38]; and (3) scene datasets [39-41]. To establish a comprehensive and rigorous comparison with baselines, we consider different dataset splits utilized in prior works with the same input-view configuration, unless specified as our split (O). These include splits used in 4DiM [12] (D), ViewCrafter [9] (V), + +![](images/a4b8491584ad3d6abe80aa4aa7086eb632e6d09a48470af878ef5e46c8e001c0.jpg) +Figure 6. SOTA comparison on set NVs (top) and trajectory NVs (bottom) across varying numbers of input views. From left to right and top to bottom, inputs views are (top, i.e. set NVS) 1, 2, 3, 32; (bottom, i.e. trajectory NVS) 1, 2, 1, 1, 3, 3. For cases with multiple input views, the view closest to the target is shown over the rest. We compare with open-source approaches—ViewCrafter [9] (VC) and DepthSplat [31] (DS)—as well as proprietary ones including LVSM [11], Long-LRM [17] (LLRM), 4DiM [12], and CAT3D [8]. + +pixelSplat [15] (P), ReconFusion [7] (R), SV3D [18] (S), and Long-LRM [17] (L). + +Small-viewpoint versus large-viewpoint NVS. Sweeping across all datasets, splits, and input-view configurations reveals a diverse benchmark of setups. To better evaluate models' generation capacity and interpolation smoothness (Sec. 2.1), we propose to categorize this setup into two groups—small-viewpoint NVS and large-viewpoint NVS—depending on the disparity between $\mathbf{I}^{\mathrm{tgt}}$ and $\mathbf{I}^{\mathrm{inp}}$ . + +Baselines. We consider a range of proprietary models, including ReconFusion [7], CAT3D [8], 4DiM [12], LVSM [11], and Long-LRM [17]. We also consider various open-source models, including SV3D [18], MVSplat [16], depthSplat [31], MotionCtrl [42], and ViewCrafter [9]. As outlined in Sec. 2.2, these baselines encompass both regression-based and diffusion-based approaches, providing a comprehensive framework for comparison. Please refer to the appendix for more details of the benchmark. + +# 4.2. Set NVs + +Quantitative comparison. The input and target views are chosen following the splits used in previous methods. The order of target views is not preserved. We use standard metrics of peak signal-to-noise ratio (PSNR), structural simi + +larity index measure [44] (SSIM), and learned perceptual image patch similarity [45] (LPIPS). Only PSNR is showcased here due to space limits, with the rest deferred to the appendix. Empirically, our method shows a greater performance improvement on LPIPS, reflecting the photorealism of our results. + +For small-viewpoint set NVs, Tab. 2 shows that SEVA sets state-of-the-art results in the majority of splits. In the sparse-view regime (i.e., $P \leqslant 8$ ), SEVA excels across different datasets when $P > 1$ . For example, a performance gain of +6.0 dB PSNR is achieved on LLFF with $P = 3$ . In the semi-dense-view regime (e.g., $P = 32$ ), SEVA surprisingly performs favorably against the specialized model [17], despite not being specifically designed for this setup. For example, SEVA lags behind the state-of-the-art method [17] by only 1.7 dB on T&T. A detailed discussion on the RE10K dataset can be found in the appendix. On object datasets OO3D and GSO, SEVA achieves a significantly higher state-of-the-art PSNR compared to all other methods. + +For large-viewpoint set NVs, Tab. 3 shows that SEVA's quantitative advantages are even more prominent here, revealing clear benefits of SEVA in terms of generation capacity when the camera spans a large spatial range. On Mip360 with $P = 3$ , SEVA improves over previous state-of-the-art method CAT3D [8] by $+0.6$ dB PSNR. On harder scenes like DL3DV and T&T with different input-view configurations, SEVA obtains a clear performance lead. + +
Methoddataset split POO3D OGSO ORE10KLLFF R [7] 1 3DTU R [7] 1 3CO3D V [9] R [7] 1 3WRGBD Oe 3 6Mip360 O L [17] 6 32DL3DV T & T V [9] L [17] 1 32
O 3O 3D [12] 1V [9] 1P [15] 2R [7] 13
Regression-based models
Long-LRM [17]-------------23.86 - 18.20
MVSplat [16]14.78 15.6715.21 16.5220.42 20.9020.32 19.2426.39 27.4421.56 21.8725.64 22.5411.23 12.0712.50 12.6213.87 15.52 14.1513.52 13.52 13.2314.56 12.54 15.9313.56 14.24 14.2313.24 13.22 14.35
DepthSplat [31]
LVSM [11]----29.67---------
Diffusion-based models
MotionCtrl [42]--12.74 17.0816.29 ------15.46 ----13.29 -
4DiM [12]--17.08-----------
ViewCrafter [9]14.64 30.3015.93 31.5320.43 17.9922.04 18.5621.42 25.6620.88 18.1122.81 27.5710.53 14.0313.52 14.4712.66 16.40 18.4718.96 14.72 18.4016.42 12.66 19.7514.59 13.78 16.70- 18.07 - 15.16
SEVA
+ +Table 2. PSNR† on small-viewpoint set NVS. $P$ denotes the number of input views. For all results with $P = 1$ , we sweep the unit length for camera normalization due to the model's scale ambiguity. $\mathrm{O}_e$ and $\mathrm{O}_h$ denote the easy and hard split of our split, respectively. Underlined numbers are run by us using the official released code. + +
Methoddataset split POO3D S [18] 1GSO S [18] 1CO3D R [7] 1WRGBD Oh 1Mip360 R [7] 3DL3DV O 1T&T O 31369
SV3D [18]19.2820.38----------
DepthSplat [31]11.5612.3210.429.3513.5310.4912.549.6312.528.639.7810.12
CAT3D [8]------15.15-----
ViewCrafter [9]10.5611.4210.119.1213.459.7910.348.9711.509.239.8810.32
SEVA19.2520.6515.3014.3717.2812.9315.7813.0115.9511.2812.6513.80
+ +
Methodsmall-viewpointlarge-viewpoint
RE10KLLFFDTUCO3DMip360
ZipNeRF [43]20.7717.239.1814.3412.77
ZeroNVS [13]19.1115.9116.7117.1314.44
ReconFusion [7]25.8421.3420.7419.5915.50
CAT3D [8]26.7821.5822.0220.5716.62
SEVA27.9521.8822.6821.8817.82
+ +Table 3. PSNR† on large-viewpoint set NVs. For all results with $P = 1$ , we sweep the Table 4. PSNR† on 3DGS renderings for set unit length for camera normalization due to the model's scale ambiguity. Underlined NVS. Results are reported on the ReconFusion [7] numbers are run by us using the officially released code. split with $P = 3$ . + +Comparison of 3D reconstruction. To enable a direct quantitative comparison with prior works [7, 8], we adopt the few-view 3D reconstruction pipeline described in [8]. We optimize the camera parameters and apply LPIPS loss [46] during the distillation. Finally, we render the distilled 3D model on the test views and report the performance in Tab. 4. SEVA shows a consistent performance lead. + +Qualitative comparison. Fig. 6 top panel shows a qualitative comparison with diverse baselines. For small-viewpoint set NVs, the output from SEVA with the best scale exhibits desirable alignment with the ground truth while being more photorealistic in details. Compared with LVSM [11] on the P split of RE10K, SEVA produces sharper images, also corroborating that lower PSNR arises from scale ambiguity rather than interpolation quality. Similar trends hold when compared to Long-LRM [17] on DL3DV with $P = 32$ . For large-viewpoint set NVs, we compare with DepthSplat [31] on DL3DV with $P = 3$ . DepthSplat fails to produce reasonable results when the viewpoint change is too large and falls short in overall visual quality. + +# 4.3. Trajectory NVS + +Qualitative results. Fig. 2 presents qualitative results, illustrating trajectories of varying complexities with different numbers of input views across diverse types. + +In the single-view regime (i.e., $P = 1$ ), we manually craft a set of common camera movements/effects, e.g., look-at 360, spiral, panning, zoom-in, zoom-out, dolly zoom, etc. We observe that SEVA generalizes to a wide range of images and demonstrates accurate camera-following capacity. An extensive sweeping of camera movements on 4 types of images is provided in the appendix. + +In the sparse-view regime with few input views (i.e., $1 < P \leqslant 8$ ), we observe that SEVA demonstrates strong generalization to in-the-wild real-world images and versatility in adapting to different numbers of input views. The output forms a smooth trajectory video with subtle temporal flickering, revealing its capacity to interpolate between views smoothly. In the last row of Fig. 2, our model generates plausible results at the end of the trajectory—an area unseen in the input observations—demonstrating its strong generation capacity. In the semi-dense-view regime (i.e., $P > 9$ ), we similarly find that SEVA is surprisingly able to produce a smooth trajectory video with minimal artifacts. Please check the supplementary website for video results. + +Qualitative comparison. Fig. 6 bottom panel presents a qualitative comparison with diverse baselines. In the single-view regime (i.e., $P = 1$ ), we compare to 4DiM [12] and CAT3D [8]. We observe more photo-realistic and sharper + +
Methodsplit datasetsmall-viewpointlarge-viewpoint
V [9]O
RECO3DT&TREDTUWRDLT&T
MotionCtrl [42]16.2915.4613.29------
DepthSplat [31]19.2413.2314.3525.2314.6812.4511.329.11
ViewCrafter [9]22.0418.9618.0726.5418.9913.4411.459.68
SEVA18.5618.4015.1627.3419.9917.7915.7611.92
SEVA (+ temp.)18.6218.4315.1327.3620.1917.9315.7811.99
+ +Table 5. PSNR $\uparrow$ on trajectory NVS. temp. denotes optional temporal pathway. RE, WR, and DL denotes RE10K, WRGBD, and DL3DV, respectively. For the V [9] split, $P = 1$ with unit length swept; for the O split, $P = 3$ . Underlined numbers are run by us using the officially released code. + +![](images/7429eaa8fd38800a41e78b4526e383df46bbff51fb6a7508b0cdc85b3a1575dd.jpg) +Figure 7. Temporal quality. Vertical slices of a rendered novel camera path on the BONSAI scene from Mip-NeRF360 [38] illustrate the temporal quality across adjacent viewpoints. One-pass or gt + nearest procedural sampling results in notable flickering, whereas interp procedural sampling ensures temporally smooth rendering. + +output from our model, especially in the background area for object-centric scenes. 4DiM outputs tend to be cartoonish and over-simplistic, given that the model is only trained on RE10K. In the sparse-view regime with few input views (i.e., $P = 3$ ), we compare with CAT3D and observe that our model demonstrates more photo-realistic textures, especially in the background. + +Quantitative comparison. We use the same input views as in set NVS for each split. We use all frames from each scene as target views such that they form a smoothly transitioning trajectory video, i.e., $\mathbf{I}^{\mathrm{gt}}\sim \mathcal{V}$ . For small-viewpoint trajectory NVS, Tab. 5 compares SEVA with baselines on PSNR. SEVA performs favorably against other methods in V split with $P = 1$ . The performance lead of ViewCrafter is mainly attributed to its training on high-resolution images. For large-viewpoint trajectory NVS with $P = 3$ , SEVA consistency sets new state-of-the-art results. Applying the temporal pathway further boosts performance and improves smoothness, indicating the benefits of the gated + +![](images/1116035c33588665ed1055470f28fc4ffbe5e890116e24c52949be06a628600d.jpg) +Figure 8. Long-range 3D consistency. We visualize samples following a camera path looping three times around the TELEPHONE-BOOTH scene. Lookup using spatial neighbors from the memory bank (ours) notably improves view consistency and reduces artifacts in recurring locations across different loops, compared to lookup using temporal neighbors (baseline). + +architecture. + +Ablation on two-pass procedural sampling. Fig. 7 shows that one-pass sampling introduces visible temporal flickering and abrupt scene changes. In contrast, interp procedural sampling produces the smoothest transitions, outperforming the alternative $\mathrm{gt} +$ nearest strategy and mitigating noticeable flickering. Additional quantitative ablation is deferred to the appendix, including 3DGS distillation and video quality metrics. + +# 4.4. Long-Trajectory NVS + +Fig. 8 presents a qualitative demonstration of NVS over a long trajectory of up to 1000 frames. With the memory bank maintaining previously generated anchors, SEVA achieves robust 3D consistency for long-trajectory NVS, e.g., the building in front of and the plantation after the booth. Comparing it to using temporal nearest anchors previously generated, using spatially nearest ones demonstrates a clear advantage in terms of long-term consistency. + +# 5. Conclusion + +We introduce SEVA, a generalist diffusion model for novel view synthesis that balances large viewpoint changes and smooth interpolation while supporting flexible input and target configurations. By designing a diffusion-based architecture without 3D representation, a structured training strategy, and a two-pass procedural sampling approach, SEVA achieves 3D consistent rendering across diverse NVs tasks. Extensive benchmarking demonstrates its superiority over existing methods, with strong generalization to real-world scenes. For broader impact and limitations, please refer to the appendix. + +# References + +[1] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, pages 405-421, 2020. 1 +[2] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 2022. +[3] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4), July 2023. 1, 3 +[4] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In CVPR, 2021. 1, 2 +[5] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3D using 2D diffusion. arXiv, 2022. 1 +[6] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3D object. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 2 +[7] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21551-21561, 2024. 1, 2, 6, 7 +[8] Ruiqi Gao*, Aleksander Holynski*, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul P. Srinivasan, Jonathan T. Barron, and Ben Poole*. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv, 2024. 1, 2, 4, 5, 6, 7 +[9] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. ViewCrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 1, 2, 5, 6, 7, 8 +[10] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1, 2, 3 +[11] Haian Jin, Hanwen Jiang, Hao Tan, Kai Zhang, Sai Bi, Tianyuan Zhang, Fujun Luan, Noah Snavely, and Zexiang Xu. Lvsm: A large view synthesis model with minimal 3d inductive bias, 2024. 1, 2, 6, 7 +[12] Daniel Watson, Saurabh Saxena, Lala Li, Andrea Tagliasacchi, and David J Fleet. Controlling space and time with diffusion models. arXiv preprint arXiv:2407.07860, 2024. 1, 2, 5, 6, 7 +[13] Kyle Sargent, Zizhang Li, Tanmay Shah, Charles Herrmann, Hong-Xing Yu, Yunzhi Zhang, Eric Ryan Chan, Dmitry Lagun, Li Fei-Fei, Deqing Sun, et al. Zeronvs: Zero-shot + +360-degree view synthesis from a single real image. arXiv preprint arXiv:2310.17994, 2023. 1, 2, 7 +[14] Virtual camera system. https://en.wikipedia.org/wiki/Virtual_camera_system. Accessed: 5-Mar-2025. 1 +[15] David Charatan, Sizhe Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In arXiv, 2023. 2, 6, 7 +[16] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In European Conference on Computer Vision, pages 370-386. Springer, 2025. 2, 6, 7 +[17] Chen Ziwen, Hao Tan, Kai Zhang, Sai Bi, Fujun Luan, Yicong Hong, Li Fuxin, and Zexiang Xu. Long-lrm: Long-sequence large reconstruction model for wide-coverage gaussian splats. arXiv preprint arXiv:2410.12781, 2024. 2, 6, 7 +[18] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitrii Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. SV3D: Novel multi-view synthesis and 3D generation from a single image using latent video diffusion. In European Conference on Computer Vision, 2024. 2, 3, 6, 7 +[19] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 3 +[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3, 4 +[21] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrl: Enabling camera control for text-to-video generation. arXiv preprint arXiv:2404.02101, 2024. 3 +[22] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. arXiv, 2022. 3, 5 +[23] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, June 2022. 4 +[24] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. MVDream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 4 +[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4 +[26] P Goyal. Accurate, large minibatch sg d: training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 4 +[27] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Content + +ference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 4 +[28] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 4 +[29] Emiel Hoogeboom, Jonathan Heek, and Tim Salimans. simple diffusion: End-to-end diffusion for high resolution images. arXiv preprint arXiv:2301.11093, 2023. 4 +[30] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICLR, 2024. 4 +[31] Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. Depthsplat: Connecting gaussian splatting and depth. arXiv preprint arXiv:2410.13862, 2024. 6, 7, 8 +[32] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, et al. Omniobject3D: Large-vocabulary 3D object dataset for realistic perception, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 803-814, 2023. 5 +[33] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google Scanned Objects: A high-quality dataset of 3D scanned household items. In 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560. IEEE, 2022. 5 +[34] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (ToG), 38(4):1-14, 2019. 5 +[35] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 406-413, 2014. +[36] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In International Conference on Computer Vision, 2021. +[37] Hongchi Xia, Yang Fu, Sifei Liu, and Xiaolong Wang. RGBd objects in the wild: Scaling real-world 3d object learning from rgb-d videos, 2024. +[38] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-NeRF: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 5, 8 +[39] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018. 5 + +[40] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024. +[41] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 5 +[42] Tsai-Shien Chen, Chieh Hubert Lin, Hung-Yu Tseng, Tsung-Yi Lin, and Ming-Hsuan Yang. Motion-conditioned diffusion model for controllable video synthesis. arXiv preprint arXiv:2304.14404, 2023. 6, 7, 8 +[43] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 7 +[44] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6 +[45] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6 +[46] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric, 2018. 7 \ No newline at end of file diff --git a/2025/Stable Virtual Camera_ Generative View Synthesis with Diffusion Models/images.zip b/2025/Stable Virtual Camera_ Generative View Synthesis with Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6661436232402e9002483663837a797656490808 --- /dev/null +++ b/2025/Stable Virtual Camera_ Generative View Synthesis with Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6c730a8f07c20a0a61175110bfbab779f5a4c7f4eee2944023bac46400eb6687 +size 1038704 diff --git a/2025/Stable Virtual Camera_ Generative View Synthesis with Diffusion Models/layout.json b/2025/Stable Virtual Camera_ Generative View Synthesis with Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c511b69ded4e5a1837bedb2442e4f51f2b9a642e --- /dev/null +++ b/2025/Stable Virtual Camera_ Generative View Synthesis with Diffusion Models/layout.json @@ -0,0 +1,8893 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 103, + 548, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 103, + 548, + 120 + ], + "spans": [ + { + "bbox": [ + 59, + 103, + 548, + 120 + ], + "type": "text", + "content": "STABLE VIRTUAL CAMERA: Generative View Synthesis with Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 129, + 128, + 479, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 128, + 479, + 174 + ], + "spans": [ + { + "bbox": [ + 129, + 128, + 479, + 174 + ], + "type": "text", + "content": "Jensen (Jinghao) Zhou" + }, + { + "bbox": [ + 129, + 128, + 479, + 174 + ], + "type": "inline_equation", + "content": "^{1,2,*,\\dagger}" + }, + { + "bbox": [ + 129, + 128, + 479, + 174 + ], + "type": "text", + "content": " Hang Gao" + }, + { + "bbox": [ + 129, + 128, + 479, + 174 + ], + "type": "inline_equation", + "content": "^{1,3,*,\\dagger}" + }, + { + "bbox": [ + 129, + 128, + 479, + 174 + ], + "type": "text", + "content": " \nVikram Voleti" + }, + { + "bbox": [ + 129, + 128, + 479, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 129, + 128, + 479, + 174 + ], + "type": "text", + "content": " Aaryaman Vasishta" + }, + { + "bbox": [ + 129, + 128, + 479, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 129, + 128, + 479, + 174 + ], + "type": "text", + "content": " Chun-Han Yao" + }, + { + "bbox": [ + 129, + 128, + 479, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 129, + 128, + 479, + 174 + ], + "type": "text", + "content": " Mark Boss" + }, + { + "bbox": [ + 129, + 128, + 479, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 129, + 128, + 479, + 174 + ], + "type": "text", + "content": " Philip Torr" + }, + { + "bbox": [ + 129, + 128, + 479, + 174 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 129, + 128, + 479, + 174 + ], + "type": "text", + "content": " Christian Rupprecht" + }, + { + "bbox": [ + 129, + 128, + 479, + 174 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 129, + 128, + 479, + 174 + ], + "type": "text", + "content": " Varun Jampani" + }, + { + "bbox": [ + 129, + 128, + 479, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 121, + 178, + 488, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 178, + 488, + 194 + ], + "spans": [ + { + "bbox": [ + 121, + 178, + 488, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 121, + 178, + 488, + 194 + ], + "type": "text", + "content": "Stability AI " + }, + { + "bbox": [ + 121, + 178, + 488, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 121, + 178, + 488, + 194 + ], + "type": "text", + "content": "University of Oxford " + }, + { + "bbox": [ + 121, + 178, + 488, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 121, + 178, + 488, + 194 + ], + "type": "text", + "content": "University of California, Berkeley" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 230, + 200, + 378, + 215 + ], + "blocks": [ + { + "bbox": [ + 230, + 200, + 378, + 215 + ], + "lines": [ + { + "bbox": [ + 230, + 200, + 378, + 215 + ], + "spans": [ + { + "bbox": [ + 230, + 200, + 378, + 215 + ], + "type": "image", + "image_path": "7972dd8723f41d3ffe2accb551f716bcfa11ff6f527fc96112d36828f2b1c496.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 253, + 296, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 253, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 296, + 446 + ], + "type": "text", + "content": "We present STABLE VIRTUAL CAMERA (SEVA), a generalist diffusion model that creates novel views of a scene, given any number of input views and target cameras. Existing works struggle to generate either large viewpoint changes or temporally smooth samples, while relying on specific task configurations. Our approach overcomes these limitations through simple model design, optimized training recipe, and flexible sampling strategy that generalize across view synthesis tasks at test time. As a result, our samples maintain high consistency without requiring additional 3D representation-based distillation, thus streamlining view synthesis in the wild. Furthermore, we show that our method can generate high-quality videos lasting up to half a minute with seamless loop closure. Extensive benchmarking demonstrates that SEVA outperforms existing methods across different datasets and settings." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 463, + 135, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 135, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 135, + 475 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 484, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 296, + 628 + ], + "type": "text", + "content": "Novel view synthesis (NVS) aims to generate realistic, 3D-consistent images of a scene from arbitrary camera viewpoints given any number of camera-posed input views. Traditional methods, which rely on dense input views, treat NVS as a 3D reconstruction and rendering problem [1-3], but this approach fails with sparse inputs. Generative view synthesis addresses this limitation by leveraging modern deep network priors [4, 5], enabling immersive 3D interactions in uncontrolled environments without the need to capture large image sets per scene. In this work, we focus on generative view synthesis and, unless otherwise specified, refer to it simply as NVS for clarity." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": "Despite recent progress [6-12], NVS in the wild remains limited due to two key challenges: First, existing methods struggle to generate both large viewpoint changes [9, 10] and temporally smooth samples [6-8, 13] while being constrained by rigid task configurations, such as a fixed number of input and target views [7, 9, 11, 12], reviewed in Tab. 1. Second, their sampling consistency is often insufficient, ne" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 315, + 225, + 555, + 390 + ], + "blocks": [ + { + "bbox": [ + 315, + 225, + 555, + 390 + ], + "lines": [ + { + "bbox": [ + 315, + 225, + 555, + 390 + ], + "spans": [ + { + "bbox": [ + 315, + 225, + 555, + 390 + ], + "type": "image", + "image_path": "fb92f49ac881fc46a4d8f22a85a7508e73995c0848ced9210adf4443401260a8.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 396, + 555, + 472 + ], + "lines": [ + { + "bbox": [ + 313, + 396, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 555, + 472 + ], + "type": "text", + "content": "Figure 1. Generative view synthesis. STABLE VIRTUAL CAMERA generates novel views from any number of input views and target cameras, which the user can specify anywhere. We show three examples: single view with simple orbit camera trajectory (top); two views with long camera trajectory (middle); and nine views with large spatial range (bottom). Please refer to the supplementary webpage and video for the video results." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 479, + 554, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 554, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 554, + 525 + ], + "type": "text", + "content": "cessitating additional NeRF distillation to fuse inconsistent results into a coherent representation [7, 8, 13]. These limitations hinder their applicability across diverse NVS tasks, which we address in this work." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 528, + 556, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 556, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 556, + 673 + ], + "type": "text", + "content": "We present _STABLE_VIRTUAL CAMERA1 (SEVA), a diffusion-based NVS model that generalizes across a spectrum of view synthesis tasks without requiring NeRF distillation. With a single network, SEVA generates high-quality novel views that strike both large viewpoint changes and temporal smoothness while supporting any number of input and target views. Our approach simplifies the NVS pipeline without requiring distillation from a 3D representation, thus streamlining it for real-world applications. For the first time, we demonstrate high-quality videos lasting up to half a minute with precise camera control and seamless loop closure in 3D. We highlight these results in Fig. 1 and" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 313, + 683, + 555, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 683, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 683, + 555, + 712 + ], + "type": "text", + "content": "1We are naming this model in tribute to the Virtual Camera [14] cinematography technology, a pre-visualization technique to simulate real-world camera movements." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12405" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 271, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 271, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 271, + 83 + ], + "type": "text", + "content": "showcase more examples of camera control in Fig. 2." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 84, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 84, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 54, + 84, + 294, + 251 + ], + "type": "text", + "content": "To achieve this, we carefully design our pipeline in three key aspects: model design, training recipe, and sampling method at inference. First, SEVA avoids explicit 3D representations within the network, allowing the model to inherit strong priors from pre-trained 2D models. Second, during training, we carefully craft our view selection strategy to cover both small and large viewpoint changes, ensuring strong generalization to diverse NVS tasks. Third, at inference, we introduce a two-pass procedural sampling approach that supports flexible input-target configurations. Together, these design choices create a versatile 3D \"virtual camera simulation system\" capable of synthesizing novel views along arbitrary camera trajectories with any number of input and target views, without using a 3D representation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 251, + 294, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 251, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 294, + 358 + ], + "type": "text", + "content": "We conducted a unified benchmark across 10 datasets and a variety of experimental settings, including both open-source and proprietary models. It reflects the diversity of real-world NVS tasks across the board and systematically evaluates existing methods beyond their comfort zones. We find that SEVA consistently outperforms previous works, achieving " + }, + { + "bbox": [ + 54, + 251, + 294, + 358 + ], + "type": "inline_equation", + "content": "+1.5" + }, + { + "bbox": [ + 54, + 251, + 294, + 358 + ], + "type": "text", + "content": " dB PSNR over the state of the art CAT3D [8]. Our method generalizes well to in-the-wild user captures, with input views ranging from 1 to 32." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 359, + 295, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 359, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 54, + 359, + 295, + 455 + ], + "type": "text", + "content": "In summary, our key contributions with the SEVA model include: (1) a training strategy for jointly modeling large viewpoint changes and temporal smoothness, (2) a two-pass procedural sampling method for smooth video generation along arbitrary long camera trajectories, (3) a comprehensive benchmark that evaluates NVS methods across different datasets and settings, and (4) an open-source release of model weights to support future research." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 465, + 134, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 465, + 134, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 465, + 134, + 479 + ], + "type": "text", + "content": "2. Background" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 485, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 485, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 485, + 295, + 581 + ], + "type": "text", + "content": "We consider the evaluation of an NVS model across three key criteria: (1) generation capacity—the ability to synthesize missing regions for large viewpoint changes; (2) interpolation smoothness—the ability to produce seamless transitions between views; and (3) input flexibility—the ability to handle a variable number of input and target views. We review existing NVS models based on these criteria in Tab. 1, including the types of training data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 588, + 171, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 171, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 171, + 601 + ], + "type": "text", + "content": "2.1. Types of NVS Tasks" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 604, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 295, + 714 + ], + "type": "text", + "content": "Given " + }, + { + "bbox": [ + 55, + 604, + 295, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 604, + 295, + 714 + ], + "type": "text", + "content": " input view images " + }, + { + "bbox": [ + 55, + 604, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{inp}}\\in \\mathbb{R}^{M\\times H\\times W\\times 3}" + }, + { + "bbox": [ + 55, + 604, + 295, + 714 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 604, + 295, + 714 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 55, + 604, + 295, + 714 + ], + "type": "text", + "content": " resolution, along with their corresponding cameras " + }, + { + "bbox": [ + 55, + 604, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\pi^{\\mathrm{inp}}" + }, + { + "bbox": [ + 55, + 604, + 295, + 714 + ], + "type": "text", + "content": ", NVS involves predicting " + }, + { + "bbox": [ + 55, + 604, + 295, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 604, + 295, + 714 + ], + "type": "text", + "content": " targets views " + }, + { + "bbox": [ + 55, + 604, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{tgt}}\\in \\mathbb{R}^{N\\times H\\times W\\times 3}" + }, + { + "bbox": [ + 55, + 604, + 295, + 714 + ], + "type": "text", + "content": ", specified by their respective cameras " + }, + { + "bbox": [ + 55, + 604, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\pi^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 55, + 604, + 295, + 714 + ], + "type": "text", + "content": ". For each camera, we assume we know both intrinsics and extrinsics. Based on the number of input views, we define the \"sparse-view regime\" as having up to 8 input views, and the \"semi-dense-view regime\" as an intermediate state bridging the sparse-view regime and dense captures, which" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 316, + 70, + 553, + 292 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 553, + 292 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 553, + 292 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 553, + 292 + ], + "type": "table", + "html": "
modeltraining datageneration capacityinterpolation smoothnessinput flexibility
Regression-based
pixelNeRF [4]▲X✓sparse (1)
pixelSplat [15]▲X✓sparse (2)
MVSplat [16]▲X✓sparse (2)
Long-LRM [17]▲X✓semi-dense (\\{16, 32\\})
LVSM [11]/▲X✓sparse (\\{2, 4\\})
Diffusion-based: image models
Zero123 [6]▲✓Xsparse (1)
ZeroNVS [13]▲✓Xsparse (1)
ReconFusion [7]▲✓Xsparse (3)
CAT3D [8]▲✓Xsparse ([1, 9])
Diffusion-based: video models
SV3D [18]▲X✓sparse (1)
MotionCtrl [10]▲X✓sparse (1)
ViewCrafter [9]▲X✓sparse (2)
4DiM [12]▲✓✓sparse (\\{1, 2, 8\\})
SEVA▲✓✓sparse ([1, 8]), semi-dense ([9, 32*])
", + "image_path": "127edfaca72ede73b9e414e914fa8866af67936bbd80405eb52b0938d618241b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 297, + 555, + 386 + ], + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 386 + ], + "type": "text", + "content": "Table 1. Comparison of existing NVS models based on the source of training data and key attributes. SEVA is trained on both object-level (●) and scene-level (▲) data, offering flexible input conditioning, strong generation capacity, and smooth view interpolation. We define generation capacity and interpolation smoothness of each work based on their evaluation setting and our benchmark results. *This upper-bound can be up to hundreds for dense captures, we test our model up to 32 views in practice." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 394, + 555, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 555, + 430 + ], + "type": "text", + "content": "typically involve hundreds of views. Based on the nature of their target views, we bucket a broad range of NVs tasks into \"set NVS\" and \"trajectory NVS\", as shown in Fig. 3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 438, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 554, + 498 + ], + "type": "text", + "content": "Set NVs considers a set of target views in arbitrary order, usually across a large spatial range. The order of views is often not helpful here, and a good NVS model requires great generation capacity to excel at this task. We note that some works address only this task (e.g. ReconFusion [7])." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 506, + 554, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 554, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 554, + 591 + ], + "type": "text", + "content": "Trajectory NVs regards target views along a smooth camera trajectory, such that they form a video sequence. However, they are often sampled within a small spatial range in a shorter video. To solve this task, a good NVS model requires great interpolation smoothness to produce consistent and non-flickering results. We note that some existing works address only this task (e.g. ViewCrafter [9])." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 599, + 411, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 411, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 411, + 612 + ], + "type": "text", + "content": "2.2. Existing Models" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 617, + 554, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 666 + ], + "type": "text", + "content": "We group existing approaches into regression- and diffusion-based models based on their high-level design choices. A more detailed discussion of related works can be found in the appendix." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 674, + 553, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 674, + 553, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 674, + 553, + 687 + ], + "type": "text", + "content": "Regression-based models learn a deterministic mapping:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 393, + 696, + 474, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 696, + 474, + 716 + ], + "spans": [ + { + "bbox": [ + 393, + 696, + 474, + 716 + ], + "type": "interline_equation", + "content": "f _ {\\theta} \\left(\\mathbf {I} ^ {\\text {i n p}}, \\boldsymbol {\\pi} ^ {\\text {i n p}}, \\boldsymbol {\\pi} ^ {\\text {t g t}}\\right),", + "image_path": "ba59ed0f1d6c51e3cbc8a115f75cb065be8974bed97e8c8a0b970c5836afc18a.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12406" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 79, + 553, + 401 + ], + "blocks": [ + { + "bbox": [ + 58, + 79, + 553, + 401 + ], + "lines": [ + { + "bbox": [ + 58, + 79, + 553, + 401 + ], + "spans": [ + { + "bbox": [ + 58, + 79, + 553, + 401 + ], + "type": "image", + "image_path": "d5b834cf8c302efb60c3cc213a069ca1a0f3c04730ee044608b07528d09ade78.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 404, + 555, + 427 + ], + "lines": [ + { + "bbox": [ + 55, + 404, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 555, + 427 + ], + "type": "text", + "content": "Figure 2. Diverse camera control. SEVA generates photorealistic novel views following diverse camera trajectories. This includes orbit, spiral, zoom out, dolly zooms, and any user-specified trajectories. Please visit our website for more visual results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 434, + 295, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 517 + ], + "type": "text", + "content": "to directly generate " + }, + { + "bbox": [ + 55, + 434, + 295, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 55, + 434, + 295, + 517 + ], + "type": "text", + "content": " deterministically from " + }, + { + "bbox": [ + 55, + 434, + 295, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{inp}}, \\boldsymbol{\\pi}^{\\mathrm{inp}}, \\boldsymbol{\\pi}^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 55, + 434, + 295, + 517 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 434, + 295, + 517 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 55, + 434, + 295, + 517 + ], + "type": "text", + "content": " can be either an end-to-end network parameterized by " + }, + { + "bbox": [ + 55, + 434, + 295, + 517 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 434, + 295, + 517 + ], + "type": "text", + "content": ", or a composition of a feed-forward prediction of an intermediate 3D representation and then a neural renderer (e.g., NeRF [19] or 3DGS [3]). For the latter case, set NVs and trajectory NVs are solved in the same way since there exists a persistent 3D representation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 524, + 295, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 524, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 55, + 524, + 295, + 546 + ], + "type": "text", + "content": "Diffusion-based models capture the conditional distribution:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 124, + 544, + 227, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 544, + 227, + 564 + ], + "spans": [ + { + "bbox": [ + 124, + 544, + 227, + 564 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(\\mathbf {I} ^ {\\mathrm {t g t}} \\mid \\mathbf {I} ^ {\\mathrm {i n p}}, \\boldsymbol {\\pi} ^ {\\mathrm {i n p}}, \\boldsymbol {\\pi} ^ {\\mathrm {t g t}}\\right),", + "image_path": "42e9679f313a23aab3ef1692e58d7a67d67f6e54dea6f9e6c4c33842c9d6547f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": "from which " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": " are sampled [20] iteratively. We highlight two types of models within this scope: Image and Video models. Image models are trained on unordered image sets, such that " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "(\\mathbf{I}^{\\mathrm{inp}},\\mathbf{I}^{\\mathrm{tgt}})\\sim \\mathcal{I}" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = \\{\\mathbf{I}_{\\rho (1)},\\mathbf{I}_{\\rho (2)},\\dots ,\\mathbf{I}_{\\rho (M + N)}\\}" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": " is an image batch, and " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\rho (\\cdot)" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": " is a random permutation function (camera parameters are omitted for simplicity). Image models usually thrive at set NVs but struggle at trajectory NVS since they are designed to generate images and not videos. Moreover, the unordered nature of all views solicits flexible input conditioning. Video models are instead trained on ordered views, such that " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "(\\mathbf{I}^{\\mathrm{inp}},\\mathbf{I}^{\\mathrm{tgt}})\\sim \\mathcal{V}" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{\\mathbf{I}_1,\\mathbf{I}_2,\\dots ,\\mathbf{I}_{M + N}\\}" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 432, + 421, + 487 + ], + "blocks": [ + { + "bbox": [ + 317, + 432, + 421, + 487 + ], + "lines": [ + { + "bbox": [ + 317, + 432, + 421, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 432, + 421, + 487 + ], + "type": "image", + "image_path": "4518ec28f6c75a7e52b1e70835d8bce7d8bba039f0fbb7a69b5b8a76f6374b69.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 490, + 383, + 498 + ], + "lines": [ + { + "bbox": [ + 353, + 490, + 383, + 498 + ], + "spans": [ + { + "bbox": [ + 353, + 490, + 383, + 498 + ], + "type": "text", + "content": "SetNVS" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 424, + 432, + 526, + 487 + ], + "blocks": [ + { + "bbox": [ + 424, + 432, + 526, + 487 + ], + "lines": [ + { + "bbox": [ + 424, + 432, + 526, + 487 + ], + "spans": [ + { + "bbox": [ + 424, + 432, + 526, + 487 + ], + "type": "image", + "image_path": "d3b77916ee65256010f7877cfd0d879c34e6fca09a92b63d881f6cc0b5880459.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 489, + 501, + 499 + ], + "lines": [ + { + "bbox": [ + 447, + 489, + 501, + 499 + ], + "spans": [ + { + "bbox": [ + 447, + 489, + 501, + 499 + ], + "type": "text", + "content": "Trajectory NVS" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 502, + 555, + 536 + ], + "lines": [ + { + "bbox": [ + 313, + 502, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 555, + 536 + ], + "type": "text", + "content": "Figure 3. Set NVS versus trajectory NVS. Set NVS generates target views as an image set, whereas trajectory NVS produces them as a trajectory video." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 527, + 432, + 553, + 487 + ], + "blocks": [ + { + "bbox": [ + 527, + 432, + 553, + 487 + ], + "lines": [ + { + "bbox": [ + 527, + 432, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 527, + 432, + 553, + 487 + ], + "type": "image", + "image_path": "ce0250ea6326add0094e34b1c625f739992a3a56a17efd99669ca3f462d11661.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 539, + 555, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 555, + 635 + ], + "type": "text", + "content": "is a set of video frames with preserved ordering. Additional temporal operators may also be used to improve temporal smoothness, such as temporal positional encoding and temporal attention. In contrast with image models, video models thrive at trajectory NVS but struggle at set NVS. Moreover, all existing video models require both input and target views to be ordered (input views followed by target ones), constraining their input flexibility [10, 18, 21, 22]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 646, + 370, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 646, + 370, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 646, + 370, + 658 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "It is challenging to develop a unified model that works for different NVS tasks while being flexible in input and output domains. Our design choices enable SEVA to achieve high generation capacity, smooth view interpolation, and flexi" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12407" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 227, + 194 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 227, + 194 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 227, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 227, + 194 + ], + "type": "image", + "image_path": "1bf2f170e5012697b960197a8454e34e52415c909a54c43e646dc0e6d9483f57.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 196, + 555, + 240 + ], + "lines": [ + { + "bbox": [ + 55, + 196, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 196, + 555, + 240 + ], + "type": "text", + "content": "Figure 4. Method. SEVA is trained with fixed sequence length as a “" + }, + { + "bbox": [ + 55, + 196, + 555, + 240 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 196, + 555, + 240 + ], + "type": "text", + "content": "-in " + }, + { + "bbox": [ + 55, + 196, + 555, + 240 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 196, + 555, + 240 + ], + "type": "text", + "content": "-out” multi-view diffusion model with standard architecture. During sampling, SEVA can be cast as a generative “" + }, + { + "bbox": [ + 55, + 196, + 555, + 240 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 196, + 555, + 240 + ], + "type": "text", + "content": "-in " + }, + { + "bbox": [ + 55, + 196, + 555, + 240 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 196, + 555, + 240 + ], + "type": "text", + "content": "-out” renderer that works with variable sequence length, where " + }, + { + "bbox": [ + 55, + 196, + 555, + 240 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 196, + 555, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 196, + 555, + 240 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 196, + 555, + 240 + ], + "type": "text", + "content": " need not be equal to " + }, + { + "bbox": [ + 55, + 196, + 555, + 240 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 196, + 555, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 196, + 555, + 240 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 196, + 555, + 240 + ], + "type": "text", + "content": ". To enhance temporal and 3D consistency across generated views, especially when generating along a trajectory, we present procedural two-pass sampling as a general strategy." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 230, + 69, + 555, + 195 + ], + "blocks": [ + { + "bbox": [ + 230, + 69, + 555, + 195 + ], + "lines": [ + { + "bbox": [ + 230, + 69, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 230, + 69, + 555, + 195 + ], + "type": "image", + "image_path": "801909587f252e64b3dd1adecbc53ed19c318cae5441d2991497717a92097bac.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 247, + 295, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 295, + 272 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 295, + 272 + ], + "type": "text", + "content": "ble input conditioning (Tab. 1) in both set NVs as well as trajectory NVs. A system overview is provided in Fig. 4." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 277, + 207, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 207, + 290 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 207, + 290 + ], + "type": "text", + "content": "3.1. Model Design and Training" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 295, + 296, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 296, + 331 + ], + "type": "text", + "content": "We consider an \" " + }, + { + "bbox": [ + 55, + 295, + 296, + 331 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 295, + 296, + 331 + ], + "type": "text", + "content": "-in " + }, + { + "bbox": [ + 55, + 295, + 296, + 331 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 295, + 296, + 331 + ], + "type": "text", + "content": "-out\" multi-view diffusion model " + }, + { + "bbox": [ + 55, + 295, + 296, + 331 + ], + "type": "inline_equation", + "content": "p_{\\theta}" + }, + { + "bbox": [ + 55, + 295, + 296, + 331 + ], + "type": "text", + "content": ", as noted in Sec. 2.2. We formulate this learning problem as a standard diffusion process [20] without any change." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 337, + 296, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 296, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 296, + 517 + ], + "type": "text", + "content": "Architecture. Our model is based on the publicly available SD 2.1 [23], which consists of an auto-encoder and a latent denoising U-Net. Following [8], we inflate the 2D self-attention of each low-resolution residual block into 3D self-attention [24] within the U-Net. To improve model capacity, we add 1D self-attention along the view axis after each self-attention block via skip connection [25, 26], bumping the model parameters from 870M to 1.3B. Optionally, we further tame this model into a video model by introducing 3D convolutions after each residual block via skip connection, similar to [27, 28], yielding 1.5B total parameters. The temporal pathway can be enabled during inference when frames within one forward pass are known to be spatially ordered, enhancing output's smoothness. For the conditioning of our model, please refer to the appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": "Training. Let us define the training context window length " + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "inline_equation", + "content": "T = |\\mathbf{I}^{\\mathrm{inp}}| + |\\mathbf{I}^{\\mathrm{tgt}}| = M + N" + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": ". One natural goal is to support large " + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": " such that we can generate larger set of frames. However, we find that naive training is prone to divergence, and we thus employ a two-stage training curriculum. During the first stage, we train our model with " + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "inline_equation", + "content": "T = 8" + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": " with a batch size of 1472 for " + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "inline_equation", + "content": "100K" + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": " iterations. In the second stage, we train our model with " + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "inline_equation", + "content": "T = 21" + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": " with a batch size of 512 for " + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "inline_equation", + "content": "600K" + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": " iterations. Given a training video sequence, we randomly sample the number of input frames " + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "inline_equation", + "content": "M \\in [1,T - 1]" + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": " and the frames " + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "inline_equation", + "content": "(\\mathbf{I}^{\\mathrm{inp}},\\mathbf{I}^{\\mathrm{gt}})" + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": " across the entire sequence to encourage diversity. We find it important to jointly sample with a smaller subsampling stride to ensure sufficient temporal granularity and avoid missing critical transitions with a small probability (0.2 is used in practice). In the optional video training stage, we only train" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 247, + 555, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 555, + 354 + ], + "type": "text", + "content": "temporal weights with data sampled with a small subsampling stride and a batch size of 512 for " + }, + { + "bbox": [ + 313, + 247, + 555, + 354 + ], + "type": "inline_equation", + "content": "200K" + }, + { + "bbox": [ + 313, + 247, + 555, + 354 + ], + "type": "text", + "content": " iterations. We shift the signal-to-noise ratio (SNR) in all stages as more noise is necessary to remove the input information, corroborating findings from [8, 29, 30]. The model is trained with squared images with " + }, + { + "bbox": [ + 313, + 247, + 555, + 354 + ], + "type": "inline_equation", + "content": "H = W = 576" + }, + { + "bbox": [ + 313, + 247, + 555, + 354 + ], + "type": "text", + "content": ". The training data is a superset of data used in [8] and derives from all public academic datasets, consisting of around " + }, + { + "bbox": [ + 313, + 247, + 555, + 354 + ], + "type": "inline_equation", + "content": "340K" + }, + { + "bbox": [ + 313, + 247, + 555, + 354 + ], + "type": "text", + "content": " objects and " + }, + { + "bbox": [ + 313, + 247, + 555, + 354 + ], + "type": "inline_equation", + "content": "80K" + }, + { + "bbox": [ + 313, + 247, + 555, + 354 + ], + "type": "text", + "content": " scenes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 364, + 441, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 441, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 441, + 376 + ], + "type": "text", + "content": "3.2. Sampling Novel Views" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 382, + 554, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 554, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 554, + 466 + ], + "type": "text", + "content": "Once the diffusion model is trained, we can sample it for a wide range of NVS tasks during test time. Formally, let us consider a \" " + }, + { + "bbox": [ + 313, + 382, + 554, + 466 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 382, + 554, + 466 + ], + "type": "text", + "content": "-in " + }, + { + "bbox": [ + 313, + 382, + 554, + 466 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 382, + 554, + 466 + ], + "type": "text", + "content": "-out\" NVS task during testing, where we are given " + }, + { + "bbox": [ + 313, + 382, + 554, + 466 + ], + "type": "inline_equation", + "content": "P = |\\mathbf{I}^{\\mathrm{inp}}|" + }, + { + "bbox": [ + 313, + 382, + 554, + 466 + ], + "type": "text", + "content": " input frames and aim to produce " + }, + { + "bbox": [ + 313, + 382, + 554, + 466 + ], + "type": "inline_equation", + "content": "Q = |\\mathbf{I}^{\\mathrm{tgt}}|" + }, + { + "bbox": [ + 313, + 382, + 554, + 466 + ], + "type": "text", + "content": " target frames. Our goal is to design a generic sampling strategy that works for all " + }, + { + "bbox": [ + 313, + 382, + 554, + 466 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 382, + 554, + 466 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 382, + 554, + 466 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 382, + 554, + 466 + ], + "type": "text", + "content": " configurations, where " + }, + { + "bbox": [ + 313, + 382, + 554, + 466 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 382, + 554, + 466 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 382, + 554, + 466 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 382, + 554, + 466 + ], + "type": "text", + "content": " need not be equal to " + }, + { + "bbox": [ + 313, + 382, + 554, + 466 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 382, + 554, + 466 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 382, + 554, + 466 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 382, + 554, + 466 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 467, + 554, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 554, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 554, + 598 + ], + "type": "text", + "content": "We make two key observations: First, within a single forward pass, predictions are 3D consistent, provided the model is well-trained. Second, when " + }, + { + "bbox": [ + 313, + 467, + 554, + 598 + ], + "type": "inline_equation", + "content": "P + Q > T" + }, + { + "bbox": [ + 313, + 467, + 554, + 598 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 467, + 554, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{gt}}" + }, + { + "bbox": [ + 313, + 467, + 554, + 598 + ], + "type": "text", + "content": " must be split into smaller chunks of size " + }, + { + "bbox": [ + 313, + 467, + 554, + 598 + ], + "type": "inline_equation", + "content": "Q_{i}" + }, + { + "bbox": [ + 313, + 467, + 554, + 598 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 313, + 467, + 554, + 598 + ], + "type": "inline_equation", + "content": "P + Q_{i} \\leqslant T" + }, + { + "bbox": [ + 313, + 467, + 554, + 598 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 313, + 467, + 554, + 598 + ], + "type": "inline_equation", + "content": "i^{\\mathrm{th}}" + }, + { + "bbox": [ + 313, + 467, + 554, + 598 + ], + "type": "text", + "content": " forward pass. We term this practice one-pass sampling. However, predictions across these forward passes would be inconsistent unless they share common frames to maintain local consistency within a spatial neighborhood. Building on these observations, we summarize our sampling process under two scenarios: " + }, + { + "bbox": [ + 313, + 467, + 554, + 598 + ], + "type": "inline_equation", + "content": "P + Q \\leqslant T" + }, + { + "bbox": [ + 313, + 467, + 554, + 598 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 467, + 554, + 598 + ], + "type": "inline_equation", + "content": "P + Q > T" + }, + { + "bbox": [ + 313, + 467, + 554, + 598 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 607, + 554, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 607, + 554, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 607, + 554, + 668 + ], + "type": "inline_equation", + "content": "P + Q \\leqslant T" + }, + { + "bbox": [ + 313, + 607, + 554, + 668 + ], + "type": "text", + "content": ". We fit the task within one forward pass for simplicity and consistency. As shown in the appendix, we find it works better to pad the forward pass to have exactly " + }, + { + "bbox": [ + 313, + 607, + 554, + 668 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 607, + 554, + 668 + ], + "type": "text", + "content": " frames by repeating the first input image, compared to changing the context window " + }, + { + "bbox": [ + 313, + 607, + 554, + 668 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 607, + 554, + 668 + ], + "type": "text", + "content": " zero-shot." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "P + Q > T" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": ". We propose procedural two-pass sampling: In the first pass, we generate anchor frames " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{acr}}" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": " using all input frames " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{inp}}" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": ". In the second pass, we divide " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": " into" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12408" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "chunks and generate them using " + }, + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{acr}}" + }, + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": " (and optionally " + }, + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{inp}}" + }, + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": ") according to the spatial distribution of " + }, + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{acr}}" + }, + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{gt}}" + }, + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 295, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 295, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 295, + 144 + ], + "type": "text", + "content": "Given the distinct nature of the two tasks of interest—set NVS and trajectory NVS—e.g., differences in the availability of views' ordering, we design tailored chunking strategies for each task." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 144, + 296, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 296, + 216 + ], + "type": "text", + "content": "For set NVS, we consider nearest procedural sampling. We first generate " + }, + { + "bbox": [ + 55, + 144, + 296, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{acr}}" + }, + { + "bbox": [ + 55, + 144, + 296, + 216 + ], + "type": "text", + "content": " based on pre-defined trajectory priors, similar to [8], e.g., 360 trajectories for object-centric scenes or spiral trajectories for forward-facing scenes. We then divide " + }, + { + "bbox": [ + 55, + 144, + 296, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{gt}}" + }, + { + "bbox": [ + 55, + 144, + 296, + 216 + ], + "type": "text", + "content": " into chunks w.r.t. " + }, + { + "bbox": [ + 55, + 144, + 296, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{acr}}" + }, + { + "bbox": [ + 55, + 144, + 296, + 216 + ], + "type": "text", + "content": " using nearest neighbor. Specifically, the " + }, + { + "bbox": [ + 55, + 144, + 296, + 216 + ], + "type": "inline_equation", + "content": "i^{\\mathrm{th}}" + }, + { + "bbox": [ + 55, + 144, + 296, + 216 + ], + "type": "text", + "content": " forward pass involves:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 224, + 271, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 224, + 271, + 239 + ], + "spans": [ + { + "bbox": [ + 80, + 224, + 271, + 239 + ], + "type": "interline_equation", + "content": "\\operatorname {n e a r e s t}: \\left\\{\\mathbf {I} _ {i} ^ {\\mathrm {a c r}} \\right\\} \\cup \\left\\{\\mathbf {I} _ {j} ^ {\\mathrm {t g t}} \\mid \\mathrm {N N} (\\mathbf {I} _ {j} ^ {\\mathrm {t g t}}, \\mathbf {I} ^ {\\mathrm {a c r}}) = \\mathbf {I} _ {i} ^ {\\mathrm {a c r}} \\right\\}.", + "image_path": "51c3cdbff2dca9c8a656db6b0645cfbc829e465660fdd902309873b040fefa0b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 246, + 296, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 246, + 296, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 246, + 296, + 341 + ], + "type": "text", + "content": "We considered two strategies of procedural sampling: nearest as described above, and " + }, + { + "bbox": [ + 55, + 246, + 296, + 341 + ], + "type": "inline_equation", + "content": "\\mathrm{gt} +" + }, + { + "bbox": [ + 55, + 246, + 296, + 341 + ], + "type": "text", + "content": " nearest strategy by appending " + }, + { + "bbox": [ + 55, + 246, + 296, + 341 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{inp}}" + }, + { + "bbox": [ + 55, + 246, + 296, + 341 + ], + "type": "text", + "content": " into each forward pass. We find that the " + }, + { + "bbox": [ + 55, + 246, + 296, + 341 + ], + "type": "inline_equation", + "content": "\\mathrm{gt} +" + }, + { + "bbox": [ + 55, + 246, + 296, + 341 + ], + "type": "text", + "content": " nearest strategy performs better than nearest, and thus default to it instead. In the absence of trajectory priors, we revert to one-pass sampling. In practice, employing nearest anchors enhances qualitative consistency, albeit on a limited scale." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 342, + 296, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 401 + ], + "type": "text", + "content": "For trajectory NVS, we consider interp procedural sampling. We first generate a subset of target frames as " + }, + { + "bbox": [ + 55, + 342, + 296, + 401 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{acr}}" + }, + { + "bbox": [ + 55, + 342, + 296, + 401 + ], + "type": "text", + "content": " by uniformly spanning the target camera path with a stride " + }, + { + "bbox": [ + 55, + 342, + 296, + 401 + ], + "type": "inline_equation", + "content": "\\Delta = \\left\\lfloor \\frac{Q}{T - 2}\\right\\rfloor" + }, + { + "bbox": [ + 55, + 342, + 296, + 401 + ], + "type": "text", + "content": ". We then generate the rest of " + }, + { + "bbox": [ + 55, + 342, + 296, + 401 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 55, + 342, + 296, + 401 + ], + "type": "text", + "content": " as segments between those anchors:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 83, + 409, + 268, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 409, + 268, + 426 + ], + "spans": [ + { + "bbox": [ + 83, + 409, + 268, + 426 + ], + "type": "interline_equation", + "content": "\\mathrm {i n t e r p}: \\{\\mathbf {I} _ {i} ^ {\\mathrm {a c r}}, \\mathbf {I} _ {i \\cdot \\Delta + 1} ^ {\\mathrm {t g t}}, \\dots , \\mathbf {I} _ {(i + 1) \\cdot \\Delta - 1} ^ {\\mathrm {t g t}}, \\mathbf {I} _ {i + 1} ^ {\\mathrm {a c r}} \\}.", + "image_path": "42c553a4c368a29997de6441bf5e280d646e77495fa6f5b9f6578f7f27358db7.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 432, + 296, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 296, + 527 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 296, + 527 + ], + "type": "text", + "content": "Since the input to the model is ordered, we can leverage temporal weights to further improve smoothness (Sec. 4.3). Similarly, " + }, + { + "bbox": [ + 55, + 432, + 296, + 527 + ], + "type": "inline_equation", + "content": "\\mathrm{gt} + \\mathrm{interp}" + }, + { + "bbox": [ + 55, + 432, + 296, + 527 + ], + "type": "text", + "content": " is possible by appending " + }, + { + "bbox": [ + 55, + 432, + 296, + 527 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{inp}}" + }, + { + "bbox": [ + 55, + 432, + 296, + 527 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 432, + 296, + 527 + ], + "type": "inline_equation", + "content": "\\Delta = \\left\\lfloor \\frac{Q}{T - P - 2} \\right\\rfloor" + }, + { + "bbox": [ + 55, + 432, + 296, + 527 + ], + "type": "text", + "content": ". We find that interp is sufficiently robust, and choose it as the default option. The interp strategy drastically outperforms its counterparts (e.g., one-pass, or " + }, + { + "bbox": [ + 55, + 432, + 296, + 527 + ], + "type": "inline_equation", + "content": "\\mathrm{gt} + \\mathrm{nearest}" + }, + { + "bbox": [ + 55, + 432, + 296, + 527 + ], + "type": "text", + "content": " procedural sampling) in terms of temporal smoothness." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 534, + 253, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 253, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 253, + 548 + ], + "type": "text", + "content": "3.3. Scaling Sampling for Large " + }, + { + "bbox": [ + 55, + 534, + 253, + 548 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 534, + 253, + 548 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 534, + 253, + 548 + ], + "type": "inline_equation", + "content": "Q" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 552, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 296, + 601 + ], + "type": "text", + "content": "Next, we examine two special cases when " + }, + { + "bbox": [ + 55, + 552, + 296, + 601 + ], + "type": "inline_equation", + "content": "P + Q > T" + }, + { + "bbox": [ + 55, + 552, + 296, + 601 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 55, + 552, + 296, + 601 + ], + "type": "inline_equation", + "content": "P > T" + }, + { + "bbox": [ + 55, + 552, + 296, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 552, + 296, + 601 + ], + "type": "inline_equation", + "content": "Q \\gg T" + }, + { + "bbox": [ + 55, + 552, + 296, + 601 + ], + "type": "text", + "content": ". Here, we make a tailored design for anchor generation in the first pass, while keeping target generation in the second pass unchanged." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "P > T" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": ". In the semi-dense-view regime (e.g., " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "P = 32" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "), we extend the context window length " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " zero-shot to accommodate all " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " input views and anchor views in one pass during anchor generation. Empirically, " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " can even be extended up to hundreds without severe degradation in photorealism in the generated outputs. We find that the diffusion model generalizes well in this case as long as the input views cover the majority of the scene, shifting the task from generation to primarily interpolation. In the sparse-view" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 553, + 171 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 553, + 171 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 553, + 171 + ], + "type": "image", + "image_path": "a43cdb8665181349f6de4fcef3de81638119f70716dba8864ecbf09cb5e37967.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 177, + 555, + 221 + ], + "lines": [ + { + "bbox": [ + 313, + 177, + 555, + 221 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 555, + 221 + ], + "type": "text", + "content": "Figure 5. Anchor generation when " + }, + { + "bbox": [ + 313, + 177, + 555, + 221 + ], + "type": "inline_equation", + "content": "Q \\gg T" + }, + { + "bbox": [ + 313, + 177, + 555, + 221 + ], + "type": "text", + "content": ". We introduce a memory bank composed of previously-generated anchor views and their corresponding camera poses. The lookup of spatial neighbors helps improve long-term 3D consistency." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 229, + 555, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 229, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 555, + 277 + ], + "type": "text", + "content": "regime " + }, + { + "bbox": [ + 313, + 229, + 555, + 277 + ], + "type": "inline_equation", + "content": "(P \\leq 8)" + }, + { + "bbox": [ + 313, + 229, + 555, + 277 + ], + "type": "text", + "content": ", we observe similar performance degradation caused by zero-shot extension of " + }, + { + "bbox": [ + 313, + 229, + 555, + 277 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 229, + 555, + 277 + ], + "type": "text", + "content": " compared to what we have found when " + }, + { + "bbox": [ + 313, + 229, + 555, + 277 + ], + "type": "inline_equation", + "content": "P + Q \\leqslant T" + }, + { + "bbox": [ + 313, + 229, + 555, + 277 + ], + "type": "text", + "content": ". Refer to the appendix for a detailed discussion." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 296, + 556, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 296, + 556, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 296, + 556, + 452 + ], + "type": "inline_equation", + "content": "Q \\gg T" + }, + { + "bbox": [ + 313, + 296, + 556, + 452 + ], + "type": "text", + "content": ". When the number of target views " + }, + { + "bbox": [ + 313, + 296, + 556, + 452 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 296, + 556, + 452 + ], + "type": "text", + "content": " is large, e.g., in large-set NVS or long-trajectory NVS, even anchors will be chunked into different forwards in the first pass, leading to the inconsistency of anchors. To this end, we maintain a memory bank of anchor views previously generated, as shown in Fig. 5. We generate anchors autoregressively by retrieving their spatially nearest ones from the memory bank, similar to the nearest strategy introduced above for the second pass. In Sec. 4.4, we show that this strategy drastically outperforms the standard practice of reusing temporally nearest anchors previously generated in long video literature [22], in terms of long-range 3D consistency, especially for hard trajectories." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 461, + 395, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 395, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 395, + 475 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 482, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 555, + 567 + ], + "type": "text", + "content": "We employ a single model for a spectrum of settings and find that SEVA model generalizes well under the three criteria (Tab. 1). We cover different NVS tasks (set NVS and trajectory NVS) and examine one special task of interest—long trajectory NVS. We also cover different input regimes (single-view, sparse-view, and semi-dense-view). A discussion about key properties is presented in the appendix." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 575, + 392, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 392, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 392, + 588 + ], + "type": "text", + "content": "4.1. Benchmark" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": "Datasets, splits, and the number of input views. We collect 10 commonly used datasets to benchmark NVS, encompassing a diverse range of scene distributions and complexities. We consider (1) object datasets [32, 33]; (2) object-centric scene datasets [34-38]; and (3) scene datasets [39-41]. To establish a comprehensive and rigorous comparison with baselines, we consider different dataset splits utilized in prior works with the same input-view configuration, unless specified as our split (O). These include splits used in 4DiM [12] (D), ViewCrafter [9] (V)," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12409" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 555, + 337 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 337 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 337 + ], + "type": "image", + "image_path": "a4b8491584ad3d6abe80aa4aa7086eb632e6d09a48470af878ef5e46c8e001c0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 338, + 555, + 384 + ], + "lines": [ + { + "bbox": [ + 55, + 338, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 555, + 384 + ], + "type": "text", + "content": "Figure 6. SOTA comparison on set NVs (top) and trajectory NVs (bottom) across varying numbers of input views. From left to right and top to bottom, inputs views are (top, i.e. set NVS) 1, 2, 3, 32; (bottom, i.e. trajectory NVS) 1, 2, 1, 1, 3, 3. For cases with multiple input views, the view closest to the target is shown over the rest. We compare with open-source approaches—ViewCrafter [9] (VC) and DepthSplat [31] (DS)—as well as proprietary ones including LVSM [11], Long-LRM [17] (LLRM), 4DiM [12], and CAT3D [8]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 389, + 295, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 295, + 415 + ], + "type": "text", + "content": "pixelSplat [15] (P), ReconFusion [7] (R), SV3D [18] (S), and Long-LRM [17] (L)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 430, + 295, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 295, + 515 + ], + "type": "text", + "content": "Small-viewpoint versus large-viewpoint NVS. Sweeping across all datasets, splits, and input-view configurations reveals a diverse benchmark of setups. To better evaluate models' generation capacity and interpolation smoothness (Sec. 2.1), we propose to categorize this setup into two groups—small-viewpoint NVS and large-viewpoint NVS—depending on the disparity between " + }, + { + "bbox": [ + 55, + 430, + 295, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 55, + 430, + 295, + 515 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 430, + 295, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{inp}}" + }, + { + "bbox": [ + 55, + 430, + 295, + 515 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 531, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 295, + 639 + ], + "type": "text", + "content": "Baselines. We consider a range of proprietary models, including ReconFusion [7], CAT3D [8], 4DiM [12], LVSM [11], and Long-LRM [17]. We also consider various open-source models, including SV3D [18], MVSplat [16], depthSplat [31], MotionCtrl [42], and ViewCrafter [9]. As outlined in Sec. 2.2, these baselines encompass both regression-based and diffusion-based approaches, providing a comprehensive framework for comparison. Please refer to the appendix for more details of the benchmark." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 647, + 118, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 118, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 118, + 659 + ], + "type": "text", + "content": "4.2. Set NVs" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": "Quantitative comparison. The input and target views are chosen following the splits used in previous methods. The order of target views is not preserved. We use standard metrics of peak signal-to-noise ratio (PSNR), structural simi" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 389, + 555, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 461 + ], + "type": "text", + "content": "larity index measure [44] (SSIM), and learned perceptual image patch similarity [45] (LPIPS). Only PSNR is showcased here due to space limits, with the rest deferred to the appendix. Empirically, our method shows a greater performance improvement on LPIPS, reflecting the photorealism of our results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 462, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 617 + ], + "type": "text", + "content": "For small-viewpoint set NVs, Tab. 2 shows that SEVA sets state-of-the-art results in the majority of splits. In the sparse-view regime (i.e., " + }, + { + "bbox": [ + 313, + 462, + 555, + 617 + ], + "type": "inline_equation", + "content": "P \\leqslant 8" + }, + { + "bbox": [ + 313, + 462, + 555, + 617 + ], + "type": "text", + "content": "), SEVA excels across different datasets when " + }, + { + "bbox": [ + 313, + 462, + 555, + 617 + ], + "type": "inline_equation", + "content": "P > 1" + }, + { + "bbox": [ + 313, + 462, + 555, + 617 + ], + "type": "text", + "content": ". For example, a performance gain of +6.0 dB PSNR is achieved on LLFF with " + }, + { + "bbox": [ + 313, + 462, + 555, + 617 + ], + "type": "inline_equation", + "content": "P = 3" + }, + { + "bbox": [ + 313, + 462, + 555, + 617 + ], + "type": "text", + "content": ". In the semi-dense-view regime (e.g., " + }, + { + "bbox": [ + 313, + 462, + 555, + 617 + ], + "type": "inline_equation", + "content": "P = 32" + }, + { + "bbox": [ + 313, + 462, + 555, + 617 + ], + "type": "text", + "content": "), SEVA surprisingly performs favorably against the specialized model [17], despite not being specifically designed for this setup. For example, SEVA lags behind the state-of-the-art method [17] by only 1.7 dB on T&T. A detailed discussion on the RE10K dataset can be found in the appendix. On object datasets OO3D and GSO, SEVA achieves a significantly higher state-of-the-art PSNR compared to all other methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": "For large-viewpoint set NVs, Tab. 3 shows that SEVA's quantitative advantages are even more prominent here, revealing clear benefits of SEVA in terms of generation capacity when the camera spans a large spatial range. On Mip360 with " + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "inline_equation", + "content": "P = 3" + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": ", SEVA improves over previous state-of-the-art method CAT3D [8] by " + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "inline_equation", + "content": "+0.6" + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": " dB PSNR. On harder scenes like DL3DV and T&T with different input-view configurations, SEVA obtains a clear performance lead." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12410" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 70, + 547, + 220 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 547, + 220 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 547, + 220 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 547, + 220 + ], + "type": "table", + "html": "
Methoddataset split POO3D OGSO ORE10KLLFF R [7] 1 3DTU R [7] 1 3CO3D V [9] R [7] 1 3WRGBD Oe 3 6Mip360 O L [17] 6 32DL3DV T & T V [9] L [17] 1 32
O 3O 3D [12] 1V [9] 1P [15] 2R [7] 13
Regression-based models
Long-LRM [17]-------------23.86 - 18.20
MVSplat [16]14.78 15.6715.21 16.5220.42 20.9020.32 19.2426.39 27.4421.56 21.8725.64 22.5411.23 12.0712.50 12.6213.87 15.52 14.1513.52 13.52 13.2314.56 12.54 15.9313.56 14.24 14.2313.24 13.22 14.35
DepthSplat [31]
LVSM [11]----29.67---------
Diffusion-based models
MotionCtrl [42]--12.74 17.0816.29 ------15.46 ----13.29 -
4DiM [12]--17.08-----------
ViewCrafter [9]14.64 30.3015.93 31.5320.43 17.9922.04 18.5621.42 25.6620.88 18.1122.81 27.5710.53 14.0313.52 14.4712.66 16.40 18.4718.96 14.72 18.4016.42 12.66 19.7514.59 13.78 16.70- 18.07 - 15.16
SEVA
", + "image_path": "48ef1fdff6384acc2831050f6090b14d45fbf5a3b56359e23f650fa4eb8d9db0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 262, + 369, + 361 + ], + "blocks": [ + { + "bbox": [ + 55, + 223, + 555, + 257 + ], + "lines": [ + { + "bbox": [ + 55, + 223, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 555, + 257 + ], + "type": "text", + "content": "Table 2. PSNR† on small-viewpoint set NVS. " + }, + { + "bbox": [ + 55, + 223, + 555, + 257 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 223, + 555, + 257 + ], + "type": "text", + "content": " denotes the number of input views. For all results with " + }, + { + "bbox": [ + 55, + 223, + 555, + 257 + ], + "type": "inline_equation", + "content": "P = 1" + }, + { + "bbox": [ + 55, + 223, + 555, + 257 + ], + "type": "text", + "content": ", we sweep the unit length for camera normalization due to the model's scale ambiguity. " + }, + { + "bbox": [ + 55, + 223, + 555, + 257 + ], + "type": "inline_equation", + "content": "\\mathrm{O}_e" + }, + { + "bbox": [ + 55, + 223, + 555, + 257 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 223, + 555, + 257 + ], + "type": "inline_equation", + "content": "\\mathrm{O}_h" + }, + { + "bbox": [ + 55, + 223, + 555, + 257 + ], + "type": "text", + "content": " denote the easy and hard split of our split, respectively. Underlined numbers are run by us using the official released code." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 262, + 369, + 361 + ], + "lines": [ + { + "bbox": [ + 58, + 262, + 369, + 361 + ], + "spans": [ + { + "bbox": [ + 58, + 262, + 369, + 361 + ], + "type": "table", + "html": "
Methoddataset split POO3D S [18] 1GSO S [18] 1CO3D R [7] 1WRGBD Oh 1Mip360 R [7] 3DL3DV O 1T&T O 31369
SV3D [18]19.2820.38----------
DepthSplat [31]11.5612.3210.429.3513.5310.4912.549.6312.528.639.7810.12
CAT3D [8]------15.15-----
ViewCrafter [9]10.5611.4210.119.1213.459.7910.348.9711.509.239.8810.32
SEVA19.2520.6515.3014.3717.2812.9315.7813.0115.9511.2812.6513.80
", + "image_path": "fd88ef218fbbf8ca821867ff3dec15305457e250dea210e8fc17864eeb982b71.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 372, + 264, + 552, + 358 + ], + "blocks": [ + { + "bbox": [ + 372, + 264, + 552, + 358 + ], + "lines": [ + { + "bbox": [ + 372, + 264, + 552, + 358 + ], + "spans": [ + { + "bbox": [ + 372, + 264, + 552, + 358 + ], + "type": "table", + "html": "
Methodsmall-viewpointlarge-viewpoint
RE10KLLFFDTUCO3DMip360
ZipNeRF [43]20.7717.239.1814.3412.77
ZeroNVS [13]19.1115.9116.7117.1314.44
ReconFusion [7]25.8421.3420.7419.5915.50
CAT3D [8]26.7821.5822.0220.5716.62
SEVA27.9521.8822.6821.8817.82
", + "image_path": "508a3fc5dbfb74647a83806a2d668e2fd53b711dbb112cf0fc88f955522aefb8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 369, + 553, + 403 + ], + "lines": [ + { + "bbox": [ + 55, + 369, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 553, + 403 + ], + "type": "text", + "content": "Table 3. PSNR† on large-viewpoint set NVs. For all results with " + }, + { + "bbox": [ + 55, + 369, + 553, + 403 + ], + "type": "inline_equation", + "content": "P = 1" + }, + { + "bbox": [ + 55, + 369, + 553, + 403 + ], + "type": "text", + "content": ", we sweep the Table 4. PSNR† on 3DGS renderings for set unit length for camera normalization due to the model's scale ambiguity. Underlined NVS. Results are reported on the ReconFusion [7] numbers are run by us using the officially released code. split with " + }, + { + "bbox": [ + 55, + 369, + 553, + 403 + ], + "type": "inline_equation", + "content": "P = 3" + }, + { + "bbox": [ + 55, + 369, + 553, + 403 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 409, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 409, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 54, + 409, + 295, + 495 + ], + "type": "text", + "content": "Comparison of 3D reconstruction. To enable a direct quantitative comparison with prior works [7, 8], we adopt the few-view 3D reconstruction pipeline described in [8]. We optimize the camera parameters and apply LPIPS loss [46] during the distillation. Finally, we render the distilled 3D model on the test views and report the performance in Tab. 4. SEVA shows a consistent performance lead." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 498, + 295, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 498, + 295, + 654 + ], + "spans": [ + { + "bbox": [ + 54, + 498, + 295, + 654 + ], + "type": "text", + "content": "Qualitative comparison. Fig. 6 top panel shows a qualitative comparison with diverse baselines. For small-viewpoint set NVs, the output from SEVA with the best scale exhibits desirable alignment with the ground truth while being more photorealistic in details. Compared with LVSM [11] on the P split of RE10K, SEVA produces sharper images, also corroborating that lower PSNR arises from scale ambiguity rather than interpolation quality. Similar trends hold when compared to Long-LRM [17] on DL3DV with " + }, + { + "bbox": [ + 54, + 498, + 295, + 654 + ], + "type": "inline_equation", + "content": "P = 32" + }, + { + "bbox": [ + 54, + 498, + 295, + 654 + ], + "type": "text", + "content": ". For large-viewpoint set NVs, we compare with DepthSplat [31] on DL3DV with " + }, + { + "bbox": [ + 54, + 498, + 295, + 654 + ], + "type": "inline_equation", + "content": "P = 3" + }, + { + "bbox": [ + 54, + 498, + 295, + 654 + ], + "type": "text", + "content": ". DepthSplat fails to produce reasonable results when the viewpoint change is too large and falls short in overall visual quality." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 660, + 153, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 153, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 153, + 673 + ], + "type": "text", + "content": "4.3. Trajectory NVS" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Qualitative results. Fig. 2 presents qualitative results, illustrating trajectories of varying complexities with different numbers of input views across diverse types." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 409, + 555, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 555, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 555, + 493 + ], + "type": "text", + "content": "In the single-view regime (i.e., " + }, + { + "bbox": [ + 313, + 409, + 555, + 493 + ], + "type": "inline_equation", + "content": "P = 1" + }, + { + "bbox": [ + 313, + 409, + 555, + 493 + ], + "type": "text", + "content": "), we manually craft a set of common camera movements/effects, e.g., look-at 360, spiral, panning, zoom-in, zoom-out, dolly zoom, etc. We observe that SEVA generalizes to a wide range of images and demonstrates accurate camera-following capacity. An extensive sweeping of camera movements on 4 types of images is provided in the appendix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 494, + 556, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 494, + 556, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 556, + 649 + ], + "type": "text", + "content": "In the sparse-view regime with few input views (i.e., " + }, + { + "bbox": [ + 313, + 494, + 556, + 649 + ], + "type": "inline_equation", + "content": "1 < P \\leqslant 8" + }, + { + "bbox": [ + 313, + 494, + 556, + 649 + ], + "type": "text", + "content": "), we observe that SEVA demonstrates strong generalization to in-the-wild real-world images and versatility in adapting to different numbers of input views. The output forms a smooth trajectory video with subtle temporal flickering, revealing its capacity to interpolate between views smoothly. In the last row of Fig. 2, our model generates plausible results at the end of the trajectory—an area unseen in the input observations—demonstrating its strong generation capacity. In the semi-dense-view regime (i.e., " + }, + { + "bbox": [ + 313, + 494, + 556, + 649 + ], + "type": "inline_equation", + "content": "P > 9" + }, + { + "bbox": [ + 313, + 494, + 556, + 649 + ], + "type": "text", + "content": "), we similarly find that SEVA is surprisingly able to produce a smooth trajectory video with minimal artifacts. Please check the supplementary website for video results." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Qualitative comparison. Fig. 6 bottom panel presents a qualitative comparison with diverse baselines. In the single-view regime (i.e., " + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "inline_equation", + "content": "P = 1" + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "), we compare to 4DiM [12] and CAT3D [8]. We observe more photo-realistic and sharper" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12411" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 71, + 292, + 173 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 292, + 173 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 292, + 173 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 292, + 173 + ], + "type": "table", + "html": "
Methodsplit datasetsmall-viewpointlarge-viewpoint
V [9]O
RECO3DT&TREDTUWRDLT&T
MotionCtrl [42]16.2915.4613.29------
DepthSplat [31]19.2413.2314.3525.2314.6812.4511.329.11
ViewCrafter [9]22.0418.9618.0726.5418.9913.4411.459.68
SEVA18.5618.4015.1627.3419.9917.7915.7611.92
SEVA (+ temp.)18.6218.4315.1327.3620.1917.9315.7811.99
", + "image_path": "e132f50a1e0a1699e7008268e344305c60b18995ea4027079a8baf39bc357971.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 182, + 294, + 237 + ], + "lines": [ + { + "bbox": [ + 55, + 182, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 294, + 237 + ], + "type": "text", + "content": "Table 5. PSNR" + }, + { + "bbox": [ + 55, + 182, + 294, + 237 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 55, + 182, + 294, + 237 + ], + "type": "text", + "content": " on trajectory NVS. temp. denotes optional temporal pathway. RE, WR, and DL denotes RE10K, WRGBD, and DL3DV, respectively. For the V [9] split, " + }, + { + "bbox": [ + 55, + 182, + 294, + 237 + ], + "type": "inline_equation", + "content": "P = 1" + }, + { + "bbox": [ + 55, + 182, + 294, + 237 + ], + "type": "text", + "content": " with unit length swept; for the O split, " + }, + { + "bbox": [ + 55, + 182, + 294, + 237 + ], + "type": "inline_equation", + "content": "P = 3" + }, + { + "bbox": [ + 55, + 182, + 294, + 237 + ], + "type": "text", + "content": ". Underlined numbers are run by us using the officially released code." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 57, + 243, + 294, + 392 + ], + "blocks": [ + { + "bbox": [ + 57, + 243, + 294, + 392 + ], + "lines": [ + { + "bbox": [ + 57, + 243, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 57, + 243, + 294, + 392 + ], + "type": "image", + "image_path": "7429eaa8fd38800a41e78b4526e383df46bbff51fb6a7508b0cdc85b3a1575dd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 395, + 295, + 462 + ], + "lines": [ + { + "bbox": [ + 55, + 395, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 395, + 295, + 462 + ], + "type": "text", + "content": "Figure 7. Temporal quality. Vertical slices of a rendered novel camera path on the BONSAI scene from Mip-NeRF360 [38] illustrate the temporal quality across adjacent viewpoints. One-pass or gt + nearest procedural sampling results in notable flickering, whereas interp procedural sampling ensures temporally smooth rendering." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 469, + 296, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 469, + 296, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 296, + 554 + ], + "type": "text", + "content": "output from our model, especially in the background area for object-centric scenes. 4DiM outputs tend to be cartoonish and over-simplistic, given that the model is only trained on RE10K. In the sparse-view regime with few input views (i.e., " + }, + { + "bbox": [ + 55, + 469, + 296, + 554 + ], + "type": "inline_equation", + "content": "P = 3" + }, + { + "bbox": [ + 55, + 469, + 296, + 554 + ], + "type": "text", + "content": "), we compare with CAT3D and observe that our model demonstrates more photo-realistic textures, especially in the background." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "Quantitative comparison. We use the same input views as in set NVS for each split. We use all frames from each scene as target views such that they form a smoothly transitioning trajectory video, i.e., " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{gt}}\\sim \\mathcal{V}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": ". For small-viewpoint trajectory NVS, Tab. 5 compares SEVA with baselines on PSNR. SEVA performs favorably against other methods in V split with " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "P = 1" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": ". The performance lead of ViewCrafter is mainly attributed to its training on high-resolution images. For large-viewpoint trajectory NVS with " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "P = 3" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": ", SEVA consistency sets new state-of-the-art results. Applying the temporal pathway further boosts performance and improves smoothness, indicating the benefits of the gated" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 555, + 230 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 555, + 230 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 555, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 555, + 230 + ], + "type": "image", + "image_path": "1116035c33588665ed1055470f28fc4ffbe5e890116e24c52949be06a628600d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 236, + 555, + 304 + ], + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 304 + ], + "type": "text", + "content": "Figure 8. Long-range 3D consistency. We visualize samples following a camera path looping three times around the TELEPHONE-BOOTH scene. Lookup using spatial neighbors from the memory bank (ours) notably improves view consistency and reduces artifacts in recurring locations across different loops, compared to lookup using temporal neighbors (baseline)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 309, + 367, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 309, + 367, + 319 + ], + "spans": [ + { + "bbox": [ + 314, + 309, + 367, + 319 + ], + "type": "text", + "content": "architecture." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 335, + 555, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 555, + 431 + ], + "type": "text", + "content": "Ablation on two-pass procedural sampling. Fig. 7 shows that one-pass sampling introduces visible temporal flickering and abrupt scene changes. In contrast, interp procedural sampling produces the smoothest transitions, outperforming the alternative " + }, + { + "bbox": [ + 313, + 335, + 555, + 431 + ], + "type": "inline_equation", + "content": "\\mathrm{gt} +" + }, + { + "bbox": [ + 313, + 335, + 555, + 431 + ], + "type": "text", + "content": " nearest strategy and mitigating noticeable flickering. Additional quantitative ablation is deferred to the appendix, including 3DGS distillation and video quality metrics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 437, + 440, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 440, + 451 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 440, + 451 + ], + "type": "text", + "content": "4.4. Long-Trajectory NVS" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 456, + 554, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 554, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 554, + 552 + ], + "type": "text", + "content": "Fig. 8 presents a qualitative demonstration of NVS over a long trajectory of up to 1000 frames. With the memory bank maintaining previously generated anchors, SEVA achieves robust 3D consistency for long-trajectory NVS, e.g., the building in front of and the plantation after the booth. Comparing it to using temporal nearest anchors previously generated, using spatially nearest ones demonstrates a clear advantage in terms of long-term consistency." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 562, + 388, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 562, + 388, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 562, + 388, + 574 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": "We introduce SEVA, a generalist diffusion model for novel view synthesis that balances large viewpoint changes and smooth interpolation while supporting flexible input and target configurations. By designing a diffusion-based architecture without 3D representation, a structured training strategy, and a two-pass procedural sampling approach, SEVA achieves 3D consistent rendering across diverse NVs tasks. Extensive benchmarking demonstrates its superiority over existing methods, with strong generalization to real-world scenes. For broader impact and limitations, please refer to the appendix." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12412" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, pages 405-421, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 180 + ], + "type": "text", + "content": "[2] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 295, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 295, + 225 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 295, + 225 + ], + "type": "text", + "content": "[3] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4), July 2023. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 228, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 228, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 62, + 228, + 294, + 259 + ], + "type": "text", + "content": "[4] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In CVPR, 2021. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 262, + 294, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 262, + 294, + 294 + ], + "spans": [ + { + "bbox": [ + 62, + 262, + 294, + 294 + ], + "type": "text", + "content": "[5] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3D using 2D diffusion. arXiv, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 297, + 295, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 297, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 297, + 295, + 350 + ], + "type": "text", + "content": "[6] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3D object. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 353, + 296, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 353, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 353, + 296, + 418 + ], + "type": "text", + "content": "[7] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21551-21561, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 420, + 295, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 295, + 474 + ], + "type": "text", + "content": "[8] Ruiqi Gao*, Aleksander Holynski*, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul P. Srinivasan, Jonathan T. Barron, and Ben Poole*. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv, 2024. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 476, + 295, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 476, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 476, + 295, + 531 + ], + "type": "text", + "content": "[9] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. ViewCrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 533, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 294, + 588 + ], + "type": "text", + "content": "[10] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1, 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 590, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 294, + 633 + ], + "type": "text", + "content": "[11] Haian Jin, Hanwen Jiang, Hao Tan, Kai Zhang, Sai Bi, Tianyuan Zhang, Fujun Luan, Noah Snavely, and Zexiang Xu. Lvsm: A large view synthesis model with minimal 3d inductive bias, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 294, + 678 + ], + "type": "text", + "content": "[12] Daniel Watson, Saurabh Saxena, Lala Li, Andrea Tagliasacchi, and David J Fleet. Controlling space and time with diffusion models. arXiv preprint arXiv:2407.07860, 2024. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 680, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 294, + 714 + ], + "type": "text", + "content": "[13] Kyle Sargent, Zizhang Li, Tanmay Shah, Charles Herrmann, Hong-Xing Yu, Yunzhi Zhang, Eric Ryan Chan, Dmitry Lagun, Li Fei-Fei, Deqing Sun, et al. Zeronvs: Zero-shot" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "360-degree view synthesis from a single real image. arXiv preprint arXiv:2310.17994, 2023. 1, 2, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 97, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 128 + ], + "type": "text", + "content": "[14] Virtual camera system. https://en.wikipedia.org/wiki/Virtual_camera_system. Accessed: 5-Mar-2025. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 130, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 173 + ], + "type": "text", + "content": "[15] David Charatan, Sizhe Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In arXiv, 2023. 2, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 175, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 230 + ], + "type": "text", + "content": "[16] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In European Conference on Computer Vision, pages 370-386. Springer, 2025. 2, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 232, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 285 + ], + "type": "text", + "content": "[17] Chen Ziwen, Hao Tan, Kai Zhang, Sai Bi, Fujun Luan, Yicong Hong, Li Fuxin, and Zexiang Xu. Long-lrm: Long-sequence large reconstruction model for wide-coverage gaussian splats. arXiv preprint arXiv:2410.12781, 2024. 2, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 287, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 352 + ], + "type": "text", + "content": "[18] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitrii Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. SV3D: Novel multi-view synthesis and 3D generation from a single image using latent video diffusion. In European Conference on Computer Vision, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 354, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 397 + ], + "type": "text", + "content": "[19] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 399, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 431 + ], + "type": "text", + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 433, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 476 + ], + "type": "text", + "content": "[21] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrl: Enabling camera control for text-to-video generation. arXiv preprint arXiv:2404.02101, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "type": "text", + "content": "[22] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. arXiv, 2022. 3, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 512, + 555, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 512, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 317, + 512, + 555, + 566 + ], + "type": "text", + "content": "[23] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, June 2022. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 567, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 600 + ], + "type": "text", + "content": "[24] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. MVDream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "type": "text", + "content": "[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 647, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 669 + ], + "type": "text", + "content": "[26] P Goyal. Accurate, large minibatch sg d: training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[27] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Content" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12413" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 75, + 73, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 73, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 75, + 73, + 294, + 94 + ], + "type": "text", + "content": "ference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "text", + "content": "[28] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 294, + 183 + ], + "type": "text", + "content": "[29] Emiel Hoogeboom, Jonathan Heek, and Tim Salimans. simple diffusion: End-to-end diffusion for high resolution images. arXiv preprint arXiv:2301.11093, 2023. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 237 + ], + "type": "text", + "content": "[30] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICLR, 2024. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "text", + "content": "[31] Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. Depthsplat: Connecting gaussian splatting and depth. arXiv preprint arXiv:2410.13862, 2024. 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 284, + 294, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 284, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 294, + 348 + ], + "type": "text", + "content": "[32] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, et al. Omniobject3D: Large-vocabulary 3D object dataset for realistic perception, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 803-814, 2023. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 350, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 415 + ], + "type": "text", + "content": "[33] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google Scanned Objects: A high-quality dataset of 3D scanned household items. In 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560. IEEE, 2022. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 416, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 294, + 470 + ], + "type": "text", + "content": "[34] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (ToG), 38(4):1-14, 2019. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 472, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 472, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 472, + 294, + 514 + ], + "type": "text", + "content": "[35] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 406-413, 2014." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 515, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 294, + 569 + ], + "type": "text", + "content": "[36] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In International Conference on Computer Vision, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 571, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 571, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 571, + 294, + 602 + ], + "type": "text", + "content": "[37] Hongchi Xia, Yang Fu, Sifei Liu, and Xiaolong Wang. RGBd objects in the wild: Scaling real-world 3d object learning from rgb-d videos, 2024." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 604, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 604, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 604, + 294, + 668 + ], + "type": "text", + "content": "[38] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-NeRF: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 5, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[39] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018. 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 419 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "type": "text", + "content": "[40] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "type": "text", + "content": "[41] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 185, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 228 + ], + "type": "text", + "content": "[42] Tsai-Shien Chen, Chieh Hubert Lin, Hung-Yu Tseng, Tsung-Yi Lin, and Ming-Hsuan Yang. Motion-conditioned diffusion model for controllable video synthesis. arXiv preprint arXiv:2304.14404, 2023. 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 230, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 284 + ], + "type": "text", + "content": "[43] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 285, + 553, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 553, + 328 + ], + "type": "text", + "content": "[44] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 331, + 553, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 384 + ], + "type": "text", + "content": "[45] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 386, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 386, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 386, + 553, + 419 + ], + "type": "text", + "content": "[46] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric, 2018. 7" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12414" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Stable-Sim2Real_ Exploring Simulation of Real-Captured 3D Data with Two-Stage Depth Diffusion/ac3096b4-f6df-4b0e-b9d1-43e46507be43_content_list.json b/2025/Stable-Sim2Real_ Exploring Simulation of Real-Captured 3D Data with Two-Stage Depth Diffusion/ac3096b4-f6df-4b0e-b9d1-43e46507be43_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6b26e00bdb48d87d9d7aa87be3a981e847e835bb --- /dev/null +++ b/2025/Stable-Sim2Real_ Exploring Simulation of Real-Captured 3D Data with Two-Stage Depth Diffusion/ac3096b4-f6df-4b0e-b9d1-43e46507be43_content_list.json @@ -0,0 +1,1652 @@ +[ + { + "type": "text", + "text": "Stable-Sim2Real: Exploring Simulation of Real-Captured 3D Data with Two-Stage Depth Diffusion", + "text_level": 1, + "bbox": [ + 137, + 128, + 859, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mutian $\\mathrm{Xu}^{1}$ Chongjie $\\mathrm{Ye}^{2,3,1}$ Haolin Liu $^{4}$ Yushuang Wu $^{5}$ Jiahao Chang $^{1}$ Xiaoguang Han $^{1,2,3*}$", + "bbox": [ + 99, + 200, + 924, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ SSE, CUHKSZ $^{2}$ FNii-Shenzhen", + "bbox": [ + 383, + 227, + 640, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Guangdong Provincial Key Laboratory of Future Networks of Intelligence, CUHKSZ", + "bbox": [ + 227, + 244, + 797, + 262 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4Tencent Hunyuan3D 5ByteDance Games", + "bbox": [ + 354, + 263, + 665, + 280 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 314, + 326, + 330 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D data simulation aims to bridge the gap between simulated and real-captured 3D data, which is a fundamental problem for real-world 3D visual tasks. Most 3D data simulation methods inject predefined physical priors but struggle to capture the full complexity of real data. An optimal approach involves learning an implicit mapping from synthetic to realistic data in a data-driven manner, but progress in this solution has met stagnation in recent studies. This work explores a new solution path of data-driven 3D simulation, called Stable-Sim2Real, based on a novel two-stage depth diffusion model. The initial stage finetunes Stable-Diffusion to generate the residual between the real and synthetic paired depth, producing a stable but coarse depth, where some local regions may deviate from realistic patterns. To enhance this, both the synthetic and initial output depth are fed into a second-stage diffusion, where diffusion loss is adjusted to prioritize these distinct areas identified by a 3D discriminator. We provide a new benchmark scheme to evaluate 3D data simulation methods. Extensive experiments show that training the network with the 3D simulated data derived from our method significantly enhances performance in real-world 3D visual tasks. Moreover, the evaluation demonstrates the high similarity between our 3D simulated data and real-captured patterns. Project page: mutianxu.github.io/stable-sim2real.", + "bbox": [ + 89, + 348, + 483, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 756, + 220, + 772 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, real-world 3D datasets have played a crucial role in addressing a wide range of tasks in 3D vision and robotics [1, 7, 11, 15, 23, 52, 69, 70]. However, collecting real 3D data is often labor-intensive and time-consuming, and is further complicated by the recent concerns regarding data privacy. In this scenario, synthetic (i.e., simulated)", + "bbox": [ + 89, + 782, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c7f7612d866edb84c842c072879108ad28fd72abb03ac19e782c62dd0c19aab0.jpg", + "image_caption": [ + "Figure 1. We delve back into a substantial yet stagnant problem of simulating real-captured 3D data in a data-driven manner, introducing Stable-Sim2Real. Given the depth maps of the 3D synthetic data, our approach applies a two-stage depth diffusion model based on Stable Diffusion [42] to generate realistic depth maps that are fused to produce real 3D point clouds. The initial stage generates stable but coarse depth maps, followed by a second-stage diffusion to enhance regional details." + ], + "image_footnote": [], + "bbox": [ + 532, + 311, + 888, + 445 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "data [8, 10, 12, 16, 68] emerges as an alternative data resource that is cost-effective, rapid, and scalable. Nevertheless, models trained using synthetic data may not perform robustly against the real-world signal.", + "bbox": [ + 511, + 582, + 906, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This problem has catalyzed the development of 3D data simulation techniques, to minimize the gap between simulated and real-captured 3D data. While some studies [2, 30, 33, 39, 75] have explored grounding physical priors to simulate depth sensors, they struggle to capture the full complexity of real-world patterns and are confined to simulation environments, due to the reliance on predefined and explicit physical modeling. A better solution involves learning an implicit mapping from synthetic to realistic data in a data-driven manner, offering greater flexibility and adaptability to real-world diversities. However, only a handful of earlier works [4, 19, 47, 56] have ventured down this path. Hindered by the lack of synthetic-real paired datasets and the dependence on conventional networks like GANs [18], which are prone to training instability and mode collapse [29], these methods have yet to demonstrate the anticipated performance levels. As a result, developing data-driven 3D", + "bbox": [ + 511, + 643, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 233, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2609", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sim2Real methods has long been a challenge within the community, and has shown little progress.", + "bbox": [ + 89, + 90, + 480, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we aim to explore the data-driven 3D Sim2Real and refocus the community's attention on this critical problem. As a prerequisite, we choose a recent synthetic-real paired dataset, LASA [31], containing 10,412 high-quality 3D shape CAD annotations aligned with the real-world object scans. Driven by LASA, our efforts focus on designing an effective data-driven 3D Sim2Real algorithm. Given the inherent indeterminacy and diversity in real-captured data patterns, deep generative models emerge as a natural fit. Among these models, diffusion models [20, 53-55, 62] recently offer easier training dynamics and stronger generation ability. Yet, due to the lack of 3D data, it is hard to train a 3D diffusion model to get strong 3D priors for 3D simulation. Therefore, we opt to leverage the strong generalization priors of $2D$ diffusion foundation models (e.g., SD - Stable Diffusion [42]) to simulate real 2D depth maps, which are then fused to gain 3D data. This strategy is similar to real 3D data acquisition, where 2D depth is captured and fused into 3D.", + "bbox": [ + 89, + 123, + 482, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, an intuitive baseline is to sample CAD (i.e., synthetic) depth maps and their paired real depth images from LASA, and then finetune SD to learn an implicit mapping between them. Although such a solution powered by large-scale paired data and the foundation diffusion model has shown effectiveness in conventional image translations, it faces special challenges in our task: while conventional image translations (e.g., denoising/restoration) normally remove noise to produce clean images, our task aims to output noisy depth that is highly uncertain, thereby making the distribution that needs to be learned more complex.", + "bbox": [ + 89, + 412, + 482, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose Stable-Sim2Real. The core is a novel two-stage depth diffusion model to simulate real-captured noise (a sample is shown in Fig. 1). At Stage-I diffusion, instead of directly generating the corresponding real-world depth map, our model generates the residual (i.e., difference) between real and CAD depth maps. The simulated depth map is then yielded by adding the generated residual with the CAD depth. Empirically, this residual represents the noise added to clean (i.e., synthetic) depth during real-world scanning. Moreover, compared to directly generating noisy real depth, adding noise to CAD depth, which is inherently clean and view-consistent, can produce more stable depth with smaller view variation and better preservation of the original geometry (see Sec. 3.2 for more discussions and analysis).", + "bbox": [ + 89, + 580, + 482, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, according to our observations and experiments, while some regions within the Stage-I generated depth map successfully fit the realistic pattern, some unsatisfactory areas still exhibit conspicuous 3D geometric discrepancies between the generation and real captures (shown in Fig. 6). To tackle this issue, our model then revolves", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "around discriminating unsatisfactory areas and enhancing the Stage-I generation at local regions. We train a 3D-Aware Discriminator to distinguish between Stage-I generation and real captures at local geometries, which is expected to delineate the distribution boundary between them. If a generated patch can be successfully discriminated, it signifies an unsatisfactory area. Next, we introduce a Stage-II diffusion, conditioned not only on the CAD depth but also on Stage-I generation. More importantly, the diffusion loss is adjusted to prioritize on unsatisfactory regions. In this way, Stage-II diffusion is specialized on learning the distribution of unsatisfactory areas, thereby enhancing Stage-I generation. Finally, the generated depth maps are fused to derive the simulated 3D data. Notably, the binary classifier is only used during the training phase.", + "bbox": [ + 511, + 90, + 903, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Last, we provide a comprehensive benchmark scheme for 3D data simulation. The key logic is: if a model's real-world performance can be improved after being trained with simulated data, it would validate the effectiveness of simulation methods. We focus on fundamental real-world 3D tasks: 3D shape reconstruction and 3D object/scene understanding. As for shape reconstruction, we pretrain a 3D reconstruction network that takes simulated 3D data as input and outputs clean 3D surfaces. For 3D object/scene understanding, the simulated 3D data is used to pretrain a self-supervised point cloud learning framework. To better assess the performance gain purely from simulated data itself, we conduct few-shot evaluation on pre-trained networks.", + "bbox": [ + 511, + 319, + 903, + 515 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as:", + "bbox": [ + 529, + 516, + 782, + 529 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We delve back into a substantial yet stagnant problem, data-driven 3D real data simulation, paving a new solution path of training generative foundation models on extensive synthetic-real paired data.", + "- We present Stable-Sim2Real, a novel 3D real data simulation method based on a two-stage depth diffusion model that can stably simulate real-captured noise.", + "- We provide a comprehensive scheme to benchmark 3D data simulation, based on few-shot evaluation on networks trained with simulated data for real-world 3D tasks.", + "- Extensive qualitative and quantitative results show that the data simulated by our method significantly boosts model performance on real-world 3D tasks." + ], + "bbox": [ + 511, + 532, + 903, + 728 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 744, + 653, + 761 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Simulation of Real-World Data", + "text_level": 1, + "bbox": [ + 511, + 771, + 787, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Simulation of 2D real-world data. In the realm of 2D real-world data simulation, two predominant approaches have emerged: domain randomization and domain adaptation. The first line of works [57, 58, 60] randomly perturbs textures, lighting, and material of synthetic data, thereby modeling the synthetic-real domain gap as one kind of variation to endow models with robustness to real-world data.", + "bbox": [ + 511, + 794, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2610", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, it brought the challenge of achieving a delicate balance between ensuring that the distribution of augmentations adequately covers that of real-world data and preventing the augmentation from being too extensive to hinder effective model training [24, 37, 72]. Domain adaptation, on the other hand, aims to minimize the domain gap by transferring knowledge from synthetic data to real-world scenarios, leveraging techniques such as adversarial training and self-supervised learning [5, 49, 71].", + "bbox": [ + 89, + 90, + 480, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Simulation of 3D real-captured data. Acquiring real-world 3D data and their annotation is labor intensive and inefficient, deriving methods for the simulation of 3D data. Existing simulation methods primarily focus on enhancing the realism of depth information by grounding physical priors in simulators to close the domain gap with 3D data from the physical world [2, 30, 33, 39, 75]. However, these approaches are usually limited by their reliance on predefined and explicit physical modeling. Alternatively, data-driven methods turn to implicitly learn from the distribution in real-captured data, independent of physical modeling [4, 19, 47, 56], offering enhanced adaptability and generalization capabilities to various real-world scenarios. Nonetheless, due to the absence of extensive synthetic-real paired datasets and the reliance on traditional GANs [18], which suffer from training instability and mode collapse [29], these approaches have not achieved the expected performance. Consequently, the research progress of data-driven 3D Sim2Real techniques faces a recent stagnation.", + "bbox": [ + 91, + 244, + 483, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Image Generation and Translation.", + "text_level": 1, + "bbox": [ + 89, + 542, + 397, + 558 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Conventional generative models. We formulate the 3D Sim2Real task as learning the mapping from synthetic to real-world captured noise, based on depth images. Existing approaches in the domain of image generation and translation include a variety of generative models such as Generative Adversarial Networks (GANs) [18], Variational Autoencoders (VAEs) [28], Cycle-Consistent Adversarial Networks (Cycle-GAN) [77], and Pix2Pix [22]. While GAN-based methods, which have been predominant in image generation before diffusion models, achieved significant milestones in Sim2Real tasks [6, 65, 71], they often struggle with maintaining local details and semantic consistency, leading to sub-optimal performance [13, 29, 32, 36, 38].", + "bbox": [ + 89, + 564, + 482, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models. Diffusion models, also known as denoising diffusion probabilistic models (DDPMs) [50], have emerged as a powerful alternative to GANs for image synthesis [13, 32, 34, 35, 43, 44, 74]. Unlike GANs, diffusion models offer a more stable training procedure and have been shown to generate images with higher fidelity and better preservation of fine details [21, 42]. Depth diffusion models recently track great attention. They mainly focus on the task", + "bbox": [ + 89, + 779, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of depth estimation in a monocular [14, 26, 45, 46, 59, 64] or multi-view [27, 67] manner. Other works also cover conventional image translation tasks, such as depth map superresolution [48] and refinement [76], which produce clean and certain output images. Differently, our work leverages generative models to \"generate noise\", where the expected output is highly uncertain. Thus, the distribution that needs to be learned is more complex, bringing fundamentally different challenges.", + "bbox": [ + 511, + 90, + 903, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 239, + 604, + 255 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We hope to exploit the strong prior of the diffusion-based foundation model, so we take depth map as our bridge towards the simulation of real-captured 3D data. The preliminaries on the diffusion model are provided in Sec. 3.1. Our Stable-Sim2Real is a two-stage depth diffusion framework, as illustrated in Fig. 2. The first-stage model (Sec. 3.2) learns to generate the depth residual, producing stable but coarse depth maps, where some areas are distinct from the realistic pattern. Therefore, a 3D binary classifier (Sec. 3.3) is introduced to discriminate these distinct areas. The second-stage model (Sec. 3.4) is specialized to enhance the generation of these areas. Finally, the enhanced depth maps are fused to obtain the 3D simulated data (Sec. 3.5).", + "bbox": [ + 511, + 263, + 906, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries on Diffusion Models", + "text_level": 1, + "bbox": [ + 511, + 469, + 812, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our model is based on Denoising Diffusion Probabilistic Models (DDPM) [20, 51]. The key of DDPM is to model a data distribution $p(x)$ via building a Markov chain in the data space. It can be divided into two main processes, forward process and backward process.", + "bbox": [ + 511, + 491, + 905, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Forward diffusion process. Given an initial data $x_0$ , we gradually add random noise $\\epsilon \\sim \\mathcal{N}(0, \\mathbf{I})$ to $x_0$ , yielding a series of noisy data, which can be simply formulated as $x_{t+1} = x_t + \\epsilon_t$ . After adding noise for multiple steps, the initial distribution of $p(x)$ is expected to be transformed into a standard Gaussian distribution.", + "bbox": [ + 511, + 585, + 905, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "According to DDPM [20], this process can be directly achieved by the following forward diffusion process:", + "bbox": [ + 511, + 675, + 905, + 705 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(x _ {t}\\right) = \\sqrt {\\alpha_ {t}} x _ {0} + \\sqrt {1 - \\alpha_ {t}} \\boldsymbol {\\epsilon}, x _ {0} \\sim p (x), t \\in \\{0, 1, \\dots , T \\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 714, + 903, + 746 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $T$ denotes the number of the time step, $t$ is the current time step, and $\\alpha_{t}$ regulates the noise schedule deciding the pace at which the initial data distribution transforms into a standard Gaussian distribution.", + "bbox": [ + 511, + 747, + 905, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Backward diffusion process. Given a noisy $x_{t}$ from a standard Gaussian distribution as the initial state, $x_{t}$ is denoised to a cleaner $x_{t-1}$ at each time step, which can also be simply formulated as $x_{t-1} = x_{t} - \\mu_{\\theta}^{\\epsilon}(x_{t}, t) + \\epsilon_{t}$ in which $\\epsilon_{t} \\sim \\mathcal{N}(0, \\sigma_{\\mathbf{t}} \\mathbf{I})$ and $\\mu_{\\theta}^{\\epsilon}$ is a learnable network to predict", + "bbox": [ + 511, + 825, + 905, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2611", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1d460675239563ae4e71c0ad18d35994344d4f1bde9eeaf6a4f9ba27d8b9c2a2.jpg", + "image_caption": [ + "Figure 2. An overview of our Stable-Sim2Real. Conditioned on CAD (i.e., synthetic) depth maps, our model first generates the residual (i.e., difference) between real-captured and CAD depth maps. By adding the generated residual to the CAD depth, Stage-I diffusion produces stable but coarse depth maps (Sec. 3.2). Next, the Stage-I output is projected into 3D and sent to a 3D discriminator, to identify the unsatisfactory local areas that are distinct from real-captured patterns (Sec. 3.3). Later, our Stage-II diffusion is conditioned on both CAD and Stage-I output, where the training loss of Stage-II is re-weighted to be specialized on enhancing unsatisfactory areas of Stage-I output (Sec. 3.4). Notably, the local discriminating and loss re-weighting are only used during training." + ], + "image_footnote": [], + "bbox": [ + 135, + 88, + 867, + 286 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the injected noise. By repeating this reverse process until the maximum number of steps $T$ , we can obtain the final generated data.", + "bbox": [ + 89, + 395, + 482, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In DDPM [20], the backward diffusion process is proven to be:", + "bbox": [ + 89, + 441, + 482, + 470 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} x _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {\\alpha_ {t} \\left(1 - \\Pi_ {\\tau = 0} ^ {t} \\alpha_ {\\tau}\\right)}} \\mu_ {\\theta} ^ {\\epsilon} \\left(x _ {t}, t;\\right) + \\sigma_ {t} \\epsilon , \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 477, + 480, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sigma_{t}$ denotes the standard deviation of $q(x_{t - 1}|x_t,x_0)$ in the backward diffusion to sample random noise. The loss for training $\\mu_{\\theta}^{\\epsilon}$ is:", + "bbox": [ + 89, + 537, + 482, + 584 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\theta} = \\mathbb {E} _ {\\epsilon , t} \\| \\epsilon - \\mu_ {\\theta} ^ {\\epsilon} (x _ {t}, t) \\| ^ {2} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 593, + 482, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Stage-I: Initial Depth Generation", + "text_level": 1, + "bbox": [ + 89, + 622, + 382, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Residual depth generation with SD. Different from unconditional diffusion models, text-to-image diffusion models focus on generating images with optional text prompts. A representative is Stable Diffusion (SD) [42]. SD trains a denoising U-Net architecture $\\mu_{\\theta}(x_t,t,c)$ on the latent space of a pre-trained Variational Auto-Encoder (VAE). Here, $c$ represents the supplementary text prompt embedding, usually acquired through methods like CLIP [41].", + "bbox": [ + 89, + 643, + 482, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We finetune SD model on synthetic-real paired depth data from LASA [31]. Specifically, given an initial synthetic depth map $S$ and its paired real-captured depth map $R$ from a pre-defined dataset, we first get the initial ground truth residual (i.e., difference) depth map $D_{res}^{gt} = S - R$ . Then we follow ControlNet [73] to add an additional encoder $En$ and decoder $\\mu_{\\theta}^{\\epsilon}$ for finetuning SD. Specifically, we run the forward process to add noise on $D_{res}^{gt}$ following Eq. (1) and encode with $En$ to obtain $x_{t}$ . Next, we send $S$", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to a pretrained VAE encoder $En$ , generating the latent condition signal. This latent signal $En(S)$ is further processed by an additional encoder $\\pmb{f}_{\\phi}$ and sent to the decoder blocks of denoising U-Net. We train both the decoder blocks $\\mu_{\\theta}^{\\epsilon}$ and encoder $\\pmb{f}_{\\phi}$ using the following loss (based on Eq. (3)):", + "bbox": [ + 511, + 395, + 906, + 472 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\theta , \\phi} = \\mathbb {E} _ {\\epsilon , S, t} \\| \\epsilon - \\boldsymbol {\\mu} _ {\\theta} ^ {\\epsilon} (\\boldsymbol {x} _ {t}, t, \\boldsymbol {f} _ {\\phi} (E n (\\boldsymbol {S}))) \\| ^ {2} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 483, + 906, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Learning depth residual helps to produce stable depth. Empirically, this residual represents the noise added to clean (i.e., synthetic) depth during real-world scanning. Moreover, compared to directly generating noisy real depth, adding noise to CAD depth, which is inherently clean and view-consistent, can produce more stable depth with smaller view variation and better preservation of the original geometry. From experimental results, learning residual indeed performs better (see Tab. 5).", + "bbox": [ + 511, + 518, + 906, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Learning depth residual is theoretically identical to learning real depth. First, assume that real depth is a sample from a Gaussian distribution, denoted by $R \\sim \\mathcal{N}(\\alpha_{\\mathbf{R}}, \\vartheta)$ , where $\\alpha$ denotes the mean and $\\vartheta$ represents the variance. Notably, $\\vartheta$ can naturally represent real-captured noise. Therefore, synthetic depth can be denoted by $S \\sim \\mathcal{N}(\\alpha_{\\mathbf{S}}, 0)$ , where the variance is zero as there is no real-captured noise.", + "bbox": [ + 511, + 672, + 906, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our design, as the diffusion model is inherently designed to learn the transformation from a random noise $\\mathcal{N}(\\mathbf{0},\\mathbf{I})$ to any other Gaussian distribution [53], we leverage it to model the distribution of the depth residual that can be denoted by $\\mathcal{N}(\\alpha_{\\mathbf{R}} - \\alpha_{\\mathbf{S}},\\vartheta)$ . Last, by adding the depth residual to synthetic depth, the output distribution can be formulated as: $\\mathcal{N}(\\alpha_{\\mathbf{R}} - \\alpha_{\\mathbf{S}},\\vartheta) + \\mathcal{N}(\\alpha_{\\mathbf{S}},0) = \\mathcal{N}(\\alpha_{\\mathbf{R}},\\vartheta)$ ,", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2612", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which is exactly identical to the previously assumed real-captured noise distribution.", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. 3D-Aware Local Discriminating", + "text_level": 1, + "bbox": [ + 89, + 128, + 372, + 143 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unsatisfactory areas distinct from real patterns. Based on our observations, the noise captured in real data is spatially-variant, meaning different areas within a real depth map may display varying levels of perturbation. However, our Stage-I diffusion tends to introduce spatially-uniform noise across the entire input CAD depth. This discrepancy may stem from a bias within SD, as also discussed in [9]. As a result, although certain regions of the generated depth map align well with real perturbation patterns, there remain noticeable areas that deviate from reality. This leads to evident 3D geometric discrepancies between Stage-I outputs and real captures (as illustrated in Fig. 6).", + "bbox": [ + 89, + 150, + 483, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Identifying unsatisfactory areas with 3D discriminator.", + "text_level": 1, + "bbox": [ + 89, + 345, + 483, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our design logic to discriminate unsatisfactory areas is that if the generated areas diverge significantly from real patterns, there should exist a distribution boundary of their 3D geometries. To this end, we propose employing a 3D binary classifier to discriminate this boundary. Initially, we split all generated and real-captured depth maps into different patches representing different areas, which are then projected to form 3D point patches. Subsequently, we leverage PointNet [40] as a 3D binary classifier to differentiate whether the input point patches come from model generation or real captures. The CAD depth maps are also concatenated in the input as additional guidance. The successful classification of a generated patch indicates that its corresponding area is distinct from real captures. Notably, the binary classifier is solely used during training.", + "bbox": [ + 88, + 361, + 482, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Stage-II: Unsatisfactory Area Enhancement", + "text_level": 1, + "bbox": [ + 89, + 594, + 465, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Improving unsatisfactory areas by re-weighting denoising loss. After identifying these local patches, we introduce a second-stage depth refinement model that differs from the Stage-I diffusion in two crucial aspects: 1) The diffusion model is now conditioned on the concatenation of synthetic depth $S$ and the previously generated depth residual $D^{I}$ from Stage I. Here, $D^{I} = D_{res}^{I} + S$ , where $D_{res}^{I}$ is the output residual depth from Stage-I model. Thus, the training loss defined in Eq. (4) is updated to:", + "bbox": [ + 89, + 616, + 483, + 752 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\chi , \\psi} = \\mathbb {E} _ {\\boldsymbol {\\epsilon}, \\boldsymbol {D} ^ {I}, \\boldsymbol {S}, t} \\left\\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\mu} _ {\\chi} ^ {\\epsilon} \\left(\\boldsymbol {x} _ {t}, t, \\boldsymbol {f} _ {\\psi} \\left(E n \\left(\\boldsymbol {D} _ {r e s} ^ {I}, \\boldsymbol {S}\\right)\\right)\\right) \\right\\| ^ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 758, + 480, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2) More importantly, the diffusion loss is re-weighted to prioritize the enhancement of the identified unsatisfactory regions, making the final Stage-II loss $L_{\\chi ,\\psi}$ to be:", + "bbox": [ + 89, + 792, + 482, + 838 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\chi , \\psi} = \\omega L _ {\\chi , \\psi} ^ {s i m} + \\lambda L _ {\\chi , \\psi} ^ {d i s}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 842, + 482, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $L_{\\chi, \\psi}^{sim}$ indicates the loss on similar regions, while $L_{\\chi, \\psi}^{dis}$ denotes the loss on previously identified distinct (i.e.,", + "bbox": [ + 89, + 867, + 483, + 904 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "unsatisfactory) areas. $\\omega$ and $\\lambda$ are weight coefficients, in which $\\lambda >\\omega$ . In our experiment, setting $\\omega = 0.5$ and $\\lambda = 1.5$ produces the best result.", + "bbox": [ + 511, + 90, + 903, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our Stage-II diffusion model is specialized on learning the distribution of unsatisfactory areas in Stage-I generation. Meanwhile, by conditioning on Stage-I output and assigning less weight to satisfied areas, the generation on satisfied outputs is also preserved in Stage-II, as in Fig. 6.", + "bbox": [ + 511, + 136, + 903, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Creating Real-World 3D Data", + "text_level": 1, + "bbox": [ + 511, + 219, + 779, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Two-stage real depth simulation. In the initial stage, we sample a random distribution $\\epsilon$ from $\\mathcal{N}(\\mathbf{0},\\mathbf{I})$ and feed it into the denoising U-Net $\\mu_{\\theta}$ of Stage-I diffusion model. The model runs a backward diffusion process for $T$ steps to generate the initial depth residual $D_{res}^{I}$ , following Eq. (2) with the synthetic depth $S$ serving as the condition. Now we can obtain the initial depth map $D^{I} = D_{res}^{I} + S$ .", + "bbox": [ + 511, + 241, + 903, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Moving to the second stage, we repeat a similar backward diffusion process to generate the residual depth $D_{res}^{II}$ , using the denoising U-Net $\\mu_{\\chi}$ of our Stage-II diffusion model, conditioned on both synthetic depth $S$ and Stage-I output $D^{I}$ . The final depth map $D^{II} = D_{res}^{II} + S$ . DDIM sampler [51] is utilized during the whole inference phase.", + "bbox": [ + 511, + 348, + 903, + 438 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Depth fusion to 3D data. Given the 3D synthetic object, we follow LASA [31] to embed CAD objects into 3D scenes. This allows us to render synthetic depth images, simulating the occlusions in real-world captures. Then we conduct the aforementioned two-stage real depth simulation on rendered synthetic depth maps, followed by a depth fusion algorithm to create simulated 3D data.", + "bbox": [ + 511, + 455, + 905, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 573, + 645, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We provide a new benchmark scheme for 3D data simulation. To begin with, different simulation methods are performed on synthetic datasets. The 3D simulated data are then utilized to train networks for tackling fundamental real-world 3D visual tasks: 3D instance reconstruction, 3D object/scene understanding. The key logic of our evaluation is: if the model's real-world performance can be improved after being trained with simulated data, it would validate that the simulated data is close to real distributions, demonstrating the effectiveness of simulation methods.", + "bbox": [ + 511, + 599, + 905, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 757, + 653, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training of Stable-Sim2Real. We train our model by finetuning the official StableDiffusion V2.1 [42]. The training dataset is a recent 3D synthetic-real paired dataset, LASA [31]. LASA is a large-scale dataset that contains 10,412 unique CAD models across 17 categories. Skilled artists are employed to meticulously craft aligned CAD models from 3D object real-world scans of [3]. LASA provides CAD-real depth map pairs, rendered based on camera", + "bbox": [ + 511, + 779, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2613", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e8b2ccee144e33d3fdae5008b96e4e0aa286e3bbb045bc59c378bb27bdf5e207.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodChairSofaTableCabinetBedShelf
w/o sim. data24.5 / 12.3 / 20.155.6 / 7.02 / 19.825.8 / 19.3 / 20.357.1 / 8.83 / 24.140.1 / 6.14 / 17.312.8 / 6.74 / 23.7
Cycle-GAN [77]25.1 / 12.1 / 20.955.2 / 6.98 / 20.526.1 / 19.1 / 19.556.5 / 8.24 / 25.140.6 / 5.98 / 16.913.1 / 6.08 / 24.9
Pix2Pix [22]24.9 / 11.0 / 20.556.4 / 6.74 / 21.326.6 / 18.5 / 22.157.4 / 7.83 / 25.541.3 / 6.01 / 17.814.2 / 6.15 / 26.2
Giga-GAN [25]26.2 / 10.3 / 22.158.9 / 5.87 / 22.429.3 / 16.4 / 24.360.5 / 7.01 / 26.944.2 / 5.25 / 19.318.2 / 5.87 / 27.8
Stable-Diffusion [42]27.3 / 9.51 / 25.559.4 / 5.93 / 23.629.7 / 15.8 / 24.761.6 / 6.92 / 27.545.1 / 5.31 / 20.119.1 / 5.85 / 28.3
Ours29.2 / 7.83 / 27.562.9 / 4.09 / 25.131.2 / 14.1 / 26.863.2 / 5.58 / 28.646.3 / 4.23 / 21.818.8 / 5.02 / 31.6
", + "bbox": [ + 135, + 88, + 867, + 188 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1862514e1cf0661cdfd021754bc5447ce38b48fa1fe74f0f4a693dea0ddbb046.jpg", + "table_caption": [ + "Table 1. Quantitative comparison of few-shot evaluation on 3D instance reconstruction, where the model is pretrained using simulated data generated by different methods. \"w/o sim. data\" indicates pretraining using no simulated data but only synthetic data. Evaluation metrics are mIoU / Chamfer L2 / F-score respectively. Higher mIoU and F-score are better while lower Chamfer L2 is better. Chamfer L2 is scaled by 1,000. All results are reproduced using the official code of DisCo [31]." + ], + "table_footnote": [], + "table_body": "
MethodChairSofaTableCabinetBedShelf
w/o sim. data19.4 / 14.7 / 15.846.2 / 9.92 / 15.720.5 / 23.6 / 15.951.2 / 11.3 / 20.233.6 / 13.1 / 10.911.1 / 9.23 / 21.6
Cycle-GAN [77]20.3 / 13.1 / 16.548.9 / 9.03 / 18.923.2 / 22.4 / 17.551.9 / 10.8 / 21.435.2 / 12.5 / 12.111.7 / 8.87 / 23.5
Pix2Pix [22]19.3 / 12.3 / 17.449.7 / 9.51 / 18.822.6 / 21.5 / 18.852.7 / 11.2 / 21.137.7 / 10.0 / 13.812.5 / 8.91 / 24.3
Giga-GAN [25]22.4 / 11.8 / 19.553.0 / 7.69 / 19.423.1 / 19.3 / 21.654.3 / 10.1 / 21.940.3 / 7.95 / 15.414.2 / 7.81 / 26.9
Stable-Diffusion [42]23.3 / 11.7 / 20.454.1 / 7.98 / 19.224.2 / 17.9 / 22.655.1 / 9.07 / 22.440.4 / 7.89 / 16.215.8 / 6.72 / 28.3
Ours27.1 / 9.31 / 24.860.6 / 5.29 / 22.328.5 / 16.1 / 24.760.6 / 7.51 / 25.943.2 / 6.23 / 18.017.3 / 5.93 / 30.5
", + "bbox": [ + 151, + 268, + 849, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative comparison of zero-shot evaluation on 3D instance reconstruction, when the model is solely pretrained using simulated data generated by different methods.", + "bbox": [ + 89, + 375, + 906, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "information during the real-world collection. See the supplementary material for more implementation details.", + "bbox": [ + 89, + 417, + 482, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Methods in comparison. Since our method operates on depth images, we select several image-to-image generation methods for comparison: 1) Cycle-Consistent Adversarial Networks (Cycle-GAN) [77], which is a representative generative network when paired training data are not available, so we train Cycle-GAN on the pair-shuffled LASA training data. 2) Pix2Pix [22], which is a classical image generation method. 3) Giga-GAN [25], a state-of-the-art GAN-based method. 4) Stable Diffusion [42], our main baseline model. Both 2), 3), 4) are trained using the paired data from LASA.", + "bbox": [ + 89, + 472, + 482, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.3D Instance Reconstruction", + "text_level": 1, + "bbox": [ + 89, + 635, + 339, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Setup. 3D instance reconstruction aims to recover the full geometry from real-captured partial point clouds or multiview images. We utilize a recent state-of-the-art diffusion-based reconstruction network, DisCo, which is proposed in LASA [31]. It is originally pretrained on three synthetic datasets (i.e., ShapeNet [8], ABO [10], and 3D-Future [17]) and finetuned on LASA real training data. During pretraining, DisCo simply constructs partial point clouds by adding random perturbation to synthetic data. Instead, in our experiment, to simulate partial point clouds for pretraining, we perform simulation using our method and other baseline models based on the original three synthetic datasets used to pretrain DisCo, and replace half of the original partial point clouds with the simulated 3D data. The other configurations all follow the original DisCo.", + "bbox": [ + 89, + 657, + 482, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Subsequently, we employ few-shot and zero-shot learn", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing configurations to evaluate: We randomly sample only a few LASA real training data to finetune the pretrained network. In this way, the real-world knowledge of the network will mostly come from the pretraining data (i.e., the simulated data produced by our method and others) instead of the finetuning data, which can better reflect the performance gain from the simulated data itself. The pretrained model is finetuned with $20\\sim 30\\%$ of the original training samples for few-shot learning and is evaluated on the original LASA validation set. For zero-shot learning, we directly evaluate the pretrained network without fine-tuning.", + "bbox": [ + 511, + 417, + 906, + 584 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Why not perform simulation on LASA synthetic data and compare the performance using the simulated data with that using LASA paired real GT? Since our StableSim2Real and other baseline techniques are trained on the LASA training set, it is unfeasible to conduct further inference on LASA, as this would not verify the actual model performance but only reflect the performance caused by overfitting. We can also not conduct simulation on a third-party paired dataset similar to LASA for 3D reconstruction, as there is currently no such paired dataset.", + "bbox": [ + 511, + 585, + 908, + 737 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative results. Following [31], we utilize mean Intersection over Union (mIoU), L2 chamfer distance, and $1\\%$ F-score metrics. Initially, we normalize both the results and the ground-truth CAD objects to fall within the range of -0.5 to 0.5, according to which we calculate the aforementioned metrics between them. As shown in Tab. 1, under all few-shot settings, the model pretrained using the simulated data from our method achieves the best performance, and significantly surpasses the model using no simulated data (i.e., only add random perturbation to synthetic data) for", + "bbox": [ + 511, + 750, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2614", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/414b82cce4a2b9fa53f937d90d7bdcabbc409f262a00e4fe4c768b326cc6e2da.jpg", + "image_caption": [ + "Figure 3. Qualitative comparison of few-shot evaluation on 3D instance reconstruction. \"w/o sim.\" denotes using no simulated data for pretraining." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 480, + 231 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "pretraining. Moreover, using the simulated data produced by Cycle-GAN or Pix2Pix does not improve the model performance compared to that using no simulated data, possibly attributed to their inferior simulation quality, introducing ambiguity and complexity during pretraining, making it hard for the model to learn useful real-world knowledge. Moreover, as listed in Tab. 2, under zero-shot setting, when solely using simulated data for pretraining without any finetuning, the model trained with our simulated data demonstrates more conspicuous effectiveness.", + "bbox": [ + 88, + 299, + 483, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative results. Fig. 3 shows the qualitative comparison, where our simulated data can help the model to gain the best reconstruction accuracy and diversity. More qualitative results are provided in the supplementary material.", + "bbox": [ + 88, + 459, + 483, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. 3D Object/Scene Understanding", + "text_level": 1, + "bbox": [ + 89, + 530, + 375, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Setup. We select OcCo [63] as our backbone, which is a self-supervised point cloud pretraining framework. OcCo is originally pretrained on ModelNet [68], and finetuned on real-scanned datasets - ScanObjectNN [61] for object classification, and S3DIS [1] for scene semantic segmentation. ModelNet contains 12,311 synthetic CAD objects from 40 classes. ScanObjectNN includes $\\sim 15,000$ real-scanned objects across 15 categories. S3DIS encompasses 271 indoor rooms and 13 classes. The classification task outputs the class of the whole 3D object, and the semantic segmentation predicts the category of each point in the 3D scene.", + "bbox": [ + 88, + 551, + 482, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In our experiment, we apply our method and other baselines on ModelNet to generate simulated data that are further used to replace half of the original ModelNet synthetic data to pretrain OcCo. Next, the pretrained model is again evaluated under the few-shot learning setting. A typical setting is \"K-way $N$ -shot\", which indicates that $K$ classes are randomly selected during finetuning, and each category contains $N$ samples. The finetuned model is evaluated on the unseen samples picked from each of the $K$ classes. We select DGCNN [66] as the backbone for OcCo. For few-shot learning, we employ the \"5-way, 10-shot\" setting. For S3DIS, we follow the original OcCo to report the results", + "bbox": [ + 88, + 719, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8d040676f80bf275281d976a4104927d21a993e25e333704013cf643a3f52223.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methodw/o sim.C-GAN [77]P2P [22]G-GAN [25]SD [42]Ours
Acc. (%)72.472.572.174.274.677.5
mIoU (%)47.948.047.749.148.951.2
", + "bbox": [ + 534, + 89, + 883, + 135 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/408c2c2736329846d7b13f2e54908d51e2fbc90eaf7aceec492f30b4ffc57c42.jpg", + "table_caption": [ + "Table 3. Quantitative comparison of few-shot evaluation on 3D object classification and 3D scene semantic segmentation. \"w/o sim\" means using no simulated data for pretraining." + ], + "table_footnote": [], + "table_body": "
Methodw/o sim.C-GAN [77]P2P [22]G-GAN [25]SD [42]Ours
Metrics59.2/3.54/25.360.4/3.21/25.261.5/3.02/26.462.1/2.86/26.762.3/2.83/27.562.9/2.57/28.2
", + "bbox": [ + 514, + 193, + 903, + 220 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. The mIoU / Chamfer L2 / F-score between simulated data and real-captures on LASA [31] validation set.", + "bbox": [ + 511, + 228, + 906, + 257 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "using 6-fold cross-validation. All the other configurations follow the original OcCo.", + "bbox": [ + 511, + 273, + 906, + 306 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative results. We report the classification accuracy and segmentation mean Intersection over Union (mIoU) in Tab. 3. The model pretrained using our simulated data gets the best results across both 3D object classification and 3D scene understanding tasks. By pretraining with our simulated data, the model can acquire valuable real-world knowledge that helps downstream tasks. In contrast, pretraining with the low-quality simulated data from CycleGAN or Pix2Pix may even degrade the performance.", + "bbox": [ + 511, + 316, + 906, + 454 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Is the simulated data close to real captures?", + "text_level": 1, + "bbox": [ + 511, + 465, + 883, + 482 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As previously mentioned, our Stable-Sim2Real is trained on LASA training set. In this section, we aim to evaluate the data simulation quality by directly performing inference and evaluation on LASA validation set. Tab. 4 compares the mIoU, Chamfer L2, and F-score metrics between the simulated data generated by different methods and the real captured Ground Truth (GT). Although not precise, this result can partially prove that our simulated data closely resembles real-world captures. Fig. 4 shows simulation results on depth maps and 3D point clouds. It demonstrates that while our simulated data closely aligns with real-world captures, subtle differences persist, highlighting the diversity.", + "bbox": [ + 511, + 487, + 905, + 669 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also provide the simulation results on ShapeNet [8] in Fig. 5. Despite the absence of real-captured Ground Truth (GT) in ShapeNet, the outcomes still show that the simulated results from our output closely resemble real captures, where the generated geometry is not largely broken, with certain realistic perturbations introduced.", + "bbox": [ + 511, + 669, + 906, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Studies and Analysis", + "text_level": 1, + "bbox": [ + 511, + 771, + 782, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Module impacts. To evaluate the impact of each key design in our framework, we conduct ablation studies on the 3D instance reconstruction task, specifically on the table category. Similar to Tab. 2, we conduct experiments under zero-shot protocol. This allows for a focused evaluation of the module's impact without the influence of additional factors. Other setups remain consistent with Sec. 4.2.", + "bbox": [ + 511, + 794, + 905, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2615", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/67b8575547711c1a0e0a0c8bc82a52321364a1a7a3c0690e3cdec57bd75550dd.jpg", + "image_caption": [ + "Figure 4. Qualitative simulation results on LASA [31]." + ], + "image_footnote": [], + "bbox": [ + 93, + 85, + 480, + 227 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e9db5cfe08a8b76f9d87cca1c13c177fcb31b0ffe4d5424382446be9a81cddb7.jpg", + "image_caption": [ + "Figure 5. Qualitative simulation results on ShapeNet [8]." + ], + "image_footnote": [], + "bbox": [ + 93, + 263, + 478, + 470 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) Residual generation. In Tab. 5 (w/o res.), directly generating the real-world depth map without using residual generation degrades the model performance, verifying the efficacy of our residual generation.", + "2) Loss re-weighting. In our Stage-II model, we reweight the diffusion loss to focus more on unsatisfactory areas. In Tab. 5 (w/o re-w.), the result w/o loss re-weighting is provided, showing its effectiveness.", + "3) 3D binary discriminator. Instead of training a 3D binary discriminator, another simple way is to directly set a threshold of the absolute difference between Stage-I generation and real-world GT to decide distinct areas. We tried various difference thresholds and reported the best result in Tab. 5 (w/o bin.), which hurts the model performance. To explain, our output depth is gained by adding subtle noise to synthetic depth. Thus, the magnitude of the difference between our output and its corresponding real GT is similar at 2D/pixel space, where it is impractical to find an appropriate threshold to identify distinct/unsatisfactory areas. Instead, their difference in distribution space is relatively more significant, so using a 3D binary classifier to delineate their distribution boundary is a more feasible way.", + "4) Number of stages. Theoretically, we can iteratively improve the generation quality by adding more stages. However, as shown in Tab. 5, adding Stage-III model does" + ], + "bbox": [ + 88, + 517, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/442c5b995a84f9f0cac3839eae86a18d120d115511d351525e6a187545e82139.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methodw/o res.w/o re-w.w/o bin.Stage-IIIStage-IVStage-II
Metrics26.4/18.2/22.127.1/17.9/23.327.3/17.5/23.628.6/ 16.1/24.528.0/16.9/24.128.5/16.1/24.7
", + "bbox": [ + 517, + 89, + 905, + 113 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Quantitative ablations of zero-shot evaluation on 3D instance reconstruction of table category. Evaluation metrics are mIoU / Chamfer L2 / F-score respectively.", + "bbox": [ + 511, + 126, + 906, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2ab4154292e764e9b073cdd8aaee375cb02008785b2a33da79a2a12fd61425a9.jpg", + "image_caption": [ + "Figure 6. Qualitative simulation results at the patch level." + ], + "image_footnote": [], + "bbox": [ + 575, + 181, + 848, + 311 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "not improve the model performance, and Stage-IV model even slightly degrades the result. To analyze, Stage-II model has been sufficient to identify most of the unsatisfactory areas and improve their generation quality. In addition, adding redundant stages may also cause overfitting issues that hurt the performance.", + "bbox": [ + 511, + 353, + 906, + 444 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "Analyzing Stage-II model at 3D patch-level. 1) We compare simulation results with real-captured Ground Truth (GT) at the patch level. As illustrated in Fig. 6, our Stage-II model effectively refines previously unsatisfactory (i.e., dissimilar to GT) patches from Stage-I output to be similar to GT, while keeping previously satisfactory (i.e., similar to GT) patches almost unchanged.", + "2) When sending the newly generated output patches from Stage-II to our 3D binary discriminator, the classification accuracy notably drops from $67.8\\%$ to $32.5\\%$ , suggesting that the generated patches successfully approximate the real-captured distributions, making it challenging for the classifier to discriminate. This again verifies the effectiveness of our Stage-II model. See the supplementary material for more ablations of the binary classifier." + ], + "bbox": [ + 511, + 452, + 905, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 694, + 633, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce Stable-Sim2Real, a two-stage depth diffusion model for 3D real data simulation. Initially, our method learns the residual between synthetic and real depth to generate a stable but coarse depth, where some areas may deviate from realistic patterns. Next, both synthetic and initial output depth are fed into the second-stage model, where diffusion loss prioritizes enhancing distinct regions identified by a binary classifier. Experiments show that the 3D simulated data generated from our method can improve the network performance in real-world 3D tasks, and our simulated data is highly similar to real captures. More limitations are discussed in the supplementary material.", + "bbox": [ + 509, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2616", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The work was supported in part by Guangdong S&T Programme with Grant No. 2024B0101030002, the Basic Research Project No. HZQB-KCZYZ-2021067 of Hetao Shenzhen-HK S&T Cooperation Zone, Guangdong Provincial Outstanding Youth Fund (No. 2023B1515020055), No. NSFC 62172348, the Shenzhen Outstanding Talents Training Fund 202002, the Guangdong Research Projects No. 2017ZT07X152 and No. 2019CX01X104, the Guangdong Provincial Key Laboratory of Future Networks of Intelligence (Grant No. 2022B1212010001), and the Shenzhen Key Laboratory of Big Data and Artificial Intelligence (Grant No. ZDSYS201707251409055).", + "bbox": [ + 89, + 116, + 485, + 297 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 314, + 187, + 330 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] I. Armeni, O. Sener, A. R. Zamir, H. Jiang, I. Brilakis, M. Fischer, and S. Savarese. 3d semantic parsing of large-scale indoor spaces. In CVPR, 2016. 1, 7", + "[2] Kaixin Bai, Lei Zhang, Zhaopeng Chen, Fang Wan, and Jianwei Zhang. Close the sim2real gap via physically-based structured light synthetic data simulation. In ICRA, 2024. 1, 3", + "[3] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Yuri Feigin, Peter Fu, Thomas Gebauer, Daniel Kurz, Tal Dimry, Brandon Joffe, Arik Schwartz, and Elad Shulman. ARK-scenes: A diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In NeurlPS, 2021. 5", + "[4] Konstantinos Bousmalis, Nathan Silberman, David Dohan, Dumitru Erhan, and Dilip Krishnan. Unsupervised pixel-level domain adaptation with generative adversarial networks. In CVPR, 2017. 1, 3", + "[5] Konstantinos Bousmalis, Alex Irpan, Paul Wohlhart, Yunfei Bai, Matthew Kelcey, Mrinal Kalakrishnan, Laura Downs, Julian Ibarz, Peter Pastor, Kurt Konolige, et al. Using simulation and domain adaptation to improve efficiency of deep robotic grasping. In ICRA, 2018. 3", + "[6] Marc Bresson, Yang Xing, and Weisi Guo. Sim2real: Generative ai to enhance photorealism through domain transfer with gan and seven-chanel- $360^{\\circ}$ -paired-images dataset. Sensors, 24(1):94, 2023. 3", + "[7] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. In 3DV, 2017. 1", + "[8] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 1, 6, 7, 8", + "[9] Hansheng Chen, Ruoxi Shi, Yulin Liu, Bokui Shen, Jiayuan Gu, Gordon Wetzstein, Hao Su, and Leonidas Guibas. Generic 3d diffusion adapter using controlled multi-view editing. arXiv preprint arXiv:2403.12032, 2024. 5" + ], + "bbox": [ + 99, + 339, + 480, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksen, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In CVPR, 2022. 1, 6", + "[11] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 1", + "[12] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, 2023. 1", + "[13] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. In NeurlPS, 2021. 3", + "[14] Yiquan Duan, Xianda Guo, and Zheng Zhu. Diffusiondepth: Diffusion denoising approach for monocular depth estimation. In ECCV, 2025. 3", + "[15] Hao-Shu Fang, Chenxi Wang, Hongjie Fang, Minghao Gou, Jirong Liu, Hengxu Yan, Wenhai Liu, Yichen Xie, and Cewu Lu. Anygrasp: Robust and efficient grasp perception in spatial and temporal domains. T-RO, 2023. 1", + "[16] Huan Fu, Bowen Cai, Lin Gao, Ling-Xiao Zhang, Jiaming Wang, Cao Li, Qixun Zeng, Chengyue Sun, Rongfei Jia, Binqiang Zhao, et al. 3d-front: 3d furnished rooms with layouts and semantics. In ICCV, 2021. 1", + "[17] Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3d-future: 3d furniture shape with texture. *IJCT*, 2021. 6", + "[18] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 2020. 1, 3", + "[19] Xiao Gu, Yao Guo, Fani Deligianni, and Guang-Zhong Yang. Coupled real-synthetic domain adaptation for real-world deep depth enhancement. TIP, 2020. 1, 3", + "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3, 4", + "[21] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. Journal of Machine Learning Research, 23(47):1-33, 2022. 3", + "[22] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In CVPR, 2017. 3, 6, 7", + "[23] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engil Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In CVPR, 2014. 1", + "[24] Cheongwoong Kang, Wonjoon Chang, and Jaesik Choi. Balanced domain randomization for safe reinforcement learning. Applied Sciences, 14(21):9710, 2024. 3", + "[25] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In CVPR, 2023. 6, 7", + "[26] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In CVPR, 2024. 3" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2617", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Numair Khan, Min H Kim, and James Tompkin. Differentiable diffusion for dense depth estimation from multi-view images. In CVPR, 2021. 3", + "[28] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. STAT, 1050:1, 2014. 3", + "[29] Naveen Kodali, Jacob D. Abernethy, James Hays, and Zsolt Kira. On convergence and stability of gans. arXiv preprint arXiv:1705.07215, 2017. 1, 3", + "[30] Michael J Landau, Benjamin Y Choo, and Peter A Beling. Simulating kinect infrared and depth images. IEEE transactions on cybernetics, 2015. 1, 3", + "[31] Haolin Liu, Chongjie Ye, Yinyu Nie, Yingfan He, and Xiaoguang Han. Lasa: Instance reconstruction from real scans using a large-scale aligned shape annotation dataset. In CVPR, 2024. 2, 4, 5, 6, 7, 8", + "[32] Naoki Matsunaga, Masato Ishii, Akio Hayakawa, Kenji Suzuki, and Takuya Narihira. Fine-grained image editing by pixel-wise guidance using diffusion models. arXiv preprint arXiv:2212.02024, 2022. 3", + "[33] Stephan Meister, Rahul Nair, and Daniel Kondermann. Simulation of time-of-flight sensors using global illumination. In VMV, 2013. 1, 3", + "[34] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2021. 3", + "[35] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3", + "[36] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 3", + "[37] Alexander Pashevich, Robin Strudel, Igor Kalevatykh, Ivan Laptev, and Cordelia Schmid. Learning to augment synthetic images for sim2real policy transfer. In IROS, 2019. 3", + "[38] Duo Peng, Ping Hu, Qiuhong Ke, and Jun Liu. Diffusion-based image translation with label guidance for domain adaptive semantic segmentation. In ICCV, 2023. 3", + "[39] Benjamin Planche, Ziyan Wu, Kai Ma, Shanhui Sun, Stefan Kluckner, Oliver Lehmann, Terrence Chen, Andreas Hutter, Sergey Zakharov, Harald Kosch, et al. Depthsynth: Realtime realistic synthetic data generation from cad models for 2.5 d recognition. In 3DV, 2017. 1, 3", + "[40] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 5", + "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 4", + "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 3, 4, 5, 6, 7" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 3", + "[44] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. TPAMI, 2022. 3", + "[45] Saurabh Saxena, Abhishek Kar, Mohammad Norouzi, and David J Fleet. Monocular depth estimation using diffusion models. arXiv preprint arXiv:2302.14816, 2023. 3", + "[46] Saurabh Saxena, Charles Herrmann, Junhwa Hur, Abhishek Kar, Mohammad Norouzi, Deqing Sun, and David J Fleet. The surprising effectiveness of diffusion models for optical flow and monocular depth estimation. NeurIPS, 2024. 3", + "[47] Yuefan Shen, Yanchao Yang, Youyi Zheng, C Karen Liu, and Leonidas J Guibas. Dcl: Differential contrastive learning for geometry-aware depth synthesis. IEEE Robotics and Automation Letters, 2022. 1, 3", + "[48] Yuan Shi, Huiyun Cao, Bin Xia, Rui Zhu, Qingmin Liao, and Wenming Yang. Dsr-diff: Depth map super-resolution with diffusion model. Pattern Recognition Letters, 184:225-231, 2024. 3", + "[49] Ashish Shrivastava, Tomas Pfister, Oncel Tuzel, Joshua Susskind, Wenda Wang, and Russell Webb. Learning from simulated and unsupervised images through adversarial training. In CVPR, 2017. 3", + "[50] Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 3", + "[51] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3, 5", + "[52] Shuran Song, Samuel P. Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In CVPR, 2015. 1", + "[53] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In NeurIPS, 2019, 2, 4", + "[54] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. In NeurIPS, 2020.", + "[55] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2", + "[56] Chris Sweeney, Greg Izatt, and Russ Tedrake. A supervised approach to predicting noise in depth images. In ICRA, 2019. 1, 3", + "[57] Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. In IROS, 2017. 2", + "[58] Josh Tobin, Lukas Biewald, Rocky Duan, Marcin Andrychowicz, Ankur Handa, Vikash Kumar, Bob McGrew, Alex Ray, Jonas Schneider, Peter Welinder, et al. Domain randomization and generative models for robotic grasping. In IROS, 2018. 2" + ], + "bbox": [ + 514, + 90, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2618", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[59] Fabio Tosi, Pierluigi Zama Ramirez, and Matteo Poggi. Diffusion models for monocular depth estimation: Overcoming challenging conditions. arXiv preprint arXiv:2407.16698, 2024.3", + "[60] Jonathan Tremblay, Aayush Prakash, David Acuna, Mark Brophy, Varun Jampani, Cem Anil, Thang To, Eric Cameracci, Shaad Boochoon, and Stan Birchfield. Training deep networks with synthetic data: Bridging the reality gap by domain randomization. In CVPRW, 2018. 2", + "[61] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. 7", + "[62] Arash Vahdat, Karsten Kreis, and Jan Kautz. Score-based generative modeling in latent space. In NeurIPS, 2021. 2", + "[63] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matthew J. Kusner. Unsupervised point cloud pre-training via occlusion completion. In ICCV, 2021. 7", + "[64] Jiyuan Wang, Chunyu Lin, Lang Nie, Kang Liao, Shuwei Shao, and Yao Zhao. Digging into contrastive learning for robust depth estimation with diffusion models. In ACM MM, 2024. 3", + "[65] Leichen Wang, Bastian Goldluecke, and Carsten Anklam. L2r gan: Lidar-to-radar translation. In ACCV, 2020. 3", + "[66] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic graph cnn for learning on point clouds. TOG, 2019. 7", + "[67] Zhen Wang, Qiangeng Xu, Feitong Tan, Mengelei Chai, Shichen Liu, Rohit Pandey, Sean Fanello, Achuta Kadambi, and Yinda Zhang. Mvdd: Multi-view depth diffusion models. In ECCV, 2025. 3", + "[68] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In CVPR, 2015. 1, 7", + "[69] Mutian Xu, Pei Chen, Haolin Liu, and Xiaoguang Han. Toscane: A large-scale dataset for understanding 3d tabletop scenes. In ECCV, 2022. 1", + "[70] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Tianyou Liang, Guanying Chen, Shuguang Cui, and Xiaoguang Han. Mvimgnet: A large-scale dataset of multi-view images. In CVPR, 2023. 1", + "[71] Chengjie Yuan, Yunlei Shi, Qian Feng, Chunyang Chang, Michael Liu, Zhaopeng Chen, Alois Christian Knoll, and Jianwei Zhang. Sim-to-real transfer of robotic assembly with visual inputs using cyclegan and force control. In 2022 IEEE International Conference on Robotics and Biomimetics (ROBIO), pages 1426-1432. IEEE, 2022. 3", + "[72] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2100-2110, 2019. 3", + "[73] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 4" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[74] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3", + "[75] Xiaoshuai Zhang, Rui Chen, Ang Li, Fanbo Xiang, Yuzhe Qin, Jiayuan Gu, Zhan Ling, Minghua Liu, Peiyu Zeng, Songfang Han, Zhiao Huang, Tongzhou Mu, Jing Xu, and Hao Su. Close the optical sensing domain gap by physics-grounded active stereo sensor simulation. IEEE Transactions on Robotics, 2023. 1, 3", + "[76] Xiang Zhang, Bingxin Ke, Hayko Riemenschneider, Nando Metzger, Anton Obukhov, Markus Gross, Konrad Schindler, and Christopher Schroers. Betterdepth: Plug-and-play diffusion refiner for zero-shot monocular depth estimation. arXiv preprint arXiv:2407.17952, 2024. 3", + "[77] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, 2017. 3, 6, 7" + ], + "bbox": [ + 514, + 90, + 903, + 330 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2619", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Stable-Sim2Real_ Exploring Simulation of Real-Captured 3D Data with Two-Stage Depth Diffusion/ac3096b4-f6df-4b0e-b9d1-43e46507be43_model.json b/2025/Stable-Sim2Real_ Exploring Simulation of Real-Captured 3D Data with Two-Stage Depth Diffusion/ac3096b4-f6df-4b0e-b9d1-43e46507be43_model.json new file mode 100644 index 0000000000000000000000000000000000000000..40d47b6c2c7244be188f6224b785266b47fb7d24 --- /dev/null +++ b/2025/Stable-Sim2Real_ Exploring Simulation of Real-Captured 3D Data with Two-Stage Depth Diffusion/ac3096b4-f6df-4b0e-b9d1-43e46507be43_model.json @@ -0,0 +1,2543 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.138, + 0.13, + 0.861, + 0.177 + ], + "angle": 0, + "content": "Stable-Sim2Real: Exploring Simulation of Real-Captured 3D Data with Two-Stage Depth Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.202, + 0.925, + 0.224 + ], + "angle": 0, + "content": "Mutian \\(\\mathrm{Xu}^{1}\\) Chongjie \\(\\mathrm{Ye}^{2,3,1}\\) Haolin Liu\\(^{4}\\) Yushuang Wu\\(^{5}\\) Jiahao Chang\\(^{1}\\) Xiaoguang Han\\(^{1,2,3*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.385, + 0.228, + 0.641, + 0.244 + ], + "angle": 0, + "content": "\\(^{1}\\)SSE, CUHKSZ \\(^{2}\\)FNii-Shenzhen" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.246, + 0.798, + 0.263 + ], + "angle": 0, + "content": "3Guangdong Provincial Key Laboratory of Future Networks of Intelligence, CUHKSZ" + }, + { + "type": "text", + "bbox": [ + 0.355, + 0.265, + 0.666, + 0.281 + ], + "angle": 0, + "content": "4Tencent Hunyuan3D 5ByteDance Games" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.315, + 0.327, + 0.332 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.349, + 0.485, + 0.726 + ], + "angle": 0, + "content": "3D data simulation aims to bridge the gap between simulated and real-captured 3D data, which is a fundamental problem for real-world 3D visual tasks. Most 3D data simulation methods inject predefined physical priors but struggle to capture the full complexity of real data. An optimal approach involves learning an implicit mapping from synthetic to realistic data in a data-driven manner, but progress in this solution has met stagnation in recent studies. This work explores a new solution path of data-driven 3D simulation, called Stable-Sim2Real, based on a novel two-stage depth diffusion model. The initial stage finetunes Stable-Diffusion to generate the residual between the real and synthetic paired depth, producing a stable but coarse depth, where some local regions may deviate from realistic patterns. To enhance this, both the synthetic and initial output depth are fed into a second-stage diffusion, where diffusion loss is adjusted to prioritize these distinct areas identified by a 3D discriminator. We provide a new benchmark scheme to evaluate 3D data simulation methods. Extensive experiments show that training the network with the 3D simulated data derived from our method significantly enhances performance in real-world 3D visual tasks. Moreover, the evaluation demonstrates the high similarity between our 3D simulated data and real-captured patterns. Project page: mutianxu.github.io/stable-sim2real." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.757, + 0.222, + 0.773 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.784, + 0.484, + 0.875 + ], + "angle": 0, + "content": "In recent years, real-world 3D datasets have played a crucial role in addressing a wide range of tasks in 3D vision and robotics [1, 7, 11, 15, 23, 52, 69, 70]. However, collecting real 3D data is often labor-intensive and time-consuming, and is further complicated by the recent concerns regarding data privacy. In this scenario, synthetic (i.e., simulated)" + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.313, + 0.889, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.456, + 0.908, + 0.568 + ], + "angle": 0, + "content": "Figure 1. We delve back into a substantial yet stagnant problem of simulating real-captured 3D data in a data-driven manner, introducing Stable-Sim2Real. Given the depth maps of the 3D synthetic data, our approach applies a two-stage depth diffusion model based on Stable Diffusion [42] to generate realistic depth maps that are fused to produce real 3D point clouds. The initial stage generates stable but coarse depth maps, followed by a second-stage diffusion to enhance regional details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.907, + 0.645 + ], + "angle": 0, + "content": "data [8, 10, 12, 16, 68] emerges as an alternative data resource that is cost-effective, rapid, and scalable. Nevertheless, models trained using synthetic data may not perform robustly against the real-world signal." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.909, + 0.903 + ], + "angle": 0, + "content": "This problem has catalyzed the development of 3D data simulation techniques, to minimize the gap between simulated and real-captured 3D data. While some studies [2, 30, 33, 39, 75] have explored grounding physical priors to simulate depth sensors, they struggle to capture the full complexity of real-world patterns and are confined to simulation environments, due to the reliance on predefined and explicit physical modeling. A better solution involves learning an implicit mapping from synthetic to realistic data in a data-driven manner, offering greater flexibility and adaptability to real-world diversities. However, only a handful of earlier works [4, 19, 47, 56] have ventured down this path. Hindered by the lack of synthetic-real paired datasets and the dependence on conventional networks like GANs [18], which are prone to training instability and mode collapse [29], these methods have yet to demonstrate the anticipated performance levels. As a result, developing data-driven 3D" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.235, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2609" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.122 + ], + "angle": 0, + "content": "Sim2Real methods has long been a challenge within the community, and has shown little progress." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.124, + 0.483, + 0.412 + ], + "angle": 0, + "content": "In this paper, we aim to explore the data-driven 3D Sim2Real and refocus the community's attention on this critical problem. As a prerequisite, we choose a recent synthetic-real paired dataset, LASA [31], containing 10,412 high-quality 3D shape CAD annotations aligned with the real-world object scans. Driven by LASA, our efforts focus on designing an effective data-driven 3D Sim2Real algorithm. Given the inherent indeterminacy and diversity in real-captured data patterns, deep generative models emerge as a natural fit. Among these models, diffusion models [20, 53-55, 62] recently offer easier training dynamics and stronger generation ability. Yet, due to the lack of 3D data, it is hard to train a 3D diffusion model to get strong 3D priors for 3D simulation. Therefore, we opt to leverage the strong generalization priors of \\(2D\\) diffusion foundation models (e.g., SD - Stable Diffusion [42]) to simulate real 2D depth maps, which are then fused to gain 3D data. This strategy is similar to real 3D data acquisition, where 2D depth is captured and fused into 3D." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.483, + 0.581 + ], + "angle": 0, + "content": "To this end, an intuitive baseline is to sample CAD (i.e., synthetic) depth maps and their paired real depth images from LASA, and then finetune SD to learn an implicit mapping between them. Although such a solution powered by large-scale paired data and the foundation diffusion model has shown effectiveness in conventional image translations, it faces special challenges in our task: while conventional image translations (e.g., denoising/restoration) normally remove noise to produce clean images, our task aims to output noisy depth that is highly uncertain, thereby making the distribution that needs to be learned more complex." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.483, + 0.809 + ], + "angle": 0, + "content": "To address these challenges, we propose Stable-Sim2Real. The core is a novel two-stage depth diffusion model to simulate real-captured noise (a sample is shown in Fig. 1). At Stage-I diffusion, instead of directly generating the corresponding real-world depth map, our model generates the residual (i.e., difference) between real and CAD depth maps. The simulated depth map is then yielded by adding the generated residual with the CAD depth. Empirically, this residual represents the noise added to clean (i.e., synthetic) depth during real-world scanning. Moreover, compared to directly generating noisy real depth, adding noise to CAD depth, which is inherently clean and view-consistent, can produce more stable depth with smaller view variation and better preservation of the original geometry (see Sec. 3.2 for more discussions and analysis)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "However, according to our observations and experiments, while some regions within the Stage-I generated depth map successfully fit the realistic pattern, some unsatisfactory areas still exhibit conspicuous 3D geometric discrepancies between the generation and real captures (shown in Fig. 6). To tackle this issue, our model then revolves" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.319 + ], + "angle": 0, + "content": "around discriminating unsatisfactory areas and enhancing the Stage-I generation at local regions. We train a 3D-Aware Discriminator to distinguish between Stage-I generation and real captures at local geometries, which is expected to delineate the distribution boundary between them. If a generated patch can be successfully discriminated, it signifies an unsatisfactory area. Next, we introduce a Stage-II diffusion, conditioned not only on the CAD depth but also on Stage-I generation. More importantly, the diffusion loss is adjusted to prioritize on unsatisfactory regions. In this way, Stage-II diffusion is specialized on learning the distribution of unsatisfactory areas, thereby enhancing Stage-I generation. Finally, the generated depth maps are fused to derive the simulated 3D data. Notably, the binary classifier is only used during the training phase." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.32, + 0.905, + 0.516 + ], + "angle": 0, + "content": "Last, we provide a comprehensive benchmark scheme for 3D data simulation. The key logic is: if a model's real-world performance can be improved after being trained with simulated data, it would validate the effectiveness of simulation methods. We focus on fundamental real-world 3D tasks: 3D shape reconstruction and 3D object/scene understanding. As for shape reconstruction, we pretrain a 3D reconstruction network that takes simulated 3D data as input and outputs clean 3D surfaces. For 3D object/scene understanding, the simulated 3D data is used to pretrain a self-supervised point cloud learning framework. To better assess the performance gain purely from simulated data itself, we conduct few-shot evaluation on pre-trained networks." + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.517, + 0.783, + 0.53 + ], + "angle": 0, + "content": "Our contributions are summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.533, + 0.904, + 0.593 + ], + "angle": 0, + "content": "- We delve back into a substantial yet stagnant problem, data-driven 3D real data simulation, paving a new solution path of training generative foundation models on extensive synthetic-real paired data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.904, + 0.638 + ], + "angle": 0, + "content": "- We present Stable-Sim2Real, a novel 3D real data simulation method based on a two-stage depth diffusion model that can stably simulate real-captured noise." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.639, + 0.904, + 0.682 + ], + "angle": 0, + "content": "- We provide a comprehensive scheme to benchmark 3D data simulation, based on few-shot evaluation on networks trained with simulated data for real-world 3D tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.684, + 0.904, + 0.729 + ], + "angle": 0, + "content": "- Extensive qualitative and quantitative results show that the data simulated by our method significantly boosts model performance on real-world 3D tasks." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.533, + 0.904, + 0.729 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.746, + 0.655, + 0.762 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.788, + 0.787 + ], + "angle": 0, + "content": "2.1. Simulation of Real-World Data" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Simulation of 2D real-world data. In the realm of 2D real-world data simulation, two predominant approaches have emerged: domain randomization and domain adaptation. The first line of works [57, 58, 60] randomly perturbs textures, lighting, and material of synthetic data, thereby modeling the synthetic-real domain gap as one kind of variation to endow models with robustness to real-world data." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2610" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.228 + ], + "angle": 0, + "content": "However, it brought the challenge of achieving a delicate balance between ensuring that the distribution of augmentations adequately covers that of real-world data and preventing the augmentation from being too extensive to hinder effective model training [24, 37, 72]. Domain adaptation, on the other hand, aims to minimize the domain gap by transferring knowledge from synthetic data to real-world scenarios, leveraging techniques such as adversarial training and self-supervised learning [5, 49, 71]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.246, + 0.485, + 0.534 + ], + "angle": 0, + "content": "Simulation of 3D real-captured data. Acquiring real-world 3D data and their annotation is labor intensive and inefficient, deriving methods for the simulation of 3D data. Existing simulation methods primarily focus on enhancing the realism of depth information by grounding physical priors in simulators to close the domain gap with 3D data from the physical world [2, 30, 33, 39, 75]. However, these approaches are usually limited by their reliance on predefined and explicit physical modeling. Alternatively, data-driven methods turn to implicitly learn from the distribution in real-captured data, independent of physical modeling [4, 19, 47, 56], offering enhanced adaptability and generalization capabilities to various real-world scenarios. Nonetheless, due to the absence of extensive synthetic-real paired datasets and the reliance on traditional GANs [18], which suffer from training instability and mode collapse [29], these approaches have not achieved the expected performance. Consequently, the research progress of data-driven 3D Sim2Real techniques faces a recent stagnation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.543, + 0.398, + 0.559 + ], + "angle": 0, + "content": "2.2. Image Generation and Translation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Conventional generative models. We formulate the 3D Sim2Real task as learning the mapping from synthetic to real-world captured noise, based on depth images. Existing approaches in the domain of image generation and translation include a variety of generative models such as Generative Adversarial Networks (GANs) [18], Variational Autoencoders (VAEs) [28], Cycle-Consistent Adversarial Networks (Cycle-GAN) [77], and Pix2Pix [22]. While GAN-based methods, which have been predominant in image generation before diffusion models, achieved significant milestones in Sim2Real tasks [6, 65, 71], they often struggle with maintaining local details and semantic consistency, leading to sub-optimal performance [13, 29, 32, 36, 38]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Diffusion models. Diffusion models, also known as denoising diffusion probabilistic models (DDPMs) [50], have emerged as a powerful alternative to GANs for image synthesis [13, 32, 34, 35, 43, 44, 74]. Unlike GANs, diffusion models offer a more stable training procedure and have been shown to generate images with higher fidelity and better preservation of fine details [21, 42]. Depth diffusion models recently track great attention. They mainly focus on the task" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.228 + ], + "angle": 0, + "content": "of depth estimation in a monocular [14, 26, 45, 46, 59, 64] or multi-view [27, 67] manner. Other works also cover conventional image translation tasks, such as depth map superresolution [48] and refinement [76], which produce clean and certain output images. Differently, our work leverages generative models to \"generate noise\", where the expected output is highly uncertain. Thus, the distribution that needs to be learned is more complex, bringing fundamentally different challenges." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.24, + 0.605, + 0.256 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.265, + 0.907, + 0.462 + ], + "angle": 0, + "content": "We hope to exploit the strong prior of the diffusion-based foundation model, so we take depth map as our bridge towards the simulation of real-captured 3D data. The preliminaries on the diffusion model are provided in Sec. 3.1. Our Stable-Sim2Real is a two-stage depth diffusion framework, as illustrated in Fig. 2. The first-stage model (Sec. 3.2) learns to generate the depth residual, producing stable but coarse depth maps, where some areas are distinct from the realistic pattern. Therefore, a 3D binary classifier (Sec. 3.3) is introduced to discriminate these distinct areas. The second-stage model (Sec. 3.4) is specialized to enhance the generation of these areas. Finally, the enhanced depth maps are fused to obtain the 3D simulated data (Sec. 3.5)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.47, + 0.813, + 0.485 + ], + "angle": 0, + "content": "3.1. Preliminaries on Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.906, + 0.569 + ], + "angle": 0, + "content": "Our model is based on Denoising Diffusion Probabilistic Models (DDPM) [20, 51]. The key of DDPM is to model a data distribution \\( p(x) \\) via building a Markov chain in the data space. It can be divided into two main processes, forward process and backward process." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.586, + 0.906, + 0.675 + ], + "angle": 0, + "content": "Forward diffusion process. Given an initial data \\( x_0 \\), we gradually add random noise \\( \\epsilon \\sim \\mathcal{N}(0, \\mathbf{I}) \\) to \\( x_0 \\), yielding a series of noisy data, which can be simply formulated as \\( x_{t+1} = x_t + \\epsilon_t \\). After adding noise for multiple steps, the initial distribution of \\( p(x) \\) is expected to be transformed into a standard Gaussian distribution." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.676, + 0.906, + 0.707 + ], + "angle": 0, + "content": "According to DDPM [20], this process can be directly achieved by the following forward diffusion process:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.715, + 0.905, + 0.747 + ], + "angle": 0, + "content": "\\[\nq \\left(x _ {t}\\right) = \\sqrt {\\alpha_ {t}} x _ {0} + \\sqrt {1 - \\alpha_ {t}} \\boldsymbol {\\epsilon}, x _ {0} \\sim p (x), t \\in \\{0, 1, \\dots , T \\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.748, + 0.906, + 0.807 + ], + "angle": 0, + "content": "where \\(T\\) denotes the number of the time step, \\(t\\) is the current time step, and \\(\\alpha_{t}\\) regulates the noise schedule deciding the pace at which the initial data distribution transforms into a standard Gaussian distribution." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Backward diffusion process. Given a noisy \\( x_{t} \\) from a standard Gaussian distribution as the initial state, \\( x_{t} \\) is denoised to a cleaner \\( x_{t-1} \\) at each time step, which can also be simply formulated as \\( x_{t-1} = x_{t} - \\mu_{\\theta}^{\\epsilon}(x_{t}, t) + \\epsilon_{t} \\) in which \\( \\epsilon_{t} \\sim \\mathcal{N}(0, \\sigma_{\\mathbf{t}} \\mathbf{I}) \\) and \\( \\mu_{\\theta}^{\\epsilon} \\) is a learnable network to predict" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2611" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.136, + 0.089, + 0.868, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.296, + 0.908, + 0.385 + ], + "angle": 0, + "content": "Figure 2. An overview of our Stable-Sim2Real. Conditioned on CAD (i.e., synthetic) depth maps, our model first generates the residual (i.e., difference) between real-captured and CAD depth maps. By adding the generated residual to the CAD depth, Stage-I diffusion produces stable but coarse depth maps (Sec. 3.2). Next, the Stage-I output is projected into 3D and sent to a 3D discriminator, to identify the unsatisfactory local areas that are distinct from real-captured patterns (Sec. 3.3). Later, our Stage-II diffusion is conditioned on both CAD and Stage-I output, where the training loss of Stage-II is re-weighted to be specialized on enhancing unsatisfactory areas of Stage-I output (Sec. 3.4). Notably, the local discriminating and loss re-weighting are only used during training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.396, + 0.483, + 0.442 + ], + "angle": 0, + "content": "the injected noise. By repeating this reverse process until the maximum number of steps \\( T \\), we can obtain the final generated data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.442, + 0.483, + 0.471 + ], + "angle": 0, + "content": "In DDPM [20], the backward diffusion process is proven to be:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.478, + 0.482, + 0.527 + ], + "angle": 0, + "content": "\\[\nx _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} x _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {\\alpha_ {t} \\left(1 - \\Pi_ {\\tau = 0} ^ {t} \\alpha_ {\\tau}\\right)}} \\mu_ {\\theta} ^ {\\epsilon} \\left(x _ {t}, t;\\right) + \\sigma_ {t} \\epsilon , \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.483, + 0.585 + ], + "angle": 0, + "content": "where \\(\\sigma_{t}\\) denotes the standard deviation of \\(q(x_{t - 1}|x_t,x_0)\\) in the backward diffusion to sample random noise. The loss for training \\(\\mu_{\\theta}^{\\epsilon}\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.594, + 0.483, + 0.613 + ], + "angle": 0, + "content": "\\[\nL _ {\\theta} = \\mathbb {E} _ {\\epsilon , t} \\| \\epsilon - \\mu_ {\\theta} ^ {\\epsilon} (x _ {t}, t) \\| ^ {2} \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.623, + 0.383, + 0.639 + ], + "angle": 0, + "content": "3.2. Stage-I: Initial Depth Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.765 + ], + "angle": 0, + "content": "Residual depth generation with SD. Different from unconditional diffusion models, text-to-image diffusion models focus on generating images with optional text prompts. A representative is Stable Diffusion (SD) [42]. SD trains a denoising U-Net architecture \\(\\mu_{\\theta}(x_t,t,c)\\) on the latent space of a pre-trained Variational Auto-Encoder (VAE). Here, \\(c\\) represents the supplementary text prompt embedding, usually acquired through methods like CLIP [41]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We finetune SD model on synthetic-real paired depth data from LASA [31]. Specifically, given an initial synthetic depth map \\( S \\) and its paired real-captured depth map \\( R \\) from a pre-defined dataset, we first get the initial ground truth residual (i.e., difference) depth map \\( D_{res}^{gt} = S - R \\). Then we follow ControlNet [73] to add an additional encoder \\( En \\) and decoder \\( \\mu_{\\theta}^{\\epsilon} \\) for finetuning SD. Specifically, we run the forward process to add noise on \\( D_{res}^{gt} \\) following Eq. (1) and encode with \\( En \\) to obtain \\( x_{t} \\). Next, we send \\( S \\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.396, + 0.907, + 0.473 + ], + "angle": 0, + "content": "to a pretrained VAE encoder \\(En\\), generating the latent condition signal. This latent signal \\(En(S)\\) is further processed by an additional encoder \\(\\pmb{f}_{\\phi}\\) and sent to the decoder blocks of denoising U-Net. We train both the decoder blocks \\(\\mu_{\\theta}^{\\epsilon}\\) and encoder \\(\\pmb{f}_{\\phi}\\) using the following loss (based on Eq. (3)):" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.484, + 0.907, + 0.503 + ], + "angle": 0, + "content": "\\[\nL _ {\\theta , \\phi} = \\mathbb {E} _ {\\epsilon , S, t} \\| \\epsilon - \\boldsymbol {\\mu} _ {\\theta} ^ {\\epsilon} (\\boldsymbol {x} _ {t}, t, \\boldsymbol {f} _ {\\phi} (E n (\\boldsymbol {S}))) \\| ^ {2} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.907, + 0.657 + ], + "angle": 0, + "content": "Learning depth residual helps to produce stable depth. Empirically, this residual represents the noise added to clean (i.e., synthetic) depth during real-world scanning. Moreover, compared to directly generating noisy real depth, adding noise to CAD depth, which is inherently clean and view-consistent, can produce more stable depth with smaller view variation and better preservation of the original geometry. From experimental results, learning residual indeed performs better (see Tab. 5)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.907, + 0.795 + ], + "angle": 0, + "content": "Learning depth residual is theoretically identical to learning real depth. First, assume that real depth is a sample from a Gaussian distribution, denoted by \\( R \\sim \\mathcal{N}(\\alpha_{\\mathbf{R}}, \\vartheta) \\), where \\( \\alpha \\) denotes the mean and \\( \\vartheta \\) represents the variance. Notably, \\( \\vartheta \\) can naturally represent real-captured noise. Therefore, synthetic depth can be denoted by \\( S \\sim \\mathcal{N}(\\alpha_{\\mathbf{S}}, 0) \\), where the variance is zero as there is no real-captured noise." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In our design, as the diffusion model is inherently designed to learn the transformation from a random noise \\(\\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) to any other Gaussian distribution [53], we leverage it to model the distribution of the depth residual that can be denoted by \\(\\mathcal{N}(\\alpha_{\\mathbf{R}} - \\alpha_{\\mathbf{S}},\\vartheta)\\). Last, by adding the depth residual to synthetic depth, the output distribution can be formulated as: \\(\\mathcal{N}(\\alpha_{\\mathbf{R}} - \\alpha_{\\mathbf{S}},\\vartheta) + \\mathcal{N}(\\alpha_{\\mathbf{S}},0) = \\mathcal{N}(\\alpha_{\\mathbf{R}},\\vartheta)\\)," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2612" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "which is exactly identical to the previously assumed real-captured noise distribution." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.129, + 0.373, + 0.145 + ], + "angle": 0, + "content": "3.3. 3D-Aware Local Discriminating" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.151, + 0.484, + 0.332 + ], + "angle": 0, + "content": "Unsatisfactory areas distinct from real patterns. Based on our observations, the noise captured in real data is spatially-variant, meaning different areas within a real depth map may display varying levels of perturbation. However, our Stage-I diffusion tends to introduce spatially-uniform noise across the entire input CAD depth. This discrepancy may stem from a bias within SD, as also discussed in [9]. As a result, although certain regions of the generated depth map align well with real perturbation patterns, there remain noticeable areas that deviate from reality. This leads to evident 3D geometric discrepancies between Stage-I outputs and real captures (as illustrated in Fig. 6)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.346, + 0.484, + 0.36 + ], + "angle": 0, + "content": "Identifying unsatisfactory areas with 3D discriminator." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.362, + 0.483, + 0.589 + ], + "angle": 0, + "content": "Our design logic to discriminate unsatisfactory areas is that if the generated areas diverge significantly from real patterns, there should exist a distribution boundary of their 3D geometries. To this end, we propose employing a 3D binary classifier to discriminate this boundary. Initially, we split all generated and real-captured depth maps into different patches representing different areas, which are then projected to form 3D point patches. Subsequently, we leverage PointNet [40] as a 3D binary classifier to differentiate whether the input point patches come from model generation or real captures. The CAD depth maps are also concatenated in the input as additional guidance. The successful classification of a generated patch indicates that its corresponding area is distinct from real captures. Notably, the binary classifier is solely used during training." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.595, + 0.466, + 0.611 + ], + "angle": 0, + "content": "3.4. Stage-II: Unsatisfactory Area Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.617, + 0.484, + 0.753 + ], + "angle": 0, + "content": "Improving unsatisfactory areas by re-weighting denoising loss. After identifying these local patches, we introduce a second-stage depth refinement model that differs from the Stage-I diffusion in two crucial aspects: 1) The diffusion model is now conditioned on the concatenation of synthetic depth \\( S \\) and the previously generated depth residual \\( D^{I} \\) from Stage I. Here, \\( D^{I} = D_{res}^{I} + S \\), where \\( D_{res}^{I} \\) is the output residual depth from Stage-I model. Thus, the training loss defined in Eq. (4) is updated to:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.759, + 0.482, + 0.793 + ], + "angle": 0, + "content": "\\[\nL _ {\\chi , \\psi} = \\mathbb {E} _ {\\boldsymbol {\\epsilon}, \\boldsymbol {D} ^ {I}, \\boldsymbol {S}, t} \\left\\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\mu} _ {\\chi} ^ {\\epsilon} \\left(\\boldsymbol {x} _ {t}, t, \\boldsymbol {f} _ {\\psi} \\left(E n \\left(\\boldsymbol {D} _ {r e s} ^ {I}, \\boldsymbol {S}\\right)\\right)\\right) \\right\\| ^ {2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.793, + 0.483, + 0.839 + ], + "angle": 0, + "content": "2) More importantly, the diffusion loss is re-weighted to prioritize the enhancement of the identified unsatisfactory regions, making the final Stage-II loss \\(L_{\\chi ,\\psi}\\) to be:" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.843, + 0.483, + 0.864 + ], + "angle": 0, + "content": "\\[\nL _ {\\chi , \\psi} = \\omega L _ {\\chi , \\psi} ^ {s i m} + \\lambda L _ {\\chi , \\psi} ^ {d i s}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.868, + 0.484, + 0.905 + ], + "angle": 0, + "content": "where \\(L_{\\chi, \\psi}^{sim}\\) indicates the loss on similar regions, while \\(L_{\\chi, \\psi}^{dis}\\) denotes the loss on previously identified distinct (i.e.," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.136 + ], + "angle": 0, + "content": "unsatisfactory) areas. \\(\\omega\\) and \\(\\lambda\\) are weight coefficients, in which \\(\\lambda >\\omega\\). In our experiment, setting \\(\\omega = 0.5\\) and \\(\\lambda = 1.5\\) produces the best result." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.137, + 0.905, + 0.213 + ], + "angle": 0, + "content": "Our Stage-II diffusion model is specialized on learning the distribution of unsatisfactory areas in Stage-I generation. Meanwhile, by conditioning on Stage-I output and assigning less weight to satisfied areas, the generation on satisfied outputs is also preserved in Stage-II, as in Fig. 6." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.22, + 0.78, + 0.236 + ], + "angle": 0, + "content": "3.5. Creating Real-World 3D Data" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.242, + 0.905, + 0.348 + ], + "angle": 0, + "content": "Two-stage real depth simulation. In the initial stage, we sample a random distribution \\(\\epsilon\\) from \\(\\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) and feed it into the denoising U-Net \\(\\mu_{\\theta}\\) of Stage-I diffusion model. The model runs a backward diffusion process for \\(T\\) steps to generate the initial depth residual \\(D_{res}^{I}\\), following Eq. (2) with the synthetic depth \\(S\\) serving as the condition. Now we can obtain the initial depth map \\(D^{I} = D_{res}^{I} + S\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.349, + 0.905, + 0.439 + ], + "angle": 0, + "content": "Moving to the second stage, we repeat a similar backward diffusion process to generate the residual depth \\(D_{res}^{II}\\), using the denoising U-Net \\(\\mu_{\\chi}\\) of our Stage-II diffusion model, conditioned on both synthetic depth \\(S\\) and Stage-I output \\(D^{I}\\). The final depth map \\(D^{II} = D_{res}^{II} + S\\). DDIM sampler [51] is utilized during the whole inference phase." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.906, + 0.562 + ], + "angle": 0, + "content": "Depth fusion to 3D data. Given the 3D synthetic object, we follow LASA [31] to embed CAD objects into 3D scenes. This allows us to render synthetic depth images, simulating the occlusions in real-world captures. Then we conduct the aforementioned two-stage real depth simulation on rendered synthetic depth maps, followed by a depth fusion algorithm to create simulated 3D data." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.574, + 0.646, + 0.591 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.906, + 0.75 + ], + "angle": 0, + "content": "We provide a new benchmark scheme for 3D data simulation. To begin with, different simulation methods are performed on synthetic datasets. The 3D simulated data are then utilized to train networks for tackling fundamental real-world 3D visual tasks: 3D instance reconstruction, 3D object/scene understanding. The key logic of our evaluation is: if the model's real-world performance can be improved after being trained with simulated data, it would validate that the simulated data is close to real distributions, demonstrating the effectiveness of simulation methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.655, + 0.772 + ], + "angle": 0, + "content": "4.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Training of Stable-Sim2Real. We train our model by finetuning the official StableDiffusion V2.1 [42]. The training dataset is a recent 3D synthetic-real paired dataset, LASA [31]. LASA is a large-scale dataset that contains 10,412 unique CAD models across 17 categories. Skilled artists are employed to meticulously craft aligned CAD models from 3D object real-world scans of [3]. LASA provides CAD-real depth map pairs, rendered based on camera" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2613" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.137, + 0.089, + 0.868, + 0.189 + ], + "angle": 0, + "content": "
MethodChairSofaTableCabinetBedShelf
w/o sim. data24.5 / 12.3 / 20.155.6 / 7.02 / 19.825.8 / 19.3 / 20.357.1 / 8.83 / 24.140.1 / 6.14 / 17.312.8 / 6.74 / 23.7
Cycle-GAN [77]25.1 / 12.1 / 20.955.2 / 6.98 / 20.526.1 / 19.1 / 19.556.5 / 8.24 / 25.140.6 / 5.98 / 16.913.1 / 6.08 / 24.9
Pix2Pix [22]24.9 / 11.0 / 20.556.4 / 6.74 / 21.326.6 / 18.5 / 22.157.4 / 7.83 / 25.541.3 / 6.01 / 17.814.2 / 6.15 / 26.2
Giga-GAN [25]26.2 / 10.3 / 22.158.9 / 5.87 / 22.429.3 / 16.4 / 24.360.5 / 7.01 / 26.944.2 / 5.25 / 19.318.2 / 5.87 / 27.8
Stable-Diffusion [42]27.3 / 9.51 / 25.559.4 / 5.93 / 23.629.7 / 15.8 / 24.761.6 / 6.92 / 27.545.1 / 5.31 / 20.119.1 / 5.85 / 28.3
Ours29.2 / 7.83 / 27.562.9 / 4.09 / 25.131.2 / 14.1 / 26.863.2 / 5.58 / 28.646.3 / 4.23 / 21.818.8 / 5.02 / 31.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.2, + 0.907, + 0.256 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of few-shot evaluation on 3D instance reconstruction, where the model is pretrained using simulated data generated by different methods. \"w/o sim. data\" indicates pretraining using no simulated data but only synthetic data. Evaluation metrics are mIoU / Chamfer L2 / F-score respectively. Higher mIoU and F-score are better while lower Chamfer L2 is better. Chamfer L2 is scaled by 1,000. All results are reproduced using the official code of DisCo [31]." + }, + { + "type": "table", + "bbox": [ + 0.153, + 0.269, + 0.851, + 0.365 + ], + "angle": 0, + "content": "
MethodChairSofaTableCabinetBedShelf
w/o sim. data19.4 / 14.7 / 15.846.2 / 9.92 / 15.720.5 / 23.6 / 15.951.2 / 11.3 / 20.233.6 / 13.1 / 10.911.1 / 9.23 / 21.6
Cycle-GAN [77]20.3 / 13.1 / 16.548.9 / 9.03 / 18.923.2 / 22.4 / 17.551.9 / 10.8 / 21.435.2 / 12.5 / 12.111.7 / 8.87 / 23.5
Pix2Pix [22]19.3 / 12.3 / 17.449.7 / 9.51 / 18.822.6 / 21.5 / 18.852.7 / 11.2 / 21.137.7 / 10.0 / 13.812.5 / 8.91 / 24.3
Giga-GAN [25]22.4 / 11.8 / 19.553.0 / 7.69 / 19.423.1 / 19.3 / 21.654.3 / 10.1 / 21.940.3 / 7.95 / 15.414.2 / 7.81 / 26.9
Stable-Diffusion [42]23.3 / 11.7 / 20.454.1 / 7.98 / 19.224.2 / 17.9 / 22.655.1 / 9.07 / 22.440.4 / 7.89 / 16.215.8 / 6.72 / 28.3
Ours27.1 / 9.31 / 24.860.6 / 5.29 / 22.328.5 / 16.1 / 24.760.6 / 7.51 / 25.943.2 / 6.23 / 18.017.3 / 5.93 / 30.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.375, + 0.907, + 0.404 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison of zero-shot evaluation on 3D instance reconstruction, when the model is solely pretrained using simulated data generated by different methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.419, + 0.483, + 0.451 + ], + "angle": 0, + "content": "information during the real-world collection. See the supplementary material for more implementation details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.473, + 0.483, + 0.624 + ], + "angle": 0, + "content": "Methods in comparison. Since our method operates on depth images, we select several image-to-image generation methods for comparison: 1) Cycle-Consistent Adversarial Networks (Cycle-GAN) [77], which is a representative generative network when paired training data are not available, so we train Cycle-GAN on the pair-shuffled LASA training data. 2) Pix2Pix [22], which is a classical image generation method. 3) Giga-GAN [25], a state-of-the-art GAN-based method. 4) Stable Diffusion [42], our main baseline model. Both 2), 3), 4) are trained using the paired data from LASA." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.636, + 0.341, + 0.65 + ], + "angle": 0, + "content": "4.2.3D Instance Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Setup. 3D instance reconstruction aims to recover the full geometry from real-captured partial point clouds or multiview images. We utilize a recent state-of-the-art diffusion-based reconstruction network, DisCo, which is proposed in LASA [31]. It is originally pretrained on three synthetic datasets (i.e., ShapeNet [8], ABO [10], and 3D-Future [17]) and finetuned on LASA real training data. During pretraining, DisCo simply constructs partial point clouds by adding random perturbation to synthetic data. Instead, in our experiment, to simulate partial point clouds for pretraining, we perform simulation using our method and other baseline models based on the original three synthetic datasets used to pretrain DisCo, and replace half of the original partial point clouds with the simulated 3D data. The other configurations all follow the original DisCo." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Subsequently, we employ few-shot and zero-shot learn" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.419, + 0.907, + 0.585 + ], + "angle": 0, + "content": "ing configurations to evaluate: We randomly sample only a few LASA real training data to finetune the pretrained network. In this way, the real-world knowledge of the network will mostly come from the pretraining data (i.e., the simulated data produced by our method and others) instead of the finetuning data, which can better reflect the performance gain from the simulated data itself. The pretrained model is finetuned with \\(20\\sim 30\\%\\) of the original training samples for few-shot learning and is evaluated on the original LASA validation set. For zero-shot learning, we directly evaluate the pretrained network without fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.587, + 0.909, + 0.738 + ], + "angle": 0, + "content": "Why not perform simulation on LASA synthetic data and compare the performance using the simulated data with that using LASA paired real GT? Since our StableSim2Real and other baseline techniques are trained on the LASA training set, it is unfeasible to conduct further inference on LASA, as this would not verify the actual model performance but only reflect the performance caused by overfitting. We can also not conduct simulation on a third-party paired dataset similar to LASA for 3D reconstruction, as there is currently no such paired dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Quantitative results. Following [31], we utilize mean Intersection over Union (mIoU), L2 chamfer distance, and \\(1\\%\\) F-score metrics. Initially, we normalize both the results and the ground-truth CAD objects to fall within the range of -0.5 to 0.5, according to which we calculate the aforementioned metrics between them. As shown in Tab. 1, under all few-shot settings, the model pretrained using the simulated data from our method achieves the best performance, and significantly surpasses the model using no simulated data (i.e., only add random perturbation to synthetic data) for" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2614" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.482, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.241, + 0.485, + 0.286 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparison of few-shot evaluation on 3D instance reconstruction. \"w/o sim.\" denotes using no simulated data for pretraining." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.3, + 0.485, + 0.452 + ], + "angle": 0, + "content": "pretraining. Moreover, using the simulated data produced by Cycle-GAN or Pix2Pix does not improve the model performance compared to that using no simulated data, possibly attributed to their inferior simulation quality, introducing ambiguity and complexity during pretraining, making it hard for the model to learn useful real-world knowledge. Moreover, as listed in Tab. 2, under zero-shot setting, when solely using simulated data for pretraining without any finetuning, the model trained with our simulated data demonstrates more conspicuous effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.46, + 0.484, + 0.522 + ], + "angle": 0, + "content": "Qualitative results. Fig. 3 shows the qualitative comparison, where our simulated data can help the model to gain the best reconstruction accuracy and diversity. More qualitative results are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.531, + 0.377, + 0.549 + ], + "angle": 0, + "content": "4.3. 3D Object/Scene Understanding" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.553, + 0.483, + 0.719 + ], + "angle": 0, + "content": "Setup. We select OcCo [63] as our backbone, which is a self-supervised point cloud pretraining framework. OcCo is originally pretrained on ModelNet [68], and finetuned on real-scanned datasets - ScanObjectNN [61] for object classification, and S3DIS [1] for scene semantic segmentation. ModelNet contains 12,311 synthetic CAD objects from 40 classes. ScanObjectNN includes \\(\\sim 15,000\\) real-scanned objects across 15 categories. S3DIS encompasses 271 indoor rooms and 13 classes. The classification task outputs the class of the whole 3D object, and the semantic segmentation predicts the category of each point in the 3D scene." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In our experiment, we apply our method and other baselines on ModelNet to generate simulated data that are further used to replace half of the original ModelNet synthetic data to pretrain OcCo. Next, the pretrained model is again evaluated under the few-shot learning setting. A typical setting is \"K-way \\(N\\)-shot\", which indicates that \\(K\\) classes are randomly selected during finetuning, and each category contains \\(N\\) samples. The finetuned model is evaluated on the unseen samples picked from each of the \\(K\\) classes. We select DGCNN [66] as the backbone for OcCo. For few-shot learning, we employ the \"5-way, 10-shot\" setting. For S3DIS, we follow the original OcCo to report the results" + }, + { + "type": "table", + "bbox": [ + 0.535, + 0.09, + 0.885, + 0.136 + ], + "angle": 0, + "content": "
Methodw/o sim.C-GAN [77]P2P [22]G-GAN [25]SD [42]Ours
Acc. (%)72.472.572.174.274.677.5
mIoU (%)47.948.047.749.148.951.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.14, + 0.907, + 0.184 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison of few-shot evaluation on 3D object classification and 3D scene semantic segmentation. \"w/o sim\" means using no simulated data for pretraining." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.194, + 0.905, + 0.222 + ], + "angle": 0, + "content": "
Methodw/o sim.C-GAN [77]P2P [22]G-GAN [25]SD [42]Ours
Metrics59.2/3.54/25.360.4/3.21/25.261.5/3.02/26.462.1/2.86/26.762.3/2.83/27.562.9/2.57/28.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.229, + 0.907, + 0.258 + ], + "angle": 0, + "content": "Table 4. The mIoU / Chamfer L2 / F-score between simulated data and real-captures on LASA [31] validation set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.275, + 0.907, + 0.307 + ], + "angle": 0, + "content": "using 6-fold cross-validation. All the other configurations follow the original OcCo." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.317, + 0.907, + 0.455 + ], + "angle": 0, + "content": "Qualitative results. We report the classification accuracy and segmentation mean Intersection over Union (mIoU) in Tab. 3. The model pretrained using our simulated data gets the best results across both 3D object classification and 3D scene understanding tasks. By pretraining with our simulated data, the model can acquire valuable real-world knowledge that helps downstream tasks. In contrast, pretraining with the low-quality simulated data from CycleGAN or Pix2Pix may even degrade the performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.466, + 0.885, + 0.483 + ], + "angle": 0, + "content": "4.4. Is the simulated data close to real captures?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.488, + 0.906, + 0.67 + ], + "angle": 0, + "content": "As previously mentioned, our Stable-Sim2Real is trained on LASA training set. In this section, we aim to evaluate the data simulation quality by directly performing inference and evaluation on LASA validation set. Tab. 4 compares the mIoU, Chamfer L2, and F-score metrics between the simulated data generated by different methods and the real captured Ground Truth (GT). Although not precise, this result can partially prove that our simulated data closely resembles real-world captures. Fig. 4 shows simulation results on depth maps and 3D point clouds. It demonstrates that while our simulated data closely aligns with real-world captures, subtle differences persist, highlighting the diversity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.67, + 0.907, + 0.761 + ], + "angle": 0, + "content": "We also provide the simulation results on ShapeNet [8] in Fig. 5. Despite the absence of real-captured Ground Truth (GT) in ShapeNet, the outcomes still show that the simulated results from our output closely resemble real captures, where the generated geometry is not largely broken, with certain realistic perturbations introduced." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.783, + 0.789 + ], + "angle": 0, + "content": "4.5. Ablation Studies and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Module impacts. To evaluate the impact of each key design in our framework, we conduct ablation studies on the 3D instance reconstruction task, specifically on the table category. Similar to Tab. 2, we conduct experiments under zero-shot protocol. This allows for a focused evaluation of the module's impact without the influence of additional factors. Other setups remain consistent with Sec. 4.2." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "2615" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.087, + 0.482, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.238, + 0.45, + 0.253 + ], + "angle": 0, + "content": "Figure 4. Qualitative simulation results on LASA [31]." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.264, + 0.48, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.484, + 0.455, + 0.499 + ], + "angle": 0, + "content": "Figure 5. Qualitative simulation results on ShapeNet [8]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.483, + 0.579 + ], + "angle": 0, + "content": "1) Residual generation. In Tab. 5 (w/o res.), directly generating the real-world depth map without using residual generation degrades the model performance, verifying the efficacy of our residual generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.483, + 0.641 + ], + "angle": 0, + "content": "2) Loss re-weighting. In our Stage-II model, we reweight the diffusion loss to focus more on unsatisfactory areas. In Tab. 5 (w/o re-w.), the result w/o loss re-weighting is provided, showing its effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.643, + 0.483, + 0.853 + ], + "angle": 0, + "content": "3) 3D binary discriminator. Instead of training a 3D binary discriminator, another simple way is to directly set a threshold of the absolute difference between Stage-I generation and real-world GT to decide distinct areas. We tried various difference thresholds and reported the best result in Tab. 5 (w/o bin.), which hurts the model performance. To explain, our output depth is gained by adding subtle noise to synthetic depth. Thus, the magnitude of the difference between our output and its corresponding real GT is similar at 2D/pixel space, where it is impractical to find an appropriate threshold to identify distinct/unsatisfactory areas. Instead, their difference in distribution space is relatively more significant, so using a 3D binary classifier to delineate their distribution boundary is a more feasible way." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "4) Number of stages. Theoretically, we can iteratively improve the generation quality by adding more stages. However, as shown in Tab. 5, adding Stage-III model does" + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.518, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.09, + 0.906, + 0.114 + ], + "angle": 0, + "content": "
Methodw/o res.w/o re-w.w/o bin.Stage-IIIStage-IVStage-II
Metrics26.4/18.2/22.127.1/17.9/23.327.3/17.5/23.628.6/ 16.1/24.528.0/16.9/24.128.5/16.1/24.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.127, + 0.907, + 0.169 + ], + "angle": 0, + "content": "Table 5. Quantitative ablations of zero-shot evaluation on 3D instance reconstruction of table category. Evaluation metrics are mIoU / Chamfer L2 / F-score respectively." + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.182, + 0.849, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.324, + 0.88, + 0.338 + ], + "angle": 0, + "content": "Figure 6. Qualitative simulation results at the patch level." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.354, + 0.907, + 0.445 + ], + "angle": 0, + "content": "not improve the model performance, and Stage-IV model even slightly degrades the result. To analyze, Stage-II model has been sufficient to identify most of the unsatisfactory areas and improve their generation quality. In addition, adding redundant stages may also cause overfitting issues that hurt the performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.453, + 0.906, + 0.559 + ], + "angle": 0, + "content": "Analyzing Stage-II model at 3D patch-level. 1) We compare simulation results with real-captured Ground Truth (GT) at the patch level. As illustrated in Fig. 6, our Stage-II model effectively refines previously unsatisfactory (i.e., dissimilar to GT) patches from Stage-I output to be similar to GT, while keeping previously satisfactory (i.e., similar to GT) patches almost unchanged." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.56, + 0.906, + 0.68 + ], + "angle": 0, + "content": "2) When sending the newly generated output patches from Stage-II to our 3D binary discriminator, the classification accuracy notably drops from \\(67.8\\%\\) to \\(32.5\\%\\), suggesting that the generated patches successfully approximate the real-captured distributions, making it challenging for the classifier to discriminate. This again verifies the effectiveness of our Stage-II model. See the supplementary material for more ablations of the binary classifier." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.453, + 0.906, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.634, + 0.71 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We introduce Stable-Sim2Real, a two-stage depth diffusion model for 3D real data simulation. Initially, our method learns the residual between synthetic and real depth to generate a stable but coarse depth, where some areas may deviate from realistic patterns. Next, both synthetic and initial output depth are fed into the second-stage model, where diffusion loss prioritizes enhancing distinct regions identified by a binary classifier. Experiments show that the 3D simulated data generated from our method can improve the network performance in real-world 3D tasks, and our simulated data is highly similar to real captures. More limitations are discussed in the supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2616" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.117, + 0.486, + 0.299 + ], + "angle": 0, + "content": "The work was supported in part by Guangdong S&T Programme with Grant No. 2024B0101030002, the Basic Research Project No. HZQB-KCZYZ-2021067 of Hetao Shenzhen-HK S&T Cooperation Zone, Guangdong Provincial Outstanding Youth Fund (No. 2023B1515020055), No. NSFC 62172348, the Shenzhen Outstanding Talents Training Fund 202002, the Guangdong Research Projects No. 2017ZT07X152 and No. 2019CX01X104, the Guangdong Provincial Key Laboratory of Future Networks of Intelligence (Grant No. 2022B1212010001), and the Shenzhen Key Laboratory of Big Data and Artificial Intelligence (Grant No. ZDSYS201707251409055)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.315, + 0.188, + 0.332 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.34, + 0.482, + 0.381 + ], + "angle": 0, + "content": "[1] I. Armeni, O. Sener, A. R. Zamir, H. Jiang, I. Brilakis, M. Fischer, and S. Savarese. 3d semantic parsing of large-scale indoor spaces. In CVPR, 2016. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.385, + 0.482, + 0.438 + ], + "angle": 0, + "content": "[2] Kaixin Bai, Lei Zhang, Zhaopeng Chen, Fang Wan, and Jianwei Zhang. Close the sim2real gap via physically-based structured light synthetic data simulation. In ICRA, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.443, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[3] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Yuri Feigin, Peter Fu, Thomas Gebauer, Daniel Kurz, Tal Dimry, Brandon Joffe, Arik Schwartz, and Elad Shulman. ARK-scenes: A diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In NeurlPS, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.529, + 0.482, + 0.582 + ], + "angle": 0, + "content": "[4] Konstantinos Bousmalis, Nathan Silberman, David Dohan, Dumitru Erhan, and Dilip Krishnan. Unsupervised pixel-level domain adaptation with generative adversarial networks. In CVPR, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.586, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[5] Konstantinos Bousmalis, Alex Irpan, Paul Wohlhart, Yunfei Bai, Matthew Kelcey, Mrinal Kalakrishnan, Laura Downs, Julian Ibarz, Peter Pastor, Kurt Konolige, et al. Using simulation and domain adaptation to improve efficiency of deep robotic grasping. In ICRA, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.659, + 0.482, + 0.712 + ], + "angle": 0, + "content": "[6] Marc Bresson, Yang Xing, and Weisi Guo. Sim2real: Generative ai to enhance photorealism through domain transfer with gan and seven-chanel- \\(360^{\\circ}\\)-paired-images dataset. Sensors, 24(1):94, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.716, + 0.482, + 0.77 + ], + "angle": 0, + "content": "[7] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. In 3DV, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.774, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[8] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 1, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[9] Hansheng Chen, Ruoxi Shi, Yulin Liu, Bokui Shen, Jiayuan Gu, Gordon Wetzstein, Hao Su, and Leonidas Guibas. Generic 3d diffusion adapter using controlled multi-view editing. arXiv preprint arXiv:2403.12032, 2024. 5" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.34, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[10] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksen, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In CVPR, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.163, + 0.905, + 0.216 + ], + "angle": 0, + "content": "[11] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[12] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.904, + 0.3 + ], + "angle": 0, + "content": "[13] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. In NeurlPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.302, + 0.904, + 0.342 + ], + "angle": 0, + "content": "[14] Yiquan Duan, Xianda Guo, and Zheng Zhu. Diffusiondepth: Diffusion denoising approach for monocular depth estimation. In ECCV, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.344, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[15] Hao-Shu Fang, Chenxi Wang, Hongjie Fang, Minghao Gou, Jirong Liu, Hengxu Yan, Wenhai Liu, Yichen Xie, and Cewu Lu. Anygrasp: Robust and efficient grasp perception in spatial and temporal domains. T-RO, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.4, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[16] Huan Fu, Bowen Cai, Lin Gao, Ling-Xiao Zhang, Jiaming Wang, Cao Li, Qixun Zeng, Chengyue Sun, Rongfei Jia, Binqiang Zhao, et al. 3d-front: 3d furnished rooms with layouts and semantics. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.455, + 0.905, + 0.495 + ], + "angle": 0, + "content": "[17] Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3d-future: 3d furniture shape with texture. *IJCT*, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.498, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[18] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.553, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[19] Xiao Gu, Yao Guo, Fani Deligianni, and Guang-Zhong Yang. Coupled real-synthetic domain adaptation for real-world deep depth enhancement. TIP, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.624, + 0.907, + 0.677 + ], + "angle": 0, + "content": "[21] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. Journal of Machine Learning Research, 23(47):1-33, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[22] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In CVPR, 2017. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.72, + 0.905, + 0.76 + ], + "angle": 0, + "content": "[23] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engil Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In CVPR, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.904, + 0.802 + ], + "angle": 0, + "content": "[24] Cheongwoong Kang, Wonjoon Chang, and Jaesik Choi. Balanced domain randomization for safe reinforcement learning. Applied Sciences, 14(21):9710, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[25] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In CVPR, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[26] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In CVPR, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2617" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[27] Numair Khan, Min H Kim, and James Tompkin. Differentiable diffusion for dense depth estimation from multi-view images. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.163 + ], + "angle": 0, + "content": "[28] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. STAT, 1050:1, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[29] Naveen Kodali, Jacob D. Abernethy, James Hays, and Zsolt Kira. On convergence and stability of gans. arXiv preprint arXiv:1705.07215, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[30] Michael J Landau, Benjamin Y Choo, and Peter A Beling. Simulating kinect infrared and depth images. IEEE transactions on cybernetics, 2015. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.305 + ], + "angle": 0, + "content": "[31] Haolin Liu, Chongjie Ye, Yinyu Nie, Yingfan He, and Xiaoguang Han. Lasa: Instance reconstruction from real scans using a large-scale aligned shape annotation dataset. In CVPR, 2024. 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.361 + ], + "angle": 0, + "content": "[32] Naoki Matsunaga, Masato Ishii, Akio Hayakawa, Kenji Suzuki, and Takuya Narihira. Fine-grained image editing by pixel-wise guidance using diffusion models. arXiv preprint arXiv:2212.02024, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.364, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[33] Stephan Meister, Rahul Nair, and Daniel Kondermann. Simulation of time-of-flight sensors using global illumination. In VMV, 2013. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.461 + ], + "angle": 0, + "content": "[34] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[35] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.482, + 0.561 + ], + "angle": 0, + "content": "[36] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.563, + 0.482, + 0.604 + ], + "angle": 0, + "content": "[37] Alexander Pashevich, Robin Strudel, Igor Kalevatykh, Ivan Laptev, and Cordelia Schmid. Learning to augment synthetic images for sim2real policy transfer. In IROS, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.606, + 0.482, + 0.646 + ], + "angle": 0, + "content": "[38] Duo Peng, Ping Hu, Qiuhong Ke, and Jun Liu. Diffusion-based image translation with label guidance for domain adaptive semantic segmentation. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.649, + 0.482, + 0.717 + ], + "angle": 0, + "content": "[39] Benjamin Planche, Ziyan Wu, Kai Ma, Shanhui Sun, Stefan Kluckner, Oliver Lehmann, Terrence Chen, Andreas Hutter, Sergey Zakharov, Harald Kosch, et al. Depthsynth: Realtime realistic synthetic data generation from cad models for 2.5 d recognition. In 3DV, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.719, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[40] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.841 + ], + "angle": 0, + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[43] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.149, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[44] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. TPAMI, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.192, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[45] Saurabh Saxena, Abhishek Kar, Mohammad Norouzi, and David J Fleet. Monocular depth estimation using diffusion models. arXiv preprint arXiv:2302.14816, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.236, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[46] Saurabh Saxena, Charles Herrmann, Junhwa Hur, Abhishek Kar, Mohammad Norouzi, Deqing Sun, and David J Fleet. The surprising effectiveness of diffusion models for optical flow and monocular depth estimation. NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.292, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[47] Yuefan Shen, Yanchao Yang, Youyi Zheng, C Karen Liu, and Leonidas J Guibas. Dcl: Differential contrastive learning for geometry-aware depth synthesis. IEEE Robotics and Automation Letters, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.349, + 0.905, + 0.402 + ], + "angle": 0, + "content": "[48] Yuan Shi, Huiyun Cao, Bin Xia, Rui Zhu, Qingmin Liao, and Wenming Yang. Dsr-diff: Depth map super-resolution with diffusion model. Pattern Recognition Letters, 184:225-231, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.404, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[49] Ashish Shrivastava, Tomas Pfister, Oncel Tuzel, Joshua Susskind, Wenda Wang, and Russell Webb. Learning from simulated and unsupervised images through adversarial training. In CVPR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.462, + 0.905, + 0.515 + ], + "angle": 0, + "content": "[50] Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.518, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[51] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.562, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[52] Shuran Song, Samuel P. Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In CVPR, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.604, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[53] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In NeurIPS, 2019, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.647, + 0.905, + 0.674 + ], + "angle": 0, + "content": "[54] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. In NeurIPS, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.676, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[55] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.733, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[56] Chris Sweeney, Greg Izatt, and Russ Tedrake. A supervised approach to predicting noise in depth images. In ICRA, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.775, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[57] Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. In IROS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[58] Josh Tobin, Lukas Biewald, Rocky Duan, Marcin Andrychowicz, Ankur Handa, Vikash Kumar, Bob McGrew, Alex Ray, Jonas Schneider, Peter Welinder, et al. Domain randomization and generative models for robotic grasping. In IROS, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2618" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[59] Fabio Tosi, Pierluigi Zama Ramirez, and Matteo Poggi. Diffusion models for monocular depth estimation: Overcoming challenging conditions. arXiv preprint arXiv:2407.16698, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.148, + 0.482, + 0.217 + ], + "angle": 0, + "content": "[60] Jonathan Tremblay, Aayush Prakash, David Acuna, Mark Brophy, Varun Jampani, Cem Anil, Thang To, Eric Cameracci, Shaad Boochoon, and Stan Birchfield. Training deep networks with synthetic data: Bridging the reality gap by domain randomization. In CVPRW, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.482, + 0.272 + ], + "angle": 0, + "content": "[61] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[62] Arash Vahdat, Karsten Kreis, and Jan Kautz. Score-based generative modeling in latent space. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.343 + ], + "angle": 0, + "content": "[63] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matthew J. Kusner. Unsupervised point cloud pre-training via occlusion completion. In ICCV, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.397 + ], + "angle": 0, + "content": "[64] Jiyuan Wang, Chunyu Lin, Lang Nie, Kang Liao, Shuwei Shao, and Yao Zhao. Digging into contrastive learning for robust depth estimation with diffusion models. In ACM MM, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[65] Leichen Wang, Bastian Goldluecke, and Carsten Anklam. L2r gan: Lidar-to-radar translation. In ACCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.469 + ], + "angle": 0, + "content": "[66] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic graph cnn for learning on point clouds. TOG, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.482, + 0.523 + ], + "angle": 0, + "content": "[67] Zhen Wang, Qiangeng Xu, Feitong Tan, Mengelei Chai, Shichen Liu, Rohit Pandey, Sean Fanello, Achuta Kadambi, and Yinda Zhang. Mvdd: Multi-view depth diffusion models. In ECCV, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[68] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In CVPR, 2015. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[69] Mutian Xu, Pei Chen, Haolin Liu, and Xiaoguang Han. Toscane: A large-scale dataset for understanding 3d tabletop scenes. In ECCV, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[70] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Tianyou Liang, Guanying Chen, Shuguang Cui, and Xiaoguang Han. Mvimgnet: A large-scale dataset of multi-view images. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.774 + ], + "angle": 0, + "content": "[71] Chengjie Yuan, Yunlei Shi, Qian Feng, Chunyang Chang, Michael Liu, Zhaopeng Chen, Alois Christian Knoll, and Jianwei Zhang. Sim-to-real transfer of robotic assembly with visual inputs using cyclegan and force control. In 2022 IEEE International Conference on Robotics and Biomimetics (ROBIO), pages 1426-1432. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[72] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2100-2110, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[73] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[74] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[75] Xiaoshuai Zhang, Rui Chen, Ang Li, Fanbo Xiang, Yuzhe Qin, Jiayuan Gu, Zhan Ling, Minghua Liu, Peiyu Zeng, Songfang Han, Zhiao Huang, Tongzhou Mu, Jing Xu, and Hao Su. Close the optical sensing domain gap by physics-grounded active stereo sensor simulation. IEEE Transactions on Robotics, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[76] Xiang Zhang, Bingxin Ke, Hayko Riemenschneider, Nando Metzger, Anton Obukhov, Markus Gross, Konrad Schindler, and Christopher Schroers. Betterdepth: Plug-and-play diffusion refiner for zero-shot monocular depth estimation. arXiv preprint arXiv:2407.17952, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.29, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[77] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, 2017. 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2619" + } + ] +] \ No newline at end of file diff --git a/2025/Stable-Sim2Real_ Exploring Simulation of Real-Captured 3D Data with Two-Stage Depth Diffusion/ac3096b4-f6df-4b0e-b9d1-43e46507be43_origin.pdf b/2025/Stable-Sim2Real_ Exploring Simulation of Real-Captured 3D Data with Two-Stage Depth Diffusion/ac3096b4-f6df-4b0e-b9d1-43e46507be43_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5263ba46deaff8a0b4aec507194a37468fe5cf78 --- /dev/null +++ b/2025/Stable-Sim2Real_ Exploring Simulation of Real-Captured 3D Data with Two-Stage Depth Diffusion/ac3096b4-f6df-4b0e-b9d1-43e46507be43_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6180375dbdfcf8f7f459a5f0189ef929a8f76f613a20cccd136dc48760499e53 +size 11948117 diff --git a/2025/Stable-Sim2Real_ Exploring Simulation of Real-Captured 3D Data with Two-Stage Depth Diffusion/full.md b/2025/Stable-Sim2Real_ Exploring Simulation of Real-Captured 3D Data with Two-Stage Depth Diffusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..86cd030407fedd2949274943dff37cb2b2e8b6ea --- /dev/null +++ b/2025/Stable-Sim2Real_ Exploring Simulation of Real-Captured 3D Data with Two-Stage Depth Diffusion/full.md @@ -0,0 +1,339 @@ +# Stable-Sim2Real: Exploring Simulation of Real-Captured 3D Data with Two-Stage Depth Diffusion + +Mutian $\mathrm{Xu}^{1}$ Chongjie $\mathrm{Ye}^{2,3,1}$ Haolin Liu $^{4}$ Yushuang Wu $^{5}$ Jiahao Chang $^{1}$ Xiaoguang Han $^{1,2,3*}$ + +$^{1}$ SSE, CUHKSZ $^{2}$ FNii-Shenzhen + +3Guangdong Provincial Key Laboratory of Future Networks of Intelligence, CUHKSZ + +4Tencent Hunyuan3D 5ByteDance Games + +# Abstract + +3D data simulation aims to bridge the gap between simulated and real-captured 3D data, which is a fundamental problem for real-world 3D visual tasks. Most 3D data simulation methods inject predefined physical priors but struggle to capture the full complexity of real data. An optimal approach involves learning an implicit mapping from synthetic to realistic data in a data-driven manner, but progress in this solution has met stagnation in recent studies. This work explores a new solution path of data-driven 3D simulation, called Stable-Sim2Real, based on a novel two-stage depth diffusion model. The initial stage finetunes Stable-Diffusion to generate the residual between the real and synthetic paired depth, producing a stable but coarse depth, where some local regions may deviate from realistic patterns. To enhance this, both the synthetic and initial output depth are fed into a second-stage diffusion, where diffusion loss is adjusted to prioritize these distinct areas identified by a 3D discriminator. We provide a new benchmark scheme to evaluate 3D data simulation methods. Extensive experiments show that training the network with the 3D simulated data derived from our method significantly enhances performance in real-world 3D visual tasks. Moreover, the evaluation demonstrates the high similarity between our 3D simulated data and real-captured patterns. Project page: mutianxu.github.io/stable-sim2real. + +# 1. Introduction + +In recent years, real-world 3D datasets have played a crucial role in addressing a wide range of tasks in 3D vision and robotics [1, 7, 11, 15, 23, 52, 69, 70]. However, collecting real 3D data is often labor-intensive and time-consuming, and is further complicated by the recent concerns regarding data privacy. In this scenario, synthetic (i.e., simulated) + +![](images/c7f7612d866edb84c842c072879108ad28fd72abb03ac19e782c62dd0c19aab0.jpg) +Figure 1. We delve back into a substantial yet stagnant problem of simulating real-captured 3D data in a data-driven manner, introducing Stable-Sim2Real. Given the depth maps of the 3D synthetic data, our approach applies a two-stage depth diffusion model based on Stable Diffusion [42] to generate realistic depth maps that are fused to produce real 3D point clouds. The initial stage generates stable but coarse depth maps, followed by a second-stage diffusion to enhance regional details. + +data [8, 10, 12, 16, 68] emerges as an alternative data resource that is cost-effective, rapid, and scalable. Nevertheless, models trained using synthetic data may not perform robustly against the real-world signal. + +This problem has catalyzed the development of 3D data simulation techniques, to minimize the gap between simulated and real-captured 3D data. While some studies [2, 30, 33, 39, 75] have explored grounding physical priors to simulate depth sensors, they struggle to capture the full complexity of real-world patterns and are confined to simulation environments, due to the reliance on predefined and explicit physical modeling. A better solution involves learning an implicit mapping from synthetic to realistic data in a data-driven manner, offering greater flexibility and adaptability to real-world diversities. However, only a handful of earlier works [4, 19, 47, 56] have ventured down this path. Hindered by the lack of synthetic-real paired datasets and the dependence on conventional networks like GANs [18], which are prone to training instability and mode collapse [29], these methods have yet to demonstrate the anticipated performance levels. As a result, developing data-driven 3D + +Sim2Real methods has long been a challenge within the community, and has shown little progress. + +In this paper, we aim to explore the data-driven 3D Sim2Real and refocus the community's attention on this critical problem. As a prerequisite, we choose a recent synthetic-real paired dataset, LASA [31], containing 10,412 high-quality 3D shape CAD annotations aligned with the real-world object scans. Driven by LASA, our efforts focus on designing an effective data-driven 3D Sim2Real algorithm. Given the inherent indeterminacy and diversity in real-captured data patterns, deep generative models emerge as a natural fit. Among these models, diffusion models [20, 53-55, 62] recently offer easier training dynamics and stronger generation ability. Yet, due to the lack of 3D data, it is hard to train a 3D diffusion model to get strong 3D priors for 3D simulation. Therefore, we opt to leverage the strong generalization priors of $2D$ diffusion foundation models (e.g., SD - Stable Diffusion [42]) to simulate real 2D depth maps, which are then fused to gain 3D data. This strategy is similar to real 3D data acquisition, where 2D depth is captured and fused into 3D. + +To this end, an intuitive baseline is to sample CAD (i.e., synthetic) depth maps and their paired real depth images from LASA, and then finetune SD to learn an implicit mapping between them. Although such a solution powered by large-scale paired data and the foundation diffusion model has shown effectiveness in conventional image translations, it faces special challenges in our task: while conventional image translations (e.g., denoising/restoration) normally remove noise to produce clean images, our task aims to output noisy depth that is highly uncertain, thereby making the distribution that needs to be learned more complex. + +To address these challenges, we propose Stable-Sim2Real. The core is a novel two-stage depth diffusion model to simulate real-captured noise (a sample is shown in Fig. 1). At Stage-I diffusion, instead of directly generating the corresponding real-world depth map, our model generates the residual (i.e., difference) between real and CAD depth maps. The simulated depth map is then yielded by adding the generated residual with the CAD depth. Empirically, this residual represents the noise added to clean (i.e., synthetic) depth during real-world scanning. Moreover, compared to directly generating noisy real depth, adding noise to CAD depth, which is inherently clean and view-consistent, can produce more stable depth with smaller view variation and better preservation of the original geometry (see Sec. 3.2 for more discussions and analysis). + +However, according to our observations and experiments, while some regions within the Stage-I generated depth map successfully fit the realistic pattern, some unsatisfactory areas still exhibit conspicuous 3D geometric discrepancies between the generation and real captures (shown in Fig. 6). To tackle this issue, our model then revolves + +around discriminating unsatisfactory areas and enhancing the Stage-I generation at local regions. We train a 3D-Aware Discriminator to distinguish between Stage-I generation and real captures at local geometries, which is expected to delineate the distribution boundary between them. If a generated patch can be successfully discriminated, it signifies an unsatisfactory area. Next, we introduce a Stage-II diffusion, conditioned not only on the CAD depth but also on Stage-I generation. More importantly, the diffusion loss is adjusted to prioritize on unsatisfactory regions. In this way, Stage-II diffusion is specialized on learning the distribution of unsatisfactory areas, thereby enhancing Stage-I generation. Finally, the generated depth maps are fused to derive the simulated 3D data. Notably, the binary classifier is only used during the training phase. + +Last, we provide a comprehensive benchmark scheme for 3D data simulation. The key logic is: if a model's real-world performance can be improved after being trained with simulated data, it would validate the effectiveness of simulation methods. We focus on fundamental real-world 3D tasks: 3D shape reconstruction and 3D object/scene understanding. As for shape reconstruction, we pretrain a 3D reconstruction network that takes simulated 3D data as input and outputs clean 3D surfaces. For 3D object/scene understanding, the simulated 3D data is used to pretrain a self-supervised point cloud learning framework. To better assess the performance gain purely from simulated data itself, we conduct few-shot evaluation on pre-trained networks. + +Our contributions are summarized as: + +- We delve back into a substantial yet stagnant problem, data-driven 3D real data simulation, paving a new solution path of training generative foundation models on extensive synthetic-real paired data. +- We present Stable-Sim2Real, a novel 3D real data simulation method based on a two-stage depth diffusion model that can stably simulate real-captured noise. +- We provide a comprehensive scheme to benchmark 3D data simulation, based on few-shot evaluation on networks trained with simulated data for real-world 3D tasks. +- Extensive qualitative and quantitative results show that the data simulated by our method significantly boosts model performance on real-world 3D tasks. + +# 2. Related Work + +# 2.1. Simulation of Real-World Data + +Simulation of 2D real-world data. In the realm of 2D real-world data simulation, two predominant approaches have emerged: domain randomization and domain adaptation. The first line of works [57, 58, 60] randomly perturbs textures, lighting, and material of synthetic data, thereby modeling the synthetic-real domain gap as one kind of variation to endow models with robustness to real-world data. + +However, it brought the challenge of achieving a delicate balance between ensuring that the distribution of augmentations adequately covers that of real-world data and preventing the augmentation from being too extensive to hinder effective model training [24, 37, 72]. Domain adaptation, on the other hand, aims to minimize the domain gap by transferring knowledge from synthetic data to real-world scenarios, leveraging techniques such as adversarial training and self-supervised learning [5, 49, 71]. + +Simulation of 3D real-captured data. Acquiring real-world 3D data and their annotation is labor intensive and inefficient, deriving methods for the simulation of 3D data. Existing simulation methods primarily focus on enhancing the realism of depth information by grounding physical priors in simulators to close the domain gap with 3D data from the physical world [2, 30, 33, 39, 75]. However, these approaches are usually limited by their reliance on predefined and explicit physical modeling. Alternatively, data-driven methods turn to implicitly learn from the distribution in real-captured data, independent of physical modeling [4, 19, 47, 56], offering enhanced adaptability and generalization capabilities to various real-world scenarios. Nonetheless, due to the absence of extensive synthetic-real paired datasets and the reliance on traditional GANs [18], which suffer from training instability and mode collapse [29], these approaches have not achieved the expected performance. Consequently, the research progress of data-driven 3D Sim2Real techniques faces a recent stagnation. + +# 2.2. Image Generation and Translation. + +Conventional generative models. We formulate the 3D Sim2Real task as learning the mapping from synthetic to real-world captured noise, based on depth images. Existing approaches in the domain of image generation and translation include a variety of generative models such as Generative Adversarial Networks (GANs) [18], Variational Autoencoders (VAEs) [28], Cycle-Consistent Adversarial Networks (Cycle-GAN) [77], and Pix2Pix [22]. While GAN-based methods, which have been predominant in image generation before diffusion models, achieved significant milestones in Sim2Real tasks [6, 65, 71], they often struggle with maintaining local details and semantic consistency, leading to sub-optimal performance [13, 29, 32, 36, 38]. + +Diffusion models. Diffusion models, also known as denoising diffusion probabilistic models (DDPMs) [50], have emerged as a powerful alternative to GANs for image synthesis [13, 32, 34, 35, 43, 44, 74]. Unlike GANs, diffusion models offer a more stable training procedure and have been shown to generate images with higher fidelity and better preservation of fine details [21, 42]. Depth diffusion models recently track great attention. They mainly focus on the task + +of depth estimation in a monocular [14, 26, 45, 46, 59, 64] or multi-view [27, 67] manner. Other works also cover conventional image translation tasks, such as depth map superresolution [48] and refinement [76], which produce clean and certain output images. Differently, our work leverages generative models to "generate noise", where the expected output is highly uncertain. Thus, the distribution that needs to be learned is more complex, bringing fundamentally different challenges. + +# 3. Method + +We hope to exploit the strong prior of the diffusion-based foundation model, so we take depth map as our bridge towards the simulation of real-captured 3D data. The preliminaries on the diffusion model are provided in Sec. 3.1. Our Stable-Sim2Real is a two-stage depth diffusion framework, as illustrated in Fig. 2. The first-stage model (Sec. 3.2) learns to generate the depth residual, producing stable but coarse depth maps, where some areas are distinct from the realistic pattern. Therefore, a 3D binary classifier (Sec. 3.3) is introduced to discriminate these distinct areas. The second-stage model (Sec. 3.4) is specialized to enhance the generation of these areas. Finally, the enhanced depth maps are fused to obtain the 3D simulated data (Sec. 3.5). + +# 3.1. Preliminaries on Diffusion Models + +Our model is based on Denoising Diffusion Probabilistic Models (DDPM) [20, 51]. The key of DDPM is to model a data distribution $p(x)$ via building a Markov chain in the data space. It can be divided into two main processes, forward process and backward process. + +Forward diffusion process. Given an initial data $x_0$ , we gradually add random noise $\epsilon \sim \mathcal{N}(0, \mathbf{I})$ to $x_0$ , yielding a series of noisy data, which can be simply formulated as $x_{t+1} = x_t + \epsilon_t$ . After adding noise for multiple steps, the initial distribution of $p(x)$ is expected to be transformed into a standard Gaussian distribution. + +According to DDPM [20], this process can be directly achieved by the following forward diffusion process: + +$$ +q \left(x _ {t}\right) = \sqrt {\alpha_ {t}} x _ {0} + \sqrt {1 - \alpha_ {t}} \boldsymbol {\epsilon}, x _ {0} \sim p (x), t \in \{0, 1, \dots , T \}, \tag {1} +$$ + +where $T$ denotes the number of the time step, $t$ is the current time step, and $\alpha_{t}$ regulates the noise schedule deciding the pace at which the initial data distribution transforms into a standard Gaussian distribution. + +Backward diffusion process. Given a noisy $x_{t}$ from a standard Gaussian distribution as the initial state, $x_{t}$ is denoised to a cleaner $x_{t-1}$ at each time step, which can also be simply formulated as $x_{t-1} = x_{t} - \mu_{\theta}^{\epsilon}(x_{t}, t) + \epsilon_{t}$ in which $\epsilon_{t} \sim \mathcal{N}(0, \sigma_{\mathbf{t}} \mathbf{I})$ and $\mu_{\theta}^{\epsilon}$ is a learnable network to predict + +![](images/1d460675239563ae4e71c0ad18d35994344d4f1bde9eeaf6a4f9ba27d8b9c2a2.jpg) +Figure 2. An overview of our Stable-Sim2Real. Conditioned on CAD (i.e., synthetic) depth maps, our model first generates the residual (i.e., difference) between real-captured and CAD depth maps. By adding the generated residual to the CAD depth, Stage-I diffusion produces stable but coarse depth maps (Sec. 3.2). Next, the Stage-I output is projected into 3D and sent to a 3D discriminator, to identify the unsatisfactory local areas that are distinct from real-captured patterns (Sec. 3.3). Later, our Stage-II diffusion is conditioned on both CAD and Stage-I output, where the training loss of Stage-II is re-weighted to be specialized on enhancing unsatisfactory areas of Stage-I output (Sec. 3.4). Notably, the local discriminating and loss re-weighting are only used during training. + +the injected noise. By repeating this reverse process until the maximum number of steps $T$ , we can obtain the final generated data. + +In DDPM [20], the backward diffusion process is proven to be: + +$$ +x _ {t - 1} = \frac {1}{\sqrt {\alpha_ {t}}} x _ {t} - \frac {1 - \alpha_ {t}}{\sqrt {\alpha_ {t} \left(1 - \Pi_ {\tau = 0} ^ {t} \alpha_ {\tau}\right)}} \mu_ {\theta} ^ {\epsilon} \left(x _ {t}, t;\right) + \sigma_ {t} \epsilon , \tag {2} +$$ + +where $\sigma_{t}$ denotes the standard deviation of $q(x_{t - 1}|x_t,x_0)$ in the backward diffusion to sample random noise. The loss for training $\mu_{\theta}^{\epsilon}$ is: + +$$ +L _ {\theta} = \mathbb {E} _ {\epsilon , t} \| \epsilon - \mu_ {\theta} ^ {\epsilon} (x _ {t}, t) \| ^ {2} \tag {3} +$$ + +# 3.2. Stage-I: Initial Depth Generation + +Residual depth generation with SD. Different from unconditional diffusion models, text-to-image diffusion models focus on generating images with optional text prompts. A representative is Stable Diffusion (SD) [42]. SD trains a denoising U-Net architecture $\mu_{\theta}(x_t,t,c)$ on the latent space of a pre-trained Variational Auto-Encoder (VAE). Here, $c$ represents the supplementary text prompt embedding, usually acquired through methods like CLIP [41]. + +We finetune SD model on synthetic-real paired depth data from LASA [31]. Specifically, given an initial synthetic depth map $S$ and its paired real-captured depth map $R$ from a pre-defined dataset, we first get the initial ground truth residual (i.e., difference) depth map $D_{res}^{gt} = S - R$ . Then we follow ControlNet [73] to add an additional encoder $En$ and decoder $\mu_{\theta}^{\epsilon}$ for finetuning SD. Specifically, we run the forward process to add noise on $D_{res}^{gt}$ following Eq. (1) and encode with $En$ to obtain $x_{t}$ . Next, we send $S$ + +to a pretrained VAE encoder $En$ , generating the latent condition signal. This latent signal $En(S)$ is further processed by an additional encoder $\pmb{f}_{\phi}$ and sent to the decoder blocks of denoising U-Net. We train both the decoder blocks $\mu_{\theta}^{\epsilon}$ and encoder $\pmb{f}_{\phi}$ using the following loss (based on Eq. (3)): + +$$ +L _ {\theta , \phi} = \mathbb {E} _ {\epsilon , S, t} \| \epsilon - \boldsymbol {\mu} _ {\theta} ^ {\epsilon} (\boldsymbol {x} _ {t}, t, \boldsymbol {f} _ {\phi} (E n (\boldsymbol {S}))) \| ^ {2} \tag {4} +$$ + +Learning depth residual helps to produce stable depth. Empirically, this residual represents the noise added to clean (i.e., synthetic) depth during real-world scanning. Moreover, compared to directly generating noisy real depth, adding noise to CAD depth, which is inherently clean and view-consistent, can produce more stable depth with smaller view variation and better preservation of the original geometry. From experimental results, learning residual indeed performs better (see Tab. 5). + +Learning depth residual is theoretically identical to learning real depth. First, assume that real depth is a sample from a Gaussian distribution, denoted by $R \sim \mathcal{N}(\alpha_{\mathbf{R}}, \vartheta)$ , where $\alpha$ denotes the mean and $\vartheta$ represents the variance. Notably, $\vartheta$ can naturally represent real-captured noise. Therefore, synthetic depth can be denoted by $S \sim \mathcal{N}(\alpha_{\mathbf{S}}, 0)$ , where the variance is zero as there is no real-captured noise. + +In our design, as the diffusion model is inherently designed to learn the transformation from a random noise $\mathcal{N}(\mathbf{0},\mathbf{I})$ to any other Gaussian distribution [53], we leverage it to model the distribution of the depth residual that can be denoted by $\mathcal{N}(\alpha_{\mathbf{R}} - \alpha_{\mathbf{S}},\vartheta)$ . Last, by adding the depth residual to synthetic depth, the output distribution can be formulated as: $\mathcal{N}(\alpha_{\mathbf{R}} - \alpha_{\mathbf{S}},\vartheta) + \mathcal{N}(\alpha_{\mathbf{S}},0) = \mathcal{N}(\alpha_{\mathbf{R}},\vartheta)$ , + +which is exactly identical to the previously assumed real-captured noise distribution. + +# 3.3. 3D-Aware Local Discriminating + +Unsatisfactory areas distinct from real patterns. Based on our observations, the noise captured in real data is spatially-variant, meaning different areas within a real depth map may display varying levels of perturbation. However, our Stage-I diffusion tends to introduce spatially-uniform noise across the entire input CAD depth. This discrepancy may stem from a bias within SD, as also discussed in [9]. As a result, although certain regions of the generated depth map align well with real perturbation patterns, there remain noticeable areas that deviate from reality. This leads to evident 3D geometric discrepancies between Stage-I outputs and real captures (as illustrated in Fig. 6). + +# Identifying unsatisfactory areas with 3D discriminator. + +Our design logic to discriminate unsatisfactory areas is that if the generated areas diverge significantly from real patterns, there should exist a distribution boundary of their 3D geometries. To this end, we propose employing a 3D binary classifier to discriminate this boundary. Initially, we split all generated and real-captured depth maps into different patches representing different areas, which are then projected to form 3D point patches. Subsequently, we leverage PointNet [40] as a 3D binary classifier to differentiate whether the input point patches come from model generation or real captures. The CAD depth maps are also concatenated in the input as additional guidance. The successful classification of a generated patch indicates that its corresponding area is distinct from real captures. Notably, the binary classifier is solely used during training. + +# 3.4. Stage-II: Unsatisfactory Area Enhancement + +Improving unsatisfactory areas by re-weighting denoising loss. After identifying these local patches, we introduce a second-stage depth refinement model that differs from the Stage-I diffusion in two crucial aspects: 1) The diffusion model is now conditioned on the concatenation of synthetic depth $S$ and the previously generated depth residual $D^{I}$ from Stage I. Here, $D^{I} = D_{res}^{I} + S$ , where $D_{res}^{I}$ is the output residual depth from Stage-I model. Thus, the training loss defined in Eq. (4) is updated to: + +$$ +L _ {\chi , \psi} = \mathbb {E} _ {\boldsymbol {\epsilon}, \boldsymbol {D} ^ {I}, \boldsymbol {S}, t} \left\| \boldsymbol {\epsilon} - \boldsymbol {\mu} _ {\chi} ^ {\epsilon} \left(\boldsymbol {x} _ {t}, t, \boldsymbol {f} _ {\psi} \left(E n \left(\boldsymbol {D} _ {r e s} ^ {I}, \boldsymbol {S}\right)\right)\right) \right\| ^ {2}, \tag {5} +$$ + +2) More importantly, the diffusion loss is re-weighted to prioritize the enhancement of the identified unsatisfactory regions, making the final Stage-II loss $L_{\chi ,\psi}$ to be: + +$$ +L _ {\chi , \psi} = \omega L _ {\chi , \psi} ^ {s i m} + \lambda L _ {\chi , \psi} ^ {d i s}, \tag {6} +$$ + +where $L_{\chi, \psi}^{sim}$ indicates the loss on similar regions, while $L_{\chi, \psi}^{dis}$ denotes the loss on previously identified distinct (i.e., + +unsatisfactory) areas. $\omega$ and $\lambda$ are weight coefficients, in which $\lambda >\omega$ . In our experiment, setting $\omega = 0.5$ and $\lambda = 1.5$ produces the best result. + +Our Stage-II diffusion model is specialized on learning the distribution of unsatisfactory areas in Stage-I generation. Meanwhile, by conditioning on Stage-I output and assigning less weight to satisfied areas, the generation on satisfied outputs is also preserved in Stage-II, as in Fig. 6. + +# 3.5. Creating Real-World 3D Data + +Two-stage real depth simulation. In the initial stage, we sample a random distribution $\epsilon$ from $\mathcal{N}(\mathbf{0},\mathbf{I})$ and feed it into the denoising U-Net $\mu_{\theta}$ of Stage-I diffusion model. The model runs a backward diffusion process for $T$ steps to generate the initial depth residual $D_{res}^{I}$ , following Eq. (2) with the synthetic depth $S$ serving as the condition. Now we can obtain the initial depth map $D^{I} = D_{res}^{I} + S$ . + +Moving to the second stage, we repeat a similar backward diffusion process to generate the residual depth $D_{res}^{II}$ , using the denoising U-Net $\mu_{\chi}$ of our Stage-II diffusion model, conditioned on both synthetic depth $S$ and Stage-I output $D^{I}$ . The final depth map $D^{II} = D_{res}^{II} + S$ . DDIM sampler [51] is utilized during the whole inference phase. + +Depth fusion to 3D data. Given the 3D synthetic object, we follow LASA [31] to embed CAD objects into 3D scenes. This allows us to render synthetic depth images, simulating the occlusions in real-world captures. Then we conduct the aforementioned two-stage real depth simulation on rendered synthetic depth maps, followed by a depth fusion algorithm to create simulated 3D data. + +# 4. Experiments + +We provide a new benchmark scheme for 3D data simulation. To begin with, different simulation methods are performed on synthetic datasets. The 3D simulated data are then utilized to train networks for tackling fundamental real-world 3D visual tasks: 3D instance reconstruction, 3D object/scene understanding. The key logic of our evaluation is: if the model's real-world performance can be improved after being trained with simulated data, it would validate that the simulated data is close to real distributions, demonstrating the effectiveness of simulation methods. + +# 4.1. Preliminaries + +Training of Stable-Sim2Real. We train our model by finetuning the official StableDiffusion V2.1 [42]. The training dataset is a recent 3D synthetic-real paired dataset, LASA [31]. LASA is a large-scale dataset that contains 10,412 unique CAD models across 17 categories. Skilled artists are employed to meticulously craft aligned CAD models from 3D object real-world scans of [3]. LASA provides CAD-real depth map pairs, rendered based on camera + +
MethodChairSofaTableCabinetBedShelf
w/o sim. data24.5 / 12.3 / 20.155.6 / 7.02 / 19.825.8 / 19.3 / 20.357.1 / 8.83 / 24.140.1 / 6.14 / 17.312.8 / 6.74 / 23.7
Cycle-GAN [77]25.1 / 12.1 / 20.955.2 / 6.98 / 20.526.1 / 19.1 / 19.556.5 / 8.24 / 25.140.6 / 5.98 / 16.913.1 / 6.08 / 24.9
Pix2Pix [22]24.9 / 11.0 / 20.556.4 / 6.74 / 21.326.6 / 18.5 / 22.157.4 / 7.83 / 25.541.3 / 6.01 / 17.814.2 / 6.15 / 26.2
Giga-GAN [25]26.2 / 10.3 / 22.158.9 / 5.87 / 22.429.3 / 16.4 / 24.360.5 / 7.01 / 26.944.2 / 5.25 / 19.318.2 / 5.87 / 27.8
Stable-Diffusion [42]27.3 / 9.51 / 25.559.4 / 5.93 / 23.629.7 / 15.8 / 24.761.6 / 6.92 / 27.545.1 / 5.31 / 20.119.1 / 5.85 / 28.3
Ours29.2 / 7.83 / 27.562.9 / 4.09 / 25.131.2 / 14.1 / 26.863.2 / 5.58 / 28.646.3 / 4.23 / 21.818.8 / 5.02 / 31.6
+ +Table 1. Quantitative comparison of few-shot evaluation on 3D instance reconstruction, where the model is pretrained using simulated data generated by different methods. "w/o sim. data" indicates pretraining using no simulated data but only synthetic data. Evaluation metrics are mIoU / Chamfer L2 / F-score respectively. Higher mIoU and F-score are better while lower Chamfer L2 is better. Chamfer L2 is scaled by 1,000. All results are reproduced using the official code of DisCo [31]. + +
MethodChairSofaTableCabinetBedShelf
w/o sim. data19.4 / 14.7 / 15.846.2 / 9.92 / 15.720.5 / 23.6 / 15.951.2 / 11.3 / 20.233.6 / 13.1 / 10.911.1 / 9.23 / 21.6
Cycle-GAN [77]20.3 / 13.1 / 16.548.9 / 9.03 / 18.923.2 / 22.4 / 17.551.9 / 10.8 / 21.435.2 / 12.5 / 12.111.7 / 8.87 / 23.5
Pix2Pix [22]19.3 / 12.3 / 17.449.7 / 9.51 / 18.822.6 / 21.5 / 18.852.7 / 11.2 / 21.137.7 / 10.0 / 13.812.5 / 8.91 / 24.3
Giga-GAN [25]22.4 / 11.8 / 19.553.0 / 7.69 / 19.423.1 / 19.3 / 21.654.3 / 10.1 / 21.940.3 / 7.95 / 15.414.2 / 7.81 / 26.9
Stable-Diffusion [42]23.3 / 11.7 / 20.454.1 / 7.98 / 19.224.2 / 17.9 / 22.655.1 / 9.07 / 22.440.4 / 7.89 / 16.215.8 / 6.72 / 28.3
Ours27.1 / 9.31 / 24.860.6 / 5.29 / 22.328.5 / 16.1 / 24.760.6 / 7.51 / 25.943.2 / 6.23 / 18.017.3 / 5.93 / 30.5
+ +Table 2. Quantitative comparison of zero-shot evaluation on 3D instance reconstruction, when the model is solely pretrained using simulated data generated by different methods. + +information during the real-world collection. See the supplementary material for more implementation details. + +Methods in comparison. Since our method operates on depth images, we select several image-to-image generation methods for comparison: 1) Cycle-Consistent Adversarial Networks (Cycle-GAN) [77], which is a representative generative network when paired training data are not available, so we train Cycle-GAN on the pair-shuffled LASA training data. 2) Pix2Pix [22], which is a classical image generation method. 3) Giga-GAN [25], a state-of-the-art GAN-based method. 4) Stable Diffusion [42], our main baseline model. Both 2), 3), 4) are trained using the paired data from LASA. + +# 4.2.3D Instance Reconstruction + +Setup. 3D instance reconstruction aims to recover the full geometry from real-captured partial point clouds or multiview images. We utilize a recent state-of-the-art diffusion-based reconstruction network, DisCo, which is proposed in LASA [31]. It is originally pretrained on three synthetic datasets (i.e., ShapeNet [8], ABO [10], and 3D-Future [17]) and finetuned on LASA real training data. During pretraining, DisCo simply constructs partial point clouds by adding random perturbation to synthetic data. Instead, in our experiment, to simulate partial point clouds for pretraining, we perform simulation using our method and other baseline models based on the original three synthetic datasets used to pretrain DisCo, and replace half of the original partial point clouds with the simulated 3D data. The other configurations all follow the original DisCo. + +Subsequently, we employ few-shot and zero-shot learn + +ing configurations to evaluate: We randomly sample only a few LASA real training data to finetune the pretrained network. In this way, the real-world knowledge of the network will mostly come from the pretraining data (i.e., the simulated data produced by our method and others) instead of the finetuning data, which can better reflect the performance gain from the simulated data itself. The pretrained model is finetuned with $20\sim 30\%$ of the original training samples for few-shot learning and is evaluated on the original LASA validation set. For zero-shot learning, we directly evaluate the pretrained network without fine-tuning. + +Why not perform simulation on LASA synthetic data and compare the performance using the simulated data with that using LASA paired real GT? Since our StableSim2Real and other baseline techniques are trained on the LASA training set, it is unfeasible to conduct further inference on LASA, as this would not verify the actual model performance but only reflect the performance caused by overfitting. We can also not conduct simulation on a third-party paired dataset similar to LASA for 3D reconstruction, as there is currently no such paired dataset. + +Quantitative results. Following [31], we utilize mean Intersection over Union (mIoU), L2 chamfer distance, and $1\%$ F-score metrics. Initially, we normalize both the results and the ground-truth CAD objects to fall within the range of -0.5 to 0.5, according to which we calculate the aforementioned metrics between them. As shown in Tab. 1, under all few-shot settings, the model pretrained using the simulated data from our method achieves the best performance, and significantly surpasses the model using no simulated data (i.e., only add random perturbation to synthetic data) for + +![](images/414b82cce4a2b9fa53f937d90d7bdcabbc409f262a00e4fe4c768b326cc6e2da.jpg) +Figure 3. Qualitative comparison of few-shot evaluation on 3D instance reconstruction. "w/o sim." denotes using no simulated data for pretraining. + +pretraining. Moreover, using the simulated data produced by Cycle-GAN or Pix2Pix does not improve the model performance compared to that using no simulated data, possibly attributed to their inferior simulation quality, introducing ambiguity and complexity during pretraining, making it hard for the model to learn useful real-world knowledge. Moreover, as listed in Tab. 2, under zero-shot setting, when solely using simulated data for pretraining without any finetuning, the model trained with our simulated data demonstrates more conspicuous effectiveness. + +Qualitative results. Fig. 3 shows the qualitative comparison, where our simulated data can help the model to gain the best reconstruction accuracy and diversity. More qualitative results are provided in the supplementary material. + +# 4.3. 3D Object/Scene Understanding + +Setup. We select OcCo [63] as our backbone, which is a self-supervised point cloud pretraining framework. OcCo is originally pretrained on ModelNet [68], and finetuned on real-scanned datasets - ScanObjectNN [61] for object classification, and S3DIS [1] for scene semantic segmentation. ModelNet contains 12,311 synthetic CAD objects from 40 classes. ScanObjectNN includes $\sim 15,000$ real-scanned objects across 15 categories. S3DIS encompasses 271 indoor rooms and 13 classes. The classification task outputs the class of the whole 3D object, and the semantic segmentation predicts the category of each point in the 3D scene. + +In our experiment, we apply our method and other baselines on ModelNet to generate simulated data that are further used to replace half of the original ModelNet synthetic data to pretrain OcCo. Next, the pretrained model is again evaluated under the few-shot learning setting. A typical setting is "K-way $N$ -shot", which indicates that $K$ classes are randomly selected during finetuning, and each category contains $N$ samples. The finetuned model is evaluated on the unseen samples picked from each of the $K$ classes. We select DGCNN [66] as the backbone for OcCo. For few-shot learning, we employ the "5-way, 10-shot" setting. For S3DIS, we follow the original OcCo to report the results + +
Methodw/o sim.C-GAN [77]P2P [22]G-GAN [25]SD [42]Ours
Acc. (%)72.472.572.174.274.677.5
mIoU (%)47.948.047.749.148.951.2
+ +Table 3. Quantitative comparison of few-shot evaluation on 3D object classification and 3D scene semantic segmentation. "w/o sim" means using no simulated data for pretraining. + +
Methodw/o sim.C-GAN [77]P2P [22]G-GAN [25]SD [42]Ours
Metrics59.2/3.54/25.360.4/3.21/25.261.5/3.02/26.462.1/2.86/26.762.3/2.83/27.562.9/2.57/28.2
+ +Table 4. The mIoU / Chamfer L2 / F-score between simulated data and real-captures on LASA [31] validation set. + +using 6-fold cross-validation. All the other configurations follow the original OcCo. + +Qualitative results. We report the classification accuracy and segmentation mean Intersection over Union (mIoU) in Tab. 3. The model pretrained using our simulated data gets the best results across both 3D object classification and 3D scene understanding tasks. By pretraining with our simulated data, the model can acquire valuable real-world knowledge that helps downstream tasks. In contrast, pretraining with the low-quality simulated data from CycleGAN or Pix2Pix may even degrade the performance. + +# 4.4. Is the simulated data close to real captures? + +As previously mentioned, our Stable-Sim2Real is trained on LASA training set. In this section, we aim to evaluate the data simulation quality by directly performing inference and evaluation on LASA validation set. Tab. 4 compares the mIoU, Chamfer L2, and F-score metrics between the simulated data generated by different methods and the real captured Ground Truth (GT). Although not precise, this result can partially prove that our simulated data closely resembles real-world captures. Fig. 4 shows simulation results on depth maps and 3D point clouds. It demonstrates that while our simulated data closely aligns with real-world captures, subtle differences persist, highlighting the diversity. + +We also provide the simulation results on ShapeNet [8] in Fig. 5. Despite the absence of real-captured Ground Truth (GT) in ShapeNet, the outcomes still show that the simulated results from our output closely resemble real captures, where the generated geometry is not largely broken, with certain realistic perturbations introduced. + +# 4.5. Ablation Studies and Analysis + +Module impacts. To evaluate the impact of each key design in our framework, we conduct ablation studies on the 3D instance reconstruction task, specifically on the table category. Similar to Tab. 2, we conduct experiments under zero-shot protocol. This allows for a focused evaluation of the module's impact without the influence of additional factors. Other setups remain consistent with Sec. 4.2. + +![](images/67b8575547711c1a0e0a0c8bc82a52321364a1a7a3c0690e3cdec57bd75550dd.jpg) +Figure 4. Qualitative simulation results on LASA [31]. + +![](images/e9db5cfe08a8b76f9d87cca1c13c177fcb31b0ffe4d5424382446be9a81cddb7.jpg) +Figure 5. Qualitative simulation results on ShapeNet [8]. + +1) Residual generation. In Tab. 5 (w/o res.), directly generating the real-world depth map without using residual generation degrades the model performance, verifying the efficacy of our residual generation. +2) Loss re-weighting. In our Stage-II model, we reweight the diffusion loss to focus more on unsatisfactory areas. In Tab. 5 (w/o re-w.), the result w/o loss re-weighting is provided, showing its effectiveness. +3) 3D binary discriminator. Instead of training a 3D binary discriminator, another simple way is to directly set a threshold of the absolute difference between Stage-I generation and real-world GT to decide distinct areas. We tried various difference thresholds and reported the best result in Tab. 5 (w/o bin.), which hurts the model performance. To explain, our output depth is gained by adding subtle noise to synthetic depth. Thus, the magnitude of the difference between our output and its corresponding real GT is similar at 2D/pixel space, where it is impractical to find an appropriate threshold to identify distinct/unsatisfactory areas. Instead, their difference in distribution space is relatively more significant, so using a 3D binary classifier to delineate their distribution boundary is a more feasible way. +4) Number of stages. Theoretically, we can iteratively improve the generation quality by adding more stages. However, as shown in Tab. 5, adding Stage-III model does + +
Methodw/o res.w/o re-w.w/o bin.Stage-IIIStage-IVStage-II
Metrics26.4/18.2/22.127.1/17.9/23.327.3/17.5/23.628.6/ 16.1/24.528.0/16.9/24.128.5/16.1/24.7
+ +Table 5. Quantitative ablations of zero-shot evaluation on 3D instance reconstruction of table category. Evaluation metrics are mIoU / Chamfer L2 / F-score respectively. + +![](images/2ab4154292e764e9b073cdd8aaee375cb02008785b2a33da79a2a12fd61425a9.jpg) +Figure 6. Qualitative simulation results at the patch level. + +not improve the model performance, and Stage-IV model even slightly degrades the result. To analyze, Stage-II model has been sufficient to identify most of the unsatisfactory areas and improve their generation quality. In addition, adding redundant stages may also cause overfitting issues that hurt the performance. + +Analyzing Stage-II model at 3D patch-level. 1) We compare simulation results with real-captured Ground Truth (GT) at the patch level. As illustrated in Fig. 6, our Stage-II model effectively refines previously unsatisfactory (i.e., dissimilar to GT) patches from Stage-I output to be similar to GT, while keeping previously satisfactory (i.e., similar to GT) patches almost unchanged. +2) When sending the newly generated output patches from Stage-II to our 3D binary discriminator, the classification accuracy notably drops from $67.8\%$ to $32.5\%$ , suggesting that the generated patches successfully approximate the real-captured distributions, making it challenging for the classifier to discriminate. This again verifies the effectiveness of our Stage-II model. See the supplementary material for more ablations of the binary classifier. + +# 5. Conclusion + +We introduce Stable-Sim2Real, a two-stage depth diffusion model for 3D real data simulation. Initially, our method learns the residual between synthetic and real depth to generate a stable but coarse depth, where some areas may deviate from realistic patterns. Next, both synthetic and initial output depth are fed into the second-stage model, where diffusion loss prioritizes enhancing distinct regions identified by a binary classifier. Experiments show that the 3D simulated data generated from our method can improve the network performance in real-world 3D tasks, and our simulated data is highly similar to real captures. More limitations are discussed in the supplementary material. + +# Acknowledgement + +The work was supported in part by Guangdong S&T Programme with Grant No. 2024B0101030002, the Basic Research Project No. HZQB-KCZYZ-2021067 of Hetao Shenzhen-HK S&T Cooperation Zone, Guangdong Provincial Outstanding Youth Fund (No. 2023B1515020055), No. NSFC 62172348, the Shenzhen Outstanding Talents Training Fund 202002, the Guangdong Research Projects No. 2017ZT07X152 and No. 2019CX01X104, the Guangdong Provincial Key Laboratory of Future Networks of Intelligence (Grant No. 2022B1212010001), and the Shenzhen Key Laboratory of Big Data and Artificial Intelligence (Grant No. ZDSYS201707251409055). + +# References + +[1] I. Armeni, O. Sener, A. R. Zamir, H. Jiang, I. Brilakis, M. Fischer, and S. Savarese. 3d semantic parsing of large-scale indoor spaces. In CVPR, 2016. 1, 7 +[2] Kaixin Bai, Lei Zhang, Zhaopeng Chen, Fang Wan, and Jianwei Zhang. Close the sim2real gap via physically-based structured light synthetic data simulation. In ICRA, 2024. 1, 3 +[3] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Yuri Feigin, Peter Fu, Thomas Gebauer, Daniel Kurz, Tal Dimry, Brandon Joffe, Arik Schwartz, and Elad Shulman. ARK-scenes: A diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In NeurlPS, 2021. 5 +[4] Konstantinos Bousmalis, Nathan Silberman, David Dohan, Dumitru Erhan, and Dilip Krishnan. Unsupervised pixel-level domain adaptation with generative adversarial networks. In CVPR, 2017. 1, 3 +[5] Konstantinos Bousmalis, Alex Irpan, Paul Wohlhart, Yunfei Bai, Matthew Kelcey, Mrinal Kalakrishnan, Laura Downs, Julian Ibarz, Peter Pastor, Kurt Konolige, et al. Using simulation and domain adaptation to improve efficiency of deep robotic grasping. In ICRA, 2018. 3 +[6] Marc Bresson, Yang Xing, and Weisi Guo. Sim2real: Generative ai to enhance photorealism through domain transfer with gan and seven-chanel- $360^{\circ}$ -paired-images dataset. Sensors, 24(1):94, 2023. 3 +[7] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. In 3DV, 2017. 1 +[8] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 1, 6, 7, 8 +[9] Hansheng Chen, Ruoxi Shi, Yulin Liu, Bokui Shen, Jiayuan Gu, Gordon Wetzstein, Hao Su, and Leonidas Guibas. Generic 3d diffusion adapter using controlled multi-view editing. arXiv preprint arXiv:2403.12032, 2024. 5 + +[10] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksen, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In CVPR, 2022. 1, 6 +[11] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 1 +[12] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, 2023. 1 +[13] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. In NeurlPS, 2021. 3 +[14] Yiquan Duan, Xianda Guo, and Zheng Zhu. Diffusiondepth: Diffusion denoising approach for monocular depth estimation. In ECCV, 2025. 3 +[15] Hao-Shu Fang, Chenxi Wang, Hongjie Fang, Minghao Gou, Jirong Liu, Hengxu Yan, Wenhai Liu, Yichen Xie, and Cewu Lu. Anygrasp: Robust and efficient grasp perception in spatial and temporal domains. T-RO, 2023. 1 +[16] Huan Fu, Bowen Cai, Lin Gao, Ling-Xiao Zhang, Jiaming Wang, Cao Li, Qixun Zeng, Chengyue Sun, Rongfei Jia, Binqiang Zhao, et al. 3d-front: 3d furnished rooms with layouts and semantics. In ICCV, 2021. 1 +[17] Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3d-future: 3d furniture shape with texture. *IJCT*, 2021. 6 +[18] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 2020. 1, 3 +[19] Xiao Gu, Yao Guo, Fani Deligianni, and Guang-Zhong Yang. Coupled real-synthetic domain adaptation for real-world deep depth enhancement. TIP, 2020. 1, 3 +[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3, 4 +[21] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. Journal of Machine Learning Research, 23(47):1-33, 2022. 3 +[22] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In CVPR, 2017. 3, 6, 7 +[23] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engil Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In CVPR, 2014. 1 +[24] Cheongwoong Kang, Wonjoon Chang, and Jaesik Choi. Balanced domain randomization for safe reinforcement learning. Applied Sciences, 14(21):9710, 2024. 3 +[25] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In CVPR, 2023. 6, 7 +[26] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In CVPR, 2024. 3 + +[27] Numair Khan, Min H Kim, and James Tompkin. Differentiable diffusion for dense depth estimation from multi-view images. In CVPR, 2021. 3 +[28] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. STAT, 1050:1, 2014. 3 +[29] Naveen Kodali, Jacob D. Abernethy, James Hays, and Zsolt Kira. On convergence and stability of gans. arXiv preprint arXiv:1705.07215, 2017. 1, 3 +[30] Michael J Landau, Benjamin Y Choo, and Peter A Beling. Simulating kinect infrared and depth images. IEEE transactions on cybernetics, 2015. 1, 3 +[31] Haolin Liu, Chongjie Ye, Yinyu Nie, Yingfan He, and Xiaoguang Han. Lasa: Instance reconstruction from real scans using a large-scale aligned shape annotation dataset. In CVPR, 2024. 2, 4, 5, 6, 7, 8 +[32] Naoki Matsunaga, Masato Ishii, Akio Hayakawa, Kenji Suzuki, and Takuya Narihira. Fine-grained image editing by pixel-wise guidance using diffusion models. arXiv preprint arXiv:2212.02024, 2022. 3 +[33] Stephan Meister, Rahul Nair, and Daniel Kondermann. Simulation of time-of-flight sensors using global illumination. In VMV, 2013. 1, 3 +[34] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2021. 3 +[35] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3 +[36] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 3 +[37] Alexander Pashevich, Robin Strudel, Igor Kalevatykh, Ivan Laptev, and Cordelia Schmid. Learning to augment synthetic images for sim2real policy transfer. In IROS, 2019. 3 +[38] Duo Peng, Ping Hu, Qiuhong Ke, and Jun Liu. Diffusion-based image translation with label guidance for domain adaptive semantic segmentation. In ICCV, 2023. 3 +[39] Benjamin Planche, Ziyan Wu, Kai Ma, Shanhui Sun, Stefan Kluckner, Oliver Lehmann, Terrence Chen, Andreas Hutter, Sergey Zakharov, Harald Kosch, et al. Depthsynth: Realtime realistic synthetic data generation from cad models for 2.5 d recognition. In 3DV, 2017. 1, 3 +[40] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 5 +[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 4 +[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 3, 4, 5, 6, 7 + +[43] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 3 +[44] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. TPAMI, 2022. 3 +[45] Saurabh Saxena, Abhishek Kar, Mohammad Norouzi, and David J Fleet. Monocular depth estimation using diffusion models. arXiv preprint arXiv:2302.14816, 2023. 3 +[46] Saurabh Saxena, Charles Herrmann, Junhwa Hur, Abhishek Kar, Mohammad Norouzi, Deqing Sun, and David J Fleet. The surprising effectiveness of diffusion models for optical flow and monocular depth estimation. NeurIPS, 2024. 3 +[47] Yuefan Shen, Yanchao Yang, Youyi Zheng, C Karen Liu, and Leonidas J Guibas. Dcl: Differential contrastive learning for geometry-aware depth synthesis. IEEE Robotics and Automation Letters, 2022. 1, 3 +[48] Yuan Shi, Huiyun Cao, Bin Xia, Rui Zhu, Qingmin Liao, and Wenming Yang. Dsr-diff: Depth map super-resolution with diffusion model. Pattern Recognition Letters, 184:225-231, 2024. 3 +[49] Ashish Shrivastava, Tomas Pfister, Oncel Tuzel, Joshua Susskind, Wenda Wang, and Russell Webb. Learning from simulated and unsupervised images through adversarial training. In CVPR, 2017. 3 +[50] Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 3 +[51] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3, 5 +[52] Shuran Song, Samuel P. Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In CVPR, 2015. 1 +[53] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In NeurIPS, 2019, 2, 4 +[54] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. In NeurIPS, 2020. +[55] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2 +[56] Chris Sweeney, Greg Izatt, and Russ Tedrake. A supervised approach to predicting noise in depth images. In ICRA, 2019. 1, 3 +[57] Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. In IROS, 2017. 2 +[58] Josh Tobin, Lukas Biewald, Rocky Duan, Marcin Andrychowicz, Ankur Handa, Vikash Kumar, Bob McGrew, Alex Ray, Jonas Schneider, Peter Welinder, et al. Domain randomization and generative models for robotic grasping. In IROS, 2018. 2 + +[59] Fabio Tosi, Pierluigi Zama Ramirez, and Matteo Poggi. Diffusion models for monocular depth estimation: Overcoming challenging conditions. arXiv preprint arXiv:2407.16698, 2024.3 +[60] Jonathan Tremblay, Aayush Prakash, David Acuna, Mark Brophy, Varun Jampani, Cem Anil, Thang To, Eric Cameracci, Shaad Boochoon, and Stan Birchfield. Training deep networks with synthetic data: Bridging the reality gap by domain randomization. In CVPRW, 2018. 2 +[61] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. 7 +[62] Arash Vahdat, Karsten Kreis, and Jan Kautz. Score-based generative modeling in latent space. In NeurIPS, 2021. 2 +[63] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matthew J. Kusner. Unsupervised point cloud pre-training via occlusion completion. In ICCV, 2021. 7 +[64] Jiyuan Wang, Chunyu Lin, Lang Nie, Kang Liao, Shuwei Shao, and Yao Zhao. Digging into contrastive learning for robust depth estimation with diffusion models. In ACM MM, 2024. 3 +[65] Leichen Wang, Bastian Goldluecke, and Carsten Anklam. L2r gan: Lidar-to-radar translation. In ACCV, 2020. 3 +[66] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic graph cnn for learning on point clouds. TOG, 2019. 7 +[67] Zhen Wang, Qiangeng Xu, Feitong Tan, Mengelei Chai, Shichen Liu, Rohit Pandey, Sean Fanello, Achuta Kadambi, and Yinda Zhang. Mvdd: Multi-view depth diffusion models. In ECCV, 2025. 3 +[68] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In CVPR, 2015. 1, 7 +[69] Mutian Xu, Pei Chen, Haolin Liu, and Xiaoguang Han. Toscane: A large-scale dataset for understanding 3d tabletop scenes. In ECCV, 2022. 1 +[70] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Tianyou Liang, Guanying Chen, Shuguang Cui, and Xiaoguang Han. Mvimgnet: A large-scale dataset of multi-view images. In CVPR, 2023. 1 +[71] Chengjie Yuan, Yunlei Shi, Qian Feng, Chunyang Chang, Michael Liu, Zhaopeng Chen, Alois Christian Knoll, and Jianwei Zhang. Sim-to-real transfer of robotic assembly with visual inputs using cyclegan and force control. In 2022 IEEE International Conference on Robotics and Biomimetics (ROBIO), pages 1426-1432. IEEE, 2022. 3 +[72] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2100-2110, 2019. 3 +[73] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 4 + +[74] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3 +[75] Xiaoshuai Zhang, Rui Chen, Ang Li, Fanbo Xiang, Yuzhe Qin, Jiayuan Gu, Zhan Ling, Minghua Liu, Peiyu Zeng, Songfang Han, Zhiao Huang, Tongzhou Mu, Jing Xu, and Hao Su. Close the optical sensing domain gap by physics-grounded active stereo sensor simulation. IEEE Transactions on Robotics, 2023. 1, 3 +[76] Xiang Zhang, Bingxin Ke, Hayko Riemenschneider, Nando Metzger, Anton Obukhov, Markus Gross, Konrad Schindler, and Christopher Schroers. Betterdepth: Plug-and-play diffusion refiner for zero-shot monocular depth estimation. arXiv preprint arXiv:2407.17952, 2024. 3 +[77] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, 2017. 3, 6, 7 \ No newline at end of file diff --git a/2025/Stable-Sim2Real_ Exploring Simulation of Real-Captured 3D Data with Two-Stage Depth Diffusion/images.zip b/2025/Stable-Sim2Real_ Exploring Simulation of Real-Captured 3D Data with Two-Stage Depth Diffusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a53a3d5e17350c065ddbff3e7fb28def0fe855f6 --- /dev/null +++ b/2025/Stable-Sim2Real_ Exploring Simulation of Real-Captured 3D Data with Two-Stage Depth Diffusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:69d158ed5338fc362893b049ffc544ba08f8e0a26776710a571864616971a3ea +size 427146 diff --git a/2025/Stable-Sim2Real_ Exploring Simulation of Real-Captured 3D Data with Two-Stage Depth Diffusion/layout.json b/2025/Stable-Sim2Real_ Exploring Simulation of Real-Captured 3D Data with Two-Stage Depth Diffusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5be576979d88184b9c8b0305ab0d725556d10dfa --- /dev/null +++ b/2025/Stable-Sim2Real_ Exploring Simulation of Real-Captured 3D Data with Two-Stage Depth Diffusion/layout.json @@ -0,0 +1,9346 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 84, + 102, + 526, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 102, + 526, + 140 + ], + "spans": [ + { + "bbox": [ + 84, + 102, + 526, + 140 + ], + "type": "text", + "content": "Stable-Sim2Real: Exploring Simulation of Real-Captured 3D Data with Two-Stage Depth Diffusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 159, + 566, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 159, + 566, + 177 + ], + "spans": [ + { + "bbox": [ + 61, + 159, + 566, + 177 + ], + "type": "text", + "content": "Mutian " + }, + { + "bbox": [ + 61, + 159, + 566, + 177 + ], + "type": "inline_equation", + "content": "\\mathrm{Xu}^{1}" + }, + { + "bbox": [ + 61, + 159, + 566, + 177 + ], + "type": "text", + "content": " Chongjie " + }, + { + "bbox": [ + 61, + 159, + 566, + 177 + ], + "type": "inline_equation", + "content": "\\mathrm{Ye}^{2,3,1}" + }, + { + "bbox": [ + 61, + 159, + 566, + 177 + ], + "type": "text", + "content": " Haolin Liu" + }, + { + "bbox": [ + 61, + 159, + 566, + 177 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 159, + 566, + 177 + ], + "type": "text", + "content": " Yushuang Wu" + }, + { + "bbox": [ + 61, + 159, + 566, + 177 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 61, + 159, + 566, + 177 + ], + "type": "text", + "content": " Jiahao Chang" + }, + { + "bbox": [ + 61, + 159, + 566, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 159, + 566, + 177 + ], + "type": "text", + "content": " Xiaoguang Han" + }, + { + "bbox": [ + 61, + 159, + 566, + 177 + ], + "type": "inline_equation", + "content": "^{1,2,3*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 235, + 180, + 392, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 180, + 392, + 193 + ], + "spans": [ + { + "bbox": [ + 235, + 180, + 392, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 235, + 180, + 392, + 193 + ], + "type": "text", + "content": "SSE, CUHKSZ " + }, + { + "bbox": [ + 235, + 180, + 392, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 235, + 180, + 392, + 193 + ], + "type": "text", + "content": "FNii-Shenzhen" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 139, + 194, + 488, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 194, + 488, + 208 + ], + "spans": [ + { + "bbox": [ + 139, + 194, + 488, + 208 + ], + "type": "text", + "content": "3Guangdong Provincial Key Laboratory of Future Networks of Intelligence, CUHKSZ" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 217, + 209, + 407, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 209, + 407, + 222 + ], + "spans": [ + { + "bbox": [ + 217, + 209, + 407, + 222 + ], + "type": "text", + "content": "4Tencent Hunyuan3D 5ByteDance Games" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 249, + 200, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 249, + 200, + 262 + ], + "spans": [ + { + "bbox": [ + 151, + 249, + 200, + 262 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 276, + 296, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 276, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 296, + 574 + ], + "type": "text", + "content": "3D data simulation aims to bridge the gap between simulated and real-captured 3D data, which is a fundamental problem for real-world 3D visual tasks. Most 3D data simulation methods inject predefined physical priors but struggle to capture the full complexity of real data. An optimal approach involves learning an implicit mapping from synthetic to realistic data in a data-driven manner, but progress in this solution has met stagnation in recent studies. This work explores a new solution path of data-driven 3D simulation, called Stable-Sim2Real, based on a novel two-stage depth diffusion model. The initial stage finetunes Stable-Diffusion to generate the residual between the real and synthetic paired depth, producing a stable but coarse depth, where some local regions may deviate from realistic patterns. To enhance this, both the synthetic and initial output depth are fed into a second-stage diffusion, where diffusion loss is adjusted to prioritize these distinct areas identified by a 3D discriminator. We provide a new benchmark scheme to evaluate 3D data simulation methods. Extensive experiments show that training the network with the 3D simulated data derived from our method significantly enhances performance in real-world 3D visual tasks. Moreover, the evaluation demonstrates the high similarity between our 3D simulated data and real-captured patterns. Project page: mutianxu.github.io/stable-sim2real." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 599, + 135, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 599, + 135, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 599, + 135, + 612 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 620, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 296, + 693 + ], + "type": "text", + "content": "In recent years, real-world 3D datasets have played a crucial role in addressing a wide range of tasks in 3D vision and robotics [1, 7, 11, 15, 23, 52, 69, 70]. However, collecting real 3D data is often labor-intensive and time-consuming, and is further complicated by the recent concerns regarding data privacy. In this scenario, synthetic (i.e., simulated)" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 326, + 247, + 544, + 353 + ], + "blocks": [ + { + "bbox": [ + 326, + 247, + 544, + 353 + ], + "lines": [ + { + "bbox": [ + 326, + 247, + 544, + 353 + ], + "spans": [ + { + "bbox": [ + 326, + 247, + 544, + 353 + ], + "type": "image", + "image_path": "c7f7612d866edb84c842c072879108ad28fd72abb03ac19e782c62dd0c19aab0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 361, + 555, + 449 + ], + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 449 + ], + "type": "text", + "content": "Figure 1. We delve back into a substantial yet stagnant problem of simulating real-captured 3D data in a data-driven manner, introducing Stable-Sim2Real. Given the depth maps of the 3D synthetic data, our approach applies a two-stage depth diffusion model based on Stable Diffusion [42] to generate realistic depth maps that are fused to produce real 3D point clouds. The initial stage generates stable but coarse depth maps, followed by a second-stage diffusion to enhance regional details." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 461, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 555, + 510 + ], + "type": "text", + "content": "data [8, 10, 12, 16, 68] emerges as an alternative data resource that is cost-effective, rapid, and scalable. Nevertheless, models trained using synthetic data may not perform robustly against the real-world signal." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": "This problem has catalyzed the development of 3D data simulation techniques, to minimize the gap between simulated and real-captured 3D data. While some studies [2, 30, 33, 39, 75] have explored grounding physical priors to simulate depth sensors, they struggle to capture the full complexity of real-world patterns and are confined to simulation environments, due to the reliance on predefined and explicit physical modeling. A better solution involves learning an implicit mapping from synthetic to realistic data in a data-driven manner, offering greater flexibility and adaptability to real-world diversities. However, only a handful of earlier works [4, 19, 47, 56] have ventured down this path. Hindered by the lack of synthetic-real paired datasets and the dependence on conventional networks like GANs [18], which are prone to training instability and mode collapse [29], these methods have yet to demonstrate the anticipated performance levels. As a result, developing data-driven 3D" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2609" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 96 + ], + "type": "text", + "content": "Sim2Real methods has long been a challenge within the community, and has shown little progress." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 98, + 295, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 98, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 98, + 295, + 326 + ], + "type": "text", + "content": "In this paper, we aim to explore the data-driven 3D Sim2Real and refocus the community's attention on this critical problem. As a prerequisite, we choose a recent synthetic-real paired dataset, LASA [31], containing 10,412 high-quality 3D shape CAD annotations aligned with the real-world object scans. Driven by LASA, our efforts focus on designing an effective data-driven 3D Sim2Real algorithm. Given the inherent indeterminacy and diversity in real-captured data patterns, deep generative models emerge as a natural fit. Among these models, diffusion models [20, 53-55, 62] recently offer easier training dynamics and stronger generation ability. Yet, due to the lack of 3D data, it is hard to train a 3D diffusion model to get strong 3D priors for 3D simulation. Therefore, we opt to leverage the strong generalization priors of " + }, + { + "bbox": [ + 55, + 98, + 295, + 326 + ], + "type": "inline_equation", + "content": "2D" + }, + { + "bbox": [ + 55, + 98, + 295, + 326 + ], + "type": "text", + "content": " diffusion foundation models (e.g., SD - Stable Diffusion [42]) to simulate real 2D depth maps, which are then fused to gain 3D data. This strategy is similar to real 3D data acquisition, where 2D depth is captured and fused into 3D." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 327, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 295, + 460 + ], + "type": "text", + "content": "To this end, an intuitive baseline is to sample CAD (i.e., synthetic) depth maps and their paired real depth images from LASA, and then finetune SD to learn an implicit mapping between them. Although such a solution powered by large-scale paired data and the foundation diffusion model has shown effectiveness in conventional image translations, it faces special challenges in our task: while conventional image translations (e.g., denoising/restoration) normally remove noise to produce clean images, our task aims to output noisy depth that is highly uncertain, thereby making the distribution that needs to be learned more complex." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 460, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 640 + ], + "type": "text", + "content": "To address these challenges, we propose Stable-Sim2Real. The core is a novel two-stage depth diffusion model to simulate real-captured noise (a sample is shown in Fig. 1). At Stage-I diffusion, instead of directly generating the corresponding real-world depth map, our model generates the residual (i.e., difference) between real and CAD depth maps. The simulated depth map is then yielded by adding the generated residual with the CAD depth. Empirically, this residual represents the noise added to clean (i.e., synthetic) depth during real-world scanning. Moreover, compared to directly generating noisy real depth, adding noise to CAD depth, which is inherently clean and view-consistent, can produce more stable depth with smaller view variation and better preservation of the original geometry (see Sec. 3.2 for more discussions and analysis)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "However, according to our observations and experiments, while some regions within the Stage-I generated depth map successfully fit the realistic pattern, some unsatisfactory areas still exhibit conspicuous 3D geometric discrepancies between the generation and real captures (shown in Fig. 6). To tackle this issue, our model then revolves" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "type": "text", + "content": "around discriminating unsatisfactory areas and enhancing the Stage-I generation at local regions. We train a 3D-Aware Discriminator to distinguish between Stage-I generation and real captures at local geometries, which is expected to delineate the distribution boundary between them. If a generated patch can be successfully discriminated, it signifies an unsatisfactory area. Next, we introduce a Stage-II diffusion, conditioned not only on the CAD depth but also on Stage-I generation. More importantly, the diffusion loss is adjusted to prioritize on unsatisfactory regions. In this way, Stage-II diffusion is specialized on learning the distribution of unsatisfactory areas, thereby enhancing Stage-I generation. Finally, the generated depth maps are fused to derive the simulated 3D data. Notably, the binary classifier is only used during the training phase." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 253, + 553, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 553, + 408 + ], + "type": "text", + "content": "Last, we provide a comprehensive benchmark scheme for 3D data simulation. The key logic is: if a model's real-world performance can be improved after being trained with simulated data, it would validate the effectiveness of simulation methods. We focus on fundamental real-world 3D tasks: 3D shape reconstruction and 3D object/scene understanding. As for shape reconstruction, we pretrain a 3D reconstruction network that takes simulated 3D data as input and outputs clean 3D surfaces. For 3D object/scene understanding, the simulated 3D data is used to pretrain a self-supervised point cloud learning framework. To better assess the performance gain purely from simulated data itself, we conduct few-shot evaluation on pre-trained networks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 324, + 409, + 479, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 409, + 479, + 419 + ], + "spans": [ + { + "bbox": [ + 324, + 409, + 479, + 419 + ], + "type": "text", + "content": "Our contributions are summarized as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 422, + 553, + 577 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 313, + 422, + 553, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 553, + 469 + ], + "type": "text", + "content": "- We delve back into a substantial yet stagnant problem, data-driven 3D real data simulation, paving a new solution path of training generative foundation models on extensive synthetic-real paired data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 470, + 553, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 553, + 505 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 553, + 505 + ], + "type": "text", + "content": "- We present Stable-Sim2Real, a novel 3D real data simulation method based on a two-stage depth diffusion model that can stably simulate real-captured noise." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 506, + 553, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 553, + 540 + ], + "type": "text", + "content": "- We provide a comprehensive scheme to benchmark 3D data simulation, based on few-shot evaluation on networks trained with simulated data for real-world 3D tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 541, + 553, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 541, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 541, + 553, + 577 + ], + "type": "text", + "content": "- Extensive qualitative and quantitative results show that the data simulated by our method significantly boosts model performance on real-world 3D tasks." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 590, + 400, + 603 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 400, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 400, + 603 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 611, + 482, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 482, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 482, + 623 + ], + "type": "text", + "content": "2.1. Simulation of Real-World Data" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "content": "Simulation of 2D real-world data. In the realm of 2D real-world data simulation, two predominant approaches have emerged: domain randomization and domain adaptation. The first line of works [57, 58, 60] randomly perturbs textures, lighting, and material of synthetic data, thereby modeling the synthetic-real domain gap as one kind of variation to endow models with robustness to real-world data." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2610" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": "However, it brought the challenge of achieving a delicate balance between ensuring that the distribution of augmentations adequately covers that of real-world data and preventing the augmentation from being too extensive to hinder effective model training [24, 37, 72]. Domain adaptation, on the other hand, aims to minimize the domain gap by transferring knowledge from synthetic data to real-world scenarios, leveraging techniques such as adversarial training and self-supervised learning [5, 49, 71]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 194, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 296, + 422 + ], + "type": "text", + "content": "Simulation of 3D real-captured data. Acquiring real-world 3D data and their annotation is labor intensive and inefficient, deriving methods for the simulation of 3D data. Existing simulation methods primarily focus on enhancing the realism of depth information by grounding physical priors in simulators to close the domain gap with 3D data from the physical world [2, 30, 33, 39, 75]. However, these approaches are usually limited by their reliance on predefined and explicit physical modeling. Alternatively, data-driven methods turn to implicitly learn from the distribution in real-captured data, independent of physical modeling [4, 19, 47, 56], offering enhanced adaptability and generalization capabilities to various real-world scenarios. Nonetheless, due to the absence of extensive synthetic-real paired datasets and the reliance on traditional GANs [18], which suffer from training instability and mode collapse [29], these approaches have not achieved the expected performance. Consequently, the research progress of data-driven 3D Sim2Real techniques faces a recent stagnation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 430, + 243, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 243, + 442 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 243, + 442 + ], + "type": "text", + "content": "2.2. Image Generation and Translation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 447, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 295, + 604 + ], + "type": "text", + "content": "Conventional generative models. We formulate the 3D Sim2Real task as learning the mapping from synthetic to real-world captured noise, based on depth images. Existing approaches in the domain of image generation and translation include a variety of generative models such as Generative Adversarial Networks (GANs) [18], Variational Autoencoders (VAEs) [28], Cycle-Consistent Adversarial Networks (Cycle-GAN) [77], and Pix2Pix [22]. While GAN-based methods, which have been predominant in image generation before diffusion models, achieved significant milestones in Sim2Real tasks [6, 65, 71], they often struggle with maintaining local details and semantic consistency, leading to sub-optimal performance [13, 29, 32, 36, 38]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": "Diffusion models. Diffusion models, also known as denoising diffusion probabilistic models (DDPMs) [50], have emerged as a powerful alternative to GANs for image synthesis [13, 32, 34, 35, 43, 44, 74]. Unlike GANs, diffusion models offer a more stable training procedure and have been shown to generate images with higher fidelity and better preservation of fine details [21, 42]. Depth diffusion models recently track great attention. They mainly focus on the task" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "of depth estimation in a monocular [14, 26, 45, 46, 59, 64] or multi-view [27, 67] manner. Other works also cover conventional image translation tasks, such as depth map superresolution [48] and refinement [76], which produce clean and certain output images. Differently, our work leverages generative models to \"generate noise\", where the expected output is highly uncertain. Thus, the distribution that needs to be learned is more complex, bringing fundamentally different challenges." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 190, + 370, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 190, + 370, + 202 + ], + "spans": [ + { + "bbox": [ + 314, + 190, + 370, + 202 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 209, + 555, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 209, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 555, + 365 + ], + "type": "text", + "content": "We hope to exploit the strong prior of the diffusion-based foundation model, so we take depth map as our bridge towards the simulation of real-captured 3D data. The preliminaries on the diffusion model are provided in Sec. 3.1. Our Stable-Sim2Real is a two-stage depth diffusion framework, as illustrated in Fig. 2. The first-stage model (Sec. 3.2) learns to generate the depth residual, producing stable but coarse depth maps, where some areas are distinct from the realistic pattern. Therefore, a 3D binary classifier (Sec. 3.3) is introduced to discriminate these distinct areas. The second-stage model (Sec. 3.4) is specialized to enhance the generation of these areas. Finally, the enhanced depth maps are fused to obtain the 3D simulated data (Sec. 3.5)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 372, + 497, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 497, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 497, + 384 + ], + "type": "text", + "content": "3.1. Preliminaries on Diffusion Models" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 389, + 554, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 554, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 554, + 450 + ], + "type": "text", + "content": "Our model is based on Denoising Diffusion Probabilistic Models (DDPM) [20, 51]. The key of DDPM is to model a data distribution " + }, + { + "bbox": [ + 313, + 389, + 554, + 450 + ], + "type": "inline_equation", + "content": "p(x)" + }, + { + "bbox": [ + 313, + 389, + 554, + 450 + ], + "type": "text", + "content": " via building a Markov chain in the data space. It can be divided into two main processes, forward process and backward process." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 464, + 554, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 554, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 554, + 534 + ], + "type": "text", + "content": "Forward diffusion process. Given an initial data " + }, + { + "bbox": [ + 313, + 464, + 554, + 534 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 464, + 554, + 534 + ], + "type": "text", + "content": ", we gradually add random noise " + }, + { + "bbox": [ + 313, + 464, + 554, + 534 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 313, + 464, + 554, + 534 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 464, + 554, + 534 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 464, + 554, + 534 + ], + "type": "text", + "content": ", yielding a series of noisy data, which can be simply formulated as " + }, + { + "bbox": [ + 313, + 464, + 554, + 534 + ], + "type": "inline_equation", + "content": "x_{t+1} = x_t + \\epsilon_t" + }, + { + "bbox": [ + 313, + 464, + 554, + 534 + ], + "type": "text", + "content": ". After adding noise for multiple steps, the initial distribution of " + }, + { + "bbox": [ + 313, + 464, + 554, + 534 + ], + "type": "inline_equation", + "content": "p(x)" + }, + { + "bbox": [ + 313, + 464, + 554, + 534 + ], + "type": "text", + "content": " is expected to be transformed into a standard Gaussian distribution." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 535, + 554, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 535, + 554, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 535, + 554, + 559 + ], + "type": "text", + "content": "According to DDPM [20], this process can be directly achieved by the following forward diffusion process:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 566, + 553, + 591 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 566, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 315, + 566, + 553, + 591 + ], + "type": "interline_equation", + "content": "q \\left(x _ {t}\\right) = \\sqrt {\\alpha_ {t}} x _ {0} + \\sqrt {1 - \\alpha_ {t}} \\boldsymbol {\\epsilon}, x _ {0} \\sim p (x), t \\in \\{0, 1, \\dots , T \\}, \\tag {1}", + "image_path": "1f3481030dcf733a3f4a48c52628671fda27edd76d33eafcf4edf338dec3e52e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 592, + 554, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 554, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 554, + 639 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 592, + 554, + 639 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 592, + 554, + 639 + ], + "type": "text", + "content": " denotes the number of the time step, " + }, + { + "bbox": [ + 313, + 592, + 554, + 639 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 592, + 554, + 639 + ], + "type": "text", + "content": " is the current time step, and " + }, + { + "bbox": [ + 313, + 592, + 554, + 639 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 313, + 592, + 554, + 639 + ], + "type": "text", + "content": " regulates the noise schedule deciding the pace at which the initial data distribution transforms into a standard Gaussian distribution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "content": "Backward diffusion process. Given a noisy " + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "content": " from a standard Gaussian distribution as the initial state, " + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "content": " is denoised to a cleaner " + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "inline_equation", + "content": "x_{t-1}" + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "content": " at each time step, which can also be simply formulated as " + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "inline_equation", + "content": "x_{t-1} = x_{t} - \\mu_{\\theta}^{\\epsilon}(x_{t}, t) + \\epsilon_{t}" + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "content": " in which " + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon_{t} \\sim \\mathcal{N}(0, \\sigma_{\\mathbf{t}} \\mathbf{I})" + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}^{\\epsilon}" + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "content": " is a learnable network to predict" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2611" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 70, + 531, + 227 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 531, + 227 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 531, + 227 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 531, + 227 + ], + "type": "image", + "image_path": "1d460675239563ae4e71c0ad18d35994344d4f1bde9eeaf6a4f9ba27d8b9c2a2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 234, + 555, + 304 + ], + "lines": [ + { + "bbox": [ + 54, + 234, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 54, + 234, + 555, + 304 + ], + "type": "text", + "content": "Figure 2. An overview of our Stable-Sim2Real. Conditioned on CAD (i.e., synthetic) depth maps, our model first generates the residual (i.e., difference) between real-captured and CAD depth maps. By adding the generated residual to the CAD depth, Stage-I diffusion produces stable but coarse depth maps (Sec. 3.2). Next, the Stage-I output is projected into 3D and sent to a 3D discriminator, to identify the unsatisfactory local areas that are distinct from real-captured patterns (Sec. 3.3). Later, our Stage-II diffusion is conditioned on both CAD and Stage-I output, where the training loss of Stage-II is re-weighted to be specialized on enhancing unsatisfactory areas of Stage-I output (Sec. 3.4). Notably, the local discriminating and loss re-weighting are only used during training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 313, + 295, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 313, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 295, + 350 + ], + "type": "text", + "content": "the injected noise. By repeating this reverse process until the maximum number of steps " + }, + { + "bbox": [ + 55, + 313, + 295, + 350 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 313, + 295, + 350 + ], + "type": "text", + "content": ", we can obtain the final generated data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 350, + 295, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 295, + 373 + ], + "type": "text", + "content": "In DDPM [20], the backward diffusion process is proven to be:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 378, + 294, + 417 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 378, + 294, + 417 + ], + "spans": [ + { + "bbox": [ + 61, + 378, + 294, + 417 + ], + "type": "interline_equation", + "content": "x _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} x _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {\\alpha_ {t} \\left(1 - \\Pi_ {\\tau = 0} ^ {t} \\alpha_ {\\tau}\\right)}} \\mu_ {\\theta} ^ {\\epsilon} \\left(x _ {t}, t;\\right) + \\sigma_ {t} \\epsilon , \\tag {2}", + "image_path": "5673d1cd90d93de3bc68282363e9c6d16eb659fe389e4b54ea20959b12dff806.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 426, + 295, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 463 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 463 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 426, + 295, + 463 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 55, + 426, + 295, + 463 + ], + "type": "text", + "content": " denotes the standard deviation of " + }, + { + "bbox": [ + 55, + 426, + 295, + 463 + ], + "type": "inline_equation", + "content": "q(x_{t - 1}|x_t,x_0)" + }, + { + "bbox": [ + 55, + 426, + 295, + 463 + ], + "type": "text", + "content": " in the backward diffusion to sample random noise. The loss for training " + }, + { + "bbox": [ + 55, + 426, + 295, + 463 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}^{\\epsilon}" + }, + { + "bbox": [ + 55, + 426, + 295, + 463 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 470, + 295, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 470, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 116, + 470, + 295, + 485 + ], + "type": "interline_equation", + "content": "L _ {\\theta} = \\mathbb {E} _ {\\epsilon , t} \\| \\epsilon - \\mu_ {\\theta} ^ {\\epsilon} (x _ {t}, t) \\| ^ {2} \\tag {3}", + "image_path": "1540512f13871b2ed4f68c8d1198e96d8a84f2ba07dc58146bb6bdb0b9ac11ee.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 493, + 234, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 234, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 234, + 506 + ], + "type": "text", + "content": "3.2. Stage-I: Initial Depth Generation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 510, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 605 + ], + "type": "text", + "content": "Residual depth generation with SD. Different from unconditional diffusion models, text-to-image diffusion models focus on generating images with optional text prompts. A representative is Stable Diffusion (SD) [42]. SD trains a denoising U-Net architecture " + }, + { + "bbox": [ + 55, + 510, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}(x_t,t,c)" + }, + { + "bbox": [ + 55, + 510, + 295, + 605 + ], + "type": "text", + "content": " on the latent space of a pre-trained Variational Auto-Encoder (VAE). Here, " + }, + { + "bbox": [ + 55, + 510, + 295, + 605 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 510, + 295, + 605 + ], + "type": "text", + "content": " represents the supplementary text prompt embedding, usually acquired through methods like CLIP [41]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "We finetune SD model on synthetic-real paired depth data from LASA [31]. Specifically, given an initial synthetic depth map " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " and its paired real-captured depth map " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " from a pre-defined dataset, we first get the initial ground truth residual (i.e., difference) depth map " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "D_{res}^{gt} = S - R" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": ". Then we follow ControlNet [73] to add an additional encoder " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "En" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " and decoder " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}^{\\epsilon}" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " for finetuning SD. Specifically, we run the forward process to add noise on " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "D_{res}^{gt}" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " following Eq. (1) and encode with " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "En" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": ". Next, we send " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "S" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 313, + 555, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 313, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 555, + 374 + ], + "type": "text", + "content": "to a pretrained VAE encoder " + }, + { + "bbox": [ + 313, + 313, + 555, + 374 + ], + "type": "inline_equation", + "content": "En" + }, + { + "bbox": [ + 313, + 313, + 555, + 374 + ], + "type": "text", + "content": ", generating the latent condition signal. This latent signal " + }, + { + "bbox": [ + 313, + 313, + 555, + 374 + ], + "type": "inline_equation", + "content": "En(S)" + }, + { + "bbox": [ + 313, + 313, + 555, + 374 + ], + "type": "text", + "content": " is further processed by an additional encoder " + }, + { + "bbox": [ + 313, + 313, + 555, + 374 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{\\phi}" + }, + { + "bbox": [ + 313, + 313, + 555, + 374 + ], + "type": "text", + "content": " and sent to the decoder blocks of denoising U-Net. We train both the decoder blocks " + }, + { + "bbox": [ + 313, + 313, + 555, + 374 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}^{\\epsilon}" + }, + { + "bbox": [ + 313, + 313, + 555, + 374 + ], + "type": "text", + "content": " and encoder " + }, + { + "bbox": [ + 313, + 313, + 555, + 374 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{\\phi}" + }, + { + "bbox": [ + 313, + 313, + 555, + 374 + ], + "type": "text", + "content": " using the following loss (based on Eq. (3)):" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 340, + 383, + 555, + 398 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 383, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 340, + 383, + 555, + 398 + ], + "type": "interline_equation", + "content": "L _ {\\theta , \\phi} = \\mathbb {E} _ {\\epsilon , S, t} \\| \\epsilon - \\boldsymbol {\\mu} _ {\\theta} ^ {\\epsilon} (\\boldsymbol {x} _ {t}, t, \\boldsymbol {f} _ {\\phi} (E n (\\boldsymbol {S}))) \\| ^ {2} \\tag {4}", + "image_path": "b0051426104fac86b89c0c7b17571c59619d79c5d4b2b991b195e5ff5de9e984.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 411, + 555, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 520 + ], + "type": "text", + "content": "Learning depth residual helps to produce stable depth. Empirically, this residual represents the noise added to clean (i.e., synthetic) depth during real-world scanning. Moreover, compared to directly generating noisy real depth, adding noise to CAD depth, which is inherently clean and view-consistent, can produce more stable depth with smaller view variation and better preservation of the original geometry. From experimental results, learning residual indeed performs better (see Tab. 5)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 533, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 629 + ], + "type": "text", + "content": "Learning depth residual is theoretically identical to learning real depth. First, assume that real depth is a sample from a Gaussian distribution, denoted by " + }, + { + "bbox": [ + 313, + 533, + 555, + 629 + ], + "type": "inline_equation", + "content": "R \\sim \\mathcal{N}(\\alpha_{\\mathbf{R}}, \\vartheta)" + }, + { + "bbox": [ + 313, + 533, + 555, + 629 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 533, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 533, + 555, + 629 + ], + "type": "text", + "content": " denotes the mean and " + }, + { + "bbox": [ + 313, + 533, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\vartheta" + }, + { + "bbox": [ + 313, + 533, + 555, + 629 + ], + "type": "text", + "content": " represents the variance. Notably, " + }, + { + "bbox": [ + 313, + 533, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\vartheta" + }, + { + "bbox": [ + 313, + 533, + 555, + 629 + ], + "type": "text", + "content": " can naturally represent real-captured noise. Therefore, synthetic depth can be denoted by " + }, + { + "bbox": [ + 313, + 533, + 555, + 629 + ], + "type": "inline_equation", + "content": "S \\sim \\mathcal{N}(\\alpha_{\\mathbf{S}}, 0)" + }, + { + "bbox": [ + 313, + 533, + 555, + 629 + ], + "type": "text", + "content": ", where the variance is zero as there is no real-captured noise." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "In our design, as the diffusion model is inherently designed to learn the transformation from a random noise " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": " to any other Gaussian distribution [53], we leverage it to model the distribution of the depth residual that can be denoted by " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\alpha_{\\mathbf{R}} - \\alpha_{\\mathbf{S}},\\vartheta)" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": ". Last, by adding the depth residual to synthetic depth, the output distribution can be formulated as: " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\alpha_{\\mathbf{R}} - \\alpha_{\\mathbf{S}},\\vartheta) + \\mathcal{N}(\\alpha_{\\mathbf{S}},0) = \\mathcal{N}(\\alpha_{\\mathbf{R}},\\vartheta)" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2612" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "which is exactly identical to the previously assumed real-captured noise distribution." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 102, + 228, + 114 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 102, + 228, + 114 + ], + "spans": [ + { + "bbox": [ + 55, + 102, + 228, + 114 + ], + "type": "text", + "content": "3.3. 3D-Aware Local Discriminating" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 119, + 296, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 119, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 119, + 296, + 262 + ], + "type": "text", + "content": "Unsatisfactory areas distinct from real patterns. Based on our observations, the noise captured in real data is spatially-variant, meaning different areas within a real depth map may display varying levels of perturbation. However, our Stage-I diffusion tends to introduce spatially-uniform noise across the entire input CAD depth. This discrepancy may stem from a bias within SD, as also discussed in [9]. As a result, although certain regions of the generated depth map align well with real perturbation patterns, there remain noticeable areas that deviate from reality. This leads to evident 3D geometric discrepancies between Stage-I outputs and real captures (as illustrated in Fig. 6)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 274, + 296, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 296, + 285 + ], + "type": "text", + "content": "Identifying unsatisfactory areas with 3D discriminator." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 286, + 295, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 286, + 295, + 466 + ], + "spans": [ + { + "bbox": [ + 54, + 286, + 295, + 466 + ], + "type": "text", + "content": "Our design logic to discriminate unsatisfactory areas is that if the generated areas diverge significantly from real patterns, there should exist a distribution boundary of their 3D geometries. To this end, we propose employing a 3D binary classifier to discriminate this boundary. Initially, we split all generated and real-captured depth maps into different patches representing different areas, which are then projected to form 3D point patches. Subsequently, we leverage PointNet [40] as a 3D binary classifier to differentiate whether the input point patches come from model generation or real captures. The CAD depth maps are also concatenated in the input as additional guidance. The successful classification of a generated patch indicates that its corresponding area is distinct from real captures. Notably, the binary classifier is solely used during training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 471, + 285, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 285, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 285, + 483 + ], + "type": "text", + "content": "3.4. Stage-II: Unsatisfactory Area Enhancement" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 488, + 296, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 488, + 296, + 596 + ], + "spans": [ + { + "bbox": [ + 55, + 488, + 296, + 596 + ], + "type": "text", + "content": "Improving unsatisfactory areas by re-weighting denoising loss. After identifying these local patches, we introduce a second-stage depth refinement model that differs from the Stage-I diffusion in two crucial aspects: 1) The diffusion model is now conditioned on the concatenation of synthetic depth " + }, + { + "bbox": [ + 55, + 488, + 296, + 596 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 488, + 296, + 596 + ], + "type": "text", + "content": " and the previously generated depth residual " + }, + { + "bbox": [ + 55, + 488, + 296, + 596 + ], + "type": "inline_equation", + "content": "D^{I}" + }, + { + "bbox": [ + 55, + 488, + 296, + 596 + ], + "type": "text", + "content": " from Stage I. Here, " + }, + { + "bbox": [ + 55, + 488, + 296, + 596 + ], + "type": "inline_equation", + "content": "D^{I} = D_{res}^{I} + S" + }, + { + "bbox": [ + 55, + 488, + 296, + 596 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 488, + 296, + 596 + ], + "type": "inline_equation", + "content": "D_{res}^{I}" + }, + { + "bbox": [ + 55, + 488, + 296, + 596 + ], + "type": "text", + "content": " is the output residual depth from Stage-I model. Thus, the training loss defined in Eq. (4) is updated to:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 601, + 294, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 601, + 294, + 628 + ], + "spans": [ + { + "bbox": [ + 58, + 601, + 294, + 628 + ], + "type": "interline_equation", + "content": "L _ {\\chi , \\psi} = \\mathbb {E} _ {\\boldsymbol {\\epsilon}, \\boldsymbol {D} ^ {I}, \\boldsymbol {S}, t} \\left\\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\mu} _ {\\chi} ^ {\\epsilon} \\left(\\boldsymbol {x} _ {t}, t, \\boldsymbol {f} _ {\\psi} \\left(E n \\left(\\boldsymbol {D} _ {r e s} ^ {I}, \\boldsymbol {S}\\right)\\right)\\right) \\right\\| ^ {2}, \\tag {5}", + "image_path": "8934f121e8476c6152677127a14af423ac7f5b0c4589320c5525c29a01eb9943.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 628, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 295, + 664 + ], + "type": "text", + "content": "2) More importantly, the diffusion loss is re-weighted to prioritize the enhancement of the identified unsatisfactory regions, making the final Stage-II loss " + }, + { + "bbox": [ + 55, + 628, + 295, + 664 + ], + "type": "inline_equation", + "content": "L_{\\chi ,\\psi}" + }, + { + "bbox": [ + 55, + 628, + 295, + 664 + ], + "type": "text", + "content": " to be:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 123, + 667, + 295, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 667, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 123, + 667, + 295, + 684 + ], + "type": "interline_equation", + "content": "L _ {\\chi , \\psi} = \\omega L _ {\\chi , \\psi} ^ {s i m} + \\lambda L _ {\\chi , \\psi} ^ {d i s}, \\tag {6}", + "image_path": "2c7d558768137fe08064671dbe1b9fdfe6d0f5038dd5bbe74f763703849367e7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 687, + 296, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 687, + 296, + 716 + ], + "spans": [ + { + "bbox": [ + 55, + 687, + 296, + 716 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 687, + 296, + 716 + ], + "type": "inline_equation", + "content": "L_{\\chi, \\psi}^{sim}" + }, + { + "bbox": [ + 55, + 687, + 296, + 716 + ], + "type": "text", + "content": " indicates the loss on similar regions, while " + }, + { + "bbox": [ + 55, + 687, + 296, + 716 + ], + "type": "inline_equation", + "content": "L_{\\chi, \\psi}^{dis}" + }, + { + "bbox": [ + 55, + 687, + 296, + 716 + ], + "type": "text", + "content": " denotes the loss on previously identified distinct (i.e.," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "content": "unsatisfactory) areas. " + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "content": " are weight coefficients, in which " + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "inline_equation", + "content": "\\lambda >\\omega" + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "content": ". In our experiment, setting " + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "inline_equation", + "content": "\\omega = 0.5" + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "inline_equation", + "content": "\\lambda = 1.5" + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "content": " produces the best result." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 108, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 553, + 168 + ], + "type": "text", + "content": "Our Stage-II diffusion model is specialized on learning the distribution of unsatisfactory areas in Stage-I generation. Meanwhile, by conditioning on Stage-I output and assigning less weight to satisfied areas, the generation on satisfied outputs is also preserved in Stage-II, as in Fig. 6." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 174, + 477, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 174, + 477, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 477, + 186 + ], + "type": "text", + "content": "3.5. Creating Real-World 3D Data" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 191, + 553, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 191, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 191, + 553, + 275 + ], + "type": "text", + "content": "Two-stage real depth simulation. In the initial stage, we sample a random distribution " + }, + { + "bbox": [ + 313, + 191, + 553, + 275 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 191, + 553, + 275 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 191, + 553, + 275 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 313, + 191, + 553, + 275 + ], + "type": "text", + "content": " and feed it into the denoising U-Net " + }, + { + "bbox": [ + 313, + 191, + 553, + 275 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}" + }, + { + "bbox": [ + 313, + 191, + 553, + 275 + ], + "type": "text", + "content": " of Stage-I diffusion model. The model runs a backward diffusion process for " + }, + { + "bbox": [ + 313, + 191, + 553, + 275 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 191, + 553, + 275 + ], + "type": "text", + "content": " steps to generate the initial depth residual " + }, + { + "bbox": [ + 313, + 191, + 553, + 275 + ], + "type": "inline_equation", + "content": "D_{res}^{I}" + }, + { + "bbox": [ + 313, + 191, + 553, + 275 + ], + "type": "text", + "content": ", following Eq. (2) with the synthetic depth " + }, + { + "bbox": [ + 313, + 191, + 553, + 275 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 191, + 553, + 275 + ], + "type": "text", + "content": " serving as the condition. Now we can obtain the initial depth map " + }, + { + "bbox": [ + 313, + 191, + 553, + 275 + ], + "type": "inline_equation", + "content": "D^{I} = D_{res}^{I} + S" + }, + { + "bbox": [ + 313, + 191, + 553, + 275 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 276, + 553, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 553, + 347 + ], + "type": "text", + "content": "Moving to the second stage, we repeat a similar backward diffusion process to generate the residual depth " + }, + { + "bbox": [ + 313, + 276, + 553, + 347 + ], + "type": "inline_equation", + "content": "D_{res}^{II}" + }, + { + "bbox": [ + 313, + 276, + 553, + 347 + ], + "type": "text", + "content": ", using the denoising U-Net " + }, + { + "bbox": [ + 313, + 276, + 553, + 347 + ], + "type": "inline_equation", + "content": "\\mu_{\\chi}" + }, + { + "bbox": [ + 313, + 276, + 553, + 347 + ], + "type": "text", + "content": " of our Stage-II diffusion model, conditioned on both synthetic depth " + }, + { + "bbox": [ + 313, + 276, + 553, + 347 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 276, + 553, + 347 + ], + "type": "text", + "content": " and Stage-I output " + }, + { + "bbox": [ + 313, + 276, + 553, + 347 + ], + "type": "inline_equation", + "content": "D^{I}" + }, + { + "bbox": [ + 313, + 276, + 553, + 347 + ], + "type": "text", + "content": ". The final depth map " + }, + { + "bbox": [ + 313, + 276, + 553, + 347 + ], + "type": "inline_equation", + "content": "D^{II} = D_{res}^{II} + S" + }, + { + "bbox": [ + 313, + 276, + 553, + 347 + ], + "type": "text", + "content": ". DDIM sampler [51] is utilized during the whole inference phase." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 361, + 554, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 554, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 554, + 445 + ], + "type": "text", + "content": "Depth fusion to 3D data. Given the 3D synthetic object, we follow LASA [31] to embed CAD objects into 3D scenes. This allows us to render synthetic depth images, simulating the occlusions in real-world captures. Then we conduct the aforementioned two-stage real depth simulation on rendered synthetic depth maps, followed by a depth fusion algorithm to create simulated 3D data." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 454, + 395, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 395, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 395, + 468 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 475, + 554, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 554, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 554, + 594 + ], + "type": "text", + "content": "We provide a new benchmark scheme for 3D data simulation. To begin with, different simulation methods are performed on synthetic datasets. The 3D simulated data are then utilized to train networks for tackling fundamental real-world 3D visual tasks: 3D instance reconstruction, 3D object/scene understanding. The key logic of our evaluation is: if the model's real-world performance can be improved after being trained with simulated data, it would validate that the simulated data is close to real distributions, demonstrating the effectiveness of simulation methods." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 600, + 400, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 400, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 400, + 611 + ], + "type": "text", + "content": "4.1. Preliminaries" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": "Training of Stable-Sim2Real. We train our model by finetuning the official StableDiffusion V2.1 [42]. The training dataset is a recent 3D synthetic-real paired dataset, LASA [31]. LASA is a large-scale dataset that contains 10,412 unique CAD models across 17 categories. Skilled artists are employed to meticulously craft aligned CAD models from 3D object real-world scans of [3]. LASA provides CAD-real depth map pairs, rendered based on camera" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2613" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 70, + 531, + 149 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 531, + 149 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 531, + 149 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 531, + 149 + ], + "type": "table", + "html": "
MethodChairSofaTableCabinetBedShelf
w/o sim. data24.5 / 12.3 / 20.155.6 / 7.02 / 19.825.8 / 19.3 / 20.357.1 / 8.83 / 24.140.1 / 6.14 / 17.312.8 / 6.74 / 23.7
Cycle-GAN [77]25.1 / 12.1 / 20.955.2 / 6.98 / 20.526.1 / 19.1 / 19.556.5 / 8.24 / 25.140.6 / 5.98 / 16.913.1 / 6.08 / 24.9
Pix2Pix [22]24.9 / 11.0 / 20.556.4 / 6.74 / 21.326.6 / 18.5 / 22.157.4 / 7.83 / 25.541.3 / 6.01 / 17.814.2 / 6.15 / 26.2
Giga-GAN [25]26.2 / 10.3 / 22.158.9 / 5.87 / 22.429.3 / 16.4 / 24.360.5 / 7.01 / 26.944.2 / 5.25 / 19.318.2 / 5.87 / 27.8
Stable-Diffusion [42]27.3 / 9.51 / 25.559.4 / 5.93 / 23.629.7 / 15.8 / 24.761.6 / 6.92 / 27.545.1 / 5.31 / 20.119.1 / 5.85 / 28.3
Ours29.2 / 7.83 / 27.562.9 / 4.09 / 25.131.2 / 14.1 / 26.863.2 / 5.58 / 28.646.3 / 4.23 / 21.818.8 / 5.02 / 31.6
", + "image_path": "e8b2ccee144e33d3fdae5008b96e4e0aa286e3bbb045bc59c378bb27bdf5e207.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 93, + 213, + 520, + 289 + ], + "blocks": [ + { + "bbox": [ + 55, + 158, + 555, + 202 + ], + "lines": [ + { + "bbox": [ + 55, + 158, + 555, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 555, + 202 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of few-shot evaluation on 3D instance reconstruction, where the model is pretrained using simulated data generated by different methods. \"w/o sim. data\" indicates pretraining using no simulated data but only synthetic data. Evaluation metrics are mIoU / Chamfer L2 / F-score respectively. Higher mIoU and F-score are better while lower Chamfer L2 is better. Chamfer L2 is scaled by 1,000. All results are reproduced using the official code of DisCo [31]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 93, + 213, + 520, + 289 + ], + "lines": [ + { + "bbox": [ + 93, + 213, + 520, + 289 + ], + "spans": [ + { + "bbox": [ + 93, + 213, + 520, + 289 + ], + "type": "table", + "html": "
MethodChairSofaTableCabinetBedShelf
w/o sim. data19.4 / 14.7 / 15.846.2 / 9.92 / 15.720.5 / 23.6 / 15.951.2 / 11.3 / 20.233.6 / 13.1 / 10.911.1 / 9.23 / 21.6
Cycle-GAN [77]20.3 / 13.1 / 16.548.9 / 9.03 / 18.923.2 / 22.4 / 17.551.9 / 10.8 / 21.435.2 / 12.5 / 12.111.7 / 8.87 / 23.5
Pix2Pix [22]19.3 / 12.3 / 17.449.7 / 9.51 / 18.822.6 / 21.5 / 18.852.7 / 11.2 / 21.137.7 / 10.0 / 13.812.5 / 8.91 / 24.3
Giga-GAN [25]22.4 / 11.8 / 19.553.0 / 7.69 / 19.423.1 / 19.3 / 21.654.3 / 10.1 / 21.940.3 / 7.95 / 15.414.2 / 7.81 / 26.9
Stable-Diffusion [42]23.3 / 11.7 / 20.454.1 / 7.98 / 19.224.2 / 17.9 / 22.655.1 / 9.07 / 22.440.4 / 7.89 / 16.215.8 / 6.72 / 28.3
Ours27.1 / 9.31 / 24.860.6 / 5.29 / 22.328.5 / 16.1 / 24.760.6 / 7.51 / 25.943.2 / 6.23 / 18.017.3 / 5.93 / 30.5
", + "image_path": "1862514e1cf0661cdfd021754bc5447ce38b48fa1fe74f0f4a693dea0ddbb046.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 297, + 555, + 319 + ], + "lines": [ + { + "bbox": [ + 55, + 297, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 555, + 319 + ], + "type": "text", + "content": "Table 2. Quantitative comparison of zero-shot evaluation on 3D instance reconstruction, when the model is solely pretrained using simulated data generated by different methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 331, + 295, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 295, + 357 + ], + "type": "text", + "content": "information during the real-world collection. See the supplementary material for more implementation details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 374, + 295, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 295, + 494 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 295, + 494 + ], + "type": "text", + "content": "Methods in comparison. Since our method operates on depth images, we select several image-to-image generation methods for comparison: 1) Cycle-Consistent Adversarial Networks (Cycle-GAN) [77], which is a representative generative network when paired training data are not available, so we train Cycle-GAN on the pair-shuffled LASA training data. 2) Pix2Pix [22], which is a classical image generation method. 3) Giga-GAN [25], a state-of-the-art GAN-based method. 4) Stable Diffusion [42], our main baseline model. Both 2), 3), 4) are trained using the paired data from LASA." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 503, + 208, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 208, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 208, + 514 + ], + "type": "text", + "content": "4.2.3D Instance Reconstruction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 521, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 700 + ], + "type": "text", + "content": "Setup. 3D instance reconstruction aims to recover the full geometry from real-captured partial point clouds or multiview images. We utilize a recent state-of-the-art diffusion-based reconstruction network, DisCo, which is proposed in LASA [31]. It is originally pretrained on three synthetic datasets (i.e., ShapeNet [8], ABO [10], and 3D-Future [17]) and finetuned on LASA real training data. During pretraining, DisCo simply constructs partial point clouds by adding random perturbation to synthetic data. Instead, in our experiment, to simulate partial point clouds for pretraining, we perform simulation using our method and other baseline models based on the original three synthetic datasets used to pretrain DisCo, and replace half of the original partial point clouds with the simulated 3D data. The other configurations all follow the original DisCo." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "Subsequently, we employ few-shot and zero-shot learn" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 331, + 555, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 463 + ], + "type": "text", + "content": "ing configurations to evaluate: We randomly sample only a few LASA real training data to finetune the pretrained network. In this way, the real-world knowledge of the network will mostly come from the pretraining data (i.e., the simulated data produced by our method and others) instead of the finetuning data, which can better reflect the performance gain from the simulated data itself. The pretrained model is finetuned with " + }, + { + "bbox": [ + 313, + 331, + 555, + 463 + ], + "type": "inline_equation", + "content": "20\\sim 30\\%" + }, + { + "bbox": [ + 313, + 331, + 555, + 463 + ], + "type": "text", + "content": " of the original training samples for few-shot learning and is evaluated on the original LASA validation set. For zero-shot learning, we directly evaluate the pretrained network without fine-tuning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 464, + 556, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 556, + 584 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 556, + 584 + ], + "type": "text", + "content": "Why not perform simulation on LASA synthetic data and compare the performance using the simulated data with that using LASA paired real GT? Since our StableSim2Real and other baseline techniques are trained on the LASA training set, it is unfeasible to conduct further inference on LASA, as this would not verify the actual model performance but only reflect the performance caused by overfitting. We can also not conduct simulation on a third-party paired dataset similar to LASA for 3D reconstruction, as there is currently no such paired dataset." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": "Quantitative results. Following [31], we utilize mean Intersection over Union (mIoU), L2 chamfer distance, and " + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": " F-score metrics. Initially, we normalize both the results and the ground-truth CAD objects to fall within the range of -0.5 to 0.5, according to which we calculate the aforementioned metrics between them. As shown in Tab. 1, under all few-shot settings, the model pretrained using the simulated data from our method achieves the best performance, and significantly surpasses the model using no simulated data (i.e., only add random perturbation to synthetic data) for" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2614" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 294, + 183 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 294, + 183 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 294, + 183 + ], + "type": "image", + "image_path": "414b82cce4a2b9fa53f937d90d7bdcabbc409f262a00e4fe4c768b326cc6e2da.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 190, + 296, + 226 + ], + "lines": [ + { + "bbox": [ + 55, + 190, + 296, + 226 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 296, + 226 + ], + "type": "text", + "content": "Figure 3. Qualitative comparison of few-shot evaluation on 3D instance reconstruction. \"w/o sim.\" denotes using no simulated data for pretraining." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 237, + 296, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 237, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 54, + 237, + 296, + 357 + ], + "type": "text", + "content": "pretraining. Moreover, using the simulated data produced by Cycle-GAN or Pix2Pix does not improve the model performance compared to that using no simulated data, possibly attributed to their inferior simulation quality, introducing ambiguity and complexity during pretraining, making it hard for the model to learn useful real-world knowledge. Moreover, as listed in Tab. 2, under zero-shot setting, when solely using simulated data for pretraining without any finetuning, the model trained with our simulated data demonstrates more conspicuous effectiveness." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 364, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 364, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 54, + 364, + 296, + 413 + ], + "type": "text", + "content": "Qualitative results. Fig. 3 shows the qualitative comparison, where our simulated data can help the model to gain the best reconstruction accuracy and diversity. More qualitative results are provided in the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 420, + 230, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 230, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 230, + 434 + ], + "type": "text", + "content": "4.3. 3D Object/Scene Understanding" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 437, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 437, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 54, + 437, + 295, + 569 + ], + "type": "text", + "content": "Setup. We select OcCo [63] as our backbone, which is a self-supervised point cloud pretraining framework. OcCo is originally pretrained on ModelNet [68], and finetuned on real-scanned datasets - ScanObjectNN [61] for object classification, and S3DIS [1] for scene semantic segmentation. ModelNet contains 12,311 synthetic CAD objects from 40 classes. ScanObjectNN includes " + }, + { + "bbox": [ + 54, + 437, + 295, + 569 + ], + "type": "inline_equation", + "content": "\\sim 15,000" + }, + { + "bbox": [ + 54, + 437, + 295, + 569 + ], + "type": "text", + "content": " real-scanned objects across 15 categories. S3DIS encompasses 271 indoor rooms and 13 classes. The classification task outputs the class of the whole 3D object, and the semantic segmentation predicts the category of each point in the 3D scene." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "content": "In our experiment, we apply our method and other baselines on ModelNet to generate simulated data that are further used to replace half of the original ModelNet synthetic data to pretrain OcCo. Next, the pretrained model is again evaluated under the few-shot learning setting. A typical setting is \"K-way " + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "content": "-shot\", which indicates that " + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "content": " classes are randomly selected during finetuning, and each category contains " + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "content": " samples. The finetuned model is evaluated on the unseen samples picked from each of the " + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "content": " classes. We select DGCNN [66] as the backbone for OcCo. For few-shot learning, we employ the \"5-way, 10-shot\" setting. For S3DIS, we follow the original OcCo to report the results" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 327, + 71, + 541, + 107 + ], + "blocks": [ + { + "bbox": [ + 327, + 71, + 541, + 107 + ], + "lines": [ + { + "bbox": [ + 327, + 71, + 541, + 107 + ], + "spans": [ + { + "bbox": [ + 327, + 71, + 541, + 107 + ], + "type": "table", + "html": "
Methodw/o sim.C-GAN [77]P2P [22]G-GAN [25]SD [42]Ours
Acc. (%)72.472.572.174.274.677.5
mIoU (%)47.948.047.749.148.951.2
", + "image_path": "8d040676f80bf275281d976a4104927d21a993e25e333704013cf643a3f52223.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 153, + 553, + 175 + ], + "blocks": [ + { + "bbox": [ + 313, + 110, + 555, + 145 + ], + "lines": [ + { + "bbox": [ + 313, + 110, + 555, + 145 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 555, + 145 + ], + "type": "text", + "content": "Table 3. Quantitative comparison of few-shot evaluation on 3D object classification and 3D scene semantic segmentation. \"w/o sim\" means using no simulated data for pretraining." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 153, + 553, + 175 + ], + "lines": [ + { + "bbox": [ + 315, + 153, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 315, + 153, + 553, + 175 + ], + "type": "table", + "html": "
Methodw/o sim.C-GAN [77]P2P [22]G-GAN [25]SD [42]Ours
Metrics59.2/3.54/25.360.4/3.21/25.261.5/3.02/26.462.1/2.86/26.762.3/2.83/27.562.9/2.57/28.2
", + "image_path": "408c2c2736329846d7b13f2e54908d51e2fbc90eaf7aceec492f30b4ffc57c42.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 181, + 555, + 204 + ], + "lines": [ + { + "bbox": [ + 313, + 181, + 555, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 181, + 555, + 204 + ], + "type": "text", + "content": "Table 4. The mIoU / Chamfer L2 / F-score between simulated data and real-captures on LASA [31] validation set." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 217, + 555, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 243 + ], + "type": "text", + "content": "using 6-fold cross-validation. All the other configurations follow the original OcCo." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 251, + 555, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 360 + ], + "type": "text", + "content": "Qualitative results. We report the classification accuracy and segmentation mean Intersection over Union (mIoU) in Tab. 3. The model pretrained using our simulated data gets the best results across both 3D object classification and 3D scene understanding tasks. By pretraining with our simulated data, the model can acquire valuable real-world knowledge that helps downstream tasks. In contrast, pretraining with the low-quality simulated data from CycleGAN or Pix2Pix may even degrade the performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 369, + 541, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 541, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 541, + 382 + ], + "type": "text", + "content": "4.4. Is the simulated data close to real captures?" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 386, + 554, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 554, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 554, + 530 + ], + "type": "text", + "content": "As previously mentioned, our Stable-Sim2Real is trained on LASA training set. In this section, we aim to evaluate the data simulation quality by directly performing inference and evaluation on LASA validation set. Tab. 4 compares the mIoU, Chamfer L2, and F-score metrics between the simulated data generated by different methods and the real captured Ground Truth (GT). Although not precise, this result can partially prove that our simulated data closely resembles real-world captures. Fig. 4 shows simulation results on depth maps and 3D point clouds. It demonstrates that while our simulated data closely aligns with real-world captures, subtle differences persist, highlighting the diversity." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 530, + 555, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 555, + 602 + ], + "type": "text", + "content": "We also provide the simulation results on ShapeNet [8] in Fig. 5. Despite the absence of real-captured Ground Truth (GT) in ShapeNet, the outcomes still show that the simulated results from our output closely resemble real captures, where the generated geometry is not largely broken, with certain realistic perturbations introduced." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 611, + 479, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 479, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 479, + 624 + ], + "type": "text", + "content": "4.5. Ablation Studies and Analysis" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": "Module impacts. To evaluate the impact of each key design in our framework, we conduct ablation studies on the 3D instance reconstruction task, specifically on the table category. Similar to Tab. 2, we conduct experiments under zero-shot protocol. This allows for a focused evaluation of the module's impact without the influence of additional factors. Other setups remain consistent with Sec. 4.2." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "2615" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 68, + 294, + 180 + ], + "blocks": [ + { + "bbox": [ + 57, + 68, + 294, + 180 + ], + "lines": [ + { + "bbox": [ + 57, + 68, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 57, + 68, + 294, + 180 + ], + "type": "image", + "image_path": "67b8575547711c1a0e0a0c8bc82a52321364a1a7a3c0690e3cdec57bd75550dd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 188, + 275, + 200 + ], + "lines": [ + { + "bbox": [ + 75, + 188, + 275, + 200 + ], + "spans": [ + { + "bbox": [ + 75, + 188, + 275, + 200 + ], + "type": "text", + "content": "Figure 4. Qualitative simulation results on LASA [31]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 209, + 293, + 373 + ], + "blocks": [ + { + "bbox": [ + 57, + 209, + 293, + 373 + ], + "lines": [ + { + "bbox": [ + 57, + 209, + 293, + 373 + ], + "spans": [ + { + "bbox": [ + 57, + 209, + 293, + 373 + ], + "type": "image", + "image_path": "e9db5cfe08a8b76f9d87cca1c13c177fcb31b0ffe4d5424382446be9a81cddb7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 383, + 278, + 395 + ], + "lines": [ + { + "bbox": [ + 71, + 383, + 278, + 395 + ], + "spans": [ + { + "bbox": [ + 71, + 383, + 278, + 395 + ], + "type": "text", + "content": "Figure 5. Qualitative simulation results on ShapeNet [8]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 410, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 410, + 295, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 295, + 458 + ], + "type": "text", + "content": "1) Residual generation. In Tab. 5 (w/o res.), directly generating the real-world depth map without using residual generation degrades the model performance, verifying the efficacy of our residual generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 460, + 295, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 507 + ], + "type": "text", + "content": "2) Loss re-weighting. In our Stage-II model, we reweight the diffusion loss to focus more on unsatisfactory areas. In Tab. 5 (w/o re-w.), the result w/o loss re-weighting is provided, showing its effectiveness." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 509, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 509, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 54, + 509, + 295, + 675 + ], + "type": "text", + "content": "3) 3D binary discriminator. Instead of training a 3D binary discriminator, another simple way is to directly set a threshold of the absolute difference between Stage-I generation and real-world GT to decide distinct areas. We tried various difference thresholds and reported the best result in Tab. 5 (w/o bin.), which hurts the model performance. To explain, our output depth is gained by adding subtle noise to synthetic depth. Thus, the magnitude of the difference between our output and its corresponding real GT is similar at 2D/pixel space, where it is impractical to find an appropriate threshold to identify distinct/unsatisfactory areas. Instead, their difference in distribution space is relatively more significant, so using a 3D binary classifier to delineate their distribution boundary is a more feasible way." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "4) Number of stages. Theoretically, we can iteratively improve the generation quality by adding more stages. However, as shown in Tab. 5, adding Stage-III model does" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "type": "table", + "bbox": [ + 317, + 71, + 554, + 90 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 554, + 90 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 554, + 90 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 554, + 90 + ], + "type": "table", + "html": "
Methodw/o res.w/o re-w.w/o bin.Stage-IIIStage-IVStage-II
Metrics26.4/18.2/22.127.1/17.9/23.327.3/17.5/23.628.6/ 16.1/24.528.0/16.9/24.128.5/16.1/24.7
", + "image_path": "442c5b995a84f9f0cac3839eae86a18d120d115511d351525e6a187545e82139.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 100, + 555, + 133 + ], + "lines": [ + { + "bbox": [ + 313, + 100, + 555, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 100, + 555, + 133 + ], + "type": "text", + "content": "Table 5. Quantitative ablations of zero-shot evaluation on 3D instance reconstruction of table category. Evaluation metrics are mIoU / Chamfer L2 / F-score respectively." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 352, + 144, + 519, + 247 + ], + "blocks": [ + { + "bbox": [ + 352, + 144, + 519, + 247 + ], + "lines": [ + { + "bbox": [ + 352, + 144, + 519, + 247 + ], + "spans": [ + { + "bbox": [ + 352, + 144, + 519, + 247 + ], + "type": "image", + "image_path": "2ab4154292e764e9b073cdd8aaee375cb02008785b2a33da79a2a12fd61425a9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 256, + 538, + 267 + ], + "lines": [ + { + "bbox": [ + 329, + 256, + 538, + 267 + ], + "spans": [ + { + "bbox": [ + 329, + 256, + 538, + 267 + ], + "type": "text", + "content": "Figure 6. Qualitative simulation results at the patch level." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 280, + 555, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 280, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 555, + 352 + ], + "type": "text", + "content": "not improve the model performance, and Stage-IV model even slightly degrades the result. To analyze, Stage-II model has been sufficient to identify most of the unsatisfactory areas and improve their generation quality. In addition, adding redundant stages may also cause overfitting issues that hurt the performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 358, + 554, + 538 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 313, + 358, + 554, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 554, + 442 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 554, + 442 + ], + "type": "text", + "content": "Analyzing Stage-II model at 3D patch-level. 1) We compare simulation results with real-captured Ground Truth (GT) at the patch level. As illustrated in Fig. 6, our Stage-II model effectively refines previously unsatisfactory (i.e., dissimilar to GT) patches from Stage-I output to be similar to GT, while keeping previously satisfactory (i.e., similar to GT) patches almost unchanged." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 443, + 554, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 554, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 554, + 538 + ], + "type": "text", + "content": "2) When sending the newly generated output patches from Stage-II to our 3D binary discriminator, the classification accuracy notably drops from " + }, + { + "bbox": [ + 313, + 443, + 554, + 538 + ], + "type": "inline_equation", + "content": "67.8\\%" + }, + { + "bbox": [ + 313, + 443, + 554, + 538 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 443, + 554, + 538 + ], + "type": "inline_equation", + "content": "32.5\\%" + }, + { + "bbox": [ + 313, + 443, + 554, + 538 + ], + "type": "text", + "content": ", suggesting that the generated patches successfully approximate the real-captured distributions, making it challenging for the classifier to discriminate. This again verifies the effectiveness of our Stage-II model. See the supplementary material for more ablations of the binary classifier." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 570, + 555, + 713 + ], + "type": "text", + "content": "We introduce Stable-Sim2Real, a two-stage depth diffusion model for 3D real data simulation. Initially, our method learns the residual between synthetic and real depth to generate a stable but coarse depth, where some areas may deviate from realistic patterns. Next, both synthetic and initial output depth are fed into the second-stage model, where diffusion loss prioritizes enhancing distinct regions identified by a binary classifier. Experiments show that the 3D simulated data generated from our method can improve the network performance in real-world 3D tasks, and our simulated data is highly similar to real captures. More limitations are discussed in the supplementary material." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2616" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 92, + 297, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 92, + 297, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 92, + 297, + 236 + ], + "type": "text", + "content": "The work was supported in part by Guangdong S&T Programme with Grant No. 2024B0101030002, the Basic Research Project No. HZQB-KCZYZ-2021067 of Hetao Shenzhen-HK S&T Cooperation Zone, Guangdong Provincial Outstanding Youth Fund (No. 2023B1515020055), No. NSFC 62172348, the Shenzhen Outstanding Talents Training Fund 202002, the Guangdong Research Projects No. 2017ZT07X152 and No. 2019CX01X104, the Guangdong Provincial Key Laboratory of Future Networks of Intelligence (Grant No. 2022B1212010001), and the Shenzhen Key Laboratory of Big Data and Artificial Intelligence (Grant No. ZDSYS201707251409055)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 249, + 115, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 249, + 115, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 249, + 115, + 262 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 269, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 269, + 294, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 269, + 294, + 301 + ], + "spans": [ + { + "bbox": [ + 61, + 269, + 294, + 301 + ], + "type": "text", + "content": "[1] I. Armeni, O. Sener, A. R. Zamir, H. Jiang, I. Brilakis, M. Fischer, and S. Savarese. 3d semantic parsing of large-scale indoor spaces. In CVPR, 2016. 1, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 304, + 294, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 304, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 61, + 304, + 294, + 346 + ], + "type": "text", + "content": "[2] Kaixin Bai, Lei Zhang, Zhaopeng Chen, Fang Wan, and Jianwei Zhang. Close the sim2real gap via physically-based structured light synthetic data simulation. In ICRA, 2024. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 350, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 350, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 62, + 350, + 294, + 415 + ], + "type": "text", + "content": "[3] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Yuri Feigin, Peter Fu, Thomas Gebauer, Daniel Kurz, Tal Dimry, Brandon Joffe, Arik Schwartz, and Elad Shulman. ARK-scenes: A diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In NeurlPS, 2021. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 418, + 294, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 418, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 62, + 418, + 294, + 460 + ], + "type": "text", + "content": "[4] Konstantinos Bousmalis, Nathan Silberman, David Dohan, Dumitru Erhan, and Dilip Krishnan. Unsupervised pixel-level domain adaptation with generative adversarial networks. In CVPR, 2017. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 464, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 464, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 62, + 464, + 294, + 518 + ], + "type": "text", + "content": "[5] Konstantinos Bousmalis, Alex Irpan, Paul Wohlhart, Yunfei Bai, Matthew Kelcey, Mrinal Kalakrishnan, Laura Downs, Julian Ibarz, Peter Pastor, Kurt Konolige, et al. Using simulation and domain adaptation to improve efficiency of deep robotic grasping. In ICRA, 2018. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 521, + 294, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 521, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 62, + 521, + 294, + 563 + ], + "type": "text", + "content": "[6] Marc Bresson, Yang Xing, and Weisi Guo. Sim2real: Generative ai to enhance photorealism through domain transfer with gan and seven-chanel- " + }, + { + "bbox": [ + 62, + 521, + 294, + 563 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 62, + 521, + 294, + 563 + ], + "type": "text", + "content": "-paired-images dataset. Sensors, 24(1):94, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 567, + 294, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 294, + 609 + ], + "type": "text", + "content": "[7] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. In 3DV, 2017. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 613, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 613, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 62, + 613, + 294, + 666 + ], + "type": "text", + "content": "[8] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 1, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 294, + 713 + ], + "type": "text", + "content": "[9] Hansheng Chen, Ruoxi Shi, Yulin Liu, Bokui Shen, Jiayuan Gu, Gordon Wetzstein, Hao Su, and Leonidas Guibas. Generic 3d diffusion adapter using controlled multi-view editing. arXiv preprint arXiv:2403.12032, 2024. 5" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[10] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksen, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In CVPR, 2022. 1, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 129, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 171 + ], + "type": "text", + "content": "[11] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 173, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 215 + ], + "type": "text", + "content": "[12] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 217, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 237 + ], + "type": "text", + "content": "[13] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. In NeurlPS, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 239, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 239, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 317, + 239, + 553, + 270 + ], + "type": "text", + "content": "[14] Yiquan Duan, Xianda Guo, and Zheng Zhu. Diffusiondepth: Diffusion denoising approach for monocular depth estimation. In ECCV, 2025. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 272, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 272, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 317, + 272, + 553, + 315 + ], + "type": "text", + "content": "[15] Hao-Shu Fang, Chenxi Wang, Hongjie Fang, Minghao Gou, Jirong Liu, Hengxu Yan, Wenhai Liu, Yichen Xie, and Cewu Lu. Anygrasp: Robust and efficient grasp perception in spatial and temporal domains. T-RO, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 316, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 316, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 317, + 316, + 553, + 358 + ], + "type": "text", + "content": "[16] Huan Fu, Bowen Cai, Lin Gao, Ling-Xiao Zhang, Jiaming Wang, Cao Li, Qixun Zeng, Chengyue Sun, Rongfei Jia, Binqiang Zhao, et al. 3d-front: 3d furnished rooms with layouts and semantics. In ICCV, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 360, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 360, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 317, + 360, + 553, + 392 + ], + "type": "text", + "content": "[17] Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3d-future: 3d furniture shape with texture. *IJCT*, 2021. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 394, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 394, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 317, + 394, + 553, + 436 + ], + "type": "text", + "content": "[18] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 2020. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 437, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 437, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 317, + 437, + 553, + 469 + ], + "type": "text", + "content": "[19] Xiao Gu, Yao Guo, Fani Deligianni, and Guang-Zhong Yang. Coupled real-synthetic domain adaptation for real-world deep depth enhancement. TIP, 2020. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 472, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 553, + 491 + ], + "type": "text", + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 494, + 555, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 494, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 494, + 555, + 536 + ], + "type": "text", + "content": "[21] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. Journal of Machine Learning Research, 23(47):1-33, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 537, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 553, + 568 + ], + "type": "text", + "content": "[22] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In CVPR, 2017. 3, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 570, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 570, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 317, + 570, + 553, + 601 + ], + "type": "text", + "content": "[23] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engil Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In CVPR, 2014. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 604, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 553, + 635 + ], + "type": "text", + "content": "[24] Cheongwoong Kang, Wonjoon Chang, and Jaesik Choi. Balanced domain randomization for safe reinforcement learning. Applied Sciences, 14(21):9710, 2024. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 636, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 553, + 669 + ], + "type": "text", + "content": "[25] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In CVPR, 2023. 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[26] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In CVPR, 2024. 3" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2617" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[27] Numair Khan, Min H Kim, and James Tompkin. Differentiable diffusion for dense depth estimation from multi-view images. In CVPR, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 129 + ], + "type": "text", + "content": "[28] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. STAT, 1050:1, 2014. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 294, + 162 + ], + "type": "text", + "content": "[29] Naveen Kodali, Jacob D. Abernethy, James Hays, and Zsolt Kira. On convergence and stability of gans. arXiv preprint arXiv:1705.07215, 2017. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 164, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 294, + 196 + ], + "type": "text", + "content": "[30] Michael J Landau, Benjamin Y Choo, and Peter A Beling. Simulating kinect infrared and depth images. IEEE transactions on cybernetics, 2015. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "type": "text", + "content": "[31] Haolin Liu, Chongjie Ye, Yinyu Nie, Yingfan He, and Xiaoguang Han. Lasa: Instance reconstruction from real scans using a large-scale aligned shape annotation dataset. In CVPR, 2024. 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 243, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 285 + ], + "type": "text", + "content": "[32] Naoki Matsunaga, Masato Ishii, Akio Hayakawa, Kenji Suzuki, and Takuya Narihira. Fine-grained image editing by pixel-wise guidance using diffusion models. arXiv preprint arXiv:2212.02024, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 288, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 288, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 288, + 294, + 319 + ], + "type": "text", + "content": "[33] Stephan Meister, Rahul Nair, and Daniel Kondermann. Simulation of time-of-flight sensors using global illumination. In VMV, 2013. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 321, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 365 + ], + "type": "text", + "content": "[34] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 366, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 294, + 421 + ], + "type": "text", + "content": "[35] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 422, + 294, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 444 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 444 + ], + "type": "text", + "content": "[36] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 445, + 294, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 478 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 478 + ], + "type": "text", + "content": "[37] Alexander Pashevich, Robin Strudel, Igor Kalevatykh, Ivan Laptev, and Cordelia Schmid. Learning to augment synthetic images for sim2real policy transfer. In IROS, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 479, + 294, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 511 + ], + "type": "text", + "content": "[38] Duo Peng, Ping Hu, Qiuhong Ke, and Jun Liu. Diffusion-based image translation with label guidance for domain adaptive semantic segmentation. In ICCV, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 514, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 294, + 567 + ], + "type": "text", + "content": "[39] Benjamin Planche, Ziyan Wu, Kai Ma, Shanhui Sun, Stefan Kluckner, Oliver Lehmann, Terrence Chen, Andreas Hutter, Sergey Zakharov, Harald Kosch, et al. Depthsynth: Realtime realistic synthetic data generation from cad models for 2.5 d recognition. In 3DV, 2017. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 569, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 569, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 569, + 294, + 602 + ], + "type": "text", + "content": "[40] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 603, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 666 + ], + "type": "text", + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "text", + "content": "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "type": "text", + "content": "[43] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 553, + 150 + ], + "type": "text", + "content": "[44] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. TPAMI, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 152, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 152, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 315, + 152, + 553, + 185 + ], + "type": "text", + "content": "[45] Saurabh Saxena, Abhishek Kar, Mohammad Norouzi, and David J Fleet. Monocular depth estimation using diffusion models. arXiv preprint arXiv:2302.14816, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 186, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 186, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 315, + 186, + 553, + 229 + ], + "type": "text", + "content": "[46] Saurabh Saxena, Charles Herrmann, Junhwa Hur, Abhishek Kar, Mohammad Norouzi, Deqing Sun, and David J Fleet. The surprising effectiveness of diffusion models for optical flow and monocular depth estimation. NeurIPS, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 231, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 231, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 315, + 231, + 553, + 274 + ], + "type": "text", + "content": "[47] Yuefan Shen, Yanchao Yang, Youyi Zheng, C Karen Liu, and Leonidas J Guibas. Dcl: Differential contrastive learning for geometry-aware depth synthesis. IEEE Robotics and Automation Letters, 2022. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 276, + 553, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 276, + 553, + 318 + ], + "spans": [ + { + "bbox": [ + 315, + 276, + 553, + 318 + ], + "type": "text", + "content": "[48] Yuan Shi, Huiyun Cao, Bin Xia, Rui Zhu, Qingmin Liao, and Wenming Yang. Dsr-diff: Depth map super-resolution with diffusion model. Pattern Recognition Letters, 184:225-231, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 319, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 319, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 315, + 319, + 553, + 364 + ], + "type": "text", + "content": "[49] Ashish Shrivastava, Tomas Pfister, Oncel Tuzel, Joshua Susskind, Wenda Wang, and Russell Webb. Learning from simulated and unsupervised images through adversarial training. In CVPR, 2017. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 365, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 365, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 315, + 365, + 553, + 407 + ], + "type": "text", + "content": "[50] Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 410, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 410, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 315, + 410, + 553, + 442 + ], + "type": "text", + "content": "[51] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 445, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 445, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 315, + 445, + 553, + 476 + ], + "type": "text", + "content": "[52] Shuran Song, Samuel P. Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In CVPR, 2015. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 478, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 478, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 315, + 478, + 553, + 510 + ], + "type": "text", + "content": "[53] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In NeurIPS, 2019, 2, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 315, + 512, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 512, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 315, + 512, + 553, + 533 + ], + "type": "text", + "content": "[54] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. In NeurIPS, 2020." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 315, + 535, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 535, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 315, + 535, + 553, + 578 + ], + "type": "text", + "content": "[55] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 315, + 580, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 580, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 315, + 580, + 553, + 611 + ], + "type": "text", + "content": "[56] Chris Sweeney, Greg Izatt, and Russ Tedrake. A supervised approach to predicting noise in depth images. In ICRA, 2019. 1, 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 315, + 613, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 613, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 315, + 613, + 553, + 656 + ], + "type": "text", + "content": "[57] Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. In IROS, 2017. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 315, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 658, + 553, + 712 + ], + "type": "text", + "content": "[58] Josh Tobin, Lukas Biewald, Rocky Duan, Marcin Andrychowicz, Ankur Handa, Vikash Kumar, Bob McGrew, Alex Ray, Jonas Schneider, Peter Welinder, et al. Domain randomization and generative models for robotic grasping. In IROS, 2018. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2618" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "text", + "content": "[59] Fabio Tosi, Pierluigi Zama Ramirez, and Matteo Poggi. Diffusion models for monocular depth estimation: Overcoming challenging conditions. arXiv preprint arXiv:2407.16698, 2024.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 117, + 294, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 117, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 117, + 294, + 171 + ], + "type": "text", + "content": "[60] Jonathan Tremblay, Aayush Prakash, David Acuna, Mark Brophy, Varun Jampani, Cem Anil, Thang To, Eric Cameracci, Shaad Boochoon, and Stan Birchfield. Training deep networks with synthetic data: Bridging the reality gap by domain randomization. In CVPRW, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 294, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 294, + 215 + ], + "type": "text", + "content": "[61] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 237 + ], + "type": "text", + "content": "[62] Arash Vahdat, Karsten Kreis, and Jan Kautz. Score-based generative modeling in latent space. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 294, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 271 + ], + "type": "text", + "content": "[63] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matthew J. Kusner. Unsupervised point cloud pre-training via occlusion completion. In ICCV, 2021. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 272, + 294, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 314 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 314 + ], + "type": "text", + "content": "[64] Jiyuan Wang, Chunyu Lin, Lang Nie, Kang Liao, Shuwei Shao, and Yao Zhao. Digging into contrastive learning for robust depth estimation with diffusion models. In ACM MM, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 316, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 337 + ], + "type": "text", + "content": "[65] Leichen Wang, Bastian Goldluecke, and Carsten Anklam. L2r gan: Lidar-to-radar translation. In ACCV, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 338, + 294, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 371 + ], + "type": "text", + "content": "[66] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic graph cnn for learning on point clouds. TOG, 2019. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 372, + 294, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 294, + 414 + ], + "type": "text", + "content": "[67] Zhen Wang, Qiangeng Xu, Feitong Tan, Mengelei Chai, Shichen Liu, Rohit Pandey, Sean Fanello, Achuta Kadambi, and Yinda Zhang. Mvdd: Multi-view depth diffusion models. In ECCV, 2025. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 415, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 458 + ], + "type": "text", + "content": "[68] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In CVPR, 2015. 1, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 460, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 491 + ], + "type": "text", + "content": "[69] Mutian Xu, Pei Chen, Haolin Liu, and Xiaoguang Han. Toscane: A large-scale dataset for understanding 3d tabletop scenes. In ECCV, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 493, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 546 + ], + "type": "text", + "content": "[70] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Tianyou Liang, Guanying Chen, Shuguang Cui, and Xiaoguang Han. Mvimgnet: A large-scale dataset of multi-view images. In CVPR, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 548, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 613 + ], + "type": "text", + "content": "[71] Chengjie Yuan, Yunlei Shi, Qian Feng, Chunyang Chang, Michael Liu, Zhaopeng Chen, Alois Christian Knoll, and Jianwei Zhang. Sim-to-real transfer of robotic assembly with visual inputs using cyclegan and force control. In 2022 IEEE International Conference on Robotics and Biomimetics (ROBIO), pages 1426-1432. IEEE, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 614, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 679 + ], + "type": "text", + "content": "[72] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2100-2110, 2019. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "text", + "content": "[73] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 4" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 262 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 105 + ], + "type": "text", + "content": "[74] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 171 + ], + "type": "text", + "content": "[75] Xiaoshuai Zhang, Rui Chen, Ang Li, Fanbo Xiang, Yuzhe Qin, Jiayuan Gu, Zhan Ling, Minghua Liu, Peiyu Zeng, Songfang Han, Zhiao Huang, Tongzhou Mu, Jing Xu, and Hao Su. Close the optical sensing domain gap by physics-grounded active stereo sensor simulation. IEEE Transactions on Robotics, 2023. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 174, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 228 + ], + "type": "text", + "content": "[76] Xiang Zhang, Bingxin Ke, Hayko Riemenschneider, Nando Metzger, Anton Obukhov, Markus Gross, Konrad Schindler, and Christopher Schroers. Betterdepth: Plug-and-play diffusion refiner for zero-shot monocular depth estimation. arXiv preprint arXiv:2407.17952, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 229, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 229, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 229, + 553, + 262 + ], + "type": "text", + "content": "[77] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, 2017. 3, 6, 7" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2619" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/StableCodec_ Taming One-Step Diffusion for Extreme Image Compression/5269d8ed-d6d6-4961-a311-a0e841365f77_content_list.json b/2025/StableCodec_ Taming One-Step Diffusion for Extreme Image Compression/5269d8ed-d6d6-4961-a311-a0e841365f77_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..eb89230984a600c1095450cd23b0d9bde836b790 --- /dev/null +++ b/2025/StableCodec_ Taming One-Step Diffusion for Extreme Image Compression/5269d8ed-d6d6-4961-a311-a0e841365f77_content_list.json @@ -0,0 +1,1627 @@ +[ + { + "type": "text", + "text": "StableCodec: Taming One-Step Diffusion for Extreme Image Compression", + "text_level": 1, + "bbox": [ + 122, + 130, + 875, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tianyu Zhang Xin Luo Li Li Dong Liu", + "bbox": [ + 282, + 181, + 712, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "MOE Key Laboratory of Brain-Inspired Intelligent Perception and Cognition, University of Science and Technology of China, Hefei 230093, China", + "bbox": [ + 191, + 199, + 803, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhangtianyu, xinluo}@mail.ustc.edu.cn, {lil1, dongeliu}@ustc.edu.cn", + "bbox": [ + 197, + 236, + 794, + 251 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/72f1cd571559f186a71ddf82a499acc1730b5a3132351d5605ad61029481f7a6.jpg", + "image_caption": [ + "Figure 1. Visual examples and comparisons when compressing a 4K-resolution image [38] at ultra-low bitrates. The proposed Stable-Coded produces more realistic and consistent details with fewer bits. In contrast, VVC [10], ELIC [20] and MS-ILLM [46] reconstructions are blurry, while PerCo [11] and DiffEIC [40] generate inconsistent details against the original images. Best viewed on screen for details." + ], + "image_footnote": [], + "bbox": [ + 93, + 268, + 903, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 669, + 795, + 750, + 811 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work is supported by the Natural Science Foundation of China under Grant 62021001. We acknowledge the support of GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC. (Corresponding author: Dong Liu.)", + "bbox": [ + 89, + 849, + 486, + 902 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Diffusion-based image compression has shown remarkable potential for achieving ultra-low bitrate coding (less than 0.05 bits per pixel) with high realism, by leveraging the generative priors of large pre-trained text-to-image diffu", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17379", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sion models. However, current approaches require a large number of denoising steps at the decoder to generate realistic results under extreme bitrate constraints, limiting their application in real-time compression scenarios. Additionally, these methods often sacrifice reconstruction fidelity, as diffusion models typically fail to guarantee pixel-level consistency. To address these challenges, we introduce StableCodec, which enables one-step diffusion for high-fidelity and high-realism extreme image compression with improved coding efficiency. To achieve ultra-low bitrates, we first develop an efficient Deep Compression Latent Codec to transmit a noisy latent representation for a single-step denoising process. We then propose a Dual-Branch Coding Structure, consisting of a pair of auxiliary encoder and decoder, to enhance reconstruction fidelity. Furthermore, we adopt end-to-end optimization with joint bitrate and pixel-level constraints. Extensive experiments on the CLIC 2020, DIV2K, and Kodak datasets demonstrate that StableCodec outperforms existing methods in terms of FID, KID and DISTS by a significant margin, even at bitrates as low as 0.005 bits per pixel, while maintaining strong fidelity. Additionally, StableCodec achieves inference speeds comparable to mainstream transform coding schemes. Code is available at our project repository.", + "bbox": [ + 89, + 90, + 483, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 482, + 222, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image compression is a foundational problem in signal processing. Driven by advances in digital imaging and the widespread use of social platforms, the volume of image data in modern multimedia has grown exponentially, placing increasing demands on the coding efficiency of image compression techniques. Over the past few decades, traditional CODECs such as JPEG [62] and H.266/VVC [10], along with emerging learning-based methods [4, 5, 12, 18, 19, 28, 42, 44, 45], have been widely adopted in real-world image compression applications. However, these methods are typically optimized for rate-distortion performance, and often produce unrealistic and blurry reconstructions, particularly under severe bitrate constraints, as shown in Fig. 1.", + "bbox": [ + 89, + 508, + 483, + 704 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle this issue, generative image compression [2, 43] optimized for human perceptual performance has gained increasing attention. These methods are evaluated based on the rate-distortion-perception tradeoff [8, 9, 67, 68], and progressively demonstrate their advantages in producing visually appealing reconstructions at lower bitrates compared to traditional CODEs or common neural CODEs. A prominent research direction [3, 21, 31, 43, 43, 46, 66] involves integrating a discriminator into the transform coding pipeline [5, 20, 44], employing adversarial training to enhance the perceptual quality of reconstructions. Motivated by the impressive generative capabilities, more researchers [11, 24, 35, 40, 51, 59, 65, 69] have begun exploring the po", + "bbox": [ + 89, + 705, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3c5ee440181aaca1b32bec22899a91ae3be7cb8eaa5c5c3a71b5a7fd47924454.jpg", + "image_caption": [ + "Figure 2. (Top) Illustration of our motivation. One-step diffusion can produce perceptually consistent results given severely corrupted images and a general prompt. (Bottom) Challenges in StableCodec: how to compress a noisy latent for one-step diffusion using ultra-low bitrates, and how to improve fidelity." + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 905, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tential of diffusion models, particularly the generative priors in large pre-trained text-to-image (T2I) models, to compensate for severely distorted information at ultra-low bitrates while ensuring perceptually consistent generation. A recent study, PerCo [11], produces realistic results at an extreme bitrate as low as 0.003 bits per pixel (bpp) using a pretrained latent diffusion model (LDM), highlighting the potential of diffusion-based generative CODEs on image compression at more severe bitrates.", + "bbox": [ + 511, + 393, + 906, + 530 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite these promising advancements, existing methods face two primary limitations inherent to diffusion models. First, they require dozens of denoising steps at the decoder to produce results with sufficient perceptual quality. Second, the reconstructions often deviate from the original images, as diffusion models typically do not guarantee reconstruction consistency. To address the first challenge, we consider leveraging the generative priors in SD-Turbo [54], a distilled version of Stable Diffusion 2.1 [52] that enables real-time image synthesis in 1 to 4 denoising steps. Following [73], we demonstrate that SD-Turbo can produce perceptually consistent reconstructions with a single-step denoising process, even for severely corrupted inputs and a general positive prompt, as shown in Fig. 2. We thus pose an intuitive question: Can we compress a noisy latent representation of the original image, which can be effectively denoised in a single-step diffusion process, using an ultra-low bitrate? Building on these insights, we present StableCodec for extreme image compression, which integrates SD-Turbo with the proposed Deep Compression Latent Codec to compress noisy latents at ultra-low bitrates for a single-step diffusion process.", + "bbox": [ + 511, + 534, + 908, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In response to the second challenge, we introduce a Dual-Branch Coding Structure with a pair of auxiliary en", + "bbox": [ + 511, + 869, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17380", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "coder and decoder to further enhance reconstruction fidelity. Considering the limitations of the pre-trained VAE encoder on practical entropy coding and reconstruction consistency, we employ a rate-distortion-oriented auxiliary encoder to embed more entropy-aware semantic information for coding decisions. In parallel, we add an auxiliary decoder to perform structure apportionment during the decoding process, improving the generation guidance on the one-step denoising process for more consistent details. To enable end-to-end optimization, we design a two-stage training objective that jointly optimizes bitrate and pixel-level constraints.", + "bbox": [ + 89, + 90, + 480, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Benefit from these designs, StableCodec produces high-fidelity and high-realism reconstructions at ultra-low bitrates as low as 0.005 bpp. Extensive experiments on CLIC 2020 [60], DIV2K [1] and Kodak [16] demonstrate that StableCodec sets up a new state-of-the-art performance in terms of FID [22], KID [6], and DISTS [14], outperforming existing methods by significant margins. In terms of computational complexity, StableCodec offers much faster decoding speeds compared to other diffusion-based competitors, achieving inference times comparable to those of mainstream transform coding schemes. For practical deployment, StableCodec supports inference at arbitrary resolutions with memory consumption less than 9 GB.", + "bbox": [ + 89, + 258, + 482, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We summarize our contributions as follows:", + "bbox": [ + 109, + 455, + 401, + 469 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present StableCodec, an extreme image codec integrating one-step diffusion and Deep Compression Latent Codec to achieve ultra-low bitrate compression with high realism, high fidelity and superior coding efficiency.", + "- We develop Dual-Branch Coding Structure to improve reconstruction fidelity. A pair of auxiliary encoder and decoder is introduced for semantic enhancement and structure apportionment.", + "- StableCodec obtains SOTA FID, KID and DISTS performance on CLIC 2020 and DIV2K dataset, significantly outperforms existing methods at bitrates as low as 0.005 bpp with well-preserved fidelity, and achieves comparable inference speeds with mainstream neural CODECs." + ], + "bbox": [ + 91, + 470, + 482, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 684, + 232, + 699 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Generative Image Compression", + "text_level": 1, + "bbox": [ + 89, + 710, + 367, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning-based image compression has shown competitive potential compared to traditional methods [10, 37, 39, 58, 62], leveraging non-linear transforms and joint rate-distortion optimization. Balle et al. [4] introduced the first end-to-end learned image compression framework, which was subsequently enhanced with the hyperprior [5] and context model [45]. Building on this foundation, much work [12, 18, 19, 28, 42, 44, 49] has been devoted to improving both rate-distortion performance and model practicality.", + "bbox": [ + 89, + 733, + 482, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In practical scenarios, a key challenge is achieving extreme image compression at ultra-low bitrates while main", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "taining both fidelity and realism [2]. Traditional image codes optimized for rate-distortion often produce blurry reconstructions and noticeable artifacts. To address this, Mentzer et al. [43] introduced HiFiC and the concept of generative image compression, integrating GANs into codec optimization and evaluating performance in terms of the rate-distortion-perception tradeoff [8, 9, 67, 68]. Subsequent research can be broadly categorized into two main approaches. The first category [3, 21, 31, 43, 46] focuses on enhancing transform coding [5, 20, 44] for human perception by incorporating adversarial losses and optimized discriminator architectures, which typically can be extended to a wide range of bitrate and a flexible decoding control between fidelity and realism [3, 31]. The second category [11, 24, 35, 40, 51, 59, 65, 69] leverages diffusion models for generative image compression. Although these methods show promise for ultra-low bitrate compression, they are often constrained by reconstruction fidelity and inference efficiency due to the multi-step denoising process. Recently, GLC [26] introduced transform coding in the generative latent space of VQ-VAE [15, 61], achieving more visually appealing results at ultra-low bitrates.", + "bbox": [ + 511, + 90, + 903, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Generative Models and Few-Step Diffusions", + "text_level": 1, + "bbox": [ + 511, + 435, + 885, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generative models play a crucial role in image generation. While many architectures, such as VAEs [30] and GANs [17], have been explored, diffusion models [55] have emerged as a powerful alternative, achieving state-of-the-art synthesis quality. Inspired by non-equilibrium statistical physics [55], diffusion models learn to reverse a noise perturbation process through a Markovian framework. Recent advancements, such as DDPM [23], DDIM [56], and LDM [52], have significantly reduced computational complexity and improved image synthesis quality, making diffusion-based approaches a dominant force in generative modeling.", + "bbox": [ + 511, + 458, + 903, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the inefficiency of iterative denoising, several approaches [53, 54, 57, 70] aim to reduce the number of denoising steps while maintaining generation quality. These methods train diffusion models to approximate the full denoising trajectory in a single or a few steps, significantly improving inference efficiency. Notably, SD/SDXL-Turbo [54] demonstrates image generation in 1 to 4 steps with near-parity quality compared to multi-step models, making real-time diffusion-based applications [48, 63, 73] feasible.", + "bbox": [ + 511, + 625, + 903, + 761 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 776, + 604, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 511, + 801, + 622, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce the overall framework of the proposed StableCodec, built upon SD-Turbo [54] with a VAE encoder $\\mathcal{E}_{\\mathrm{SD}}$ , a VAE decoder $\\mathcal{D}_{\\mathrm{SD}}$ and a denoising Unet $\\epsilon_{\\mathrm{SD}}$ . As shown in Fig. 3, we incorporate a Deep Compression Latent Codec to perform extreme transform", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17381", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fb19566377769b1588ee11eede0d60fe9da173e08c2d511ccb2532a752adc0e2.jpg", + "image_caption": [ + "Figure 3. The framework of StableCodec. We incorporate the proposed Deep Compression Latent Codec to transmit a noisy latent $l_{T}$ for one-step denoising, where $64 \\times$ denotes a spatial compression ratio of 64. To adjust the latent resolution, we deploy DownSample block and Conv $3 \\times 3$ as adapters after the VAE encoder $\\mathcal{E}_{\\mathrm{SD}}$ and auxiliary encoder $\\mathcal{E}_{\\mathrm{Aux}}$ , respectively. We use a general prompt in both training and inference. The auxiliary decoder $\\mathcal{D}_{\\mathrm{Aux}}$ shares a similar structure with $g_{s}$ . More details on networks are provided in the supplementary." + ], + "image_footnote": [], + "bbox": [ + 89, + 85, + 911, + 358 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "coding in the VAE latent space. To adapt SD-Turbo for image compression, we integrate LoRA [25] into $\\mathcal{E}_{\\mathrm{SD}}$ and $\\epsilon_{\\mathrm{SD}}$ , while keeping $\\mathcal{D}_{\\mathrm{SD}}$ unchanged to preserve the generative priors [73]. Additionally, we introduce a Dual-Branch Coding Structure to enhance reconstruction fidelity, utilizing an auxiliary encoder $\\mathcal{E}_{\\mathrm{Aux}}$ to embed rich semantic information and an auxiliary decoder $\\mathcal{D}_{\\mathrm{Aux}}$ to perform structure apportionment. Finally, we optimize StableCodec end-to-end with joint bitrate and pixel-level constraints, achieving high-fidelity and high-realism extreme image compression.", + "bbox": [ + 88, + 441, + 480, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Deep Compression Latent Codec", + "text_level": 1, + "bbox": [ + 89, + 604, + 377, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We design our latent codec using the extreme analysis transform $g_{a}$ , extreme synthesis transform $g_{s}$ and a 4-step autoregressive entropy model. To reach ultra-low bitrates, we employ deep compression transform networks for both $g_{a}$ and $g_{s}$ . Specifically, $\\mathcal{E}_{\\mathrm{SD}}$ and $\\mathcal{D}_{\\mathrm{SD}}$ provide a latent space with a spatial compression ratio of 8 (abbreviated as $8\\times$ ). Unlike mainstream schemes [18, 20, 26, 28, 40, 42, 49] that perform entropy coding at $16\\times$ , we further downsample and apply entropy coding for $\\hat{y}$ at $64\\times$ and the hyperprior [5] at $256\\times$ . Consequently, we use $g_{s}$ to restore the spatial compression ratio to $8\\times$ for $\\epsilon_{\\mathrm{SD}}$ and $\\mathcal{D}_{\\mathrm{SD}}$ . The entire coding process can be formulated as follows:", + "bbox": [ + 89, + 627, + 483, + 808 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nl = \\operatorname {c o n c a t} \\left[ \\mathcal {E} _ {\\mathrm {S D}} (x), \\mathcal {E} _ {\\text {A u x}} (x) \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 821, + 482, + 838 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ny = g _ {a} (l), \\hat {y} = Q (y), l _ {T} = g _ {s} (\\hat {y}) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 840, + 482, + 857 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nl _ {0} = \\left[ l _ {T} - \\sqrt {1 - \\bar {\\alpha} _ {T}} \\cdot \\epsilon_ {\\mathrm {S D}} \\left(l _ {T}, T\\right) \\right] / \\sqrt {\\bar {\\alpha} _ {T}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 861, + 480, + 878 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} = \\mathcal {D} _ {\\mathrm {S D}} \\left(l _ {0} + \\mathcal {D} _ {\\mathrm {A u x}} (\\hat {y})\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 880, + 482, + 896 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Eq. (1), we first obtain an intermediate latent $l$ from the input image $x$ through $\\mathcal{E}_{\\mathrm{SD}}$ and $\\mathcal{E}_{\\mathrm{Aux}}$ . Eq. (2) is the latent-space transform coding process to produce a noisy latent $l_{T}$ using ultra-low bitrates. Eq. (3) displays the one-step denoising process with the noise schedule $\\{\\bar{\\alpha}_t\\}$ [23] in the $T$ -th timestep. Finally, in Eq. (4), the reconstruction $\\hat{x}$ is obtained from $l_{0}$ and $\\hat{y}$ using $\\mathcal{D}_{\\mathrm{SD}}$ and $\\mathcal{D}_{\\mathrm{Aux}}$ . To balance performance and coding latency, we construct efficient $g_{a}$ and $g_{s}$ with InceptionNeXt [72] and GatedCNN [71], and build a 4-step antoregressive entropy model with quadtree partition [36] and latent residual prediction [44].", + "bbox": [ + 511, + 441, + 906, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Dual-Branch Coding Structure", + "text_level": 1, + "bbox": [ + 511, + 619, + 790, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While deploying the proposed latent codec with LoRA enables image compression with SD-Turbo at ultra-low bitrates, the reconstruction fidelity is limited. In this section, we analyze the reasons and introduce Dual-Branch Coding Structure with a pair of auxiliary encoder and decoder, $\\mathcal{E}_{\\mathrm{Aux}}$ and $\\mathcal{D}_{\\mathrm{Aux}}$ , to further enhance compression performance.", + "bbox": [ + 511, + 642, + 906, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1. Entropy-Aware Semantic Enhancement", + "text_level": 1, + "bbox": [ + 511, + 744, + 836, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We observed that the VAE in SD-Turbo has several limitations when reconstruction fidelity and practical coding are required. As noted in [52] and Table 1, this VAE is pretrained for perceptual compression, which does not preserve pixel-level fidelity as well as a rate-distortion-oriented autoencoder from a typical neural codec [20]. Additionally, while this VAE provides a compressed representation of the original image, it is still in floating-point format and not optimized for practical entropy coding, making it less suitable", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17382", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/97af191850c02dfc6d01df514b96753e378bba34b929e6d7fda6cd7255e46969.jpg", + "table_caption": [ + "Table 1. Reconstruction quality of VAEs on Kodak [16]. The pre-trained VAE in SD performs perceptual compression, while the one in ELIC [20] preserves more pixel-level information." + ], + "table_footnote": [], + "table_body": "
VAEPSNR↑MS-SSIM↑LPIPS↓DISTS↓
SD26.650.93180.07260.0415
ELIC40.400.99610.05550.0707
", + "bbox": [ + 102, + 133, + 470, + 188 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/65ec16377b6b7a4e67fe570150e2341b979702bc630cc024376c5898daf53581.jpg", + "image_caption": [ + "Figure 4. Top-energy channels learned from different encoders. $\\mathcal{E}_{\\mathrm{Aux}}$ embeds more pixel-level semantic information into latents." + ], + "image_footnote": [], + "bbox": [ + 96, + 200, + 485, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for further latent-space ultra-low bitrate compression.", + "bbox": [ + 89, + 406, + 444, + 420 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Building on these insights, we introduce the analysis transform of a pre-trained high-bitrate ELIC model [20] to serve as an auxiliary encoder $\\mathcal{E}_{\\mathrm{Aux}}$ . $\\mathcal{E}_{\\mathrm{Aux}}$ remains frozen, and provides rich, entropy-aware semantic information of the input images. To combine the latents from different encoders, we introduce trainable adapters that align the latent resolutions to a spatial compression ratio of 16, followed by channel-wise concatenation. During optimization, various pieces of information are learned from different encoders as shown in Fig. 4, where more pixel-level semantic information is embedded into the latent codec through $\\mathcal{E}_{\\mathrm{Aux}}$ .", + "bbox": [ + 88, + 421, + 483, + 587 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2. Structure Apportionment", + "text_level": 1, + "bbox": [ + 89, + 595, + 316, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The reconstruction quality of StableCodec heavily depends on how the denoising Unet is conditioned. Since a fixed prompt is used for both training and inference, the one-step denoising process is primarily guided by $l_{T}$ , which is produced by $g_{s}$ . This places high demands on the capability of $g_{s}$ , resulting in unsatisfactory denoising guidance, as reflected in the reconstructions shown in Fig. 5 (b).", + "bbox": [ + 89, + 613, + 483, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To alleviate the decoding burden on $g_{s}$ , we introduce an auxiliary decoder $\\mathcal{D}_{\\mathrm{Aux}}$ to perform an additional decoding branch from $\\hat{y}$ , bypassing the Unet. This design is motivated by the observation that an extremely compressed bitstream contains mainly the basic structure of images. We distribute and decode these components directly from the bitstream using the auxiliary branch, allowing $g_{s}$ to focus primarily on producing guidance to generate realistic details. Fig. 5 compares StableCodec that is trained with or without $\\mathcal{D}_{\\mathrm{Aux}}$ . When trained without $\\mathcal{D}_{\\mathrm{Aux}}$ , $g_{s}$ produces all types of information as it is the only decoding branch. When trained with $\\mathcal{D}_{\\mathrm{Aux}}$ , structural information is routed", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/83e18420fabd9eae154e1ae6c87af76a385445279d9f2cc6b87ff24af0f64499.jpg", + "image_caption": [ + "(a) Latents from different decoders, trained w. or w/o $\\mathcal{D}_{\\mathrm{Aux}}$" + ], + "image_footnote": [], + "bbox": [ + 517, + 87, + 910, + 311 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/25ba7824b0160ae6f6194eedcf32c522d1eb40166ef0338a518723382f6cf250.jpg", + "image_caption": [ + "(b) Reconstruction comparison, trained w. or w/o $\\mathcal{D}_{\\mathrm{Aux}}$", + "Figure 5. Impact of $\\mathcal{D}_{\\mathrm{Aux}}$ on latents and reconstructions." + ], + "image_footnote": [], + "bbox": [ + 532, + 325, + 893, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "through $\\mathcal{D}_{\\mathrm{Aux}}$ , while the energy in $g_{s}$ latents drops significantly. Meanwhile, more semantically aligned details are reconstructed in Fig. 5 (b), suggesting $g_{s}$ now provides better high-frequency guidance for the denoising Unet $\\epsilon_{\\mathrm{SD}}$ .", + "bbox": [ + 511, + 450, + 905, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. End-to-End Training Objective", + "text_level": 1, + "bbox": [ + 511, + 521, + 790, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We adopt end-to-end optimization with joint bitrate and pixel-level restrictions to train StableCodec. Given the original image $x$ , the quantized latent $\\hat{y}$ and the reconstructed image $\\hat{x}$ , we construct our training objective based on the standard rate-distortion loss:", + "bbox": [ + 511, + 542, + 905, + 617 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda \\mathcal {R} (\\hat {y}) + \\mathcal {D} (x, \\hat {x}) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 648, + 631, + 903, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the bitrate $\\mathcal{R}$ and pixel-level distortion $\\mathcal{D}$ are balanced by the Lagrange multiplier $\\lambda$ .", + "bbox": [ + 511, + 657, + 903, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by [20, 44], we train StableCodec with a 2-stage implicit bitrate pruning (IBP) strategy. We first train a base model using a smaller $\\lambda_{base}$ , adapting the latent codec into SD-Turbo under a relaxed bitrate constraint, and warming up with a more expressive transform. In the second stage, we finetune the shared base model with larger $\\lambda_{target}$ to reach ultra-low target bitrates. IBP facilitates efficient and stable training, resulting in improved performance.", + "bbox": [ + 511, + 689, + 905, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The distortion term $\\mathcal{D}$ includes MSE, LPIPS (using VGG features) [74] and a CLIP [50] distance $\\mathcal{L}_{CLIP}$ , for which we compute the L2-distance between the CLIP embeddings of $x$ and $\\hat{x}$ . Note that this term is a simplified version from [34], and we find it beneficial for reconstruction at ultralow bitrates. Additionally, we follow [73] and incorporate", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17383", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7f9d7f01a147b11667990979775e443d604673ba0ed3b3a58bac5fc972014eef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 192, + 93, + 802, + 132 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8b2210b11c21bd44326ced02d09aa107f91421454626268e90c0dbe2af2f5d21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 137, + 292, + 258 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/11fe770e88e21dfb7f2d021c8cee65889c7c31f74963780c0ba1a74b7f2ee2f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 137, + 493, + 258 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/81019f026d7b49fccdc154c011b104618da55869cf16c5e6061794d685f35b0a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 137, + 697, + 258 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1a69dd8cbcd53a10f0feaa9ddb2a0fcc11234f4ae36a01790ddf1039dd387bcb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 137, + 900, + 258 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c1bbb274d9ea3a973b75d174502c6912932e225677744d441bb5a0ffcc364d71.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 260, + 292, + 381 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/69ecf9d1bfe2780365a7fe85fd7d63aac2d4e435fe6a5f3a98b0468cdbe5aba4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 260, + 493, + 381 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/af1daffea9066d94a7acaec322648900023351dcf045c3c56974928194027c9f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 260, + 697, + 381 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6ccbb99bd2b025a4802db51eae451d9d8a947ba3c9a63076e7855d37fdbd6fb5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 260, + 900, + 381 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/50664a381cec20a3f97596b197e1f90dbe127b7408e8841b6ff3534c926bdd49.jpg", + "image_caption": [ + "Figure 6. Rate-distortion and rate-perception curve comparisons of different methods on the CLIC 2020, DIV2K and Kodak dataset." + ], + "image_footnote": [], + "bbox": [ + 96, + 381, + 292, + 503 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a89037520d9c1e38d270c748a76e5e136c40ee1fc101cf61a03570260c0459b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 381, + 495, + 503 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/741544a4de1adbd1d53be9b8cbbd9d8f8f73924c67ff2c2141275ab3d6765d6b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 381, + 697, + 503 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b0d4541c6ae912b0dd7cecea8c98d296195ecaed058b50c6f5e62bb0fd41e1ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 381, + 900, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "a GAN loss $\\mathcal{L}_{adv}$ to narrow the distribution gap between $x$ and $\\hat{x}$ . We use DINOv2 [47] with registers [13] as the discriminator backbone [33]. To ensure stable training, we incorporate the GAN only in the second training stage. The full objective can be formulated as:", + "bbox": [ + 89, + 549, + 483, + 625 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\text {S t a g e I :} \\underset {\\theta} {\\arg \\min } \\lambda_ {\\text {b a s e}} \\mathcal {R} (\\hat {y}) + \\mathcal {D} (x, \\hat {x}) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 637, + 482, + 661 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\text {S t a g e I I :} \\underset {\\theta} {\\arg \\min } \\lambda_ {\\text {t a r g e t}} \\mathcal {R} (\\hat {y}) + \\mathcal {D} (x, \\hat {x}) + \\beta \\mathcal {L} _ {\\text {a d v}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 664, + 482, + 689 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {D} (x, \\hat {x}) = d _ {1} M S E (x, \\hat {x}) + d _ {2} L P I P S (x, \\hat {x}) \\tag {8} \\\\ + d _ {3} \\mathcal {L} _ {C L I P} (x, \\hat {x}) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 705, + 483, + 741 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\theta$ represents all trainable parameters in StableCodec (Fig. 3), $d_{1}, d_{2}, d_{3}$ and $\\beta$ are balancing weights.", + "bbox": [ + 89, + 748, + 482, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 792, + 223, + 810 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation", + "text_level": 1, + "bbox": [ + 89, + 818, + 250, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Details. We use the training set of DF2K [41] and CLIC 2020 Professional [60] to train StableCodec. During training, we use $512 \\times 512$ patches with a batch size of 8. The first training stage takes over 100k iterations with a", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "learning rate of $1e^{-4}$ and a $\\lambda_{base}$ of 0.5 (about 0.05bpp). In the second stage, we finetune the base model for another 20k iterations with GAN incorporated and $\\lambda_{target} \\in \\{2,3,4,6,8,12,16,24,32\\}$ , while the learning rate undergoes $5e^{-5}, 2e^{-5}, 1e^{-5}$ and $1e^{-6}$ for 5k iterations each. We set $d_1, d_2, d_3$ and $\\beta$ to 2, 1, 0.1 and 0.1, respectively. All models are trained using 2 RTX 3090 GPUs.", + "bbox": [ + 511, + 549, + 903, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Test Data. We evaluate StableCodec on the test set of CLIC 2020 Professional [60] (CLIC 2020 Test), the validation set of DIV2K [1] (DIV2K Val) and Kodak [16] following [26, 40]. CLIC 2020 Test and DIV2K Val contain 428 and 100 high-quality 2K-resolution natural images, respectively, while Kodak contains 24 natural images with a smaller resolution of $768 \\times 512$ . We evaluate all images with the original resolution as detailed in the supplementary.", + "bbox": [ + 511, + 656, + 905, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We employ established metrics to assess the rate-distortion-perception performance of StableCodec. Concretely, we measure bitrate by bits per pixel (bpp), and evaluate perceptual quality using FID [22], KID [6], DISTS [14] and LPIPS [74] (using AlexNet features by default). Meanwhile, we use PSNR and MS-SSIM [64] to measure the reconstruction fidelity. We follow [26, 40, 46] to calculate FID and KID on $256 \\times 256$ patches, and neglect", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17384", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3e87a132c57a35619e00a779af9e427ef509fd54c56ae3a052d0c677b4768983.jpg", + "image_caption": [ + "Figure 7. Qualitative comparisons of different methods on Kodak. Best viewed on screen for details." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 903, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the results on Kodak as it is too small for calculating. Note that pixel-level distortion metrics like LPIPS, PSNR and MS-SSIM have strong limitations when evaluating image compression at ultra-low bitrates [11, 14, 26, 35]. Therefore, for StableCodec, we focus primarily on FID, KID, and DISTS, which offer a more accurate assessment of quality in severely compressed images. We also provide user study in the supplementary to visually validate the results.", + "bbox": [ + 88, + 429, + 480, + 549 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Compared Methods. We compare StableCodec with advanced image compression methods: (1) Traditional Codec H.266/VVC [10] by VTM-23.0 intra coding, (2) Neural Codec ELIC [20], (3) Generative Codec HiFiC [43], Text+Sketch [35], MS-ILLM [46], PerCo [11], EGIC [31], DiffEIC [40], TACO [34] and GLC [26]. Note that some methods do not release models for ultra-low bitrates, we either re-train or finetune existing weights to reach specific bitrates. For PerCo and GLC that do not have official codes, we use PerCo (SD) [32] as a substitute, and request for the results of GLC under the same evaluation approach1. We equip PerCo and Text+Sketch with the same inference strategy for a fair comparison on high-resolution images.", + "bbox": [ + 89, + 551, + 482, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 89, + 758, + 228, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.1. Rate-Distortion-Perception Performance", + "text_level": 1, + "bbox": [ + 89, + 781, + 419, + 796 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 6 presents the rate-perception and rate-distortion curves of various methods at ultra-low bitrates over CLIC 2020 Test and DIV2K Val. The proposed StableCodec outperforms all compared methods in terms of FID, KID, and DISTS. Specifically, StableCodec shows a significant im", + "bbox": [ + 89, + 800, + 483, + 876 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9a5813169cff0e242d3c808d3751acd31f1532f975ce5096cbd79b9093b58597.jpg", + "table_caption": [ + "Table 2. Encoding and decoding seconds averaged on Kodak." + ], + "table_footnote": [], + "table_body": "
TypeMethod#StepsEnc. TDec. T
VAE- basedELIC [20]-0.1550.245
MLIC++ [27]-0.3640.319
GAN- basedHiFiC [43]-0.1430.337
MS-ILLM [46]-0.1390.316
Diffusion- basedText+Sketch [35]25113.25233.560
PerCo [11]200.2873.742
DiffEIC [40]500.6767.423
StableCodec (Ours)10.1590.326
", + "bbox": [ + 517, + 452, + 903, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "provement over H.266/VVC and ELIC on all perceptual metrics. Compared to generative codec especially previous SOTA GLC [26], StableCodec demonstrates superiority and stability on FID and KID performance with well-preserved fidelity, and reaches extreme bitrates as low as 0.005 bpp. Although TACO [34] achieves the best LPIPS performance, it fails to ensure visual quality as FID, KID and DISTS scores are high. The PSNR and MS-SSIM results on CLIC and DIV2K are displayed in the supplementary.", + "bbox": [ + 511, + 627, + 906, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2. Qualitative Comparisons", + "text_level": 1, + "bbox": [ + 511, + 773, + 733, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We provide qualitative results among compared methods in Fig. 7. Notably, StableCodec generates more visually-aligned details and realistic textures at ultra-low bitrates, such as the teeth and murals shown in the first and third rows. In contrast, all other methods fail to produce high-realism results with well-preserved fidelity due to the severe bitrate restriction. For example, ELIC and MS-ILLM", + "bbox": [ + 511, + 795, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "1We acknowledge the authors of [26] for kindly providing their results.", + "bbox": [ + 107, + 886, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17385", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9a11a5ff3405640dec50ef82f96b99b8dfeb2aca84b8e7ea3a560138446d89fe.jpg", + "table_caption": [ + "Table 3. Ablation study on ${\\mathcal{E}}_{\\text{Aux }}$ and ${\\mathcal{D}}_{\\text{Aux }}$ ." + ], + "table_footnote": [], + "table_body": "
Model \nVariantsBD-rate↓ on the Rate-Y curves
PSNRMS-SSIMLPIPSDISTS
Base0000
+ EAux-20.63%-22.05%-23.04%-28.13%
+ EAux & DAux-23.96%-28.12%-40.66%-54.89%
", + "bbox": [ + 93, + 114, + 480, + 199 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/49f051536a4a38af15dc8c7a94f0125cb83ebc964047476cd01cca6f151b5ef2.jpg", + "table_caption": [ + "Table 4. Ablation study on LoRA ranks $\\left( {{\\mathcal{E}}_{\\mathrm{{SD}}}/{\\epsilon }_{\\mathrm{{SD}}}/{\\mathcal{D}}_{\\mathrm{{SD}}}}\\right)$ ." + ], + "table_footnote": [], + "table_body": "
LoRA RanksBD-rate↓ on the Rate-Y curves
PSNRMS-SSIMLPIPSDISTS
8/8/-0000
8/16/-1.21%-0.35%-4.67%-5.28%
16/16/--5.62%-3.27%-6.41%-12.78%
16/32/--7.12%-5.31%-13.43%-17.96%
32/32/--5.98%-5.69%-11.21%-16.70%
32/64/--5.15%-4.98%-12.95%-17.34%
16/32/4-21.77%-12.28%-8.27%-5.13%
", + "bbox": [ + 96, + 237, + 478, + 385 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "produce blurry reconstructions, while PerCo and DiffEIC deviate from the original images. EGIC and TACO exhibit noticeable artifacts, particularly in detailed areas.", + "bbox": [ + 89, + 409, + 482, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2.3. Computational Complexity", + "text_level": 1, + "bbox": [ + 89, + 463, + 326, + 478 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compare the practical complexity of StableCodec with representative schemes in Table 2 using a single RTX 3090 GPU. Among representative image compression schemes, Diffusion-based methods [11, 35, 40] typically suffer from a much longer decoding time compared to VAE-based [20, 27] or GAN-based [43, 46] competitors due to multi-step denoising. In contrast, StableCodec reaches comparable encoding and decoding speed against these methods exploiting one-step denoising, deep compression transforms and efficient entropy model, while achieving significantly better performance at ultra-low bitrates. Detailed runtime analysis is provided in the supplementary. In terms of memory, StableCodec consumes less than 9 GB VRAM with tiling techniques [29, 63], supporting arbitrary-resolution inference on a single GTX 1080Ti GPU.", + "bbox": [ + 88, + 482, + 482, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 723, + 254, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we conduct ablations to validate the proposed components. For reliable comparison, we compute the BD-rate [7] with Rate-Y curves on Kodak [16] using four target bitrates, where Y denotes specific metrics among PSNR, MS-SSIM, LPIPS, and DISTS.", + "bbox": [ + 89, + 750, + 482, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Auxiliary Encoder and Decoder. We begin by providing numerical results for the auxiliary coding branch in Table 3. Specifically, we construct a base model without $\\mathcal{E}_{\\mathrm{Aux}}$ and $\\mathcal{D}_{\\mathrm{Aux}}$ , and a variant with $\\mathcal{E}_{\\mathrm{Aux}}$ only. When $\\mathcal{E}_{\\mathrm{Aux}}$ is incorporated, more than $20\\%$ bits can be saved to reach the", + "bbox": [ + 89, + 824, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cfd78d7064f691ee806e17d851d081508028d0c2b8a2cbf1965c04256803c956.jpg", + "table_caption": [ + "Table 5. Ablation study on the training strategy components. Implicit bitrate pruning is abbreviated as IBP. The base one-stage objective only contains bitrate, MSE and LPIPS." + ], + "table_footnote": [], + "table_body": "
Training StrategyBD-rate↓ on the Rate-Y curves
IBPLadvLCIPPSNRMS-SSIMLPIPSDISTS
---0000
✓---24.12%-13.67%-21.37%-16.60%
✓✓-24.41%9.88%-36.18%-49.24%
✓✓✓13.29%5.70%-38.99%-52.95%
", + "bbox": [ + 514, + 142, + 911, + 242 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "same reconstruction quality. Subsequently, $\\mathcal{D}_{\\mathrm{Aux}}$ further improves the performance particularly in perceptual quality. As shown in Fig. 5, the purified $g_{s}$ latent provides better guidance for reconstruction consistency.", + "bbox": [ + 511, + 268, + 906, + 329 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "LoRA Ranks. We explore the impact of LoRA ranks in Table 4. Positive results are observed in both distortion and perception as the ranks increase to 16/32, which become our final choice. For larger ranks like 32 and 64, we observe performance degradation as the pre-trained priors may be corrupted. Besides, adding LoRA to the VAE decoder introduces a distortion-perception tradeoff, where PSNR and MS-SSIM improve at the cost of LPIPS and DISTS. To preserve perceptual quality, we leave the decoder unchanged.", + "bbox": [ + 511, + 330, + 908, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training Strategy. We perform ablations on our training strategy components in Table 5. We construct a simplified one-stage objective with bitrate, MSE and LPIPS, then progressively integrate the two-stage implicit bitrate pruning (IBP), adversarial training $\\mathcal{L}_{adv}$ and the CLIP distance term $\\mathcal{L}_{CLIP}$ . We find that IBP improves performance considerably by first adapting the latent codec into the T2I model with relaxed bitrate constraint. Furthermore, incorporating GAN introduces a significant distortion-perception tradeoff since we primarily focus on the perceptual quality. Besides, the CLIP distance alleviates the degradation in distortion and slightly improves perceptual quality.", + "bbox": [ + 511, + 467, + 908, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 662, + 633, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduce StableCodec, a novel diffusion-based extreme image compression approach that addresses key limitations of existing methods. By leveraging one-step diffusion in combination with Deep Compression Latent Codec and Dual-Branch Coding Structure, StableCodec achieves ultra-low bitrate compression with high realism, fidelity, and coding efficiency. Extensive experimental evaluations on benchmark datasets demonstrate the superiority of StableCodec in terms of FID, KID, and DISTS, even at extreme bitrates as low as 0.005 bpp, while enabling competitive speeds with mainstream transform coding methods. These results underscore the potential of diffusion models for practical image compression, particularly in real-time coding scenarios where bitrate is severely constrained.", + "bbox": [ + 511, + 688, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17386", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 126-135, 2017. 3, 6", + "[2] Eirikur Agustsson, Michael Tschannen, Fabian Mentzer, Radu Timofte, and Luc Van Gool. Generative adversarial networks for extreme learned image compression. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 221-231, 2019. 2, 3", + "[3] Eirikur Agustsson, David Minnen, George Toderici, and Fabian Mentzer. Multi-realism image compression with a conditional generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22324-22333, 2023. 2, 3", + "[4] Johannes Ballé, Valero Laparra, and Eero P Simoncelli. End-to-end optimized image compression. arXiv preprint arXiv:1611.01704, 2016. 2, 3", + "[5] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436, 2018. 2, 3, 4", + "[6] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 3, 6", + "[7] Gisle Bjontegaard. Calculation of average psnr differences between rd-curves. Tech. Rep. VCEG-m33, 2001. 8", + "[8] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6228-6237, 2018. 2, 3", + "[9] Yochai Blau and Tomer Michaeli. Rethinking lossy compression: The rate-distortion-perception tradeoff. In International Conference on Machine Learning, pages 675-685. PMLR, 2019. 2, 3", + "[10] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (vvc) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3736-3764, 2021. 1, 2, 3, 7", + "[11] Marlene Careil, Matthew J Muckley, Jakob Verbeek, and Stéphane Lathuilière. Towards image compression with perfect realism at ultra-low bitrates. In The Twelfth International Conference on Learning Representations, 2023. 1, 2, 3, 7, 8", + "[12] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7939-7948, 2020. 2, 3", + "[13] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023. 6", + "[14] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 3, 6, 7", + "[15] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Pro" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 3", + "[16] Rich Franzen. Kodak lossless true color image suite (photocd pcd0992). http://r0k.us/graphics/kodak/, 1993. 3, 5, 6, 8", + "[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 3", + "[18] Zongyu Guo, Zhizheng Zhang, Runsen Feng, and Zhibo Chen. Causal contextual prediction for learned image compression. IEEE Transactions on Circuits and Systems for Video Technology, 32(4):2329-2341, 2021. 2, 3, 4", + "[19] Dailan He, Yaoyan Zheng, Baocheng Sun, Yan Wang, and Hongwei Qin. Checkerboard context model for efficient learned image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14771-14780, 2021. 2, 3", + "[20] Dailan He, Ziming Yang, Weikun Peng, Rui Ma, Hongwei Qin, and Yan Wang. Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5718-5727, 2022. 1, 2, 3, 4, 5, 7, 8", + "[21] Dailan He, Ziming Yang, Hongjiu Yu, Tongda Xu, Jixiang Luo, Yuan Chen, Chenjian Gao, Xinjie Shi, Hongwei Qin, and Yan Wang. Po-elic: Perception-oriented efficient learned image coding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1764–1769, 2022. 2, 3", + "[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 3, 6", + "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3, 4", + "[24] Emiel Hoogeboom, Eirikur Agustsson, Fabian Mentzer, Luca Versari, George Toderici, and Lucas Theis. High-fidelity image compression with score-based generative models. arXiv preprint arXiv:2305.18231, 2023. 2, 3", + "[25] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 4", + "[26] Zhaoyang Jia, Jiahao Li, Bin Li, Houqiang Li, and Yan Lu. Generative latent coding for ultra-low bitrate image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26088-26098, 2024. 3, 4, 6, 7", + "[27] Wei Jiang and Ronggang Wang. Mlic++: Linear complexity multi-reference entropy modeling for learned image compression. In ICML 2023 Workshop Neural Compression: From Information Theory to Applications, 2023. 7, 8", + "[28] Wei Jiang, Jiayu Yang, Yongqi Zhai, Peirong Ning, Feng Gao, and Ronggang Wang. Mlic: Multi-reference entropy" + ], + "bbox": [ + 516, + 93, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17387", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "model for learned image compression. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7618-7627, 2023. 2, 3, 4", + "[29] Álvaro Barbero Jiménez. Mixture of diffusers for scene composition and high resolution image generation. arXiv preprint arXiv:2302.02412, 2023. 8", + "[30] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3", + "[31] Nikolai Körber, Eduard Kromer, Andreas Siebert, Sascha Hauke, Daniel Mueller-Gritschneder, and Björn Schuller. Egid: enhanced low-bit-rate generative image compression guided by semantic segmentation. In European Conference on Computer Vision, pages 202-220. Springer, 2024. 2, 3, 7", + "[32] Nikolai Körber, Eduard Kromer, Andreas Siebert, Sascha Hauke, Daniel Mueller-Gritschneder, and Björn Schuller. Perco (sd): Open perceptual compression. arXiv preprint arXiv:2409.20255, 2024. 7", + "[33] Nupur Kumari, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Ensembling off-the-shelf models for gan training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10651-10662, 2022. 6", + "[34] Hagyeong Lee, Minkyu Kim, Jun-Hyuk Kim, Seungeon Kim, Dokwan Oh, and Jaeho Lee. Neural image compression with text-guided encoding for both pixel-level and perceptual fidelity. arXiv preprint arXiv:2403.02944, 2024. 5, 7", + "[35] Eric Lei, Yigit Berkay Uslu, Hamed Hassani, and Shirin Saeedi Bidokhti. Text+ sketch: Image compression at ultra low rates. arXiv preprint arXiv:2307.01944, 2023. 2, 3, 7, 8", + "[36] Jiahao Li, Bin Li, and Yan Lu. Neural video compression with diverse contexts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22616-22626, 2023. 4", + "[37] Zhuoyuan Li, Yao Li, Chuanbo Tang, Li Li, Dong Liu, and Feng Wu. Uniformly accelerated motion model for inter prediction. In 2024 IEEE International Conference on Visual Communications and Image Processing (VCIP), pages 1-5. IEEE, 2024. 3", + "[38] Zhuoyuan Li, Junqi Liao, Chuanbo Tang, Haotian Zhang, Yuqi Li, Yifan Bian, Xihua Sheng, Xinmin Feng, Yao Li, Changsheng Gao, et al. Ustc-td: A test dataset and benchmark for image and video coding in 2020s. arXiv preprint arXiv:2409.08481, 2024. 1", + "[39] Zhuoyuan Li, Zikun Yuan, Li Li, Dong Liu, Xiaohu Tang, and Feng Wu. Object segmentation-assisted inter prediction for versatile video coding. IEEE Transactions on Broadcasting, 2024. 3", + "[40] Zhiyuan Li, Yanhui Zhou, Hao Wei, Chenyang Ge, and Jingwen Jiang. Towards extreme image compression with latent feature guidance and diffusion prior. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1, 2, 3, 4, 6, 7, 8", + "[41] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 6" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Jinming Liu, Heming Sun, and Jiro Katto. Learned image compression with mixed transformer-cnn architectures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14388-14397, 2023. 2, 3, 4", + "[43] Fabian Mentzer, George D Toderici, Michael Tschannen, and Eirikur Agustsson. High-fidelity generative image compression. Advances in Neural Information Processing Systems, 33:11913-11924, 2020. 2, 3, 7, 8", + "[44] David Minnen and Saurabh Singh. Channel-wise autoregressive entropy models for learned image compression. In 2020 IEEE International Conference on Image Processing (ICIP), pages 3339-3343. IEEE, 2020. 2, 3, 4, 5", + "[45] David Minnen, Johannes Balle, and George D Toderici. Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems, 31, 2018. 2, 3", + "[46] Matthew J Muckley, Alaaeldin El-Nouby, Karen Ullrich, Hervé Jégou, and Jakob Verbeek. Improving statistical fidelity for neural image compression with implicit local likelihood models. In International Conference on Machine Learning, pages 25426-25443. PMLR, 2023. 1, 2, 3, 6, 7, 8", + "[47] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6", + "[48] Gaurav Parmar, Taesung Park, Srinivasa Narasimhan, and Jun-Yan Zhu. One-step image translation with text-to-image models. arXiv preprint arXiv:2403.12036, 2024. 3", + "[49] Yichen Qian, Ming Lin, Xiuyu Sun, Zhiyu Tan, and Rong Jin. Entroformer: A transformer-based entropy model for learned image compression. arXiv preprint arXiv:2202.05492, 2022. 3, 4", + "[50] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5", + "[51] Lucas Relic, Roberto Azevedo, Markus Gross, and Christopher Schroers. Lossy image compression with foundation diffusion models. In European Conference on Computer Vision, pages 303-319. Springer, 2024. 2, 3", + "[52] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 4", + "[53] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 3", + "[54] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87-103. Springer, 2024. 2, 3", + "[55] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17388", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 3", + "[56] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3", + "[57] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023. 3", + "[58] Gary J Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. Overview of the high efficiency video coding (hevc) standard. IEEE Transactions on circuits and systems for video technology, 22(12):1649-1668, 2012. 3", + "[59] Lucas Theis, Tim Salimans, Matthew D Hoffman, and Fabian Mentzer. Lossy compression with gaussian diffusion. arXiv preprint arXiv:2206.08889, 2022. 2, 3", + "[60] George Toderici, Lucas Theis, Nick Johnston, Eirikur Agustsson, Fabian Mentzer, Johannes Ballé, Wenzhe Shi, and Radu Timofte. Clic 2020: Challenge on learned image compression, 2020, 2020. 3, 6", + "[61] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 3", + "[62] Gregory K Wallace. The jpeg still picture compression standard. Communications of the ACM, 34(4):30-44, 1991. 2, 3", + "[63] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, pages 1-21, 2024. 3, 8", + "[64] Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for image quality assessment. In The Thrity-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, pages 1398-1402. IEEE, 2003. 6", + "[65] Tongda Xu, Ziran Zhu, Dailan He, Yanghao Li, Lina Guo, Yuanyuan Wang, Zhe Wang, Hongwei Qin, Yan Wang, Jingjing Liu, et al. Idempotence and perceptual image compression. arXiv preprint arXiv:2401.08920, 2024. 2, 3", + "[66] Naifu Xue, Zhaoyang Jia, Jiahao Li, Bin Li, Yuan Zhang, and Yan Lu. Dlf: Extreme image compression with dual-generative latent fusion. arXiv preprint arXiv:2503.01428, 2025. 2", + "[67] Zeyu Yan, Fei Wen, Rendong Ying, Chao Ma, and Peilin Liu. On perceptual lossy compression: The cost of perceptual reconstruction and an optimal training framework. In International Conference on Machine Learning, pages 11682-11692. PMLR, 2021. 2, 3", + "[68] Zeyu Yan, Fei Wen, and Peilin Liu. Optimally controllable perceptual lossy compression. arXiv preprint arXiv:2206.10082, 2022. 2, 3", + "[69] Ruihan Yang and Stephan Mandt. Lossy image compression with conditional diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2, 3", + "[70] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. arXiv preprint arXiv:2405.14867, 2024. 3" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[71] Weihao Yu and Xinchao Wang. Mambaout: Do we really need mamba for vision? arXiv preprint arXiv:2405.07992, 2024. 4", + "[72] Weihao Yu, Pan Zhou, Shuicheng Yan, and Xinchao Wang. Inceptionnext: When inception meets convnext. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5672-5683, 2024. 4", + "[73] Aiping Zhang, Zongsheng Yue, Renjing Pei, Wenqi Ren, and Xiaochun Cao. Degradation-guided one-step image super-resolution with diffusion priors. arXiv preprint arXiv:2409.17058, 2024. 2, 3, 4, 5", + "[74] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5, 6" + ], + "bbox": [ + 514, + 92, + 903, + 316 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "17389", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/StableCodec_ Taming One-Step Diffusion for Extreme Image Compression/5269d8ed-d6d6-4961-a311-a0e841365f77_model.json b/2025/StableCodec_ Taming One-Step Diffusion for Extreme Image Compression/5269d8ed-d6d6-4961-a311-a0e841365f77_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a7ca52679dbdb69bf0bd532db418b77d441ef4a3 --- /dev/null +++ b/2025/StableCodec_ Taming One-Step Diffusion for Extreme Image Compression/5269d8ed-d6d6-4961-a311-a0e841365f77_model.json @@ -0,0 +1,2477 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.123, + 0.131, + 0.877, + 0.155 + ], + "angle": 0, + "content": "StableCodec: Taming One-Step Diffusion for Extreme Image Compression" + }, + { + "type": "text", + "bbox": [ + 0.284, + 0.182, + 0.714, + 0.2 + ], + "angle": 0, + "content": "Tianyu Zhang Xin Luo Li Li Dong Liu" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.2, + 0.804, + 0.234 + ], + "angle": 0, + "content": "MOE Key Laboratory of Brain-Inspired Intelligent Perception and Cognition, University of Science and Technology of China, Hefei 230093, China" + }, + { + "type": "text", + "bbox": [ + 0.198, + 0.237, + 0.795, + 0.252 + ], + "angle": 0, + "content": "{zhangtianyu, xinluo}@mail.ustc.edu.cn, {lil1, dongeliu}@ustc.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.269, + 0.905, + 0.74 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.742, + 0.908, + 0.786 + ], + "angle": 0, + "content": "Figure 1. Visual examples and comparisons when compressing a 4K-resolution image [38] at ultra-low bitrates. The proposed Stable-Coded produces more realistic and consistent details with fewer bits. In contrast, VVC [10], ELIC [20] and MS-ILLM [46] reconstructions are blurry, while PerCo [11] and DiffEIC [40] generate inconsistent details against the original images. Best viewed on screen for details." + }, + { + "type": "title", + "bbox": [ + 0.671, + 0.796, + 0.751, + 0.812 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.851, + 0.487, + 0.903 + ], + "angle": 0, + "content": "This work is supported by the Natural Science Foundation of China under Grant 62021001. We acknowledge the support of GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC. (Corresponding author: Dong Liu.)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Diffusion-based image compression has shown remarkable potential for achieving ultra-low bitrate coding (less than 0.05 bits per pixel) with high realism, by leveraging the generative priors of large pre-trained text-to-image diffu" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17379" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.455 + ], + "angle": 0, + "content": "sion models. However, current approaches require a large number of denoising steps at the decoder to generate realistic results under extreme bitrate constraints, limiting their application in real-time compression scenarios. Additionally, these methods often sacrifice reconstruction fidelity, as diffusion models typically fail to guarantee pixel-level consistency. To address these challenges, we introduce StableCodec, which enables one-step diffusion for high-fidelity and high-realism extreme image compression with improved coding efficiency. To achieve ultra-low bitrates, we first develop an efficient Deep Compression Latent Codec to transmit a noisy latent representation for a single-step denoising process. We then propose a Dual-Branch Coding Structure, consisting of a pair of auxiliary encoder and decoder, to enhance reconstruction fidelity. Furthermore, we adopt end-to-end optimization with joint bitrate and pixel-level constraints. Extensive experiments on the CLIC 2020, DIV2K, and Kodak datasets demonstrate that StableCodec outperforms existing methods in terms of FID, KID and DISTS by a significant margin, even at bitrates as low as 0.005 bits per pixel, while maintaining strong fidelity. Additionally, StableCodec achieves inference speeds comparable to mainstream transform coding schemes. Code is available at our project repository." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.483, + 0.223, + 0.499 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.484, + 0.705 + ], + "angle": 0, + "content": "Image compression is a foundational problem in signal processing. Driven by advances in digital imaging and the widespread use of social platforms, the volume of image data in modern multimedia has grown exponentially, placing increasing demands on the coding efficiency of image compression techniques. Over the past few decades, traditional CODECs such as JPEG [62] and H.266/VVC [10], along with emerging learning-based methods [4, 5, 12, 18, 19, 28, 42, 44, 45], have been widely adopted in real-world image compression applications. However, these methods are typically optimized for rate-distortion performance, and often produce unrealistic and blurry reconstructions, particularly under severe bitrate constraints, as shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.706, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To tackle this issue, generative image compression [2, 43] optimized for human perceptual performance has gained increasing attention. These methods are evaluated based on the rate-distortion-perception tradeoff [8, 9, 67, 68], and progressively demonstrate their advantages in producing visually appealing reconstructions at lower bitrates compared to traditional CODEs or common neural CODEs. A prominent research direction [3, 21, 31, 43, 43, 46, 66] involves integrating a discriminator into the transform coding pipeline [5, 20, 44], employing adversarial training to enhance the perceptual quality of reconstructions. Motivated by the impressive generative capabilities, more researchers [11, 24, 35, 40, 51, 59, 65, 69] have begun exploring the po" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.906, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.291, + 0.907, + 0.362 + ], + "angle": 0, + "content": "Figure 2. (Top) Illustration of our motivation. One-step diffusion can produce perceptually consistent results given severely corrupted images and a general prompt. (Bottom) Challenges in StableCodec: how to compress a noisy latent for one-step diffusion using ultra-low bitrates, and how to improve fidelity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.907, + 0.531 + ], + "angle": 0, + "content": "tential of diffusion models, particularly the generative priors in large pre-trained text-to-image (T2I) models, to compensate for severely distorted information at ultra-low bitrates while ensuring perceptually consistent generation. A recent study, PerCo [11], produces realistic results at an extreme bitrate as low as 0.003 bits per pixel (bpp) using a pretrained latent diffusion model (LDM), highlighting the potential of diffusion-based generative CODEs on image compression at more severe bitrates." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.909, + 0.868 + ], + "angle": 0, + "content": "Despite these promising advancements, existing methods face two primary limitations inherent to diffusion models. First, they require dozens of denoising steps at the decoder to produce results with sufficient perceptual quality. Second, the reconstructions often deviate from the original images, as diffusion models typically do not guarantee reconstruction consistency. To address the first challenge, we consider leveraging the generative priors in SD-Turbo [54], a distilled version of Stable Diffusion 2.1 [52] that enables real-time image synthesis in 1 to 4 denoising steps. Following [73], we demonstrate that SD-Turbo can produce perceptually consistent reconstructions with a single-step denoising process, even for severely corrupted inputs and a general positive prompt, as shown in Fig. 2. We thus pose an intuitive question: Can we compress a noisy latent representation of the original image, which can be effectively denoised in a single-step diffusion process, using an ultra-low bitrate? Building on these insights, we present StableCodec for extreme image compression, which integrates SD-Turbo with the proposed Deep Compression Latent Codec to compress noisy latents at ultra-low bitrates for a single-step diffusion process." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.909, + 0.902 + ], + "angle": 0, + "content": "In response to the second challenge, we introduce a Dual-Branch Coding Structure with a pair of auxiliary en" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17380" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.258 + ], + "angle": 0, + "content": "coder and decoder to further enhance reconstruction fidelity. Considering the limitations of the pre-trained VAE encoder on practical entropy coding and reconstruction consistency, we employ a rate-distortion-oriented auxiliary encoder to embed more entropy-aware semantic information for coding decisions. In parallel, we add an auxiliary decoder to perform structure apportionment during the decoding process, improving the generation guidance on the one-step denoising process for more consistent details. To enable end-to-end optimization, we design a two-stage training objective that jointly optimizes bitrate and pixel-level constraints." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.259, + 0.483, + 0.456 + ], + "angle": 0, + "content": "Benefit from these designs, StableCodec produces high-fidelity and high-realism reconstructions at ultra-low bitrates as low as 0.005 bpp. Extensive experiments on CLIC 2020 [60], DIV2K [1] and Kodak [16] demonstrate that StableCodec sets up a new state-of-the-art performance in terms of FID [22], KID [6], and DISTS [14], outperforming existing methods by significant margins. In terms of computational complexity, StableCodec offers much faster decoding speeds compared to other diffusion-based competitors, achieving inference times comparable to those of mainstream transform coding schemes. For practical deployment, StableCodec supports inference at arbitrary resolutions with memory consumption less than 9 GB." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.457, + 0.402, + 0.47 + ], + "angle": 0, + "content": "We summarize our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.472, + 0.483, + 0.533 + ], + "angle": 0, + "content": "- We present StableCodec, an extreme image codec integrating one-step diffusion and Deep Compression Latent Codec to achieve ultra-low bitrate compression with high realism, high fidelity and superior coding efficiency." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.533, + 0.483, + 0.593 + ], + "angle": 0, + "content": "- We develop Dual-Branch Coding Structure to improve reconstruction fidelity. A pair of auxiliary encoder and decoder is introduced for semantic enhancement and structure apportionment." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.593, + 0.483, + 0.669 + ], + "angle": 0, + "content": "- StableCodec obtains SOTA FID, KID and DISTS performance on CLIC 2020 and DIV2K dataset, significantly outperforms existing methods at bitrates as low as 0.005 bpp with well-preserved fidelity, and achieves comparable inference speeds with mainstream neural CODECs." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.472, + 0.483, + 0.669 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.685, + 0.233, + 0.7 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.711, + 0.368, + 0.727 + ], + "angle": 0, + "content": "2.1. Generative Image Compression" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Learning-based image compression has shown competitive potential compared to traditional methods [10, 37, 39, 58, 62], leveraging non-linear transforms and joint rate-distortion optimization. Balle et al. [4] introduced the first end-to-end learned image compression framework, which was subsequently enhanced with the hyperprior [5] and context model [45]. Building on this foundation, much work [12, 18, 19, 28, 42, 44, 49] has been devoted to improving both rate-distortion performance and model practicality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In practical scenarios, a key challenge is achieving extreme image compression at ultra-low bitrates while main" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.424 + ], + "angle": 0, + "content": "taining both fidelity and realism [2]. Traditional image codes optimized for rate-distortion often produce blurry reconstructions and noticeable artifacts. To address this, Mentzer et al. [43] introduced HiFiC and the concept of generative image compression, integrating GANs into codec optimization and evaluating performance in terms of the rate-distortion-perception tradeoff [8, 9, 67, 68]. Subsequent research can be broadly categorized into two main approaches. The first category [3, 21, 31, 43, 46] focuses on enhancing transform coding [5, 20, 44] for human perception by incorporating adversarial losses and optimized discriminator architectures, which typically can be extended to a wide range of bitrate and a flexible decoding control between fidelity and realism [3, 31]. The second category [11, 24, 35, 40, 51, 59, 65, 69] leverages diffusion models for generative image compression. Although these methods show promise for ultra-low bitrate compression, they are often constrained by reconstruction fidelity and inference efficiency due to the multi-step denoising process. Recently, GLC [26] introduced transform coding in the generative latent space of VQ-VAE [15, 61], achieving more visually appealing results at ultra-low bitrates." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.436, + 0.886, + 0.452 + ], + "angle": 0, + "content": "2.2. Generative Models and Few-Step Diffusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.905, + 0.626 + ], + "angle": 0, + "content": "Generative models play a crucial role in image generation. While many architectures, such as VAEs [30] and GANs [17], have been explored, diffusion models [55] have emerged as a powerful alternative, achieving state-of-the-art synthesis quality. Inspired by non-equilibrium statistical physics [55], diffusion models learn to reverse a noise perturbation process through a Markovian framework. Recent advancements, such as DDPM [23], DDIM [56], and LDM [52], have significantly reduced computational complexity and improved image synthesis quality, making diffusion-based approaches a dominant force in generative modeling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.626, + 0.905, + 0.762 + ], + "angle": 0, + "content": "To address the inefficiency of iterative denoising, several approaches [53, 54, 57, 70] aim to reduce the number of denoising steps while maintaining generation quality. These methods train diffusion models to approximate the full denoising trajectory in a single or a few steps, significantly improving inference efficiency. Notably, SD/SDXL-Turbo [54] demonstrates image generation in 1 to 4 steps with near-parity quality compared to multi-step models, making real-time diffusion-based applications [48, 63, 73] feasible." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.777, + 0.605, + 0.793 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.624, + 0.817 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In this section, we introduce the overall framework of the proposed StableCodec, built upon SD-Turbo [54] with a VAE encoder \\(\\mathcal{E}_{\\mathrm{SD}}\\), a VAE decoder \\(\\mathcal{D}_{\\mathrm{SD}}\\) and a denoising Unet \\(\\epsilon_{\\mathrm{SD}}\\). As shown in Fig. 3, we incorporate a Deep Compression Latent Codec to perform extreme transform" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17381" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.09, + 0.087, + 0.912, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.36, + 0.908, + 0.417 + ], + "angle": 0, + "content": "Figure 3. The framework of StableCodec. We incorporate the proposed Deep Compression Latent Codec to transmit a noisy latent \\( l_{T} \\) for one-step denoising, where \\( 64 \\times \\) denotes a spatial compression ratio of 64. To adjust the latent resolution, we deploy DownSample block and Conv \\( 3 \\times 3 \\) as adapters after the VAE encoder \\( \\mathcal{E}_{\\mathrm{SD}} \\) and auxiliary encoder \\( \\mathcal{E}_{\\mathrm{Aux}} \\), respectively. We use a general prompt in both training and inference. The auxiliary decoder \\( \\mathcal{D}_{\\mathrm{Aux}} \\) shares a similar structure with \\( g_{s} \\). More details on networks are provided in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.442, + 0.482, + 0.593 + ], + "angle": 0, + "content": "coding in the VAE latent space. To adapt SD-Turbo for image compression, we integrate LoRA [25] into \\(\\mathcal{E}_{\\mathrm{SD}}\\) and \\(\\epsilon_{\\mathrm{SD}}\\), while keeping \\(\\mathcal{D}_{\\mathrm{SD}}\\) unchanged to preserve the generative priors [73]. Additionally, we introduce a Dual-Branch Coding Structure to enhance reconstruction fidelity, utilizing an auxiliary encoder \\(\\mathcal{E}_{\\mathrm{Aux}}\\) to embed rich semantic information and an auxiliary decoder \\(\\mathcal{D}_{\\mathrm{Aux}}\\) to perform structure apportionment. Finally, we optimize StableCodec end-to-end with joint bitrate and pixel-level constraints, achieving high-fidelity and high-realism extreme image compression." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.605, + 0.379, + 0.621 + ], + "angle": 0, + "content": "3.2. Deep Compression Latent Codec" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.484, + 0.809 + ], + "angle": 0, + "content": "We design our latent codec using the extreme analysis transform \\( g_{a} \\), extreme synthesis transform \\( g_{s} \\) and a 4-step autoregressive entropy model. To reach ultra-low bitrates, we employ deep compression transform networks for both \\( g_{a} \\) and \\( g_{s} \\). Specifically, \\( \\mathcal{E}_{\\mathrm{SD}} \\) and \\( \\mathcal{D}_{\\mathrm{SD}} \\) provide a latent space with a spatial compression ratio of 8 (abbreviated as \\( 8\\times \\)). Unlike mainstream schemes [18, 20, 26, 28, 40, 42, 49] that perform entropy coding at \\( 16\\times \\), we further downsample and apply entropy coding for \\( \\hat{y} \\) at \\( 64\\times \\) and the hyperprior [5] at \\( 256\\times \\). Consequently, we use \\( g_{s} \\) to restore the spatial compression ratio to \\( 8\\times \\) for \\( \\epsilon_{\\mathrm{SD}} \\) and \\( \\mathcal{D}_{\\mathrm{SD}} \\). The entire coding process can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.823, + 0.483, + 0.839 + ], + "angle": 0, + "content": "\\[\nl = \\operatorname {c o n c a t} \\left[ \\mathcal {E} _ {\\mathrm {S D}} (x), \\mathcal {E} _ {\\text {A u x}} (x) \\right] \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.841, + 0.483, + 0.858 + ], + "angle": 0, + "content": "\\[\ny = g _ {a} (l), \\hat {y} = Q (y), l _ {T} = g _ {s} (\\hat {y}) \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.862, + 0.482, + 0.879 + ], + "angle": 0, + "content": "\\[\nl _ {0} = \\left[ l _ {T} - \\sqrt {1 - \\bar {\\alpha} _ {T}} \\cdot \\epsilon_ {\\mathrm {S D}} \\left(l _ {T}, T\\right) \\right] / \\sqrt {\\bar {\\alpha} _ {T}} \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.881, + 0.483, + 0.897 + ], + "angle": 0, + "content": "\\[\n\\hat {x} = \\mathcal {D} _ {\\mathrm {S D}} \\left(l _ {0} + \\mathcal {D} _ {\\mathrm {A u x}} (\\hat {y})\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.442, + 0.907, + 0.61 + ], + "angle": 0, + "content": "In Eq. (1), we first obtain an intermediate latent \\(l\\) from the input image \\(x\\) through \\(\\mathcal{E}_{\\mathrm{SD}}\\) and \\(\\mathcal{E}_{\\mathrm{Aux}}\\). Eq. (2) is the latent-space transform coding process to produce a noisy latent \\(l_{T}\\) using ultra-low bitrates. Eq. (3) displays the one-step denoising process with the noise schedule \\(\\{\\bar{\\alpha}_t\\}\\) [23] in the \\(T\\)-th timestep. Finally, in Eq. (4), the reconstruction \\(\\hat{x}\\) is obtained from \\(l_{0}\\) and \\(\\hat{y}\\) using \\(\\mathcal{D}_{\\mathrm{SD}}\\) and \\(\\mathcal{D}_{\\mathrm{Aux}}\\). To balance performance and coding latency, we construct efficient \\(g_{a}\\) and \\(g_{s}\\) with InceptionNeXt [72] and GatedCNN [71], and build a 4-step antoregressive entropy model with quadtree partition [36] and latent residual prediction [44]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.621, + 0.792, + 0.637 + ], + "angle": 0, + "content": "3.3. Dual-Branch Coding Structure" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.737 + ], + "angle": 0, + "content": "While deploying the proposed latent codec with LoRA enables image compression with SD-Turbo at ultra-low bitrates, the reconstruction fidelity is limited. In this section, we analyze the reasons and introduce Dual-Branch Coding Structure with a pair of auxiliary encoder and decoder, \\(\\mathcal{E}_{\\mathrm{Aux}}\\) and \\(\\mathcal{D}_{\\mathrm{Aux}}\\), to further enhance compression performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.745, + 0.837, + 0.76 + ], + "angle": 0, + "content": "3.3.1. Entropy-Aware Semantic Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We observed that the VAE in SD-Turbo has several limitations when reconstruction fidelity and practical coding are required. As noted in [52] and Table 1, this VAE is pretrained for perceptual compression, which does not preserve pixel-level fidelity as well as a rate-distortion-oriented autoencoder from a typical neural codec [20]. Additionally, while this VAE provides a compressed representation of the original image, it is still in floating-point format and not optimized for practical entropy coding, making it less suitable" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17382" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.486, + 0.131 + ], + "angle": 0, + "content": "Table 1. Reconstruction quality of VAEs on Kodak [16]. The pre-trained VAE in SD performs perceptual compression, while the one in ELIC [20] preserves more pixel-level information." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.134, + 0.472, + 0.189 + ], + "angle": 0, + "content": "
VAEPSNR↑MS-SSIM↑LPIPS↓DISTS↓
SD26.650.93180.07260.0415
ELIC40.400.99610.05550.0707
" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.201, + 0.486, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.352, + 0.483, + 0.38 + ], + "angle": 0, + "content": "Figure 4. Top-energy channels learned from different encoders. \\(\\mathcal{E}_{\\mathrm{Aux}}\\) embeds more pixel-level semantic information into latents." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.407, + 0.445, + 0.421 + ], + "angle": 0, + "content": "for further latent-space ultra-low bitrate compression." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.422, + 0.484, + 0.588 + ], + "angle": 0, + "content": "Building on these insights, we introduce the analysis transform of a pre-trained high-bitrate ELIC model [20] to serve as an auxiliary encoder \\(\\mathcal{E}_{\\mathrm{Aux}}\\). \\(\\mathcal{E}_{\\mathrm{Aux}}\\) remains frozen, and provides rich, entropy-aware semantic information of the input images. To combine the latents from different encoders, we introduce trainable adapters that align the latent resolutions to a spatial compression ratio of 16, followed by channel-wise concatenation. During optimization, various pieces of information are learned from different encoders as shown in Fig. 4, where more pixel-level semantic information is embedded into the latent codec through \\(\\mathcal{E}_{\\mathrm{Aux}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.596, + 0.317, + 0.611 + ], + "angle": 0, + "content": "3.3.2. Structure Apportionment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.719 + ], + "angle": 0, + "content": "The reconstruction quality of StableCodec heavily depends on how the denoising Unet is conditioned. Since a fixed prompt is used for both training and inference, the one-step denoising process is primarily guided by \\( l_{T} \\), which is produced by \\( g_{s} \\). This places high demands on the capability of \\( g_{s} \\), resulting in unsatisfactory denoising guidance, as reflected in the reconstructions shown in Fig. 5 (b)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To alleviate the decoding burden on \\( g_{s} \\), we introduce an auxiliary decoder \\( \\mathcal{D}_{\\mathrm{Aux}} \\) to perform an additional decoding branch from \\( \\hat{y} \\), bypassing the Unet. This design is motivated by the observation that an extremely compressed bitstream contains mainly the basic structure of images. We distribute and decode these components directly from the bitstream using the auxiliary branch, allowing \\( g_{s} \\) to focus primarily on producing guidance to generate realistic details. Fig. 5 compares StableCodec that is trained with or without \\( \\mathcal{D}_{\\mathrm{Aux}} \\). When trained without \\( \\mathcal{D}_{\\mathrm{Aux}} \\), \\( g_{s} \\) produces all types of information as it is the only decoding branch. When trained with \\( \\mathcal{D}_{\\mathrm{Aux}} \\), structural information is routed" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.088, + 0.911, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.312, + 0.864, + 0.324 + ], + "angle": 0, + "content": "(a) Latents from different decoders, trained w. or w/o \\(\\mathcal{D}_{\\mathrm{Aux}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.326, + 0.895, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.575, + 0.399, + 0.854, + 0.41 + ], + "angle": 0, + "content": "(b) Reconstruction comparison, trained w. or w/o \\(\\mathcal{D}_{\\mathrm{Aux}}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.531, + 0.412, + 0.888, + 0.426 + ], + "angle": 0, + "content": "Figure 5. Impact of \\(\\mathcal{D}_{\\mathrm{Aux}}\\) on latents and reconstructions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.452, + 0.906, + 0.514 + ], + "angle": 0, + "content": "through \\(\\mathcal{D}_{\\mathrm{Aux}}\\), while the energy in \\(g_{s}\\) latents drops significantly. Meanwhile, more semantically aligned details are reconstructed in Fig. 5 (b), suggesting \\(g_{s}\\) now provides better high-frequency guidance for the denoising Unet \\(\\epsilon_{\\mathrm{SD}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.522, + 0.792, + 0.539 + ], + "angle": 0, + "content": "3.4. End-to-End Training Objective" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.544, + 0.906, + 0.618 + ], + "angle": 0, + "content": "We adopt end-to-end optimization with joint bitrate and pixel-level restrictions to train StableCodec. Given the original image \\( x \\), the quantized latent \\( \\hat{y} \\) and the reconstructed image \\( \\hat{x} \\), we construct our training objective based on the standard rate-distortion loss:" + }, + { + "type": "equation", + "bbox": [ + 0.649, + 0.632, + 0.905, + 0.648 + ], + "angle": 0, + "content": "\\[\n\\lambda \\mathcal {R} (\\hat {y}) + \\mathcal {D} (x, \\hat {x}) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.904, + 0.689 + ], + "angle": 0, + "content": "where the bitrate \\(\\mathcal{R}\\) and pixel-level distortion \\(\\mathcal{D}\\) are balanced by the Lagrange multiplier \\(\\lambda\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.906, + 0.81 + ], + "angle": 0, + "content": "Inspired by [20, 44], we train StableCodec with a 2-stage implicit bitrate pruning (IBP) strategy. We first train a base model using a smaller \\(\\lambda_{base}\\), adapting the latent codec into SD-Turbo under a relaxed bitrate constraint, and warming up with a more expressive transform. In the second stage, we finetune the shared base model with larger \\(\\lambda_{target}\\) to reach ultra-low target bitrates. IBP facilitates efficient and stable training, resulting in improved performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "The distortion term \\(\\mathcal{D}\\) includes MSE, LPIPS (using VGG features) [74] and a CLIP [50] distance \\(\\mathcal{L}_{CLIP}\\), for which we compute the L2-distance between the CLIP embeddings of \\(x\\) and \\(\\hat{x}\\). Note that this term is a simplified version from [34], and we find it beneficial for reconstruction at ultralow bitrates. Additionally, we follow [73] and incorporate" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17383" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.193, + 0.094, + 0.803, + 0.133 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.138, + 0.293, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.138, + 0.495, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.138, + 0.698, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.138, + 0.901, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.261, + 0.294, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.261, + 0.495, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.261, + 0.698, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.261, + 0.901, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.382, + 0.294, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.382, + 0.496, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.382, + 0.699, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.382, + 0.901, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.509, + 0.9, + 0.523 + ], + "angle": 0, + "content": "Figure 6. Rate-distortion and rate-perception curve comparisons of different methods on the CLIC 2020, DIV2K and Kodak dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.55, + 0.484, + 0.626 + ], + "angle": 0, + "content": "a GAN loss \\(\\mathcal{L}_{adv}\\) to narrow the distribution gap between \\(x\\) and \\(\\hat{x}\\). We use DINOv2 [47] with registers [13] as the discriminator backbone [33]. To ensure stable training, we incorporate the GAN only in the second training stage. The full objective can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.109, + 0.638, + 0.483, + 0.662 + ], + "angle": 0, + "content": "\\[\n\\text {S t a g e I :} \\underset {\\theta} {\\arg \\min } \\lambda_ {\\text {b a s e}} \\mathcal {R} (\\hat {y}) + \\mathcal {D} (x, \\hat {x}) \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.665, + 0.483, + 0.69 + ], + "angle": 0, + "content": "\\[\n\\text {S t a g e I I :} \\underset {\\theta} {\\arg \\min } \\lambda_ {\\text {t a r g e t}} \\mathcal {R} (\\hat {y}) + \\mathcal {D} (x, \\hat {x}) + \\beta \\mathcal {L} _ {\\text {a d v}} \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.706, + 0.484, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {D} (x, \\hat {x}) = d _ {1} M S E (x, \\hat {x}) + d _ {2} L P I P S (x, \\hat {x}) \\tag {8} \\\\ + d _ {3} \\mathcal {L} _ {C L I P} (x, \\hat {x}) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.749, + 0.483, + 0.78 + ], + "angle": 0, + "content": "where \\(\\theta\\) represents all trainable parameters in StableCodec (Fig. 3), \\(d_{1}, d_{2}, d_{3}\\) and \\(\\beta\\) are balancing weights." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.794, + 0.224, + 0.811 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.819, + 0.25, + 0.835 + ], + "angle": 0, + "content": "4.1. Implementation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Training Details. We use the training set of DF2K [41] and CLIC 2020 Professional [60] to train StableCodec. During training, we use \\(512 \\times 512\\) patches with a batch size of 8. The first training stage takes over 100k iterations with a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.905, + 0.656 + ], + "angle": 0, + "content": "learning rate of \\( 1e^{-4} \\) and a \\( \\lambda_{base} \\) of 0.5 (about 0.05bpp). In the second stage, we finetune the base model for another 20k iterations with GAN incorporated and \\( \\lambda_{target} \\in \\{2,3,4,6,8,12,16,24,32\\} \\), while the learning rate undergoes \\( 5e^{-5}, 2e^{-5}, 1e^{-5} \\) and \\( 1e^{-6} \\) for 5k iterations each. We set \\( d_1, d_2, d_3 \\) and \\( \\beta \\) to 2, 1, 0.1 and 0.1, respectively. All models are trained using 2 RTX 3090 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.906, + 0.779 + ], + "angle": 0, + "content": "Test Data. We evaluate StableCodec on the test set of CLIC 2020 Professional [60] (CLIC 2020 Test), the validation set of DIV2K [1] (DIV2K Val) and Kodak [16] following [26, 40]. CLIC 2020 Test and DIV2K Val contain 428 and 100 high-quality 2K-resolution natural images, respectively, while Kodak contains 24 natural images with a smaller resolution of \\(768 \\times 512\\). We evaluate all images with the original resolution as detailed in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Evaluation Metrics. We employ established metrics to assess the rate-distortion-perception performance of StableCodec. Concretely, we measure bitrate by bits per pixel (bpp), and evaluate perceptual quality using FID [22], KID [6], DISTS [14] and LPIPS [74] (using AlexNet features by default). Meanwhile, we use PSNR and MS-SSIM [64] to measure the reconstruction fidelity. We follow [26, 40, 46] to calculate FID and KID on \\(256 \\times 256\\) patches, and neglect" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17384" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.905, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.196, + 0.389, + 0.799, + 0.404 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparisons of different methods on Kodak. Best viewed on screen for details." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.43, + 0.482, + 0.55 + ], + "angle": 0, + "content": "the results on Kodak as it is too small for calculating. Note that pixel-level distortion metrics like LPIPS, PSNR and MS-SSIM have strong limitations when evaluating image compression at ultra-low bitrates [11, 14, 26, 35]. Therefore, for StableCodec, we focus primarily on FID, KID, and DISTS, which offer a more accurate assessment of quality in severely compressed images. We also provide user study in the supplementary to visually validate the results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.552, + 0.483, + 0.749 + ], + "angle": 0, + "content": "Compared Methods. We compare StableCodec with advanced image compression methods: (1) Traditional Codec H.266/VVC [10] by VTM-23.0 intra coding, (2) Neural Codec ELIC [20], (3) Generative Codec HiFiC [43], Text+Sketch [35], MS-ILLM [46], PerCo [11], EGIC [31], DiffEIC [40], TACO [34] and GLC [26]. Note that some methods do not release models for ultra-low bitrates, we either re-train or finetune existing weights to reach specific bitrates. For PerCo and GLC that do not have official codes, we use PerCo (SD) [32] as a substitute, and request for the results of GLC under the same evaluation approach1. We equip PerCo and Text+Sketch with the same inference strategy for a fair comparison on high-resolution images." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.759, + 0.229, + 0.773 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.782, + 0.42, + 0.797 + ], + "angle": 0, + "content": "4.2.1. Rate-Distortion-Perception Performance" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.801, + 0.484, + 0.877 + ], + "angle": 0, + "content": "Fig. 6 presents the rate-perception and rate-distortion curves of various methods at ultra-low bitrates over CLIC 2020 Test and DIV2K Val. The proposed StableCodec outperforms all compared methods in terms of FID, KID, and DISTS. Specifically, StableCodec shows a significant im" + }, + { + "type": "table_caption", + "bbox": [ + 0.521, + 0.428, + 0.897, + 0.442 + ], + "angle": 0, + "content": "Table 2. Encoding and decoding seconds averaged on Kodak." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.453, + 0.905, + 0.6 + ], + "angle": 0, + "content": "
TypeMethod#StepsEnc. TDec. T
VAE- basedELIC [20]-0.1550.245
MLIC++ [27]-0.3640.319
GAN- basedHiFiC [43]-0.1430.337
MS-ILLM [46]-0.1390.316
Diffusion- basedText+Sketch [35]25113.25233.560
PerCo [11]200.2873.742
DiffEIC [40]500.6767.423
StableCodec (Ours)10.1590.326
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.907, + 0.765 + ], + "angle": 0, + "content": "provement over H.266/VVC and ELIC on all perceptual metrics. Compared to generative codec especially previous SOTA GLC [26], StableCodec demonstrates superiority and stability on FID and KID performance with well-preserved fidelity, and reaches extreme bitrates as low as 0.005 bpp. Although TACO [34] achieves the best LPIPS performance, it fails to ensure visual quality as FID, KID and DISTS scores are high. The PSNR and MS-SSIM results on CLIC and DIV2K are displayed in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.775, + 0.734, + 0.79 + ], + "angle": 0, + "content": "4.2.2. Qualitative Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We provide qualitative results among compared methods in Fig. 7. Notably, StableCodec generates more visually-aligned details and realistic textures at ultra-low bitrates, such as the teeth and murals shown in the first and third rows. In contrast, all other methods fail to produce high-realism results with well-preserved fidelity due to the severe bitrate restriction. For example, ELIC and MS-ILLM" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1We acknowledge the authors of [26] for kindly providing their results." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17385" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.152, + 0.09, + 0.424, + 0.104 + ], + "angle": 0, + "content": "Table 3. Ablation study on \\( {\\mathcal{E}}_{\\text{Aux }} \\) and \\( {\\mathcal{D}}_{\\text{Aux }} \\) ." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.115, + 0.482, + 0.2 + ], + "angle": 0, + "content": "
Model \nVariantsBD-rate↓ on the Rate-Y curves
PSNRMS-SSIMLPIPSDISTS
Base0000
+ EAux-20.63%-22.05%-23.04%-28.13%
+ EAux & DAux-23.96%-28.12%-40.66%-54.89%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.111, + 0.213, + 0.462, + 0.227 + ], + "angle": 0, + "content": "Table 4. Ablation study on LoRA ranks \\( \\left( {{\\mathcal{E}}_{\\mathrm{{SD}}}/{\\epsilon }_{\\mathrm{{SD}}}/{\\mathcal{D}}_{\\mathrm{{SD}}}}\\right) \\) ." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.238, + 0.48, + 0.386 + ], + "angle": 0, + "content": "
LoRA RanksBD-rate↓ on the Rate-Y curves
PSNRMS-SSIMLPIPSDISTS
8/8/-0000
8/16/-1.21%-0.35%-4.67%-5.28%
16/16/--5.62%-3.27%-6.41%-12.78%
16/32/--7.12%-5.31%-13.43%-17.96%
32/32/--5.98%-5.69%-11.21%-16.70%
32/64/--5.15%-4.98%-12.95%-17.34%
16/32/4-21.77%-12.28%-8.27%-5.13%
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.41, + 0.483, + 0.456 + ], + "angle": 0, + "content": "produce blurry reconstructions, while PerCo and DiffEIC deviate from the original images. EGIC and TACO exhibit noticeable artifacts, particularly in detailed areas." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.464, + 0.327, + 0.479 + ], + "angle": 0, + "content": "4.2.3. Computational Complexity" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.483, + 0.483, + 0.71 + ], + "angle": 0, + "content": "We compare the practical complexity of StableCodec with representative schemes in Table 2 using a single RTX 3090 GPU. Among representative image compression schemes, Diffusion-based methods [11, 35, 40] typically suffer from a much longer decoding time compared to VAE-based [20, 27] or GAN-based [43, 46] competitors due to multi-step denoising. In contrast, StableCodec reaches comparable encoding and decoding speed against these methods exploiting one-step denoising, deep compression transforms and efficient entropy model, while achieving significantly better performance at ultra-low bitrates. Detailed runtime analysis is provided in the supplementary. In terms of memory, StableCodec consumes less than 9 GB VRAM with tiling techniques [29, 63], supporting arbitrary-resolution inference on a single GTX 1080Ti GPU." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.724, + 0.255, + 0.74 + ], + "angle": 0, + "content": "5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.825 + ], + "angle": 0, + "content": "In this section, we conduct ablations to validate the proposed components. For reliable comparison, we compute the BD-rate [7] with Rate-Y curves on Kodak [16] using four target bitrates, where Y denotes specific metrics among PSNR, MS-SSIM, LPIPS, and DISTS." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Auxiliary Encoder and Decoder. We begin by providing numerical results for the auxiliary coding branch in Table 3. Specifically, we construct a base model without \\(\\mathcal{E}_{\\mathrm{Aux}}\\) and \\(\\mathcal{D}_{\\mathrm{Aux}}\\), and a variant with \\(\\mathcal{E}_{\\mathrm{Aux}}\\) only. When \\(\\mathcal{E}_{\\mathrm{Aux}}\\) is incorporated, more than \\(20\\%\\) bits can be saved to reach the" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 5. Ablation study on the training strategy components. Implicit bitrate pruning is abbreviated as IBP. The base one-stage objective only contains bitrate, MSE and LPIPS." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.143, + 0.912, + 0.243 + ], + "angle": 0, + "content": "
Training StrategyBD-rate↓ on the Rate-Y curves
IBPLadvLCIPPSNRMS-SSIMLPIPSDISTS
---0000
✓---24.12%-13.67%-21.37%-16.60%
✓✓-24.41%9.88%-36.18%-49.24%
✓✓✓13.29%5.70%-38.99%-52.95%
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.27, + 0.907, + 0.33 + ], + "angle": 0, + "content": "same reconstruction quality. Subsequently, \\( \\mathcal{D}_{\\mathrm{Aux}} \\) further improves the performance particularly in perceptual quality. As shown in Fig. 5, the purified \\( g_{s} \\) latent provides better guidance for reconstruction consistency." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.331, + 0.909, + 0.467 + ], + "angle": 0, + "content": "LoRA Ranks. We explore the impact of LoRA ranks in Table 4. Positive results are observed in both distortion and perception as the ranks increase to 16/32, which become our final choice. For larger ranks like 32 and 64, we observe performance degradation as the pre-trained priors may be corrupted. Besides, adding LoRA to the VAE decoder introduces a distortion-perception tradeoff, where PSNR and MS-SSIM improve at the cost of LPIPS and DISTS. To preserve perceptual quality, we leave the decoder unchanged." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.468, + 0.909, + 0.649 + ], + "angle": 0, + "content": "Training Strategy. We perform ablations on our training strategy components in Table 5. We construct a simplified one-stage objective with bitrate, MSE and LPIPS, then progressively integrate the two-stage implicit bitrate pruning (IBP), adversarial training \\(\\mathcal{L}_{adv}\\) and the CLIP distance term \\(\\mathcal{L}_{CLIP}\\). We find that IBP improves performance considerably by first adapting the latent codec into the T2I model with relaxed bitrate constraint. Furthermore, incorporating GAN introduces a significant distortion-perception tradeoff since we primarily focus on the perceptual quality. Besides, the CLIP distance alleviates the degradation in distortion and slightly improves perceptual quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.664, + 0.634, + 0.68 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this work, we introduce StableCodec, a novel diffusion-based extreme image compression approach that addresses key limitations of existing methods. By leveraging one-step diffusion in combination with Deep Compression Latent Codec and Dual-Branch Coding Structure, StableCodec achieves ultra-low bitrate compression with high realism, fidelity, and coding efficiency. Extensive experimental evaluations on benchmark datasets demonstrate the superiority of StableCodec in terms of FID, KID, and DISTS, even at extreme bitrates as low as 0.005 bpp, while enabling competitive speeds with mainstream transform coding methods. These results underscore the potential of diffusion models for practical image compression, particularly in real-time coding scenarios where bitrate is severely constrained." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17386" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 126-135, 2017. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.242 + ], + "angle": 0, + "content": "[2] Eirikur Agustsson, Michael Tschannen, Fabian Mentzer, Radu Timofte, and Luc Van Gool. Generative adversarial networks for extreme learned image compression. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 221-231, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.242, + 0.483, + 0.311 + ], + "angle": 0, + "content": "[3] Eirikur Agustsson, David Minnen, George Toderici, and Fabian Mentzer. Multi-realism image compression with a conditional generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22324-22333, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.483, + 0.353 + ], + "angle": 0, + "content": "[4] Johannes Ballé, Valero Laparra, and Eero P Simoncelli. End-to-end optimized image compression. arXiv preprint arXiv:1611.01704, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.483, + 0.409 + ], + "angle": 0, + "content": "[5] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436, 2018. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.41, + 0.483, + 0.451 + ], + "angle": 0, + "content": "[6] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.452, + 0.483, + 0.479 + ], + "angle": 0, + "content": "[7] Gisle Bjontegaard. Calculation of average psnr differences between rd-curves. Tech. Rep. VCEG-m33, 2001. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.483, + 0.522 + ], + "angle": 0, + "content": "[8] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6228-6237, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.523, + 0.483, + 0.578 + ], + "angle": 0, + "content": "[9] Yochai Blau and Tomer Michaeli. Rethinking lossy compression: The rate-distortion-perception tradeoff. In International Conference on Machine Learning, pages 675-685. PMLR, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.58, + 0.483, + 0.648 + ], + "angle": 0, + "content": "[10] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (vvc) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3736-3764, 2021. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.649, + 0.483, + 0.704 + ], + "angle": 0, + "content": "[11] Marlene Careil, Matthew J Muckley, Jakob Verbeek, and Stéphane Lathuilière. Towards image compression with perfect realism at ultra-low bitrates. In The Twelfth International Conference on Learning Representations, 2023. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[12] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7939-7948, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[13] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[14] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[15] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Pro" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.094, + 0.905, + 0.121 + ], + "angle": 0, + "content": "ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.164 + ], + "angle": 0, + "content": "[16] Rich Franzen. Kodak lossless true color image suite (photocd pcd0992). http://r0k.us/graphics/kodak/, 1993. 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.221 + ], + "angle": 0, + "content": "[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.277 + ], + "angle": 0, + "content": "[18] Zongyu Guo, Zhizheng Zhang, Runsen Feng, and Zhibo Chen. Causal contextual prediction for learned image compression. IEEE Transactions on Circuits and Systems for Video Technology, 32(4):2329-2341, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.279, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[19] Dailan He, Yaoyan Zheng, Baocheng Sun, Yan Wang, and Hongwei Qin. Checkerboard context model for efficient learned image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14771-14780, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.35, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[20] Dailan He, Ziming Yang, Weikun Peng, Rui Ma, Hongwei Qin, and Yan Wang. Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5718-5727, 2022. 1, 2, 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[21] Dailan He, Ziming Yang, Hongjiu Yu, Tongda Xu, Jixiang Luo, Yuan Chen, Chenjian Gao, Xinjie Shi, Hongwei Qin, and Yan Wang. Po-elic: Perception-oriented efficient learned image coding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1764–1769, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[24] Emiel Hoogeboom, Eirikur Agustsson, Fabian Mentzer, Luca Versari, George Toderici, and Lucas Theis. High-fidelity image compression with score-based generative models. arXiv preprint arXiv:2305.18231, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[25] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[26] Zhaoyang Jia, Jiahao Li, Bin Li, Houqiang Li, and Yan Lu. Generative latent coding for ultra-low bitrate image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26088-26098, 2024. 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[27] Wei Jiang and Ronggang Wang. Mlic++: Linear complexity multi-reference entropy modeling for learned image compression. In ICML 2023 Workshop Neural Compression: From Information Theory to Applications, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[28] Wei Jiang, Jiayu Yang, Yongqi Zhai, Peirong Ning, Feng Gao, and Ronggang Wang. Mlic: Multi-reference entropy" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17387" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "model for learned image compression. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7618-7627, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.135, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[29] Álvaro Barbero Jiménez. Mixture of diffusers for scene composition and high resolution image generation. arXiv preprint arXiv:2302.02412, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.177, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[30] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.205, + 0.483, + 0.273 + ], + "angle": 0, + "content": "[31] Nikolai Körber, Eduard Kromer, Andreas Siebert, Sascha Hauke, Daniel Mueller-Gritschneder, and Björn Schuller. Egid: enhanced low-bit-rate generative image compression guided by semantic segmentation. In European Conference on Computer Vision, pages 202-220. Springer, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.274, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[32] Nikolai Körber, Eduard Kromer, Andreas Siebert, Sascha Hauke, Daniel Mueller-Gritschneder, and Björn Schuller. Perco (sd): Open perceptual compression. arXiv preprint arXiv:2409.20255, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.483, + 0.385 + ], + "angle": 0, + "content": "[33] Nupur Kumari, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Ensembling off-the-shelf models for gan training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10651-10662, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.386, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[34] Hagyeong Lee, Minkyu Kim, Jun-Hyuk Kim, Seungeon Kim, Dokwan Oh, and Jaeho Lee. Neural image compression with text-guided encoding for both pixel-level and perceptual fidelity. arXiv preprint arXiv:2403.02944, 2024. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.483, + 0.509 + ], + "angle": 0, + "content": "[35] Eric Lei, Yigit Berkay Uslu, Hamed Hassani, and Shirin Saeedi Bidokhti. Text+ sketch: Image compression at ultra low rates. arXiv preprint arXiv:2307.01944, 2023. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.483, + 0.565 + ], + "angle": 0, + "content": "[36] Jiahao Li, Bin Li, and Yan Lu. Neural video compression with diverse contexts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22616-22626, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.483, + 0.635 + ], + "angle": 0, + "content": "[37] Zhuoyuan Li, Yao Li, Chuanbo Tang, Li Li, Dong Liu, and Feng Wu. Uniformly accelerated motion model for inter prediction. In 2024 IEEE International Conference on Visual Communications and Image Processing (VCIP), pages 1-5. IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.483, + 0.704 + ], + "angle": 0, + "content": "[38] Zhuoyuan Li, Junqi Liao, Chuanbo Tang, Haotian Zhang, Yuqi Li, Yifan Bian, Xihua Sheng, Xinmin Feng, Yao Li, Changsheng Gao, et al. Ustc-td: A test dataset and benchmark for image and video coding in 2020s. arXiv preprint arXiv:2409.08481, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.706, + 0.483, + 0.761 + ], + "angle": 0, + "content": "[39] Zhuoyuan Li, Zikun Yuan, Li Li, Dong Liu, Xiaohu Tang, and Feng Wu. Object segmentation-assisted inter prediction for versatile video coding. IEEE Transactions on Broadcasting, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[40] Zhiyuan Li, Yanhui Zhou, Hao Wei, Chenyang Ge, and Jingwen Jiang. Towards extreme image compression with latent feature guidance and diffusion prior. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[41] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.16 + ], + "angle": 0, + "content": "[42] Jinming Liu, Heming Sun, and Jiro Katto. Learned image compression with mixed transformer-cnn architectures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14388-14397, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[43] Fabian Mentzer, George D Toderici, Michael Tschannen, and Eirikur Agustsson. High-fidelity generative image compression. Advances in Neural Information Processing Systems, 33:11913-11924, 2020. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.218, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[44] David Minnen and Saurabh Singh. Channel-wise autoregressive entropy models for learned image compression. In 2020 IEEE International Conference on Image Processing (ICIP), pages 3339-3343. IEEE, 2020. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[45] David Minnen, Johannes Balle, and George D Toderici. Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems, 31, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[46] Matthew J Muckley, Alaaeldin El-Nouby, Karen Ullrich, Hervé Jégou, and Jakob Verbeek. Improving statistical fidelity for neural image compression with implicit local likelihood models. In International Conference on Machine Learning, pages 25426-25443. PMLR, 2023. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[47] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[48] Gaurav Parmar, Taesung Park, Srinivasa Narasimhan, and Jun-Yan Zhu. One-step image translation with text-to-image models. arXiv preprint arXiv:2403.12036, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.565 + ], + "angle": 0, + "content": "[49] Yichen Qian, Ming Lin, Xiuyu Sun, Zhiyu Tan, and Rong Jin. Entroformer: A transformer-based entropy model for learned image compression. arXiv preprint arXiv:2202.05492, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.648 + ], + "angle": 0, + "content": "[50] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.65, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[51] Lucas Relic, Roberto Azevedo, Markus Gross, and Christopher Schroers. Lossy image compression with foundation diffusion models. In European Conference on Computer Vision, pages 303-319. Springer, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.706, + 0.907, + 0.774 + ], + "angle": 0, + "content": "[52] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[53] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[54] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87-103. Springer, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[55] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17388" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[56] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[57] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.483, + 0.276 + ], + "angle": 0, + "content": "[58] Gary J Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. Overview of the high efficiency video coding (hevc) standard. IEEE Transactions on circuits and systems for video technology, 22(12):1649-1668, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[59] Lucas Theis, Tim Salimans, Matthew D Hoffman, and Fabian Mentzer. Lossy compression with gaussian diffusion. arXiv preprint arXiv:2206.08889, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[60] George Toderici, Lucas Theis, Nick Johnston, Eirikur Agustsson, Fabian Mentzer, Johannes Ballé, Wenzhe Shi, and Radu Timofte. Clic 2020: Challenge on learned image compression, 2020, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.418 + ], + "angle": 0, + "content": "[61] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.421, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[62] Gregory K Wallace. The jpeg still picture compression standard. Communications of the ACM, 34(4):30-44, 1991. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.483, + 0.517 + ], + "angle": 0, + "content": "[63] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, pages 1-21, 2024. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.575 + ], + "angle": 0, + "content": "[64] Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for image quality assessment. In The Thrity-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, pages 1398-1402. IEEE, 2003. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.577, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[65] Tongda Xu, Ziran Zhu, Dailan He, Yanghao Li, Lina Guo, Yuanyuan Wang, Zhe Wang, Hongwei Qin, Yan Wang, Jingjing Liu, et al. Idempotence and perceptual image compression. arXiv preprint arXiv:2401.08920, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[66] Naifu Xue, Zhaoyang Jia, Jiahao Li, Bin Li, Yuan Zhang, and Yan Lu. Dlf: Extreme image compression with dual-generative latent fusion. arXiv preprint arXiv:2503.01428, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[67] Zeyu Yan, Fei Wen, Rendong Ying, Chao Ma, and Peilin Liu. On perceptual lossy compression: The cost of perceptual reconstruction and an optimal training framework. In International Conference on Machine Learning, pages 11682-11692. PMLR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[68] Zeyu Yan, Fei Wen, and Peilin Liu. Optimally controllable perceptual lossy compression. arXiv preprint arXiv:2206.10082, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[69] Ruihan Yang and Stephan Mandt. Lossy image compression with conditional diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[70] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. arXiv preprint arXiv:2405.14867, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[71] Weihao Yu and Xinchao Wang. Mambaout: Do we really need mamba for vision? arXiv preprint arXiv:2405.07992, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[72] Weihao Yu, Pan Zhou, Shuicheng Yan, and Xinchao Wang. Inceptionnext: When inception meets convnext. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5672-5683, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.246 + ], + "angle": 0, + "content": "[73] Aiping Zhang, Zongsheng Yue, Renjing Pei, Wenqi Ren, and Xiaochun Cao. Degradation-guided one-step image super-resolution with diffusion priors. arXiv preprint arXiv:2409.17058, 2024. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[74] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.093, + 0.905, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17389" + } + ] +] \ No newline at end of file diff --git a/2025/StableCodec_ Taming One-Step Diffusion for Extreme Image Compression/5269d8ed-d6d6-4961-a311-a0e841365f77_origin.pdf b/2025/StableCodec_ Taming One-Step Diffusion for Extreme Image Compression/5269d8ed-d6d6-4961-a311-a0e841365f77_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fcafc8c99e106d24240c044469ed9f2946b2e088 --- /dev/null +++ b/2025/StableCodec_ Taming One-Step Diffusion for Extreme Image Compression/5269d8ed-d6d6-4961-a311-a0e841365f77_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c2c1058c42815c1a507508c9a0b82825b48abc168f99e218ba139c2c39975d5e +size 7202586 diff --git a/2025/StableCodec_ Taming One-Step Diffusion for Extreme Image Compression/full.md b/2025/StableCodec_ Taming One-Step Diffusion for Extreme Image Compression/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b9dbf52d24bd1432a498f0715da3e10a1f1cbf19 --- /dev/null +++ b/2025/StableCodec_ Taming One-Step Diffusion for Extreme Image Compression/full.md @@ -0,0 +1,334 @@ +# StableCodec: Taming One-Step Diffusion for Extreme Image Compression + +Tianyu Zhang Xin Luo Li Li Dong Liu + +MOE Key Laboratory of Brain-Inspired Intelligent Perception and Cognition, University of Science and Technology of China, Hefei 230093, China + +{zhangtianyu, xinluo}@mail.ustc.edu.cn, {lil1, dongeliu}@ustc.edu.cn + +![](images/72f1cd571559f186a71ddf82a499acc1730b5a3132351d5605ad61029481f7a6.jpg) +Figure 1. Visual examples and comparisons when compressing a 4K-resolution image [38] at ultra-low bitrates. The proposed Stable-Coded produces more realistic and consistent details with fewer bits. In contrast, VVC [10], ELIC [20] and MS-ILLM [46] reconstructions are blurry, while PerCo [11] and DiffEIC [40] generate inconsistent details against the original images. Best viewed on screen for details. + +# Abstract + +sion models. However, current approaches require a large number of denoising steps at the decoder to generate realistic results under extreme bitrate constraints, limiting their application in real-time compression scenarios. Additionally, these methods often sacrifice reconstruction fidelity, as diffusion models typically fail to guarantee pixel-level consistency. To address these challenges, we introduce StableCodec, which enables one-step diffusion for high-fidelity and high-realism extreme image compression with improved coding efficiency. To achieve ultra-low bitrates, we first develop an efficient Deep Compression Latent Codec to transmit a noisy latent representation for a single-step denoising process. We then propose a Dual-Branch Coding Structure, consisting of a pair of auxiliary encoder and decoder, to enhance reconstruction fidelity. Furthermore, we adopt end-to-end optimization with joint bitrate and pixel-level constraints. Extensive experiments on the CLIC 2020, DIV2K, and Kodak datasets demonstrate that StableCodec outperforms existing methods in terms of FID, KID and DISTS by a significant margin, even at bitrates as low as 0.005 bits per pixel, while maintaining strong fidelity. Additionally, StableCodec achieves inference speeds comparable to mainstream transform coding schemes. Code is available at our project repository. + +# 1. Introduction + +Image compression is a foundational problem in signal processing. Driven by advances in digital imaging and the widespread use of social platforms, the volume of image data in modern multimedia has grown exponentially, placing increasing demands on the coding efficiency of image compression techniques. Over the past few decades, traditional CODECs such as JPEG [62] and H.266/VVC [10], along with emerging learning-based methods [4, 5, 12, 18, 19, 28, 42, 44, 45], have been widely adopted in real-world image compression applications. However, these methods are typically optimized for rate-distortion performance, and often produce unrealistic and blurry reconstructions, particularly under severe bitrate constraints, as shown in Fig. 1. + +To tackle this issue, generative image compression [2, 43] optimized for human perceptual performance has gained increasing attention. These methods are evaluated based on the rate-distortion-perception tradeoff [8, 9, 67, 68], and progressively demonstrate their advantages in producing visually appealing reconstructions at lower bitrates compared to traditional CODEs or common neural CODEs. A prominent research direction [3, 21, 31, 43, 43, 46, 66] involves integrating a discriminator into the transform coding pipeline [5, 20, 44], employing adversarial training to enhance the perceptual quality of reconstructions. Motivated by the impressive generative capabilities, more researchers [11, 24, 35, 40, 51, 59, 65, 69] have begun exploring the po + +![](images/3c5ee440181aaca1b32bec22899a91ae3be7cb8eaa5c5c3a71b5a7fd47924454.jpg) +Figure 2. (Top) Illustration of our motivation. One-step diffusion can produce perceptually consistent results given severely corrupted images and a general prompt. (Bottom) Challenges in StableCodec: how to compress a noisy latent for one-step diffusion using ultra-low bitrates, and how to improve fidelity. + +tential of diffusion models, particularly the generative priors in large pre-trained text-to-image (T2I) models, to compensate for severely distorted information at ultra-low bitrates while ensuring perceptually consistent generation. A recent study, PerCo [11], produces realistic results at an extreme bitrate as low as 0.003 bits per pixel (bpp) using a pretrained latent diffusion model (LDM), highlighting the potential of diffusion-based generative CODEs on image compression at more severe bitrates. + +Despite these promising advancements, existing methods face two primary limitations inherent to diffusion models. First, they require dozens of denoising steps at the decoder to produce results with sufficient perceptual quality. Second, the reconstructions often deviate from the original images, as diffusion models typically do not guarantee reconstruction consistency. To address the first challenge, we consider leveraging the generative priors in SD-Turbo [54], a distilled version of Stable Diffusion 2.1 [52] that enables real-time image synthesis in 1 to 4 denoising steps. Following [73], we demonstrate that SD-Turbo can produce perceptually consistent reconstructions with a single-step denoising process, even for severely corrupted inputs and a general positive prompt, as shown in Fig. 2. We thus pose an intuitive question: Can we compress a noisy latent representation of the original image, which can be effectively denoised in a single-step diffusion process, using an ultra-low bitrate? Building on these insights, we present StableCodec for extreme image compression, which integrates SD-Turbo with the proposed Deep Compression Latent Codec to compress noisy latents at ultra-low bitrates for a single-step diffusion process. + +In response to the second challenge, we introduce a Dual-Branch Coding Structure with a pair of auxiliary en + +coder and decoder to further enhance reconstruction fidelity. Considering the limitations of the pre-trained VAE encoder on practical entropy coding and reconstruction consistency, we employ a rate-distortion-oriented auxiliary encoder to embed more entropy-aware semantic information for coding decisions. In parallel, we add an auxiliary decoder to perform structure apportionment during the decoding process, improving the generation guidance on the one-step denoising process for more consistent details. To enable end-to-end optimization, we design a two-stage training objective that jointly optimizes bitrate and pixel-level constraints. + +Benefit from these designs, StableCodec produces high-fidelity and high-realism reconstructions at ultra-low bitrates as low as 0.005 bpp. Extensive experiments on CLIC 2020 [60], DIV2K [1] and Kodak [16] demonstrate that StableCodec sets up a new state-of-the-art performance in terms of FID [22], KID [6], and DISTS [14], outperforming existing methods by significant margins. In terms of computational complexity, StableCodec offers much faster decoding speeds compared to other diffusion-based competitors, achieving inference times comparable to those of mainstream transform coding schemes. For practical deployment, StableCodec supports inference at arbitrary resolutions with memory consumption less than 9 GB. + +We summarize our contributions as follows: + +- We present StableCodec, an extreme image codec integrating one-step diffusion and Deep Compression Latent Codec to achieve ultra-low bitrate compression with high realism, high fidelity and superior coding efficiency. +- We develop Dual-Branch Coding Structure to improve reconstruction fidelity. A pair of auxiliary encoder and decoder is introduced for semantic enhancement and structure apportionment. +- StableCodec obtains SOTA FID, KID and DISTS performance on CLIC 2020 and DIV2K dataset, significantly outperforms existing methods at bitrates as low as 0.005 bpp with well-preserved fidelity, and achieves comparable inference speeds with mainstream neural CODECs. + +# 2. Related Work + +# 2.1. Generative Image Compression + +Learning-based image compression has shown competitive potential compared to traditional methods [10, 37, 39, 58, 62], leveraging non-linear transforms and joint rate-distortion optimization. Balle et al. [4] introduced the first end-to-end learned image compression framework, which was subsequently enhanced with the hyperprior [5] and context model [45]. Building on this foundation, much work [12, 18, 19, 28, 42, 44, 49] has been devoted to improving both rate-distortion performance and model practicality. + +In practical scenarios, a key challenge is achieving extreme image compression at ultra-low bitrates while main + +taining both fidelity and realism [2]. Traditional image codes optimized for rate-distortion often produce blurry reconstructions and noticeable artifacts. To address this, Mentzer et al. [43] introduced HiFiC and the concept of generative image compression, integrating GANs into codec optimization and evaluating performance in terms of the rate-distortion-perception tradeoff [8, 9, 67, 68]. Subsequent research can be broadly categorized into two main approaches. The first category [3, 21, 31, 43, 46] focuses on enhancing transform coding [5, 20, 44] for human perception by incorporating adversarial losses and optimized discriminator architectures, which typically can be extended to a wide range of bitrate and a flexible decoding control between fidelity and realism [3, 31]. The second category [11, 24, 35, 40, 51, 59, 65, 69] leverages diffusion models for generative image compression. Although these methods show promise for ultra-low bitrate compression, they are often constrained by reconstruction fidelity and inference efficiency due to the multi-step denoising process. Recently, GLC [26] introduced transform coding in the generative latent space of VQ-VAE [15, 61], achieving more visually appealing results at ultra-low bitrates. + +# 2.2. Generative Models and Few-Step Diffusions + +Generative models play a crucial role in image generation. While many architectures, such as VAEs [30] and GANs [17], have been explored, diffusion models [55] have emerged as a powerful alternative, achieving state-of-the-art synthesis quality. Inspired by non-equilibrium statistical physics [55], diffusion models learn to reverse a noise perturbation process through a Markovian framework. Recent advancements, such as DDPM [23], DDIM [56], and LDM [52], have significantly reduced computational complexity and improved image synthesis quality, making diffusion-based approaches a dominant force in generative modeling. + +To address the inefficiency of iterative denoising, several approaches [53, 54, 57, 70] aim to reduce the number of denoising steps while maintaining generation quality. These methods train diffusion models to approximate the full denoising trajectory in a single or a few steps, significantly improving inference efficiency. Notably, SD/SDXL-Turbo [54] demonstrates image generation in 1 to 4 steps with near-parity quality compared to multi-step models, making real-time diffusion-based applications [48, 63, 73] feasible. + +# 3. Method + +# 3.1. Overview + +In this section, we introduce the overall framework of the proposed StableCodec, built upon SD-Turbo [54] with a VAE encoder $\mathcal{E}_{\mathrm{SD}}$ , a VAE decoder $\mathcal{D}_{\mathrm{SD}}$ and a denoising Unet $\epsilon_{\mathrm{SD}}$ . As shown in Fig. 3, we incorporate a Deep Compression Latent Codec to perform extreme transform + +![](images/fb19566377769b1588ee11eede0d60fe9da173e08c2d511ccb2532a752adc0e2.jpg) +Figure 3. The framework of StableCodec. We incorporate the proposed Deep Compression Latent Codec to transmit a noisy latent $l_{T}$ for one-step denoising, where $64 \times$ denotes a spatial compression ratio of 64. To adjust the latent resolution, we deploy DownSample block and Conv $3 \times 3$ as adapters after the VAE encoder $\mathcal{E}_{\mathrm{SD}}$ and auxiliary encoder $\mathcal{E}_{\mathrm{Aux}}$ , respectively. We use a general prompt in both training and inference. The auxiliary decoder $\mathcal{D}_{\mathrm{Aux}}$ shares a similar structure with $g_{s}$ . More details on networks are provided in the supplementary. + +coding in the VAE latent space. To adapt SD-Turbo for image compression, we integrate LoRA [25] into $\mathcal{E}_{\mathrm{SD}}$ and $\epsilon_{\mathrm{SD}}$ , while keeping $\mathcal{D}_{\mathrm{SD}}$ unchanged to preserve the generative priors [73]. Additionally, we introduce a Dual-Branch Coding Structure to enhance reconstruction fidelity, utilizing an auxiliary encoder $\mathcal{E}_{\mathrm{Aux}}$ to embed rich semantic information and an auxiliary decoder $\mathcal{D}_{\mathrm{Aux}}$ to perform structure apportionment. Finally, we optimize StableCodec end-to-end with joint bitrate and pixel-level constraints, achieving high-fidelity and high-realism extreme image compression. + +# 3.2. Deep Compression Latent Codec + +We design our latent codec using the extreme analysis transform $g_{a}$ , extreme synthesis transform $g_{s}$ and a 4-step autoregressive entropy model. To reach ultra-low bitrates, we employ deep compression transform networks for both $g_{a}$ and $g_{s}$ . Specifically, $\mathcal{E}_{\mathrm{SD}}$ and $\mathcal{D}_{\mathrm{SD}}$ provide a latent space with a spatial compression ratio of 8 (abbreviated as $8\times$ ). Unlike mainstream schemes [18, 20, 26, 28, 40, 42, 49] that perform entropy coding at $16\times$ , we further downsample and apply entropy coding for $\hat{y}$ at $64\times$ and the hyperprior [5] at $256\times$ . Consequently, we use $g_{s}$ to restore the spatial compression ratio to $8\times$ for $\epsilon_{\mathrm{SD}}$ and $\mathcal{D}_{\mathrm{SD}}$ . The entire coding process can be formulated as follows: + +$$ +l = \operatorname {c o n c a t} \left[ \mathcal {E} _ {\mathrm {S D}} (x), \mathcal {E} _ {\text {A u x}} (x) \right] \tag {1} +$$ + +$$ +y = g _ {a} (l), \hat {y} = Q (y), l _ {T} = g _ {s} (\hat {y}) \tag {2} +$$ + +$$ +l _ {0} = \left[ l _ {T} - \sqrt {1 - \bar {\alpha} _ {T}} \cdot \epsilon_ {\mathrm {S D}} \left(l _ {T}, T\right) \right] / \sqrt {\bar {\alpha} _ {T}} \tag {3} +$$ + +$$ +\hat {x} = \mathcal {D} _ {\mathrm {S D}} \left(l _ {0} + \mathcal {D} _ {\mathrm {A u x}} (\hat {y})\right) \tag {4} +$$ + +In Eq. (1), we first obtain an intermediate latent $l$ from the input image $x$ through $\mathcal{E}_{\mathrm{SD}}$ and $\mathcal{E}_{\mathrm{Aux}}$ . Eq. (2) is the latent-space transform coding process to produce a noisy latent $l_{T}$ using ultra-low bitrates. Eq. (3) displays the one-step denoising process with the noise schedule $\{\bar{\alpha}_t\}$ [23] in the $T$ -th timestep. Finally, in Eq. (4), the reconstruction $\hat{x}$ is obtained from $l_{0}$ and $\hat{y}$ using $\mathcal{D}_{\mathrm{SD}}$ and $\mathcal{D}_{\mathrm{Aux}}$ . To balance performance and coding latency, we construct efficient $g_{a}$ and $g_{s}$ with InceptionNeXt [72] and GatedCNN [71], and build a 4-step antoregressive entropy model with quadtree partition [36] and latent residual prediction [44]. + +# 3.3. Dual-Branch Coding Structure + +While deploying the proposed latent codec with LoRA enables image compression with SD-Turbo at ultra-low bitrates, the reconstruction fidelity is limited. In this section, we analyze the reasons and introduce Dual-Branch Coding Structure with a pair of auxiliary encoder and decoder, $\mathcal{E}_{\mathrm{Aux}}$ and $\mathcal{D}_{\mathrm{Aux}}$ , to further enhance compression performance. + +# 3.3.1. Entropy-Aware Semantic Enhancement + +We observed that the VAE in SD-Turbo has several limitations when reconstruction fidelity and practical coding are required. As noted in [52] and Table 1, this VAE is pretrained for perceptual compression, which does not preserve pixel-level fidelity as well as a rate-distortion-oriented autoencoder from a typical neural codec [20]. Additionally, while this VAE provides a compressed representation of the original image, it is still in floating-point format and not optimized for practical entropy coding, making it less suitable + +Table 1. Reconstruction quality of VAEs on Kodak [16]. The pre-trained VAE in SD performs perceptual compression, while the one in ELIC [20] preserves more pixel-level information. + +
VAEPSNR↑MS-SSIM↑LPIPS↓DISTS↓
SD26.650.93180.07260.0415
ELIC40.400.99610.05550.0707
+ +![](images/65ec16377b6b7a4e67fe570150e2341b979702bc630cc024376c5898daf53581.jpg) +Figure 4. Top-energy channels learned from different encoders. $\mathcal{E}_{\mathrm{Aux}}$ embeds more pixel-level semantic information into latents. + +for further latent-space ultra-low bitrate compression. + +Building on these insights, we introduce the analysis transform of a pre-trained high-bitrate ELIC model [20] to serve as an auxiliary encoder $\mathcal{E}_{\mathrm{Aux}}$ . $\mathcal{E}_{\mathrm{Aux}}$ remains frozen, and provides rich, entropy-aware semantic information of the input images. To combine the latents from different encoders, we introduce trainable adapters that align the latent resolutions to a spatial compression ratio of 16, followed by channel-wise concatenation. During optimization, various pieces of information are learned from different encoders as shown in Fig. 4, where more pixel-level semantic information is embedded into the latent codec through $\mathcal{E}_{\mathrm{Aux}}$ . + +# 3.3.2. Structure Apportionment + +The reconstruction quality of StableCodec heavily depends on how the denoising Unet is conditioned. Since a fixed prompt is used for both training and inference, the one-step denoising process is primarily guided by $l_{T}$ , which is produced by $g_{s}$ . This places high demands on the capability of $g_{s}$ , resulting in unsatisfactory denoising guidance, as reflected in the reconstructions shown in Fig. 5 (b). + +To alleviate the decoding burden on $g_{s}$ , we introduce an auxiliary decoder $\mathcal{D}_{\mathrm{Aux}}$ to perform an additional decoding branch from $\hat{y}$ , bypassing the Unet. This design is motivated by the observation that an extremely compressed bitstream contains mainly the basic structure of images. We distribute and decode these components directly from the bitstream using the auxiliary branch, allowing $g_{s}$ to focus primarily on producing guidance to generate realistic details. Fig. 5 compares StableCodec that is trained with or without $\mathcal{D}_{\mathrm{Aux}}$ . When trained without $\mathcal{D}_{\mathrm{Aux}}$ , $g_{s}$ produces all types of information as it is the only decoding branch. When trained with $\mathcal{D}_{\mathrm{Aux}}$ , structural information is routed + +![](images/83e18420fabd9eae154e1ae6c87af76a385445279d9f2cc6b87ff24af0f64499.jpg) +(a) Latents from different decoders, trained w. or w/o $\mathcal{D}_{\mathrm{Aux}}$ + +![](images/25ba7824b0160ae6f6194eedcf32c522d1eb40166ef0338a518723382f6cf250.jpg) +(b) Reconstruction comparison, trained w. or w/o $\mathcal{D}_{\mathrm{Aux}}$ +Figure 5. Impact of $\mathcal{D}_{\mathrm{Aux}}$ on latents and reconstructions. + +through $\mathcal{D}_{\mathrm{Aux}}$ , while the energy in $g_{s}$ latents drops significantly. Meanwhile, more semantically aligned details are reconstructed in Fig. 5 (b), suggesting $g_{s}$ now provides better high-frequency guidance for the denoising Unet $\epsilon_{\mathrm{SD}}$ . + +# 3.4. End-to-End Training Objective + +We adopt end-to-end optimization with joint bitrate and pixel-level restrictions to train StableCodec. Given the original image $x$ , the quantized latent $\hat{y}$ and the reconstructed image $\hat{x}$ , we construct our training objective based on the standard rate-distortion loss: + +$$ +\lambda \mathcal {R} (\hat {y}) + \mathcal {D} (x, \hat {x}) \tag {5} +$$ + +where the bitrate $\mathcal{R}$ and pixel-level distortion $\mathcal{D}$ are balanced by the Lagrange multiplier $\lambda$ . + +Inspired by [20, 44], we train StableCodec with a 2-stage implicit bitrate pruning (IBP) strategy. We first train a base model using a smaller $\lambda_{base}$ , adapting the latent codec into SD-Turbo under a relaxed bitrate constraint, and warming up with a more expressive transform. In the second stage, we finetune the shared base model with larger $\lambda_{target}$ to reach ultra-low target bitrates. IBP facilitates efficient and stable training, resulting in improved performance. + +The distortion term $\mathcal{D}$ includes MSE, LPIPS (using VGG features) [74] and a CLIP [50] distance $\mathcal{L}_{CLIP}$ , for which we compute the L2-distance between the CLIP embeddings of $x$ and $\hat{x}$ . Note that this term is a simplified version from [34], and we find it beneficial for reconstruction at ultralow bitrates. Additionally, we follow [73] and incorporate + +![](images/7f9d7f01a147b11667990979775e443d604673ba0ed3b3a58bac5fc972014eef.jpg) + +![](images/8b2210b11c21bd44326ced02d09aa107f91421454626268e90c0dbe2af2f5d21.jpg) + +![](images/11fe770e88e21dfb7f2d021c8cee65889c7c31f74963780c0ba1a74b7f2ee2f4.jpg) + +![](images/81019f026d7b49fccdc154c011b104618da55869cf16c5e6061794d685f35b0a.jpg) + +![](images/1a69dd8cbcd53a10f0feaa9ddb2a0fcc11234f4ae36a01790ddf1039dd387bcb.jpg) + +![](images/c1bbb274d9ea3a973b75d174502c6912932e225677744d441bb5a0ffcc364d71.jpg) + +![](images/69ecf9d1bfe2780365a7fe85fd7d63aac2d4e435fe6a5f3a98b0468cdbe5aba4.jpg) + +![](images/af1daffea9066d94a7acaec322648900023351dcf045c3c56974928194027c9f.jpg) + +![](images/6ccbb99bd2b025a4802db51eae451d9d8a947ba3c9a63076e7855d37fdbd6fb5.jpg) + +![](images/50664a381cec20a3f97596b197e1f90dbe127b7408e8841b6ff3534c926bdd49.jpg) +Figure 6. Rate-distortion and rate-perception curve comparisons of different methods on the CLIC 2020, DIV2K and Kodak dataset. + +![](images/a89037520d9c1e38d270c748a76e5e136c40ee1fc101cf61a03570260c0459b0.jpg) + +![](images/741544a4de1adbd1d53be9b8cbbd9d8f8f73924c67ff2c2141275ab3d6765d6b.jpg) + +![](images/b0d4541c6ae912b0dd7cecea8c98d296195ecaed058b50c6f5e62bb0fd41e1ca.jpg) + +a GAN loss $\mathcal{L}_{adv}$ to narrow the distribution gap between $x$ and $\hat{x}$ . We use DINOv2 [47] with registers [13] as the discriminator backbone [33]. To ensure stable training, we incorporate the GAN only in the second training stage. The full objective can be formulated as: + +$$ +\text {S t a g e I :} \underset {\theta} {\arg \min } \lambda_ {\text {b a s e}} \mathcal {R} (\hat {y}) + \mathcal {D} (x, \hat {x}) \tag {6} +$$ + +$$ +\text {S t a g e I I :} \underset {\theta} {\arg \min } \lambda_ {\text {t a r g e t}} \mathcal {R} (\hat {y}) + \mathcal {D} (x, \hat {x}) + \beta \mathcal {L} _ {\text {a d v}} \tag {7} +$$ + +$$ +\begin{array}{l} \mathcal {D} (x, \hat {x}) = d _ {1} M S E (x, \hat {x}) + d _ {2} L P I P S (x, \hat {x}) \tag {8} \\ + d _ {3} \mathcal {L} _ {C L I P} (x, \hat {x}) \\ \end{array} +$$ + +where $\theta$ represents all trainable parameters in StableCodec (Fig. 3), $d_{1}, d_{2}, d_{3}$ and $\beta$ are balancing weights. + +# 4. Experiments + +# 4.1. Implementation + +Training Details. We use the training set of DF2K [41] and CLIC 2020 Professional [60] to train StableCodec. During training, we use $512 \times 512$ patches with a batch size of 8. The first training stage takes over 100k iterations with a + +learning rate of $1e^{-4}$ and a $\lambda_{base}$ of 0.5 (about 0.05bpp). In the second stage, we finetune the base model for another 20k iterations with GAN incorporated and $\lambda_{target} \in \{2,3,4,6,8,12,16,24,32\}$ , while the learning rate undergoes $5e^{-5}, 2e^{-5}, 1e^{-5}$ and $1e^{-6}$ for 5k iterations each. We set $d_1, d_2, d_3$ and $\beta$ to 2, 1, 0.1 and 0.1, respectively. All models are trained using 2 RTX 3090 GPUs. + +Test Data. We evaluate StableCodec on the test set of CLIC 2020 Professional [60] (CLIC 2020 Test), the validation set of DIV2K [1] (DIV2K Val) and Kodak [16] following [26, 40]. CLIC 2020 Test and DIV2K Val contain 428 and 100 high-quality 2K-resolution natural images, respectively, while Kodak contains 24 natural images with a smaller resolution of $768 \times 512$ . We evaluate all images with the original resolution as detailed in the supplementary. + +Evaluation Metrics. We employ established metrics to assess the rate-distortion-perception performance of StableCodec. Concretely, we measure bitrate by bits per pixel (bpp), and evaluate perceptual quality using FID [22], KID [6], DISTS [14] and LPIPS [74] (using AlexNet features by default). Meanwhile, we use PSNR and MS-SSIM [64] to measure the reconstruction fidelity. We follow [26, 40, 46] to calculate FID and KID on $256 \times 256$ patches, and neglect + +![](images/3e87a132c57a35619e00a779af9e427ef509fd54c56ae3a052d0c677b4768983.jpg) +Figure 7. Qualitative comparisons of different methods on Kodak. Best viewed on screen for details. + +the results on Kodak as it is too small for calculating. Note that pixel-level distortion metrics like LPIPS, PSNR and MS-SSIM have strong limitations when evaluating image compression at ultra-low bitrates [11, 14, 26, 35]. Therefore, for StableCodec, we focus primarily on FID, KID, and DISTS, which offer a more accurate assessment of quality in severely compressed images. We also provide user study in the supplementary to visually validate the results. + +Compared Methods. We compare StableCodec with advanced image compression methods: (1) Traditional Codec H.266/VVC [10] by VTM-23.0 intra coding, (2) Neural Codec ELIC [20], (3) Generative Codec HiFiC [43], Text+Sketch [35], MS-ILLM [46], PerCo [11], EGIC [31], DiffEIC [40], TACO [34] and GLC [26]. Note that some methods do not release models for ultra-low bitrates, we either re-train or finetune existing weights to reach specific bitrates. For PerCo and GLC that do not have official codes, we use PerCo (SD) [32] as a substitute, and request for the results of GLC under the same evaluation approach1. We equip PerCo and Text+Sketch with the same inference strategy for a fair comparison on high-resolution images. + +# 4.2. Main Results + +# 4.2.1. Rate-Distortion-Perception Performance + +Fig. 6 presents the rate-perception and rate-distortion curves of various methods at ultra-low bitrates over CLIC 2020 Test and DIV2K Val. The proposed StableCodec outperforms all compared methods in terms of FID, KID, and DISTS. Specifically, StableCodec shows a significant im + +Table 2. Encoding and decoding seconds averaged on Kodak. + +
TypeMethod#StepsEnc. TDec. T
VAE- basedELIC [20]-0.1550.245
MLIC++ [27]-0.3640.319
GAN- basedHiFiC [43]-0.1430.337
MS-ILLM [46]-0.1390.316
Diffusion- basedText+Sketch [35]25113.25233.560
PerCo [11]200.2873.742
DiffEIC [40]500.6767.423
StableCodec (Ours)10.1590.326
+ +provement over H.266/VVC and ELIC on all perceptual metrics. Compared to generative codec especially previous SOTA GLC [26], StableCodec demonstrates superiority and stability on FID and KID performance with well-preserved fidelity, and reaches extreme bitrates as low as 0.005 bpp. Although TACO [34] achieves the best LPIPS performance, it fails to ensure visual quality as FID, KID and DISTS scores are high. The PSNR and MS-SSIM results on CLIC and DIV2K are displayed in the supplementary. + +# 4.2.2. Qualitative Comparisons + +We provide qualitative results among compared methods in Fig. 7. Notably, StableCodec generates more visually-aligned details and realistic textures at ultra-low bitrates, such as the teeth and murals shown in the first and third rows. In contrast, all other methods fail to produce high-realism results with well-preserved fidelity due to the severe bitrate restriction. For example, ELIC and MS-ILLM + +Table 3. Ablation study on ${\mathcal{E}}_{\text{Aux }}$ and ${\mathcal{D}}_{\text{Aux }}$ . + +
Model +VariantsBD-rate↓ on the Rate-Y curves
PSNRMS-SSIMLPIPSDISTS
Base0000
+ EAux-20.63%-22.05%-23.04%-28.13%
+ EAux & DAux-23.96%-28.12%-40.66%-54.89%
+ +Table 4. Ablation study on LoRA ranks $\left( {{\mathcal{E}}_{\mathrm{{SD}}}/{\epsilon }_{\mathrm{{SD}}}/{\mathcal{D}}_{\mathrm{{SD}}}}\right)$ . + +
LoRA RanksBD-rate↓ on the Rate-Y curves
PSNRMS-SSIMLPIPSDISTS
8/8/-0000
8/16/-1.21%-0.35%-4.67%-5.28%
16/16/--5.62%-3.27%-6.41%-12.78%
16/32/--7.12%-5.31%-13.43%-17.96%
32/32/--5.98%-5.69%-11.21%-16.70%
32/64/--5.15%-4.98%-12.95%-17.34%
16/32/4-21.77%-12.28%-8.27%-5.13%
+ +produce blurry reconstructions, while PerCo and DiffEIC deviate from the original images. EGIC and TACO exhibit noticeable artifacts, particularly in detailed areas. + +# 4.2.3. Computational Complexity + +We compare the practical complexity of StableCodec with representative schemes in Table 2 using a single RTX 3090 GPU. Among representative image compression schemes, Diffusion-based methods [11, 35, 40] typically suffer from a much longer decoding time compared to VAE-based [20, 27] or GAN-based [43, 46] competitors due to multi-step denoising. In contrast, StableCodec reaches comparable encoding and decoding speed against these methods exploiting one-step denoising, deep compression transforms and efficient entropy model, while achieving significantly better performance at ultra-low bitrates. Detailed runtime analysis is provided in the supplementary. In terms of memory, StableCodec consumes less than 9 GB VRAM with tiling techniques [29, 63], supporting arbitrary-resolution inference on a single GTX 1080Ti GPU. + +# 5. Ablation Studies + +In this section, we conduct ablations to validate the proposed components. For reliable comparison, we compute the BD-rate [7] with Rate-Y curves on Kodak [16] using four target bitrates, where Y denotes specific metrics among PSNR, MS-SSIM, LPIPS, and DISTS. + +Auxiliary Encoder and Decoder. We begin by providing numerical results for the auxiliary coding branch in Table 3. Specifically, we construct a base model without $\mathcal{E}_{\mathrm{Aux}}$ and $\mathcal{D}_{\mathrm{Aux}}$ , and a variant with $\mathcal{E}_{\mathrm{Aux}}$ only. When $\mathcal{E}_{\mathrm{Aux}}$ is incorporated, more than $20\%$ bits can be saved to reach the + +Table 5. Ablation study on the training strategy components. Implicit bitrate pruning is abbreviated as IBP. The base one-stage objective only contains bitrate, MSE and LPIPS. + +
Training StrategyBD-rate↓ on the Rate-Y curves
IBPLadvLCIPPSNRMS-SSIMLPIPSDISTS
---0000
✓---24.12%-13.67%-21.37%-16.60%
✓✓-24.41%9.88%-36.18%-49.24%
✓✓✓13.29%5.70%-38.99%-52.95%
+ +same reconstruction quality. Subsequently, $\mathcal{D}_{\mathrm{Aux}}$ further improves the performance particularly in perceptual quality. As shown in Fig. 5, the purified $g_{s}$ latent provides better guidance for reconstruction consistency. + +LoRA Ranks. We explore the impact of LoRA ranks in Table 4. Positive results are observed in both distortion and perception as the ranks increase to 16/32, which become our final choice. For larger ranks like 32 and 64, we observe performance degradation as the pre-trained priors may be corrupted. Besides, adding LoRA to the VAE decoder introduces a distortion-perception tradeoff, where PSNR and MS-SSIM improve at the cost of LPIPS and DISTS. To preserve perceptual quality, we leave the decoder unchanged. + +Training Strategy. We perform ablations on our training strategy components in Table 5. We construct a simplified one-stage objective with bitrate, MSE and LPIPS, then progressively integrate the two-stage implicit bitrate pruning (IBP), adversarial training $\mathcal{L}_{adv}$ and the CLIP distance term $\mathcal{L}_{CLIP}$ . We find that IBP improves performance considerably by first adapting the latent codec into the T2I model with relaxed bitrate constraint. Furthermore, incorporating GAN introduces a significant distortion-perception tradeoff since we primarily focus on the perceptual quality. Besides, the CLIP distance alleviates the degradation in distortion and slightly improves perceptual quality. + +# 6. Conclusion + +In this work, we introduce StableCodec, a novel diffusion-based extreme image compression approach that addresses key limitations of existing methods. By leveraging one-step diffusion in combination with Deep Compression Latent Codec and Dual-Branch Coding Structure, StableCodec achieves ultra-low bitrate compression with high realism, fidelity, and coding efficiency. Extensive experimental evaluations on benchmark datasets demonstrate the superiority of StableCodec in terms of FID, KID, and DISTS, even at extreme bitrates as low as 0.005 bpp, while enabling competitive speeds with mainstream transform coding methods. These results underscore the potential of diffusion models for practical image compression, particularly in real-time coding scenarios where bitrate is severely constrained. + +# References + +[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 126-135, 2017. 3, 6 +[2] Eirikur Agustsson, Michael Tschannen, Fabian Mentzer, Radu Timofte, and Luc Van Gool. Generative adversarial networks for extreme learned image compression. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 221-231, 2019. 2, 3 +[3] Eirikur Agustsson, David Minnen, George Toderici, and Fabian Mentzer. Multi-realism image compression with a conditional generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22324-22333, 2023. 2, 3 +[4] Johannes Ballé, Valero Laparra, and Eero P Simoncelli. End-to-end optimized image compression. arXiv preprint arXiv:1611.01704, 2016. 2, 3 +[5] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436, 2018. 2, 3, 4 +[6] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 3, 6 +[7] Gisle Bjontegaard. Calculation of average psnr differences between rd-curves. Tech. Rep. VCEG-m33, 2001. 8 +[8] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6228-6237, 2018. 2, 3 +[9] Yochai Blau and Tomer Michaeli. Rethinking lossy compression: The rate-distortion-perception tradeoff. In International Conference on Machine Learning, pages 675-685. PMLR, 2019. 2, 3 +[10] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (vvc) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3736-3764, 2021. 1, 2, 3, 7 +[11] Marlene Careil, Matthew J Muckley, Jakob Verbeek, and Stéphane Lathuilière. Towards image compression with perfect realism at ultra-low bitrates. In The Twelfth International Conference on Learning Representations, 2023. 1, 2, 3, 7, 8 +[12] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7939-7948, 2020. 2, 3 +[13] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023. 6 +[14] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 3, 6, 7 +[15] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Pro + +ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 3 +[16] Rich Franzen. Kodak lossless true color image suite (photocd pcd0992). http://r0k.us/graphics/kodak/, 1993. 3, 5, 6, 8 +[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 3 +[18] Zongyu Guo, Zhizheng Zhang, Runsen Feng, and Zhibo Chen. Causal contextual prediction for learned image compression. IEEE Transactions on Circuits and Systems for Video Technology, 32(4):2329-2341, 2021. 2, 3, 4 +[19] Dailan He, Yaoyan Zheng, Baocheng Sun, Yan Wang, and Hongwei Qin. Checkerboard context model for efficient learned image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14771-14780, 2021. 2, 3 +[20] Dailan He, Ziming Yang, Weikun Peng, Rui Ma, Hongwei Qin, and Yan Wang. Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5718-5727, 2022. 1, 2, 3, 4, 5, 7, 8 +[21] Dailan He, Ziming Yang, Hongjiu Yu, Tongda Xu, Jixiang Luo, Yuan Chen, Chenjian Gao, Xinjie Shi, Hongwei Qin, and Yan Wang. Po-elic: Perception-oriented efficient learned image coding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1764–1769, 2022. 2, 3 +[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 3, 6 +[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3, 4 +[24] Emiel Hoogeboom, Eirikur Agustsson, Fabian Mentzer, Luca Versari, George Toderici, and Lucas Theis. High-fidelity image compression with score-based generative models. arXiv preprint arXiv:2305.18231, 2023. 2, 3 +[25] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 4 +[26] Zhaoyang Jia, Jiahao Li, Bin Li, Houqiang Li, and Yan Lu. Generative latent coding for ultra-low bitrate image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26088-26098, 2024. 3, 4, 6, 7 +[27] Wei Jiang and Ronggang Wang. Mlic++: Linear complexity multi-reference entropy modeling for learned image compression. In ICML 2023 Workshop Neural Compression: From Information Theory to Applications, 2023. 7, 8 +[28] Wei Jiang, Jiayu Yang, Yongqi Zhai, Peirong Ning, Feng Gao, and Ronggang Wang. Mlic: Multi-reference entropy + +model for learned image compression. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7618-7627, 2023. 2, 3, 4 +[29] Álvaro Barbero Jiménez. Mixture of diffusers for scene composition and high resolution image generation. arXiv preprint arXiv:2302.02412, 2023. 8 +[30] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3 +[31] Nikolai Körber, Eduard Kromer, Andreas Siebert, Sascha Hauke, Daniel Mueller-Gritschneder, and Björn Schuller. Egid: enhanced low-bit-rate generative image compression guided by semantic segmentation. In European Conference on Computer Vision, pages 202-220. Springer, 2024. 2, 3, 7 +[32] Nikolai Körber, Eduard Kromer, Andreas Siebert, Sascha Hauke, Daniel Mueller-Gritschneder, and Björn Schuller. Perco (sd): Open perceptual compression. arXiv preprint arXiv:2409.20255, 2024. 7 +[33] Nupur Kumari, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Ensembling off-the-shelf models for gan training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10651-10662, 2022. 6 +[34] Hagyeong Lee, Minkyu Kim, Jun-Hyuk Kim, Seungeon Kim, Dokwan Oh, and Jaeho Lee. Neural image compression with text-guided encoding for both pixel-level and perceptual fidelity. arXiv preprint arXiv:2403.02944, 2024. 5, 7 +[35] Eric Lei, Yigit Berkay Uslu, Hamed Hassani, and Shirin Saeedi Bidokhti. Text+ sketch: Image compression at ultra low rates. arXiv preprint arXiv:2307.01944, 2023. 2, 3, 7, 8 +[36] Jiahao Li, Bin Li, and Yan Lu. Neural video compression with diverse contexts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22616-22626, 2023. 4 +[37] Zhuoyuan Li, Yao Li, Chuanbo Tang, Li Li, Dong Liu, and Feng Wu. Uniformly accelerated motion model for inter prediction. In 2024 IEEE International Conference on Visual Communications and Image Processing (VCIP), pages 1-5. IEEE, 2024. 3 +[38] Zhuoyuan Li, Junqi Liao, Chuanbo Tang, Haotian Zhang, Yuqi Li, Yifan Bian, Xihua Sheng, Xinmin Feng, Yao Li, Changsheng Gao, et al. Ustc-td: A test dataset and benchmark for image and video coding in 2020s. arXiv preprint arXiv:2409.08481, 2024. 1 +[39] Zhuoyuan Li, Zikun Yuan, Li Li, Dong Liu, Xiaohu Tang, and Feng Wu. Object segmentation-assisted inter prediction for versatile video coding. IEEE Transactions on Broadcasting, 2024. 3 +[40] Zhiyuan Li, Yanhui Zhou, Hao Wei, Chenyang Ge, and Jingwen Jiang. Towards extreme image compression with latent feature guidance and diffusion prior. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1, 2, 3, 4, 6, 7, 8 +[41] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 6 + +[42] Jinming Liu, Heming Sun, and Jiro Katto. Learned image compression with mixed transformer-cnn architectures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14388-14397, 2023. 2, 3, 4 +[43] Fabian Mentzer, George D Toderici, Michael Tschannen, and Eirikur Agustsson. High-fidelity generative image compression. Advances in Neural Information Processing Systems, 33:11913-11924, 2020. 2, 3, 7, 8 +[44] David Minnen and Saurabh Singh. Channel-wise autoregressive entropy models for learned image compression. In 2020 IEEE International Conference on Image Processing (ICIP), pages 3339-3343. IEEE, 2020. 2, 3, 4, 5 +[45] David Minnen, Johannes Balle, and George D Toderici. Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems, 31, 2018. 2, 3 +[46] Matthew J Muckley, Alaaeldin El-Nouby, Karen Ullrich, Hervé Jégou, and Jakob Verbeek. Improving statistical fidelity for neural image compression with implicit local likelihood models. In International Conference on Machine Learning, pages 25426-25443. PMLR, 2023. 1, 2, 3, 6, 7, 8 +[47] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6 +[48] Gaurav Parmar, Taesung Park, Srinivasa Narasimhan, and Jun-Yan Zhu. One-step image translation with text-to-image models. arXiv preprint arXiv:2403.12036, 2024. 3 +[49] Yichen Qian, Ming Lin, Xiuyu Sun, Zhiyu Tan, and Rong Jin. Entroformer: A transformer-based entropy model for learned image compression. arXiv preprint arXiv:2202.05492, 2022. 3, 4 +[50] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5 +[51] Lucas Relic, Roberto Azevedo, Markus Gross, and Christopher Schroers. Lossy image compression with foundation diffusion models. In European Conference on Computer Vision, pages 303-319. Springer, 2024. 2, 3 +[52] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 4 +[53] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 3 +[54] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87-103. Springer, 2024. 2, 3 +[55] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using + +nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 3 +[56] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3 +[57] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023. 3 +[58] Gary J Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. Overview of the high efficiency video coding (hevc) standard. IEEE Transactions on circuits and systems for video technology, 22(12):1649-1668, 2012. 3 +[59] Lucas Theis, Tim Salimans, Matthew D Hoffman, and Fabian Mentzer. Lossy compression with gaussian diffusion. arXiv preprint arXiv:2206.08889, 2022. 2, 3 +[60] George Toderici, Lucas Theis, Nick Johnston, Eirikur Agustsson, Fabian Mentzer, Johannes Ballé, Wenzhe Shi, and Radu Timofte. Clic 2020: Challenge on learned image compression, 2020, 2020. 3, 6 +[61] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 3 +[62] Gregory K Wallace. The jpeg still picture compression standard. Communications of the ACM, 34(4):30-44, 1991. 2, 3 +[63] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, pages 1-21, 2024. 3, 8 +[64] Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for image quality assessment. In The Thrity-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, pages 1398-1402. IEEE, 2003. 6 +[65] Tongda Xu, Ziran Zhu, Dailan He, Yanghao Li, Lina Guo, Yuanyuan Wang, Zhe Wang, Hongwei Qin, Yan Wang, Jingjing Liu, et al. Idempotence and perceptual image compression. arXiv preprint arXiv:2401.08920, 2024. 2, 3 +[66] Naifu Xue, Zhaoyang Jia, Jiahao Li, Bin Li, Yuan Zhang, and Yan Lu. Dlf: Extreme image compression with dual-generative latent fusion. arXiv preprint arXiv:2503.01428, 2025. 2 +[67] Zeyu Yan, Fei Wen, Rendong Ying, Chao Ma, and Peilin Liu. On perceptual lossy compression: The cost of perceptual reconstruction and an optimal training framework. In International Conference on Machine Learning, pages 11682-11692. PMLR, 2021. 2, 3 +[68] Zeyu Yan, Fei Wen, and Peilin Liu. Optimally controllable perceptual lossy compression. arXiv preprint arXiv:2206.10082, 2022. 2, 3 +[69] Ruihan Yang and Stephan Mandt. Lossy image compression with conditional diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2, 3 +[70] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. arXiv preprint arXiv:2405.14867, 2024. 3 + +[71] Weihao Yu and Xinchao Wang. Mambaout: Do we really need mamba for vision? arXiv preprint arXiv:2405.07992, 2024. 4 +[72] Weihao Yu, Pan Zhou, Shuicheng Yan, and Xinchao Wang. Inceptionnext: When inception meets convnext. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5672-5683, 2024. 4 +[73] Aiping Zhang, Zongsheng Yue, Renjing Pei, Wenqi Ren, and Xiaochun Cao. Degradation-guided one-step image super-resolution with diffusion priors. arXiv preprint arXiv:2409.17058, 2024. 2, 3, 4, 5 +[74] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5, 6 \ No newline at end of file diff --git a/2025/StableCodec_ Taming One-Step Diffusion for Extreme Image Compression/images.zip b/2025/StableCodec_ Taming One-Step Diffusion for Extreme Image Compression/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b336c7a9599388641026fef44c1f6fcdc5e9c222 --- /dev/null +++ b/2025/StableCodec_ Taming One-Step Diffusion for Extreme Image Compression/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:879468103b3c28bc855b7bb1a5b0ef4c8e288c99cbdd544114e82555e810cd62 +size 1139542 diff --git a/2025/StableCodec_ Taming One-Step Diffusion for Extreme Image Compression/layout.json b/2025/StableCodec_ Taming One-Step Diffusion for Extreme Image Compression/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9390423a8cf1e53a83a3f7e7993d42f08299fe3d --- /dev/null +++ b/2025/StableCodec_ Taming One-Step Diffusion for Extreme Image Compression/layout.json @@ -0,0 +1,9806 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 75, + 103, + 536, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 103, + 536, + 122 + ], + "spans": [ + { + "bbox": [ + 75, + 103, + 536, + 122 + ], + "type": "text", + "content": "StableCodec: Taming One-Step Diffusion for Extreme Image Compression" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 173, + 144, + 436, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 144, + 436, + 158 + ], + "spans": [ + { + "bbox": [ + 173, + 144, + 436, + 158 + ], + "type": "text", + "content": "Tianyu Zhang Xin Luo Li Li Dong Liu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 117, + 158, + 492, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 158, + 492, + 185 + ], + "spans": [ + { + "bbox": [ + 117, + 158, + 492, + 185 + ], + "type": "text", + "content": "MOE Key Laboratory of Brain-Inspired Intelligent Perception and Cognition, University of Science and Technology of China, Hefei 230093, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 121, + 187, + 486, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 187, + 486, + 199 + ], + "spans": [ + { + "bbox": [ + 121, + 187, + 486, + 199 + ], + "type": "text", + "content": "{zhangtianyu, xinluo}@mail.ustc.edu.cn, {lil1, dongeliu}@ustc.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 57, + 213, + 553, + 586 + ], + "blocks": [ + { + "bbox": [ + 57, + 213, + 553, + 586 + ], + "lines": [ + { + "bbox": [ + 57, + 213, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 57, + 213, + 553, + 586 + ], + "type": "image", + "image_path": "72f1cd571559f186a71ddf82a499acc1730b5a3132351d5605ad61029481f7a6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 587, + 555, + 622 + ], + "lines": [ + { + "bbox": [ + 55, + 587, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 555, + 622 + ], + "type": "text", + "content": "Figure 1. Visual examples and comparisons when compressing a 4K-resolution image [38] at ultra-low bitrates. The proposed Stable-Coded produces more realistic and consistent details with fewer bits. In contrast, VVC [10], ELIC [20] and MS-ILLM [46] reconstructions are blurry, while PerCo [11] and DiffEIC [40] generate inconsistent details against the original images. Best viewed on screen for details." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 410, + 630, + 459, + 643 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 410, + 630, + 459, + 643 + ], + "spans": [ + { + "bbox": [ + 410, + 630, + 459, + 643 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 673, + 298, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 673, + 298, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 673, + 298, + 715 + ], + "type": "text", + "content": "This work is supported by the Natural Science Foundation of China under Grant 62021001. We acknowledge the support of GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC. (Corresponding author: Dong Liu.)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Diffusion-based image compression has shown remarkable potential for achieving ultra-low bitrate coding (less than 0.05 bits per pixel) with high realism, by leveraging the generative priors of large pre-trained text-to-image diffu" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17379" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 360 + ], + "type": "text", + "content": "sion models. However, current approaches require a large number of denoising steps at the decoder to generate realistic results under extreme bitrate constraints, limiting their application in real-time compression scenarios. Additionally, these methods often sacrifice reconstruction fidelity, as diffusion models typically fail to guarantee pixel-level consistency. To address these challenges, we introduce StableCodec, which enables one-step diffusion for high-fidelity and high-realism extreme image compression with improved coding efficiency. To achieve ultra-low bitrates, we first develop an efficient Deep Compression Latent Codec to transmit a noisy latent representation for a single-step denoising process. We then propose a Dual-Branch Coding Structure, consisting of a pair of auxiliary encoder and decoder, to enhance reconstruction fidelity. Furthermore, we adopt end-to-end optimization with joint bitrate and pixel-level constraints. Extensive experiments on the CLIC 2020, DIV2K, and Kodak datasets demonstrate that StableCodec outperforms existing methods in terms of FID, KID and DISTS by a significant margin, even at bitrates as low as 0.005 bits per pixel, while maintaining strong fidelity. Additionally, StableCodec achieves inference speeds comparable to mainstream transform coding schemes. Code is available at our project repository." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 382, + 136, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 136, + 395 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 136, + 395 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 403, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 296, + 558 + ], + "type": "text", + "content": "Image compression is a foundational problem in signal processing. Driven by advances in digital imaging and the widespread use of social platforms, the volume of image data in modern multimedia has grown exponentially, placing increasing demands on the coding efficiency of image compression techniques. Over the past few decades, traditional CODECs such as JPEG [62] and H.266/VVC [10], along with emerging learning-based methods [4, 5, 12, 18, 19, 28, 42, 44, 45], have been widely adopted in real-world image compression applications. However, these methods are typically optimized for rate-distortion performance, and often produce unrealistic and blurry reconstructions, particularly under severe bitrate constraints, as shown in Fig. 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 559, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 296, + 714 + ], + "type": "text", + "content": "To tackle this issue, generative image compression [2, 43] optimized for human perceptual performance has gained increasing attention. These methods are evaluated based on the rate-distortion-perception tradeoff [8, 9, 67, 68], and progressively demonstrate their advantages in producing visually appealing reconstructions at lower bitrates compared to traditional CODEs or common neural CODEs. A prominent research direction [3, 21, 31, 43, 43, 46, 66] involves integrating a discriminator into the transform coding pipeline [5, 20, 44], employing adversarial training to enhance the perceptual quality of reconstructions. Motivated by the impressive generative capabilities, more researchers [11, 24, 35, 40, 51, 59, 65, 69] have begun exploring the po" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 554, + 228 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 554, + 228 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 554, + 228 + ], + "type": "image", + "image_path": "3c5ee440181aaca1b32bec22899a91ae3be7cb8eaa5c5c3a71b5a7fd47924454.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 230, + 555, + 286 + ], + "lines": [ + { + "bbox": [ + 313, + 230, + 555, + 286 + ], + "spans": [ + { + "bbox": [ + 313, + 230, + 555, + 286 + ], + "type": "text", + "content": "Figure 2. (Top) Illustration of our motivation. One-step diffusion can produce perceptually consistent results given severely corrupted images and a general prompt. (Bottom) Challenges in StableCodec: how to compress a noisy latent for one-step diffusion using ultra-low bitrates, and how to improve fidelity." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 312, + 555, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 420 + ], + "type": "text", + "content": "tential of diffusion models, particularly the generative priors in large pre-trained text-to-image (T2I) models, to compensate for severely distorted information at ultra-low bitrates while ensuring perceptually consistent generation. A recent study, PerCo [11], produces realistic results at an extreme bitrate as low as 0.003 bits per pixel (bpp) using a pretrained latent diffusion model (LDM), highlighting the potential of diffusion-based generative CODEs on image compression at more severe bitrates." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 423, + 556, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 556, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 556, + 687 + ], + "type": "text", + "content": "Despite these promising advancements, existing methods face two primary limitations inherent to diffusion models. First, they require dozens of denoising steps at the decoder to produce results with sufficient perceptual quality. Second, the reconstructions often deviate from the original images, as diffusion models typically do not guarantee reconstruction consistency. To address the first challenge, we consider leveraging the generative priors in SD-Turbo [54], a distilled version of Stable Diffusion 2.1 [52] that enables real-time image synthesis in 1 to 4 denoising steps. Following [73], we demonstrate that SD-Turbo can produce perceptually consistent reconstructions with a single-step denoising process, even for severely corrupted inputs and a general positive prompt, as shown in Fig. 2. We thus pose an intuitive question: Can we compress a noisy latent representation of the original image, which can be effectively denoised in a single-step diffusion process, using an ultra-low bitrate? Building on these insights, we present StableCodec for extreme image compression, which integrates SD-Turbo with the proposed Deep Compression Latent Codec to compress noisy latents at ultra-low bitrates for a single-step diffusion process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": "In response to the second challenge, we introduce a Dual-Branch Coding Structure with a pair of auxiliary en" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17380" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": "coder and decoder to further enhance reconstruction fidelity. Considering the limitations of the pre-trained VAE encoder on practical entropy coding and reconstruction consistency, we employ a rate-distortion-oriented auxiliary encoder to embed more entropy-aware semantic information for coding decisions. In parallel, we add an auxiliary decoder to perform structure apportionment during the decoding process, improving the generation guidance on the one-step denoising process for more consistent details. To enable end-to-end optimization, we design a two-stage training objective that jointly optimizes bitrate and pixel-level constraints." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 205, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 295, + 361 + ], + "type": "text", + "content": "Benefit from these designs, StableCodec produces high-fidelity and high-realism reconstructions at ultra-low bitrates as low as 0.005 bpp. Extensive experiments on CLIC 2020 [60], DIV2K [1] and Kodak [16] demonstrate that StableCodec sets up a new state-of-the-art performance in terms of FID [22], KID [6], and DISTS [14], outperforming existing methods by significant margins. In terms of computational complexity, StableCodec offers much faster decoding speeds compared to other diffusion-based competitors, achieving inference times comparable to those of mainstream transform coding schemes. For practical deployment, StableCodec supports inference at arbitrary resolutions with memory consumption less than 9 GB." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 361, + 246, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 361, + 246, + 372 + ], + "spans": [ + { + "bbox": [ + 67, + 361, + 246, + 372 + ], + "type": "text", + "content": "We summarize our contributions as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 373, + 295, + 529 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 373, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 373, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 56, + 373, + 295, + 422 + ], + "type": "text", + "content": "- We present StableCodec, an extreme image codec integrating one-step diffusion and Deep Compression Latent Codec to achieve ultra-low bitrate compression with high realism, high fidelity and superior coding efficiency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 422, + 295, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 295, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 295, + 469 + ], + "type": "text", + "content": "- We develop Dual-Branch Coding Structure to improve reconstruction fidelity. A pair of auxiliary encoder and decoder is introduced for semantic enhancement and structure apportionment." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 469, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 469, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 469, + 295, + 529 + ], + "type": "text", + "content": "- StableCodec obtains SOTA FID, KID and DISTS performance on CLIC 2020 and DIV2K dataset, significantly outperforms existing methods at bitrates as low as 0.005 bpp with well-preserved fidelity, and achieves comparable inference speeds with mainstream neural CODECs." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 542, + 142, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 142, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 142, + 554 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 563, + 225, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 225, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 225, + 575 + ], + "type": "text", + "content": "2.1. Generative Image Compression" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 581, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 295, + 689 + ], + "type": "text", + "content": "Learning-based image compression has shown competitive potential compared to traditional methods [10, 37, 39, 58, 62], leveraging non-linear transforms and joint rate-distortion optimization. Balle et al. [4] introduced the first end-to-end learned image compression framework, which was subsequently enhanced with the hyperprior [5] and context model [45]. Building on this foundation, much work [12, 18, 19, 28, 42, 44, 49] has been devoted to improving both rate-distortion performance and model practicality." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "In practical scenarios, a key challenge is achieving extreme image compression at ultra-low bitrates while main" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 335 + ], + "type": "text", + "content": "taining both fidelity and realism [2]. Traditional image codes optimized for rate-distortion often produce blurry reconstructions and noticeable artifacts. To address this, Mentzer et al. [43] introduced HiFiC and the concept of generative image compression, integrating GANs into codec optimization and evaluating performance in terms of the rate-distortion-perception tradeoff [8, 9, 67, 68]. Subsequent research can be broadly categorized into two main approaches. The first category [3, 21, 31, 43, 46] focuses on enhancing transform coding [5, 20, 44] for human perception by incorporating adversarial losses and optimized discriminator architectures, which typically can be extended to a wide range of bitrate and a flexible decoding control between fidelity and realism [3, 31]. The second category [11, 24, 35, 40, 51, 59, 65, 69] leverages diffusion models for generative image compression. Although these methods show promise for ultra-low bitrate compression, they are often constrained by reconstruction fidelity and inference efficiency due to the multi-step denoising process. Recently, GLC [26] introduced transform coding in the generative latent space of VQ-VAE [15, 61], achieving more visually appealing results at ultra-low bitrates." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 345, + 542, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 542, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 542, + 357 + ], + "type": "text", + "content": "2.2. Generative Models and Few-Step Diffusions" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 363, + 553, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 553, + 495 + ], + "type": "text", + "content": "Generative models play a crucial role in image generation. While many architectures, such as VAEs [30] and GANs [17], have been explored, diffusion models [55] have emerged as a powerful alternative, achieving state-of-the-art synthesis quality. Inspired by non-equilibrium statistical physics [55], diffusion models learn to reverse a noise perturbation process through a Markovian framework. Recent advancements, such as DDPM [23], DDIM [56], and LDM [52], have significantly reduced computational complexity and improved image synthesis quality, making diffusion-based approaches a dominant force in generative modeling." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 495, + 553, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 553, + 603 + ], + "type": "text", + "content": "To address the inefficiency of iterative denoising, several approaches [53, 54, 57, 70] aim to reduce the number of denoising steps while maintaining generation quality. These methods train diffusion models to approximate the full denoising trajectory in a single or a few steps, significantly improving inference efficiency. Notably, SD/SDXL-Turbo [54] demonstrates image generation in 1 to 4 steps with near-parity quality compared to multi-step models, making real-time diffusion-based applications [48, 63, 73] feasible." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 615, + 370, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 370, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 370, + 628 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 635, + 381, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 381, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 381, + 647 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "In this section, we introduce the overall framework of the proposed StableCodec, built upon SD-Turbo [54] with a VAE encoder " + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{SD}}" + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": ", a VAE decoder " + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{SD}}" + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": " and a denoising Unet " + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{SD}}" + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": ". As shown in Fig. 3, we incorporate a Deep Compression Latent Codec to perform extreme transform" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17381" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 558, + 284 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 558, + 284 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 558, + 284 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 558, + 284 + ], + "type": "image", + "image_path": "fb19566377769b1588ee11eede0d60fe9da173e08c2d511ccb2532a752adc0e2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 285, + 555, + 330 + ], + "lines": [ + { + "bbox": [ + 54, + 285, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 54, + 285, + 555, + 330 + ], + "type": "text", + "content": "Figure 3. The framework of StableCodec. We incorporate the proposed Deep Compression Latent Codec to transmit a noisy latent " + }, + { + "bbox": [ + 54, + 285, + 555, + 330 + ], + "type": "inline_equation", + "content": "l_{T}" + }, + { + "bbox": [ + 54, + 285, + 555, + 330 + ], + "type": "text", + "content": " for one-step denoising, where " + }, + { + "bbox": [ + 54, + 285, + 555, + 330 + ], + "type": "inline_equation", + "content": "64 \\times" + }, + { + "bbox": [ + 54, + 285, + 555, + 330 + ], + "type": "text", + "content": " denotes a spatial compression ratio of 64. To adjust the latent resolution, we deploy DownSample block and Conv " + }, + { + "bbox": [ + 54, + 285, + 555, + 330 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 54, + 285, + 555, + 330 + ], + "type": "text", + "content": " as adapters after the VAE encoder " + }, + { + "bbox": [ + 54, + 285, + 555, + 330 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{SD}}" + }, + { + "bbox": [ + 54, + 285, + 555, + 330 + ], + "type": "text", + "content": " and auxiliary encoder " + }, + { + "bbox": [ + 54, + 285, + 555, + 330 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{Aux}}" + }, + { + "bbox": [ + 54, + 285, + 555, + 330 + ], + "type": "text", + "content": ", respectively. We use a general prompt in both training and inference. The auxiliary decoder " + }, + { + "bbox": [ + 54, + 285, + 555, + 330 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{Aux}}" + }, + { + "bbox": [ + 54, + 285, + 555, + 330 + ], + "type": "text", + "content": " shares a similar structure with " + }, + { + "bbox": [ + 54, + 285, + 555, + 330 + ], + "type": "inline_equation", + "content": "g_{s}" + }, + { + "bbox": [ + 54, + 285, + 555, + 330 + ], + "type": "text", + "content": ". More details on networks are provided in the supplementary." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 350, + 294, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 350, + 294, + 469 + ], + "spans": [ + { + "bbox": [ + 54, + 350, + 294, + 469 + ], + "type": "text", + "content": "coding in the VAE latent space. To adapt SD-Turbo for image compression, we integrate LoRA [25] into " + }, + { + "bbox": [ + 54, + 350, + 294, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{SD}}" + }, + { + "bbox": [ + 54, + 350, + 294, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 350, + 294, + 469 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{SD}}" + }, + { + "bbox": [ + 54, + 350, + 294, + 469 + ], + "type": "text", + "content": ", while keeping " + }, + { + "bbox": [ + 54, + 350, + 294, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{SD}}" + }, + { + "bbox": [ + 54, + 350, + 294, + 469 + ], + "type": "text", + "content": " unchanged to preserve the generative priors [73]. Additionally, we introduce a Dual-Branch Coding Structure to enhance reconstruction fidelity, utilizing an auxiliary encoder " + }, + { + "bbox": [ + 54, + 350, + 294, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{Aux}}" + }, + { + "bbox": [ + 54, + 350, + 294, + 469 + ], + "type": "text", + "content": " to embed rich semantic information and an auxiliary decoder " + }, + { + "bbox": [ + 54, + 350, + 294, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{Aux}}" + }, + { + "bbox": [ + 54, + 350, + 294, + 469 + ], + "type": "text", + "content": " to perform structure apportionment. Finally, we optimize StableCodec end-to-end with joint bitrate and pixel-level constraints, achieving high-fidelity and high-realism extreme image compression." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 479, + 231, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 231, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 231, + 491 + ], + "type": "text", + "content": "3.2. Deep Compression Latent Codec" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 497, + 296, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 296, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 296, + 640 + ], + "type": "text", + "content": "We design our latent codec using the extreme analysis transform " + }, + { + "bbox": [ + 55, + 497, + 296, + 640 + ], + "type": "inline_equation", + "content": "g_{a}" + }, + { + "bbox": [ + 55, + 497, + 296, + 640 + ], + "type": "text", + "content": ", extreme synthesis transform " + }, + { + "bbox": [ + 55, + 497, + 296, + 640 + ], + "type": "inline_equation", + "content": "g_{s}" + }, + { + "bbox": [ + 55, + 497, + 296, + 640 + ], + "type": "text", + "content": " and a 4-step autoregressive entropy model. To reach ultra-low bitrates, we employ deep compression transform networks for both " + }, + { + "bbox": [ + 55, + 497, + 296, + 640 + ], + "type": "inline_equation", + "content": "g_{a}" + }, + { + "bbox": [ + 55, + 497, + 296, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 497, + 296, + 640 + ], + "type": "inline_equation", + "content": "g_{s}" + }, + { + "bbox": [ + 55, + 497, + 296, + 640 + ], + "type": "text", + "content": ". Specifically, " + }, + { + "bbox": [ + 55, + 497, + 296, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{SD}}" + }, + { + "bbox": [ + 55, + 497, + 296, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 497, + 296, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{SD}}" + }, + { + "bbox": [ + 55, + 497, + 296, + 640 + ], + "type": "text", + "content": " provide a latent space with a spatial compression ratio of 8 (abbreviated as " + }, + { + "bbox": [ + 55, + 497, + 296, + 640 + ], + "type": "inline_equation", + "content": "8\\times" + }, + { + "bbox": [ + 55, + 497, + 296, + 640 + ], + "type": "text", + "content": "). Unlike mainstream schemes [18, 20, 26, 28, 40, 42, 49] that perform entropy coding at " + }, + { + "bbox": [ + 55, + 497, + 296, + 640 + ], + "type": "inline_equation", + "content": "16\\times" + }, + { + "bbox": [ + 55, + 497, + 296, + 640 + ], + "type": "text", + "content": ", we further downsample and apply entropy coding for " + }, + { + "bbox": [ + 55, + 497, + 296, + 640 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 55, + 497, + 296, + 640 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 55, + 497, + 296, + 640 + ], + "type": "inline_equation", + "content": "64\\times" + }, + { + "bbox": [ + 55, + 497, + 296, + 640 + ], + "type": "text", + "content": " and the hyperprior [5] at " + }, + { + "bbox": [ + 55, + 497, + 296, + 640 + ], + "type": "inline_equation", + "content": "256\\times" + }, + { + "bbox": [ + 55, + 497, + 296, + 640 + ], + "type": "text", + "content": ". Consequently, we use " + }, + { + "bbox": [ + 55, + 497, + 296, + 640 + ], + "type": "inline_equation", + "content": "g_{s}" + }, + { + "bbox": [ + 55, + 497, + 296, + 640 + ], + "type": "text", + "content": " to restore the spatial compression ratio to " + }, + { + "bbox": [ + 55, + 497, + 296, + 640 + ], + "type": "inline_equation", + "content": "8\\times" + }, + { + "bbox": [ + 55, + 497, + 296, + 640 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 497, + 296, + 640 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{SD}}" + }, + { + "bbox": [ + 55, + 497, + 296, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 497, + 296, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{SD}}" + }, + { + "bbox": [ + 55, + 497, + 296, + 640 + ], + "type": "text", + "content": ". The entire coding process can be formulated as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 651, + 295, + 664 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 651, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 89, + 651, + 295, + 664 + ], + "type": "interline_equation", + "content": "l = \\operatorname {c o n c a t} \\left[ \\mathcal {E} _ {\\mathrm {S D}} (x), \\mathcal {E} _ {\\text {A u x}} (x) \\right] \\tag {1}", + "image_path": "1bbc3dff7bb817bed2d270d90fb8ff724dc07cf175dab47552609a3dd541d373.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 88, + 666, + 295, + 679 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 666, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 88, + 666, + 295, + 679 + ], + "type": "interline_equation", + "content": "y = g _ {a} (l), \\hat {y} = Q (y), l _ {T} = g _ {s} (\\hat {y}) \\tag {2}", + "image_path": "b591e2065fb6437b7ef35339d76efa968bc2fae1a7d114f4e2bafccbc8837a8d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 88, + 682, + 294, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 682, + 294, + 696 + ], + "spans": [ + { + "bbox": [ + 88, + 682, + 294, + 696 + ], + "type": "interline_equation", + "content": "l _ {0} = \\left[ l _ {T} - \\sqrt {1 - \\bar {\\alpha} _ {T}} \\cdot \\epsilon_ {\\mathrm {S D}} \\left(l _ {T}, T\\right) \\right] / \\sqrt {\\bar {\\alpha} _ {T}} \\tag {3}", + "image_path": "60f0cf4e0933d8cd964fce43dfe8b7f10b84f75951d22dc808b2f3ec6c167342.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 88, + 697, + 295, + 710 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 697, + 295, + 710 + ], + "spans": [ + { + "bbox": [ + 88, + 697, + 295, + 710 + ], + "type": "interline_equation", + "content": "\\hat {x} = \\mathcal {D} _ {\\mathrm {S D}} \\left(l _ {0} + \\mathcal {D} _ {\\mathrm {A u x}} (\\hat {y})\\right) \\tag {4}", + "image_path": "9420617c14f17b93cf04bc52c9efe0e239446058ac2d82500d577cf7a8cb12e7.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "type": "text", + "content": "In Eq. (1), we first obtain an intermediate latent " + }, + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "type": "text", + "content": " from the input image " + }, + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{SD}}" + }, + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{Aux}}" + }, + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "type": "text", + "content": ". Eq. (2) is the latent-space transform coding process to produce a noisy latent " + }, + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "type": "inline_equation", + "content": "l_{T}" + }, + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "type": "text", + "content": " using ultra-low bitrates. Eq. (3) displays the one-step denoising process with the noise schedule " + }, + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\{\\bar{\\alpha}_t\\}" + }, + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "type": "text", + "content": " [23] in the " + }, + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "type": "text", + "content": "-th timestep. Finally, in Eq. (4), the reconstruction " + }, + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "type": "text", + "content": " is obtained from " + }, + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "type": "inline_equation", + "content": "l_{0}" + }, + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{SD}}" + }, + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{Aux}}" + }, + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "type": "text", + "content": ". To balance performance and coding latency, we construct efficient " + }, + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "type": "inline_equation", + "content": "g_{a}" + }, + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "type": "inline_equation", + "content": "g_{s}" + }, + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "type": "text", + "content": " with InceptionNeXt [72] and GatedCNN [71], and build a 4-step antoregressive entropy model with quadtree partition [36] and latent residual prediction [44]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 491, + 484, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 484, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 484, + 504 + ], + "type": "text", + "content": "3.3. Dual-Branch Coding Structure" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 509, + 555, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 583 + ], + "type": "text", + "content": "While deploying the proposed latent codec with LoRA enables image compression with SD-Turbo at ultra-low bitrates, the reconstruction fidelity is limited. In this section, we analyze the reasons and introduce Dual-Branch Coding Structure with a pair of auxiliary encoder and decoder, " + }, + { + "bbox": [ + 313, + 509, + 555, + 583 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{Aux}}" + }, + { + "bbox": [ + 313, + 509, + 555, + 583 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 509, + 555, + 583 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{Aux}}" + }, + { + "bbox": [ + 313, + 509, + 555, + 583 + ], + "type": "text", + "content": ", to further enhance compression performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 590, + 512, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 512, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 512, + 601 + ], + "type": "text", + "content": "3.3.1. Entropy-Aware Semantic Enhancement" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "We observed that the VAE in SD-Turbo has several limitations when reconstruction fidelity and practical coding are required. As noted in [52] and Table 1, this VAE is pretrained for perceptual compression, which does not preserve pixel-level fidelity as well as a rate-distortion-oriented autoencoder from a typical neural codec [20]. Additionally, while this VAE provides a compressed representation of the original image, it is still in floating-point format and not optimized for practical entropy coding, making it less suitable" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "17382" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 106, + 288, + 149 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 103 + ], + "type": "text", + "content": "Table 1. Reconstruction quality of VAEs on Kodak [16]. The pre-trained VAE in SD performs perceptual compression, while the one in ELIC [20] preserves more pixel-level information." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 106, + 288, + 149 + ], + "lines": [ + { + "bbox": [ + 63, + 106, + 288, + 149 + ], + "spans": [ + { + "bbox": [ + 63, + 106, + 288, + 149 + ], + "type": "table", + "html": "
VAEPSNR↑MS-SSIM↑LPIPS↓DISTS↓
SD26.650.93180.07260.0415
ELIC40.400.99610.05550.0707
", + "image_path": "97af191850c02dfc6d01df514b96753e378bba34b929e6d7fda6cd7255e46969.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 59, + 159, + 297, + 277 + ], + "blocks": [ + { + "bbox": [ + 59, + 159, + 297, + 277 + ], + "lines": [ + { + "bbox": [ + 59, + 159, + 297, + 277 + ], + "spans": [ + { + "bbox": [ + 59, + 159, + 297, + 277 + ], + "type": "image", + "image_path": "65ec16377b6b7a4e67fe570150e2341b979702bc630cc024376c5898daf53581.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 278, + 295, + 300 + ], + "lines": [ + { + "bbox": [ + 55, + 278, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 295, + 300 + ], + "type": "text", + "content": "Figure 4. Top-energy channels learned from different encoders. " + }, + { + "bbox": [ + 55, + 278, + 295, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{Aux}}" + }, + { + "bbox": [ + 55, + 278, + 295, + 300 + ], + "type": "text", + "content": " embeds more pixel-level semantic information into latents." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 322, + 272, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 322, + 272, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 272, + 333 + ], + "type": "text", + "content": "for further latent-space ultra-low bitrate compression." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 334, + 296, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 334, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 54, + 334, + 296, + 465 + ], + "type": "text", + "content": "Building on these insights, we introduce the analysis transform of a pre-trained high-bitrate ELIC model [20] to serve as an auxiliary encoder " + }, + { + "bbox": [ + 54, + 334, + 296, + 465 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{Aux}}" + }, + { + "bbox": [ + 54, + 334, + 296, + 465 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 54, + 334, + 296, + 465 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{Aux}}" + }, + { + "bbox": [ + 54, + 334, + 296, + 465 + ], + "type": "text", + "content": " remains frozen, and provides rich, entropy-aware semantic information of the input images. To combine the latents from different encoders, we introduce trainable adapters that align the latent resolutions to a spatial compression ratio of 16, followed by channel-wise concatenation. During optimization, various pieces of information are learned from different encoders as shown in Fig. 4, where more pixel-level semantic information is embedded into the latent codec through " + }, + { + "bbox": [ + 54, + 334, + 296, + 465 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{Aux}}" + }, + { + "bbox": [ + 54, + 334, + 296, + 465 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 472, + 194, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 194, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 194, + 483 + ], + "type": "text", + "content": "3.3.2. Structure Apportionment" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 486, + 296, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 569 + ], + "type": "text", + "content": "The reconstruction quality of StableCodec heavily depends on how the denoising Unet is conditioned. Since a fixed prompt is used for both training and inference, the one-step denoising process is primarily guided by " + }, + { + "bbox": [ + 55, + 486, + 296, + 569 + ], + "type": "inline_equation", + "content": "l_{T}" + }, + { + "bbox": [ + 55, + 486, + 296, + 569 + ], + "type": "text", + "content": ", which is produced by " + }, + { + "bbox": [ + 55, + 486, + 296, + 569 + ], + "type": "inline_equation", + "content": "g_{s}" + }, + { + "bbox": [ + 55, + 486, + 296, + 569 + ], + "type": "text", + "content": ". This places high demands on the capability of " + }, + { + "bbox": [ + 55, + 486, + 296, + 569 + ], + "type": "inline_equation", + "content": "g_{s}" + }, + { + "bbox": [ + 55, + 486, + 296, + 569 + ], + "type": "text", + "content": ", resulting in unsatisfactory denoising guidance, as reflected in the reconstructions shown in Fig. 5 (b)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "To alleviate the decoding burden on " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "g_{s}" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": ", we introduce an auxiliary decoder " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{Aux}}" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " to perform an additional decoding branch from " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": ", bypassing the Unet. This design is motivated by the observation that an extremely compressed bitstream contains mainly the basic structure of images. We distribute and decode these components directly from the bitstream using the auxiliary branch, allowing " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "g_{s}" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " to focus primarily on producing guidance to generate realistic details. Fig. 5 compares StableCodec that is trained with or without " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{Aux}}" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": ". When trained without " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{Aux}}" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "g_{s}" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " produces all types of information as it is the only decoding branch. When trained with " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{Aux}}" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": ", structural information is routed" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 69, + 557, + 247 + ], + "blocks": [ + { + "bbox": [ + 317, + 69, + 557, + 247 + ], + "lines": [ + { + "bbox": [ + 317, + 69, + 557, + 247 + ], + "spans": [ + { + "bbox": [ + 317, + 69, + 557, + 247 + ], + "type": "image", + "image_path": "83e18420fabd9eae154e1ae6c87af76a385445279d9f2cc6b87ff24af0f64499.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 247, + 528, + 256 + ], + "lines": [ + { + "bbox": [ + 345, + 247, + 528, + 256 + ], + "spans": [ + { + "bbox": [ + 345, + 247, + 528, + 256 + ], + "type": "text", + "content": "(a) Latents from different decoders, trained w. or w/o " + }, + { + "bbox": [ + 345, + 247, + 528, + 256 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{Aux}}" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 326, + 258, + 547, + 316 + ], + "blocks": [ + { + "bbox": [ + 326, + 258, + 547, + 316 + ], + "lines": [ + { + "bbox": [ + 326, + 258, + 547, + 316 + ], + "spans": [ + { + "bbox": [ + 326, + 258, + 547, + 316 + ], + "type": "image", + "image_path": "25ba7824b0160ae6f6194eedcf32c522d1eb40166ef0338a518723382f6cf250.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 316, + 522, + 324 + ], + "lines": [ + { + "bbox": [ + 351, + 316, + 522, + 324 + ], + "spans": [ + { + "bbox": [ + 351, + 316, + 522, + 324 + ], + "type": "text", + "content": "(b) Reconstruction comparison, trained w. or w/o " + }, + { + "bbox": [ + 351, + 316, + 522, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{Aux}}" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 324, + 326, + 543, + 337 + ], + "lines": [ + { + "bbox": [ + 324, + 326, + 543, + 337 + ], + "spans": [ + { + "bbox": [ + 324, + 326, + 543, + 337 + ], + "type": "text", + "content": "Figure 5. Impact of " + }, + { + "bbox": [ + 324, + 326, + 543, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{Aux}}" + }, + { + "bbox": [ + 324, + 326, + 543, + 337 + ], + "type": "text", + "content": " on latents and reconstructions." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 357, + 554, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 554, + 407 + ], + "type": "text", + "content": "through " + }, + { + "bbox": [ + 313, + 357, + 554, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{Aux}}" + }, + { + "bbox": [ + 313, + 357, + 554, + 407 + ], + "type": "text", + "content": ", while the energy in " + }, + { + "bbox": [ + 313, + 357, + 554, + 407 + ], + "type": "inline_equation", + "content": "g_{s}" + }, + { + "bbox": [ + 313, + 357, + 554, + 407 + ], + "type": "text", + "content": " latents drops significantly. Meanwhile, more semantically aligned details are reconstructed in Fig. 5 (b), suggesting " + }, + { + "bbox": [ + 313, + 357, + 554, + 407 + ], + "type": "inline_equation", + "content": "g_{s}" + }, + { + "bbox": [ + 313, + 357, + 554, + 407 + ], + "type": "text", + "content": " now provides better high-frequency guidance for the denoising Unet " + }, + { + "bbox": [ + 313, + 357, + 554, + 407 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{SD}}" + }, + { + "bbox": [ + 313, + 357, + 554, + 407 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 413, + 484, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 484, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 484, + 426 + ], + "type": "text", + "content": "3.4. End-to-End Training Objective" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 430, + 554, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 554, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 554, + 489 + ], + "type": "text", + "content": "We adopt end-to-end optimization with joint bitrate and pixel-level restrictions to train StableCodec. Given the original image " + }, + { + "bbox": [ + 313, + 430, + 554, + 489 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 430, + 554, + 489 + ], + "type": "text", + "content": ", the quantized latent " + }, + { + "bbox": [ + 313, + 430, + 554, + 489 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 313, + 430, + 554, + 489 + ], + "type": "text", + "content": " and the reconstructed image " + }, + { + "bbox": [ + 313, + 430, + 554, + 489 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 313, + 430, + 554, + 489 + ], + "type": "text", + "content": ", we construct our training objective based on the standard rate-distortion loss:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 397, + 500, + 553, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 500, + 553, + 513 + ], + "spans": [ + { + "bbox": [ + 397, + 500, + 553, + 513 + ], + "type": "interline_equation", + "content": "\\lambda \\mathcal {R} (\\hat {y}) + \\mathcal {D} (x, \\hat {x}) \\tag {5}", + "image_path": "71f29e5e25cea10bfa2c6477d2e50bd74ef135dc5ad288077804dcc08b67df41.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "type": "text", + "content": "where the bitrate " + }, + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "type": "text", + "content": " and pixel-level distortion " + }, + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "type": "text", + "content": " are balanced by the Lagrange multiplier " + }, + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 546, + 554, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 554, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 554, + 641 + ], + "type": "text", + "content": "Inspired by [20, 44], we train StableCodec with a 2-stage implicit bitrate pruning (IBP) strategy. We first train a base model using a smaller " + }, + { + "bbox": [ + 313, + 546, + 554, + 641 + ], + "type": "inline_equation", + "content": "\\lambda_{base}" + }, + { + "bbox": [ + 313, + 546, + 554, + 641 + ], + "type": "text", + "content": ", adapting the latent codec into SD-Turbo under a relaxed bitrate constraint, and warming up with a more expressive transform. In the second stage, we finetune the shared base model with larger " + }, + { + "bbox": [ + 313, + 546, + 554, + 641 + ], + "type": "inline_equation", + "content": "\\lambda_{target}" + }, + { + "bbox": [ + 313, + 546, + 554, + 641 + ], + "type": "text", + "content": " to reach ultra-low target bitrates. IBP facilitates efficient and stable training, resulting in improved performance." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "The distortion term " + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": " includes MSE, LPIPS (using VGG features) [74] and a CLIP [50] distance " + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CLIP}" + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": ", for which we compute the L2-distance between the CLIP embeddings of " + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": ". Note that this term is a simplified version from [34], and we find it beneficial for reconstruction at ultralow bitrates. Additionally, we follow [73] and incorporate" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17383" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 118, + 74, + 491, + 105 + ], + "blocks": [ + { + "bbox": [ + 118, + 74, + 491, + 105 + ], + "lines": [ + { + "bbox": [ + 118, + 74, + 491, + 105 + ], + "spans": [ + { + "bbox": [ + 118, + 74, + 491, + 105 + ], + "type": "image", + "image_path": "7f9d7f01a147b11667990979775e443d604673ba0ed3b3a58bac5fc972014eef.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 109, + 179, + 205 + ], + "blocks": [ + { + "bbox": [ + 59, + 109, + 179, + 205 + ], + "lines": [ + { + "bbox": [ + 59, + 109, + 179, + 205 + ], + "spans": [ + { + "bbox": [ + 59, + 109, + 179, + 205 + ], + "type": "image", + "image_path": "8b2210b11c21bd44326ced02d09aa107f91421454626268e90c0dbe2af2f5d21.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 179, + 109, + 302, + 205 + ], + "blocks": [ + { + "bbox": [ + 179, + 109, + 302, + 205 + ], + "lines": [ + { + "bbox": [ + 179, + 109, + 302, + 205 + ], + "spans": [ + { + "bbox": [ + 179, + 109, + 302, + 205 + ], + "type": "image", + "image_path": "11fe770e88e21dfb7f2d021c8cee65889c7c31f74963780c0ba1a74b7f2ee2f4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 303, + 109, + 427, + 205 + ], + "blocks": [ + { + "bbox": [ + 303, + 109, + 427, + 205 + ], + "lines": [ + { + "bbox": [ + 303, + 109, + 427, + 205 + ], + "spans": [ + { + "bbox": [ + 303, + 109, + 427, + 205 + ], + "type": "image", + "image_path": "81019f026d7b49fccdc154c011b104618da55869cf16c5e6061794d685f35b0a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 429, + 109, + 551, + 205 + ], + "blocks": [ + { + "bbox": [ + 429, + 109, + 551, + 205 + ], + "lines": [ + { + "bbox": [ + 429, + 109, + 551, + 205 + ], + "spans": [ + { + "bbox": [ + 429, + 109, + 551, + 205 + ], + "type": "image", + "image_path": "1a69dd8cbcd53a10f0feaa9ddb2a0fcc11234f4ae36a01790ddf1039dd387bcb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 58, + 206, + 179, + 302 + ], + "blocks": [ + { + "bbox": [ + 58, + 206, + 179, + 302 + ], + "lines": [ + { + "bbox": [ + 58, + 206, + 179, + 302 + ], + "spans": [ + { + "bbox": [ + 58, + 206, + 179, + 302 + ], + "type": "image", + "image_path": "c1bbb274d9ea3a973b75d174502c6912932e225677744d441bb5a0ffcc364d71.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 179, + 206, + 302, + 302 + ], + "blocks": [ + { + "bbox": [ + 179, + 206, + 302, + 302 + ], + "lines": [ + { + "bbox": [ + 179, + 206, + 302, + 302 + ], + "spans": [ + { + "bbox": [ + 179, + 206, + 302, + 302 + ], + "type": "image", + "image_path": "69ecf9d1bfe2780365a7fe85fd7d63aac2d4e435fe6a5f3a98b0468cdbe5aba4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 304, + 206, + 427, + 302 + ], + "blocks": [ + { + "bbox": [ + 304, + 206, + 427, + 302 + ], + "lines": [ + { + "bbox": [ + 304, + 206, + 427, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 206, + 427, + 302 + ], + "type": "image", + "image_path": "af1daffea9066d94a7acaec322648900023351dcf045c3c56974928194027c9f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 429, + 206, + 551, + 302 + ], + "blocks": [ + { + "bbox": [ + 429, + 206, + 551, + 302 + ], + "lines": [ + { + "bbox": [ + 429, + 206, + 551, + 302 + ], + "spans": [ + { + "bbox": [ + 429, + 206, + 551, + 302 + ], + "type": "image", + "image_path": "6ccbb99bd2b025a4802db51eae451d9d8a947ba3c9a63076e7855d37fdbd6fb5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 59, + 302, + 179, + 399 + ], + "blocks": [ + { + "bbox": [ + 59, + 302, + 179, + 399 + ], + "lines": [ + { + "bbox": [ + 59, + 302, + 179, + 399 + ], + "spans": [ + { + "bbox": [ + 59, + 302, + 179, + 399 + ], + "type": "image", + "image_path": "50664a381cec20a3f97596b197e1f90dbe127b7408e8841b6ff3534c926bdd49.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 403, + 550, + 414 + ], + "lines": [ + { + "bbox": [ + 58, + 403, + 550, + 414 + ], + "spans": [ + { + "bbox": [ + 58, + 403, + 550, + 414 + ], + "type": "text", + "content": "Figure 6. Rate-distortion and rate-perception curve comparisons of different methods on the CLIC 2020, DIV2K and Kodak dataset." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 181, + 302, + 303, + 399 + ], + "blocks": [ + { + "bbox": [ + 181, + 302, + 303, + 399 + ], + "lines": [ + { + "bbox": [ + 181, + 302, + 303, + 399 + ], + "spans": [ + { + "bbox": [ + 181, + 302, + 303, + 399 + ], + "type": "image", + "image_path": "a89037520d9c1e38d270c748a76e5e136c40ee1fc101cf61a03570260c0459b0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 304, + 302, + 427, + 399 + ], + "blocks": [ + { + "bbox": [ + 304, + 302, + 427, + 399 + ], + "lines": [ + { + "bbox": [ + 304, + 302, + 427, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 302, + 427, + 399 + ], + "type": "image", + "image_path": "741544a4de1adbd1d53be9b8cbbd9d8f8f73924c67ff2c2141275ab3d6765d6b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 429, + 302, + 551, + 399 + ], + "blocks": [ + { + "bbox": [ + 429, + 302, + 551, + 399 + ], + "lines": [ + { + "bbox": [ + 429, + 302, + 551, + 399 + ], + "spans": [ + { + "bbox": [ + 429, + 302, + 551, + 399 + ], + "type": "image", + "image_path": "b0d4541c6ae912b0dd7cecea8c98d296195ecaed058b50c6f5e62bb0fd41e1ca.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 435, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 296, + 495 + ], + "type": "text", + "content": "a GAN loss " + }, + { + "bbox": [ + 55, + 435, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{adv}" + }, + { + "bbox": [ + 55, + 435, + 296, + 495 + ], + "type": "text", + "content": " to narrow the distribution gap between " + }, + { + "bbox": [ + 55, + 435, + 296, + 495 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 435, + 296, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 435, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 55, + 435, + 296, + 495 + ], + "type": "text", + "content": ". We use DINOv2 [47] with registers [13] as the discriminator backbone [33]. To ensure stable training, we incorporate the GAN only in the second training stage. The full objective can be formulated as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 66, + 505, + 295, + 524 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 505, + 295, + 524 + ], + "spans": [ + { + "bbox": [ + 66, + 505, + 295, + 524 + ], + "type": "interline_equation", + "content": "\\text {S t a g e I :} \\underset {\\theta} {\\arg \\min } \\lambda_ {\\text {b a s e}} \\mathcal {R} (\\hat {y}) + \\mathcal {D} (x, \\hat {x}) \\tag {6}", + "image_path": "09375080ad0ff43da60c6fc72cf6bcd7549a09b03881a482146e892ed218aa87.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 63, + 526, + 295, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 526, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 63, + 526, + 295, + 546 + ], + "type": "interline_equation", + "content": "\\text {S t a g e I I :} \\underset {\\theta} {\\arg \\min } \\lambda_ {\\text {t a r g e t}} \\mathcal {R} (\\hat {y}) + \\mathcal {D} (x, \\hat {x}) + \\beta \\mathcal {L} _ {\\text {a d v}} \\tag {7}", + "image_path": "df825d0331d4b94a74dc75ea75893996c2bc80e94603fd5dd8669cb9d57e847f.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 82, + 559, + 296, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 559, + 296, + 587 + ], + "spans": [ + { + "bbox": [ + 82, + 559, + 296, + 587 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {D} (x, \\hat {x}) = d _ {1} M S E (x, \\hat {x}) + d _ {2} L P I P S (x, \\hat {x}) \\tag {8} \\\\ + d _ {3} \\mathcal {L} _ {C L I P} (x, \\hat {x}) \\\\ \\end{array}", + "image_path": "a7ea806f3c04cf8f00881948ffb07a8c7cec87d2bb20f17a990206fe04f40102.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 593, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 295, + 617 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 593, + 295, + 617 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 593, + 295, + 617 + ], + "type": "text", + "content": " represents all trainable parameters in StableCodec (Fig. 3), " + }, + { + "bbox": [ + 55, + 593, + 295, + 617 + ], + "type": "inline_equation", + "content": "d_{1}, d_{2}, d_{3}" + }, + { + "bbox": [ + 55, + 593, + 295, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 593, + 295, + 617 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 593, + 295, + 617 + ], + "type": "text", + "content": " are balancing weights." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 628, + 137, + 642 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 137, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 137, + 642 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 648, + 153, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 153, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 153, + 661 + ], + "type": "text", + "content": "4.1. Implementation" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Training Details. We use the training set of DF2K [41] and CLIC 2020 Professional [60] to train StableCodec. During training, we use " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": " patches with a batch size of 8. The first training stage takes over 100k iterations with a" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 435, + 553, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 553, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 553, + 519 + ], + "type": "text", + "content": "learning rate of " + }, + { + "bbox": [ + 313, + 435, + 553, + 519 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 313, + 435, + 553, + 519 + ], + "type": "text", + "content": " and a " + }, + { + "bbox": [ + 313, + 435, + 553, + 519 + ], + "type": "inline_equation", + "content": "\\lambda_{base}" + }, + { + "bbox": [ + 313, + 435, + 553, + 519 + ], + "type": "text", + "content": " of 0.5 (about 0.05bpp). In the second stage, we finetune the base model for another 20k iterations with GAN incorporated and " + }, + { + "bbox": [ + 313, + 435, + 553, + 519 + ], + "type": "inline_equation", + "content": "\\lambda_{target} \\in \\{2,3,4,6,8,12,16,24,32\\}" + }, + { + "bbox": [ + 313, + 435, + 553, + 519 + ], + "type": "text", + "content": ", while the learning rate undergoes " + }, + { + "bbox": [ + 313, + 435, + 553, + 519 + ], + "type": "inline_equation", + "content": "5e^{-5}, 2e^{-5}, 1e^{-5}" + }, + { + "bbox": [ + 313, + 435, + 553, + 519 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 435, + 553, + 519 + ], + "type": "inline_equation", + "content": "1e^{-6}" + }, + { + "bbox": [ + 313, + 435, + 553, + 519 + ], + "type": "text", + "content": " for 5k iterations each. We set " + }, + { + "bbox": [ + 313, + 435, + 553, + 519 + ], + "type": "inline_equation", + "content": "d_1, d_2, d_3" + }, + { + "bbox": [ + 313, + 435, + 553, + 519 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 435, + 553, + 519 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 435, + 553, + 519 + ], + "type": "text", + "content": " to 2, 1, 0.1 and 0.1, respectively. All models are trained using 2 RTX 3090 GPUs." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 520, + 554, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 554, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 554, + 616 + ], + "type": "text", + "content": "Test Data. We evaluate StableCodec on the test set of CLIC 2020 Professional [60] (CLIC 2020 Test), the validation set of DIV2K [1] (DIV2K Val) and Kodak [16] following [26, 40]. CLIC 2020 Test and DIV2K Val contain 428 and 100 high-quality 2K-resolution natural images, respectively, while Kodak contains 24 natural images with a smaller resolution of " + }, + { + "bbox": [ + 313, + 520, + 554, + 616 + ], + "type": "inline_equation", + "content": "768 \\times 512" + }, + { + "bbox": [ + 313, + 520, + 554, + 616 + ], + "type": "text", + "content": ". We evaluate all images with the original resolution as detailed in the supplementary." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "Evaluation Metrics. We employ established metrics to assess the rate-distortion-perception performance of StableCodec. Concretely, we measure bitrate by bits per pixel (bpp), and evaluate perceptual quality using FID [22], KID [6], DISTS [14] and LPIPS [74] (using AlexNet features by default). Meanwhile, we use PSNR and MS-SSIM [64] to measure the reconstruction fidelity. We follow [26, 40, 46] to calculate FID and KID on " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": " patches, and neglect" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17384" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 553, + 300 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 553, + 300 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 553, + 300 + ], + "type": "image", + "image_path": "3e87a132c57a35619e00a779af9e427ef509fd54c56ae3a052d0c677b4768983.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 119, + 308, + 488, + 319 + ], + "lines": [ + { + "bbox": [ + 119, + 308, + 488, + 319 + ], + "spans": [ + { + "bbox": [ + 119, + 308, + 488, + 319 + ], + "type": "text", + "content": "Figure 7. Qualitative comparisons of different methods on Kodak. Best viewed on screen for details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 340, + 294, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 340, + 294, + 435 + ], + "spans": [ + { + "bbox": [ + 54, + 340, + 294, + 435 + ], + "type": "text", + "content": "the results on Kodak as it is too small for calculating. Note that pixel-level distortion metrics like LPIPS, PSNR and MS-SSIM have strong limitations when evaluating image compression at ultra-low bitrates [11, 14, 26, 35]. Therefore, for StableCodec, we focus primarily on FID, KID, and DISTS, which offer a more accurate assessment of quality in severely compressed images. We also provide user study in the supplementary to visually validate the results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 437, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 593 + ], + "type": "text", + "content": "Compared Methods. We compare StableCodec with advanced image compression methods: (1) Traditional Codec H.266/VVC [10] by VTM-23.0 intra coding, (2) Neural Codec ELIC [20], (3) Generative Codec HiFiC [43], Text+Sketch [35], MS-ILLM [46], PerCo [11], EGIC [31], DiffEIC [40], TACO [34] and GLC [26]. Note that some methods do not release models for ultra-low bitrates, we either re-train or finetune existing weights to reach specific bitrates. For PerCo and GLC that do not have official codes, we use PerCo (SD) [32] as a substitute, and request for the results of GLC under the same evaluation approach1. We equip PerCo and Text+Sketch with the same inference strategy for a fair comparison on high-resolution images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 601, + 140, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 140, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 140, + 612 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 619, + 257, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 619, + 257, + 631 + ], + "spans": [ + { + "bbox": [ + 55, + 619, + 257, + 631 + ], + "type": "text", + "content": "4.2.1. Rate-Distortion-Perception Performance" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 634, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 634, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 634, + 296, + 694 + ], + "type": "text", + "content": "Fig. 6 presents the rate-perception and rate-distortion curves of various methods at ultra-low bitrates over CLIC 2020 Test and DIV2K Val. The proposed StableCodec outperforms all compared methods in terms of FID, KID, and DISTS. Specifically, StableCodec shows a significant im" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 358, + 553, + 475 + ], + "blocks": [ + { + "bbox": [ + 318, + 338, + 548, + 350 + ], + "lines": [ + { + "bbox": [ + 318, + 338, + 548, + 350 + ], + "spans": [ + { + "bbox": [ + 318, + 338, + 548, + 350 + ], + "type": "text", + "content": "Table 2. Encoding and decoding seconds averaged on Kodak." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 358, + 553, + 475 + ], + "lines": [ + { + "bbox": [ + 317, + 358, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 317, + 358, + 553, + 475 + ], + "type": "table", + "html": "
TypeMethod#StepsEnc. TDec. T
VAE- basedELIC [20]-0.1550.245
MLIC++ [27]-0.3640.319
GAN- basedHiFiC [43]-0.1430.337
MS-ILLM [46]-0.1390.316
Diffusion- basedText+Sketch [35]25113.25233.560
PerCo [11]200.2873.742
DiffEIC [40]500.6767.423
StableCodec (Ours)10.1590.326
", + "image_path": "9a5813169cff0e242d3c808d3751acd31f1532f975ce5096cbd79b9093b58597.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 497, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 555, + 605 + ], + "type": "text", + "content": "provement over H.266/VVC and ELIC on all perceptual metrics. Compared to generative codec especially previous SOTA GLC [26], StableCodec demonstrates superiority and stability on FID and KID performance with well-preserved fidelity, and reaches extreme bitrates as low as 0.005 bpp. Although TACO [34] achieves the best LPIPS performance, it fails to ensure visual quality as FID, KID and DISTS scores are high. The PSNR and MS-SSIM results on CLIC and DIV2K are displayed in the supplementary." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 613, + 449, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 449, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 449, + 625 + ], + "type": "text", + "content": "4.2.2. Qualitative Comparisons" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "content": "We provide qualitative results among compared methods in Fig. 7. Notably, StableCodec generates more visually-aligned details and realistic textures at ultra-low bitrates, such as the teeth and murals shown in the first and third rows. In contrast, all other methods fail to produce high-realism results with well-preserved fidelity due to the severe bitrate restriction. For example, ELIC and MS-ILLM" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 295, + 713 + ], + "type": "text", + "content": "1We acknowledge the authors of [26] for kindly providing their results." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17385" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 91, + 294, + 158 + ], + "blocks": [ + { + "bbox": [ + 93, + 71, + 259, + 82 + ], + "lines": [ + { + "bbox": [ + 93, + 71, + 259, + 82 + ], + "spans": [ + { + "bbox": [ + 93, + 71, + 259, + 82 + ], + "type": "text", + "content": "Table 3. Ablation study on " + }, + { + "bbox": [ + 93, + 71, + 259, + 82 + ], + "type": "inline_equation", + "content": "{\\mathcal{E}}_{\\text{Aux }}" + }, + { + "bbox": [ + 93, + 71, + 259, + 82 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 93, + 71, + 259, + 82 + ], + "type": "inline_equation", + "content": "{\\mathcal{D}}_{\\text{Aux }}" + }, + { + "bbox": [ + 93, + 71, + 259, + 82 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 91, + 294, + 158 + ], + "lines": [ + { + "bbox": [ + 57, + 91, + 294, + 158 + ], + "spans": [ + { + "bbox": [ + 57, + 91, + 294, + 158 + ], + "type": "table", + "html": "
Model \nVariantsBD-rate↓ on the Rate-Y curves
PSNRMS-SSIMLPIPSDISTS
Base0000
+ EAux-20.63%-22.05%-23.04%-28.13%
+ EAux & DAux-23.96%-28.12%-40.66%-54.89%
", + "image_path": "9a11a5ff3405640dec50ef82f96b99b8dfeb2aca84b8e7ea3a560138446d89fe.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 59, + 188, + 293, + 305 + ], + "blocks": [ + { + "bbox": [ + 67, + 168, + 282, + 179 + ], + "lines": [ + { + "bbox": [ + 67, + 168, + 282, + 179 + ], + "spans": [ + { + "bbox": [ + 67, + 168, + 282, + 179 + ], + "type": "text", + "content": "Table 4. Ablation study on LoRA ranks " + }, + { + "bbox": [ + 67, + 168, + 282, + 179 + ], + "type": "inline_equation", + "content": "\\left( {{\\mathcal{E}}_{\\mathrm{{SD}}}/{\\epsilon }_{\\mathrm{{SD}}}/{\\mathcal{D}}_{\\mathrm{{SD}}}}\\right)" + }, + { + "bbox": [ + 67, + 168, + 282, + 179 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 188, + 293, + 305 + ], + "lines": [ + { + "bbox": [ + 59, + 188, + 293, + 305 + ], + "spans": [ + { + "bbox": [ + 59, + 188, + 293, + 305 + ], + "type": "table", + "html": "
LoRA RanksBD-rate↓ on the Rate-Y curves
PSNRMS-SSIMLPIPSDISTS
8/8/-0000
8/16/-1.21%-0.35%-4.67%-5.28%
16/16/--5.62%-3.27%-6.41%-12.78%
16/32/--7.12%-5.31%-13.43%-17.96%
32/32/--5.98%-5.69%-11.21%-16.70%
32/64/--5.15%-4.98%-12.95%-17.34%
16/32/4-21.77%-12.28%-8.27%-5.13%
", + "image_path": "49f051536a4a38af15dc8c7a94f0125cb83ebc964047476cd01cca6f151b5ef2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 324, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 295, + 361 + ], + "type": "text", + "content": "produce blurry reconstructions, while PerCo and DiffEIC deviate from the original images. EGIC and TACO exhibit noticeable artifacts, particularly in detailed areas." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 367, + 200, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 200, + 379 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 200, + 379 + ], + "type": "text", + "content": "4.2.3. Computational Complexity" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 382, + 295, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 382, + 295, + 562 + ], + "spans": [ + { + "bbox": [ + 54, + 382, + 295, + 562 + ], + "type": "text", + "content": "We compare the practical complexity of StableCodec with representative schemes in Table 2 using a single RTX 3090 GPU. Among representative image compression schemes, Diffusion-based methods [11, 35, 40] typically suffer from a much longer decoding time compared to VAE-based [20, 27] or GAN-based [43, 46] competitors due to multi-step denoising. In contrast, StableCodec reaches comparable encoding and decoding speed against these methods exploiting one-step denoising, deep compression transforms and efficient entropy model, while achieving significantly better performance at ultra-low bitrates. Detailed runtime analysis is provided in the supplementary. In terms of memory, StableCodec consumes less than 9 GB VRAM with tiling techniques [29, 63], supporting arbitrary-resolution inference on a single GTX 1080Ti GPU." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 573, + 156, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 573, + 156, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 573, + 156, + 586 + ], + "type": "text", + "content": "5. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "type": "text", + "content": "In this section, we conduct ablations to validate the proposed components. For reliable comparison, we compute the BD-rate [7] with Rate-Y curves on Kodak [16] using four target bitrates, where Y denotes specific metrics among PSNR, MS-SSIM, LPIPS, and DISTS." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": "Auxiliary Encoder and Decoder. We begin by providing numerical results for the auxiliary coding branch in Table 3. Specifically, we construct a base model without " + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{Aux}}" + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{Aux}}" + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": ", and a variant with " + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{Aux}}" + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": " only. When " + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{Aux}}" + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": " is incorporated, more than " + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": " bits can be saved to reach the" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 113, + 558, + 192 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 105 + ], + "type": "text", + "content": "Table 5. Ablation study on the training strategy components. Implicit bitrate pruning is abbreviated as IBP. The base one-stage objective only contains bitrate, MSE and LPIPS." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 113, + 558, + 192 + ], + "lines": [ + { + "bbox": [ + 315, + 113, + 558, + 192 + ], + "spans": [ + { + "bbox": [ + 315, + 113, + 558, + 192 + ], + "type": "table", + "html": "
Training StrategyBD-rate↓ on the Rate-Y curves
IBPLadvLCIPPSNRMS-SSIMLPIPSDISTS
---0000
✓---24.12%-13.67%-21.37%-16.60%
✓✓-24.41%9.88%-36.18%-49.24%
✓✓✓13.29%5.70%-38.99%-52.95%
", + "image_path": "cfd78d7064f691ee806e17d851d081508028d0c2b8a2cbf1965c04256803c956.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 213, + 555, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 261 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 261 + ], + "type": "text", + "content": "same reconstruction quality. Subsequently, " + }, + { + "bbox": [ + 313, + 213, + 555, + 261 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{Aux}}" + }, + { + "bbox": [ + 313, + 213, + 555, + 261 + ], + "type": "text", + "content": " further improves the performance particularly in perceptual quality. As shown in Fig. 5, the purified " + }, + { + "bbox": [ + 313, + 213, + 555, + 261 + ], + "type": "inline_equation", + "content": "g_{s}" + }, + { + "bbox": [ + 313, + 213, + 555, + 261 + ], + "type": "text", + "content": " latent provides better guidance for reconstruction consistency." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 262, + 556, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 556, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 556, + 369 + ], + "type": "text", + "content": "LoRA Ranks. We explore the impact of LoRA ranks in Table 4. Positive results are observed in both distortion and perception as the ranks increase to 16/32, which become our final choice. For larger ranks like 32 and 64, we observe performance degradation as the pre-trained priors may be corrupted. Besides, adding LoRA to the VAE decoder introduces a distortion-perception tradeoff, where PSNR and MS-SSIM improve at the cost of LPIPS and DISTS. To preserve perceptual quality, we leave the decoder unchanged." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 370, + 556, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 556, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 556, + 514 + ], + "type": "text", + "content": "Training Strategy. We perform ablations on our training strategy components in Table 5. We construct a simplified one-stage objective with bitrate, MSE and LPIPS, then progressively integrate the two-stage implicit bitrate pruning (IBP), adversarial training " + }, + { + "bbox": [ + 313, + 370, + 556, + 514 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{adv}" + }, + { + "bbox": [ + 313, + 370, + 556, + 514 + ], + "type": "text", + "content": " and the CLIP distance term " + }, + { + "bbox": [ + 313, + 370, + 556, + 514 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CLIP}" + }, + { + "bbox": [ + 313, + 370, + 556, + 514 + ], + "type": "text", + "content": ". We find that IBP improves performance considerably by first adapting the latent codec into the T2I model with relaxed bitrate constraint. Furthermore, incorporating GAN introduces a significant distortion-perception tradeoff since we primarily focus on the perceptual quality. Besides, the CLIP distance alleviates the degradation in distortion and slightly improves perceptual quality." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": "In this work, we introduce StableCodec, a novel diffusion-based extreme image compression approach that addresses key limitations of existing methods. By leveraging one-step diffusion in combination with Deep Compression Latent Codec and Dual-Branch Coding Structure, StableCodec achieves ultra-low bitrate compression with high realism, fidelity, and coding efficiency. Extensive experimental evaluations on benchmark datasets demonstrate the superiority of StableCodec in terms of FID, KID, and DISTS, even at extreme bitrates as low as 0.005 bpp, while enabling competitive speeds with mainstream transform coding methods. These results underscore the potential of diffusion models for practical image compression, particularly in real-time coding scenarios where bitrate is severely constrained." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17386" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 126-135, 2017. 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 191 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 191 + ], + "type": "text", + "content": "[2] Eirikur Agustsson, Michael Tschannen, Fabian Mentzer, Radu Timofte, and Luc Van Gool. Generative adversarial networks for extreme learned image compression. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 221-231, 2019. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 191, + 295, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 295, + 246 + ], + "type": "text", + "content": "[3] Eirikur Agustsson, David Minnen, George Toderici, and Fabian Mentzer. Multi-realism image compression with a conditional generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22324-22333, 2023. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 247, + 295, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 295, + 279 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 295, + 279 + ], + "type": "text", + "content": "[4] Johannes Ballé, Valero Laparra, and Eero P Simoncelli. End-to-end optimized image compression. arXiv preprint arXiv:1611.01704, 2016. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 280, + 295, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 295, + 323 + ], + "type": "text", + "content": "[5] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436, 2018. 2, 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 324, + 295, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 324, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 324, + 295, + 357 + ], + "type": "text", + "content": "[6] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 357, + 295, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 295, + 379 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 295, + 379 + ], + "type": "text", + "content": "[7] Gisle Bjontegaard. Calculation of average psnr differences between rd-curves. Tech. Rep. VCEG-m33, 2001. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 380, + 295, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 295, + 413 + ], + "type": "text", + "content": "[8] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6228-6237, 2018. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 414, + 295, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 295, + 457 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 295, + 457 + ], + "type": "text", + "content": "[9] Yochai Blau and Tomer Michaeli. Rethinking lossy compression: The rate-distortion-perception tradeoff. In International Conference on Machine Learning, pages 675-685. PMLR, 2019. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 459, + 295, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 459, + 295, + 513 + ], + "spans": [ + { + "bbox": [ + 57, + 459, + 295, + 513 + ], + "type": "text", + "content": "[10] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (vvc) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3736-3764, 2021. 1, 2, 3, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 514, + 295, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 295, + 557 + ], + "type": "text", + "content": "[11] Marlene Careil, Matthew J Muckley, Jakob Verbeek, and Stéphane Lathuilière. Towards image compression with perfect realism at ultra-low bitrates. In The Twelfth International Conference on Learning Representations, 2023. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 558, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 295, + 612 + ], + "type": "text", + "content": "[12] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7939-7948, 2020. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 613, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 646 + ], + "type": "text", + "content": "[13] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 647, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 690 + ], + "type": "text", + "content": "[14] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 3, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[15] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Pro" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "text", + "content": "ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "text", + "content": "[16] Rich Franzen. Kodak lossless true color image suite (photocd pcd0992). http://r0k.us/graphics/kodak/, 1993. 3, 5, 6, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 130, + 553, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 175 + ], + "type": "text", + "content": "[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 175, + 553, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 219 + ], + "type": "text", + "content": "[18] Zongyu Guo, Zhizheng Zhang, Runsen Feng, and Zhibo Chen. Causal contextual prediction for learned image compression. IEEE Transactions on Circuits and Systems for Video Technology, 32(4):2329-2341, 2021. 2, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 220, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 275 + ], + "type": "text", + "content": "[19] Dailan He, Yaoyan Zheng, Baocheng Sun, Yan Wang, and Hongwei Qin. Checkerboard context model for efficient learned image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14771-14780, 2021. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 277, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 553, + 342 + ], + "type": "text", + "content": "[20] Dailan He, Ziming Yang, Weikun Peng, Rui Ma, Hongwei Qin, and Yan Wang. Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5718-5727, 2022. 1, 2, 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 343, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 408 + ], + "type": "text", + "content": "[21] Dailan He, Ziming Yang, Hongjiu Yu, Tongda Xu, Jixiang Luo, Yuan Chen, Chenjian Gao, Xinjie Shi, Hongwei Qin, and Yan Wang. Po-elic: Perception-oriented efficient learned image coding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1764–1769, 2022. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 410, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 553, + 464 + ], + "type": "text", + "content": "[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 466, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 553, + 499 + ], + "type": "text", + "content": "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "type": "text", + "content": "[24] Emiel Hoogeboom, Eirikur Agustsson, Fabian Mentzer, Luca Versari, George Toderici, and Lucas Theis. High-fidelity image compression with score-based generative models. arXiv preprint arXiv:2305.18231, 2023. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 545, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 589 + ], + "type": "text", + "content": "[25] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 590, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 644 + ], + "type": "text", + "content": "[26] Zhaoyang Jia, Jiahao Li, Bin Li, Houqiang Li, and Yan Lu. Generative latent coding for ultra-low bitrate image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26088-26098, 2024. 3, 4, 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "text", + "content": "[27] Wei Jiang and Ronggang Wang. Mlic++: Linear complexity multi-reference entropy modeling for learned image compression. In ICML 2023 Workshop Neural Compression: From Information Theory to Applications, 2023. 7, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[28] Wei Jiang, Jiayu Yang, Yongqi Zhai, Peirong Ning, Feng Gao, and Ronggang Wang. Mlic: Multi-reference entropy" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17387" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 106 + ], + "type": "text", + "content": "model for learned image compression. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7618-7627, 2023. 2, 3, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 139 + ], + "type": "text", + "content": "[29] Álvaro Barbero Jiménez. Mixture of diffusers for scene composition and high resolution image generation. arXiv preprint arXiv:2302.02412, 2023. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 140, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 140, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 57, + 140, + 294, + 161 + ], + "type": "text", + "content": "[30] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 162, + 295, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 162, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 57, + 162, + 295, + 216 + ], + "type": "text", + "content": "[31] Nikolai Körber, Eduard Kromer, Andreas Siebert, Sascha Hauke, Daniel Mueller-Gritschneder, and Björn Schuller. Egid: enhanced low-bit-rate generative image compression guided by semantic segmentation. In European Conference on Computer Vision, pages 202-220. Springer, 2024. 2, 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 217, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 217, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 57, + 217, + 294, + 259 + ], + "type": "text", + "content": "[32] Nikolai Körber, Eduard Kromer, Andreas Siebert, Sascha Hauke, Daniel Mueller-Gritschneder, and Björn Schuller. Perco (sd): Open perceptual compression. arXiv preprint arXiv:2409.20255, 2024. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 261, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 295, + 304 + ], + "type": "text", + "content": "[33] Nupur Kumari, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Ensembling off-the-shelf models for gan training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10651-10662, 2022. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 305, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 305, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 57, + 305, + 294, + 358 + ], + "type": "text", + "content": "[34] Hagyeong Lee, Minkyu Kim, Jun-Hyuk Kim, Seungeon Kim, Dokwan Oh, and Jaeho Lee. Neural image compression with text-guided encoding for both pixel-level and perceptual fidelity. arXiv preprint arXiv:2403.02944, 2024. 5, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 360, + 295, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 295, + 403 + ], + "type": "text", + "content": "[35] Eric Lei, Yigit Berkay Uslu, Hamed Hassani, and Shirin Saeedi Bidokhti. Text+ sketch: Image compression at ultra low rates. arXiv preprint arXiv:2307.01944, 2023. 2, 3, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 404, + 295, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 295, + 447 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 295, + 447 + ], + "type": "text", + "content": "[36] Jiahao Li, Bin Li, and Yan Lu. Neural video compression with diverse contexts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22616-22626, 2023. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 449, + 295, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 295, + 502 + ], + "type": "text", + "content": "[37] Zhuoyuan Li, Yao Li, Chuanbo Tang, Li Li, Dong Liu, and Feng Wu. Uniformly accelerated motion model for inter prediction. In 2024 IEEE International Conference on Visual Communications and Image Processing (VCIP), pages 1-5. IEEE, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 504, + 295, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 295, + 557 + ], + "type": "text", + "content": "[38] Zhuoyuan Li, Junqi Liao, Chuanbo Tang, Haotian Zhang, Yuqi Li, Yifan Bian, Xihua Sheng, Xinmin Feng, Yao Li, Changsheng Gao, et al. Ustc-td: A test dataset and benchmark for image and video coding in 2020s. arXiv preprint arXiv:2409.08481, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 559, + 295, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 295, + 602 + ], + "type": "text", + "content": "[39] Zhuoyuan Li, Zikun Yuan, Li Li, Dong Liu, Xiaohu Tang, and Feng Wu. Object segmentation-assisted inter prediction for versatile video coding. IEEE Transactions on Broadcasting, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 603, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 295, + 657 + ], + "type": "text", + "content": "[40] Zhiyuan Li, Yanhui Zhou, Hao Wei, Chenyang Ge, and Jingwen Jiang. Towards extreme image compression with latent feature guidance and diffusion prior. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "text", + "content": "[41] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "type": "text", + "content": "[42] Jinming Liu, Heming Sun, and Jiro Katto. Learned image compression with mixed transformer-cnn architectures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14388-14397, 2023. 2, 3, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 171 + ], + "type": "text", + "content": "[43] Fabian Mentzer, George D Toderici, Michael Tschannen, and Eirikur Agustsson. High-fidelity generative image compression. Advances in Neural Information Processing Systems, 33:11913-11924, 2020. 2, 3, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 172, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 172, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 316, + 172, + 553, + 216 + ], + "type": "text", + "content": "[44] David Minnen and Saurabh Singh. Channel-wise autoregressive entropy models for learned image compression. In 2020 IEEE International Conference on Image Processing (ICIP), pages 3339-3343. IEEE, 2020. 2, 3, 4, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "type": "text", + "content": "[45] David Minnen, Johannes Balle, and George D Toderici. Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems, 31, 2018. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "type": "text", + "content": "[46] Matthew J Muckley, Alaaeldin El-Nouby, Karen Ullrich, Hervé Jégou, and Jakob Verbeek. Improving statistical fidelity for neural image compression with implicit local likelihood models. In International Conference on Machine Learning, pages 25426-25443. PMLR, 2023. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "type": "text", + "content": "[47] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 371, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 403 + ], + "type": "text", + "content": "[48] Gaurav Parmar, Taesung Park, Srinivasa Narasimhan, and Jun-Yan Zhu. One-step image translation with text-to-image models. arXiv preprint arXiv:2403.12036, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 404, + 553, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 447 + ], + "type": "text", + "content": "[49] Yichen Qian, Ming Lin, Xiuyu Sun, Zhiyu Tan, and Rong Jin. Entroformer: A transformer-based entropy model for learned image compression. arXiv preprint arXiv:2202.05492, 2022. 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 449, + 553, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 513 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 513 + ], + "type": "text", + "content": "[50] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 514, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 553, + 558 + ], + "type": "text", + "content": "[51] Lucas Relic, Roberto Azevedo, Markus Gross, and Christopher Schroers. Lossy image compression with foundation diffusion models. In European Conference on Computer Vision, pages 303-319. Springer, 2024. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 559, + 555, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 555, + 613 + ], + "type": "text", + "content": "[52] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 614, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 645 + ], + "type": "text", + "content": "[53] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[54] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87-103. Springer, 2024. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[55] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17388" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "text", + "content": "nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "text", + "content": "[56] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 173 + ], + "type": "text", + "content": "[57] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "type": "text", + "content": "[58] Gary J Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. Overview of the high efficiency video coding (hevc) standard. IEEE Transactions on circuits and systems for video technology, 22(12):1649-1668, 2012. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 252 + ], + "type": "text", + "content": "[59] Lucas Theis, Tim Salimans, Matthew D Hoffman, and Fabian Mentzer. Lossy compression with gaussian diffusion. arXiv preprint arXiv:2206.08889, 2022. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "text", + "content": "[60] George Toderici, Lucas Theis, Nick Johnston, Eirikur Agustsson, Fabian Mentzer, Johannes Ballé, Wenzhe Shi, and Radu Timofte. Clic 2020: Challenge on learned image compression, 2020, 2020. 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 299, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 331 + ], + "type": "text", + "content": "[61] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 333, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 294, + 364 + ], + "type": "text", + "content": "[62] Gregory K Wallace. The jpeg still picture compression standard. Communications of the ACM, 34(4):30-44, 1991. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 366, + 295, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 295, + 409 + ], + "type": "text", + "content": "[63] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, pages 1-21, 2024. 3, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 411, + 294, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 455 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 455 + ], + "type": "text", + "content": "[64] Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for image quality assessment. In The Thrity-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, pages 1398-1402. IEEE, 2003. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 456, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 499 + ], + "type": "text", + "content": "[65] Tongda Xu, Ziran Zhu, Dailan He, Yanghao Li, Lina Guo, Yuanyuan Wang, Zhe Wang, Hongwei Qin, Yan Wang, Jingjing Liu, et al. Idempotence and perceptual image compression. arXiv preprint arXiv:2401.08920, 2024. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 501, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 544 + ], + "type": "text", + "content": "[66] Naifu Xue, Zhaoyang Jia, Jiahao Li, Bin Li, Yuan Zhang, and Yan Lu. Dlf: Extreme image compression with dual-generative latent fusion. arXiv preprint arXiv:2503.01428, 2025. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 545, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 599 + ], + "type": "text", + "content": "[67] Zeyu Yan, Fei Wen, Rendong Ying, Chao Ma, and Peilin Liu. On perceptual lossy compression: The cost of perceptual reconstruction and an optimal training framework. In International Conference on Machine Learning, pages 11682-11692. PMLR, 2021. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 601, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 633 + ], + "type": "text", + "content": "[68] Zeyu Yan, Fei Wen, and Peilin Liu. Optimally controllable perceptual lossy compression. arXiv preprint arXiv:2206.10082, 2022. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 635, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 667 + ], + "type": "text", + "content": "[69] Ruihan Yang and Stephan Mandt. Lossy image compression with conditional diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[70] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. arXiv preprint arXiv:2405.14867, 2024. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 251 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 315, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 553, + 105 + ], + "type": "text", + "content": "[71] Weihao Yu and Xinchao Wang. Mambaout: Do we really need mamba for vision? arXiv preprint arXiv:2405.07992, 2024. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[72] Weihao Yu, Pan Zhou, Shuicheng Yan, and Xinchao Wang. Inceptionnext: When inception meets convnext. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5672-5683, 2024. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 152, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 194 + ], + "type": "text", + "content": "[73] Aiping Zhang, Zongsheng Yue, Renjing Pei, Wenqi Ren, and Xiaochun Cao. Degradation-guided one-step image super-resolution with diffusion priors. arXiv preprint arXiv:2409.17058, 2024. 2, 3, 4, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "type": "text", + "content": "[74] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5, 6" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17389" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/StableDepth_ Scene-Consistent and Scale-Invariant Monocular Depth/73601462-88ee-4d47-a92f-ac7545019662_content_list.json b/2025/StableDepth_ Scene-Consistent and Scale-Invariant Monocular Depth/73601462-88ee-4d47-a92f-ac7545019662_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..73f953dd483e288343d50b63794bfb42dc1731eb --- /dev/null +++ b/2025/StableDepth_ Scene-Consistent and Scale-Invariant Monocular Depth/73601462-88ee-4d47-a92f-ac7545019662_content_list.json @@ -0,0 +1,1565 @@ +[ + { + "type": "text", + "text": "StableDepth: Scene-Consistent and Scale-Invariant Monocular Depth", + "text_level": 1, + "bbox": [ + 147, + 130, + 849, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zheng Zhang $^{1,2,\\star,*}$ Lihe Yang $^{1,\\star}$ Tianyu Yang $^{2,4,\\dagger}$ Chaohui Yu $^{2,4}$ Xiaoyang Guo $^{3}$ Yixing Lao $^{1}$ Hengshuang Zhao $^{1,\\dagger}$", + "bbox": [ + 212, + 178, + 776, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The University of Hong Kong 2DAMO Academy, Alibaba Group", + "bbox": [ + 228, + 218, + 769, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3The Chinese University of Hong Kong 4Hupan Lab", + "bbox": [ + 282, + 237, + 714, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://stabledepth.github.io", + "bbox": [ + 411, + 257, + 581, + 271 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ed2e4f41cfe3e1b6c05366775acc5ff9c0cacb43f52a9052db53c6ff90c1157b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 306, + 893, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4490ab9a6e715390594f56e20f7355d2a50fa3fbd9a8b2d7edbdf10e61e6fb89.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 395, + 893, + 483 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2ae26fe8b485b7a3119433d3c2cc07c3041cbfa3aa7570aa45756eb54e360b6b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 484, + 893, + 571 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/883a9c2ab0c65c01db85179a41d31bc8ae28ec877c628825def7489cca6442c9.jpg", + "image_caption": [ + "Figure 1. StableDepth achieves efficient online monocular depth estimation that produces scene-consistent and scale-invariant predictions frame by frame. As shown in the above challenging scenarios, StableDepth generates stable depth maps with precise geometric details and effectively handles dynamic motions, without the need of accessing future frames." + ], + "image_footnote": [], + "bbox": [ + 101, + 571, + 893, + 660 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 729, + 326, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in monocular depth estimation significantly improve robustness and accuracy. However, relative depth models exhibit flickering and 3D inconsistency in video data, limiting 3D reconstruction applications. We introduce StableDepth, a scene-consistent and scale-invariant depth estimation method achieving scene-level 3D consistency. Our dual-decoder architecture learns from", + "bbox": [ + 89, + 761, + 485, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "large-scale unlabeled video data, enhancing generalization and reducing flickering. Unlike previous methods requiring full video sequences, StableDepth enables online inference at $13 \\times$ faster speed, achieving significant improvements across benchmarks with comparable temporal consistency to video diffusion-based estimators.", + "bbox": [ + 511, + 731, + 906, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 828, + 645, + 844 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Monocular Depth Estimation (MDE) [12, 45] is pivotal in bridging the gap between 2D and 3D data representation, serving critical roles in a wide range of downstream applica", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\star$ Joint first authors. *Work done when interning at DAMO Academy. † Corresponding author.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7069", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tions, such as autonomous driving [51], robotics [54], medical imaging [32], and virtual reality [13]. Advancements in foundational models [35, 43] have greatly enhanced the accuracy and robustness of MDE [25, 56]. Despite these advances, the intrinsic ambiguity of monocular cues still poses substantial challenges, particularly for downstream applications that require 3D-consistent data and maintaining stability for uncontrolled in-the-wild data.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "MDE is primarily categorized into two main task settings: relative depth estimation and metric depth estimation. Relative depth estimation, as explored in studies [15, 18, 25, 39, 56], involves predicting depth values within a normalized depth range of $[0, 1]$ . This scale- and shift-invariant regression target facilitates robust joint training across diverse data domains. However, the predicted depth values differ from the actual depth due to scale and shift discrepancies, leading to inconsistency across video frames. This inconsistency can complicate advanced downstream tasks like 3D reconstruction. In contrast, metric depth estimation [4, 22, 37, 59] aims to predict physical distances in metric scale. However, these approaches face significant challenges. Scaled versions of a scene's depth values might all represent plausible depth estimations of the actual depth. Such ambiguity complicates the learning process of depth regression, hindering both generalization capabilities and precise 3D structures. Severe flickering and 3D inconsistency are also observed given video inputs. This challenge prompts us to reconsider what is an effective representation of MDE foundation models.", + "bbox": [ + 91, + 214, + 483, + 531 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The limitations of existing approaches are summarized in Tab. 1, which compares different depth estimation paradigms. While relative depth models [25, 39, 56] offer good generalization across datasets but suffer from per-frame inconsistency, metric depth models [4, 59] provide real-world measurements but struggle with generalization across diverse scenes. Scale-invariant models [20] improve upon relative models by eliminating shift ambiguity while maintaining scale invariance, but they still lack consistency across frames in the same scene.", + "bbox": [ + 89, + 535, + 482, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose that the ideal task setting should be scene-consistent and scale-invariant (SCSI) depth estimation, as shown in Tab. 1. Our approach strikes a balance between per-frame scale-invariant methods that lack consistency and purely metric approaches that struggle with generalization. Specifically, our model primarily generates scale-invariant depth maps, but in the same scene or video, predictions across multiple frames demonstrate 3D consistency. Additionally, the scale factors for multi-frame data remain consistent compared with the actual depth values. In this work, the development of our model is primarily guided by the following objectives: 1) ensuring high-quality, scene-consistent, and scale-invariant depth estimations to achieve robust video consis", + "bbox": [ + 89, + 689, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/3fb730c9fb83678fb4f6d3f9836626a91056d215a9833fbdf667df335e09a04b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Depth TypeScaleShiftGeneralReconsConsist
RelativePer-framePer-frame✓✗✗
Scale-invariantPer-frameNone✓✓✗
MetricNoneNone✗✓✗
SCSI (Ours)Per-sceneNone✓✓✓
", + "bbox": [ + 517, + 90, + 903, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Comparison of different depth types. Our proposed SCSI depth combines the strengths of relative and metric depth while overcoming their weaknesses. General: generalization capability. Recons: reconstruction quality. Consist: scene consistency.", + "bbox": [ + 511, + 175, + 906, + 231 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tency; 2) enhancing strong generalization capabilities and precise depth details.", + "bbox": [ + 511, + 247, + 903, + 277 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our framework introduces three key innovations to achieve scene-consistent depth estimation. First, we leverage temporal priors from pre-trained video diffusion models to generate pseudo-labels with inherent 3D consistency, replacing traditional unstable photometric constraints [3, 23, 47, 48]. Second, we design a dual-branch architecture where one branch processes labeled data for metric accuracy, while the other aligns unlabeled video frames to pseudo-labels through scene-level normalization, decoupling conflicting objectives. Third, our semi-supervised paradigm jointly optimizes metric supervision and temporal alignment, enabling the model to generalize across domains while eliminating flickering artifacts. This unified approach supports real-time inference without requiring future frames, overcoming the limitations of both single-frame and video-based methods.", + "bbox": [ + 511, + 279, + 906, + 520 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate our model on multiple datasets under zero-shot settings, aligning each scene using only a single scale (SCSI). As demonstrated in Figure 1, our method robustly handles challenging scenarios including dynamic motions and complex textures while preserving geometric fidelity frame-by-frame. Compared to DepthCrafter [23], we achieve $13 \\times$ faster inference speed with online processing capability, improving prediction accuracy on Sintel [5], Bonn [36], ScanNet v2 [9], and KITTI [17] by $13.2\\%$ , $86.8\\%$ , $39.3\\%$ , and $8.2\\%$ , respectively. Our contributions are summarized as follows:", + "bbox": [ + 511, + 521, + 905, + 686 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce StableDepth, a novel monocular depth estimation framework that achieves scene-consistent and scale-invariant (SCSI) depth estimation, bridging the gap between relative and metric depth prediction while maintaining temporal consistency in video sequences.", + "- We propose a video diffusion baking strategy that leverages pre-trained video diffusion models for generating high-quality pseudo-labels, enabling effective semi-supervised learning from unlabeled video data while preserving temporal consistency.", + "- We develop a dual-decoder architecture that simultaneously handles supervised metric depth estimation and unsupervised depth alignment, effectively addressing the target misalignment issue while maintaining accuracy." + ], + "bbox": [ + 511, + 689, + 903, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7070", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We demonstrate state-of-the-art performance across multiple benchmarks, achieving up to $86.8\\%$ improvement in accuracy while providing $13 \\times$ faster inference compared to previous methods. Our method enables efficient, online inference while maintaining comparable temporal stability to video diffusion models.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 196, + 230, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Monocular depth estimation. MDE focuses on inferring depth from a single image. Early works [8, 31, 55] learn ordinal relationships from coarse annotations but cannot recover geometric structure. Recent methods [11, 39, 56, 57] leverage large-scale data to learn affine-invariant depth with improved generalization. However, unknown depth shifts cause geometric distortions, hindering 3D applications. To address this, recent works attempt zero-shot metric depth estimation [4, 19, 22, 37, 38, 59, 61] and explore self-supervised scale-consistent learning [50, 60], though maintaining consistent scale across scenes remains challenging.", + "bbox": [ + 89, + 223, + 483, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Towards consistent video depth. Video depth estimation requires both temporal consistency and per-frame precision. Traditional approaches use post-processing optimization [27, 34] or temporal modeling [29, 49], but incur computational costs or fail in dynamic scenes. Other methods leverage camera poses [46, 53] or memory-based approaches [52, 58], but often sacrifice geometric fidelity for smoothness. Recent works like Video Depth Anything [7] and video-free approaches [26] show promise for temporal consistency in long sequences. Video Depth Anything freezes the Depth Anything V2 encoder and incorporates cross-frame attention in the DPT [40] decoder for temporal consistency, though these methods still produce depth with inherent shift ambiguity.", + "bbox": [ + 89, + 390, + 483, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion priors for depth modeling. Video diffusion models [21, 24] demonstrate potential for consistent depth estimation. DepthCrafter [23] achieves temporal stability but generates only relative depths, while ChronoDepth [47] suffers from limited temporal contexts. Our proposed SCSI depth estimation leverages video diffusion baking and dual-decoder architecture for temporally consistent depth suitable for efficient online inference.", + "bbox": [ + 89, + 602, + 483, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. StableDepth", + "text_level": 1, + "bbox": [ + 89, + 738, + 218, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our work leverages both labeled images and unlabeled video data to enhance SCSI depth predictions, achieving video consistency and enhanced generalization capabilities through the distillation of video diffusion priors. Formally, we denote the sets of labeled image-depth pairs and unlabeled videos as $\\mathcal{D}^l = \\{(x_i,d_i)\\}_{i = 1}^M$ and $\\mathcal{D}^u = \\{v_i\\}_{i = 1}^N$ , respectively. Our objective is to utilize a high-quality labeled image dataset, Hypersim [42], to help the model grasp real-world scale, while employing unlabeled video data to guide", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the model's depth predictions toward achieving video consistency and improved generalization.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Mining Consistency from Unlabeled Videos", + "text_level": 1, + "bbox": [ + 511, + 132, + 880, + 148 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "StableDepth is designed to learn from both labeled images and unlabeled videos. For labeled image data, we predict the depth in metric space. The activation function of the network's last layer is sigmoid, after which its predictions are scaled to match the dataset values by multiplying with a predefined maximum depth. The SiLogLoss [12, 44, 57] is then used to minimize the error between the prediction and the ground truth:", + "bbox": [ + 511, + 152, + 905, + 266 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {l a b e l}} = \\sqrt {\\frac {1}{H W} \\sum_ {i = 1} ^ {H W} \\left(\\log \\frac {d _ {i}}{d _ {i} ^ {*}}\\right) ^ {2} - \\lambda \\left(\\frac {1}{H W} \\sum_ {i = 1} ^ {H W} \\log \\frac {d _ {i}}{d _ {i} ^ {*}}\\right) ^ {2}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 276, + 903, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $d_i^*$ and $d_{i}$ are the prediction and ground truth, respectively, and $\\lambda$ is a hyper-parameter.", + "bbox": [ + 511, + 330, + 903, + 361 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For unlabeled video data, we leverage video diffusion models, specifically DepthCrafter [23], to generate pseudolabels due to their inherent advantages in maintaining temporal consistency. These models, pre-trained on large-scale video data with cross-frame attention mechanisms, can generate stable depth predictions across entire video segments while preserving a consistent global scale and shift relationship with metric depth. Unlike conventional image-based MDE approaches that process frames independently, video diffusion models account for temporal dependencies, making them particularly effective for generating reliable pseudo-labels that exhibit strong video consistency - a crucial property for our scene-consistent training objective. For a video, we need a method to normalize the depth predicted by our model to the same space as the pseudo-labeled depth. Unlike the normalization used in single-frame relative depth training, this approach focuses on ensuring that an entire video has a unique scale and shift, rather than each frame having an independent scale and shift. This method aims to achieve video consistency in the depth predictions within the same scene. To supervise our model, we first use the following formula to normalize the pseudo-labels,", + "bbox": [ + 511, + 362, + 906, + 694 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {D} (v _ {i}) = \\frac {D (v _ {i}) - m (D (v _ {i}))}{s (D (v _ {i}))}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 705, + 903, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $D(v_{i})$ is the pseudo-label for video $v_{i}$ , and $m(D(v_{i}))$ and $s(D(v_{i}))$ are used to remove the shift and scale for this video, respectively:", + "bbox": [ + 511, + 750, + 903, + 792 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nm (D (v _ {i})) = \\operatorname {m e d i a n} (D (v _ {i})),\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 804, + 740, + 819 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ns \\left(D \\left(v _ {i}\\right)\\right) = \\frac {1}{T H W} \\sum_ {i = 1} ^ {T H W} | D \\left(v _ {i}\\right) - m \\left(D \\left(v _ {i}\\right)\\right) |. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 823, + 903, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the frame-by-frame depth predictions of video $v_{i}$ from our model, we concatenate them to obtain $P(v_{i})$ with", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7071", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/347cbf4b90ce09213c1072109f744af1d03ab5cc6bb7b8787a1157f68f170541.jpg", + "image_caption": [ + "Figure 2. Overview of StableDepth. StableDepth employs a shared encoder with dual decoders to process both labeled images and unlabeled videos. The upper branch handles labeled images with direct metric depth supervision (\"Sup\"), while the lower branch processes unlabeled video sequences using temporal-coherent pseudo-labels generated by pre-trained video diffusion priors [23] (\"Bake\"). This architecture enables scene-consistent and scale-invariant (SCSI) depth prediction by combining the benefits of metric depth supervision with temporal consistency learning from video sequences. Solid arrows indicate the primary data flow, while dashed arrows represent the auxiliary training path for unlabeled data." + ], + "image_footnote": [], + "bbox": [ + 94, + 85, + 903, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "dimensions $[T,H,W]$ . We then apply the same normalization process to the depth predictions $P(v_{i})$ to obtain its normalized version $\\hat{P} (v_{i})$ . For $\\hat{P} (v_{i})$ and $\\hat{D} (v_{i})$ , we use the following alignment loss to minimize their difference,", + "bbox": [ + 89, + 354, + 483, + 415 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {u n l a b e l}} = \\frac {1}{T H W} \\sum_ {i = 1} ^ {T H W} \\left| \\hat {P} \\left(v _ {i}\\right) - \\hat {D} \\left(v _ {i}\\right) \\right|. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 425, + 482, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In one iteration, we simultaneously sample a certain proportion of labeled images and unlabeled videos for joint training, and use the following combined loss as the final loss for this iteration:", + "bbox": [ + 89, + 477, + 483, + 536 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {l a b e l}} + \\gamma \\mathcal {L} _ {\\text {u n l a b e l}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 549, + 482, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\gamma$ gradually increases following a cosine curve as the number of iterations progresses.", + "bbox": [ + 89, + 575, + 483, + 607 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Avoiding Conflict in Depth Labels", + "text_level": 1, + "bbox": [ + 89, + 614, + 387, + 632 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the field of video consistency for depth prediction, most research [23] focuses on improving video consistency in near-field scenes. Therefore, we use depth in the disparity space for video data training.", + "bbox": [ + 89, + 637, + 482, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following the previous network structure [2, 39, 56, 57], we use the ViT-L encoder [10], denoted as $\\mathcal{E}$ , paired with the decoder from DPT [40], denoted as $\\mathcal{D}_1$ . The decoder outputs depth in SCSI space with a range of (0, 1). A natural approach is to utilize this original decoder to predict the disparity-space depth of a video sequence, and then take the reciprocal of the predictions. Specifically, the depth prediction $P(v_{i})$ is given by:", + "bbox": [ + 89, + 698, + 482, + 819 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP \\left(v _ {i}\\right) = \\frac {1}{\\mathcal {D} _ {1} \\left(\\mathcal {E} \\left(v _ {i}\\right)\\right)}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 829, + 482, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, the depth predictions are normalized and supervised in semi-supervised settings following Eq. 3 and 4.", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unfortunately, in our experiments, we found that although this modification improves the video consistency of depth predictions for videos, it reduces the accuracy of depth prediction by the model, resulting in lower metrics such as $\\delta_{1}$ . We speculate that this is mainly due to the used pseudo-labeling model [23], which, despite having good video consistency, still lags behind Depth Anything V2 [57] in terms of depth prediction accuracy, particularly in the prediction of distant objects. To address this issue, we propose to use an auxiliary network to decouple the prediction and supervision of labeled images and unlabeled videos. To this end, we experimented with two approaches:", + "bbox": [ + 511, + 354, + 906, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the first approach, both labeled images and unlabeled videos share the encoder $\\mathcal{E}$ and the decoder $\\mathcal{D}_1$ . A lightweight CNN network, denoted as $\\mathcal{C}$ , is used to directly process the prediction results of the original decoder $\\mathcal{D}_1$ with the CNN network outputting depth in disparity space:", + "bbox": [ + 511, + 536, + 905, + 611 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP \\left(v _ {i}\\right) = \\mathcal {C} \\left(\\mathcal {D} _ {1} \\left(\\mathcal {E} \\left(v _ {i}\\right)\\right)\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 618, + 903, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the second approach, labeled images and unlabeled videos share the encoder $\\mathcal{E}$ . An additional DPT [40] decoder, denoted as $\\mathcal{D}_2$ , is used to process the features extracted by the encoder $\\mathcal{E}$ for unlabeled videos, as illustrated by Fig. 2. $\\mathcal{D}_2$ directly produces depth in disparity space, while $\\mathcal{D}_1$ continues to output depth in SCSI space.", + "bbox": [ + 511, + 642, + 905, + 734 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP \\left(x _ {i}\\right) = \\mathcal {D} _ {1} \\left(\\mathcal {E} \\left(x _ {i}\\right)\\right), \\quad P \\left(v _ {i}\\right) = \\mathcal {D} _ {2} \\left(\\mathcal {E} \\left(v _ {i}\\right)\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 741, + 903, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The final results indicate that sharing the encoder between labeled images and unlabeled videos, while using two independent DPT decoders, achieves the best accuracy and video consistency.", + "bbox": [ + 511, + 765, + 905, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Scene-Consistent Depth Evaluation", + "text_level": 1, + "bbox": [ + 511, + 833, + 820, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scale alignment. For zero-shot evaluation, we enforce single-scale consistency per scene through RANSAC-based alignment with zero shift constraint. Each scene's", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7072", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "predictions are scaled by a unified factor determined through 1,000 sampling iterations, retaining inliers with $\\max \\left(\\frac{d_{pred}}{d_{gt}},\\frac{d_{gt}}{d_{pred}}\\right) < \\tau$ , where $\\tau = 1.25$ . The optimal scale is refined via least-squares optimization over all inliers, preserving geometric fidelity while eliminating cross-frame scale ambiguity. This per-scene single-scale paradigm ensures coherent depth relationships across dynamic sequences without per-frame adjustments.", + "bbox": [ + 89, + 90, + 480, + 214 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Motion-aware temporal metric. We propose the Motion-aware Temporal Difference (MTD) to quantify depth consistency in dynamic regions:", + "bbox": [ + 89, + 215, + 480, + 262 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {M T D} = \\frac {1}{T - 1} \\sum_ {t = 1} ^ {T - 1} \\frac {\\sum_ {x , y} | \\Delta D _ {t} | \\cdot W _ {t}}{H \\times W}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 276, + 482, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Delta D_{t} = D_{t} - D_{t - 1}$ and $W_{t}(x,y) = (1 + \\| \\mathbf{f}_{t}\\|_{2})^{-1}$ weights the depth variations by optical flow magnitude $\\mathbf{f}_t$ (computed via Farneback [14] algorithm). Lower MTD indicates better temporal stability, penalizing flickering in high-motion areas.", + "bbox": [ + 89, + 330, + 482, + 406 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Discussion", + "text_level": 1, + "bbox": [ + 89, + 421, + 209, + 435 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we address several important aspects of our approach to provide further clarity on methodology, design choices, and relationships to existing work.", + "bbox": [ + 89, + 445, + 482, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Decouple accuracy and consistency via dual learning. Our dual-decoder architecture decouples depth accuracy from temporal consistency learning, enabling joint optimization from diverse data sources. The primary decoder specializes in per-pixel depth precision through supervised training on high-quality labeled images, while the auxiliary decoder learns scene-level coherence from unlabeled videos via pseudo-labels. This separation resolves the inherent conflict where video-based methods like DepthCrafter [23] sacrifice depth accuracy for temporal smoothness, and single-frame approaches fail to ensure cross-frame consistency. Crucially, our design preserves absolute depth scales critical for 3D reconstruction, unlike relative-depth representations in video diffusion models.", + "bbox": [ + 88, + 491, + 482, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Beyond video diffusion priors. We repurpose video diffusion models as sources for scene consistency rather than as direct depth estimators. While models like DepthCrafter [23] excel at generating temporally consistent sequences, they produce relative depth representations with limited accuracy, particularly for distant objects, making them unsuitable for downstream tasks requiring precise geometry. Our approach extracts the temporal consistency knowledge from these models through scene-level normalization while maintaining superior depth accuracy. This strategy enables online inference $13 \\times$ faster than video-based methods while achieving both better depth precision and comparable temporal stability.", + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 513, + 90, + 638, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 513, + 114, + 730, + 131 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Network parameters and training details. We use the network architecture from Depth Anything V2 (DAv2) [57], consisting of a pre-trained ViT-L [10] encoder for feature extraction and a DPT [40] decoder for depth regression. The ViT-L encoder is initialized with parameters from the DAv2 model, while the primary decoder is randomly initialized with Sigmoid activation for metric depth output. The auxiliary decoder, also initialized randomly, uses ReLU activation for disparity (inverse depth) prediction. Each batch consists of a 1:2 ratio of labeled images to unlabeled video frames, totaling 24 frames. The learning rate for the encoder is $2.5 \\times 10^{-6}$ , while the decoder's learning rate is 10 times larger. We use the AdamW optimizer [33] with a weight decay of 0.01 and employ a polynomial decay schedule for the learning rate. For data augmentation, we apply horizontal flipping to labeled images. In Eq. (5), the maximum value of $\\gamma$ is set to 1. In Eq. (1), $\\lambda$ is set to 0.5.", + "bbox": [ + 511, + 138, + 903, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training dataset and pseudo-labeling. To train our model, we utilize DepthCrafter [23], a model finetuned on SVD [3], to generate video-consistent pseudo-labels for unlabeled video data. For metric depth supervision, we use Hypersim [42], a high-quality synthetic dataset with $\\sim 60\\mathrm{K}$ frames, as training data for decoder $\\mathcal{D}_1$ . We select 6,000 videos from SA-V [41], which features dynamic foreground objects, yielding approximately $\\sim 200\\mathrm{K}$ frames for pseudolabeling training. To enhance the model's performance on distant scenes where DepthCrafter shows limitations, we incorporate additional datasets with accurate depth labels including VKITTI2 [6, 16], LightwheelOcc [30], and MatrixCity [28], treating frames from the same scene as video sequences during training.", + "bbox": [ + 511, + 396, + 903, + 607 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Evaluation", + "text_level": 1, + "bbox": [ + 513, + 619, + 633, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation datasets. We evaluate on four datasets spanning diverse scenarios: Sintel [5] (23 synthetic videos, 50 frames each with precise depth labels), Bonn [36] (26 dynamic videos using frames 30-140 with foreground motions), ScanNet v2 [9] (100 indoor test videos sampled every third frame), and the full KITTI validation set [17] (13 outdoor driving videos with first 110 frames per sequence). This comprehensive benchmark covers synthetic animations, dynamic interactions, static indoor scenes, and real-world autonomous driving environments.", + "bbox": [ + 511, + 642, + 903, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantitative results. We evaluate our method against strong baselines: DepthCrafter [23], known for video consistency, Metric3D [59], a state-of-the-art metric depth estimator, and ZoeDepth [1], a robust monocular depth estimation method. For fair comparison, we align all predictions within each video using a single shared scale factor across the entire scene with shift set to zero. As shown in", + "bbox": [ + 511, + 795, + 903, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7073", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1a27920025350c8d0a704b45fdde248e5d80d87f775e4ce1ddcee77bccd5173e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Frame IntervalMethodSintel (~50 frames)Bonn (110 frames)ScanNet (90 frames)KITTI (110 frames)
AbsRel (↓)δ1 (↑)AbsRel (↓)δ1 (↑)AbsRel (↓)δ1 (↑)AbsRel (↓)δ1 (↑)
1DepthCrafter [23]2.9590.5450.3110.5240.3700.6610.1230.832
Zoedepth [1]0.4540.3560.2770.5620.3470.4360.2160.633
Metric3D [59]0.3720.6100.0550.9810.0590.9740.0670.977
StableDepth (Ours)0.2800.6170.0630.9790.0940.9210.1120.900
4DepthCrafter [23]6.3530.3930.5150.4170.5390.4400.2440.670
Zoedepth [1]0.4610.3570.2770.5630.3490.4360.2180.629
Metric3D [59]0.3610.6170.0550.9810.0590.9740.0670.977
StableDepth (Ours)0.2920.6060.0630.9790.0930.9210.1120.901
8DepthCrafter [23]2.0480.3980.4950.4120.5710.4230.2410.656
Zoedepth [1]0.4700.3570.2760.5620.3520.4340.2180.633
Metric3D [59]0.3630.6130.0550.9800.0600.9730.0660.977
StableDepth (Ours)0.3000.6040.0620.9800.0940.9200.1120.902
", + "bbox": [ + 91, + 88, + 911, + 318 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/358b9d6ff7641a7bc714845fdaba20a028a5e2155d990be68dde45ff668fad0a.jpg", + "table_caption": [ + "Table 2. Comparison of depth estimation methods across different frame intervals. StableDepth maintains consistent performance across varying frame intervals while achieving superior accuracy. Best results are shown in bold, second best are underlined. Metric3D results on ScanNet are shown in gray as it was trained on this dataset." + ], + "table_footnote": [], + "table_body": "
MethodBonnScanNet
AbsRel (↓)δ1(↑)AbsRel (↓)δ1(↑)
Baseline0.0710.9460.1020.903
DPT0.1090.9140.1260.861
DPT + CNN0.1080.9260.1380.829
Dual DPT0.0570.9760.0970.913
", + "bbox": [ + 91, + 378, + 480, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Ablation on decoders. The baseline model is trained only on labeled image data. We compare three approaches for handling unlabeled video data: using a single DPT decoder with reciprocal output (DPT), adding a lightweight CNN head (DPT + CNN), and using dual DPT decoders (Dual DPT).", + "bbox": [ + 89, + 491, + 482, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 2, our method achieves superior performance across most datasets and metrics, with particularly strong improvements in AbsRel and $\\delta_{1}$ metrics compared to video-based approaches. We examine the impact of frame intervals on prediction accuracy to simulate varying temporal densities in real-world scenarios. Note that compared to our ablation studies, these results reflect extended training with additional iterations for optimal performance.", + "bbox": [ + 89, + 565, + 482, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 696, + 243, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison of different decoders. In Tab. 3, we compare four methods using the Bonn and ScanNet v2 datasets: training with only labeled Hypersim data, and training with both labeled Hypersim [42] and unlabeled SA-V [41] data processed by three decoder designs. The first row (Baseline) shows results with only labeled Hypersim data, while the following rows show results with both labeled and unlabeled data. For labeled Hypersim data, depth in SCSI space is obtained directly from the DPT head output. For the unlabeled SA-V data, the method for obtaining disparity space depth varies: in the second row (DPT), it's the reciprocal of the original DPT output; in the third row (DPT", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "+ CNN), a lightweight CNN processes the DPT output; in the fourth row (Dual DPT), an additional DPT head provides the disparity space depth. The results show that using a single DPT head, either by reciprocating its output or adding a lightweight CNN, reduces prediction accuracy due to the partial inaccuracy of the pseudo-labels. However, using two independent DPT heads improves frame-to-frame consistency without impacting accuracy, validating the effectiveness of our disentangled approach.", + "bbox": [ + 511, + 382, + 906, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Encoder feature stability. We analyze encoder feature stability across consecutive DAVIS frames to understand our dual-decoder effectiveness. Compared to DAv2, our method achieves $7.1\\%$ lower mean differences, $13.5\\%$ reduced standard deviation, and $2.3\\%$ higher cosine similarity. These improvements confirm that auxiliary decoder supervision enhances encoder stability, contributing to temporal consistency in depth predictions.", + "bbox": [ + 511, + 518, + 908, + 641 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness of unlabeled video data. As shown in Tab. 4, we evaluate incremental training data impact. Using only Hypersim [42] (row 1), then adding pseudo-labeled SA-V [41] data (row 2) shows mixed results - improving Bonn and ScanNet v2 but decreasing Sintel and KITTI performance. This reveals DepthCrafter's video consistency benefits but limitations in distant regions. Progressively incorporating VKITTI2 [6, 16], LightWheel [30], Hypersim [42], and MatrixCity [28] (rows 3-6) successfully addresses these limitations, achieving best performance across all datasets.", + "bbox": [ + 511, + 642, + 908, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison of video consistency. Tab. 5 showcases the performance comparison of depth estimation methods on the DAVIS dataset evaluated using Motion-aware temporal Difference (MTD (Eq. (9)), lower is better). The video diffusion-based DepthCrafter [23] achieves strong consistency $(\\mathrm{MTD} = 0.003479)$ but requires processing the entire video sequence and runs at only 0.45 FPS. While", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7074", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/26f4f6fe90a9b8d2c1293a0843f558d906dce92089d7881c0505814eedee670d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Training DataSintel (~50 frames)Bonn (110 frames)ScanNet (90 frames)KITTI (110 frames)
AbsRel (↓)δ1 (↑)AbsRel (↓)δ1 (↑)AbsRel (↓)δ1 (↑)AbsRel (↓)δ1 (↑)
Hypersim only [42]0.3090.5780.0710.9460.1020.9030.1220.871
+ SA-V [41]0.3260.5610.0570.9760.0970.9130.1430.823
+ VKITTI2 [6, 16]0.3180.5770.0580.9760.0970.9110.1290.839
+ LightWheel [30]0.3120.5730.0600.9760.0940.9160.1340.847
+ Hypersim-U [42]0.3140.5720.0640.9760.0960.9180.1230.870
+ Matrixcity [28]0.3150.5800.0620.9790.0920.9240.1140.895
", + "bbox": [ + 91, + 88, + 906, + 223 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/517d05a13a0817eb8f234dc7db953a936edae9d9fd565f6c8693aa1dbec0d07e.jpg", + "image_caption": [ + "Figure 3. Ablation on the number of unlabeled video frames in each batch. From 0 to 16, more frames bring better scene-level consistency. (Out of memory for ViT-Large under 32 frames)" + ], + "image_footnote": [], + "bbox": [ + 91, + 311, + 480, + 429 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/35cba8b4c6254d0c1318149139c75a716725c491e5b6b02e29846e6f3d3447c0.jpg", + "table_caption": [ + "Table 4. Comparison of depth estimation performance with different training data configurations. The baseline model is trained on labeled Hypersim data only, with subsequent rows showing the cumulative effect of adding different unlabeled datasets (VKITTI2: Virtual KITTI 2, Hypersim-U: treat Hypersim as unlabeled video data). Lower AbsRel $(\\downarrow)$ and higher $\\delta 1$ $(\\uparrow)$ values indicate better performance. For each video segment, we align using only a single scale, setting the shift to zero." + ], + "table_footnote": [], + "table_body": "
MethodMTD (↓)Video Model?Online?FPS (↑)
DepthCrafter [23]0.003479✓×0.45
Baseline0.003777×✓5.77
StableDepth (Ours)0.003482×✓5.77
", + "bbox": [ + 94, + 496, + 478, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "our baseline (metric depth version of DAv2-Large) enables online inference at 5.77 FPS, it shows relatively higher inconsistency (MTD=0.003777). StableDepth significantly improves upon our baseline with a $7.8\\%$ reduction in MTD (0.003482), achieving comparable consistency to DepthCrafter. Crucially, StableDepth maintains online inference capability, processing frames independently without requiring past or future frames, and operates at 5.77 FPS on a H20 GPU - $13 \\times$ faster than DepthCrafter. Both methods were evaluated on $1920 \\times 1080$ resolution images using half-precision (FP16) for inference speed testing. These results demonstrate that StableDepth successfully combines strong temporal consistency with practical advantages in", + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/277ed3ad43b53cfa9eafb921540613bc0b19ab0efacb43d181812e8b98b914ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 311, + 645, + 368 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6fbe2d51ca205d8f78e9fda27356074c98e6068528a039780413474603878c8d.jpg", + "image_caption": [ + "RGB" + ], + "image_footnote": [], + "bbox": [ + 519, + 368, + 645, + 424 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b0b71a1817c0e70862ef88b7e68b82c3c904b8a719798575cd20183c448a8e58.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 311, + 772, + 368 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8b42e1b69ca34ea54d9dec765e23f973b34f040f5bca98ac35f7c497922f82d2.jpg", + "image_caption": [ + "DepthCrafter [23]" + ], + "image_footnote": [], + "bbox": [ + 645, + 368, + 772, + 424 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/94b3ea5978e55dc36625501f657bb921d531e3db558defe960306e67a32f359e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 311, + 900, + 368 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/14f8d1b0091cd4f616700935be367bdb5481ed2c9f9cdfa82d77f0e10049c710.jpg", + "image_caption": [ + "StableDepth (Ours)", + "Figure 4. Qualitative comparisons. StableDepth shows improved robustness in handling ambiguous regions such as water surfaces and sky boundaries while maintaining accurate depth estimation for the main subjects (flamingo, horse) compared to DepthCrafter. The highlighted region in DepthCrafter results shows inconsistent depth predictions in these challenging regions." + ], + "image_footnote": [], + "bbox": [ + 774, + 369, + 900, + 424 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/46c2bf6a72ad81905606bdab6b168807f9b3322b356a81706caa0fcb611a2bbd.jpg", + "table_caption": [ + "Table 5. Performance comparison of different depth estimation methods on DAVIS dataset. MTD: Motion-aware Temporal Difference (Eq. (9)). Online: whether the method can process data without requiring the entire video. FPS: inference speed in frames per second. StableDepth achieves comparable consistency to DepthCrafter that based on heavy SVD [3], while maintaining online capability and significantly faster inference." + ], + "table_footnote": [], + "table_body": "
DataPretrainedBonnScanNet
AbsRel (↓)δ1(↑)AbsRel (↓)δ1(↑)
LDINOv20.0860.9290.0980.911
LDAv20.0710.9460.1020.903
L + UDINOv20.0680.9490.0990.909
L + UDAv20.0570.9760.0970.913
", + "bbox": [ + 514, + 547, + 901, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Performance comparison on Bonn and ScanNet. We compare different initialization strategies: DINOv2 pre-trained weights vs. DAv2's well-trained parameters. L: labeled data only; L+U: both labeled and unlabeled data with dual-decoder.", + "bbox": [ + 511, + 659, + 906, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "deployment, addressing a key limitation in current depth estimation approaches.", + "bbox": [ + 511, + 732, + 903, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison of unlabeled video frames. As shown in Fig. 3, MTD scores improve steadily with more unlabeled video frames in each mini-batch and gradually converge at 16 frames, with negligible gains beyond this point, supporting our choice of 16-frame clips as an optimal trade-off.", + "bbox": [ + 511, + 763, + 905, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison of different pre-trained encoders. Tab. 6 shows the effectiveness of our semi-supervised training strategy on the Bonn and ScanNet datasets. We compare two initialization strategies: DINOv2 and DAv2 pre-trained", + "bbox": [ + 511, + 840, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7075", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bb0ac1a715090252d9205b98f030c4d49333ce8d6b00bda453f229ff5b994036.jpg", + "image_caption": [ + "Figure 5. Qualitative visualizations of depth predictions across diverse scenarios. Each row pair shows an input sequence and corresponding depth predictions from our method. Our approach maintains consistent depth estimation while preserving fine geometric details across challenging cases including dynamic motions (dancer), reflective surfaces (water), and large objects (rhinoceros)." + ], + "image_footnote": [], + "bbox": [ + 102, + 88, + 893, + 443 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "weights. With only labeled data (L), DAv2 initialization outperforms DINOv2, particularly on Bonn (0.071 vs 0.086 AbsRel). Incorporating unlabeled video data $(\\mathrm{L} + \\mathrm{U})$ through our dual-decoder approach yields consistent improvements across both datasets. DAv2 with $\\mathrm{L} + \\mathrm{U}$ achieves the best performance, reducing AbsRel to 0.057 on Bonn and 0.097 on ScanNet, demonstrating the effectiveness of our semi-supervised approach.", + "bbox": [ + 89, + 525, + 483, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Qualitative Results", + "text_level": 1, + "bbox": [ + 89, + 664, + 272, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method demonstrates robust performance across diverse real-world scenarios. As shown in Fig. 1, it generates stable depth maps even in challenging conditions, capturing dynamic motions in dog agility courses, street performances, and natural scenes while maintaining consistent geometry frame-by-frame. Fig. 4 shows our method's superiority over DepthCrafter in ambiguous depth cue cases. In the flamingo scene, DepthCrafter produces inconsistent depth predictions for water reflections, while our method maintains coherent depth across the water surface. In the horse sequence, DepthCrafter mistakenly estimates the sky as foreground, whereas our approach handles the background-sky transition more accurately. These results highlight our model's robustness to common depth estima", + "bbox": [ + 89, + 689, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tion challenges like reflective surfaces and infinite-depth regions, while preserving scene consistency. Fig. 5 further demonstrates the effectiveness of our method in challenging scenarios, including fast-moving subjects (dancer) and large objects (rinoceros). Our approach consistently preserves fine geometric details and maintains stable depth relationships across frames, even under complex lighting conditions and dynamic motions. The temporal smoothness achieved enables reliable depth estimation for practical applications requiring consistent 3D understanding.", + "bbox": [ + 511, + 525, + 906, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 690, + 633, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present StableDepth, a framework that bridges the gap between relative and metric depth estimation through our scene-consistent and scale-invariant (SCSI) paradigm. By leveraging our dual-decoder architecture and video diffusion baking strategy, we effectively decouple depth accuracy from temporal consistency learning, while harnessing knowledge from unlabeled video data. StableDepth enables online inference at $13 \\times$ faster speed than previous video-based methods while maintaining comparable temporal stability. This approach advances monocular depth estimation toward more practical applications in 3D reconstruction, robotics, and augmented reality.", + "bbox": [ + 511, + 715, + 906, + 897 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7076", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work is supported by the National Natural Science Foundation of China (No. 62422606, 62201484) and DAMO Academy Research Intern Program.", + "bbox": [ + 89, + 90, + 485, + 154 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 180, + 187, + 195 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv:2302.12288, 2023.5, 6", + "[2] Reiner Birkl, Diana Wofk, and Matthias Muller. Midas v3. 1-a model zoo for robust monocular relative depth estimation. arXiv:2307.14460, 2023. 4", + "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv:2311.15127, 2023. 2, 5, 7", + "[4] Aleksei Bochkovskii, Amaël Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. In ICLR, 2025. 2, 3", + "[5] D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black. A naturalistic open source movie for optical flow evaluation. In ECCV, 2012. 2, 5", + "[6] Yohann Cabon, Naila Murray, and Martin Humenberger. Virtual katii 2. arXiv:2001.10773, 2020. 5, 6, 7", + "[7] Sili Chen, Hengkai Guo, Shengnan Zhu, Feihu Zhang, Zi long Huang, Jiashi Feng, and Bingyi Kang. Video depth anything: Consistent depth estimation for super-long videos. In CVPR, 2025. 3", + "[8] Weifeng Chen, Zhao Fu, Dawei Yang, and Jia Deng. Single-image depth perception in the wild. In NeurIPS, 2016. 3", + "[9] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 2, 5", + "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 4, 5", + "[11] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In ICCV, 2021. 3", + "[12] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. In NeurIPS, 2014. 1, 3", + "[13] Fatima El Jamiy and Ronald Marsh. Survey on depth perception in head mounted displays: distance estimation in virtual reality, augmented reality, and mixed reality. IET Image Processing, 2019. 2", + "[14] Gunnar Farnebäck. Two-frame motion estimation based on polynomial expansion. In SCIA, 2003. 5" + ], + "bbox": [ + 93, + 205, + 480, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Xiao Fu, Wei Yin, Mu Hu, Kaixuan Wang, Yuexin Ma, Ping Tan, Shaojie Shen, Dahua Lin, and Xiaoxiao Long. Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image. In ECCV, 2024. 2", + "[16] Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as proxy for multi-object tracking analysis. In CVPR, 2016. 5, 6, 7", + "[17] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. *IJRR*, 2013. 2, 5", + "[18] Ming Gui, Johannes S Fischer, Ulrich Prestel, Pingchuan Ma, Dmytro Kotovenko, Olga Grebenkova, Stefan Andreas Baumann, Vincent Tao Hu, and Björn Ommer. Depthfm: Fast monocular depth estimation with flow matching. In AAAI, 2025. 2", + "[19] Vitor Guizilini, Igor Vasiljevic, Dian Chen, Rares Ambrus, and Adrien Gaidon. Towards zero-shot scale-aware monocular depth estimation. In ICCV, 2023. 3", + "[20] S Mahdi H. Miangoleh, Mahesh Reddy, and Yaqiz Aksoy. Scale-invariant monocular depth estimation via ssi depth. In SIGGRAPH, 2024. 2", + "[21] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In NeurIPS, 2022. 3", + "[22] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. TPAMI, 2024. 2, 3", + "[23] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. In CVPR, 2025. 2, 3, 4, 5, 6, 7", + "[24] Yaosi Hu, Chong Luo, and Zhenzhong Chen. Make it move: controllable image-to-video generation with text descriptions. In CVPR, 2022. 3", + "[25] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In CVPR, 2024. 2", + "[26] Bingxin Ke, Dominik Narnhofer, Shengyu Huang, Lei Ke, Torben Peters, Katerina Fragkiadaki, Anton Obukhov, and Konrad Schindler. Video depth without video models. In CVPR, 2025. 3", + "[27] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In CVPR, 2021. 3", + "[28] Yixuan Li, Lihan Jiang, Linning Xu, Yuanbo Xiangli, Zhenzhi Wang, Dahua Lin, and Bo Dai. Matrixcity: A large-scale city dataset for city-scale neural rendering and beyond. In ICCV, 2023. 5, 6, 7", + "[29] Zhaoshuo Li, Wei Ye, Dilin Wang, Francis X Creighton, Russell H Taylor, Ganesh Venkatesh, and Mathias Unberath. Temporally consistent online depth estimation in dynamic scenes. In WACV, 2023. 3", + "[30] LightwheelAI and LightwheelOcc contributors. LightwheelOcc: A 3d occupancy synthetic dataset in autonomous driving. https://github.com/OpenDriveLab/LightwheelOcc, 2024.5,6,7" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7077", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Nian Liu, Ni Zhang, Ling Shao, and Junwei Han. Learning selective mutual attention and contrast for rgb-d saliency detection. TPAMI, 2021. 3", + "[32] Xingtong Liu, Ayushi Sinha, Masaru Ishii, Gregory D Hager, Austin Reiter, Russell H Taylor, and Mathias Unberath. Dense depth estimation in monocular endoscopy with self-supervised learning methods. TMI, 2019. 2", + "[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 5", + "[34] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. In SIGGRAPH, 2020. 3", + "[35] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. TMLR, 2023. 2", + "[36] E. Palazzolo, J. Behley, P. Lottes, P. Giguère, and C. Stachniss. ReFusion: 3D Reconstruction in Dynamic Environments for RGB-D Cameras Exploiting Residuals. In IROS, 2019. 2, 5", + "[37] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Unidepth: Universal monocular metric depth estimation. In CVPR, 2024. 2, 3", + "[38] Luigi Piccinelli, Christos Sakaridis, Yung-Hsu Yang, Mattia Segu, Siyuan Li, Wim Abbeloos, and Luc Van Gool. Unidepthv2: Universal monocular metric depth estimation made simpler. arXiv:2502.20110, 2025. 3", + "[39] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. TPAMI, 2020. 2, 3, 4", + "[40] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, 2021. 3, 4, 5", + "[41] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. In ICLR, 2025. 5, 6, 7", + "[42] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In ICCV, 2021. 3, 5, 6, 7", + "[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2", + "[44] Abhinav Sagar. Monocular depth estimation using multi-scale neural network and feature fusion. In WACV, 2022. 3", + "[45] Ashutosh Saxena, Min Sun, and Andrew Y Ng. Make3d: Learning 3d scene structure from a single still image. TPAMI, 2008. 1", + "[46] Mohamed Sayed, John Gibson, Jamie Watson, Victor Prisacariu, Michael Firman, and Clément Godard. Simplere" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "con: 3d reconstruction without 3d convolutions. In ECCV, 2022. 3", + "[47] Jiahao Shao, Yuanbo Yang, Hongyu Zhou, Youmin Zhang, Yujun Shen, Matteo Poggi, and Yiyi Liao. Learning temporally consistent video depth from video diffusion priors. 2025. 2, 3", + "[48] Chang Shu, Kun Yu, Zhixiang Duan, and Kuiyuan Yang. Feature-metric loss for self-supervised learning of depth and egomotion. In ECCV, 2020. 2", + "[49] Chaoyang Wang, Simon Lucey, Federico Perazzi, and Oliver Wang. Web stereo video supervision for depth prediction from dynamic scenes. In 3DV, 2019. 3", + "[50] Lijun Wang, Yifan Wang, Linzhao Wang, Yunlong Zhan, Ying Wang, and Huchuan Lu. Can scale-consistent monocular depth be learned in a self-supervised scale-invariant manner? In ICCV, 2021. 3", + "[51] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In CVPR, 2019. 2", + "[52] Yiran Wang, Min Shi, Jiaqi Li, Zihao Huang, Zhiguo Cao, Jianming Zhang, Ke Xian, and Guosheng Lin. Neural video depth stabilizer. In ICCV, 2023. 3", + "[53] Jamie Watson, Mohamed Sayed, Zawar Qureshi, Gabriel J Brostow, Sara Vicente, Oisin Mac Aodha, and Michael Firman. Virtual occlusions through implicit depth. In CVPR, 2023. 3", + "[54] Diana Wofk, Fangchang Ma, Tien-Ju Yang, Sertac Karaman, and Vivienne Sze. Fastdepth: Fast monocular depth estimation on embedded systems. In ICRA, 2019. 2", + "[55] Ke Xian, Jianming Zhang, Oliver Wang, Long Mai, Zhe Lin, and Zhiguo Cao. Structure-guided ranking loss for single image depth prediction. In CVPR, 2020. 3", + "[56] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 2, 3, 4", + "[57] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. In NeurIPS, 2024. 3, 4, 5", + "[58] Rajeev Yasarla, Hong Cai, Jisoo Jeong, Yunxiao Shi, Risheek Garrepalli, and Fatih Porikli. Mamo: Leveraging memory and attention for monocular video depth estimation. In ICCV, 2023. 3", + "[59] Wei Yin, Chi Zhang, Hao Chen, Zhipeng Cai, Gang Yu, Kaixuan Wang, Xiaozhi Chen, and Chunhua Shen. Metric3d: Towards zero-shot metric 3d prediction from a single image. In ICCV, 2023. 2, 3, 5, 6", + "[60] Sen Zhang, Jing Zhang, and Dacheng Tao. Towards scale-aware, robust, and generalizable unsupervised monocular depth estimation by integrating imu motion dynamics. In ECCV, 2022. 3", + "[61] Ruijie Zhu, Chuxin Wang, Ziyang Song, Li Liu, Tianzhu Zhang, and Yongdong Zhang. Scaledepth: Decomposing metric depth estimation into scale prediction and relative depth estimation. arXiv:2407.08187, 2024. 3" + ], + "bbox": [ + 516, + 92, + 903, + 885 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7078", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/StableDepth_ Scene-Consistent and Scale-Invariant Monocular Depth/73601462-88ee-4d47-a92f-ac7545019662_model.json b/2025/StableDepth_ Scene-Consistent and Scale-Invariant Monocular Depth/73601462-88ee-4d47-a92f-ac7545019662_model.json new file mode 100644 index 0000000000000000000000000000000000000000..236d4dd8b37566e50809d53c1475e0f577b25151 --- /dev/null +++ b/2025/StableDepth_ Scene-Consistent and Scale-Invariant Monocular Depth/73601462-88ee-4d47-a92f-ac7545019662_model.json @@ -0,0 +1,2255 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.148, + 0.131, + 0.851, + 0.154 + ], + "angle": 0, + "content": "StableDepth: Scene-Consistent and Scale-Invariant Monocular Depth" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.777, + 0.218 + ], + "angle": 0, + "content": "Zheng Zhang\\(^{1,2,\\star,*}\\) Lihe Yang\\(^{1,\\star}\\) Tianyu Yang\\(^{2,4,\\dagger}\\) Chaohui Yu\\(^{2,4}\\) Xiaoyang Guo\\(^{3}\\) Yixing Lao\\(^{1}\\) Hengshuang Zhao\\(^{1,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.219, + 0.77, + 0.238 + ], + "angle": 0, + "content": "1The University of Hong Kong 2DAMO Academy, Alibaba Group" + }, + { + "type": "text", + "bbox": [ + 0.284, + 0.238, + 0.715, + 0.256 + ], + "angle": 0, + "content": "3The Chinese University of Hong Kong 4Hupan Lab" + }, + { + "type": "text", + "bbox": [ + 0.412, + 0.258, + 0.582, + 0.272 + ], + "angle": 0, + "content": "https://stabledepth.github.io" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.307, + 0.895, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.396, + 0.895, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.485, + 0.895, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.573, + 0.895, + 0.661 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.676, + 0.908, + 0.719 + ], + "angle": 0, + "content": "Figure 1. StableDepth achieves efficient online monocular depth estimation that produces scene-consistent and scale-invariant predictions frame by frame. As shown in the above challenging scenarios, StableDepth generates stable depth maps with precise geometric details and effectively handles dynamic motions, without the need of accessing future frames." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.73, + 0.327, + 0.746 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.762, + 0.486, + 0.869 + ], + "angle": 0, + "content": "Recent advances in monocular depth estimation significantly improve robustness and accuracy. However, relative depth models exhibit flickering and 3D inconsistency in video data, limiting 3D reconstruction applications. We introduce StableDepth, a scene-consistent and scale-invariant depth estimation method achieving scene-level 3D consistency. Our dual-decoder architecture learns from" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.876, + 0.483, + 0.901 + ], + "angle": 0, + "content": "\\(\\star\\) Joint first authors. *Work done when interning at DAMO Academy. † Corresponding author." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.732, + 0.907, + 0.824 + ], + "angle": 0, + "content": "large-scale unlabeled video data, enhancing generalization and reducing flickering. Unlike previous methods requiring full video sequences, StableDepth enables online inference at \\(13 \\times\\) faster speed, achieving significant improvements across benchmarks with comparable temporal consistency to video diffusion-based estimators." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.829, + 0.646, + 0.845 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Monocular Depth Estimation (MDE) [12, 45] is pivotal in bridging the gap between 2D and 3D data representation, serving critical roles in a wide range of downstream applica" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7069" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "tions, such as autonomous driving [51], robotics [54], medical imaging [32], and virtual reality [13]. Advancements in foundational models [35, 43] have greatly enhanced the accuracy and robustness of MDE [25, 56]. Despite these advances, the intrinsic ambiguity of monocular cues still poses substantial challenges, particularly for downstream applications that require 3D-consistent data and maintaining stability for uncontrolled in-the-wild data." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.215, + 0.485, + 0.532 + ], + "angle": 0, + "content": "MDE is primarily categorized into two main task settings: relative depth estimation and metric depth estimation. Relative depth estimation, as explored in studies [15, 18, 25, 39, 56], involves predicting depth values within a normalized depth range of \\([0, 1]\\). This scale- and shift-invariant regression target facilitates robust joint training across diverse data domains. However, the predicted depth values differ from the actual depth due to scale and shift discrepancies, leading to inconsistency across video frames. This inconsistency can complicate advanced downstream tasks like 3D reconstruction. In contrast, metric depth estimation [4, 22, 37, 59] aims to predict physical distances in metric scale. However, these approaches face significant challenges. Scaled versions of a scene's depth values might all represent plausible depth estimations of the actual depth. Such ambiguity complicates the learning process of depth regression, hindering both generalization capabilities and precise 3D structures. Severe flickering and 3D inconsistency are also observed given video inputs. This challenge prompts us to reconsider what is an effective representation of MDE foundation models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.536, + 0.483, + 0.686 + ], + "angle": 0, + "content": "The limitations of existing approaches are summarized in Tab. 1, which compares different depth estimation paradigms. While relative depth models [25, 39, 56] offer good generalization across datasets but suffer from per-frame inconsistency, metric depth models [4, 59] provide real-world measurements but struggle with generalization across diverse scenes. Scale-invariant models [20] improve upon relative models by eliminating shift ambiguity while maintaining scale invariance, but they still lack consistency across frames in the same scene." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To address these challenges, we propose that the ideal task setting should be scene-consistent and scale-invariant (SCSI) depth estimation, as shown in Tab. 1. Our approach strikes a balance between per-frame scale-invariant methods that lack consistency and purely metric approaches that struggle with generalization. Specifically, our model primarily generates scale-invariant depth maps, but in the same scene or video, predictions across multiple frames demonstrate 3D consistency. Additionally, the scale factors for multi-frame data remain consistent compared with the actual depth values. In this work, the development of our model is primarily guided by the following objectives: 1) ensuring high-quality, scene-consistent, and scale-invariant depth estimations to achieve robust video consis" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.091, + 0.905, + 0.168 + ], + "angle": 0, + "content": "
Depth TypeScaleShiftGeneralReconsConsist
RelativePer-framePer-frame✓✗✗
Scale-invariantPer-frameNone✓✓✗
MetricNoneNone✗✓✗
SCSI (Ours)Per-sceneNone✓✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.176, + 0.907, + 0.232 + ], + "angle": 0, + "content": "Table 1. Comparison of different depth types. Our proposed SCSI depth combines the strengths of relative and metric depth while overcoming their weaknesses. General: generalization capability. Recons: reconstruction quality. Consist: scene consistency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.248, + 0.905, + 0.278 + ], + "angle": 0, + "content": "tency; 2) enhancing strong generalization capabilities and precise depth details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.28, + 0.907, + 0.521 + ], + "angle": 0, + "content": "Our framework introduces three key innovations to achieve scene-consistent depth estimation. First, we leverage temporal priors from pre-trained video diffusion models to generate pseudo-labels with inherent 3D consistency, replacing traditional unstable photometric constraints [3, 23, 47, 48]. Second, we design a dual-branch architecture where one branch processes labeled data for metric accuracy, while the other aligns unlabeled video frames to pseudo-labels through scene-level normalization, decoupling conflicting objectives. Third, our semi-supervised paradigm jointly optimizes metric supervision and temporal alignment, enabling the model to generalize across domains while eliminating flickering artifacts. This unified approach supports real-time inference without requiring future frames, overcoming the limitations of both single-frame and video-based methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.906, + 0.688 + ], + "angle": 0, + "content": "We evaluate our model on multiple datasets under zero-shot settings, aligning each scene using only a single scale (SCSI). As demonstrated in Figure 1, our method robustly handles challenging scenarios including dynamic motions and complex textures while preserving geometric fidelity frame-by-frame. Compared to DepthCrafter [23], we achieve \\(13 \\times\\) faster inference speed with online processing capability, improving prediction accuracy on Sintel [5], Bonn [36], ScanNet v2 [9], and KITTI [17] by \\(13.2\\%\\), \\(86.8\\%\\), \\(39.3\\%\\), and \\(8.2\\%\\), respectively. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.69, + 0.905, + 0.765 + ], + "angle": 0, + "content": "- We introduce StableDepth, a novel monocular depth estimation framework that achieves scene-consistent and scale-invariant (SCSI) depth estimation, bridging the gap between relative and metric depth prediction while maintaining temporal consistency in video sequences." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.766, + 0.905, + 0.841 + ], + "angle": 0, + "content": "- We propose a video diffusion baking strategy that leverages pre-trained video diffusion models for generating high-quality pseudo-labels, enabling effective semi-supervised learning from unlabeled video data while preserving temporal consistency." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.841, + 0.905, + 0.902 + ], + "angle": 0, + "content": "- We develop a dual-decoder architecture that simultaneously handles supervised metric depth estimation and unsupervised depth alignment, effectively addressing the target misalignment issue while maintaining accuracy." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.69, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7070" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "- We demonstrate state-of-the-art performance across multiple benchmarks, achieving up to \\(86.8\\%\\) improvement in accuracy while providing \\(13 \\times\\) faster inference compared to previous methods. Our method enables efficient, online inference while maintaining comparable temporal stability to video diffusion models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.198, + 0.232, + 0.213 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.224, + 0.484, + 0.389 + ], + "angle": 0, + "content": "Monocular depth estimation. MDE focuses on inferring depth from a single image. Early works [8, 31, 55] learn ordinal relationships from coarse annotations but cannot recover geometric structure. Recent methods [11, 39, 56, 57] leverage large-scale data to learn affine-invariant depth with improved generalization. However, unknown depth shifts cause geometric distortions, hindering 3D applications. To address this, recent works attempt zero-shot metric depth estimation [4, 19, 22, 37, 38, 59, 61] and explore self-supervised scale-consistent learning [50, 60], though maintaining consistent scale across scenes remains challenging." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.391, + 0.484, + 0.601 + ], + "angle": 0, + "content": "Towards consistent video depth. Video depth estimation requires both temporal consistency and per-frame precision. Traditional approaches use post-processing optimization [27, 34] or temporal modeling [29, 49], but incur computational costs or fail in dynamic scenes. Other methods leverage camera poses [46, 53] or memory-based approaches [52, 58], but often sacrifice geometric fidelity for smoothness. Recent works like Video Depth Anything [7] and video-free approaches [26] show promise for temporal consistency in long sequences. Video Depth Anything freezes the Depth Anything V2 encoder and incorporates cross-frame attention in the DPT [40] decoder for temporal consistency, though these methods still produce depth with inherent shift ambiguity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.603, + 0.484, + 0.722 + ], + "angle": 0, + "content": "Diffusion priors for depth modeling. Video diffusion models [21, 24] demonstrate potential for consistent depth estimation. DepthCrafter [23] achieves temporal stability but generates only relative depths, while ChronoDepth [47] suffers from limited temporal contexts. Our proposed SCSI depth estimation leverages video diffusion baking and dual-decoder architecture for temporally consistent depth suitable for efficient online inference." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.739, + 0.219, + 0.755 + ], + "angle": 0, + "content": "3. StableDepth" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Our work leverages both labeled images and unlabeled video data to enhance SCSI depth predictions, achieving video consistency and enhanced generalization capabilities through the distillation of video diffusion priors. Formally, we denote the sets of labeled image-depth pairs and unlabeled videos as \\(\\mathcal{D}^l = \\{(x_i,d_i)\\}_{i = 1}^M\\) and \\(\\mathcal{D}^u = \\{v_i\\}_{i = 1}^N\\), respectively. Our objective is to utilize a high-quality labeled image dataset, Hypersim [42], to help the model grasp real-world scale, while employing unlabeled video data to guide" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "the model's depth predictions toward achieving video consistency and improved generalization." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.133, + 0.882, + 0.149 + ], + "angle": 0, + "content": "3.1. Mining Consistency from Unlabeled Videos" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.906, + 0.267 + ], + "angle": 0, + "content": "StableDepth is designed to learn from both labeled images and unlabeled videos. For labeled image data, we predict the depth in metric space. The activation function of the network's last layer is sigmoid, after which its predictions are scaled to match the dataset values by multiplying with a predefined maximum depth. The SiLogLoss [12, 44, 57] is then used to minimize the error between the prediction and the ground truth:" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.277, + 0.905, + 0.331 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {l a b e l}} = \\sqrt {\\frac {1}{H W} \\sum_ {i = 1} ^ {H W} \\left(\\log \\frac {d _ {i}}{d _ {i} ^ {*}}\\right) ^ {2} - \\lambda \\left(\\frac {1}{H W} \\sum_ {i = 1} ^ {H W} \\log \\frac {d _ {i}}{d _ {i} ^ {*}}\\right) ^ {2}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.332, + 0.905, + 0.362 + ], + "angle": 0, + "content": "where \\(d_i^*\\) and \\(d_{i}\\) are the prediction and ground truth, respectively, and \\(\\lambda\\) is a hyper-parameter." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.363, + 0.907, + 0.695 + ], + "angle": 0, + "content": "For unlabeled video data, we leverage video diffusion models, specifically DepthCrafter [23], to generate pseudolabels due to their inherent advantages in maintaining temporal consistency. These models, pre-trained on large-scale video data with cross-frame attention mechanisms, can generate stable depth predictions across entire video segments while preserving a consistent global scale and shift relationship with metric depth. Unlike conventional image-based MDE approaches that process frames independently, video diffusion models account for temporal dependencies, making them particularly effective for generating reliable pseudo-labels that exhibit strong video consistency - a crucial property for our scene-consistent training objective. For a video, we need a method to normalize the depth predicted by our model to the same space as the pseudo-labeled depth. Unlike the normalization used in single-frame relative depth training, this approach focuses on ensuring that an entire video has a unique scale and shift, rather than each frame having an independent scale and shift. This method aims to achieve video consistency in the depth predictions within the same scene. To supervise our model, we first use the following formula to normalize the pseudo-labels," + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.706, + 0.905, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\hat {D} (v _ {i}) = \\frac {D (v _ {i}) - m (D (v _ {i}))}{s (D (v _ {i}))}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.75, + 0.905, + 0.794 + ], + "angle": 0, + "content": "where \\(D(v_{i})\\) is the pseudo-label for video \\(v_{i}\\), and \\(m(D(v_{i}))\\) and \\(s(D(v_{i}))\\) are used to remove the shift and scale for this video, respectively:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.805, + 0.741, + 0.82 + ], + "angle": 0, + "content": "\\[\nm (D (v _ {i})) = \\operatorname {m e d i a n} (D (v _ {i})),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.824, + 0.905, + 0.86 + ], + "angle": 0, + "content": "\\[\ns \\left(D \\left(v _ {i}\\right)\\right) = \\frac {1}{T H W} \\sum_ {i = 1} ^ {T H W} | D \\left(v _ {i}\\right) - m \\left(D \\left(v _ {i}\\right)\\right) |. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "For the frame-by-frame depth predictions of video \\( v_{i} \\) from our model, we concatenate them to obtain \\( P(v_{i}) \\) with" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7071" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.087, + 0.905, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.255, + 0.907, + 0.34 + ], + "angle": 0, + "content": "Figure 2. Overview of StableDepth. StableDepth employs a shared encoder with dual decoders to process both labeled images and unlabeled videos. The upper branch handles labeled images with direct metric depth supervision (\"Sup\"), while the lower branch processes unlabeled video sequences using temporal-coherent pseudo-labels generated by pre-trained video diffusion priors [23] (\"Bake\"). This architecture enables scene-consistent and scale-invariant (SCSI) depth prediction by combining the benefits of metric depth supervision with temporal consistency learning from video sequences. Solid arrows indicate the primary data flow, while dashed arrows represent the auxiliary training path for unlabeled data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.355, + 0.485, + 0.416 + ], + "angle": 0, + "content": "dimensions \\([T,H,W]\\). We then apply the same normalization process to the depth predictions \\(P(v_{i})\\) to obtain its normalized version \\(\\hat{P} (v_{i})\\). For \\(\\hat{P} (v_{i})\\) and \\(\\hat{D} (v_{i})\\), we use the following alignment loss to minimize their difference," + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.426, + 0.483, + 0.468 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {u n l a b e l}} = \\frac {1}{T H W} \\sum_ {i = 1} ^ {T H W} \\left| \\hat {P} \\left(v _ {i}\\right) - \\hat {D} \\left(v _ {i}\\right) \\right|. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.484, + 0.537 + ], + "angle": 0, + "content": "In one iteration, we simultaneously sample a certain proportion of labeled images and unlabeled videos for joint training, and use the following combined loss as the final loss for this iteration:" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.55, + 0.483, + 0.567 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {l a b e l}} + \\gamma \\mathcal {L} _ {\\text {u n l a b e l}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.577, + 0.484, + 0.608 + ], + "angle": 0, + "content": "where \\(\\gamma\\) gradually increases following a cosine curve as the number of iterations progresses." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.616, + 0.388, + 0.633 + ], + "angle": 0, + "content": "3.2. Avoiding Conflict in Depth Labels" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.638, + 0.483, + 0.698 + ], + "angle": 0, + "content": "In the field of video consistency for depth prediction, most research [23] focuses on improving video consistency in near-field scenes. Therefore, we use depth in the disparity space for video data training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.699, + 0.483, + 0.82 + ], + "angle": 0, + "content": "Following the previous network structure [2, 39, 56, 57], we use the ViT-L encoder [10], denoted as \\(\\mathcal{E}\\), paired with the decoder from DPT [40], denoted as \\(\\mathcal{D}_1\\). The decoder outputs depth in SCSI space with a range of (0, 1). A natural approach is to utilize this original decoder to predict the disparity-space depth of a video sequence, and then take the reciprocal of the predictions. Specifically, the depth prediction \\(P(v_{i})\\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.83, + 0.483, + 0.862 + ], + "angle": 0, + "content": "\\[\nP \\left(v _ {i}\\right) = \\frac {1}{\\mathcal {D} _ {1} \\left(\\mathcal {E} \\left(v _ {i}\\right)\\right)}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Finally, the depth predictions are normalized and supervised in semi-supervised settings following Eq. 3 and 4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.355, + 0.907, + 0.536 + ], + "angle": 0, + "content": "Unfortunately, in our experiments, we found that although this modification improves the video consistency of depth predictions for videos, it reduces the accuracy of depth prediction by the model, resulting in lower metrics such as \\(\\delta_{1}\\). We speculate that this is mainly due to the used pseudo-labeling model [23], which, despite having good video consistency, still lags behind Depth Anything V2 [57] in terms of depth prediction accuracy, particularly in the prediction of distant objects. To address this issue, we propose to use an auxiliary network to decouple the prediction and supervision of labeled images and unlabeled videos. To this end, we experimented with two approaches:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.906, + 0.612 + ], + "angle": 0, + "content": "In the first approach, both labeled images and unlabeled videos share the encoder \\(\\mathcal{E}\\) and the decoder \\(\\mathcal{D}_1\\). A lightweight CNN network, denoted as \\(\\mathcal{C}\\), is used to directly process the prediction results of the original decoder \\(\\mathcal{D}_1\\) with the CNN network outputting depth in disparity space:" + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.619, + 0.905, + 0.637 + ], + "angle": 0, + "content": "\\[\nP \\left(v _ {i}\\right) = \\mathcal {C} \\left(\\mathcal {D} _ {1} \\left(\\mathcal {E} \\left(v _ {i}\\right)\\right)\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.906, + 0.735 + ], + "angle": 0, + "content": "In the second approach, labeled images and unlabeled videos share the encoder \\(\\mathcal{E}\\). An additional DPT [40] decoder, denoted as \\(\\mathcal{D}_2\\), is used to process the features extracted by the encoder \\(\\mathcal{E}\\) for unlabeled videos, as illustrated by Fig. 2. \\(\\mathcal{D}_2\\) directly produces depth in disparity space, while \\(\\mathcal{D}_1\\) continues to output depth in SCSI space." + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.742, + 0.905, + 0.759 + ], + "angle": 0, + "content": "\\[\nP \\left(x _ {i}\\right) = \\mathcal {D} _ {1} \\left(\\mathcal {E} \\left(x _ {i}\\right)\\right), \\quad P \\left(v _ {i}\\right) = \\mathcal {D} _ {2} \\left(\\mathcal {E} \\left(v _ {i}\\right)\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.906, + 0.827 + ], + "angle": 0, + "content": "The final results indicate that sharing the encoder between labeled images and unlabeled videos, while using two independent DPT decoders, achieves the best accuracy and video consistency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.821, + 0.85 + ], + "angle": 0, + "content": "3.3. Scene-Consistent Depth Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Scale alignment. For zero-shot evaluation, we enforce single-scale consistency per scene through RANSAC-based alignment with zero shift constraint. Each scene's" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7072" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.215 + ], + "angle": 0, + "content": "predictions are scaled by a unified factor determined through 1,000 sampling iterations, retaining inliers with \\(\\max \\left(\\frac{d_{pred}}{d_{gt}},\\frac{d_{gt}}{d_{pred}}\\right) < \\tau\\), where \\(\\tau = 1.25\\). The optimal scale is refined via least-squares optimization over all inliers, preserving geometric fidelity while eliminating cross-frame scale ambiguity. This per-scene single-scale paradigm ensures coherent depth relationships across dynamic sequences without per-frame adjustments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.217, + 0.482, + 0.263 + ], + "angle": 0, + "content": "Motion-aware temporal metric. We propose the Motion-aware Temporal Difference (MTD) to quantify depth consistency in dynamic regions:" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.277, + 0.483, + 0.318 + ], + "angle": 0, + "content": "\\[\n\\mathrm {M T D} = \\frac {1}{T - 1} \\sum_ {t = 1} ^ {T - 1} \\frac {\\sum_ {x , y} | \\Delta D _ {t} | \\cdot W _ {t}}{H \\times W}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.332, + 0.483, + 0.407 + ], + "angle": 0, + "content": "where \\(\\Delta D_{t} = D_{t} - D_{t - 1}\\) and \\(W_{t}(x,y) = (1 + \\| \\mathbf{f}_{t}\\|_{2})^{-1}\\) weights the depth variations by optical flow magnitude \\(\\mathbf{f}_t\\) (computed via Farneback [14] algorithm). Lower MTD indicates better temporal stability, penalizing flickering in high-motion areas." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.422, + 0.21, + 0.436 + ], + "angle": 0, + "content": "3.4. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.446, + 0.483, + 0.49 + ], + "angle": 0, + "content": "In this section, we address several important aspects of our approach to provide further clarity on methodology, design choices, and relationships to existing work." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.492, + 0.483, + 0.702 + ], + "angle": 0, + "content": "Decouple accuracy and consistency via dual learning. Our dual-decoder architecture decouples depth accuracy from temporal consistency learning, enabling joint optimization from diverse data sources. The primary decoder specializes in per-pixel depth precision through supervised training on high-quality labeled images, while the auxiliary decoder learns scene-level coherence from unlabeled videos via pseudo-labels. This separation resolves the inherent conflict where video-based methods like DepthCrafter [23] sacrifice depth accuracy for temporal smoothness, and single-frame approaches fail to ensure cross-frame consistency. Crucially, our design preserves absolute depth scales critical for 3D reconstruction, unlike relative-depth representations in video diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Beyond video diffusion priors. We repurpose video diffusion models as sources for scene consistency rather than as direct depth estimators. While models like DepthCrafter [23] excel at generating temporally consistent sequences, they produce relative depth representations with limited accuracy, particularly for distant objects, making them unsuitable for downstream tasks requiring precise geometry. Our approach extracts the temporal consistency knowledge from these models through scene-level normalization while maintaining superior depth accuracy. This strategy enables online inference \\(13 \\times\\) faster than video-based methods while achieving both better depth precision and comparable temporal stability." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.639, + 0.108 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.116, + 0.731, + 0.132 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.139, + 0.905, + 0.395 + ], + "angle": 0, + "content": "Network parameters and training details. We use the network architecture from Depth Anything V2 (DAv2) [57], consisting of a pre-trained ViT-L [10] encoder for feature extraction and a DPT [40] decoder for depth regression. The ViT-L encoder is initialized with parameters from the DAv2 model, while the primary decoder is randomly initialized with Sigmoid activation for metric depth output. The auxiliary decoder, also initialized randomly, uses ReLU activation for disparity (inverse depth) prediction. Each batch consists of a 1:2 ratio of labeled images to unlabeled video frames, totaling 24 frames. The learning rate for the encoder is \\(2.5 \\times 10^{-6}\\), while the decoder's learning rate is 10 times larger. We use the AdamW optimizer [33] with a weight decay of 0.01 and employ a polynomial decay schedule for the learning rate. For data augmentation, we apply horizontal flipping to labeled images. In Eq. (5), the maximum value of \\(\\gamma\\) is set to 1. In Eq. (1), \\(\\lambda\\) is set to 0.5." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.397, + 0.905, + 0.608 + ], + "angle": 0, + "content": "Training dataset and pseudo-labeling. To train our model, we utilize DepthCrafter [23], a model finetuned on SVD [3], to generate video-consistent pseudo-labels for unlabeled video data. For metric depth supervision, we use Hypersim [42], a high-quality synthetic dataset with \\(\\sim 60\\mathrm{K}\\) frames, as training data for decoder \\(\\mathcal{D}_1\\). We select 6,000 videos from SA-V [41], which features dynamic foreground objects, yielding approximately \\(\\sim 200\\mathrm{K}\\) frames for pseudolabeling training. To enhance the model's performance on distant scenes where DepthCrafter shows limitations, we incorporate additional datasets with accurate depth labels including VKITTI2 [6, 16], LightwheelOcc [30], and MatrixCity [28], treating frames from the same scene as video sequences during training." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.62, + 0.634, + 0.635 + ], + "angle": 0, + "content": "4.2. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.643, + 0.905, + 0.794 + ], + "angle": 0, + "content": "Evaluation datasets. We evaluate on four datasets spanning diverse scenarios: Sintel [5] (23 synthetic videos, 50 frames each with precise depth labels), Bonn [36] (26 dynamic videos using frames 30-140 with foreground motions), ScanNet v2 [9] (100 indoor test videos sampled every third frame), and the full KITTI validation set [17] (13 outdoor driving videos with first 110 frames per sequence). This comprehensive benchmark covers synthetic animations, dynamic interactions, static indoor scenes, and real-world autonomous driving environments." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.796, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Quantitative results. We evaluate our method against strong baselines: DepthCrafter [23], known for video consistency, Metric3D [59], a state-of-the-art metric depth estimator, and ZoeDepth [1], a robust monocular depth estimation method. For fair comparison, we align all predictions within each video using a single shared scale factor across the entire scene with shift set to zero. As shown in" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7073" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.912, + 0.319 + ], + "angle": 0, + "content": "
Frame IntervalMethodSintel (~50 frames)Bonn (110 frames)ScanNet (90 frames)KITTI (110 frames)
AbsRel (↓)δ1 (↑)AbsRel (↓)δ1 (↑)AbsRel (↓)δ1 (↑)AbsRel (↓)δ1 (↑)
1DepthCrafter [23]2.9590.5450.3110.5240.3700.6610.1230.832
Zoedepth [1]0.4540.3560.2770.5620.3470.4360.2160.633
Metric3D [59]0.3720.6100.0550.9810.0590.9740.0670.977
StableDepth (Ours)0.2800.6170.0630.9790.0940.9210.1120.900
4DepthCrafter [23]6.3530.3930.5150.4170.5390.4400.2440.670
Zoedepth [1]0.4610.3570.2770.5630.3490.4360.2180.629
Metric3D [59]0.3610.6170.0550.9810.0590.9740.0670.977
StableDepth (Ours)0.2920.6060.0630.9790.0930.9210.1120.901
8DepthCrafter [23]2.0480.3980.4950.4120.5710.4230.2410.656
Zoedepth [1]0.4700.3570.2760.5620.3520.4340.2180.633
Metric3D [59]0.3630.6130.0550.9800.0600.9730.0660.977
StableDepth (Ours)0.3000.6040.0620.9800.0940.9200.1120.902
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.328, + 0.908, + 0.371 + ], + "angle": 0, + "content": "Table 2. Comparison of depth estimation methods across different frame intervals. StableDepth maintains consistent performance across varying frame intervals while achieving superior accuracy. Best results are shown in bold, second best are underlined. Metric3D results on ScanNet are shown in gray as it was trained on this dataset." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.38, + 0.482, + 0.486 + ], + "angle": 0, + "content": "
MethodBonnScanNet
AbsRel (↓)δ1(↑)AbsRel (↓)δ1(↑)
Baseline0.0710.9460.1020.903
DPT0.1090.9140.1260.861
DPT + CNN0.1080.9260.1380.829
Dual DPT0.0570.9760.0970.913
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.492, + 0.483, + 0.563 + ], + "angle": 0, + "content": "Table 3. Ablation on decoders. The baseline model is trained only on labeled image data. We compare three approaches for handling unlabeled video data: using a single DPT decoder with reciprocal output (DPT), adding a lightweight CNN head (DPT + CNN), and using dual DPT decoders (Dual DPT)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.566, + 0.483, + 0.689 + ], + "angle": 0, + "content": "Tab. 2, our method achieves superior performance across most datasets and metrics, with particularly strong improvements in AbsRel and \\(\\delta_{1}\\) metrics compared to video-based approaches. We examine the impact of frame intervals on prediction accuracy to simulate varying temporal densities in real-world scenarios. Note that compared to our ablation studies, these results reflect extended training with additional iterations for optimal performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.698, + 0.245, + 0.714 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Comparison of different decoders. In Tab. 3, we compare four methods using the Bonn and ScanNet v2 datasets: training with only labeled Hypersim data, and training with both labeled Hypersim [42] and unlabeled SA-V [41] data processed by three decoder designs. The first row (Baseline) shows results with only labeled Hypersim data, while the following rows show results with both labeled and unlabeled data. For labeled Hypersim data, depth in SCSI space is obtained directly from the DPT head output. For the unlabeled SA-V data, the method for obtaining disparity space depth varies: in the second row (DPT), it's the reciprocal of the original DPT output; in the third row (DPT" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.907, + 0.519 + ], + "angle": 0, + "content": "+ CNN), a lightweight CNN processes the DPT output; in the fourth row (Dual DPT), an additional DPT head provides the disparity space depth. The results show that using a single DPT head, either by reciprocating its output or adding a lightweight CNN, reduces prediction accuracy due to the partial inaccuracy of the pseudo-labels. However, using two independent DPT heads improves frame-to-frame consistency without impacting accuracy, validating the effectiveness of our disentangled approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.909, + 0.642 + ], + "angle": 0, + "content": "Encoder feature stability. We analyze encoder feature stability across consecutive DAVIS frames to understand our dual-decoder effectiveness. Compared to DAv2, our method achieves \\(7.1\\%\\) lower mean differences, \\(13.5\\%\\) reduced standard deviation, and \\(2.3\\%\\) higher cosine similarity. These improvements confirm that auxiliary decoder supervision enhances encoder stability, contributing to temporal consistency in depth predictions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.909, + 0.794 + ], + "angle": 0, + "content": "Effectiveness of unlabeled video data. As shown in Tab. 4, we evaluate incremental training data impact. Using only Hypersim [42] (row 1), then adding pseudo-labeled SA-V [41] data (row 2) shows mixed results - improving Bonn and ScanNet v2 but decreasing Sintel and KITTI performance. This reveals DepthCrafter's video consistency benefits but limitations in distant regions. Progressively incorporating VKITTI2 [6, 16], LightWheel [30], Hypersim [42], and MatrixCity [28] (rows 3-6) successfully addresses these limitations, achieving best performance across all datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Comparison of video consistency. Tab. 5 showcases the performance comparison of depth estimation methods on the DAVIS dataset evaluated using Motion-aware temporal Difference (MTD (Eq. (9)), lower is better). The video diffusion-based DepthCrafter [23] achieves strong consistency \\((\\mathrm{MTD} = 0.003479)\\) but requires processing the entire video sequence and runs at only 0.45 FPS. While" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7074" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.907, + 0.224 + ], + "angle": 0, + "content": "
Training DataSintel (~50 frames)Bonn (110 frames)ScanNet (90 frames)KITTI (110 frames)
AbsRel (↓)δ1 (↑)AbsRel (↓)δ1 (↑)AbsRel (↓)δ1 (↑)AbsRel (↓)δ1 (↑)
Hypersim only [42]0.3090.5780.0710.9460.1020.9030.1220.871
+ SA-V [41]0.3260.5610.0570.9760.0970.9130.1430.823
+ VKITTI2 [6, 16]0.3180.5770.0580.9760.0970.9110.1290.839
+ LightWheel [30]0.3120.5730.0600.9760.0940.9160.1340.847
+ Hypersim-U [42]0.3140.5720.0640.9760.0960.9180.1230.870
+ Matrixcity [28]0.3150.5800.0620.9790.0920.9240.1140.895
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.233, + 0.907, + 0.29 + ], + "angle": 0, + "content": "Table 4. Comparison of depth estimation performance with different training data configurations. The baseline model is trained on labeled Hypersim data only, with subsequent rows showing the cumulative effect of adding different unlabeled datasets (VKITTI2: Virtual KITTI 2, Hypersim-U: treat Hypersim as unlabeled video data). Lower AbsRel \\((\\downarrow)\\) and higher \\(\\delta 1\\) \\((\\uparrow)\\) values indicate better performance. For each video segment, we align using only a single scale, setting the shift to zero." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.313, + 0.482, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.441, + 0.483, + 0.482 + ], + "angle": 0, + "content": "Figure 3. Ablation on the number of unlabeled video frames in each batch. From 0 to 16, more frames bring better scene-level consistency. (Out of memory for ViT-Large under 32 frames)" + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.497, + 0.48, + 0.577 + ], + "angle": 0, + "content": "
MethodMTD (↓)Video Model?Online?FPS (↑)
DepthCrafter [23]0.003479✓×0.45
Baseline0.003777×✓5.77
StableDepth (Ours)0.003482×✓5.77
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.587, + 0.483, + 0.684 + ], + "angle": 0, + "content": "Table 5. Performance comparison of different depth estimation methods on DAVIS dataset. MTD: Motion-aware Temporal Difference (Eq. (9)). Online: whether the method can process data without requiring the entire video. FPS: inference speed in frames per second. StableDepth achieves comparable consistency to DepthCrafter that based on heavy SVD [3], while maintaining online capability and significantly faster inference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "our baseline (metric depth version of DAv2-Large) enables online inference at 5.77 FPS, it shows relatively higher inconsistency (MTD=0.003777). StableDepth significantly improves upon our baseline with a \\(7.8\\%\\) reduction in MTD (0.003482), achieving comparable consistency to DepthCrafter. Crucially, StableDepth maintains online inference capability, processing frames independently without requiring past or future frames, and operates at 5.77 FPS on a H20 GPU - \\(13 \\times\\) faster than DepthCrafter. Both methods were evaluated on \\(1920 \\times 1080\\) resolution images using half-precision (FP16) for inference speed testing. These results demonstrate that StableDepth successfully combines strong temporal consistency with practical advantages in" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.312, + 0.646, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.369, + 0.646, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.428, + 0.598, + 0.439 + ], + "angle": 0, + "content": "RGB" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.312, + 0.773, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.369, + 0.773, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.663, + 0.428, + 0.758, + 0.44 + ], + "angle": 0, + "content": "DepthCrafter [23]" + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.312, + 0.901, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.37, + 0.901, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.786, + 0.428, + 0.889, + 0.44 + ], + "angle": 0, + "content": "StableDepth (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.451, + 0.907, + 0.535 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparisons. StableDepth shows improved robustness in handling ambiguous regions such as water surfaces and sky boundaries while maintaining accurate depth estimation for the main subjects (flamingo, horse) compared to DepthCrafter. The highlighted region in DepthCrafter results shows inconsistent depth predictions in these challenging regions." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.549, + 0.902, + 0.656 + ], + "angle": 0, + "content": "
DataPretrainedBonnScanNet
AbsRel (↓)δ1(↑)AbsRel (↓)δ1(↑)
LDINOv20.0860.9290.0980.911
LDAv20.0710.9460.1020.903
L + UDINOv20.0680.9490.0990.909
L + UDAv20.0570.9760.0970.913
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.66, + 0.907, + 0.716 + ], + "angle": 0, + "content": "Table 6. Performance comparison on Bonn and ScanNet. We compare different initialization strategies: DINOv2 pre-trained weights vs. DAv2's well-trained parameters. L: labeled data only; L+U: both labeled and unlabeled data with dual-decoder." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.905, + 0.763 + ], + "angle": 0, + "content": "deployment, addressing a key limitation in current depth estimation approaches." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.84 + ], + "angle": 0, + "content": "Comparison of unlabeled video frames. As shown in Fig. 3, MTD scores improve steadily with more unlabeled video frames in each mini-batch and gradually converge at 16 frames, with negligible gains beyond this point, supporting our choice of 16-frame clips as an optimal trade-off." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Comparison of different pre-trained encoders. Tab. 6 shows the effectiveness of our semi-supervised training strategy on the Bonn and ScanNet datasets. We compare two initialization strategies: DINOv2 and DAv2 pre-trained" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7075" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.089, + 0.895, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.457, + 0.908, + 0.501 + ], + "angle": 0, + "content": "Figure 5. Qualitative visualizations of depth predictions across diverse scenarios. Each row pair shows an input sequence and corresponding depth predictions from our method. Our approach maintains consistent depth estimation while preserving fine geometric details across challenging cases including dynamic motions (dancer), reflective surfaces (water), and large objects (rhinoceros)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.526, + 0.484, + 0.648 + ], + "angle": 0, + "content": "weights. With only labeled data (L), DAv2 initialization outperforms DINOv2, particularly on Bonn (0.071 vs 0.086 AbsRel). Incorporating unlabeled video data \\((\\mathrm{L} + \\mathrm{U})\\) through our dual-decoder approach yields consistent improvements across both datasets. DAv2 with \\(\\mathrm{L} + \\mathrm{U}\\) achieves the best performance, reducing AbsRel to 0.057 on Bonn and 0.097 on ScanNet, demonstrating the effectiveness of our semi-supervised approach." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.665, + 0.274, + 0.68 + ], + "angle": 0, + "content": "4.4. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Our method demonstrates robust performance across diverse real-world scenarios. As shown in Fig. 1, it generates stable depth maps even in challenging conditions, capturing dynamic motions in dog agility courses, street performances, and natural scenes while maintaining consistent geometry frame-by-frame. Fig. 4 shows our method's superiority over DepthCrafter in ambiguous depth cue cases. In the flamingo scene, DepthCrafter produces inconsistent depth predictions for water reflections, while our method maintains coherent depth across the water surface. In the horse sequence, DepthCrafter mistakenly estimates the sky as foreground, whereas our approach handles the background-sky transition more accurately. These results highlight our model's robustness to common depth estima" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.526, + 0.907, + 0.679 + ], + "angle": 0, + "content": "tion challenges like reflective surfaces and infinite-depth regions, while preserving scene consistency. Fig. 5 further demonstrates the effectiveness of our method in challenging scenarios, including fast-moving subjects (dancer) and large objects (rinoceros). Our approach consistently preserves fine geometric details and maintains stable depth relationships across frames, even under complex lighting conditions and dynamic motions. The temporal smoothness achieved enables reliable depth estimation for practical applications requiring consistent 3D understanding." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.691, + 0.634, + 0.707 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.716, + 0.907, + 0.898 + ], + "angle": 0, + "content": "We present StableDepth, a framework that bridges the gap between relative and metric depth estimation through our scene-consistent and scale-invariant (SCSI) paradigm. By leveraging our dual-decoder architecture and video diffusion baking strategy, we effectively decouple depth accuracy from temporal consistency learning, while harnessing knowledge from unlabeled video data. StableDepth enables online inference at \\(13 \\times\\) faster speed than previous video-based methods while maintaining comparable temporal stability. This approach advances monocular depth estimation toward more practical applications in 3D reconstruction, robotics, and augmented reality." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7076" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.155 + ], + "angle": 0, + "content": "Acknowledgements. This work is supported by the National Natural Science Foundation of China (No. 62422606, 62201484) and DAMO Academy Research Intern Program." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.181, + 0.188, + 0.196 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.206, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[1] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv:2302.12288, 2023.5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.263, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[2] Reiner Birkl, Diana Wofk, and Matthias Muller. Midas v3. 1-a model zoo for robust monocular relative depth estimation. arXiv:2307.14460, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.306, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv:2311.15127, 2023. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.377, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[4] Aleksei Bochkovskii, Amaël Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. In ICLR, 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.433, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[5] D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black. A naturalistic open source movie for optical flow evaluation. In ECCV, 2012. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.476, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[6] Yohann Cabon, Naila Murray, and Martin Humenberger. Virtual katii 2. arXiv:2001.10773, 2020. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.505, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[7] Sili Chen, Hengkai Guo, Shengnan Zhu, Feihu Zhang, Zi long Huang, Jiashi Feng, and Bingyi Kang. Video depth anything: Consistent depth estimation for super-long videos. In CVPR, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.562, + 0.482, + 0.589 + ], + "angle": 0, + "content": "[8] Weifeng Chen, Zhao Fu, Dawei Yang, and Jia Deng. Single-image depth perception in the wild. In NeurIPS, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.591, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[9] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.647, + 0.482, + 0.728 + ], + "angle": 0, + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.732, + 0.482, + 0.772 + ], + "angle": 0, + "content": "[11] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[12] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. In NeurIPS, 2014. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[13] Fatima El Jamiy and Ronald Marsh. Survey on depth perception in head mounted displays: distance estimation in virtual reality, augmented reality, and mixed reality. IET Image Processing, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[14] Gunnar Farnebäck. Two-frame motion estimation based on polynomial expansion. In SCIA, 2003. 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.206, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[15] Xiao Fu, Wei Yin, Mu Hu, Kaixuan Wang, Yuexin Ma, Ping Tan, Shaojie Shen, Dahua Lin, and Xiaoxiao Long. Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[16] Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as proxy for multi-object tracking analysis. In CVPR, 2016. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[17] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. *IJRR*, 2013. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[18] Ming Gui, Johannes S Fischer, Ulrich Prestel, Pingchuan Ma, Dmytro Kotovenko, Olga Grebenkova, Stefan Andreas Baumann, Vincent Tao Hu, and Björn Ommer. Depthfm: Fast monocular depth estimation with flow matching. In AAAI, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[19] Vitor Guizilini, Igor Vasiljevic, Dian Chen, Rares Ambrus, and Adrien Gaidon. Towards zero-shot scale-aware monocular depth estimation. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.345, + 0.905, + 0.384 + ], + "angle": 0, + "content": "[20] S Mahdi H. Miangoleh, Mahesh Reddy, and Yaqiz Aksoy. Scale-invariant monocular depth estimation via ssi depth. In SIGGRAPH, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[21] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.428, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[22] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. TPAMI, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[23] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. In CVPR, 2025. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[24] Yaosi Hu, Chong Luo, and Zhenzhong Chen. Make it move: controllable image-to-video generation with text descriptions. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[25] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.651, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[26] Bingxin Ke, Dominik Narnhofer, Shengyu Huang, Lei Ke, Torben Peters, Katerina Fragkiadaki, Anton Obukhov, and Konrad Schindler. Video depth without video models. In CVPR, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[27] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[28] Yixuan Li, Lihan Jiang, Linning Xu, Yuanbo Xiangli, Zhenzhi Wang, Dahua Lin, and Bo Dai. Matrixcity: A large-scale city dataset for city-scale neural rendering and beyond. In ICCV, 2023. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[29] Zhaoshuo Li, Wei Ye, Dilin Wang, Francis X Creighton, Russell H Taylor, Ganesh Venkatesh, and Mathias Unberath. Temporally consistent online depth estimation in dynamic scenes. In WACV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[30] LightwheelAI and LightwheelOcc contributors. LightwheelOcc: A 3d occupancy synthetic dataset in autonomous driving. https://github.com/OpenDriveLab/LightwheelOcc, 2024.5,6,7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7077" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[31] Nian Liu, Ni Zhang, Ling Shao, and Junwei Han. Learning selective mutual attention and contrast for rgb-d saliency detection. TPAMI, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[32] Xingtong Liu, Ayushi Sinha, Masaru Ishii, Gregory D Hager, Austin Reiter, Russell H Taylor, and Mathias Unberath. Dense depth estimation in monocular endoscopy with self-supervised learning methods. TMI, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.192, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.222, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[34] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. In SIGGRAPH, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[35] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. TMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[36] E. Palazzolo, J. Behley, P. Lottes, P. Giguère, and C. Stachniss. ReFusion: 3D Reconstruction in Dynamic Environments for RGB-D Cameras Exploiting Residuals. In IROS, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[37] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Unidepth: Universal monocular metric depth estimation. In CVPR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[38] Luigi Piccinelli, Christos Sakaridis, Yung-Hsu Yang, Mattia Segu, Siyuan Li, Wim Abbeloos, and Luc Van Gool. Unidepthv2: Universal monocular metric depth estimation made simpler. arXiv:2502.20110, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[39] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. TPAMI, 2020. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.601 + ], + "angle": 0, + "content": "[40] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, 2021. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.671 + ], + "angle": 0, + "content": "[41] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. In ICLR, 2025. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[42] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In ICCV, 2021. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.482, + 0.828 + ], + "angle": 0, + "content": "[44] Abhinav Sagar. Monocular depth estimation using multi-scale neural network and feature fusion. In WACV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[45] Ashutosh Saxena, Min Sun, and Andrew Y Ng. Make3d: Learning 3d scene structure from a single still image. TPAMI, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[46] Mohamed Sayed, John Gibson, Jamie Watson, Victor Prisacariu, Michael Firman, and Clément Godard. Simplere" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "con: 3d reconstruction without 3d convolutions. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[47] Jiahao Shao, Yuanbo Yang, Hongyu Zhou, Youmin Zhang, Yujun Shen, Matteo Poggi, and Yiyi Liao. Learning temporally consistent video depth from video diffusion priors. 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[48] Chang Shu, Kun Yu, Zhixiang Duan, and Kuiyuan Yang. Feature-metric loss for self-supervised learning of depth and egomotion. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[49] Chaoyang Wang, Simon Lucey, Federico Perazzi, and Oliver Wang. Web stereo video supervision for depth prediction from dynamic scenes. In 3DV, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[50] Lijun Wang, Yifan Wang, Linzhao Wang, Yunlong Zhan, Ying Wang, and Huchuan Lu. Can scale-consistent monocular depth be learned in a self-supervised scale-invariant manner? In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[51] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.378, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[52] Yiran Wang, Min Shi, Jiaqi Li, Zihao Huang, Zhiguo Cao, Jianming Zhang, Ke Xian, and Guosheng Lin. Neural video depth stabilizer. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.421, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[53] Jamie Watson, Mohamed Sayed, Zawar Qureshi, Gabriel J Brostow, Sara Vicente, Oisin Mac Aodha, and Michael Firman. Virtual occlusions through implicit depth. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[54] Diana Wofk, Fangchang Ma, Tien-Ju Yang, Sertac Karaman, and Vivienne Sze. Fastdepth: Fast monocular depth estimation on embedded systems. In ICRA, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.52, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[55] Ke Xian, Jianming Zhang, Oliver Wang, Long Mai, Zhe Lin, and Zhiguo Cao. Structure-guided ranking loss for single image depth prediction. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.563, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[56] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[57] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. In NeurIPS, 2024. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.662, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[58] Rajeev Yasarla, Hong Cai, Jisoo Jeong, Yunxiao Shi, Risheek Garrepalli, and Fatih Porikli. Mamo: Leveraging memory and attention for monocular video depth estimation. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.719, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[59] Wei Yin, Chi Zhang, Hao Chen, Zhipeng Cai, Gang Yu, Kaixuan Wang, Xiaozhi Chen, and Chunhua Shen. Metric3d: Towards zero-shot metric 3d prediction from a single image. In ICCV, 2023. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[60] Sen Zhang, Jing Zhang, and Dacheng Tao. Towards scale-aware, robust, and generalizable unsupervised monocular depth estimation by integrating imu motion dynamics. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.886 + ], + "angle": 0, + "content": "[61] Ruijie Zhu, Chuxin Wang, Ziyang Song, Li Liu, Tianzhu Zhang, and Yongdong Zhang. Scaledepth: Decomposing metric depth estimation into scale prediction and relative depth estimation. arXiv:2407.08187, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.886 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7078" + } + ] +] \ No newline at end of file diff --git a/2025/StableDepth_ Scene-Consistent and Scale-Invariant Monocular Depth/73601462-88ee-4d47-a92f-ac7545019662_origin.pdf b/2025/StableDepth_ Scene-Consistent and Scale-Invariant Monocular Depth/73601462-88ee-4d47-a92f-ac7545019662_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..900113d27b4cd1642587a351294602356887d0da --- /dev/null +++ b/2025/StableDepth_ Scene-Consistent and Scale-Invariant Monocular Depth/73601462-88ee-4d47-a92f-ac7545019662_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8691788b5f996d12153bf07c817a24f2bcfbeff180520df6a1d7ac6b92c7dc38 +size 6622275 diff --git a/2025/StableDepth_ Scene-Consistent and Scale-Invariant Monocular Depth/full.md b/2025/StableDepth_ Scene-Consistent and Scale-Invariant Monocular Depth/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1b2f0edc892aa8ac98aa65c451ab507f57585767 --- /dev/null +++ b/2025/StableDepth_ Scene-Consistent and Scale-Invariant Monocular Depth/full.md @@ -0,0 +1,321 @@ +# StableDepth: Scene-Consistent and Scale-Invariant Monocular Depth + +Zheng Zhang $^{1,2,\star,*}$ Lihe Yang $^{1,\star}$ Tianyu Yang $^{2,4,\dagger}$ Chaohui Yu $^{2,4}$ Xiaoyang Guo $^{3}$ Yixing Lao $^{1}$ Hengshuang Zhao $^{1,\dagger}$ + +1The University of Hong Kong 2DAMO Academy, Alibaba Group + +3The Chinese University of Hong Kong 4Hupan Lab + +https://stabledepth.github.io + +![](images/ed2e4f41cfe3e1b6c05366775acc5ff9c0cacb43f52a9052db53c6ff90c1157b.jpg) + +![](images/4490ab9a6e715390594f56e20f7355d2a50fa3fbd9a8b2d7edbdf10e61e6fb89.jpg) + +![](images/2ae26fe8b485b7a3119433d3c2cc07c3041cbfa3aa7570aa45756eb54e360b6b.jpg) + +![](images/883a9c2ab0c65c01db85179a41d31bc8ae28ec877c628825def7489cca6442c9.jpg) +Figure 1. StableDepth achieves efficient online monocular depth estimation that produces scene-consistent and scale-invariant predictions frame by frame. As shown in the above challenging scenarios, StableDepth generates stable depth maps with precise geometric details and effectively handles dynamic motions, without the need of accessing future frames. + +# Abstract + +Recent advances in monocular depth estimation significantly improve robustness and accuracy. However, relative depth models exhibit flickering and 3D inconsistency in video data, limiting 3D reconstruction applications. We introduce StableDepth, a scene-consistent and scale-invariant depth estimation method achieving scene-level 3D consistency. Our dual-decoder architecture learns from + +large-scale unlabeled video data, enhancing generalization and reducing flickering. Unlike previous methods requiring full video sequences, StableDepth enables online inference at $13 \times$ faster speed, achieving significant improvements across benchmarks with comparable temporal consistency to video diffusion-based estimators. + +# 1. Introduction + +Monocular Depth Estimation (MDE) [12, 45] is pivotal in bridging the gap between 2D and 3D data representation, serving critical roles in a wide range of downstream applica + +tions, such as autonomous driving [51], robotics [54], medical imaging [32], and virtual reality [13]. Advancements in foundational models [35, 43] have greatly enhanced the accuracy and robustness of MDE [25, 56]. Despite these advances, the intrinsic ambiguity of monocular cues still poses substantial challenges, particularly for downstream applications that require 3D-consistent data and maintaining stability for uncontrolled in-the-wild data. + +MDE is primarily categorized into two main task settings: relative depth estimation and metric depth estimation. Relative depth estimation, as explored in studies [15, 18, 25, 39, 56], involves predicting depth values within a normalized depth range of $[0, 1]$ . This scale- and shift-invariant regression target facilitates robust joint training across diverse data domains. However, the predicted depth values differ from the actual depth due to scale and shift discrepancies, leading to inconsistency across video frames. This inconsistency can complicate advanced downstream tasks like 3D reconstruction. In contrast, metric depth estimation [4, 22, 37, 59] aims to predict physical distances in metric scale. However, these approaches face significant challenges. Scaled versions of a scene's depth values might all represent plausible depth estimations of the actual depth. Such ambiguity complicates the learning process of depth regression, hindering both generalization capabilities and precise 3D structures. Severe flickering and 3D inconsistency are also observed given video inputs. This challenge prompts us to reconsider what is an effective representation of MDE foundation models. + +The limitations of existing approaches are summarized in Tab. 1, which compares different depth estimation paradigms. While relative depth models [25, 39, 56] offer good generalization across datasets but suffer from per-frame inconsistency, metric depth models [4, 59] provide real-world measurements but struggle with generalization across diverse scenes. Scale-invariant models [20] improve upon relative models by eliminating shift ambiguity while maintaining scale invariance, but they still lack consistency across frames in the same scene. + +To address these challenges, we propose that the ideal task setting should be scene-consistent and scale-invariant (SCSI) depth estimation, as shown in Tab. 1. Our approach strikes a balance between per-frame scale-invariant methods that lack consistency and purely metric approaches that struggle with generalization. Specifically, our model primarily generates scale-invariant depth maps, but in the same scene or video, predictions across multiple frames demonstrate 3D consistency. Additionally, the scale factors for multi-frame data remain consistent compared with the actual depth values. In this work, the development of our model is primarily guided by the following objectives: 1) ensuring high-quality, scene-consistent, and scale-invariant depth estimations to achieve robust video consis + +
Depth TypeScaleShiftGeneralReconsConsist
RelativePer-framePer-frame✓✗✗
Scale-invariantPer-frameNone✓✓✗
MetricNoneNone✗✓✗
SCSI (Ours)Per-sceneNone✓✓✓
+ +Table 1. Comparison of different depth types. Our proposed SCSI depth combines the strengths of relative and metric depth while overcoming their weaknesses. General: generalization capability. Recons: reconstruction quality. Consist: scene consistency. + +tency; 2) enhancing strong generalization capabilities and precise depth details. + +Our framework introduces three key innovations to achieve scene-consistent depth estimation. First, we leverage temporal priors from pre-trained video diffusion models to generate pseudo-labels with inherent 3D consistency, replacing traditional unstable photometric constraints [3, 23, 47, 48]. Second, we design a dual-branch architecture where one branch processes labeled data for metric accuracy, while the other aligns unlabeled video frames to pseudo-labels through scene-level normalization, decoupling conflicting objectives. Third, our semi-supervised paradigm jointly optimizes metric supervision and temporal alignment, enabling the model to generalize across domains while eliminating flickering artifacts. This unified approach supports real-time inference without requiring future frames, overcoming the limitations of both single-frame and video-based methods. + +We evaluate our model on multiple datasets under zero-shot settings, aligning each scene using only a single scale (SCSI). As demonstrated in Figure 1, our method robustly handles challenging scenarios including dynamic motions and complex textures while preserving geometric fidelity frame-by-frame. Compared to DepthCrafter [23], we achieve $13 \times$ faster inference speed with online processing capability, improving prediction accuracy on Sintel [5], Bonn [36], ScanNet v2 [9], and KITTI [17] by $13.2\%$ , $86.8\%$ , $39.3\%$ , and $8.2\%$ , respectively. Our contributions are summarized as follows: + +- We introduce StableDepth, a novel monocular depth estimation framework that achieves scene-consistent and scale-invariant (SCSI) depth estimation, bridging the gap between relative and metric depth prediction while maintaining temporal consistency in video sequences. +- We propose a video diffusion baking strategy that leverages pre-trained video diffusion models for generating high-quality pseudo-labels, enabling effective semi-supervised learning from unlabeled video data while preserving temporal consistency. +- We develop a dual-decoder architecture that simultaneously handles supervised metric depth estimation and unsupervised depth alignment, effectively addressing the target misalignment issue while maintaining accuracy. + +- We demonstrate state-of-the-art performance across multiple benchmarks, achieving up to $86.8\%$ improvement in accuracy while providing $13 \times$ faster inference compared to previous methods. Our method enables efficient, online inference while maintaining comparable temporal stability to video diffusion models. + +# 2. Related Work + +Monocular depth estimation. MDE focuses on inferring depth from a single image. Early works [8, 31, 55] learn ordinal relationships from coarse annotations but cannot recover geometric structure. Recent methods [11, 39, 56, 57] leverage large-scale data to learn affine-invariant depth with improved generalization. However, unknown depth shifts cause geometric distortions, hindering 3D applications. To address this, recent works attempt zero-shot metric depth estimation [4, 19, 22, 37, 38, 59, 61] and explore self-supervised scale-consistent learning [50, 60], though maintaining consistent scale across scenes remains challenging. + +Towards consistent video depth. Video depth estimation requires both temporal consistency and per-frame precision. Traditional approaches use post-processing optimization [27, 34] or temporal modeling [29, 49], but incur computational costs or fail in dynamic scenes. Other methods leverage camera poses [46, 53] or memory-based approaches [52, 58], but often sacrifice geometric fidelity for smoothness. Recent works like Video Depth Anything [7] and video-free approaches [26] show promise for temporal consistency in long sequences. Video Depth Anything freezes the Depth Anything V2 encoder and incorporates cross-frame attention in the DPT [40] decoder for temporal consistency, though these methods still produce depth with inherent shift ambiguity. + +Diffusion priors for depth modeling. Video diffusion models [21, 24] demonstrate potential for consistent depth estimation. DepthCrafter [23] achieves temporal stability but generates only relative depths, while ChronoDepth [47] suffers from limited temporal contexts. Our proposed SCSI depth estimation leverages video diffusion baking and dual-decoder architecture for temporally consistent depth suitable for efficient online inference. + +# 3. StableDepth + +Our work leverages both labeled images and unlabeled video data to enhance SCSI depth predictions, achieving video consistency and enhanced generalization capabilities through the distillation of video diffusion priors. Formally, we denote the sets of labeled image-depth pairs and unlabeled videos as $\mathcal{D}^l = \{(x_i,d_i)\}_{i = 1}^M$ and $\mathcal{D}^u = \{v_i\}_{i = 1}^N$ , respectively. Our objective is to utilize a high-quality labeled image dataset, Hypersim [42], to help the model grasp real-world scale, while employing unlabeled video data to guide + +the model's depth predictions toward achieving video consistency and improved generalization. + +# 3.1. Mining Consistency from Unlabeled Videos + +StableDepth is designed to learn from both labeled images and unlabeled videos. For labeled image data, we predict the depth in metric space. The activation function of the network's last layer is sigmoid, after which its predictions are scaled to match the dataset values by multiplying with a predefined maximum depth. The SiLogLoss [12, 44, 57] is then used to minimize the error between the prediction and the ground truth: + +$$ +\mathcal {L} _ {\text {l a b e l}} = \sqrt {\frac {1}{H W} \sum_ {i = 1} ^ {H W} \left(\log \frac {d _ {i}}{d _ {i} ^ {*}}\right) ^ {2} - \lambda \left(\frac {1}{H W} \sum_ {i = 1} ^ {H W} \log \frac {d _ {i}}{d _ {i} ^ {*}}\right) ^ {2}}, \tag {1} +$$ + +where $d_i^*$ and $d_{i}$ are the prediction and ground truth, respectively, and $\lambda$ is a hyper-parameter. + +For unlabeled video data, we leverage video diffusion models, specifically DepthCrafter [23], to generate pseudolabels due to their inherent advantages in maintaining temporal consistency. These models, pre-trained on large-scale video data with cross-frame attention mechanisms, can generate stable depth predictions across entire video segments while preserving a consistent global scale and shift relationship with metric depth. Unlike conventional image-based MDE approaches that process frames independently, video diffusion models account for temporal dependencies, making them particularly effective for generating reliable pseudo-labels that exhibit strong video consistency - a crucial property for our scene-consistent training objective. For a video, we need a method to normalize the depth predicted by our model to the same space as the pseudo-labeled depth. Unlike the normalization used in single-frame relative depth training, this approach focuses on ensuring that an entire video has a unique scale and shift, rather than each frame having an independent scale and shift. This method aims to achieve video consistency in the depth predictions within the same scene. To supervise our model, we first use the following formula to normalize the pseudo-labels, + +$$ +\hat {D} (v _ {i}) = \frac {D (v _ {i}) - m (D (v _ {i}))}{s (D (v _ {i}))}, \tag {2} +$$ + +where $D(v_{i})$ is the pseudo-label for video $v_{i}$ , and $m(D(v_{i}))$ and $s(D(v_{i}))$ are used to remove the shift and scale for this video, respectively: + +$$ +m (D (v _ {i})) = \operatorname {m e d i a n} (D (v _ {i})), +$$ + +$$ +s \left(D \left(v _ {i}\right)\right) = \frac {1}{T H W} \sum_ {i = 1} ^ {T H W} | D \left(v _ {i}\right) - m \left(D \left(v _ {i}\right)\right) |. \tag {3} +$$ + +For the frame-by-frame depth predictions of video $v_{i}$ from our model, we concatenate them to obtain $P(v_{i})$ with + +![](images/347cbf4b90ce09213c1072109f744af1d03ab5cc6bb7b8787a1157f68f170541.jpg) +Figure 2. Overview of StableDepth. StableDepth employs a shared encoder with dual decoders to process both labeled images and unlabeled videos. The upper branch handles labeled images with direct metric depth supervision ("Sup"), while the lower branch processes unlabeled video sequences using temporal-coherent pseudo-labels generated by pre-trained video diffusion priors [23] ("Bake"). This architecture enables scene-consistent and scale-invariant (SCSI) depth prediction by combining the benefits of metric depth supervision with temporal consistency learning from video sequences. Solid arrows indicate the primary data flow, while dashed arrows represent the auxiliary training path for unlabeled data. + +dimensions $[T,H,W]$ . We then apply the same normalization process to the depth predictions $P(v_{i})$ to obtain its normalized version $\hat{P} (v_{i})$ . For $\hat{P} (v_{i})$ and $\hat{D} (v_{i})$ , we use the following alignment loss to minimize their difference, + +$$ +\mathcal {L} _ {\text {u n l a b e l}} = \frac {1}{T H W} \sum_ {i = 1} ^ {T H W} \left| \hat {P} \left(v _ {i}\right) - \hat {D} \left(v _ {i}\right) \right|. \tag {4} +$$ + +In one iteration, we simultaneously sample a certain proportion of labeled images and unlabeled videos for joint training, and use the following combined loss as the final loss for this iteration: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {l a b e l}} + \gamma \mathcal {L} _ {\text {u n l a b e l}}, \tag {5} +$$ + +where $\gamma$ gradually increases following a cosine curve as the number of iterations progresses. + +# 3.2. Avoiding Conflict in Depth Labels + +In the field of video consistency for depth prediction, most research [23] focuses on improving video consistency in near-field scenes. Therefore, we use depth in the disparity space for video data training. + +Following the previous network structure [2, 39, 56, 57], we use the ViT-L encoder [10], denoted as $\mathcal{E}$ , paired with the decoder from DPT [40], denoted as $\mathcal{D}_1$ . The decoder outputs depth in SCSI space with a range of (0, 1). A natural approach is to utilize this original decoder to predict the disparity-space depth of a video sequence, and then take the reciprocal of the predictions. Specifically, the depth prediction $P(v_{i})$ is given by: + +$$ +P \left(v _ {i}\right) = \frac {1}{\mathcal {D} _ {1} \left(\mathcal {E} \left(v _ {i}\right)\right)}. \tag {6} +$$ + +Finally, the depth predictions are normalized and supervised in semi-supervised settings following Eq. 3 and 4. + +Unfortunately, in our experiments, we found that although this modification improves the video consistency of depth predictions for videos, it reduces the accuracy of depth prediction by the model, resulting in lower metrics such as $\delta_{1}$ . We speculate that this is mainly due to the used pseudo-labeling model [23], which, despite having good video consistency, still lags behind Depth Anything V2 [57] in terms of depth prediction accuracy, particularly in the prediction of distant objects. To address this issue, we propose to use an auxiliary network to decouple the prediction and supervision of labeled images and unlabeled videos. To this end, we experimented with two approaches: + +In the first approach, both labeled images and unlabeled videos share the encoder $\mathcal{E}$ and the decoder $\mathcal{D}_1$ . A lightweight CNN network, denoted as $\mathcal{C}$ , is used to directly process the prediction results of the original decoder $\mathcal{D}_1$ with the CNN network outputting depth in disparity space: + +$$ +P \left(v _ {i}\right) = \mathcal {C} \left(\mathcal {D} _ {1} \left(\mathcal {E} \left(v _ {i}\right)\right)\right). \tag {7} +$$ + +In the second approach, labeled images and unlabeled videos share the encoder $\mathcal{E}$ . An additional DPT [40] decoder, denoted as $\mathcal{D}_2$ , is used to process the features extracted by the encoder $\mathcal{E}$ for unlabeled videos, as illustrated by Fig. 2. $\mathcal{D}_2$ directly produces depth in disparity space, while $\mathcal{D}_1$ continues to output depth in SCSI space. + +$$ +P \left(x _ {i}\right) = \mathcal {D} _ {1} \left(\mathcal {E} \left(x _ {i}\right)\right), \quad P \left(v _ {i}\right) = \mathcal {D} _ {2} \left(\mathcal {E} \left(v _ {i}\right)\right). \tag {8} +$$ + +The final results indicate that sharing the encoder between labeled images and unlabeled videos, while using two independent DPT decoders, achieves the best accuracy and video consistency. + +# 3.3. Scene-Consistent Depth Evaluation + +Scale alignment. For zero-shot evaluation, we enforce single-scale consistency per scene through RANSAC-based alignment with zero shift constraint. Each scene's + +predictions are scaled by a unified factor determined through 1,000 sampling iterations, retaining inliers with $\max \left(\frac{d_{pred}}{d_{gt}},\frac{d_{gt}}{d_{pred}}\right) < \tau$ , where $\tau = 1.25$ . The optimal scale is refined via least-squares optimization over all inliers, preserving geometric fidelity while eliminating cross-frame scale ambiguity. This per-scene single-scale paradigm ensures coherent depth relationships across dynamic sequences without per-frame adjustments. + +Motion-aware temporal metric. We propose the Motion-aware Temporal Difference (MTD) to quantify depth consistency in dynamic regions: + +$$ +\mathrm {M T D} = \frac {1}{T - 1} \sum_ {t = 1} ^ {T - 1} \frac {\sum_ {x , y} | \Delta D _ {t} | \cdot W _ {t}}{H \times W}, \tag {9} +$$ + +where $\Delta D_{t} = D_{t} - D_{t - 1}$ and $W_{t}(x,y) = (1 + \| \mathbf{f}_{t}\|_{2})^{-1}$ weights the depth variations by optical flow magnitude $\mathbf{f}_t$ (computed via Farneback [14] algorithm). Lower MTD indicates better temporal stability, penalizing flickering in high-motion areas. + +# 3.4. Discussion + +In this section, we address several important aspects of our approach to provide further clarity on methodology, design choices, and relationships to existing work. + +Decouple accuracy and consistency via dual learning. Our dual-decoder architecture decouples depth accuracy from temporal consistency learning, enabling joint optimization from diverse data sources. The primary decoder specializes in per-pixel depth precision through supervised training on high-quality labeled images, while the auxiliary decoder learns scene-level coherence from unlabeled videos via pseudo-labels. This separation resolves the inherent conflict where video-based methods like DepthCrafter [23] sacrifice depth accuracy for temporal smoothness, and single-frame approaches fail to ensure cross-frame consistency. Crucially, our design preserves absolute depth scales critical for 3D reconstruction, unlike relative-depth representations in video diffusion models. + +Beyond video diffusion priors. We repurpose video diffusion models as sources for scene consistency rather than as direct depth estimators. While models like DepthCrafter [23] excel at generating temporally consistent sequences, they produce relative depth representations with limited accuracy, particularly for distant objects, making them unsuitable for downstream tasks requiring precise geometry. Our approach extracts the temporal consistency knowledge from these models through scene-level normalization while maintaining superior depth accuracy. This strategy enables online inference $13 \times$ faster than video-based methods while achieving both better depth precision and comparable temporal stability. + +# 4. Experiment + +# 4.1. Implementation Details + +Network parameters and training details. We use the network architecture from Depth Anything V2 (DAv2) [57], consisting of a pre-trained ViT-L [10] encoder for feature extraction and a DPT [40] decoder for depth regression. The ViT-L encoder is initialized with parameters from the DAv2 model, while the primary decoder is randomly initialized with Sigmoid activation for metric depth output. The auxiliary decoder, also initialized randomly, uses ReLU activation for disparity (inverse depth) prediction. Each batch consists of a 1:2 ratio of labeled images to unlabeled video frames, totaling 24 frames. The learning rate for the encoder is $2.5 \times 10^{-6}$ , while the decoder's learning rate is 10 times larger. We use the AdamW optimizer [33] with a weight decay of 0.01 and employ a polynomial decay schedule for the learning rate. For data augmentation, we apply horizontal flipping to labeled images. In Eq. (5), the maximum value of $\gamma$ is set to 1. In Eq. (1), $\lambda$ is set to 0.5. + +Training dataset and pseudo-labeling. To train our model, we utilize DepthCrafter [23], a model finetuned on SVD [3], to generate video-consistent pseudo-labels for unlabeled video data. For metric depth supervision, we use Hypersim [42], a high-quality synthetic dataset with $\sim 60\mathrm{K}$ frames, as training data for decoder $\mathcal{D}_1$ . We select 6,000 videos from SA-V [41], which features dynamic foreground objects, yielding approximately $\sim 200\mathrm{K}$ frames for pseudolabeling training. To enhance the model's performance on distant scenes where DepthCrafter shows limitations, we incorporate additional datasets with accurate depth labels including VKITTI2 [6, 16], LightwheelOcc [30], and MatrixCity [28], treating frames from the same scene as video sequences during training. + +# 4.2. Evaluation + +Evaluation datasets. We evaluate on four datasets spanning diverse scenarios: Sintel [5] (23 synthetic videos, 50 frames each with precise depth labels), Bonn [36] (26 dynamic videos using frames 30-140 with foreground motions), ScanNet v2 [9] (100 indoor test videos sampled every third frame), and the full KITTI validation set [17] (13 outdoor driving videos with first 110 frames per sequence). This comprehensive benchmark covers synthetic animations, dynamic interactions, static indoor scenes, and real-world autonomous driving environments. + +Quantitative results. We evaluate our method against strong baselines: DepthCrafter [23], known for video consistency, Metric3D [59], a state-of-the-art metric depth estimator, and ZoeDepth [1], a robust monocular depth estimation method. For fair comparison, we align all predictions within each video using a single shared scale factor across the entire scene with shift set to zero. As shown in + +
Frame IntervalMethodSintel (~50 frames)Bonn (110 frames)ScanNet (90 frames)KITTI (110 frames)
AbsRel (↓)δ1 (↑)AbsRel (↓)δ1 (↑)AbsRel (↓)δ1 (↑)AbsRel (↓)δ1 (↑)
1DepthCrafter [23]2.9590.5450.3110.5240.3700.6610.1230.832
Zoedepth [1]0.4540.3560.2770.5620.3470.4360.2160.633
Metric3D [59]0.3720.6100.0550.9810.0590.9740.0670.977
StableDepth (Ours)0.2800.6170.0630.9790.0940.9210.1120.900
4DepthCrafter [23]6.3530.3930.5150.4170.5390.4400.2440.670
Zoedepth [1]0.4610.3570.2770.5630.3490.4360.2180.629
Metric3D [59]0.3610.6170.0550.9810.0590.9740.0670.977
StableDepth (Ours)0.2920.6060.0630.9790.0930.9210.1120.901
8DepthCrafter [23]2.0480.3980.4950.4120.5710.4230.2410.656
Zoedepth [1]0.4700.3570.2760.5620.3520.4340.2180.633
Metric3D [59]0.3630.6130.0550.9800.0600.9730.0660.977
StableDepth (Ours)0.3000.6040.0620.9800.0940.9200.1120.902
+ +Table 2. Comparison of depth estimation methods across different frame intervals. StableDepth maintains consistent performance across varying frame intervals while achieving superior accuracy. Best results are shown in bold, second best are underlined. Metric3D results on ScanNet are shown in gray as it was trained on this dataset. + +
MethodBonnScanNet
AbsRel (↓)δ1(↑)AbsRel (↓)δ1(↑)
Baseline0.0710.9460.1020.903
DPT0.1090.9140.1260.861
DPT + CNN0.1080.9260.1380.829
Dual DPT0.0570.9760.0970.913
+ +Table 3. Ablation on decoders. The baseline model is trained only on labeled image data. We compare three approaches for handling unlabeled video data: using a single DPT decoder with reciprocal output (DPT), adding a lightweight CNN head (DPT + CNN), and using dual DPT decoders (Dual DPT). + +Tab. 2, our method achieves superior performance across most datasets and metrics, with particularly strong improvements in AbsRel and $\delta_{1}$ metrics compared to video-based approaches. We examine the impact of frame intervals on prediction accuracy to simulate varying temporal densities in real-world scenarios. Note that compared to our ablation studies, these results reflect extended training with additional iterations for optimal performance. + +# 4.3. Ablation Study + +Comparison of different decoders. In Tab. 3, we compare four methods using the Bonn and ScanNet v2 datasets: training with only labeled Hypersim data, and training with both labeled Hypersim [42] and unlabeled SA-V [41] data processed by three decoder designs. The first row (Baseline) shows results with only labeled Hypersim data, while the following rows show results with both labeled and unlabeled data. For labeled Hypersim data, depth in SCSI space is obtained directly from the DPT head output. For the unlabeled SA-V data, the method for obtaining disparity space depth varies: in the second row (DPT), it's the reciprocal of the original DPT output; in the third row (DPT + ++ CNN), a lightweight CNN processes the DPT output; in the fourth row (Dual DPT), an additional DPT head provides the disparity space depth. The results show that using a single DPT head, either by reciprocating its output or adding a lightweight CNN, reduces prediction accuracy due to the partial inaccuracy of the pseudo-labels. However, using two independent DPT heads improves frame-to-frame consistency without impacting accuracy, validating the effectiveness of our disentangled approach. + +Encoder feature stability. We analyze encoder feature stability across consecutive DAVIS frames to understand our dual-decoder effectiveness. Compared to DAv2, our method achieves $7.1\%$ lower mean differences, $13.5\%$ reduced standard deviation, and $2.3\%$ higher cosine similarity. These improvements confirm that auxiliary decoder supervision enhances encoder stability, contributing to temporal consistency in depth predictions. + +Effectiveness of unlabeled video data. As shown in Tab. 4, we evaluate incremental training data impact. Using only Hypersim [42] (row 1), then adding pseudo-labeled SA-V [41] data (row 2) shows mixed results - improving Bonn and ScanNet v2 but decreasing Sintel and KITTI performance. This reveals DepthCrafter's video consistency benefits but limitations in distant regions. Progressively incorporating VKITTI2 [6, 16], LightWheel [30], Hypersim [42], and MatrixCity [28] (rows 3-6) successfully addresses these limitations, achieving best performance across all datasets. + +Comparison of video consistency. Tab. 5 showcases the performance comparison of depth estimation methods on the DAVIS dataset evaluated using Motion-aware temporal Difference (MTD (Eq. (9)), lower is better). The video diffusion-based DepthCrafter [23] achieves strong consistency $(\mathrm{MTD} = 0.003479)$ but requires processing the entire video sequence and runs at only 0.45 FPS. While + +
Training DataSintel (~50 frames)Bonn (110 frames)ScanNet (90 frames)KITTI (110 frames)
AbsRel (↓)δ1 (↑)AbsRel (↓)δ1 (↑)AbsRel (↓)δ1 (↑)AbsRel (↓)δ1 (↑)
Hypersim only [42]0.3090.5780.0710.9460.1020.9030.1220.871
+ SA-V [41]0.3260.5610.0570.9760.0970.9130.1430.823
+ VKITTI2 [6, 16]0.3180.5770.0580.9760.0970.9110.1290.839
+ LightWheel [30]0.3120.5730.0600.9760.0940.9160.1340.847
+ Hypersim-U [42]0.3140.5720.0640.9760.0960.9180.1230.870
+ Matrixcity [28]0.3150.5800.0620.9790.0920.9240.1140.895
+ +![](images/517d05a13a0817eb8f234dc7db953a936edae9d9fd565f6c8693aa1dbec0d07e.jpg) +Figure 3. Ablation on the number of unlabeled video frames in each batch. From 0 to 16, more frames bring better scene-level consistency. (Out of memory for ViT-Large under 32 frames) + +Table 4. Comparison of depth estimation performance with different training data configurations. The baseline model is trained on labeled Hypersim data only, with subsequent rows showing the cumulative effect of adding different unlabeled datasets (VKITTI2: Virtual KITTI 2, Hypersim-U: treat Hypersim as unlabeled video data). Lower AbsRel $(\downarrow)$ and higher $\delta 1$ $(\uparrow)$ values indicate better performance. For each video segment, we align using only a single scale, setting the shift to zero. + +
MethodMTD (↓)Video Model?Online?FPS (↑)
DepthCrafter [23]0.003479✓×0.45
Baseline0.003777×✓5.77
StableDepth (Ours)0.003482×✓5.77
+ +our baseline (metric depth version of DAv2-Large) enables online inference at 5.77 FPS, it shows relatively higher inconsistency (MTD=0.003777). StableDepth significantly improves upon our baseline with a $7.8\%$ reduction in MTD (0.003482), achieving comparable consistency to DepthCrafter. Crucially, StableDepth maintains online inference capability, processing frames independently without requiring past or future frames, and operates at 5.77 FPS on a H20 GPU - $13 \times$ faster than DepthCrafter. Both methods were evaluated on $1920 \times 1080$ resolution images using half-precision (FP16) for inference speed testing. These results demonstrate that StableDepth successfully combines strong temporal consistency with practical advantages in + +![](images/277ed3ad43b53cfa9eafb921540613bc0b19ab0efacb43d181812e8b98b914ff.jpg) + +![](images/6fbe2d51ca205d8f78e9fda27356074c98e6068528a039780413474603878c8d.jpg) +RGB + +![](images/b0b71a1817c0e70862ef88b7e68b82c3c904b8a719798575cd20183c448a8e58.jpg) + +![](images/8b42e1b69ca34ea54d9dec765e23f973b34f040f5bca98ac35f7c497922f82d2.jpg) +DepthCrafter [23] + +![](images/94b3ea5978e55dc36625501f657bb921d531e3db558defe960306e67a32f359e.jpg) + +![](images/14f8d1b0091cd4f616700935be367bdb5481ed2c9f9cdfa82d77f0e10049c710.jpg) +StableDepth (Ours) +Figure 4. Qualitative comparisons. StableDepth shows improved robustness in handling ambiguous regions such as water surfaces and sky boundaries while maintaining accurate depth estimation for the main subjects (flamingo, horse) compared to DepthCrafter. The highlighted region in DepthCrafter results shows inconsistent depth predictions in these challenging regions. + +Table 5. Performance comparison of different depth estimation methods on DAVIS dataset. MTD: Motion-aware Temporal Difference (Eq. (9)). Online: whether the method can process data without requiring the entire video. FPS: inference speed in frames per second. StableDepth achieves comparable consistency to DepthCrafter that based on heavy SVD [3], while maintaining online capability and significantly faster inference. + +
DataPretrainedBonnScanNet
AbsRel (↓)δ1(↑)AbsRel (↓)δ1(↑)
LDINOv20.0860.9290.0980.911
LDAv20.0710.9460.1020.903
L + UDINOv20.0680.9490.0990.909
L + UDAv20.0570.9760.0970.913
+ +Table 6. Performance comparison on Bonn and ScanNet. We compare different initialization strategies: DINOv2 pre-trained weights vs. DAv2's well-trained parameters. L: labeled data only; L+U: both labeled and unlabeled data with dual-decoder. + +deployment, addressing a key limitation in current depth estimation approaches. + +Comparison of unlabeled video frames. As shown in Fig. 3, MTD scores improve steadily with more unlabeled video frames in each mini-batch and gradually converge at 16 frames, with negligible gains beyond this point, supporting our choice of 16-frame clips as an optimal trade-off. + +Comparison of different pre-trained encoders. Tab. 6 shows the effectiveness of our semi-supervised training strategy on the Bonn and ScanNet datasets. We compare two initialization strategies: DINOv2 and DAv2 pre-trained + +![](images/bb0ac1a715090252d9205b98f030c4d49333ce8d6b00bda453f229ff5b994036.jpg) +Figure 5. Qualitative visualizations of depth predictions across diverse scenarios. Each row pair shows an input sequence and corresponding depth predictions from our method. Our approach maintains consistent depth estimation while preserving fine geometric details across challenging cases including dynamic motions (dancer), reflective surfaces (water), and large objects (rhinoceros). + +weights. With only labeled data (L), DAv2 initialization outperforms DINOv2, particularly on Bonn (0.071 vs 0.086 AbsRel). Incorporating unlabeled video data $(\mathrm{L} + \mathrm{U})$ through our dual-decoder approach yields consistent improvements across both datasets. DAv2 with $\mathrm{L} + \mathrm{U}$ achieves the best performance, reducing AbsRel to 0.057 on Bonn and 0.097 on ScanNet, demonstrating the effectiveness of our semi-supervised approach. + +# 4.4. Qualitative Results + +Our method demonstrates robust performance across diverse real-world scenarios. As shown in Fig. 1, it generates stable depth maps even in challenging conditions, capturing dynamic motions in dog agility courses, street performances, and natural scenes while maintaining consistent geometry frame-by-frame. Fig. 4 shows our method's superiority over DepthCrafter in ambiguous depth cue cases. In the flamingo scene, DepthCrafter produces inconsistent depth predictions for water reflections, while our method maintains coherent depth across the water surface. In the horse sequence, DepthCrafter mistakenly estimates the sky as foreground, whereas our approach handles the background-sky transition more accurately. These results highlight our model's robustness to common depth estima + +tion challenges like reflective surfaces and infinite-depth regions, while preserving scene consistency. Fig. 5 further demonstrates the effectiveness of our method in challenging scenarios, including fast-moving subjects (dancer) and large objects (rinoceros). Our approach consistently preserves fine geometric details and maintains stable depth relationships across frames, even under complex lighting conditions and dynamic motions. The temporal smoothness achieved enables reliable depth estimation for practical applications requiring consistent 3D understanding. + +# 5. Conclusion + +We present StableDepth, a framework that bridges the gap between relative and metric depth estimation through our scene-consistent and scale-invariant (SCSI) paradigm. By leveraging our dual-decoder architecture and video diffusion baking strategy, we effectively decouple depth accuracy from temporal consistency learning, while harnessing knowledge from unlabeled video data. StableDepth enables online inference at $13 \times$ faster speed than previous video-based methods while maintaining comparable temporal stability. This approach advances monocular depth estimation toward more practical applications in 3D reconstruction, robotics, and augmented reality. + +Acknowledgements. This work is supported by the National Natural Science Foundation of China (No. 62422606, 62201484) and DAMO Academy Research Intern Program. + +# References + +[1] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv:2302.12288, 2023.5, 6 +[2] Reiner Birkl, Diana Wofk, and Matthias Muller. Midas v3. 1-a model zoo for robust monocular relative depth estimation. arXiv:2307.14460, 2023. 4 +[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv:2311.15127, 2023. 2, 5, 7 +[4] Aleksei Bochkovskii, Amaël Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. In ICLR, 2025. 2, 3 +[5] D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black. A naturalistic open source movie for optical flow evaluation. In ECCV, 2012. 2, 5 +[6] Yohann Cabon, Naila Murray, and Martin Humenberger. Virtual katii 2. arXiv:2001.10773, 2020. 5, 6, 7 +[7] Sili Chen, Hengkai Guo, Shengnan Zhu, Feihu Zhang, Zi long Huang, Jiashi Feng, and Bingyi Kang. Video depth anything: Consistent depth estimation for super-long videos. In CVPR, 2025. 3 +[8] Weifeng Chen, Zhao Fu, Dawei Yang, and Jia Deng. Single-image depth perception in the wild. In NeurIPS, 2016. 3 +[9] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 2, 5 +[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 4, 5 +[11] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In ICCV, 2021. 3 +[12] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. In NeurIPS, 2014. 1, 3 +[13] Fatima El Jamiy and Ronald Marsh. Survey on depth perception in head mounted displays: distance estimation in virtual reality, augmented reality, and mixed reality. IET Image Processing, 2019. 2 +[14] Gunnar Farnebäck. Two-frame motion estimation based on polynomial expansion. In SCIA, 2003. 5 + +[15] Xiao Fu, Wei Yin, Mu Hu, Kaixuan Wang, Yuexin Ma, Ping Tan, Shaojie Shen, Dahua Lin, and Xiaoxiao Long. Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image. In ECCV, 2024. 2 +[16] Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as proxy for multi-object tracking analysis. In CVPR, 2016. 5, 6, 7 +[17] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. *IJRR*, 2013. 2, 5 +[18] Ming Gui, Johannes S Fischer, Ulrich Prestel, Pingchuan Ma, Dmytro Kotovenko, Olga Grebenkova, Stefan Andreas Baumann, Vincent Tao Hu, and Björn Ommer. Depthfm: Fast monocular depth estimation with flow matching. In AAAI, 2025. 2 +[19] Vitor Guizilini, Igor Vasiljevic, Dian Chen, Rares Ambrus, and Adrien Gaidon. Towards zero-shot scale-aware monocular depth estimation. In ICCV, 2023. 3 +[20] S Mahdi H. Miangoleh, Mahesh Reddy, and Yaqiz Aksoy. Scale-invariant monocular depth estimation via ssi depth. In SIGGRAPH, 2024. 2 +[21] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In NeurIPS, 2022. 3 +[22] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. TPAMI, 2024. 2, 3 +[23] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. In CVPR, 2025. 2, 3, 4, 5, 6, 7 +[24] Yaosi Hu, Chong Luo, and Zhenzhong Chen. Make it move: controllable image-to-video generation with text descriptions. In CVPR, 2022. 3 +[25] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In CVPR, 2024. 2 +[26] Bingxin Ke, Dominik Narnhofer, Shengyu Huang, Lei Ke, Torben Peters, Katerina Fragkiadaki, Anton Obukhov, and Konrad Schindler. Video depth without video models. In CVPR, 2025. 3 +[27] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In CVPR, 2021. 3 +[28] Yixuan Li, Lihan Jiang, Linning Xu, Yuanbo Xiangli, Zhenzhi Wang, Dahua Lin, and Bo Dai. Matrixcity: A large-scale city dataset for city-scale neural rendering and beyond. In ICCV, 2023. 5, 6, 7 +[29] Zhaoshuo Li, Wei Ye, Dilin Wang, Francis X Creighton, Russell H Taylor, Ganesh Venkatesh, and Mathias Unberath. Temporally consistent online depth estimation in dynamic scenes. In WACV, 2023. 3 +[30] LightwheelAI and LightwheelOcc contributors. LightwheelOcc: A 3d occupancy synthetic dataset in autonomous driving. https://github.com/OpenDriveLab/LightwheelOcc, 2024.5,6,7 + +[31] Nian Liu, Ni Zhang, Ling Shao, and Junwei Han. Learning selective mutual attention and contrast for rgb-d saliency detection. TPAMI, 2021. 3 +[32] Xingtong Liu, Ayushi Sinha, Masaru Ishii, Gregory D Hager, Austin Reiter, Russell H Taylor, and Mathias Unberath. Dense depth estimation in monocular endoscopy with self-supervised learning methods. TMI, 2019. 2 +[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 5 +[34] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. In SIGGRAPH, 2020. 3 +[35] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. TMLR, 2023. 2 +[36] E. Palazzolo, J. Behley, P. Lottes, P. Giguère, and C. Stachniss. ReFusion: 3D Reconstruction in Dynamic Environments for RGB-D Cameras Exploiting Residuals. In IROS, 2019. 2, 5 +[37] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Unidepth: Universal monocular metric depth estimation. In CVPR, 2024. 2, 3 +[38] Luigi Piccinelli, Christos Sakaridis, Yung-Hsu Yang, Mattia Segu, Siyuan Li, Wim Abbeloos, and Luc Van Gool. Unidepthv2: Universal monocular metric depth estimation made simpler. arXiv:2502.20110, 2025. 3 +[39] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. TPAMI, 2020. 2, 3, 4 +[40] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, 2021. 3, 4, 5 +[41] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. In ICLR, 2025. 5, 6, 7 +[42] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In ICCV, 2021. 3, 5, 6, 7 +[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2 +[44] Abhinav Sagar. Monocular depth estimation using multi-scale neural network and feature fusion. In WACV, 2022. 3 +[45] Ashutosh Saxena, Min Sun, and Andrew Y Ng. Make3d: Learning 3d scene structure from a single still image. TPAMI, 2008. 1 +[46] Mohamed Sayed, John Gibson, Jamie Watson, Victor Prisacariu, Michael Firman, and Clément Godard. Simplere + +con: 3d reconstruction without 3d convolutions. In ECCV, 2022. 3 +[47] Jiahao Shao, Yuanbo Yang, Hongyu Zhou, Youmin Zhang, Yujun Shen, Matteo Poggi, and Yiyi Liao. Learning temporally consistent video depth from video diffusion priors. 2025. 2, 3 +[48] Chang Shu, Kun Yu, Zhixiang Duan, and Kuiyuan Yang. Feature-metric loss for self-supervised learning of depth and egomotion. In ECCV, 2020. 2 +[49] Chaoyang Wang, Simon Lucey, Federico Perazzi, and Oliver Wang. Web stereo video supervision for depth prediction from dynamic scenes. In 3DV, 2019. 3 +[50] Lijun Wang, Yifan Wang, Linzhao Wang, Yunlong Zhan, Ying Wang, and Huchuan Lu. Can scale-consistent monocular depth be learned in a self-supervised scale-invariant manner? In ICCV, 2021. 3 +[51] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In CVPR, 2019. 2 +[52] Yiran Wang, Min Shi, Jiaqi Li, Zihao Huang, Zhiguo Cao, Jianming Zhang, Ke Xian, and Guosheng Lin. Neural video depth stabilizer. In ICCV, 2023. 3 +[53] Jamie Watson, Mohamed Sayed, Zawar Qureshi, Gabriel J Brostow, Sara Vicente, Oisin Mac Aodha, and Michael Firman. Virtual occlusions through implicit depth. In CVPR, 2023. 3 +[54] Diana Wofk, Fangchang Ma, Tien-Ju Yang, Sertac Karaman, and Vivienne Sze. Fastdepth: Fast monocular depth estimation on embedded systems. In ICRA, 2019. 2 +[55] Ke Xian, Jianming Zhang, Oliver Wang, Long Mai, Zhe Lin, and Zhiguo Cao. Structure-guided ranking loss for single image depth prediction. In CVPR, 2020. 3 +[56] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 2, 3, 4 +[57] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. In NeurIPS, 2024. 3, 4, 5 +[58] Rajeev Yasarla, Hong Cai, Jisoo Jeong, Yunxiao Shi, Risheek Garrepalli, and Fatih Porikli. Mamo: Leveraging memory and attention for monocular video depth estimation. In ICCV, 2023. 3 +[59] Wei Yin, Chi Zhang, Hao Chen, Zhipeng Cai, Gang Yu, Kaixuan Wang, Xiaozhi Chen, and Chunhua Shen. Metric3d: Towards zero-shot metric 3d prediction from a single image. In ICCV, 2023. 2, 3, 5, 6 +[60] Sen Zhang, Jing Zhang, and Dacheng Tao. Towards scale-aware, robust, and generalizable unsupervised monocular depth estimation by integrating imu motion dynamics. In ECCV, 2022. 3 +[61] Ruijie Zhu, Chuxin Wang, Ziyang Song, Li Liu, Tianzhu Zhang, and Yongdong Zhang. Scaledepth: Decomposing metric depth estimation into scale prediction and relative depth estimation. arXiv:2407.08187, 2024. 3 \ No newline at end of file diff --git a/2025/StableDepth_ Scene-Consistent and Scale-Invariant Monocular Depth/images.zip b/2025/StableDepth_ Scene-Consistent and Scale-Invariant Monocular Depth/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..77359706c92c3e905642644ed4a6614508a48da9 --- /dev/null +++ b/2025/StableDepth_ Scene-Consistent and Scale-Invariant Monocular Depth/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95182edf4d9d5180d4187b13024e382ca4bc776ce215b7f325002a435aa04c45 +size 746116 diff --git a/2025/StableDepth_ Scene-Consistent and Scale-Invariant Monocular Depth/layout.json b/2025/StableDepth_ Scene-Consistent and Scale-Invariant Monocular Depth/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..98db9b16bf796cc7d93b00f38f713b595c30885c --- /dev/null +++ b/2025/StableDepth_ Scene-Consistent and Scale-Invariant Monocular Depth/layout.json @@ -0,0 +1,8394 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 90, + 103, + 520, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 103, + 520, + 121 + ], + "spans": [ + { + "bbox": [ + 90, + 103, + 520, + 121 + ], + "type": "text", + "content": "StableDepth: Scene-Consistent and Scale-Invariant Monocular Depth" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 475, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 475, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 475, + 172 + ], + "type": "text", + "content": "Zheng Zhang" + }, + { + "bbox": [ + 130, + 141, + 475, + 172 + ], + "type": "inline_equation", + "content": "^{1,2,\\star,*}" + }, + { + "bbox": [ + 130, + 141, + 475, + 172 + ], + "type": "text", + "content": " Lihe Yang" + }, + { + "bbox": [ + 130, + 141, + 475, + 172 + ], + "type": "inline_equation", + "content": "^{1,\\star}" + }, + { + "bbox": [ + 130, + 141, + 475, + 172 + ], + "type": "text", + "content": " Tianyu Yang" + }, + { + "bbox": [ + 130, + 141, + 475, + 172 + ], + "type": "inline_equation", + "content": "^{2,4,\\dagger}" + }, + { + "bbox": [ + 130, + 141, + 475, + 172 + ], + "type": "text", + "content": " Chaohui Yu" + }, + { + "bbox": [ + 130, + 141, + 475, + 172 + ], + "type": "inline_equation", + "content": "^{2,4}" + }, + { + "bbox": [ + 130, + 141, + 475, + 172 + ], + "type": "text", + "content": " Xiaoyang Guo" + }, + { + "bbox": [ + 130, + 141, + 475, + 172 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 130, + 141, + 475, + 172 + ], + "type": "text", + "content": " Yixing Lao" + }, + { + "bbox": [ + 130, + 141, + 475, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 130, + 141, + 475, + 172 + ], + "type": "text", + "content": " Hengshuang Zhao" + }, + { + "bbox": [ + 130, + 141, + 475, + 172 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 173, + 471, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 173, + 471, + 188 + ], + "spans": [ + { + "bbox": [ + 140, + 173, + 471, + 188 + ], + "type": "text", + "content": "1The University of Hong Kong 2DAMO Academy, Alibaba Group" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 173, + 188, + 437, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 188, + 437, + 202 + ], + "spans": [ + { + "bbox": [ + 173, + 188, + 437, + 202 + ], + "type": "text", + "content": "3The Chinese University of Hong Kong 4Hupan Lab" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 252, + 204, + 356, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 204, + 356, + 215 + ], + "spans": [ + { + "bbox": [ + 252, + 204, + 356, + 215 + ], + "type": "text", + "content": "https://stabledepth.github.io" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 62, + 243, + 547, + 312 + ], + "blocks": [ + { + "bbox": [ + 62, + 243, + 547, + 312 + ], + "lines": [ + { + "bbox": [ + 62, + 243, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 243, + 547, + 312 + ], + "type": "image", + "image_path": "ed2e4f41cfe3e1b6c05366775acc5ff9c0cacb43f52a9052db53c6ff90c1157b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 62, + 313, + 547, + 383 + ], + "blocks": [ + { + "bbox": [ + 62, + 313, + 547, + 383 + ], + "lines": [ + { + "bbox": [ + 62, + 313, + 547, + 383 + ], + "spans": [ + { + "bbox": [ + 62, + 313, + 547, + 383 + ], + "type": "image", + "image_path": "4490ab9a6e715390594f56e20f7355d2a50fa3fbd9a8b2d7edbdf10e61e6fb89.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 62, + 384, + 547, + 453 + ], + "blocks": [ + { + "bbox": [ + 62, + 384, + 547, + 453 + ], + "lines": [ + { + "bbox": [ + 62, + 384, + 547, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 384, + 547, + 453 + ], + "type": "image", + "image_path": "2ae26fe8b485b7a3119433d3c2cc07c3041cbfa3aa7570aa45756eb54e360b6b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 62, + 453, + 547, + 523 + ], + "blocks": [ + { + "bbox": [ + 62, + 453, + 547, + 523 + ], + "lines": [ + { + "bbox": [ + 62, + 453, + 547, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 453, + 547, + 523 + ], + "type": "image", + "image_path": "883a9c2ab0c65c01db85179a41d31bc8ae28ec877c628825def7489cca6442c9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 535, + 555, + 569 + ], + "lines": [ + { + "bbox": [ + 54, + 535, + 555, + 569 + ], + "spans": [ + { + "bbox": [ + 54, + 535, + 555, + 569 + ], + "type": "text", + "content": "Figure 1. StableDepth achieves efficient online monocular depth estimation that produces scene-consistent and scale-invariant predictions frame by frame. As shown in the above challenging scenarios, StableDepth generates stable depth maps with precise geometric details and effectively handles dynamic motions, without the need of accessing future frames." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 151, + 578, + 200, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 578, + 200, + 590 + ], + "spans": [ + { + "bbox": [ + 151, + 578, + 200, + 590 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 603, + 297, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 297, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 297, + 688 + ], + "type": "text", + "content": "Recent advances in monocular depth estimation significantly improve robustness and accuracy. However, relative depth models exhibit flickering and 3D inconsistency in video data, limiting 3D reconstruction applications. We introduce StableDepth, a scene-consistent and scale-invariant depth estimation method achieving scene-level 3D consistency. Our dual-decoder architecture learns from" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 579, + 555, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 555, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 555, + 652 + ], + "type": "text", + "content": "large-scale unlabeled video data, enhancing generalization and reducing flickering. Unlike previous methods requiring full video sequences, StableDepth enables online inference at " + }, + { + "bbox": [ + 313, + 579, + 555, + 652 + ], + "type": "inline_equation", + "content": "13 \\times" + }, + { + "bbox": [ + 313, + 579, + 555, + 652 + ], + "type": "text", + "content": " faster speed, achieving significant improvements across benchmarks with comparable temporal consistency to video diffusion-based estimators." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 656, + 395, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 656, + 395, + 669 + ], + "spans": [ + { + "bbox": [ + 314, + 656, + 395, + 669 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "Monocular Depth Estimation (MDE) [12, 45] is pivotal in bridging the gap between 2D and 3D data representation, serving critical roles in a wide range of downstream applica" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": " Joint first authors. *Work done when interning at DAMO Academy. † Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7069" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "tions, such as autonomous driving [51], robotics [54], medical imaging [32], and virtual reality [13]. Advancements in foundational models [35, 43] have greatly enhanced the accuracy and robustness of MDE [25, 56]. Despite these advances, the intrinsic ambiguity of monocular cues still poses substantial challenges, particularly for downstream applications that require 3D-consistent data and maintaining stability for uncontrolled in-the-wild data." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 170, + 296, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 170, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 170, + 296, + 421 + ], + "type": "text", + "content": "MDE is primarily categorized into two main task settings: relative depth estimation and metric depth estimation. Relative depth estimation, as explored in studies [15, 18, 25, 39, 56], involves predicting depth values within a normalized depth range of " + }, + { + "bbox": [ + 56, + 170, + 296, + 421 + ], + "type": "inline_equation", + "content": "[0, 1]" + }, + { + "bbox": [ + 56, + 170, + 296, + 421 + ], + "type": "text", + "content": ". This scale- and shift-invariant regression target facilitates robust joint training across diverse data domains. However, the predicted depth values differ from the actual depth due to scale and shift discrepancies, leading to inconsistency across video frames. This inconsistency can complicate advanced downstream tasks like 3D reconstruction. In contrast, metric depth estimation [4, 22, 37, 59] aims to predict physical distances in metric scale. However, these approaches face significant challenges. Scaled versions of a scene's depth values might all represent plausible depth estimations of the actual depth. Such ambiguity complicates the learning process of depth regression, hindering both generalization capabilities and precise 3D structures. Severe flickering and 3D inconsistency are also observed given video inputs. This challenge prompts us to reconsider what is an effective representation of MDE foundation models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 424, + 295, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 424, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 295, + 543 + ], + "type": "text", + "content": "The limitations of existing approaches are summarized in Tab. 1, which compares different depth estimation paradigms. While relative depth models [25, 39, 56] offer good generalization across datasets but suffer from per-frame inconsistency, metric depth models [4, 59] provide real-world measurements but struggle with generalization across diverse scenes. Scale-invariant models [20] improve upon relative models by eliminating shift ambiguity while maintaining scale invariance, but they still lack consistency across frames in the same scene." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": "To address these challenges, we propose that the ideal task setting should be scene-consistent and scale-invariant (SCSI) depth estimation, as shown in Tab. 1. Our approach strikes a balance between per-frame scale-invariant methods that lack consistency and purely metric approaches that struggle with generalization. Specifically, our model primarily generates scale-invariant depth maps, but in the same scene or video, predictions across multiple frames demonstrate 3D consistency. Additionally, the scale factors for multi-frame data remain consistent compared with the actual depth values. In this work, the development of our model is primarily guided by the following objectives: 1) ensuring high-quality, scene-consistent, and scale-invariant depth estimations to achieve robust video consis" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 317, + 72, + 553, + 133 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 553, + 133 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 553, + 133 + ], + "type": "table", + "html": "
Depth TypeScaleShiftGeneralReconsConsist
RelativePer-framePer-frame✓✗✗
Scale-invariantPer-frameNone✓✓✗
MetricNoneNone✗✓✗
SCSI (Ours)Per-sceneNone✓✓✓
", + "image_path": "3fb730c9fb83678fb4f6d3f9836626a91056d215a9833fbdf667df335e09a04b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 139, + 555, + 183 + ], + "lines": [ + { + "bbox": [ + 313, + 139, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 313, + 139, + 555, + 183 + ], + "type": "text", + "content": "Table 1. Comparison of different depth types. Our proposed SCSI depth combines the strengths of relative and metric depth while overcoming their weaknesses. General: generalization capability. Recons: reconstruction quality. Consist: scene consistency." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 196, + 553, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 196, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 196, + 553, + 220 + ], + "type": "text", + "content": "tency; 2) enhancing strong generalization capabilities and precise depth details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 221, + 555, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 221, + 555, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 555, + 412 + ], + "type": "text", + "content": "Our framework introduces three key innovations to achieve scene-consistent depth estimation. First, we leverage temporal priors from pre-trained video diffusion models to generate pseudo-labels with inherent 3D consistency, replacing traditional unstable photometric constraints [3, 23, 47, 48]. Second, we design a dual-branch architecture where one branch processes labeled data for metric accuracy, while the other aligns unlabeled video frames to pseudo-labels through scene-level normalization, decoupling conflicting objectives. Third, our semi-supervised paradigm jointly optimizes metric supervision and temporal alignment, enabling the model to generalize across domains while eliminating flickering artifacts. This unified approach supports real-time inference without requiring future frames, overcoming the limitations of both single-frame and video-based methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 413, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 554, + 544 + ], + "type": "text", + "content": "We evaluate our model on multiple datasets under zero-shot settings, aligning each scene using only a single scale (SCSI). As demonstrated in Figure 1, our method robustly handles challenging scenarios including dynamic motions and complex textures while preserving geometric fidelity frame-by-frame. Compared to DepthCrafter [23], we achieve " + }, + { + "bbox": [ + 313, + 413, + 554, + 544 + ], + "type": "inline_equation", + "content": "13 \\times" + }, + { + "bbox": [ + 313, + 413, + 554, + 544 + ], + "type": "text", + "content": " faster inference speed with online processing capability, improving prediction accuracy on Sintel [5], Bonn [36], ScanNet v2 [9], and KITTI [17] by " + }, + { + "bbox": [ + 313, + 413, + 554, + 544 + ], + "type": "inline_equation", + "content": "13.2\\%" + }, + { + "bbox": [ + 313, + 413, + 554, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 413, + 554, + 544 + ], + "type": "inline_equation", + "content": "86.8\\%" + }, + { + "bbox": [ + 313, + 413, + 554, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 413, + 554, + 544 + ], + "type": "inline_equation", + "content": "39.3\\%" + }, + { + "bbox": [ + 313, + 413, + 554, + 544 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 413, + 554, + 544 + ], + "type": "inline_equation", + "content": "8.2\\%" + }, + { + "bbox": [ + 313, + 413, + 554, + 544 + ], + "type": "text", + "content": ", respectively. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 546, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 313, + 546, + 553, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 553, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 553, + 605 + ], + "type": "text", + "content": "- We introduce StableDepth, a novel monocular depth estimation framework that achieves scene-consistent and scale-invariant (SCSI) depth estimation, bridging the gap between relative and metric depth prediction while maintaining temporal consistency in video sequences." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 606, + 553, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 553, + 666 + ], + "type": "text", + "content": "- We propose a video diffusion baking strategy that leverages pre-trained video diffusion models for generating high-quality pseudo-labels, enabling effective semi-supervised learning from unlabeled video data while preserving temporal consistency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 666, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 553, + 714 + ], + "type": "text", + "content": "- We develop a dual-decoder architecture that simultaneously handles supervised metric depth estimation and unsupervised depth alignment, effectively addressing the target misalignment issue while maintaining accuracy." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7070" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "- We demonstrate state-of-the-art performance across multiple benchmarks, achieving up to " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "86.8\\%" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " improvement in accuracy while providing " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "13 \\times" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " faster inference compared to previous methods. Our method enables efficient, online inference while maintaining comparable temporal stability to video diffusion models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 156, + 141, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 141, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 141, + 168 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 177, + 296, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 177, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 177, + 296, + 308 + ], + "type": "text", + "content": "Monocular depth estimation. MDE focuses on inferring depth from a single image. Early works [8, 31, 55] learn ordinal relationships from coarse annotations but cannot recover geometric structure. Recent methods [11, 39, 56, 57] leverage large-scale data to learn affine-invariant depth with improved generalization. However, unknown depth shifts cause geometric distortions, hindering 3D applications. To address this, recent works attempt zero-shot metric depth estimation [4, 19, 22, 37, 38, 59, 61] and explore self-supervised scale-consistent learning [50, 60], though maintaining consistent scale across scenes remains challenging." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 309, + 296, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 309, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 296, + 475 + ], + "type": "text", + "content": "Towards consistent video depth. Video depth estimation requires both temporal consistency and per-frame precision. Traditional approaches use post-processing optimization [27, 34] or temporal modeling [29, 49], but incur computational costs or fail in dynamic scenes. Other methods leverage camera poses [46, 53] or memory-based approaches [52, 58], but often sacrifice geometric fidelity for smoothness. Recent works like Video Depth Anything [7] and video-free approaches [26] show promise for temporal consistency in long sequences. Video Depth Anything freezes the Depth Anything V2 encoder and incorporates cross-frame attention in the DPT [40] decoder for temporal consistency, though these methods still produce depth with inherent shift ambiguity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 477, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 477, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 477, + 296, + 571 + ], + "type": "text", + "content": "Diffusion priors for depth modeling. Video diffusion models [21, 24] demonstrate potential for consistent depth estimation. DepthCrafter [23] achieves temporal stability but generates only relative depths, while ChronoDepth [47] suffers from limited temporal contexts. Our proposed SCSI depth estimation leverages video diffusion baking and dual-decoder architecture for temporally consistent depth suitable for efficient online inference." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 585, + 134, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 585, + 134, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 585, + 134, + 597 + ], + "type": "text", + "content": "3. StableDepth" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "Our work leverages both labeled images and unlabeled video data to enhance SCSI depth predictions, achieving video consistency and enhanced generalization capabilities through the distillation of video diffusion priors. Formally, we denote the sets of labeled image-depth pairs and unlabeled videos as " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^l = \\{(x_i,d_i)\\}_{i = 1}^M" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^u = \\{v_i\\}_{i = 1}^N" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": ", respectively. Our objective is to utilize a high-quality labeled image dataset, Hypersim [42], to help the model grasp real-world scale, while employing unlabeled video data to guide" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "the model's depth predictions toward achieving video consistency and improved generalization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 105, + 539, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 105, + 539, + 118 + ], + "spans": [ + { + "bbox": [ + 313, + 105, + 539, + 118 + ], + "type": "text", + "content": "3.1. Mining Consistency from Unlabeled Videos" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 121, + 554, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 554, + 211 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 554, + 211 + ], + "type": "text", + "content": "StableDepth is designed to learn from both labeled images and unlabeled videos. For labeled image data, we predict the depth in metric space. The activation function of the network's last layer is sigmoid, after which its predictions are scaled to match the dataset values by multiplying with a predefined maximum depth. The SiLogLoss [12, 44, 57] is then used to minimize the error between the prediction and the ground truth:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 320, + 219, + 553, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 219, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 320, + 219, + 553, + 262 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {l a b e l}} = \\sqrt {\\frac {1}{H W} \\sum_ {i = 1} ^ {H W} \\left(\\log \\frac {d _ {i}}{d _ {i} ^ {*}}\\right) ^ {2} - \\lambda \\left(\\frac {1}{H W} \\sum_ {i = 1} ^ {H W} \\log \\frac {d _ {i}}{d _ {i} ^ {*}}\\right) ^ {2}}, \\tag {1}", + "image_path": "cfbfa8202c71c0b44a1574d0c70f5d8765aeaa3681fb5c27fe890817ebde617b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 262, + 553, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 553, + 286 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 262, + 553, + 286 + ], + "type": "inline_equation", + "content": "d_i^*" + }, + { + "bbox": [ + 313, + 262, + 553, + 286 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 262, + 553, + 286 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 313, + 262, + 553, + 286 + ], + "type": "text", + "content": " are the prediction and ground truth, respectively, and " + }, + { + "bbox": [ + 313, + 262, + 553, + 286 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 262, + 553, + 286 + ], + "type": "text", + "content": " is a hyper-parameter." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 287, + 555, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 555, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 555, + 550 + ], + "type": "text", + "content": "For unlabeled video data, we leverage video diffusion models, specifically DepthCrafter [23], to generate pseudolabels due to their inherent advantages in maintaining temporal consistency. These models, pre-trained on large-scale video data with cross-frame attention mechanisms, can generate stable depth predictions across entire video segments while preserving a consistent global scale and shift relationship with metric depth. Unlike conventional image-based MDE approaches that process frames independently, video diffusion models account for temporal dependencies, making them particularly effective for generating reliable pseudo-labels that exhibit strong video consistency - a crucial property for our scene-consistent training objective. For a video, we need a method to normalize the depth predicted by our model to the same space as the pseudo-labeled depth. Unlike the normalization used in single-frame relative depth training, this approach focuses on ensuring that an entire video has a unique scale and shift, rather than each frame having an independent scale and shift. This method aims to achieve video consistency in the depth predictions within the same scene. To supervise our model, we first use the following formula to normalize the pseudo-labels," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 372, + 559, + 553, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 559, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 372, + 559, + 553, + 586 + ], + "type": "interline_equation", + "content": "\\hat {D} (v _ {i}) = \\frac {D (v _ {i}) - m (D (v _ {i}))}{s (D (v _ {i}))}, \\tag {2}", + "image_path": "8a1c53c62e7ecf2d421a69b951723351ae7f8c54c5c79f879c0b6bc5a8ab27b9.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 594, + 553, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 594, + 553, + 628 + ], + "type": "inline_equation", + "content": "D(v_{i})" + }, + { + "bbox": [ + 313, + 594, + 553, + 628 + ], + "type": "text", + "content": " is the pseudo-label for video " + }, + { + "bbox": [ + 313, + 594, + 553, + 628 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 313, + 594, + 553, + 628 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 594, + 553, + 628 + ], + "type": "inline_equation", + "content": "m(D(v_{i}))" + }, + { + "bbox": [ + 313, + 594, + 553, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 553, + 628 + ], + "type": "inline_equation", + "content": "s(D(v_{i}))" + }, + { + "bbox": [ + 313, + 594, + 553, + 628 + ], + "type": "text", + "content": " are used to remove the shift and scale for this video, respectively:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 341, + 637, + 453, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 637, + 453, + 649 + ], + "spans": [ + { + "bbox": [ + 341, + 637, + 453, + 649 + ], + "type": "interline_equation", + "content": "m (D (v _ {i})) = \\operatorname {m e d i a n} (D (v _ {i})),", + "image_path": "d9be84aa16138055565be6fe33eb4afe05c8c6b236b98d4d98b0cfc19c4ffb62.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 345, + 652, + 553, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 652, + 553, + 681 + ], + "spans": [ + { + "bbox": [ + 345, + 652, + 553, + 681 + ], + "type": "interline_equation", + "content": "s \\left(D \\left(v _ {i}\\right)\\right) = \\frac {1}{T H W} \\sum_ {i = 1} ^ {T H W} | D \\left(v _ {i}\\right) - m \\left(D \\left(v _ {i}\\right)\\right) |. \\tag {3}", + "image_path": "d96d924dff4f661b6a0203b6de747ff4ee5c62d7316f5e193203a986f34511bb.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "For the frame-by-frame depth predictions of video " + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": " from our model, we concatenate them to obtain " + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "inline_equation", + "content": "P(v_{i})" + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": " with" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7071" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 68, + 553, + 201 + ], + "blocks": [ + { + "bbox": [ + 58, + 68, + 553, + 201 + ], + "lines": [ + { + "bbox": [ + 58, + 68, + 553, + 201 + ], + "spans": [ + { + "bbox": [ + 58, + 68, + 553, + 201 + ], + "type": "image", + "image_path": "347cbf4b90ce09213c1072109f744af1d03ab5cc6bb7b8787a1157f68f170541.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 201, + 555, + 269 + ], + "lines": [ + { + "bbox": [ + 54, + 201, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 54, + 201, + 555, + 269 + ], + "type": "text", + "content": "Figure 2. Overview of StableDepth. StableDepth employs a shared encoder with dual decoders to process both labeled images and unlabeled videos. The upper branch handles labeled images with direct metric depth supervision (\"Sup\"), while the lower branch processes unlabeled video sequences using temporal-coherent pseudo-labels generated by pre-trained video diffusion priors [23] (\"Bake\"). This architecture enables scene-consistent and scale-invariant (SCSI) depth prediction by combining the benefits of metric depth supervision with temporal consistency learning from video sequences. Solid arrows indicate the primary data flow, while dashed arrows represent the auxiliary training path for unlabeled data." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 281, + 296, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 296, + 329 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 296, + 329 + ], + "type": "text", + "content": "dimensions " + }, + { + "bbox": [ + 55, + 281, + 296, + 329 + ], + "type": "inline_equation", + "content": "[T,H,W]" + }, + { + "bbox": [ + 55, + 281, + 296, + 329 + ], + "type": "text", + "content": ". We then apply the same normalization process to the depth predictions " + }, + { + "bbox": [ + 55, + 281, + 296, + 329 + ], + "type": "inline_equation", + "content": "P(v_{i})" + }, + { + "bbox": [ + 55, + 281, + 296, + 329 + ], + "type": "text", + "content": " to obtain its normalized version " + }, + { + "bbox": [ + 55, + 281, + 296, + 329 + ], + "type": "inline_equation", + "content": "\\hat{P} (v_{i})" + }, + { + "bbox": [ + 55, + 281, + 296, + 329 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 55, + 281, + 296, + 329 + ], + "type": "inline_equation", + "content": "\\hat{P} (v_{i})" + }, + { + "bbox": [ + 55, + 281, + 296, + 329 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 281, + 296, + 329 + ], + "type": "inline_equation", + "content": "\\hat{D} (v_{i})" + }, + { + "bbox": [ + 55, + 281, + 296, + 329 + ], + "type": "text", + "content": ", we use the following alignment loss to minimize their difference," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 90, + 337, + 295, + 370 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 337, + 295, + 370 + ], + "spans": [ + { + "bbox": [ + 90, + 337, + 295, + 370 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {u n l a b e l}} = \\frac {1}{T H W} \\sum_ {i = 1} ^ {T H W} \\left| \\hat {P} \\left(v _ {i}\\right) - \\hat {D} \\left(v _ {i}\\right) \\right|. \\tag {4}", + "image_path": "c3a2a07599cdf4736e1dd55487b070082f1305a065f4d9a844bed1588a41ef67.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 378, + 296, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 296, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 296, + 425 + ], + "type": "text", + "content": "In one iteration, we simultaneously sample a certain proportion of labeled images and unlabeled videos for joint training, and use the following combined loss as the final loss for this iteration:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 129, + 435, + 295, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 435, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 129, + 435, + 295, + 449 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {l a b e l}} + \\gamma \\mathcal {L} _ {\\text {u n l a b e l}}, \\tag {5}", + "image_path": "5dffd0ee8a8258e36f69e31d304fd3b5b2efd25d5a94877fa88d124caca8d067.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 456, + 296, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 296, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 296, + 481 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 456, + 296, + 481 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 456, + 296, + 481 + ], + "type": "text", + "content": " gradually increases following a cosine curve as the number of iterations progresses." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 487, + 237, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 237, + 501 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 237, + 501 + ], + "type": "text", + "content": "3.2. Avoiding Conflict in Depth Labels" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 505, + 295, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 295, + 552 + ], + "type": "text", + "content": "In the field of video consistency for depth prediction, most research [23] focuses on improving video consistency in near-field scenes. Therefore, we use depth in the disparity space for video data training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 553, + 295, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 553, + 295, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 553, + 295, + 649 + ], + "type": "text", + "content": "Following the previous network structure [2, 39, 56, 57], we use the ViT-L encoder [10], denoted as " + }, + { + "bbox": [ + 55, + 553, + 295, + 649 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 553, + 295, + 649 + ], + "type": "text", + "content": ", paired with the decoder from DPT [40], denoted as " + }, + { + "bbox": [ + 55, + 553, + 295, + 649 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_1" + }, + { + "bbox": [ + 55, + 553, + 295, + 649 + ], + "type": "text", + "content": ". The decoder outputs depth in SCSI space with a range of (0, 1). A natural approach is to utilize this original decoder to predict the disparity-space depth of a video sequence, and then take the reciprocal of the predictions. Specifically, the depth prediction " + }, + { + "bbox": [ + 55, + 553, + 295, + 649 + ], + "type": "inline_equation", + "content": "P(v_{i})" + }, + { + "bbox": [ + 55, + 553, + 295, + 649 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 657, + 295, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 657, + 295, + 682 + ], + "spans": [ + { + "bbox": [ + 132, + 657, + 295, + 682 + ], + "type": "interline_equation", + "content": "P \\left(v _ {i}\\right) = \\frac {1}{\\mathcal {D} _ {1} \\left(\\mathcal {E} \\left(v _ {i}\\right)\\right)}. \\tag {6}", + "image_path": "1d60af933627f0be159f26f35b96780a9e3e1fc7ba0f08cba890e1090114d6a3.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Finally, the depth predictions are normalized and supervised in semi-supervised settings following Eq. 3 and 4." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 281, + 555, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 424 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 424 + ], + "type": "text", + "content": "Unfortunately, in our experiments, we found that although this modification improves the video consistency of depth predictions for videos, it reduces the accuracy of depth prediction by the model, resulting in lower metrics such as " + }, + { + "bbox": [ + 313, + 281, + 555, + 424 + ], + "type": "inline_equation", + "content": "\\delta_{1}" + }, + { + "bbox": [ + 313, + 281, + 555, + 424 + ], + "type": "text", + "content": ". We speculate that this is mainly due to the used pseudo-labeling model [23], which, despite having good video consistency, still lags behind Depth Anything V2 [57] in terms of depth prediction accuracy, particularly in the prediction of distant objects. To address this issue, we propose to use an auxiliary network to decouple the prediction and supervision of labeled images and unlabeled videos. To this end, we experimented with two approaches:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 425, + 554, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 554, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 554, + 484 + ], + "type": "text", + "content": "In the first approach, both labeled images and unlabeled videos share the encoder " + }, + { + "bbox": [ + 313, + 425, + 554, + 484 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 425, + 554, + 484 + ], + "type": "text", + "content": " and the decoder " + }, + { + "bbox": [ + 313, + 425, + 554, + 484 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_1" + }, + { + "bbox": [ + 313, + 425, + 554, + 484 + ], + "type": "text", + "content": ". A lightweight CNN network, denoted as " + }, + { + "bbox": [ + 313, + 425, + 554, + 484 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 425, + 554, + 484 + ], + "type": "text", + "content": ", is used to directly process the prediction results of the original decoder " + }, + { + "bbox": [ + 313, + 425, + 554, + 484 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_1" + }, + { + "bbox": [ + 313, + 425, + 554, + 484 + ], + "type": "text", + "content": " with the CNN network outputting depth in disparity space:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 384, + 490, + 553, + 504 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 490, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 384, + 490, + 553, + 504 + ], + "type": "interline_equation", + "content": "P \\left(v _ {i}\\right) = \\mathcal {C} \\left(\\mathcal {D} _ {1} \\left(\\mathcal {E} \\left(v _ {i}\\right)\\right)\\right). \\tag {7}", + "image_path": "aec30edcbc91232efff868103a230f793c1952ae39572b066de2290c2d88849c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 509, + 554, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 554, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 554, + 582 + ], + "type": "text", + "content": "In the second approach, labeled images and unlabeled videos share the encoder " + }, + { + "bbox": [ + 313, + 509, + 554, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 509, + 554, + 582 + ], + "type": "text", + "content": ". An additional DPT [40] decoder, denoted as " + }, + { + "bbox": [ + 313, + 509, + 554, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_2" + }, + { + "bbox": [ + 313, + 509, + 554, + 582 + ], + "type": "text", + "content": ", is used to process the features extracted by the encoder " + }, + { + "bbox": [ + 313, + 509, + 554, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 509, + 554, + 582 + ], + "type": "text", + "content": " for unlabeled videos, as illustrated by Fig. 2. " + }, + { + "bbox": [ + 313, + 509, + 554, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_2" + }, + { + "bbox": [ + 313, + 509, + 554, + 582 + ], + "type": "text", + "content": " directly produces depth in disparity space, while " + }, + { + "bbox": [ + 313, + 509, + 554, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_1" + }, + { + "bbox": [ + 313, + 509, + 554, + 582 + ], + "type": "text", + "content": " continues to output depth in SCSI space." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 346, + 587, + 553, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 587, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 346, + 587, + 553, + 601 + ], + "type": "interline_equation", + "content": "P \\left(x _ {i}\\right) = \\mathcal {D} _ {1} \\left(\\mathcal {E} \\left(x _ {i}\\right)\\right), \\quad P \\left(v _ {i}\\right) = \\mathcal {D} _ {2} \\left(\\mathcal {E} \\left(v _ {i}\\right)\\right). \\tag {8}", + "image_path": "e1f372d63ab908a31de0daf7ed063154d2ea5d5751b96922dbb71772edf52e2f.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 606, + 554, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 554, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 554, + 654 + ], + "type": "text", + "content": "The final results indicate that sharing the encoder between labeled images and unlabeled videos, while using two independent DPT decoders, achieves the best accuracy and video consistency." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 660, + 502, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 502, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 502, + 673 + ], + "type": "text", + "content": "3.3. Scene-Consistent Depth Evaluation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Scale alignment. For zero-shot evaluation, we enforce single-scale consistency per scene through RANSAC-based alignment with zero shift constraint. Each scene's" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7072" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 170 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 170 + ], + "type": "text", + "content": "predictions are scaled by a unified factor determined through 1,000 sampling iterations, retaining inliers with " + }, + { + "bbox": [ + 55, + 72, + 294, + 170 + ], + "type": "inline_equation", + "content": "\\max \\left(\\frac{d_{pred}}{d_{gt}},\\frac{d_{gt}}{d_{pred}}\\right) < \\tau" + }, + { + "bbox": [ + 55, + 72, + 294, + 170 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 72, + 294, + 170 + ], + "type": "inline_equation", + "content": "\\tau = 1.25" + }, + { + "bbox": [ + 55, + 72, + 294, + 170 + ], + "type": "text", + "content": ". The optimal scale is refined via least-squares optimization over all inliers, preserving geometric fidelity while eliminating cross-frame scale ambiguity. This per-scene single-scale paradigm ensures coherent depth relationships across dynamic sequences without per-frame adjustments." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 171, + 294, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 294, + 208 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 294, + 208 + ], + "type": "text", + "content": "Motion-aware temporal metric. We propose the Motion-aware Temporal Difference (MTD) to quantify depth consistency in dynamic regions:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 95, + 219, + 295, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 219, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 95, + 219, + 295, + 251 + ], + "type": "interline_equation", + "content": "\\mathrm {M T D} = \\frac {1}{T - 1} \\sum_ {t = 1} ^ {T - 1} \\frac {\\sum_ {x , y} | \\Delta D _ {t} | \\cdot W _ {t}}{H \\times W}, \\tag {9}", + "image_path": "9432f8e3dbaee6bbfb020bcc1920d69c7162f8f8d0262ab12ee95a50dc2a9ed4.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 262, + 295, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 295, + 322 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 295, + 322 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 262, + 295, + 322 + ], + "type": "inline_equation", + "content": "\\Delta D_{t} = D_{t} - D_{t - 1}" + }, + { + "bbox": [ + 55, + 262, + 295, + 322 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 262, + 295, + 322 + ], + "type": "inline_equation", + "content": "W_{t}(x,y) = (1 + \\| \\mathbf{f}_{t}\\|_{2})^{-1}" + }, + { + "bbox": [ + 55, + 262, + 295, + 322 + ], + "type": "text", + "content": " weights the depth variations by optical flow magnitude " + }, + { + "bbox": [ + 55, + 262, + 295, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t" + }, + { + "bbox": [ + 55, + 262, + 295, + 322 + ], + "type": "text", + "content": " (computed via Farneback [14] algorithm). Lower MTD indicates better temporal stability, penalizing flickering in high-motion areas." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 334, + 128, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 334, + 128, + 345 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 128, + 345 + ], + "type": "text", + "content": "3.4. Discussion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 353, + 295, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 353, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 353, + 295, + 388 + ], + "type": "text", + "content": "In this section, we address several important aspects of our approach to provide further clarity on methodology, design choices, and relationships to existing work." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 389, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 389, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 54, + 389, + 295, + 555 + ], + "type": "text", + "content": "Decouple accuracy and consistency via dual learning. Our dual-decoder architecture decouples depth accuracy from temporal consistency learning, enabling joint optimization from diverse data sources. The primary decoder specializes in per-pixel depth precision through supervised training on high-quality labeled images, while the auxiliary decoder learns scene-level coherence from unlabeled videos via pseudo-labels. This separation resolves the inherent conflict where video-based methods like DepthCrafter [23] sacrifice depth accuracy for temporal smoothness, and single-frame approaches fail to ensure cross-frame consistency. Crucially, our design preserves absolute depth scales critical for 3D reconstruction, unlike relative-depth representations in video diffusion models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": "Beyond video diffusion priors. We repurpose video diffusion models as sources for scene consistency rather than as direct depth estimators. While models like DepthCrafter [23] excel at generating temporally consistent sequences, they produce relative depth representations with limited accuracy, particularly for distant objects, making them unsuitable for downstream tasks requiring precise geometry. Our approach extracts the temporal consistency knowledge from these models through scene-level normalization while maintaining superior depth accuracy. This strategy enables online inference " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "13 \\times" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": " faster than video-based methods while achieving both better depth precision and comparable temporal stability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 72, + 391, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 391, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 391, + 85 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 91, + 447, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 91, + 447, + 104 + ], + "spans": [ + { + "bbox": [ + 314, + 91, + 447, + 104 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 110, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 553, + 312 + ], + "type": "text", + "content": "Network parameters and training details. We use the network architecture from Depth Anything V2 (DAv2) [57], consisting of a pre-trained ViT-L [10] encoder for feature extraction and a DPT [40] decoder for depth regression. The ViT-L encoder is initialized with parameters from the DAv2 model, while the primary decoder is randomly initialized with Sigmoid activation for metric depth output. The auxiliary decoder, also initialized randomly, uses ReLU activation for disparity (inverse depth) prediction. Each batch consists of a 1:2 ratio of labeled images to unlabeled video frames, totaling 24 frames. The learning rate for the encoder is " + }, + { + "bbox": [ + 313, + 110, + 553, + 312 + ], + "type": "inline_equation", + "content": "2.5 \\times 10^{-6}" + }, + { + "bbox": [ + 313, + 110, + 553, + 312 + ], + "type": "text", + "content": ", while the decoder's learning rate is 10 times larger. We use the AdamW optimizer [33] with a weight decay of 0.01 and employ a polynomial decay schedule for the learning rate. For data augmentation, we apply horizontal flipping to labeled images. In Eq. (5), the maximum value of " + }, + { + "bbox": [ + 313, + 110, + 553, + 312 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 110, + 553, + 312 + ], + "type": "text", + "content": " is set to 1. In Eq. (1), " + }, + { + "bbox": [ + 313, + 110, + 553, + 312 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 110, + 553, + 312 + ], + "type": "text", + "content": " is set to 0.5." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 314, + 553, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 553, + 481 + ], + "type": "text", + "content": "Training dataset and pseudo-labeling. To train our model, we utilize DepthCrafter [23], a model finetuned on SVD [3], to generate video-consistent pseudo-labels for unlabeled video data. For metric depth supervision, we use Hypersim [42], a high-quality synthetic dataset with " + }, + { + "bbox": [ + 313, + 314, + 553, + 481 + ], + "type": "inline_equation", + "content": "\\sim 60\\mathrm{K}" + }, + { + "bbox": [ + 313, + 314, + 553, + 481 + ], + "type": "text", + "content": " frames, as training data for decoder " + }, + { + "bbox": [ + 313, + 314, + 553, + 481 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_1" + }, + { + "bbox": [ + 313, + 314, + 553, + 481 + ], + "type": "text", + "content": ". We select 6,000 videos from SA-V [41], which features dynamic foreground objects, yielding approximately " + }, + { + "bbox": [ + 313, + 314, + 553, + 481 + ], + "type": "inline_equation", + "content": "\\sim 200\\mathrm{K}" + }, + { + "bbox": [ + 313, + 314, + 553, + 481 + ], + "type": "text", + "content": " frames for pseudolabeling training. To enhance the model's performance on distant scenes where DepthCrafter shows limitations, we incorporate additional datasets with accurate depth labels including VKITTI2 [6, 16], LightwheelOcc [30], and MatrixCity [28], treating frames from the same scene as video sequences during training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 491, + 388, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 491, + 388, + 502 + ], + "spans": [ + { + "bbox": [ + 314, + 491, + 388, + 502 + ], + "type": "text", + "content": "4.2. Evaluation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 509, + 553, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 553, + 628 + ], + "type": "text", + "content": "Evaluation datasets. We evaluate on four datasets spanning diverse scenarios: Sintel [5] (23 synthetic videos, 50 frames each with precise depth labels), Bonn [36] (26 dynamic videos using frames 30-140 with foreground motions), ScanNet v2 [9] (100 indoor test videos sampled every third frame), and the full KITTI validation set [17] (13 outdoor driving videos with first 110 frames per sequence). This comprehensive benchmark covers synthetic animations, dynamic interactions, static indoor scenes, and real-world autonomous driving environments." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 630, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 712 + ], + "type": "text", + "content": "Quantitative results. We evaluate our method against strong baselines: DepthCrafter [23], known for video consistency, Metric3D [59], a state-of-the-art metric depth estimator, and ZoeDepth [1], a robust monocular depth estimation method. For fair comparison, we align all predictions within each video using a single shared scale factor across the entire scene with shift set to zero. As shown in" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7073" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 558, + 252 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 558, + 252 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 558, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 558, + 252 + ], + "type": "table", + "html": "
Frame IntervalMethodSintel (~50 frames)Bonn (110 frames)ScanNet (90 frames)KITTI (110 frames)
AbsRel (↓)δ1 (↑)AbsRel (↓)δ1 (↑)AbsRel (↓)δ1 (↑)AbsRel (↓)δ1 (↑)
1DepthCrafter [23]2.9590.5450.3110.5240.3700.6610.1230.832
Zoedepth [1]0.4540.3560.2770.5620.3470.4360.2160.633
Metric3D [59]0.3720.6100.0550.9810.0590.9740.0670.977
StableDepth (Ours)0.2800.6170.0630.9790.0940.9210.1120.900
4DepthCrafter [23]6.3530.3930.5150.4170.5390.4400.2440.670
Zoedepth [1]0.4610.3570.2770.5630.3490.4360.2180.629
Metric3D [59]0.3610.6170.0550.9810.0590.9740.0670.977
StableDepth (Ours)0.2920.6060.0630.9790.0930.9210.1120.901
8DepthCrafter [23]2.0480.3980.4950.4120.5710.4230.2410.656
Zoedepth [1]0.4700.3570.2760.5620.3520.4340.2180.633
Metric3D [59]0.3630.6130.0550.9800.0600.9730.0660.977
StableDepth (Ours)0.3000.6040.0620.9800.0940.9200.1120.902
", + "image_path": "1a27920025350c8d0a704b45fdde248e5d80d87f775e4ce1ddcee77bccd5173e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 300, + 294, + 384 + ], + "blocks": [ + { + "bbox": [ + 55, + 259, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 55, + 259, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 555, + 293 + ], + "type": "text", + "content": "Table 2. Comparison of depth estimation methods across different frame intervals. StableDepth maintains consistent performance across varying frame intervals while achieving superior accuracy. Best results are shown in bold, second best are underlined. Metric3D results on ScanNet are shown in gray as it was trained on this dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 300, + 294, + 384 + ], + "lines": [ + { + "bbox": [ + 56, + 300, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 300, + 294, + 384 + ], + "type": "table", + "html": "
MethodBonnScanNet
AbsRel (↓)δ1(↑)AbsRel (↓)δ1(↑)
Baseline0.0710.9460.1020.903
DPT0.1090.9140.1260.861
DPT + CNN0.1080.9260.1380.829
Dual DPT0.0570.9760.0970.913
", + "image_path": "358b9d6ff7641a7bc714845fdaba20a028a5e2155d990be68dde45ff668fad0a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 389, + 295, + 445 + ], + "lines": [ + { + "bbox": [ + 55, + 389, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 295, + 445 + ], + "type": "text", + "content": "Table 3. Ablation on decoders. The baseline model is trained only on labeled image data. We compare three approaches for handling unlabeled video data: using a single DPT decoder with reciprocal output (DPT), adding a lightweight CNN head (DPT + CNN), and using dual DPT decoders (Dual DPT)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 448, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 295, + 545 + ], + "type": "text", + "content": "Tab. 2, our method achieves superior performance across most datasets and metrics, with particularly strong improvements in AbsRel and " + }, + { + "bbox": [ + 55, + 448, + 295, + 545 + ], + "type": "inline_equation", + "content": "\\delta_{1}" + }, + { + "bbox": [ + 55, + 448, + 295, + 545 + ], + "type": "text", + "content": " metrics compared to video-based approaches. We examine the impact of frame intervals on prediction accuracy to simulate varying temporal densities in real-world scenarios. Note that compared to our ablation studies, these results reflect extended training with additional iterations for optimal performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 552, + 149, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 149, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 149, + 565 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "Comparison of different decoders. In Tab. 3, we compare four methods using the Bonn and ScanNet v2 datasets: training with only labeled Hypersim data, and training with both labeled Hypersim [42] and unlabeled SA-V [41] data processed by three decoder designs. The first row (Baseline) shows results with only labeled Hypersim data, while the following rows show results with both labeled and unlabeled data. For labeled Hypersim data, depth in SCSI space is obtained directly from the DPT head output. For the unlabeled SA-V data, the method for obtaining disparity space depth varies: in the second row (DPT), it's the reciprocal of the original DPT output; in the third row (DPT" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 303, + 555, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 555, + 411 + ], + "type": "text", + "content": "+ CNN), a lightweight CNN processes the DPT output; in the fourth row (Dual DPT), an additional DPT head provides the disparity space depth. The results show that using a single DPT head, either by reciprocating its output or adding a lightweight CNN, reduces prediction accuracy due to the partial inaccuracy of the pseudo-labels. However, using two independent DPT heads improves frame-to-frame consistency without impacting accuracy, validating the effectiveness of our disentangled approach." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 411, + 556, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 556, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 556, + 508 + ], + "type": "text", + "content": "Encoder feature stability. We analyze encoder feature stability across consecutive DAVIS frames to understand our dual-decoder effectiveness. Compared to DAv2, our method achieves " + }, + { + "bbox": [ + 313, + 411, + 556, + 508 + ], + "type": "inline_equation", + "content": "7.1\\%" + }, + { + "bbox": [ + 313, + 411, + 556, + 508 + ], + "type": "text", + "content": " lower mean differences, " + }, + { + "bbox": [ + 313, + 411, + 556, + 508 + ], + "type": "inline_equation", + "content": "13.5\\%" + }, + { + "bbox": [ + 313, + 411, + 556, + 508 + ], + "type": "text", + "content": " reduced standard deviation, and " + }, + { + "bbox": [ + 313, + 411, + 556, + 508 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 313, + 411, + 556, + 508 + ], + "type": "text", + "content": " higher cosine similarity. These improvements confirm that auxiliary decoder supervision enhances encoder stability, contributing to temporal consistency in depth predictions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 509, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 556, + 628 + ], + "type": "text", + "content": "Effectiveness of unlabeled video data. As shown in Tab. 4, we evaluate incremental training data impact. Using only Hypersim [42] (row 1), then adding pseudo-labeled SA-V [41] data (row 2) shows mixed results - improving Bonn and ScanNet v2 but decreasing Sintel and KITTI performance. This reveals DepthCrafter's video consistency benefits but limitations in distant regions. Progressively incorporating VKITTI2 [6, 16], LightWheel [30], Hypersim [42], and MatrixCity [28] (rows 3-6) successfully addresses these limitations, achieving best performance across all datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Comparison of video consistency. Tab. 5 showcases the performance comparison of depth estimation methods on the DAVIS dataset evaluated using Motion-aware temporal Difference (MTD (Eq. (9)), lower is better). The video diffusion-based DepthCrafter [23] achieves strong consistency " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "(\\mathrm{MTD} = 0.003479)" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": " but requires processing the entire video sequence and runs at only 0.45 FPS. While" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7074" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 555, + 177 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 177 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 177 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 177 + ], + "type": "table", + "html": "
Training DataSintel (~50 frames)Bonn (110 frames)ScanNet (90 frames)KITTI (110 frames)
AbsRel (↓)δ1 (↑)AbsRel (↓)δ1 (↑)AbsRel (↓)δ1 (↑)AbsRel (↓)δ1 (↑)
Hypersim only [42]0.3090.5780.0710.9460.1020.9030.1220.871
+ SA-V [41]0.3260.5610.0570.9760.0970.9130.1430.823
+ VKITTI2 [6, 16]0.3180.5770.0580.9760.0970.9110.1290.839
+ LightWheel [30]0.3120.5730.0600.9760.0940.9160.1340.847
+ Hypersim-U [42]0.3140.5720.0640.9760.0960.9180.1230.870
+ Matrixcity [28]0.3150.5800.0620.9790.0920.9240.1140.895
", + "image_path": "26f4f6fe90a9b8d2c1293a0843f558d906dce92089d7881c0505814eedee670d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 247, + 294, + 340 + ], + "blocks": [ + { + "bbox": [ + 56, + 247, + 294, + 340 + ], + "lines": [ + { + "bbox": [ + 56, + 247, + 294, + 340 + ], + "spans": [ + { + "bbox": [ + 56, + 247, + 294, + 340 + ], + "type": "image", + "image_path": "517d05a13a0817eb8f234dc7db953a936edae9d9fd565f6c8693aa1dbec0d07e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 349, + 295, + 381 + ], + "lines": [ + { + "bbox": [ + 55, + 349, + 295, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 295, + 381 + ], + "type": "text", + "content": "Figure 3. Ablation on the number of unlabeled video frames in each batch. From 0 to 16, more frames bring better scene-level consistency. (Out of memory for ViT-Large under 32 frames)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 393, + 293, + 456 + ], + "blocks": [ + { + "bbox": [ + 55, + 184, + 555, + 229 + ], + "lines": [ + { + "bbox": [ + 55, + 184, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 55, + 184, + 555, + 229 + ], + "type": "text", + "content": "Table 4. Comparison of depth estimation performance with different training data configurations. The baseline model is trained on labeled Hypersim data only, with subsequent rows showing the cumulative effect of adding different unlabeled datasets (VKITTI2: Virtual KITTI 2, Hypersim-U: treat Hypersim as unlabeled video data). Lower AbsRel " + }, + { + "bbox": [ + 55, + 184, + 555, + 229 + ], + "type": "inline_equation", + "content": "(\\downarrow)" + }, + { + "bbox": [ + 55, + 184, + 555, + 229 + ], + "type": "text", + "content": " and higher " + }, + { + "bbox": [ + 55, + 184, + 555, + 229 + ], + "type": "inline_equation", + "content": "\\delta 1" + }, + { + "bbox": [ + 55, + 184, + 555, + 229 + ], + "type": "inline_equation", + "content": "(\\uparrow)" + }, + { + "bbox": [ + 55, + 184, + 555, + 229 + ], + "type": "text", + "content": " values indicate better performance. For each video segment, we align using only a single scale, setting the shift to zero." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 393, + 293, + 456 + ], + "lines": [ + { + "bbox": [ + 58, + 393, + 293, + 456 + ], + "spans": [ + { + "bbox": [ + 58, + 393, + 293, + 456 + ], + "type": "table", + "html": "
MethodMTD (↓)Video Model?Online?FPS (↑)
DepthCrafter [23]0.003479✓×0.45
Baseline0.003777×✓5.77
StableDepth (Ours)0.003482×✓5.77
", + "image_path": "35cba8b4c6254d0c1318149139c75a716725c491e5b6b02e29846e6f3d3447c0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": "our baseline (metric depth version of DAv2-Large) enables online inference at 5.77 FPS, it shows relatively higher inconsistency (MTD=0.003777). StableDepth significantly improves upon our baseline with a " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "7.8\\%" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": " reduction in MTD (0.003482), achieving comparable consistency to DepthCrafter. Crucially, StableDepth maintains online inference capability, processing frames independently without requiring past or future frames, and operates at 5.77 FPS on a H20 GPU - " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "13 \\times" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": " faster than DepthCrafter. Both methods were evaluated on " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "1920 \\times 1080" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": " resolution images using half-precision (FP16) for inference speed testing. These results demonstrate that StableDepth successfully combines strong temporal consistency with practical advantages in" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 318, + 247, + 395, + 292 + ], + "blocks": [ + { + "bbox": [ + 318, + 247, + 395, + 292 + ], + "lines": [ + { + "bbox": [ + 318, + 247, + 395, + 292 + ], + "spans": [ + { + "bbox": [ + 318, + 247, + 395, + 292 + ], + "type": "image", + "image_path": "277ed3ad43b53cfa9eafb921540613bc0b19ab0efacb43d181812e8b98b914ff.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 292, + 395, + 336 + ], + "blocks": [ + { + "bbox": [ + 318, + 292, + 395, + 336 + ], + "lines": [ + { + "bbox": [ + 318, + 292, + 395, + 336 + ], + "spans": [ + { + "bbox": [ + 318, + 292, + 395, + 336 + ], + "type": "image", + "image_path": "6fbe2d51ca205d8f78e9fda27356074c98e6068528a039780413474603878c8d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 338, + 365, + 347 + ], + "lines": [ + { + "bbox": [ + 348, + 338, + 365, + 347 + ], + "spans": [ + { + "bbox": [ + 348, + 338, + 365, + 347 + ], + "type": "text", + "content": "RGB" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 395, + 247, + 473, + 292 + ], + "blocks": [ + { + "bbox": [ + 395, + 247, + 473, + 292 + ], + "lines": [ + { + "bbox": [ + 395, + 247, + 473, + 292 + ], + "spans": [ + { + "bbox": [ + 395, + 247, + 473, + 292 + ], + "type": "image", + "image_path": "b0b71a1817c0e70862ef88b7e68b82c3c904b8a719798575cd20183c448a8e58.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 395, + 292, + 473, + 336 + ], + "blocks": [ + { + "bbox": [ + 395, + 292, + 473, + 336 + ], + "lines": [ + { + "bbox": [ + 395, + 292, + 473, + 336 + ], + "spans": [ + { + "bbox": [ + 395, + 292, + 473, + 336 + ], + "type": "image", + "image_path": "8b42e1b69ca34ea54d9dec765e23f973b34f040f5bca98ac35f7c497922f82d2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 405, + 338, + 463, + 348 + ], + "lines": [ + { + "bbox": [ + 405, + 338, + 463, + 348 + ], + "spans": [ + { + "bbox": [ + 405, + 338, + 463, + 348 + ], + "type": "text", + "content": "DepthCrafter [23]" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 474, + 247, + 551, + 292 + ], + "blocks": [ + { + "bbox": [ + 474, + 247, + 551, + 292 + ], + "lines": [ + { + "bbox": [ + 474, + 247, + 551, + 292 + ], + "spans": [ + { + "bbox": [ + 474, + 247, + 551, + 292 + ], + "type": "image", + "image_path": "94b3ea5978e55dc36625501f657bb921d531e3db558defe960306e67a32f359e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 474, + 293, + 551, + 336 + ], + "blocks": [ + { + "bbox": [ + 474, + 293, + 551, + 336 + ], + "lines": [ + { + "bbox": [ + 474, + 293, + 551, + 336 + ], + "spans": [ + { + "bbox": [ + 474, + 293, + 551, + 336 + ], + "type": "image", + "image_path": "14f8d1b0091cd4f616700935be367bdb5481ed2c9f9cdfa82d77f0e10049c710.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 338, + 544, + 348 + ], + "lines": [ + { + "bbox": [ + 481, + 338, + 544, + 348 + ], + "spans": [ + { + "bbox": [ + 481, + 338, + 544, + 348 + ], + "type": "text", + "content": "StableDepth (Ours)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 357, + 555, + 423 + ], + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 423 + ], + "type": "text", + "content": "Figure 4. Qualitative comparisons. StableDepth shows improved robustness in handling ambiguous regions such as water surfaces and sky boundaries while maintaining accurate depth estimation for the main subjects (flamingo, horse) compared to DepthCrafter. The highlighted region in DepthCrafter results shows inconsistent depth predictions in these challenging regions." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 315, + 434, + 552, + 519 + ], + "blocks": [ + { + "bbox": [ + 55, + 464, + 295, + 541 + ], + "lines": [ + { + "bbox": [ + 55, + 464, + 295, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 295, + 541 + ], + "type": "text", + "content": "Table 5. Performance comparison of different depth estimation methods on DAVIS dataset. MTD: Motion-aware Temporal Difference (Eq. (9)). Online: whether the method can process data without requiring the entire video. FPS: inference speed in frames per second. StableDepth achieves comparable consistency to DepthCrafter that based on heavy SVD [3], while maintaining online capability and significantly faster inference." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 434, + 552, + 519 + ], + "lines": [ + { + "bbox": [ + 315, + 434, + 552, + 519 + ], + "spans": [ + { + "bbox": [ + 315, + 434, + 552, + 519 + ], + "type": "table", + "html": "
DataPretrainedBonnScanNet
AbsRel (↓)δ1(↑)AbsRel (↓)δ1(↑)
LDINOv20.0860.9290.0980.911
LDAv20.0710.9460.1020.903
L + UDINOv20.0680.9490.0990.909
L + UDAv20.0570.9760.0970.913
", + "image_path": "46c2bf6a72ad81905606bdab6b168807f9b3322b356a81706caa0fcb611a2bbd.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 522, + 555, + 567 + ], + "lines": [ + { + "bbox": [ + 313, + 522, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 555, + 567 + ], + "type": "text", + "content": "Table 6. Performance comparison on Bonn and ScanNet. We compare different initialization strategies: DINOv2 pre-trained weights vs. DAv2's well-trained parameters. L: labeled data only; L+U: both labeled and unlabeled data with dual-decoder." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 580, + 553, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 553, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 553, + 604 + ], + "type": "text", + "content": "deployment, addressing a key limitation in current depth estimation approaches." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "text", + "content": "Comparison of unlabeled video frames. As shown in Fig. 3, MTD scores improve steadily with more unlabeled video frames in each mini-batch and gradually converge at 16 frames, with negligible gains beyond this point, supporting our choice of 16-frame clips as an optimal trade-off." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "content": "Comparison of different pre-trained encoders. Tab. 6 shows the effectiveness of our semi-supervised training strategy on the Bonn and ScanNet datasets. We compare two initialization strategies: DINOv2 and DAv2 pre-trained" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7075" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 70, + 547, + 351 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 547, + 351 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 547, + 351 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 547, + 351 + ], + "type": "image", + "image_path": "bb0ac1a715090252d9205b98f030c4d49333ce8d6b00bda453f229ff5b994036.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 361, + 555, + 396 + ], + "lines": [ + { + "bbox": [ + 55, + 361, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 555, + 396 + ], + "type": "text", + "content": "Figure 5. Qualitative visualizations of depth predictions across diverse scenarios. Each row pair shows an input sequence and corresponding depth predictions from our method. Our approach maintains consistent depth estimation while preserving fine geometric details across challenging cases including dynamic motions (dancer), reflective surfaces (water), and large objects (rhinoceros)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 416, + 296, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 416, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 416, + 296, + 513 + ], + "type": "text", + "content": "weights. With only labeled data (L), DAv2 initialization outperforms DINOv2, particularly on Bonn (0.071 vs 0.086 AbsRel). Incorporating unlabeled video data " + }, + { + "bbox": [ + 55, + 416, + 296, + 513 + ], + "type": "inline_equation", + "content": "(\\mathrm{L} + \\mathrm{U})" + }, + { + "bbox": [ + 55, + 416, + 296, + 513 + ], + "type": "text", + "content": " through our dual-decoder approach yields consistent improvements across both datasets. DAv2 with " + }, + { + "bbox": [ + 55, + 416, + 296, + 513 + ], + "type": "inline_equation", + "content": "\\mathrm{L} + \\mathrm{U}" + }, + { + "bbox": [ + 55, + 416, + 296, + 513 + ], + "type": "text", + "content": " achieves the best performance, reducing AbsRel to 0.057 on Bonn and 0.097 on ScanNet, demonstrating the effectiveness of our semi-supervised approach." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 526, + 167, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 167, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 167, + 538 + ], + "type": "text", + "content": "4.4. Qualitative Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "content": "Our method demonstrates robust performance across diverse real-world scenarios. As shown in Fig. 1, it generates stable depth maps even in challenging conditions, capturing dynamic motions in dog agility courses, street performances, and natural scenes while maintaining consistent geometry frame-by-frame. Fig. 4 shows our method's superiority over DepthCrafter in ambiguous depth cue cases. In the flamingo scene, DepthCrafter produces inconsistent depth predictions for water reflections, while our method maintains coherent depth across the water surface. In the horse sequence, DepthCrafter mistakenly estimates the sky as foreground, whereas our approach handles the background-sky transition more accurately. These results highlight our model's robustness to common depth estima" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 416, + 555, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 416, + 555, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 555, + 537 + ], + "type": "text", + "content": "tion challenges like reflective surfaces and infinite-depth regions, while preserving scene consistency. Fig. 5 further demonstrates the effectiveness of our method in challenging scenarios, including fast-moving subjects (dancer) and large objects (rinoceros). Our approach consistently preserves fine geometric details and maintains stable depth relationships across frames, even under complex lighting conditions and dynamic motions. The temporal smoothness achieved enables reliable depth estimation for practical applications requiring consistent 3D understanding." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 547, + 388, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 547, + 388, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 547, + 388, + 559 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 567, + 555, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 555, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 555, + 711 + ], + "type": "text", + "content": "We present StableDepth, a framework that bridges the gap between relative and metric depth estimation through our scene-consistent and scale-invariant (SCSI) paradigm. By leveraging our dual-decoder architecture and video diffusion baking strategy, we effectively decouple depth accuracy from temporal consistency learning, while harnessing knowledge from unlabeled video data. StableDepth enables online inference at " + }, + { + "bbox": [ + 313, + 567, + 555, + 711 + ], + "type": "inline_equation", + "content": "13 \\times" + }, + { + "bbox": [ + 313, + 567, + 555, + 711 + ], + "type": "text", + "content": " faster speed than previous video-based methods while maintaining comparable temporal stability. This approach advances monocular depth estimation toward more practical applications in 3D reconstruction, robotics, and augmented reality." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7076" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 122 + ], + "type": "text", + "content": "Acknowledgements. This work is supported by the National Natural Science Foundation of China (No. 62422606, 62201484) and DAMO Academy Research Intern Program." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 143, + 115, + 155 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 143, + 115, + 155 + ], + "spans": [ + { + "bbox": [ + 56, + 143, + 115, + 155 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 163, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 163, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 163, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 61, + 163, + 294, + 205 + ], + "type": "text", + "content": "[1] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv:2302.12288, 2023.5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 208, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 208, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 62, + 208, + 294, + 239 + ], + "type": "text", + "content": "[2] Reiner Birkl, Diana Wofk, and Matthias Muller. Midas v3. 1-a model zoo for robust monocular relative depth estimation. arXiv:2307.14460, 2023. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 242, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 242, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 62, + 242, + 294, + 296 + ], + "type": "text", + "content": "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv:2311.15127, 2023. 2, 5, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 298, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 298, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 62, + 298, + 294, + 341 + ], + "type": "text", + "content": "[4] Aleksei Bochkovskii, Amaël Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. In ICLR, 2025. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 342, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 294, + 374 + ], + "type": "text", + "content": "[5] D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black. A naturalistic open source movie for optical flow evaluation. In ECCV, 2012. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 376, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 376, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 62, + 376, + 294, + 398 + ], + "type": "text", + "content": "[6] Yohann Cabon, Naila Murray, and Martin Humenberger. Virtual katii 2. arXiv:2001.10773, 2020. 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 399, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 294, + 442 + ], + "type": "text", + "content": "[7] Sili Chen, Hengkai Guo, Shengnan Zhu, Feihu Zhang, Zi long Huang, Jiashi Feng, and Bingyi Kang. Video depth anything: Consistent depth estimation for super-long videos. In CVPR, 2025. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 445, + 294, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 294, + 466 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 294, + 466 + ], + "type": "text", + "content": "[8] Weifeng Chen, Zhao Fu, Dawei Yang, and Jia Deng. Single-image depth perception in the wild. In NeurIPS, 2016. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 468, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 468, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 468, + 294, + 510 + ], + "type": "text", + "content": "[9] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 512, + 294, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 512, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 57, + 512, + 294, + 576 + ], + "type": "text", + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 4, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 579, + 294, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 579, + 294, + 611 + ], + "spans": [ + { + "bbox": [ + 57, + 579, + 294, + 611 + ], + "type": "text", + "content": "[11] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In ICCV, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 613, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 294, + 644 + ], + "type": "text", + "content": "[12] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. In NeurIPS, 2014. 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "type": "text", + "content": "[13] Fatima El Jamiy and Ronald Marsh. Survey on depth perception in head mounted displays: distance estimation in virtual reality, augmented reality, and mixed reality. IET Image Processing, 2019. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "text", + "content": "[14] Gunnar Farnebäck. Two-frame motion estimation based on polynomial expansion. In SCIA, 2003. 5" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[15] Xiao Fu, Wei Yin, Mu Hu, Kaixuan Wang, Yuexin Ma, Ping Tan, Shaojie Shen, Dahua Lin, and Xiaoxiao Long. Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image. In ECCV, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 118, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 149 + ], + "type": "text", + "content": "[16] Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as proxy for multi-object tracking analysis. In CVPR, 2016. 5, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 151, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 553, + 182 + ], + "type": "text", + "content": "[17] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. *IJRR*, 2013. 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 184, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 553, + 237 + ], + "type": "text", + "content": "[18] Ming Gui, Johannes S Fischer, Ulrich Prestel, Pingchuan Ma, Dmytro Kotovenko, Olga Grebenkova, Stefan Andreas Baumann, Vincent Tao Hu, and Björn Ommer. Depthfm: Fast monocular depth estimation with flow matching. In AAAI, 2025. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 239, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 271 + ], + "type": "text", + "content": "[19] Vitor Guizilini, Igor Vasiljevic, Dian Chen, Rares Ambrus, and Adrien Gaidon. Towards zero-shot scale-aware monocular depth estimation. In ICCV, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 273, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 273, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 273, + 553, + 304 + ], + "type": "text", + "content": "[20] S Mahdi H. Miangoleh, Mahesh Reddy, and Yaqiz Aksoy. Scale-invariant monocular depth estimation via ssi depth. In SIGGRAPH, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 305, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 337 + ], + "type": "text", + "content": "[21] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 338, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 553, + 392 + ], + "type": "text", + "content": "[22] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. TPAMI, 2024. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 394, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 436 + ], + "type": "text", + "content": "[23] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. In CVPR, 2025. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 438, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 469 + ], + "type": "text", + "content": "[24] Yaosi Hu, Chong Luo, and Zhenzhong Chen. Make it move: controllable image-to-video generation with text descriptions. In CVPR, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "type": "text", + "content": "[25] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In CVPR, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 515, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 515, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 553, + 557 + ], + "type": "text", + "content": "[26] Bingxin Ke, Dominik Narnhofer, Shengyu Huang, Lei Ke, Torben Peters, Katerina Fragkiadaki, Anton Obukhov, and Konrad Schindler. Video depth without video models. In CVPR, 2025. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 559, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 580 + ], + "type": "text", + "content": "[27] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In CVPR, 2021. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "type": "text", + "content": "[28] Yixuan Li, Lihan Jiang, Linning Xu, Yuanbo Xiangli, Zhenzhi Wang, Dahua Lin, and Bo Dai. Matrixcity: A large-scale city dataset for city-scale neural rendering and beyond. In ICCV, 2023. 5, 6, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "type": "text", + "content": "[29] Zhaoshuo Li, Wei Ye, Dilin Wang, Francis X Creighton, Russell H Taylor, Ganesh Venkatesh, and Mathias Unberath. Temporally consistent online depth estimation in dynamic scenes. In WACV, 2023. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[30] LightwheelAI and LightwheelOcc contributors. LightwheelOcc: A 3d occupancy synthetic dataset in autonomous driving. https://github.com/OpenDriveLab/LightwheelOcc, 2024.5,6,7" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7077" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[31] Nian Liu, Ni Zhang, Ling Shao, and Junwei Han. Learning selective mutual attention and contrast for rgb-d saliency detection. TPAMI, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "text", + "content": "[32] Xingtong Liu, Ayushi Sinha, Masaru Ishii, Gregory D Hager, Austin Reiter, Russell H Taylor, and Mathias Unberath. Dense depth estimation in monocular endoscopy with self-supervised learning methods. TMI, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 152, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 152, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 57, + 152, + 294, + 174 + ], + "type": "text", + "content": "[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 175, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 175, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 57, + 175, + 294, + 207 + ], + "type": "text", + "content": "[34] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. In SIGGRAPH, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "type": "text", + "content": "[35] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. TMLR, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "type": "text", + "content": "[36] E. Palazzolo, J. Behley, P. Lottes, P. Giguère, and C. Stachniss. ReFusion: 3D Reconstruction in Dynamic Environments for RGB-D Cameras Exploiting Residuals. In IROS, 2019. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 309, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 352 + ], + "type": "text", + "content": "[37] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Unidepth: Universal monocular metric depth estimation. In CVPR, 2024. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 354, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 398 + ], + "type": "text", + "content": "[38] Luigi Piccinelli, Christos Sakaridis, Yung-Hsu Yang, Mattia Segu, Siyuan Li, Wim Abbeloos, and Luc Van Gool. Unidepthv2: Universal monocular metric depth estimation made simpler. arXiv:2502.20110, 2025. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 399, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 442 + ], + "type": "text", + "content": "[39] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. TPAMI, 2020. 2, 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 445, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 475 + ], + "type": "text", + "content": "[40] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, 2021. 3, 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 478, + 294, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 531 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 531 + ], + "type": "text", + "content": "[41] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. In ICLR, 2025. 5, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "type": "text", + "content": "[42] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In ICCV, 2021. 3, 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 590, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 622 + ], + "type": "text", + "content": "[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 624, + 294, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 655 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 655 + ], + "type": "text", + "content": "[44] Abhinav Sagar. Monocular depth estimation using multi-scale neural network and feature fusion. In WACV, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "type": "text", + "content": "[45] Ashutosh Saxena, Min Sun, and Andrew Y Ng. Make3d: Learning 3d scene structure from a single still image. TPAMI, 2008. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[46] Mohamed Sayed, John Gibson, Jamie Watson, Victor Prisacariu, Michael Firman, and Clément Godard. Simplere" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 701 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "text", + "content": "con: 3d reconstruction without 3d convolutions. In ECCV, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[47] Jiahao Shao, Yuanbo Yang, Hongyu Zhou, Youmin Zhang, Yujun Shen, Matteo Poggi, and Yiyi Liao. Learning temporally consistent video depth from video diffusion priors. 2025. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "type": "text", + "content": "[48] Chang Shu, Kun Yu, Zhixiang Duan, and Kuiyuan Yang. Feature-metric loss for self-supervised learning of depth and egomotion. In ECCV, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "type": "text", + "content": "[49] Chaoyang Wang, Simon Lucey, Federico Perazzi, and Oliver Wang. Web stereo video supervision for depth prediction from dynamic scenes. In 3DV, 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 209, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 251 + ], + "type": "text", + "content": "[50] Lijun Wang, Yifan Wang, Linzhao Wang, Yunlong Zhan, Ying Wang, and Huchuan Lu. Can scale-consistent monocular depth be learned in a self-supervised scale-invariant manner? In ICCV, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "type": "text", + "content": "[51] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In CVPR, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 299, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 299, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 553, + 331 + ], + "type": "text", + "content": "[52] Yiran Wang, Min Shi, Jiaqi Li, Zihao Huang, Zhiguo Cao, Jianming Zhang, Ke Xian, and Guosheng Lin. Neural video depth stabilizer. In ICCV, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 333, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 333, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 333, + 553, + 374 + ], + "type": "text", + "content": "[53] Jamie Watson, Mohamed Sayed, Zawar Qureshi, Gabriel J Brostow, Sara Vicente, Oisin Mac Aodha, and Michael Firman. Virtual occlusions through implicit depth. In CVPR, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 377, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 553, + 409 + ], + "type": "text", + "content": "[54] Diana Wofk, Fangchang Ma, Tien-Ju Yang, Sertac Karaman, and Vivienne Sze. Fastdepth: Fast monocular depth estimation on embedded systems. In ICRA, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 411, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 443 + ], + "type": "text", + "content": "[55] Ke Xian, Jianming Zhang, Oliver Wang, Long Mai, Zhe Lin, and Zhiguo Cao. Structure-guided ranking loss for single image depth prediction. In CVPR, 2020. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 445, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 487 + ], + "type": "text", + "content": "[56] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 2, 3, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 490, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 553, + 521 + ], + "type": "text", + "content": "[57] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. In NeurIPS, 2024. 3, 4, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 524, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 524, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 524, + 553, + 567 + ], + "type": "text", + "content": "[58] Rajeev Yasarla, Hong Cai, Jisoo Jeong, Yunxiao Shi, Risheek Garrepalli, and Fatih Porikli. Mamo: Leveraging memory and attention for monocular video depth estimation. In ICCV, 2023. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 569, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 569, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 569, + 553, + 611 + ], + "type": "text", + "content": "[59] Wei Yin, Chi Zhang, Hao Chen, Zhipeng Cai, Gang Yu, Kaixuan Wang, Xiaozhi Chen, and Chunhua Shen. Metric3d: Towards zero-shot metric 3d prediction from a single image. In ICCV, 2023. 2, 3, 5, 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 613, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 655 + ], + "type": "text", + "content": "[60] Sen Zhang, Jing Zhang, and Dacheng Tao. Towards scale-aware, robust, and generalizable unsupervised monocular depth estimation by integrating imu motion dynamics. In ECCV, 2022. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 658, + 553, + 701 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 701 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 701 + ], + "type": "text", + "content": "[61] Ruijie Zhu, Chuxin Wang, Ziyang Song, Li Liu, Tianzhu Zhang, and Yongdong Zhang. Scaledepth: Decomposing metric depth estimation into scale prediction and relative depth estimation. arXiv:2407.08187, 2024. 3" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7078" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Staining and Locking Computer Vision Models Without Retraining/6d43a5d3-0678-494f-aae6-04c44e76e1c5_content_list.json b/2025/Staining and Locking Computer Vision Models Without Retraining/6d43a5d3-0678-494f-aae6-04c44e76e1c5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..148d6e4b11cdfd855997219cb2e1b77b0e9b5f87 --- /dev/null +++ b/2025/Staining and Locking Computer Vision Models Without Retraining/6d43a5d3-0678-494f-aae6-04c44e76e1c5_content_list.json @@ -0,0 +1,1704 @@ +[ + { + "type": "text", + "text": "Staining and locking computer vision models without retraining", + "text_level": 1, + "bbox": [ + 176, + 130, + 821, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Oliver J. Sutton*", + "bbox": [ + 228, + 181, + 364, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Synoptix Ltd. and", + "bbox": [ + 220, + 199, + 367, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "King's College London", + "bbox": [ + 200, + 217, + 387, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qinghua Zhou*", + "bbox": [ + 486, + 181, + 611, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "King's College London", + "bbox": [ + 452, + 199, + 638, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "qinghua.zhou@kcl.ac.uk", + "bbox": [ + 447, + 219, + 643, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "George Leete", + "bbox": [ + 694, + 181, + 805, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Synoptix Ltd.", + "bbox": [ + 694, + 199, + 805, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "oliver.sutton@kcl.ac.uk", + "bbox": [ + 189, + 237, + 395, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Alexander N. Gorban", + "bbox": [ + 116, + 258, + 295, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Central Univ. Moscow", + "bbox": [ + 116, + 277, + 297, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ivan Y. Tyukin \nKing's College London and AI Center, Bol'shoi Bul'var 30, Moscow \nand ISP RAS Research Center for Trusted Artificial Intelligence", + "bbox": [ + 334, + 258, + 879, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 345, + 326, + 363 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce new methods of staining and locking computer vision models, to protect their owners' intellectual property. Staining, also known as watermarking, embeds secret behaviour into a model which can later be used to identify it, while locking aims to make a model unusable unless a secret trigger is inserted into input images. Unlike existing methods, our algorithms can be used to stain and lock pretrained models without requiring fine-tuning or retraining, and come with provable, computable guarantees bounding their worst-case false positive rates. The stain and lock are implemented by directly modifying a small number of the model's weights and have minimal impact on the (unlocked) model's performance. Locked models are unlocked by inserting a small 'trigger patch' into the corner of the input image. We present experimental results showing the efficacy of our methods and demonstrating their practical performance on a variety of computer vision models.", + "bbox": [ + 89, + 378, + 485, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 666, + 220, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Designing, training and validating a computer vision model can be an expensive undertaking. The effect on an organisation if such a valuable investment is leaked or stolen, can be deleterious, particularly since model weights can be difficult for a human judge to distinguish confidently. Developing methods for protecting such intellectual property, which may be deployed on a vulnerable edge device, is therefore crucial.", + "bbox": [ + 89, + 691, + 483, + 797 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Two natural approaches that a model's owner may wish to take are either to imprint an identifiable $\\text{stain}$ within its weights, or to add a lock to make the model unusable without adding a separate secret unlocking trigger to the model's input. Staining, also known as watermarking in the liter-", + "bbox": [ + 89, + 799, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ature [5, 23], can be viewed as a form of 'copyright trap'. By embedding specific (unnatural) behavioural quirks into the model, an identifying fingerprint is produced which can later be used to determine the provenance of a stolen or copied model. Locking, on the other hand, has a more direct aim [6, 14, 16]: unless the thief knows the secret trigger, they are unable to use the model to its full capability.", + "bbox": [ + 511, + 347, + 906, + 455 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Here, we present new methods for both staining and locking neural network models in computer vision. Unlike existing approaches in the literature (reviewed in Sec. 2), our algorithms do not require model training or fine-tuning. They directly modify the model's weights, and can therefore be used to stain or lock pre-trained models with minimal performance impact. Since re-training is not required, a single model can be supported and updated across multiple customers, while providing each with a separate stain or lock. A further key advantage of our staining and locking techniques is that implementing them does not require access to training or validation data. We also provide two distinct novel theoretical guarantees on the worst-case false positive rate of our staining and locking algorithms, analogues of which do not appear to be present in the literature.", + "bbox": [ + 511, + 457, + 908, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Staining methods are broadly composed of two parts: a schema describing the information encoded in the stain and how it may later be read, and a mechanism through which this information is implanted into the model. Numerous schemas have been proposed in the literature, e.g. hiding a message in the model's weights, its activations or its class predictions for specific trigger inputs. Until now, however, only two families of mechanisms have been available, both requiring model retraining: through special loss functions [7, 8, 11, 28, 45, 46], or by manipulating the training data [1, 17, 26, 32]. Our novel training-free staining mechanism implants highly selective detector neurons into the model, which produce a strong output only when a special trigger is present in the model's input. Sec. 3.1 shows how several standard staining", + "bbox": [ + 511, + 688, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally and are listed alphabetically.", + "bbox": [ + 107, + 886, + 447, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2346", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "schemas may be implemented using this mechanism.", + "bbox": [ + 89, + 90, + 439, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our lock may be seen as an 'active' variant of the stain mechanism. Additional disruptors are also implanted to pollute the model's activations and prevent it from performing well. A specially constructed 'trigger patch' is added to one corner of the input image during inference is detected by the detector neuron. The detector neuron then produces a signal which deactivates the disruptors, recovering the full model performance. Examples of such trigger patches are shown in Appendix B. Our locking methods are presented with a focus on their fundamental building blocks, which can then be used to design bespoke algorithms to lock specific models.", + "bbox": [ + 89, + 108, + 480, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our goal is to present these new techniques as clearly as possible in a simple setting. For this, we focus primarily on convolutional models, although extensions of the core algorithms to generative adversarial networks (GANs) and vision transformers (ViTs) are presented in Appendix E. Our staining methods are generic and apply directly to many different model architectures. Our basic lock is more architecture-dependent since it interacts with the fundamental action of the model. To provide architectural-independence, we also show how the lock may be embedded within a Squeeze-and-Excite block [21], a common component of many models. We demonstrate, in Sec. 6, that such a block can be added to many pre-trained models with minimal impact on performance. For simplicity and brevity, we do not focus too strongly on techniques for hiding and obfuscating the stains and locks; we provide ideas for achieving this throughout, but do not pursue it in detail to present the core concepts as clearly as possible in the space available. For practical success, we also feel it is preferable for a diversity of obfuscation techniques and algorithmic variations to proliferate. Instead, we focus on the simplest case where a single detector and disruptor are added to the model; in practice, multiple detectors should be combined to deactivate multiple disruptors. This straightforward extension of the methods presented here makes reverse-engineering the lock a significant challenge.", + "bbox": [ + 91, + 292, + 483, + 670 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approaches to staining and locking are inspired by stealth attacks [41, 43, 44], which have recently been shown to enable an attacker to specify the output of models from many modern families for a given piece of input data by directly modifying just a few weights. By harnessing this core sensitivity of modern models, and exploiting the fundamental concentration properties of their feature spaces, the highly selective detector neuron of our stain and lock can be implanted without retraining the model. This detector forms the heart of our methods, around which completely new mechanisms are built to facilitate staining and locking. This link means that the novel theoretical guarantees we present in Sec. 5 can be adapted back to study stealth attacks, and complement those available in the literature [41, 43, 44].", + "bbox": [ + 89, + 672, + 482, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The paper is organised as follows. Sec. 2 places our", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "approach in the context of related work. Our staining and locking algorithms are presented in Sec. 3 and Sec. 4. Sec. 5 presents novel theoretical guarantees on the detector's false positive rate. Experimental results demonstrate the practical performance of our algorithms in Sec. 6. A discussion of our findings is in Sec. 7, and some final conclusions are given in Sec. 8. Our mathematical notation is summarised in Appendix A (in the supplementary material).", + "bbox": [ + 511, + 90, + 903, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 511, + 232, + 651, + 247 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Staining. Several staining methods have been proposed in the literature and mainly fall within two thematic groups; see [5, 23] for a thorough review. Loss-based methods encode a secret message into the model's weights during training using a special loss function. This message may then be revealed using a special trigger input to the network. Common approaches encode the message as a binary string in the model's weights [7, 45], activations [28] or outputs [8, 11], either directly or to be decoded by a special model [10, 46]. Backdoor methods, on the other hand, manipulate training data to produce a unique behavioural signature. One method is to force the model to learn specific non-natural labels for certain distinctive inputs [1, 15, 25, 32]. Alternatively, other work has considered the task of training the model to learn to associate a specific class with certain deliberate changes to natural images [17], such as the presence of a logo [26].", + "bbox": [ + 511, + 258, + 906, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Locking. Two key families of approaches have been proposed in the literature. Intrinsic methods embed the lock into the model itself during training. This may be by training the model's output to depend on an additional (secret) input [14, 42], a cryptographic scheme [33], a secret global mask applied to inputs [6], or by converting training and inference images using a 'style-transfer' model (e.g. to oil paintings) [16]. Extrinsic methods offer additional orthogonal security to intrinsic methods via techniques from outside the neural network domain, such as encryption [2] or secure infrastructure [35, 40].", + "bbox": [ + 511, + 503, + 908, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Novelty. In contrast to those in the literature, our staining and locking algorithms directly modify a model's weights, so require no model training, fine-tuning or external infrastructure. This is a key advantage because a single base model can be supplied to many clients with different locks/stains, without re-training each time. Moreover, re-training a model carries the inherent risk of changing its behaviour in unidentifiable ways (especially if the model is tuned to give incorrect responses to natural images as in backdoor methods), which is avoided through our direct approach. Being embedded in the model itself, a lock also remains active even if an adversary gains access to the hardware on which it runs. Our provable performance guarantees further differentiate our methods; comparable results are not available for any of the methods in the literature.", + "bbox": [ + 511, + 674, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2347", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Staining", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section presents the building blocks of our novel training-free staining mechanism. The stain implants a highly selective detector neuron into the model, which is activated by a specially optimised trigger input to the model. We show how this mechanism can be used to implement standard staining schemas in Sec. 3.1.", + "bbox": [ + 89, + 116, + 483, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multilayer perceptrons. Algorithm 1 outlines our mechanism for staining a generic multilayer perceptron neural network. A detector neuron with a weight vector sampled from $\\mathcal{U}(\\mathbb{S}^{m - 1})$ (i.e. uniformly from the surface of a sphere) is inserted into the model. Concentration properties in the model's feature space ensure that this randomly sampled neuron is very unlikely to respond strongly to natural inputs (see Sec. 5). It is possible, however, to optimise a trigger input (e.g. using gradient descent) to strongly activate the detector. The strong response to the trigger input, therefore, becomes an identifying characteristic of the model and can be used as the basis for implementing standard staining schemas. To ensure that the detector neuron has no impact on the model's performance, the parameters $\\delta$ and $\\Delta$ can be used in conjunction with the activation function to ensure that the detector's response is zero for natural inputs.", + "bbox": [ + 89, + 210, + 483, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The optimisation of the trigger input $x^{*}$ is important here: a strong detector response to $x^{*}$ is very distinct from its response to natural inputs, and the guarantees on the false positive rate are improved (see Theorem 1). It also means the response $\\Delta$ can be achieved by scaling down the weight vector, which can be beneficial for evading detection.", + "bbox": [ + 89, + 453, + 483, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Additive staining. Alternatively, an additive stain adds the detector weight to an existing neuron's weight. Once again, the detector component of the weight likely only responds weakly to natural inputs, meaning that it has little impact on the model's operation. The strong response to the trigger input, however, is still preserved. Since this method does not produce a 'silent' detector neuron, it is much less likely that the detector will be identified and removed by an adversary.", + "bbox": [ + 89, + 547, + 483, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Convolutional networks. Algorithm 2 demonstrates how this framework applies to a convolutional neural network. For simplicity, this algorithm focuses on the non-additive stain, but also applies to models incorporating batch normalisation. A single convolutional kernel is used as the detector, and a reduction operator $r$ is used for the optimisation to convert the kernel's activation map into a single scalar response value. This could, for example, be the response of the kernel at a particular position in the image, or its average response. Larger kernels are helpful in reducing the false positive rate, as made explicit by Theorem 1: because a $\\kappa \\times \\kappa$ convolutional kernel with $c$ input channels has $\\kappa^2 c$ parameters, large kernels operate in a higher dimensional setting.", + "bbox": [ + 89, + 671, + 483, + 867 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Extensions. Extensions to apply these staining algorithms to GANs and ViTs are presented in Section E.", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Multilayer perceptron staining" + ], + "code_body": "Input:Trained multilayer perceptron network $\\mathcal{N}$ Index $k$ of neuron to stain in layer $j$ Detector response to trigger $\\Delta \\in \\mathbb{R}$ Desired non-trigger detector response $\\delta \\in \\mathbb{R}$ Whether the stain should be additive 1 Define the map $\\phi$ from network inputs to inputs to layer $j$ 2 Let $W_{j}\\in \\mathbb{R}^{n\\times m}$ $b_{j}\\in \\mathbb{R}^{n}$ be the layer $j$ weight and bias. 3 Sample a detector weight vector $v\\sim \\mathcal{U}(\\mathbb{S}^{m - 1})$ 4 Optimise the trigger input $x^{*}\\in \\arg \\max_{z\\in S}v\\cdot \\phi (z)$ 5 if stain is additive then 6 Let $\\beta$ be entry $k$ of $b_{j}$ and $w$ be row $k$ of $W_{j}$ 7 Define $u = w + \\frac{\\Delta - \\beta - w\\cdot\\phi(x^{*})}{v\\cdot\\phi(x^{*})} v.$ 8 else 9 Define $\\beta = \\delta$ and $u = \\frac{\\Delta - \\beta}{v\\cdot\\phi(x^*)} v.$ 10 Replace row $k$ of $W_{j}$ with $u$ and entry $k$ of $b_{j}$ with $\\beta$ Output: Stained model $\\mathcal{N}$ and trigger input $x^{*}$", + "bbox": [ + 516, + 111, + 903, + 345 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 2: Convolutional network staining" + ], + "code_body": "Input:Trained convolutional network $\\mathcal{N}$ Parameters $(k,j,\\Delta ,\\delta)$ as in Alg.1 Matrix to scalar reduction operation $r$ Whether layer $j$ performs batch normalisation \n1 Define the map $\\phi :S = \\mathbb{R}^{3\\times m_0\\times n_0}\\to \\mathbb{R}^{c_j\\times m_j\\times n_j}$ from network inputs to inputs to layer $j$ \n2 Let $W_{j}\\in \\mathbb{R}^{c_{j} + 1\\times c_{j}\\times \\kappa_{j}\\times \\kappa_{j}}$ be the layer $j$ weight, with kernel shape $\\kappa_{j}\\times \\kappa_{j}$ \n3 Sample a kernel $v\\in \\mathbb{R}^{c_j\\times \\kappa_j\\times \\kappa_j}$ from $\\mathcal{U}(\\mathbb{S}^{c_j\\kappa_j^2 -1})$ viewed as a tensor with shape $(c_{j},\\kappa_{j},\\kappa_{j})$ \n4 Optimise the trigger image $x^{*}\\in \\arg \\max_{z\\in S}r(v*\\phi (z))$ \n5 if layer $j$ performs batch normalisation then \n6 Let $\\mu_j,\\sigma_j^2,w_j,b_j\\in \\mathbb{R}^{c_j + 1}$ be the mean, variance, weight and bias of the layer $j$ batch normalisation. \n7 Define $\\alpha = \\frac{(\\Delta - \\delta)(\\sigma_j)_k}{(w_j)_kr(v*\\phi(x^*))}$ and $\\beta = \\delta +\\frac{(w_j)_k(\\mu_j)_k}{(\\sigma_j)_k}$ \n8 else \n9 Let $b_{j}\\in \\mathbb{R}^{c_{j} + 1}$ be the bias of layer $j$ \n10 Define $\\alpha = \\frac{(\\Delta - \\delta)}{r(v*\\phi(x^*))}$ and $\\beta = \\delta$ \n11 Replace $(W_{j})_{k,\\dots ,}$ with $\\alpha v$ ,and entry $k$ of $b_{j}$ with $\\beta$ Output: Stained model $\\mathcal{N}$ and trigger input $x^{*}$", + "bbox": [ + 516, + 386, + 903, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Implementing standard staining schemas", + "text_level": 1, + "bbox": [ + 511, + 710, + 864, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The algorithms above introduce a novel training-free staining mechanism, and here we briefly discuss how it can be used to implement standard schemas from the literature.", + "bbox": [ + 511, + 732, + 903, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Weight staining. Originally proposed by [7, 45], this schema encodes a secret message as a binary string, which is then implanted into the model as the signs of the weights of a target neuron (e.g. $-$ encodes 0 and $+$ encodes 1). The stain can be checked by decoding the message from the signs of the neuron's weight vector's components. This can be embedded into a non-additive stain by sampling the detector vector $v$ from the orthant of the sphere with the specified", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2348", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "component signs, and the rest of the staining algorithm can then proceed as usual. This restriction on the sampling may increase the false positive rate of the detector (by decreasing the entropy of the distribution), but this should not present a problem if the dimension is sufficiently high.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Activation staining. This family of schemas, introduced in [28], instead embeds the secret message into the signs of the model's activations for a target input, rather than the weights. Our additive stain can be used to directly implement this. Suppose a detector vector is sampled for a given layer of the model, and a trigger input is optimised to produce a strong response to this detector vector. Then, this detector vector is weighted and added to each neuron in the layer, to ensure that the response to the trigger image from that neuron has the specified sign.", + "bbox": [ + 89, + 169, + 482, + 319 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Output staining. Another standard schema family specifies (non-natural) class predictions for certain trigger input images [1, 8, 11, 32]. The stain can, therefore, be verified without access to model weights or activations. This can be implemented post-training by starting with a standard non-additive stain from Algorithm 1. If the stain detector neuron is implanted into row $k$ of the weight matrix in layer $j$ , then the detector's response (which will be zero for non-trigger inputs after a ReLU activation if $\\delta \\ll 0$ ) will multiply the entries of column $k$ in layer $j + 1$ . For any chosen output class $C$ , it is typically possible to find a vector (e.g. using gradient descent) which, if it were the output of layer $j + 1$ , would cause the model to confidently predict class $C$ . Replacing column $k$ of layer $j + 1$ with this vector, therefore, causes the model to predict the chosen class for the trigger input; the selectivity of the detector ensures that the new column will not otherwise affect the model's performance. This technique extends analogously to convolutional stains.", + "bbox": [ + 89, + 323, + 486, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Locking", + "text_level": 1, + "bbox": [ + 89, + 614, + 187, + 631 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our locking methods build on the staining mechanism by inserting additional disruptors into the model to pollute its latent activations. The signal from the stain's detector neuron is used to disable the disruptors when the trigger is present in the input. Fig. 8 and Fig. 9 (in the supplementary material) show examples of images containing these patches. We show how the underlying structure of a convolutional detector can be exploited to produce a transferable trigger 'patch' which can be inserted into any image to unlock the full model performance. We present two variants of the locking mechanism. Internal locking embeds the lock into the weights of convolutional networks with a particular structure, while squeeze-and-excite locking uses a standard Squeeze-and-Excite block [21], common in computer vision models, as the foundation of a more generic procedure which can be added to any convolutional model.", + "bbox": [ + 89, + 641, + 483, + 881 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Internal locking. Algorithm 3 adds an internal lock to a", + "bbox": [ + 89, + 885, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "convolutional neural network. To present the idea in the simplest possible setting, we suppose that the model consists of a sequence of convolutional layers with biases (batch normalisation can be handled as in Algorithm 2), ReLU activations, followed by a single dense logits layer. For brevity, we do not explore here how the many additional model components present in practice may alter this algorithm: while they complicate the implementation, they do not change the fundamental concept. The algorithm consists of three main steps: implanting the detector and building the trigger (lines Line 1-4), propagating the detector signal through the model (lines 5-10), and building the disruptor (lines 11-14).", + "bbox": [ + 511, + 90, + 908, + 273 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 3: Convolutional network internal lock" + ], + "code_body": "Input : Parameters $(\\mathcal{N},k,j,\\Delta >0,\\delta < 0)$ as in Alg.2 Image coordinates $(a,b)$ for the trigger patch Disruption scale $s > 0$ and offset $t\\in \\mathbb{R}^c$ \n1 Define the map $\\phi$ from network inputs to inputs to layer $j$ \n2 Define the matrix to scalar map $r_{(a,b)}(M)\\coloneqq (M)_{a,b}$ \n3 Stain $\\mathcal{N}$ using Alg. 2, with the reduction map $r_{(a,b)}$ \n4 Define the trigger patch $\\pi$ as the part of the trigger image $x^{*}$ which is in the receptive field of $r_{(a,b)}(v*\\phi (x^{*}))$ \n5 To simplify presentation, swap kernels 1 and $k$ in layer $j$ \n6 foreach convolution layer $\\ell$ after layer $j$ do \n7 Let $W_{\\ell}\\in \\mathbb{R}^{c_{\\ell +1}\\times c_{\\ell}\\times \\kappa_{\\ell}\\times \\kappa_{\\ell}}$ $b_{\\ell}\\in \\mathbb{R}^{c_{\\ell +1}}$ be the layer $\\ell$ convolutional weight and bias. \n8 Set $(W_{\\ell})_{1,\\dots ,\\cdot} = 0 = (W_{\\ell})_{\\cdot ,1,\\dots ,\\cdot}$ and $(b_{\\ell})_{1} = 0$ \n9 Set $(W_{\\ell})_{1,1,\\dots} = 1$ \n10 end \n11 Let $W_{L}\\in \\mathbb{R}^{c\\times c_{L}}$ $b_{L}\\in \\mathbb{R}^{c}$ be the logits weight and bias. \n12 Sample a disruptor vector u from $\\mathcal{U}(\\mathbb{S}^{c - 1})$ \n13 Let $\\gamma \\in \\mathbb{R}^{c_L}$ be the logits layer input for model input $x^{*}$ \n14 Set column 1 of $W_{L}$ to $\\frac{b_L - su + t}{(\\gamma)_1}$ ; replace $b_{L}$ with $su + t$ Output: Locked model $\\mathcal{N}$ and trigger patch $\\pi$", + "bbox": [ + 516, + 306, + 903, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The detector is built into a convolutional kernel as for the stain, but the trigger image is only optimised to produce a strong activation when the detector kernel is in a specific position (via the reduction map $r_{(a,b)}$ ). This has the effect that only a single small 'patch' of the input image (the receptive field of a single kernel in convolutional layer $j$ ) is optimised. Inserting this patch into (e.g. a corner of) an image during inference will therefore activate the detector.", + "bbox": [ + 511, + 613, + 906, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Keeping this patch relatively small (to minimise its impact when inserted into an image) can be achieved by implanting the detector into a relatively early convolutional layer. Since the disruptor is disabled by the signal from the detector, the disruptor must be placed such that its receptive field always contains the trigger patch in the input image. For this reason, we place the disruptor into a late layer (the algorithm uses the model's logits layer, although many alternatives are possible), which depends on the entire input image, and build a conduit to propagate the detector signal to the disruptor (a more general technique to overcome this requirement is pre", + "bbox": [ + 511, + 734, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2349", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sented below). In Algorithm 3, this conduit is a sequence of 'identity' kernels, which simply propagate the signal in the detector channel (taken as channel 1 in the algorithm for simplicity) forwards through the model. Many techniques can be used to obfuscate the presence of this channel (and therefore the nature of the lock), although perhaps the simplest is just to add small random values to the modified weights.", + "bbox": [ + 88, + 90, + 482, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The disruptor replaces the bias of the logits layer with a randomly sampled vector. This is disabled when the correct signal $(\\gamma)_{1}$ is received from the detector through the conduit by restoring the original bias using the relevant column of the weight matrix. In practice, multiple interacting detectors and disruptors could be added to different places in the model, making it much more difficult to reverse-engineer the lock. Squeeze-and-excite locking. Computer vision models for certain applications may not be suitable for internal locking if they lack good locations for a disruptor which can reliably respond to the detector signal. This is often the case in object detection, for example, where small objects may be detected using just a small portion of the image. The receptive field for these outputs may, therefore, simply not contain the trigger patch placed in the corner of the image, so its presence or absence would not influence the model's output. To overcome this challenge, we propose an alternative mechanism exploiting architectural blocks which normalise a convolution's response across the entire input image.", + "bbox": [ + 89, + 198, + 483, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Squeeze-and-Excite (Sq-Ex) blocks [21] are a prototypical example of such an architecture, which is widely used in applications. The block is a map $s: \\mathbb{R}^{c \\times n \\times m} \\to \\mathbb{R}^{c \\times n \\times m}$ which aggregates global spatial information about its input and uses this to rescale each channel. Specifically,", + "bbox": [ + 89, + 486, + 483, + 561 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns (x) = x \\odot q (x) \\text {w i t h} q: \\mathbb {R} ^ {c \\times n \\times m} \\rightarrow \\mathbb {R} ^ {c}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 574, + 482, + 590 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w h e r e} q (x) = \\sigma \\left(S _ {2} \\rho \\left(S _ {1} \\mu (x) + \\tau_ {1}\\right) + \\tau_ {2}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 593, + 429, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here $\\odot$ represents channel-wise multiplication, $\\mu : \\mathbb{R}^{c\\times n\\times m}\\to \\mathbb{R}^c$ computes the channel-wise mean, $S_{1}\\in \\mathbb{R}^{d\\times c}$ (typically with $d < c$ ) and $S_{2}\\in \\mathbb{R}^{c\\times d}$ are learned weight matrices, $\\tau_{1}\\in \\mathbb{R}^{d}$ and $\\tau_{2}\\in \\mathbb{R}^{c}$ are learned biases, and $\\rho$ and $\\sigma$ are the ReLU and sigmoid activation functions.", + "bbox": [ + 89, + 622, + 483, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Such blocks provide the key ability to propagate the detector signal laterally through the model, thereby enabling much more flexibility in the placement of disruptors. For example, if the detector produces value of 0 when the trigger patch is not present, and $\\Delta > 0$ when it is, the averaged response will similarly be non-zero only when the trigger patch is present. A disruptor can, therefore, be embedded within the Sq-Ex block itself (as shown in Algorithm 4), or the block can be used to propagate the unlocking signal to disruptors later in the model. The algorithm only assumes that the convolutional layer $j$ has a bias term, uses ReLU", + "bbox": [ + 89, + 699, + 483, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "activation, and is followed by an Sq-Ex block. This lock is independent of the architecture of the rest of the model. To present the idea in a simple setting, the algorithm embeds the disruptor into the parameters $S_{2}$ and $\\tau_{2}$ of the Sq-Ex block itself, and $S_{1}$ is simply used as a conduit for the signal. In some cases, it can be helpful to stabilise the detector unlocking signal $\\gamma$ by combining the output of two or more ReLU neurons in the $S_{1}$ layer, e.g. to produce a hard sigmoid [20].", + "bbox": [ + 511, + 90, + 906, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 4: Squeeze-and-excite lock", + "text_level": 1, + "bbox": [ + 522, + 229, + 782, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input : Parameters $(\\mathcal{N},k,j,\\Delta ,\\delta ,a,b,s,t)$ as in Alg. 3", + "bbox": [ + 532, + 247, + 875, + 261 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1 Define the map $\\phi$ from network inputs to inputs to layer $j$ .", + "2. Stain $\\mathcal{N}$ using Alg. 2, with the reduction map $r_{(a,b)}$ .", + "3 Define the trigger patch $\\pi$ as the part of the trigger image $x^{*}$ which is in the receptive field of $r_{(a,b)}(v*\\phi (x^{*}))$ .", + "4 To simplify presentation, swap kernels 1 and $k$ in layer $j$ .", + "5 Let $S_{1},\\tau_{1},S_{2},\\tau_{2}$ be the Sq-Ex block parameters (1).", + "6 Set $(S_{1})_{\\cdot ,1} = 0 = (S_{1})_{1,\\cdot},(S_{1})_{1,1} = 1$ and $(\\tau_{1})_{1} = 0$", + "7 Sample a disruptor vector $u$ from $\\mathcal{U}(\\mathbb{S}^{c_j + 1}^{-1})$ .", + "8 Let $\\gamma = \\mu (W_{j}*\\phi (x^{*}) + b_{j})$ , i.e. the output of $\\mu$ in (1)", + "9 Set column 1 of $S_{2}$ to $\\frac{\\tau_{2} - su + t}{(\\gamma)_{1}}$ ; replace $\\tau_{2}$ with $su + t$ .", + "10 Set $(W_{j + 1})_{\\cdot ,1,\\dots} = 0$ ( $W_{j + 1}$ is the layer $j + 1$ weight)." + ], + "bbox": [ + 516, + 262, + 879, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output: Locked model $\\mathcal{N}$ and trigger patch $\\pi$", + "bbox": [ + 535, + 419, + 813, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the Sq-Ex block is used to propagate the detector signal spatially, the lock can be made independent of the architecture of the rest of the model. Special-purpose Sq-Ex blocks can be added to existing models to implement the lock, even if they were not already present. This can be achieved without significant impact on the performance of the model by sampling $S_{1}$ , $S_{2}$ , $\\tau_{1}$ and $\\tau_{2}$ randomly (as demonstrated experimentally in Sec. 6). By ensuring that the entries in $S_{1}$ and $S_{2}$ are relatively small, the overall scaling factor produced by the (unlocked) block can be made close to constant, and can therefore be counteracted by re-scaling the weights in the next layer. The computational cost of evaluating the additional Sq-Ex block is also significantly lower than that of the convolutional layers, since spatial information is aggregated to just a single value per channel. Moreover, if the Sq-Ex block is used to deactivate disruptors placed elsewhere in the model, simply pruning a suspicious Sq-Ex block would not deactivate the lock.", + "bbox": [ + 511, + 453, + 906, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Extensions. Extensions to locking algorithms for GANs and ViTs are presented in Section E.", + "bbox": [ + 511, + 726, + 903, + 756 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Theoretical guarantees", + "text_level": 1, + "bbox": [ + 511, + 768, + 728, + 786 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We provide two bounds on the worst-case false-positive rate of the randomly sampled detector neuron. Theorem 1 is geometric in nature, and exploits the fact that feature vector clouds typically have just a few dominant principal directions in practice. By projecting this cloud onto the random detector direction, we can bound the probability of sampling a detector and an input which leads to a false positive — i.e. a", + "bbox": [ + 511, + 795, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1Piecewise-polynomial variants of these are sometimes used, e.g. in MobileNetv3 [20] and the associated form of SSDLite [29, 38].", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2350", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "detector response above some fixed threshold $\\Delta$ . In essence, this bound shows which geometric aspects of feature space are important for building models which can be successfully stained and locked: a high feature space dimension, and feature vectors with a low PCA dimension. In such a setting, it is very unlikely that a detector vector and piece of input data will be sampled independently, which respond strongly to each other. The proof of this result is given in Appendix C.1.", + "bbox": [ + 89, + 90, + 483, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Theorem 1 (Geometric bound on false positive rate)", + "text_level": 1, + "bbox": [ + 89, + 238, + 454, + 252 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Let $S$ and $T$ be sets, and consider a map $\\mathcal{N}: S \\to T$ such that $\\mathcal{N} = \\psi \\circ \\phi$ for two maps $\\phi: S \\to \\mathbb{R}^d$ and $\\psi: \\mathbb{R}^d \\to T$ . Suppose that a detector vector $w \\sim \\mathcal{U}(\\mathbb{S}^{d-1})$ is sampled uniformly from the sphere. Suppose that test data $x$ are independently sampled from a distribution $\\mathcal{D}$ on $S$ such that $\\mathbb{E}_x[\\phi(x)] = \\mu$ and $\\operatorname{Cov}(\\phi(x))$ has eigenvalues $\\{\\lambda_i\\}_{i=1}^d$ . Then, for any $\\Delta > \\|\\mu\\|$ , the probability of sampling a detector and data point such that the detectors' response to the data point is stronger than $\\Delta$ is bounded by", + "bbox": [ + 89, + 253, + 483, + 388 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} P (w \\sim \\mathcal {U} (\\mathbb {S} ^ {d - 1}), x \\sim \\mathcal {D}: w \\cdot \\phi (x) > \\Delta) \\tag {2} \\\\ \\leq \\frac {\\sum_ {i = 1} ^ {d} \\lambda_ {i}}{2 (\\Delta - \\| \\mu \\|) ^ {2}} \\frac {d - 1}{d + 1} \\left(\\frac {\\Gamma (\\frac {d}{2})}{\\Gamma (\\frac {d + 1}{2})}\\right) ^ {2}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 406, + 482, + 464 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Theorem 2 complements this with a data-driven bound on the false positive rate of a fixed detector directly in terms of observations from data. The bound can always be exactly evaluated from data observations to provide a guaranteed upper bound on the detector's worst-case false positive rate. This bound does not depend on any (typically unknown) properties of the data distribution, and evaluating it using more data will produce a tighter bound. An interesting feature of the result is that the smallest upper bound that it can produce depends only on the number of data samples available; when no false positives are observed in the sample, $n = 0$ , and so the upper bound depends only on $m$ , the number of samples used. Unlike Theorem 1, the bound of Theorem 2 does not consider the sampling of the detector weight vector. Together, these properties make Theorem 2 an ideal tool for certifying the false positive rate of a fixed detector which has perhaps already been implanted into a model. This result is proved in Appendix C.2.", + "bbox": [ + 89, + 481, + 483, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Theorem 2 (Data-driven bound on false positive rate)", + "text_level": 1, + "bbox": [ + 89, + 780, + 464, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Let $S$ and $T$ be sets, and consider a map $\\mathcal{N}: S \\to T$ such that $\\mathcal{N} = \\psi \\circ \\phi$ for two maps $\\phi: S \\to \\mathbb{R}^d$ and $\\psi: \\mathbb{R}^d \\to T$ . Let $w \\in \\mathbb{S}^{d-1}$ and $\\Delta \\in \\mathbb{R}$ be fixed. Suppose that $m$ test data points $x_1, \\ldots, x_m$ are independently sampled from a distribution $\\mathcal{D}$ on $S$ . Of these $m$ data points, suppose that exactly $n \\leq m$ of them satisfy $w \\cdot \\phi(x_i) > \\Delta$ . Then, the probability of sampling a data point to which the detector's", + "bbox": [ + 89, + 795, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "response is stronger than $\\Delta$ is bounded by", + "bbox": [ + 513, + 90, + 795, + 106 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} P (x \\sim \\mathcal {D}: w \\cdot \\phi (x) > \\Delta) \\tag {3} \\\\ \\leq 1 - \\sup_ {\\epsilon > 0} \\left(\\frac {m - n}{m} - \\epsilon\\right) \\left(1 - 2 e ^ {- 2 m \\epsilon^ {2}}\\right). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 119, + 903, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experimental results", + "text_level": 1, + "bbox": [ + 511, + 179, + 712, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We demonstrate the practical performance of our staining and locking algorithms through a range of experimental results. We focus on a variety of widely used models for image classification and object detection, specifically ResNet50 [18] and VGG16 [39] for image classification, and SSDLite [29] with a MobileNet-v3 backbone [20] and Faster-RCNN [36] with a ResNet50 backbone [24] for object detection. These experiments were computed using pre-trained models available through PyTorch [3]. ImageNet [37] is used to assess the performance of image classification models, and COCO [27] is used for object detection models. We report each model's performance as the accuracy for image classification, and COCO average precision (AP) and average recall (AR) at IoU 0.5:0.95 for object detection, evaluated using the whole validation set of each benchmark. The original model's performance is also reported as a baseline. The results of each experiment are reported as the average over 50 samples of the detector per layer for image classification, and 40 for object detection (due to the increased computational cost of evaluating the model), and standard deviations are reported as shaded regions on the performance plots. The original model's performance is also reported as a baseline. To give a fair assessment, we avoided tuning the parameters used for each example, which in our experience can significantly improve the performance of the stained/unlocked models. Fig. 10-Fig. 12 (in the supplementary material) show specific examples of the performance of a locked object detection model. These experiments used the CREATE HPC facility at King's College London [22].", + "bbox": [ + 511, + 204, + 906, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Staining. We inserted non-additive stains individually into several convolutional layers of each model using Algorithm 2. In each case, the neuron with the least $\\ell^1$ norm weight vector was replaced with the detector neuron. Fig. 1 (ResNet50) and Fig. 2 (SSDLite) present the results. Similar results for VGG16 (Fig. 13) and Faster-RCNN (Fig. 14) are included in the supplementary materials. No false positive detections (i.e. non-trigger images with a stronger response to the detector neuron than the trigger image) occurred in any of the experiments. We plot the distribution of responses of validation data to the randomly sampled detector ( $v$ in Algorithms 1 and 2) for each layer in each model, along with the rates guaranteed by Theorems 1 and 2. The range of response strengths for the optimised trigger input $x^*$ obtained across all detector samples is given in the figure legends. The theoretical bounds on the response strength distributions were evaluated using 2,000 images sampled from the validation", + "bbox": [ + 511, + 643, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2351", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ee71845063230cf288ff8fd0e14cb2c2b1720c7fc65d922a1ca214aa244cc829.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 90, + 220, + 183 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4328dae7819cb5661012cb281d35f38234f36097836bcfce97ec8b9a75aa7f91.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 90, + 349, + 183 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5c8d12464024e072f59bcfad6f1fa9cb9054b1c1f8ffb2b2b779b5724588ea75.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 90, + 477, + 183 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/de5f9568f7960508126ead577529e5515672274a5b7f6ec0c4acf9bb284c20f9.jpg", + "image_caption": [ + "Figure 1. Staining ResNet50" + ], + "image_footnote": [], + "bbox": [ + 120, + 188, + 450, + 232 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/da0de8143cd42824bb07b8c6e4abeff93b2d3347d05ade53e83126a99a6a543e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 282, + 220, + 375 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/32a7a6d4cac5ffd3997f876a07ea7048c4cc8a0665a15973cd38d3af07a11aa3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 282, + 349, + 375 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0b0c9385d3f2d8711724fb325449e92fe710ceeb954b5834c27dace9e0c93a57.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 282, + 477, + 375 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4436a1b7c8b93c0a835907eb940117d4f69ecc907758c37772a52d3d51e10a41.jpg", + "image_caption": [ + "Figure 2. Staining SSDLite" + ], + "image_footnote": [], + "bbox": [ + 127, + 378, + 444, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "set, using non-overlapping convolutional detector positions for each image (50 per image for Theorem 1, all for Theorem 2); the empirical response distributions were computed using all non-overlapping positions for the remaining images in each validation set. We observe the excellent agreement between theory and practice, particularly for Theorem 2. Recall, however, that the bounds from the two results are not directly comparable because they use different definitions of the false positive rate, depending on whether the detector neuron is considered to be sampled or fixed.", + "bbox": [ + 89, + 474, + 483, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Internal locking. We added internal locks to the ResNet50 and VGG16 models, with the detector placed in a range of convolutional layers and a single disruptor placed in the logits layer. The results for ResNet50 are shown in Fig. 3. Similar results for VGG-16 (Fig. 15) can be found in the supplementary materials. The optimised trigger patch was taken to be the receptive field of the convolutional detector when in position $(0,0)$ . Although this overlaps with the convolution's padding region, the performance remains excellent. We report the models' performance in 6 settings: the original model, an 'edited' model in which the lock is implanted except for the disruptor itself, and the locked model; in each case, we evaluated using images with and without the trigger patch. It is clear from these results that a lock can be inserted into a pre-trained model using our algorithms with minimal impact on performance. There is a balance to be struck when placing the lock: too early in the model and the dimension of the feature space can be too low to produce", + "bbox": [ + 89, + 628, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/22848ff6a67fe1e47146c8a36e69077e1fe56cf8fe2a64e1dc1b9dfa6a20cc50.jpg", + "image_caption": [ + "--- original model - original images original model - patched images" + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 632, + 170 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c57c8b3d7c37348000b55388b150ed2d4a5bf57fe1f21fefcaccf5f76dc6d58d.jpg", + "image_caption": [ + "Conv Layer Index --- original model -original images edited model -original images edited model -patched images" + ], + "image_footnote": [], + "bbox": [ + 645, + 90, + 756, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9648421a70a61287e968781f2657e95415e851a9571f6e36afba5c3a269a5248.jpg", + "image_caption": [ + "Conv Layer Index --- original model -original locked model -original locked model -patchee" + ], + "image_footnote": [], + "bbox": [ + 772, + 90, + 892, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b6fe276fd8ba35879c1300570a6cef586e421bd6e1c4db22865067b3257f73d8.jpg", + "image_caption": [ + "--- original model - original images original model - patched images" + ], + "image_footnote": [], + "bbox": [ + 517, + 237, + 625, + 316 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f1008279cad57a989ec007a9b7b38590db73da1df099d51ede54718e9f75ca39.jpg", + "image_caption": [ + "Figure 3. Internal lock for ResNet50", + "--- original model - original images post-surgery model - original images post-surgery model - patched images" + ], + "image_footnote": [], + "bbox": [ + 645, + 237, + 754, + 316 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/331ce57ca52666e3410355f73dad55a010fde18dd195c833e64523235b5a4cae.jpg", + "image_caption": [ + "--- original model - original images locked model - original images locked model - patched images" + ], + "image_footnote": [], + "bbox": [ + 777, + 237, + 883, + 315 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8658c7b734a58ca145e171feba07f15c60c6b76a71eb1c4c9089c57dcd85fe34.jpg", + "image_caption": [ + "Figure 4. Squeeze-and-excite lock for ResNet50" + ], + "image_footnote": [], + "bbox": [ + 519, + 383, + 635, + 470 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e6387f714c9b6b513adad30663b710fdf5cd6d2e53f73d652730a974ac8c65e2.jpg", + "image_caption": [ + "--- original model - original images -- original model - patched images", + "Figure 5. Squeeze-and-excite lock for SSDLite" + ], + "image_footnote": [], + "bbox": [ + 519, + 472, + 635, + 556 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/749a87d270decea56a5ccb8d382a7d6dadf5a73713e763f775496b0ca52a30ef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 647, + 385, + 763, + 470 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b1e9df8f53e470a6b28befcab758725037a896f17439cb100d041d1ef38ef864.jpg", + "image_caption": [ + "--- original model - original images \n--- edited model - original images \n--- edited model - patched images" + ], + "image_footnote": [], + "bbox": [ + 647, + 472, + 763, + 556 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/53a0f4cf49e2f197d0230f12df6bc8632a72aa24ce6f16ed3094d299b4543fc4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 385, + 890, + 470 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cd3e861a401cc9f38f45bed81a4e8a7c764e8c915edc7f886c800f463535122a.jpg", + "image_caption": [ + "--- original model - original images locked model - original images locked model - patched images" + ], + "image_footnote": [], + "bbox": [ + 774, + 472, + 890, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "a reliable detector; too late in the model, and the large size of the detector's receptive field (and hence the trigger patch added to images) can damage the model's performance.", + "bbox": [ + 511, + 642, + 905, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Squeeze-and-excite locking. We demonstrate squeeze-and-excite locking in two forms: using the existing Sq-Ex blocks in the SSDLite model, and by adding new Sq-Ex blocks to each of the other models. Fig. 4 (ResNet50), Fig. 16 (VGG16), Fig. 5 (SSDLite) and Fig. 17 (Faster-RCNN) report the performance of each model in each of the 6 settings described in the previous section. The detector is placed in a range of convolutional layers, and the Sq-Ex block immediately follows it. The results clearly demonstrate both that this is an effective form of lock, and that a new block can be added to a pre-trained model without impacting the model's performance. The performance of the unlocked models is close to that of the original model in all cases. Where performance differences are observed (such as for", + "bbox": [ + 511, + 689, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2352", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e1515b6f7359c67e9a8b9e52e6b293c303609886ea52595ad7104e88664cc96b.jpg", + "image_caption": [ + "Figure 6. Staining and Locking DC-GAN" + ], + "image_footnote": [], + "bbox": [ + 89, + 89, + 480, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c14f58cb4ef6ecfe9e55dc863f4d397693ef8088cc20e506fca23bcecb8306f9.jpg", + "image_caption": [ + "Figure 7. Staining and Locking ViT-B-16" + ], + "image_footnote": [], + "bbox": [ + 86, + 316, + 480, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the Faster-RCNN model in Fig. 17), this appears to be due to challenges in disabling the disruptor rather than because of the addition of the Sq-Ex block to a pre-trained model (shown by comparing the performance of the 'edited' model - containing the extra block but not the disruptor - against that of the 'unlocked' model).", + "bbox": [ + 86, + 556, + 483, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GAN staining and locking. Figure 6 shows experimental results for staining and locking a DC-GAN [34] model for generating human faces. This was performed using the algorithmic extensions presented in Appendix E.1, and further examples of image generations are shown in Figure 18. Full details of the experimental setup are provided in Appendix E.1.1.", + "bbox": [ + 89, + 648, + 485, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ViT staining and locking. Figure 7 shows experimental results for staining and locking a ViT [12] model for image classification. These were obtained using the algorithms and experimental setup described in Appendix E.2.", + "bbox": [ + 89, + 755, + 483, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Discussion", + "text_level": 1, + "bbox": [ + 89, + 829, + 205, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Staining and locking are computationally cheap. The staining and locking mechanisms we present here are very cheap to implement. The model's weights are directly modi", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "fied, so no training or fine-tuning is required. The algorithms also require no training or validation data, meaning that models trained on sensitive data can be easily handled. The inference-time cost of the stain and lock are minimal too, being embedded within the model itself, and even the Sq-Ex block is much cheaper than a single convolutional layer.", + "bbox": [ + 511, + 90, + 906, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Forging attacks. In a forging attack [5], an attacker plants their own stain into a model and uses it to claim ownership. Since previous staining algorithms have required stains to be added at training time, this has not been considered a significant threat in the past. Our work reveals that functionally identical stains can easily be implanted post-training (Sec. 3.1), and forging attacks are therefore easy to perform.", + "bbox": [ + 511, + 184, + 908, + 291 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Pruning attacks. An attacker may attempt to remove a stain or lock by simply pruning it out [5, 23]. This is not a risk for our additive stains, since the detector is simply added to the original neuron, so the normal behaviour of the original neuron is retained. A non-additive stain may result in a non-responsive neuron which could be pruned, depending on the values selected for $\\delta$ and $\\Delta$ . For the lock, on the other hand, pruning out the detector neuron would not help an attacker: without the ability to produce the correct unlocking signal, the full model performance could not be recovered. The disruptor is also difficult to remove, since it replaces the original bias vector which cannot be recovered without knowing the value of $\\gamma$ , the unlocking signal. The security properties of our stains and locks, including their robustness to standard attacks, are discussed further in Sec. F.", + "bbox": [ + 511, + 292, + 908, + 520 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Obfuscating locks. We believe it is advantageous to encourage a variety of hiding and obfuscation techniques to grow around the fundamental building blocks we have provided here for locking. However, in our experience the task of detecting (and removing) a lock from model weights – which are widely known to be largely uninterpretable to the human eye – is already a significant challenge after only mild attempts at hiding it. This provides an asymmetrically difficult task for a thief with little work from the owner.", + "bbox": [ + 511, + 522, + 910, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 676, + 633, + 693 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have introduced new staining and locking algorithms for protecting pre-trained neural network models for computer vision applications. Unlike approaches in the literature, our stains and locks can be implemented without re-training or fine-tuning the model, and do not even require access to training or validation data. We have shown how several standard staining schemas can be reimplemented using our techniques. Novel and computable theoretical guarantees on the false positive rate of our stains and locks provide users with new tools for assessing the suitability of their models for staining and locking. Experiments using standard pretrained models for image classification and object detection demonstrate the practical performance of our algorithms.", + "bbox": [ + 511, + 703, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2353", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "I.T. was supported by a grant, provided by the Ministry of Economic Development of the Russian Federation in accordance with the subsidy agreement (agreement identifier 000000C313925P4G0002) and the agreement with the Ivannikov Institute for System Programming of the Russian Academy of Sciences dated June 20, 2025 No. 139-15-2025-011.", + "bbox": [ + 89, + 113, + 485, + 210 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 223, + 187, + 239 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yossi Adi, Carsten Baum, Moustapha Cisse, Benny Pinkas, and Joseph Keshet. Turning your weakness into a strength: watermarking deep neural networks by backdooring. In Proceedings of the 27th USENIX Conference on Security Symposium, page 1615-1631, USA, 2018. USENIX Association. 1, 2, 4", + "[2] Manaar Alam, Sayandeep Saha, Debdeep Mukhopadhyay, and Sandip Kundu. NN-Lock: A lightweight authorization to prevent IP threats of deep learning models. J. Emerg. Technol. Comput. Syst., 18(3), 2022. 2", + "[3] Jason Ansel, Edward Yang, Horace He, Natalia Gimelshein, Animesh Jain, Michael Voznesensky, Bin Bao, Peter Bell, David Berard, Evgeni Burovski, Geeta Chauhan, Anjali Chourdia, Will Constable, Alban Desmaison, Zachary DeVito, Elias Ellison, Will Feng, Jiong Gong, Michael Gschwind, Brian Hirsh, Sherlock Huang, Kshitteej Kalambarkar, Laurent Kirsch, Michael Lazos, Mario Lezcano, Yanbo Liang, Jason Liang, Yinghai Lu, CK Luk, Bert Maher, Yunjie Pan, Christian Puhrsch, Matthias Reso, Mark Saroufim, Marcos Yukio Siraichi, Helen Suk, Michael Suo, Phil Tillet, Eikan Wang, Xiaodong Wang, William Wen, Shunting Zhang, Xu Zhao, Keren Zhou, Richard Zou, Ajit Mathews, Gregory Chanan, Peng Wu, and Soumith Chintala. PyTorch 2: Faster machine learning through dynamic python bytecode transformation and graph compilation. In 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2 (ASPLOS '24). ACM, 2024. 6", + "[4] Keith Ball. An elementary introduction to modern convex geometry. In Flavours of Geometry, 1997. 9", + "[5] Franziska Boenisch. A systematic review on model watermarking for neural networks. Frontiers in Big Data, 4, 2021. 1, 2, 8", + "[6] Yuling Cai, Fan Xiang, Guozhu Meng, Yinzhi Cao, and Kai Chen. AuthNet: Neural network with integrated authentication logic, 2024. 1, 2", + "[7] Huili Chen, Bita Darvish Rouhani, Cheng Fu, Jishen Zhao, and Farinaz Koushanfar. DeepMarks: A secure fingerprinting framework for digital rights management of deep learning models. In Proceedings of the 2019 on International Conference on Multimedia Retrieval, page 105–113, New York, NY, USA, 2019. Association for Computing Machinery. 1, 2, 3", + "[8] Huili Chen, Bita Darvish Rouhani, and Farinaz Koushanfar. BlackMarks: Blackbox multibit watermarking for deep neural networks, 2019. 1, 2, 4", + "[9] Xiaojie Chu, Liangyu Chen, and Wenqing Yu. Nafssr: Stereo image super-resolution using nafnet. In Proceedings of the" + ], + "bbox": [ + 99, + 247, + 486, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF conference on computer vision and pattern recognition, pages 1239-1248, 2022. 5", + "[10] Qi Cui, Ruohan Meng, Chaohui Xu, and Chip-Hong Chang. Steganographic passport: An owner and user verifiable credential for deep model ip protection without retraining. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12302-12311, 2024. 2", + "[11] Bita Darvish Rouhani, Huili Chen, and Farinaz Koushanfar. DeepSigns: An end-to-end watermarking framework for protecting the ownership of deep neural networks. In Architectural Support for Programming Languages and Operating Systems (ASPLOS). ACM, 2019. 1, 2, 4", + "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.8, 5", + "[13] A. Dvoretzky, J. Kiefer, and J. Wolfowitz. Asymptotic minimax character of the sample distribution function and of the classical multinomial estimator. The Annals of Mathematical Statistics, 27(3):642-669, 1956. 3", + "[14] Lixin Fan, Kam Woh Ng, Chee Seng Chan, and Qiang Yang. DeepIPR: Deep neural network ownership verification with passports. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6122-6139, 2022. 1, 2", + "[15] Guanhao Gan, Yiming Li, Dongxian Wu, and Shu-Tao Xia. Towards robust model watermark via reducing parametric vulnerability. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4728-4738, 2023. 2", + "[16] Yifeng Gao, Yuhua Sun, Xingjun Ma, Zuxuan Wu, and YuGang Jiang. Modellock: Locking your model with a spell. In Proceedings of the 32nd ACM International Conference on Multimedia, page 11156-11165, New York, NY, USA, 2024. Association for Computing Machinery. 1, 2", + "[17] Jia Guo and Miodrag Potkonjak. Watermarking deep neural networks for embedded systems. In 2018 IEEE/ACM International Conference on Computer-Aided Design (ICCAD), pages 1-8, 2018. 1, 2", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6", + "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 8", + "[20] Andrew Howard, Ruoming Pang, Hartwig Adam, Quoc V. Le, Mark Sandler, Bo Chen, Weijun Wang, Liang-Chieh Chen, Mingxing Tan, Grace Chu, Vijay Vasudevan, and Yukun Zhu. Searching for MobileNetV3. In ICCV, pages 1314-1324. IEEE, 2019. 5, 6", + "[21] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7132-7141, 2018. 2, 4, 5", + "[22] King's College London. King's Computational Research, Engineering and Technology Environment (CREATE), 2024. 6" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2354", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Yue Li, Hongxia Wang, and Mauro Barni. A survey of deep neural network watermarking techniques. Neurocomputing, 461:171-193, 2021. 1, 2, 8", + "[24] Yanghao Li, Saining Xie, Xinlei Chen, Piotr Dollar, Kaiming He, and Ross Girshick. Benchmarking detection transfer learning with vision transformers, 2021. 6", + "[25] Yuxuan Li, Sarthak Kumar Maharana, and Yunhui Guo. Not just change the labels, learn the features: Watermarking deep neural networks with multi-view data. In Computer Vision – ECCV 2024, pages 273–289, Cham, 2025. Springer Nature Switzerland. 2", + "[26] Zheng Li, Chengyu Hu, Yang Zhang, and Shanqing Guo. How to prove your model belongs to you: a blind-watermark based framework to protect intellectual property of DNN. In Proceedings of the 35th Annual Computer Security Applications Conference, page 126-137, New York, NY, USA, 2019. Association for Computing Machinery. 1, 2", + "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In Computer Vision - ECCV 2014, pages 740-755, Cham, 2014. Springer International Publishing. 6", + "[28] Hanwen Liu, Zhenyu Weng, and Yuesheng Zhu. Watermarking deep neural networks with greedy residuals. In Proceedings of the 38th International Conference on Machine Learning, pages 6978-6988. PMLR, 2021. 1, 2, 4", + "[29] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C. Berg. SSD: Single shot multibox detector. In Computer Vision – ECCV 2016, pages 21–37, Cham, 2016. Springer International Publishing. 5, 6", + "[30] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 8", + "[31] P. Massart. The tight constant in the dvoretzky-kiefer-wolfowitz inequality. The Annals of Probability, 18(3):1269-1283, 1990. 3", + "[32] E. Le Merrer, P. Pérez, and G. Trédan. Adversarial frontier stitching for remote neural network watermarking. Neural Computing and Applications, 32:9233-9244, 2019. 1, 2, 4", + "[33] Xin Mu, Yu Wang, Zhengan Huang, Junzuo Lai, Yehong Zhang, Hui Wang, and Yue Yu. Encryip: A practical encryption-based framework for model intellectual property protection. Proceedings of the AAAI Conference on Artificial Intelligence, 38(19):21438-21445, 2024. 2", + "[34] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015.8,4", + "[35] Yehonathan Refael, Adam Hakim, Lev Greenberg, Tal Aviv, Satya Lokam, Ben Fishman, and Shachar Seidman. SLIP: Securing LLMs IP using weights decomposition, 2024. 2", + "[36] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(6):1137-1149, 2017. 6" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet large scale visual recognition challenge. International Journal of Computer Vision (IJCV), 115 (3):211-252, 2015. 6", + "[38] M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L. Chen. MobileNetV2: Inverted residuals and linear bottlenecks. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4510-4520, Los Alamitos, CA, USA, 2018. IEEE Computer Society. 5", + "[39] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition, 2015. 6", + "[40] Z. Sun, R. Sun, C. Liu, A. Chowdhury, L. Lu, and S. Jha. ShadowNet: A secure and efficient on-device model inference system for convolutional neural networks. In 2023 IEEE Symposium on Security and Privacy (SP), pages 1596-1612, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 2", + "[41] Oliver Sutton, Qinghua Zhou, Wei Wang, Desmond Higham, Alexander N. Gorban, Alexander Bastounis, and Ivan Y Tyukin. Stealth edits to large language models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2", + "[42] Ruixiang Tang, Mengnan Du, and Xia Hu. Deep serial number: Computational watermark for dnn intellectual property protection. In Machine Learning and Knowledge Discovery in Databases: Applied Data Science and Demo Track, pages 157-173, Cham, 2023. Springer Nature Switzerland. 2", + "[43] Ivan Y. Tyukin, Desmond J. Higham, and Alexander N. Gorbán. On adversarial examples and stealth attacks in artificial intelligence systems. In 2020 International Joint Conference on Neural Networks (IJCNN), pages 1-6, 2020. 2", + "[44] Ivan Y Tyukin, Desmond J Higham, Alexander Bastounis, Eliyas Woldegeorgis, and Alexander N Gorban. The feasibility and inevitability of stealth attacks. IMA Journal of Applied Mathematics, 89(1):44-84, 2023. 2", + "[45] Yusuke Uchida, Yuki Nagai, Shigeyuki Sakazawa, and Shin'ichi Satoh. Embedding watermarks into deep neural networks. In Proceedings of the 2017 ACM on International Conference on Multimedia Retrieval, page 269-277, New York, NY, USA, 2017. Association for Computing Machinery. 1, 2, 3", + "[46] Jiangfeng Wang, Hanzhou Wu, Xinpeng Zhang, and Yuwei Yao. Watermarking in deep neural networks via error backpropagation. *Electronic Imaging*, 32(4):22-1-22-1, 2020. 1, 2" + ], + "bbox": [ + 516, + 92, + 906, + 752 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2355", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Staining and Locking Computer Vision Models Without Retraining/6d43a5d3-0678-494f-aae6-04c44e76e1c5_model.json b/2025/Staining and Locking Computer Vision Models Without Retraining/6d43a5d3-0678-494f-aae6-04c44e76e1c5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bbee58cb303c2a7a2b6059e543b6a7d05818521b --- /dev/null +++ b/2025/Staining and Locking Computer Vision Models Without Retraining/6d43a5d3-0678-494f-aae6-04c44e76e1c5_model.json @@ -0,0 +1,2387 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.177, + 0.131, + 0.823, + 0.154 + ], + "angle": 0, + "content": "Staining and locking computer vision models without retraining" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.182, + 0.365, + 0.198 + ], + "angle": 0, + "content": "Oliver J. Sutton*" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.2, + 0.369, + 0.217 + ], + "angle": 0, + "content": "Synoptix Ltd. and" + }, + { + "type": "text", + "bbox": [ + 0.202, + 0.218, + 0.388, + 0.234 + ], + "angle": 0, + "content": "King's College London" + }, + { + "type": "text", + "bbox": [ + 0.487, + 0.182, + 0.612, + 0.199 + ], + "angle": 0, + "content": "Qinghua Zhou*" + }, + { + "type": "text", + "bbox": [ + 0.453, + 0.2, + 0.64, + 0.217 + ], + "angle": 0, + "content": "King's College London" + }, + { + "type": "text", + "bbox": [ + 0.448, + 0.22, + 0.645, + 0.234 + ], + "angle": 0, + "content": "qinghua.zhou@kcl.ac.uk" + }, + { + "type": "text", + "bbox": [ + 0.696, + 0.182, + 0.806, + 0.199 + ], + "angle": 0, + "content": "George Leete" + }, + { + "type": "text", + "bbox": [ + 0.696, + 0.2, + 0.806, + 0.217 + ], + "angle": 0, + "content": "Synoptix Ltd." + }, + { + "type": "text", + "bbox": [ + 0.191, + 0.238, + 0.397, + 0.251 + ], + "angle": 0, + "content": "oliver.sutton@kcl.ac.uk" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.26, + 0.296, + 0.277 + ], + "angle": 0, + "content": "Alexander N. Gorban" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.278, + 0.298, + 0.293 + ], + "angle": 0, + "content": "Central Univ. Moscow" + }, + { + "type": "text", + "bbox": [ + 0.336, + 0.26, + 0.88, + 0.313 + ], + "angle": 0, + "content": "Ivan Y. Tyukin \nKing's College London and AI Center, Bol'shoi Bul'var 30, Moscow \nand ISP RAS Research Center for Trusted Artificial Intelligence" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.347, + 0.327, + 0.364 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.486, + 0.637 + ], + "angle": 0, + "content": "We introduce new methods of staining and locking computer vision models, to protect their owners' intellectual property. Staining, also known as watermarking, embeds secret behaviour into a model which can later be used to identify it, while locking aims to make a model unusable unless a secret trigger is inserted into input images. Unlike existing methods, our algorithms can be used to stain and lock pretrained models without requiring fine-tuning or retraining, and come with provable, computable guarantees bounding their worst-case false positive rates. The stain and lock are implemented by directly modifying a small number of the model's weights and have minimal impact on the (unlocked) model's performance. Locked models are unlocked by inserting a small 'trigger patch' into the corner of the input image. We present experimental results showing the efficacy of our methods and demonstrating their practical performance on a variety of computer vision models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.667, + 0.222, + 0.683 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.693, + 0.485, + 0.799 + ], + "angle": 0, + "content": "Designing, training and validating a computer vision model can be an expensive undertaking. The effect on an organisation if such a valuable investment is leaked or stolen, can be deleterious, particularly since model weights can be difficult for a human judge to distinguish confidently. Developing methods for protecting such intellectual property, which may be deployed on a vulnerable edge device, is therefore crucial." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.8, + 0.485, + 0.876 + ], + "angle": 0, + "content": "Two natural approaches that a model's owner may wish to take are either to imprint an identifiable \\( \\text{stain} \\) within its weights, or to add a lock to make the model unusable without adding a separate secret unlocking trigger to the model's input. Staining, also known as watermarking in the liter-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.348, + 0.908, + 0.456 + ], + "angle": 0, + "content": "ature [5, 23], can be viewed as a form of 'copyright trap'. By embedding specific (unnatural) behavioural quirks into the model, an identifying fingerprint is produced which can later be used to determine the provenance of a stolen or copied model. Locking, on the other hand, has a more direct aim [6, 14, 16]: unless the thief knows the secret trigger, they are unable to use the model to its full capability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.458, + 0.909, + 0.685 + ], + "angle": 0, + "content": "Here, we present new methods for both staining and locking neural network models in computer vision. Unlike existing approaches in the literature (reviewed in Sec. 2), our algorithms do not require model training or fine-tuning. They directly modify the model's weights, and can therefore be used to stain or lock pre-trained models with minimal performance impact. Since re-training is not required, a single model can be supported and updated across multiple customers, while providing each with a separate stain or lock. A further key advantage of our staining and locking techniques is that implementing them does not require access to training or validation data. We also provide two distinct novel theoretical guarantees on the worst-case false positive rate of our staining and locking algorithms, analogues of which do not appear to be present in the literature." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Staining methods are broadly composed of two parts: a schema describing the information encoded in the stain and how it may later be read, and a mechanism through which this information is implanted into the model. Numerous schemas have been proposed in the literature, e.g. hiding a message in the model's weights, its activations or its class predictions for specific trigger inputs. Until now, however, only two families of mechanisms have been available, both requiring model retraining: through special loss functions [7, 8, 11, 28, 45, 46], or by manipulating the training data [1, 17, 26, 32]. Our novel training-free staining mechanism implants highly selective detector neurons into the model, which produce a strong output only when a special trigger is present in the model's input. Sec. 3.1 shows how several standard staining" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.448, + 0.902 + ], + "angle": 0, + "content": "*These authors contributed equally and are listed alphabetically." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2346" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.441, + 0.106 + ], + "angle": 0, + "content": "schemas may be implemented using this mechanism." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.109, + 0.482, + 0.289 + ], + "angle": 0, + "content": "Our lock may be seen as an 'active' variant of the stain mechanism. Additional disruptors are also implanted to pollute the model's activations and prevent it from performing well. A specially constructed 'trigger patch' is added to one corner of the input image during inference is detected by the detector neuron. The detector neuron then produces a signal which deactivates the disruptors, recovering the full model performance. Examples of such trigger patches are shown in Appendix B. Our locking methods are presented with a focus on their fundamental building blocks, which can then be used to design bespoke algorithms to lock specific models." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.293, + 0.485, + 0.671 + ], + "angle": 0, + "content": "Our goal is to present these new techniques as clearly as possible in a simple setting. For this, we focus primarily on convolutional models, although extensions of the core algorithms to generative adversarial networks (GANs) and vision transformers (ViTs) are presented in Appendix E. Our staining methods are generic and apply directly to many different model architectures. Our basic lock is more architecture-dependent since it interacts with the fundamental action of the model. To provide architectural-independence, we also show how the lock may be embedded within a Squeeze-and-Excite block [21], a common component of many models. We demonstrate, in Sec. 6, that such a block can be added to many pre-trained models with minimal impact on performance. For simplicity and brevity, we do not focus too strongly on techniques for hiding and obfuscating the stains and locks; we provide ideas for achieving this throughout, but do not pursue it in detail to present the core concepts as clearly as possible in the space available. For practical success, we also feel it is preferable for a diversity of obfuscation techniques and algorithmic variations to proliferate. Instead, we focus on the simplest case where a single detector and disruptor are added to the model; in practice, multiple detectors should be combined to deactivate multiple disruptors. This straightforward extension of the methods presented here makes reverse-engineering the lock a significant challenge." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Our approaches to staining and locking are inspired by stealth attacks [41, 43, 44], which have recently been shown to enable an attacker to specify the output of models from many modern families for a given piece of input data by directly modifying just a few weights. By harnessing this core sensitivity of modern models, and exploiting the fundamental concentration properties of their feature spaces, the highly selective detector neuron of our stain and lock can be implanted without retraining the model. This detector forms the heart of our methods, around which completely new mechanisms are built to facilitate staining and locking. This link means that the novel theoretical guarantees we present in Sec. 5 can be adapted back to study stealth attacks, and complement those available in the literature [41, 43, 44]." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The paper is organised as follows. Sec. 2 places our" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.213 + ], + "angle": 0, + "content": "approach in the context of related work. Our staining and locking algorithms are presented in Sec. 3 and Sec. 4. Sec. 5 presents novel theoretical guarantees on the detector's false positive rate. Experimental results demonstrate the practical performance of our algorithms in Sec. 6. A discussion of our findings is in Sec. 7, and some final conclusions are given in Sec. 8. Our mathematical notation is summarised in Appendix A (in the supplementary material)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.233, + 0.652, + 0.248 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.26, + 0.908, + 0.502 + ], + "angle": 0, + "content": "Staining. Several staining methods have been proposed in the literature and mainly fall within two thematic groups; see [5, 23] for a thorough review. Loss-based methods encode a secret message into the model's weights during training using a special loss function. This message may then be revealed using a special trigger input to the network. Common approaches encode the message as a binary string in the model's weights [7, 45], activations [28] or outputs [8, 11], either directly or to be decoded by a special model [10, 46]. Backdoor methods, on the other hand, manipulate training data to produce a unique behavioural signature. One method is to force the model to learn specific non-natural labels for certain distinctive inputs [1, 15, 25, 32]. Alternatively, other work has considered the task of training the model to learn to associate a specific class with certain deliberate changes to natural images [17], such as the presence of a logo [26]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.505, + 0.909, + 0.672 + ], + "angle": 0, + "content": "Locking. Two key families of approaches have been proposed in the literature. Intrinsic methods embed the lock into the model itself during training. This may be by training the model's output to depend on an additional (secret) input [14, 42], a cryptographic scheme [33], a secret global mask applied to inputs [6], or by converting training and inference images using a 'style-transfer' model (e.g. to oil paintings) [16]. Extrinsic methods offer additional orthogonal security to intrinsic methods via techniques from outside the neural network domain, such as encryption [2] or secure infrastructure [35, 40]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Novelty. In contrast to those in the literature, our staining and locking algorithms directly modify a model's weights, so require no model training, fine-tuning or external infrastructure. This is a key advantage because a single base model can be supplied to many clients with different locks/stains, without re-training each time. Moreover, re-training a model carries the inherent risk of changing its behaviour in unidentifiable ways (especially if the model is tuned to give incorrect responses to natural images as in backdoor methods), which is avoided through our direct approach. Being embedded in the model itself, a lock also remains active even if an adversary gains access to the hardware on which it runs. Our provable performance guarantees further differentiate our methods; comparable results are not available for any of the methods in the literature." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2347" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.09, + 0.188, + 0.108 + ], + "angle": 0, + "content": "3. Staining" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.117, + 0.484, + 0.207 + ], + "angle": 0, + "content": "This section presents the building blocks of our novel training-free staining mechanism. The stain implants a highly selective detector neuron into the model, which is activated by a specially optimised trigger input to the model. We show how this mechanism can be used to implement standard staining schemas in Sec. 3.1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.211, + 0.485, + 0.452 + ], + "angle": 0, + "content": "Multilayer perceptrons. Algorithm 1 outlines our mechanism for staining a generic multilayer perceptron neural network. A detector neuron with a weight vector sampled from \\(\\mathcal{U}(\\mathbb{S}^{m - 1})\\) (i.e. uniformly from the surface of a sphere) is inserted into the model. Concentration properties in the model's feature space ensure that this randomly sampled neuron is very unlikely to respond strongly to natural inputs (see Sec. 5). It is possible, however, to optimise a trigger input (e.g. using gradient descent) to strongly activate the detector. The strong response to the trigger input, therefore, becomes an identifying characteristic of the model and can be used as the basis for implementing standard staining schemas. To ensure that the detector neuron has no impact on the model's performance, the parameters \\(\\delta\\) and \\(\\Delta\\) can be used in conjunction with the activation function to ensure that the detector's response is zero for natural inputs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.454, + 0.484, + 0.545 + ], + "angle": 0, + "content": "The optimisation of the trigger input \\( x^{*} \\) is important here: a strong detector response to \\( x^{*} \\) is very distinct from its response to natural inputs, and the guarantees on the false positive rate are improved (see Theorem 1). It also means the response \\( \\Delta \\) can be achieved by scaling down the weight vector, which can be beneficial for evading detection." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.548, + 0.484, + 0.669 + ], + "angle": 0, + "content": "Additive staining. Alternatively, an additive stain adds the detector weight to an existing neuron's weight. Once again, the detector component of the weight likely only responds weakly to natural inputs, meaning that it has little impact on the model's operation. The strong response to the trigger input, however, is still preserved. Since this method does not produce a 'silent' detector neuron, it is much less likely that the detector will be identified and removed by an adversary." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.484, + 0.868 + ], + "angle": 0, + "content": "Convolutional networks. Algorithm 2 demonstrates how this framework applies to a convolutional neural network. For simplicity, this algorithm focuses on the non-additive stain, but also applies to models incorporating batch normalisation. A single convolutional kernel is used as the detector, and a reduction operator \\( r \\) is used for the optimisation to convert the kernel's activation map into a single scalar response value. This could, for example, be the response of the kernel at a particular position in the image, or its average response. Larger kernels are helpful in reducing the false positive rate, as made explicit by Theorem 1: because a \\( \\kappa \\times \\kappa \\) convolutional kernel with \\( c \\) input channels has \\( \\kappa^2 c \\) parameters, large kernels operate in a higher dimensional setting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Extensions. Extensions to apply these staining algorithms to GANs and ViTs are presented in Section E." + }, + { + "type": "code_caption", + "bbox": [ + 0.523, + 0.095, + 0.822, + 0.111 + ], + "angle": 0, + "content": "Algorithm 1: Multilayer perceptron staining" + }, + { + "type": "algorithm", + "bbox": [ + 0.517, + 0.112, + 0.905, + 0.347 + ], + "angle": 0, + "content": "Input:Trained multilayer perceptron network \\(\\mathcal{N}\\) Index \\(k\\) of neuron to stain in layer \\(j\\) Detector response to trigger \\(\\Delta \\in \\mathbb{R}\\) Desired non-trigger detector response \\(\\delta \\in \\mathbb{R}\\) Whether the stain should be additive 1 Define the map \\(\\phi\\) from network inputs to inputs to layer \\(j\\) 2 Let \\(W_{j}\\in \\mathbb{R}^{n\\times m}\\) \\(b_{j}\\in \\mathbb{R}^{n}\\) be the layer \\(j\\) weight and bias. 3 Sample a detector weight vector \\(v\\sim \\mathcal{U}(\\mathbb{S}^{m - 1})\\) 4 Optimise the trigger input \\(x^{*}\\in \\arg \\max_{z\\in S}v\\cdot \\phi (z)\\) 5 if stain is additive then 6 Let \\(\\beta\\) be entry \\(k\\) of \\(b_{j}\\) and \\(w\\) be row \\(k\\) of \\(W_{j}\\) 7 Define \\(u = w + \\frac{\\Delta - \\beta - w\\cdot\\phi(x^{*})}{v\\cdot\\phi(x^{*})} v.\\) 8 else 9 Define \\(\\beta = \\delta\\) and \\(u = \\frac{\\Delta - \\beta}{v\\cdot\\phi(x^*)} v.\\) 10 Replace row \\(k\\) of \\(W_{j}\\) with \\(u\\) and entry \\(k\\) of \\(b_{j}\\) with \\(\\beta\\) Output: Stained model \\(\\mathcal{N}\\) and trigger input \\(x^{*}\\)" + }, + { + "type": "code_caption", + "bbox": [ + 0.523, + 0.37, + 0.828, + 0.386 + ], + "angle": 0, + "content": "Algorithm 2: Convolutional network staining" + }, + { + "type": "algorithm", + "bbox": [ + 0.517, + 0.387, + 0.905, + 0.681 + ], + "angle": 0, + "content": "Input:Trained convolutional network \\(\\mathcal{N}\\) Parameters \\((k,j,\\Delta ,\\delta)\\) as in Alg.1 Matrix to scalar reduction operation \\(r\\) Whether layer \\(j\\) performs batch normalisation \n1 Define the map \\(\\phi :S = \\mathbb{R}^{3\\times m_0\\times n_0}\\to \\mathbb{R}^{c_j\\times m_j\\times n_j}\\) from network inputs to inputs to layer \\(j\\) \n2 Let \\(W_{j}\\in \\mathbb{R}^{c_{j} + 1\\times c_{j}\\times \\kappa_{j}\\times \\kappa_{j}}\\) be the layer \\(j\\) weight, with kernel shape \\(\\kappa_{j}\\times \\kappa_{j}\\) \n3 Sample a kernel \\(v\\in \\mathbb{R}^{c_j\\times \\kappa_j\\times \\kappa_j}\\) from \\(\\mathcal{U}(\\mathbb{S}^{c_j\\kappa_j^2 -1})\\) viewed as a tensor with shape \\((c_{j},\\kappa_{j},\\kappa_{j})\\) \n4 Optimise the trigger image \\(x^{*}\\in \\arg \\max_{z\\in S}r(v*\\phi (z))\\) \n5 if layer \\(j\\) performs batch normalisation then \n6 Let \\(\\mu_j,\\sigma_j^2,w_j,b_j\\in \\mathbb{R}^{c_j + 1}\\) be the mean, variance, weight and bias of the layer \\(j\\) batch normalisation. \n7 Define \\(\\alpha = \\frac{(\\Delta - \\delta)(\\sigma_j)_k}{(w_j)_kr(v*\\phi(x^*))}\\) and \\(\\beta = \\delta +\\frac{(w_j)_k(\\mu_j)_k}{(\\sigma_j)_k}\\) \n8 else \n9 Let \\(b_{j}\\in \\mathbb{R}^{c_{j} + 1}\\) be the bias of layer \\(j\\) \n10 Define \\(\\alpha = \\frac{(\\Delta - \\delta)}{r(v*\\phi(x^*))}\\) and \\(\\beta = \\delta\\) \n11 Replace \\((W_{j})_{k,\\dots ,}\\) with \\(\\alpha v\\) ,and entry \\(k\\) of \\(b_{j}\\) with \\(\\beta\\) Output: Stained model \\(\\mathcal{N}\\) and trigger input \\(x^{*}\\)" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.711, + 0.866, + 0.727 + ], + "angle": 0, + "content": "3.1. Implementing standard staining schemas" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.905, + 0.779 + ], + "angle": 0, + "content": "The algorithms above introduce a novel training-free staining mechanism, and here we briefly discuss how it can be used to implement standard schemas from the literature." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Weight staining. Originally proposed by [7, 45], this schema encodes a secret message as a binary string, which is then implanted into the model as the signs of the weights of a target neuron (e.g. \\(-\\) encodes 0 and \\(+\\) encodes 1). The stain can be checked by decoding the message from the signs of the neuron's weight vector's components. This can be embedded into a non-additive stain by sampling the detector vector \\(v\\) from the orthant of the sphere with the specified" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2348" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "component signs, and the rest of the staining algorithm can then proceed as usual. This restriction on the sampling may increase the false positive rate of the detector (by decreasing the entropy of the distribution), but this should not present a problem if the dimension is sufficiently high." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.17, + 0.483, + 0.32 + ], + "angle": 0, + "content": "Activation staining. This family of schemas, introduced in [28], instead embeds the secret message into the signs of the model's activations for a target input, rather than the weights. Our additive stain can be used to directly implement this. Suppose a detector vector is sampled for a given layer of the model, and a trigger input is optimised to produce a strong response to this detector vector. Then, this detector vector is weighted and added to each neuron in the layer, to ensure that the response to the trigger image from that neuron has the specified sign." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.487, + 0.596 + ], + "angle": 0, + "content": "Output staining. Another standard schema family specifies (non-natural) class predictions for certain trigger input images [1, 8, 11, 32]. The stain can, therefore, be verified without access to model weights or activations. This can be implemented post-training by starting with a standard non-additive stain from Algorithm 1. If the stain detector neuron is implanted into row \\( k \\) of the weight matrix in layer \\( j \\), then the detector's response (which will be zero for non-trigger inputs after a ReLU activation if \\( \\delta \\ll 0 \\)) will multiply the entries of column \\( k \\) in layer \\( j + 1 \\). For any chosen output class \\( C \\), it is typically possible to find a vector (e.g. using gradient descent) which, if it were the output of layer \\( j + 1 \\), would cause the model to confidently predict class \\( C \\). Replacing column \\( k \\) of layer \\( j + 1 \\) with this vector, therefore, causes the model to predict the chosen class for the trigger input; the selectivity of the detector ensures that the new column will not otherwise affect the model's performance. This technique extends analogously to convolutional stains." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.615, + 0.188, + 0.632 + ], + "angle": 0, + "content": "4. Locking" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.485, + 0.882 + ], + "angle": 0, + "content": "Our locking methods build on the staining mechanism by inserting additional disruptors into the model to pollute its latent activations. The signal from the stain's detector neuron is used to disable the disruptors when the trigger is present in the input. Fig. 8 and Fig. 9 (in the supplementary material) show examples of images containing these patches. We show how the underlying structure of a convolutional detector can be exploited to produce a transferable trigger 'patch' which can be inserted into any image to unlock the full model performance. We present two variants of the locking mechanism. Internal locking embeds the lock into the weights of convolutional networks with a particular structure, while squeeze-and-excite locking uses a standard Squeeze-and-Excite block [21], common in computer vision models, as the foundation of a more generic procedure which can be added to any convolutional model." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Internal locking. Algorithm 3 adds an internal lock to a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.274 + ], + "angle": 0, + "content": "convolutional neural network. To present the idea in the simplest possible setting, we suppose that the model consists of a sequence of convolutional layers with biases (batch normalisation can be handled as in Algorithm 2), ReLU activations, followed by a single dense logits layer. For brevity, we do not explore here how the many additional model components present in practice may alter this algorithm: while they complicate the implementation, they do not change the fundamental concept. The algorithm consists of three main steps: implanting the detector and building the trigger (lines Line 1-4), propagating the detector signal through the model (lines 5-10), and building the disruptor (lines 11-14)." + }, + { + "type": "code_caption", + "bbox": [ + 0.526, + 0.292, + 0.858, + 0.307 + ], + "angle": 0, + "content": "Algorithm 3: Convolutional network internal lock" + }, + { + "type": "algorithm", + "bbox": [ + 0.517, + 0.308, + 0.905, + 0.593 + ], + "angle": 0, + "content": "Input : Parameters \\((\\mathcal{N},k,j,\\Delta >0,\\delta < 0)\\) as in Alg.2 Image coordinates \\((a,b)\\) for the trigger patch Disruption scale \\(s > 0\\) and offset \\(t\\in \\mathbb{R}^c\\) \n1 Define the map \\(\\phi\\) from network inputs to inputs to layer \\(j\\) \n2 Define the matrix to scalar map \\(r_{(a,b)}(M)\\coloneqq (M)_{a,b}\\) \n3 Stain \\(\\mathcal{N}\\) using Alg. 2, with the reduction map \\(r_{(a,b)}\\) \n4 Define the trigger patch \\(\\pi\\) as the part of the trigger image \\(x^{*}\\) which is in the receptive field of \\(r_{(a,b)}(v*\\phi (x^{*}))\\) \n5 To simplify presentation, swap kernels 1 and \\(k\\) in layer \\(j\\) \n6 foreach convolution layer \\(\\ell\\) after layer \\(j\\) do \n7 Let \\(W_{\\ell}\\in \\mathbb{R}^{c_{\\ell +1}\\times c_{\\ell}\\times \\kappa_{\\ell}\\times \\kappa_{\\ell}}\\) \\(b_{\\ell}\\in \\mathbb{R}^{c_{\\ell +1}}\\) be the layer \\(\\ell\\) convolutional weight and bias. \n8 Set \\((W_{\\ell})_{1,\\dots ,\\cdot} = 0 = (W_{\\ell})_{\\cdot ,1,\\dots ,\\cdot}\\) and \\((b_{\\ell})_{1} = 0\\) \n9 Set \\((W_{\\ell})_{1,1,\\dots} = 1\\) \n10 end \n11 Let \\(W_{L}\\in \\mathbb{R}^{c\\times c_{L}}\\) \\(b_{L}\\in \\mathbb{R}^{c}\\) be the logits weight and bias. \n12 Sample a disruptor vector u from \\(\\mathcal{U}(\\mathbb{S}^{c - 1})\\) \n13 Let \\(\\gamma \\in \\mathbb{R}^{c_L}\\) be the logits layer input for model input \\(x^{*}\\) \n14 Set column 1 of \\(W_{L}\\) to \\(\\frac{b_L - su + t}{(\\gamma)_1}\\) ; replace \\(b_{L}\\) with \\(su + t\\) Output: Locked model \\(\\mathcal{N}\\) and trigger patch \\(\\pi\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.734 + ], + "angle": 0, + "content": "The detector is built into a convolutional kernel as for the stain, but the trigger image is only optimised to produce a strong activation when the detector kernel is in a specific position (via the reduction map \\( r_{(a,b)} \\)). This has the effect that only a single small 'patch' of the input image (the receptive field of a single kernel in convolutional layer \\( j \\)) is optimised. Inserting this patch into (e.g. a corner of) an image during inference will therefore activate the detector." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Keeping this patch relatively small (to minimise its impact when inserted into an image) can be achieved by implanting the detector into a relatively early convolutional layer. Since the disruptor is disabled by the signal from the detector, the disruptor must be placed such that its receptive field always contains the trigger patch in the input image. For this reason, we place the disruptor into a late layer (the algorithm uses the model's logits layer, although many alternatives are possible), which depends on the entire input image, and build a conduit to propagate the detector signal to the disruptor (a more general technique to overcome this requirement is pre" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2349" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.483, + 0.198 + ], + "angle": 0, + "content": "sented below). In Algorithm 3, this conduit is a sequence of 'identity' kernels, which simply propagate the signal in the detector channel (taken as channel 1 in the algorithm for simplicity) forwards through the model. Many techniques can be used to obfuscate the presence of this channel (and therefore the nature of the lock), although perhaps the simplest is just to add small random values to the modified weights." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.199, + 0.484, + 0.487 + ], + "angle": 0, + "content": "The disruptor replaces the bias of the logits layer with a randomly sampled vector. This is disabled when the correct signal \\((\\gamma)_{1}\\) is received from the detector through the conduit by restoring the original bias using the relevant column of the weight matrix. In practice, multiple interacting detectors and disruptors could be added to different places in the model, making it much more difficult to reverse-engineer the lock. Squeeze-and-excite locking. Computer vision models for certain applications may not be suitable for internal locking if they lack good locations for a disruptor which can reliably respond to the detector signal. This is often the case in object detection, for example, where small objects may be detected using just a small portion of the image. The receptive field for these outputs may, therefore, simply not contain the trigger patch placed in the corner of the image, so its presence or absence would not influence the model's output. To overcome this challenge, we propose an alternative mechanism exploiting architectural blocks which normalise a convolution's response across the entire input image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.487, + 0.484, + 0.563 + ], + "angle": 0, + "content": "Squeeze-and-Excite (Sq-Ex) blocks [21] are a prototypical example of such an architecture, which is widely used in applications. The block is a map \\( s: \\mathbb{R}^{c \\times n \\times m} \\to \\mathbb{R}^{c \\times n \\times m} \\) which aggregates global spatial information about its input and uses this to rescale each channel. Specifically," + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.575, + 0.483, + 0.591 + ], + "angle": 0, + "content": "\\[\ns (x) = x \\odot q (x) \\text {w i t h} q: \\mathbb {R} ^ {c \\times n \\times m} \\rightarrow \\mathbb {R} ^ {c}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.594, + 0.431, + 0.61 + ], + "angle": 0, + "content": "\\[\n\\text {w h e r e} q (x) = \\sigma \\left(S _ {2} \\rho \\left(S _ {1} \\mu (x) + \\tau_ {1}\\right) + \\tau_ {2}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.623, + 0.484, + 0.698 + ], + "angle": 0, + "content": "Here \\(\\odot\\) represents channel-wise multiplication, \\(\\mu : \\mathbb{R}^{c\\times n\\times m}\\to \\mathbb{R}^c\\) computes the channel-wise mean, \\(S_{1}\\in \\mathbb{R}^{d\\times c}\\) (typically with \\(d < c\\)) and \\(S_{2}\\in \\mathbb{R}^{c\\times d}\\) are learned weight matrices, \\(\\tau_{1}\\in \\mathbb{R}^{d}\\) and \\(\\tau_{2}\\in \\mathbb{R}^{c}\\) are learned biases, and \\(\\rho\\) and \\(\\sigma\\) are the ReLU and sigmoid activation functions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.7, + 0.484, + 0.865 + ], + "angle": 0, + "content": "Such blocks provide the key ability to propagate the detector signal laterally through the model, thereby enabling much more flexibility in the placement of disruptors. For example, if the detector produces value of 0 when the trigger patch is not present, and \\(\\Delta > 0\\) when it is, the averaged response will similarly be non-zero only when the trigger patch is present. A disruptor can, therefore, be embedded within the Sq-Ex block itself (as shown in Algorithm 4), or the block can be used to propagate the unlocking signal to disruptors later in the model. The algorithm only assumes that the convolutional layer \\(j\\) has a bias term, uses ReLU" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.213 + ], + "angle": 0, + "content": "activation, and is followed by an Sq-Ex block. This lock is independent of the architecture of the rest of the model. To present the idea in a simple setting, the algorithm embeds the disruptor into the parameters \\( S_{2} \\) and \\( \\tau_{2} \\) of the Sq-Ex block itself, and \\( S_{1} \\) is simply used as a conduit for the signal. In some cases, it can be helpful to stabilise the detector unlocking signal \\( \\gamma \\) by combining the output of two or more ReLU neurons in the \\( S_{1} \\) layer, e.g. to produce a hard sigmoid [20]." + }, + { + "type": "title", + "bbox": [ + 0.523, + 0.23, + 0.783, + 0.244 + ], + "angle": 0, + "content": "Algorithm 4: Squeeze-and-excite lock" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.248, + 0.877, + 0.262 + ], + "angle": 0, + "content": "Input : Parameters \\((\\mathcal{N},k,j,\\Delta ,\\delta ,a,b,s,t)\\) as in Alg. 3" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.263, + 0.88, + 0.277 + ], + "angle": 0, + "content": "1 Define the map \\(\\phi\\) from network inputs to inputs to layer \\(j\\)." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.277, + 0.844, + 0.29 + ], + "angle": 0, + "content": "2. Stain \\(\\mathcal{N}\\) using Alg. 2, with the reduction map \\(r_{(a,b)}\\)." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.291, + 0.873, + 0.318 + ], + "angle": 0, + "content": "3 Define the trigger patch \\(\\pi\\) as the part of the trigger image \\(x^{*}\\) which is in the receptive field of \\(r_{(a,b)}(v*\\phi (x^{*}))\\)." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.319, + 0.874, + 0.332 + ], + "angle": 0, + "content": "4 To simplify presentation, swap kernels 1 and \\( k \\) in layer \\( j \\)." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.333, + 0.849, + 0.345 + ], + "angle": 0, + "content": "5 Let \\(S_{1},\\tau_{1},S_{2},\\tau_{2}\\) be the Sq-Ex block parameters (1)." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.346, + 0.863, + 0.36 + ], + "angle": 0, + "content": "6 Set \\((S_{1})_{\\cdot ,1} = 0 = (S_{1})_{1,\\cdot},(S_{1})_{1,1} = 1\\) and \\((\\tau_{1})_{1} = 0\\)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.361, + 0.814, + 0.373 + ], + "angle": 0, + "content": "7 Sample a disruptor vector \\(u\\) from \\(\\mathcal{U}(\\mathbb{S}^{c_j + 1}^{-1})\\)." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.374, + 0.866, + 0.387 + ], + "angle": 0, + "content": "8 Let \\(\\gamma = \\mu (W_{j}*\\phi (x^{*}) + b_{j})\\) , i.e. the output of \\(\\mu\\) in (1)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.388, + 0.866, + 0.405 + ], + "angle": 0, + "content": "9 Set column 1 of \\( S_{2} \\) to \\( \\frac{\\tau_{2} - su + t}{(\\gamma)_{1}} \\); replace \\( \\tau_{2} \\) with \\( su + t \\)." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.406, + 0.865, + 0.419 + ], + "angle": 0, + "content": "10 Set \\((W_{j + 1})_{\\cdot ,1,\\dots} = 0\\) (\\(W_{j + 1}\\) is the layer \\(j + 1\\) weight)." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.263, + 0.88, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.536, + 0.42, + 0.814, + 0.433 + ], + "angle": 0, + "content": "Output: Locked model \\(\\mathcal{N}\\) and trigger patch \\(\\pi\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.454, + 0.907, + 0.726 + ], + "angle": 0, + "content": "Since the Sq-Ex block is used to propagate the detector signal spatially, the lock can be made independent of the architecture of the rest of the model. Special-purpose Sq-Ex blocks can be added to existing models to implement the lock, even if they were not already present. This can be achieved without significant impact on the performance of the model by sampling \\( S_{1} \\), \\( S_{2} \\), \\( \\tau_{1} \\) and \\( \\tau_{2} \\) randomly (as demonstrated experimentally in Sec. 6). By ensuring that the entries in \\( S_{1} \\) and \\( S_{2} \\) are relatively small, the overall scaling factor produced by the (unlocked) block can be made close to constant, and can therefore be counteracted by re-scaling the weights in the next layer. The computational cost of evaluating the additional Sq-Ex block is also significantly lower than that of the convolutional layers, since spatial information is aggregated to just a single value per channel. Moreover, if the Sq-Ex block is used to deactivate disruptors placed elsewhere in the model, simply pruning a suspicious Sq-Ex block would not deactivate the lock." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.727, + 0.905, + 0.757 + ], + "angle": 0, + "content": "Extensions. Extensions to locking algorithms for GANs and ViTs are presented in Section E." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.77, + 0.73, + 0.787 + ], + "angle": 0, + "content": "5. Theoretical guarantees" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We provide two bounds on the worst-case false-positive rate of the randomly sampled detector neuron. Theorem 1 is geometric in nature, and exploits the fact that feature vector clouds typically have just a few dominant principal directions in practice. By projecting this cloud onto the random detector direction, we can bound the probability of sampling a detector and an input which leads to a false positive — i.e. a" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1Piecewise-polynomial variants of these are sometimes used, e.g. in MobileNetv3 [20] and the associated form of SSDLite [29, 38]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2350" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.213 + ], + "angle": 0, + "content": "detector response above some fixed threshold \\(\\Delta\\). In essence, this bound shows which geometric aspects of feature space are important for building models which can be successfully stained and locked: a high feature space dimension, and feature vectors with a low PCA dimension. In such a setting, it is very unlikely that a detector vector and piece of input data will be sampled independently, which respond strongly to each other. The proof of this result is given in Appendix C.1." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.239, + 0.455, + 0.253 + ], + "angle": 0, + "content": "Theorem 1 (Geometric bound on false positive rate)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.254, + 0.484, + 0.39 + ], + "angle": 0, + "content": "Let \\(S\\) and \\(T\\) be sets, and consider a map \\(\\mathcal{N}: S \\to T\\) such that \\(\\mathcal{N} = \\psi \\circ \\phi\\) for two maps \\(\\phi: S \\to \\mathbb{R}^d\\) and \\(\\psi: \\mathbb{R}^d \\to T\\). Suppose that a detector vector \\(w \\sim \\mathcal{U}(\\mathbb{S}^{d-1})\\) is sampled uniformly from the sphere. Suppose that test data \\(x\\) are independently sampled from a distribution \\(\\mathcal{D}\\) on \\(S\\) such that \\(\\mathbb{E}_x[\\phi(x)] = \\mu\\) and \\(\\operatorname{Cov}(\\phi(x))\\) has eigenvalues \\(\\{\\lambda_i\\}_{i=1}^d\\). Then, for any \\(\\Delta > \\|\\mu\\|\\), the probability of sampling a detector and data point such that the detectors' response to the data point is stronger than \\(\\Delta\\) is bounded by" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.407, + 0.483, + 0.465 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} P (w \\sim \\mathcal {U} (\\mathbb {S} ^ {d - 1}), x \\sim \\mathcal {D}: w \\cdot \\phi (x) > \\Delta) \\tag {2} \\\\ \\leq \\frac {\\sum_ {i = 1} ^ {d} \\lambda_ {i}}{2 (\\Delta - \\| \\mu \\|) ^ {2}} \\frac {d - 1}{d + 1} \\left(\\frac {\\Gamma (\\frac {d}{2})}{\\Gamma (\\frac {d + 1}{2})}\\right) ^ {2}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.482, + 0.484, + 0.755 + ], + "angle": 0, + "content": "Theorem 2 complements this with a data-driven bound on the false positive rate of a fixed detector directly in terms of observations from data. The bound can always be exactly evaluated from data observations to provide a guaranteed upper bound on the detector's worst-case false positive rate. This bound does not depend on any (typically unknown) properties of the data distribution, and evaluating it using more data will produce a tighter bound. An interesting feature of the result is that the smallest upper bound that it can produce depends only on the number of data samples available; when no false positives are observed in the sample, \\( n = 0 \\), and so the upper bound depends only on \\( m \\), the number of samples used. Unlike Theorem 1, the bound of Theorem 2 does not consider the sampling of the detector weight vector. Together, these properties make Theorem 2 an ideal tool for certifying the false positive rate of a fixed detector which has perhaps already been implanted into a model. This result is proved in Appendix C.2." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.781, + 0.465, + 0.795 + ], + "angle": 0, + "content": "Theorem 2 (Data-driven bound on false positive rate)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Let \\( S \\) and \\( T \\) be sets, and consider a map \\( \\mathcal{N}: S \\to T \\) such that \\( \\mathcal{N} = \\psi \\circ \\phi \\) for two maps \\( \\phi: S \\to \\mathbb{R}^d \\) and \\( \\psi: \\mathbb{R}^d \\to T \\). Let \\( w \\in \\mathbb{S}^{d-1} \\) and \\( \\Delta \\in \\mathbb{R} \\) be fixed. Suppose that \\( m \\) test data points \\( x_1, \\ldots, x_m \\) are independently sampled from a distribution \\( \\mathcal{D} \\) on \\( S \\). Of these \\( m \\) data points, suppose that exactly \\( n \\leq m \\) of them satisfy \\( w \\cdot \\phi(x_i) > \\Delta \\). Then, the probability of sampling a data point to which the detector's" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.796, + 0.107 + ], + "angle": 0, + "content": "response is stronger than \\(\\Delta\\) is bounded by" + }, + { + "type": "equation", + "bbox": [ + 0.559, + 0.121, + 0.905, + 0.168 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} P (x \\sim \\mathcal {D}: w \\cdot \\phi (x) > \\Delta) \\tag {3} \\\\ \\leq 1 - \\sup_ {\\epsilon > 0} \\left(\\frac {m - n}{m} - \\epsilon\\right) \\left(1 - 2 e ^ {- 2 m \\epsilon^ {2}}\\right). \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.18, + 0.714, + 0.197 + ], + "angle": 0, + "content": "6. Experimental results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.205, + 0.908, + 0.643 + ], + "angle": 0, + "content": "We demonstrate the practical performance of our staining and locking algorithms through a range of experimental results. We focus on a variety of widely used models for image classification and object detection, specifically ResNet50 [18] and VGG16 [39] for image classification, and SSDLite [29] with a MobileNet-v3 backbone [20] and Faster-RCNN [36] with a ResNet50 backbone [24] for object detection. These experiments were computed using pre-trained models available through PyTorch [3]. ImageNet [37] is used to assess the performance of image classification models, and COCO [27] is used for object detection models. We report each model's performance as the accuracy for image classification, and COCO average precision (AP) and average recall (AR) at IoU 0.5:0.95 for object detection, evaluated using the whole validation set of each benchmark. The original model's performance is also reported as a baseline. The results of each experiment are reported as the average over 50 samples of the detector per layer for image classification, and 40 for object detection (due to the increased computational cost of evaluating the model), and standard deviations are reported as shaded regions on the performance plots. The original model's performance is also reported as a baseline. To give a fair assessment, we avoided tuning the parameters used for each example, which in our experience can significantly improve the performance of the stained/unlocked models. Fig. 10-Fig. 12 (in the supplementary material) show specific examples of the performance of a locked object detection model. These experiments used the CREATE HPC facility at King's College London [22]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Staining. We inserted non-additive stains individually into several convolutional layers of each model using Algorithm 2. In each case, the neuron with the least \\(\\ell^1\\) norm weight vector was replaced with the detector neuron. Fig. 1 (ResNet50) and Fig. 2 (SSDLite) present the results. Similar results for VGG16 (Fig. 13) and Faster-RCNN (Fig. 14) are included in the supplementary materials. No false positive detections (i.e. non-trigger images with a stronger response to the detector neuron than the trigger image) occurred in any of the experiments. We plot the distribution of responses of validation data to the randomly sampled detector (\\(v\\) in Algorithms 1 and 2) for each layer in each model, along with the rates guaranteed by Theorems 1 and 2. The range of response strengths for the optimised trigger input \\(x^*\\) obtained across all detector samples is given in the figure legends. The theoretical bounds on the response strength distributions were evaluated using 2,000 images sampled from the validation" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2351" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.091, + 0.222, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.092, + 0.351, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.092, + 0.478, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.121, + 0.189, + 0.452, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.2, + 0.247, + 0.375, + 0.262 + ], + "angle": 0, + "content": "Figure 1. Staining ResNet50" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.283, + 0.222, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.284, + 0.351, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.284, + 0.478, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.38, + 0.446, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.204, + 0.429, + 0.371, + 0.443 + ], + "angle": 0, + "content": "Figure 2. Staining SSDLite" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.475, + 0.484, + 0.626 + ], + "angle": 0, + "content": "set, using non-overlapping convolutional detector positions for each image (50 per image for Theorem 1, all for Theorem 2); the empirical response distributions were computed using all non-overlapping positions for the remaining images in each validation set. We observe the excellent agreement between theory and practice, particularly for Theorem 2. Recall, however, that the bounds from the two results are not directly comparable because they use different definitions of the false positive rate, depending on whether the detector neuron is considered to be sampled or fixed." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Internal locking. We added internal locks to the ResNet50 and VGG16 models, with the detector placed in a range of convolutional layers and a single disruptor placed in the logits layer. The results for ResNet50 are shown in Fig. 3. Similar results for VGG-16 (Fig. 15) can be found in the supplementary materials. The optimised trigger patch was taken to be the receptive field of the convolutional detector when in position \\((0,0)\\). Although this overlaps with the convolution's padding region, the performance remains excellent. We report the models' performance in 6 settings: the original model, an 'edited' model in which the lock is implanted except for the disruptor itself, and the locked model; in each case, we evaluated using images with and without the trigger patch. It is clear from these results that a lock can be inserted into a pre-trained model using our algorithms with minimal impact on performance. There is a balance to be struck when placing the lock: too early in the model and the dimension of the feature space can be too low to produce" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.633, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.177, + 0.645, + 0.192 + ], + "angle": 0, + "content": "--- original model - original images original model - patched images" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.091, + 0.758, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.655, + 0.164, + 0.771, + 0.192 + ], + "angle": 0, + "content": "Conv Layer Index --- original model -original images edited model -original images edited model -patched images" + }, + { + "type": "image", + "bbox": [ + 0.774, + 0.091, + 0.893, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.783, + 0.164, + 0.872, + 0.192 + ], + "angle": 0, + "content": "Conv Layer Index --- original model -original locked model -original locked model -patchee" + }, + { + "type": "image_caption", + "bbox": [ + 0.6, + 0.207, + 0.82, + 0.221 + ], + "angle": 0, + "content": "Figure 3. Internal lock for ResNet50" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.238, + 0.626, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.322, + 0.641, + 0.339 + ], + "angle": 0, + "content": "--- original model - original images original model - patched images" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.238, + 0.755, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.646, + 0.318, + 0.779, + 0.339 + ], + "angle": 0, + "content": "--- original model - original images post-surgery model - original images post-surgery model - patched images" + }, + { + "type": "image", + "bbox": [ + 0.779, + 0.238, + 0.885, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.782, + 0.318, + 0.9, + 0.339 + ], + "angle": 0, + "content": "--- original model - original images locked model - original images locked model - patched images" + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.353, + 0.855, + 0.368 + ], + "angle": 0, + "content": "Figure 4. Squeeze-and-excite lock for ResNet50" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.385, + 0.637, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.473, + 0.637, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.569, + 0.645, + 0.584 + ], + "angle": 0, + "content": "--- original model - original images -- original model - patched images" + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.386, + 0.764, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.473, + 0.764, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.662, + 0.563, + 0.771, + 0.583 + ], + "angle": 0, + "content": "--- original model - original images \n--- edited model - original images \n--- edited model - patched images" + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.386, + 0.891, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.473, + 0.891, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.789, + 0.563, + 0.898, + 0.583 + ], + "angle": 0, + "content": "--- original model - original images locked model - original images locked model - patched images" + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.6, + 0.851, + 0.614 + ], + "angle": 0, + "content": "Figure 5. Squeeze-and-excite lock for SSDLite" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.906, + 0.688 + ], + "angle": 0, + "content": "a reliable detector; too late in the model, and the large size of the detector's receptive field (and hence the trigger patch added to images) can damage the model's performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Squeeze-and-excite locking. We demonstrate squeeze-and-excite locking in two forms: using the existing Sq-Ex blocks in the SSDLite model, and by adding new Sq-Ex blocks to each of the other models. Fig. 4 (ResNet50), Fig. 16 (VGG16), Fig. 5 (SSDLite) and Fig. 17 (Faster-RCNN) report the performance of each model in each of the 6 settings described in the previous section. The detector is placed in a range of convolutional layers, and the Sq-Ex block immediately follows it. The results clearly demonstrate both that this is an effective form of lock, and that a new block can be added to a pre-trained model without impacting the model's performance. The performance of the unlocked models is close to that of the original model in all cases. Where performance differences are observed (such as for" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2352" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.09, + 0.09, + 0.482, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.162, + 0.289, + 0.414, + 0.305 + ], + "angle": 0, + "content": "Figure 6. Staining and Locking DC-GAN" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.317, + 0.482, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.163, + 0.516, + 0.414, + 0.531 + ], + "angle": 0, + "content": "Figure 7. Staining and Locking ViT-B-16" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.557, + 0.485, + 0.648 + ], + "angle": 0, + "content": "the Faster-RCNN model in Fig. 17), this appears to be due to challenges in disabling the disruptor rather than because of the addition of the Sq-Ex block to a pre-trained model (shown by comparing the performance of the 'edited' model - containing the extra block but not the disruptor - against that of the 'unlocked' model)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.649, + 0.486, + 0.754 + ], + "angle": 0, + "content": "GAN staining and locking. Figure 6 shows experimental results for staining and locking a DC-GAN [34] model for generating human faces. This was performed using the algorithmic extensions presented in Appendix E.1, and further examples of image generations are shown in Figure 18. Full details of the experimental setup are provided in Appendix E.1.1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.756, + 0.485, + 0.817 + ], + "angle": 0, + "content": "ViT staining and locking. Figure 7 shows experimental results for staining and locking a ViT [12] model for image classification. These were obtained using the algorithms and experimental setup described in Appendix E.2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.83, + 0.206, + 0.847 + ], + "angle": 0, + "content": "7. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Staining and locking are computationally cheap. The staining and locking mechanisms we present here are very cheap to implement. The model's weights are directly modi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.183 + ], + "angle": 0, + "content": "fied, so no training or fine-tuning is required. The algorithms also require no training or validation data, meaning that models trained on sensitive data can be easily handled. The inference-time cost of the stain and lock are minimal too, being embedded within the model itself, and even the Sq-Ex block is much cheaper than a single convolutional layer." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.185, + 0.909, + 0.292 + ], + "angle": 0, + "content": "Forging attacks. In a forging attack [5], an attacker plants their own stain into a model and uses it to claim ownership. Since previous staining algorithms have required stains to be added at training time, this has not been considered a significant threat in the past. Our work reveals that functionally identical stains can easily be implanted post-training (Sec. 3.1), and forging attacks are therefore easy to perform." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.909, + 0.521 + ], + "angle": 0, + "content": "Pruning attacks. An attacker may attempt to remove a stain or lock by simply pruning it out [5, 23]. This is not a risk for our additive stains, since the detector is simply added to the original neuron, so the normal behaviour of the original neuron is retained. A non-additive stain may result in a non-responsive neuron which could be pruned, depending on the values selected for \\(\\delta\\) and \\(\\Delta\\). For the lock, on the other hand, pruning out the detector neuron would not help an attacker: without the ability to produce the correct unlocking signal, the full model performance could not be recovered. The disruptor is also difficult to remove, since it replaces the original bias vector which cannot be recovered without knowing the value of \\(\\gamma\\), the unlocking signal. The security properties of our stains and locks, including their robustness to standard attacks, are discussed further in Sec. F." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.911, + 0.659 + ], + "angle": 0, + "content": "Obfuscating locks. We believe it is advantageous to encourage a variety of hiding and obfuscation techniques to grow around the fundamental building blocks we have provided here for locking. However, in our experience the task of detecting (and removing) a lock from model weights – which are widely known to be largely uninterpretable to the human eye – is already a significant challenge after only mild attempts at hiding it. This provides an asymmetrically difficult task for a thief with little work from the owner." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.678, + 0.634, + 0.694 + ], + "angle": 0, + "content": "8. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We have introduced new staining and locking algorithms for protecting pre-trained neural network models for computer vision applications. Unlike approaches in the literature, our stains and locks can be implemented without re-training or fine-tuning the model, and do not even require access to training or validation data. We have shown how several standard staining schemas can be reimplemented using our techniques. Novel and computable theoretical guarantees on the false positive rate of our stains and locks provide users with new tools for assessing the suitability of their models for staining and locking. Experiments using standard pretrained models for image classification and object detection demonstrate the practical performance of our algorithms." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2353" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.486, + 0.212 + ], + "angle": 0, + "content": "I.T. was supported by a grant, provided by the Ministry of Economic Development of the Russian Federation in accordance with the subsidy agreement (agreement identifier 000000C313925P4G0002) and the agreement with the Ivannikov Institute for System Programming of the Russian Academy of Sciences dated June 20, 2025 No. 139-15-2025-011." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.224, + 0.188, + 0.24 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.248, + 0.486, + 0.331 + ], + "angle": 0, + "content": "[1] Yossi Adi, Carsten Baum, Moustapha Cisse, Benny Pinkas, and Joseph Keshet. Turning your weakness into a strength: watermarking deep neural networks by backdooring. In Proceedings of the 27th USENIX Conference on Security Symposium, page 1615-1631, USA, 2018. USENIX Association. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.333, + 0.484, + 0.387 + ], + "angle": 0, + "content": "[2] Manaar Alam, Sayandeep Saha, Debdeep Mukhopadhyay, and Sandip Kundu. NN-Lock: A lightweight authorization to prevent IP threats of deep learning models. J. Emerg. Technol. Comput. Syst., 18(3), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.389, + 0.487, + 0.636 + ], + "angle": 0, + "content": "[3] Jason Ansel, Edward Yang, Horace He, Natalia Gimelshein, Animesh Jain, Michael Voznesensky, Bin Bao, Peter Bell, David Berard, Evgeni Burovski, Geeta Chauhan, Anjali Chourdia, Will Constable, Alban Desmaison, Zachary DeVito, Elias Ellison, Will Feng, Jiong Gong, Michael Gschwind, Brian Hirsh, Sherlock Huang, Kshitteej Kalambarkar, Laurent Kirsch, Michael Lazos, Mario Lezcano, Yanbo Liang, Jason Liang, Yinghai Lu, CK Luk, Bert Maher, Yunjie Pan, Christian Puhrsch, Matthias Reso, Mark Saroufim, Marcos Yukio Siraichi, Helen Suk, Michael Suo, Phil Tillet, Eikan Wang, Xiaodong Wang, William Wen, Shunting Zhang, Xu Zhao, Keren Zhou, Richard Zou, Ajit Mathews, Gregory Chanan, Peng Wu, and Soumith Chintala. PyTorch 2: Faster machine learning through dynamic python bytecode transformation and graph compilation. In 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2 (ASPLOS '24). ACM, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.638, + 0.484, + 0.665 + ], + "angle": 0, + "content": "[4] Keith Ball. An elementary introduction to modern convex geometry. In Flavours of Geometry, 1997. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.666, + 0.484, + 0.705 + ], + "angle": 0, + "content": "[5] Franziska Boenisch. A systematic review on model watermarking for neural networks. Frontiers in Big Data, 4, 2021. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.707, + 0.484, + 0.747 + ], + "angle": 0, + "content": "[6] Yuling Cai, Fan Xiang, Guozhu Meng, Yinzhi Cao, and Kai Chen. AuthNet: Neural network with integrated authentication logic, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.749, + 0.484, + 0.831 + ], + "angle": 0, + "content": "[7] Huili Chen, Bita Darvish Rouhani, Cheng Fu, Jishen Zhao, and Farinaz Koushanfar. DeepMarks: A secure fingerprinting framework for digital rights management of deep learning models. In Proceedings of the 2019 on International Conference on Multimedia Retrieval, page 105–113, New York, NY, USA, 2019. Association for Computing Machinery. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.833, + 0.484, + 0.872 + ], + "angle": 0, + "content": "[8] Huili Chen, Bita Darvish Rouhani, and Farinaz Koushanfar. BlackMarks: Blackbox multibit watermarking for deep neural networks, 2019. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.874, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[9] Xiaojie Chu, Liangyu Chen, and Wenqing Yu. Nafssr: Stereo image super-resolution using nafnet. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.248, + 0.487, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.12 + ], + "angle": 0, + "content": "IEEE/CVF conference on computer vision and pattern recognition, pages 1239-1248, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.908, + 0.191 + ], + "angle": 0, + "content": "[10] Qi Cui, Ruohan Meng, Chaohui Xu, and Chip-Hong Chang. Steganographic passport: An owner and user verifiable credential for deep model ip protection without retraining. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12302-12311, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.907, + 0.261 + ], + "angle": 0, + "content": "[11] Bita Darvish Rouhani, Huili Chen, and Farinaz Koushanfar. DeepSigns: An end-to-end watermarking framework for protecting the ownership of deep neural networks. In Architectural Support for Programming Languages and Operating Systems (ASPLOS). ACM, 2019. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.908, + 0.345 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.8, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.907, + 0.403 + ], + "angle": 0, + "content": "[13] A. Dvoretzky, J. Kiefer, and J. Wolfowitz. Asymptotic minimax character of the sample distribution function and of the classical multinomial estimator. The Annals of Mathematical Statistics, 27(3):642-669, 1956. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.406, + 0.907, + 0.46 + ], + "angle": 0, + "content": "[14] Lixin Fan, Kam Woh Ng, Chee Seng Chan, and Qiang Yang. DeepIPR: Deep neural network ownership verification with passports. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6122-6139, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.462, + 0.906, + 0.517 + ], + "angle": 0, + "content": "[15] Guanhao Gan, Yiming Li, Dongxian Wu, and Shu-Tao Xia. Towards robust model watermark via reducing parametric vulnerability. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4728-4738, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.519, + 0.908, + 0.587 + ], + "angle": 0, + "content": "[16] Yifeng Gao, Yuhua Sun, Xingjun Ma, Zuxuan Wu, and YuGang Jiang. Modellock: Locking your model with a spell. In Proceedings of the 32nd ACM International Conference on Multimedia, page 11156-11165, New York, NY, USA, 2024. Association for Computing Machinery. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.59, + 0.908, + 0.644 + ], + "angle": 0, + "content": "[17] Jia Guo and Miodrag Potkonjak. Watermarking deep neural networks for embedded systems. In 2018 IEEE/ACM International Conference on Computer-Aided Design (ICCAD), pages 1-8, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.647, + 0.907, + 0.701 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.907, + 0.744 + ], + "angle": 0, + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.747, + 0.908, + 0.814 + ], + "angle": 0, + "content": "[20] Andrew Howard, Ruoming Pang, Hartwig Adam, Quoc V. Le, Mark Sandler, Bo Chen, Weijun Wang, Liang-Chieh Chen, Mingxing Tan, Grace Chu, Vijay Vasudevan, and Yukun Zhu. Searching for MobileNetV3. In ICCV, pages 1314-1324. IEEE, 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.906, + 0.858 + ], + "angle": 0, + "content": "[21] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7132-7141, 2018. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.907, + 0.899 + ], + "angle": 0, + "content": "[22] King's College London. King's Computational Research, Engineering and Technology Environment (CREATE), 2024. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2354" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.133 + ], + "angle": 0, + "content": "[23] Yue Li, Hongxia Wang, and Mauro Barni. A survey of deep neural network watermarking techniques. Neurocomputing, 461:171-193, 2021. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.484, + 0.176 + ], + "angle": 0, + "content": "[24] Yanghao Li, Saining Xie, Xinlei Chen, Piotr Dollar, Kaiming He, and Ross Girshick. Benchmarking detection transfer learning with vision transformers, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.484, + 0.245 + ], + "angle": 0, + "content": "[25] Yuxuan Li, Sarthak Kumar Maharana, and Yunhui Guo. Not just change the labels, learn the features: Watermarking deep neural networks with multi-view data. In Computer Vision – ECCV 2024, pages 273–289, Cham, 2025. Springer Nature Switzerland. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.484, + 0.329 + ], + "angle": 0, + "content": "[26] Zheng Li, Chengyu Hu, Yang Zhang, and Shanqing Guo. How to prove your model belongs to you: a blind-watermark based framework to protect intellectual property of DNN. In Proceedings of the 35th Annual Computer Security Applications Conference, page 126-137, New York, NY, USA, 2019. Association for Computing Machinery. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.484, + 0.399 + ], + "angle": 0, + "content": "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In Computer Vision - ECCV 2014, pages 740-755, Cham, 2014. Springer International Publishing. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.484, + 0.455 + ], + "angle": 0, + "content": "[28] Hanwen Liu, Zhenyu Weng, and Yuesheng Zhu. Watermarking deep neural networks with greedy residuals. In Proceedings of the 38th International Conference on Machine Learning, pages 6978-6988. PMLR, 2021. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.484, + 0.524 + ], + "angle": 0, + "content": "[29] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C. Berg. SSD: Single shot multibox detector. In Computer Vision – ECCV 2016, pages 21–37, Cham, 2016. Springer International Publishing. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.484, + 0.594 + ], + "angle": 0, + "content": "[30] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.484, + 0.635 + ], + "angle": 0, + "content": "[31] P. Massart. The tight constant in the dvoretzky-kiefer-wolfowitz inequality. The Annals of Probability, 18(3):1269-1283, 1990. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.484, + 0.678 + ], + "angle": 0, + "content": "[32] E. Le Merrer, P. Pérez, and G. Trédan. Adversarial frontier stitching for remote neural network watermarking. Neural Computing and Applications, 32:9233-9244, 2019. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.484, + 0.747 + ], + "angle": 0, + "content": "[33] Xin Mu, Yu Wang, Zhengan Huang, Junzuo Lai, Yehong Zhang, Hui Wang, and Yue Yu. Encryip: A practical encryption-based framework for model intellectual property protection. Proceedings of the AAAI Conference on Artificial Intelligence, 38(19):21438-21445, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.484, + 0.802 + ], + "angle": 0, + "content": "[34] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015.8,4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[35] Yehonathan Refael, Adam Hakim, Lev Greenberg, Tal Aviv, Satya Lokam, Ben Fishman, and Shachar Seidman. SLIP: Securing LLMs IP using weights decomposition, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[36] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(6):1137-1149, 2017. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.176 + ], + "angle": 0, + "content": "[37] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet large scale visual recognition challenge. International Journal of Computer Vision (IJCV), 115 (3):211-252, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.908, + 0.247 + ], + "angle": 0, + "content": "[38] M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L. Chen. MobileNetV2: Inverted residuals and linear bottlenecks. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4510-4520, Los Alamitos, CA, USA, 2018. IEEE Computer Society. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.248, + 0.908, + 0.288 + ], + "angle": 0, + "content": "[39] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.29, + 0.908, + 0.359 + ], + "angle": 0, + "content": "[40] Z. Sun, R. Sun, C. Liu, A. Chowdhury, L. Lu, and S. Jha. ShadowNet: A secure and efficient on-device model inference system for convolutional neural networks. In 2023 IEEE Symposium on Security and Privacy (SP), pages 1596-1612, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.36, + 0.908, + 0.43 + ], + "angle": 0, + "content": "[41] Oliver Sutton, Qinghua Zhou, Wei Wang, Desmond Higham, Alexander N. Gorban, Alexander Bastounis, and Ivan Y Tyukin. Stealth edits to large language models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.431, + 0.908, + 0.5 + ], + "angle": 0, + "content": "[42] Ruixiang Tang, Mengnan Du, and Xia Hu. Deep serial number: Computational watermark for dnn intellectual property protection. In Machine Learning and Knowledge Discovery in Databases: Applied Data Science and Demo Track, pages 157-173, Cham, 2023. Springer Nature Switzerland. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.501, + 0.908, + 0.557 + ], + "angle": 0, + "content": "[43] Ivan Y. Tyukin, Desmond J. Higham, and Alexander N. Gorbán. On adversarial examples and stealth attacks in artificial intelligence systems. In 2020 International Joint Conference on Neural Networks (IJCNN), pages 1-6, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.558, + 0.908, + 0.613 + ], + "angle": 0, + "content": "[44] Ivan Y Tyukin, Desmond J Higham, Alexander Bastounis, Eliyas Woldegeorgis, and Alexander N Gorban. The feasibility and inevitability of stealth attacks. IMA Journal of Applied Mathematics, 89(1):44-84, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.614, + 0.908, + 0.697 + ], + "angle": 0, + "content": "[45] Yusuke Uchida, Yuki Nagai, Shigeyuki Sakazawa, and Shin'ichi Satoh. Embedding watermarks into deep neural networks. In Proceedings of the 2017 ACM on International Conference on Multimedia Retrieval, page 269-277, New York, NY, USA, 2017. Association for Computing Machinery. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.699, + 0.908, + 0.753 + ], + "angle": 0, + "content": "[46] Jiangfeng Wang, Hanzhou Wu, Xinpeng Zhang, and Yuwei Yao. Watermarking in deep neural networks via error backpropagation. *Electronic Imaging*, 32(4):22-1-22-1, 2020. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.753 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2355" + } + ] +] \ No newline at end of file diff --git a/2025/Staining and Locking Computer Vision Models Without Retraining/6d43a5d3-0678-494f-aae6-04c44e76e1c5_origin.pdf b/2025/Staining and Locking Computer Vision Models Without Retraining/6d43a5d3-0678-494f-aae6-04c44e76e1c5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e8f03cb24a05bc0704b8efc0724d3847e6ec3675 --- /dev/null +++ b/2025/Staining and Locking Computer Vision Models Without Retraining/6d43a5d3-0678-494f-aae6-04c44e76e1c5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb8b39d3468b1fff1e77fcd3aa43017b8d064be3cefea3e8924197df38007852 +size 2927759 diff --git a/2025/Staining and Locking Computer Vision Models Without Retraining/full.md b/2025/Staining and Locking Computer Vision Models Without Retraining/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3743e047afb03f4d3aa61226881b5ba9fc2532b8 --- /dev/null +++ b/2025/Staining and Locking Computer Vision Models Without Retraining/full.md @@ -0,0 +1,357 @@ +# Staining and locking computer vision models without retraining + +Oliver J. Sutton* + +Synoptix Ltd. and + +King's College London + +Qinghua Zhou* + +King's College London + +qinghua.zhou@kcl.ac.uk + +George Leete + +Synoptix Ltd. + +oliver.sutton@kcl.ac.uk + +Alexander N. Gorban + +Central Univ. Moscow + +Ivan Y. Tyukin +King's College London and AI Center, Bol'shoi Bul'var 30, Moscow +and ISP RAS Research Center for Trusted Artificial Intelligence + +# Abstract + +We introduce new methods of staining and locking computer vision models, to protect their owners' intellectual property. Staining, also known as watermarking, embeds secret behaviour into a model which can later be used to identify it, while locking aims to make a model unusable unless a secret trigger is inserted into input images. Unlike existing methods, our algorithms can be used to stain and lock pretrained models without requiring fine-tuning or retraining, and come with provable, computable guarantees bounding their worst-case false positive rates. The stain and lock are implemented by directly modifying a small number of the model's weights and have minimal impact on the (unlocked) model's performance. Locked models are unlocked by inserting a small 'trigger patch' into the corner of the input image. We present experimental results showing the efficacy of our methods and demonstrating their practical performance on a variety of computer vision models. + +# 1. Introduction + +Designing, training and validating a computer vision model can be an expensive undertaking. The effect on an organisation if such a valuable investment is leaked or stolen, can be deleterious, particularly since model weights can be difficult for a human judge to distinguish confidently. Developing methods for protecting such intellectual property, which may be deployed on a vulnerable edge device, is therefore crucial. + +Two natural approaches that a model's owner may wish to take are either to imprint an identifiable $\text{stain}$ within its weights, or to add a lock to make the model unusable without adding a separate secret unlocking trigger to the model's input. Staining, also known as watermarking in the liter- + +ature [5, 23], can be viewed as a form of 'copyright trap'. By embedding specific (unnatural) behavioural quirks into the model, an identifying fingerprint is produced which can later be used to determine the provenance of a stolen or copied model. Locking, on the other hand, has a more direct aim [6, 14, 16]: unless the thief knows the secret trigger, they are unable to use the model to its full capability. + +Here, we present new methods for both staining and locking neural network models in computer vision. Unlike existing approaches in the literature (reviewed in Sec. 2), our algorithms do not require model training or fine-tuning. They directly modify the model's weights, and can therefore be used to stain or lock pre-trained models with minimal performance impact. Since re-training is not required, a single model can be supported and updated across multiple customers, while providing each with a separate stain or lock. A further key advantage of our staining and locking techniques is that implementing them does not require access to training or validation data. We also provide two distinct novel theoretical guarantees on the worst-case false positive rate of our staining and locking algorithms, analogues of which do not appear to be present in the literature. + +Staining methods are broadly composed of two parts: a schema describing the information encoded in the stain and how it may later be read, and a mechanism through which this information is implanted into the model. Numerous schemas have been proposed in the literature, e.g. hiding a message in the model's weights, its activations or its class predictions for specific trigger inputs. Until now, however, only two families of mechanisms have been available, both requiring model retraining: through special loss functions [7, 8, 11, 28, 45, 46], or by manipulating the training data [1, 17, 26, 32]. Our novel training-free staining mechanism implants highly selective detector neurons into the model, which produce a strong output only when a special trigger is present in the model's input. Sec. 3.1 shows how several standard staining + +schemas may be implemented using this mechanism. + +Our lock may be seen as an 'active' variant of the stain mechanism. Additional disruptors are also implanted to pollute the model's activations and prevent it from performing well. A specially constructed 'trigger patch' is added to one corner of the input image during inference is detected by the detector neuron. The detector neuron then produces a signal which deactivates the disruptors, recovering the full model performance. Examples of such trigger patches are shown in Appendix B. Our locking methods are presented with a focus on their fundamental building blocks, which can then be used to design bespoke algorithms to lock specific models. + +Our goal is to present these new techniques as clearly as possible in a simple setting. For this, we focus primarily on convolutional models, although extensions of the core algorithms to generative adversarial networks (GANs) and vision transformers (ViTs) are presented in Appendix E. Our staining methods are generic and apply directly to many different model architectures. Our basic lock is more architecture-dependent since it interacts with the fundamental action of the model. To provide architectural-independence, we also show how the lock may be embedded within a Squeeze-and-Excite block [21], a common component of many models. We demonstrate, in Sec. 6, that such a block can be added to many pre-trained models with minimal impact on performance. For simplicity and brevity, we do not focus too strongly on techniques for hiding and obfuscating the stains and locks; we provide ideas for achieving this throughout, but do not pursue it in detail to present the core concepts as clearly as possible in the space available. For practical success, we also feel it is preferable for a diversity of obfuscation techniques and algorithmic variations to proliferate. Instead, we focus on the simplest case where a single detector and disruptor are added to the model; in practice, multiple detectors should be combined to deactivate multiple disruptors. This straightforward extension of the methods presented here makes reverse-engineering the lock a significant challenge. + +Our approaches to staining and locking are inspired by stealth attacks [41, 43, 44], which have recently been shown to enable an attacker to specify the output of models from many modern families for a given piece of input data by directly modifying just a few weights. By harnessing this core sensitivity of modern models, and exploiting the fundamental concentration properties of their feature spaces, the highly selective detector neuron of our stain and lock can be implanted without retraining the model. This detector forms the heart of our methods, around which completely new mechanisms are built to facilitate staining and locking. This link means that the novel theoretical guarantees we present in Sec. 5 can be adapted back to study stealth attacks, and complement those available in the literature [41, 43, 44]. + +The paper is organised as follows. Sec. 2 places our + +approach in the context of related work. Our staining and locking algorithms are presented in Sec. 3 and Sec. 4. Sec. 5 presents novel theoretical guarantees on the detector's false positive rate. Experimental results demonstrate the practical performance of our algorithms in Sec. 6. A discussion of our findings is in Sec. 7, and some final conclusions are given in Sec. 8. Our mathematical notation is summarised in Appendix A (in the supplementary material). + +# 2. Related work + +Staining. Several staining methods have been proposed in the literature and mainly fall within two thematic groups; see [5, 23] for a thorough review. Loss-based methods encode a secret message into the model's weights during training using a special loss function. This message may then be revealed using a special trigger input to the network. Common approaches encode the message as a binary string in the model's weights [7, 45], activations [28] or outputs [8, 11], either directly or to be decoded by a special model [10, 46]. Backdoor methods, on the other hand, manipulate training data to produce a unique behavioural signature. One method is to force the model to learn specific non-natural labels for certain distinctive inputs [1, 15, 25, 32]. Alternatively, other work has considered the task of training the model to learn to associate a specific class with certain deliberate changes to natural images [17], such as the presence of a logo [26]. + +Locking. Two key families of approaches have been proposed in the literature. Intrinsic methods embed the lock into the model itself during training. This may be by training the model's output to depend on an additional (secret) input [14, 42], a cryptographic scheme [33], a secret global mask applied to inputs [6], or by converting training and inference images using a 'style-transfer' model (e.g. to oil paintings) [16]. Extrinsic methods offer additional orthogonal security to intrinsic methods via techniques from outside the neural network domain, such as encryption [2] or secure infrastructure [35, 40]. + +Novelty. In contrast to those in the literature, our staining and locking algorithms directly modify a model's weights, so require no model training, fine-tuning or external infrastructure. This is a key advantage because a single base model can be supplied to many clients with different locks/stains, without re-training each time. Moreover, re-training a model carries the inherent risk of changing its behaviour in unidentifiable ways (especially if the model is tuned to give incorrect responses to natural images as in backdoor methods), which is avoided through our direct approach. Being embedded in the model itself, a lock also remains active even if an adversary gains access to the hardware on which it runs. Our provable performance guarantees further differentiate our methods; comparable results are not available for any of the methods in the literature. + +# 3. Staining + +This section presents the building blocks of our novel training-free staining mechanism. The stain implants a highly selective detector neuron into the model, which is activated by a specially optimised trigger input to the model. We show how this mechanism can be used to implement standard staining schemas in Sec. 3.1. + +Multilayer perceptrons. Algorithm 1 outlines our mechanism for staining a generic multilayer perceptron neural network. A detector neuron with a weight vector sampled from $\mathcal{U}(\mathbb{S}^{m - 1})$ (i.e. uniformly from the surface of a sphere) is inserted into the model. Concentration properties in the model's feature space ensure that this randomly sampled neuron is very unlikely to respond strongly to natural inputs (see Sec. 5). It is possible, however, to optimise a trigger input (e.g. using gradient descent) to strongly activate the detector. The strong response to the trigger input, therefore, becomes an identifying characteristic of the model and can be used as the basis for implementing standard staining schemas. To ensure that the detector neuron has no impact on the model's performance, the parameters $\delta$ and $\Delta$ can be used in conjunction with the activation function to ensure that the detector's response is zero for natural inputs. + +The optimisation of the trigger input $x^{*}$ is important here: a strong detector response to $x^{*}$ is very distinct from its response to natural inputs, and the guarantees on the false positive rate are improved (see Theorem 1). It also means the response $\Delta$ can be achieved by scaling down the weight vector, which can be beneficial for evading detection. + +Additive staining. Alternatively, an additive stain adds the detector weight to an existing neuron's weight. Once again, the detector component of the weight likely only responds weakly to natural inputs, meaning that it has little impact on the model's operation. The strong response to the trigger input, however, is still preserved. Since this method does not produce a 'silent' detector neuron, it is much less likely that the detector will be identified and removed by an adversary. + +Convolutional networks. Algorithm 2 demonstrates how this framework applies to a convolutional neural network. For simplicity, this algorithm focuses on the non-additive stain, but also applies to models incorporating batch normalisation. A single convolutional kernel is used as the detector, and a reduction operator $r$ is used for the optimisation to convert the kernel's activation map into a single scalar response value. This could, for example, be the response of the kernel at a particular position in the image, or its average response. Larger kernels are helpful in reducing the false positive rate, as made explicit by Theorem 1: because a $\kappa \times \kappa$ convolutional kernel with $c$ input channels has $\kappa^2 c$ parameters, large kernels operate in a higher dimensional setting. + +Extensions. Extensions to apply these staining algorithms to GANs and ViTs are presented in Section E. + +Algorithm 1: Multilayer perceptron staining +Input:Trained multilayer perceptron network $\mathcal{N}$ Index $k$ of neuron to stain in layer $j$ Detector response to trigger $\Delta \in \mathbb{R}$ Desired non-trigger detector response $\delta \in \mathbb{R}$ Whether the stain should be additive 1 Define the map $\phi$ from network inputs to inputs to layer $j$ 2 Let $W_{j}\in \mathbb{R}^{n\times m}$ $b_{j}\in \mathbb{R}^{n}$ be the layer $j$ weight and bias. 3 Sample a detector weight vector $v\sim \mathcal{U}(\mathbb{S}^{m - 1})$ 4 Optimise the trigger input $x^{*}\in \arg \max_{z\in S}v\cdot \phi (z)$ 5 if stain is additive then 6 Let $\beta$ be entry $k$ of $b_{j}$ and $w$ be row $k$ of $W_{j}$ 7 Define $u = w + \frac{\Delta - \beta - w\cdot\phi(x^{*})}{v\cdot\phi(x^{*})} v.$ 8 else 9 Define $\beta = \delta$ and $u = \frac{\Delta - \beta}{v\cdot\phi(x^*)} v.$ 10 Replace row $k$ of $W_{j}$ with $u$ and entry $k$ of $b_{j}$ with $\beta$ Output: Stained model $\mathcal{N}$ and trigger input $x^{*}$ + +Algorithm 2: Convolutional network staining +Input:Trained convolutional network $\mathcal{N}$ Parameters $(k,j,\Delta ,\delta)$ as in Alg.1 Matrix to scalar reduction operation $r$ Whether layer $j$ performs batch normalisation +1 Define the map $\phi :S = \mathbb{R}^{3\times m_0\times n_0}\to \mathbb{R}^{c_j\times m_j\times n_j}$ from network inputs to inputs to layer $j$ +2 Let $W_{j}\in \mathbb{R}^{c_{j} + 1\times c_{j}\times \kappa_{j}\times \kappa_{j}}$ be the layer $j$ weight, with kernel shape $\kappa_{j}\times \kappa_{j}$ +3 Sample a kernel $v\in \mathbb{R}^{c_j\times \kappa_j\times \kappa_j}$ from $\mathcal{U}(\mathbb{S}^{c_j\kappa_j^2 -1})$ viewed as a tensor with shape $(c_{j},\kappa_{j},\kappa_{j})$ +4 Optimise the trigger image $x^{*}\in \arg \max_{z\in S}r(v*\phi (z))$ +5 if layer $j$ performs batch normalisation then +6 Let $\mu_j,\sigma_j^2,w_j,b_j\in \mathbb{R}^{c_j + 1}$ be the mean, variance, weight and bias of the layer $j$ batch normalisation. +7 Define $\alpha = \frac{(\Delta - \delta)(\sigma_j)_k}{(w_j)_kr(v*\phi(x^*))}$ and $\beta = \delta +\frac{(w_j)_k(\mu_j)_k}{(\sigma_j)_k}$ +8 else +9 Let $b_{j}\in \mathbb{R}^{c_{j} + 1}$ be the bias of layer $j$ +10 Define $\alpha = \frac{(\Delta - \delta)}{r(v*\phi(x^*))}$ and $\beta = \delta$ +11 Replace $(W_{j})_{k,\dots ,}$ with $\alpha v$ ,and entry $k$ of $b_{j}$ with $\beta$ Output: Stained model $\mathcal{N}$ and trigger input $x^{*}$ + +# 3.1. Implementing standard staining schemas + +The algorithms above introduce a novel training-free staining mechanism, and here we briefly discuss how it can be used to implement standard schemas from the literature. + +Weight staining. Originally proposed by [7, 45], this schema encodes a secret message as a binary string, which is then implanted into the model as the signs of the weights of a target neuron (e.g. $-$ encodes 0 and $+$ encodes 1). The stain can be checked by decoding the message from the signs of the neuron's weight vector's components. This can be embedded into a non-additive stain by sampling the detector vector $v$ from the orthant of the sphere with the specified + +component signs, and the rest of the staining algorithm can then proceed as usual. This restriction on the sampling may increase the false positive rate of the detector (by decreasing the entropy of the distribution), but this should not present a problem if the dimension is sufficiently high. + +Activation staining. This family of schemas, introduced in [28], instead embeds the secret message into the signs of the model's activations for a target input, rather than the weights. Our additive stain can be used to directly implement this. Suppose a detector vector is sampled for a given layer of the model, and a trigger input is optimised to produce a strong response to this detector vector. Then, this detector vector is weighted and added to each neuron in the layer, to ensure that the response to the trigger image from that neuron has the specified sign. + +Output staining. Another standard schema family specifies (non-natural) class predictions for certain trigger input images [1, 8, 11, 32]. The stain can, therefore, be verified without access to model weights or activations. This can be implemented post-training by starting with a standard non-additive stain from Algorithm 1. If the stain detector neuron is implanted into row $k$ of the weight matrix in layer $j$ , then the detector's response (which will be zero for non-trigger inputs after a ReLU activation if $\delta \ll 0$ ) will multiply the entries of column $k$ in layer $j + 1$ . For any chosen output class $C$ , it is typically possible to find a vector (e.g. using gradient descent) which, if it were the output of layer $j + 1$ , would cause the model to confidently predict class $C$ . Replacing column $k$ of layer $j + 1$ with this vector, therefore, causes the model to predict the chosen class for the trigger input; the selectivity of the detector ensures that the new column will not otherwise affect the model's performance. This technique extends analogously to convolutional stains. + +# 4. Locking + +Our locking methods build on the staining mechanism by inserting additional disruptors into the model to pollute its latent activations. The signal from the stain's detector neuron is used to disable the disruptors when the trigger is present in the input. Fig. 8 and Fig. 9 (in the supplementary material) show examples of images containing these patches. We show how the underlying structure of a convolutional detector can be exploited to produce a transferable trigger 'patch' which can be inserted into any image to unlock the full model performance. We present two variants of the locking mechanism. Internal locking embeds the lock into the weights of convolutional networks with a particular structure, while squeeze-and-excite locking uses a standard Squeeze-and-Excite block [21], common in computer vision models, as the foundation of a more generic procedure which can be added to any convolutional model. + +Internal locking. Algorithm 3 adds an internal lock to a + +convolutional neural network. To present the idea in the simplest possible setting, we suppose that the model consists of a sequence of convolutional layers with biases (batch normalisation can be handled as in Algorithm 2), ReLU activations, followed by a single dense logits layer. For brevity, we do not explore here how the many additional model components present in practice may alter this algorithm: while they complicate the implementation, they do not change the fundamental concept. The algorithm consists of three main steps: implanting the detector and building the trigger (lines Line 1-4), propagating the detector signal through the model (lines 5-10), and building the disruptor (lines 11-14). + +Algorithm 3: Convolutional network internal lock +Input : Parameters $(\mathcal{N},k,j,\Delta >0,\delta < 0)$ as in Alg.2 Image coordinates $(a,b)$ for the trigger patch Disruption scale $s > 0$ and offset $t\in \mathbb{R}^c$ +1 Define the map $\phi$ from network inputs to inputs to layer $j$ +2 Define the matrix to scalar map $r_{(a,b)}(M)\coloneqq (M)_{a,b}$ +3 Stain $\mathcal{N}$ using Alg. 2, with the reduction map $r_{(a,b)}$ +4 Define the trigger patch $\pi$ as the part of the trigger image $x^{*}$ which is in the receptive field of $r_{(a,b)}(v*\phi (x^{*}))$ +5 To simplify presentation, swap kernels 1 and $k$ in layer $j$ +6 foreach convolution layer $\ell$ after layer $j$ do +7 Let $W_{\ell}\in \mathbb{R}^{c_{\ell +1}\times c_{\ell}\times \kappa_{\ell}\times \kappa_{\ell}}$ $b_{\ell}\in \mathbb{R}^{c_{\ell +1}}$ be the layer $\ell$ convolutional weight and bias. +8 Set $(W_{\ell})_{1,\dots ,\cdot} = 0 = (W_{\ell})_{\cdot ,1,\dots ,\cdot}$ and $(b_{\ell})_{1} = 0$ +9 Set $(W_{\ell})_{1,1,\dots} = 1$ +10 end +11 Let $W_{L}\in \mathbb{R}^{c\times c_{L}}$ $b_{L}\in \mathbb{R}^{c}$ be the logits weight and bias. +12 Sample a disruptor vector u from $\mathcal{U}(\mathbb{S}^{c - 1})$ +13 Let $\gamma \in \mathbb{R}^{c_L}$ be the logits layer input for model input $x^{*}$ +14 Set column 1 of $W_{L}$ to $\frac{b_L - su + t}{(\gamma)_1}$ ; replace $b_{L}$ with $su + t$ Output: Locked model $\mathcal{N}$ and trigger patch $\pi$ + +The detector is built into a convolutional kernel as for the stain, but the trigger image is only optimised to produce a strong activation when the detector kernel is in a specific position (via the reduction map $r_{(a,b)}$ ). This has the effect that only a single small 'patch' of the input image (the receptive field of a single kernel in convolutional layer $j$ ) is optimised. Inserting this patch into (e.g. a corner of) an image during inference will therefore activate the detector. + +Keeping this patch relatively small (to minimise its impact when inserted into an image) can be achieved by implanting the detector into a relatively early convolutional layer. Since the disruptor is disabled by the signal from the detector, the disruptor must be placed such that its receptive field always contains the trigger patch in the input image. For this reason, we place the disruptor into a late layer (the algorithm uses the model's logits layer, although many alternatives are possible), which depends on the entire input image, and build a conduit to propagate the detector signal to the disruptor (a more general technique to overcome this requirement is pre + +sented below). In Algorithm 3, this conduit is a sequence of 'identity' kernels, which simply propagate the signal in the detector channel (taken as channel 1 in the algorithm for simplicity) forwards through the model. Many techniques can be used to obfuscate the presence of this channel (and therefore the nature of the lock), although perhaps the simplest is just to add small random values to the modified weights. + +The disruptor replaces the bias of the logits layer with a randomly sampled vector. This is disabled when the correct signal $(\gamma)_{1}$ is received from the detector through the conduit by restoring the original bias using the relevant column of the weight matrix. In practice, multiple interacting detectors and disruptors could be added to different places in the model, making it much more difficult to reverse-engineer the lock. Squeeze-and-excite locking. Computer vision models for certain applications may not be suitable for internal locking if they lack good locations for a disruptor which can reliably respond to the detector signal. This is often the case in object detection, for example, where small objects may be detected using just a small portion of the image. The receptive field for these outputs may, therefore, simply not contain the trigger patch placed in the corner of the image, so its presence or absence would not influence the model's output. To overcome this challenge, we propose an alternative mechanism exploiting architectural blocks which normalise a convolution's response across the entire input image. + +Squeeze-and-Excite (Sq-Ex) blocks [21] are a prototypical example of such an architecture, which is widely used in applications. The block is a map $s: \mathbb{R}^{c \times n \times m} \to \mathbb{R}^{c \times n \times m}$ which aggregates global spatial information about its input and uses this to rescale each channel. Specifically, + +$$ +s (x) = x \odot q (x) \text {w i t h} q: \mathbb {R} ^ {c \times n \times m} \rightarrow \mathbb {R} ^ {c}, \tag {1} +$$ + +$$ +\text {w h e r e} q (x) = \sigma \left(S _ {2} \rho \left(S _ {1} \mu (x) + \tau_ {1}\right) + \tau_ {2}\right). +$$ + +Here $\odot$ represents channel-wise multiplication, $\mu : \mathbb{R}^{c\times n\times m}\to \mathbb{R}^c$ computes the channel-wise mean, $S_{1}\in \mathbb{R}^{d\times c}$ (typically with $d < c$ ) and $S_{2}\in \mathbb{R}^{c\times d}$ are learned weight matrices, $\tau_{1}\in \mathbb{R}^{d}$ and $\tau_{2}\in \mathbb{R}^{c}$ are learned biases, and $\rho$ and $\sigma$ are the ReLU and sigmoid activation functions. + +Such blocks provide the key ability to propagate the detector signal laterally through the model, thereby enabling much more flexibility in the placement of disruptors. For example, if the detector produces value of 0 when the trigger patch is not present, and $\Delta > 0$ when it is, the averaged response will similarly be non-zero only when the trigger patch is present. A disruptor can, therefore, be embedded within the Sq-Ex block itself (as shown in Algorithm 4), or the block can be used to propagate the unlocking signal to disruptors later in the model. The algorithm only assumes that the convolutional layer $j$ has a bias term, uses ReLU + +activation, and is followed by an Sq-Ex block. This lock is independent of the architecture of the rest of the model. To present the idea in a simple setting, the algorithm embeds the disruptor into the parameters $S_{2}$ and $\tau_{2}$ of the Sq-Ex block itself, and $S_{1}$ is simply used as a conduit for the signal. In some cases, it can be helpful to stabilise the detector unlocking signal $\gamma$ by combining the output of two or more ReLU neurons in the $S_{1}$ layer, e.g. to produce a hard sigmoid [20]. + +# Algorithm 4: Squeeze-and-excite lock + +Input : Parameters $(\mathcal{N},k,j,\Delta ,\delta ,a,b,s,t)$ as in Alg. 3 + +1 Define the map $\phi$ from network inputs to inputs to layer $j$ . +2. Stain $\mathcal{N}$ using Alg. 2, with the reduction map $r_{(a,b)}$ . +3 Define the trigger patch $\pi$ as the part of the trigger image $x^{*}$ which is in the receptive field of $r_{(a,b)}(v*\phi (x^{*}))$ . +4 To simplify presentation, swap kernels 1 and $k$ in layer $j$ . +5 Let $S_{1},\tau_{1},S_{2},\tau_{2}$ be the Sq-Ex block parameters (1). +6 Set $(S_{1})_{\cdot ,1} = 0 = (S_{1})_{1,\cdot},(S_{1})_{1,1} = 1$ and $(\tau_{1})_{1} = 0$ +7 Sample a disruptor vector $u$ from $\mathcal{U}(\mathbb{S}^{c_j + 1}^{-1})$ . +8 Let $\gamma = \mu (W_{j}*\phi (x^{*}) + b_{j})$ , i.e. the output of $\mu$ in (1) +9 Set column 1 of $S_{2}$ to $\frac{\tau_{2} - su + t}{(\gamma)_{1}}$ ; replace $\tau_{2}$ with $su + t$ . +10 Set $(W_{j + 1})_{\cdot ,1,\dots} = 0$ ( $W_{j + 1}$ is the layer $j + 1$ weight). + +Output: Locked model $\mathcal{N}$ and trigger patch $\pi$ + +Since the Sq-Ex block is used to propagate the detector signal spatially, the lock can be made independent of the architecture of the rest of the model. Special-purpose Sq-Ex blocks can be added to existing models to implement the lock, even if they were not already present. This can be achieved without significant impact on the performance of the model by sampling $S_{1}$ , $S_{2}$ , $\tau_{1}$ and $\tau_{2}$ randomly (as demonstrated experimentally in Sec. 6). By ensuring that the entries in $S_{1}$ and $S_{2}$ are relatively small, the overall scaling factor produced by the (unlocked) block can be made close to constant, and can therefore be counteracted by re-scaling the weights in the next layer. The computational cost of evaluating the additional Sq-Ex block is also significantly lower than that of the convolutional layers, since spatial information is aggregated to just a single value per channel. Moreover, if the Sq-Ex block is used to deactivate disruptors placed elsewhere in the model, simply pruning a suspicious Sq-Ex block would not deactivate the lock. + +Extensions. Extensions to locking algorithms for GANs and ViTs are presented in Section E. + +# 5. Theoretical guarantees + +We provide two bounds on the worst-case false-positive rate of the randomly sampled detector neuron. Theorem 1 is geometric in nature, and exploits the fact that feature vector clouds typically have just a few dominant principal directions in practice. By projecting this cloud onto the random detector direction, we can bound the probability of sampling a detector and an input which leads to a false positive — i.e. a + +detector response above some fixed threshold $\Delta$ . In essence, this bound shows which geometric aspects of feature space are important for building models which can be successfully stained and locked: a high feature space dimension, and feature vectors with a low PCA dimension. In such a setting, it is very unlikely that a detector vector and piece of input data will be sampled independently, which respond strongly to each other. The proof of this result is given in Appendix C.1. + +# Theorem 1 (Geometric bound on false positive rate) + +Let $S$ and $T$ be sets, and consider a map $\mathcal{N}: S \to T$ such that $\mathcal{N} = \psi \circ \phi$ for two maps $\phi: S \to \mathbb{R}^d$ and $\psi: \mathbb{R}^d \to T$ . Suppose that a detector vector $w \sim \mathcal{U}(\mathbb{S}^{d-1})$ is sampled uniformly from the sphere. Suppose that test data $x$ are independently sampled from a distribution $\mathcal{D}$ on $S$ such that $\mathbb{E}_x[\phi(x)] = \mu$ and $\operatorname{Cov}(\phi(x))$ has eigenvalues $\{\lambda_i\}_{i=1}^d$ . Then, for any $\Delta > \|\mu\|$ , the probability of sampling a detector and data point such that the detectors' response to the data point is stronger than $\Delta$ is bounded by + +$$ +\begin{array}{l} P (w \sim \mathcal {U} (\mathbb {S} ^ {d - 1}), x \sim \mathcal {D}: w \cdot \phi (x) > \Delta) \tag {2} \\ \leq \frac {\sum_ {i = 1} ^ {d} \lambda_ {i}}{2 (\Delta - \| \mu \|) ^ {2}} \frac {d - 1}{d + 1} \left(\frac {\Gamma (\frac {d}{2})}{\Gamma (\frac {d + 1}{2})}\right) ^ {2}. \\ \end{array} +$$ + +Theorem 2 complements this with a data-driven bound on the false positive rate of a fixed detector directly in terms of observations from data. The bound can always be exactly evaluated from data observations to provide a guaranteed upper bound on the detector's worst-case false positive rate. This bound does not depend on any (typically unknown) properties of the data distribution, and evaluating it using more data will produce a tighter bound. An interesting feature of the result is that the smallest upper bound that it can produce depends only on the number of data samples available; when no false positives are observed in the sample, $n = 0$ , and so the upper bound depends only on $m$ , the number of samples used. Unlike Theorem 1, the bound of Theorem 2 does not consider the sampling of the detector weight vector. Together, these properties make Theorem 2 an ideal tool for certifying the false positive rate of a fixed detector which has perhaps already been implanted into a model. This result is proved in Appendix C.2. + +# Theorem 2 (Data-driven bound on false positive rate) + +Let $S$ and $T$ be sets, and consider a map $\mathcal{N}: S \to T$ such that $\mathcal{N} = \psi \circ \phi$ for two maps $\phi: S \to \mathbb{R}^d$ and $\psi: \mathbb{R}^d \to T$ . Let $w \in \mathbb{S}^{d-1}$ and $\Delta \in \mathbb{R}$ be fixed. Suppose that $m$ test data points $x_1, \ldots, x_m$ are independently sampled from a distribution $\mathcal{D}$ on $S$ . Of these $m$ data points, suppose that exactly $n \leq m$ of them satisfy $w \cdot \phi(x_i) > \Delta$ . Then, the probability of sampling a data point to which the detector's + +response is stronger than $\Delta$ is bounded by + +$$ +\begin{array}{l} P (x \sim \mathcal {D}: w \cdot \phi (x) > \Delta) \tag {3} \\ \leq 1 - \sup_ {\epsilon > 0} \left(\frac {m - n}{m} - \epsilon\right) \left(1 - 2 e ^ {- 2 m \epsilon^ {2}}\right). \\ \end{array} +$$ + +# 6. Experimental results + +We demonstrate the practical performance of our staining and locking algorithms through a range of experimental results. We focus on a variety of widely used models for image classification and object detection, specifically ResNet50 [18] and VGG16 [39] for image classification, and SSDLite [29] with a MobileNet-v3 backbone [20] and Faster-RCNN [36] with a ResNet50 backbone [24] for object detection. These experiments were computed using pre-trained models available through PyTorch [3]. ImageNet [37] is used to assess the performance of image classification models, and COCO [27] is used for object detection models. We report each model's performance as the accuracy for image classification, and COCO average precision (AP) and average recall (AR) at IoU 0.5:0.95 for object detection, evaluated using the whole validation set of each benchmark. The original model's performance is also reported as a baseline. The results of each experiment are reported as the average over 50 samples of the detector per layer for image classification, and 40 for object detection (due to the increased computational cost of evaluating the model), and standard deviations are reported as shaded regions on the performance plots. The original model's performance is also reported as a baseline. To give a fair assessment, we avoided tuning the parameters used for each example, which in our experience can significantly improve the performance of the stained/unlocked models. Fig. 10-Fig. 12 (in the supplementary material) show specific examples of the performance of a locked object detection model. These experiments used the CREATE HPC facility at King's College London [22]. + +Staining. We inserted non-additive stains individually into several convolutional layers of each model using Algorithm 2. In each case, the neuron with the least $\ell^1$ norm weight vector was replaced with the detector neuron. Fig. 1 (ResNet50) and Fig. 2 (SSDLite) present the results. Similar results for VGG16 (Fig. 13) and Faster-RCNN (Fig. 14) are included in the supplementary materials. No false positive detections (i.e. non-trigger images with a stronger response to the detector neuron than the trigger image) occurred in any of the experiments. We plot the distribution of responses of validation data to the randomly sampled detector ( $v$ in Algorithms 1 and 2) for each layer in each model, along with the rates guaranteed by Theorems 1 and 2. The range of response strengths for the optimised trigger input $x^*$ obtained across all detector samples is given in the figure legends. The theoretical bounds on the response strength distributions were evaluated using 2,000 images sampled from the validation + +![](images/ee71845063230cf288ff8fd0e14cb2c2b1720c7fc65d922a1ca214aa244cc829.jpg) + +![](images/4328dae7819cb5661012cb281d35f38234f36097836bcfce97ec8b9a75aa7f91.jpg) + +![](images/5c8d12464024e072f59bcfad6f1fa9cb9054b1c1f8ffb2b2b779b5724588ea75.jpg) + +![](images/de5f9568f7960508126ead577529e5515672274a5b7f6ec0c4acf9bb284c20f9.jpg) +Figure 1. Staining ResNet50 + +![](images/da0de8143cd42824bb07b8c6e4abeff93b2d3347d05ade53e83126a99a6a543e.jpg) + +![](images/32a7a6d4cac5ffd3997f876a07ea7048c4cc8a0665a15973cd38d3af07a11aa3.jpg) + +![](images/0b0c9385d3f2d8711724fb325449e92fe710ceeb954b5834c27dace9e0c93a57.jpg) + +![](images/4436a1b7c8b93c0a835907eb940117d4f69ecc907758c37772a52d3d51e10a41.jpg) +Figure 2. Staining SSDLite + +set, using non-overlapping convolutional detector positions for each image (50 per image for Theorem 1, all for Theorem 2); the empirical response distributions were computed using all non-overlapping positions for the remaining images in each validation set. We observe the excellent agreement between theory and practice, particularly for Theorem 2. Recall, however, that the bounds from the two results are not directly comparable because they use different definitions of the false positive rate, depending on whether the detector neuron is considered to be sampled or fixed. + +Internal locking. We added internal locks to the ResNet50 and VGG16 models, with the detector placed in a range of convolutional layers and a single disruptor placed in the logits layer. The results for ResNet50 are shown in Fig. 3. Similar results for VGG-16 (Fig. 15) can be found in the supplementary materials. The optimised trigger patch was taken to be the receptive field of the convolutional detector when in position $(0,0)$ . Although this overlaps with the convolution's padding region, the performance remains excellent. We report the models' performance in 6 settings: the original model, an 'edited' model in which the lock is implanted except for the disruptor itself, and the locked model; in each case, we evaluated using images with and without the trigger patch. It is clear from these results that a lock can be inserted into a pre-trained model using our algorithms with minimal impact on performance. There is a balance to be struck when placing the lock: too early in the model and the dimension of the feature space can be too low to produce + +![](images/22848ff6a67fe1e47146c8a36e69077e1fe56cf8fe2a64e1dc1b9dfa6a20cc50.jpg) +--- original model - original images original model - patched images + +![](images/c57c8b3d7c37348000b55388b150ed2d4a5bf57fe1f21fefcaccf5f76dc6d58d.jpg) +Conv Layer Index --- original model -original images edited model -original images edited model -patched images + +![](images/9648421a70a61287e968781f2657e95415e851a9571f6e36afba5c3a269a5248.jpg) +Conv Layer Index --- original model -original locked model -original locked model -patchee + +![](images/b6fe276fd8ba35879c1300570a6cef586e421bd6e1c4db22865067b3257f73d8.jpg) +--- original model - original images original model - patched images + +![](images/f1008279cad57a989ec007a9b7b38590db73da1df099d51ede54718e9f75ca39.jpg) +Figure 3. Internal lock for ResNet50 +--- original model - original images post-surgery model - original images post-surgery model - patched images + +![](images/331ce57ca52666e3410355f73dad55a010fde18dd195c833e64523235b5a4cae.jpg) +--- original model - original images locked model - original images locked model - patched images + +![](images/8658c7b734a58ca145e171feba07f15c60c6b76a71eb1c4c9089c57dcd85fe34.jpg) +Figure 4. Squeeze-and-excite lock for ResNet50 + +![](images/e6387f714c9b6b513adad30663b710fdf5cd6d2e53f73d652730a974ac8c65e2.jpg) +--- original model - original images -- original model - patched images +Figure 5. Squeeze-and-excite lock for SSDLite + +![](images/749a87d270decea56a5ccb8d382a7d6dadf5a73713e763f775496b0ca52a30ef.jpg) + +![](images/b1e9df8f53e470a6b28befcab758725037a896f17439cb100d041d1ef38ef864.jpg) +--- original model - original images +--- edited model - original images +--- edited model - patched images + +![](images/53a0f4cf49e2f197d0230f12df6bc8632a72aa24ce6f16ed3094d299b4543fc4.jpg) + +![](images/cd3e861a401cc9f38f45bed81a4e8a7c764e8c915edc7f886c800f463535122a.jpg) +--- original model - original images locked model - original images locked model - patched images + +a reliable detector; too late in the model, and the large size of the detector's receptive field (and hence the trigger patch added to images) can damage the model's performance. + +Squeeze-and-excite locking. We demonstrate squeeze-and-excite locking in two forms: using the existing Sq-Ex blocks in the SSDLite model, and by adding new Sq-Ex blocks to each of the other models. Fig. 4 (ResNet50), Fig. 16 (VGG16), Fig. 5 (SSDLite) and Fig. 17 (Faster-RCNN) report the performance of each model in each of the 6 settings described in the previous section. The detector is placed in a range of convolutional layers, and the Sq-Ex block immediately follows it. The results clearly demonstrate both that this is an effective form of lock, and that a new block can be added to a pre-trained model without impacting the model's performance. The performance of the unlocked models is close to that of the original model in all cases. Where performance differences are observed (such as for + +![](images/e1515b6f7359c67e9a8b9e52e6b293c303609886ea52595ad7104e88664cc96b.jpg) +Figure 6. Staining and Locking DC-GAN + +![](images/c14f58cb4ef6ecfe9e55dc863f4d397693ef8088cc20e506fca23bcecb8306f9.jpg) +Figure 7. Staining and Locking ViT-B-16 + +the Faster-RCNN model in Fig. 17), this appears to be due to challenges in disabling the disruptor rather than because of the addition of the Sq-Ex block to a pre-trained model (shown by comparing the performance of the 'edited' model - containing the extra block but not the disruptor - against that of the 'unlocked' model). + +GAN staining and locking. Figure 6 shows experimental results for staining and locking a DC-GAN [34] model for generating human faces. This was performed using the algorithmic extensions presented in Appendix E.1, and further examples of image generations are shown in Figure 18. Full details of the experimental setup are provided in Appendix E.1.1. + +ViT staining and locking. Figure 7 shows experimental results for staining and locking a ViT [12] model for image classification. These were obtained using the algorithms and experimental setup described in Appendix E.2. + +# 7. Discussion + +Staining and locking are computationally cheap. The staining and locking mechanisms we present here are very cheap to implement. The model's weights are directly modi + +fied, so no training or fine-tuning is required. The algorithms also require no training or validation data, meaning that models trained on sensitive data can be easily handled. The inference-time cost of the stain and lock are minimal too, being embedded within the model itself, and even the Sq-Ex block is much cheaper than a single convolutional layer. + +Forging attacks. In a forging attack [5], an attacker plants their own stain into a model and uses it to claim ownership. Since previous staining algorithms have required stains to be added at training time, this has not been considered a significant threat in the past. Our work reveals that functionally identical stains can easily be implanted post-training (Sec. 3.1), and forging attacks are therefore easy to perform. + +Pruning attacks. An attacker may attempt to remove a stain or lock by simply pruning it out [5, 23]. This is not a risk for our additive stains, since the detector is simply added to the original neuron, so the normal behaviour of the original neuron is retained. A non-additive stain may result in a non-responsive neuron which could be pruned, depending on the values selected for $\delta$ and $\Delta$ . For the lock, on the other hand, pruning out the detector neuron would not help an attacker: without the ability to produce the correct unlocking signal, the full model performance could not be recovered. The disruptor is also difficult to remove, since it replaces the original bias vector which cannot be recovered without knowing the value of $\gamma$ , the unlocking signal. The security properties of our stains and locks, including their robustness to standard attacks, are discussed further in Sec. F. + +Obfuscating locks. We believe it is advantageous to encourage a variety of hiding and obfuscation techniques to grow around the fundamental building blocks we have provided here for locking. However, in our experience the task of detecting (and removing) a lock from model weights – which are widely known to be largely uninterpretable to the human eye – is already a significant challenge after only mild attempts at hiding it. This provides an asymmetrically difficult task for a thief with little work from the owner. + +# 8. Conclusion + +We have introduced new staining and locking algorithms for protecting pre-trained neural network models for computer vision applications. Unlike approaches in the literature, our stains and locks can be implemented without re-training or fine-tuning the model, and do not even require access to training or validation data. We have shown how several standard staining schemas can be reimplemented using our techniques. Novel and computable theoretical guarantees on the false positive rate of our stains and locks provide users with new tools for assessing the suitability of their models for staining and locking. Experiments using standard pretrained models for image classification and object detection demonstrate the practical performance of our algorithms. + +# Acknowledgement + +I.T. was supported by a grant, provided by the Ministry of Economic Development of the Russian Federation in accordance with the subsidy agreement (agreement identifier 000000C313925P4G0002) and the agreement with the Ivannikov Institute for System Programming of the Russian Academy of Sciences dated June 20, 2025 No. 139-15-2025-011. + +# References + +[1] Yossi Adi, Carsten Baum, Moustapha Cisse, Benny Pinkas, and Joseph Keshet. Turning your weakness into a strength: watermarking deep neural networks by backdooring. In Proceedings of the 27th USENIX Conference on Security Symposium, page 1615-1631, USA, 2018. USENIX Association. 1, 2, 4 +[2] Manaar Alam, Sayandeep Saha, Debdeep Mukhopadhyay, and Sandip Kundu. NN-Lock: A lightweight authorization to prevent IP threats of deep learning models. J. Emerg. Technol. Comput. Syst., 18(3), 2022. 2 +[3] Jason Ansel, Edward Yang, Horace He, Natalia Gimelshein, Animesh Jain, Michael Voznesensky, Bin Bao, Peter Bell, David Berard, Evgeni Burovski, Geeta Chauhan, Anjali Chourdia, Will Constable, Alban Desmaison, Zachary DeVito, Elias Ellison, Will Feng, Jiong Gong, Michael Gschwind, Brian Hirsh, Sherlock Huang, Kshitteej Kalambarkar, Laurent Kirsch, Michael Lazos, Mario Lezcano, Yanbo Liang, Jason Liang, Yinghai Lu, CK Luk, Bert Maher, Yunjie Pan, Christian Puhrsch, Matthias Reso, Mark Saroufim, Marcos Yukio Siraichi, Helen Suk, Michael Suo, Phil Tillet, Eikan Wang, Xiaodong Wang, William Wen, Shunting Zhang, Xu Zhao, Keren Zhou, Richard Zou, Ajit Mathews, Gregory Chanan, Peng Wu, and Soumith Chintala. PyTorch 2: Faster machine learning through dynamic python bytecode transformation and graph compilation. In 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2 (ASPLOS '24). ACM, 2024. 6 +[4] Keith Ball. An elementary introduction to modern convex geometry. In Flavours of Geometry, 1997. 9 +[5] Franziska Boenisch. A systematic review on model watermarking for neural networks. Frontiers in Big Data, 4, 2021. 1, 2, 8 +[6] Yuling Cai, Fan Xiang, Guozhu Meng, Yinzhi Cao, and Kai Chen. AuthNet: Neural network with integrated authentication logic, 2024. 1, 2 +[7] Huili Chen, Bita Darvish Rouhani, Cheng Fu, Jishen Zhao, and Farinaz Koushanfar. DeepMarks: A secure fingerprinting framework for digital rights management of deep learning models. In Proceedings of the 2019 on International Conference on Multimedia Retrieval, page 105–113, New York, NY, USA, 2019. Association for Computing Machinery. 1, 2, 3 +[8] Huili Chen, Bita Darvish Rouhani, and Farinaz Koushanfar. BlackMarks: Blackbox multibit watermarking for deep neural networks, 2019. 1, 2, 4 +[9] Xiaojie Chu, Liangyu Chen, and Wenqing Yu. Nafssr: Stereo image super-resolution using nafnet. In Proceedings of the + +IEEE/CVF conference on computer vision and pattern recognition, pages 1239-1248, 2022. 5 +[10] Qi Cui, Ruohan Meng, Chaohui Xu, and Chip-Hong Chang. Steganographic passport: An owner and user verifiable credential for deep model ip protection without retraining. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12302-12311, 2024. 2 +[11] Bita Darvish Rouhani, Huili Chen, and Farinaz Koushanfar. DeepSigns: An end-to-end watermarking framework for protecting the ownership of deep neural networks. In Architectural Support for Programming Languages and Operating Systems (ASPLOS). ACM, 2019. 1, 2, 4 +[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.8, 5 +[13] A. Dvoretzky, J. Kiefer, and J. Wolfowitz. Asymptotic minimax character of the sample distribution function and of the classical multinomial estimator. The Annals of Mathematical Statistics, 27(3):642-669, 1956. 3 +[14] Lixin Fan, Kam Woh Ng, Chee Seng Chan, and Qiang Yang. DeepIPR: Deep neural network ownership verification with passports. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6122-6139, 2022. 1, 2 +[15] Guanhao Gan, Yiming Li, Dongxian Wu, and Shu-Tao Xia. Towards robust model watermark via reducing parametric vulnerability. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4728-4738, 2023. 2 +[16] Yifeng Gao, Yuhua Sun, Xingjun Ma, Zuxuan Wu, and YuGang Jiang. Modellock: Locking your model with a spell. In Proceedings of the 32nd ACM International Conference on Multimedia, page 11156-11165, New York, NY, USA, 2024. Association for Computing Machinery. 1, 2 +[17] Jia Guo and Miodrag Potkonjak. Watermarking deep neural networks for embedded systems. In 2018 IEEE/ACM International Conference on Computer-Aided Design (ICCAD), pages 1-8, 2018. 1, 2 +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6 +[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 8 +[20] Andrew Howard, Ruoming Pang, Hartwig Adam, Quoc V. Le, Mark Sandler, Bo Chen, Weijun Wang, Liang-Chieh Chen, Mingxing Tan, Grace Chu, Vijay Vasudevan, and Yukun Zhu. Searching for MobileNetV3. In ICCV, pages 1314-1324. IEEE, 2019. 5, 6 +[21] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7132-7141, 2018. 2, 4, 5 +[22] King's College London. King's Computational Research, Engineering and Technology Environment (CREATE), 2024. 6 + +[23] Yue Li, Hongxia Wang, and Mauro Barni. A survey of deep neural network watermarking techniques. Neurocomputing, 461:171-193, 2021. 1, 2, 8 +[24] Yanghao Li, Saining Xie, Xinlei Chen, Piotr Dollar, Kaiming He, and Ross Girshick. Benchmarking detection transfer learning with vision transformers, 2021. 6 +[25] Yuxuan Li, Sarthak Kumar Maharana, and Yunhui Guo. Not just change the labels, learn the features: Watermarking deep neural networks with multi-view data. In Computer Vision – ECCV 2024, pages 273–289, Cham, 2025. Springer Nature Switzerland. 2 +[26] Zheng Li, Chengyu Hu, Yang Zhang, and Shanqing Guo. How to prove your model belongs to you: a blind-watermark based framework to protect intellectual property of DNN. In Proceedings of the 35th Annual Computer Security Applications Conference, page 126-137, New York, NY, USA, 2019. Association for Computing Machinery. 1, 2 +[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In Computer Vision - ECCV 2014, pages 740-755, Cham, 2014. Springer International Publishing. 6 +[28] Hanwen Liu, Zhenyu Weng, and Yuesheng Zhu. Watermarking deep neural networks with greedy residuals. In Proceedings of the 38th International Conference on Machine Learning, pages 6978-6988. PMLR, 2021. 1, 2, 4 +[29] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C. Berg. SSD: Single shot multibox detector. In Computer Vision – ECCV 2016, pages 21–37, Cham, 2016. Springer International Publishing. 5, 6 +[30] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 8 +[31] P. Massart. The tight constant in the dvoretzky-kiefer-wolfowitz inequality. The Annals of Probability, 18(3):1269-1283, 1990. 3 +[32] E. Le Merrer, P. Pérez, and G. Trédan. Adversarial frontier stitching for remote neural network watermarking. Neural Computing and Applications, 32:9233-9244, 2019. 1, 2, 4 +[33] Xin Mu, Yu Wang, Zhengan Huang, Junzuo Lai, Yehong Zhang, Hui Wang, and Yue Yu. Encryip: A practical encryption-based framework for model intellectual property protection. Proceedings of the AAAI Conference on Artificial Intelligence, 38(19):21438-21445, 2024. 2 +[34] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015.8,4 +[35] Yehonathan Refael, Adam Hakim, Lev Greenberg, Tal Aviv, Satya Lokam, Ben Fishman, and Shachar Seidman. SLIP: Securing LLMs IP using weights decomposition, 2024. 2 +[36] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(6):1137-1149, 2017. 6 + +[37] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet large scale visual recognition challenge. International Journal of Computer Vision (IJCV), 115 (3):211-252, 2015. 6 +[38] M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L. Chen. MobileNetV2: Inverted residuals and linear bottlenecks. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4510-4520, Los Alamitos, CA, USA, 2018. IEEE Computer Society. 5 +[39] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition, 2015. 6 +[40] Z. Sun, R. Sun, C. Liu, A. Chowdhury, L. Lu, and S. Jha. ShadowNet: A secure and efficient on-device model inference system for convolutional neural networks. In 2023 IEEE Symposium on Security and Privacy (SP), pages 1596-1612, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 2 +[41] Oliver Sutton, Qinghua Zhou, Wei Wang, Desmond Higham, Alexander N. Gorban, Alexander Bastounis, and Ivan Y Tyukin. Stealth edits to large language models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2 +[42] Ruixiang Tang, Mengnan Du, and Xia Hu. Deep serial number: Computational watermark for dnn intellectual property protection. In Machine Learning and Knowledge Discovery in Databases: Applied Data Science and Demo Track, pages 157-173, Cham, 2023. Springer Nature Switzerland. 2 +[43] Ivan Y. Tyukin, Desmond J. Higham, and Alexander N. Gorbán. On adversarial examples and stealth attacks in artificial intelligence systems. In 2020 International Joint Conference on Neural Networks (IJCNN), pages 1-6, 2020. 2 +[44] Ivan Y Tyukin, Desmond J Higham, Alexander Bastounis, Eliyas Woldegeorgis, and Alexander N Gorban. The feasibility and inevitability of stealth attacks. IMA Journal of Applied Mathematics, 89(1):44-84, 2023. 2 +[45] Yusuke Uchida, Yuki Nagai, Shigeyuki Sakazawa, and Shin'ichi Satoh. Embedding watermarks into deep neural networks. In Proceedings of the 2017 ACM on International Conference on Multimedia Retrieval, page 269-277, New York, NY, USA, 2017. Association for Computing Machinery. 1, 2, 3 +[46] Jiangfeng Wang, Hanzhou Wu, Xinpeng Zhang, and Yuwei Yao. Watermarking in deep neural networks via error backpropagation. *Electronic Imaging*, 32(4):22-1-22-1, 2020. 1, 2 \ No newline at end of file diff --git a/2025/Staining and Locking Computer Vision Models Without Retraining/images.zip b/2025/Staining and Locking Computer Vision Models Without Retraining/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f17ff334c7b45dedde09098ccb49e0542536cf93 --- /dev/null +++ b/2025/Staining and Locking Computer Vision Models Without Retraining/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:998a41305c80d16a6dd18659727eb2bd57ba22a61daa1e15e8f1d2903c947e8f +size 288340 diff --git a/2025/Staining and Locking Computer Vision Models Without Retraining/layout.json b/2025/Staining and Locking Computer Vision Models Without Retraining/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9256397dbb26ba592eee0279fd02ae4defccb9d7 --- /dev/null +++ b/2025/Staining and Locking Computer Vision Models Without Retraining/layout.json @@ -0,0 +1,11503 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 108, + 103, + 503, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 103, + 503, + 121 + ], + "spans": [ + { + "bbox": [ + 108, + 103, + 503, + 121 + ], + "type": "text", + "content": "Staining and locking computer vision models without retraining" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 144, + 223, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 144, + 223, + 156 + ], + "spans": [ + { + "bbox": [ + 140, + 144, + 223, + 156 + ], + "type": "text", + "content": "Oliver J. Sutton*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 158, + 225, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 158, + 225, + 171 + ], + "spans": [ + { + "bbox": [ + 135, + 158, + 225, + 171 + ], + "type": "text", + "content": "Synoptix Ltd. and" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 123, + 172, + 237, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 172, + 237, + 185 + ], + "spans": [ + { + "bbox": [ + 123, + 172, + 237, + 185 + ], + "type": "text", + "content": "King's College London" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 298, + 144, + 374, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 144, + 374, + 157 + ], + "spans": [ + { + "bbox": [ + 298, + 144, + 374, + 157 + ], + "type": "text", + "content": "Qinghua Zhou*" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 277, + 158, + 391, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 158, + 391, + 171 + ], + "spans": [ + { + "bbox": [ + 277, + 158, + 391, + 171 + ], + "type": "text", + "content": "King's College London" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 274, + 174, + 394, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 174, + 394, + 185 + ], + "spans": [ + { + "bbox": [ + 274, + 174, + 394, + 185 + ], + "type": "text", + "content": "qinghua.zhou@kcl.ac.uk" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 425, + 144, + 493, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 144, + 493, + 157 + ], + "spans": [ + { + "bbox": [ + 425, + 144, + 493, + 157 + ], + "type": "text", + "content": "George Leete" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 425, + 158, + 493, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 158, + 493, + 171 + ], + "spans": [ + { + "bbox": [ + 425, + 158, + 493, + 171 + ], + "type": "text", + "content": "Synoptix Ltd." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 188, + 242, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 188, + 242, + 198 + ], + "spans": [ + { + "bbox": [ + 116, + 188, + 242, + 198 + ], + "type": "text", + "content": "oliver.sutton@kcl.ac.uk" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 71, + 205, + 181, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 205, + 181, + 219 + ], + "spans": [ + { + "bbox": [ + 71, + 205, + 181, + 219 + ], + "type": "text", + "content": "Alexander N. Gorban" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 71, + 220, + 182, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 220, + 182, + 232 + ], + "spans": [ + { + "bbox": [ + 71, + 220, + 182, + 232 + ], + "type": "text", + "content": "Central Univ. Moscow" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 205, + 205, + 538, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 205, + 538, + 247 + ], + "spans": [ + { + "bbox": [ + 205, + 205, + 538, + 247 + ], + "type": "text", + "content": "Ivan Y. Tyukin \nKing's College London and AI Center, Bol'shoi Bul'var 30, Moscow \nand ISP RAS Research Center for Trusted Artificial Intelligence" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 152, + 274, + 200, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 274, + 200, + 288 + ], + "spans": [ + { + "bbox": [ + 152, + 274, + 200, + 288 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 300, + 297, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 297, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 297, + 504 + ], + "type": "text", + "content": "We introduce new methods of staining and locking computer vision models, to protect their owners' intellectual property. Staining, also known as watermarking, embeds secret behaviour into a model which can later be used to identify it, while locking aims to make a model unusable unless a secret trigger is inserted into input images. Unlike existing methods, our algorithms can be used to stain and lock pretrained models without requiring fine-tuning or retraining, and come with provable, computable guarantees bounding their worst-case false positive rates. The stain and lock are implemented by directly modifying a small number of the model's weights and have minimal impact on the (unlocked) model's performance. Locked models are unlocked by inserting a small 'trigger patch' into the corner of the input image. We present experimental results showing the efficacy of our methods and demonstrating their practical performance on a variety of computer vision models." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 528, + 135, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 135, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 135, + 540 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 548, + 296, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 296, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 296, + 632 + ], + "type": "text", + "content": "Designing, training and validating a computer vision model can be an expensive undertaking. The effect on an organisation if such a valuable investment is leaked or stolen, can be deleterious, particularly since model weights can be difficult for a human judge to distinguish confidently. Developing methods for protecting such intellectual property, which may be deployed on a vulnerable edge device, is therefore crucial." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 633, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 296, + 693 + ], + "type": "text", + "content": "Two natural approaches that a model's owner may wish to take are either to imprint an identifiable " + }, + { + "bbox": [ + 55, + 633, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\text{stain}" + }, + { + "bbox": [ + 55, + 633, + 296, + 693 + ], + "type": "text", + "content": " within its weights, or to add a lock to make the model unusable without adding a separate secret unlocking trigger to the model's input. Staining, also known as watermarking in the liter-" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 275, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 555, + 361 + ], + "type": "text", + "content": "ature [5, 23], can be viewed as a form of 'copyright trap'. By embedding specific (unnatural) behavioural quirks into the model, an identifying fingerprint is produced which can later be used to determine the provenance of a stolen or copied model. Locking, on the other hand, has a more direct aim [6, 14, 16]: unless the thief knows the secret trigger, they are unable to use the model to its full capability." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 362, + 556, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 362, + 556, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 556, + 542 + ], + "type": "text", + "content": "Here, we present new methods for both staining and locking neural network models in computer vision. Unlike existing approaches in the literature (reviewed in Sec. 2), our algorithms do not require model training or fine-tuning. They directly modify the model's weights, and can therefore be used to stain or lock pre-trained models with minimal performance impact. Since re-training is not required, a single model can be supported and updated across multiple customers, while providing each with a separate stain or lock. A further key advantage of our staining and locking techniques is that implementing them does not require access to training or validation data. We also provide two distinct novel theoretical guarantees on the worst-case false positive rate of our staining and locking algorithms, analogues of which do not appear to be present in the literature." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 545, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 714 + ], + "type": "text", + "content": "Staining methods are broadly composed of two parts: a schema describing the information encoded in the stain and how it may later be read, and a mechanism through which this information is implanted into the model. Numerous schemas have been proposed in the literature, e.g. hiding a message in the model's weights, its activations or its class predictions for specific trigger inputs. Until now, however, only two families of mechanisms have been available, both requiring model retraining: through special loss functions [7, 8, 11, 28, 45, 46], or by manipulating the training data [1, 17, 26, 32]. Our novel training-free staining mechanism implants highly selective detector neurons into the model, which produce a strong output only when a special trigger is present in the model's input. Sec. 3.1 shows how several standard staining" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 274, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 274, + 714 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 274, + 714 + ], + "type": "text", + "content": "*These authors contributed equally and are listed alphabetically." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2346" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 269, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 269, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 269, + 83 + ], + "type": "text", + "content": "schemas may be implemented using this mechanism." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 86, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 86, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 86, + 294, + 228 + ], + "type": "text", + "content": "Our lock may be seen as an 'active' variant of the stain mechanism. Additional disruptors are also implanted to pollute the model's activations and prevent it from performing well. A specially constructed 'trigger patch' is added to one corner of the input image during inference is detected by the detector neuron. The detector neuron then produces a signal which deactivates the disruptors, recovering the full model performance. Examples of such trigger patches are shown in Appendix B. Our locking methods are presented with a focus on their fundamental building blocks, which can then be used to design bespoke algorithms to lock specific models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 232, + 296, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 296, + 531 + ], + "type": "text", + "content": "Our goal is to present these new techniques as clearly as possible in a simple setting. For this, we focus primarily on convolutional models, although extensions of the core algorithms to generative adversarial networks (GANs) and vision transformers (ViTs) are presented in Appendix E. Our staining methods are generic and apply directly to many different model architectures. Our basic lock is more architecture-dependent since it interacts with the fundamental action of the model. To provide architectural-independence, we also show how the lock may be embedded within a Squeeze-and-Excite block [21], a common component of many models. We demonstrate, in Sec. 6, that such a block can be added to many pre-trained models with minimal impact on performance. For simplicity and brevity, we do not focus too strongly on techniques for hiding and obfuscating the stains and locks; we provide ideas for achieving this throughout, but do not pursue it in detail to present the core concepts as clearly as possible in the space available. For practical success, we also feel it is preferable for a diversity of obfuscation techniques and algorithmic variations to proliferate. Instead, we focus on the simplest case where a single detector and disruptor are added to the model; in practice, multiple detectors should be combined to deactivate multiple disruptors. This straightforward extension of the methods presented here makes reverse-engineering the lock a significant challenge." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 533, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 700 + ], + "type": "text", + "content": "Our approaches to staining and locking are inspired by stealth attacks [41, 43, 44], which have recently been shown to enable an attacker to specify the output of models from many modern families for a given piece of input data by directly modifying just a few weights. By harnessing this core sensitivity of modern models, and exploiting the fundamental concentration properties of their feature spaces, the highly selective detector neuron of our stain and lock can be implanted without retraining the model. This detector forms the heart of our methods, around which completely new mechanisms are built to facilitate staining and locking. This link means that the novel theoretical guarantees we present in Sec. 5 can be adapted back to study stealth attacks, and complement those available in the literature [41, 43, 44]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "The paper is organised as follows. Sec. 2 places our" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "content": "approach in the context of related work. Our staining and locking algorithms are presented in Sec. 3 and Sec. 4. Sec. 5 presents novel theoretical guarantees on the detector's false positive rate. Experimental results demonstrate the practical performance of our algorithms in Sec. 6. A discussion of our findings is in Sec. 7, and some final conclusions are given in Sec. 8. Our mathematical notation is summarised in Appendix A (in the supplementary material)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 184, + 399, + 196 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 184, + 399, + 196 + ], + "spans": [ + { + "bbox": [ + 313, + 184, + 399, + 196 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 205, + 555, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 555, + 397 + ], + "type": "text", + "content": "Staining. Several staining methods have been proposed in the literature and mainly fall within two thematic groups; see [5, 23] for a thorough review. Loss-based methods encode a secret message into the model's weights during training using a special loss function. This message may then be revealed using a special trigger input to the network. Common approaches encode the message as a binary string in the model's weights [7, 45], activations [28] or outputs [8, 11], either directly or to be decoded by a special model [10, 46]. Backdoor methods, on the other hand, manipulate training data to produce a unique behavioural signature. One method is to force the model to learn specific non-natural labels for certain distinctive inputs [1, 15, 25, 32]. Alternatively, other work has considered the task of training the model to learn to associate a specific class with certain deliberate changes to natural images [17], such as the presence of a logo [26]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 399, + 556, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 556, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 556, + 532 + ], + "type": "text", + "content": "Locking. Two key families of approaches have been proposed in the literature. Intrinsic methods embed the lock into the model itself during training. This may be by training the model's output to depend on an additional (secret) input [14, 42], a cryptographic scheme [33], a secret global mask applied to inputs [6], or by converting training and inference images using a 'style-transfer' model (e.g. to oil paintings) [16]. Extrinsic methods offer additional orthogonal security to intrinsic methods via techniques from outside the neural network domain, such as encryption [2] or secure infrastructure [35, 40]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "content": "Novelty. In contrast to those in the literature, our staining and locking algorithms directly modify a model's weights, so require no model training, fine-tuning or external infrastructure. This is a key advantage because a single base model can be supplied to many clients with different locks/stains, without re-training each time. Moreover, re-training a model carries the inherent risk of changing its behaviour in unidentifiable ways (especially if the model is tuned to give incorrect responses to natural images as in backdoor methods), which is avoided through our direct approach. Being embedded in the model itself, a lock also remains active even if an adversary gains access to the hardware on which it runs. Our provable performance guarantees further differentiate our methods; comparable results are not available for any of the methods in the literature." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2347" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 85 + ], + "type": "text", + "content": "3. Staining" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 92, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 92, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 92, + 296, + 163 + ], + "type": "text", + "content": "This section presents the building blocks of our novel training-free staining mechanism. The stain implants a highly selective detector neuron into the model, which is activated by a specially optimised trigger input to the model. We show how this mechanism can be used to implement standard staining schemas in Sec. 3.1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 167, + 296, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 167, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 296, + 357 + ], + "type": "text", + "content": "Multilayer perceptrons. Algorithm 1 outlines our mechanism for staining a generic multilayer perceptron neural network. A detector neuron with a weight vector sampled from " + }, + { + "bbox": [ + 55, + 167, + 296, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(\\mathbb{S}^{m - 1})" + }, + { + "bbox": [ + 55, + 167, + 296, + 357 + ], + "type": "text", + "content": " (i.e. uniformly from the surface of a sphere) is inserted into the model. Concentration properties in the model's feature space ensure that this randomly sampled neuron is very unlikely to respond strongly to natural inputs (see Sec. 5). It is possible, however, to optimise a trigger input (e.g. using gradient descent) to strongly activate the detector. The strong response to the trigger input, therefore, becomes an identifying characteristic of the model and can be used as the basis for implementing standard staining schemas. To ensure that the detector neuron has no impact on the model's performance, the parameters " + }, + { + "bbox": [ + 55, + 167, + 296, + 357 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 167, + 296, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 167, + 296, + 357 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 55, + 167, + 296, + 357 + ], + "type": "text", + "content": " can be used in conjunction with the activation function to ensure that the detector's response is zero for natural inputs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "text", + "content": "The optimisation of the trigger input " + }, + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "inline_equation", + "content": "x^{*}" + }, + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "text", + "content": " is important here: a strong detector response to " + }, + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "inline_equation", + "content": "x^{*}" + }, + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "text", + "content": " is very distinct from its response to natural inputs, and the guarantees on the false positive rate are improved (see Theorem 1). It also means the response " + }, + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "text", + "content": " can be achieved by scaling down the weight vector, which can be beneficial for evading detection." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 434, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 296, + 529 + ], + "type": "text", + "content": "Additive staining. Alternatively, an additive stain adds the detector weight to an existing neuron's weight. Once again, the detector component of the weight likely only responds weakly to natural inputs, meaning that it has little impact on the model's operation. The strong response to the trigger input, however, is still preserved. Since this method does not produce a 'silent' detector neuron, it is much less likely that the detector will be identified and removed by an adversary." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 532, + 296, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 296, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 296, + 687 + ], + "type": "text", + "content": "Convolutional networks. Algorithm 2 demonstrates how this framework applies to a convolutional neural network. For simplicity, this algorithm focuses on the non-additive stain, but also applies to models incorporating batch normalisation. A single convolutional kernel is used as the detector, and a reduction operator " + }, + { + "bbox": [ + 55, + 532, + 296, + 687 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 532, + 296, + 687 + ], + "type": "text", + "content": " is used for the optimisation to convert the kernel's activation map into a single scalar response value. This could, for example, be the response of the kernel at a particular position in the image, or its average response. Larger kernels are helpful in reducing the false positive rate, as made explicit by Theorem 1: because a " + }, + { + "bbox": [ + 55, + 532, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\kappa \\times \\kappa" + }, + { + "bbox": [ + 55, + 532, + 296, + 687 + ], + "type": "text", + "content": " convolutional kernel with " + }, + { + "bbox": [ + 55, + 532, + 296, + 687 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 532, + 296, + 687 + ], + "type": "text", + "content": " input channels has " + }, + { + "bbox": [ + 55, + 532, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\kappa^2 c" + }, + { + "bbox": [ + 55, + 532, + 296, + 687 + ], + "type": "text", + "content": " parameters, large kernels operate in a higher dimensional setting." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Extensions. Extensions to apply these staining algorithms to GANs and ViTs are presented in Section E." + } + ] + } + ], + "index": 6 + }, + { + "type": "code", + "bbox": [ + 316, + 88, + 553, + 274 + ], + "blocks": [ + { + "bbox": [ + 320, + 75, + 503, + 87 + ], + "lines": [ + { + "bbox": [ + 320, + 75, + 503, + 87 + ], + "spans": [ + { + "bbox": [ + 320, + 75, + 503, + 87 + ], + "type": "text", + "content": "Algorithm 1: Multilayer perceptron staining" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "lines": [ + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "text", + "content": "Input:Trained multilayer perceptron network " + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "text", + "content": " Index " + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "text", + "content": " of neuron to stain in layer " + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "text", + "content": " Detector response to trigger " + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "inline_equation", + "content": "\\Delta \\in \\mathbb{R}" + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "text", + "content": " Desired non-trigger detector response " + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "inline_equation", + "content": "\\delta \\in \\mathbb{R}" + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "text", + "content": " Whether the stain should be additive 1 Define the map " + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "text", + "content": " from network inputs to inputs to layer " + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "text", + "content": " 2 Let " + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "inline_equation", + "content": "W_{j}\\in \\mathbb{R}^{n\\times m}" + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "inline_equation", + "content": "b_{j}\\in \\mathbb{R}^{n}" + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "text", + "content": " be the layer " + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "text", + "content": " weight and bias. 3 Sample a detector weight vector " + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "inline_equation", + "content": "v\\sim \\mathcal{U}(\\mathbb{S}^{m - 1})" + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "text", + "content": " 4 Optimise the trigger input " + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "inline_equation", + "content": "x^{*}\\in \\arg \\max_{z\\in S}v\\cdot \\phi (z)" + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "text", + "content": " 5 if stain is additive then 6 Let " + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "text", + "content": " be entry " + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "inline_equation", + "content": "b_{j}" + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "text", + "content": " be row " + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "inline_equation", + "content": "W_{j}" + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "text", + "content": " 7 Define " + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "inline_equation", + "content": "u = w + \\frac{\\Delta - \\beta - w\\cdot\\phi(x^{*})}{v\\cdot\\phi(x^{*})} v." + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "text", + "content": " 8 else 9 Define " + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "inline_equation", + "content": "\\beta = \\delta" + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "inline_equation", + "content": "u = \\frac{\\Delta - \\beta}{v\\cdot\\phi(x^*)} v." + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "text", + "content": " 10 Replace row " + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "inline_equation", + "content": "W_{j}" + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "text", + "content": " and entry " + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "inline_equation", + "content": "b_{j}" + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "text", + "content": " Output: Stained model " + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "text", + "content": " and trigger input " + }, + { + "bbox": [ + 316, + 88, + 553, + 274 + ], + "type": "inline_equation", + "content": "x^{*}" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "code_body" + } + ], + "index": 8, + "sub_type": "algorithm" + }, + { + "type": "code", + "bbox": [ + 316, + 306, + 553, + 539 + ], + "blocks": [ + { + "bbox": [ + 320, + 293, + 506, + 305 + ], + "lines": [ + { + "bbox": [ + 320, + 293, + 506, + 305 + ], + "spans": [ + { + "bbox": [ + 320, + 293, + 506, + 305 + ], + "type": "text", + "content": "Algorithm 2: Convolutional network staining" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "lines": [ + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "spans": [ + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "text", + "content": "Input:Trained convolutional network " + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "text", + "content": " Parameters " + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "inline_equation", + "content": "(k,j,\\Delta ,\\delta)" + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "text", + "content": " as in Alg.1 Matrix to scalar reduction operation " + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "text", + "content": " Whether layer " + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "text", + "content": " performs batch normalisation \n1 Define the map " + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "inline_equation", + "content": "\\phi :S = \\mathbb{R}^{3\\times m_0\\times n_0}\\to \\mathbb{R}^{c_j\\times m_j\\times n_j}" + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "text", + "content": " from network inputs to inputs to layer " + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "text", + "content": " \n2 Let " + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "inline_equation", + "content": "W_{j}\\in \\mathbb{R}^{c_{j} + 1\\times c_{j}\\times \\kappa_{j}\\times \\kappa_{j}}" + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "text", + "content": " be the layer " + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "text", + "content": " weight, with kernel shape " + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "inline_equation", + "content": "\\kappa_{j}\\times \\kappa_{j}" + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "text", + "content": " \n3 Sample a kernel " + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "inline_equation", + "content": "v\\in \\mathbb{R}^{c_j\\times \\kappa_j\\times \\kappa_j}" + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(\\mathbb{S}^{c_j\\kappa_j^2 -1})" + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "text", + "content": " viewed as a tensor with shape " + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "inline_equation", + "content": "(c_{j},\\kappa_{j},\\kappa_{j})" + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "text", + "content": " \n4 Optimise the trigger image " + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "inline_equation", + "content": "x^{*}\\in \\arg \\max_{z\\in S}r(v*\\phi (z))" + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "text", + "content": " \n5 if layer " + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "text", + "content": " performs batch normalisation then \n6 Let " + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "inline_equation", + "content": "\\mu_j,\\sigma_j^2,w_j,b_j\\in \\mathbb{R}^{c_j + 1}" + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "text", + "content": " be the mean, variance, weight and bias of the layer " + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "text", + "content": " batch normalisation. \n7 Define " + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "inline_equation", + "content": "\\alpha = \\frac{(\\Delta - \\delta)(\\sigma_j)_k}{(w_j)_kr(v*\\phi(x^*))}" + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "inline_equation", + "content": "\\beta = \\delta +\\frac{(w_j)_k(\\mu_j)_k}{(\\sigma_j)_k}" + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "text", + "content": " \n8 else \n9 Let " + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "inline_equation", + "content": "b_{j}\\in \\mathbb{R}^{c_{j} + 1}" + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "text", + "content": " be the bias of layer " + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "text", + "content": " \n10 Define " + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "inline_equation", + "content": "\\alpha = \\frac{(\\Delta - \\delta)}{r(v*\\phi(x^*))}" + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "inline_equation", + "content": "\\beta = \\delta" + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "text", + "content": " \n11 Replace " + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "inline_equation", + "content": "(W_{j})_{k,\\dots ,}" + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "inline_equation", + "content": "\\alpha v" + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "text", + "content": " ,and entry " + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "inline_equation", + "content": "b_{j}" + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "text", + "content": " Output: Stained model " + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "text", + "content": " and trigger input " + }, + { + "bbox": [ + 316, + 306, + 553, + 539 + ], + "type": "inline_equation", + "content": "x^{*}" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "code_body" + } + ], + "index": 10, + "sub_type": "algorithm" + }, + { + "bbox": [ + 313, + 563, + 529, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 529, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 529, + 575 + ], + "type": "text", + "content": "3.1. Implementing standard staining schemas" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 580, + 553, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 553, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 553, + 616 + ], + "type": "text", + "content": "The algorithms above introduce a novel training-free staining mechanism, and here we briefly discuss how it can be used to implement standard schemas from the literature." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "Weight staining. Originally proposed by [7, 45], this schema encodes a secret message as a binary string, which is then implanted into the model as the signs of the weights of a target neuron (e.g. " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "-" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": " encodes 0 and " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "+" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": " encodes 1). The stain can be checked by decoding the message from the signs of the neuron's weight vector's components. This can be embedded into a non-additive stain by sampling the detector vector " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": " from the orthant of the sphere with the specified" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2348" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "component signs, and the rest of the staining algorithm can then proceed as usual. This restriction on the sampling may increase the false positive rate of the detector (by decreasing the entropy of the distribution), but this should not present a problem if the dimension is sufficiently high." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 134, + 295, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 134, + 295, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 295, + 253 + ], + "type": "text", + "content": "Activation staining. This family of schemas, introduced in [28], instead embeds the secret message into the signs of the model's activations for a target input, rather than the weights. Our additive stain can be used to directly implement this. Suppose a detector vector is sampled for a given layer of the model, and a trigger input is optimised to produce a strong response to this detector vector. Then, this detector vector is weighted and added to each neuron in the layer, to ensure that the response to the trigger image from that neuron has the specified sign." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 256, + 298, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 298, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 298, + 472 + ], + "type": "text", + "content": "Output staining. Another standard schema family specifies (non-natural) class predictions for certain trigger input images [1, 8, 11, 32]. The stain can, therefore, be verified without access to model weights or activations. This can be implemented post-training by starting with a standard non-additive stain from Algorithm 1. If the stain detector neuron is implanted into row " + }, + { + "bbox": [ + 55, + 256, + 298, + 472 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 256, + 298, + 472 + ], + "type": "text", + "content": " of the weight matrix in layer " + }, + { + "bbox": [ + 55, + 256, + 298, + 472 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 256, + 298, + 472 + ], + "type": "text", + "content": ", then the detector's response (which will be zero for non-trigger inputs after a ReLU activation if " + }, + { + "bbox": [ + 55, + 256, + 298, + 472 + ], + "type": "inline_equation", + "content": "\\delta \\ll 0" + }, + { + "bbox": [ + 55, + 256, + 298, + 472 + ], + "type": "text", + "content": ") will multiply the entries of column " + }, + { + "bbox": [ + 55, + 256, + 298, + 472 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 256, + 298, + 472 + ], + "type": "text", + "content": " in layer " + }, + { + "bbox": [ + 55, + 256, + 298, + 472 + ], + "type": "inline_equation", + "content": "j + 1" + }, + { + "bbox": [ + 55, + 256, + 298, + 472 + ], + "type": "text", + "content": ". For any chosen output class " + }, + { + "bbox": [ + 55, + 256, + 298, + 472 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 256, + 298, + 472 + ], + "type": "text", + "content": ", it is typically possible to find a vector (e.g. using gradient descent) which, if it were the output of layer " + }, + { + "bbox": [ + 55, + 256, + 298, + 472 + ], + "type": "inline_equation", + "content": "j + 1" + }, + { + "bbox": [ + 55, + 256, + 298, + 472 + ], + "type": "text", + "content": ", would cause the model to confidently predict class " + }, + { + "bbox": [ + 55, + 256, + 298, + 472 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 256, + 298, + 472 + ], + "type": "text", + "content": ". Replacing column " + }, + { + "bbox": [ + 55, + 256, + 298, + 472 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 256, + 298, + 472 + ], + "type": "text", + "content": " of layer " + }, + { + "bbox": [ + 55, + 256, + 298, + 472 + ], + "type": "inline_equation", + "content": "j + 1" + }, + { + "bbox": [ + 55, + 256, + 298, + 472 + ], + "type": "text", + "content": " with this vector, therefore, causes the model to predict the chosen class for the trigger input; the selectivity of the detector ensures that the new column will not otherwise affect the model's performance. This technique extends analogously to convolutional stains." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 487, + 115, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 115, + 500 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 115, + 500 + ], + "type": "text", + "content": "4. Locking" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 508, + 296, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 296, + 698 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 296, + 698 + ], + "type": "text", + "content": "Our locking methods build on the staining mechanism by inserting additional disruptors into the model to pollute its latent activations. The signal from the stain's detector neuron is used to disable the disruptors when the trigger is present in the input. Fig. 8 and Fig. 9 (in the supplementary material) show examples of images containing these patches. We show how the underlying structure of a convolutional detector can be exploited to produce a transferable trigger 'patch' which can be inserted into any image to unlock the full model performance. We present two variants of the locking mechanism. Internal locking embeds the lock into the weights of convolutional networks with a particular structure, while squeeze-and-excite locking uses a standard Squeeze-and-Excite block [21], common in computer vision models, as the foundation of a more generic procedure which can be added to any convolutional model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "type": "text", + "content": "Internal locking. Algorithm 3 adds an internal lock to a" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 556, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 217 + ], + "type": "text", + "content": "convolutional neural network. To present the idea in the simplest possible setting, we suppose that the model consists of a sequence of convolutional layers with biases (batch normalisation can be handled as in Algorithm 2), ReLU activations, followed by a single dense logits layer. For brevity, we do not explore here how the many additional model components present in practice may alter this algorithm: while they complicate the implementation, they do not change the fundamental concept. The algorithm consists of three main steps: implanting the detector and building the trigger (lines Line 1-4), propagating the detector signal through the model (lines 5-10), and building the disruptor (lines 11-14)." + } + ] + } + ], + "index": 6 + }, + { + "type": "code", + "bbox": [ + 316, + 243, + 553, + 469 + ], + "blocks": [ + { + "bbox": [ + 321, + 231, + 525, + 243 + ], + "lines": [ + { + "bbox": [ + 321, + 231, + 525, + 243 + ], + "spans": [ + { + "bbox": [ + 321, + 231, + 525, + 243 + ], + "type": "text", + "content": "Algorithm 3: Convolutional network internal lock" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "text", + "content": "Input : Parameters " + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "inline_equation", + "content": "(\\mathcal{N},k,j,\\Delta >0,\\delta < 0)" + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "text", + "content": " as in Alg.2 Image coordinates " + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "inline_equation", + "content": "(a,b)" + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "text", + "content": " for the trigger patch Disruption scale " + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "inline_equation", + "content": "s > 0" + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "text", + "content": " and offset " + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "inline_equation", + "content": "t\\in \\mathbb{R}^c" + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "text", + "content": " \n1 Define the map " + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "text", + "content": " from network inputs to inputs to layer " + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "text", + "content": " \n2 Define the matrix to scalar map " + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "inline_equation", + "content": "r_{(a,b)}(M)\\coloneqq (M)_{a,b}" + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "text", + "content": " \n3 Stain " + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "text", + "content": " using Alg. 2, with the reduction map " + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "inline_equation", + "content": "r_{(a,b)}" + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "text", + "content": " \n4 Define the trigger patch " + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "text", + "content": " as the part of the trigger image " + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "inline_equation", + "content": "x^{*}" + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "text", + "content": " which is in the receptive field of " + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "inline_equation", + "content": "r_{(a,b)}(v*\\phi (x^{*}))" + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "text", + "content": " \n5 To simplify presentation, swap kernels 1 and " + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "text", + "content": " in layer " + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "text", + "content": " \n6 foreach convolution layer " + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "text", + "content": " after layer " + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "text", + "content": " do \n7 Let " + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "inline_equation", + "content": "W_{\\ell}\\in \\mathbb{R}^{c_{\\ell +1}\\times c_{\\ell}\\times \\kappa_{\\ell}\\times \\kappa_{\\ell}}" + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "inline_equation", + "content": "b_{\\ell}\\in \\mathbb{R}^{c_{\\ell +1}}" + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "text", + "content": " be the layer " + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "text", + "content": " convolutional weight and bias. \n8 Set " + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "inline_equation", + "content": "(W_{\\ell})_{1,\\dots ,\\cdot} = 0 = (W_{\\ell})_{\\cdot ,1,\\dots ,\\cdot}" + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "inline_equation", + "content": "(b_{\\ell})_{1} = 0" + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "text", + "content": " \n9 Set " + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "inline_equation", + "content": "(W_{\\ell})_{1,1,\\dots} = 1" + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "text", + "content": " \n10 end \n11 Let " + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "inline_equation", + "content": "W_{L}\\in \\mathbb{R}^{c\\times c_{L}}" + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "inline_equation", + "content": "b_{L}\\in \\mathbb{R}^{c}" + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "text", + "content": " be the logits weight and bias. \n12 Sample a disruptor vector u from " + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(\\mathbb{S}^{c - 1})" + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "text", + "content": " \n13 Let " + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "inline_equation", + "content": "\\gamma \\in \\mathbb{R}^{c_L}" + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "text", + "content": " be the logits layer input for model input " + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "inline_equation", + "content": "x^{*}" + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "text", + "content": " \n14 Set column 1 of " + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "inline_equation", + "content": "W_{L}" + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "inline_equation", + "content": "\\frac{b_L - su + t}{(\\gamma)_1}" + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "text", + "content": " ; replace " + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "inline_equation", + "content": "b_{L}" + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "inline_equation", + "content": "su + t" + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "text", + "content": " Output: Locked model " + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "text", + "content": " and trigger patch " + }, + { + "bbox": [ + 316, + 243, + 553, + 469 + ], + "type": "inline_equation", + "content": "\\pi" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "code_body" + } + ], + "index": 8, + "sub_type": "algorithm" + }, + { + "bbox": [ + 313, + 486, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 581 + ], + "type": "text", + "content": "The detector is built into a convolutional kernel as for the stain, but the trigger image is only optimised to produce a strong activation when the detector kernel is in a specific position (via the reduction map " + }, + { + "bbox": [ + 313, + 486, + 555, + 581 + ], + "type": "inline_equation", + "content": "r_{(a,b)}" + }, + { + "bbox": [ + 313, + 486, + 555, + 581 + ], + "type": "text", + "content": "). This has the effect that only a single small 'patch' of the input image (the receptive field of a single kernel in convolutional layer " + }, + { + "bbox": [ + 313, + 486, + 555, + 581 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 486, + 555, + 581 + ], + "type": "text", + "content": ") is optimised. Inserting this patch into (e.g. a corner of) an image during inference will therefore activate the detector." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": "Keeping this patch relatively small (to minimise its impact when inserted into an image) can be achieved by implanting the detector into a relatively early convolutional layer. Since the disruptor is disabled by the signal from the detector, the disruptor must be placed such that its receptive field always contains the trigger patch in the input image. For this reason, we place the disruptor into a late layer (the algorithm uses the model's logits layer, although many alternatives are possible), which depends on the entire input image, and build a conduit to propagate the detector signal to the disruptor (a more general technique to overcome this requirement is pre" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2349" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 295, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 295, + 156 + ], + "type": "text", + "content": "sented below). In Algorithm 3, this conduit is a sequence of 'identity' kernels, which simply propagate the signal in the detector channel (taken as channel 1 in the algorithm for simplicity) forwards through the model. Many techniques can be used to obfuscate the presence of this channel (and therefore the nature of the lock), although perhaps the simplest is just to add small random values to the modified weights." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 157, + 296, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 157, + 296, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 296, + 385 + ], + "type": "text", + "content": "The disruptor replaces the bias of the logits layer with a randomly sampled vector. This is disabled when the correct signal " + }, + { + "bbox": [ + 55, + 157, + 296, + 385 + ], + "type": "inline_equation", + "content": "(\\gamma)_{1}" + }, + { + "bbox": [ + 55, + 157, + 296, + 385 + ], + "type": "text", + "content": " is received from the detector through the conduit by restoring the original bias using the relevant column of the weight matrix. In practice, multiple interacting detectors and disruptors could be added to different places in the model, making it much more difficult to reverse-engineer the lock. Squeeze-and-excite locking. Computer vision models for certain applications may not be suitable for internal locking if they lack good locations for a disruptor which can reliably respond to the detector signal. This is often the case in object detection, for example, where small objects may be detected using just a small portion of the image. The receptive field for these outputs may, therefore, simply not contain the trigger patch placed in the corner of the image, so its presence or absence would not influence the model's output. To overcome this challenge, we propose an alternative mechanism exploiting architectural blocks which normalise a convolution's response across the entire input image." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 385, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 296, + 445 + ], + "type": "text", + "content": "Squeeze-and-Excite (Sq-Ex) blocks [21] are a prototypical example of such an architecture, which is widely used in applications. The block is a map " + }, + { + "bbox": [ + 55, + 385, + 296, + 445 + ], + "type": "inline_equation", + "content": "s: \\mathbb{R}^{c \\times n \\times m} \\to \\mathbb{R}^{c \\times n \\times m}" + }, + { + "bbox": [ + 55, + 385, + 296, + 445 + ], + "type": "text", + "content": " which aggregates global spatial information about its input and uses this to rescale each channel. Specifically," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 455, + 295, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 455, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 91, + 455, + 295, + 468 + ], + "type": "interline_equation", + "content": "s (x) = x \\odot q (x) \\text {w i t h} q: \\mathbb {R} ^ {c \\times n \\times m} \\rightarrow \\mathbb {R} ^ {c}, \\tag {1}", + "image_path": "eade99aadb895e3f7f523077c4139783c8eb189016283f33ee7da5f999460a37.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 91, + 470, + 263, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 470, + 263, + 483 + ], + "spans": [ + { + "bbox": [ + 91, + 470, + 263, + 483 + ], + "type": "interline_equation", + "content": "\\text {w h e r e} q (x) = \\sigma \\left(S _ {2} \\rho \\left(S _ {1} \\mu (x) + \\tau_ {1}\\right) + \\tau_ {2}\\right).", + "image_path": "372521b6d35b04e65f98e8c439ed933aabc61f23bf10c5a2ee76e2d8435ab83a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 493, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 296, + 552 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 55, + 493, + 296, + 552 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 493, + 296, + 552 + ], + "type": "text", + "content": " represents channel-wise multiplication, " + }, + { + "bbox": [ + 55, + 493, + 296, + 552 + ], + "type": "inline_equation", + "content": "\\mu : \\mathbb{R}^{c\\times n\\times m}\\to \\mathbb{R}^c" + }, + { + "bbox": [ + 55, + 493, + 296, + 552 + ], + "type": "text", + "content": " computes the channel-wise mean, " + }, + { + "bbox": [ + 55, + 493, + 296, + 552 + ], + "type": "inline_equation", + "content": "S_{1}\\in \\mathbb{R}^{d\\times c}" + }, + { + "bbox": [ + 55, + 493, + 296, + 552 + ], + "type": "text", + "content": " (typically with " + }, + { + "bbox": [ + 55, + 493, + 296, + 552 + ], + "type": "inline_equation", + "content": "d < c" + }, + { + "bbox": [ + 55, + 493, + 296, + 552 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 55, + 493, + 296, + 552 + ], + "type": "inline_equation", + "content": "S_{2}\\in \\mathbb{R}^{c\\times d}" + }, + { + "bbox": [ + 55, + 493, + 296, + 552 + ], + "type": "text", + "content": " are learned weight matrices, " + }, + { + "bbox": [ + 55, + 493, + 296, + 552 + ], + "type": "inline_equation", + "content": "\\tau_{1}\\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 55, + 493, + 296, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 493, + 296, + 552 + ], + "type": "inline_equation", + "content": "\\tau_{2}\\in \\mathbb{R}^{c}" + }, + { + "bbox": [ + 55, + 493, + 296, + 552 + ], + "type": "text", + "content": " are learned biases, and " + }, + { + "bbox": [ + 55, + 493, + 296, + 552 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 55, + 493, + 296, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 493, + 296, + 552 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 493, + 296, + 552 + ], + "type": "text", + "content": " are the ReLU and sigmoid activation functions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 554, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 296, + 685 + ], + "type": "text", + "content": "Such blocks provide the key ability to propagate the detector signal laterally through the model, thereby enabling much more flexibility in the placement of disruptors. For example, if the detector produces value of 0 when the trigger patch is not present, and " + }, + { + "bbox": [ + 55, + 554, + 296, + 685 + ], + "type": "inline_equation", + "content": "\\Delta > 0" + }, + { + "bbox": [ + 55, + 554, + 296, + 685 + ], + "type": "text", + "content": " when it is, the averaged response will similarly be non-zero only when the trigger patch is present. A disruptor can, therefore, be embedded within the Sq-Ex block itself (as shown in Algorithm 4), or the block can be used to propagate the unlocking signal to disruptors later in the model. The algorithm only assumes that the convolutional layer " + }, + { + "bbox": [ + 55, + 554, + 296, + 685 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 554, + 296, + 685 + ], + "type": "text", + "content": " has a bias term, uses ReLU" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": "activation, and is followed by an Sq-Ex block. This lock is independent of the architecture of the rest of the model. To present the idea in a simple setting, the algorithm embeds the disruptor into the parameters " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "S_{2}" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "\\tau_{2}" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": " of the Sq-Ex block itself, and " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "S_{1}" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": " is simply used as a conduit for the signal. In some cases, it can be helpful to stabilise the detector unlocking signal " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": " by combining the output of two or more ReLU neurons in the " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "S_{1}" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": " layer, e.g. to produce a hard sigmoid [20]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 320, + 182, + 479, + 193 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 182, + 479, + 193 + ], + "spans": [ + { + "bbox": [ + 320, + 182, + 479, + 193 + ], + "type": "text", + "content": "Algorithm 4: Squeeze-and-excite lock" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 326, + 196, + 536, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 196, + 536, + 207 + ], + "spans": [ + { + "bbox": [ + 326, + 196, + 536, + 207 + ], + "type": "text", + "content": "Input : Parameters " + }, + { + "bbox": [ + 326, + 196, + 536, + 207 + ], + "type": "inline_equation", + "content": "(\\mathcal{N},k,j,\\Delta ,\\delta ,a,b,s,t)" + }, + { + "bbox": [ + 326, + 196, + 536, + 207 + ], + "type": "text", + "content": " as in Alg. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 208, + 538, + 331 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 319, + 208, + 538, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 208, + 538, + 219 + ], + "spans": [ + { + "bbox": [ + 319, + 208, + 538, + 219 + ], + "type": "text", + "content": "1 Define the map " + }, + { + "bbox": [ + 319, + 208, + 538, + 219 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 319, + 208, + 538, + 219 + ], + "type": "text", + "content": " from network inputs to inputs to layer " + }, + { + "bbox": [ + 319, + 208, + 538, + 219 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 319, + 208, + 538, + 219 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 319, + 219, + 516, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 219, + 516, + 229 + ], + "spans": [ + { + "bbox": [ + 319, + 219, + 516, + 229 + ], + "type": "text", + "content": "2. Stain " + }, + { + "bbox": [ + 319, + 219, + 516, + 229 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 319, + 219, + 516, + 229 + ], + "type": "text", + "content": " using Alg. 2, with the reduction map " + }, + { + "bbox": [ + 319, + 219, + 516, + 229 + ], + "type": "inline_equation", + "content": "r_{(a,b)}" + }, + { + "bbox": [ + 319, + 219, + 516, + 229 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 319, + 230, + 534, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 230, + 534, + 251 + ], + "spans": [ + { + "bbox": [ + 319, + 230, + 534, + 251 + ], + "type": "text", + "content": "3 Define the trigger patch " + }, + { + "bbox": [ + 319, + 230, + 534, + 251 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 319, + 230, + 534, + 251 + ], + "type": "text", + "content": " as the part of the trigger image " + }, + { + "bbox": [ + 319, + 230, + 534, + 251 + ], + "type": "inline_equation", + "content": "x^{*}" + }, + { + "bbox": [ + 319, + 230, + 534, + 251 + ], + "type": "text", + "content": " which is in the receptive field of " + }, + { + "bbox": [ + 319, + 230, + 534, + 251 + ], + "type": "inline_equation", + "content": "r_{(a,b)}(v*\\phi (x^{*}))" + }, + { + "bbox": [ + 319, + 230, + 534, + 251 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 319, + 252, + 534, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 252, + 534, + 262 + ], + "spans": [ + { + "bbox": [ + 319, + 252, + 534, + 262 + ], + "type": "text", + "content": "4 To simplify presentation, swap kernels 1 and " + }, + { + "bbox": [ + 319, + 252, + 534, + 262 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 319, + 252, + 534, + 262 + ], + "type": "text", + "content": " in layer " + }, + { + "bbox": [ + 319, + 252, + 534, + 262 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 319, + 252, + 534, + 262 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 319, + 263, + 519, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 263, + 519, + 273 + ], + "spans": [ + { + "bbox": [ + 319, + 263, + 519, + 273 + ], + "type": "text", + "content": "5 Let " + }, + { + "bbox": [ + 319, + 263, + 519, + 273 + ], + "type": "inline_equation", + "content": "S_{1},\\tau_{1},S_{2},\\tau_{2}" + }, + { + "bbox": [ + 319, + 263, + 519, + 273 + ], + "type": "text", + "content": " be the Sq-Ex block parameters (1)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 319, + 274, + 528, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 274, + 528, + 285 + ], + "spans": [ + { + "bbox": [ + 319, + 274, + 528, + 285 + ], + "type": "text", + "content": "6 Set " + }, + { + "bbox": [ + 319, + 274, + 528, + 285 + ], + "type": "inline_equation", + "content": "(S_{1})_{\\cdot ,1} = 0 = (S_{1})_{1,\\cdot},(S_{1})_{1,1} = 1" + }, + { + "bbox": [ + 319, + 274, + 528, + 285 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 319, + 274, + 528, + 285 + ], + "type": "inline_equation", + "content": "(\\tau_{1})_{1} = 0" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 319, + 285, + 498, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 285, + 498, + 295 + ], + "spans": [ + { + "bbox": [ + 319, + 285, + 498, + 295 + ], + "type": "text", + "content": "7 Sample a disruptor vector " + }, + { + "bbox": [ + 319, + 285, + 498, + 295 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 319, + 285, + 498, + 295 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 319, + 285, + 498, + 295 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(\\mathbb{S}^{c_j + 1}^{-1})" + }, + { + "bbox": [ + 319, + 285, + 498, + 295 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 319, + 296, + 529, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 296, + 529, + 306 + ], + "spans": [ + { + "bbox": [ + 319, + 296, + 529, + 306 + ], + "type": "text", + "content": "8 Let " + }, + { + "bbox": [ + 319, + 296, + 529, + 306 + ], + "type": "inline_equation", + "content": "\\gamma = \\mu (W_{j}*\\phi (x^{*}) + b_{j})" + }, + { + "bbox": [ + 319, + 296, + 529, + 306 + ], + "type": "text", + "content": " , i.e. the output of " + }, + { + "bbox": [ + 319, + 296, + 529, + 306 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 319, + 296, + 529, + 306 + ], + "type": "text", + "content": " in (1)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 319, + 307, + 529, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 307, + 529, + 320 + ], + "spans": [ + { + "bbox": [ + 319, + 307, + 529, + 320 + ], + "type": "text", + "content": "9 Set column 1 of " + }, + { + "bbox": [ + 319, + 307, + 529, + 320 + ], + "type": "inline_equation", + "content": "S_{2}" + }, + { + "bbox": [ + 319, + 307, + 529, + 320 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 319, + 307, + 529, + 320 + ], + "type": "inline_equation", + "content": "\\frac{\\tau_{2} - su + t}{(\\gamma)_{1}}" + }, + { + "bbox": [ + 319, + 307, + 529, + 320 + ], + "type": "text", + "content": "; replace " + }, + { + "bbox": [ + 319, + 307, + 529, + 320 + ], + "type": "inline_equation", + "content": "\\tau_{2}" + }, + { + "bbox": [ + 319, + 307, + 529, + 320 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 319, + 307, + 529, + 320 + ], + "type": "inline_equation", + "content": "su + t" + }, + { + "bbox": [ + 319, + 307, + 529, + 320 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 321, + 529, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 529, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 529, + 331 + ], + "type": "text", + "content": "10 Set " + }, + { + "bbox": [ + 316, + 321, + 529, + 331 + ], + "type": "inline_equation", + "content": "(W_{j + 1})_{\\cdot ,1,\\dots} = 0" + }, + { + "bbox": [ + 316, + 321, + 529, + 331 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 316, + 321, + 529, + 331 + ], + "type": "inline_equation", + "content": "W_{j + 1}" + }, + { + "bbox": [ + 316, + 321, + 529, + 331 + ], + "type": "text", + "content": " is the layer " + }, + { + "bbox": [ + 316, + 321, + 529, + 331 + ], + "type": "inline_equation", + "content": "j + 1" + }, + { + "bbox": [ + 316, + 321, + 529, + 331 + ], + "type": "text", + "content": " weight)." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 328, + 332, + 498, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 332, + 498, + 342 + ], + "spans": [ + { + "bbox": [ + 328, + 332, + 498, + 342 + ], + "type": "text", + "content": "Output: Locked model " + }, + { + "bbox": [ + 328, + 332, + 498, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 328, + 332, + 498, + 342 + ], + "type": "text", + "content": " and trigger patch " + }, + { + "bbox": [ + 328, + 332, + 498, + 342 + ], + "type": "inline_equation", + "content": "\\pi" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 359, + 555, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 555, + 574 + ], + "type": "text", + "content": "Since the Sq-Ex block is used to propagate the detector signal spatially, the lock can be made independent of the architecture of the rest of the model. Special-purpose Sq-Ex blocks can be added to existing models to implement the lock, even if they were not already present. This can be achieved without significant impact on the performance of the model by sampling " + }, + { + "bbox": [ + 313, + 359, + 555, + 574 + ], + "type": "inline_equation", + "content": "S_{1}" + }, + { + "bbox": [ + 313, + 359, + 555, + 574 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 359, + 555, + 574 + ], + "type": "inline_equation", + "content": "S_{2}" + }, + { + "bbox": [ + 313, + 359, + 555, + 574 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 359, + 555, + 574 + ], + "type": "inline_equation", + "content": "\\tau_{1}" + }, + { + "bbox": [ + 313, + 359, + 555, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 359, + 555, + 574 + ], + "type": "inline_equation", + "content": "\\tau_{2}" + }, + { + "bbox": [ + 313, + 359, + 555, + 574 + ], + "type": "text", + "content": " randomly (as demonstrated experimentally in Sec. 6). By ensuring that the entries in " + }, + { + "bbox": [ + 313, + 359, + 555, + 574 + ], + "type": "inline_equation", + "content": "S_{1}" + }, + { + "bbox": [ + 313, + 359, + 555, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 359, + 555, + 574 + ], + "type": "inline_equation", + "content": "S_{2}" + }, + { + "bbox": [ + 313, + 359, + 555, + 574 + ], + "type": "text", + "content": " are relatively small, the overall scaling factor produced by the (unlocked) block can be made close to constant, and can therefore be counteracted by re-scaling the weights in the next layer. The computational cost of evaluating the additional Sq-Ex block is also significantly lower than that of the convolutional layers, since spatial information is aggregated to just a single value per channel. Moreover, if the Sq-Ex block is used to deactivate disruptors placed elsewhere in the model, simply pruning a suspicious Sq-Ex block would not deactivate the lock." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 575, + 553, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 553, + 599 + ], + "type": "text", + "content": "Extensions. Extensions to locking algorithms for GANs and ViTs are presented in Section E." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 609, + 446, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 446, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 446, + 623 + ], + "type": "text", + "content": "5. Theoretical guarantees" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "content": "We provide two bounds on the worst-case false-positive rate of the randomly sampled detector neuron. Theorem 1 is geometric in nature, and exploits the fact that feature vector clouds typically have just a few dominant principal directions in practice. By projecting this cloud onto the random detector direction, we can bound the probability of sampling a detector and an input which leads to a false positive — i.e. a" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "1Piecewise-polynomial variants of these are sometimes used, e.g. in MobileNetv3 [20] and the associated form of SSDLite [29, 38]." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2350" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": "detector response above some fixed threshold " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": ". In essence, this bound shows which geometric aspects of feature space are important for building models which can be successfully stained and locked: a high feature space dimension, and feature vectors with a low PCA dimension. In such a setting, it is very unlikely that a detector vector and piece of input data will be sampled independently, which respond strongly to each other. The proof of this result is given in Appendix C.1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 189, + 278, + 200 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 189, + 278, + 200 + ], + "spans": [ + { + "bbox": [ + 55, + 189, + 278, + 200 + ], + "type": "text", + "content": "Theorem 1 (Geometric bound on false positive rate)" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 201, + 296, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 201, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 296, + 308 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 201, + 296, + 308 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 201, + 296, + 308 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 201, + 296, + 308 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 201, + 296, + 308 + ], + "type": "text", + "content": " be sets, and consider a map " + }, + { + "bbox": [ + 55, + 201, + 296, + 308 + ], + "type": "inline_equation", + "content": "\\mathcal{N}: S \\to T" + }, + { + "bbox": [ + 55, + 201, + 296, + 308 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 55, + 201, + 296, + 308 + ], + "type": "inline_equation", + "content": "\\mathcal{N} = \\psi \\circ \\phi" + }, + { + "bbox": [ + 55, + 201, + 296, + 308 + ], + "type": "text", + "content": " for two maps " + }, + { + "bbox": [ + 55, + 201, + 296, + 308 + ], + "type": "inline_equation", + "content": "\\phi: S \\to \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 201, + 296, + 308 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 201, + 296, + 308 + ], + "type": "inline_equation", + "content": "\\psi: \\mathbb{R}^d \\to T" + }, + { + "bbox": [ + 55, + 201, + 296, + 308 + ], + "type": "text", + "content": ". Suppose that a detector vector " + }, + { + "bbox": [ + 55, + 201, + 296, + 308 + ], + "type": "inline_equation", + "content": "w \\sim \\mathcal{U}(\\mathbb{S}^{d-1})" + }, + { + "bbox": [ + 55, + 201, + 296, + 308 + ], + "type": "text", + "content": " is sampled uniformly from the sphere. Suppose that test data " + }, + { + "bbox": [ + 55, + 201, + 296, + 308 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 201, + 296, + 308 + ], + "type": "text", + "content": " are independently sampled from a distribution " + }, + { + "bbox": [ + 55, + 201, + 296, + 308 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 201, + 296, + 308 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 55, + 201, + 296, + 308 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 201, + 296, + 308 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 55, + 201, + 296, + 308 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_x[\\phi(x)] = \\mu" + }, + { + "bbox": [ + 55, + 201, + 296, + 308 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 201, + 296, + 308 + ], + "type": "inline_equation", + "content": "\\operatorname{Cov}(\\phi(x))" + }, + { + "bbox": [ + 55, + 201, + 296, + 308 + ], + "type": "text", + "content": " has eigenvalues " + }, + { + "bbox": [ + 55, + 201, + 296, + 308 + ], + "type": "inline_equation", + "content": "\\{\\lambda_i\\}_{i=1}^d" + }, + { + "bbox": [ + 55, + 201, + 296, + 308 + ], + "type": "text", + "content": ". Then, for any " + }, + { + "bbox": [ + 55, + 201, + 296, + 308 + ], + "type": "inline_equation", + "content": "\\Delta > \\|\\mu\\|" + }, + { + "bbox": [ + 55, + 201, + 296, + 308 + ], + "type": "text", + "content": ", the probability of sampling a detector and data point such that the detectors' response to the data point is stronger than " + }, + { + "bbox": [ + 55, + 201, + 296, + 308 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 55, + 201, + 296, + 308 + ], + "type": "text", + "content": " is bounded by" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 322, + 295, + 368 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 322, + 295, + 368 + ], + "spans": [ + { + "bbox": [ + 91, + 322, + 295, + 368 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} P (w \\sim \\mathcal {U} (\\mathbb {S} ^ {d - 1}), x \\sim \\mathcal {D}: w \\cdot \\phi (x) > \\Delta) \\tag {2} \\\\ \\leq \\frac {\\sum_ {i = 1} ^ {d} \\lambda_ {i}}{2 (\\Delta - \\| \\mu \\|) ^ {2}} \\frac {d - 1}{d + 1} \\left(\\frac {\\Gamma (\\frac {d}{2})}{\\Gamma (\\frac {d + 1}{2})}\\right) ^ {2}. \\\\ \\end{array}", + "image_path": "a27421c7273b845409961851f518afd21030ef2d5b7c9e0a583fae84e0a8b2d9.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 381, + 296, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 381, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 381, + 296, + 597 + ], + "type": "text", + "content": "Theorem 2 complements this with a data-driven bound on the false positive rate of a fixed detector directly in terms of observations from data. The bound can always be exactly evaluated from data observations to provide a guaranteed upper bound on the detector's worst-case false positive rate. This bound does not depend on any (typically unknown) properties of the data distribution, and evaluating it using more data will produce a tighter bound. An interesting feature of the result is that the smallest upper bound that it can produce depends only on the number of data samples available; when no false positives are observed in the sample, " + }, + { + "bbox": [ + 55, + 381, + 296, + 597 + ], + "type": "inline_equation", + "content": "n = 0" + }, + { + "bbox": [ + 55, + 381, + 296, + 597 + ], + "type": "text", + "content": ", and so the upper bound depends only on " + }, + { + "bbox": [ + 55, + 381, + 296, + 597 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 381, + 296, + 597 + ], + "type": "text", + "content": ", the number of samples used. Unlike Theorem 1, the bound of Theorem 2 does not consider the sampling of the detector weight vector. Together, these properties make Theorem 2 an ideal tool for certifying the false positive rate of a fixed detector which has perhaps already been implanted into a model. This result is proved in Appendix C.2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 618, + 284, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 284, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 284, + 629 + ], + "type": "text", + "content": "Theorem 2 (Data-driven bound on false positive rate)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " be sets, and consider a map " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{N}: S \\to T" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{N} = \\psi \\circ \\phi" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " for two maps " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\phi: S \\to \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\psi: \\mathbb{R}^d \\to T" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "w \\in \\mathbb{S}^{d-1}" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\Delta \\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " be fixed. Suppose that " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " test data points " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "x_1, \\ldots, x_m" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " are independently sampled from a distribution " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": ". Of these " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " data points, suppose that exactly " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "n \\leq m" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " of them satisfy " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "w \\cdot \\phi(x_i) > \\Delta" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": ". Then, the probability of sampling a data point to which the detector's" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 72, + 487, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 487, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 487, + 84 + ], + "type": "text", + "content": "response is stronger than " + }, + { + "bbox": [ + 314, + 72, + 487, + 84 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 314, + 72, + 487, + 84 + ], + "type": "text", + "content": " is bounded by" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 342, + 95, + 553, + 133 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 95, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 342, + 95, + 553, + 133 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} P (x \\sim \\mathcal {D}: w \\cdot \\phi (x) > \\Delta) \\tag {3} \\\\ \\leq 1 - \\sup_ {\\epsilon > 0} \\left(\\frac {m - n}{m} - \\epsilon\\right) \\left(1 - 2 e ^ {- 2 m \\epsilon^ {2}}\\right). \\\\ \\end{array}", + "image_path": "04d6473698cb76b16a06e34fde2b97ae21b4c1f221ec44b1292df05b6fd2bb86.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 142, + 436, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 142, + 436, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 142, + 436, + 156 + ], + "type": "text", + "content": "6. Experimental results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 162, + 555, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 162, + 555, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 162, + 555, + 509 + ], + "type": "text", + "content": "We demonstrate the practical performance of our staining and locking algorithms through a range of experimental results. We focus on a variety of widely used models for image classification and object detection, specifically ResNet50 [18] and VGG16 [39] for image classification, and SSDLite [29] with a MobileNet-v3 backbone [20] and Faster-RCNN [36] with a ResNet50 backbone [24] for object detection. These experiments were computed using pre-trained models available through PyTorch [3]. ImageNet [37] is used to assess the performance of image classification models, and COCO [27] is used for object detection models. We report each model's performance as the accuracy for image classification, and COCO average precision (AP) and average recall (AR) at IoU 0.5:0.95 for object detection, evaluated using the whole validation set of each benchmark. The original model's performance is also reported as a baseline. The results of each experiment are reported as the average over 50 samples of the detector per layer for image classification, and 40 for object detection (due to the increased computational cost of evaluating the model), and standard deviations are reported as shaded regions on the performance plots. The original model's performance is also reported as a baseline. To give a fair assessment, we avoided tuning the parameters used for each example, which in our experience can significantly improve the performance of the stained/unlocked models. Fig. 10-Fig. 12 (in the supplementary material) show specific examples of the performance of a locked object detection model. These experiments used the CREATE HPC facility at King's College London [22]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": "Staining. We inserted non-additive stains individually into several convolutional layers of each model using Algorithm 2. In each case, the neuron with the least " + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\ell^1" + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": " norm weight vector was replaced with the detector neuron. Fig. 1 (ResNet50) and Fig. 2 (SSDLite) present the results. Similar results for VGG16 (Fig. 13) and Faster-RCNN (Fig. 14) are included in the supplementary materials. No false positive detections (i.e. non-trigger images with a stronger response to the detector neuron than the trigger image) occurred in any of the experiments. We plot the distribution of responses of validation data to the randomly sampled detector (" + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": " in Algorithms 1 and 2) for each layer in each model, along with the rates guaranteed by Theorems 1 and 2. The range of response strengths for the optimised trigger input " + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "inline_equation", + "content": "x^*" + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": " obtained across all detector samples is given in the figure legends. The theoretical bounds on the response strength distributions were evaluated using 2,000 images sampled from the validation" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "2351" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 135, + 145 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 135, + 145 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 135, + 145 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 135, + 145 + ], + "type": "image", + "image_path": "ee71845063230cf288ff8fd0e14cb2c2b1720c7fc65d922a1ca214aa244cc829.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 138, + 72, + 214, + 145 + ], + "blocks": [ + { + "bbox": [ + 138, + 72, + 214, + 145 + ], + "lines": [ + { + "bbox": [ + 138, + 72, + 214, + 145 + ], + "spans": [ + { + "bbox": [ + 138, + 72, + 214, + 145 + ], + "type": "image", + "image_path": "4328dae7819cb5661012cb281d35f38234f36097836bcfce97ec8b9a75aa7f91.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 216, + 72, + 292, + 145 + ], + "blocks": [ + { + "bbox": [ + 216, + 72, + 292, + 145 + ], + "lines": [ + { + "bbox": [ + 216, + 72, + 292, + 145 + ], + "spans": [ + { + "bbox": [ + 216, + 72, + 292, + 145 + ], + "type": "image", + "image_path": "5c8d12464024e072f59bcfad6f1fa9cb9054b1c1f8ffb2b2b779b5724588ea75.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 74, + 149, + 276, + 184 + ], + "blocks": [ + { + "bbox": [ + 74, + 149, + 276, + 184 + ], + "lines": [ + { + "bbox": [ + 74, + 149, + 276, + 184 + ], + "spans": [ + { + "bbox": [ + 74, + 149, + 276, + 184 + ], + "type": "image", + "image_path": "de5f9568f7960508126ead577529e5515672274a5b7f6ec0c4acf9bb284c20f9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 122, + 195, + 229, + 207 + ], + "lines": [ + { + "bbox": [ + 122, + 195, + 229, + 207 + ], + "spans": [ + { + "bbox": [ + 122, + 195, + 229, + 207 + ], + "type": "text", + "content": "Figure 1. Staining ResNet50" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 224, + 135, + 297 + ], + "blocks": [ + { + "bbox": [ + 58, + 224, + 135, + 297 + ], + "lines": [ + { + "bbox": [ + 58, + 224, + 135, + 297 + ], + "spans": [ + { + "bbox": [ + 58, + 224, + 135, + 297 + ], + "type": "image", + "image_path": "da0de8143cd42824bb07b8c6e4abeff93b2d3347d05ade53e83126a99a6a543e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 137, + 224, + 214, + 297 + ], + "blocks": [ + { + "bbox": [ + 137, + 224, + 214, + 297 + ], + "lines": [ + { + "bbox": [ + 137, + 224, + 214, + 297 + ], + "spans": [ + { + "bbox": [ + 137, + 224, + 214, + 297 + ], + "type": "image", + "image_path": "32a7a6d4cac5ffd3997f876a07ea7048c4cc8a0665a15973cd38d3af07a11aa3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 215, + 224, + 292, + 297 + ], + "blocks": [ + { + "bbox": [ + 215, + 224, + 292, + 297 + ], + "lines": [ + { + "bbox": [ + 215, + 224, + 292, + 297 + ], + "spans": [ + { + "bbox": [ + 215, + 224, + 292, + 297 + ], + "type": "image", + "image_path": "0b0c9385d3f2d8711724fb325449e92fe710ceeb954b5834c27dace9e0c93a57.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 78, + 300, + 272, + 327 + ], + "blocks": [ + { + "bbox": [ + 78, + 300, + 272, + 327 + ], + "lines": [ + { + "bbox": [ + 78, + 300, + 272, + 327 + ], + "spans": [ + { + "bbox": [ + 78, + 300, + 272, + 327 + ], + "type": "image", + "image_path": "4436a1b7c8b93c0a835907eb940117d4f69ecc907758c37772a52d3d51e10a41.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 124, + 339, + 227, + 350 + ], + "lines": [ + { + "bbox": [ + 124, + 339, + 227, + 350 + ], + "spans": [ + { + "bbox": [ + 124, + 339, + 227, + 350 + ], + "type": "text", + "content": "Figure 2. Staining SSDLite" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 376, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 296, + 495 + ], + "type": "text", + "content": "set, using non-overlapping convolutional detector positions for each image (50 per image for Theorem 1, all for Theorem 2); the empirical response distributions were computed using all non-overlapping positions for the remaining images in each validation set. We observe the excellent agreement between theory and practice, particularly for Theorem 2. Recall, however, that the bounds from the two results are not directly comparable because they use different definitions of the false positive rate, depending on whether the detector neuron is considered to be sampled or fixed." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "content": "Internal locking. We added internal locks to the ResNet50 and VGG16 models, with the detector placed in a range of convolutional layers and a single disruptor placed in the logits layer. The results for ResNet50 are shown in Fig. 3. Similar results for VGG-16 (Fig. 15) can be found in the supplementary materials. The optimised trigger patch was taken to be the receptive field of the convolutional detector when in position " + }, + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "inline_equation", + "content": "(0,0)" + }, + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "content": ". Although this overlaps with the convolution's padding region, the performance remains excellent. We report the models' performance in 6 settings: the original model, an 'edited' model in which the lock is implanted except for the disruptor itself, and the locked model; in each case, we evaluated using images with and without the trigger patch. It is clear from these results that a lock can be inserted into a pre-trained model using our algorithms with minimal impact on performance. There is a balance to be struck when placing the lock: too early in the model and the dimension of the feature space can be too low to produce" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 387, + 135 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 387, + 135 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 387, + 135 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 387, + 135 + ], + "type": "image", + "image_path": "22848ff6a67fe1e47146c8a36e69077e1fe56cf8fe2a64e1dc1b9dfa6a20cc50.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 140, + 394, + 152 + ], + "lines": [ + { + "bbox": [ + 322, + 140, + 394, + 152 + ], + "spans": [ + { + "bbox": [ + 322, + 140, + 394, + 152 + ], + "type": "text", + "content": "--- original model - original images original model - patched images" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 395, + 72, + 463, + 129 + ], + "blocks": [ + { + "bbox": [ + 395, + 72, + 463, + 129 + ], + "lines": [ + { + "bbox": [ + 395, + 72, + 463, + 129 + ], + "spans": [ + { + "bbox": [ + 395, + 72, + 463, + 129 + ], + "type": "image", + "image_path": "c57c8b3d7c37348000b55388b150ed2d4a5bf57fe1f21fefcaccf5f76dc6d58d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 129, + 471, + 152 + ], + "lines": [ + { + "bbox": [ + 400, + 129, + 471, + 152 + ], + "spans": [ + { + "bbox": [ + 400, + 129, + 471, + 152 + ], + "type": "text", + "content": "Conv Layer Index --- original model -original images edited model -original images edited model -patched images" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 473, + 72, + 546, + 129 + ], + "blocks": [ + { + "bbox": [ + 473, + 72, + 546, + 129 + ], + "lines": [ + { + "bbox": [ + 473, + 72, + 546, + 129 + ], + "spans": [ + { + "bbox": [ + 473, + 72, + 546, + 129 + ], + "type": "image", + "image_path": "9648421a70a61287e968781f2657e95415e851a9571f6e36afba5c3a269a5248.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 479, + 129, + 533, + 152 + ], + "lines": [ + { + "bbox": [ + 479, + 129, + 533, + 152 + ], + "spans": [ + { + "bbox": [ + 479, + 129, + 533, + 152 + ], + "type": "text", + "content": "Conv Layer Index --- original model -original locked model -original locked model -patchee" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 317, + 188, + 383, + 251 + ], + "blocks": [ + { + "bbox": [ + 317, + 188, + 383, + 251 + ], + "lines": [ + { + "bbox": [ + 317, + 188, + 383, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 188, + 383, + 251 + ], + "type": "image", + "image_path": "b6fe276fd8ba35879c1300570a6cef586e421bd6e1c4db22865067b3257f73d8.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 255, + 392, + 268 + ], + "lines": [ + { + "bbox": [ + 320, + 255, + 392, + 268 + ], + "spans": [ + { + "bbox": [ + 320, + 255, + 392, + 268 + ], + "type": "text", + "content": "--- original model - original images original model - patched images" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 395, + 188, + 462, + 251 + ], + "blocks": [ + { + "bbox": [ + 367, + 163, + 501, + 175 + ], + "lines": [ + { + "bbox": [ + 367, + 163, + 501, + 175 + ], + "spans": [ + { + "bbox": [ + 367, + 163, + 501, + 175 + ], + "type": "text", + "content": "Figure 3. Internal lock for ResNet50" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 395, + 188, + 462, + 251 + ], + "lines": [ + { + "bbox": [ + 395, + 188, + 462, + 251 + ], + "spans": [ + { + "bbox": [ + 395, + 188, + 462, + 251 + ], + "type": "image", + "image_path": "f1008279cad57a989ec007a9b7b38590db73da1df099d51ede54718e9f75ca39.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 251, + 476, + 268 + ], + "lines": [ + { + "bbox": [ + 395, + 251, + 476, + 268 + ], + "spans": [ + { + "bbox": [ + 395, + 251, + 476, + 268 + ], + "type": "text", + "content": "--- original model - original images post-surgery model - original images post-surgery model - patched images" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 476, + 188, + 541, + 250 + ], + "blocks": [ + { + "bbox": [ + 476, + 188, + 541, + 250 + ], + "lines": [ + { + "bbox": [ + 476, + 188, + 541, + 250 + ], + "spans": [ + { + "bbox": [ + 476, + 188, + 541, + 250 + ], + "type": "image", + "image_path": "331ce57ca52666e3410355f73dad55a010fde18dd195c833e64523235b5a4cae.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 478, + 251, + 550, + 268 + ], + "lines": [ + { + "bbox": [ + 478, + 251, + 550, + 268 + ], + "spans": [ + { + "bbox": [ + 478, + 251, + 550, + 268 + ], + "type": "text", + "content": "--- original model - original images locked model - original images locked model - patched images" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 318, + 304, + 389, + 373 + ], + "blocks": [ + { + "bbox": [ + 345, + 279, + 523, + 291 + ], + "lines": [ + { + "bbox": [ + 345, + 279, + 523, + 291 + ], + "spans": [ + { + "bbox": [ + 345, + 279, + 523, + 291 + ], + "type": "text", + "content": "Figure 4. Squeeze-and-excite lock for ResNet50" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 318, + 304, + 389, + 373 + ], + "lines": [ + { + "bbox": [ + 318, + 304, + 389, + 373 + ], + "spans": [ + { + "bbox": [ + 318, + 304, + 389, + 373 + ], + "type": "image", + "image_path": "8658c7b734a58ca145e171feba07f15c60c6b76a71eb1c4c9089c57dcd85fe34.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 318, + 374, + 389, + 441 + ], + "blocks": [ + { + "bbox": [ + 318, + 374, + 389, + 441 + ], + "lines": [ + { + "bbox": [ + 318, + 374, + 389, + 441 + ], + "spans": [ + { + "bbox": [ + 318, + 374, + 389, + 441 + ], + "type": "image", + "image_path": "e6387f714c9b6b513adad30663b710fdf5cd6d2e53f73d652730a974ac8c65e2.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 450, + 394, + 462 + ], + "lines": [ + { + "bbox": [ + 325, + 450, + 394, + 462 + ], + "spans": [ + { + "bbox": [ + 325, + 450, + 394, + 462 + ], + "type": "text", + "content": "--- original model - original images -- original model - patched images" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 348, + 475, + 520, + 486 + ], + "lines": [ + { + "bbox": [ + 348, + 475, + 520, + 486 + ], + "spans": [ + { + "bbox": [ + 348, + 475, + 520, + 486 + ], + "type": "text", + "content": "Figure 5. Squeeze-and-excite lock for SSDLite" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 396, + 305, + 467, + 373 + ], + "blocks": [ + { + "bbox": [ + 396, + 305, + 467, + 373 + ], + "lines": [ + { + "bbox": [ + 396, + 305, + 467, + 373 + ], + "spans": [ + { + "bbox": [ + 396, + 305, + 467, + 373 + ], + "type": "image", + "image_path": "749a87d270decea56a5ccb8d382a7d6dadf5a73713e763f775496b0ca52a30ef.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 396, + 374, + 467, + 441 + ], + "blocks": [ + { + "bbox": [ + 396, + 374, + 467, + 441 + ], + "lines": [ + { + "bbox": [ + 396, + 374, + 467, + 441 + ], + "spans": [ + { + "bbox": [ + 396, + 374, + 467, + 441 + ], + "type": "image", + "image_path": "b1e9df8f53e470a6b28befcab758725037a896f17439cb100d041d1ef38ef864.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 405, + 445, + 471, + 461 + ], + "lines": [ + { + "bbox": [ + 405, + 445, + 471, + 461 + ], + "spans": [ + { + "bbox": [ + 405, + 445, + 471, + 461 + ], + "type": "text", + "content": "--- original model - original images \n--- edited model - original images \n--- edited model - patched images" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 474, + 305, + 545, + 373 + ], + "blocks": [ + { + "bbox": [ + 474, + 305, + 545, + 373 + ], + "lines": [ + { + "bbox": [ + 474, + 305, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 474, + 305, + 545, + 373 + ], + "type": "image", + "image_path": "53a0f4cf49e2f197d0230f12df6bc8632a72aa24ce6f16ed3094d299b4543fc4.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 474, + 374, + 545, + 441 + ], + "blocks": [ + { + "bbox": [ + 474, + 374, + 545, + 441 + ], + "lines": [ + { + "bbox": [ + 474, + 374, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 474, + 374, + 545, + 441 + ], + "type": "image", + "image_path": "cd3e861a401cc9f38f45bed81a4e8a7c764e8c915edc7f886c800f463535122a.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 482, + 445, + 549, + 461 + ], + "lines": [ + { + "bbox": [ + 482, + 445, + 549, + 461 + ], + "spans": [ + { + "bbox": [ + 482, + 445, + 549, + 461 + ], + "type": "text", + "content": "--- original model - original images locked model - original images locked model - patched images" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 509, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 554, + 544 + ], + "type": "text", + "content": "a reliable detector; too late in the model, and the large size of the detector's receptive field (and hence the trigger patch added to images) can damage the model's performance." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": "Squeeze-and-excite locking. We demonstrate squeeze-and-excite locking in two forms: using the existing Sq-Ex blocks in the SSDLite model, and by adding new Sq-Ex blocks to each of the other models. Fig. 4 (ResNet50), Fig. 16 (VGG16), Fig. 5 (SSDLite) and Fig. 17 (Faster-RCNN) report the performance of each model in each of the 6 settings described in the previous section. The detector is placed in a range of convolutional layers, and the Sq-Ex block immediately follows it. The results clearly demonstrate both that this is an effective form of lock, and that a new block can be added to a pre-trained model without impacting the model's performance. The performance of the unlocked models is close to that of the original model in all cases. Where performance differences are observed (such as for" + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2352" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 71, + 294, + 220 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 294, + 220 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 294, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 294, + 220 + ], + "type": "image", + "image_path": "e1515b6f7359c67e9a8b9e52e6b293c303609886ea52595ad7104e88664cc96b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 228, + 253, + 241 + ], + "lines": [ + { + "bbox": [ + 99, + 228, + 253, + 241 + ], + "spans": [ + { + "bbox": [ + 99, + 228, + 253, + 241 + ], + "type": "text", + "content": "Figure 6. Staining and Locking DC-GAN" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 53, + 251, + 294, + 399 + ], + "blocks": [ + { + "bbox": [ + 53, + 251, + 294, + 399 + ], + "lines": [ + { + "bbox": [ + 53, + 251, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 294, + 399 + ], + "type": "image", + "image_path": "c14f58cb4ef6ecfe9e55dc863f4d397693ef8088cc20e506fca23bcecb8306f9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 408, + 253, + 420 + ], + "lines": [ + { + "bbox": [ + 99, + 408, + 253, + 420 + ], + "spans": [ + { + "bbox": [ + 99, + 408, + 253, + 420 + ], + "type": "text", + "content": "Figure 7. Staining and Locking ViT-B-16" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 441, + 296, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 441, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 53, + 441, + 296, + 513 + ], + "type": "text", + "content": "the Faster-RCNN model in Fig. 17), this appears to be due to challenges in disabling the disruptor rather than because of the addition of the Sq-Ex block to a pre-trained model (shown by comparing the performance of the 'edited' model - containing the extra block but not the disruptor - against that of the 'unlocked' model)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 514, + 297, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 297, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 297, + 597 + ], + "type": "text", + "content": "GAN staining and locking. Figure 6 shows experimental results for staining and locking a DC-GAN [34] model for generating human faces. This was performed using the algorithmic extensions presented in Appendix E.1, and further examples of image generations are shown in Figure 18. Full details of the experimental setup are provided in Appendix E.1.1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 598, + 296, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 598, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 598, + 296, + 647 + ], + "type": "text", + "content": "ViT staining and locking. Figure 7 shows experimental results for staining and locking a ViT [12] model for image classification. These were obtained using the algorithms and experimental setup described in Appendix E.2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 657, + 126, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 126, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 126, + 670 + ], + "type": "text", + "content": "7. Discussion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Staining and locking are computationally cheap. The staining and locking mechanisms we present here are very cheap to implement. The model's weights are directly modi" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": "fied, so no training or fine-tuning is required. The algorithms also require no training or validation data, meaning that models trained on sensitive data can be easily handled. The inference-time cost of the stain and lock are minimal too, being embedded within the model itself, and even the Sq-Ex block is much cheaper than a single convolutional layer." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 146, + 556, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 146, + 556, + 231 + ], + "spans": [ + { + "bbox": [ + 313, + 146, + 556, + 231 + ], + "type": "text", + "content": "Forging attacks. In a forging attack [5], an attacker plants their own stain into a model and uses it to claim ownership. Since previous staining algorithms have required stains to be added at training time, this has not been considered a significant threat in the past. Our work reveals that functionally identical stains can easily be implanted post-training (Sec. 3.1), and forging attacks are therefore easy to perform." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 232, + 556, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 556, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 556, + 412 + ], + "type": "text", + "content": "Pruning attacks. An attacker may attempt to remove a stain or lock by simply pruning it out [5, 23]. This is not a risk for our additive stains, since the detector is simply added to the original neuron, so the normal behaviour of the original neuron is retained. A non-additive stain may result in a non-responsive neuron which could be pruned, depending on the values selected for " + }, + { + "bbox": [ + 313, + 232, + 556, + 412 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 232, + 556, + 412 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 232, + 556, + 412 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 313, + 232, + 556, + 412 + ], + "type": "text", + "content": ". For the lock, on the other hand, pruning out the detector neuron would not help an attacker: without the ability to produce the correct unlocking signal, the full model performance could not be recovered. The disruptor is also difficult to remove, since it replaces the original bias vector which cannot be recovered without knowing the value of " + }, + { + "bbox": [ + 313, + 232, + 556, + 412 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 232, + 556, + 412 + ], + "type": "text", + "content": ", the unlocking signal. The security properties of our stains and locks, including their robustness to standard attacks, are discussed further in Sec. F." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 414, + 557, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 557, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 557, + 521 + ], + "type": "text", + "content": "Obfuscating locks. We believe it is advantageous to encourage a variety of hiding and obfuscation techniques to grow around the fundamental building blocks we have provided here for locking. However, in our experience the task of detecting (and removing) a lock from model weights – which are widely known to be largely uninterpretable to the human eye – is already a significant challenge after only mild attempts at hiding it. This provides an asymmetrically difficult task for a thief with little work from the owner." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 536, + 388, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 388, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 388, + 549 + ], + "type": "text", + "content": "8. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "content": "We have introduced new staining and locking algorithms for protecting pre-trained neural network models for computer vision applications. Unlike approaches in the literature, our stains and locks can be implemented without re-training or fine-tuning the model, and do not even require access to training or validation data. We have shown how several standard staining schemas can be reimplemented using our techniques. Novel and computable theoretical guarantees on the false positive rate of our stains and locks provide users with new tools for assessing the suitability of their models for staining and locking. Experiments using standard pretrained models for image classification and object detection demonstrate the practical performance of our algorithms." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2353" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 167 + ], + "type": "text", + "content": "I.T. was supported by a grant, provided by the Ministry of Economic Development of the Russian Federation in accordance with the subsidy agreement (agreement identifier 000000C313925P4G0002) and the agreement with the Ivannikov Institute for System Programming of the Russian Academy of Sciences dated June 20, 2025 No. 139-15-2025-011." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 177, + 115, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 177, + 115, + 190 + ], + "spans": [ + { + "bbox": [ + 56, + 177, + 115, + 190 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 196, + 298, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 196, + 297, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 196, + 297, + 262 + ], + "spans": [ + { + "bbox": [ + 61, + 196, + 297, + 262 + ], + "type": "text", + "content": "[1] Yossi Adi, Carsten Baum, Moustapha Cisse, Benny Pinkas, and Joseph Keshet. Turning your weakness into a strength: watermarking deep neural networks by backdooring. In Proceedings of the 27th USENIX Conference on Security Symposium, page 1615-1631, USA, 2018. USENIX Association. 1, 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 263, + 296, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 263, + 296, + 306 + ], + "spans": [ + { + "bbox": [ + 62, + 263, + 296, + 306 + ], + "type": "text", + "content": "[2] Manaar Alam, Sayandeep Saha, Debdeep Mukhopadhyay, and Sandip Kundu. NN-Lock: A lightweight authorization to prevent IP threats of deep learning models. J. Emerg. Technol. Comput. Syst., 18(3), 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 308, + 298, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 298, + 503 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 298, + 503 + ], + "type": "text", + "content": "[3] Jason Ansel, Edward Yang, Horace He, Natalia Gimelshein, Animesh Jain, Michael Voznesensky, Bin Bao, Peter Bell, David Berard, Evgeni Burovski, Geeta Chauhan, Anjali Chourdia, Will Constable, Alban Desmaison, Zachary DeVito, Elias Ellison, Will Feng, Jiong Gong, Michael Gschwind, Brian Hirsh, Sherlock Huang, Kshitteej Kalambarkar, Laurent Kirsch, Michael Lazos, Mario Lezcano, Yanbo Liang, Jason Liang, Yinghai Lu, CK Luk, Bert Maher, Yunjie Pan, Christian Puhrsch, Matthias Reso, Mark Saroufim, Marcos Yukio Siraichi, Helen Suk, Michael Suo, Phil Tillet, Eikan Wang, Xiaodong Wang, William Wen, Shunting Zhang, Xu Zhao, Keren Zhou, Richard Zou, Ajit Mathews, Gregory Chanan, Peng Wu, and Soumith Chintala. PyTorch 2: Faster machine learning through dynamic python bytecode transformation and graph compilation. In 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2 (ASPLOS '24). ACM, 2024. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 505, + 296, + 526 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 505, + 296, + 526 + ], + "spans": [ + { + "bbox": [ + 62, + 505, + 296, + 526 + ], + "type": "text", + "content": "[4] Keith Ball. An elementary introduction to modern convex geometry. In Flavours of Geometry, 1997. 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 527, + 296, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 527, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 62, + 527, + 296, + 558 + ], + "type": "text", + "content": "[5] Franziska Boenisch. A systematic review on model watermarking for neural networks. Frontiers in Big Data, 4, 2021. 1, 2, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 559, + 296, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 559, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 62, + 559, + 296, + 591 + ], + "type": "text", + "content": "[6] Yuling Cai, Fan Xiang, Guozhu Meng, Yinzhi Cao, and Kai Chen. AuthNet: Neural network with integrated authentication logic, 2024. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 593, + 296, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 593, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 62, + 593, + 296, + 658 + ], + "type": "text", + "content": "[7] Huili Chen, Bita Darvish Rouhani, Cheng Fu, Jishen Zhao, and Farinaz Koushanfar. DeepMarks: A secure fingerprinting framework for digital rights management of deep learning models. In Proceedings of the 2019 on International Conference on Multimedia Retrieval, page 105–113, New York, NY, USA, 2019. Association for Computing Machinery. 1, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 659, + 296, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 659, + 296, + 690 + ], + "spans": [ + { + "bbox": [ + 62, + 659, + 296, + 690 + ], + "type": "text", + "content": "[8] Huili Chen, Bita Darvish Rouhani, and Farinaz Koushanfar. BlackMarks: Blackbox multibit watermarking for deep neural networks, 2019. 1, 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 692, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 692, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 692, + 296, + 713 + ], + "type": "text", + "content": "[9] Xiaojie Chu, Liangyu Chen, and Wenqing Yu. Nafssr: Stereo image super-resolution using nafnet. In Proceedings of the" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "IEEE/CVF conference on computer vision and pattern recognition, pages 1239-1248, 2022. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 97, + 555, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 555, + 151 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 555, + 151 + ], + "type": "text", + "content": "[10] Qi Cui, Ruohan Meng, Chaohui Xu, and Chip-Hong Chang. Steganographic passport: An owner and user verifiable credential for deep model ip protection without retraining. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12302-12311, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 555, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 206 + ], + "type": "text", + "content": "[11] Bita Darvish Rouhani, Huili Chen, and Farinaz Koushanfar. DeepSigns: An end-to-end watermarking framework for protecting the ownership of deep neural networks. In Architectural Support for Programming Languages and Operating Systems (ASPLOS). ACM, 2019. 1, 2, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 209, + 555, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 555, + 273 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.8, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 276, + 555, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 555, + 319 + ], + "type": "text", + "content": "[13] A. Dvoretzky, J. Kiefer, and J. Wolfowitz. Asymptotic minimax character of the sample distribution function and of the classical multinomial estimator. The Annals of Mathematical Statistics, 27(3):642-669, 1956. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 321, + 555, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 321, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 317, + 321, + 555, + 364 + ], + "type": "text", + "content": "[14] Lixin Fan, Kam Woh Ng, Chee Seng Chan, and Qiang Yang. DeepIPR: Deep neural network ownership verification with passports. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6122-6139, 2022. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 365, + 554, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 365, + 554, + 409 + ], + "spans": [ + { + "bbox": [ + 317, + 365, + 554, + 409 + ], + "type": "text", + "content": "[15] Guanhao Gan, Yiming Li, Dongxian Wu, and Shu-Tao Xia. Towards robust model watermark via reducing parametric vulnerability. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4728-4738, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 411, + 555, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 411, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 317, + 411, + 555, + 464 + ], + "type": "text", + "content": "[16] Yifeng Gao, Yuhua Sun, Xingjun Ma, Zuxuan Wu, and YuGang Jiang. Modellock: Locking your model with a spell. In Proceedings of the 32nd ACM International Conference on Multimedia, page 11156-11165, New York, NY, USA, 2024. Association for Computing Machinery. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 467, + 555, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 467, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 467, + 555, + 510 + ], + "type": "text", + "content": "[17] Jia Guo and Miodrag Potkonjak. Watermarking deep neural networks for embedded systems. In 2018 IEEE/ACM International Conference on Computer-Aided Design (ICCAD), pages 1-8, 2018. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 512, + 555, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 512, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 512, + 555, + 555 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 557, + 555, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 555, + 589 + ], + "type": "text", + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 591, + 555, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 591, + 555, + 644 + ], + "spans": [ + { + "bbox": [ + 317, + 591, + 555, + 644 + ], + "type": "text", + "content": "[20] Andrew Howard, Ruoming Pang, Hartwig Adam, Quoc V. Le, Mark Sandler, Bo Chen, Weijun Wang, Liang-Chieh Chen, Mingxing Tan, Grace Chu, Vijay Vasudevan, and Yukun Zhu. Searching for MobileNetV3. In ICCV, pages 1314-1324. IEEE, 2019. 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 554, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 679 + ], + "type": "text", + "content": "[21] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7132-7141, 2018. 2, 4, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 681, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 555, + 712 + ], + "type": "text", + "content": "[22] King's College London. King's Computational Research, Engineering and Technology Environment (CREATE), 2024. 6" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2354" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 105 + ], + "type": "text", + "content": "[23] Yue Li, Hongxia Wang, and Mauro Barni. A survey of deep neural network watermarking techniques. Neurocomputing, 461:171-193, 2021. 1, 2, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 296, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 296, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 296, + 139 + ], + "type": "text", + "content": "[24] Yanghao Li, Saining Xie, Xinlei Chen, Piotr Dollar, Kaiming He, and Ross Girshick. Benchmarking detection transfer learning with vision transformers, 2021. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 296, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 296, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 296, + 194 + ], + "type": "text", + "content": "[25] Yuxuan Li, Sarthak Kumar Maharana, and Yunhui Guo. Not just change the labels, learn the features: Watermarking deep neural networks with multi-view data. In Computer Vision – ECCV 2024, pages 273–289, Cham, 2025. Springer Nature Switzerland. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 296, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 296, + 260 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 296, + 260 + ], + "type": "text", + "content": "[26] Zheng Li, Chengyu Hu, Yang Zhang, and Shanqing Guo. How to prove your model belongs to you: a blind-watermark based framework to protect intellectual property of DNN. In Proceedings of the 35th Annual Computer Security Applications Conference, page 126-137, New York, NY, USA, 2019. Association for Computing Machinery. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 296, + 316 + ], + "type": "text", + "content": "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In Computer Vision - ECCV 2014, pages 740-755, Cham, 2014. Springer International Publishing. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 296, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 296, + 360 + ], + "type": "text", + "content": "[28] Hanwen Liu, Zhenyu Weng, and Yuesheng Zhu. Watermarking deep neural networks with greedy residuals. In Proceedings of the 38th International Conference on Machine Learning, pages 6978-6988. PMLR, 2021. 1, 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 361, + 296, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 296, + 415 + ], + "type": "text", + "content": "[29] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C. Berg. SSD: Single shot multibox detector. In Computer Vision – ECCV 2016, pages 21–37, Cham, 2016. Springer International Publishing. 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 415, + 296, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 296, + 470 + ], + "type": "text", + "content": "[30] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 296, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 296, + 502 + ], + "type": "text", + "content": "[31] P. Massart. The tight constant in the dvoretzky-kiefer-wolfowitz inequality. The Annals of Probability, 18(3):1269-1283, 1990. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 503, + 296, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 296, + 536 + ], + "type": "text", + "content": "[32] E. Le Merrer, P. Pérez, and G. Trédan. Adversarial frontier stitching for remote neural network watermarking. Neural Computing and Applications, 32:9233-9244, 2019. 1, 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 537, + 296, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 296, + 591 + ], + "type": "text", + "content": "[33] Xin Mu, Yu Wang, Zhengan Huang, Junzuo Lai, Yehong Zhang, Hui Wang, and Yue Yu. Encryip: A practical encryption-based framework for model intellectual property protection. Proceedings of the AAAI Conference on Artificial Intelligence, 38(19):21438-21445, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 592, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 296, + 635 + ], + "type": "text", + "content": "[34] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015.8,4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 636, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 296, + 669 + ], + "type": "text", + "content": "[35] Yehonathan Refael, Adam Hakim, Lev Greenberg, Tal Aviv, Satya Lokam, Ben Fishman, and Shachar Seidman. SLIP: Securing LLMs IP using weights decomposition, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "text", + "content": "[36] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(6):1137-1149, 2017. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 596 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 139 + ], + "type": "text", + "content": "[37] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet large scale visual recognition challenge. International Journal of Computer Vision (IJCV), 115 (3):211-252, 2015. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 555, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 555, + 195 + ], + "type": "text", + "content": "[38] M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L. Chen. MobileNetV2: Inverted residuals and linear bottlenecks. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4510-4520, Los Alamitos, CA, USA, 2018. IEEE Computer Society. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 196, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 196, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 196, + 555, + 228 + ], + "type": "text", + "content": "[39] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition, 2015. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 229, + 555, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 229, + 555, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 229, + 555, + 284 + ], + "type": "text", + "content": "[40] Z. Sun, R. Sun, C. Liu, A. Chowdhury, L. Lu, and S. Jha. ShadowNet: A secure and efficient on-device model inference system for convolutional neural networks. In 2023 IEEE Symposium on Security and Privacy (SP), pages 1596-1612, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 285, + 555, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 555, + 340 + ], + "type": "text", + "content": "[41] Oliver Sutton, Qinghua Zhou, Wei Wang, Desmond Higham, Alexander N. Gorban, Alexander Bastounis, and Ivan Y Tyukin. Stealth edits to large language models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 341, + 555, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 341, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 341, + 555, + 396 + ], + "type": "text", + "content": "[42] Ruixiang Tang, Mengnan Du, and Xia Hu. Deep serial number: Computational watermark for dnn intellectual property protection. In Machine Learning and Knowledge Discovery in Databases: Applied Data Science and Demo Track, pages 157-173, Cham, 2023. Springer Nature Switzerland. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 396, + 555, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 396, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 396, + 555, + 441 + ], + "type": "text", + "content": "[43] Ivan Y. Tyukin, Desmond J. Higham, and Alexander N. Gorbán. On adversarial examples and stealth attacks in artificial intelligence systems. In 2020 International Joint Conference on Neural Networks (IJCNN), pages 1-6, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 441, + 555, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 441, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 316, + 441, + 555, + 485 + ], + "type": "text", + "content": "[44] Ivan Y Tyukin, Desmond J Higham, Alexander Bastounis, Eliyas Woldegeorgis, and Alexander N Gorban. The feasibility and inevitability of stealth attacks. IMA Journal of Applied Mathematics, 89(1):44-84, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 486, + 555, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 486, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 316, + 486, + 555, + 552 + ], + "type": "text", + "content": "[45] Yusuke Uchida, Yuki Nagai, Shigeyuki Sakazawa, and Shin'ichi Satoh. Embedding watermarks into deep neural networks. In Proceedings of the 2017 ACM on International Conference on Multimedia Retrieval, page 269-277, New York, NY, USA, 2017. Association for Computing Machinery. 1, 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 553, + 555, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 553, + 555, + 596 + ], + "spans": [ + { + "bbox": [ + 316, + 553, + 555, + 596 + ], + "type": "text", + "content": "[46] Jiangfeng Wang, Hanzhou Wu, Xinpeng Zhang, and Yuwei Yao. Watermarking in deep neural networks via error backpropagation. *Electronic Imaging*, 32(4):22-1-22-1, 2020. 1, 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2355" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Statistical Confidence Rescoring for Robust 3D Scene Graph Generation from Multi-View Images/00c53575-a21f-463b-8889-7a2370d2bff7_content_list.json b/2025/Statistical Confidence Rescoring for Robust 3D Scene Graph Generation from Multi-View Images/00c53575-a21f-463b-8889-7a2370d2bff7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..822f68149c8b6706966dd87b3d4f8ccfdc2782bd --- /dev/null +++ b/2025/Statistical Confidence Rescoring for Robust 3D Scene Graph Generation from Multi-View Images/00c53575-a21f-463b-8889-7a2370d2bff7_content_list.json @@ -0,0 +1,1550 @@ +[ + { + "type": "text", + "text": "Statistical Confidence Rescoring for Robust 3D Scene Graph Generation from Multi-View Images", + "text_level": 1, + "bbox": [ + 106, + 128, + 890, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qi Xun Yeo Yanyan Li Gim Hee Lee", + "bbox": [ + 331, + 203, + 666, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Department of Computer Science, National University of Singapore", + "bbox": [ + 230, + 220, + 764, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{qixunyeo,yan.li,gimhee.lee}@nus.edu.sg", + "bbox": [ + 323, + 241, + 668, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Modern 3D semantic scene graph estimation methods utilize ground truth 3D annotations to accurately predict target objects, predicates, and relationships. In the absence of given 3D ground truth representations, we explore leveraging only multi-view RGB images to tackle this task. To attain robust features for accurate scene graph estimation, we must overcome the noisy reconstructed pseudo point-based geometry from predicted depth maps and reduce the amount of background noise present in multi-view image features. The key is to enrich node and edge features with accurate semantic and spatial information and through neighboring relations. We obtain semantic masks to guide feature aggregation to filter background features and design a novel method to incorporate neighboring node information to aid robustness of our scene graph estimates. Furthermore, we leverage on explicit statistical priors calculated from the training summary statistics to refine node and edge predictions based on their one-hop neighborhood. Our experiments show that our method outperforms current methods purely using multi-view images as the initial input. Our project page is available at https://qixun1.github.io/projects/SCRSSG.", + "bbox": [ + 89, + 325, + 483, + 659 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 691, + 220, + 707 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The semantic scene graph (SSG) is a crucial intermediate representation that enhances higher-level scene understanding. It plays a key role in tasks such as image captioning [17, 34, 37], image retrieval [13, 22, 23, 38], image editing [3, 5, 43], and medical applications [7, 12, 15, 20, 39] by capturing both semantic and, more importantly, relational information between objects and their surroundings. Initially developed for 2D images, SSG has since expanded to the 3D domain [6, 9, 19]. 3D scene graphs provide a high-level representation of an entire 3D scene using inputs such as multi-view RGB images or LiDAR point clouds. Unlike their 2D counterparts, they incorporate spatial relation", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/10f9bb1c1d968dcecc34a9179ce6fe0f4eb8083b7c106fca028feadbbe5c4a55.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 292, + 903, + 395 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6cf51e49573ec2724c7423f0cc28af1d864f451b002fcf1e2f6d3975a3199906.jpg", + "image_caption": [ + "Figure 1. Existing multi-view RGB methods aggregate features using bounding boxes, introducing background noise that hinders accuracy. We use pretrained segmentation masks to reduce this interference and thus improving prediction accuracy." + ], + "image_footnote": [], + "bbox": [ + 517, + 400, + 903, + 500 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ships that extend beyond the visible image plane, allowing a holistic understanding of complex environments.", + "bbox": [ + 511, + 594, + 905, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Many approaches [1, 28, 29] rely on ground-truth 3D LiDAR point clouds as input. However, LiDAR sensors are resource-intensive and lack inherent semantic richness. Recent works [31, 32] have shifted toward using semantically rich multi-view RGB images, which offer a balance between computational efficiency and improved semantic fidelity. Although point cloud-based methods remain valuable for capturing geometric information, our approach focuses on leveraging multi-view RGB images. By incorporating pretrained depth estimators, we generate pseudo point-based geometry to estimate the 3D semantic scene graph, reducing reliance on LiDAR while preserving spatial and structural details.", + "bbox": [ + 509, + 626, + 906, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A key challenge in using only multi-view RGB images is ensuring that node features remain free from distractors. As shown in Fig. 1, prior approaches that rely on entity detectors to extract 2D bounding box proposals or regions of interest often fail to guarantee robustness since distrac", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24999", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tors within the bounding box are often mistakenly aggregated. Ensuring robust initial multi-view features is crucial without ground truth point-based geometry for error correction. No explicit mechanism exists to correct incorrectly bounded objects or misclassified rare classes that are not well trained on the model. To alleviate this problem, a mechanism should exist to refine estimates using prior knowledge instead of depending solely on implicit interactions between nodes and edges to achieve confident and accurate predictions.", + "bbox": [ + 88, + 90, + 480, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a framework to enhance the robustness of the model. First, we introduce a masked feature initialization (MFI), which leverages segmentation masks to aggregate image features instead of relying on the bounding box proposals. This reduces background noise, which results in cleaner multi-view image features. Next, we design a robust residual spatial neighbor graph neural network (RSN-GNN) to encode spatial information into node features. This network filters highly activated regions from neighboring nodes, refining target node features for improved predicate estimation. Finally, we propose a confidence rescoring (CR) module, which refines object and predicate estimates using an inverse softmax-weighted contribution of neighboring node-to-node and node-to-edge co-occurrence counts. By integrating this explicit inductive bias, our approach improves the accuracy of the prediction, particularly in low-confidence scenarios. Extensive experimental results on the benchmark 3RScan dataset show the competitiveness of our proposed approach compared to existing 2D and 3D approaches.", + "bbox": [ + 88, + 242, + 486, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions can be summarized as follows:", + "bbox": [ + 89, + 545, + 454, + 560 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a masked feature initialization to enhance the robustness of node features by reducing background distractors, yielding cleaner multi-view image features.", + "- We design a novel GCN architecture that integrates highly activated neighboring features into the target node to increase the robustness of the edge features.", + "- We propose a new refinement module to explicitly refine predictions based on statistical prior knowledge.", + "- Experiments show that we outperform previous state-of-the-art approaches on the 3RScan dataset, particularly on metrics that deal with low-tail imbalanced classes." + ], + "bbox": [ + 99, + 563, + 483, + 737 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 753, + 233, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2D Semantic Scene Graph. 2D semantic scene graph prediction is typically categorized into two-step and one-step approaches. The two-step approach first detects objects and then classifies their relationships [2, 4, 35, 40]. In contrast, the one-step approach jointly infers object and relationship classes [33]. Xu et al. [33] pioneers the problem of scene graph generation and they tried to solve this problem via iterative message passing. Baier et al. [2] first showed", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "how semantic models can be improved by incorporating triplet frequencies. Zellers et al. [40] analyzes the usefulness of statistical co-occurrences for the Visual Genome dataset and concluded that such statistical priors serves as strong regularization for the task. Chen et al. [4] formalizes the first approach (KERN) to incorporate the statistical prior directly into graph neural networks. Sharifzadeh et. al. proposes Schemata to assimilate image-based relational prior knowledge into the representations within the neural network [24]. Compared to previous methods, our method combines predictions from the prediction head with statistical co-occurrence of the current node and neighbors to determine the final class using a confidence score. Unlike Sonogashira et. al. [25] which uses hard thresholding, our approach uses adaptive relationship and node refinements.", + "bbox": [ + 511, + 90, + 906, + 320 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Semantic Scene Graph. Methods for generating a 3D semantic scene graph can be categorized into two types based on input data. The first relies on 3D inputs such as ground-truth 3D geometry to learn contextual information [28, 30, 41]. The second relies on multi-view RGB images with some methods incorporating depth information [31] while others do not [32]. Our approach follows the latter, leveraging the estimated depth from off-the-shelf depth estimators [36] to enhance the information provided by multi-view RGB images. Wald et al. [28] first proposed 3RScan: a richly annotated 3D scene graph dataset and a method that focuses on pairwise relationships to predict the 3D scene graphs. Wu et al. [31] uses SLAM to obtain dense point representations and a novel graph convolutional network (GCN) based aggregation function to enhance 3D scene graph prediction. Zhang et al. [41] proposes an edge-oriented GCN to incorporate multi-dimensional edge features for explicit inter-node relationship modeling. Zhang et al. [42] integrates prior commonsense knowledge by learning meta-embeddings only from the class labels with their graphical structures to avoid perceptual errors. Wang et al. [30] incorporated additional information by distilling knowledge learned by the multi-modal oracle model to the 3D model. Wu et al. [32] introduces a novel entity association method to obtain the 3D entities from the predicted 2D entities and a geometric gate to incorporate geometric information to multi-view image features. Feng et al. [8] uses ConceptNet external knowledge base to accumulate both contextualized visual content and textual facts to form a 3D spatial multimodal knowledge graph. Our method leverages on the explicit statistical priors similar to 2D scene graph method KERN and Schemata while guiding low confidence predictions made by the initial node and edge predictors using both visual and geometric cues.", + "bbox": [ + 511, + 386, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25000", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b8012a394b2241f11960424a8067920d6cd7ff88f68ec8065dd998a40112353f.jpg", + "image_caption": [ + "Figure 2. The architecture of our proposed framework consists of three main components: 1) A masked feature initialization step leverages backprojected semantic masks to aggregate multi-view image, geometric, and spatial features. 2) A message passing step utilizes our residual spatial neighbor GNN module to refine node and edge features through feature interaction. 3) A confidence rescoring step adjusts logits based on summary statistics of object and predicate co-occurrence from the training dataset." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 903, + 244 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Our Methodology", + "text_level": 1, + "bbox": [ + 89, + 335, + 264, + 353 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Definition. Given either a set of 3D point clouds $P$ or multi-view RGB images $I$ of a 3D scene as input, the goal is to estimate the 3D scene graph denoted as:", + "bbox": [ + 89, + 361, + 482, + 407 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nG = (O, R), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 241, + 420, + 482, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where", + "text_level": 1, + "bbox": [ + 89, + 449, + 137, + 462 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- $N$ : Number of multi-view input images;", + "- $O = \\{o_m\\}_{m=1}^M$ : All objects in the scene;", + "- $R = \\{r_{k\\to m}\\}_{m,k = 1}^{M,M}$ : Relationships between all objects;", + "- $M = |O|$ : Number of objects in the scene." + ], + "bbox": [ + 99, + 465, + 480, + 525 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Furthermore, we refer to $o_m$ as the object class for node instance $m$ , $r_{k \\to m}$ as the predicate class for the edge connecting nodes $k$ and $m$ . Alternatively, the 3D scene graph can be defined as a set of triplets given by: $\\{o_k, r_{k \\to m}, o_m\\} := \\{node, predicate, node\\}$ .", + "bbox": [ + 89, + 529, + 483, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. Fig. 2 shows an illustration of the architecture of our proposed framework. The input to our framework is a set of $N$ multi-view RGB images $\\{I_i\\}_{i=1}^N$ of the 3D scene. Since the focus of our work is not on multi-view 3D reconstruction, we recover the 3D point cloud of the scene by running an off-the-shelf RGB-D SLAM system [26] with the metric depth maps of the input images $\\{\\hat{D}_i\\}_{i=1}^N$ obtained from the pretrained Depth Anything [36] model.", + "bbox": [ + 89, + 613, + 483, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "From the predicted 3D point cloud, we initialize multiview image features using segmentation masks, extract geometric features with a point-based feature extractor, and derive spatial features via a linear layer. These multi-view image features serve as the initial input to our RSN-GNN module. At the GCN initialization, we enhance node features by integrating geometric and spatial features through geometric and spatial gates. We additionally enhance edge features by integrating max pooled neighboring nodes features. These enhanced node and edge features undergo message passing, with the integration process repeating at each", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "layer. After $k$ iterations, the final node and edge features are processed by predictors to generate initial estimates, which are then refined using precomputed co-occurrence statistics via our CR module, producing the final scene graph output.", + "bbox": [ + 511, + 337, + 903, + 397 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We elaborate on the feature initialization step in Sec. 3.1, the message passing step in Sec. 3.2, and the confidence rescoring step in Sec. 3.3.", + "bbox": [ + 511, + 398, + 903, + 444 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Masked Feature Initialization (MFI)", + "text_level": 1, + "bbox": [ + 511, + 454, + 828, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We initialize the node features of the scene graph by combining multi-view image representations from a pretrained image encoder with PointNet features derived from the predicted points in the RGB-D SLAM reconstruction step that utilizes predicted depth. To eliminate background clutter, we employ the Segment Anything Model (SAM) [14] to produce precise 2D semantic masks for each object. For each node, we backproject the associated points onto the 2D instance mask to identify its corresponding region. We then generate object proposals from these instance masks, resize them to $224 \\times 224$ , and extract their multi-view image features using the pretrained encoder. Finally, we apply these semantic masks to the multi-view image features before aggregating them into our node representations.", + "bbox": [ + 511, + 477, + 905, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method to get the object proposals, i.e. the nodes of the scene graph offers two key advantages over relying solely on the entity detector from SceneGraphFusion and JointSSG [31, 32]. First, the entity detector often captures only part of an object. Although SceneGraphFusion propagates the label for this segment, the feature extractor can operate only on the detected portion. In contrast, our method employs more complete object masks, enabling the feature extractor to derive features from the entire object. Second, we exclude extraneous background features by applying these masks before feature aggregation. In SceneGraphFusion and JointSSG, the bounding boxes of the pretrained entity detector frequently include background regions, which can dilute the quality of the features and hinder", + "bbox": [ + 511, + 689, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25001", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the performance. Our method is closer to the open vocabulary method ConceptGraphs [10] which uses SAM masks but with the aggregation method provided by JointSSG.", + "bbox": [ + 89, + 90, + 482, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formally, we compute a corresponding multi-view image feature $\\mathbf{v}_{i,img}$ , a geometric feature $\\mathbf{v}_{i,geo}$ , and a spatial feature $\\mathbf{v}_{i,spat}$ for each node $v_{i}$ . We aggregate features only for nodes that are visible in the可视ability graph $G_{v}$ denoted by $\\mathbf{1}(v_{i} \\in I_{k})$ and have corresponding pixels within the semantic mask provided by SAM. The aggregated multi-view image feature $\\mathbf{v}_{i,img}$ is given by:", + "bbox": [ + 89, + 136, + 483, + 243 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {i} ^ {0} = \\frac {\\sum_ {k = 1} ^ {N} \\mathbf {1} \\left(v _ {i} \\in I _ {k}\\right) f \\left(\\eta_ {k} \\odot x _ {i , k}\\right)}{\\left| k \\in G _ {v} \\right|}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 253, + 482, + 289 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which also serves as the initial node feature. $f(\\cdot)$ is the image feature encoder and $\\eta_{k}$ refers to the 2D semantic mask. $x_{i,k}$ refers to the cropped image patch containing the node $i$ from the $k$ -th image. $\\odot$ refers to the element-wise multiplication operator. The aggregation is the arithmetic mean of the node features for each image in the可视ability graph.", + "bbox": [ + 89, + 299, + 483, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The geometric feature $\\mathbf{v}_{i,geo}$ is extracted using a vanilla PointNet [21], with inputs derived from the 3D instance masks generated during the RGB-D SLAM step. For the spatial features, we employ a simple linear layer to transform the attributes of each 3D bounding box: dimensions $\\mathbf{b}_i$ , volume $s_i$ , length $l_i$ , and the ratio of its x- and y-axes $r_i$ into the spatial feature as:", + "bbox": [ + 89, + 388, + 483, + 494 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {i, s p a t} = g _ {s} ([ \\mathbf {b} _ {i}, s _ {i}, l _ {i}, r _ {i} ]), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 506, + 482, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $g_{s}$ refers to a set of learnable weights for spatial features. $[\\cdot]$ refers to the concatenation of the inputs within the operator.", + "bbox": [ + 89, + 534, + 483, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Residual Spatial Neighbor GNN Module", + "text_level": 1, + "bbox": [ + 89, + 588, + 436, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following a similar design to the learnable geometric gate in JointSSG [32], we integrate both geometric and spatial features into the initial node feature, which originally consists only of multi-view image features. The spatial features serve as additional information for individual nodes, analogous to the embeddings used for token encoding in Transformer architectures [27]. We define the geometric gate as:", + "bbox": [ + 89, + 609, + 483, + 717 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {v}} _ {i} ^ {l} = \\mathbf {v} _ {i} ^ {l} + \\sigma \\left(\\mathbf {w} _ {g} ^ {T} \\left[ \\mathbf {v} _ {i} ^ {l}, \\mathbf {v} _ {i, g e o} \\right]\\right) \\cdot \\sigma \\left(\\mathbf {v} _ {i, g e o}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 724, + 482, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{w}_g$ refers to a set of learnable weights, $\\sigma$ refers to the sigmoid operator and $l$ refers to the layer of the node feature. Furthermore, we define the spatial gate as:", + "bbox": [ + 89, + 755, + 483, + 800 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\ddot {\\mathbf {v}} _ {i} ^ {l} = \\hat {\\mathbf {v}} _ {i} ^ {l} + \\sigma \\left(\\mathbf {w} _ {s} ^ {T} [ \\hat {\\mathbf {v}} _ {i} ^ {l}, \\mathbf {v} _ {i, s p a t} ]\\right) \\cdot \\sigma \\left(\\mathbf {v} _ {i, s p a t}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 809, + 482, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{w}_s$ is a set of learnable weights for spatial features.", + "bbox": [ + 89, + 839, + 473, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by the residual connections in ResNet [11], we introduce a mechanism to integrate highly activated neighboring features into the target node feature using max pooling during the message-passing process for edge features. This approach implicitly encodes neighboring information, improving edge prediction for scene graph estimation. Specifically, we augment the edge messages in the GCN by integrating max-pooled neighboring features before message passing as follows:", + "bbox": [ + 511, + 90, + 905, + 227 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {v}} _ {i} ^ {l} = \\ddot {\\mathbf {v}} _ {i} ^ {l} + \\max ([ \\ddot {\\mathbf {v}} _ {j} ^ {l} ] _ {j \\in n (i)}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 232, + 903, + 252 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\max (\\cdot)$ refers to the max pooling operator and $n(i)$ refers to the neighboring nodes of node $i$ . Max pooling allows highly activated regions in the different neighboring node features to be fused into the target node feature via the residual connection.", + "bbox": [ + 511, + 257, + 905, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We apply max-pooled neighboring features to the calculation of edge messages without using them in node message updates. Since node features are highly susceptible to noise and spurious correlations from other nodes [16], incorporating max-pooled neighboring features into node message computations can introduce confusion instead of improving performance. The edge message $m_{i\\rightarrow j}$ is thus defined as:", + "bbox": [ + 511, + 333, + 905, + 450 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nm _ {i \\rightarrow j} = g _ {e} ([ \\tilde {\\mathbf {v}} _ {i} ^ {l}, \\mathbf {e} _ {i j} ^ {l}, \\tilde {\\mathbf {v}} _ {j} ^ {l} ]), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 452, + 903, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $g_{e}(\\cdot)$ is an MLP and $\\mathbf{e}_{ij}^{l}$ is the edge descriptor used in [32] containing relative node properties such as centroid displacement, relative bounding box size difference, and a relative pose descriptor.", + "bbox": [ + 511, + 474, + 905, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Confidence Rescoring (CR) Module", + "text_level": 1, + "bbox": [ + 511, + 542, + 823, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As discussed in Sec. 1, modern scene graph methods depend heavily on the entity detector and image feature extractor to produce an accurate initial object prediction. However, there is no mechanism to correct a target bounding box prediction that is initially wrong. To address this limitation, we leverage relationships with neighboring nodes, allowing us to mitigate the impact of incorrect initial predictions by weighting contributions from neighboring node statistics to refine target object classification. Fig. 3 illustrates the logits of the final prediction after applying our confidence rescoring module. This module incorporates prior knowledge to improve prediction confidence beyond that of the original node and edge multi-layer perceptron (MLP) predictor.", + "bbox": [ + 511, + 565, + 905, + 775 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The probability of predicting node $i$ given its neighboring nodes $n(i)$ is approximated by the output of our node predictor $h_v(\\cdot)$ as:", + "bbox": [ + 511, + 776, + 905, + 821 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} P \\left(o _ {i} \\mid \\left\\{o _ {j}: j \\in n (i) \\right\\}\\right) = \\frac {1}{Z _ {i}} \\psi_ {i} \\left(o _ {i}\\right) \\prod_ {j \\in n (i)} \\psi_ {i j} \\left(o _ {i}, o _ {j}\\right) \\\\ \\approx \\tau \\left(h _ {v} \\left(\\mathbf {v} ^ {L}\\right)\\right), \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 827, + 903, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1The covisibility graph is obtained as intermediate output after running the RGB-D SLAM system. It refers to a bipartite graph where each image $I_{k}$ is connected with the specific node instance $v_{i}$ .", + "bbox": [ + 89, + 862, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25002", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\psi_i(o_i)$ refers to the potential associated with node $i$ and $\\psi_{ij}(o_i,o_j)$ refers to the potential associated with the edge between node $i$ and its neighbor node $j$ . The potentials are implicitly learned through the GCN unlike classical methods, which explicitly learn these potentials. $h_v(\\cdot)$ is a small MLP that predicts the class of each node and $\\tau$ refers to the softmax operator. $L$ refers to the last layer of the GCN message passing output. The partition function given by:", + "bbox": [ + 89, + 90, + 483, + 227 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {i} = \\sum_ {o _ {i}} \\psi_ {i} \\left(o _ {i}\\right) \\prod_ {j \\in n (i)} \\psi_ {i j} \\left(o _ {i}, o _ {j}\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 234, + 482, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ensures the probabilities sum to 1. A similar equation can be obtained for the probability of predicting edge $ij$ given its neighboring nodes $i$ and $j$ , i.e.:", + "bbox": [ + 89, + 277, + 483, + 325 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} P \\left(r _ {i \\rightarrow j} \\mid o _ {i}, o _ {j}\\right) = \\frac {1}{Z _ {i j}} \\psi_ {i j} \\left(o _ {i}, o _ {j}, r _ {i \\rightarrow j}\\right) \\tag {10} \\\\ \\approx \\tau (h _ {e} (\\mathbf {e} ^ {L})), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 330, + 480, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $h_e(\\cdot)$ is a small MLP which predicts the class of each edge, and the partition function to normalize the probabilities is given by:", + "bbox": [ + 89, + 392, + 483, + 438 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {i j} = \\sum_ {r _ {i \\rightarrow j}} \\psi_ {i j} \\left(o _ {i}, o _ {j}, r _ {i \\rightarrow j}\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 446, + 482, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We begin by computing summary statistics from the training dataset to node-to-node co-occurrences denoted as $c_{ij}(o_i, o_j)$ , and node-to-edge co-occurrences denoted as $d_{ij}(o_i, e_j)$ . Subsequently, we determine the prediction confidence of the model, where $\\alpha_{\\mathbf{v}}$ represents node confidence and $\\alpha_{\\mathbf{e}}$ represents edge confidence. Next, we obtain the node $\\alpha_{\\mathbf{v}}$ and edge $\\alpha_{\\mathbf{e}}$ confidences as:", + "bbox": [ + 89, + 488, + 483, + 594 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {\\mathbf {v}} = \\max _ {\\mathbf {v}} \\tau \\left(h _ {v} \\left(\\mathbf {v} ^ {L}\\right)\\right), \\quad \\alpha_ {\\mathbf {e}} = \\max _ {\\mathbf {e}} \\tau \\left(h _ {e} \\left(\\mathbf {e} ^ {L}\\right)\\right). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 602, + 482, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first estimate the marginal probability of a specific instantiation of node $i$ , conditioned on node $j$ using node-to-node co-occurrences, and on edge $ij$ using node-to-edge co-occurrences defined as:", + "bbox": [ + 89, + 632, + 483, + 691 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} \\left(o _ {i} \\mid o _ {j}\\right) = \\left[ c \\left(o _ {i} = c _ {k} \\mid o _ {j} = c _ {l}\\right) \\right] _ {k = 0, l = 0} ^ {C, C}, \\quad \\text {w h e r e}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 699, + 434, + 719 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nc \\left(o _ {i} = c _ {k} \\mid o _ {j} = c _ {l}\\right) = \\frac {c \\left(o _ {i} = c _ {k} , o _ {j} = c _ {l}\\right)}{\\sum_ {k = 0} ^ {C} c \\left(o _ {i} = c _ {k} , o _ {j} = c _ {l}\\right)}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 720, + 482, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $C$ refers to the number of object classes and $c_{k}$ refers to the $k$ -th class in the set of object classes. We then integrate the summary statistics via an inverse softmax operation denoted as $\\gamma(\\cdot)$ into the node probability in Eq. 14 to get:", + "bbox": [ + 89, + 763, + 483, + 840 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} P \\left(o _ {i} \\mid \\left\\{o _ {j}: j \\in n (i) \\right\\}\\right) = \\tau \\left(\\alpha_ {\\mathbf {v}} \\cdot h _ {v} (\\mathbf {v} _ {i}) \\right. \\\\ + \\left(1 - \\alpha_ {\\mathbf {v}}\\right) \\cdot \\sum_ {j \\in n (i)} \\alpha_ {j} \\cdot \\gamma (\\mathbf {c} \\left(o _ {i} \\mid o _ {j}\\right)). \\tag {14} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 848, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This formulation is chosen because directly adding the estimated marginal probability to the predictions of the model after softmax would disregard the required normalization, resulting in incompatible terms.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarly, we compute the estimated marginal probability of an edge $ij$ instantiation conditioned on nodes $i$ and $j$ using node-to-edge co-occurrence statistics for predicate estimation:", + "bbox": [ + 511, + 151, + 905, + 210 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {d} \\left(r _ {i \\rightarrow j} \\mid o _ {j}\\right) = \\left[ d \\left(r _ {i \\rightarrow j} = e _ {k} \\mid o _ {j} = c _ {l}\\right)\\right] _ {k = 0, l = 0} ^ {P, C},\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 220, + 892, + 241 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd \\left(r _ {i \\rightarrow j = k} \\mid o _ {j = l}\\right) = \\frac {d \\left(r _ {i \\rightarrow j} = e _ {k} , o _ {j} = c _ {l}\\right)}{\\sum_ {k = 0} ^ {P} d \\left(r _ {i \\rightarrow j} = e _ {k} , o _ {j} = c _ {l}\\right)}. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 243, + 903, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $P$ refers to the total number of predicate classes and $e_k$ refers to the $k$ -th class in the set of predicate class. The refined logits for predicate estimation are then given by:", + "bbox": [ + 511, + 306, + 905, + 353 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} P \\left(r _ {i \\rightarrow j} \\mid o _ {i}, o _ {j}\\right) = \\tau \\left(\\alpha_ {i j} \\cdot h _ {e} (\\mathbf {e} _ {i j}) \\right. \\\\ + \\left(1 - \\alpha_ {i j}\\right) \\cdot \\left(\\alpha_ {i} \\cdot \\gamma (\\mathbf {d} \\left(r _ {i \\rightarrow j} \\mid o _ {i}\\right))\\right) \\tag {16} \\\\ \\cdot \\left(\\alpha_ {j} \\cdot \\gamma (\\mathbf {d} (r _ {i \\rightarrow j} | o _ {j})))\\right). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 364, + 903, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remarks. As shown in Fig. 3, the original node probability for the correct class is already relatively high. However, our CR module further enhances the predicted probability as evidenced by the upward shift in the box plot from \"Original\" to \"After CR\" in the red boxes, with a notable increase in the lower quartile. Moreover, a larger proportion of low node probability instances are flagged as outliers as shown by the greater number of red circles at \"After CR\" compared to \"Original\". Most notable gains occur when original logits are low since CR increases the probability of the correct class by down-weighting misclassified classes $c'$ via lower conditional probabilities $P(c'|c)$ given conditioned class $c$ . This targeted adjustment raises suppressed logits for rare classes, thus improving prediction performance for underrepresented categories.", + "bbox": [ + 511, + 429, + 906, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 670, + 645, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 695, + 702, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate on the 3RScan dataset following the same training and test split from [28]. The image quality of the dataset is rather poor as each scene constitutes a variable number of short videos spliced together with different motion trajectories. As the images are obtained from consumer-level cameras, the images possess a low frame rate of $10\\mathrm{Hz}$ , image blur, and image jitter from sudden camera motion which makes this dataset challenging.", + "bbox": [ + 511, + 717, + 905, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We utilize the codebase provided from the 3DSSG GitHub repository $^2$ to reproduce the prior", + "bbox": [ + 511, + 845, + 905, + 876 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://github.com/ShunChengWu/3DSSG", + "bbox": [ + 529, + 886, + 763, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25003", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9750957ed9fb272c754841d3f8a1602addff684b3c4eca4deb3627bd5b0ef2b8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methodbath.bedbkshfcab.chaircntr.curt.deskdoorfloorofurnpic.refri.show.sinksofatabletoil.wallwind.Mean.
IMP0.0001.0000.0000.3410.4670.0000.6060.2000.5280.9000.1250.1430.0000.0000.3330.4500.5250.8000.6220.1250.358
VGfM0.5000.0000.0000.4150.5430.0830.5450.0000.3890.9200.2320.0860.0000.0000.3810.4500.5740.8000.6440.1670.336
3DSSG0.0000.6670.0000.2070.3480.1670.5760.2000.3610.7800.1250.0570.0000.0000.3330.5000.2790.0000.4670.1250.260
SGFN0.5000.3330.0000.2930.6850.2500.5150.2000.4440.8800.2320.1710.0000.0000.4290.5000.5080.4000.7070.1250.359
JointSSG1.0001.0000.0000.4760.6740.2500.5760.2000.5000.9000.1070.1430.0000.0000.4290.5500.5410.8000.6800.2500.454
Ours1.0001.0000.3330.6220.6300.3330.5450.6000.5280.8800.3040.1710.6670.6670.3330.6000.6230.8000.6930.3750.605
", + "bbox": [ + 94, + 89, + 908, + 159 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison with state-of-the-art methods on the 3RScan dataset for each object class. The Best results are highlighted.", + "bbox": [ + 114, + 171, + 880, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b5f5e82be139d5c212c4fc229a7d5c51692abb22563db3642c14ebd2d4ee705b.jpg", + "image_caption": [ + "Figure 3. Box plot for predicted node probability for correct class after our CR module. The lower quartile visibly moves upwards due to an increase in predicted node probability for correct class." + ], + "image_footnote": [], + "bbox": [ + 122, + 214, + 428, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "methods. For JointSSG[32], we replace their image feature extractor from a pretrained ResNet18 to a pretrained DINOv2 with a ViT-L backbone to ensure fair comparison.", + "bbox": [ + 89, + 467, + 482, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following [32], we use PointNet [21] as the point encoder and employ two message-passing layers. For RGB-D SLAM reconstruction, we utilize the method from [26] and our predicted depth to generate predicted points. Multiview feature extraction is performed using a pretrained DINOv2 [18] with a ViT-L backbone. To obtain object segmentation masks, we use a pretrained SAM predictor with a text prompt containing the class set of 20 NYU object classes following the subset defined in [28]. Node-to-node and node-to-edge co-occurrences are precomputed on the training set. We train the model on a single RTX 3090 GPU for two days with early stopping.", + "bbox": [ + 88, + 513, + 482, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following [29] and JointSSG[32], we report the overall top-1 recall (Recall) for the object class estimation (Obj.), the predicate estimation (Pred.), and the relationship triplet estimation (Rel.). We also report the mean recall (mRecall) for the object class estimation (Obj.), the predicate estimation (Pred.) only. We map all predictions on the estimated segments to the ground truth segments for a fair comparison between the different methods.", + "bbox": [ + 89, + 700, + 482, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative Results", + "text_level": 1, + "bbox": [ + 89, + 832, + 284, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present the per-class recall for 20 object classes in the 3RScan dataset in Tab. 1, demonstrating that our method significantly outperforms others across most object cate", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "gories. Notably, we are the only approach that achieves a recall above $0\\%$ for all object classes, indicating that we robustly learn all classes instead of just a subset. Our method tends to underperform slightly compared to the other methods on majority classes such as floor and wall classes in return, although not by a significant margin. The main experimental results for the top-1 recall and the mean recall are shown in Tab. 2. Importantly, our approach significantly outperforms others in mean recall (mRecall) for both objects and predicates, suggesting that it handles class imbalance issues [29] more effectively than competing methods. Additionally, our method surpasses all baselines in relationship and object estimation metrics while achieving comparable performance in predicate estimation. More impressively, our method with the weaker ResNet18 multiview image features already surpasses all baselines in relationship and object estimation metrics except for predicate estimation. Our method with ResNet18 image features even manages to surpass the performance of JointSSG with DINOv2 multi-view image features by non-trivial margins. This shows that our improved performance is not purely due to the more powerful DINOv2 features. Refer to our supplementary material for more quantitative results.", + "bbox": [ + 511, + 204, + 906, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Qualitative Results", + "text_level": 1, + "bbox": [ + 511, + 561, + 694, + 577 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We show qualitative results comparing our method with other competitive methods in Fig. 4. Since the scene contains a large number of objects, we display only a subset with fewer objects and predicates for better readability. The correct prediction of the None predicate class is omitted in the visualization due to its prevalence in the dataset. As seen in Fig. 4, our method can distinguish the relatively more ambiguous wall which the other methods fail to distinguish. The shower curtain is tricky to classify since it is a subset of a curtain that is another class of objects present in the dataset. Understandably, all methods fail to correctly classify this object. Both our method and JointSSG attain good performance on the predicate class compared to SGFN as both methods only misclassify the predicate between the curtain object and the wall.", + "bbox": [ + 511, + 583, + 906, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We additionally evaluate on the entire scene for scan 43b8cae1 without comparison to other methods. Our method can recognize rare classes present in the scene, which are more challenging to learn. As seen in Fig. 5, our approach successfully classifies the rare class bookshelf instance denoted by $bkshf$ which no other method managed", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25004", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5936e487f036635e02696b434b01b76f7101e3528d0de93e1a3f253bae481cbf.jpg", + "image_caption": [ + "Figure 4. The qualitative results on 3RScan dataset of previous works and our proposed framework on scan 4d3d82b0. SGFN also fails to classify majority of non-background object classes except the sink and cabinet. It also fails to classify all predicate classes. JointSSG performs better but misclassifies the shower curtain as a curtain and the ambiguous blue wall as a cabinet. Our method correctly classifies the wall but misidentifies the shower curtain as a curtain." + ], + "image_footnote": [], + "bbox": [ + 135, + 92, + 861, + 417 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0fc87d00a72f2183bee837af9985e58af734d76da85d780d9600e08a3b7dd392.jpg", + "image_caption": [ + "Figure 5. The qualitative results on 3RScan dataset of our proposed framework on the entire scene of scan 43b8cae1. Our method can correctly classify most objects in the scene except for the light blue dress table at the top right of the scan. Our method can also correctly classify most predicates." + ], + "image_footnote": [], + "bbox": [ + 151, + 486, + 846, + 816 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25005", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ed1bc5ddf4c41762169967b9dc63e5c5df5bd3df3ced45623864b2a54dc05f58.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2. Comparison with state-of-the-art methods on the 3RScan dataset with 20 object classes and 8 predicate classes. The top group of results are reported in JointSSG [32]. The middle group of results are reproduced via 3DSSG GitHub repository. JointSSG‡ refers to JointSSG but with DINOv2 multi-view image features. Ours† uses ResNet18 multi-view image features; Ours uses the DINOv2 multi-view image features instead. The Best and Second Best results are highlighted, respectively." + ], + "table_body": "
MethodRecall%mRecall%
RelObj.Pred.Obj.Pred
IMP [32]25.851.890.430.023.0
VGfM [32]28.353.390.731.624.4
3DSSG [32]17.541.488.231.926.6
SGFN [32]31.456.789.638.330.5
JointSSG [32]34.158.189.943.033.3
IMP25.447.789.935.819.9
VGfM27.450.190.633.622.3
3DSSG13.035.686.826.024.3
SGFN29.952.289.535.927.1
JointSSG32.954.588.745.435.2
JointSSG‡34.356.686.552.936.2
Ours†38.659.389.157.436.8
Ours40.561.890.460.539.2
", + "bbox": [ + 98, + 89, + 477, + 325 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to predict based on the results shown in Tab. 1. Our method also predicts multiple reasonable predicates of standing on between various objects and floor. Additional qualitative results can be found in the supplementary material.", + "bbox": [ + 89, + 470, + 482, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 544, + 254, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We analyze the effects of our feature initialization, robust neighbor residuals, and confidence rescoring module in Tab. 3. Our confidence rescoring (CR) module significantly improves performance across all metrics, indicating that it effectively leverages statistical priors to address class imbalance issues as noted in [32]. Apart from predicate mRecall, incorporating our masked feature initialization further improves performance across all metrics. This outcome is expected since masking primarily aims to refine multi-view image features, which do not directly contribute to edge feature improvements. The combination of the masked feature initialization and the CR module act as a force multiplier since the CR module benefits from cleaner and more accurate neighboring node features. Additionally, our robust neighbor residual gate maintains comparable performance across most metrics except for predicate mRecall, which benefits from additional neighbor information to better handle class imbalance in predicate estimation.", + "bbox": [ + 89, + 566, + 482, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We additionally study the effects of using our CR module during inference in Tab. 4. The addition of the CR module with the prior knowledge of the dataset is akin to calculating statistics during transductive setting of inference instead", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8c033161e40a4f9cac6af0b2a6911fb55407a69b2a8dc8dc5d2624ceb18880f3.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Ablation study on the 3RScan dataset. Conf. refers to the usage of summary statistic priors for confidence rescoring via the CR module. Feat. refers to the usage of DINOv2 image features with refinement from SAM semantic masks via MFI. Neigh. refers to the usage of RSN-GNN. The Best results are shown in bold." + ], + "table_body": "
MethodConf.Feat.Neigh.Recall%mRecall%
Rel.Obj.Pred.Obj.Pred.
JointSSG×××32.954.588.245.435.2
+ Conf.✓××37.957.289.654.136.3
+ Feat.✓✓×41.161.390.861.335.8
Ours✓✓✓40.561.890.460.539.2
", + "bbox": [ + 517, + 89, + 911, + 167 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e7ad321f39aa38eecd67d717360a08ded85bd33a8645109259c7a242cbf64b3c.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4. Ablation study on the 3RScan dataset. CR refers to the usage of summary statistic priors for confidence rescoring via the CR module during inference. The Best results are shown in bold." + ], + "table_body": "
MethodCR.Recall%mRecall%
Rel.Obj.Pred.Obj.Pred.
Ours×41.161.190.660.235.7
Ours✓40.561.890.460.539.2
", + "bbox": [ + 566, + 266, + 857, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of a true overfitting to the dataset. Interestingly, the removal of the CR module during inference does not significantly reduce performance. It even leads to a slight increase in predicate Recall and overall relationship triplet Recall. The slight increase in predicate Recall, coupled with a drastic decrease in predicate mRecall, suggests that removing the CR module benefits the predicate classes that appear frequently in the dataset, including the None class, with the trade off of decreasing performance on the tail classes. Nonetheless, the use of the CR module during evaluation benefits object Recall, object mRecall, and predicate mRecall.", + "bbox": [ + 511, + 405, + 906, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. The effectiveness of initial feature aggregation relies on the quality of the segmentation mask. Although the CR module performs well when there is no significant domain shift, the statistical prior can negatively impact performance when applied to datasets with a drastically different training data distribution. Similarly, the prior becomes less useful on datasets with very limited data.", + "bbox": [ + 511, + 580, + 905, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 707, + 633, + 722 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce a straightforward yet effective approach for enhancing 3D semantic scene graph estimation through our MFI, RSN-GNN, and CR modules. Specifically, our masked feature initialization (MFI) enhances node features by filtering out background distractors and leveraging more than just part-based image features. The RSN-GNN module strengthens edge features by integrating highly activated neighboring features into the target node feature, improving robustness. Finally, we introduce a refinement module that explicitly refines predictions using statistical prior knowledge, further enhancing performance.", + "bbox": [ + 511, + 734, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25006", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research work is supported by the Agency for Science, Technology and Research (A\\*STAR) under its MTC Programmatic Funds (Grant No. M23L7b0021), and the Tier 2 grant MOE-T2EP20124-0015 from the Singapore Ministry of Education. We also thank Shun-Cheng Wu for the helpful discussions on running the SLAM code and reproducing the results in the 3DSSG repository.", + "bbox": [ + 89, + 113, + 483, + 227 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 239, + 187, + 255 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Iro Armeni, Zhi-Yang He, JunYoung Gwak, Amir R Zamir, Martin Fischer, Jitendra Malik, and Silvio Savarese. 3d scene graph: A structure for unified semantics, 3d space, and camera. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5664-5673, 2019. 1", + "[2] Stephan Baier, Yunpu Ma, and Volker Tresp. Improving visual relationship detection using semantic modeling of scene descriptions. In The Semantic Web-ISWC 2017: 16th International Semantic Web Conference, Vienna, Austria, October 21-25, 2017, Proceedings, Part I 16, pages 53-68. Springer, 2017. 2", + "[3] Lichang Chen, Guosheng Lin, Shijie Wang, and Qingyao Wu. Graph edit distance reward: Learning to edit scene graph. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX 16, pages 539-554. Springer, 2020. 1", + "[4] Tianshui Chen, Weihao Yu, Riquan Chen, and Liang Lin. Knowledge-embedded routing network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6163–6171, 2019. 2", + "[5] Helisa Dhamo, Azade Farshad, Iro Laina, Nassir Navab, Gregory D Hager, Federico Tombari, and Christian Rupprecht. Semantic image manipulation using scene graphs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5213-5222, 2020. 1", + "[6] Helisa Dhamo, Fabian Manhardt, Nassir Navab, and Federico Tombari. Graph-to-3d: End-to-end generation and manipulation of 3d scenes using scene graphs. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16352–16361, 2021. 1", + "[7] Jessica D'souza, PK Aleema, S Dhanyashree, Clita Fernandes, KM Kavitha, and Chandra Naik. Knowledge-based scene graph generation in medical field. In 2023 IEEE International Conference on Distributed Computing, VLSI, Electrical Circuits and Robotics (DISCOVER), pages 232-237. IEEE, 2023. 1", + "[8] Mingtao Feng, Haoran Hou, Liang Zhang, Zijie Wu, Yulan Guo, and Ajmal Mian. 3d spatial multimodal knowledge accumulation for scene graph prediction in point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9182-9191, 2023. 2", + "[9] Gege Gao, Weiyang Liu, Anpei Chen, Andreas Geiger, and Bernhard Scholkopf. Graphdreamer: Compositional 3d scene synthesis from scene graphs. In Proceedings of" + ], + "bbox": [ + 99, + 265, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21295-21304, 2024. 1", + "[10] Qiao Gu, Ali Kuwajerwala, Sacha Morin, Krishna Murthy Jatavallabhula, Bipasha Sen, Aditya Agarwal, Corban Rivera, William Paul, Kirsty Ellis, Rama Chellappa, et al. Conceptgraphs: Open-vocabulary 3d scene graphs for perception and planning. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 5021-5028. IEEE, 2024. 4", + "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4", + "[12] Felix Holm, Ghazal Ghazaei, Tobias Czempiel, Ege Özsoy, Stefan Saur, and Nassir Navab. Dynamic scene graph representation for surgical video. In Proceedings of the IEEE/CVF international conference on computer vision, pages 81-87, 2023. 1", + "[13] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3668–3678, 2015. 1", + "[14] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 3", + "[15] Chen Lin, Shuai Zheng, Zhizhe Liu, Youru Li, Zhenfeng Zhu, and Yao Zhao. Sgt: Scene graph-guided transformer for surgical report generation. In International conference on medical image computing and computer-assisted intervention, pages 507-518. Springer, 2022. 1", + "[16] Xin Lin, Changxing Ding, Jing Zhang, Yibing Zhan, and Dacheng Tao. Ru-net: Regularized unrolling network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19466, 2022. 4", + "[17] Kien Nguyen, Subarna Tripathi, Bang Du, Tanaya Guha, and Truong Q Nguyen. In defense of scene graphs for image captioning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1407-1416, 2021. 1", + "[18] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research Journal, pages 1-31, 2024. 6", + "[19] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2856-2865, 2021. 1", + "[20] Ege Özsoy, Evin Pinar Örnek, Ulrich Eck, Tobias Czempiel, Federico Tombari, and Nassir Navab. 4d-or: Semantic scene graphs for or domain modeling. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 475–485. Springer, 2022. 1" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25007", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[21] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 4, 6", + "[22] Brigit Schroeder and Subarna Tripathi. Structured query-based image retrieval using scene graphs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 178-179, 2020. 1", + "[23] Sebastian Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In Proceedings of the fourth workshop on vision and language, pages 70–80, 2015. 1", + "[24] Sahand Sharifzadeh, Sina Moayed Baharlou, and Volker Tresp. Classification by attention: Scene graph classification with prior knowledge. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5025-5033, 2021. 2", + "[25] Motoharu Sonogashira, Masaaki Iiyama, and Yasutomo Kawanishi. Towards open-set scene graph generation with unknown objects. IEEE Access, 10:11574-11583, 2022. 2", + "[26] Keisuke Tateno, Federico Tombari, and Nassir Navab. Real-time and scalable incremental segmentation on dense slam. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4465-4472. IEEE, 2015. 3, 6", + "[27] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4", + "[28] Johanna Wald, Helisa Dhamo, Nassir Navab, and Federico Tombari. Learning 3d semantic scene graphs from 3d indoor reconstructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3961-3970, 2020. 1, 2, 5, 6", + "[29] Johanna Wald, Nassir Navab, and Federico Tombari. Learning 3d semantic scene graphs with instance embeddings. International Journal of Computer Vision, 130(3):630-651, 2022. 1, 6", + "[30] Ziqin Wang, Bowen Cheng, Lichen Zhao, Dong Xu, Yang Tang, and Lu Sheng. Vl-sat: Visual-linguistic semantics assisted training for 3d semantic scene graph prediction in point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21560–21569, 2023. 2", + "[31] Shun-Cheng Wu, Johanna Wald, Keisuke Tateno, Nassir Navab, and Federico Tombari. Scenographfusion: Incremental 3d scene graph prediction from rgb-d sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7515-7525, 2021. 1, 2, 3", + "[32] Shun-Cheng Wu, Keisuke Tateno, Nassir Navab, and Federico Tombari. Incremental 3d semantic scene graph prediction from rgb sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5064-5074, 2023. 1, 2, 3, 4, 6, 8", + "[33] Danfei Xu, Yuke Zhu, Christopher B Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In Pro" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ceedings of the IEEE conference on computer vision and pattern recognition, pages 5410-5419, 2017. 2", + "[34] Ning Xu, An-An Liu, Jing Liu, Weizhi Nie, and Yuting Su. Scene graph captioner: Image captioning based on structural visual representation. Journal of Visual Communication and Image Representation, 58:477-485, 2019. 1", + "[35] Jianwei Yang, Jiasen Lu, Stefan Lee, Dhruv Batra, and Devi Parikh. Graph r-cnn for scene graph generation. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 2", + "[36] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 2, 3", + "[37] Xu Yang, Kaihua Tang, Hanwang Zhang, and Jianfei Cai. Auto-encoding scene graphs for image captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10685–10694, 2019. 1", + "[38] Sangwoong Yoon, Woo Young Kang, Sungwook Jeon, SeongEun Lee, Changjin Han, Jonghun Park, and Eun-Sol Kim. Image-to-image retrieval by learning similarity between scene graphs. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10718-10726, 2021. 1", + "[39] Kun Yuan, Manasi Kattel, Joel L Lavanchy, Nassir Navab, Vinkle Srivastav, and Nicolas Padoy. Advancing surgical vqa with scene graph knowledge. International Journal of Computer Assisted Radiology and Surgery, 19(7):1409-1417, 2024. 1", + "[40] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural motifs: Scene graph parsing with global context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5831-5840, 2018. 2", + "[41] Chaoyi Zhang, Jianhui Yu, Yang Song, and Weidong Cai. Exploiting edge-oriented reasoning for 3d point-based scene graph analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9705-9715, 2021. 2", + "[42] Shoulong Zhang, Aimin Hao, Hong Qin, et al. Knowledge-inspired 3d scene graph prediction in point cloud. Advances in Neural Information Processing Systems, 34:18620-18632, 2021. 2", + "[43] Zhongping Zhang, Huiwen He, Bryan A Plummer, Zhenyu Liao, and Huayan Wang. Complex scene image editing by scene graph comprehension. In BMVC, 2023. 1" + ], + "bbox": [ + 516, + 92, + 906, + 728 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25008", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Statistical Confidence Rescoring for Robust 3D Scene Graph Generation from Multi-View Images/00c53575-a21f-463b-8889-7a2370d2bff7_model.json b/2025/Statistical Confidence Rescoring for Robust 3D Scene Graph Generation from Multi-View Images/00c53575-a21f-463b-8889-7a2370d2bff7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5101ce8cad1fe59b1336455ff0931347a13a5888 --- /dev/null +++ b/2025/Statistical Confidence Rescoring for Robust 3D Scene Graph Generation from Multi-View Images/00c53575-a21f-463b-8889-7a2370d2bff7_model.json @@ -0,0 +1,2101 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.13, + 0.892, + 0.177 + ], + "angle": 0, + "content": "Statistical Confidence Rescoring for Robust 3D Scene Graph Generation from Multi-View Images" + }, + { + "type": "text", + "bbox": [ + 0.332, + 0.204, + 0.667, + 0.221 + ], + "angle": 0, + "content": "Qi Xun Yeo Yanyan Li Gim Hee Lee" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.222, + 0.766, + 0.24 + ], + "angle": 0, + "content": "Department of Computer Science, National University of Singapore" + }, + { + "type": "text", + "bbox": [ + 0.324, + 0.242, + 0.669, + 0.256 + ], + "angle": 0, + "content": "{qixunyeo,yan.li,gimhee.lee}@nus.edu.sg" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.327, + 0.485, + 0.66 + ], + "angle": 0, + "content": "Modern 3D semantic scene graph estimation methods utilize ground truth 3D annotations to accurately predict target objects, predicates, and relationships. In the absence of given 3D ground truth representations, we explore leveraging only multi-view RGB images to tackle this task. To attain robust features for accurate scene graph estimation, we must overcome the noisy reconstructed pseudo point-based geometry from predicted depth maps and reduce the amount of background noise present in multi-view image features. The key is to enrich node and edge features with accurate semantic and spatial information and through neighboring relations. We obtain semantic masks to guide feature aggregation to filter background features and design a novel method to incorporate neighboring node information to aid robustness of our scene graph estimates. Furthermore, we leverage on explicit statistical priors calculated from the training summary statistics to refine node and edge predictions based on their one-hop neighborhood. Our experiments show that our method outperforms current methods purely using multi-view images as the initial input. Our project page is available at https://qixun1.github.io/projects/SCRSSG." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.692, + 0.222, + 0.708 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The semantic scene graph (SSG) is a crucial intermediate representation that enhances higher-level scene understanding. It plays a key role in tasks such as image captioning [17, 34, 37], image retrieval [13, 22, 23, 38], image editing [3, 5, 43], and medical applications [7, 12, 15, 20, 39] by capturing both semantic and, more importantly, relational information between objects and their surroundings. Initially developed for 2D images, SSG has since expanded to the 3D domain [6, 9, 19]. 3D scene graphs provide a high-level representation of an entire 3D scene using inputs such as multi-view RGB images or LiDAR point clouds. Unlike their 2D counterparts, they incorporate spatial relation" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.294, + 0.905, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.401, + 0.904, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.518, + 0.907, + 0.575 + ], + "angle": 0, + "content": "Figure 1. Existing multi-view RGB methods aggregate features using bounding boxes, introducing background noise that hinders accuracy. We use pretrained segmentation masks to reduce this interference and thus improving prediction accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.906, + 0.626 + ], + "angle": 0, + "content": "ships that extend beyond the visible image plane, allowing a holistic understanding of complex environments." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.627, + 0.907, + 0.822 + ], + "angle": 0, + "content": "Many approaches [1, 28, 29] rely on ground-truth 3D LiDAR point clouds as input. However, LiDAR sensors are resource-intensive and lack inherent semantic richness. Recent works [31, 32] have shifted toward using semantically rich multi-view RGB images, which offer a balance between computational efficiency and improved semantic fidelity. Although point cloud-based methods remain valuable for capturing geometric information, our approach focuses on leveraging multi-view RGB images. By incorporating pretrained depth estimators, we generate pseudo point-based geometry to estimate the 3D semantic scene graph, reducing reliance on LiDAR while preserving spatial and structural details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "A key challenge in using only multi-view RGB images is ensuring that node features remain free from distractors. As shown in Fig. 1, prior approaches that rely on entity detectors to extract 2D bounding box proposals or regions of interest often fail to guarantee robustness since distrac" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24999" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "tors within the bounding box are often mistakenly aggregated. Ensuring robust initial multi-view features is crucial without ground truth point-based geometry for error correction. No explicit mechanism exists to correct incorrectly bounded objects or misclassified rare classes that are not well trained on the model. To alleviate this problem, a mechanism should exist to refine estimates using prior knowledge instead of depending solely on implicit interactions between nodes and edges to achieve confident and accurate predictions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.243, + 0.487, + 0.545 + ], + "angle": 0, + "content": "In this paper, we propose a framework to enhance the robustness of the model. First, we introduce a masked feature initialization (MFI), which leverages segmentation masks to aggregate image features instead of relying on the bounding box proposals. This reduces background noise, which results in cleaner multi-view image features. Next, we design a robust residual spatial neighbor graph neural network (RSN-GNN) to encode spatial information into node features. This network filters highly activated regions from neighboring nodes, refining target node features for improved predicate estimation. Finally, we propose a confidence rescoring (CR) module, which refines object and predicate estimates using an inverse softmax-weighted contribution of neighboring node-to-node and node-to-edge co-occurrence counts. By integrating this explicit inductive bias, our approach improves the accuracy of the prediction, particularly in low-confidence scenarios. Extensive experimental results on the benchmark 3RScan dataset show the competitiveness of our proposed approach compared to existing 2D and 3D approaches." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.546, + 0.455, + 0.561 + ], + "angle": 0, + "content": "Our main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.564, + 0.483, + 0.61 + ], + "angle": 0, + "content": "- We introduce a masked feature initialization to enhance the robustness of node features by reducing background distractors, yielding cleaner multi-view image features." + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.612, + 0.484, + 0.657 + ], + "angle": 0, + "content": "- We design a novel GCN architecture that integrates highly activated neighboring features into the target node to increase the robustness of the edge features." + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.66, + 0.484, + 0.691 + ], + "angle": 0, + "content": "- We propose a new refinement module to explicitly refine predictions based on statistical prior knowledge." + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.694, + 0.484, + 0.738 + ], + "angle": 0, + "content": "- Experiments show that we outperform previous state-of-the-art approaches on the 3RScan dataset, particularly on metrics that deal with low-tail imbalanced classes." + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.564, + 0.484, + 0.738 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.754, + 0.235, + 0.77 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "2D Semantic Scene Graph. 2D semantic scene graph prediction is typically categorized into two-step and one-step approaches. The two-step approach first detects objects and then classifies their relationships [2, 4, 35, 40]. In contrast, the one-step approach jointly infers object and relationship classes [33]. Xu et al. [33] pioneers the problem of scene graph generation and they tried to solve this problem via iterative message passing. Baier et al. [2] first showed" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.321 + ], + "angle": 0, + "content": "how semantic models can be improved by incorporating triplet frequencies. Zellers et al. [40] analyzes the usefulness of statistical co-occurrences for the Visual Genome dataset and concluded that such statistical priors serves as strong regularization for the task. Chen et al. [4] formalizes the first approach (KERN) to incorporate the statistical prior directly into graph neural networks. Sharifzadeh et. al. proposes Schemata to assimilate image-based relational prior knowledge into the representations within the neural network [24]. Compared to previous methods, our method combines predictions from the prediction head with statistical co-occurrence of the current node and neighbors to determine the final class using a confidence score. Unlike Sonogashira et. al. [25] which uses hard thresholding, our approach uses adaptive relationship and node refinements." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.909, + 0.903 + ], + "angle": 0, + "content": "3D Semantic Scene Graph. Methods for generating a 3D semantic scene graph can be categorized into two types based on input data. The first relies on 3D inputs such as ground-truth 3D geometry to learn contextual information [28, 30, 41]. The second relies on multi-view RGB images with some methods incorporating depth information [31] while others do not [32]. Our approach follows the latter, leveraging the estimated depth from off-the-shelf depth estimators [36] to enhance the information provided by multi-view RGB images. Wald et al. [28] first proposed 3RScan: a richly annotated 3D scene graph dataset and a method that focuses on pairwise relationships to predict the 3D scene graphs. Wu et al. [31] uses SLAM to obtain dense point representations and a novel graph convolutional network (GCN) based aggregation function to enhance 3D scene graph prediction. Zhang et al. [41] proposes an edge-oriented GCN to incorporate multi-dimensional edge features for explicit inter-node relationship modeling. Zhang et al. [42] integrates prior commonsense knowledge by learning meta-embeddings only from the class labels with their graphical structures to avoid perceptual errors. Wang et al. [30] incorporated additional information by distilling knowledge learned by the multi-modal oracle model to the 3D model. Wu et al. [32] introduces a novel entity association method to obtain the 3D entities from the predicted 2D entities and a geometric gate to incorporate geometric information to multi-view image features. Feng et al. [8] uses ConceptNet external knowledge base to accumulate both contextualized visual content and textual facts to form a 3D spatial multimodal knowledge graph. Our method leverages on the explicit statistical priors similar to 2D scene graph method KERN and Schemata while guiding low confidence predictions made by the initial node and edge predictors using both visual and geometric cues." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25000" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.905, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.255, + 0.908, + 0.314 + ], + "angle": 0, + "content": "Figure 2. The architecture of our proposed framework consists of three main components: 1) A masked feature initialization step leverages backprojected semantic masks to aggregate multi-view image, geometric, and spatial features. 2) A message passing step utilizes our residual spatial neighbor GNN module to refine node and edge features through feature interaction. 3) A confidence rescoring step adjusts logits based on summary statistics of object and predicate co-occurrence from the training dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.336, + 0.266, + 0.354 + ], + "angle": 0, + "content": "3. Our Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.362, + 0.483, + 0.409 + ], + "angle": 0, + "content": "Problem Definition. Given either a set of 3D point clouds \\( P \\) or multi-view RGB images \\( I \\) of a 3D scene as input, the goal is to estimate the 3D scene graph denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.242, + 0.421, + 0.483, + 0.437 + ], + "angle": 0, + "content": "\\[\nG = (O, R), \\tag {1}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.45, + 0.138, + 0.463 + ], + "angle": 0, + "content": "where" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.466, + 0.384, + 0.48 + ], + "angle": 0, + "content": "- \\(N\\): Number of multi-view input images;" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.48, + 0.385, + 0.495 + ], + "angle": 0, + "content": "- \\( O = \\{o_m\\}_{m=1}^M \\): All objects in the scene;" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.495, + 0.481, + 0.512 + ], + "angle": 0, + "content": "- \\( R = \\{r_{k\\to m}\\}_{m,k = 1}^{M,M} \\): Relationships between all objects;" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.512, + 0.396, + 0.526 + ], + "angle": 0, + "content": "- \\( M = |O| \\): Number of objects in the scene." + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.466, + 0.481, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.484, + 0.607 + ], + "angle": 0, + "content": "Furthermore, we refer to \\( o_m \\) as the object class for node instance \\( m \\), \\( r_{k \\to m} \\) as the predicate class for the edge connecting nodes \\( k \\) and \\( m \\). Alternatively, the 3D scene graph can be defined as a set of triplets given by: \\( \\{o_k, r_{k \\to m}, o_m\\} := \\{node, predicate, node\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.734 + ], + "angle": 0, + "content": "Overview. Fig. 2 shows an illustration of the architecture of our proposed framework. The input to our framework is a set of \\(N\\) multi-view RGB images \\(\\{I_i\\}_{i=1}^N\\) of the 3D scene. Since the focus of our work is not on multi-view 3D reconstruction, we recover the 3D point cloud of the scene by running an off-the-shelf RGB-D SLAM system [26] with the metric depth maps of the input images \\(\\{\\hat{D}_i\\}_{i=1}^N\\) obtained from the pretrained Depth Anything [36] model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "From the predicted 3D point cloud, we initialize multiview image features using segmentation masks, extract geometric features with a point-based feature extractor, and derive spatial features via a linear layer. These multi-view image features serve as the initial input to our RSN-GNN module. At the GCN initialization, we enhance node features by integrating geometric and spatial features through geometric and spatial gates. We additionally enhance edge features by integrating max pooled neighboring nodes features. These enhanced node and edge features undergo message passing, with the integration process repeating at each" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.338, + 0.905, + 0.398 + ], + "angle": 0, + "content": "layer. After \\( k \\) iterations, the final node and edge features are processed by predictors to generate initial estimates, which are then refined using precomputed co-occurrence statistics via our CR module, producing the final scene graph output." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.905, + 0.445 + ], + "angle": 0, + "content": "We elaborate on the feature initialization step in Sec. 3.1, the message passing step in Sec. 3.2, and the confidence rescoring step in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.455, + 0.829, + 0.47 + ], + "angle": 0, + "content": "3.1. Masked Feature Initialization (MFI)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.906, + 0.689 + ], + "angle": 0, + "content": "We initialize the node features of the scene graph by combining multi-view image representations from a pretrained image encoder with PointNet features derived from the predicted points in the RGB-D SLAM reconstruction step that utilizes predicted depth. To eliminate background clutter, we employ the Segment Anything Model (SAM) [14] to produce precise 2D semantic masks for each object. For each node, we backproject the associated points onto the 2D instance mask to identify its corresponding region. We then generate object proposals from these instance masks, resize them to \\(224 \\times 224\\), and extract their multi-view image features using the pretrained encoder. Finally, we apply these semantic masks to the multi-view image features before aggregating them into our node representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Our method to get the object proposals, i.e. the nodes of the scene graph offers two key advantages over relying solely on the entity detector from SceneGraphFusion and JointSSG [31, 32]. First, the entity detector often captures only part of an object. Although SceneGraphFusion propagates the label for this segment, the feature extractor can operate only on the detected portion. In contrast, our method employs more complete object masks, enabling the feature extractor to derive features from the entire object. Second, we exclude extraneous background features by applying these masks before feature aggregation. In SceneGraphFusion and JointSSG, the bounding boxes of the pretrained entity detector frequently include background regions, which can dilute the quality of the features and hinder" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25001" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.137 + ], + "angle": 0, + "content": "the performance. Our method is closer to the open vocabulary method ConceptGraphs [10] which uses SAM masks but with the aggregation method provided by JointSSG." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.137, + 0.484, + 0.244 + ], + "angle": 0, + "content": "Formally, we compute a corresponding multi-view image feature \\(\\mathbf{v}_{i,img}\\), a geometric feature \\(\\mathbf{v}_{i,geo}\\), and a spatial feature \\(\\mathbf{v}_{i,spat}\\) for each node \\(v_{i}\\). We aggregate features only for nodes that are visible in the可视ability graph \\(G_{v}\\) denoted by \\(\\mathbf{1}(v_{i} \\in I_{k})\\) and have corresponding pixels within the semantic mask provided by SAM. The aggregated multi-view image feature \\(\\mathbf{v}_{i,img}\\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.254, + 0.483, + 0.29 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {i} ^ {0} = \\frac {\\sum_ {k = 1} ^ {N} \\mathbf {1} \\left(v _ {i} \\in I _ {k}\\right) f \\left(\\eta_ {k} \\odot x _ {i , k}\\right)}{\\left| k \\in G _ {v} \\right|}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.3, + 0.484, + 0.39 + ], + "angle": 0, + "content": "which also serves as the initial node feature. \\( f(\\cdot) \\) is the image feature encoder and \\( \\eta_{k} \\) refers to the 2D semantic mask. \\( x_{i,k} \\) refers to the cropped image patch containing the node \\( i \\) from the \\( k \\)-th image. \\( \\odot \\) refers to the element-wise multiplication operator. The aggregation is the arithmetic mean of the node features for each image in the可视ability graph." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.39, + 0.484, + 0.496 + ], + "angle": 0, + "content": "The geometric feature \\(\\mathbf{v}_{i,geo}\\) is extracted using a vanilla PointNet [21], with inputs derived from the 3D instance masks generated during the RGB-D SLAM step. For the spatial features, we employ a simple linear layer to transform the attributes of each 3D bounding box: dimensions \\(\\mathbf{b}_i\\), volume \\(s_i\\), length \\(l_i\\), and the ratio of its x- and y-axes \\(r_i\\) into the spatial feature as:" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.507, + 0.483, + 0.525 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {i, s p a t} = g _ {s} ([ \\mathbf {b} _ {i}, s _ {i}, l _ {i}, r _ {i} ]), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.484, + 0.58 + ], + "angle": 0, + "content": "where \\( g_{s} \\) refers to a set of learnable weights for spatial features. \\([\\cdot]\\) refers to the concatenation of the inputs within the operator." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.589, + 0.437, + 0.605 + ], + "angle": 0, + "content": "3.2. Residual Spatial Neighbor GNN Module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.484, + 0.718 + ], + "angle": 0, + "content": "Following a similar design to the learnable geometric gate in JointSSG [32], we integrate both geometric and spatial features into the initial node feature, which originally consists only of multi-view image features. The spatial features serve as additional information for individual nodes, analogous to the embeddings used for token encoding in Transformer architectures [27]. We define the geometric gate as:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.726, + 0.483, + 0.747 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {v}} _ {i} ^ {l} = \\mathbf {v} _ {i} ^ {l} + \\sigma \\left(\\mathbf {w} _ {g} ^ {T} \\left[ \\mathbf {v} _ {i} ^ {l}, \\mathbf {v} _ {i, g e o} \\right]\\right) \\cdot \\sigma \\left(\\mathbf {v} _ {i, g e o}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.756, + 0.484, + 0.801 + ], + "angle": 0, + "content": "where \\(\\mathbf{w}_g\\) refers to a set of learnable weights, \\(\\sigma\\) refers to the sigmoid operator and \\(l\\) refers to the layer of the node feature. Furthermore, we define the spatial gate as:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.81, + 0.483, + 0.83 + ], + "angle": 0, + "content": "\\[\n\\ddot {\\mathbf {v}} _ {i} ^ {l} = \\hat {\\mathbf {v}} _ {i} ^ {l} + \\sigma \\left(\\mathbf {w} _ {s} ^ {T} [ \\hat {\\mathbf {v}} _ {i} ^ {l}, \\mathbf {v} _ {i, s p a t} ]\\right) \\cdot \\sigma \\left(\\mathbf {v} _ {i, s p a t}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.84, + 0.475, + 0.855 + ], + "angle": 0, + "content": "where \\(\\mathbf{w}_s\\) is a set of learnable weights for spatial features." + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.484, + 0.901 + ], + "angle": 0, + "content": "1The covisibility graph is obtained as intermediate output after running the RGB-D SLAM system. It refers to a bipartite graph where each image \\( I_{k} \\) is connected with the specific node instance \\( v_{i} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.228 + ], + "angle": 0, + "content": "Inspired by the residual connections in ResNet [11], we introduce a mechanism to integrate highly activated neighboring features into the target node feature using max pooling during the message-passing process for edge features. This approach implicitly encodes neighboring information, improving edge prediction for scene graph estimation. Specifically, we augment the edge messages in the GCN by integrating max-pooled neighboring features before message passing as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.233, + 0.905, + 0.253 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {v}} _ {i} ^ {l} = \\ddot {\\mathbf {v}} _ {i} ^ {l} + \\max ([ \\ddot {\\mathbf {v}} _ {j} ^ {l} ] _ {j \\in n (i)}), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.906, + 0.332 + ], + "angle": 0, + "content": "where \\(\\max (\\cdot)\\) refers to the max pooling operator and \\(n(i)\\) refers to the neighboring nodes of node \\(i\\). Max pooling allows highly activated regions in the different neighboring node features to be fused into the target node feature via the residual connection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.334, + 0.906, + 0.452 + ], + "angle": 0, + "content": "We apply max-pooled neighboring features to the calculation of edge messages without using them in node message updates. Since node features are highly susceptible to noise and spurious correlations from other nodes [16], incorporating max-pooled neighboring features into node message computations can introduce confusion instead of improving performance. The edge message \\( m_{i\\rightarrow j} \\) is thus defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.453, + 0.905, + 0.472 + ], + "angle": 0, + "content": "\\[\nm _ {i \\rightarrow j} = g _ {e} ([ \\tilde {\\mathbf {v}} _ {i} ^ {l}, \\mathbf {e} _ {i j} ^ {l}, \\tilde {\\mathbf {v}} _ {j} ^ {l} ]), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.906, + 0.536 + ], + "angle": 0, + "content": "where \\( g_{e}(\\cdot) \\) is an MLP and \\( \\mathbf{e}_{ij}^{l} \\) is the edge descriptor used in [32] containing relative node properties such as centroid displacement, relative bounding box size difference, and a relative pose descriptor." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.544, + 0.824, + 0.56 + ], + "angle": 0, + "content": "3.3. Confidence Rescoring (CR) Module" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.566, + 0.906, + 0.776 + ], + "angle": 0, + "content": "As discussed in Sec. 1, modern scene graph methods depend heavily on the entity detector and image feature extractor to produce an accurate initial object prediction. However, there is no mechanism to correct a target bounding box prediction that is initially wrong. To address this limitation, we leverage relationships with neighboring nodes, allowing us to mitigate the impact of incorrect initial predictions by weighting contributions from neighboring node statistics to refine target object classification. Fig. 3 illustrates the logits of the final prediction after applying our confidence rescoring module. This module incorporates prior knowledge to improve prediction confidence beyond that of the original node and edge multi-layer perceptron (MLP) predictor." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.777, + 0.906, + 0.823 + ], + "angle": 0, + "content": "The probability of predicting node \\(i\\) given its neighboring nodes \\(n(i)\\) is approximated by the output of our node predictor \\(h_v(\\cdot)\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.828, + 0.905, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} P \\left(o _ {i} \\mid \\left\\{o _ {j}: j \\in n (i) \\right\\}\\right) = \\frac {1}{Z _ {i}} \\psi_ {i} \\left(o _ {i}\\right) \\prod_ {j \\in n (i)} \\psi_ {i j} \\left(o _ {i}, o _ {j}\\right) \\\\ \\approx \\tau \\left(h _ {v} \\left(\\mathbf {v} ^ {L}\\right)\\right), \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25002" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.228 + ], + "angle": 0, + "content": "where \\(\\psi_i(o_i)\\) refers to the potential associated with node \\(i\\) and \\(\\psi_{ij}(o_i,o_j)\\) refers to the potential associated with the edge between node \\(i\\) and its neighbor node \\(j\\). The potentials are implicitly learned through the GCN unlike classical methods, which explicitly learn these potentials. \\(h_v(\\cdot)\\) is a small MLP that predicts the class of each node and \\(\\tau\\) refers to the softmax operator. \\(L\\) refers to the last layer of the GCN message passing output. The partition function given by:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.236, + 0.483, + 0.27 + ], + "angle": 0, + "content": "\\[\nZ _ {i} = \\sum_ {o _ {i}} \\psi_ {i} \\left(o _ {i}\\right) \\prod_ {j \\in n (i)} \\psi_ {i j} \\left(o _ {i}, o _ {j}\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.279, + 0.484, + 0.326 + ], + "angle": 0, + "content": "ensures the probabilities sum to 1. A similar equation can be obtained for the probability of predicting edge \\( ij \\) given its neighboring nodes \\( i \\) and \\( j \\), i.e.:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.332, + 0.482, + 0.385 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} P \\left(r _ {i \\rightarrow j} \\mid o _ {i}, o _ {j}\\right) = \\frac {1}{Z _ {i j}} \\psi_ {i j} \\left(o _ {i}, o _ {j}, r _ {i \\rightarrow j}\\right) \\tag {10} \\\\ \\approx \\tau (h _ {e} (\\mathbf {e} ^ {L})), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.393, + 0.484, + 0.439 + ], + "angle": 0, + "content": "where \\( h_e(\\cdot) \\) is a small MLP which predicts the class of each edge, and the partition function to normalize the probabilities is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.447, + 0.483, + 0.48 + ], + "angle": 0, + "content": "\\[\nZ _ {i j} = \\sum_ {r _ {i \\rightarrow j}} \\psi_ {i j} \\left(o _ {i}, o _ {j}, r _ {i \\rightarrow j}\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.484, + 0.595 + ], + "angle": 0, + "content": "We begin by computing summary statistics from the training dataset to node-to-node co-occurrences denoted as \\( c_{ij}(o_i, o_j) \\), and node-to-edge co-occurrences denoted as \\( d_{ij}(o_i, e_j) \\). Subsequently, we determine the prediction confidence of the model, where \\( \\alpha_{\\mathbf{v}} \\) represents node confidence and \\( \\alpha_{\\mathbf{e}} \\) represents edge confidence. Next, we obtain the node \\( \\alpha_{\\mathbf{v}} \\) and edge \\( \\alpha_{\\mathbf{e}} \\) confidences as:" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.603, + 0.483, + 0.626 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {\\mathbf {v}} = \\max _ {\\mathbf {v}} \\tau \\left(h _ {v} \\left(\\mathbf {v} ^ {L}\\right)\\right), \\quad \\alpha_ {\\mathbf {e}} = \\max _ {\\mathbf {e}} \\tau \\left(h _ {e} \\left(\\mathbf {e} ^ {L}\\right)\\right). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.633, + 0.484, + 0.693 + ], + "angle": 0, + "content": "We first estimate the marginal probability of a specific instantiation of node \\( i \\), conditioned on node \\( j \\) using node-to-node co-occurrences, and on edge \\( ij \\) using node-to-edge co-occurrences defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.7, + 0.435, + 0.72 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} \\left(o _ {i} \\mid o _ {j}\\right) = \\left[ c \\left(o _ {i} = c _ {k} \\mid o _ {j} = c _ {l}\\right) \\right] _ {k = 0, l = 0} ^ {C, C}, \\quad \\text {w h e r e}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.722, + 0.483, + 0.759 + ], + "angle": 0, + "content": "\\[\nc \\left(o _ {i} = c _ {k} \\mid o _ {j} = c _ {l}\\right) = \\frac {c \\left(o _ {i} = c _ {k} , o _ {j} = c _ {l}\\right)}{\\sum_ {k = 0} ^ {C} c \\left(o _ {i} = c _ {k} , o _ {j} = c _ {l}\\right)}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.841 + ], + "angle": 0, + "content": "Here, \\( C \\) refers to the number of object classes and \\( c_{k} \\) refers to the \\( k \\)-th class in the set of object classes. We then integrate the summary statistics via an inverse softmax operation denoted as \\( \\gamma(\\cdot) \\) into the node probability in Eq. 14 to get:" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.849, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} P \\left(o _ {i} \\mid \\left\\{o _ {j}: j \\in n (i) \\right\\}\\right) = \\tau \\left(\\alpha_ {\\mathbf {v}} \\cdot h _ {v} (\\mathbf {v} _ {i}) \\right. \\\\ + \\left(1 - \\alpha_ {\\mathbf {v}}\\right) \\cdot \\sum_ {j \\in n (i)} \\alpha_ {j} \\cdot \\gamma (\\mathbf {c} \\left(o _ {i} \\mid o _ {j}\\right)). \\tag {14} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "This formulation is chosen because directly adding the estimated marginal probability to the predictions of the model after softmax would disregard the required normalization, resulting in incompatible terms." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.152, + 0.906, + 0.212 + ], + "angle": 0, + "content": "Similarly, we compute the estimated marginal probability of an edge \\( ij \\) instantiation conditioned on nodes \\( i \\) and \\( j \\) using node-to-edge co-occurrence statistics for predicate estimation:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.221, + 0.893, + 0.242 + ], + "angle": 0, + "content": "\\[\n\\mathbf {d} \\left(r _ {i \\rightarrow j} \\mid o _ {j}\\right) = \\left[ d \\left(r _ {i \\rightarrow j} = e _ {k} \\mid o _ {j} = c _ {l}\\right)\\right] _ {k = 0, l = 0} ^ {P, C},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.244, + 0.905, + 0.295 + ], + "angle": 0, + "content": "\\[\nd \\left(r _ {i \\rightarrow j = k} \\mid o _ {j = l}\\right) = \\frac {d \\left(r _ {i \\rightarrow j} = e _ {k} , o _ {j} = c _ {l}\\right)}{\\sum_ {k = 0} ^ {P} d \\left(r _ {i \\rightarrow j} = e _ {k} , o _ {j} = c _ {l}\\right)}. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.307, + 0.906, + 0.354 + ], + "angle": 0, + "content": "Here, \\( P \\) refers to the total number of predicate classes and \\( e_k \\) refers to the \\( k \\)-th class in the set of predicate class. The refined logits for predicate estimation are then given by:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.365, + 0.905, + 0.42 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} P \\left(r _ {i \\rightarrow j} \\mid o _ {i}, o _ {j}\\right) = \\tau \\left(\\alpha_ {i j} \\cdot h _ {e} (\\mathbf {e} _ {i j}) \\right. \\\\ + \\left(1 - \\alpha_ {i j}\\right) \\cdot \\left(\\alpha_ {i} \\cdot \\gamma (\\mathbf {d} \\left(r _ {i \\rightarrow j} \\mid o _ {i}\\right))\\right) \\tag {16} \\\\ \\cdot \\left(\\alpha_ {j} \\cdot \\gamma (\\mathbf {d} (r _ {i \\rightarrow j} | o _ {j})))\\right). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.907, + 0.657 + ], + "angle": 0, + "content": "Remarks. As shown in Fig. 3, the original node probability for the correct class is already relatively high. However, our CR module further enhances the predicted probability as evidenced by the upward shift in the box plot from \"Original\" to \"After CR\" in the red boxes, with a notable increase in the lower quartile. Moreover, a larger proportion of low node probability instances are flagged as outliers as shown by the greater number of red circles at \"After CR\" compared to \"Original\". Most notable gains occur when original logits are low since CR increases the probability of the correct class by down-weighting misclassified classes \\( c' \\) via lower conditional probabilities \\( P(c'|c) \\) given conditioned class \\( c \\). This targeted adjustment raises suppressed logits for rare classes, thus improving prediction performance for underrepresented categories." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.671, + 0.646, + 0.688 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.696, + 0.704, + 0.712 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.906, + 0.84 + ], + "angle": 0, + "content": "Datasets. We evaluate on the 3RScan dataset following the same training and test split from [28]. The image quality of the dataset is rather poor as each scene constitutes a variable number of short videos spliced together with different motion trajectories. As the images are obtained from consumer-level cameras, the images possess a low frame rate of \\(10\\mathrm{Hz}\\), image blur, and image jitter from sudden camera motion which makes this dataset challenging." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.847, + 0.906, + 0.877 + ], + "angle": 0, + "content": "Implementation Details. We utilize the codebase provided from the 3DSSG GitHub repository\\(^2\\) to reproduce the prior" + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.887, + 0.764, + 0.901 + ], + "angle": 0, + "content": "\\(^{2}\\)https://github.com/ShunChengWu/3DSSG" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25003" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.09, + 0.909, + 0.16 + ], + "angle": 0, + "content": "
Methodbath.bedbkshfcab.chaircntr.curt.deskdoorfloorofurnpic.refri.show.sinksofatabletoil.wallwind.Mean.
IMP0.0001.0000.0000.3410.4670.0000.6060.2000.5280.9000.1250.1430.0000.0000.3330.4500.5250.8000.6220.1250.358
VGfM0.5000.0000.0000.4150.5430.0830.5450.0000.3890.9200.2320.0860.0000.0000.3810.4500.5740.8000.6440.1670.336
3DSSG0.0000.6670.0000.2070.3480.1670.5760.2000.3610.7800.1250.0570.0000.0000.3330.5000.2790.0000.4670.1250.260
SGFN0.5000.3330.0000.2930.6850.2500.5150.2000.4440.8800.2320.1710.0000.0000.4290.5000.5080.4000.7070.1250.359
JointSSG1.0001.0000.0000.4760.6740.2500.5760.2000.5000.9000.1070.1430.0000.0000.4290.5500.5410.8000.6800.2500.454
Ours1.0001.0000.3330.6220.6300.3330.5450.6000.5280.8800.3040.1710.6670.6670.3330.6000.6230.8000.6930.3750.605
" + }, + { + "type": "table_caption", + "bbox": [ + 0.115, + 0.172, + 0.882, + 0.187 + ], + "angle": 0, + "content": "Table 1. Comparison with state-of-the-art methods on the 3RScan dataset for each object class. The Best results are highlighted." + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.215, + 0.429, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.407, + 0.484, + 0.451 + ], + "angle": 0, + "content": "Figure 3. Box plot for predicted node probability for correct class after our CR module. The lower quartile visibly moves upwards due to an increase in predicted node probability for correct class." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.468, + 0.483, + 0.514 + ], + "angle": 0, + "content": "methods. For JointSSG[32], we replace their image feature extractor from a pretrained ResNet18 to a pretrained DINOv2 with a ViT-L backbone to ensure fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.514, + 0.483, + 0.696 + ], + "angle": 0, + "content": "Following [32], we use PointNet [21] as the point encoder and employ two message-passing layers. For RGB-D SLAM reconstruction, we utilize the method from [26] and our predicted depth to generate predicted points. Multiview feature extraction is performed using a pretrained DINOv2 [18] with a ViT-L backbone. To obtain object segmentation masks, we use a pretrained SAM predictor with a text prompt containing the class set of 20 NYU object classes following the subset defined in [28]. Node-to-node and node-to-edge co-occurrences are precomputed on the training set. We train the model on a single RTX 3090 GPU for two days with early stopping." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.701, + 0.483, + 0.822 + ], + "angle": 0, + "content": "Evaluation Metrics. Following [29] and JointSSG[32], we report the overall top-1 recall (Recall) for the object class estimation (Obj.), the predicate estimation (Pred.), and the relationship triplet estimation (Rel.). We also report the mean recall (mRecall) for the object class estimation (Obj.), the predicate estimation (Pred.) only. We map all predictions on the estimated segments to the ground truth segments for a fair comparison between the different methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.285, + 0.849 + ], + "angle": 0, + "content": "4.2. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We present the per-class recall for 20 object classes in the 3RScan dataset in Tab. 1, demonstrating that our method significantly outperforms others across most object cate" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.205, + 0.907, + 0.553 + ], + "angle": 0, + "content": "gories. Notably, we are the only approach that achieves a recall above \\(0\\%\\) for all object classes, indicating that we robustly learn all classes instead of just a subset. Our method tends to underperform slightly compared to the other methods on majority classes such as floor and wall classes in return, although not by a significant margin. The main experimental results for the top-1 recall and the mean recall are shown in Tab. 2. Importantly, our approach significantly outperforms others in mean recall (mRecall) for both objects and predicates, suggesting that it handles class imbalance issues [29] more effectively than competing methods. Additionally, our method surpasses all baselines in relationship and object estimation metrics while achieving comparable performance in predicate estimation. More impressively, our method with the weaker ResNet18 multiview image features already surpasses all baselines in relationship and object estimation metrics except for predicate estimation. Our method with ResNet18 image features even manages to surpass the performance of JointSSG with DINOv2 multi-view image features by non-trivial margins. This shows that our improved performance is not purely due to the more powerful DINOv2 features. Refer to our supplementary material for more quantitative results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.562, + 0.696, + 0.578 + ], + "angle": 0, + "content": "4.3. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.81 + ], + "angle": 0, + "content": "We show qualitative results comparing our method with other competitive methods in Fig. 4. Since the scene contains a large number of objects, we display only a subset with fewer objects and predicates for better readability. The correct prediction of the None predicate class is omitted in the visualization due to its prevalence in the dataset. As seen in Fig. 4, our method can distinguish the relatively more ambiguous wall which the other methods fail to distinguish. The shower curtain is tricky to classify since it is a subset of a curtain that is another class of objects present in the dataset. Understandably, all methods fail to correctly classify this object. Both our method and JointSSG attain good performance on the predicate class compared to SGFN as both methods only misclassify the predicate between the curtain object and the wall." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "We additionally evaluate on the entire scene for scan 43b8cae1 without comparison to other methods. Our method can recognize rare classes present in the scene, which are more challenging to learn. As seen in Fig. 5, our approach successfully classifies the rare class bookshelf instance denoted by \\(bkshf\\) which no other method managed" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "25004" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.137, + 0.093, + 0.862, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.424, + 0.908, + 0.48 + ], + "angle": 0, + "content": "Figure 4. The qualitative results on 3RScan dataset of previous works and our proposed framework on scan 4d3d82b0. SGFN also fails to classify majority of non-background object classes except the sink and cabinet. It also fails to classify all predicate classes. JointSSG performs better but misclassifies the shower curtain as a curtain and the ambiguous blue wall as a cabinet. Our method correctly classifies the wall but misidentifies the shower curtain as a curtain." + }, + { + "type": "image", + "bbox": [ + 0.152, + 0.487, + 0.847, + 0.817 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.83, + 0.908, + 0.873 + ], + "angle": 0, + "content": "Figure 5. The qualitative results on 3RScan dataset of our proposed framework on the entire scene of scan 43b8cae1. Our method can correctly classify most objects in the scene except for the light blue dress table at the top right of the scan. Our method can also correctly classify most predicates." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "25005" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.099, + 0.09, + 0.478, + 0.326 + ], + "angle": 0, + "content": "
MethodRecall%mRecall%
RelObj.Pred.Obj.Pred
IMP [32]25.851.890.430.023.0
VGfM [32]28.353.390.731.624.4
3DSSG [32]17.541.488.231.926.6
SGFN [32]31.456.789.638.330.5
JointSSG [32]34.158.189.943.033.3
IMP25.447.789.935.819.9
VGfM27.450.190.633.622.3
3DSSG13.035.686.826.024.3
SGFN29.952.289.535.927.1
JointSSG32.954.588.745.435.2
JointSSG‡34.356.686.552.936.2
Ours†38.659.389.157.436.8
Ours40.561.890.460.539.2
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.337, + 0.482, + 0.455 + ], + "angle": 0, + "content": "Table 2. Comparison with state-of-the-art methods on the 3RScan dataset with 20 object classes and 8 predicate classes. The top group of results are reported in JointSSG [32]. The middle group of results are reproduced via 3DSSG GitHub repository. JointSSG‡ refers to JointSSG but with DINOv2 multi-view image features. Ours† uses ResNet18 multi-view image features; Ours uses the DINOv2 multi-view image features instead. The Best and Second Best results are highlighted, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.472, + 0.483, + 0.534 + ], + "angle": 0, + "content": "to predict based on the results shown in Tab. 1. Our method also predicts multiple reasonable predicates of standing on between various objects and floor. Additional qualitative results can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.545, + 0.255, + 0.56 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.483, + 0.839 + ], + "angle": 0, + "content": "We analyze the effects of our feature initialization, robust neighbor residuals, and confidence rescoring module in Tab. 3. Our confidence rescoring (CR) module significantly improves performance across all metrics, indicating that it effectively leverages statistical priors to address class imbalance issues as noted in [32]. Apart from predicate mRecall, incorporating our masked feature initialization further improves performance across all metrics. This outcome is expected since masking primarily aims to refine multi-view image features, which do not directly contribute to edge feature improvements. The combination of the masked feature initialization and the CR module act as a force multiplier since the CR module benefits from cleaner and more accurate neighboring node features. Additionally, our robust neighbor residual gate maintains comparable performance across most metrics except for predicate mRecall, which benefits from additional neighbor information to better handle class imbalance in predicate estimation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We additionally study the effects of using our CR module during inference in Tab. 4. The addition of the CR module with the prior knowledge of the dataset is akin to calculating statistics during transductive setting of inference instead" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.09, + 0.912, + 0.168 + ], + "angle": 0, + "content": "
MethodConf.Feat.Neigh.Recall%mRecall%
Rel.Obj.Pred.Obj.Pred.
JointSSG×××32.954.588.245.435.2
+ Conf.✓××37.957.289.654.136.3
+ Feat.✓✓×41.161.390.861.335.8
Ours✓✓✓40.561.890.460.539.2
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.179, + 0.907, + 0.248 + ], + "angle": 0, + "content": "Table 3. Ablation study on the 3RScan dataset. Conf. refers to the usage of summary statistic priors for confidence rescoring via the CR module. Feat. refers to the usage of DINOv2 image features with refinement from SAM semantic masks via MFI. Neigh. refers to the usage of RSN-GNN. The Best results are shown in bold." + }, + { + "type": "table", + "bbox": [ + 0.568, + 0.267, + 0.858, + 0.322 + ], + "angle": 0, + "content": "
MethodCR.Recall%mRecall%
Rel.Obj.Pred.Obj.Pred.
Ours×41.161.190.660.235.7
Ours✓40.561.890.460.539.2
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.333, + 0.907, + 0.376 + ], + "angle": 0, + "content": "Table 4. Ablation study on the 3RScan dataset. CR refers to the usage of summary statistic priors for confidence rescoring via the CR module during inference. The Best results are shown in bold." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.406, + 0.907, + 0.572 + ], + "angle": 0, + "content": "of a true overfitting to the dataset. Interestingly, the removal of the CR module during inference does not significantly reduce performance. It even leads to a slight increase in predicate Recall and overall relationship triplet Recall. The slight increase in predicate Recall, coupled with a drastic decrease in predicate mRecall, suggests that removing the CR module benefits the predicate classes that appear frequently in the dataset, including the None class, with the trade off of decreasing performance on the tail classes. Nonetheless, the use of the CR module during evaluation benefits object Recall, object mRecall, and predicate mRecall." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.906, + 0.688 + ], + "angle": 0, + "content": "Limitations. The effectiveness of initial feature aggregation relies on the quality of the segmentation mask. Although the CR module performs well when there is no significant domain shift, the statistical prior can negatively impact performance when applied to datasets with a drastically different training data distribution. Similarly, the prior becomes less useful on datasets with very limited data." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.708, + 0.634, + 0.723 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We introduce a straightforward yet effective approach for enhancing 3D semantic scene graph estimation through our MFI, RSN-GNN, and CR modules. Specifically, our masked feature initialization (MFI) enhances node features by filtering out background distractors and leveraging more than just part-based image features. The RSN-GNN module strengthens edge features by integrating highly activated neighboring features into the target node feature, improving robustness. Finally, we introduce a refinement module that explicitly refines predictions using statistical prior knowledge, further enhancing performance." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25006" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.228 + ], + "angle": 0, + "content": "This research work is supported by the Agency for Science, Technology and Research (A\\*STAR) under its MTC Programmatic Funds (Grant No. M23L7b0021), and the Tier 2 grant MOE-T2EP20124-0015 from the Singapore Ministry of Education. We also thank Shun-Cheng Wu for the helpful discussions on running the SLAM code and reproducing the results in the 3DSSG repository." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.241, + 0.188, + 0.256 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.266, + 0.484, + 0.336 + ], + "angle": 0, + "content": "[1] Iro Armeni, Zhi-Yang He, JunYoung Gwak, Amir R Zamir, Martin Fischer, Jitendra Malik, and Silvio Savarese. 3d scene graph: A structure for unified semantics, 3d space, and camera. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5664-5673, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.338, + 0.484, + 0.42 + ], + "angle": 0, + "content": "[2] Stephan Baier, Yunpu Ma, and Volker Tresp. Improving visual relationship detection using semantic modeling of scene descriptions. In The Semantic Web-ISWC 2017: 16th International Semantic Web Conference, Vienna, Austria, October 21-25, 2017, Proceedings, Part I 16, pages 53-68. Springer, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.422, + 0.484, + 0.492 + ], + "angle": 0, + "content": "[3] Lichang Chen, Guosheng Lin, Shijie Wang, and Qingyao Wu. Graph edit distance reward: Learning to edit scene graph. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX 16, pages 539-554. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.493, + 0.484, + 0.56 + ], + "angle": 0, + "content": "[4] Tianshui Chen, Weihao Yu, Riquan Chen, and Liang Lin. Knowledge-embedded routing network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6163–6171, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.564, + 0.484, + 0.632 + ], + "angle": 0, + "content": "[5] Helisa Dhamo, Azade Farshad, Iro Laina, Nassir Navab, Gregory D Hager, Federico Tombari, and Christian Rupprecht. Semantic image manipulation using scene graphs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5213-5222, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.634, + 0.484, + 0.702 + ], + "angle": 0, + "content": "[6] Helisa Dhamo, Fabian Manhardt, Nassir Navab, and Federico Tombari. Graph-to-3d: End-to-end generation and manipulation of 3d scenes using scene graphs. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16352–16361, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.484, + 0.787 + ], + "angle": 0, + "content": "[7] Jessica D'souza, PK Aleema, S Dhanyashree, Clita Fernandes, KM Kavitha, and Chandra Naik. Knowledge-based scene graph generation in medical field. In 2023 IEEE International Conference on Distributed Computing, VLSI, Electrical Circuits and Robotics (DISCOVER), pages 232-237. IEEE, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.789, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[8] Mingtao Feng, Haoran Hou, Liang Zhang, Zijie Wu, Yulan Guo, and Ajmal Mian. 3d spatial multimodal knowledge accumulation for scene graph prediction in point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9182-9191, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.86, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[9] Gege Gao, Weiyang Liu, Anpei Chen, Andreas Geiger, and Bernhard Scholkopf. Graphdreamer: Compositional 3d scene synthesis from scene graphs. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.266, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21295-21304, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[10] Qiao Gu, Ali Kuwajerwala, Sacha Morin, Krishna Murthy Jatavallabhula, Bipasha Sen, Aditya Agarwal, Corban Rivera, William Paul, Kirsty Ellis, Rama Chellappa, et al. Conceptgraphs: Open-vocabulary 3d scene graphs for perception and planning. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 5021-5028. IEEE, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.277 + ], + "angle": 0, + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[12] Felix Holm, Ghazal Ghazaei, Tobias Czempiel, Ege Özsoy, Stefan Saur, and Nassir Navab. Dynamic scene graph representation for surgical video. In Proceedings of the IEEE/CVF international conference on computer vision, pages 81-87, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[13] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3668–3678, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.49 + ], + "angle": 0, + "content": "[14] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[15] Chen Lin, Shuai Zheng, Zhizhe Liu, Youru Li, Zhenfeng Zhu, and Yao Zhao. Sgt: Scene graph-guided transformer for surgical report generation. In International conference on medical image computing and computer-assisted intervention, pages 507-518. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[16] Xin Lin, Changxing Ding, Jing Zhang, Yibing Zhan, and Dacheng Tao. Ru-net: Regularized unrolling network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19466, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.689 + ], + "angle": 0, + "content": "[17] Kien Nguyen, Subarna Tripathi, Bang Du, Tanaya Guha, and Truong Q Nguyen. In defense of scene graphs for image captioning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1407-1416, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.69, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[18] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research Journal, pages 1-31, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[19] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2856-2865, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[20] Ege Özsoy, Evin Pinar Örnek, Ulrich Eck, Tobias Czempiel, Federico Tombari, and Nassir Navab. 4d-or: Semantic scene graphs for or domain modeling. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 475–485. Springer, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25007" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[21] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[22] Brigit Schroeder and Subarna Tripathi. Structured query-based image retrieval using scene graphs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 178-179, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[23] Sebastian Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In Proceedings of the fourth workshop on vision and language, pages 70–80, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[24] Sahand Sharifzadeh, Sina Moayed Baharlou, and Volker Tresp. Classification by attention: Scene graph classification with prior knowledge. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5025-5033, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.39 + ], + "angle": 0, + "content": "[25] Motoharu Sonogashira, Masaaki Iiyama, and Yasutomo Kawanishi. Towards open-set scene graph generation with unknown objects. IEEE Access, 10:11574-11583, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[26] Keisuke Tateno, Federico Tombari, and Nassir Navab. Real-time and scalable incremental segmentation on dense slam. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4465-4472. IEEE, 2015. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.518 + ], + "angle": 0, + "content": "[27] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.52, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[28] Johanna Wald, Helisa Dhamo, Nassir Navab, and Federico Tombari. Learning 3d semantic scene graphs from 3d indoor reconstructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3961-3970, 2020. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[29] Johanna Wald, Nassir Navab, and Federico Tombari. Learning 3d semantic scene graphs with instance embeddings. International Journal of Computer Vision, 130(3):630-651, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[30] Ziqin Wang, Bowen Cheng, Lichen Zhao, Dong Xu, Yang Tang, and Lu Sheng. Vl-sat: Visual-linguistic semantics assisted training for 3d semantic scene graph prediction in point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21560–21569, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[31] Shun-Cheng Wu, Johanna Wald, Keisuke Tateno, Nassir Navab, and Federico Tombari. Scenographfusion: Incremental 3d scene graph prediction from rgb-d sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7515-7525, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[32] Shun-Cheng Wu, Keisuke Tateno, Nassir Navab, and Federico Tombari. Incremental 3d semantic scene graph prediction from rgb sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5064-5074, 2023. 1, 2, 3, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[33] Danfei Xu, Yuke Zhu, Christopher B Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In Pro" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "ceedings of the IEEE conference on computer vision and pattern recognition, pages 5410-5419, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[34] Ning Xu, An-An Liu, Jing Liu, Weizhi Nie, and Yuting Su. Scene graph captioner: Image captioning based on structural visual representation. Journal of Visual Communication and Image Representation, 58:477-485, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[35] Jianwei Yang, Jiasen Lu, Stefan Lee, Dhruv Batra, and Devi Parikh. Graph r-cnn for scene graph generation. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.907, + 0.304 + ], + "angle": 0, + "content": "[36] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[37] Xu Yang, Kaihua Tang, Hanwang Zhang, and Jianfei Cai. Auto-encoding scene graphs for image captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10685–10694, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[38] Sangwoong Yoon, Woo Young Kang, Sungwook Jeon, SeongEun Lee, Changjin Han, Jonghun Park, and Eun-Sol Kim. Image-to-image retrieval by learning similarity between scene graphs. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10718-10726, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.501 + ], + "angle": 0, + "content": "[39] Kun Yuan, Manasi Kattel, Joel L Lavanchy, Nassir Navab, Vinkle Srivastav, and Nicolas Padoy. Advancing surgical vqa with scene graph knowledge. International Journal of Computer Assisted Radiology and Surgery, 19(7):1409-1417, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.503, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[40] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural motifs: Scene graph parsing with global context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5831-5840, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.56, + 0.905, + 0.627 + ], + "angle": 0, + "content": "[41] Chaoyi Zhang, Jianhui Yu, Yang Song, and Weidong Cai. Exploiting edge-oriented reasoning for 3d point-based scene graph analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9705-9715, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.63, + 0.905, + 0.684 + ], + "angle": 0, + "content": "[42] Shoulong Zhang, Aimin Hao, Hong Qin, et al. Knowledge-inspired 3d scene graph prediction in point cloud. Advances in Neural Information Processing Systems, 34:18620-18632, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.687, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[43] Zhongping Zhang, Huiwen He, Bryan A Plummer, Zhenyu Liao, and Huayan Wang. Complex scene image editing by scene graph comprehension. In BMVC, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.729 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25008" + } + ] +] \ No newline at end of file diff --git a/2025/Statistical Confidence Rescoring for Robust 3D Scene Graph Generation from Multi-View Images/00c53575-a21f-463b-8889-7a2370d2bff7_origin.pdf b/2025/Statistical Confidence Rescoring for Robust 3D Scene Graph Generation from Multi-View Images/00c53575-a21f-463b-8889-7a2370d2bff7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..93380436bb93697a850bc92ad9ccd63d8294d9b9 --- /dev/null +++ b/2025/Statistical Confidence Rescoring for Robust 3D Scene Graph Generation from Multi-View Images/00c53575-a21f-463b-8889-7a2370d2bff7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd29e97c145c018025874c268ad10c4dac4f0f3f1762b6a30e2dfbe3aaecbc5b +size 1217618 diff --git a/2025/Statistical Confidence Rescoring for Robust 3D Scene Graph Generation from Multi-View Images/full.md b/2025/Statistical Confidence Rescoring for Robust 3D Scene Graph Generation from Multi-View Images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5fbbe31f508dcc83ee0360384b40953ca5492bbc --- /dev/null +++ b/2025/Statistical Confidence Rescoring for Robust 3D Scene Graph Generation from Multi-View Images/full.md @@ -0,0 +1,321 @@ +# Statistical Confidence Rescoring for Robust 3D Scene Graph Generation from Multi-View Images + +Qi Xun Yeo Yanyan Li Gim Hee Lee + +Department of Computer Science, National University of Singapore + +{qixunyeo,yan.li,gimhee.lee}@nus.edu.sg + +# Abstract + +Modern 3D semantic scene graph estimation methods utilize ground truth 3D annotations to accurately predict target objects, predicates, and relationships. In the absence of given 3D ground truth representations, we explore leveraging only multi-view RGB images to tackle this task. To attain robust features for accurate scene graph estimation, we must overcome the noisy reconstructed pseudo point-based geometry from predicted depth maps and reduce the amount of background noise present in multi-view image features. The key is to enrich node and edge features with accurate semantic and spatial information and through neighboring relations. We obtain semantic masks to guide feature aggregation to filter background features and design a novel method to incorporate neighboring node information to aid robustness of our scene graph estimates. Furthermore, we leverage on explicit statistical priors calculated from the training summary statistics to refine node and edge predictions based on their one-hop neighborhood. Our experiments show that our method outperforms current methods purely using multi-view images as the initial input. Our project page is available at https://qixun1.github.io/projects/SCRSSG. + +# 1. Introduction + +The semantic scene graph (SSG) is a crucial intermediate representation that enhances higher-level scene understanding. It plays a key role in tasks such as image captioning [17, 34, 37], image retrieval [13, 22, 23, 38], image editing [3, 5, 43], and medical applications [7, 12, 15, 20, 39] by capturing both semantic and, more importantly, relational information between objects and their surroundings. Initially developed for 2D images, SSG has since expanded to the 3D domain [6, 9, 19]. 3D scene graphs provide a high-level representation of an entire 3D scene using inputs such as multi-view RGB images or LiDAR point clouds. Unlike their 2D counterparts, they incorporate spatial relation + +![](images/10f9bb1c1d968dcecc34a9179ce6fe0f4eb8083b7c106fca028feadbbe5c4a55.jpg) + +![](images/6cf51e49573ec2724c7423f0cc28af1d864f451b002fcf1e2f6d3975a3199906.jpg) +Figure 1. Existing multi-view RGB methods aggregate features using bounding boxes, introducing background noise that hinders accuracy. We use pretrained segmentation masks to reduce this interference and thus improving prediction accuracy. + +ships that extend beyond the visible image plane, allowing a holistic understanding of complex environments. + +Many approaches [1, 28, 29] rely on ground-truth 3D LiDAR point clouds as input. However, LiDAR sensors are resource-intensive and lack inherent semantic richness. Recent works [31, 32] have shifted toward using semantically rich multi-view RGB images, which offer a balance between computational efficiency and improved semantic fidelity. Although point cloud-based methods remain valuable for capturing geometric information, our approach focuses on leveraging multi-view RGB images. By incorporating pretrained depth estimators, we generate pseudo point-based geometry to estimate the 3D semantic scene graph, reducing reliance on LiDAR while preserving spatial and structural details. + +A key challenge in using only multi-view RGB images is ensuring that node features remain free from distractors. As shown in Fig. 1, prior approaches that rely on entity detectors to extract 2D bounding box proposals or regions of interest often fail to guarantee robustness since distrac + +tors within the bounding box are often mistakenly aggregated. Ensuring robust initial multi-view features is crucial without ground truth point-based geometry for error correction. No explicit mechanism exists to correct incorrectly bounded objects or misclassified rare classes that are not well trained on the model. To alleviate this problem, a mechanism should exist to refine estimates using prior knowledge instead of depending solely on implicit interactions between nodes and edges to achieve confident and accurate predictions. + +In this paper, we propose a framework to enhance the robustness of the model. First, we introduce a masked feature initialization (MFI), which leverages segmentation masks to aggregate image features instead of relying on the bounding box proposals. This reduces background noise, which results in cleaner multi-view image features. Next, we design a robust residual spatial neighbor graph neural network (RSN-GNN) to encode spatial information into node features. This network filters highly activated regions from neighboring nodes, refining target node features for improved predicate estimation. Finally, we propose a confidence rescoring (CR) module, which refines object and predicate estimates using an inverse softmax-weighted contribution of neighboring node-to-node and node-to-edge co-occurrence counts. By integrating this explicit inductive bias, our approach improves the accuracy of the prediction, particularly in low-confidence scenarios. Extensive experimental results on the benchmark 3RScan dataset show the competitiveness of our proposed approach compared to existing 2D and 3D approaches. + +Our main contributions can be summarized as follows: + +- We introduce a masked feature initialization to enhance the robustness of node features by reducing background distractors, yielding cleaner multi-view image features. +- We design a novel GCN architecture that integrates highly activated neighboring features into the target node to increase the robustness of the edge features. +- We propose a new refinement module to explicitly refine predictions based on statistical prior knowledge. +- Experiments show that we outperform previous state-of-the-art approaches on the 3RScan dataset, particularly on metrics that deal with low-tail imbalanced classes. + +# 2. Related Work + +2D Semantic Scene Graph. 2D semantic scene graph prediction is typically categorized into two-step and one-step approaches. The two-step approach first detects objects and then classifies their relationships [2, 4, 35, 40]. In contrast, the one-step approach jointly infers object and relationship classes [33]. Xu et al. [33] pioneers the problem of scene graph generation and they tried to solve this problem via iterative message passing. Baier et al. [2] first showed + +how semantic models can be improved by incorporating triplet frequencies. Zellers et al. [40] analyzes the usefulness of statistical co-occurrences for the Visual Genome dataset and concluded that such statistical priors serves as strong regularization for the task. Chen et al. [4] formalizes the first approach (KERN) to incorporate the statistical prior directly into graph neural networks. Sharifzadeh et. al. proposes Schemata to assimilate image-based relational prior knowledge into the representations within the neural network [24]. Compared to previous methods, our method combines predictions from the prediction head with statistical co-occurrence of the current node and neighbors to determine the final class using a confidence score. Unlike Sonogashira et. al. [25] which uses hard thresholding, our approach uses adaptive relationship and node refinements. + +3D Semantic Scene Graph. Methods for generating a 3D semantic scene graph can be categorized into two types based on input data. The first relies on 3D inputs such as ground-truth 3D geometry to learn contextual information [28, 30, 41]. The second relies on multi-view RGB images with some methods incorporating depth information [31] while others do not [32]. Our approach follows the latter, leveraging the estimated depth from off-the-shelf depth estimators [36] to enhance the information provided by multi-view RGB images. Wald et al. [28] first proposed 3RScan: a richly annotated 3D scene graph dataset and a method that focuses on pairwise relationships to predict the 3D scene graphs. Wu et al. [31] uses SLAM to obtain dense point representations and a novel graph convolutional network (GCN) based aggregation function to enhance 3D scene graph prediction. Zhang et al. [41] proposes an edge-oriented GCN to incorporate multi-dimensional edge features for explicit inter-node relationship modeling. Zhang et al. [42] integrates prior commonsense knowledge by learning meta-embeddings only from the class labels with their graphical structures to avoid perceptual errors. Wang et al. [30] incorporated additional information by distilling knowledge learned by the multi-modal oracle model to the 3D model. Wu et al. [32] introduces a novel entity association method to obtain the 3D entities from the predicted 2D entities and a geometric gate to incorporate geometric information to multi-view image features. Feng et al. [8] uses ConceptNet external knowledge base to accumulate both contextualized visual content and textual facts to form a 3D spatial multimodal knowledge graph. Our method leverages on the explicit statistical priors similar to 2D scene graph method KERN and Schemata while guiding low confidence predictions made by the initial node and edge predictors using both visual and geometric cues. + +![](images/b8012a394b2241f11960424a8067920d6cd7ff88f68ec8065dd998a40112353f.jpg) +Figure 2. The architecture of our proposed framework consists of three main components: 1) A masked feature initialization step leverages backprojected semantic masks to aggregate multi-view image, geometric, and spatial features. 2) A message passing step utilizes our residual spatial neighbor GNN module to refine node and edge features through feature interaction. 3) A confidence rescoring step adjusts logits based on summary statistics of object and predicate co-occurrence from the training dataset. + +# 3. Our Methodology + +Problem Definition. Given either a set of 3D point clouds $P$ or multi-view RGB images $I$ of a 3D scene as input, the goal is to estimate the 3D scene graph denoted as: + +$$ +G = (O, R), \tag {1} +$$ + +# where + +- $N$ : Number of multi-view input images; +- $O = \{o_m\}_{m=1}^M$ : All objects in the scene; +- $R = \{r_{k\to m}\}_{m,k = 1}^{M,M}$ : Relationships between all objects; +- $M = |O|$ : Number of objects in the scene. + +Furthermore, we refer to $o_m$ as the object class for node instance $m$ , $r_{k \to m}$ as the predicate class for the edge connecting nodes $k$ and $m$ . Alternatively, the 3D scene graph can be defined as a set of triplets given by: $\{o_k, r_{k \to m}, o_m\} := \{node, predicate, node\}$ . + +Overview. Fig. 2 shows an illustration of the architecture of our proposed framework. The input to our framework is a set of $N$ multi-view RGB images $\{I_i\}_{i=1}^N$ of the 3D scene. Since the focus of our work is not on multi-view 3D reconstruction, we recover the 3D point cloud of the scene by running an off-the-shelf RGB-D SLAM system [26] with the metric depth maps of the input images $\{\hat{D}_i\}_{i=1}^N$ obtained from the pretrained Depth Anything [36] model. + +From the predicted 3D point cloud, we initialize multiview image features using segmentation masks, extract geometric features with a point-based feature extractor, and derive spatial features via a linear layer. These multi-view image features serve as the initial input to our RSN-GNN module. At the GCN initialization, we enhance node features by integrating geometric and spatial features through geometric and spatial gates. We additionally enhance edge features by integrating max pooled neighboring nodes features. These enhanced node and edge features undergo message passing, with the integration process repeating at each + +layer. After $k$ iterations, the final node and edge features are processed by predictors to generate initial estimates, which are then refined using precomputed co-occurrence statistics via our CR module, producing the final scene graph output. + +We elaborate on the feature initialization step in Sec. 3.1, the message passing step in Sec. 3.2, and the confidence rescoring step in Sec. 3.3. + +# 3.1. Masked Feature Initialization (MFI) + +We initialize the node features of the scene graph by combining multi-view image representations from a pretrained image encoder with PointNet features derived from the predicted points in the RGB-D SLAM reconstruction step that utilizes predicted depth. To eliminate background clutter, we employ the Segment Anything Model (SAM) [14] to produce precise 2D semantic masks for each object. For each node, we backproject the associated points onto the 2D instance mask to identify its corresponding region. We then generate object proposals from these instance masks, resize them to $224 \times 224$ , and extract their multi-view image features using the pretrained encoder. Finally, we apply these semantic masks to the multi-view image features before aggregating them into our node representations. + +Our method to get the object proposals, i.e. the nodes of the scene graph offers two key advantages over relying solely on the entity detector from SceneGraphFusion and JointSSG [31, 32]. First, the entity detector often captures only part of an object. Although SceneGraphFusion propagates the label for this segment, the feature extractor can operate only on the detected portion. In contrast, our method employs more complete object masks, enabling the feature extractor to derive features from the entire object. Second, we exclude extraneous background features by applying these masks before feature aggregation. In SceneGraphFusion and JointSSG, the bounding boxes of the pretrained entity detector frequently include background regions, which can dilute the quality of the features and hinder + +the performance. Our method is closer to the open vocabulary method ConceptGraphs [10] which uses SAM masks but with the aggregation method provided by JointSSG. + +Formally, we compute a corresponding multi-view image feature $\mathbf{v}_{i,img}$ , a geometric feature $\mathbf{v}_{i,geo}$ , and a spatial feature $\mathbf{v}_{i,spat}$ for each node $v_{i}$ . We aggregate features only for nodes that are visible in the可视ability graph $G_{v}$ denoted by $\mathbf{1}(v_{i} \in I_{k})$ and have corresponding pixels within the semantic mask provided by SAM. The aggregated multi-view image feature $\mathbf{v}_{i,img}$ is given by: + +$$ +\mathbf {v} _ {i} ^ {0} = \frac {\sum_ {k = 1} ^ {N} \mathbf {1} \left(v _ {i} \in I _ {k}\right) f \left(\eta_ {k} \odot x _ {i , k}\right)}{\left| k \in G _ {v} \right|}, \tag {2} +$$ + +which also serves as the initial node feature. $f(\cdot)$ is the image feature encoder and $\eta_{k}$ refers to the 2D semantic mask. $x_{i,k}$ refers to the cropped image patch containing the node $i$ from the $k$ -th image. $\odot$ refers to the element-wise multiplication operator. The aggregation is the arithmetic mean of the node features for each image in the可视ability graph. + +The geometric feature $\mathbf{v}_{i,geo}$ is extracted using a vanilla PointNet [21], with inputs derived from the 3D instance masks generated during the RGB-D SLAM step. For the spatial features, we employ a simple linear layer to transform the attributes of each 3D bounding box: dimensions $\mathbf{b}_i$ , volume $s_i$ , length $l_i$ , and the ratio of its x- and y-axes $r_i$ into the spatial feature as: + +$$ +\mathbf {v} _ {i, s p a t} = g _ {s} ([ \mathbf {b} _ {i}, s _ {i}, l _ {i}, r _ {i} ]), \tag {3} +$$ + +where $g_{s}$ refers to a set of learnable weights for spatial features. $[\cdot]$ refers to the concatenation of the inputs within the operator. + +# 3.2. Residual Spatial Neighbor GNN Module + +Following a similar design to the learnable geometric gate in JointSSG [32], we integrate both geometric and spatial features into the initial node feature, which originally consists only of multi-view image features. The spatial features serve as additional information for individual nodes, analogous to the embeddings used for token encoding in Transformer architectures [27]. We define the geometric gate as: + +$$ +\hat {\mathbf {v}} _ {i} ^ {l} = \mathbf {v} _ {i} ^ {l} + \sigma \left(\mathbf {w} _ {g} ^ {T} \left[ \mathbf {v} _ {i} ^ {l}, \mathbf {v} _ {i, g e o} \right]\right) \cdot \sigma \left(\mathbf {v} _ {i, g e o}\right), \tag {4} +$$ + +where $\mathbf{w}_g$ refers to a set of learnable weights, $\sigma$ refers to the sigmoid operator and $l$ refers to the layer of the node feature. Furthermore, we define the spatial gate as: + +$$ +\ddot {\mathbf {v}} _ {i} ^ {l} = \hat {\mathbf {v}} _ {i} ^ {l} + \sigma \left(\mathbf {w} _ {s} ^ {T} [ \hat {\mathbf {v}} _ {i} ^ {l}, \mathbf {v} _ {i, s p a t} ]\right) \cdot \sigma \left(\mathbf {v} _ {i, s p a t}\right), \tag {5} +$$ + +where $\mathbf{w}_s$ is a set of learnable weights for spatial features. + +Inspired by the residual connections in ResNet [11], we introduce a mechanism to integrate highly activated neighboring features into the target node feature using max pooling during the message-passing process for edge features. This approach implicitly encodes neighboring information, improving edge prediction for scene graph estimation. Specifically, we augment the edge messages in the GCN by integrating max-pooled neighboring features before message passing as follows: + +$$ +\tilde {\mathbf {v}} _ {i} ^ {l} = \ddot {\mathbf {v}} _ {i} ^ {l} + \max ([ \ddot {\mathbf {v}} _ {j} ^ {l} ] _ {j \in n (i)}), \tag {6} +$$ + +where $\max (\cdot)$ refers to the max pooling operator and $n(i)$ refers to the neighboring nodes of node $i$ . Max pooling allows highly activated regions in the different neighboring node features to be fused into the target node feature via the residual connection. + +We apply max-pooled neighboring features to the calculation of edge messages without using them in node message updates. Since node features are highly susceptible to noise and spurious correlations from other nodes [16], incorporating max-pooled neighboring features into node message computations can introduce confusion instead of improving performance. The edge message $m_{i\rightarrow j}$ is thus defined as: + +$$ +m _ {i \rightarrow j} = g _ {e} ([ \tilde {\mathbf {v}} _ {i} ^ {l}, \mathbf {e} _ {i j} ^ {l}, \tilde {\mathbf {v}} _ {j} ^ {l} ]), \tag {7} +$$ + +where $g_{e}(\cdot)$ is an MLP and $\mathbf{e}_{ij}^{l}$ is the edge descriptor used in [32] containing relative node properties such as centroid displacement, relative bounding box size difference, and a relative pose descriptor. + +# 3.3. Confidence Rescoring (CR) Module + +As discussed in Sec. 1, modern scene graph methods depend heavily on the entity detector and image feature extractor to produce an accurate initial object prediction. However, there is no mechanism to correct a target bounding box prediction that is initially wrong. To address this limitation, we leverage relationships with neighboring nodes, allowing us to mitigate the impact of incorrect initial predictions by weighting contributions from neighboring node statistics to refine target object classification. Fig. 3 illustrates the logits of the final prediction after applying our confidence rescoring module. This module incorporates prior knowledge to improve prediction confidence beyond that of the original node and edge multi-layer perceptron (MLP) predictor. + +The probability of predicting node $i$ given its neighboring nodes $n(i)$ is approximated by the output of our node predictor $h_v(\cdot)$ as: + +$$ +\begin{array}{l} P \left(o _ {i} \mid \left\{o _ {j}: j \in n (i) \right\}\right) = \frac {1}{Z _ {i}} \psi_ {i} \left(o _ {i}\right) \prod_ {j \in n (i)} \psi_ {i j} \left(o _ {i}, o _ {j}\right) \\ \approx \tau \left(h _ {v} \left(\mathbf {v} ^ {L}\right)\right), \tag {8} \\ \end{array} +$$ + +where $\psi_i(o_i)$ refers to the potential associated with node $i$ and $\psi_{ij}(o_i,o_j)$ refers to the potential associated with the edge between node $i$ and its neighbor node $j$ . The potentials are implicitly learned through the GCN unlike classical methods, which explicitly learn these potentials. $h_v(\cdot)$ is a small MLP that predicts the class of each node and $\tau$ refers to the softmax operator. $L$ refers to the last layer of the GCN message passing output. The partition function given by: + +$$ +Z _ {i} = \sum_ {o _ {i}} \psi_ {i} \left(o _ {i}\right) \prod_ {j \in n (i)} \psi_ {i j} \left(o _ {i}, o _ {j}\right) \tag {9} +$$ + +ensures the probabilities sum to 1. A similar equation can be obtained for the probability of predicting edge $ij$ given its neighboring nodes $i$ and $j$ , i.e.: + +$$ +\begin{array}{l} P \left(r _ {i \rightarrow j} \mid o _ {i}, o _ {j}\right) = \frac {1}{Z _ {i j}} \psi_ {i j} \left(o _ {i}, o _ {j}, r _ {i \rightarrow j}\right) \tag {10} \\ \approx \tau (h _ {e} (\mathbf {e} ^ {L})), \\ \end{array} +$$ + +where $h_e(\cdot)$ is a small MLP which predicts the class of each edge, and the partition function to normalize the probabilities is given by: + +$$ +Z _ {i j} = \sum_ {r _ {i \rightarrow j}} \psi_ {i j} \left(o _ {i}, o _ {j}, r _ {i \rightarrow j}\right). \tag {11} +$$ + +We begin by computing summary statistics from the training dataset to node-to-node co-occurrences denoted as $c_{ij}(o_i, o_j)$ , and node-to-edge co-occurrences denoted as $d_{ij}(o_i, e_j)$ . Subsequently, we determine the prediction confidence of the model, where $\alpha_{\mathbf{v}}$ represents node confidence and $\alpha_{\mathbf{e}}$ represents edge confidence. Next, we obtain the node $\alpha_{\mathbf{v}}$ and edge $\alpha_{\mathbf{e}}$ confidences as: + +$$ +\alpha_ {\mathbf {v}} = \max _ {\mathbf {v}} \tau \left(h _ {v} \left(\mathbf {v} ^ {L}\right)\right), \quad \alpha_ {\mathbf {e}} = \max _ {\mathbf {e}} \tau \left(h _ {e} \left(\mathbf {e} ^ {L}\right)\right). \tag {12} +$$ + +We first estimate the marginal probability of a specific instantiation of node $i$ , conditioned on node $j$ using node-to-node co-occurrences, and on edge $ij$ using node-to-edge co-occurrences defined as: + +$$ +\mathbf {c} \left(o _ {i} \mid o _ {j}\right) = \left[ c \left(o _ {i} = c _ {k} \mid o _ {j} = c _ {l}\right) \right] _ {k = 0, l = 0} ^ {C, C}, \quad \text {w h e r e} +$$ + +$$ +c \left(o _ {i} = c _ {k} \mid o _ {j} = c _ {l}\right) = \frac {c \left(o _ {i} = c _ {k} , o _ {j} = c _ {l}\right)}{\sum_ {k = 0} ^ {C} c \left(o _ {i} = c _ {k} , o _ {j} = c _ {l}\right)}. \tag {13} +$$ + +Here, $C$ refers to the number of object classes and $c_{k}$ refers to the $k$ -th class in the set of object classes. We then integrate the summary statistics via an inverse softmax operation denoted as $\gamma(\cdot)$ into the node probability in Eq. 14 to get: + +$$ +\begin{array}{l} P \left(o _ {i} \mid \left\{o _ {j}: j \in n (i) \right\}\right) = \tau \left(\alpha_ {\mathbf {v}} \cdot h _ {v} (\mathbf {v} _ {i}) \right. \\ + \left(1 - \alpha_ {\mathbf {v}}\right) \cdot \sum_ {j \in n (i)} \alpha_ {j} \cdot \gamma (\mathbf {c} \left(o _ {i} \mid o _ {j}\right)). \tag {14} \\ \end{array} +$$ + +This formulation is chosen because directly adding the estimated marginal probability to the predictions of the model after softmax would disregard the required normalization, resulting in incompatible terms. + +Similarly, we compute the estimated marginal probability of an edge $ij$ instantiation conditioned on nodes $i$ and $j$ using node-to-edge co-occurrence statistics for predicate estimation: + +$$ +\mathbf {d} \left(r _ {i \rightarrow j} \mid o _ {j}\right) = \left[ d \left(r _ {i \rightarrow j} = e _ {k} \mid o _ {j} = c _ {l}\right)\right] _ {k = 0, l = 0} ^ {P, C}, +$$ + +$$ +d \left(r _ {i \rightarrow j = k} \mid o _ {j = l}\right) = \frac {d \left(r _ {i \rightarrow j} = e _ {k} , o _ {j} = c _ {l}\right)}{\sum_ {k = 0} ^ {P} d \left(r _ {i \rightarrow j} = e _ {k} , o _ {j} = c _ {l}\right)}. \tag {15} +$$ + +Here, $P$ refers to the total number of predicate classes and $e_k$ refers to the $k$ -th class in the set of predicate class. The refined logits for predicate estimation are then given by: + +$$ +\begin{array}{l} P \left(r _ {i \rightarrow j} \mid o _ {i}, o _ {j}\right) = \tau \left(\alpha_ {i j} \cdot h _ {e} (\mathbf {e} _ {i j}) \right. \\ + \left(1 - \alpha_ {i j}\right) \cdot \left(\alpha_ {i} \cdot \gamma (\mathbf {d} \left(r _ {i \rightarrow j} \mid o _ {i}\right))\right) \tag {16} \\ \cdot \left(\alpha_ {j} \cdot \gamma (\mathbf {d} (r _ {i \rightarrow j} | o _ {j})))\right). \\ \end{array} +$$ + +Remarks. As shown in Fig. 3, the original node probability for the correct class is already relatively high. However, our CR module further enhances the predicted probability as evidenced by the upward shift in the box plot from "Original" to "After CR" in the red boxes, with a notable increase in the lower quartile. Moreover, a larger proportion of low node probability instances are flagged as outliers as shown by the greater number of red circles at "After CR" compared to "Original". Most notable gains occur when original logits are low since CR increases the probability of the correct class by down-weighting misclassified classes $c'$ via lower conditional probabilities $P(c'|c)$ given conditioned class $c$ . This targeted adjustment raises suppressed logits for rare classes, thus improving prediction performance for underrepresented categories. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. We evaluate on the 3RScan dataset following the same training and test split from [28]. The image quality of the dataset is rather poor as each scene constitutes a variable number of short videos spliced together with different motion trajectories. As the images are obtained from consumer-level cameras, the images possess a low frame rate of $10\mathrm{Hz}$ , image blur, and image jitter from sudden camera motion which makes this dataset challenging. + +Implementation Details. We utilize the codebase provided from the 3DSSG GitHub repository $^2$ to reproduce the prior + +
Methodbath.bedbkshfcab.chaircntr.curt.deskdoorfloorofurnpic.refri.show.sinksofatabletoil.wallwind.Mean.
IMP0.0001.0000.0000.3410.4670.0000.6060.2000.5280.9000.1250.1430.0000.0000.3330.4500.5250.8000.6220.1250.358
VGfM0.5000.0000.0000.4150.5430.0830.5450.0000.3890.9200.2320.0860.0000.0000.3810.4500.5740.8000.6440.1670.336
3DSSG0.0000.6670.0000.2070.3480.1670.5760.2000.3610.7800.1250.0570.0000.0000.3330.5000.2790.0000.4670.1250.260
SGFN0.5000.3330.0000.2930.6850.2500.5150.2000.4440.8800.2320.1710.0000.0000.4290.5000.5080.4000.7070.1250.359
JointSSG1.0001.0000.0000.4760.6740.2500.5760.2000.5000.9000.1070.1430.0000.0000.4290.5500.5410.8000.6800.2500.454
Ours1.0001.0000.3330.6220.6300.3330.5450.6000.5280.8800.3040.1710.6670.6670.3330.6000.6230.8000.6930.3750.605
+ +Table 1. Comparison with state-of-the-art methods on the 3RScan dataset for each object class. The Best results are highlighted. + +![](images/b5f5e82be139d5c212c4fc229a7d5c51692abb22563db3642c14ebd2d4ee705b.jpg) +Figure 3. Box plot for predicted node probability for correct class after our CR module. The lower quartile visibly moves upwards due to an increase in predicted node probability for correct class. + +methods. For JointSSG[32], we replace their image feature extractor from a pretrained ResNet18 to a pretrained DINOv2 with a ViT-L backbone to ensure fair comparison. + +Following [32], we use PointNet [21] as the point encoder and employ two message-passing layers. For RGB-D SLAM reconstruction, we utilize the method from [26] and our predicted depth to generate predicted points. Multiview feature extraction is performed using a pretrained DINOv2 [18] with a ViT-L backbone. To obtain object segmentation masks, we use a pretrained SAM predictor with a text prompt containing the class set of 20 NYU object classes following the subset defined in [28]. Node-to-node and node-to-edge co-occurrences are precomputed on the training set. We train the model on a single RTX 3090 GPU for two days with early stopping. + +Evaluation Metrics. Following [29] and JointSSG[32], we report the overall top-1 recall (Recall) for the object class estimation (Obj.), the predicate estimation (Pred.), and the relationship triplet estimation (Rel.). We also report the mean recall (mRecall) for the object class estimation (Obj.), the predicate estimation (Pred.) only. We map all predictions on the estimated segments to the ground truth segments for a fair comparison between the different methods. + +# 4.2. Quantitative Results + +We present the per-class recall for 20 object classes in the 3RScan dataset in Tab. 1, demonstrating that our method significantly outperforms others across most object cate + +gories. Notably, we are the only approach that achieves a recall above $0\%$ for all object classes, indicating that we robustly learn all classes instead of just a subset. Our method tends to underperform slightly compared to the other methods on majority classes such as floor and wall classes in return, although not by a significant margin. The main experimental results for the top-1 recall and the mean recall are shown in Tab. 2. Importantly, our approach significantly outperforms others in mean recall (mRecall) for both objects and predicates, suggesting that it handles class imbalance issues [29] more effectively than competing methods. Additionally, our method surpasses all baselines in relationship and object estimation metrics while achieving comparable performance in predicate estimation. More impressively, our method with the weaker ResNet18 multiview image features already surpasses all baselines in relationship and object estimation metrics except for predicate estimation. Our method with ResNet18 image features even manages to surpass the performance of JointSSG with DINOv2 multi-view image features by non-trivial margins. This shows that our improved performance is not purely due to the more powerful DINOv2 features. Refer to our supplementary material for more quantitative results. + +# 4.3. Qualitative Results + +We show qualitative results comparing our method with other competitive methods in Fig. 4. Since the scene contains a large number of objects, we display only a subset with fewer objects and predicates for better readability. The correct prediction of the None predicate class is omitted in the visualization due to its prevalence in the dataset. As seen in Fig. 4, our method can distinguish the relatively more ambiguous wall which the other methods fail to distinguish. The shower curtain is tricky to classify since it is a subset of a curtain that is another class of objects present in the dataset. Understandably, all methods fail to correctly classify this object. Both our method and JointSSG attain good performance on the predicate class compared to SGFN as both methods only misclassify the predicate between the curtain object and the wall. + +We additionally evaluate on the entire scene for scan 43b8cae1 without comparison to other methods. Our method can recognize rare classes present in the scene, which are more challenging to learn. As seen in Fig. 5, our approach successfully classifies the rare class bookshelf instance denoted by $bkshf$ which no other method managed + +![](images/5936e487f036635e02696b434b01b76f7101e3528d0de93e1a3f253bae481cbf.jpg) +Figure 4. The qualitative results on 3RScan dataset of previous works and our proposed framework on scan 4d3d82b0. SGFN also fails to classify majority of non-background object classes except the sink and cabinet. It also fails to classify all predicate classes. JointSSG performs better but misclassifies the shower curtain as a curtain and the ambiguous blue wall as a cabinet. Our method correctly classifies the wall but misidentifies the shower curtain as a curtain. + +![](images/0fc87d00a72f2183bee837af9985e58af734d76da85d780d9600e08a3b7dd392.jpg) +Figure 5. The qualitative results on 3RScan dataset of our proposed framework on the entire scene of scan 43b8cae1. Our method can correctly classify most objects in the scene except for the light blue dress table at the top right of the scan. Our method can also correctly classify most predicates. + +
MethodRecall%mRecall%
RelObj.Pred.Obj.Pred
IMP [32]25.851.890.430.023.0
VGfM [32]28.353.390.731.624.4
3DSSG [32]17.541.488.231.926.6
SGFN [32]31.456.789.638.330.5
JointSSG [32]34.158.189.943.033.3
IMP25.447.789.935.819.9
VGfM27.450.190.633.622.3
3DSSG13.035.686.826.024.3
SGFN29.952.289.535.927.1
JointSSG32.954.588.745.435.2
JointSSG‡34.356.686.552.936.2
Ours†38.659.389.157.436.8
Ours40.561.890.460.539.2
+ +Table 2. Comparison with state-of-the-art methods on the 3RScan dataset with 20 object classes and 8 predicate classes. The top group of results are reported in JointSSG [32]. The middle group of results are reproduced via 3DSSG GitHub repository. JointSSG‡ refers to JointSSG but with DINOv2 multi-view image features. Ours† uses ResNet18 multi-view image features; Ours uses the DINOv2 multi-view image features instead. The Best and Second Best results are highlighted, respectively. + +to predict based on the results shown in Tab. 1. Our method also predicts multiple reasonable predicates of standing on between various objects and floor. Additional qualitative results can be found in the supplementary material. + +# 4.4. Ablation Studies + +We analyze the effects of our feature initialization, robust neighbor residuals, and confidence rescoring module in Tab. 3. Our confidence rescoring (CR) module significantly improves performance across all metrics, indicating that it effectively leverages statistical priors to address class imbalance issues as noted in [32]. Apart from predicate mRecall, incorporating our masked feature initialization further improves performance across all metrics. This outcome is expected since masking primarily aims to refine multi-view image features, which do not directly contribute to edge feature improvements. The combination of the masked feature initialization and the CR module act as a force multiplier since the CR module benefits from cleaner and more accurate neighboring node features. Additionally, our robust neighbor residual gate maintains comparable performance across most metrics except for predicate mRecall, which benefits from additional neighbor information to better handle class imbalance in predicate estimation. + +We additionally study the effects of using our CR module during inference in Tab. 4. The addition of the CR module with the prior knowledge of the dataset is akin to calculating statistics during transductive setting of inference instead + +
MethodConf.Feat.Neigh.Recall%mRecall%
Rel.Obj.Pred.Obj.Pred.
JointSSG×××32.954.588.245.435.2
+ Conf.✓××37.957.289.654.136.3
+ Feat.✓✓×41.161.390.861.335.8
Ours✓✓✓40.561.890.460.539.2
+ +Table 3. Ablation study on the 3RScan dataset. Conf. refers to the usage of summary statistic priors for confidence rescoring via the CR module. Feat. refers to the usage of DINOv2 image features with refinement from SAM semantic masks via MFI. Neigh. refers to the usage of RSN-GNN. The Best results are shown in bold. + +
MethodCR.Recall%mRecall%
Rel.Obj.Pred.Obj.Pred.
Ours×41.161.190.660.235.7
Ours✓40.561.890.460.539.2
+ +Table 4. Ablation study on the 3RScan dataset. CR refers to the usage of summary statistic priors for confidence rescoring via the CR module during inference. The Best results are shown in bold. + +of a true overfitting to the dataset. Interestingly, the removal of the CR module during inference does not significantly reduce performance. It even leads to a slight increase in predicate Recall and overall relationship triplet Recall. The slight increase in predicate Recall, coupled with a drastic decrease in predicate mRecall, suggests that removing the CR module benefits the predicate classes that appear frequently in the dataset, including the None class, with the trade off of decreasing performance on the tail classes. Nonetheless, the use of the CR module during evaluation benefits object Recall, object mRecall, and predicate mRecall. + +Limitations. The effectiveness of initial feature aggregation relies on the quality of the segmentation mask. Although the CR module performs well when there is no significant domain shift, the statistical prior can negatively impact performance when applied to datasets with a drastically different training data distribution. Similarly, the prior becomes less useful on datasets with very limited data. + +# 5. Conclusion + +We introduce a straightforward yet effective approach for enhancing 3D semantic scene graph estimation through our MFI, RSN-GNN, and CR modules. Specifically, our masked feature initialization (MFI) enhances node features by filtering out background distractors and leveraging more than just part-based image features. The RSN-GNN module strengthens edge features by integrating highly activated neighboring features into the target node feature, improving robustness. Finally, we introduce a refinement module that explicitly refines predictions using statistical prior knowledge, further enhancing performance. + +# Acknowledgments + +This research work is supported by the Agency for Science, Technology and Research (A\*STAR) under its MTC Programmatic Funds (Grant No. M23L7b0021), and the Tier 2 grant MOE-T2EP20124-0015 from the Singapore Ministry of Education. We also thank Shun-Cheng Wu for the helpful discussions on running the SLAM code and reproducing the results in the 3DSSG repository. + +# References + +[1] Iro Armeni, Zhi-Yang He, JunYoung Gwak, Amir R Zamir, Martin Fischer, Jitendra Malik, and Silvio Savarese. 3d scene graph: A structure for unified semantics, 3d space, and camera. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5664-5673, 2019. 1 +[2] Stephan Baier, Yunpu Ma, and Volker Tresp. Improving visual relationship detection using semantic modeling of scene descriptions. In The Semantic Web-ISWC 2017: 16th International Semantic Web Conference, Vienna, Austria, October 21-25, 2017, Proceedings, Part I 16, pages 53-68. Springer, 2017. 2 +[3] Lichang Chen, Guosheng Lin, Shijie Wang, and Qingyao Wu. Graph edit distance reward: Learning to edit scene graph. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX 16, pages 539-554. Springer, 2020. 1 +[4] Tianshui Chen, Weihao Yu, Riquan Chen, and Liang Lin. Knowledge-embedded routing network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6163–6171, 2019. 2 +[5] Helisa Dhamo, Azade Farshad, Iro Laina, Nassir Navab, Gregory D Hager, Federico Tombari, and Christian Rupprecht. Semantic image manipulation using scene graphs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5213-5222, 2020. 1 +[6] Helisa Dhamo, Fabian Manhardt, Nassir Navab, and Federico Tombari. Graph-to-3d: End-to-end generation and manipulation of 3d scenes using scene graphs. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16352–16361, 2021. 1 +[7] Jessica D'souza, PK Aleema, S Dhanyashree, Clita Fernandes, KM Kavitha, and Chandra Naik. Knowledge-based scene graph generation in medical field. In 2023 IEEE International Conference on Distributed Computing, VLSI, Electrical Circuits and Robotics (DISCOVER), pages 232-237. IEEE, 2023. 1 +[8] Mingtao Feng, Haoran Hou, Liang Zhang, Zijie Wu, Yulan Guo, and Ajmal Mian. 3d spatial multimodal knowledge accumulation for scene graph prediction in point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9182-9191, 2023. 2 +[9] Gege Gao, Weiyang Liu, Anpei Chen, Andreas Geiger, and Bernhard Scholkopf. Graphdreamer: Compositional 3d scene synthesis from scene graphs. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21295-21304, 2024. 1 +[10] Qiao Gu, Ali Kuwajerwala, Sacha Morin, Krishna Murthy Jatavallabhula, Bipasha Sen, Aditya Agarwal, Corban Rivera, William Paul, Kirsty Ellis, Rama Chellappa, et al. Conceptgraphs: Open-vocabulary 3d scene graphs for perception and planning. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 5021-5028. IEEE, 2024. 4 +[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4 +[12] Felix Holm, Ghazal Ghazaei, Tobias Czempiel, Ege Özsoy, Stefan Saur, and Nassir Navab. Dynamic scene graph representation for surgical video. In Proceedings of the IEEE/CVF international conference on computer vision, pages 81-87, 2023. 1 +[13] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3668–3678, 2015. 1 +[14] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 3 +[15] Chen Lin, Shuai Zheng, Zhizhe Liu, Youru Li, Zhenfeng Zhu, and Yao Zhao. Sgt: Scene graph-guided transformer for surgical report generation. In International conference on medical image computing and computer-assisted intervention, pages 507-518. Springer, 2022. 1 +[16] Xin Lin, Changxing Ding, Jing Zhang, Yibing Zhan, and Dacheng Tao. Ru-net: Regularized unrolling network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19466, 2022. 4 +[17] Kien Nguyen, Subarna Tripathi, Bang Du, Tanaya Guha, and Truong Q Nguyen. In defense of scene graphs for image captioning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1407-1416, 2021. 1 +[18] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research Journal, pages 1-31, 2024. 6 +[19] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2856-2865, 2021. 1 +[20] Ege Özsoy, Evin Pinar Örnek, Ulrich Eck, Tobias Czempiel, Federico Tombari, and Nassir Navab. 4d-or: Semantic scene graphs for or domain modeling. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 475–485. Springer, 2022. 1 + +[21] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 4, 6 +[22] Brigit Schroeder and Subarna Tripathi. Structured query-based image retrieval using scene graphs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 178-179, 2020. 1 +[23] Sebastian Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In Proceedings of the fourth workshop on vision and language, pages 70–80, 2015. 1 +[24] Sahand Sharifzadeh, Sina Moayed Baharlou, and Volker Tresp. Classification by attention: Scene graph classification with prior knowledge. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5025-5033, 2021. 2 +[25] Motoharu Sonogashira, Masaaki Iiyama, and Yasutomo Kawanishi. Towards open-set scene graph generation with unknown objects. IEEE Access, 10:11574-11583, 2022. 2 +[26] Keisuke Tateno, Federico Tombari, and Nassir Navab. Real-time and scalable incremental segmentation on dense slam. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4465-4472. IEEE, 2015. 3, 6 +[27] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4 +[28] Johanna Wald, Helisa Dhamo, Nassir Navab, and Federico Tombari. Learning 3d semantic scene graphs from 3d indoor reconstructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3961-3970, 2020. 1, 2, 5, 6 +[29] Johanna Wald, Nassir Navab, and Federico Tombari. Learning 3d semantic scene graphs with instance embeddings. International Journal of Computer Vision, 130(3):630-651, 2022. 1, 6 +[30] Ziqin Wang, Bowen Cheng, Lichen Zhao, Dong Xu, Yang Tang, and Lu Sheng. Vl-sat: Visual-linguistic semantics assisted training for 3d semantic scene graph prediction in point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21560–21569, 2023. 2 +[31] Shun-Cheng Wu, Johanna Wald, Keisuke Tateno, Nassir Navab, and Federico Tombari. Scenographfusion: Incremental 3d scene graph prediction from rgb-d sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7515-7525, 2021. 1, 2, 3 +[32] Shun-Cheng Wu, Keisuke Tateno, Nassir Navab, and Federico Tombari. Incremental 3d semantic scene graph prediction from rgb sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5064-5074, 2023. 1, 2, 3, 4, 6, 8 +[33] Danfei Xu, Yuke Zhu, Christopher B Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In Pro + +ceedings of the IEEE conference on computer vision and pattern recognition, pages 5410-5419, 2017. 2 +[34] Ning Xu, An-An Liu, Jing Liu, Weizhi Nie, and Yuting Su. Scene graph captioner: Image captioning based on structural visual representation. Journal of Visual Communication and Image Representation, 58:477-485, 2019. 1 +[35] Jianwei Yang, Jiasen Lu, Stefan Lee, Dhruv Batra, and Devi Parikh. Graph r-cnn for scene graph generation. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 2 +[36] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 2, 3 +[37] Xu Yang, Kaihua Tang, Hanwang Zhang, and Jianfei Cai. Auto-encoding scene graphs for image captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10685–10694, 2019. 1 +[38] Sangwoong Yoon, Woo Young Kang, Sungwook Jeon, SeongEun Lee, Changjin Han, Jonghun Park, and Eun-Sol Kim. Image-to-image retrieval by learning similarity between scene graphs. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10718-10726, 2021. 1 +[39] Kun Yuan, Manasi Kattel, Joel L Lavanchy, Nassir Navab, Vinkle Srivastav, and Nicolas Padoy. Advancing surgical vqa with scene graph knowledge. International Journal of Computer Assisted Radiology and Surgery, 19(7):1409-1417, 2024. 1 +[40] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural motifs: Scene graph parsing with global context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5831-5840, 2018. 2 +[41] Chaoyi Zhang, Jianhui Yu, Yang Song, and Weidong Cai. Exploiting edge-oriented reasoning for 3d point-based scene graph analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9705-9715, 2021. 2 +[42] Shoulong Zhang, Aimin Hao, Hong Qin, et al. Knowledge-inspired 3d scene graph prediction in point cloud. Advances in Neural Information Processing Systems, 34:18620-18632, 2021. 2 +[43] Zhongping Zhang, Huiwen He, Bryan A Plummer, Zhenyu Liao, and Huayan Wang. Complex scene image editing by scene graph comprehension. In BMVC, 2023. 1 \ No newline at end of file diff --git a/2025/Statistical Confidence Rescoring for Robust 3D Scene Graph Generation from Multi-View Images/images.zip b/2025/Statistical Confidence Rescoring for Robust 3D Scene Graph Generation from Multi-View Images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ec8076f729016c2aed37b9b463c4564b52d3b2f2 --- /dev/null +++ b/2025/Statistical Confidence Rescoring for Robust 3D Scene Graph Generation from Multi-View Images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b71f8f90df417a62313d6a5971b4d3ae34b031da766e68a7646b84bb275a6387 +size 615414 diff --git a/2025/Statistical Confidence Rescoring for Robust 3D Scene Graph Generation from Multi-View Images/layout.json b/2025/Statistical Confidence Rescoring for Robust 3D Scene Graph Generation from Multi-View Images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4a3281727c6b45ce7c91debc20b8b5e08e07ae58 --- /dev/null +++ b/2025/Statistical Confidence Rescoring for Robust 3D Scene Graph Generation from Multi-View Images/layout.json @@ -0,0 +1,8191 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 102, + 545, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 102, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 65, + 102, + 545, + 140 + ], + "type": "text", + "content": "Statistical Confidence Rescoring for Robust 3D Scene Graph Generation from Multi-View Images" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 203, + 161, + 408, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 161, + 408, + 175 + ], + "spans": [ + { + "bbox": [ + 203, + 161, + 408, + 175 + ], + "type": "text", + "content": "Qi Xun Yeo Yanyan Li Gim Hee Lee" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 141, + 175, + 468, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 175, + 468, + 190 + ], + "spans": [ + { + "bbox": [ + 141, + 175, + 468, + 190 + ], + "type": "text", + "content": "Department of Computer Science, National University of Singapore" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 198, + 191, + 409, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 191, + 409, + 202 + ], + "spans": [ + { + "bbox": [ + 198, + 191, + 409, + 202 + ], + "type": "text", + "content": "{qixunyeo,yan.li,gimhee.lee}@nus.edu.sg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 258, + 296, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 522 + ], + "type": "text", + "content": "Modern 3D semantic scene graph estimation methods utilize ground truth 3D annotations to accurately predict target objects, predicates, and relationships. In the absence of given 3D ground truth representations, we explore leveraging only multi-view RGB images to tackle this task. To attain robust features for accurate scene graph estimation, we must overcome the noisy reconstructed pseudo point-based geometry from predicted depth maps and reduce the amount of background noise present in multi-view image features. The key is to enrich node and edge features with accurate semantic and spatial information and through neighboring relations. We obtain semantic masks to guide feature aggregation to filter background features and design a novel method to incorporate neighboring node information to aid robustness of our scene graph estimates. Furthermore, we leverage on explicit statistical priors calculated from the training summary statistics to refine node and edge predictions based on their one-hop neighborhood. Our experiments show that our method outperforms current methods purely using multi-view images as the initial input. Our project page is available at https://qixun1.github.io/projects/SCRSSG." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 548, + 135, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 135, + 560 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 135, + 560 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "The semantic scene graph (SSG) is a crucial intermediate representation that enhances higher-level scene understanding. It plays a key role in tasks such as image captioning [17, 34, 37], image retrieval [13, 22, 23, 38], image editing [3, 5, 43], and medical applications [7, 12, 15, 20, 39] by capturing both semantic and, more importantly, relational information between objects and their surroundings. Initially developed for 2D images, SSG has since expanded to the 3D domain [6, 9, 19]. 3D scene graphs provide a high-level representation of an entire 3D scene using inputs such as multi-view RGB images or LiDAR point clouds. Unlike their 2D counterparts, they incorporate spatial relation" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 232, + 553, + 313 + ], + "blocks": [ + { + "bbox": [ + 317, + 232, + 553, + 313 + ], + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 313 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 313 + ], + "type": "image", + "image_path": "10f9bb1c1d968dcecc34a9179ce6fe0f4eb8083b7c106fca028feadbbe5c4a55.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 317, + 553, + 396 + ], + "blocks": [ + { + "bbox": [ + 317, + 317, + 553, + 396 + ], + "lines": [ + { + "bbox": [ + 317, + 317, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 317, + 317, + 553, + 396 + ], + "type": "image", + "image_path": "6cf51e49573ec2724c7423f0cc28af1d864f451b002fcf1e2f6d3975a3199906.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 410, + 555, + 455 + ], + "lines": [ + { + "bbox": [ + 313, + 410, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 555, + 455 + ], + "type": "text", + "content": "Figure 1. Existing multi-view RGB methods aggregate features using bounding boxes, introducing background noise that hinders accuracy. We use pretrained segmentation masks to reduce this interference and thus improving prediction accuracy." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 471, + 554, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 554, + 495 + ], + "type": "text", + "content": "ships that extend beyond the visible image plane, allowing a holistic understanding of complex environments." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 496, + 555, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 496, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 312, + 496, + 555, + 651 + ], + "type": "text", + "content": "Many approaches [1, 28, 29] rely on ground-truth 3D LiDAR point clouds as input. However, LiDAR sensors are resource-intensive and lack inherent semantic richness. Recent works [31, 32] have shifted toward using semantically rich multi-view RGB images, which offer a balance between computational efficiency and improved semantic fidelity. Although point cloud-based methods remain valuable for capturing geometric information, our approach focuses on leveraging multi-view RGB images. By incorporating pretrained depth estimators, we generate pseudo point-based geometry to estimate the 3D semantic scene graph, reducing reliance on LiDAR while preserving spatial and structural details." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "A key challenge in using only multi-view RGB images is ensuring that node features remain free from distractors. As shown in Fig. 1, prior approaches that rely on entity detectors to extract 2D bounding box proposals or regions of interest often fail to guarantee robustness since distrac" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24999" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 191 + ], + "type": "text", + "content": "tors within the bounding box are often mistakenly aggregated. Ensuring robust initial multi-view features is crucial without ground truth point-based geometry for error correction. No explicit mechanism exists to correct incorrectly bounded objects or misclassified rare classes that are not well trained on the model. To alleviate this problem, a mechanism should exist to refine estimates using prior knowledge instead of depending solely on implicit interactions between nodes and edges to achieve confident and accurate predictions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 192, + 298, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 192, + 298, + 431 + ], + "spans": [ + { + "bbox": [ + 54, + 192, + 298, + 431 + ], + "type": "text", + "content": "In this paper, we propose a framework to enhance the robustness of the model. First, we introduce a masked feature initialization (MFI), which leverages segmentation masks to aggregate image features instead of relying on the bounding box proposals. This reduces background noise, which results in cleaner multi-view image features. Next, we design a robust residual spatial neighbor graph neural network (RSN-GNN) to encode spatial information into node features. This network filters highly activated regions from neighboring nodes, refining target node features for improved predicate estimation. Finally, we propose a confidence rescoring (CR) module, which refines object and predicate estimates using an inverse softmax-weighted contribution of neighboring node-to-node and node-to-edge co-occurrence counts. By integrating this explicit inductive bias, our approach improves the accuracy of the prediction, particularly in low-confidence scenarios. Extensive experimental results on the benchmark 3RScan dataset show the competitiveness of our proposed approach compared to existing 2D and 3D approaches." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 432, + 278, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 278, + 444 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 278, + 444 + ], + "type": "text", + "content": "Our main contributions can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 446, + 296, + 584 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 61, + 446, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 446, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 61, + 446, + 295, + 483 + ], + "type": "text", + "content": "- We introduce a masked feature initialization to enhance the robustness of node features by reducing background distractors, yielding cleaner multi-view image features." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 484, + 296, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 484, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 61, + 484, + 296, + 520 + ], + "type": "text", + "content": "- We design a novel GCN architecture that integrates highly activated neighboring features into the target node to increase the robustness of the edge features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 522, + 296, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 522, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 61, + 522, + 296, + 547 + ], + "type": "text", + "content": "- We propose a new refinement module to explicitly refine predictions based on statistical prior knowledge." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 549, + 296, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 549, + 296, + 584 + ], + "spans": [ + { + "bbox": [ + 61, + 549, + 296, + 584 + ], + "type": "text", + "content": "- Experiments show that we outperform previous state-of-the-art approaches on the 3RScan dataset, particularly on metrics that deal with low-tail imbalanced classes." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 597, + 143, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 143, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 143, + 609 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "2D Semantic Scene Graph. 2D semantic scene graph prediction is typically categorized into two-step and one-step approaches. The two-step approach first detects objects and then classifies their relationships [2, 4, 35, 40]. In contrast, the one-step approach jointly infers object and relationship classes [33]. Xu et al. [33] pioneers the problem of scene graph generation and they tried to solve this problem via iterative message passing. Baier et al. [2] first showed" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 254 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 254 + ], + "type": "text", + "content": "how semantic models can be improved by incorporating triplet frequencies. Zellers et al. [40] analyzes the usefulness of statistical co-occurrences for the Visual Genome dataset and concluded that such statistical priors serves as strong regularization for the task. Chen et al. [4] formalizes the first approach (KERN) to incorporate the statistical prior directly into graph neural networks. Sharifzadeh et. al. proposes Schemata to assimilate image-based relational prior knowledge into the representations within the neural network [24]. Compared to previous methods, our method combines predictions from the prediction head with statistical co-occurrence of the current node and neighbors to determine the final class using a confidence score. Unlike Sonogashira et. al. [25] which uses hard thresholding, our approach uses adaptive relationship and node refinements." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 306, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 556, + 715 + ], + "type": "text", + "content": "3D Semantic Scene Graph. Methods for generating a 3D semantic scene graph can be categorized into two types based on input data. The first relies on 3D inputs such as ground-truth 3D geometry to learn contextual information [28, 30, 41]. The second relies on multi-view RGB images with some methods incorporating depth information [31] while others do not [32]. Our approach follows the latter, leveraging the estimated depth from off-the-shelf depth estimators [36] to enhance the information provided by multi-view RGB images. Wald et al. [28] first proposed 3RScan: a richly annotated 3D scene graph dataset and a method that focuses on pairwise relationships to predict the 3D scene graphs. Wu et al. [31] uses SLAM to obtain dense point representations and a novel graph convolutional network (GCN) based aggregation function to enhance 3D scene graph prediction. Zhang et al. [41] proposes an edge-oriented GCN to incorporate multi-dimensional edge features for explicit inter-node relationship modeling. Zhang et al. [42] integrates prior commonsense knowledge by learning meta-embeddings only from the class labels with their graphical structures to avoid perceptual errors. Wang et al. [30] incorporated additional information by distilling knowledge learned by the multi-modal oracle model to the 3D model. Wu et al. [32] introduces a novel entity association method to obtain the 3D entities from the predicted 2D entities and a geometric gate to incorporate geometric information to multi-view image features. Feng et al. [8] uses ConceptNet external knowledge base to accumulate both contextualized visual content and textual facts to form a 3D spatial multimodal knowledge graph. Our method leverages on the explicit statistical priors similar to 2D scene graph method KERN and Schemata while guiding low confidence predictions made by the initial node and edge predictors using both visual and geometric cues." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25000" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 553, + 194 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 194 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 194 + ], + "type": "image", + "image_path": "b8012a394b2241f11960424a8067920d6cd7ff88f68ec8065dd998a40112353f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 201, + 555, + 248 + ], + "lines": [ + { + "bbox": [ + 54, + 201, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 54, + 201, + 555, + 248 + ], + "type": "text", + "content": "Figure 2. The architecture of our proposed framework consists of three main components: 1) A masked feature initialization step leverages backprojected semantic masks to aggregate multi-view image, geometric, and spatial features. 2) A message passing step utilizes our residual spatial neighbor GNN module to refine node and edge features through feature interaction. 3) A confidence rescoring step adjusts logits based on summary statistics of object and predicate co-occurrence from the training dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 266, + 162, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 162, + 280 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 162, + 280 + ], + "type": "text", + "content": "3. Our Methodology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 286, + 295, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 286, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 295, + 323 + ], + "type": "text", + "content": "Problem Definition. Given either a set of 3D point clouds " + }, + { + "bbox": [ + 55, + 286, + 295, + 323 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 286, + 295, + 323 + ], + "type": "text", + "content": " or multi-view RGB images " + }, + { + "bbox": [ + 55, + 286, + 295, + 323 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 286, + 295, + 323 + ], + "type": "text", + "content": " of a 3D scene as input, the goal is to estimate the 3D scene graph denoted as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 148, + 333, + 295, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 333, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 148, + 333, + 295, + 346 + ], + "type": "interline_equation", + "content": "G = (O, R), \\tag {1}", + "image_path": "825a5f5851223ba0d6e79e908280e0339495a9247b6476ec1f7a8ea8d3d562ce.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 356, + 84, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 356, + 84, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 84, + 366 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 369, + 294, + 416 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 61, + 369, + 235, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 369, + 235, + 380 + ], + "spans": [ + { + "bbox": [ + 61, + 369, + 235, + 380 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 61, + 369, + 235, + 380 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 61, + 369, + 235, + 380 + ], + "type": "text", + "content": ": Number of multi-view input images;" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 380, + 235, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 380, + 235, + 392 + ], + "spans": [ + { + "bbox": [ + 61, + 380, + 235, + 392 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 61, + 380, + 235, + 392 + ], + "type": "inline_equation", + "content": "O = \\{o_m\\}_{m=1}^M" + }, + { + "bbox": [ + 61, + 380, + 235, + 392 + ], + "type": "text", + "content": ": All objects in the scene;" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 392, + 294, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 392, + 294, + 405 + ], + "spans": [ + { + "bbox": [ + 61, + 392, + 294, + 405 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 61, + 392, + 294, + 405 + ], + "type": "inline_equation", + "content": "R = \\{r_{k\\to m}\\}_{m,k = 1}^{M,M}" + }, + { + "bbox": [ + 61, + 392, + 294, + 405 + ], + "type": "text", + "content": ": Relationships between all objects;" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 405, + 242, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 405, + 242, + 416 + ], + "spans": [ + { + "bbox": [ + 61, + 405, + 242, + 416 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 61, + 405, + 242, + 416 + ], + "type": "inline_equation", + "content": "M = |O|" + }, + { + "bbox": [ + 61, + 405, + 242, + 416 + ], + "type": "text", + "content": ": Number of objects in the scene." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 419, + 296, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 296, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 296, + 480 + ], + "type": "text", + "content": "Furthermore, we refer to " + }, + { + "bbox": [ + 55, + 419, + 296, + 480 + ], + "type": "inline_equation", + "content": "o_m" + }, + { + "bbox": [ + 55, + 419, + 296, + 480 + ], + "type": "text", + "content": " as the object class for node instance " + }, + { + "bbox": [ + 55, + 419, + 296, + 480 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 419, + 296, + 480 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 419, + 296, + 480 + ], + "type": "inline_equation", + "content": "r_{k \\to m}" + }, + { + "bbox": [ + 55, + 419, + 296, + 480 + ], + "type": "text", + "content": " as the predicate class for the edge connecting nodes " + }, + { + "bbox": [ + 55, + 419, + 296, + 480 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 419, + 296, + 480 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 419, + 296, + 480 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 419, + 296, + 480 + ], + "type": "text", + "content": ". Alternatively, the 3D scene graph can be defined as a set of triplets given by: " + }, + { + "bbox": [ + 55, + 419, + 296, + 480 + ], + "type": "inline_equation", + "content": "\\{o_k, r_{k \\to m}, o_m\\} := \\{node, predicate, node\\}" + }, + { + "bbox": [ + 55, + 419, + 296, + 480 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 486, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 581 + ], + "type": "text", + "content": "Overview. Fig. 2 shows an illustration of the architecture of our proposed framework. The input to our framework is a set of " + }, + { + "bbox": [ + 55, + 486, + 296, + 581 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 486, + 296, + 581 + ], + "type": "text", + "content": " multi-view RGB images " + }, + { + "bbox": [ + 55, + 486, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\{I_i\\}_{i=1}^N" + }, + { + "bbox": [ + 55, + 486, + 296, + 581 + ], + "type": "text", + "content": " of the 3D scene. Since the focus of our work is not on multi-view 3D reconstruction, we recover the 3D point cloud of the scene by running an off-the-shelf RGB-D SLAM system [26] with the metric depth maps of the input images " + }, + { + "bbox": [ + 55, + 486, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\{\\hat{D}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 55, + 486, + 296, + 581 + ], + "type": "text", + "content": " obtained from the pretrained Depth Anything [36] model." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "From the predicted 3D point cloud, we initialize multiview image features using segmentation masks, extract geometric features with a point-based feature extractor, and derive spatial features via a linear layer. These multi-view image features serve as the initial input to our RSN-GNN module. At the GCN initialization, we enhance node features by integrating geometric and spatial features through geometric and spatial gates. We additionally enhance edge features by integrating max pooled neighboring nodes features. These enhanced node and edge features undergo message passing, with the integration process repeating at each" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 267, + 553, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 267, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 267, + 553, + 315 + ], + "type": "text", + "content": "layer. After " + }, + { + "bbox": [ + 313, + 267, + 553, + 315 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 267, + 553, + 315 + ], + "type": "text", + "content": " iterations, the final node and edge features are processed by predictors to generate initial estimates, which are then refined using precomputed co-occurrence statistics via our CR module, producing the final scene graph output." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 316, + 553, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 553, + 352 + ], + "type": "text", + "content": "We elaborate on the feature initialization step in Sec. 3.1, the message passing step in Sec. 3.2, and the confidence rescoring step in Sec. 3.3." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 360, + 507, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 507, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 507, + 372 + ], + "type": "text", + "content": "3.1. Masked Feature Initialization (MFI)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 378, + 554, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 554, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 554, + 545 + ], + "type": "text", + "content": "We initialize the node features of the scene graph by combining multi-view image representations from a pretrained image encoder with PointNet features derived from the predicted points in the RGB-D SLAM reconstruction step that utilizes predicted depth. To eliminate background clutter, we employ the Segment Anything Model (SAM) [14] to produce precise 2D semantic masks for each object. For each node, we backproject the associated points onto the 2D instance mask to identify its corresponding region. We then generate object proposals from these instance masks, resize them to " + }, + { + "bbox": [ + 313, + 378, + 554, + 545 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 313, + 378, + 554, + 545 + ], + "type": "text", + "content": ", and extract their multi-view image features using the pretrained encoder. Finally, we apply these semantic masks to the multi-view image features before aggregating them into our node representations." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": "Our method to get the object proposals, i.e. the nodes of the scene graph offers two key advantages over relying solely on the entity detector from SceneGraphFusion and JointSSG [31, 32]. First, the entity detector often captures only part of an object. Although SceneGraphFusion propagates the label for this segment, the feature extractor can operate only on the detected portion. In contrast, our method employs more complete object masks, enabling the feature extractor to derive features from the entire object. Second, we exclude extraneous background features by applying these masks before feature aggregation. In SceneGraphFusion and JointSSG, the bounding boxes of the pretrained entity detector frequently include background regions, which can dilute the quality of the features and hinder" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25001" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": "the performance. Our method is closer to the open vocabulary method ConceptGraphs [10] which uses SAM masks but with the aggregation method provided by JointSSG." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 108, + 296, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 296, + 193 + ], + "type": "text", + "content": "Formally, we compute a corresponding multi-view image feature " + }, + { + "bbox": [ + 55, + 108, + 296, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{i,img}" + }, + { + "bbox": [ + 55, + 108, + 296, + 193 + ], + "type": "text", + "content": ", a geometric feature " + }, + { + "bbox": [ + 55, + 108, + 296, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{i,geo}" + }, + { + "bbox": [ + 55, + 108, + 296, + 193 + ], + "type": "text", + "content": ", and a spatial feature " + }, + { + "bbox": [ + 55, + 108, + 296, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{i,spat}" + }, + { + "bbox": [ + 55, + 108, + 296, + 193 + ], + "type": "text", + "content": " for each node " + }, + { + "bbox": [ + 55, + 108, + 296, + 193 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 55, + 108, + 296, + 193 + ], + "type": "text", + "content": ". We aggregate features only for nodes that are visible in the可视ability graph " + }, + { + "bbox": [ + 55, + 108, + 296, + 193 + ], + "type": "inline_equation", + "content": "G_{v}" + }, + { + "bbox": [ + 55, + 108, + 296, + 193 + ], + "type": "text", + "content": " denoted by " + }, + { + "bbox": [ + 55, + 108, + 296, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{1}(v_{i} \\in I_{k})" + }, + { + "bbox": [ + 55, + 108, + 296, + 193 + ], + "type": "text", + "content": " and have corresponding pixels within the semantic mask provided by SAM. The aggregated multi-view image feature " + }, + { + "bbox": [ + 55, + 108, + 296, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{i,img}" + }, + { + "bbox": [ + 55, + 108, + 296, + 193 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 97, + 201, + 295, + 229 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 201, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 97, + 201, + 295, + 229 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {i} ^ {0} = \\frac {\\sum_ {k = 1} ^ {N} \\mathbf {1} \\left(v _ {i} \\in I _ {k}\\right) f \\left(\\eta_ {k} \\odot x _ {i , k}\\right)}{\\left| k \\in G _ {v} \\right|}, \\tag {2}", + "image_path": "609f324545070c07333c561beb96df969deb4263062ce5f8a96639e881d729bf.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 237, + 296, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 237, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 296, + 308 + ], + "type": "text", + "content": "which also serves as the initial node feature. " + }, + { + "bbox": [ + 55, + 237, + 296, + 308 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 55, + 237, + 296, + 308 + ], + "type": "text", + "content": " is the image feature encoder and " + }, + { + "bbox": [ + 55, + 237, + 296, + 308 + ], + "type": "inline_equation", + "content": "\\eta_{k}" + }, + { + "bbox": [ + 55, + 237, + 296, + 308 + ], + "type": "text", + "content": " refers to the 2D semantic mask. " + }, + { + "bbox": [ + 55, + 237, + 296, + 308 + ], + "type": "inline_equation", + "content": "x_{i,k}" + }, + { + "bbox": [ + 55, + 237, + 296, + 308 + ], + "type": "text", + "content": " refers to the cropped image patch containing the node " + }, + { + "bbox": [ + 55, + 237, + 296, + 308 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 237, + 296, + 308 + ], + "type": "text", + "content": " from the " + }, + { + "bbox": [ + 55, + 237, + 296, + 308 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 237, + 296, + 308 + ], + "type": "text", + "content": "-th image. " + }, + { + "bbox": [ + 55, + 237, + 296, + 308 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 237, + 296, + 308 + ], + "type": "text", + "content": " refers to the element-wise multiplication operator. The aggregation is the arithmetic mean of the node features for each image in the可视ability graph." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 308, + 296, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 296, + 392 + ], + "type": "text", + "content": "The geometric feature " + }, + { + "bbox": [ + 55, + 308, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{i,geo}" + }, + { + "bbox": [ + 55, + 308, + 296, + 392 + ], + "type": "text", + "content": " is extracted using a vanilla PointNet [21], with inputs derived from the 3D instance masks generated during the RGB-D SLAM step. For the spatial features, we employ a simple linear layer to transform the attributes of each 3D bounding box: dimensions " + }, + { + "bbox": [ + 55, + 308, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_i" + }, + { + "bbox": [ + 55, + 308, + 296, + 392 + ], + "type": "text", + "content": ", volume " + }, + { + "bbox": [ + 55, + 308, + 296, + 392 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 55, + 308, + 296, + 392 + ], + "type": "text", + "content": ", length " + }, + { + "bbox": [ + 55, + 308, + 296, + 392 + ], + "type": "inline_equation", + "content": "l_i" + }, + { + "bbox": [ + 55, + 308, + 296, + 392 + ], + "type": "text", + "content": ", and the ratio of its x- and y-axes " + }, + { + "bbox": [ + 55, + 308, + 296, + 392 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 55, + 308, + 296, + 392 + ], + "type": "text", + "content": " into the spatial feature as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 118, + 401, + 295, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 401, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 118, + 401, + 295, + 415 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {i, s p a t} = g _ {s} ([ \\mathbf {b} _ {i}, s _ {i}, l _ {i}, r _ {i} ]), \\tag {3}", + "image_path": "83c218cbb8f8937ed3a56905740f7e6e73b74596b373e60ebfa64922f954eeff.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 423, + 296, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 296, + 459 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 423, + 296, + 459 + ], + "type": "inline_equation", + "content": "g_{s}" + }, + { + "bbox": [ + 55, + 423, + 296, + 459 + ], + "type": "text", + "content": " refers to a set of learnable weights for spatial features. " + }, + { + "bbox": [ + 55, + 423, + 296, + 459 + ], + "type": "inline_equation", + "content": "[\\cdot]" + }, + { + "bbox": [ + 55, + 423, + 296, + 459 + ], + "type": "text", + "content": " refers to the concatenation of the inputs within the operator." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 466, + 267, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 466, + 267, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 267, + 479 + ], + "type": "text", + "content": "3.2. Residual Spatial Neighbor GNN Module" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 483, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 568 + ], + "type": "text", + "content": "Following a similar design to the learnable geometric gate in JointSSG [32], we integrate both geometric and spatial features into the initial node feature, which originally consists only of multi-view image features. The spatial features serve as additional information for individual nodes, analogous to the embeddings used for token encoding in Transformer architectures [27]. We define the geometric gate as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 92, + 574, + 295, + 591 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 574, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 92, + 574, + 295, + 591 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {v}} _ {i} ^ {l} = \\mathbf {v} _ {i} ^ {l} + \\sigma \\left(\\mathbf {w} _ {g} ^ {T} \\left[ \\mathbf {v} _ {i} ^ {l}, \\mathbf {v} _ {i, g e o} \\right]\\right) \\cdot \\sigma \\left(\\mathbf {v} _ {i, g e o}\\right), \\tag {4}", + "image_path": "39db9ddce400062a23039242effe3a8e12cc67d585f8adbaa5561c06ccb8d169.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 598, + 296, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 598, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 598, + 296, + 634 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 598, + 296, + 634 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_g" + }, + { + "bbox": [ + 55, + 598, + 296, + 634 + ], + "type": "text", + "content": " refers to a set of learnable weights, " + }, + { + "bbox": [ + 55, + 598, + 296, + 634 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 598, + 296, + 634 + ], + "type": "text", + "content": " refers to the sigmoid operator and " + }, + { + "bbox": [ + 55, + 598, + 296, + 634 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 598, + 296, + 634 + ], + "type": "text", + "content": " refers to the layer of the node feature. Furthermore, we define the spatial gate as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 88, + 641, + 295, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 641, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 88, + 641, + 295, + 657 + ], + "type": "interline_equation", + "content": "\\ddot {\\mathbf {v}} _ {i} ^ {l} = \\hat {\\mathbf {v}} _ {i} ^ {l} + \\sigma \\left(\\mathbf {w} _ {s} ^ {T} [ \\hat {\\mathbf {v}} _ {i} ^ {l}, \\mathbf {v} _ {i, s p a t} ]\\right) \\cdot \\sigma \\left(\\mathbf {v} _ {i, s p a t}\\right), \\tag {5}", + "image_path": "05d6cdbf6e59f68d158df1bd5e95602410cb1471da466fd89f59da75a4a0a63b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 665, + 290, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 290, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 290, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 665, + 290, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_s" + }, + { + "bbox": [ + 55, + 665, + 290, + 677 + ], + "type": "text", + "content": " is a set of learnable weights for spatial features." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 554, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 180 + ], + "type": "text", + "content": "Inspired by the residual connections in ResNet [11], we introduce a mechanism to integrate highly activated neighboring features into the target node feature using max pooling during the message-passing process for edge features. This approach implicitly encodes neighboring information, improving edge prediction for scene graph estimation. Specifically, we augment the edge messages in the GCN by integrating max-pooled neighboring features before message passing as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 376, + 184, + 553, + 200 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 184, + 553, + 200 + ], + "spans": [ + { + "bbox": [ + 376, + 184, + 553, + 200 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {v}} _ {i} ^ {l} = \\ddot {\\mathbf {v}} _ {i} ^ {l} + \\max ([ \\ddot {\\mathbf {v}} _ {j} ^ {l} ] _ {j \\in n (i)}), \\tag {6}", + "image_path": "45a8f1baf4e5f1370f6718281f477cee05955d32c92ea82e69464096fb857730.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 204, + 554, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 554, + 262 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 204, + 554, + 262 + ], + "type": "inline_equation", + "content": "\\max (\\cdot)" + }, + { + "bbox": [ + 313, + 204, + 554, + 262 + ], + "type": "text", + "content": " refers to the max pooling operator and " + }, + { + "bbox": [ + 313, + 204, + 554, + 262 + ], + "type": "inline_equation", + "content": "n(i)" + }, + { + "bbox": [ + 313, + 204, + 554, + 262 + ], + "type": "text", + "content": " refers to the neighboring nodes of node " + }, + { + "bbox": [ + 313, + 204, + 554, + 262 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 204, + 554, + 262 + ], + "type": "text", + "content": ". Max pooling allows highly activated regions in the different neighboring node features to be fused into the target node feature via the residual connection." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 264, + 554, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 554, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 554, + 357 + ], + "type": "text", + "content": "We apply max-pooled neighboring features to the calculation of edge messages without using them in node message updates. Since node features are highly susceptible to noise and spurious correlations from other nodes [16], incorporating max-pooled neighboring features into node message computations can introduce confusion instead of improving performance. The edge message " + }, + { + "bbox": [ + 313, + 264, + 554, + 357 + ], + "type": "inline_equation", + "content": "m_{i\\rightarrow j}" + }, + { + "bbox": [ + 313, + 264, + 554, + 357 + ], + "type": "text", + "content": " is thus defined as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 380, + 358, + 553, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 358, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 380, + 358, + 553, + 373 + ], + "type": "interline_equation", + "content": "m _ {i \\rightarrow j} = g _ {e} ([ \\tilde {\\mathbf {v}} _ {i} ^ {l}, \\mathbf {e} _ {i j} ^ {l}, \\tilde {\\mathbf {v}} _ {j} ^ {l} ]), \\tag {7}", + "image_path": "ded497e1fd600be1ad30721dd5bdb54194ac6e5b59ec604c9d19ff1659e99c91.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 376, + 554, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 554, + 424 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 554, + 424 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 376, + 554, + 424 + ], + "type": "inline_equation", + "content": "g_{e}(\\cdot)" + }, + { + "bbox": [ + 313, + 376, + 554, + 424 + ], + "type": "text", + "content": " is an MLP and " + }, + { + "bbox": [ + 313, + 376, + 554, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{ij}^{l}" + }, + { + "bbox": [ + 313, + 376, + 554, + 424 + ], + "type": "text", + "content": " is the edge descriptor used in [32] containing relative node properties such as centroid displacement, relative bounding box size difference, and a relative pose descriptor." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 430, + 504, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 504, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 504, + 443 + ], + "type": "text", + "content": "3.3. Confidence Rescoring (CR) Module" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 448, + 554, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 554, + 614 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 554, + 614 + ], + "type": "text", + "content": "As discussed in Sec. 1, modern scene graph methods depend heavily on the entity detector and image feature extractor to produce an accurate initial object prediction. However, there is no mechanism to correct a target bounding box prediction that is initially wrong. To address this limitation, we leverage relationships with neighboring nodes, allowing us to mitigate the impact of incorrect initial predictions by weighting contributions from neighboring node statistics to refine target object classification. Fig. 3 illustrates the logits of the final prediction after applying our confidence rescoring module. This module incorporates prior knowledge to improve prediction confidence beyond that of the original node and edge multi-layer perceptron (MLP) predictor." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 615, + 554, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 554, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 554, + 651 + ], + "type": "text", + "content": "The probability of predicting node " + }, + { + "bbox": [ + 313, + 615, + 554, + 651 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 615, + 554, + 651 + ], + "type": "text", + "content": " given its neighboring nodes " + }, + { + "bbox": [ + 313, + 615, + 554, + 651 + ], + "type": "inline_equation", + "content": "n(i)" + }, + { + "bbox": [ + 313, + 615, + 554, + 651 + ], + "type": "text", + "content": " is approximated by the output of our node predictor " + }, + { + "bbox": [ + 313, + 615, + 554, + 651 + ], + "type": "inline_equation", + "content": "h_v(\\cdot)" + }, + { + "bbox": [ + 313, + 615, + 554, + 651 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 323, + 655, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 655, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 323, + 655, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} P \\left(o _ {i} \\mid \\left\\{o _ {j}: j \\in n (i) \\right\\}\\right) = \\frac {1}{Z _ {i}} \\psi_ {i} \\left(o _ {i}\\right) \\prod_ {j \\in n (i)} \\psi_ {i j} \\left(o _ {i}, o _ {j}\\right) \\\\ \\approx \\tau \\left(h _ {v} \\left(\\mathbf {v} ^ {L}\\right)\\right), \\tag {8} \\\\ \\end{array}", + "image_path": "579aaa3f3cf13e6df9989fdeef28b790c859c8bf479eb60d5cfb56d0b7630b0a.jpg" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 683, + 296, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 296, + 713 + ], + "type": "text", + "content": "1The covisibility graph is obtained as intermediate output after running the RGB-D SLAM system. It refers to a bipartite graph where each image " + }, + { + "bbox": [ + 55, + 683, + 296, + 713 + ], + "type": "inline_equation", + "content": "I_{k}" + }, + { + "bbox": [ + 55, + 683, + 296, + 713 + ], + "type": "text", + "content": " is connected with the specific node instance " + }, + { + "bbox": [ + 55, + 683, + 296, + 713 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 55, + 683, + 296, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25002" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\psi_i(o_i)" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " refers to the potential associated with node " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\psi_{ij}(o_i,o_j)" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " refers to the potential associated with the edge between node " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " and its neighbor node " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": ". The potentials are implicitly learned through the GCN unlike classical methods, which explicitly learn these potentials. " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "h_v(\\cdot)" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " is a small MLP that predicts the class of each node and " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " refers to the softmax operator. " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " refers to the last layer of the GCN message passing output. The partition function given by:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 186, + 295, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 186, + 295, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 186, + 295, + 213 + ], + "type": "interline_equation", + "content": "Z _ {i} = \\sum_ {o _ {i}} \\psi_ {i} \\left(o _ {i}\\right) \\prod_ {j \\in n (i)} \\psi_ {i j} \\left(o _ {i}, o _ {j}\\right) \\tag {9}", + "image_path": "80697be59f67dcb665bbb5be79dd363245eb62021efeeb68fba65411d7e431a5.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 220, + 296, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 220, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 296, + 258 + ], + "type": "text", + "content": "ensures the probabilities sum to 1. A similar equation can be obtained for the probability of predicting edge " + }, + { + "bbox": [ + 55, + 220, + 296, + 258 + ], + "type": "inline_equation", + "content": "ij" + }, + { + "bbox": [ + 55, + 220, + 296, + 258 + ], + "type": "text", + "content": " given its neighboring nodes " + }, + { + "bbox": [ + 55, + 220, + 296, + 258 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 220, + 296, + 258 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 220, + 296, + 258 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 220, + 296, + 258 + ], + "type": "text", + "content": ", i.e.:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 92, + 262, + 294, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 262, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 92, + 262, + 294, + 304 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} P \\left(r _ {i \\rightarrow j} \\mid o _ {i}, o _ {j}\\right) = \\frac {1}{Z _ {i j}} \\psi_ {i j} \\left(o _ {i}, o _ {j}, r _ {i \\rightarrow j}\\right) \\tag {10} \\\\ \\approx \\tau (h _ {e} (\\mathbf {e} ^ {L})), \\\\ \\end{array}", + "image_path": "0bca4b7d0b8c21f14bca033f8fd7d4898c3a1ae27cc6f48df782db7e84c81cd9.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 311, + 296, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 311, + 296, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 296, + 347 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 311, + 296, + 347 + ], + "type": "inline_equation", + "content": "h_e(\\cdot)" + }, + { + "bbox": [ + 55, + 311, + 296, + 347 + ], + "type": "text", + "content": " is a small MLP which predicts the class of each edge, and the partition function to normalize the probabilities is given by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 354, + 295, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 354, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 116, + 354, + 295, + 380 + ], + "type": "interline_equation", + "content": "Z _ {i j} = \\sum_ {r _ {i \\rightarrow j}} \\psi_ {i j} \\left(o _ {i}, o _ {j}, r _ {i \\rightarrow j}\\right). \\tag {11}", + "image_path": "2dc6d44565923d907894a652eed8d57765d9aaeebc1b53ffed43c2f27776098d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 387, + 296, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 296, + 471 + ], + "type": "text", + "content": "We begin by computing summary statistics from the training dataset to node-to-node co-occurrences denoted as " + }, + { + "bbox": [ + 55, + 387, + 296, + 471 + ], + "type": "inline_equation", + "content": "c_{ij}(o_i, o_j)" + }, + { + "bbox": [ + 55, + 387, + 296, + 471 + ], + "type": "text", + "content": ", and node-to-edge co-occurrences denoted as " + }, + { + "bbox": [ + 55, + 387, + 296, + 471 + ], + "type": "inline_equation", + "content": "d_{ij}(o_i, e_j)" + }, + { + "bbox": [ + 55, + 387, + 296, + 471 + ], + "type": "text", + "content": ". Subsequently, we determine the prediction confidence of the model, where " + }, + { + "bbox": [ + 55, + 387, + 296, + 471 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathbf{v}}" + }, + { + "bbox": [ + 55, + 387, + 296, + 471 + ], + "type": "text", + "content": " represents node confidence and " + }, + { + "bbox": [ + 55, + 387, + 296, + 471 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathbf{e}}" + }, + { + "bbox": [ + 55, + 387, + 296, + 471 + ], + "type": "text", + "content": " represents edge confidence. Next, we obtain the node " + }, + { + "bbox": [ + 55, + 387, + 296, + 471 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathbf{v}}" + }, + { + "bbox": [ + 55, + 387, + 296, + 471 + ], + "type": "text", + "content": " and edge " + }, + { + "bbox": [ + 55, + 387, + 296, + 471 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathbf{e}}" + }, + { + "bbox": [ + 55, + 387, + 296, + 471 + ], + "type": "text", + "content": " confidences as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 477, + 295, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 477, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 69, + 477, + 295, + 495 + ], + "type": "interline_equation", + "content": "\\alpha_ {\\mathbf {v}} = \\max _ {\\mathbf {v}} \\tau \\left(h _ {v} \\left(\\mathbf {v} ^ {L}\\right)\\right), \\quad \\alpha_ {\\mathbf {e}} = \\max _ {\\mathbf {e}} \\tau \\left(h _ {e} \\left(\\mathbf {e} ^ {L}\\right)\\right). \\tag {12}", + "image_path": "d8f7253304851708dced08490c805cf2cc43aac292f9bdb1a7e3d4c05601b798.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 501, + 296, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 296, + 548 + ], + "type": "text", + "content": "We first estimate the marginal probability of a specific instantiation of node " + }, + { + "bbox": [ + 55, + 501, + 296, + 548 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 501, + 296, + 548 + ], + "type": "text", + "content": ", conditioned on node " + }, + { + "bbox": [ + 55, + 501, + 296, + 548 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 501, + 296, + 548 + ], + "type": "text", + "content": " using node-to-node co-occurrences, and on edge " + }, + { + "bbox": [ + 55, + 501, + 296, + 548 + ], + "type": "inline_equation", + "content": "ij" + }, + { + "bbox": [ + 55, + 501, + 296, + 548 + ], + "type": "text", + "content": " using node-to-edge co-occurrences defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 554, + 266, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 554, + 266, + 570 + ], + "spans": [ + { + "bbox": [ + 66, + 554, + 266, + 570 + ], + "type": "interline_equation", + "content": "\\mathbf {c} \\left(o _ {i} \\mid o _ {j}\\right) = \\left[ c \\left(o _ {i} = c _ {k} \\mid o _ {j} = c _ {l}\\right) \\right] _ {k = 0, l = 0} ^ {C, C}, \\quad \\text {w h e r e}", + "image_path": "272603e8f7b2d2aad320be9ff3c89571218b47365beb6f512f7ef4412c6a1513.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 571, + 295, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 571, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 67, + 571, + 295, + 601 + ], + "type": "interline_equation", + "content": "c \\left(o _ {i} = c _ {k} \\mid o _ {j} = c _ {l}\\right) = \\frac {c \\left(o _ {i} = c _ {k} , o _ {j} = c _ {l}\\right)}{\\sum_ {k = 0} ^ {C} c \\left(o _ {i} = c _ {k} , o _ {j} = c _ {l}\\right)}. \\tag {13}", + "image_path": "c4ce363e714bccb486a7fe9c674e6f6dffbd23a9ff3348ddcdfadd1aab6638c8.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 605, + 296, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 666 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 605, + 296, + 666 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 605, + 296, + 666 + ], + "type": "text", + "content": " refers to the number of object classes and " + }, + { + "bbox": [ + 55, + 605, + 296, + 666 + ], + "type": "inline_equation", + "content": "c_{k}" + }, + { + "bbox": [ + 55, + 605, + 296, + 666 + ], + "type": "text", + "content": " refers to the " + }, + { + "bbox": [ + 55, + 605, + 296, + 666 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 605, + 296, + 666 + ], + "type": "text", + "content": "-th class in the set of object classes. We then integrate the summary statistics via an inverse softmax operation denoted as " + }, + { + "bbox": [ + 55, + 605, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\gamma(\\cdot)" + }, + { + "bbox": [ + 55, + 605, + 296, + 666 + ], + "type": "text", + "content": " into the node probability in Eq. 14 to get:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 69, + 672, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 672, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 69, + 672, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} P \\left(o _ {i} \\mid \\left\\{o _ {j}: j \\in n (i) \\right\\}\\right) = \\tau \\left(\\alpha_ {\\mathbf {v}} \\cdot h _ {v} (\\mathbf {v} _ {i}) \\right. \\\\ + \\left(1 - \\alpha_ {\\mathbf {v}}\\right) \\cdot \\sum_ {j \\in n (i)} \\alpha_ {j} \\cdot \\gamma (\\mathbf {c} \\left(o _ {i} \\mid o _ {j}\\right)). \\tag {14} \\\\ \\end{array}", + "image_path": "64ca6284120589d182160a588e06fc0107a850f0245d2c8aa29050c40185a6ab.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "This formulation is chosen because directly adding the estimated marginal probability to the predictions of the model after softmax would disregard the required normalization, resulting in incompatible terms." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 120, + 554, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 120, + 554, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 120, + 554, + 167 + ], + "type": "text", + "content": "Similarly, we compute the estimated marginal probability of an edge " + }, + { + "bbox": [ + 313, + 120, + 554, + 167 + ], + "type": "inline_equation", + "content": "ij" + }, + { + "bbox": [ + 313, + 120, + 554, + 167 + ], + "type": "text", + "content": " instantiation conditioned on nodes " + }, + { + "bbox": [ + 313, + 120, + 554, + 167 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 120, + 554, + 167 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 120, + 554, + 167 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 120, + 554, + 167 + ], + "type": "text", + "content": " using node-to-edge co-occurrence statistics for predicate estimation:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 175, + 546, + 191 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 175, + 546, + 191 + ], + "spans": [ + { + "bbox": [ + 321, + 175, + 546, + 191 + ], + "type": "interline_equation", + "content": "\\mathbf {d} \\left(r _ {i \\rightarrow j} \\mid o _ {j}\\right) = \\left[ d \\left(r _ {i \\rightarrow j} = e _ {k} \\mid o _ {j} = c _ {l}\\right)\\right] _ {k = 0, l = 0} ^ {P, C},", + "image_path": "7a722474b8bfe11ce836f8eab125a88c071542cdfab6e2eb3c5d84ab7086e54a.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 193, + 553, + 233 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 193, + 553, + 233 + ], + "spans": [ + { + "bbox": [ + 321, + 193, + 553, + 233 + ], + "type": "interline_equation", + "content": "d \\left(r _ {i \\rightarrow j = k} \\mid o _ {j = l}\\right) = \\frac {d \\left(r _ {i \\rightarrow j} = e _ {k} , o _ {j} = c _ {l}\\right)}{\\sum_ {k = 0} ^ {P} d \\left(r _ {i \\rightarrow j} = e _ {k} , o _ {j} = c _ {l}\\right)}. \\tag {15}", + "image_path": "c16aa68ba57e2b1f626462124234962da0ff3a2bf557b675401fdff581acb949.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 243, + 554, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 554, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 554, + 280 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 243, + 554, + 280 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 243, + 554, + 280 + ], + "type": "text", + "content": " refers to the total number of predicate classes and " + }, + { + "bbox": [ + 313, + 243, + 554, + 280 + ], + "type": "inline_equation", + "content": "e_k" + }, + { + "bbox": [ + 313, + 243, + 554, + 280 + ], + "type": "text", + "content": " refers to the " + }, + { + "bbox": [ + 313, + 243, + 554, + 280 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 243, + 554, + 280 + ], + "type": "text", + "content": "-th class in the set of predicate class. The refined logits for predicate estimation are then given by:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 289, + 553, + 332 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 289, + 553, + 332 + ], + "spans": [ + { + "bbox": [ + 321, + 289, + 553, + 332 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} P \\left(r _ {i \\rightarrow j} \\mid o _ {i}, o _ {j}\\right) = \\tau \\left(\\alpha_ {i j} \\cdot h _ {e} (\\mathbf {e} _ {i j}) \\right. \\\\ + \\left(1 - \\alpha_ {i j}\\right) \\cdot \\left(\\alpha_ {i} \\cdot \\gamma (\\mathbf {d} \\left(r _ {i \\rightarrow j} \\mid o _ {i}\\right))\\right) \\tag {16} \\\\ \\cdot \\left(\\alpha_ {j} \\cdot \\gamma (\\mathbf {d} (r _ {i \\rightarrow j} | o _ {j})))\\right). \\\\ \\end{array}", + "image_path": "64a789d8ae71c97cd1791f6d6d8a139b38ec38f94adde9ff215d9cf1d1b83418.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 340, + 555, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 520 + ], + "type": "text", + "content": "Remarks. As shown in Fig. 3, the original node probability for the correct class is already relatively high. However, our CR module further enhances the predicted probability as evidenced by the upward shift in the box plot from \"Original\" to \"After CR\" in the red boxes, with a notable increase in the lower quartile. Moreover, a larger proportion of low node probability instances are flagged as outliers as shown by the greater number of red circles at \"After CR\" compared to \"Original\". Most notable gains occur when original logits are low since CR increases the probability of the correct class by down-weighting misclassified classes " + }, + { + "bbox": [ + 313, + 340, + 555, + 520 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 313, + 340, + 555, + 520 + ], + "type": "text", + "content": " via lower conditional probabilities " + }, + { + "bbox": [ + 313, + 340, + 555, + 520 + ], + "type": "inline_equation", + "content": "P(c'|c)" + }, + { + "bbox": [ + 313, + 340, + 555, + 520 + ], + "type": "text", + "content": " given conditioned class " + }, + { + "bbox": [ + 313, + 340, + 555, + 520 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 340, + 555, + 520 + ], + "type": "text", + "content": ". This targeted adjustment raises suppressed logits for rare classes, thus improving prediction performance for underrepresented categories." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 531, + 395, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 395, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 395, + 544 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 551, + 430, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 551, + 430, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 551, + 430, + 563 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 568, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 554, + 665 + ], + "type": "text", + "content": "Datasets. We evaluate on the 3RScan dataset following the same training and test split from [28]. The image quality of the dataset is rather poor as each scene constitutes a variable number of short videos spliced together with different motion trajectories. As the images are obtained from consumer-level cameras, the images possess a low frame rate of " + }, + { + "bbox": [ + 313, + 568, + 554, + 665 + ], + "type": "inline_equation", + "content": "10\\mathrm{Hz}" + }, + { + "bbox": [ + 313, + 568, + 554, + 665 + ], + "type": "text", + "content": ", image blur, and image jitter from sudden camera motion which makes this dataset challenging." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 670, + 554, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 554, + 694 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 554, + 694 + ], + "type": "text", + "content": "Implementation Details. We utilize the codebase provided from the 3DSSG GitHub repository" + }, + { + "bbox": [ + 313, + 670, + 554, + 694 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 313, + 670, + 554, + 694 + ], + "type": "text", + "content": " to reproduce the prior" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 702, + 467, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 702, + 467, + 713 + ], + "spans": [ + { + "bbox": [ + 324, + 702, + 467, + 713 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 324, + 702, + 467, + 713 + ], + "type": "text", + "content": "https://github.com/ShunChengWu/3DSSG" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25003" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 71, + 556, + 126 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 556, + 126 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 556, + 126 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 556, + 126 + ], + "type": "table", + "html": "
Methodbath.bedbkshfcab.chaircntr.curt.deskdoorfloorofurnpic.refri.show.sinksofatabletoil.wallwind.Mean.
IMP0.0001.0000.0000.3410.4670.0000.6060.2000.5280.9000.1250.1430.0000.0000.3330.4500.5250.8000.6220.1250.358
VGfM0.5000.0000.0000.4150.5430.0830.5450.0000.3890.9200.2320.0860.0000.0000.3810.4500.5740.8000.6440.1670.336
3DSSG0.0000.6670.0000.2070.3480.1670.5760.2000.3610.7800.1250.0570.0000.0000.3330.5000.2790.0000.4670.1250.260
SGFN0.5000.3330.0000.2930.6850.2500.5150.2000.4440.8800.2320.1710.0000.0000.4290.5000.5080.4000.7070.1250.359
JointSSG1.0001.0000.0000.4760.6740.2500.5760.2000.5000.9000.1070.1430.0000.0000.4290.5500.5410.8000.6800.2500.454
Ours1.0001.0000.3330.6220.6300.3330.5450.6000.5280.8800.3040.1710.6670.6670.3330.6000.6230.8000.6930.3750.605
", + "image_path": "9750957ed9fb272c754841d3f8a1602addff684b3c4eca4deb3627bd5b0ef2b8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 70, + 136, + 539, + 148 + ], + "lines": [ + { + "bbox": [ + 70, + 136, + 539, + 148 + ], + "spans": [ + { + "bbox": [ + 70, + 136, + 539, + 148 + ], + "type": "text", + "content": "Table 1. Comparison with state-of-the-art methods on the 3RScan dataset for each object class. The Best results are highlighted." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 75, + 170, + 262, + 312 + ], + "blocks": [ + { + "bbox": [ + 75, + 170, + 262, + 312 + ], + "lines": [ + { + "bbox": [ + 75, + 170, + 262, + 312 + ], + "spans": [ + { + "bbox": [ + 75, + 170, + 262, + 312 + ], + "type": "image", + "image_path": "b5f5e82be139d5c212c4fc229a7d5c51692abb22563db3642c14ebd2d4ee705b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 322, + 296, + 357 + ], + "lines": [ + { + "bbox": [ + 55, + 322, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 296, + 357 + ], + "type": "text", + "content": "Figure 3. Box plot for predicted node probability for correct class after our CR module. The lower quartile visibly moves upwards due to an increase in predicted node probability for correct class." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 370, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 295, + 407 + ], + "type": "text", + "content": "methods. For JointSSG[32], we replace their image feature extractor from a pretrained ResNet18 to a pretrained DINOv2 with a ViT-L backbone to ensure fair comparison." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 407, + 295, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 407, + 295, + 551 + ], + "spans": [ + { + "bbox": [ + 54, + 407, + 295, + 551 + ], + "type": "text", + "content": "Following [32], we use PointNet [21] as the point encoder and employ two message-passing layers. For RGB-D SLAM reconstruction, we utilize the method from [26] and our predicted depth to generate predicted points. Multiview feature extraction is performed using a pretrained DINOv2 [18] with a ViT-L backbone. To obtain object segmentation masks, we use a pretrained SAM predictor with a text prompt containing the class set of 20 NYU object classes following the subset defined in [28]. Node-to-node and node-to-edge co-occurrences are precomputed on the training set. We train the model on a single RTX 3090 GPU for two days with early stopping." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 555, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 295, + 651 + ], + "type": "text", + "content": "Evaluation Metrics. Following [29] and JointSSG[32], we report the overall top-1 recall (Recall) for the object class estimation (Obj.), the predicate estimation (Pred.), and the relationship triplet estimation (Rel.). We also report the mean recall (mRecall) for the object class estimation (Obj.), the predicate estimation (Pred.) only. We map all predictions on the estimated segments to the ground truth segments for a fair comparison between the different methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 659, + 174, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 174, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 174, + 672 + ], + "type": "text", + "content": "4.2. Quantitative Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "We present the per-class recall for 20 object classes in the 3RScan dataset in Tab. 1, demonstrating that our method significantly outperforms others across most object cate" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 162, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 162, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 162, + 555, + 437 + ], + "type": "text", + "content": "gories. Notably, we are the only approach that achieves a recall above " + }, + { + "bbox": [ + 313, + 162, + 555, + 437 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 313, + 162, + 555, + 437 + ], + "type": "text", + "content": " for all object classes, indicating that we robustly learn all classes instead of just a subset. Our method tends to underperform slightly compared to the other methods on majority classes such as floor and wall classes in return, although not by a significant margin. The main experimental results for the top-1 recall and the mean recall are shown in Tab. 2. Importantly, our approach significantly outperforms others in mean recall (mRecall) for both objects and predicates, suggesting that it handles class imbalance issues [29] more effectively than competing methods. Additionally, our method surpasses all baselines in relationship and object estimation metrics while achieving comparable performance in predicate estimation. More impressively, our method with the weaker ResNet18 multiview image features already surpasses all baselines in relationship and object estimation metrics except for predicate estimation. Our method with ResNet18 image features even manages to surpass the performance of JointSSG with DINOv2 multi-view image features by non-trivial margins. This shows that our improved performance is not purely due to the more powerful DINOv2 features. Refer to our supplementary material for more quantitative results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 445, + 425, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 425, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 425, + 457 + ], + "type": "text", + "content": "4.3. Qualitative Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 462, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 641 + ], + "type": "text", + "content": "We show qualitative results comparing our method with other competitive methods in Fig. 4. Since the scene contains a large number of objects, we display only a subset with fewer objects and predicates for better readability. The correct prediction of the None predicate class is omitted in the visualization due to its prevalence in the dataset. As seen in Fig. 4, our method can distinguish the relatively more ambiguous wall which the other methods fail to distinguish. The shower curtain is tricky to classify since it is a subset of a curtain that is another class of objects present in the dataset. Understandably, all methods fail to correctly classify this object. Both our method and JointSSG attain good performance on the predicate class compared to SGFN as both methods only misclassify the predicate between the curtain object and the wall." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "We additionally evaluate on the entire scene for scan 43b8cae1 without comparison to other methods. Our method can recognize rare classes present in the scene, which are more challenging to learn. As seen in Fig. 5, our approach successfully classifies the rare class bookshelf instance denoted by " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "bkshf" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " which no other method managed" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "25004" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 73, + 527, + 331 + ], + "blocks": [ + { + "bbox": [ + 83, + 73, + 527, + 331 + ], + "lines": [ + { + "bbox": [ + 83, + 73, + 527, + 331 + ], + "spans": [ + { + "bbox": [ + 83, + 73, + 527, + 331 + ], + "type": "image", + "image_path": "5936e487f036635e02696b434b01b76f7101e3528d0de93e1a3f253bae481cbf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 335, + 555, + 380 + ], + "lines": [ + { + "bbox": [ + 55, + 335, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 555, + 380 + ], + "type": "text", + "content": "Figure 4. The qualitative results on 3RScan dataset of previous works and our proposed framework on scan 4d3d82b0. SGFN also fails to classify majority of non-background object classes except the sink and cabinet. It also fails to classify all predicate classes. JointSSG performs better but misclassifies the shower curtain as a curtain and the ambiguous blue wall as a cabinet. Our method correctly classifies the wall but misidentifies the shower curtain as a curtain." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 93, + 385, + 518, + 647 + ], + "blocks": [ + { + "bbox": [ + 93, + 385, + 518, + 647 + ], + "lines": [ + { + "bbox": [ + 93, + 385, + 518, + 647 + ], + "spans": [ + { + "bbox": [ + 93, + 385, + 518, + 647 + ], + "type": "image", + "image_path": "0fc87d00a72f2183bee837af9985e58af734d76da85d780d9600e08a3b7dd392.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 657, + 555, + 691 + ], + "lines": [ + { + "bbox": [ + 55, + 657, + 555, + 691 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 555, + 691 + ], + "type": "text", + "content": "Figure 5. The qualitative results on 3RScan dataset of our proposed framework on the entire scene of scan 43b8cae1. Our method can correctly classify most objects in the scene except for the light blue dress table at the top right of the scan. Our method can also correctly classify most predicates." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25005" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 71, + 292, + 258 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 292, + 258 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 292, + 258 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 292, + 258 + ], + "type": "table", + "html": "
MethodRecall%mRecall%
RelObj.Pred.Obj.Pred
IMP [32]25.851.890.430.023.0
VGfM [32]28.353.390.731.624.4
3DSSG [32]17.541.488.231.926.6
SGFN [32]31.456.789.638.330.5
JointSSG [32]34.158.189.943.033.3
IMP25.447.789.935.819.9
VGfM27.450.190.633.622.3
3DSSG13.035.686.826.024.3
SGFN29.952.289.535.927.1
JointSSG32.954.588.745.435.2
JointSSG‡34.356.686.552.936.2
Ours†38.659.389.157.436.8
Ours40.561.890.460.539.2
", + "image_path": "ed1bc5ddf4c41762169967b9dc63e5c5df5bd3df3ced45623864b2a54dc05f58.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 266, + 294, + 360 + ], + "lines": [ + { + "bbox": [ + 55, + 266, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 294, + 360 + ], + "type": "text", + "content": "Table 2. Comparison with state-of-the-art methods on the 3RScan dataset with 20 object classes and 8 predicate classes. The top group of results are reported in JointSSG [32]. The middle group of results are reproduced via 3DSSG GitHub repository. JointSSG‡ refers to JointSSG but with DINOv2 multi-view image features. Ours† uses ResNet18 multi-view image features; Ours uses the DINOv2 multi-view image features instead. The Best and Second Best results are highlighted, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 373, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 295, + 422 + ], + "type": "text", + "content": "to predict based on the results shown in Tab. 1. Our method also predicts multiple reasonable predicates of standing on between various objects and floor. Additional qualitative results can be found in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 431, + 156, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 156, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 156, + 443 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 449, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 664 + ], + "type": "text", + "content": "We analyze the effects of our feature initialization, robust neighbor residuals, and confidence rescoring module in Tab. 3. Our confidence rescoring (CR) module significantly improves performance across all metrics, indicating that it effectively leverages statistical priors to address class imbalance issues as noted in [32]. Apart from predicate mRecall, incorporating our masked feature initialization further improves performance across all metrics. This outcome is expected since masking primarily aims to refine multi-view image features, which do not directly contribute to edge feature improvements. The combination of the masked feature initialization and the CR module act as a force multiplier since the CR module benefits from cleaner and more accurate neighboring node features. Additionally, our robust neighbor residual gate maintains comparable performance across most metrics except for predicate mRecall, which benefits from additional neighbor information to better handle class imbalance in predicate estimation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "We additionally study the effects of using our CR module during inference in Tab. 4. The addition of the CR module with the prior knowledge of the dataset is akin to calculating statistics during transductive setting of inference instead" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 71, + 558, + 133 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 558, + 133 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 558, + 133 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 558, + 133 + ], + "type": "table", + "html": "
MethodConf.Feat.Neigh.Recall%mRecall%
Rel.Obj.Pred.Obj.Pred.
JointSSG×××32.954.588.245.435.2
+ Conf.✓××37.957.289.654.136.3
+ Feat.✓✓×41.161.390.861.335.8
Ours✓✓✓40.561.890.460.539.2
", + "image_path": "8c033161e40a4f9cac6af0b2a6911fb55407a69b2a8dc8dc5d2624ceb18880f3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 141, + 555, + 196 + ], + "lines": [ + { + "bbox": [ + 313, + 141, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 313, + 141, + 555, + 196 + ], + "type": "text", + "content": "Table 3. Ablation study on the 3RScan dataset. Conf. refers to the usage of summary statistic priors for confidence rescoring via the CR module. Feat. refers to the usage of DINOv2 image features with refinement from SAM semantic masks via MFI. Neigh. refers to the usage of RSN-GNN. The Best results are shown in bold." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 347, + 211, + 525, + 255 + ], + "blocks": [ + { + "bbox": [ + 347, + 211, + 525, + 255 + ], + "lines": [ + { + "bbox": [ + 347, + 211, + 525, + 255 + ], + "spans": [ + { + "bbox": [ + 347, + 211, + 525, + 255 + ], + "type": "table", + "html": "
MethodCR.Recall%mRecall%
Rel.Obj.Pred.Obj.Pred.
Ours×41.161.190.660.235.7
Ours✓40.561.890.460.539.2
", + "image_path": "e7ad321f39aa38eecd67d717360a08ded85bd33a8645109259c7a242cbf64b3c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 263, + 555, + 297 + ], + "lines": [ + { + "bbox": [ + 313, + 263, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 555, + 297 + ], + "type": "text", + "content": "Table 4. Ablation study on the 3RScan dataset. CR refers to the usage of summary statistic priors for confidence rescoring via the CR module during inference. The Best results are shown in bold." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 321, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 321, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 555, + 453 + ], + "type": "text", + "content": "of a true overfitting to the dataset. Interestingly, the removal of the CR module during inference does not significantly reduce performance. It even leads to a slight increase in predicate Recall and overall relationship triplet Recall. The slight increase in predicate Recall, coupled with a drastic decrease in predicate mRecall, suggests that removing the CR module benefits the predicate classes that appear frequently in the dataset, including the None class, with the trade off of decreasing performance on the tail classes. Nonetheless, the use of the CR module during evaluation benefits object Recall, object mRecall, and predicate mRecall." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 460, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 544 + ], + "type": "text", + "content": "Limitations. The effectiveness of initial feature aggregation relies on the quality of the segmentation mask. Although the CR module performs well when there is no significant domain shift, the statistical prior can negatively impact performance when applied to datasets with a drastically different training data distribution. Similarly, the prior becomes less useful on datasets with very limited data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 560, + 388, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 560, + 388, + 572 + ], + "spans": [ + { + "bbox": [ + 313, + 560, + 388, + 572 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": "We introduce a straightforward yet effective approach for enhancing 3D semantic scene graph estimation through our MFI, RSN-GNN, and CR modules. Specifically, our masked feature initialization (MFI) enhances node features by filtering out background distractors and leveraging more than just part-based image features. The RSN-GNN module strengthens edge features by integrating highly activated neighboring features into the target node feature, improving robustness. Finally, we introduce a refinement module that explicitly refines predictions using statistical prior knowledge, further enhancing performance." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25006" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 180 + ], + "type": "text", + "content": "This research work is supported by the Agency for Science, Technology and Research (A\\*STAR) under its MTC Programmatic Funds (Grant No. M23L7b0021), and the Tier 2 grant MOE-T2EP20124-0015 from the Singapore Ministry of Education. We also thank Shun-Cheng Wu for the helpful discussions on running the SLAM code and reproducing the results in the 3DSSG repository." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 190, + 115, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 190, + 115, + 202 + ], + "spans": [ + { + "bbox": [ + 56, + 190, + 115, + 202 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 210, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 210, + 296, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 210, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 61, + 210, + 296, + 266 + ], + "type": "text", + "content": "[1] Iro Armeni, Zhi-Yang He, JunYoung Gwak, Amir R Zamir, Martin Fischer, Jitendra Malik, and Silvio Savarese. 3d scene graph: A structure for unified semantics, 3d space, and camera. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5664-5673, 2019. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 267, + 296, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 267, + 296, + 332 + ], + "spans": [ + { + "bbox": [ + 61, + 267, + 296, + 332 + ], + "type": "text", + "content": "[2] Stephan Baier, Yunpu Ma, and Volker Tresp. Improving visual relationship detection using semantic modeling of scene descriptions. In The Semantic Web-ISWC 2017: 16th International Semantic Web Conference, Vienna, Austria, October 21-25, 2017, Proceedings, Part I 16, pages 53-68. Springer, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 334, + 296, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 334, + 296, + 389 + ], + "spans": [ + { + "bbox": [ + 62, + 334, + 296, + 389 + ], + "type": "text", + "content": "[3] Lichang Chen, Guosheng Lin, Shijie Wang, and Qingyao Wu. Graph edit distance reward: Learning to edit scene graph. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX 16, pages 539-554. Springer, 2020. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 390, + 296, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 390, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 62, + 390, + 296, + 443 + ], + "type": "text", + "content": "[4] Tianshui Chen, Weihao Yu, Riquan Chen, and Liang Lin. Knowledge-embedded routing network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6163–6171, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 446, + 296, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 446, + 296, + 500 + ], + "spans": [ + { + "bbox": [ + 62, + 446, + 296, + 500 + ], + "type": "text", + "content": "[5] Helisa Dhamo, Azade Farshad, Iro Laina, Nassir Navab, Gregory D Hager, Federico Tombari, and Christian Rupprecht. Semantic image manipulation using scene graphs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5213-5222, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 502, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 296, + 555 + ], + "type": "text", + "content": "[6] Helisa Dhamo, Fabian Manhardt, Nassir Navab, and Federico Tombari. Graph-to-3d: End-to-end generation and manipulation of 3d scenes using scene graphs. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16352–16361, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 557, + 296, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 296, + 623 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 296, + 623 + ], + "type": "text", + "content": "[7] Jessica D'souza, PK Aleema, S Dhanyashree, Clita Fernandes, KM Kavitha, and Chandra Naik. Knowledge-based scene graph generation in medical field. In 2023 IEEE International Conference on Distributed Computing, VLSI, Electrical Circuits and Robotics (DISCOVER), pages 232-237. IEEE, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 624, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 296, + 679 + ], + "type": "text", + "content": "[8] Mingtao Feng, Haoran Hou, Liang Zhang, Zijie Wu, Yulan Guo, and Ajmal Mian. 3d spatial multimodal knowledge accumulation for scene graph prediction in point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9182-9191, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 681, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 681, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 62, + 681, + 296, + 714 + ], + "type": "text", + "content": "[9] Gege Gao, Weiyang Liu, Anpei Chen, Andreas Geiger, and Bernhard Scholkopf. Graphdreamer: Compositional 3d scene synthesis from scene graphs. In Proceedings of" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21295-21304, 2024. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 173 + ], + "type": "text", + "content": "[10] Qiao Gu, Ali Kuwajerwala, Sacha Morin, Krishna Murthy Jatavallabhula, Bipasha Sen, Aditya Agarwal, Corban Rivera, William Paul, Kirsty Ellis, Rama Chellappa, et al. Conceptgraphs: Open-vocabulary 3d scene graphs for perception and planning. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 5021-5028. IEEE, 2024. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 175, + 553, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 219 + ], + "type": "text", + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 220, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 274 + ], + "type": "text", + "content": "[12] Felix Holm, Ghazal Ghazaei, Tobias Czempiel, Ege Özsoy, Stefan Saur, and Nassir Navab. Dynamic scene graph representation for surgical video. In Proceedings of the IEEE/CVF international conference on computer vision, pages 81-87, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 276, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 330 + ], + "type": "text", + "content": "[13] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3668–3678, 2015. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 332, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 388 + ], + "type": "text", + "content": "[14] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 388, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 443 + ], + "type": "text", + "content": "[15] Chen Lin, Shuai Zheng, Zhizhe Liu, Youru Li, Zhenfeng Zhu, and Yao Zhao. Sgt: Scene graph-guided transformer for surgical report generation. In International conference on medical image computing and computer-assisted intervention, pages 507-518. Springer, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 445, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 499 + ], + "type": "text", + "content": "[16] Xin Lin, Changxing Ding, Jing Zhang, Yibing Zhan, and Dacheng Tao. Ru-net: Regularized unrolling network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19466, 2022. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 501, + 553, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 545 + ], + "type": "text", + "content": "[17] Kien Nguyen, Subarna Tripathi, Bang Du, Tanaya Guha, and Truong Q Nguyen. In defense of scene graphs for image captioning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1407-1416, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 546, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 546, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 317, + 546, + 553, + 611 + ], + "type": "text", + "content": "[18] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research Journal, pages 1-31, 2024. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 613, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 657 + ], + "type": "text", + "content": "[19] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2856-2865, 2021. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 658, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 714 + ], + "type": "text", + "content": "[20] Ege Özsoy, Evin Pinar Örnek, Ulrich Eck, Tobias Czempiel, Federico Tombari, and Nassir Navab. 4d-or: Semantic scene graphs for or domain modeling. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 475–485. Springer, 2022. 1" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25007" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[21] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 4, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 174 + ], + "type": "text", + "content": "[22] Brigit Schroeder and Subarna Tripathi. Structured query-based image retrieval using scene graphs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 178-179, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 229 + ], + "type": "text", + "content": "[23] Sebastian Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In Proceedings of the fourth workshop on vision and language, pages 70–80, 2015. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "type": "text", + "content": "[24] Sahand Sharifzadeh, Sina Moayed Baharlou, and Volker Tresp. Classification by attention: Scene graph classification with prior knowledge. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5025-5033, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 308 + ], + "type": "text", + "content": "[25] Motoharu Sonogashira, Masaaki Iiyama, and Yasutomo Kawanishi. Towards open-set scene graph generation with unknown objects. IEEE Access, 10:11574-11583, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 310, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 294, + 363 + ], + "type": "text", + "content": "[26] Keisuke Tateno, Federico Tombari, and Nassir Navab. Real-time and scalable incremental segmentation on dense slam. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4465-4472. IEEE, 2015. 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 410 + ], + "type": "text", + "content": "[27] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 465 + ], + "type": "text", + "content": "[28] Johanna Wald, Helisa Dhamo, Nassir Navab, and Federico Tombari. Learning 3d semantic scene graphs from 3d indoor reconstructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3961-3970, 2020. 1, 2, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 468, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 294, + 510 + ], + "type": "text", + "content": "[29] Johanna Wald, Nassir Navab, and Federico Tombari. Learning 3d semantic scene graphs with instance embeddings. International Journal of Computer Vision, 130(3):630-651, 2022. 1, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 577 + ], + "type": "text", + "content": "[30] Ziqin Wang, Bowen Cheng, Lichen Zhao, Dong Xu, Yang Tang, and Lu Sheng. Vl-sat: Visual-linguistic semantics assisted training for 3d semantic scene graph prediction in point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21560–21569, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 634 + ], + "type": "text", + "content": "[31] Shun-Cheng Wu, Johanna Wald, Keisuke Tateno, Nassir Navab, and Federico Tombari. Scenographfusion: Incremental 3d scene graph prediction from rgb-d sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7515-7525, 2021. 1, 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[32] Shun-Cheng Wu, Keisuke Tateno, Nassir Navab, and Federico Tombari. Incremental 3d semantic scene graph prediction from rgb sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5064-5074, 2023. 1, 2, 3, 4, 6, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[33] Danfei Xu, Yuke Zhu, Christopher B Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In Pro" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 577 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "ceedings of the IEEE conference on computer vision and pattern recognition, pages 5410-5419, 2017. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[34] Ning Xu, An-An Liu, Jing Liu, Weizhi Nie, and Yuting Su. Scene graph captioner: Image captioning based on structural visual representation. Journal of Visual Communication and Image Representation, 58:477-485, 2019. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "text", + "content": "[35] Jianwei Yang, Jiasen Lu, Stefan Lee, Dhruv Batra, and Devi Parikh. Graph r-cnn for scene graph generation. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 186, + 555, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 555, + 240 + ], + "type": "text", + "content": "[36] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "text", + "content": "[37] Xu Yang, Kaihua Tang, Hanwang Zhang, and Jianfei Cai. Auto-encoding scene graphs for image captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10685–10694, 2019. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "type": "text", + "content": "[38] Sangwoong Yoon, Woo Young Kang, Sungwook Jeon, SeongEun Lee, Changjin Han, Jonghun Park, and Eun-Sol Kim. Image-to-image retrieval by learning similarity between scene graphs. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10718-10726, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 342, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 396 + ], + "type": "text", + "content": "[39] Kun Yuan, Manasi Kattel, Joel L Lavanchy, Nassir Navab, Vinkle Srivastav, and Nicolas Padoy. Advancing surgical vqa with scene graph knowledge. International Journal of Computer Assisted Radiology and Surgery, 19(7):1409-1417, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 398, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 398, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 398, + 553, + 442 + ], + "type": "text", + "content": "[40] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural motifs: Scene graph parsing with global context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5831-5840, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 443, + 553, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 443, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 316, + 443, + 553, + 496 + ], + "type": "text", + "content": "[41] Chaoyi Zhang, Jianhui Yu, Yang Song, and Weidong Cai. Exploiting edge-oriented reasoning for 3d point-based scene graph analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9705-9715, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 498, + 553, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 541 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 541 + ], + "type": "text", + "content": "[42] Shoulong Zhang, Aimin Hao, Hong Qin, et al. Knowledge-inspired 3d scene graph prediction in point cloud. Advances in Neural Information Processing Systems, 34:18620-18632, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 544, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 577 + ], + "type": "text", + "content": "[43] Zhongping Zhang, Huiwen He, Bryan A Plummer, Zhenyu Liao, and Huayan Wang. Complex scene image editing by scene graph comprehension. In BMVC, 2023. 1" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25008" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/StealthAttack_ Robust 3D Gaussian Splatting Poisoning via Density-Guided Illusions/b6c26c82-2907-466d-9aed-934611879ad9_content_list.json b/2025/StealthAttack_ Robust 3D Gaussian Splatting Poisoning via Density-Guided Illusions/b6c26c82-2907-466d-9aed-934611879ad9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..aea494573664f91d94e8d09104a24d6e88faad72 --- /dev/null +++ b/2025/StealthAttack_ Robust 3D Gaussian Splatting Poisoning via Density-Guided Illusions/b6c26c82-2907-466d-9aed-934611879ad9_content_list.json @@ -0,0 +1,1733 @@ +[ + { + "type": "text", + "text": "StealthAttack: Robust 3D Gaussian Splatting Poisoning via Density-Guided Illusions", + "text_level": 1, + "bbox": [ + 148, + 130, + 849, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bo-Hsu Ke You-Zhe Xie Yu-Lun Liu Wei-Chen Chiu", + "bbox": [ + 269, + 203, + 750, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "National Yang Ming Chiao Tung University", + "bbox": [ + 333, + 228, + 684, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 281, + 326, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D scene representation methods like Neural Radience Fields (NeRF) and 3D Gaussian Splatting (3DGS) have significantly advanced novel view synthesis. As these methods become prevalent, addressing their vulnerabilities becomes critical. We analyze 3DGS robustness against image-level poisoning attacks and propose a novel density-guided poisoning method. Our method strategically injects Gaussian points into low-density regions identified via Kernel Density Estimation (KDE), embedding viewpoint-dependent illusory objects clearly visible from poisoned views while minimally affecting innocent views. Additionally, we introduce an adaptive noise strategy to disrupt multi-view consistency, further enhancing attack effectiveness. We propose a KDE-based evaluation protocol to assess attack difficulty systematically, enabling objective benchmarking for future research. Extensive experiments demonstrate our method's superior performance compared to state-of-the-art techniques. Project page: https://hentci.github.io/stealthattack/", + "bbox": [ + 89, + 314, + 485, + 587 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 632, + 220, + 647 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D scene representation methods, such as Neural Radiance Fields (NeRF)[71] and 3D Gaussian Splatting (3DGS)[42], have significantly advanced novel view synthesis, accurately modeling complex scene geometry and appearance. Along with their popularity, the protection of 3D digital content encoded in these representations has become a matter of concern, where we have witnessed the corresponding watermarking or steganography techniques being proposed in recent years. For instance, GaussianMarker [35] and 3DGSW [38] embed watermarks (mainly binary messages) into Gaussian parameters of 3DGS with minimal visual impact, coupled with dedicated decoders to decipher the hidden messages. Moreover, embedding or hiding extraneous information/messages into 3D scene representations is also directly connected to the risk of data poisoning (where the extraneous information as the poison appears in the training data", + "bbox": [ + 89, + 657, + 485, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/097bb94581b239f08eff96dac2236fd0d5e7ecd172d3872445a90e5f15c8cfde.jpg", + "image_caption": [ + "Figure 1. Illustration of our proposed Density-Guided Poisoning Attack for 3D Gaussian Splatting (3DGS). Our method strategically distribute the Gaussian points of the illusory object (i.e. the red vehicle) among the low-density regions which are discovered along the rays casted from the virtual camera of the poisoned view (i.e. the target view that we would like to attack), making the illusory object clearly visible from the poisoned view while having the minimal interference for the rendering quality on the other non-target/innocent views." + ], + "image_footnote": [], + "bbox": [ + 517, + 277, + 906, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and is encoded into the 3D representations during model training), in which the further utilization or visualization of the poisoned 3D representations would lead to abnormal or even malicious model behaviours (i.e. the negative impact stemmed from the poison is triggered). To this end, in this work, we focus on the investigation of poisoning attacks on 3D scene representation methods, as they become integral to safety-critical applications. Hence, addressing their security vulnerabilities is not only of utmost importance but also urgent.", + "bbox": [ + 511, + 623, + 908, + 775 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While there exists a prior work of studying the poisoning attack upon NeRF, i.e. IPA-NeRF [40] which effectively exploits NeRF's implicit representations to embed targeted visual illusions (i.e. the illusion as poison will appear while rendering the scene from a certain viewing direction), its applicability to explicit 3D scene representations (particularly 3DGS) however remains limited and not directly transferable. Considering the rapidly growing application scenarios", + "bbox": [ + 511, + 779, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27400", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of 3DGS (thanks to its capability of fast rendering and accurately capturing scene geometry), we devote our research effort to realizing poisoning attacks on 3D Gaussian splatt. To the best of our knowledge, this is the first work of its kind. While concurrent work, Poison-Splat [62] focuses on computational cost attacks, our work targets visible illusion embedding. In particular, we would like to inject the visible illusory objects (i.e., poison) onto a target view (named as poisoned view) while keeping the other non-target views (named as innocent views) unaffected, as shown in Figure 1. Our work starts from conducting an investigation (cf. Figure 2) upon the robustness of 3DGS against the prior image-level poisoning methods such as IPA-NeRF, where we find that the attempts of directly adopting IPA-NeRF's approach or naively injecting illusory content into training images easily fail, as 3DGS's inherent multi-view consistency and densification processes effectively neutralize or significantly weaken these attacks.", + "bbox": [ + 89, + 90, + 483, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by the aforementioned investigation, we propose a density-guided poisoning method for 3DGS. Our approach (cf. Figure 1) strategically identifies low-density regions in the initial Gaussian point cloud using Kernel Density Estimation (KDE), in which the points of illusory objects are then distributed among the low-density regions along the rays casting from the virtual camera of target view (i.e. the rays are casted from the virtual camera with the target viewing direction). These points effectively embed illusory objects which would be clearly visible from targeted views, while having minimal impact on other innocent views (i.e. being less perceptible). Moreover, we introduce the adaptive Gaussian noise into innocent views during training for disrupting the property of multi-view consistency in 3DGS, in order to further enhance the overall efficacy of the attack. We conduct extensive experiments, and the results demonstrate the consistent superiority of our proposed poisoning method in comparison to several baselines. The contribution of our work can be summarized as follows:", + "bbox": [ + 91, + 363, + 486, + 650 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We are the first work to address data poisoning attacks upon 3D Gaussian Splatting for illusory objects injection.", + "- We identify and analyze the robustness of 3DGS against the prior poisoning attack techniques.", + "- We propose a density-guided poisoning method tailored for 3DGS, introducing adaptive noise scheduling to disrupt the multi-view consistency of 3DGS and better realize the entire attack." + ], + "bbox": [ + 109, + 656, + 483, + 804 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 825, + 232, + 840 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adversarial Attack. Adversarial attacks [8, 27, 54, 66, 79, 86, 90] are a critical research area in machine learning and computer vision [1, 9, 107]. These attacks exploit", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e48a29e7ea137b218d8d0ed871da30b110101e26111fb35634de8036813eb3de.jpg", + "image_caption": [ + "Figure 2. Limitations of existing poisoning methods on 3DGS. Existing poisoning methods (e.g., IPA-NeRF [40] designed for NeRF or the one adapted to 3DGS, denoted as IPA-Splat) produce weak or absent illusions due to 3DGS's robustness and multi-view consistency. In contrast, our proposed approach successfully injects clearly visible illusory objects (i.e., the dog)." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ML model vulnerabilities by creating inputs that cause misclassification while appearing normal to humans. Goodfellow et al. [27] introduced FGSM, showing how small perturbations could transform panda images into gibbons, while Madry et al. [66] established PGD as a stronger iterative method. Adversarial attacks divide into black-box attacks [6, 13, 18, 30, 36, 75, 108], which operate without model access, and white-box attacks [47, 53, 72, 95, 101], which use full model knowledge. Black-box methods include transfer-based attacks [20, 58] exploiting cross-model transferability and query-based approaches [2, 12] that iteratively refine perturbations. White-box methods like Carlini-Wagner attacks [8] formulate adversarial generation as optimization problems, achieving high success rates while maintaining imperceptibility. Recent research has expanded to complex domains including object detection [70, 99], semantic segmentation [3], and 3D point clouds [56, 96]. Our approach resembles black-box attacks, modifying point clouds and applying simple perturbations to input images, extending adversarial concepts to 3D Gaussian Splitting representations.", + "bbox": [ + 511, + 319, + 908, + 623 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Data Poisoning. Data poisoning [14, 26, 80, 98] represents a critical vulnerability in machine learning systems. Unlike adversarial attacks targeting inference, poisoning attacks manipulate training by injecting crafted samples that exploit learning mechanisms, causing defective statistical distributions [7, 37]. Two primary poisoning strategies exist: creating indistinguishable malicious samples that blend with normal data [85, 91, 98], or employing surgical precision by polluting minimal data subsets [39, 80]. These approaches prove effective across computer vision [34, 105], NLP [48, 92], and recommender systems [23, 50]. They categorize as availability attacks degrading overall performance [74, 83] or integrity attacks targeting specific inputs [14, 29]. More sophisticated approaches formulate poisoning as bi-level optimization problems [46, 52, 59, 61], allowing attackers to optimize poisoned samples while anticipating victim model behavior. Common defenses include robust statistics [19, 89], data sanitization [17, 76], and differential privacy [65]. Our", + "bbox": [ + 511, + 628, + 910, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27401", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "work employs the surgical precision strategy, conducting a stealth attack with minimal contaminated data while exploring 3D Gaussian Splatting's unique vulnerabilities to poisoning.", + "bbox": [ + 89, + 90, + 485, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural Rendering. Novel view synthesis (NVS)[10, 24, 88, 109] has evolved from traditional graphics to learning-based methods. Neural Radiance Field (NeRF)[71] and 3D Gaussian Splatting (3DGS)[42] have transformed 3D scene representation. NeRF uses MLP-based networks to implicitly represent 3D scenes, leveraging differential alpha blending[43, 77] for high-quality volumetric rendering [41, 68]. Extensions like Mip-NeRF [4], Instant-NGP [73], and NeRF-W [67] address anti-aliasing, speed, and real-world limitations. Recent advances include few-shot synthesis [55], MVS-based approaches for large scenes [84], and improved robustness for dynamic content [60]. Robustness enhancements include progressive optimization [69] and joint camera pose optimization [11]. However, NeRF's implicit representation lacks explicit geometric constraints, creating vulnerabilities to view-specific perturbations [57, 93]. IPA-NeRF [40] exploits this for data poisoning attacks. In contrast, 3DGS [42] uses explicit representation with discrete 3D Gaussian primitives inspired by point-based rendering [28, 110]. This provides stronger geometry constraints and multi-view consistency with superior rendering efficiency [15, 102]. 3DGS has expanded to dynamic scenes [63, 100], human avatars [45, 49], efficient implementations [16, 22], compression [104], robustness for unconstrained images [33], and specular reconstruction [21]. Our work addresses the challenge of attacking 3DGS's robust explicit representation, contributing to understanding security implications in modern neural rendering techniques.", + "bbox": [ + 91, + 155, + 486, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data Poisoning on Neural Rendering. As neural rendering gains adoption, security vulnerabilities have attracted attention. IPA-NeRF [40] pioneered poisoning attacks against NeRF by inserting crafted samples at specific viewing angles, formulating data poisoning as bi-level optimization. Lu et al.'s Poison-Splat [62] targeted 3DGS efficiency by generating samples that dramatically increase memory consumption, demonstrating attacks targeting resource utilization [81] rather than accuracy. The security landscape extends to privacy concerns [64] and adversarial examples [25, 32, 103]. Zeybey et al. [103] showed how adversarial noise in 3D objects misleads vision-language models like CLIP [78]. Song et al.'s Geometry Cloak [82] prevents unauthorized 3D reconstruction from copyrighted images, while security studies examine point clouds [31, 56], meshes [97], and broader vision systems [1, 9]. Similar to steganography approaches embedding hidden information [51, 106], our method injects view-dependent content but enables decoder-free extraction through standard rendering from specific viewpoints. Our work introduces a density-guided attack methodology targeting 3DGS's initial point cloud prior, exploiting its robustness", + "bbox": [ + 89, + 583, + 486, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "characteristics. We propose an evaluation protocol based on scene density analysis to identify optimal positions for poison injection, contributing insights that could inform future defense mechanisms against such attacks.", + "bbox": [ + 511, + 90, + 908, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 162, + 604, + 180 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 511, + 188, + 718, + 204 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a dataset $\\mathcal{D}$ composed of multiple images $\\{I_k\\}_{k=1}^N$ viewing a scene $\\mathcal{E}$ from different viewing directions, 3D Gaussian Splatting is originally proposed to construct a 3D Gaussian point cloud $G$ (each Gaussian has properties in terms of position, covariance, opacity, and color factors) for representing the 3D scene $\\mathcal{E}$ , where we are able to render the image observation of $\\mathcal{E}$ from any arbitrary view via projection and differentiable tile rasterizer. Basically, the goal of our poisoning attack upon 3DGS is to inject illusory/poison object $O_{\\mathrm{ILL}}$ onto the target view $v_{\\mathbf{p}}$ , where we denoted the resultant Gaussian point cloud after being poisoned as $\\tilde{G}$ , such that the image $\\tilde{I}_{\\mathrm{ILL}} = \\Re(\\tilde{G}, v_{\\mathbf{p}})$ obtained by rendering $\\tilde{G}$ from $v_{\\mathbf{p}}$ would contain $O_{\\mathrm{ILL}}$ while keeping the images of $\\tilde{G}$ rendered from any other views $v_{k}$ (i.e. non-target views $v_{k} \\neq v_{\\mathbf{p}}$ ) being identical to the ones of $G$ of the same view $v_{k}$ . Please note that, when we denote the image of $G$ rendered from $v_{\\mathbf{p}}$ as $\\Re(G, v_{\\mathbf{p}})$ , the ideal appearance of $\\tilde{I}_{\\mathrm{ILL}}$ should be the combination of $\\Re(G, v_{\\mathbf{p}})$ and $O_{\\mathrm{ILL}}$ (in which such combination is denoted as $I_{\\mathrm{ILL}}$ ). The core objective of our 3DGS poisoning is formulated as:", + "bbox": [ + 511, + 209, + 908, + 517 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\tilde {G}} \\| \\tilde {I} _ {\\mathbf {I L L}} - I _ {\\mathbf {I L L}} \\| _ {2} ^ {2} + \\sum_ {v _ {k} \\neq v _ {\\mathbf {p}}} \\| \\Re (\\tilde {G}, v _ {k}) - \\Re (G, v _ {k}) \\| _ {2} ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 525, + 908, + 560 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve this under different threat models, we propose two strategies: density-guided point cloud attack (Section 3.3) following classical data poisoning, and view consistency disruption (Section 3.4) as a backdoor attack with minimal training modification via noise scheduling.", + "bbox": [ + 511, + 568, + 908, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Potential Naive Approaches and Limitations", + "text_level": 1, + "bbox": [ + 511, + 652, + 888, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first explore two potential naive approaches, followed by discussing their limitations and motivating our attack design: 1) Directly injecting the illusory object onto the training image in $\\mathcal{D}$ of the target view $v_{p}$ and following the typical training procedure of 3DGS would easily fail since the property of multi-view consistency in 3DGS would treat the illusory object as the noise (i.e. which violates the consistency) and eliminate it from the resultant Gaussian cloud.", + "bbox": [ + 511, + 674, + 908, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2) Directly backprojecting the illusory object into the Gaussian point cloud $G$ reconstructed from $\\mathcal{D}$ also faces the challenge of determining proper depth for the illusory object, otherwise it could be occluded by existing geometry in $G$ .", + "bbox": [ + 511, + 795, + 908, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Our Density-Guided Point Cloud Attack", + "text_level": 1, + "bbox": [ + 511, + 863, + 859, + 878 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With learning the lessons from the aforementioned naive", + "bbox": [ + 511, + 885, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27402", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5c3837895aada4145de0d50d445ea9dfb6ee805f9afb3c540432ceba479f136b.jpg", + "image_caption": [ + "(a) Density-guided Point Cloud Attack" + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 390, + 250 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/943cdd96767ca9aba11a37f3c3269df6f9c8c4ac9ee0c050fcbdd68690929f6e.jpg", + "image_caption": [ + "(b) Normal 3DGS Training Iterations" + ], + "image_footnote": [], + "bbox": [ + 392, + 88, + 651, + 250 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b109103375f0e9e80568ea1b3a8e1c760d869f44751e18961723b7226aa18fd1.jpg", + "image_caption": [ + "(c) View Consistency Disruption Attack" + ], + "image_footnote": [], + "bbox": [ + 651, + 89, + 903, + 250 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a1972bfd8fe7408d8735a0483626eab09b86aa6ab9e516355880c04b30c8067d.jpg", + "image_caption": [ + "Figure 3. Overview of our proposed poisoning attack framework. Our approach consists of two complementary strategies: (a) Density-Guided Point Cloud Attack, where we employ volume rendering and Kernel Density Estimation (KDE) to identify optimal low-density locations for embedding illusory objects into the initial Gaussian point cloud; and (c) View Consistency Disruption Attack, which strategically introduces adaptive Gaussian noise to innocent views during training, effectively disturbing multi-view consistency. (b) illustrates the standard 3D Gaussian Splitting (3DGS) training pipeline for reference. The combined strategies successfully inject convincing illusions from poisoned views while maintaining high fidelity in innocent viewpoints.", + "Figure 4. Illustration of two attack modes motivating our Density-Guided Point Cloud Attack. (a) Points placed outside the coverage of innocent viewpoints can effectively embed illusions visible only from the poisoned view. (b) Points occluded from innocent viewpoints also provide viable hidden locations. These scenarios motivate our Density-Guided strategy for robust and stealthy attacks." + ], + "image_footnote": [], + "bbox": [ + 94, + 369, + 478, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "approaches, we conclude that effective 3DGS poisoning must consider the inherent properties of 3DGS's explicit representation and multi-view consistency, while ensuring illusion visibility on the target view and minimal perception on innocent views. To this end, we propose two simple but effective ideas for discovering optimal 3D positions in $G$ to place backprojected Gaussian points of the illusory object, as illustrated in Figure 4: First, poison points can be placed in regions invisible to innocent views; Second, regions occluded for innocent views by existing geometry in $G$ are effective candidates (visible for target view but invisible for innocent views due to occluders). To identify regions satisfying these ideas, we propose a density-guided point placement strategy, detailed below.", + "bbox": [ + 88, + 595, + 482, + 805 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scene Space Analysis. Given the Gaussian point cloud $G$ (reconstructed from $\\mathcal{D}$ via the typical 3DGS procedure), we start from determining the Axis-Aligned Bounding Box (AABB) of $G$ (basically, finding the minimum and maximum coordinates across all points in $G$ ) and creating a rectangular box that fully encloses the entire 3D scene $\\mathcal{E}$ in $G$ , with its", + "bbox": [ + 88, + 809, + 483, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "edges being aligned to the coordinate axes. Such bounding box is then first decomposed/voxelized into a uniform grid $S$ , where we denote each cell in the grid as $s$ .", + "bbox": [ + 511, + 371, + 903, + 415 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As the opacity $\\alpha(g)$ of each Gaussian point $g$ in $G$ can be estimate via volume rendering technique (depending on our target view $v_{\\mathbf{p}}$ ), we can easily compute the density $\\rho(s)$ of each voxel $t$ by $\\sum_{g \\in s} \\alpha(g)$ as the opacity and density are highly related, where $g \\in s$ indicates that Gaussian point $g$ is located within voxel $s$ .", + "bbox": [ + 511, + 416, + 905, + 505 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Continuous Density Estimation. Based on the per-voxel density $\\rho(s)$ , we apply Gaussian Kernel Density Estimation (KDE) to obtain a continuous density estimate for any arbitrary 3D position $x$ :", + "bbox": [ + 511, + 510, + 906, + 571 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf (x) = \\frac {1}{| \\mathcal {S} |} \\sum_ {s \\in \\mathcal {S}} K _ {h} (x - c (s)) \\cdot \\rho (s), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 580, + 906, + 616 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c(s)$ denotes the centroid of a voxel $s$ and $K_{h}$ is the Gaussian kernel with bandwidth $h$ :", + "bbox": [ + 511, + 627, + 905, + 656 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nK _ {h} (x) = \\frac {1}{(2 \\pi h ^ {2}) ^ {3 / 2}} \\exp \\left(- \\frac {\\| x \\| ^ {2}}{2 h ^ {2}}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 664, + 906, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Optimal Position Selection. With the illusory object placed on the image plane of the virtual camera rendering the Gaussian point cloud from target view $v_{\\mathbf{p}}$ , backprojection starts by casting rays from camera center $C$ through all pixels of the illusory object. We sample points along each ray to find regions in the Gaussian cloud with minimum density, via querying the KDE result described above. Given a casting ray with direction $d$ , any sampled 3D position along the ray can be written as $C + t \\cdot d$ , where $t$ ranges from $t_{\\mathrm{min}}$ to $t_{\\mathrm{max}}$ . We set $t_{\\mathrm{min}}$ to 0.3 (as points near camera often appear as floaters in 3DGS optimization) and $t_{\\mathrm{max}}$ represents the original scene depth at each pixel in the poisoned view.", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27403", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The sampled point in the minimum density region can be computed by:", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nx _ {\\min } = \\underset {x \\in C + t \\cdot d, t \\in [ t _ {\\min }, t _ {\\max } ]} {\\arg \\min } f (x). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 131, + 483, + 159 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We then insert new Gaussian poison points at position $x_{\\mathrm{min}}$ , assigning colors from the illusory object (i.e. the color value for the new Gaussian point is obtained from the corresponding illusory object pixel). Our proposed density-guided method hence strategically places Gaussian points of poison to embed the illusory object prominently from the poisoned view while minimizing its visibility from innocent views.", + "bbox": [ + 89, + 166, + 483, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. View Consistency Disruption Attack", + "text_level": 1, + "bbox": [ + 89, + 280, + 406, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While our Density-Guided Point Cloud Attack (cf. Section 3.3) effectively places Gaussian points of poison in many cases, scenes with high view overlap (i.e. the field-of-views of the training image $\\{I_k\\}_{k=1}^N$ in $\\mathcal{D}$ have high overlaps) remain challenging. To address this, we introduce the View Consistency Disruption Attack, which strategically adds controlled noise to innocent views, thus weakening the multi-view consistency of 3DGS and better preserving our injected illusions.", + "bbox": [ + 89, + 301, + 483, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We selectively apply Gaussian noise to innocent views, leaving the poisoned view clean. For a training image $I_{k}$ with view direction $v_{k}$ , the noise is applied as:", + "bbox": [ + 89, + 438, + 483, + 484 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI _ {k} ^ {\\prime} = \\mathbf {1} _ {v _ {k} = v _ {\\mathbf {p}}} \\cdot I _ {k} + \\mathbf {1} _ {v _ {k} \\neq v _ {\\mathbf {p}}} \\cdot \\operatorname {C L I P} (I _ {k} + \\eta), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 493, + 483, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{1}$ is an indicator function, CLIP prevents $I_{k} + \\eta$ from exceeding the pixel value range, and $\\eta \\sim \\mathcal{N}(0,\\sigma_t^2)$ denotes noise with strength $\\sigma_{t}$ adjusted according to 3DGS iteration. The $\\sigma_{t}$ scheduling follows the principle of having strong noise in early 3DGS optimization to disrupt multi-view consistency (as noise injected into training images are independent) while gradually reducing noise strength to maintain high-quality reconstruction for innocent views in late optimization. We explore three noise decay strategies:", + "bbox": [ + 89, + 518, + 483, + 654 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {\\text {l i n e a r}} (t) = \\sigma_ {0} \\cdot \\left(1 - \\frac {t}{T}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 662, + 483, + 691 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {\\text {c o s i n e}} (t) = \\sigma_ {0} \\cdot \\cos \\left(\\frac {\\pi t}{2 T}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 694, + 483, + 722 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {\\mathrm {s q r t}} (t) = \\sigma_ {0} \\cdot \\sqrt {1 - \\frac {t}{T}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 209, + 723, + 483, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma_0$ is the initial noise strength, $t$ is the current iteration, and $T$ is the total training iterations. Linear decay reduces noise evenly, cosine decay provides a smooth start and accelerates later reduction, and square root decay maintains higher noise longer before a rapid decrease.", + "bbox": [ + 89, + 763, + 483, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our controlled noise injection creates intentional imbalance during training, enabling preservation of illusory objects from the poisoned viewpoint while ensuring scene fidelity as noise diminishes. The overview of our proposed", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "method, composed of both density-guided point cloud attack (cf. Section 3.3) and view consistency disruption attack (cf. Section 3.4), is provided in Figure 3.", + "bbox": [ + 511, + 90, + 906, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 152, + 645, + 169 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 513, + 178, + 702, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate our method on three common datasets: (1) Mip-NeRF360 [5] with complex $360^{\\circ}$ scenes, (2) Tanks & Temples [44] containing realistic indoor and outdoor captures, and (3) Free [94], featuring unbounded scenes with free camera trajectories. These datasets provide diverse benchmarks for novel view synthesis evaluation.", + "bbox": [ + 511, + 205, + 905, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compared Methods. We evaluate our method against three baselines: (1) IPA-NeRF (Nerfacto) [40]: The original backdoor attack applied to Nerfacto [87], featuring advanced static scene reconstruction techniques. (2) IPA-NeRF (Instant-NGP) [40]: The original backdoor attack on Instant-NGP [73], known for accelerated training and rendering. (3) IPA-Splat: Our adaptation of IPA-NeRF specifically for 3D Gaussian Splatting.", + "bbox": [ + 511, + 300, + 906, + 420 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For IPA-NeRF baselines, we maintain original settings but reduce total iterations to $O = 15,000$ for faster convergence. Other parameters remain unchanged: $T = 200$ iterations per epoch, $O / T = 75$ attack epochs, $A = 10$ attack iterations per epoch, $K = 100$ perturbation renderings, distortion budget $\\epsilon = 32$ , constraint parameter $\\eta = 1$ , and view constraints $(13^{\\circ}$ and $15^{\\circ}$ ).", + "bbox": [ + 511, + 421, + 906, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For our IPA-Splat method, we adapt IPA-NeRF to 3D Gaussian Splatting with $O = 30,000$ total iterations and $T = 200$ normal training iterations per epoch, resulting in $O / T = 150$ epochs with $A = 10$ attack iterations each. Other settings match IPA-NeRF. Due to 3DGS's explicit representation, we implement separate parameter constraints (xyz coordinates, feature vectors, scaling, rotation, and opacity) for precise control.", + "bbox": [ + 511, + 527, + 906, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following IPA-NeRF [40], we evaluate using PSNR, SSIM, and LPIPS metrics on two view sets: (1) V-ILLUSORY, focusing on masked metrics for illusory objects, and (2) V-TEST, assessing performance on unseen viewpoints. Attack success is defined as achieving PSNR $> 25$ on V-ILLUSORY while maintaining V-TEST PSNR drop $\\leq 3$ , ensuring effective illusion generation and preserved innocent view quality.", + "bbox": [ + 511, + 652, + 905, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Protocol. Our evaluation accounts for varying attack difficulty across camera positions. As shown in Figure 5, datasets exhibit distinct camera patterns. For uniform datasets like Mip-NeRF 360 (e.g., the \"bicycle\" scene) and Tanks-and-Temple, we select the median-indexed frame as the attack viewpoint.", + "bbox": [ + 511, + 779, + 906, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For Free dataset scenes with irregular camera trajectories (e.g., the \"stair\" scene), we quantify varying attack difficulty", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27404", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8ac6d8aefe7938395711ce139d2c170f800a2745b28e275cd41f10547fe907a8.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons on single-view attack. Metrics evaluated on Mip-NeRF 360 [5], Tanks & Temples [44], and Free datasets. Our method significantly outperforms baseline attacks in embedding illusory objects (V-IILLUSORY) while maintaining high fidelity in other views (V-TEST)." + ], + "table_footnote": [], + "table_body": "
MethodMip-NeRF 360 [5] datasetTanks & Temples [44] datasetFree [94] dataset
V-IILLUSORYV-TESTV-IILLUSORYV-TESTV-IILLUSORYV-TEST
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Naive 3DGS (w/o attack)13.210.5210.73129.450.8830.16513.150.6160.73230.600.9150.13512.000.3150.90526.800.8260.228
IPA-NeRF [40] (Nerfacto [87])16.000.5820.68521.940.5860.41513.510.6360.71123.880.7300.21813.930.4430.69920.280.4970.532
IPA-NeRF [40] (Instant-NGP [73])17.600.6180.64120.000.5170.47916.050.6930.61620.290.6690.35018.940.5080.51920.430.5030.548
IPA-Splat13.230.5180.74027.390.8290.24713.430.6250.72428.530.8910.19012.600.3720.74424.710.7490.341
Ours27.040.8130.36927.760.8050.28621.330.8090.37127.580.8520.23926.660.7540.31725.250.7280.382
", + "bbox": [ + 93, + 132, + 906, + 234 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5c9f07d2003fc2392168871a9e7ad08bfd7dff760543fae3d5707438d2556549.jpg", + "table_caption": [ + "Table 2. Quantitative comparisons on single-view attack with different difficulty levels on the Free [94] dataset. We evaluate attack effectiveness (V-ILLUSORY) at varying difficulty levels, defined by our KDE-based evaluation protocol. Our method consistently achieves superior results across all metrics, clearly outperforming state-of-the-art methods, especially in EASY and MEDIAN scenarios, while remaining effective in the challenging HARD scenario." + ], + "table_footnote": [], + "table_body": "
MethodEASYMEDIANHARD
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
IPA-NeRF (nerfacto)15.040.4820.66213.930.4430.69914.250.4500.728
IPA-NeRF (instant-ngp)18.170.5180.54118.940.5080.51917.950.4870.557
IPA-Splat13.940.4790.65812.600.3720.74313.060.3400.796
Ours29.940.8530.18826.660.7540.31717.530.5260.581
", + "bbox": [ + 156, + 304, + 839, + 398 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c70dbf187a8f9d618d7c54ede62f6ac63d0682c8b495aa2859a408a6cb17e5e5.jpg", + "image_caption": [ + "Figure 5. Our evaluation protocol. We evaluate two scenes with varying difficulties. Left: The \"bicycle\" scene (Mip-NeRF 360 [5]) has uniform camera coverage, providing similar difficulty across views. Right: The \"stair\" scene (Free [94]) has increasing difficulty as later views are visible from more prior viewpoints." + ], + "image_footnote": [], + "bbox": [ + 93, + 419, + 480, + 553 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "using our KDE-based protocol:", + "bbox": [ + 89, + 659, + 302, + 674 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Compute overall scene density distribution using KDE.", + "2. Calculate camera viewpoint densities within their FOV using camera intrinsics and a $10\\%$ sampling radius.", + "3. Sort cameras by density, select three representative viewpoints:" + ], + "bbox": [ + 89, + 679, + 483, + 753 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "EASY: Minimum density (lowest coverage)", + "- MEDIAN: Median density (average coverage)", + "- HARD: Maximum density (highest coverage)" + ], + "bbox": [ + 132, + 761, + 447, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experiments in Tab. 2 confirm negative correlation between scene density and attack success (V-IILLUSORY), validating that higher scene coverage increases attack difficulty. This protocol enables fair evaluation across scenes and provides a benchmark for future 3DGS poisoning research.", + "bbox": [ + 89, + 825, + 485, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/70ca45540e64c8eb7f57c76830c3950dcde8df4cedc7d5fc9cba4ddba2efb2c3.jpg", + "image_caption": [ + "Figure 6. Qualitative comparisons on single-view attack. Our method generates significantly clearer and more convincing illusory objects from the poisoned viewpoint, demonstrating better multiview consistency and fewer artifacts compared to other state-of-the-art methods." + ], + "image_footnote": [], + "bbox": [ + 516, + 422, + 903, + 555 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5aceb08329024fc42e7b73406b283418722f50762b03c7b502820d50bb85edf3.jpg", + "image_caption": [ + "Figure 7. Qualitative comparisons of different poisoning methods on innocent views. Unlike baseline methods, our method effectively maintains high visual fidelity across innocent viewpoints, introducing significantly fewer artifacts and ensuring minimal disruption of the original scene appearance." + ], + "image_footnote": [], + "bbox": [ + 516, + 647, + 903, + 781 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27405", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/997f2dde5faf1f2c68ea31e45b4dfd6a967b889b223bc1d337c31abf93dbc1cf.jpg", + "table_caption": [ + "Table 3. Multi-view Attack Evaluation. We quantitatively evaluate our method on multiple poisoned views (V-ILLUSORY) and innocent views (V-TEST). Our approach consistently outperforms state-of-the-art methods, embedding clear illusions in targeted views while maintaining high fidelity in innocent views." + ], + "table_footnote": [], + "table_body": "
# of viewsMethodV-ILLUSORY (poisoned avg.)V-TEST
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
2IPA-NeRF (Nerfacto)16.170.5830.68019.640.4570.548
IPA-NeRF (Instant-NGP)19.190.6240.61618.390.4400.539
IPA-Splat13.240.4970.75227.450.8320.243
Ours27.490.8420.29927.770.8040.286
3IPA-NeRF (Nerfacto)18.480.5840.66019.830.4620.545
IPA-NeRF (Instant-NGP)18.090.6040.64318.630.4580.524
IPA-Splat13.760.5380.73227.970.8580.223
Ours27.040.8330.31127.720.8030.287
4IPA-NeRF (Nerfacto)17.060.6260.67619.610.4670.538
IPA-NeRF (Instant-NGP)19.060.6570.63218.510.4580.523
IPA-Splat13.090.4890.79627.600.8380.228
Ours26.950.8550.30527.590.8020.287
", + "bbox": [ + 91, + 159, + 483, + 301 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e0b33b6899861cd9b605d9dc39e467727c5efc879babeebe7d37d823462db10d.jpg", + "image_caption": [ + "Figure 8. Qualitative comparisons on multi-view attack. Our density-guided method produces sharper, more consistent illusions across multiple poisoned views, clearly outperforming baseline methods that yield faint or inconsistent results." + ], + "image_footnote": [], + "bbox": [ + 94, + 323, + 480, + 496 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Single-view Attack", + "text_level": 1, + "bbox": [ + 89, + 590, + 272, + 607 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 1 shows quantitative comparisons of single-view attacks. Our method outperforms baselines across all datasets. On V-ILLUSORY views, our approach significantly improves PSNR, SSIM, and LPIPS, effectively embedding convincing illusions. Performance on V-TEST remains consistently high, indicating minimal impact on innocent views.", + "bbox": [ + 89, + 618, + 485, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative results (Figs. 6 and 7) further highlight our advantages. In Fig. 6, our method produces clearer and more convincing illusory objects compared to baselines, which often yield faint or inconsistent illusions. Fig. 7 emphasizes our superior visual fidelity in innocent views, with fewer artifacts and better overall scene quality.", + "bbox": [ + 89, + 714, + 483, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Tab. 2, we analyze performance across difficulty levels (EASY, MEDIAN, HARD) on the Free dataset. As expected, attack effectiveness decreases with difficulty. Nonetheless, our density-guided approach achieves superior results, particularly at EASY and MEDIAN levels, demonstrating robust performance even in challenging conditions.", + "bbox": [ + 89, + 810, + 485, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c6b58f940e43952e32a55b066d2b90e2fc351751ab42bcdf50e55825f576e713.jpg", + "table_caption": [ + "Table 4. Effect of KDE bandwidth $h$ on attack performance. We examine how KDE bandwidth impacts our density-guided attack. A moderate bandwidth ( $h = 7.5$ ) achieves the best balance, maximizing effectiveness on poisoned views while preserving quality in innocent views." + ], + "table_footnote": [], + "table_body": "
Bandwidth hV-ILLUSORYV-TEST
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
0.127.000.8110.37327.830.8050.286
2.526.920.8090.37527.810.8050.286
5.026.950.8110.37527.250.7860.297
7.527.040.8130.36927.760.8050.286
10.026.890.8070.38027.720.8050.286
", + "bbox": [ + 514, + 159, + 906, + 258 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0932271001648d75d8532364bdf2fd18a3b8598900964ba0049090988e445954.jpg", + "table_caption": [ + "Table 5. Effect of noise scheduling parameters. We analyze initial noise strength $\\sigma_0$ and decay strategies. Higher initial noise $(\\sigma_0 = 100)$ with linear decay provides the best balance, maximizing illusory quality (V-ILLUSORY) while preserving fidelity in innocent views (V-TEST)." + ], + "table_footnote": [], + "table_body": "
Initial noise σ0Decay strategyV-ILLUSORYV-TEST
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
30Linear26.470.7950.39828.430.8510.203
30Cosine26.700.8010.38828.300.8450.213
30Square root26.620.7990.39328.270.8450.212
100Linear27.040.8130.36927.760.8050.286
100Cosine26.930.8120.37326.960.7710.315
100Square root26.900.8130.37326.810.7670.319
", + "bbox": [ + 514, + 342, + 905, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Multi-view Attack", + "text_level": 1, + "bbox": [ + 511, + 464, + 691, + 479 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In realistic scenarios, attackers may need to embed illusory objects into multiple viewpoints simultaneously. Unlike single-view attacks, multi-view poisoning balances multiple objectives while preserving scene consistency and fidelity. We rigorously evaluate our approach using the Mip-NeRF 360 dataset [5], poisoned views at $0^{\\circ}$ , $90^{\\circ}$ , $180^{\\circ}$ , and $270^{\\circ}$ with the median view at $0^{\\circ}$ as reference.", + "bbox": [ + 511, + 486, + 908, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate our density-guided method against baselines (IPA-NeRF [40] with Nerfacto, IPA-NeRF with Instant-NGP, IPA-Splat) on the Mip-NeRF-360 dataset under multi-view attacks (2, 3, and 4 poisoned views). Results in Tab. 3 show our approach consistently outperforms baselines, effectively embedding illusory objects (V-ILLUSORY) while minimally affecting innocent views.", + "bbox": [ + 511, + 592, + 908, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 8 qualitatively demonstrates our method's superiority. It consistently generates clear, visually convincing illusions with minimal artifacts, significantly outperforming baselines and maintaining high quality from innocent views.", + "bbox": [ + 511, + 698, + 908, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 768, + 679, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "KDE bandwidth. The KDE bandwidth significantly affects density estimation smoothness. Evaluations (Tab. 4) show medium bandwidth ( $h = 7.5$ ) achieves optimal balance, maximizing V-IILLUSORY effectiveness while preserving V-TEST quality. Smaller bandwidths ( $h = 0.1$ ) overly restrict point placement, while larger bandwidths ( $h = 10.0$ ) decrease attack precision.", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27406", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d7d858e236346b6c42ebfa1bfe2fe9de31a8208e8b3ff9d77e0e9ba42b357fe1.jpg", + "table_caption": [ + "Table 6. Comparison of different attack strategy combinations. We evaluate the impact of each component of our proposed method on poisoning effectiveness. Combining all strategies achieves the best results, significantly improving rendering quality of the illusory object (V-ILLUSORY) and maintaining satisfactory performance in innocent views (V-TEST), resulting in optimal Attack Success Rate (ASR). The combination of point cloud poisoning and noise scheduling is crucial for successful attacks, highlighting their complementary nature." + ], + "table_footnote": [], + "table_body": "
Poisoned View GT ReplacementDensity-Guided Point Cloud AttackView Consistency Disruption AttackV-ILLUSORYV-TESTASR (PSNR)
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓V-ILLUSORY > 25 & V-TEST drop ≤ 3
✓13.220.5210.73029.460.8840.1640/7
✓✓26.010.7750.42729.400.8830.1646/7
✓✓13.310.5220.74727.790.8050.2860/7
✓✓✓27.040.8130.36927.760.8050.2867/7
", + "bbox": [ + 93, + 146, + 903, + 255 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4b276a6bd6b2d176ef1a317d32aa75ab3193898add151222c5904bad82e9ba7a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 277, + 171, + 366 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6fc2c2107f298b3b032bd4ef4099fd40ab1b1cf16a1a32bba6d90ad25ba289ad.jpg", + "image_caption": [ + "Figure 9. Qualitative analysis of attack component combinations. We compare three attack strategies: (1) direct image poisoning, (2) density-guided point cloud poisoning, and (3) multi-view consistency disruption. Combining all three achieves the most realistic illusions across various scenes from the Mip-NeRF 360 [5] dataset, highlighting their complementary effectiveness." + ], + "image_footnote": [], + "bbox": [ + 171, + 277, + 248, + 367 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/53b147586ae134144335c0f8a7bdae34a97b6c366f5b01e6bb12401f759aac53.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 250, + 277, + 325, + 368 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/620656d3e6f56d3dc5206c331e6833a2199d73e123ebd216af6a2139854ebb56.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 326, + 277, + 401, + 368 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8a678fb1441cd47f0ee305568c5e07268231260746b7c44e1d66f67ab2ca1c02.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 401, + 277, + 478, + 368 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Noise scheduling. We evaluated noise scheduling strategies, varying initial noise intensities $(\\sigma_0)$ and decay rates (linear, cosine, and square root), summarized in Tab. 5. Higher initial noise $(\\sigma_0 = 100)$ with slower linear decay achieved optimal balance, greatly enhancing attack effectiveness with moderate impact on innocent views.", + "bbox": [ + 89, + 484, + 483, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Attack components. We analyzed combinations of direct target-view image poisoning, density-guided point cloud poisoning (KDE-based), and noise-based view consistency disruption. Quantitative results (Tab. 6) show that direct image poisoning alone is ineffective. Combining image poisoning with density-guided poisoning notably improves outcomes. Integrating all three components achieves the best results, embedding robust illusions and preserving rendering quality. Qualitative results (Fig. 9) visually confirm the superior clarity and effectiveness of this combined approach.", + "bbox": [ + 89, + 580, + 483, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 752, + 209, + 767 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented a density-guided poisoning method for 3DGS, strategically injecting illusory objects and disrupting multiview consistency via adaptive noise. Experiments show our approach outperforms existing baselines, effectively embedding convincing illusions with minimal impact on innocent views. Our work highlights critical vulnerabilities in 3D representation models, providing a robust framework for future security research.", + "bbox": [ + 89, + 779, + 485, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ac6904c291ae49c4d9122d3a419964a36f6b548d7c42c759b3381d4143a436e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 277, + 901, + 392 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3788b58fac4075bf70ae621fba7e0fffc1ef18f7101bbbb3ddce44e69662445b.jpg", + "image_caption": [ + "Figure 10. Effect on neighboring views (\"counter\" scene, Mip-NeRF 360 [5]). Clockwise and counterclockwise shifts from the attack view (0 degrees) up to 20 degrees show PSNR between clean and poisoned renderings. IPA-NeRF significantly lowers PSNR across most angles, whereas our method mainly impacts views within five degrees, preserving quality beyond this range." + ], + "image_footnote": [], + "bbox": [ + 517, + 393, + 901, + 501 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e89b3d20713e0a837135cbfe406baeacb12f7f74858763e36098b0e585220c9e.jpg", + "image_caption": [ + "Figure 11. Visualization of evaluation protocol (\"grass\" scene, Free [94]). Our method achieves clearly visible illusory objects in EASY and MEDIAN scenarios and maintains robust performance even under challenging HARD conditions." + ], + "image_footnote": [], + "bbox": [ + 516, + 602, + 602, + 729 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bf275cad4ceec9cd93c91d11c0035c166c80f75c700c97ad40c0c47051b5b770.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 602, + 676, + 729 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b6d263976ec595f19403e8f3ec1eddeab024b21d91cfbf64cb5019f842d3bb3a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 602, + 750, + 729 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/408582928a04e3fb6d0f5f2ecd4b05fcee8a7182fae8b9ede3abda540aca5aa7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 602, + 826, + 723 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c652c9b10ac033be566f40d72f80016cff636c63aba390e6c2fbd92c4a417eb7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 828, + 602, + 903, + 723 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Our method struggles with scenes having highly overlapping views or complex camera trajectories due to 3DGS's strict multi-view consistency. Future work should address this balance between consistency and attack effectiveness.", + "bbox": [ + 511, + 815, + 906, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27407", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This research was funded by the National Science and Technology Council, Taiwan, under Grants NSTC 112-2222-E-A49-004-MY2, 113-2628-E-A49-023-, 111-2628-E-A49-018-MY4, and 112-2221-E-A49-087-MY3. The authors are grateful to Google, NVIDIA, and MediaTek Inc. for their generous donations. Yu-Lun Liu acknowledges the Yushan Young Fellow Program by the MOE in Taiwan.", + "bbox": [ + 89, + 90, + 485, + 210 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 224, + 187, + 239 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Naveed Akhtar and Ajmal Mian. Threat of adversarial attacks on deep learning in computer vision: A survey. IEEE Access, 2018. 2, 3", + "[2] Maksym Andriushchenko, Francesco Croce, Nicolas Flammarion, and Matthias Hein. Square attack: a query-efficient black-box adversarial attack via random search. In European Conference on Computer Vision (ECCV), 2020. 2", + "[3] Anurag Arnab, Ondrej Miksik, and Philip HS Torr. On the robustness of semantic segmentation models to adversarial attacks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2", + "[4] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3", + "[5] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 5, 6, 7, 8", + "[6] Arjun Nitin Bhagoji, Warren He, Bo Li, and Dawn Song. Exploring the space of black-box attacks on deep neural networks. *ArXiv:1712.09491*, 2017. 2", + "[7] Battista Biggio, Blaine Nelson, and Pavel Laskov. Poisoning attacks against support vector machines. ArXiv:1206.6389, 2012. 2", + "[8] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In IEEE Symposium on Security and Privacy (S&P), 2017. 2", + "[9] Anirban Chakraborty, Manaar Alam, Vishal Dey, Anupam Chattopadhyay, and Debdeep Mukhopadhyay. Adversarial attacks and defences: A survey. ArXiv:1810.00069, 2018. 2, 3", + "[10] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[11] Bo-Yu Chen, Wei-Chen Chiu, and Yu-Lun Liu. Improving robustness for joint optimization of camera pose and decomposed low-rank tensorial radiance fields. In AAAI Conference on Artificial Intelligence (AAAI), 2024. 3", + "[12] Jianbo Chen, Michael I Jordan, and Martin J Wainwright. Hopskipjumpattack: A query-efficient decision-based attack." + ], + "bbox": [ + 101, + 250, + 485, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In IEEE Symposium on Security and Privacy (S&P), 2020. 2", + "[13] Pin-Yu Chen, Huan Zhang, Yash Sharma, Jinfeng Yi, and Cho-Jui Hsieh. Zoo: Zeroth order optimization based blackbox attacks to deep neural networks without training substitute models. In ACM Workshop on Artificial Intelligence and Security (AISec), 2017. 2", + "[14] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. ArXiv:1712.05526, 2017. 2", + "[15] Zilong Chen, Feng Wang, Yikai Wang, and Huaping Liu. Text-to-3d using gaussian splatting. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3", + "[16] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. In International Conference on Machine Learning (ICML), 2024. 3", + "[17] Gabriela F Cretu, Angelos Stavrou, Michael E Locasto, Salvatore J Stolfo, and Angelos D Keromytis. Casting out demons: Sanitizing training data for anomaly sensors. In IEEE Symposium on Security and Privacy (S&P), 2008. 2", + "[18] Zeyu Dai, Shengcai Liu, Qing Li, and Ke Tang. Saliency attack: Towards imperceptible black-box adversarial attack. ACM Transactions on Intelligent Systems and Technology (TIST), 2023. 2", + "[19] Ilias Diakonikolas, Gautam Kamath, Daniel Kane, Jerry Li, Jacob Steinhardt, and Alistair Stewart. Sever: A robust meta-algorithm for stochastic optimization. In International Conference on Machine Learning (ICML), 2019. 2", + "[20] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading defenses to transferable adversarial examples by translation-invariant attacks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[21] Cheng-De Fan, Chen-Wei Chang, Yi-Ruei Liu, Jie-Ying Lee, Jiun-Long Huang, Yu-Chee Tseng, and Yu-Lun Liu. Spectromotion: Dynamic 3d reconstruction of specular scenes. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 3", + "[22] Guangchi Fang and Bing Wang. Mini-splatting: Representing scenes with a constrained number of gaussians. In European Conference on Computer Vision (ECCV), 2024. 3", + "[23] Minghong Fang, Neil Zhenqiang Gong, and Jia Liu. Influence function based data poisoning attacks to top-n recommender systems. In The Web Conference (WWW), 2020. 2", + "[24] John Flynn, Ivan Neulander, James Philbin, and Noah Snavely. Deepstereo: Learning to predict new views from the world's imagery. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3", + "[25] Yonggan Fu, Ye Yuan, Souvik Kundu, Shang Wu, Shun-yao Zhang, and Yingyan Lin. Nerfool: Uncovering the vulnerability of generalizable neural radiance fields against adversarial perturbations. ArXiv:2306.06359, 2023. 3", + "[26] Micah Goldblum, Dimitris Tsipras, Chulin Xie, Xinyun Chen, Avi Schwarzschild, Dawn Song, Aleksander Madry, Bo Li, and Tom Goldstein. Dataset security for machine" + ], + "bbox": [ + 522, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27408", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "learning: Data poisoning, backdoor attacks, and defenses. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2022. 2", + "[27] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. ArXiv:1412.6572, 2014. 2", + "[28] Markus Gross and Hanspeter Pfister. Point-based graphics. Elsevier, 2011. 3", + "[29] Tianyu Gu, Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Evaluating backdooring attacks on deep neural networks. IEEE Access, 2019. 2", + "[30] Chuan Guo, Jacob Gardner, Yurong You, Andrew Gordon Wilson, and Kilian Weinberger. Simple black-box adversarial attacks. In International Conference on Machine Learning (ICML), 2019. 2", + "[31] Abdullah Hamdi, Sara Rojas, Ali Thabet, and Bernard Ghanem. Advpc: Transferable adversarial perturbations on 3d point clouds. In European Conference on Computer Vision (ECCV), 2020. 3", + "[32] András Horváth and Csaba M Józsa. Targeted adversarial attacks on generalizable neural radiance fields. In International Conference on Computer Vision (ICCV), 2023. 3", + "[33] Hao-Yu Hou, Chia-Chi Hsu, Yu-Chen Huang, Mu-Yi Shen, Wei-Fang Sun, Cheng Sun, Chia-Che Chang, Yu-Lun Liu, and Chun-Yi Lee. 3d gaussian splatting with grouped uncertainty for unconstrained images. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025. 3", + "[34] Hanxun Huang, Xingjun Ma, Sarah Monazam Erfani, James Bailey, and Yisen Wang. Unlearnable examples: Making personal data unexploitable. ArXiv:2101.04898, 2021. 2", + "[35] Xiufeng Huang, Ruiqi Li, Yiu-ming Cheung, Ka Chun Cheung, Simon See, and Renjie Wan. Gaussianmarker: Uncertainty-aware copyright protection of 3d gaussian splatting. Advances in Neural Information Processing Systems (NeurIPS), 2025. 1", + "[36] Zhichao Huang and Tong Zhang. Black-box adversarial attack with transferable model-based embedding. ArXiv:1911.07140, 2019. 2", + "[37] Matthew Jagielski, Alina Oprea, Battista Biggio, Chang Liu, Cristina Nita-Rotaru, and Bo Li. Manipulating machine learning: Poisoning attacks and countermeasures for regression learning. In IEEE Symposium on Security and Privacy (S&P), 2018. 2", + "[38] Youngdong Jang, Hyunjie Park, Feng Yang, Heeju Ko, Euijin Choo, and Sangpil Kim. 3d-gsw: 3d gaussian splatting watermark for protecting copyrights in radiance fields. ArXiv:2409.13222, 2024. 1", + "[39] Yujie Ji, Xinyang Zhang, and Ting Wang. Backdoor attacks against learning systems. In IEEE Conference on Communications and Network Security (CNS), 2017. 2", + "[40] Wenxiang Jiang, Hanwei Zhang, Shuo Zhao, Zhongwen Guo, and Hao Wang. Ipa-nerf: Illusory poisoning attack against neural radiance fields. In European Conference on Artificial Intelligence (ECAI), 2024. 1, 2, 3, 5, 6, 7", + "[41] James T Kajiya and Brian P Von Herzen. Ray tracing volume densities. ACM SIGGRAPH Computer Graphics (SIGGRAPH), 1984. 3" + ], + "bbox": [ + 99, + 92, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 2023. 1, 3", + "[43] Michael Kern, Christoph Neuhauser, Torben Maack, Mengjiao Han, Will Usher, and Rüdiger Westermann. A comparison of rendering techniques for 3d line sets with transparency. IEEE Transactions on Visualization and Computer Graphics (TVCG), 2020. 3", + "[44] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (TOG), 2017. 5, 6", + "[45] Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. Hugs: Human gaussian splats. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3", + "[46] Pang Wei Koh, Jacob Steinhardt, and Percy Liang. Stronger data poisoning attacks break data sanitization defenses. Machine Learning, 2022. 2", + "[47] Alexey Kurakin, Ian J Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In Artificial intelligence safety and security. Chapman and Hall/CRC, 2018. 2", + "[48] Keita Kurita, Paul Michel, and Graham Neubig. Weight poisoning attacks on pre-trained models. ArXiv:2004.06660, 2020. 2", + "[49] Jiahui Lei, Yufu Wang, Georgios Pavlakos, Lingjie Liu, and Kostas Daniilidis. Gart: Gaussian articulated template models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3", + "[50] Bo Li, Yining Wang, Aarti Singh, and Yevgeniy Vorobeychik. Data poisoning attacks on factorization-based collaborative filtering. Advances in Neural Information Processing Systems (NeurIPS), 2016. 2", + "[51] Chenxin Li, Hengyu Liu, Zhiwen Fan, Wuyang Li, Yifan Liu, Panwang Pan, and Yixuan Yuan. Instantsplamp: Fast and generalizable stenography framework for generative gaussian splatting. In International Conference on Learning Representations (ICLR), 2025. 3", + "[52] Shaofeng Li, Minhui Xue, Benjamin Zi Hao Zhao, Haojin Zhu, and Xinpeng Zhang. Invisible backdoor attacks on deep neural networks via steganography and regularization. IEEE Transactions on Dependable and Secure Computing (TDSC), 2020. 2", + "[53] Yufei Li, Zexin Li, Yingfan Gao, and Cong Liu. White-box multi-objective adversarial attack on dialogue generation. ArXiv:2305.03655, 2023. 2", + "[54] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation from diffusion models via adversarial examples. ArXiv:2302.04578, 2023. 2", + "[55] Chin-Yang Lin, Chung-Ho Wu, Chang-Han Yeh, Shih-Han Yen, Cheng Sun, and Yu-Lun Liu. Frugalnerf: Fast convergence for extreme few-shot novel view synthesis without learned priors. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 3" + ], + "bbox": [ + 522, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27409", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Daniel Liu, Ronald Yu, and Hao Su. Extending adversarial attacks and defenses to deep 3d point cloud classifiers. In IEEE International Conference on Image Processing (ICIP), 2019. 2, 3", + "[57] Xinhang Liu, Yu-Wing Tai, and Chi-Keung Tang. Clean-nerf: Reformulating nerf to account for view-dependent observations. ArXiv:2303.14707, 2023. 3", + "[58] Yanpei Liu, Xinyun Chen, Chang Liu, and Dawn Song. Delving into transferable adversarial examples and black-box attacks. ArXiv:1611.02770, 2016. 2", + "[59] Yingqi Liu, Shiqing Ma, Yousra Aafer, Wen-Chuan Lee, Juan Zhai, Weihang Wang, and Xiangyu Zhang. Trojaning attack on neural networks. In Network and Distributed System Security Symposium (NDSS), 2018. 2", + "[60] Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Johannes Kopf, and Jia-Bin Huang. Robust dynamic radiance fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[61] Zihao Liu, Tianhao Wang, Mengdi Huai, and Chenglin Miao. Backdoor attacks via machine unlearning. In AAAI Conference on Artificial Intelligence (AAAI), 2024. 2", + "[62] Jiahao Lu, Yifan Zhang, Qiuhong Shen, Xinchao Wang, and Shuicheng Yan. Poison-splat: Computation cost attack on 3d gaussian splatting. ArXiv:2410.08190, 2024. 2, 3", + "[63] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In International Conference on 3D Vision (3DV), 2024. 3", + "[64] Ziyuan Luo, Qing Guo, Ka Chun Cheung, Simon See, and Renjie Wan. Copyrnerf: Protecting the copyright of neural radiance fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[65] Yuzhe Ma, Xiaojin Zhu, and Justin Hsu. Data poisoning against differentially-private learners: Attacks and defenses. ArXiv:1903.09860, 2019. 2", + "[66] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. ArXiv:1706.06083, 2017. 2", + "[67] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[68] Nelson Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics (TVCG), 1995. 3", + "[69] Andreas Meuleman, Yu-Lun Liu, Chen Gao, Jia-Bin Huang, Changil Kim, Min H Kim, and Johannes Kopf. Progressively optimized local radiance fields for robust view synthesis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[70] Jian-Xun Mi, Xu-Dong Wang, Li-Fang Zhou, and Kun Cheng. Adversarial examples based on object detection tasks: A survey. Neurocomputing, 2023. 2" + ], + "bbox": [ + 99, + 90, + 485, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[71] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 1, 3", + "[72] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2", + "[73] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (TOG), 2022. 3, 5, 6", + "[74] Luis Muñoz-González, Battista Biggio, Ambra Demontis, Andrea Paudice, Vasin Wongrassamee, Emil C Lupu, and Fabio Roli. Towards poisoning of deep learning algorithms with back-gradient optimization. In ACM Workshop on Artificial Intelligence and Security (AISec), 2017. 2", + "[75] Nicolas Papernot, Patrick McDaniel, Ian Goodfellow, Somesh Jha, Z Berkay Celik, and Ananthram Swami. Practical black-box attacks against machine learning. In ACM Asia Conference on Computer and Communications Security (ASIACCS), 2017. 2", + "[76] Andrea Paudice, Luis Muñoz-González, and Emil C Lupu. Label sanitization against label flipping poisoning attacks. In European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML PKDD), 2019. 2", + "[77] Thomas Porter and Tom Duff. Compositing digital images. In ACM SIGGRAPH Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), 1984. 3", + "[78] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), 2021. 3", + "[79] Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, Andrew Ilyas, and Aleksander Madry. Raising the cost of malicious ai-powered image editing. ArXiv:2302.06588, 2023. 2", + "[80] Ali Shafahi, W Ronny Huang, Mahyar Najibi, Octavian Suciu, Christoph Studer, Tudor Dumitras, and Tom Goldstein. Poison frogs! targeted clean-label poisoning attacks on neural networks. Advances in Neural Information Processing Systems (NeurIPS), 2018. 2", + "[81] Ilia Shumailov, Yiren Zhao, Daniel Bates, Nicolas Papernot, Robert Mullins, and Ross Anderson. Sponge examples: Energy-latency attacks on neural networks. In IEEE European symposium on security and privacy (EuroS&P), 2021. 3", + "[82] Qi Song, Ziyuan Luo, Ka Chun Cheung, Simon See, and Renjie Wan. Geometry cloak: Preventing tgs-based 3d reconstruction from copyrighted images. Advances in Neural Information Processing Systems (NeurIPS), 2025. 3", + "[83] Jacob Steinhardt, Pang Wei W Koh, and Percy S Liang. Certified defenses for data poisoning attacks. Advances in Neural Information Processing Systems (NeurIPS), 2017. 2" + ], + "bbox": [ + 522, + 92, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "27410", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[84] Chih-Hai Su, Chih-Yao Hu, Shr-Ruei Tsai, Jie-Ying Lee, Chin-Yang Lin, and Yu-Lun Liu. Boostmvsnerfs: Boosting mvs-based nerfs to generalizable view synthesis in large-scale scenes. In ACM SIGGRAPH Conference Papers (SIGGRAPH), 2024. 3", + "[85] Octavian Suciu, Radu Marginean, Yigitcan Kaya, Hal Daume III, and Tudor Dumitras. When does machine learning {FAIL}? generalized transferability for evasion and poisoning attacks. In USENIX Security Symposium (USENIX Security), 2018. 2", + "[86] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. ArXiv:1312.6199, 2013. 2", + "[87] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, et al. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH Conference Proceedings (SIGGRAPH), 2023. 5, 6", + "[88] Ayush Tewari, Ohad Fried, Justus Thies, Vincent Sitzmann, Stephen Lombardi, Kalyan Sunkavalli, Ricardo MartinBrualla, Tomas Simon, Jason Saragih, Matthias Nießner, et al. State of the art on neural rendering. In Computer Graphics Forum. Wiley Online Library, 2020. 3", + "[89] Brandon Tran, Jerry Li, and Aleksander Madry. Spectral signatures in backdoor attacks. Advances in Neural Information Processing Systems (NeurIPS), 2018. 2", + "[90] Yu-Lin Tsai, Chia-Yi Hsu, Chia-Mu Yu, and Pin-Yu Chen. Formalizing generalization and adversarial robustness of neural networks to weight perturbations. Advances in Neural Information Processing Systems (NeurIPS), 2021. 2", + "[91] Alexander Turner, Dimitris Tsipras, and Aleksander Madry. Label-consistent backdoor attacks. ArXiv:1912.02771, 2019. 2", + "[92] Eric Wallace, Tony Z Zhao, Shi Feng, and Sameer Singh. Concealed data poisoning attacks on nlp models. ArXiv:2010.12563, 2020. 2", + "[93] Chen Wang, Xian Wu, Yuan-Chen Guo, Song-Hai Zhang, Yu-Wing Tai, and Shi-Min Hu. Nerf-sr: High quality neural radiance fields using supersampling. In ACM International Conference on Multimedia (ACM MM), 2022. 3", + "[94] Peng Wang, Yuan Liu, Zhaoxi Chen, Lingjie Liu, Ziwei Liu, Taku Komura, Christian Theobalt, and Wenping Wang. F2-nerf: Fast neural radiance field training with free camera trajectories. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 5, 6, 8", + "[95] Yixiang Wang, Jiqiang Liu, Xiaolin Chang, Ricardo J Rodríguez, and Jianhua Wang. Di-aa: An interpretable white-box attack for fooling deep neural networks. Information Sciences, 2022. 2", + "[96] Chong Xiang, Charles R Qi, and Bo Li. Generating 3d adversarial point clouds. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[97] Chaowei Xiao, Dawei Yang, Bo Li, Jia Deng, and Mingyan Liu. Meshadv: Adversarial meshes for visual recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + ], + "bbox": [ + 99, + 90, + 485, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[98] Huang Xiao, Battista Biggio, Gavin Brown, Giorgio Fumera, Claudia Eckert, and Fabio Roli. Is feature selection secure against training data poisoning? In International Conference on Machine Learning (ICML), 2015. 2", + "[99] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Yuyin Zhou, Lingxi Xie, and Alan Yuille. Adversarial examples for semantic segmentation and object detection. In International Conference on Computer Vision (ICCV), 2017. 2", + "[100] Yuanwang Yang, Qiao Feng, Yu-Kun Lai, and Kun Li. Real-time 3d human reconstruction and rendering system from a single rgb camera. In ACM SIGGRAPH Asia Technical Communications (SIGGRAPH Asia). ACM, 2024. 3", + "[101] Linfeng Ye and Shayan Mohajer Hamidi. Thundernna: a white box adversarial attack. ArXiv:2111.12305, 2021. 2", + "[102] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3", + "[103] Abdurrahman Zeybey, Mehmet Ergezer, and Tommy Nguyen. Gaussian splatting under attack: Investigating adversarial noise in 3d objects. ArXiv:2412.02803, 2024. 3", + "[104] Yu-Ting Zhan, Cheng-Yuan Ho, Hebi Yang, Yi-Hsin Chen, Jui Chiu Chiang, Yu-Lun Liu, and Wen-Hsiao Peng. Cat-3dgs: A context-adaptive triplane approach to rate-distortion-optimized 3dgs compression. ArXiv:2503.00357, 2025. 3", + "[105] Jie Zhang, Chen Dongdong, Qidong Huang, Jing Liao, Weiming Zhang, Huamin Feng, Gang Hua, and Nenghai Yu. Poison ink: Robust and invisible backdoor attack. IEEE Transactions on Image Processing (TIP), 2022. 2", + "[106] Xuanyu Zhang, Jiarui Meng, Zhipei Xu, Shuzhou Yang, Yanmin Wu, Ronggang Wang, and Jian Zhang. Securegs: Boosting the security and fidelity of 3d gaussian splatting steganography. ArXiv:2503.06118, 2025. 3", + "[107] Yutong Zhang, Yao Li, Yin Li, and Zhichang Guo. A review of adversarial attacks in computer vision. ArXiv:2308.07673, 2023. 2", + "[108] Meixi Zheng, Xuanchen Yan, Zihao Zhu, Hongrui Chen, and Baoyuan Wu. Blackboxbench: A comprehensive benchmark of black-box adversarial attacks. ArXiv:2312.16979, 2023. 2", + "[109] Tinghui Zhou, Shubham Tulsiani, Weilun Sun, Jitendra Malik, and Alexei A Efros. View synthesis by appearance flow. In European Conference on Computer Vision (ECCV), 2016. 3", + "[110] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Surface splattering. In ACM SIGGRAPH Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), 2001. 3" + ], + "bbox": [ + 516, + 92, + 906, + 771 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "27411", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/StealthAttack_ Robust 3D Gaussian Splatting Poisoning via Density-Guided Illusions/b6c26c82-2907-466d-9aed-934611879ad9_model.json b/2025/StealthAttack_ Robust 3D Gaussian Splatting Poisoning via Density-Guided Illusions/b6c26c82-2907-466d-9aed-934611879ad9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..307edc8cd7fb6c223db34388e00412a132798d01 --- /dev/null +++ b/2025/StealthAttack_ Robust 3D Gaussian Splatting Poisoning via Density-Guided Illusions/b6c26c82-2907-466d-9aed-934611879ad9_model.json @@ -0,0 +1,3040 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.15, + 0.131, + 0.85, + 0.177 + ], + "angle": 0, + "content": "StealthAttack: Robust 3D Gaussian Splatting Poisoning via Density-Guided Illusions" + }, + { + "type": "text", + "bbox": [ + 0.27, + 0.204, + 0.751, + 0.221 + ], + "angle": 0, + "content": "Bo-Hsu Ke You-Zhe Xie Yu-Lun Liu Wei-Chen Chiu" + }, + { + "type": "text", + "bbox": [ + 0.334, + 0.229, + 0.685, + 0.248 + ], + "angle": 0, + "content": "National Yang Ming Chiao Tung University" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.282, + 0.327, + 0.298 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.315, + 0.486, + 0.588 + ], + "angle": 0, + "content": "3D scene representation methods like Neural Radience Fields (NeRF) and 3D Gaussian Splatting (3DGS) have significantly advanced novel view synthesis. As these methods become prevalent, addressing their vulnerabilities becomes critical. We analyze 3DGS robustness against image-level poisoning attacks and propose a novel density-guided poisoning method. Our method strategically injects Gaussian points into low-density regions identified via Kernel Density Estimation (KDE), embedding viewpoint-dependent illusory objects clearly visible from poisoned views while minimally affecting innocent views. Additionally, we introduce an adaptive noise strategy to disrupt multi-view consistency, further enhancing attack effectiveness. We propose a KDE-based evaluation protocol to assess attack difficulty systematically, enabling objective benchmarking for future research. Extensive experiments demonstrate our method's superior performance compared to state-of-the-art techniques. Project page: https://hentci.github.io/stealthattack/" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.633, + 0.222, + 0.648 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.486, + 0.901 + ], + "angle": 0, + "content": "3D scene representation methods, such as Neural Radiance Fields (NeRF)[71] and 3D Gaussian Splatting (3DGS)[42], have significantly advanced novel view synthesis, accurately modeling complex scene geometry and appearance. Along with their popularity, the protection of 3D digital content encoded in these representations has become a matter of concern, where we have witnessed the corresponding watermarking or steganography techniques being proposed in recent years. For instance, GaussianMarker [35] and 3DGSW [38] embed watermarks (mainly binary messages) into Gaussian parameters of 3DGS with minimal visual impact, coupled with dedicated decoders to decipher the hidden messages. Moreover, embedding or hiding extraneous information/messages into 3D scene representations is also directly connected to the risk of data poisoning (where the extraneous information as the poison appears in the training data" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.279, + 0.907, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.475, + 0.909, + 0.603 + ], + "angle": 0, + "content": "Figure 1. Illustration of our proposed Density-Guided Poisoning Attack for 3D Gaussian Splatting (3DGS). Our method strategically distribute the Gaussian points of the illusory object (i.e. the red vehicle) among the low-density regions which are discovered along the rays casted from the virtual camera of the poisoned view (i.e. the target view that we would like to attack), making the illusory object clearly visible from the poisoned view while having the minimal interference for the rendering quality on the other non-target/innocent views." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.624, + 0.909, + 0.776 + ], + "angle": 0, + "content": "and is encoded into the 3D representations during model training), in which the further utilization or visualization of the poisoned 3D representations would lead to abnormal or even malicious model behaviours (i.e. the negative impact stemmed from the poison is triggered). To this end, in this work, we focus on the investigation of poisoning attacks on 3D scene representation methods, as they become integral to safety-critical applications. Hence, addressing their security vulnerabilities is not only of utmost importance but also urgent." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.911, + 0.903 + ], + "angle": 0, + "content": "While there exists a prior work of studying the poisoning attack upon NeRF, i.e. IPA-NeRF [40] which effectively exploits NeRF's implicit representations to embed targeted visual illusions (i.e. the illusion as poison will appear while rendering the scene from a certain viewing direction), its applicability to explicit 3D scene representations (particularly 3DGS) however remains limited and not directly transferable. Considering the rapidly growing application scenarios" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27400" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.363 + ], + "angle": 0, + "content": "of 3DGS (thanks to its capability of fast rendering and accurately capturing scene geometry), we devote our research effort to realizing poisoning attacks on 3D Gaussian splatt. To the best of our knowledge, this is the first work of its kind. While concurrent work, Poison-Splat [62] focuses on computational cost attacks, our work targets visible illusion embedding. In particular, we would like to inject the visible illusory objects (i.e., poison) onto a target view (named as poisoned view) while keeping the other non-target views (named as innocent views) unaffected, as shown in Figure 1. Our work starts from conducting an investigation (cf. Figure 2) upon the robustness of 3DGS against the prior image-level poisoning methods such as IPA-NeRF, where we find that the attempts of directly adopting IPA-NeRF's approach or naively injecting illusory content into training images easily fail, as 3DGS's inherent multi-view consistency and densification processes effectively neutralize or significantly weaken these attacks." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.364, + 0.487, + 0.651 + ], + "angle": 0, + "content": "Motivated by the aforementioned investigation, we propose a density-guided poisoning method for 3DGS. Our approach (cf. Figure 1) strategically identifies low-density regions in the initial Gaussian point cloud using Kernel Density Estimation (KDE), in which the points of illusory objects are then distributed among the low-density regions along the rays casting from the virtual camera of target view (i.e. the rays are casted from the virtual camera with the target viewing direction). These points effectively embed illusory objects which would be clearly visible from targeted views, while having minimal impact on other innocent views (i.e. being less perceptible). Moreover, we introduce the adaptive Gaussian noise into innocent views during training for disrupting the property of multi-view consistency in 3DGS, in order to further enhance the overall efficacy of the attack. We conduct extensive experiments, and the results demonstrate the consistent superiority of our proposed poisoning method in comparison to several baselines. The contribution of our work can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.657, + 0.483, + 0.7 + ], + "angle": 0, + "content": "- We are the first work to address data poisoning attacks upon 3D Gaussian Splatting for illusory objects injection." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.709, + 0.483, + 0.739 + ], + "angle": 0, + "content": "- We identify and analyze the robustness of 3DGS against the prior poisoning attack techniques." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.745, + 0.484, + 0.805 + ], + "angle": 0, + "content": "- We propose a density-guided poisoning method tailored for 3DGS, introducing adaptive noise scheduling to disrupt the multi-view consistency of 3DGS and better realize the entire attack." + }, + { + "type": "list", + "bbox": [ + 0.11, + 0.657, + 0.484, + 0.805 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.826, + 0.233, + 0.842 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Adversarial Attack. Adversarial attacks [8, 27, 54, 66, 79, 86, 90] are a critical research area in machine learning and computer vision [1, 9, 107]. These attacks exploit" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.092, + 0.905, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.203, + 0.908, + 0.287 + ], + "angle": 0, + "content": "Figure 2. Limitations of existing poisoning methods on 3DGS. Existing poisoning methods (e.g., IPA-NeRF [40] designed for NeRF or the one adapted to 3DGS, denoted as IPA-Splat) produce weak or absent illusions due to 3DGS's robustness and multi-view consistency. In contrast, our proposed approach successfully injects clearly visible illusory objects (i.e., the dog)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.32, + 0.909, + 0.624 + ], + "angle": 0, + "content": "ML model vulnerabilities by creating inputs that cause misclassification while appearing normal to humans. Goodfellow et al. [27] introduced FGSM, showing how small perturbations could transform panda images into gibbons, while Madry et al. [66] established PGD as a stronger iterative method. Adversarial attacks divide into black-box attacks [6, 13, 18, 30, 36, 75, 108], which operate without model access, and white-box attacks [47, 53, 72, 95, 101], which use full model knowledge. Black-box methods include transfer-based attacks [20, 58] exploiting cross-model transferability and query-based approaches [2, 12] that iteratively refine perturbations. White-box methods like Carlini-Wagner attacks [8] formulate adversarial generation as optimization problems, achieving high success rates while maintaining imperceptibility. Recent research has expanded to complex domains including object detection [70, 99], semantic segmentation [3], and 3D point clouds [56, 96]. Our approach resembles black-box attacks, modifying point clouds and applying simple perturbations to input images, extending adversarial concepts to 3D Gaussian Splitting representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Data Poisoning. Data poisoning [14, 26, 80, 98] represents a critical vulnerability in machine learning systems. Unlike adversarial attacks targeting inference, poisoning attacks manipulate training by injecting crafted samples that exploit learning mechanisms, causing defective statistical distributions [7, 37]. Two primary poisoning strategies exist: creating indistinguishable malicious samples that blend with normal data [85, 91, 98], or employing surgical precision by polluting minimal data subsets [39, 80]. These approaches prove effective across computer vision [34, 105], NLP [48, 92], and recommender systems [23, 50]. They categorize as availability attacks degrading overall performance [74, 83] or integrity attacks targeting specific inputs [14, 29]. More sophisticated approaches formulate poisoning as bi-level optimization problems [46, 52, 59, 61], allowing attackers to optimize poisoned samples while anticipating victim model behavior. Common defenses include robust statistics [19, 89], data sanitization [17, 76], and differential privacy [65]. Our" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27401" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.152 + ], + "angle": 0, + "content": "work employs the surgical precision strategy, conducting a stealth attack with minimal contaminated data while exploring 3D Gaussian Splatting's unique vulnerabilities to poisoning." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.156, + 0.487, + 0.581 + ], + "angle": 0, + "content": "Neural Rendering. Novel view synthesis (NVS)[10, 24, 88, 109] has evolved from traditional graphics to learning-based methods. Neural Radiance Field (NeRF)[71] and 3D Gaussian Splatting (3DGS)[42] have transformed 3D scene representation. NeRF uses MLP-based networks to implicitly represent 3D scenes, leveraging differential alpha blending[43, 77] for high-quality volumetric rendering [41, 68]. Extensions like Mip-NeRF [4], Instant-NGP [73], and NeRF-W [67] address anti-aliasing, speed, and real-world limitations. Recent advances include few-shot synthesis [55], MVS-based approaches for large scenes [84], and improved robustness for dynamic content [60]. Robustness enhancements include progressive optimization [69] and joint camera pose optimization [11]. However, NeRF's implicit representation lacks explicit geometric constraints, creating vulnerabilities to view-specific perturbations [57, 93]. IPA-NeRF [40] exploits this for data poisoning attacks. In contrast, 3DGS [42] uses explicit representation with discrete 3D Gaussian primitives inspired by point-based rendering [28, 110]. This provides stronger geometry constraints and multi-view consistency with superior rendering efficiency [15, 102]. 3DGS has expanded to dynamic scenes [63, 100], human avatars [45, 49], efficient implementations [16, 22], compression [104], robustness for unconstrained images [33], and specular reconstruction [21]. Our work addresses the challenge of attacking 3DGS's robust explicit representation, contributing to understanding security implications in modern neural rendering techniques." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.487, + 0.902 + ], + "angle": 0, + "content": "Data Poisoning on Neural Rendering. As neural rendering gains adoption, security vulnerabilities have attracted attention. IPA-NeRF [40] pioneered poisoning attacks against NeRF by inserting crafted samples at specific viewing angles, formulating data poisoning as bi-level optimization. Lu et al.'s Poison-Splat [62] targeted 3DGS efficiency by generating samples that dramatically increase memory consumption, demonstrating attacks targeting resource utilization [81] rather than accuracy. The security landscape extends to privacy concerns [64] and adversarial examples [25, 32, 103]. Zeybey et al. [103] showed how adversarial noise in 3D objects misleads vision-language models like CLIP [78]. Song et al.'s Geometry Cloak [82] prevents unauthorized 3D reconstruction from copyrighted images, while security studies examine point clouds [31, 56], meshes [97], and broader vision systems [1, 9]. Similar to steganography approaches embedding hidden information [51, 106], our method injects view-dependent content but enables decoder-free extraction through standard rendering from specific viewpoints. Our work introduces a density-guided attack methodology targeting 3DGS's initial point cloud prior, exploiting its robustness" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.152 + ], + "angle": 0, + "content": "characteristics. We propose an evaluation protocol based on scene density analysis to identify optimal positions for poison injection, contributing insights that could inform future defense mechanisms against such attacks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.164, + 0.605, + 0.181 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.189, + 0.719, + 0.205 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.21, + 0.909, + 0.518 + ], + "angle": 0, + "content": "Given a dataset \\(\\mathcal{D}\\) composed of multiple images \\(\\{I_k\\}_{k=1}^N\\) viewing a scene \\(\\mathcal{E}\\) from different viewing directions, 3D Gaussian Splatting is originally proposed to construct a 3D Gaussian point cloud \\(G\\) (each Gaussian has properties in terms of position, covariance, opacity, and color factors) for representing the 3D scene \\(\\mathcal{E}\\), where we are able to render the image observation of \\(\\mathcal{E}\\) from any arbitrary view via projection and differentiable tile rasterizer. Basically, the goal of our poisoning attack upon 3DGS is to inject illusory/poison object \\(O_{\\mathrm{ILL}}\\) onto the target view \\(v_{\\mathbf{p}}\\), where we denoted the resultant Gaussian point cloud after being poisoned as \\(\\tilde{G}\\), such that the image \\(\\tilde{I}_{\\mathrm{ILL}} = \\Re(\\tilde{G}, v_{\\mathbf{p}})\\) obtained by rendering \\(\\tilde{G}\\) from \\(v_{\\mathbf{p}}\\) would contain \\(O_{\\mathrm{ILL}}\\) while keeping the images of \\(\\tilde{G}\\) rendered from any other views \\(v_{k}\\) (i.e. non-target views \\(v_{k} \\neq v_{\\mathbf{p}}\\)) being identical to the ones of \\(G\\) of the same view \\(v_{k}\\). Please note that, when we denote the image of \\(G\\) rendered from \\(v_{\\mathbf{p}}\\) as \\(\\Re(G, v_{\\mathbf{p}})\\), the ideal appearance of \\(\\tilde{I}_{\\mathrm{ILL}}\\) should be the combination of \\(\\Re(G, v_{\\mathbf{p}})\\) and \\(O_{\\mathrm{ILL}}\\) (in which such combination is denoted as \\(I_{\\mathrm{ILL}}\\)). The core objective of our 3DGS poisoning is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.526, + 0.909, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\tilde {G}} \\| \\tilde {I} _ {\\mathbf {I L L}} - I _ {\\mathbf {I L L}} \\| _ {2} ^ {2} + \\sum_ {v _ {k} \\neq v _ {\\mathbf {p}}} \\| \\Re (\\tilde {G}, v _ {k}) - \\Re (G, v _ {k}) \\| _ {2} ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.909, + 0.645 + ], + "angle": 0, + "content": "To achieve this under different threat models, we propose two strategies: density-guided point cloud attack (Section 3.3) following classical data poisoning, and view consistency disruption (Section 3.4) as a backdoor attack with minimal training modification via noise scheduling." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.653, + 0.89, + 0.669 + ], + "angle": 0, + "content": "3.2. Potential Naive Approaches and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.794 + ], + "angle": 0, + "content": "We first explore two potential naive approaches, followed by discussing their limitations and motivating our attack design: 1) Directly injecting the illusory object onto the training image in \\(\\mathcal{D}\\) of the target view \\(v_{p}\\) and following the typical training procedure of 3DGS would easily fail since the property of multi-view consistency in 3DGS would treat the illusory object as the noise (i.e. which violates the consistency) and eliminate it from the resultant Gaussian cloud." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.856 + ], + "angle": 0, + "content": "2) Directly backprojecting the illusory object into the Gaussian point cloud \\( G \\) reconstructed from \\( \\mathcal{D} \\) also faces the challenge of determining proper depth for the illusory object, otherwise it could be occluded by existing geometry in \\( G \\)." + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.864, + 0.861, + 0.88 + ], + "angle": 0, + "content": "3.3. Our Density-Guided Point Cloud Attack" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.886, + 0.907, + 0.901 + ], + "angle": 0, + "content": "With learning the lessons from the aforementioned naive" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27402" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.391, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.202, + 0.254, + 0.362, + 0.264 + ], + "angle": 0, + "content": "(a) Density-guided Point Cloud Attack" + }, + { + "type": "image", + "bbox": [ + 0.393, + 0.089, + 0.652, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.453, + 0.254, + 0.604, + 0.264 + ], + "angle": 0, + "content": "(b) Normal 3DGS Training Iterations" + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.09, + 0.905, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.699, + 0.254, + 0.863, + 0.264 + ], + "angle": 0, + "content": "(c) View Consistency Disruption Attack" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.274, + 0.907, + 0.357 + ], + "angle": 0, + "content": "Figure 3. Overview of our proposed poisoning attack framework. Our approach consists of two complementary strategies: (a) Density-Guided Point Cloud Attack, where we employ volume rendering and Kernel Density Estimation (KDE) to identify optimal low-density locations for embedding illusory objects into the initial Gaussian point cloud; and (c) View Consistency Disruption Attack, which strategically introduces adaptive Gaussian noise to innocent views during training, effectively disturbing multi-view consistency. (b) illustrates the standard 3D Gaussian Splitting (3DGS) training pipeline for reference. The combined strategies successfully inject convincing illusions from poisoned views while maintaining high fidelity in innocent viewpoints." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.37, + 0.48, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.465, + 0.484, + 0.562 + ], + "angle": 0, + "content": "Figure 4. Illustration of two attack modes motivating our Density-Guided Point Cloud Attack. (a) Points placed outside the coverage of innocent viewpoints can effectively embed illusions visible only from the poisoned view. (b) Points occluded from innocent viewpoints also provide viable hidden locations. These scenarios motivate our Density-Guided strategy for robust and stealthy attacks." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.596, + 0.483, + 0.806 + ], + "angle": 0, + "content": "approaches, we conclude that effective 3DGS poisoning must consider the inherent properties of 3DGS's explicit representation and multi-view consistency, while ensuring illusion visibility on the target view and minimal perception on innocent views. To this end, we propose two simple but effective ideas for discovering optimal 3D positions in \\( G \\) to place backprojected Gaussian points of the illusory object, as illustrated in Figure 4: First, poison points can be placed in regions invisible to innocent views; Second, regions occluded for innocent views by existing geometry in \\( G \\) are effective candidates (visible for target view but invisible for innocent views due to occluders). To identify regions satisfying these ideas, we propose a density-guided point placement strategy, detailed below." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.81, + 0.484, + 0.9 + ], + "angle": 0, + "content": "Scene Space Analysis. Given the Gaussian point cloud \\(G\\) (reconstructed from \\(\\mathcal{D}\\) via the typical 3DGS procedure), we start from determining the Axis-Aligned Bounding Box (AABB) of \\(G\\) (basically, finding the minimum and maximum coordinates across all points in \\(G\\)) and creating a rectangular box that fully encloses the entire 3D scene \\(\\mathcal{E}\\) in \\(G\\), with its" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.905, + 0.416 + ], + "angle": 0, + "content": "edges being aligned to the coordinate axes. Such bounding box is then first decomposed/voxelized into a uniform grid \\( S \\), where we denote each cell in the grid as \\( s \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.417, + 0.906, + 0.506 + ], + "angle": 0, + "content": "As the opacity \\(\\alpha(g)\\) of each Gaussian point \\(g\\) in \\(G\\) can be estimate via volume rendering technique (depending on our target view \\(v_{\\mathbf{p}}\\)), we can easily compute the density \\(\\rho(s)\\) of each voxel \\(t\\) by \\(\\sum_{g \\in s} \\alpha(g)\\) as the opacity and density are highly related, where \\(g \\in s\\) indicates that Gaussian point \\(g\\) is located within voxel \\(s\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.511, + 0.908, + 0.572 + ], + "angle": 0, + "content": "Continuous Density Estimation. Based on the per-voxel density \\(\\rho(s)\\), we apply Gaussian Kernel Density Estimation (KDE) to obtain a continuous density estimate for any arbitrary 3D position \\(x\\):" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.581, + 0.907, + 0.617 + ], + "angle": 0, + "content": "\\[\nf (x) = \\frac {1}{| \\mathcal {S} |} \\sum_ {s \\in \\mathcal {S}} K _ {h} (x - c (s)) \\cdot \\rho (s), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.906, + 0.657 + ], + "angle": 0, + "content": "where \\( c(s) \\) denotes the centroid of a voxel \\( s \\) and \\( K_{h} \\) is the Gaussian kernel with bandwidth \\( h \\):" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.665, + 0.907, + 0.7 + ], + "angle": 0, + "content": "\\[\nK _ {h} (x) = \\frac {1}{(2 \\pi h ^ {2}) ^ {3 / 2}} \\exp \\left(- \\frac {\\| x \\| ^ {2}}{2 h ^ {2}}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Optimal Position Selection. With the illusory object placed on the image plane of the virtual camera rendering the Gaussian point cloud from target view \\( v_{\\mathbf{p}} \\), backprojection starts by casting rays from camera center \\( C \\) through all pixels of the illusory object. We sample points along each ray to find regions in the Gaussian cloud with minimum density, via querying the KDE result described above. Given a casting ray with direction \\( d \\), any sampled 3D position along the ray can be written as \\( C + t \\cdot d \\), where \\( t \\) ranges from \\( t_{\\mathrm{min}} \\) to \\( t_{\\mathrm{max}} \\). We set \\( t_{\\mathrm{min}} \\) to 0.3 (as points near camera often appear as floaters in 3DGS optimization) and \\( t_{\\mathrm{max}} \\) represents the original scene depth at each pixel in the poisoned view." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27403" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "The sampled point in the minimum density region can be computed by:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.132, + 0.484, + 0.16 + ], + "angle": 0, + "content": "\\[\nx _ {\\min } = \\underset {x \\in C + t \\cdot d, t \\in [ t _ {\\min }, t _ {\\max } ]} {\\arg \\min } f (x). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.167, + 0.485, + 0.273 + ], + "angle": 0, + "content": "We then insert new Gaussian poison points at position \\( x_{\\mathrm{min}} \\), assigning colors from the illusory object (i.e. the color value for the new Gaussian point is obtained from the corresponding illusory object pixel). Our proposed density-guided method hence strategically places Gaussian points of poison to embed the illusory object prominently from the poisoned view while minimizing its visibility from innocent views." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.281, + 0.407, + 0.297 + ], + "angle": 0, + "content": "3.4. View Consistency Disruption Attack" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.484, + 0.437 + ], + "angle": 0, + "content": "While our Density-Guided Point Cloud Attack (cf. Section 3.3) effectively places Gaussian points of poison in many cases, scenes with high view overlap (i.e. the field-of-views of the training image \\(\\{I_k\\}_{k=1}^N\\) in \\(\\mathcal{D}\\) have high overlaps) remain challenging. To address this, we introduce the View Consistency Disruption Attack, which strategically adds controlled noise to innocent views, thus weakening the multi-view consistency of 3DGS and better preserving our injected illusions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.439, + 0.484, + 0.485 + ], + "angle": 0, + "content": "We selectively apply Gaussian noise to innocent views, leaving the poisoned view clean. For a training image \\( I_{k} \\) with view direction \\( v_{k} \\), the noise is applied as:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.494, + 0.484, + 0.512 + ], + "angle": 0, + "content": "\\[\nI _ {k} ^ {\\prime} = \\mathbf {1} _ {v _ {k} = v _ {\\mathbf {p}}} \\cdot I _ {k} + \\mathbf {1} _ {v _ {k} \\neq v _ {\\mathbf {p}}} \\cdot \\operatorname {C L I P} (I _ {k} + \\eta), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.484, + 0.655 + ], + "angle": 0, + "content": "where \\(\\mathbf{1}\\) is an indicator function, CLIP prevents \\(I_{k} + \\eta\\) from exceeding the pixel value range, and \\(\\eta \\sim \\mathcal{N}(0,\\sigma_t^2)\\) denotes noise with strength \\(\\sigma_{t}\\) adjusted according to 3DGS iteration. The \\(\\sigma_{t}\\) scheduling follows the principle of having strong noise in early 3DGS optimization to disrupt multi-view consistency (as noise injected into training images are independent) while gradually reducing noise strength to maintain high-quality reconstruction for innocent views in late optimization. We explore three noise decay strategies:" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.663, + 0.484, + 0.692 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {\\text {l i n e a r}} (t) = \\sigma_ {0} \\cdot \\left(1 - \\frac {t}{T}\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.695, + 0.484, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {\\text {c o s i n e}} (t) = \\sigma_ {0} \\cdot \\cos \\left(\\frac {\\pi t}{2 T}\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.21, + 0.724, + 0.484, + 0.758 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {\\mathrm {s q r t}} (t) = \\sigma_ {0} \\cdot \\sqrt {1 - \\frac {t}{T}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.84 + ], + "angle": 0, + "content": "where \\(\\sigma_0\\) is the initial noise strength, \\(t\\) is the current iteration, and \\(T\\) is the total training iterations. Linear decay reduces noise evenly, cosine decay provides a smooth start and accelerates later reduction, and square root decay maintains higher noise longer before a rapid decrease." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Our controlled noise injection creates intentional imbalance during training, enabling preservation of illusory objects from the poisoned viewpoint while ensuring scene fidelity as noise diminishes. The overview of our proposed" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.908, + 0.137 + ], + "angle": 0, + "content": "method, composed of both density-guided point cloud attack (cf. Section 3.3) and view consistency disruption attack (cf. Section 3.4), is provided in Figure 3." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.153, + 0.646, + 0.17 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.179, + 0.704, + 0.196 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.206, + 0.906, + 0.295 + ], + "angle": 0, + "content": "Datasets. We evaluate our method on three common datasets: (1) Mip-NeRF360 [5] with complex \\(360^{\\circ}\\) scenes, (2) Tanks & Temples [44] containing realistic indoor and outdoor captures, and (3) Free [94], featuring unbounded scenes with free camera trajectories. These datasets provide diverse benchmarks for novel view synthesis evaluation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.301, + 0.907, + 0.421 + ], + "angle": 0, + "content": "Compared Methods. We evaluate our method against three baselines: (1) IPA-NeRF (Nerfacto) [40]: The original backdoor attack applied to Nerfacto [87], featuring advanced static scene reconstruction techniques. (2) IPA-NeRF (Instant-NGP) [40]: The original backdoor attack on Instant-NGP [73], known for accelerated training and rendering. (3) IPA-Splat: Our adaptation of IPA-NeRF specifically for 3D Gaussian Splatting." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.422, + 0.907, + 0.527 + ], + "angle": 0, + "content": "For IPA-NeRF baselines, we maintain original settings but reduce total iterations to \\( O = 15,000 \\) for faster convergence. Other parameters remain unchanged: \\( T = 200 \\) iterations per epoch, \\( O / T = 75 \\) attack epochs, \\( A = 10 \\) attack iterations per epoch, \\( K = 100 \\) perturbation renderings, distortion budget \\( \\epsilon = 32 \\), constraint parameter \\( \\eta = 1 \\), and view constraints \\( (13^{\\circ} \\) and \\( 15^{\\circ} \\))." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.529, + 0.907, + 0.649 + ], + "angle": 0, + "content": "For our IPA-Splat method, we adapt IPA-NeRF to 3D Gaussian Splatting with \\( O = 30,000 \\) total iterations and \\( T = 200 \\) normal training iterations per epoch, resulting in \\( O / T = 150 \\) epochs with \\( A = 10 \\) attack iterations each. Other settings match IPA-NeRF. Due to 3DGS's explicit representation, we implement separate parameter constraints (xyz coordinates, feature vectors, scaling, rotation, and opacity) for precise control." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.654, + 0.906, + 0.774 + ], + "angle": 0, + "content": "Evaluation Metrics. Following IPA-NeRF [40], we evaluate using PSNR, SSIM, and LPIPS metrics on two view sets: (1) V-ILLUSORY, focusing on masked metrics for illusory objects, and (2) V-TEST, assessing performance on unseen viewpoints. Attack success is defined as achieving PSNR \\(> 25\\) on V-ILLUSORY while maintaining V-TEST PSNR drop \\(\\leq 3\\), ensuring effective illusion generation and preserved innocent view quality." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.78, + 0.907, + 0.87 + ], + "angle": 0, + "content": "Evaluation Protocol. Our evaluation accounts for varying attack difficulty across camera positions. As shown in Figure 5, datasets exhibit distinct camera patterns. For uniform datasets like Mip-NeRF 360 (e.g., the \"bicycle\" scene) and Tanks-and-Temple, we select the median-indexed frame as the attack viewpoint." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "For Free dataset scenes with irregular camera trajectories (e.g., the \"stair\" scene), we quantify varying attack difficulty" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27404" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.908, + 0.131 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons on single-view attack. Metrics evaluated on Mip-NeRF 360 [5], Tanks & Temples [44], and Free datasets. Our method significantly outperforms baseline attacks in embedding illusory objects (V-IILLUSORY) while maintaining high fidelity in other views (V-TEST)." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.133, + 0.907, + 0.236 + ], + "angle": 0, + "content": "
MethodMip-NeRF 360 [5] datasetTanks & Temples [44] datasetFree [94] dataset
V-IILLUSORYV-TESTV-IILLUSORYV-TESTV-IILLUSORYV-TEST
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Naive 3DGS (w/o attack)13.210.5210.73129.450.8830.16513.150.6160.73230.600.9150.13512.000.3150.90526.800.8260.228
IPA-NeRF [40] (Nerfacto [87])16.000.5820.68521.940.5860.41513.510.6360.71123.880.7300.21813.930.4430.69920.280.4970.532
IPA-NeRF [40] (Instant-NGP [73])17.600.6180.64120.000.5170.47916.050.6930.61620.290.6690.35018.940.5080.51920.430.5030.548
IPA-Splat13.230.5180.74027.390.8290.24713.430.6250.72428.530.8910.19012.600.3720.74424.710.7490.341
Ours27.040.8130.36927.760.8050.28621.330.8090.37127.580.8520.23926.660.7540.31725.250.7280.382
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.249, + 0.908, + 0.304 + ], + "angle": 0, + "content": "Table 2. Quantitative comparisons on single-view attack with different difficulty levels on the Free [94] dataset. We evaluate attack effectiveness (V-ILLUSORY) at varying difficulty levels, defined by our KDE-based evaluation protocol. Our method consistently achieves superior results across all metrics, clearly outperforming state-of-the-art methods, especially in EASY and MEDIAN scenarios, while remaining effective in the challenging HARD scenario." + }, + { + "type": "table", + "bbox": [ + 0.158, + 0.305, + 0.84, + 0.4 + ], + "angle": 0, + "content": "
MethodEASYMEDIANHARD
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
IPA-NeRF (nerfacto)15.040.4820.66213.930.4430.69914.250.4500.728
IPA-NeRF (instant-ngp)18.170.5180.54118.940.5080.51917.950.4870.557
IPA-Splat13.940.4790.65812.600.3720.74313.060.3400.796
Ours29.940.8530.18826.660.7540.31717.530.5260.581
" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.42, + 0.482, + 0.554 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.556, + 0.484, + 0.626 + ], + "angle": 0, + "content": "Figure 5. Our evaluation protocol. We evaluate two scenes with varying difficulties. Left: The \"bicycle\" scene (Mip-NeRF 360 [5]) has uniform camera coverage, providing similar difficulty across views. Right: The \"stair\" scene (Free [94]) has increasing difficulty as later views are visible from more prior viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.66, + 0.303, + 0.675 + ], + "angle": 0, + "content": "using our KDE-based protocol:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.68, + 0.478, + 0.695 + ], + "angle": 0, + "content": "1. Compute overall scene density distribution using KDE." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.695, + 0.483, + 0.725 + ], + "angle": 0, + "content": "2. Calculate camera viewpoint densities within their FOV using camera intrinsics and a \\(10\\%\\) sampling radius." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.726, + 0.484, + 0.755 + ], + "angle": 0, + "content": "3. Sort cameras by density, select three representative viewpoints:" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.68, + 0.484, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.762, + 0.436, + 0.776 + ], + "angle": 0, + "content": "EASY: Minimum density (lowest coverage)" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.783, + 0.449, + 0.798 + ], + "angle": 0, + "content": "- MEDIAN: Median density (average coverage)" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.805, + 0.447, + 0.82 + ], + "angle": 0, + "content": "- HARD: Maximum density (highest coverage)" + }, + { + "type": "list", + "bbox": [ + 0.133, + 0.762, + 0.449, + 0.82 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Experiments in Tab. 2 confirm negative correlation between scene density and attack success (V-IILLUSORY), validating that higher scene coverage increases attack difficulty. This protocol enables fair evaluation across scenes and provides a benchmark for future 3DGS poisoning research." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.424, + 0.905, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.559, + 0.908, + 0.628 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparisons on single-view attack. Our method generates significantly clearer and more convincing illusory objects from the poisoned viewpoint, demonstrating better multiview consistency and fewer artifacts compared to other state-of-the-art methods." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.648, + 0.905, + 0.782 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.786, + 0.909, + 0.855 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparisons of different poisoning methods on innocent views. Unlike baseline methods, our method effectively maintains high visual fidelity across innocent viewpoints, introducing significantly fewer artifacts and ensuring minimal disruption of the original scene appearance." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27405" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.159 + ], + "angle": 0, + "content": "Table 3. Multi-view Attack Evaluation. We quantitatively evaluate our method on multiple poisoned views (V-ILLUSORY) and innocent views (V-TEST). Our approach consistently outperforms state-of-the-art methods, embedding clear illusions in targeted views while maintaining high fidelity in innocent views." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.16, + 0.485, + 0.303 + ], + "angle": 0, + "content": "
# of viewsMethodV-ILLUSORY (poisoned avg.)V-TEST
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
2IPA-NeRF (Nerfacto)16.170.5830.68019.640.4570.548
IPA-NeRF (Instant-NGP)19.190.6240.61618.390.4400.539
IPA-Splat13.240.4970.75227.450.8320.243
Ours27.490.8420.29927.770.8040.286
3IPA-NeRF (Nerfacto)18.480.5840.66019.830.4620.545
IPA-NeRF (Instant-NGP)18.090.6040.64318.630.4580.524
IPA-Splat13.760.5380.73227.970.8580.223
Ours27.040.8330.31127.720.8030.287
4IPA-NeRF (Nerfacto)17.060.6260.67619.610.4670.538
IPA-NeRF (Instant-NGP)19.060.6570.63218.510.4580.523
IPA-Splat13.090.4890.79627.600.8380.228
Ours26.950.8550.30527.590.8020.287
" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.324, + 0.482, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.5, + 0.485, + 0.556 + ], + "angle": 0, + "content": "Figure 8. Qualitative comparisons on multi-view attack. Our density-guided method produces sharper, more consistent illusions across multiple poisoned views, clearly outperforming baseline methods that yield faint or inconsistent results." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.592, + 0.273, + 0.608 + ], + "angle": 0, + "content": "4.2. Single-view Attack" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.619, + 0.486, + 0.71 + ], + "angle": 0, + "content": "Tab. 1 shows quantitative comparisons of single-view attacks. Our method outperforms baselines across all datasets. On V-ILLUSORY views, our approach significantly improves PSNR, SSIM, and LPIPS, effectively embedding convincing illusions. Performance on V-TEST remains consistently high, indicating minimal impact on innocent views." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.715, + 0.485, + 0.807 + ], + "angle": 0, + "content": "Qualitative results (Figs. 6 and 7) further highlight our advantages. In Fig. 6, our method produces clearer and more convincing illusory objects compared to baselines, which often yield faint or inconsistent illusions. Fig. 7 emphasizes our superior visual fidelity in innocent views, with fewer artifacts and better overall scene quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.486, + 0.903 + ], + "angle": 0, + "content": "In Tab. 2, we analyze performance across difficulty levels (EASY, MEDIAN, HARD) on the Free dataset. As expected, attack effectiveness decreases with difficulty. Nonetheless, our density-guided approach achieves superior results, particularly at EASY and MEDIAN levels, demonstrating robust performance even in challenging conditions." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.158 + ], + "angle": 0, + "content": "Table 4. Effect of KDE bandwidth \\( h \\) on attack performance. We examine how KDE bandwidth impacts our density-guided attack. A moderate bandwidth (\\( h = 7.5 \\)) achieves the best balance, maximizing effectiveness on poisoned views while preserving quality in innocent views." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.16, + 0.907, + 0.26 + ], + "angle": 0, + "content": "
Bandwidth hV-ILLUSORYV-TEST
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
0.127.000.8110.37327.830.8050.286
2.526.920.8090.37527.810.8050.286
5.026.950.8110.37527.250.7860.297
7.527.040.8130.36927.760.8050.286
10.026.890.8070.38027.720.8050.286
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.273, + 0.907, + 0.342 + ], + "angle": 0, + "content": "Table 5. Effect of noise scheduling parameters. We analyze initial noise strength \\(\\sigma_0\\) and decay strategies. Higher initial noise \\((\\sigma_0 = 100)\\) with linear decay provides the best balance, maximizing illusory quality (V-ILLUSORY) while preserving fidelity in innocent views (V-TEST)." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.343, + 0.906, + 0.44 + ], + "angle": 0, + "content": "
Initial noise σ0Decay strategyV-ILLUSORYV-TEST
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
30Linear26.470.7950.39828.430.8510.203
30Cosine26.700.8010.38828.300.8450.213
30Square root26.620.7990.39328.270.8450.212
100Linear27.040.8130.36927.760.8050.286
100Cosine26.930.8120.37326.960.7710.315
100Square root26.900.8130.37326.810.7670.319
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.465, + 0.692, + 0.48 + ], + "angle": 0, + "content": "4.3. Multi-view Attack" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.487, + 0.909, + 0.592 + ], + "angle": 0, + "content": "In realistic scenarios, attackers may need to embed illusory objects into multiple viewpoints simultaneously. Unlike single-view attacks, multi-view poisoning balances multiple objectives while preserving scene consistency and fidelity. We rigorously evaluate our approach using the Mip-NeRF 360 dataset [5], poisoned views at \\(0^{\\circ}\\), \\(90^{\\circ}\\), \\(180^{\\circ}\\), and \\(270^{\\circ}\\) with the median view at \\(0^{\\circ}\\) as reference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.593, + 0.909, + 0.699 + ], + "angle": 0, + "content": "We evaluate our density-guided method against baselines (IPA-NeRF [40] with Nerfacto, IPA-NeRF with Instant-NGP, IPA-Splat) on the Mip-NeRF-360 dataset under multi-view attacks (2, 3, and 4 poisoned views). Results in Tab. 3 show our approach consistently outperforms baselines, effectively embedding illusory objects (V-ILLUSORY) while minimally affecting innocent views." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.699, + 0.909, + 0.76 + ], + "angle": 0, + "content": "Fig. 8 qualitatively demonstrates our method's superiority. It consistently generates clear, visually convincing illusions with minimal artifacts, significantly outperforming baselines and maintaining high quality from innocent views." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.769, + 0.681, + 0.785 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "KDE bandwidth. The KDE bandwidth significantly affects density estimation smoothness. Evaluations (Tab. 4) show medium bandwidth (\\(h = 7.5\\)) achieves optimal balance, maximizing V-IILLUSORY effectiveness while preserving V-TEST quality. Smaller bandwidths (\\(h = 0.1\\)) overly restrict point placement, while larger bandwidths (\\(h = 10.0\\)) decrease attack precision." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27406" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.146 + ], + "angle": 0, + "content": "Table 6. Comparison of different attack strategy combinations. We evaluate the impact of each component of our proposed method on poisoning effectiveness. Combining all strategies achieves the best results, significantly improving rendering quality of the illusory object (V-ILLUSORY) and maintaining satisfactory performance in innocent views (V-TEST), resulting in optimal Attack Success Rate (ASR). The combination of point cloud poisoning and noise scheduling is crucial for successful attacks, highlighting their complementary nature." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.147, + 0.905, + 0.256 + ], + "angle": 0, + "content": "
Poisoned View GT ReplacementDensity-Guided Point Cloud AttackView Consistency Disruption AttackV-ILLUSORYV-TESTASR (PSNR)
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓V-ILLUSORY > 25 & V-TEST drop ≤ 3
✓13.220.5210.73029.460.8840.1640/7
✓✓26.010.7750.42729.400.8830.1646/7
✓✓13.310.5220.74727.790.8050.2860/7
✓✓✓27.040.8130.36927.760.8050.2867/7
" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.279, + 0.172, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.173, + 0.279, + 0.249, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.25, + 0.279, + 0.326, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.279, + 0.403, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.403, + 0.279, + 0.48, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.371, + 0.485, + 0.455 + ], + "angle": 0, + "content": "Figure 9. Qualitative analysis of attack component combinations. We compare three attack strategies: (1) direct image poisoning, (2) density-guided point cloud poisoning, and (3) multi-view consistency disruption. Combining all three achieves the most realistic illusions across various scenes from the Mip-NeRF 360 [5] dataset, highlighting their complementary effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.485, + 0.484, + 0.576 + ], + "angle": 0, + "content": "Noise scheduling. We evaluated noise scheduling strategies, varying initial noise intensities \\((\\sigma_0)\\) and decay rates (linear, cosine, and square root), summarized in Tab. 5. Higher initial noise \\((\\sigma_0 = 100)\\) with slower linear decay achieved optimal balance, greatly enhancing attack effectiveness with moderate impact on innocent views." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.484, + 0.734 + ], + "angle": 0, + "content": "Attack components. We analyzed combinations of direct target-view image poisoning, density-guided point cloud poisoning (KDE-based), and noise-based view consistency disruption. Quantitative results (Tab. 6) show that direct image poisoning alone is ineffective. Combining image poisoning with density-guided poisoning notably improves outcomes. Integrating all three components achieves the best results, embedding robust illusions and preserving rendering quality. Qualitative results (Fig. 9) visually confirm the superior clarity and effectiveness of this combined approach." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.753, + 0.21, + 0.768 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.486, + 0.901 + ], + "angle": 0, + "content": "We presented a density-guided poisoning method for 3DGS, strategically injecting illusory objects and disrupting multiview consistency via adaptive noise. Experiments show our approach outperforms existing baselines, effectively embedding convincing illusions with minimal impact on innocent views. Our work highlights critical vulnerabilities in 3D representation models, providing a robust framework for future security research." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.279, + 0.903, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.394, + 0.903, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.504, + 0.907, + 0.588 + ], + "angle": 0, + "content": "Figure 10. Effect on neighboring views (\"counter\" scene, Mip-NeRF 360 [5]). Clockwise and counterclockwise shifts from the attack view (0 degrees) up to 20 degrees show PSNR between clean and poisoned renderings. IPA-NeRF significantly lowers PSNR across most angles, whereas our method mainly impacts views within five degrees, preserving quality beyond this range." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.603, + 0.603, + 0.731 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.603, + 0.678, + 0.731 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.678, + 0.603, + 0.751, + 0.731 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.603, + 0.828, + 0.724 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.829, + 0.603, + 0.904, + 0.724 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.733, + 0.907, + 0.789 + ], + "angle": 0, + "content": "Figure 11. Visualization of evaluation protocol (\"grass\" scene, Free [94]). Our method achieves clearly visible illusory objects in EASY and MEDIAN scenarios and maintains robust performance even under challenging HARD conditions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.816, + 0.907, + 0.89 + ], + "angle": 0, + "content": "Limitations. Our method struggles with scenes having highly overlapping views or complex camera trajectories due to 3DGS's strict multi-view consistency. Future work should address this balance between consistency and attack effectiveness." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27407" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.212 + ], + "angle": 0, + "content": "Acknowledgements. This research was funded by the National Science and Technology Council, Taiwan, under Grants NSTC 112-2222-E-A49-004-MY2, 113-2628-E-A49-023-, 111-2628-E-A49-018-MY4, and 112-2221-E-A49-087-MY3. The authors are grateful to Google, NVIDIA, and MediaTek Inc. for their generous donations. Yu-Lun Liu acknowledges the Yushan Young Fellow Program by the MOE in Taiwan." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.226, + 0.188, + 0.241 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.251, + 0.486, + 0.293 + ], + "angle": 0, + "content": "[1] Naveed Akhtar and Ajmal Mian. Threat of adversarial attacks on deep learning in computer vision: A survey. IEEE Access, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.293, + 0.484, + 0.349 + ], + "angle": 0, + "content": "[2] Maksym Andriushchenko, Francesco Croce, Nicolas Flammarion, and Matthias Hein. Square attack: a query-efficient black-box adversarial attack via random search. In European Conference on Computer Vision (ECCV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.35, + 0.483, + 0.406 + ], + "angle": 0, + "content": "[3] Anurag Arnab, Ondrej Miksik, and Philip HS Torr. On the robustness of semantic segmentation models to adversarial attacks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.407, + 0.485, + 0.476 + ], + "angle": 0, + "content": "[4] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.478, + 0.485, + 0.546 + ], + "angle": 0, + "content": "[5] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.548, + 0.484, + 0.589 + ], + "angle": 0, + "content": "[6] Arjun Nitin Bhagoji, Warren He, Bo Li, and Dawn Song. Exploring the space of black-box attacks on deep neural networks. *ArXiv:1712.09491*, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.591, + 0.484, + 0.631 + ], + "angle": 0, + "content": "[7] Battista Biggio, Blaine Nelson, and Pavel Laskov. Poisoning attacks against support vector machines. ArXiv:1206.6389, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.633, + 0.482, + 0.675 + ], + "angle": 0, + "content": "[8] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In IEEE Symposium on Security and Privacy (S&P), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.676, + 0.485, + 0.731 + ], + "angle": 0, + "content": "[9] Anirban Chakraborty, Manaar Alam, Vishal Dey, Anupam Chattopadhyay, and Debdeep Mukhopadhyay. Adversarial attacks and defences: A survey. ArXiv:1810.00069, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.733, + 0.484, + 0.815 + ], + "angle": 0, + "content": "[10] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.817, + 0.484, + 0.872 + ], + "angle": 0, + "content": "[11] Bo-Yu Chen, Wei-Chen Chiu, and Yu-Lun Liu. Improving robustness for joint optimization of camera pose and decomposed low-rank tensorial radiance fields. In AAAI Conference on Artificial Intelligence (AAAI), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.873, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[12] Jianbo Chen, Michael I Jordan, and Martin J Wainwright. Hopskipjumpattack: A query-efficient decision-based attack." + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.251, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.908, + 0.12 + ], + "angle": 0, + "content": "In IEEE Symposium on Security and Privacy (S&P), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.122, + 0.908, + 0.191 + ], + "angle": 0, + "content": "[13] Pin-Yu Chen, Huan Zhang, Yash Sharma, Jinfeng Yi, and Cho-Jui Hsieh. Zoo: Zeroth order optimization based blackbox attacks to deep neural networks without training substitute models. In ACM Workshop on Artificial Intelligence and Security (AISec), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.192, + 0.907, + 0.235 + ], + "angle": 0, + "content": "[14] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. ArXiv:1712.05526, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.236, + 0.907, + 0.277 + ], + "angle": 0, + "content": "[15] Zilong Chen, Feng Wang, Yikai Wang, and Huaping Liu. Text-to-3d using gaussian splatting. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.278, + 0.907, + 0.346 + ], + "angle": 0, + "content": "[16] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. In International Conference on Machine Learning (ICML), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.349, + 0.907, + 0.405 + ], + "angle": 0, + "content": "[17] Gabriela F Cretu, Angelos Stavrou, Michael E Locasto, Salvatore J Stolfo, and Angelos D Keromytis. Casting out demons: Sanitizing training data for anomaly sensors. In IEEE Symposium on Security and Privacy (S&P), 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.406, + 0.907, + 0.461 + ], + "angle": 0, + "content": "[18] Zeyu Dai, Shengcai Liu, Qing Li, and Ke Tang. Saliency attack: Towards imperceptible black-box adversarial attack. ACM Transactions on Intelligent Systems and Technology (TIST), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.462, + 0.906, + 0.518 + ], + "angle": 0, + "content": "[19] Ilias Diakonikolas, Gautam Kamath, Daniel Kane, Jerry Li, Jacob Steinhardt, and Alistair Stewart. Sever: A robust meta-algorithm for stochastic optimization. In International Conference on Machine Learning (ICML), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.519, + 0.907, + 0.575 + ], + "angle": 0, + "content": "[20] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading defenses to transferable adversarial examples by translation-invariant attacks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.576, + 0.907, + 0.645 + ], + "angle": 0, + "content": "[21] Cheng-De Fan, Chen-Wei Chang, Yi-Ruei Liu, Jie-Ying Lee, Jiun-Long Huang, Yu-Chee Tseng, and Yu-Lun Liu. Spectromotion: Dynamic 3d reconstruction of specular scenes. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.647, + 0.907, + 0.688 + ], + "angle": 0, + "content": "[22] Guangchi Fang and Bing Wang. Mini-splatting: Representing scenes with a constrained number of gaussians. In European Conference on Computer Vision (ECCV), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.689, + 0.907, + 0.744 + ], + "angle": 0, + "content": "[23] Minghong Fang, Neil Zhenqiang Gong, and Jia Liu. Influence function based data poisoning attacks to top-n recommender systems. In The Web Conference (WWW), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.746, + 0.906, + 0.801 + ], + "angle": 0, + "content": "[24] John Flynn, Ivan Neulander, James Philbin, and Noah Snavely. Deepstereo: Learning to predict new views from the world's imagery. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.802, + 0.907, + 0.859 + ], + "angle": 0, + "content": "[25] Yonggan Fu, Ye Yuan, Souvik Kundu, Shang Wu, Shun-yao Zhang, and Yingyan Lin. Nerfool: Uncovering the vulnerability of generalizable neural radiance fields against adversarial perturbations. ArXiv:2306.06359, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.86, + 0.907, + 0.902 + ], + "angle": 0, + "content": "[26] Micah Goldblum, Dimitris Tsipras, Chulin Xie, Xinyun Chen, Avi Schwarzschild, Dawn Song, Aleksander Madry, Bo Li, and Tom Goldstein. Dataset security for machine" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.908, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27408" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.484, + 0.134 + ], + "angle": 0, + "content": "learning: Data poisoning, backdoor attacks, and defenses. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.135, + 0.484, + 0.175 + ], + "angle": 0, + "content": "[27] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. ArXiv:1412.6572, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.177, + 0.484, + 0.203 + ], + "angle": 0, + "content": "[28] Markus Gross and Hanspeter Pfister. Point-based graphics. Elsevier, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.205, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[29] Tianyu Gu, Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Evaluating backdooring attacks on deep neural networks. IEEE Access, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.247, + 0.484, + 0.3 + ], + "angle": 0, + "content": "[30] Chuan Guo, Jacob Gardner, Yurong You, Andrew Gordon Wilson, and Kilian Weinberger. Simple black-box adversarial attacks. In International Conference on Machine Learning (ICML), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.302, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[31] Abdullah Hamdi, Sara Rojas, Ali Thabet, and Bernard Ghanem. Advpc: Transferable adversarial perturbations on 3d point clouds. In European Conference on Computer Vision (ECCV), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.359, + 0.484, + 0.399 + ], + "angle": 0, + "content": "[32] András Horváth and Csaba M Józsa. Targeted adversarial attacks on generalizable neural radiance fields. In International Conference on Computer Vision (ICCV), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.401, + 0.484, + 0.482 + ], + "angle": 0, + "content": "[33] Hao-Yu Hou, Chia-Chi Hsu, Yu-Chen Huang, Mu-Yi Shen, Wei-Fang Sun, Cheng Sun, Chia-Che Chang, Yu-Lun Liu, and Chun-Yi Lee. 3d gaussian splatting with grouped uncertainty for unconstrained images. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.484, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[34] Hanxun Huang, Xingjun Ma, Sarah Monazam Erfani, James Bailey, and Yisen Wang. Unlearnable examples: Making personal data unexploitable. ArXiv:2101.04898, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.526, + 0.484, + 0.594 + ], + "angle": 0, + "content": "[35] Xiufeng Huang, Ruiqi Li, Yiu-ming Cheung, Ka Chun Cheung, Simon See, and Renjie Wan. Gaussianmarker: Uncertainty-aware copyright protection of 3d gaussian splatting. Advances in Neural Information Processing Systems (NeurIPS), 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.596, + 0.484, + 0.635 + ], + "angle": 0, + "content": "[36] Zhichao Huang and Tong Zhang. Black-box adversarial attack with transferable model-based embedding. ArXiv:1911.07140, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.637, + 0.484, + 0.705 + ], + "angle": 0, + "content": "[37] Matthew Jagielski, Alina Oprea, Battista Biggio, Chang Liu, Cristina Nita-Rotaru, and Bo Li. Manipulating machine learning: Poisoning attacks and countermeasures for regression learning. In IEEE Symposium on Security and Privacy (S&P), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.707, + 0.484, + 0.761 + ], + "angle": 0, + "content": "[38] Youngdong Jang, Hyunjie Park, Feng Yang, Heeju Ko, Euijin Choo, and Sangpil Kim. 3d-gsw: 3d gaussian splatting watermark for protecting copyrights in radiance fields. ArXiv:2409.13222, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.763, + 0.484, + 0.802 + ], + "angle": 0, + "content": "[39] Yujie Ji, Xinyang Zhang, and Ting Wang. Backdoor attacks against learning systems. In IEEE Conference on Communications and Network Security (CNS), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.804, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[40] Wenxiang Jiang, Hanwei Zhang, Shuo Zhao, Zhongwen Guo, and Hao Wang. Ipa-nerf: Illusory poisoning attack against neural radiance fields. In European Conference on Artificial Intelligence (ECAI), 2024. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.86, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[41] James T Kajiya and Brian P Von Herzen. Ray tracing volume densities. ACM SIGGRAPH Computer Graphics (SIGGRAPH), 1984. 3" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.093, + 0.907, + 0.147 + ], + "angle": 0, + "content": "[42] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.149, + 0.907, + 0.217 + ], + "angle": 0, + "content": "[43] Michael Kern, Christoph Neuhauser, Torben Maack, Mengjiao Han, Will Usher, and Rüdiger Westermann. A comparison of rendering techniques for 3d line sets with transparency. IEEE Transactions on Visualization and Computer Graphics (TVCG), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.219, + 0.907, + 0.272 + ], + "angle": 0, + "content": "[44] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (TOG), 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.274, + 0.906, + 0.328 + ], + "angle": 0, + "content": "[45] Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. Hugs: Human gaussian splats. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.33, + 0.907, + 0.371 + ], + "angle": 0, + "content": "[46] Pang Wei Koh, Jacob Steinhardt, and Percy Liang. Stronger data poisoning attacks break data sanitization defenses. Machine Learning, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.372, + 0.907, + 0.425 + ], + "angle": 0, + "content": "[47] Alexey Kurakin, Ian J Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In Artificial intelligence safety and security. Chapman and Hall/CRC, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.428, + 0.907, + 0.467 + ], + "angle": 0, + "content": "[48] Keita Kurita, Paul Michel, and Graham Neubig. Weight poisoning attacks on pre-trained models. ArXiv:2004.06660, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.47, + 0.907, + 0.524 + ], + "angle": 0, + "content": "[49] Jiahui Lei, Yufu Wang, Georgios Pavlakos, Lingjie Liu, and Kostas Daniilidis. Gart: Gaussian articulated template models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.526, + 0.907, + 0.579 + ], + "angle": 0, + "content": "[50] Bo Li, Yining Wang, Aarti Singh, and Yevgeniy Vorobeychik. Data poisoning attacks on factorization-based collaborative filtering. Advances in Neural Information Processing Systems (NeurIPS), 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.581, + 0.906, + 0.649 + ], + "angle": 0, + "content": "[51] Chenxin Li, Hengyu Liu, Zhiwen Fan, Wuyang Li, Yifan Liu, Panwang Pan, and Yixuan Yuan. Instantsplamp: Fast and generalizable stenography framework for generative gaussian splatting. In International Conference on Learning Representations (ICLR), 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.651, + 0.907, + 0.718 + ], + "angle": 0, + "content": "[52] Shaofeng Li, Minhui Xue, Benjamin Zi Hao Zhao, Haojin Zhu, and Xinpeng Zhang. Invisible backdoor attacks on deep neural networks via steganography and regularization. IEEE Transactions on Dependable and Secure Computing (TDSC), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.72, + 0.907, + 0.761 + ], + "angle": 0, + "content": "[53] Yufei Li, Zexin Li, Yingfan Gao, and Cong Liu. White-box multi-objective adversarial attack on dialogue generation. ArXiv:2305.03655, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.763, + 0.907, + 0.83 + ], + "angle": 0, + "content": "[54] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation from diffusion models via adversarial examples. ArXiv:2302.04578, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.832, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[55] Chin-Yang Lin, Chung-Ho Wu, Chang-Han Yeh, Shih-Han Yen, Cheng Sun, and Yu-Lun Liu. Frugalnerf: Fast convergence for extreme few-shot novel view synthesis without learned priors. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 3" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27409" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.092, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[56] Daniel Liu, Ronald Yu, and Hao Su. Extending adversarial attacks and defenses to deep 3d point cloud classifiers. In IEEE International Conference on Image Processing (ICIP), 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.15, + 0.485, + 0.191 + ], + "angle": 0, + "content": "[57] Xinhang Liu, Yu-Wing Tai, and Chi-Keung Tang. Clean-nerf: Reformulating nerf to account for view-dependent observations. ArXiv:2303.14707, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.194, + 0.484, + 0.235 + ], + "angle": 0, + "content": "[58] Yanpei Liu, Xinyun Chen, Chang Liu, and Dawn Song. Delving into transferable adversarial examples and black-box attacks. ArXiv:1611.02770, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.238, + 0.484, + 0.293 + ], + "angle": 0, + "content": "[59] Yingqi Liu, Shiqing Ma, Yousra Aafer, Wen-Chuan Lee, Juan Zhai, Weihang Wang, and Xiangyu Zhang. Trojaning attack on neural networks. In Network and Distributed System Security Symposium (NDSS), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.296, + 0.484, + 0.364 + ], + "angle": 0, + "content": "[60] Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Johannes Kopf, and Jia-Bin Huang. Robust dynamic radiance fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.484, + 0.409 + ], + "angle": 0, + "content": "[61] Zihao Liu, Tianhao Wang, Mengdi Huai, and Chenglin Miao. Backdoor attacks via machine unlearning. In AAAI Conference on Artificial Intelligence (AAAI), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.412, + 0.484, + 0.453 + ], + "angle": 0, + "content": "[62] Jiahao Lu, Yifan Zhang, Qiuhong Shen, Xinchao Wang, and Shuicheng Yan. Poison-splat: Computation cost attack on 3d gaussian splatting. ArXiv:2410.08190, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.456, + 0.484, + 0.51 + ], + "angle": 0, + "content": "[63] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In International Conference on 3D Vision (3DV), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.513, + 0.484, + 0.568 + ], + "angle": 0, + "content": "[64] Ziyuan Luo, Qing Guo, Ka Chun Cheung, Simon See, and Renjie Wan. Copyrnerf: Protecting the copyright of neural radiance fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.571, + 0.484, + 0.611 + ], + "angle": 0, + "content": "[65] Yuzhe Ma, Xiaojin Zhu, and Justin Hsu. Data poisoning against differentially-private learners: Attacks and defenses. ArXiv:1903.09860, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.615, + 0.484, + 0.668 + ], + "angle": 0, + "content": "[66] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. ArXiv:1706.06083, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.673, + 0.484, + 0.74 + ], + "angle": 0, + "content": "[67] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.744, + 0.484, + 0.785 + ], + "angle": 0, + "content": "[68] Nelson Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics (TVCG), 1995. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.788, + 0.484, + 0.856 + ], + "angle": 0, + "content": "[69] Andreas Meuleman, Yu-Lun Liu, Chen Gao, Jia-Bin Huang, Changil Kim, Min H Kim, and Johannes Kopf. Progressively optimized local radiance fields for robust view synthesis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[70] Jian-Xun Mi, Xu-Dong Wang, Li-Fang Zhou, and Kun Cheng. Adversarial examples based on object detection tasks: A survey. Neurocomputing, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.908, + 0.16 + ], + "angle": 0, + "content": "[71] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.164, + 0.907, + 0.219 + ], + "angle": 0, + "content": "[72] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.221, + 0.908, + 0.275 + ], + "angle": 0, + "content": "[73] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (TOG), 2022. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.278, + 0.907, + 0.346 + ], + "angle": 0, + "content": "[74] Luis Muñoz-González, Battista Biggio, Ambra Demontis, Andrea Paudice, Vasin Wongrassamee, Emil C Lupu, and Fabio Roli. Towards poisoning of deep learning algorithms with back-gradient optimization. In ACM Workshop on Artificial Intelligence and Security (AISec), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.349, + 0.908, + 0.417 + ], + "angle": 0, + "content": "[75] Nicolas Papernot, Patrick McDaniel, Ian Goodfellow, Somesh Jha, Z Berkay Celik, and Ananthram Swami. Practical black-box attacks against machine learning. In ACM Asia Conference on Computer and Communications Security (ASIACCS), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.42, + 0.907, + 0.487 + ], + "angle": 0, + "content": "[76] Andrea Paudice, Luis Muñoz-González, and Emil C Lupu. Label sanitization against label flipping poisoning attacks. In European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML PKDD), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.49, + 0.907, + 0.531 + ], + "angle": 0, + "content": "[77] Thomas Porter and Tom Duff. Compositing digital images. In ACM SIGGRAPH Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), 1984. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.533, + 0.908, + 0.615 + ], + "angle": 0, + "content": "[78] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.618, + 0.907, + 0.658 + ], + "angle": 0, + "content": "[79] Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, Andrew Ilyas, and Aleksander Madry. Raising the cost of malicious ai-powered image editing. ArXiv:2302.06588, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.661, + 0.908, + 0.729 + ], + "angle": 0, + "content": "[80] Ali Shafahi, W Ronny Huang, Mahyar Najibi, Octavian Suciu, Christoph Studer, Tudor Dumitras, and Tom Goldstein. Poison frogs! targeted clean-label poisoning attacks on neural networks. Advances in Neural Information Processing Systems (NeurIPS), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.732, + 0.908, + 0.799 + ], + "angle": 0, + "content": "[81] Ilia Shumailov, Yiren Zhao, Daniel Bates, Nicolas Papernot, Robert Mullins, and Ross Anderson. Sponge examples: Energy-latency attacks on neural networks. In IEEE European symposium on security and privacy (EuroS&P), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.803, + 0.908, + 0.857 + ], + "angle": 0, + "content": "[82] Qi Song, Ziyuan Luo, Ka Chun Cheung, Simon See, and Renjie Wan. Geometry cloak: Preventing tgs-based 3d reconstruction from copyrighted images. Advances in Neural Information Processing Systems (NeurIPS), 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.86, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[83] Jacob Steinhardt, Pang Wei W Koh, and Percy S Liang. Certified defenses for data poisoning attacks. Advances in Neural Information Processing Systems (NeurIPS), 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27410" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[84] Chih-Hai Su, Chih-Yao Hu, Shr-Ruei Tsai, Jie-Ying Lee, Chin-Yang Lin, and Yu-Lun Liu. Boostmvsnerfs: Boosting mvs-based nerfs to generalizable view synthesis in large-scale scenes. In ACM SIGGRAPH Conference Papers (SIGGRAPH), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.162, + 0.485, + 0.232 + ], + "angle": 0, + "content": "[85] Octavian Suciu, Radu Marginean, Yigitcan Kaya, Hal Daume III, and Tudor Dumitras. When does machine learning {FAIL}? generalized transferability for evasion and poisoning attacks. In USENIX Security Symposium (USENIX Security), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.233, + 0.484, + 0.286 + ], + "angle": 0, + "content": "[86] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. ArXiv:1312.6199, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.288, + 0.484, + 0.37 + ], + "angle": 0, + "content": "[87] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, et al. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH Conference Proceedings (SIGGRAPH), 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.371, + 0.484, + 0.44 + ], + "angle": 0, + "content": "[88] Ayush Tewari, Ohad Fried, Justus Thies, Vincent Sitzmann, Stephen Lombardi, Kalyan Sunkavalli, Ricardo MartinBrualla, Tomas Simon, Jason Saragih, Matthias Nießner, et al. State of the art on neural rendering. In Computer Graphics Forum. Wiley Online Library, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.441, + 0.484, + 0.482 + ], + "angle": 0, + "content": "[89] Brandon Tran, Jerry Li, and Aleksander Madry. Spectral signatures in backdoor attacks. Advances in Neural Information Processing Systems (NeurIPS), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.484, + 0.484, + 0.538 + ], + "angle": 0, + "content": "[90] Yu-Lin Tsai, Chia-Yi Hsu, Chia-Mu Yu, and Pin-Yu Chen. Formalizing generalization and adversarial robustness of neural networks to weight perturbations. Advances in Neural Information Processing Systems (NeurIPS), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.539, + 0.484, + 0.579 + ], + "angle": 0, + "content": "[91] Alexander Turner, Dimitris Tsipras, and Aleksander Madry. Label-consistent backdoor attacks. ArXiv:1912.02771, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.581, + 0.484, + 0.621 + ], + "angle": 0, + "content": "[92] Eric Wallace, Tony Z Zhao, Shi Feng, and Sameer Singh. Concealed data poisoning attacks on nlp models. ArXiv:2010.12563, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.622, + 0.484, + 0.678 + ], + "angle": 0, + "content": "[93] Chen Wang, Xian Wu, Yuan-Chen Guo, Song-Hai Zhang, Yu-Wing Tai, and Shi-Min Hu. Nerf-sr: High quality neural radiance fields using supersampling. In ACM International Conference on Multimedia (ACM MM), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.679, + 0.484, + 0.747 + ], + "angle": 0, + "content": "[94] Peng Wang, Yuan Liu, Zhaoxi Chen, Lingjie Liu, Ziwei Liu, Taku Komura, Christian Theobalt, and Wenping Wang. F2-nerf: Fast neural radiance field training with free camera trajectories. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.748, + 0.484, + 0.802 + ], + "angle": 0, + "content": "[95] Yixiang Wang, Jiqiang Liu, Xiaolin Chang, Ricardo J Rodríguez, and Jianhua Wang. Di-aa: An interpretable white-box attack for fooling deep neural networks. Information Sciences, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.804, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[96] Chong Xiang, Charles R Qi, and Bo Li. Generating 3d adversarial point clouds. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[97] Chaowei Xiao, Dawei Yang, Bo Li, Jia Deng, and Mingyan Liu. Meshadv: Adversarial meshes for visual recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.092, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.907, + 0.147 + ], + "angle": 0, + "content": "[98] Huang Xiao, Battista Biggio, Gavin Brown, Giorgio Fumera, Claudia Eckert, and Fabio Roli. Is feature selection secure against training data poisoning? In International Conference on Machine Learning (ICML), 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.149, + 0.907, + 0.205 + ], + "angle": 0, + "content": "[99] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Yuyin Zhou, Lingxi Xie, and Alan Yuille. Adversarial examples for semantic segmentation and object detection. In International Conference on Computer Vision (ICCV), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.907, + 0.261 + ], + "angle": 0, + "content": "[100] Yuanwang Yang, Qiao Feng, Yu-Kun Lai, and Kun Li. Real-time 3d human reconstruction and rendering system from a single rgb camera. In ACM SIGGRAPH Asia Technical Communications (SIGGRAPH Asia). ACM, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.906, + 0.29 + ], + "angle": 0, + "content": "[101] Linfeng Ye and Shayan Mohajer Hamidi. Thundernna: a white box adversarial attack. ArXiv:2111.12305, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.907, + 0.347 + ], + "angle": 0, + "content": "[102] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.348, + 0.907, + 0.389 + ], + "angle": 0, + "content": "[103] Abdurrahman Zeybey, Mehmet Ergezer, and Tommy Nguyen. Gaussian splatting under attack: Investigating adversarial noise in 3d objects. ArXiv:2412.02803, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.907, + 0.446 + ], + "angle": 0, + "content": "[104] Yu-Ting Zhan, Cheng-Yuan Ho, Hebi Yang, Yi-Hsin Chen, Jui Chiu Chiang, Yu-Lun Liu, and Wen-Hsiao Peng. Cat-3dgs: A context-adaptive triplane approach to rate-distortion-optimized 3dgs compression. ArXiv:2503.00357, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.906, + 0.503 + ], + "angle": 0, + "content": "[105] Jie Zhang, Chen Dongdong, Qidong Huang, Jing Liao, Weiming Zhang, Huamin Feng, Gang Hua, and Nenghai Yu. Poison ink: Robust and invisible backdoor attack. IEEE Transactions on Image Processing (TIP), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.906, + 0.56 + ], + "angle": 0, + "content": "[106] Xuanyu Zhang, Jiarui Meng, Zhipei Xu, Shuzhou Yang, Yanmin Wu, Ronggang Wang, and Jian Zhang. Securegs: Boosting the security and fidelity of 3d gaussian splatting steganography. ArXiv:2503.06118, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.907, + 0.601 + ], + "angle": 0, + "content": "[107] Yutong Zhang, Yao Li, Yin Li, and Zhichang Guo. A review of adversarial attacks in computer vision. ArXiv:2308.07673, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.603, + 0.907, + 0.657 + ], + "angle": 0, + "content": "[108] Meixi Zheng, Xuanchen Yan, Zihao Zhu, Hongrui Chen, and Baoyuan Wu. Blackboxbench: A comprehensive benchmark of black-box adversarial attacks. ArXiv:2312.16979, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.907, + 0.715 + ], + "angle": 0, + "content": "[109] Tinghui Zhou, Shubham Tulsiani, Weilun Sun, Jitendra Malik, and Alexei A Efros. View synthesis by appearance flow. In European Conference on Computer Vision (ECCV), 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.716, + 0.907, + 0.772 + ], + "angle": 0, + "content": "[110] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Surface splattering. In ACM SIGGRAPH Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), 2001. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.772 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27411" + } + ] +] \ No newline at end of file diff --git a/2025/StealthAttack_ Robust 3D Gaussian Splatting Poisoning via Density-Guided Illusions/b6c26c82-2907-466d-9aed-934611879ad9_origin.pdf b/2025/StealthAttack_ Robust 3D Gaussian Splatting Poisoning via Density-Guided Illusions/b6c26c82-2907-466d-9aed-934611879ad9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..45d62e4733ad99be8f3fa4f574d4c76ee6711d71 --- /dev/null +++ b/2025/StealthAttack_ Robust 3D Gaussian Splatting Poisoning via Density-Guided Illusions/b6c26c82-2907-466d-9aed-934611879ad9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4b1159555c965f802f693f5d95adc03f15ddb8b3a0d94901b60b8f0d7cf812ac +size 5967923 diff --git a/2025/StealthAttack_ Robust 3D Gaussian Splatting Poisoning via Density-Guided Illusions/full.md b/2025/StealthAttack_ Robust 3D Gaussian Splatting Poisoning via Density-Guided Illusions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5b0d3375de044da4f84f28e080b0c145486936db --- /dev/null +++ b/2025/StealthAttack_ Robust 3D Gaussian Splatting Poisoning via Density-Guided Illusions/full.md @@ -0,0 +1,389 @@ +# StealthAttack: Robust 3D Gaussian Splatting Poisoning via Density-Guided Illusions + +Bo-Hsu Ke You-Zhe Xie Yu-Lun Liu Wei-Chen Chiu + +National Yang Ming Chiao Tung University + +# Abstract + +3D scene representation methods like Neural Radience Fields (NeRF) and 3D Gaussian Splatting (3DGS) have significantly advanced novel view synthesis. As these methods become prevalent, addressing their vulnerabilities becomes critical. We analyze 3DGS robustness against image-level poisoning attacks and propose a novel density-guided poisoning method. Our method strategically injects Gaussian points into low-density regions identified via Kernel Density Estimation (KDE), embedding viewpoint-dependent illusory objects clearly visible from poisoned views while minimally affecting innocent views. Additionally, we introduce an adaptive noise strategy to disrupt multi-view consistency, further enhancing attack effectiveness. We propose a KDE-based evaluation protocol to assess attack difficulty systematically, enabling objective benchmarking for future research. Extensive experiments demonstrate our method's superior performance compared to state-of-the-art techniques. Project page: https://hentci.github.io/stealthattack/ + +# 1. Introduction + +3D scene representation methods, such as Neural Radiance Fields (NeRF)[71] and 3D Gaussian Splatting (3DGS)[42], have significantly advanced novel view synthesis, accurately modeling complex scene geometry and appearance. Along with their popularity, the protection of 3D digital content encoded in these representations has become a matter of concern, where we have witnessed the corresponding watermarking or steganography techniques being proposed in recent years. For instance, GaussianMarker [35] and 3DGSW [38] embed watermarks (mainly binary messages) into Gaussian parameters of 3DGS with minimal visual impact, coupled with dedicated decoders to decipher the hidden messages. Moreover, embedding or hiding extraneous information/messages into 3D scene representations is also directly connected to the risk of data poisoning (where the extraneous information as the poison appears in the training data + +![](images/097bb94581b239f08eff96dac2236fd0d5e7ecd172d3872445a90e5f15c8cfde.jpg) +Figure 1. Illustration of our proposed Density-Guided Poisoning Attack for 3D Gaussian Splatting (3DGS). Our method strategically distribute the Gaussian points of the illusory object (i.e. the red vehicle) among the low-density regions which are discovered along the rays casted from the virtual camera of the poisoned view (i.e. the target view that we would like to attack), making the illusory object clearly visible from the poisoned view while having the minimal interference for the rendering quality on the other non-target/innocent views. + +and is encoded into the 3D representations during model training), in which the further utilization or visualization of the poisoned 3D representations would lead to abnormal or even malicious model behaviours (i.e. the negative impact stemmed from the poison is triggered). To this end, in this work, we focus on the investigation of poisoning attacks on 3D scene representation methods, as they become integral to safety-critical applications. Hence, addressing their security vulnerabilities is not only of utmost importance but also urgent. + +While there exists a prior work of studying the poisoning attack upon NeRF, i.e. IPA-NeRF [40] which effectively exploits NeRF's implicit representations to embed targeted visual illusions (i.e. the illusion as poison will appear while rendering the scene from a certain viewing direction), its applicability to explicit 3D scene representations (particularly 3DGS) however remains limited and not directly transferable. Considering the rapidly growing application scenarios + +of 3DGS (thanks to its capability of fast rendering and accurately capturing scene geometry), we devote our research effort to realizing poisoning attacks on 3D Gaussian splatt. To the best of our knowledge, this is the first work of its kind. While concurrent work, Poison-Splat [62] focuses on computational cost attacks, our work targets visible illusion embedding. In particular, we would like to inject the visible illusory objects (i.e., poison) onto a target view (named as poisoned view) while keeping the other non-target views (named as innocent views) unaffected, as shown in Figure 1. Our work starts from conducting an investigation (cf. Figure 2) upon the robustness of 3DGS against the prior image-level poisoning methods such as IPA-NeRF, where we find that the attempts of directly adopting IPA-NeRF's approach or naively injecting illusory content into training images easily fail, as 3DGS's inherent multi-view consistency and densification processes effectively neutralize or significantly weaken these attacks. + +Motivated by the aforementioned investigation, we propose a density-guided poisoning method for 3DGS. Our approach (cf. Figure 1) strategically identifies low-density regions in the initial Gaussian point cloud using Kernel Density Estimation (KDE), in which the points of illusory objects are then distributed among the low-density regions along the rays casting from the virtual camera of target view (i.e. the rays are casted from the virtual camera with the target viewing direction). These points effectively embed illusory objects which would be clearly visible from targeted views, while having minimal impact on other innocent views (i.e. being less perceptible). Moreover, we introduce the adaptive Gaussian noise into innocent views during training for disrupting the property of multi-view consistency in 3DGS, in order to further enhance the overall efficacy of the attack. We conduct extensive experiments, and the results demonstrate the consistent superiority of our proposed poisoning method in comparison to several baselines. The contribution of our work can be summarized as follows: + +- We are the first work to address data poisoning attacks upon 3D Gaussian Splatting for illusory objects injection. +- We identify and analyze the robustness of 3DGS against the prior poisoning attack techniques. +- We propose a density-guided poisoning method tailored for 3DGS, introducing adaptive noise scheduling to disrupt the multi-view consistency of 3DGS and better realize the entire attack. + +# 2. Related Work + +Adversarial Attack. Adversarial attacks [8, 27, 54, 66, 79, 86, 90] are a critical research area in machine learning and computer vision [1, 9, 107]. These attacks exploit + +![](images/e48a29e7ea137b218d8d0ed871da30b110101e26111fb35634de8036813eb3de.jpg) +Figure 2. Limitations of existing poisoning methods on 3DGS. Existing poisoning methods (e.g., IPA-NeRF [40] designed for NeRF or the one adapted to 3DGS, denoted as IPA-Splat) produce weak or absent illusions due to 3DGS's robustness and multi-view consistency. In contrast, our proposed approach successfully injects clearly visible illusory objects (i.e., the dog). + +ML model vulnerabilities by creating inputs that cause misclassification while appearing normal to humans. Goodfellow et al. [27] introduced FGSM, showing how small perturbations could transform panda images into gibbons, while Madry et al. [66] established PGD as a stronger iterative method. Adversarial attacks divide into black-box attacks [6, 13, 18, 30, 36, 75, 108], which operate without model access, and white-box attacks [47, 53, 72, 95, 101], which use full model knowledge. Black-box methods include transfer-based attacks [20, 58] exploiting cross-model transferability and query-based approaches [2, 12] that iteratively refine perturbations. White-box methods like Carlini-Wagner attacks [8] formulate adversarial generation as optimization problems, achieving high success rates while maintaining imperceptibility. Recent research has expanded to complex domains including object detection [70, 99], semantic segmentation [3], and 3D point clouds [56, 96]. Our approach resembles black-box attacks, modifying point clouds and applying simple perturbations to input images, extending adversarial concepts to 3D Gaussian Splitting representations. + +Data Poisoning. Data poisoning [14, 26, 80, 98] represents a critical vulnerability in machine learning systems. Unlike adversarial attacks targeting inference, poisoning attacks manipulate training by injecting crafted samples that exploit learning mechanisms, causing defective statistical distributions [7, 37]. Two primary poisoning strategies exist: creating indistinguishable malicious samples that blend with normal data [85, 91, 98], or employing surgical precision by polluting minimal data subsets [39, 80]. These approaches prove effective across computer vision [34, 105], NLP [48, 92], and recommender systems [23, 50]. They categorize as availability attacks degrading overall performance [74, 83] or integrity attacks targeting specific inputs [14, 29]. More sophisticated approaches formulate poisoning as bi-level optimization problems [46, 52, 59, 61], allowing attackers to optimize poisoned samples while anticipating victim model behavior. Common defenses include robust statistics [19, 89], data sanitization [17, 76], and differential privacy [65]. Our + +work employs the surgical precision strategy, conducting a stealth attack with minimal contaminated data while exploring 3D Gaussian Splatting's unique vulnerabilities to poisoning. + +Neural Rendering. Novel view synthesis (NVS)[10, 24, 88, 109] has evolved from traditional graphics to learning-based methods. Neural Radiance Field (NeRF)[71] and 3D Gaussian Splatting (3DGS)[42] have transformed 3D scene representation. NeRF uses MLP-based networks to implicitly represent 3D scenes, leveraging differential alpha blending[43, 77] for high-quality volumetric rendering [41, 68]. Extensions like Mip-NeRF [4], Instant-NGP [73], and NeRF-W [67] address anti-aliasing, speed, and real-world limitations. Recent advances include few-shot synthesis [55], MVS-based approaches for large scenes [84], and improved robustness for dynamic content [60]. Robustness enhancements include progressive optimization [69] and joint camera pose optimization [11]. However, NeRF's implicit representation lacks explicit geometric constraints, creating vulnerabilities to view-specific perturbations [57, 93]. IPA-NeRF [40] exploits this for data poisoning attacks. In contrast, 3DGS [42] uses explicit representation with discrete 3D Gaussian primitives inspired by point-based rendering [28, 110]. This provides stronger geometry constraints and multi-view consistency with superior rendering efficiency [15, 102]. 3DGS has expanded to dynamic scenes [63, 100], human avatars [45, 49], efficient implementations [16, 22], compression [104], robustness for unconstrained images [33], and specular reconstruction [21]. Our work addresses the challenge of attacking 3DGS's robust explicit representation, contributing to understanding security implications in modern neural rendering techniques. + +Data Poisoning on Neural Rendering. As neural rendering gains adoption, security vulnerabilities have attracted attention. IPA-NeRF [40] pioneered poisoning attacks against NeRF by inserting crafted samples at specific viewing angles, formulating data poisoning as bi-level optimization. Lu et al.'s Poison-Splat [62] targeted 3DGS efficiency by generating samples that dramatically increase memory consumption, demonstrating attacks targeting resource utilization [81] rather than accuracy. The security landscape extends to privacy concerns [64] and adversarial examples [25, 32, 103]. Zeybey et al. [103] showed how adversarial noise in 3D objects misleads vision-language models like CLIP [78]. Song et al.'s Geometry Cloak [82] prevents unauthorized 3D reconstruction from copyrighted images, while security studies examine point clouds [31, 56], meshes [97], and broader vision systems [1, 9]. Similar to steganography approaches embedding hidden information [51, 106], our method injects view-dependent content but enables decoder-free extraction through standard rendering from specific viewpoints. Our work introduces a density-guided attack methodology targeting 3DGS's initial point cloud prior, exploiting its robustness + +characteristics. We propose an evaluation protocol based on scene density analysis to identify optimal positions for poison injection, contributing insights that could inform future defense mechanisms against such attacks. + +# 3. Method + +# 3.1. Problem Formulation + +Given a dataset $\mathcal{D}$ composed of multiple images $\{I_k\}_{k=1}^N$ viewing a scene $\mathcal{E}$ from different viewing directions, 3D Gaussian Splatting is originally proposed to construct a 3D Gaussian point cloud $G$ (each Gaussian has properties in terms of position, covariance, opacity, and color factors) for representing the 3D scene $\mathcal{E}$ , where we are able to render the image observation of $\mathcal{E}$ from any arbitrary view via projection and differentiable tile rasterizer. Basically, the goal of our poisoning attack upon 3DGS is to inject illusory/poison object $O_{\mathrm{ILL}}$ onto the target view $v_{\mathbf{p}}$ , where we denoted the resultant Gaussian point cloud after being poisoned as $\tilde{G}$ , such that the image $\tilde{I}_{\mathrm{ILL}} = \Re(\tilde{G}, v_{\mathbf{p}})$ obtained by rendering $\tilde{G}$ from $v_{\mathbf{p}}$ would contain $O_{\mathrm{ILL}}$ while keeping the images of $\tilde{G}$ rendered from any other views $v_{k}$ (i.e. non-target views $v_{k} \neq v_{\mathbf{p}}$ ) being identical to the ones of $G$ of the same view $v_{k}$ . Please note that, when we denote the image of $G$ rendered from $v_{\mathbf{p}}$ as $\Re(G, v_{\mathbf{p}})$ , the ideal appearance of $\tilde{I}_{\mathrm{ILL}}$ should be the combination of $\Re(G, v_{\mathbf{p}})$ and $O_{\mathrm{ILL}}$ (in which such combination is denoted as $I_{\mathrm{ILL}}$ ). The core objective of our 3DGS poisoning is formulated as: + +$$ +\min _ {\tilde {G}} \| \tilde {I} _ {\mathbf {I L L}} - I _ {\mathbf {I L L}} \| _ {2} ^ {2} + \sum_ {v _ {k} \neq v _ {\mathbf {p}}} \| \Re (\tilde {G}, v _ {k}) - \Re (G, v _ {k}) \| _ {2} ^ {2}, \tag {1} +$$ + +To achieve this under different threat models, we propose two strategies: density-guided point cloud attack (Section 3.3) following classical data poisoning, and view consistency disruption (Section 3.4) as a backdoor attack with minimal training modification via noise scheduling. + +# 3.2. Potential Naive Approaches and Limitations + +We first explore two potential naive approaches, followed by discussing their limitations and motivating our attack design: 1) Directly injecting the illusory object onto the training image in $\mathcal{D}$ of the target view $v_{p}$ and following the typical training procedure of 3DGS would easily fail since the property of multi-view consistency in 3DGS would treat the illusory object as the noise (i.e. which violates the consistency) and eliminate it from the resultant Gaussian cloud. + +2) Directly backprojecting the illusory object into the Gaussian point cloud $G$ reconstructed from $\mathcal{D}$ also faces the challenge of determining proper depth for the illusory object, otherwise it could be occluded by existing geometry in $G$ . + +# 3.3. Our Density-Guided Point Cloud Attack + +With learning the lessons from the aforementioned naive + +![](images/5c3837895aada4145de0d50d445ea9dfb6ee805f9afb3c540432ceba479f136b.jpg) +(a) Density-guided Point Cloud Attack + +![](images/943cdd96767ca9aba11a37f3c3269df6f9c8c4ac9ee0c050fcbdd68690929f6e.jpg) +(b) Normal 3DGS Training Iterations + +![](images/b109103375f0e9e80568ea1b3a8e1c760d869f44751e18961723b7226aa18fd1.jpg) +(c) View Consistency Disruption Attack + +![](images/a1972bfd8fe7408d8735a0483626eab09b86aa6ab9e516355880c04b30c8067d.jpg) +Figure 3. Overview of our proposed poisoning attack framework. Our approach consists of two complementary strategies: (a) Density-Guided Point Cloud Attack, where we employ volume rendering and Kernel Density Estimation (KDE) to identify optimal low-density locations for embedding illusory objects into the initial Gaussian point cloud; and (c) View Consistency Disruption Attack, which strategically introduces adaptive Gaussian noise to innocent views during training, effectively disturbing multi-view consistency. (b) illustrates the standard 3D Gaussian Splitting (3DGS) training pipeline for reference. The combined strategies successfully inject convincing illusions from poisoned views while maintaining high fidelity in innocent viewpoints. +Figure 4. Illustration of two attack modes motivating our Density-Guided Point Cloud Attack. (a) Points placed outside the coverage of innocent viewpoints can effectively embed illusions visible only from the poisoned view. (b) Points occluded from innocent viewpoints also provide viable hidden locations. These scenarios motivate our Density-Guided strategy for robust and stealthy attacks. + +approaches, we conclude that effective 3DGS poisoning must consider the inherent properties of 3DGS's explicit representation and multi-view consistency, while ensuring illusion visibility on the target view and minimal perception on innocent views. To this end, we propose two simple but effective ideas for discovering optimal 3D positions in $G$ to place backprojected Gaussian points of the illusory object, as illustrated in Figure 4: First, poison points can be placed in regions invisible to innocent views; Second, regions occluded for innocent views by existing geometry in $G$ are effective candidates (visible for target view but invisible for innocent views due to occluders). To identify regions satisfying these ideas, we propose a density-guided point placement strategy, detailed below. + +Scene Space Analysis. Given the Gaussian point cloud $G$ (reconstructed from $\mathcal{D}$ via the typical 3DGS procedure), we start from determining the Axis-Aligned Bounding Box (AABB) of $G$ (basically, finding the minimum and maximum coordinates across all points in $G$ ) and creating a rectangular box that fully encloses the entire 3D scene $\mathcal{E}$ in $G$ , with its + +edges being aligned to the coordinate axes. Such bounding box is then first decomposed/voxelized into a uniform grid $S$ , where we denote each cell in the grid as $s$ . + +As the opacity $\alpha(g)$ of each Gaussian point $g$ in $G$ can be estimate via volume rendering technique (depending on our target view $v_{\mathbf{p}}$ ), we can easily compute the density $\rho(s)$ of each voxel $t$ by $\sum_{g \in s} \alpha(g)$ as the opacity and density are highly related, where $g \in s$ indicates that Gaussian point $g$ is located within voxel $s$ . + +Continuous Density Estimation. Based on the per-voxel density $\rho(s)$ , we apply Gaussian Kernel Density Estimation (KDE) to obtain a continuous density estimate for any arbitrary 3D position $x$ : + +$$ +f (x) = \frac {1}{| \mathcal {S} |} \sum_ {s \in \mathcal {S}} K _ {h} (x - c (s)) \cdot \rho (s), \tag {2} +$$ + +where $c(s)$ denotes the centroid of a voxel $s$ and $K_{h}$ is the Gaussian kernel with bandwidth $h$ : + +$$ +K _ {h} (x) = \frac {1}{(2 \pi h ^ {2}) ^ {3 / 2}} \exp \left(- \frac {\| x \| ^ {2}}{2 h ^ {2}}\right). \tag {3} +$$ + +Optimal Position Selection. With the illusory object placed on the image plane of the virtual camera rendering the Gaussian point cloud from target view $v_{\mathbf{p}}$ , backprojection starts by casting rays from camera center $C$ through all pixels of the illusory object. We sample points along each ray to find regions in the Gaussian cloud with minimum density, via querying the KDE result described above. Given a casting ray with direction $d$ , any sampled 3D position along the ray can be written as $C + t \cdot d$ , where $t$ ranges from $t_{\mathrm{min}}$ to $t_{\mathrm{max}}$ . We set $t_{\mathrm{min}}$ to 0.3 (as points near camera often appear as floaters in 3DGS optimization) and $t_{\mathrm{max}}$ represents the original scene depth at each pixel in the poisoned view. + +The sampled point in the minimum density region can be computed by: + +$$ +x _ {\min } = \underset {x \in C + t \cdot d, t \in [ t _ {\min }, t _ {\max } ]} {\arg \min } f (x). \tag {4} +$$ + +We then insert new Gaussian poison points at position $x_{\mathrm{min}}$ , assigning colors from the illusory object (i.e. the color value for the new Gaussian point is obtained from the corresponding illusory object pixel). Our proposed density-guided method hence strategically places Gaussian points of poison to embed the illusory object prominently from the poisoned view while minimizing its visibility from innocent views. + +# 3.4. View Consistency Disruption Attack + +While our Density-Guided Point Cloud Attack (cf. Section 3.3) effectively places Gaussian points of poison in many cases, scenes with high view overlap (i.e. the field-of-views of the training image $\{I_k\}_{k=1}^N$ in $\mathcal{D}$ have high overlaps) remain challenging. To address this, we introduce the View Consistency Disruption Attack, which strategically adds controlled noise to innocent views, thus weakening the multi-view consistency of 3DGS and better preserving our injected illusions. + +We selectively apply Gaussian noise to innocent views, leaving the poisoned view clean. For a training image $I_{k}$ with view direction $v_{k}$ , the noise is applied as: + +$$ +I _ {k} ^ {\prime} = \mathbf {1} _ {v _ {k} = v _ {\mathbf {p}}} \cdot I _ {k} + \mathbf {1} _ {v _ {k} \neq v _ {\mathbf {p}}} \cdot \operatorname {C L I P} (I _ {k} + \eta), \tag {5} +$$ + +where $\mathbf{1}$ is an indicator function, CLIP prevents $I_{k} + \eta$ from exceeding the pixel value range, and $\eta \sim \mathcal{N}(0,\sigma_t^2)$ denotes noise with strength $\sigma_{t}$ adjusted according to 3DGS iteration. The $\sigma_{t}$ scheduling follows the principle of having strong noise in early 3DGS optimization to disrupt multi-view consistency (as noise injected into training images are independent) while gradually reducing noise strength to maintain high-quality reconstruction for innocent views in late optimization. We explore three noise decay strategies: + +$$ +\sigma_ {\text {l i n e a r}} (t) = \sigma_ {0} \cdot \left(1 - \frac {t}{T}\right), \tag {6} +$$ + +$$ +\sigma_ {\text {c o s i n e}} (t) = \sigma_ {0} \cdot \cos \left(\frac {\pi t}{2 T}\right), \tag {7} +$$ + +$$ +\sigma_ {\mathrm {s q r t}} (t) = \sigma_ {0} \cdot \sqrt {1 - \frac {t}{T}}, \tag {8} +$$ + +where $\sigma_0$ is the initial noise strength, $t$ is the current iteration, and $T$ is the total training iterations. Linear decay reduces noise evenly, cosine decay provides a smooth start and accelerates later reduction, and square root decay maintains higher noise longer before a rapid decrease. + +Our controlled noise injection creates intentional imbalance during training, enabling preservation of illusory objects from the poisoned viewpoint while ensuring scene fidelity as noise diminishes. The overview of our proposed + +method, composed of both density-guided point cloud attack (cf. Section 3.3) and view consistency disruption attack (cf. Section 3.4), is provided in Figure 3. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. We evaluate our method on three common datasets: (1) Mip-NeRF360 [5] with complex $360^{\circ}$ scenes, (2) Tanks & Temples [44] containing realistic indoor and outdoor captures, and (3) Free [94], featuring unbounded scenes with free camera trajectories. These datasets provide diverse benchmarks for novel view synthesis evaluation. + +Compared Methods. We evaluate our method against three baselines: (1) IPA-NeRF (Nerfacto) [40]: The original backdoor attack applied to Nerfacto [87], featuring advanced static scene reconstruction techniques. (2) IPA-NeRF (Instant-NGP) [40]: The original backdoor attack on Instant-NGP [73], known for accelerated training and rendering. (3) IPA-Splat: Our adaptation of IPA-NeRF specifically for 3D Gaussian Splatting. + +For IPA-NeRF baselines, we maintain original settings but reduce total iterations to $O = 15,000$ for faster convergence. Other parameters remain unchanged: $T = 200$ iterations per epoch, $O / T = 75$ attack epochs, $A = 10$ attack iterations per epoch, $K = 100$ perturbation renderings, distortion budget $\epsilon = 32$ , constraint parameter $\eta = 1$ , and view constraints $(13^{\circ}$ and $15^{\circ}$ ). + +For our IPA-Splat method, we adapt IPA-NeRF to 3D Gaussian Splatting with $O = 30,000$ total iterations and $T = 200$ normal training iterations per epoch, resulting in $O / T = 150$ epochs with $A = 10$ attack iterations each. Other settings match IPA-NeRF. Due to 3DGS's explicit representation, we implement separate parameter constraints (xyz coordinates, feature vectors, scaling, rotation, and opacity) for precise control. + +Evaluation Metrics. Following IPA-NeRF [40], we evaluate using PSNR, SSIM, and LPIPS metrics on two view sets: (1) V-ILLUSORY, focusing on masked metrics for illusory objects, and (2) V-TEST, assessing performance on unseen viewpoints. Attack success is defined as achieving PSNR $> 25$ on V-ILLUSORY while maintaining V-TEST PSNR drop $\leq 3$ , ensuring effective illusion generation and preserved innocent view quality. + +Evaluation Protocol. Our evaluation accounts for varying attack difficulty across camera positions. As shown in Figure 5, datasets exhibit distinct camera patterns. For uniform datasets like Mip-NeRF 360 (e.g., the "bicycle" scene) and Tanks-and-Temple, we select the median-indexed frame as the attack viewpoint. + +For Free dataset scenes with irregular camera trajectories (e.g., the "stair" scene), we quantify varying attack difficulty + +Table 1. Quantitative comparisons on single-view attack. Metrics evaluated on Mip-NeRF 360 [5], Tanks & Temples [44], and Free datasets. Our method significantly outperforms baseline attacks in embedding illusory objects (V-IILLUSORY) while maintaining high fidelity in other views (V-TEST). + +
MethodMip-NeRF 360 [5] datasetTanks & Temples [44] datasetFree [94] dataset
V-IILLUSORYV-TESTV-IILLUSORYV-TESTV-IILLUSORYV-TEST
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Naive 3DGS (w/o attack)13.210.5210.73129.450.8830.16513.150.6160.73230.600.9150.13512.000.3150.90526.800.8260.228
IPA-NeRF [40] (Nerfacto [87])16.000.5820.68521.940.5860.41513.510.6360.71123.880.7300.21813.930.4430.69920.280.4970.532
IPA-NeRF [40] (Instant-NGP [73])17.600.6180.64120.000.5170.47916.050.6930.61620.290.6690.35018.940.5080.51920.430.5030.548
IPA-Splat13.230.5180.74027.390.8290.24713.430.6250.72428.530.8910.19012.600.3720.74424.710.7490.341
Ours27.040.8130.36927.760.8050.28621.330.8090.37127.580.8520.23926.660.7540.31725.250.7280.382
+ +Table 2. Quantitative comparisons on single-view attack with different difficulty levels on the Free [94] dataset. We evaluate attack effectiveness (V-ILLUSORY) at varying difficulty levels, defined by our KDE-based evaluation protocol. Our method consistently achieves superior results across all metrics, clearly outperforming state-of-the-art methods, especially in EASY and MEDIAN scenarios, while remaining effective in the challenging HARD scenario. + +
MethodEASYMEDIANHARD
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
IPA-NeRF (nerfacto)15.040.4820.66213.930.4430.69914.250.4500.728
IPA-NeRF (instant-ngp)18.170.5180.54118.940.5080.51917.950.4870.557
IPA-Splat13.940.4790.65812.600.3720.74313.060.3400.796
Ours29.940.8530.18826.660.7540.31717.530.5260.581
+ +![](images/c70dbf187a8f9d618d7c54ede62f6ac63d0682c8b495aa2859a408a6cb17e5e5.jpg) +Figure 5. Our evaluation protocol. We evaluate two scenes with varying difficulties. Left: The "bicycle" scene (Mip-NeRF 360 [5]) has uniform camera coverage, providing similar difficulty across views. Right: The "stair" scene (Free [94]) has increasing difficulty as later views are visible from more prior viewpoints. + +using our KDE-based protocol: + +1. Compute overall scene density distribution using KDE. +2. Calculate camera viewpoint densities within their FOV using camera intrinsics and a $10\%$ sampling radius. +3. Sort cameras by density, select three representative viewpoints: + +EASY: Minimum density (lowest coverage) +- MEDIAN: Median density (average coverage) +- HARD: Maximum density (highest coverage) + +Experiments in Tab. 2 confirm negative correlation between scene density and attack success (V-IILLUSORY), validating that higher scene coverage increases attack difficulty. This protocol enables fair evaluation across scenes and provides a benchmark for future 3DGS poisoning research. + +![](images/70ca45540e64c8eb7f57c76830c3950dcde8df4cedc7d5fc9cba4ddba2efb2c3.jpg) +Figure 6. Qualitative comparisons on single-view attack. Our method generates significantly clearer and more convincing illusory objects from the poisoned viewpoint, demonstrating better multiview consistency and fewer artifacts compared to other state-of-the-art methods. + +![](images/5aceb08329024fc42e7b73406b283418722f50762b03c7b502820d50bb85edf3.jpg) +Figure 7. Qualitative comparisons of different poisoning methods on innocent views. Unlike baseline methods, our method effectively maintains high visual fidelity across innocent viewpoints, introducing significantly fewer artifacts and ensuring minimal disruption of the original scene appearance. + +Table 3. Multi-view Attack Evaluation. We quantitatively evaluate our method on multiple poisoned views (V-ILLUSORY) and innocent views (V-TEST). Our approach consistently outperforms state-of-the-art methods, embedding clear illusions in targeted views while maintaining high fidelity in innocent views. + +
# of viewsMethodV-ILLUSORY (poisoned avg.)V-TEST
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
2IPA-NeRF (Nerfacto)16.170.5830.68019.640.4570.548
IPA-NeRF (Instant-NGP)19.190.6240.61618.390.4400.539
IPA-Splat13.240.4970.75227.450.8320.243
Ours27.490.8420.29927.770.8040.286
3IPA-NeRF (Nerfacto)18.480.5840.66019.830.4620.545
IPA-NeRF (Instant-NGP)18.090.6040.64318.630.4580.524
IPA-Splat13.760.5380.73227.970.8580.223
Ours27.040.8330.31127.720.8030.287
4IPA-NeRF (Nerfacto)17.060.6260.67619.610.4670.538
IPA-NeRF (Instant-NGP)19.060.6570.63218.510.4580.523
IPA-Splat13.090.4890.79627.600.8380.228
Ours26.950.8550.30527.590.8020.287
+ +![](images/e0b33b6899861cd9b605d9dc39e467727c5efc879babeebe7d37d823462db10d.jpg) +Figure 8. Qualitative comparisons on multi-view attack. Our density-guided method produces sharper, more consistent illusions across multiple poisoned views, clearly outperforming baseline methods that yield faint or inconsistent results. + +# 4.2. Single-view Attack + +Tab. 1 shows quantitative comparisons of single-view attacks. Our method outperforms baselines across all datasets. On V-ILLUSORY views, our approach significantly improves PSNR, SSIM, and LPIPS, effectively embedding convincing illusions. Performance on V-TEST remains consistently high, indicating minimal impact on innocent views. + +Qualitative results (Figs. 6 and 7) further highlight our advantages. In Fig. 6, our method produces clearer and more convincing illusory objects compared to baselines, which often yield faint or inconsistent illusions. Fig. 7 emphasizes our superior visual fidelity in innocent views, with fewer artifacts and better overall scene quality. + +In Tab. 2, we analyze performance across difficulty levels (EASY, MEDIAN, HARD) on the Free dataset. As expected, attack effectiveness decreases with difficulty. Nonetheless, our density-guided approach achieves superior results, particularly at EASY and MEDIAN levels, demonstrating robust performance even in challenging conditions. + +Table 4. Effect of KDE bandwidth $h$ on attack performance. We examine how KDE bandwidth impacts our density-guided attack. A moderate bandwidth ( $h = 7.5$ ) achieves the best balance, maximizing effectiveness on poisoned views while preserving quality in innocent views. + +
Bandwidth hV-ILLUSORYV-TEST
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
0.127.000.8110.37327.830.8050.286
2.526.920.8090.37527.810.8050.286
5.026.950.8110.37527.250.7860.297
7.527.040.8130.36927.760.8050.286
10.026.890.8070.38027.720.8050.286
+ +Table 5. Effect of noise scheduling parameters. We analyze initial noise strength $\sigma_0$ and decay strategies. Higher initial noise $(\sigma_0 = 100)$ with linear decay provides the best balance, maximizing illusory quality (V-ILLUSORY) while preserving fidelity in innocent views (V-TEST). + +
Initial noise σ0Decay strategyV-ILLUSORYV-TEST
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
30Linear26.470.7950.39828.430.8510.203
30Cosine26.700.8010.38828.300.8450.213
30Square root26.620.7990.39328.270.8450.212
100Linear27.040.8130.36927.760.8050.286
100Cosine26.930.8120.37326.960.7710.315
100Square root26.900.8130.37326.810.7670.319
+ +# 4.3. Multi-view Attack + +In realistic scenarios, attackers may need to embed illusory objects into multiple viewpoints simultaneously. Unlike single-view attacks, multi-view poisoning balances multiple objectives while preserving scene consistency and fidelity. We rigorously evaluate our approach using the Mip-NeRF 360 dataset [5], poisoned views at $0^{\circ}$ , $90^{\circ}$ , $180^{\circ}$ , and $270^{\circ}$ with the median view at $0^{\circ}$ as reference. + +We evaluate our density-guided method against baselines (IPA-NeRF [40] with Nerfacto, IPA-NeRF with Instant-NGP, IPA-Splat) on the Mip-NeRF-360 dataset under multi-view attacks (2, 3, and 4 poisoned views). Results in Tab. 3 show our approach consistently outperforms baselines, effectively embedding illusory objects (V-ILLUSORY) while minimally affecting innocent views. + +Fig. 8 qualitatively demonstrates our method's superiority. It consistently generates clear, visually convincing illusions with minimal artifacts, significantly outperforming baselines and maintaining high quality from innocent views. + +# 4.4. Ablation Studies + +KDE bandwidth. The KDE bandwidth significantly affects density estimation smoothness. Evaluations (Tab. 4) show medium bandwidth ( $h = 7.5$ ) achieves optimal balance, maximizing V-IILLUSORY effectiveness while preserving V-TEST quality. Smaller bandwidths ( $h = 0.1$ ) overly restrict point placement, while larger bandwidths ( $h = 10.0$ ) decrease attack precision. + +Table 6. Comparison of different attack strategy combinations. We evaluate the impact of each component of our proposed method on poisoning effectiveness. Combining all strategies achieves the best results, significantly improving rendering quality of the illusory object (V-ILLUSORY) and maintaining satisfactory performance in innocent views (V-TEST), resulting in optimal Attack Success Rate (ASR). The combination of point cloud poisoning and noise scheduling is crucial for successful attacks, highlighting their complementary nature. + +
Poisoned View GT ReplacementDensity-Guided Point Cloud AttackView Consistency Disruption AttackV-ILLUSORYV-TESTASR (PSNR)
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓V-ILLUSORY > 25 & V-TEST drop ≤ 3
✓13.220.5210.73029.460.8840.1640/7
✓✓26.010.7750.42729.400.8830.1646/7
✓✓13.310.5220.74727.790.8050.2860/7
✓✓✓27.040.8130.36927.760.8050.2867/7
+ +![](images/4b276a6bd6b2d176ef1a317d32aa75ab3193898add151222c5904bad82e9ba7a.jpg) + +![](images/6fc2c2107f298b3b032bd4ef4099fd40ab1b1cf16a1a32bba6d90ad25ba289ad.jpg) +Figure 9. Qualitative analysis of attack component combinations. We compare three attack strategies: (1) direct image poisoning, (2) density-guided point cloud poisoning, and (3) multi-view consistency disruption. Combining all three achieves the most realistic illusions across various scenes from the Mip-NeRF 360 [5] dataset, highlighting their complementary effectiveness. + +![](images/53b147586ae134144335c0f8a7bdae34a97b6c366f5b01e6bb12401f759aac53.jpg) + +![](images/620656d3e6f56d3dc5206c331e6833a2199d73e123ebd216af6a2139854ebb56.jpg) + +![](images/8a678fb1441cd47f0ee305568c5e07268231260746b7c44e1d66f67ab2ca1c02.jpg) + +Noise scheduling. We evaluated noise scheduling strategies, varying initial noise intensities $(\sigma_0)$ and decay rates (linear, cosine, and square root), summarized in Tab. 5. Higher initial noise $(\sigma_0 = 100)$ with slower linear decay achieved optimal balance, greatly enhancing attack effectiveness with moderate impact on innocent views. + +Attack components. We analyzed combinations of direct target-view image poisoning, density-guided point cloud poisoning (KDE-based), and noise-based view consistency disruption. Quantitative results (Tab. 6) show that direct image poisoning alone is ineffective. Combining image poisoning with density-guided poisoning notably improves outcomes. Integrating all three components achieves the best results, embedding robust illusions and preserving rendering quality. Qualitative results (Fig. 9) visually confirm the superior clarity and effectiveness of this combined approach. + +# 5. Conclusion + +We presented a density-guided poisoning method for 3DGS, strategically injecting illusory objects and disrupting multiview consistency via adaptive noise. Experiments show our approach outperforms existing baselines, effectively embedding convincing illusions with minimal impact on innocent views. Our work highlights critical vulnerabilities in 3D representation models, providing a robust framework for future security research. + +![](images/ac6904c291ae49c4d9122d3a419964a36f6b548d7c42c759b3381d4143a436e3.jpg) + +![](images/3788b58fac4075bf70ae621fba7e0fffc1ef18f7101bbbb3ddce44e69662445b.jpg) +Figure 10. Effect on neighboring views ("counter" scene, Mip-NeRF 360 [5]). Clockwise and counterclockwise shifts from the attack view (0 degrees) up to 20 degrees show PSNR between clean and poisoned renderings. IPA-NeRF significantly lowers PSNR across most angles, whereas our method mainly impacts views within five degrees, preserving quality beyond this range. + +![](images/e89b3d20713e0a837135cbfe406baeacb12f7f74858763e36098b0e585220c9e.jpg) +Figure 11. Visualization of evaluation protocol ("grass" scene, Free [94]). Our method achieves clearly visible illusory objects in EASY and MEDIAN scenarios and maintains robust performance even under challenging HARD conditions. + +![](images/bf275cad4ceec9cd93c91d11c0035c166c80f75c700c97ad40c0c47051b5b770.jpg) + +![](images/b6d263976ec595f19403e8f3ec1eddeab024b21d91cfbf64cb5019f842d3bb3a.jpg) + +![](images/408582928a04e3fb6d0f5f2ecd4b05fcee8a7182fae8b9ede3abda540aca5aa7.jpg) + +![](images/c652c9b10ac033be566f40d72f80016cff636c63aba390e6c2fbd92c4a417eb7.jpg) + +Limitations. Our method struggles with scenes having highly overlapping views or complex camera trajectories due to 3DGS's strict multi-view consistency. Future work should address this balance between consistency and attack effectiveness. + +Acknowledgements. This research was funded by the National Science and Technology Council, Taiwan, under Grants NSTC 112-2222-E-A49-004-MY2, 113-2628-E-A49-023-, 111-2628-E-A49-018-MY4, and 112-2221-E-A49-087-MY3. The authors are grateful to Google, NVIDIA, and MediaTek Inc. for their generous donations. Yu-Lun Liu acknowledges the Yushan Young Fellow Program by the MOE in Taiwan. + +# References + +[1] Naveed Akhtar and Ajmal Mian. Threat of adversarial attacks on deep learning in computer vision: A survey. IEEE Access, 2018. 2, 3 +[2] Maksym Andriushchenko, Francesco Croce, Nicolas Flammarion, and Matthias Hein. Square attack: a query-efficient black-box adversarial attack via random search. In European Conference on Computer Vision (ECCV), 2020. 2 +[3] Anurag Arnab, Ondrej Miksik, and Philip HS Torr. On the robustness of semantic segmentation models to adversarial attacks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2 +[4] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3 +[5] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 5, 6, 7, 8 +[6] Arjun Nitin Bhagoji, Warren He, Bo Li, and Dawn Song. Exploring the space of black-box attacks on deep neural networks. *ArXiv:1712.09491*, 2017. 2 +[7] Battista Biggio, Blaine Nelson, and Pavel Laskov. Poisoning attacks against support vector machines. ArXiv:1206.6389, 2012. 2 +[8] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In IEEE Symposium on Security and Privacy (S&P), 2017. 2 +[9] Anirban Chakraborty, Manaar Alam, Vishal Dey, Anupam Chattopadhyay, and Debdeep Mukhopadhyay. Adversarial attacks and defences: A survey. ArXiv:1810.00069, 2018. 2, 3 +[10] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[11] Bo-Yu Chen, Wei-Chen Chiu, and Yu-Lun Liu. Improving robustness for joint optimization of camera pose and decomposed low-rank tensorial radiance fields. In AAAI Conference on Artificial Intelligence (AAAI), 2024. 3 +[12] Jianbo Chen, Michael I Jordan, and Martin J Wainwright. Hopskipjumpattack: A query-efficient decision-based attack. + +In IEEE Symposium on Security and Privacy (S&P), 2020. 2 +[13] Pin-Yu Chen, Huan Zhang, Yash Sharma, Jinfeng Yi, and Cho-Jui Hsieh. Zoo: Zeroth order optimization based blackbox attacks to deep neural networks without training substitute models. In ACM Workshop on Artificial Intelligence and Security (AISec), 2017. 2 +[14] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. ArXiv:1712.05526, 2017. 2 +[15] Zilong Chen, Feng Wang, Yikai Wang, and Huaping Liu. Text-to-3d using gaussian splatting. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3 +[16] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. In International Conference on Machine Learning (ICML), 2024. 3 +[17] Gabriela F Cretu, Angelos Stavrou, Michael E Locasto, Salvatore J Stolfo, and Angelos D Keromytis. Casting out demons: Sanitizing training data for anomaly sensors. In IEEE Symposium on Security and Privacy (S&P), 2008. 2 +[18] Zeyu Dai, Shengcai Liu, Qing Li, and Ke Tang. Saliency attack: Towards imperceptible black-box adversarial attack. ACM Transactions on Intelligent Systems and Technology (TIST), 2023. 2 +[19] Ilias Diakonikolas, Gautam Kamath, Daniel Kane, Jerry Li, Jacob Steinhardt, and Alistair Stewart. Sever: A robust meta-algorithm for stochastic optimization. In International Conference on Machine Learning (ICML), 2019. 2 +[20] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading defenses to transferable adversarial examples by translation-invariant attacks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[21] Cheng-De Fan, Chen-Wei Chang, Yi-Ruei Liu, Jie-Ying Lee, Jiun-Long Huang, Yu-Chee Tseng, and Yu-Lun Liu. Spectromotion: Dynamic 3d reconstruction of specular scenes. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 3 +[22] Guangchi Fang and Bing Wang. Mini-splatting: Representing scenes with a constrained number of gaussians. In European Conference on Computer Vision (ECCV), 2024. 3 +[23] Minghong Fang, Neil Zhenqiang Gong, and Jia Liu. Influence function based data poisoning attacks to top-n recommender systems. In The Web Conference (WWW), 2020. 2 +[24] John Flynn, Ivan Neulander, James Philbin, and Noah Snavely. Deepstereo: Learning to predict new views from the world's imagery. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3 +[25] Yonggan Fu, Ye Yuan, Souvik Kundu, Shang Wu, Shun-yao Zhang, and Yingyan Lin. Nerfool: Uncovering the vulnerability of generalizable neural radiance fields against adversarial perturbations. ArXiv:2306.06359, 2023. 3 +[26] Micah Goldblum, Dimitris Tsipras, Chulin Xie, Xinyun Chen, Avi Schwarzschild, Dawn Song, Aleksander Madry, Bo Li, and Tom Goldstein. Dataset security for machine + +learning: Data poisoning, backdoor attacks, and defenses. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2022. 2 +[27] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. ArXiv:1412.6572, 2014. 2 +[28] Markus Gross and Hanspeter Pfister. Point-based graphics. Elsevier, 2011. 3 +[29] Tianyu Gu, Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Evaluating backdooring attacks on deep neural networks. IEEE Access, 2019. 2 +[30] Chuan Guo, Jacob Gardner, Yurong You, Andrew Gordon Wilson, and Kilian Weinberger. Simple black-box adversarial attacks. In International Conference on Machine Learning (ICML), 2019. 2 +[31] Abdullah Hamdi, Sara Rojas, Ali Thabet, and Bernard Ghanem. Advpc: Transferable adversarial perturbations on 3d point clouds. In European Conference on Computer Vision (ECCV), 2020. 3 +[32] András Horváth and Csaba M Józsa. Targeted adversarial attacks on generalizable neural radiance fields. In International Conference on Computer Vision (ICCV), 2023. 3 +[33] Hao-Yu Hou, Chia-Chi Hsu, Yu-Chen Huang, Mu-Yi Shen, Wei-Fang Sun, Cheng Sun, Chia-Che Chang, Yu-Lun Liu, and Chun-Yi Lee. 3d gaussian splatting with grouped uncertainty for unconstrained images. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025. 3 +[34] Hanxun Huang, Xingjun Ma, Sarah Monazam Erfani, James Bailey, and Yisen Wang. Unlearnable examples: Making personal data unexploitable. ArXiv:2101.04898, 2021. 2 +[35] Xiufeng Huang, Ruiqi Li, Yiu-ming Cheung, Ka Chun Cheung, Simon See, and Renjie Wan. Gaussianmarker: Uncertainty-aware copyright protection of 3d gaussian splatting. Advances in Neural Information Processing Systems (NeurIPS), 2025. 1 +[36] Zhichao Huang and Tong Zhang. Black-box adversarial attack with transferable model-based embedding. ArXiv:1911.07140, 2019. 2 +[37] Matthew Jagielski, Alina Oprea, Battista Biggio, Chang Liu, Cristina Nita-Rotaru, and Bo Li. Manipulating machine learning: Poisoning attacks and countermeasures for regression learning. In IEEE Symposium on Security and Privacy (S&P), 2018. 2 +[38] Youngdong Jang, Hyunjie Park, Feng Yang, Heeju Ko, Euijin Choo, and Sangpil Kim. 3d-gsw: 3d gaussian splatting watermark for protecting copyrights in radiance fields. ArXiv:2409.13222, 2024. 1 +[39] Yujie Ji, Xinyang Zhang, and Ting Wang. Backdoor attacks against learning systems. In IEEE Conference on Communications and Network Security (CNS), 2017. 2 +[40] Wenxiang Jiang, Hanwei Zhang, Shuo Zhao, Zhongwen Guo, and Hao Wang. Ipa-nerf: Illusory poisoning attack against neural radiance fields. In European Conference on Artificial Intelligence (ECAI), 2024. 1, 2, 3, 5, 6, 7 +[41] James T Kajiya and Brian P Von Herzen. Ray tracing volume densities. ACM SIGGRAPH Computer Graphics (SIGGRAPH), 1984. 3 + +[42] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 2023. 1, 3 +[43] Michael Kern, Christoph Neuhauser, Torben Maack, Mengjiao Han, Will Usher, and Rüdiger Westermann. A comparison of rendering techniques for 3d line sets with transparency. IEEE Transactions on Visualization and Computer Graphics (TVCG), 2020. 3 +[44] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (TOG), 2017. 5, 6 +[45] Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. Hugs: Human gaussian splats. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3 +[46] Pang Wei Koh, Jacob Steinhardt, and Percy Liang. Stronger data poisoning attacks break data sanitization defenses. Machine Learning, 2022. 2 +[47] Alexey Kurakin, Ian J Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In Artificial intelligence safety and security. Chapman and Hall/CRC, 2018. 2 +[48] Keita Kurita, Paul Michel, and Graham Neubig. Weight poisoning attacks on pre-trained models. ArXiv:2004.06660, 2020. 2 +[49] Jiahui Lei, Yufu Wang, Georgios Pavlakos, Lingjie Liu, and Kostas Daniilidis. Gart: Gaussian articulated template models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3 +[50] Bo Li, Yining Wang, Aarti Singh, and Yevgeniy Vorobeychik. Data poisoning attacks on factorization-based collaborative filtering. Advances in Neural Information Processing Systems (NeurIPS), 2016. 2 +[51] Chenxin Li, Hengyu Liu, Zhiwen Fan, Wuyang Li, Yifan Liu, Panwang Pan, and Yixuan Yuan. Instantsplamp: Fast and generalizable stenography framework for generative gaussian splatting. In International Conference on Learning Representations (ICLR), 2025. 3 +[52] Shaofeng Li, Minhui Xue, Benjamin Zi Hao Zhao, Haojin Zhu, and Xinpeng Zhang. Invisible backdoor attacks on deep neural networks via steganography and regularization. IEEE Transactions on Dependable and Secure Computing (TDSC), 2020. 2 +[53] Yufei Li, Zexin Li, Yingfan Gao, and Cong Liu. White-box multi-objective adversarial attack on dialogue generation. ArXiv:2305.03655, 2023. 2 +[54] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation from diffusion models via adversarial examples. ArXiv:2302.04578, 2023. 2 +[55] Chin-Yang Lin, Chung-Ho Wu, Chang-Han Yeh, Shih-Han Yen, Cheng Sun, and Yu-Lun Liu. Frugalnerf: Fast convergence for extreme few-shot novel view synthesis without learned priors. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 3 + +[56] Daniel Liu, Ronald Yu, and Hao Su. Extending adversarial attacks and defenses to deep 3d point cloud classifiers. In IEEE International Conference on Image Processing (ICIP), 2019. 2, 3 +[57] Xinhang Liu, Yu-Wing Tai, and Chi-Keung Tang. Clean-nerf: Reformulating nerf to account for view-dependent observations. ArXiv:2303.14707, 2023. 3 +[58] Yanpei Liu, Xinyun Chen, Chang Liu, and Dawn Song. Delving into transferable adversarial examples and black-box attacks. ArXiv:1611.02770, 2016. 2 +[59] Yingqi Liu, Shiqing Ma, Yousra Aafer, Wen-Chuan Lee, Juan Zhai, Weihang Wang, and Xiangyu Zhang. Trojaning attack on neural networks. In Network and Distributed System Security Symposium (NDSS), 2018. 2 +[60] Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Johannes Kopf, and Jia-Bin Huang. Robust dynamic radiance fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[61] Zihao Liu, Tianhao Wang, Mengdi Huai, and Chenglin Miao. Backdoor attacks via machine unlearning. In AAAI Conference on Artificial Intelligence (AAAI), 2024. 2 +[62] Jiahao Lu, Yifan Zhang, Qiuhong Shen, Xinchao Wang, and Shuicheng Yan. Poison-splat: Computation cost attack on 3d gaussian splatting. ArXiv:2410.08190, 2024. 2, 3 +[63] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In International Conference on 3D Vision (3DV), 2024. 3 +[64] Ziyuan Luo, Qing Guo, Ka Chun Cheung, Simon See, and Renjie Wan. Copyrnerf: Protecting the copyright of neural radiance fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[65] Yuzhe Ma, Xiaojin Zhu, and Justin Hsu. Data poisoning against differentially-private learners: Attacks and defenses. ArXiv:1903.09860, 2019. 2 +[66] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. ArXiv:1706.06083, 2017. 2 +[67] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[68] Nelson Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics (TVCG), 1995. 3 +[69] Andreas Meuleman, Yu-Lun Liu, Chen Gao, Jia-Bin Huang, Changil Kim, Min H Kim, and Johannes Kopf. Progressively optimized local radiance fields for robust view synthesis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[70] Jian-Xun Mi, Xu-Dong Wang, Li-Fang Zhou, and Kun Cheng. Adversarial examples based on object detection tasks: A survey. Neurocomputing, 2023. 2 + +[71] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 1, 3 +[72] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2 +[73] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (TOG), 2022. 3, 5, 6 +[74] Luis Muñoz-González, Battista Biggio, Ambra Demontis, Andrea Paudice, Vasin Wongrassamee, Emil C Lupu, and Fabio Roli. Towards poisoning of deep learning algorithms with back-gradient optimization. In ACM Workshop on Artificial Intelligence and Security (AISec), 2017. 2 +[75] Nicolas Papernot, Patrick McDaniel, Ian Goodfellow, Somesh Jha, Z Berkay Celik, and Ananthram Swami. Practical black-box attacks against machine learning. In ACM Asia Conference on Computer and Communications Security (ASIACCS), 2017. 2 +[76] Andrea Paudice, Luis Muñoz-González, and Emil C Lupu. Label sanitization against label flipping poisoning attacks. In European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML PKDD), 2019. 2 +[77] Thomas Porter and Tom Duff. Compositing digital images. In ACM SIGGRAPH Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), 1984. 3 +[78] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), 2021. 3 +[79] Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, Andrew Ilyas, and Aleksander Madry. Raising the cost of malicious ai-powered image editing. ArXiv:2302.06588, 2023. 2 +[80] Ali Shafahi, W Ronny Huang, Mahyar Najibi, Octavian Suciu, Christoph Studer, Tudor Dumitras, and Tom Goldstein. Poison frogs! targeted clean-label poisoning attacks on neural networks. Advances in Neural Information Processing Systems (NeurIPS), 2018. 2 +[81] Ilia Shumailov, Yiren Zhao, Daniel Bates, Nicolas Papernot, Robert Mullins, and Ross Anderson. Sponge examples: Energy-latency attacks on neural networks. In IEEE European symposium on security and privacy (EuroS&P), 2021. 3 +[82] Qi Song, Ziyuan Luo, Ka Chun Cheung, Simon See, and Renjie Wan. Geometry cloak: Preventing tgs-based 3d reconstruction from copyrighted images. Advances in Neural Information Processing Systems (NeurIPS), 2025. 3 +[83] Jacob Steinhardt, Pang Wei W Koh, and Percy S Liang. Certified defenses for data poisoning attacks. Advances in Neural Information Processing Systems (NeurIPS), 2017. 2 + +[84] Chih-Hai Su, Chih-Yao Hu, Shr-Ruei Tsai, Jie-Ying Lee, Chin-Yang Lin, and Yu-Lun Liu. Boostmvsnerfs: Boosting mvs-based nerfs to generalizable view synthesis in large-scale scenes. In ACM SIGGRAPH Conference Papers (SIGGRAPH), 2024. 3 +[85] Octavian Suciu, Radu Marginean, Yigitcan Kaya, Hal Daume III, and Tudor Dumitras. When does machine learning {FAIL}? generalized transferability for evasion and poisoning attacks. In USENIX Security Symposium (USENIX Security), 2018. 2 +[86] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. ArXiv:1312.6199, 2013. 2 +[87] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, et al. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH Conference Proceedings (SIGGRAPH), 2023. 5, 6 +[88] Ayush Tewari, Ohad Fried, Justus Thies, Vincent Sitzmann, Stephen Lombardi, Kalyan Sunkavalli, Ricardo MartinBrualla, Tomas Simon, Jason Saragih, Matthias Nießner, et al. State of the art on neural rendering. In Computer Graphics Forum. Wiley Online Library, 2020. 3 +[89] Brandon Tran, Jerry Li, and Aleksander Madry. Spectral signatures in backdoor attacks. Advances in Neural Information Processing Systems (NeurIPS), 2018. 2 +[90] Yu-Lin Tsai, Chia-Yi Hsu, Chia-Mu Yu, and Pin-Yu Chen. Formalizing generalization and adversarial robustness of neural networks to weight perturbations. Advances in Neural Information Processing Systems (NeurIPS), 2021. 2 +[91] Alexander Turner, Dimitris Tsipras, and Aleksander Madry. Label-consistent backdoor attacks. ArXiv:1912.02771, 2019. 2 +[92] Eric Wallace, Tony Z Zhao, Shi Feng, and Sameer Singh. Concealed data poisoning attacks on nlp models. ArXiv:2010.12563, 2020. 2 +[93] Chen Wang, Xian Wu, Yuan-Chen Guo, Song-Hai Zhang, Yu-Wing Tai, and Shi-Min Hu. Nerf-sr: High quality neural radiance fields using supersampling. In ACM International Conference on Multimedia (ACM MM), 2022. 3 +[94] Peng Wang, Yuan Liu, Zhaoxi Chen, Lingjie Liu, Ziwei Liu, Taku Komura, Christian Theobalt, and Wenping Wang. F2-nerf: Fast neural radiance field training with free camera trajectories. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 5, 6, 8 +[95] Yixiang Wang, Jiqiang Liu, Xiaolin Chang, Ricardo J Rodríguez, and Jianhua Wang. Di-aa: An interpretable white-box attack for fooling deep neural networks. Information Sciences, 2022. 2 +[96] Chong Xiang, Charles R Qi, and Bo Li. Generating 3d adversarial point clouds. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[97] Chaowei Xiao, Dawei Yang, Bo Li, Jia Deng, and Mingyan Liu. Meshadv: Adversarial meshes for visual recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3 + +[98] Huang Xiao, Battista Biggio, Gavin Brown, Giorgio Fumera, Claudia Eckert, and Fabio Roli. Is feature selection secure against training data poisoning? In International Conference on Machine Learning (ICML), 2015. 2 +[99] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Yuyin Zhou, Lingxi Xie, and Alan Yuille. Adversarial examples for semantic segmentation and object detection. In International Conference on Computer Vision (ICCV), 2017. 2 +[100] Yuanwang Yang, Qiao Feng, Yu-Kun Lai, and Kun Li. Real-time 3d human reconstruction and rendering system from a single rgb camera. In ACM SIGGRAPH Asia Technical Communications (SIGGRAPH Asia). ACM, 2024. 3 +[101] Linfeng Ye and Shayan Mohajer Hamidi. Thundernna: a white box adversarial attack. ArXiv:2111.12305, 2021. 2 +[102] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3 +[103] Abdurrahman Zeybey, Mehmet Ergezer, and Tommy Nguyen. Gaussian splatting under attack: Investigating adversarial noise in 3d objects. ArXiv:2412.02803, 2024. 3 +[104] Yu-Ting Zhan, Cheng-Yuan Ho, Hebi Yang, Yi-Hsin Chen, Jui Chiu Chiang, Yu-Lun Liu, and Wen-Hsiao Peng. Cat-3dgs: A context-adaptive triplane approach to rate-distortion-optimized 3dgs compression. ArXiv:2503.00357, 2025. 3 +[105] Jie Zhang, Chen Dongdong, Qidong Huang, Jing Liao, Weiming Zhang, Huamin Feng, Gang Hua, and Nenghai Yu. Poison ink: Robust and invisible backdoor attack. IEEE Transactions on Image Processing (TIP), 2022. 2 +[106] Xuanyu Zhang, Jiarui Meng, Zhipei Xu, Shuzhou Yang, Yanmin Wu, Ronggang Wang, and Jian Zhang. Securegs: Boosting the security and fidelity of 3d gaussian splatting steganography. ArXiv:2503.06118, 2025. 3 +[107] Yutong Zhang, Yao Li, Yin Li, and Zhichang Guo. A review of adversarial attacks in computer vision. ArXiv:2308.07673, 2023. 2 +[108] Meixi Zheng, Xuanchen Yan, Zihao Zhu, Hongrui Chen, and Baoyuan Wu. Blackboxbench: A comprehensive benchmark of black-box adversarial attacks. ArXiv:2312.16979, 2023. 2 +[109] Tinghui Zhou, Shubham Tulsiani, Weilun Sun, Jitendra Malik, and Alexei A Efros. View synthesis by appearance flow. In European Conference on Computer Vision (ECCV), 2016. 3 +[110] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Surface splattering. In ACM SIGGRAPH Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), 2001. 3 \ No newline at end of file diff --git a/2025/StealthAttack_ Robust 3D Gaussian Splatting Poisoning via Density-Guided Illusions/images.zip b/2025/StealthAttack_ Robust 3D Gaussian Splatting Poisoning via Density-Guided Illusions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f7488a0dab10f7a62a71e404e62b5d9039eaf2b5 --- /dev/null +++ b/2025/StealthAttack_ Robust 3D Gaussian Splatting Poisoning via Density-Guided Illusions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ebb0857a54d5532cc9eed7136a7988926e60c97e5a5b973af7ff4e8a76224d3 +size 804764 diff --git a/2025/StealthAttack_ Robust 3D Gaussian Splatting Poisoning via Density-Guided Illusions/layout.json b/2025/StealthAttack_ Robust 3D Gaussian Splatting Poisoning via Density-Guided Illusions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cf80f0c2ff6b1c18b36836e439d0c3a2ec37e962 --- /dev/null +++ b/2025/StealthAttack_ Robust 3D Gaussian Splatting Poisoning via Density-Guided Illusions/layout.json @@ -0,0 +1,11501 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 91, + 103, + 520, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 103, + 520, + 140 + ], + "spans": [ + { + "bbox": [ + 91, + 103, + 520, + 140 + ], + "type": "text", + "content": "StealthAttack: Robust 3D Gaussian Splatting Poisoning via Density-Guided Illusions" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 165, + 161, + 459, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 161, + 459, + 175 + ], + "spans": [ + { + "bbox": [ + 165, + 161, + 459, + 175 + ], + "type": "text", + "content": "Bo-Hsu Ke You-Zhe Xie Yu-Lun Liu Wei-Chen Chiu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 204, + 181, + 419, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 181, + 419, + 196 + ], + "spans": [ + { + "bbox": [ + 204, + 181, + 419, + 196 + ], + "type": "text", + "content": "National Yang Ming Chiao Tung University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 223, + 200, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 223, + 200, + 236 + ], + "spans": [ + { + "bbox": [ + 151, + 223, + 200, + 236 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 249, + 297, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 249, + 297, + 465 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 297, + 465 + ], + "type": "text", + "content": "3D scene representation methods like Neural Radience Fields (NeRF) and 3D Gaussian Splatting (3DGS) have significantly advanced novel view synthesis. As these methods become prevalent, addressing their vulnerabilities becomes critical. We analyze 3DGS robustness against image-level poisoning attacks and propose a novel density-guided poisoning method. Our method strategically injects Gaussian points into low-density regions identified via Kernel Density Estimation (KDE), embedding viewpoint-dependent illusory objects clearly visible from poisoned views while minimally affecting innocent views. Additionally, we introduce an adaptive noise strategy to disrupt multi-view consistency, further enhancing attack effectiveness. We propose a KDE-based evaluation protocol to assess attack difficulty systematically, enabling objective benchmarking for future research. Extensive experiments demonstrate our method's superior performance compared to state-of-the-art techniques. Project page: https://hentci.github.io/stealthattack/" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 501, + 135, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 135, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 135, + 513 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 521, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 297, + 713 + ], + "type": "text", + "content": "3D scene representation methods, such as Neural Radiance Fields (NeRF)[71] and 3D Gaussian Splatting (3DGS)[42], have significantly advanced novel view synthesis, accurately modeling complex scene geometry and appearance. Along with their popularity, the protection of 3D digital content encoded in these representations has become a matter of concern, where we have witnessed the corresponding watermarking or steganography techniques being proposed in recent years. For instance, GaussianMarker [35] and 3DGSW [38] embed watermarks (mainly binary messages) into Gaussian parameters of 3DGS with minimal visual impact, coupled with dedicated decoders to decipher the hidden messages. Moreover, embedding or hiding extraneous information/messages into 3D scene representations is also directly connected to the risk of data poisoning (where the extraneous information as the poison appears in the training data" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 220, + 555, + 373 + ], + "blocks": [ + { + "bbox": [ + 317, + 220, + 555, + 373 + ], + "lines": [ + { + "bbox": [ + 317, + 220, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 317, + 220, + 555, + 373 + ], + "type": "image", + "image_path": "097bb94581b239f08eff96dac2236fd0d5e7ecd172d3872445a90e5f15c8cfde.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 376, + 556, + 477 + ], + "lines": [ + { + "bbox": [ + 313, + 376, + 556, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 556, + 477 + ], + "type": "text", + "content": "Figure 1. Illustration of our proposed Density-Guided Poisoning Attack for 3D Gaussian Splatting (3DGS). Our method strategically distribute the Gaussian points of the illusory object (i.e. the red vehicle) among the low-density regions which are discovered along the rays casted from the virtual camera of the poisoned view (i.e. the target view that we would like to attack), making the illusory object clearly visible from the poisoned view while having the minimal interference for the rendering quality on the other non-target/innocent views." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 494, + 556, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 494, + 556, + 614 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 556, + 614 + ], + "type": "text", + "content": "and is encoded into the 3D representations during model training), in which the further utilization or visualization of the poisoned 3D representations would lead to abnormal or even malicious model behaviours (i.e. the negative impact stemmed from the poison is triggered). To this end, in this work, we focus on the investigation of poisoning attacks on 3D scene representation methods, as they become integral to safety-critical applications. Hence, addressing their security vulnerabilities is not only of utmost importance but also urgent." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "type": "text", + "content": "While there exists a prior work of studying the poisoning attack upon NeRF, i.e. IPA-NeRF [40] which effectively exploits NeRF's implicit representations to embed targeted visual illusions (i.e. the illusion as poison will appear while rendering the scene from a certain viewing direction), its applicability to explicit 3D scene representations (particularly 3DGS) however remains limited and not directly transferable. Considering the rapidly growing application scenarios" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27400" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 287 + ], + "type": "text", + "content": "of 3DGS (thanks to its capability of fast rendering and accurately capturing scene geometry), we devote our research effort to realizing poisoning attacks on 3D Gaussian splatt. To the best of our knowledge, this is the first work of its kind. While concurrent work, Poison-Splat [62] focuses on computational cost attacks, our work targets visible illusion embedding. In particular, we would like to inject the visible illusory objects (i.e., poison) onto a target view (named as poisoned view) while keeping the other non-target views (named as innocent views) unaffected, as shown in Figure 1. Our work starts from conducting an investigation (cf. Figure 2) upon the robustness of 3DGS against the prior image-level poisoning methods such as IPA-NeRF, where we find that the attempts of directly adopting IPA-NeRF's approach or naively injecting illusory content into training images easily fail, as 3DGS's inherent multi-view consistency and densification processes effectively neutralize or significantly weaken these attacks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 288, + 298, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 288, + 298, + 515 + ], + "spans": [ + { + "bbox": [ + 56, + 288, + 298, + 515 + ], + "type": "text", + "content": "Motivated by the aforementioned investigation, we propose a density-guided poisoning method for 3DGS. Our approach (cf. Figure 1) strategically identifies low-density regions in the initial Gaussian point cloud using Kernel Density Estimation (KDE), in which the points of illusory objects are then distributed among the low-density regions along the rays casting from the virtual camera of target view (i.e. the rays are casted from the virtual camera with the target viewing direction). These points effectively embed illusory objects which would be clearly visible from targeted views, while having minimal impact on other innocent views (i.e. being less perceptible). Moreover, we introduce the adaptive Gaussian noise into innocent views during training for disrupting the property of multi-view consistency in 3DGS, in order to further enhance the overall efficacy of the attack. We conduct extensive experiments, and the results demonstrate the consistent superiority of our proposed poisoning method in comparison to several baselines. The contribution of our work can be summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 520, + 296, + 637 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 67, + 520, + 295, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 520, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 67, + 520, + 295, + 554 + ], + "type": "text", + "content": "- We are the first work to address data poisoning attacks upon 3D Gaussian Splatting for illusory objects injection." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 561, + 295, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 561, + 295, + 585 + ], + "spans": [ + { + "bbox": [ + 67, + 561, + 295, + 585 + ], + "type": "text", + "content": "- We identify and analyze the robustness of 3DGS against the prior poisoning attack techniques." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 590, + 296, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 590, + 296, + 637 + ], + "spans": [ + { + "bbox": [ + 67, + 590, + 296, + 637 + ], + "type": "text", + "content": "- We propose a density-guided poisoning method tailored for 3DGS, introducing adaptive noise scheduling to disrupt the multi-view consistency of 3DGS and better realize the entire attack." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 654, + 142, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 142, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 142, + 666 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Adversarial Attack. Adversarial attacks [8, 27, 54, 66, 79, 86, 90] are a critical research area in machine learning and computer vision [1, 9, 107]. These attacks exploit" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 553, + 156 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 553, + 156 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 553, + 156 + ], + "type": "image", + "image_path": "e48a29e7ea137b218d8d0ed871da30b110101e26111fb35634de8036813eb3de.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 160, + 555, + 227 + ], + "lines": [ + { + "bbox": [ + 313, + 160, + 555, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 160, + 555, + 227 + ], + "type": "text", + "content": "Figure 2. Limitations of existing poisoning methods on 3DGS. Existing poisoning methods (e.g., IPA-NeRF [40] designed for NeRF or the one adapted to 3DGS, denoted as IPA-Splat) produce weak or absent illusions due to 3DGS's robustness and multi-view consistency. In contrast, our proposed approach successfully injects clearly visible illusory objects (i.e., the dog)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 253, + 556, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 556, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 556, + 494 + ], + "type": "text", + "content": "ML model vulnerabilities by creating inputs that cause misclassification while appearing normal to humans. Goodfellow et al. [27] introduced FGSM, showing how small perturbations could transform panda images into gibbons, while Madry et al. [66] established PGD as a stronger iterative method. Adversarial attacks divide into black-box attacks [6, 13, 18, 30, 36, 75, 108], which operate without model access, and white-box attacks [47, 53, 72, 95, 101], which use full model knowledge. Black-box methods include transfer-based attacks [20, 58] exploiting cross-model transferability and query-based approaches [2, 12] that iteratively refine perturbations. White-box methods like Carlini-Wagner attacks [8] formulate adversarial generation as optimization problems, achieving high success rates while maintaining imperceptibility. Recent research has expanded to complex domains including object detection [70, 99], semantic segmentation [3], and 3D point clouds [56, 96]. Our approach resembles black-box attacks, modifying point clouds and applying simple perturbations to input images, extending adversarial concepts to 3D Gaussian Splitting representations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 498, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 557, + 714 + ], + "type": "text", + "content": "Data Poisoning. Data poisoning [14, 26, 80, 98] represents a critical vulnerability in machine learning systems. Unlike adversarial attacks targeting inference, poisoning attacks manipulate training by injecting crafted samples that exploit learning mechanisms, causing defective statistical distributions [7, 37]. Two primary poisoning strategies exist: creating indistinguishable malicious samples that blend with normal data [85, 91, 98], or employing surgical precision by polluting minimal data subsets [39, 80]. These approaches prove effective across computer vision [34, 105], NLP [48, 92], and recommender systems [23, 50]. They categorize as availability attacks degrading overall performance [74, 83] or integrity attacks targeting specific inputs [14, 29]. More sophisticated approaches formulate poisoning as bi-level optimization problems [46, 52, 59, 61], allowing attackers to optimize poisoned samples while anticipating victim model behavior. Common defenses include robust statistics [19, 89], data sanitization [17, 76], and differential privacy [65]. Our" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27401" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "content": "work employs the surgical precision strategy, conducting a stealth attack with minimal contaminated data while exploring 3D Gaussian Splatting's unique vulnerabilities to poisoning." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 123, + 298, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 123, + 298, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 123, + 298, + 460 + ], + "type": "text", + "content": "Neural Rendering. Novel view synthesis (NVS)[10, 24, 88, 109] has evolved from traditional graphics to learning-based methods. Neural Radiance Field (NeRF)[71] and 3D Gaussian Splatting (3DGS)[42] have transformed 3D scene representation. NeRF uses MLP-based networks to implicitly represent 3D scenes, leveraging differential alpha blending[43, 77] for high-quality volumetric rendering [41, 68]. Extensions like Mip-NeRF [4], Instant-NGP [73], and NeRF-W [67] address anti-aliasing, speed, and real-world limitations. Recent advances include few-shot synthesis [55], MVS-based approaches for large scenes [84], and improved robustness for dynamic content [60]. Robustness enhancements include progressive optimization [69] and joint camera pose optimization [11]. However, NeRF's implicit representation lacks explicit geometric constraints, creating vulnerabilities to view-specific perturbations [57, 93]. IPA-NeRF [40] exploits this for data poisoning attacks. In contrast, 3DGS [42] uses explicit representation with discrete 3D Gaussian primitives inspired by point-based rendering [28, 110]. This provides stronger geometry constraints and multi-view consistency with superior rendering efficiency [15, 102]. 3DGS has expanded to dynamic scenes [63, 100], human avatars [45, 49], efficient implementations [16, 22], compression [104], robustness for unconstrained images [33], and specular reconstruction [21]. Our work addresses the challenge of attacking 3DGS's robust explicit representation, contributing to understanding security implications in modern neural rendering techniques." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 462, + 298, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 298, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 298, + 714 + ], + "type": "text", + "content": "Data Poisoning on Neural Rendering. As neural rendering gains adoption, security vulnerabilities have attracted attention. IPA-NeRF [40] pioneered poisoning attacks against NeRF by inserting crafted samples at specific viewing angles, formulating data poisoning as bi-level optimization. Lu et al.'s Poison-Splat [62] targeted 3DGS efficiency by generating samples that dramatically increase memory consumption, demonstrating attacks targeting resource utilization [81] rather than accuracy. The security landscape extends to privacy concerns [64] and adversarial examples [25, 32, 103]. Zeybey et al. [103] showed how adversarial noise in 3D objects misleads vision-language models like CLIP [78]. Song et al.'s Geometry Cloak [82] prevents unauthorized 3D reconstruction from copyrighted images, while security studies examine point clouds [31, 56], meshes [97], and broader vision systems [1, 9]. Similar to steganography approaches embedding hidden information [51, 106], our method injects view-dependent content but enables decoder-free extraction through standard rendering from specific viewpoints. Our work introduces a density-guided attack methodology targeting 3DGS's initial point cloud prior, exploiting its robustness" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 556, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 120 + ], + "type": "text", + "content": "characteristics. We propose an evaluation protocol based on scene density analysis to identify optimal positions for poison injection, contributing insights that could inform future defense mechanisms against such attacks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 129, + 370, + 143 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 129, + 370, + 143 + ], + "spans": [ + { + "bbox": [ + 313, + 129, + 370, + 143 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 149, + 440, + 162 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 149, + 440, + 162 + ], + "spans": [ + { + "bbox": [ + 313, + 149, + 440, + 162 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "text", + "content": "Given a dataset " + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "text", + "content": " composed of multiple images " + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "inline_equation", + "content": "\\{I_k\\}_{k=1}^N" + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "text", + "content": " viewing a scene " + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "text", + "content": " from different viewing directions, 3D Gaussian Splatting is originally proposed to construct a 3D Gaussian point cloud " + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "text", + "content": " (each Gaussian has properties in terms of position, covariance, opacity, and color factors) for representing the 3D scene " + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "text", + "content": ", where we are able to render the image observation of " + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "text", + "content": " from any arbitrary view via projection and differentiable tile rasterizer. Basically, the goal of our poisoning attack upon 3DGS is to inject illusory/poison object " + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "inline_equation", + "content": "O_{\\mathrm{ILL}}" + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "text", + "content": " onto the target view " + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "inline_equation", + "content": "v_{\\mathbf{p}}" + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "text", + "content": ", where we denoted the resultant Gaussian point cloud after being poisoned as " + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "inline_equation", + "content": "\\tilde{G}" + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "text", + "content": ", such that the image " + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "inline_equation", + "content": "\\tilde{I}_{\\mathrm{ILL}} = \\Re(\\tilde{G}, v_{\\mathbf{p}})" + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "text", + "content": " obtained by rendering " + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "inline_equation", + "content": "\\tilde{G}" + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "inline_equation", + "content": "v_{\\mathbf{p}}" + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "text", + "content": " would contain " + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "inline_equation", + "content": "O_{\\mathrm{ILL}}" + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "text", + "content": " while keeping the images of " + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "inline_equation", + "content": "\\tilde{G}" + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "text", + "content": " rendered from any other views " + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "inline_equation", + "content": "v_{k}" + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "text", + "content": " (i.e. non-target views " + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "inline_equation", + "content": "v_{k} \\neq v_{\\mathbf{p}}" + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "text", + "content": ") being identical to the ones of " + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "text", + "content": " of the same view " + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "inline_equation", + "content": "v_{k}" + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "text", + "content": ". Please note that, when we denote the image of " + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "text", + "content": " rendered from " + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "inline_equation", + "content": "v_{\\mathbf{p}}" + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "inline_equation", + "content": "\\Re(G, v_{\\mathbf{p}})" + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "text", + "content": ", the ideal appearance of " + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "inline_equation", + "content": "\\tilde{I}_{\\mathrm{ILL}}" + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "text", + "content": " should be the combination of " + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "inline_equation", + "content": "\\Re(G, v_{\\mathbf{p}})" + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "inline_equation", + "content": "O_{\\mathrm{ILL}}" + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "text", + "content": " (in which such combination is denoted as " + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{ILL}}" + }, + { + "bbox": [ + 313, + 166, + 556, + 410 + ], + "type": "text", + "content": "). The core objective of our 3DGS poisoning is formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 318, + 416, + 556, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 416, + 556, + 444 + ], + "spans": [ + { + "bbox": [ + 318, + 416, + 556, + 444 + ], + "type": "interline_equation", + "content": "\\min _ {\\tilde {G}} \\| \\tilde {I} _ {\\mathbf {I L L}} - I _ {\\mathbf {I L L}} \\| _ {2} ^ {2} + \\sum_ {v _ {k} \\neq v _ {\\mathbf {p}}} \\| \\Re (\\tilde {G}, v _ {k}) - \\Re (G, v _ {k}) \\| _ {2} ^ {2}, \\tag {1}", + "image_path": "f7827b520a690585253152d853f1df10a714e51aee9232a2967ce6bb4ddbf553.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 450, + 556, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 556, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 556, + 510 + ], + "type": "text", + "content": "To achieve this under different threat models, we propose two strategies: density-guided point cloud attack (Section 3.3) following classical data poisoning, and view consistency disruption (Section 3.4) as a backdoor attack with minimal training modification via noise scheduling." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 517, + 544, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 544, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 544, + 529 + ], + "type": "text", + "content": "3.2. Potential Naive Approaches and Limitations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 534, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 628 + ], + "type": "text", + "content": "We first explore two potential naive approaches, followed by discussing their limitations and motivating our attack design: 1) Directly injecting the illusory object onto the training image in " + }, + { + "bbox": [ + 313, + 534, + 556, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 534, + 556, + 628 + ], + "type": "text", + "content": " of the target view " + }, + { + "bbox": [ + 313, + 534, + 556, + 628 + ], + "type": "inline_equation", + "content": "v_{p}" + }, + { + "bbox": [ + 313, + 534, + 556, + 628 + ], + "type": "text", + "content": " and following the typical training procedure of 3DGS would easily fail since the property of multi-view consistency in 3DGS would treat the illusory object as the noise (i.e. which violates the consistency) and eliminate it from the resultant Gaussian cloud." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 630, + 556, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 677 + ], + "type": "text", + "content": "2) Directly backprojecting the illusory object into the Gaussian point cloud " + }, + { + "bbox": [ + 313, + 630, + 556, + 677 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 630, + 556, + 677 + ], + "type": "text", + "content": " reconstructed from " + }, + { + "bbox": [ + 313, + 630, + 556, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 630, + 556, + 677 + ], + "type": "text", + "content": " also faces the challenge of determining proper depth for the illusory object, otherwise it could be occluded by existing geometry in " + }, + { + "bbox": [ + 313, + 630, + 556, + 677 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 630, + 556, + 677 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 684, + 526, + 696 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 684, + 526, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 684, + 526, + 696 + ], + "type": "text", + "content": "3.3. Our Density-Guided Point Cloud Attack" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 701, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 555, + 713 + ], + "type": "text", + "content": "With learning the lessons from the aforementioned naive" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27402" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 239, + 198 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 239, + 198 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 239, + 198 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 239, + 198 + ], + "type": "image", + "image_path": "5c3837895aada4145de0d50d445ea9dfb6ee805f9afb3c540432ceba479f136b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 123, + 201, + 221, + 209 + ], + "lines": [ + { + "bbox": [ + 123, + 201, + 221, + 209 + ], + "spans": [ + { + "bbox": [ + 123, + 201, + 221, + 209 + ], + "type": "text", + "content": "(a) Density-guided Point Cloud Attack" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 240, + 70, + 399, + 198 + ], + "blocks": [ + { + "bbox": [ + 240, + 70, + 399, + 198 + ], + "lines": [ + { + "bbox": [ + 240, + 70, + 399, + 198 + ], + "spans": [ + { + "bbox": [ + 240, + 70, + 399, + 198 + ], + "type": "image", + "image_path": "943cdd96767ca9aba11a37f3c3269df6f9c8c4ac9ee0c050fcbdd68690929f6e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 201, + 369, + 209 + ], + "lines": [ + { + "bbox": [ + 277, + 201, + 369, + 209 + ], + "spans": [ + { + "bbox": [ + 277, + 201, + 369, + 209 + ], + "type": "text", + "content": "(b) Normal 3DGS Training Iterations" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 399, + 71, + 553, + 198 + ], + "blocks": [ + { + "bbox": [ + 399, + 71, + 553, + 198 + ], + "lines": [ + { + "bbox": [ + 399, + 71, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 399, + 71, + 553, + 198 + ], + "type": "image", + "image_path": "b109103375f0e9e80568ea1b3a8e1c760d869f44751e18961723b7226aa18fd1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 201, + 528, + 209 + ], + "lines": [ + { + "bbox": [ + 427, + 201, + 528, + 209 + ], + "spans": [ + { + "bbox": [ + 427, + 201, + 528, + 209 + ], + "type": "text", + "content": "(c) View Consistency Disruption Attack" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 58, + 293, + 293, + 365 + ], + "blocks": [ + { + "bbox": [ + 54, + 217, + 555, + 282 + ], + "lines": [ + { + "bbox": [ + 54, + 217, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 54, + 217, + 555, + 282 + ], + "type": "text", + "content": "Figure 3. Overview of our proposed poisoning attack framework. Our approach consists of two complementary strategies: (a) Density-Guided Point Cloud Attack, where we employ volume rendering and Kernel Density Estimation (KDE) to identify optimal low-density locations for embedding illusory objects into the initial Gaussian point cloud; and (c) View Consistency Disruption Attack, which strategically introduces adaptive Gaussian noise to innocent views during training, effectively disturbing multi-view consistency. (b) illustrates the standard 3D Gaussian Splitting (3DGS) training pipeline for reference. The combined strategies successfully inject convincing illusions from poisoned views while maintaining high fidelity in innocent viewpoints." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 293, + 293, + 365 + ], + "lines": [ + { + "bbox": [ + 58, + 293, + 293, + 365 + ], + "spans": [ + { + "bbox": [ + 58, + 293, + 293, + 365 + ], + "type": "image", + "image_path": "a1972bfd8fe7408d8735a0483626eab09b86aa6ab9e516355880c04b30c8067d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 368, + 296, + 445 + ], + "lines": [ + { + "bbox": [ + 55, + 368, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 368, + 296, + 445 + ], + "type": "text", + "content": "Figure 4. Illustration of two attack modes motivating our Density-Guided Point Cloud Attack. (a) Points placed outside the coverage of innocent viewpoints can effectively embed illusions visible only from the poisoned view. (b) Points occluded from innocent viewpoints also provide viable hidden locations. These scenarios motivate our Density-Guided strategy for robust and stealthy attacks." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 472, + 295, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 472, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 54, + 472, + 295, + 638 + ], + "type": "text", + "content": "approaches, we conclude that effective 3DGS poisoning must consider the inherent properties of 3DGS's explicit representation and multi-view consistency, while ensuring illusion visibility on the target view and minimal perception on innocent views. To this end, we propose two simple but effective ideas for discovering optimal 3D positions in " + }, + { + "bbox": [ + 54, + 472, + 295, + 638 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 54, + 472, + 295, + 638 + ], + "type": "text", + "content": " to place backprojected Gaussian points of the illusory object, as illustrated in Figure 4: First, poison points can be placed in regions invisible to innocent views; Second, regions occluded for innocent views by existing geometry in " + }, + { + "bbox": [ + 54, + 472, + 295, + 638 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 54, + 472, + 295, + 638 + ], + "type": "text", + "content": " are effective candidates (visible for target view but invisible for innocent views due to occluders). To identify regions satisfying these ideas, we propose a density-guided point placement strategy, detailed below." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 641, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 641, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 54, + 641, + 296, + 712 + ], + "type": "text", + "content": "Scene Space Analysis. Given the Gaussian point cloud " + }, + { + "bbox": [ + 54, + 641, + 296, + 712 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 54, + 641, + 296, + 712 + ], + "type": "text", + "content": " (reconstructed from " + }, + { + "bbox": [ + 54, + 641, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 54, + 641, + 296, + 712 + ], + "type": "text", + "content": " via the typical 3DGS procedure), we start from determining the Axis-Aligned Bounding Box (AABB) of " + }, + { + "bbox": [ + 54, + 641, + 296, + 712 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 54, + 641, + 296, + 712 + ], + "type": "text", + "content": " (basically, finding the minimum and maximum coordinates across all points in " + }, + { + "bbox": [ + 54, + 641, + 296, + 712 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 54, + 641, + 296, + 712 + ], + "type": "text", + "content": ") and creating a rectangular box that fully encloses the entire 3D scene " + }, + { + "bbox": [ + 54, + 641, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 54, + 641, + 296, + 712 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 54, + 641, + 296, + 712 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 54, + 641, + 296, + 712 + ], + "type": "text", + "content": ", with its" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 294, + 553, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 553, + 329 + ], + "type": "text", + "content": "edges being aligned to the coordinate axes. Such bounding box is then first decomposed/voxelized into a uniform grid " + }, + { + "bbox": [ + 313, + 294, + 553, + 329 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 294, + 553, + 329 + ], + "type": "text", + "content": ", where we denote each cell in the grid as " + }, + { + "bbox": [ + 313, + 294, + 553, + 329 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 294, + 553, + 329 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 330, + 554, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 554, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 554, + 400 + ], + "type": "text", + "content": "As the opacity " + }, + { + "bbox": [ + 313, + 330, + 554, + 400 + ], + "type": "inline_equation", + "content": "\\alpha(g)" + }, + { + "bbox": [ + 313, + 330, + 554, + 400 + ], + "type": "text", + "content": " of each Gaussian point " + }, + { + "bbox": [ + 313, + 330, + 554, + 400 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 330, + 554, + 400 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 330, + 554, + 400 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 330, + 554, + 400 + ], + "type": "text", + "content": " can be estimate via volume rendering technique (depending on our target view " + }, + { + "bbox": [ + 313, + 330, + 554, + 400 + ], + "type": "inline_equation", + "content": "v_{\\mathbf{p}}" + }, + { + "bbox": [ + 313, + 330, + 554, + 400 + ], + "type": "text", + "content": "), we can easily compute the density " + }, + { + "bbox": [ + 313, + 330, + 554, + 400 + ], + "type": "inline_equation", + "content": "\\rho(s)" + }, + { + "bbox": [ + 313, + 330, + 554, + 400 + ], + "type": "text", + "content": " of each voxel " + }, + { + "bbox": [ + 313, + 330, + 554, + 400 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 330, + 554, + 400 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 313, + 330, + 554, + 400 + ], + "type": "inline_equation", + "content": "\\sum_{g \\in s} \\alpha(g)" + }, + { + "bbox": [ + 313, + 330, + 554, + 400 + ], + "type": "text", + "content": " as the opacity and density are highly related, where " + }, + { + "bbox": [ + 313, + 330, + 554, + 400 + ], + "type": "inline_equation", + "content": "g \\in s" + }, + { + "bbox": [ + 313, + 330, + 554, + 400 + ], + "type": "text", + "content": " indicates that Gaussian point " + }, + { + "bbox": [ + 313, + 330, + 554, + 400 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 330, + 554, + 400 + ], + "type": "text", + "content": " is located within voxel " + }, + { + "bbox": [ + 313, + 330, + 554, + 400 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 330, + 554, + 400 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 404, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 404, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 404, + 555, + 453 + ], + "type": "text", + "content": "Continuous Density Estimation. Based on the per-voxel density " + }, + { + "bbox": [ + 313, + 404, + 555, + 453 + ], + "type": "inline_equation", + "content": "\\rho(s)" + }, + { + "bbox": [ + 313, + 404, + 555, + 453 + ], + "type": "text", + "content": ", we apply Gaussian Kernel Density Estimation (KDE) to obtain a continuous density estimate for any arbitrary 3D position " + }, + { + "bbox": [ + 313, + 404, + 555, + 453 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 404, + 555, + 453 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 358, + 460, + 555, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 460, + 555, + 488 + ], + "spans": [ + { + "bbox": [ + 358, + 460, + 555, + 488 + ], + "type": "interline_equation", + "content": "f (x) = \\frac {1}{| \\mathcal {S} |} \\sum_ {s \\in \\mathcal {S}} K _ {h} (x - c (s)) \\cdot \\rho (s), \\tag {2}", + "image_path": "5016a3f1b531e102f7ea4d97d63422e72d3c07ef0429852bf4cf08c0fd1ff599.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 497, + 554, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 554, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 554, + 520 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 497, + 554, + 520 + ], + "type": "inline_equation", + "content": "c(s)" + }, + { + "bbox": [ + 313, + 497, + 554, + 520 + ], + "type": "text", + "content": " denotes the centroid of a voxel " + }, + { + "bbox": [ + 313, + 497, + 554, + 520 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 497, + 554, + 520 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 497, + 554, + 520 + ], + "type": "inline_equation", + "content": "K_{h}" + }, + { + "bbox": [ + 313, + 497, + 554, + 520 + ], + "type": "text", + "content": " is the Gaussian kernel with bandwidth " + }, + { + "bbox": [ + 313, + 497, + 554, + 520 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 497, + 554, + 520 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 356, + 526, + 555, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 526, + 555, + 554 + ], + "spans": [ + { + "bbox": [ + 356, + 526, + 555, + 554 + ], + "type": "interline_equation", + "content": "K _ {h} (x) = \\frac {1}{(2 \\pi h ^ {2}) ^ {3 / 2}} \\exp \\left(- \\frac {\\| x \\| ^ {2}}{2 h ^ {2}}\\right). \\tag {3}", + "image_path": "5edced8b97bfbde9cfd1edf89578d13353a853e6c2a59ec3565975c0083606df.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "Optimal Position Selection. With the illusory object placed on the image plane of the virtual camera rendering the Gaussian point cloud from target view " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "v_{\\mathbf{p}}" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": ", backprojection starts by casting rays from camera center " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " through all pixels of the illusory object. We sample points along each ray to find regions in the Gaussian cloud with minimum density, via querying the KDE result described above. Given a casting ray with direction " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": ", any sampled 3D position along the ray can be written as " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "C + t \\cdot d" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " ranges from " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{min}}" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{max}}" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": ". We set " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{min}}" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " to 0.3 (as points near camera often appear as floaters in 3DGS optimization) and " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{max}}" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " represents the original scene depth at each pixel in the poisoned view." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27403" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "The sampled point in the minimum density region can be computed by:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 110, + 104, + 296, + 126 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 104, + 296, + 126 + ], + "spans": [ + { + "bbox": [ + 110, + 104, + 296, + 126 + ], + "type": "interline_equation", + "content": "x _ {\\min } = \\underset {x \\in C + t \\cdot d, t \\in [ t _ {\\min }, t _ {\\max } ]} {\\arg \\min } f (x). \\tag {4}", + "image_path": "3972d906407b6961f4fff78af91e34ce113ce1caf9210c6fe8898262523bae8b.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 132, + 296, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 132, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 132, + 296, + 216 + ], + "type": "text", + "content": "We then insert new Gaussian poison points at position " + }, + { + "bbox": [ + 55, + 132, + 296, + 216 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{min}}" + }, + { + "bbox": [ + 55, + 132, + 296, + 216 + ], + "type": "text", + "content": ", assigning colors from the illusory object (i.e. the color value for the new Gaussian point is obtained from the corresponding illusory object pixel). Our proposed density-guided method hence strategically places Gaussian points of poison to embed the illusory object prominently from the poisoned view while minimizing its visibility from innocent views." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 222, + 249, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 222, + 249, + 235 + ], + "spans": [ + { + "bbox": [ + 55, + 222, + 249, + 235 + ], + "type": "text", + "content": "3.4. View Consistency Disruption Attack" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 239, + 296, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 346 + ], + "type": "text", + "content": "While our Density-Guided Point Cloud Attack (cf. Section 3.3) effectively places Gaussian points of poison in many cases, scenes with high view overlap (i.e. the field-of-views of the training image " + }, + { + "bbox": [ + 55, + 239, + 296, + 346 + ], + "type": "inline_equation", + "content": "\\{I_k\\}_{k=1}^N" + }, + { + "bbox": [ + 55, + 239, + 296, + 346 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 239, + 296, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 239, + 296, + 346 + ], + "type": "text", + "content": " have high overlaps) remain challenging. To address this, we introduce the View Consistency Disruption Attack, which strategically adds controlled noise to innocent views, thus weakening the multi-view consistency of 3DGS and better preserving our injected illusions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 347, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 296, + 384 + ], + "type": "text", + "content": "We selectively apply Gaussian noise to innocent views, leaving the poisoned view clean. For a training image " + }, + { + "bbox": [ + 55, + 347, + 296, + 384 + ], + "type": "inline_equation", + "content": "I_{k}" + }, + { + "bbox": [ + 55, + 347, + 296, + 384 + ], + "type": "text", + "content": " with view direction " + }, + { + "bbox": [ + 55, + 347, + 296, + 384 + ], + "type": "inline_equation", + "content": "v_{k}" + }, + { + "bbox": [ + 55, + 347, + 296, + 384 + ], + "type": "text", + "content": ", the noise is applied as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 88, + 391, + 296, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 391, + 296, + 405 + ], + "spans": [ + { + "bbox": [ + 88, + 391, + 296, + 405 + ], + "type": "interline_equation", + "content": "I _ {k} ^ {\\prime} = \\mathbf {1} _ {v _ {k} = v _ {\\mathbf {p}}} \\cdot I _ {k} + \\mathbf {1} _ {v _ {k} \\neq v _ {\\mathbf {p}}} \\cdot \\operatorname {C L I P} (I _ {k} + \\eta), \\tag {5}", + "image_path": "46243eb648c62ca646ce9719e5235918854e3054a0357b371555936cc43285c1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 411, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 296, + 518 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 411, + 296, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{1}" + }, + { + "bbox": [ + 55, + 411, + 296, + 518 + ], + "type": "text", + "content": " is an indicator function, CLIP prevents " + }, + { + "bbox": [ + 55, + 411, + 296, + 518 + ], + "type": "inline_equation", + "content": "I_{k} + \\eta" + }, + { + "bbox": [ + 55, + 411, + 296, + 518 + ], + "type": "text", + "content": " from exceeding the pixel value range, and " + }, + { + "bbox": [ + 55, + 411, + 296, + 518 + ], + "type": "inline_equation", + "content": "\\eta \\sim \\mathcal{N}(0,\\sigma_t^2)" + }, + { + "bbox": [ + 55, + 411, + 296, + 518 + ], + "type": "text", + "content": " denotes noise with strength " + }, + { + "bbox": [ + 55, + 411, + 296, + 518 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 55, + 411, + 296, + 518 + ], + "type": "text", + "content": " adjusted according to 3DGS iteration. The " + }, + { + "bbox": [ + 55, + 411, + 296, + 518 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 55, + 411, + 296, + 518 + ], + "type": "text", + "content": " scheduling follows the principle of having strong noise in early 3DGS optimization to disrupt multi-view consistency (as noise injected into training images are independent) while gradually reducing noise strength to maintain high-quality reconstruction for innocent views in late optimization. We explore three noise decay strategies:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 122, + 525, + 296, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 525, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 122, + 525, + 296, + 548 + ], + "type": "interline_equation", + "content": "\\sigma_ {\\text {l i n e a r}} (t) = \\sigma_ {0} \\cdot \\left(1 - \\frac {t}{T}\\right), \\tag {6}", + "image_path": "8463a0938d2fb8ea44437f041beeecede3ed08a5421e5d23e5af10077094e725.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 121, + 550, + 296, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 550, + 296, + 572 + ], + "spans": [ + { + "bbox": [ + 121, + 550, + 296, + 572 + ], + "type": "interline_equation", + "content": "\\sigma_ {\\text {c o s i n e}} (t) = \\sigma_ {0} \\cdot \\cos \\left(\\frac {\\pi t}{2 T}\\right), \\tag {7}", + "image_path": "c07265129f39dddcbc224510f30c57113fc8afb679f54fc496cc913617c507cb.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 128, + 573, + 296, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 573, + 296, + 600 + ], + "spans": [ + { + "bbox": [ + 128, + 573, + 296, + 600 + ], + "type": "interline_equation", + "content": "\\sigma_ {\\mathrm {s q r t}} (t) = \\sigma_ {0} \\cdot \\sqrt {1 - \\frac {t}{T}}, \\tag {8}", + "image_path": "e6516425021e19994f163354bcb9ac20be02058d6b58e8be0b626b58f750dbcf.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 605, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 605, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\sigma_0" + }, + { + "bbox": [ + 55, + 605, + 296, + 665 + ], + "type": "text", + "content": " is the initial noise strength, " + }, + { + "bbox": [ + 55, + 605, + 296, + 665 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 605, + 296, + 665 + ], + "type": "text", + "content": " is the current iteration, and " + }, + { + "bbox": [ + 55, + 605, + 296, + 665 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 605, + 296, + 665 + ], + "type": "text", + "content": " is the total training iterations. Linear decay reduces noise evenly, cosine decay provides a smooth start and accelerates later reduction, and square root decay maintains higher noise longer before a rapid decrease." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Our controlled noise injection creates intentional imbalance during training, enabling preservation of illusory objects from the poisoned viewpoint while ensuring scene fidelity as noise diminishes. The overview of our proposed" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "content": "method, composed of both density-guided point cloud attack (cf. Section 3.3) and view consistency disruption attack (cf. Section 3.4), is provided in Figure 3." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 121, + 395, + 134 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 121, + 395, + 134 + ], + "spans": [ + { + "bbox": [ + 314, + 121, + 395, + 134 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 141, + 430, + 155 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 141, + 430, + 155 + ], + "spans": [ + { + "bbox": [ + 314, + 141, + 430, + 155 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 163, + 554, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 554, + 233 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 554, + 233 + ], + "type": "text", + "content": "Datasets. We evaluate our method on three common datasets: (1) Mip-NeRF360 [5] with complex " + }, + { + "bbox": [ + 313, + 163, + 554, + 233 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 163, + 554, + 233 + ], + "type": "text", + "content": " scenes, (2) Tanks & Temples [44] containing realistic indoor and outdoor captures, and (3) Free [94], featuring unbounded scenes with free camera trajectories. These datasets provide diverse benchmarks for novel view synthesis evaluation." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 238, + 555, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 238, + 555, + 333 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 555, + 333 + ], + "type": "text", + "content": "Compared Methods. We evaluate our method against three baselines: (1) IPA-NeRF (Nerfacto) [40]: The original backdoor attack applied to Nerfacto [87], featuring advanced static scene reconstruction techniques. (2) IPA-NeRF (Instant-NGP) [40]: The original backdoor attack on Instant-NGP [73], known for accelerated training and rendering. (3) IPA-Splat: Our adaptation of IPA-NeRF specifically for 3D Gaussian Splatting." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 334, + 555, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 555, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 555, + 417 + ], + "type": "text", + "content": "For IPA-NeRF baselines, we maintain original settings but reduce total iterations to " + }, + { + "bbox": [ + 313, + 334, + 555, + 417 + ], + "type": "inline_equation", + "content": "O = 15,000" + }, + { + "bbox": [ + 313, + 334, + 555, + 417 + ], + "type": "text", + "content": " for faster convergence. Other parameters remain unchanged: " + }, + { + "bbox": [ + 313, + 334, + 555, + 417 + ], + "type": "inline_equation", + "content": "T = 200" + }, + { + "bbox": [ + 313, + 334, + 555, + 417 + ], + "type": "text", + "content": " iterations per epoch, " + }, + { + "bbox": [ + 313, + 334, + 555, + 417 + ], + "type": "inline_equation", + "content": "O / T = 75" + }, + { + "bbox": [ + 313, + 334, + 555, + 417 + ], + "type": "text", + "content": " attack epochs, " + }, + { + "bbox": [ + 313, + 334, + 555, + 417 + ], + "type": "inline_equation", + "content": "A = 10" + }, + { + "bbox": [ + 313, + 334, + 555, + 417 + ], + "type": "text", + "content": " attack iterations per epoch, " + }, + { + "bbox": [ + 313, + 334, + 555, + 417 + ], + "type": "inline_equation", + "content": "K = 100" + }, + { + "bbox": [ + 313, + 334, + 555, + 417 + ], + "type": "text", + "content": " perturbation renderings, distortion budget " + }, + { + "bbox": [ + 313, + 334, + 555, + 417 + ], + "type": "inline_equation", + "content": "\\epsilon = 32" + }, + { + "bbox": [ + 313, + 334, + 555, + 417 + ], + "type": "text", + "content": ", constraint parameter " + }, + { + "bbox": [ + 313, + 334, + 555, + 417 + ], + "type": "inline_equation", + "content": "\\eta = 1" + }, + { + "bbox": [ + 313, + 334, + 555, + 417 + ], + "type": "text", + "content": ", and view constraints " + }, + { + "bbox": [ + 313, + 334, + 555, + 417 + ], + "type": "inline_equation", + "content": "(13^{\\circ}" + }, + { + "bbox": [ + 313, + 334, + 555, + 417 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 334, + 555, + 417 + ], + "type": "inline_equation", + "content": "15^{\\circ}" + }, + { + "bbox": [ + 313, + 334, + 555, + 417 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 418, + 555, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 555, + 514 + ], + "type": "text", + "content": "For our IPA-Splat method, we adapt IPA-NeRF to 3D Gaussian Splatting with " + }, + { + "bbox": [ + 313, + 418, + 555, + 514 + ], + "type": "inline_equation", + "content": "O = 30,000" + }, + { + "bbox": [ + 313, + 418, + 555, + 514 + ], + "type": "text", + "content": " total iterations and " + }, + { + "bbox": [ + 313, + 418, + 555, + 514 + ], + "type": "inline_equation", + "content": "T = 200" + }, + { + "bbox": [ + 313, + 418, + 555, + 514 + ], + "type": "text", + "content": " normal training iterations per epoch, resulting in " + }, + { + "bbox": [ + 313, + 418, + 555, + 514 + ], + "type": "inline_equation", + "content": "O / T = 150" + }, + { + "bbox": [ + 313, + 418, + 555, + 514 + ], + "type": "text", + "content": " epochs with " + }, + { + "bbox": [ + 313, + 418, + 555, + 514 + ], + "type": "inline_equation", + "content": "A = 10" + }, + { + "bbox": [ + 313, + 418, + 555, + 514 + ], + "type": "text", + "content": " attack iterations each. Other settings match IPA-NeRF. Due to 3DGS's explicit representation, we implement separate parameter constraints (xyz coordinates, feature vectors, scaling, rotation, and opacity) for precise control." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 517, + 554, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 554, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 554, + 613 + ], + "type": "text", + "content": "Evaluation Metrics. Following IPA-NeRF [40], we evaluate using PSNR, SSIM, and LPIPS metrics on two view sets: (1) V-ILLUSORY, focusing on masked metrics for illusory objects, and (2) V-TEST, assessing performance on unseen viewpoints. Attack success is defined as achieving PSNR " + }, + { + "bbox": [ + 313, + 517, + 554, + 613 + ], + "type": "inline_equation", + "content": "> 25" + }, + { + "bbox": [ + 313, + 517, + 554, + 613 + ], + "type": "text", + "content": " on V-ILLUSORY while maintaining V-TEST PSNR drop " + }, + { + "bbox": [ + 313, + 517, + 554, + 613 + ], + "type": "inline_equation", + "content": "\\leq 3" + }, + { + "bbox": [ + 313, + 517, + 554, + 613 + ], + "type": "text", + "content": ", ensuring effective illusion generation and preserved innocent view quality." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 617, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 689 + ], + "type": "text", + "content": "Evaluation Protocol. Our evaluation accounts for varying attack difficulty across camera positions. As shown in Figure 5, datasets exhibit distinct camera patterns. For uniform datasets like Mip-NeRF 360 (e.g., the \"bicycle\" scene) and Tanks-and-Temple, we select the median-indexed frame as the attack viewpoint." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "For Free dataset scenes with irregular camera trajectories (e.g., the \"stair\" scene), we quantify varying attack difficulty" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27404" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 105, + 555, + 186 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 103 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons on single-view attack. Metrics evaluated on Mip-NeRF 360 [5], Tanks & Temples [44], and Free datasets. Our method significantly outperforms baseline attacks in embedding illusory objects (V-IILLUSORY) while maintaining high fidelity in other views (V-TEST)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 105, + 555, + 186 + ], + "lines": [ + { + "bbox": [ + 57, + 105, + 555, + 186 + ], + "spans": [ + { + "bbox": [ + 57, + 105, + 555, + 186 + ], + "type": "table", + "html": "
MethodMip-NeRF 360 [5] datasetTanks & Temples [44] datasetFree [94] dataset
V-IILLUSORYV-TESTV-IILLUSORYV-TESTV-IILLUSORYV-TEST
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Naive 3DGS (w/o attack)13.210.5210.73129.450.8830.16513.150.6160.73230.600.9150.13512.000.3150.90526.800.8260.228
IPA-NeRF [40] (Nerfacto [87])16.000.5820.68521.940.5860.41513.510.6360.71123.880.7300.21813.930.4430.69920.280.4970.532
IPA-NeRF [40] (Instant-NGP [73])17.600.6180.64120.000.5170.47916.050.6930.61620.290.6690.35018.940.5080.51920.430.5030.548
IPA-Splat13.230.5180.74027.390.8290.24713.430.6250.72428.530.8910.19012.600.3720.74424.710.7490.341
Ours27.040.8130.36927.760.8050.28621.330.8090.37127.580.8520.23926.660.7540.31725.250.7280.382
", + "image_path": "8ac6d8aefe7938395711ce139d2c170f800a2745b28e275cd41f10547fe907a8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 96, + 241, + 514, + 316 + ], + "blocks": [ + { + "bbox": [ + 55, + 197, + 555, + 240 + ], + "lines": [ + { + "bbox": [ + 55, + 197, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 197, + 555, + 240 + ], + "type": "text", + "content": "Table 2. Quantitative comparisons on single-view attack with different difficulty levels on the Free [94] dataset. We evaluate attack effectiveness (V-ILLUSORY) at varying difficulty levels, defined by our KDE-based evaluation protocol. Our method consistently achieves superior results across all metrics, clearly outperforming state-of-the-art methods, especially in EASY and MEDIAN scenarios, while remaining effective in the challenging HARD scenario." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 96, + 241, + 514, + 316 + ], + "lines": [ + { + "bbox": [ + 96, + 241, + 514, + 316 + ], + "spans": [ + { + "bbox": [ + 96, + 241, + 514, + 316 + ], + "type": "table", + "html": "
MethodEASYMEDIANHARD
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
IPA-NeRF (nerfacto)15.040.4820.66213.930.4430.69914.250.4500.728
IPA-NeRF (instant-ngp)18.170.5180.54118.940.5080.51917.950.4870.557
IPA-Splat13.940.4790.65812.600.3720.74313.060.3400.796
Ours29.940.8530.18826.660.7540.31717.530.5260.581
", + "image_path": "5c9f07d2003fc2392168871a9e7ad08bfd7dff760543fae3d5707438d2556549.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 57, + 332, + 294, + 438 + ], + "blocks": [ + { + "bbox": [ + 57, + 332, + 294, + 438 + ], + "lines": [ + { + "bbox": [ + 57, + 332, + 294, + 438 + ], + "spans": [ + { + "bbox": [ + 57, + 332, + 294, + 438 + ], + "type": "image", + "image_path": "c70dbf187a8f9d618d7c54ede62f6ac63d0682c8b495aa2859a408a6cb17e5e5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 440, + 296, + 495 + ], + "lines": [ + { + "bbox": [ + 55, + 440, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 440, + 296, + 495 + ], + "type": "text", + "content": "Figure 5. Our evaluation protocol. We evaluate two scenes with varying difficulties. Left: The \"bicycle\" scene (Mip-NeRF 360 [5]) has uniform camera coverage, providing similar difficulty across views. Right: The \"stair\" scene (Free [94]) has increasing difficulty as later views are visible from more prior viewpoints." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 522, + 185, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 185, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 185, + 534 + ], + "type": "text", + "content": "using our KDE-based protocol:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 538, + 296, + 597 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 56, + 538, + 292, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 538, + 292, + 550 + ], + "spans": [ + { + "bbox": [ + 56, + 538, + 292, + 550 + ], + "type": "text", + "content": "1. Compute overall scene density distribution using KDE." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 550, + 295, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 295, + 574 + ], + "type": "text", + "content": "2. Calculate camera viewpoint densities within their FOV using camera intrinsics and a " + }, + { + "bbox": [ + 55, + 550, + 295, + 574 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 550, + 295, + 574 + ], + "type": "text", + "content": " sampling radius." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 574, + 296, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 296, + 597 + ], + "type": "text", + "content": "3. Sort cameras by density, select three representative viewpoints:" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 81, + 603, + 274, + 649 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 81, + 603, + 266, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 603, + 266, + 614 + ], + "spans": [ + { + "bbox": [ + 81, + 603, + 266, + 614 + ], + "type": "text", + "content": "EASY: Minimum density (lowest coverage)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 81, + 620, + 274, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 620, + 274, + 632 + ], + "spans": [ + { + "bbox": [ + 81, + 620, + 274, + 632 + ], + "type": "text", + "content": "- MEDIAN: Median density (average coverage)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 81, + 637, + 273, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 637, + 273, + 649 + ], + "spans": [ + { + "bbox": [ + 81, + 637, + 273, + 649 + ], + "type": "text", + "content": "- HARD: Maximum density (highest coverage)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 654, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 297, + 713 + ], + "type": "text", + "content": "Experiments in Tab. 2 confirm negative correlation between scene density and attack success (V-IILLUSORY), validating that higher scene coverage increases attack difficulty. This protocol enables fair evaluation across scenes and provides a benchmark for future 3DGS poisoning research." + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 316, + 335, + 553, + 440 + ], + "blocks": [ + { + "bbox": [ + 316, + 335, + 553, + 440 + ], + "lines": [ + { + "bbox": [ + 316, + 335, + 553, + 440 + ], + "spans": [ + { + "bbox": [ + 316, + 335, + 553, + 440 + ], + "type": "image", + "image_path": "70ca45540e64c8eb7f57c76830c3950dcde8df4cedc7d5fc9cba4ddba2efb2c3.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 442, + 555, + 497 + ], + "lines": [ + { + "bbox": [ + 313, + 442, + 555, + 497 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 555, + 497 + ], + "type": "text", + "content": "Figure 6. Qualitative comparisons on single-view attack. Our method generates significantly clearer and more convincing illusory objects from the poisoned viewpoint, demonstrating better multiview consistency and fewer artifacts compared to other state-of-the-art methods." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 316, + 513, + 553, + 619 + ], + "blocks": [ + { + "bbox": [ + 316, + 513, + 553, + 619 + ], + "lines": [ + { + "bbox": [ + 316, + 513, + 553, + 619 + ], + "spans": [ + { + "bbox": [ + 316, + 513, + 553, + 619 + ], + "type": "image", + "image_path": "5aceb08329024fc42e7b73406b283418722f50762b03c7b502820d50bb85edf3.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 622, + 556, + 677 + ], + "lines": [ + { + "bbox": [ + 313, + 622, + 556, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 622, + 556, + 677 + ], + "type": "text", + "content": "Figure 7. Qualitative comparisons of different poisoning methods on innocent views. Unlike baseline methods, our method effectively maintains high visual fidelity across innocent viewpoints, introducing significantly fewer artifacts and ensuring minimal disruption of the original scene appearance." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27405" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 126, + 296, + 239 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 125 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 125 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 125 + ], + "type": "text", + "content": "Table 3. Multi-view Attack Evaluation. We quantitatively evaluate our method on multiple poisoned views (V-ILLUSORY) and innocent views (V-TEST). Our approach consistently outperforms state-of-the-art methods, embedding clear illusions in targeted views while maintaining high fidelity in innocent views." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 126, + 296, + 239 + ], + "lines": [ + { + "bbox": [ + 56, + 126, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 126, + 296, + 239 + ], + "type": "table", + "html": "
# of viewsMethodV-ILLUSORY (poisoned avg.)V-TEST
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
2IPA-NeRF (Nerfacto)16.170.5830.68019.640.4570.548
IPA-NeRF (Instant-NGP)19.190.6240.61618.390.4400.539
IPA-Splat13.240.4970.75227.450.8320.243
Ours27.490.8420.29927.770.8040.286
3IPA-NeRF (Nerfacto)18.480.5840.66019.830.4620.545
IPA-NeRF (Instant-NGP)18.090.6040.64318.630.4580.524
IPA-Splat13.760.5380.73227.970.8580.223
Ours27.040.8330.31127.720.8030.287
4IPA-NeRF (Nerfacto)17.060.6260.67619.610.4670.538
IPA-NeRF (Instant-NGP)19.060.6570.63218.510.4580.523
IPA-Splat13.090.4890.79627.600.8380.228
Ours26.950.8550.30527.590.8020.287
", + "image_path": "997f2dde5faf1f2c68ea31e45b4dfd6a967b889b223bc1d337c31abf93dbc1cf.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 256, + 294, + 393 + ], + "blocks": [ + { + "bbox": [ + 58, + 256, + 294, + 393 + ], + "lines": [ + { + "bbox": [ + 58, + 256, + 294, + 393 + ], + "spans": [ + { + "bbox": [ + 58, + 256, + 294, + 393 + ], + "type": "image", + "image_path": "e0b33b6899861cd9b605d9dc39e467727c5efc879babeebe7d37d823462db10d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 396, + 296, + 440 + ], + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 440 + ], + "type": "text", + "content": "Figure 8. Qualitative comparisons on multi-view attack. Our density-guided method produces sharper, more consistent illusions across multiple poisoned views, clearly outperforming baseline methods that yield faint or inconsistent results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 468, + 167, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 167, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 167, + 481 + ], + "type": "text", + "content": "4.2. Single-view Attack" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 490, + 297, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 297, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 297, + 562 + ], + "type": "text", + "content": "Tab. 1 shows quantitative comparisons of single-view attacks. Our method outperforms baselines across all datasets. On V-ILLUSORY views, our approach significantly improves PSNR, SSIM, and LPIPS, effectively embedding convincing illusions. Performance on V-TEST remains consistently high, indicating minimal impact on innocent views." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 566, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 296, + 639 + ], + "type": "text", + "content": "Qualitative results (Figs. 6 and 7) further highlight our advantages. In Fig. 6, our method produces clearer and more convincing illusory objects compared to baselines, which often yield faint or inconsistent illusions. Fig. 7 emphasizes our superior visual fidelity in innocent views, with fewer artifacts and better overall scene quality." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 297, + 715 + ], + "type": "text", + "content": "In Tab. 2, we analyze performance across difficulty levels (EASY, MEDIAN, HARD) on the Free dataset. As expected, attack effectiveness decreases with difficulty. Nonetheless, our density-guided approach achieves superior results, particularly at EASY and MEDIAN levels, demonstrating robust performance even in challenging conditions." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 126, + 555, + 205 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 125 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 125 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 125 + ], + "type": "text", + "content": "Table 4. Effect of KDE bandwidth " + }, + { + "bbox": [ + 313, + 70, + 555, + 125 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 70, + 555, + 125 + ], + "type": "text", + "content": " on attack performance. We examine how KDE bandwidth impacts our density-guided attack. A moderate bandwidth (" + }, + { + "bbox": [ + 313, + 70, + 555, + 125 + ], + "type": "inline_equation", + "content": "h = 7.5" + }, + { + "bbox": [ + 313, + 70, + 555, + 125 + ], + "type": "text", + "content": ") achieves the best balance, maximizing effectiveness on poisoned views while preserving quality in innocent views." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 126, + 555, + 205 + ], + "lines": [ + { + "bbox": [ + 315, + 126, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 315, + 126, + 555, + 205 + ], + "type": "table", + "html": "
Bandwidth hV-ILLUSORYV-TEST
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
0.127.000.8110.37327.830.8050.286
2.526.920.8090.37527.810.8050.286
5.026.950.8110.37527.250.7860.297
7.527.040.8130.36927.760.8050.286
10.026.890.8070.38027.720.8050.286
", + "image_path": "c6b58f940e43952e32a55b066d2b90e2fc351751ab42bcdf50e55825f576e713.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 271, + 554, + 348 + ], + "blocks": [ + { + "bbox": [ + 313, + 216, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 313, + 216, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 555, + 270 + ], + "type": "text", + "content": "Table 5. Effect of noise scheduling parameters. We analyze initial noise strength " + }, + { + "bbox": [ + 313, + 216, + 555, + 270 + ], + "type": "inline_equation", + "content": "\\sigma_0" + }, + { + "bbox": [ + 313, + 216, + 555, + 270 + ], + "type": "text", + "content": " and decay strategies. Higher initial noise " + }, + { + "bbox": [ + 313, + 216, + 555, + 270 + ], + "type": "inline_equation", + "content": "(\\sigma_0 = 100)" + }, + { + "bbox": [ + 313, + 216, + 555, + 270 + ], + "type": "text", + "content": " with linear decay provides the best balance, maximizing illusory quality (V-ILLUSORY) while preserving fidelity in innocent views (V-TEST)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 271, + 554, + 348 + ], + "lines": [ + { + "bbox": [ + 315, + 271, + 554, + 348 + ], + "spans": [ + { + "bbox": [ + 315, + 271, + 554, + 348 + ], + "type": "table", + "html": "
Initial noise σ0Decay strategyV-ILLUSORYV-TEST
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
30Linear26.470.7950.39828.430.8510.203
30Cosine26.700.8010.38828.300.8450.213
30Square root26.620.7990.39328.270.8450.212
100Linear27.040.8130.36927.760.8050.286
100Cosine26.930.8120.37326.960.7710.315
100Square root26.900.8130.37326.810.7670.319
", + "image_path": "0932271001648d75d8532364bdf2fd18a3b8598900964ba0049090988e445954.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 368, + 423, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 423, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 423, + 380 + ], + "type": "text", + "content": "4.3. Multi-view Attack" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 385, + 556, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 556, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 556, + 468 + ], + "type": "text", + "content": "In realistic scenarios, attackers may need to embed illusory objects into multiple viewpoints simultaneously. Unlike single-view attacks, multi-view poisoning balances multiple objectives while preserving scene consistency and fidelity. We rigorously evaluate our approach using the Mip-NeRF 360 dataset [5], poisoned views at " + }, + { + "bbox": [ + 313, + 385, + 556, + 468 + ], + "type": "inline_equation", + "content": "0^{\\circ}" + }, + { + "bbox": [ + 313, + 385, + 556, + 468 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 385, + 556, + 468 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 313, + 385, + 556, + 468 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 385, + 556, + 468 + ], + "type": "inline_equation", + "content": "180^{\\circ}" + }, + { + "bbox": [ + 313, + 385, + 556, + 468 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 385, + 556, + 468 + ], + "type": "inline_equation", + "content": "270^{\\circ}" + }, + { + "bbox": [ + 313, + 385, + 556, + 468 + ], + "type": "text", + "content": " with the median view at " + }, + { + "bbox": [ + 313, + 385, + 556, + 468 + ], + "type": "inline_equation", + "content": "0^{\\circ}" + }, + { + "bbox": [ + 313, + 385, + 556, + 468 + ], + "type": "text", + "content": " as reference." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 469, + 556, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 469, + 556, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 469, + 556, + 553 + ], + "type": "text", + "content": "We evaluate our density-guided method against baselines (IPA-NeRF [40] with Nerfacto, IPA-NeRF with Instant-NGP, IPA-Splat) on the Mip-NeRF-360 dataset under multi-view attacks (2, 3, and 4 poisoned views). Results in Tab. 3 show our approach consistently outperforms baselines, effectively embedding illusory objects (V-ILLUSORY) while minimally affecting innocent views." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 553, + 556, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 553, + 556, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 553, + 556, + 601 + ], + "type": "text", + "content": "Fig. 8 qualitatively demonstrates our method's superiority. It consistently generates clear, visually convincing illusions with minimal artifacts, significantly outperforming baselines and maintaining high quality from innocent views." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 609, + 416, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 416, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 416, + 621 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "KDE bandwidth. The KDE bandwidth significantly affects density estimation smoothness. Evaluations (Tab. 4) show medium bandwidth (" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "h = 7.5" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": ") achieves optimal balance, maximizing V-IILLUSORY effectiveness while preserving V-TEST quality. Smaller bandwidths (" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "h = 0.1" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": ") overly restrict point placement, while larger bandwidths (" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "h = 10.0" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": ") decrease attack precision." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27406" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 116, + 553, + 202 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": "Table 6. Comparison of different attack strategy combinations. We evaluate the impact of each component of our proposed method on poisoning effectiveness. Combining all strategies achieves the best results, significantly improving rendering quality of the illusory object (V-ILLUSORY) and maintaining satisfactory performance in innocent views (V-TEST), resulting in optimal Attack Success Rate (ASR). The combination of point cloud poisoning and noise scheduling is crucial for successful attacks, highlighting their complementary nature." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 116, + 553, + 202 + ], + "lines": [ + { + "bbox": [ + 57, + 116, + 553, + 202 + ], + "spans": [ + { + "bbox": [ + 57, + 116, + 553, + 202 + ], + "type": "table", + "html": "
Poisoned View GT ReplacementDensity-Guided Point Cloud AttackView Consistency Disruption AttackV-ILLUSORYV-TESTASR (PSNR)
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓V-ILLUSORY > 25 & V-TEST drop ≤ 3
✓13.220.5210.73029.460.8840.1640/7
✓✓26.010.7750.42729.400.8830.1646/7
✓✓13.310.5220.74727.790.8050.2860/7
✓✓✓27.040.8130.36927.760.8050.2867/7
", + "image_path": "d7d858e236346b6c42ebfa1bfe2fe9de31a8208e8b3ff9d77e0e9ba42b357fe1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 57, + 220, + 105, + 290 + ], + "blocks": [ + { + "bbox": [ + 57, + 220, + 105, + 290 + ], + "lines": [ + { + "bbox": [ + 57, + 220, + 105, + 290 + ], + "spans": [ + { + "bbox": [ + 57, + 220, + 105, + 290 + ], + "type": "image", + "image_path": "4b276a6bd6b2d176ef1a317d32aa75ab3193898add151222c5904bad82e9ba7a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 105, + 220, + 152, + 291 + ], + "blocks": [ + { + "bbox": [ + 105, + 220, + 152, + 291 + ], + "lines": [ + { + "bbox": [ + 105, + 220, + 152, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 152, + 291 + ], + "type": "image", + "image_path": "6fc2c2107f298b3b032bd4ef4099fd40ab1b1cf16a1a32bba6d90ad25ba289ad.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 293, + 296, + 360 + ], + "lines": [ + { + "bbox": [ + 55, + 293, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 296, + 360 + ], + "type": "text", + "content": "Figure 9. Qualitative analysis of attack component combinations. We compare three attack strategies: (1) direct image poisoning, (2) density-guided point cloud poisoning, and (3) multi-view consistency disruption. Combining all three achieves the most realistic illusions across various scenes from the Mip-NeRF 360 [5] dataset, highlighting their complementary effectiveness." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 153, + 220, + 199, + 292 + ], + "blocks": [ + { + "bbox": [ + 153, + 220, + 199, + 292 + ], + "lines": [ + { + "bbox": [ + 153, + 220, + 199, + 292 + ], + "spans": [ + { + "bbox": [ + 153, + 220, + 199, + 292 + ], + "type": "image", + "image_path": "53b147586ae134144335c0f8a7bdae34a97b6c366f5b01e6bb12401f759aac53.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 200, + 220, + 246, + 292 + ], + "blocks": [ + { + "bbox": [ + 200, + 220, + 246, + 292 + ], + "lines": [ + { + "bbox": [ + 200, + 220, + 246, + 292 + ], + "spans": [ + { + "bbox": [ + 200, + 220, + 246, + 292 + ], + "type": "image", + "image_path": "620656d3e6f56d3dc5206c331e6833a2199d73e123ebd216af6a2139854ebb56.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 246, + 220, + 293, + 292 + ], + "blocks": [ + { + "bbox": [ + 246, + 220, + 293, + 292 + ], + "lines": [ + { + "bbox": [ + 246, + 220, + 293, + 292 + ], + "spans": [ + { + "bbox": [ + 246, + 220, + 293, + 292 + ], + "type": "image", + "image_path": "8a678fb1441cd47f0ee305568c5e07268231260746b7c44e1d66f67ab2ca1c02.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 384, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 296, + 456 + ], + "type": "text", + "content": "Noise scheduling. We evaluated noise scheduling strategies, varying initial noise intensities " + }, + { + "bbox": [ + 55, + 384, + 296, + 456 + ], + "type": "inline_equation", + "content": "(\\sigma_0)" + }, + { + "bbox": [ + 55, + 384, + 296, + 456 + ], + "type": "text", + "content": " and decay rates (linear, cosine, and square root), summarized in Tab. 5. Higher initial noise " + }, + { + "bbox": [ + 55, + 384, + 296, + 456 + ], + "type": "inline_equation", + "content": "(\\sigma_0 = 100)" + }, + { + "bbox": [ + 55, + 384, + 296, + 456 + ], + "type": "text", + "content": " with slower linear decay achieved optimal balance, greatly enhancing attack effectiveness with moderate impact on innocent views." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "text", + "content": "Attack components. We analyzed combinations of direct target-view image poisoning, density-guided point cloud poisoning (KDE-based), and noise-based view consistency disruption. Quantitative results (Tab. 6) show that direct image poisoning alone is ineffective. Combining image poisoning with density-guided poisoning notably improves outcomes. Integrating all three components achieves the best results, embedding robust illusions and preserving rendering quality. Qualitative results (Fig. 9) visually confirm the superior clarity and effectiveness of this combined approach." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 596, + 128, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 596, + 128, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 596, + 128, + 608 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 617, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 297, + 713 + ], + "type": "text", + "content": "We presented a density-guided poisoning method for 3DGS, strategically injecting illusory objects and disrupting multiview consistency via adaptive noise. Experiments show our approach outperforms existing baselines, effectively embedding convincing illusions with minimal impact on innocent views. Our work highlights critical vulnerabilities in 3D representation models, providing a robust framework for future security research." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 220, + 552, + 311 + ], + "blocks": [ + { + "bbox": [ + 317, + 220, + 552, + 311 + ], + "lines": [ + { + "bbox": [ + 317, + 220, + 552, + 311 + ], + "spans": [ + { + "bbox": [ + 317, + 220, + 552, + 311 + ], + "type": "image", + "image_path": "ac6904c291ae49c4d9122d3a419964a36f6b548d7c42c759b3381d4143a436e3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 312, + 552, + 397 + ], + "blocks": [ + { + "bbox": [ + 317, + 312, + 552, + 397 + ], + "lines": [ + { + "bbox": [ + 317, + 312, + 552, + 397 + ], + "spans": [ + { + "bbox": [ + 317, + 312, + 552, + 397 + ], + "type": "image", + "image_path": "3788b58fac4075bf70ae621fba7e0fffc1ef18f7101bbbb3ddce44e69662445b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 399, + 555, + 465 + ], + "lines": [ + { + "bbox": [ + 313, + 399, + 555, + 465 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 555, + 465 + ], + "type": "text", + "content": "Figure 10. Effect on neighboring views (\"counter\" scene, Mip-NeRF 360 [5]). Clockwise and counterclockwise shifts from the attack view (0 degrees) up to 20 degrees show PSNR between clean and poisoned renderings. IPA-NeRF significantly lowers PSNR across most angles, whereas our method mainly impacts views within five degrees, preserving quality beyond this range." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 316, + 477, + 369, + 578 + ], + "blocks": [ + { + "bbox": [ + 316, + 477, + 369, + 578 + ], + "lines": [ + { + "bbox": [ + 316, + 477, + 369, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 477, + 369, + 578 + ], + "type": "image", + "image_path": "e89b3d20713e0a837135cbfe406baeacb12f7f74858763e36098b0e585220c9e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 580, + 555, + 624 + ], + "lines": [ + { + "bbox": [ + 313, + 580, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 555, + 624 + ], + "type": "text", + "content": "Figure 11. Visualization of evaluation protocol (\"grass\" scene, Free [94]). Our method achieves clearly visible illusory objects in EASY and MEDIAN scenarios and maintains robust performance even under challenging HARD conditions." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 369, + 477, + 414, + 578 + ], + "blocks": [ + { + "bbox": [ + 369, + 477, + 414, + 578 + ], + "lines": [ + { + "bbox": [ + 369, + 477, + 414, + 578 + ], + "spans": [ + { + "bbox": [ + 369, + 477, + 414, + 578 + ], + "type": "image", + "image_path": "bf275cad4ceec9cd93c91d11c0035c166c80f75c700c97ad40c0c47051b5b770.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 414, + 477, + 459, + 578 + ], + "blocks": [ + { + "bbox": [ + 414, + 477, + 459, + 578 + ], + "lines": [ + { + "bbox": [ + 414, + 477, + 459, + 578 + ], + "spans": [ + { + "bbox": [ + 414, + 477, + 459, + 578 + ], + "type": "image", + "image_path": "b6d263976ec595f19403e8f3ec1eddeab024b21d91cfbf64cb5019f842d3bb3a.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 460, + 477, + 506, + 573 + ], + "blocks": [ + { + "bbox": [ + 460, + 477, + 506, + 573 + ], + "lines": [ + { + "bbox": [ + 460, + 477, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 460, + 477, + 506, + 573 + ], + "type": "image", + "image_path": "408582928a04e3fb6d0f5f2ecd4b05fcee8a7182fae8b9ede3abda540aca5aa7.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 507, + 477, + 553, + 573 + ], + "blocks": [ + { + "bbox": [ + 507, + 477, + 553, + 573 + ], + "lines": [ + { + "bbox": [ + 507, + 477, + 553, + 573 + ], + "spans": [ + { + "bbox": [ + 507, + 477, + 553, + 573 + ], + "type": "image", + "image_path": "c652c9b10ac033be566f40d72f80016cff636c63aba390e6c2fbd92c4a417eb7.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 646, + 555, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 646, + 555, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 646, + 555, + 704 + ], + "type": "text", + "content": "Limitations. Our method struggles with scenes having highly overlapping views or complex camera trajectories due to 3DGS's strict multi-view consistency. Future work should address this balance between consistency and attack effectiveness." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27407" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 167 + ], + "type": "text", + "content": "Acknowledgements. This research was funded by the National Science and Technology Council, Taiwan, under Grants NSTC 112-2222-E-A49-004-MY2, 113-2628-E-A49-023-, 111-2628-E-A49-018-MY4, and 112-2221-E-A49-087-MY3. The authors are grateful to Google, NVIDIA, and MediaTek Inc. for their generous donations. Yu-Lun Liu acknowledges the Yushan Young Fellow Program by the MOE in Taiwan." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 178, + 115, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 178, + 115, + 190 + ], + "spans": [ + { + "bbox": [ + 56, + 178, + 115, + 190 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 198, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 198, + 297, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 198, + 297, + 232 + ], + "spans": [ + { + "bbox": [ + 66, + 198, + 297, + 232 + ], + "type": "text", + "content": "[1] Naveed Akhtar and Ajmal Mian. Threat of adversarial attacks on deep learning in computer vision: A survey. IEEE Access, 2018. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 232, + 296, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 232, + 296, + 276 + ], + "spans": [ + { + "bbox": [ + 66, + 232, + 296, + 276 + ], + "type": "text", + "content": "[2] Maksym Andriushchenko, Francesco Croce, Nicolas Flammarion, and Matthias Hein. Square attack: a query-efficient black-box adversarial attack via random search. In European Conference on Computer Vision (ECCV), 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 277, + 295, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 277, + 295, + 321 + ], + "spans": [ + { + "bbox": [ + 66, + 277, + 295, + 321 + ], + "type": "text", + "content": "[3] Anurag Arnab, Ondrej Miksik, and Philip HS Torr. On the robustness of semantic segmentation models to adversarial attacks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 322, + 296, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 322, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 66, + 322, + 296, + 376 + ], + "type": "text", + "content": "[4] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 378, + 296, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 378, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 66, + 378, + 296, + 432 + ], + "type": "text", + "content": "[5] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 5, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 434, + 296, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 434, + 296, + 466 + ], + "spans": [ + { + "bbox": [ + 66, + 434, + 296, + 466 + ], + "type": "text", + "content": "[6] Arjun Nitin Bhagoji, Warren He, Bo Li, and Dawn Song. Exploring the space of black-box attacks on deep neural networks. *ArXiv:1712.09491*, 2017. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 468, + 296, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 468, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 66, + 468, + 296, + 499 + ], + "type": "text", + "content": "[7] Battista Biggio, Blaine Nelson, and Pavel Laskov. Poisoning attacks against support vector machines. ArXiv:1206.6389, 2012. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 501, + 294, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 501, + 294, + 534 + ], + "spans": [ + { + "bbox": [ + 66, + 501, + 294, + 534 + ], + "type": "text", + "content": "[8] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In IEEE Symposium on Security and Privacy (S&P), 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 66, + 535, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 535, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 66, + 535, + 296, + 578 + ], + "type": "text", + "content": "[9] Anirban Chakraborty, Manaar Alam, Vishal Dey, Anupam Chattopadhyay, and Debdeep Mukhopadhyay. Adversarial attacks and defences: A survey. ArXiv:1810.00069, 2018. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 580, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 580, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 62, + 580, + 296, + 645 + ], + "type": "text", + "content": "[10] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 647, + 296, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 296, + 690 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 296, + 690 + ], + "type": "text", + "content": "[11] Bo-Yu Chen, Wei-Chen Chiu, and Yu-Lun Liu. Improving robustness for joint optimization of camera pose and decomposed low-rank tensorial radiance fields. In AAAI Conference on Artificial Intelligence (AAAI), 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 691, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 691, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 62, + 691, + 296, + 714 + ], + "type": "text", + "content": "[12] Jianbo Chen, Michael I Jordan, and Martin J Wainwright. Hopskipjumpattack: A query-efficient decision-based attack." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 339, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 555, + 95 + ], + "type": "text", + "content": "In IEEE Symposium on Security and Privacy (S&P), 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 96, + 555, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 96, + 555, + 151 + ], + "spans": [ + { + "bbox": [ + 320, + 96, + 555, + 151 + ], + "type": "text", + "content": "[13] Pin-Yu Chen, Huan Zhang, Yash Sharma, Jinfeng Yi, and Cho-Jui Hsieh. Zoo: Zeroth order optimization based blackbox attacks to deep neural networks without training substitute models. In ACM Workshop on Artificial Intelligence and Security (AISec), 2017. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 152, + 555, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 152, + 555, + 186 + ], + "spans": [ + { + "bbox": [ + 320, + 152, + 555, + 186 + ], + "type": "text", + "content": "[14] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. ArXiv:1712.05526, 2017. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 186, + 555, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 186, + 555, + 219 + ], + "spans": [ + { + "bbox": [ + 320, + 186, + 555, + 219 + ], + "type": "text", + "content": "[15] Zilong Chen, Feng Wang, Yikai Wang, and Huaping Liu. Text-to-3d using gaussian splatting. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 220, + 555, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 220, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 320, + 220, + 555, + 274 + ], + "type": "text", + "content": "[16] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. In International Conference on Machine Learning (ICML), 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 276, + 555, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 276, + 555, + 320 + ], + "spans": [ + { + "bbox": [ + 320, + 276, + 555, + 320 + ], + "type": "text", + "content": "[17] Gabriela F Cretu, Angelos Stavrou, Michael E Locasto, Salvatore J Stolfo, and Angelos D Keromytis. Casting out demons: Sanitizing training data for anomaly sensors. In IEEE Symposium on Security and Privacy (S&P), 2008. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 321, + 555, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 321, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 320, + 321, + 555, + 365 + ], + "type": "text", + "content": "[18] Zeyu Dai, Shengcai Liu, Qing Li, and Ke Tang. Saliency attack: Towards imperceptible black-box adversarial attack. ACM Transactions on Intelligent Systems and Technology (TIST), 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 365, + 554, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 365, + 554, + 410 + ], + "spans": [ + { + "bbox": [ + 320, + 365, + 554, + 410 + ], + "type": "text", + "content": "[19] Ilias Diakonikolas, Gautam Kamath, Daniel Kane, Jerry Li, Jacob Steinhardt, and Alistair Stewart. Sever: A robust meta-algorithm for stochastic optimization. In International Conference on Machine Learning (ICML), 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 411, + 555, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 411, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 320, + 411, + 555, + 455 + ], + "type": "text", + "content": "[20] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading defenses to transferable adversarial examples by translation-invariant attacks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 456, + 555, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 456, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 320, + 456, + 555, + 510 + ], + "type": "text", + "content": "[21] Cheng-De Fan, Chen-Wei Chang, Yi-Ruei Liu, Jie-Ying Lee, Jiun-Long Huang, Yu-Chee Tseng, and Yu-Lun Liu. Spectromotion: Dynamic 3d reconstruction of specular scenes. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 512, + 555, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 512, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 320, + 512, + 555, + 544 + ], + "type": "text", + "content": "[22] Guangchi Fang and Bing Wang. Mini-splatting: Representing scenes with a constrained number of gaussians. In European Conference on Computer Vision (ECCV), 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 545, + 555, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 545, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 320, + 545, + 555, + 589 + ], + "type": "text", + "content": "[23] Minghong Fang, Neil Zhenqiang Gong, and Jia Liu. Influence function based data poisoning attacks to top-n recommender systems. In The Web Conference (WWW), 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 590, + 554, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 590, + 554, + 634 + ], + "spans": [ + { + "bbox": [ + 320, + 590, + 554, + 634 + ], + "type": "text", + "content": "[24] John Flynn, Ivan Neulander, James Philbin, and Noah Snavely. Deepstereo: Learning to predict new views from the world's imagery. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 635, + 555, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 635, + 555, + 680 + ], + "spans": [ + { + "bbox": [ + 320, + 635, + 555, + 680 + ], + "type": "text", + "content": "[25] Yonggan Fu, Ye Yuan, Souvik Kundu, Shang Wu, Shun-yao Zhang, and Yingyan Lin. Nerfool: Uncovering the vulnerability of generalizable neural radiance fields against adversarial perturbations. ArXiv:2306.06359, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 681, + 555, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 681, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 320, + 681, + 555, + 714 + ], + "type": "text", + "content": "[26] Micah Goldblum, Dimitris Tsipras, Chulin Xie, Xinyun Chen, Avi Schwarzschild, Dawn Song, Aleksander Madry, Bo Li, and Tom Goldstein. Dataset security for machine" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27408" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 81, + 73, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 296, + 106 + ], + "type": "text", + "content": "learning: Data poisoning, backdoor attacks, and defenses. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 106, + 296, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 106, + 296, + 138 + ], + "spans": [ + { + "bbox": [ + 61, + 106, + 296, + 138 + ], + "type": "text", + "content": "[27] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. ArXiv:1412.6572, 2014. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 140, + 296, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 140, + 296, + 160 + ], + "spans": [ + { + "bbox": [ + 61, + 140, + 296, + 160 + ], + "type": "text", + "content": "[28] Markus Gross and Hanspeter Pfister. Point-based graphics. Elsevier, 2011. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 162, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 162, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 61, + 162, + 294, + 194 + ], + "type": "text", + "content": "[29] Tianyu Gu, Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Evaluating backdooring attacks on deep neural networks. IEEE Access, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 195, + 296, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 195, + 296, + 237 + ], + "spans": [ + { + "bbox": [ + 61, + 195, + 296, + 237 + ], + "type": "text", + "content": "[30] Chuan Guo, Jacob Gardner, Yurong You, Andrew Gordon Wilson, and Kilian Weinberger. Simple black-box adversarial attacks. In International Conference on Machine Learning (ICML), 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 239, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 239, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 61, + 239, + 294, + 282 + ], + "type": "text", + "content": "[31] Abdullah Hamdi, Sara Rojas, Ali Thabet, and Bernard Ghanem. Advpc: Transferable adversarial perturbations on 3d point clouds. In European Conference on Computer Vision (ECCV), 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 284, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 284, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 61, + 284, + 296, + 316 + ], + "type": "text", + "content": "[32] András Horváth and Csaba M Józsa. Targeted adversarial attacks on generalizable neural radiance fields. In International Conference on Computer Vision (ICCV), 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 317, + 296, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 317, + 296, + 381 + ], + "spans": [ + { + "bbox": [ + 61, + 317, + 296, + 381 + ], + "type": "text", + "content": "[33] Hao-Yu Hou, Chia-Chi Hsu, Yu-Chen Huang, Mu-Yi Shen, Wei-Fang Sun, Cheng Sun, Chia-Che Chang, Yu-Lun Liu, and Chun-Yi Lee. 3d gaussian splatting with grouped uncertainty for unconstrained images. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 383, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 383, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 61, + 383, + 294, + 415 + ], + "type": "text", + "content": "[34] Hanxun Huang, Xingjun Ma, Sarah Monazam Erfani, James Bailey, and Yisen Wang. Unlearnable examples: Making personal data unexploitable. ArXiv:2101.04898, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 416, + 296, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 416, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 61, + 416, + 296, + 470 + ], + "type": "text", + "content": "[35] Xiufeng Huang, Ruiqi Li, Yiu-ming Cheung, Ka Chun Cheung, Simon See, and Renjie Wan. Gaussianmarker: Uncertainty-aware copyright protection of 3d gaussian splatting. Advances in Neural Information Processing Systems (NeurIPS), 2025. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 472, + 296, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 472, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 61, + 472, + 296, + 502 + ], + "type": "text", + "content": "[36] Zhichao Huang and Tong Zhang. Black-box adversarial attack with transferable model-based embedding. ArXiv:1911.07140, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 61, + 504, + 296, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 504, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 61, + 504, + 296, + 558 + ], + "type": "text", + "content": "[37] Matthew Jagielski, Alina Oprea, Battista Biggio, Chang Liu, Cristina Nita-Rotaru, and Bo Li. Manipulating machine learning: Poisoning attacks and countermeasures for regression learning. In IEEE Symposium on Security and Privacy (S&P), 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 61, + 559, + 296, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 559, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 61, + 559, + 296, + 602 + ], + "type": "text", + "content": "[38] Youngdong Jang, Hyunjie Park, Feng Yang, Heeju Ko, Euijin Choo, and Sangpil Kim. 3d-gsw: 3d gaussian splatting watermark for protecting copyrights in radiance fields. ArXiv:2409.13222, 2024. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 61, + 604, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 604, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 61, + 604, + 296, + 635 + ], + "type": "text", + "content": "[39] Yujie Ji, Xinyang Zhang, and Ting Wang. Backdoor attacks against learning systems. In IEEE Conference on Communications and Network Security (CNS), 2017. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 61, + 636, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 636, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 61, + 636, + 296, + 679 + ], + "type": "text", + "content": "[40] Wenxiang Jiang, Hanwei Zhang, Shuo Zhao, Zhongwen Guo, and Hao Wang. Ipa-nerf: Illusory poisoning attack against neural radiance fields. In European Conference on Artificial Intelligence (ECAI), 2024. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 61, + 681, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 681, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 61, + 681, + 296, + 712 + ], + "type": "text", + "content": "[41] James T Kajiya and Brian P Von Herzen. Ray tracing volume densities. ACM SIGGRAPH Computer Graphics (SIGGRAPH), 1984. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 320, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 555, + 116 + ], + "type": "text", + "content": "[42] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 2023. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 118, + 555, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 118, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 320, + 118, + 555, + 171 + ], + "type": "text", + "content": "[43] Michael Kern, Christoph Neuhauser, Torben Maack, Mengjiao Han, Will Usher, and Rüdiger Westermann. A comparison of rendering techniques for 3d line sets with transparency. IEEE Transactions on Visualization and Computer Graphics (TVCG), 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 173, + 555, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 173, + 555, + 215 + ], + "spans": [ + { + "bbox": [ + 320, + 173, + 555, + 215 + ], + "type": "text", + "content": "[44] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (TOG), 2017. 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 217, + 554, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 217, + 554, + 259 + ], + "spans": [ + { + "bbox": [ + 320, + 217, + 554, + 259 + ], + "type": "text", + "content": "[45] Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. Hugs: Human gaussian splats. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 261, + 555, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 261, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 320, + 261, + 555, + 293 + ], + "type": "text", + "content": "[46] Pang Wei Koh, Jacob Steinhardt, and Percy Liang. Stronger data poisoning attacks break data sanitization defenses. Machine Learning, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 294, + 555, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 294, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 320, + 294, + 555, + 336 + ], + "type": "text", + "content": "[47] Alexey Kurakin, Ian J Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In Artificial intelligence safety and security. Chapman and Hall/CRC, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 338, + 555, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 338, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 320, + 338, + 555, + 369 + ], + "type": "text", + "content": "[48] Keita Kurita, Paul Michel, and Graham Neubig. Weight poisoning attacks on pre-trained models. ArXiv:2004.06660, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 372, + 555, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 372, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 320, + 372, + 555, + 415 + ], + "type": "text", + "content": "[49] Jiahui Lei, Yufu Wang, Georgios Pavlakos, Lingjie Liu, and Kostas Daniilidis. Gart: Gaussian articulated template models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 416, + 555, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 416, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 320, + 416, + 555, + 458 + ], + "type": "text", + "content": "[50] Bo Li, Yining Wang, Aarti Singh, and Yevgeniy Vorobeychik. Data poisoning attacks on factorization-based collaborative filtering. Advances in Neural Information Processing Systems (NeurIPS), 2016. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 460, + 554, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 460, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 320, + 460, + 554, + 514 + ], + "type": "text", + "content": "[51] Chenxin Li, Hengyu Liu, Zhiwen Fan, Wuyang Li, Yifan Liu, Panwang Pan, and Yixuan Yuan. Instantsplamp: Fast and generalizable stenography framework for generative gaussian splatting. In International Conference on Learning Representations (ICLR), 2025. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 515, + 555, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 515, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 320, + 515, + 555, + 568 + ], + "type": "text", + "content": "[52] Shaofeng Li, Minhui Xue, Benjamin Zi Hao Zhao, Haojin Zhu, and Xinpeng Zhang. Invisible backdoor attacks on deep neural networks via steganography and regularization. IEEE Transactions on Dependable and Secure Computing (TDSC), 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 570, + 555, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 570, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 320, + 570, + 555, + 602 + ], + "type": "text", + "content": "[53] Yufei Li, Zexin Li, Yingfan Gao, and Cong Liu. White-box multi-objective adversarial attack on dialogue generation. ArXiv:2305.03655, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 604, + 555, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 604, + 555, + 657 + ], + "spans": [ + { + "bbox": [ + 320, + 604, + 555, + 657 + ], + "type": "text", + "content": "[54] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation from diffusion models via adversarial examples. ArXiv:2302.04578, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 320, + 658, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 658, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 320, + 658, + 555, + 712 + ], + "type": "text", + "content": "[55] Chin-Yang Lin, Chung-Ho Wu, Chang-Han Yeh, Shih-Han Yen, Cheng Sun, and Yu-Lun Liu. Frugalnerf: Fast convergence for extreme few-shot novel view synthesis without learned priors. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27409" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 72, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 297, + 116 + ], + "type": "text", + "content": "[56] Daniel Liu, Ronald Yu, and Hao Su. Extending adversarial attacks and defenses to deep 3d point cloud classifiers. In IEEE International Conference on Image Processing (ICIP), 2019. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 118, + 296, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 296, + 151 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 296, + 151 + ], + "type": "text", + "content": "[57] Xinhang Liu, Yu-Wing Tai, and Chi-Keung Tang. Clean-nerf: Reformulating nerf to account for view-dependent observations. ArXiv:2303.14707, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 153, + 296, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 153, + 296, + 186 + ], + "spans": [ + { + "bbox": [ + 62, + 153, + 296, + 186 + ], + "type": "text", + "content": "[58] Yanpei Liu, Xinyun Chen, Chang Liu, and Dawn Song. Delving into transferable adversarial examples and black-box attacks. ArXiv:1611.02770, 2016. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 188, + 296, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 188, + 296, + 232 + ], + "spans": [ + { + "bbox": [ + 62, + 188, + 296, + 232 + ], + "type": "text", + "content": "[59] Yingqi Liu, Shiqing Ma, Yousra Aafer, Wen-Chuan Lee, Juan Zhai, Weihang Wang, and Xiangyu Zhang. Trojaning attack on neural networks. In Network and Distributed System Security Symposium (NDSS), 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 234, + 296, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 234, + 296, + 288 + ], + "spans": [ + { + "bbox": [ + 62, + 234, + 296, + 288 + ], + "type": "text", + "content": "[60] Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Johannes Kopf, and Jia-Bin Huang. Robust dynamic radiance fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 291, + 296, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 296, + 323 + ], + "type": "text", + "content": "[61] Zihao Liu, Tianhao Wang, Mengdi Huai, and Chenglin Miao. Backdoor attacks via machine unlearning. In AAAI Conference on Artificial Intelligence (AAAI), 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 326, + 296, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 326, + 296, + 358 + ], + "spans": [ + { + "bbox": [ + 62, + 326, + 296, + 358 + ], + "type": "text", + "content": "[62] Jiahao Lu, Yifan Zhang, Qiuhong Shen, Xinchao Wang, and Shuicheng Yan. Poison-splat: Computation cost attack on 3d gaussian splatting. ArXiv:2410.08190, 2024. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 361, + 296, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 361, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 62, + 361, + 296, + 403 + ], + "type": "text", + "content": "[63] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In International Conference on 3D Vision (3DV), 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 406, + 296, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 406, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 62, + 406, + 296, + 449 + ], + "type": "text", + "content": "[64] Ziyuan Luo, Qing Guo, Ka Chun Cheung, Simon See, and Renjie Wan. Copyrnerf: Protecting the copyright of neural radiance fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 452, + 296, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 452, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 62, + 452, + 296, + 483 + ], + "type": "text", + "content": "[65] Yuzhe Ma, Xiaojin Zhu, and Justin Hsu. Data poisoning against differentially-private learners: Attacks and defenses. ArXiv:1903.09860, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 487, + 296, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 487, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 62, + 487, + 296, + 529 + ], + "type": "text", + "content": "[66] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. ArXiv:1706.06083, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 533, + 296, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 296, + 586 + ], + "type": "text", + "content": "[67] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 589, + 296, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 589, + 296, + 621 + ], + "spans": [ + { + "bbox": [ + 62, + 589, + 296, + 621 + ], + "type": "text", + "content": "[68] Nelson Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics (TVCG), 1995. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 624, + 296, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 296, + 677 + ], + "type": "text", + "content": "[69] Andreas Meuleman, Yu-Lun Liu, Chen Gao, Jia-Bin Huang, Changil Kim, Min H Kim, and Johannes Kopf. Progressively optimized local radiance fields for robust view synthesis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 680, + 296, + 713 + ], + "type": "text", + "content": "[70] Jian-Xun Mi, Xu-Dong Wang, Li-Fang Zhou, and Kun Cheng. Adversarial examples based on object detection tasks: A survey. Neurocomputing, 2023. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 320, + 73, + 555, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 555, + 126 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 555, + 126 + ], + "type": "text", + "content": "[71] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 129, + 555, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 129, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 320, + 129, + 555, + 173 + ], + "type": "text", + "content": "[72] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 175, + 555, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 175, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 320, + 175, + 555, + 217 + ], + "type": "text", + "content": "[73] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (TOG), 2022. 3, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 220, + 555, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 220, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 320, + 220, + 555, + 274 + ], + "type": "text", + "content": "[74] Luis Muñoz-González, Battista Biggio, Ambra Demontis, Andrea Paudice, Vasin Wongrassamee, Emil C Lupu, and Fabio Roli. Towards poisoning of deep learning algorithms with back-gradient optimization. In ACM Workshop on Artificial Intelligence and Security (AISec), 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 276, + 555, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 276, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 320, + 276, + 555, + 330 + ], + "type": "text", + "content": "[75] Nicolas Papernot, Patrick McDaniel, Ian Goodfellow, Somesh Jha, Z Berkay Celik, and Ananthram Swami. Practical black-box attacks against machine learning. In ACM Asia Conference on Computer and Communications Security (ASIACCS), 2017. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 332, + 555, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 332, + 555, + 385 + ], + "spans": [ + { + "bbox": [ + 320, + 332, + 555, + 385 + ], + "type": "text", + "content": "[76] Andrea Paudice, Luis Muñoz-González, and Emil C Lupu. Label sanitization against label flipping poisoning attacks. In European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML PKDD), 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 388, + 555, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 388, + 555, + 420 + ], + "spans": [ + { + "bbox": [ + 320, + 388, + 555, + 420 + ], + "type": "text", + "content": "[77] Thomas Porter and Tom Duff. Compositing digital images. In ACM SIGGRAPH Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), 1984. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 422, + 555, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 422, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 320, + 422, + 555, + 487 + ], + "type": "text", + "content": "[78] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 489, + 555, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 489, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 320, + 489, + 555, + 521 + ], + "type": "text", + "content": "[79] Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, Andrew Ilyas, and Aleksander Madry. Raising the cost of malicious ai-powered image editing. ArXiv:2302.06588, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 523, + 555, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 523, + 555, + 577 + ], + "spans": [ + { + "bbox": [ + 320, + 523, + 555, + 577 + ], + "type": "text", + "content": "[80] Ali Shafahi, W Ronny Huang, Mahyar Najibi, Octavian Suciu, Christoph Studer, Tudor Dumitras, and Tom Goldstein. Poison frogs! targeted clean-label poisoning attacks on neural networks. Advances in Neural Information Processing Systems (NeurIPS), 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 579, + 555, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 579, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 320, + 579, + 555, + 632 + ], + "type": "text", + "content": "[81] Ilia Shumailov, Yiren Zhao, Daniel Bates, Nicolas Papernot, Robert Mullins, and Ross Anderson. Sponge examples: Energy-latency attacks on neural networks. In IEEE European symposium on security and privacy (EuroS&P), 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 635, + 555, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 635, + 555, + 678 + ], + "spans": [ + { + "bbox": [ + 320, + 635, + 555, + 678 + ], + "type": "text", + "content": "[82] Qi Song, Ziyuan Luo, Ka Chun Cheung, Simon See, and Renjie Wan. Geometry cloak: Preventing tgs-based 3d reconstruction from copyrighted images. Advances in Neural Information Processing Systems (NeurIPS), 2025. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 681, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 681, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 681, + 555, + 713 + ], + "type": "text", + "content": "[83] Jacob Steinhardt, Pang Wei W Koh, and Percy S Liang. Certified defenses for data poisoning attacks. Advances in Neural Information Processing Systems (NeurIPS), 2017. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27410" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 297, + 127 + ], + "type": "text", + "content": "[84] Chih-Hai Su, Chih-Yao Hu, Shr-Ruei Tsai, Jie-Ying Lee, Chin-Yang Lin, and Yu-Lun Liu. Boostmvsnerfs: Boosting mvs-based nerfs to generalizable view synthesis in large-scale scenes. In ACM SIGGRAPH Conference Papers (SIGGRAPH), 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 128, + 296, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 128, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 61, + 128, + 296, + 183 + ], + "type": "text", + "content": "[85] Octavian Suciu, Radu Marginean, Yigitcan Kaya, Hal Daume III, and Tudor Dumitras. When does machine learning {FAIL}? generalized transferability for evasion and poisoning attacks. In USENIX Security Symposium (USENIX Security), 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 184, + 296, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 184, + 296, + 226 + ], + "spans": [ + { + "bbox": [ + 62, + 184, + 296, + 226 + ], + "type": "text", + "content": "[86] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. ArXiv:1312.6199, 2013. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 228, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 228, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 228, + 296, + 293 + ], + "type": "text", + "content": "[87] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, et al. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH Conference Proceedings (SIGGRAPH), 2023. 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 293, + 296, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 293, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 62, + 293, + 296, + 348 + ], + "type": "text", + "content": "[88] Ayush Tewari, Ohad Fried, Justus Thies, Vincent Sitzmann, Stephen Lombardi, Kalyan Sunkavalli, Ricardo MartinBrualla, Tomas Simon, Jason Saragih, Matthias Nießner, et al. State of the art on neural rendering. In Computer Graphics Forum. Wiley Online Library, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 349, + 296, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 349, + 296, + 381 + ], + "spans": [ + { + "bbox": [ + 62, + 349, + 296, + 381 + ], + "type": "text", + "content": "[89] Brandon Tran, Jerry Li, and Aleksander Madry. Spectral signatures in backdoor attacks. Advances in Neural Information Processing Systems (NeurIPS), 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 383, + 296, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 383, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 62, + 383, + 296, + 426 + ], + "type": "text", + "content": "[90] Yu-Lin Tsai, Chia-Yi Hsu, Chia-Mu Yu, and Pin-Yu Chen. Formalizing generalization and adversarial robustness of neural networks to weight perturbations. Advances in Neural Information Processing Systems (NeurIPS), 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 426, + 296, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 426, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 62, + 426, + 296, + 458 + ], + "type": "text", + "content": "[91] Alexander Turner, Dimitris Tsipras, and Aleksander Madry. Label-consistent backdoor attacks. ArXiv:1912.02771, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 460, + 296, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 460, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 460, + 296, + 491 + ], + "type": "text", + "content": "[92] Eric Wallace, Tony Z Zhao, Shi Feng, and Sameer Singh. Concealed data poisoning attacks on nlp models. ArXiv:2010.12563, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 492, + 296, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 492, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 62, + 492, + 296, + 536 + ], + "type": "text", + "content": "[93] Chen Wang, Xian Wu, Yuan-Chen Guo, Song-Hai Zhang, Yu-Wing Tai, and Shi-Min Hu. Nerf-sr: High quality neural radiance fields using supersampling. In ACM International Conference on Multimedia (ACM MM), 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 537, + 296, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 537, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 62, + 537, + 296, + 591 + ], + "type": "text", + "content": "[94] Peng Wang, Yuan Liu, Zhaoxi Chen, Lingjie Liu, Ziwei Liu, Taku Komura, Christian Theobalt, and Wenping Wang. F2-nerf: Fast neural radiance field training with free camera trajectories. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 5, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 592, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 592, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 592, + 296, + 635 + ], + "type": "text", + "content": "[95] Yixiang Wang, Jiqiang Liu, Xiaolin Chang, Ricardo J Rodríguez, and Jianhua Wang. Di-aa: An interpretable white-box attack for fooling deep neural networks. Information Sciences, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 636, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 636, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 62, + 636, + 296, + 669 + ], + "type": "text", + "content": "[96] Chong Xiang, Charles R Qi, and Bo Li. Generating 3d adversarial point clouds. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 296, + 712 + ], + "type": "text", + "content": "[97] Chaowei Xiao, Dawei Yang, Bo Li, Jia Deng, and Mingyan Liu. Meshadv: Adversarial meshes for visual recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 611 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 320, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 555, + 116 + ], + "type": "text", + "content": "[98] Huang Xiao, Battista Biggio, Gavin Brown, Giorgio Fumera, Claudia Eckert, and Fabio Roli. Is feature selection secure against training data poisoning? In International Conference on Machine Learning (ICML), 2015. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 118, + 555, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 118, + 555, + 162 + ], + "spans": [ + { + "bbox": [ + 320, + 118, + 555, + 162 + ], + "type": "text", + "content": "[99] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Yuyin Zhou, Lingxi Xie, and Alan Yuille. Adversarial examples for semantic segmentation and object detection. In International Conference on Computer Vision (ICCV), 2017. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 555, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 555, + 206 + ], + "type": "text", + "content": "[100] Yuanwang Yang, Qiao Feng, Yu-Kun Lai, and Kun Li. Real-time 3d human reconstruction and rendering system from a single rgb camera. In ACM SIGGRAPH Asia Technical Communications (SIGGRAPH Asia). ACM, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 207, + 554, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 554, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 554, + 229 + ], + "type": "text", + "content": "[101] Linfeng Ye and Shayan Mohajer Hamidi. Thundernna: a white box adversarial attack. ArXiv:2111.12305, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 230, + 555, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 555, + 274 + ], + "type": "text", + "content": "[102] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 275, + 555, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 275, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 275, + 555, + 308 + ], + "type": "text", + "content": "[103] Abdurrahman Zeybey, Mehmet Ergezer, and Tommy Nguyen. Gaussian splatting under attack: Investigating adversarial noise in 3d objects. ArXiv:2412.02803, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 309, + 555, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 555, + 353 + ], + "type": "text", + "content": "[104] Yu-Ting Zhan, Cheng-Yuan Ho, Hebi Yang, Yi-Hsin Chen, Jui Chiu Chiang, Yu-Lun Liu, and Wen-Hsiao Peng. Cat-3dgs: A context-adaptive triplane approach to rate-distortion-optimized 3dgs compression. ArXiv:2503.00357, 2025. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 354, + 554, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 554, + 398 + ], + "type": "text", + "content": "[105] Jie Zhang, Chen Dongdong, Qidong Huang, Jing Liao, Weiming Zhang, Huamin Feng, Gang Hua, and Nenghai Yu. Poison ink: Robust and invisible backdoor attack. IEEE Transactions on Image Processing (TIP), 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 399, + 554, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 554, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 554, + 443 + ], + "type": "text", + "content": "[106] Xuanyu Zhang, Jiarui Meng, Zhipei Xu, Shuzhou Yang, Yanmin Wu, Ronggang Wang, and Jian Zhang. Securegs: Boosting the security and fidelity of 3d gaussian splatting steganography. ArXiv:2503.06118, 2025. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 444, + 555, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 555, + 475 + ], + "type": "text", + "content": "[107] Yutong Zhang, Yao Li, Yin Li, and Zhichang Guo. A review of adversarial attacks in computer vision. ArXiv:2308.07673, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 477, + 555, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 477, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 477, + 555, + 520 + ], + "type": "text", + "content": "[108] Meixi Zheng, Xuanchen Yan, Zihao Zhu, Hongrui Chen, and Baoyuan Wu. Blackboxbench: A comprehensive benchmark of black-box adversarial attacks. ArXiv:2312.16979, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 522, + 555, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 555, + 566 + ], + "type": "text", + "content": "[109] Tinghui Zhou, Shubham Tulsiani, Weilun Sun, Jitendra Malik, and Alexei A Efros. View synthesis by appearance flow. In European Conference on Computer Vision (ECCV), 2016. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 567, + 555, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 555, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 555, + 611 + ], + "type": "text", + "content": "[110] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Surface splattering. In ACM SIGGRAPH Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), 2001. 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27411" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Stealthy Backdoor Attack in Federated Learning via Adaptive Layer-wise Gradient Alignment/f7dfb2e6-c547-4b0c-bb2e-4bfc0b02a71f_content_list.json b/2025/Stealthy Backdoor Attack in Federated Learning via Adaptive Layer-wise Gradient Alignment/f7dfb2e6-c547-4b0c-bb2e-4bfc0b02a71f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..647cfb8c4a51a73a6e55b2f79f399d8cbcf2f322 --- /dev/null +++ b/2025/Stealthy Backdoor Attack in Federated Learning via Adaptive Layer-wise Gradient Alignment/f7dfb2e6-c547-4b0c-bb2e-4bfc0b02a71f_content_list.json @@ -0,0 +1,1590 @@ +[ + { + "type": "text", + "text": "Stealthy Backdoor Attack in Federated Learning via Adaptive Layer-wise Gradient Alignment", + "text_level": 1, + "bbox": [ + 251, + 130, + 746, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qingqian Yang $^{1*}$ Peishen Yan $^{2}$ Xiaoyu Wu $^{2}$ Jiaru Zhang $^{2}$ Tao Song $^{2\\dagger}$ Yang Hua $^{3}$ Hao Wang $^{4}$ Liangliang Wang $^{1\\dagger}$ Haibing Guan $^{2}$ $^{1}$ Shanghai University of Electric Power $^{2}$ Shanghai Jiao Tong University \n $^{3}$ Queen's University Belfast $^{4}$ Stevens Institute of Technology", + "bbox": [ + 138, + 202, + 854, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "y23108013@mail.shiep.edu.cn", + "bbox": [ + 377, + 277, + 620, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{peishenyan, wuxiaoyu2000, jiaruzhang, songt333, hbguan}@sjtu.edu.cn", + "bbox": [ + 199, + 294, + 797, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Y.Hua@qub.ac.uk hwang9@stevens.edu llwang@shiep.edu.cn", + "bbox": [ + 243, + 311, + 745, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 361, + 326, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The distributed nature of federated learning exposes it to significant security threats, among which backdoor attacks are one of the most prevalent. However, existing backdoor attacks face a trade-off between attack strength and stealthiness: attacks maximizing the attack strength are often detectable, while stealthier approaches significantly reduce the effectiveness of the attack itself. Both of them result in ineffective backdoor injection. In this paper, we propose an adaptive layer-wise gradient alignment strategy to effectively evade various robust defense mechanisms while preserving attack strength. Without requiring additional knowledge, we leverage the previous global update as a reference for alignment to ensure stealthiness during dynamic FL training. This fine-grained alignment strategy applies appropriate constraints to each layer, which helps significantly maintain attack strength. To demonstrate the effectiveness of our method, we conduct extensive evaluations across a wide range of datasets and networks. Our experimental results show that the proposed attack effectively bypasses eight state-of-the-art defenses and achieves high backdoor accuracy, outperforming existing attacks by up to $54.76\\%$ . Additionally, it significantly preserves attack strength and maintains robust performance across diverse scenarios, highlighting its adaptability and generalizability. Code implementation is available at https://github.com/yqqhyqq/LGA.", + "bbox": [ + 88, + 393, + 485, + 772 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 799, + 220, + 814 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Federated Learning (FL) has emerged as a promising paradigm for privacy-preserving distributed machine learn", + "bbox": [ + 89, + 824, + 483, + 854 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing, allowing multiple decentralized clients to collaboratively train a global model without directly sharing their raw data [14, 21]. FL leverages local computation on edge devices (e.g., smartphones and IoT sensors) and aggregates model updates via a central server, making it widely adopted in various real-world applications [27, 39]. However, the distributed architecture of FL systems significantly increases the potential attack surface, introducing a wide range of security threats, among which backdoor attacks are a commonly studied threat in recent research. In a backdoor attack [2, 32, 33, 35, 41, 42], adversarial clients intentionally inject a backdoor into the global model by uploading a poisoning local update. As a result, the global model behaves normally on benign inputs but misclassifies triggered inputs into predefined target labels.", + "bbox": [ + 511, + 363, + 908, + 589 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing backdoor defenses primarily focus on detecting and filtering malicious updates by leveraging information from multiple dimensions. Specifically, some methods rely on distance-based measures between benign and malicious updates [4, 5, 12, 23], while others detect inconsistencies in parameter signs [25] or exploit the spectral separability [29]. These defenses significantly restrict the feasibility of backdoor attacks.", + "bbox": [ + 511, + 594, + 908, + 714 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To counter these defenses, an effective backdoor attack depends on two key factors: attack strength and attack stealthiness. Some attacks [2] achieve high attack strength by scaling up the update, which increases their detectability and makes them easily filtered by defense mechanisms. Others [35, 42] prioritize stealthiness to evade detection but show significantly lower performance when no defense is present, highlighting their sacrifice of attack strength. These attacks essentially constrain the malicious update by discarding certain backdoor-related information [42] or using coarse-grained scaling [33]. Although constraining malicious updates helps maintain stealthiness, this inappropriate", + "bbox": [ + 511, + 719, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*The work was done during her internship at Shanghai Jiao Tong University, Shanghai Key Laboratory of Scalable Computing and Systems. † Corresponding authors.", + "bbox": [ + 89, + 862, + 485, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "29163", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "constraint leads to attack strength degradation. Therefore, balancing the attack stealthiness and attack strength with limited knowledge of the system's defenses is challenging.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the challenge, we propose a layer-wise adaptive gradient alignment strategy that dynamically adjusts the norm bound for each layer of malicious updates. This adjustment enhances their similarity to the benign update distribution in a fine-grained manner, thereby achieving both attack stealthiness and attack effectiveness. Without requiring any additional information, our approach uses the previous global update as a reference point for adaptive alignment. By doing so, the proposed strategy not only constrains the malicious update within a small norm ball, but also enables the attack to adapt to the evolving global model and training dynamics, thereby significantly improving its stealthiness. Inspired by previous observations [15] that parameters within the same layer tend to share similar gradient magnitudes, we further conduct empirical analysis revealing that the magnitudes of both malicious and benign updates vary across layers, and there is no consistent proportional relationship between them. This discrepancy highlights the need for layer-specific norm constraints. Such constraints are essential to prevent over-constraining layers with inherently smaller magnitudes, thereby preserving attack strength.", + "bbox": [ + 91, + 137, + 485, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments validate the effectiveness of our attack, demonstrating its ability to evade eight state-of-the-art defenses while maintaining a high attack success rate. Additionally, we investigate its performance across various attack scenarios, demonstrating its generalizability. We summarize our main contributions as follows:", + "bbox": [ + 89, + 454, + 483, + 542 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose an adaptive layer-wise gradient alignment strategy to effectively evade various robust defense mechanisms while preserving attack strength. Without requiring any additional knowledge, we leverage the previous global update as a reference for gradient alignment, ensuring that the attack maintains the stealthiness while adapting to the training dynamics. This fine-grained alignment strategy applies layer-specific constraints, allowing the attack to retain its strength.", + "- We conduct extensive evaluations across diverse models and datasets, demonstrating that the proposed attack successfully bypasses eight SOTA defenses and maintains the attack impact. Our backdoor attack outperforms SOTA attacks, with an improvement in backdoor accuracy by up to $54.76\\%$ . Furthermore, our attack exhibits excellent performance across various scenarios, such as different attack patterns, highlighting its generalizability." + ], + "bbox": [ + 89, + 546, + 482, + 800 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 814, + 232, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Backdoor attacks in FL. Backdoor attacks present a serious security threat in FL [16]. Typically, attackers compromise a small subset of client devices to upload poisoned updates. These updates are crafted to inject the backdoor into", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the global model, which then enables the model to obtain predetermined predictions for samples with specific trigger features. The backdoor attack process generally consists of two stages: trigger embedding and model manipulation. To successfully inject a backdoor, the attacker can carefully design either the trigger or the model update to bypass detection. Accordingly, existing methods can be divided into two categories: model-optimization attacks [2, 18, 41, 42] and trigger-optimization attacks [8, 20, 24, 33, 35, 40].", + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Model-optimization attacks [2, 3, 18, 33, 41, 42] focus on fine-tuning the model to achieve the desired attack. Scaling attack [2] amplifies malicious updates to achieve a one-shot attack. However, the abnormally large gradient magnitudes make it exposed to detection. Neurotoxin [41] projects malicious gradients onto infrequently updated neurons, improving the persistence of the attack and preventing the backdoor from being overwritten by benign updates. However, as an unconstrained backdoor attack, it's vulnerable to robust detection. The LP attack [42] adaptively injects the backdoor into backdoor-critical (BC) layers, achieving high stealthiness. However, by discarding backdoor-related information in non-BC layers, the attack suffers from reduced strength. PGD attack [33] exploits knowledge of server-side defenses to craft malicious updates that bypass detection. However, this strategy is only effective when the attacker has access to detailed information about the system. Additionally, 3DFed [18] introduces three distinct attack strategies targeting different types of defenses and utilizes an indicator for adaptive strategy adjustments. However, it relies on client collaboration.", + "bbox": [ + 511, + 227, + 906, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Trigger-optimization backdoor attacks aim to design triggers that are both subtle and adaptable, ensuring their stealthiness and effectiveness. Edge-case attack [33] selects marginal data as poison samples to circumvent detection mechanisms. DBA [35] firstly introduces distributed backdoor attacks, decomposing a global trigger into smaller, less detectable components. These approaches craft triggers at the pre-training stage and then fix them during training. Recently, a range of studies [8, 19, 20, 24, 40] dynamically optimize the trigger throughout the training process and obtain better attack results than the former ones.", + "bbox": [ + 511, + 545, + 908, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Defense strategies against backdoor attacks in FL. In FL [16], the server only receives uploaded models without access to any information about client data. As a result, the majority of defense strategies are similarity-based, aiming to detect and mitigate backdoor attacks by identifying differences between malicious and benign models.", + "bbox": [ + 511, + 733, + 906, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Among them, distance-based defenses [1, 4, 5, 9, 12, 17, 22, 23, 28, 29, 31, 34] exploit the observation that malicious models often deviate from benign ones in their distribution within high-dimensional space, making distance-based metrics an effective tool for detection. MultiKrum [4]", + "bbox": [ + 511, + 825, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "29164", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "only aggregates updates that are selected with the smallest pairwise Euclidean distances. FLTrust [5] weights upload gradients based on their cosine similarity to a trusted gradient to mitigate the influence of malicious models. The trusted gradient is obtained from a root dataset collected by the server. FLAME [23] and DeepSight [28] use pairwise cosine distance between client gradients for clustering, identifying malicious gradients as outliers and further refining the defense through additional modules. Multi-Metrics [12] dynamically adjusts the weights of three distance metrics to detect backdoor models. MESAS [17] leverages six distance metrics for detection.", + "bbox": [ + 89, + 90, + 483, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Other similarity-based defenses [13, 26, 29, 36, 38] identify malicious updates from more diverse perspectives. DnC [29] leverages singular value decomposition (SVD)-based spectral methods to remove outliers. RLR [25] detects inconsistencies in parameter signs across training rounds and adjusts learning rates accordingly to suppress backdoors. MASA [37] leverages the observation that a malicious model unlearns the main task more quickly than a benign model for detection. FedREDefense [36] relies on the finding that a malicious model is more resistant to distillation. In addition, some methods recently focus on the difference in update behavior between benign models and malicious models at the parameter level. For instance, SparseFed [26] only aggregates Top-k highest magnitude parameters because the backdoor updates are usually injected into coordinates that are unimportant for benign updates. Meanwhile, Lockdown [13] restrains local training in a subspace to further reduce the poison-coupling effect.", + "bbox": [ + 91, + 272, + 485, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 556, + 225, + 573 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Background. Federated Learning (FL) distributes the training task across multiple clients to preserve data privacy. At the beginning, a global model $\\theta_g^0$ is initialized on the server and distributed to a set of $N$ clients. In each round $t$ , a standard FL framework involves the following steps:", + "bbox": [ + 89, + 582, + 483, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "1. Local Training: Each client $i$ receives the global model $\\theta_{g}^{t - 1}$ and performs $E$ local epochs of stochastic gradient descent (SGD) on its private dataset $\\hat{\\mathcal{D}}_i$ with batch size $B$ , obtaining an updated model $\\theta_{i}^{t}$ . The client computes the model update as:", + "bbox": [ + 91, + 657, + 483, + 736 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\theta_ {i} ^ {t} = \\theta_ {i} ^ {t} - \\theta_ {g} ^ {t - 1}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 235, + 744, + 483, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and uploads it to the server.", + "bbox": [ + 109, + 772, + 294, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Global aggregation: The server randomly selects a subset of clients and aggregates their model updates $\\{\\Delta \\theta^t\\}$ to compute the global update:", + "bbox": [ + 89, + 786, + 483, + 832 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\theta_ {g} ^ {t} = \\sum_ {i \\in [ n ]} \\Delta \\theta_ {i} ^ {t}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 227, + 840, + 483, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and updates the global model as $\\theta_g^t = \\theta_g^{t - 1} + \\Delta \\theta_g^t$ . The new global model is then distributed to clients for the", + "bbox": [ + 109, + 869, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "next round. Before updating the global model, the server typically performs backdoor detection during aggregation to mitigate the impact of malicious updates.", + "bbox": [ + 532, + 90, + 906, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In FL, backdoor attacks exploit the system's decentralized nature. Specifically, an attacker, denoted as $A$ , compromises a subset of clients and injects malicious updates to implant a backdoor into the global model. We consider a continuous attack scenario, following prior studies [20, 35, 42].", + "bbox": [ + 511, + 136, + 906, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adversarial Objectives: The attacker $A$ aims to inject a backdoor into the global model while minimizing the impact on the model's performance on non-poisoned data. The goal of the backdoor injection is to manipulate the model so that it misclassifies inputs with a specific trigger pattern as an incorrect target label. We denote the malicious update as $\\Delta \\theta_{m}$ , which is trained on the poisoned dataset. By carefully crafting $\\Delta \\theta_{m}$ , the attacker can introduce the backdoor without significantly degrading the model's overall accuracy. The server's aggregation mechanism, which aggregates updates from multiple clients, may unintentionally include these poisoned updates, thereby corrupting the global model.", + "bbox": [ + 511, + 212, + 908, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adversarial Capability: The attacker $A$ has the ability to control a small subset of clients and gain access to their datasets. Additionally, the attacker can manipulate the models of these compromised clients [2, 35, 41]. The attacker does not have access to the data or models of benign clients, nor to the specific defense methods implemented by the defender.", + "bbox": [ + 511, + 393, + 908, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. We introduce an adaptive layer-wise gradient alignment to enhance stealthiness while preserving attack strength. In each round, we train the malicious update on a poisoned local dataset and iteratively align gradients to resemble benign updates before uploading them to the server. This increases the likelihood of the server aggregating the malicious update, effectively injecting the backdoor into the global model.", + "bbox": [ + 511, + 518, + 908, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adaptive reference selection. Traditionally, the Unconstrained malicious update $\\Delta \\theta_{m}$ , which is trained on the poisoned dataset including out-of-distribution backdoor samples, tends to exhibit a larger magnitude compared to benign updates [32]. To ensure that the malicious update remains similar to the distribution of benign updates, it should be carefully constrained within an appropriately small norm ball. Notably, the selection of the norm bound significantly impacts the effectiveness of the attack's concealment. Therefore, it is crucial to rigorously determine an appropriate reference point for alignment. Some defense methods, such as MultiKrum [4], select a fixed number of client updates that appear most benign in each round. However, due to data heterogeneity, such methods may inadvertently exclude genuinely benign clients whose updates deviate from the majority. To mitigate this issue, we adopt the previous global", + "bbox": [ + 511, + 657, + 910, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "29165", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/299b4ddf2a78efc4cc680d2adcb185cccf9d779d87cdda40acbe2c18330a06e3.jpg", + "image_caption": [ + "Figure 1. Layer-wise Magnitude Analysis of Benign vs. Malicious Updates in VGG19. To better visualize, we focus on weight layers and divide them into A-layers and B-layers based on $\\|\\Delta\\theta^l\\|$ : those with norms above 0.5 are A-layers; others are B-layers. The experiment uses the default setting and records $\\|\\Delta\\theta^l\\|$ every 50 rounds." + ], + "image_footnote": [], + "bbox": [ + 133, + 87, + 433, + 236 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "update $\\Delta \\theta_{g}^{t - 1}$ as a reference point. Since it is computed by aggregating multiple client updates, most of which are expected to be benign, it serves as a reasonable approximation of benign behavior.", + "bbox": [ + 89, + 318, + 483, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By aligning the malicious update with the global update, the attack remains within the expected range of benign variation, thereby enhancing its stealthiness and increasing the likelihood of successfully influencing the global model.", + "bbox": [ + 89, + 380, + 483, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Layer-wise gradient alignment. Intuitively, applying a norm constraint weakens the impact of the malicious update, which potentially leads to ineffective backdoor injection. We introduce a layer-wise scaling strategy to refine the malicious update at a more granular level to not only ensure the attack remains stealthy but also preserve its strength. A key observation shown in Figure 1 is that the layer-wise magnitudes of the malicious and benign updates do not maintain a fixed proportional relationship; instead, their patterns differ. Therefore, we align each layer individually with $\\Delta \\theta_{g}^{t - 1}$ . Specifically, we define the scaling factor $S^{t,l}$ for each layer $l$ as:", + "bbox": [ + 89, + 458, + 483, + 638 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS ^ {t, l} = \\min (1, \\frac {\\| \\Delta \\theta_ {g} ^ {t - 1 , l} \\|}{\\| \\Delta \\theta_ {m} ^ {t , l} \\|}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 637, + 483, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The scaled malicious update after each local epoch is given by:", + "bbox": [ + 89, + 676, + 482, + 705 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\theta_ {m} ^ {t, l} = S ^ {t, l} \\cdot \\Delta \\theta_ {m} ^ {t, l}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 704, + 482, + 723 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The adaptive layer-wise gradient alignment, implemented in just four lines (Lines 10-13 in Algorithm 1), extends the standard FL backdoor pipeline. This approach prevents excessive constraints on layers with inherently smaller magnitudes and constrains the update for each layer within an $\\ell 2$ norm ball, thereby preserving attack strength while enhancing stealthiness.", + "bbox": [ + 89, + 727, + 485, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 844, + 223, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our empirical study evaluates the effectiveness of the proposed backdoor attack through experiments conducted on", + "bbox": [ + 89, + 869, + 485, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 Layer-wise Gradient Alignment for Attackers \nRequire: Batch size $\\ell$ , local epochs $E$ , malicious model parameters $\\theta_{i}^{t}$ , global parameters $\\theta_{g}^{t-1}$ , poisoned dataset $\\hat{D}$ , and previous global parameter $\\theta_{g}^{t-2}$", + "bbox": [ + 513, + 90, + 906, + 154 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Procedure:", + "text_level": 1, + "bbox": [ + 514, + 154, + 596, + 166 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: // Attacker's local training:", + "2: Calculate the previous global update:", + "3: $\\Delta \\theta_{g}^{t - 1}\\gets \\theta_{g}^{t - 1} - \\theta_{g}^{t - 2}$", + "4: Initialize local model $\\theta_{i,0}^{t}\\gets \\theta_{g}^{t - 1}$", + "5: for $e = 1$ to $E$ do" + ], + "bbox": [ + 522, + 170, + 787, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "6: for each batch $b$ in $\\hat{\\mathcal{D}}$ do", + "bbox": [ + 522, + 246, + 723, + 260 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "7: Compute gradient for batch $b$ :", + "8: $\\Delta \\theta_{i,e}^{t}\\gets \\bar{\\Delta}\\theta_{i,e}^{t} + \\nabla_{\\theta}\\mathcal{L}(\\theta_{i,e - 1}^{t},\\hat{\\mathcal{D}}_{b})$", + "9: end for" + ], + "bbox": [ + 522, + 261, + 805, + 304 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "10: // Layer-wise gradient alignment", + "11: for each layer $l$ in $\\Delta \\theta_{i,e}^{t}$ do", + "12: $S_{i,e}^{t,l}\\gets min(1,\\frac{\\|\\Delta\\theta_{g}^{t - 1,l}\\|}{\\|\\Delta\\theta_{i,e}^{t,l}\\|})$", + "13: $\\theta_{i,e}^{t,l}\\gets \\Delta \\theta_{i,e}^{t,l}\\cdot S_{i,e}^{t,l} + \\theta_{g}^{t - 1,l}$", + "14: end for", + "15: end for", + "16:Upload update $\\Delta \\theta_{i}^{t}$ to server" + ], + "bbox": [ + 519, + 306, + 805, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "real-world datasets, simulating an FL environment using the NVIDIA RTX 4090 GPU.", + "bbox": [ + 511, + 458, + 906, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Experimental setup", + "text_level": 1, + "bbox": [ + 511, + 497, + 700, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets and Models. We evaluate the attack using two widely used benchmark datasets: CIFAR10 (50,000 training and 10,000 test samples across 10 classes) and CIFAR100 (50,000 training and 10,000 test samples across 100 classes). Following previous studies [41, 42], we use the VGG19 [30] network and ResNet18 network for CIFAR10, while we use the ResNet18 [11] network for CIFAR100. We simulate a non-IID data distribution following prior methods [5, 7, 42], with $q = 0.5 / 0.2$ for CIFAR10/100, respectively.", + "bbox": [ + 511, + 518, + 908, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Attack setup. In each round, we select $n = 10$ clients among $N = 100$ clients to participate in the aggregation, with the proportion of compromised clients set to $C = 0.1$ by default. The FL training runs for 300 rounds to make it converge. We train for 6 local epochs (2 for the benign clients) with a learning rate 0.1 for both CIFAR10 and CIFAR100. Additionally, we assess the effectiveness of our attack in an IID setting, as defense strategies are more likely to identify malicious models as outliers in this case [42]. The trigger is a $5 \\times 5$ pixel square located at the bottom-right corner of each image by default. We further discuss the impact of various shapes and sizes of the trigger. We evaluate our attack in the fixed-frequency attack scenario. The attacker controls a fixed number of compromised clients, which participate in FL training at a fixed frequency. Following prior", + "bbox": [ + 511, + 674, + 910, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "29166", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/664947691c5df6c6c5cbeb8f1b04e534fa265586e5a00dc775acb01354b3e307.jpg", + "table_caption": [ + "Table 1. The effectiveness of our approach against the baseline attack across various state-of-the-art (SOTA) defenses under a fixed-frequency attack setting on non-IID datasets. BA values below $10\\%$ in CIFAR10 (10 classes) and below $1\\%$ in CIFAR100 (100 classes) are highlighted in red, indicating a failed attack. The highest BA achieved in each setting is presented in bold. Our attack results are reported as $a \\pm b$ , where $a$ denotes the mean and $b$ represents the standard deviation. MA and BA unit: %. Avg indicates the average. Each result is obtained as the average over five independent runs." + ], + "table_footnote": [], + "table_body": "
DefenseVGG19 (CIFAR10)ResNet18 (CIFAR10)ResNet18 (CIFAR100)
BadNetsDBALPOursBadNetsDBALPOursBadNetsDBALPOurs
No DefenseMA80.9781.2682.3881.81±0.3477.6678.2678.4378.70±0.8263.0562.465.1764.31±0.72
Best BA98.3346.1795.3897.05±0.1998.5920.9696.6997.60±0.21100.01.1672.2498.51±1.06
Avg BA97.7033.9690.5696.16±0.0997.9114.9395.0897.22±0.2199.730.8061.8796.70±2.46
MK [4]MA74.1775.4674.3078.03±0.3375.2476.2176.0776.49±0.8657.8458.6458.6558.73±0.55
Best BA24.063.1693.5498.01±0.806.047.5295.8998.61±0.140.880.9997.5299.94±0.01
Avg BA5.991.3376.4195.98±0.363.643.8592.4097.03±1.570.620.5489.2294.42±2.87
FLTrust [5]MA83.1483.2582.9883.84±0.2080.7580.2280.2280.33±0.2961.5861.8061.8961.14±1.81
Best BA56.236.5396.9697.95±0.3491.814.1897.1997.78±0.2730.120.6995.4299.98±0.05
Avg BA42.454.9196.6096.74±0.4789.123.6795.4891.94±8.4119.870.4687.5198.52±1.59
FLAME [23]MA58.0757.5258.6363.46±0.9174.2173.5271.9073.06±0.7655.9354.2855.6255.87±0.45
Best BA21.2037.8492.4398.38±0.6312.702.9794.4698.69±0.170.830.5377.6799.99±0.00
Avg BA7.7512.7678.8997.46±0.447.012.2789.7297.54±0.580.620.4472.8999.88±0.08
MM [12]MA69.3166.4370.7275.42±0.9771.4372.6172.8073.16±1.0753.0353.8253.4053.55±0.33
Best BA29.0816.0294.9598.31±0.4415.541.7196.0198.47±0.090.970.8592.7399.99±0.00
Avg BA7.045.4284.6194.55±4.203.911.3793.1297.52±0.670.370.6683.4697.05±1.58
DnC [29]MA81.0982.682.3281.30±1.4577.9879.3278.6378.83±0.3063.0161.4363.8963.43±0.73
Best BA5.942.8395.0497.85±0.276.5811.2096.2598.19±0.140.434.8880.9399.67±0.29
Avg BA3.681.6793.5796.94±0.534.039.7294.4597.29±0.370.412.6963.3798.07±0.59
RLR [25]MA72.2876.2277.7177.90±0.6875.7275.0674.5174.77±1.3754.7756.7856.1456.58±0.58
Best BA96.9323.0694.1197.78±0.1197.2917.1694.8398.39±0.130.140.3488.1299.98±0.01
Avg BA95.6515.1391.6197.39±0.2997.0210.8589.4197.95±0.170.050.0983.8199.97±0.02
FLARE [34]MA84.4584.2484.5384.56±0.1177.7978.2177.5277.82±2.0560.4456.5859.5459.07±1.41
Best BA98.2143.8794.7997.85±0.1298.185.8894.9196.75±0.5099.711.1488.5699.56±0.41
Avg BA98.1136.3294.1097.28±0.3098.025.4093.5596.30±0.8298.600.6380.7398.17±2.42
", + "bbox": [ + 119, + 161, + 877, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "approaches [2, 41, 42], we assume that the attacker controls exactly one client per round and sets the attack interval to $F = 1$ . This setup allows us to analyze the attack in isolation, without client collusion. We also evaluate the attack's performance across different frequencies, comparing it to the baseline.", + "bbox": [ + 89, + 560, + 482, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baseline Defenses and Attacks. We assess our backdoor attack on eight state-of-the-art or representative defenses in FL: MultiKrum (MK) [4], FLTrust [5], FLAME [23], Multi-Metrics (MM) [12], DnC [29], RLR [25], FLARE [34], and DeepSight [28], ensuring a comprehensive evaluation of the attack's effectiveness.", + "bbox": [ + 89, + 670, + 482, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare our backdoor attack against three state-of-the-art or representative attacks in FL: BadNets [10], DBA [35], and LP attack [42].", + "bbox": [ + 89, + 760, + 483, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. Following prior work on FL backdoor attacks [10, 35, 42], we evaluate the global model using two primary metrics: main task accuracy (MA) and backdoor accuracy (BA). Given the dynamic nature of the global model in FL, we report both the average and the best BA over the last", + "bbox": [ + 89, + 824, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "10 communication rounds. Furthermore, to evaluate the stealthiness of the attack, we adopt two metrics introduced in [42]: the malicious client acceptance rate (MAR) and the benign client acceptance rate (BAR). MAR measures the proportion of rounds in which malicious clients successfully bypass defense and are selected for aggregation, while BAR represents the average proportion of rounds in which benign clients are selected for aggregation.", + "bbox": [ + 511, + 560, + 905, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. The effectiveness of our attack", + "text_level": 1, + "bbox": [ + 511, + 694, + 782, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first evaluate our attack in the fixed-frequency attack setting under different SOTA defenses. Table 1 presents the performance comparison between our attack and the baseline attacks. We also evaluate our attack in the IID setting in supplementary §B.1 and on a larger dataset TinyImageNet in supplementary §B.11. The results show that our attack achieves the highest BA in most cases. These results demonstrate the broad effectiveness of our attack, effectively bypassing defenses while maintaining superior attack performance.", + "bbox": [ + 511, + 717, + 906, + 868 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "BadNets only works well under FLARE and consistently fails to bypass defenses such as MK, FLAME, MM, and", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "29167", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/df455fb8d4e898343907ca4ca722560cfcf39de61c3db0d9e9f12e23bf380456.jpg", + "table_caption": [ + "Table 2. MAR and BAR comparison under outlier-based defenses: MK, FLAME, MM, and DnC." + ], + "table_footnote": [], + "table_body": "
Dataset (Model)AttackMKFLAMEMMDnC
BARMARBARMARBARMARBARMAR
CIFAR10 (ResNets18)BadNets0.420.010.750.010.330.010.970.01
DBA0.410.010.710.010.340.010.950.01
LP attack0.360.720.650.900.130.80.390.82
Ours0.380.680.640.880.250.790.780.99
CIFAR10 (VGG19)BadNets0.410.010.720.010.310.010.980.01
DBA0.420.010.750.010.330.030.970.01
LP attack0.350.630.610.750.130.80.410.99
Ours0.380.560.60.90.260.600.790.99
CIFAR100 (ResNets18)BadNets0.430.010.680.010.330.010.990.01
DBA0.420.010.760.010.330.010.970.01
LP attack0.320.910.560.990.120.930.350.94
Ours0.340.830.580.970.230.920.720.99
", + "bbox": [ + 94, + 119, + 478, + 291 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DnC across all settings, and it also fails against RLR on CIFAR100 in both non-IID and IID settings. Similarly, DBA also fails under several defenses, including MK, FLTrust, MM, and DnC under both non-IID and IID settings. LP attack achieves comparable BA across most defenses, except when evaluated under the FLARE defense in the IID setting. Specifically, for CIFAR100, our attack achieves significantly higher BA than the LP attack, demonstrating superior adaptability to complex data distributions. We also evaluate our attack performance compared with the LP attack under DeepSight in supplementary §B.3. Our attack achieves a higher BA than the LP attack across all scenarios and the LP attack fails on CIFAR100.", + "bbox": [ + 89, + 304, + 483, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our attack improves the attack stealthiness and maintains strength. Under the FedAvg setting, our attack achieves significantly higher BA compared to both DBA and LP attack, particularly on CIFAR100, approaching the performance of unconstrained attacks like BadNets. This demonstrates the strong strength of our attack.", + "bbox": [ + 89, + 522, + 483, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2 presents the comparison of MAR and BAR under outlier-based defenses such as MK, FLAME, MM, and DnC. A higher MAR indicates that malicious updates are more likely to bypass detection. We observe that BadNets and DBA fail to bypass defenses, as their MAR remains close to zero in all cases. This suggests that BadNets and DBA are consistently detected as outliers during training, rendering them ineffective against robust defenses. While BadNets demonstrates strong attack strength, its failure to bypass defenses undermines its overall effectiveness. In contrast, our attack achieves MAR consistently higher than BAR, meaning our malicious updates are more likely to be selected for aggregation, which reflects the stealthiness capability of our method.", + "bbox": [ + 89, + 613, + 483, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When compared to the LP attack, our method maintains nearly the same or even higher MAR in defenses such as FLAME and DnC, further demonstrating its stealthiness. MK and MM impose stricter selection criteria by choosing only a small proportion of the most benign clients, making", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "it more challenging for an attacker to be selected. LP attack achieves a higher MAR than our method under these defenses by restricting the backdoor to BC layers, which significantly enhances its stealthiness. However, our method consistently achieves higher BA than LP attack across all settings in FLAME, DnC, MK, and MM, demonstrating that while LP attack sacrifices attack strength for stealthiness, our approach achieves a better balance, ensuring both strong attack strength and high stealthiness.", + "bbox": [ + 511, + 90, + 906, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "More evaluation of attack stealthiness. We leverage Krum distance, cosine distance, and Manhattan distance to further illustrate the stealthiness of our attack. Specifically, we calculate the sum of squared Euclidean distances, cosine distances, and Manhattan distances between a model update and other local model updates, referring to them as Krum distance, Cosine distance, and Manhattan distance, respectively.", + "bbox": [ + 511, + 244, + 908, + 366 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A larger Krum distance indicates that the update deviates significantly from benign updates, making it less likely to be accepted by the aggregation mechanism. Figure 2 shows the comparison of Krum distance with baseline attacks. The Krum distances of BadNets and DBA are significantly larger than those of benign updates, indicating that these attacks produce model updates that are easily distinguishable from benign ones. In contrast, our attack and the LP attack exhibit Krum distances close to those of benign updates, suggesting that they are more stealthy. Although DBA demonstrates slightly better stealthiness than BadNets, its Krum distance remains considerably larger than that of benign updates. This suggests that DBA is still insufficiently stealthy against robust defenses. In supplementary §B.4, we further discuss the Cosine distance and Manhattan distance between the malicious update and the benign update. The result exhibits similar trends across all attacks.", + "bbox": [ + 511, + 366, + 908, + 622 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e849e5a29d08a96d8565e8754f48b0feea385e10d664fb29277a5bd388eee77d.jpg", + "table_caption": [ + "Table 3. The BA of various backdoor attacks combined without/with ours." + ], + "table_footnote": [], + "table_body": "
AttackIBA [24]IBA+oursCerp [20]Cerp+oursA3FL [40]A3FL+ours
FLAME2.7198.04↑29.8499.64↑94.8799.87↑
MM31.6399.99↑24.7299.98↑86.2999.89↑
", + "bbox": [ + 516, + 664, + 903, + 722 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our attack demonstrates strong generalizability. It can be easily integrated into existing FL backdoor attacks, particularly those based on trigger optimization. As shown in Table 3, incorporating our method further enhances attack effectiveness, underscoring its practical value.", + "bbox": [ + 511, + 750, + 906, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further evaluate our attack trained on other networks: ResNet50 [11], ResNet101 [11] and ViT [6] in supplementary §B.5. Our attack achieves high BA on those networks, demonstrating its generalizability to more complex architectures.", + "bbox": [ + 511, + 824, + 908, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "29168", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/575f9f1063f4b2c8a1843a068f3c31fda3c8d70ac9a1cf73ca23e5626fb40014.jpg", + "image_caption": [ + "(a) Ours" + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 339, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b9e4c51957637677699b7c6e072df031c4c715311495a22cab8096573ff4ce2e.jpg", + "image_caption": [ + "(b) LP attack" + ], + "image_footnote": [], + "bbox": [ + 361, + 88, + 609, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fba678557aa3469d28691cc5b762a1d605f02e32a1de260d54bfdf4b4ed37a6b.jpg", + "image_caption": [ + "(c) BadNets and DBA" + ], + "image_footnote": [], + "bbox": [ + 632, + 89, + 877, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3d67cbc95e873a5b62984da63b35810ffb0f5807b195437aa53988a87fc86a78.jpg", + "image_caption": [ + "Figure 2. Krum distance of malicious and benign model updates. A larger distance indicates that the update deviates significantly from benign updates. The result shows the stealthiness of our attack." + ], + "image_footnote": [], + "bbox": [ + 93, + 257, + 284, + 368 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e190363b426e6b8ea4a385ac74797fe89d1f3fca396f68ccbd83b2052448cc1d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 258, + 480, + 367 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2105f793fb1b4346702c883ffd7a17a248aa45429ff879b3f2e018264b1c88b4.jpg", + "image_caption": [ + "(a) MK (CIFAR-10)", + "(c) MK (CIFAR-100)", + "Figure 3. Impacts of different attack intervals $F$ on the attack performance. The $F$ indicates that an attack is performed every $F$ rounds." + ], + "image_footnote": [], + "bbox": [ + 93, + 383, + 282, + 494 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e37ff7d581a43f51ac3de764c871d10da6f8db87eb5364c40f377c340ce83975.jpg", + "image_caption": [ + "(b) FLAME (CIFAR-10)", + "(d) FLAME (CIFAR-100)" + ], + "image_footnote": [], + "bbox": [ + 287, + 383, + 475, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition, we analyze the impact of varying degrees of data heterogeneity in supplementary §B.6. Results show that our attack remains effective under highly non-IID conditions and consistently outperforms the LP attack across all scenarios, further confirming its generalizability.", + "bbox": [ + 89, + 571, + 483, + 647 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our attack maintains the global model's performance. Supplementary §B.2 shows the MA with the unpoisoned global model to assess whether our attack affects the utility of the global model. The results demonstrate that our attack does not lead to a significant decrease in MA compared to the unpoisoned global model.", + "bbox": [ + 89, + 669, + 483, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Our attack under different scenarios", + "text_level": 1, + "bbox": [ + 89, + 772, + 408, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of attack intervals. We evaluate the performance of our attack compared to baseline attacks under different attack intervals, as shown in Figure 3. The attack interval $F$ ranges from 1 to 5, where $F$ indicates that an attack is performed once every $F$ rounds. This setting is designed to simulate scenarios where the frequency of malicious updates is limited, requiring more effective and persistent attack", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0b9544a783f459438d10ac4f5052f3dd28b7887cb617d1e40eec390aef14ed8a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 258, + 705, + 368 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f14ebdacb58da45c5f06eefbe201609d3b5da69e8a508ab25c746bf5752b4f26.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 258, + 901, + 368 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dc08f1af510875b2d0889e0f97dc31b2b018d876875e3843f2843c2fe0c05cbe.jpg", + "image_caption": [ + "(a) MK (CIFAR-10)", + "(c) MK (CIFAR-100)", + "Figure 4. Effectiveness of our attack under different proportions of compromised clients ( $C = 0.02, 0.04, 0.06, 0.1$ ) in a fixed-pool attack setting." + ], + "image_footnote": [], + "bbox": [ + 516, + 383, + 705, + 494 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/01b7a387c13fcdeb60c840ffdd1b29e753cf21c6a92b4b3832c720fc349a8e08.jpg", + "image_caption": [ + "(b) FLAME (CIFAR-10)", + "(d) FLAME (CIFAR-100)" + ], + "image_footnote": [], + "bbox": [ + 710, + 383, + 897, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "strategies.", + "bbox": [ + 513, + 573, + 584, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results show that as $F$ increases, the performance of all attacks degrades as the attack frequency decreases. This trend is consistent across all evaluated methods, where a lower frequency results in reduced BA. A larger $F$ imposes greater demands on the attack's strength, persistence, and stealthiness. The attack must be sufficiently powerful and stealthy to maintain its effectiveness with fewer opportunities to update the model. We observe that BadNets fails in all cases. Notably, our attack maintains a comparable BA even when $F = 5$ . Whereas the LP attack loses nearly all effectiveness, particularly when trained on CIFAR100, demonstrating its vulnerability to defenses when attack opportunities are limited. We also show the results of BA under all defenses trained on CIFAR10 (ResNet18) in supplementary §B.8. We observe that our attack achieves comparable BA across all defenses when $F = 5$ .", + "bbox": [ + 511, + 588, + 908, + 829 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of different proportions of compromised clients. To evaluate our attack in a more practical scenario, we conduct experiments in a fixed-pool attack setting. To further", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "29169", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e6df169697eb6a1fd9071d261d5f77ec159239becf7ef8b431999abefab8aab5.jpg", + "table_caption": [ + "Table 4. Ablation study on alignment strategy." + ], + "table_footnote": [], + "table_body": "
Dataset (Model)AttackNo DefenseFLAMEMM
BABAMARBAMAR
CIFAR-10 (VGG19)no alignment97.707.050.017.040.01
model-wise90.1245.790.9083.580.65
layer-wise (ours)96.1697.460.9097.090.67
CIFAR-10 (ResNet18)no alignment97.917.010.013.910.01
model-wise93.3985.320.8995.530.80
layer-wise (ours)97.2198.160.8897.850.79
CIFAR-100 (ResNet18)no alignment99.730.620.010.370.01
model-wise73.2985.860.9595.050.90
layer-wise (ours)96.7099.880.9797.050.92
", + "bbox": [ + 94, + 104, + 480, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "demonstrate the strength of our attack, we also analyze its performance under varying proportions of compromised clients $C$ . In a fixed-pool attack setting, the attacker controls a pool of compromised clients, from which participants are randomly selected in each round. This setup allows multiple malicious clients to be chosen in a single round, while in some cases, several consecutive rounds may pass without selecting any malicious clients.", + "bbox": [ + 89, + 290, + 483, + 411 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Figure 4, our attack achieves the highest BA compared to the baseline across all cases. Notably, even with a very low proportion of malicious clients ( $C = 0.02$ , meaning only $2\\%$ of clients are compromised), our attack successfully bypasses defenses and maintains a comparable BA. This demonstrates that our attack is both powerful and persistent, as it continues to perform well even when the pool of malicious clients is small.", + "bbox": [ + 89, + 412, + 483, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additionally, BadNets performs better when $C = 0.1$ due to the increased likelihood of malicious clients colluding in the same round. This collusion significantly enhances the attack's effectiveness, as multiple attackers working together are more likely to successfully inject a backdoor. This is particularly evident under MK, where BadNets achieves a higher BA with a higher $C$ , compared to when it fails in the fixed-frequency setting. However, this improvement diminishes as the compromised client ratio decreases. At lower $C$ , the attack becomes more reliant on its strength. In contrast, as shown in supplementary §B.9, our attack achieves high BA across all defenses when $C > 0.04$ , demonstrating its robustness even with fewer compromised clients.", + "bbox": [ + 89, + 534, + 483, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation study", + "text_level": 1, + "bbox": [ + 89, + 742, + 240, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of different triggers. We further evaluate the impact of various trigger shapes. Supplementary §B.10 presents three different trigger shapes (Apple, Watermark, and Square) applied to backdoor a ResNet18 model trained on the CIFAR10 dataset. Figure 5a shows the impact of various trigger shapes. Under the default setting $(F = 1)$ , all three trigger shapes achieve high BA. The results demonstrate that our attack remains effective across different trigger shapes, indicating its robustness and generalizability.", + "bbox": [ + 89, + 763, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/75ec940596f787780c656c100a20fc75b8b01c81b1653b92be4b3b53a1c1660c.jpg", + "image_caption": [ + "(a) various trigger shapes" + ], + "image_footnote": [], + "bbox": [ + 517, + 87, + 705, + 198 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4a21bf3029a24c2a85ea2eb8ed226d8840c5fd01fdc3975c6ac11db85c3a90e1.jpg", + "image_caption": [ + "(b) various trigger sizes", + "Figure 5. Impact of trigger shape (left) and trigger size (right) in Flame." + ], + "image_footnote": [], + "bbox": [ + 709, + 88, + 898, + 198 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 5b illustrates the performance of different trigger sizes (3, 5, 7, and 10). Regardless of trigger size, our attack consistently achieves high BA. Notably, when the trigger size is set to 10, the attack effectiveness is significantly enhanced. Larger triggers strengthen the attack, and their ability to bypass defenses is improved through layer-wise norm constraints, leading to superior attack performance.", + "bbox": [ + 511, + 271, + 906, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of alignment strategy. To demonstrate how dynamic layer-wise gradient alignment enhances the effectiveness of our attack, we compare it with BadNets (no alignment) and the model-wise gradient alignment attack, which adjusts the overall gradient magnitude to align with the previous global update. As shown in Table 4, our attack consistently outperforms both BadNets and the model-wise method across all settings.", + "bbox": [ + 511, + 417, + 906, + 540 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "First, both model-wise and layer-wise gradient alignment show higher MAR compared to BadNets, indicating that aligning the update with the previous global model significantly improves the stealthiness of the attack. Second, when compared to model-wise alignment, layer-wise alignment achieves better attack performance. Specifically, even in the absence of an attack, particularly on CIFAR100, the BA for model-wise alignment is much lower than that for layer-wise alignment.", + "bbox": [ + 511, + 542, + 908, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 705, + 633, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper proposes a new stealthy backdoor attack in FL, which leverages adaptive layer-wise gradient alignment to enhance stealthiness while preserving attack strength. By dynamically aligning malicious gradients to historical global updates at the layer level, the proposed attack reduces statistical anomalies caused by malicious updates, making them harder to detect. Extensive experiments across diverse datasets, models, and defenses demonstrate that the proposed attack achieves superior backdoor success rates while evading SOTA defenses, highlighting its effectiveness and stealthiness in various FL scenarios.", + "bbox": [ + 511, + 734, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "29170", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China (NO. 62472284, No. 62302288), Shanghai Key Laboratory of Scalable Computing and Systems, SJTU Kunpeng & Ascend Center of Excellence.", + "bbox": [ + 89, + 113, + 485, + 172 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 196, + 187, + 213 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sana Awan, Bo Luo, and Fengjun Li. CONTRA: Defending Against Poisoning Attacks in Federated Learning. In ESORICS, 2021. 2", + "[2] Eugene Bagdasaryan, Andreas Veit, Yiqing Hua, Deborah Estrin, and Vitaly Shmatikov. How To Backdoor Federated Learning. In AISTATS, 2020. 1, 2, 3, 5", + "[3] Akshay Batheja and Pushpak Bhattacharyya. \"A Little is Enough\": Few-Shot Quality Estimation based Corpus Filtering improves Machine Translation. In ACL (Findings), 2023. 2", + "[4] Peva Blanchard, El Mahdi El Mhamdi, Rachid Guerraoui, and Julien Stainer. Machine Learning with Adversaries: Byzantine Tolerant Gradient Descent. In NeurIPS, 2017. 1, 2, 3, 5", + "[5] Xiaoyu Cao, Minghong Fang, Jia Liu, and Neil Zhenqiang Gong. FLTrust: Byzantine-robust Federated Learning via Trust Bootstrapping. In NDSS, 2021. 1, 2, 3, 4, 5", + "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 6", + "[7] Minghong Fang, Xiaoyu Cao, Jinyuan Jia, and Neil Gong. Local Model Poisoning Attacks to Byzantine-Robust Federated Learning. In USENIX Security, 2020. 4", + "[8] Pei Fang and Jinghui Chen. On the Vulnerability of Backdoor Defenses for Federated Learning. In AAAI, 2023. 2", + "[9] Clement Fung, Chris JM Yoon, and Ivan Beschastnikh. The Limitations of Federated Learning in Sybil Settings. In RAID, 2020. 2", + "[10] Tianyu Gu, Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. BadNets: Evaluating Backdooring Attacks on Deep Neural Networks. IEEE Access, 7:47230-47244, 2019. 5", + "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In CVPR, 2016. 4, 6", + "[12] Siquan Huang, Yijiang Li, Chong Chen, Leyu Shi, and Ying Gao. Multi-Metrics Adaptively Identifies Backdoors in Federated Learning. In ICCV, 2023. 1, 2, 3, 5", + "[13] Tiansheng Huang, Sihao Hu, Ka Ho Chow, Fatih Ilhan, Selim F. Tekin, and Ling Liu. Lockdown: Backdoor Defense for Federated Learning with Isolated Subspace Training. In NeurIPS, 2023. 3", + "[14] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and Open Problems in Federated" + ], + "bbox": [ + 93, + 222, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Learning. Foundations and trends® in machine learning, 14 (1-2):1-210, 2021. 1", + "[15] Yun-Yong Ko, Dongwon Lee, and Sang-Wook Kim. Not All Layers Are Equal: A Layer-Wise Adaptive Approach Toward Large-Scale DNN Training. In WWW, 2022. 2", + "[16] Jakub Konečny. Federated learning: Strategies for improving communication efficiency. arXiv preprint arXiv:1610.05492, 2016. 2", + "[17] Torsten Krauß and Alexandra Dmitrienko. MESAS: Poisoning Defense for Federated Learning Resilient against Adaptive Attackers. In CCS, 2023. 2, 3", + "[18] Haoyang Li, Qingqing Ye, Haibo Hu, Jin Li, Leixia Wang, Chengfang Fang, and Jie Shi. 3DFed: Adaptive and Extensible Framework for Covert Backdoor Attack in Federated Learning. In S&P, 2023. 2", + "[19] Ye Li, Yanchao Zhao, Chengcheng Zhu, and Jiale Zhang. Infighting in the dark: Multi-labels backdoor attack in federated learning. arXiv preprint arXiv:2409.19601, 2024. 2", + "[20] Xiaoting Lyu, Yufei Han, Wei Wang, Jingkai Liu, Bin Wang, Jiqiang Liu, and Xiangliang Zhang. Poisoning with Cerberus: Stealthy and Colluded Backdoor Attack against Federated Learning. In AAAI, 2023. 2, 3, 6", + "[21] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y Arcas. Communication-Efficient Learning of Deep Networks from Decentralized Data. In AISTATS, 2017. 1", + "[22] El Mahdi El Mhamdi, Rachid Guerraoui, and Sébastien Rouault. The Hidden Vulnerability of Distributed Learning in Byzantium. In ICML, 2018. 2", + "[23] Thien Duc Nguyen, Phillip Rieger, Huili Chen, Hossein Yalame, Helen Möllering, Hossein Fereidooni, Samuel Marchal, Markus Miettinen, Azalia Mirhoseini, Shaza Zeitouni, Farinaz Koushanfar, Ahmad-Reza Sadeghi, and Thomas Schneider. FLAME: Taming Backdoors in Federated Learning. In USENIX Security 22, 2022. 1, 2, 3, 5", + "[24] Thuy Dung Nguyen, Tuan Nguyen, Anh Tran, Khoa D. Doan, and Kok-Seng Wong. IBA: Towards Irreversible Backdoor Attacks in Federated Learning. In NeurIPS, 2023. 2, 6", + "[25] Mustafa Safa Ozdayi, Murat Kantarcioglu, and Yulia R Gel. Defending against backdoors in federated learning with robust learning rate. In AAAI, 2021. 1, 3, 5", + "[26] Ashwinee Panda, Saeed Mahloujifar, Arjun Nitin Bhagoji, Supriyo Chakraborty, and Prateek Mittal. SparseFed: Mitigating Model Poisoning Attacks in Federated Learning with Sparsification. In AISTATS, 2022. 3", + "[27] Matthias Paulik, Matt Seigel, Henry Mason, Dominic Telaar, Joris Kluivers, Rogier Van Dalen, Chi Wai Lau, Luke Carlson, Filip Granqvist, Chris Vandevelde, et al. Federated Evaluation and Tuning for On-Device Personalization: System Design & Applications. arXiv preprint arXiv:2102.08503, 2021. 1", + "[28] Phillip Rieger, Thien Duc Nguyen, Markus Miettinen, and Ahmad-Reza Sadeghi. DeepSight: Mitigating Backdoor Attacks in Federated Learning Through Deep Model Inspection. In NDSS, 2022. 2, 3, 5", + "[29] Virat Shejwalkar and Amir Houmansadr. Manipulating the Byzantine: Optimizing Model Poisoning Attacks and Defenses for Federated Learning. In NDSS, 2021. 1, 2, 3, 5" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "29171", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Karen Simonyan. Very Deep Convolutional Networks for Large-Scale Image Recognition. arXiv preprint arXiv:1409.1556, 2014. 4", + "[31] Peng Sun, Xinyang Liu, Zhibo Wang, and Bo Liu. Byzantine robust Decentralized Federated Learning via Dual-domain Clustering and Trust Bootstrapping. In CVPR, 2024. 2", + "[32] Ziteng Sun, Peter Kairouz, Ananda Theertha Suresh, and H Brendan McMahan. Can You Really Backdoor Federated Learning? arXiv preprint arXiv:1911.07963, 2019. 1, 3", + "[33] Hongyi Wang, Kartik Sreenivasan, Shashank Rajput, Harit Vishwakarma, Saurabh Agarwal, Jy-yong Sohn, Kangwook Lee, and Dimitris S. Papailiopoulos. Attack of the Tails: Yes, You Really Can Backdoor Federated Learning. In NeurIPS, 2020. 1, 2", + "[34] Ning Wang, Yang Xiao, Yimin Chen, Yang Hu, Wenjing Lou, and Y. Thomas Hou. FLARE: Defending Federated Learning against Model Poisoning Attacks via Latent Space Representations. In AsiaCCS, 2022. 2, 5", + "[35] Chulin Xie, Keli Huang, Pin-Yu Chen, and Bo Li. DBA: Distributed Backdoor Attacks against Federated Learning. In ICLR, 2020. 1, 2, 3, 5", + "[36] Yueqi Xie, Minghong Fang, and Neil Zhenqiang Gong. FedREDefence: Defending against Model Poisoning Attacks for Federated Learning using Model Update Reconstruction Error. In ICML, 2024. 3", + "[37] Jiahao Xu, Zikai Zhang, and Rui Hu. Identify Backdoored Model in Federated Learning via Individual Unlearning. arXiv preprint arXiv:2411.01040, 2024. 3", + "[38] Peishen Yan, Hao Wang, Tao Song, Yang Hua, Ruhui Ma, Ningxin Hu, Mohammad Reza Haghighat, and Haibing Guan. SkyMask: Attack-Agnostic Robust Federated Learning With Fine-Grained Learnable Masks. In ECCV, 2025. 3", + "[39] Timothy Yang, Galen Andrew, Hubert Eichner, Haicheng Sun, Wei Li, Nicholas Kong, Daniel Ramage, and Françoise Bea-fays. Applied Federated Learning: Improving Google Keyboard Query Suggestions. arXiv preprint arXiv:1812.02903, 2018. 1", + "[40] Hangfan Zhang, Jinyuan Jia, Jinghui Chen, Lu Lin, and Ding-hao Wu. A3FL: Adversarily Adaptive Backdoor Attacks to Federated Learning. In NeurIPS, 2023. 2, 6", + "[41] Zhengming Zhang, Ashwinee Panda, Linyue Song, Yaoqing Yang, Michael W. Mahoney, Prateek Mittal, Kannan Ramchandran, and Joseph Gonzalez. Neurotoxin: Durable Backdoors in Federated Learning. In ICML, 2022. 1, 2, 3, 4, 5", + "[42] Haomin Zhuang, Mingxian Yu, Hao Wang, Yang Hua, Jian Li, and Xu Yuan. Backdoor Federated Learning by Poisoning Backdoor-Critical Layers. In ICLR, 2024. 1, 2, 3, 4, 5" + ], + "bbox": [ + 91, + 92, + 482, + 771 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "29172", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Stealthy Backdoor Attack in Federated Learning via Adaptive Layer-wise Gradient Alignment/f7dfb2e6-c547-4b0c-bb2e-4bfc0b02a71f_model.json b/2025/Stealthy Backdoor Attack in Federated Learning via Adaptive Layer-wise Gradient Alignment/f7dfb2e6-c547-4b0c-bb2e-4bfc0b02a71f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b4a8d601382d0dd1818406456ff2d6e6b5c2973c --- /dev/null +++ b/2025/Stealthy Backdoor Attack in Federated Learning via Adaptive Layer-wise Gradient Alignment/f7dfb2e6-c547-4b0c-bb2e-4bfc0b02a71f_model.json @@ -0,0 +1,2343 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.252, + 0.131, + 0.748, + 0.177 + ], + "angle": 0, + "content": "Stealthy Backdoor Attack in Federated Learning via Adaptive Layer-wise Gradient Alignment" + }, + { + "type": "text", + "bbox": [ + 0.14, + 0.203, + 0.856, + 0.277 + ], + "angle": 0, + "content": "Qingqian Yang\\(^{1*}\\) Peishen Yan\\(^{2}\\) Xiaoyu Wu\\(^{2}\\) Jiaru Zhang\\(^{2}\\) Tao Song\\(^{2\\dagger}\\) Yang Hua\\(^{3}\\) Hao Wang\\(^{4}\\) Liangliang Wang\\(^{1\\dagger}\\) Haibing Guan\\(^{2}\\) \n\\(^{1}\\)Shanghai University of Electric Power \\(^{2}\\)Shanghai Jiao Tong University \n\\(^{3}\\)Queen's University Belfast \\(^{4}\\)Stevens Institute of Technology" + }, + { + "type": "text", + "bbox": [ + 0.378, + 0.278, + 0.621, + 0.292 + ], + "angle": 0, + "content": "y23108013@mail.shiep.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.2, + 0.295, + 0.798, + 0.311 + ], + "angle": 0, + "content": "{peishenyan, wuxiaoyu2000, jiaruzhang, songt333, hbguan}@sjtu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.313, + 0.746, + 0.327 + ], + "angle": 0, + "content": "Y.Hua@qub.ac.uk hwang9@stevens.edu llwang@shiep.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.362, + 0.327, + 0.379 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.394, + 0.486, + 0.773 + ], + "angle": 0, + "content": "The distributed nature of federated learning exposes it to significant security threats, among which backdoor attacks are one of the most prevalent. However, existing backdoor attacks face a trade-off between attack strength and stealthiness: attacks maximizing the attack strength are often detectable, while stealthier approaches significantly reduce the effectiveness of the attack itself. Both of them result in ineffective backdoor injection. In this paper, we propose an adaptive layer-wise gradient alignment strategy to effectively evade various robust defense mechanisms while preserving attack strength. Without requiring additional knowledge, we leverage the previous global update as a reference for alignment to ensure stealthiness during dynamic FL training. This fine-grained alignment strategy applies appropriate constraints to each layer, which helps significantly maintain attack strength. To demonstrate the effectiveness of our method, we conduct extensive evaluations across a wide range of datasets and networks. Our experimental results show that the proposed attack effectively bypasses eight state-of-the-art defenses and achieves high backdoor accuracy, outperforming existing attacks by up to \\(54.76\\%\\). Additionally, it significantly preserves attack strength and maintains robust performance across diverse scenarios, highlighting its adaptability and generalizability. Code implementation is available at https://github.com/yqqhyqq/LGA." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.8, + 0.222, + 0.815 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.485, + 0.856 + ], + "angle": 0, + "content": "Federated Learning (FL) has emerged as a promising paradigm for privacy-preserving distributed machine learn" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.909, + 0.59 + ], + "angle": 0, + "content": "ing, allowing multiple decentralized clients to collaboratively train a global model without directly sharing their raw data [14, 21]. FL leverages local computation on edge devices (e.g., smartphones and IoT sensors) and aggregates model updates via a central server, making it widely adopted in various real-world applications [27, 39]. However, the distributed architecture of FL systems significantly increases the potential attack surface, introducing a wide range of security threats, among which backdoor attacks are a commonly studied threat in recent research. In a backdoor attack [2, 32, 33, 35, 41, 42], adversarial clients intentionally inject a backdoor into the global model by uploading a poisoning local update. As a result, the global model behaves normally on benign inputs but misclassifies triggered inputs into predefined target labels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.91, + 0.715 + ], + "angle": 0, + "content": "Existing backdoor defenses primarily focus on detecting and filtering malicious updates by leveraging information from multiple dimensions. Specifically, some methods rely on distance-based measures between benign and malicious updates [4, 5, 12, 23], while others detect inconsistencies in parameter signs [25] or exploit the spectral separability [29]. These defenses significantly restrict the feasibility of backdoor attacks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.911, + 0.903 + ], + "angle": 0, + "content": "To counter these defenses, an effective backdoor attack depends on two key factors: attack strength and attack stealthiness. Some attacks [2] achieve high attack strength by scaling up the update, which increases their detectability and makes them easily filtered by defense mechanisms. Others [35, 42] prioritize stealthiness to evade detection but show significantly lower performance when no defense is present, highlighting their sacrifice of attack strength. These attacks essentially constrain the malicious update by discarding certain backdoor-related information [42] or using coarse-grained scaling [33]. Although constraining malicious updates helps maintain stealthiness, this inappropriate" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.486, + 0.901 + ], + "angle": 0, + "content": "*The work was done during her internship at Shanghai Jiao Tong University, Shanghai Key Laboratory of Scalable Computing and Systems. † Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29163" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "constraint leads to attack strength degradation. Therefore, balancing the attack stealthiness and attack strength with limited knowledge of the system's defenses is challenging." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.138, + 0.486, + 0.453 + ], + "angle": 0, + "content": "To address the challenge, we propose a layer-wise adaptive gradient alignment strategy that dynamically adjusts the norm bound for each layer of malicious updates. This adjustment enhances their similarity to the benign update distribution in a fine-grained manner, thereby achieving both attack stealthiness and attack effectiveness. Without requiring any additional information, our approach uses the previous global update as a reference point for adaptive alignment. By doing so, the proposed strategy not only constrains the malicious update within a small norm ball, but also enables the attack to adapt to the evolving global model and training dynamics, thereby significantly improving its stealthiness. Inspired by previous observations [15] that parameters within the same layer tend to share similar gradient magnitudes, we further conduct empirical analysis revealing that the magnitudes of both malicious and benign updates vary across layers, and there is no consistent proportional relationship between them. This discrepancy highlights the need for layer-specific norm constraints. Such constraints are essential to prevent over-constraining layers with inherently smaller magnitudes, thereby preserving attack strength." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.455, + 0.485, + 0.543 + ], + "angle": 0, + "content": "Extensive experiments validate the effectiveness of our attack, demonstrating its ability to evade eight state-of-the-art defenses while maintaining a high attack success rate. Additionally, we investigate its performance across various attack scenarios, demonstrating its generalizability. We summarize our main contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.547, + 0.483, + 0.68 + ], + "angle": 0, + "content": "- We propose an adaptive layer-wise gradient alignment strategy to effectively evade various robust defense mechanisms while preserving attack strength. Without requiring any additional knowledge, we leverage the previous global update as a reference for gradient alignment, ensuring that the attack maintains the stealthiness while adapting to the training dynamics. This fine-grained alignment strategy applies layer-specific constraints, allowing the attack to retain its strength." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.683, + 0.483, + 0.801 + ], + "angle": 0, + "content": "- We conduct extensive evaluations across diverse models and datasets, demonstrating that the proposed attack successfully bypasses eight SOTA defenses and maintains the attack impact. Our backdoor attack outperforms SOTA attacks, with an improvement in backdoor accuracy by up to \\(54.76\\%\\). Furthermore, our attack exhibits excellent performance across various scenarios, such as different attack patterns, highlighting its generalizability." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.547, + 0.483, + 0.801 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.815, + 0.233, + 0.831 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Backdoor attacks in FL. Backdoor attacks present a serious security threat in FL [16]. Typically, attackers compromise a small subset of client devices to upload poisoned updates. These updates are crafted to inject the backdoor into" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": "the global model, which then enables the model to obtain predetermined predictions for samples with specific trigger features. The backdoor attack process generally consists of two stages: trigger embedding and model manipulation. To successfully inject a backdoor, the attacker can carefully design either the trigger or the model update to bypass detection. Accordingly, existing methods can be divided into two categories: model-optimization attacks [2, 18, 41, 42] and trigger-optimization attacks [8, 20, 24, 33, 35, 40]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.907, + 0.544 + ], + "angle": 0, + "content": "Model-optimization attacks [2, 3, 18, 33, 41, 42] focus on fine-tuning the model to achieve the desired attack. Scaling attack [2] amplifies malicious updates to achieve a one-shot attack. However, the abnormally large gradient magnitudes make it exposed to detection. Neurotoxin [41] projects malicious gradients onto infrequently updated neurons, improving the persistence of the attack and preventing the backdoor from being overwritten by benign updates. However, as an unconstrained backdoor attack, it's vulnerable to robust detection. The LP attack [42] adaptively injects the backdoor into backdoor-critical (BC) layers, achieving high stealthiness. However, by discarding backdoor-related information in non-BC layers, the attack suffers from reduced strength. PGD attack [33] exploits knowledge of server-side defenses to craft malicious updates that bypass detection. However, this strategy is only effective when the attacker has access to detailed information about the system. Additionally, 3DFed [18] introduces three distinct attack strategies targeting different types of defenses and utilizes an indicator for adaptive strategy adjustments. However, it relies on client collaboration." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.546, + 0.909, + 0.711 + ], + "angle": 0, + "content": "Trigger-optimization backdoor attacks aim to design triggers that are both subtle and adaptable, ensuring their stealthiness and effectiveness. Edge-case attack [33] selects marginal data as poison samples to circumvent detection mechanisms. DBA [35] firstly introduces distributed backdoor attacks, decomposing a global trigger into smaller, less detectable components. These approaches craft triggers at the pre-training stage and then fix them during training. Recently, a range of studies [8, 19, 20, 24, 40] dynamically optimize the trigger throughout the training process and obtain better attack results than the former ones." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.908, + 0.824 + ], + "angle": 0, + "content": "Defense strategies against backdoor attacks in FL. In FL [16], the server only receives uploaded models without access to any information about client data. As a result, the majority of defense strategies are similarity-based, aiming to detect and mitigate backdoor attacks by identifying differences between malicious and benign models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Among them, distance-based defenses [1, 4, 5, 9, 12, 17, 22, 23, 28, 29, 31, 34] exploit the observation that malicious models often deviate from benign ones in their distribution within high-dimensional space, making distance-based metrics an effective tool for detection. MultiKrum [4]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "29164" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.272 + ], + "angle": 0, + "content": "only aggregates updates that are selected with the smallest pairwise Euclidean distances. FLTrust [5] weights upload gradients based on their cosine similarity to a trusted gradient to mitigate the influence of malicious models. The trusted gradient is obtained from a root dataset collected by the server. FLAME [23] and DeepSight [28] use pairwise cosine distance between client gradients for clustering, identifying malicious gradients as outliers and further refining the defense through additional modules. Multi-Metrics [12] dynamically adjusts the weights of three distance metrics to detect backdoor models. MESAS [17] leverages six distance metrics for detection." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.273, + 0.486, + 0.545 + ], + "angle": 0, + "content": "Other similarity-based defenses [13, 26, 29, 36, 38] identify malicious updates from more diverse perspectives. DnC [29] leverages singular value decomposition (SVD)-based spectral methods to remove outliers. RLR [25] detects inconsistencies in parameter signs across training rounds and adjusts learning rates accordingly to suppress backdoors. MASA [37] leverages the observation that a malicious model unlearns the main task more quickly than a benign model for detection. FedREDefense [36] relies on the finding that a malicious model is more resistant to distillation. In addition, some methods recently focus on the difference in update behavior between benign models and malicious models at the parameter level. For instance, SparseFed [26] only aggregates Top-k highest magnitude parameters because the backdoor updates are usually injected into coordinates that are unimportant for benign updates. Meanwhile, Lockdown [13] restrains local training in a subspace to further reduce the poison-coupling effect." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.558, + 0.226, + 0.574 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.484, + 0.658 + ], + "angle": 0, + "content": "Background. Federated Learning (FL) distributes the training task across multiple clients to preserve data privacy. At the beginning, a global model \\(\\theta_g^0\\) is initialized on the server and distributed to a set of \\(N\\) clients. In each round \\(t\\), a standard FL framework involves the following steps:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.658, + 0.484, + 0.737 + ], + "angle": 0, + "content": "1. Local Training: Each client \\(i\\) receives the global model \\(\\theta_{g}^{t - 1}\\) and performs \\(E\\) local epochs of stochastic gradient descent (SGD) on its private dataset \\(\\hat{\\mathcal{D}}_i\\) with batch size \\(B\\), obtaining an updated model \\(\\theta_{i}^{t}\\). The client computes the model update as:" + }, + { + "type": "equation", + "bbox": [ + 0.236, + 0.745, + 0.484, + 0.765 + ], + "angle": 0, + "content": "\\[\n\\Delta \\theta_ {i} ^ {t} = \\theta_ {i} ^ {t} - \\theta_ {g} ^ {t - 1}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.773, + 0.295, + 0.787 + ], + "angle": 0, + "content": "and uploads it to the server." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.787, + 0.484, + 0.833 + ], + "angle": 0, + "content": "2. Global aggregation: The server randomly selects a subset of clients and aggregates their model updates \\(\\{\\Delta \\theta^t\\}\\) to compute the global update:" + }, + { + "type": "equation", + "bbox": [ + 0.228, + 0.842, + 0.484, + 0.862 + ], + "angle": 0, + "content": "\\[\n\\Delta \\theta_ {g} ^ {t} = \\sum_ {i \\in [ n ]} \\Delta \\theta_ {i} ^ {t}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.87, + 0.484, + 0.901 + ], + "angle": 0, + "content": "and updates the global model as \\(\\theta_g^t = \\theta_g^{t - 1} + \\Delta \\theta_g^t\\). The new global model is then distributed to clients for the" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.092, + 0.907, + 0.137 + ], + "angle": 0, + "content": "next round. Before updating the global model, the server typically performs backdoor detection during aggregation to mitigate the impact of malicious updates." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.137, + 0.907, + 0.213 + ], + "angle": 0, + "content": "In FL, backdoor attacks exploit the system's decentralized nature. Specifically, an attacker, denoted as \\( A \\), compromises a subset of clients and injects malicious updates to implant a backdoor into the global model. We consider a continuous attack scenario, following prior studies [20, 35, 42]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.909, + 0.394 + ], + "angle": 0, + "content": "Adversarial Objectives: The attacker \\( A \\) aims to inject a backdoor into the global model while minimizing the impact on the model's performance on non-poisoned data. The goal of the backdoor injection is to manipulate the model so that it misclassifies inputs with a specific trigger pattern as an incorrect target label. We denote the malicious update as \\( \\Delta \\theta_{m} \\), which is trained on the poisoned dataset. By carefully crafting \\( \\Delta \\theta_{m} \\), the attacker can introduce the backdoor without significantly degrading the model's overall accuracy. The server's aggregation mechanism, which aggregates updates from multiple clients, may unintentionally include these poisoned updates, thereby corrupting the global model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.909, + 0.5 + ], + "angle": 0, + "content": "Adversarial Capability: The attacker \\( A \\) has the ability to control a small subset of clients and gain access to their datasets. Additionally, the attacker can manipulate the models of these compromised clients [2, 35, 41]. The attacker does not have access to the data or models of benign clients, nor to the specific defense methods implemented by the defender." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.519, + 0.909, + 0.641 + ], + "angle": 0, + "content": "Overview. We introduce an adaptive layer-wise gradient alignment to enhance stealthiness while preserving attack strength. In each round, we train the malicious update on a poisoned local dataset and iteratively align gradients to resemble benign updates before uploading them to the server. This increases the likelihood of the server aggregating the malicious update, effectively injecting the backdoor into the global model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Adaptive reference selection. Traditionally, the Unconstrained malicious update \\(\\Delta \\theta_{m}\\), which is trained on the poisoned dataset including out-of-distribution backdoor samples, tends to exhibit a larger magnitude compared to benign updates [32]. To ensure that the malicious update remains similar to the distribution of benign updates, it should be carefully constrained within an appropriately small norm ball. Notably, the selection of the norm bound significantly impacts the effectiveness of the attack's concealment. Therefore, it is crucial to rigorously determine an appropriate reference point for alignment. Some defense methods, such as MultiKrum [4], select a fixed number of client updates that appear most benign in each round. However, due to data heterogeneity, such methods may inadvertently exclude genuinely benign clients whose updates deviate from the majority. To mitigate this issue, we adopt the previous global" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29165" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.088, + 0.434, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.24, + 0.486, + 0.312 + ], + "angle": 0, + "content": "Figure 1. Layer-wise Magnitude Analysis of Benign vs. Malicious Updates in VGG19. To better visualize, we focus on weight layers and divide them into A-layers and B-layers based on \\(\\|\\Delta\\theta^l\\|\\): those with norms above 0.5 are A-layers; others are B-layers. The experiment uses the default setting and records \\(\\|\\Delta\\theta^l\\|\\) every 50 rounds." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.319, + 0.485, + 0.38 + ], + "angle": 0, + "content": "update \\(\\Delta \\theta_{g}^{t - 1}\\) as a reference point. Since it is computed by aggregating multiple client updates, most of which are expected to be benign, it serves as a reasonable approximation of benign behavior." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.381, + 0.485, + 0.442 + ], + "angle": 0, + "content": "By aligning the malicious update with the global update, the attack remains within the expected range of benign variation, thereby enhancing its stealthiness and increasing the likelihood of successfully influencing the global model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.485, + 0.639 + ], + "angle": 0, + "content": "Layer-wise gradient alignment. Intuitively, applying a norm constraint weakens the impact of the malicious update, which potentially leads to ineffective backdoor injection. We introduce a layer-wise scaling strategy to refine the malicious update at a more granular level to not only ensure the attack remains stealthy but also preserve its strength. A key observation shown in Figure 1 is that the layer-wise magnitudes of the malicious and benign updates do not maintain a fixed proportional relationship; instead, their patterns differ. Therefore, we align each layer individually with \\(\\Delta \\theta_{g}^{t - 1}\\). Specifically, we define the scaling factor \\(S^{t,l}\\) for each layer \\(l\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.638, + 0.484, + 0.675 + ], + "angle": 0, + "content": "\\[\nS ^ {t, l} = \\min (1, \\frac {\\| \\Delta \\theta_ {g} ^ {t - 1 , l} \\|}{\\| \\Delta \\theta_ {m} ^ {t , l} \\|}). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.678, + 0.483, + 0.707 + ], + "angle": 0, + "content": "The scaled malicious update after each local epoch is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.705, + 0.483, + 0.724 + ], + "angle": 0, + "content": "\\[\n\\Delta \\theta_ {m} ^ {t, l} = S ^ {t, l} \\cdot \\Delta \\theta_ {m} ^ {t, l}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.728, + 0.486, + 0.833 + ], + "angle": 0, + "content": "The adaptive layer-wise gradient alignment, implemented in just four lines (Lines 10-13 in Algorithm 1), extends the standard FL backdoor pipeline. This approach prevents excessive constraints on layers with inherently smaller magnitudes and constrains the update for each layer within an \\(\\ell 2\\) norm ball, thereby preserving attack strength while enhancing stealthiness." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.845, + 0.224, + 0.862 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Our empirical study evaluates the effectiveness of the proposed backdoor attack through experiments conducted on" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.091, + 0.907, + 0.155 + ], + "angle": 0, + "content": "Algorithm 1 Layer-wise Gradient Alignment for Attackers \nRequire: Batch size \\(\\ell\\), local epochs \\(E\\), malicious model parameters \\(\\theta_{i}^{t}\\), global parameters \\(\\theta_{g}^{t-1}\\), poisoned dataset \\(\\hat{D}\\), and previous global parameter \\(\\theta_{g}^{t-2}\\)" + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.155, + 0.597, + 0.167 + ], + "angle": 0, + "content": "Procedure:" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.171, + 0.722, + 0.186 + ], + "angle": 0, + "content": "1: // Attacker's local training:" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.187, + 0.788, + 0.201 + ], + "angle": 0, + "content": "2: Calculate the previous global update:" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.201, + 0.697, + 0.218 + ], + "angle": 0, + "content": "3: \\(\\Delta \\theta_{g}^{t - 1}\\gets \\theta_{g}^{t - 1} - \\theta_{g}^{t - 2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.218, + 0.767, + 0.233 + ], + "angle": 0, + "content": "4: Initialize local model \\(\\theta_{i,0}^{t}\\gets \\theta_{g}^{t - 1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.233, + 0.664, + 0.245 + ], + "angle": 0, + "content": "5: for \\( e = 1 \\) to \\( E \\) do" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.171, + 0.788, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.247, + 0.724, + 0.261 + ], + "angle": 0, + "content": "6: for each batch \\(b\\) in \\(\\hat{\\mathcal{D}}\\) do" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.262, + 0.774, + 0.276 + ], + "angle": 0, + "content": "7: Compute gradient for batch \\( b \\):" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.276, + 0.807, + 0.293 + ], + "angle": 0, + "content": "8: \\(\\Delta \\theta_{i,e}^{t}\\gets \\bar{\\Delta}\\theta_{i,e}^{t} + \\nabla_{\\theta}\\mathcal{L}(\\theta_{i,e - 1}^{t},\\hat{\\mathcal{D}}_{b})\\)" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.293, + 0.613, + 0.305 + ], + "angle": 0, + "content": "9: end for" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.262, + 0.807, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.307, + 0.776, + 0.322 + ], + "angle": 0, + "content": "10: // Layer-wise gradient alignment" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.323, + 0.742, + 0.339 + ], + "angle": 0, + "content": "11: for each layer \\(l\\) in \\(\\Delta \\theta_{i,e}^{t}\\) do" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.339, + 0.748, + 0.366 + ], + "angle": 0, + "content": "12: \\(S_{i,e}^{t,l}\\gets min(1,\\frac{\\|\\Delta\\theta_{g}^{t - 1,l}\\|}{\\|\\Delta\\theta_{i,e}^{t,l}\\|})\\)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.366, + 0.759, + 0.385 + ], + "angle": 0, + "content": "13: \\(\\theta_{i,e}^{t,l}\\gets \\Delta \\theta_{i,e}^{t,l}\\cdot S_{i,e}^{t,l} + \\theta_{g}^{t - 1,l}\\)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.386, + 0.612, + 0.396 + ], + "angle": 0, + "content": "14: end for" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.399, + 0.596, + 0.411 + ], + "angle": 0, + "content": "15: end for" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.413, + 0.734, + 0.429 + ], + "angle": 0, + "content": "16:Upload update \\(\\Delta \\theta_{i}^{t}\\) to server" + }, + { + "type": "list", + "bbox": [ + 0.52, + 0.307, + 0.807, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.907, + 0.489 + ], + "angle": 0, + "content": "real-world datasets, simulating an FL environment using the NVIDIA RTX 4090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.498, + 0.702, + 0.515 + ], + "angle": 0, + "content": "4.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.909, + 0.658 + ], + "angle": 0, + "content": "Datasets and Models. We evaluate the attack using two widely used benchmark datasets: CIFAR10 (50,000 training and 10,000 test samples across 10 classes) and CIFAR100 (50,000 training and 10,000 test samples across 100 classes). Following previous studies [41, 42], we use the VGG19 [30] network and ResNet18 network for CIFAR10, while we use the ResNet18 [11] network for CIFAR100. We simulate a non-IID data distribution following prior methods [5, 7, 42], with \\( q = 0.5 / 0.2 \\) for CIFAR10/100, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Attack setup. In each round, we select \\( n = 10 \\) clients among \\( N = 100 \\) clients to participate in the aggregation, with the proportion of compromised clients set to \\( C = 0.1 \\) by default. The FL training runs for 300 rounds to make it converge. We train for 6 local epochs (2 for the benign clients) with a learning rate 0.1 for both CIFAR10 and CIFAR100. Additionally, we assess the effectiveness of our attack in an IID setting, as defense strategies are more likely to identify malicious models as outliers in this case [42]. The trigger is a \\( 5 \\times 5 \\) pixel square located at the bottom-right corner of each image by default. We further discuss the impact of various shapes and sizes of the trigger. We evaluate our attack in the fixed-frequency attack scenario. The attacker controls a fixed number of compromised clients, which participate in FL training at a fixed frequency. Following prior" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29166" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.159 + ], + "angle": 0, + "content": "Table 1. The effectiveness of our approach against the baseline attack across various state-of-the-art (SOTA) defenses under a fixed-frequency attack setting on non-IID datasets. BA values below \\(10\\%\\) in CIFAR10 (10 classes) and below \\(1\\%\\) in CIFAR100 (100 classes) are highlighted in red, indicating a failed attack. The highest BA achieved in each setting is presented in bold. Our attack results are reported as \\(a \\pm b\\), where \\(a\\) denotes the mean and \\(b\\) represents the standard deviation. MA and BA unit: %. Avg indicates the average. Each result is obtained as the average over five independent runs." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.162, + 0.878, + 0.543 + ], + "angle": 0, + "content": "
DefenseVGG19 (CIFAR10)ResNet18 (CIFAR10)ResNet18 (CIFAR100)
BadNetsDBALPOursBadNetsDBALPOursBadNetsDBALPOurs
No DefenseMA80.9781.2682.3881.81±0.3477.6678.2678.4378.70±0.8263.0562.465.1764.31±0.72
Best BA98.3346.1795.3897.05±0.1998.5920.9696.6997.60±0.21100.01.1672.2498.51±1.06
Avg BA97.7033.9690.5696.16±0.0997.9114.9395.0897.22±0.2199.730.8061.8796.70±2.46
MK [4]MA74.1775.4674.3078.03±0.3375.2476.2176.0776.49±0.8657.8458.6458.6558.73±0.55
Best BA24.063.1693.5498.01±0.806.047.5295.8998.61±0.140.880.9997.5299.94±0.01
Avg BA5.991.3376.4195.98±0.363.643.8592.4097.03±1.570.620.5489.2294.42±2.87
FLTrust [5]MA83.1483.2582.9883.84±0.2080.7580.2280.2280.33±0.2961.5861.8061.8961.14±1.81
Best BA56.236.5396.9697.95±0.3491.814.1897.1997.78±0.2730.120.6995.4299.98±0.05
Avg BA42.454.9196.6096.74±0.4789.123.6795.4891.94±8.4119.870.4687.5198.52±1.59
FLAME [23]MA58.0757.5258.6363.46±0.9174.2173.5271.9073.06±0.7655.9354.2855.6255.87±0.45
Best BA21.2037.8492.4398.38±0.6312.702.9794.4698.69±0.170.830.5377.6799.99±0.00
Avg BA7.7512.7678.8997.46±0.447.012.2789.7297.54±0.580.620.4472.8999.88±0.08
MM [12]MA69.3166.4370.7275.42±0.9771.4372.6172.8073.16±1.0753.0353.8253.4053.55±0.33
Best BA29.0816.0294.9598.31±0.4415.541.7196.0198.47±0.090.970.8592.7399.99±0.00
Avg BA7.045.4284.6194.55±4.203.911.3793.1297.52±0.670.370.6683.4697.05±1.58
DnC [29]MA81.0982.682.3281.30±1.4577.9879.3278.6378.83±0.3063.0161.4363.8963.43±0.73
Best BA5.942.8395.0497.85±0.276.5811.2096.2598.19±0.140.434.8880.9399.67±0.29
Avg BA3.681.6793.5796.94±0.534.039.7294.4597.29±0.370.412.6963.3798.07±0.59
RLR [25]MA72.2876.2277.7177.90±0.6875.7275.0674.5174.77±1.3754.7756.7856.1456.58±0.58
Best BA96.9323.0694.1197.78±0.1197.2917.1694.8398.39±0.130.140.3488.1299.98±0.01
Avg BA95.6515.1391.6197.39±0.2997.0210.8589.4197.95±0.170.050.0983.8199.97±0.02
FLARE [34]MA84.4584.2484.5384.56±0.1177.7978.2177.5277.82±2.0560.4456.5859.5459.07±1.41
Best BA98.2143.8794.7997.85±0.1298.185.8894.9196.75±0.5099.711.1488.5699.56±0.41
Avg BA98.1136.3294.1097.28±0.3098.025.4093.5596.30±0.8298.600.6380.7398.17±2.42
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.561, + 0.483, + 0.651 + ], + "angle": 0, + "content": "approaches [2, 41, 42], we assume that the attacker controls exactly one client per round and sets the attack interval to \\( F = 1 \\). This setup allows us to analyze the attack in isolation, without client collusion. We also evaluate the attack's performance across different frequencies, comparing it to the baseline." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.671, + 0.483, + 0.76 + ], + "angle": 0, + "content": "Baseline Defenses and Attacks. We assess our backdoor attack on eight state-of-the-art or representative defenses in FL: MultiKrum (MK) [4], FLTrust [5], FLAME [23], Multi-Metrics (MM) [12], DnC [29], RLR [25], FLARE [34], and DeepSight [28], ensuring a comprehensive evaluation of the attack's effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.761, + 0.484, + 0.807 + ], + "angle": 0, + "content": "We compare our backdoor attack against three state-of-the-art or representative attacks in FL: BadNets [10], DBA [35], and LP attack [42]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Metrics. Following prior work on FL backdoor attacks [10, 35, 42], we evaluate the global model using two primary metrics: main task accuracy (MA) and backdoor accuracy (BA). Given the dynamic nature of the global model in FL, we report both the average and the best BA over the last" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.561, + 0.906, + 0.681 + ], + "angle": 0, + "content": "10 communication rounds. Furthermore, to evaluate the stealthiness of the attack, we adopt two metrics introduced in [42]: the malicious client acceptance rate (MAR) and the benign client acceptance rate (BAR). MAR measures the proportion of rounds in which malicious clients successfully bypass defense and are selected for aggregation, while BAR represents the average proportion of rounds in which benign clients are selected for aggregation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.783, + 0.71 + ], + "angle": 0, + "content": "4.2. The effectiveness of our attack" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.907, + 0.869 + ], + "angle": 0, + "content": "We first evaluate our attack in the fixed-frequency attack setting under different SOTA defenses. Table 1 presents the performance comparison between our attack and the baseline attacks. We also evaluate our attack in the IID setting in supplementary §B.1 and on a larger dataset TinyImageNet in supplementary §B.11. The results show that our attack achieves the highest BA in most cases. These results demonstrate the broad effectiveness of our attack, effectively bypassing defenses while maintaining superior attack performance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "BadNets only works well under FLARE and consistently fails to bypass defenses such as MK, FLAME, MM, and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29167" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.09, + 0.483, + 0.117 + ], + "angle": 0, + "content": "Table 2. MAR and BAR comparison under outlier-based defenses: MK, FLAME, MM, and DnC." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.12, + 0.48, + 0.292 + ], + "angle": 0, + "content": "
Dataset (Model)AttackMKFLAMEMMDnC
BARMARBARMARBARMARBARMAR
CIFAR10 (ResNets18)BadNets0.420.010.750.010.330.010.970.01
DBA0.410.010.710.010.340.010.950.01
LP attack0.360.720.650.900.130.80.390.82
Ours0.380.680.640.880.250.790.780.99
CIFAR10 (VGG19)BadNets0.410.010.720.010.310.010.980.01
DBA0.420.010.750.010.330.030.970.01
LP attack0.350.630.610.750.130.80.410.99
Ours0.380.560.60.90.260.600.790.99
CIFAR100 (ResNets18)BadNets0.430.010.680.010.330.010.990.01
DBA0.420.010.760.010.330.010.970.01
LP attack0.320.910.560.990.120.930.350.94
Ours0.340.830.580.970.230.920.720.99
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.305, + 0.484, + 0.501 + ], + "angle": 0, + "content": "DnC across all settings, and it also fails against RLR on CIFAR100 in both non-IID and IID settings. Similarly, DBA also fails under several defenses, including MK, FLTrust, MM, and DnC under both non-IID and IID settings. LP attack achieves comparable BA across most defenses, except when evaluated under the FLARE defense in the IID setting. Specifically, for CIFAR100, our attack achieves significantly higher BA than the LP attack, demonstrating superior adaptability to complex data distributions. We also evaluate our attack performance compared with the LP attack under DeepSight in supplementary §B.3. Our attack achieves a higher BA than the LP attack across all scenarios and the LP attack fails on CIFAR100." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.523, + 0.484, + 0.613 + ], + "angle": 0, + "content": "Our attack improves the attack stealthiness and maintains strength. Under the FedAvg setting, our attack achieves significantly higher BA compared to both DBA and LP attack, particularly on CIFAR100, approaching the performance of unconstrained attacks like BadNets. This demonstrates the strong strength of our attack." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.824 + ], + "angle": 0, + "content": "Table 2 presents the comparison of MAR and BAR under outlier-based defenses such as MK, FLAME, MM, and DnC. A higher MAR indicates that malicious updates are more likely to bypass detection. We observe that BadNets and DBA fail to bypass defenses, as their MAR remains close to zero in all cases. This suggests that BadNets and DBA are consistently detected as outliers during training, rendering them ineffective against robust defenses. While BadNets demonstrates strong attack strength, its failure to bypass defenses undermines its overall effectiveness. In contrast, our attack achieves MAR consistently higher than BAR, meaning our malicious updates are more likely to be selected for aggregation, which reflects the stealthiness capability of our method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "When compared to the LP attack, our method maintains nearly the same or even higher MAR in defenses such as FLAME and DnC, further demonstrating its stealthiness. MK and MM impose stricter selection criteria by choosing only a small proportion of the most benign clients, making" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.228 + ], + "angle": 0, + "content": "it more challenging for an attacker to be selected. LP attack achieves a higher MAR than our method under these defenses by restricting the backdoor to BC layers, which significantly enhances its stealthiness. However, our method consistently achieves higher BA than LP attack across all settings in FLAME, DnC, MK, and MM, demonstrating that while LP attack sacrifices attack strength for stealthiness, our approach achieves a better balance, ensuring both strong attack strength and high stealthiness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.246, + 0.909, + 0.367 + ], + "angle": 0, + "content": "More evaluation of attack stealthiness. We leverage Krum distance, cosine distance, and Manhattan distance to further illustrate the stealthiness of our attack. Specifically, we calculate the sum of squared Euclidean distances, cosine distances, and Manhattan distances between a model update and other local model updates, referring to them as Krum distance, Cosine distance, and Manhattan distance, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.367, + 0.909, + 0.623 + ], + "angle": 0, + "content": "A larger Krum distance indicates that the update deviates significantly from benign updates, making it less likely to be accepted by the aggregation mechanism. Figure 2 shows the comparison of Krum distance with baseline attacks. The Krum distances of BadNets and DBA are significantly larger than those of benign updates, indicating that these attacks produce model updates that are easily distinguishable from benign ones. In contrast, our attack and the LP attack exhibit Krum distances close to those of benign updates, suggesting that they are more stealthy. Although DBA demonstrates slightly better stealthiness than BadNets, its Krum distance remains considerably larger than that of benign updates. This suggests that DBA is still insufficiently stealthy against robust defenses. In supplementary §B.4, we further discuss the Cosine distance and Manhattan distance between the malicious update and the benign update. The result exhibits similar trends across all attacks." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.634, + 0.908, + 0.661 + ], + "angle": 0, + "content": "Table 3. The BA of various backdoor attacks combined without/with ours." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.723 + ], + "angle": 0, + "content": "
AttackIBA [24]IBA+oursCerp [20]Cerp+oursA3FL [40]A3FL+ours
FLAME2.7198.04↑29.8499.64↑94.8799.87↑
MM31.6399.99↑24.7299.98↑86.2999.89↑
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.825 + ], + "angle": 0, + "content": "Our attack demonstrates strong generalizability. It can be easily integrated into existing FL backdoor attacks, particularly those based on trigger optimization. As shown in Table 3, incorporating our method further enhances attack effectiveness, underscoring its practical value." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.9 + ], + "angle": 0, + "content": "We further evaluate our attack trained on other networks: ResNet50 [11], ResNet101 [11] and ViT [6] in supplementary §B.5. Our attack achieves high BA on those networks, demonstrating its generalizability to more complex architectures." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29168" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.341, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.192, + 0.206, + 0.24, + 0.217 + ], + "angle": 0, + "content": "(a) Ours" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.089, + 0.61, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.45, + 0.205, + 0.521, + 0.217 + ], + "angle": 0, + "content": "(b) LP attack" + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.09, + 0.878, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.697, + 0.205, + 0.813, + 0.217 + ], + "angle": 0, + "content": "(c) BadNets and DBA" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.221, + 0.908, + 0.249 + ], + "angle": 0, + "content": "Figure 2. Krum distance of malicious and benign model updates. A larger distance indicates that the update deviates significantly from benign updates. The result shows the stealthiness of our attack." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.258, + 0.285, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.373, + 0.243, + 0.385 + ], + "angle": 0, + "content": "(a) MK (CIFAR-10)" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.259, + 0.482, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.321, + 0.373, + 0.452, + 0.385 + ], + "angle": 0, + "content": "(b) FLAME (CIFAR-10)" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.385, + 0.284, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.499, + 0.246, + 0.51 + ], + "angle": 0, + "content": "(c) MK (CIFAR-100)" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.385, + 0.477, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.313, + 0.499, + 0.451, + 0.511 + ], + "angle": 0, + "content": "(d) FLAME (CIFAR-100)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.515, + 0.483, + 0.556 + ], + "angle": 0, + "content": "Figure 3. Impacts of different attack intervals \\( F \\) on the attack performance. The \\( F \\) indicates that an attack is performed every \\( F \\) rounds." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.572, + 0.484, + 0.648 + ], + "angle": 0, + "content": "In addition, we analyze the impact of varying degrees of data heterogeneity in supplementary §B.6. Results show that our attack remains effective under highly non-IID conditions and consistently outperforms the LP attack across all scenarios, further confirming its generalizability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.67, + 0.485, + 0.761 + ], + "angle": 0, + "content": "Our attack maintains the global model's performance. Supplementary §B.2 shows the MA with the unpoisoned global model to assess whether our attack affects the utility of the global model. The results demonstrate that our attack does not lead to a significant decrease in MA compared to the unpoisoned global model." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.773, + 0.409, + 0.788 + ], + "angle": 0, + "content": "4.3. Our attack under different scenarios" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Impact of attack intervals. We evaluate the performance of our attack compared to baseline attacks under different attack intervals, as shown in Figure 3. The attack interval \\( F \\) ranges from 1 to 5, where \\( F \\) indicates that an attack is performed once every \\( F \\) rounds. This setting is designed to simulate scenarios where the frequency of malicious updates is limited, requiring more effective and persistent attack" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.259, + 0.707, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.558, + 0.373, + 0.666, + 0.385 + ], + "angle": 0, + "content": "(a) MK (CIFAR-10)" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.259, + 0.902, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.373, + 0.875, + 0.384 + ], + "angle": 0, + "content": "(b) FLAME (CIFAR-10)" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.385, + 0.706, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.555, + 0.499, + 0.669, + 0.511 + ], + "angle": 0, + "content": "(c) MK (CIFAR-100)" + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.385, + 0.898, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.736, + 0.499, + 0.874, + 0.511 + ], + "angle": 0, + "content": "(d) FLAME (CIFAR-100)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.515, + 0.908, + 0.557 + ], + "angle": 0, + "content": "Figure 4. Effectiveness of our attack under different proportions of compromised clients (\\(C = 0.02, 0.04, 0.06, 0.1\\)) in a fixed-pool attack setting." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.574, + 0.585, + 0.587 + ], + "angle": 0, + "content": "strategies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.589, + 0.909, + 0.83 + ], + "angle": 0, + "content": "The results show that as \\( F \\) increases, the performance of all attacks degrades as the attack frequency decreases. This trend is consistent across all evaluated methods, where a lower frequency results in reduced BA. A larger \\( F \\) imposes greater demands on the attack's strength, persistence, and stealthiness. The attack must be sufficiently powerful and stealthy to maintain its effectiveness with fewer opportunities to update the model. We observe that BadNets fails in all cases. Notably, our attack maintains a comparable BA even when \\( F = 5 \\). Whereas the LP attack loses nearly all effectiveness, particularly when trained on CIFAR100, demonstrating its vulnerability to defenses when attack opportunities are limited. We also show the results of BA under all defenses trained on CIFAR10 (ResNet18) in supplementary §B.8. We observe that our attack achieves comparable BA across all defenses when \\( F = 5 \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.91, + 0.901 + ], + "angle": 0, + "content": "Impact of different proportions of compromised clients. To evaluate our attack in a more practical scenario, we conduct experiments in a fixed-pool attack setting. To further" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29169" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.149, + 0.09, + 0.424, + 0.104 + ], + "angle": 0, + "content": "Table 4. Ablation study on alignment strategy." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.106, + 0.481, + 0.273 + ], + "angle": 0, + "content": "
Dataset (Model)AttackNo DefenseFLAMEMM
BABAMARBAMAR
CIFAR-10 (VGG19)no alignment97.707.050.017.040.01
model-wise90.1245.790.9083.580.65
layer-wise (ours)96.1697.460.9097.090.67
CIFAR-10 (ResNet18)no alignment97.917.010.013.910.01
model-wise93.3985.320.8995.530.80
layer-wise (ours)97.2198.160.8897.850.79
CIFAR-100 (ResNet18)no alignment99.730.620.010.370.01
model-wise73.2985.860.9595.050.90
layer-wise (ours)96.7099.880.9797.050.92
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.291, + 0.484, + 0.412 + ], + "angle": 0, + "content": "demonstrate the strength of our attack, we also analyze its performance under varying proportions of compromised clients \\( C \\). In a fixed-pool attack setting, the attacker controls a pool of compromised clients, from which participants are randomly selected in each round. This setup allows multiple malicious clients to be chosen in a single round, while in some cases, several consecutive rounds may pass without selecting any malicious clients." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.484, + 0.534 + ], + "angle": 0, + "content": "As shown in Figure 4, our attack achieves the highest BA compared to the baseline across all cases. Notably, even with a very low proportion of malicious clients (\\(C = 0.02\\), meaning only \\(2\\%\\) of clients are compromised), our attack successfully bypasses defenses and maintains a comparable BA. This demonstrates that our attack is both powerful and persistent, as it continues to perform well even when the pool of malicious clients is small." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.484, + 0.732 + ], + "angle": 0, + "content": "Additionally, BadNets performs better when \\( C = 0.1 \\) due to the increased likelihood of malicious clients colluding in the same round. This collusion significantly enhances the attack's effectiveness, as multiple attackers working together are more likely to successfully inject a backdoor. This is particularly evident under MK, where BadNets achieves a higher BA with a higher \\( C \\), compared to when it fails in the fixed-frequency setting. However, this improvement diminishes as the compromised client ratio decreases. At lower \\( C \\), the attack becomes more reliant on its strength. In contrast, as shown in supplementary §B.9, our attack achieves high BA across all defenses when \\( C > 0.04 \\), demonstrating its robustness even with fewer compromised clients." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.241, + 0.759 + ], + "angle": 0, + "content": "4.4. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Impact of different triggers. We further evaluate the impact of various trigger shapes. Supplementary §B.10 presents three different trigger shapes (Apple, Watermark, and Square) applied to backdoor a ResNet18 model trained on the CIFAR10 dataset. Figure 5a shows the impact of various trigger shapes. Under the default setting \\((F = 1)\\), all three trigger shapes achieve high BA. The results demonstrate that our attack remains effective across different trigger shapes, indicating its robustness and generalizability." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.088, + 0.707, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.203, + 0.68, + 0.216 + ], + "angle": 0, + "content": "(a) various trigger shapes" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.089, + 0.9, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.203, + 0.868, + 0.215 + ], + "angle": 0, + "content": "(b) various trigger sizes" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.218, + 0.905, + 0.245 + ], + "angle": 0, + "content": "Figure 5. Impact of trigger shape (left) and trigger size (right) in Flame." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.272, + 0.907, + 0.379 + ], + "angle": 0, + "content": "Figure 5b illustrates the performance of different trigger sizes (3, 5, 7, and 10). Regardless of trigger size, our attack consistently achieves high BA. Notably, when the trigger size is set to 10, the attack effectiveness is significantly enhanced. Larger triggers strengthen the attack, and their ability to bypass defenses is improved through layer-wise norm constraints, leading to superior attack performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.419, + 0.908, + 0.541 + ], + "angle": 0, + "content": "Impact of alignment strategy. To demonstrate how dynamic layer-wise gradient alignment enhances the effectiveness of our attack, we compare it with BadNets (no alignment) and the model-wise gradient alignment attack, which adjusts the overall gradient magnitude to align with the previous global update. As shown in Table 4, our attack consistently outperforms both BadNets and the model-wise method across all settings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.543, + 0.909, + 0.68 + ], + "angle": 0, + "content": "First, both model-wise and layer-wise gradient alignment show higher MAR compared to BadNets, indicating that aligning the update with the previous global model significantly improves the stealthiness of the attack. Second, when compared to model-wise alignment, layer-wise alignment achieves better attack performance. Specifically, even in the absence of an attack, particularly on CIFAR100, the BA for model-wise alignment is much lower than that for layer-wise alignment." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.706, + 0.634, + 0.721 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.901 + ], + "angle": 0, + "content": "This paper proposes a new stealthy backdoor attack in FL, which leverages adaptive layer-wise gradient alignment to enhance stealthiness while preserving attack strength. By dynamically aligning malicious gradients to historical global updates at the layer level, the proposed attack reduces statistical anomalies caused by malicious updates, making them harder to detect. Extensive experiments across diverse datasets, models, and defenses demonstrate that the proposed attack achieves superior backdoor success rates while evading SOTA defenses, highlighting its effectiveness and stealthiness in various FL scenarios." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29170" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.486, + 0.173 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China (NO. 62472284, No. 62302288), Shanghai Key Laboratory of Scalable Computing and Systems, SJTU Kunpeng & Ascend Center of Excellence." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.198, + 0.188, + 0.214 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.223, + 0.484, + 0.264 + ], + "angle": 0, + "content": "[1] Sana Awan, Bo Luo, and Fengjun Li. CONTRA: Defending Against Poisoning Attacks in Federated Learning. In ESORICS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.266, + 0.484, + 0.307 + ], + "angle": 0, + "content": "[2] Eugene Bagdasaryan, Andreas Veit, Yiqing Hua, Deborah Estrin, and Vitaly Shmatikov. How To Backdoor Federated Learning. In AISTATS, 2020. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.309, + 0.484, + 0.362 + ], + "angle": 0, + "content": "[3] Akshay Batheja and Pushpak Bhattacharyya. \"A Little is Enough\": Few-Shot Quality Estimation based Corpus Filtering improves Machine Translation. In ACL (Findings), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.365, + 0.484, + 0.419 + ], + "angle": 0, + "content": "[4] Peva Blanchard, El Mahdi El Mhamdi, Rachid Guerraoui, and Julien Stainer. Machine Learning with Adversaries: Byzantine Tolerant Gradient Descent. In NeurIPS, 2017. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.422, + 0.483, + 0.463 + ], + "angle": 0, + "content": "[5] Xiaoyu Cao, Minghong Fang, Jia Liu, and Neil Zhenqiang Gong. FLTrust: Byzantine-robust Federated Learning via Trust Bootstrapping. In NDSS, 2021. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.464, + 0.484, + 0.546 + ], + "angle": 0, + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.548, + 0.484, + 0.589 + ], + "angle": 0, + "content": "[7] Minghong Fang, Xiaoyu Cao, Jinyuan Jia, and Neil Gong. Local Model Poisoning Attacks to Byzantine-Robust Federated Learning. In USENIX Security, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.591, + 0.484, + 0.618 + ], + "angle": 0, + "content": "[8] Pei Fang and Jinghui Chen. On the Vulnerability of Backdoor Defenses for Federated Learning. In AAAI, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.62, + 0.484, + 0.659 + ], + "angle": 0, + "content": "[9] Clement Fung, Chris JM Yoon, and Ivan Beschastnikh. The Limitations of Federated Learning in Sybil Settings. In RAID, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.662, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[10] Tianyu Gu, Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. BadNets: Evaluating Backdooring Attacks on Deep Neural Networks. IEEE Access, 7:47230-47244, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.704, + 0.484, + 0.745 + ], + "angle": 0, + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In CVPR, 2016. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.484, + 0.788 + ], + "angle": 0, + "content": "[12] Siquan Huang, Yijiang Li, Chong Chen, Leyu Shi, and Ying Gao. Multi-Metrics Adaptively Identifies Backdoors in Federated Learning. In ICCV, 2023. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.484, + 0.844 + ], + "angle": 0, + "content": "[13] Tiansheng Huang, Sihao Hu, Ka Ho Chow, Fatih Ilhan, Selim F. Tekin, and Ling Liu. Lockdown: Backdoor Defense for Federated Learning with Isolated Subspace Training. In NeurIPS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[14] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and Open Problems in Federated" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.223, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.906, + 0.12 + ], + "angle": 0, + "content": "Learning. Foundations and trends® in machine learning, 14 (1-2):1-210, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.906, + 0.164 + ], + "angle": 0, + "content": "[15] Yun-Yong Ko, Dongwon Lee, and Sang-Wook Kim. Not All Layers Are Equal: A Layer-Wise Adaptive Approach Toward Large-Scale DNN Training. In WWW, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.906, + 0.205 + ], + "angle": 0, + "content": "[16] Jakub Konečny. Federated learning: Strategies for improving communication efficiency. arXiv preprint arXiv:1610.05492, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.906, + 0.248 + ], + "angle": 0, + "content": "[17] Torsten Krauß and Alexandra Dmitrienko. MESAS: Poisoning Defense for Federated Learning Resilient against Adaptive Attackers. In CCS, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.906, + 0.305 + ], + "angle": 0, + "content": "[18] Haoyang Li, Qingqing Ye, Haibo Hu, Jin Li, Leixia Wang, Chengfang Fang, and Jie Shi. 3DFed: Adaptive and Extensible Framework for Covert Backdoor Attack in Federated Learning. In S&P, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.906, + 0.349 + ], + "angle": 0, + "content": "[19] Ye Li, Yanchao Zhao, Chengcheng Zhu, and Jiale Zhang. Infighting in the dark: Multi-labels backdoor attack in federated learning. arXiv preprint arXiv:2409.19601, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.35, + 0.906, + 0.405 + ], + "angle": 0, + "content": "[20] Xiaoting Lyu, Yufei Han, Wei Wang, Jingkai Liu, Bin Wang, Jiqiang Liu, and Xiangliang Zhang. Poisoning with Cerberus: Stealthy and Colluded Backdoor Attack against Federated Learning. In AAAI, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.407, + 0.906, + 0.462 + ], + "angle": 0, + "content": "[21] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y Arcas. Communication-Efficient Learning of Deep Networks from Decentralized Data. In AISTATS, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.906, + 0.504 + ], + "angle": 0, + "content": "[22] El Mahdi El Mhamdi, Rachid Guerraoui, and Sébastien Rouault. The Hidden Vulnerability of Distributed Learning in Byzantium. In ICML, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.506, + 0.906, + 0.589 + ], + "angle": 0, + "content": "[23] Thien Duc Nguyen, Phillip Rieger, Huili Chen, Hossein Yalame, Helen Möllering, Hossein Fereidooni, Samuel Marchal, Markus Miettinen, Azalia Mirhoseini, Shaza Zeitouni, Farinaz Koushanfar, Ahmad-Reza Sadeghi, and Thomas Schneider. FLAME: Taming Backdoors in Federated Learning. In USENIX Security 22, 2022. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.906, + 0.631 + ], + "angle": 0, + "content": "[24] Thuy Dung Nguyen, Tuan Nguyen, Anh Tran, Khoa D. Doan, and Kok-Seng Wong. IBA: Towards Irreversible Backdoor Attacks in Federated Learning. In NeurIPS, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.906, + 0.675 + ], + "angle": 0, + "content": "[25] Mustafa Safa Ozdayi, Murat Kantarcioglu, and Yulia R Gel. Defending against backdoors in federated learning with robust learning rate. In AAAI, 2021. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.906, + 0.731 + ], + "angle": 0, + "content": "[26] Ashwinee Panda, Saeed Mahloujifar, Arjun Nitin Bhagoji, Supriyo Chakraborty, and Prateek Mittal. SparseFed: Mitigating Model Poisoning Attacks in Federated Learning with Sparsification. In AISTATS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.733, + 0.906, + 0.802 + ], + "angle": 0, + "content": "[27] Matthias Paulik, Matt Seigel, Henry Mason, Dominic Telaar, Joris Kluivers, Rogier Van Dalen, Chi Wai Lau, Luke Carlson, Filip Granqvist, Chris Vandevelde, et al. Federated Evaluation and Tuning for On-Device Personalization: System Design & Applications. arXiv preprint arXiv:2102.08503, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.906, + 0.858 + ], + "angle": 0, + "content": "[28] Phillip Rieger, Thien Duc Nguyen, Markus Miettinen, and Ahmad-Reza Sadeghi. DeepSight: Mitigating Backdoor Attacks in Federated Learning Through Deep Model Inspection. In NDSS, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[29] Virat Shejwalkar and Amir Houmansadr. Manipulating the Byzantine: Optimizing Model Poisoning Attacks and Defenses for Federated Learning. In NDSS, 2021. 1, 2, 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29171" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[30] Karen Simonyan. Very Deep Convolutional Networks for Large-Scale Image Recognition. arXiv preprint arXiv:1409.1556, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[31] Peng Sun, Xinyang Liu, Zhibo Wang, and Bo Liu. Byzantine robust Decentralized Federated Learning via Dual-domain Clustering and Trust Bootstrapping. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[32] Ziteng Sun, Peter Kairouz, Ananda Theertha Suresh, and H Brendan McMahan. Can You Really Backdoor Federated Learning? arXiv preprint arXiv:1911.07963, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.483, + 0.289 + ], + "angle": 0, + "content": "[33] Hongyi Wang, Kartik Sreenivasan, Shashank Rajput, Harit Vishwakarma, Saurabh Agarwal, Jy-yong Sohn, Kangwook Lee, and Dimitris S. Papailiopoulos. Attack of the Tails: Yes, You Really Can Backdoor Federated Learning. In NeurIPS, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[34] Ning Wang, Yang Xiao, Yimin Chen, Yang Hu, Wenjing Lou, and Y. Thomas Hou. FLARE: Defending Federated Learning against Model Poisoning Attacks via Latent Space Representations. In AsiaCCS, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.483, + 0.388 + ], + "angle": 0, + "content": "[35] Chulin Xie, Keli Huang, Pin-Yu Chen, and Bo Li. DBA: Distributed Backdoor Attacks against Federated Learning. In ICLR, 2020. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.483, + 0.445 + ], + "angle": 0, + "content": "[36] Yueqi Xie, Minghong Fang, and Neil Zhenqiang Gong. FedREDefence: Defending against Model Poisoning Attacks for Federated Learning using Model Update Reconstruction Error. In ICML, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.483, + 0.489 + ], + "angle": 0, + "content": "[37] Jiahao Xu, Zikai Zhang, and Rui Hu. Identify Backdoored Model in Federated Learning via Individual Unlearning. arXiv preprint arXiv:2411.01040, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.483, + 0.544 + ], + "angle": 0, + "content": "[38] Peishen Yan, Hao Wang, Tao Song, Yang Hua, Ruhui Ma, Ningxin Hu, Mohammad Reza Haghighat, and Haibing Guan. SkyMask: Attack-Agnostic Robust Federated Learning With Fine-Grained Learnable Masks. In ECCV, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.483, + 0.614 + ], + "angle": 0, + "content": "[39] Timothy Yang, Galen Andrew, Hubert Eichner, Haicheng Sun, Wei Li, Nicholas Kong, Daniel Ramage, and Françoise Bea-fays. Applied Federated Learning: Improving Google Keyboard Query Suggestions. arXiv preprint arXiv:1812.02903, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.483, + 0.659 + ], + "angle": 0, + "content": "[40] Hangfan Zhang, Jinyuan Jia, Jinghui Chen, Lu Lin, and Ding-hao Wu. A3FL: Adversarily Adaptive Backdoor Attacks to Federated Learning. In NeurIPS, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.483, + 0.728 + ], + "angle": 0, + "content": "[41] Zhengming Zhang, Ashwinee Panda, Linyue Song, Yaoqing Yang, Michael W. Mahoney, Prateek Mittal, Kannan Ramchandran, and Joseph Gonzalez. Neurotoxin: Durable Backdoors in Federated Learning. In ICML, 2022. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.731, + 0.483, + 0.772 + ], + "angle": 0, + "content": "[42] Haomin Zhuang, Mingxian Yu, Hao Wang, Yang Hua, Jian Li, and Xu Yuan. Backdoor Federated Learning by Poisoning Backdoor-Critical Layers. In ICLR, 2024. 1, 2, 3, 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.772 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29172" + } + ] +] \ No newline at end of file diff --git a/2025/Stealthy Backdoor Attack in Federated Learning via Adaptive Layer-wise Gradient Alignment/f7dfb2e6-c547-4b0c-bb2e-4bfc0b02a71f_origin.pdf b/2025/Stealthy Backdoor Attack in Federated Learning via Adaptive Layer-wise Gradient Alignment/f7dfb2e6-c547-4b0c-bb2e-4bfc0b02a71f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..06dc5e77423a6682fac956510d08e1313016b8b4 --- /dev/null +++ b/2025/Stealthy Backdoor Attack in Federated Learning via Adaptive Layer-wise Gradient Alignment/f7dfb2e6-c547-4b0c-bb2e-4bfc0b02a71f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e2e647eecc77b6c514dd83d7fbf032a170a923056486552d067e2ea6b5e30a6d +size 792042 diff --git a/2025/Stealthy Backdoor Attack in Federated Learning via Adaptive Layer-wise Gradient Alignment/full.md b/2025/Stealthy Backdoor Attack in Federated Learning via Adaptive Layer-wise Gradient Alignment/full.md new file mode 100644 index 0000000000000000000000000000000000000000..36244c3d58160e4ea9f6fb03ea653ad3dae16504 --- /dev/null +++ b/2025/Stealthy Backdoor Attack in Federated Learning via Adaptive Layer-wise Gradient Alignment/full.md @@ -0,0 +1,320 @@ +# Stealthy Backdoor Attack in Federated Learning via Adaptive Layer-wise Gradient Alignment + +Qingqian Yang $^{1*}$ Peishen Yan $^{2}$ Xiaoyu Wu $^{2}$ Jiaru Zhang $^{2}$ Tao Song $^{2\dagger}$ Yang Hua $^{3}$ Hao Wang $^{4}$ Liangliang Wang $^{1\dagger}$ Haibing Guan $^{2}$ $^{1}$ Shanghai University of Electric Power $^{2}$ Shanghai Jiao Tong University + $^{3}$ Queen's University Belfast $^{4}$ Stevens Institute of Technology + +y23108013@mail.shiep.edu.cn + +{peishenyan, wuxiaoyu2000, jiaruzhang, songt333, hbguan}@sjtu.edu.cn + +Y.Hua@qub.ac.uk hwang9@stevens.edu llwang@shiep.edu.cn + +# Abstract + +The distributed nature of federated learning exposes it to significant security threats, among which backdoor attacks are one of the most prevalent. However, existing backdoor attacks face a trade-off between attack strength and stealthiness: attacks maximizing the attack strength are often detectable, while stealthier approaches significantly reduce the effectiveness of the attack itself. Both of them result in ineffective backdoor injection. In this paper, we propose an adaptive layer-wise gradient alignment strategy to effectively evade various robust defense mechanisms while preserving attack strength. Without requiring additional knowledge, we leverage the previous global update as a reference for alignment to ensure stealthiness during dynamic FL training. This fine-grained alignment strategy applies appropriate constraints to each layer, which helps significantly maintain attack strength. To demonstrate the effectiveness of our method, we conduct extensive evaluations across a wide range of datasets and networks. Our experimental results show that the proposed attack effectively bypasses eight state-of-the-art defenses and achieves high backdoor accuracy, outperforming existing attacks by up to $54.76\%$ . Additionally, it significantly preserves attack strength and maintains robust performance across diverse scenarios, highlighting its adaptability and generalizability. Code implementation is available at https://github.com/yqqhyqq/LGA. + +# 1. Introduction + +Federated Learning (FL) has emerged as a promising paradigm for privacy-preserving distributed machine learn + +ing, allowing multiple decentralized clients to collaboratively train a global model without directly sharing their raw data [14, 21]. FL leverages local computation on edge devices (e.g., smartphones and IoT sensors) and aggregates model updates via a central server, making it widely adopted in various real-world applications [27, 39]. However, the distributed architecture of FL systems significantly increases the potential attack surface, introducing a wide range of security threats, among which backdoor attacks are a commonly studied threat in recent research. In a backdoor attack [2, 32, 33, 35, 41, 42], adversarial clients intentionally inject a backdoor into the global model by uploading a poisoning local update. As a result, the global model behaves normally on benign inputs but misclassifies triggered inputs into predefined target labels. + +Existing backdoor defenses primarily focus on detecting and filtering malicious updates by leveraging information from multiple dimensions. Specifically, some methods rely on distance-based measures between benign and malicious updates [4, 5, 12, 23], while others detect inconsistencies in parameter signs [25] or exploit the spectral separability [29]. These defenses significantly restrict the feasibility of backdoor attacks. + +To counter these defenses, an effective backdoor attack depends on two key factors: attack strength and attack stealthiness. Some attacks [2] achieve high attack strength by scaling up the update, which increases their detectability and makes them easily filtered by defense mechanisms. Others [35, 42] prioritize stealthiness to evade detection but show significantly lower performance when no defense is present, highlighting their sacrifice of attack strength. These attacks essentially constrain the malicious update by discarding certain backdoor-related information [42] or using coarse-grained scaling [33]. Although constraining malicious updates helps maintain stealthiness, this inappropriate + +constraint leads to attack strength degradation. Therefore, balancing the attack stealthiness and attack strength with limited knowledge of the system's defenses is challenging. + +To address the challenge, we propose a layer-wise adaptive gradient alignment strategy that dynamically adjusts the norm bound for each layer of malicious updates. This adjustment enhances their similarity to the benign update distribution in a fine-grained manner, thereby achieving both attack stealthiness and attack effectiveness. Without requiring any additional information, our approach uses the previous global update as a reference point for adaptive alignment. By doing so, the proposed strategy not only constrains the malicious update within a small norm ball, but also enables the attack to adapt to the evolving global model and training dynamics, thereby significantly improving its stealthiness. Inspired by previous observations [15] that parameters within the same layer tend to share similar gradient magnitudes, we further conduct empirical analysis revealing that the magnitudes of both malicious and benign updates vary across layers, and there is no consistent proportional relationship between them. This discrepancy highlights the need for layer-specific norm constraints. Such constraints are essential to prevent over-constraining layers with inherently smaller magnitudes, thereby preserving attack strength. + +Extensive experiments validate the effectiveness of our attack, demonstrating its ability to evade eight state-of-the-art defenses while maintaining a high attack success rate. Additionally, we investigate its performance across various attack scenarios, demonstrating its generalizability. We summarize our main contributions as follows: + +- We propose an adaptive layer-wise gradient alignment strategy to effectively evade various robust defense mechanisms while preserving attack strength. Without requiring any additional knowledge, we leverage the previous global update as a reference for gradient alignment, ensuring that the attack maintains the stealthiness while adapting to the training dynamics. This fine-grained alignment strategy applies layer-specific constraints, allowing the attack to retain its strength. +- We conduct extensive evaluations across diverse models and datasets, demonstrating that the proposed attack successfully bypasses eight SOTA defenses and maintains the attack impact. Our backdoor attack outperforms SOTA attacks, with an improvement in backdoor accuracy by up to $54.76\%$ . Furthermore, our attack exhibits excellent performance across various scenarios, such as different attack patterns, highlighting its generalizability. + +# 2. Related Work + +Backdoor attacks in FL. Backdoor attacks present a serious security threat in FL [16]. Typically, attackers compromise a small subset of client devices to upload poisoned updates. These updates are crafted to inject the backdoor into + +the global model, which then enables the model to obtain predetermined predictions for samples with specific trigger features. The backdoor attack process generally consists of two stages: trigger embedding and model manipulation. To successfully inject a backdoor, the attacker can carefully design either the trigger or the model update to bypass detection. Accordingly, existing methods can be divided into two categories: model-optimization attacks [2, 18, 41, 42] and trigger-optimization attacks [8, 20, 24, 33, 35, 40]. + +Model-optimization attacks [2, 3, 18, 33, 41, 42] focus on fine-tuning the model to achieve the desired attack. Scaling attack [2] amplifies malicious updates to achieve a one-shot attack. However, the abnormally large gradient magnitudes make it exposed to detection. Neurotoxin [41] projects malicious gradients onto infrequently updated neurons, improving the persistence of the attack and preventing the backdoor from being overwritten by benign updates. However, as an unconstrained backdoor attack, it's vulnerable to robust detection. The LP attack [42] adaptively injects the backdoor into backdoor-critical (BC) layers, achieving high stealthiness. However, by discarding backdoor-related information in non-BC layers, the attack suffers from reduced strength. PGD attack [33] exploits knowledge of server-side defenses to craft malicious updates that bypass detection. However, this strategy is only effective when the attacker has access to detailed information about the system. Additionally, 3DFed [18] introduces three distinct attack strategies targeting different types of defenses and utilizes an indicator for adaptive strategy adjustments. However, it relies on client collaboration. + +Trigger-optimization backdoor attacks aim to design triggers that are both subtle and adaptable, ensuring their stealthiness and effectiveness. Edge-case attack [33] selects marginal data as poison samples to circumvent detection mechanisms. DBA [35] firstly introduces distributed backdoor attacks, decomposing a global trigger into smaller, less detectable components. These approaches craft triggers at the pre-training stage and then fix them during training. Recently, a range of studies [8, 19, 20, 24, 40] dynamically optimize the trigger throughout the training process and obtain better attack results than the former ones. + +Defense strategies against backdoor attacks in FL. In FL [16], the server only receives uploaded models without access to any information about client data. As a result, the majority of defense strategies are similarity-based, aiming to detect and mitigate backdoor attacks by identifying differences between malicious and benign models. + +Among them, distance-based defenses [1, 4, 5, 9, 12, 17, 22, 23, 28, 29, 31, 34] exploit the observation that malicious models often deviate from benign ones in their distribution within high-dimensional space, making distance-based metrics an effective tool for detection. MultiKrum [4] + +only aggregates updates that are selected with the smallest pairwise Euclidean distances. FLTrust [5] weights upload gradients based on their cosine similarity to a trusted gradient to mitigate the influence of malicious models. The trusted gradient is obtained from a root dataset collected by the server. FLAME [23] and DeepSight [28] use pairwise cosine distance between client gradients for clustering, identifying malicious gradients as outliers and further refining the defense through additional modules. Multi-Metrics [12] dynamically adjusts the weights of three distance metrics to detect backdoor models. MESAS [17] leverages six distance metrics for detection. + +Other similarity-based defenses [13, 26, 29, 36, 38] identify malicious updates from more diverse perspectives. DnC [29] leverages singular value decomposition (SVD)-based spectral methods to remove outliers. RLR [25] detects inconsistencies in parameter signs across training rounds and adjusts learning rates accordingly to suppress backdoors. MASA [37] leverages the observation that a malicious model unlearns the main task more quickly than a benign model for detection. FedREDefense [36] relies on the finding that a malicious model is more resistant to distillation. In addition, some methods recently focus on the difference in update behavior between benign models and malicious models at the parameter level. For instance, SparseFed [26] only aggregates Top-k highest magnitude parameters because the backdoor updates are usually injected into coordinates that are unimportant for benign updates. Meanwhile, Lockdown [13] restrains local training in a subspace to further reduce the poison-coupling effect. + +# 3. Methodology + +Background. Federated Learning (FL) distributes the training task across multiple clients to preserve data privacy. At the beginning, a global model $\theta_g^0$ is initialized on the server and distributed to a set of $N$ clients. In each round $t$ , a standard FL framework involves the following steps: + +1. Local Training: Each client $i$ receives the global model $\theta_{g}^{t - 1}$ and performs $E$ local epochs of stochastic gradient descent (SGD) on its private dataset $\hat{\mathcal{D}}_i$ with batch size $B$ , obtaining an updated model $\theta_{i}^{t}$ . The client computes the model update as: + +$$ +\Delta \theta_ {i} ^ {t} = \theta_ {i} ^ {t} - \theta_ {g} ^ {t - 1}, \tag {1} +$$ + +and uploads it to the server. + +2. Global aggregation: The server randomly selects a subset of clients and aggregates their model updates $\{\Delta \theta^t\}$ to compute the global update: + +$$ +\Delta \theta_ {g} ^ {t} = \sum_ {i \in [ n ]} \Delta \theta_ {i} ^ {t}, \tag {2} +$$ + +and updates the global model as $\theta_g^t = \theta_g^{t - 1} + \Delta \theta_g^t$ . The new global model is then distributed to clients for the + +next round. Before updating the global model, the server typically performs backdoor detection during aggregation to mitigate the impact of malicious updates. + +In FL, backdoor attacks exploit the system's decentralized nature. Specifically, an attacker, denoted as $A$ , compromises a subset of clients and injects malicious updates to implant a backdoor into the global model. We consider a continuous attack scenario, following prior studies [20, 35, 42]. + +Adversarial Objectives: The attacker $A$ aims to inject a backdoor into the global model while minimizing the impact on the model's performance on non-poisoned data. The goal of the backdoor injection is to manipulate the model so that it misclassifies inputs with a specific trigger pattern as an incorrect target label. We denote the malicious update as $\Delta \theta_{m}$ , which is trained on the poisoned dataset. By carefully crafting $\Delta \theta_{m}$ , the attacker can introduce the backdoor without significantly degrading the model's overall accuracy. The server's aggregation mechanism, which aggregates updates from multiple clients, may unintentionally include these poisoned updates, thereby corrupting the global model. + +Adversarial Capability: The attacker $A$ has the ability to control a small subset of clients and gain access to their datasets. Additionally, the attacker can manipulate the models of these compromised clients [2, 35, 41]. The attacker does not have access to the data or models of benign clients, nor to the specific defense methods implemented by the defender. + +Overview. We introduce an adaptive layer-wise gradient alignment to enhance stealthiness while preserving attack strength. In each round, we train the malicious update on a poisoned local dataset and iteratively align gradients to resemble benign updates before uploading them to the server. This increases the likelihood of the server aggregating the malicious update, effectively injecting the backdoor into the global model. + +Adaptive reference selection. Traditionally, the Unconstrained malicious update $\Delta \theta_{m}$ , which is trained on the poisoned dataset including out-of-distribution backdoor samples, tends to exhibit a larger magnitude compared to benign updates [32]. To ensure that the malicious update remains similar to the distribution of benign updates, it should be carefully constrained within an appropriately small norm ball. Notably, the selection of the norm bound significantly impacts the effectiveness of the attack's concealment. Therefore, it is crucial to rigorously determine an appropriate reference point for alignment. Some defense methods, such as MultiKrum [4], select a fixed number of client updates that appear most benign in each round. However, due to data heterogeneity, such methods may inadvertently exclude genuinely benign clients whose updates deviate from the majority. To mitigate this issue, we adopt the previous global + +![](images/299b4ddf2a78efc4cc680d2adcb185cccf9d779d87cdda40acbe2c18330a06e3.jpg) +Figure 1. Layer-wise Magnitude Analysis of Benign vs. Malicious Updates in VGG19. To better visualize, we focus on weight layers and divide them into A-layers and B-layers based on $\|\Delta\theta^l\|$ : those with norms above 0.5 are A-layers; others are B-layers. The experiment uses the default setting and records $\|\Delta\theta^l\|$ every 50 rounds. + +update $\Delta \theta_{g}^{t - 1}$ as a reference point. Since it is computed by aggregating multiple client updates, most of which are expected to be benign, it serves as a reasonable approximation of benign behavior. + +By aligning the malicious update with the global update, the attack remains within the expected range of benign variation, thereby enhancing its stealthiness and increasing the likelihood of successfully influencing the global model. + +Layer-wise gradient alignment. Intuitively, applying a norm constraint weakens the impact of the malicious update, which potentially leads to ineffective backdoor injection. We introduce a layer-wise scaling strategy to refine the malicious update at a more granular level to not only ensure the attack remains stealthy but also preserve its strength. A key observation shown in Figure 1 is that the layer-wise magnitudes of the malicious and benign updates do not maintain a fixed proportional relationship; instead, their patterns differ. Therefore, we align each layer individually with $\Delta \theta_{g}^{t - 1}$ . Specifically, we define the scaling factor $S^{t,l}$ for each layer $l$ as: + +$$ +S ^ {t, l} = \min (1, \frac {\| \Delta \theta_ {g} ^ {t - 1 , l} \|}{\| \Delta \theta_ {m} ^ {t , l} \|}). \tag {3} +$$ + +The scaled malicious update after each local epoch is given by: + +$$ +\Delta \theta_ {m} ^ {t, l} = S ^ {t, l} \cdot \Delta \theta_ {m} ^ {t, l}. \tag {4} +$$ + +The adaptive layer-wise gradient alignment, implemented in just four lines (Lines 10-13 in Algorithm 1), extends the standard FL backdoor pipeline. This approach prevents excessive constraints on layers with inherently smaller magnitudes and constrains the update for each layer within an $\ell 2$ norm ball, thereby preserving attack strength while enhancing stealthiness. + +# 4. Experiments + +Our empirical study evaluates the effectiveness of the proposed backdoor attack through experiments conducted on + +Algorithm 1 Layer-wise Gradient Alignment for Attackers +Require: Batch size $\ell$ , local epochs $E$ , malicious model parameters $\theta_{i}^{t}$ , global parameters $\theta_{g}^{t-1}$ , poisoned dataset $\hat{D}$ , and previous global parameter $\theta_{g}^{t-2}$ + +# Procedure: + +1: // Attacker's local training: +2: Calculate the previous global update: +3: $\Delta \theta_{g}^{t - 1}\gets \theta_{g}^{t - 1} - \theta_{g}^{t - 2}$ +4: Initialize local model $\theta_{i,0}^{t}\gets \theta_{g}^{t - 1}$ +5: for $e = 1$ to $E$ do + +6: for each batch $b$ in $\hat{\mathcal{D}}$ do + +7: Compute gradient for batch $b$ : +8: $\Delta \theta_{i,e}^{t}\gets \bar{\Delta}\theta_{i,e}^{t} + \nabla_{\theta}\mathcal{L}(\theta_{i,e - 1}^{t},\hat{\mathcal{D}}_{b})$ +9: end for + +10: // Layer-wise gradient alignment +11: for each layer $l$ in $\Delta \theta_{i,e}^{t}$ do +12: $S_{i,e}^{t,l}\gets min(1,\frac{\|\Delta\theta_{g}^{t - 1,l}\|}{\|\Delta\theta_{i,e}^{t,l}\|})$ +13: $\theta_{i,e}^{t,l}\gets \Delta \theta_{i,e}^{t,l}\cdot S_{i,e}^{t,l} + \theta_{g}^{t - 1,l}$ +14: end for +15: end for +16:Upload update $\Delta \theta_{i}^{t}$ to server + +real-world datasets, simulating an FL environment using the NVIDIA RTX 4090 GPU. + +# 4.1. Experimental setup + +Datasets and Models. We evaluate the attack using two widely used benchmark datasets: CIFAR10 (50,000 training and 10,000 test samples across 10 classes) and CIFAR100 (50,000 training and 10,000 test samples across 100 classes). Following previous studies [41, 42], we use the VGG19 [30] network and ResNet18 network for CIFAR10, while we use the ResNet18 [11] network for CIFAR100. We simulate a non-IID data distribution following prior methods [5, 7, 42], with $q = 0.5 / 0.2$ for CIFAR10/100, respectively. + +Attack setup. In each round, we select $n = 10$ clients among $N = 100$ clients to participate in the aggregation, with the proportion of compromised clients set to $C = 0.1$ by default. The FL training runs for 300 rounds to make it converge. We train for 6 local epochs (2 for the benign clients) with a learning rate 0.1 for both CIFAR10 and CIFAR100. Additionally, we assess the effectiveness of our attack in an IID setting, as defense strategies are more likely to identify malicious models as outliers in this case [42]. The trigger is a $5 \times 5$ pixel square located at the bottom-right corner of each image by default. We further discuss the impact of various shapes and sizes of the trigger. We evaluate our attack in the fixed-frequency attack scenario. The attacker controls a fixed number of compromised clients, which participate in FL training at a fixed frequency. Following prior + +Table 1. The effectiveness of our approach against the baseline attack across various state-of-the-art (SOTA) defenses under a fixed-frequency attack setting on non-IID datasets. BA values below $10\%$ in CIFAR10 (10 classes) and below $1\%$ in CIFAR100 (100 classes) are highlighted in red, indicating a failed attack. The highest BA achieved in each setting is presented in bold. Our attack results are reported as $a \pm b$ , where $a$ denotes the mean and $b$ represents the standard deviation. MA and BA unit: %. Avg indicates the average. Each result is obtained as the average over five independent runs. + +
DefenseVGG19 (CIFAR10)ResNet18 (CIFAR10)ResNet18 (CIFAR100)
BadNetsDBALPOursBadNetsDBALPOursBadNetsDBALPOurs
No DefenseMA80.9781.2682.3881.81±0.3477.6678.2678.4378.70±0.8263.0562.465.1764.31±0.72
Best BA98.3346.1795.3897.05±0.1998.5920.9696.6997.60±0.21100.01.1672.2498.51±1.06
Avg BA97.7033.9690.5696.16±0.0997.9114.9395.0897.22±0.2199.730.8061.8796.70±2.46
MK [4]MA74.1775.4674.3078.03±0.3375.2476.2176.0776.49±0.8657.8458.6458.6558.73±0.55
Best BA24.063.1693.5498.01±0.806.047.5295.8998.61±0.140.880.9997.5299.94±0.01
Avg BA5.991.3376.4195.98±0.363.643.8592.4097.03±1.570.620.5489.2294.42±2.87
FLTrust [5]MA83.1483.2582.9883.84±0.2080.7580.2280.2280.33±0.2961.5861.8061.8961.14±1.81
Best BA56.236.5396.9697.95±0.3491.814.1897.1997.78±0.2730.120.6995.4299.98±0.05
Avg BA42.454.9196.6096.74±0.4789.123.6795.4891.94±8.4119.870.4687.5198.52±1.59
FLAME [23]MA58.0757.5258.6363.46±0.9174.2173.5271.9073.06±0.7655.9354.2855.6255.87±0.45
Best BA21.2037.8492.4398.38±0.6312.702.9794.4698.69±0.170.830.5377.6799.99±0.00
Avg BA7.7512.7678.8997.46±0.447.012.2789.7297.54±0.580.620.4472.8999.88±0.08
MM [12]MA69.3166.4370.7275.42±0.9771.4372.6172.8073.16±1.0753.0353.8253.4053.55±0.33
Best BA29.0816.0294.9598.31±0.4415.541.7196.0198.47±0.090.970.8592.7399.99±0.00
Avg BA7.045.4284.6194.55±4.203.911.3793.1297.52±0.670.370.6683.4697.05±1.58
DnC [29]MA81.0982.682.3281.30±1.4577.9879.3278.6378.83±0.3063.0161.4363.8963.43±0.73
Best BA5.942.8395.0497.85±0.276.5811.2096.2598.19±0.140.434.8880.9399.67±0.29
Avg BA3.681.6793.5796.94±0.534.039.7294.4597.29±0.370.412.6963.3798.07±0.59
RLR [25]MA72.2876.2277.7177.90±0.6875.7275.0674.5174.77±1.3754.7756.7856.1456.58±0.58
Best BA96.9323.0694.1197.78±0.1197.2917.1694.8398.39±0.130.140.3488.1299.98±0.01
Avg BA95.6515.1391.6197.39±0.2997.0210.8589.4197.95±0.170.050.0983.8199.97±0.02
FLARE [34]MA84.4584.2484.5384.56±0.1177.7978.2177.5277.82±2.0560.4456.5859.5459.07±1.41
Best BA98.2143.8794.7997.85±0.1298.185.8894.9196.75±0.5099.711.1488.5699.56±0.41
Avg BA98.1136.3294.1097.28±0.3098.025.4093.5596.30±0.8298.600.6380.7398.17±2.42
+ +approaches [2, 41, 42], we assume that the attacker controls exactly one client per round and sets the attack interval to $F = 1$ . This setup allows us to analyze the attack in isolation, without client collusion. We also evaluate the attack's performance across different frequencies, comparing it to the baseline. + +Baseline Defenses and Attacks. We assess our backdoor attack on eight state-of-the-art or representative defenses in FL: MultiKrum (MK) [4], FLTrust [5], FLAME [23], Multi-Metrics (MM) [12], DnC [29], RLR [25], FLARE [34], and DeepSight [28], ensuring a comprehensive evaluation of the attack's effectiveness. + +We compare our backdoor attack against three state-of-the-art or representative attacks in FL: BadNets [10], DBA [35], and LP attack [42]. + +Metrics. Following prior work on FL backdoor attacks [10, 35, 42], we evaluate the global model using two primary metrics: main task accuracy (MA) and backdoor accuracy (BA). Given the dynamic nature of the global model in FL, we report both the average and the best BA over the last + +10 communication rounds. Furthermore, to evaluate the stealthiness of the attack, we adopt two metrics introduced in [42]: the malicious client acceptance rate (MAR) and the benign client acceptance rate (BAR). MAR measures the proportion of rounds in which malicious clients successfully bypass defense and are selected for aggregation, while BAR represents the average proportion of rounds in which benign clients are selected for aggregation. + +# 4.2. The effectiveness of our attack + +We first evaluate our attack in the fixed-frequency attack setting under different SOTA defenses. Table 1 presents the performance comparison between our attack and the baseline attacks. We also evaluate our attack in the IID setting in supplementary §B.1 and on a larger dataset TinyImageNet in supplementary §B.11. The results show that our attack achieves the highest BA in most cases. These results demonstrate the broad effectiveness of our attack, effectively bypassing defenses while maintaining superior attack performance. + +BadNets only works well under FLARE and consistently fails to bypass defenses such as MK, FLAME, MM, and + +Table 2. MAR and BAR comparison under outlier-based defenses: MK, FLAME, MM, and DnC. + +
Dataset (Model)AttackMKFLAMEMMDnC
BARMARBARMARBARMARBARMAR
CIFAR10 (ResNets18)BadNets0.420.010.750.010.330.010.970.01
DBA0.410.010.710.010.340.010.950.01
LP attack0.360.720.650.900.130.80.390.82
Ours0.380.680.640.880.250.790.780.99
CIFAR10 (VGG19)BadNets0.410.010.720.010.310.010.980.01
DBA0.420.010.750.010.330.030.970.01
LP attack0.350.630.610.750.130.80.410.99
Ours0.380.560.60.90.260.600.790.99
CIFAR100 (ResNets18)BadNets0.430.010.680.010.330.010.990.01
DBA0.420.010.760.010.330.010.970.01
LP attack0.320.910.560.990.120.930.350.94
Ours0.340.830.580.970.230.920.720.99
+ +DnC across all settings, and it also fails against RLR on CIFAR100 in both non-IID and IID settings. Similarly, DBA also fails under several defenses, including MK, FLTrust, MM, and DnC under both non-IID and IID settings. LP attack achieves comparable BA across most defenses, except when evaluated under the FLARE defense in the IID setting. Specifically, for CIFAR100, our attack achieves significantly higher BA than the LP attack, demonstrating superior adaptability to complex data distributions. We also evaluate our attack performance compared with the LP attack under DeepSight in supplementary §B.3. Our attack achieves a higher BA than the LP attack across all scenarios and the LP attack fails on CIFAR100. + +Our attack improves the attack stealthiness and maintains strength. Under the FedAvg setting, our attack achieves significantly higher BA compared to both DBA and LP attack, particularly on CIFAR100, approaching the performance of unconstrained attacks like BadNets. This demonstrates the strong strength of our attack. + +Table 2 presents the comparison of MAR and BAR under outlier-based defenses such as MK, FLAME, MM, and DnC. A higher MAR indicates that malicious updates are more likely to bypass detection. We observe that BadNets and DBA fail to bypass defenses, as their MAR remains close to zero in all cases. This suggests that BadNets and DBA are consistently detected as outliers during training, rendering them ineffective against robust defenses. While BadNets demonstrates strong attack strength, its failure to bypass defenses undermines its overall effectiveness. In contrast, our attack achieves MAR consistently higher than BAR, meaning our malicious updates are more likely to be selected for aggregation, which reflects the stealthiness capability of our method. + +When compared to the LP attack, our method maintains nearly the same or even higher MAR in defenses such as FLAME and DnC, further demonstrating its stealthiness. MK and MM impose stricter selection criteria by choosing only a small proportion of the most benign clients, making + +it more challenging for an attacker to be selected. LP attack achieves a higher MAR than our method under these defenses by restricting the backdoor to BC layers, which significantly enhances its stealthiness. However, our method consistently achieves higher BA than LP attack across all settings in FLAME, DnC, MK, and MM, demonstrating that while LP attack sacrifices attack strength for stealthiness, our approach achieves a better balance, ensuring both strong attack strength and high stealthiness. + +More evaluation of attack stealthiness. We leverage Krum distance, cosine distance, and Manhattan distance to further illustrate the stealthiness of our attack. Specifically, we calculate the sum of squared Euclidean distances, cosine distances, and Manhattan distances between a model update and other local model updates, referring to them as Krum distance, Cosine distance, and Manhattan distance, respectively. + +A larger Krum distance indicates that the update deviates significantly from benign updates, making it less likely to be accepted by the aggregation mechanism. Figure 2 shows the comparison of Krum distance with baseline attacks. The Krum distances of BadNets and DBA are significantly larger than those of benign updates, indicating that these attacks produce model updates that are easily distinguishable from benign ones. In contrast, our attack and the LP attack exhibit Krum distances close to those of benign updates, suggesting that they are more stealthy. Although DBA demonstrates slightly better stealthiness than BadNets, its Krum distance remains considerably larger than that of benign updates. This suggests that DBA is still insufficiently stealthy against robust defenses. In supplementary §B.4, we further discuss the Cosine distance and Manhattan distance between the malicious update and the benign update. The result exhibits similar trends across all attacks. + +Table 3. The BA of various backdoor attacks combined without/with ours. + +
AttackIBA [24]IBA+oursCerp [20]Cerp+oursA3FL [40]A3FL+ours
FLAME2.7198.04↑29.8499.64↑94.8799.87↑
MM31.6399.99↑24.7299.98↑86.2999.89↑
+ +Our attack demonstrates strong generalizability. It can be easily integrated into existing FL backdoor attacks, particularly those based on trigger optimization. As shown in Table 3, incorporating our method further enhances attack effectiveness, underscoring its practical value. + +We further evaluate our attack trained on other networks: ResNet50 [11], ResNet101 [11] and ViT [6] in supplementary §B.5. Our attack achieves high BA on those networks, demonstrating its generalizability to more complex architectures. + +![](images/575f9f1063f4b2c8a1843a068f3c31fda3c8d70ac9a1cf73ca23e5626fb40014.jpg) +(a) Ours + +![](images/b9e4c51957637677699b7c6e072df031c4c715311495a22cab8096573ff4ce2e.jpg) +(b) LP attack + +![](images/fba678557aa3469d28691cc5b762a1d605f02e32a1de260d54bfdf4b4ed37a6b.jpg) +(c) BadNets and DBA + +![](images/3d67cbc95e873a5b62984da63b35810ffb0f5807b195437aa53988a87fc86a78.jpg) +Figure 2. Krum distance of malicious and benign model updates. A larger distance indicates that the update deviates significantly from benign updates. The result shows the stealthiness of our attack. + +![](images/e190363b426e6b8ea4a385ac74797fe89d1f3fca396f68ccbd83b2052448cc1d.jpg) + +![](images/2105f793fb1b4346702c883ffd7a17a248aa45429ff879b3f2e018264b1c88b4.jpg) +(a) MK (CIFAR-10) +(c) MK (CIFAR-100) +Figure 3. Impacts of different attack intervals $F$ on the attack performance. The $F$ indicates that an attack is performed every $F$ rounds. + +![](images/e37ff7d581a43f51ac3de764c871d10da6f8db87eb5364c40f377c340ce83975.jpg) +(b) FLAME (CIFAR-10) +(d) FLAME (CIFAR-100) + +In addition, we analyze the impact of varying degrees of data heterogeneity in supplementary §B.6. Results show that our attack remains effective under highly non-IID conditions and consistently outperforms the LP attack across all scenarios, further confirming its generalizability. + +Our attack maintains the global model's performance. Supplementary §B.2 shows the MA with the unpoisoned global model to assess whether our attack affects the utility of the global model. The results demonstrate that our attack does not lead to a significant decrease in MA compared to the unpoisoned global model. + +# 4.3. Our attack under different scenarios + +Impact of attack intervals. We evaluate the performance of our attack compared to baseline attacks under different attack intervals, as shown in Figure 3. The attack interval $F$ ranges from 1 to 5, where $F$ indicates that an attack is performed once every $F$ rounds. This setting is designed to simulate scenarios where the frequency of malicious updates is limited, requiring more effective and persistent attack + +![](images/0b9544a783f459438d10ac4f5052f3dd28b7887cb617d1e40eec390aef14ed8a.jpg) + +![](images/f14ebdacb58da45c5f06eefbe201609d3b5da69e8a508ab25c746bf5752b4f26.jpg) + +![](images/dc08f1af510875b2d0889e0f97dc31b2b018d876875e3843f2843c2fe0c05cbe.jpg) +(a) MK (CIFAR-10) +(c) MK (CIFAR-100) +Figure 4. Effectiveness of our attack under different proportions of compromised clients ( $C = 0.02, 0.04, 0.06, 0.1$ ) in a fixed-pool attack setting. + +![](images/01b7a387c13fcdeb60c840ffdd1b29e753cf21c6a92b4b3832c720fc349a8e08.jpg) +(b) FLAME (CIFAR-10) +(d) FLAME (CIFAR-100) + +strategies. + +The results show that as $F$ increases, the performance of all attacks degrades as the attack frequency decreases. This trend is consistent across all evaluated methods, where a lower frequency results in reduced BA. A larger $F$ imposes greater demands on the attack's strength, persistence, and stealthiness. The attack must be sufficiently powerful and stealthy to maintain its effectiveness with fewer opportunities to update the model. We observe that BadNets fails in all cases. Notably, our attack maintains a comparable BA even when $F = 5$ . Whereas the LP attack loses nearly all effectiveness, particularly when trained on CIFAR100, demonstrating its vulnerability to defenses when attack opportunities are limited. We also show the results of BA under all defenses trained on CIFAR10 (ResNet18) in supplementary §B.8. We observe that our attack achieves comparable BA across all defenses when $F = 5$ . + +Impact of different proportions of compromised clients. To evaluate our attack in a more practical scenario, we conduct experiments in a fixed-pool attack setting. To further + +Table 4. Ablation study on alignment strategy. + +
Dataset (Model)AttackNo DefenseFLAMEMM
BABAMARBAMAR
CIFAR-10 (VGG19)no alignment97.707.050.017.040.01
model-wise90.1245.790.9083.580.65
layer-wise (ours)96.1697.460.9097.090.67
CIFAR-10 (ResNet18)no alignment97.917.010.013.910.01
model-wise93.3985.320.8995.530.80
layer-wise (ours)97.2198.160.8897.850.79
CIFAR-100 (ResNet18)no alignment99.730.620.010.370.01
model-wise73.2985.860.9595.050.90
layer-wise (ours)96.7099.880.9797.050.92
+ +demonstrate the strength of our attack, we also analyze its performance under varying proportions of compromised clients $C$ . In a fixed-pool attack setting, the attacker controls a pool of compromised clients, from which participants are randomly selected in each round. This setup allows multiple malicious clients to be chosen in a single round, while in some cases, several consecutive rounds may pass without selecting any malicious clients. + +As shown in Figure 4, our attack achieves the highest BA compared to the baseline across all cases. Notably, even with a very low proportion of malicious clients ( $C = 0.02$ , meaning only $2\%$ of clients are compromised), our attack successfully bypasses defenses and maintains a comparable BA. This demonstrates that our attack is both powerful and persistent, as it continues to perform well even when the pool of malicious clients is small. + +Additionally, BadNets performs better when $C = 0.1$ due to the increased likelihood of malicious clients colluding in the same round. This collusion significantly enhances the attack's effectiveness, as multiple attackers working together are more likely to successfully inject a backdoor. This is particularly evident under MK, where BadNets achieves a higher BA with a higher $C$ , compared to when it fails in the fixed-frequency setting. However, this improvement diminishes as the compromised client ratio decreases. At lower $C$ , the attack becomes more reliant on its strength. In contrast, as shown in supplementary §B.9, our attack achieves high BA across all defenses when $C > 0.04$ , demonstrating its robustness even with fewer compromised clients. + +# 4.4. Ablation study + +Impact of different triggers. We further evaluate the impact of various trigger shapes. Supplementary §B.10 presents three different trigger shapes (Apple, Watermark, and Square) applied to backdoor a ResNet18 model trained on the CIFAR10 dataset. Figure 5a shows the impact of various trigger shapes. Under the default setting $(F = 1)$ , all three trigger shapes achieve high BA. The results demonstrate that our attack remains effective across different trigger shapes, indicating its robustness and generalizability. + +![](images/75ec940596f787780c656c100a20fc75b8b01c81b1653b92be4b3b53a1c1660c.jpg) +(a) various trigger shapes + +![](images/4a21bf3029a24c2a85ea2eb8ed226d8840c5fd01fdc3975c6ac11db85c3a90e1.jpg) +(b) various trigger sizes +Figure 5. Impact of trigger shape (left) and trigger size (right) in Flame. + +Figure 5b illustrates the performance of different trigger sizes (3, 5, 7, and 10). Regardless of trigger size, our attack consistently achieves high BA. Notably, when the trigger size is set to 10, the attack effectiveness is significantly enhanced. Larger triggers strengthen the attack, and their ability to bypass defenses is improved through layer-wise norm constraints, leading to superior attack performance. + +Impact of alignment strategy. To demonstrate how dynamic layer-wise gradient alignment enhances the effectiveness of our attack, we compare it with BadNets (no alignment) and the model-wise gradient alignment attack, which adjusts the overall gradient magnitude to align with the previous global update. As shown in Table 4, our attack consistently outperforms both BadNets and the model-wise method across all settings. + +First, both model-wise and layer-wise gradient alignment show higher MAR compared to BadNets, indicating that aligning the update with the previous global model significantly improves the stealthiness of the attack. Second, when compared to model-wise alignment, layer-wise alignment achieves better attack performance. Specifically, even in the absence of an attack, particularly on CIFAR100, the BA for model-wise alignment is much lower than that for layer-wise alignment. + +# 5. Conclusion + +This paper proposes a new stealthy backdoor attack in FL, which leverages adaptive layer-wise gradient alignment to enhance stealthiness while preserving attack strength. By dynamically aligning malicious gradients to historical global updates at the layer level, the proposed attack reduces statistical anomalies caused by malicious updates, making them harder to detect. Extensive experiments across diverse datasets, models, and defenses demonstrate that the proposed attack achieves superior backdoor success rates while evading SOTA defenses, highlighting its effectiveness and stealthiness in various FL scenarios. + +# Acknowledgements + +This work was supported in part by the National Natural Science Foundation of China (NO. 62472284, No. 62302288), Shanghai Key Laboratory of Scalable Computing and Systems, SJTU Kunpeng & Ascend Center of Excellence. + +# References + +[1] Sana Awan, Bo Luo, and Fengjun Li. CONTRA: Defending Against Poisoning Attacks in Federated Learning. In ESORICS, 2021. 2 +[2] Eugene Bagdasaryan, Andreas Veit, Yiqing Hua, Deborah Estrin, and Vitaly Shmatikov. How To Backdoor Federated Learning. In AISTATS, 2020. 1, 2, 3, 5 +[3] Akshay Batheja and Pushpak Bhattacharyya. "A Little is Enough": Few-Shot Quality Estimation based Corpus Filtering improves Machine Translation. In ACL (Findings), 2023. 2 +[4] Peva Blanchard, El Mahdi El Mhamdi, Rachid Guerraoui, and Julien Stainer. Machine Learning with Adversaries: Byzantine Tolerant Gradient Descent. In NeurIPS, 2017. 1, 2, 3, 5 +[5] Xiaoyu Cao, Minghong Fang, Jia Liu, and Neil Zhenqiang Gong. FLTrust: Byzantine-robust Federated Learning via Trust Bootstrapping. In NDSS, 2021. 1, 2, 3, 4, 5 +[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 6 +[7] Minghong Fang, Xiaoyu Cao, Jinyuan Jia, and Neil Gong. Local Model Poisoning Attacks to Byzantine-Robust Federated Learning. In USENIX Security, 2020. 4 +[8] Pei Fang and Jinghui Chen. On the Vulnerability of Backdoor Defenses for Federated Learning. In AAAI, 2023. 2 +[9] Clement Fung, Chris JM Yoon, and Ivan Beschastnikh. The Limitations of Federated Learning in Sybil Settings. In RAID, 2020. 2 +[10] Tianyu Gu, Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. BadNets: Evaluating Backdooring Attacks on Deep Neural Networks. IEEE Access, 7:47230-47244, 2019. 5 +[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In CVPR, 2016. 4, 6 +[12] Siquan Huang, Yijiang Li, Chong Chen, Leyu Shi, and Ying Gao. Multi-Metrics Adaptively Identifies Backdoors in Federated Learning. In ICCV, 2023. 1, 2, 3, 5 +[13] Tiansheng Huang, Sihao Hu, Ka Ho Chow, Fatih Ilhan, Selim F. Tekin, and Ling Liu. Lockdown: Backdoor Defense for Federated Learning with Isolated Subspace Training. In NeurIPS, 2023. 3 +[14] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and Open Problems in Federated + +Learning. Foundations and trends® in machine learning, 14 (1-2):1-210, 2021. 1 +[15] Yun-Yong Ko, Dongwon Lee, and Sang-Wook Kim. Not All Layers Are Equal: A Layer-Wise Adaptive Approach Toward Large-Scale DNN Training. In WWW, 2022. 2 +[16] Jakub Konečny. Federated learning: Strategies for improving communication efficiency. arXiv preprint arXiv:1610.05492, 2016. 2 +[17] Torsten Krauß and Alexandra Dmitrienko. MESAS: Poisoning Defense for Federated Learning Resilient against Adaptive Attackers. In CCS, 2023. 2, 3 +[18] Haoyang Li, Qingqing Ye, Haibo Hu, Jin Li, Leixia Wang, Chengfang Fang, and Jie Shi. 3DFed: Adaptive and Extensible Framework for Covert Backdoor Attack in Federated Learning. In S&P, 2023. 2 +[19] Ye Li, Yanchao Zhao, Chengcheng Zhu, and Jiale Zhang. Infighting in the dark: Multi-labels backdoor attack in federated learning. arXiv preprint arXiv:2409.19601, 2024. 2 +[20] Xiaoting Lyu, Yufei Han, Wei Wang, Jingkai Liu, Bin Wang, Jiqiang Liu, and Xiangliang Zhang. Poisoning with Cerberus: Stealthy and Colluded Backdoor Attack against Federated Learning. In AAAI, 2023. 2, 3, 6 +[21] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y Arcas. Communication-Efficient Learning of Deep Networks from Decentralized Data. In AISTATS, 2017. 1 +[22] El Mahdi El Mhamdi, Rachid Guerraoui, and Sébastien Rouault. The Hidden Vulnerability of Distributed Learning in Byzantium. In ICML, 2018. 2 +[23] Thien Duc Nguyen, Phillip Rieger, Huili Chen, Hossein Yalame, Helen Möllering, Hossein Fereidooni, Samuel Marchal, Markus Miettinen, Azalia Mirhoseini, Shaza Zeitouni, Farinaz Koushanfar, Ahmad-Reza Sadeghi, and Thomas Schneider. FLAME: Taming Backdoors in Federated Learning. In USENIX Security 22, 2022. 1, 2, 3, 5 +[24] Thuy Dung Nguyen, Tuan Nguyen, Anh Tran, Khoa D. Doan, and Kok-Seng Wong. IBA: Towards Irreversible Backdoor Attacks in Federated Learning. In NeurIPS, 2023. 2, 6 +[25] Mustafa Safa Ozdayi, Murat Kantarcioglu, and Yulia R Gel. Defending against backdoors in federated learning with robust learning rate. In AAAI, 2021. 1, 3, 5 +[26] Ashwinee Panda, Saeed Mahloujifar, Arjun Nitin Bhagoji, Supriyo Chakraborty, and Prateek Mittal. SparseFed: Mitigating Model Poisoning Attacks in Federated Learning with Sparsification. In AISTATS, 2022. 3 +[27] Matthias Paulik, Matt Seigel, Henry Mason, Dominic Telaar, Joris Kluivers, Rogier Van Dalen, Chi Wai Lau, Luke Carlson, Filip Granqvist, Chris Vandevelde, et al. Federated Evaluation and Tuning for On-Device Personalization: System Design & Applications. arXiv preprint arXiv:2102.08503, 2021. 1 +[28] Phillip Rieger, Thien Duc Nguyen, Markus Miettinen, and Ahmad-Reza Sadeghi. DeepSight: Mitigating Backdoor Attacks in Federated Learning Through Deep Model Inspection. In NDSS, 2022. 2, 3, 5 +[29] Virat Shejwalkar and Amir Houmansadr. Manipulating the Byzantine: Optimizing Model Poisoning Attacks and Defenses for Federated Learning. In NDSS, 2021. 1, 2, 3, 5 + +[30] Karen Simonyan. Very Deep Convolutional Networks for Large-Scale Image Recognition. arXiv preprint arXiv:1409.1556, 2014. 4 +[31] Peng Sun, Xinyang Liu, Zhibo Wang, and Bo Liu. Byzantine robust Decentralized Federated Learning via Dual-domain Clustering and Trust Bootstrapping. In CVPR, 2024. 2 +[32] Ziteng Sun, Peter Kairouz, Ananda Theertha Suresh, and H Brendan McMahan. Can You Really Backdoor Federated Learning? arXiv preprint arXiv:1911.07963, 2019. 1, 3 +[33] Hongyi Wang, Kartik Sreenivasan, Shashank Rajput, Harit Vishwakarma, Saurabh Agarwal, Jy-yong Sohn, Kangwook Lee, and Dimitris S. Papailiopoulos. Attack of the Tails: Yes, You Really Can Backdoor Federated Learning. In NeurIPS, 2020. 1, 2 +[34] Ning Wang, Yang Xiao, Yimin Chen, Yang Hu, Wenjing Lou, and Y. Thomas Hou. FLARE: Defending Federated Learning against Model Poisoning Attacks via Latent Space Representations. In AsiaCCS, 2022. 2, 5 +[35] Chulin Xie, Keli Huang, Pin-Yu Chen, and Bo Li. DBA: Distributed Backdoor Attacks against Federated Learning. In ICLR, 2020. 1, 2, 3, 5 +[36] Yueqi Xie, Minghong Fang, and Neil Zhenqiang Gong. FedREDefence: Defending against Model Poisoning Attacks for Federated Learning using Model Update Reconstruction Error. In ICML, 2024. 3 +[37] Jiahao Xu, Zikai Zhang, and Rui Hu. Identify Backdoored Model in Federated Learning via Individual Unlearning. arXiv preprint arXiv:2411.01040, 2024. 3 +[38] Peishen Yan, Hao Wang, Tao Song, Yang Hua, Ruhui Ma, Ningxin Hu, Mohammad Reza Haghighat, and Haibing Guan. SkyMask: Attack-Agnostic Robust Federated Learning With Fine-Grained Learnable Masks. In ECCV, 2025. 3 +[39] Timothy Yang, Galen Andrew, Hubert Eichner, Haicheng Sun, Wei Li, Nicholas Kong, Daniel Ramage, and Françoise Bea-fays. Applied Federated Learning: Improving Google Keyboard Query Suggestions. arXiv preprint arXiv:1812.02903, 2018. 1 +[40] Hangfan Zhang, Jinyuan Jia, Jinghui Chen, Lu Lin, and Ding-hao Wu. A3FL: Adversarily Adaptive Backdoor Attacks to Federated Learning. In NeurIPS, 2023. 2, 6 +[41] Zhengming Zhang, Ashwinee Panda, Linyue Song, Yaoqing Yang, Michael W. Mahoney, Prateek Mittal, Kannan Ramchandran, and Joseph Gonzalez. Neurotoxin: Durable Backdoors in Federated Learning. In ICML, 2022. 1, 2, 3, 4, 5 +[42] Haomin Zhuang, Mingxian Yu, Hao Wang, Yang Hua, Jian Li, and Xu Yuan. Backdoor Federated Learning by Poisoning Backdoor-Critical Layers. In ICLR, 2024. 1, 2, 3, 4, 5 \ No newline at end of file diff --git a/2025/Stealthy Backdoor Attack in Federated Learning via Adaptive Layer-wise Gradient Alignment/images.zip b/2025/Stealthy Backdoor Attack in Federated Learning via Adaptive Layer-wise Gradient Alignment/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..af3df67737b0e646caa208133376330aafae96f6 --- /dev/null +++ b/2025/Stealthy Backdoor Attack in Federated Learning via Adaptive Layer-wise Gradient Alignment/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34968a643d6c899b77507db43174074233e7a1cea89a8b6890e78c2e4e63f757 +size 582451 diff --git a/2025/Stealthy Backdoor Attack in Federated Learning via Adaptive Layer-wise Gradient Alignment/layout.json b/2025/Stealthy Backdoor Attack in Federated Learning via Adaptive Layer-wise Gradient Alignment/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b817c78f3c2c0f1b1b43b3c1fac72d7752bf920d --- /dev/null +++ b/2025/Stealthy Backdoor Attack in Federated Learning via Adaptive Layer-wise Gradient Alignment/layout.json @@ -0,0 +1,8860 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 154, + 103, + 457, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 103, + 457, + 140 + ], + "spans": [ + { + "bbox": [ + 154, + 103, + 457, + 140 + ], + "type": "text", + "content": "Stealthy Backdoor Attack in Federated Learning via Adaptive Layer-wise Gradient Alignment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 160, + 523, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 160, + 523, + 219 + ], + "spans": [ + { + "bbox": [ + 85, + 160, + 523, + 219 + ], + "type": "text", + "content": "Qingqian Yang" + }, + { + "bbox": [ + 85, + 160, + 523, + 219 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 85, + 160, + 523, + 219 + ], + "type": "text", + "content": " Peishen Yan" + }, + { + "bbox": [ + 85, + 160, + 523, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 85, + 160, + 523, + 219 + ], + "type": "text", + "content": " Xiaoyu Wu" + }, + { + "bbox": [ + 85, + 160, + 523, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 85, + 160, + 523, + 219 + ], + "type": "text", + "content": " Jiaru Zhang" + }, + { + "bbox": [ + 85, + 160, + 523, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 85, + 160, + 523, + 219 + ], + "type": "text", + "content": " Tao Song" + }, + { + "bbox": [ + 85, + 160, + 523, + 219 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 85, + 160, + 523, + 219 + ], + "type": "text", + "content": " Yang Hua" + }, + { + "bbox": [ + 85, + 160, + 523, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 85, + 160, + 523, + 219 + ], + "type": "text", + "content": " Hao Wang" + }, + { + "bbox": [ + 85, + 160, + 523, + 219 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 85, + 160, + 523, + 219 + ], + "type": "text", + "content": " Liangliang Wang" + }, + { + "bbox": [ + 85, + 160, + 523, + 219 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 85, + 160, + 523, + 219 + ], + "type": "text", + "content": " Haibing Guan" + }, + { + "bbox": [ + 85, + 160, + 523, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 85, + 160, + 523, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 85, + 160, + 523, + 219 + ], + "type": "text", + "content": "Shanghai University of Electric Power " + }, + { + "bbox": [ + 85, + 160, + 523, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 85, + 160, + 523, + 219 + ], + "type": "text", + "content": "Shanghai Jiao Tong University \n" + }, + { + "bbox": [ + 85, + 160, + 523, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 85, + 160, + 523, + 219 + ], + "type": "text", + "content": "Queen's University Belfast " + }, + { + "bbox": [ + 85, + 160, + 523, + 219 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 85, + 160, + 523, + 219 + ], + "type": "text", + "content": "Stevens Institute of Technology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 231, + 220, + 380, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 220, + 380, + 231 + ], + "spans": [ + { + "bbox": [ + 231, + 220, + 380, + 231 + ], + "type": "text", + "content": "y23108013@mail.shiep.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 122, + 233, + 488, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 233, + 488, + 246 + ], + "spans": [ + { + "bbox": [ + 122, + 233, + 488, + 246 + ], + "type": "text", + "content": "{peishenyan, wuxiaoyu2000, jiaruzhang, songt333, hbguan}@sjtu.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 149, + 247, + 456, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 247, + 456, + 258 + ], + "spans": [ + { + "bbox": [ + 149, + 247, + 456, + 258 + ], + "type": "text", + "content": "Y.Hua@qub.ac.uk hwang9@stevens.edu llwang@shiep.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 286, + 200, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 286, + 200, + 300 + ], + "spans": [ + { + "bbox": [ + 151, + 286, + 200, + 300 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 312, + 297, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 312, + 297, + 612 + ], + "spans": [ + { + "bbox": [ + 54, + 312, + 297, + 612 + ], + "type": "text", + "content": "The distributed nature of federated learning exposes it to significant security threats, among which backdoor attacks are one of the most prevalent. However, existing backdoor attacks face a trade-off between attack strength and stealthiness: attacks maximizing the attack strength are often detectable, while stealthier approaches significantly reduce the effectiveness of the attack itself. Both of them result in ineffective backdoor injection. In this paper, we propose an adaptive layer-wise gradient alignment strategy to effectively evade various robust defense mechanisms while preserving attack strength. Without requiring additional knowledge, we leverage the previous global update as a reference for alignment to ensure stealthiness during dynamic FL training. This fine-grained alignment strategy applies appropriate constraints to each layer, which helps significantly maintain attack strength. To demonstrate the effectiveness of our method, we conduct extensive evaluations across a wide range of datasets and networks. Our experimental results show that the proposed attack effectively bypasses eight state-of-the-art defenses and achieves high backdoor accuracy, outperforming existing attacks by up to " + }, + { + "bbox": [ + 54, + 312, + 297, + 612 + ], + "type": "inline_equation", + "content": "54.76\\%" + }, + { + "bbox": [ + 54, + 312, + 297, + 612 + ], + "type": "text", + "content": ". Additionally, it significantly preserves attack strength and maintains robust performance across diverse scenarios, highlighting its adaptability and generalizability. Code implementation is available at https://github.com/yqqhyqq/LGA." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 633, + 135, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 633, + 135, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 633, + 135, + 645 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 653, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 677 + ], + "type": "text", + "content": "Federated Learning (FL) has emerged as a promising paradigm for privacy-preserving distributed machine learn" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 288, + 556, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 556, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 556, + 467 + ], + "type": "text", + "content": "ing, allowing multiple decentralized clients to collaboratively train a global model without directly sharing their raw data [14, 21]. FL leverages local computation on edge devices (e.g., smartphones and IoT sensors) and aggregates model updates via a central server, making it widely adopted in various real-world applications [27, 39]. However, the distributed architecture of FL systems significantly increases the potential attack surface, introducing a wide range of security threats, among which backdoor attacks are a commonly studied threat in recent research. In a backdoor attack [2, 32, 33, 35, 41, 42], adversarial clients intentionally inject a backdoor into the global model by uploading a poisoning local update. As a result, the global model behaves normally on benign inputs but misclassifies triggered inputs into predefined target labels." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 471, + 556, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 556, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 556, + 566 + ], + "type": "text", + "content": "Existing backdoor defenses primarily focus on detecting and filtering malicious updates by leveraging information from multiple dimensions. Specifically, some methods rely on distance-based measures between benign and malicious updates [4, 5, 12, 23], while others detect inconsistencies in parameter signs [25] or exploit the spectral separability [29]. These defenses significantly restrict the feasibility of backdoor attacks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "type": "text", + "content": "To counter these defenses, an effective backdoor attack depends on two key factors: attack strength and attack stealthiness. Some attacks [2] achieve high attack strength by scaling up the update, which increases their detectability and makes them easily filtered by defense mechanisms. Others [35, 42] prioritize stealthiness to evade detection but show significantly lower performance when no defense is present, highlighting their sacrifice of attack strength. These attacks essentially constrain the malicious update by discarding certain backdoor-related information [42] or using coarse-grained scaling [33]. Although constraining malicious updates helps maintain stealthiness, this inappropriate" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 683, + 297, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 297, + 713 + ], + "type": "text", + "content": "*The work was done during her internship at Shanghai Jiao Tong University, Shanghai Key Laboratory of Scalable Computing and Systems. † Corresponding authors." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29163" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "constraint leads to attack strength degradation. Therefore, balancing the attack stealthiness and attack strength with limited knowledge of the system's defenses is challenging." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 109, + 297, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 109, + 297, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 109, + 297, + 358 + ], + "type": "text", + "content": "To address the challenge, we propose a layer-wise adaptive gradient alignment strategy that dynamically adjusts the norm bound for each layer of malicious updates. This adjustment enhances their similarity to the benign update distribution in a fine-grained manner, thereby achieving both attack stealthiness and attack effectiveness. Without requiring any additional information, our approach uses the previous global update as a reference point for adaptive alignment. By doing so, the proposed strategy not only constrains the malicious update within a small norm ball, but also enables the attack to adapt to the evolving global model and training dynamics, thereby significantly improving its stealthiness. Inspired by previous observations [15] that parameters within the same layer tend to share similar gradient magnitudes, we further conduct empirical analysis revealing that the magnitudes of both malicious and benign updates vary across layers, and there is no consistent proportional relationship between them. This discrepancy highlights the need for layer-specific norm constraints. Such constraints are essential to prevent over-constraining layers with inherently smaller magnitudes, thereby preserving attack strength." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 360, + 296, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 296, + 430 + ], + "type": "text", + "content": "Extensive experiments validate the effectiveness of our attack, demonstrating its ability to evade eight state-of-the-art defenses while maintaining a high attack success rate. Additionally, we investigate its performance across various attack scenarios, demonstrating its generalizability. We summarize our main contributions as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 433, + 295, + 634 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 433, + 295, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 295, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 295, + 538 + ], + "type": "text", + "content": "- We propose an adaptive layer-wise gradient alignment strategy to effectively evade various robust defense mechanisms while preserving attack strength. Without requiring any additional knowledge, we leverage the previous global update as a reference for gradient alignment, ensuring that the attack maintains the stealthiness while adapting to the training dynamics. This fine-grained alignment strategy applies layer-specific constraints, allowing the attack to retain its strength." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 540, + 295, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 295, + 634 + ], + "type": "text", + "content": "- We conduct extensive evaluations across diverse models and datasets, demonstrating that the proposed attack successfully bypasses eight SOTA defenses and maintains the attack impact. Our backdoor attack outperforms SOTA attacks, with an improvement in backdoor accuracy by up to " + }, + { + "bbox": [ + 55, + 540, + 295, + 634 + ], + "type": "inline_equation", + "content": "54.76\\%" + }, + { + "bbox": [ + 55, + 540, + 295, + 634 + ], + "type": "text", + "content": ". Furthermore, our attack exhibits excellent performance across various scenarios, such as different attack patterns, highlighting its generalizability." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 645, + 142, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 142, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 142, + 658 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Backdoor attacks in FL. Backdoor attacks present a serious security threat in FL [16]. Typically, attackers compromise a small subset of client devices to upload poisoned updates. These updates are crafted to inject the backdoor into" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": "the global model, which then enables the model to obtain predetermined predictions for samples with specific trigger features. The backdoor attack process generally consists of two stages: trigger embedding and model manipulation. To successfully inject a backdoor, the attacker can carefully design either the trigger or the model update to bypass detection. Accordingly, existing methods can be divided into two categories: model-optimization attacks [2, 18, 41, 42] and trigger-optimization attacks [8, 20, 24, 33, 35, 40]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 180, + 555, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 555, + 430 + ], + "type": "text", + "content": "Model-optimization attacks [2, 3, 18, 33, 41, 42] focus on fine-tuning the model to achieve the desired attack. Scaling attack [2] amplifies malicious updates to achieve a one-shot attack. However, the abnormally large gradient magnitudes make it exposed to detection. Neurotoxin [41] projects malicious gradients onto infrequently updated neurons, improving the persistence of the attack and preventing the backdoor from being overwritten by benign updates. However, as an unconstrained backdoor attack, it's vulnerable to robust detection. The LP attack [42] adaptively injects the backdoor into backdoor-critical (BC) layers, achieving high stealthiness. However, by discarding backdoor-related information in non-BC layers, the attack suffers from reduced strength. PGD attack [33] exploits knowledge of server-side defenses to craft malicious updates that bypass detection. However, this strategy is only effective when the attacker has access to detailed information about the system. Additionally, 3DFed [18] introduces three distinct attack strategies targeting different types of defenses and utilizes an indicator for adaptive strategy adjustments. However, it relies on client collaboration." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 432, + 556, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 556, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 556, + 563 + ], + "type": "text", + "content": "Trigger-optimization backdoor attacks aim to design triggers that are both subtle and adaptable, ensuring their stealthiness and effectiveness. Edge-case attack [33] selects marginal data as poison samples to circumvent detection mechanisms. DBA [35] firstly introduces distributed backdoor attacks, decomposing a global trigger into smaller, less detectable components. These approaches craft triggers at the pre-training stage and then fix them during training. Recently, a range of studies [8, 19, 20, 24, 40] dynamically optimize the trigger throughout the training process and obtain better attack results than the former ones." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 581, + 555, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 555, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 555, + 652 + ], + "type": "text", + "content": "Defense strategies against backdoor attacks in FL. In FL [16], the server only receives uploaded models without access to any information about client data. As a result, the majority of defense strategies are similarity-based, aiming to detect and mitigate backdoor attacks by identifying differences between malicious and benign models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "content": "Among them, distance-based defenses [1, 4, 5, 9, 12, 17, 22, 23, 28, 29, 31, 34] exploit the observation that malicious models often deviate from benign ones in their distribution within high-dimensional space, making distance-based metrics an effective tool for detection. MultiKrum [4]" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29164" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": "only aggregates updates that are selected with the smallest pairwise Euclidean distances. FLTrust [5] weights upload gradients based on their cosine similarity to a trusted gradient to mitigate the influence of malicious models. The trusted gradient is obtained from a root dataset collected by the server. FLAME [23] and DeepSight [28] use pairwise cosine distance between client gradients for clustering, identifying malicious gradients as outliers and further refining the defense through additional modules. Multi-Metrics [12] dynamically adjusts the weights of three distance metrics to detect backdoor models. MESAS [17] leverages six distance metrics for detection." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 216, + 297, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 216, + 297, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 216, + 297, + 431 + ], + "type": "text", + "content": "Other similarity-based defenses [13, 26, 29, 36, 38] identify malicious updates from more diverse perspectives. DnC [29] leverages singular value decomposition (SVD)-based spectral methods to remove outliers. RLR [25] detects inconsistencies in parameter signs across training rounds and adjusts learning rates accordingly to suppress backdoors. MASA [37] leverages the observation that a malicious model unlearns the main task more quickly than a benign model for detection. FedREDefense [36] relies on the finding that a malicious model is more resistant to distillation. In addition, some methods recently focus on the difference in update behavior between benign models and malicious models at the parameter level. For instance, SparseFed [26] only aggregates Top-k highest magnitude parameters because the backdoor updates are usually injected into coordinates that are unimportant for benign updates. Meanwhile, Lockdown [13] restrains local training in a subspace to further reduce the poison-coupling effect." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 441, + 138, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 138, + 454 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 138, + 454 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 461, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 296, + 521 + ], + "type": "text", + "content": "Background. Federated Learning (FL) distributes the training task across multiple clients to preserve data privacy. At the beginning, a global model " + }, + { + "bbox": [ + 55, + 461, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\theta_g^0" + }, + { + "bbox": [ + 55, + 461, + 296, + 521 + ], + "type": "text", + "content": " is initialized on the server and distributed to a set of " + }, + { + "bbox": [ + 55, + 461, + 296, + 521 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 461, + 296, + 521 + ], + "type": "text", + "content": " clients. In each round " + }, + { + "bbox": [ + 55, + 461, + 296, + 521 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 461, + 296, + 521 + ], + "type": "text", + "content": ", a standard FL framework involves the following steps:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 521, + 296, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 521, + 296, + 583 + ], + "spans": [ + { + "bbox": [ + 56, + 521, + 296, + 583 + ], + "type": "text", + "content": "1. Local Training: Each client " + }, + { + "bbox": [ + 56, + 521, + 296, + 583 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 56, + 521, + 296, + 583 + ], + "type": "text", + "content": " receives the global model " + }, + { + "bbox": [ + 56, + 521, + 296, + 583 + ], + "type": "inline_equation", + "content": "\\theta_{g}^{t - 1}" + }, + { + "bbox": [ + 56, + 521, + 296, + 583 + ], + "type": "text", + "content": " and performs " + }, + { + "bbox": [ + 56, + 521, + 296, + 583 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 56, + 521, + 296, + 583 + ], + "type": "text", + "content": " local epochs of stochastic gradient descent (SGD) on its private dataset " + }, + { + "bbox": [ + 56, + 521, + 296, + 583 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{D}}_i" + }, + { + "bbox": [ + 56, + 521, + 296, + 583 + ], + "type": "text", + "content": " with batch size " + }, + { + "bbox": [ + 56, + 521, + 296, + 583 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 56, + 521, + 296, + 583 + ], + "type": "text", + "content": ", obtaining an updated model " + }, + { + "bbox": [ + 56, + 521, + 296, + 583 + ], + "type": "inline_equation", + "content": "\\theta_{i}^{t}" + }, + { + "bbox": [ + 56, + 521, + 296, + 583 + ], + "type": "text", + "content": ". The client computes the model update as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 144, + 590, + 296, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 590, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 144, + 590, + 296, + 605 + ], + "type": "interline_equation", + "content": "\\Delta \\theta_ {i} ^ {t} = \\theta_ {i} ^ {t} - \\theta_ {g} ^ {t - 1}, \\tag {1}", + "image_path": "d62f0746d41b492db43e52695f350b2845af8329a10840b0c0e1fef0c1792d2f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 612, + 180, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 612, + 180, + 623 + ], + "spans": [ + { + "bbox": [ + 67, + 612, + 180, + 623 + ], + "type": "text", + "content": "and uploads it to the server." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 623, + 296, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 296, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 296, + 659 + ], + "type": "text", + "content": "2. Global aggregation: The server randomly selects a subset of clients and aggregates their model updates " + }, + { + "bbox": [ + 55, + 623, + 296, + 659 + ], + "type": "inline_equation", + "content": "\\{\\Delta \\theta^t\\}" + }, + { + "bbox": [ + 55, + 623, + 296, + 659 + ], + "type": "text", + "content": " to compute the global update:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 139, + 666, + 296, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 666, + 296, + 682 + ], + "spans": [ + { + "bbox": [ + 139, + 666, + 296, + 682 + ], + "type": "interline_equation", + "content": "\\Delta \\theta_ {g} ^ {t} = \\sum_ {i \\in [ n ]} \\Delta \\theta_ {i} ^ {t}, \\tag {2}", + "image_path": "d7172abc22faf83bb4bedae52feb9b03889388c94baaf93d601c700ab1524043.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 689, + 296, + 713 + ], + "type": "text", + "content": "and updates the global model as " + }, + { + "bbox": [ + 67, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\theta_g^t = \\theta_g^{t - 1} + \\Delta \\theta_g^t" + }, + { + "bbox": [ + 67, + 689, + 296, + 713 + ], + "type": "text", + "content": ". The new global model is then distributed to clients for the" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 326, + 72, + 555, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 555, + 108 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 555, + 108 + ], + "type": "text", + "content": "next round. Before updating the global model, the server typically performs backdoor detection during aggregation to mitigate the impact of malicious updates." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 108, + 555, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 555, + 168 + ], + "type": "text", + "content": "In FL, backdoor attacks exploit the system's decentralized nature. Specifically, an attacker, denoted as " + }, + { + "bbox": [ + 313, + 108, + 555, + 168 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 108, + 555, + 168 + ], + "type": "text", + "content": ", compromises a subset of clients and injects malicious updates to implant a backdoor into the global model. We consider a continuous attack scenario, following prior studies [20, 35, 42]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 168, + 556, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 556, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 556, + 312 + ], + "type": "text", + "content": "Adversarial Objectives: The attacker " + }, + { + "bbox": [ + 313, + 168, + 556, + 312 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 168, + 556, + 312 + ], + "type": "text", + "content": " aims to inject a backdoor into the global model while minimizing the impact on the model's performance on non-poisoned data. The goal of the backdoor injection is to manipulate the model so that it misclassifies inputs with a specific trigger pattern as an incorrect target label. We denote the malicious update as " + }, + { + "bbox": [ + 313, + 168, + 556, + 312 + ], + "type": "inline_equation", + "content": "\\Delta \\theta_{m}" + }, + { + "bbox": [ + 313, + 168, + 556, + 312 + ], + "type": "text", + "content": ", which is trained on the poisoned dataset. By carefully crafting " + }, + { + "bbox": [ + 313, + 168, + 556, + 312 + ], + "type": "inline_equation", + "content": "\\Delta \\theta_{m}" + }, + { + "bbox": [ + 313, + 168, + 556, + 312 + ], + "type": "text", + "content": ", the attacker can introduce the backdoor without significantly degrading the model's overall accuracy. The server's aggregation mechanism, which aggregates updates from multiple clients, may unintentionally include these poisoned updates, thereby corrupting the global model." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 312, + 556, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 556, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 556, + 396 + ], + "type": "text", + "content": "Adversarial Capability: The attacker " + }, + { + "bbox": [ + 313, + 312, + 556, + 396 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 312, + 556, + 396 + ], + "type": "text", + "content": " has the ability to control a small subset of clients and gain access to their datasets. Additionally, the attacker can manipulate the models of these compromised clients [2, 35, 41]. The attacker does not have access to the data or models of benign clients, nor to the specific defense methods implemented by the defender." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 411, + 556, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 556, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 556, + 507 + ], + "type": "text", + "content": "Overview. We introduce an adaptive layer-wise gradient alignment to enhance stealthiness while preserving attack strength. In each round, we train the malicious update on a poisoned local dataset and iteratively align gradients to resemble benign updates before uploading them to the server. This increases the likelihood of the server aggregating the malicious update, effectively injecting the backdoor into the global model." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 521, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 557, + 715 + ], + "type": "text", + "content": "Adaptive reference selection. Traditionally, the Unconstrained malicious update " + }, + { + "bbox": [ + 313, + 521, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\Delta \\theta_{m}" + }, + { + "bbox": [ + 313, + 521, + 557, + 715 + ], + "type": "text", + "content": ", which is trained on the poisoned dataset including out-of-distribution backdoor samples, tends to exhibit a larger magnitude compared to benign updates [32]. To ensure that the malicious update remains similar to the distribution of benign updates, it should be carefully constrained within an appropriately small norm ball. Notably, the selection of the norm bound significantly impacts the effectiveness of the attack's concealment. Therefore, it is crucial to rigorously determine an appropriate reference point for alignment. Some defense methods, such as MultiKrum [4], select a fixed number of client updates that appear most benign in each round. However, due to data heterogeneity, such methods may inadvertently exclude genuinely benign clients whose updates deviate from the majority. To mitigate this issue, we adopt the previous global" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29165" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 69, + 265, + 187 + ], + "blocks": [ + { + "bbox": [ + 82, + 69, + 265, + 187 + ], + "lines": [ + { + "bbox": [ + 82, + 69, + 265, + 187 + ], + "spans": [ + { + "bbox": [ + 82, + 69, + 265, + 187 + ], + "type": "image", + "image_path": "299b4ddf2a78efc4cc680d2adcb185cccf9d779d87cdda40acbe2c18330a06e3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 190, + 297, + 247 + ], + "lines": [ + { + "bbox": [ + 55, + 190, + 297, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 297, + 247 + ], + "type": "text", + "content": "Figure 1. Layer-wise Magnitude Analysis of Benign vs. Malicious Updates in VGG19. To better visualize, we focus on weight layers and divide them into A-layers and B-layers based on " + }, + { + "bbox": [ + 55, + 190, + 297, + 247 + ], + "type": "inline_equation", + "content": "\\|\\Delta\\theta^l\\|" + }, + { + "bbox": [ + 55, + 190, + 297, + 247 + ], + "type": "text", + "content": ": those with norms above 0.5 are A-layers; others are B-layers. The experiment uses the default setting and records " + }, + { + "bbox": [ + 55, + 190, + 297, + 247 + ], + "type": "inline_equation", + "content": "\\|\\Delta\\theta^l\\|" + }, + { + "bbox": [ + 55, + 190, + 297, + 247 + ], + "type": "text", + "content": " every 50 rounds." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 252, + 296, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 296, + 300 + ], + "type": "text", + "content": "update " + }, + { + "bbox": [ + 55, + 252, + 296, + 300 + ], + "type": "inline_equation", + "content": "\\Delta \\theta_{g}^{t - 1}" + }, + { + "bbox": [ + 55, + 252, + 296, + 300 + ], + "type": "text", + "content": " as a reference point. Since it is computed by aggregating multiple client updates, most of which are expected to be benign, it serves as a reasonable approximation of benign behavior." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 301, + 296, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 296, + 350 + ], + "type": "text", + "content": "By aligning the malicious update with the global update, the attack remains within the expected range of benign variation, thereby enhancing its stealthiness and increasing the likelihood of successfully influencing the global model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 363, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 296, + 506 + ], + "type": "text", + "content": "Layer-wise gradient alignment. Intuitively, applying a norm constraint weakens the impact of the malicious update, which potentially leads to ineffective backdoor injection. We introduce a layer-wise scaling strategy to refine the malicious update at a more granular level to not only ensure the attack remains stealthy but also preserve its strength. A key observation shown in Figure 1 is that the layer-wise magnitudes of the malicious and benign updates do not maintain a fixed proportional relationship; instead, their patterns differ. Therefore, we align each layer individually with " + }, + { + "bbox": [ + 55, + 363, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\Delta \\theta_{g}^{t - 1}" + }, + { + "bbox": [ + 55, + 363, + 296, + 506 + ], + "type": "text", + "content": ". Specifically, we define the scaling factor " + }, + { + "bbox": [ + 55, + 363, + 296, + 506 + ], + "type": "inline_equation", + "content": "S^{t,l}" + }, + { + "bbox": [ + 55, + 363, + 296, + 506 + ], + "type": "text", + "content": " for each layer " + }, + { + "bbox": [ + 55, + 363, + 296, + 506 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 363, + 296, + 506 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 118, + 505, + 296, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 505, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 118, + 505, + 296, + 534 + ], + "type": "interline_equation", + "content": "S ^ {t, l} = \\min (1, \\frac {\\| \\Delta \\theta_ {g} ^ {t - 1 , l} \\|}{\\| \\Delta \\theta_ {m} ^ {t , l} \\|}). \\tag {3}", + "image_path": "d87bb12b2343aa31fe451b26bbd0ce3d8730c21c1f4cd9331ad83235625c6b71.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 536, + 295, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 295, + 559 + ], + "type": "text", + "content": "The scaled malicious update after each local epoch is given by:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 558, + 295, + 573 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 558, + 295, + 573 + ], + "spans": [ + { + "bbox": [ + 133, + 558, + 295, + 573 + ], + "type": "interline_equation", + "content": "\\Delta \\theta_ {m} ^ {t, l} = S ^ {t, l} \\cdot \\Delta \\theta_ {m} ^ {t, l}. \\tag {4}", + "image_path": "d2318e27eaba869e63f35ecf1ef400abeaaa3262d5eb77484fe5fb066ce9ae5e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 576, + 297, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 297, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 297, + 659 + ], + "type": "text", + "content": "The adaptive layer-wise gradient alignment, implemented in just four lines (Lines 10-13 in Algorithm 1), extends the standard FL backdoor pipeline. This approach prevents excessive constraints on layers with inherently smaller magnitudes and constrains the update for each layer within an " + }, + { + "bbox": [ + 55, + 576, + 297, + 659 + ], + "type": "inline_equation", + "content": "\\ell 2" + }, + { + "bbox": [ + 55, + 576, + 297, + 659 + ], + "type": "text", + "content": " norm ball, thereby preserving attack strength while enhancing stealthiness." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 669, + 137, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 137, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 137, + 682 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 297, + 714 + ], + "type": "text", + "content": "Our empirical study evaluates the effectiveness of the proposed backdoor attack through experiments conducted on" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 72, + 555, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 555, + 122 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 555, + 122 + ], + "type": "text", + "content": "Algorithm 1 Layer-wise Gradient Alignment for Attackers \nRequire: Batch size " + }, + { + "bbox": [ + 314, + 72, + 555, + 122 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 314, + 72, + 555, + 122 + ], + "type": "text", + "content": ", local epochs " + }, + { + "bbox": [ + 314, + 72, + 555, + 122 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 314, + 72, + 555, + 122 + ], + "type": "text", + "content": ", malicious model parameters " + }, + { + "bbox": [ + 314, + 72, + 555, + 122 + ], + "type": "inline_equation", + "content": "\\theta_{i}^{t}" + }, + { + "bbox": [ + 314, + 72, + 555, + 122 + ], + "type": "text", + "content": ", global parameters " + }, + { + "bbox": [ + 314, + 72, + 555, + 122 + ], + "type": "inline_equation", + "content": "\\theta_{g}^{t-1}" + }, + { + "bbox": [ + 314, + 72, + 555, + 122 + ], + "type": "text", + "content": ", poisoned dataset " + }, + { + "bbox": [ + 314, + 72, + 555, + 122 + ], + "type": "inline_equation", + "content": "\\hat{D}" + }, + { + "bbox": [ + 314, + 72, + 555, + 122 + ], + "type": "text", + "content": ", and previous global parameter " + }, + { + "bbox": [ + 314, + 72, + 555, + 122 + ], + "type": "inline_equation", + "content": "\\theta_{g}^{t-2}" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 122, + 365, + 132 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 122, + 365, + 132 + ], + "spans": [ + { + "bbox": [ + 315, + 122, + 365, + 132 + ], + "type": "text", + "content": "Procedure:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 135, + 482, + 194 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 320, + 135, + 441, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 135, + 441, + 147 + ], + "spans": [ + { + "bbox": [ + 320, + 135, + 441, + 147 + ], + "type": "text", + "content": "1: // Attacker's local training:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 320, + 148, + 482, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 148, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 320, + 148, + 482, + 159 + ], + "type": "text", + "content": "2: Calculate the previous global update:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 159, + 426, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 159, + 426, + 172 + ], + "spans": [ + { + "bbox": [ + 320, + 159, + 426, + 172 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 320, + 159, + 426, + 172 + ], + "type": "inline_equation", + "content": "\\Delta \\theta_{g}^{t - 1}\\gets \\theta_{g}^{t - 1} - \\theta_{g}^{t - 2}" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 172, + 469, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 172, + 469, + 184 + ], + "spans": [ + { + "bbox": [ + 320, + 172, + 469, + 184 + ], + "type": "text", + "content": "4: Initialize local model " + }, + { + "bbox": [ + 320, + 172, + 469, + 184 + ], + "type": "inline_equation", + "content": "\\theta_{i,0}^{t}\\gets \\theta_{g}^{t - 1}" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 184, + 406, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 184, + 406, + 194 + ], + "spans": [ + { + "bbox": [ + 320, + 184, + 406, + 194 + ], + "type": "text", + "content": "5: for " + }, + { + "bbox": [ + 320, + 184, + 406, + 194 + ], + "type": "inline_equation", + "content": "e = 1" + }, + { + "bbox": [ + 320, + 184, + 406, + 194 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 320, + 184, + 406, + 194 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 320, + 184, + 406, + 194 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 320, + 195, + 443, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 195, + 443, + 206 + ], + "spans": [ + { + "bbox": [ + 320, + 195, + 443, + 206 + ], + "type": "text", + "content": "6: for each batch " + }, + { + "bbox": [ + 320, + 195, + 443, + 206 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 320, + 195, + 443, + 206 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 320, + 195, + 443, + 206 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{D}}" + }, + { + "bbox": [ + 320, + 195, + 443, + 206 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 207, + 493, + 241 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 320, + 207, + 473, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 207, + 473, + 218 + ], + "spans": [ + { + "bbox": [ + 320, + 207, + 473, + 218 + ], + "type": "text", + "content": "7: Compute gradient for batch " + }, + { + "bbox": [ + 320, + 207, + 473, + 218 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 320, + 207, + 473, + 218 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 218, + 493, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 218, + 493, + 232 + ], + "spans": [ + { + "bbox": [ + 320, + 218, + 493, + 232 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 320, + 218, + 493, + 232 + ], + "type": "inline_equation", + "content": "\\Delta \\theta_{i,e}^{t}\\gets \\bar{\\Delta}\\theta_{i,e}^{t} + \\nabla_{\\theta}\\mathcal{L}(\\theta_{i,e - 1}^{t},\\hat{\\mathcal{D}}_{b})" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 232, + 375, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 232, + 375, + 241 + ], + "spans": [ + { + "bbox": [ + 320, + 232, + 375, + 241 + ], + "type": "text", + "content": "9: end for" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 318, + 243, + 493, + 339 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 318, + 243, + 474, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 243, + 474, + 255 + ], + "spans": [ + { + "bbox": [ + 318, + 243, + 474, + 255 + ], + "type": "text", + "content": "10: // Layer-wise gradient alignment" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 318, + 255, + 454, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 255, + 454, + 268 + ], + "spans": [ + { + "bbox": [ + 318, + 255, + 454, + 268 + ], + "type": "text", + "content": "11: for each layer " + }, + { + "bbox": [ + 318, + 255, + 454, + 268 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 318, + 255, + 454, + 268 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 318, + 255, + 454, + 268 + ], + "type": "inline_equation", + "content": "\\Delta \\theta_{i,e}^{t}" + }, + { + "bbox": [ + 318, + 255, + 454, + 268 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 318, + 268, + 457, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 268, + 457, + 289 + ], + "spans": [ + { + "bbox": [ + 318, + 268, + 457, + 289 + ], + "type": "text", + "content": "12: " + }, + { + "bbox": [ + 318, + 268, + 457, + 289 + ], + "type": "inline_equation", + "content": "S_{i,e}^{t,l}\\gets min(1,\\frac{\\|\\Delta\\theta_{g}^{t - 1,l}\\|}{\\|\\Delta\\theta_{i,e}^{t,l}\\|})" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 318, + 289, + 464, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 289, + 464, + 304 + ], + "spans": [ + { + "bbox": [ + 318, + 289, + 464, + 304 + ], + "type": "text", + "content": "13: " + }, + { + "bbox": [ + 318, + 289, + 464, + 304 + ], + "type": "inline_equation", + "content": "\\theta_{i,e}^{t,l}\\gets \\Delta \\theta_{i,e}^{t,l}\\cdot S_{i,e}^{t,l} + \\theta_{g}^{t - 1,l}" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 318, + 305, + 374, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 305, + 374, + 313 + ], + "spans": [ + { + "bbox": [ + 318, + 305, + 374, + 313 + ], + "type": "text", + "content": "14: end for" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 318, + 316, + 364, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 316, + 364, + 325 + ], + "spans": [ + { + "bbox": [ + 318, + 316, + 364, + 325 + ], + "type": "text", + "content": "15: end for" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 318, + 327, + 449, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 327, + 449, + 339 + ], + "spans": [ + { + "bbox": [ + 318, + 327, + 449, + 339 + ], + "type": "text", + "content": "16:Upload update " + }, + { + "bbox": [ + 318, + 327, + 449, + 339 + ], + "type": "inline_equation", + "content": "\\Delta \\theta_{i}^{t}" + }, + { + "bbox": [ + 318, + 327, + 449, + 339 + ], + "type": "text", + "content": " to server" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 363, + 555, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 555, + 387 + ], + "type": "text", + "content": "real-world datasets, simulating an FL environment using the NVIDIA RTX 4090 GPU." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 394, + 429, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 429, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 429, + 407 + ], + "type": "text", + "content": "4.1. Experimental setup" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 411, + 556, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 556, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 556, + 521 + ], + "type": "text", + "content": "Datasets and Models. We evaluate the attack using two widely used benchmark datasets: CIFAR10 (50,000 training and 10,000 test samples across 10 classes) and CIFAR100 (50,000 training and 10,000 test samples across 100 classes). Following previous studies [41, 42], we use the VGG19 [30] network and ResNet18 network for CIFAR10, while we use the ResNet18 [11] network for CIFAR100. We simulate a non-IID data distribution following prior methods [5, 7, 42], with " + }, + { + "bbox": [ + 313, + 411, + 556, + 521 + ], + "type": "inline_equation", + "content": "q = 0.5 / 0.2" + }, + { + "bbox": [ + 313, + 411, + 556, + 521 + ], + "type": "text", + "content": " for CIFAR10/100, respectively." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "text", + "content": "Attack setup. In each round, we select " + }, + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "inline_equation", + "content": "n = 10" + }, + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "text", + "content": " clients among " + }, + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "inline_equation", + "content": "N = 100" + }, + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "text", + "content": " clients to participate in the aggregation, with the proportion of compromised clients set to " + }, + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "inline_equation", + "content": "C = 0.1" + }, + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "text", + "content": " by default. The FL training runs for 300 rounds to make it converge. We train for 6 local epochs (2 for the benign clients) with a learning rate 0.1 for both CIFAR10 and CIFAR100. Additionally, we assess the effectiveness of our attack in an IID setting, as defense strategies are more likely to identify malicious models as outliers in this case [42]. The trigger is a " + }, + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "inline_equation", + "content": "5 \\times 5" + }, + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "text", + "content": " pixel square located at the bottom-right corner of each image by default. We further discuss the impact of various shapes and sizes of the trigger. We evaluate our attack in the fixed-frequency attack scenario. The attacker controls a fixed number of compromised clients, which participate in FL training at a fixed frequency. Following prior" + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29166" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 128, + 537, + 430 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 125 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 125 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 125 + ], + "type": "text", + "content": "Table 1. The effectiveness of our approach against the baseline attack across various state-of-the-art (SOTA) defenses under a fixed-frequency attack setting on non-IID datasets. BA values below " + }, + { + "bbox": [ + 55, + 70, + 555, + 125 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 70, + 555, + 125 + ], + "type": "text", + "content": " in CIFAR10 (10 classes) and below " + }, + { + "bbox": [ + 55, + 70, + 555, + 125 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 55, + 70, + 555, + 125 + ], + "type": "text", + "content": " in CIFAR100 (100 classes) are highlighted in red, indicating a failed attack. The highest BA achieved in each setting is presented in bold. Our attack results are reported as " + }, + { + "bbox": [ + 55, + 70, + 555, + 125 + ], + "type": "inline_equation", + "content": "a \\pm b" + }, + { + "bbox": [ + 55, + 70, + 555, + 125 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 70, + 555, + 125 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 55, + 70, + 555, + 125 + ], + "type": "text", + "content": " denotes the mean and " + }, + { + "bbox": [ + 55, + 70, + 555, + 125 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 70, + 555, + 125 + ], + "type": "text", + "content": " represents the standard deviation. MA and BA unit: %. Avg indicates the average. Each result is obtained as the average over five independent runs." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 73, + 128, + 537, + 430 + ], + "lines": [ + { + "bbox": [ + 73, + 128, + 537, + 430 + ], + "spans": [ + { + "bbox": [ + 73, + 128, + 537, + 430 + ], + "type": "table", + "html": "
DefenseVGG19 (CIFAR10)ResNet18 (CIFAR10)ResNet18 (CIFAR100)
BadNetsDBALPOursBadNetsDBALPOursBadNetsDBALPOurs
No DefenseMA80.9781.2682.3881.81±0.3477.6678.2678.4378.70±0.8263.0562.465.1764.31±0.72
Best BA98.3346.1795.3897.05±0.1998.5920.9696.6997.60±0.21100.01.1672.2498.51±1.06
Avg BA97.7033.9690.5696.16±0.0997.9114.9395.0897.22±0.2199.730.8061.8796.70±2.46
MK [4]MA74.1775.4674.3078.03±0.3375.2476.2176.0776.49±0.8657.8458.6458.6558.73±0.55
Best BA24.063.1693.5498.01±0.806.047.5295.8998.61±0.140.880.9997.5299.94±0.01
Avg BA5.991.3376.4195.98±0.363.643.8592.4097.03±1.570.620.5489.2294.42±2.87
FLTrust [5]MA83.1483.2582.9883.84±0.2080.7580.2280.2280.33±0.2961.5861.8061.8961.14±1.81
Best BA56.236.5396.9697.95±0.3491.814.1897.1997.78±0.2730.120.6995.4299.98±0.05
Avg BA42.454.9196.6096.74±0.4789.123.6795.4891.94±8.4119.870.4687.5198.52±1.59
FLAME [23]MA58.0757.5258.6363.46±0.9174.2173.5271.9073.06±0.7655.9354.2855.6255.87±0.45
Best BA21.2037.8492.4398.38±0.6312.702.9794.4698.69±0.170.830.5377.6799.99±0.00
Avg BA7.7512.7678.8997.46±0.447.012.2789.7297.54±0.580.620.4472.8999.88±0.08
MM [12]MA69.3166.4370.7275.42±0.9771.4372.6172.8073.16±1.0753.0353.8253.4053.55±0.33
Best BA29.0816.0294.9598.31±0.4415.541.7196.0198.47±0.090.970.8592.7399.99±0.00
Avg BA7.045.4284.6194.55±4.203.911.3793.1297.52±0.670.370.6683.4697.05±1.58
DnC [29]MA81.0982.682.3281.30±1.4577.9879.3278.6378.83±0.3063.0161.4363.8963.43±0.73
Best BA5.942.8395.0497.85±0.276.5811.2096.2598.19±0.140.434.8880.9399.67±0.29
Avg BA3.681.6793.5796.94±0.534.039.7294.4597.29±0.370.412.6963.3798.07±0.59
RLR [25]MA72.2876.2277.7177.90±0.6875.7275.0674.5174.77±1.3754.7756.7856.1456.58±0.58
Best BA96.9323.0694.1197.78±0.1197.2917.1694.8398.39±0.130.140.3488.1299.98±0.01
Avg BA95.6515.1391.6197.39±0.2997.0210.8589.4197.95±0.170.050.0983.8199.97±0.02
FLARE [34]MA84.4584.2484.5384.56±0.1177.7978.2177.5277.82±2.0560.4456.5859.5459.07±1.41
Best BA98.2143.8794.7997.85±0.1298.185.8894.9196.75±0.5099.711.1488.5699.56±0.41
Avg BA98.1136.3294.1097.28±0.3098.025.4093.5596.30±0.8298.600.6380.7398.17±2.42
", + "image_path": "664947691c5df6c6c5cbeb8f1b04e534fa265586e5a00dc775acb01354b3e307.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 444, + 295, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 295, + 515 + ], + "type": "text", + "content": "approaches [2, 41, 42], we assume that the attacker controls exactly one client per round and sets the attack interval to " + }, + { + "bbox": [ + 55, + 444, + 295, + 515 + ], + "type": "inline_equation", + "content": "F = 1" + }, + { + "bbox": [ + 55, + 444, + 295, + 515 + ], + "type": "text", + "content": ". This setup allows us to analyze the attack in isolation, without client collusion. We also evaluate the attack's performance across different frequencies, comparing it to the baseline." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 531, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 295, + 601 + ], + "type": "text", + "content": "Baseline Defenses and Attacks. We assess our backdoor attack on eight state-of-the-art or representative defenses in FL: MultiKrum (MK) [4], FLTrust [5], FLAME [23], Multi-Metrics (MM) [12], DnC [29], RLR [25], FLARE [34], and DeepSight [28], ensuring a comprehensive evaluation of the attack's effectiveness." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 602, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 296, + 639 + ], + "type": "text", + "content": "We compare our backdoor attack against three state-of-the-art or representative attacks in FL: BadNets [10], DBA [35], and LP attack [42]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": "Metrics. Following prior work on FL backdoor attacks [10, 35, 42], we evaluate the global model using two primary metrics: main task accuracy (MA) and backdoor accuracy (BA). Given the dynamic nature of the global model in FL, we report both the average and the best BA over the last" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 444, + 554, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 554, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 554, + 539 + ], + "type": "text", + "content": "10 communication rounds. Furthermore, to evaluate the stealthiness of the attack, we adopt two metrics introduced in [42]: the malicious client acceptance rate (MAR) and the benign client acceptance rate (BAR). MAR measures the proportion of rounds in which malicious clients successfully bypass defense and are selected for aggregation, while BAR represents the average proportion of rounds in which benign clients are selected for aggregation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 550, + 479, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 479, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 479, + 562 + ], + "type": "text", + "content": "4.2. The effectiveness of our attack" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 568, + 555, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 555, + 688 + ], + "type": "text", + "content": "We first evaluate our attack in the fixed-frequency attack setting under different SOTA defenses. Table 1 presents the performance comparison between our attack and the baseline attacks. We also evaluate our attack in the IID setting in supplementary §B.1 and on a larger dataset TinyImageNet in supplementary §B.11. The results show that our attack achieves the highest BA in most cases. These results demonstrate the broad effectiveness of our attack, effectively bypassing defenses while maintaining superior attack performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "BadNets only works well under FLARE and consistently fails to bypass defenses such as MK, FLAME, MM, and" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29167" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 95, + 293, + 231 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 295, + 92 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 295, + 92 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 295, + 92 + ], + "type": "text", + "content": "Table 2. MAR and BAR comparison under outlier-based defenses: MK, FLAME, MM, and DnC." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 95, + 293, + 231 + ], + "lines": [ + { + "bbox": [ + 58, + 95, + 293, + 231 + ], + "spans": [ + { + "bbox": [ + 58, + 95, + 293, + 231 + ], + "type": "table", + "html": "
Dataset (Model)AttackMKFLAMEMMDnC
BARMARBARMARBARMARBARMAR
CIFAR10 (ResNets18)BadNets0.420.010.750.010.330.010.970.01
DBA0.410.010.710.010.340.010.950.01
LP attack0.360.720.650.900.130.80.390.82
Ours0.380.680.640.880.250.790.780.99
CIFAR10 (VGG19)BadNets0.410.010.720.010.310.010.980.01
DBA0.420.010.750.010.330.030.970.01
LP attack0.350.630.610.750.130.80.410.99
Ours0.380.560.60.90.260.600.790.99
CIFAR100 (ResNets18)BadNets0.430.010.680.010.330.010.990.01
DBA0.420.010.760.010.330.010.970.01
LP attack0.320.910.560.990.120.930.350.94
Ours0.340.830.580.970.230.920.720.99
", + "image_path": "df455fb8d4e898343907ca4ca722560cfcf39de61c3db0d9e9f12e23bf380456.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 241, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 241, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 241, + 296, + 396 + ], + "type": "text", + "content": "DnC across all settings, and it also fails against RLR on CIFAR100 in both non-IID and IID settings. Similarly, DBA also fails under several defenses, including MK, FLTrust, MM, and DnC under both non-IID and IID settings. LP attack achieves comparable BA across most defenses, except when evaluated under the FLARE defense in the IID setting. Specifically, for CIFAR100, our attack achieves significantly higher BA than the LP attack, demonstrating superior adaptability to complex data distributions. We also evaluate our attack performance compared with the LP attack under DeepSight in supplementary §B.3. Our attack achieves a higher BA than the LP attack across all scenarios and the LP attack fails on CIFAR100." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 414, + 296, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 414, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 414, + 296, + 485 + ], + "type": "text", + "content": "Our attack improves the attack stealthiness and maintains strength. Under the FedAvg setting, our attack achieves significantly higher BA compared to both DBA and LP attack, particularly on CIFAR100, approaching the performance of unconstrained attacks like BadNets. This demonstrates the strong strength of our attack." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 486, + 296, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 652 + ], + "type": "text", + "content": "Table 2 presents the comparison of MAR and BAR under outlier-based defenses such as MK, FLAME, MM, and DnC. A higher MAR indicates that malicious updates are more likely to bypass detection. We observe that BadNets and DBA fail to bypass defenses, as their MAR remains close to zero in all cases. This suggests that BadNets and DBA are consistently detected as outliers during training, rendering them ineffective against robust defenses. While BadNets demonstrates strong attack strength, its failure to bypass defenses undermines its overall effectiveness. In contrast, our attack achieves MAR consistently higher than BAR, meaning our malicious updates are more likely to be selected for aggregation, which reflects the stealthiness capability of our method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "When compared to the LP attack, our method maintains nearly the same or even higher MAR in defenses such as FLAME and DnC, further demonstrating its stealthiness. MK and MM impose stricter selection criteria by choosing only a small proportion of the most benign clients, making" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": "it more challenging for an attacker to be selected. LP attack achieves a higher MAR than our method under these defenses by restricting the backdoor to BC layers, which significantly enhances its stealthiness. However, our method consistently achieves higher BA than LP attack across all settings in FLAME, DnC, MK, and MM, demonstrating that while LP attack sacrifices attack strength for stealthiness, our approach achieves a better balance, ensuring both strong attack strength and high stealthiness." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 194, + 556, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 556, + 290 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 556, + 290 + ], + "type": "text", + "content": "More evaluation of attack stealthiness. We leverage Krum distance, cosine distance, and Manhattan distance to further illustrate the stealthiness of our attack. Specifically, we calculate the sum of squared Euclidean distances, cosine distances, and Manhattan distances between a model update and other local model updates, referring to them as Krum distance, Cosine distance, and Manhattan distance, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 290, + 556, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 556, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 556, + 493 + ], + "type": "text", + "content": "A larger Krum distance indicates that the update deviates significantly from benign updates, making it less likely to be accepted by the aggregation mechanism. Figure 2 shows the comparison of Krum distance with baseline attacks. The Krum distances of BadNets and DBA are significantly larger than those of benign updates, indicating that these attacks produce model updates that are easily distinguishable from benign ones. In contrast, our attack and the LP attack exhibit Krum distances close to those of benign updates, suggesting that they are more stealthy. Although DBA demonstrates slightly better stealthiness than BadNets, its Krum distance remains considerably larger than that of benign updates. This suggests that DBA is still insufficiently stealthy against robust defenses. In supplementary §B.4, we further discuss the Cosine distance and Manhattan distance between the malicious update and the benign update. The result exhibits similar trends across all attacks." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 316, + 526, + 553, + 572 + ], + "blocks": [ + { + "bbox": [ + 313, + 502, + 555, + 523 + ], + "lines": [ + { + "bbox": [ + 313, + 502, + 555, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 555, + 523 + ], + "type": "text", + "content": "Table 3. The BA of various backdoor attacks combined without/with ours." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 526, + 553, + 572 + ], + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 572 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 572 + ], + "type": "table", + "html": "
AttackIBA [24]IBA+oursCerp [20]Cerp+oursA3FL [40]A3FL+ours
FLAME2.7198.04↑29.8499.64↑94.8799.87↑
MM31.6399.99↑24.7299.98↑86.2999.89↑
", + "image_path": "e849e5a29d08a96d8565e8754f48b0feea385e10d664fb29277a5bd388eee77d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 653 + ], + "type": "text", + "content": "Our attack demonstrates strong generalizability. It can be easily integrated into existing FL backdoor attacks, particularly those based on trigger optimization. As shown in Table 3, incorporating our method further enhances attack effectiveness, underscoring its practical value." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 653, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 712 + ], + "type": "text", + "content": "We further evaluate our attack trained on other networks: ResNet50 [11], ResNet101 [11] and ViT [6] in supplementary §B.5. Our attack achieves high BA on those networks, demonstrating its generalizability to more complex architectures." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29168" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 208, + 159 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 208, + 159 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 208, + 159 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 208, + 159 + ], + "type": "image", + "image_path": "575f9f1063f4b2c8a1843a068f3c31fda3c8d70ac9a1cf73ca23e5626fb40014.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 117, + 163, + 146, + 171 + ], + "lines": [ + { + "bbox": [ + 117, + 163, + 146, + 171 + ], + "spans": [ + { + "bbox": [ + 117, + 163, + 146, + 171 + ], + "type": "text", + "content": "(a) Ours" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 221, + 70, + 373, + 159 + ], + "blocks": [ + { + "bbox": [ + 221, + 70, + 373, + 159 + ], + "lines": [ + { + "bbox": [ + 221, + 70, + 373, + 159 + ], + "spans": [ + { + "bbox": [ + 221, + 70, + 373, + 159 + ], + "type": "image", + "image_path": "b9e4c51957637677699b7c6e072df031c4c715311495a22cab8096573ff4ce2e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 275, + 162, + 318, + 171 + ], + "lines": [ + { + "bbox": [ + 275, + 162, + 318, + 171 + ], + "spans": [ + { + "bbox": [ + 275, + 162, + 318, + 171 + ], + "type": "text", + "content": "(b) LP attack" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 387, + 71, + 537, + 159 + ], + "blocks": [ + { + "bbox": [ + 387, + 71, + 537, + 159 + ], + "lines": [ + { + "bbox": [ + 387, + 71, + 537, + 159 + ], + "spans": [ + { + "bbox": [ + 387, + 71, + 537, + 159 + ], + "type": "image", + "image_path": "fba678557aa3469d28691cc5b762a1d605f02e32a1de260d54bfdf4b4ed37a6b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 426, + 162, + 497, + 171 + ], + "lines": [ + { + "bbox": [ + 426, + 162, + 497, + 171 + ], + "spans": [ + { + "bbox": [ + 426, + 162, + 497, + 171 + ], + "type": "text", + "content": "(c) BadNets and DBA" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 57, + 204, + 174, + 292 + ], + "blocks": [ + { + "bbox": [ + 55, + 175, + 555, + 197 + ], + "lines": [ + { + "bbox": [ + 55, + 175, + 555, + 197 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 555, + 197 + ], + "type": "text", + "content": "Figure 2. Krum distance of malicious and benign model updates. A larger distance indicates that the update deviates significantly from benign updates. The result shows the stealthiness of our attack." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 204, + 174, + 292 + ], + "lines": [ + { + "bbox": [ + 57, + 204, + 174, + 292 + ], + "spans": [ + { + "bbox": [ + 57, + 204, + 174, + 292 + ], + "type": "image", + "image_path": "3d67cbc95e873a5b62984da63b35810ffb0f5807b195437aa53988a87fc86a78.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 178, + 205, + 294, + 291 + ], + "blocks": [ + { + "bbox": [ + 178, + 205, + 294, + 291 + ], + "lines": [ + { + "bbox": [ + 178, + 205, + 294, + 291 + ], + "spans": [ + { + "bbox": [ + 178, + 205, + 294, + 291 + ], + "type": "image", + "image_path": "e190363b426e6b8ea4a385ac74797fe89d1f3fca396f68ccbd83b2052448cc1d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 57, + 304, + 173, + 392 + ], + "blocks": [ + { + "bbox": [ + 83, + 295, + 148, + 304 + ], + "lines": [ + { + "bbox": [ + 83, + 295, + 148, + 304 + ], + "spans": [ + { + "bbox": [ + 83, + 295, + 148, + 304 + ], + "type": "text", + "content": "(a) MK (CIFAR-10)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 304, + 173, + 392 + ], + "lines": [ + { + "bbox": [ + 57, + 304, + 173, + 392 + ], + "spans": [ + { + "bbox": [ + 57, + 304, + 173, + 392 + ], + "type": "image", + "image_path": "2105f793fb1b4346702c883ffd7a17a248aa45429ff879b3f2e018264b1c88b4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 395, + 150, + 403 + ], + "lines": [ + { + "bbox": [ + 80, + 395, + 150, + 403 + ], + "spans": [ + { + "bbox": [ + 80, + 395, + 150, + 403 + ], + "type": "text", + "content": "(c) MK (CIFAR-100)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 407, + 295, + 440 + ], + "lines": [ + { + "bbox": [ + 55, + 407, + 295, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 295, + 440 + ], + "type": "text", + "content": "Figure 3. Impacts of different attack intervals " + }, + { + "bbox": [ + 55, + 407, + 295, + 440 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 407, + 295, + 440 + ], + "type": "text", + "content": " on the attack performance. The " + }, + { + "bbox": [ + 55, + 407, + 295, + 440 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 407, + 295, + 440 + ], + "type": "text", + "content": " indicates that an attack is performed every " + }, + { + "bbox": [ + 55, + 407, + 295, + 440 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 407, + 295, + 440 + ], + "type": "text", + "content": " rounds." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 176, + 304, + 291, + 392 + ], + "blocks": [ + { + "bbox": [ + 196, + 295, + 276, + 304 + ], + "lines": [ + { + "bbox": [ + 196, + 295, + 276, + 304 + ], + "spans": [ + { + "bbox": [ + 196, + 295, + 276, + 304 + ], + "type": "text", + "content": "(b) FLAME (CIFAR-10)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 176, + 304, + 291, + 392 + ], + "lines": [ + { + "bbox": [ + 176, + 304, + 291, + 392 + ], + "spans": [ + { + "bbox": [ + 176, + 304, + 291, + 392 + ], + "type": "image", + "image_path": "e37ff7d581a43f51ac3de764c871d10da6f8db87eb5364c40f377c340ce83975.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 191, + 395, + 276, + 404 + ], + "lines": [ + { + "bbox": [ + 191, + 395, + 276, + 404 + ], + "spans": [ + { + "bbox": [ + 191, + 395, + 276, + 404 + ], + "type": "text", + "content": "(d) FLAME (CIFAR-100)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 453, + 296, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 296, + 513 + ], + "type": "text", + "content": "In addition, we analyze the impact of varying degrees of data heterogeneity in supplementary §B.6. Results show that our attack remains effective under highly non-IID conditions and consistently outperforms the LP attack across all scenarios, further confirming its generalizability." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 530, + 296, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 530, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 296, + 602 + ], + "type": "text", + "content": "Our attack maintains the global model's performance. Supplementary §B.2 shows the MA with the unpoisoned global model to assess whether our attack affects the utility of the global model. The results demonstrate that our attack does not lead to a significant decrease in MA compared to the unpoisoned global model." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 612, + 250, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 250, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 250, + 624 + ], + "type": "text", + "content": "4.3. Our attack under different scenarios" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "Impact of attack intervals. We evaluate the performance of our attack compared to baseline attacks under different attack intervals, as shown in Figure 3. The attack interval " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " ranges from 1 to 5, where " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " indicates that an attack is performed once every " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " rounds. This setting is designed to simulate scenarios where the frequency of malicious updates is limited, requiring more effective and persistent attack" + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 316, + 205, + 432, + 292 + ], + "blocks": [ + { + "bbox": [ + 316, + 205, + 432, + 292 + ], + "lines": [ + { + "bbox": [ + 316, + 205, + 432, + 292 + ], + "spans": [ + { + "bbox": [ + 316, + 205, + 432, + 292 + ], + "type": "image", + "image_path": "0b9544a783f459438d10ac4f5052f3dd28b7887cb617d1e40eec390aef14ed8a.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 436, + 205, + 552, + 292 + ], + "blocks": [ + { + "bbox": [ + 436, + 205, + 552, + 292 + ], + "lines": [ + { + "bbox": [ + 436, + 205, + 552, + 292 + ], + "spans": [ + { + "bbox": [ + 436, + 205, + 552, + 292 + ], + "type": "image", + "image_path": "f14ebdacb58da45c5f06eefbe201609d3b5da69e8a508ab25c746bf5752b4f26.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 316, + 304, + 432, + 392 + ], + "blocks": [ + { + "bbox": [ + 341, + 295, + 407, + 304 + ], + "lines": [ + { + "bbox": [ + 341, + 295, + 407, + 304 + ], + "spans": [ + { + "bbox": [ + 341, + 295, + 407, + 304 + ], + "type": "text", + "content": "(a) MK (CIFAR-10)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 304, + 432, + 392 + ], + "lines": [ + { + "bbox": [ + 316, + 304, + 432, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 432, + 392 + ], + "type": "image", + "image_path": "dc08f1af510875b2d0889e0f97dc31b2b018d876875e3843f2843c2fe0c05cbe.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 395, + 409, + 404 + ], + "lines": [ + { + "bbox": [ + 339, + 395, + 409, + 404 + ], + "spans": [ + { + "bbox": [ + 339, + 395, + 409, + 404 + ], + "type": "text", + "content": "(c) MK (CIFAR-100)" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 407, + 555, + 441 + ], + "lines": [ + { + "bbox": [ + 313, + 407, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 555, + 441 + ], + "type": "text", + "content": "Figure 4. Effectiveness of our attack under different proportions of compromised clients (" + }, + { + "bbox": [ + 313, + 407, + 555, + 441 + ], + "type": "inline_equation", + "content": "C = 0.02, 0.04, 0.06, 0.1" + }, + { + "bbox": [ + 313, + 407, + 555, + 441 + ], + "type": "text", + "content": ") in a fixed-pool attack setting." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 435, + 304, + 549, + 392 + ], + "blocks": [ + { + "bbox": [ + 454, + 295, + 535, + 304 + ], + "lines": [ + { + "bbox": [ + 454, + 295, + 535, + 304 + ], + "spans": [ + { + "bbox": [ + 454, + 295, + 535, + 304 + ], + "type": "text", + "content": "(b) FLAME (CIFAR-10)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 435, + 304, + 549, + 392 + ], + "lines": [ + { + "bbox": [ + 435, + 304, + 549, + 392 + ], + "spans": [ + { + "bbox": [ + 435, + 304, + 549, + 392 + ], + "type": "image", + "image_path": "01b7a387c13fcdeb60c840ffdd1b29e753cf21c6a92b4b3832c720fc349a8e08.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 450, + 395, + 534, + 404 + ], + "lines": [ + { + "bbox": [ + 450, + 395, + 534, + 404 + ], + "spans": [ + { + "bbox": [ + 450, + 395, + 534, + 404 + ], + "type": "text", + "content": "(d) FLAME (CIFAR-100)" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "bbox": [ + 314, + 454, + 358, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 454, + 358, + 464 + ], + "spans": [ + { + "bbox": [ + 314, + 454, + 358, + 464 + ], + "type": "text", + "content": "strategies." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 466, + 556, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 556, + 657 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 556, + 657 + ], + "type": "text", + "content": "The results show that as " + }, + { + "bbox": [ + 313, + 466, + 556, + 657 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 466, + 556, + 657 + ], + "type": "text", + "content": " increases, the performance of all attacks degrades as the attack frequency decreases. This trend is consistent across all evaluated methods, where a lower frequency results in reduced BA. A larger " + }, + { + "bbox": [ + 313, + 466, + 556, + 657 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 466, + 556, + 657 + ], + "type": "text", + "content": " imposes greater demands on the attack's strength, persistence, and stealthiness. The attack must be sufficiently powerful and stealthy to maintain its effectiveness with fewer opportunities to update the model. We observe that BadNets fails in all cases. Notably, our attack maintains a comparable BA even when " + }, + { + "bbox": [ + 313, + 466, + 556, + 657 + ], + "type": "inline_equation", + "content": "F = 5" + }, + { + "bbox": [ + 313, + 466, + 556, + 657 + ], + "type": "text", + "content": ". Whereas the LP attack loses nearly all effectiveness, particularly when trained on CIFAR100, demonstrating its vulnerability to defenses when attack opportunities are limited. We also show the results of BA under all defenses trained on CIFAR10 (ResNet18) in supplementary §B.8. We observe that our attack achieves comparable BA across all defenses when " + }, + { + "bbox": [ + 313, + 466, + 556, + 657 + ], + "type": "inline_equation", + "content": "F = 5" + }, + { + "bbox": [ + 313, + 466, + 556, + 657 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "Impact of different proportions of compromised clients. To evaluate our attack in a more practical scenario, we conduct experiments in a fixed-pool attack setting. To further" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29169" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 83, + 294, + 216 + ], + "blocks": [ + { + "bbox": [ + 91, + 71, + 259, + 82 + ], + "lines": [ + { + "bbox": [ + 91, + 71, + 259, + 82 + ], + "spans": [ + { + "bbox": [ + 91, + 71, + 259, + 82 + ], + "type": "text", + "content": "Table 4. Ablation study on alignment strategy." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 83, + 294, + 216 + ], + "lines": [ + { + "bbox": [ + 58, + 83, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 58, + 83, + 294, + 216 + ], + "type": "table", + "html": "
Dataset (Model)AttackNo DefenseFLAMEMM
BABAMARBAMAR
CIFAR-10 (VGG19)no alignment97.707.050.017.040.01
model-wise90.1245.790.9083.580.65
layer-wise (ours)96.1697.460.9097.090.67
CIFAR-10 (ResNet18)no alignment97.917.010.013.910.01
model-wise93.3985.320.8995.530.80
layer-wise (ours)97.2198.160.8897.850.79
CIFAR-100 (ResNet18)no alignment99.730.620.010.370.01
model-wise73.2985.860.9595.050.90
layer-wise (ours)96.7099.880.9797.050.92
", + "image_path": "e6df169697eb6a1fd9071d261d5f77ec159239becf7ef8b431999abefab8aab5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 230, + 296, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 230, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 230, + 296, + 326 + ], + "type": "text", + "content": "demonstrate the strength of our attack, we also analyze its performance under varying proportions of compromised clients " + }, + { + "bbox": [ + 55, + 230, + 296, + 326 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 230, + 296, + 326 + ], + "type": "text", + "content": ". In a fixed-pool attack setting, the attacker controls a pool of compromised clients, from which participants are randomly selected in each round. This setup allows multiple malicious clients to be chosen in a single round, while in some cases, several consecutive rounds may pass without selecting any malicious clients." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 327, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 422 + ], + "type": "text", + "content": "As shown in Figure 4, our attack achieves the highest BA compared to the baseline across all cases. Notably, even with a very low proportion of malicious clients (" + }, + { + "bbox": [ + 55, + 327, + 296, + 422 + ], + "type": "inline_equation", + "content": "C = 0.02" + }, + { + "bbox": [ + 55, + 327, + 296, + 422 + ], + "type": "text", + "content": ", meaning only " + }, + { + "bbox": [ + 55, + 327, + 296, + 422 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 55, + 327, + 296, + 422 + ], + "type": "text", + "content": " of clients are compromised), our attack successfully bypasses defenses and maintains a comparable BA. This demonstrates that our attack is both powerful and persistent, as it continues to perform well even when the pool of malicious clients is small." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 423, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 296, + 579 + ], + "type": "text", + "content": "Additionally, BadNets performs better when " + }, + { + "bbox": [ + 55, + 423, + 296, + 579 + ], + "type": "inline_equation", + "content": "C = 0.1" + }, + { + "bbox": [ + 55, + 423, + 296, + 579 + ], + "type": "text", + "content": " due to the increased likelihood of malicious clients colluding in the same round. This collusion significantly enhances the attack's effectiveness, as multiple attackers working together are more likely to successfully inject a backdoor. This is particularly evident under MK, where BadNets achieves a higher BA with a higher " + }, + { + "bbox": [ + 55, + 423, + 296, + 579 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 423, + 296, + 579 + ], + "type": "text", + "content": ", compared to when it fails in the fixed-frequency setting. However, this improvement diminishes as the compromised client ratio decreases. At lower " + }, + { + "bbox": [ + 55, + 423, + 296, + 579 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 423, + 296, + 579 + ], + "type": "text", + "content": ", the attack becomes more reliant on its strength. In contrast, as shown in supplementary §B.9, our attack achieves high BA across all defenses when " + }, + { + "bbox": [ + 55, + 423, + 296, + 579 + ], + "type": "inline_equation", + "content": "C > 0.04" + }, + { + "bbox": [ + 55, + 423, + 296, + 579 + ], + "type": "text", + "content": ", demonstrating its robustness even with fewer compromised clients." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 588, + 147, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 147, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 147, + 601 + ], + "type": "text", + "content": "4.4. Ablation study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": "Impact of different triggers. We further evaluate the impact of various trigger shapes. Supplementary §B.10 presents three different trigger shapes (Apple, Watermark, and Square) applied to backdoor a ResNet18 model trained on the CIFAR10 dataset. Figure 5a shows the impact of various trigger shapes. Under the default setting " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "(F = 1)" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": ", all three trigger shapes achieve high BA. The results demonstrate that our attack remains effective across different trigger shapes, indicating its robustness and generalizability." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 69, + 432, + 157 + ], + "blocks": [ + { + "bbox": [ + 317, + 69, + 432, + 157 + ], + "lines": [ + { + "bbox": [ + 317, + 69, + 432, + 157 + ], + "spans": [ + { + "bbox": [ + 317, + 69, + 432, + 157 + ], + "type": "image", + "image_path": "75ec940596f787780c656c100a20fc75b8b01c81b1653b92be4b3b53a1c1660c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 160, + 416, + 171 + ], + "lines": [ + { + "bbox": [ + 333, + 160, + 416, + 171 + ], + "spans": [ + { + "bbox": [ + 333, + 160, + 416, + 171 + ], + "type": "text", + "content": "(a) various trigger shapes" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 434, + 70, + 550, + 157 + ], + "blocks": [ + { + "bbox": [ + 434, + 70, + 550, + 157 + ], + "lines": [ + { + "bbox": [ + 434, + 70, + 550, + 157 + ], + "spans": [ + { + "bbox": [ + 434, + 70, + 550, + 157 + ], + "type": "image", + "image_path": "4a21bf3029a24c2a85ea2eb8ed226d8840c5fd01fdc3975c6ac11db85c3a90e1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 454, + 160, + 531, + 170 + ], + "lines": [ + { + "bbox": [ + 454, + 160, + 531, + 170 + ], + "spans": [ + { + "bbox": [ + 454, + 160, + 531, + 170 + ], + "type": "text", + "content": "(b) various trigger sizes" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 172, + 553, + 194 + ], + "lines": [ + { + "bbox": [ + 313, + 172, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 313, + 172, + 553, + 194 + ], + "type": "text", + "content": "Figure 5. Impact of trigger shape (left) and trigger size (right) in Flame." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 215, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 215, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 215, + 555, + 300 + ], + "type": "text", + "content": "Figure 5b illustrates the performance of different trigger sizes (3, 5, 7, and 10). Regardless of trigger size, our attack consistently achieves high BA. Notably, when the trigger size is set to 10, the attack effectiveness is significantly enhanced. Larger triggers strengthen the attack, and their ability to bypass defenses is improved through layer-wise norm constraints, leading to superior attack performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 331, + 555, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 428 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 428 + ], + "type": "text", + "content": "Impact of alignment strategy. To demonstrate how dynamic layer-wise gradient alignment enhances the effectiveness of our attack, we compare it with BadNets (no alignment) and the model-wise gradient alignment attack, which adjusts the overall gradient magnitude to align with the previous global update. As shown in Table 4, our attack consistently outperforms both BadNets and the model-wise method across all settings." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 430, + 556, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 556, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 556, + 538 + ], + "type": "text", + "content": "First, both model-wise and layer-wise gradient alignment show higher MAR compared to BadNets, indicating that aligning the update with the previous global model significantly improves the stealthiness of the attack. Second, when compared to model-wise alignment, layer-wise alignment achieves better attack performance. Specifically, even in the absence of an attack, particularly on CIFAR100, the BA for model-wise alignment is much lower than that for layer-wise alignment." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 559, + 388, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 388, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 388, + 571 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": "This paper proposes a new stealthy backdoor attack in FL, which leverages adaptive layer-wise gradient alignment to enhance stealthiness while preserving attack strength. By dynamically aligning malicious gradients to historical global updates at the layer level, the proposed attack reduces statistical anomalies caused by malicious updates, making them harder to detect. Extensive experiments across diverse datasets, models, and defenses demonstrate that the proposed attack achieves superior backdoor success rates while evading SOTA defenses, highlighting its effectiveness and stealthiness in various FL scenarios." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29170" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 137 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China (NO. 62472284, No. 62302288), Shanghai Key Laboratory of Scalable Computing and Systems, SJTU Kunpeng & Ascend Center of Excellence." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 156, + 115, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 156, + 115, + 169 + ], + "spans": [ + { + "bbox": [ + 56, + 156, + 115, + 169 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 176, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 176, + 296, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 176, + 296, + 209 + ], + "spans": [ + { + "bbox": [ + 61, + 176, + 296, + 209 + ], + "type": "text", + "content": "[1] Sana Awan, Bo Luo, and Fengjun Li. CONTRA: Defending Against Poisoning Attacks in Federated Learning. In ESORICS, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 210, + 296, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 210, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 61, + 210, + 296, + 243 + ], + "type": "text", + "content": "[2] Eugene Bagdasaryan, Andreas Veit, Yiqing Hua, Deborah Estrin, and Vitaly Shmatikov. How To Backdoor Federated Learning. In AISTATS, 2020. 1, 2, 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 244, + 296, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 244, + 296, + 286 + ], + "spans": [ + { + "bbox": [ + 62, + 244, + 296, + 286 + ], + "type": "text", + "content": "[3] Akshay Batheja and Pushpak Bhattacharyya. \"A Little is Enough\": Few-Shot Quality Estimation based Corpus Filtering improves Machine Translation. In ACL (Findings), 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 289, + 296, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 289, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 62, + 289, + 296, + 331 + ], + "type": "text", + "content": "[4] Peva Blanchard, El Mahdi El Mhamdi, Rachid Guerraoui, and Julien Stainer. Machine Learning with Adversaries: Byzantine Tolerant Gradient Descent. In NeurIPS, 2017. 1, 2, 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 334, + 295, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 334, + 295, + 366 + ], + "spans": [ + { + "bbox": [ + 62, + 334, + 295, + 366 + ], + "type": "text", + "content": "[5] Xiaoyu Cao, Minghong Fang, Jia Liu, and Neil Zhenqiang Gong. FLTrust: Byzantine-robust Federated Learning via Trust Bootstrapping. In NDSS, 2021. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 367, + 296, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 367, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 62, + 367, + 296, + 432 + ], + "type": "text", + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 434, + 296, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 434, + 296, + 466 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 296, + 466 + ], + "type": "text", + "content": "[7] Minghong Fang, Xiaoyu Cao, Jinyuan Jia, and Neil Gong. Local Model Poisoning Attacks to Byzantine-Robust Federated Learning. In USENIX Security, 2020. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 468, + 296, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 468, + 296, + 489 + ], + "spans": [ + { + "bbox": [ + 62, + 468, + 296, + 489 + ], + "type": "text", + "content": "[8] Pei Fang and Jinghui Chen. On the Vulnerability of Backdoor Defenses for Federated Learning. In AAAI, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 491, + 296, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 296, + 521 + ], + "type": "text", + "content": "[9] Clement Fung, Chris JM Yoon, and Ivan Beschastnikh. The Limitations of Federated Learning in Sybil Settings. In RAID, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 524, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 524, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 57, + 524, + 295, + 555 + ], + "type": "text", + "content": "[10] Tianyu Gu, Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. BadNets: Evaluating Backdooring Attacks on Deep Neural Networks. IEEE Access, 7:47230-47244, 2019. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 557, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 557, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 557, + 296, + 590 + ], + "type": "text", + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In CVPR, 2016. 4, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 591, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 296, + 624 + ], + "type": "text", + "content": "[12] Siquan Huang, Yijiang Li, Chong Chen, Leyu Shi, and Ying Gao. Multi-Metrics Adaptively Identifies Backdoors in Federated Learning. In ICCV, 2023. 1, 2, 3, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 625, + 296, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 296, + 668 + ], + "type": "text", + "content": "[13] Tiansheng Huang, Sihao Hu, Ka Ho Chow, Fatih Ilhan, Selim F. Tekin, and Ling Liu. Lockdown: Backdoor Defense for Federated Learning with Isolated Subspace Training. In NeurIPS, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "text", + "content": "[14] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and Open Problems in Federated" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 335, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 554, + 95 + ], + "type": "text", + "content": "Learning. Foundations and trends® in machine learning, 14 (1-2):1-210, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 96, + 554, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 554, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 554, + 129 + ], + "type": "text", + "content": "[15] Yun-Yong Ko, Dongwon Lee, and Sang-Wook Kim. Not All Layers Are Equal: A Layer-Wise Adaptive Approach Toward Large-Scale DNN Training. In WWW, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 130, + 554, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 554, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 554, + 162 + ], + "type": "text", + "content": "[16] Jakub Konečny. Federated learning: Strategies for improving communication efficiency. arXiv preprint arXiv:1610.05492, 2016. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 164, + 554, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 554, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 554, + 196 + ], + "type": "text", + "content": "[17] Torsten Krauß and Alexandra Dmitrienko. MESAS: Poisoning Defense for Federated Learning Resilient against Adaptive Attackers. In CCS, 2023. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 198, + 554, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 554, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 554, + 241 + ], + "type": "text", + "content": "[18] Haoyang Li, Qingqing Ye, Haibo Hu, Jin Li, Leixia Wang, Chengfang Fang, and Jie Shi. 3DFed: Adaptive and Extensible Framework for Covert Backdoor Attack in Federated Learning. In S&P, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 243, + 554, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 554, + 276 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 554, + 276 + ], + "type": "text", + "content": "[19] Ye Li, Yanchao Zhao, Chengcheng Zhu, and Jiale Zhang. Infighting in the dark: Multi-labels backdoor attack in federated learning. arXiv preprint arXiv:2409.19601, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 277, + 554, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 554, + 320 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 554, + 320 + ], + "type": "text", + "content": "[20] Xiaoting Lyu, Yufei Han, Wei Wang, Jingkai Liu, Bin Wang, Jiqiang Liu, and Xiangliang Zhang. Poisoning with Cerberus: Stealthy and Colluded Backdoor Attack against Federated Learning. In AAAI, 2023. 2, 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 322, + 554, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 322, + 554, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 322, + 554, + 365 + ], + "type": "text", + "content": "[21] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y Arcas. Communication-Efficient Learning of Deep Networks from Decentralized Data. In AISTATS, 2017. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 366, + 554, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 554, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 554, + 399 + ], + "type": "text", + "content": "[22] El Mahdi El Mhamdi, Rachid Guerraoui, and Sébastien Rouault. The Hidden Vulnerability of Distributed Learning in Byzantium. In ICML, 2018. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 400, + 554, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 400, + 554, + 466 + ], + "spans": [ + { + "bbox": [ + 316, + 400, + 554, + 466 + ], + "type": "text", + "content": "[23] Thien Duc Nguyen, Phillip Rieger, Huili Chen, Hossein Yalame, Helen Möllering, Hossein Fereidooni, Samuel Marchal, Markus Miettinen, Azalia Mirhoseini, Shaza Zeitouni, Farinaz Koushanfar, Ahmad-Reza Sadeghi, and Thomas Schneider. FLAME: Taming Backdoors in Federated Learning. In USENIX Security 22, 2022. 1, 2, 3, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 468, + 554, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 554, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 554, + 499 + ], + "type": "text", + "content": "[24] Thuy Dung Nguyen, Tuan Nguyen, Anh Tran, Khoa D. Doan, and Kok-Seng Wong. IBA: Towards Irreversible Backdoor Attacks in Federated Learning. In NeurIPS, 2023. 2, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 501, + 554, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 554, + 534 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 554, + 534 + ], + "type": "text", + "content": "[25] Mustafa Safa Ozdayi, Murat Kantarcioglu, and Yulia R Gel. Defending against backdoors in federated learning with robust learning rate. In AAAI, 2021. 1, 3, 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 535, + 554, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 554, + 578 + ], + "type": "text", + "content": "[26] Ashwinee Panda, Saeed Mahloujifar, Arjun Nitin Bhagoji, Supriyo Chakraborty, and Prateek Mittal. SparseFed: Mitigating Model Poisoning Attacks in Federated Learning with Sparsification. In AISTATS, 2022. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 580, + 554, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 580, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 580, + 554, + 635 + ], + "type": "text", + "content": "[27] Matthias Paulik, Matt Seigel, Henry Mason, Dominic Telaar, Joris Kluivers, Rogier Van Dalen, Chi Wai Lau, Luke Carlson, Filip Granqvist, Chris Vandevelde, et al. Federated Evaluation and Tuning for On-Device Personalization: System Design & Applications. arXiv preprint arXiv:2102.08503, 2021. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 635, + 554, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 554, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 554, + 679 + ], + "type": "text", + "content": "[28] Phillip Rieger, Thien Duc Nguyen, Markus Miettinen, and Ahmad-Reza Sadeghi. DeepSight: Mitigating Backdoor Attacks in Federated Learning Through Deep Model Inspection. In NDSS, 2022. 2, 3, 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "text", + "content": "[29] Virat Shejwalkar and Amir Houmansadr. Manipulating the Byzantine: Optimizing Model Poisoning Attacks and Defenses for Federated Learning. In NDSS, 2021. 1, 2, 3, 5" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29171" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 611 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "text", + "content": "[30] Karen Simonyan. Very Deep Convolutional Networks for Large-Scale Image Recognition. arXiv preprint arXiv:1409.1556, 2014. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "text", + "content": "[31] Peng Sun, Xinyang Liu, Zhibo Wang, and Bo Liu. Byzantine robust Decentralized Federated Learning via Dual-domain Clustering and Trust Bootstrapping. In CVPR, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "type": "text", + "content": "[32] Ziteng Sun, Peter Kairouz, Ananda Theertha Suresh, and H Brendan McMahan. Can You Really Backdoor Federated Learning? arXiv preprint arXiv:1911.07963, 2019. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 228 + ], + "type": "text", + "content": "[33] Hongyi Wang, Kartik Sreenivasan, Shashank Rajput, Harit Vishwakarma, Saurabh Agarwal, Jy-yong Sohn, Kangwook Lee, and Dimitris S. Papailiopoulos. Attack of the Tails: Yes, You Really Can Backdoor Federated Learning. In NeurIPS, 2020. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "type": "text", + "content": "[34] Ning Wang, Yang Xiao, Yimin Chen, Yang Hu, Wenjing Lou, and Y. Thomas Hou. FLARE: Defending Federated Learning against Model Poisoning Attacks via Latent Space Representations. In AsiaCCS, 2022. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 295, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 295, + 307 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 295, + 307 + ], + "type": "text", + "content": "[35] Chulin Xie, Keli Huang, Pin-Yu Chen, and Bo Li. DBA: Distributed Backdoor Attacks against Federated Learning. In ICLR, 2020. 1, 2, 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 310, + 295, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 295, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 295, + 352 + ], + "type": "text", + "content": "[36] Yueqi Xie, Minghong Fang, and Neil Zhenqiang Gong. FedREDefence: Defending against Model Poisoning Attacks for Federated Learning using Model Update Reconstruction Error. In ICML, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 354, + 295, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 295, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 295, + 387 + ], + "type": "text", + "content": "[37] Jiahao Xu, Zikai Zhang, and Rui Hu. Identify Backdoored Model in Federated Learning via Individual Unlearning. arXiv preprint arXiv:2411.01040, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 388, + 295, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 295, + 430 + ], + "type": "text", + "content": "[38] Peishen Yan, Hao Wang, Tao Song, Yang Hua, Ruhui Ma, Ningxin Hu, Mohammad Reza Haghighat, and Haibing Guan. SkyMask: Attack-Agnostic Robust Federated Learning With Fine-Grained Learnable Masks. In ECCV, 2025. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 434, + 295, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 295, + 486 + ], + "type": "text", + "content": "[39] Timothy Yang, Galen Andrew, Hubert Eichner, Haicheng Sun, Wei Li, Nicholas Kong, Daniel Ramage, and Françoise Bea-fays. Applied Federated Learning: Improving Google Keyboard Query Suggestions. arXiv preprint arXiv:1812.02903, 2018. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 489, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 295, + 521 + ], + "type": "text", + "content": "[40] Hangfan Zhang, Jinyuan Jia, Jinghui Chen, Lu Lin, and Ding-hao Wu. A3FL: Adversarily Adaptive Backdoor Attacks to Federated Learning. In NeurIPS, 2023. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 523, + 295, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 295, + 576 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 295, + 576 + ], + "type": "text", + "content": "[41] Zhengming Zhang, Ashwinee Panda, Linyue Song, Yaoqing Yang, Michael W. Mahoney, Prateek Mittal, Kannan Ramchandran, and Joseph Gonzalez. Neurotoxin: Durable Backdoors in Federated Learning. In ICML, 2022. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 578, + 295, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 295, + 611 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 295, + 611 + ], + "type": "text", + "content": "[42] Haomin Zhuang, Mingxian Yu, Hao Wang, Yang Hua, Jian Li, and Xu Yuan. Backdoor Federated Learning by Poisoning Backdoor-Critical Layers. In ICLR, 2024. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29172" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SteerX_ Creating Any Camera-Free 3D and 4D Scenes with Geometric Steering/a06f04c2-df21-4c58-83ac-45d3ea205c44_content_list.json b/2025/SteerX_ Creating Any Camera-Free 3D and 4D Scenes with Geometric Steering/a06f04c2-df21-4c58-83ac-45d3ea205c44_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4ce0c6f1789dce7a5dcb96c150d47b8339153c16 --- /dev/null +++ b/2025/SteerX_ Creating Any Camera-Free 3D and 4D Scenes with Geometric Steering/a06f04c2-df21-4c58-83ac-45d3ea205c44_content_list.json @@ -0,0 +1,2120 @@ +[ + { + "type": "text", + "text": "SteerX: Creating Any Camera-Free 3D and 4D Scenes with Geometric Steering", + "text_level": 1, + "bbox": [ + 98, + 130, + 898, + 152 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fac72fcde383d71a8cd37a08dbe546fb0ec531de0c396a6ad5b940595e7c277f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 179, + 178, + 816, + 265 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/af83eea19b4ebc5216dbfba59eabd762a2347e84f80d75e241d21a4118168a25.jpg", + "image_caption": [ + "Text-to-4D" + ], + "image_footnote": [ + "A carpet covers the living room floor, and a puppy is lying on it playing with a kitten. To the right of the carpet, there is a toy ball." + ], + "bbox": [ + 102, + 299, + 488, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/709b9410f7dd3068975e60ebb2e7ee39d15211698082fdfc378200689da6d189.jpg", + "image_caption": [ + "Text-to-3D" + ], + "image_footnote": [ + "A pink bicycle leaning against a fence", + "A pair of sneakers" + ], + "bbox": [ + 506, + 299, + 890, + 392 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5a9ab954fb11ca1e13d26d72a5ae6712baf93bd1084f2ecdf131db87e3334cc4.jpg", + "image_caption": [ + "SteerX" + ], + "image_footnote": [], + "bbox": [ + 102, + 436, + 498, + 578 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f5caf242268bd3e7d583e2c4b85817feba91542361dc4e8f4fa48e20abd960be.jpg", + "image_caption": [ + "Image-to-3D", + "Figure 1. SteerX is a zero-shot inference-time steering approach that seamlessly integrates video generative models [24, 35, 59, 62, 80] and feed-forward scene reconstruction models [24, 61, 86], enabling any 3D and 4D scene generation without explicit camera conditions." + ], + "image_footnote": [], + "bbox": [ + 500, + 455, + 898, + 578 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 633, + 326, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent progress in 3D/4D scene generation emphasizes the importance of physical alignment throughout video generation and scene reconstruction. However, existing methods improve the alignment separately at each stage, making it difficult to manage subtle misalignments arising from another stage. Here, we present SteerX, a zero-shot inference-time steering method that unifies scene reconstruction into the generation process, tilting data distributions toward better geometric alignment. To this end, we introduce two geometric reward functions for 3D/4D scene generation by using pose-free feed-forward scene reconstruction models. Through extensive experiments, we demonstrate the effectiveness of SteerX in improving 3D/4D scene generation.", + "bbox": [ + 89, + 667, + 483, + 863 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 633, + 643, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generating 3D and 4D scenes from images or text prompts has attracted significant attention due to its potential applications in AR/VR and robotics [10, 63, 76]. This progress is largely driven by the advancement of generative models [7, 25, 35, 62, 80] and neural scene representations [33, 46, 73]. Generative models learn the underlying distribution of large-scale and high-quality video data, leveraging their scalability without imposing explicit physical constraints. In contrast, neural scene representations lift these distributions into structured 3D or 4D spaces, enforcing physical consistency and enabling more faithful scene modeling.", + "bbox": [ + 509, + 660, + 906, + 825 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this context, recent efforts [2, 3, 23, 27, 59, 71, 75, 87] have focused on producing geometrically consistent images by fine-tuning generative models with camera pose parameters, where the generation process follows user-defined or pre-defined camera trajectories. This facilitates seamless 3D", + "bbox": [ + 511, + 825, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution, †Corresponding author", + "bbox": [ + 107, + 887, + 344, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27326", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and 4D scene reconstructions but requires complex optimization to regress neural scene representations, increasing computational overhead and hindering practical adaptation.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this inefficiency, another line of work [24, 37] has introduced text-conditioned camera pose generation and feed-forward decoding of pixel-aligned 3DGS [33]. This approach trains 3DGS decoders to learn a mapping function that directly reconstructs 3D scenes from multi-view images, leveraging multiple 3D scene datasets for improved 3D scene modeling. However, the limited datasets used to train pose-conditioned decoders fail to capture the various camera trajectories produced by video generative models. This results in slight misalignments between video generation and scene reconstruction, requiring further refinement steps [53, 79, 89] to enhance geometric consistency.", + "bbox": [ + 88, + 136, + 480, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To sum up, previous works handle geometric alignment separately in either video generation or scene reconstruction. This makes it difficult to address cross-stage misalignments, as inconsistencies in one stage may not be fully corrected in the other. Despite recent efforts to mitigate this issue, achieving precise alignment remains an ongoing challenge due to the indistinct link between the two stages.", + "bbox": [ + 89, + 318, + 480, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the other hand, zero-shot guidance methods that alter the sampling trajectory of generative models to enforce physical constraints have been widely explored in recent literature [17, 64]. These methods have shown great generalizability and performance, contingent upon a well-defined reward. For instance, in inverse imaging problems, a closed-form likelihood function can guide prior sampling processes toward posterior sampling [13, 57]. However, directly applying zero-shot guidance to 3D/4D scene generation remains challenging due to the ambiguity in defining reward functions and the significant computational overhead.", + "bbox": [ + 89, + 424, + 482, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we address this gap by introducing SteerX, a zero-shot inference-time steering method that seamlessly integrates video generation and scene reconstruction, generating geometrically aligned high-quality 3D and 4D scenes. To achieve this, we define geometric reward functions to assess physical consistency across video frames, drawing inspiration from cycle consistency [90] to guide the generation process toward high-reward outputs. We propose two geometric reward functions tailored for 3D and 4D scene generation. To evaluate geometric consistency across multiple video frames, we extend MEt3R [1], a recent evaluation metric for image pairs, by incorporating advanced pose-free feed-forward scene reconstruction methods such as MV-DUSt3R+ [61] and MonST3R [86]. These reconstruction methods lift intermediate generated video frames' during the reverse sampling process into 3D and 4D spaces. The reconstructed scenes are then projected back into the original image space for consistency evaluation.", + "bbox": [ + 89, + 590, + 482, + 862 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, even when the reward is defined, one cannot", + "bbox": [ + 109, + 863, + 480, + 876 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "directly adopt standard gradient guidance approaches [13, 14, 57] for several reasons. For one, gradient guidance can be used only when the reward is fully differentiable, which hampers flexibility in the design of reward functions. Moreover, there exist substantial memory constraints to compute gradients for dozens of video frames, limiting their scalability for long video sequences. Therefore, we propose a steering algorithm based on sequential Monte Carlo (SMC) [20], which has gained recent traction [19, 36, 55, 74] due to its applicability to non-differentiable reward functions and favorable inference-time scaling properties.", + "bbox": [ + 511, + 90, + 903, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By designing tailored reward functions for geometrically plausible and accurate 3D/4D scene generation, along with a guided sampling process based on SMC, SteerX serves as a fully general framework that integrates any generative video models with any 3D reconstruction models, enabling diverse tasks, including Image-to-3D, Image-to-4D, Text-to-3D, and Text-to-4D generation. Through extensive experiments on both 3D and 4D scene generation with various pre-trained video generative models [24, 35, 59, 62, 80], we demonstrate the effectiveness and broad applicability of our approach. Furthermore, we show that by increasing the number of particles, we see favorable scaling properties, opening up new possibilities that have been rather untouched in 3D/4D scene generation: test-time scaling.", + "bbox": [ + 511, + 262, + 903, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 503, + 661, + 520 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. 3D and 4D Scene Reconstruction", + "text_level": 1, + "bbox": [ + 511, + 535, + 800, + 550 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Various neural scene representations have been explored for 3D and 4D scene reconstruction. Most recent works tend to use Neural Radiance Fields (NeRF) [46] and 3D Gaussian Splats (3DGS) [33] for its high-quality rendering of underlying static scenes [4, 5, 85]. These scene representations have been extended to reconstruct 4D scenes by incorporating additional supervisions (e.g., depth maps [78] and segmentation masks [34]) along with deformable fields [38, 47, 50, 51, 67] and Gaussian splats [65, 73].", + "bbox": [ + 511, + 561, + 903, + 699 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advancements in 3D and 4D scene datasets [16, 40, 41, 54, 82, 84, 88] have enabled the development of generalizable scene representations, allowing feed-forward scene reconstruction methods [9, 12, 15, 60, 69]. Further advancements have extended these techniques to pose-free feed-forward 3D reconstruction models [56, 61, 68, 81, 86], eliminating the need for camera pose conditions and enabling the scene reconstruction from hundreds of unposed images, making the process more flexible and scalable. We leverage these powerful pose-free 3D reconstruction models, specifically MV-DUSt3R+ [61] and MonST3R [86], to assess geometric consistency and lift generated video frames into 3D and 4D spaces.", + "bbox": [ + 511, + 704, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Denoised predictions, as used in [13]", + "bbox": [ + 91, + 886, + 295, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27327", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. 3D and 4D Scene Generation", + "text_level": 1, + "bbox": [ + 89, + 90, + 349, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent 3D and 4D scene generation approaches commonly adopt a two-stage framework, where video generation is followed by scene reconstruction, with a focus on enhancing geometric consistency across both stages. Some works fine-tune video generative models with camera pose parameters [2, 23, 27, 42, 71, 75, 83, 87], where the generation process is conditioned on a user-defined or pre-defined camera trajectory to produce geometrically aligned video frames. These camera-conditioned video frames are used as training data to regress neural scene representations. While they produce plausible scenes, optimizing scene representations from scratch incurs high computational costs.", + "bbox": [ + 89, + 112, + 483, + 294 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast, camera-free approaches [24, 37] reconstruct 3D scenes without camera conditions, improving geometric alignment in the scene reconstruction stage. They generate text-conditioned camera poses and use a feed-forward 3DGS decoder trained on diverse 3D scene datasets to estimate underlying 3D structures from posed multi-view images. However, these methods often fail to fully capture the diverse camera trajectories produced by video generative models. Instead of addressing alignment separately at each stage, SteerX integrates scene reconstruction models directly into the video generation process, ensuring that video frames are optimally structured for accurate scene reconstruction and enabling seamless integration of both stages.", + "bbox": [ + 89, + 294, + 483, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Guided Sampling", + "text_level": 1, + "bbox": [ + 89, + 497, + 264, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generative models are trained to represent data distributions across various domains, including images [30, 48, 49, 52], videos [26, 31], and 3D [43, 72]. However, sampling from tilted distributions (e.g., posterior distributions from measurements) is often preferable to drawing from naive prior.", + "bbox": [ + 89, + 520, + 482, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent works [6, 77] have encoded user preferences through reward functions and fine-tuned generative models to maximize these rewards, effectively sampling from a tilted distribution toward high-reward outputs. However, fine-tuning approaches require extensive training and hampers the generalizability by modifying the base distribution.", + "bbox": [ + 89, + 595, + 483, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast, guided sampling methods [13, 17, 57, 58, 64] allow sampling from a tilted distribution in a zero-shot manner while keeping the base distribution intact. Among various paradigms, two classes stand out: gradient-based guidance [13, 57] and particle filtering[19, 36, 74]. Gradient-based guidance is simple to implement and widely applied but requires differentiable reward functions and is memory-intensive. In contrast, particle filtering is amenable to non-differentiable rewards, and is more memory-efficient, as it does not rely on backpropagation. Recently, Feynman-Kac Steering (FKS) [55] has unified previous particle filtering techniques, introducing a general steering framework that", + "bbox": [ + 89, + 686, + 483, + 867 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "can be applied to any reward function. SteerX builds upon FKS, integrating geometric reward functions and exploring a tailored design space for geometric steering, making it applicable to any video generative models.", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary: Feynman-Kac Steering", + "text_level": 1, + "bbox": [ + 511, + 165, + 844, + 184 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models [18, 58] are learned to reverse a forward process $\\mathbf{x}_t \\sim q(\\mathbf{x}_t | \\mathbf{x}_0 = \\mathbf{x})$ with $T$ timesteps, that gradually transforms a data sample $\\mathbf{x}$ into Gaussian noise. This reverse process $p_{\\theta}(\\mathbf{x}_{t-1} | \\mathbf{x}_t)$ is defined as:", + "bbox": [ + 511, + 191, + 905, + 252 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(\\mathbf {x} _ {T}, \\mathbf {x} _ {T - 1}, \\dots , \\mathbf {x} _ {1}, \\mathbf {x} _ {0}\\right) \\propto \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 263, + 905, + 305 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main goal of inference-time steering methods is to sample tilted distributions toward maximizing an exponential of user-specified reward function $r_{\\phi}(\\mathbf{x}_0)$ as:", + "bbox": [ + 511, + 316, + 905, + 362 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {p} _ {\\theta} \\left(\\mathbf {x} _ {0}\\right) \\propto p _ {\\theta} \\left(\\mathbf {x} _ {0}\\right) \\exp \\left(\\lambda r _ {\\phi} \\left(\\mathbf {x} _ {0}\\right)\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 373, + 905, + 390 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Feynman-Kac Steering (FKS) [55] introduces a system of multiple interacting diffusion paths (i.e. particles), where each path $\\mathbf{x}_{T:0} = (\\mathbf{x}_T,\\dots ,\\mathbf{x}_0)$ is resampled at intermediate steps based on potentials $G_{t}(\\mathbf{x}_{T:t})$ , resulting in a sequence of tilted distributions $\\tilde{p}_{\\theta}(\\mathbf{x}_{T:t})$ :", + "bbox": [ + 511, + 402, + 905, + 479 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {p} _ {\\theta} \\left(\\mathbf {x} _ {T: t}\\right) \\propto \\prod_ {s = T} ^ {t} p _ {\\theta} \\left(\\mathbf {x} _ {T: s}\\right) G _ {s} \\left(\\mathbf {x} _ {T: s}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 489, + 903, + 531 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the product of potentials matches the total tilt of the base distribution as $\\prod_{t=T}^{0} G_{t}(\\mathbf{x}_{T:t}) = \\exp(\\lambda r_{\\phi}(\\mathbf{x}_{0}))$ . These potentials are generally defined as SMC [74] to score the particles at each transition step $\\mathbf{x}_{t+1} \\rightarrow \\mathbf{x}_{t}$ based on intermediate rewards $r_{\\phi}(\\mathbf{x}_{t})$ .", + "bbox": [ + 511, + 541, + 906, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methods", + "text_level": 1, + "bbox": [ + 511, + 631, + 612, + 645 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce SteerX, a zero-shot inference-time steering method for 3D/4D scene generation. SteerX unifies pose-free feed-forward scene reconstruction models into the video generation process, iteratively tilting the data distribution towards geometrically aligned samples. In Section 4.1, we define two geometric reward functions to evaluate geometric consistency in generated multi-view images and dynamic videos, respectively. In Section 4.2, we detail our SMC-based steering algorithm.", + "bbox": [ + 511, + 656, + 905, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Geometric Rewards", + "text_level": 1, + "bbox": [ + 511, + 801, + 702, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our geometric rewards build upon MEt3R [1], which utilizes DUSt3R [68] to measure feature similarity in overlapping regions between image pairs. It extracts upscaled DINO features [8, 22] and computes the cosine similarity between reference view features and the rendered features", + "bbox": [ + 511, + 825, + 905, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2 Throughout the paper, we flexibly use the term particle filtering and SMC to refer to this class.", + "bbox": [ + 89, + 875, + 482, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27328", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/646add79ba657e36d048a37e04351cbde2394b7c39bd44d61bd6119ccf47364b.jpg", + "image_caption": [ + "Figure 2. An overview of geometric rewards. Our reward functions assess the geometric consistency of intermediate generated video frames by computing the feature similarity of upscaled DINO features. (a) GS-MEt3R evaluates feature similarity between the original video frames and their corresponding rendered images from 3DGS. (b) Dyn-MEt3R focuses on background regions by unprojecting background features from half of the video frames and reprojecting them onto the remaining frames to compute feature similarity." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 897, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "from another viewpoint. It can evaluate multiple images by computing feature similarity across all possible image pairs and averaging the scores. However, as the number of images increases, this becomes computationally expensive, making the overall generation process inefficient.", + "bbox": [ + 88, + 421, + 480, + 496 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this, as shown in Fig. 2, we introduce two geometric reward functions, GS-MEt3R and Dyn-MEt3R, which measure geometric consistency across video frames tailored for 3D/4D scene generation, respectively. We employ pose-free feed-forward scene reconstruction methods, MV-DUSt3R+ [61] and MonST3R [86], to reconstruct 3D and 4D scenes, which are subsequently projected back into image space to assess consistency with the original frames. This mitigates computational bottlenecks while efficiently evaluating global consistency in the generated video frames.", + "bbox": [ + 88, + 497, + 482, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D Scene Reward. We extend MEt3R [1] in distinct ways to better suit 3D and 4D scene reconstruction models. For 3D scenes, recent methods [61, 81] introduce a mapping function $f_{\\phi}$ that directly reconstructs 3DGS and camera poses from $N$ sparse views $\\{I_i \\in \\mathbb{R}^{H \\times W \\times 3}\\}_{i=1}^N$ as:", + "bbox": [ + 89, + 654, + 482, + 734 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\phi}: \\left\\{I _ {i} \\right\\} _ {i = 1} ^ {N} \\rightarrow \\left\\{\\begin{array}{l}\\left\\{\\boldsymbol {\\mu} _ {j}, \\mathbf {o} _ {j}, \\boldsymbol {\\Sigma} _ {j}, \\mathbf {c} _ {j} \\right\\} _ {j = 1} ^ {N \\times H \\times W}\\\\\\left\\{P _ {i} \\right\\} _ {i = 1} ^ {N}\\end{array}, \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 741, + 482, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the 3D scene is represented as Gaussian parameters, including position $\\pmb{\\mu}$ , volume density $\\mathbf{o}$ , covariance $\\pmb{\\Sigma}$ , and color $\\mathbf{c}$ . Then, we produce images $\\{\\tilde{I}_i\\}_{i=1}^N$ by rendering the scene with estimated camera poses $\\{P_i\\}_{i=1}^N$ , ensuring they correspond to the same viewpoints as the input images. Finally, GS-MEt3R is measured by computing the cosine similarity between upscaled DINO [8, 22] features of input im", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ages $\\{F_i\\}_{i = 1}^N$ and rendered images $\\{\\hat{F}_i\\}_{i = 1}^N$ as:", + "bbox": [ + 511, + 419, + 823, + 439 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nr _ {\\phi} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {H} \\sum_ {k = 1} ^ {W} \\frac {F _ {i} ^ {j k} \\cdot \\hat {F} _ {i} ^ {j k}}{| | F _ {i} ^ {j k} | | | | \\hat {F} _ {i} ^ {j k} | |}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 446, + 906, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This approach not only offers a more direct evaluation of the physical consistency of the 3D scene compared to averaging consistency across all image pair combinations but also indirectly assesses the rendering quality of 3DGS. In other words, high GS-MEt3R scores indicate both geometric alignment and visually realistic 3D scenes.", + "bbox": [ + 511, + 497, + 906, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4D Scene Reward. While the 3D scene reward function is based on 3DGS, feed-forward dynamic scene reconstruction with Gaussians remains underexplored, making it difficult to directly apply 3DGS-based rewards. Instead, we employ 3D point cloud representations, where MonST3R [86] reconstructs it with point maps $\\{X_{i}\\}_{i = 1}^{N}$ , binary dynamic masks $\\{M_i\\}_{i = 1}^N$ , and camera poses $\\{P_i\\}_{i = 1}^N$ as:", + "bbox": [ + 511, + 592, + 905, + 700 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\phi}: \\left\\{I _ {i} \\right\\} _ {i = 1} ^ {N} \\rightarrow \\left\\{X _ {i}, M _ {i}, P _ {i} \\right\\} _ {i = 1} ^ {N}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 705, + 903, + 726 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we leverage these time-varying point clouds as 4D scene representations and design a reward function for evaluating geometric consistency in dynamic videos.", + "bbox": [ + 511, + 734, + 903, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since dynamic masks are produced in the camera pose estimation process to retain only high-confidence points, a well-reconstructed 4D scene should effectively filter out dynamic objects while preserving geometric consistency in the background regions. Therefore, we evaluate the consistency only for background regions of video frames, which are not filtered out by the dynamic mask. To this end, we first split $N$ video frames into two subsets: $\\mathcal{I}_{src} = \\{I_1, I_3, \\ldots, I_N\\}$", + "bbox": [ + 511, + 780, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27329", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 SteerX (v-prediction)", + "text_level": 1, + "bbox": [ + 91, + 93, + 325, + 108 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Required: v-parametrized diffusion model $\\mathbf{v}_{\\theta}$ , reward function $r_{\\phi}$ , number of particles $k$ , and initial noise $\\{\\mathbf{x}_T^i\\}_{i=1}^k \\sim \\mathcal{N}(0, I)$ .", + "bbox": [ + 91, + 109, + 482, + 157 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sampling:", + "text_level": 1, + "bbox": [ + 91, + 156, + 166, + 170 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: for $t \\in \\{T - 1, \\dots, 0\\}$ do", + "2: for $i\\in \\{1\\dots k\\}$ do", + "3: $\\hat{\\mathbf{x}}_0^i\\gets \\sqrt{\\bar{\\alpha}_{t + 1}}\\mathbf{x}_{t + 1}^i -\\sqrt{1 - \\bar{\\alpha}_{t + 1}}\\mathbf{v}_\\theta (\\mathbf{x}_{t + 1}^i)$", + "4: $\\mathbf{x}_t^i\\gets dpm\\text{- solver} (\\hat{\\mathbf{x}}_0^i,\\mathbf{x}_{t + 1}^i)$", + "5: $\\mathbf{s}_t^i\\gets r_\\phi (\\hat{\\mathbf{x}}_0^i)$ Intermediaterewards", + "6: $G_{t}^{i}\\gets \\exp (\\lambda \\max_{j = t}^{T}(\\mathbf{s}_{j}^{i}))$ $\\triangleright$ Potential", + "7: end for", + "8: $\\{\\mathbf{x}_t^i\\}_{i=1}^k \\sim \\text{Multinomial}(\\{\\mathbf{x}_t^i, G_t^i\\}_{i=1}^k) \\triangleright \\text{Resample}$", + "9: end for", + "10: $l\\gets \\arg \\max_{i\\in \\{1,\\ldots ,k\\}}r_{\\phi}(\\mathbf{x}_0^i)$", + "11: return $\\mathbf{x}_0^l$" + ], + "bbox": [ + 96, + 172, + 478, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and $\\mathcal{I}_{tgt} = \\{I_2, I_4, \\ldots, I_{N-1}\\}$ . Then, we unproject the upsampled DINO features of background regions in $\\mathcal{I}_{src}$ into 3D space using MonST3R. Finally, we reproject them onto the viewpoint of $\\mathcal{I}_{tgt}$ , where the rendered features $\\hat{\\mathcal{F}}_{tgt} = \\{\\hat{F}_1, \\hat{F}_3, \\ldots, \\hat{F}_N\\}$ are used to compute the feature similarity with background regions in $\\mathcal{I}_{tgt}$ as:", + "bbox": [ + 89, + 396, + 483, + 489 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr _ {i} = \\sum_ {j = 1} ^ {H} \\sum_ {k = 1} ^ {W} \\left(1 - M _ {i} ^ {j k}\\right) \\frac {F _ {i} ^ {j k} \\cdot \\hat {F} _ {i} ^ {j k}}{| | F _ {i} ^ {j k} | | | | \\hat {F} _ {i} ^ {j k} | |}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 503, + 482, + 547 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr _ {\\phi} = \\frac {1}{(N / / 2)} \\sum_ {i = 1} ^ {N / / 2} r _ {i}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 569, + 482, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With these geometric rewards, we can effectively steer pre-trained video generative models to produce geometrically consistent video frames, which are then directly reconstructed into 3D and 4D spaces using the feed-forward 3D reconstruction models employed in the geometric rewards.", + "bbox": [ + 89, + 625, + 483, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Geometric Steering", + "text_level": 1, + "bbox": [ + 89, + 715, + 277, + 732 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Using the rewards defined in Section 4.1, let $\\tilde{p}_{\\theta}$ in (2) be the distribution we wish to sample from. SMC operates with the three following steps:", + "bbox": [ + 89, + 739, + 482, + 786 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. (Proposal) For each particle $i$ , sample from the proposal distribution $\\mathbf{x}_t^i \\sim q_t(\\mathbf{x}_t|\\mathbf{x}_{t+1}^i)$", + "2. (Weighting) Compute weights from reward-based potentials $\\omega_{t}^{i} = \\frac{p_{\\theta}(\\mathbf{x}_{t}^{i}|\\mathbf{x}_{t + 1}^{i})}{q_{t}(\\mathbf{x}_{t}^{i}|\\mathbf{x}_{t + 1}^{i})} G_{t}(\\mathbf{x}_{T:t}^{i})$", + "3. (Resampling) Draw new particles from the multinomial distribution $\\{\\mathbf{x}_t^i\\}_{i=1}^k \\sim \\text{Multinomial}(\\{\\mathbf{x}_t^i, G_t^i\\}_{i=1}^k)$" + ], + "bbox": [ + 89, + 787, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Two choices should be made: the potential $G_{t}$ , and the proposal distribution $q_{t}$ . For the potential, we use max potential", + "bbox": [ + 511, + 90, + 906, + 122 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nG _ {t} \\left(\\mathbf {x} _ {T: t}\\right) ^ {i} := \\exp \\left(\\lambda \\max _ {j = t} ^ {T} \\left[ r _ {\\phi} \\left(\\hat {\\mathbf {x}} _ {0}\\right) \\right]\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 127, + 906, + 160 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with", + "bbox": [ + 513, + 164, + 547, + 178 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nG _ {0} \\left(\\mathbf {x} _ {T: 0}\\right) := \\exp \\left(\\lambda r _ {\\phi} (\\mathbf {x} _ {0})\\right) \\left(\\prod_ {t = 1} ^ {T} G _ {t} \\left(\\mathbf {x} _ {T: t}\\right)\\right) ^ {- 1}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 181, + 905, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "such that the particle with the highest reward is preferred. Notice that we use the Tweedie estimate $\\hat{\\mathbf{x}}_0 = \\mathbb{E}[\\mathbf{x}_0|\\mathbf{x}_t]$ [13, 21, 55] in intermediate steps to avoid full reverse sampling. For the proposal kernel, to save computation, we leverage DPM-solver++ [44], which approximates the true sampling trajectory limited to small neural function evaluation (NFE). These choices lead to SteerX, as shown in Alg. 1.", + "bbox": [ + 511, + 231, + 906, + 337 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposition 1. Given the reverse generative process in (1), let $q_{t}$ be the transition kernel satisfying", + "bbox": [ + 511, + 343, + 905, + 373 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right)}{q _ {t} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right)} = 1 + \\epsilon_ {t} \\left(\\mathbf {x} _ {T: t}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 378, + 905, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with $|\\epsilon_t(\\mathbf{x}_{T:t})| \\leq \\varepsilon$ uniformly. Also assume that the error from the reward computed at the approximate state $\\hat{\\mathbf{x}}_0$ is bounded, i.e. $|r_{\\phi}(\\hat{\\mathbf{x}}_0) - r_{\\phi}(\\mathbf{x}_0)| \\leq \\eta$ . Then, given the defined max potentials in (9), (10), Alg. 1 samples from", + "bbox": [ + 511, + 416, + 906, + 477 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {p} _ {\\theta} \\left(\\mathbf {x} _ {0}\\right) \\propto p _ {\\theta} \\left(\\mathbf {x} _ {0}\\right) \\exp \\left(\\lambda r _ {\\phi} \\left(\\mathbf {x} _ {0}\\right)\\right) \\left(1 + \\mathcal {O} \\left(T \\varepsilon + \\lambda \\eta\\right)\\right) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 484, + 905, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "See Appendix A for the proof. Prop. 1 states that when the approximation errors are sufficiently small, then we can sample from the desired tilted distribution with SteerX.", + "bbox": [ + 511, + 506, + 905, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 563, + 718, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we conduct extensive experiments to verify the scalability and effectiveness of SteerX across various video generative models in four scene generation scenarios: Text-to-4D, Image-to-4D, Text-to-3D, and Image-to-3D. We briefly explain our experimental setup in Section 5.1 and present both qualitative and quantitative results, as well as design choices of SteerX in Section 5.2.", + "bbox": [ + 511, + 589, + 906, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 700, + 702, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. For video generative models, we generate 25 frames at a $480 \\times 480$ resolution, except for models that require a fixed video length. Since reconstructing 3D scenes using all generated video frames is impractical, we use only eight frames for the reconstruction. We follow the default settings of MV-DUST3R+ [61] and MonST3R [86]. We set $\\lambda = 10$ for the potential. To encourage exploration and reduce computation overhead, we adopt interval resampling, as in FKS [55], and apply the steering procedure at $M$ selected timesteps out of $T$ reverse diffusion steps, setting $M = 4$ for 3D scene generation and $M = 2$ for 4D scene generation.", + "bbox": [ + 511, + 723, + 906, + 905 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27330", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f5061d47d90e7fce6062e396346ab58f785bc1621882e8c56922f04aba5e23e7.jpg", + "image_caption": [ + "Figure 3. Qualitative results of video generation in PenguinVideoBenchmark [35]. We visualize for Mochi [62] (top) and Hunyuan-Video [35] (bottom). SteerX achieves the best alignment with camera motion instructions while maintaining high video quality." + ], + "image_footnote": [], + "bbox": [ + 94, + 85, + 903, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines and Benchmarks. We evaluate our SteerX with Mochi [62] and HunyuanVideo [35] for Text-to-4D generation, and CogVideoX [80] for Image-to-4D. For Image-to-3D, we employ S-Director with orbit-left from DimensionX [59]. For Text-to-3D, we utilize the video generation and scene reconstruction models proposed in SplatFlow [24]. We conduct evaluations on 98 samples from the PenguinVideoBenchmark [35] with camera descriptions for Text-to-4D generation, and 355 samples from VBench-I2V [32] for Image-to-3D and Image-to-4D generation. Additionally, we use 100 samples from the Single-Object-with-Surrounding set of T3Bench [28] for Text-to-3D evaluation. For all evaluations, we compare against the best-of-N approach, where $k$ particles are generated independently, and the one with the highest reward is selected. Additionally, we include a baseline ( $k = 1$ ) that generates a video and reconstructs the scene without inference-time steering.", + "bbox": [ + 88, + 421, + 485, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. For 4D generation, we evaluated the Aesthetic Score and Subject Consistency in VBench [32] for video quality and semantic consistency, respectively. Also, we evaluated Temporal Consistency to measure text alignment in Text-to-4D generation, including adherence to camera motion instructions. We evaluated Dynamic Degree to measure the overall dynamicness in Image-to-4D generation. For 3D generation, we followed the evaluation protocols in Director3D [37] and SplatFlow [24], measuring the image quality and the CLIP score [29] for both generated multi-view images and rendered images of 3DGS. Finally, for all generation scenarios, we evaluated the geometric consistency of multi-view images and dynamic videos using GS-MEt3R and Dyn-MEt3R, respectively.", + "bbox": [ + 89, + 688, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f5520b64dd3e29c763553668d7c9057f501db4c72e1ea70c16e7c8372ed0890e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodkAesthetic↑Subject↑Temporal↑Dyn-MEt3R↑
Mochi [62]10.4910.9500.2430.884
+ BoN40.4980.9410.2440.912
+ SteerX (later)40.4880.9370.2420.910
+ SteerX (linear)40.4900.9490.2440.918
+ SteerX (early)40.5000.9550.2480.929
HunyuanVideo [35]10.5490.9670.2410.911
+ BoN40.5510.9780.2390.931
+ SteerX (later)40.5550.9760.2370.931
+ SteerX (linear)40.5560.9730.2410.943
+ SteerX (early)40.5550.9800.2430.964
", + "bbox": [ + 517, + 419, + 903, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative results in PenguinVideoBenchmark [35]. SteerX with annealed resampling timesteps to the early sampling process significantly improves the overall video quality.", + "bbox": [ + 511, + 583, + 906, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Main Results", + "text_level": 1, + "bbox": [ + 513, + 643, + 653, + 659 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Text-to-4D Generation. Figure 3 and Table 1 present results of text-conditioned video generation. Notably, SteerX significantly outperforms both the baseline and the best-of-N, confirming that our geometric steering effectively resamples particles to guide the data distribution toward geometrically aligned samples. Moreover, quantitative results show a strong correlation between Dyn-MEt3R and other metrics, suggesting that the geometric steering can contribute to both physical consistency and overall video quality.", + "bbox": [ + 511, + 670, + 906, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We observe that annealed resampling timesteps, where the steering procedure at the early sampling stages outperforms linear resampling, whereas resampling at later stages tends to degrade performance. This observation aligns with previous findings [2, 59], where camera poses for video frames are largely determined early in the sampling process", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27331", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ee45af8a7e241095acc82e50c751535804dd7a12fe92f4cde4a60b01d7b25727.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 88, + 176, + 150 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bcf97e68a83e7c14e10d0f5e94237808b29a4f17d1b1aa7f376f4c1a3789e809.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 96, + 152, + 176, + 213 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ecadeb1c4079f1a3511ce20995afb3f1d43ee7962258d2d74753088e153ab0c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 178, + 88, + 256, + 150 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8a0956ed05cfa5a60368219f950f215715aef7518c15cd595eeb868649ed6c17.jpg", + "image_caption": [ + "CogVideoX [80]" + ], + "image_footnote": [], + "bbox": [ + 178, + 152, + 256, + 213 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1c45db49760c6149d82ab4ecd209aa54bd1800b3dff0d68875c03936b22f0976.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 256, + 88, + 336, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ba17b7765c03118d83b021179e4364c54ee4d6c43bdc608f240d83faac7d94df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 256, + 152, + 336, + 213 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/982b968d258ce237842b6c2aed398ee1e9b99bbed83d5b63cb2fdf4c38743576.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 338, + 88, + 418, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/466e437bb0013c49fa561d7086243876faa82ed619d69e13fa1ec92231875992.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 338, + 152, + 418, + 213 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6e811a60c73167e59857e283b7f10cd3876fb91b26c46ddecfbc5a3920e4758f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 88, + 500, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e46a43182df5084ddbede4cf1c2b5c4ad2fe4a5096d65521df97898a475c5811.jpg", + "image_caption": [ + "+Best-of-N" + ], + "image_footnote": [], + "bbox": [ + 419, + 152, + 500, + 213 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6a385882fdc381c7082d2f3b668416fc7d25668713ab4e3bb095469d8ef6c57c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 88, + 578, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e622220c3b33a3918c8189caf5ab11c5f132f3c3a1d48dd6f61b77b9f73a5187.jpg", + "image_caption": [ + "" + ], + "image_footnote": [], + "bbox": [ + 500, + 152, + 604, + 213 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6576cde83e0ade8241f1a4f8cd62198407309dfbda2c77f367ac54d1c4c00ef8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 580, + 88, + 658, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4a9a9412f541315a38a8799109ed3ca16927e6768924b2971ed9f964958614ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 580, + 152, + 658, + 213 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e8abb12535553004d9599928fe56f7558a09d656d56d42169f8faab183116de7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 88, + 738, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f8d22e0d534ce1a77e6c2e8f4c4d026fedbae33dc8bc6c695ac86b93e76825c7.jpg", + "image_caption": [ + "+", + "+ SteerX (early)" + ], + "image_footnote": [], + "bbox": [ + 660, + 152, + 738, + 213 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/839473ecd6b6cfda4099eea2caf07a2f8bce7be7b83078b6b52d7b40509e1c5c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 740, + 88, + 820, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4c261f6a2d834073a74ac687ead7e57e2c93ccbfbcecf97abb2a2fc43406aa03.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 740, + 152, + 820, + 213 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/231b2ef285506b99170364dac01d2b99ec0fa50b5f71d4d564bd22b9f7d56a76.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 820, + 88, + 901, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0e2142af8b404a66c8656c5084606171d3c56cbc6613060154c8b03c00e550a5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 820, + 152, + 901, + 213 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e7d36517131aac71bef94157cf52bef874e6f92395918606317829051654fe82.jpg", + "image_caption": [ + "Figure 4. Qualitative results of video generation in VBench-I2V [32]. SteerX enhances vividness and overall visual quality.", + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 94, + 255, + 220, + 354 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cc064c04aba8a7f50b9c88ea584fdc1d50818ad364045be0c43d4be86bede913.jpg", + "image_caption": [ + "CogVideoX", + "Figure 5. Qualitative results in Image-to-4D. SteerX naturally lifts object motion into 4D spaces, while preserving geometric alignments." + ], + "image_footnote": [], + "bbox": [ + 248, + 255, + 344, + 358 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a3a62951dcc98557cf040428f95041591c2a93de2a2b08855aabe0c8385e6606.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 255, + 450, + 358 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d13790f3bb9c0ccb395dca33c02ea81afeb7e1ccf33a99996d3313f7fff7c3be.jpg", + "image_caption": [ + "+Best-of-N" + ], + "image_footnote": [], + "bbox": [ + 480, + 257, + 578, + 358 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6d8e77e74264f9871ea0c0a79e110ef591fdb98280dcb72ab702e669d59949ec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 257, + 674, + 357 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2810a864dc46e4fa0fdf53a60ca6af54fde743458903575f60bf1590d3393bb0.jpg", + "image_caption": [ + "+SteerX" + ], + "image_footnote": [], + "bbox": [ + 705, + 255, + 803, + 354 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5afe9d78bdb1c380a7e1bed8d80cc2ec80967c497afb6d479b942a92c107534a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 805, + 255, + 903, + 354 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/55aa1f8a1453425da741d00f8419f0a10c20590fa60b0be403ef1ad8d5ee0937.jpg", + "image_caption": [ + "\"Filmed from a first-person perspective, the camera passes through the graffiti alley in Melbourne, Australia, where the graffiti walls are covered with artwork from many artists.\"", + "HunyuanVideo", + "Figure 6. Qualitative results in Text-to-4D." + ], + "image_footnote": [], + "bbox": [ + 96, + 424, + 210, + 549 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/989a8e2ad5d2d6eccafae54599f5088f8e6ed865584572d9dd5d8cae42ed6137.jpg", + "image_caption": [ + "+BoN" + ], + "image_footnote": [], + "bbox": [ + 230, + 424, + 341, + 545 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a8a5463465a6f306c5212cccec1a788bd51c39cca942a43a505be904cc3354a5.jpg", + "image_caption": [ + "+SteerX" + ], + "image_footnote": [], + "bbox": [ + 367, + 424, + 470, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "as the diffusion model conceptualizes global structures. As a result, resampling at early timesteps coarsely aligns the geometry by tilting the data distribution, laying the foundation for a more precise geometric alignment in later stages.", + "bbox": [ + 89, + 598, + 482, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The geometrically aligned video frames contribute to better 4D scene reconstruction, as shown in Fig. 6. While the baseline struggles to capture accurate camera poses and the best-of-N approach results in blurry 4D scenes, our approach precisely estimates camera poses and generates visually realistic 4D scenes. This highlights the effectiveness of SteerX in the seamless integration of video generation and scene reconstruction, ensuring that generated video frames are optimally structured for precise scene reconstruction.", + "bbox": [ + 89, + 659, + 483, + 795 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Image-to-4D Generation. Figure 4 and Table 2 show the results in image-conditioned video generation, and our SteerX performs the best for all metrics, and we observe that SteerX generates dynamic and realistic video frames, whereas the baseline produces almost motionless frames, and the best-of-N approach results in blurry frames. Also,", + "bbox": [ + 89, + 809, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8df505d2f84997743dbaec395125722ef8f25273dadb2631b164408a3bf627a5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodkAesthetic↑Subject↑Dynamic↑Dyn-MEt3R↑
CogVideoX [80]10.5920.9450.1580.880
+ BoN20.5910.9410.1410.882
+ SteerX (early)20.5930.9450.1610.894
+ BoN40.5940.9440.1430.901
+ SteerX (early)40.5960.9570.1700.909
", + "bbox": [ + 517, + 398, + 901, + 497 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0c276462cff6ae5c76a7582cecd3ae0e211dc58b50545d77a7c77c5f2fb5e549.jpg", + "table_caption": [ + "Table 2. Quantitative results in VBench-I2V [32]. Results demonstrate the scalability of geometric steering as $k$ increases." + ], + "table_footnote": [], + "table_body": "
MethodkRenderingSample
BRISQUE↓NIQE↓CLIP-I↑GS-MEt3R↑
DimensionX [59]137.34.2582.40.708
+ BoN429.84.3383.20.745
+ SteerX (early)429.74.2483.70.749
", + "bbox": [ + 517, + 537, + 901, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Quantitative results in VBench-I2V [32]. For rendered images, we report results after the refinement process in [24, 37].", + "bbox": [ + 511, + 622, + 903, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "we validate the scalability of SteerX, as increasing the number of particles $k$ consistently enhances performance across all metrics. As shown in Fig. 5, geometrically aligned video frames can be reconstructed to better 4D scenes, producing realistic dynamic objects and well-aligned backgrounds.", + "bbox": [ + 511, + 667, + 906, + 743 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Image-to-3D Generation. We verify the effectiveness of SteerX in the Image-to-3D scene generation. While multiview images generated by SteerX closely resemble those from the best-of-N approach, SteerX reconstructs 3D scenes with finer details and fewer discontinuities in novel views, as shown in Fig. 8. Furthermore, quantitative results in Table 3 demonstrate that SteerX achieves better visual quality and geometric consistency in both rendered images and generated images compared to other baselines.", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27332", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c3fd68343b530168c1e529c434e32f085f21e34acc9418d2f878a38dcd1df63d.jpg", + "image_caption": [ + "Figure 7. Qualitative results of rendered images in T3Bench [28]. SteerX significantly enhances the visual quality of rendered images and textual alignment, demonstrating its compatibility with various video generation and scene reconstruction models." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 893, + 333 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4e2def90e8c2fcd5cad7440959e2b65a7e4683dfdf92e104dc5c4a9404925294.jpg", + "image_caption": [ + "(a) Input Image" + ], + "image_footnote": [], + "bbox": [ + 99, + 373, + 209, + 539 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/820bdb7269a762b7a3b1189562b133d970d4e4b222c07d5be8cf2530e1210fdb.jpg", + "image_caption": [ + "(b) DimensionX + BoN" + ], + "image_footnote": [], + "bbox": [ + 227, + 373, + 339, + 539 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/18469372ca97bdffd3ab3b6f87a17c59a6b587540ad8ac11aead7a0c7f5cb5dd.jpg", + "image_caption": [ + "(c) DimensionX + SteerX" + ], + "image_footnote": [], + "bbox": [ + 359, + 373, + 470, + 539 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a15c32776e1ec7863ea312a20ccbffe00e8c8c87393e8cc28b217c8b05820d58.jpg", + "table_caption": [ + "Figure 8. Qualitative results in of 3DGS in VBench-I2V [32]." + ], + "table_footnote": [], + "table_body": "
MethodkBRISQUE↓CLIPScore↑GS-MEt3R↑
SplatFlow [24]133.2529.660.727
+ BoN236.8229.550.756
+ SteerX (early)227.4729.560.767
+ BoN428.1429.280.768
+ SteerX (early)426.6529.730.775
", + "bbox": [ + 94, + 575, + 478, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Text-to-3D Generation. Notice that SteerX can be applied to any generative model and scene reconstruction model, including the multi-view rectified flow model and the feed-forward 3DGS decoder in SplatFlow [24], enabling geometric steering through GS-MEt3R rewards. Table 4 presents quantitative results of generated multi-view images, demonstrating that SteerX effectively enhances overall performance and further improves scalability as the number of particles increases. Furthermore, as shown in Table 5, these 3D-aligned multi-view images are integrated with existing refinement processes, achieving state-of-the-art performance in 3D scene generation. Figure 7 presents qualita", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/89755f25e0b4dc182d45c9a39a34b74d5e9542d922c2754241cbd023327ed289.jpg", + "table_caption": [ + "Table 4. Quantitative results of multi-view generation in Single-Object-with-Surrounding set of T3Bench [28]." + ], + "table_footnote": [], + "table_body": "
MethodBRISQUE↓NIQE↓CLIPScore↑
DreamFusion [53]90.210.48-
Magic3D [39]92.811.20-
LatentNeRF [45]88.69.19-
SJC [66]82.010.15-
Fantasia3D [11]69.67.65-
ProlificDreamer [70]61.57.07-
Director3D [37]32.34.3532.9
SplatFlow [24]19.64.2433.2
SplatFlow [24]†23.44.8432.7
+ BoN (k=4)17.24.4132.3
+ SteerX (k=4)13.14.3033.4
", + "bbox": [ + 517, + 372, + 901, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Quantitative results in T3Bench [28]. SplatFlow† is re-implemented results without the stop ray strategy, which is incompatible with our geometric steering pipeline.", + "bbox": [ + 511, + 558, + 903, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tive results, demonstrating that SteerX produces more text-aligned 3D scenes and visually realistic rendered images.", + "bbox": [ + 511, + 614, + 903, + 646 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 662, + 633, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have introduced SteerX, a zero-shot inference-time steering method for camera-free 3D and 4D scene generation. Instead of addressing physical alignment separately in either video generation or scene reconstruction, SteerX unifies both stages and iteratively tilts the data distribution toward geometrically consistent samples. Extensive experiments across diverse scene generation tasks verify that SteerX effectively enhances visual quality, textual alignment, and geometric consistency. SteerX is practical, enabling zero-shot 3D and 4D scene generation, and is scalable, with geometric alignment improving as the number of particles increases. We believe this efficient scaling property holds great potential, opening new avenues for 3D and 4D scene generation, particularly in test-time scaling.", + "bbox": [ + 509, + 688, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27333", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mohammad Asim, Christopher Wewer, Thomas Wimmer, Bernt Schiele, and Jan Eric Lenssen. Met3r: Measuring multi-view consistency in generated images. arXiv preprint arXiv:2501.06336, 2025. 2, 3, 4", + "[2] Sherwin Bahmani, Ivan Skorokhodov, Guocheng Qian, Aliaksandr Siarohin, Willi Menapace, Andrea Tagliasacchi, David B Lindell, and Sergey Tulyakov. Ac3d: Analyzing and improving 3d camera control in video diffusion transformers. arXiv preprint arXiv:2411.18673, 2024. 1, 3, 6", + "[3] Sherwin Bahmani, Ivan Skorokhodov, Aliaksandr Siarohin, Willi Menapace, Guocheng Qian, Michael Vasilkovsky, Hsin-Ying Lee, Chaoyang Wang, Jiaxu Zou, Andrea Tagliasacchi, et al. Vd3d: Taming large video diffusion transformers for 3d camera control. arXiv preprint arXiv:2407.12781, 2024. 1", + "[4] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5855–5864, 2021. 2", + "[5] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 2", + "[6] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. In The Twelfth International Conference on Learning Representations, 2024. 3", + "[7] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1", + "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 3, 4", + "[9] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 2", + "[10] Guikun Chen and Wenguan Wang. A survey on 3d gaussian splatting. arXiv preprint arXiv:2401.03890, 2024. 1", + "[11] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22246–22256, 2023. 8", + "[12] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splattering from sparse" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "multi-view images. In European Conference on Computer Vision, pages 370-386. Springer, 2024. 2", + "[13] Hyungjin Chung, Jeongsol Kim, Michael T McCann, Marc L Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. arXiv preprint arXiv:2209.14687, 2022. 2, 3, 5", + "[14] Hyungjin Chung, Suhyeon Lee, and Jong Chul Ye. Decomposed diffusion sampler for accelerating large-scale inverse problems. arXiv preprint arXiv:2303.05754, 2023. 2", + "[15] Wenyan Cong, Hanxue Liang, Peihao Wang, Zhiwen Fan, Tianlong Chen, Mukund Varma, Yi Wang, and Zhangyang Wang. Enhancing nerf akin to enhancing llms: Generalizable nerf transformer with mixture-of-view-experts. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3193-3204, 2023. 2", + "[16] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2", + "[17] Giannis Daras, Hyungjin Chung, Chieh-Hsin Lai, Yuki Mitsufuji, Jong Chul Ye, Peyman Milanfar, Alexandros G Dimakis, and Mauricio Delbracio. A survey on diffusion models for inverse problems. arXiv preprint arXiv:2410.00083, 2024. 2, 3", + "[18] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 3", + "[19] Zehao Dou and Yang Song. Diffusion posterior sampling for linear inverse problem solving: A filtering perspective. In The Twelfth International Conference on Learning Representations, 2024. 2, 3", + "[20] Arnaud Doucet, Nando De Freitas, and Neil Gordon. An introduction to sequential monte carlo methods. Sequential Monte Carlo methods in practice, pages 3-14, 2001. 2", + "[21] Bradley Efron. Tweedie's formula and selection bias. Journal of the American Statistical Association, 106(496):1602-1614, 2011. 5", + "[22] Stephanie Fu, Mark Hamilton, Laura Brandt, Axel Feldman, Zhoutong Zhang, and William T Freeman. Fitup: A model-agnostic framework for features at any resolution. arXiv preprint arXiv:2403.10516, 2024. 3, 4", + "[23] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 1, 3", + "[24] Hyojun Go, Byeongjun Park, Jiho Jang, Jin-Young Kim, Soonwoo Kwon, and Changick Kim. Splatflow: Multiview rectified flow model for 3d gaussian splatting synthesis. arXiv preprint arXiv:2411.16443, 2024. 1, 2, 3, 6, 7, 8, 13", + "[25] Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, Poriya Panet, Sapir Weissbuch, Victor Kulikov, Yaki Bitterman, Zeev Melumian, and Ofir Bibi. Ltx-video: Realtime video latent diffusion. arXiv preprint arXiv:2501.00103, 2024. 1" + ], + "bbox": [ + 517, + 93, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27334", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible diffusion modeling of long videos. Advances in Neural Information Processing Systems, 35:27953-27965, 2022. 3", + "[27] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrl: Enabling camera control for text-to-video generation. arXiv preprint arXiv:2404.02101, 2024. 1, 3", + "[28] Yuze He, Yushi Bai, Matthieu Lin, Wang Zhao, Yubin Hu, Jenny Sheng, Ran Yi, Juanzi Li, and Yong-Jin Liu. T $^3$ bench: Benchmarking current progress in text-to-3d generation, 2023. 6, 8", + "[29] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 6", + "[30] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3", + "[31] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 3", + "[32] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 6, 7, 8", + "[33] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 2", + "[34] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 2", + "[35] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. 1, 2, 6, 13, 14", + "[36] Xiner Li, Yulai Zhao, Chenyu Wang, Gabriele Scalia, Gokcen Eraslan, Surag Nair, Tommaso Biancalani, Shuiwang Ji, Aviv Regev, Sergey Levine, et al. Derivative-free guidance in continuous and discrete diffusion models with soft value-based decoding. arXiv preprint arXiv:2408.08252, 2024. 2, 3", + "[37] Xinyang Li, Zhangyu Lai, Linning Xu, Yansong Qu, Liujuan Cao, Shengchuan Zhang, Bo Dai, and Rongrong Ji. Director3d: Real-world camera trajectory and 3d scene generation from text. Advances in Neural Information Processing Systems, 37:75125-75151, 2025. 2, 3, 6, 7, 8", + "[38] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 2", + "[39] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 300–309, 2023. 8", + "[40] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024. 2", + "[41] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14458-14467, 2021. 2", + "[42] Fangfu Liu, Wenqiang Sun, Hanyang Wang, Yikai Wang, Haowen Sun, Junliang Ye, Jun Zhang, and Yueqi Duan. Reconx: Reconstruct any scene from sparse views with video diffusion model. arXiv preprint arXiv:2408.16767, 2024. 3", + "[43] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 3", + "[44] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 5", + "[45] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 8", + "[46] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2", + "[47] Byeongjun Park and Changick Kim. Point-dynrf: Point-based dynamic radiance fields from a monocular video. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3171-3181, 2024. 2", + "[48] Byeongjun Park, Sangmin Woo, Hyojun Go, Jin-Young Kim, and Changick Kim. Denoising task routing for diffusion models. arXiv preprint arXiv:2310.07138, 2023. 3", + "[49] Byeongjun Park, Hyojun Go, Jin-Young Kim, Sangmin Woo, Seokil Ham, and Changick Kim. Switch diffusion transformer: Synergizing denoising tasks with sparse mixture-of-experts. arXiv preprint arXiv:2403.09176, 2024. 3", + "[50] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5865-5874, 2021. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27335", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 2", + "[52] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195-4205, 2023. 3", + "[53] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 8", + "[54] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10901-10911, 2021. 2", + "[55] Raghav Singhal, Zachary Horvitz, Ryan Teehan, Mengye Ren, Zhou Yu, Kathleen McKeown, and Rajesh Ranganath. A general framework for inference-time scaling and steering of diffusion models. arXiv preprint arXiv:2501.06848, 2025. 2, 3, 5", + "[56] Brandon Smart, Chuanxia Zheng, Iro Laina, and Victor Adrian Prisacariu. Splatt3r: Zero-shot gaussian splatting from uncalibrated image pairs. arXiv preprint arXiv:2408.13912, 2024. 2", + "[57] Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. Pseudoinverse-guided diffusion models for inverse problems. In International Conference on Learning Representations, 2023. 2, 3", + "[58] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3", + "[59] Wenqiang Sun, Shuo Chen, Fangfu Liu, Zilong Chen, Yueqi Duan, Jun Zhang, and Yikai Wang. Dimensionx: Create any 3d and 4d scenes from a single image with controllable video diffusion. arXiv preprint arXiv:2411.04928, 2024. 1, 2, 6, 7", + "[60] Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10208-10217, 2024. 2", + "[61] Zhenggang Tang, Yuchen Fan, Dilin Wang, Hongyu Xu, Rakesh Ranjan, Alexander Schwing, and Zhicheng Yan. Mvdust3r+: Single-stage scene reconstruction from sparse views in 2 seconds. arXiv preprint arXiv:2412.06974, 2024. 1, 2, 4, 5", + "[62] Genmo Team. Mochi 1. https://github.com/genmoai/models, 2024. 1, 2, 6, 13, 14", + "[63] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Wang Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. In Computer Graphics Forum, pages 703-735. Wiley Online Library, 2022. 1" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Masatoshi Uehara, Xingyu Su, Yulai Zhao, Xiner Li, Aviv Regev, Shuiwang Ji, Sergey Levine, and Tommaso Biancalani. Reward-guided iterative refinement in diffusion models at test-time with applications to protein and dna design. arXiv preprint arXiv:2502.14944, 2025. 2, 3", + "[65] Diwen Wan, Ruijie Lu, and Gang Zeng. Superpoint gaussian splatting for real-time high-fidelity dynamic scene reconstruction. arXiv preprint arXiv:2406.03697, 2024. 2", + "[66] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 8", + "[67] Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. 2", + "[68] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 2, 3", + "[69] Yunsong Wang, Tianxin Huang, Hanlin Chen, and Gim Hee Lee. Freesplat: Generalizable 3d gaussian splatting towards free view synthesis of indoor scenes. Advances in Neural Information Processing Systems, 37:107326-107349, 2025. 2", + "[70] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 8", + "[71] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1, 3", + "[72] Sangmin Woo, Byeongjun Park, Hyojun Go, Jin-Young Kim, and Changick Kim. Harmonyview: Harmonizing consistency and diversity in one-image-to-3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10574–10584, 2024. 3", + "[73] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20310-20320, 2024. 1, 2", + "[74] Luhuan Wu, Brian Trippe, Christian Naesseth, David Blei, and John P Cunningham. Practical and asymptotically exact conditional sampling in diffusion models. Advances in Neural Information Processing Systems, 36:31372-31403, 2023. 2, 3", + "[75] Rundi Wu, Ruiqi Gao, Ben Poole, Alex Trevithick, Changxi Zheng, Jonathan T Barron, and Aleksander Holynski. Cat4d: Create anything in 4d with multi-view video diffusion models. arXiv preprint arXiv:2411.18613, 2024. 1, 3" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "27336", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[76] Tong Wu, Yu-Jie Yuan, Ling-Xiao Zhang, Jie Yang, Yan-Pei Cao, Ling-Qi Yan, and Lin Gao. Recent advances in 3d gaussian splatting. Computational Visual Media, 10(4):613-642, 2024. 1", + "[77] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: learning and evaluating human preferences for text-to-image generation. In Proceedings of the 37th International Conference on Neural Information Processing Systems, pages 15903-15935, 2023. 3", + "[78] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 2", + "[79] Xiaofeng Yang, Yiwen Chen, Cheng Chen, Chi Zhang, Yi Xu, Xulei Yang, Fayao Liu, and Guosheng Lin. Learn to optimize denoising scores for 3d generation: A unified and improved diffusion prior on nerf and 3d gaussian splatting. arXiv preprint arXiv:2312.04820, 2023. 2", + "[80] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 2, 6, 7, 14", + "[81] Botao Ye, Sifei Liu, Haofei Xu, Xueting Li, Marc Pollefeys, Ming-Hsuan Yang, and Songyou Peng. No pose, no problem: Surprisingly simple 3d gaussian splats from sparse unposed images. arXiv preprint arXiv:2410.24207, 2024. 2, 4", + "[82] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-22, 2023. 2", + "[83] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 3", + "[84] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Chenming Zhu, Zhangyang Xiong, Tianyou Liang, et al. Mvmgnet: A large-scale dataset of multi-view images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9150-9161, 2023. 2", + "[85] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19447-19456, 2024. 2", + "[86] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arXiv:2410.03825, 2024. 1, 2, 4, 5", + "[87] Yuyang Zhao, Chung-Ching Lin, Kevin Lin, Zhiwen Yan, Linjie Li, Zhengyuan Yang, Jianfeng Wang, Gim Hee Lee, and Lijuan Wang. Genxd: Generating any 3d and 4d scenes. arXiv preprint arXiv:2411.02319, 2024. 1, 3" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[88] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018. 2", + "[89] Junzhe Zhu, Peiye Zhuang, and Sanmi Koyejo. Hifa: High-fidelity text-to-3d generation with advanced diffusion guidance. arXiv preprint arXiv:2305.18766, 2023. 2", + "[90] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2" + ], + "bbox": [ + 516, + 92, + 903, + 258 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "27337", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/SteerX_ Creating Any Camera-Free 3D and 4D Scenes with Geometric Steering/a06f04c2-df21-4c58-83ac-45d3ea205c44_model.json b/2025/SteerX_ Creating Any Camera-Free 3D and 4D Scenes with Geometric Steering/a06f04c2-df21-4c58-83ac-45d3ea205c44_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d0075b4f21141db5b9e28713099537ca6a3e031e --- /dev/null +++ b/2025/SteerX_ Creating Any Camera-Free 3D and 4D Scenes with Geometric Steering/a06f04c2-df21-4c58-83ac-45d3ea205c44_model.json @@ -0,0 +1,3370 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.099, + 0.131, + 0.9, + 0.154 + ], + "angle": 0, + "content": "SteerX: Creating Any Camera-Free 3D and 4D Scenes with Geometric Steering" + }, + { + "type": "image", + "bbox": [ + 0.18, + 0.179, + 0.818, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.259, + 0.281, + 0.347, + 0.295 + ], + "angle": 0, + "content": "Text-to-4D" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.3, + 0.489, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.139, + 0.4, + 0.448, + 0.422 + ], + "angle": 0, + "content": "A carpet covers the living room floor, and a puppy is lying on it playing with a kitten. To the right of the carpet, there is a toy ball." + }, + { + "type": "image_caption", + "bbox": [ + 0.454, + 0.415, + 0.543, + 0.464 + ], + "angle": 0, + "content": "SteerX" + }, + { + "type": "image_caption", + "bbox": [ + 0.658, + 0.282, + 0.745, + 0.296 + ], + "angle": 0, + "content": "Text-to-3D" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.3, + 0.892, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.544, + 0.399, + 0.724, + 0.411 + ], + "angle": 0, + "content": "A pink bicycle leaning against a fence" + }, + { + "type": "image_footnote", + "bbox": [ + 0.783, + 0.398, + 0.871, + 0.409 + ], + "angle": 0, + "content": "A pair of sneakers" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.438, + 0.499, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.649, + 0.439, + 0.75, + 0.455 + ], + "angle": 0, + "content": "Image-to-3D" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.457, + 0.899, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.59, + 0.907, + 0.619 + ], + "angle": 0, + "content": "Figure 1. SteerX is a zero-shot inference-time steering approach that seamlessly integrates video generative models [24, 35, 59, 62, 80] and feed-forward scene reconstruction models [24, 61, 86], enabling any 3D and 4D scene generation without explicit camera conditions." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.635, + 0.327, + 0.65 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.668, + 0.484, + 0.864 + ], + "angle": 0, + "content": "Recent progress in 3D/4D scene generation emphasizes the importance of physical alignment throughout video generation and scene reconstruction. However, existing methods improve the alignment separately at each stage, making it difficult to manage subtle misalignments arising from another stage. Here, we present SteerX, a zero-shot inference-time steering method that unifies scene reconstruction into the generation process, tilting data distributions toward better geometric alignment. To this end, we introduce two geometric reward functions for 3D/4D scene generation by using pose-free feed-forward scene reconstruction models. Through extensive experiments, we demonstrate the effectiveness of SteerX in improving 3D/4D scene generation." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.635, + 0.645, + 0.65 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.661, + 0.907, + 0.827 + ], + "angle": 0, + "content": "Generating 3D and 4D scenes from images or text prompts has attracted significant attention due to its potential applications in AR/VR and robotics [10, 63, 76]. This progress is largely driven by the advancement of generative models [7, 25, 35, 62, 80] and neural scene representations [33, 46, 73]. Generative models learn the underlying distribution of large-scale and high-quality video data, leveraging their scalability without imposing explicit physical constraints. In contrast, neural scene representations lift these distributions into structured 3D or 4D spaces, enforcing physical consistency and enabling more faithful scene modeling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.827, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this context, recent efforts [2, 3, 23, 27, 59, 71, 75, 87] have focused on producing geometrically consistent images by fine-tuning generative models with camera pose parameters, where the generation process follows user-defined or pre-defined camera trajectories. This facilitates seamless 3D" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.888, + 0.345, + 0.902 + ], + "angle": 0, + "content": "*Equal contribution, †Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "27326" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "and 4D scene reconstructions but requires complex optimization to regress neural scene representations, increasing computational overhead and hindering practical adaptation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.137, + 0.482, + 0.318 + ], + "angle": 0, + "content": "To address this inefficiency, another line of work [24, 37] has introduced text-conditioned camera pose generation and feed-forward decoding of pixel-aligned 3DGS [33]. This approach trains 3DGS decoders to learn a mapping function that directly reconstructs 3D scenes from multi-view images, leveraging multiple 3D scene datasets for improved 3D scene modeling. However, the limited datasets used to train pose-conditioned decoders fail to capture the various camera trajectories produced by video generative models. This results in slight misalignments between video generation and scene reconstruction, requiring further refinement steps [53, 79, 89] to enhance geometric consistency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.319, + 0.482, + 0.424 + ], + "angle": 0, + "content": "To sum up, previous works handle geometric alignment separately in either video generation or scene reconstruction. This makes it difficult to address cross-stage misalignments, as inconsistencies in one stage may not be fully corrected in the other. Despite recent efforts to mitigate this issue, achieving precise alignment remains an ongoing challenge due to the indistinct link between the two stages." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.425, + 0.483, + 0.59 + ], + "angle": 0, + "content": "On the other hand, zero-shot guidance methods that alter the sampling trajectory of generative models to enforce physical constraints have been widely explored in recent literature [17, 64]. These methods have shown great generalizability and performance, contingent upon a well-defined reward. For instance, in inverse imaging problems, a closed-form likelihood function can guide prior sampling processes toward posterior sampling [13, 57]. However, directly applying zero-shot guidance to 3D/4D scene generation remains challenging due to the ambiguity in defining reward functions and the significant computational overhead." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.591, + 0.483, + 0.863 + ], + "angle": 0, + "content": "In this work, we address this gap by introducing SteerX, a zero-shot inference-time steering method that seamlessly integrates video generation and scene reconstruction, generating geometrically aligned high-quality 3D and 4D scenes. To achieve this, we define geometric reward functions to assess physical consistency across video frames, drawing inspiration from cycle consistency [90] to guide the generation process toward high-reward outputs. We propose two geometric reward functions tailored for 3D and 4D scene generation. To evaluate geometric consistency across multiple video frames, we extend MEt3R [1], a recent evaluation metric for image pairs, by incorporating advanced pose-free feed-forward scene reconstruction methods such as MV-DUSt3R+ [61] and MonST3R [86]. These reconstruction methods lift intermediate generated video frames' during the reverse sampling process into 3D and 4D spaces. The reconstructed scenes are then projected back into the original image space for consistency evaluation." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.864, + 0.482, + 0.877 + ], + "angle": 0, + "content": "However, even when the reward is defined, one cannot" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.258 + ], + "angle": 0, + "content": "directly adopt standard gradient guidance approaches [13, 14, 57] for several reasons. For one, gradient guidance can be used only when the reward is fully differentiable, which hampers flexibility in the design of reward functions. Moreover, there exist substantial memory constraints to compute gradients for dozens of video frames, limiting their scalability for long video sequences. Therefore, we propose a steering algorithm based on sequential Monte Carlo (SMC) [20], which has gained recent traction [19, 36, 55, 74] due to its applicability to non-differentiable reward functions and favorable inference-time scaling properties." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.263, + 0.905, + 0.475 + ], + "angle": 0, + "content": "By designing tailored reward functions for geometrically plausible and accurate 3D/4D scene generation, along with a guided sampling process based on SMC, SteerX serves as a fully general framework that integrates any generative video models with any 3D reconstruction models, enabling diverse tasks, including Image-to-3D, Image-to-4D, Text-to-3D, and Text-to-4D generation. Through extensive experiments on both 3D and 4D scene generation with various pre-trained video generative models [24, 35, 59, 62, 80], we demonstrate the effectiveness and broad applicability of our approach. Furthermore, we show that by increasing the number of particles, we see favorable scaling properties, opening up new possibilities that have been rather untouched in 3D/4D scene generation: test-time scaling." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.505, + 0.663, + 0.521 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.536, + 0.802, + 0.551 + ], + "angle": 0, + "content": "2.1. 3D and 4D Scene Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.563, + 0.905, + 0.7 + ], + "angle": 0, + "content": "Various neural scene representations have been explored for 3D and 4D scene reconstruction. Most recent works tend to use Neural Radiance Fields (NeRF) [46] and 3D Gaussian Splats (3DGS) [33] for its high-quality rendering of underlying static scenes [4, 5, 85]. These scene representations have been extended to reconstruct 4D scenes by incorporating additional supervisions (e.g., depth maps [78] and segmentation masks [34]) along with deformable fields [38, 47, 50, 51, 67] and Gaussian splats [65, 73]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Recent advancements in 3D and 4D scene datasets [16, 40, 41, 54, 82, 84, 88] have enabled the development of generalizable scene representations, allowing feed-forward scene reconstruction methods [9, 12, 15, 60, 69]. Further advancements have extended these techniques to pose-free feed-forward 3D reconstruction models [56, 61, 68, 81, 86], eliminating the need for camera pose conditions and enabling the scene reconstruction from hundreds of unposed images, making the process more flexible and scalable. We leverage these powerful pose-free 3D reconstruction models, specifically MV-DUSt3R+ [61] and MonST3R [86], to assess geometric consistency and lift generated video frames into 3D and 4D spaces." + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.887, + 0.296, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\)Denoised predictions, as used in [13]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27327" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.35, + 0.106 + ], + "angle": 0, + "content": "2.2. 3D and 4D Scene Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.295 + ], + "angle": 0, + "content": "Recent 3D and 4D scene generation approaches commonly adopt a two-stage framework, where video generation is followed by scene reconstruction, with a focus on enhancing geometric consistency across both stages. Some works fine-tune video generative models with camera pose parameters [2, 23, 27, 42, 71, 75, 83, 87], where the generation process is conditioned on a user-defined or pre-defined camera trajectory to produce geometrically aligned video frames. These camera-conditioned video frames are used as training data to regress neural scene representations. While they produce plausible scenes, optimizing scene representations from scratch incurs high computational costs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.295, + 0.484, + 0.492 + ], + "angle": 0, + "content": "In contrast, camera-free approaches [24, 37] reconstruct 3D scenes without camera conditions, improving geometric alignment in the scene reconstruction stage. They generate text-conditioned camera poses and use a feed-forward 3DGS decoder trained on diverse 3D scene datasets to estimate underlying 3D structures from posed multi-view images. However, these methods often fail to fully capture the diverse camera trajectories produced by video generative models. Instead of addressing alignment separately at each stage, SteerX integrates scene reconstruction models directly into the video generation process, ensuring that video frames are optimally structured for accurate scene reconstruction and enabling seamless integration of both stages." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.498, + 0.265, + 0.515 + ], + "angle": 0, + "content": "2.3. Guided Sampling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.521, + 0.483, + 0.596 + ], + "angle": 0, + "content": "Generative models are trained to represent data distributions across various domains, including images [30, 48, 49, 52], videos [26, 31], and 3D [43, 72]. However, sampling from tilted distributions (e.g., posterior distributions from measurements) is often preferable to drawing from naive prior." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.596, + 0.484, + 0.687 + ], + "angle": 0, + "content": "Recent works [6, 77] have encoded user preferences through reward functions and fine-tuned generative models to maximize these rewards, effectively sampling from a tilted distribution toward high-reward outputs. However, fine-tuning approaches require extensive training and hampers the generalizability by modifying the base distribution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.484, + 0.868 + ], + "angle": 0, + "content": "In contrast, guided sampling methods [13, 17, 57, 58, 64] allow sampling from a tilted distribution in a zero-shot manner while keeping the base distribution intact. Among various paradigms, two classes stand out: gradient-based guidance [13, 57] and particle filtering[19, 36, 74]. Gradient-based guidance is simple to implement and widely applied but requires differentiable reward functions and is memory-intensive. In contrast, particle filtering is amenable to non-differentiable rewards, and is more memory-efficient, as it does not rely on backpropagation. Recently, Feynman-Kac Steering (FKS) [55] has unified previous particle filtering techniques, introducing a general steering framework that" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "can be applied to any reward function. SteerX builds upon FKS, integrating geometric reward functions and exploring a tailored design space for geometric steering, making it applicable to any video generative models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.166, + 0.845, + 0.185 + ], + "angle": 0, + "content": "3. Preliminary: Feynman-Kac Steering" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.192, + 0.906, + 0.253 + ], + "angle": 0, + "content": "Diffusion models [18, 58] are learned to reverse a forward process \\(\\mathbf{x}_t \\sim q(\\mathbf{x}_t | \\mathbf{x}_0 = \\mathbf{x})\\) with \\(T\\) timesteps, that gradually transforms a data sample \\(\\mathbf{x}\\) into Gaussian noise. This reverse process \\(p_{\\theta}(\\mathbf{x}_{t-1} | \\mathbf{x}_t)\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.264, + 0.906, + 0.306 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(\\mathbf {x} _ {T}, \\mathbf {x} _ {T - 1}, \\dots , \\mathbf {x} _ {1}, \\mathbf {x} _ {0}\\right) \\propto \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.317, + 0.906, + 0.363 + ], + "angle": 0, + "content": "The main goal of inference-time steering methods is to sample tilted distributions toward maximizing an exponential of user-specified reward function \\( r_{\\phi}(\\mathbf{x}_0) \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.374, + 0.906, + 0.391 + ], + "angle": 0, + "content": "\\[\n\\tilde {p} _ {\\theta} \\left(\\mathbf {x} _ {0}\\right) \\propto p _ {\\theta} \\left(\\mathbf {x} _ {0}\\right) \\exp \\left(\\lambda r _ {\\phi} \\left(\\mathbf {x} _ {0}\\right)\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.906, + 0.48 + ], + "angle": 0, + "content": "Feynman-Kac Steering (FKS) [55] introduces a system of multiple interacting diffusion paths (i.e. particles), where each path \\(\\mathbf{x}_{T:0} = (\\mathbf{x}_T,\\dots ,\\mathbf{x}_0)\\) is resampled at intermediate steps based on potentials \\(G_{t}(\\mathbf{x}_{T:t})\\), resulting in a sequence of tilted distributions \\(\\tilde{p}_{\\theta}(\\mathbf{x}_{T:t})\\):" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.491, + 0.905, + 0.532 + ], + "angle": 0, + "content": "\\[\n\\tilde {p} _ {\\theta} \\left(\\mathbf {x} _ {T: t}\\right) \\propto \\prod_ {s = T} ^ {t} p _ {\\theta} \\left(\\mathbf {x} _ {T: s}\\right) G _ {s} \\left(\\mathbf {x} _ {T: s}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.542, + 0.907, + 0.618 + ], + "angle": 0, + "content": "where the product of potentials matches the total tilt of the base distribution as \\(\\prod_{t=T}^{0} G_{t}(\\mathbf{x}_{T:t}) = \\exp(\\lambda r_{\\phi}(\\mathbf{x}_{0}))\\). These potentials are generally defined as SMC [74] to score the particles at each transition step \\(\\mathbf{x}_{t+1} \\rightarrow \\mathbf{x}_{t}\\) based on intermediate rewards \\(r_{\\phi}(\\mathbf{x}_{t})\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.632, + 0.613, + 0.646 + ], + "angle": 0, + "content": "4. Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.906, + 0.794 + ], + "angle": 0, + "content": "In this section, we introduce SteerX, a zero-shot inference-time steering method for 3D/4D scene generation. SteerX unifies pose-free feed-forward scene reconstruction models into the video generation process, iteratively tilting the data distribution towards geometrically aligned samples. In Section 4.1, we define two geometric reward functions to evaluate geometric consistency in generated multi-view images and dynamic videos, respectively. In Section 4.2, we detail our SMC-based steering algorithm." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.704, + 0.818 + ], + "angle": 0, + "content": "4.1. Geometric Rewards" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.9 + ], + "angle": 0, + "content": "Our geometric rewards build upon MEt3R [1], which utilizes DUSt3R [68] to measure feature similarity in overlapping regions between image pairs. It extracts upscaled DINO features [8, 22] and computes the cosine similarity between reference view features and the rendered features" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.9 + ], + "angle": 0, + "content": "2 Throughout the paper, we flexibly use the term particle filtering and SMC to refer to this class." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27328" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.898, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.354, + 0.908, + 0.412 + ], + "angle": 0, + "content": "Figure 2. An overview of geometric rewards. Our reward functions assess the geometric consistency of intermediate generated video frames by computing the feature similarity of upscaled DINO features. (a) GS-MEt3R evaluates feature similarity between the original video frames and their corresponding rendered images from 3DGS. (b) Dyn-MEt3R focuses on background regions by unprojecting background features from half of the video frames and reprojecting them onto the remaining frames to compute feature similarity." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.422, + 0.482, + 0.497 + ], + "angle": 0, + "content": "from another viewpoint. It can evaluate multiple images by computing feature similarity across all possible image pairs and averaging the scores. However, as the number of images increases, this becomes computationally expensive, making the overall generation process inefficient." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.498, + 0.483, + 0.651 + ], + "angle": 0, + "content": "To address this, as shown in Fig. 2, we introduce two geometric reward functions, GS-MEt3R and Dyn-MEt3R, which measure geometric consistency across video frames tailored for 3D/4D scene generation, respectively. We employ pose-free feed-forward scene reconstruction methods, MV-DUSt3R+ [61] and MonST3R [86], to reconstruct 3D and 4D scenes, which are subsequently projected back into image space to assess consistency with the original frames. This mitigates computational bottlenecks while efficiently evaluating global consistency in the generated video frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.655, + 0.483, + 0.735 + ], + "angle": 0, + "content": "3D Scene Reward. We extend MEt3R [1] in distinct ways to better suit 3D and 4D scene reconstruction models. For 3D scenes, recent methods [61, 81] introduce a mapping function \\( f_{\\phi} \\) that directly reconstructs 3DGS and camera poses from \\( N \\) sparse views \\( \\{I_i \\in \\mathbb{R}^{H \\times W \\times 3}\\}_{i=1}^N \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.742, + 0.483, + 0.786 + ], + "angle": 0, + "content": "\\[\nf _ {\\phi}: \\left\\{I _ {i} \\right\\} _ {i = 1} ^ {N} \\rightarrow \\left\\{\\begin{array}{l}\\left\\{\\boldsymbol {\\mu} _ {j}, \\mathbf {o} _ {j}, \\boldsymbol {\\Sigma} _ {j}, \\mathbf {c} _ {j} \\right\\} _ {j = 1} ^ {N \\times H \\times W}\\\\\\left\\{P _ {i} \\right\\} _ {i = 1} ^ {N}\\end{array}, \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "where the 3D scene is represented as Gaussian parameters, including position \\(\\pmb{\\mu}\\), volume density \\(\\mathbf{o}\\), covariance \\(\\pmb{\\Sigma}\\), and color \\(\\mathbf{c}\\). Then, we produce images \\(\\{\\tilde{I}_i\\}_{i=1}^N\\) by rendering the scene with estimated camera poses \\(\\{P_i\\}_{i=1}^N\\), ensuring they correspond to the same viewpoints as the input images. Finally, GS-MEt3R is measured by computing the cosine similarity between upscaled DINO [8, 22] features of input im" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.42, + 0.825, + 0.44 + ], + "angle": 0, + "content": "ages \\(\\{F_i\\}_{i = 1}^N\\) and rendered images \\(\\{\\hat{F}_i\\}_{i = 1}^N\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.447, + 0.907, + 0.49 + ], + "angle": 0, + "content": "\\[\nr _ {\\phi} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {H} \\sum_ {k = 1} ^ {W} \\frac {F _ {i} ^ {j k} \\cdot \\hat {F} _ {i} ^ {j k}}{| | F _ {i} ^ {j k} | | | | \\hat {F} _ {i} ^ {j k} | |}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.498, + 0.907, + 0.59 + ], + "angle": 0, + "content": "This approach not only offers a more direct evaluation of the physical consistency of the 3D scene compared to averaging consistency across all image pair combinations but also indirectly assesses the rendering quality of 3DGS. In other words, high GS-MEt3R scores indicate both geometric alignment and visually realistic 3D scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.593, + 0.906, + 0.701 + ], + "angle": 0, + "content": "4D Scene Reward. While the 3D scene reward function is based on 3DGS, feed-forward dynamic scene reconstruction with Gaussians remains underexplored, making it difficult to directly apply 3DGS-based rewards. Instead, we employ 3D point cloud representations, where MonST3R [86] reconstructs it with point maps \\(\\{X_{i}\\}_{i = 1}^{N}\\), binary dynamic masks \\(\\{M_i\\}_{i = 1}^N\\), and camera poses \\(\\{P_i\\}_{i = 1}^N\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.707, + 0.905, + 0.727 + ], + "angle": 0, + "content": "\\[\nf _ {\\phi}: \\left\\{I _ {i} \\right\\} _ {i = 1} ^ {N} \\rightarrow \\left\\{X _ {i}, M _ {i}, P _ {i} \\right\\} _ {i = 1} ^ {N}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.905, + 0.78 + ], + "angle": 0, + "content": "where we leverage these time-varying point clouds as 4D scene representations and design a reward function for evaluating geometric consistency in dynamic videos." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Since dynamic masks are produced in the camera pose estimation process to retain only high-confidence points, a well-reconstructed 4D scene should effectively filter out dynamic objects while preserving geometric consistency in the background regions. Therefore, we evaluate the consistency only for background regions of video frames, which are not filtered out by the dynamic mask. To this end, we first split \\(N\\) video frames into two subsets: \\(\\mathcal{I}_{src} = \\{I_1, I_3, \\ldots, I_N\\}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27329" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.094, + 0.326, + 0.109 + ], + "angle": 0, + "content": "Algorithm 1 SteerX (v-prediction)" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.111, + 0.483, + 0.158 + ], + "angle": 0, + "content": "Required: v-parametrized diffusion model \\(\\mathbf{v}_{\\theta}\\), reward function \\(r_{\\phi}\\), number of particles \\(k\\), and initial noise \\(\\{\\mathbf{x}_T^i\\}_{i=1}^k \\sim \\mathcal{N}(0, I)\\)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.157, + 0.168, + 0.171 + ], + "angle": 0, + "content": "Sampling:" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.174, + 0.297, + 0.189 + ], + "angle": 0, + "content": "1: for \\( t \\in \\{T - 1, \\dots, 0\\} \\) do" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.193, + 0.278, + 0.208 + ], + "angle": 0, + "content": "2: for \\(i\\in \\{1\\dots k\\}\\) do" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.211, + 0.449, + 0.227 + ], + "angle": 0, + "content": "3: \\(\\hat{\\mathbf{x}}_0^i\\gets \\sqrt{\\bar{\\alpha}_{t + 1}}\\mathbf{x}_{t + 1}^i -\\sqrt{1 - \\bar{\\alpha}_{t + 1}}\\mathbf{v}_\\theta (\\mathbf{x}_{t + 1}^i)\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.23, + 0.356, + 0.246 + ], + "angle": 0, + "content": "4: \\(\\mathbf{x}_t^i\\gets dpm\\text{- solver} (\\hat{\\mathbf{x}}_0^i,\\mathbf{x}_{t + 1}^i)\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.248, + 0.478, + 0.265 + ], + "angle": 0, + "content": "5: \\(\\mathbf{s}_t^i\\gets r_\\phi (\\hat{\\mathbf{x}}_0^i)\\) Intermediaterewards" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.267, + 0.479, + 0.284 + ], + "angle": 0, + "content": "6: \\(G_{t}^{i}\\gets \\exp (\\lambda \\max_{j = t}^{T}(\\mathbf{s}_{j}^{i}))\\) \\(\\triangleright\\) Potential" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.287, + 0.198, + 0.299 + ], + "angle": 0, + "content": "7: end for" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.303, + 0.478, + 0.321 + ], + "angle": 0, + "content": "8: \\(\\{\\mathbf{x}_t^i\\}_{i=1}^k \\sim \\text{Multinomial}(\\{\\mathbf{x}_t^i, G_t^i\\}_{i=1}^k) \\triangleright \\text{Resample}\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.325, + 0.174, + 0.337 + ], + "angle": 0, + "content": "9: end for" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.342, + 0.318, + 0.359 + ], + "angle": 0, + "content": "10: \\(l\\gets \\arg \\max_{i\\in \\{1,\\ldots ,k\\}}r_{\\phi}(\\mathbf{x}_0^i)\\)" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.36, + 0.19, + 0.375 + ], + "angle": 0, + "content": "11: return \\(\\mathbf{x}_0^l\\)" + }, + { + "type": "list", + "bbox": [ + 0.098, + 0.174, + 0.479, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.397, + 0.484, + 0.491 + ], + "angle": 0, + "content": "and \\(\\mathcal{I}_{tgt} = \\{I_2, I_4, \\ldots, I_{N-1}\\}\\). Then, we unproject the upsampled DINO features of background regions in \\(\\mathcal{I}_{src}\\) into 3D space using MonST3R. Finally, we reproject them onto the viewpoint of \\(\\mathcal{I}_{tgt}\\), where the rendered features \\(\\hat{\\mathcal{F}}_{tgt} = \\{\\hat{F}_1, \\hat{F}_3, \\ldots, \\hat{F}_N\\}\\) are used to compute the feature similarity with background regions in \\(\\mathcal{I}_{tgt}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.505, + 0.483, + 0.548 + ], + "angle": 0, + "content": "\\[\nr _ {i} = \\sum_ {j = 1} ^ {H} \\sum_ {k = 1} ^ {W} \\left(1 - M _ {i} ^ {j k}\\right) \\frac {F _ {i} ^ {j k} \\cdot \\hat {F} _ {i} ^ {j k}}{| | F _ {i} ^ {j k} | | | | \\hat {F} _ {i} ^ {j k} | |}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.57, + 0.483, + 0.613 + ], + "angle": 0, + "content": "\\[\nr _ {\\phi} = \\frac {1}{(N / / 2)} \\sum_ {i = 1} ^ {N / / 2} r _ {i}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.626, + 0.484, + 0.703 + ], + "angle": 0, + "content": "With these geometric rewards, we can effectively steer pre-trained video generative models to produce geometrically consistent video frames, which are then directly reconstructed into 3D and 4D spaces using the feed-forward 3D reconstruction models employed in the geometric rewards." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.717, + 0.279, + 0.733 + ], + "angle": 0, + "content": "4.2. Geometric Steering" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.741, + 0.483, + 0.787 + ], + "angle": 0, + "content": "Using the rewards defined in Section 4.1, let \\(\\tilde{p}_{\\theta}\\) in (2) be the distribution we wish to sample from. SMC operates with the three following steps:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.788, + 0.483, + 0.82 + ], + "angle": 0, + "content": "1. (Proposal) For each particle \\(i\\), sample from the proposal distribution \\(\\mathbf{x}_t^i \\sim q_t(\\mathbf{x}_t|\\mathbf{x}_{t+1}^i)\\)" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.826, + 0.483, + 0.866 + ], + "angle": 0, + "content": "2. (Weighting) Compute weights from reward-based potentials \\(\\omega_{t}^{i} = \\frac{p_{\\theta}(\\mathbf{x}_{t}^{i}|\\mathbf{x}_{t + 1}^{i})}{q_{t}(\\mathbf{x}_{t}^{i}|\\mathbf{x}_{t + 1}^{i})} G_{t}(\\mathbf{x}_{T:t}^{i})\\)" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "3. (Resampling) Draw new particles from the multinomial distribution \\(\\{\\mathbf{x}_t^i\\}_{i=1}^k \\sim \\text{Multinomial}(\\{\\mathbf{x}_t^i, G_t^i\\}_{i=1}^k)\\)" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.788, + 0.483, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.123 + ], + "angle": 0, + "content": "Two choices should be made: the potential \\( G_{t} \\), and the proposal distribution \\( q_{t} \\). For the potential, we use max potential" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.128, + 0.907, + 0.161 + ], + "angle": 0, + "content": "\\[\nG _ {t} \\left(\\mathbf {x} _ {T: t}\\right) ^ {i} := \\exp \\left(\\lambda \\max _ {j = t} ^ {T} \\left[ r _ {\\phi} \\left(\\hat {\\mathbf {x}} _ {0}\\right) \\right]\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.165, + 0.549, + 0.179 + ], + "angle": 0, + "content": "with" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.183, + 0.906, + 0.228 + ], + "angle": 0, + "content": "\\[\nG _ {0} \\left(\\mathbf {x} _ {T: 0}\\right) := \\exp \\left(\\lambda r _ {\\phi} (\\mathbf {x} _ {0})\\right) \\left(\\prod_ {t = 1} ^ {T} G _ {t} \\left(\\mathbf {x} _ {T: t}\\right)\\right) ^ {- 1}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.232, + 0.907, + 0.338 + ], + "angle": 0, + "content": "such that the particle with the highest reward is preferred. Notice that we use the Tweedie estimate \\(\\hat{\\mathbf{x}}_0 = \\mathbb{E}[\\mathbf{x}_0|\\mathbf{x}_t]\\) [13, 21, 55] in intermediate steps to avoid full reverse sampling. For the proposal kernel, to save computation, we leverage DPM-solver++ [44], which approximates the true sampling trajectory limited to small neural function evaluation (NFE). These choices lead to SteerX, as shown in Alg. 1." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.344, + 0.906, + 0.374 + ], + "angle": 0, + "content": "Proposition 1. Given the reverse generative process in (1), let \\( q_{t} \\) be the transition kernel satisfying" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.38, + 0.906, + 0.413 + ], + "angle": 0, + "content": "\\[\n\\frac {p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right)}{q _ {t} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right)} = 1 + \\epsilon_ {t} \\left(\\mathbf {x} _ {T: t}\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.417, + 0.907, + 0.478 + ], + "angle": 0, + "content": "with \\(|\\epsilon_t(\\mathbf{x}_{T:t})| \\leq \\varepsilon\\) uniformly. Also assume that the error from the reward computed at the approximate state \\(\\hat{\\mathbf{x}}_0\\) is bounded, i.e. \\(|r_{\\phi}(\\hat{\\mathbf{x}}_0) - r_{\\phi}(\\mathbf{x}_0)| \\leq \\eta\\). Then, given the defined max potentials in (9), (10), Alg. 1 samples from" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.485, + 0.906, + 0.502 + ], + "angle": 0, + "content": "\\[\n\\tilde {p} _ {\\theta} \\left(\\mathbf {x} _ {0}\\right) \\propto p _ {\\theta} \\left(\\mathbf {x} _ {0}\\right) \\exp \\left(\\lambda r _ {\\phi} \\left(\\mathbf {x} _ {0}\\right)\\right) \\left(1 + \\mathcal {O} \\left(T \\varepsilon + \\lambda \\eta\\right)\\right) \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.906, + 0.552 + ], + "angle": 0, + "content": "See Appendix A for the proof. Prop. 1 states that when the approximation errors are sufficiently small, then we can sample from the desired tilted distribution with SteerX." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.564, + 0.72, + 0.581 + ], + "angle": 0, + "content": "5. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.59, + 0.907, + 0.696 + ], + "angle": 0, + "content": "In this section, we conduct extensive experiments to verify the scalability and effectiveness of SteerX across various video generative models in four scene generation scenarios: Text-to-4D, Image-to-4D, Text-to-3D, and Image-to-3D. We briefly explain our experimental setup in Section 5.1 and present both qualitative and quantitative results, as well as design choices of SteerX in Section 5.2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.702, + 0.704, + 0.719 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.724, + 0.907, + 0.906 + ], + "angle": 0, + "content": "Implementation Details. For video generative models, we generate 25 frames at a \\(480 \\times 480\\) resolution, except for models that require a fixed video length. Since reconstructing 3D scenes using all generated video frames is impractical, we use only eight frames for the reconstruction. We follow the default settings of MV-DUST3R+ [61] and MonST3R [86]. We set \\(\\lambda = 10\\) for the potential. To encourage exploration and reduce computation overhead, we adopt interval resampling, as in FKS [55], and apply the steering procedure at \\(M\\) selected timesteps out of \\(T\\) reverse diffusion steps, setting \\(M = 4\\) for 3D scene generation and \\(M = 2\\) for 4D scene generation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27330" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.087, + 0.904, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.382, + 0.907, + 0.412 + ], + "angle": 0, + "content": "Figure 3. Qualitative results of video generation in PenguinVideoBenchmark [35]. We visualize for Mochi [62] (top) and Hunyuan-Video [35] (bottom). SteerX achieves the best alignment with camera motion instructions while maintaining high video quality." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.422, + 0.486, + 0.681 + ], + "angle": 0, + "content": "Baselines and Benchmarks. We evaluate our SteerX with Mochi [62] and HunyuanVideo [35] for Text-to-4D generation, and CogVideoX [80] for Image-to-4D. For Image-to-3D, we employ S-Director with orbit-left from DimensionX [59]. For Text-to-3D, we utilize the video generation and scene reconstruction models proposed in SplatFlow [24]. We conduct evaluations on 98 samples from the PenguinVideoBenchmark [35] with camera descriptions for Text-to-4D generation, and 355 samples from VBench-I2V [32] for Image-to-3D and Image-to-4D generation. Additionally, we use 100 samples from the Single-Object-with-Surrounding set of T3Bench [28] for Text-to-3D evaluation. For all evaluations, we compare against the best-of-N approach, where \\( k \\) particles are generated independently, and the one with the highest reward is selected. Additionally, we include a baseline (\\( k = 1 \\)) that generates a video and reconstructs the scene without inference-time steering." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Evaluation Metrics. For 4D generation, we evaluated the Aesthetic Score and Subject Consistency in VBench [32] for video quality and semantic consistency, respectively. Also, we evaluated Temporal Consistency to measure text alignment in Text-to-4D generation, including adherence to camera motion instructions. We evaluated Dynamic Degree to measure the overall dynamicness in Image-to-4D generation. For 3D generation, we followed the evaluation protocols in Director3D [37] and SplatFlow [24], measuring the image quality and the CLIP score [29] for both generated multi-view images and rendered images of 3DGS. Finally, for all generation scenarios, we evaluated the geometric consistency of multi-view images and dynamic videos using GS-MEt3R and Dyn-MEt3R, respectively." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.42, + 0.905, + 0.582 + ], + "angle": 0, + "content": "
MethodkAesthetic↑Subject↑Temporal↑Dyn-MEt3R↑
Mochi [62]10.4910.9500.2430.884
+ BoN40.4980.9410.2440.912
+ SteerX (later)40.4880.9370.2420.910
+ SteerX (linear)40.4900.9490.2440.918
+ SteerX (early)40.5000.9550.2480.929
HunyuanVideo [35]10.5490.9670.2410.911
+ BoN40.5510.9780.2390.931
+ SteerX (later)40.5550.9760.2370.931
+ SteerX (linear)40.5560.9730.2410.943
+ SteerX (early)40.5550.9800.2430.964
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.584, + 0.907, + 0.628 + ], + "angle": 0, + "content": "Table 1. Quantitative results in PenguinVideoBenchmark [35]. SteerX with annealed resampling timesteps to the early sampling process significantly improves the overall video quality." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.645, + 0.655, + 0.66 + ], + "angle": 0, + "content": "5.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.671, + 0.907, + 0.808 + ], + "angle": 0, + "content": "Text-to-4D Generation. Figure 3 and Table 1 present results of text-conditioned video generation. Notably, SteerX significantly outperforms both the baseline and the best-of-N, confirming that our geometric steering effectively resamples particles to guide the data distribution toward geometrically aligned samples. Moreover, quantitative results show a strong correlation between Dyn-MEt3R and other metrics, suggesting that the geometric steering can contribute to both physical consistency and overall video quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We observe that annealed resampling timesteps, where the steering procedure at the early sampling stages outperforms linear resampling, whereas resampling at later stages tends to degrade performance. This observation aligns with previous findings [2, 59], where camera poses for video frames are largely determined early in the sampling process" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27331" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.177, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.153, + 0.178, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.089, + 0.258, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.153, + 0.258, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.089, + 0.338, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.153, + 0.338, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.339, + 0.089, + 0.419, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.339, + 0.153, + 0.419, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.42, + 0.089, + 0.5, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.42, + 0.153, + 0.5, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.089, + 0.58, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.153, + 0.605, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.219, + 0.577, + 0.231 + ], + "angle": 0, + "content": "+Best-of-N" + }, + { + "type": "image", + "bbox": [ + 0.581, + 0.089, + 0.659, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.581, + 0.153, + 0.659, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.219, + 0.659, + 0.231 + ], + "angle": 0, + "content": "" + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.089, + 0.74, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.661, + 0.153, + 0.74, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.661, + 0.219, + 0.74, + 0.231 + ], + "angle": 0, + "content": "+" + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.089, + 0.821, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.153, + 0.821, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.733, + 0.219, + 0.83, + 0.232 + ], + "angle": 0, + "content": "+ SteerX (early)" + }, + { + "type": "image", + "bbox": [ + 0.821, + 0.089, + 0.902, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.821, + 0.153, + 0.902, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.218, + 0.174, + 0.232 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.247, + 0.218, + 0.349, + 0.232 + ], + "angle": 0, + "content": "CogVideoX [80]" + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.236, + 0.87, + 0.251 + ], + "angle": 0, + "content": "Figure 4. Qualitative results of video generation in VBench-I2V [32]. SteerX enhances vividness and overall visual quality." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.256, + 0.221, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.359, + 0.191, + 0.372 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.256, + 0.346, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.256, + 0.451, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.316, + 0.359, + 0.383, + 0.372 + ], + "angle": 0, + "content": "CogVideoX" + }, + { + "type": "image", + "bbox": [ + 0.482, + 0.258, + 0.579, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.543, + 0.36, + 0.612, + 0.371 + ], + "angle": 0, + "content": "+Best-of-N" + }, + { + "type": "image", + "bbox": [ + 0.588, + 0.258, + 0.675, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.707, + 0.256, + 0.804, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.807, + 0.256, + 0.904, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.782, + 0.359, + 0.835, + 0.371 + ], + "angle": 0, + "content": "+SteerX" + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.376, + 0.904, + 0.391 + ], + "angle": 0, + "content": "Figure 5. Qualitative results in Image-to-4D. SteerX naturally lifts object motion into 4D spaces, while preserving geometric alignments." + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.403, + 0.468, + 0.422 + ], + "angle": 0, + "content": "\"Filmed from a first-person perspective, the camera passes through the graffiti alley in Melbourne, Australia, where the graffiti walls are covered with artwork from many artists.\"" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.425, + 0.212, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.552, + 0.192, + 0.564 + ], + "angle": 0, + "content": "HunyuanVideo" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.425, + 0.343, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.269, + 0.552, + 0.306, + 0.563 + ], + "angle": 0, + "content": "+BoN" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.425, + 0.472, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.402, + 0.552, + 0.448, + 0.563 + ], + "angle": 0, + "content": "+SteerX" + }, + { + "type": "image_caption", + "bbox": [ + 0.154, + 0.571, + 0.421, + 0.585 + ], + "angle": 0, + "content": "Figure 6. Qualitative results in Text-to-4D." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.659 + ], + "angle": 0, + "content": "as the diffusion model conceptualizes global structures. As a result, resampling at early timesteps coarsely aligns the geometry by tilting the data distribution, laying the foundation for a more precise geometric alignment in later stages." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.796 + ], + "angle": 0, + "content": "The geometrically aligned video frames contribute to better 4D scene reconstruction, as shown in Fig. 6. While the baseline struggles to capture accurate camera poses and the best-of-N approach results in blurry 4D scenes, our approach precisely estimates camera poses and generates visually realistic 4D scenes. This highlights the effectiveness of SteerX in the seamless integration of video generation and scene reconstruction, ensuring that generated video frames are optimally structured for precise scene reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Image-to-4D Generation. Figure 4 and Table 2 show the results in image-conditioned video generation, and our SteerX performs the best for all metrics, and we observe that SteerX generates dynamic and realistic video frames, whereas the baseline produces almost motionless frames, and the best-of-N approach results in blurry frames. Also," + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.4, + 0.903, + 0.498 + ], + "angle": 0, + "content": "
MethodkAesthetic↑Subject↑Dynamic↑Dyn-MEt3R↑
CogVideoX [80]10.5920.9450.1580.880
+ BoN20.5910.9410.1410.882
+ SteerX (early)20.5930.9450.1610.894
+ BoN40.5940.9440.1430.901
+ SteerX (early)40.5960.9570.1700.909
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.5, + 0.905, + 0.529 + ], + "angle": 0, + "content": "Table 2. Quantitative results in VBench-I2V [32]. Results demonstrate the scalability of geometric steering as \\( k \\) increases." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.538, + 0.903, + 0.621 + ], + "angle": 0, + "content": "
MethodkRenderingSample
BRISQUE↓NIQE↓CLIP-I↑GS-MEt3R↑
DimensionX [59]137.34.2582.40.708
+ BoN429.84.3383.20.745
+ SteerX (early)429.74.2483.70.749
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.623, + 0.905, + 0.652 + ], + "angle": 0, + "content": "Table 3. Quantitative results in VBench-I2V [32]. For rendered images, we report results after the refinement process in [24, 37]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.668, + 0.907, + 0.744 + ], + "angle": 0, + "content": "we validate the scalability of SteerX, as increasing the number of particles \\( k \\) consistently enhances performance across all metrics. As shown in Fig. 5, geometrically aligned video frames can be reconstructed to better 4D scenes, producing realistic dynamic objects and well-aligned backgrounds." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Image-to-3D Generation. We verify the effectiveness of SteerX in the Image-to-3D scene generation. While multiview images generated by SteerX closely resemble those from the best-of-N approach, SteerX reconstructs 3D scenes with finer details and fewer discontinuities in novel views, as shown in Fig. 8. Furthermore, quantitative results in Table 3 demonstrate that SteerX achieves better visual quality and geometric consistency in both rendered images and generated images compared to other baselines." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27332" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.895, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.336, + 0.905, + 0.364 + ], + "angle": 0, + "content": "Figure 7. Qualitative results of rendered images in T3Bench [28]. SteerX significantly enhances the visual quality of rendered images and textual alignment, demonstrating its compatibility with various video generation and scene reconstruction models." + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.374, + 0.21, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.543, + 0.188, + 0.553 + ], + "angle": 0, + "content": "(a) Input Image" + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.374, + 0.34, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.543, + 0.328, + 0.552 + ], + "angle": 0, + "content": "(b) DimensionX + BoN" + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.374, + 0.471, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.35, + 0.543, + 0.463, + 0.552 + ], + "angle": 0, + "content": "(c) DimensionX + SteerX" + }, + { + "type": "table_caption", + "bbox": [ + 0.098, + 0.558, + 0.476, + 0.572 + ], + "angle": 0, + "content": "Figure 8. Qualitative results in of 3DGS in VBench-I2V [32]." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.577, + 0.48, + 0.679 + ], + "angle": 0, + "content": "
MethodkBRISQUE↓CLIPScore↑GS-MEt3R↑
SplatFlow [24]133.2529.660.727
+ BoN236.8229.550.756
+ SteerX (early)227.4729.560.767
+ BoN428.1429.280.768
+ SteerX (early)426.6529.730.775
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.681, + 0.483, + 0.709 + ], + "angle": 0, + "content": "Table 4. Quantitative results of multi-view generation in Single-Object-with-Surrounding set of T3Bench [28]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Text-to-3D Generation. Notice that SteerX can be applied to any generative model and scene reconstruction model, including the multi-view rectified flow model and the feed-forward 3DGS decoder in SplatFlow [24], enabling geometric steering through GS-MEt3R rewards. Table 4 presents quantitative results of generated multi-view images, demonstrating that SteerX effectively enhances overall performance and further improves scalability as the number of particles increases. Furthermore, as shown in Table 5, these 3D-aligned multi-view images are integrated with existing refinement processes, achieving state-of-the-art performance in 3D scene generation. Figure 7 presents qualita" + }, + { + "type": "table", + "bbox": [ + 0.519, + 0.373, + 0.903, + 0.557 + ], + "angle": 0, + "content": "
MethodBRISQUE↓NIQE↓CLIPScore↑
DreamFusion [53]90.210.48-
Magic3D [39]92.811.20-
LatentNeRF [45]88.69.19-
SJC [66]82.010.15-
Fantasia3D [11]69.67.65-
ProlificDreamer [70]61.57.07-
Director3D [37]32.34.3532.9
SplatFlow [24]19.64.2433.2
SplatFlow [24]†23.44.8432.7
+ BoN (k=4)17.24.4132.3
+ SteerX (k=4)13.14.3033.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.559, + 0.905, + 0.603 + ], + "angle": 0, + "content": "Table 5. Quantitative results in T3Bench [28]. SplatFlow† is re-implemented results without the stop ray strategy, which is incompatible with our geometric steering pipeline." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.616, + 0.905, + 0.647 + ], + "angle": 0, + "content": "tive results, demonstrating that SteerX produces more text-aligned 3D scenes and visually realistic rendered images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.663, + 0.634, + 0.679 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we have introduced SteerX, a zero-shot inference-time steering method for camera-free 3D and 4D scene generation. Instead of addressing physical alignment separately in either video generation or scene reconstruction, SteerX unifies both stages and iteratively tilts the data distribution toward geometrically consistent samples. Extensive experiments across diverse scene generation tasks verify that SteerX effectively enhances visual quality, textual alignment, and geometric consistency. SteerX is practical, enabling zero-shot 3D and 4D scene generation, and is scalable, with geometric alignment improving as the number of particles increases. We believe this efficient scaling property holds great potential, opening new avenues for 3D and 4D scene generation, particularly in test-time scaling." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27333" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Mohammad Asim, Christopher Wewer, Thomas Wimmer, Bernt Schiele, and Jan Eric Lenssen. Met3r: Measuring multi-view consistency in generated images. arXiv preprint arXiv:2501.06336, 2025. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.243 + ], + "angle": 0, + "content": "[2] Sherwin Bahmani, Ivan Skorokhodov, Guocheng Qian, Aliaksandr Siarohin, Willi Menapace, Andrea Tagliasacchi, David B Lindell, and Sergey Tulyakov. Ac3d: Analyzing and improving 3d camera control in video diffusion transformers. arXiv preprint arXiv:2411.18673, 2024. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.484, + 0.327 + ], + "angle": 0, + "content": "[3] Sherwin Bahmani, Ivan Skorokhodov, Aliaksandr Siarohin, Willi Menapace, Guocheng Qian, Michael Vasilkovsky, Hsin-Ying Lee, Chaoyang Wang, Jiaxu Zou, Andrea Tagliasacchi, et al. Vd3d: Taming large video diffusion transformers for 3d camera control. arXiv preprint arXiv:2407.12781, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.484, + 0.411 + ], + "angle": 0, + "content": "[4] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5855–5864, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.415, + 0.484, + 0.484 + ], + "angle": 0, + "content": "[5] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.487, + 0.484, + 0.542 + ], + "angle": 0, + "content": "[6] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. In The Twelfth International Conference on Learning Representations, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.544, + 0.484, + 0.613 + ], + "angle": 0, + "content": "[7] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.616, + 0.484, + 0.684 + ], + "angle": 0, + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.687, + 0.484, + 0.757 + ], + "angle": 0, + "content": "[9] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.758, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[10] Guikun Chen and Wenguan Wang. A survey on 3d gaussian splatting. arXiv preprint arXiv:2401.03890, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.788, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[11] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22246–22256, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splattering from sparse" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.906, + 0.12 + ], + "angle": 0, + "content": "multi-view images. In European Conference on Computer Vision, pages 370-386. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.906, + 0.177 + ], + "angle": 0, + "content": "[13] Hyungjin Chung, Jeongsol Kim, Michael T McCann, Marc L Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. arXiv preprint arXiv:2209.14687, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[14] Hyungjin Chung, Suhyeon Lee, and Jong Chul Ye. Decomposed diffusion sampler for accelerating large-scale inverse problems. arXiv preprint arXiv:2303.05754, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.222, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[15] Wenyan Cong, Hanxue Liang, Peihao Wang, Zhiwen Fan, Tianlong Chen, Mukund Varma, Yi Wang, and Zhangyang Wang. Enhancing nerf akin to enhancing llms: Generalizable nerf transformer with mixture-of-view-experts. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3193-3204, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.307, + 0.905, + 0.376 + ], + "angle": 0, + "content": "[16] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.378, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[17] Giannis Daras, Hyungjin Chung, Chieh-Hsin Lai, Yuki Mitsufuji, Jong Chul Ye, Peyman Milanfar, Alexandros G Dimakis, and Mauricio Delbracio. A survey on diffusion models for inverse problems. arXiv preprint arXiv:2410.00083, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.448, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[18] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.491, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[19] Zehao Dou and Yang Song. Diffusion posterior sampling for linear inverse problem solving: A filtering perspective. In The Twelfth International Conference on Learning Representations, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.548, + 0.905, + 0.589 + ], + "angle": 0, + "content": "[20] Arnaud Doucet, Nando De Freitas, and Neil Gordon. An introduction to sequential monte carlo methods. Sequential Monte Carlo methods in practice, pages 3-14, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[21] Bradley Efron. Tweedie's formula and selection bias. Journal of the American Statistical Association, 106(496):1602-1614, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.634, + 0.905, + 0.689 + ], + "angle": 0, + "content": "[22] Stephanie Fu, Mark Hamilton, Laura Brandt, Axel Feldman, Zhoutong Zhang, and William T Freeman. Fitup: A model-agnostic framework for features at any resolution. arXiv preprint arXiv:2403.10516, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.691, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[23] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.762, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[24] Hyojun Go, Byeongjun Park, Jiho Jang, Jin-Young Kim, Soonwoo Kwon, and Changick Kim. Splatflow: Multiview rectified flow model for 3d gaussian splatting synthesis. arXiv preprint arXiv:2411.16443, 2024. 1, 2, 3, 6, 7, 8, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, Poriya Panet, Sapir Weissbuch, Victor Kulikov, Yaki Bitterman, Zeev Melumian, and Ofir Bibi. Ltx-video: Realtime video latent diffusion. arXiv preprint arXiv:2501.00103, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27334" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[26] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible diffusion modeling of long videos. Advances in Neural Information Processing Systems, 35:27953-27965, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[27] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrl: Enabling camera control for text-to-video generation. arXiv preprint arXiv:2404.02101, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.206, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[28] Yuze He, Yushi Bai, Matthieu Lin, Wang Zhao, Yubin Hu, Jenny Sheng, Ran Yi, Juanzi Li, and Yong-Jin Liu. T\\(^3\\)bench: Benchmarking current progress in text-to-3d generation, 2023. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[29] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.362 + ], + "angle": 0, + "content": "[30] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.364, + 0.482, + 0.418 + ], + "angle": 0, + "content": "[31] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.421, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[32] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[33] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.563, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[34] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[35] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. 1, 2, 6, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[36] Xiner Li, Yulai Zhao, Chenyu Wang, Gabriele Scalia, Gokcen Eraslan, Surag Nair, Tommaso Biancalani, Shuiwang Ji, Aviv Regev, Sergey Levine, et al. Derivative-free guidance in continuous and discrete diffusion models with soft value-based decoding. arXiv preprint arXiv:2408.08252, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[37] Xinyang Li, Zhangyu Lai, Linning Xu, Yansong Qu, Liujuan Cao, Shengchuan Zhang, Bo Dai, and Rongrong Ji. Director3d: Real-world camera trajectory and 3d scene generation from text. Advances in Neural Information Processing Systems, 37:75125-75151, 2025. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[38] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[39] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 300–309, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[40] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.369 + ], + "angle": 0, + "content": "[41] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14458-14467, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.371, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[42] Fangfu Liu, Wenqiang Sun, Hanyang Wang, Yikai Wang, Haowen Sun, Junliang Ye, Jun Zhang, and Yueqi Duan. Reconx: Reconstruct any scene from sparse views with video diffusion model. arXiv preprint arXiv:2408.16767, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[43] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[44] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[45] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.675 + ], + "angle": 0, + "content": "[46] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.678, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[47] Byeongjun Park and Changick Kim. Point-dynrf: Point-based dynamic radiance fields from a monocular video. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3171-3181, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[48] Byeongjun Park, Sangmin Woo, Hyojun Go, Jin-Young Kim, and Changick Kim. Denoising task routing for diffusion models. arXiv preprint arXiv:2310.07138, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[49] Byeongjun Park, Hyojun Go, Jin-Young Kim, Sangmin Woo, Seokil Ham, and Changick Kim. Switch diffusion transformer: Synergizing denoising tasks with sparse mixture-of-experts. arXiv preprint arXiv:2403.09176, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[50] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5865-5874, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27335" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[51] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[52] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195-4205, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.223, + 0.482, + 0.264 + ], + "angle": 0, + "content": "[53] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.267, + 0.482, + 0.349 + ], + "angle": 0, + "content": "[54] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10901-10911, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.353, + 0.482, + 0.421 + ], + "angle": 0, + "content": "[55] Raghav Singhal, Zachary Horvitz, Ryan Teehan, Mengye Ren, Zhou Yu, Kathleen McKeown, and Rajesh Ranganath. A general framework for inference-time scaling and steering of diffusion models. arXiv preprint arXiv:2501.06848, 2025. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.424, + 0.482, + 0.479 + ], + "angle": 0, + "content": "[56] Brandon Smart, Chuanxia Zheng, Iro Laina, and Victor Adrian Prisacariu. Splatt3r: Zero-shot gaussian splatting from uncalibrated image pairs. arXiv preprint arXiv:2408.13912, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.482, + 0.537 + ], + "angle": 0, + "content": "[57] Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. Pseudoinverse-guided diffusion models for inverse problems. In International Conference on Learning Representations, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.541, + 0.482, + 0.596 + ], + "angle": 0, + "content": "[58] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.599, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[59] Wenqiang Sun, Shuo Chen, Fangfu Liu, Zilong Chen, Yueqi Duan, Jun Zhang, and Yikai Wang. Dimensionx: Create any 3d and 4d scenes from a single image with controllable video diffusion. arXiv preprint arXiv:2411.04928, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.482, + 0.724 + ], + "angle": 0, + "content": "[60] Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10208-10217, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.729, + 0.482, + 0.797 + ], + "angle": 0, + "content": "[61] Zhenggang Tang, Yuchen Fan, Dilin Wang, Hongyu Xu, Rakesh Ranjan, Alexander Schwing, and Zhicheng Yan. Mvdust3r+: Single-stage scene reconstruction from sparse views in 2 seconds. arXiv preprint arXiv:2412.06974, 2024. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[62] Genmo Team. Mochi 1. https://github.com/genmoai/models, 2024. 1, 2, 6, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[63] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Wang Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. In Computer Graphics Forum, pages 703-735. Wiley Online Library, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[64] Masatoshi Uehara, Xingyu Su, Yulai Zhao, Xiner Li, Aviv Regev, Shuiwang Ji, Sergey Levine, and Tommaso Biancalani. Reward-guided iterative refinement in diffusion models at test-time with applications to protein and dna design. arXiv preprint arXiv:2502.14944, 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[65] Diwen Wan, Ruijie Lu, and Gang Zeng. Superpoint gaussian splatting for real-time high-fidelity dynamic scene reconstruction. arXiv preprint arXiv:2406.03697, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[66] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[67] Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.335, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[68] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[69] Yunsong Wang, Tianxin Huang, Hanlin Chen, and Gim Hee Lee. Freesplat: Generalizable 3d gaussian splatting towards free view synthesis of indoor scenes. Advances in Neural Information Processing Systems, 37:107326-107349, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[70] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[71] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.62, + 0.907, + 0.687 + ], + "angle": 0, + "content": "[72] Sangmin Woo, Byeongjun Park, Hyojun Go, Jin-Young Kim, and Changick Kim. Harmonyview: Harmonizing consistency and diversity in one-image-to-3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10574–10584, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.69, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[73] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20310-20320, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.905, + 0.842 + ], + "angle": 0, + "content": "[74] Luhuan Wu, Brian Trippe, Christian Naesseth, David Blei, and John P Cunningham. Practical and asymptotically exact conditional sampling in diffusion models. Advances in Neural Information Processing Systems, 36:31372-31403, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[75] Rundi Wu, Ruiqi Gao, Ben Poole, Alex Trevithick, Changxi Zheng, Jonathan T Barron, and Aleksander Holynski. Cat4d: Create anything in 4d with multi-view video diffusion models. arXiv preprint arXiv:2411.18613, 2024. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27336" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[76] Tong Wu, Yu-Jie Yuan, Ling-Xiao Zhang, Jie Yang, Yan-Pei Cao, Ling-Qi Yan, and Lin Gao. Recent advances in 3d gaussian splatting. Computational Visual Media, 10(4):613-642, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.231 + ], + "angle": 0, + "content": "[77] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: learning and evaluating human preferences for text-to-image generation. In Proceedings of the 37th International Conference on Neural Information Processing Systems, pages 15903-15935, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.483, + 0.302 + ], + "angle": 0, + "content": "[78] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.303, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[79] Xiaofeng Yang, Yiwen Chen, Cheng Chen, Chi Zhang, Yi Xu, Xulei Yang, Fayao Liu, and Guosheng Lin. Learn to optimize denoising scores for 3d generation: A unified and improved diffusion prior on nerf and 3d gaussian splatting. arXiv preprint arXiv:2312.04820, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.44 + ], + "angle": 0, + "content": "[80] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 2, 6, 7, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[81] Botao Ye, Sifei Liu, Haofei Xu, Xueting Li, Marc Pollefeys, Ming-Hsuan Yang, and Songyou Peng. No pose, no problem: Surprisingly simple 3d gaussian splats from sparse unposed images. arXiv preprint arXiv:2410.24207, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[82] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-22, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.554, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[83] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.705 + ], + "angle": 0, + "content": "[84] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Chenming Zhu, Zhangyang Xiong, Tianyou Liang, et al. Mvmgnet: A large-scale dataset of multi-view images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9150-9161, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[85] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19447-19456, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[86] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arXiv:2410.03825, 2024. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[87] Yuyang Zhao, Chung-Ching Lin, Kevin Lin, Zhiwen Yan, Linjie Li, Zhengyuan Yang, Jianfeng Wang, Gim Hee Lee, and Lijuan Wang. Genxd: Generating any 3d and 4d scenes. arXiv preprint arXiv:2411.02319, 2024. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.146 + ], + "angle": 0, + "content": "[88] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[89] Junzhe Zhu, Peiye Zhuang, and Sanmi Koyejo. Hifa: High-fidelity text-to-3d generation with advanced diffusion guidance. arXiv preprint arXiv:2305.18766, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[90] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "27337" + } + ] +] \ No newline at end of file diff --git a/2025/SteerX_ Creating Any Camera-Free 3D and 4D Scenes with Geometric Steering/a06f04c2-df21-4c58-83ac-45d3ea205c44_origin.pdf b/2025/SteerX_ Creating Any Camera-Free 3D and 4D Scenes with Geometric Steering/a06f04c2-df21-4c58-83ac-45d3ea205c44_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6f13094ea87fa7346352d282693330d6e5278137 --- /dev/null +++ b/2025/SteerX_ Creating Any Camera-Free 3D and 4D Scenes with Geometric Steering/a06f04c2-df21-4c58-83ac-45d3ea205c44_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2e8db71d77c6423b5ba26f8a11f192f27d2fdd92b042bec920b51c461f4fb88e +size 19029959 diff --git a/2025/SteerX_ Creating Any Camera-Free 3D and 4D Scenes with Geometric Steering/full.md b/2025/SteerX_ Creating Any Camera-Free 3D and 4D Scenes with Geometric Steering/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4c293036f4e43605f087b28a0bbe3d9c189584bd --- /dev/null +++ b/2025/SteerX_ Creating Any Camera-Free 3D and 4D Scenes with Geometric Steering/full.md @@ -0,0 +1,449 @@ +# SteerX: Creating Any Camera-Free 3D and 4D Scenes with Geometric Steering + +![](images/fac72fcde383d71a8cd37a08dbe546fb0ec531de0c396a6ad5b940595e7c277f.jpg) + +Text-to-4D +![](images/af83eea19b4ebc5216dbfba59eabd762a2347e84f80d75e241d21a4118168a25.jpg) +A carpet covers the living room floor, and a puppy is lying on it playing with a kitten. To the right of the carpet, there is a toy ball. + +Text-to-3D +![](images/709b9410f7dd3068975e60ebb2e7ee39d15211698082fdfc378200689da6d189.jpg) +A pink bicycle leaning against a fence +A pair of sneakers + +![](images/5a9ab954fb11ca1e13d26d72a5ae6712baf93bd1084f2ecdf131db87e3334cc4.jpg) +SteerX + +![](images/f5caf242268bd3e7d583e2c4b85817feba91542361dc4e8f4fa48e20abd960be.jpg) +Image-to-3D +Figure 1. SteerX is a zero-shot inference-time steering approach that seamlessly integrates video generative models [24, 35, 59, 62, 80] and feed-forward scene reconstruction models [24, 61, 86], enabling any 3D and 4D scene generation without explicit camera conditions. + +# Abstract + +Recent progress in 3D/4D scene generation emphasizes the importance of physical alignment throughout video generation and scene reconstruction. However, existing methods improve the alignment separately at each stage, making it difficult to manage subtle misalignments arising from another stage. Here, we present SteerX, a zero-shot inference-time steering method that unifies scene reconstruction into the generation process, tilting data distributions toward better geometric alignment. To this end, we introduce two geometric reward functions for 3D/4D scene generation by using pose-free feed-forward scene reconstruction models. Through extensive experiments, we demonstrate the effectiveness of SteerX in improving 3D/4D scene generation. + +# 1. Introduction + +Generating 3D and 4D scenes from images or text prompts has attracted significant attention due to its potential applications in AR/VR and robotics [10, 63, 76]. This progress is largely driven by the advancement of generative models [7, 25, 35, 62, 80] and neural scene representations [33, 46, 73]. Generative models learn the underlying distribution of large-scale and high-quality video data, leveraging their scalability without imposing explicit physical constraints. In contrast, neural scene representations lift these distributions into structured 3D or 4D spaces, enforcing physical consistency and enabling more faithful scene modeling. + +In this context, recent efforts [2, 3, 23, 27, 59, 71, 75, 87] have focused on producing geometrically consistent images by fine-tuning generative models with camera pose parameters, where the generation process follows user-defined or pre-defined camera trajectories. This facilitates seamless 3D + +and 4D scene reconstructions but requires complex optimization to regress neural scene representations, increasing computational overhead and hindering practical adaptation. + +To address this inefficiency, another line of work [24, 37] has introduced text-conditioned camera pose generation and feed-forward decoding of pixel-aligned 3DGS [33]. This approach trains 3DGS decoders to learn a mapping function that directly reconstructs 3D scenes from multi-view images, leveraging multiple 3D scene datasets for improved 3D scene modeling. However, the limited datasets used to train pose-conditioned decoders fail to capture the various camera trajectories produced by video generative models. This results in slight misalignments between video generation and scene reconstruction, requiring further refinement steps [53, 79, 89] to enhance geometric consistency. + +To sum up, previous works handle geometric alignment separately in either video generation or scene reconstruction. This makes it difficult to address cross-stage misalignments, as inconsistencies in one stage may not be fully corrected in the other. Despite recent efforts to mitigate this issue, achieving precise alignment remains an ongoing challenge due to the indistinct link between the two stages. + +On the other hand, zero-shot guidance methods that alter the sampling trajectory of generative models to enforce physical constraints have been widely explored in recent literature [17, 64]. These methods have shown great generalizability and performance, contingent upon a well-defined reward. For instance, in inverse imaging problems, a closed-form likelihood function can guide prior sampling processes toward posterior sampling [13, 57]. However, directly applying zero-shot guidance to 3D/4D scene generation remains challenging due to the ambiguity in defining reward functions and the significant computational overhead. + +In this work, we address this gap by introducing SteerX, a zero-shot inference-time steering method that seamlessly integrates video generation and scene reconstruction, generating geometrically aligned high-quality 3D and 4D scenes. To achieve this, we define geometric reward functions to assess physical consistency across video frames, drawing inspiration from cycle consistency [90] to guide the generation process toward high-reward outputs. We propose two geometric reward functions tailored for 3D and 4D scene generation. To evaluate geometric consistency across multiple video frames, we extend MEt3R [1], a recent evaluation metric for image pairs, by incorporating advanced pose-free feed-forward scene reconstruction methods such as MV-DUSt3R+ [61] and MonST3R [86]. These reconstruction methods lift intermediate generated video frames' during the reverse sampling process into 3D and 4D spaces. The reconstructed scenes are then projected back into the original image space for consistency evaluation. + +However, even when the reward is defined, one cannot + +directly adopt standard gradient guidance approaches [13, 14, 57] for several reasons. For one, gradient guidance can be used only when the reward is fully differentiable, which hampers flexibility in the design of reward functions. Moreover, there exist substantial memory constraints to compute gradients for dozens of video frames, limiting their scalability for long video sequences. Therefore, we propose a steering algorithm based on sequential Monte Carlo (SMC) [20], which has gained recent traction [19, 36, 55, 74] due to its applicability to non-differentiable reward functions and favorable inference-time scaling properties. + +By designing tailored reward functions for geometrically plausible and accurate 3D/4D scene generation, along with a guided sampling process based on SMC, SteerX serves as a fully general framework that integrates any generative video models with any 3D reconstruction models, enabling diverse tasks, including Image-to-3D, Image-to-4D, Text-to-3D, and Text-to-4D generation. Through extensive experiments on both 3D and 4D scene generation with various pre-trained video generative models [24, 35, 59, 62, 80], we demonstrate the effectiveness and broad applicability of our approach. Furthermore, we show that by increasing the number of particles, we see favorable scaling properties, opening up new possibilities that have been rather untouched in 3D/4D scene generation: test-time scaling. + +# 2. Related Works + +# 2.1. 3D and 4D Scene Reconstruction + +Various neural scene representations have been explored for 3D and 4D scene reconstruction. Most recent works tend to use Neural Radiance Fields (NeRF) [46] and 3D Gaussian Splats (3DGS) [33] for its high-quality rendering of underlying static scenes [4, 5, 85]. These scene representations have been extended to reconstruct 4D scenes by incorporating additional supervisions (e.g., depth maps [78] and segmentation masks [34]) along with deformable fields [38, 47, 50, 51, 67] and Gaussian splats [65, 73]. + +Recent advancements in 3D and 4D scene datasets [16, 40, 41, 54, 82, 84, 88] have enabled the development of generalizable scene representations, allowing feed-forward scene reconstruction methods [9, 12, 15, 60, 69]. Further advancements have extended these techniques to pose-free feed-forward 3D reconstruction models [56, 61, 68, 81, 86], eliminating the need for camera pose conditions and enabling the scene reconstruction from hundreds of unposed images, making the process more flexible and scalable. We leverage these powerful pose-free 3D reconstruction models, specifically MV-DUSt3R+ [61] and MonST3R [86], to assess geometric consistency and lift generated video frames into 3D and 4D spaces. + +# 2.2. 3D and 4D Scene Generation + +Recent 3D and 4D scene generation approaches commonly adopt a two-stage framework, where video generation is followed by scene reconstruction, with a focus on enhancing geometric consistency across both stages. Some works fine-tune video generative models with camera pose parameters [2, 23, 27, 42, 71, 75, 83, 87], where the generation process is conditioned on a user-defined or pre-defined camera trajectory to produce geometrically aligned video frames. These camera-conditioned video frames are used as training data to regress neural scene representations. While they produce plausible scenes, optimizing scene representations from scratch incurs high computational costs. + +In contrast, camera-free approaches [24, 37] reconstruct 3D scenes without camera conditions, improving geometric alignment in the scene reconstruction stage. They generate text-conditioned camera poses and use a feed-forward 3DGS decoder trained on diverse 3D scene datasets to estimate underlying 3D structures from posed multi-view images. However, these methods often fail to fully capture the diverse camera trajectories produced by video generative models. Instead of addressing alignment separately at each stage, SteerX integrates scene reconstruction models directly into the video generation process, ensuring that video frames are optimally structured for accurate scene reconstruction and enabling seamless integration of both stages. + +# 2.3. Guided Sampling + +Generative models are trained to represent data distributions across various domains, including images [30, 48, 49, 52], videos [26, 31], and 3D [43, 72]. However, sampling from tilted distributions (e.g., posterior distributions from measurements) is often preferable to drawing from naive prior. + +Recent works [6, 77] have encoded user preferences through reward functions and fine-tuned generative models to maximize these rewards, effectively sampling from a tilted distribution toward high-reward outputs. However, fine-tuning approaches require extensive training and hampers the generalizability by modifying the base distribution. + +In contrast, guided sampling methods [13, 17, 57, 58, 64] allow sampling from a tilted distribution in a zero-shot manner while keeping the base distribution intact. Among various paradigms, two classes stand out: gradient-based guidance [13, 57] and particle filtering[19, 36, 74]. Gradient-based guidance is simple to implement and widely applied but requires differentiable reward functions and is memory-intensive. In contrast, particle filtering is amenable to non-differentiable rewards, and is more memory-efficient, as it does not rely on backpropagation. Recently, Feynman-Kac Steering (FKS) [55] has unified previous particle filtering techniques, introducing a general steering framework that + +can be applied to any reward function. SteerX builds upon FKS, integrating geometric reward functions and exploring a tailored design space for geometric steering, making it applicable to any video generative models. + +# 3. Preliminary: Feynman-Kac Steering + +Diffusion models [18, 58] are learned to reverse a forward process $\mathbf{x}_t \sim q(\mathbf{x}_t | \mathbf{x}_0 = \mathbf{x})$ with $T$ timesteps, that gradually transforms a data sample $\mathbf{x}$ into Gaussian noise. This reverse process $p_{\theta}(\mathbf{x}_{t-1} | \mathbf{x}_t)$ is defined as: + +$$ +p _ {\theta} \left(\mathbf {x} _ {T}, \mathbf {x} _ {T - 1}, \dots , \mathbf {x} _ {1}, \mathbf {x} _ {0}\right) \propto \prod_ {t = 1} ^ {T} p _ {\theta} \left(\mathbf {x} _ {t - 1} \mid \mathbf {x} _ {t}\right). \tag {1} +$$ + +The main goal of inference-time steering methods is to sample tilted distributions toward maximizing an exponential of user-specified reward function $r_{\phi}(\mathbf{x}_0)$ as: + +$$ +\tilde {p} _ {\theta} \left(\mathbf {x} _ {0}\right) \propto p _ {\theta} \left(\mathbf {x} _ {0}\right) \exp \left(\lambda r _ {\phi} \left(\mathbf {x} _ {0}\right)\right). \tag {2} +$$ + +Feynman-Kac Steering (FKS) [55] introduces a system of multiple interacting diffusion paths (i.e. particles), where each path $\mathbf{x}_{T:0} = (\mathbf{x}_T,\dots ,\mathbf{x}_0)$ is resampled at intermediate steps based on potentials $G_{t}(\mathbf{x}_{T:t})$ , resulting in a sequence of tilted distributions $\tilde{p}_{\theta}(\mathbf{x}_{T:t})$ : + +$$ +\tilde {p} _ {\theta} \left(\mathbf {x} _ {T: t}\right) \propto \prod_ {s = T} ^ {t} p _ {\theta} \left(\mathbf {x} _ {T: s}\right) G _ {s} \left(\mathbf {x} _ {T: s}\right), \tag {3} +$$ + +where the product of potentials matches the total tilt of the base distribution as $\prod_{t=T}^{0} G_{t}(\mathbf{x}_{T:t}) = \exp(\lambda r_{\phi}(\mathbf{x}_{0}))$ . These potentials are generally defined as SMC [74] to score the particles at each transition step $\mathbf{x}_{t+1} \rightarrow \mathbf{x}_{t}$ based on intermediate rewards $r_{\phi}(\mathbf{x}_{t})$ . + +# 4. Methods + +In this section, we introduce SteerX, a zero-shot inference-time steering method for 3D/4D scene generation. SteerX unifies pose-free feed-forward scene reconstruction models into the video generation process, iteratively tilting the data distribution towards geometrically aligned samples. In Section 4.1, we define two geometric reward functions to evaluate geometric consistency in generated multi-view images and dynamic videos, respectively. In Section 4.2, we detail our SMC-based steering algorithm. + +# 4.1. Geometric Rewards + +Our geometric rewards build upon MEt3R [1], which utilizes DUSt3R [68] to measure feature similarity in overlapping regions between image pairs. It extracts upscaled DINO features [8, 22] and computes the cosine similarity between reference view features and the rendered features + +![](images/646add79ba657e36d048a37e04351cbde2394b7c39bd44d61bd6119ccf47364b.jpg) +Figure 2. An overview of geometric rewards. Our reward functions assess the geometric consistency of intermediate generated video frames by computing the feature similarity of upscaled DINO features. (a) GS-MEt3R evaluates feature similarity between the original video frames and their corresponding rendered images from 3DGS. (b) Dyn-MEt3R focuses on background regions by unprojecting background features from half of the video frames and reprojecting them onto the remaining frames to compute feature similarity. + +from another viewpoint. It can evaluate multiple images by computing feature similarity across all possible image pairs and averaging the scores. However, as the number of images increases, this becomes computationally expensive, making the overall generation process inefficient. + +To address this, as shown in Fig. 2, we introduce two geometric reward functions, GS-MEt3R and Dyn-MEt3R, which measure geometric consistency across video frames tailored for 3D/4D scene generation, respectively. We employ pose-free feed-forward scene reconstruction methods, MV-DUSt3R+ [61] and MonST3R [86], to reconstruct 3D and 4D scenes, which are subsequently projected back into image space to assess consistency with the original frames. This mitigates computational bottlenecks while efficiently evaluating global consistency in the generated video frames. + +3D Scene Reward. We extend MEt3R [1] in distinct ways to better suit 3D and 4D scene reconstruction models. For 3D scenes, recent methods [61, 81] introduce a mapping function $f_{\phi}$ that directly reconstructs 3DGS and camera poses from $N$ sparse views $\{I_i \in \mathbb{R}^{H \times W \times 3}\}_{i=1}^N$ as: + +$$ +f _ {\phi}: \left\{I _ {i} \right\} _ {i = 1} ^ {N} \rightarrow \left\{\begin{array}{l}\left\{\boldsymbol {\mu} _ {j}, \mathbf {o} _ {j}, \boldsymbol {\Sigma} _ {j}, \mathbf {c} _ {j} \right\} _ {j = 1} ^ {N \times H \times W}\\\left\{P _ {i} \right\} _ {i = 1} ^ {N}\end{array}, \right. \tag {4} +$$ + +where the 3D scene is represented as Gaussian parameters, including position $\pmb{\mu}$ , volume density $\mathbf{o}$ , covariance $\pmb{\Sigma}$ , and color $\mathbf{c}$ . Then, we produce images $\{\tilde{I}_i\}_{i=1}^N$ by rendering the scene with estimated camera poses $\{P_i\}_{i=1}^N$ , ensuring they correspond to the same viewpoints as the input images. Finally, GS-MEt3R is measured by computing the cosine similarity between upscaled DINO [8, 22] features of input im + +ages $\{F_i\}_{i = 1}^N$ and rendered images $\{\hat{F}_i\}_{i = 1}^N$ as: + +$$ +r _ {\phi} = \frac {1}{N} \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {H} \sum_ {k = 1} ^ {W} \frac {F _ {i} ^ {j k} \cdot \hat {F} _ {i} ^ {j k}}{| | F _ {i} ^ {j k} | | | | \hat {F} _ {i} ^ {j k} | |}. \tag {5} +$$ + +This approach not only offers a more direct evaluation of the physical consistency of the 3D scene compared to averaging consistency across all image pair combinations but also indirectly assesses the rendering quality of 3DGS. In other words, high GS-MEt3R scores indicate both geometric alignment and visually realistic 3D scenes. + +4D Scene Reward. While the 3D scene reward function is based on 3DGS, feed-forward dynamic scene reconstruction with Gaussians remains underexplored, making it difficult to directly apply 3DGS-based rewards. Instead, we employ 3D point cloud representations, where MonST3R [86] reconstructs it with point maps $\{X_{i}\}_{i = 1}^{N}$ , binary dynamic masks $\{M_i\}_{i = 1}^N$ , and camera poses $\{P_i\}_{i = 1}^N$ as: + +$$ +f _ {\phi}: \left\{I _ {i} \right\} _ {i = 1} ^ {N} \rightarrow \left\{X _ {i}, M _ {i}, P _ {i} \right\} _ {i = 1} ^ {N}, \tag {6} +$$ + +where we leverage these time-varying point clouds as 4D scene representations and design a reward function for evaluating geometric consistency in dynamic videos. + +Since dynamic masks are produced in the camera pose estimation process to retain only high-confidence points, a well-reconstructed 4D scene should effectively filter out dynamic objects while preserving geometric consistency in the background regions. Therefore, we evaluate the consistency only for background regions of video frames, which are not filtered out by the dynamic mask. To this end, we first split $N$ video frames into two subsets: $\mathcal{I}_{src} = \{I_1, I_3, \ldots, I_N\}$ + +# Algorithm 1 SteerX (v-prediction) + +Required: v-parametrized diffusion model $\mathbf{v}_{\theta}$ , reward function $r_{\phi}$ , number of particles $k$ , and initial noise $\{\mathbf{x}_T^i\}_{i=1}^k \sim \mathcal{N}(0, I)$ . + +# Sampling: + +1: for $t \in \{T - 1, \dots, 0\}$ do +2: for $i\in \{1\dots k\}$ do +3: $\hat{\mathbf{x}}_0^i\gets \sqrt{\bar{\alpha}_{t + 1}}\mathbf{x}_{t + 1}^i -\sqrt{1 - \bar{\alpha}_{t + 1}}\mathbf{v}_\theta (\mathbf{x}_{t + 1}^i)$ +4: $\mathbf{x}_t^i\gets dpm\text{- solver} (\hat{\mathbf{x}}_0^i,\mathbf{x}_{t + 1}^i)$ +5: $\mathbf{s}_t^i\gets r_\phi (\hat{\mathbf{x}}_0^i)$ Intermediaterewards +6: $G_{t}^{i}\gets \exp (\lambda \max_{j = t}^{T}(\mathbf{s}_{j}^{i}))$ $\triangleright$ Potential +7: end for +8: $\{\mathbf{x}_t^i\}_{i=1}^k \sim \text{Multinomial}(\{\mathbf{x}_t^i, G_t^i\}_{i=1}^k) \triangleright \text{Resample}$ +9: end for +10: $l\gets \arg \max_{i\in \{1,\ldots ,k\}}r_{\phi}(\mathbf{x}_0^i)$ +11: return $\mathbf{x}_0^l$ + +and $\mathcal{I}_{tgt} = \{I_2, I_4, \ldots, I_{N-1}\}$ . Then, we unproject the upsampled DINO features of background regions in $\mathcal{I}_{src}$ into 3D space using MonST3R. Finally, we reproject them onto the viewpoint of $\mathcal{I}_{tgt}$ , where the rendered features $\hat{\mathcal{F}}_{tgt} = \{\hat{F}_1, \hat{F}_3, \ldots, \hat{F}_N\}$ are used to compute the feature similarity with background regions in $\mathcal{I}_{tgt}$ as: + +$$ +r _ {i} = \sum_ {j = 1} ^ {H} \sum_ {k = 1} ^ {W} \left(1 - M _ {i} ^ {j k}\right) \frac {F _ {i} ^ {j k} \cdot \hat {F} _ {i} ^ {j k}}{| | F _ {i} ^ {j k} | | | | \hat {F} _ {i} ^ {j k} | |}, \tag {7} +$$ + +$$ +r _ {\phi} = \frac {1}{(N / / 2)} \sum_ {i = 1} ^ {N / / 2} r _ {i}. \tag {8} +$$ + +With these geometric rewards, we can effectively steer pre-trained video generative models to produce geometrically consistent video frames, which are then directly reconstructed into 3D and 4D spaces using the feed-forward 3D reconstruction models employed in the geometric rewards. + +# 4.2. Geometric Steering + +Using the rewards defined in Section 4.1, let $\tilde{p}_{\theta}$ in (2) be the distribution we wish to sample from. SMC operates with the three following steps: + +1. (Proposal) For each particle $i$ , sample from the proposal distribution $\mathbf{x}_t^i \sim q_t(\mathbf{x}_t|\mathbf{x}_{t+1}^i)$ +2. (Weighting) Compute weights from reward-based potentials $\omega_{t}^{i} = \frac{p_{\theta}(\mathbf{x}_{t}^{i}|\mathbf{x}_{t + 1}^{i})}{q_{t}(\mathbf{x}_{t}^{i}|\mathbf{x}_{t + 1}^{i})} G_{t}(\mathbf{x}_{T:t}^{i})$ +3. (Resampling) Draw new particles from the multinomial distribution $\{\mathbf{x}_t^i\}_{i=1}^k \sim \text{Multinomial}(\{\mathbf{x}_t^i, G_t^i\}_{i=1}^k)$ + +Two choices should be made: the potential $G_{t}$ , and the proposal distribution $q_{t}$ . For the potential, we use max potential + +$$ +G _ {t} \left(\mathbf {x} _ {T: t}\right) ^ {i} := \exp \left(\lambda \max _ {j = t} ^ {T} \left[ r _ {\phi} \left(\hat {\mathbf {x}} _ {0}\right) \right]\right), \tag {9} +$$ + +with + +$$ +G _ {0} \left(\mathbf {x} _ {T: 0}\right) := \exp \left(\lambda r _ {\phi} (\mathbf {x} _ {0})\right) \left(\prod_ {t = 1} ^ {T} G _ {t} \left(\mathbf {x} _ {T: t}\right)\right) ^ {- 1}, \tag {10} +$$ + +such that the particle with the highest reward is preferred. Notice that we use the Tweedie estimate $\hat{\mathbf{x}}_0 = \mathbb{E}[\mathbf{x}_0|\mathbf{x}_t]$ [13, 21, 55] in intermediate steps to avoid full reverse sampling. For the proposal kernel, to save computation, we leverage DPM-solver++ [44], which approximates the true sampling trajectory limited to small neural function evaluation (NFE). These choices lead to SteerX, as shown in Alg. 1. + +Proposition 1. Given the reverse generative process in (1), let $q_{t}$ be the transition kernel satisfying + +$$ +\frac {p _ {\theta} \left(\mathbf {x} _ {t - 1} \mid \mathbf {x} _ {t}\right)}{q _ {t} \left(\mathbf {x} _ {t - 1} \mid \mathbf {x} _ {t}\right)} = 1 + \epsilon_ {t} \left(\mathbf {x} _ {T: t}\right), \tag {11} +$$ + +with $|\epsilon_t(\mathbf{x}_{T:t})| \leq \varepsilon$ uniformly. Also assume that the error from the reward computed at the approximate state $\hat{\mathbf{x}}_0$ is bounded, i.e. $|r_{\phi}(\hat{\mathbf{x}}_0) - r_{\phi}(\mathbf{x}_0)| \leq \eta$ . Then, given the defined max potentials in (9), (10), Alg. 1 samples from + +$$ +\tilde {p} _ {\theta} \left(\mathbf {x} _ {0}\right) \propto p _ {\theta} \left(\mathbf {x} _ {0}\right) \exp \left(\lambda r _ {\phi} \left(\mathbf {x} _ {0}\right)\right) \left(1 + \mathcal {O} \left(T \varepsilon + \lambda \eta\right)\right) \tag {12} +$$ + +See Appendix A for the proof. Prop. 1 states that when the approximation errors are sufficiently small, then we can sample from the desired tilted distribution with SteerX. + +# 5. Experimental Results + +In this section, we conduct extensive experiments to verify the scalability and effectiveness of SteerX across various video generative models in four scene generation scenarios: Text-to-4D, Image-to-4D, Text-to-3D, and Image-to-3D. We briefly explain our experimental setup in Section 5.1 and present both qualitative and quantitative results, as well as design choices of SteerX in Section 5.2. + +# 5.1. Experimental Setup + +Implementation Details. For video generative models, we generate 25 frames at a $480 \times 480$ resolution, except for models that require a fixed video length. Since reconstructing 3D scenes using all generated video frames is impractical, we use only eight frames for the reconstruction. We follow the default settings of MV-DUST3R+ [61] and MonST3R [86]. We set $\lambda = 10$ for the potential. To encourage exploration and reduce computation overhead, we adopt interval resampling, as in FKS [55], and apply the steering procedure at $M$ selected timesteps out of $T$ reverse diffusion steps, setting $M = 4$ for 3D scene generation and $M = 2$ for 4D scene generation. + +![](images/f5061d47d90e7fce6062e396346ab58f785bc1621882e8c56922f04aba5e23e7.jpg) +Figure 3. Qualitative results of video generation in PenguinVideoBenchmark [35]. We visualize for Mochi [62] (top) and Hunyuan-Video [35] (bottom). SteerX achieves the best alignment with camera motion instructions while maintaining high video quality. + +Baselines and Benchmarks. We evaluate our SteerX with Mochi [62] and HunyuanVideo [35] for Text-to-4D generation, and CogVideoX [80] for Image-to-4D. For Image-to-3D, we employ S-Director with orbit-left from DimensionX [59]. For Text-to-3D, we utilize the video generation and scene reconstruction models proposed in SplatFlow [24]. We conduct evaluations on 98 samples from the PenguinVideoBenchmark [35] with camera descriptions for Text-to-4D generation, and 355 samples from VBench-I2V [32] for Image-to-3D and Image-to-4D generation. Additionally, we use 100 samples from the Single-Object-with-Surrounding set of T3Bench [28] for Text-to-3D evaluation. For all evaluations, we compare against the best-of-N approach, where $k$ particles are generated independently, and the one with the highest reward is selected. Additionally, we include a baseline ( $k = 1$ ) that generates a video and reconstructs the scene without inference-time steering. + +Evaluation Metrics. For 4D generation, we evaluated the Aesthetic Score and Subject Consistency in VBench [32] for video quality and semantic consistency, respectively. Also, we evaluated Temporal Consistency to measure text alignment in Text-to-4D generation, including adherence to camera motion instructions. We evaluated Dynamic Degree to measure the overall dynamicness in Image-to-4D generation. For 3D generation, we followed the evaluation protocols in Director3D [37] and SplatFlow [24], measuring the image quality and the CLIP score [29] for both generated multi-view images and rendered images of 3DGS. Finally, for all generation scenarios, we evaluated the geometric consistency of multi-view images and dynamic videos using GS-MEt3R and Dyn-MEt3R, respectively. + +
MethodkAesthetic↑Subject↑Temporal↑Dyn-MEt3R↑
Mochi [62]10.4910.9500.2430.884
+ BoN40.4980.9410.2440.912
+ SteerX (later)40.4880.9370.2420.910
+ SteerX (linear)40.4900.9490.2440.918
+ SteerX (early)40.5000.9550.2480.929
HunyuanVideo [35]10.5490.9670.2410.911
+ BoN40.5510.9780.2390.931
+ SteerX (later)40.5550.9760.2370.931
+ SteerX (linear)40.5560.9730.2410.943
+ SteerX (early)40.5550.9800.2430.964
+ +Table 1. Quantitative results in PenguinVideoBenchmark [35]. SteerX with annealed resampling timesteps to the early sampling process significantly improves the overall video quality. + +# 5.2. Main Results + +Text-to-4D Generation. Figure 3 and Table 1 present results of text-conditioned video generation. Notably, SteerX significantly outperforms both the baseline and the best-of-N, confirming that our geometric steering effectively resamples particles to guide the data distribution toward geometrically aligned samples. Moreover, quantitative results show a strong correlation between Dyn-MEt3R and other metrics, suggesting that the geometric steering can contribute to both physical consistency and overall video quality. + +We observe that annealed resampling timesteps, where the steering procedure at the early sampling stages outperforms linear resampling, whereas resampling at later stages tends to degrade performance. This observation aligns with previous findings [2, 59], where camera poses for video frames are largely determined early in the sampling process + +![](images/ee45af8a7e241095acc82e50c751535804dd7a12fe92f4cde4a60b01d7b25727.jpg) + +![](images/bcf97e68a83e7c14e10d0f5e94237808b29a4f17d1b1aa7f376f4c1a3789e809.jpg) +Input Image + +![](images/ecadeb1c4079f1a3511ce20995afb3f1d43ee7962258d2d74753088e153ab0c5.jpg) + +![](images/8a0956ed05cfa5a60368219f950f215715aef7518c15cd595eeb868649ed6c17.jpg) +CogVideoX [80] + +![](images/1c45db49760c6149d82ab4ecd209aa54bd1800b3dff0d68875c03936b22f0976.jpg) + +![](images/ba17b7765c03118d83b021179e4364c54ee4d6c43bdc608f240d83faac7d94df.jpg) + +![](images/982b968d258ce237842b6c2aed398ee1e9b99bbed83d5b63cb2fdf4c38743576.jpg) + +![](images/466e437bb0013c49fa561d7086243876faa82ed619d69e13fa1ec92231875992.jpg) + +![](images/6e811a60c73167e59857e283b7f10cd3876fb91b26c46ddecfbc5a3920e4758f.jpg) + +![](images/e46a43182df5084ddbede4cf1c2b5c4ad2fe4a5096d65521df97898a475c5811.jpg) ++Best-of-N + +![](images/6a385882fdc381c7082d2f3b668416fc7d25668713ab4e3bb095469d8ef6c57c.jpg) + +![](images/e622220c3b33a3918c8189caf5ab11c5f132f3c3a1d48dd6f61b77b9f73a5187.jpg) + +![](images/6576cde83e0ade8241f1a4f8cd62198407309dfbda2c77f367ac54d1c4c00ef8.jpg) + +![](images/4a9a9412f541315a38a8799109ed3ca16927e6768924b2971ed9f964958614ac.jpg) + +![](images/e8abb12535553004d9599928fe56f7558a09d656d56d42169f8faab183116de7.jpg) + +![](images/f8d22e0d534ce1a77e6c2e8f4c4d026fedbae33dc8bc6c695ac86b93e76825c7.jpg) ++ ++ SteerX (early) + +![](images/839473ecd6b6cfda4099eea2caf07a2f8bce7be7b83078b6b52d7b40509e1c5c.jpg) + +![](images/4c261f6a2d834073a74ac687ead7e57e2c93ccbfbcecf97abb2a2fc43406aa03.jpg) + +![](images/231b2ef285506b99170364dac01d2b99ec0fa50b5f71d4d564bd22b9f7d56a76.jpg) + +![](images/0e2142af8b404a66c8656c5084606171d3c56cbc6613060154c8b03c00e550a5.jpg) + +![](images/e7d36517131aac71bef94157cf52bef874e6f92395918606317829051654fe82.jpg) +Figure 4. Qualitative results of video generation in VBench-I2V [32]. SteerX enhances vividness and overall visual quality. +Input Image + +![](images/cc064c04aba8a7f50b9c88ea584fdc1d50818ad364045be0c43d4be86bede913.jpg) +CogVideoX +Figure 5. Qualitative results in Image-to-4D. SteerX naturally lifts object motion into 4D spaces, while preserving geometric alignments. + +![](images/a3a62951dcc98557cf040428f95041591c2a93de2a2b08855aabe0c8385e6606.jpg) + +![](images/d13790f3bb9c0ccb395dca33c02ea81afeb7e1ccf33a99996d3313f7fff7c3be.jpg) ++Best-of-N + +![](images/6d8e77e74264f9871ea0c0a79e110ef591fdb98280dcb72ab702e669d59949ec.jpg) + +![](images/2810a864dc46e4fa0fdf53a60ca6af54fde743458903575f60bf1590d3393bb0.jpg) ++SteerX + +![](images/5afe9d78bdb1c380a7e1bed8d80cc2ec80967c497afb6d479b942a92c107534a.jpg) + +![](images/55aa1f8a1453425da741d00f8419f0a10c20590fa60b0be403ef1ad8d5ee0937.jpg) +"Filmed from a first-person perspective, the camera passes through the graffiti alley in Melbourne, Australia, where the graffiti walls are covered with artwork from many artists." +HunyuanVideo +Figure 6. Qualitative results in Text-to-4D. + +![](images/989a8e2ad5d2d6eccafae54599f5088f8e6ed865584572d9dd5d8cae42ed6137.jpg) ++BoN + +![](images/a8a5463465a6f306c5212cccec1a788bd51c39cca942a43a505be904cc3354a5.jpg) ++SteerX + +as the diffusion model conceptualizes global structures. As a result, resampling at early timesteps coarsely aligns the geometry by tilting the data distribution, laying the foundation for a more precise geometric alignment in later stages. + +The geometrically aligned video frames contribute to better 4D scene reconstruction, as shown in Fig. 6. While the baseline struggles to capture accurate camera poses and the best-of-N approach results in blurry 4D scenes, our approach precisely estimates camera poses and generates visually realistic 4D scenes. This highlights the effectiveness of SteerX in the seamless integration of video generation and scene reconstruction, ensuring that generated video frames are optimally structured for precise scene reconstruction. + +Image-to-4D Generation. Figure 4 and Table 2 show the results in image-conditioned video generation, and our SteerX performs the best for all metrics, and we observe that SteerX generates dynamic and realistic video frames, whereas the baseline produces almost motionless frames, and the best-of-N approach results in blurry frames. Also, + +
MethodkAesthetic↑Subject↑Dynamic↑Dyn-MEt3R↑
CogVideoX [80]10.5920.9450.1580.880
+ BoN20.5910.9410.1410.882
+ SteerX (early)20.5930.9450.1610.894
+ BoN40.5940.9440.1430.901
+ SteerX (early)40.5960.9570.1700.909
+ +Table 2. Quantitative results in VBench-I2V [32]. Results demonstrate the scalability of geometric steering as $k$ increases. + +
MethodkRenderingSample
BRISQUE↓NIQE↓CLIP-I↑GS-MEt3R↑
DimensionX [59]137.34.2582.40.708
+ BoN429.84.3383.20.745
+ SteerX (early)429.74.2483.70.749
+ +Table 3. Quantitative results in VBench-I2V [32]. For rendered images, we report results after the refinement process in [24, 37]. + +we validate the scalability of SteerX, as increasing the number of particles $k$ consistently enhances performance across all metrics. As shown in Fig. 5, geometrically aligned video frames can be reconstructed to better 4D scenes, producing realistic dynamic objects and well-aligned backgrounds. + +Image-to-3D Generation. We verify the effectiveness of SteerX in the Image-to-3D scene generation. While multiview images generated by SteerX closely resemble those from the best-of-N approach, SteerX reconstructs 3D scenes with finer details and fewer discontinuities in novel views, as shown in Fig. 8. Furthermore, quantitative results in Table 3 demonstrate that SteerX achieves better visual quality and geometric consistency in both rendered images and generated images compared to other baselines. + +![](images/c3fd68343b530168c1e529c434e32f085f21e34acc9418d2f878a38dcd1df63d.jpg) +Figure 7. Qualitative results of rendered images in T3Bench [28]. SteerX significantly enhances the visual quality of rendered images and textual alignment, demonstrating its compatibility with various video generation and scene reconstruction models. + +![](images/4e2def90e8c2fcd5cad7440959e2b65a7e4683dfdf92e104dc5c4a9404925294.jpg) +(a) Input Image + +![](images/820bdb7269a762b7a3b1189562b133d970d4e4b222c07d5be8cf2530e1210fdb.jpg) +(b) DimensionX + BoN + +![](images/18469372ca97bdffd3ab3b6f87a17c59a6b587540ad8ac11aead7a0c7f5cb5dd.jpg) +(c) DimensionX + SteerX + +Figure 8. Qualitative results in of 3DGS in VBench-I2V [32]. + +
MethodkBRISQUE↓CLIPScore↑GS-MEt3R↑
SplatFlow [24]133.2529.660.727
+ BoN236.8229.550.756
+ SteerX (early)227.4729.560.767
+ BoN428.1429.280.768
+ SteerX (early)426.6529.730.775
+ +Text-to-3D Generation. Notice that SteerX can be applied to any generative model and scene reconstruction model, including the multi-view rectified flow model and the feed-forward 3DGS decoder in SplatFlow [24], enabling geometric steering through GS-MEt3R rewards. Table 4 presents quantitative results of generated multi-view images, demonstrating that SteerX effectively enhances overall performance and further improves scalability as the number of particles increases. Furthermore, as shown in Table 5, these 3D-aligned multi-view images are integrated with existing refinement processes, achieving state-of-the-art performance in 3D scene generation. Figure 7 presents qualita + +Table 4. Quantitative results of multi-view generation in Single-Object-with-Surrounding set of T3Bench [28]. + +
MethodBRISQUE↓NIQE↓CLIPScore↑
DreamFusion [53]90.210.48-
Magic3D [39]92.811.20-
LatentNeRF [45]88.69.19-
SJC [66]82.010.15-
Fantasia3D [11]69.67.65-
ProlificDreamer [70]61.57.07-
Director3D [37]32.34.3532.9
SplatFlow [24]19.64.2433.2
SplatFlow [24]†23.44.8432.7
+ BoN (k=4)17.24.4132.3
+ SteerX (k=4)13.14.3033.4
+ +Table 5. Quantitative results in T3Bench [28]. SplatFlow† is re-implemented results without the stop ray strategy, which is incompatible with our geometric steering pipeline. + +tive results, demonstrating that SteerX produces more text-aligned 3D scenes and visually realistic rendered images. + +# 6. Conclusion + +In this paper, we have introduced SteerX, a zero-shot inference-time steering method for camera-free 3D and 4D scene generation. Instead of addressing physical alignment separately in either video generation or scene reconstruction, SteerX unifies both stages and iteratively tilts the data distribution toward geometrically consistent samples. Extensive experiments across diverse scene generation tasks verify that SteerX effectively enhances visual quality, textual alignment, and geometric consistency. SteerX is practical, enabling zero-shot 3D and 4D scene generation, and is scalable, with geometric alignment improving as the number of particles increases. We believe this efficient scaling property holds great potential, opening new avenues for 3D and 4D scene generation, particularly in test-time scaling. + +# References + +[1] Mohammad Asim, Christopher Wewer, Thomas Wimmer, Bernt Schiele, and Jan Eric Lenssen. Met3r: Measuring multi-view consistency in generated images. arXiv preprint arXiv:2501.06336, 2025. 2, 3, 4 +[2] Sherwin Bahmani, Ivan Skorokhodov, Guocheng Qian, Aliaksandr Siarohin, Willi Menapace, Andrea Tagliasacchi, David B Lindell, and Sergey Tulyakov. Ac3d: Analyzing and improving 3d camera control in video diffusion transformers. arXiv preprint arXiv:2411.18673, 2024. 1, 3, 6 +[3] Sherwin Bahmani, Ivan Skorokhodov, Aliaksandr Siarohin, Willi Menapace, Guocheng Qian, Michael Vasilkovsky, Hsin-Ying Lee, Chaoyang Wang, Jiaxu Zou, Andrea Tagliasacchi, et al. Vd3d: Taming large video diffusion transformers for 3d camera control. arXiv preprint arXiv:2407.12781, 2024. 1 +[4] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5855–5864, 2021. 2 +[5] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 2 +[6] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. In The Twelfth International Conference on Learning Representations, 2024. 3 +[7] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1 +[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 3, 4 +[9] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 2 +[10] Guikun Chen and Wenguan Wang. A survey on 3d gaussian splatting. arXiv preprint arXiv:2401.03890, 2024. 1 +[11] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22246–22256, 2023. 8 +[12] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splattering from sparse + +multi-view images. In European Conference on Computer Vision, pages 370-386. Springer, 2024. 2 +[13] Hyungjin Chung, Jeongsol Kim, Michael T McCann, Marc L Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. arXiv preprint arXiv:2209.14687, 2022. 2, 3, 5 +[14] Hyungjin Chung, Suhyeon Lee, and Jong Chul Ye. Decomposed diffusion sampler for accelerating large-scale inverse problems. arXiv preprint arXiv:2303.05754, 2023. 2 +[15] Wenyan Cong, Hanxue Liang, Peihao Wang, Zhiwen Fan, Tianlong Chen, Mukund Varma, Yi Wang, and Zhangyang Wang. Enhancing nerf akin to enhancing llms: Generalizable nerf transformer with mixture-of-view-experts. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3193-3204, 2023. 2 +[16] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2 +[17] Giannis Daras, Hyungjin Chung, Chieh-Hsin Lai, Yuki Mitsufuji, Jong Chul Ye, Peyman Milanfar, Alexandros G Dimakis, and Mauricio Delbracio. A survey on diffusion models for inverse problems. arXiv preprint arXiv:2410.00083, 2024. 2, 3 +[18] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 3 +[19] Zehao Dou and Yang Song. Diffusion posterior sampling for linear inverse problem solving: A filtering perspective. In The Twelfth International Conference on Learning Representations, 2024. 2, 3 +[20] Arnaud Doucet, Nando De Freitas, and Neil Gordon. An introduction to sequential monte carlo methods. Sequential Monte Carlo methods in practice, pages 3-14, 2001. 2 +[21] Bradley Efron. Tweedie's formula and selection bias. Journal of the American Statistical Association, 106(496):1602-1614, 2011. 5 +[22] Stephanie Fu, Mark Hamilton, Laura Brandt, Axel Feldman, Zhoutong Zhang, and William T Freeman. Fitup: A model-agnostic framework for features at any resolution. arXiv preprint arXiv:2403.10516, 2024. 3, 4 +[23] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 1, 3 +[24] Hyojun Go, Byeongjun Park, Jiho Jang, Jin-Young Kim, Soonwoo Kwon, and Changick Kim. Splatflow: Multiview rectified flow model for 3d gaussian splatting synthesis. arXiv preprint arXiv:2411.16443, 2024. 1, 2, 3, 6, 7, 8, 13 +[25] Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, Poriya Panet, Sapir Weissbuch, Victor Kulikov, Yaki Bitterman, Zeev Melumian, and Ofir Bibi. Ltx-video: Realtime video latent diffusion. arXiv preprint arXiv:2501.00103, 2024. 1 + +[26] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible diffusion modeling of long videos. Advances in Neural Information Processing Systems, 35:27953-27965, 2022. 3 +[27] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrl: Enabling camera control for text-to-video generation. arXiv preprint arXiv:2404.02101, 2024. 1, 3 +[28] Yuze He, Yushi Bai, Matthieu Lin, Wang Zhao, Yubin Hu, Jenny Sheng, Ran Yi, Juanzi Li, and Yong-Jin Liu. T $^3$ bench: Benchmarking current progress in text-to-3d generation, 2023. 6, 8 +[29] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 6 +[30] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3 +[31] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 3 +[32] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 6, 7, 8 +[33] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 2 +[34] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 2 +[35] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. 1, 2, 6, 13, 14 +[36] Xiner Li, Yulai Zhao, Chenyu Wang, Gabriele Scalia, Gokcen Eraslan, Surag Nair, Tommaso Biancalani, Shuiwang Ji, Aviv Regev, Sergey Levine, et al. Derivative-free guidance in continuous and discrete diffusion models with soft value-based decoding. arXiv preprint arXiv:2408.08252, 2024. 2, 3 +[37] Xinyang Li, Zhangyu Lai, Linning Xu, Yansong Qu, Liujuan Cao, Shengchuan Zhang, Bo Dai, and Rongrong Ji. Director3d: Real-world camera trajectory and 3d scene generation from text. Advances in Neural Information Processing Systems, 37:75125-75151, 2025. 2, 3, 6, 7, 8 +[38] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference + +on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 2 +[39] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 300–309, 2023. 8 +[40] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024. 2 +[41] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14458-14467, 2021. 2 +[42] Fangfu Liu, Wenqiang Sun, Hanyang Wang, Yikai Wang, Haowen Sun, Junliang Ye, Jun Zhang, and Yueqi Duan. Reconx: Reconstruct any scene from sparse views with video diffusion model. arXiv preprint arXiv:2408.16767, 2024. 3 +[43] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 3 +[44] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 5 +[45] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 8 +[46] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2 +[47] Byeongjun Park and Changick Kim. Point-dynrf: Point-based dynamic radiance fields from a monocular video. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3171-3181, 2024. 2 +[48] Byeongjun Park, Sangmin Woo, Hyojun Go, Jin-Young Kim, and Changick Kim. Denoising task routing for diffusion models. arXiv preprint arXiv:2310.07138, 2023. 3 +[49] Byeongjun Park, Hyojun Go, Jin-Young Kim, Sangmin Woo, Seokil Ham, and Changick Kim. Switch diffusion transformer: Synergizing denoising tasks with sparse mixture-of-experts. arXiv preprint arXiv:2403.09176, 2024. 3 +[50] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5865-5874, 2021. 2 + +[51] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 2 +[52] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195-4205, 2023. 3 +[53] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 8 +[54] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10901-10911, 2021. 2 +[55] Raghav Singhal, Zachary Horvitz, Ryan Teehan, Mengye Ren, Zhou Yu, Kathleen McKeown, and Rajesh Ranganath. A general framework for inference-time scaling and steering of diffusion models. arXiv preprint arXiv:2501.06848, 2025. 2, 3, 5 +[56] Brandon Smart, Chuanxia Zheng, Iro Laina, and Victor Adrian Prisacariu. Splatt3r: Zero-shot gaussian splatting from uncalibrated image pairs. arXiv preprint arXiv:2408.13912, 2024. 2 +[57] Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. Pseudoinverse-guided diffusion models for inverse problems. In International Conference on Learning Representations, 2023. 2, 3 +[58] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3 +[59] Wenqiang Sun, Shuo Chen, Fangfu Liu, Zilong Chen, Yueqi Duan, Jun Zhang, and Yikai Wang. Dimensionx: Create any 3d and 4d scenes from a single image with controllable video diffusion. arXiv preprint arXiv:2411.04928, 2024. 1, 2, 6, 7 +[60] Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10208-10217, 2024. 2 +[61] Zhenggang Tang, Yuchen Fan, Dilin Wang, Hongyu Xu, Rakesh Ranjan, Alexander Schwing, and Zhicheng Yan. Mvdust3r+: Single-stage scene reconstruction from sparse views in 2 seconds. arXiv preprint arXiv:2412.06974, 2024. 1, 2, 4, 5 +[62] Genmo Team. Mochi 1. https://github.com/genmoai/models, 2024. 1, 2, 6, 13, 14 +[63] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Wang Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. In Computer Graphics Forum, pages 703-735. Wiley Online Library, 2022. 1 + +[64] Masatoshi Uehara, Xingyu Su, Yulai Zhao, Xiner Li, Aviv Regev, Shuiwang Ji, Sergey Levine, and Tommaso Biancalani. Reward-guided iterative refinement in diffusion models at test-time with applications to protein and dna design. arXiv preprint arXiv:2502.14944, 2025. 2, 3 +[65] Diwen Wan, Ruijie Lu, and Gang Zeng. Superpoint gaussian splatting for real-time high-fidelity dynamic scene reconstruction. arXiv preprint arXiv:2406.03697, 2024. 2 +[66] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 8 +[67] Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. 2 +[68] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 2, 3 +[69] Yunsong Wang, Tianxin Huang, Hanlin Chen, and Gim Hee Lee. Freesplat: Generalizable 3d gaussian splatting towards free view synthesis of indoor scenes. Advances in Neural Information Processing Systems, 37:107326-107349, 2025. 2 +[70] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 8 +[71] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1, 3 +[72] Sangmin Woo, Byeongjun Park, Hyojun Go, Jin-Young Kim, and Changick Kim. Harmonyview: Harmonizing consistency and diversity in one-image-to-3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10574–10584, 2024. 3 +[73] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20310-20320, 2024. 1, 2 +[74] Luhuan Wu, Brian Trippe, Christian Naesseth, David Blei, and John P Cunningham. Practical and asymptotically exact conditional sampling in diffusion models. Advances in Neural Information Processing Systems, 36:31372-31403, 2023. 2, 3 +[75] Rundi Wu, Ruiqi Gao, Ben Poole, Alex Trevithick, Changxi Zheng, Jonathan T Barron, and Aleksander Holynski. Cat4d: Create anything in 4d with multi-view video diffusion models. arXiv preprint arXiv:2411.18613, 2024. 1, 3 + +[76] Tong Wu, Yu-Jie Yuan, Ling-Xiao Zhang, Jie Yang, Yan-Pei Cao, Ling-Qi Yan, and Lin Gao. Recent advances in 3d gaussian splatting. Computational Visual Media, 10(4):613-642, 2024. 1 +[77] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: learning and evaluating human preferences for text-to-image generation. In Proceedings of the 37th International Conference on Neural Information Processing Systems, pages 15903-15935, 2023. 3 +[78] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 2 +[79] Xiaofeng Yang, Yiwen Chen, Cheng Chen, Chi Zhang, Yi Xu, Xulei Yang, Fayao Liu, and Guosheng Lin. Learn to optimize denoising scores for 3d generation: A unified and improved diffusion prior on nerf and 3d gaussian splatting. arXiv preprint arXiv:2312.04820, 2023. 2 +[80] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 2, 6, 7, 14 +[81] Botao Ye, Sifei Liu, Haofei Xu, Xueting Li, Marc Pollefeys, Ming-Hsuan Yang, and Songyou Peng. No pose, no problem: Surprisingly simple 3d gaussian splats from sparse unposed images. arXiv preprint arXiv:2410.24207, 2024. 2, 4 +[82] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-22, 2023. 2 +[83] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 3 +[84] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Chenming Zhu, Zhangyang Xiong, Tianyou Liang, et al. Mvmgnet: A large-scale dataset of multi-view images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9150-9161, 2023. 2 +[85] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19447-19456, 2024. 2 +[86] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arXiv:2410.03825, 2024. 1, 2, 4, 5 +[87] Yuyang Zhao, Chung-Ching Lin, Kevin Lin, Zhiwen Yan, Linjie Li, Zhengyuan Yang, Jianfeng Wang, Gim Hee Lee, and Lijuan Wang. Genxd: Generating any 3d and 4d scenes. arXiv preprint arXiv:2411.02319, 2024. 1, 3 + +[88] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018. 2 +[89] Junzhe Zhu, Peiye Zhuang, and Sanmi Koyejo. Hifa: High-fidelity text-to-3d generation with advanced diffusion guidance. arXiv preprint arXiv:2305.18766, 2023. 2 +[90] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2 \ No newline at end of file diff --git a/2025/SteerX_ Creating Any Camera-Free 3D and 4D Scenes with Geometric Steering/images.zip b/2025/SteerX_ Creating Any Camera-Free 3D and 4D Scenes with Geometric Steering/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..41a81d904490da5ab1c71293fea290260c450bd8 --- /dev/null +++ b/2025/SteerX_ Creating Any Camera-Free 3D and 4D Scenes with Geometric Steering/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b7cf6aefd1e0e1a29c0fa707510d97a46edb3d0132699c88e5eed254eb225dbf +size 1053938 diff --git a/2025/SteerX_ Creating Any Camera-Free 3D and 4D Scenes with Geometric Steering/layout.json b/2025/SteerX_ Creating Any Camera-Free 3D and 4D Scenes with Geometric Steering/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..76d740932b1861804107eb6fd1f9ba919696a224 --- /dev/null +++ b/2025/SteerX_ Creating Any Camera-Free 3D and 4D Scenes with Geometric Steering/layout.json @@ -0,0 +1,11839 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 60, + 103, + 550, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 103, + 550, + 121 + ], + "spans": [ + { + "bbox": [ + 60, + 103, + 550, + 121 + ], + "type": "text", + "content": "SteerX: Creating Any Camera-Free 3D and 4D Scenes with Geometric Steering" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 110, + 141, + 500, + 210 + ], + "blocks": [ + { + "bbox": [ + 110, + 141, + 500, + 210 + ], + "lines": [ + { + "bbox": [ + 110, + 141, + 500, + 210 + ], + "spans": [ + { + "bbox": [ + 110, + 141, + 500, + 210 + ], + "type": "image", + "image_path": "fac72fcde383d71a8cd37a08dbe546fb0ec531de0c396a6ad5b940595e7c277f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 63, + 237, + 299, + 312 + ], + "blocks": [ + { + "bbox": [ + 158, + 222, + 212, + 233 + ], + "lines": [ + { + "bbox": [ + 158, + 222, + 212, + 233 + ], + "spans": [ + { + "bbox": [ + 158, + 222, + 212, + 233 + ], + "type": "text", + "content": "Text-to-4D" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 63, + 237, + 299, + 312 + ], + "lines": [ + { + "bbox": [ + 63, + 237, + 299, + 312 + ], + "spans": [ + { + "bbox": [ + 63, + 237, + 299, + 312 + ], + "type": "image", + "image_path": "af83eea19b4ebc5216dbfba59eabd762a2347e84f80d75e241d21a4118168a25.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 316, + 274, + 334 + ], + "lines": [ + { + "bbox": [ + 85, + 316, + 274, + 334 + ], + "spans": [ + { + "bbox": [ + 85, + 316, + 274, + 334 + ], + "type": "text", + "content": "A carpet covers the living room floor, and a puppy is lying on it playing with a kitten. To the right of the carpet, there is a toy ball." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 310, + 237, + 545, + 311 + ], + "blocks": [ + { + "bbox": [ + 402, + 223, + 455, + 234 + ], + "lines": [ + { + "bbox": [ + 402, + 223, + 455, + 234 + ], + "spans": [ + { + "bbox": [ + 402, + 223, + 455, + 234 + ], + "type": "text", + "content": "Text-to-3D" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 237, + 545, + 311 + ], + "lines": [ + { + "bbox": [ + 310, + 237, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 310, + 237, + 545, + 311 + ], + "type": "image", + "image_path": "709b9410f7dd3068975e60ebb2e7ee39d15211698082fdfc378200689da6d189.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 316, + 443, + 325 + ], + "lines": [ + { + "bbox": [ + 332, + 316, + 443, + 325 + ], + "spans": [ + { + "bbox": [ + 332, + 316, + 443, + 325 + ], + "type": "text", + "content": "A pink bicycle leaning against a fence" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 479, + 315, + 533, + 323 + ], + "lines": [ + { + "bbox": [ + 479, + 315, + 533, + 323 + ], + "spans": [ + { + "bbox": [ + 479, + 315, + 533, + 323 + ], + "type": "text", + "content": "A pair of sneakers" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 63, + 346, + 305, + 458 + ], + "blocks": [ + { + "bbox": [ + 277, + 328, + 332, + 367 + ], + "lines": [ + { + "bbox": [ + 277, + 328, + 332, + 367 + ], + "spans": [ + { + "bbox": [ + 277, + 328, + 332, + 367 + ], + "type": "text", + "content": "SteerX" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 63, + 346, + 305, + 458 + ], + "lines": [ + { + "bbox": [ + 63, + 346, + 305, + 458 + ], + "spans": [ + { + "bbox": [ + 63, + 346, + 305, + 458 + ], + "type": "image", + "image_path": "5a9ab954fb11ca1e13d26d72a5ae6712baf93bd1084f2ecdf131db87e3334cc4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 306, + 361, + 550, + 458 + ], + "blocks": [ + { + "bbox": [ + 397, + 347, + 459, + 360 + ], + "lines": [ + { + "bbox": [ + 397, + 347, + 459, + 360 + ], + "spans": [ + { + "bbox": [ + 397, + 347, + 459, + 360 + ], + "type": "text", + "content": "Image-to-3D" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 361, + 550, + 458 + ], + "lines": [ + { + "bbox": [ + 306, + 361, + 550, + 458 + ], + "spans": [ + { + "bbox": [ + 306, + 361, + 550, + 458 + ], + "type": "image", + "image_path": "f5caf242268bd3e7d583e2c4b85817feba91542361dc4e8f4fa48e20abd960be.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 467, + 555, + 490 + ], + "lines": [ + { + "bbox": [ + 55, + 467, + 555, + 490 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 555, + 490 + ], + "type": "text", + "content": "Figure 1. SteerX is a zero-shot inference-time steering approach that seamlessly integrates video generative models [24, 35, 59, 62, 80] and feed-forward scene reconstruction models [24, 61, 86], enabling any 3D and 4D scene generation without explicit camera conditions." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 152, + 502, + 200, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 502, + 200, + 514 + ], + "spans": [ + { + "bbox": [ + 152, + 502, + 200, + 514 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 529, + 296, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 296, + 684 + ], + "type": "text", + "content": "Recent progress in 3D/4D scene generation emphasizes the importance of physical alignment throughout video generation and scene reconstruction. However, existing methods improve the alignment separately at each stage, making it difficult to manage subtle misalignments arising from another stage. Here, we present SteerX, a zero-shot inference-time steering method that unifies scene reconstruction into the generation process, tilting data distributions toward better geometric alignment. To this end, we introduce two geometric reward functions for 3D/4D scene generation by using pose-free feed-forward scene reconstruction models. Through extensive experiments, we demonstrate the effectiveness of SteerX in improving 3D/4D scene generation." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 502, + 394, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 502, + 394, + 514 + ], + "spans": [ + { + "bbox": [ + 315, + 502, + 394, + 514 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 523, + 555, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 523, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 312, + 523, + 555, + 654 + ], + "type": "text", + "content": "Generating 3D and 4D scenes from images or text prompts has attracted significant attention due to its potential applications in AR/VR and robotics [10, 63, 76]. This progress is largely driven by the advancement of generative models [7, 25, 35, 62, 80] and neural scene representations [33, 46, 73]. Generative models learn the underlying distribution of large-scale and high-quality video data, leveraging their scalability without imposing explicit physical constraints. In contrast, neural scene representations lift these distributions into structured 3D or 4D spaces, enforcing physical consistency and enabling more faithful scene modeling." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": "In this context, recent efforts [2, 3, 23, 27, 59, 71, 75, 87] have focused on producing geometrically consistent images by fine-tuning generative models with camera pose parameters, where the generation process follows user-defined or pre-defined camera trajectories. This facilitates seamless 3D" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 703, + 211, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 703, + 211, + 714 + ], + "spans": [ + { + "bbox": [ + 66, + 703, + 211, + 714 + ], + "type": "text", + "content": "*Equal contribution, †Corresponding author" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27326" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "and 4D scene reconstructions but requires complex optimization to regress neural scene representations, increasing computational overhead and hindering practical adaptation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 108, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 108, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 54, + 108, + 294, + 251 + ], + "type": "text", + "content": "To address this inefficiency, another line of work [24, 37] has introduced text-conditioned camera pose generation and feed-forward decoding of pixel-aligned 3DGS [33]. This approach trains 3DGS decoders to learn a mapping function that directly reconstructs 3D scenes from multi-view images, leveraging multiple 3D scene datasets for improved 3D scene modeling. However, the limited datasets used to train pose-conditioned decoders fail to capture the various camera trajectories produced by video generative models. This results in slight misalignments between video generation and scene reconstruction, requiring further refinement steps [53, 79, 89] to enhance geometric consistency." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 252, + 294, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 294, + 335 + ], + "type": "text", + "content": "To sum up, previous works handle geometric alignment separately in either video generation or scene reconstruction. This makes it difficult to address cross-stage misalignments, as inconsistencies in one stage may not be fully corrected in the other. Despite recent efforts to mitigate this issue, achieving precise alignment remains an ongoing challenge due to the indistinct link between the two stages." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 336, + 295, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 295, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 295, + 467 + ], + "type": "text", + "content": "On the other hand, zero-shot guidance methods that alter the sampling trajectory of generative models to enforce physical constraints have been widely explored in recent literature [17, 64]. These methods have shown great generalizability and performance, contingent upon a well-defined reward. For instance, in inverse imaging problems, a closed-form likelihood function can guide prior sampling processes toward posterior sampling [13, 57]. However, directly applying zero-shot guidance to 3D/4D scene generation remains challenging due to the ambiguity in defining reward functions and the significant computational overhead." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 468, + 295, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 295, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 295, + 683 + ], + "type": "text", + "content": "In this work, we address this gap by introducing SteerX, a zero-shot inference-time steering method that seamlessly integrates video generation and scene reconstruction, generating geometrically aligned high-quality 3D and 4D scenes. To achieve this, we define geometric reward functions to assess physical consistency across video frames, drawing inspiration from cycle consistency [90] to guide the generation process toward high-reward outputs. We propose two geometric reward functions tailored for 3D and 4D scene generation. To evaluate geometric consistency across multiple video frames, we extend MEt3R [1], a recent evaluation metric for image pairs, by incorporating advanced pose-free feed-forward scene reconstruction methods such as MV-DUSt3R+ [61] and MonST3R [86]. These reconstruction methods lift intermediate generated video frames' during the reverse sampling process into 3D and 4D spaces. The reconstructed scenes are then projected back into the original image space for consistency evaluation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 684, + 294, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 684, + 294, + 694 + ], + "spans": [ + { + "bbox": [ + 67, + 684, + 294, + 694 + ], + "type": "text", + "content": "However, even when the reward is defined, one cannot" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": "directly adopt standard gradient guidance approaches [13, 14, 57] for several reasons. For one, gradient guidance can be used only when the reward is fully differentiable, which hampers flexibility in the design of reward functions. Moreover, there exist substantial memory constraints to compute gradients for dozens of video frames, limiting their scalability for long video sequences. Therefore, we propose a steering algorithm based on sequential Monte Carlo (SMC) [20], which has gained recent traction [19, 36, 55, 74] due to its applicability to non-differentiable reward functions and favorable inference-time scaling properties." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 208, + 553, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 208, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 208, + 553, + 376 + ], + "type": "text", + "content": "By designing tailored reward functions for geometrically plausible and accurate 3D/4D scene generation, along with a guided sampling process based on SMC, SteerX serves as a fully general framework that integrates any generative video models with any 3D reconstruction models, enabling diverse tasks, including Image-to-3D, Image-to-4D, Text-to-3D, and Text-to-4D generation. Through extensive experiments on both 3D and 4D scene generation with various pre-trained video generative models [24, 35, 59, 62, 80], we demonstrate the effectiveness and broad applicability of our approach. Furthermore, we show that by increasing the number of particles, we see favorable scaling properties, opening up new possibilities that have been rather untouched in 3D/4D scene generation: test-time scaling." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 399, + 405, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 399, + 405, + 412 + ], + "spans": [ + { + "bbox": [ + 314, + 399, + 405, + 412 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 424, + 490, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 490, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 490, + 436 + ], + "type": "text", + "content": "2.1. 3D and 4D Scene Reconstruction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 445, + 553, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 553, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 553, + 554 + ], + "type": "text", + "content": "Various neural scene representations have been explored for 3D and 4D scene reconstruction. Most recent works tend to use Neural Radiance Fields (NeRF) [46] and 3D Gaussian Splats (3DGS) [33] for its high-quality rendering of underlying static scenes [4, 5, 85]. These scene representations have been extended to reconstruct 4D scenes by incorporating additional supervisions (e.g., depth maps [78] and segmentation masks [34]) along with deformable fields [38, 47, 50, 51, 67] and Gaussian splats [65, 73]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "text", + "content": "Recent advancements in 3D and 4D scene datasets [16, 40, 41, 54, 82, 84, 88] have enabled the development of generalizable scene representations, allowing feed-forward scene reconstruction methods [9, 12, 15, 60, 69]. Further advancements have extended these techniques to pose-free feed-forward 3D reconstruction models [56, 61, 68, 81, 86], eliminating the need for camera pose conditions and enabling the scene reconstruction from hundreds of unposed images, making the process more flexible and scalable. We leverage these powerful pose-free 3D reconstruction models, specifically MV-DUSt3R+ [61] and MonST3R [86], to assess geometric consistency and lift generated video frames into 3D and 4D spaces." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 56, + 702, + 181, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 702, + 181, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 702, + 181, + 713 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 56, + 702, + 181, + 713 + ], + "type": "text", + "content": "Denoised predictions, as used in [13]" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27327" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 214, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 214, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 214, + 83 + ], + "type": "text", + "content": "2.2. 3D and 4D Scene Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 233 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 233 + ], + "type": "text", + "content": "Recent 3D and 4D scene generation approaches commonly adopt a two-stage framework, where video generation is followed by scene reconstruction, with a focus on enhancing geometric consistency across both stages. Some works fine-tune video generative models with camera pose parameters [2, 23, 27, 42, 71, 75, 83, 87], where the generation process is conditioned on a user-defined or pre-defined camera trajectory to produce geometrically aligned video frames. These camera-conditioned video frames are used as training data to regress neural scene representations. While they produce plausible scenes, optimizing scene representations from scratch incurs high computational costs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 233, + 296, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 233, + 296, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 233, + 296, + 389 + ], + "type": "text", + "content": "In contrast, camera-free approaches [24, 37] reconstruct 3D scenes without camera conditions, improving geometric alignment in the scene reconstruction stage. They generate text-conditioned camera poses and use a feed-forward 3DGS decoder trained on diverse 3D scene datasets to estimate underlying 3D structures from posed multi-view images. However, these methods often fail to fully capture the diverse camera trajectories produced by video generative models. Instead of addressing alignment separately at each stage, SteerX integrates scene reconstruction models directly into the video generation process, ensuring that video frames are optimally structured for accurate scene reconstruction and enabling seamless integration of both stages." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 394, + 162, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 394, + 162, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 162, + 407 + ], + "type": "text", + "content": "2.3. Guided Sampling" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 412, + 295, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 295, + 472 + ], + "type": "text", + "content": "Generative models are trained to represent data distributions across various domains, including images [30, 48, 49, 52], videos [26, 31], and 3D [43, 72]. However, sampling from tilted distributions (e.g., posterior distributions from measurements) is often preferable to drawing from naive prior." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 472, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 544 + ], + "type": "text", + "content": "Recent works [6, 77] have encoded user preferences through reward functions and fine-tuned generative models to maximize these rewards, effectively sampling from a tilted distribution toward high-reward outputs. However, fine-tuning approaches require extensive training and hampers the generalizability by modifying the base distribution." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 544, + 296, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 296, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 296, + 687 + ], + "type": "text", + "content": "In contrast, guided sampling methods [13, 17, 57, 58, 64] allow sampling from a tilted distribution in a zero-shot manner while keeping the base distribution intact. Among various paradigms, two classes stand out: gradient-based guidance [13, 57] and particle filtering[19, 36, 74]. Gradient-based guidance is simple to implement and widely applied but requires differentiable reward functions and is memory-intensive. In contrast, particle filtering is amenable to non-differentiable rewards, and is more memory-efficient, as it does not rely on backpropagation. Recently, Feynman-Kac Steering (FKS) [55] has unified previous particle filtering techniques, introducing a general steering framework that" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "can be applied to any reward function. SteerX builds upon FKS, integrating geometric reward functions and exploring a tailored design space for geometric steering, making it applicable to any video generative models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 131, + 517, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 131, + 517, + 146 + ], + "spans": [ + { + "bbox": [ + 313, + 131, + 517, + 146 + ], + "type": "text", + "content": "3. Preliminary: Feynman-Kac Steering" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 152, + 554, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 554, + 200 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 554, + 200 + ], + "type": "text", + "content": "Diffusion models [18, 58] are learned to reverse a forward process " + }, + { + "bbox": [ + 313, + 152, + 554, + 200 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t \\sim q(\\mathbf{x}_t | \\mathbf{x}_0 = \\mathbf{x})" + }, + { + "bbox": [ + 313, + 152, + 554, + 200 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 152, + 554, + 200 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 152, + 554, + 200 + ], + "type": "text", + "content": " timesteps, that gradually transforms a data sample " + }, + { + "bbox": [ + 313, + 152, + 554, + 200 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 152, + 554, + 200 + ], + "type": "text", + "content": " into Gaussian noise. This reverse process " + }, + { + "bbox": [ + 313, + 152, + 554, + 200 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}_{t-1} | \\mathbf{x}_t)" + }, + { + "bbox": [ + 313, + 152, + 554, + 200 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 339, + 209, + 554, + 242 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 209, + 554, + 242 + ], + "spans": [ + { + "bbox": [ + 339, + 209, + 554, + 242 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(\\mathbf {x} _ {T}, \\mathbf {x} _ {T - 1}, \\dots , \\mathbf {x} _ {1}, \\mathbf {x} _ {0}\\right) \\propto \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right). \\tag {1}", + "image_path": "2287cf129f95880f21060cf27ba65e511219ed34e037ff113e7ae8af24718acf.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 251, + 554, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 554, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 554, + 287 + ], + "type": "text", + "content": "The main goal of inference-time steering methods is to sample tilted distributions toward maximizing an exponential of user-specified reward function " + }, + { + "bbox": [ + 313, + 251, + 554, + 287 + ], + "type": "inline_equation", + "content": "r_{\\phi}(\\mathbf{x}_0)" + }, + { + "bbox": [ + 313, + 251, + 554, + 287 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 367, + 296, + 554, + 309 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 296, + 554, + 309 + ], + "spans": [ + { + "bbox": [ + 367, + 296, + 554, + 309 + ], + "type": "interline_equation", + "content": "\\tilde {p} _ {\\theta} \\left(\\mathbf {x} _ {0}\\right) \\propto p _ {\\theta} \\left(\\mathbf {x} _ {0}\\right) \\exp \\left(\\lambda r _ {\\phi} \\left(\\mathbf {x} _ {0}\\right)\\right). \\tag {2}", + "image_path": "9709dbaa929b3cdcb8483660024c1000d977f471502b42507f87d9743fbe089f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 319, + 554, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 554, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 554, + 380 + ], + "type": "text", + "content": "Feynman-Kac Steering (FKS) [55] introduces a system of multiple interacting diffusion paths (i.e. particles), where each path " + }, + { + "bbox": [ + 313, + 319, + 554, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{T:0} = (\\mathbf{x}_T,\\dots ,\\mathbf{x}_0)" + }, + { + "bbox": [ + 313, + 319, + 554, + 380 + ], + "type": "text", + "content": " is resampled at intermediate steps based on potentials " + }, + { + "bbox": [ + 313, + 319, + 554, + 380 + ], + "type": "inline_equation", + "content": "G_{t}(\\mathbf{x}_{T:t})" + }, + { + "bbox": [ + 313, + 319, + 554, + 380 + ], + "type": "text", + "content": ", resulting in a sequence of tilted distributions " + }, + { + "bbox": [ + 313, + 319, + 554, + 380 + ], + "type": "inline_equation", + "content": "\\tilde{p}_{\\theta}(\\mathbf{x}_{T:t})" + }, + { + "bbox": [ + 313, + 319, + 554, + 380 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 361, + 388, + 553, + 421 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 388, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 361, + 388, + 553, + 421 + ], + "type": "interline_equation", + "content": "\\tilde {p} _ {\\theta} \\left(\\mathbf {x} _ {T: t}\\right) \\propto \\prod_ {s = T} ^ {t} p _ {\\theta} \\left(\\mathbf {x} _ {T: s}\\right) G _ {s} \\left(\\mathbf {x} _ {T: s}\\right), \\tag {3}", + "image_path": "ccac995c7ee5c0b118f43daaea7f97e35fc24f28d9d73304b48b8d46c995bec8.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 429, + 555, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 555, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 555, + 489 + ], + "type": "text", + "content": "where the product of potentials matches the total tilt of the base distribution as " + }, + { + "bbox": [ + 313, + 429, + 555, + 489 + ], + "type": "inline_equation", + "content": "\\prod_{t=T}^{0} G_{t}(\\mathbf{x}_{T:t}) = \\exp(\\lambda r_{\\phi}(\\mathbf{x}_{0}))" + }, + { + "bbox": [ + 313, + 429, + 555, + 489 + ], + "type": "text", + "content": ". These potentials are generally defined as SMC [74] to score the particles at each transition step " + }, + { + "bbox": [ + 313, + 429, + 555, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t+1} \\rightarrow \\mathbf{x}_{t}" + }, + { + "bbox": [ + 313, + 429, + 555, + 489 + ], + "type": "text", + "content": " based on intermediate rewards " + }, + { + "bbox": [ + 313, + 429, + 555, + 489 + ], + "type": "inline_equation", + "content": "r_{\\phi}(\\mathbf{x}_{t})" + }, + { + "bbox": [ + 313, + 429, + 555, + 489 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 500, + 375, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 500, + 375, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 500, + 375, + 511 + ], + "type": "text", + "content": "4. Methods" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 520, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 554, + 628 + ], + "type": "text", + "content": "In this section, we introduce SteerX, a zero-shot inference-time steering method for 3D/4D scene generation. SteerX unifies pose-free feed-forward scene reconstruction models into the video generation process, iteratively tilting the data distribution towards geometrically aligned samples. In Section 4.1, we define two geometric reward functions to evaluate geometric consistency in generated multi-view images and dynamic videos, respectively. In Section 4.2, we detail our SMC-based steering algorithm." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 635, + 430, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 430, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 430, + 647 + ], + "type": "text", + "content": "4.1. Geometric Rewards" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "type": "text", + "content": "Our geometric rewards build upon MEt3R [1], which utilizes DUSt3R [68] to measure feature similarity in overlapping regions between image pairs. It extracts upscaled DINO features [8, 22] and computes the cosine similarity between reference view features and the rendered features" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 712 + ], + "type": "text", + "content": "2 Throughout the paper, we flexibly use the term particle filtering and SMC to refer to this class." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27328" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 549, + 277 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 549, + 277 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 549, + 277 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 549, + 277 + ], + "type": "image", + "image_path": "646add79ba657e36d048a37e04351cbde2394b7c39bd44d61bd6119ccf47364b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 280, + 555, + 326 + ], + "lines": [ + { + "bbox": [ + 54, + 280, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 54, + 280, + 555, + 326 + ], + "type": "text", + "content": "Figure 2. An overview of geometric rewards. Our reward functions assess the geometric consistency of intermediate generated video frames by computing the feature similarity of upscaled DINO features. (a) GS-MEt3R evaluates feature similarity between the original video frames and their corresponding rendered images from 3DGS. (b) Dyn-MEt3R focuses on background regions by unprojecting background features from half of the video frames and reprojecting them onto the remaining frames to compute feature similarity." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 334, + 294, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 334, + 294, + 393 + ], + "spans": [ + { + "bbox": [ + 54, + 334, + 294, + 393 + ], + "type": "text", + "content": "from another viewpoint. It can evaluate multiple images by computing feature similarity across all possible image pairs and averaging the scores. However, as the number of images increases, this becomes computationally expensive, making the overall generation process inefficient." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 394, + 295, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 394, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 54, + 394, + 295, + 515 + ], + "type": "text", + "content": "To address this, as shown in Fig. 2, we introduce two geometric reward functions, GS-MEt3R and Dyn-MEt3R, which measure geometric consistency across video frames tailored for 3D/4D scene generation, respectively. We employ pose-free feed-forward scene reconstruction methods, MV-DUSt3R+ [61] and MonST3R [86], to reconstruct 3D and 4D scenes, which are subsequently projected back into image space to assess consistency with the original frames. This mitigates computational bottlenecks while efficiently evaluating global consistency in the generated video frames." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 518, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 295, + 582 + ], + "type": "text", + "content": "3D Scene Reward. We extend MEt3R [1] in distinct ways to better suit 3D and 4D scene reconstruction models. For 3D scenes, recent methods [61, 81] introduce a mapping function " + }, + { + "bbox": [ + 55, + 518, + 295, + 582 + ], + "type": "inline_equation", + "content": "f_{\\phi}" + }, + { + "bbox": [ + 55, + 518, + 295, + 582 + ], + "type": "text", + "content": " that directly reconstructs 3DGS and camera poses from " + }, + { + "bbox": [ + 55, + 518, + 295, + 582 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 518, + 295, + 582 + ], + "type": "text", + "content": " sparse views " + }, + { + "bbox": [ + 55, + 518, + 295, + 582 + ], + "type": "inline_equation", + "content": "\\{I_i \\in \\mathbb{R}^{H \\times W \\times 3}\\}_{i=1}^N" + }, + { + "bbox": [ + 55, + 518, + 295, + 582 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 83, + 587, + 295, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 587, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 83, + 587, + 295, + 622 + ], + "type": "interline_equation", + "content": "f _ {\\phi}: \\left\\{I _ {i} \\right\\} _ {i = 1} ^ {N} \\rightarrow \\left\\{\\begin{array}{l}\\left\\{\\boldsymbol {\\mu} _ {j}, \\mathbf {o} _ {j}, \\boldsymbol {\\Sigma} _ {j}, \\mathbf {c} _ {j} \\right\\} _ {j = 1} ^ {N \\times H \\times W}\\\\\\left\\{P _ {i} \\right\\} _ {i = 1} ^ {N}\\end{array}, \\right. \\tag {4}", + "image_path": "486c3f7c159c36f96ddd0d2d24f3cc1e085b0f5844a277b2d21c966208169856.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "where the 3D scene is represented as Gaussian parameters, including position " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": ", volume density " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": ", covariance " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": ", and color " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": ". Then, we produce images " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\{\\tilde{I}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": " by rendering the scene with estimated camera poses " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\{P_i\\}_{i=1}^N" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": ", ensuring they correspond to the same viewpoints as the input images. Finally, GS-MEt3R is measured by computing the cosine similarity between upscaled DINO [8, 22] features of input im" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 332, + 504, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 504, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 504, + 348 + ], + "type": "text", + "content": "ages " + }, + { + "bbox": [ + 313, + 332, + 504, + 348 + ], + "type": "inline_equation", + "content": "\\{F_i\\}_{i = 1}^N" + }, + { + "bbox": [ + 313, + 332, + 504, + 348 + ], + "type": "text", + "content": " and rendered images " + }, + { + "bbox": [ + 313, + 332, + 504, + 348 + ], + "type": "inline_equation", + "content": "\\{\\hat{F}_i\\}_{i = 1}^N" + }, + { + "bbox": [ + 313, + 332, + 504, + 348 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 359, + 354, + 555, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 354, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 359, + 354, + 555, + 388 + ], + "type": "interline_equation", + "content": "r _ {\\phi} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {H} \\sum_ {k = 1} ^ {W} \\frac {F _ {i} ^ {j k} \\cdot \\hat {F} _ {i} ^ {j k}}{| | F _ {i} ^ {j k} | | | | \\hat {F} _ {i} ^ {j k} | |}. \\tag {5}", + "image_path": "0a07fc3b0e42c2c88ba65bbda141a007a462cc7e71fd04ebaa0d8d7db970305f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 394, + 555, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 555, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 555, + 467 + ], + "type": "text", + "content": "This approach not only offers a more direct evaluation of the physical consistency of the 3D scene compared to averaging consistency across all image pair combinations but also indirectly assesses the rendering quality of 3DGS. In other words, high GS-MEt3R scores indicate both geometric alignment and visually realistic 3D scenes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 469, + 554, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 469, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 469, + 554, + 555 + ], + "type": "text", + "content": "4D Scene Reward. While the 3D scene reward function is based on 3DGS, feed-forward dynamic scene reconstruction with Gaussians remains underexplored, making it difficult to directly apply 3DGS-based rewards. Instead, we employ 3D point cloud representations, where MonST3R [86] reconstructs it with point maps " + }, + { + "bbox": [ + 313, + 469, + 554, + 555 + ], + "type": "inline_equation", + "content": "\\{X_{i}\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 313, + 469, + 554, + 555 + ], + "type": "text", + "content": ", binary dynamic masks " + }, + { + "bbox": [ + 313, + 469, + 554, + 555 + ], + "type": "inline_equation", + "content": "\\{M_i\\}_{i = 1}^N" + }, + { + "bbox": [ + 313, + 469, + 554, + 555 + ], + "type": "text", + "content": ", and camera poses " + }, + { + "bbox": [ + 313, + 469, + 554, + 555 + ], + "type": "inline_equation", + "content": "\\{P_i\\}_{i = 1}^N" + }, + { + "bbox": [ + 313, + 469, + 554, + 555 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 365, + 559, + 553, + 575 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 559, + 553, + 575 + ], + "spans": [ + { + "bbox": [ + 365, + 559, + 553, + 575 + ], + "type": "interline_equation", + "content": "f _ {\\phi}: \\left\\{I _ {i} \\right\\} _ {i = 1} ^ {N} \\rightarrow \\left\\{X _ {i}, M _ {i}, P _ {i} \\right\\} _ {i = 1} ^ {N}, \\tag {6}", + "image_path": "cfa4bfbffad527300e22ab197f6fd19b901c88c000035738b10154c2895d1374.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 553, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 617 + ], + "type": "text", + "content": "where we leverage these time-varying point clouds as 4D scene representations and design a reward function for evaluating geometric consistency in dynamic videos." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "type": "text", + "content": "Since dynamic masks are produced in the camera pose estimation process to retain only high-confidence points, a well-reconstructed 4D scene should effectively filter out dynamic objects while preserving geometric consistency in the background regions. Therefore, we evaluate the consistency only for background regions of video frames, which are not filtered out by the dynamic mask. To this end, we first split " + }, + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "type": "text", + "content": " video frames into two subsets: " + }, + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{src} = \\{I_1, I_3, \\ldots, I_N\\}" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27329" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 74, + 199, + 86 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 74, + 199, + 86 + ], + "spans": [ + { + "bbox": [ + 56, + 74, + 199, + 86 + ], + "type": "text", + "content": "Algorithm 1 SteerX (v-prediction)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 87, + 295, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 87, + 295, + 125 + ], + "spans": [ + { + "bbox": [ + 56, + 87, + 295, + 125 + ], + "type": "text", + "content": "Required: v-parametrized diffusion model " + }, + { + "bbox": [ + 56, + 87, + 295, + 125 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\theta}" + }, + { + "bbox": [ + 56, + 87, + 295, + 125 + ], + "type": "text", + "content": ", reward function " + }, + { + "bbox": [ + 56, + 87, + 295, + 125 + ], + "type": "inline_equation", + "content": "r_{\\phi}" + }, + { + "bbox": [ + 56, + 87, + 295, + 125 + ], + "type": "text", + "content": ", number of particles " + }, + { + "bbox": [ + 56, + 87, + 295, + 125 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 56, + 87, + 295, + 125 + ], + "type": "text", + "content": ", and initial noise " + }, + { + "bbox": [ + 56, + 87, + 295, + 125 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_T^i\\}_{i=1}^k \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 56, + 87, + 295, + 125 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 124, + 102, + 135 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 124, + 102, + 135 + ], + "spans": [ + { + "bbox": [ + 56, + 124, + 102, + 135 + ], + "type": "text", + "content": "Sampling:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 137, + 293, + 297 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 62, + 137, + 181, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 137, + 181, + 149 + ], + "spans": [ + { + "bbox": [ + 62, + 137, + 181, + 149 + ], + "type": "text", + "content": "1: for " + }, + { + "bbox": [ + 62, + 137, + 181, + 149 + ], + "type": "inline_equation", + "content": "t \\in \\{T - 1, \\dots, 0\\}" + }, + { + "bbox": [ + 62, + 137, + 181, + 149 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 152, + 170, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 152, + 170, + 164 + ], + "spans": [ + { + "bbox": [ + 62, + 152, + 170, + 164 + ], + "type": "text", + "content": "2: for " + }, + { + "bbox": [ + 62, + 152, + 170, + 164 + ], + "type": "inline_equation", + "content": "i\\in \\{1\\dots k\\}" + }, + { + "bbox": [ + 62, + 152, + 170, + 164 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 167, + 274, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 167, + 274, + 179 + ], + "spans": [ + { + "bbox": [ + 62, + 167, + 274, + 179 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 62, + 167, + 274, + 179 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0^i\\gets \\sqrt{\\bar{\\alpha}_{t + 1}}\\mathbf{x}_{t + 1}^i -\\sqrt{1 - \\bar{\\alpha}_{t + 1}}\\mathbf{v}_\\theta (\\mathbf{x}_{t + 1}^i)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 182, + 217, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 217, + 194 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 217, + 194 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 62, + 182, + 217, + 194 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t^i\\gets dpm\\text{- solver} (\\hat{\\mathbf{x}}_0^i,\\mathbf{x}_{t + 1}^i)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 196, + 292, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 196, + 292, + 209 + ], + "spans": [ + { + "bbox": [ + 62, + 196, + 292, + 209 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 62, + 196, + 292, + 209 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_t^i\\gets r_\\phi (\\hat{\\mathbf{x}}_0^i)" + }, + { + "bbox": [ + 62, + 196, + 292, + 209 + ], + "type": "text", + "content": " Intermediaterewards" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 211, + 293, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 211, + 293, + 224 + ], + "spans": [ + { + "bbox": [ + 62, + 211, + 293, + 224 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 62, + 211, + 293, + 224 + ], + "type": "inline_equation", + "content": "G_{t}^{i}\\gets \\exp (\\lambda \\max_{j = t}^{T}(\\mathbf{s}_{j}^{i}))" + }, + { + "bbox": [ + 62, + 211, + 293, + 224 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 62, + 211, + 293, + 224 + ], + "type": "text", + "content": " Potential" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 227, + 121, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 227, + 121, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 227, + 121, + 236 + ], + "type": "text", + "content": "7: end for" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 239, + 292, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 292, + 254 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 292, + 254 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 62, + 239, + 292, + 254 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_t^i\\}_{i=1}^k \\sim \\text{Multinomial}(\\{\\mathbf{x}_t^i, G_t^i\\}_{i=1}^k) \\triangleright \\text{Resample}" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 257, + 106, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 257, + 106, + 266 + ], + "spans": [ + { + "bbox": [ + 62, + 257, + 106, + 266 + ], + "type": "text", + "content": "9: end for" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 61, + 270, + 194, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 270, + 194, + 284 + ], + "spans": [ + { + "bbox": [ + 61, + 270, + 194, + 284 + ], + "type": "text", + "content": "10: " + }, + { + "bbox": [ + 61, + 270, + 194, + 284 + ], + "type": "inline_equation", + "content": "l\\gets \\arg \\max_{i\\in \\{1,\\ldots ,k\\}}r_{\\phi}(\\mathbf{x}_0^i)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 59, + 285, + 116, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 285, + 116, + 297 + ], + "spans": [ + { + "bbox": [ + 59, + 285, + 116, + 297 + ], + "type": "text", + "content": "11: return " + }, + { + "bbox": [ + 59, + 285, + 116, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^l" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 314, + 296, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 314, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 314, + 296, + 388 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 55, + 314, + 296, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{tgt} = \\{I_2, I_4, \\ldots, I_{N-1}\\}" + }, + { + "bbox": [ + 55, + 314, + 296, + 388 + ], + "type": "text", + "content": ". Then, we unproject the upsampled DINO features of background regions in " + }, + { + "bbox": [ + 55, + 314, + 296, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{src}" + }, + { + "bbox": [ + 55, + 314, + 296, + 388 + ], + "type": "text", + "content": " into 3D space using MonST3R. Finally, we reproject them onto the viewpoint of " + }, + { + "bbox": [ + 55, + 314, + 296, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{tgt}" + }, + { + "bbox": [ + 55, + 314, + 296, + 388 + ], + "type": "text", + "content": ", where the rendered features " + }, + { + "bbox": [ + 55, + 314, + 296, + 388 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}_{tgt} = \\{\\hat{F}_1, \\hat{F}_3, \\ldots, \\hat{F}_N\\}" + }, + { + "bbox": [ + 55, + 314, + 296, + 388 + ], + "type": "text", + "content": " are used to compute the feature similarity with background regions in " + }, + { + "bbox": [ + 55, + 314, + 296, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{tgt}" + }, + { + "bbox": [ + 55, + 314, + 296, + 388 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 94, + 399, + 295, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 399, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 94, + 399, + 295, + 434 + ], + "type": "interline_equation", + "content": "r _ {i} = \\sum_ {j = 1} ^ {H} \\sum_ {k = 1} ^ {W} \\left(1 - M _ {i} ^ {j k}\\right) \\frac {F _ {i} ^ {j k} \\cdot \\hat {F} _ {i} ^ {j k}}{| | F _ {i} ^ {j k} | | | | \\hat {F} _ {i} ^ {j k} | |}, \\tag {7}", + "image_path": "564839a9ef94c2e5faa7761d802c808b2c99f5feff4ec5e5d9efd9090bc706d0.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 129, + 451, + 295, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 451, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 129, + 451, + 295, + 485 + ], + "type": "interline_equation", + "content": "r _ {\\phi} = \\frac {1}{(N / / 2)} \\sum_ {i = 1} ^ {N / / 2} r _ {i}. \\tag {8}", + "image_path": "0645d2c182e4c2e6d641ac58108c9242790ddb5139b6651c8938c84a06ba68fb.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 495, + 296, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 296, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 296, + 556 + ], + "type": "text", + "content": "With these geometric rewards, we can effectively steer pre-trained video generative models to produce geometrically consistent video frames, which are then directly reconstructed into 3D and 4D spaces using the feed-forward 3D reconstruction models employed in the geometric rewards." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 567, + 170, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 170, + 580 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 170, + 580 + ], + "type": "text", + "content": "4.2. Geometric Steering" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 586, + 295, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 295, + 623 + ], + "type": "text", + "content": "Using the rewards defined in Section 4.1, let " + }, + { + "bbox": [ + 55, + 586, + 295, + 623 + ], + "type": "inline_equation", + "content": "\\tilde{p}_{\\theta}" + }, + { + "bbox": [ + 55, + 586, + 295, + 623 + ], + "type": "text", + "content": " in (2) be the distribution we wish to sample from. SMC operates with the three following steps:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 624, + 295, + 715 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 56, + 624, + 295, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 649 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 649 + ], + "type": "text", + "content": "1. (Proposal) For each particle " + }, + { + "bbox": [ + 56, + 624, + 295, + 649 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 56, + 624, + 295, + 649 + ], + "type": "text", + "content": ", sample from the proposal distribution " + }, + { + "bbox": [ + 56, + 624, + 295, + 649 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t^i \\sim q_t(\\mathbf{x}_t|\\mathbf{x}_{t+1}^i)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 654, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 685 + ], + "type": "text", + "content": "2. (Weighting) Compute weights from reward-based potentials " + }, + { + "bbox": [ + 55, + 654, + 295, + 685 + ], + "type": "inline_equation", + "content": "\\omega_{t}^{i} = \\frac{p_{\\theta}(\\mathbf{x}_{t}^{i}|\\mathbf{x}_{t + 1}^{i})}{q_{t}(\\mathbf{x}_{t}^{i}|\\mathbf{x}_{t + 1}^{i})} G_{t}(\\mathbf{x}_{T:t}^{i})" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "3. (Resampling) Draw new particles from the multinomial distribution " + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_t^i\\}_{i=1}^k \\sim \\text{Multinomial}(\\{\\mathbf{x}_t^i, G_t^i\\}_{i=1}^k)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": "Two choices should be made: the potential " + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "inline_equation", + "content": "G_{t}" + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": ", and the proposal distribution " + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "inline_equation", + "content": "q_{t}" + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": ". For the potential, we use max potential" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 356, + 101, + 555, + 127 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 101, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 356, + 101, + 555, + 127 + ], + "type": "interline_equation", + "content": "G _ {t} \\left(\\mathbf {x} _ {T: t}\\right) ^ {i} := \\exp \\left(\\lambda \\max _ {j = t} ^ {T} \\left[ r _ {\\phi} \\left(\\hat {\\mathbf {x}} _ {0}\\right) \\right]\\right), \\tag {9}", + "image_path": "46a6b27f2530494dae2282c804ad2e5bd59c99e3152519317d84fca4a7f63a2e.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 314, + 130, + 335, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 130, + 335, + 141 + ], + "spans": [ + { + "bbox": [ + 314, + 130, + 335, + 141 + ], + "type": "text", + "content": "with" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 325, + 144, + 554, + 180 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 144, + 554, + 180 + ], + "spans": [ + { + "bbox": [ + 325, + 144, + 554, + 180 + ], + "type": "interline_equation", + "content": "G _ {0} \\left(\\mathbf {x} _ {T: 0}\\right) := \\exp \\left(\\lambda r _ {\\phi} (\\mathbf {x} _ {0})\\right) \\left(\\prod_ {t = 1} ^ {T} G _ {t} \\left(\\mathbf {x} _ {T: t}\\right)\\right) ^ {- 1}, \\tag {10}", + "image_path": "004c90a94132ad268a136cc1dce87a93860ba68baaab72063acd5694a509baa4.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 183, + 555, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 183, + 555, + 267 + ], + "spans": [ + { + "bbox": [ + 313, + 183, + 555, + 267 + ], + "type": "text", + "content": "such that the particle with the highest reward is preferred. Notice that we use the Tweedie estimate " + }, + { + "bbox": [ + 313, + 183, + 555, + 267 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0 = \\mathbb{E}[\\mathbf{x}_0|\\mathbf{x}_t]" + }, + { + "bbox": [ + 313, + 183, + 555, + 267 + ], + "type": "text", + "content": " [13, 21, 55] in intermediate steps to avoid full reverse sampling. For the proposal kernel, to save computation, we leverage DPM-solver++ [44], which approximates the true sampling trajectory limited to small neural function evaluation (NFE). These choices lead to SteerX, as shown in Alg. 1." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 272, + 554, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 272, + 554, + 296 + ], + "spans": [ + { + "bbox": [ + 313, + 272, + 554, + 296 + ], + "type": "text", + "content": "Proposition 1. Given the reverse generative process in (1), let " + }, + { + "bbox": [ + 313, + 272, + 554, + 296 + ], + "type": "inline_equation", + "content": "q_{t}" + }, + { + "bbox": [ + 313, + 272, + 554, + 296 + ], + "type": "text", + "content": " be the transition kernel satisfying" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 373, + 300, + 554, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 300, + 554, + 327 + ], + "spans": [ + { + "bbox": [ + 373, + 300, + 554, + 327 + ], + "type": "interline_equation", + "content": "\\frac {p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right)}{q _ {t} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right)} = 1 + \\epsilon_ {t} \\left(\\mathbf {x} _ {T: t}\\right), \\tag {11}", + "image_path": "e2136f092a2aee96bbc9d77e3f0056a55ac65eae6136371e32b9545307af9358.jpg" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 330, + 555, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 555, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 555, + 378 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 313, + 330, + 555, + 378 + ], + "type": "inline_equation", + "content": "|\\epsilon_t(\\mathbf{x}_{T:t})| \\leq \\varepsilon" + }, + { + "bbox": [ + 313, + 330, + 555, + 378 + ], + "type": "text", + "content": " uniformly. Also assume that the error from the reward computed at the approximate state " + }, + { + "bbox": [ + 313, + 330, + 555, + 378 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0" + }, + { + "bbox": [ + 313, + 330, + 555, + 378 + ], + "type": "text", + "content": " is bounded, i.e. " + }, + { + "bbox": [ + 313, + 330, + 555, + 378 + ], + "type": "inline_equation", + "content": "|r_{\\phi}(\\hat{\\mathbf{x}}_0) - r_{\\phi}(\\mathbf{x}_0)| \\leq \\eta" + }, + { + "bbox": [ + 313, + 330, + 555, + 378 + ], + "type": "text", + "content": ". Then, given the defined max potentials in (9), (10), Alg. 1 samples from" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 321, + 384, + 554, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 384, + 554, + 397 + ], + "spans": [ + { + "bbox": [ + 321, + 384, + 554, + 397 + ], + "type": "interline_equation", + "content": "\\tilde {p} _ {\\theta} \\left(\\mathbf {x} _ {0}\\right) \\propto p _ {\\theta} \\left(\\mathbf {x} _ {0}\\right) \\exp \\left(\\lambda r _ {\\phi} \\left(\\mathbf {x} _ {0}\\right)\\right) \\left(1 + \\mathcal {O} \\left(T \\varepsilon + \\lambda \\eta\\right)\\right) \\tag {12}", + "image_path": "fe2110028e9f756bee41cf6ca8e696d2dd556260eefa9935ca3e95b069de2169.jpg" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 401, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 554, + 437 + ], + "type": "text", + "content": "See Appendix A for the proof. Prop. 1 states that when the approximation errors are sufficiently small, then we can sample from the desired tilted distribution with SteerX." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 446, + 440, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 446, + 440, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 446, + 440, + 460 + ], + "type": "text", + "content": "5. Experimental Results" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 467, + 555, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 555, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 555, + 551 + ], + "type": "text", + "content": "In this section, we conduct extensive experiments to verify the scalability and effectiveness of SteerX across various video generative models in four scene generation scenarios: Text-to-4D, Image-to-4D, Text-to-3D, and Image-to-3D. We briefly explain our experimental setup in Section 5.1 and present both qualitative and quantitative results, as well as design choices of SteerX in Section 5.2." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 555, + 430, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 430, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 430, + 569 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 573, + 555, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 573, + 555, + 717 + ], + "spans": [ + { + "bbox": [ + 313, + 573, + 555, + 717 + ], + "type": "text", + "content": "Implementation Details. For video generative models, we generate 25 frames at a " + }, + { + "bbox": [ + 313, + 573, + 555, + 717 + ], + "type": "inline_equation", + "content": "480 \\times 480" + }, + { + "bbox": [ + 313, + 573, + 555, + 717 + ], + "type": "text", + "content": " resolution, except for models that require a fixed video length. Since reconstructing 3D scenes using all generated video frames is impractical, we use only eight frames for the reconstruction. We follow the default settings of MV-DUST3R+ [61] and MonST3R [86]. We set " + }, + { + "bbox": [ + 313, + 573, + 555, + 717 + ], + "type": "inline_equation", + "content": "\\lambda = 10" + }, + { + "bbox": [ + 313, + 573, + 555, + 717 + ], + "type": "text", + "content": " for the potential. To encourage exploration and reduce computation overhead, we adopt interval resampling, as in FKS [55], and apply the steering procedure at " + }, + { + "bbox": [ + 313, + 573, + 555, + 717 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 573, + 555, + 717 + ], + "type": "text", + "content": " selected timesteps out of " + }, + { + "bbox": [ + 313, + 573, + 555, + 717 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 573, + 555, + 717 + ], + "type": "text", + "content": " reverse diffusion steps, setting " + }, + { + "bbox": [ + 313, + 573, + 555, + 717 + ], + "type": "inline_equation", + "content": "M = 4" + }, + { + "bbox": [ + 313, + 573, + 555, + 717 + ], + "type": "text", + "content": " for 3D scene generation and " + }, + { + "bbox": [ + 313, + 573, + 555, + 717 + ], + "type": "inline_equation", + "content": "M = 2" + }, + { + "bbox": [ + 313, + 573, + 555, + 717 + ], + "type": "text", + "content": " for 4D scene generation." + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27330" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 68, + 553, + 300 + ], + "blocks": [ + { + "bbox": [ + 58, + 68, + 553, + 300 + ], + "lines": [ + { + "bbox": [ + 58, + 68, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 58, + 68, + 553, + 300 + ], + "type": "image", + "image_path": "f5061d47d90e7fce6062e396346ab58f785bc1621882e8c56922f04aba5e23e7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 302, + 555, + 326 + ], + "lines": [ + { + "bbox": [ + 55, + 302, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 302, + 555, + 326 + ], + "type": "text", + "content": "Figure 3. Qualitative results of video generation in PenguinVideoBenchmark [35]. We visualize for Mochi [62] (top) and Hunyuan-Video [35] (bottom). SteerX achieves the best alignment with camera motion instructions while maintaining high video quality." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 334, + 297, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 334, + 297, + 539 + ], + "spans": [ + { + "bbox": [ + 54, + 334, + 297, + 539 + ], + "type": "text", + "content": "Baselines and Benchmarks. We evaluate our SteerX with Mochi [62] and HunyuanVideo [35] for Text-to-4D generation, and CogVideoX [80] for Image-to-4D. For Image-to-3D, we employ S-Director with orbit-left from DimensionX [59]. For Text-to-3D, we utilize the video generation and scene reconstruction models proposed in SplatFlow [24]. We conduct evaluations on 98 samples from the PenguinVideoBenchmark [35] with camera descriptions for Text-to-4D generation, and 355 samples from VBench-I2V [32] for Image-to-3D and Image-to-4D generation. Additionally, we use 100 samples from the Single-Object-with-Surrounding set of T3Bench [28] for Text-to-3D evaluation. For all evaluations, we compare against the best-of-N approach, where " + }, + { + "bbox": [ + 54, + 334, + 297, + 539 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 54, + 334, + 297, + 539 + ], + "type": "text", + "content": " particles are generated independently, and the one with the highest reward is selected. Additionally, we include a baseline (" + }, + { + "bbox": [ + 54, + 334, + 297, + 539 + ], + "type": "inline_equation", + "content": "k = 1" + }, + { + "bbox": [ + 54, + 334, + 297, + 539 + ], + "type": "text", + "content": ") that generates a video and reconstructs the scene without inference-time steering." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "text", + "content": "Evaluation Metrics. For 4D generation, we evaluated the Aesthetic Score and Subject Consistency in VBench [32] for video quality and semantic consistency, respectively. Also, we evaluated Temporal Consistency to measure text alignment in Text-to-4D generation, including adherence to camera motion instructions. We evaluated Dynamic Degree to measure the overall dynamicness in Image-to-4D generation. For 3D generation, we followed the evaluation protocols in Director3D [37] and SplatFlow [24], measuring the image quality and the CLIP score [29] for both generated multi-view images and rendered images of 3DGS. Finally, for all generation scenarios, we evaluated the geometric consistency of multi-view images and dynamic videos using GS-MEt3R and Dyn-MEt3R, respectively." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 317, + 332, + 553, + 460 + ], + "blocks": [ + { + "bbox": [ + 317, + 332, + 553, + 460 + ], + "lines": [ + { + "bbox": [ + 317, + 332, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 317, + 332, + 553, + 460 + ], + "type": "table", + "html": "
MethodkAesthetic↑Subject↑Temporal↑Dyn-MEt3R↑
Mochi [62]10.4910.9500.2430.884
+ BoN40.4980.9410.2440.912
+ SteerX (later)40.4880.9370.2420.910
+ SteerX (linear)40.4900.9490.2440.918
+ SteerX (early)40.5000.9550.2480.929
HunyuanVideo [35]10.5490.9670.2410.911
+ BoN40.5510.9780.2390.931
+ SteerX (later)40.5550.9760.2370.931
+ SteerX (linear)40.5560.9730.2410.943
+ SteerX (early)40.5550.9800.2430.964
", + "image_path": "f5520b64dd3e29c763553668d7c9057f501db4c72e1ea70c16e7c8372ed0890e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 462, + 555, + 497 + ], + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 497 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 497 + ], + "type": "text", + "content": "Table 1. Quantitative results in PenguinVideoBenchmark [35]. SteerX with annealed resampling timesteps to the early sampling process significantly improves the overall video quality." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 510, + 400, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 510, + 400, + 522 + ], + "spans": [ + { + "bbox": [ + 314, + 510, + 400, + 522 + ], + "type": "text", + "content": "5.2. Main Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 531, + 555, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 555, + 639 + ], + "type": "text", + "content": "Text-to-4D Generation. Figure 3 and Table 1 present results of text-conditioned video generation. Notably, SteerX significantly outperforms both the baseline and the best-of-N, confirming that our geometric steering effectively resamples particles to guide the data distribution toward geometrically aligned samples. Moreover, quantitative results show a strong correlation between Dyn-MEt3R and other metrics, suggesting that the geometric steering can contribute to both physical consistency and overall video quality." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "We observe that annealed resampling timesteps, where the steering procedure at the early sampling stages outperforms linear resampling, whereas resampling at later stages tends to degrade performance. This observation aligns with previous findings [2, 59], where camera poses for video frames are largely determined early in the sampling process" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27331" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 108, + 119 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 108, + 119 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 108, + 119 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 108, + 119 + ], + "type": "image", + "image_path": "ee45af8a7e241095acc82e50c751535804dd7a12fe92f4cde4a60b01d7b25727.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 121, + 108, + 169 + ], + "blocks": [ + { + "bbox": [ + 59, + 121, + 108, + 169 + ], + "lines": [ + { + "bbox": [ + 59, + 121, + 108, + 169 + ], + "spans": [ + { + "bbox": [ + 59, + 121, + 108, + 169 + ], + "type": "image", + "image_path": "bcf97e68a83e7c14e10d0f5e94237808b29a4f17d1b1aa7f376f4c1a3789e809.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 172, + 106, + 183 + ], + "lines": [ + { + "bbox": [ + 61, + 172, + 106, + 183 + ], + "spans": [ + { + "bbox": [ + 61, + 172, + 106, + 183 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 109, + 70, + 157, + 119 + ], + "blocks": [ + { + "bbox": [ + 109, + 70, + 157, + 119 + ], + "lines": [ + { + "bbox": [ + 109, + 70, + 157, + 119 + ], + "spans": [ + { + "bbox": [ + 109, + 70, + 157, + 119 + ], + "type": "image", + "image_path": "ecadeb1c4079f1a3511ce20995afb3f1d43ee7962258d2d74753088e153ab0c5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 109, + 121, + 157, + 169 + ], + "blocks": [ + { + "bbox": [ + 109, + 121, + 157, + 169 + ], + "lines": [ + { + "bbox": [ + 109, + 121, + 157, + 169 + ], + "spans": [ + { + "bbox": [ + 109, + 121, + 157, + 169 + ], + "type": "image", + "image_path": "8a0956ed05cfa5a60368219f950f215715aef7518c15cd595eeb868649ed6c17.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 172, + 213, + 183 + ], + "lines": [ + { + "bbox": [ + 151, + 172, + 213, + 183 + ], + "spans": [ + { + "bbox": [ + 151, + 172, + 213, + 183 + ], + "type": "text", + "content": "CogVideoX [80]" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 157, + 70, + 206, + 118 + ], + "blocks": [ + { + "bbox": [ + 157, + 70, + 206, + 118 + ], + "lines": [ + { + "bbox": [ + 157, + 70, + 206, + 118 + ], + "spans": [ + { + "bbox": [ + 157, + 70, + 206, + 118 + ], + "type": "image", + "image_path": "1c45db49760c6149d82ab4ecd209aa54bd1800b3dff0d68875c03936b22f0976.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 157, + 121, + 206, + 169 + ], + "blocks": [ + { + "bbox": [ + 157, + 121, + 206, + 169 + ], + "lines": [ + { + "bbox": [ + 157, + 121, + 206, + 169 + ], + "spans": [ + { + "bbox": [ + 157, + 121, + 206, + 169 + ], + "type": "image", + "image_path": "ba17b7765c03118d83b021179e4364c54ee4d6c43bdc608f240d83faac7d94df.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 207, + 70, + 256, + 118 + ], + "blocks": [ + { + "bbox": [ + 207, + 70, + 256, + 118 + ], + "lines": [ + { + "bbox": [ + 207, + 70, + 256, + 118 + ], + "spans": [ + { + "bbox": [ + 207, + 70, + 256, + 118 + ], + "type": "image", + "image_path": "982b968d258ce237842b6c2aed398ee1e9b99bbed83d5b63cb2fdf4c38743576.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 207, + 121, + 256, + 169 + ], + "blocks": [ + { + "bbox": [ + 207, + 121, + 256, + 169 + ], + "lines": [ + { + "bbox": [ + 207, + 121, + 256, + 169 + ], + "spans": [ + { + "bbox": [ + 207, + 121, + 256, + 169 + ], + "type": "image", + "image_path": "466e437bb0013c49fa561d7086243876faa82ed619d69e13fa1ec92231875992.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 257, + 70, + 306, + 118 + ], + "blocks": [ + { + "bbox": [ + 257, + 70, + 306, + 118 + ], + "lines": [ + { + "bbox": [ + 257, + 70, + 306, + 118 + ], + "spans": [ + { + "bbox": [ + 257, + 70, + 306, + 118 + ], + "type": "image", + "image_path": "6e811a60c73167e59857e283b7f10cd3876fb91b26c46ddecfbc5a3920e4758f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 257, + 121, + 306, + 169 + ], + "blocks": [ + { + "bbox": [ + 257, + 121, + 306, + 169 + ], + "lines": [ + { + "bbox": [ + 257, + 121, + 306, + 169 + ], + "spans": [ + { + "bbox": [ + 257, + 121, + 306, + 169 + ], + "type": "image", + "image_path": "e46a43182df5084ddbede4cf1c2b5c4ad2fe4a5096d65521df97898a475c5811.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 173, + 353, + 182 + ], + "lines": [ + { + "bbox": [ + 307, + 173, + 353, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 353, + 182 + ], + "type": "text", + "content": "+Best-of-N" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 306, + 70, + 354, + 118 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 354, + 118 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 354, + 118 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 354, + 118 + ], + "type": "image", + "image_path": "6a385882fdc381c7082d2f3b668416fc7d25668713ab4e3bb095469d8ef6c57c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 306, + 121, + 370, + 169 + ], + "blocks": [ + { + "bbox": [ + 306, + 121, + 370, + 169 + ], + "lines": [ + { + "bbox": [ + 306, + 121, + 370, + 169 + ], + "spans": [ + { + "bbox": [ + 306, + 121, + 370, + 169 + ], + "type": "image", + "image_path": "e622220c3b33a3918c8189caf5ab11c5f132f3c3a1d48dd6f61b77b9f73a5187.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 355, + 173, + 403, + 182 + ], + "lines": [ + { + "bbox": [ + 355, + 173, + 403, + 182 + ], + "spans": [ + { + "bbox": [ + 355, + 173, + 403, + 182 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 355, + 70, + 403, + 118 + ], + "blocks": [ + { + "bbox": [ + 355, + 70, + 403, + 118 + ], + "lines": [ + { + "bbox": [ + 355, + 70, + 403, + 118 + ], + "spans": [ + { + "bbox": [ + 355, + 70, + 403, + 118 + ], + "type": "image", + "image_path": "6576cde83e0ade8241f1a4f8cd62198407309dfbda2c77f367ac54d1c4c00ef8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 355, + 121, + 403, + 169 + ], + "blocks": [ + { + "bbox": [ + 355, + 121, + 403, + 169 + ], + "lines": [ + { + "bbox": [ + 355, + 121, + 403, + 169 + ], + "spans": [ + { + "bbox": [ + 355, + 121, + 403, + 169 + ], + "type": "image", + "image_path": "4a9a9412f541315a38a8799109ed3ca16927e6768924b2971ed9f964958614ac.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 403, + 70, + 452, + 118 + ], + "blocks": [ + { + "bbox": [ + 403, + 70, + 452, + 118 + ], + "lines": [ + { + "bbox": [ + 403, + 70, + 452, + 118 + ], + "spans": [ + { + "bbox": [ + 403, + 70, + 452, + 118 + ], + "type": "image", + "image_path": "e8abb12535553004d9599928fe56f7558a09d656d56d42169f8faab183116de7.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 404, + 121, + 452, + 169 + ], + "blocks": [ + { + "bbox": [ + 404, + 121, + 452, + 169 + ], + "lines": [ + { + "bbox": [ + 404, + 121, + 452, + 169 + ], + "spans": [ + { + "bbox": [ + 404, + 121, + 452, + 169 + ], + "type": "image", + "image_path": "f8d22e0d534ce1a77e6c2e8f4c4d026fedbae33dc8bc6c695ac86b93e76825c7.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 404, + 173, + 452, + 182 + ], + "lines": [ + { + "bbox": [ + 404, + 173, + 452, + 182 + ], + "spans": [ + { + "bbox": [ + 404, + 173, + 452, + 182 + ], + "type": "text", + "content": "+" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 448, + 173, + 507, + 183 + ], + "lines": [ + { + "bbox": [ + 448, + 173, + 507, + 183 + ], + "spans": [ + { + "bbox": [ + 448, + 173, + 507, + 183 + ], + "type": "text", + "content": "+ SteerX (early)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 453, + 70, + 502, + 118 + ], + "blocks": [ + { + "bbox": [ + 453, + 70, + 502, + 118 + ], + "lines": [ + { + "bbox": [ + 453, + 70, + 502, + 118 + ], + "spans": [ + { + "bbox": [ + 453, + 70, + 502, + 118 + ], + "type": "image", + "image_path": "839473ecd6b6cfda4099eea2caf07a2f8bce7be7b83078b6b52d7b40509e1c5c.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 453, + 121, + 502, + 169 + ], + "blocks": [ + { + "bbox": [ + 453, + 121, + 502, + 169 + ], + "lines": [ + { + "bbox": [ + 453, + 121, + 502, + 169 + ], + "spans": [ + { + "bbox": [ + 453, + 121, + 502, + 169 + ], + "type": "image", + "image_path": "4c261f6a2d834073a74ac687ead7e57e2c93ccbfbcecf97abb2a2fc43406aa03.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 502, + 70, + 552, + 118 + ], + "blocks": [ + { + "bbox": [ + 502, + 70, + 552, + 118 + ], + "lines": [ + { + "bbox": [ + 502, + 70, + 552, + 118 + ], + "spans": [ + { + "bbox": [ + 502, + 70, + 552, + 118 + ], + "type": "image", + "image_path": "231b2ef285506b99170364dac01d2b99ec0fa50b5f71d4d564bd22b9f7d56a76.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 502, + 121, + 552, + 169 + ], + "blocks": [ + { + "bbox": [ + 502, + 121, + 552, + 169 + ], + "lines": [ + { + "bbox": [ + 502, + 121, + 552, + 169 + ], + "spans": [ + { + "bbox": [ + 502, + 121, + 552, + 169 + ], + "type": "image", + "image_path": "0e2142af8b404a66c8656c5084606171d3c56cbc6613060154c8b03c00e550a5.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 58, + 202, + 135, + 281 + ], + "blocks": [ + { + "bbox": [ + 76, + 186, + 532, + 198 + ], + "lines": [ + { + "bbox": [ + 76, + 186, + 532, + 198 + ], + "spans": [ + { + "bbox": [ + 76, + 186, + 532, + 198 + ], + "type": "text", + "content": "Figure 4. Qualitative results of video generation in VBench-I2V [32]. SteerX enhances vividness and overall visual quality." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 202, + 135, + 281 + ], + "lines": [ + { + "bbox": [ + 58, + 202, + 135, + 281 + ], + "spans": [ + { + "bbox": [ + 58, + 202, + 135, + 281 + ], + "type": "image", + "image_path": "e7d36517131aac71bef94157cf52bef874e6f92395918606317829051654fe82.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 284, + 116, + 294 + ], + "lines": [ + { + "bbox": [ + 74, + 284, + 116, + 294 + ], + "spans": [ + { + "bbox": [ + 74, + 284, + 116, + 294 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 152, + 202, + 211, + 284 + ], + "blocks": [ + { + "bbox": [ + 152, + 202, + 211, + 284 + ], + "lines": [ + { + "bbox": [ + 152, + 202, + 211, + 284 + ], + "spans": [ + { + "bbox": [ + 152, + 202, + 211, + 284 + ], + "type": "image", + "image_path": "cc064c04aba8a7f50b9c88ea584fdc1d50818ad364045be0c43d4be86bede913.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 193, + 284, + 234, + 294 + ], + "lines": [ + { + "bbox": [ + 193, + 284, + 234, + 294 + ], + "spans": [ + { + "bbox": [ + 193, + 284, + 234, + 294 + ], + "type": "text", + "content": "CogVideoX" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 297, + 553, + 309 + ], + "lines": [ + { + "bbox": [ + 55, + 297, + 553, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 553, + 309 + ], + "type": "text", + "content": "Figure 5. Qualitative results in Image-to-4D. SteerX naturally lifts object motion into 4D spaces, while preserving geometric alignments." + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 216, + 202, + 276, + 284 + ], + "blocks": [ + { + "bbox": [ + 216, + 202, + 276, + 284 + ], + "lines": [ + { + "bbox": [ + 216, + 202, + 276, + 284 + ], + "spans": [ + { + "bbox": [ + 216, + 202, + 276, + 284 + ], + "type": "image", + "image_path": "a3a62951dcc98557cf040428f95041591c2a93de2a2b08855aabe0c8385e6606.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 294, + 204, + 354, + 284 + ], + "blocks": [ + { + "bbox": [ + 294, + 204, + 354, + 284 + ], + "lines": [ + { + "bbox": [ + 294, + 204, + 354, + 284 + ], + "spans": [ + { + "bbox": [ + 294, + 204, + 354, + 284 + ], + "type": "image", + "image_path": "d13790f3bb9c0ccb395dca33c02ea81afeb7e1ccf33a99996d3313f7fff7c3be.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 285, + 374, + 293 + ], + "lines": [ + { + "bbox": [ + 332, + 285, + 374, + 293 + ], + "spans": [ + { + "bbox": [ + 332, + 285, + 374, + 293 + ], + "type": "text", + "content": "+Best-of-N" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 359, + 204, + 413, + 283 + ], + "blocks": [ + { + "bbox": [ + 359, + 204, + 413, + 283 + ], + "lines": [ + { + "bbox": [ + 359, + 204, + 413, + 283 + ], + "spans": [ + { + "bbox": [ + 359, + 204, + 413, + 283 + ], + "type": "image", + "image_path": "6d8e77e74264f9871ea0c0a79e110ef591fdb98280dcb72ab702e669d59949ec.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 432, + 202, + 492, + 281 + ], + "blocks": [ + { + "bbox": [ + 432, + 202, + 492, + 281 + ], + "lines": [ + { + "bbox": [ + 432, + 202, + 492, + 281 + ], + "spans": [ + { + "bbox": [ + 432, + 202, + 492, + 281 + ], + "type": "image", + "image_path": "2810a864dc46e4fa0fdf53a60ca6af54fde743458903575f60bf1590d3393bb0.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 478, + 284, + 511, + 293 + ], + "lines": [ + { + "bbox": [ + 478, + 284, + 511, + 293 + ], + "spans": [ + { + "bbox": [ + 478, + 284, + 511, + 293 + ], + "type": "text", + "content": "+SteerX" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 493, + 202, + 553, + 281 + ], + "blocks": [ + { + "bbox": [ + 493, + 202, + 553, + 281 + ], + "lines": [ + { + "bbox": [ + 493, + 202, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 493, + 202, + 553, + 281 + ], + "type": "image", + "image_path": "5afe9d78bdb1c380a7e1bed8d80cc2ec80967c497afb6d479b942a92c107534a.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 59, + 336, + 129, + 435 + ], + "blocks": [ + { + "bbox": [ + 66, + 319, + 286, + 334 + ], + "lines": [ + { + "bbox": [ + 66, + 319, + 286, + 334 + ], + "spans": [ + { + "bbox": [ + 66, + 319, + 286, + 334 + ], + "type": "text", + "content": "\"Filmed from a first-person perspective, the camera passes through the graffiti alley in Melbourne, Australia, where the graffiti walls are covered with artwork from many artists.\"" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 336, + 129, + 435 + ], + "lines": [ + { + "bbox": [ + 59, + 336, + 129, + 435 + ], + "spans": [ + { + "bbox": [ + 59, + 336, + 129, + 435 + ], + "type": "image", + "image_path": "55aa1f8a1453425da741d00f8419f0a10c20590fa60b0be403ef1ad8d5ee0937.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 437, + 117, + 446 + ], + "lines": [ + { + "bbox": [ + 70, + 437, + 117, + 446 + ], + "spans": [ + { + "bbox": [ + 70, + 437, + 117, + 446 + ], + "type": "text", + "content": "HunyuanVideo" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 94, + 452, + 257, + 463 + ], + "lines": [ + { + "bbox": [ + 94, + 452, + 257, + 463 + ], + "spans": [ + { + "bbox": [ + 94, + 452, + 257, + 463 + ], + "type": "text", + "content": "Figure 6. Qualitative results in Text-to-4D." + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 141, + 336, + 209, + 432 + ], + "blocks": [ + { + "bbox": [ + 141, + 336, + 209, + 432 + ], + "lines": [ + { + "bbox": [ + 141, + 336, + 209, + 432 + ], + "spans": [ + { + "bbox": [ + 141, + 336, + 209, + 432 + ], + "type": "image", + "image_path": "989a8e2ad5d2d6eccafae54599f5088f8e6ed865584572d9dd5d8cae42ed6137.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 437, + 187, + 445 + ], + "lines": [ + { + "bbox": [ + 164, + 437, + 187, + 445 + ], + "spans": [ + { + "bbox": [ + 164, + 437, + 187, + 445 + ], + "type": "text", + "content": "+BoN" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 225, + 336, + 288, + 434 + ], + "blocks": [ + { + "bbox": [ + 225, + 336, + 288, + 434 + ], + "lines": [ + { + "bbox": [ + 225, + 336, + 288, + 434 + ], + "spans": [ + { + "bbox": [ + 225, + 336, + 288, + 434 + ], + "type": "image", + "image_path": "a8a5463465a6f306c5212cccec1a788bd51c39cca942a43a505be904cc3354a5.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 246, + 437, + 274, + 445 + ], + "lines": [ + { + "bbox": [ + 246, + 437, + 274, + 445 + ], + "spans": [ + { + "bbox": [ + 246, + 437, + 274, + 445 + ], + "type": "text", + "content": "+SteerX" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "bbox": [ + 55, + 474, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 521 + ], + "type": "text", + "content": "as the diffusion model conceptualizes global structures. As a result, resampling at early timesteps coarsely aligns the geometry by tilting the data distribution, laying the foundation for a more precise geometric alignment in later stages." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 55, + 522, + 296, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 630 + ], + "type": "text", + "content": "The geometrically aligned video frames contribute to better 4D scene reconstruction, as shown in Fig. 6. While the baseline struggles to capture accurate camera poses and the best-of-N approach results in blurry 4D scenes, our approach precisely estimates camera poses and generates visually realistic 4D scenes. This highlights the effectiveness of SteerX in the seamless integration of video generation and scene reconstruction, ensuring that generated video frames are optimally structured for precise scene reconstruction." + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": "Image-to-4D Generation. Figure 4 and Table 2 show the results in image-conditioned video generation, and our SteerX performs the best for all metrics, and we observe that SteerX generates dynamic and realistic video frames, whereas the baseline produces almost motionless frames, and the best-of-N approach results in blurry frames. Also," + } + ] + } + ], + "index": 49 + }, + { + "type": "table", + "bbox": [ + 317, + 316, + 552, + 394 + ], + "blocks": [ + { + "bbox": [ + 317, + 316, + 552, + 394 + ], + "lines": [ + { + "bbox": [ + 317, + 316, + 552, + 394 + ], + "spans": [ + { + "bbox": [ + 317, + 316, + 552, + 394 + ], + "type": "table", + "html": "
MethodkAesthetic↑Subject↑Dynamic↑Dyn-MEt3R↑
CogVideoX [80]10.5920.9450.1580.880
+ BoN20.5910.9410.1410.882
+ SteerX (early)20.5930.9450.1610.894
+ BoN40.5940.9440.1430.901
+ SteerX (early)40.5960.9570.1700.909
", + "image_path": "8df505d2f84997743dbaec395125722ef8f25273dadb2631b164408a3bf627a5.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "table_body" + } + ], + "index": 50 + }, + { + "type": "table", + "bbox": [ + 317, + 426, + 552, + 491 + ], + "blocks": [ + { + "bbox": [ + 314, + 396, + 553, + 418 + ], + "lines": [ + { + "bbox": [ + 314, + 396, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 314, + 396, + 553, + 418 + ], + "type": "text", + "content": "Table 2. Quantitative results in VBench-I2V [32]. Results demonstrate the scalability of geometric steering as " + }, + { + "bbox": [ + 314, + 396, + 553, + 418 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 314, + 396, + 553, + 418 + ], + "type": "text", + "content": " increases." + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 426, + 552, + 491 + ], + "lines": [ + { + "bbox": [ + 317, + 426, + 552, + 491 + ], + "spans": [ + { + "bbox": [ + 317, + 426, + 552, + 491 + ], + "type": "table", + "html": "
MethodkRenderingSample
BRISQUE↓NIQE↓CLIP-I↑GS-MEt3R↑
DimensionX [59]137.34.2582.40.708
+ BoN429.84.3383.20.745
+ SteerX (early)429.74.2483.70.749
", + "image_path": "0c276462cff6ae5c76a7582cecd3ae0e211dc58b50545d77a7c77c5f2fb5e549.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "table_body" + } + ], + "index": 52 + }, + { + "bbox": [ + 313, + 493, + 553, + 516 + ], + "lines": [ + { + "bbox": [ + 313, + 493, + 553, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 553, + 516 + ], + "type": "text", + "content": "Table 3. Quantitative results in VBench-I2V [32]. For rendered images, we report results after the refinement process in [24, 37]." + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 529, + 555, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 555, + 589 + ], + "type": "text", + "content": "we validate the scalability of SteerX, as increasing the number of particles " + }, + { + "bbox": [ + 313, + 529, + 555, + 589 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 529, + 555, + 589 + ], + "type": "text", + "content": " consistently enhances performance across all metrics. As shown in Fig. 5, geometrically aligned video frames can be reconstructed to better 4D scenes, producing realistic dynamic objects and well-aligned backgrounds." + } + ] + } + ], + "index": 54 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "Image-to-3D Generation. We verify the effectiveness of SteerX in the Image-to-3D scene generation. While multiview images generated by SteerX closely resemble those from the best-of-N approach, SteerX reconstructs 3D scenes with finer details and fewer discontinuities in novel views, as shown in Fig. 8. Furthermore, quantitative results in Table 3 demonstrate that SteerX achieves better visual quality and geometric consistency in both rendered images and generated images compared to other baselines." + } + ] + } + ], + "index": 55 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27332" + } + ] + } + ], + "index": 56 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 547, + 264 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 547, + 264 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 547, + 264 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 547, + 264 + ], + "type": "image", + "image_path": "c3fd68343b530168c1e529c434e32f085f21e34acc9418d2f878a38dcd1df63d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 266, + 553, + 288 + ], + "lines": [ + { + "bbox": [ + 55, + 266, + 553, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 553, + 288 + ], + "type": "text", + "content": "Figure 7. Qualitative results of rendered images in T3Bench [28]. SteerX significantly enhances the visual quality of rendered images and textual alignment, demonstrating its compatibility with various video generation and scene reconstruction models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 61, + 296, + 128, + 427 + ], + "blocks": [ + { + "bbox": [ + 61, + 296, + 128, + 427 + ], + "lines": [ + { + "bbox": [ + 61, + 296, + 128, + 427 + ], + "spans": [ + { + "bbox": [ + 61, + 296, + 128, + 427 + ], + "type": "image", + "image_path": "4e2def90e8c2fcd5cad7440959e2b65a7e4683dfdf92e104dc5c4a9404925294.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 430, + 115, + 437 + ], + "lines": [ + { + "bbox": [ + 72, + 430, + 115, + 437 + ], + "spans": [ + { + "bbox": [ + 72, + 430, + 115, + 437 + ], + "type": "text", + "content": "(a) Input Image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 139, + 296, + 208, + 427 + ], + "blocks": [ + { + "bbox": [ + 139, + 296, + 208, + 427 + ], + "lines": [ + { + "bbox": [ + 139, + 296, + 208, + 427 + ], + "spans": [ + { + "bbox": [ + 139, + 296, + 208, + 427 + ], + "type": "image", + "image_path": "820bdb7269a762b7a3b1189562b133d970d4e4b222c07d5be8cf2530e1210fdb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 430, + 200, + 437 + ], + "lines": [ + { + "bbox": [ + 137, + 430, + 200, + 437 + ], + "spans": [ + { + "bbox": [ + 137, + 430, + 200, + 437 + ], + "type": "text", + "content": "(b) DimensionX + BoN" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 220, + 296, + 288, + 427 + ], + "blocks": [ + { + "bbox": [ + 220, + 296, + 288, + 427 + ], + "lines": [ + { + "bbox": [ + 220, + 296, + 288, + 427 + ], + "spans": [ + { + "bbox": [ + 220, + 296, + 288, + 427 + ], + "type": "image", + "image_path": "18469372ca97bdffd3ab3b6f87a17c59a6b587540ad8ac11aead7a0c7f5cb5dd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 430, + 283, + 437 + ], + "lines": [ + { + "bbox": [ + 214, + 430, + 283, + 437 + ], + "spans": [ + { + "bbox": [ + 214, + 430, + 283, + 437 + ], + "type": "text", + "content": "(c) DimensionX + SteerX" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 58, + 456, + 293, + 537 + ], + "blocks": [ + { + "bbox": [ + 59, + 441, + 291, + 453 + ], + "lines": [ + { + "bbox": [ + 59, + 441, + 291, + 453 + ], + "spans": [ + { + "bbox": [ + 59, + 441, + 291, + 453 + ], + "type": "text", + "content": "Figure 8. Qualitative results in of 3DGS in VBench-I2V [32]." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 456, + 293, + 537 + ], + "lines": [ + { + "bbox": [ + 58, + 456, + 293, + 537 + ], + "spans": [ + { + "bbox": [ + 58, + 456, + 293, + 537 + ], + "type": "table", + "html": "
MethodkBRISQUE↓CLIPScore↑GS-MEt3R↑
SplatFlow [24]133.2529.660.727
+ BoN236.8229.550.756
+ SteerX (early)227.4729.560.767
+ BoN428.1429.280.768
+ SteerX (early)426.6529.730.775
", + "image_path": "a15c32776e1ec7863ea312a20ccbffe00e8c8c87393e8cc28b217c8b05820d58.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "Text-to-3D Generation. Notice that SteerX can be applied to any generative model and scene reconstruction model, including the multi-view rectified flow model and the feed-forward 3DGS decoder in SplatFlow [24], enabling geometric steering through GS-MEt3R rewards. Table 4 presents quantitative results of generated multi-view images, demonstrating that SteerX effectively enhances overall performance and further improves scalability as the number of particles increases. Furthermore, as shown in Table 5, these 3D-aligned multi-view images are integrated with existing refinement processes, achieving state-of-the-art performance in 3D scene generation. Figure 7 presents qualita" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 317, + 295, + 552, + 441 + ], + "blocks": [ + { + "bbox": [ + 55, + 539, + 295, + 561 + ], + "lines": [ + { + "bbox": [ + 55, + 539, + 295, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 295, + 561 + ], + "type": "text", + "content": "Table 4. Quantitative results of multi-view generation in Single-Object-with-Surrounding set of T3Bench [28]." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 295, + 552, + 441 + ], + "lines": [ + { + "bbox": [ + 317, + 295, + 552, + 441 + ], + "spans": [ + { + "bbox": [ + 317, + 295, + 552, + 441 + ], + "type": "table", + "html": "
MethodBRISQUE↓NIQE↓CLIPScore↑
DreamFusion [53]90.210.48-
Magic3D [39]92.811.20-
LatentNeRF [45]88.69.19-
SJC [66]82.010.15-
Fantasia3D [11]69.67.65-
ProlificDreamer [70]61.57.07-
Director3D [37]32.34.3532.9
SplatFlow [24]19.64.2433.2
SplatFlow [24]†23.44.8432.7
+ BoN (k=4)17.24.4132.3
+ SteerX (k=4)13.14.3033.4
", + "image_path": "89755f25e0b4dc182d45c9a39a34b74d5e9542d922c2754241cbd023327ed289.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 442, + 553, + 477 + ], + "lines": [ + { + "bbox": [ + 313, + 442, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 553, + 477 + ], + "type": "text", + "content": "Table 5. Quantitative results in T3Bench [28]. SplatFlow† is re-implemented results without the stop ray strategy, which is incompatible with our geometric steering pipeline." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 487, + 553, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 553, + 512 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 553, + 512 + ], + "type": "text", + "content": "tive results, demonstrating that SteerX produces more text-aligned 3D scenes and visually realistic rendered images." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 525, + 388, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 388, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 388, + 537 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we have introduced SteerX, a zero-shot inference-time steering method for camera-free 3D and 4D scene generation. Instead of addressing physical alignment separately in either video generation or scene reconstruction, SteerX unifies both stages and iteratively tilts the data distribution toward geometrically consistent samples. Extensive experiments across diverse scene generation tasks verify that SteerX effectively enhances visual quality, textual alignment, and geometric consistency. SteerX is practical, enabling zero-shot 3D and 4D scene generation, and is scalable, with geometric alignment improving as the number of particles increases. We believe this efficient scaling property holds great potential, opening new avenues for 3D and 4D scene generation, particularly in test-time scaling." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27333" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Mohammad Asim, Christopher Wewer, Thomas Wimmer, Bernt Schiele, and Jan Eric Lenssen. Met3r: Measuring multi-view consistency in generated images. arXiv preprint arXiv:2501.06336, 2025. 2, 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "type": "text", + "content": "[2] Sherwin Bahmani, Ivan Skorokhodov, Guocheng Qian, Aliaksandr Siarohin, Willi Menapace, Andrea Tagliasacchi, David B Lindell, and Sergey Tulyakov. Ac3d: Analyzing and improving 3d camera control in video diffusion transformers. arXiv preprint arXiv:2411.18673, 2024. 1, 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 296, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 296, + 258 + ], + "type": "text", + "content": "[3] Sherwin Bahmani, Ivan Skorokhodov, Aliaksandr Siarohin, Willi Menapace, Guocheng Qian, Michael Vasilkovsky, Hsin-Ying Lee, Chaoyang Wang, Jiaxu Zou, Andrea Tagliasacchi, et al. Vd3d: Taming large video diffusion transformers for 3d camera control. arXiv preprint arXiv:2407.12781, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 261, + 296, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 296, + 325 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 296, + 325 + ], + "type": "text", + "content": "[4] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5855–5864, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 328, + 296, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 328, + 296, + 383 + ], + "spans": [ + { + "bbox": [ + 62, + 328, + 296, + 383 + ], + "type": "text", + "content": "[5] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 385, + 296, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 385, + 296, + 429 + ], + "spans": [ + { + "bbox": [ + 62, + 385, + 296, + 429 + ], + "type": "text", + "content": "[6] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. In The Twelfth International Conference on Learning Representations, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 430, + 296, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 430, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 62, + 430, + 296, + 485 + ], + "type": "text", + "content": "[7] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 487, + 296, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 487, + 296, + 541 + ], + "spans": [ + { + "bbox": [ + 62, + 487, + 296, + 541 + ], + "type": "text", + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 544, + 296, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 544, + 296, + 599 + ], + "spans": [ + { + "bbox": [ + 62, + 544, + 296, + 599 + ], + "type": "text", + "content": "[9] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 600, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 600, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 57, + 600, + 295, + 622 + ], + "type": "text", + "content": "[10] Guikun Chen and Wenguan Wang. A survey on 3d gaussian splatting. arXiv preprint arXiv:2401.03890, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 624, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 678 + ], + "type": "text", + "content": "[11] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22246–22256, 2023. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "text", + "content": "[12] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splattering from sparse" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 335, + 74, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 554, + 95 + ], + "type": "text", + "content": "multi-view images. In European Conference on Computer Vision, pages 370-386. Springer, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 97, + 554, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 554, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 554, + 140 + ], + "type": "text", + "content": "[13] Hyungjin Chung, Jeongsol Kim, Michael T McCann, Marc L Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. arXiv preprint arXiv:2209.14687, 2022. 2, 3, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 141, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 174 + ], + "type": "text", + "content": "[14] Hyungjin Chung, Suhyeon Lee, and Jong Chul Ye. Decomposed diffusion sampler for accelerating large-scale inverse problems. arXiv preprint arXiv:2303.05754, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 175, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 241 + ], + "type": "text", + "content": "[15] Wenyan Cong, Hanxue Liang, Peihao Wang, Zhiwen Fan, Tianlong Chen, Mukund Varma, Yi Wang, and Zhangyang Wang. Enhancing nerf akin to enhancing llms: Generalizable nerf transformer with mixture-of-view-experts. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3193-3204, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 243, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 243, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 243, + 553, + 297 + ], + "type": "text", + "content": "[16] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 299, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 299, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 317, + 299, + 553, + 352 + ], + "type": "text", + "content": "[17] Giannis Daras, Hyungjin Chung, Chieh-Hsin Lai, Yuki Mitsufuji, Jong Chul Ye, Peyman Milanfar, Alexandros G Dimakis, and Mauricio Delbracio. A survey on diffusion models for inverse problems. arXiv preprint arXiv:2410.00083, 2024. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 354, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 553, + 387 + ], + "type": "text", + "content": "[18] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 388, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 553, + 431 + ], + "type": "text", + "content": "[19] Zehao Dou and Yang Song. Diffusion posterior sampling for linear inverse problem solving: A filtering perspective. In The Twelfth International Conference on Learning Representations, 2024. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 434, + 553, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 434, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 317, + 434, + 553, + 466 + ], + "type": "text", + "content": "[20] Arnaud Doucet, Nando De Freitas, and Neil Gordon. An introduction to sequential monte carlo methods. Sequential Monte Carlo methods in practice, pages 3-14, 2001. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 468, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 553, + 499 + ], + "type": "text", + "content": "[21] Bradley Efron. Tweedie's formula and selection bias. Journal of the American Statistical Association, 106(496):1602-1614, 2011. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 502, + 553, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 502, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 317, + 502, + 553, + 545 + ], + "type": "text", + "content": "[22] Stephanie Fu, Mark Hamilton, Laura Brandt, Axel Feldman, Zhoutong Zhang, and William T Freeman. Fitup: A model-agnostic framework for features at any resolution. arXiv preprint arXiv:2403.10516, 2024. 3, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 547, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 547, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 317, + 547, + 553, + 601 + ], + "type": "text", + "content": "[23] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 603, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 603, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 317, + 603, + 553, + 646 + ], + "type": "text", + "content": "[24] Hyojun Go, Byeongjun Park, Jiho Jang, Jin-Young Kim, Soonwoo Kwon, and Changick Kim. Splatflow: Multiview rectified flow model for 3d gaussian splatting synthesis. arXiv preprint arXiv:2411.16443, 2024. 1, 2, 3, 6, 7, 8, 13" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 647, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 713 + ], + "type": "text", + "content": "[25] Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, Poriya Panet, Sapir Weissbuch, Victor Kulikov, Yaki Bitterman, Zeev Melumian, and Ofir Bibi. Ltx-video: Realtime video latent diffusion. arXiv preprint arXiv:2501.00103, 2024. 1" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27334" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[26] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible diffusion modeling of long videos. Advances in Neural Information Processing Systems, 35:27953-27965, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "text", + "content": "[27] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrl: Enabling camera control for text-to-video generation. arXiv preprint arXiv:2404.02101, 2024. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "text", + "content": "[28] Yuze He, Yushi Bai, Matthieu Lin, Wang Zhao, Yubin Hu, Jenny Sheng, Ran Yi, Juanzi Li, and Yong-Jin Liu. T" + }, + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "text", + "content": "bench: Benchmarking current progress in text-to-3d generation, 2023. 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 251 + ], + "type": "text", + "content": "[29] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 286 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 286 + ], + "type": "text", + "content": "[30] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 288, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 288, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 288, + 294, + 331 + ], + "type": "text", + "content": "[31] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 333, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 294, + 398 + ], + "type": "text", + "content": "[32] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 442 + ], + "type": "text", + "content": "[33] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 499 + ], + "type": "text", + "content": "[34] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "type": "text", + "content": "[35] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. 1, 2, 6, 13, 14" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 557, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 621 + ], + "type": "text", + "content": "[36] Xiner Li, Yulai Zhao, Chenyu Wang, Gabriele Scalia, Gokcen Eraslan, Surag Nair, Tommaso Biancalani, Shuiwang Ji, Aviv Regev, Sergey Levine, et al. Derivative-free guidance in continuous and discrete diffusion models with soft value-based decoding. arXiv preprint arXiv:2408.08252, 2024. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "type": "text", + "content": "[37] Xinyang Li, Zhangyu Lai, Linning Xu, Yansong Qu, Liujuan Cao, Shengchuan Zhang, Bo Dai, and Rongrong Ji. Director3d: Real-world camera trajectory and 3d scene generation from text. Advances in Neural Information Processing Systems, 37:75125-75151, 2025. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[38] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 161 + ], + "type": "text", + "content": "[39] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 300–309, 2023. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 162, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 227 + ], + "type": "text", + "content": "[40] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 228, + 553, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 292 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 292 + ], + "type": "text", + "content": "[41] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14458-14467, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 293, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 293, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 293, + 553, + 337 + ], + "type": "text", + "content": "[42] Fangfu Liu, Wenqiang Sun, Hanyang Wang, Yikai Wang, Haowen Sun, Junliang Ye, Jun Zhang, and Yueqi Duan. Reconx: Reconstruct any scene from sparse views with video diffusion model. arXiv preprint arXiv:2408.16767, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "text", + "content": "[43] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "type": "text", + "content": "[44] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "type": "text", + "content": "[45] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 482, + 553, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 534 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 534 + ], + "type": "text", + "content": "[46] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 536, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 536, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 536, + 553, + 581 + ], + "type": "text", + "content": "[47] Byeongjun Park and Changick Kim. Point-dynrf: Point-based dynamic radiance fields from a monocular video. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3171-3181, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 582, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 613 + ], + "type": "text", + "content": "[48] Byeongjun Park, Sangmin Woo, Hyojun Go, Jin-Young Kim, and Changick Kim. Denoising task routing for diffusion models. arXiv preprint arXiv:2310.07138, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "type": "text", + "content": "[49] Byeongjun Park, Hyojun Go, Jin-Young Kim, Sangmin Woo, Seokil Ham, and Changick Kim. Switch diffusion transformer: Synergizing denoising tasks with sparse mixture-of-experts. arXiv preprint arXiv:2403.09176, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[50] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5865-5874, 2021. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27335" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[51] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 130, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 294, + 173 + ], + "type": "text", + "content": "[52] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195-4205, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 176, + 294, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 176, + 294, + 209 + ], + "spans": [ + { + "bbox": [ + 56, + 176, + 294, + 209 + ], + "type": "text", + "content": "[53] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 211, + 294, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 211, + 294, + 276 + ], + "spans": [ + { + "bbox": [ + 56, + 211, + 294, + 276 + ], + "type": "text", + "content": "[54] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10901-10911, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 279, + 294, + 333 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 279, + 294, + 333 + ], + "spans": [ + { + "bbox": [ + 56, + 279, + 294, + 333 + ], + "type": "text", + "content": "[55] Raghav Singhal, Zachary Horvitz, Ryan Teehan, Mengye Ren, Zhou Yu, Kathleen McKeown, and Rajesh Ranganath. A general framework for inference-time scaling and steering of diffusion models. arXiv preprint arXiv:2501.06848, 2025. 2, 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 335, + 294, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 335, + 294, + 379 + ], + "spans": [ + { + "bbox": [ + 56, + 335, + 294, + 379 + ], + "type": "text", + "content": "[56] Brandon Smart, Chuanxia Zheng, Iro Laina, and Victor Adrian Prisacariu. Splatt3r: Zero-shot gaussian splatting from uncalibrated image pairs. arXiv preprint arXiv:2408.13912, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 382, + 294, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 294, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 294, + 425 + ], + "type": "text", + "content": "[57] Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. Pseudoinverse-guided diffusion models for inverse problems. In International Conference on Learning Representations, 2023. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 428, + 294, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 428, + 294, + 472 + ], + "spans": [ + { + "bbox": [ + 56, + 428, + 294, + 472 + ], + "type": "text", + "content": "[58] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 474, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 474, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 474, + 294, + 518 + ], + "type": "text", + "content": "[59] Wenqiang Sun, Shuo Chen, Fangfu Liu, Zilong Chen, Yueqi Duan, Jun Zhang, and Yikai Wang. Dimensionx: Create any 3d and 4d scenes from a single image with controllable video diffusion. arXiv preprint arXiv:2411.04928, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 520, + 294, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 294, + 573 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 294, + 573 + ], + "type": "text", + "content": "[60] Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10208-10217, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 577, + 294, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 577, + 294, + 631 + ], + "spans": [ + { + "bbox": [ + 56, + 577, + 294, + 631 + ], + "type": "text", + "content": "[61] Zhenggang Tang, Yuchen Fan, Dilin Wang, Hongyu Xu, Rakesh Ranjan, Alexander Schwing, and Zhicheng Yan. Mvdust3r+: Single-stage scene reconstruction from sparse views in 2 seconds. arXiv preprint arXiv:2412.06974, 2024. 1, 2, 4, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 656 + ], + "type": "text", + "content": "[62] Genmo Team. Mochi 1. https://github.com/genmoai/models, 2024. 1, 2, 6, 13, 14" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[63] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Wang Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. In Computer Graphics Forum, pages 703-735. Wiley Online Library, 2022. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "text", + "content": "[64] Masatoshi Uehara, Xingyu Su, Yulai Zhao, Xiner Li, Aviv Regev, Shuiwang Ji, Sergey Levine, and Tommaso Biancalani. Reward-guided iterative refinement in diffusion models at test-time with applications to protein and dna design. arXiv preprint arXiv:2502.14944, 2025. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 162 + ], + "type": "text", + "content": "[65] Diwen Wan, Ruijie Lu, and Gang Zeng. Superpoint gaussian splatting for real-time high-fidelity dynamic scene reconstruction. arXiv preprint arXiv:2406.03697, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 164, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 553, + 217 + ], + "type": "text", + "content": "[66] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "text", + "content": "[67] Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 265, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 265, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 265, + 553, + 319 + ], + "type": "text", + "content": "[68] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 321, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 374 + ], + "type": "text", + "content": "[69] Yunsong Wang, Tianxin Huang, Hanlin Chen, and Gim Hee Lee. Freesplat: Generalizable 3d gaussian splatting towards free view synthesis of indoor scenes. Advances in Neural Information Processing Systems, 37:107326-107349, 2025. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 377, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 553, + 431 + ], + "type": "text", + "content": "[70] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 434, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 488 + ], + "type": "text", + "content": "[71] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 491, + 555, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 491, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 491, + 555, + 544 + ], + "type": "text", + "content": "[72] Sangmin Woo, Byeongjun Park, Hyojun Go, Jin-Young Kim, and Changick Kim. Harmonyview: Harmonizing consistency and diversity in one-image-to-3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10574–10584, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 546, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 546, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 546, + 553, + 610 + ], + "type": "text", + "content": "[73] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20310-20320, 2024. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 613, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 666 + ], + "type": "text", + "content": "[74] Luhuan Wu, Brian Trippe, Christian Naesseth, David Blei, and John P Cunningham. Practical and asymptotically exact conditional sampling in diffusion models. Advances in Neural Information Processing Systems, 36:31372-31403, 2023. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[75] Rundi Wu, Ruiqi Gao, Ben Poole, Alex Trevithick, Changxi Zheng, Jonathan T Barron, and Aleksander Holynski. Cat4d: Create anything in 4d with multi-view video diffusion models. arXiv preprint arXiv:2411.18613, 2024. 1, 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27336" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "text", + "content": "[76] Tong Wu, Yu-Jie Yuan, Ling-Xiao Zhang, Jie Yang, Yan-Pei Cao, Ling-Qi Yan, and Lin Gao. Recent advances in 3d gaussian splatting. Computational Visual Media, 10(4):613-642, 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 182 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 182 + ], + "type": "text", + "content": "[77] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: learning and evaluating human preferences for text-to-image generation. In Proceedings of the 37th International Conference on Neural Information Processing Systems, pages 15903-15935, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 295, + 239 + ], + "type": "text", + "content": "[78] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "type": "text", + "content": "[79] Xiaofeng Yang, Yiwen Chen, Cheng Chen, Chi Zhang, Yi Xu, Xulei Yang, Fayao Liu, and Guosheng Lin. Learn to optimize denoising scores for 3d generation: A unified and improved diffusion prior on nerf and 3d gaussian splatting. arXiv preprint arXiv:2312.04820, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 294, + 294, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 348 + ], + "type": "text", + "content": "[80] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 2, 6, 7, 14" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "type": "text", + "content": "[81] Botao Ye, Sifei Liu, Haofei Xu, Xueting Li, Marc Pollefeys, Ming-Hsuan Yang, and Songyou Peng. No pose, no problem: Surprisingly simple 3d gaussian splats from sparse unposed images. arXiv preprint arXiv:2410.24207, 2024. 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 394, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 294, + 437 + ], + "type": "text", + "content": "[82] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-22, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 438, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 438, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 438, + 294, + 491 + ], + "type": "text", + "content": "[83] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 493, + 294, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 558 + ], + "type": "text", + "content": "[84] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Chenming Zhu, Zhangyang Xiong, Tianyou Liang, et al. Mvmgnet: A large-scale dataset of multi-view images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9150-9161, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 559, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 612 + ], + "type": "text", + "content": "[85] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19447-19456, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 614, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 668 + ], + "type": "text", + "content": "[86] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arXiv:2410.03825, 2024. 1, 2, 4, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[87] Yuyang Zhao, Chung-Ching Lin, Kevin Lin, Zhiwen Yan, Linjie Li, Zhengyuan Yang, Jianfeng Wang, Gim Hee Lee, and Lijuan Wang. Genxd: Generating any 3d and 4d scenes. arXiv preprint arXiv:2411.02319, 2024. 1, 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 205 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "type": "text", + "content": "[88] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "type": "text", + "content": "[89] Junzhe Zhu, Peiye Zhuang, and Sanmi Koyejo. Hifa: High-fidelity text-to-3d generation with advanced diffusion guidance. arXiv preprint arXiv:2305.18766, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "type": "text", + "content": "[90] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "27337" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Steering Guidance for Personalized Text-to-Image Diffusion Models/34e2c957-5535-42df-a5b2-c4f7cd9d6450_content_list.json b/2025/Steering Guidance for Personalized Text-to-Image Diffusion Models/34e2c957-5535-42df-a5b2-c4f7cd9d6450_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..30d829690b5685bb09f553c758a002bb09df601e --- /dev/null +++ b/2025/Steering Guidance for Personalized Text-to-Image Diffusion Models/34e2c957-5535-42df-a5b2-c4f7cd9d6450_content_list.json @@ -0,0 +1,1614 @@ +[ + { + "type": "text", + "text": "Steering Guidance for Personalized Text-to-Image Diffusion Models", + "text_level": 1, + "bbox": [ + 156, + 130, + 841, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sunghyun Park* Seokeon Choi* Hyoungwoo Park Sungrack Yun Qualcomm AI Research†", + "bbox": [ + 217, + 180, + 779, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{sunpar, seokchoi, hwoopark, sungrack}@qti=qualcomm.com", + "bbox": [ + 279, + 220, + 712, + 233 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/23db2504435ec8c5e07cd7363b0933debf27211c658fddc7dd6f63cf5b626403.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 253, + 187, + 391 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/aac4c5beaae87aa768d036c94516d336fc2d62c597bac7c81b2e0e1609698162.jpg", + "image_caption": [ + "A $\\star$ cat in a chef outfit" + ], + "image_footnote": [], + "bbox": [ + 191, + 253, + 367, + 391 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7dae5fcbd0bec05bbf8130b6cdf42f0ae06f049200608ce42fec42b4a0914837.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 253, + 544, + 391 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fff0b5ad4afed71db9bbc380374de9d6e20017c9f9415c4f38857ca605120bf9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 545, + 253, + 720, + 391 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9cedbbe1c3a7dcf4835071f82f91fd2a5982cf516be0380ff5f4547ca7b610ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 722, + 253, + 898, + 391 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ed49c68166a3371a73a1db68dd3ff105df920ff28d5cb82ea5512f78f55581cb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 405, + 187, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/119fa937dfe5f87617e53075cd78d8963591897ca23c47ced8268c54d6ca7e1a.jpg", + "image_caption": [ + "A $\\star$ dog in a movie theater" + ], + "image_footnote": [], + "bbox": [ + 189, + 405, + 544, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8ca313a3fd2efad70cddccc58943cef2fa6ef9d821e56fad310a1c768d276ac6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 545, + 405, + 720, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6e5a1016bb96b56f1c7cc5f3ed8116e58617475e19b6ad540b4ca961da886e05.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 722, + 405, + 898, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cbfdfbd2181e8f59b59228a03aeef53ec85be25f8e84eb32496cf1989ff1085d.jpg", + "image_caption": [ + "Reference", + "Figure 1. Comparison of generated image quality between Classifier-Free Guidance (CFG) and our Personalization Guidance." + ], + "image_footnote": [], + "bbox": [ + 96, + 556, + 187, + 695 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f71a04de42d6201e407401c0d6691d3fc6469428b7bfff3acf5b634f8ce5c51d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 189, + 556, + 367, + 694 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8a5befc83390fb14d03018c0f1bf0a55ec5e9d05210da6a831e121b4fb100df9.jpg", + "image_caption": [ + "A duck toy with the Eiffel Towel in the background" + ], + "image_footnote": [], + "bbox": [ + 367, + 556, + 542, + 695 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/52eec6116cbdea1d926e6c41cf7403a25d3e8edbc24a692b9e55ee3d3c16c390.jpg", + "image_caption": [ + "A duck toy among the skyscrapers in New York city" + ], + "image_footnote": [], + "bbox": [ + 544, + 556, + 720, + 695 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/74ef1e47cef28154d4c886aa7597eefa719541e434c44e5551438f4888b8414e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 722, + 556, + 898, + 695 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 752, + 326, + 767 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Personalizing text-to-image diffusion models is crucial for adapting the pre-trained models to specific target concepts, enabling diverse image generation. However, fine-tuning with few images introduces an inherent trade-off between aligning with the target distribution (e.g., subject fidelity)", + "bbox": [ + 89, + 785, + 482, + 861 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and preserving the broad knowledge of the original model (e.g., text editability). Existing sampling guidance methods, such as classifier-free guidance (CFG) and autoguidance (AG), fail to effectively guide the output toward well-balanced space: CFG restricts the adaptation to the target distribution, while AG compromises text alignment. To address these limitations, we propose personalization guidance, a simple yet effective method leveraging an unlearned weak model conditioned on a null text prompt. Moreover, our method dynamically controls the extent of unlearning", + "bbox": [ + 511, + 738, + 906, + 891 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These two authors contributed equally to this work.", + "bbox": [ + 89, + 875, + 372, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Qualcomm AI Research is an initiative of Qualcomm Technologies, Inc.", + "bbox": [ + 91, + 887, + 478, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15907", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in a weak model through weight interpolation between pretrained and fine-tuned models during inference. Unlike existing guidance methods, which depend solely on guidance scales, our method explicitly steers the outputs toward a balanced latent space without additional computational overhead. Experimental results demonstrate that our proposed guidance can improve text alignment and target distribution fidelity, integrating seamlessly with various fine-tuning strategies.", + "bbox": [ + 89, + 90, + 483, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 256, + 222, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models [11, 32] have emerged as powerful generative models capable of representing complex data distributions, driving significant advancements in high-fidelity image generation [24, 26, 29, 38]. However, real-world use cases often demand personalized generation, where the model generates not only generic concepts but also specific concepts like a novel character or person [8, 20, 27]. This personalization technique is particularly useful for applications such as portrait editing and specific content creation.", + "bbox": [ + 89, + 280, + 482, + 416 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Many approaches have been proposed to personalize diffusion models for novel concepts. One line of work [6, 28, 30, 34] learns a general-purpose encoder capable of adapting to new concepts without test-time finetuning. However, this approach requires large-scale text-image datasets and considerable computational resources, and must often be retrained from scratch for each pre-trained model, making it expensive to adapt to the growing variety of text-to-image or text-to-video diffusion architectures. An alternative solution is to directly fine-tune a pre-trained diffusion model on the target images, which can be roughly categorized into fine-tuning low-rank adapters [5, 13], text embeddings [8], specific layers [14, 20], or the full model parameters [27]. This strategy can be applied to any pre-trained diffusion model and only requires a few target images, proving effective for personalization in various scenarios.", + "bbox": [ + 89, + 417, + 482, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite these advancements, optimization-based personalization remains a nontrivial challenge. When fine-tuning a model to a specific, often much smaller, target data distribution, there is an inevitable tradeoff between aligning with the target distribution (e.g., learning a novel concept with a few target images) and preserving broad knowledge of the original model (e.g., preserving text editability). This tradeoff manifests in various ways: a fine-tuned model for personalization may generate high-quality domain-specific images but lose text editability, making it difficult to generalize across different prompts. Conversely, a model that retains broad knowledge from its original training distribution may struggle to fully align with new concepts introduced in fine-tuning. Existing approaches attempt to mitigate this tradeoff by adjusting the number of fine-tuning iterations [8, 27], leveraging regularization [14, 27], reducing the model pa", + "bbox": [ + 89, + 659, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ae36c689934c7060d99a5fbf9318cdb4c55f44b49ddd1b30e6247d248d0fe863.jpg", + "image_caption": [ + "Figure 2. Motivation of Personalization Guidance. Classifier-free guidance [10] can interfere with subject fidelity due to its alignment with the text prompt. In contrast, autoguidance [18] fails to reflect the text prompt but generates the subject more similarly. Our guidance complements the limitations of both methods." + ], + "image_footnote": [], + "bbox": [ + 519, + 90, + 602, + 190 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bb9b851d5c3a3ac9ea1411833aa6e6d8639d714b23a8cf5d3b886b22841cdb78.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 90, + 702, + 191 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6d21b32247e247b671aa7ca9ce6b146e58087326080219977d753112269c9cfe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 90, + 802, + 191 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1f73207d1918fa94259aa2f031677f7275b0051240112abeb37a68f3021e8ab7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 802, + 90, + 903, + 191 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "rameters of updates [5, 13], yet they primarily operate at the fine-tuning stage. Moreover, balancing the trade-off between comprehensive text editability and precise alignment with a new target distribution remains an open problem.", + "bbox": [ + 511, + 275, + 903, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A parallel line of research focuses on sampling guidance techniques, such as Classifier-Free Guidance (CFG) [10] and AutoGuidance (AG) [18]. These methods rely on \"weak models\" to detect and push the outputs away from poor trajectories, thereby guiding samples toward higher-quality regions of the data manifold. For instance, CFG uses an unconditional model using the null text prompt as the weak model, while AG uses an unlearned model—one with fewer parameters or less training—conditioned on the same text prompt. However, these methods have limitations when applied to personalized diffusion models, as shown in Fig. 2. CFG can interfere with the model's adaptation to the target distribution, restricting its ability to fully utilize fine-tuned knowledge. Moreover, while the guidance scale of CFG adjusts the degree of text fidelity, it does not improve alignment with the target data. On the other hand, AG, which primarily enhances alignment with the target distribution, however, lacks an explicit mechanism for improving text alignment, making it unsuitable for fine-tuned text-to-image models.", + "bbox": [ + 511, + 338, + 906, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these issues, we propose a simple yet effective guidance technique, called Personalization Guidance, for personalized diffusion models, which can dynamically balance text alignment and adaptation to the target distribution. Our method leverages an \"unlearned\" weak model with a null text prompt, without incurring additional computational overhead during inference. Here, we aim to balance the alignment with the target distribution (e.g., subject fidelity) and the preservation of the broad knowledge of the original model (e.g., text fidelity) by controlling the extent of \"unlearning\" in a weak model. Here, the degree of unlearning refers to the extent to which the target data distribution has not been learned compared to the fine-tuned model. The rationale behind adjusting this degree of unlearning is to steer the output toward an optimal, well-balanced latent space. Specifically, during inference, we interpolate the weight parameters between the pre-trained and fine-tuned", + "bbox": [ + 511, + 643, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15908", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "diffusion models to derive an optimal weak model, effectively controlling the extent of unlearning. Unlike guidance scales, which adjust the intensity of guidance, weight interpolation for the weak model serves to steer outputs in an appropriate direction. This straightforward approach can be readily integrated into various fine-tuning diffusion models.", + "bbox": [ + 89, + 90, + 483, + 183 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 195, + 233, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Personalization of Diffusion Models. Personalizing diffusion models aim to generate novel images by adapting to specific concepts using a set of reference images. Optimization-based methods personalize models by finetuning textual embeddings [8, 14, 20], entire model parameters with regularization [27], or low-rank adapters [5, 13]. Another line of research leverages extensive text-image datasets to train dedicated text-to-image personalization models, by fine-tuning additional modules [6, 28, 30, 34]. Recently, several studies [4, 22, 31] have achieved better personalization without additional fine-tuning or training. Orthogonal to these existing approaches, we introduce a simple yet effective personalization guidance technique that seamlessly integrates with various personalization methods by modifying the classifier-free guidance. Moreover, our guidance enables us to adjust the degree of subject and text fidelity without incurring additional computational costs.", + "bbox": [ + 91, + 220, + 483, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Guidance for Diffusion Models. Classifier-Free Guidance (CFG) [10] enhances conditional image generation by leveraging the unconditional model as a weak model to guide the outputs toward better adherence to the given condition. Various guidance techniques have been explored, such as the guidance using perturbations [1, 16] or blurring to the attention maps [12]. Autoguidance [18] improves generation by using a weak model—one with fewer parameters or less training—to push outputs toward a well-trained direction. While InstructPix2Pix [2] introduces separate CFG using both image and text conditions, it cannot be directly applied to DB-LoRA due to its reliance on image-conditioned training. For personalization, subject-agnostic guidance [4] enhances subject-agnostic attributes by employing a weak model conditioned on subject-specific text prompts. In this study, we aim to address the fundamental personalization trade-off—aligning a model with a smaller target data distribution while preserving the broad text fidelity of the original model. We derive a weak model that dynamically adjusts the degree of adaptation, balancing comprehensive text editability with precise alignment to new concepts.", + "bbox": [ + 91, + 478, + 483, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Weight Interpolation. There are previous studies on exploring the use of weight interpolation. WISE-FT [36] improves the robustness of fine-tuned classifiers against distribution shifts through weight-space ensembling. Model-soups [35] maximizes model accuracy and robustness by interpolating the weights of multiple fine-tuned classifiers trained with different hyperparameter configurations.", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 89, + 604, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 513, + 114, + 653, + 128 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Models. Diffusion models [11, 32] aim to learn reversing a forward noising process that gradually transforms data samples into noise. Concretely, the forward process starts with a data distribution $p_0(\\mathbf{x})$ and ends at a high-noise distribution $p_T(\\mathbf{x}) \\approx \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ . Sampling from the data distribution proceeds by solving the equivalent probability-flow ordinary differential equation (PF-ODE). Under the variance-preserving parameterization where $p(\\mathbf{x}_t|\\mathbf{x}_0) = \\mathcal{N}(\\mathbf{x}_0,\\sigma_t^2\\mathbf{I})$ , the PF-ODE takes the form:", + "bbox": [ + 511, + 135, + 906, + 286 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {d} \\mathbf {x} _ {t} = - \\sigma_ {t} \\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {x} _ {t}) \\mathrm {d} t, \\quad \\mathbf {x} _ {T} \\sim p _ {T} (\\mathbf {x} _ {T}). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 297, + 906, + 314 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In diffusion models, the score function is parameterized by $\\epsilon_{\\theta}$ . Here, the noise estimation of the diffusion model can be considered as $\\epsilon_{\\theta}(\\mathbf{x}_t) \\approx -\\sigma_t \\nabla_{\\mathbf{x}_t} \\log p(\\mathbf{x}_t)$ .", + "bbox": [ + 511, + 321, + 906, + 368 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Personalization. Suppose we have a pre-trained diffusion model $\\epsilon_{\\theta}$ , where $\\theta$ is the parameters of the model. We wish to personalize it to a smaller target dataset $\\mathcal{D}_{\\mathrm{target}} = \\left\\{(\\mathbf{x}^i, \\mathbf{c}^i)\\right\\}_{i=1}^N$ , where each $\\mathbf{x}^i$ and $\\mathbf{c}^i$ is an image and its associated text prompt, respectively. The goal is to obtain new parameters $\\theta'$ so that the fine-tuned model $\\epsilon_{\\theta'}$ can generate samples more faithfully matched the $\\mathcal{D}_{\\mathrm{target}}$ .", + "bbox": [ + 511, + 368, + 905, + 473 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fine-tuning typically follows the same denoising objective used by the original diffusion model [11]. Concretely, let $\\sigma$ be a noise level sampled from a training schedule, and let $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\sigma^2\\mathbf{I})$ . For each $(\\mathbf{x},\\mathbf{c})$ in $\\mathcal{D}_{\\mathrm{target}}$ , we form the noisy input $\\mathbf{x}_t = \\mathbf{x} + \\epsilon$ and minimize:", + "bbox": [ + 511, + 474, + 905, + 549 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} _ {(\\mathbf {x}, \\mathbf {c}) \\sim \\mathcal {D} _ {\\text {t a r g e t}}, \\sigma , \\boldsymbol {\\epsilon}} \\left\\| \\epsilon_ {\\theta^ {\\prime}} (\\mathbf {x} _ {t}, \\sigma , \\mathbf {c}) - \\boldsymbol {\\epsilon} \\right\\| _ {2} ^ {2}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 556, + 906, + 587 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This teaches the model to reconstruct $\\mathbf{x}$ from noisy inputs in the domain of $\\mathcal{D}_{\\mathrm{target}}$ , thereby adapting it to the new concept. In practice, additional regularization terms are often used to prevent overfitting [14, 27].", + "bbox": [ + 511, + 594, + 905, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Classifier-Free Guidance. To enhance the generation towards an arbitrary condition $\\mathbf{c}$ , classifier-free guidance (CFG) [10] introduces a new sampling distribution $p^{\\lambda}(\\mathbf{x}|\\mathbf{c})$ composed with both $p(\\mathbf{x})$ and the classifier distribution $p(c|\\mathbf{x})^{\\lambda}$ , which is described as:", + "bbox": [ + 511, + 654, + 905, + 731 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np ^ {\\lambda} (\\mathbf {x} | \\mathbf {c}) \\propto p (\\mathbf {x}) p (\\mathbf {c} | \\mathbf {x}) ^ {\\lambda}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 738, + 906, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\lambda > 1$ amplifies the conditional likelihood. By applying Bayes' rule for some timestep $t$ , the corresponding score function is:", + "bbox": [ + 511, + 765, + 905, + 810 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {\\mathbf {x} _ {t}} \\log p ^ {\\lambda} (\\mathbf {x} _ {t} | \\mathbf {c}) = \\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {x} _ {t}) \\\\ + \\lambda \\left(\\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {x} _ {t} | \\mathbf {c}) - \\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {x} _ {t})\\right). \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 823, + 903, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enable conditional generation such as text prompts and class labels, CFG jointly trains the unconditional model", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15909", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f9f9e426c6ceb3e0cc05f6b58e917c2a3355395365832500f44d8242372c3b0a.jpg", + "image_caption": [ + "Figure 3. Comparison between Classifier-Free Guidance (CFG), AutoGuidance (AG), and our Personalization Guidance (PG), with the band conceptually representing the noisy data manifold. By leveraging weight interpolation, guidance can be steered toward the optimal space between CFG and PG." + ], + "image_footnote": [], + "bbox": [ + 106, + 88, + 475, + 319 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\epsilon_{\\theta}(\\mathbf{x}_t,\\phi)$ and the conditional model $\\epsilon_{\\theta}(\\mathbf{x}_t,\\mathbf{c})$ within a single model by dropping $\\mathbf{c}$ with certain probability to allow null conditioning with $c = \\phi$ . In the personalization setting, parameterizing the score function with the fine-tuned diffusion model $\\epsilon_{\\theta'}$ is described as:", + "bbox": [ + 89, + 407, + 483, + 482 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\epsilon} _ {\\theta^ {\\prime}} ^ {\\lambda} \\left(\\mathbf {x} _ {t} \\mid \\mathbf {c}\\right) = \\epsilon_ {\\theta^ {\\prime}} \\left(\\mathbf {x} _ {t} \\mid \\phi\\right) + \\lambda \\left(\\epsilon_ {\\theta^ {\\prime}} \\left(\\mathbf {x} _ {t} \\mid \\mathbf {c}\\right) - \\epsilon_ {\\theta^ {\\prime}} \\left(\\mathbf {x} _ {t} \\mid \\phi\\right)\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 489, + 482, + 510 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, while CFG improves alignment with the conditions, such as the text prompt, it struggles to enhance adaptation to the target data in the personalization task.", + "bbox": [ + 89, + 517, + 483, + 563 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "AutoGuidance. AutoGuidance (AG) [18] is another guidance technique that employs a weaker version of the same model to identify suboptimal generation trajectories. Rather than using the unconditional model $\\epsilon_{\\theta}(\\mathbf{x}_t|\\phi)$ , AG directly guides a model with the weak model $\\epsilon_{\\theta_{weak}}$ , which is trained on the same task and conditioning, but degraded by under-training or less-parameters. However, obtaining a weaker version $\\epsilon_{\\theta_{weak}}$ of a fully trained diffusion model like Stable Diffusion [26] is challenging in practice. Instead, we found that a less trained model could serve as the pre-trained model for personalization, where $\\theta_{weak} = \\theta$ , effectively shifts the output away in the direction of fine-tuning, enhancing the adaptation to the target data. In the personalization, AG is formulated as:", + "bbox": [ + 89, + 564, + 483, + 775 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\epsilon} _ {\\theta^ {\\prime}} ^ {\\lambda} \\left(\\mathbf {x} _ {t} | \\mathbf {c}\\right) = \\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t} | \\mathbf {c}\\right) + \\lambda \\left(\\epsilon_ {\\theta^ {\\prime}} \\left(\\mathbf {x} _ {t} | \\mathbf {c}\\right) - \\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t} | \\mathbf {c}\\right)\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 782, + 482, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, this approach has a major drawback—it scarcely reflects the text prompt, as shown in Fig. 2.", + "bbox": [ + 89, + 809, + 482, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Steering Personalization Guidance", + "text_level": 1, + "bbox": [ + 89, + 848, + 393, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by autoguidance [18], we simply modify the CFG [10], by leveraging a pre-trained diffusion model $\\epsilon_{\\theta}$", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c12daafe985ba09788877b81777b62340a8302bd27ec89c843108668947cdf82.jpg", + "image_caption": [ + "Subject Fidelity (DINO) vs Text Fidelity (CLIP-T) (weight interpolation scale $\\omega \\in [0,1]$ , step $= 0.1$ )", + "Figure 4. Ablation study on the weight interpolation scale $\\omega \\in [0.0,1.0]$ with measurements taken at intervals of 0.1. Here, we use DreamBooth (DB) LoRA based on SD 1.5, SD 2.1, and SANA. By adjusting $\\omega$ , it is possible to improve subject fidelity or text fidelity, while preserving the other." + ], + "image_footnote": [], + "bbox": [ + 516, + 116, + 901, + 395 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "instead of the fine-tuned model $\\epsilon_{\\theta^{\\prime}}$ as the weak model.", + "bbox": [ + 511, + 482, + 872, + 497 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\epsilon} _ {\\theta^ {\\prime}} ^ {\\lambda} \\left(\\mathbf {x} _ {t} | \\mathbf {c}\\right) = \\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t} | \\phi\\right) + \\lambda \\left(\\epsilon_ {\\theta^ {\\prime}} \\left(\\mathbf {x} _ {t} | \\mathbf {c}\\right) - \\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t} | \\phi\\right)\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 508, + 905, + 529 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Despite being a simple extension, we found that this approach significantly improves subject fidelity.", + "bbox": [ + 511, + 540, + 905, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 3 shows the rationale behind our personalization guidance. Since this simply modified guidance can steer the output toward the intersection space where both text fidelity and subject fidelity are high. As illustrated in Fig. 3, we explore the feasibility of identifying a superior space within the noisy data manifold. We hypothesize that an optimal point for text and subject fidelity could be located between CFG and PG, and we propose a method to facilitate the steering of guidance. Steering the guidance involves the adjustment of an appropriate weak model. By identifying a suitable weak model, we can steer $\\epsilon_{\\theta}(\\mathbf{x}_t|\\phi)$ towards a more favorable point. We propose employing a weight interpolation to obtain an appropriate weak model. Specifically, by interpolating weights between the parameters of the pretrained and the fine-tuned models, we can modulate the degree of adaptation to the target data, thereby obtaining an appropriate weak model $\\epsilon_{\\theta_{\\omega}}$ .", + "bbox": [ + 511, + 571, + 906, + 828 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {\\omega} = \\omega \\theta^ {\\prime} + (1 - \\omega) \\theta , \\quad \\omega \\in [ 0, 1 ], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 840, + 903, + 858 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\omega$ denotes the scale of weight interpolation. Increasing $\\omega$ moves parameters closer to the fine-tuned model $\\theta^{\\prime}$", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15910", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/679f8766c1ed0e46e7e51f40a44cbceda0c02dbb581e0d4426c5ddf020f388e2.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Comparison with other guidance methods across different base diffusion models including SD1.5, SD2.1, and SANA. Here, we use DreamBooth-LoRA (DB-LoRA) only." + ], + "table_body": "
MethodλDINOCLIP-ICLIP-T
SD 1.5DB-LoRA-0.37410.67970.2834
+ CFG7.50.47010.73490.3345
+ AG2.00.48310.74080.3236
+ Ours (ω = 0)7.50.49850.75160.3260
+ Ours (ω = 0.6)7.50.50240.75100.3288
SD 2.1DB-LoRA-0.42020.70760.2929
+ CFG7.50.49760.75190.3323
+ AG2.00.50720.75710.3212
+ Ours (ω = 0)7.50.52480.76550.3254
+ Ours (ω = 0.6)7.50.52810.76600.3277
SANADB-LoRA-0.41900.69390.3064
+ CFG4.50.48190.72910.3363
+ AG2.00.47920.72840.3355
+ Ours (ω = 0)4.50.53660.75220.3357
", + "bbox": [ + 96, + 88, + 477, + 324 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "reducing the loss $\\mathcal{L}$ on $\\mathcal{D}_{\\mathrm{target}}$ . Finally, applying the weight interpolation to the weak model in the guidance:", + "bbox": [ + 89, + 372, + 482, + 402 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\epsilon} _ {\\theta^ {\\prime}} ^ {\\lambda} (\\mathbf {x} _ {t} | \\mathbf {c}) = \\epsilon_ {\\theta_ {\\omega}} (\\mathbf {x} _ {t} | \\phi) + \\lambda \\bigl (\\epsilon_ {\\theta^ {\\prime}} (\\mathbf {x} _ {t} | \\mathbf {c}) - \\epsilon_ {\\theta_ {\\omega}} (\\mathbf {x} _ {t} | \\phi) \\bigr). (9)\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 409, + 482, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that $\\omega = 1$ is equal to CFG. As shown in Fig. 4, it is possible to easily control the degree of subject and text fidelity. If $\\theta$ includes only linear classifier, the personalization guidance with weight interpolation is equal to the combination of CFG and AG, which is the output-space interpolation. However, as neural networks are non-linear with respect to parameters, guidance with weight interpolation is different from the output interpolation of CFG and AG [36].", + "bbox": [ + 89, + 434, + 483, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 566, + 223, + 584 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setting", + "text_level": 1, + "bbox": [ + 89, + 590, + 290, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We conduct our experiments on the ViCo dataset [9], which includes 16 concepts and 31 text prompts. Each concept includes a small set of reference images (typically 4-7) and associated text prompts. Following the previous work [22], we fine-tune the pre-trained text-to-image diffusion models per concept and then generate 8 images per concept and text prompt, resulting in a total of 3,968 images for evaluation.", + "bbox": [ + 89, + 613, + 482, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following the existing personalization studies [14, 22, 27], we evaluate subject and text fidelity. For the quantitative assessment, we adopt three evaluation metrics: (i) CLIP-T score [25], which uses CLIP's text-to-image similarity to evaluate text prompt alignment; (ii) CLIP-I score, which computes image-to-image similarity using CLIP embeddings to gauge how closely generated samples match reference images; and (iii) DINO score [3], which calculates image-to-image similarity between reference and generated images, offering a complementary perspective on subject fidelity.", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c3e2b1e44ebbe1820b900d27d2a8547efaec04fb2cbca863c7c8be35fbcf8ea2.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2. Comparison with other guidance methods across different base diffusion models including SD1.5 and SD2.1. Here, we use DreamBooth-LoRA (DB-LoRA) + Textual Inversion (TI) and ClassDiffusion. Since these personalization methods fine-tune text embedding, subject-agnostic guidance (SAG) is also compared with our method." + ], + "table_body": "
MethodλDINOCLIP-ICLIP-T
SD 1.5DB-LoRA+TI-0.37250.67750.2817
+ CFG7.50.46180.72920.3316
+ SAG1.00.46000.72730.3307
+ AG2.00.48060.74080.3264
+ Ours (ω = 0)7.50.48140.74590.3233
+ Ours (ω = 0.7)7.50.48800.74610.3272
ClassDiffusion-0.37030.66690.2757
+ CFG7.50.52870.75170.3305
+ SAG1.00.51720.74650.3296
+ AG2.00.55700.76280.3280
+ Ours (ω = 0)7.50.57000.76560.3297
SD 2.1DB-LoRA+TI-0.45140.72510.2823
+ CFG7.50.52910.77490.3244
+ SAG1.00.53280.77660.3223
+ AG2.00.55570.78770.3206
+ Ours (ω = 0)7.50.56830.79180.3163
+ Ours (ω = 0.6)7.50.56890.79190.3185
ClassDiffusion-0.42730.69950.2843
+ CFG7.50.58400.77570.3277
+ SAG1.00.57830.77390.3244
+ AG2.00.60600.78560.3246
+ Ours (ω = 0)7.50.61660.78960.3246
", + "bbox": [ + 517, + 88, + 901, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. In our experiments, we utilize three pre-trained text-to-image diffusion models as the base models: Stable Diffusion 1.5 (SD 1.5) [26], Stable Diffusion 2.1 (SD 2.1) [26], and SANA-1.6B [38]. SD 1.5 and SD 2.1 generate $512 \\times 512$ resolution images, which comprise U-Net based architectures. On the other hand, SANA is a recent text-to-image diffusion model, which is based on the diffusion transformer [23] that can generate images up to $4096 \\times 4096$ resolution. In the experiments, SANA generates $1024 \\times 1024$ images.", + "bbox": [ + 511, + 536, + 906, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To fine-tune the diffusion models, we leverage several personalization techniques: (i) DreamBoooth-LoRA (DB-LoRA) [13, 27], (ii) DB-LoRA + Textual Inversion (TI), and (iii) ClassDiffusion [14]. We fine-tune each model for up to 500 steps on the reference images from the ViCo dataset. We set the number of sampling steps (e.g., DDIM) between 20-50. Further details regarding additional hyperparameters are described in the supplementary.", + "bbox": [ + 511, + 688, + 905, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We evaluate our guidance method against other guidance techniques: (i) Classifier-Free Guidance (CFG) [10], (ii) Subject-Agnostic Guidance (SAG) [4], and (iii) AutoGuidance (AG) [18]. Since SAG and AG alone struggle to generate images that accurately reflect the text prompt, they are used in conjunction with CFG to produce", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15911", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1d705a80131c1bae71e8c79f1a6b8494f5312ce07dd18e15e4dac37e4555b5e6.jpg", + "image_caption": [ + "Figure 5. Comparison with other guidance techniques, including without guidance, CFG, CFG+SAG, and CFG+AG. These images are generated by the fine-tuned SD 2.1 using DB-LoRA and ClassDiffusion." + ], + "image_footnote": [], + "bbox": [ + 99, + 93, + 906, + 353 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/50c376ece40a93b7ade99a7c6ab85e351bf8e8c86ddc1916c68038b40127be71.jpg", + "image_caption": [ + "Figure 6. Comparison with other guidance techniques. These images are generated by fine-tuned SANA [38] using DB-LoRA." + ], + "image_footnote": [], + "bbox": [ + 96, + 393, + 906, + 621 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "images. Specifically, AG experiments are conducted by referring to recent work [19] that combines CFG and AG.", + "bbox": [ + 89, + 656, + 482, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results", + "text_level": 1, + "bbox": [ + 89, + 696, + 183, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Results. Table 1 and Table 2 showcase the personalization performance on the ViCo dataset. Table 1 provides the quantitative results for DB-LoRA based on SD 1.5, SD 2.1, and SANA. Since DB-LoRA+TI and Class-Diffusion involve fine-tuning textual embeddings, we also evaluate subject-agnostic guidance (SAG) [4], which necessitates learned textual embeddings for specific concepts. As illustrated in both Table 1 and Table 2, our guidance method significantly enhances subject fidelity (DINO and CLIP-I scores), while maintaining text fidelity (CLIP-T score) with minimal change. Additionally, by adjusting $\\omega$ , the model can further optimize personalization performance across all", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "evaluation metrics. We select the optimal $\\omega$ based on the best DINO and CLIP-I scores. For SANA with DB-LoRA and ClassDiffusion, the best subject fidelity was observed at $\\omega = 0$ . These results demonstrate that our approach significantly improves subject fidelity while maintaining text fidelity, and that steering the guidance enables a balanced trade-off between the two.", + "bbox": [ + 511, + 656, + 906, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Interestingly, the impact of personalized guidance varies depending on the personalization method. ClassDiffusion [14] employs a semantic preservation loss that maintains the text embedding space, leading to better overall text fidelity retention. As a result, our guidance method maintains the CLIP-T score more effectively when applied to ClassDiffusion than to other personalization methods. This observation highlights a key insight: the better the fine-tuning direction, the more effective our guidance becomes.", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15912", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/17e0c5f46402d9a73a2f9565946968999bb38872429f0f2a1a45f7f67f3dbf86.jpg", + "image_caption": [ + "Figure 7. Generated images by changing $\\omega$ , where SANA with DB-LoRA is used." + ], + "image_footnote": [], + "bbox": [ + 130, + 90, + 864, + 281 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/91489bc95955dc44d77e01a3eed13a0fdfe422e08c3aabf176fc9dbff4303b07.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSubject Fidelity ↑Text Fidelity ↑
Stable Diffusion 2.1 + ClassDiffusion
CFG10.45%15.24%
CFG + SAG8.01%6.20%
CFG + AG12.60%6.20%
Ours55.22%32.04%
Undecided13.67%40.31%
SANA + DB-LoRA
CFG2.48%3.08%
CFG + AG2.17%2.77%
Ours68.32%8.64%
Undecided27.01%85.49%
", + "bbox": [ + 117, + 309, + 455, + 512 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. User preference on subject fidelity and text fidelity.", + "bbox": [ + 107, + 515, + 464, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results. Fig. 5 and Fig. 6 show the comparison results based on SD 2.1 and SANA, respectively. As shown in Fig. 5, CFG often fails to capture fine details of the reference subject (e.g., cat's appearance and the teddybear's eye). Additionally, applying SAG or AG has minimal impact in some cases (first row of Fig. 5) or even degrades image quality, introducing artifacts such as a teddy bear with three eyes (CFG+AG, second row of Fig. 5). In contrast, our method consistently achieves better subject fidelity with the same noise input. Similarly, in Fig. 6, our method better preserves subject details compared to CFG and CFG+AG, as evidenced by the significant improvement in subject fidelity in quantitative results. For instance, it accurately retains the ears and color of a dog (first row of Fig. 6) and the shape and numerical details of a clock (second row of Fig. 6). In contrast, applying AG with CFG shows little to no improvement despite incurring a higher computational cost compared to both CFG and our guidance method.", + "bbox": [ + 89, + 537, + 482, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User Study. Table 3 presents the user preference results comparing our method with baseline approaches. The results show that our guidance technique achieves significantly higher subject fidelity than other guidance methods. Our approach also shows superior text fidelity in user evaluations. Further details can be found in the supplementary.", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1a917af5ed98a8aec4c07d008536465cd3e144fa6ba21d34b16f34e6522c3167.jpg", + "image_caption": [ + "Subject Fidelity (DINO) vs Text Fidelity (CLIP-T) (guidance scale $\\lambda \\in [2.5, 5.0]$ , step = 0.5)", + "Figure 8. Ablation study on the guidance scale $\\lambda \\in [2.5, 5.0]$ , measuring performance at 0.5 intervals using DB-LoRA based on SANA. When $\\omega$ is set to 0.2, we can improve both subject and text fidelity compared to CFG, regardless of guidance scale $\\lambda$ . Note that CLIP-T score is slightly changed in the graph $(max(CLIP-T) - min(CLIP-T) < 0.004)$ ." + ], + "image_footnote": [], + "bbox": [ + 516, + 327, + 898, + 554 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Further Analysis", + "text_level": 1, + "bbox": [ + 511, + 648, + 681, + 664 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Study on $\\omega$ . Fig. 4 illustrates how DINO and CLIP-T scores vary with changes in $\\omega$ . For both SD 1.5 and SD 2.1, personalization guidance slightly reduces text fidelity but significantly enhances subject fidelity compared to CFG ( $\\omega = 1$ ). By appropriately adjusting $\\omega$ , it is possible to either boost text fidelity while maintaining subject fidelity (as indicated by the pink arrow) or enhance subject fidelity while preserving text fidelity (as indicated by the blue arrow). It is important to note that these trends may vary depending on the personalization technique or the pre-trained diffusion model. However, the ability to easily fine-tune results via weight interpolation at inference, without modifying training steps or learning rates, offers a major practical advantage over traditional methods for controlling overfitting. Fig. 7 further visualizes the impact of $\\omega$ on gen", + "bbox": [ + 509, + 672, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15913", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0cadc9c90ad9a3a236e20a924c06e815efbd2aa75916588431776a2c1f0e4c2d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPickScore↑
DPO+CFG21.53
DPO+Ours21.58
", + "bbox": [ + 96, + 80, + 263, + 141 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3349415b8f3464d29f6a27bc26c2387fd9df7eb55e43b94c91ea0fedc54b6b34.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDreamSim↓
PairCustom+CFG0.7993
PairCustom+Ours0.7344
", + "bbox": [ + 272, + 80, + 483, + 140 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Results of combining Diffusion-DPO with ours.", + "bbox": [ + 91, + 142, + 269, + 170 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Results of combining Pair-Custom with ours.", + "bbox": [ + 272, + 143, + 480, + 170 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c2b3fb3b5cfe9e7d732fe412e893c8b52684fff8cbef965c1f757c093092cb5d.jpg", + "image_caption": [ + "Figure 9. Qualitative results of our guidance across diverse tasks." + ], + "image_footnote": [], + "bbox": [ + 96, + 179, + 473, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "eration quality. The first row shows that higher $\\omega$ values retain fine details, such as the intricate patterns on the vase, at the cost of weaker adherence to the \"purple\" text prompt. Additionally, as $\\omega$ decreases, subject fidelity improves significantly, particularly in the range of 0.6 to 1.0.", + "bbox": [ + 89, + 506, + 482, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Study on Guidance Scale. Fig. 8 shows how personalization performance varies with different guidance scales $\\lambda$ . We evaluate our guidance at $\\omega = 0$ and $\\omega = 0.2$ , where text fidelity remains preserved. As $\\lambda$ decreases, text fidelity deteriorates for both guidance methods. However, as $\\lambda$ increases, our guidance method shows improvements in both the DINO score and CLIP-T score, whereas the DINO score of CFG declines. This suggests that CFG reduces subject fidelity, and adjusting the guidance scale alone is insufficient to correct this effect.", + "bbox": [ + 89, + 580, + 483, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Application to Fine-tuned Diffusion Models.", + "text_level": 1, + "bbox": [ + 89, + 742, + 465, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fine-tuning diffusion model extends beyond personalization tasks to applications such as style personalization [17], human preference learning [21, 33], and specific task learning [2, 15]. Our personalization guidance framework leverages the knowledge gap between the pre-trained and fine-tuned models, emphasizing the information better captured by the fine-tuned model. This concept can be generalized to control the degree of adaptation for other fine-tuned models, such as adjusting the extent of human preference alignment", + "bbox": [ + 89, + 763, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "or style personalization during inference.", + "bbox": [ + 511, + 90, + 785, + 106 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To this end, to investigate our method's generalizability across diverse tasks, Fig. 9 show the following experiments: (a) Human Preference Optimization (Diffusion-DPO [33]), (b) Style Personalization (PairCustom [17]), and (c) Instruction-based Editing (InstructPix2Pix [2]).", + "bbox": [ + 511, + 106, + 903, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(a) Human Preference Optimization: For evaluation with the Diffusion-DPO model [33], we use the PartiPrompts dataset [39] and report PickScore [39], which reflects human-perceived image quality. Table 4 shows that integrating our method into the SD 1.5-based Diffusion-DPO model yields superior performance compared to CFG, confirming the effectiveness of our approach on human preference optimization tasks. As shown in Fig. 9 (a), our method helps generate more natural-looking images and corrects artifacts that would otherwise appear unnatural (e.g., additional finger).", + "bbox": [ + 511, + 181, + 906, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(b) Style Personalization: To evaluate style alignment, we use two pre-trained style-LoRAs released by PairCustom, combining each with both CFG and our method. Using text prompts from the StyleAlign [37] paper, we generate 800 images and compute the DreamSim score [7] to quantify the similarity between the target style and generated outputs. As shown in Table 5, applying our method significantly improves style alignment, demonstrating its effectiveness for style personalization. In Fig. 9 (b), combining PairCustom with our guidance leads to a more faithful transfer of the painting style from the reference image.", + "bbox": [ + 511, + 347, + 906, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(c) Instruction-based Editing: In Fig. 9 (c), our guidance improves the performance of the InstructPix2Pix [2], which is designed to edit images based on text instructions, to better follow the given instruction. As a result, the generated outputs are more accurately aligned with the intended edits.", + "bbox": [ + 511, + 513, + 905, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "These results demonstrate that applying our guidance to various fine-tuned diffusion models allows them to generate images more faithfully aligned with their respective finetuning objectives. This highlights the generalizability of our method across diverse tasks. Further implementation details are provided in the supplementary.", + "bbox": [ + 511, + 589, + 905, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 693, + 633, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper tackles the critical challenge of balancing adaptation to the target distribution and knowledge preservation from the original model in personalized diffusion models. Our approach simply modifies the existing guidance by adjusting the degree of unlearning in the weak model through weight interpolation between the pre-trained and fine-tuned models. This simple yet effective guidance enhances subject fidelity while preserving text editability, ensuring an optimal trajectory for personalization. Furthermore, we demonstrate that our guidance can be effectively applied to a wide range of fine-tuned diffusion models, such as human preference learning.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15914", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement We would like to thank the Qualcomm AI Research team for their valuable discussions.", + "bbox": [ + 91, + 90, + 480, + 119 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 136, + 187, + 151 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Donghoon Ahn, Hyoungwon Cho, Jaewon Min, Wooseok Jang, Jungwoo Kim, SeonHwa Kim, Hyun Hee Park, Kyong Hwan Jin, and Seungryong Kim. Self-rectifying diffusion sampling with perturbed-attention guidance. In European Conference on Computer Vision, pages 1-17. Springer, 2024. 3", + "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18392-18402, 2023. 3, 8", + "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 5", + "[4] Kelvin CK Chan, Yang Zhao, Xuhui Jia, Ming-Hsuan Yang, and Huisheng Wang. Improving subject-driven image synthesis with subject-agnostic guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6733–6742, 2024. 3, 5, 6", + "[5] Shangyu Chen, Zizheng Pan, Jianfei Cai, and Dinh Phung. Para: Personalizing text-to-image diffusion via parameter rank reduction. arXiv preprint arXiv:2406.05641, 2024. 2, 3", + "[6] Wenhu Chen, Hexiang Hu, Yandong Li, Nataniel Ruiz, Xuhui Jia, Ming-Wei Chang, and William W Cohen. Subject-driven text-to-image generation via apprenticeship learning. Advances in Neural Information Processing Systems, 36:30286–30305, 2023. 2, 3", + "[7] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. arXiv preprint arXiv:2306.09344, 2023.8", + "[8] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In The Eleventh International Conference on Learning Representations. 2, 3", + "[9] Shaozhe Hao, Kai Han, Shihao Zhao, and Kwan-Yee K Wong. Vico: Plug-and-play visual condition for personalized text-to-image generation. arXiv preprint arXiv:2306.00971, 2023.5", + "[10] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 2, 3, 4, 5", + "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 2, 3", + "[12] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seunglyon Kim. Improving sample quality of diffusion models using self-attention guidance. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 93, + 161, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "International Conference on Computer Vision, pages 7462-7471, 2023. 3", + "[13] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 2, 3, 5", + "[14] Jiannan Huang, Jun Hao Liew, Hanshu Yan, Yuyang Yin, Yao Zhao, and Yunchao Wei. Classdiffusion: More aligned personalization tuning with explicit class guidance. arXiv preprint arXiv:2405.17532, 2024. 2, 3, 5, 6", + "[15] Lianghua Huang, Wei Wang, Zhi-Fan Wu, Yupeng Shi, Huanzhang Dou, Chen Liang, Yutong Feng, Yu Liu, and Jingren Zhou. In-context lora for diffusion transformers. arXiv preprint arXiv:2410.23775, 2024. 8", + "[16] Junha Hyung, Kinam Kim, Susung Hong, Min-Jung Kim, and Jaegul Choo. Spatiotemporal skip guidance for enhanced video diffusion sampling. arXiv preprint arXiv:2411.18664, 2024.3", + "[17] Maxwell Jones, Sheng-Yu Wang, Nupur Kumari, David Bau, and Jun-Yan Zhu. Customizing text-to-image models with a single image pair. In SIGGRAPH Asia 2024 Conference Papers, pages 1-13, 2024. 8", + "[18] Tero Karras, Miika Aittala, Tuomas Kynkänniemi, Jaakko Lehtinen, Timo Aila, and Samuli Laine. Guiding a diffusion model with a bad version of itself. Advances in Neural Information Processing Systems, 37:52996-53021, 2025. 2, 3, 4, 5", + "[19] Artur Kasymov, Marcin Sendera, Michal Stypulkowski, Macej Zięba, and Przemysław Spurek. Autolora: Autoguidance meets low-rank adaptation for diffusion models. arXiv preprint arXiv:2410.03941, 2024. 6", + "[20] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1931-1941, 2023. 2, 3", + "[21] Shufan Li, Konstantinos Kallidromitis, Akash Gokul, Yusuke Kato, and Kazuki Kozuka. Aligning diffusion models by optimizing human utility. Advances in Neural Information Processing Systems, 37:24897-24925, 2025. 8", + "[22] Jisu Nam, Heesu Kim, DongJae Lee, Siyoon Jin, Seungryong Kim, and Seunggyu Chang. Dreamcatcher: appearance matching self-attention for semantically-consistent text-to-image personalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8100–8110, 2024. 3, 5", + "[23] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195-4205, 2023. 5", + "[24] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2", + "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15915", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5", + "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2, 4, 5", + "[27] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, pages 22500-22510, 2023. 2, 3, 5", + "[28] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdream booth: Hypermetworks for fast personalization of text-to-image models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6527-6536, 2024. 2, 3", + "[29] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 35:36479-36494, 2022. 2", + "[30] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8543-8552, 2024. 2, 3", + "[31] Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. Freeu: Free lunch in diffusion u-net. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4733-4743, 2024. 3", + "[32] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 3", + "[33] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228-8238, 2024. 8", + "[34] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15943-15953, 2023. 2, 3", + "[35] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International conference on machine learning, pages 23965-23998. PMLR, 2022. 3", + "[36] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok" + ], + "bbox": [ + 91, + 90, + 483, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Namkoong, et al. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7959-7971, 2022. 3, 5", + "[37] Zongze Wu, Yotam Nitzan, Eli Shechtman, and Dani Lischinski. Stylealign: Analysis and applications of aligned stylegan models. arXiv preprint arXiv:2110.11323, 2021. 8", + "[38] Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, et al. Sana: Efficient high-resolution image synthesis with linear diffusion transformers. arXiv preprint arXiv:2410.10629, 2024. 2, 5, 6", + "[39] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 8" + ], + "bbox": [ + 516, + 92, + 903, + 318 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15916", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Steering Guidance for Personalized Text-to-Image Diffusion Models/34e2c957-5535-42df-a5b2-c4f7cd9d6450_model.json b/2025/Steering Guidance for Personalized Text-to-Image Diffusion Models/34e2c957-5535-42df-a5b2-c4f7cd9d6450_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fc539e2ca28adaf08a3406e651a423c176c428ef --- /dev/null +++ b/2025/Steering Guidance for Personalized Text-to-Image Diffusion Models/34e2c957-5535-42df-a5b2-c4f7cd9d6450_model.json @@ -0,0 +1,2134 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.157, + 0.131, + 0.842, + 0.154 + ], + "angle": 0, + "content": "Steering Guidance for Personalized Text-to-Image Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.181, + 0.78, + 0.217 + ], + "angle": 0, + "content": "Sunghyun Park* Seokeon Choi* Hyoungwoo Park Sungrack Yun Qualcomm AI Research†" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.221, + 0.714, + 0.234 + ], + "angle": 0, + "content": "{sunpar, seokchoi, hwoopark, sungrack}@qti=qualcomm.com" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.254, + 0.189, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.254, + 0.368, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.254, + 0.545, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.254, + 0.722, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.254, + 0.899, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.392, + 0.439, + 0.403 + ], + "angle": 0, + "content": "A \\(\\star\\) cat in a chef outfit" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.406, + 0.188, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.191, + 0.406, + 0.545, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.406, + 0.722, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.406, + 0.899, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.287, + 0.544, + 0.449, + 0.556 + ], + "angle": 0, + "content": "A \\(\\star\\) dog in a movie theater" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.558, + 0.188, + 0.696 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.191, + 0.558, + 0.368, + 0.695 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.558, + 0.544, + 0.696 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.558, + 0.722, + 0.696 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.558, + 0.899, + 0.696 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.696, + 0.177, + 0.707 + ], + "angle": 0, + "content": "Reference" + }, + { + "type": "image_caption", + "bbox": [ + 0.195, + 0.696, + 0.519, + 0.708 + ], + "angle": 0, + "content": "A duck toy with the Eiffel Towel in the background" + }, + { + "type": "image_caption", + "bbox": [ + 0.56, + 0.696, + 0.885, + 0.709 + ], + "angle": 0, + "content": "A duck toy among the skyscrapers in New York city" + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.713, + 0.871, + 0.727 + ], + "angle": 0, + "content": "Figure 1. Comparison of generated image quality between Classifier-Free Guidance (CFG) and our Personalization Guidance." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.753, + 0.327, + 0.768 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.786, + 0.483, + 0.862 + ], + "angle": 0, + "content": "Personalizing text-to-image diffusion models is crucial for adapting the pre-trained models to specific target concepts, enabling diverse image generation. However, fine-tuning with few images introduces an inherent trade-off between aligning with the target distribution (e.g., subject fidelity)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.739, + 0.907, + 0.892 + ], + "angle": 0, + "content": "and preserving the broad knowledge of the original model (e.g., text editability). Existing sampling guidance methods, such as classifier-free guidance (CFG) and autoguidance (AG), fail to effectively guide the output toward well-balanced space: CFG restricts the adaptation to the target distribution, while AG compromises text alignment. To address these limitations, we propose personalization guidance, a simple yet effective method leveraging an unlearned weak model conditioned on a null text prompt. Moreover, our method dynamically controls the extent of unlearning" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.374, + 0.888 + ], + "angle": 0, + "content": "*These two authors contributed equally to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.888, + 0.48, + 0.9 + ], + "angle": 0, + "content": "\\(\\dagger\\)Qualcomm AI Research is an initiative of Qualcomm Technologies, Inc." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.875, + 0.48, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15907" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.228 + ], + "angle": 0, + "content": "in a weak model through weight interpolation between pretrained and fine-tuned models during inference. Unlike existing guidance methods, which depend solely on guidance scales, our method explicitly steers the outputs toward a balanced latent space without additional computational overhead. Experimental results demonstrate that our proposed guidance can improve text alignment and target distribution fidelity, integrating seamlessly with various fine-tuning strategies." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.257, + 0.223, + 0.273 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.281, + 0.483, + 0.417 + ], + "angle": 0, + "content": "Diffusion models [11, 32] have emerged as powerful generative models capable of representing complex data distributions, driving significant advancements in high-fidelity image generation [24, 26, 29, 38]. However, real-world use cases often demand personalized generation, where the model generates not only generic concepts but also specific concepts like a novel character or person [8, 20, 27]. This personalization technique is particularly useful for applications such as portrait editing and specific content creation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.418, + 0.483, + 0.658 + ], + "angle": 0, + "content": "Many approaches have been proposed to personalize diffusion models for novel concepts. One line of work [6, 28, 30, 34] learns a general-purpose encoder capable of adapting to new concepts without test-time finetuning. However, this approach requires large-scale text-image datasets and considerable computational resources, and must often be retrained from scratch for each pre-trained model, making it expensive to adapt to the growing variety of text-to-image or text-to-video diffusion architectures. An alternative solution is to directly fine-tune a pre-trained diffusion model on the target images, which can be roughly categorized into fine-tuning low-rank adapters [5, 13], text embeddings [8], specific layers [14, 20], or the full model parameters [27]. This strategy can be applied to any pre-trained diffusion model and only requires a few target images, proving effective for personalization in various scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Despite these advancements, optimization-based personalization remains a nontrivial challenge. When fine-tuning a model to a specific, often much smaller, target data distribution, there is an inevitable tradeoff between aligning with the target distribution (e.g., learning a novel concept with a few target images) and preserving broad knowledge of the original model (e.g., preserving text editability). This tradeoff manifests in various ways: a fine-tuned model for personalization may generate high-quality domain-specific images but lose text editability, making it difficult to generalize across different prompts. Conversely, a model that retains broad knowledge from its original training distribution may struggle to fully align with new concepts introduced in fine-tuning. Existing approaches attempt to mitigate this tradeoff by adjusting the number of fine-tuning iterations [8, 27], leveraging regularization [14, 27], reducing the model pa" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.092, + 0.603, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.092, + 0.704, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.092, + 0.803, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.803, + 0.092, + 0.905, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.196, + 0.905, + 0.265 + ], + "angle": 0, + "content": "Figure 2. Motivation of Personalization Guidance. Classifier-free guidance [10] can interfere with subject fidelity due to its alignment with the text prompt. In contrast, autoguidance [18] fails to reflect the text prompt but generates the subject more similarly. Our guidance complements the limitations of both methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.276, + 0.905, + 0.337 + ], + "angle": 0, + "content": "rameters of updates [5, 13], yet they primarily operate at the fine-tuning stage. Moreover, balancing the trade-off between comprehensive text editability and precise alignment with a new target distribution remains an open problem." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.339, + 0.907, + 0.642 + ], + "angle": 0, + "content": "A parallel line of research focuses on sampling guidance techniques, such as Classifier-Free Guidance (CFG) [10] and AutoGuidance (AG) [18]. These methods rely on \"weak models\" to detect and push the outputs away from poor trajectories, thereby guiding samples toward higher-quality regions of the data manifold. For instance, CFG uses an unconditional model using the null text prompt as the weak model, while AG uses an unlearned model—one with fewer parameters or less training—conditioned on the same text prompt. However, these methods have limitations when applied to personalized diffusion models, as shown in Fig. 2. CFG can interfere with the model's adaptation to the target distribution, restricting its ability to fully utilize fine-tuned knowledge. Moreover, while the guidance scale of CFG adjusts the degree of text fidelity, it does not improve alignment with the target data. On the other hand, AG, which primarily enhances alignment with the target distribution, however, lacks an explicit mechanism for improving text alignment, making it unsuitable for fine-tuned text-to-image models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.909, + 0.902 + ], + "angle": 0, + "content": "To address these issues, we propose a simple yet effective guidance technique, called Personalization Guidance, for personalized diffusion models, which can dynamically balance text alignment and adaptation to the target distribution. Our method leverages an \"unlearned\" weak model with a null text prompt, without incurring additional computational overhead during inference. Here, we aim to balance the alignment with the target distribution (e.g., subject fidelity) and the preservation of the broad knowledge of the original model (e.g., text fidelity) by controlling the extent of \"unlearning\" in a weak model. Here, the degree of unlearning refers to the extent to which the target data distribution has not been learned compared to the fine-tuned model. The rationale behind adjusting this degree of unlearning is to steer the output toward an optimal, well-balanced latent space. Specifically, during inference, we interpolate the weight parameters between the pre-trained and fine-tuned" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15908" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.184 + ], + "angle": 0, + "content": "diffusion models to derive an optimal weak model, effectively controlling the extent of unlearning. Unlike guidance scales, which adjust the intensity of guidance, weight interpolation for the weak model serves to steer outputs in an appropriate direction. This straightforward approach can be readily integrated into various fine-tuning diffusion models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.196, + 0.234, + 0.212 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.221, + 0.485, + 0.478 + ], + "angle": 0, + "content": "Personalization of Diffusion Models. Personalizing diffusion models aim to generate novel images by adapting to specific concepts using a set of reference images. Optimization-based methods personalize models by finetuning textual embeddings [8, 14, 20], entire model parameters with regularization [27], or low-rank adapters [5, 13]. Another line of research leverages extensive text-image datasets to train dedicated text-to-image personalization models, by fine-tuning additional modules [6, 28, 30, 34]. Recently, several studies [4, 22, 31] have achieved better personalization without additional fine-tuning or training. Orthogonal to these existing approaches, we introduce a simple yet effective personalization guidance technique that seamlessly integrates with various personalization methods by modifying the classifier-free guidance. Moreover, our guidance enables us to adjust the degree of subject and text fidelity without incurring additional computational costs." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.479, + 0.484, + 0.795 + ], + "angle": 0, + "content": "Guidance for Diffusion Models. Classifier-Free Guidance (CFG) [10] enhances conditional image generation by leveraging the unconditional model as a weak model to guide the outputs toward better adherence to the given condition. Various guidance techniques have been explored, such as the guidance using perturbations [1, 16] or blurring to the attention maps [12]. Autoguidance [18] improves generation by using a weak model—one with fewer parameters or less training—to push outputs toward a well-trained direction. While InstructPix2Pix [2] introduces separate CFG using both image and text conditions, it cannot be directly applied to DB-LoRA due to its reliance on image-conditioned training. For personalization, subject-agnostic guidance [4] enhances subject-agnostic attributes by employing a weak model conditioned on subject-specific text prompts. In this study, we aim to address the fundamental personalization trade-off—aligning a model with a smaller target data distribution while preserving the broad text fidelity of the original model. We derive a weak model that dynamically adjusts the degree of adaptation, balancing comprehensive text editability with precise alignment to new concepts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Weight Interpolation. There are previous studies on exploring the use of weight interpolation. WISE-FT [36] improves the robustness of fine-tuned classifiers against distribution shifts through weight-space ensembling. Model-soups [35] maximizes model accuracy and robustness by interpolating the weights of multiple fine-tuned classifiers trained with different hyperparameter configurations." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.606, + 0.105 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.115, + 0.655, + 0.13 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.136, + 0.907, + 0.287 + ], + "angle": 0, + "content": "Diffusion Models. Diffusion models [11, 32] aim to learn reversing a forward noising process that gradually transforms data samples into noise. Concretely, the forward process starts with a data distribution \\( p_0(\\mathbf{x}) \\) and ends at a high-noise distribution \\( p_T(\\mathbf{x}) \\approx \\mathcal{N}(\\mathbf{0},\\mathbf{I}) \\). Sampling from the data distribution proceeds by solving the equivalent probability-flow ordinary differential equation (PF-ODE). Under the variance-preserving parameterization where \\( p(\\mathbf{x}_t|\\mathbf{x}_0) = \\mathcal{N}(\\mathbf{x}_0,\\sigma_t^2\\mathbf{I}) \\), the PF-ODE takes the form:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.298, + 0.907, + 0.315 + ], + "angle": 0, + "content": "\\[\n\\mathrm {d} \\mathbf {x} _ {t} = - \\sigma_ {t} \\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {x} _ {t}) \\mathrm {d} t, \\quad \\mathbf {x} _ {T} \\sim p _ {T} (\\mathbf {x} _ {T}). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.323, + 0.907, + 0.369 + ], + "angle": 0, + "content": "In diffusion models, the score function is parameterized by \\(\\epsilon_{\\theta}\\). Here, the noise estimation of the diffusion model can be considered as \\(\\epsilon_{\\theta}(\\mathbf{x}_t) \\approx -\\sigma_t \\nabla_{\\mathbf{x}_t} \\log p(\\mathbf{x}_t)\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.369, + 0.906, + 0.474 + ], + "angle": 0, + "content": "Personalization. Suppose we have a pre-trained diffusion model \\(\\epsilon_{\\theta}\\), where \\(\\theta\\) is the parameters of the model. We wish to personalize it to a smaller target dataset \\(\\mathcal{D}_{\\mathrm{target}} = \\left\\{(\\mathbf{x}^i, \\mathbf{c}^i)\\right\\}_{i=1}^N\\), where each \\(\\mathbf{x}^i\\) and \\(\\mathbf{c}^i\\) is an image and its associated text prompt, respectively. The goal is to obtain new parameters \\(\\theta'\\) so that the fine-tuned model \\(\\epsilon_{\\theta'}\\) can generate samples more faithfully matched the \\(\\mathcal{D}_{\\mathrm{target}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.475, + 0.906, + 0.55 + ], + "angle": 0, + "content": "Fine-tuning typically follows the same denoising objective used by the original diffusion model [11]. Concretely, let \\(\\sigma\\) be a noise level sampled from a training schedule, and let \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\sigma^2\\mathbf{I})\\). For each \\((\\mathbf{x},\\mathbf{c})\\) in \\(\\mathcal{D}_{\\mathrm{target}}\\), we form the noisy input \\(\\mathbf{x}_t = \\mathbf{x} + \\epsilon\\) and minimize:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.558, + 0.907, + 0.588 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} _ {(\\mathbf {x}, \\mathbf {c}) \\sim \\mathcal {D} _ {\\text {t a r g e t}}, \\sigma , \\boldsymbol {\\epsilon}} \\left\\| \\epsilon_ {\\theta^ {\\prime}} (\\mathbf {x} _ {t}, \\sigma , \\mathbf {c}) - \\boldsymbol {\\epsilon} \\right\\| _ {2} ^ {2}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.906, + 0.655 + ], + "angle": 0, + "content": "This teaches the model to reconstruct \\( \\mathbf{x} \\) from noisy inputs in the domain of \\( \\mathcal{D}_{\\mathrm{target}} \\), thereby adapting it to the new concept. In practice, additional regularization terms are often used to prevent overfitting [14, 27]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.655, + 0.906, + 0.732 + ], + "angle": 0, + "content": "Classifier-Free Guidance. To enhance the generation towards an arbitrary condition \\(\\mathbf{c}\\), classifier-free guidance (CFG) [10] introduces a new sampling distribution \\(p^{\\lambda}(\\mathbf{x}|\\mathbf{c})\\) composed with both \\(p(\\mathbf{x})\\) and the classifier distribution \\(p(c|\\mathbf{x})^{\\lambda}\\), which is described as:" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.739, + 0.907, + 0.758 + ], + "angle": 0, + "content": "\\[\np ^ {\\lambda} (\\mathbf {x} | \\mathbf {c}) \\propto p (\\mathbf {x}) p (\\mathbf {c} | \\mathbf {x}) ^ {\\lambda}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.906, + 0.811 + ], + "angle": 0, + "content": "where \\(\\lambda > 1\\) amplifies the conditional likelihood. By applying Bayes' rule for some timestep \\(t\\), the corresponding score function is:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.824, + 0.905, + 0.862 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {\\mathbf {x} _ {t}} \\log p ^ {\\lambda} (\\mathbf {x} _ {t} | \\mathbf {c}) = \\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {x} _ {t}) \\\\ + \\lambda \\left(\\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {x} _ {t} | \\mathbf {c}) - \\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {x} _ {t})\\right). \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "To enable conditional generation such as text prompts and class labels, CFG jointly trains the unconditional model" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15909" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.089, + 0.476, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.328, + 0.485, + 0.397 + ], + "angle": 0, + "content": "Figure 3. Comparison between Classifier-Free Guidance (CFG), AutoGuidance (AG), and our Personalization Guidance (PG), with the band conceptually representing the noisy data manifold. By leveraging weight interpolation, guidance can be steered toward the optimal space between CFG and PG." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.408, + 0.484, + 0.483 + ], + "angle": 0, + "content": "\\(\\epsilon_{\\theta}(\\mathbf{x}_t,\\phi)\\) and the conditional model \\(\\epsilon_{\\theta}(\\mathbf{x}_t,\\mathbf{c})\\) within a single model by dropping \\(\\mathbf{c}\\) with certain probability to allow null conditioning with \\(c = \\phi\\). In the personalization setting, parameterizing the score function with the fine-tuned diffusion model \\(\\epsilon_{\\theta'}\\) is described as:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.491, + 0.483, + 0.511 + ], + "angle": 0, + "content": "\\[\n\\hat {\\epsilon} _ {\\theta^ {\\prime}} ^ {\\lambda} \\left(\\mathbf {x} _ {t} \\mid \\mathbf {c}\\right) = \\epsilon_ {\\theta^ {\\prime}} \\left(\\mathbf {x} _ {t} \\mid \\phi\\right) + \\lambda \\left(\\epsilon_ {\\theta^ {\\prime}} \\left(\\mathbf {x} _ {t} \\mid \\mathbf {c}\\right) - \\epsilon_ {\\theta^ {\\prime}} \\left(\\mathbf {x} _ {t} \\mid \\phi\\right)\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.484, + 0.564 + ], + "angle": 0, + "content": "However, while CFG improves alignment with the conditions, such as the text prompt, it struggles to enhance adaptation to the target data in the personalization task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.484, + 0.776 + ], + "angle": 0, + "content": "AutoGuidance. AutoGuidance (AG) [18] is another guidance technique that employs a weaker version of the same model to identify suboptimal generation trajectories. Rather than using the unconditional model \\(\\epsilon_{\\theta}(\\mathbf{x}_t|\\phi)\\), AG directly guides a model with the weak model \\(\\epsilon_{\\theta_{weak}}\\), which is trained on the same task and conditioning, but degraded by under-training or less-parameters. However, obtaining a weaker version \\(\\epsilon_{\\theta_{weak}}\\) of a fully trained diffusion model like Stable Diffusion [26] is challenging in practice. Instead, we found that a less trained model could serve as the pre-trained model for personalization, where \\(\\theta_{weak} = \\theta\\), effectively shifts the output away in the direction of fine-tuning, enhancing the adaptation to the target data. In the personalization, AG is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.784, + 0.483, + 0.803 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\epsilon} _ {\\theta^ {\\prime}} ^ {\\lambda} \\left(\\mathbf {x} _ {t} | \\mathbf {c}\\right) = \\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t} | \\mathbf {c}\\right) + \\lambda \\left(\\epsilon_ {\\theta^ {\\prime}} \\left(\\mathbf {x} _ {t} | \\mathbf {c}\\right) - \\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t} | \\mathbf {c}\\right)\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.841 + ], + "angle": 0, + "content": "However, this approach has a major drawback—it scarcely reflects the text prompt, as shown in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.394, + 0.866 + ], + "angle": 0, + "content": "3.2. Steering Personalization Guidance" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Inspired by autoguidance [18], we simply modify the CFG [10], by leveraging a pre-trained diffusion model \\(\\epsilon_{\\theta}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.094, + 0.851, + 0.116 + ], + "angle": 0, + "content": "Subject Fidelity (DINO) vs Text Fidelity (CLIP-T) (weight interpolation scale \\(\\omega \\in [0,1]\\), step \\(= 0.1\\))" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.117, + 0.903, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.4, + 0.907, + 0.47 + ], + "angle": 0, + "content": "Figure 4. Ablation study on the weight interpolation scale \\(\\omega \\in [0.0,1.0]\\) with measurements taken at intervals of 0.1. Here, we use DreamBooth (DB) LoRA based on SD 1.5, SD 2.1, and SANA. By adjusting \\(\\omega\\), it is possible to improve subject fidelity or text fidelity, while preserving the other." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.483, + 0.874, + 0.498 + ], + "angle": 0, + "content": "instead of the fine-tuned model \\(\\epsilon_{\\theta^{\\prime}}\\) as the weak model." + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.51, + 0.906, + 0.53 + ], + "angle": 0, + "content": "\\[\n\\hat {\\epsilon} _ {\\theta^ {\\prime}} ^ {\\lambda} \\left(\\mathbf {x} _ {t} | \\mathbf {c}\\right) = \\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t} | \\phi\\right) + \\lambda \\left(\\epsilon_ {\\theta^ {\\prime}} \\left(\\mathbf {x} _ {t} | \\mathbf {c}\\right) - \\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t} | \\phi\\right)\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.541, + 0.906, + 0.571 + ], + "angle": 0, + "content": "Despite being a simple extension, we found that this approach significantly improves subject fidelity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.572, + 0.907, + 0.829 + ], + "angle": 0, + "content": "Fig. 3 shows the rationale behind our personalization guidance. Since this simply modified guidance can steer the output toward the intersection space where both text fidelity and subject fidelity are high. As illustrated in Fig. 3, we explore the feasibility of identifying a superior space within the noisy data manifold. We hypothesize that an optimal point for text and subject fidelity could be located between CFG and PG, and we propose a method to facilitate the steering of guidance. Steering the guidance involves the adjustment of an appropriate weak model. By identifying a suitable weak model, we can steer \\(\\epsilon_{\\theta}(\\mathbf{x}_t|\\phi)\\) towards a more favorable point. We propose employing a weight interpolation to obtain an appropriate weak model. Specifically, by interpolating weights between the parameters of the pretrained and the fine-tuned models, we can modulate the degree of adaptation to the target data, thereby obtaining an appropriate weak model \\(\\epsilon_{\\theta_{\\omega}}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.841, + 0.905, + 0.859 + ], + "angle": 0, + "content": "\\[\n\\theta_ {\\omega} = \\omega \\theta^ {\\prime} + (1 - \\omega) \\theta , \\quad \\omega \\in [ 0, 1 ], \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where \\(\\omega\\) denotes the scale of weight interpolation. Increasing \\(\\omega\\) moves parameters closer to the fine-tuned model \\(\\theta^{\\prime}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15910" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.478, + 0.325 + ], + "angle": 0, + "content": "
MethodλDINOCLIP-ICLIP-T
SD 1.5DB-LoRA-0.37410.67970.2834
+ CFG7.50.47010.73490.3345
+ AG2.00.48310.74080.3236
+ Ours (ω = 0)7.50.49850.75160.3260
+ Ours (ω = 0.6)7.50.50240.75100.3288
SD 2.1DB-LoRA-0.42020.70760.2929
+ CFG7.50.49760.75190.3323
+ AG2.00.50720.75710.3212
+ Ours (ω = 0)7.50.52480.76550.3254
+ Ours (ω = 0.6)7.50.52810.76600.3277
SANADB-LoRA-0.41900.69390.3064
+ CFG4.50.48190.72910.3363
+ AG2.00.47920.72840.3355
+ Ours (ω = 0)4.50.53660.75220.3357
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.326, + 0.483, + 0.368 + ], + "angle": 0, + "content": "Table 1. Comparison with other guidance methods across different base diffusion models including SD1.5, SD2.1, and SANA. Here, we use DreamBooth-LoRA (DB-LoRA) only." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.373, + 0.483, + 0.404 + ], + "angle": 0, + "content": "reducing the loss \\(\\mathcal{L}\\) on \\(\\mathcal{D}_{\\mathrm{target}}\\). Finally, applying the weight interpolation to the weak model in the guidance:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.41, + 0.483, + 0.43 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\epsilon} _ {\\theta^ {\\prime}} ^ {\\lambda} (\\mathbf {x} _ {t} | \\mathbf {c}) = \\epsilon_ {\\theta_ {\\omega}} (\\mathbf {x} _ {t} | \\phi) + \\lambda \\bigl (\\epsilon_ {\\theta^ {\\prime}} (\\mathbf {x} _ {t} | \\mathbf {c}) - \\epsilon_ {\\theta_ {\\omega}} (\\mathbf {x} _ {t} | \\phi) \\bigr). (9)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.435, + 0.484, + 0.556 + ], + "angle": 0, + "content": "Note that \\(\\omega = 1\\) is equal to CFG. As shown in Fig. 4, it is possible to easily control the degree of subject and text fidelity. If \\(\\theta\\) includes only linear classifier, the personalization guidance with weight interpolation is equal to the combination of CFG and AG, which is the output-space interpolation. However, as neural networks are non-linear with respect to parameters, guidance with weight interpolation is different from the output interpolation of CFG and AG [36]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.568, + 0.224, + 0.585 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.592, + 0.292, + 0.609 + ], + "angle": 0, + "content": "4.1. Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Datasets. We conduct our experiments on the ViCo dataset [9], which includes 16 concepts and 31 text prompts. Each concept includes a small set of reference images (typically 4-7) and associated text prompts. Following the previous work [22], we fine-tune the pre-trained text-to-image diffusion models per concept and then generate 8 images per concept and text prompt, resulting in a total of 3,968 images for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Evaluation Metrics. Following the existing personalization studies [14, 22, 27], we evaluate subject and text fidelity. For the quantitative assessment, we adopt three evaluation metrics: (i) CLIP-T score [25], which uses CLIP's text-to-image similarity to evaluate text prompt alignment; (ii) CLIP-I score, which computes image-to-image similarity using CLIP embeddings to gauge how closely generated samples match reference images; and (iii) DINO score [3], which calculates image-to-image similarity between reference and generated images, offering a complementary perspective on subject fidelity." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.902, + 0.442 + ], + "angle": 0, + "content": "
MethodλDINOCLIP-ICLIP-T
SD 1.5DB-LoRA+TI-0.37250.67750.2817
+ CFG7.50.46180.72920.3316
+ SAG1.00.46000.72730.3307
+ AG2.00.48060.74080.3264
+ Ours (ω = 0)7.50.48140.74590.3233
+ Ours (ω = 0.7)7.50.48800.74610.3272
ClassDiffusion-0.37030.66690.2757
+ CFG7.50.52870.75170.3305
+ SAG1.00.51720.74650.3296
+ AG2.00.55700.76280.3280
+ Ours (ω = 0)7.50.57000.76560.3297
SD 2.1DB-LoRA+TI-0.45140.72510.2823
+ CFG7.50.52910.77490.3244
+ SAG1.00.53280.77660.3223
+ AG2.00.55570.78770.3206
+ Ours (ω = 0)7.50.56830.79180.3163
+ Ours (ω = 0.6)7.50.56890.79190.3185
ClassDiffusion-0.42730.69950.2843
+ CFG7.50.58400.77570.3277
+ SAG1.00.57830.77390.3244
+ AG2.00.60600.78560.3246
+ Ours (ω = 0)7.50.61660.78960.3246
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.442, + 0.907, + 0.526 + ], + "angle": 0, + "content": "Table 2. Comparison with other guidance methods across different base diffusion models including SD1.5 and SD2.1. Here, we use DreamBooth-LoRA (DB-LoRA) + Textual Inversion (TI) and ClassDiffusion. Since these personalization methods fine-tune text embedding, subject-agnostic guidance (SAG) is also compared with our method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.907, + 0.688 + ], + "angle": 0, + "content": "Implementation Details. In our experiments, we utilize three pre-trained text-to-image diffusion models as the base models: Stable Diffusion 1.5 (SD 1.5) [26], Stable Diffusion 2.1 (SD 2.1) [26], and SANA-1.6B [38]. SD 1.5 and SD 2.1 generate \\(512 \\times 512\\) resolution images, which comprise U-Net based architectures. On the other hand, SANA is a recent text-to-image diffusion model, which is based on the diffusion transformer [23] that can generate images up to \\(4096 \\times 4096\\) resolution. In the experiments, SANA generates \\(1024 \\times 1024\\) images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.809 + ], + "angle": 0, + "content": "To fine-tune the diffusion models, we leverage several personalization techniques: (i) DreamBoooth-LoRA (DB-LoRA) [13, 27], (ii) DB-LoRA + Textual Inversion (TI), and (iii) ClassDiffusion [14]. We fine-tune each model for up to 500 steps on the reference images from the ViCo dataset. We set the number of sampling steps (e.g., DDIM) between 20-50. Further details regarding additional hyperparameters are described in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Baselines. We evaluate our guidance method against other guidance techniques: (i) Classifier-Free Guidance (CFG) [10], (ii) Subject-Agnostic Guidance (SAG) [4], and (iii) AutoGuidance (AG) [18]. Since SAG and AG alone struggle to generate images that accurately reflect the text prompt, they are used in conjunction with CFG to produce" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15911" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.094, + 0.907, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.357, + 0.907, + 0.387 + ], + "angle": 0, + "content": "Figure 5. Comparison with other guidance techniques, including without guidance, CFG, CFG+SAG, and CFG+AG. These images are generated by the fine-tuned SD 2.1 using DB-LoRA and ClassDiffusion." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.395, + 0.907, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.624, + 0.871, + 0.638 + ], + "angle": 0, + "content": "Figure 6. Comparison with other guidance techniques. These images are generated by fine-tuned SANA [38] using DB-LoRA." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.483, + 0.687 + ], + "angle": 0, + "content": "images. Specifically, AG experiments are conducted by referring to recent work [19] that combines CFG and AG." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.698, + 0.184, + 0.712 + ], + "angle": 0, + "content": "4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Quantitative Results. Table 1 and Table 2 showcase the personalization performance on the ViCo dataset. Table 1 provides the quantitative results for DB-LoRA based on SD 1.5, SD 2.1, and SANA. Since DB-LoRA+TI and Class-Diffusion involve fine-tuning textual embeddings, we also evaluate subject-agnostic guidance (SAG) [4], which necessitates learned textual embeddings for specific concepts. As illustrated in both Table 1 and Table 2, our guidance method significantly enhances subject fidelity (DINO and CLIP-I scores), while maintaining text fidelity (CLIP-T score) with minimal change. Additionally, by adjusting \\(\\omega\\), the model can further optimize personalization performance across all" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.907, + 0.761 + ], + "angle": 0, + "content": "evaluation metrics. We select the optimal \\(\\omega\\) based on the best DINO and CLIP-I scores. For SANA with DB-LoRA and ClassDiffusion, the best subject fidelity was observed at \\(\\omega = 0\\). These results demonstrate that our approach significantly improves subject fidelity while maintaining text fidelity, and that steering the guidance enables a balanced trade-off between the two." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Interestingly, the impact of personalized guidance varies depending on the personalization method. ClassDiffusion [14] employs a semantic preservation loss that maintains the text embedding space, leading to better overall text fidelity retention. As a result, our guidance method maintains the CLIP-T score more effectively when applied to ClassDiffusion than to other personalization methods. This observation highlights a key insight: the better the fine-tuning direction, the more effective our guidance becomes." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15912" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.132, + 0.092, + 0.865, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.284, + 0.741, + 0.298 + ], + "angle": 0, + "content": "Figure 7. Generated images by changing \\(\\omega\\), where SANA with DB-LoRA is used." + }, + { + "type": "table", + "bbox": [ + 0.118, + 0.31, + 0.456, + 0.513 + ], + "angle": 0, + "content": "
MethodSubject Fidelity ↑Text Fidelity ↑
Stable Diffusion 2.1 + ClassDiffusion
CFG10.45%15.24%
CFG + SAG8.01%6.20%
CFG + AG12.60%6.20%
Ours55.22%32.04%
Undecided13.67%40.31%
SANA + DB-LoRA
CFG2.48%3.08%
CFG + AG2.17%2.77%
Ours68.32%8.64%
Undecided27.01%85.49%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.109, + 0.516, + 0.465, + 0.53 + ], + "angle": 0, + "content": "Table 3. User preference on subject fidelity and text fidelity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Qualitative Results. Fig. 5 and Fig. 6 show the comparison results based on SD 2.1 and SANA, respectively. As shown in Fig. 5, CFG often fails to capture fine details of the reference subject (e.g., cat's appearance and the teddybear's eye). Additionally, applying SAG or AG has minimal impact in some cases (first row of Fig. 5) or even degrades image quality, introducing artifacts such as a teddy bear with three eyes (CFG+AG, second row of Fig. 5). In contrast, our method consistently achieves better subject fidelity with the same noise input. Similarly, in Fig. 6, our method better preserves subject details compared to CFG and CFG+AG, as evidenced by the significant improvement in subject fidelity in quantitative results. For instance, it accurately retains the ears and color of a dog (first row of Fig. 6) and the shape and numerical details of a clock (second row of Fig. 6). In contrast, applying AG with CFG shows little to no improvement despite incurring a higher computational cost compared to both CFG and our guidance method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "User Study. Table 3 presents the user preference results comparing our method with baseline approaches. The results show that our guidance technique achieves significantly higher subject fidelity than other guidance methods. Our approach also shows superior text fidelity in user evaluations. Further details can be found in the supplementary." + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.308, + 0.815, + 0.328 + ], + "angle": 0, + "content": "Subject Fidelity (DINO) vs Text Fidelity (CLIP-T) (guidance scale \\(\\lambda \\in [2.5, 5.0]\\), step = 0.5)" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.328, + 0.9, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.556, + 0.906, + 0.641 + ], + "angle": 0, + "content": "Figure 8. Ablation study on the guidance scale \\(\\lambda \\in [2.5, 5.0]\\), measuring performance at 0.5 intervals using DB-LoRA based on SANA. When \\(\\omega\\) is set to 0.2, we can improve both subject and text fidelity compared to CFG, regardless of guidance scale \\(\\lambda\\). Note that CLIP-T score is slightly changed in the graph \\((max(CLIP-T) - min(CLIP-T) < 0.004)\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.649, + 0.682, + 0.665 + ], + "angle": 0, + "content": "4.3. Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.674, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Ablation Study on \\(\\omega\\). Fig. 4 illustrates how DINO and CLIP-T scores vary with changes in \\(\\omega\\). For both SD 1.5 and SD 2.1, personalization guidance slightly reduces text fidelity but significantly enhances subject fidelity compared to CFG (\\(\\omega = 1\\)). By appropriately adjusting \\(\\omega\\), it is possible to either boost text fidelity while maintaining subject fidelity (as indicated by the pink arrow) or enhance subject fidelity while preserving text fidelity (as indicated by the blue arrow). It is important to note that these trends may vary depending on the personalization technique or the pre-trained diffusion model. However, the ability to easily fine-tune results via weight interpolation at inference, without modifying training steps or learning rates, offers a major practical advantage over traditional methods for controlling overfitting. Fig. 7 further visualizes the impact of \\(\\omega\\) on gen" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15913" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.082, + 0.264, + 0.142 + ], + "angle": 0, + "content": "
MethodPickScore↑
DPO+CFG21.53
DPO+Ours21.58
" + }, + { + "type": "table", + "bbox": [ + 0.274, + 0.082, + 0.484, + 0.141 + ], + "angle": 0, + "content": "
MethodDreamSim↓
PairCustom+CFG0.7993
PairCustom+Ours0.7344
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.143, + 0.27, + 0.171 + ], + "angle": 0, + "content": "Table 4. Results of combining Diffusion-DPO with ours." + }, + { + "type": "table_caption", + "bbox": [ + 0.274, + 0.144, + 0.482, + 0.171 + ], + "angle": 0, + "content": "Table 5. Results of combining Pair-Custom with ours." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.18, + 0.475, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.484, + 0.48, + 0.497 + ], + "angle": 0, + "content": "Figure 9. Qualitative results of our guidance across diverse tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.483, + 0.582 + ], + "angle": 0, + "content": "eration quality. The first row shows that higher \\(\\omega\\) values retain fine details, such as the intricate patterns on the vase, at the cost of weaker adherence to the \"purple\" text prompt. Additionally, as \\(\\omega\\) decreases, subject fidelity improves significantly, particularly in the range of 0.6 to 1.0." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.484, + 0.734 + ], + "angle": 0, + "content": "Ablation Study on Guidance Scale. Fig. 8 shows how personalization performance varies with different guidance scales \\(\\lambda\\). We evaluate our guidance at \\(\\omega = 0\\) and \\(\\omega = 0.2\\), where text fidelity remains preserved. As \\(\\lambda\\) decreases, text fidelity deteriorates for both guidance methods. However, as \\(\\lambda\\) increases, our guidance method shows improvements in both the DINO score and CLIP-T score, whereas the DINO score of CFG declines. This suggests that CFG reduces subject fidelity, and adjusting the guidance scale alone is insufficient to correct this effect." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.743, + 0.466, + 0.76 + ], + "angle": 0, + "content": "4.4. Application to Fine-tuned Diffusion Models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Fine-tuning diffusion model extends beyond personalization tasks to applications such as style personalization [17], human preference learning [21, 33], and specific task learning [2, 15]. Our personalization guidance framework leverages the knowledge gap between the pre-trained and fine-tuned models, emphasizing the information better captured by the fine-tuned model. This concept can be generalized to control the degree of adaptation for other fine-tuned models, such as adjusting the extent of human preference alignment" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.787, + 0.107 + ], + "angle": 0, + "content": "or style personalization during inference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.107, + 0.905, + 0.182 + ], + "angle": 0, + "content": "To this end, to investigate our method's generalizability across diverse tasks, Fig. 9 show the following experiments: (a) Human Preference Optimization (Diffusion-DPO [33]), (b) Style Personalization (PairCustom [17]), and (c) Instruction-based Editing (InstructPix2Pix [2])." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.182, + 0.907, + 0.348 + ], + "angle": 0, + "content": "(a) Human Preference Optimization: For evaluation with the Diffusion-DPO model [33], we use the PartiPrompts dataset [39] and report PickScore [39], which reflects human-perceived image quality. Table 4 shows that integrating our method into the SD 1.5-based Diffusion-DPO model yields superior performance compared to CFG, confirming the effectiveness of our approach on human preference optimization tasks. As shown in Fig. 9 (a), our method helps generate more natural-looking images and corrects artifacts that would otherwise appear unnatural (e.g., additional finger)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.348, + 0.907, + 0.514 + ], + "angle": 0, + "content": "(b) Style Personalization: To evaluate style alignment, we use two pre-trained style-LoRAs released by PairCustom, combining each with both CFG and our method. Using text prompts from the StyleAlign [37] paper, we generate 800 images and compute the DreamSim score [7] to quantify the similarity between the target style and generated outputs. As shown in Table 5, applying our method significantly improves style alignment, demonstrating its effectiveness for style personalization. In Fig. 9 (b), combining PairCustom with our guidance leads to a more faithful transfer of the painting style from the reference image." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.515, + 0.906, + 0.59 + ], + "angle": 0, + "content": "(c) Instruction-based Editing: In Fig. 9 (c), our guidance improves the performance of the InstructPix2Pix [2], which is designed to edit images based on text instructions, to better follow the given instruction. As a result, the generated outputs are more accurately aligned with the intended edits." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.59, + 0.906, + 0.681 + ], + "angle": 0, + "content": "These results demonstrate that applying our guidance to various fine-tuned diffusion models allows them to generate images more faithfully aligned with their respective finetuning objectives. This highlights the generalizability of our method across diverse tasks. Further implementation details are provided in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.694, + 0.634, + 0.71 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This paper tackles the critical challenge of balancing adaptation to the target distribution and knowledge preservation from the original model in personalized diffusion models. Our approach simply modifies the existing guidance by adjusting the degree of unlearning in the weak model through weight interpolation between the pre-trained and fine-tuned models. This simple yet effective guidance enhances subject fidelity while preserving text editability, ensuring an optimal trajectory for personalization. Furthermore, we demonstrate that our guidance can be effectively applied to a wide range of fine-tuned diffusion models, such as human preference learning." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15914" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.092, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "Acknowledgement We would like to thank the Qualcomm AI Research team for their valuable discussions." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.137, + 0.188, + 0.152 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.162, + 0.483, + 0.244 + ], + "angle": 0, + "content": "[1] Donghoon Ahn, Hyoungwon Cho, Jaewon Min, Wooseok Jang, Jungwoo Kim, SeonHwa Kim, Hyun Hee Park, Kyong Hwan Jin, and Seungryong Kim. Self-rectifying diffusion sampling with perturbed-attention guidance. In European Conference on Computer Vision, pages 1-17. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.247, + 0.483, + 0.315 + ], + "angle": 0, + "content": "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18392-18402, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.318, + 0.483, + 0.387 + ], + "angle": 0, + "content": "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.389, + 0.484, + 0.457 + ], + "angle": 0, + "content": "[4] Kelvin CK Chan, Yang Zhao, Xuhui Jia, Ming-Hsuan Yang, and Huisheng Wang. Improving subject-driven image synthesis with subject-agnostic guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6733–6742, 2024. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.46, + 0.483, + 0.5 + ], + "angle": 0, + "content": "[5] Shangyu Chen, Zizheng Pan, Jianfei Cai, and Dinh Phung. Para: Personalizing text-to-image diffusion via parameter rank reduction. arXiv preprint arXiv:2406.05641, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.503, + 0.483, + 0.571 + ], + "angle": 0, + "content": "[6] Wenhu Chen, Hexiang Hu, Yandong Li, Nataniel Ruiz, Xuhui Jia, Ming-Wei Chang, and William W Cohen. Subject-driven text-to-image generation via apprenticeship learning. Advances in Neural Information Processing Systems, 36:30286–30305, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.574, + 0.483, + 0.641 + ], + "angle": 0, + "content": "[7] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. arXiv preprint arXiv:2306.09344, 2023.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.645, + 0.483, + 0.713 + ], + "angle": 0, + "content": "[8] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In The Eleventh International Conference on Learning Representations. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.716, + 0.483, + 0.769 + ], + "angle": 0, + "content": "[9] Shaozhe Hao, Kai Han, Shihao Zhao, and Kwan-Yee K Wong. Vico: Plug-and-play visual condition for personalized text-to-image generation. arXiv preprint arXiv:2306.00971, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.773, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[10] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.856 + ], + "angle": 0, + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seunglyon Kim. Improving sample quality of diffusion models using self-attention guidance. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.162, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "International Conference on Computer Vision, pages 7462-7471, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[13] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[14] Jiannan Huang, Jun Hao Liew, Hanshu Yan, Yuyang Yin, Yao Zhao, and Yunchao Wei. Classdiffusion: More aligned personalization tuning with explicit class guidance. arXiv preprint arXiv:2405.17532, 2024. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[15] Lianghua Huang, Wei Wang, Zhi-Fan Wu, Yupeng Shi, Huanzhang Dou, Chen Liang, Yutong Feng, Yu Liu, and Jingren Zhou. In-context lora for diffusion transformers. arXiv preprint arXiv:2410.23775, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.293, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[16] Junha Hyung, Kinam Kim, Susung Hong, Min-Jung Kim, and Jaegul Choo. Spatiotemporal skip guidance for enhanced video diffusion sampling. arXiv preprint arXiv:2411.18664, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.35, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[17] Maxwell Jones, Sheng-Yu Wang, Nupur Kumari, David Bau, and Jun-Yan Zhu. Customizing text-to-image models with a single image pair. In SIGGRAPH Asia 2024 Conference Papers, pages 1-13, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.407, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[18] Tero Karras, Miika Aittala, Tuomas Kynkänniemi, Jaakko Lehtinen, Timo Aila, and Samuli Laine. Guiding a diffusion model with a bad version of itself. Advances in Neural Information Processing Systems, 37:52996-53021, 2025. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.477, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[19] Artur Kasymov, Marcin Sendera, Michal Stypulkowski, Macej Zięba, and Przemysław Spurek. Autolora: Autoguidance meets low-rank adaptation for diffusion models. arXiv preprint arXiv:2410.03941, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.534, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[20] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1931-1941, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.605, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[21] Shufan Li, Konstantinos Kallidromitis, Akash Gokul, Yusuke Kato, and Kazuki Kozuka. Aligning diffusion models by optimizing human utility. Advances in Neural Information Processing Systems, 37:24897-24925, 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.661, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[22] Jisu Nam, Heesu Kim, DongJae Lee, Siyoon Jin, Seungryong Kim, and Seunggyu Chang. Dreamcatcher: appearance matching self-attention for semantically-consistent text-to-image personalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8100–8110, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[23] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195-4205, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.802, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[24] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15915" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.147 + ], + "angle": 0, + "content": "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.484, + 0.205 + ], + "angle": 0, + "content": "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.484, + 0.262 + ], + "angle": 0, + "content": "[27] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, pages 22500-22510, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.484, + 0.347 + ], + "angle": 0, + "content": "[28] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdream booth: Hypermetworks for fast personalization of text-to-image models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6527-6536, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.484, + 0.418 + ], + "angle": 0, + "content": "[29] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 35:36479-36494, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.484, + 0.488 + ], + "angle": 0, + "content": "[30] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8543-8552, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.484, + 0.546 + ], + "angle": 0, + "content": "[31] Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. Freeu: Free lunch in diffusion u-net. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4733-4743, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.484, + 0.589 + ], + "angle": 0, + "content": "[32] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.484, + 0.674 + ], + "angle": 0, + "content": "[33] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228-8238, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.484, + 0.758 + ], + "angle": 0, + "content": "[34] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15943-15953, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[35] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International conference on machine learning, pages 23965-23998. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[36] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "Namkoong, et al. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7959-7971, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[37] Zongze Wu, Yotam Nitzan, Eli Shechtman, and Dani Lischinski. Stylealign: Analysis and applications of aligned stylegan models. arXiv preprint arXiv:2110.11323, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[38] Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, et al. Sana: Efficient high-resolution image synthesis with linear diffusion transformers. arXiv preprint arXiv:2410.10629, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[39] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15916" + } + ] +] \ No newline at end of file diff --git a/2025/Steering Guidance for Personalized Text-to-Image Diffusion Models/34e2c957-5535-42df-a5b2-c4f7cd9d6450_origin.pdf b/2025/Steering Guidance for Personalized Text-to-Image Diffusion Models/34e2c957-5535-42df-a5b2-c4f7cd9d6450_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a236538b0e9c7c4c05c133d480c9c97625533e7a --- /dev/null +++ b/2025/Steering Guidance for Personalized Text-to-Image Diffusion Models/34e2c957-5535-42df-a5b2-c4f7cd9d6450_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:760ba7ecce2f249475ab0f894e35c468f686299ce049844c529292c0cc763edc +size 6521601 diff --git a/2025/Steering Guidance for Personalized Text-to-Image Diffusion Models/full.md b/2025/Steering Guidance for Personalized Text-to-Image Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..80c6d97eba22df7ab40d347a704475d7189dd636 --- /dev/null +++ b/2025/Steering Guidance for Personalized Text-to-Image Diffusion Models/full.md @@ -0,0 +1,309 @@ +# Steering Guidance for Personalized Text-to-Image Diffusion Models + +Sunghyun Park* Seokeon Choi* Hyoungwoo Park Sungrack Yun Qualcomm AI Research† + +{sunpar, seokchoi, hwoopark, sungrack}@qti=qualcomm.com + +![](images/23db2504435ec8c5e07cd7363b0933debf27211c658fddc7dd6f63cf5b626403.jpg) + +![](images/aac4c5beaae87aa768d036c94516d336fc2d62c597bac7c81b2e0e1609698162.jpg) +A $\star$ cat in a chef outfit + +![](images/7dae5fcbd0bec05bbf8130b6cdf42f0ae06f049200608ce42fec42b4a0914837.jpg) + +![](images/fff0b5ad4afed71db9bbc380374de9d6e20017c9f9415c4f38857ca605120bf9.jpg) + +![](images/9cedbbe1c3a7dcf4835071f82f91fd2a5982cf516be0380ff5f4547ca7b610ce.jpg) + +![](images/ed49c68166a3371a73a1db68dd3ff105df920ff28d5cb82ea5512f78f55581cb.jpg) + +![](images/119fa937dfe5f87617e53075cd78d8963591897ca23c47ced8268c54d6ca7e1a.jpg) +A $\star$ dog in a movie theater + +![](images/8ca313a3fd2efad70cddccc58943cef2fa6ef9d821e56fad310a1c768d276ac6.jpg) + +![](images/6e5a1016bb96b56f1c7cc5f3ed8116e58617475e19b6ad540b4ca961da886e05.jpg) + +![](images/cbfdfbd2181e8f59b59228a03aeef53ec85be25f8e84eb32496cf1989ff1085d.jpg) +Reference +Figure 1. Comparison of generated image quality between Classifier-Free Guidance (CFG) and our Personalization Guidance. + +![](images/f71a04de42d6201e407401c0d6691d3fc6469428b7bfff3acf5b634f8ce5c51d.jpg) + +![](images/8a5befc83390fb14d03018c0f1bf0a55ec5e9d05210da6a831e121b4fb100df9.jpg) +A duck toy with the Eiffel Towel in the background + +![](images/52eec6116cbdea1d926e6c41cf7403a25d3e8edbc24a692b9e55ee3d3c16c390.jpg) +A duck toy among the skyscrapers in New York city + +![](images/74ef1e47cef28154d4c886aa7597eefa719541e434c44e5551438f4888b8414e.jpg) + +# Abstract + +Personalizing text-to-image diffusion models is crucial for adapting the pre-trained models to specific target concepts, enabling diverse image generation. However, fine-tuning with few images introduces an inherent trade-off between aligning with the target distribution (e.g., subject fidelity) + +and preserving the broad knowledge of the original model (e.g., text editability). Existing sampling guidance methods, such as classifier-free guidance (CFG) and autoguidance (AG), fail to effectively guide the output toward well-balanced space: CFG restricts the adaptation to the target distribution, while AG compromises text alignment. To address these limitations, we propose personalization guidance, a simple yet effective method leveraging an unlearned weak model conditioned on a null text prompt. Moreover, our method dynamically controls the extent of unlearning + +in a weak model through weight interpolation between pretrained and fine-tuned models during inference. Unlike existing guidance methods, which depend solely on guidance scales, our method explicitly steers the outputs toward a balanced latent space without additional computational overhead. Experimental results demonstrate that our proposed guidance can improve text alignment and target distribution fidelity, integrating seamlessly with various fine-tuning strategies. + +# 1. Introduction + +Diffusion models [11, 32] have emerged as powerful generative models capable of representing complex data distributions, driving significant advancements in high-fidelity image generation [24, 26, 29, 38]. However, real-world use cases often demand personalized generation, where the model generates not only generic concepts but also specific concepts like a novel character or person [8, 20, 27]. This personalization technique is particularly useful for applications such as portrait editing and specific content creation. + +Many approaches have been proposed to personalize diffusion models for novel concepts. One line of work [6, 28, 30, 34] learns a general-purpose encoder capable of adapting to new concepts without test-time finetuning. However, this approach requires large-scale text-image datasets and considerable computational resources, and must often be retrained from scratch for each pre-trained model, making it expensive to adapt to the growing variety of text-to-image or text-to-video diffusion architectures. An alternative solution is to directly fine-tune a pre-trained diffusion model on the target images, which can be roughly categorized into fine-tuning low-rank adapters [5, 13], text embeddings [8], specific layers [14, 20], or the full model parameters [27]. This strategy can be applied to any pre-trained diffusion model and only requires a few target images, proving effective for personalization in various scenarios. + +Despite these advancements, optimization-based personalization remains a nontrivial challenge. When fine-tuning a model to a specific, often much smaller, target data distribution, there is an inevitable tradeoff between aligning with the target distribution (e.g., learning a novel concept with a few target images) and preserving broad knowledge of the original model (e.g., preserving text editability). This tradeoff manifests in various ways: a fine-tuned model for personalization may generate high-quality domain-specific images but lose text editability, making it difficult to generalize across different prompts. Conversely, a model that retains broad knowledge from its original training distribution may struggle to fully align with new concepts introduced in fine-tuning. Existing approaches attempt to mitigate this tradeoff by adjusting the number of fine-tuning iterations [8, 27], leveraging regularization [14, 27], reducing the model pa + +![](images/ae36c689934c7060d99a5fbf9318cdb4c55f44b49ddd1b30e6247d248d0fe863.jpg) +Figure 2. Motivation of Personalization Guidance. Classifier-free guidance [10] can interfere with subject fidelity due to its alignment with the text prompt. In contrast, autoguidance [18] fails to reflect the text prompt but generates the subject more similarly. Our guidance complements the limitations of both methods. + +![](images/bb9b851d5c3a3ac9ea1411833aa6e6d8639d714b23a8cf5d3b886b22841cdb78.jpg) + +![](images/6d21b32247e247b671aa7ca9ce6b146e58087326080219977d753112269c9cfe.jpg) + +![](images/1f73207d1918fa94259aa2f031677f7275b0051240112abeb37a68f3021e8ab7.jpg) + +rameters of updates [5, 13], yet they primarily operate at the fine-tuning stage. Moreover, balancing the trade-off between comprehensive text editability and precise alignment with a new target distribution remains an open problem. + +A parallel line of research focuses on sampling guidance techniques, such as Classifier-Free Guidance (CFG) [10] and AutoGuidance (AG) [18]. These methods rely on "weak models" to detect and push the outputs away from poor trajectories, thereby guiding samples toward higher-quality regions of the data manifold. For instance, CFG uses an unconditional model using the null text prompt as the weak model, while AG uses an unlearned model—one with fewer parameters or less training—conditioned on the same text prompt. However, these methods have limitations when applied to personalized diffusion models, as shown in Fig. 2. CFG can interfere with the model's adaptation to the target distribution, restricting its ability to fully utilize fine-tuned knowledge. Moreover, while the guidance scale of CFG adjusts the degree of text fidelity, it does not improve alignment with the target data. On the other hand, AG, which primarily enhances alignment with the target distribution, however, lacks an explicit mechanism for improving text alignment, making it unsuitable for fine-tuned text-to-image models. + +To address these issues, we propose a simple yet effective guidance technique, called Personalization Guidance, for personalized diffusion models, which can dynamically balance text alignment and adaptation to the target distribution. Our method leverages an "unlearned" weak model with a null text prompt, without incurring additional computational overhead during inference. Here, we aim to balance the alignment with the target distribution (e.g., subject fidelity) and the preservation of the broad knowledge of the original model (e.g., text fidelity) by controlling the extent of "unlearning" in a weak model. Here, the degree of unlearning refers to the extent to which the target data distribution has not been learned compared to the fine-tuned model. The rationale behind adjusting this degree of unlearning is to steer the output toward an optimal, well-balanced latent space. Specifically, during inference, we interpolate the weight parameters between the pre-trained and fine-tuned + +diffusion models to derive an optimal weak model, effectively controlling the extent of unlearning. Unlike guidance scales, which adjust the intensity of guidance, weight interpolation for the weak model serves to steer outputs in an appropriate direction. This straightforward approach can be readily integrated into various fine-tuning diffusion models. + +# 2. Related Work + +Personalization of Diffusion Models. Personalizing diffusion models aim to generate novel images by adapting to specific concepts using a set of reference images. Optimization-based methods personalize models by finetuning textual embeddings [8, 14, 20], entire model parameters with regularization [27], or low-rank adapters [5, 13]. Another line of research leverages extensive text-image datasets to train dedicated text-to-image personalization models, by fine-tuning additional modules [6, 28, 30, 34]. Recently, several studies [4, 22, 31] have achieved better personalization without additional fine-tuning or training. Orthogonal to these existing approaches, we introduce a simple yet effective personalization guidance technique that seamlessly integrates with various personalization methods by modifying the classifier-free guidance. Moreover, our guidance enables us to adjust the degree of subject and text fidelity without incurring additional computational costs. + +Guidance for Diffusion Models. Classifier-Free Guidance (CFG) [10] enhances conditional image generation by leveraging the unconditional model as a weak model to guide the outputs toward better adherence to the given condition. Various guidance techniques have been explored, such as the guidance using perturbations [1, 16] or blurring to the attention maps [12]. Autoguidance [18] improves generation by using a weak model—one with fewer parameters or less training—to push outputs toward a well-trained direction. While InstructPix2Pix [2] introduces separate CFG using both image and text conditions, it cannot be directly applied to DB-LoRA due to its reliance on image-conditioned training. For personalization, subject-agnostic guidance [4] enhances subject-agnostic attributes by employing a weak model conditioned on subject-specific text prompts. In this study, we aim to address the fundamental personalization trade-off—aligning a model with a smaller target data distribution while preserving the broad text fidelity of the original model. We derive a weak model that dynamically adjusts the degree of adaptation, balancing comprehensive text editability with precise alignment to new concepts. + +Weight Interpolation. There are previous studies on exploring the use of weight interpolation. WISE-FT [36] improves the robustness of fine-tuned classifiers against distribution shifts through weight-space ensembling. Model-soups [35] maximizes model accuracy and robustness by interpolating the weights of multiple fine-tuned classifiers trained with different hyperparameter configurations. + +# 3. Method + +# 3.1. Preliminaries + +Diffusion Models. Diffusion models [11, 32] aim to learn reversing a forward noising process that gradually transforms data samples into noise. Concretely, the forward process starts with a data distribution $p_0(\mathbf{x})$ and ends at a high-noise distribution $p_T(\mathbf{x}) \approx \mathcal{N}(\mathbf{0},\mathbf{I})$ . Sampling from the data distribution proceeds by solving the equivalent probability-flow ordinary differential equation (PF-ODE). Under the variance-preserving parameterization where $p(\mathbf{x}_t|\mathbf{x}_0) = \mathcal{N}(\mathbf{x}_0,\sigma_t^2\mathbf{I})$ , the PF-ODE takes the form: + +$$ +\mathrm {d} \mathbf {x} _ {t} = - \sigma_ {t} \nabla_ {\mathbf {x} _ {t}} \log p (\mathbf {x} _ {t}) \mathrm {d} t, \quad \mathbf {x} _ {T} \sim p _ {T} (\mathbf {x} _ {T}). \tag {1} +$$ + +In diffusion models, the score function is parameterized by $\epsilon_{\theta}$ . Here, the noise estimation of the diffusion model can be considered as $\epsilon_{\theta}(\mathbf{x}_t) \approx -\sigma_t \nabla_{\mathbf{x}_t} \log p(\mathbf{x}_t)$ . + +Personalization. Suppose we have a pre-trained diffusion model $\epsilon_{\theta}$ , where $\theta$ is the parameters of the model. We wish to personalize it to a smaller target dataset $\mathcal{D}_{\mathrm{target}} = \left\{(\mathbf{x}^i, \mathbf{c}^i)\right\}_{i=1}^N$ , where each $\mathbf{x}^i$ and $\mathbf{c}^i$ is an image and its associated text prompt, respectively. The goal is to obtain new parameters $\theta'$ so that the fine-tuned model $\epsilon_{\theta'}$ can generate samples more faithfully matched the $\mathcal{D}_{\mathrm{target}}$ . + +Fine-tuning typically follows the same denoising objective used by the original diffusion model [11]. Concretely, let $\sigma$ be a noise level sampled from a training schedule, and let $\epsilon \sim \mathcal{N}(\mathbf{0},\sigma^2\mathbf{I})$ . For each $(\mathbf{x},\mathbf{c})$ in $\mathcal{D}_{\mathrm{target}}$ , we form the noisy input $\mathbf{x}_t = \mathbf{x} + \epsilon$ and minimize: + +$$ +\mathcal {L} = \mathbb {E} _ {(\mathbf {x}, \mathbf {c}) \sim \mathcal {D} _ {\text {t a r g e t}}, \sigma , \boldsymbol {\epsilon}} \left\| \epsilon_ {\theta^ {\prime}} (\mathbf {x} _ {t}, \sigma , \mathbf {c}) - \boldsymbol {\epsilon} \right\| _ {2} ^ {2}. \tag {2} +$$ + +This teaches the model to reconstruct $\mathbf{x}$ from noisy inputs in the domain of $\mathcal{D}_{\mathrm{target}}$ , thereby adapting it to the new concept. In practice, additional regularization terms are often used to prevent overfitting [14, 27]. + +Classifier-Free Guidance. To enhance the generation towards an arbitrary condition $\mathbf{c}$ , classifier-free guidance (CFG) [10] introduces a new sampling distribution $p^{\lambda}(\mathbf{x}|\mathbf{c})$ composed with both $p(\mathbf{x})$ and the classifier distribution $p(c|\mathbf{x})^{\lambda}$ , which is described as: + +$$ +p ^ {\lambda} (\mathbf {x} | \mathbf {c}) \propto p (\mathbf {x}) p (\mathbf {c} | \mathbf {x}) ^ {\lambda}, \tag {3} +$$ + +where $\lambda > 1$ amplifies the conditional likelihood. By applying Bayes' rule for some timestep $t$ , the corresponding score function is: + +$$ +\begin{array}{l} \nabla_ {\mathbf {x} _ {t}} \log p ^ {\lambda} (\mathbf {x} _ {t} | \mathbf {c}) = \nabla_ {\mathbf {x} _ {t}} \log p (\mathbf {x} _ {t}) \\ + \lambda \left(\nabla_ {\mathbf {x} _ {t}} \log p (\mathbf {x} _ {t} | \mathbf {c}) - \nabla_ {\mathbf {x} _ {t}} \log p (\mathbf {x} _ {t})\right). \tag {4} \\ \end{array} +$$ + +To enable conditional generation such as text prompts and class labels, CFG jointly trains the unconditional model + +![](images/f9f9e426c6ceb3e0cc05f6b58e917c2a3355395365832500f44d8242372c3b0a.jpg) +Figure 3. Comparison between Classifier-Free Guidance (CFG), AutoGuidance (AG), and our Personalization Guidance (PG), with the band conceptually representing the noisy data manifold. By leveraging weight interpolation, guidance can be steered toward the optimal space between CFG and PG. + +$\epsilon_{\theta}(\mathbf{x}_t,\phi)$ and the conditional model $\epsilon_{\theta}(\mathbf{x}_t,\mathbf{c})$ within a single model by dropping $\mathbf{c}$ with certain probability to allow null conditioning with $c = \phi$ . In the personalization setting, parameterizing the score function with the fine-tuned diffusion model $\epsilon_{\theta'}$ is described as: + +$$ +\hat {\epsilon} _ {\theta^ {\prime}} ^ {\lambda} \left(\mathbf {x} _ {t} \mid \mathbf {c}\right) = \epsilon_ {\theta^ {\prime}} \left(\mathbf {x} _ {t} \mid \phi\right) + \lambda \left(\epsilon_ {\theta^ {\prime}} \left(\mathbf {x} _ {t} \mid \mathbf {c}\right) - \epsilon_ {\theta^ {\prime}} \left(\mathbf {x} _ {t} \mid \phi\right)\right). \tag {5} +$$ + +However, while CFG improves alignment with the conditions, such as the text prompt, it struggles to enhance adaptation to the target data in the personalization task. + +AutoGuidance. AutoGuidance (AG) [18] is another guidance technique that employs a weaker version of the same model to identify suboptimal generation trajectories. Rather than using the unconditional model $\epsilon_{\theta}(\mathbf{x}_t|\phi)$ , AG directly guides a model with the weak model $\epsilon_{\theta_{weak}}$ , which is trained on the same task and conditioning, but degraded by under-training or less-parameters. However, obtaining a weaker version $\epsilon_{\theta_{weak}}$ of a fully trained diffusion model like Stable Diffusion [26] is challenging in practice. Instead, we found that a less trained model could serve as the pre-trained model for personalization, where $\theta_{weak} = \theta$ , effectively shifts the output away in the direction of fine-tuning, enhancing the adaptation to the target data. In the personalization, AG is formulated as: + +$$ +\tilde {\epsilon} _ {\theta^ {\prime}} ^ {\lambda} \left(\mathbf {x} _ {t} | \mathbf {c}\right) = \epsilon_ {\theta} \left(\mathbf {x} _ {t} | \mathbf {c}\right) + \lambda \left(\epsilon_ {\theta^ {\prime}} \left(\mathbf {x} _ {t} | \mathbf {c}\right) - \epsilon_ {\theta} \left(\mathbf {x} _ {t} | \mathbf {c}\right)\right). \tag {6} +$$ + +However, this approach has a major drawback—it scarcely reflects the text prompt, as shown in Fig. 2. + +# 3.2. Steering Personalization Guidance + +Inspired by autoguidance [18], we simply modify the CFG [10], by leveraging a pre-trained diffusion model $\epsilon_{\theta}$ + +![](images/c12daafe985ba09788877b81777b62340a8302bd27ec89c843108668947cdf82.jpg) +Subject Fidelity (DINO) vs Text Fidelity (CLIP-T) (weight interpolation scale $\omega \in [0,1]$ , step $= 0.1$ ) +Figure 4. Ablation study on the weight interpolation scale $\omega \in [0.0,1.0]$ with measurements taken at intervals of 0.1. Here, we use DreamBooth (DB) LoRA based on SD 1.5, SD 2.1, and SANA. By adjusting $\omega$ , it is possible to improve subject fidelity or text fidelity, while preserving the other. + +instead of the fine-tuned model $\epsilon_{\theta^{\prime}}$ as the weak model. + +$$ +\hat {\epsilon} _ {\theta^ {\prime}} ^ {\lambda} \left(\mathbf {x} _ {t} | \mathbf {c}\right) = \epsilon_ {\theta} \left(\mathbf {x} _ {t} | \phi\right) + \lambda \left(\epsilon_ {\theta^ {\prime}} \left(\mathbf {x} _ {t} | \mathbf {c}\right) - \epsilon_ {\theta} \left(\mathbf {x} _ {t} | \phi\right)\right). \tag {7} +$$ + +Despite being a simple extension, we found that this approach significantly improves subject fidelity. + +Fig. 3 shows the rationale behind our personalization guidance. Since this simply modified guidance can steer the output toward the intersection space where both text fidelity and subject fidelity are high. As illustrated in Fig. 3, we explore the feasibility of identifying a superior space within the noisy data manifold. We hypothesize that an optimal point for text and subject fidelity could be located between CFG and PG, and we propose a method to facilitate the steering of guidance. Steering the guidance involves the adjustment of an appropriate weak model. By identifying a suitable weak model, we can steer $\epsilon_{\theta}(\mathbf{x}_t|\phi)$ towards a more favorable point. We propose employing a weight interpolation to obtain an appropriate weak model. Specifically, by interpolating weights between the parameters of the pretrained and the fine-tuned models, we can modulate the degree of adaptation to the target data, thereby obtaining an appropriate weak model $\epsilon_{\theta_{\omega}}$ . + +$$ +\theta_ {\omega} = \omega \theta^ {\prime} + (1 - \omega) \theta , \quad \omega \in [ 0, 1 ], \tag {8} +$$ + +where $\omega$ denotes the scale of weight interpolation. Increasing $\omega$ moves parameters closer to the fine-tuned model $\theta^{\prime}$ + +
MethodλDINOCLIP-ICLIP-T
SD 1.5DB-LoRA-0.37410.67970.2834
+ CFG7.50.47010.73490.3345
+ AG2.00.48310.74080.3236
+ Ours (ω = 0)7.50.49850.75160.3260
+ Ours (ω = 0.6)7.50.50240.75100.3288
SD 2.1DB-LoRA-0.42020.70760.2929
+ CFG7.50.49760.75190.3323
+ AG2.00.50720.75710.3212
+ Ours (ω = 0)7.50.52480.76550.3254
+ Ours (ω = 0.6)7.50.52810.76600.3277
SANADB-LoRA-0.41900.69390.3064
+ CFG4.50.48190.72910.3363
+ AG2.00.47920.72840.3355
+ Ours (ω = 0)4.50.53660.75220.3357
+ +Table 1. Comparison with other guidance methods across different base diffusion models including SD1.5, SD2.1, and SANA. Here, we use DreamBooth-LoRA (DB-LoRA) only. + +reducing the loss $\mathcal{L}$ on $\mathcal{D}_{\mathrm{target}}$ . Finally, applying the weight interpolation to the weak model in the guidance: + +$$ +\tilde {\epsilon} _ {\theta^ {\prime}} ^ {\lambda} (\mathbf {x} _ {t} | \mathbf {c}) = \epsilon_ {\theta_ {\omega}} (\mathbf {x} _ {t} | \phi) + \lambda \bigl (\epsilon_ {\theta^ {\prime}} (\mathbf {x} _ {t} | \mathbf {c}) - \epsilon_ {\theta_ {\omega}} (\mathbf {x} _ {t} | \phi) \bigr). (9) +$$ + +Note that $\omega = 1$ is equal to CFG. As shown in Fig. 4, it is possible to easily control the degree of subject and text fidelity. If $\theta$ includes only linear classifier, the personalization guidance with weight interpolation is equal to the combination of CFG and AG, which is the output-space interpolation. However, as neural networks are non-linear with respect to parameters, guidance with weight interpolation is different from the output interpolation of CFG and AG [36]. + +# 4. Experiments + +# 4.1. Experimental Setting + +Datasets. We conduct our experiments on the ViCo dataset [9], which includes 16 concepts and 31 text prompts. Each concept includes a small set of reference images (typically 4-7) and associated text prompts. Following the previous work [22], we fine-tune the pre-trained text-to-image diffusion models per concept and then generate 8 images per concept and text prompt, resulting in a total of 3,968 images for evaluation. + +Evaluation Metrics. Following the existing personalization studies [14, 22, 27], we evaluate subject and text fidelity. For the quantitative assessment, we adopt three evaluation metrics: (i) CLIP-T score [25], which uses CLIP's text-to-image similarity to evaluate text prompt alignment; (ii) CLIP-I score, which computes image-to-image similarity using CLIP embeddings to gauge how closely generated samples match reference images; and (iii) DINO score [3], which calculates image-to-image similarity between reference and generated images, offering a complementary perspective on subject fidelity. + +
MethodλDINOCLIP-ICLIP-T
SD 1.5DB-LoRA+TI-0.37250.67750.2817
+ CFG7.50.46180.72920.3316
+ SAG1.00.46000.72730.3307
+ AG2.00.48060.74080.3264
+ Ours (ω = 0)7.50.48140.74590.3233
+ Ours (ω = 0.7)7.50.48800.74610.3272
ClassDiffusion-0.37030.66690.2757
+ CFG7.50.52870.75170.3305
+ SAG1.00.51720.74650.3296
+ AG2.00.55700.76280.3280
+ Ours (ω = 0)7.50.57000.76560.3297
SD 2.1DB-LoRA+TI-0.45140.72510.2823
+ CFG7.50.52910.77490.3244
+ SAG1.00.53280.77660.3223
+ AG2.00.55570.78770.3206
+ Ours (ω = 0)7.50.56830.79180.3163
+ Ours (ω = 0.6)7.50.56890.79190.3185
ClassDiffusion-0.42730.69950.2843
+ CFG7.50.58400.77570.3277
+ SAG1.00.57830.77390.3244
+ AG2.00.60600.78560.3246
+ Ours (ω = 0)7.50.61660.78960.3246
+ +Table 2. Comparison with other guidance methods across different base diffusion models including SD1.5 and SD2.1. Here, we use DreamBooth-LoRA (DB-LoRA) + Textual Inversion (TI) and ClassDiffusion. Since these personalization methods fine-tune text embedding, subject-agnostic guidance (SAG) is also compared with our method. + +Implementation Details. In our experiments, we utilize three pre-trained text-to-image diffusion models as the base models: Stable Diffusion 1.5 (SD 1.5) [26], Stable Diffusion 2.1 (SD 2.1) [26], and SANA-1.6B [38]. SD 1.5 and SD 2.1 generate $512 \times 512$ resolution images, which comprise U-Net based architectures. On the other hand, SANA is a recent text-to-image diffusion model, which is based on the diffusion transformer [23] that can generate images up to $4096 \times 4096$ resolution. In the experiments, SANA generates $1024 \times 1024$ images. + +To fine-tune the diffusion models, we leverage several personalization techniques: (i) DreamBoooth-LoRA (DB-LoRA) [13, 27], (ii) DB-LoRA + Textual Inversion (TI), and (iii) ClassDiffusion [14]. We fine-tune each model for up to 500 steps on the reference images from the ViCo dataset. We set the number of sampling steps (e.g., DDIM) between 20-50. Further details regarding additional hyperparameters are described in the supplementary. + +Baselines. We evaluate our guidance method against other guidance techniques: (i) Classifier-Free Guidance (CFG) [10], (ii) Subject-Agnostic Guidance (SAG) [4], and (iii) AutoGuidance (AG) [18]. Since SAG and AG alone struggle to generate images that accurately reflect the text prompt, they are used in conjunction with CFG to produce + +![](images/1d705a80131c1bae71e8c79f1a6b8494f5312ce07dd18e15e4dac37e4555b5e6.jpg) +Figure 5. Comparison with other guidance techniques, including without guidance, CFG, CFG+SAG, and CFG+AG. These images are generated by the fine-tuned SD 2.1 using DB-LoRA and ClassDiffusion. + +![](images/50c376ece40a93b7ade99a7c6ab85e351bf8e8c86ddc1916c68038b40127be71.jpg) +Figure 6. Comparison with other guidance techniques. These images are generated by fine-tuned SANA [38] using DB-LoRA. + +images. Specifically, AG experiments are conducted by referring to recent work [19] that combines CFG and AG. + +# 4.2. Results + +Quantitative Results. Table 1 and Table 2 showcase the personalization performance on the ViCo dataset. Table 1 provides the quantitative results for DB-LoRA based on SD 1.5, SD 2.1, and SANA. Since DB-LoRA+TI and Class-Diffusion involve fine-tuning textual embeddings, we also evaluate subject-agnostic guidance (SAG) [4], which necessitates learned textual embeddings for specific concepts. As illustrated in both Table 1 and Table 2, our guidance method significantly enhances subject fidelity (DINO and CLIP-I scores), while maintaining text fidelity (CLIP-T score) with minimal change. Additionally, by adjusting $\omega$ , the model can further optimize personalization performance across all + +evaluation metrics. We select the optimal $\omega$ based on the best DINO and CLIP-I scores. For SANA with DB-LoRA and ClassDiffusion, the best subject fidelity was observed at $\omega = 0$ . These results demonstrate that our approach significantly improves subject fidelity while maintaining text fidelity, and that steering the guidance enables a balanced trade-off between the two. + +Interestingly, the impact of personalized guidance varies depending on the personalization method. ClassDiffusion [14] employs a semantic preservation loss that maintains the text embedding space, leading to better overall text fidelity retention. As a result, our guidance method maintains the CLIP-T score more effectively when applied to ClassDiffusion than to other personalization methods. This observation highlights a key insight: the better the fine-tuning direction, the more effective our guidance becomes. + +![](images/17e0c5f46402d9a73a2f9565946968999bb38872429f0f2a1a45f7f67f3dbf86.jpg) +Figure 7. Generated images by changing $\omega$ , where SANA with DB-LoRA is used. + +
MethodSubject Fidelity ↑Text Fidelity ↑
Stable Diffusion 2.1 + ClassDiffusion
CFG10.45%15.24%
CFG + SAG8.01%6.20%
CFG + AG12.60%6.20%
Ours55.22%32.04%
Undecided13.67%40.31%
SANA + DB-LoRA
CFG2.48%3.08%
CFG + AG2.17%2.77%
Ours68.32%8.64%
Undecided27.01%85.49%
+ +Table 3. User preference on subject fidelity and text fidelity. + +Qualitative Results. Fig. 5 and Fig. 6 show the comparison results based on SD 2.1 and SANA, respectively. As shown in Fig. 5, CFG often fails to capture fine details of the reference subject (e.g., cat's appearance and the teddybear's eye). Additionally, applying SAG or AG has minimal impact in some cases (first row of Fig. 5) or even degrades image quality, introducing artifacts such as a teddy bear with three eyes (CFG+AG, second row of Fig. 5). In contrast, our method consistently achieves better subject fidelity with the same noise input. Similarly, in Fig. 6, our method better preserves subject details compared to CFG and CFG+AG, as evidenced by the significant improvement in subject fidelity in quantitative results. For instance, it accurately retains the ears and color of a dog (first row of Fig. 6) and the shape and numerical details of a clock (second row of Fig. 6). In contrast, applying AG with CFG shows little to no improvement despite incurring a higher computational cost compared to both CFG and our guidance method. + +User Study. Table 3 presents the user preference results comparing our method with baseline approaches. The results show that our guidance technique achieves significantly higher subject fidelity than other guidance methods. Our approach also shows superior text fidelity in user evaluations. Further details can be found in the supplementary. + +![](images/1a917af5ed98a8aec4c07d008536465cd3e144fa6ba21d34b16f34e6522c3167.jpg) +Subject Fidelity (DINO) vs Text Fidelity (CLIP-T) (guidance scale $\lambda \in [2.5, 5.0]$ , step = 0.5) +Figure 8. Ablation study on the guidance scale $\lambda \in [2.5, 5.0]$ , measuring performance at 0.5 intervals using DB-LoRA based on SANA. When $\omega$ is set to 0.2, we can improve both subject and text fidelity compared to CFG, regardless of guidance scale $\lambda$ . Note that CLIP-T score is slightly changed in the graph $(max(CLIP-T) - min(CLIP-T) < 0.004)$ . + +# 4.3. Further Analysis + +Ablation Study on $\omega$ . Fig. 4 illustrates how DINO and CLIP-T scores vary with changes in $\omega$ . For both SD 1.5 and SD 2.1, personalization guidance slightly reduces text fidelity but significantly enhances subject fidelity compared to CFG ( $\omega = 1$ ). By appropriately adjusting $\omega$ , it is possible to either boost text fidelity while maintaining subject fidelity (as indicated by the pink arrow) or enhance subject fidelity while preserving text fidelity (as indicated by the blue arrow). It is important to note that these trends may vary depending on the personalization technique or the pre-trained diffusion model. However, the ability to easily fine-tune results via weight interpolation at inference, without modifying training steps or learning rates, offers a major practical advantage over traditional methods for controlling overfitting. Fig. 7 further visualizes the impact of $\omega$ on gen + +
MethodPickScore↑
DPO+CFG21.53
DPO+Ours21.58
+ +
MethodDreamSim↓
PairCustom+CFG0.7993
PairCustom+Ours0.7344
+ +Table 4. Results of combining Diffusion-DPO with ours. + +Table 5. Results of combining Pair-Custom with ours. + +![](images/c2b3fb3b5cfe9e7d732fe412e893c8b52684fff8cbef965c1f757c093092cb5d.jpg) +Figure 9. Qualitative results of our guidance across diverse tasks. + +eration quality. The first row shows that higher $\omega$ values retain fine details, such as the intricate patterns on the vase, at the cost of weaker adherence to the "purple" text prompt. Additionally, as $\omega$ decreases, subject fidelity improves significantly, particularly in the range of 0.6 to 1.0. + +Ablation Study on Guidance Scale. Fig. 8 shows how personalization performance varies with different guidance scales $\lambda$ . We evaluate our guidance at $\omega = 0$ and $\omega = 0.2$ , where text fidelity remains preserved. As $\lambda$ decreases, text fidelity deteriorates for both guidance methods. However, as $\lambda$ increases, our guidance method shows improvements in both the DINO score and CLIP-T score, whereas the DINO score of CFG declines. This suggests that CFG reduces subject fidelity, and adjusting the guidance scale alone is insufficient to correct this effect. + +# 4.4. Application to Fine-tuned Diffusion Models. + +Fine-tuning diffusion model extends beyond personalization tasks to applications such as style personalization [17], human preference learning [21, 33], and specific task learning [2, 15]. Our personalization guidance framework leverages the knowledge gap between the pre-trained and fine-tuned models, emphasizing the information better captured by the fine-tuned model. This concept can be generalized to control the degree of adaptation for other fine-tuned models, such as adjusting the extent of human preference alignment + +or style personalization during inference. + +To this end, to investigate our method's generalizability across diverse tasks, Fig. 9 show the following experiments: (a) Human Preference Optimization (Diffusion-DPO [33]), (b) Style Personalization (PairCustom [17]), and (c) Instruction-based Editing (InstructPix2Pix [2]). + +(a) Human Preference Optimization: For evaluation with the Diffusion-DPO model [33], we use the PartiPrompts dataset [39] and report PickScore [39], which reflects human-perceived image quality. Table 4 shows that integrating our method into the SD 1.5-based Diffusion-DPO model yields superior performance compared to CFG, confirming the effectiveness of our approach on human preference optimization tasks. As shown in Fig. 9 (a), our method helps generate more natural-looking images and corrects artifacts that would otherwise appear unnatural (e.g., additional finger). + +(b) Style Personalization: To evaluate style alignment, we use two pre-trained style-LoRAs released by PairCustom, combining each with both CFG and our method. Using text prompts from the StyleAlign [37] paper, we generate 800 images and compute the DreamSim score [7] to quantify the similarity between the target style and generated outputs. As shown in Table 5, applying our method significantly improves style alignment, demonstrating its effectiveness for style personalization. In Fig. 9 (b), combining PairCustom with our guidance leads to a more faithful transfer of the painting style from the reference image. + +(c) Instruction-based Editing: In Fig. 9 (c), our guidance improves the performance of the InstructPix2Pix [2], which is designed to edit images based on text instructions, to better follow the given instruction. As a result, the generated outputs are more accurately aligned with the intended edits. + +These results demonstrate that applying our guidance to various fine-tuned diffusion models allows them to generate images more faithfully aligned with their respective finetuning objectives. This highlights the generalizability of our method across diverse tasks. Further implementation details are provided in the supplementary. + +# 5. Conclusion + +This paper tackles the critical challenge of balancing adaptation to the target distribution and knowledge preservation from the original model in personalized diffusion models. Our approach simply modifies the existing guidance by adjusting the degree of unlearning in the weak model through weight interpolation between the pre-trained and fine-tuned models. This simple yet effective guidance enhances subject fidelity while preserving text editability, ensuring an optimal trajectory for personalization. Furthermore, we demonstrate that our guidance can be effectively applied to a wide range of fine-tuned diffusion models, such as human preference learning. + +Acknowledgement We would like to thank the Qualcomm AI Research team for their valuable discussions. + +# References + +[1] Donghoon Ahn, Hyoungwon Cho, Jaewon Min, Wooseok Jang, Jungwoo Kim, SeonHwa Kim, Hyun Hee Park, Kyong Hwan Jin, and Seungryong Kim. Self-rectifying diffusion sampling with perturbed-attention guidance. In European Conference on Computer Vision, pages 1-17. Springer, 2024. 3 +[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18392-18402, 2023. 3, 8 +[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 5 +[4] Kelvin CK Chan, Yang Zhao, Xuhui Jia, Ming-Hsuan Yang, and Huisheng Wang. Improving subject-driven image synthesis with subject-agnostic guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6733–6742, 2024. 3, 5, 6 +[5] Shangyu Chen, Zizheng Pan, Jianfei Cai, and Dinh Phung. Para: Personalizing text-to-image diffusion via parameter rank reduction. arXiv preprint arXiv:2406.05641, 2024. 2, 3 +[6] Wenhu Chen, Hexiang Hu, Yandong Li, Nataniel Ruiz, Xuhui Jia, Ming-Wei Chang, and William W Cohen. Subject-driven text-to-image generation via apprenticeship learning. Advances in Neural Information Processing Systems, 36:30286–30305, 2023. 2, 3 +[7] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. arXiv preprint arXiv:2306.09344, 2023.8 +[8] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In The Eleventh International Conference on Learning Representations. 2, 3 +[9] Shaozhe Hao, Kai Han, Shihao Zhao, and Kwan-Yee K Wong. Vico: Plug-and-play visual condition for personalized text-to-image generation. arXiv preprint arXiv:2306.00971, 2023.5 +[10] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 2, 3, 4, 5 +[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 2, 3 +[12] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seunglyon Kim. Improving sample quality of diffusion models using self-attention guidance. In Proceedings of the IEEE/CVF + +International Conference on Computer Vision, pages 7462-7471, 2023. 3 +[13] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 2, 3, 5 +[14] Jiannan Huang, Jun Hao Liew, Hanshu Yan, Yuyang Yin, Yao Zhao, and Yunchao Wei. Classdiffusion: More aligned personalization tuning with explicit class guidance. arXiv preprint arXiv:2405.17532, 2024. 2, 3, 5, 6 +[15] Lianghua Huang, Wei Wang, Zhi-Fan Wu, Yupeng Shi, Huanzhang Dou, Chen Liang, Yutong Feng, Yu Liu, and Jingren Zhou. In-context lora for diffusion transformers. arXiv preprint arXiv:2410.23775, 2024. 8 +[16] Junha Hyung, Kinam Kim, Susung Hong, Min-Jung Kim, and Jaegul Choo. Spatiotemporal skip guidance for enhanced video diffusion sampling. arXiv preprint arXiv:2411.18664, 2024.3 +[17] Maxwell Jones, Sheng-Yu Wang, Nupur Kumari, David Bau, and Jun-Yan Zhu. Customizing text-to-image models with a single image pair. In SIGGRAPH Asia 2024 Conference Papers, pages 1-13, 2024. 8 +[18] Tero Karras, Miika Aittala, Tuomas Kynkänniemi, Jaakko Lehtinen, Timo Aila, and Samuli Laine. Guiding a diffusion model with a bad version of itself. Advances in Neural Information Processing Systems, 37:52996-53021, 2025. 2, 3, 4, 5 +[19] Artur Kasymov, Marcin Sendera, Michal Stypulkowski, Macej Zięba, and Przemysław Spurek. Autolora: Autoguidance meets low-rank adaptation for diffusion models. arXiv preprint arXiv:2410.03941, 2024. 6 +[20] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1931-1941, 2023. 2, 3 +[21] Shufan Li, Konstantinos Kallidromitis, Akash Gokul, Yusuke Kato, and Kazuki Kozuka. Aligning diffusion models by optimizing human utility. Advances in Neural Information Processing Systems, 37:24897-24925, 2025. 8 +[22] Jisu Nam, Heesu Kim, DongJae Lee, Siyoon Jin, Seungryong Kim, and Seunggyu Chang. Dreamcatcher: appearance matching self-attention for semantically-consistent text-to-image personalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8100–8110, 2024. 3, 5 +[23] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195-4205, 2023. 5 +[24] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2 +[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, + +Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5 +[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2, 4, 5 +[27] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, pages 22500-22510, 2023. 2, 3, 5 +[28] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdream booth: Hypermetworks for fast personalization of text-to-image models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6527-6536, 2024. 2, 3 +[29] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 35:36479-36494, 2022. 2 +[30] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8543-8552, 2024. 2, 3 +[31] Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. Freeu: Free lunch in diffusion u-net. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4733-4743, 2024. 3 +[32] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 3 +[33] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228-8238, 2024. 8 +[34] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15943-15953, 2023. 2, 3 +[35] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International conference on machine learning, pages 23965-23998. PMLR, 2022. 3 +[36] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok + +Namkoong, et al. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7959-7971, 2022. 3, 5 +[37] Zongze Wu, Yotam Nitzan, Eli Shechtman, and Dani Lischinski. Stylealign: Analysis and applications of aligned stylegan models. arXiv preprint arXiv:2110.11323, 2021. 8 +[38] Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, et al. Sana: Efficient high-resolution image synthesis with linear diffusion transformers. arXiv preprint arXiv:2410.10629, 2024. 2, 5, 6 +[39] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 8 \ No newline at end of file diff --git a/2025/Steering Guidance for Personalized Text-to-Image Diffusion Models/images.zip b/2025/Steering Guidance for Personalized Text-to-Image Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f2a426f81daca10f05105d4c810dea6b617f9c32 --- /dev/null +++ b/2025/Steering Guidance for Personalized Text-to-Image Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9b75c994eda02a9aaadac28d5f3ab9d62b1bf52b33833dcdaf9fa59ea3b7b734 +size 1092843 diff --git a/2025/Steering Guidance for Personalized Text-to-Image Diffusion Models/layout.json b/2025/Steering Guidance for Personalized Text-to-Image Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d90f8a39578578cfa1a4f51107aa1b84c66e8777 --- /dev/null +++ b/2025/Steering Guidance for Personalized Text-to-Image Diffusion Models/layout.json @@ -0,0 +1,8217 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 96, + 103, + 515, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 103, + 515, + 121 + ], + "spans": [ + { + "bbox": [ + 96, + 103, + 515, + 121 + ], + "type": "text", + "content": "Steering Guidance for Personalized Text-to-Image Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 143, + 477, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 143, + 477, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 143, + 477, + 171 + ], + "type": "text", + "content": "Sunghyun Park* Seokeon Choi* Hyoungwoo Park Sungrack Yun Qualcomm AI Research†" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 171, + 175, + 436, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 175, + 436, + 185 + ], + "spans": [ + { + "bbox": [ + 171, + 175, + 436, + 185 + ], + "type": "text", + "content": "{sunpar, seokchoi, hwoopark, sungrack}@qti=qualcomm.com" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 59, + 201, + 115, + 310 + ], + "blocks": [ + { + "bbox": [ + 59, + 201, + 115, + 310 + ], + "lines": [ + { + "bbox": [ + 59, + 201, + 115, + 310 + ], + "spans": [ + { + "bbox": [ + 59, + 201, + 115, + 310 + ], + "type": "image", + "image_path": "23db2504435ec8c5e07cd7363b0933debf27211c658fddc7dd6f63cf5b626403.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 117, + 201, + 225, + 310 + ], + "blocks": [ + { + "bbox": [ + 117, + 201, + 225, + 310 + ], + "lines": [ + { + "bbox": [ + 117, + 201, + 225, + 310 + ], + "spans": [ + { + "bbox": [ + 117, + 201, + 225, + 310 + ], + "type": "image", + "image_path": "aac4c5beaae87aa768d036c94516d336fc2d62c597bac7c81b2e0e1609698162.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 310, + 268, + 319 + ], + "lines": [ + { + "bbox": [ + 181, + 310, + 268, + 319 + ], + "spans": [ + { + "bbox": [ + 181, + 310, + 268, + 319 + ], + "type": "text", + "content": "A " + }, + { + "bbox": [ + 181, + 310, + 268, + 319 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 181, + 310, + 268, + 319 + ], + "type": "text", + "content": " cat in a chef outfit" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 225, + 201, + 333, + 310 + ], + "blocks": [ + { + "bbox": [ + 225, + 201, + 333, + 310 + ], + "lines": [ + { + "bbox": [ + 225, + 201, + 333, + 310 + ], + "spans": [ + { + "bbox": [ + 225, + 201, + 333, + 310 + ], + "type": "image", + "image_path": "7dae5fcbd0bec05bbf8130b6cdf42f0ae06f049200608ce42fec42b4a0914837.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 334, + 201, + 441, + 310 + ], + "blocks": [ + { + "bbox": [ + 334, + 201, + 441, + 310 + ], + "lines": [ + { + "bbox": [ + 334, + 201, + 441, + 310 + ], + "spans": [ + { + "bbox": [ + 334, + 201, + 441, + 310 + ], + "type": "image", + "image_path": "fff0b5ad4afed71db9bbc380374de9d6e20017c9f9415c4f38857ca605120bf9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 442, + 201, + 550, + 310 + ], + "blocks": [ + { + "bbox": [ + 442, + 201, + 550, + 310 + ], + "lines": [ + { + "bbox": [ + 442, + 201, + 550, + 310 + ], + "spans": [ + { + "bbox": [ + 442, + 201, + 550, + 310 + ], + "type": "image", + "image_path": "9cedbbe1c3a7dcf4835071f82f91fd2a5982cf516be0380ff5f4547ca7b610ce.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 59, + 321, + 115, + 430 + ], + "blocks": [ + { + "bbox": [ + 59, + 321, + 115, + 430 + ], + "lines": [ + { + "bbox": [ + 59, + 321, + 115, + 430 + ], + "spans": [ + { + "bbox": [ + 59, + 321, + 115, + 430 + ], + "type": "image", + "image_path": "ed49c68166a3371a73a1db68dd3ff105df920ff28d5cb82ea5512f78f55581cb.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 116, + 321, + 333, + 430 + ], + "blocks": [ + { + "bbox": [ + 116, + 321, + 333, + 430 + ], + "lines": [ + { + "bbox": [ + 116, + 321, + 333, + 430 + ], + "spans": [ + { + "bbox": [ + 116, + 321, + 333, + 430 + ], + "type": "image", + "image_path": "119fa937dfe5f87617e53075cd78d8963591897ca23c47ced8268c54d6ca7e1a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 175, + 430, + 274, + 440 + ], + "lines": [ + { + "bbox": [ + 175, + 430, + 274, + 440 + ], + "spans": [ + { + "bbox": [ + 175, + 430, + 274, + 440 + ], + "type": "text", + "content": "A " + }, + { + "bbox": [ + 175, + 430, + 274, + 440 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 175, + 430, + 274, + 440 + ], + "type": "text", + "content": " dog in a movie theater" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 334, + 321, + 441, + 430 + ], + "blocks": [ + { + "bbox": [ + 334, + 321, + 441, + 430 + ], + "lines": [ + { + "bbox": [ + 334, + 321, + 441, + 430 + ], + "spans": [ + { + "bbox": [ + 334, + 321, + 441, + 430 + ], + "type": "image", + "image_path": "8ca313a3fd2efad70cddccc58943cef2fa6ef9d821e56fad310a1c768d276ac6.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 442, + 321, + 550, + 430 + ], + "blocks": [ + { + "bbox": [ + 442, + 321, + 550, + 430 + ], + "lines": [ + { + "bbox": [ + 442, + 321, + 550, + 430 + ], + "spans": [ + { + "bbox": [ + 442, + 321, + 550, + 430 + ], + "type": "image", + "image_path": "6e5a1016bb96b56f1c7cc5f3ed8116e58617475e19b6ad540b4ca961da886e05.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 59, + 441, + 115, + 551 + ], + "blocks": [ + { + "bbox": [ + 59, + 441, + 115, + 551 + ], + "lines": [ + { + "bbox": [ + 59, + 441, + 115, + 551 + ], + "spans": [ + { + "bbox": [ + 59, + 441, + 115, + 551 + ], + "type": "image", + "image_path": "cbfdfbd2181e8f59b59228a03aeef53ec85be25f8e84eb32496cf1989ff1085d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 551, + 108, + 559 + ], + "lines": [ + { + "bbox": [ + 67, + 551, + 108, + 559 + ], + "spans": [ + { + "bbox": [ + 67, + 551, + 108, + 559 + ], + "type": "text", + "content": "Reference" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 76, + 564, + 533, + 575 + ], + "lines": [ + { + "bbox": [ + 76, + 564, + 533, + 575 + ], + "spans": [ + { + "bbox": [ + 76, + 564, + 533, + 575 + ], + "type": "text", + "content": "Figure 1. Comparison of generated image quality between Classifier-Free Guidance (CFG) and our Personalization Guidance." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 116, + 441, + 225, + 550 + ], + "blocks": [ + { + "bbox": [ + 116, + 441, + 225, + 550 + ], + "lines": [ + { + "bbox": [ + 116, + 441, + 225, + 550 + ], + "spans": [ + { + "bbox": [ + 116, + 441, + 225, + 550 + ], + "type": "image", + "image_path": "f71a04de42d6201e407401c0d6691d3fc6469428b7bfff3acf5b634f8ce5c51d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 225, + 441, + 332, + 551 + ], + "blocks": [ + { + "bbox": [ + 225, + 441, + 332, + 551 + ], + "lines": [ + { + "bbox": [ + 225, + 441, + 332, + 551 + ], + "spans": [ + { + "bbox": [ + 225, + 441, + 332, + 551 + ], + "type": "image", + "image_path": "8a5befc83390fb14d03018c0f1bf0a55ec5e9d05210da6a831e121b4fb100df9.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 119, + 551, + 317, + 560 + ], + "lines": [ + { + "bbox": [ + 119, + 551, + 317, + 560 + ], + "spans": [ + { + "bbox": [ + 119, + 551, + 317, + 560 + ], + "type": "text", + "content": "A duck toy with the Eiffel Towel in the background" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 333, + 441, + 441, + 551 + ], + "blocks": [ + { + "bbox": [ + 333, + 441, + 441, + 551 + ], + "lines": [ + { + "bbox": [ + 333, + 441, + 441, + 551 + ], + "spans": [ + { + "bbox": [ + 333, + 441, + 441, + 551 + ], + "type": "image", + "image_path": "52eec6116cbdea1d926e6c41cf7403a25d3e8edbc24a692b9e55ee3d3c16c390.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 551, + 541, + 561 + ], + "lines": [ + { + "bbox": [ + 342, + 551, + 541, + 561 + ], + "spans": [ + { + "bbox": [ + 342, + 551, + 541, + 561 + ], + "type": "text", + "content": "A duck toy among the skyscrapers in New York city" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 442, + 441, + 550, + 551 + ], + "blocks": [ + { + "bbox": [ + 442, + 441, + 550, + 551 + ], + "lines": [ + { + "bbox": [ + 442, + 441, + 550, + 551 + ], + "spans": [ + { + "bbox": [ + 442, + 441, + 550, + 551 + ], + "type": "image", + "image_path": "74ef1e47cef28154d4c886aa7597eefa719541e434c44e5551438f4888b8414e.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 152, + 596, + 200, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 596, + 200, + 608 + ], + "spans": [ + { + "bbox": [ + 152, + 596, + 200, + 608 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 622, + 295, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 622, + 295, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 622, + 295, + 682 + ], + "type": "text", + "content": "Personalizing text-to-image diffusion models is crucial for adapting the pre-trained models to specific target concepts, enabling diverse image generation. However, fine-tuning with few images introduces an inherent trade-off between aligning with the target distribution (e.g., subject fidelity)" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 585, + 555, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 585, + 555, + 706 + ], + "spans": [ + { + "bbox": [ + 313, + 585, + 555, + 706 + ], + "type": "text", + "content": "and preserving the broad knowledge of the original model (e.g., text editability). Existing sampling guidance methods, such as classifier-free guidance (CFG) and autoguidance (AG), fail to effectively guide the output toward well-balanced space: CFG restricts the adaptation to the target distribution, while AG compromises text alignment. To address these limitations, we propose personalization guidance, a simple yet effective method leveraging an unlearned weak model conditioned on a null text prompt. Moreover, our method dynamically controls the extent of unlearning" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 693, + 228, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 228, + 703 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 228, + 703 + ], + "type": "text", + "content": "*These two authors contributed equally to this work." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 56, + 703, + 293, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 703, + 293, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 703, + 293, + 712 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 56, + 703, + 293, + 712 + ], + "type": "text", + "content": "Qualcomm AI Research is an initiative of Qualcomm Technologies, Inc." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15907" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "in a weak model through weight interpolation between pretrained and fine-tuned models during inference. Unlike existing guidance methods, which depend solely on guidance scales, our method explicitly steers the outputs toward a balanced latent space without additional computational overhead. Experimental results demonstrate that our proposed guidance can improve text alignment and target distribution fidelity, integrating seamlessly with various fine-tuning strategies." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 203, + 136, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 203, + 136, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 203, + 136, + 216 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 222, + 295, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 222, + 295, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 222, + 295, + 330 + ], + "type": "text", + "content": "Diffusion models [11, 32] have emerged as powerful generative models capable of representing complex data distributions, driving significant advancements in high-fidelity image generation [24, 26, 29, 38]. However, real-world use cases often demand personalized generation, where the model generates not only generic concepts but also specific concepts like a novel character or person [8, 20, 27]. This personalization technique is particularly useful for applications such as portrait editing and specific content creation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 331, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 295, + 521 + ], + "type": "text", + "content": "Many approaches have been proposed to personalize diffusion models for novel concepts. One line of work [6, 28, 30, 34] learns a general-purpose encoder capable of adapting to new concepts without test-time finetuning. However, this approach requires large-scale text-image datasets and considerable computational resources, and must often be retrained from scratch for each pre-trained model, making it expensive to adapt to the growing variety of text-to-image or text-to-video diffusion architectures. An alternative solution is to directly fine-tune a pre-trained diffusion model on the target images, which can be roughly categorized into fine-tuning low-rank adapters [5, 13], text embeddings [8], specific layers [14, 20], or the full model parameters [27]. This strategy can be applied to any pre-trained diffusion model and only requires a few target images, proving effective for personalization in various scenarios." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 522, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 714 + ], + "type": "text", + "content": "Despite these advancements, optimization-based personalization remains a nontrivial challenge. When fine-tuning a model to a specific, often much smaller, target data distribution, there is an inevitable tradeoff between aligning with the target distribution (e.g., learning a novel concept with a few target images) and preserving broad knowledge of the original model (e.g., preserving text editability). This tradeoff manifests in various ways: a fine-tuned model for personalization may generate high-quality domain-specific images but lose text editability, making it difficult to generalize across different prompts. Conversely, a model that retains broad knowledge from its original training distribution may struggle to fully align with new concepts introduced in fine-tuning. Existing approaches attempt to mitigate this tradeoff by adjusting the number of fine-tuning iterations [8, 27], leveraging regularization [14, 27], reducing the model pa" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 318, + 72, + 369, + 151 + ], + "blocks": [ + { + "bbox": [ + 318, + 72, + 369, + 151 + ], + "lines": [ + { + "bbox": [ + 318, + 72, + 369, + 151 + ], + "spans": [ + { + "bbox": [ + 318, + 72, + 369, + 151 + ], + "type": "image", + "image_path": "ae36c689934c7060d99a5fbf9318cdb4c55f44b49ddd1b30e6247d248d0fe863.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 155, + 553, + 209 + ], + "lines": [ + { + "bbox": [ + 313, + 155, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 155, + 553, + 209 + ], + "type": "text", + "content": "Figure 2. Motivation of Personalization Guidance. Classifier-free guidance [10] can interfere with subject fidelity due to its alignment with the text prompt. In contrast, autoguidance [18] fails to reflect the text prompt but generates the subject more similarly. Our guidance complements the limitations of both methods." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 370, + 72, + 430, + 152 + ], + "blocks": [ + { + "bbox": [ + 370, + 72, + 430, + 152 + ], + "lines": [ + { + "bbox": [ + 370, + 72, + 430, + 152 + ], + "spans": [ + { + "bbox": [ + 370, + 72, + 430, + 152 + ], + "type": "image", + "image_path": "bb9b851d5c3a3ac9ea1411833aa6e6d8639d714b23a8cf5d3b886b22841cdb78.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 431, + 72, + 491, + 152 + ], + "blocks": [ + { + "bbox": [ + 431, + 72, + 491, + 152 + ], + "lines": [ + { + "bbox": [ + 431, + 72, + 491, + 152 + ], + "spans": [ + { + "bbox": [ + 431, + 72, + 491, + 152 + ], + "type": "image", + "image_path": "6d21b32247e247b671aa7ca9ce6b146e58087326080219977d753112269c9cfe.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 491, + 72, + 553, + 152 + ], + "blocks": [ + { + "bbox": [ + 491, + 72, + 553, + 152 + ], + "lines": [ + { + "bbox": [ + 491, + 72, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 491, + 72, + 553, + 152 + ], + "type": "image", + "image_path": "1f73207d1918fa94259aa2f031677f7275b0051240112abeb37a68f3021e8ab7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 218, + 553, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 218, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 553, + 266 + ], + "type": "text", + "content": "rameters of updates [5, 13], yet they primarily operate at the fine-tuning stage. Moreover, balancing the trade-off between comprehensive text editability and precise alignment with a new target distribution remains an open problem." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 268, + 555, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 555, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 555, + 508 + ], + "type": "text", + "content": "A parallel line of research focuses on sampling guidance techniques, such as Classifier-Free Guidance (CFG) [10] and AutoGuidance (AG) [18]. These methods rely on \"weak models\" to detect and push the outputs away from poor trajectories, thereby guiding samples toward higher-quality regions of the data manifold. For instance, CFG uses an unconditional model using the null text prompt as the weak model, while AG uses an unlearned model—one with fewer parameters or less training—conditioned on the same text prompt. However, these methods have limitations when applied to personalized diffusion models, as shown in Fig. 2. CFG can interfere with the model's adaptation to the target distribution, restricting its ability to fully utilize fine-tuned knowledge. Moreover, while the guidance scale of CFG adjusts the degree of text fidelity, it does not improve alignment with the target data. On the other hand, AG, which primarily enhances alignment with the target distribution, however, lacks an explicit mechanism for improving text alignment, making it unsuitable for fine-tuned text-to-image models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": "To address these issues, we propose a simple yet effective guidance technique, called Personalization Guidance, for personalized diffusion models, which can dynamically balance text alignment and adaptation to the target distribution. Our method leverages an \"unlearned\" weak model with a null text prompt, without incurring additional computational overhead during inference. Here, we aim to balance the alignment with the target distribution (e.g., subject fidelity) and the preservation of the broad knowledge of the original model (e.g., text fidelity) by controlling the extent of \"unlearning\" in a weak model. Here, the degree of unlearning refers to the extent to which the target data distribution has not been learned compared to the fine-tuned model. The rationale behind adjusting this degree of unlearning is to steer the output toward an optimal, well-balanced latent space. Specifically, during inference, we interpolate the weight parameters between the pre-trained and fine-tuned" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15908" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": "diffusion models to derive an optimal weak model, effectively controlling the extent of unlearning. Unlike guidance scales, which adjust the intensity of guidance, weight interpolation for the weak model serves to steer outputs in an appropriate direction. This straightforward approach can be readily integrated into various fine-tuning diffusion models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 155, + 143, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 155, + 143, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 155, + 143, + 167 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 296, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 296, + 378 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 296, + 378 + ], + "type": "text", + "content": "Personalization of Diffusion Models. Personalizing diffusion models aim to generate novel images by adapting to specific concepts using a set of reference images. Optimization-based methods personalize models by finetuning textual embeddings [8, 14, 20], entire model parameters with regularization [27], or low-rank adapters [5, 13]. Another line of research leverages extensive text-image datasets to train dedicated text-to-image personalization models, by fine-tuning additional modules [6, 28, 30, 34]. Recently, several studies [4, 22, 31] have achieved better personalization without additional fine-tuning or training. Orthogonal to these existing approaches, we introduce a simple yet effective personalization guidance technique that seamlessly integrates with various personalization methods by modifying the classifier-free guidance. Moreover, our guidance enables us to adjust the degree of subject and text fidelity without incurring additional computational costs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 379, + 296, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 379, + 296, + 629 + ], + "spans": [ + { + "bbox": [ + 56, + 379, + 296, + 629 + ], + "type": "text", + "content": "Guidance for Diffusion Models. Classifier-Free Guidance (CFG) [10] enhances conditional image generation by leveraging the unconditional model as a weak model to guide the outputs toward better adherence to the given condition. Various guidance techniques have been explored, such as the guidance using perturbations [1, 16] or blurring to the attention maps [12]. Autoguidance [18] improves generation by using a weak model—one with fewer parameters or less training—to push outputs toward a well-trained direction. While InstructPix2Pix [2] introduces separate CFG using both image and text conditions, it cannot be directly applied to DB-LoRA due to its reliance on image-conditioned training. For personalization, subject-agnostic guidance [4] enhances subject-agnostic attributes by employing a weak model conditioned on subject-specific text prompts. In this study, we aim to address the fundamental personalization trade-off—aligning a model with a smaller target data distribution while preserving the broad text fidelity of the original model. We derive a weak model that dynamically adjusts the degree of adaptation, balancing comprehensive text editability with precise alignment to new concepts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "Weight Interpolation. There are previous studies on exploring the use of weight interpolation. WISE-FT [36] improves the robustness of fine-tuned classifiers against distribution shifts through weight-space ensembling. Model-soups [35] maximizes model accuracy and robustness by interpolating the weights of multiple fine-tuned classifiers trained with different hyperparameter configurations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 71, + 370, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 370, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 370, + 83 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 91, + 400, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 91, + 400, + 102 + ], + "spans": [ + { + "bbox": [ + 314, + 91, + 400, + 102 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 107, + 555, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 107, + 555, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 107, + 555, + 227 + ], + "type": "text", + "content": "Diffusion Models. Diffusion models [11, 32] aim to learn reversing a forward noising process that gradually transforms data samples into noise. Concretely, the forward process starts with a data distribution " + }, + { + "bbox": [ + 313, + 107, + 555, + 227 + ], + "type": "inline_equation", + "content": "p_0(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 107, + 555, + 227 + ], + "type": "text", + "content": " and ends at a high-noise distribution " + }, + { + "bbox": [ + 313, + 107, + 555, + 227 + ], + "type": "inline_equation", + "content": "p_T(\\mathbf{x}) \\approx \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 313, + 107, + 555, + 227 + ], + "type": "text", + "content": ". Sampling from the data distribution proceeds by solving the equivalent probability-flow ordinary differential equation (PF-ODE). Under the variance-preserving parameterization where " + }, + { + "bbox": [ + 313, + 107, + 555, + 227 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_t|\\mathbf{x}_0) = \\mathcal{N}(\\mathbf{x}_0,\\sigma_t^2\\mathbf{I})" + }, + { + "bbox": [ + 313, + 107, + 555, + 227 + ], + "type": "text", + "content": ", the PF-ODE takes the form:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 339, + 236, + 555, + 249 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 236, + 555, + 249 + ], + "spans": [ + { + "bbox": [ + 339, + 236, + 555, + 249 + ], + "type": "interline_equation", + "content": "\\mathrm {d} \\mathbf {x} _ {t} = - \\sigma_ {t} \\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {x} _ {t}) \\mathrm {d} t, \\quad \\mathbf {x} _ {T} \\sim p _ {T} (\\mathbf {x} _ {T}). \\tag {1}", + "image_path": "42d932461f163e20e30953338de0feeee43bc2703dcd14f9ca2ecc94c9186679.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 255, + 555, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 292 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 292 + ], + "type": "text", + "content": "In diffusion models, the score function is parameterized by " + }, + { + "bbox": [ + 313, + 255, + 555, + 292 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 255, + 555, + 292 + ], + "type": "text", + "content": ". Here, the noise estimation of the diffusion model can be considered as " + }, + { + "bbox": [ + 313, + 255, + 555, + 292 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x}_t) \\approx -\\sigma_t \\nabla_{\\mathbf{x}_t} \\log p(\\mathbf{x}_t)" + }, + { + "bbox": [ + 313, + 255, + 555, + 292 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "text", + "content": "Personalization. Suppose we have a pre-trained diffusion model " + }, + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "text", + "content": " is the parameters of the model. We wish to personalize it to a smaller target dataset " + }, + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{target}} = \\left\\{(\\mathbf{x}^i, \\mathbf{c}^i)\\right\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "text", + "content": ", where each " + }, + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^i" + }, + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^i" + }, + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "text", + "content": " is an image and its associated text prompt, respectively. The goal is to obtain new parameters " + }, + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "inline_equation", + "content": "\\theta'" + }, + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "text", + "content": " so that the fine-tuned model " + }, + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta'}" + }, + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "text", + "content": " can generate samples more faithfully matched the " + }, + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{target}}" + }, + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 376, + 554, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 554, + 435 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 554, + 435 + ], + "type": "text", + "content": "Fine-tuning typically follows the same denoising objective used by the original diffusion model [11]. Concretely, let " + }, + { + "bbox": [ + 313, + 376, + 554, + 435 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 376, + 554, + 435 + ], + "type": "text", + "content": " be a noise level sampled from a training schedule, and let " + }, + { + "bbox": [ + 313, + 376, + 554, + 435 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\sigma^2\\mathbf{I})" + }, + { + "bbox": [ + 313, + 376, + 554, + 435 + ], + "type": "text", + "content": ". For each " + }, + { + "bbox": [ + 313, + 376, + 554, + 435 + ], + "type": "inline_equation", + "content": "(\\mathbf{x},\\mathbf{c})" + }, + { + "bbox": [ + 313, + 376, + 554, + 435 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 376, + 554, + 435 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{target}}" + }, + { + "bbox": [ + 313, + 376, + 554, + 435 + ], + "type": "text", + "content": ", we form the noisy input " + }, + { + "bbox": [ + 313, + 376, + 554, + 435 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t = \\mathbf{x} + \\epsilon" + }, + { + "bbox": [ + 313, + 376, + 554, + 435 + ], + "type": "text", + "content": " and minimize:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 348, + 441, + 555, + 465 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 441, + 555, + 465 + ], + "spans": [ + { + "bbox": [ + 348, + 441, + 555, + 465 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} _ {(\\mathbf {x}, \\mathbf {c}) \\sim \\mathcal {D} _ {\\text {t a r g e t}}, \\sigma , \\boldsymbol {\\epsilon}} \\left\\| \\epsilon_ {\\theta^ {\\prime}} (\\mathbf {x} _ {t}, \\sigma , \\mathbf {c}) - \\boldsymbol {\\epsilon} \\right\\| _ {2} ^ {2}. \\tag {2}", + "image_path": "857b2025d273a4c806e7d265a98e5d56fa75a35a9ac829481258d9035ab00e47.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 471, + 554, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 554, + 518 + ], + "type": "text", + "content": "This teaches the model to reconstruct " + }, + { + "bbox": [ + 313, + 471, + 554, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 471, + 554, + 518 + ], + "type": "text", + "content": " from noisy inputs in the domain of " + }, + { + "bbox": [ + 313, + 471, + 554, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{target}}" + }, + { + "bbox": [ + 313, + 471, + 554, + 518 + ], + "type": "text", + "content": ", thereby adapting it to the new concept. In practice, additional regularization terms are often used to prevent overfitting [14, 27]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 518, + 554, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 554, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 554, + 579 + ], + "type": "text", + "content": "Classifier-Free Guidance. To enhance the generation towards an arbitrary condition " + }, + { + "bbox": [ + 313, + 518, + 554, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 518, + 554, + 579 + ], + "type": "text", + "content": ", classifier-free guidance (CFG) [10] introduces a new sampling distribution " + }, + { + "bbox": [ + 313, + 518, + 554, + 579 + ], + "type": "inline_equation", + "content": "p^{\\lambda}(\\mathbf{x}|\\mathbf{c})" + }, + { + "bbox": [ + 313, + 518, + 554, + 579 + ], + "type": "text", + "content": " composed with both " + }, + { + "bbox": [ + 313, + 518, + 554, + 579 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 518, + 554, + 579 + ], + "type": "text", + "content": " and the classifier distribution " + }, + { + "bbox": [ + 313, + 518, + 554, + 579 + ], + "type": "inline_equation", + "content": "p(c|\\mathbf{x})^{\\lambda}" + }, + { + "bbox": [ + 313, + 518, + 554, + 579 + ], + "type": "text", + "content": ", which is described as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 381, + 585, + 555, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 585, + 555, + 600 + ], + "spans": [ + { + "bbox": [ + 381, + 585, + 555, + 600 + ], + "type": "interline_equation", + "content": "p ^ {\\lambda} (\\mathbf {x} | \\mathbf {c}) \\propto p (\\mathbf {x}) p (\\mathbf {c} | \\mathbf {x}) ^ {\\lambda}, \\tag {3}", + "image_path": "4005293e38ab3802bbbcb13abcd44c92ee779670b8b6fe5dd14695a318d8f22e.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 606, + 554, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 554, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 554, + 642 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 606, + 554, + 642 + ], + "type": "inline_equation", + "content": "\\lambda > 1" + }, + { + "bbox": [ + 313, + 606, + 554, + 642 + ], + "type": "text", + "content": " amplifies the conditional likelihood. By applying Bayes' rule for some timestep " + }, + { + "bbox": [ + 313, + 606, + 554, + 642 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 606, + 554, + 642 + ], + "type": "text", + "content": ", the corresponding score function is:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 324, + 652, + 553, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 652, + 553, + 682 + ], + "spans": [ + { + "bbox": [ + 324, + 652, + 553, + 682 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {\\mathbf {x} _ {t}} \\log p ^ {\\lambda} (\\mathbf {x} _ {t} | \\mathbf {c}) = \\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {x} _ {t}) \\\\ + \\lambda \\left(\\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {x} _ {t} | \\mathbf {c}) - \\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {x} _ {t})\\right). \\tag {4} \\\\ \\end{array}", + "image_path": "746186c7cadea5cf576413bd59f5e76f62efd7ee9832499d96ed59ce50a12d4a.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "To enable conditional generation such as text prompts and class labels, CFG jointly trains the unconditional model" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15909" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 70, + 291, + 253 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 291, + 253 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 291, + 253 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 291, + 253 + ], + "type": "image", + "image_path": "f9f9e426c6ceb3e0cc05f6b58e917c2a3355395365832500f44d8242372c3b0a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 259, + 296, + 314 + ], + "lines": [ + { + "bbox": [ + 55, + 259, + 296, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 296, + 314 + ], + "type": "text", + "content": "Figure 3. Comparison between Classifier-Free Guidance (CFG), AutoGuidance (AG), and our Personalization Guidance (PG), with the band conceptually representing the noisy data manifold. By leveraging weight interpolation, guidance can be steered toward the optimal space between CFG and PG." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 323, + 296, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 382 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 382 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x}_t,\\phi)" + }, + { + "bbox": [ + 55, + 323, + 296, + 382 + ], + "type": "text", + "content": " and the conditional model " + }, + { + "bbox": [ + 55, + 323, + 296, + 382 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x}_t,\\mathbf{c})" + }, + { + "bbox": [ + 55, + 323, + 296, + 382 + ], + "type": "text", + "content": " within a single model by dropping " + }, + { + "bbox": [ + 55, + 323, + 296, + 382 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 323, + 296, + 382 + ], + "type": "text", + "content": " with certain probability to allow null conditioning with " + }, + { + "bbox": [ + 55, + 323, + 296, + 382 + ], + "type": "inline_equation", + "content": "c = \\phi" + }, + { + "bbox": [ + 55, + 323, + 296, + 382 + ], + "type": "text", + "content": ". In the personalization setting, parameterizing the score function with the fine-tuned diffusion model " + }, + { + "bbox": [ + 55, + 323, + 296, + 382 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta'}" + }, + { + "bbox": [ + 55, + 323, + 296, + 382 + ], + "type": "text", + "content": " is described as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 388, + 295, + 404 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 388, + 295, + 404 + ], + "spans": [ + { + "bbox": [ + 66, + 388, + 295, + 404 + ], + "type": "interline_equation", + "content": "\\hat {\\epsilon} _ {\\theta^ {\\prime}} ^ {\\lambda} \\left(\\mathbf {x} _ {t} \\mid \\mathbf {c}\\right) = \\epsilon_ {\\theta^ {\\prime}} \\left(\\mathbf {x} _ {t} \\mid \\phi\\right) + \\lambda \\left(\\epsilon_ {\\theta^ {\\prime}} \\left(\\mathbf {x} _ {t} \\mid \\mathbf {c}\\right) - \\epsilon_ {\\theta^ {\\prime}} \\left(\\mathbf {x} _ {t} \\mid \\phi\\right)\\right). \\tag {5}", + "image_path": "1de8789ee9443204f191b2064e70c1e46ed1e380069706b5814edde7bb8026c5.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 410, + 296, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 296, + 446 + ], + "type": "text", + "content": "However, while CFG improves alignment with the conditions, such as the text prompt, it struggles to enhance adaptation to the target data in the personalization task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 447, + 296, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 296, + 614 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 296, + 614 + ], + "type": "text", + "content": "AutoGuidance. AutoGuidance (AG) [18] is another guidance technique that employs a weaker version of the same model to identify suboptimal generation trajectories. Rather than using the unconditional model " + }, + { + "bbox": [ + 55, + 447, + 296, + 614 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x}_t|\\phi)" + }, + { + "bbox": [ + 55, + 447, + 296, + 614 + ], + "type": "text", + "content": ", AG directly guides a model with the weak model " + }, + { + "bbox": [ + 55, + 447, + 296, + 614 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta_{weak}}" + }, + { + "bbox": [ + 55, + 447, + 296, + 614 + ], + "type": "text", + "content": ", which is trained on the same task and conditioning, but degraded by under-training or less-parameters. However, obtaining a weaker version " + }, + { + "bbox": [ + 55, + 447, + 296, + 614 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta_{weak}}" + }, + { + "bbox": [ + 55, + 447, + 296, + 614 + ], + "type": "text", + "content": " of a fully trained diffusion model like Stable Diffusion [26] is challenging in practice. Instead, we found that a less trained model could serve as the pre-trained model for personalization, where " + }, + { + "bbox": [ + 55, + 447, + 296, + 614 + ], + "type": "inline_equation", + "content": "\\theta_{weak} = \\theta" + }, + { + "bbox": [ + 55, + 447, + 296, + 614 + ], + "type": "text", + "content": ", effectively shifts the output away in the direction of fine-tuning, enhancing the adaptation to the target data. In the personalization, AG is formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 620, + 295, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 620, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 70, + 620, + 295, + 635 + ], + "type": "interline_equation", + "content": "\\tilde {\\epsilon} _ {\\theta^ {\\prime}} ^ {\\lambda} \\left(\\mathbf {x} _ {t} | \\mathbf {c}\\right) = \\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t} | \\mathbf {c}\\right) + \\lambda \\left(\\epsilon_ {\\theta^ {\\prime}} \\left(\\mathbf {x} _ {t} | \\mathbf {c}\\right) - \\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t} | \\mathbf {c}\\right)\\right). \\tag {6}", + "image_path": "23ca468b8ee177b9ce88cd64e3fcaca5695dd26b151e160bf6b39c8cd83552f5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 641, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 666 + ], + "type": "text", + "content": "However, this approach has a major drawback—it scarcely reflects the text prompt, as shown in Fig. 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 672, + 241, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 241, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 241, + 685 + ], + "type": "text", + "content": "3.2. Steering Personalization Guidance" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Inspired by autoguidance [18], we simply modify the CFG [10], by leveraging a pre-trained diffusion model " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 92, + 552, + 313 + ], + "blocks": [ + { + "bbox": [ + 365, + 74, + 520, + 91 + ], + "lines": [ + { + "bbox": [ + 365, + 74, + 520, + 91 + ], + "spans": [ + { + "bbox": [ + 365, + 74, + 520, + 91 + ], + "type": "text", + "content": "Subject Fidelity (DINO) vs Text Fidelity (CLIP-T) (weight interpolation scale " + }, + { + "bbox": [ + 365, + 74, + 520, + 91 + ], + "type": "inline_equation", + "content": "\\omega \\in [0,1]" + }, + { + "bbox": [ + 365, + 74, + 520, + 91 + ], + "type": "text", + "content": ", step " + }, + { + "bbox": [ + 365, + 74, + 520, + 91 + ], + "type": "inline_equation", + "content": "= 0.1" + }, + { + "bbox": [ + 365, + 74, + 520, + 91 + ], + "type": "text", + "content": ")" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 92, + 552, + 313 + ], + "lines": [ + { + "bbox": [ + 316, + 92, + 552, + 313 + ], + "spans": [ + { + "bbox": [ + 316, + 92, + 552, + 313 + ], + "type": "image", + "image_path": "c12daafe985ba09788877b81777b62340a8302bd27ec89c843108668947cdf82.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 316, + 555, + 372 + ], + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 372 + ], + "type": "text", + "content": "Figure 4. Ablation study on the weight interpolation scale " + }, + { + "bbox": [ + 313, + 316, + 555, + 372 + ], + "type": "inline_equation", + "content": "\\omega \\in [0.0,1.0]" + }, + { + "bbox": [ + 313, + 316, + 555, + 372 + ], + "type": "text", + "content": " with measurements taken at intervals of 0.1. Here, we use DreamBooth (DB) LoRA based on SD 1.5, SD 2.1, and SANA. By adjusting " + }, + { + "bbox": [ + 313, + 316, + 555, + 372 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 316, + 555, + 372 + ], + "type": "text", + "content": ", it is possible to improve subject fidelity or text fidelity, while preserving the other." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 382, + 534, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 534, + 394 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 534, + 394 + ], + "type": "text", + "content": "instead of the fine-tuned model " + }, + { + "bbox": [ + 313, + 382, + 534, + 394 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta^{\\prime}}" + }, + { + "bbox": [ + 313, + 382, + 534, + 394 + ], + "type": "text", + "content": " as the weak model." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 327, + 403, + 554, + 419 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 403, + 554, + 419 + ], + "spans": [ + { + "bbox": [ + 327, + 403, + 554, + 419 + ], + "type": "interline_equation", + "content": "\\hat {\\epsilon} _ {\\theta^ {\\prime}} ^ {\\lambda} \\left(\\mathbf {x} _ {t} | \\mathbf {c}\\right) = \\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t} | \\phi\\right) + \\lambda \\left(\\epsilon_ {\\theta^ {\\prime}} \\left(\\mathbf {x} _ {t} | \\mathbf {c}\\right) - \\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t} | \\phi\\right)\\right). \\tag {7}", + "image_path": "fad9017e3c31bc8859a99e34d913dba7f5bf39fb00d35d91cec2166570c585e8.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 428, + 554, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 428, + 554, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 428, + 554, + 452 + ], + "type": "text", + "content": "Despite being a simple extension, we found that this approach significantly improves subject fidelity." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 453, + 555, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 555, + 656 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 555, + 656 + ], + "type": "text", + "content": "Fig. 3 shows the rationale behind our personalization guidance. Since this simply modified guidance can steer the output toward the intersection space where both text fidelity and subject fidelity are high. As illustrated in Fig. 3, we explore the feasibility of identifying a superior space within the noisy data manifold. We hypothesize that an optimal point for text and subject fidelity could be located between CFG and PG, and we propose a method to facilitate the steering of guidance. Steering the guidance involves the adjustment of an appropriate weak model. By identifying a suitable weak model, we can steer " + }, + { + "bbox": [ + 313, + 453, + 555, + 656 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x}_t|\\phi)" + }, + { + "bbox": [ + 313, + 453, + 555, + 656 + ], + "type": "text", + "content": " towards a more favorable point. We propose employing a weight interpolation to obtain an appropriate weak model. Specifically, by interpolating weights between the parameters of the pretrained and the fine-tuned models, we can modulate the degree of adaptation to the target data, thereby obtaining an appropriate weak model " + }, + { + "bbox": [ + 313, + 453, + 555, + 656 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta_{\\omega}}" + }, + { + "bbox": [ + 313, + 453, + 555, + 656 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 359, + 666, + 553, + 680 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 666, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 359, + 666, + 553, + 680 + ], + "type": "interline_equation", + "content": "\\theta_ {\\omega} = \\omega \\theta^ {\\prime} + (1 - \\omega) \\theta , \\quad \\omega \\in [ 0, 1 ], \\tag {8}", + "image_path": "a59bfff863a9190ff589f83d16db7373671edb604692223f9d1ae164503cf33c.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": " denotes the scale of weight interpolation. Increasing " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": " moves parameters closer to the fine-tuned model " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\theta^{\\prime}" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15910" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 292, + 257 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 292, + 257 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 292, + 257 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 292, + 257 + ], + "type": "table", + "html": "
MethodλDINOCLIP-ICLIP-T
SD 1.5DB-LoRA-0.37410.67970.2834
+ CFG7.50.47010.73490.3345
+ AG2.00.48310.74080.3236
+ Ours (ω = 0)7.50.49850.75160.3260
+ Ours (ω = 0.6)7.50.50240.75100.3288
SD 2.1DB-LoRA-0.42020.70760.2929
+ CFG7.50.49760.75190.3323
+ AG2.00.50720.75710.3212
+ Ours (ω = 0)7.50.52480.76550.3254
+ Ours (ω = 0.6)7.50.52810.76600.3277
SANADB-LoRA-0.41900.69390.3064
+ CFG4.50.48190.72910.3363
+ AG2.00.47920.72840.3355
+ Ours (ω = 0)4.50.53660.75220.3357
", + "image_path": "679f8766c1ed0e46e7e51f40a44cbceda0c02dbb581e0d4426c5ddf020f388e2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 258, + 295, + 291 + ], + "lines": [ + { + "bbox": [ + 55, + 258, + 295, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 295, + 291 + ], + "type": "text", + "content": "Table 1. Comparison with other guidance methods across different base diffusion models including SD1.5, SD2.1, and SANA. Here, we use DreamBooth-LoRA (DB-LoRA) only." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 295, + 295, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 295, + 319 + ], + "type": "text", + "content": "reducing the loss " + }, + { + "bbox": [ + 55, + 295, + 295, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 295, + 295, + 319 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 55, + 295, + 295, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{target}}" + }, + { + "bbox": [ + 55, + 295, + 295, + 319 + ], + "type": "text", + "content": ". Finally, applying the weight interpolation to the weak model in the guidance:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 324, + 295, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 324, + 295, + 340 + ], + "spans": [ + { + "bbox": [ + 64, + 324, + 295, + 340 + ], + "type": "interline_equation", + "content": "\\tilde {\\epsilon} _ {\\theta^ {\\prime}} ^ {\\lambda} (\\mathbf {x} _ {t} | \\mathbf {c}) = \\epsilon_ {\\theta_ {\\omega}} (\\mathbf {x} _ {t} | \\phi) + \\lambda \\bigl (\\epsilon_ {\\theta^ {\\prime}} (\\mathbf {x} _ {t} | \\mathbf {c}) - \\epsilon_ {\\theta_ {\\omega}} (\\mathbf {x} _ {t} | \\phi) \\bigr). (9)", + "image_path": "e77e8050678ff842dc79f83835e5441041b9ac0d5a7d60e3a80c3dc12de7cbde.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 344, + 296, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 344, + 296, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 296, + 440 + ], + "type": "text", + "content": "Note that " + }, + { + "bbox": [ + 55, + 344, + 296, + 440 + ], + "type": "inline_equation", + "content": "\\omega = 1" + }, + { + "bbox": [ + 55, + 344, + 296, + 440 + ], + "type": "text", + "content": " is equal to CFG. As shown in Fig. 4, it is possible to easily control the degree of subject and text fidelity. If " + }, + { + "bbox": [ + 55, + 344, + 296, + 440 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 344, + 296, + 440 + ], + "type": "text", + "content": " includes only linear classifier, the personalization guidance with weight interpolation is equal to the combination of CFG and AG, which is the output-space interpolation. However, as neural networks are non-linear with respect to parameters, guidance with weight interpolation is different from the output interpolation of CFG and AG [36]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 449, + 137, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 137, + 463 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 137, + 463 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 468, + 178, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 178, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 178, + 482 + ], + "type": "text", + "content": "4.1. Experimental Setting" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 486, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 581 + ], + "type": "text", + "content": "Datasets. We conduct our experiments on the ViCo dataset [9], which includes 16 concepts and 31 text prompts. Each concept includes a small set of reference images (typically 4-7) and associated text prompts. Following the previous work [22], we fine-tune the pre-trained text-to-image diffusion models per concept and then generate 8 images per concept and text prompt, resulting in a total of 3,968 images for evaluation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "Evaluation Metrics. Following the existing personalization studies [14, 22, 27], we evaluate subject and text fidelity. For the quantitative assessment, we adopt three evaluation metrics: (i) CLIP-T score [25], which uses CLIP's text-to-image similarity to evaluate text prompt alignment; (ii) CLIP-I score, which computes image-to-image similarity using CLIP embeddings to gauge how closely generated samples match reference images; and (iii) DINO score [3], which calculates image-to-image similarity between reference and generated images, offering a complementary perspective on subject fidelity." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 552, + 350 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 552, + 350 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 552, + 350 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 552, + 350 + ], + "type": "table", + "html": "
MethodλDINOCLIP-ICLIP-T
SD 1.5DB-LoRA+TI-0.37250.67750.2817
+ CFG7.50.46180.72920.3316
+ SAG1.00.46000.72730.3307
+ AG2.00.48060.74080.3264
+ Ours (ω = 0)7.50.48140.74590.3233
+ Ours (ω = 0.7)7.50.48800.74610.3272
ClassDiffusion-0.37030.66690.2757
+ CFG7.50.52870.75170.3305
+ SAG1.00.51720.74650.3296
+ AG2.00.55700.76280.3280
+ Ours (ω = 0)7.50.57000.76560.3297
SD 2.1DB-LoRA+TI-0.45140.72510.2823
+ CFG7.50.52910.77490.3244
+ SAG1.00.53280.77660.3223
+ AG2.00.55570.78770.3206
+ Ours (ω = 0)7.50.56830.79180.3163
+ Ours (ω = 0.6)7.50.56890.79190.3185
ClassDiffusion-0.42730.69950.2843
+ CFG7.50.58400.77570.3277
+ SAG1.00.57830.77390.3244
+ AG2.00.60600.78560.3246
+ Ours (ω = 0)7.50.61660.78960.3246
", + "image_path": "c3e2b1e44ebbe1820b900d27d2a8547efaec04fb2cbca863c7c8be35fbcf8ea2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 350, + 555, + 416 + ], + "lines": [ + { + "bbox": [ + 313, + 350, + 555, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 555, + 416 + ], + "type": "text", + "content": "Table 2. Comparison with other guidance methods across different base diffusion models including SD1.5 and SD2.1. Here, we use DreamBooth-LoRA (DB-LoRA) + Textual Inversion (TI) and ClassDiffusion. Since these personalization methods fine-tune text embedding, subject-agnostic guidance (SAG) is also compared with our method." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 425, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 555, + 544 + ], + "type": "text", + "content": "Implementation Details. In our experiments, we utilize three pre-trained text-to-image diffusion models as the base models: Stable Diffusion 1.5 (SD 1.5) [26], Stable Diffusion 2.1 (SD 2.1) [26], and SANA-1.6B [38]. SD 1.5 and SD 2.1 generate " + }, + { + "bbox": [ + 313, + 425, + 555, + 544 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 425, + 555, + 544 + ], + "type": "text", + "content": " resolution images, which comprise U-Net based architectures. On the other hand, SANA is a recent text-to-image diffusion model, which is based on the diffusion transformer [23] that can generate images up to " + }, + { + "bbox": [ + 313, + 425, + 555, + 544 + ], + "type": "inline_equation", + "content": "4096 \\times 4096" + }, + { + "bbox": [ + 313, + 425, + 555, + 544 + ], + "type": "text", + "content": " resolution. In the experiments, SANA generates " + }, + { + "bbox": [ + 313, + 425, + 555, + 544 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 313, + 425, + 555, + 544 + ], + "type": "text", + "content": " images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 545, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 640 + ], + "type": "text", + "content": "To fine-tune the diffusion models, we leverage several personalization techniques: (i) DreamBoooth-LoRA (DB-LoRA) [13, 27], (ii) DB-LoRA + Textual Inversion (TI), and (iii) ClassDiffusion [14]. We fine-tune each model for up to 500 steps on the reference images from the ViCo dataset. We set the number of sampling steps (e.g., DDIM) between 20-50. Further details regarding additional hyperparameters are described in the supplementary." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Baselines. We evaluate our guidance method against other guidance techniques: (i) Classifier-Free Guidance (CFG) [10], (ii) Subject-Agnostic Guidance (SAG) [4], and (iii) AutoGuidance (AG) [18]. Since SAG and AG alone struggle to generate images that accurately reflect the text prompt, they are used in conjunction with CFG to produce" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15911" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 74, + 555, + 280 + ], + "blocks": [ + { + "bbox": [ + 61, + 74, + 555, + 280 + ], + "lines": [ + { + "bbox": [ + 61, + 74, + 555, + 280 + ], + "spans": [ + { + "bbox": [ + 61, + 74, + 555, + 280 + ], + "type": "image", + "image_path": "1d705a80131c1bae71e8c79f1a6b8494f5312ce07dd18e15e4dac37e4555b5e6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 282, + 555, + 306 + ], + "lines": [ + { + "bbox": [ + 55, + 282, + 555, + 306 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 555, + 306 + ], + "type": "text", + "content": "Figure 5. Comparison with other guidance techniques, including without guidance, CFG, CFG+SAG, and CFG+AG. These images are generated by the fine-tuned SD 2.1 using DB-LoRA and ClassDiffusion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 312, + 555, + 492 + ], + "blocks": [ + { + "bbox": [ + 59, + 312, + 555, + 492 + ], + "lines": [ + { + "bbox": [ + 59, + 312, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 59, + 312, + 555, + 492 + ], + "type": "image", + "image_path": "50c376ece40a93b7ade99a7c6ab85e351bf8e8c86ddc1916c68038b40127be71.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 494, + 533, + 505 + ], + "lines": [ + { + "bbox": [ + 75, + 494, + 533, + 505 + ], + "spans": [ + { + "bbox": [ + 75, + 494, + 533, + 505 + ], + "type": "text", + "content": "Figure 6. Comparison with other guidance techniques. These images are generated by fine-tuned SANA [38] using DB-LoRA." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 520, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 295, + 544 + ], + "type": "text", + "content": "images. Specifically, AG experiments are conducted by referring to recent work [19] that combines CFG and AG." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 552, + 112, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 112, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 112, + 563 + ], + "type": "text", + "content": "4.2. Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "Quantitative Results. Table 1 and Table 2 showcase the personalization performance on the ViCo dataset. Table 1 provides the quantitative results for DB-LoRA based on SD 1.5, SD 2.1, and SANA. Since DB-LoRA+TI and Class-Diffusion involve fine-tuning textual embeddings, we also evaluate subject-agnostic guidance (SAG) [4], which necessitates learned textual embeddings for specific concepts. As illustrated in both Table 1 and Table 2, our guidance method significantly enhances subject fidelity (DINO and CLIP-I scores), while maintaining text fidelity (CLIP-T score) with minimal change. Additionally, by adjusting " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": ", the model can further optimize personalization performance across all" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 520, + 555, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 555, + 602 + ], + "type": "text", + "content": "evaluation metrics. We select the optimal " + }, + { + "bbox": [ + 313, + 520, + 555, + 602 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 520, + 555, + 602 + ], + "type": "text", + "content": " based on the best DINO and CLIP-I scores. For SANA with DB-LoRA and ClassDiffusion, the best subject fidelity was observed at " + }, + { + "bbox": [ + 313, + 520, + 555, + 602 + ], + "type": "inline_equation", + "content": "\\omega = 0" + }, + { + "bbox": [ + 313, + 520, + 555, + 602 + ], + "type": "text", + "content": ". These results demonstrate that our approach significantly improves subject fidelity while maintaining text fidelity, and that steering the guidance enables a balanced trade-off between the two." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Interestingly, the impact of personalized guidance varies depending on the personalization method. ClassDiffusion [14] employs a semantic preservation loss that maintains the text embedding space, leading to better overall text fidelity retention. As a result, our guidance method maintains the CLIP-T score more effectively when applied to ClassDiffusion than to other personalization methods. This observation highlights a key insight: the better the fine-tuning direction, the more effective our guidance becomes." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15912" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 72, + 529, + 223 + ], + "blocks": [ + { + "bbox": [ + 80, + 72, + 529, + 223 + ], + "lines": [ + { + "bbox": [ + 80, + 72, + 529, + 223 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 529, + 223 + ], + "type": "image", + "image_path": "17e0c5f46402d9a73a2f9565946968999bb38872429f0f2a1a45f7f67f3dbf86.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 224, + 453, + 236 + ], + "lines": [ + { + "bbox": [ + 156, + 224, + 453, + 236 + ], + "spans": [ + { + "bbox": [ + 156, + 224, + 453, + 236 + ], + "type": "text", + "content": "Figure 7. Generated images by changing " + }, + { + "bbox": [ + 156, + 224, + 453, + 236 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 156, + 224, + 453, + 236 + ], + "type": "text", + "content": ", where SANA with DB-LoRA is used." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 72, + 245, + 279, + 406 + ], + "blocks": [ + { + "bbox": [ + 72, + 245, + 279, + 406 + ], + "lines": [ + { + "bbox": [ + 72, + 245, + 279, + 406 + ], + "spans": [ + { + "bbox": [ + 72, + 245, + 279, + 406 + ], + "type": "table", + "html": "
MethodSubject Fidelity ↑Text Fidelity ↑
Stable Diffusion 2.1 + ClassDiffusion
CFG10.45%15.24%
CFG + SAG8.01%6.20%
CFG + AG12.60%6.20%
Ours55.22%32.04%
Undecided13.67%40.31%
SANA + DB-LoRA
CFG2.48%3.08%
CFG + AG2.17%2.77%
Ours68.32%8.64%
Undecided27.01%85.49%
", + "image_path": "91489bc95955dc44d77e01a3eed13a0fdfe422e08c3aabf176fc9dbff4303b07.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 408, + 284, + 419 + ], + "lines": [ + { + "bbox": [ + 66, + 408, + 284, + 419 + ], + "spans": [ + { + "bbox": [ + 66, + 408, + 284, + 419 + ], + "type": "text", + "content": "Table 3. User preference on subject fidelity and text fidelity." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 426, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 640 + ], + "type": "text", + "content": "Qualitative Results. Fig. 5 and Fig. 6 show the comparison results based on SD 2.1 and SANA, respectively. As shown in Fig. 5, CFG often fails to capture fine details of the reference subject (e.g., cat's appearance and the teddybear's eye). Additionally, applying SAG or AG has minimal impact in some cases (first row of Fig. 5) or even degrades image quality, introducing artifacts such as a teddy bear with three eyes (CFG+AG, second row of Fig. 5). In contrast, our method consistently achieves better subject fidelity with the same noise input. Similarly, in Fig. 6, our method better preserves subject details compared to CFG and CFG+AG, as evidenced by the significant improvement in subject fidelity in quantitative results. For instance, it accurately retains the ears and color of a dog (first row of Fig. 6) and the shape and numerical details of a clock (second row of Fig. 6). In contrast, applying AG with CFG shows little to no improvement despite incurring a higher computational cost compared to both CFG and our guidance method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "User Study. Table 3 presents the user preference results comparing our method with baseline approaches. The results show that our guidance technique achieves significantly higher subject fidelity than other guidance methods. Our approach also shows superior text fidelity in user evaluations. Further details can be found in the supplementary." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 259, + 550, + 439 + ], + "blocks": [ + { + "bbox": [ + 370, + 243, + 498, + 259 + ], + "lines": [ + { + "bbox": [ + 370, + 243, + 498, + 259 + ], + "spans": [ + { + "bbox": [ + 370, + 243, + 498, + 259 + ], + "type": "text", + "content": "Subject Fidelity (DINO) vs Text Fidelity (CLIP-T) (guidance scale " + }, + { + "bbox": [ + 370, + 243, + 498, + 259 + ], + "type": "inline_equation", + "content": "\\lambda \\in [2.5, 5.0]" + }, + { + "bbox": [ + 370, + 243, + 498, + 259 + ], + "type": "text", + "content": ", step = 0.5)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 259, + 550, + 439 + ], + "lines": [ + { + "bbox": [ + 316, + 259, + 550, + 439 + ], + "spans": [ + { + "bbox": [ + 316, + 259, + 550, + 439 + ], + "type": "image", + "image_path": "1a917af5ed98a8aec4c07d008536465cd3e144fa6ba21d34b16f34e6522c3167.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 440, + 554, + 507 + ], + "lines": [ + { + "bbox": [ + 313, + 440, + 554, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 440, + 554, + 507 + ], + "type": "text", + "content": "Figure 8. Ablation study on the guidance scale " + }, + { + "bbox": [ + 313, + 440, + 554, + 507 + ], + "type": "inline_equation", + "content": "\\lambda \\in [2.5, 5.0]" + }, + { + "bbox": [ + 313, + 440, + 554, + 507 + ], + "type": "text", + "content": ", measuring performance at 0.5 intervals using DB-LoRA based on SANA. When " + }, + { + "bbox": [ + 313, + 440, + 554, + 507 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 440, + 554, + 507 + ], + "type": "text", + "content": " is set to 0.2, we can improve both subject and text fidelity compared to CFG, regardless of guidance scale " + }, + { + "bbox": [ + 313, + 440, + 554, + 507 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 440, + 554, + 507 + ], + "type": "text", + "content": ". Note that CLIP-T score is slightly changed in the graph " + }, + { + "bbox": [ + 313, + 440, + 554, + 507 + ], + "type": "inline_equation", + "content": "(max(CLIP-T) - min(CLIP-T) < 0.004)" + }, + { + "bbox": [ + 313, + 440, + 554, + 507 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 514, + 417, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 417, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 417, + 526 + ], + "type": "text", + "content": "4.3. Further Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 533, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 533, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 533, + 555, + 715 + ], + "type": "text", + "content": "Ablation Study on " + }, + { + "bbox": [ + 312, + 533, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 312, + 533, + 555, + 715 + ], + "type": "text", + "content": ". Fig. 4 illustrates how DINO and CLIP-T scores vary with changes in " + }, + { + "bbox": [ + 312, + 533, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 312, + 533, + 555, + 715 + ], + "type": "text", + "content": ". For both SD 1.5 and SD 2.1, personalization guidance slightly reduces text fidelity but significantly enhances subject fidelity compared to CFG (" + }, + { + "bbox": [ + 312, + 533, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\omega = 1" + }, + { + "bbox": [ + 312, + 533, + 555, + 715 + ], + "type": "text", + "content": "). By appropriately adjusting " + }, + { + "bbox": [ + 312, + 533, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 312, + 533, + 555, + 715 + ], + "type": "text", + "content": ", it is possible to either boost text fidelity while maintaining subject fidelity (as indicated by the pink arrow) or enhance subject fidelity while preserving text fidelity (as indicated by the blue arrow). It is important to note that these trends may vary depending on the personalization technique or the pre-trained diffusion model. However, the ability to easily fine-tune results via weight interpolation at inference, without modifying training steps or learning rates, offers a major practical advantage over traditional methods for controlling overfitting. Fig. 7 further visualizes the impact of " + }, + { + "bbox": [ + 312, + 533, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 312, + 533, + 555, + 715 + ], + "type": "text", + "content": " on gen" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15913" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 64, + 161, + 112 + ], + "blocks": [ + { + "bbox": [ + 59, + 64, + 161, + 112 + ], + "lines": [ + { + "bbox": [ + 59, + 64, + 161, + 112 + ], + "spans": [ + { + "bbox": [ + 59, + 64, + 161, + 112 + ], + "type": "table", + "html": "
MethodPickScore↑
DPO+CFG21.53
DPO+Ours21.58
", + "image_path": "0cadc9c90ad9a3a236e20a924c06e815efbd2aa75916588431776a2c1f0e4c2d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 167, + 64, + 296, + 111 + ], + "blocks": [ + { + "bbox": [ + 167, + 64, + 296, + 111 + ], + "lines": [ + { + "bbox": [ + 167, + 64, + 296, + 111 + ], + "spans": [ + { + "bbox": [ + 167, + 64, + 296, + 111 + ], + "type": "table", + "html": "
MethodDreamSim↓
PairCustom+CFG0.7993
PairCustom+Ours0.7344
", + "image_path": "3349415b8f3464d29f6a27bc26c2387fd9df7eb55e43b94c91ea0fedc54b6b34.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 113, + 165, + 135 + ], + "lines": [ + { + "bbox": [ + 56, + 113, + 165, + 135 + ], + "spans": [ + { + "bbox": [ + 56, + 113, + 165, + 135 + ], + "type": "text", + "content": "Table 4. Results of combining Diffusion-DPO with ours." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 167, + 114, + 294, + 135 + ], + "lines": [ + { + "bbox": [ + 167, + 114, + 294, + 135 + ], + "spans": [ + { + "bbox": [ + 167, + 114, + 294, + 135 + ], + "type": "text", + "content": "Table 5. Results of combining Pair-Custom with ours." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 59, + 142, + 290, + 380 + ], + "blocks": [ + { + "bbox": [ + 59, + 142, + 290, + 380 + ], + "lines": [ + { + "bbox": [ + 59, + 142, + 290, + 380 + ], + "spans": [ + { + "bbox": [ + 59, + 142, + 290, + 380 + ], + "type": "image", + "image_path": "c2b3fb3b5cfe9e7d732fe412e893c8b52684fff8cbef965c1f757c093092cb5d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 383, + 293, + 393 + ], + "lines": [ + { + "bbox": [ + 56, + 383, + 293, + 393 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 293, + 393 + ], + "type": "text", + "content": "Figure 9. Qualitative results of our guidance across diverse tasks." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 401, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 295, + 460 + ], + "type": "text", + "content": "eration quality. The first row shows that higher " + }, + { + "bbox": [ + 55, + 401, + 295, + 460 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 55, + 401, + 295, + 460 + ], + "type": "text", + "content": " values retain fine details, such as the intricate patterns on the vase, at the cost of weaker adherence to the \"purple\" text prompt. Additionally, as " + }, + { + "bbox": [ + 55, + 401, + 295, + 460 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 55, + 401, + 295, + 460 + ], + "type": "text", + "content": " decreases, subject fidelity improves significantly, particularly in the range of 0.6 to 1.0." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "text", + "content": "Ablation Study on Guidance Scale. Fig. 8 shows how personalization performance varies with different guidance scales " + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "text", + "content": ". We evaluate our guidance at " + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\omega = 0" + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\omega = 0.2" + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "text", + "content": ", where text fidelity remains preserved. As " + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "text", + "content": " decreases, text fidelity deteriorates for both guidance methods. However, as " + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "text", + "content": " increases, our guidance method shows improvements in both the DINO score and CLIP-T score, whereas the DINO score of CFG declines. This suggests that CFG reduces subject fidelity, and adjusting the guidance scale alone is insufficient to correct this effect." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 588, + 285, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 285, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 285, + 601 + ], + "type": "text", + "content": "4.4. Application to Fine-tuned Diffusion Models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": "Fine-tuning diffusion model extends beyond personalization tasks to applications such as style personalization [17], human preference learning [21, 33], and specific task learning [2, 15]. Our personalization guidance framework leverages the knowledge gap between the pre-trained and fine-tuned models, emphasizing the information better captured by the fine-tuned model. This concept can be generalized to control the degree of adaptation for other fine-tuned models, such as adjusting the extent of human preference alignment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 481, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 481, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 481, + 84 + ], + "type": "text", + "content": "or style personalization during inference." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 84, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 84, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 84, + 553, + 144 + ], + "type": "text", + "content": "To this end, to investigate our method's generalizability across diverse tasks, Fig. 9 show the following experiments: (a) Human Preference Optimization (Diffusion-DPO [33]), (b) Style Personalization (PairCustom [17]), and (c) Instruction-based Editing (InstructPix2Pix [2])." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 144, + 555, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 275 + ], + "type": "text", + "content": "(a) Human Preference Optimization: For evaluation with the Diffusion-DPO model [33], we use the PartiPrompts dataset [39] and report PickScore [39], which reflects human-perceived image quality. Table 4 shows that integrating our method into the SD 1.5-based Diffusion-DPO model yields superior performance compared to CFG, confirming the effectiveness of our approach on human preference optimization tasks. As shown in Fig. 9 (a), our method helps generate more natural-looking images and corrects artifacts that would otherwise appear unnatural (e.g., additional finger)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 275, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 555, + 407 + ], + "type": "text", + "content": "(b) Style Personalization: To evaluate style alignment, we use two pre-trained style-LoRAs released by PairCustom, combining each with both CFG and our method. Using text prompts from the StyleAlign [37] paper, we generate 800 images and compute the DreamSim score [7] to quantify the similarity between the target style and generated outputs. As shown in Table 5, applying our method significantly improves style alignment, demonstrating its effectiveness for style personalization. In Fig. 9 (b), combining PairCustom with our guidance leads to a more faithful transfer of the painting style from the reference image." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 407, + 554, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 554, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 554, + 467 + ], + "type": "text", + "content": "(c) Instruction-based Editing: In Fig. 9 (c), our guidance improves the performance of the InstructPix2Pix [2], which is designed to edit images based on text instructions, to better follow the given instruction. As a result, the generated outputs are more accurately aligned with the intended edits." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 467, + 554, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 554, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 554, + 539 + ], + "type": "text", + "content": "These results demonstrate that applying our guidance to various fine-tuned diffusion models allows them to generate images more faithfully aligned with their respective finetuning objectives. This highlights the generalizability of our method across diverse tasks. Further implementation details are provided in the supplementary." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 549, + 388, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 549, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 549, + 388, + 562 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "This paper tackles the critical challenge of balancing adaptation to the target distribution and knowledge preservation from the original model in personalized diffusion models. Our approach simply modifies the existing guidance by adjusting the degree of unlearning in the weak model through weight interpolation between the pre-trained and fine-tuned models. This simple yet effective guidance enhances subject fidelity while preserving text editability, ensuring an optimal trajectory for personalization. Furthermore, we demonstrate that our guidance can be effectively applied to a wide range of fine-tuned diffusion models, such as human preference learning." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15914" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 95 + ], + "type": "text", + "content": "Acknowledgement We would like to thank the Qualcomm AI Research team for their valuable discussions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 108, + 115, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 108, + 115, + 120 + ], + "spans": [ + { + "bbox": [ + 56, + 108, + 115, + 120 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 128, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 128, + 295, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 128, + 295, + 193 + ], + "spans": [ + { + "bbox": [ + 61, + 128, + 295, + 193 + ], + "type": "text", + "content": "[1] Donghoon Ahn, Hyoungwon Cho, Jaewon Min, Wooseok Jang, Jungwoo Kim, SeonHwa Kim, Hyun Hee Park, Kyong Hwan Jin, and Seungryong Kim. Self-rectifying diffusion sampling with perturbed-attention guidance. In European Conference on Computer Vision, pages 1-17. Springer, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 195, + 295, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 195, + 295, + 249 + ], + "spans": [ + { + "bbox": [ + 62, + 195, + 295, + 249 + ], + "type": "text", + "content": "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18392-18402, 2023. 3, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 251, + 295, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 295, + 306 + ], + "type": "text", + "content": "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 308, + 296, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 296, + 361 + ], + "type": "text", + "content": "[4] Kelvin CK Chan, Yang Zhao, Xuhui Jia, Ming-Hsuan Yang, and Huisheng Wang. Improving subject-driven image synthesis with subject-agnostic guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6733–6742, 2024. 3, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 364, + 295, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 364, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 364, + 295, + 396 + ], + "type": "text", + "content": "[5] Shangyu Chen, Zizheng Pan, Jianfei Cai, and Dinh Phung. Para: Personalizing text-to-image diffusion via parameter rank reduction. arXiv preprint arXiv:2406.05641, 2024. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 398, + 295, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 398, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 62, + 398, + 295, + 452 + ], + "type": "text", + "content": "[6] Wenhu Chen, Hexiang Hu, Yandong Li, Nataniel Ruiz, Xuhui Jia, Ming-Wei Chang, and William W Cohen. Subject-driven text-to-image generation via apprenticeship learning. Advances in Neural Information Processing Systems, 36:30286–30305, 2023. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 454, + 295, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 454, + 295, + 507 + ], + "spans": [ + { + "bbox": [ + 62, + 454, + 295, + 507 + ], + "type": "text", + "content": "[7] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. arXiv preprint arXiv:2306.09344, 2023.8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 510, + 295, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 510, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 62, + 510, + 295, + 564 + ], + "type": "text", + "content": "[8] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In The Eleventh International Conference on Learning Representations. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 567, + 295, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 295, + 609 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 295, + 609 + ], + "type": "text", + "content": "[9] Shaozhe Hao, Kai Han, Shihao Zhao, and Kwan-Yee K Wong. Vico: Plug-and-play visual condition for personalized text-to-image generation. arXiv preprint arXiv:2306.00971, 2023.5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 612, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 612, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 612, + 295, + 644 + ], + "type": "text", + "content": "[10] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 2, 3, 4, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 647, + 295, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 677 + ], + "type": "text", + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "text", + "content": "[12] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seunglyon Kim. Improving sample quality of diffusion models using self-attention guidance. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "International Conference on Computer Vision, pages 7462-7471, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 97, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 139 + ], + "type": "text", + "content": "[13] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 2, 3, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 141, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 185 + ], + "type": "text", + "content": "[14] Jiannan Huang, Jun Hao Liew, Hanshu Yan, Yuyang Yin, Yao Zhao, and Yunchao Wei. Classdiffusion: More aligned personalization tuning with explicit class guidance. arXiv preprint arXiv:2405.17532, 2024. 2, 3, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 186, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 230 + ], + "type": "text", + "content": "[15] Lianghua Huang, Wei Wang, Zhi-Fan Wu, Yupeng Shi, Huanzhang Dou, Chen Liang, Yutong Feng, Yu Liu, and Jingren Zhou. In-context lora for diffusion transformers. arXiv preprint arXiv:2410.23775, 2024. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "type": "text", + "content": "[16] Junha Hyung, Kinam Kim, Susung Hong, Min-Jung Kim, and Jaegul Choo. Spatiotemporal skip guidance for enhanced video diffusion sampling. arXiv preprint arXiv:2411.18664, 2024.3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 277, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 277, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 277, + 553, + 319 + ], + "type": "text", + "content": "[17] Maxwell Jones, Sheng-Yu Wang, Nupur Kumari, David Bau, and Jun-Yan Zhu. Customizing text-to-image models with a single image pair. In SIGGRAPH Asia 2024 Conference Papers, pages 1-13, 2024. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 322, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 322, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 317, + 322, + 553, + 374 + ], + "type": "text", + "content": "[18] Tero Karras, Miika Aittala, Tuomas Kynkänniemi, Jaakko Lehtinen, Timo Aila, and Samuli Laine. Guiding a diffusion model with a bad version of itself. Advances in Neural Information Processing Systems, 37:52996-53021, 2025. 2, 3, 4, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 377, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 377, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 317, + 377, + 553, + 421 + ], + "type": "text", + "content": "[19] Artur Kasymov, Marcin Sendera, Michal Stypulkowski, Macej Zięba, and Przemysław Spurek. Autolora: Autoguidance meets low-rank adaptation for diffusion models. arXiv preprint arXiv:2410.03941, 2024. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 422, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 553, + 477 + ], + "type": "text", + "content": "[20] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1931-1941, 2023. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 479, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 479, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 479, + 553, + 521 + ], + "type": "text", + "content": "[21] Shufan Li, Konstantinos Kallidromitis, Akash Gokul, Yusuke Kato, and Kazuki Kozuka. Aligning diffusion models by optimizing human utility. Advances in Neural Information Processing Systems, 37:24897-24925, 2025. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 523, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 553, + 589 + ], + "type": "text", + "content": "[22] Jisu Nam, Heesu Kim, DongJae Lee, Siyoon Jin, Seungryong Kim, and Seunggyu Chang. Dreamcatcher: appearance matching self-attention for semantically-consistent text-to-image personalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8100–8110, 2024. 3, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 590, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 633 + ], + "type": "text", + "content": "[23] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195-4205, 2023. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "text", + "content": "[24] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15915" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "text", + "content": "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "type": "text", + "content": "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2, 4, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 207 + ], + "type": "text", + "content": "[27] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, pages 22500-22510, 2023. 2, 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 296, + 274 + ], + "type": "text", + "content": "[28] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdream booth: Hypermetworks for fast personalization of text-to-image models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6527-6536, 2024. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 296, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 296, + 331 + ], + "type": "text", + "content": "[29] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 35:36479-36494, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 296, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 296, + 386 + ], + "type": "text", + "content": "[30] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8543-8552, 2024. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 296, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 296, + 432 + ], + "type": "text", + "content": "[31] Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. Freeu: Free lunch in diffusion u-net. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4733-4743, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 296, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 296, + 466 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 296, + 466 + ], + "type": "text", + "content": "[32] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 468, + 296, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 296, + 533 + ], + "type": "text", + "content": "[33] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228-8238, 2024. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 535, + 296, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 296, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 296, + 600 + ], + "type": "text", + "content": "[34] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15943-15953, 2023. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 602, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 296, + 678 + ], + "type": "text", + "content": "[35] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International conference on machine learning, pages 23965-23998. PMLR, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 680, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 296, + 714 + ], + "type": "text", + "content": "[36] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 252 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "Namkoong, et al. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7959-7971, 2022. 3, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "text", + "content": "[37] Zongze Wu, Yotam Nitzan, Eli Shechtman, and Dani Lischinski. Stylealign: Analysis and applications of aligned stylegan models. arXiv preprint arXiv:2110.11323, 2021. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "text", + "content": "[38] Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, et al. Sana: Efficient high-resolution image synthesis with linear diffusion transformers. arXiv preprint arXiv:2410.10629, 2024. 2, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 197, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 252 + ], + "type": "text", + "content": "[39] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 8" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "15916" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Stepping Out of Similar Semantic Space for Open-Vocabulary Segmentation/f495e889-2c8e-4e64-ba9d-86b1aec5716d_content_list.json b/2025/Stepping Out of Similar Semantic Space for Open-Vocabulary Segmentation/f495e889-2c8e-4e64-ba9d-86b1aec5716d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..97a29f8b0bb42ac1d11e13757b9db49a497ee178 --- /dev/null +++ b/2025/Stepping Out of Similar Semantic Space for Open-Vocabulary Segmentation/f495e889-2c8e-4e64-ba9d-86b1aec5716d_content_list.json @@ -0,0 +1,1534 @@ +[ + { + "type": "text", + "text": "Stepping Out of Similar Semantic Space for Open-Vocabulary Segmentation", + "text_level": 1, + "bbox": [ + 114, + 130, + 883, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yong Liu $^{1,2*}$ , Song-Li Wu $^{1,2*}$ , Sule Bai $^{1,2*}$ , Jiahao Wang $^{3}$ , Yitong Wang $^{4}$ , Yansong Tang $^{1,2\\dagger}$ $^{1}$ Tsinghua Shenzhen International Graduate School $^{2}$ Tsinghua University \n $^{3}$ The University of Hong Kong $^{4}$ ByteDance Inc.", + "bbox": [ + 142, + 178, + 861, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "liuyong23@mails.tsinghua.edu.cn, tang.yansong@sz.tsinghua.edu.cn", + "bbox": [ + 210, + 236, + 779, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 301, + 326, + 316 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Open-vocabulary segmentation aims to achieve segmentation of arbitrary categories given unlimited text inputs as guidance. To achieve this, recent works have focused on developing various technical routes to exploit the potential of large-scale pre-trained vision-language models and have made significant progress on existing benchmarks. However, we find that existing test sets are limited in measuring the models' comprehension of \"open-vocabulary\" concepts, as their semantic space closely resembles the training space, even with many overlapping categories. To this end, we present a new benchmark named OpenBench that differs significantly from the training semantics. It is designed to better assess the model's ability to understand and segment a wide range of real-world concepts. When testing existing methods on OpenBench, we find that their performance diverges from the conclusions drawn on existing test sets. In addition, we propose a method named OVSNet to improve the segmentation performance for diverse and open scenarios. Through elaborate fusion of heterogeneous features and cost-free expansion of the training space, OVSNet achieves state-of-the-art results on both existing datasets and our proposed OpenBench. Corresponding analysis demonstrates the rationality and effectiveness of our proposed benchmark and method.", + "bbox": [ + 89, + 334, + 483, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 715, + 220, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semantic segmentation is one of the most fundamental tasks in computer vision, which targets at assigning semantic category to pixels in an image. Despite achieving excellent progress in recent years [2, 6, 20, 21, 33-36, 38, 42, 49-52, 58], traditional semantic segmentation approaches rely on predefined categories sets and tend to falter when encountering categories absent during the training phase, significantly impeding their real-world application. Such chal", + "bbox": [ + 89, + 741, + 482, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/272a77adb469063800da3e373da97c34de40af323c96325c224e0ec6c28273b3.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 514, + 285, + 728, + 415 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b8593b55c6cc94d7cfe44a2bb7cad13e6c9440d3dfe8f2dca35986c7e7469c92.jpg", + "image_caption": [ + "(b)", + "Figure 1. (a) shows the similarity distribution of the existing test set (blue) and our OpenBench (orange) with the training semantic space. (b) illustrates the performance gap between the methods finetuning [10] and frozen [61] CLIP on different benchmarks." + ], + "image_footnote": [], + "bbox": [ + 730, + 286, + 903, + 412 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lenge has inspired the exploration of Open-Vocabulary Segmentation (OVS) task [12, 15, 17, 18, 27, 37, 55, 64]. Different from traditional closed-set segmentation, OVS methods aim to segment arbitrary categories under the guidance of given text inputs, which has many potential applications such as auto-driving and human-robot interaction [19, 39].", + "bbox": [ + 511, + 503, + 905, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, it is challenging to identify unseen categories without the intervention of external knowledge. Therefore, an intuitive idea is to introduce large-scale vision-language model [24, 46] trained with numerous sources to extend the semantic space of segmentation models. Following this idea, existing OVS methods [10, 31, 37, 57, 61, 62] concentrate on exploring how to exploit the pretrained knowledge prior more effectively and derive several technical routes. Early methods [12, 31, 60] adopt a two-stage pipeline. This approach first generates class-agnostic mask proposals with segmentation models, following which the pre-trained CLIP [46] serves as an additional classifier to execute region-level open-vocabulary classification. However, this strategy incurs heavy computation overhead, as it requires repeated forward passes of the CLIP vision encoder for each mask. This can be prohibitively expensive for real-world applications. To address this problem, researchers [25, 61–63] propose to unify classification and segmentation within a shared space, i.e., integrating frozen CLIP features into the segmenter and extracting CLIP embeddings from potential", + "bbox": [ + 511, + 598, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 107, + 875, + 215, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 109, + 887, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22664", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "target regions to enable simultaneous classification and segmentation. These methods achieve remarkable performance in a more efficient one-stage pipeline. Subsequently, some works [10, 57] discover that finetuning the CLIP encoder could further enhance the model's performance on existing test sets. They develop an early fusion paradigm based on cost map and achieve outstanding segmentation results. However, this performance boost is somewhat counterintuitive, as fine-tuning CLIP typically adapts it to specific training semantic spaces, which could reduce its original generalization ability. Yet, the ability of generalized visual-linguistic alignment is precisely what the open-vocabulary task requires. So why does fine-tuning CLIP appear to benefit the performance of OVS?", + "bbox": [ + 89, + 90, + 480, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This question drives us to revisit the existing evaluation benchmarks of the cross-dataset evaluation paradigm. We count the maximum similarity of each category in test set to the training semantics. Figure 1 (a) illustrates the similarity distribution of all test categories to the training space (blue color). Results indicate that the existing test data is, in fact, highly similar to the training semantic space. Specifically, even the most challenging ADE-847 [66] and PC-459 [44] datasets have an average similarity to the training classes of 0.79 and 0.83. PC-59 [44] and VOC [13] datasets even have a similarity to the training semantics of 0.95 and 0.97. This explains why finetuning CLIP on training set can lead to further performance gains on existing benchmarks. However, it somewhat conflicts with the purpose of the open-vocabulary setting, which aims to achieve segmentation of unlimited categories in the real world.", + "bbox": [ + 91, + 305, + 483, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above problem and provide a more comprehensive measure of OVS models, we propose a new evaluation benchmark named OpenBench, which is characterized by significant semantic differences from the training data. Figure 1 (a) also demonstrates the similarity distribution of our OpenBench (orange color). It can be seen that our benchmark contains more novel classes compared to the existing test sets. We evaluate the performance of approaches across different technical routes on our benchmark, and the findings diverge from those obtained on existing test sets. As shown in Figure 1 (b), while finetuning CLIP yields steady performance gains on existing datasets, it shows a significant drop on our benchmark. This outcome also supports our motivation for proposing a benchmark that diverges substantially from the training space.", + "bbox": [ + 89, + 550, + 482, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, we also present a framework named OVSNet to improve the segmentation performance for diverse and open scenarios. In OVSNet, we adopt a Proxy Calibration (PC) training strategy to enhance the robustness of model representations by broadening the training space with no additional cost. This strategy leverages the synthesis of proxy embeddings, which approximate unseen semantics through convex combinations of in-vocabulary classes. Addition", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ally, inspired by random walk [16], we propose a gradient-free fusion of CLIP embeddings and segmentation decoder features to model a robust joint space and mitigate potential overfitting issues related to training semantics. With the above proposed designs calibrating model knowledge from the aspects of training space and features separately, our OVSNet demonstrates strong performance across a wide range of scenarios, both on existing test sets similar to the training semantics and on our proposed novel OpenBench. Related ablations also demonstrate the effectiveness and soundness of our proposed strategies.", + "bbox": [ + 511, + 90, + 903, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 532, + 257, + 857, + 271 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a new benchmark named OpenBench for OVS, which is characterized by significant semantic differences from the training space. While using it to evaluate the comprehension of OVS methods in more open scenarios, we reach a different conclusion from that of the existing test sets.", + "- We propose OVSNet to improve the segmentation performance on diverse and open scenarios. It adopts a Proxy Calibration training strategy to broaden training space without additional cost. Besides, a gradient-free algorithm is leveraged to fuse CLIP embeddings and segmentation features for modeling robust joint space.", + "- Our method achieves state-of-the-art performance on existing benchmarks and the proposed OpenBench. Extensive experiments are conducted to prove the effectiveness and soundness of the proposed algorithms." + ], + "bbox": [ + 513, + 272, + 903, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 527, + 653, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Open-Vocabulary Segmentation. This task aims to segment an image and identify regions with arbitrary given text queries [1, 4, 15, 41, 55]. Pioneering work [55] replaces the output convolution layer by computing the similarity between visual features and linguistic embeddings, which has become common practice. Later, some methods [12, 15, 31, 59, 60] propose a two-stage pipeline: the model first generates class-agnostic mask proposals, then a pretrained CLIP [46] is utilized to perform sub-image classification by cropping and masking corresponding regions. With the combination of both in-vocabulary and out-vocabulary classification, these methods obtain great performance. However, this strategy incurs heavy computation overhead, as it requires repeated forward passes of the CLIP vision encoder for each mask. This can be prohibitively expensive for real-world applications. To address this problem, subsequent researchers propose to unify classification and segmentation within a shared space. Among them, SAN [61] designs a side-adapter network to leverage CLIP features for reorganizing segmentation and classification. GKC [17] presents text-guided knowledge distillation strategy to transfer CLIP knowledge to specific classification layer. FCCLIP [63] proposes to leverage", + "bbox": [ + 511, + 553, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22665", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8709d79e14e101d89d52c4c2862f7e91ac20ccbf8812adc9082fc19c1e860acc.jpg", + "image_caption": [ + "Figure 2. Illustration of the samples in the proposed OpenBench. Compared to existing test sets, our OpenBench has semantic categories that differ more from the training space, which helps to provide a more comprehensive measure of the model's ability to perform open-vocabulary segmentation. The corresponding category name is labeled in the lower right corner of each image." + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 898, + 296 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/0fb306c615cc2275c9e3d5422476faacc1569d96d62de2d35193f5136e179158.jpg", + "table_caption": [ + "Table 1. Comparison of the statistics of our proposed OpenBench with existing datasets. 'Cls', 'Img', and 'Sim' indicate class, image, and similarity, respectively. ADE, PC, and VOC denote ADE20K [66], Pascal Context [44], and Pascal VOC [13]" + ], + "table_footnote": [], + "table_body": "
DatasetCls Num.Img Num.Mean Sim.Median Sim.Min Sim.Max Sim.Fine GranularitySemantic Duplication
VOC [13]2014490.97561.00000.84491.0000××
PC-59 [44]5951050.94811.00000.76581.0000××
ADE-150 [66]15020000.80860.82280.19601.0000×✓
Cityscapes [11]331000.84880.83700.67461.0000××
PC-459 [44]45951050.83740.81240.67181.0000✓✓
ADE-847 [66]84720000.79260.79130.19601.0000✓✓
OpenBench (Ours)28660560.61420.64520.26080.7947✓×
", + "bbox": [ + 91, + 386, + 898, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the local-awareness advantage of convolution backbones. Some works [10, 57] then discover that finetuning CLIP encoder could further enhance the model's performance on existing test sets and develop effective one-stage framework.", + "bbox": [ + 89, + 537, + 482, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, we find that existing test sets are highly similar to the training set, which limits their ability to effectively evaluate the model's comprehension of the open vocabulary semantics. Therefore, in this paper we propose a new benchmark OpenBench that is characterized by significant semantic differences from the training data. We also propose a method named OVSNet to improve the segmentation performance for diverse and open scenarios.", + "bbox": [ + 89, + 599, + 482, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision-Language Pre-training. Vision-language pretraining aims to learn a unified visual-linguistic representation space. Early approaches [7, 28, 30, 40, 45, 47, 48, 65], constrained by small-scale datasets, struggle to achieve strong performance and require fine-tuning for downstream tasks. However, with the advent of large-scale web data, recent works [23, 46] have demonstrated the benefits of leveraging such data to train a more robust multi-modal representation space. Among these, CLIP [46], the most widely used vision-language model, employs contrastive learning to link images with their corresponding captions,", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "achieving impressive cross-modal alignment. Building upon previous works [12, 22, 27, 29, 60], we also utilize the well-aligned and generalized space of CLIP to enhance open-vocabulary segmentation.", + "bbox": [ + 511, + 537, + 906, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Benchmark", + "text_level": 1, + "bbox": [ + 511, + 612, + 635, + 627 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Dataset Introduction", + "text_level": 1, + "bbox": [ + 511, + 637, + 709, + 651 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Table 1, we present the statistics comparison of our proposed OpenBench with existing popular open-vocabulary segmentation datasets, including ADE-150 [66], ADE-847 [66], PC-59 [44], PC-459 [44], VOC [13], and Cityscapes [11]. Specifically, our OpenBench contains 286 categories, of which the maximum similarity to the training categories is 0.7947, the minimum is 0.2608, and the average similarity is 0.6142. It can be observed that, compared to existing datasets, our OpenBench exhibits a greater disparity from the training set, which better evaluates the model's segmentation ability for arbitrary open categories.", + "bbox": [ + 511, + 657, + 906, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Besides, our OpenBench contains fine-grained categories, e.g., \"oystercatcher\", rather than just generic broad category such as \"bird\" (corresponds to the \"Fine Granularity\" characterization in Table 1). This is important for evaluating whether the OVS models preserve the general", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22666", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ization capability of CLIP [46], which is inherently capable of detecting these fine-grained categories. Actually, PC-459 [44] and ADE-847 [66] also have fine-grained categories, but they suffer from semantic duplication problem. In detail, semantic duplication means that there will be both fine-grained categories as well as corresponding broad categories in these datasets, which would cause bias in the metrics calculation. For instance, if a model assigns the \"door\" tag to a region whose ground truth label is \"double door\", it will be considered incorrect. For open-vocabulary segmentation setting, we argue that the responsibility of the model is to discern the correct semantic. The distribution of such duplication categories and the corresponding performance statistics is unreasonable. SCAN [37] also points similar observations and try to solve this by designing new metric. However, this requires manual judgment of the relationships between categories, introducing additional cost and uncertainty. Our OpenBench, on the other hand, has fine-grained semantics without the problem of semantic duplication, which helps to more accurately reflect the real capabilities of related models.", + "bbox": [ + 93, + 90, + 480, + 406 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 2 illustrates part of the categories contained in OpenBench, where some categories such as food related even differ significantly from the visual domain of the training set. Furthermore, our OpenBench also accounts for \"other\" or \"background\" pixels in images, which are often overlooked by existing datasets. This is important for open-vocabulary segmentation because the model is essentially forced to choose the most appropriate category from a predefined set. If there is no correct class and no \"others\" option in the candidate set, the model would be compelled to make an incorrect prediction. Moreover, \"others\" or \"background\" require the model to consider not only the direct visual-linguistic correspondence between the image content and each class, but also the relative relationships among different classes. This adds complexity and makes the task more relevant to real-world scenarios.", + "bbox": [ + 93, + 410, + 480, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Dataset Collection", + "text_level": 1, + "bbox": [ + 94, + 666, + 264, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our OpenBench is primarily derived from existing accessible segmentation datasets. To obtain semantic categories that differ significantly from the training set, i.e., COCO [32], we first organize a group of segmentation datasets that possess category annotations, e.g., Food103 [53], Imagenet-S [14], and CamVid [3]. For all categories in the collected data, we leverage CLIP-L/14 [46] to compute their maximum similarity with COCO categories. Then we calculate the minimum similarity among the categories present in each image as the image's similarity to the training semantics. If the image similarity is greater than the set threshold $\\sigma_{1}$ , the image will be filtered. Conversely, if the similarity of the image is less than the threshold $\\sigma_{1}$ , we would set the category exist in the image", + "bbox": [ + 93, + 689, + 480, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "whose similarity is greater than the threshold $\\sigma_{2}$ to \"others\" and save the related images and annotation. Finally, after manually filtering the categories for potential conflicts or duplicates, we get the OpenBench.", + "bbox": [ + 517, + 92, + 903, + 150 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 516, + 167, + 602, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Overview", + "text_level": 1, + "bbox": [ + 516, + 194, + 620, + 207 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall pipeline of our proposed OVSNet is shown in Figure 3. We leverage the vision and text encoder of pretraind CLIP [46] to encode the input image and category names, respectively. Then, a segmentation decoder is trained to decode the visual features, generating masks for potential targets and their corresponding query embeddings. We then leverage the predicted masks to extract related CLIP features with Mask Pooling operation. To integrate heterogeneous knowledge spaces and improve the representation robustness, Gradient-Free Aggregation fuses the query embeddings with pooled CLIP features. The aggregated visual embeddings are combined with linguistic embeddings to perform visual-linguistic alignment and generate final output. During training, a Proxy Calibration strategy is utilized to expand training space by applying supervision on the convex combination of training semantics.", + "bbox": [ + 516, + 215, + 903, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Gradient-Free Aggregation", + "text_level": 1, + "bbox": [ + 516, + 470, + 756, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although CLIP [46] exhibits strong generalization ability, it optimizes for image-level visual-linguistic alignment during pretraining, which creates a great domain gap compared to mask-level alignment. As a result, directly using CLIP features obtained through mask pooling for mask classification may lead to suboptimal performance. The query embeddings generated by the segmentation decoder, on the other hand, have a strong region-level alignment prior owing to their suitability for the training space. However, they struggle to recognize novel semantics.", + "bbox": [ + 516, + 492, + 903, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, we propose to reorganize them to obtain more robust domain-adapted representations. An intuitive approach is to learn the intrinsic connection between them through self-attention or cross-attention mechanisms. However, we observe that such learning-based strategy tends to develop excessive reliance on query embeddings fitted to seen semantics during the training process while neglecting the utilization of more generalizable information provided by CLIP, ultimately affecting the model's ability to understand unlimited categories. Therefore, we propose to combine CLIP features and query embeddings via a gradient-free approach to avoid potential overfitting problems during the learning of the joint space. This is inspired by Random Walk algorithm [16, 56]. Specifically, there are multiple iterations in this process where two features learn collaboratively in shared embedding space until convergence. Suppose we have $N$ queries. For convenience, we omit the", + "bbox": [ + 516, + 643, + 903, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22667", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c7a64c0b7842a0c3ebf6f55d5287a827062903b0d7a95554faaefef99e4654c1.jpg", + "image_caption": [ + "Figure 3. Pipeline of our OVSNet. The input image and corresponding categories are first encoded by pretrained CLIP [46]. A segmentation decoder is trained to decode the visual features and generate potential masks as well as corresponding query embeddings. We then leverage the predicted masks to extract related CLIP features with Mask Pooling operation. To integrate different knowledge spaces and improve the representation robustness, Gradient-Free Aggregation fuses the query embeddings with CLIP features. The aggregated visual embeddings are combined with class embeddings to perform visual-linguistic alignment and generate final output. During training, we apply Proxy Calibration strategy to expand training space in a cost-free manner." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 903, + 308 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "query index and take a random query as the example. The corresponding query embeddings and CLIP features can be represented by $F_{Q}$ and $F_{C}$ . $F_{Q}$ and $F_{C}$ are defined as the initial state $F_{Q}^{0}$ and $F_{C}^{0}$ at the 0-th iteration. Then we formulate affinity $\\mathcal{Z}$ by scaled dot product: $\\mathcal{Z} = \\lambda F_{C}^{0}(F_{Q}^{0})^{\\top}$ , where $\\lambda$ is the scaling factor. At $t$ -th iteration, the query embedding $F_{Q}^{t}$ is derived by the original embedding $F_{Q}^{0}$ and the updated CLIP feature $F_{C}^{t-1}$ from previous iteration, which can be formulated as:", + "bbox": [ + 88, + 420, + 480, + 559 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {Q} ^ {t} = \\omega \\operatorname {N o r m} (\\mathcal {Z}) ^ {\\top} F _ {C} ^ {t - 1} + (1 - \\omega) F _ {Q} ^ {0}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 565, + 482, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Subsequently, it is utilized to boost the CLIP feature of this iteration:", + "bbox": [ + 89, + 589, + 482, + 617 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {C} ^ {t} = \\omega \\mathcal {Z} F _ {Q} ^ {t} + (1 - \\omega) F _ {C} ^ {0}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 617, + 480, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\omega \\in (0,1)$ is the factor that controls the degree of feature fusion. By combining the above equations we can obtain the overall iterative update formula of $F_{C}^{t}$ :", + "bbox": [ + 89, + 638, + 483, + 685 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {C} ^ {t} = \\left(\\omega^ {2} A\\right) ^ {t} F _ {C} ^ {0} + (1 - \\omega) \\sum_ {i = 0} ^ {t - 1} \\left(\\omega^ {2} A\\right) ^ {i} \\left(\\omega \\mathcal {Z} F _ {Q} ^ {0} + F _ {C} ^ {0}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 689, + 482, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $A$ denotes $\\mathcal{Z}\\mathrm{Norm}(\\mathcal{Z})^{\\top}$ . Moreover, to avoid the potential issue of unexpected gradient and simplify the calculation process, we use an approximate inference function based on Neumann Series [43] when $t\\to \\infty$ . Then the update formula can be present as:", + "bbox": [ + 89, + 736, + 482, + 813 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {C} ^ {\\infty} = (1 - \\omega) \\left(I - \\omega^ {2} A\\right) ^ {- 1} \\left(\\omega \\mathcal {Z} F _ {Q} ^ {0} + F _ {C} ^ {0}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 816, + 482, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $I$ represents the identity matrix. With this gradient-free aggregation strategy, the model can recombine query embeddings with CLIP features to produce more robust region-level representations.", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Proxy Calibration", + "text_level": 1, + "bbox": [ + 513, + 419, + 689, + 435 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A broader semantic training space tends to result in more generalized model representations. However, expanding the training space for OVS is costly due to the intensive prediction nature of segmentation tasks. To alleviate this problem, we propose a cost-free approach named Proxy Calibration (PC) to expand the training space and enhance the generalization ability of the model. It leverages the synthesis of proxy embeddings, which approximates novel semantics via convex combination between the in-vocabulary classes.", + "bbox": [ + 511, + 450, + 906, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, having the aggregated query embeddings $F_{Q}$ , we perform random convex combination of them during training to simulate embedding space for undefined classes, as shown by the black dotted line in Figure 3. This process can be formulated as:", + "bbox": [ + 511, + 595, + 906, + 674 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {Q m n} ^ {\\prime} = \\alpha * F _ {Q m} + (1 - \\alpha) * F _ {Q n} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 702, + 906, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $m, n \\in \\{1, 2, \\dots, N\\}$ . $\\alpha$ is a random weight sampled from the distribution of $Beta(\\gamma, \\gamma)$ . $F'$ denotes the generated proxy embeddings. The same enhancement is performed for the corresponding CLIP feature $F_C$ and linguistic embeddings $F_T$ . Note that the text features here refer to the class representations of the ground truth regions corresponding to each query, which are obtained after Hungarian Matching [5]. Having the generated $F_Q'$ , $F_C'$ , and $F_T'$ , we perform distance supervision on them to promote region-level visual-linguistic alignment performance from the per", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22668", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c519e915a6a847400c6b085e623bc8297e063f203e1f7fd2067ffd780b7ea571.jpg", + "table_caption": [ + "Table 2. Performance comparison with state-of-the-art methods on existing benchmarks and the proposed OpenBench. ADE, PC, and VOC denote ADE20K [66], Pascal Context [44], and Pascal VOC [13], respectively. 171 and 133 indicate the number of training categories. * denotes the re-trained version according to the official code for fair comparison." + ], + "table_footnote": [], + "table_body": "
MethodTraining DatasetADE-150ADE-847PC-59PC-459VOCOpenBenchAverage Score
Base-level CLIP Backbone
SimSeg* [60]COCO-Stuff-17121.16.951.99.791.819.433.5
OVSeg [31]COCO-Stuff-17124.87.153.311.092.624.935.6
SAN [61]COCO-Stuff-17127.510.153.812.694.039.639.6
SCAN [37]COCO-Stuff-17130.810.858.413.297.038.741.5
CATSeg [10]COCO-Stuff-17131.812.057.519.094.636.141.8
SED [57]COCO-Stuff-17131.611.457.318.694.433.341.1
FCCLIP* [63]COCO-Panoptic-13331.113.554.812.893.240.340.9
MAFT [25]COCO-Panoptic-13329.110.153.512.890.037.538.8
MAFT+ [26]COCO-Panoptic-13334.613.857.516.295.443.743.5
OVSNet (Ours)COCO-Panoptic-13335.814.558.619.195.744.944.8
Large-level CLIP Backbone
SimSeg* [60]COCO-Stuff-17121.77.152.210.292.327.835.2
OVSeg [31]COCO-Stuff-17129.69.055.712.494.535.639.5
SAN [61]COCO-Stuff-17132.112.457.715.794.643.442.7
SCAN [37]COCO-Stuff-17133.514.059.316.397.149.444.9
SED [57]COCO-Stuff-17135.213.960.622.696.145.645.6
FCCLIP [63]COCO-Panoptic-13334.014.858.418.295.444.844.3
MAFT [25]COCO-Panoptic-13334.413.157.517.093.041.742.8
MAFT+ [26]COCO-Panoptic-13336.115.159.421.696.547.346.0
OVSNet (Ours)COCO-Panoptic-13337.116.262.023.596.948.647.4
", + "bbox": [ + 89, + 133, + 910, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "spective of proxy semantics. That is:", + "bbox": [ + 89, + 452, + 334, + 467 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {P Q} = 1 - \\frac {\\mathbf {F} _ {\\mathbf {Q}} ^ {\\prime} \\cdot \\mathbf {F} _ {\\mathbf {T}} ^ {\\prime}}{\\| \\mathbf {F} _ {\\mathbf {Q}} ^ {\\prime} \\| _ {2} \\| \\mathbf {F} _ {\\mathbf {T}} ^ {\\prime} \\| _ {2}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 474, + 482, + 512 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {P C} = 1 - \\frac {\\mathbf {F} _ {\\mathbf {C}} ^ {\\prime} \\cdot \\mathbf {F} _ {\\mathbf {T}} ^ {\\prime}}{\\| \\mathbf {F} _ {\\mathbf {C}} ^ {\\prime} \\| _ {2} \\| \\mathbf {F} _ {\\mathbf {T}} ^ {\\prime} \\| _ {2}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 513, + 482, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{PQ}$ and $\\mathcal{L}_{PC}$ denote the applied proxy supervision for query embeddings and pooled CLIP features, respectively. The final proxy loss is the sum of $\\mathcal{L}_{PQ}$ and $\\mathcal{L}_{PC}$ . By applying additional supervision in this way, the model is not limited to a specific training space and can learn more robust visual representations.", + "bbox": [ + 89, + 554, + 483, + 645 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 89, + 657, + 215, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Evaluation Benchmark and Metric", + "text_level": 1, + "bbox": [ + 89, + 681, + 395, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the effectiveness and generalization ability of our method on diverse scenarios, we conduct extensive experiments on the popular existing benchmarks, ADE20K150 [66], ADE20K847 [66], Pascal VOC [13], Pascal Context-59 [44], Pascal Context-459 [44] and our proposed OpenBench. ADE20K is a large-scale scene understanding benchmark, containing 20k training images, 2k validation images, and 3k testing images. There are two splits of this dataset. ADE20K-150 contains 150 semantic classes whereas ADE20K-847 has 847 classes. The images of both are the same. Pascal Context is an extension of Pascal VOC 2010, containing 5,005 validation images. We take the commonly used PC-59 and challenging PC-459", + "bbox": [ + 89, + 704, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "version for validation. Pascal VOC contains 11,185 training images and 1,449 validation images from 20 classes. We use the provided augmented annotations. Following previous works [12, 15, 60], we take the mean-intersection-over-union (mIoU) as the metric to compare our model with previous state-of-the-art methods. In ablation, we assume the use of base-level CLIP as the backbone for comparison.", + "bbox": [ + 511, + 450, + 906, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 569, + 732, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We utilize the Mask2Former [8] as the segmentation decoder and the implementation is based on detectron2 [54]. The model is trained with batch size of 8 and total iteration of $60\\mathrm{k}$ . The base learning rate is 0.0001 with a polynomial schedule. Following previous works [26, 63], we take CLIP [46] of ConvNeXt as the backbone. The input image is resized and cropped to $1024\\times 1024$ . For data augmentation, random flip and color augmentation are adopted. The weight decay of the model is 0.05. $\\lambda$ and $\\gamma$ of GFA and PC are set to 0.2 and 2 by default. The segmentation loss consists of dice loss and cross entropy loss. The classification loss is cross entropy loss. For the weights of the loss function, we set 5 and 2 for segmentation loss and classification loss, respectively. Other hyperparameters are the same as Mask2Former [9].", + "bbox": [ + 511, + 593, + 906, + 820 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Main Results", + "text_level": 1, + "bbox": [ + 511, + 830, + 653, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our model with existing state-of-the-art approaches in Table 2. It can be seen that with proxy calibration expanding training semantics and the gradient-free", + "bbox": [ + 511, + 854, + 905, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22669", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/13244781382e9943c88b03c348c971a72a78dcb831f904419d2747b33e447a95.jpg", + "table_caption": [ + "Table 3. Ablation experiments about the proposed methods. GFA and PC indicate the Gradient-Free Aggregation and Proxy Calibration strategies, respectively." + ], + "table_footnote": [], + "table_body": "
MethodADE-150PC-459OpenBench
Baseline33.114.342.3
+ GFA34.7↑ (1.6)16.0↑ (1.7)43.7↑ (1.4)
+ PC33.9↑ (0.8)17.2↑ (2.9)44.3↑ (2.0)
+ Both35.8↑ (2.7)19.1↑ (4.8)44.9↑ (2.6)
", + "bbox": [ + 89, + 133, + 480, + 209 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "aggregation algorithm modeling the joint space, our model achieves the state-of-the-art performance on most popular benchmarks. The last column counts the average performance across multiple datasets, which shows that our method performs well for both in-vocabulary as well as out-vocabulary diverse scenarios. The relatively inferior performance of our method on VOC is due to the fact that the semantics of this dataset are almost contained in the training set, which tends to benefit models such as SCAN [37] that finetunes the CLIP backbone on training set. Besides, the greater divergence of OpenBench from training semantics does not imply that it is totally more challenging than other benchmarks, because the final performance is influenced by various factors such as image content complexity, image clarity, and the number of given categories. More discussion can be seen in the supplementary materials.", + "bbox": [ + 88, + 214, + 486, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 467, + 243, + 483 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Component Analysis Table 3 shows the ablations about the proposed Gradient-Free Aggregation (GFA) and Proxy Calibration (PC) strategy. We evaluate their effectiveness on both existing benchmarks and our proposed OpenBench. It can be seen that with GFA fusing CLIP feature and query embeddings, the model's performance is greatly improved on both seen and unseen semantics. Besides, PC can facilitate the model's ability to understand various scenarios by cost-free expansion of the training semantic space. When applying both, the model achieves the best performance.", + "bbox": [ + 89, + 489, + 482, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Embedding Distribution Analysis To demonstrate that our proposed OpenBench is more open-vocabulary compared to the existing test sets, we perform t-SNE visualization of the category embeddings of the training set, the existing test sets, and our OpenBench in Figure 4. Besides, we also visualize the proxy embeddings generated by PC. We can see that the existing test sets (green dots) are overall similar to the training semantics (blue dots). Our OpenBench (orange dots), on the other hand, shows significant differences. In addition, the proxy embeddings (red dots) generated by PC expand the training set and thus help to improve the robustness and generalization of the model.", + "bbox": [ + 89, + 648, + 482, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis about Inference Class Number In this part we explore an interesting phenomenon: why the OVS models perform much worse on ADE-847 and PC-459 than on other test sets? Intuitively, this is because these datasets", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ae6d8df7e54015ea14855309442f3fc89d438b782f13954996469bbf9ad68bb1.jpg", + "image_caption": [ + "Figure 4. Visualization of category embeddings across the training space, existing evaluation space, our OpenBench space, and the generated proxy space." + ], + "image_footnote": [], + "bbox": [ + 535, + 89, + 882, + 276 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2f9ffae676bc6c1a4e9b664692e25a8682fe827cae87d8c2d18c096cd7d52636.jpg", + "image_caption": [ + "Figure 5. Impact of the number of inference categories on model performance. As the number of irrelevant categories increases, the model's performance on the same image degrades." + ], + "image_footnote": [], + "bbox": [ + 532, + 330, + 883, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "contain more novel categories. However, we find that another factor significantly impacts model performance, that is the number of categories fed into the model during inference. Let's start with the conclusion: for the same model and test image, the model performance is worse when there are more inference categories given. In other words, the model may be able to make correct predictions when only a small number of categories are given. However, when more categories are prompted, the model's output will be incorrect. Here we take two typical methods, SAN [61] and MAFT-Plus [26], as examples and show how their performance on the ADE-150 dataset changes as the number of categories increases. As shown in Figure 5, with the number of given irrelevant categories increases, the models of different technology routes show a consistent decline.", + "bbox": [ + 511, + 561, + 906, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Internal Analysis of Proposed Strategies In Table 4 we conduct more experiments about the proposed methods. Specifically, we first compare our gradient-free aggregation designs with the typical learning-based feature fusion approaches in (a). Results show that vanilla attention performs inferior than our GFA, especially on the more open-vocabulary set OpenBench (degraded by more than", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22670", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eac635c0d5e9fcf77f4ef38cc457d980ef13348b47609cec0fab18099c59cdc6.jpg", + "image_caption": [ + "(a) Seen: building treefence bicycle" + ], + "image_footnote": [], + "bbox": [ + 99, + 89, + 297, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6c46b639d0000c7d046b194c73a0b7df5d90e39b16b95a00b9df546125c05e8e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 90, + 496, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8ebda229eee540b5321463bc44485c601e2283e02a91ea233c57af6c629916de.jpg", + "image_caption": [ + "Unseen: sidewalk streetlight" + ], + "image_footnote": [], + "bbox": [ + 99, + 198, + 297, + 295 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e712fefb1e0a5711f184bc4a71101150a4f76838d596280e8c0aa697dcc1b676.jpg", + "image_caption": [ + "(c) Seen: ceiling desk chair" + ], + "image_footnote": [], + "bbox": [ + 297, + 198, + 496, + 295 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/217dfb670f55b5179e12cbf69a31b51589f5cf93124a108cb2ddf12dd600fe0f.jpg", + "image_caption": [ + "Unseen: computer chandelier", + "→ opera house" + ], + "image_footnote": [], + "bbox": [ + 99, + 310, + 297, + 405 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/16aa424d28b7f6a7aae259f3fb3bc024884a8c8eeae3cbf2fa86affbf7b304fa.jpg", + "image_caption": [ + "(e) building" + ], + "image_footnote": [], + "bbox": [ + 297, + 310, + 496, + 405 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3ee052e628fdbc2b67d490ead5016756613b7ad75c6edcba61f0a542d94a3e42.jpg", + "image_caption": [ + "(b) Seen: sea boat mountain" + ], + "image_footnote": [], + "bbox": [ + 501, + 90, + 697, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a99f3606ca78fda54e006e50ca67b981ed5b989ea0fedd8d3833088e8486c955.jpg", + "image_caption": [ + "Unseen: pier pole" + ], + "image_footnote": [], + "bbox": [ + 699, + 90, + 898, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9fbc96ee036cac8c76f0d7f9c34d4a1bbe6eef3a0c55f03c5b3e6070cf778e20.jpg", + "image_caption": [ + "(d) Seen: plant rock person" + ], + "image_footnote": [], + "bbox": [ + 501, + 198, + 697, + 295 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fa1a4772f0f95bc1b6a5bbc312259b27de77c8c33cc45d533b8dd2bc5cd558f7.jpg", + "image_caption": [ + "Unseen: waterfall" + ], + "image_footnote": [], + "bbox": [ + 699, + 198, + 898, + 295 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0c4788019177927f18722285b24801a6d2619541eebe6fe75b0369a5694e23a3.jpg", + "image_caption": [ + "(f)", + "Figure 6. Visualization of segmentation results. (a) - (d) demonstrate the successful segmentation cases of our method for seen and unseen categories. (e) and (f) display the adaptability for flexible text query. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 501, + 310, + 699, + 405 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/08cc198f1b540bd4c57fca47bdb0e53b7258f6b724f49f6953e5716f86ab39a2.jpg", + "image_caption": [ + "→ ferrari" + ], + "image_footnote": [], + "bbox": [ + 699, + 310, + 898, + 405 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c9fc5f2aec0f5413300f658a15fb4a387d8c40a7572e79315b3b16958a68cc1f.jpg", + "table_caption": [ + "Table 4. Ablation experiments within each modules. (a) shows the comparison of our gradient-free algorithm with learning-based method, e.g., cross attention. (b) demonstrates the effect of $\\gamma$ of Beta distribution in the proxy learning strategy." + ], + "table_footnote": [], + "table_body": "
ADE-150PC-459OpenBench
(a) Comparisons of GFA and learning-based strategy
Self Attn34.014.740.6
Cross Attn34.214.841.2
GFA34.716.043.7
(b) Effect of sampling distribution of PC
γ = 133.114.342.3
γ = 0.533.6↑ (0.5)16.1↑ (1.8)43.7↑ (1.4)
γ = 233.9↑ (0.8)17.2↑ (2.9)44.3↑ (2.0)
", + "bbox": [ + 94, + 521, + 470, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "2.5 mIoU). This is because learning-based methods tend to develop a reliance on the segmentation query embeddings during training, while overlooking the CLIP features that inherently exhibit stronger generalization capabilities, thereby introducing biases.", + "bbox": [ + 89, + 654, + 482, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also experiment with the influence of different $\\gamma$ in Beta distribution to the performance of proxy calibration. The results are shown in Table 4 (b), where $\\gamma = 1$ is equivalent to do not utilize PC. We find that different $\\gamma$ benefit the model performance while the performance is better when the intermediate probability density is high ( $\\gamma = 2$ ).", + "bbox": [ + 89, + 729, + 482, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Visualization", + "text_level": 1, + "bbox": [ + 89, + 832, + 227, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 6 shows some segmentation cases of our OVSNet. It can be seen that our method achieves excellent segmentation performance on various scenarios. Specifically, (a) - (d)", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "demonstrate the excellent segmentation of our method for seen and unseen categories. (e) and (f) display the adaptability for flexible text query, e.g., change \"building\" to \"opera house\" and \"car\" to \"ferrari\".", + "bbox": [ + 511, + 463, + 906, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 540, + 633, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we identify that the existing open-vocabulary segmentation evaluation sets suffer from excessive semantic similarity to the training set, which limits their ability to comprehensively assess the model's open vocabulary understanding capabilities. To address this, we propose a new benchmark, OpenBench, which features significant differences from the training set and includes more fine-grained categories. Comparison on OpenBench further demonstrates that maintaining the generalization space of CLIP is crucial for OVS. Additionally, we introduce OVSNet, a method that leverages proxy calibration for cost-free augmentation of the training space and gradient-free aggregation for fusing heterogeneous features. Experiments show that the proposed OVSNet achieves excellent performance on both existing datasets and the newly proposed OpenBench. We hope that our benchmark and model design could inspire further researches in this area.", + "bbox": [ + 511, + 566, + 906, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by Shenzhen Science and Technology Program under Grant CJGJZD20220517142402006 and Guangdong Natural Science Funds for Distinguished Young Scholar (No. 2025B1515020012).", + "bbox": [ + 511, + 825, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22671", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sule Bai, Yong Liu, Yifei Han, Haoji Zhang, and Yansong Tang. Self-calibrated clip for training-free open-vocabulary segmentation. arXiv preprint arXiv:2411.15869, 2024. 2", + "[2] Sule Bai, Mingxing Li, Yong Liu, Jing Tang, Haoji Zhang, Lei Sun, Xiangxiang Chu, and Yansong Tang. Univg-r1: Reasoning guided universal visual grounding with reinforcement learning. arXiv preprint arXiv:2505.14231, 2025. 1", + "[3] Gabriel J. Brostow, Jamie Shotton, Julien Fauqueur, and Roberto Cipolla. Segmentation and recognition using structure from motion point clouds. In ECCV, pages 44-57, 2008. 4", + "[4] Maxime Bucher, Tuan-Hung Vu, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. In NeurlPS, 2019. 2", + "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 5", + "[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. TPAMI, 2018. 1", + "[7] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. UNITER: learning universal image-text representations. arXiv preprint arXiv:1909.11740, 2019. 3", + "[8] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G. Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021. 6", + "[9] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 6", + "[10] Seokju Cho, Heeseong Shin, Sunghwan Hong, Anurag Arnab, Paul Hongsuck Seo, and Seungryong Kim. Cat-seg: Cost aggregation for open-vocabulary semantic segmentation. In CVPR, pages 4113-4123, 2024. 1, 2, 3, 6", + "[11] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 3", + "[12] Jian Ding, Nan Xue, Gui-Song Xia, and Dengxin Dai. Decoupling zero-shot semantic segmentation. In CVPR, 2022. 1, 2, 3, 6", + "[13] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes challenge: A retrospective. IJCV, 111:98-136, 2015. 2, 3, 6", + "[14] Shanghua Gao, Zhong-Yu Li, Ming-Hsuan Yang, Ming-Ming Cheng, Junwei Han, and Philip Torr. Large-scale unsupervised semantic segmentation. TPAMI, 45(6):7457-7476, 2022. 4", + "[15] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Open-vocabulary image segmentation. arXiv preprint arXiv:2112.12143, 2021. 1, 2, 6" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Leo Grady. Random walks for image segmentation. IEEE TPAMI, 28(11):1768-1783, 2006. 2, 4", + "[17] Kunyang Han, Yong Liu, Jun Hao Liew, Henghui Ding, Jiajun Liu, Yitong Wang, Yansong Tang, Yujiu Yang, Jiashi Feng, Yao Zhao, et al. Global knowledge calibration for fast open-vocabulary segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 797-807, 2023. 1, 2", + "[18] Shuting He, Henghui Ding, and Wei Jiang. Primitive generation and semantic-related alignment for universal zero-shot segmentation. In CVPR, 2023. 1", + "[19] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1", + "[20] Jiaqi Huang, Zunnan Xu, Ting Liu, Yong Liu, Haonan Han, Kehong Yuan, and Xiu Li. Densely connected parameter-efficient tuning for referring image segmentation. arXiv preprint arXiv:2501.08580, 2025. 1", + "[21] Xiaoke Huang, Jianfeng Wang, Yansong Tang, Zheng Zhang, Han Hu, Jiwen Lu, Lijuan Wang, and Zicheng Liu. Segment and caption anything. arXiv preprint arXiv:2312.00869, 2023. 1", + "[22] Joonhyun Jeong, Geondo Park, Jayeon Yoo, Hyungsik Jung, and Heesu Kim. Proxydet: Synthesizing proxy novel classes via classwise mixup for open-vocabulary object detection. In AAAI, pages 2462-2470, 2024. 3", + "[23] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 3", + "[24] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916, 2021. 1", + "[25] Siyu Jiao, Yunchao Wei, Yaowei Wang, Yao Zhao, and Humphrey Shi. Learning mask-aware clip representations for zero-shot segmentation. arXiv preprint arXiv:2310.00240, 2023. 1, 6", + "[26] Siyu Jiao, Hongguang Zhu, Jiannan Huang, Yao Zhao, Yunchao Wei, and Humphrey Shi. Collaborative vision-text representation optimizing for open-vocabulary segmentation. In ECCV, pages 399-416, 2025. 6, 7", + "[27] Boyi Li, Kilian Q. Weinberger, Serge J. Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. In ICLR, 2022. 1, 3", + "[28] Gen Li, Nan Duan, Yuejian Fang, Ming Gong, and Daxin Jiang. Unicoder-vl: A universal encoder for vision and language by cross-modal pre-training. In AAAI, 2020. 3", + "[29] Wanhua Li, Xiaoke Huang, Zheng Zhu, Yansong Tang, Xiu Li, Jie Zhou, and Jiwen Lu. Ordinalclip: Learning rank prompts for language-guided ordinal regression. NeurIPS, pages 35313-35325, 2022. 3", + "[30] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu" + ], + "bbox": [ + 516, + 92, + 905, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22672", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-semantics aligned pre-training for vision-language tasks. In ECCV, 2020. 3", + "[31] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yi-nan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted CLIP. arXiv preprint arXiv:2210.04150, 2022. 1, 2, 6", + "[32] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In ECCV, 2014. 4", + "[33] Yong Liu, Ran Yu, Xinyuan Zhao, and Yujiu Yang. Quality-aware and selective prior enhancement memory network for video object segmentation. In CVPR Workshop, 2021. 1", + "[34] Yong Liu, Ran Yu, Jiahao Wang, Xinyuan Zhao, Yitong Wang, Yansong Tang, and Yujiu Yang. Global spectral filter memory network for video object segmentation. In ECCV, pages 648-665, 2022.", + "[35] Yong Liu, Ran Yu, Fei Yin, Xinyuan Zhao, Wei Zhao, Weihao Xia, and Yujiu Yang. Learning quality-aware dynamic memory for video object segmentation. In ECCV, pages 468-486, 2022.", + "[36] Yong Liu, Cairong Zhang, Yitong Wang, Jiahao Wang, Yujiu Yang, and Yansong Tang. Universal segmentation at arbitrary granularity with language instruction. arXiv preprint arXiv:2312.01623, 2023. 1", + "[37] Yong Liu, Sule Bai, Guanbin Li, Yitong Wang, and Yansong Tang. Open-vocabulary segmentation with semantic-assisted calibration. In CVPR, pages 3491-3500, 2024. 1, 4, 6, 7", + "[38] Yong Liu, Ran Yu, Fei Yin, Xinyuan Zhao, Wei Zhao, Weihao Xia, Jiahao Wang, Yitong Wang, Yansong Tang, and Yujiu Yang. Learning high-quality dynamic memory for video object segmentation. IEEE TPAMI, 2025. 1", + "[39] Guanxing Lu, Ziwei Wang, Changliu Liu, Jiwen Lu, and Yansong Tang. Thinkbot: Embodied instruction following with thought chain reasoning. arXiv preprint arXiv:2312.07062, 2023. 1", + "[40] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In NeurIPS, 2019. 3", + "[41] Zhuoyan Luo, Yicheng Xiao, Yong Liu, Shuyan Li, Yitong Wang, Yansong Tang, Xiu Li, and Yujiu Yang. Soc: Semantic-assisted object cluster for referring video object segmentation. arXiv preprint arXiv:2305.17011, 2023. 2", + "[42] Zhuoyan Luo, Yinghao Wu, Yong Liu, Yicheng Xiao, Xiaoping Zhang, and Yujiu Yang. Hdc: Hierarchical semantic decoding with counting assistance for generalized referring expression segmentation. arXiv preprint arXiv:2405.15658, 2024. 1", + "[43] Carl D Meyer and Ian Stewart. Matrix analysis and applied linear algebra. SIAM, 2023. 5", + "[44] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan L. Yuille. The role of context for object detection and semantic segmentation in the wild. In CVPR, 2014. 2, 3, 4, 6" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Mengxue Qu, Yu Wu, Wu Liu, Qiqi Gong, Xiaodan Liang, Olga Russakovsky, Yao Zhao, and Yunchao Wei. Siri: A simple selective retraining mechanism for transformer-based visual grounding. In ECCV, 2022. 3", + "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 4, 5, 6", + "[47] Yiqin Wang, Haoji Zhang, Yansong Tang, Yong Liu, Jiashi Feng, Jifeng Dai, and Xiaojie Jin. Hierarchical memory for long video qa. arXiv preprint arXiv:2407.00603, 2024. 3", + "[48] Yulin Wang, Haoji Zhang, Yang Yue, Shiji Song, Chao Deng, Junlan Feng, and Gao Huang. Uni-adafocus: Spatial-temporal dynamic computation for video recognition. IEEE TPAMI, 2024. 3", + "[49] Yuji Wang, Jingchen Ni, Yong Liu, Chun Yuan, and Yansong Tang. Iterprime: Zero-shot referring image segmentation with iterative grad-cam refinement and primary word emphasis. In AAAI, pages 8159-8168, 2025. 1", + "[50] Yuji Wang, Haoran Xu, Yong Liu, Jiaze Li, and Yansong Tang. Sam2-love: Segment anything model 2 in language-aided audio-visual scenes. In CVPR, pages 28932-28941, 2025.", + "[51] Cong Wei, Yujie Zhong, Haoxian Tan, Yong Liu, Zheng Zhao, Jie Hu, and Yujiu Yang. Hyperseg: Towards universal visual segmentation with large language model. arXiv preprint arXiv:2411.17606, 2024.", + "[52] Cong Wei, Yujie Zhong, Haoxian Tan, Yingsen Zeng, Yong Liu, Zheng Zhao, and Yujiu Yang. Instructseg: Unifying instructed visual segmentation with multi-modal large language models. arXiv preprint arXiv:2412.14006, 2024. 1", + "[53] Xiongwei Wu, Xin Fu, Ying Liu, Ee-Peng Lim, Steven CH Hoi, and Qianru Sun. A large-scale benchmark for food image segmentation. In ACM MM, 2021. 4", + "[54] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.6", + "[55] Yongqin Xian, Subhabrata Choudhury, Yang He, Bernt Schiele, and Zeynep Akata. Semantic projection network for zero- and few-label semantic segmentation. In CVPR, 2019. 1, 2", + "[56] Yicheng Xiao, Zhuoyan Luo, Yong Liu, Yue Ma, Hengwei Bian, Yatai Ji, Yujiu Yang, and Xiu Li. Bridging the gap: A unified video comprehension framework for moment retrieval and highlight detection. In CVPR, pages 18709-18719, 2024. 4", + "[57] Bin Xie, Jiale Cao, Jin Xie, Fahad Shahbaz Khan, and Yanwei Pang. Sed: A simple encoder-decoder for open-vocabulary semantic segmentation. In CVPR, pages 3426-3436, 2024. 1, 2, 3, 6", + "[58] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In NIPS, 2021. 1" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22673", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[59] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, pages 2955-2966, 2023. 2", + "[60] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A simple baseline for zero-shot semantic segmentation with pre-trained vision-language model. arXiv preprint arXiv:2112.14757, 2021. 1, 2, 3, 6", + "[61] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xi-ang Bai. Side adapter network for open-vocabulary semantic segmentation. In CVPR, pages 2945–2954, 2023. 1, 2, 6, 7", + "[62] Xin Xu, Tianyi Xiong, Zheng Ding, and Zhuowen Tu. Masqclip for open-vocabulary universal image segmentation. In ICCV, pages 887-898, 2023. 1", + "[63] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. NeurIPS, pages 32215-32234, 2023. 1, 2, 6", + "[64] Hui Zhang and Henghui Ding. Prototypical matching and open set rejection for zero-shot semantic segmentation. In ICCV, 2021. 1", + "[65] Haoji Zhang, Yiqin Wang, Yansong Tang, Yong Liu, Jiashi Feng, Jifeng Dai, and Xiaojie Jin. Flash-vstream: Memory-based real-time understanding for long video streams. arXiv preprint arXiv:2406.08085, 2024. 3", + "[66] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ADE20K dataset. In CVPR, 2017. 2, 3, 4, 6" + ], + "bbox": [ + 91, + 92, + 480, + 487 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22674", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Stepping Out of Similar Semantic Space for Open-Vocabulary Segmentation/f495e889-2c8e-4e64-ba9d-86b1aec5716d_model.json b/2025/Stepping Out of Similar Semantic Space for Open-Vocabulary Segmentation/f495e889-2c8e-4e64-ba9d-86b1aec5716d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e61fb57357e1af0d2e0bd8233c7df174e9fd7c69 --- /dev/null +++ b/2025/Stepping Out of Similar Semantic Space for Open-Vocabulary Segmentation/f495e889-2c8e-4e64-ba9d-86b1aec5716d_model.json @@ -0,0 +1,2400 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.131, + 0.884, + 0.153 + ], + "angle": 0, + "content": "Stepping Out of Similar Semantic Space for Open-Vocabulary Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.143, + 0.179, + 0.862, + 0.235 + ], + "angle": 0, + "content": "Yong Liu\\(^{1,2*}\\), Song-Li Wu\\(^{1,2*}\\), Sule Bai\\(^{1,2*}\\), Jiahao Wang\\(^{3}\\), Yitong Wang\\(^{4}\\), Yansong Tang\\(^{1,2\\dagger}\\) \n\\(^{1}\\)Tsinghua Shenzhen International Graduate School \\(^{2}\\)Tsinghua University \n\\(^{3}\\)The University of Hong Kong \\(^{4}\\)ByteDance Inc." + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.237, + 0.78, + 0.252 + ], + "angle": 0, + "content": "liuyong23@mails.tsinghua.edu.cn, tang.yansong@sz.tsinghua.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.302, + 0.327, + 0.318 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.335, + 0.485, + 0.699 + ], + "angle": 0, + "content": "Open-vocabulary segmentation aims to achieve segmentation of arbitrary categories given unlimited text inputs as guidance. To achieve this, recent works have focused on developing various technical routes to exploit the potential of large-scale pre-trained vision-language models and have made significant progress on existing benchmarks. However, we find that existing test sets are limited in measuring the models' comprehension of \"open-vocabulary\" concepts, as their semantic space closely resembles the training space, even with many overlapping categories. To this end, we present a new benchmark named OpenBench that differs significantly from the training semantics. It is designed to better assess the model's ability to understand and segment a wide range of real-world concepts. When testing existing methods on OpenBench, we find that their performance diverges from the conclusions drawn on existing test sets. In addition, we propose a method named OVSNet to improve the segmentation performance for diverse and open scenarios. Through elaborate fusion of heterogeneous features and cost-free expansion of the training space, OVSNet achieves state-of-the-art results on both existing datasets and our proposed OpenBench. Corresponding analysis demonstrates the rationality and effectiveness of our proposed benchmark and method." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.716, + 0.222, + 0.731 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.742, + 0.483, + 0.864 + ], + "angle": 0, + "content": "Semantic segmentation is one of the most fundamental tasks in computer vision, which targets at assigning semantic category to pixels in an image. Despite achieving excellent progress in recent years [2, 6, 20, 21, 33-36, 38, 42, 49-52, 58], traditional semantic segmentation approaches rely on predefined categories sets and tend to falter when encountering categories absent during the training phase, significantly impeding their real-world application. Such chal" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.286, + 0.73, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.624, + 0.421, + 0.641, + 0.432 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.287, + 0.905, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.789, + 0.42, + 0.807, + 0.432 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.434, + 0.907, + 0.489 + ], + "angle": 0, + "content": "Figure 1. (a) shows the similarity distribution of the existing test set (blue) and our OpenBench (orange) with the training semantic space. (b) illustrates the performance gap between the methods finetuning [10] and frozen [61] CLIP on different benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.505, + 0.906, + 0.596 + ], + "angle": 0, + "content": "lenge has inspired the exploration of Open-Vocabulary Segmentation (OVS) task [12, 15, 17, 18, 27, 37, 55, 64]. Different from traditional closed-set segmentation, OVS methods aim to segment arbitrary categories under the guidance of given text inputs, which has many potential applications such as auto-driving and human-robot interaction [19, 39]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.903 + ], + "angle": 0, + "content": "However, it is challenging to identify unseen categories without the intervention of external knowledge. Therefore, an intuitive idea is to introduce large-scale vision-language model [24, 46] trained with numerous sources to extend the semantic space of segmentation models. Following this idea, existing OVS methods [10, 31, 37, 57, 61, 62] concentrate on exploring how to exploit the pretrained knowledge prior more effectively and derive several technical routes. Early methods [12, 31, 60] adopt a two-stage pipeline. This approach first generates class-agnostic mask proposals with segmentation models, following which the pre-trained CLIP [46] serves as an additional classifier to execute region-level open-vocabulary classification. However, this strategy incurs heavy computation overhead, as it requires repeated forward passes of the CLIP vision encoder for each mask. This can be prohibitively expensive for real-world applications. To address this problem, researchers [25, 61–63] propose to unify classification and segmentation within a shared space, i.e., integrating frozen CLIP features into the segmenter and extracting CLIP embeddings from potential" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.217, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.234, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.234, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22664" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.302 + ], + "angle": 0, + "content": "target regions to enable simultaneous classification and segmentation. These methods achieve remarkable performance in a more efficient one-stage pipeline. Subsequently, some works [10, 57] discover that finetuning the CLIP encoder could further enhance the model's performance on existing test sets. They develop an early fusion paradigm based on cost map and achieve outstanding segmentation results. However, this performance boost is somewhat counterintuitive, as fine-tuning CLIP typically adapts it to specific training semantic spaces, which could reduce its original generalization ability. Yet, the ability of generalized visual-linguistic alignment is precisely what the open-vocabulary task requires. So why does fine-tuning CLIP appear to benefit the performance of OVS?" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.306, + 0.485, + 0.547 + ], + "angle": 0, + "content": "This question drives us to revisit the existing evaluation benchmarks of the cross-dataset evaluation paradigm. We count the maximum similarity of each category in test set to the training semantics. Figure 1 (a) illustrates the similarity distribution of all test categories to the training space (blue color). Results indicate that the existing test data is, in fact, highly similar to the training semantic space. Specifically, even the most challenging ADE-847 [66] and PC-459 [44] datasets have an average similarity to the training classes of 0.79 and 0.83. PC-59 [44] and VOC [13] datasets even have a similarity to the training semantics of 0.95 and 0.97. This explains why finetuning CLIP on training set can lead to further performance gains on existing benchmarks. However, it somewhat conflicts with the purpose of the open-vocabulary setting, which aims to achieve segmentation of unlimited categories in the real world." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.483, + 0.778 + ], + "angle": 0, + "content": "To address the above problem and provide a more comprehensive measure of OVS models, we propose a new evaluation benchmark named OpenBench, which is characterized by significant semantic differences from the training data. Figure 1 (a) also demonstrates the similarity distribution of our OpenBench (orange color). It can be seen that our benchmark contains more novel classes compared to the existing test sets. We evaluate the performance of approaches across different technical routes on our benchmark, and the findings diverge from those obtained on existing test sets. As shown in Figure 1 (b), while finetuning CLIP yields steady performance gains on existing datasets, it shows a significant drop on our benchmark. This outcome also supports our motivation for proposing a benchmark that diverges substantially from the training space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In addition, we also present a framework named OVSNet to improve the segmentation performance for diverse and open scenarios. In OVSNet, we adopt a Proxy Calibration (PC) training strategy to enhance the robustness of model representations by broadening the training space with no additional cost. This strategy leverages the synthesis of proxy embeddings, which approximate unseen semantics through convex combinations of in-vocabulary classes. Addition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.258 + ], + "angle": 0, + "content": "ally, inspired by random walk [16], we propose a gradient-free fusion of CLIP embeddings and segmentation decoder features to model a robust joint space and mitigate potential overfitting issues related to training semantics. With the above proposed designs calibrating model knowledge from the aspects of training space and features separately, our OVSNet demonstrates strong performance across a wide range of scenarios, both on existing test sets similar to the training semantics and on our proposed novel OpenBench. Related ablations also demonstrate the effectiveness and soundness of our proposed strategies." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.258, + 0.858, + 0.272 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.273, + 0.905, + 0.363 + ], + "angle": 0, + "content": "- We present a new benchmark named OpenBench for OVS, which is characterized by significant semantic differences from the training space. While using it to evaluate the comprehension of OVS methods in more open scenarios, we reach a different conclusion from that of the existing test sets." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.364, + 0.905, + 0.453 + ], + "angle": 0, + "content": "- We propose OVSNet to improve the segmentation performance on diverse and open scenarios. It adopts a Proxy Calibration training strategy to broaden training space without additional cost. Besides, a gradient-free algorithm is leveraged to fuse CLIP embeddings and segmentation features for modeling robust joint space." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.455, + 0.905, + 0.515 + ], + "angle": 0, + "content": "- Our method achieves state-of-the-art performance on existing benchmarks and the proposed OpenBench. Extensive experiments are conducted to prove the effectiveness and soundness of the proposed algorithms." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.273, + 0.905, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.528, + 0.655, + 0.544 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Open-Vocabulary Segmentation. This task aims to segment an image and identify regions with arbitrary given text queries [1, 4, 15, 41, 55]. Pioneering work [55] replaces the output convolution layer by computing the similarity between visual features and linguistic embeddings, which has become common practice. Later, some methods [12, 15, 31, 59, 60] propose a two-stage pipeline: the model first generates class-agnostic mask proposals, then a pretrained CLIP [46] is utilized to perform sub-image classification by cropping and masking corresponding regions. With the combination of both in-vocabulary and out-vocabulary classification, these methods obtain great performance. However, this strategy incurs heavy computation overhead, as it requires repeated forward passes of the CLIP vision encoder for each mask. This can be prohibitively expensive for real-world applications. To address this problem, subsequent researchers propose to unify classification and segmentation within a shared space. Among them, SAN [61] designs a side-adapter network to leverage CLIP features for reorganizing segmentation and classification. GKC [17] presents text-guided knowledge distillation strategy to transfer CLIP knowledge to specific classification layer. FCCLIP [63] proposes to leverage" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22665" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.9, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.299, + 0.907, + 0.342 + ], + "angle": 0, + "content": "Figure 2. Illustration of the samples in the proposed OpenBench. Compared to existing test sets, our OpenBench has semantic categories that differ more from the training space, which helps to provide a more comprehensive measure of the model's ability to perform open-vocabulary segmentation. The corresponding category name is labeled in the lower right corner of each image." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.355, + 0.907, + 0.384 + ], + "angle": 0, + "content": "Table 1. Comparison of the statistics of our proposed OpenBench with existing datasets. 'Cls', 'Img', and 'Sim' indicate class, image, and similarity, respectively. ADE, PC, and VOC denote ADE20K [66], Pascal Context [44], and Pascal VOC [13]" + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.387, + 0.9, + 0.525 + ], + "angle": 0, + "content": "
DatasetCls Num.Img Num.Mean Sim.Median Sim.Min Sim.Max Sim.Fine GranularitySemantic Duplication
VOC [13]2014490.97561.00000.84491.0000××
PC-59 [44]5951050.94811.00000.76581.0000××
ADE-150 [66]15020000.80860.82280.19601.0000×✓
Cityscapes [11]331000.84880.83700.67461.0000××
PC-459 [44]45951050.83740.81240.67181.0000✓✓
ADE-847 [66]84720000.79260.79130.19601.0000✓✓
OpenBench (Ours)28660560.61420.64520.26080.7947✓×
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.599 + ], + "angle": 0, + "content": "the local-awareness advantage of convolution backbones. Some works [10, 57] then discover that finetuning CLIP encoder could further enhance the model's performance on existing test sets and develop effective one-stage framework." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.601, + 0.483, + 0.722 + ], + "angle": 0, + "content": "However, we find that existing test sets are highly similar to the training set, which limits their ability to effectively evaluate the model's comprehension of the open vocabulary semantics. Therefore, in this paper we propose a new benchmark OpenBench that is characterized by significant semantic differences from the training data. We also propose a method named OVSNet to improve the segmentation performance for diverse and open scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Vision-Language Pre-training. Vision-language pretraining aims to learn a unified visual-linguistic representation space. Early approaches [7, 28, 30, 40, 45, 47, 48, 65], constrained by small-scale datasets, struggle to achieve strong performance and require fine-tuning for downstream tasks. However, with the advent of large-scale web data, recent works [23, 46] have demonstrated the benefits of leveraging such data to train a more robust multi-modal representation space. Among these, CLIP [46], the most widely used vision-language model, employs contrastive learning to link images with their corresponding captions," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.6 + ], + "angle": 0, + "content": "achieving impressive cross-modal alignment. Building upon previous works [12, 22, 27, 29, 60], we also utilize the well-aligned and generalized space of CLIP to enhance open-vocabulary segmentation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.613, + 0.637, + 0.628 + ], + "angle": 0, + "content": "3. Benchmark" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.638, + 0.71, + 0.652 + ], + "angle": 0, + "content": "3.1. Dataset Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.826 + ], + "angle": 0, + "content": "In Table 1, we present the statistics comparison of our proposed OpenBench with existing popular open-vocabulary segmentation datasets, including ADE-150 [66], ADE-847 [66], PC-59 [44], PC-459 [44], VOC [13], and Cityscapes [11]. Specifically, our OpenBench contains 286 categories, of which the maximum similarity to the training categories is 0.7947, the minimum is 0.2608, and the average similarity is 0.6142. It can be observed that, compared to existing datasets, our OpenBench exhibits a greater disparity from the training set, which better evaluates the model's segmentation ability for arbitrary open categories." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Besides, our OpenBench contains fine-grained categories, e.g., \"oystercatcher\", rather than just generic broad category such as \"bird\" (corresponds to the \"Fine Granularity\" characterization in Table 1). This is important for evaluating whether the OVS models preserve the general" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22666" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.482, + 0.407 + ], + "angle": 0, + "content": "ization capability of CLIP [46], which is inherently capable of detecting these fine-grained categories. Actually, PC-459 [44] and ADE-847 [66] also have fine-grained categories, but they suffer from semantic duplication problem. In detail, semantic duplication means that there will be both fine-grained categories as well as corresponding broad categories in these datasets, which would cause bias in the metrics calculation. For instance, if a model assigns the \"door\" tag to a region whose ground truth label is \"double door\", it will be considered incorrect. For open-vocabulary segmentation setting, we argue that the responsibility of the model is to discern the correct semantic. The distribution of such duplication categories and the corresponding performance statistics is unreasonable. SCAN [37] also points similar observations and try to solve this by designing new metric. However, this requires manual judgment of the relationships between categories, introducing additional cost and uncertainty. Our OpenBench, on the other hand, has fine-grained semantics without the problem of semantic duplication, which helps to more accurately reflect the real capabilities of related models." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.411, + 0.482, + 0.65 + ], + "angle": 0, + "content": "Figure 2 illustrates part of the categories contained in OpenBench, where some categories such as food related even differ significantly from the visual domain of the training set. Furthermore, our OpenBench also accounts for \"other\" or \"background\" pixels in images, which are often overlooked by existing datasets. This is important for open-vocabulary segmentation because the model is essentially forced to choose the most appropriate category from a predefined set. If there is no correct class and no \"others\" option in the candidate set, the model would be compelled to make an incorrect prediction. Moreover, \"others\" or \"background\" require the model to consider not only the direct visual-linguistic correspondence between the image content and each class, but also the relative relationships among different classes. This adds complexity and makes the task more relevant to real-world scenarios." + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.667, + 0.265, + 0.68 + ], + "angle": 0, + "content": "3.2. Dataset Collection" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.69, + 0.482, + 0.9 + ], + "angle": 0, + "content": "Our OpenBench is primarily derived from existing accessible segmentation datasets. To obtain semantic categories that differ significantly from the training set, i.e., COCO [32], we first organize a group of segmentation datasets that possess category annotations, e.g., Food103 [53], Imagenet-S [14], and CamVid [3]. For all categories in the collected data, we leverage CLIP-L/14 [46] to compute their maximum similarity with COCO categories. Then we calculate the minimum similarity among the categories present in each image as the image's similarity to the training semantics. If the image similarity is greater than the set threshold \\(\\sigma_{1}\\), the image will be filtered. Conversely, if the similarity of the image is less than the threshold \\(\\sigma_{1}\\), we would set the category exist in the image" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.093, + 0.905, + 0.151 + ], + "angle": 0, + "content": "whose similarity is greater than the threshold \\(\\sigma_{2}\\) to \"others\" and save the related images and annotation. Finally, after manually filtering the categories for potential conflicts or duplicates, we get the OpenBench." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.169, + 0.603, + 0.183 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.195, + 0.622, + 0.208 + ], + "angle": 0, + "content": "4.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.217, + 0.905, + 0.457 + ], + "angle": 0, + "content": "The overall pipeline of our proposed OVSNet is shown in Figure 3. We leverage the vision and text encoder of pretraind CLIP [46] to encode the input image and category names, respectively. Then, a segmentation decoder is trained to decode the visual features, generating masks for potential targets and their corresponding query embeddings. We then leverage the predicted masks to extract related CLIP features with Mask Pooling operation. To integrate heterogeneous knowledge spaces and improve the representation robustness, Gradient-Free Aggregation fuses the query embeddings with pooled CLIP features. The aggregated visual embeddings are combined with linguistic embeddings to perform visual-linguistic alignment and generate final output. During training, a Proxy Calibration strategy is utilized to expand training space by applying supervision on the convex combination of training semantics." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.471, + 0.758, + 0.485 + ], + "angle": 0, + "content": "4.2. Gradient-Free Aggregation" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.493, + 0.905, + 0.643 + ], + "angle": 0, + "content": "Although CLIP [46] exhibits strong generalization ability, it optimizes for image-level visual-linguistic alignment during pretraining, which creates a great domain gap compared to mask-level alignment. As a result, directly using CLIP features obtained through mask pooling for mask classification may lead to suboptimal performance. The query embeddings generated by the segmentation decoder, on the other hand, have a strong region-level alignment prior owing to their suitability for the training space. However, they struggle to recognize novel semantics." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.645, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Therefore, we propose to reorganize them to obtain more robust domain-adapted representations. An intuitive approach is to learn the intrinsic connection between them through self-attention or cross-attention mechanisms. However, we observe that such learning-based strategy tends to develop excessive reliance on query embeddings fitted to seen semantics during the training process while neglecting the utilization of more generalizable information provided by CLIP, ultimately affecting the model's ability to understand unlimited categories. Therefore, we propose to combine CLIP features and query embeddings via a gradient-free approach to avoid potential overfitting problems during the learning of the joint space. This is inspired by Random Walk algorithm [16, 56]. Specifically, there are multiple iterations in this process where two features learn collaboratively in shared embedding space until convergence. Suppose we have \\(N\\) queries. For convenience, we omit the" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22667" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.904, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.318, + 0.907, + 0.404 + ], + "angle": 0, + "content": "Figure 3. Pipeline of our OVSNet. The input image and corresponding categories are first encoded by pretrained CLIP [46]. A segmentation decoder is trained to decode the visual features and generate potential masks as well as corresponding query embeddings. We then leverage the predicted masks to extract related CLIP features with Mask Pooling operation. To integrate different knowledge spaces and improve the representation robustness, Gradient-Free Aggregation fuses the query embeddings with CLIP features. The aggregated visual embeddings are combined with class embeddings to perform visual-linguistic alignment and generate final output. During training, we apply Proxy Calibration strategy to expand training space in a cost-free manner." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.421, + 0.482, + 0.56 + ], + "angle": 0, + "content": "query index and take a random query as the example. The corresponding query embeddings and CLIP features can be represented by \\( F_{Q} \\) and \\( F_{C} \\). \\( F_{Q} \\) and \\( F_{C} \\) are defined as the initial state \\( F_{Q}^{0} \\) and \\( F_{C}^{0} \\) at the 0-th iteration. Then we formulate affinity \\( \\mathcal{Z} \\) by scaled dot product: \\( \\mathcal{Z} = \\lambda F_{C}^{0}(F_{Q}^{0})^{\\top} \\), where \\( \\lambda \\) is the scaling factor. At \\( t \\)-th iteration, the query embedding \\( F_{Q}^{t} \\) is derived by the original embedding \\( F_{Q}^{0} \\) and the updated CLIP feature \\( F_{C}^{t-1} \\) from previous iteration, which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.566, + 0.483, + 0.587 + ], + "angle": 0, + "content": "\\[\nF _ {Q} ^ {t} = \\omega \\operatorname {N o r m} (\\mathcal {Z}) ^ {\\top} F _ {C} ^ {t - 1} + (1 - \\omega) F _ {Q} ^ {0}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.59, + 0.483, + 0.618 + ], + "angle": 0, + "content": "Subsequently, it is utilized to boost the CLIP feature of this iteration:" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.618, + 0.482, + 0.637 + ], + "angle": 0, + "content": "\\[\nF _ {C} ^ {t} = \\omega \\mathcal {Z} F _ {Q} ^ {t} + (1 - \\omega) F _ {C} ^ {0}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.639, + 0.485, + 0.686 + ], + "angle": 0, + "content": "where \\(\\omega \\in (0,1)\\) is the factor that controls the degree of feature fusion. By combining the above equations we can obtain the overall iterative update formula of \\(F_{C}^{t}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.69, + 0.483, + 0.731 + ], + "angle": 0, + "content": "\\[\nF _ {C} ^ {t} = \\left(\\omega^ {2} A\\right) ^ {t} F _ {C} ^ {0} + (1 - \\omega) \\sum_ {i = 0} ^ {t - 1} \\left(\\omega^ {2} A\\right) ^ {i} \\left(\\omega \\mathcal {Z} F _ {Q} ^ {0} + F _ {C} ^ {0}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.737, + 0.483, + 0.814 + ], + "angle": 0, + "content": "where \\(A\\) denotes \\(\\mathcal{Z}\\mathrm{Norm}(\\mathcal{Z})^{\\top}\\). Moreover, to avoid the potential issue of unexpected gradient and simplify the calculation process, we use an approximate inference function based on Neumann Series [43] when \\(t\\to \\infty\\). Then the update formula can be present as:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.818, + 0.483, + 0.837 + ], + "angle": 0, + "content": "\\[\nF _ {C} ^ {\\infty} = (1 - \\omega) \\left(I - \\omega^ {2} A\\right) ^ {- 1} \\left(\\omega \\mathcal {Z} F _ {Q} ^ {0} + F _ {C} ^ {0}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\( I \\) represents the identity matrix. With this gradient-free aggregation strategy, the model can recombine query embeddings with CLIP features to produce more robust region-level representations." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.42, + 0.691, + 0.436 + ], + "angle": 0, + "content": "4.3. Proxy Calibration" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.452, + 0.907, + 0.59 + ], + "angle": 0, + "content": "A broader semantic training space tends to result in more generalized model representations. However, expanding the training space for OVS is costly due to the intensive prediction nature of segmentation tasks. To alleviate this problem, we propose a cost-free approach named Proxy Calibration (PC) to expand the training space and enhance the generalization ability of the model. It leverages the synthesis of proxy embeddings, which approximates novel semantics via convex combination between the in-vocabulary classes." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.597, + 0.907, + 0.675 + ], + "angle": 0, + "content": "Specifically, having the aggregated query embeddings \\( F_{Q} \\), we perform random convex combination of them during training to simulate embedding space for undefined classes, as shown by the black dotted line in Figure 3. This process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.703, + 0.907, + 0.723 + ], + "angle": 0, + "content": "\\[\nF _ {Q m n} ^ {\\prime} = \\alpha * F _ {Q m} + (1 - \\alpha) * F _ {Q n} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "where \\(m, n \\in \\{1, 2, \\dots, N\\}\\). \\(\\alpha\\) is a random weight sampled from the distribution of \\(Beta(\\gamma, \\gamma)\\). \\(F'\\) denotes the generated proxy embeddings. The same enhancement is performed for the corresponding CLIP feature \\(F_C\\) and linguistic embeddings \\(F_T\\). Note that the text features here refer to the class representations of the ground truth regions corresponding to each query, which are obtained after Hungarian Matching [5]. Having the generated \\(F_Q'\\), \\(F_C'\\), and \\(F_T'\\), we perform distance supervision on them to promote region-level visual-linguistic alignment performance from the per" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22668" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.133 + ], + "angle": 0, + "content": "Table 2. Performance comparison with state-of-the-art methods on existing benchmarks and the proposed OpenBench. ADE, PC, and VOC denote ADE20K [66], Pascal Context [44], and Pascal VOC [13], respectively. 171 and 133 indicate the number of training categories. * denotes the re-trained version according to the official code for fair comparison." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.135, + 0.911, + 0.44 + ], + "angle": 0, + "content": "
MethodTraining DatasetADE-150ADE-847PC-59PC-459VOCOpenBenchAverage Score
Base-level CLIP Backbone
SimSeg* [60]COCO-Stuff-17121.16.951.99.791.819.433.5
OVSeg [31]COCO-Stuff-17124.87.153.311.092.624.935.6
SAN [61]COCO-Stuff-17127.510.153.812.694.039.639.6
SCAN [37]COCO-Stuff-17130.810.858.413.297.038.741.5
CATSeg [10]COCO-Stuff-17131.812.057.519.094.636.141.8
SED [57]COCO-Stuff-17131.611.457.318.694.433.341.1
FCCLIP* [63]COCO-Panoptic-13331.113.554.812.893.240.340.9
MAFT [25]COCO-Panoptic-13329.110.153.512.890.037.538.8
MAFT+ [26]COCO-Panoptic-13334.613.857.516.295.443.743.5
OVSNet (Ours)COCO-Panoptic-13335.814.558.619.195.744.944.8
Large-level CLIP Backbone
SimSeg* [60]COCO-Stuff-17121.77.152.210.292.327.835.2
OVSeg [31]COCO-Stuff-17129.69.055.712.494.535.639.5
SAN [61]COCO-Stuff-17132.112.457.715.794.643.442.7
SCAN [37]COCO-Stuff-17133.514.059.316.397.149.444.9
SED [57]COCO-Stuff-17135.213.960.622.696.145.645.6
FCCLIP [63]COCO-Panoptic-13334.014.858.418.295.444.844.3
MAFT [25]COCO-Panoptic-13334.413.157.517.093.041.742.8
MAFT+ [26]COCO-Panoptic-13336.115.159.421.696.547.346.0
OVSNet (Ours)COCO-Panoptic-13337.116.262.023.596.948.647.4
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.453, + 0.336, + 0.468 + ], + "angle": 0, + "content": "spective of proxy semantics. That is:" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.475, + 0.483, + 0.513 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {P Q} = 1 - \\frac {\\mathbf {F} _ {\\mathbf {Q}} ^ {\\prime} \\cdot \\mathbf {F} _ {\\mathbf {T}} ^ {\\prime}}{\\| \\mathbf {F} _ {\\mathbf {Q}} ^ {\\prime} \\| _ {2} \\| \\mathbf {F} _ {\\mathbf {T}} ^ {\\prime} \\| _ {2}}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.514, + 0.483, + 0.55 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {P C} = 1 - \\frac {\\mathbf {F} _ {\\mathbf {C}} ^ {\\prime} \\cdot \\mathbf {F} _ {\\mathbf {T}} ^ {\\prime}}{\\| \\mathbf {F} _ {\\mathbf {C}} ^ {\\prime} \\| _ {2} \\| \\mathbf {F} _ {\\mathbf {T}} ^ {\\prime} \\| _ {2}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.555, + 0.484, + 0.646 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{PQ}\\) and \\(\\mathcal{L}_{PC}\\) denote the applied proxy supervision for query embeddings and pooled CLIP features, respectively. The final proxy loss is the sum of \\(\\mathcal{L}_{PQ}\\) and \\(\\mathcal{L}_{PC}\\). By applying additional supervision in this way, the model is not limited to a specific training space and can learn more robust visual representations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.658, + 0.217, + 0.675 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.396, + 0.698 + ], + "angle": 0, + "content": "5.1. Evaluation Benchmark and Metric" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To evaluate the effectiveness and generalization ability of our method on diverse scenarios, we conduct extensive experiments on the popular existing benchmarks, ADE20K150 [66], ADE20K847 [66], Pascal VOC [13], Pascal Context-59 [44], Pascal Context-459 [44] and our proposed OpenBench. ADE20K is a large-scale scene understanding benchmark, containing 20k training images, 2k validation images, and 3k testing images. There are two splits of this dataset. ADE20K-150 contains 150 semantic classes whereas ADE20K-847 has 847 classes. The images of both are the same. Pascal Context is an extension of Pascal VOC 2010, containing 5,005 validation images. We take the commonly used PC-59 and challenging PC-459" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.452, + 0.907, + 0.558 + ], + "angle": 0, + "content": "version for validation. Pascal VOC contains 11,185 training images and 1,449 validation images from 20 classes. We use the provided augmented annotations. Following previous works [12, 15, 60], we take the mean-intersection-over-union (mIoU) as the metric to compare our model with previous state-of-the-art methods. In ablation, we assume the use of base-level CLIP as the backbone for comparison." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.57, + 0.733, + 0.587 + ], + "angle": 0, + "content": "5.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.907, + 0.821 + ], + "angle": 0, + "content": "We utilize the Mask2Former [8] as the segmentation decoder and the implementation is based on detectron2 [54]. The model is trained with batch size of 8 and total iteration of \\(60\\mathrm{k}\\). The base learning rate is 0.0001 with a polynomial schedule. Following previous works [26, 63], we take CLIP [46] of ConvNeXt as the backbone. The input image is resized and cropped to \\(1024\\times 1024\\). For data augmentation, random flip and color augmentation are adopted. The weight decay of the model is 0.05. \\(\\lambda\\) and \\(\\gamma\\) of GFA and PC are set to 0.2 and 2 by default. The segmentation loss consists of dice loss and cross entropy loss. The classification loss is cross entropy loss. For the weights of the loss function, we set 5 and 2 for segmentation loss and classification loss, respectively. Other hyperparameters are the same as Mask2Former [9]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.832, + 0.655, + 0.848 + ], + "angle": 0, + "content": "5.3. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.903 + ], + "angle": 0, + "content": "We compare our model with existing state-of-the-art approaches in Table 2. It can be seen that with proxy calibration expanding training semantics and the gradient-free" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22669" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.483, + 0.132 + ], + "angle": 0, + "content": "Table 3. Ablation experiments about the proposed methods. GFA and PC indicate the Gradient-Free Aggregation and Proxy Calibration strategies, respectively." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.134, + 0.482, + 0.21 + ], + "angle": 0, + "content": "
MethodADE-150PC-459OpenBench
Baseline33.114.342.3
+ GFA34.7↑ (1.6)16.0↑ (1.7)43.7↑ (1.4)
+ PC33.9↑ (0.8)17.2↑ (2.9)44.3↑ (2.0)
+ Both35.8↑ (2.7)19.1↑ (4.8)44.9↑ (2.6)
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.215, + 0.487, + 0.457 + ], + "angle": 0, + "content": "aggregation algorithm modeling the joint space, our model achieves the state-of-the-art performance on most popular benchmarks. The last column counts the average performance across multiple datasets, which shows that our method performs well for both in-vocabulary as well as out-vocabulary diverse scenarios. The relatively inferior performance of our method on VOC is due to the fact that the semantics of this dataset are almost contained in the training set, which tends to benefit models such as SCAN [37] that finetunes the CLIP backbone on training set. Besides, the greater divergence of OpenBench from training semantics does not imply that it is totally more challenging than other benchmarks, because the final performance is influenced by various factors such as image content complexity, image clarity, and the number of given categories. More discussion can be seen in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.468, + 0.245, + 0.484 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.49, + 0.483, + 0.643 + ], + "angle": 0, + "content": "Component Analysis Table 3 shows the ablations about the proposed Gradient-Free Aggregation (GFA) and Proxy Calibration (PC) strategy. We evaluate their effectiveness on both existing benchmarks and our proposed OpenBench. It can be seen that with GFA fusing CLIP feature and query embeddings, the model's performance is greatly improved on both seen and unseen semantics. Besides, PC can facilitate the model's ability to understand various scenarios by cost-free expansion of the training semantic space. When applying both, the model achieves the best performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.65, + 0.483, + 0.833 + ], + "angle": 0, + "content": "Embedding Distribution Analysis To demonstrate that our proposed OpenBench is more open-vocabulary compared to the existing test sets, we perform t-SNE visualization of the category embeddings of the training set, the existing test sets, and our OpenBench in Figure 4. Besides, we also visualize the proxy embeddings generated by PC. We can see that the existing test sets (green dots) are overall similar to the training semantics (blue dots). Our OpenBench (orange dots), on the other hand, shows significant differences. In addition, the proxy embeddings (red dots) generated by PC expand the training set and thus help to improve the robustness and generalization of the model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Analysis about Inference Class Number In this part we explore an interesting phenomenon: why the OVS models perform much worse on ADE-847 and PC-459 than on other test sets? Intuitively, this is because these datasets" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.09, + 0.883, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.282, + 0.907, + 0.325 + ], + "angle": 0, + "content": "Figure 4. Visualization of category embeddings across the training space, existing evaluation space, our OpenBench space, and the generated proxy space." + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.332, + 0.885, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.501, + 0.907, + 0.543 + ], + "angle": 0, + "content": "Figure 5. Impact of the number of inference categories on model performance. As the number of irrelevant categories increases, the model's performance on the same image degrades." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.563, + 0.907, + 0.791 + ], + "angle": 0, + "content": "contain more novel categories. However, we find that another factor significantly impacts model performance, that is the number of categories fed into the model during inference. Let's start with the conclusion: for the same model and test image, the model performance is worse when there are more inference categories given. In other words, the model may be able to make correct predictions when only a small number of categories are given. However, when more categories are prompted, the model's output will be incorrect. Here we take two typical methods, SAN [61] and MAFT-Plus [26], as examples and show how their performance on the ADE-150 dataset changes as the number of categories increases. As shown in Figure 5, with the number of given irrelevant categories increases, the models of different technology routes show a consistent decline." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Internal Analysis of Proposed Strategies In Table 4 we conduct more experiments about the proposed methods. Specifically, we first compare our gradient-free aggregation designs with the typical learning-based feature fusion approaches in (a). Results show that vanilla attention performs inferior than our GFA, especially on the more open-vocabulary set OpenBench (degraded by more than" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22670" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.09, + 0.298, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.186, + 0.298, + 0.197 + ], + "angle": 0, + "content": "(a) Seen: building treefence bicycle" + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.091, + 0.497, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.298, + 0.186, + 0.496, + 0.197 + ], + "angle": 0, + "content": "Unseen: sidewalk streetlight" + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.199, + 0.298, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.297, + 0.298, + 0.308 + ], + "angle": 0, + "content": "(c) Seen: ceiling desk chair" + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.199, + 0.497, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.298, + 0.297, + 0.496, + 0.31 + ], + "angle": 0, + "content": "Unseen: computer chandelier" + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.311, + 0.298, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.408, + 0.298, + 0.419 + ], + "angle": 0, + "content": "(e) building" + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.311, + 0.497, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.298, + 0.408, + 0.441, + 0.419 + ], + "angle": 0, + "content": "→ opera house" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.091, + 0.699, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.505, + 0.186, + 0.699, + 0.197 + ], + "angle": 0, + "content": "(b) Seen: sea boat mountain" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.091, + 0.899, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.705, + 0.186, + 0.899, + 0.197 + ], + "angle": 0, + "content": "Unseen: pier pole" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.199, + 0.699, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.505, + 0.297, + 0.698, + 0.308 + ], + "angle": 0, + "content": "(d) Seen: plant rock person" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.199, + 0.899, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.705, + 0.297, + 0.899, + 0.308 + ], + "angle": 0, + "content": "Unseen: waterfall" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.311, + 0.7, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.408, + 0.525, + 0.419 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.311, + 0.899, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.705, + 0.408, + 0.805, + 0.419 + ], + "angle": 0, + "content": "→ ferrari" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.422, + 0.907, + 0.451 + ], + "angle": 0, + "content": "Figure 6. Visualization of segmentation results. (a) - (d) demonstrate the successful segmentation cases of our method for seen and unseen categories. (e) and (f) display the adaptability for flexible text query. Best viewed in color." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.462, + 0.484, + 0.518 + ], + "angle": 0, + "content": "Table 4. Ablation experiments within each modules. (a) shows the comparison of our gradient-free algorithm with learning-based method, e.g., cross attention. (b) demonstrates the effect of \\(\\gamma\\) of Beta distribution in the proxy learning strategy." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.522, + 0.472, + 0.641 + ], + "angle": 0, + "content": "
ADE-150PC-459OpenBench
(a) Comparisons of GFA and learning-based strategy
Self Attn34.014.740.6
Cross Attn34.214.841.2
GFA34.716.043.7
(b) Effect of sampling distribution of PC
γ = 133.114.342.3
γ = 0.533.6↑ (0.5)16.1↑ (1.8)43.7↑ (1.4)
γ = 233.9↑ (0.8)17.2↑ (2.9)44.3↑ (2.0)
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.655, + 0.483, + 0.73 + ], + "angle": 0, + "content": "2.5 mIoU). This is because learning-based methods tend to develop a reliance on the segmentation query embeddings during training, while overlooking the CLIP features that inherently exhibit stronger generalization capabilities, thereby introducing biases." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.731, + 0.483, + 0.822 + ], + "angle": 0, + "content": "We also experiment with the influence of different \\(\\gamma\\) in Beta distribution to the performance of proxy calibration. The results are shown in Table 4 (b), where \\(\\gamma = 1\\) is equivalent to do not utilize PC. We find that different \\(\\gamma\\) benefit the model performance while the performance is better when the intermediate probability density is high (\\(\\gamma = 2\\))." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.228, + 0.847 + ], + "angle": 0, + "content": "5.5. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Figure 6 shows some segmentation cases of our OVSNet. It can be seen that our method achieves excellent segmentation performance on various scenarios. Specifically, (a) - (d)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.464, + 0.907, + 0.525 + ], + "angle": 0, + "content": "demonstrate the excellent segmentation of our method for seen and unseen categories. (e) and (f) display the adaptability for flexible text query, e.g., change \"building\" to \"opera house\" and \"car\" to \"ferrari\"." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.541, + 0.634, + 0.557 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.907, + 0.824 + ], + "angle": 0, + "content": "In this paper, we identify that the existing open-vocabulary segmentation evaluation sets suffer from excessive semantic similarity to the training set, which limits their ability to comprehensively assess the model's open vocabulary understanding capabilities. To address this, we propose a new benchmark, OpenBench, which features significant differences from the training set and includes more fine-grained categories. Comparison on OpenBench further demonstrates that maintaining the generalization space of CLIP is crucial for OVS. Additionally, we introduce OVSNet, a method that leverages proxy calibration for cost-free augmentation of the training space and gradient-free aggregation for fusing heterogeneous features. Experiments show that the proposed OVSNet achieves excellent performance on both existing datasets and the newly proposed OpenBench. We hope that our benchmark and model design could inspire further researches in this area." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by Shenzhen Science and Technology Program under Grant CJGJZD20220517142402006 and Guangdong Natural Science Funds for Distinguished Young Scholar (No. 2025B1515020012)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22671" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.483, + 0.157 + ], + "angle": 0, + "content": "[1] Sule Bai, Yong Liu, Yifei Han, Haoji Zhang, and Yansong Tang. Self-calibrated clip for training-free open-vocabulary segmentation. arXiv preprint arXiv:2411.15869, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.157, + 0.483, + 0.213 + ], + "angle": 0, + "content": "[2] Sule Bai, Mingxing Li, Yong Liu, Jing Tang, Haoji Zhang, Lei Sun, Xiangxiang Chu, and Yansong Tang. Univg-r1: Reasoning guided universal visual grounding with reinforcement learning. arXiv preprint arXiv:2505.14231, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.213, + 0.483, + 0.267 + ], + "angle": 0, + "content": "[3] Gabriel J. Brostow, Jamie Shotton, Julien Fauqueur, and Roberto Cipolla. Segmentation and recognition using structure from motion point clouds. In ECCV, pages 44-57, 2008. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.269, + 0.482, + 0.31 + ], + "angle": 0, + "content": "[4] Maxime Bucher, Tuan-Hung Vu, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. In NeurlPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.311, + 0.482, + 0.354 + ], + "angle": 0, + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.482, + 0.41 + ], + "angle": 0, + "content": "[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. TPAMI, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.411, + 0.482, + 0.466 + ], + "angle": 0, + "content": "[7] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. UNITER: learning universal image-text representations. arXiv preprint arXiv:1909.11740, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.466, + 0.482, + 0.522 + ], + "angle": 0, + "content": "[8] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G. Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.523, + 0.482, + 0.577 + ], + "angle": 0, + "content": "[9] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.578, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[10] Seokju Cho, Heeseong Shin, Sunghwan Hong, Anurag Arnab, Paul Hongsuck Seo, and Seungryong Kim. Cat-seg: Cost aggregation for open-vocabulary semantic segmentation. In CVPR, pages 4113-4123, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.634, + 0.482, + 0.703 + ], + "angle": 0, + "content": "[11] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.704, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[12] Jian Ding, Nan Xue, Gui-Song Xia, and Dengxin Dai. Decoupling zero-shot semantic segmentation. In CVPR, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[13] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes challenge: A retrospective. IJCV, 111:98-136, 2015. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[14] Shanghua Gao, Zhong-Yu Li, Ming-Hsuan Yang, Ming-Ming Cheng, Junwei Han, and Philip Torr. Large-scale unsupervised semantic segmentation. TPAMI, 45(6):7457-7476, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.858, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[15] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Open-vocabulary image segmentation. arXiv preprint arXiv:2112.12143, 2021. 1, 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.12 + ], + "angle": 0, + "content": "[16] Leo Grady. Random walks for image segmentation. IEEE TPAMI, 28(11):1768-1783, 2006. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.906, + 0.205 + ], + "angle": 0, + "content": "[17] Kunyang Han, Yong Liu, Jun Hao Liew, Henghui Ding, Jiajun Liu, Yitong Wang, Yansong Tang, Yujiu Yang, Jiashi Feng, Yao Zhao, et al. Global knowledge calibration for fast open-vocabulary segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 797-807, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[18] Shuting He, Henghui Ding, and Wei Jiang. Primitive generation and semantic-related alignment for universal zero-shot segmentation. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[19] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[20] Jiaqi Huang, Zunnan Xu, Ting Liu, Yong Liu, Haonan Han, Kehong Yuan, and Xiu Li. Densely connected parameter-efficient tuning for referring image segmentation. arXiv preprint arXiv:2501.08580, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[21] Xiaoke Huang, Jianfeng Wang, Yansong Tang, Zheng Zhang, Han Hu, Jiwen Lu, Lijuan Wang, and Zicheng Liu. Segment and caption anything. arXiv preprint arXiv:2312.00869, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[22] Joonhyun Jeong, Geondo Park, Jayeon Yoo, Hyungsik Jung, and Heesu Kim. Proxydet: Synthesizing proxy novel classes via classwise mixup for open-vocabulary object detection. In AAAI, pages 2462-2470, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.546 + ], + "angle": 0, + "content": "[23] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.546, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[24] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.616, + 0.905, + 0.672 + ], + "angle": 0, + "content": "[25] Siyu Jiao, Yunchao Wei, Yaowei Wang, Yao Zhao, and Humphrey Shi. Learning mask-aware clip representations for zero-shot segmentation. arXiv preprint arXiv:2310.00240, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.673, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[26] Siyu Jiao, Hongguang Zhu, Jiannan Huang, Yao Zhao, Yunchao Wei, and Humphrey Shi. Collaborative vision-text representation optimizing for open-vocabulary segmentation. In ECCV, pages 399-416, 2025. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.73, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[27] Boyi Li, Kilian Q. Weinberger, Serge J. Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. In ICLR, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.772, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[28] Gen Li, Nan Duan, Yuejian Fang, Ming Gong, and Daxin Jiang. Unicoder-vl: A universal encoder for vision and language by cross-modal pre-training. In AAAI, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[29] Wanhua Li, Xiaoke Huang, Zheng Zhu, Yansong Tang, Xiu Li, Jie Zhou, and Jiwen Lu. Ordinalclip: Learning rank prompts for language-guided ordinal regression. NeurIPS, pages 35313-35325, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.872, + 0.905, + 0.903 + ], + "angle": 0, + "content": "[30] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22672" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-semantics aligned pre-training for vision-language tasks. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[31] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yi-nan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted CLIP. arXiv preprint arXiv:2210.04150, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.483, + 0.261 + ], + "angle": 0, + "content": "[32] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In ECCV, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.483, + 0.304 + ], + "angle": 0, + "content": "[33] Yong Liu, Ran Yu, Xinyuan Zhao, and Yujiu Yang. Quality-aware and selective prior enhancement memory network for video object segmentation. In CVPR Workshop, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.483, + 0.361 + ], + "angle": 0, + "content": "[34] Yong Liu, Ran Yu, Jiahao Wang, Xinyuan Zhao, Yitong Wang, Yansong Tang, and Yujiu Yang. Global spectral filter memory network for video object segmentation. In ECCV, pages 648-665, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.363, + 0.483, + 0.417 + ], + "angle": 0, + "content": "[35] Yong Liu, Ran Yu, Fei Yin, Xinyuan Zhao, Wei Zhao, Weihao Xia, and Yujiu Yang. Learning quality-aware dynamic memory for video object segmentation. In ECCV, pages 468-486, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.483, + 0.473 + ], + "angle": 0, + "content": "[36] Yong Liu, Cairong Zhang, Yitong Wang, Jiahao Wang, Yujiu Yang, and Yansong Tang. Universal segmentation at arbitrary granularity with language instruction. arXiv preprint arXiv:2312.01623, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.483, + 0.518 + ], + "angle": 0, + "content": "[37] Yong Liu, Sule Bai, Guanbin Li, Yitong Wang, and Yansong Tang. Open-vocabulary segmentation with semantic-assisted calibration. In CVPR, pages 3491-3500, 2024. 1, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.52, + 0.483, + 0.574 + ], + "angle": 0, + "content": "[38] Yong Liu, Ran Yu, Fei Yin, Xinyuan Zhao, Wei Zhao, Weihao Xia, Jiahao Wang, Yitong Wang, Yansong Tang, and Yujiu Yang. Learning high-quality dynamic memory for video object segmentation. IEEE TPAMI, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.577, + 0.483, + 0.63 + ], + "angle": 0, + "content": "[39] Guanxing Lu, Ziwei Wang, Changliu Liu, Jiwen Lu, and Yansong Tang. Thinkbot: Embodied instruction following with thought chain reasoning. arXiv preprint arXiv:2312.07062, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.483, + 0.674 + ], + "angle": 0, + "content": "[40] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In NeurIPS, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[41] Zhuoyan Luo, Yicheng Xiao, Yong Liu, Shuyan Li, Yitong Wang, Yansong Tang, Xiu Li, and Yujiu Yang. Soc: Semantic-assisted object cluster for referring video object segmentation. arXiv preprint arXiv:2305.17011, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[42] Zhuoyan Luo, Yinghao Wu, Yong Liu, Yicheng Xiao, Xiaoping Zhang, and Yujiu Yang. Hdc: Hierarchical semantic decoding with counting assistance for generalized referring expression segmentation. arXiv preprint arXiv:2405.15658, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[43] Carl D Meyer and Ian Stewart. Matrix analysis and applied linear algebra. SIAM, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[44] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan L. Yuille. The role of context for object detection and semantic segmentation in the wild. In CVPR, 2014. 2, 3, 4, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[45] Mengxue Qu, Yu Wu, Wu Liu, Qiqi Gong, Xiaodan Liang, Olga Russakovsky, Yao Zhao, and Yunchao Wei. Siri: A simple selective retraining mechanism for transformer-based visual grounding. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.237, + 0.905, + 0.278 + ], + "angle": 0, + "content": "[47] Yiqin Wang, Haoji Zhang, Yansong Tang, Yong Liu, Jiashi Feng, Jifeng Dai, and Xiaojie Jin. Hierarchical memory for long video qa. arXiv preprint arXiv:2407.00603, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.281, + 0.905, + 0.334 + ], + "angle": 0, + "content": "[48] Yulin Wang, Haoji Zhang, Yang Yue, Shiji Song, Chao Deng, Junlan Feng, and Gao Huang. Uni-adafocus: Spatial-temporal dynamic computation for video recognition. IEEE TPAMI, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.339, + 0.905, + 0.393 + ], + "angle": 0, + "content": "[49] Yuji Wang, Jingchen Ni, Yong Liu, Chun Yuan, and Yansong Tang. Iterprime: Zero-shot referring image segmentation with iterative grad-cam refinement and primary word emphasis. In AAAI, pages 8159-8168, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.397, + 0.905, + 0.45 + ], + "angle": 0, + "content": "[50] Yuji Wang, Haoran Xu, Yong Liu, Jiaze Li, and Yansong Tang. Sam2-love: Segment anything model 2 in language-aided audio-visual scenes. In CVPR, pages 28932-28941, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.454, + 0.905, + 0.509 + ], + "angle": 0, + "content": "[51] Cong Wei, Yujie Zhong, Haoxian Tan, Yong Liu, Zheng Zhao, Jie Hu, and Yujiu Yang. Hyperseg: Towards universal visual segmentation with large language model. arXiv preprint arXiv:2411.17606, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.512, + 0.905, + 0.567 + ], + "angle": 0, + "content": "[52] Cong Wei, Yujie Zhong, Haoxian Tan, Yingsen Zeng, Yong Liu, Zheng Zhao, and Yujiu Yang. Instructseg: Unifying instructed visual segmentation with multi-modal large language models. arXiv preprint arXiv:2412.14006, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.57, + 0.905, + 0.611 + ], + "angle": 0, + "content": "[53] Xiongwei Wu, Xin Fu, Ying Liu, Ee-Peng Lim, Steven CH Hoi, and Qianru Sun. A large-scale benchmark for food image segmentation. In ACM MM, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.614, + 0.905, + 0.655 + ], + "angle": 0, + "content": "[54] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.658, + 0.905, + 0.712 + ], + "angle": 0, + "content": "[55] Yongqin Xian, Subhabrata Choudhury, Yang He, Bernt Schiele, and Zeynep Akata. Semantic projection network for zero- and few-label semantic segmentation. In CVPR, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.716, + 0.905, + 0.783 + ], + "angle": 0, + "content": "[56] Yicheng Xiao, Zhuoyan Luo, Yong Liu, Yue Ma, Hengwei Bian, Yatai Ji, Yujiu Yang, and Xiu Li. Bridging the gap: A unified video comprehension framework for moment retrieval and highlight detection. In CVPR, pages 18709-18719, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.787, + 0.905, + 0.842 + ], + "angle": 0, + "content": "[57] Bin Xie, Jiale Cao, Jin Xie, Fahad Shahbaz Khan, and Yanwei Pang. Sed: A simple encoder-decoder for open-vocabulary semantic segmentation. In CVPR, pages 3426-3436, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[58] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In NIPS, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22673" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[59] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, pages 2955-2966, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[60] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A simple baseline for zero-shot semantic segmentation with pre-trained vision-language model. arXiv preprint arXiv:2112.14757, 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[61] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xi-ang Bai. Side adapter network for open-vocabulary semantic segmentation. In CVPR, pages 2945–2954, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[62] Xin Xu, Tianyi Xiong, Zheng Ding, and Zhuowen Tu. Masqclip for open-vocabulary universal image segmentation. In ICCV, pages 887-898, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[63] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. NeurIPS, pages 32215-32234, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[64] Hui Zhang and Henghui Ding. Prototypical matching and open set rejection for zero-shot semantic segmentation. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[65] Haoji Zhang, Yiqin Wang, Yansong Tang, Yong Liu, Jiashi Feng, Jifeng Dai, and Xiaojie Jin. Flash-vstream: Memory-based real-time understanding for long video streams. arXiv preprint arXiv:2406.08085, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[66] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ADE20K dataset. In CVPR, 2017. 2, 3, 4, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22674" + } + ] +] \ No newline at end of file diff --git a/2025/Stepping Out of Similar Semantic Space for Open-Vocabulary Segmentation/f495e889-2c8e-4e64-ba9d-86b1aec5716d_origin.pdf b/2025/Stepping Out of Similar Semantic Space for Open-Vocabulary Segmentation/f495e889-2c8e-4e64-ba9d-86b1aec5716d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c4844222ceb1ad2f52f9cb239b02d3c592f4d9db --- /dev/null +++ b/2025/Stepping Out of Similar Semantic Space for Open-Vocabulary Segmentation/f495e889-2c8e-4e64-ba9d-86b1aec5716d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7f68f92da61fe58bc3c514088d02f58153e8bdcebc62baec2580ae10091882a6 +size 2926313 diff --git a/2025/Stepping Out of Similar Semantic Space for Open-Vocabulary Segmentation/full.md b/2025/Stepping Out of Similar Semantic Space for Open-Vocabulary Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e6f3887cb60cf814a99b07fa3b07d65f955bd1d7 --- /dev/null +++ b/2025/Stepping Out of Similar Semantic Space for Open-Vocabulary Segmentation/full.md @@ -0,0 +1,318 @@ +# Stepping Out of Similar Semantic Space for Open-Vocabulary Segmentation + +Yong Liu $^{1,2*}$ , Song-Li Wu $^{1,2*}$ , Sule Bai $^{1,2*}$ , Jiahao Wang $^{3}$ , Yitong Wang $^{4}$ , Yansong Tang $^{1,2\dagger}$ $^{1}$ Tsinghua Shenzhen International Graduate School $^{2}$ Tsinghua University + $^{3}$ The University of Hong Kong $^{4}$ ByteDance Inc. + +liuyong23@mails.tsinghua.edu.cn, tang.yansong@sz.tsinghua.edu.cn + +# Abstract + +Open-vocabulary segmentation aims to achieve segmentation of arbitrary categories given unlimited text inputs as guidance. To achieve this, recent works have focused on developing various technical routes to exploit the potential of large-scale pre-trained vision-language models and have made significant progress on existing benchmarks. However, we find that existing test sets are limited in measuring the models' comprehension of "open-vocabulary" concepts, as their semantic space closely resembles the training space, even with many overlapping categories. To this end, we present a new benchmark named OpenBench that differs significantly from the training semantics. It is designed to better assess the model's ability to understand and segment a wide range of real-world concepts. When testing existing methods on OpenBench, we find that their performance diverges from the conclusions drawn on existing test sets. In addition, we propose a method named OVSNet to improve the segmentation performance for diverse and open scenarios. Through elaborate fusion of heterogeneous features and cost-free expansion of the training space, OVSNet achieves state-of-the-art results on both existing datasets and our proposed OpenBench. Corresponding analysis demonstrates the rationality and effectiveness of our proposed benchmark and method. + +# 1. Introduction + +Semantic segmentation is one of the most fundamental tasks in computer vision, which targets at assigning semantic category to pixels in an image. Despite achieving excellent progress in recent years [2, 6, 20, 21, 33-36, 38, 42, 49-52, 58], traditional semantic segmentation approaches rely on predefined categories sets and tend to falter when encountering categories absent during the training phase, significantly impeding their real-world application. Such chal + +![](images/272a77adb469063800da3e373da97c34de40af323c96325c224e0ec6c28273b3.jpg) +(a) + +![](images/b8593b55c6cc94d7cfe44a2bb7cad13e6c9440d3dfe8f2dca35986c7e7469c92.jpg) +(b) +Figure 1. (a) shows the similarity distribution of the existing test set (blue) and our OpenBench (orange) with the training semantic space. (b) illustrates the performance gap between the methods finetuning [10] and frozen [61] CLIP on different benchmarks. + +lenge has inspired the exploration of Open-Vocabulary Segmentation (OVS) task [12, 15, 17, 18, 27, 37, 55, 64]. Different from traditional closed-set segmentation, OVS methods aim to segment arbitrary categories under the guidance of given text inputs, which has many potential applications such as auto-driving and human-robot interaction [19, 39]. + +However, it is challenging to identify unseen categories without the intervention of external knowledge. Therefore, an intuitive idea is to introduce large-scale vision-language model [24, 46] trained with numerous sources to extend the semantic space of segmentation models. Following this idea, existing OVS methods [10, 31, 37, 57, 61, 62] concentrate on exploring how to exploit the pretrained knowledge prior more effectively and derive several technical routes. Early methods [12, 31, 60] adopt a two-stage pipeline. This approach first generates class-agnostic mask proposals with segmentation models, following which the pre-trained CLIP [46] serves as an additional classifier to execute region-level open-vocabulary classification. However, this strategy incurs heavy computation overhead, as it requires repeated forward passes of the CLIP vision encoder for each mask. This can be prohibitively expensive for real-world applications. To address this problem, researchers [25, 61–63] propose to unify classification and segmentation within a shared space, i.e., integrating frozen CLIP features into the segmenter and extracting CLIP embeddings from potential + +target regions to enable simultaneous classification and segmentation. These methods achieve remarkable performance in a more efficient one-stage pipeline. Subsequently, some works [10, 57] discover that finetuning the CLIP encoder could further enhance the model's performance on existing test sets. They develop an early fusion paradigm based on cost map and achieve outstanding segmentation results. However, this performance boost is somewhat counterintuitive, as fine-tuning CLIP typically adapts it to specific training semantic spaces, which could reduce its original generalization ability. Yet, the ability of generalized visual-linguistic alignment is precisely what the open-vocabulary task requires. So why does fine-tuning CLIP appear to benefit the performance of OVS? + +This question drives us to revisit the existing evaluation benchmarks of the cross-dataset evaluation paradigm. We count the maximum similarity of each category in test set to the training semantics. Figure 1 (a) illustrates the similarity distribution of all test categories to the training space (blue color). Results indicate that the existing test data is, in fact, highly similar to the training semantic space. Specifically, even the most challenging ADE-847 [66] and PC-459 [44] datasets have an average similarity to the training classes of 0.79 and 0.83. PC-59 [44] and VOC [13] datasets even have a similarity to the training semantics of 0.95 and 0.97. This explains why finetuning CLIP on training set can lead to further performance gains on existing benchmarks. However, it somewhat conflicts with the purpose of the open-vocabulary setting, which aims to achieve segmentation of unlimited categories in the real world. + +To address the above problem and provide a more comprehensive measure of OVS models, we propose a new evaluation benchmark named OpenBench, which is characterized by significant semantic differences from the training data. Figure 1 (a) also demonstrates the similarity distribution of our OpenBench (orange color). It can be seen that our benchmark contains more novel classes compared to the existing test sets. We evaluate the performance of approaches across different technical routes on our benchmark, and the findings diverge from those obtained on existing test sets. As shown in Figure 1 (b), while finetuning CLIP yields steady performance gains on existing datasets, it shows a significant drop on our benchmark. This outcome also supports our motivation for proposing a benchmark that diverges substantially from the training space. + +In addition, we also present a framework named OVSNet to improve the segmentation performance for diverse and open scenarios. In OVSNet, we adopt a Proxy Calibration (PC) training strategy to enhance the robustness of model representations by broadening the training space with no additional cost. This strategy leverages the synthesis of proxy embeddings, which approximate unseen semantics through convex combinations of in-vocabulary classes. Addition + +ally, inspired by random walk [16], we propose a gradient-free fusion of CLIP embeddings and segmentation decoder features to model a robust joint space and mitigate potential overfitting issues related to training semantics. With the above proposed designs calibrating model knowledge from the aspects of training space and features separately, our OVSNet demonstrates strong performance across a wide range of scenarios, both on existing test sets similar to the training semantics and on our proposed novel OpenBench. Related ablations also demonstrate the effectiveness and soundness of our proposed strategies. + +Our contributions can be summarized as follows: + +- We present a new benchmark named OpenBench for OVS, which is characterized by significant semantic differences from the training space. While using it to evaluate the comprehension of OVS methods in more open scenarios, we reach a different conclusion from that of the existing test sets. +- We propose OVSNet to improve the segmentation performance on diverse and open scenarios. It adopts a Proxy Calibration training strategy to broaden training space without additional cost. Besides, a gradient-free algorithm is leveraged to fuse CLIP embeddings and segmentation features for modeling robust joint space. +- Our method achieves state-of-the-art performance on existing benchmarks and the proposed OpenBench. Extensive experiments are conducted to prove the effectiveness and soundness of the proposed algorithms. + +# 2. Related Work + +Open-Vocabulary Segmentation. This task aims to segment an image and identify regions with arbitrary given text queries [1, 4, 15, 41, 55]. Pioneering work [55] replaces the output convolution layer by computing the similarity between visual features and linguistic embeddings, which has become common practice. Later, some methods [12, 15, 31, 59, 60] propose a two-stage pipeline: the model first generates class-agnostic mask proposals, then a pretrained CLIP [46] is utilized to perform sub-image classification by cropping and masking corresponding regions. With the combination of both in-vocabulary and out-vocabulary classification, these methods obtain great performance. However, this strategy incurs heavy computation overhead, as it requires repeated forward passes of the CLIP vision encoder for each mask. This can be prohibitively expensive for real-world applications. To address this problem, subsequent researchers propose to unify classification and segmentation within a shared space. Among them, SAN [61] designs a side-adapter network to leverage CLIP features for reorganizing segmentation and classification. GKC [17] presents text-guided knowledge distillation strategy to transfer CLIP knowledge to specific classification layer. FCCLIP [63] proposes to leverage + +![](images/8709d79e14e101d89d52c4c2862f7e91ac20ccbf8812adc9082fc19c1e860acc.jpg) +Figure 2. Illustration of the samples in the proposed OpenBench. Compared to existing test sets, our OpenBench has semantic categories that differ more from the training space, which helps to provide a more comprehensive measure of the model's ability to perform open-vocabulary segmentation. The corresponding category name is labeled in the lower right corner of each image. + +Table 1. Comparison of the statistics of our proposed OpenBench with existing datasets. 'Cls', 'Img', and 'Sim' indicate class, image, and similarity, respectively. ADE, PC, and VOC denote ADE20K [66], Pascal Context [44], and Pascal VOC [13] + +
DatasetCls Num.Img Num.Mean Sim.Median Sim.Min Sim.Max Sim.Fine GranularitySemantic Duplication
VOC [13]2014490.97561.00000.84491.0000××
PC-59 [44]5951050.94811.00000.76581.0000××
ADE-150 [66]15020000.80860.82280.19601.0000×✓
Cityscapes [11]331000.84880.83700.67461.0000××
PC-459 [44]45951050.83740.81240.67181.0000✓✓
ADE-847 [66]84720000.79260.79130.19601.0000✓✓
OpenBench (Ours)28660560.61420.64520.26080.7947✓×
+ +the local-awareness advantage of convolution backbones. Some works [10, 57] then discover that finetuning CLIP encoder could further enhance the model's performance on existing test sets and develop effective one-stage framework. + +However, we find that existing test sets are highly similar to the training set, which limits their ability to effectively evaluate the model's comprehension of the open vocabulary semantics. Therefore, in this paper we propose a new benchmark OpenBench that is characterized by significant semantic differences from the training data. We also propose a method named OVSNet to improve the segmentation performance for diverse and open scenarios. + +Vision-Language Pre-training. Vision-language pretraining aims to learn a unified visual-linguistic representation space. Early approaches [7, 28, 30, 40, 45, 47, 48, 65], constrained by small-scale datasets, struggle to achieve strong performance and require fine-tuning for downstream tasks. However, with the advent of large-scale web data, recent works [23, 46] have demonstrated the benefits of leveraging such data to train a more robust multi-modal representation space. Among these, CLIP [46], the most widely used vision-language model, employs contrastive learning to link images with their corresponding captions, + +achieving impressive cross-modal alignment. Building upon previous works [12, 22, 27, 29, 60], we also utilize the well-aligned and generalized space of CLIP to enhance open-vocabulary segmentation. + +# 3. Benchmark + +# 3.1. Dataset Introduction + +In Table 1, we present the statistics comparison of our proposed OpenBench with existing popular open-vocabulary segmentation datasets, including ADE-150 [66], ADE-847 [66], PC-59 [44], PC-459 [44], VOC [13], and Cityscapes [11]. Specifically, our OpenBench contains 286 categories, of which the maximum similarity to the training categories is 0.7947, the minimum is 0.2608, and the average similarity is 0.6142. It can be observed that, compared to existing datasets, our OpenBench exhibits a greater disparity from the training set, which better evaluates the model's segmentation ability for arbitrary open categories. + +Besides, our OpenBench contains fine-grained categories, e.g., "oystercatcher", rather than just generic broad category such as "bird" (corresponds to the "Fine Granularity" characterization in Table 1). This is important for evaluating whether the OVS models preserve the general + +ization capability of CLIP [46], which is inherently capable of detecting these fine-grained categories. Actually, PC-459 [44] and ADE-847 [66] also have fine-grained categories, but they suffer from semantic duplication problem. In detail, semantic duplication means that there will be both fine-grained categories as well as corresponding broad categories in these datasets, which would cause bias in the metrics calculation. For instance, if a model assigns the "door" tag to a region whose ground truth label is "double door", it will be considered incorrect. For open-vocabulary segmentation setting, we argue that the responsibility of the model is to discern the correct semantic. The distribution of such duplication categories and the corresponding performance statistics is unreasonable. SCAN [37] also points similar observations and try to solve this by designing new metric. However, this requires manual judgment of the relationships between categories, introducing additional cost and uncertainty. Our OpenBench, on the other hand, has fine-grained semantics without the problem of semantic duplication, which helps to more accurately reflect the real capabilities of related models. + +Figure 2 illustrates part of the categories contained in OpenBench, where some categories such as food related even differ significantly from the visual domain of the training set. Furthermore, our OpenBench also accounts for "other" or "background" pixels in images, which are often overlooked by existing datasets. This is important for open-vocabulary segmentation because the model is essentially forced to choose the most appropriate category from a predefined set. If there is no correct class and no "others" option in the candidate set, the model would be compelled to make an incorrect prediction. Moreover, "others" or "background" require the model to consider not only the direct visual-linguistic correspondence between the image content and each class, but also the relative relationships among different classes. This adds complexity and makes the task more relevant to real-world scenarios. + +# 3.2. Dataset Collection + +Our OpenBench is primarily derived from existing accessible segmentation datasets. To obtain semantic categories that differ significantly from the training set, i.e., COCO [32], we first organize a group of segmentation datasets that possess category annotations, e.g., Food103 [53], Imagenet-S [14], and CamVid [3]. For all categories in the collected data, we leverage CLIP-L/14 [46] to compute their maximum similarity with COCO categories. Then we calculate the minimum similarity among the categories present in each image as the image's similarity to the training semantics. If the image similarity is greater than the set threshold $\sigma_{1}$ , the image will be filtered. Conversely, if the similarity of the image is less than the threshold $\sigma_{1}$ , we would set the category exist in the image + +whose similarity is greater than the threshold $\sigma_{2}$ to "others" and save the related images and annotation. Finally, after manually filtering the categories for potential conflicts or duplicates, we get the OpenBench. + +# 4. Method + +# 4.1. Overview + +The overall pipeline of our proposed OVSNet is shown in Figure 3. We leverage the vision and text encoder of pretraind CLIP [46] to encode the input image and category names, respectively. Then, a segmentation decoder is trained to decode the visual features, generating masks for potential targets and their corresponding query embeddings. We then leverage the predicted masks to extract related CLIP features with Mask Pooling operation. To integrate heterogeneous knowledge spaces and improve the representation robustness, Gradient-Free Aggregation fuses the query embeddings with pooled CLIP features. The aggregated visual embeddings are combined with linguistic embeddings to perform visual-linguistic alignment and generate final output. During training, a Proxy Calibration strategy is utilized to expand training space by applying supervision on the convex combination of training semantics. + +# 4.2. Gradient-Free Aggregation + +Although CLIP [46] exhibits strong generalization ability, it optimizes for image-level visual-linguistic alignment during pretraining, which creates a great domain gap compared to mask-level alignment. As a result, directly using CLIP features obtained through mask pooling for mask classification may lead to suboptimal performance. The query embeddings generated by the segmentation decoder, on the other hand, have a strong region-level alignment prior owing to their suitability for the training space. However, they struggle to recognize novel semantics. + +Therefore, we propose to reorganize them to obtain more robust domain-adapted representations. An intuitive approach is to learn the intrinsic connection between them through self-attention or cross-attention mechanisms. However, we observe that such learning-based strategy tends to develop excessive reliance on query embeddings fitted to seen semantics during the training process while neglecting the utilization of more generalizable information provided by CLIP, ultimately affecting the model's ability to understand unlimited categories. Therefore, we propose to combine CLIP features and query embeddings via a gradient-free approach to avoid potential overfitting problems during the learning of the joint space. This is inspired by Random Walk algorithm [16, 56]. Specifically, there are multiple iterations in this process where two features learn collaboratively in shared embedding space until convergence. Suppose we have $N$ queries. For convenience, we omit the + +![](images/c7a64c0b7842a0c3ebf6f55d5287a827062903b0d7a95554faaefef99e4654c1.jpg) +Figure 3. Pipeline of our OVSNet. The input image and corresponding categories are first encoded by pretrained CLIP [46]. A segmentation decoder is trained to decode the visual features and generate potential masks as well as corresponding query embeddings. We then leverage the predicted masks to extract related CLIP features with Mask Pooling operation. To integrate different knowledge spaces and improve the representation robustness, Gradient-Free Aggregation fuses the query embeddings with CLIP features. The aggregated visual embeddings are combined with class embeddings to perform visual-linguistic alignment and generate final output. During training, we apply Proxy Calibration strategy to expand training space in a cost-free manner. + +query index and take a random query as the example. The corresponding query embeddings and CLIP features can be represented by $F_{Q}$ and $F_{C}$ . $F_{Q}$ and $F_{C}$ are defined as the initial state $F_{Q}^{0}$ and $F_{C}^{0}$ at the 0-th iteration. Then we formulate affinity $\mathcal{Z}$ by scaled dot product: $\mathcal{Z} = \lambda F_{C}^{0}(F_{Q}^{0})^{\top}$ , where $\lambda$ is the scaling factor. At $t$ -th iteration, the query embedding $F_{Q}^{t}$ is derived by the original embedding $F_{Q}^{0}$ and the updated CLIP feature $F_{C}^{t-1}$ from previous iteration, which can be formulated as: + +$$ +F _ {Q} ^ {t} = \omega \operatorname {N o r m} (\mathcal {Z}) ^ {\top} F _ {C} ^ {t - 1} + (1 - \omega) F _ {Q} ^ {0}, \tag {1} +$$ + +Subsequently, it is utilized to boost the CLIP feature of this iteration: + +$$ +F _ {C} ^ {t} = \omega \mathcal {Z} F _ {Q} ^ {t} + (1 - \omega) F _ {C} ^ {0}, \tag {2} +$$ + +where $\omega \in (0,1)$ is the factor that controls the degree of feature fusion. By combining the above equations we can obtain the overall iterative update formula of $F_{C}^{t}$ : + +$$ +F _ {C} ^ {t} = \left(\omega^ {2} A\right) ^ {t} F _ {C} ^ {0} + (1 - \omega) \sum_ {i = 0} ^ {t - 1} \left(\omega^ {2} A\right) ^ {i} \left(\omega \mathcal {Z} F _ {Q} ^ {0} + F _ {C} ^ {0}\right), \tag {3} +$$ + +where $A$ denotes $\mathcal{Z}\mathrm{Norm}(\mathcal{Z})^{\top}$ . Moreover, to avoid the potential issue of unexpected gradient and simplify the calculation process, we use an approximate inference function based on Neumann Series [43] when $t\to \infty$ . Then the update formula can be present as: + +$$ +F _ {C} ^ {\infty} = (1 - \omega) \left(I - \omega^ {2} A\right) ^ {- 1} \left(\omega \mathcal {Z} F _ {Q} ^ {0} + F _ {C} ^ {0}\right), \tag {4} +$$ + +where $I$ represents the identity matrix. With this gradient-free aggregation strategy, the model can recombine query embeddings with CLIP features to produce more robust region-level representations. + +# 4.3. Proxy Calibration + +A broader semantic training space tends to result in more generalized model representations. However, expanding the training space for OVS is costly due to the intensive prediction nature of segmentation tasks. To alleviate this problem, we propose a cost-free approach named Proxy Calibration (PC) to expand the training space and enhance the generalization ability of the model. It leverages the synthesis of proxy embeddings, which approximates novel semantics via convex combination between the in-vocabulary classes. + +Specifically, having the aggregated query embeddings $F_{Q}$ , we perform random convex combination of them during training to simulate embedding space for undefined classes, as shown by the black dotted line in Figure 3. This process can be formulated as: + +$$ +F _ {Q m n} ^ {\prime} = \alpha * F _ {Q m} + (1 - \alpha) * F _ {Q n} \tag {5} +$$ + +where $m, n \in \{1, 2, \dots, N\}$ . $\alpha$ is a random weight sampled from the distribution of $Beta(\gamma, \gamma)$ . $F'$ denotes the generated proxy embeddings. The same enhancement is performed for the corresponding CLIP feature $F_C$ and linguistic embeddings $F_T$ . Note that the text features here refer to the class representations of the ground truth regions corresponding to each query, which are obtained after Hungarian Matching [5]. Having the generated $F_Q'$ , $F_C'$ , and $F_T'$ , we perform distance supervision on them to promote region-level visual-linguistic alignment performance from the per + +Table 2. Performance comparison with state-of-the-art methods on existing benchmarks and the proposed OpenBench. ADE, PC, and VOC denote ADE20K [66], Pascal Context [44], and Pascal VOC [13], respectively. 171 and 133 indicate the number of training categories. * denotes the re-trained version according to the official code for fair comparison. + +
MethodTraining DatasetADE-150ADE-847PC-59PC-459VOCOpenBenchAverage Score
Base-level CLIP Backbone
SimSeg* [60]COCO-Stuff-17121.16.951.99.791.819.433.5
OVSeg [31]COCO-Stuff-17124.87.153.311.092.624.935.6
SAN [61]COCO-Stuff-17127.510.153.812.694.039.639.6
SCAN [37]COCO-Stuff-17130.810.858.413.297.038.741.5
CATSeg [10]COCO-Stuff-17131.812.057.519.094.636.141.8
SED [57]COCO-Stuff-17131.611.457.318.694.433.341.1
FCCLIP* [63]COCO-Panoptic-13331.113.554.812.893.240.340.9
MAFT [25]COCO-Panoptic-13329.110.153.512.890.037.538.8
MAFT+ [26]COCO-Panoptic-13334.613.857.516.295.443.743.5
OVSNet (Ours)COCO-Panoptic-13335.814.558.619.195.744.944.8
Large-level CLIP Backbone
SimSeg* [60]COCO-Stuff-17121.77.152.210.292.327.835.2
OVSeg [31]COCO-Stuff-17129.69.055.712.494.535.639.5
SAN [61]COCO-Stuff-17132.112.457.715.794.643.442.7
SCAN [37]COCO-Stuff-17133.514.059.316.397.149.444.9
SED [57]COCO-Stuff-17135.213.960.622.696.145.645.6
FCCLIP [63]COCO-Panoptic-13334.014.858.418.295.444.844.3
MAFT [25]COCO-Panoptic-13334.413.157.517.093.041.742.8
MAFT+ [26]COCO-Panoptic-13336.115.159.421.696.547.346.0
OVSNet (Ours)COCO-Panoptic-13337.116.262.023.596.948.647.4
+ +spective of proxy semantics. That is: + +$$ +\mathcal {L} _ {P Q} = 1 - \frac {\mathbf {F} _ {\mathbf {Q}} ^ {\prime} \cdot \mathbf {F} _ {\mathbf {T}} ^ {\prime}}{\| \mathbf {F} _ {\mathbf {Q}} ^ {\prime} \| _ {2} \| \mathbf {F} _ {\mathbf {T}} ^ {\prime} \| _ {2}}, \tag {6} +$$ + +$$ +\mathcal {L} _ {P C} = 1 - \frac {\mathbf {F} _ {\mathbf {C}} ^ {\prime} \cdot \mathbf {F} _ {\mathbf {T}} ^ {\prime}}{\| \mathbf {F} _ {\mathbf {C}} ^ {\prime} \| _ {2} \| \mathbf {F} _ {\mathbf {T}} ^ {\prime} \| _ {2}}, \tag {7} +$$ + +where $\mathcal{L}_{PQ}$ and $\mathcal{L}_{PC}$ denote the applied proxy supervision for query embeddings and pooled CLIP features, respectively. The final proxy loss is the sum of $\mathcal{L}_{PQ}$ and $\mathcal{L}_{PC}$ . By applying additional supervision in this way, the model is not limited to a specific training space and can learn more robust visual representations. + +# 5. Experiment + +# 5.1. Evaluation Benchmark and Metric + +To evaluate the effectiveness and generalization ability of our method on diverse scenarios, we conduct extensive experiments on the popular existing benchmarks, ADE20K150 [66], ADE20K847 [66], Pascal VOC [13], Pascal Context-59 [44], Pascal Context-459 [44] and our proposed OpenBench. ADE20K is a large-scale scene understanding benchmark, containing 20k training images, 2k validation images, and 3k testing images. There are two splits of this dataset. ADE20K-150 contains 150 semantic classes whereas ADE20K-847 has 847 classes. The images of both are the same. Pascal Context is an extension of Pascal VOC 2010, containing 5,005 validation images. We take the commonly used PC-59 and challenging PC-459 + +version for validation. Pascal VOC contains 11,185 training images and 1,449 validation images from 20 classes. We use the provided augmented annotations. Following previous works [12, 15, 60], we take the mean-intersection-over-union (mIoU) as the metric to compare our model with previous state-of-the-art methods. In ablation, we assume the use of base-level CLIP as the backbone for comparison. + +# 5.2. Implementation Details + +We utilize the Mask2Former [8] as the segmentation decoder and the implementation is based on detectron2 [54]. The model is trained with batch size of 8 and total iteration of $60\mathrm{k}$ . The base learning rate is 0.0001 with a polynomial schedule. Following previous works [26, 63], we take CLIP [46] of ConvNeXt as the backbone. The input image is resized and cropped to $1024\times 1024$ . For data augmentation, random flip and color augmentation are adopted. The weight decay of the model is 0.05. $\lambda$ and $\gamma$ of GFA and PC are set to 0.2 and 2 by default. The segmentation loss consists of dice loss and cross entropy loss. The classification loss is cross entropy loss. For the weights of the loss function, we set 5 and 2 for segmentation loss and classification loss, respectively. Other hyperparameters are the same as Mask2Former [9]. + +# 5.3. Main Results + +We compare our model with existing state-of-the-art approaches in Table 2. It can be seen that with proxy calibration expanding training semantics and the gradient-free + +Table 3. Ablation experiments about the proposed methods. GFA and PC indicate the Gradient-Free Aggregation and Proxy Calibration strategies, respectively. + +
MethodADE-150PC-459OpenBench
Baseline33.114.342.3
+ GFA34.7↑ (1.6)16.0↑ (1.7)43.7↑ (1.4)
+ PC33.9↑ (0.8)17.2↑ (2.9)44.3↑ (2.0)
+ Both35.8↑ (2.7)19.1↑ (4.8)44.9↑ (2.6)
+ +aggregation algorithm modeling the joint space, our model achieves the state-of-the-art performance on most popular benchmarks. The last column counts the average performance across multiple datasets, which shows that our method performs well for both in-vocabulary as well as out-vocabulary diverse scenarios. The relatively inferior performance of our method on VOC is due to the fact that the semantics of this dataset are almost contained in the training set, which tends to benefit models such as SCAN [37] that finetunes the CLIP backbone on training set. Besides, the greater divergence of OpenBench from training semantics does not imply that it is totally more challenging than other benchmarks, because the final performance is influenced by various factors such as image content complexity, image clarity, and the number of given categories. More discussion can be seen in the supplementary materials. + +# 5.4. Ablation Study + +Component Analysis Table 3 shows the ablations about the proposed Gradient-Free Aggregation (GFA) and Proxy Calibration (PC) strategy. We evaluate their effectiveness on both existing benchmarks and our proposed OpenBench. It can be seen that with GFA fusing CLIP feature and query embeddings, the model's performance is greatly improved on both seen and unseen semantics. Besides, PC can facilitate the model's ability to understand various scenarios by cost-free expansion of the training semantic space. When applying both, the model achieves the best performance. + +Embedding Distribution Analysis To demonstrate that our proposed OpenBench is more open-vocabulary compared to the existing test sets, we perform t-SNE visualization of the category embeddings of the training set, the existing test sets, and our OpenBench in Figure 4. Besides, we also visualize the proxy embeddings generated by PC. We can see that the existing test sets (green dots) are overall similar to the training semantics (blue dots). Our OpenBench (orange dots), on the other hand, shows significant differences. In addition, the proxy embeddings (red dots) generated by PC expand the training set and thus help to improve the robustness and generalization of the model. + +Analysis about Inference Class Number In this part we explore an interesting phenomenon: why the OVS models perform much worse on ADE-847 and PC-459 than on other test sets? Intuitively, this is because these datasets + +![](images/ae6d8df7e54015ea14855309442f3fc89d438b782f13954996469bbf9ad68bb1.jpg) +Figure 4. Visualization of category embeddings across the training space, existing evaluation space, our OpenBench space, and the generated proxy space. + +![](images/2f9ffae676bc6c1a4e9b664692e25a8682fe827cae87d8c2d18c096cd7d52636.jpg) +Figure 5. Impact of the number of inference categories on model performance. As the number of irrelevant categories increases, the model's performance on the same image degrades. + +contain more novel categories. However, we find that another factor significantly impacts model performance, that is the number of categories fed into the model during inference. Let's start with the conclusion: for the same model and test image, the model performance is worse when there are more inference categories given. In other words, the model may be able to make correct predictions when only a small number of categories are given. However, when more categories are prompted, the model's output will be incorrect. Here we take two typical methods, SAN [61] and MAFT-Plus [26], as examples and show how their performance on the ADE-150 dataset changes as the number of categories increases. As shown in Figure 5, with the number of given irrelevant categories increases, the models of different technology routes show a consistent decline. + +Internal Analysis of Proposed Strategies In Table 4 we conduct more experiments about the proposed methods. Specifically, we first compare our gradient-free aggregation designs with the typical learning-based feature fusion approaches in (a). Results show that vanilla attention performs inferior than our GFA, especially on the more open-vocabulary set OpenBench (degraded by more than + +![](images/eac635c0d5e9fcf77f4ef38cc457d980ef13348b47609cec0fab18099c59cdc6.jpg) +(a) Seen: building treefence bicycle + +![](images/6c46b639d0000c7d046b194c73a0b7df5d90e39b16b95a00b9df546125c05e8e.jpg) + +![](images/8ebda229eee540b5321463bc44485c601e2283e02a91ea233c57af6c629916de.jpg) +Unseen: sidewalk streetlight + +![](images/e712fefb1e0a5711f184bc4a71101150a4f76838d596280e8c0aa697dcc1b676.jpg) +(c) Seen: ceiling desk chair + +![](images/217dfb670f55b5179e12cbf69a31b51589f5cf93124a108cb2ddf12dd600fe0f.jpg) +Unseen: computer chandelier +→ opera house + +![](images/16aa424d28b7f6a7aae259f3fb3bc024884a8c8eeae3cbf2fa86affbf7b304fa.jpg) +(e) building + +![](images/3ee052e628fdbc2b67d490ead5016756613b7ad75c6edcba61f0a542d94a3e42.jpg) +(b) Seen: sea boat mountain + +![](images/a99f3606ca78fda54e006e50ca67b981ed5b989ea0fedd8d3833088e8486c955.jpg) +Unseen: pier pole + +![](images/9fbc96ee036cac8c76f0d7f9c34d4a1bbe6eef3a0c55f03c5b3e6070cf778e20.jpg) +(d) Seen: plant rock person + +![](images/fa1a4772f0f95bc1b6a5bbc312259b27de77c8c33cc45d533b8dd2bc5cd558f7.jpg) +Unseen: waterfall + +![](images/0c4788019177927f18722285b24801a6d2619541eebe6fe75b0369a5694e23a3.jpg) +(f) +Figure 6. Visualization of segmentation results. (a) - (d) demonstrate the successful segmentation cases of our method for seen and unseen categories. (e) and (f) display the adaptability for flexible text query. Best viewed in color. + +![](images/08cc198f1b540bd4c57fca47bdb0e53b7258f6b724f49f6953e5716f86ab39a2.jpg) +→ ferrari + +Table 4. Ablation experiments within each modules. (a) shows the comparison of our gradient-free algorithm with learning-based method, e.g., cross attention. (b) demonstrates the effect of $\gamma$ of Beta distribution in the proxy learning strategy. + +
ADE-150PC-459OpenBench
(a) Comparisons of GFA and learning-based strategy
Self Attn34.014.740.6
Cross Attn34.214.841.2
GFA34.716.043.7
(b) Effect of sampling distribution of PC
γ = 133.114.342.3
γ = 0.533.6↑ (0.5)16.1↑ (1.8)43.7↑ (1.4)
γ = 233.9↑ (0.8)17.2↑ (2.9)44.3↑ (2.0)
+ +2.5 mIoU). This is because learning-based methods tend to develop a reliance on the segmentation query embeddings during training, while overlooking the CLIP features that inherently exhibit stronger generalization capabilities, thereby introducing biases. + +We also experiment with the influence of different $\gamma$ in Beta distribution to the performance of proxy calibration. The results are shown in Table 4 (b), where $\gamma = 1$ is equivalent to do not utilize PC. We find that different $\gamma$ benefit the model performance while the performance is better when the intermediate probability density is high ( $\gamma = 2$ ). + +# 5.5. Visualization + +Figure 6 shows some segmentation cases of our OVSNet. It can be seen that our method achieves excellent segmentation performance on various scenarios. Specifically, (a) - (d) + +demonstrate the excellent segmentation of our method for seen and unseen categories. (e) and (f) display the adaptability for flexible text query, e.g., change "building" to "opera house" and "car" to "ferrari". + +# 6. Conclusion + +In this paper, we identify that the existing open-vocabulary segmentation evaluation sets suffer from excessive semantic similarity to the training set, which limits their ability to comprehensively assess the model's open vocabulary understanding capabilities. To address this, we propose a new benchmark, OpenBench, which features significant differences from the training set and includes more fine-grained categories. Comparison on OpenBench further demonstrates that maintaining the generalization space of CLIP is crucial for OVS. Additionally, we introduce OVSNet, a method that leverages proxy calibration for cost-free augmentation of the training space and gradient-free aggregation for fusing heterogeneous features. Experiments show that the proposed OVSNet achieves excellent performance on both existing datasets and the newly proposed OpenBench. We hope that our benchmark and model design could inspire further researches in this area. + +Acknowledgements. This work was supported by Shenzhen Science and Technology Program under Grant CJGJZD20220517142402006 and Guangdong Natural Science Funds for Distinguished Young Scholar (No. 2025B1515020012). + +# References + +[1] Sule Bai, Yong Liu, Yifei Han, Haoji Zhang, and Yansong Tang. Self-calibrated clip for training-free open-vocabulary segmentation. arXiv preprint arXiv:2411.15869, 2024. 2 +[2] Sule Bai, Mingxing Li, Yong Liu, Jing Tang, Haoji Zhang, Lei Sun, Xiangxiang Chu, and Yansong Tang. Univg-r1: Reasoning guided universal visual grounding with reinforcement learning. arXiv preprint arXiv:2505.14231, 2025. 1 +[3] Gabriel J. Brostow, Jamie Shotton, Julien Fauqueur, and Roberto Cipolla. Segmentation and recognition using structure from motion point clouds. In ECCV, pages 44-57, 2008. 4 +[4] Maxime Bucher, Tuan-Hung Vu, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. In NeurlPS, 2019. 2 +[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 5 +[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. TPAMI, 2018. 1 +[7] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. UNITER: learning universal image-text representations. arXiv preprint arXiv:1909.11740, 2019. 3 +[8] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G. Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021. 6 +[9] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 6 +[10] Seokju Cho, Heeseong Shin, Sunghwan Hong, Anurag Arnab, Paul Hongsuck Seo, and Seungryong Kim. Cat-seg: Cost aggregation for open-vocabulary semantic segmentation. In CVPR, pages 4113-4123, 2024. 1, 2, 3, 6 +[11] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 3 +[12] Jian Ding, Nan Xue, Gui-Song Xia, and Dengxin Dai. Decoupling zero-shot semantic segmentation. In CVPR, 2022. 1, 2, 3, 6 +[13] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes challenge: A retrospective. IJCV, 111:98-136, 2015. 2, 3, 6 +[14] Shanghua Gao, Zhong-Yu Li, Ming-Hsuan Yang, Ming-Ming Cheng, Junwei Han, and Philip Torr. Large-scale unsupervised semantic segmentation. TPAMI, 45(6):7457-7476, 2022. 4 +[15] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Open-vocabulary image segmentation. arXiv preprint arXiv:2112.12143, 2021. 1, 2, 6 + +[16] Leo Grady. Random walks for image segmentation. IEEE TPAMI, 28(11):1768-1783, 2006. 2, 4 +[17] Kunyang Han, Yong Liu, Jun Hao Liew, Henghui Ding, Jiajun Liu, Yitong Wang, Yansong Tang, Yujiu Yang, Jiashi Feng, Yao Zhao, et al. Global knowledge calibration for fast open-vocabulary segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 797-807, 2023. 1, 2 +[18] Shuting He, Henghui Ding, and Wei Jiang. Primitive generation and semantic-related alignment for universal zero-shot segmentation. In CVPR, 2023. 1 +[19] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1 +[20] Jiaqi Huang, Zunnan Xu, Ting Liu, Yong Liu, Haonan Han, Kehong Yuan, and Xiu Li. Densely connected parameter-efficient tuning for referring image segmentation. arXiv preprint arXiv:2501.08580, 2025. 1 +[21] Xiaoke Huang, Jianfeng Wang, Yansong Tang, Zheng Zhang, Han Hu, Jiwen Lu, Lijuan Wang, and Zicheng Liu. Segment and caption anything. arXiv preprint arXiv:2312.00869, 2023. 1 +[22] Joonhyun Jeong, Geondo Park, Jayeon Yoo, Hyungsik Jung, and Heesu Kim. Proxydet: Synthesizing proxy novel classes via classwise mixup for open-vocabulary object detection. In AAAI, pages 2462-2470, 2024. 3 +[23] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 3 +[24] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916, 2021. 1 +[25] Siyu Jiao, Yunchao Wei, Yaowei Wang, Yao Zhao, and Humphrey Shi. Learning mask-aware clip representations for zero-shot segmentation. arXiv preprint arXiv:2310.00240, 2023. 1, 6 +[26] Siyu Jiao, Hongguang Zhu, Jiannan Huang, Yao Zhao, Yunchao Wei, and Humphrey Shi. Collaborative vision-text representation optimizing for open-vocabulary segmentation. In ECCV, pages 399-416, 2025. 6, 7 +[27] Boyi Li, Kilian Q. Weinberger, Serge J. Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. In ICLR, 2022. 1, 3 +[28] Gen Li, Nan Duan, Yuejian Fang, Ming Gong, and Daxin Jiang. Unicoder-vl: A universal encoder for vision and language by cross-modal pre-training. In AAAI, 2020. 3 +[29] Wanhua Li, Xiaoke Huang, Zheng Zhu, Yansong Tang, Xiu Li, Jie Zhou, and Jiwen Lu. Ordinalclip: Learning rank prompts for language-guided ordinal regression. NeurIPS, pages 35313-35325, 2022. 3 +[30] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu + +Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-semantics aligned pre-training for vision-language tasks. In ECCV, 2020. 3 +[31] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yi-nan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted CLIP. arXiv preprint arXiv:2210.04150, 2022. 1, 2, 6 +[32] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In ECCV, 2014. 4 +[33] Yong Liu, Ran Yu, Xinyuan Zhao, and Yujiu Yang. Quality-aware and selective prior enhancement memory network for video object segmentation. In CVPR Workshop, 2021. 1 +[34] Yong Liu, Ran Yu, Jiahao Wang, Xinyuan Zhao, Yitong Wang, Yansong Tang, and Yujiu Yang. Global spectral filter memory network for video object segmentation. In ECCV, pages 648-665, 2022. +[35] Yong Liu, Ran Yu, Fei Yin, Xinyuan Zhao, Wei Zhao, Weihao Xia, and Yujiu Yang. Learning quality-aware dynamic memory for video object segmentation. In ECCV, pages 468-486, 2022. +[36] Yong Liu, Cairong Zhang, Yitong Wang, Jiahao Wang, Yujiu Yang, and Yansong Tang. Universal segmentation at arbitrary granularity with language instruction. arXiv preprint arXiv:2312.01623, 2023. 1 +[37] Yong Liu, Sule Bai, Guanbin Li, Yitong Wang, and Yansong Tang. Open-vocabulary segmentation with semantic-assisted calibration. In CVPR, pages 3491-3500, 2024. 1, 4, 6, 7 +[38] Yong Liu, Ran Yu, Fei Yin, Xinyuan Zhao, Wei Zhao, Weihao Xia, Jiahao Wang, Yitong Wang, Yansong Tang, and Yujiu Yang. Learning high-quality dynamic memory for video object segmentation. IEEE TPAMI, 2025. 1 +[39] Guanxing Lu, Ziwei Wang, Changliu Liu, Jiwen Lu, and Yansong Tang. Thinkbot: Embodied instruction following with thought chain reasoning. arXiv preprint arXiv:2312.07062, 2023. 1 +[40] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In NeurIPS, 2019. 3 +[41] Zhuoyan Luo, Yicheng Xiao, Yong Liu, Shuyan Li, Yitong Wang, Yansong Tang, Xiu Li, and Yujiu Yang. Soc: Semantic-assisted object cluster for referring video object segmentation. arXiv preprint arXiv:2305.17011, 2023. 2 +[42] Zhuoyan Luo, Yinghao Wu, Yong Liu, Yicheng Xiao, Xiaoping Zhang, and Yujiu Yang. Hdc: Hierarchical semantic decoding with counting assistance for generalized referring expression segmentation. arXiv preprint arXiv:2405.15658, 2024. 1 +[43] Carl D Meyer and Ian Stewart. Matrix analysis and applied linear algebra. SIAM, 2023. 5 +[44] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan L. Yuille. The role of context for object detection and semantic segmentation in the wild. In CVPR, 2014. 2, 3, 4, 6 + +[45] Mengxue Qu, Yu Wu, Wu Liu, Qiqi Gong, Xiaodan Liang, Olga Russakovsky, Yao Zhao, and Yunchao Wei. Siri: A simple selective retraining mechanism for transformer-based visual grounding. In ECCV, 2022. 3 +[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 4, 5, 6 +[47] Yiqin Wang, Haoji Zhang, Yansong Tang, Yong Liu, Jiashi Feng, Jifeng Dai, and Xiaojie Jin. Hierarchical memory for long video qa. arXiv preprint arXiv:2407.00603, 2024. 3 +[48] Yulin Wang, Haoji Zhang, Yang Yue, Shiji Song, Chao Deng, Junlan Feng, and Gao Huang. Uni-adafocus: Spatial-temporal dynamic computation for video recognition. IEEE TPAMI, 2024. 3 +[49] Yuji Wang, Jingchen Ni, Yong Liu, Chun Yuan, and Yansong Tang. Iterprime: Zero-shot referring image segmentation with iterative grad-cam refinement and primary word emphasis. In AAAI, pages 8159-8168, 2025. 1 +[50] Yuji Wang, Haoran Xu, Yong Liu, Jiaze Li, and Yansong Tang. Sam2-love: Segment anything model 2 in language-aided audio-visual scenes. In CVPR, pages 28932-28941, 2025. +[51] Cong Wei, Yujie Zhong, Haoxian Tan, Yong Liu, Zheng Zhao, Jie Hu, and Yujiu Yang. Hyperseg: Towards universal visual segmentation with large language model. arXiv preprint arXiv:2411.17606, 2024. +[52] Cong Wei, Yujie Zhong, Haoxian Tan, Yingsen Zeng, Yong Liu, Zheng Zhao, and Yujiu Yang. Instructseg: Unifying instructed visual segmentation with multi-modal large language models. arXiv preprint arXiv:2412.14006, 2024. 1 +[53] Xiongwei Wu, Xin Fu, Ying Liu, Ee-Peng Lim, Steven CH Hoi, and Qianru Sun. A large-scale benchmark for food image segmentation. In ACM MM, 2021. 4 +[54] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.6 +[55] Yongqin Xian, Subhabrata Choudhury, Yang He, Bernt Schiele, and Zeynep Akata. Semantic projection network for zero- and few-label semantic segmentation. In CVPR, 2019. 1, 2 +[56] Yicheng Xiao, Zhuoyan Luo, Yong Liu, Yue Ma, Hengwei Bian, Yatai Ji, Yujiu Yang, and Xiu Li. Bridging the gap: A unified video comprehension framework for moment retrieval and highlight detection. In CVPR, pages 18709-18719, 2024. 4 +[57] Bin Xie, Jiale Cao, Jin Xie, Fahad Shahbaz Khan, and Yanwei Pang. Sed: A simple encoder-decoder for open-vocabulary semantic segmentation. In CVPR, pages 3426-3436, 2024. 1, 2, 3, 6 +[58] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In NIPS, 2021. 1 + +[59] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, pages 2955-2966, 2023. 2 +[60] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A simple baseline for zero-shot semantic segmentation with pre-trained vision-language model. arXiv preprint arXiv:2112.14757, 2021. 1, 2, 3, 6 +[61] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xi-ang Bai. Side adapter network for open-vocabulary semantic segmentation. In CVPR, pages 2945–2954, 2023. 1, 2, 6, 7 +[62] Xin Xu, Tianyi Xiong, Zheng Ding, and Zhuowen Tu. Masqclip for open-vocabulary universal image segmentation. In ICCV, pages 887-898, 2023. 1 +[63] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. NeurIPS, pages 32215-32234, 2023. 1, 2, 6 +[64] Hui Zhang and Henghui Ding. Prototypical matching and open set rejection for zero-shot semantic segmentation. In ICCV, 2021. 1 +[65] Haoji Zhang, Yiqin Wang, Yansong Tang, Yong Liu, Jiashi Feng, Jifeng Dai, and Xiaojie Jin. Flash-vstream: Memory-based real-time understanding for long video streams. arXiv preprint arXiv:2406.08085, 2024. 3 +[66] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ADE20K dataset. In CVPR, 2017. 2, 3, 4, 6 \ No newline at end of file diff --git a/2025/Stepping Out of Similar Semantic Space for Open-Vocabulary Segmentation/images.zip b/2025/Stepping Out of Similar Semantic Space for Open-Vocabulary Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9d09736fb86cc2167c511e78d7ccde0ddf85fa3c --- /dev/null +++ b/2025/Stepping Out of Similar Semantic Space for Open-Vocabulary Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5e740ed505e42526283f653450f32c27bfc61208ff9a0acde41ba128e178ff70 +size 741399 diff --git a/2025/Stepping Out of Similar Semantic Space for Open-Vocabulary Segmentation/layout.json b/2025/Stepping Out of Similar Semantic Space for Open-Vocabulary Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..26c98a526d1f6c65cb28256051acdc79bcb593d0 --- /dev/null +++ b/2025/Stepping Out of Similar Semantic Space for Open-Vocabulary Segmentation/layout.json @@ -0,0 +1,8493 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 103, + 541, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 103, + 541, + 121 + ], + "spans": [ + { + "bbox": [ + 70, + 103, + 541, + 121 + ], + "type": "text", + "content": "Stepping Out of Similar Semantic Space for Open-Vocabulary Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 87, + 141, + 527, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 141, + 527, + 186 + ], + "spans": [ + { + "bbox": [ + 87, + 141, + 527, + 186 + ], + "type": "text", + "content": "Yong Liu" + }, + { + "bbox": [ + 87, + 141, + 527, + 186 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 87, + 141, + 527, + 186 + ], + "type": "text", + "content": ", Song-Li Wu" + }, + { + "bbox": [ + 87, + 141, + 527, + 186 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 87, + 141, + 527, + 186 + ], + "type": "text", + "content": ", Sule Bai" + }, + { + "bbox": [ + 87, + 141, + 527, + 186 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 87, + 141, + 527, + 186 + ], + "type": "text", + "content": ", Jiahao Wang" + }, + { + "bbox": [ + 87, + 141, + 527, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 87, + 141, + 527, + 186 + ], + "type": "text", + "content": ", Yitong Wang" + }, + { + "bbox": [ + 87, + 141, + 527, + 186 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 87, + 141, + 527, + 186 + ], + "type": "text", + "content": ", Yansong Tang" + }, + { + "bbox": [ + 87, + 141, + 527, + 186 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 87, + 141, + 527, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 87, + 141, + 527, + 186 + ], + "type": "text", + "content": "Tsinghua Shenzhen International Graduate School " + }, + { + "bbox": [ + 87, + 141, + 527, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 87, + 141, + 527, + 186 + ], + "type": "text", + "content": "Tsinghua University \n" + }, + { + "bbox": [ + 87, + 141, + 527, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 87, + 141, + 527, + 186 + ], + "type": "text", + "content": "The University of Hong Kong " + }, + { + "bbox": [ + 87, + 141, + 527, + 186 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 87, + 141, + 527, + 186 + ], + "type": "text", + "content": "ByteDance Inc." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 129, + 187, + 477, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 187, + 477, + 199 + ], + "spans": [ + { + "bbox": [ + 129, + 187, + 477, + 199 + ], + "type": "text", + "content": "liuyong23@mails.tsinghua.edu.cn, tang.yansong@sz.tsinghua.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 239, + 200, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 239, + 200, + 251 + ], + "spans": [ + { + "bbox": [ + 151, + 239, + 200, + 251 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 265, + 296, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 296, + 553 + ], + "type": "text", + "content": "Open-vocabulary segmentation aims to achieve segmentation of arbitrary categories given unlimited text inputs as guidance. To achieve this, recent works have focused on developing various technical routes to exploit the potential of large-scale pre-trained vision-language models and have made significant progress on existing benchmarks. However, we find that existing test sets are limited in measuring the models' comprehension of \"open-vocabulary\" concepts, as their semantic space closely resembles the training space, even with many overlapping categories. To this end, we present a new benchmark named OpenBench that differs significantly from the training semantics. It is designed to better assess the model's ability to understand and segment a wide range of real-world concepts. When testing existing methods on OpenBench, we find that their performance diverges from the conclusions drawn on existing test sets. In addition, we propose a method named OVSNet to improve the segmentation performance for diverse and open scenarios. Through elaborate fusion of heterogeneous features and cost-free expansion of the training space, OVSNet achieves state-of-the-art results on both existing datasets and our proposed OpenBench. Corresponding analysis demonstrates the rationality and effectiveness of our proposed benchmark and method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 567, + 135, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 135, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 135, + 578 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 587, + 295, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 295, + 684 + ], + "type": "text", + "content": "Semantic segmentation is one of the most fundamental tasks in computer vision, which targets at assigning semantic category to pixels in an image. Despite achieving excellent progress in recent years [2, 6, 20, 21, 33-36, 38, 42, 49-52, 58], traditional semantic segmentation approaches rely on predefined categories sets and tend to falter when encountering categories absent during the training phase, significantly impeding their real-world application. Such chal" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 226, + 446, + 329 + ], + "blocks": [ + { + "bbox": [ + 315, + 226, + 446, + 329 + ], + "lines": [ + { + "bbox": [ + 315, + 226, + 446, + 329 + ], + "spans": [ + { + "bbox": [ + 315, + 226, + 446, + 329 + ], + "type": "image", + "image_path": "272a77adb469063800da3e373da97c34de40af323c96325c224e0ec6c28273b3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 333, + 392, + 342 + ], + "lines": [ + { + "bbox": [ + 381, + 333, + 392, + 342 + ], + "spans": [ + { + "bbox": [ + 381, + 333, + 392, + 342 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 447, + 227, + 553, + 327 + ], + "blocks": [ + { + "bbox": [ + 447, + 227, + 553, + 327 + ], + "lines": [ + { + "bbox": [ + 447, + 227, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 447, + 227, + 553, + 327 + ], + "type": "image", + "image_path": "b8593b55c6cc94d7cfe44a2bb7cad13e6c9440d3dfe8f2dca35986c7e7469c92.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 482, + 332, + 493, + 342 + ], + "lines": [ + { + "bbox": [ + 482, + 332, + 493, + 342 + ], + "spans": [ + { + "bbox": [ + 482, + 332, + 493, + 342 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 343, + 555, + 387 + ], + "lines": [ + { + "bbox": [ + 313, + 343, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 555, + 387 + ], + "type": "text", + "content": "Figure 1. (a) shows the similarity distribution of the existing test set (blue) and our OpenBench (orange) with the training semantic space. (b) illustrates the performance gap between the methods finetuning [10] and frozen [61] CLIP on different benchmarks." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 399, + 554, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 554, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 554, + 472 + ], + "type": "text", + "content": "lenge has inspired the exploration of Open-Vocabulary Segmentation (OVS) task [12, 15, 17, 18, 27, 37, 55, 64]. Different from traditional closed-set segmentation, OVS methods aim to segment arbitrary categories under the guidance of given text inputs, which has many potential applications such as auto-driving and human-robot interaction [19, 39]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "content": "However, it is challenging to identify unseen categories without the intervention of external knowledge. Therefore, an intuitive idea is to introduce large-scale vision-language model [24, 46] trained with numerous sources to extend the semantic space of segmentation models. Following this idea, existing OVS methods [10, 31, 37, 57, 61, 62] concentrate on exploring how to exploit the pretrained knowledge prior more effectively and derive several technical routes. Early methods [12, 31, 60] adopt a two-stage pipeline. This approach first generates class-agnostic mask proposals with segmentation models, following which the pre-trained CLIP [46] serves as an additional classifier to execute region-level open-vocabulary classification. However, this strategy incurs heavy computation overhead, as it requires repeated forward passes of the CLIP vision encoder for each mask. This can be prohibitively expensive for real-world applications. To address this problem, researchers [25, 61–63] propose to unify classification and segmentation within a shared space, i.e., integrating frozen CLIP features into the segmenter and extracting CLIP embeddings from potential" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 132, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 132, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 132, + 703 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22664" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "type": "text", + "content": "target regions to enable simultaneous classification and segmentation. These methods achieve remarkable performance in a more efficient one-stage pipeline. Subsequently, some works [10, 57] discover that finetuning the CLIP encoder could further enhance the model's performance on existing test sets. They develop an early fusion paradigm based on cost map and achieve outstanding segmentation results. However, this performance boost is somewhat counterintuitive, as fine-tuning CLIP typically adapts it to specific training semantic spaces, which could reduce its original generalization ability. Yet, the ability of generalized visual-linguistic alignment is precisely what the open-vocabulary task requires. So why does fine-tuning CLIP appear to benefit the performance of OVS?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 242, + 296, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 242, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 56, + 242, + 296, + 433 + ], + "type": "text", + "content": "This question drives us to revisit the existing evaluation benchmarks of the cross-dataset evaluation paradigm. We count the maximum similarity of each category in test set to the training semantics. Figure 1 (a) illustrates the similarity distribution of all test categories to the training space (blue color). Results indicate that the existing test data is, in fact, highly similar to the training semantic space. Specifically, even the most challenging ADE-847 [66] and PC-459 [44] datasets have an average similarity to the training classes of 0.79 and 0.83. PC-59 [44] and VOC [13] datasets even have a similarity to the training semantics of 0.95 and 0.97. This explains why finetuning CLIP on training set can lead to further performance gains on existing benchmarks. However, it somewhat conflicts with the purpose of the open-vocabulary setting, which aims to achieve segmentation of unlimited categories in the real world." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 436, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 295, + 616 + ], + "type": "text", + "content": "To address the above problem and provide a more comprehensive measure of OVS models, we propose a new evaluation benchmark named OpenBench, which is characterized by significant semantic differences from the training data. Figure 1 (a) also demonstrates the similarity distribution of our OpenBench (orange color). It can be seen that our benchmark contains more novel classes compared to the existing test sets. We evaluate the performance of approaches across different technical routes on our benchmark, and the findings diverge from those obtained on existing test sets. As shown in Figure 1 (b), while finetuning CLIP yields steady performance gains on existing datasets, it shows a significant drop on our benchmark. This outcome also supports our motivation for proposing a benchmark that diverges substantially from the training space." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "In addition, we also present a framework named OVSNet to improve the segmentation performance for diverse and open scenarios. In OVSNet, we adopt a Proxy Calibration (PC) training strategy to enhance the robustness of model representations by broadening the training space with no additional cost. This strategy leverages the synthesis of proxy embeddings, which approximate unseen semantics through convex combinations of in-vocabulary classes. Addition" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": "ally, inspired by random walk [16], we propose a gradient-free fusion of CLIP embeddings and segmentation decoder features to model a robust joint space and mitigate potential overfitting issues related to training semantics. With the above proposed designs calibrating model knowledge from the aspects of training space and features separately, our OVSNet demonstrates strong performance across a wide range of scenarios, both on existing test sets similar to the training semantics and on our proposed novel OpenBench. Related ablations also demonstrate the effectiveness and soundness of our proposed strategies." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 326, + 204, + 525, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 204, + 525, + 215 + ], + "spans": [ + { + "bbox": [ + 326, + 204, + 525, + 215 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 216, + 553, + 407 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 216, + 553, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 216, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 314, + 216, + 553, + 287 + ], + "type": "text", + "content": "- We present a new benchmark named OpenBench for OVS, which is characterized by significant semantic differences from the training space. While using it to evaluate the comprehension of OVS methods in more open scenarios, we reach a different conclusion from that of the existing test sets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 288, + 553, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 288, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 315, + 288, + 553, + 358 + ], + "type": "text", + "content": "- We propose OVSNet to improve the segmentation performance on diverse and open scenarios. It adopts a Proxy Calibration training strategy to broaden training space without additional cost. Besides, a gradient-free algorithm is leveraged to fuse CLIP embeddings and segmentation features for modeling robust joint space." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 360, + 553, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 360, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 315, + 360, + 553, + 407 + ], + "type": "text", + "content": "- Our method achieves state-of-the-art performance on existing benchmarks and the proposed OpenBench. Extensive experiments are conducted to prove the effectiveness and soundness of the proposed algorithms." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 418, + 400, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 418, + 400, + 430 + ], + "spans": [ + { + "bbox": [ + 314, + 418, + 400, + 430 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "text", + "content": "Open-Vocabulary Segmentation. This task aims to segment an image and identify regions with arbitrary given text queries [1, 4, 15, 41, 55]. Pioneering work [55] replaces the output convolution layer by computing the similarity between visual features and linguistic embeddings, which has become common practice. Later, some methods [12, 15, 31, 59, 60] propose a two-stage pipeline: the model first generates class-agnostic mask proposals, then a pretrained CLIP [46] is utilized to perform sub-image classification by cropping and masking corresponding regions. With the combination of both in-vocabulary and out-vocabulary classification, these methods obtain great performance. However, this strategy incurs heavy computation overhead, as it requires repeated forward passes of the CLIP vision encoder for each mask. This can be prohibitively expensive for real-world applications. To address this problem, subsequent researchers propose to unify classification and segmentation within a shared space. Among them, SAN [61] designs a side-adapter network to leverage CLIP features for reorganizing segmentation and classification. GKC [17] presents text-guided knowledge distillation strategy to transfer CLIP knowledge to specific classification layer. FCCLIP [63] proposes to leverage" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22665" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 550, + 235 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 550, + 235 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 550, + 235 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 550, + 235 + ], + "type": "image", + "image_path": "8709d79e14e101d89d52c4c2862f7e91ac20ccbf8812adc9082fc19c1e860acc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 236, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 236, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 555, + 270 + ], + "type": "text", + "content": "Figure 2. Illustration of the samples in the proposed OpenBench. Compared to existing test sets, our OpenBench has semantic categories that differ more from the training space, which helps to provide a more comprehensive measure of the model's ability to perform open-vocabulary segmentation. The corresponding category name is labeled in the lower right corner of each image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 306, + 550, + 415 + ], + "blocks": [ + { + "bbox": [ + 55, + 281, + 555, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 281, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 555, + 304 + ], + "type": "text", + "content": "Table 1. Comparison of the statistics of our proposed OpenBench with existing datasets. 'Cls', 'Img', and 'Sim' indicate class, image, and similarity, respectively. ADE, PC, and VOC denote ADE20K [66], Pascal Context [44], and Pascal VOC [13]" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 306, + 550, + 415 + ], + "lines": [ + { + "bbox": [ + 56, + 306, + 550, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 306, + 550, + 415 + ], + "type": "table", + "html": "
DatasetCls Num.Img Num.Mean Sim.Median Sim.Min Sim.Max Sim.Fine GranularitySemantic Duplication
VOC [13]2014490.97561.00000.84491.0000××
PC-59 [44]5951050.94811.00000.76581.0000××
ADE-150 [66]15020000.80860.82280.19601.0000×✓
Cityscapes [11]331000.84880.83700.67461.0000××
PC-459 [44]45951050.83740.81240.67181.0000✓✓
ADE-847 [66]84720000.79260.79130.19601.0000✓✓
OpenBench (Ours)28660560.61420.64520.26080.7947✓×
", + "image_path": "0fb306c615cc2275c9e3d5422476faacc1569d96d62de2d35193f5136e179158.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 426, + 295, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 474 + ], + "type": "text", + "content": "the local-awareness advantage of convolution backbones. Some works [10, 57] then discover that finetuning CLIP encoder could further enhance the model's performance on existing test sets and develop effective one-stage framework." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 475, + 295, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 295, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 295, + 571 + ], + "type": "text", + "content": "However, we find that existing test sets are highly similar to the training set, which limits their ability to effectively evaluate the model's comprehension of the open vocabulary semantics. Therefore, in this paper we propose a new benchmark OpenBench that is characterized by significant semantic differences from the training data. We also propose a method named OVSNet to improve the segmentation performance for diverse and open scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "Vision-Language Pre-training. Vision-language pretraining aims to learn a unified visual-linguistic representation space. Early approaches [7, 28, 30, 40, 45, 47, 48, 65], constrained by small-scale datasets, struggle to achieve strong performance and require fine-tuning for downstream tasks. However, with the advent of large-scale web data, recent works [23, 46] have demonstrated the benefits of leveraging such data to train a more robust multi-modal representation space. Among these, CLIP [46], the most widely used vision-language model, employs contrastive learning to link images with their corresponding captions," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 426, + 555, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 475 + ], + "type": "text", + "content": "achieving impressive cross-modal alignment. Building upon previous works [12, 22, 27, 29, 60], we also utilize the well-aligned and generalized space of CLIP to enhance open-vocabulary segmentation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 485, + 389, + 497 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 389, + 497 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 389, + 497 + ], + "type": "text", + "content": "3. Benchmark" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 505, + 434, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 434, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 434, + 516 + ], + "type": "text", + "content": "3.1. Dataset Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 521, + 555, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 654 + ], + "type": "text", + "content": "In Table 1, we present the statistics comparison of our proposed OpenBench with existing popular open-vocabulary segmentation datasets, including ADE-150 [66], ADE-847 [66], PC-59 [44], PC-459 [44], VOC [13], and Cityscapes [11]. Specifically, our OpenBench contains 286 categories, of which the maximum similarity to the training categories is 0.7947, the minimum is 0.2608, and the average similarity is 0.6142. It can be observed that, compared to existing datasets, our OpenBench exhibits a greater disparity from the training set, which better evaluates the model's segmentation ability for arbitrary open categories." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "Besides, our OpenBench contains fine-grained categories, e.g., \"oystercatcher\", rather than just generic broad category such as \"bird\" (corresponds to the \"Fine Granularity\" characterization in Table 1). This is important for evaluating whether the OVS models preserve the general" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22666" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 294, + 322 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 294, + 322 + ], + "type": "text", + "content": "ization capability of CLIP [46], which is inherently capable of detecting these fine-grained categories. Actually, PC-459 [44] and ADE-847 [66] also have fine-grained categories, but they suffer from semantic duplication problem. In detail, semantic duplication means that there will be both fine-grained categories as well as corresponding broad categories in these datasets, which would cause bias in the metrics calculation. For instance, if a model assigns the \"door\" tag to a region whose ground truth label is \"double door\", it will be considered incorrect. For open-vocabulary segmentation setting, we argue that the responsibility of the model is to discern the correct semantic. The distribution of such duplication categories and the corresponding performance statistics is unreasonable. SCAN [37] also points similar observations and try to solve this by designing new metric. However, this requires manual judgment of the relationships between categories, introducing additional cost and uncertainty. Our OpenBench, on the other hand, has fine-grained semantics without the problem of semantic duplication, which helps to more accurately reflect the real capabilities of related models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 325, + 294, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 325, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 57, + 325, + 294, + 514 + ], + "type": "text", + "content": "Figure 2 illustrates part of the categories contained in OpenBench, where some categories such as food related even differ significantly from the visual domain of the training set. Furthermore, our OpenBench also accounts for \"other\" or \"background\" pixels in images, which are often overlooked by existing datasets. This is important for open-vocabulary segmentation because the model is essentially forced to choose the most appropriate category from a predefined set. If there is no correct class and no \"others\" option in the candidate set, the model would be compelled to make an incorrect prediction. Moreover, \"others\" or \"background\" require the model to consider not only the direct visual-linguistic correspondence between the image content and each class, but also the relative relationships among different classes. This adds complexity and makes the task more relevant to real-world scenarios." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 528, + 162, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 528, + 162, + 538 + ], + "spans": [ + { + "bbox": [ + 58, + 528, + 162, + 538 + ], + "type": "text", + "content": "3.2. Dataset Collection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 546, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 546, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 546, + 294, + 712 + ], + "type": "text", + "content": "Our OpenBench is primarily derived from existing accessible segmentation datasets. To obtain semantic categories that differ significantly from the training set, i.e., COCO [32], we first organize a group of segmentation datasets that possess category annotations, e.g., Food103 [53], Imagenet-S [14], and CamVid [3]. For all categories in the collected data, we leverage CLIP-L/14 [46] to compute their maximum similarity with COCO categories. Then we calculate the minimum similarity among the categories present in each image as the image's similarity to the training semantics. If the image similarity is greater than the set threshold " + }, + { + "bbox": [ + 57, + 546, + 294, + 712 + ], + "type": "inline_equation", + "content": "\\sigma_{1}" + }, + { + "bbox": [ + 57, + 546, + 294, + 712 + ], + "type": "text", + "content": ", the image will be filtered. Conversely, if the similarity of the image is less than the threshold " + }, + { + "bbox": [ + 57, + 546, + 294, + 712 + ], + "type": "inline_equation", + "content": "\\sigma_{1}" + }, + { + "bbox": [ + 57, + 546, + 294, + 712 + ], + "type": "text", + "content": ", we would set the category exist in the image" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 73, + 553, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 73, + 553, + 119 + ], + "spans": [ + { + "bbox": [ + 317, + 73, + 553, + 119 + ], + "type": "text", + "content": "whose similarity is greater than the threshold " + }, + { + "bbox": [ + 317, + 73, + 553, + 119 + ], + "type": "inline_equation", + "content": "\\sigma_{2}" + }, + { + "bbox": [ + 317, + 73, + 553, + 119 + ], + "type": "text", + "content": " to \"others\" and save the related images and annotation. Finally, after manually filtering the categories for potential conflicts or duplicates, we get the OpenBench." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 316, + 133, + 369, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 133, + 369, + 144 + ], + "spans": [ + { + "bbox": [ + 316, + 133, + 369, + 144 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 316, + 154, + 380, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 154, + 380, + 164 + ], + "spans": [ + { + "bbox": [ + 316, + 154, + 380, + 164 + ], + "type": "text", + "content": "4.1. Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 316, + 171, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 171, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 316, + 171, + 553, + 361 + ], + "type": "text", + "content": "The overall pipeline of our proposed OVSNet is shown in Figure 3. We leverage the vision and text encoder of pretraind CLIP [46] to encode the input image and category names, respectively. Then, a segmentation decoder is trained to decode the visual features, generating masks for potential targets and their corresponding query embeddings. We then leverage the predicted masks to extract related CLIP features with Mask Pooling operation. To integrate heterogeneous knowledge spaces and improve the representation robustness, Gradient-Free Aggregation fuses the query embeddings with pooled CLIP features. The aggregated visual embeddings are combined with linguistic embeddings to perform visual-linguistic alignment and generate final output. During training, a Proxy Calibration strategy is utilized to expand training space by applying supervision on the convex combination of training semantics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 373, + 463, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 373, + 463, + 384 + ], + "spans": [ + { + "bbox": [ + 316, + 373, + 463, + 384 + ], + "type": "text", + "content": "4.2. Gradient-Free Aggregation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 316, + 390, + 553, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 390, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 316, + 390, + 553, + 509 + ], + "type": "text", + "content": "Although CLIP [46] exhibits strong generalization ability, it optimizes for image-level visual-linguistic alignment during pretraining, which creates a great domain gap compared to mask-level alignment. As a result, directly using CLIP features obtained through mask pooling for mask classification may lead to suboptimal performance. The query embeddings generated by the segmentation decoder, on the other hand, have a strong region-level alignment prior owing to their suitability for the training space. However, they struggle to recognize novel semantics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 510, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 553, + 712 + ], + "type": "text", + "content": "Therefore, we propose to reorganize them to obtain more robust domain-adapted representations. An intuitive approach is to learn the intrinsic connection between them through self-attention or cross-attention mechanisms. However, we observe that such learning-based strategy tends to develop excessive reliance on query embeddings fitted to seen semantics during the training process while neglecting the utilization of more generalizable information provided by CLIP, ultimately affecting the model's ability to understand unlimited categories. Therefore, we propose to combine CLIP features and query embeddings via a gradient-free approach to avoid potential overfitting problems during the learning of the joint space. This is inspired by Random Walk algorithm [16, 56]. Specifically, there are multiple iterations in this process where two features learn collaboratively in shared embedding space until convergence. Suppose we have " + }, + { + "bbox": [ + 316, + 510, + 553, + 712 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 316, + 510, + 553, + 712 + ], + "type": "text", + "content": " queries. For convenience, we omit the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22667" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 553, + 244 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 553, + 244 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 553, + 244 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 553, + 244 + ], + "type": "image", + "image_path": "c7a64c0b7842a0c3ebf6f55d5287a827062903b0d7a95554faaefef99e4654c1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 251, + 555, + 319 + ], + "lines": [ + { + "bbox": [ + 54, + 251, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 555, + 319 + ], + "type": "text", + "content": "Figure 3. Pipeline of our OVSNet. The input image and corresponding categories are first encoded by pretrained CLIP [46]. A segmentation decoder is trained to decode the visual features and generate potential masks as well as corresponding query embeddings. We then leverage the predicted masks to extract related CLIP features with Mask Pooling operation. To integrate different knowledge spaces and improve the representation robustness, Gradient-Free Aggregation fuses the query embeddings with CLIP features. The aggregated visual embeddings are combined with class embeddings to perform visual-linguistic alignment and generate final output. During training, we apply Proxy Calibration strategy to expand training space in a cost-free manner." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 333, + 294, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 333, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 54, + 333, + 294, + 443 + ], + "type": "text", + "content": "query index and take a random query as the example. The corresponding query embeddings and CLIP features can be represented by " + }, + { + "bbox": [ + 54, + 333, + 294, + 443 + ], + "type": "inline_equation", + "content": "F_{Q}" + }, + { + "bbox": [ + 54, + 333, + 294, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 333, + 294, + 443 + ], + "type": "inline_equation", + "content": "F_{C}" + }, + { + "bbox": [ + 54, + 333, + 294, + 443 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 54, + 333, + 294, + 443 + ], + "type": "inline_equation", + "content": "F_{Q}" + }, + { + "bbox": [ + 54, + 333, + 294, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 333, + 294, + 443 + ], + "type": "inline_equation", + "content": "F_{C}" + }, + { + "bbox": [ + 54, + 333, + 294, + 443 + ], + "type": "text", + "content": " are defined as the initial state " + }, + { + "bbox": [ + 54, + 333, + 294, + 443 + ], + "type": "inline_equation", + "content": "F_{Q}^{0}" + }, + { + "bbox": [ + 54, + 333, + 294, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 333, + 294, + 443 + ], + "type": "inline_equation", + "content": "F_{C}^{0}" + }, + { + "bbox": [ + 54, + 333, + 294, + 443 + ], + "type": "text", + "content": " at the 0-th iteration. Then we formulate affinity " + }, + { + "bbox": [ + 54, + 333, + 294, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 54, + 333, + 294, + 443 + ], + "type": "text", + "content": " by scaled dot product: " + }, + { + "bbox": [ + 54, + 333, + 294, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{Z} = \\lambda F_{C}^{0}(F_{Q}^{0})^{\\top}" + }, + { + "bbox": [ + 54, + 333, + 294, + 443 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 333, + 294, + 443 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 54, + 333, + 294, + 443 + ], + "type": "text", + "content": " is the scaling factor. At " + }, + { + "bbox": [ + 54, + 333, + 294, + 443 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 333, + 294, + 443 + ], + "type": "text", + "content": "-th iteration, the query embedding " + }, + { + "bbox": [ + 54, + 333, + 294, + 443 + ], + "type": "inline_equation", + "content": "F_{Q}^{t}" + }, + { + "bbox": [ + 54, + 333, + 294, + 443 + ], + "type": "text", + "content": " is derived by the original embedding " + }, + { + "bbox": [ + 54, + 333, + 294, + 443 + ], + "type": "inline_equation", + "content": "F_{Q}^{0}" + }, + { + "bbox": [ + 54, + 333, + 294, + 443 + ], + "type": "text", + "content": " and the updated CLIP feature " + }, + { + "bbox": [ + 54, + 333, + 294, + 443 + ], + "type": "inline_equation", + "content": "F_{C}^{t-1}" + }, + { + "bbox": [ + 54, + 333, + 294, + 443 + ], + "type": "text", + "content": " from previous iteration, which can be formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 93, + 448, + 295, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 448, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 93, + 448, + 295, + 464 + ], + "type": "interline_equation", + "content": "F _ {Q} ^ {t} = \\omega \\operatorname {N o r m} (\\mathcal {Z}) ^ {\\top} F _ {C} ^ {t - 1} + (1 - \\omega) F _ {Q} ^ {0}, \\tag {1}", + "image_path": "85848fc3d0cfbc2b82e9968f43ac3e9c0821fd251df6721670a535473bb02e7f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 467, + 295, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 295, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 295, + 489 + ], + "type": "text", + "content": "Subsequently, it is utilized to boost the CLIP feature of this iteration:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 489, + 294, + 504 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 489, + 294, + 504 + ], + "spans": [ + { + "bbox": [ + 116, + 489, + 294, + 504 + ], + "type": "interline_equation", + "content": "F _ {C} ^ {t} = \\omega \\mathcal {Z} F _ {Q} ^ {t} + (1 - \\omega) F _ {C} ^ {0}, \\tag {2}", + "image_path": "ed77d67993341f7936262ce7777b0b57cf76887b3960001fb3acaad6cb5c12bf.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 506, + 296, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 296, + 543 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 296, + 543 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 506, + 296, + 543 + ], + "type": "inline_equation", + "content": "\\omega \\in (0,1)" + }, + { + "bbox": [ + 55, + 506, + 296, + 543 + ], + "type": "text", + "content": " is the factor that controls the degree of feature fusion. By combining the above equations we can obtain the overall iterative update formula of " + }, + { + "bbox": [ + 55, + 506, + 296, + 543 + ], + "type": "inline_equation", + "content": "F_{C}^{t}" + }, + { + "bbox": [ + 55, + 506, + 296, + 543 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 546, + 295, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 546, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 61, + 546, + 295, + 578 + ], + "type": "interline_equation", + "content": "F _ {C} ^ {t} = \\left(\\omega^ {2} A\\right) ^ {t} F _ {C} ^ {0} + (1 - \\omega) \\sum_ {i = 0} ^ {t - 1} \\left(\\omega^ {2} A\\right) ^ {i} \\left(\\omega \\mathcal {Z} F _ {Q} ^ {0} + F _ {C} ^ {0}\\right), \\tag {3}", + "image_path": "497dfba9b91dce455a69ea7c879fdb1494a166c56ea8c414b8cb9a69bebb298a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 583, + 295, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 583, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 55, + 583, + 295, + 644 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 583, + 295, + 644 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 583, + 295, + 644 + ], + "type": "text", + "content": " denotes " + }, + { + "bbox": [ + 55, + 583, + 295, + 644 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}\\mathrm{Norm}(\\mathcal{Z})^{\\top}" + }, + { + "bbox": [ + 55, + 583, + 295, + 644 + ], + "type": "text", + "content": ". Moreover, to avoid the potential issue of unexpected gradient and simplify the calculation process, we use an approximate inference function based on Neumann Series [43] when " + }, + { + "bbox": [ + 55, + 583, + 295, + 644 + ], + "type": "inline_equation", + "content": "t\\to \\infty" + }, + { + "bbox": [ + 55, + 583, + 295, + 644 + ], + "type": "text", + "content": ". Then the update formula can be present as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 84, + 647, + 295, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 647, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 84, + 647, + 295, + 662 + ], + "type": "interline_equation", + "content": "F _ {C} ^ {\\infty} = (1 - \\omega) \\left(I - \\omega^ {2} A\\right) ^ {- 1} \\left(\\omega \\mathcal {Z} F _ {Q} ^ {0} + F _ {C} ^ {0}\\right), \\tag {4}", + "image_path": "e6bcd1f51f6ae722a472fc626b9102e1c72fd05ef279779cca1ab48a58283ef2.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " represents the identity matrix. With this gradient-free aggregation strategy, the model can recombine query embeddings with CLIP features to produce more robust region-level representations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 332, + 422, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 332, + 422, + 345 + ], + "spans": [ + { + "bbox": [ + 314, + 332, + 422, + 345 + ], + "type": "text", + "content": "4.3. Proxy Calibration" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 357, + 555, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 467 + ], + "type": "text", + "content": "A broader semantic training space tends to result in more generalized model representations. However, expanding the training space for OVS is costly due to the intensive prediction nature of segmentation tasks. To alleviate this problem, we propose a cost-free approach named Proxy Calibration (PC) to expand the training space and enhance the generalization ability of the model. It leverages the synthesis of proxy embeddings, which approximates novel semantics via convex combination between the in-vocabulary classes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 472, + 555, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 555, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 555, + 534 + ], + "type": "text", + "content": "Specifically, having the aggregated query embeddings " + }, + { + "bbox": [ + 313, + 472, + 555, + 534 + ], + "type": "inline_equation", + "content": "F_{Q}" + }, + { + "bbox": [ + 313, + 472, + 555, + 534 + ], + "type": "text", + "content": ", we perform random convex combination of them during training to simulate embedding space for undefined classes, as shown by the black dotted line in Figure 3. This process can be formulated as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 359, + 556, + 555, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 556, + 555, + 572 + ], + "spans": [ + { + "bbox": [ + 359, + 556, + 555, + 572 + ], + "type": "interline_equation", + "content": "F _ {Q m n} ^ {\\prime} = \\alpha * F _ {Q m} + (1 - \\alpha) * F _ {Q n} \\tag {5}", + "image_path": "af80e911677922c4f4cc4c881dfd1bc1acf7ca5afc8edd3d69b932d54de8c004.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "m, n \\in \\{1, 2, \\dots, N\\}" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": " is a random weight sampled from the distribution of " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "Beta(\\gamma, \\gamma)" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "F'" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": " denotes the generated proxy embeddings. The same enhancement is performed for the corresponding CLIP feature " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "F_C" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": " and linguistic embeddings " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "F_T" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": ". Note that the text features here refer to the class representations of the ground truth regions corresponding to each query, which are obtained after Hungarian Matching [5]. Having the generated " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "F_Q'" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "F_C'" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "F_T'" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": ", we perform distance supervision on them to promote region-level visual-linguistic alignment performance from the per" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22668" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 106, + 557, + 348 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 2. Performance comparison with state-of-the-art methods on existing benchmarks and the proposed OpenBench. ADE, PC, and VOC denote ADE20K [66], Pascal Context [44], and Pascal VOC [13], respectively. 171 and 133 indicate the number of training categories. * denotes the re-trained version according to the official code for fair comparison." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 106, + 557, + 348 + ], + "lines": [ + { + "bbox": [ + 55, + 106, + 557, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 106, + 557, + 348 + ], + "type": "table", + "html": "
MethodTraining DatasetADE-150ADE-847PC-59PC-459VOCOpenBenchAverage Score
Base-level CLIP Backbone
SimSeg* [60]COCO-Stuff-17121.16.951.99.791.819.433.5
OVSeg [31]COCO-Stuff-17124.87.153.311.092.624.935.6
SAN [61]COCO-Stuff-17127.510.153.812.694.039.639.6
SCAN [37]COCO-Stuff-17130.810.858.413.297.038.741.5
CATSeg [10]COCO-Stuff-17131.812.057.519.094.636.141.8
SED [57]COCO-Stuff-17131.611.457.318.694.433.341.1
FCCLIP* [63]COCO-Panoptic-13331.113.554.812.893.240.340.9
MAFT [25]COCO-Panoptic-13329.110.153.512.890.037.538.8
MAFT+ [26]COCO-Panoptic-13334.613.857.516.295.443.743.5
OVSNet (Ours)COCO-Panoptic-13335.814.558.619.195.744.944.8
Large-level CLIP Backbone
SimSeg* [60]COCO-Stuff-17121.77.152.210.292.327.835.2
OVSeg [31]COCO-Stuff-17129.69.055.712.494.535.639.5
SAN [61]COCO-Stuff-17132.112.457.715.794.643.442.7
SCAN [37]COCO-Stuff-17133.514.059.316.397.149.444.9
SED [57]COCO-Stuff-17135.213.960.622.696.145.645.6
FCCLIP [63]COCO-Panoptic-13334.014.858.418.295.444.844.3
MAFT [25]COCO-Panoptic-13334.413.157.517.093.041.742.8
MAFT+ [26]COCO-Panoptic-13336.115.159.421.696.547.346.0
OVSNet (Ours)COCO-Panoptic-13337.116.262.023.596.948.647.4
", + "image_path": "c519e915a6a847400c6b085e623bc8297e063f203e1f7fd2067ffd780b7ea571.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 358, + 205, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 205, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 205, + 370 + ], + "type": "text", + "content": "spective of proxy semantics. That is:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 376, + 295, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 376, + 295, + 406 + ], + "spans": [ + { + "bbox": [ + 116, + 376, + 295, + 406 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {P Q} = 1 - \\frac {\\mathbf {F} _ {\\mathbf {Q}} ^ {\\prime} \\cdot \\mathbf {F} _ {\\mathbf {T}} ^ {\\prime}}{\\| \\mathbf {F} _ {\\mathbf {Q}} ^ {\\prime} \\| _ {2} \\| \\mathbf {F} _ {\\mathbf {T}} ^ {\\prime} \\| _ {2}}, \\tag {6}", + "image_path": "77684d9b164da9069585f57439451c966c68e17ad93902597caf2922e1328811.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 118, + 407, + 295, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 407, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 118, + 407, + 295, + 435 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {P C} = 1 - \\frac {\\mathbf {F} _ {\\mathbf {C}} ^ {\\prime} \\cdot \\mathbf {F} _ {\\mathbf {T}} ^ {\\prime}}{\\| \\mathbf {F} _ {\\mathbf {C}} ^ {\\prime} \\| _ {2} \\| \\mathbf {F} _ {\\mathbf {T}} ^ {\\prime} \\| _ {2}}, \\tag {7}", + "image_path": "b57fc75b297835895eb50c4c7e1bcd8396ed601f5b3ae8767ca5afe91ae70f2a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 439, + 296, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 439, + 296, + 511 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 296, + 511 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 439, + 296, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{PQ}" + }, + { + "bbox": [ + 55, + 439, + 296, + 511 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 439, + 296, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{PC}" + }, + { + "bbox": [ + 55, + 439, + 296, + 511 + ], + "type": "text", + "content": " denote the applied proxy supervision for query embeddings and pooled CLIP features, respectively. The final proxy loss is the sum of " + }, + { + "bbox": [ + 55, + 439, + 296, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{PQ}" + }, + { + "bbox": [ + 55, + 439, + 296, + 511 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 439, + 296, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{PC}" + }, + { + "bbox": [ + 55, + 439, + 296, + 511 + ], + "type": "text", + "content": ". By applying additional supervision in this way, the model is not limited to a specific training space and can learn more robust visual representations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 521, + 132, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 132, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 132, + 534 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 540, + 242, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 242, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 242, + 552 + ], + "type": "text", + "content": "5.1. Evaluation Benchmark and Metric" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": "To evaluate the effectiveness and generalization ability of our method on diverse scenarios, we conduct extensive experiments on the popular existing benchmarks, ADE20K150 [66], ADE20K847 [66], Pascal VOC [13], Pascal Context-59 [44], Pascal Context-459 [44] and our proposed OpenBench. ADE20K is a large-scale scene understanding benchmark, containing 20k training images, 2k validation images, and 3k testing images. There are two splits of this dataset. ADE20K-150 contains 150 semantic classes whereas ADE20K-847 has 847 classes. The images of both are the same. Pascal Context is an extension of Pascal VOC 2010, containing 5,005 validation images. We take the commonly used PC-59 and challenging PC-459" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 357, + 555, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 441 + ], + "type": "text", + "content": "version for validation. Pascal VOC contains 11,185 training images and 1,449 validation images from 20 classes. We use the provided augmented annotations. Following previous works [12, 15, 60], we take the mean-intersection-over-union (mIoU) as the metric to compare our model with previous state-of-the-art methods. In ablation, we assume the use of base-level CLIP as the backbone for comparison." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 451, + 448, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 448, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 448, + 464 + ], + "type": "text", + "content": "5.2. Implementation Details" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 470, + 555, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 555, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 555, + 650 + ], + "type": "text", + "content": "We utilize the Mask2Former [8] as the segmentation decoder and the implementation is based on detectron2 [54]. The model is trained with batch size of 8 and total iteration of " + }, + { + "bbox": [ + 313, + 470, + 555, + 650 + ], + "type": "inline_equation", + "content": "60\\mathrm{k}" + }, + { + "bbox": [ + 313, + 470, + 555, + 650 + ], + "type": "text", + "content": ". The base learning rate is 0.0001 with a polynomial schedule. Following previous works [26, 63], we take CLIP [46] of ConvNeXt as the backbone. The input image is resized and cropped to " + }, + { + "bbox": [ + 313, + 470, + 555, + 650 + ], + "type": "inline_equation", + "content": "1024\\times 1024" + }, + { + "bbox": [ + 313, + 470, + 555, + 650 + ], + "type": "text", + "content": ". For data augmentation, random flip and color augmentation are adopted. The weight decay of the model is 0.05. " + }, + { + "bbox": [ + 313, + 470, + 555, + 650 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 470, + 555, + 650 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 470, + 555, + 650 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 470, + 555, + 650 + ], + "type": "text", + "content": " of GFA and PC are set to 0.2 and 2 by default. The segmentation loss consists of dice loss and cross entropy loss. The classification loss is cross entropy loss. For the weights of the loss function, we set 5 and 2 for segmentation loss and classification loss, respectively. Other hyperparameters are the same as Mask2Former [9]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 658, + 400, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 400, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 400, + 671 + ], + "type": "text", + "content": "5.3. Main Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": "We compare our model with existing state-of-the-art approaches in Table 2. It can be seen that with proxy calibration expanding training semantics and the gradient-free" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22669" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 106, + 294, + 166 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 295, + 104 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 295, + 104 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 295, + 104 + ], + "type": "text", + "content": "Table 3. Ablation experiments about the proposed methods. GFA and PC indicate the Gradient-Free Aggregation and Proxy Calibration strategies, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 106, + 294, + 166 + ], + "lines": [ + { + "bbox": [ + 55, + 106, + 294, + 166 + ], + "spans": [ + { + "bbox": [ + 55, + 106, + 294, + 166 + ], + "type": "table", + "html": "
MethodADE-150PC-459OpenBench
Baseline33.114.342.3
+ GFA34.7↑ (1.6)16.0↑ (1.7)43.7↑ (1.4)
+ PC33.9↑ (0.8)17.2↑ (2.9)44.3↑ (2.0)
+ Both35.8↑ (2.7)19.1↑ (4.8)44.9↑ (2.6)
", + "image_path": "13244781382e9943c88b03c348c971a72a78dcb831f904419d2747b33e447a95.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 170, + 298, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 170, + 298, + 361 + ], + "spans": [ + { + "bbox": [ + 54, + 170, + 298, + 361 + ], + "type": "text", + "content": "aggregation algorithm modeling the joint space, our model achieves the state-of-the-art performance on most popular benchmarks. The last column counts the average performance across multiple datasets, which shows that our method performs well for both in-vocabulary as well as out-vocabulary diverse scenarios. The relatively inferior performance of our method on VOC is due to the fact that the semantics of this dataset are almost contained in the training set, which tends to benefit models such as SCAN [37] that finetunes the CLIP backbone on training set. Besides, the greater divergence of OpenBench from training semantics does not imply that it is totally more challenging than other benchmarks, because the final performance is influenced by various factors such as image content complexity, image clarity, and the number of given categories. More discussion can be seen in the supplementary materials." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 370, + 149, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 149, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 149, + 383 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 388, + 295, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 295, + 509 + ], + "type": "text", + "content": "Component Analysis Table 3 shows the ablations about the proposed Gradient-Free Aggregation (GFA) and Proxy Calibration (PC) strategy. We evaluate their effectiveness on both existing benchmarks and our proposed OpenBench. It can be seen that with GFA fusing CLIP feature and query embeddings, the model's performance is greatly improved on both seen and unseen semantics. Besides, PC can facilitate the model's ability to understand various scenarios by cost-free expansion of the training semantic space. When applying both, the model achieves the best performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 514, + 295, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 295, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 295, + 659 + ], + "type": "text", + "content": "Embedding Distribution Analysis To demonstrate that our proposed OpenBench is more open-vocabulary compared to the existing test sets, we perform t-SNE visualization of the category embeddings of the training set, the existing test sets, and our OpenBench in Figure 4. Besides, we also visualize the proxy embeddings generated by PC. We can see that the existing test sets (green dots) are overall similar to the training semantics (blue dots). Our OpenBench (orange dots), on the other hand, shows significant differences. In addition, the proxy embeddings (red dots) generated by PC expand the training set and thus help to improve the robustness and generalization of the model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Analysis about Inference Class Number In this part we explore an interesting phenomenon: why the OVS models perform much worse on ADE-847 and PC-459 than on other test sets? Intuitively, this is because these datasets" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 328, + 71, + 540, + 219 + ], + "blocks": [ + { + "bbox": [ + 328, + 71, + 540, + 219 + ], + "lines": [ + { + "bbox": [ + 328, + 71, + 540, + 219 + ], + "spans": [ + { + "bbox": [ + 328, + 71, + 540, + 219 + ], + "type": "image", + "image_path": "ae6d8df7e54015ea14855309442f3fc89d438b782f13954996469bbf9ad68bb1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 223, + 555, + 257 + ], + "lines": [ + { + "bbox": [ + 313, + 223, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 555, + 257 + ], + "type": "text", + "content": "Figure 4. Visualization of category embeddings across the training space, existing evaluation space, our OpenBench space, and the generated proxy space." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 326, + 262, + 541, + 396 + ], + "blocks": [ + { + "bbox": [ + 326, + 262, + 541, + 396 + ], + "lines": [ + { + "bbox": [ + 326, + 262, + 541, + 396 + ], + "spans": [ + { + "bbox": [ + 326, + 262, + 541, + 396 + ], + "type": "image", + "image_path": "2f9ffae676bc6c1a4e9b664692e25a8682fe827cae87d8c2d18c096cd7d52636.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 396, + 555, + 430 + ], + "lines": [ + { + "bbox": [ + 313, + 396, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 555, + 430 + ], + "type": "text", + "content": "Figure 5. Impact of the number of inference categories on model performance. As the number of irrelevant categories increases, the model's performance on the same image degrades." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 445, + 555, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 626 + ], + "type": "text", + "content": "contain more novel categories. However, we find that another factor significantly impacts model performance, that is the number of categories fed into the model during inference. Let's start with the conclusion: for the same model and test image, the model performance is worse when there are more inference categories given. In other words, the model may be able to make correct predictions when only a small number of categories are given. However, when more categories are prompted, the model's output will be incorrect. Here we take two typical methods, SAN [61] and MAFT-Plus [26], as examples and show how their performance on the ADE-150 dataset changes as the number of categories increases. As shown in Figure 5, with the number of given irrelevant categories increases, the models of different technology routes show a consistent decline." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "Internal Analysis of Proposed Strategies In Table 4 we conduct more experiments about the proposed methods. Specifically, we first compare our gradient-free aggregation designs with the typical learning-based feature fusion approaches in (a). Results show that vanilla attention performs inferior than our GFA, especially on the more open-vocabulary set OpenBench (degraded by more than" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22670" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 71, + 182, + 145 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 182, + 145 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 182, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 182, + 145 + ], + "type": "image", + "image_path": "eac635c0d5e9fcf77f4ef38cc457d980ef13348b47609cec0fab18099c59cdc6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 147, + 182, + 156 + ], + "lines": [ + { + "bbox": [ + 63, + 147, + 182, + 156 + ], + "spans": [ + { + "bbox": [ + 63, + 147, + 182, + 156 + ], + "type": "text", + "content": "(a) Seen: building treefence bicycle" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 182, + 72, + 304, + 145 + ], + "blocks": [ + { + "bbox": [ + 182, + 72, + 304, + 145 + ], + "lines": [ + { + "bbox": [ + 182, + 72, + 304, + 145 + ], + "spans": [ + { + "bbox": [ + 182, + 72, + 304, + 145 + ], + "type": "image", + "image_path": "6c46b639d0000c7d046b194c73a0b7df5d90e39b16b95a00b9df546125c05e8e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 61, + 157, + 182, + 234 + ], + "blocks": [ + { + "bbox": [ + 182, + 147, + 303, + 156 + ], + "lines": [ + { + "bbox": [ + 182, + 147, + 303, + 156 + ], + "spans": [ + { + "bbox": [ + 182, + 147, + 303, + 156 + ], + "type": "text", + "content": "Unseen: sidewalk streetlight" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 157, + 182, + 234 + ], + "lines": [ + { + "bbox": [ + 61, + 157, + 182, + 234 + ], + "spans": [ + { + "bbox": [ + 61, + 157, + 182, + 234 + ], + "type": "image", + "image_path": "8ebda229eee540b5321463bc44485c601e2283e02a91ea233c57af6c629916de.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 182, + 157, + 304, + 234 + ], + "blocks": [ + { + "bbox": [ + 63, + 235, + 182, + 243 + ], + "lines": [ + { + "bbox": [ + 63, + 235, + 182, + 243 + ], + "spans": [ + { + "bbox": [ + 63, + 235, + 182, + 243 + ], + "type": "text", + "content": "(c) Seen: ceiling desk chair" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 182, + 157, + 304, + 234 + ], + "lines": [ + { + "bbox": [ + 182, + 157, + 304, + 234 + ], + "spans": [ + { + "bbox": [ + 182, + 157, + 304, + 234 + ], + "type": "image", + "image_path": "e712fefb1e0a5711f184bc4a71101150a4f76838d596280e8c0aa697dcc1b676.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 61, + 246, + 182, + 321 + ], + "blocks": [ + { + "bbox": [ + 182, + 235, + 303, + 245 + ], + "lines": [ + { + "bbox": [ + 182, + 235, + 303, + 245 + ], + "spans": [ + { + "bbox": [ + 182, + 235, + 303, + 245 + ], + "type": "text", + "content": "Unseen: computer chandelier" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 246, + 182, + 321 + ], + "lines": [ + { + "bbox": [ + 61, + 246, + 182, + 321 + ], + "spans": [ + { + "bbox": [ + 61, + 246, + 182, + 321 + ], + "type": "image", + "image_path": "217dfb670f55b5179e12cbf69a31b51589f5cf93124a108cb2ddf12dd600fe0f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 323, + 269, + 331 + ], + "lines": [ + { + "bbox": [ + 182, + 323, + 269, + 331 + ], + "spans": [ + { + "bbox": [ + 182, + 323, + 269, + 331 + ], + "type": "text", + "content": "→ opera house" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 182, + 246, + 304, + 321 + ], + "blocks": [ + { + "bbox": [ + 63, + 323, + 182, + 331 + ], + "lines": [ + { + "bbox": [ + 63, + 323, + 182, + 331 + ], + "spans": [ + { + "bbox": [ + 63, + 323, + 182, + 331 + ], + "type": "text", + "content": "(e) building" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 182, + 246, + 304, + 321 + ], + "lines": [ + { + "bbox": [ + 182, + 246, + 304, + 321 + ], + "spans": [ + { + "bbox": [ + 182, + 246, + 304, + 321 + ], + "type": "image", + "image_path": "16aa424d28b7f6a7aae259f3fb3bc024884a8c8eeae3cbf2fa86affbf7b304fa.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 427, + 145 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 427, + 145 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 427, + 145 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 427, + 145 + ], + "type": "image", + "image_path": "3ee052e628fdbc2b67d490ead5016756613b7ad75c6edcba61f0a542d94a3e42.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 147, + 427, + 156 + ], + "lines": [ + { + "bbox": [ + 309, + 147, + 427, + 156 + ], + "spans": [ + { + "bbox": [ + 309, + 147, + 427, + 156 + ], + "type": "text", + "content": "(b) Seen: sea boat mountain" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 428, + 72, + 550, + 145 + ], + "blocks": [ + { + "bbox": [ + 428, + 72, + 550, + 145 + ], + "lines": [ + { + "bbox": [ + 428, + 72, + 550, + 145 + ], + "spans": [ + { + "bbox": [ + 428, + 72, + 550, + 145 + ], + "type": "image", + "image_path": "a99f3606ca78fda54e006e50ca67b981ed5b989ea0fedd8d3833088e8486c955.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 431, + 147, + 550, + 156 + ], + "lines": [ + { + "bbox": [ + 431, + 147, + 550, + 156 + ], + "spans": [ + { + "bbox": [ + 431, + 147, + 550, + 156 + ], + "type": "text", + "content": "Unseen: pier pole" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 307, + 157, + 427, + 234 + ], + "blocks": [ + { + "bbox": [ + 307, + 157, + 427, + 234 + ], + "lines": [ + { + "bbox": [ + 307, + 157, + 427, + 234 + ], + "spans": [ + { + "bbox": [ + 307, + 157, + 427, + 234 + ], + "type": "image", + "image_path": "9fbc96ee036cac8c76f0d7f9c34d4a1bbe6eef3a0c55f03c5b3e6070cf778e20.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 235, + 427, + 243 + ], + "lines": [ + { + "bbox": [ + 309, + 235, + 427, + 243 + ], + "spans": [ + { + "bbox": [ + 309, + 235, + 427, + 243 + ], + "type": "text", + "content": "(d) Seen: plant rock person" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 428, + 157, + 550, + 234 + ], + "blocks": [ + { + "bbox": [ + 428, + 157, + 550, + 234 + ], + "lines": [ + { + "bbox": [ + 428, + 157, + 550, + 234 + ], + "spans": [ + { + "bbox": [ + 428, + 157, + 550, + 234 + ], + "type": "image", + "image_path": "fa1a4772f0f95bc1b6a5bbc312259b27de77c8c33cc45d533b8dd2bc5cd558f7.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 431, + 235, + 550, + 243 + ], + "lines": [ + { + "bbox": [ + 431, + 235, + 550, + 243 + ], + "spans": [ + { + "bbox": [ + 431, + 235, + 550, + 243 + ], + "type": "text", + "content": "Unseen: waterfall" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 307, + 246, + 428, + 321 + ], + "blocks": [ + { + "bbox": [ + 307, + 246, + 428, + 321 + ], + "lines": [ + { + "bbox": [ + 307, + 246, + 428, + 321 + ], + "spans": [ + { + "bbox": [ + 307, + 246, + 428, + 321 + ], + "type": "image", + "image_path": "0c4788019177927f18722285b24801a6d2619541eebe6fe75b0369a5694e23a3.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 323, + 321, + 331 + ], + "lines": [ + { + "bbox": [ + 312, + 323, + 321, + 331 + ], + "spans": [ + { + "bbox": [ + 312, + 323, + 321, + 331 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 334, + 555, + 357 + ], + "lines": [ + { + "bbox": [ + 55, + 334, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 555, + 357 + ], + "type": "text", + "content": "Figure 6. Visualization of segmentation results. (a) - (d) demonstrate the successful segmentation cases of our method for seen and unseen categories. (e) and (f) display the adaptability for flexible text query. Best viewed in color." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 428, + 246, + 550, + 321 + ], + "blocks": [ + { + "bbox": [ + 428, + 246, + 550, + 321 + ], + "lines": [ + { + "bbox": [ + 428, + 246, + 550, + 321 + ], + "spans": [ + { + "bbox": [ + 428, + 246, + 550, + 321 + ], + "type": "image", + "image_path": "08cc198f1b540bd4c57fca47bdb0e53b7258f6b724f49f6953e5716f86ab39a2.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 431, + 323, + 492, + 331 + ], + "lines": [ + { + "bbox": [ + 431, + 323, + 492, + 331 + ], + "spans": [ + { + "bbox": [ + 431, + 323, + 492, + 331 + ], + "type": "text", + "content": "→ ferrari" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "table", + "bbox": [ + 58, + 413, + 288, + 507 + ], + "blocks": [ + { + "bbox": [ + 55, + 365, + 296, + 410 + ], + "lines": [ + { + "bbox": [ + 55, + 365, + 296, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 296, + 410 + ], + "type": "text", + "content": "Table 4. Ablation experiments within each modules. (a) shows the comparison of our gradient-free algorithm with learning-based method, e.g., cross attention. (b) demonstrates the effect of " + }, + { + "bbox": [ + 55, + 365, + 296, + 410 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 365, + 296, + 410 + ], + "type": "text", + "content": " of Beta distribution in the proxy learning strategy." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 413, + 288, + 507 + ], + "lines": [ + { + "bbox": [ + 58, + 413, + 288, + 507 + ], + "spans": [ + { + "bbox": [ + 58, + 413, + 288, + 507 + ], + "type": "table", + "html": "
ADE-150PC-459OpenBench
(a) Comparisons of GFA and learning-based strategy
Self Attn34.014.740.6
Cross Attn34.214.841.2
GFA34.716.043.7
(b) Effect of sampling distribution of PC
γ = 133.114.342.3
γ = 0.533.6↑ (0.5)16.1↑ (1.8)43.7↑ (1.4)
γ = 233.9↑ (0.8)17.2↑ (2.9)44.3↑ (2.0)
", + "image_path": "c9fc5f2aec0f5413300f658a15fb4a387d8c40a7572e79315b3b16958a68cc1f.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 518, + 295, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 295, + 578 + ], + "type": "text", + "content": "2.5 mIoU). This is because learning-based methods tend to develop a reliance on the segmentation query embeddings during training, while overlooking the CLIP features that inherently exhibit stronger generalization capabilities, thereby introducing biases." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 578, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 295, + 651 + ], + "type": "text", + "content": "We also experiment with the influence of different " + }, + { + "bbox": [ + 55, + 578, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 578, + 295, + 651 + ], + "type": "text", + "content": " in Beta distribution to the performance of proxy calibration. The results are shown in Table 4 (b), where " + }, + { + "bbox": [ + 55, + 578, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\gamma = 1" + }, + { + "bbox": [ + 55, + 578, + 295, + 651 + ], + "type": "text", + "content": " is equivalent to do not utilize PC. We find that different " + }, + { + "bbox": [ + 55, + 578, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 578, + 295, + 651 + ], + "type": "text", + "content": " benefit the model performance while the performance is better when the intermediate probability density is high (" + }, + { + "bbox": [ + 55, + 578, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\gamma = 2" + }, + { + "bbox": [ + 55, + 578, + 295, + 651 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 55, + 659, + 139, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 139, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 139, + 670 + ], + "type": "text", + "content": "5.5. Visualization" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Figure 6 shows some segmentation cases of our OVSNet. It can be seen that our method achieves excellent segmentation performance on various scenarios. Specifically, (a) - (d)" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 367, + 555, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 555, + 415 + ], + "type": "text", + "content": "demonstrate the excellent segmentation of our method for seen and unseen categories. (e) and (f) display the adaptability for flexible text query, e.g., change \"building\" to \"opera house\" and \"car\" to \"ferrari\"." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 314, + 428, + 388, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 428, + 388, + 441 + ], + "spans": [ + { + "bbox": [ + 314, + 428, + 388, + 441 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 449, + 555, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 652 + ], + "type": "text", + "content": "In this paper, we identify that the existing open-vocabulary segmentation evaluation sets suffer from excessive semantic similarity to the training set, which limits their ability to comprehensively assess the model's open vocabulary understanding capabilities. To address this, we propose a new benchmark, OpenBench, which features significant differences from the training set and includes more fine-grained categories. Comparison on OpenBench further demonstrates that maintaining the generalization space of CLIP is crucial for OVS. Additionally, we introduce OVSNet, a method that leverages proxy calibration for cost-free augmentation of the training space and gradient-free aggregation for fusing heterogeneous features. Experiments show that the proposed OVSNet achieves excellent performance on both existing datasets and the newly proposed OpenBench. We hope that our benchmark and model design could inspire further researches in this area." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 654, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 712 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by Shenzhen Science and Technology Program under Grant CJGJZD20220517142402006 and Guangdong Natural Science Funds for Distinguished Young Scholar (No. 2025B1515020012)." + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22671" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 295, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 295, + 124 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 295, + 124 + ], + "type": "text", + "content": "[1] Sule Bai, Yong Liu, Yifei Han, Haoji Zhang, and Yansong Tang. Self-calibrated clip for training-free open-vocabulary segmentation. arXiv preprint arXiv:2411.15869, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 124, + 295, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 124, + 295, + 168 + ], + "spans": [ + { + "bbox": [ + 61, + 124, + 295, + 168 + ], + "type": "text", + "content": "[2] Sule Bai, Mingxing Li, Yong Liu, Jing Tang, Haoji Zhang, Lei Sun, Xiangxiang Chu, and Yansong Tang. Univg-r1: Reasoning guided universal visual grounding with reinforcement learning. arXiv preprint arXiv:2505.14231, 2025. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 168, + 295, + 211 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 168, + 295, + 211 + ], + "spans": [ + { + "bbox": [ + 62, + 168, + 295, + 211 + ], + "type": "text", + "content": "[3] Gabriel J. Brostow, Jamie Shotton, Julien Fauqueur, and Roberto Cipolla. Segmentation and recognition using structure from motion point clouds. In ECCV, pages 44-57, 2008. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 213, + 294, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 213, + 294, + 245 + ], + "spans": [ + { + "bbox": [ + 62, + 213, + 294, + 245 + ], + "type": "text", + "content": "[4] Maxime Bucher, Tuan-Hung Vu, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. In NeurlPS, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 246, + 294, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 246, + 294, + 280 + ], + "spans": [ + { + "bbox": [ + 62, + 246, + 294, + 280 + ], + "type": "text", + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 280, + 294, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 294, + 324 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 294, + 324 + ], + "type": "text", + "content": "[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. TPAMI, 2018. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 325, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 325, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 62, + 325, + 294, + 369 + ], + "type": "text", + "content": "[7] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. UNITER: learning universal image-text representations. arXiv preprint arXiv:1909.11740, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 369, + 294, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 294, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 294, + 413 + ], + "type": "text", + "content": "[8] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G. Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 414, + 294, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 294, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 294, + 456 + ], + "type": "text", + "content": "[9] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 457, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 457, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 57, + 457, + 294, + 502 + ], + "type": "text", + "content": "[10] Seokju Cho, Heeseong Shin, Sunghwan Hong, Anurag Arnab, Paul Hongsuck Seo, and Seungryong Kim. Cat-seg: Cost aggregation for open-vocabulary semantic segmentation. In CVPR, pages 4113-4123, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 502, + 294, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 502, + 294, + 556 + ], + "spans": [ + { + "bbox": [ + 57, + 502, + 294, + 556 + ], + "type": "text", + "content": "[11] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 557, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 557, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 557, + 294, + 590 + ], + "type": "text", + "content": "[12] Jian Ding, Nan Xue, Gui-Song Xia, and Dengxin Dai. Decoupling zero-shot semantic segmentation. In CVPR, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 590, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 294, + 635 + ], + "type": "text", + "content": "[13] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes challenge: A retrospective. IJCV, 111:98-136, 2015. 2, 3, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 635, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 294, + 678 + ], + "type": "text", + "content": "[14] Shanghua Gao, Zhong-Yu Li, Ming-Hsuan Yang, Ming-Ming Cheng, Junwei Han, and Philip Torr. Large-scale unsupervised semantic segmentation. TPAMI, 45(6):7457-7476, 2022. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 679, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 679, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 679, + 294, + 713 + ], + "type": "text", + "content": "[15] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Open-vocabulary image segmentation. arXiv preprint arXiv:2112.12143, 2021. 1, 2, 6" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 715 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 95 + ], + "type": "text", + "content": "[16] Leo Grady. Random walks for image segmentation. IEEE TPAMI, 28(11):1768-1783, 2006. 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 96, + 554, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 554, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 554, + 162 + ], + "type": "text", + "content": "[17] Kunyang Han, Yong Liu, Jun Hao Liew, Henghui Ding, Jiajun Liu, Yitong Wang, Yansong Tang, Yujiu Yang, Jiashi Feng, Yao Zhao, et al. Global knowledge calibration for fast open-vocabulary segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 797-807, 2023. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 163, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 196 + ], + "type": "text", + "content": "[18] Shuting He, Henghui Ding, and Wei Jiang. Primitive generation and semantic-related alignment for universal zero-shot segmentation. In CVPR, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 197, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 252 + ], + "type": "text", + "content": "[19] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "type": "text", + "content": "[20] Jiaqi Huang, Zunnan Xu, Ting Liu, Yong Liu, Haonan Han, Kehong Yuan, and Xiu Li. Densely connected parameter-efficient tuning for referring image segmentation. arXiv preprint arXiv:2501.08580, 2025. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 297, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 342 + ], + "type": "text", + "content": "[21] Xiaoke Huang, Jianfeng Wang, Yansong Tang, Zheng Zhang, Han Hu, Jiwen Lu, Lijuan Wang, and Zicheng Liu. Segment and caption anything. arXiv preprint arXiv:2312.00869, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 342, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 387 + ], + "type": "text", + "content": "[22] Joonhyun Jeong, Geondo Park, Jayeon Yoo, Hyungsik Jung, and Heesu Kim. Proxydet: Synthesizing proxy novel classes via classwise mixup for open-vocabulary object detection. In AAAI, pages 2462-2470, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 388, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 432 + ], + "type": "text", + "content": "[23] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 432, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 432, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 432, + 553, + 487 + ], + "type": "text", + "content": "[24] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916, 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 487, + 553, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 487, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 316, + 487, + 553, + 532 + ], + "type": "text", + "content": "[25] Siyu Jiao, Yunchao Wei, Yaowei Wang, Yao Zhao, and Humphrey Shi. Learning mask-aware clip representations for zero-shot segmentation. arXiv preprint arXiv:2310.00240, 2023. 1, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 533, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 553, + 578 + ], + "type": "text", + "content": "[26] Siyu Jiao, Hongguang Zhu, Jiannan Huang, Yao Zhao, Yunchao Wei, and Humphrey Shi. Collaborative vision-text representation optimizing for open-vocabulary segmentation. In ECCV, pages 399-416, 2025. 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 578, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 553, + 611 + ], + "type": "text", + "content": "[27] Boyi Li, Kilian Q. Weinberger, Serge J. Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. In ICLR, 2022. 1, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 611, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 611, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 611, + 553, + 646 + ], + "type": "text", + "content": "[28] Gen Li, Nan Duan, Yuejian Fang, Ming Gong, and Daxin Jiang. Unicoder-vl: A universal encoder for vision and language by cross-modal pre-training. In AAAI, 2020. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 646, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 553, + 690 + ], + "type": "text", + "content": "[29] Wanhua Li, Xiaoke Huang, Zheng Zhu, Yansong Tang, Xiu Li, Jie Zhou, and Jiwen Lu. Ordinalclip: Learning rank prompts for language-guided ordinal regression. NeurIPS, pages 35313-35325, 2022. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 690, + 553, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 690, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 316, + 690, + 553, + 715 + ], + "type": "text", + "content": "[30] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22672" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "text", + "content": "Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-semantics aligned pre-training for vision-language tasks. In ECCV, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "text", + "content": "[31] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yi-nan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted CLIP. arXiv preprint arXiv:2210.04150, 2022. 1, 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 206 + ], + "type": "text", + "content": "[32] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In ECCV, 2014. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 295, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 295, + 240 + ], + "type": "text", + "content": "[33] Yong Liu, Ran Yu, Xinyuan Zhao, and Yujiu Yang. Quality-aware and selective prior enhancement memory network for video object segmentation. In CVPR Workshop, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 295, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 295, + 285 + ], + "type": "text", + "content": "[34] Yong Liu, Ran Yu, Jiahao Wang, Xinyuan Zhao, Yitong Wang, Yansong Tang, and Yujiu Yang. Global spectral filter memory network for video object segmentation. In ECCV, pages 648-665, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 287, + 295, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 295, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 295, + 330 + ], + "type": "text", + "content": "[35] Yong Liu, Ran Yu, Fei Yin, Xinyuan Zhao, Wei Zhao, Weihao Xia, and Yujiu Yang. Learning quality-aware dynamic memory for video object segmentation. In ECCV, pages 468-486, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 332, + 295, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 295, + 374 + ], + "type": "text", + "content": "[36] Yong Liu, Cairong Zhang, Yitong Wang, Jiahao Wang, Yujiu Yang, and Yansong Tang. Universal segmentation at arbitrary granularity with language instruction. arXiv preprint arXiv:2312.01623, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 376, + 295, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 295, + 410 + ], + "type": "text", + "content": "[37] Yong Liu, Sule Bai, Guanbin Li, Yitong Wang, and Yansong Tang. Open-vocabulary segmentation with semantic-assisted calibration. In CVPR, pages 3491-3500, 2024. 1, 4, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 411, + 295, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 295, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 295, + 454 + ], + "type": "text", + "content": "[38] Yong Liu, Ran Yu, Fei Yin, Xinyuan Zhao, Wei Zhao, Weihao Xia, Jiahao Wang, Yitong Wang, Yansong Tang, and Yujiu Yang. Learning high-quality dynamic memory for video object segmentation. IEEE TPAMI, 2025. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 456, + 295, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 295, + 498 + ], + "type": "text", + "content": "[39] Guanxing Lu, Ziwei Wang, Changliu Liu, Jiwen Lu, and Yansong Tang. Thinkbot: Embodied instruction following with thought chain reasoning. arXiv preprint arXiv:2312.07062, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 501, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 295, + 533 + ], + "type": "text", + "content": "[40] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In NeurIPS, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 535, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 295, + 578 + ], + "type": "text", + "content": "[41] Zhuoyan Luo, Yicheng Xiao, Yong Liu, Shuyan Li, Yitong Wang, Yansong Tang, Xiu Li, and Yujiu Yang. Soc: Semantic-assisted object cluster for referring video object segmentation. arXiv preprint arXiv:2305.17011, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 580, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 295, + 633 + ], + "type": "text", + "content": "[42] Zhuoyan Luo, Yinghao Wu, Yong Liu, Yicheng Xiao, Xiaoping Zhang, and Yujiu Yang. Hdc: Hierarchical semantic decoding with counting assistance for generalized referring expression segmentation. arXiv preprint arXiv:2405.15658, 2024. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 295, + 657 + ], + "type": "text", + "content": "[43] Carl D Meyer and Ian Stewart. Matrix analysis and applied linear algebra. SIAM, 2023. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "text", + "content": "[44] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan L. Yuille. The role of context for object detection and semantic segmentation in the wild. In CVPR, 2014. 2, 3, 4, 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[45] Mengxue Qu, Yu Wu, Wu Liu, Qiqi Gong, Xiaodan Liang, Olga Russakovsky, Yao Zhao, and Yunchao Wei. Siri: A simple selective retraining mechanism for transformer-based visual grounding. In ECCV, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 119, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 184 + ], + "type": "text", + "content": "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 187, + 553, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 187, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 317, + 187, + 553, + 220 + ], + "type": "text", + "content": "[47] Yiqin Wang, Haoji Zhang, Yansong Tang, Yong Liu, Jiashi Feng, Jifeng Dai, and Xiaojie Jin. Hierarchical memory for long video qa. arXiv preprint arXiv:2407.00603, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 222, + 553, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 222, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 316, + 222, + 553, + 264 + ], + "type": "text", + "content": "[48] Yulin Wang, Haoji Zhang, Yang Yue, Shiji Song, Chao Deng, Junlan Feng, and Gao Huang. Uni-adafocus: Spatial-temporal dynamic computation for video recognition. IEEE TPAMI, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 268, + 553, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 268, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 316, + 268, + 553, + 311 + ], + "type": "text", + "content": "[49] Yuji Wang, Jingchen Ni, Yong Liu, Chun Yuan, and Yansong Tang. Iterprime: Zero-shot referring image segmentation with iterative grad-cam refinement and primary word emphasis. In AAAI, pages 8159-8168, 2025. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 314, + 553, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 314, + 553, + 356 + ], + "spans": [ + { + "bbox": [ + 316, + 314, + 553, + 356 + ], + "type": "text", + "content": "[50] Yuji Wang, Haoran Xu, Yong Liu, Jiaze Li, and Yansong Tang. Sam2-love: Segment anything model 2 in language-aided audio-visual scenes. In CVPR, pages 28932-28941, 2025." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 359, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 359, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 359, + 553, + 403 + ], + "type": "text", + "content": "[51] Cong Wei, Yujie Zhong, Haoxian Tan, Yong Liu, Zheng Zhao, Jie Hu, and Yujiu Yang. Hyperseg: Towards universal visual segmentation with large language model. arXiv preprint arXiv:2411.17606, 2024." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 405, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 405, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 405, + 553, + 449 + ], + "type": "text", + "content": "[52] Cong Wei, Yujie Zhong, Haoxian Tan, Yingsen Zeng, Yong Liu, Zheng Zhao, and Yujiu Yang. Instructseg: Unifying instructed visual segmentation with multi-modal large language models. arXiv preprint arXiv:2412.14006, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 451, + 553, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 451, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 316, + 451, + 553, + 483 + ], + "type": "text", + "content": "[53] Xiongwei Wu, Xin Fu, Ying Liu, Ee-Peng Lim, Steven CH Hoi, and Qianru Sun. A large-scale benchmark for food image segmentation. In ACM MM, 2021. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 486, + 553, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 486, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 316, + 486, + 553, + 518 + ], + "type": "text", + "content": "[54] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 521, + 553, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 521, + 553, + 563 + ], + "spans": [ + { + "bbox": [ + 316, + 521, + 553, + 563 + ], + "type": "text", + "content": "[55] Yongqin Xian, Subhabrata Choudhury, Yang He, Bernt Schiele, and Zeynep Akata. Semantic projection network for zero- and few-label semantic segmentation. In CVPR, 2019. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 567, + 553, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 620 + ], + "type": "text", + "content": "[56] Yicheng Xiao, Zhuoyan Luo, Yong Liu, Yue Ma, Hengwei Bian, Yatai Ji, Yujiu Yang, and Xiu Li. Bridging the gap: A unified video comprehension framework for moment retrieval and highlight detection. In CVPR, pages 18709-18719, 2024. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 623, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 623, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 316, + 623, + 553, + 666 + ], + "type": "text", + "content": "[57] Bin Xie, Jiale Cao, Jin Xie, Fahad Shahbaz Khan, and Yanwei Pang. Sed: A simple encoder-decoder for open-vocabulary semantic segmentation. In CVPR, pages 3426-3436, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[58] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In NIPS, 2021. 1" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22673" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 386 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[59] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, pages 2955-2966, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[60] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A simple baseline for zero-shot semantic segmentation with pre-trained vision-language model. arXiv preprint arXiv:2112.14757, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 196 + ], + "type": "text", + "content": "[61] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xi-ang Bai. Side adapter network for open-vocabulary semantic segmentation. In CVPR, pages 2945–2954, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 229 + ], + "type": "text", + "content": "[62] Xin Xu, Tianyi Xiong, Zheng Ding, and Zhuowen Tu. Masqclip for open-vocabulary universal image segmentation. In ICCV, pages 887-898, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "type": "text", + "content": "[63] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. NeurIPS, pages 32215-32234, 2023. 1, 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 308 + ], + "type": "text", + "content": "[64] Hui Zhang and Henghui Ding. Prototypical matching and open set rejection for zero-shot semantic segmentation. In ICCV, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "type": "text", + "content": "[65] Haoji Zhang, Yiqin Wang, Yansong Tang, Yong Liu, Jiashi Feng, Jifeng Dai, and Xiaojie Jin. Flash-vstream: Memory-based real-time understanding for long video streams. arXiv preprint arXiv:2406.08085, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 354, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 386 + ], + "type": "text", + "content": "[66] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ADE20K dataset. In CVPR, 2017. 2, 3, 4, 6" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22674" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Stereo Any Video_ Temporally Consistent Stereo Matching/dc20cc87-36ce-4c69-8c57-afd220728c97_content_list.json b/2025/Stereo Any Video_ Temporally Consistent Stereo Matching/dc20cc87-36ce-4c69-8c57-afd220728c97_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..72b24106c63f6833167c61e116df282c7c692bdf --- /dev/null +++ b/2025/Stereo Any Video_ Temporally Consistent Stereo Matching/dc20cc87-36ce-4c69-8c57-afd220728c97_content_list.json @@ -0,0 +1,2140 @@ +[ + { + "type": "text", + "text": "Stereo Any Video: Temporally Consistent Stereo Matching", + "text_level": 1, + "bbox": [ + 202, + 130, + 795, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junpeng Jing Weixun Luo Ye Mao Krystian Mikolajczyk Imperial College London", + "bbox": [ + 240, + 179, + 754, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://tomtomommi.github.io/StereoAnyVideo/", + "bbox": [ + 295, + 219, + 696, + 233 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7c752cb5b818cb77259978201f44bf7ee5c147a62dd04656f4e4508e93bbec34.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 114, + 268, + 888, + 483 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fc00e09b027f94ce0785721c9e8dba417e39d76e95c153da6a3293a6ad84455b.jpg", + "image_caption": [ + "Figure 1. Prediction on a dynamic real-world stereo video. Our method produces temporally consistent and accurate disparities without flickering, enabling the reconstruction of dense and clean metric point clouds for dynamic scenes. See video for better visualization." + ], + "image_footnote": [], + "bbox": [ + 127, + 487, + 246, + 545 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d6c877e62b44edb971474df51bd73253c50018dc3357f739b3e5343b25433b47.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 253, + 488, + 370, + 544 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/62572c3e1fc06cfb7337db396d22c885af6e05b88bb59129c7619448c56eb876.jpg", + "image_caption": [ + "time" + ], + "image_footnote": [], + "bbox": [ + 377, + 488, + 495, + 544 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b6ac54c158a8ea5b5f267eeb44f06556a7c205fe05ada34f25755f7fbfddd87b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 488, + 617, + 544 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2d46a1de5fe1d22655eae1abe29a07699a5e6f98a8bec2ad366c5fa791801589.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 488, + 743, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/82143c2df08f4a6e61c6632aeeeb9f84e71f70de2044b12f6e9a92a38deab6ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 750, + 488, + 867, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 619, + 326, + 635 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper introduces Stereo Any Video, a powerful framework for video stereo matching. It can estimate spatially accurate and temporally consistent disparities without relying on auxiliary information such as camera poses or optical flow. The strong capability is driven by rich priors from monocular video depth models, which are integrated with convolutional features to produce stable representations. To further enhance performance, key architectural innovations are introduced: all-to-all-pairs correlation, which constructs smooth and robust matching cost volumes, and temporal convex upsampling, which improves temporal coherence. These components collectively enhance robustness, accuracy, and temporal consistency, establishing a new standard in video stereo matching. Extensive experiments demonstrate that our method achieves state-of-the", + "bbox": [ + 89, + 650, + 485, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "art performance across multiple datasets both qualitatively and quantitatively in zero-shot settings, as well as strong generalization to real-world indoor and outdoor scenarios.", + "bbox": [ + 511, + 621, + 906, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 709, + 645, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video stereo matching is a task to reconstruct a 3D scene from a stereo camera. It estimates disparity by identifying the horizontal correspondences between rectified left and right images at each time step. Consistent and accurate disparity estimation is crucial for reconstructing and understanding real-world 3D scenes, benefiting various downstream tasks such as autonomous driving, robotic navigation, and virtual/augmented reality.", + "bbox": [ + 511, + 734, + 906, + 854 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the advances in deep learning, stereo matching has seen significant progress [27, 31, 55, 61], achieving promising performance on public benchmarks [13, 44, 45].", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 47 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding author: ye.mao21@imperial.ac.uk", + "bbox": [ + 107, + 886, + 421, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20836", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, most existing approaches are designed for image-based estimation and do not explicitly leverage temporal information in video sequences. When applied directly to video data, these methods often suffer from temporal inconsistencies, leading to noticeable flickering and artifacts in disparity maps and reconstructed point clouds. To mitigate this, recent video-based methods incorporate additional temporal cues, such as camera poses or optical flow, through predefined models or learnable mechanisms. For example, CODD [29] employs a pose network to estimate relative camera movements and align adjacent disparities within a fusion framework. Similarly, TemporalStereo [73] and TC-Stereo [70] follow this paradigm with variations in architectural design and alignment operations. Another approach, BiDAStereo [22], integrates optical flow [51] for temporal feature alignment and introduces a temporal aggregation module to propagate information across time.", + "bbox": [ + 89, + 90, + 480, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite these advances, existing video methods remain vulnerable to dynamic scenes, camera movement, and the accuracy of auxiliary information. In real-world scenarios, dynamic objects undergo complex motions and deformations that violate traditional multi-view constraints [18], disrupting translational invariance in disparity estimation. Moreover, models relying on auxiliary cues are constrained by the performance of camera pose or optical flow estimation modules. In environments with complex camera movements, these estimation modules degrade, becoming bottlenecks that limit the robustness of video stereo methods.", + "bbox": [ + 89, + 351, + 480, + 516 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these limitations, we propose Stereo Any Video, a robust framework for video-based stereo matching without relying on camera pose or optical flow. Inspired by recent findings in video generation [33, 48], which emphasize the role of robust feature representations in ensuring temporal coherence, we take a new perspective to design the framework for consistency: feature robustness and stability. Specifically, during feature extraction, we integrate convolutional cues with a frozen monocular video depth foundation model [7] trained on million-scale data. This ensures robust feature maps with rich priors. For correspondence matching, we introduce an all-to-all-pair correlation, which enforces smooth and reliable matching by jointly considering bidirectional correspondences from both target and reference views. For cost aggregation, we propose temporal convex upsampling within a fully temporal-based recurrent unit, playing a key role in maintaining temporal consistency. As shown in Figure 1, our method demonstrates a strong ability to produce accurate and consistent disparities, benefiting the reconstruction for dynamic scenes.", + "bbox": [ + 89, + 520, + 482, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of our work can be summarized as follows: (1) Stereo Any Video, a framework that integrates monocular video depth priors to achieve accurate and temporally consistent video stereo matching. (2) A novel all-to-all-pair correlation module and a temporal convex upsam", + "bbox": [ + 89, + 825, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "pling mechanism, ensuring robust and stable cost volume representations and aggregation. (3) State-of-the-art performance on diverse benchmarks, setting a new standard for video stereo matching.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 165, + 653, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we provide a brief literature review of stereo matching methods in the deep learning era. We first discuss image-based stereo matching methods, followed by video-based methods, and finally, we review recent advancements leveraging foundation priors.", + "bbox": [ + 511, + 191, + 903, + 266 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image-based Stereo Matching. Since the introduction of the first end-to-end stereo network [34], deep learning-based stereo matching has seen substantial advancements. Early methods typically construct 3D [16, 30, 38, 50, 65] or 4D [5, 9, 14, 25, 62, 67, 71] cost volumes from left and right image features, followed by 2D or 3D convolutions for cost aggregation. Inspired by RAFT [51], iterative refinement has become the mainstream approach. Representative advancements include multilevel recurrent units [31], adaptive correlation [27], geometry-encoding volumes [61, 64], and frequency information selection [55], all of which have significantly improved benchmark performance [13, 44, 45]. Transformer-based methods [15, 28, 49, 58, 66] have further refined stereo matching by modeling token correspondence and capturing long-range dependencies. Beyond accuracy, several methods focus on real-world challenges such as efficiency [17, 56, 60, 63], domain generalization [6, 12, 39, 42, 72], and robustness [21, 46, 47]. However, as all these methods are designed for static image pairs, they lack temporal awareness, leading to inconsistent disparity predictions in video sequences.", + "bbox": [ + 511, + 267, + 903, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video-based Stereo Matching. Video stereo matching especially in dynamic scenes presents unique challenges. Early work, such as [75], employed LSTMs with unsupervised learning for stereo videos but did not explicitly incorporate temporal information. CODD [29] introduced a per-frame stereo network, a motion network, and a fusion network designed for dynamic scenes. Neighboring-frame disparities are aligned and fused using estimated camera poses. TemporalStereo [73] extended this approach with a coarse-to-fine network that leverages past context to enhance predictions in challenging scenarios. Cheng et al. [10] developed a framework tailored for XR systems, reducing computational costs through temporal cost aggregation. TC-Stereo [70] further introduced a temporal disparity completion mechanism and a temporal state fusion module to exploit temporal information. These methods rely on auxiliary camera motion or pre-defined scene geometry to incorporate temporal information. DynamicStereo [24] introduced a transformer-based architecture that integrates temporal, stereo, and spatial attention. BiDAStereo [22] proposed an optical-flow-based alignment mechanism to propagate", + "bbox": [ + 511, + 584, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20837", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b0642654f4f12eaa27541d472af4f24e2d80d5eb84e1e5d6257aae5ee4222608.jpg", + "image_caption": [ + "Figure 2. Overview of Stereo Any Video. Given a stereo video sequence as input, the model first extracts features and context information using trainable convolutional encoders and a frozen monocular video depth encoder (Video Depth Anything [7]). At each iteration, an all-to-all-pair correlation computes feature correlations, followed by a multilayer perception (MLP) encoder for compact representation. Disparities are iteratively refined using a 3D Gated Recurrent Unit (GRU), integrated with cost volume features, and upsampled via a temporal convex upsampling layer. This process is repeated within a cascaded pipeline to progressively recover full-resolution disparities." + ], + "image_footnote": [], + "bbox": [ + 133, + 89, + 864, + 292 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "temporal information across the entire sequence. Building on this alignment, BiDAS stabilizer [23] introduced a lightweight plug-and-play stabilizer network to improve the temporal consistency of existing image-based stereo matching methods.", + "bbox": [ + 88, + 388, + 480, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Stereo Matching with Foundation Priors. Vision foundation models, such as CLIP [41], the DINO series [4, 37], and SAM [26], have been widely applied across various vision tasks due to their rich priors. In stereo matching area, FormerStereo [74] proposed feature transformation and C4 space to adapt ViT-based foundation model to stereo pipeline, learning domain invariant representations. AIOStereo [76] introduced a selective knowledge transfer module to distill knowledge from multiple heterogeneous vision foundation models into a single stereo matching model. More recent approaches primarily focused on leveraging monocular depth priors from Depth Anything [68, 69] to enhance stereo performance. For instance, Stereo Anywhere [1] directly incorporated relative depth maps into the stereo network, demonstrating robustness in challenging conditions such as non-Lambertian surfaces. DEFOM-Stereo [20] employed these depth maps alongside a novel scale updater within a recurrent module for improved alignment. FoundationStereo [59] extracted feature maps from Depth Anything and introduced a side-tuning adapter to adapt monocular features for stereo backbone. MonSter [8] utilized both depth maps and feature maps, and proposed a dualbranched refinement module to fuse stereo and monocular cues effectively. These methods highlight the effectiveness of foundation priors in image-based stereo tasks. In our work, we build upon this direction by investigating video-based stereo matching and exploring how foundation priors can enhance consistency in dynamic scenes.", + "bbox": [ + 91, + 477, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 387, + 604, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce Stereo Any Video, a feedforward video-based framework. Given a rectified stereo sequence $\\{\\mathbf{I}_L^t,\\mathbf{I}_R^t\\}_{t\\in (1,T)}\\in \\mathbb{R}^{H\\times W\\times 3}$ where $T$ denotes the total number of frames, $H$ and $W$ are the frame height and width, respectively, our goal is to estimate a temporally consistent sequence of disparity maps $\\{\\mathbf{d}^t\\}_{t\\in (1,T)}\\in \\mathbb{R}^{H\\times W}$ aligned with the left view. As shown in Figure 2, our method consists of three key components: feature extraction (Section 3.1), correlation (Section 3.2), and temporal cost aggregation (Section 3.3).", + "bbox": [ + 511, + 412, + 906, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Feature Extraction with Foundation Priors", + "text_level": 1, + "bbox": [ + 511, + 575, + 879, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike image-based tasks, maintaining temporal consistency is the primary challenge in video stereo matching. Recent studies in video generation [33, 48] indicate that substantial variations across different frames can lead to unstable feature representations, thereby degrading temporal coherence. Therefore, ensuring feature stability across adjacent frames is crucial for achieving temporally consistent disparity estimation. Towards this, we leverage Video Depth Anything (VDA) [7] for feature extraction, which offers two key advantages: (1) trained on internet-scale data, VDA provides rich priors that can enhance disparity estimation in regions with ambiguous stereo correspondence; (2) unlike image-based monocular depth models [68, 69], VDA generates temporally stable features through an explicit temporal alignment process before the final prediction head, thereby preserving coherence across frames.", + "bbox": [ + 511, + 598, + 906, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Besides, we investigate whether the depth maps from monocular video depth models can further improve stereo performance. We compare DepthCrafter [19], a representative monocular video model, with RAFTStereo [31], an", + "bbox": [ + 511, + 840, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20838", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/fb3e94e26981cda50bb30177188bc5156279e51dedf9c1170b229c46f0162601.jpg", + "table_caption": [ + "Table 1. Quantitative comparison of DepthCrafter [19] and RAFT-Stereo [31] on Sintel dataset [2]." + ], + "table_footnote": [], + "table_body": "
MethodSintel Clean
EPE ↓TEPE ↓MOS (human) ↓
DepthCrafter [19]8.682.080.212
RAFTStereo [31]1.420.840.933
", + "bbox": [ + 114, + 119, + 452, + 189 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7a70cb3d576351db04a2ea923a546ed19c954cdfdcdb42c0361bf141f1382948.jpg", + "image_caption": [ + "Input bamboo 2" + ], + "image_footnote": [], + "bbox": [ + 102, + 195, + 222, + 236 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/31f7cbdab9f83f7e21c42cf20500bbf76d372bb4821712726d60e0b428a43b00.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 195, + 346, + 236 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f95bb896f959ab1811c60258011db38988420105d2b9d06816b585a0b4a397e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 346, + 195, + 470, + 236 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6356f26fb1ae72c2f4334278f83bb72f1e390dc07642cb5c000e74b9bf8e28b3.jpg", + "image_caption": [ + "Input sleeping_1", + "Figure 3. Qualitative comparison of DepthCrafter [19] and RAFT-Stereo [31] on Sintel dataset [2]." + ], + "image_footnote": [], + "bbox": [ + 102, + 243, + 222, + 282 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0c98399bcd7f755bc365c3bbb13e7114752fdbd7b20b399494de856b3db596ea.jpg", + "image_caption": [ + "DepthCrafter", + "DepthCrafter" + ], + "image_footnote": [], + "bbox": [ + 223, + 243, + 344, + 282 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0d8c45cfda0c6876d215645a2237c716f33d00dc0669ba0a0018f750fc3e0058.jpg", + "image_caption": [ + "RAFTStereo", + "RAFTStereo" + ], + "image_footnote": [], + "bbox": [ + 346, + 243, + 470, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "image-based stereo model. Using open-source weights, we evaluate their performance on Sintel [2], as shown in Table 1 and Figure 3. Specifically, we first obtain relative depth maps from DepthCrafter (without scale and shift) and compute disparities using RAFTStereo on the same sequence. The relative depth maps are then aligned to RAFTStereo's disparities, and objective metrics are computed against the ground truth. Additionally, we conduct a subjective user study (see supplementary material for details). As observed, while DepthCrafter produces visually plausible results—reflected in better human scores—its spatial and temporal errors are significantly larger than those of RAFTStereo. This suggests that its relative depth is \"consistent but wrong\", making it unreliable for directly improving consistency within a stereo framework. Therefore, we only utilize its features as valuable priors.", + "bbox": [ + 88, + 338, + 482, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Figure 2, our feature extraction module consists of two trainable encoders, each employing a series of residual blocks [31] to extract convolutional image features and context features. To integrate VDA priors, we adopt a shallow convolutional adapter [59] that downsamples depth features from the frozen VDA. The extracted depth features are then concatenated with convolutional features to form $\\{\\mathbf{F}_L,\\mathbf{F}_R\\}$ and $\\mathbf{F}_C$ , providing more stable feature representations for subsequent processing.", + "bbox": [ + 89, + 582, + 482, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. All-to-all-pair Correlation", + "text_level": 1, + "bbox": [ + 89, + 733, + 328, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At the $n$ -th iteration for each time step, the right feature $\\mathbf{F}_R$ is first warped toward the left feature $\\mathbf{F}_L$ using the previous disparity $d_{n-1}$ , producing $\\hat{\\mathbf{F}}_R$ . Previous methods [31, 61] typically construct cost volumes $\\mathbf{C}_n$ via all-pair correlation, employing a lookup mechanism to search for potential matches within a local radius. In [27], this approach was extended to an alternating local 1D and 2D search:", + "bbox": [ + 89, + 757, + 482, + 863 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} _ {n} (x, y) = \\left\\langle \\mathbf {F} _ {L} (x, y), \\hat {\\mathbf {F}} _ {R} (x + r _ {x}, y + r _ {y}) \\right\\rangle , \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 878, + 482, + 905 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a6241fd41074fcb9c76091307049e3fb33864d44ed4e05e3c5eeacc11173b1ee.jpg", + "image_caption": [ + "Figure 4. Workflow of the temporal convex upsampling. Each pixel of the high resolution disparity (yellow box) is calculated via its 27 low resolution neighbors using learnable weights." + ], + "image_footnote": [], + "bbox": [ + 526, + 90, + 893, + 251 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $(x,y)$ denotes the position in $\\mathbf{F}_L$ , $\\langle \\cdot, \\cdot \\rangle$ represents the channel-wise product operation, and $(r_x, r_y)$ defines the search range, typically set as $(r_x, r_y) \\in \\{(\\pm 4,0), (\\pm 1, \\pm 1)\\}$ . The fundamental principle of these methods is to construct one-directional correspondence, where a reference point is selected and directly matched with target points via multiplication. Inspired by recent advances in point tracking [11], we introduce bidirectional correspondence and develop a local all-to-all-pair correlation. Instead of computing correlation in a fixed direction, we calculate the similarity between every pair of potential matching points within two search windows, formulated as:", + "bbox": [ + 511, + 306, + 906, + 489 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} _ {n} (x, y) = \\left\\langle \\mathbf {F} _ {L} \\left(x + r _ {x}, y + r _ {y}\\right), \\hat {\\mathbf {F}} _ {R} \\left(x + r _ {x}, y + r _ {y}\\right) \\right\\rangle . \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 510, + 903, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This approach enhances match verification and reduces ambiguity [32, 43]. While BiDAStereo [22] constructs temporal cost volumes by aligning adjacent frames for consistency, its performance is significantly constrained by the limitations of optical flow. In contrast, our method emphasizes robust spatial correlation, leveraging dense correspondences to enforce matching smoothness and enhance consistency and robustness across neighboring points [52, 53].", + "bbox": [ + 511, + 544, + 905, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Temporal Cost Aggregation", + "text_level": 1, + "bbox": [ + 511, + 674, + 763, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At the $n$ -th iteration, the all-to-all-pair correlation expands the potential matching points in $\\mathbf{C}_n$ , increasing its dimensionality from $\\mathbb{R}^{H' \\times W' \\times (2r_x + 1)(2r_y + 1)}$ to $\\mathbb{R}^{H' \\times W' \\times (2r_x + 1)^2(2r_y + 1)^2}$ , where $H'$ and $W'$ denote the height and width of frame feature, respectively. We first encode $\\mathbf{C}_n$ into a compact representation $\\mathbf{E}_n \\in \\mathbb{R}^{H' \\times W' \\times L}$ using a MLP:", + "bbox": [ + 511, + 696, + 905, + 801 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {E} _ {n} = \\operatorname {M L P} \\left(\\mathbf {C} _ {n}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 648, + 803, + 903, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we employ a 3D-GRU [24] to iteratively refine the disparity for each frame while preserving temporal consistency. The GRU aggregates information using separable 3D convolutions across both spatial and temporal dimensions. At iteration $n$ , it takes as input the encoded cost volume $\\mathbf{E}_n$", + "bbox": [ + 511, + 824, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20839", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/5b73c343c25623f264b13f3a0e4565677b0a17c7ee8573f8ec7dd731b2887524.jpg", + "table_caption": [ + "Table 2. Zero-shot generalization results. All methods were trained only on SceneFlow [34]. The representative image-based and video-based stereo matching models are compared on Sintel [2], Dynamic Replica [24], Infinigen SV [23], and Virtual KITTI2 [3]. The weights and parameters are fixed for evaluation. The best and second-best results are highlighted." + ], + "table_footnote": [], + "table_body": "
MethodSintel FinalDynamic ReplicaInfinigen SVVirtual KITTI2
TEPE↓δt1px↓EPE↓TEPE↓δt1px↓EPE↓TEPE↓δt1px↓EPE↓TEPE↓δt1px↓EPE↓
Image-based
RAFTStereo [31]2.0913.704.890.1452.020.382.3319.803.771.2518.391.45
IGEVStereo [61]1.7412.652.440.0810.800.202.3519.333.811.4218.881.39
Selective-IGEV [55]1.9612.952.820.1331.290.292.8521.835.691.4119.101.38
DEFOM-Stereo [20]2.3312.432.380.0800.750.132.0118.913.241.0716.631.34
MonSter [8]1.7011.063.300.1570.860.451.6515.732.270.7310.930.67
Video-based
DynamicStereo [24]1.4311.804.260.0830.730.241.9815.684.671.0615.631.47
BiDAStereo [22]1.2611.651.990.0911.040.651.9916.506.521.0214.821.43
Stereo Any Video (ours)1.079.981.750.0570.670.251.6514.273.160.7412.451.01
", + "bbox": [ + 91, + 132, + 903, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "context feature $\\mathbf{F}_c$ , and the disparities from the previous iteration $d_{n-1}$ . To enhance representation, we integrate super kernels [22], temporal and spatial attention [24] in this process. The disparity update in 3D-GRU is as follows:", + "bbox": [ + 89, + 311, + 483, + 375 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nx _ {n} = \\left[ \\operatorname {C o n v 2 d} \\left(\\mathbf {E} _ {n}\\right), \\operatorname {C o n v 2 d} \\left(d _ {n - 1}\\right), d _ {n - 1}, \\mathbf {F} _ {c} \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 381, + 480, + 400 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {x _ {n}} = \\operatorname {A t t e n t i o n} \\left(x _ {n}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 401, + 482, + 419 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {n} = \\sigma \\left(\\operatorname {C o n v 3 d} _ {z} \\left(\\left[ h _ {n - 1}, \\widetilde {x _ {n}} \\right]\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 421, + 482, + 445 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr _ {n} = \\sigma \\left(\\operatorname {C o n v 3 d} _ {r} \\left(\\left[ h _ {n - 1}, \\widetilde {x _ {n}} \\right]\\right)\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 449, + 482, + 474 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq _ {n} = \\tanh \\left(\\operatorname {C o n v} 3 \\mathrm {d} _ {q} \\left(\\left[ h _ {n - 1} \\odot r _ {n}, \\widetilde {x _ {n}} \\right]\\right)\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 477, + 482, + 502 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh _ {n} = \\left(1 - z _ {n}\\right) \\odot h _ {n - 1} + z _ {n} \\odot q _ {n}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 505, + 482, + 521 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd _ {n} = d _ {n - 1} + \\operatorname {C o n v} 3 \\mathrm {d} \\left(h _ {n}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 523, + 482, + 540 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $[\\cdot ]$ is concatenation, $\\odot$ denotes element-wise multiplication, and $\\sigma$ is the sigmoid activation function.", + "bbox": [ + 89, + 549, + 482, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While prior methods emphasize temporal consistency in correlation and cost aggregation, the upsampling process remains overlooked, despite its crucial role in achieving accurate and temporally consistent disparities. Towards this, we adapt the commonly used convex upsampling approach [51] to a temporal variant. In Figure 4, each high-resolution disparity pixel (yellow box) is computed as a convex combination of its 27 low resolution neighbors across the 3 frames (green, blue, and faint yellow boxes). The network predicts the corresponding $3 \\times 3 \\times 3$ weights using 3D convolutions, enabling a temporally consistent upsampling:", + "bbox": [ + 89, + 579, + 482, + 746 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {w} = \\operatorname {s o f t m a x} \\left(\\operatorname {C o n v 3 d} \\left(h _ {n}\\right)\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 753, + 482, + 779 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nD _ {n} ^ {t} = \\alpha \\cdot \\sum_ {i j k} \\mathbf {w} _ {i j k} \\odot \\text {u n f o l d} \\left(\\left[ d _ {n} ^ {t - 1}, d _ {n} ^ {t}, d _ {n} ^ {t + 1} \\right]\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 782, + 482, + 815 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ is upsampling rate, $D_{n}^{t}$ is the upsampled disparity.", + "bbox": [ + 89, + 824, + 482, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Loss Function", + "text_level": 1, + "bbox": [ + 89, + 848, + 236, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recent works have introduced OPW loss [57] and TGM loss [7] in monocular video depth estimation to enhance", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "temporal consistency. However, we found out that these losses impose a trade-off between accuracy and consistency without yielding substantial improvement in the video stereo task. Thus, we opt for a purely image-based $L_{1}$ loss while still achieving strong temporal consistency as:", + "bbox": [ + 511, + 311, + 906, + 388 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\sum_ {t = 1} ^ {T} \\sum_ {n = 1} ^ {N} \\gamma^ {N - n} \\left| \\left| D _ {\\mathrm {g t}} ^ {t} - D _ {n} ^ {t} \\right| \\right|, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 398, + 906, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $T$ is the number of frames, $N$ is the number of iterations, and $\\gamma$ is a weighting coefficient set to 0.9. Here, $D_{\\mathrm{gt}}^t$ denotes the ground truth disparity for the $t$ -th frame. The final output is obtained from the last iteration prediction.", + "bbox": [ + 511, + 450, + 906, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 526, + 645, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first introduce the datasets and metrics used in the evaluation, as well as implementation details. We then present the zero-shot generalization capabilities of our method, followed by ablation studies to validate the efficacy of each component in our framework. Please refer to the supplementary material for more details.", + "bbox": [ + 511, + 551, + 906, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Benchmark Datasets and Metric", + "text_level": 1, + "bbox": [ + 511, + 652, + 799, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. Our work focuses on videos captured with moving cameras, making commonly used image benchmarks unsuitable, such as Middlebury [44], ETH3D [45], and KITTI 2012/2015 [13] (two frames per scene). To evaluate our method, we use five synthetic and two real-world stereo video datasets, all featuring dynamic scenes. Sintel [2] is a synthetic dataset comprising movie sequences available in both clean and final passes. Spring [35] is a high resolution synthetic dataset. Dynamic Replica [24] is a synthetic indoor dataset featuring non-rigid objects, such as animals and people. Infinigen SV [23] is a synthetic dataset focusing on outdoor natural environments. Virtual KITTI2 [3] is a synthetic dataset simulating outdoor driving scenarios. KITTI Depth [54] is a real-world outdoor dataset captured for autonomous driving applications, with sparse", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20840", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f21618afa7672f5b7d4be33241f98f3b1b2781b4fd65bead2e2c1df713c7fefc.jpg", + "table_caption": [ + "Table 3. Quantitative comparisons on Spring [35], Sintel [2], and KITTI Depth [54] where methods are allowed to train on any existing datasets. Model weights remain fixed during evaluation. Training datasets sharing the same domain as the test sets are highlighted. “-” represents the test set has been used in training. Abbreviations: SF - SceneFlow [34], DR - Dynamic Replica [24], ISV - Infigen SV [23], M - Middlebury [44], KD - KITTI Depth [54], VK - Virtual KITTI2 [3]. The best and second-best results are highlighted." + ], + "table_footnote": [], + "table_body": "
MethodTraining DataScaleSpringSintel FinalKITTI Depth
TEPE↓δt1px↓EPE↓TEPE↓δt1px↓EPE↓TEPE↓δt1px↓EPE↓
Image-based
RAFTStereo [31]10 mixed datasets0.6M1.499.795.37---0.417.960.51
IGEVStereo [61]SF+M0.04M1.539.725.371.5313.492.230.469.120.60
Selective-IGEV [55]SF+M0.04M1.469.335.461.6413.782.140.489.660.60
DEFOM-Stereo [20]15 mixed datasets0.7M1.299.124.25---0.397.080.48
MonSter [8]10 mixed datasets0.6M1.258.655.19---0.417.730.60
FoundationStereo [59]7 mixed datasets1.3M1.789.825.40---0.407.200.57
Video-based
DynamicStereo [24]SF+DR+ISV+KD0.1M1.079.649.361.3811.843.380.428.210.57
BiDAStereo [22]SF+DR+ISV+KD0.1M0.907.903.921.3311.341.970.428.330.58
Stereo Any Video (ours)SF+DR+ISV+VK0.1M0.776.523.170.999.091.480.356.050.61
", + "bbox": [ + 91, + 148, + 906, + 328 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "depth maps obtained from a LiDAR sensor. South Kensington SV [23] is a real-world dataset capturing daily scenarios from a stereo camera for qualitative evaluation.", + "bbox": [ + 89, + 345, + 482, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metric. For accuracy evaluation, we use End-Point Error (EPE), which measures the average per-pixel disparity error. To assess temporal consistency, we employ Temporal EPE (TEPE) [24], which quantifies the variation in EPE across the temporal dimension. We also compute $\\delta_{n\\text{-px}}^t$ , which represents the proportion of pixels with TEPE exceeding $n$ . For all metrics, lower values indicate better performance.", + "bbox": [ + 89, + 391, + 483, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 508, + 308, + 525 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We implement Stereo Any Video in PyTorch and train on NVIDIA A100 GPUs. The model is trained with a batch size of 8 on SceneFlow [34] for $120k$ iterations and finetuned on a mixed dataset for $80k$ iterations consisting of Dynamic Replica [24] training set, Infinigen SV [23] training set, and Virtual KITTI2 [3]. Before being fed into the network, sequences are randomly cropped to $256 \\times 512$ . We use the AdamW optimizer with a learning rate of $2e^{-4}$ and apply a one-cycle learning rate schedule. The full training process takes approximately 6 days. Following [24], we apply various data augmentation techniques, including random cropping, rescaling, and saturation shifts. The sequence length is set to $T = 5$ during training, and $T = 20$ for evaluation. The number of iterations in GRU is set to $N = 10$ in training, and $N = 20$ in evaluation.", + "bbox": [ + 89, + 532, + 483, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Evaluation", + "text_level": 1, + "bbox": [ + 89, + 771, + 210, + 786 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Results. Table 2 presents the results when models are trained solely on SceneFlow [34] and evaluated on four dynamic synthetic datasets including diverse indoor and outdoor environments. The results clearly demonstrate that our model achieves state-of-the-art zero-shot generalization across nearly all metrics, excelling in both accuracy and temporal consistency. Table 3 reports the results", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "when models are allowed to train on any existing datasets. Notably, despite being trained solely on synthetic datasets, our model surpasses comparison methods that are trained on real-world datasets or share the same domain as the test tests by at least $15\\%$ on diverse benchmarks. Specifically, compared to the robust version of RAFT-Stereo, which is trained on 10 mixed datasets, including in-domain sets such as Sintel [2] and KITTI [13], our approach achieves superior performance across most metrics. These results highlight the strong generalization capability of our method when applied to unseen domains.", + "bbox": [ + 511, + 345, + 906, + 511 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Results. Figure 5 and Figure 6 present qualitative results using the checkpoints from Table 3 on dynamic real-world stereo videos from the South Kensington SV dataset [23]. For the indoor scenario shown in Figure 5, the upper rows display three input left frames along with their predicted disparities, while the subsequent rows visualize the globally aligned reconstructed point cloud. The red arrows highlight that our method produces temporally consistent and accurate disparities, whereas other methods exhibit noticeable artifacts. Specifically, Selective-IGEV and DynamicStereo suffer from flickering artifacts, with cabinets and walls displaying low-frequency oscillations. BiDAStereo fails to maintain an accurate shape, resulting in a distorted table. For the outdoor scenario in Figure 6, the driving car presents a challenging case due to its transparent surfaces and rapid motion. As shown, all other methods struggle with this case, failing to produce reliable disparity estimations. In contrast, our method demonstrates superior performance, maintaining accurate and stable predictions.", + "bbox": [ + 511, + 513, + 908, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 816, + 666, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Foundation Priors. As shown in Table 4, incorporating foundation priors from a monocular depth model improves performance. Notably, VDA [7] outperforms DepthAnythingV2 [68] due to its stronger temporal consistency. In", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20841", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/736c3fda992504b6dc731ac7da637f6d8941eb65923948b509378eab1b34dbb4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 143, + 89, + 261, + 141 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f160bca8a8e4544ba702f3a377734cc92121b30eedd4d6716b0233c7317f5c08.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 147, + 143, + 258, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9a565c3341946218f42901cd9a98f1610e93a7b30c14211333ae6d03cdc28801.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 148, + 169, + 258, + 195 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/44288cbd317c07b07ba5af3cd445785637d0ac5ed8d36a74199fe7ecf91b78b9.jpg", + "image_caption": [ + "Left Frames" + ], + "image_footnote": [], + "bbox": [ + 147, + 196, + 258, + 220 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b7488bb0df63d97a0e0d2df13475336313207a18ea7be9f5198ecd2d584b60e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 264, + 90, + 411, + 194 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d44f8970934e84c5523e58e54e206acdd7a26bf2aa7900b04d73d57af5942749.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 269, + 199, + 411, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/85dbea3b21284cdf3055372e945442dac28f8d1b50f8beb38c8839304edbe66d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 271, + 236, + 411, + 272 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ea10315e139192c6cbe6c8d6c96e675719601c174a345938717bcaf990e1d9a2.jpg", + "image_caption": [ + "Selective-IGEV" + ], + "image_footnote": [], + "bbox": [ + 272, + 272, + 411, + 306 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d98300b824e70c39d81ed932274f6dee03db0533785a3cd9f6dc3775dcda0d08.jpg", + "image_caption": [ + "Point Cloud" + ], + "image_footnote": [], + "bbox": [ + 143, + 247, + 267, + 297 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6fbd86428f71a0bb34023107057bbe11ba5bb9cf50283e35bad7f3994d267e26.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 90, + 558, + 194 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ff3d60f55a48ddfd65e12485e87233d2e9123539a4b332ca9850e9cec3cd6bfb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 199, + 558, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d0ce17c9f8872f60798012105e1962fb4b202bf7b14badd936987b3694ea2009.jpg", + "image_caption": [ + "DynamicStereo" + ], + "image_footnote": [], + "bbox": [ + 416, + 236, + 558, + 308 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f70935bd9499c1994e96b4ca62cd1769600fad45083657ae2d7b9367a299735f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 90, + 705, + 194 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7ace61094145e86490fcde360569048724eb0f76e67f2a905b63076d519074da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 563, + 199, + 705, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/effb0f15aa8535a7a9f4eb7dca62643cacb325ace47f162d8f5d96364b658bf8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 563, + 236, + 705, + 272 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fbc1a3a078da0899f25f0ccf9759f9b048af29d51530011b7d221e2fcb1f8500.jpg", + "image_caption": [ + "BiDAStereo" + ], + "image_footnote": [], + "bbox": [ + 563, + 272, + 705, + 308 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/619395f338c6137cad9f0e295e9f51bcd9d2e2278fcca234655b801a618d4965.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 90, + 852, + 125 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3b04867844d804312a68fe52dedb2d3a12364c35f9b7e51170ff6381a383e22d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 125, + 852, + 160 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8e75d31ec343f7b3aad94256509e56a684d3e991cf9871b6616afd8f50478e93.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 161, + 852, + 194 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d89ba3d7ac560c0ce86eb1d1f844bbbdeadc5dcd43b7e4e7f54e94d3b6c03a12.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 202, + 852, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/258905af26531a4384573582b2453fb22b0b051832302d08403475c7f09e88ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 236, + 852, + 272 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1a44549eeef01ed5e42cacff907c027ff01a812aa025558e70f6744f3c699a46.jpg", + "image_caption": [ + "Stereo Any Video (ours)" + ], + "image_footnote": [], + "bbox": [ + 710, + 272, + 852, + 308 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/11369eeeffbd1795bfc5875220b437ec0819aaabf620332545a53e83dc2a4418.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison on a dynamic indoor scenario from the South Kensington SV dataset [23]. Top: Input left frames and predicted disparities for the same region across three frames using different methods. Bottom: Disparities converted into globally aligned point clouds and rendered from a camera view for comparison. Failure regions are highlighted with red arrows. Our method produces consistent and accurate disparities without flickering. See video in the supplementary material for better visualization." + ], + "image_footnote": [], + "bbox": [ + 142, + 401, + 285, + 463 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ef3895ca7083cb8da6bc7e3838c797de5cfa34b5dd9a4cd8a21d392ae89729cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 143, + 464, + 284, + 527 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/029f957fb7a5dadd02fd8ae654dd365d50eae7dfaf63d1e680dd59b1b8db61c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 143, + 527, + 284, + 589 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3688cad9176a35789c6b6570dc82f8ce9d01f7ec5b65d8d08804543ee24b2af2.jpg", + "image_caption": [ + "Left Frames" + ], + "image_footnote": [], + "bbox": [ + 143, + 590, + 284, + 652 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cb145eceb4f686bffdba1926ab0a694e990279c00214f4652de1eff0b22d148b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 401, + 426, + 464 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bd8cc6cd940f7df338036be6fc2581f5e137bc774e97fb6234b406b2fd68c5a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 465, + 426, + 527 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2ad3a7daffc029ae491509083022a2d3c9b8d7150ccc7a37f4362ebb0ffd15a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 527, + 426, + 590 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/383f8ace1592521984b26dd8727dda2b7ce96d76fe7c207413f17bff1ba6d308.jpg", + "image_caption": [ + "Selective-IGEV" + ], + "image_footnote": [], + "bbox": [ + 287, + 590, + 426, + 652 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a259d02c94925c90bc21ef9738c90e24db2646d10b30c7073955d9273c23d6fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 426, + 401, + 568, + 464 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c698e9275cb9d9c2920f0cab98334d62bb980da10afb475a9d7e14853b9076ec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 426, + 465, + 568, + 527 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/20bbc549b77dfdfcad5b3163cf839b71ac0c3f61f16fb175677d2d447d408a1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 426, + 527, + 568, + 590 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b7ed7704e7c68d377d1939cd8ca964db0189d839b7ee4e118b2591a3f77d90a4.jpg", + "image_caption": [ + "DynamicStereo" + ], + "image_footnote": [], + "bbox": [ + 426, + 590, + 568, + 652 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9e6dc2b7bb5e82e1ed53812e7ee25e8fb09b95a60e3d51be84480b6a85325765.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 568, + 401, + 710, + 464 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/65b7044e8cad1cc88607d5bea97f09c46dd48d6da791b5b6897975e09f04df5a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 568, + 465, + 710, + 527 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ff651e9d0149255bdee8f1895e3b864f5fd2b325a04b6960320a8e872cc36265.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 568, + 527, + 710, + 590 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b199b1546b977a937308f01d412b3ca5b9ade1941e5e15d426267fadcbf3b375.jpg", + "image_caption": [ + "BiDAStereo", + "Figure 6. Qualitative comparison on a dynamic outdoor scenario from the South Kensington SV dataset [23]." + ], + "image_footnote": [], + "bbox": [ + 568, + 590, + 710, + 652 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/69cffa071f143597de1bea78beb975b0d645443dd3abb8b5c8d82eb8cf2b112d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 401, + 852, + 464 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bc471e13c904b3476c0b61295341317f454b3a536c252d6b82516dbf7929506f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 465, + 852, + 527 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/11259b7856e788a1da1da1e4ededf85c0dad292c9b26232cf19da5a0a3392e13.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 527, + 852, + 590 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/66567c4e513ab1c9adf73b6221151e9b3a6b8352ac670f4f69a23c99e1632cb9.jpg", + "image_caption": [ + "Stereo Any Video (ours)" + ], + "image_footnote": [], + "bbox": [ + 710, + 590, + 852, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "terestingly, increasing the backbone size does not yield additional performance gains. To investigate this, we conducted further experiments (Table 5) by varying depth and CNN feature channels based on VDA-L. However, the results remained similar. According to [7], VDA-L has $13 \\times$ more parameters than VDA-S (381.8M vs. 28.4M), yet the performance gain is marginal ( $\\delta_{1}(\\uparrow)$ : 0.944 vs. 0.942 on KITTI; 0.971 vs. 0.959 on NYUv2). This suggests that the larger backbone does not provide richer priors. Moreover, large backbone requires additional CNN channels and parameters for adaptation, increasing training complexity. Therefore, we adopt VDA-S in the final model.", + "bbox": [ + 89, + 715, + 483, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Correlation. To evaluate the effectiveness of our proposed all-to-all-pairs correlation, we replace it with local correlation [27] and experiment with using only 1D or 2D correlation instead of alternating between them. As shown in Table 4, all these modifications degrade performance. Notably, constructing 3D correlations—where the search window spans height, width, and time—further worsens results. This suggests that frames cannot be effectively utilized without proper alignment [23].", + "bbox": [ + 511, + 715, + 906, + 852 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cost Volume Encoder. Instead of directly encoding cost volumes with convolutions, we find that introducing an MLP layer to extract a compact representation before ap", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20842", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6c01104a90d18c151e7ee8638e9fbb20d6656e02d2b54be447894c030dc8159e.jpg", + "table_caption": [ + "Table 4. Ablation study on the foundation priors, correlation, cost volume encoder, convex upsampling, attention, and inference settings of our model on Sintel [2] and Dynamic Replica [24]. The approach used in our final model is underlined." + ], + "table_footnote": [], + "table_body": "
Setting VariationsSintel FinalDynamic Replica
TEPE↓δt1px ↓TEPE↓δt1px ↓
Foundation Priors
w/o Depth1.4712.350.0921.05
Depth AnythingV2-S [68]1.3311.270.0880.99
VDA-S [7]1.2710.980.0830.78
VDA-L [7]1.2012.470.0830.82
Correlation
1D + 2D local [27]1.2710.980.0830.78
1D all-to-all-pairs1.2311.360.0790.78
2D all-to-all-pairs1.2911.950.0850.77
3D all-to-all-pairs4.5122.901.3132.01
1D + 2D all-to-all-pairs1.2111.330.0760.75
Cost Volume Encoder
Convolution1.2611.540.0810.77
MLP + Convolution1.2111.330.0760.75
Convex Upsampling
Bilinear1.2612.740.0851.05
Convex [51]1.2111.330.0760.75
Temporal Convex1.0410.210.0670.79
Attention in GRU
w/o Attention1.0410.210.0670.79
Temporal1.0710.230.0640.79
Temporal + Spatial1.079.980.0570.67
Inference Stage
1/16 → 1/8 → 1/41.079.980.0570.67
1/32 → 1/16 → 1/8 → 1/41.0810.820.0580.70
", + "bbox": [ + 91, + 148, + 480, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "plying convolutions enhances cost aggregation, leading to improved final results.", + "bbox": [ + 89, + 515, + 482, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Convex Upsampling. We compare bilinear upsampling and general convex upsampling [51] to our proposed temporal convex upsampling module. As observed, our approach achieves better temporal consistency by incorporating learnable weights with adjacent frames.", + "bbox": [ + 89, + 547, + 482, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Attention in GRU. We further investigate different attention mechanisms within the GRU update block. As shown in Table 4, combining temporal and spatial attention together leads to the best performance.", + "bbox": [ + 89, + 627, + 482, + 688 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inference Stage. We also explore inference starting at different resolutions. As can be seen in Table 4, a three-stage inference setup already achieves strong results.", + "bbox": [ + 89, + 690, + 482, + 737 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Application", + "text_level": 1, + "bbox": [ + 89, + 753, + 217, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method enhances downstream applications that rely on accurate metric depth, such as 3D point tracking. As shown in Figure 7, we visualize results using the state-of-the-art 3D point tracking method DELTA [36]. Our predicted metric depth yields more points with precise trajectories than UniDepth [40], which gives less missing points as highlighted by the red arrows. Additional application, such as visual effects, are available in the supplementary material.", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c0bdf386ff01b859ba62db2547b4deb9977cab337786cd1ff7ddd451cdc1d662.jpg", + "table_caption": [ + "Table 5. Ablation study on the effect of different channel configurations for depth features and CNN features, based on VDA-L [7], on Sintel [2]. Trainable parameters are also reported." + ], + "table_footnote": [], + "table_body": "
Depth & CNNSintel FinalPara. (M)
Channel NumberTEPE↓δt1px ↓
VDA-L [7]
(32, 96)1.2012.478.4
(128, 64)1.5814.5213.3
(128, 128)1.3113.2119.3
UniDepthStereo Any Video (ours)
", + "bbox": [ + 544, + 133, + 870, + 383 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/90834854a545c3bc71a285b0d0ffb8836b99c5c11bde7b33f1cc1f3eee8c972e.jpg", + "table_caption": [ + "Figure 7. Visualization of 3D point tracking based on metric depth generated from UniDepth and our method.", + "Table 6. Comparison of different methods on model trainable [non-trainable] parameters, inference GPU memory, and multiply-accumulate (MAC) on 20 frames with resolution $720 \\times 1280$ ." + ], + "table_footnote": [], + "table_body": "
MethodsPara. (M)Mem. (G)MACs (T)
DynamicStereo [24]20.535.1182.3
BiDASTereo [22]12.241.1186.6
Stereo Any Video (ours)9.4 [28.4]41.1303.4
", + "bbox": [ + 514, + 478, + 903, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Limitation and Future Work", + "text_level": 1, + "bbox": [ + 511, + 566, + 769, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We report the inference efficiency results on a single NVIDIA A6000 GPU in Table 6. As shown, due to the use of foundation priors, our method incurs relatively high computational and memory costs. Looking ahead, we aim to develop a model zoo for this method, including both a large and a light version.", + "bbox": [ + 511, + 589, + 905, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 694, + 633, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce Stereo Any Video, a robust framework for video stereo matching. By leveraging rich monocular priors, an all-to-all-pairs correlation mechanism, and a temporal convex upsampling layer, our approach produces temporally consistent and accurate disparities without relying on additional inputs such as camera pose or optical flow. Extensive evaluations show that our model achieves state-of-the-art performance on multiple standard benchmarks under zero-shot settings. Furthermore, our method generalizes well to real-world stereo videos, handling diverse content, motion, and camera movements effectively. Acknowledgments. This work was funded by the IC-CSC.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20843", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Luca Bartolomei, Fabio Tosi, Matteo Poggi, and Stefano Mattoccia. Stereo anywhere: Robust zero-shot deep stereo matching even where either stereo or mono fail. arXiv preprint arXiv:2412.04472, 2024.3", + "[2] D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black. A naturalistic open source movie for optical flow evaluation. In ECCV, pages 611-625, 2012. 4, 5, 6, 8", + "[3] Yohann Cabon, Naila Murray, and Martin Humenberger. Virtual kitti 2, 2020. 5, 6", + "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3", + "[5] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In CVPR, pages 5410-5418, 2018. 2", + "[6] Tianyu Chang, Xun Yang, Tianzhu Zhang, and Meng Wang. Domain generalized stereo matching via hierarchical visual transformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9559-9568, 2023. 2", + "[7] Sili Chen, Hengkai Guo, Shengnan Zhu, Feihu Zhang, Zi long Huang, Jiashi Feng, and Bingyi Kang. Video depth anything: Consistent depth estimation for super-long videos, 2025. 2, 3, 5, 6, 7, 8", + "[8] Junda Cheng, Longliang Liu, Gangwei Xu, Xianqi Wang, Zhaoxing Zhang, Yong Deng, Jinliang Zang, Yurui Chen, Zhipeng Cai, and Xin Yang. Monster: Marry monodepth to stereo unleashes power, 2025. 3, 5, 6", + "[9] Xuelian Cheng, Yiran Zhong, Mehrtash Harandi, Yuchao Dai, Xiaojun Chang, Tom Drummond, Hongdong Li, and Zongyuan Ge. Hierarchical neural architecture search for deep stereo matching. arXiv preprint arXiv:2010.13501, 2020.2", + "[10] Ziang Cheng, Jiayu Yang, and Hongdong Li. Stereo matching in time: $100+$ fps video stereo matching for extended reality. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 8719-8728, 2024. 2", + "[11] Seokju Cho, Jiahui Huang, Jisu Nam, Honggyu An, Seun-gryong Kim, and Joon-Young Lee. Local all-pair correspondence for point tracking. arXiv preprint arXiv:2407.15420, 2024. 4", + "[12] WeiQin Chuah, Ruwan Tennakoon, Reza Hoseinnezhad, Alireza Bab-Hadiashar, and David Suter. Itsa: An information-theoretic approach to automatic shortcut avoidance and domain generalization in stereo matching networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13022-13032, 2022. 2", + "[13] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, pages 3354-3361, 2012. 1, 2, 5, 6", + "[14] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In Proceedings of" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2495-2504, 2020. 2", + "[15] Weiyu Guo, Zhaoshuo Li, Yongkui Yang, Zheng Wang, Russell H Taylor, Mathias Unberath, Alan Yuille, and Yingwei Li. Context-enhanced stereo transformer. In European Conference on Computer Vision, pages 263-279. Springer, 2022. 2", + "[16] Xiaoyang Guo, Kai Yang, Wukui Yang, Xiaogang Wang, and Hongsheng Li. Group-wise correlation stereo network. In CVPR, pages 3273-3282, 2019. 2", + "[17] Xianda Guo, Chenming Zhang, Youmin Zhang, Wenzhao Zheng, Dujun Nie, Matteo Poggi, and Long Chen. Light-stereo: Channel boost is all your need for efficient 2d cost aggregation, 2024. 2", + "[18] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 2", + "[19] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. arXiv preprint arXiv:2409.02095, 2024. 3, 4", + "[20] Hualie Jiang, Zhiqiang Lou, Laiyan Ding, Rui Xu, Minglang Tan, Wenjie Jiang, and Rui Huang. Defom-stereo: Depth foundation model based stereo matching, 2025. 3, 5, 6", + "[21] Junpeng Jing, Jiankun Li, Pengfei Xiong, Jiangyu Liu, Shuaicheng Liu, Yichen Guo, Xin Deng, Mai Xu, Lai Jiang, and Leonid Sigal. Uncertainty guided adaptive warping for robust and efficient stereo matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3318-3327, 2023. 2", + "[22] Junpeng Jing, Ye Mao, and Krystian Mikolajczyk. Match-stereo-videos: Bidirectional alignment for consistent dynamic stereo matching. In European Conference on Computer Vision, pages 415–432. Springer, 2024. 2, 4, 5, 6, 8", + "[23] Junpeng Jing, Ye Mao, Anlan Qiu, and Krystian Mikolajczyk. Match stereo videos via bidirectional alignment, 2024. 3, 5, 6, 7", + "[24] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Dynamic stereo: Consistent dynamic depth from stereo videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13229-13239, 2023. 2, 4, 5, 6, 8", + "[25] Alex Kendall, Hayk Martirosyan, Saumitro Dasgupta, Peter Henry, Ryan Kennedy, Abraham Bachrach, and Adam Bry. End-to-end learning of geometry and context for deep stereo regression. In CVPR, pages 66-75, 2017. 2", + "[26] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 3", + "[27] Jiankun Li, Peisen Wang, Pengfei Xiong, Tao Cai, Ziwei Yan, Lei Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. Practical stereo matching via cascaded recurrent network with adaptive correlation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16263-16272, 2022. 1, 2, 4, 7, 8" + ], + "bbox": [ + 517, + 93, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20844", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Zhaoshuo Li, Xingtong Liu, Nathan Drenkow, Andy Ding, Francis X Creighton, Russell H Taylor, and Mathias Unberath. Revisiting stereo depth estimation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6197-6206, 2021. 2", + "[29] Zhaoshuo Li, Wei Ye, Dilin Wang, Francis X Creighton, Russell H Taylor, Ganesh Venkatesh, and Mathias Unberath. Temporally consistent online depth estimation in dynamic scenes. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 3018-3027, 2023. 2", + "[30] Zhengfa Liang, Yiliu Feng, Yulan Guo, Hengzhu Liu, Wei Chen, Linbo Qiao, Li Zhou, and Jianfeng Zhang. Learning for disparity estimation through feature constancy. In CVPR, pages 2811-2820, 2018. 2", + "[31] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. arXiv preprint arXiv:2109.07547, 2021. 1, 2, 3, 4, 5, 6", + "[32] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60:91-110, 2004. 4", + "[33] Yang Luo, Xuanlei Zhao, Mengzhao Chen, Kaipeng Zhang, Wenqi Shao, Kai Wang, Zhangyang Wang, and Yang You. Enhance-a-video: Better generated video for free, 2024. 2, 3", + "[34] Nikolaus Mayer, Eddy Ig, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In CVPR, pages 4040-4048, 2016. 2, 5, 6", + "[35] Lukas Mehl, Jenny Schmalfuss, Azin Jahedi, Yaroslava Nalivayko, and Andres Bruhn. Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 5, 6", + "[36] Tuan Duc Ngo, Peiye Zhuang, Chuang Gan, Evangelos Kalogerakis, Sergey Tulyakov, Hsin-Ying Lee, and Chaoyang Wang. Delta: Dense efficient long-range 3d tracking for any video, 2025. 8", + "[37] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 3", + "[38] Jiahao Pang, Wenxiu Sun, Jimmy SJ Ren, Chengxi Yang, and Qiong Yan. Cascade residual learning: A two-stage convolutional neural network for stereo matching. In CVPRW, pages 887-895, 2017. 2", + "[39] Jiahao Pang, Wenxiu Sun, Chengxi Yang, Jimmy Ren, Ruichao Xiao, Jin Zeng, and Liang Lin. Zoom and learn: Generalizing deep stereo matching to novel domains. In CVPR, pages 2070-2079, 2018. 2", + "[40] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Unidepth: Universal monocular metric depth estimation, 2024. 8", + "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3", + "[42] Zhibo Rao, Bangshu Xiong, Mingyi He, Yuchao Dai, Renjie He, Zhelun Shen, and Xing Li. Masked representation learning for domain generalized stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5435-5444, 2023. 2", + "[43] Ignacio Rocco, Mircea Cimpoi, Relja Arandjelović, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Neighbourhood consensus networks. Advances in neural information processing systems, 31, 2018. 4", + "[44] Daniel Scharstein and Richard Szeliski. A taxonomy and evaluation of dense two-frame stereo correspondence algorithms. IJCV, 47(1):7-42, 2002. 1, 2, 5, 6", + "[45] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In CVPR, pages 3260-3269, 2017. 1, 2, 5", + "[46] Zhelun Shen, Yuchao Dai, and Zhibo Rao. Cfnet: Cascade and fused cost volume for robust stereo matching. In CVPR, pages 13906-13915, 2021. 2", + "[47] Zhelun Shen, Xibin Song, Yuchao Dai, Dingfu Zhou, Zhibo Rao, and Liangjun Zhang. Digging into uncertainty-based pseudo-label for robust stereo matching. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-18, 2023. 2", + "[48] Chenyang Si, Weichen Fan, Zhengyao Lv, Ziqi Huang, Yu Qiao, and Ziwei Liu. Repvideo: Rethinking cross-layer representation for video generation, 2025. 2, 3", + "[49] Qing Su and Shihao Ji. Chitransformer: Towards reliable stereo from cues. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1939-1949, 2022. 2", + "[50] Vladimir Tankovich, Christian Hane, Yinda Zhang, Adarsh Kowdle, Sean Fanello, and Sofien Bouaziz. Hitnet: Hierarchical iterative tile refinement network for real-time stereo matching. In CVPR, pages 14362-14372, 2021. 2", + "[51] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, pages 402-419, 2020. 2, 5, 8", + "[52] Prune Truong, Martin Danelljan, Luc V Gool, and Radu Timofte. Gocor: Bringing globally optimized correspondence volumes into your neural network. Advances in Neural Information Processing Systems, 33:14278-14290, 2020. 4", + "[53] Prune Truong, Martin Danelljan, and Radu Timofte. Glunet: Global-local universal network for dense flow and correspondences. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6258-6268, 2020. 4", + "[54] Jonas Uhrig, Nick Schneider, Lukas Schneider, Uwe Franke, Thomas Brox, and Andreas Geiger. Sparsity invariant cnns. In 2017 international conference on 3D Vision (3DV), pages 11-20. IEEE, 2017. 5, 6", + "[55] Xianqi Wang, Gangwei Xu, Hao Jia, and Xin Yang. Selective-stereo: Adaptive frequency information selection" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20845", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19701-19710, 2024. 1, 2, 5, 6", + "[56] Yan Wang, Zihang Lai, Gao Huang, Brian H Wang, Laurens Van Der Maaten, Mark Campbell, and Kilian Q Weinberger. Anytime stereo image depth estimation on mobile devices. In 2019 international conference on robotics and automation (ICRA), pages 5893-5900. IEEE, 2019. 2", + "[57] Yiran Wang, Min Shi, Jiaqi Li, Zihao Huang, Zhiguo Cao, Jianming Zhang, Ke Xian, and Guosheng Lin. Neural video depth stabilizer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9466-9476, 2023. 5", + "[58] Philippe Weinzaepfel, Thomas Lucas, Vincent Leroy, Yohann Cabon, Vaibhav Arora, Romain Brégier, Gabriela Csurka, Leonid Antsfeld, Boris Chidlovskii, and Jérôme Revaud. Croco v2: Improved cross-view completion pretraining for stereo matching and optical flow. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17969-17980, 2023. 2", + "[59] Bowen Wen, Matthew Trepte, Joseph Aribido, Jan Kautz, Orazio Gallo, and Stan Birchfield. Foundationstereo: Zeroshot stereo matching, 2025. 3, 4, 6", + "[60] Bin Xu, Yuhua Xu, Xiaoli Yang, Wei Jia, and Yulan Guo. Bilateral grid learning for stereo matching networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12497-12506, 2021. 2", + "[61] Gangwei Xu, Xianqi Wang, Xiaohuan Ding, and Xin Yang. Iterative geometry encoding volume for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21919-21928, 2023. 1, 2, 4, 5, 6", + "[62] Gangwei Xu, Yun Wang, Junda Cheng, Jinhui Tang, and Xin Yang. Accurate and efficient stereo matching via attention concatenation volume. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2", + "[63] Gangwei Xu, Huan Zhou, and Xin Yang. Cgi-stereo: Accurate and real-time stereo matching via context and geometry interaction. arXiv preprint arXiv:2301.02789, 2023. 2", + "[64] Gangwei Xu, Xianqi Wang, Zhaoxing Zhang, Junda Cheng, Chunyuan Liao, and Xin Yang. Igev++: Iterative multirange geometry encoding volumes for stereo matching. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 2", + "[65] Haofei Xu and Juyong Zhang. Aanet: Adaptive aggregation network for efficient stereo matching. In CVPR, pages 1959-1968, 2020. 2", + "[66] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, Fisher Yu, Dacheng Tao, and Andreas Geiger. Unifying flow, stereo and depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2", + "[67] Gengshan Yang, Joshua Manela, Michael Happold, and Deva Ramanan. Hierarchical deep stereo matching on high-resolution images. In CVPR, pages 5515-5524, 2019. 2", + "[68] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of" + ], + "bbox": [ + 91, + 90, + 482, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 3, 6, 8", + "[69] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 3", + "[70] Jiaxi Zeng, Chengtang Yao, Yuwei Wu, and Yunde Jia. Temporally consistent stereo matching. In European Conference on Computer Vision, pages 341-359. Springer, 2024. 2", + "[71] Feihu Zhang, Victor Prisacariu, Ruigang Yang, and Philip HS Torr. Ga-net: Guided aggregation net for end-to-end stereo matching. In CVPR, pages 185-194, 2019. 2", + "[72] Jiawei Zhang, Xiang Wang, Xiao Bai, Chen Wang, Lei Huang, Yimin Chen, Lin Gu, Jun Zhou, Tatsuya Harada, and Edwin R Hancock. Revisiting domain generalized stereo matching networks from a feature consistency perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13001-13011, 2022. 2", + "[73] Youmin Zhang, Matteo Poggi, and Stefano Mattoccia. Temporalstereo: Efficient spatial-temporal stereo matching network. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 9528-9535. IEEE, 2023. 2", + "[74] Yongjian Zhang, Longguang Wang, Kunhong Li, Yun Wang, and Yulan Guo. Learning representations from foundation models for domain generalized stereo matching. In European Conference on Computer Vision, pages 146-162. Springer, 2024. 3", + "[75] Yiran Zhong, Hongdong Li, and Yuchao Dai. Open-world stereo video matching with deep rnN. In Proceedings of the European Conference on Computer Vision (ECCV), pages 101-116, 2018. 2", + "[76] Jingyi Zhou, Haoyu Zhang, Jiakang Yuan, Peng Ye, Tao Chen, Hao Jiang, Meiya Chen, and Yangyang Zhang. All-in-one: Transferring vision foundation models into stereo matching. arXiv preprint arXiv:2412.09912, 2024. 3" + ], + "bbox": [ + 516, + 92, + 903, + 601 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20846", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Stereo Any Video_ Temporally Consistent Stereo Matching/dc20cc87-36ce-4c69-8c57-afd220728c97_model.json b/2025/Stereo Any Video_ Temporally Consistent Stereo Matching/dc20cc87-36ce-4c69-8c57-afd220728c97_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5fc651c03e8f5d3f27c52c99ea8b8272f6ce96e1 --- /dev/null +++ b/2025/Stereo Any Video_ Temporally Consistent Stereo Matching/dc20cc87-36ce-4c69-8c57-afd220728c97_model.json @@ -0,0 +1,3071 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.048 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.203, + 0.131, + 0.797, + 0.154 + ], + "angle": 0, + "content": "Stereo Any Video: Temporally Consistent Stereo Matching" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.18, + 0.756, + 0.218 + ], + "angle": 0, + "content": "Junpeng Jing Weixun Luo Ye Mao Krystian Mikolajczyk Imperial College London" + }, + { + "type": "text", + "bbox": [ + 0.296, + 0.22, + 0.697, + 0.234 + ], + "angle": 0, + "content": "https://tomtomommi.github.io/StereoAnyVideo/" + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.269, + 0.889, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.128, + 0.488, + 0.248, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.489, + 0.372, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.379, + 0.489, + 0.496, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.489, + 0.619, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.489, + 0.744, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.751, + 0.489, + 0.868, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.486, + 0.556, + 0.517, + 0.567 + ], + "angle": 0, + "content": "time" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.579, + 0.908, + 0.608 + ], + "angle": 0, + "content": "Figure 1. Prediction on a dynamic real-world stereo video. Our method produces temporally consistent and accurate disparities without flickering, enabling the reconstruction of dense and clean metric point clouds for dynamic scenes. See video for better visualization." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.62, + 0.327, + 0.636 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.651, + 0.486, + 0.879 + ], + "angle": 0, + "content": "This paper introduces Stereo Any Video, a powerful framework for video stereo matching. It can estimate spatially accurate and temporally consistent disparities without relying on auxiliary information such as camera poses or optical flow. The strong capability is driven by rich priors from monocular video depth models, which are integrated with convolutional features to produce stable representations. To further enhance performance, key architectural innovations are introduced: all-to-all-pairs correlation, which constructs smooth and robust matching cost volumes, and temporal convex upsampling, which improves temporal coherence. These components collectively enhance robustness, accuracy, and temporal consistency, establishing a new standard in video stereo matching. Extensive experiments demonstrate that our method achieves state-of-the" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.622, + 0.907, + 0.668 + ], + "angle": 0, + "content": "art performance across multiple datasets both qualitatively and quantitatively in zero-shot settings, as well as strong generalization to real-world indoor and outdoor scenarios." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.71, + 0.646, + 0.725 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.856 + ], + "angle": 0, + "content": "Video stereo matching is a task to reconstruct a 3D scene from a stereo camera. It estimates disparity by identifying the horizontal correspondences between rectified left and right images at each time step. Consistent and accurate disparity estimation is crucial for reconstructing and understanding real-world 3D scenes, benefiting various downstream tasks such as autonomous driving, robotic navigation, and virtual/augmented reality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "With the advances in deep learning, stereo matching has seen significant progress [27, 31, 55, 61], achieving promising performance on public benchmarks [13, 44, 45]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.422, + 0.901 + ], + "angle": 0, + "content": "†Corresponding author: ye.mao21@imperial.ac.uk" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20836" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.348 + ], + "angle": 0, + "content": "However, most existing approaches are designed for image-based estimation and do not explicitly leverage temporal information in video sequences. When applied directly to video data, these methods often suffer from temporal inconsistencies, leading to noticeable flickering and artifacts in disparity maps and reconstructed point clouds. To mitigate this, recent video-based methods incorporate additional temporal cues, such as camera poses or optical flow, through predefined models or learnable mechanisms. For example, CODD [29] employs a pose network to estimate relative camera movements and align adjacent disparities within a fusion framework. Similarly, TemporalStereo [73] and TC-Stereo [70] follow this paradigm with variations in architectural design and alignment operations. Another approach, BiDAStereo [22], integrates optical flow [51] for temporal feature alignment and introduces a temporal aggregation module to propagate information across time." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.352, + 0.482, + 0.517 + ], + "angle": 0, + "content": "Despite these advances, existing video methods remain vulnerable to dynamic scenes, camera movement, and the accuracy of auxiliary information. In real-world scenarios, dynamic objects undergo complex motions and deformations that violate traditional multi-view constraints [18], disrupting translational invariance in disparity estimation. Moreover, models relying on auxiliary cues are constrained by the performance of camera pose or optical flow estimation modules. In environments with complex camera movements, these estimation modules degrade, becoming bottlenecks that limit the robustness of video stereo methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.521, + 0.483, + 0.822 + ], + "angle": 0, + "content": "To address these limitations, we propose Stereo Any Video, a robust framework for video-based stereo matching without relying on camera pose or optical flow. Inspired by recent findings in video generation [33, 48], which emphasize the role of robust feature representations in ensuring temporal coherence, we take a new perspective to design the framework for consistency: feature robustness and stability. Specifically, during feature extraction, we integrate convolutional cues with a frozen monocular video depth foundation model [7] trained on million-scale data. This ensures robust feature maps with rich priors. For correspondence matching, we introduce an all-to-all-pair correlation, which enforces smooth and reliable matching by jointly considering bidirectional correspondences from both target and reference views. For cost aggregation, we propose temporal convex upsampling within a fully temporal-based recurrent unit, playing a key role in maintaining temporal consistency. As shown in Figure 1, our method demonstrates a strong ability to produce accurate and consistent disparities, benefiting the reconstruction for dynamic scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.482, + 0.901 + ], + "angle": 0, + "content": "The contributions of our work can be summarized as follows: (1) Stereo Any Video, a framework that integrates monocular video depth priors to achieve accurate and temporally consistent video stereo matching. (2) A novel all-to-all-pair correlation module and a temporal convex upsam" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "pling mechanism, ensuring robust and stable cost volume representations and aggregation. (3) State-of-the-art performance on diverse benchmarks, setting a new standard for video stereo matching." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.166, + 0.655, + 0.182 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.192, + 0.905, + 0.267 + ], + "angle": 0, + "content": "In this section, we provide a brief literature review of stereo matching methods in the deep learning era. We first discuss image-based stereo matching methods, followed by video-based methods, and finally, we review recent advancements leveraging foundation priors." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.268, + 0.905, + 0.584 + ], + "angle": 0, + "content": "Image-based Stereo Matching. Since the introduction of the first end-to-end stereo network [34], deep learning-based stereo matching has seen substantial advancements. Early methods typically construct 3D [16, 30, 38, 50, 65] or 4D [5, 9, 14, 25, 62, 67, 71] cost volumes from left and right image features, followed by 2D or 3D convolutions for cost aggregation. Inspired by RAFT [51], iterative refinement has become the mainstream approach. Representative advancements include multilevel recurrent units [31], adaptive correlation [27], geometry-encoding volumes [61, 64], and frequency information selection [55], all of which have significantly improved benchmark performance [13, 44, 45]. Transformer-based methods [15, 28, 49, 58, 66] have further refined stereo matching by modeling token correspondence and capturing long-range dependencies. Beyond accuracy, several methods focus on real-world challenges such as efficiency [17, 56, 60, 63], domain generalization [6, 12, 39, 42, 72], and robustness [21, 46, 47]. However, as all these methods are designed for static image pairs, they lack temporal awareness, leading to inconsistent disparity predictions in video sequences." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.585, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Video-based Stereo Matching. Video stereo matching especially in dynamic scenes presents unique challenges. Early work, such as [75], employed LSTMs with unsupervised learning for stereo videos but did not explicitly incorporate temporal information. CODD [29] introduced a per-frame stereo network, a motion network, and a fusion network designed for dynamic scenes. Neighboring-frame disparities are aligned and fused using estimated camera poses. TemporalStereo [73] extended this approach with a coarse-to-fine network that leverages past context to enhance predictions in challenging scenarios. Cheng et al. [10] developed a framework tailored for XR systems, reducing computational costs through temporal cost aggregation. TC-Stereo [70] further introduced a temporal disparity completion mechanism and a temporal state fusion module to exploit temporal information. These methods rely on auxiliary camera motion or pre-defined scene geometry to incorporate temporal information. DynamicStereo [24] introduced a transformer-based architecture that integrates temporal, stereo, and spatial attention. BiDAStereo [22] proposed an optical-flow-based alignment mechanism to propagate" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "20837" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.09, + 0.865, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.305, + 0.908, + 0.377 + ], + "angle": 0, + "content": "Figure 2. Overview of Stereo Any Video. Given a stereo video sequence as input, the model first extracts features and context information using trainable convolutional encoders and a frozen monocular video depth encoder (Video Depth Anything [7]). At each iteration, an all-to-all-pair correlation computes feature correlations, followed by a multilayer perception (MLP) encoder for compact representation. Disparities are iteratively refined using a 3D Gated Recurrent Unit (GRU), integrated with cost volume features, and upsampled via a temporal convex upsampling layer. This process is repeated within a cascaded pipeline to progressively recover full-resolution disparities." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.389, + 0.482, + 0.465 + ], + "angle": 0, + "content": "temporal information across the entire sequence. Building on this alignment, BiDAS stabilizer [23] introduced a lightweight plug-and-play stabilizer network to improve the temporal consistency of existing image-based stereo matching methods." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.478, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Stereo Matching with Foundation Priors. Vision foundation models, such as CLIP [41], the DINO series [4, 37], and SAM [26], have been widely applied across various vision tasks due to their rich priors. In stereo matching area, FormerStereo [74] proposed feature transformation and C4 space to adapt ViT-based foundation model to stereo pipeline, learning domain invariant representations. AIOStereo [76] introduced a selective knowledge transfer module to distill knowledge from multiple heterogeneous vision foundation models into a single stereo matching model. More recent approaches primarily focused on leveraging monocular depth priors from Depth Anything [68, 69] to enhance stereo performance. For instance, Stereo Anywhere [1] directly incorporated relative depth maps into the stereo network, demonstrating robustness in challenging conditions such as non-Lambertian surfaces. DEFOM-Stereo [20] employed these depth maps alongside a novel scale updater within a recurrent module for improved alignment. FoundationStereo [59] extracted feature maps from Depth Anything and introduced a side-tuning adapter to adapt monocular features for stereo backbone. MonSter [8] utilized both depth maps and feature maps, and proposed a dualbranched refinement module to fuse stereo and monocular cues effectively. These methods highlight the effectiveness of foundation priors in image-based stereo tasks. In our work, we build upon this direction by investigating video-based stereo matching and exploring how foundation priors can enhance consistency in dynamic scenes." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.388, + 0.606, + 0.403 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.413, + 0.907, + 0.566 + ], + "angle": 0, + "content": "In this section, we introduce Stereo Any Video, a feedforward video-based framework. Given a rectified stereo sequence \\(\\{\\mathbf{I}_L^t,\\mathbf{I}_R^t\\}_{t\\in (1,T)}\\in \\mathbb{R}^{H\\times W\\times 3}\\) where \\(T\\) denotes the total number of frames, \\(H\\) and \\(W\\) are the frame height and width, respectively, our goal is to estimate a temporally consistent sequence of disparity maps \\(\\{\\mathbf{d}^t\\}_{t\\in (1,T)}\\in \\mathbb{R}^{H\\times W}\\) aligned with the left view. As shown in Figure 2, our method consists of three key components: feature extraction (Section 3.1), correlation (Section 3.2), and temporal cost aggregation (Section 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.576, + 0.88, + 0.59 + ], + "angle": 0, + "content": "3.1. Feature Extraction with Foundation Priors" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.84 + ], + "angle": 0, + "content": "Unlike image-based tasks, maintaining temporal consistency is the primary challenge in video stereo matching. Recent studies in video generation [33, 48] indicate that substantial variations across different frames can lead to unstable feature representations, thereby degrading temporal coherence. Therefore, ensuring feature stability across adjacent frames is crucial for achieving temporally consistent disparity estimation. Towards this, we leverage Video Depth Anything (VDA) [7] for feature extraction, which offers two key advantages: (1) trained on internet-scale data, VDA provides rich priors that can enhance disparity estimation in regions with ambiguous stereo correspondence; (2) unlike image-based monocular depth models [68, 69], VDA generates temporally stable features through an explicit temporal alignment process before the final prediction head, thereby preserving coherence across frames." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Besides, we investigate whether the depth maps from monocular video depth models can further improve stereo performance. We compare DepthCrafter [19], a representative monocular video model, with RAFTStereo [31], an" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20838" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.483, + 0.117 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of DepthCrafter [19] and RAFT-Stereo [31] on Sintel dataset [2]." + }, + { + "type": "table", + "bbox": [ + 0.116, + 0.12, + 0.454, + 0.19 + ], + "angle": 0, + "content": "
MethodSintel Clean
EPE ↓TEPE ↓MOS (human) ↓
DepthCrafter [19]8.682.080.212
RAFTStereo [31]1.420.840.933
" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.196, + 0.223, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.237, + 0.193, + 0.244 + ], + "angle": 0, + "content": "Input bamboo 2" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.196, + 0.347, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.238, + 0.315, + 0.244 + ], + "angle": 0, + "content": "DepthCrafter" + }, + { + "type": "image", + "bbox": [ + 0.348, + 0.196, + 0.471, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.385, + 0.238, + 0.437, + 0.244 + ], + "angle": 0, + "content": "RAFTStereo" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.244, + 0.223, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.285, + 0.194, + 0.293 + ], + "angle": 0, + "content": "Input sleeping_1" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.244, + 0.346, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.259, + 0.285, + 0.314, + 0.293 + ], + "angle": 0, + "content": "DepthCrafter" + }, + { + "type": "image", + "bbox": [ + 0.348, + 0.244, + 0.471, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.385, + 0.285, + 0.437, + 0.293 + ], + "angle": 0, + "content": "RAFTStereo" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.299, + 0.483, + 0.327 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparison of DepthCrafter [19] and RAFT-Stereo [31] on Sintel dataset [2]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.339, + 0.483, + 0.581 + ], + "angle": 0, + "content": "image-based stereo model. Using open-source weights, we evaluate their performance on Sintel [2], as shown in Table 1 and Figure 3. Specifically, we first obtain relative depth maps from DepthCrafter (without scale and shift) and compute disparities using RAFTStereo on the same sequence. The relative depth maps are then aligned to RAFTStereo's disparities, and objective metrics are computed against the ground truth. Additionally, we conduct a subjective user study (see supplementary material for details). As observed, while DepthCrafter produces visually plausible results—reflected in better human scores—its spatial and temporal errors are significantly larger than those of RAFTStereo. This suggests that its relative depth is \"consistent but wrong\", making it unreliable for directly improving consistency within a stereo framework. Therefore, we only utilize its features as valuable priors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.483, + 0.72 + ], + "angle": 0, + "content": "As shown in Figure 2, our feature extraction module consists of two trainable encoders, each employing a series of residual blocks [31] to extract convolutional image features and context features. To integrate VDA priors, we adopt a shallow convolutional adapter [59] that downsamples depth features from the frozen VDA. The extracted depth features are then concatenated with convolutional features to form \\(\\{\\mathbf{F}_L,\\mathbf{F}_R\\}\\) and \\(\\mathbf{F}_C\\), providing more stable feature representations for subsequent processing." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.734, + 0.329, + 0.751 + ], + "angle": 0, + "content": "3.2. All-to-all-pair Correlation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.758, + 0.483, + 0.864 + ], + "angle": 0, + "content": "At the \\(n\\)-th iteration for each time step, the right feature \\(\\mathbf{F}_R\\) is first warped toward the left feature \\(\\mathbf{F}_L\\) using the previous disparity \\(d_{n-1}\\), producing \\(\\hat{\\mathbf{F}}_R\\). Previous methods [31, 61] typically construct cost volumes \\(\\mathbf{C}_n\\) via all-pair correlation, employing a lookup mechanism to search for potential matches within a local radius. In [27], this approach was extended to an alternating local 1D and 2D search:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.879, + 0.483, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} _ {n} (x, y) = \\left\\langle \\mathbf {F} _ {L} (x, y), \\hat {\\mathbf {F}} _ {R} (x + r _ {x}, y + r _ {y}) \\right\\rangle , \\tag {1}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.091, + 0.895, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.255, + 0.905, + 0.298 + ], + "angle": 0, + "content": "Figure 4. Workflow of the temporal convex upsampling. Each pixel of the high resolution disparity (yellow box) is calculated via its 27 low resolution neighbors using learnable weights." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.308, + 0.907, + 0.49 + ], + "angle": 0, + "content": "where \\((x,y)\\) denotes the position in \\(\\mathbf{F}_L\\), \\(\\langle \\cdot, \\cdot \\rangle\\) represents the channel-wise product operation, and \\((r_x, r_y)\\) defines the search range, typically set as \\((r_x, r_y) \\in \\{(\\pm 4,0), (\\pm 1, \\pm 1)\\}\\). The fundamental principle of these methods is to construct one-directional correspondence, where a reference point is selected and directly matched with target points via multiplication. Inspired by recent advances in point tracking [11], we introduce bidirectional correspondence and develop a local all-to-all-pair correlation. Instead of computing correlation in a fixed direction, we calculate the similarity between every pair of potential matching points within two search windows, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.511, + 0.905, + 0.537 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} _ {n} (x, y) = \\left\\langle \\mathbf {F} _ {L} \\left(x + r _ {x}, y + r _ {y}\\right), \\hat {\\mathbf {F}} _ {R} \\left(x + r _ {x}, y + r _ {y}\\right) \\right\\rangle . \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.906, + 0.667 + ], + "angle": 0, + "content": "This approach enhances match verification and reduces ambiguity [32, 43]. While BiDAStereo [22] constructs temporal cost volumes by aligning adjacent frames for consistency, its performance is significantly constrained by the limitations of optical flow. In contrast, our method emphasizes robust spatial correlation, leveraging dense correspondences to enforce matching smoothness and enhance consistency and robustness across neighboring points [52, 53]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.675, + 0.764, + 0.691 + ], + "angle": 0, + "content": "3.3. Temporal Cost Aggregation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.697, + 0.906, + 0.803 + ], + "angle": 0, + "content": "At the \\(n\\)-th iteration, the all-to-all-pair correlation expands the potential matching points in \\(\\mathbf{C}_n\\), increasing its dimensionality from \\(\\mathbb{R}^{H' \\times W' \\times (2r_x + 1)(2r_y + 1)}\\) to \\(\\mathbb{R}^{H' \\times W' \\times (2r_x + 1)^2(2r_y + 1)^2}\\), where \\(H'\\) and \\(W'\\) denote the height and width of frame feature, respectively. We first encode \\(\\mathbf{C}_n\\) into a compact representation \\(\\mathbf{E}_n \\in \\mathbb{R}^{H' \\times W' \\times L}\\) using a MLP:" + }, + { + "type": "equation", + "bbox": [ + 0.649, + 0.804, + 0.905, + 0.82 + ], + "angle": 0, + "content": "\\[\n\\mathbf {E} _ {n} = \\operatorname {M L P} \\left(\\mathbf {C} _ {n}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Next, we employ a 3D-GRU [24] to iteratively refine the disparity for each frame while preserving temporal consistency. The GRU aggregates information using separable 3D convolutions across both spatial and temporal dimensions. At iteration \\( n \\), it takes as input the encoded cost volume \\( \\mathbf{E}_n \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20839" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.131 + ], + "angle": 0, + "content": "Table 2. Zero-shot generalization results. All methods were trained only on SceneFlow [34]. The representative image-based and video-based stereo matching models are compared on Sintel [2], Dynamic Replica [24], Infinigen SV [23], and Virtual KITTI2 [3]. The weights and parameters are fixed for evaluation. The best and second-best results are highlighted." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.133, + 0.905, + 0.302 + ], + "angle": 0, + "content": "
MethodSintel FinalDynamic ReplicaInfinigen SVVirtual KITTI2
TEPE↓δt1px↓EPE↓TEPE↓δt1px↓EPE↓TEPE↓δt1px↓EPE↓TEPE↓δt1px↓EPE↓
Image-based
RAFTStereo [31]2.0913.704.890.1452.020.382.3319.803.771.2518.391.45
IGEVStereo [61]1.7412.652.440.0810.800.202.3519.333.811.4218.881.39
Selective-IGEV [55]1.9612.952.820.1331.290.292.8521.835.691.4119.101.38
DEFOM-Stereo [20]2.3312.432.380.0800.750.132.0118.913.241.0716.631.34
MonSter [8]1.7011.063.300.1570.860.451.6515.732.270.7310.930.67
Video-based
DynamicStereo [24]1.4311.804.260.0830.730.241.9815.684.671.0615.631.47
BiDAStereo [22]1.2611.651.990.0911.040.651.9916.506.521.0214.821.43
Stereo Any Video (ours)1.079.981.750.0570.670.251.6514.273.160.7412.451.01
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.313, + 0.484, + 0.375 + ], + "angle": 0, + "content": "context feature \\(\\mathbf{F}_c\\), and the disparities from the previous iteration \\(d_{n-1}\\). To enhance representation, we integrate super kernels [22], temporal and spatial attention [24] in this process. The disparity update in 3D-GRU is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.382, + 0.482, + 0.401 + ], + "angle": 0, + "content": "\\[\nx _ {n} = \\left[ \\operatorname {C o n v 2 d} \\left(\\mathbf {E} _ {n}\\right), \\operatorname {C o n v 2 d} \\left(d _ {n - 1}\\right), d _ {n - 1}, \\mathbf {F} _ {c} \\right], \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.402, + 0.483, + 0.42 + ], + "angle": 0, + "content": "\\[\n\\widetilde {x _ {n}} = \\operatorname {A t t e n t i o n} \\left(x _ {n}\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.422, + 0.483, + 0.446 + ], + "angle": 0, + "content": "\\[\nz _ {n} = \\sigma \\left(\\operatorname {C o n v 3 d} _ {z} \\left(\\left[ h _ {n - 1}, \\widetilde {x _ {n}} \\right]\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.45, + 0.483, + 0.475 + ], + "angle": 0, + "content": "\\[\nr _ {n} = \\sigma \\left(\\operatorname {C o n v 3 d} _ {r} \\left(\\left[ h _ {n - 1}, \\widetilde {x _ {n}} \\right]\\right)\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.478, + 0.483, + 0.503 + ], + "angle": 0, + "content": "\\[\nq _ {n} = \\tanh \\left(\\operatorname {C o n v} 3 \\mathrm {d} _ {q} \\left(\\left[ h _ {n - 1} \\odot r _ {n}, \\widetilde {x _ {n}} \\right]\\right)\\right), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.506, + 0.483, + 0.522 + ], + "angle": 0, + "content": "\\[\nh _ {n} = \\left(1 - z _ {n}\\right) \\odot h _ {n - 1} + z _ {n} \\odot q _ {n}, \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.525, + 0.483, + 0.541 + ], + "angle": 0, + "content": "\\[\nd _ {n} = d _ {n - 1} + \\operatorname {C o n v} 3 \\mathrm {d} \\left(h _ {n}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.55, + 0.483, + 0.58 + ], + "angle": 0, + "content": "where \\([\\cdot ]\\) is concatenation, \\(\\odot\\) denotes element-wise multiplication, and \\(\\sigma\\) is the sigmoid activation function." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.483, + 0.747 + ], + "angle": 0, + "content": "While prior methods emphasize temporal consistency in correlation and cost aggregation, the upsampling process remains overlooked, despite its crucial role in achieving accurate and temporally consistent disparities. Towards this, we adapt the commonly used convex upsampling approach [51] to a temporal variant. In Figure 4, each high-resolution disparity pixel (yellow box) is computed as a convex combination of its 27 low resolution neighbors across the 3 frames (green, blue, and faint yellow boxes). The network predicts the corresponding \\(3 \\times 3 \\times 3\\) weights using 3D convolutions, enabling a temporally consistent upsampling:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.755, + 0.483, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\mathbf {w} = \\operatorname {s o f t m a x} \\left(\\operatorname {C o n v 3 d} \\left(h _ {n}\\right)\\right), \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.783, + 0.483, + 0.816 + ], + "angle": 0, + "content": "\\[\nD _ {n} ^ {t} = \\alpha \\cdot \\sum_ {i j k} \\mathbf {w} _ {i j k} \\odot \\text {u n f o l d} \\left(\\left[ d _ {n} ^ {t - 1}, d _ {n} ^ {t}, d _ {n} ^ {t + 1} \\right]\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.825, + 0.483, + 0.842 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is upsampling rate, \\(D_{n}^{t}\\) is the upsampled disparity." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.237, + 0.864 + ], + "angle": 0, + "content": "3.4. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Recent works have introduced OPW loss [57] and TGM loss [7] in monocular video depth estimation to enhance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.313, + 0.907, + 0.39 + ], + "angle": 0, + "content": "temporal consistency. However, we found out that these losses impose a trade-off between accuracy and consistency without yielding substantial improvement in the video stereo task. Thus, we opt for a purely image-based \\( L_{1} \\) loss while still achieving strong temporal consistency as:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.4, + 0.907, + 0.442 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\sum_ {t = 1} ^ {T} \\sum_ {n = 1} ^ {N} \\gamma^ {N - n} \\left| \\left| D _ {\\mathrm {g t}} ^ {t} - D _ {n} ^ {t} \\right| \\right|, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.452, + 0.907, + 0.514 + ], + "angle": 0, + "content": "where \\(T\\) is the number of frames, \\(N\\) is the number of iterations, and \\(\\gamma\\) is a weighting coefficient set to 0.9. Here, \\(D_{\\mathrm{gt}}^t\\) denotes the ground truth disparity for the \\(t\\)-th frame. The final output is obtained from the last iteration prediction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.527, + 0.646, + 0.544 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.907, + 0.644 + ], + "angle": 0, + "content": "In this section, we first introduce the datasets and metrics used in the evaluation, as well as implementation details. We then present the zero-shot generalization capabilities of our method, followed by ablation studies to validate the efficacy of each component in our framework. Please refer to the supplementary material for more details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.653, + 0.8, + 0.668 + ], + "angle": 0, + "content": "4.1. Benchmark Datasets and Metric" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Dataset. Our work focuses on videos captured with moving cameras, making commonly used image benchmarks unsuitable, such as Middlebury [44], ETH3D [45], and KITTI 2012/2015 [13] (two frames per scene). To evaluate our method, we use five synthetic and two real-world stereo video datasets, all featuring dynamic scenes. Sintel [2] is a synthetic dataset comprising movie sequences available in both clean and final passes. Spring [35] is a high resolution synthetic dataset. Dynamic Replica [24] is a synthetic indoor dataset featuring non-rigid objects, such as animals and people. Infinigen SV [23] is a synthetic dataset focusing on outdoor natural environments. Virtual KITTI2 [3] is a synthetic dataset simulating outdoor driving scenarios. KITTI Depth [54] is a real-world outdoor dataset captured for autonomous driving applications, with sparse" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20840" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.146 + ], + "angle": 0, + "content": "Table 3. Quantitative comparisons on Spring [35], Sintel [2], and KITTI Depth [54] where methods are allowed to train on any existing datasets. Model weights remain fixed during evaluation. Training datasets sharing the same domain as the test sets are highlighted. “-” represents the test set has been used in training. Abbreviations: SF - SceneFlow [34], DR - Dynamic Replica [24], ISV - Infigen SV [23], M - Middlebury [44], KD - KITTI Depth [54], VK - Virtual KITTI2 [3]. The best and second-best results are highlighted." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.15, + 0.907, + 0.329 + ], + "angle": 0, + "content": "
MethodTraining DataScaleSpringSintel FinalKITTI Depth
TEPE↓δt1px↓EPE↓TEPE↓δt1px↓EPE↓TEPE↓δt1px↓EPE↓
Image-based
RAFTStereo [31]10 mixed datasets0.6M1.499.795.37---0.417.960.51
IGEVStereo [61]SF+M0.04M1.539.725.371.5313.492.230.469.120.60
Selective-IGEV [55]SF+M0.04M1.469.335.461.6413.782.140.489.660.60
DEFOM-Stereo [20]15 mixed datasets0.7M1.299.124.25---0.397.080.48
MonSter [8]10 mixed datasets0.6M1.258.655.19---0.417.730.60
FoundationStereo [59]7 mixed datasets1.3M1.789.825.40---0.407.200.57
Video-based
DynamicStereo [24]SF+DR+ISV+KD0.1M1.079.649.361.3811.843.380.428.210.57
BiDAStereo [22]SF+DR+ISV+KD0.1M0.907.903.921.3311.341.970.428.330.58
Stereo Any Video (ours)SF+DR+ISV+VK0.1M0.776.523.170.999.091.480.356.050.61
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.346, + 0.483, + 0.391 + ], + "angle": 0, + "content": "depth maps obtained from a LiDAR sensor. South Kensington SV [23] is a real-world dataset capturing daily scenarios from a stereo camera for qualitative evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.392, + 0.484, + 0.499 + ], + "angle": 0, + "content": "Metric. For accuracy evaluation, we use End-Point Error (EPE), which measures the average per-pixel disparity error. To assess temporal consistency, we employ Temporal EPE (TEPE) [24], which quantifies the variation in EPE across the temporal dimension. We also compute \\(\\delta_{n\\text{-px}}^t\\), which represents the proportion of pixels with TEPE exceeding \\(n\\). For all metrics, lower values indicate better performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.51, + 0.31, + 0.526 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.533, + 0.484, + 0.76 + ], + "angle": 0, + "content": "We implement Stereo Any Video in PyTorch and train on NVIDIA A100 GPUs. The model is trained with a batch size of 8 on SceneFlow [34] for \\(120k\\) iterations and finetuned on a mixed dataset for \\(80k\\) iterations consisting of Dynamic Replica [24] training set, Infinigen SV [23] training set, and Virtual KITTI2 [3]. Before being fed into the network, sequences are randomly cropped to \\(256 \\times 512\\). We use the AdamW optimizer with a learning rate of \\(2e^{-4}\\) and apply a one-cycle learning rate schedule. The full training process takes approximately 6 days. Following [24], we apply various data augmentation techniques, including random cropping, rescaling, and saturation shifts. The sequence length is set to \\(T = 5\\) during training, and \\(T = 20\\) for evaluation. The number of iterations in GRU is set to \\(N = 10\\) in training, and \\(N = 20\\) in evaluation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.772, + 0.212, + 0.787 + ], + "angle": 0, + "content": "4.3. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Quantitative Results. Table 2 presents the results when models are trained solely on SceneFlow [34] and evaluated on four dynamic synthetic datasets including diverse indoor and outdoor environments. The results clearly demonstrate that our model achieves state-of-the-art zero-shot generalization across nearly all metrics, excelling in both accuracy and temporal consistency. Table 3 reports the results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.346, + 0.907, + 0.512 + ], + "angle": 0, + "content": "when models are allowed to train on any existing datasets. Notably, despite being trained solely on synthetic datasets, our model surpasses comparison methods that are trained on real-world datasets or share the same domain as the test tests by at least \\(15\\%\\) on diverse benchmarks. Specifically, compared to the robust version of RAFT-Stereo, which is trained on 10 mixed datasets, including in-domain sets such as Sintel [2] and KITTI [13], our approach achieves superior performance across most metrics. These results highlight the strong generalization capability of our method when applied to unseen domains." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.514, + 0.909, + 0.801 + ], + "angle": 0, + "content": "Qualitative Results. Figure 5 and Figure 6 present qualitative results using the checkpoints from Table 3 on dynamic real-world stereo videos from the South Kensington SV dataset [23]. For the indoor scenario shown in Figure 5, the upper rows display three input left frames along with their predicted disparities, while the subsequent rows visualize the globally aligned reconstructed point cloud. The red arrows highlight that our method produces temporally consistent and accurate disparities, whereas other methods exhibit noticeable artifacts. Specifically, Selective-IGEV and DynamicStereo suffer from flickering artifacts, with cabinets and walls displaying low-frequency oscillations. BiDAStereo fails to maintain an accurate shape, resulting in a distorted table. For the outdoor scenario in Figure 6, the driving car presents a challenging case due to its transparent surfaces and rapid motion. As shown, all other methods struggle with this case, failing to produce reliable disparity estimations. In contrast, our method demonstrates superior performance, maintaining accurate and stable predictions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.817, + 0.667, + 0.833 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Foundation Priors. As shown in Table 4, incorporating foundation priors from a monocular depth model improves performance. Notably, VDA [7] outperforms DepthAnythingV2 [68] due to its stronger temporal consistency. In" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20841" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.145, + 0.09, + 0.263, + 0.142 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.148, + 0.144, + 0.259, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.149, + 0.17, + 0.259, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.148, + 0.197, + 0.259, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.17, + 0.223, + 0.238, + 0.232 + ], + "angle": 0, + "content": "Left Frames" + }, + { + "type": "image", + "bbox": [ + 0.266, + 0.091, + 0.412, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.271, + 0.2, + 0.412, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.237, + 0.412, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.273, + 0.273, + 0.412, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.248, + 0.269, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.31, + 0.238, + 0.32 + ], + "angle": 0, + "content": "Point Cloud" + }, + { + "type": "image_caption", + "bbox": [ + 0.272, + 0.31, + 0.381, + 0.32 + ], + "angle": 0, + "content": "Selective-IGEV" + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.091, + 0.559, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.417, + 0.2, + 0.559, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.417, + 0.237, + 0.559, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.444, + 0.311, + 0.528, + 0.321 + ], + "angle": 0, + "content": "DynamicStereo" + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.091, + 0.706, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.564, + 0.2, + 0.707, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.237, + 0.707, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.273, + 0.707, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.601, + 0.311, + 0.666, + 0.321 + ], + "angle": 0, + "content": "BiDAStereo" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.091, + 0.854, + 0.126 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.126, + 0.853, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.162, + 0.853, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.203, + 0.854, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.237, + 0.854, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.273, + 0.854, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.31, + 0.845, + 0.321 + ], + "angle": 0, + "content": "Stereo Any Video (ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.337, + 0.907, + 0.392 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison on a dynamic indoor scenario from the South Kensington SV dataset [23]. Top: Input left frames and predicted disparities for the same region across three frames using different methods. Bottom: Disparities converted into globally aligned point clouds and rendered from a camera view for comparison. Failure regions are highlighted with red arrows. Our method produces consistent and accurate disparities without flickering. See video in the supplementary material for better visualization." + }, + { + "type": "image", + "bbox": [ + 0.143, + 0.402, + 0.287, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.465, + 0.285, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.528, + 0.285, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.591, + 0.285, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.656, + 0.247, + 0.665 + ], + "angle": 0, + "content": "Left Frames" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.402, + 0.427, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.466, + 0.427, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.528, + 0.427, + 0.591 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.591, + 0.427, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.655, + 0.396, + 0.665 + ], + "angle": 0, + "content": "Selective-IGEV" + }, + { + "type": "image", + "bbox": [ + 0.428, + 0.402, + 0.569, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.428, + 0.466, + 0.569, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.428, + 0.528, + 0.569, + 0.591 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.428, + 0.591, + 0.569, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.458, + 0.656, + 0.539, + 0.665 + ], + "angle": 0, + "content": "DynamicStereo" + }, + { + "type": "image", + "bbox": [ + 0.57, + 0.402, + 0.711, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.57, + 0.466, + 0.711, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.57, + 0.528, + 0.711, + 0.591 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.57, + 0.591, + 0.711, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.656, + 0.672, + 0.665 + ], + "angle": 0, + "content": "BiDAStereo" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.402, + 0.853, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.466, + 0.853, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.528, + 0.853, + 0.591 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.591, + 0.853, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.72, + 0.656, + 0.844, + 0.666 + ], + "angle": 0, + "content": "Stereo Any Video (ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.175, + 0.683, + 0.82, + 0.697 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison on a dynamic outdoor scenario from the South Kensington SV dataset [23]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.717, + 0.484, + 0.899 + ], + "angle": 0, + "content": "terestingly, increasing the backbone size does not yield additional performance gains. To investigate this, we conducted further experiments (Table 5) by varying depth and CNN feature channels based on VDA-L. However, the results remained similar. According to [7], VDA-L has \\(13 \\times\\) more parameters than VDA-S (381.8M vs. 28.4M), yet the performance gain is marginal (\\(\\delta_{1}(\\uparrow)\\): 0.944 vs. 0.942 on KITTI; 0.971 vs. 0.959 on NYUv2). This suggests that the larger backbone does not provide richer priors. Moreover, large backbone requires additional CNN channels and parameters for adaptation, increasing training complexity. Therefore, we adopt VDA-S in the final model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.717, + 0.907, + 0.853 + ], + "angle": 0, + "content": "Correlation. To evaluate the effectiveness of our proposed all-to-all-pairs correlation, we replace it with local correlation [27] and experiment with using only 1D or 2D correlation instead of alternating between them. As shown in Table 4, all these modifications degrade performance. Notably, constructing 3D correlations—where the search window spans height, width, and time—further worsens results. This suggests that frames cannot be effectively utilized without proper alignment [23]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Cost Volume Encoder. Instead of directly encoding cost volumes with convolutions, we find that introducing an MLP layer to extract a compact representation before ap" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20842" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.146 + ], + "angle": 0, + "content": "Table 4. Ablation study on the foundation priors, correlation, cost volume encoder, convex upsampling, attention, and inference settings of our model on Sintel [2] and Dynamic Replica [24]. The approach used in our final model is underlined." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.504 + ], + "angle": 0, + "content": "
Setting VariationsSintel FinalDynamic Replica
TEPE↓δt1px ↓TEPE↓δt1px ↓
Foundation Priors
w/o Depth1.4712.350.0921.05
Depth AnythingV2-S [68]1.3311.270.0880.99
VDA-S [7]1.2710.980.0830.78
VDA-L [7]1.2012.470.0830.82
Correlation
1D + 2D local [27]1.2710.980.0830.78
1D all-to-all-pairs1.2311.360.0790.78
2D all-to-all-pairs1.2911.950.0850.77
3D all-to-all-pairs4.5122.901.3132.01
1D + 2D all-to-all-pairs1.2111.330.0760.75
Cost Volume Encoder
Convolution1.2611.540.0810.77
MLP + Convolution1.2111.330.0760.75
Convex Upsampling
Bilinear1.2612.740.0851.05
Convex [51]1.2111.330.0760.75
Temporal Convex1.0410.210.0670.79
Attention in GRU
w/o Attention1.0410.210.0670.79
Temporal1.0710.230.0640.79
Temporal + Spatial1.079.980.0570.67
Inference Stage
1/16 → 1/8 → 1/41.079.980.0570.67
1/32 → 1/16 → 1/8 → 1/41.0810.820.0580.70
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.483, + 0.547 + ], + "angle": 0, + "content": "plying convolutions enhances cost aggregation, leading to improved final results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.549, + 0.483, + 0.626 + ], + "angle": 0, + "content": "Convex Upsampling. We compare bilinear upsampling and general convex upsampling [51] to our proposed temporal convex upsampling module. As observed, our approach achieves better temporal consistency by incorporating learnable weights with adjacent frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.483, + 0.689 + ], + "angle": 0, + "content": "Attention in GRU. We further investigate different attention mechanisms within the GRU update block. As shown in Table 4, combining temporal and spatial attention together leads to the best performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.691, + 0.483, + 0.738 + ], + "angle": 0, + "content": "Inference Stage. We also explore inference starting at different resolutions. As can be seen in Table 4, a three-stage inference setup already achieves strong results." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.755, + 0.218, + 0.772 + ], + "angle": 0, + "content": "4.5. Application" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Our method enhances downstream applications that rely on accurate metric depth, such as 3D point tracking. As shown in Figure 7, we visualize results using the state-of-the-art 3D point tracking method DELTA [36]. Our predicted metric depth yields more points with precise trajectories than UniDepth [40], which gives less missing points as highlighted by the red arrows. Additional application, such as visual effects, are available in the supplementary material." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.906, + 0.132 + ], + "angle": 0, + "content": "Table 5. Ablation study on the effect of different channel configurations for depth features and CNN features, based on VDA-L [7], on Sintel [2]. Trainable parameters are also reported." + }, + { + "type": "table", + "bbox": [ + 0.545, + 0.135, + 0.872, + 0.384 + ], + "angle": 0, + "content": "
Depth & CNNSintel FinalPara. (M)
Channel NumberTEPE↓δt1px ↓
VDA-L [7]
(32, 96)1.2012.478.4
(128, 64)1.5814.5213.3
(128, 128)1.3113.2119.3
UniDepthStereo Any Video (ours)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.392, + 0.906, + 0.421 + ], + "angle": 0, + "content": "Figure 7. Visualization of 3D point tracking based on metric depth generated from UniDepth and our method." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.431, + 0.906, + 0.473 + ], + "angle": 0, + "content": "Table 6. Comparison of different methods on model trainable [non-trainable] parameters, inference GPU memory, and multiply-accumulate (MAC) on 20 frames with resolution \\(720 \\times 1280\\)." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.479, + 0.905, + 0.545 + ], + "angle": 0, + "content": "
MethodsPara. (M)Mem. (G)MACs (T)
DynamicStereo [24]20.535.1182.3
BiDASTereo [22]12.241.1186.6
Stereo Any Video (ours)9.4 [28.4]41.1303.4
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.567, + 0.771, + 0.582 + ], + "angle": 0, + "content": "4.6. Limitation and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.59, + 0.906, + 0.68 + ], + "angle": 0, + "content": "We report the inference efficiency results on a single NVIDIA A6000 GPU in Table 6. As shown, due to the use of foundation priors, our method incurs relatively high computational and memory costs. Looking ahead, we aim to develop a model zoo for this method, including both a large and a light version." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.634, + 0.71 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduce Stereo Any Video, a robust framework for video stereo matching. By leveraging rich monocular priors, an all-to-all-pairs correlation mechanism, and a temporal convex upsampling layer, our approach produces temporally consistent and accurate disparities without relying on additional inputs such as camera pose or optical flow. Extensive evaluations show that our model achieves state-of-the-art performance on multiple standard benchmarks under zero-shot settings. Furthermore, our method generalizes well to real-world stereo videos, handling diverse content, motion, and camera movements effectively. Acknowledgments. This work was funded by the IC-CSC." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20843" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Luca Bartolomei, Fabio Tosi, Matteo Poggi, and Stefano Mattoccia. Stereo anywhere: Robust zero-shot deep stereo matching even where either stereo or mono fail. arXiv preprint arXiv:2412.04472, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.214 + ], + "angle": 0, + "content": "[2] D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black. A naturalistic open source movie for optical flow evaluation. In ECCV, pages 611-625, 2012. 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.217, + 0.482, + 0.243 + ], + "angle": 0, + "content": "[3] Yohann Cabon, Naila Murray, and Martin Humenberger. Virtual kitti 2, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.246, + 0.483, + 0.315 + ], + "angle": 0, + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.317, + 0.482, + 0.345 + ], + "angle": 0, + "content": "[5] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In CVPR, pages 5410-5418, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.347, + 0.483, + 0.414 + ], + "angle": 0, + "content": "[6] Tianyu Chang, Xun Yang, Tianzhu Zhang, and Meng Wang. Domain generalized stereo matching via hierarchical visual transformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9559-9568, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.418, + 0.482, + 0.471 + ], + "angle": 0, + "content": "[7] Sili Chen, Hengkai Guo, Shengnan Zhu, Feihu Zhang, Zi long Huang, Jiashi Feng, and Bingyi Kang. Video depth anything: Consistent depth estimation for super-long videos, 2025. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.475, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[8] Junda Cheng, Longliang Liu, Gangwei Xu, Xianqi Wang, Zhaoxing Zhang, Yong Deng, Jinliang Zang, Yurui Chen, Zhipeng Cai, and Xin Yang. Monster: Marry monodepth to stereo unleashes power, 2025. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.532, + 0.482, + 0.599 + ], + "angle": 0, + "content": "[9] Xuelian Cheng, Yiran Zhong, Mehrtash Harandi, Yuchao Dai, Xiaojun Chang, Tom Drummond, Hongdong Li, and Zongyuan Ge. Hierarchical neural architecture search for deep stereo matching. arXiv preprint arXiv:2010.13501, 2020.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.603, + 0.483, + 0.658 + ], + "angle": 0, + "content": "[10] Ziang Cheng, Jiayu Yang, and Hongdong Li. Stereo matching in time: \\(100+\\) fps video stereo matching for extended reality. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 8719-8728, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.661, + 0.482, + 0.714 + ], + "angle": 0, + "content": "[11] Seokju Cho, Jiahui Huang, Jisu Nam, Honggyu An, Seun-gryong Kim, and Joon-Young Lee. Local all-pair correspondence for point tracking. arXiv preprint arXiv:2407.15420, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.717, + 0.483, + 0.813 + ], + "angle": 0, + "content": "[12] WeiQin Chuah, Ruwan Tennakoon, Reza Hoseinnezhad, Alireza Bab-Hadiashar, and David Suter. Itsa: An information-theoretic approach to automatic shortcut avoidance and domain generalization in stereo matching networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13022-13032, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[13] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, pages 3354-3361, 2012. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[14] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2495-2504, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.188 + ], + "angle": 0, + "content": "[15] Weiyu Guo, Zhaoshuo Li, Yongkui Yang, Zheng Wang, Russell H Taylor, Mathias Unberath, Alan Yuille, and Yingwei Li. Context-enhanced stereo transformer. In European Conference on Computer Vision, pages 263-279. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[16] Xiaoyang Guo, Kai Yang, Wukui Yang, Xiaogang Wang, and Hongsheng Li. Group-wise correlation stereo network. In CVPR, pages 3273-3282, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.234, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[17] Xianda Guo, Chenming Zhang, Youmin Zhang, Wenzhao Zheng, Dujun Nie, Matteo Poggi, and Long Chen. Light-stereo: Channel boost is all your need for efficient 2d cost aggregation, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.289, + 0.904, + 0.327 + ], + "angle": 0, + "content": "[18] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.33, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[19] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. arXiv preprint arXiv:2409.02095, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.387, + 0.905, + 0.427 + ], + "angle": 0, + "content": "[20] Hualie Jiang, Zhiqiang Lou, Laiyan Ding, Rui Xu, Minglang Tan, Wenjie Jiang, and Rui Huang. Defom-stereo: Depth foundation model based stereo matching, 2025. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.429, + 0.907, + 0.509 + ], + "angle": 0, + "content": "[21] Junpeng Jing, Jiankun Li, Pengfei Xiong, Jiangyu Liu, Shuaicheng Liu, Yichen Guo, Xin Deng, Mai Xu, Lai Jiang, and Leonid Sigal. Uncertainty guided adaptive warping for robust and efficient stereo matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3318-3327, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.512, + 0.905, + 0.567 + ], + "angle": 0, + "content": "[22] Junpeng Jing, Ye Mao, and Krystian Mikolajczyk. Match-stereo-videos: Bidirectional alignment for consistent dynamic stereo matching. In European Conference on Computer Vision, pages 415–432. Springer, 2024. 2, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.905, + 0.608 + ], + "angle": 0, + "content": "[23] Junpeng Jing, Ye Mao, Anlan Qiu, and Krystian Mikolajczyk. Match stereo videos via bidirectional alignment, 2024. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.61, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[24] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Dynamic stereo: Consistent dynamic depth from stereo videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13229-13239, 2023. 2, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.693, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[25] Alex Kendall, Hayk Martirosyan, Saumitro Dasgupta, Peter Henry, Ryan Kennedy, Abraham Bachrach, and Adam Bry. End-to-end learning of geometry and context for deep stereo regression. In CVPR, pages 66-75, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[26] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Jiankun Li, Peisen Wang, Pengfei Xiong, Tao Cai, Ziwei Yan, Lei Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. Practical stereo matching via cascaded recurrent network with adaptive correlation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16263-16272, 2022. 1, 2, 4, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20844" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[28] Zhaoshuo Li, Xingtong Liu, Nathan Drenkow, Andy Ding, Francis X Creighton, Russell H Taylor, and Mathias Unberath. Revisiting stereo depth estimation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6197-6206, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[29] Zhaoshuo Li, Wei Ye, Dilin Wang, Francis X Creighton, Russell H Taylor, Ganesh Venkatesh, and Mathias Unberath. Temporally consistent online depth estimation in dynamic scenes. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 3018-3027, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.262, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[30] Zhengfa Liang, Yiliu Feng, Yulan Guo, Hengzhu Liu, Wei Chen, Linbo Qiao, Li Zhou, and Jianfeng Zhang. Learning for disparity estimation through feature constancy. In CVPR, pages 2811-2820, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.319, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[31] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. arXiv preprint arXiv:2109.07547, 2021. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.402 + ], + "angle": 0, + "content": "[32] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60:91-110, 2004. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[33] Yang Luo, Xuanlei Zhao, Mengzhao Chen, Kaipeng Zhang, Wenqi Shao, Kai Wang, Zhangyang Wang, and Yang You. Enhance-a-video: Better generated video for free, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[34] Nikolaus Mayer, Eddy Ig, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In CVPR, pages 4040-4048, 2016. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[35] Lukas Mehl, Jenny Schmalfuss, Azin Jahedi, Yaroslava Nalivayko, and Andres Bruhn. Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[36] Tuan Duc Ngo, Peiye Zhuang, Chuang Gan, Evangelos Kalogerakis, Sergey Tulyakov, Hsin-Ying Lee, and Chaoyang Wang. Delta: Dense efficient long-range 3d tracking for any video, 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.714 + ], + "angle": 0, + "content": "[37] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.717, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[38] Jiahao Pang, Wenxiu Sun, Jimmy SJ Ren, Chengxi Yang, and Qiong Yan. Cascade residual learning: A two-stage convolutional neural network for stereo matching. In CVPRW, pages 887-895, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.774, + 0.482, + 0.828 + ], + "angle": 0, + "content": "[39] Jiahao Pang, Wenxiu Sun, Chengxi Yang, Jimmy Ren, Ruichao Xiao, Jin Zeng, and Liang Lin. Zoom and learn: Generalizing deep stereo matching to novel domains. In CVPR, pages 2070-2079, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[40] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Unidepth: Universal monocular metric depth estimation, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[42] Zhibo Rao, Bangshu Xiong, Mingyi He, Yuchao Dai, Renjie He, Zhelun Shen, and Xing Li. Masked representation learning for domain generalized stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5435-5444, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[43] Ignacio Rocco, Mircea Cimpoi, Relja Arandjelović, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Neighbourhood consensus networks. Advances in neural information processing systems, 31, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[44] Daniel Scharstein and Richard Szeliski. A taxonomy and evaluation of dense two-frame stereo correspondence algorithms. IJCV, 47(1):7-42, 2002. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.383 + ], + "angle": 0, + "content": "[45] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In CVPR, pages 3260-3269, 2017. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[46] Zhelun Shen, Yuchao Dai, and Zhibo Rao. Cfnet: Cascade and fused cost volume for robust stereo matching. In CVPR, pages 13906-13915, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.494 + ], + "angle": 0, + "content": "[47] Zhelun Shen, Xibin Song, Yuchao Dai, Dingfu Zhou, Zhibo Rao, and Liangjun Zhang. Digging into uncertainty-based pseudo-label for robust stereo matching. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-18, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[48] Chenyang Si, Weichen Fan, Zhengyao Lv, Ziqi Huang, Yu Qiao, and Ziwei Liu. Repvideo: Rethinking cross-layer representation for video generation, 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[49] Qing Su and Shihao Ji. Chitransformer: Towards reliable stereo from cues. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1939-1949, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[50] Vladimir Tankovich, Christian Hane, Yinda Zhang, Adarsh Kowdle, Sean Fanello, and Sofien Bouaziz. Hitnet: Hierarchical iterative tile refinement network for real-time stereo matching. In CVPR, pages 14362-14372, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[51] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, pages 402-419, 2020. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.693, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[52] Prune Truong, Martin Danelljan, Luc V Gool, and Radu Timofte. Gocor: Bringing globally optimized correspondence volumes into your neural network. Advances in Neural Information Processing Systems, 33:14278-14290, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[53] Prune Truong, Martin Danelljan, and Radu Timofte. Glunet: Global-local universal network for dense flow and correspondences. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6258-6268, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[54] Jonas Uhrig, Nick Schneider, Lukas Schneider, Uwe Franke, Thomas Brox, and Andreas Geiger. Sparsity invariant cnns. In 2017 international conference on 3D Vision (3DV), pages 11-20. IEEE, 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[55] Xianqi Wang, Gangwei Xu, Hao Jia, and Xin Yang. Selective-stereo: Adaptive frequency information selection" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20845" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19701-19710, 2024. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[56] Yan Wang, Zihang Lai, Gao Huang, Brian H Wang, Laurens Van Der Maaten, Mark Campbell, and Kilian Q Weinberger. Anytime stereo image depth estimation on mobile devices. In 2019 international conference on robotics and automation (ICRA), pages 5893-5900. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.483, + 0.275 + ], + "angle": 0, + "content": "[57] Yiran Wang, Min Shi, Jiaqi Li, Zihao Huang, Zhiguo Cao, Jianming Zhang, Ke Xian, and Guosheng Lin. Neural video depth stabilizer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9466-9476, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.483, + 0.374 + ], + "angle": 0, + "content": "[58] Philippe Weinzaepfel, Thomas Lucas, Vincent Leroy, Yohann Cabon, Vaibhav Arora, Romain Brégier, Gabriela Csurka, Leonid Antsfeld, Boris Chidlovskii, and Jérôme Revaud. Croco v2: Improved cross-view completion pretraining for stereo matching and optical flow. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17969-17980, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.483, + 0.417 + ], + "angle": 0, + "content": "[59] Bowen Wen, Matthew Trepte, Joseph Aribido, Jan Kautz, Orazio Gallo, and Stan Birchfield. Foundationstereo: Zeroshot stereo matching, 2025. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.483, + 0.475 + ], + "angle": 0, + "content": "[60] Bin Xu, Yuhua Xu, Xiaoli Yang, Wei Jia, and Yulan Guo. Bilateral grid learning for stereo matching networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12497-12506, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.483, + 0.545 + ], + "angle": 0, + "content": "[61] Gangwei Xu, Xianqi Wang, Xiaohuan Ding, and Xin Yang. Iterative geometry encoding volume for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21919-21928, 2023. 1, 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.483, + 0.602 + ], + "angle": 0, + "content": "[62] Gangwei Xu, Yun Wang, Junda Cheng, Jinhui Tang, and Xin Yang. Accurate and efficient stereo matching via attention concatenation volume. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.603, + 0.483, + 0.645 + ], + "angle": 0, + "content": "[63] Gangwei Xu, Huan Zhou, and Xin Yang. Cgi-stereo: Accurate and real-time stereo matching via context and geometry interaction. arXiv preprint arXiv:2301.02789, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.483, + 0.713 + ], + "angle": 0, + "content": "[64] Gangwei Xu, Xianqi Wang, Zhaoxing Zhang, Junda Cheng, Chunyuan Liao, and Xin Yang. Igev++: Iterative multirange geometry encoding volumes for stereo matching. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.716, + 0.483, + 0.758 + ], + "angle": 0, + "content": "[65] Haofei Xu and Juyong Zhang. Aanet: Adaptive aggregation network for efficient stereo matching. In CVPR, pages 1959-1968, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.759, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[66] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, Fisher Yu, Dacheng Tao, and Andreas Geiger. Unifying flow, stereo and depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[67] Gengshan Yang, Joshua Manela, Michael Happold, and Deva Ramanan. Hierarchical deep stereo matching on high-resolution images. In CVPR, pages 5515-5524, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[68] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.164 + ], + "angle": 0, + "content": "[69] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.207 + ], + "angle": 0, + "content": "[70] Jiaxi Zeng, Chengtang Yao, Yuwei Wu, and Yunde Jia. Temporally consistent stereo matching. In European Conference on Computer Vision, pages 341-359. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.905, + 0.249 + ], + "angle": 0, + "content": "[71] Feihu Zhang, Victor Prisacariu, Ruigang Yang, and Philip HS Torr. Ga-net: Guided aggregation net for end-to-end stereo matching. In CVPR, pages 185-194, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.345 + ], + "angle": 0, + "content": "[72] Jiawei Zhang, Xiang Wang, Xiao Bai, Chen Wang, Lei Huang, Yimin Chen, Lin Gu, Jun Zhou, Tatsuya Harada, and Edwin R Hancock. Revisiting domain generalized stereo matching networks from a feature consistency perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13001-13011, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.347, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[73] Youmin Zhang, Matteo Poggi, and Stefano Mattoccia. Temporalstereo: Efficient spatial-temporal stereo matching network. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 9528-9535. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[74] Yongjian Zhang, Longguang Wang, Kunhong Li, Yun Wang, and Yulan Guo. Learning representations from foundation models for domain generalized stereo matching. In European Conference on Computer Vision, pages 146-162. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.489, + 0.905, + 0.544 + ], + "angle": 0, + "content": "[75] Yiran Zhong, Hongdong Li, and Yuchao Dai. Open-world stereo video matching with deep rnN. In Proceedings of the European Conference on Computer Vision (ECCV), pages 101-116, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.545, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[76] Jingyi Zhou, Haoyu Zhang, Jiakang Yuan, Peng Ye, Tao Chen, Hao Jiang, Meiya Chen, and Yangyang Zhang. All-in-one: Transferring vision foundation models into stereo matching. arXiv preprint arXiv:2412.09912, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20846" + } + ] +] \ No newline at end of file diff --git a/2025/Stereo Any Video_ Temporally Consistent Stereo Matching/dc20cc87-36ce-4c69-8c57-afd220728c97_origin.pdf b/2025/Stereo Any Video_ Temporally Consistent Stereo Matching/dc20cc87-36ce-4c69-8c57-afd220728c97_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0ce8db10e90f8a9b018e9979cb8329160bcea9cc --- /dev/null +++ b/2025/Stereo Any Video_ Temporally Consistent Stereo Matching/dc20cc87-36ce-4c69-8c57-afd220728c97_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6cb22d55a413648a644c35a2f401c0c160ccfd004519b2e13ba2bd13847c790a +size 3828038 diff --git a/2025/Stereo Any Video_ Temporally Consistent Stereo Matching/full.md b/2025/Stereo Any Video_ Temporally Consistent Stereo Matching/full.md new file mode 100644 index 0000000000000000000000000000000000000000..db60a61a82eed35fd6ad402a136c3872eae6069f --- /dev/null +++ b/2025/Stereo Any Video_ Temporally Consistent Stereo Matching/full.md @@ -0,0 +1,440 @@ +# Stereo Any Video: Temporally Consistent Stereo Matching + +Junpeng Jing Weixun Luo Ye Mao Krystian Mikolajczyk Imperial College London + +https://tomtomommi.github.io/StereoAnyVideo/ + +![](images/7c752cb5b818cb77259978201f44bf7ee5c147a62dd04656f4e4508e93bbec34.jpg) + +![](images/fc00e09b027f94ce0785721c9e8dba417e39d76e95c153da6a3293a6ad84455b.jpg) +Figure 1. Prediction on a dynamic real-world stereo video. Our method produces temporally consistent and accurate disparities without flickering, enabling the reconstruction of dense and clean metric point clouds for dynamic scenes. See video for better visualization. + +![](images/d6c877e62b44edb971474df51bd73253c50018dc3357f739b3e5343b25433b47.jpg) + +![](images/62572c3e1fc06cfb7337db396d22c885af6e05b88bb59129c7619448c56eb876.jpg) +time + +![](images/b6ac54c158a8ea5b5f267eeb44f06556a7c205fe05ada34f25755f7fbfddd87b.jpg) + +![](images/2d46a1de5fe1d22655eae1abe29a07699a5e6f98a8bec2ad366c5fa791801589.jpg) + +![](images/82143c2df08f4a6e61c6632aeeeb9f84e71f70de2044b12f6e9a92a38deab6ed.jpg) + +# Abstract + +This paper introduces Stereo Any Video, a powerful framework for video stereo matching. It can estimate spatially accurate and temporally consistent disparities without relying on auxiliary information such as camera poses or optical flow. The strong capability is driven by rich priors from monocular video depth models, which are integrated with convolutional features to produce stable representations. To further enhance performance, key architectural innovations are introduced: all-to-all-pairs correlation, which constructs smooth and robust matching cost volumes, and temporal convex upsampling, which improves temporal coherence. These components collectively enhance robustness, accuracy, and temporal consistency, establishing a new standard in video stereo matching. Extensive experiments demonstrate that our method achieves state-of-the + +art performance across multiple datasets both qualitatively and quantitatively in zero-shot settings, as well as strong generalization to real-world indoor and outdoor scenarios. + +# 1. Introduction + +Video stereo matching is a task to reconstruct a 3D scene from a stereo camera. It estimates disparity by identifying the horizontal correspondences between rectified left and right images at each time step. Consistent and accurate disparity estimation is crucial for reconstructing and understanding real-world 3D scenes, benefiting various downstream tasks such as autonomous driving, robotic navigation, and virtual/augmented reality. + +With the advances in deep learning, stereo matching has seen significant progress [27, 31, 55, 61], achieving promising performance on public benchmarks [13, 44, 45]. + +However, most existing approaches are designed for image-based estimation and do not explicitly leverage temporal information in video sequences. When applied directly to video data, these methods often suffer from temporal inconsistencies, leading to noticeable flickering and artifacts in disparity maps and reconstructed point clouds. To mitigate this, recent video-based methods incorporate additional temporal cues, such as camera poses or optical flow, through predefined models or learnable mechanisms. For example, CODD [29] employs a pose network to estimate relative camera movements and align adjacent disparities within a fusion framework. Similarly, TemporalStereo [73] and TC-Stereo [70] follow this paradigm with variations in architectural design and alignment operations. Another approach, BiDAStereo [22], integrates optical flow [51] for temporal feature alignment and introduces a temporal aggregation module to propagate information across time. + +Despite these advances, existing video methods remain vulnerable to dynamic scenes, camera movement, and the accuracy of auxiliary information. In real-world scenarios, dynamic objects undergo complex motions and deformations that violate traditional multi-view constraints [18], disrupting translational invariance in disparity estimation. Moreover, models relying on auxiliary cues are constrained by the performance of camera pose or optical flow estimation modules. In environments with complex camera movements, these estimation modules degrade, becoming bottlenecks that limit the robustness of video stereo methods. + +To address these limitations, we propose Stereo Any Video, a robust framework for video-based stereo matching without relying on camera pose or optical flow. Inspired by recent findings in video generation [33, 48], which emphasize the role of robust feature representations in ensuring temporal coherence, we take a new perspective to design the framework for consistency: feature robustness and stability. Specifically, during feature extraction, we integrate convolutional cues with a frozen monocular video depth foundation model [7] trained on million-scale data. This ensures robust feature maps with rich priors. For correspondence matching, we introduce an all-to-all-pair correlation, which enforces smooth and reliable matching by jointly considering bidirectional correspondences from both target and reference views. For cost aggregation, we propose temporal convex upsampling within a fully temporal-based recurrent unit, playing a key role in maintaining temporal consistency. As shown in Figure 1, our method demonstrates a strong ability to produce accurate and consistent disparities, benefiting the reconstruction for dynamic scenes. + +The contributions of our work can be summarized as follows: (1) Stereo Any Video, a framework that integrates monocular video depth priors to achieve accurate and temporally consistent video stereo matching. (2) A novel all-to-all-pair correlation module and a temporal convex upsam + +pling mechanism, ensuring robust and stable cost volume representations and aggregation. (3) State-of-the-art performance on diverse benchmarks, setting a new standard for video stereo matching. + +# 2. Related Work + +In this section, we provide a brief literature review of stereo matching methods in the deep learning era. We first discuss image-based stereo matching methods, followed by video-based methods, and finally, we review recent advancements leveraging foundation priors. + +Image-based Stereo Matching. Since the introduction of the first end-to-end stereo network [34], deep learning-based stereo matching has seen substantial advancements. Early methods typically construct 3D [16, 30, 38, 50, 65] or 4D [5, 9, 14, 25, 62, 67, 71] cost volumes from left and right image features, followed by 2D or 3D convolutions for cost aggregation. Inspired by RAFT [51], iterative refinement has become the mainstream approach. Representative advancements include multilevel recurrent units [31], adaptive correlation [27], geometry-encoding volumes [61, 64], and frequency information selection [55], all of which have significantly improved benchmark performance [13, 44, 45]. Transformer-based methods [15, 28, 49, 58, 66] have further refined stereo matching by modeling token correspondence and capturing long-range dependencies. Beyond accuracy, several methods focus on real-world challenges such as efficiency [17, 56, 60, 63], domain generalization [6, 12, 39, 42, 72], and robustness [21, 46, 47]. However, as all these methods are designed for static image pairs, they lack temporal awareness, leading to inconsistent disparity predictions in video sequences. + +Video-based Stereo Matching. Video stereo matching especially in dynamic scenes presents unique challenges. Early work, such as [75], employed LSTMs with unsupervised learning for stereo videos but did not explicitly incorporate temporal information. CODD [29] introduced a per-frame stereo network, a motion network, and a fusion network designed for dynamic scenes. Neighboring-frame disparities are aligned and fused using estimated camera poses. TemporalStereo [73] extended this approach with a coarse-to-fine network that leverages past context to enhance predictions in challenging scenarios. Cheng et al. [10] developed a framework tailored for XR systems, reducing computational costs through temporal cost aggregation. TC-Stereo [70] further introduced a temporal disparity completion mechanism and a temporal state fusion module to exploit temporal information. These methods rely on auxiliary camera motion or pre-defined scene geometry to incorporate temporal information. DynamicStereo [24] introduced a transformer-based architecture that integrates temporal, stereo, and spatial attention. BiDAStereo [22] proposed an optical-flow-based alignment mechanism to propagate + +![](images/b0642654f4f12eaa27541d472af4f24e2d80d5eb84e1e5d6257aae5ee4222608.jpg) +Figure 2. Overview of Stereo Any Video. Given a stereo video sequence as input, the model first extracts features and context information using trainable convolutional encoders and a frozen monocular video depth encoder (Video Depth Anything [7]). At each iteration, an all-to-all-pair correlation computes feature correlations, followed by a multilayer perception (MLP) encoder for compact representation. Disparities are iteratively refined using a 3D Gated Recurrent Unit (GRU), integrated with cost volume features, and upsampled via a temporal convex upsampling layer. This process is repeated within a cascaded pipeline to progressively recover full-resolution disparities. + +temporal information across the entire sequence. Building on this alignment, BiDAS stabilizer [23] introduced a lightweight plug-and-play stabilizer network to improve the temporal consistency of existing image-based stereo matching methods. + +Stereo Matching with Foundation Priors. Vision foundation models, such as CLIP [41], the DINO series [4, 37], and SAM [26], have been widely applied across various vision tasks due to their rich priors. In stereo matching area, FormerStereo [74] proposed feature transformation and C4 space to adapt ViT-based foundation model to stereo pipeline, learning domain invariant representations. AIOStereo [76] introduced a selective knowledge transfer module to distill knowledge from multiple heterogeneous vision foundation models into a single stereo matching model. More recent approaches primarily focused on leveraging monocular depth priors from Depth Anything [68, 69] to enhance stereo performance. For instance, Stereo Anywhere [1] directly incorporated relative depth maps into the stereo network, demonstrating robustness in challenging conditions such as non-Lambertian surfaces. DEFOM-Stereo [20] employed these depth maps alongside a novel scale updater within a recurrent module for improved alignment. FoundationStereo [59] extracted feature maps from Depth Anything and introduced a side-tuning adapter to adapt monocular features for stereo backbone. MonSter [8] utilized both depth maps and feature maps, and proposed a dualbranched refinement module to fuse stereo and monocular cues effectively. These methods highlight the effectiveness of foundation priors in image-based stereo tasks. In our work, we build upon this direction by investigating video-based stereo matching and exploring how foundation priors can enhance consistency in dynamic scenes. + +# 3. Method + +In this section, we introduce Stereo Any Video, a feedforward video-based framework. Given a rectified stereo sequence $\{\mathbf{I}_L^t,\mathbf{I}_R^t\}_{t\in (1,T)}\in \mathbb{R}^{H\times W\times 3}$ where $T$ denotes the total number of frames, $H$ and $W$ are the frame height and width, respectively, our goal is to estimate a temporally consistent sequence of disparity maps $\{\mathbf{d}^t\}_{t\in (1,T)}\in \mathbb{R}^{H\times W}$ aligned with the left view. As shown in Figure 2, our method consists of three key components: feature extraction (Section 3.1), correlation (Section 3.2), and temporal cost aggregation (Section 3.3). + +# 3.1. Feature Extraction with Foundation Priors + +Unlike image-based tasks, maintaining temporal consistency is the primary challenge in video stereo matching. Recent studies in video generation [33, 48] indicate that substantial variations across different frames can lead to unstable feature representations, thereby degrading temporal coherence. Therefore, ensuring feature stability across adjacent frames is crucial for achieving temporally consistent disparity estimation. Towards this, we leverage Video Depth Anything (VDA) [7] for feature extraction, which offers two key advantages: (1) trained on internet-scale data, VDA provides rich priors that can enhance disparity estimation in regions with ambiguous stereo correspondence; (2) unlike image-based monocular depth models [68, 69], VDA generates temporally stable features through an explicit temporal alignment process before the final prediction head, thereby preserving coherence across frames. + +Besides, we investigate whether the depth maps from monocular video depth models can further improve stereo performance. We compare DepthCrafter [19], a representative monocular video model, with RAFTStereo [31], an + +Table 1. Quantitative comparison of DepthCrafter [19] and RAFT-Stereo [31] on Sintel dataset [2]. + +
MethodSintel Clean
EPE ↓TEPE ↓MOS (human) ↓
DepthCrafter [19]8.682.080.212
RAFTStereo [31]1.420.840.933
+ +![](images/7a70cb3d576351db04a2ea923a546ed19c954cdfdcdb42c0361bf141f1382948.jpg) +Input bamboo 2 + +![](images/31f7cbdab9f83f7e21c42cf20500bbf76d372bb4821712726d60e0b428a43b00.jpg) + +![](images/f95bb896f959ab1811c60258011db38988420105d2b9d06816b585a0b4a397e2.jpg) + +![](images/6356f26fb1ae72c2f4334278f83bb72f1e390dc07642cb5c000e74b9bf8e28b3.jpg) +Input sleeping_1 +Figure 3. Qualitative comparison of DepthCrafter [19] and RAFT-Stereo [31] on Sintel dataset [2]. + +![](images/0c98399bcd7f755bc365c3bbb13e7114752fdbd7b20b399494de856b3db596ea.jpg) +DepthCrafter +DepthCrafter + +![](images/0d8c45cfda0c6876d215645a2237c716f33d00dc0669ba0a0018f750fc3e0058.jpg) +RAFTStereo +RAFTStereo + +image-based stereo model. Using open-source weights, we evaluate their performance on Sintel [2], as shown in Table 1 and Figure 3. Specifically, we first obtain relative depth maps from DepthCrafter (without scale and shift) and compute disparities using RAFTStereo on the same sequence. The relative depth maps are then aligned to RAFTStereo's disparities, and objective metrics are computed against the ground truth. Additionally, we conduct a subjective user study (see supplementary material for details). As observed, while DepthCrafter produces visually plausible results—reflected in better human scores—its spatial and temporal errors are significantly larger than those of RAFTStereo. This suggests that its relative depth is "consistent but wrong", making it unreliable for directly improving consistency within a stereo framework. Therefore, we only utilize its features as valuable priors. + +As shown in Figure 2, our feature extraction module consists of two trainable encoders, each employing a series of residual blocks [31] to extract convolutional image features and context features. To integrate VDA priors, we adopt a shallow convolutional adapter [59] that downsamples depth features from the frozen VDA. The extracted depth features are then concatenated with convolutional features to form $\{\mathbf{F}_L,\mathbf{F}_R\}$ and $\mathbf{F}_C$ , providing more stable feature representations for subsequent processing. + +# 3.2. All-to-all-pair Correlation + +At the $n$ -th iteration for each time step, the right feature $\mathbf{F}_R$ is first warped toward the left feature $\mathbf{F}_L$ using the previous disparity $d_{n-1}$ , producing $\hat{\mathbf{F}}_R$ . Previous methods [31, 61] typically construct cost volumes $\mathbf{C}_n$ via all-pair correlation, employing a lookup mechanism to search for potential matches within a local radius. In [27], this approach was extended to an alternating local 1D and 2D search: + +$$ +\mathbf {C} _ {n} (x, y) = \left\langle \mathbf {F} _ {L} (x, y), \hat {\mathbf {F}} _ {R} (x + r _ {x}, y + r _ {y}) \right\rangle , \tag {1} +$$ + +![](images/a6241fd41074fcb9c76091307049e3fb33864d44ed4e05e3c5eeacc11173b1ee.jpg) +Figure 4. Workflow of the temporal convex upsampling. Each pixel of the high resolution disparity (yellow box) is calculated via its 27 low resolution neighbors using learnable weights. + +where $(x,y)$ denotes the position in $\mathbf{F}_L$ , $\langle \cdot, \cdot \rangle$ represents the channel-wise product operation, and $(r_x, r_y)$ defines the search range, typically set as $(r_x, r_y) \in \{(\pm 4,0), (\pm 1, \pm 1)\}$ . The fundamental principle of these methods is to construct one-directional correspondence, where a reference point is selected and directly matched with target points via multiplication. Inspired by recent advances in point tracking [11], we introduce bidirectional correspondence and develop a local all-to-all-pair correlation. Instead of computing correlation in a fixed direction, we calculate the similarity between every pair of potential matching points within two search windows, formulated as: + +$$ +\mathbf {C} _ {n} (x, y) = \left\langle \mathbf {F} _ {L} \left(x + r _ {x}, y + r _ {y}\right), \hat {\mathbf {F}} _ {R} \left(x + r _ {x}, y + r _ {y}\right) \right\rangle . \tag {2} +$$ + +This approach enhances match verification and reduces ambiguity [32, 43]. While BiDAStereo [22] constructs temporal cost volumes by aligning adjacent frames for consistency, its performance is significantly constrained by the limitations of optical flow. In contrast, our method emphasizes robust spatial correlation, leveraging dense correspondences to enforce matching smoothness and enhance consistency and robustness across neighboring points [52, 53]. + +# 3.3. Temporal Cost Aggregation + +At the $n$ -th iteration, the all-to-all-pair correlation expands the potential matching points in $\mathbf{C}_n$ , increasing its dimensionality from $\mathbb{R}^{H' \times W' \times (2r_x + 1)(2r_y + 1)}$ to $\mathbb{R}^{H' \times W' \times (2r_x + 1)^2(2r_y + 1)^2}$ , where $H'$ and $W'$ denote the height and width of frame feature, respectively. We first encode $\mathbf{C}_n$ into a compact representation $\mathbf{E}_n \in \mathbb{R}^{H' \times W' \times L}$ using a MLP: + +$$ +\mathbf {E} _ {n} = \operatorname {M L P} \left(\mathbf {C} _ {n}\right). \tag {3} +$$ + +Next, we employ a 3D-GRU [24] to iteratively refine the disparity for each frame while preserving temporal consistency. The GRU aggregates information using separable 3D convolutions across both spatial and temporal dimensions. At iteration $n$ , it takes as input the encoded cost volume $\mathbf{E}_n$ + +Table 2. Zero-shot generalization results. All methods were trained only on SceneFlow [34]. The representative image-based and video-based stereo matching models are compared on Sintel [2], Dynamic Replica [24], Infinigen SV [23], and Virtual KITTI2 [3]. The weights and parameters are fixed for evaluation. The best and second-best results are highlighted. + +
MethodSintel FinalDynamic ReplicaInfinigen SVVirtual KITTI2
TEPE↓δt1px↓EPE↓TEPE↓δt1px↓EPE↓TEPE↓δt1px↓EPE↓TEPE↓δt1px↓EPE↓
Image-based
RAFTStereo [31]2.0913.704.890.1452.020.382.3319.803.771.2518.391.45
IGEVStereo [61]1.7412.652.440.0810.800.202.3519.333.811.4218.881.39
Selective-IGEV [55]1.9612.952.820.1331.290.292.8521.835.691.4119.101.38
DEFOM-Stereo [20]2.3312.432.380.0800.750.132.0118.913.241.0716.631.34
MonSter [8]1.7011.063.300.1570.860.451.6515.732.270.7310.930.67
Video-based
DynamicStereo [24]1.4311.804.260.0830.730.241.9815.684.671.0615.631.47
BiDAStereo [22]1.2611.651.990.0911.040.651.9916.506.521.0214.821.43
Stereo Any Video (ours)1.079.981.750.0570.670.251.6514.273.160.7412.451.01
+ +context feature $\mathbf{F}_c$ , and the disparities from the previous iteration $d_{n-1}$ . To enhance representation, we integrate super kernels [22], temporal and spatial attention [24] in this process. The disparity update in 3D-GRU is as follows: + +$$ +x _ {n} = \left[ \operatorname {C o n v 2 d} \left(\mathbf {E} _ {n}\right), \operatorname {C o n v 2 d} \left(d _ {n - 1}\right), d _ {n - 1}, \mathbf {F} _ {c} \right], \tag {4} +$$ + +$$ +\widetilde {x _ {n}} = \operatorname {A t t e n t i o n} \left(x _ {n}\right), \tag {5} +$$ + +$$ +z _ {n} = \sigma \left(\operatorname {C o n v 3 d} _ {z} \left(\left[ h _ {n - 1}, \widetilde {x _ {n}} \right]\right)\right), \tag {6} +$$ + +$$ +r _ {n} = \sigma \left(\operatorname {C o n v 3 d} _ {r} \left(\left[ h _ {n - 1}, \widetilde {x _ {n}} \right]\right)\right), \tag {7} +$$ + +$$ +q _ {n} = \tanh \left(\operatorname {C o n v} 3 \mathrm {d} _ {q} \left(\left[ h _ {n - 1} \odot r _ {n}, \widetilde {x _ {n}} \right]\right)\right), \tag {8} +$$ + +$$ +h _ {n} = \left(1 - z _ {n}\right) \odot h _ {n - 1} + z _ {n} \odot q _ {n}, \tag {9} +$$ + +$$ +d _ {n} = d _ {n - 1} + \operatorname {C o n v} 3 \mathrm {d} \left(h _ {n}\right), \tag {10} +$$ + +where $[\cdot ]$ is concatenation, $\odot$ denotes element-wise multiplication, and $\sigma$ is the sigmoid activation function. + +While prior methods emphasize temporal consistency in correlation and cost aggregation, the upsampling process remains overlooked, despite its crucial role in achieving accurate and temporally consistent disparities. Towards this, we adapt the commonly used convex upsampling approach [51] to a temporal variant. In Figure 4, each high-resolution disparity pixel (yellow box) is computed as a convex combination of its 27 low resolution neighbors across the 3 frames (green, blue, and faint yellow boxes). The network predicts the corresponding $3 \times 3 \times 3$ weights using 3D convolutions, enabling a temporally consistent upsampling: + +$$ +\mathbf {w} = \operatorname {s o f t m a x} \left(\operatorname {C o n v 3 d} \left(h _ {n}\right)\right), \tag {11} +$$ + +$$ +D _ {n} ^ {t} = \alpha \cdot \sum_ {i j k} \mathbf {w} _ {i j k} \odot \text {u n f o l d} \left(\left[ d _ {n} ^ {t - 1}, d _ {n} ^ {t}, d _ {n} ^ {t + 1} \right]\right), \tag {12} +$$ + +where $\alpha$ is upsampling rate, $D_{n}^{t}$ is the upsampled disparity. + +# 3.4. Loss Function + +Recent works have introduced OPW loss [57] and TGM loss [7] in monocular video depth estimation to enhance + +temporal consistency. However, we found out that these losses impose a trade-off between accuracy and consistency without yielding substantial improvement in the video stereo task. Thus, we opt for a purely image-based $L_{1}$ loss while still achieving strong temporal consistency as: + +$$ +\mathcal {L} = \sum_ {t = 1} ^ {T} \sum_ {n = 1} ^ {N} \gamma^ {N - n} \left| \left| D _ {\mathrm {g t}} ^ {t} - D _ {n} ^ {t} \right| \right|, \tag {13} +$$ + +where $T$ is the number of frames, $N$ is the number of iterations, and $\gamma$ is a weighting coefficient set to 0.9. Here, $D_{\mathrm{gt}}^t$ denotes the ground truth disparity for the $t$ -th frame. The final output is obtained from the last iteration prediction. + +# 4. Experiments + +In this section, we first introduce the datasets and metrics used in the evaluation, as well as implementation details. We then present the zero-shot generalization capabilities of our method, followed by ablation studies to validate the efficacy of each component in our framework. Please refer to the supplementary material for more details. + +# 4.1. Benchmark Datasets and Metric + +Dataset. Our work focuses on videos captured with moving cameras, making commonly used image benchmarks unsuitable, such as Middlebury [44], ETH3D [45], and KITTI 2012/2015 [13] (two frames per scene). To evaluate our method, we use five synthetic and two real-world stereo video datasets, all featuring dynamic scenes. Sintel [2] is a synthetic dataset comprising movie sequences available in both clean and final passes. Spring [35] is a high resolution synthetic dataset. Dynamic Replica [24] is a synthetic indoor dataset featuring non-rigid objects, such as animals and people. Infinigen SV [23] is a synthetic dataset focusing on outdoor natural environments. Virtual KITTI2 [3] is a synthetic dataset simulating outdoor driving scenarios. KITTI Depth [54] is a real-world outdoor dataset captured for autonomous driving applications, with sparse + +Table 3. Quantitative comparisons on Spring [35], Sintel [2], and KITTI Depth [54] where methods are allowed to train on any existing datasets. Model weights remain fixed during evaluation. Training datasets sharing the same domain as the test sets are highlighted. “-” represents the test set has been used in training. Abbreviations: SF - SceneFlow [34], DR - Dynamic Replica [24], ISV - Infigen SV [23], M - Middlebury [44], KD - KITTI Depth [54], VK - Virtual KITTI2 [3]. The best and second-best results are highlighted. + +
MethodTraining DataScaleSpringSintel FinalKITTI Depth
TEPE↓δt1px↓EPE↓TEPE↓δt1px↓EPE↓TEPE↓δt1px↓EPE↓
Image-based
RAFTStereo [31]10 mixed datasets0.6M1.499.795.37---0.417.960.51
IGEVStereo [61]SF+M0.04M1.539.725.371.5313.492.230.469.120.60
Selective-IGEV [55]SF+M0.04M1.469.335.461.6413.782.140.489.660.60
DEFOM-Stereo [20]15 mixed datasets0.7M1.299.124.25---0.397.080.48
MonSter [8]10 mixed datasets0.6M1.258.655.19---0.417.730.60
FoundationStereo [59]7 mixed datasets1.3M1.789.825.40---0.407.200.57
Video-based
DynamicStereo [24]SF+DR+ISV+KD0.1M1.079.649.361.3811.843.380.428.210.57
BiDAStereo [22]SF+DR+ISV+KD0.1M0.907.903.921.3311.341.970.428.330.58
Stereo Any Video (ours)SF+DR+ISV+VK0.1M0.776.523.170.999.091.480.356.050.61
+ +depth maps obtained from a LiDAR sensor. South Kensington SV [23] is a real-world dataset capturing daily scenarios from a stereo camera for qualitative evaluation. + +Metric. For accuracy evaluation, we use End-Point Error (EPE), which measures the average per-pixel disparity error. To assess temporal consistency, we employ Temporal EPE (TEPE) [24], which quantifies the variation in EPE across the temporal dimension. We also compute $\delta_{n\text{-px}}^t$ , which represents the proportion of pixels with TEPE exceeding $n$ . For all metrics, lower values indicate better performance. + +# 4.2. Implementation Details + +We implement Stereo Any Video in PyTorch and train on NVIDIA A100 GPUs. The model is trained with a batch size of 8 on SceneFlow [34] for $120k$ iterations and finetuned on a mixed dataset for $80k$ iterations consisting of Dynamic Replica [24] training set, Infinigen SV [23] training set, and Virtual KITTI2 [3]. Before being fed into the network, sequences are randomly cropped to $256 \times 512$ . We use the AdamW optimizer with a learning rate of $2e^{-4}$ and apply a one-cycle learning rate schedule. The full training process takes approximately 6 days. Following [24], we apply various data augmentation techniques, including random cropping, rescaling, and saturation shifts. The sequence length is set to $T = 5$ during training, and $T = 20$ for evaluation. The number of iterations in GRU is set to $N = 10$ in training, and $N = 20$ in evaluation. + +# 4.3. Evaluation + +Quantitative Results. Table 2 presents the results when models are trained solely on SceneFlow [34] and evaluated on four dynamic synthetic datasets including diverse indoor and outdoor environments. The results clearly demonstrate that our model achieves state-of-the-art zero-shot generalization across nearly all metrics, excelling in both accuracy and temporal consistency. Table 3 reports the results + +when models are allowed to train on any existing datasets. Notably, despite being trained solely on synthetic datasets, our model surpasses comparison methods that are trained on real-world datasets or share the same domain as the test tests by at least $15\%$ on diverse benchmarks. Specifically, compared to the robust version of RAFT-Stereo, which is trained on 10 mixed datasets, including in-domain sets such as Sintel [2] and KITTI [13], our approach achieves superior performance across most metrics. These results highlight the strong generalization capability of our method when applied to unseen domains. + +Qualitative Results. Figure 5 and Figure 6 present qualitative results using the checkpoints from Table 3 on dynamic real-world stereo videos from the South Kensington SV dataset [23]. For the indoor scenario shown in Figure 5, the upper rows display three input left frames along with their predicted disparities, while the subsequent rows visualize the globally aligned reconstructed point cloud. The red arrows highlight that our method produces temporally consistent and accurate disparities, whereas other methods exhibit noticeable artifacts. Specifically, Selective-IGEV and DynamicStereo suffer from flickering artifacts, with cabinets and walls displaying low-frequency oscillations. BiDAStereo fails to maintain an accurate shape, resulting in a distorted table. For the outdoor scenario in Figure 6, the driving car presents a challenging case due to its transparent surfaces and rapid motion. As shown, all other methods struggle with this case, failing to produce reliable disparity estimations. In contrast, our method demonstrates superior performance, maintaining accurate and stable predictions. + +# 4.4. Ablation Study + +Foundation Priors. As shown in Table 4, incorporating foundation priors from a monocular depth model improves performance. Notably, VDA [7] outperforms DepthAnythingV2 [68] due to its stronger temporal consistency. In + +![](images/736c3fda992504b6dc731ac7da637f6d8941eb65923948b509378eab1b34dbb4.jpg) + +![](images/f160bca8a8e4544ba702f3a377734cc92121b30eedd4d6716b0233c7317f5c08.jpg) + +![](images/9a565c3341946218f42901cd9a98f1610e93a7b30c14211333ae6d03cdc28801.jpg) + +![](images/44288cbd317c07b07ba5af3cd445785637d0ac5ed8d36a74199fe7ecf91b78b9.jpg) +Left Frames + +![](images/b7488bb0df63d97a0e0d2df13475336313207a18ea7be9f5198ecd2d584b60e9.jpg) + +![](images/d44f8970934e84c5523e58e54e206acdd7a26bf2aa7900b04d73d57af5942749.jpg) + +![](images/85dbea3b21284cdf3055372e945442dac28f8d1b50f8beb38c8839304edbe66d.jpg) + +![](images/ea10315e139192c6cbe6c8d6c96e675719601c174a345938717bcaf990e1d9a2.jpg) +Selective-IGEV + +![](images/d98300b824e70c39d81ed932274f6dee03db0533785a3cd9f6dc3775dcda0d08.jpg) +Point Cloud + +![](images/6fbd86428f71a0bb34023107057bbe11ba5bb9cf50283e35bad7f3994d267e26.jpg) + +![](images/ff3d60f55a48ddfd65e12485e87233d2e9123539a4b332ca9850e9cec3cd6bfb.jpg) + +![](images/d0ce17c9f8872f60798012105e1962fb4b202bf7b14badd936987b3694ea2009.jpg) +DynamicStereo + +![](images/f70935bd9499c1994e96b4ca62cd1769600fad45083657ae2d7b9367a299735f.jpg) + +![](images/7ace61094145e86490fcde360569048724eb0f76e67f2a905b63076d519074da.jpg) + +![](images/effb0f15aa8535a7a9f4eb7dca62643cacb325ace47f162d8f5d96364b658bf8.jpg) + +![](images/fbc1a3a078da0899f25f0ccf9759f9b048af29d51530011b7d221e2fcb1f8500.jpg) +BiDAStereo + +![](images/619395f338c6137cad9f0e295e9f51bcd9d2e2278fcca234655b801a618d4965.jpg) + +![](images/3b04867844d804312a68fe52dedb2d3a12364c35f9b7e51170ff6381a383e22d.jpg) + +![](images/8e75d31ec343f7b3aad94256509e56a684d3e991cf9871b6616afd8f50478e93.jpg) + +![](images/d89ba3d7ac560c0ce86eb1d1f844bbbdeadc5dcd43b7e4e7f54e94d3b6c03a12.jpg) + +![](images/258905af26531a4384573582b2453fb22b0b051832302d08403475c7f09e88ed.jpg) + +![](images/1a44549eeef01ed5e42cacff907c027ff01a812aa025558e70f6744f3c699a46.jpg) +Stereo Any Video (ours) + +![](images/11369eeeffbd1795bfc5875220b437ec0819aaabf620332545a53e83dc2a4418.jpg) +Figure 5. Qualitative comparison on a dynamic indoor scenario from the South Kensington SV dataset [23]. Top: Input left frames and predicted disparities for the same region across three frames using different methods. Bottom: Disparities converted into globally aligned point clouds and rendered from a camera view for comparison. Failure regions are highlighted with red arrows. Our method produces consistent and accurate disparities without flickering. See video in the supplementary material for better visualization. + +![](images/ef3895ca7083cb8da6bc7e3838c797de5cfa34b5dd9a4cd8a21d392ae89729cf.jpg) + +![](images/029f957fb7a5dadd02fd8ae654dd365d50eae7dfaf63d1e680dd59b1b8db61c5.jpg) + +![](images/3688cad9176a35789c6b6570dc82f8ce9d01f7ec5b65d8d08804543ee24b2af2.jpg) +Left Frames + +![](images/cb145eceb4f686bffdba1926ab0a694e990279c00214f4652de1eff0b22d148b.jpg) + +![](images/bd8cc6cd940f7df338036be6fc2581f5e137bc774e97fb6234b406b2fd68c5a7.jpg) + +![](images/2ad3a7daffc029ae491509083022a2d3c9b8d7150ccc7a37f4362ebb0ffd15a0.jpg) + +![](images/383f8ace1592521984b26dd8727dda2b7ce96d76fe7c207413f17bff1ba6d308.jpg) +Selective-IGEV + +![](images/a259d02c94925c90bc21ef9738c90e24db2646d10b30c7073955d9273c23d6fa.jpg) + +![](images/c698e9275cb9d9c2920f0cab98334d62bb980da10afb475a9d7e14853b9076ec.jpg) + +![](images/20bbc549b77dfdfcad5b3163cf839b71ac0c3f61f16fb175677d2d447d408a1e.jpg) + +![](images/b7ed7704e7c68d377d1939cd8ca964db0189d839b7ee4e118b2591a3f77d90a4.jpg) +DynamicStereo + +![](images/9e6dc2b7bb5e82e1ed53812e7ee25e8fb09b95a60e3d51be84480b6a85325765.jpg) + +![](images/65b7044e8cad1cc88607d5bea97f09c46dd48d6da791b5b6897975e09f04df5a.jpg) + +![](images/ff651e9d0149255bdee8f1895e3b864f5fd2b325a04b6960320a8e872cc36265.jpg) + +![](images/b199b1546b977a937308f01d412b3ca5b9ade1941e5e15d426267fadcbf3b375.jpg) +BiDAStereo +Figure 6. Qualitative comparison on a dynamic outdoor scenario from the South Kensington SV dataset [23]. + +![](images/69cffa071f143597de1bea78beb975b0d645443dd3abb8b5c8d82eb8cf2b112d.jpg) + +![](images/bc471e13c904b3476c0b61295341317f454b3a536c252d6b82516dbf7929506f.jpg) + +![](images/11259b7856e788a1da1da1e4ededf85c0dad292c9b26232cf19da5a0a3392e13.jpg) + +![](images/66567c4e513ab1c9adf73b6221151e9b3a6b8352ac670f4f69a23c99e1632cb9.jpg) +Stereo Any Video (ours) + +terestingly, increasing the backbone size does not yield additional performance gains. To investigate this, we conducted further experiments (Table 5) by varying depth and CNN feature channels based on VDA-L. However, the results remained similar. According to [7], VDA-L has $13 \times$ more parameters than VDA-S (381.8M vs. 28.4M), yet the performance gain is marginal ( $\delta_{1}(\uparrow)$ : 0.944 vs. 0.942 on KITTI; 0.971 vs. 0.959 on NYUv2). This suggests that the larger backbone does not provide richer priors. Moreover, large backbone requires additional CNN channels and parameters for adaptation, increasing training complexity. Therefore, we adopt VDA-S in the final model. + +Correlation. To evaluate the effectiveness of our proposed all-to-all-pairs correlation, we replace it with local correlation [27] and experiment with using only 1D or 2D correlation instead of alternating between them. As shown in Table 4, all these modifications degrade performance. Notably, constructing 3D correlations—where the search window spans height, width, and time—further worsens results. This suggests that frames cannot be effectively utilized without proper alignment [23]. + +Cost Volume Encoder. Instead of directly encoding cost volumes with convolutions, we find that introducing an MLP layer to extract a compact representation before ap + +Table 4. Ablation study on the foundation priors, correlation, cost volume encoder, convex upsampling, attention, and inference settings of our model on Sintel [2] and Dynamic Replica [24]. The approach used in our final model is underlined. + +
Setting VariationsSintel FinalDynamic Replica
TEPE↓δt1px ↓TEPE↓δt1px ↓
Foundation Priors
w/o Depth1.4712.350.0921.05
Depth AnythingV2-S [68]1.3311.270.0880.99
VDA-S [7]1.2710.980.0830.78
VDA-L [7]1.2012.470.0830.82
Correlation
1D + 2D local [27]1.2710.980.0830.78
1D all-to-all-pairs1.2311.360.0790.78
2D all-to-all-pairs1.2911.950.0850.77
3D all-to-all-pairs4.5122.901.3132.01
1D + 2D all-to-all-pairs1.2111.330.0760.75
Cost Volume Encoder
Convolution1.2611.540.0810.77
MLP + Convolution1.2111.330.0760.75
Convex Upsampling
Bilinear1.2612.740.0851.05
Convex [51]1.2111.330.0760.75
Temporal Convex1.0410.210.0670.79
Attention in GRU
w/o Attention1.0410.210.0670.79
Temporal1.0710.230.0640.79
Temporal + Spatial1.079.980.0570.67
Inference Stage
1/16 → 1/8 → 1/41.079.980.0570.67
1/32 → 1/16 → 1/8 → 1/41.0810.820.0580.70
+ +plying convolutions enhances cost aggregation, leading to improved final results. + +Convex Upsampling. We compare bilinear upsampling and general convex upsampling [51] to our proposed temporal convex upsampling module. As observed, our approach achieves better temporal consistency by incorporating learnable weights with adjacent frames. + +Attention in GRU. We further investigate different attention mechanisms within the GRU update block. As shown in Table 4, combining temporal and spatial attention together leads to the best performance. + +Inference Stage. We also explore inference starting at different resolutions. As can be seen in Table 4, a three-stage inference setup already achieves strong results. + +# 4.5. Application + +Our method enhances downstream applications that rely on accurate metric depth, such as 3D point tracking. As shown in Figure 7, we visualize results using the state-of-the-art 3D point tracking method DELTA [36]. Our predicted metric depth yields more points with precise trajectories than UniDepth [40], which gives less missing points as highlighted by the red arrows. Additional application, such as visual effects, are available in the supplementary material. + +Table 5. Ablation study on the effect of different channel configurations for depth features and CNN features, based on VDA-L [7], on Sintel [2]. Trainable parameters are also reported. + +
Depth & CNNSintel FinalPara. (M)
Channel NumberTEPE↓δt1px ↓
VDA-L [7]
(32, 96)1.2012.478.4
(128, 64)1.5814.5213.3
(128, 128)1.3113.2119.3
UniDepthStereo Any Video (ours)
+ +Figure 7. Visualization of 3D point tracking based on metric depth generated from UniDepth and our method. +Table 6. Comparison of different methods on model trainable [non-trainable] parameters, inference GPU memory, and multiply-accumulate (MAC) on 20 frames with resolution $720 \times 1280$ . + +
MethodsPara. (M)Mem. (G)MACs (T)
DynamicStereo [24]20.535.1182.3
BiDASTereo [22]12.241.1186.6
Stereo Any Video (ours)9.4 [28.4]41.1303.4
+ +# 4.6. Limitation and Future Work + +We report the inference efficiency results on a single NVIDIA A6000 GPU in Table 6. As shown, due to the use of foundation priors, our method incurs relatively high computational and memory costs. Looking ahead, we aim to develop a model zoo for this method, including both a large and a light version. + +# 5. Conclusion + +In this paper, we introduce Stereo Any Video, a robust framework for video stereo matching. By leveraging rich monocular priors, an all-to-all-pairs correlation mechanism, and a temporal convex upsampling layer, our approach produces temporally consistent and accurate disparities without relying on additional inputs such as camera pose or optical flow. Extensive evaluations show that our model achieves state-of-the-art performance on multiple standard benchmarks under zero-shot settings. Furthermore, our method generalizes well to real-world stereo videos, handling diverse content, motion, and camera movements effectively. Acknowledgments. This work was funded by the IC-CSC. + +# References + +[1] Luca Bartolomei, Fabio Tosi, Matteo Poggi, and Stefano Mattoccia. Stereo anywhere: Robust zero-shot deep stereo matching even where either stereo or mono fail. arXiv preprint arXiv:2412.04472, 2024.3 +[2] D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black. A naturalistic open source movie for optical flow evaluation. In ECCV, pages 611-625, 2012. 4, 5, 6, 8 +[3] Yohann Cabon, Naila Murray, and Martin Humenberger. Virtual kitti 2, 2020. 5, 6 +[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3 +[5] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In CVPR, pages 5410-5418, 2018. 2 +[6] Tianyu Chang, Xun Yang, Tianzhu Zhang, and Meng Wang. Domain generalized stereo matching via hierarchical visual transformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9559-9568, 2023. 2 +[7] Sili Chen, Hengkai Guo, Shengnan Zhu, Feihu Zhang, Zi long Huang, Jiashi Feng, and Bingyi Kang. Video depth anything: Consistent depth estimation for super-long videos, 2025. 2, 3, 5, 6, 7, 8 +[8] Junda Cheng, Longliang Liu, Gangwei Xu, Xianqi Wang, Zhaoxing Zhang, Yong Deng, Jinliang Zang, Yurui Chen, Zhipeng Cai, and Xin Yang. Monster: Marry monodepth to stereo unleashes power, 2025. 3, 5, 6 +[9] Xuelian Cheng, Yiran Zhong, Mehrtash Harandi, Yuchao Dai, Xiaojun Chang, Tom Drummond, Hongdong Li, and Zongyuan Ge. Hierarchical neural architecture search for deep stereo matching. arXiv preprint arXiv:2010.13501, 2020.2 +[10] Ziang Cheng, Jiayu Yang, and Hongdong Li. Stereo matching in time: $100+$ fps video stereo matching for extended reality. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 8719-8728, 2024. 2 +[11] Seokju Cho, Jiahui Huang, Jisu Nam, Honggyu An, Seun-gryong Kim, and Joon-Young Lee. Local all-pair correspondence for point tracking. arXiv preprint arXiv:2407.15420, 2024. 4 +[12] WeiQin Chuah, Ruwan Tennakoon, Reza Hoseinnezhad, Alireza Bab-Hadiashar, and David Suter. Itsa: An information-theoretic approach to automatic shortcut avoidance and domain generalization in stereo matching networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13022-13032, 2022. 2 +[13] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, pages 3354-3361, 2012. 1, 2, 5, 6 +[14] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2495-2504, 2020. 2 +[15] Weiyu Guo, Zhaoshuo Li, Yongkui Yang, Zheng Wang, Russell H Taylor, Mathias Unberath, Alan Yuille, and Yingwei Li. Context-enhanced stereo transformer. In European Conference on Computer Vision, pages 263-279. Springer, 2022. 2 +[16] Xiaoyang Guo, Kai Yang, Wukui Yang, Xiaogang Wang, and Hongsheng Li. Group-wise correlation stereo network. In CVPR, pages 3273-3282, 2019. 2 +[17] Xianda Guo, Chenming Zhang, Youmin Zhang, Wenzhao Zheng, Dujun Nie, Matteo Poggi, and Long Chen. Light-stereo: Channel boost is all your need for efficient 2d cost aggregation, 2024. 2 +[18] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 2 +[19] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. arXiv preprint arXiv:2409.02095, 2024. 3, 4 +[20] Hualie Jiang, Zhiqiang Lou, Laiyan Ding, Rui Xu, Minglang Tan, Wenjie Jiang, and Rui Huang. Defom-stereo: Depth foundation model based stereo matching, 2025. 3, 5, 6 +[21] Junpeng Jing, Jiankun Li, Pengfei Xiong, Jiangyu Liu, Shuaicheng Liu, Yichen Guo, Xin Deng, Mai Xu, Lai Jiang, and Leonid Sigal. Uncertainty guided adaptive warping for robust and efficient stereo matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3318-3327, 2023. 2 +[22] Junpeng Jing, Ye Mao, and Krystian Mikolajczyk. Match-stereo-videos: Bidirectional alignment for consistent dynamic stereo matching. In European Conference on Computer Vision, pages 415–432. Springer, 2024. 2, 4, 5, 6, 8 +[23] Junpeng Jing, Ye Mao, Anlan Qiu, and Krystian Mikolajczyk. Match stereo videos via bidirectional alignment, 2024. 3, 5, 6, 7 +[24] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Dynamic stereo: Consistent dynamic depth from stereo videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13229-13239, 2023. 2, 4, 5, 6, 8 +[25] Alex Kendall, Hayk Martirosyan, Saumitro Dasgupta, Peter Henry, Ryan Kennedy, Abraham Bachrach, and Adam Bry. End-to-end learning of geometry and context for deep stereo regression. In CVPR, pages 66-75, 2017. 2 +[26] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 3 +[27] Jiankun Li, Peisen Wang, Pengfei Xiong, Tao Cai, Ziwei Yan, Lei Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. Practical stereo matching via cascaded recurrent network with adaptive correlation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16263-16272, 2022. 1, 2, 4, 7, 8 + +[28] Zhaoshuo Li, Xingtong Liu, Nathan Drenkow, Andy Ding, Francis X Creighton, Russell H Taylor, and Mathias Unberath. Revisiting stereo depth estimation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6197-6206, 2021. 2 +[29] Zhaoshuo Li, Wei Ye, Dilin Wang, Francis X Creighton, Russell H Taylor, Ganesh Venkatesh, and Mathias Unberath. Temporally consistent online depth estimation in dynamic scenes. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 3018-3027, 2023. 2 +[30] Zhengfa Liang, Yiliu Feng, Yulan Guo, Hengzhu Liu, Wei Chen, Linbo Qiao, Li Zhou, and Jianfeng Zhang. Learning for disparity estimation through feature constancy. In CVPR, pages 2811-2820, 2018. 2 +[31] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. arXiv preprint arXiv:2109.07547, 2021. 1, 2, 3, 4, 5, 6 +[32] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60:91-110, 2004. 4 +[33] Yang Luo, Xuanlei Zhao, Mengzhao Chen, Kaipeng Zhang, Wenqi Shao, Kai Wang, Zhangyang Wang, and Yang You. Enhance-a-video: Better generated video for free, 2024. 2, 3 +[34] Nikolaus Mayer, Eddy Ig, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In CVPR, pages 4040-4048, 2016. 2, 5, 6 +[35] Lukas Mehl, Jenny Schmalfuss, Azin Jahedi, Yaroslava Nalivayko, and Andres Bruhn. Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 5, 6 +[36] Tuan Duc Ngo, Peiye Zhuang, Chuang Gan, Evangelos Kalogerakis, Sergey Tulyakov, Hsin-Ying Lee, and Chaoyang Wang. Delta: Dense efficient long-range 3d tracking for any video, 2025. 8 +[37] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 3 +[38] Jiahao Pang, Wenxiu Sun, Jimmy SJ Ren, Chengxi Yang, and Qiong Yan. Cascade residual learning: A two-stage convolutional neural network for stereo matching. In CVPRW, pages 887-895, 2017. 2 +[39] Jiahao Pang, Wenxiu Sun, Chengxi Yang, Jimmy Ren, Ruichao Xiao, Jin Zeng, and Liang Lin. Zoom and learn: Generalizing deep stereo matching to novel domains. In CVPR, pages 2070-2079, 2018. 2 +[40] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Unidepth: Universal monocular metric depth estimation, 2024. 8 +[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, + +Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3 +[42] Zhibo Rao, Bangshu Xiong, Mingyi He, Yuchao Dai, Renjie He, Zhelun Shen, and Xing Li. Masked representation learning for domain generalized stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5435-5444, 2023. 2 +[43] Ignacio Rocco, Mircea Cimpoi, Relja Arandjelović, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Neighbourhood consensus networks. Advances in neural information processing systems, 31, 2018. 4 +[44] Daniel Scharstein and Richard Szeliski. A taxonomy and evaluation of dense two-frame stereo correspondence algorithms. IJCV, 47(1):7-42, 2002. 1, 2, 5, 6 +[45] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In CVPR, pages 3260-3269, 2017. 1, 2, 5 +[46] Zhelun Shen, Yuchao Dai, and Zhibo Rao. Cfnet: Cascade and fused cost volume for robust stereo matching. In CVPR, pages 13906-13915, 2021. 2 +[47] Zhelun Shen, Xibin Song, Yuchao Dai, Dingfu Zhou, Zhibo Rao, and Liangjun Zhang. Digging into uncertainty-based pseudo-label for robust stereo matching. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-18, 2023. 2 +[48] Chenyang Si, Weichen Fan, Zhengyao Lv, Ziqi Huang, Yu Qiao, and Ziwei Liu. Repvideo: Rethinking cross-layer representation for video generation, 2025. 2, 3 +[49] Qing Su and Shihao Ji. Chitransformer: Towards reliable stereo from cues. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1939-1949, 2022. 2 +[50] Vladimir Tankovich, Christian Hane, Yinda Zhang, Adarsh Kowdle, Sean Fanello, and Sofien Bouaziz. Hitnet: Hierarchical iterative tile refinement network for real-time stereo matching. In CVPR, pages 14362-14372, 2021. 2 +[51] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, pages 402-419, 2020. 2, 5, 8 +[52] Prune Truong, Martin Danelljan, Luc V Gool, and Radu Timofte. Gocor: Bringing globally optimized correspondence volumes into your neural network. Advances in Neural Information Processing Systems, 33:14278-14290, 2020. 4 +[53] Prune Truong, Martin Danelljan, and Radu Timofte. Glunet: Global-local universal network for dense flow and correspondences. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6258-6268, 2020. 4 +[54] Jonas Uhrig, Nick Schneider, Lukas Schneider, Uwe Franke, Thomas Brox, and Andreas Geiger. Sparsity invariant cnns. In 2017 international conference on 3D Vision (3DV), pages 11-20. IEEE, 2017. 5, 6 +[55] Xianqi Wang, Gangwei Xu, Hao Jia, and Xin Yang. Selective-stereo: Adaptive frequency information selection + +for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19701-19710, 2024. 1, 2, 5, 6 +[56] Yan Wang, Zihang Lai, Gao Huang, Brian H Wang, Laurens Van Der Maaten, Mark Campbell, and Kilian Q Weinberger. Anytime stereo image depth estimation on mobile devices. In 2019 international conference on robotics and automation (ICRA), pages 5893-5900. IEEE, 2019. 2 +[57] Yiran Wang, Min Shi, Jiaqi Li, Zihao Huang, Zhiguo Cao, Jianming Zhang, Ke Xian, and Guosheng Lin. Neural video depth stabilizer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9466-9476, 2023. 5 +[58] Philippe Weinzaepfel, Thomas Lucas, Vincent Leroy, Yohann Cabon, Vaibhav Arora, Romain Brégier, Gabriela Csurka, Leonid Antsfeld, Boris Chidlovskii, and Jérôme Revaud. Croco v2: Improved cross-view completion pretraining for stereo matching and optical flow. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17969-17980, 2023. 2 +[59] Bowen Wen, Matthew Trepte, Joseph Aribido, Jan Kautz, Orazio Gallo, and Stan Birchfield. Foundationstereo: Zeroshot stereo matching, 2025. 3, 4, 6 +[60] Bin Xu, Yuhua Xu, Xiaoli Yang, Wei Jia, and Yulan Guo. Bilateral grid learning for stereo matching networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12497-12506, 2021. 2 +[61] Gangwei Xu, Xianqi Wang, Xiaohuan Ding, and Xin Yang. Iterative geometry encoding volume for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21919-21928, 2023. 1, 2, 4, 5, 6 +[62] Gangwei Xu, Yun Wang, Junda Cheng, Jinhui Tang, and Xin Yang. Accurate and efficient stereo matching via attention concatenation volume. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2 +[63] Gangwei Xu, Huan Zhou, and Xin Yang. Cgi-stereo: Accurate and real-time stereo matching via context and geometry interaction. arXiv preprint arXiv:2301.02789, 2023. 2 +[64] Gangwei Xu, Xianqi Wang, Zhaoxing Zhang, Junda Cheng, Chunyuan Liao, and Xin Yang. Igev++: Iterative multirange geometry encoding volumes for stereo matching. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 2 +[65] Haofei Xu and Juyong Zhang. Aanet: Adaptive aggregation network for efficient stereo matching. In CVPR, pages 1959-1968, 2020. 2 +[66] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, Fisher Yu, Dacheng Tao, and Andreas Geiger. Unifying flow, stereo and depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2 +[67] Gengshan Yang, Joshua Manela, Michael Happold, and Deva Ramanan. Hierarchical deep stereo matching on high-resolution images. In CVPR, pages 5515-5524, 2019. 2 +[68] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 3, 6, 8 +[69] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 3 +[70] Jiaxi Zeng, Chengtang Yao, Yuwei Wu, and Yunde Jia. Temporally consistent stereo matching. In European Conference on Computer Vision, pages 341-359. Springer, 2024. 2 +[71] Feihu Zhang, Victor Prisacariu, Ruigang Yang, and Philip HS Torr. Ga-net: Guided aggregation net for end-to-end stereo matching. In CVPR, pages 185-194, 2019. 2 +[72] Jiawei Zhang, Xiang Wang, Xiao Bai, Chen Wang, Lei Huang, Yimin Chen, Lin Gu, Jun Zhou, Tatsuya Harada, and Edwin R Hancock. Revisiting domain generalized stereo matching networks from a feature consistency perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13001-13011, 2022. 2 +[73] Youmin Zhang, Matteo Poggi, and Stefano Mattoccia. Temporalstereo: Efficient spatial-temporal stereo matching network. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 9528-9535. IEEE, 2023. 2 +[74] Yongjian Zhang, Longguang Wang, Kunhong Li, Yun Wang, and Yulan Guo. Learning representations from foundation models for domain generalized stereo matching. In European Conference on Computer Vision, pages 146-162. Springer, 2024. 3 +[75] Yiran Zhong, Hongdong Li, and Yuchao Dai. Open-world stereo video matching with deep rnN. In Proceedings of the European Conference on Computer Vision (ECCV), pages 101-116, 2018. 2 +[76] Jingyi Zhou, Haoyu Zhang, Jiakang Yuan, Peng Ye, Tao Chen, Hao Jiang, Meiya Chen, and Yangyang Zhang. All-in-one: Transferring vision foundation models into stereo matching. arXiv preprint arXiv:2412.09912, 2024. 3 \ No newline at end of file diff --git a/2025/Stereo Any Video_ Temporally Consistent Stereo Matching/images.zip b/2025/Stereo Any Video_ Temporally Consistent Stereo Matching/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8284a0cd376a4c2647015d6d3517f8f0a34b1082 --- /dev/null +++ b/2025/Stereo Any Video_ Temporally Consistent Stereo Matching/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ebfc1628457b3f4968043ecf75ee6c119b85d5ab2fd591cac9b37fac1a4e04f0 +size 828809 diff --git a/2025/Stereo Any Video_ Temporally Consistent Stereo Matching/layout.json b/2025/Stereo Any Video_ Temporally Consistent Stereo Matching/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a3ca9fff3b66ff9f75b0ec13faee97a9d7a19cf8 --- /dev/null +++ b/2025/Stereo Any Video_ Temporally Consistent Stereo Matching/layout.json @@ -0,0 +1,11059 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 124, + 103, + 487, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 103, + 487, + 121 + ], + "spans": [ + { + "bbox": [ + 124, + 103, + 487, + 121 + ], + "type": "text", + "content": "Stereo Any Video: Temporally Consistent Stereo Matching" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 147, + 142, + 462, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 142, + 462, + 172 + ], + "spans": [ + { + "bbox": [ + 147, + 142, + 462, + 172 + ], + "type": "text", + "content": "Junpeng Jing Weixun Luo Ye Mao Krystian Mikolajczyk Imperial College London" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 181, + 174, + 426, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 174, + 426, + 185 + ], + "spans": [ + { + "bbox": [ + 181, + 174, + 426, + 185 + ], + "type": "text", + "content": "https://tomtomommi.github.io/StereoAnyVideo/" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 70, + 213, + 544, + 383 + ], + "blocks": [ + { + "bbox": [ + 70, + 213, + 544, + 383 + ], + "lines": [ + { + "bbox": [ + 70, + 213, + 544, + 383 + ], + "spans": [ + { + "bbox": [ + 70, + 213, + 544, + 383 + ], + "type": "image", + "image_path": "7c752cb5b818cb77259978201f44bf7ee5c147a62dd04656f4e4508e93bbec34.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 78, + 386, + 151, + 432 + ], + "blocks": [ + { + "bbox": [ + 78, + 386, + 151, + 432 + ], + "lines": [ + { + "bbox": [ + 78, + 386, + 151, + 432 + ], + "spans": [ + { + "bbox": [ + 78, + 386, + 151, + 432 + ], + "type": "image", + "image_path": "fc00e09b027f94ce0785721c9e8dba417e39d76e95c153da6a3293a6ad84455b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 458, + 555, + 481 + ], + "lines": [ + { + "bbox": [ + 55, + 458, + 555, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 458, + 555, + 481 + ], + "type": "text", + "content": "Figure 1. Prediction on a dynamic real-world stereo video. Our method produces temporally consistent and accurate disparities without flickering, enabling the reconstruction of dense and clean metric point clouds for dynamic scenes. See video for better visualization." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 155, + 387, + 227, + 431 + ], + "blocks": [ + { + "bbox": [ + 155, + 387, + 227, + 431 + ], + "lines": [ + { + "bbox": [ + 155, + 387, + 227, + 431 + ], + "spans": [ + { + "bbox": [ + 155, + 387, + 227, + 431 + ], + "type": "image", + "image_path": "d6c877e62b44edb971474df51bd73253c50018dc3357f739b3e5343b25433b47.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 231, + 387, + 303, + 431 + ], + "blocks": [ + { + "bbox": [ + 231, + 387, + 303, + 431 + ], + "lines": [ + { + "bbox": [ + 231, + 387, + 303, + 431 + ], + "spans": [ + { + "bbox": [ + 231, + 387, + 303, + 431 + ], + "type": "image", + "image_path": "62572c3e1fc06cfb7337db396d22c885af6e05b88bb59129c7619448c56eb876.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 297, + 440, + 316, + 449 + ], + "lines": [ + { + "bbox": [ + 297, + 440, + 316, + 449 + ], + "spans": [ + { + "bbox": [ + 297, + 440, + 316, + 449 + ], + "type": "text", + "content": "time" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 387, + 378, + 431 + ], + "blocks": [ + { + "bbox": [ + 307, + 387, + 378, + 431 + ], + "lines": [ + { + "bbox": [ + 307, + 387, + 378, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 387, + 378, + 431 + ], + "type": "image", + "image_path": "b6ac54c158a8ea5b5f267eeb44f06556a7c205fe05ada34f25755f7fbfddd87b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 383, + 387, + 455, + 430 + ], + "blocks": [ + { + "bbox": [ + 383, + 387, + 455, + 430 + ], + "lines": [ + { + "bbox": [ + 383, + 387, + 455, + 430 + ], + "spans": [ + { + "bbox": [ + 383, + 387, + 455, + 430 + ], + "type": "image", + "image_path": "2d46a1de5fe1d22655eae1abe29a07699a5e6f98a8bec2ad366c5fa791801589.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 459, + 387, + 531, + 430 + ], + "blocks": [ + { + "bbox": [ + 459, + 387, + 531, + 430 + ], + "lines": [ + { + "bbox": [ + 459, + 387, + 531, + 430 + ], + "spans": [ + { + "bbox": [ + 459, + 387, + 531, + 430 + ], + "type": "image", + "image_path": "82143c2df08f4a6e61c6632aeeeb9f84e71f70de2044b12f6e9a92a38deab6ed.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 152, + 491, + 200, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 491, + 200, + 503 + ], + "spans": [ + { + "bbox": [ + 152, + 491, + 200, + 503 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 515, + 297, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 297, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 297, + 696 + ], + "type": "text", + "content": "This paper introduces Stereo Any Video, a powerful framework for video stereo matching. It can estimate spatially accurate and temporally consistent disparities without relying on auxiliary information such as camera poses or optical flow. The strong capability is driven by rich priors from monocular video depth models, which are integrated with convolutional features to produce stable representations. To further enhance performance, key architectural innovations are introduced: all-to-all-pairs correlation, which constructs smooth and robust matching cost volumes, and temporal convex upsampling, which improves temporal coherence. These components collectively enhance robustness, accuracy, and temporal consistency, establishing a new standard in video stereo matching. Extensive experiments demonstrate that our method achieves state-of-the" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 492, + 555, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 492, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 492, + 555, + 529 + ], + "type": "text", + "content": "art performance across multiple datasets both qualitatively and quantitatively in zero-shot settings, as well as strong generalization to real-world indoor and outdoor scenarios." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 562, + 395, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 562, + 395, + 574 + ], + "spans": [ + { + "bbox": [ + 314, + 562, + 395, + 574 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 582, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 677 + ], + "type": "text", + "content": "Video stereo matching is a task to reconstruct a 3D scene from a stereo camera. It estimates disparity by identifying the horizontal correspondences between rectified left and right images at each time step. Consistent and accurate disparity estimation is crucial for reconstructing and understanding real-world 3D scenes, benefiting various downstream tasks such as autonomous driving, robotic navigation, and virtual/augmented reality." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "With the advances in deep learning, stereo matching has seen significant progress [27, 31, 55, 61], achieving promising performance on public benchmarks [13, 44, 45]." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 258, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 258, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 258, + 713 + ], + "type": "text", + "content": "†Corresponding author: ye.mao21@imperial.ac.uk" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20836" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "type": "text", + "content": "However, most existing approaches are designed for image-based estimation and do not explicitly leverage temporal information in video sequences. When applied directly to video data, these methods often suffer from temporal inconsistencies, leading to noticeable flickering and artifacts in disparity maps and reconstructed point clouds. To mitigate this, recent video-based methods incorporate additional temporal cues, such as camera poses or optical flow, through predefined models or learnable mechanisms. For example, CODD [29] employs a pose network to estimate relative camera movements and align adjacent disparities within a fusion framework. Similarly, TemporalStereo [73] and TC-Stereo [70] follow this paradigm with variations in architectural design and alignment operations. Another approach, BiDAStereo [22], integrates optical flow [51] for temporal feature alignment and introduces a temporal aggregation module to propagate information across time." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 278, + 294, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 294, + 409 + ], + "type": "text", + "content": "Despite these advances, existing video methods remain vulnerable to dynamic scenes, camera movement, and the accuracy of auxiliary information. In real-world scenarios, dynamic objects undergo complex motions and deformations that violate traditional multi-view constraints [18], disrupting translational invariance in disparity estimation. Moreover, models relying on auxiliary cues are constrained by the performance of camera pose or optical flow estimation modules. In environments with complex camera movements, these estimation modules degrade, becoming bottlenecks that limit the robustness of video stereo methods." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 412, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 295, + 651 + ], + "type": "text", + "content": "To address these limitations, we propose Stereo Any Video, a robust framework for video-based stereo matching without relying on camera pose or optical flow. Inspired by recent findings in video generation [33, 48], which emphasize the role of robust feature representations in ensuring temporal coherence, we take a new perspective to design the framework for consistency: feature robustness and stability. Specifically, during feature extraction, we integrate convolutional cues with a frozen monocular video depth foundation model [7] trained on million-scale data. This ensures robust feature maps with rich priors. For correspondence matching, we introduce an all-to-all-pair correlation, which enforces smooth and reliable matching by jointly considering bidirectional correspondences from both target and reference views. For cost aggregation, we propose temporal convex upsampling within a fully temporal-based recurrent unit, playing a key role in maintaining temporal consistency. As shown in Figure 1, our method demonstrates a strong ability to produce accurate and consistent disparities, benefiting the reconstruction for dynamic scenes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 654, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 294, + 713 + ], + "type": "text", + "content": "The contributions of our work can be summarized as follows: (1) Stereo Any Video, a framework that integrates monocular video depth priors to achieve accurate and temporally consistent video stereo matching. (2) A novel all-to-all-pair correlation module and a temporal convex upsam" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "pling mechanism, ensuring robust and stable cost volume representations and aggregation. (3) State-of-the-art performance on diverse benchmarks, setting a new standard for video stereo matching." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 131, + 400, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 131, + 400, + 144 + ], + "spans": [ + { + "bbox": [ + 314, + 131, + 400, + 144 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 152, + 553, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 553, + 211 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 553, + 211 + ], + "type": "text", + "content": "In this section, we provide a brief literature review of stereo matching methods in the deep learning era. We first discuss image-based stereo matching methods, followed by video-based methods, and finally, we review recent advancements leveraging foundation priors." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 212, + 553, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 212, + 553, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 553, + 462 + ], + "type": "text", + "content": "Image-based Stereo Matching. Since the introduction of the first end-to-end stereo network [34], deep learning-based stereo matching has seen substantial advancements. Early methods typically construct 3D [16, 30, 38, 50, 65] or 4D [5, 9, 14, 25, 62, 67, 71] cost volumes from left and right image features, followed by 2D or 3D convolutions for cost aggregation. Inspired by RAFT [51], iterative refinement has become the mainstream approach. Representative advancements include multilevel recurrent units [31], adaptive correlation [27], geometry-encoding volumes [61, 64], and frequency information selection [55], all of which have significantly improved benchmark performance [13, 44, 45]. Transformer-based methods [15, 28, 49, 58, 66] have further refined stereo matching by modeling token correspondence and capturing long-range dependencies. Beyond accuracy, several methods focus on real-world challenges such as efficiency [17, 56, 60, 63], domain generalization [6, 12, 39, 42, 72], and robustness [21, 46, 47]. However, as all these methods are designed for static image pairs, they lack temporal awareness, leading to inconsistent disparity predictions in video sequences." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 463, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 553, + 713 + ], + "type": "text", + "content": "Video-based Stereo Matching. Video stereo matching especially in dynamic scenes presents unique challenges. Early work, such as [75], employed LSTMs with unsupervised learning for stereo videos but did not explicitly incorporate temporal information. CODD [29] introduced a per-frame stereo network, a motion network, and a fusion network designed for dynamic scenes. Neighboring-frame disparities are aligned and fused using estimated camera poses. TemporalStereo [73] extended this approach with a coarse-to-fine network that leverages past context to enhance predictions in challenging scenarios. Cheng et al. [10] developed a framework tailored for XR systems, reducing computational costs through temporal cost aggregation. TC-Stereo [70] further introduced a temporal disparity completion mechanism and a temporal state fusion module to exploit temporal information. These methods rely on auxiliary camera motion or pre-defined scene geometry to incorporate temporal information. DynamicStereo [24] introduced a transformer-based architecture that integrates temporal, stereo, and spatial attention. BiDAStereo [22] proposed an optical-flow-based alignment mechanism to propagate" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20837" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 71, + 529, + 232 + ], + "blocks": [ + { + "bbox": [ + 82, + 71, + 529, + 232 + ], + "lines": [ + { + "bbox": [ + 82, + 71, + 529, + 232 + ], + "spans": [ + { + "bbox": [ + 82, + 71, + 529, + 232 + ], + "type": "image", + "image_path": "b0642654f4f12eaa27541d472af4f24e2d80d5eb84e1e5d6257aae5ee4222608.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 241, + 555, + 298 + ], + "lines": [ + { + "bbox": [ + 54, + 241, + 555, + 298 + ], + "spans": [ + { + "bbox": [ + 54, + 241, + 555, + 298 + ], + "type": "text", + "content": "Figure 2. Overview of Stereo Any Video. Given a stereo video sequence as input, the model first extracts features and context information using trainable convolutional encoders and a frozen monocular video depth encoder (Video Depth Anything [7]). At each iteration, an all-to-all-pair correlation computes feature correlations, followed by a multilayer perception (MLP) encoder for compact representation. Disparities are iteratively refined using a 3D Gated Recurrent Unit (GRU), integrated with cost volume features, and upsampled via a temporal convex upsampling layer. This process is repeated within a cascaded pipeline to progressively recover full-resolution disparities." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 308, + 294, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 308, + 294, + 368 + ], + "spans": [ + { + "bbox": [ + 54, + 308, + 294, + 368 + ], + "type": "text", + "content": "temporal information across the entire sequence. Building on this alignment, BiDAS stabilizer [23] introduced a lightweight plug-and-play stabilizer network to improve the temporal consistency of existing image-based stereo matching methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 378, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 378, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 378, + 296, + 714 + ], + "type": "text", + "content": "Stereo Matching with Foundation Priors. Vision foundation models, such as CLIP [41], the DINO series [4, 37], and SAM [26], have been widely applied across various vision tasks due to their rich priors. In stereo matching area, FormerStereo [74] proposed feature transformation and C4 space to adapt ViT-based foundation model to stereo pipeline, learning domain invariant representations. AIOStereo [76] introduced a selective knowledge transfer module to distill knowledge from multiple heterogeneous vision foundation models into a single stereo matching model. More recent approaches primarily focused on leveraging monocular depth priors from Depth Anything [68, 69] to enhance stereo performance. For instance, Stereo Anywhere [1] directly incorporated relative depth maps into the stereo network, demonstrating robustness in challenging conditions such as non-Lambertian surfaces. DEFOM-Stereo [20] employed these depth maps alongside a novel scale updater within a recurrent module for improved alignment. FoundationStereo [59] extracted feature maps from Depth Anything and introduced a side-tuning adapter to adapt monocular features for stereo backbone. MonSter [8] utilized both depth maps and feature maps, and proposed a dualbranched refinement module to fuse stereo and monocular cues effectively. These methods highlight the effectiveness of foundation priors in image-based stereo tasks. In our work, we build upon this direction by investigating video-based stereo matching and exploring how foundation priors can enhance consistency in dynamic scenes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 307, + 370, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 307, + 370, + 319 + ], + "spans": [ + { + "bbox": [ + 314, + 307, + 370, + 319 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 327, + 555, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 448 + ], + "type": "text", + "content": "In this section, we introduce Stereo Any Video, a feedforward video-based framework. Given a rectified stereo sequence " + }, + { + "bbox": [ + 313, + 327, + 555, + 448 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_L^t,\\mathbf{I}_R^t\\}_{t\\in (1,T)}\\in \\mathbb{R}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 313, + 327, + 555, + 448 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 327, + 555, + 448 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 327, + 555, + 448 + ], + "type": "text", + "content": " denotes the total number of frames, " + }, + { + "bbox": [ + 313, + 327, + 555, + 448 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 327, + 555, + 448 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 327, + 555, + 448 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 327, + 555, + 448 + ], + "type": "text", + "content": " are the frame height and width, respectively, our goal is to estimate a temporally consistent sequence of disparity maps " + }, + { + "bbox": [ + 313, + 327, + 555, + 448 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{d}^t\\}_{t\\in (1,T)}\\in \\mathbb{R}^{H\\times W}" + }, + { + "bbox": [ + 313, + 327, + 555, + 448 + ], + "type": "text", + "content": " aligned with the left view. As shown in Figure 2, our method consists of three key components: feature extraction (Section 3.1), correlation (Section 3.2), and temporal cost aggregation (Section 3.3)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 456, + 538, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 538, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 538, + 467 + ], + "type": "text", + "content": "3.1. Feature Extraction with Foundation Priors" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 474, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 665 + ], + "type": "text", + "content": "Unlike image-based tasks, maintaining temporal consistency is the primary challenge in video stereo matching. Recent studies in video generation [33, 48] indicate that substantial variations across different frames can lead to unstable feature representations, thereby degrading temporal coherence. Therefore, ensuring feature stability across adjacent frames is crucial for achieving temporally consistent disparity estimation. Towards this, we leverage Video Depth Anything (VDA) [7] for feature extraction, which offers two key advantages: (1) trained on internet-scale data, VDA provides rich priors that can enhance disparity estimation in regions with ambiguous stereo correspondence; (2) unlike image-based monocular depth models [68, 69], VDA generates temporally stable features through an explicit temporal alignment process before the final prediction head, thereby preserving coherence across frames." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": "Besides, we investigate whether the depth maps from monocular video depth models can further improve stereo performance. We compare DepthCrafter [19], a representative monocular video model, with RAFTStereo [31], an" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20838" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 95, + 277, + 150 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 295, + 92 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 295, + 92 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 295, + 92 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of DepthCrafter [19] and RAFT-Stereo [31] on Sintel dataset [2]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 95, + 277, + 150 + ], + "lines": [ + { + "bbox": [ + 70, + 95, + 277, + 150 + ], + "spans": [ + { + "bbox": [ + 70, + 95, + 277, + 150 + ], + "type": "table", + "html": "
MethodSintel Clean
EPE ↓TEPE ↓MOS (human) ↓
DepthCrafter [19]8.682.080.212
RAFTStereo [31]1.420.840.933
", + "image_path": "fb3e94e26981cda50bb30177188bc5156279e51dedf9c1170b229c46f0162601.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 63, + 155, + 136, + 187 + ], + "blocks": [ + { + "bbox": [ + 63, + 155, + 136, + 187 + ], + "lines": [ + { + "bbox": [ + 63, + 155, + 136, + 187 + ], + "spans": [ + { + "bbox": [ + 63, + 155, + 136, + 187 + ], + "type": "image", + "image_path": "7a70cb3d576351db04a2ea923a546ed19c954cdfdcdb42c0361bf141f1382948.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 187, + 118, + 193 + ], + "lines": [ + { + "bbox": [ + 82, + 187, + 118, + 193 + ], + "spans": [ + { + "bbox": [ + 82, + 187, + 118, + 193 + ], + "type": "text", + "content": "Input bamboo 2" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 137, + 155, + 212, + 187 + ], + "blocks": [ + { + "bbox": [ + 137, + 155, + 212, + 187 + ], + "lines": [ + { + "bbox": [ + 137, + 155, + 212, + 187 + ], + "spans": [ + { + "bbox": [ + 137, + 155, + 212, + 187 + ], + "type": "image", + "image_path": "31f7cbdab9f83f7e21c42cf20500bbf76d372bb4821712726d60e0b428a43b00.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 212, + 155, + 288, + 187 + ], + "blocks": [ + { + "bbox": [ + 212, + 155, + 288, + 187 + ], + "lines": [ + { + "bbox": [ + 212, + 155, + 288, + 187 + ], + "spans": [ + { + "bbox": [ + 212, + 155, + 288, + 187 + ], + "type": "image", + "image_path": "f95bb896f959ab1811c60258011db38988420105d2b9d06816b585a0b4a397e2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 63, + 193, + 136, + 224 + ], + "blocks": [ + { + "bbox": [ + 63, + 193, + 136, + 224 + ], + "lines": [ + { + "bbox": [ + 63, + 193, + 136, + 224 + ], + "spans": [ + { + "bbox": [ + 63, + 193, + 136, + 224 + ], + "type": "image", + "image_path": "6356f26fb1ae72c2f4334278f83bb72f1e390dc07642cb5c000e74b9bf8e28b3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 225, + 118, + 232 + ], + "lines": [ + { + "bbox": [ + 82, + 225, + 118, + 232 + ], + "spans": [ + { + "bbox": [ + 82, + 225, + 118, + 232 + ], + "type": "text", + "content": "Input sleeping_1" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 236, + 295, + 258 + ], + "lines": [ + { + "bbox": [ + 55, + 236, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 295, + 258 + ], + "type": "text", + "content": "Figure 3. Qualitative comparison of DepthCrafter [19] and RAFT-Stereo [31] on Sintel dataset [2]." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 137, + 193, + 211, + 224 + ], + "blocks": [ + { + "bbox": [ + 157, + 188, + 192, + 193 + ], + "lines": [ + { + "bbox": [ + 157, + 188, + 192, + 193 + ], + "spans": [ + { + "bbox": [ + 157, + 188, + 192, + 193 + ], + "type": "text", + "content": "DepthCrafter" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 193, + 211, + 224 + ], + "lines": [ + { + "bbox": [ + 137, + 193, + 211, + 224 + ], + "spans": [ + { + "bbox": [ + 137, + 193, + 211, + 224 + ], + "type": "image", + "image_path": "0c98399bcd7f755bc365c3bbb13e7114752fdbd7b20b399494de856b3db596ea.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 158, + 225, + 192, + 232 + ], + "lines": [ + { + "bbox": [ + 158, + 225, + 192, + 232 + ], + "spans": [ + { + "bbox": [ + 158, + 225, + 192, + 232 + ], + "type": "text", + "content": "DepthCrafter" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 212, + 193, + 288, + 224 + ], + "blocks": [ + { + "bbox": [ + 235, + 188, + 267, + 193 + ], + "lines": [ + { + "bbox": [ + 235, + 188, + 267, + 193 + ], + "spans": [ + { + "bbox": [ + 235, + 188, + 267, + 193 + ], + "type": "text", + "content": "RAFTStereo" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 212, + 193, + 288, + 224 + ], + "lines": [ + { + "bbox": [ + 212, + 193, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 212, + 193, + 288, + 224 + ], + "type": "image", + "image_path": "0d8c45cfda0c6876d215645a2237c716f33d00dc0669ba0a0018f750fc3e0058.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 225, + 267, + 232 + ], + "lines": [ + { + "bbox": [ + 235, + 225, + 267, + 232 + ], + "spans": [ + { + "bbox": [ + 235, + 225, + 267, + 232 + ], + "type": "text", + "content": "RAFTStereo" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 54, + 268, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 268, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 54, + 268, + 295, + 460 + ], + "type": "text", + "content": "image-based stereo model. Using open-source weights, we evaluate their performance on Sintel [2], as shown in Table 1 and Figure 3. Specifically, we first obtain relative depth maps from DepthCrafter (without scale and shift) and compute disparities using RAFTStereo on the same sequence. The relative depth maps are then aligned to RAFTStereo's disparities, and objective metrics are computed against the ground truth. Additionally, we conduct a subjective user study (see supplementary material for details). As observed, while DepthCrafter produces visually plausible results—reflected in better human scores—its spatial and temporal errors are significantly larger than those of RAFTStereo. This suggests that its relative depth is \"consistent but wrong\", making it unreliable for directly improving consistency within a stereo framework. Therefore, we only utilize its features as valuable priors." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 461, + 295, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 295, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 295, + 570 + ], + "type": "text", + "content": "As shown in Figure 2, our feature extraction module consists of two trainable encoders, each employing a series of residual blocks [31] to extract convolutional image features and context features. To integrate VDA priors, we adopt a shallow convolutional adapter [59] that downsamples depth features from the frozen VDA. The extracted depth features are then concatenated with convolutional features to form " + }, + { + "bbox": [ + 55, + 461, + 295, + 570 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{F}_L,\\mathbf{F}_R\\}" + }, + { + "bbox": [ + 55, + 461, + 295, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 461, + 295, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_C" + }, + { + "bbox": [ + 55, + 461, + 295, + 570 + ], + "type": "text", + "content": ", providing more stable feature representations for subsequent processing." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 581, + 201, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 201, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 201, + 594 + ], + "type": "text", + "content": "3.2. All-to-all-pair Correlation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 600, + 295, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 295, + 684 + ], + "type": "text", + "content": "At the " + }, + { + "bbox": [ + 55, + 600, + 295, + 684 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 600, + 295, + 684 + ], + "type": "text", + "content": "-th iteration for each time step, the right feature " + }, + { + "bbox": [ + 55, + 600, + 295, + 684 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_R" + }, + { + "bbox": [ + 55, + 600, + 295, + 684 + ], + "type": "text", + "content": " is first warped toward the left feature " + }, + { + "bbox": [ + 55, + 600, + 295, + 684 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_L" + }, + { + "bbox": [ + 55, + 600, + 295, + 684 + ], + "type": "text", + "content": " using the previous disparity " + }, + { + "bbox": [ + 55, + 600, + 295, + 684 + ], + "type": "inline_equation", + "content": "d_{n-1}" + }, + { + "bbox": [ + 55, + 600, + 295, + 684 + ], + "type": "text", + "content": ", producing " + }, + { + "bbox": [ + 55, + 600, + 295, + 684 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_R" + }, + { + "bbox": [ + 55, + 600, + 295, + 684 + ], + "type": "text", + "content": ". Previous methods [31, 61] typically construct cost volumes " + }, + { + "bbox": [ + 55, + 600, + 295, + 684 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_n" + }, + { + "bbox": [ + 55, + 600, + 295, + 684 + ], + "type": "text", + "content": " via all-pair correlation, employing a lookup mechanism to search for potential matches within a local radius. In [27], this approach was extended to an alternating local 1D and 2D search:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 80, + 696, + 295, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 696, + 295, + 717 + ], + "spans": [ + { + "bbox": [ + 80, + 696, + 295, + 717 + ], + "type": "interline_equation", + "content": "\\mathbf {C} _ {n} (x, y) = \\left\\langle \\mathbf {F} _ {L} (x, y), \\hat {\\mathbf {F}} _ {R} (x + r _ {x}, y + r _ {y}) \\right\\rangle , \\tag {1}", + "image_path": "8be38bc2258ab40a02c70214cefd76a33ea0a8b346dfe8098d9fdab5f51ab561.jpg" + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 322, + 72, + 547, + 199 + ], + "blocks": [ + { + "bbox": [ + 322, + 72, + 547, + 199 + ], + "lines": [ + { + "bbox": [ + 322, + 72, + 547, + 199 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 547, + 199 + ], + "type": "image", + "image_path": "a6241fd41074fcb9c76091307049e3fb33864d44ed4e05e3c5eeacc11173b1ee.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 201, + 553, + 236 + ], + "lines": [ + { + "bbox": [ + 313, + 201, + 553, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 553, + 236 + ], + "type": "text", + "content": "Figure 4. Workflow of the temporal convex upsampling. Each pixel of the high resolution disparity (yellow box) is calculated via its 27 low resolution neighbors using learnable weights." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 243, + 555, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 388 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 243, + 555, + 388 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 313, + 243, + 555, + 388 + ], + "type": "text", + "content": " denotes the position in " + }, + { + "bbox": [ + 313, + 243, + 555, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_L" + }, + { + "bbox": [ + 313, + 243, + 555, + 388 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 243, + 555, + 388 + ], + "type": "inline_equation", + "content": "\\langle \\cdot, \\cdot \\rangle" + }, + { + "bbox": [ + 313, + 243, + 555, + 388 + ], + "type": "text", + "content": " represents the channel-wise product operation, and " + }, + { + "bbox": [ + 313, + 243, + 555, + 388 + ], + "type": "inline_equation", + "content": "(r_x, r_y)" + }, + { + "bbox": [ + 313, + 243, + 555, + 388 + ], + "type": "text", + "content": " defines the search range, typically set as " + }, + { + "bbox": [ + 313, + 243, + 555, + 388 + ], + "type": "inline_equation", + "content": "(r_x, r_y) \\in \\{(\\pm 4,0), (\\pm 1, \\pm 1)\\}" + }, + { + "bbox": [ + 313, + 243, + 555, + 388 + ], + "type": "text", + "content": ". The fundamental principle of these methods is to construct one-directional correspondence, where a reference point is selected and directly matched with target points via multiplication. Inspired by recent advances in point tracking [11], we introduce bidirectional correspondence and develop a local all-to-all-pair correlation. Instead of computing correlation in a fixed direction, we calculate the similarity between every pair of potential matching points within two search windows, formulated as:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 319, + 404, + 553, + 425 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 404, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 319, + 404, + 553, + 425 + ], + "type": "interline_equation", + "content": "\\mathbf {C} _ {n} (x, y) = \\left\\langle \\mathbf {F} _ {L} \\left(x + r _ {x}, y + r _ {y}\\right), \\hat {\\mathbf {F}} _ {R} \\left(x + r _ {x}, y + r _ {y}\\right) \\right\\rangle . \\tag {2}", + "image_path": "3587e040b6687e50db919c5b704243ea606f6048d69c2b31c4ba7446e015103a.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 431, + 554, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 554, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 554, + 528 + ], + "type": "text", + "content": "This approach enhances match verification and reduces ambiguity [32, 43]. While BiDAStereo [22] constructs temporal cost volumes by aligning adjacent frames for consistency, its performance is significantly constrained by the limitations of optical flow. In contrast, our method emphasizes robust spatial correlation, leveraging dense correspondences to enforce matching smoothness and enhance consistency and robustness across neighboring points [52, 53]." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 534, + 467, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 467, + 547 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 467, + 547 + ], + "type": "text", + "content": "3.3. Temporal Cost Aggregation" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 552, + 554, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 554, + 635 + ], + "type": "text", + "content": "At the " + }, + { + "bbox": [ + 313, + 552, + 554, + 635 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 552, + 554, + 635 + ], + "type": "text", + "content": "-th iteration, the all-to-all-pair correlation expands the potential matching points in " + }, + { + "bbox": [ + 313, + 552, + 554, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_n" + }, + { + "bbox": [ + 313, + 552, + 554, + 635 + ], + "type": "text", + "content": ", increasing its dimensionality from " + }, + { + "bbox": [ + 313, + 552, + 554, + 635 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{H' \\times W' \\times (2r_x + 1)(2r_y + 1)}" + }, + { + "bbox": [ + 313, + 552, + 554, + 635 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 552, + 554, + 635 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{H' \\times W' \\times (2r_x + 1)^2(2r_y + 1)^2}" + }, + { + "bbox": [ + 313, + 552, + 554, + 635 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 552, + 554, + 635 + ], + "type": "inline_equation", + "content": "H'" + }, + { + "bbox": [ + 313, + 552, + 554, + 635 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 552, + 554, + 635 + ], + "type": "inline_equation", + "content": "W'" + }, + { + "bbox": [ + 313, + 552, + 554, + 635 + ], + "type": "text", + "content": " denote the height and width of frame feature, respectively. We first encode " + }, + { + "bbox": [ + 313, + 552, + 554, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_n" + }, + { + "bbox": [ + 313, + 552, + 554, + 635 + ], + "type": "text", + "content": " into a compact representation " + }, + { + "bbox": [ + 313, + 552, + 554, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_n \\in \\mathbb{R}^{H' \\times W' \\times L}" + }, + { + "bbox": [ + 313, + 552, + 554, + 635 + ], + "type": "text", + "content": " using a MLP:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 397, + 636, + 553, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 636, + 553, + 649 + ], + "spans": [ + { + "bbox": [ + 397, + 636, + 553, + 649 + ], + "type": "interline_equation", + "content": "\\mathbf {E} _ {n} = \\operatorname {M L P} \\left(\\mathbf {C} _ {n}\\right). \\tag {3}", + "image_path": "ecc239e5ca37cf69f0fed888cd32288b3c301d6409cbc80b0f2d98f87c23bfbe.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": "Next, we employ a 3D-GRU [24] to iteratively refine the disparity for each frame while preserving temporal consistency. The GRU aggregates information using separable 3D convolutions across both spatial and temporal dimensions. At iteration " + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": ", it takes as input the encoded cost volume " + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_n" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20839" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 105, + 553, + 239 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": "Table 2. Zero-shot generalization results. All methods were trained only on SceneFlow [34]. The representative image-based and video-based stereo matching models are compared on Sintel [2], Dynamic Replica [24], Infinigen SV [23], and Virtual KITTI2 [3]. The weights and parameters are fixed for evaluation. The best and second-best results are highlighted." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 105, + 553, + 239 + ], + "lines": [ + { + "bbox": [ + 56, + 105, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 105, + 553, + 239 + ], + "type": "table", + "html": "
MethodSintel FinalDynamic ReplicaInfinigen SVVirtual KITTI2
TEPE↓δt1px↓EPE↓TEPE↓δt1px↓EPE↓TEPE↓δt1px↓EPE↓TEPE↓δt1px↓EPE↓
Image-based
RAFTStereo [31]2.0913.704.890.1452.020.382.3319.803.771.2518.391.45
IGEVStereo [61]1.7412.652.440.0810.800.202.3519.333.811.4218.881.39
Selective-IGEV [55]1.9612.952.820.1331.290.292.8521.835.691.4119.101.38
DEFOM-Stereo [20]2.3312.432.380.0800.750.132.0118.913.241.0716.631.34
MonSter [8]1.7011.063.300.1570.860.451.6515.732.270.7310.930.67
Video-based
DynamicStereo [24]1.4311.804.260.0830.730.241.9815.684.671.0615.631.47
BiDAStereo [22]1.2611.651.990.0911.040.651.9916.506.521.0214.821.43
Stereo Any Video (ours)1.079.981.750.0570.670.251.6514.273.160.7412.451.01
", + "image_path": "5b73c343c25623f264b13f3a0e4565677b0a17c7ee8573f8ec7dd731b2887524.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 247, + 296, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 296, + 297 + ], + "type": "text", + "content": "context feature " + }, + { + "bbox": [ + 55, + 247, + 296, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_c" + }, + { + "bbox": [ + 55, + 247, + 296, + 297 + ], + "type": "text", + "content": ", and the disparities from the previous iteration " + }, + { + "bbox": [ + 55, + 247, + 296, + 297 + ], + "type": "inline_equation", + "content": "d_{n-1}" + }, + { + "bbox": [ + 55, + 247, + 296, + 297 + ], + "type": "text", + "content": ". To enhance representation, we integrate super kernels [22], temporal and spatial attention [24] in this process. The disparity update in 3D-GRU is as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 302, + 294, + 317 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 302, + 294, + 317 + ], + "spans": [ + { + "bbox": [ + 70, + 302, + 294, + 317 + ], + "type": "interline_equation", + "content": "x _ {n} = \\left[ \\operatorname {C o n v 2 d} \\left(\\mathbf {E} _ {n}\\right), \\operatorname {C o n v 2 d} \\left(d _ {n - 1}\\right), d _ {n - 1}, \\mathbf {F} _ {c} \\right], \\tag {4}", + "image_path": "4a813b77da7e78fb11b3d7834731cf98fc34347da0ab1b7edaa679cf3af07a25.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 71, + 318, + 295, + 332 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 318, + 295, + 332 + ], + "spans": [ + { + "bbox": [ + 71, + 318, + 295, + 332 + ], + "type": "interline_equation", + "content": "\\widetilde {x _ {n}} = \\operatorname {A t t e n t i o n} \\left(x _ {n}\\right), \\tag {5}", + "image_path": "5e99a490ce11896a92074dd6cc1d3eeb3dbbb0d06ab7610d0b59e1d12735dd4a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 71, + 334, + 295, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 334, + 295, + 353 + ], + "spans": [ + { + "bbox": [ + 71, + 334, + 295, + 353 + ], + "type": "interline_equation", + "content": "z _ {n} = \\sigma \\left(\\operatorname {C o n v 3 d} _ {z} \\left(\\left[ h _ {n - 1}, \\widetilde {x _ {n}} \\right]\\right)\\right), \\tag {6}", + "image_path": "8ff4a0dda1f55df558c699f83076d4656ef8a06a3a9662393782e3f53a5b6147.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 71, + 356, + 295, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 356, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 71, + 356, + 295, + 376 + ], + "type": "interline_equation", + "content": "r _ {n} = \\sigma \\left(\\operatorname {C o n v 3 d} _ {r} \\left(\\left[ h _ {n - 1}, \\widetilde {x _ {n}} \\right]\\right)\\right), \\tag {7}", + "image_path": "7e145efdc06bd28d66b9fdfa1e641233811528ea063f0ca0c5ee49f2428f6a71.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 71, + 378, + 295, + 398 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 378, + 295, + 398 + ], + "spans": [ + { + "bbox": [ + 71, + 378, + 295, + 398 + ], + "type": "interline_equation", + "content": "q _ {n} = \\tanh \\left(\\operatorname {C o n v} 3 \\mathrm {d} _ {q} \\left(\\left[ h _ {n - 1} \\odot r _ {n}, \\widetilde {x _ {n}} \\right]\\right)\\right), \\tag {8}", + "image_path": "5fafcf0553f92613ee8b6f8933cf695bb36da6f40b535ee73d4910bf691c82a4.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 71, + 400, + 295, + 413 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 400, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 71, + 400, + 295, + 413 + ], + "type": "interline_equation", + "content": "h _ {n} = \\left(1 - z _ {n}\\right) \\odot h _ {n - 1} + z _ {n} \\odot q _ {n}, \\tag {9}", + "image_path": "9abacfbc0494a73d15ad46eada90affe01b7ebfa7504ead2b53bb3299f728a91.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 71, + 415, + 295, + 428 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 415, + 295, + 428 + ], + "spans": [ + { + "bbox": [ + 71, + 415, + 295, + 428 + ], + "type": "interline_equation", + "content": "d _ {n} = d _ {n - 1} + \\operatorname {C o n v} 3 \\mathrm {d} \\left(h _ {n}\\right), \\tag {10}", + "image_path": "36b2ca446ba9a148d2842f57fffbbaa121a4b2f096850a15cb7bb8b836c75029.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 435, + 295, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 295, + 459 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 435, + 295, + 459 + ], + "type": "inline_equation", + "content": "[\\cdot ]" + }, + { + "bbox": [ + 55, + 435, + 295, + 459 + ], + "type": "text", + "content": " is concatenation, " + }, + { + "bbox": [ + 55, + 435, + 295, + 459 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 435, + 295, + 459 + ], + "type": "text", + "content": " denotes element-wise multiplication, and " + }, + { + "bbox": [ + 55, + 435, + 295, + 459 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 435, + 295, + 459 + ], + "type": "text", + "content": " is the sigmoid activation function." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 459, + 295, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 295, + 591 + ], + "type": "text", + "content": "While prior methods emphasize temporal consistency in correlation and cost aggregation, the upsampling process remains overlooked, despite its crucial role in achieving accurate and temporally consistent disparities. Towards this, we adapt the commonly used convex upsampling approach [51] to a temporal variant. In Figure 4, each high-resolution disparity pixel (yellow box) is computed as a convex combination of its 27 low resolution neighbors across the 3 frames (green, blue, and faint yellow boxes). The network predicts the corresponding " + }, + { + "bbox": [ + 55, + 459, + 295, + 591 + ], + "type": "inline_equation", + "content": "3 \\times 3 \\times 3" + }, + { + "bbox": [ + 55, + 459, + 295, + 591 + ], + "type": "text", + "content": " weights using 3D convolutions, enabling a temporally consistent upsampling:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 71, + 597, + 295, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 597, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 71, + 597, + 295, + 617 + ], + "type": "interline_equation", + "content": "\\mathbf {w} = \\operatorname {s o f t m a x} \\left(\\operatorname {C o n v 3 d} \\left(h _ {n}\\right)\\right), \\tag {11}", + "image_path": "0e5e0cb53a70c072522d6458f06ceb1506e53d02e2019557f543d020f789400e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 71, + 620, + 295, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 620, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 71, + 620, + 295, + 646 + ], + "type": "interline_equation", + "content": "D _ {n} ^ {t} = \\alpha \\cdot \\sum_ {i j k} \\mathbf {w} _ {i j k} \\odot \\text {u n f o l d} \\left(\\left[ d _ {n} ^ {t - 1}, d _ {n} ^ {t}, d _ {n} ^ {t + 1} \\right]\\right), \\tag {12}", + "image_path": "fa55a764902f8d227cc1ae361d76aed1afa3601b3ebe614f426d5a108da18c23.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 653, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 653, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 653, + 295, + 666 + ], + "type": "text", + "content": " is upsampling rate, " + }, + { + "bbox": [ + 55, + 653, + 295, + 666 + ], + "type": "inline_equation", + "content": "D_{n}^{t}" + }, + { + "bbox": [ + 55, + 653, + 295, + 666 + ], + "type": "text", + "content": " is the upsampled disparity." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 672, + 145, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 145, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 145, + 684 + ], + "type": "text", + "content": "3.4. Loss Function" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Recent works have introduced OPW loss [57] and TGM loss [7] in monocular video depth estimation to enhance" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 247, + 555, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 555, + 308 + ], + "type": "text", + "content": "temporal consistency. However, we found out that these losses impose a trade-off between accuracy and consistency without yielding substantial improvement in the video stereo task. Thus, we opt for a purely image-based " + }, + { + "bbox": [ + 313, + 247, + 555, + 308 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 313, + 247, + 555, + 308 + ], + "type": "text", + "content": " loss while still achieving strong temporal consistency as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 365, + 316, + 555, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 316, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 365, + 316, + 555, + 350 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\sum_ {t = 1} ^ {T} \\sum_ {n = 1} ^ {N} \\gamma^ {N - n} \\left| \\left| D _ {\\mathrm {g t}} ^ {t} - D _ {n} ^ {t} \\right| \\right|, \\tag {13}", + "image_path": "bf780e8d313eb333dacd120b1eb81d7554aa228d7ec022ff0463f1aa54b8235b.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 357, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 357, + 555, + 407 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 357, + 555, + 407 + ], + "type": "text", + "content": " is the number of frames, " + }, + { + "bbox": [ + 313, + 357, + 555, + 407 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 357, + 555, + 407 + ], + "type": "text", + "content": " is the number of iterations, and " + }, + { + "bbox": [ + 313, + 357, + 555, + 407 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 357, + 555, + 407 + ], + "type": "text", + "content": " is a weighting coefficient set to 0.9. Here, " + }, + { + "bbox": [ + 313, + 357, + 555, + 407 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{gt}}^t" + }, + { + "bbox": [ + 313, + 357, + 555, + 407 + ], + "type": "text", + "content": " denotes the ground truth disparity for the " + }, + { + "bbox": [ + 313, + 357, + 555, + 407 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 357, + 555, + 407 + ], + "type": "text", + "content": "-th frame. The final output is obtained from the last iteration prediction." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 417, + 395, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 395, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 395, + 430 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 437, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 510 + ], + "type": "text", + "content": "In this section, we first introduce the datasets and metrics used in the evaluation, as well as implementation details. We then present the zero-shot generalization capabilities of our method, followed by ablation studies to validate the efficacy of each component in our framework. Please refer to the supplementary material for more details." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 517, + 489, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 489, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 489, + 529 + ], + "type": "text", + "content": "4.1. Benchmark Datasets and Metric" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "Dataset. Our work focuses on videos captured with moving cameras, making commonly used image benchmarks unsuitable, such as Middlebury [44], ETH3D [45], and KITTI 2012/2015 [13] (two frames per scene). To evaluate our method, we use five synthetic and two real-world stereo video datasets, all featuring dynamic scenes. Sintel [2] is a synthetic dataset comprising movie sequences available in both clean and final passes. Spring [35] is a high resolution synthetic dataset. Dynamic Replica [24] is a synthetic indoor dataset featuring non-rigid objects, such as animals and people. Infinigen SV [23] is a synthetic dataset focusing on outdoor natural environments. Virtual KITTI2 [3] is a synthetic dataset simulating outdoor driving scenarios. KITTI Depth [54] is a real-world outdoor dataset captured for autonomous driving applications, with sparse" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20840" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 118, + 555, + 260 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": "Table 3. Quantitative comparisons on Spring [35], Sintel [2], and KITTI Depth [54] where methods are allowed to train on any existing datasets. Model weights remain fixed during evaluation. Training datasets sharing the same domain as the test sets are highlighted. “-” represents the test set has been used in training. Abbreviations: SF - SceneFlow [34], DR - Dynamic Replica [24], ISV - Infigen SV [23], M - Middlebury [44], KD - KITTI Depth [54], VK - Virtual KITTI2 [3]. The best and second-best results are highlighted." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 118, + 555, + 260 + ], + "lines": [ + { + "bbox": [ + 56, + 118, + 555, + 260 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 555, + 260 + ], + "type": "table", + "html": "
MethodTraining DataScaleSpringSintel FinalKITTI Depth
TEPE↓δt1px↓EPE↓TEPE↓δt1px↓EPE↓TEPE↓δt1px↓EPE↓
Image-based
RAFTStereo [31]10 mixed datasets0.6M1.499.795.37---0.417.960.51
IGEVStereo [61]SF+M0.04M1.539.725.371.5313.492.230.469.120.60
Selective-IGEV [55]SF+M0.04M1.469.335.461.6413.782.140.489.660.60
DEFOM-Stereo [20]15 mixed datasets0.7M1.299.124.25---0.397.080.48
MonSter [8]10 mixed datasets0.6M1.258.655.19---0.417.730.60
FoundationStereo [59]7 mixed datasets1.3M1.789.825.40---0.407.200.57
Video-based
DynamicStereo [24]SF+DR+ISV+KD0.1M1.079.649.361.3811.843.380.428.210.57
BiDAStereo [22]SF+DR+ISV+KD0.1M0.907.903.921.3311.341.970.428.330.58
Stereo Any Video (ours)SF+DR+ISV+VK0.1M0.776.523.170.999.091.480.356.050.61
", + "image_path": "f21618afa7672f5b7d4be33241f98f3b1b2781b4fd65bead2e2c1df713c7fefc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 274, + 295, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 295, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 295, + 309 + ], + "type": "text", + "content": "depth maps obtained from a LiDAR sensor. South Kensington SV [23] is a real-world dataset capturing daily scenarios from a stereo camera for qualitative evaluation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 310, + 296, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 310, + 296, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 296, + 395 + ], + "type": "text", + "content": "Metric. For accuracy evaluation, we use End-Point Error (EPE), which measures the average per-pixel disparity error. To assess temporal consistency, we employ Temporal EPE (TEPE) [24], which quantifies the variation in EPE across the temporal dimension. We also compute " + }, + { + "bbox": [ + 55, + 310, + 296, + 395 + ], + "type": "inline_equation", + "content": "\\delta_{n\\text{-px}}^t" + }, + { + "bbox": [ + 55, + 310, + 296, + 395 + ], + "type": "text", + "content": ", which represents the proportion of pixels with TEPE exceeding " + }, + { + "bbox": [ + 55, + 310, + 296, + 395 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 310, + 296, + 395 + ], + "type": "text", + "content": ". For all metrics, lower values indicate better performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 403, + 189, + 416 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 189, + 416 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 189, + 416 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 422, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 601 + ], + "type": "text", + "content": "We implement Stereo Any Video in PyTorch and train on NVIDIA A100 GPUs. The model is trained with a batch size of 8 on SceneFlow [34] for " + }, + { + "bbox": [ + 55, + 422, + 296, + 601 + ], + "type": "inline_equation", + "content": "120k" + }, + { + "bbox": [ + 55, + 422, + 296, + 601 + ], + "type": "text", + "content": " iterations and finetuned on a mixed dataset for " + }, + { + "bbox": [ + 55, + 422, + 296, + 601 + ], + "type": "inline_equation", + "content": "80k" + }, + { + "bbox": [ + 55, + 422, + 296, + 601 + ], + "type": "text", + "content": " iterations consisting of Dynamic Replica [24] training set, Infinigen SV [23] training set, and Virtual KITTI2 [3]. Before being fed into the network, sequences are randomly cropped to " + }, + { + "bbox": [ + 55, + 422, + 296, + 601 + ], + "type": "inline_equation", + "content": "256 \\times 512" + }, + { + "bbox": [ + 55, + 422, + 296, + 601 + ], + "type": "text", + "content": ". We use the AdamW optimizer with a learning rate of " + }, + { + "bbox": [ + 55, + 422, + 296, + 601 + ], + "type": "inline_equation", + "content": "2e^{-4}" + }, + { + "bbox": [ + 55, + 422, + 296, + 601 + ], + "type": "text", + "content": " and apply a one-cycle learning rate schedule. The full training process takes approximately 6 days. Following [24], we apply various data augmentation techniques, including random cropping, rescaling, and saturation shifts. The sequence length is set to " + }, + { + "bbox": [ + 55, + 422, + 296, + 601 + ], + "type": "inline_equation", + "content": "T = 5" + }, + { + "bbox": [ + 55, + 422, + 296, + 601 + ], + "type": "text", + "content": " during training, and " + }, + { + "bbox": [ + 55, + 422, + 296, + 601 + ], + "type": "inline_equation", + "content": "T = 20" + }, + { + "bbox": [ + 55, + 422, + 296, + 601 + ], + "type": "text", + "content": " for evaluation. The number of iterations in GRU is set to " + }, + { + "bbox": [ + 55, + 422, + 296, + 601 + ], + "type": "inline_equation", + "content": "N = 10" + }, + { + "bbox": [ + 55, + 422, + 296, + 601 + ], + "type": "text", + "content": " in training, and " + }, + { + "bbox": [ + 55, + 422, + 296, + 601 + ], + "type": "inline_equation", + "content": "N = 20" + }, + { + "bbox": [ + 55, + 422, + 296, + 601 + ], + "type": "text", + "content": " in evaluation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 611, + 129, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 129, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 129, + 623 + ], + "type": "text", + "content": "4.3. Evaluation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "Quantitative Results. Table 2 presents the results when models are trained solely on SceneFlow [34] and evaluated on four dynamic synthetic datasets including diverse indoor and outdoor environments. The results clearly demonstrate that our model achieves state-of-the-art zero-shot generalization across nearly all metrics, excelling in both accuracy and temporal consistency. Table 3 reports the results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 274, + 555, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 555, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 555, + 405 + ], + "type": "text", + "content": "when models are allowed to train on any existing datasets. Notably, despite being trained solely on synthetic datasets, our model surpasses comparison methods that are trained on real-world datasets or share the same domain as the test tests by at least " + }, + { + "bbox": [ + 313, + 274, + 555, + 405 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 313, + 274, + 555, + 405 + ], + "type": "text", + "content": " on diverse benchmarks. Specifically, compared to the robust version of RAFT-Stereo, which is trained on 10 mixed datasets, including in-domain sets such as Sintel [2] and KITTI [13], our approach achieves superior performance across most metrics. These results highlight the strong generalization capability of our method when applied to unseen domains." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 407, + 556, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 556, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 556, + 634 + ], + "type": "text", + "content": "Qualitative Results. Figure 5 and Figure 6 present qualitative results using the checkpoints from Table 3 on dynamic real-world stereo videos from the South Kensington SV dataset [23]. For the indoor scenario shown in Figure 5, the upper rows display three input left frames along with their predicted disparities, while the subsequent rows visualize the globally aligned reconstructed point cloud. The red arrows highlight that our method produces temporally consistent and accurate disparities, whereas other methods exhibit noticeable artifacts. Specifically, Selective-IGEV and DynamicStereo suffer from flickering artifacts, with cabinets and walls displaying low-frequency oscillations. BiDAStereo fails to maintain an accurate shape, resulting in a distorted table. For the outdoor scenario in Figure 6, the driving car presents a challenging case due to its transparent surfaces and rapid motion. As shown, all other methods struggle with this case, failing to produce reliable disparity estimations. In contrast, our method demonstrates superior performance, maintaining accurate and stable predictions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 647, + 408, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 408, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 408, + 659 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "Foundation Priors. As shown in Table 4, incorporating foundation priors from a monocular depth model improves performance. Notably, VDA [7] outperforms DepthAnythingV2 [68] due to its stronger temporal consistency. In" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20841" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 88, + 71, + 160, + 112 + ], + "blocks": [ + { + "bbox": [ + 88, + 71, + 160, + 112 + ], + "lines": [ + { + "bbox": [ + 88, + 71, + 160, + 112 + ], + "spans": [ + { + "bbox": [ + 88, + 71, + 160, + 112 + ], + "type": "image", + "image_path": "736c3fda992504b6dc731ac7da637f6d8941eb65923948b509378eab1b34dbb4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 90, + 114, + 158, + 133 + ], + "blocks": [ + { + "bbox": [ + 90, + 114, + 158, + 133 + ], + "lines": [ + { + "bbox": [ + 90, + 114, + 158, + 133 + ], + "spans": [ + { + "bbox": [ + 90, + 114, + 158, + 133 + ], + "type": "image", + "image_path": "f160bca8a8e4544ba702f3a377734cc92121b30eedd4d6716b0233c7317f5c08.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 91, + 134, + 158, + 155 + ], + "blocks": [ + { + "bbox": [ + 91, + 134, + 158, + 155 + ], + "lines": [ + { + "bbox": [ + 91, + 134, + 158, + 155 + ], + "spans": [ + { + "bbox": [ + 91, + 134, + 158, + 155 + ], + "type": "image", + "image_path": "9a565c3341946218f42901cd9a98f1610e93a7b30c14211333ae6d03cdc28801.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 90, + 156, + 158, + 175 + ], + "blocks": [ + { + "bbox": [ + 90, + 156, + 158, + 175 + ], + "lines": [ + { + "bbox": [ + 90, + 156, + 158, + 175 + ], + "spans": [ + { + "bbox": [ + 90, + 156, + 158, + 175 + ], + "type": "image", + "image_path": "44288cbd317c07b07ba5af3cd445785637d0ac5ed8d36a74199fe7ecf91b78b9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 176, + 145, + 183 + ], + "lines": [ + { + "bbox": [ + 104, + 176, + 145, + 183 + ], + "spans": [ + { + "bbox": [ + 104, + 176, + 145, + 183 + ], + "type": "text", + "content": "Left Frames" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 162, + 72, + 252, + 154 + ], + "blocks": [ + { + "bbox": [ + 162, + 72, + 252, + 154 + ], + "lines": [ + { + "bbox": [ + 162, + 72, + 252, + 154 + ], + "spans": [ + { + "bbox": [ + 162, + 72, + 252, + 154 + ], + "type": "image", + "image_path": "b7488bb0df63d97a0e0d2df13475336313207a18ea7be9f5198ecd2d584b60e9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 165, + 158, + 252, + 186 + ], + "blocks": [ + { + "bbox": [ + 165, + 158, + 252, + 186 + ], + "lines": [ + { + "bbox": [ + 165, + 158, + 252, + 186 + ], + "spans": [ + { + "bbox": [ + 165, + 158, + 252, + 186 + ], + "type": "image", + "image_path": "d44f8970934e84c5523e58e54e206acdd7a26bf2aa7900b04d73d57af5942749.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 166, + 187, + 252, + 216 + ], + "blocks": [ + { + "bbox": [ + 166, + 187, + 252, + 216 + ], + "lines": [ + { + "bbox": [ + 166, + 187, + 252, + 216 + ], + "spans": [ + { + "bbox": [ + 166, + 187, + 252, + 216 + ], + "type": "image", + "image_path": "85dbea3b21284cdf3055372e945442dac28f8d1b50f8beb38c8839304edbe66d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 167, + 216, + 252, + 243 + ], + "blocks": [ + { + "bbox": [ + 167, + 216, + 252, + 243 + ], + "lines": [ + { + "bbox": [ + 167, + 216, + 252, + 243 + ], + "spans": [ + { + "bbox": [ + 167, + 216, + 252, + 243 + ], + "type": "image", + "image_path": "ea10315e139192c6cbe6c8d6c96e675719601c174a345938717bcaf990e1d9a2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 166, + 245, + 233, + 253 + ], + "lines": [ + { + "bbox": [ + 166, + 245, + 233, + 253 + ], + "spans": [ + { + "bbox": [ + 166, + 245, + 233, + 253 + ], + "type": "text", + "content": "Selective-IGEV" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 88, + 196, + 164, + 236 + ], + "blocks": [ + { + "bbox": [ + 88, + 196, + 164, + 236 + ], + "lines": [ + { + "bbox": [ + 88, + 196, + 164, + 236 + ], + "spans": [ + { + "bbox": [ + 88, + 196, + 164, + 236 + ], + "type": "image", + "image_path": "d98300b824e70c39d81ed932274f6dee03db0533785a3cd9f6dc3775dcda0d08.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 245, + 145, + 253 + ], + "lines": [ + { + "bbox": [ + 104, + 245, + 145, + 253 + ], + "spans": [ + { + "bbox": [ + 104, + 245, + 145, + 253 + ], + "type": "text", + "content": "Point Cloud" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 253, + 72, + 342, + 154 + ], + "blocks": [ + { + "bbox": [ + 253, + 72, + 342, + 154 + ], + "lines": [ + { + "bbox": [ + 253, + 72, + 342, + 154 + ], + "spans": [ + { + "bbox": [ + 253, + 72, + 342, + 154 + ], + "type": "image", + "image_path": "6fbd86428f71a0bb34023107057bbe11ba5bb9cf50283e35bad7f3994d267e26.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 255, + 158, + 342, + 186 + ], + "blocks": [ + { + "bbox": [ + 255, + 158, + 342, + 186 + ], + "lines": [ + { + "bbox": [ + 255, + 158, + 342, + 186 + ], + "spans": [ + { + "bbox": [ + 255, + 158, + 342, + 186 + ], + "type": "image", + "image_path": "ff3d60f55a48ddfd65e12485e87233d2e9123539a4b332ca9850e9cec3cd6bfb.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 255, + 187, + 342, + 244 + ], + "blocks": [ + { + "bbox": [ + 255, + 187, + 342, + 244 + ], + "lines": [ + { + "bbox": [ + 255, + 187, + 342, + 244 + ], + "spans": [ + { + "bbox": [ + 255, + 187, + 342, + 244 + ], + "type": "image", + "image_path": "d0ce17c9f8872f60798012105e1962fb4b202bf7b14badd936987b3694ea2009.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 271, + 246, + 323, + 254 + ], + "lines": [ + { + "bbox": [ + 271, + 246, + 323, + 254 + ], + "spans": [ + { + "bbox": [ + 271, + 246, + 323, + 254 + ], + "type": "text", + "content": "DynamicStereo" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 343, + 72, + 432, + 154 + ], + "blocks": [ + { + "bbox": [ + 343, + 72, + 432, + 154 + ], + "lines": [ + { + "bbox": [ + 343, + 72, + 432, + 154 + ], + "spans": [ + { + "bbox": [ + 343, + 72, + 432, + 154 + ], + "type": "image", + "image_path": "f70935bd9499c1994e96b4ca62cd1769600fad45083657ae2d7b9367a299735f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 345, + 158, + 432, + 186 + ], + "blocks": [ + { + "bbox": [ + 345, + 158, + 432, + 186 + ], + "lines": [ + { + "bbox": [ + 345, + 158, + 432, + 186 + ], + "spans": [ + { + "bbox": [ + 345, + 158, + 432, + 186 + ], + "type": "image", + "image_path": "7ace61094145e86490fcde360569048724eb0f76e67f2a905b63076d519074da.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 345, + 187, + 432, + 216 + ], + "blocks": [ + { + "bbox": [ + 345, + 187, + 432, + 216 + ], + "lines": [ + { + "bbox": [ + 345, + 187, + 432, + 216 + ], + "spans": [ + { + "bbox": [ + 345, + 187, + 432, + 216 + ], + "type": "image", + "image_path": "effb0f15aa8535a7a9f4eb7dca62643cacb325ace47f162d8f5d96364b658bf8.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 345, + 216, + 432, + 244 + ], + "blocks": [ + { + "bbox": [ + 345, + 216, + 432, + 244 + ], + "lines": [ + { + "bbox": [ + 345, + 216, + 432, + 244 + ], + "spans": [ + { + "bbox": [ + 345, + 216, + 432, + 244 + ], + "type": "image", + "image_path": "fbc1a3a078da0899f25f0ccf9759f9b048af29d51530011b7d221e2fcb1f8500.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 246, + 407, + 254 + ], + "lines": [ + { + "bbox": [ + 367, + 246, + 407, + 254 + ], + "spans": [ + { + "bbox": [ + 367, + 246, + 407, + 254 + ], + "type": "text", + "content": "BiDAStereo" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 434, + 72, + 522, + 99 + ], + "blocks": [ + { + "bbox": [ + 434, + 72, + 522, + 99 + ], + "lines": [ + { + "bbox": [ + 434, + 72, + 522, + 99 + ], + "spans": [ + { + "bbox": [ + 434, + 72, + 522, + 99 + ], + "type": "image", + "image_path": "619395f338c6137cad9f0e295e9f51bcd9d2e2278fcca234655b801a618d4965.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 435, + 99, + 522, + 127 + ], + "blocks": [ + { + "bbox": [ + 435, + 99, + 522, + 127 + ], + "lines": [ + { + "bbox": [ + 435, + 99, + 522, + 127 + ], + "spans": [ + { + "bbox": [ + 435, + 99, + 522, + 127 + ], + "type": "image", + "image_path": "3b04867844d804312a68fe52dedb2d3a12364c35f9b7e51170ff6381a383e22d.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 435, + 128, + 522, + 154 + ], + "blocks": [ + { + "bbox": [ + 435, + 128, + 522, + 154 + ], + "lines": [ + { + "bbox": [ + 435, + 128, + 522, + 154 + ], + "spans": [ + { + "bbox": [ + 435, + 128, + 522, + 154 + ], + "type": "image", + "image_path": "8e75d31ec343f7b3aad94256509e56a684d3e991cf9871b6616afd8f50478e93.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 435, + 160, + 522, + 186 + ], + "blocks": [ + { + "bbox": [ + 435, + 160, + 522, + 186 + ], + "lines": [ + { + "bbox": [ + 435, + 160, + 522, + 186 + ], + "spans": [ + { + "bbox": [ + 435, + 160, + 522, + 186 + ], + "type": "image", + "image_path": "d89ba3d7ac560c0ce86eb1d1f844bbbdeadc5dcd43b7e4e7f54e94d3b6c03a12.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 435, + 187, + 522, + 216 + ], + "blocks": [ + { + "bbox": [ + 435, + 187, + 522, + 216 + ], + "lines": [ + { + "bbox": [ + 435, + 187, + 522, + 216 + ], + "spans": [ + { + "bbox": [ + 435, + 187, + 522, + 216 + ], + "type": "image", + "image_path": "258905af26531a4384573582b2453fb22b0b051832302d08403475c7f09e88ed.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 435, + 216, + 522, + 244 + ], + "blocks": [ + { + "bbox": [ + 435, + 216, + 522, + 244 + ], + "lines": [ + { + "bbox": [ + 435, + 216, + 522, + 244 + ], + "spans": [ + { + "bbox": [ + 435, + 216, + 522, + 244 + ], + "type": "image", + "image_path": "1a44549eeef01ed5e42cacff907c027ff01a812aa025558e70f6744f3c699a46.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 437, + 245, + 517, + 254 + ], + "lines": [ + { + "bbox": [ + 437, + 245, + 517, + 254 + ], + "spans": [ + { + "bbox": [ + 437, + 245, + 517, + 254 + ], + "type": "text", + "content": "Stereo Any Video (ours)" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 87, + 318, + 175, + 367 + ], + "blocks": [ + { + "bbox": [ + 55, + 266, + 555, + 310 + ], + "lines": [ + { + "bbox": [ + 55, + 266, + 555, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 555, + 310 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison on a dynamic indoor scenario from the South Kensington SV dataset [23]. Top: Input left frames and predicted disparities for the same region across three frames using different methods. Bottom: Disparities converted into globally aligned point clouds and rendered from a camera view for comparison. Failure regions are highlighted with red arrows. Our method produces consistent and accurate disparities without flickering. See video in the supplementary material for better visualization." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 87, + 318, + 175, + 367 + ], + "lines": [ + { + "bbox": [ + 87, + 318, + 175, + 367 + ], + "spans": [ + { + "bbox": [ + 87, + 318, + 175, + 367 + ], + "type": "image", + "image_path": "11369eeeffbd1795bfc5875220b437ec0819aaabf620332545a53e83dc2a4418.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 88, + 368, + 174, + 418 + ], + "blocks": [ + { + "bbox": [ + 88, + 368, + 174, + 418 + ], + "lines": [ + { + "bbox": [ + 88, + 368, + 174, + 418 + ], + "spans": [ + { + "bbox": [ + 88, + 368, + 174, + 418 + ], + "type": "image", + "image_path": "ef3895ca7083cb8da6bc7e3838c797de5cfa34b5dd9a4cd8a21d392ae89729cf.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 88, + 418, + 174, + 467 + ], + "blocks": [ + { + "bbox": [ + 88, + 418, + 174, + 467 + ], + "lines": [ + { + "bbox": [ + 88, + 418, + 174, + 467 + ], + "spans": [ + { + "bbox": [ + 88, + 418, + 174, + 467 + ], + "type": "image", + "image_path": "029f957fb7a5dadd02fd8ae654dd365d50eae7dfaf63d1e680dd59b1b8db61c5.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 88, + 468, + 174, + 517 + ], + "blocks": [ + { + "bbox": [ + 88, + 468, + 174, + 517 + ], + "lines": [ + { + "bbox": [ + 88, + 468, + 174, + 517 + ], + "spans": [ + { + "bbox": [ + 88, + 468, + 174, + 517 + ], + "type": "image", + "image_path": "3688cad9176a35789c6b6570dc82f8ce9d01f7ec5b65d8d08804543ee24b2af2.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 111, + 519, + 151, + 526 + ], + "lines": [ + { + "bbox": [ + 111, + 519, + 151, + 526 + ], + "spans": [ + { + "bbox": [ + 111, + 519, + 151, + 526 + ], + "type": "text", + "content": "Left Frames" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 176, + 318, + 261, + 368 + ], + "blocks": [ + { + "bbox": [ + 176, + 318, + 261, + 368 + ], + "lines": [ + { + "bbox": [ + 176, + 318, + 261, + 368 + ], + "spans": [ + { + "bbox": [ + 176, + 318, + 261, + 368 + ], + "type": "image", + "image_path": "cb145eceb4f686bffdba1926ab0a694e990279c00214f4652de1eff0b22d148b.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 176, + 369, + 261, + 418 + ], + "blocks": [ + { + "bbox": [ + 176, + 369, + 261, + 418 + ], + "lines": [ + { + "bbox": [ + 176, + 369, + 261, + 418 + ], + "spans": [ + { + "bbox": [ + 176, + 369, + 261, + 418 + ], + "type": "image", + "image_path": "bd8cc6cd940f7df338036be6fc2581f5e137bc774e97fb6234b406b2fd68c5a7.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 176, + 418, + 261, + 468 + ], + "blocks": [ + { + "bbox": [ + 176, + 418, + 261, + 468 + ], + "lines": [ + { + "bbox": [ + 176, + 418, + 261, + 468 + ], + "spans": [ + { + "bbox": [ + 176, + 418, + 261, + 468 + ], + "type": "image", + "image_path": "2ad3a7daffc029ae491509083022a2d3c9b8d7150ccc7a37f4362ebb0ffd15a0.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 176, + 468, + 261, + 517 + ], + "blocks": [ + { + "bbox": [ + 176, + 468, + 261, + 517 + ], + "lines": [ + { + "bbox": [ + 176, + 468, + 261, + 517 + ], + "spans": [ + { + "bbox": [ + 176, + 468, + 261, + 517 + ], + "type": "image", + "image_path": "383f8ace1592521984b26dd8727dda2b7ce96d76fe7c207413f17bff1ba6d308.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 518, + 242, + 526 + ], + "lines": [ + { + "bbox": [ + 192, + 518, + 242, + 526 + ], + "spans": [ + { + "bbox": [ + 192, + 518, + 242, + 526 + ], + "type": "text", + "content": "Selective-IGEV" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 261, + 318, + 348, + 368 + ], + "blocks": [ + { + "bbox": [ + 261, + 318, + 348, + 368 + ], + "lines": [ + { + "bbox": [ + 261, + 318, + 348, + 368 + ], + "spans": [ + { + "bbox": [ + 261, + 318, + 348, + 368 + ], + "type": "image", + "image_path": "a259d02c94925c90bc21ef9738c90e24db2646d10b30c7073955d9273c23d6fa.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 261, + 369, + 348, + 418 + ], + "blocks": [ + { + "bbox": [ + 261, + 369, + 348, + 418 + ], + "lines": [ + { + "bbox": [ + 261, + 369, + 348, + 418 + ], + "spans": [ + { + "bbox": [ + 261, + 369, + 348, + 418 + ], + "type": "image", + "image_path": "c698e9275cb9d9c2920f0cab98334d62bb980da10afb475a9d7e14853b9076ec.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 261, + 418, + 348, + 468 + ], + "blocks": [ + { + "bbox": [ + 261, + 418, + 348, + 468 + ], + "lines": [ + { + "bbox": [ + 261, + 418, + 348, + 468 + ], + "spans": [ + { + "bbox": [ + 261, + 418, + 348, + 468 + ], + "type": "image", + "image_path": "20bbc549b77dfdfcad5b3163cf839b71ac0c3f61f16fb175677d2d447d408a1e.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 261, + 468, + 348, + 517 + ], + "blocks": [ + { + "bbox": [ + 261, + 468, + 348, + 517 + ], + "lines": [ + { + "bbox": [ + 261, + 468, + 348, + 517 + ], + "spans": [ + { + "bbox": [ + 261, + 468, + 348, + 517 + ], + "type": "image", + "image_path": "b7ed7704e7c68d377d1939cd8ca964db0189d839b7ee4e118b2591a3f77d90a4.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 280, + 519, + 329, + 526 + ], + "lines": [ + { + "bbox": [ + 280, + 519, + 329, + 526 + ], + "spans": [ + { + "bbox": [ + 280, + 519, + 329, + 526 + ], + "type": "text", + "content": "DynamicStereo" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 348, + 318, + 435, + 368 + ], + "blocks": [ + { + "bbox": [ + 348, + 318, + 435, + 368 + ], + "lines": [ + { + "bbox": [ + 348, + 318, + 435, + 368 + ], + "spans": [ + { + "bbox": [ + 348, + 318, + 435, + 368 + ], + "type": "image", + "image_path": "9e6dc2b7bb5e82e1ed53812e7ee25e8fb09b95a60e3d51be84480b6a85325765.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 348, + 369, + 435, + 418 + ], + "blocks": [ + { + "bbox": [ + 348, + 369, + 435, + 418 + ], + "lines": [ + { + "bbox": [ + 348, + 369, + 435, + 418 + ], + "spans": [ + { + "bbox": [ + 348, + 369, + 435, + 418 + ], + "type": "image", + "image_path": "65b7044e8cad1cc88607d5bea97f09c46dd48d6da791b5b6897975e09f04df5a.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 348, + 418, + 435, + 468 + ], + "blocks": [ + { + "bbox": [ + 348, + 418, + 435, + 468 + ], + "lines": [ + { + "bbox": [ + 348, + 418, + 435, + 468 + ], + "spans": [ + { + "bbox": [ + 348, + 418, + 435, + 468 + ], + "type": "image", + "image_path": "ff651e9d0149255bdee8f1895e3b864f5fd2b325a04b6960320a8e872cc36265.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 348, + 468, + 435, + 517 + ], + "blocks": [ + { + "bbox": [ + 348, + 468, + 435, + 517 + ], + "lines": [ + { + "bbox": [ + 348, + 468, + 435, + 517 + ], + "spans": [ + { + "bbox": [ + 348, + 468, + 435, + 517 + ], + "type": "image", + "image_path": "b199b1546b977a937308f01d412b3ca5b9ade1941e5e15d426267fadcbf3b375.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 519, + 411, + 526 + ], + "lines": [ + { + "bbox": [ + 372, + 519, + 411, + 526 + ], + "spans": [ + { + "bbox": [ + 372, + 519, + 411, + 526 + ], + "type": "text", + "content": "BiDAStereo" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 107, + 540, + 501, + 552 + ], + "lines": [ + { + "bbox": [ + 107, + 540, + 501, + 552 + ], + "spans": [ + { + "bbox": [ + 107, + 540, + 501, + 552 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison on a dynamic outdoor scenario from the South Kensington SV dataset [23]." + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_caption" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 435, + 318, + 522, + 368 + ], + "blocks": [ + { + "bbox": [ + 435, + 318, + 522, + 368 + ], + "lines": [ + { + "bbox": [ + 435, + 318, + 522, + 368 + ], + "spans": [ + { + "bbox": [ + 435, + 318, + 522, + 368 + ], + "type": "image", + "image_path": "69cffa071f143597de1bea78beb975b0d645443dd3abb8b5c8d82eb8cf2b112d.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 435, + 369, + 522, + 418 + ], + "blocks": [ + { + "bbox": [ + 435, + 369, + 522, + 418 + ], + "lines": [ + { + "bbox": [ + 435, + 369, + 522, + 418 + ], + "spans": [ + { + "bbox": [ + 435, + 369, + 522, + 418 + ], + "type": "image", + "image_path": "bc471e13c904b3476c0b61295341317f454b3a536c252d6b82516dbf7929506f.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 435, + 418, + 522, + 468 + ], + "blocks": [ + { + "bbox": [ + 435, + 418, + 522, + 468 + ], + "lines": [ + { + "bbox": [ + 435, + 418, + 522, + 468 + ], + "spans": [ + { + "bbox": [ + 435, + 418, + 522, + 468 + ], + "type": "image", + "image_path": "11259b7856e788a1da1da1e4ededf85c0dad292c9b26232cf19da5a0a3392e13.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 435, + 468, + 522, + 517 + ], + "blocks": [ + { + "bbox": [ + 435, + 468, + 522, + 517 + ], + "lines": [ + { + "bbox": [ + 435, + 468, + 522, + 517 + ], + "spans": [ + { + "bbox": [ + 435, + 468, + 522, + 517 + ], + "type": "image", + "image_path": "66567c4e513ab1c9adf73b6221151e9b3a6b8352ac670f4f69a23c99e1632cb9.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 519, + 516, + 527 + ], + "lines": [ + { + "bbox": [ + 440, + 519, + 516, + 527 + ], + "spans": [ + { + "bbox": [ + 440, + 519, + 516, + 527 + ], + "type": "text", + "content": "Stereo Any Video (ours)" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + } + ], + "index": 52 + }, + { + "bbox": [ + 55, + 567, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 296, + 712 + ], + "type": "text", + "content": "terestingly, increasing the backbone size does not yield additional performance gains. To investigate this, we conducted further experiments (Table 5) by varying depth and CNN feature channels based on VDA-L. However, the results remained similar. According to [7], VDA-L has " + }, + { + "bbox": [ + 55, + 567, + 296, + 712 + ], + "type": "inline_equation", + "content": "13 \\times" + }, + { + "bbox": [ + 55, + 567, + 296, + 712 + ], + "type": "text", + "content": " more parameters than VDA-S (381.8M vs. 28.4M), yet the performance gain is marginal (" + }, + { + "bbox": [ + 55, + 567, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\delta_{1}(\\uparrow)" + }, + { + "bbox": [ + 55, + 567, + 296, + 712 + ], + "type": "text", + "content": ": 0.944 vs. 0.942 on KITTI; 0.971 vs. 0.959 on NYUv2). This suggests that the larger backbone does not provide richer priors. Moreover, large backbone requires additional CNN channels and parameters for adaptation, increasing training complexity. Therefore, we adopt VDA-S in the final model." + } + ] + } + ], + "index": 55 + }, + { + "bbox": [ + 313, + 567, + 555, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 555, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 555, + 675 + ], + "type": "text", + "content": "Correlation. To evaluate the effectiveness of our proposed all-to-all-pairs correlation, we replace it with local correlation [27] and experiment with using only 1D or 2D correlation instead of alternating between them. As shown in Table 4, all these modifications degrade performance. Notably, constructing 3D correlations—where the search window spans height, width, and time—further worsens results. This suggests that frames cannot be effectively utilized without proper alignment [23]." + } + ] + } + ], + "index": 56 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Cost Volume Encoder. Instead of directly encoding cost volumes with convolutions, we find that introducing an MLP layer to extract a compact representation before ap" + } + ] + } + ], + "index": 57 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20842" + } + ] + } + ], + "index": 58 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 118, + 294, + 399 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 115 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 115 + ], + "type": "text", + "content": "Table 4. Ablation study on the foundation priors, correlation, cost volume encoder, convex upsampling, attention, and inference settings of our model on Sintel [2] and Dynamic Replica [24]. The approach used in our final model is underlined." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 118, + 294, + 399 + ], + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 399 + ], + "type": "table", + "html": "
Setting VariationsSintel FinalDynamic Replica
TEPE↓δt1px ↓TEPE↓δt1px ↓
Foundation Priors
w/o Depth1.4712.350.0921.05
Depth AnythingV2-S [68]1.3311.270.0880.99
VDA-S [7]1.2710.980.0830.78
VDA-L [7]1.2012.470.0830.82
Correlation
1D + 2D local [27]1.2710.980.0830.78
1D all-to-all-pairs1.2311.360.0790.78
2D all-to-all-pairs1.2911.950.0850.77
3D all-to-all-pairs4.5122.901.3132.01
1D + 2D all-to-all-pairs1.2111.330.0760.75
Cost Volume Encoder
Convolution1.2611.540.0810.77
MLP + Convolution1.2111.330.0760.75
Convex Upsampling
Bilinear1.2612.740.0851.05
Convex [51]1.2111.330.0760.75
Temporal Convex1.0410.210.0670.79
Attention in GRU
w/o Attention1.0410.210.0670.79
Temporal1.0710.230.0640.79
Temporal + Spatial1.079.980.0570.67
Inference Stage
1/16 → 1/8 → 1/41.079.980.0570.67
1/32 → 1/16 → 1/8 → 1/41.0810.820.0580.70
", + "image_path": "6c01104a90d18c151e7ee8638e9fbb20d6656e02d2b54be447894c030dc8159e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 408, + 295, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 295, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 295, + 433 + ], + "type": "text", + "content": "plying convolutions enhances cost aggregation, leading to improved final results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 434, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 495 + ], + "type": "text", + "content": "Convex Upsampling. We compare bilinear upsampling and general convex upsampling [51] to our proposed temporal convex upsampling module. As observed, our approach achieves better temporal consistency by incorporating learnable weights with adjacent frames." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 497, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 295, + 545 + ], + "type": "text", + "content": "Attention in GRU. We further investigate different attention mechanisms within the GRU update block. As shown in Table 4, combining temporal and spatial attention together leads to the best performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 547, + 295, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 547, + 295, + 584 + ], + "spans": [ + { + "bbox": [ + 55, + 547, + 295, + 584 + ], + "type": "text", + "content": "Inference Stage. We also explore inference starting at different resolutions. As can be seen in Table 4, a three-stage inference setup already achieves strong results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 597, + 133, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 133, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 133, + 611 + ], + "type": "text", + "content": "4.5. Application" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "Our method enhances downstream applications that rely on accurate metric depth, such as 3D point tracking. As shown in Figure 7, we visualize results using the state-of-the-art 3D point tracking method DELTA [36]. Our predicted metric depth yields more points with precise trajectories than UniDepth [40], which gives less missing points as highlighted by the red arrows. Additional application, such as visual effects, are available in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 333, + 106, + 533, + 304 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 554, + 104 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 554, + 104 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 554, + 104 + ], + "type": "text", + "content": "Table 5. Ablation study on the effect of different channel configurations for depth features and CNN features, based on VDA-L [7], on Sintel [2]. Trainable parameters are also reported." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 333, + 106, + 533, + 304 + ], + "lines": [ + { + "bbox": [ + 333, + 106, + 533, + 304 + ], + "spans": [ + { + "bbox": [ + 333, + 106, + 533, + 304 + ], + "type": "table", + "html": "
Depth & CNNSintel FinalPara. (M)
Channel NumberTEPE↓δt1px ↓
VDA-L [7]
(32, 96)1.2012.478.4
(128, 64)1.5814.5213.3
(128, 128)1.3113.2119.3
UniDepthStereo Any Video (ours)
", + "image_path": "c0bdf386ff01b859ba62db2547b4deb9977cab337786cd1ff7ddd451cdc1d662.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 379, + 553, + 431 + ], + "blocks": [ + { + "bbox": [ + 313, + 310, + 554, + 333 + ], + "lines": [ + { + "bbox": [ + 313, + 310, + 554, + 333 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 554, + 333 + ], + "type": "text", + "content": "Figure 7. Visualization of 3D point tracking based on metric depth generated from UniDepth and our method." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 341, + 554, + 374 + ], + "lines": [ + { + "bbox": [ + 313, + 341, + 554, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 554, + 374 + ], + "type": "text", + "content": "Table 6. Comparison of different methods on model trainable [non-trainable] parameters, inference GPU memory, and multiply-accumulate (MAC) on 20 frames with resolution " + }, + { + "bbox": [ + 313, + 341, + 554, + 374 + ], + "type": "inline_equation", + "content": "720 \\times 1280" + }, + { + "bbox": [ + 313, + 341, + 554, + 374 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 379, + 553, + 431 + ], + "lines": [ + { + "bbox": [ + 315, + 379, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 315, + 379, + 553, + 431 + ], + "type": "table", + "html": "
MethodsPara. (M)Mem. (G)MACs (T)
DynamicStereo [24]20.535.1182.3
BiDASTereo [22]12.241.1186.6
Stereo Any Video (ours)9.4 [28.4]41.1303.4
", + "image_path": "90834854a545c3bc71a285b0d0ffb8836b99c5c11bde7b33f1cc1f3eee8c972e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 449, + 471, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 471, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 471, + 460 + ], + "type": "text", + "content": "4.6. Limitation and Future Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 467, + 554, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 554, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 554, + 538 + ], + "type": "text", + "content": "We report the inference efficiency results on a single NVIDIA A6000 GPU in Table 6. As shown, due to the use of foundation priors, our method incurs relatively high computational and memory costs. Looking ahead, we aim to develop a model zoo for this method, including both a large and a light version." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we introduce Stereo Any Video, a robust framework for video stereo matching. By leveraging rich monocular priors, an all-to-all-pairs correlation mechanism, and a temporal convex upsampling layer, our approach produces temporally consistent and accurate disparities without relying on additional inputs such as camera pose or optical flow. Extensive evaluations show that our model achieves state-of-the-art performance on multiple standard benchmarks under zero-shot settings. Furthermore, our method generalizes well to real-world stereo videos, handling diverse content, motion, and camera movements effectively. Acknowledgments. This work was funded by the IC-CSC." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20843" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Luca Bartolomei, Fabio Tosi, Matteo Poggi, and Stefano Mattoccia. Stereo anywhere: Robust zero-shot deep stereo matching even where either stereo or mono fail. arXiv preprint arXiv:2412.04472, 2024.3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 169 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 169 + ], + "type": "text", + "content": "[2] D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black. A naturalistic open source movie for optical flow evaluation. In ECCV, pages 611-625, 2012. 4, 5, 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 171, + 294, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 171, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 62, + 171, + 294, + 192 + ], + "type": "text", + "content": "[3] Yohann Cabon, Naila Murray, and Martin Humenberger. Virtual kitti 2, 2020. 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 194, + 295, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 295, + 249 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 295, + 249 + ], + "type": "text", + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 251, + 294, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 294, + 273 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 294, + 273 + ], + "type": "text", + "content": "[5] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In CVPR, pages 5410-5418, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 274, + 295, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 295, + 327 + ], + "type": "text", + "content": "[6] Tianyu Chang, Xun Yang, Tianzhu Zhang, and Meng Wang. Domain generalized stereo matching via hierarchical visual transformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9559-9568, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 331, + 294, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 331, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 62, + 331, + 294, + 373 + ], + "type": "text", + "content": "[7] Sili Chen, Hengkai Guo, Shengnan Zhu, Feihu Zhang, Zi long Huang, Jiashi Feng, and Bingyi Kang. Video depth anything: Consistent depth estimation for super-long videos, 2025. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 376, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 376, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 62, + 376, + 294, + 419 + ], + "type": "text", + "content": "[8] Junda Cheng, Longliang Liu, Gangwei Xu, Xianqi Wang, Zhaoxing Zhang, Yong Deng, Jinliang Zang, Yurui Chen, Zhipeng Cai, and Xin Yang. Monster: Marry monodepth to stereo unleashes power, 2025. 3, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 421, + 294, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 421, + 294, + 474 + ], + "spans": [ + { + "bbox": [ + 62, + 421, + 294, + 474 + ], + "type": "text", + "content": "[9] Xuelian Cheng, Yiran Zhong, Mehrtash Harandi, Yuchao Dai, Xiaojun Chang, Tom Drummond, Hongdong Li, and Zongyuan Ge. Hierarchical neural architecture search for deep stereo matching. arXiv preprint arXiv:2010.13501, 2020.2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 477, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 477, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 57, + 477, + 295, + 521 + ], + "type": "text", + "content": "[10] Ziang Cheng, Jiayu Yang, and Hongdong Li. Stereo matching in time: " + }, + { + "bbox": [ + 57, + 477, + 295, + 521 + ], + "type": "inline_equation", + "content": "100+" + }, + { + "bbox": [ + 57, + 477, + 295, + 521 + ], + "type": "text", + "content": " fps video stereo matching for extended reality. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 8719-8728, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 523, + 294, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 523, + 294, + 565 + ], + "spans": [ + { + "bbox": [ + 57, + 523, + 294, + 565 + ], + "type": "text", + "content": "[11] Seokju Cho, Jiahui Huang, Jisu Nam, Honggyu An, Seun-gryong Kim, and Joon-Young Lee. Local all-pair correspondence for point tracking. arXiv preprint arXiv:2407.15420, 2024. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 567, + 295, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 295, + 643 + ], + "type": "text", + "content": "[12] WeiQin Chuah, Ruwan Tennakoon, Reza Hoseinnezhad, Alireza Bab-Hadiashar, and David Suter. Itsa: An information-theoretic approach to automatic shortcut avoidance and domain generalization in stereo matching networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13022-13032, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 678 + ], + "type": "text", + "content": "[13] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, pages 3354-3361, 2012. 1, 2, 5, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "text", + "content": "[14] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In Proceedings of" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2495-2504, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 96, + 553, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 148 + ], + "type": "text", + "content": "[15] Weiyu Guo, Zhaoshuo Li, Yongkui Yang, Zheng Wang, Russell H Taylor, Mathias Unberath, Alan Yuille, and Yingwei Li. Context-enhanced stereo transformer. In European Conference on Computer Vision, pages 263-279. Springer, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 151, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 553, + 183 + ], + "type": "text", + "content": "[16] Xiaoyang Guo, Kai Yang, Wukui Yang, Xiaogang Wang, and Hongsheng Li. Group-wise correlation stereo network. In CVPR, pages 3273-3282, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 185, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 553, + 227 + ], + "type": "text", + "content": "[17] Xianda Guo, Chenming Zhang, Youmin Zhang, Wenzhao Zheng, Dujun Nie, Matteo Poggi, and Long Chen. Light-stereo: Channel boost is all your need for efficient 2d cost aggregation, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 228, + 553, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 553, + 258 + ], + "type": "text", + "content": "[18] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 261, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 261, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 317, + 261, + 553, + 304 + ], + "type": "text", + "content": "[19] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. arXiv preprint arXiv:2409.02095, 2024. 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 306, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 306, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 317, + 306, + 553, + 338 + ], + "type": "text", + "content": "[20] Hualie Jiang, Zhiqiang Lou, Laiyan Ding, Rui Xu, Minglang Tan, Wenjie Jiang, and Rui Huang. Defom-stereo: Depth foundation model based stereo matching, 2025. 3, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 339, + 555, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 339, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 317, + 339, + 555, + 403 + ], + "type": "text", + "content": "[21] Junpeng Jing, Jiankun Li, Pengfei Xiong, Jiangyu Liu, Shuaicheng Liu, Yichen Guo, Xin Deng, Mai Xu, Lai Jiang, and Leonid Sigal. Uncertainty guided adaptive warping for robust and efficient stereo matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3318-3327, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 405, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 405, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 317, + 405, + 553, + 449 + ], + "type": "text", + "content": "[22] Junpeng Jing, Ye Mao, and Krystian Mikolajczyk. Match-stereo-videos: Bidirectional alignment for consistent dynamic stereo matching. In European Conference on Computer Vision, pages 415–432. Springer, 2024. 2, 4, 5, 6, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 449, + 553, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 553, + 481 + ], + "type": "text", + "content": "[23] Junpeng Jing, Ye Mao, Anlan Qiu, and Krystian Mikolajczyk. Match stereo videos via bidirectional alignment, 2024. 3, 5, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 483, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 483, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 317, + 483, + 553, + 547 + ], + "type": "text", + "content": "[24] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Dynamic stereo: Consistent dynamic depth from stereo videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13229-13239, 2023. 2, 4, 5, 6, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 548, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 553, + 592 + ], + "type": "text", + "content": "[25] Alex Kendall, Hayk Martirosyan, Saumitro Dasgupta, Peter Henry, Ryan Kennedy, Abraham Bachrach, and Adam Bry. End-to-end learning of geometry and context for deep stereo regression. In CVPR, pages 66-75, 2017. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 593, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 553, + 647 + ], + "type": "text", + "content": "[26] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 647, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 713 + ], + "type": "text", + "content": "[27] Jiankun Li, Peisen Wang, Pengfei Xiong, Tao Cai, Ziwei Yan, Lei Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. Practical stereo matching via cascaded recurrent network with adaptive correlation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16263-16272, 2022. 1, 2, 4, 7, 8" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20844" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "type": "text", + "content": "[28] Zhaoshuo Li, Xingtong Liu, Nathan Drenkow, Andy Ding, Francis X Creighton, Russell H Taylor, and Mathias Unberath. Revisiting stereo depth estimation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6197-6206, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 205 + ], + "type": "text", + "content": "[29] Zhaoshuo Li, Wei Ye, Dilin Wang, Francis X Creighton, Russell H Taylor, Ganesh Venkatesh, and Mathias Unberath. Temporally consistent online depth estimation in dynamic scenes. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 3018-3027, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 207, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 207, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 207, + 294, + 251 + ], + "type": "text", + "content": "[30] Zhengfa Liang, Yiliu Feng, Yulan Guo, Hengzhu Liu, Wei Chen, Linbo Qiao, Li Zhou, and Jianfeng Zhang. Learning for disparity estimation through feature constancy. In CVPR, pages 2811-2820, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 252, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 252, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 252, + 294, + 285 + ], + "type": "text", + "content": "[31] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. arXiv preprint arXiv:2109.07547, 2021. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 294, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 318 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 318 + ], + "type": "text", + "content": "[32] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60:91-110, 2004. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 353 + ], + "type": "text", + "content": "[33] Yang Luo, Xuanlei Zhao, Mengzhao Chen, Kaipeng Zhang, Wenqi Shao, Kai Wang, Zhangyang Wang, and Yang You. Enhance-a-video: Better generated video for free, 2024. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "type": "text", + "content": "[34] Nikolaus Mayer, Eddy Ig, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In CVPR, pages 4040-4048, 2016. 2, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "type": "text", + "content": "[35] Lukas Mehl, Jenny Schmalfuss, Azin Jahedi, Yaroslava Nalivayko, and Andres Bruhn. Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "type": "text", + "content": "[36] Tuan Duc Ngo, Peiye Zhuang, Chuang Gan, Evangelos Kalogerakis, Sergey Tulyakov, Hsin-Ying Lee, and Chaoyang Wang. Delta: Dense efficient long-range 3d tracking for any video, 2025. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 294, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 565 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 565 + ], + "type": "text", + "content": "[37] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 567, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 294, + 610 + ], + "type": "text", + "content": "[38] Jiahao Pang, Wenxiu Sun, Jimmy SJ Ren, Chengxi Yang, and Qiong Yan. Cascade residual learning: A two-stage convolutional neural network for stereo matching. In CVPRW, pages 887-895, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 613, + 294, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 655 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 655 + ], + "type": "text", + "content": "[39] Jiahao Pang, Wenxiu Sun, Chengxi Yang, Jimmy Ren, Ruichao Xiao, Jin Zeng, and Liang Lin. Zoom and learn: Generalizing deep stereo matching to novel domains. In CVPR, pages 2070-2079, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "type": "text", + "content": "[40] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Unidepth: Universal monocular metric depth estimation, 2024. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "text", + "content": "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "text", + "content": "[42] Zhibo Rao, Bangshu Xiong, Mingyi He, Yuchao Dai, Renjie He, Zhelun Shen, and Xing Li. Masked representation learning for domain generalized stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5435-5444, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 173, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 215 + ], + "type": "text", + "content": "[43] Ignacio Rocco, Mircea Cimpoi, Relja Arandjelović, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Neighbourhood consensus networks. Advances in neural information processing systems, 31, 2018. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 217, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 249 + ], + "type": "text", + "content": "[44] Daniel Scharstein and Richard Szeliski. A taxonomy and evaluation of dense two-frame stereo correspondence algorithms. IJCV, 47(1):7-42, 2002. 1, 2, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 251, + 553, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 303 + ], + "type": "text", + "content": "[45] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In CVPR, pages 3260-3269, 2017. 1, 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 304, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 553, + 337 + ], + "type": "text", + "content": "[46] Zhelun Shen, Yuchao Dai, and Zhibo Rao. Cfnet: Cascade and fused cost volume for robust stereo matching. In CVPR, pages 13906-13915, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 338, + 553, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 391 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 391 + ], + "type": "text", + "content": "[47] Zhelun Shen, Xibin Song, Yuchao Dai, Dingfu Zhou, Zhibo Rao, and Liangjun Zhang. Digging into uncertainty-based pseudo-label for robust stereo matching. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-18, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 393, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 425 + ], + "type": "text", + "content": "[48] Chenyang Si, Weichen Fan, Zhengyao Lv, Ziqi Huang, Yu Qiao, and Ziwei Liu. Repvideo: Rethinking cross-layer representation for video generation, 2025. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 426, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 469 + ], + "type": "text", + "content": "[49] Qing Su and Shihao Ji. Chitransformer: Towards reliable stereo from cues. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1939-1949, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "type": "text", + "content": "[50] Vladimir Tankovich, Christian Hane, Yinda Zhang, Adarsh Kowdle, Sean Fanello, and Sofien Bouaziz. Hitnet: Hierarchical iterative tile refinement network for real-time stereo matching. In CVPR, pages 14362-14372, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 515, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 547 + ], + "type": "text", + "content": "[51] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, pages 402-419, 2020. 2, 5, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 548, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 591 + ], + "type": "text", + "content": "[52] Prune Truong, Martin Danelljan, Luc V Gool, and Radu Timofte. Gocor: Bringing globally optimized correspondence volumes into your neural network. Advances in Neural Information Processing Systems, 33:14278-14290, 2020. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 593, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 645 + ], + "type": "text", + "content": "[53] Prune Truong, Martin Danelljan, and Radu Timofte. Glunet: Global-local universal network for dense flow and correspondences. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6258-6268, 2020. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[54] Jonas Uhrig, Nick Schneider, Lukas Schneider, Uwe Franke, Thomas Brox, and Andreas Geiger. Sparsity invariant cnns. In 2017 international conference on 3D Vision (3DV), pages 11-20. IEEE, 2017. 5, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[55] Xianqi Wang, Gangwei Xu, Hao Jia, and Xin Yang. Selective-stereo: Adaptive frequency information selection" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20845" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "type": "text", + "content": "for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19701-19710, 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "text", + "content": "[56] Yan Wang, Zihang Lai, Gao Huang, Brian H Wang, Laurens Van Der Maaten, Mark Campbell, and Kilian Q Weinberger. Anytime stereo image depth estimation on mobile devices. In 2019 international conference on robotics and automation (ICRA), pages 5893-5900. IEEE, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 217 + ], + "type": "text", + "content": "[57] Yiran Wang, Min Shi, Jiaqi Li, Zihao Huang, Zhiguo Cao, Jianming Zhang, Ke Xian, and Guosheng Lin. Neural video depth stabilizer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9466-9476, 2023. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 295, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 295, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 295, + 296 + ], + "type": "text", + "content": "[58] Philippe Weinzaepfel, Thomas Lucas, Vincent Leroy, Yohann Cabon, Vaibhav Arora, Romain Brégier, Gabriela Csurka, Leonid Antsfeld, Boris Chidlovskii, and Jérôme Revaud. Croco v2: Improved cross-view completion pretraining for stereo matching and optical flow. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17969-17980, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 297, + 295, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 295, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 295, + 330 + ], + "type": "text", + "content": "[59] Bowen Wen, Matthew Trepte, Joseph Aribido, Jan Kautz, Orazio Gallo, and Stan Birchfield. Foundationstereo: Zeroshot stereo matching, 2025. 3, 4, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 295, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 295, + 376 + ], + "type": "text", + "content": "[60] Bin Xu, Yuhua Xu, Xiaoli Yang, Wei Jia, and Yulan Guo. Bilateral grid learning for stereo matching networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12497-12506, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 295, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 295, + 431 + ], + "type": "text", + "content": "[61] Gangwei Xu, Xianqi Wang, Xiaohuan Ding, and Xin Yang. Iterative geometry encoding volume for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21919-21928, 2023. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 433, + 295, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 295, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 295, + 476 + ], + "type": "text", + "content": "[62] Gangwei Xu, Yun Wang, Junda Cheng, Jinhui Tang, and Xin Yang. Accurate and efficient stereo matching via attention concatenation volume. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 477, + 295, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 477, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 477, + 295, + 510 + ], + "type": "text", + "content": "[63] Gangwei Xu, Huan Zhou, and Xin Yang. Cgi-stereo: Accurate and real-time stereo matching via context and geometry interaction. arXiv preprint arXiv:2301.02789, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 511, + 295, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 295, + 564 + ], + "type": "text", + "content": "[64] Gangwei Xu, Xianqi Wang, Zhaoxing Zhang, Junda Cheng, Chunyuan Liao, and Xin Yang. Igev++: Iterative multirange geometry encoding volumes for stereo matching. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 567, + 295, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 295, + 600 + ], + "type": "text", + "content": "[65] Haofei Xu and Juyong Zhang. Aanet: Adaptive aggregation network for efficient stereo matching. In CVPR, pages 1959-1968, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 601, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 295, + 645 + ], + "type": "text", + "content": "[66] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, Fisher Yu, Dacheng Tao, and Andreas Geiger. Unifying flow, stereo and depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 679 + ], + "type": "text", + "content": "[67] Gengshan Yang, Joshua Manela, Michael Happold, and Deva Ramanan. Hierarchical deep stereo matching on high-resolution images. In CVPR, pages 5515-5524, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 680, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 295, + 714 + ], + "type": "text", + "content": "[68] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 476 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 3, 6, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "text", + "content": "[69] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "type": "text", + "content": "[70] Jiaxi Zeng, Chengtang Yao, Yuwei Wu, and Yunde Jia. Temporally consistent stereo matching. In European Conference on Computer Vision, pages 341-359. Springer, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 164, + 553, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 553, + 197 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 553, + 197 + ], + "type": "text", + "content": "[71] Feihu Zhang, Victor Prisacariu, Ruigang Yang, and Philip HS Torr. Ga-net: Guided aggregation net for end-to-end stereo matching. In CVPR, pages 185-194, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 198, + 553, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 273 + ], + "type": "text", + "content": "[72] Jiawei Zhang, Xiang Wang, Xiao Bai, Chen Wang, Lei Huang, Yimin Chen, Lin Gu, Jun Zhou, Tatsuya Harada, and Edwin R Hancock. Revisiting domain generalized stereo matching networks from a feature consistency perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13001-13011, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 274, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 553, + 329 + ], + "type": "text", + "content": "[73] Youmin Zhang, Matteo Poggi, and Stefano Mattoccia. Temporalstereo: Efficient spatial-temporal stereo matching network. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 9528-9535. IEEE, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 331, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 385 + ], + "type": "text", + "content": "[74] Yongjian Zhang, Longguang Wang, Kunhong Li, Yun Wang, and Yulan Guo. Learning representations from foundation models for domain generalized stereo matching. In European Conference on Computer Vision, pages 146-162. Springer, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 387, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 387, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 387, + 553, + 430 + ], + "type": "text", + "content": "[75] Yiran Zhong, Hongdong Li, and Yuchao Dai. Open-world stereo video matching with deep rnN. In Proceedings of the European Conference on Computer Vision (ECCV), pages 101-116, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 431, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 431, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 431, + 553, + 476 + ], + "type": "text", + "content": "[76] Jingyi Zhou, Haoyu Zhang, Jiakang Yuan, Peng Ye, Tao Chen, Hao Jiang, Meiya Chen, and Yangyang Zhang. All-in-one: Transferring vision foundation models into stereo matching. arXiv preprint arXiv:2412.09912, 2024. 3" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20846" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Stochastic Gradient Estimation for Higher-Order Differentiable Rendering/95d3584a-ba8b-4c05-9f1b-c00fa69cfc88_content_list.json b/2025/Stochastic Gradient Estimation for Higher-Order Differentiable Rendering/95d3584a-ba8b-4c05-9f1b-c00fa69cfc88_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..82f43f8ab51a994f61b05d436afcb6a83b7c81b1 --- /dev/null +++ b/2025/Stochastic Gradient Estimation for Higher-Order Differentiable Rendering/95d3584a-ba8b-4c05-9f1b-c00fa69cfc88_content_list.json @@ -0,0 +1,1801 @@ +[ + { + "type": "text", + "text": "Stochastic Gradient Estimation for Higher-order Differentiable Rendering", + "text_level": 1, + "bbox": [ + 122, + 130, + 875, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zican Wang \nUniversity College London \nrobert.wang.19@ucl.ac.uk", + "bbox": [ + 151, + 181, + 370, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Michael Fischer Adobe Research mifischer@adobe.com", + "bbox": [ + 411, + 181, + 584, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tobias Ritschel \nUniversity College London \nt.ritschel@ucl.ac.uk", + "bbox": [ + 627, + 181, + 844, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We derive methods to compute higher-order differentials (Hessians and Hessian-vector products) of the rendering operator. Our approach is based on importance sampling of a convolution that represents the differentials of rendering parameters and shows to be applicable to both rasterization and path tracing. We further suggest an aggregate sampling strategy to importance-sample multiple dimensions of one convolution kernel simultaneously. We demonstrate that this information improves convergence when used in higher-order optimizers such as Newton or Conjugate Gradient relative to a gradient descent baseline in several inverse rendering tasks.", + "bbox": [ + 89, + 301, + 485, + 483 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 512, + 220, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inverse rendering is concerned with optimizing the parameters of a scene to minimize a loss. This can be a useful tool when the true parameters of the scene are unknown or nontrivial to set for a human expert, and therefore need to be inferred from observations or measurements. Examples for such scenarios are multi-view reconstruction [18] or the recovery of illumination and reflectance properties [38]. Differentiable rendering has recently become a popular tool for this optimization, as automatic differentiation (AD) frameworks have become more widespread.", + "bbox": [ + 89, + 535, + 483, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, differentiating the rendering process is far from trivial, as the rendering function [12] has zero and/or undefined gradients and step edges. Moreover, as many rendering operations rely on integration, discontinuities in the rendering pipeline cause problems for AD-engines, as we can no longer naively exchange the integral- and derivative-computations. Recent research therefore has designed a plethora of differentiable renderers [16], which compute gradients in various ways, typically incurring more [15, 20] or less [5, 7, 8] implementation and compute effort.", + "bbox": [ + 89, + 686, + 483, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A surprising insight is that, while these special renderers allow deriving gradients w.r.t. the scene parameters by differentiating (and back-propagating through) the rendering operator, virtually all of them are limited to first-order", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/844af3533c6928d02b7c6fcd9d6770b066ec87fcdf554887207bf9d89748cdb0.jpg", + "image_caption": [ + "Figure 1. Our approach allows sampling the Hessian for inverse rendering, here for the task of rotating the cup around its horizontal $(x_0)$ and vertical axes $(x_{1})$ . The estimated positive and negative gradients are shown in blue and red, respectively." + ], + "image_footnote": [], + "bbox": [ + 516, + 268, + 903, + 397 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "derivatives, and hence to expressing the gradient solely locally at a point in the parameter space. At the time of writing, no attempts to derive the Hessians required for higher-order inverse rendering methods have been published. However, decades of optimization research have shown the potential of higher-order methods in convergence and robustness [22].", + "bbox": [ + 511, + 476, + 906, + 582 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we argue that these benefits are also applicable to inverse rendering scenarios and show that they translate to net-gains in optimization time and performance in differentiable rendering. We tackle the Achilles' heel of higher-order optimizers - their increased per-iteration cost and computational / storage requirements - by developing efficient Monte Carlo (MC) estimators of the required higher-order quantities that can be importance-sampled with established techniques [23].", + "bbox": [ + 511, + 582, + 905, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our method leads to speedups of $2.71 \\times$ over previous methods, and net-gains in optimization time, speed and robustness while only assuming the rendering operator to be a black box that can be point-sampled.", + "bbox": [ + 511, + 717, + 905, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2. Previous Work", + "text_level": 1, + "bbox": [ + 511, + 787, + 663, + 803 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Gradient-based Optimization, the main workhorse of modern neural network training and inverse rendering, uses the gradient of the objective function to take iterative steps in the parameter space towards an improved solution until a (local) minimum is found. The exact nature of these steps varies with the optimizer that is being used [27].", + "bbox": [ + 511, + 809, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28198", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "First-order optimizers are simple to implement and cheap to execute but disregard important higher-order information about the shape of the objective function that could aid optimization, such as the second-order derivative. This usually comes at the expense of higher iteration counts, as many small steps are needed to converge to the solution. In contrast, higher-order optimizers incorporate information about the shape of the objective function, which usually allows them to take bigger steps in parameter space, leading to fewer iterations until convergence.", + "bbox": [ + 89, + 90, + 483, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For the specific case of second-order optimization, this additional information is often provided via the Hessian $\\mathsf{H}$ (or approximations thereof) and the Hessian-vector product (HVP). The Hessian contains the second-order derivatives of the objective function w.r.t. the optimization parameters, and can be interpreted as the curvature of the objective function. As such, it can be used to inform the optimizer about how quickly the current gradient is changing and thus, in turn, about how large the optimization step should be.", + "bbox": [ + 89, + 242, + 483, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Notation The following will use lowercase boldface to denote vectors and uppercase sans to denote matrices, respectively. Operators, formally defined as functions acting on functions, will be denoted in uppercase Roman lettering in order to avoid confusion with regular functions.", + "bbox": [ + 89, + 378, + 483, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Gradient descent should be familiar to most readers, so we here recall only its terminology: The first-order Taylor expansion of the cost function $f$ at position $\\theta$ is", + "bbox": [ + 89, + 474, + 483, + 520 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nf (\\boldsymbol {\\theta}) \\approx f (\\boldsymbol {\\theta} ^ {t}) + g (\\boldsymbol {\\theta} ^ {t}) ^ {\\top} (\\boldsymbol {\\theta} - \\boldsymbol {\\theta} ^ {t}),\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 531, + 398, + 549 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $g$ is the gradient $\\nabla f = \\mathrm{d}f / \\mathrm{d}\\theta$ of $f$ . The minimum is where the derivative is zero, which we can solve for as", + "bbox": [ + 89, + 561, + 482, + 590 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\mathrm {d}}{\\mathrm {d} \\boldsymbol {\\theta}} f (\\boldsymbol {\\theta}) = 0 \\approx g \\left(\\boldsymbol {\\theta} ^ {t}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 602, + 482, + 632 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As $f$ is a linear function only locally, we only make small steps with step size $\\gamma$ , by an update direction $-g(\\pmb{\\theta}^t)$ , as in", + "bbox": [ + 89, + 642, + 482, + 672 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\pmb {\\theta} ^ {t + 1} = \\pmb {\\theta} ^ {t} - \\gamma g (\\pmb {\\theta} ^ {t}).\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 683, + 357, + 702 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Newton's method is one of the most-used second-order optimizers. This term is derived from the second-order Taylor expansion of the objective $f$ around a point $\\theta$ :", + "bbox": [ + 89, + 720, + 483, + 766 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nf (\\pmb {\\theta}) \\approx f (\\pmb {\\theta} ^ {t}) + g (\\pmb {\\theta} ^ {t}) ^ {\\top} (\\pmb {\\theta} - \\pmb {\\theta} ^ {t}) + \\frac {1}{2} (\\pmb {\\theta} - \\pmb {\\theta} ^ {t}) ^ {\\top} \\mathsf {H} (\\pmb {\\theta} ^ {t}) (\\pmb {\\theta} - \\pmb {\\theta} ^ {t}),\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 777, + 483, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\mathsf{H}$ is $\\nabla^2 f = \\mathrm{d}^2 f / \\mathrm{d}^2\\pmb{\\theta}$ , the Hessian of $f$ .", + "bbox": [ + 89, + 816, + 401, + 833 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Ideally, we would like our update step to take us to an optimum. There, the derivative is necessarily zero:", + "bbox": [ + 89, + 834, + 482, + 864 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\mathrm {d}}{\\mathrm {d} \\boldsymbol {\\theta}} f (\\boldsymbol {\\theta}) = 0 \\approx g \\left(\\boldsymbol {\\theta} ^ {t}\\right) + \\mathrm {H} \\left(\\boldsymbol {\\theta} ^ {t}\\right) \\left(\\boldsymbol {\\theta} - \\boldsymbol {\\theta} ^ {t}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 875, + 483, + 906 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Solving for $\\theta$ then yields the update rule", + "bbox": [ + 513, + 90, + 782, + 106 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\theta} ^ {t + 1} = \\boldsymbol {\\theta} ^ {t} - \\gamma \\mathrm {H} ^ {- 1} (\\boldsymbol {\\theta} ^ {t}) g (\\boldsymbol {\\theta} ^ {t}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 114, + 903, + 132 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\mathbf{v} = -\\mathsf{H}^{-1}(\\pmb{\\theta}^t)g(\\pmb{\\theta}^t)$ is called the Newton direction.", + "bbox": [ + 511, + 141, + 901, + 157 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Newton's method requires the computation, storage, and inversion of the full Hessian, which quickly becomes a bottleneck in higher dimensions, as the Hessian for an $n$ -dimensional optimization problem is in $\\mathbb{R}^{n\\times n}$ .", + "bbox": [ + 511, + 157, + 905, + 218 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Newton Conjugate Gradient [31], upgrades Newton's method in two ways. First, it solves for Newton's direction iteratively as per the linear equation $\\mathsf{H}(\\pmb{\\theta})\\mathbf{v} = -g(\\pmb{\\theta})$ using conjugate directions [31]. Second, instead of an arbitrary step length $\\gamma$ , it also decides the scalar $\\alpha$ by which we move along this direction $\\mathbf{v}$ . To derive $\\alpha$ , first consider the Taylor expansion:", + "bbox": [ + 511, + 234, + 905, + 340 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\mathrm {d}}{\\mathrm {d} \\alpha} f (\\boldsymbol {\\theta} + \\alpha \\mathbf {v}) \\approx g (\\boldsymbol {\\theta}) ^ {\\top} \\mathbf {v} + \\alpha \\mathbf {v} ^ {\\top} H (\\boldsymbol {\\theta}) \\mathbf {v} = 0 \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 349, + 903, + 380 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "which can be re-arranged to", + "bbox": [ + 513, + 387, + 702, + 402 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha = - \\frac {g ^ {\\mathsf {T}} (\\boldsymbol {\\theta}) \\mathbf {v}}{\\mathbf {v} ^ {\\mathsf {T}} \\mathsf {H} (\\boldsymbol {\\theta}) \\mathbf {v}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 648, + 410, + 903, + 445 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The update rule for the Newton Conjugate Gradient is more involved: First, we maintain the direction $\\mathbf{v}$ and a residual $\\mathbf{r}$ , which are initialized to be the gradient at the initial positions:", + "bbox": [ + 511, + 452, + 905, + 508 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} ^ {0} = \\mathbf {r} ^ {0} = - g (\\boldsymbol {\\theta} ^ {t}) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 511, + 903, + 529 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We then find the $\\alpha$ by equation Eq. 5 and update the point:", + "bbox": [ + 513, + 534, + 900, + 549 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\theta} ^ {t + 1} = \\boldsymbol {\\theta} ^ {t} + \\alpha \\mathbf {v} ^ {t}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 645, + 556, + 903, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The next direction is chosen by updating the residual $\\mathbf{r}$ and computing a new conjugate search direction $\\mathbf{v}$ with the Fletcher-Reeves formula [9]:", + "bbox": [ + 511, + 585, + 905, + 630 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {r} ^ {t + 1} = \\mathbf {r} ^ {t} - \\alpha \\mathsf {H} (\\boldsymbol {\\theta} ^ {t}) \\mathbf {v} ^ {t} \\\\ \\beta = \\frac {\\mathbf {r} ^ {t + 1 , \\top} \\mathbf {r} ^ {t + 1}}{\\mathbf {r} ^ {t , \\top} \\mathbf {r} ^ {t}} \\\\ \\mathbf {v} ^ {t + 1} = \\mathbf {r} ^ {t + 1} + \\beta \\mathbf {v} ^ {t}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 638, + 787, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since Newton's method is derived via second-order Taylor expansion (approximated by a hyper-parabola), the derived Hessian may lead to either a maximum or a minimum depending on the local curvature of the function being approximated. This is determined by the positive-definiteness of the Hessian, and a negative-definite Hessian will diverge from the gradient's descent direction. To avoid this, one can ensure that the Hessian is always positive semi-definite using the Hessian modification technique and ensure step truncation to a trust region [22], $\\alpha = \\min (\\alpha ,\\Delta \\| \\mathbf{v}\\|^{-1})$ , where $\\Delta$ is the trust region radius (see Supplemental Sec. H).", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28199", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hessian-vector product approximations go one step further by entirely avoiding to compute $\\mathsf{H}$ when producing Hv. Pearlmutter [24] and Werbos [35] discuss different options to do so, but a simple option is central differences", + "bbox": [ + 89, + 90, + 483, + 154 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathsf {H} (\\pmb {\\theta}) \\mathbf {v} \\approx \\lim _ {\\varepsilon \\to 0} \\frac {g (\\pmb {\\theta} + \\varepsilon \\mathbf {v}) - g (\\pmb {\\theta} - \\varepsilon \\mathbf {v})}{2 \\varepsilon}. \\qquad (8)\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 167, + 483, + 200 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hessians in vision and graphics have been applied to several optimization tasks, for instance, for total-variation denoising [3], texture parameterization and surface mappings with approximated Hessians [28], 3D shape manipulation [13], and image deconvolution tasks via (Hessian-free) Newton methods [14]. For optical flow estimation, both Zach et al. [39] and Werlberger et al. [36] use second-order information. Additionally, several works have explored Hessian-based algorithms in machine learning for vision tasks. Yao et al. [37] introduces a Hessian-based pruning method for CNNs for image classification, while Ramesh et al. [26] utilizes Hessians to improve the training of image generation models and Desai et al. [6] introduces an algorithm that derives Hessians for C code.", + "bbox": [ + 89, + 229, + 483, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hessians in inverse rendering, however, have received surprisingly little attention, potentially due to their additional implementation overhead and computational complexity. In addition to the fact that conventional AD systems are mostly designed for first-order gradient computations, calculating the second-order information requires storing the whole forward- and first-derivative-graph in memory, which can lead to exponential memory growth. Nicolet et al. [19] are the closest to our work by approximating second-order steps for mesh reconstruction. However, they set the Hessian to the identity to avoid computational expense and instead work with a Laplacian regularizer, which works for their formulation and the case of mesh optimization, but is unclear how to translate to general problems.", + "bbox": [ + 89, + 444, + 483, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For derivative-free gradient estimators, it is equally unclear how second-order information would be computed. The zeroth-order estimators simultaneous perturbation stochastic approximation (SPSA) [32] and finite differences (FD) estimate a first-order gradient, whose second-order derivative naturally is zero. Extending these estimators to second-order information requires prohibitive amounts of function evaluations. ZeroGrads [8], which learns a neural network that fits the cost landscape, uses ReLU non-linearities, whose second-order derivative equally decays to zero. Covariance adaptation evolution strategy (CMA-ES) is often used when typical second-order optimization fails to converge [10]. It ensures that the Hessian is always positive definite, but if second-order derivative-based methods are successful, they are usually faster than CMA-ES.", + "bbox": [ + 89, + 659, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Our approach", + "text_level": 1, + "bbox": [ + 513, + 90, + 658, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first describe the computation of gradients using importance sampling of a combined gradient-smoothing operator from previous work [4, 7] which we then extend to Hessianians, and in a next step to Hessian-vector products.", + "bbox": [ + 511, + 112, + 906, + 174 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Background", + "text_level": 1, + "bbox": [ + 513, + 183, + 645, + 199 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Rendering equation The rendering equation (RE) [12] describes the radiance $L$ leaving a point $\\mathbf{x}$ in the scene into the direction $\\omega_{\\mathrm{o}}$ as", + "bbox": [ + 511, + 202, + 906, + 247 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL (\\mathbf {x}, \\omega_ {\\mathrm {o}}; \\boldsymbol {\\theta}) = \\int_ {\\Omega} \\underbrace {f _ {\\mathrm {r}} \\left(\\omega_ {\\mathrm {i}} , \\omega_ {\\mathrm {o}}\\right) L \\left(\\mathbf {y} , \\omega_ {\\mathrm {i}} ; \\boldsymbol {\\theta}\\right)} _ {R \\left(\\omega_ {\\mathrm {i}}; \\boldsymbol {\\theta}\\right)} \\mathrm {d} \\omega_ {\\mathrm {i}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 258, + 906, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta$ are the parameters of the scene we would like to optimize, such as object geometry, reflectance, or light emission. This integral over $\\Omega$ , i.e., all incoming $\\omega_{\\mathrm{i}}$ directions that multiply the radiance field arriving from that direction from the closest other point $\\mathbf{y}$ in direction $\\omega_{\\mathrm{i}}$ with the bidirectional reflectance distribution function (BRDF) $f_{\\mathrm{r}}$ , has no analytical closed-form solution, and hence usually is approximated - both in forward and inverse rendering - via MC methods. We will shorthand the entire integrand as $R$ .", + "bbox": [ + 511, + 313, + 906, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem statement We would now like to apply a differential operator $\\mathrm{D}$ to the RE, as in", + "bbox": [ + 511, + 470, + 905, + 501 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {D} L (\\mathbf {x}, \\omega_ {\\mathrm {i}}; \\boldsymbol {\\theta}). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 660, + 513, + 906, + 531 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "If $\\mathrm{D}$ was the gradient operator $\\frac{\\partial L}{\\partial \\theta}$ , this would be differentiable rendering, for other operators, this becomes higher-order differentiable rendering.", + "bbox": [ + 511, + 542, + 905, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The trouble is that we cannot move D, be it gradient or higher-order, into the integral, as in many cases (e.g., for BRDF or spatial derivatives), the integrand is discontinuous in $\\theta$ , so", + "bbox": [ + 511, + 589, + 906, + 648 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {D} L (\\mathbf {x}, \\omega_ {\\mathrm {i}}; \\boldsymbol {\\theta}) \\neq \\int_ {\\Omega} \\mathrm {D} R (\\omega_ {\\mathrm {i}}; \\boldsymbol {\\theta}) \\mathrm {d} \\omega_ {\\mathrm {i}}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 660, + 906, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, the right-hand side of the above expression is exactly the quantity that naively-applied AD computes [21, 34], leading to wrong gradients in (any-order) differentiable rendering.", + "bbox": [ + 511, + 703, + 906, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Solution The idea is to enforce the property that prevents differentiation - smoothness -, so that we actually can differentiate. To that end, assume a further linear operator $S$ that is smoothing any function in $\\theta$ . This provides a smooth rendering equation $\\bar{L}$", + "bbox": [ + 511, + 785, + 906, + 861 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {L} (\\mathbf {x}, \\omega_ {\\mathrm {o}}; \\boldsymbol {\\theta}) = \\int_ {\\Omega} \\mathrm {S} R \\left(\\omega_ {\\mathrm {i}}; \\boldsymbol {\\theta}\\right) \\mathrm {d} \\omega_ {\\mathrm {i}}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 872, + 906, + 905 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28200", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a9ec933250838921e9b5c001d23d67037c74810a16b7660bd6c2dc79676bc363.jpg", + "image_caption": [ + "Figure 2. Comparison of classic (yellow) and smooth (red) gradients, as well as our Hessians (blue) on an inverse rendering problem to change the initial parameter $\\theta$ so that the left triangle overlaps the right one. Classic gradients are zero almost everywhere (plateaus) except where the triangles already overlap. These methods do not converge. Smooth gradients point into the right direction, but make steps far from the optimum (dotted line). An update taking into account the curvature of the loss landscape lands at a point very close to the target. We incorporate this curvature information via our Hessians." + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 911, + 193 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Smoothing can be achieved by convolution, so for any function $f$ in any dimension", + "bbox": [ + 89, + 292, + 482, + 321 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {S} f (\\boldsymbol {\\theta}) = \\int_ {\\Theta} \\kappa (\\tau) f (\\boldsymbol {\\theta} - \\tau) \\mathrm {d} \\tau , \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 325, + 480, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\kappa$ is a smoothing kernel, such as a Gaussian, which we use in this work. This convolved integrand is now smooth, which, according to Leibniz' rule, allows us to move the differential operator into the integral", + "bbox": [ + 89, + 362, + 483, + 422 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {D} \\bar {L} (\\mathbf {x}, \\omega_ {\\mathrm {o}}; \\boldsymbol {\\theta}) = \\int_ {\\Omega} \\mathrm {D S} R \\left(\\omega_ {\\mathrm {i}}; \\boldsymbol {\\theta}\\right) \\mathrm {d} \\omega_ {\\mathrm {i}}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 426, + 482, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which, after rearranging and expanding, yields an integral that can be approximated via MC:", + "bbox": [ + 89, + 463, + 483, + 494 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {D} \\bar {L} (\\mathbf {x}, \\omega_ {\\mathrm {o}}; \\boldsymbol {\\theta}) = \\int_ {\\Omega} \\int_ {\\Theta} \\mathrm {D} \\kappa (\\boldsymbol {\\tau}) R \\left(\\omega_ {\\mathrm {i}}; \\boldsymbol {\\theta} - \\boldsymbol {\\tau}\\right) \\mathrm {d} \\boldsymbol {\\tau} \\mathrm {d} \\omega_ {\\mathrm {i}}. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 498, + 482, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MC here means to take random samples from the product space of light paths and scene parameters. This works best if we can importance-sample for the integrand. The integrand here is a product of four terms. Sampling for the incoming radiance and BRDF terms has been investigated in the rendering community [33] and is not our consideration here, so we simply adopt these strategies. Sampling for the application of the differential operator to the smoothing kernel is the essence of the problem at hand. Depending on the choice of differential operator, we will derive three sampling strategies for the three resulting estimators next.", + "bbox": [ + 89, + 534, + 483, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Conclusion In conclusion, to perform efficient and practical any-order differentiation of the RE, we would need to implement two functions: first, a convolution kernel that combines smoothing and the desired differentiation, and second, a function to sample from that kernel for importance sampling. We will now do so for the gradient (Sec. 3.2), Hessian (Sec. 3.3) and HVP (Sec. 3.4).", + "bbox": [ + 89, + 717, + 483, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Gradients", + "text_level": 1, + "bbox": [ + 89, + 830, + 205, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Operator: For first-order gradient descent, Fischer and Ritschel [7] have differentiated using the gradient operator", + "bbox": [ + 89, + 848, + 482, + 878 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {D} ^ {\\mathrm {G}} = \\nabla = \\partial / \\partial \\tau_ {i} \\in (\\mathbb {R} ^ {n} \\rightarrow \\mathbb {R}) \\rightarrow (\\mathbb {R} ^ {n} \\rightarrow \\mathbb {R} ^ {n}), \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 883, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/06540bd316991f2321951872ad779f1dceafe2acf26d1847f48d812f8fd2268e.jpg", + "image_caption": [ + "Figure 3. Derivation of smooth differentiation by convolution for three differential operators (rows) involves three steps (columns): Defining the operator to combine smoothing and differentiation (1st col.), positivization and normalization to become a probability density function (PDF) (2nd col.), creating an inverse mapping (3rd col.), which finally allows sampling (4th col.)." + ], + "image_footnote": [], + "bbox": [ + 517, + 290, + 903, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which maps a scalar function of $n$ dimensions to an $n$ -dimensional gradient vector field.", + "bbox": [ + 511, + 571, + 905, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The combination of smoothing and differentiation is", + "bbox": [ + 532, + 601, + 879, + 616 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {D} ^ {\\mathrm {G}} \\kappa (\\boldsymbol {\\tau}) = \\nabla_ {i} \\mathcal {N} (\\boldsymbol {\\tau}, \\sigma) = - \\frac {\\tau_ {i}}{\\sigma^ {2}} \\cdot \\mathcal {N} (\\boldsymbol {\\tau}, \\sigma). \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 622, + 903, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For a derivation, please see the supplemental, Sec. A.", + "bbox": [ + 532, + 654, + 885, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sampling: For sampling a one-dimensional Gaussian gradient, we can use inverse transform sampling via the Smirnov transform [7]. To this end, the integrand has to be a PDF, i.e., positive and integrating-to-1 [23]. Subsequently, we compute the integral of a positivized version of $p^{\\mathrm{G}}$ , the cumulative distribution function (CDF) $P^{\\mathrm{G}}$ , and invert it as", + "bbox": [ + 511, + 669, + 905, + 758 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP ^ {\\mathrm {G}, - 1} (\\xi) = \\left\\{ \\begin{array}{l l} - \\sqrt {2 \\sigma^ {2} \\log (2 \\xi)} & \\text {i f} \\xi \\leq 0. 5, \\\\ + \\sqrt {2 \\sigma^ {2} \\log (1 - \\xi)} & \\text {e l s e .} \\end{array} \\right. \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 777, + 903, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This derivation applies to the dimension which is being differentiated. The separability of multi-dimensional Gaussians ensures that the other dimensions remain a Gaussian distribution, and we can sample these dimensions independently. For a derivation of this, see Suppl. Sec. B.", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28201", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Hessians", + "text_level": 1, + "bbox": [ + 91, + 90, + 194, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Operator: The differential operator for Hessian is", + "bbox": [ + 89, + 109, + 447, + 125 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {D} ^ {\\mathrm {H}} = \\nabla^ {2} = \\partial^ {2} / \\partial \\tau_ {i} \\partial \\tau_ {j} \\in (\\mathbb {R} ^ {n} \\rightarrow \\mathbb {R}) \\rightarrow (\\mathbb {R} ^ {n} \\rightarrow \\mathbb {R} ^ {n \\times n}), \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 131, + 480, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which maps a scalar function in $n$ dimensions to its $n \\times n$ -element Hessian field. The combination of second-order derivatives and Gaussian smoothing is", + "bbox": [ + 89, + 164, + 482, + 209 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {D} ^ {\\mathrm {H}} \\kappa_ {i, j} (\\boldsymbol {\\tau}) = \\left\\{ \\begin{array}{l l} \\left(- \\frac {1}{\\sigma^ {2}} + \\frac {\\tau_ {i} ^ {2}}{\\sigma^ {4}}\\right) \\cdot \\mathcal {N} (\\boldsymbol {\\tau}, \\sigma) & \\text {i f} i = j, \\\\ \\frac {\\tau_ {i} \\tau_ {j}}{\\sigma^ {4}} \\cdot \\mathcal {N} (\\boldsymbol {\\tau}, \\sigma) & \\text {e l s e .} \\end{array} \\right. \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 215, + 480, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For a derivation, see supplemental Sec. C.", + "bbox": [ + 89, + 291, + 369, + 305 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sampling: We first need to positivize as the function is signed, then scale the function so that it is a valid distribution. This is done differently for diagonal and off-diagonal elements. For the diagonal case, we construct the CDF of a second-order derivative $P_{ii}^{\\mathrm{H}}$ of the 2D Gaussian:", + "bbox": [ + 89, + 306, + 483, + 381 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {i i} ^ {\\mathrm {H}} \\left(\\tau_ {i} = u\\right) = \\left\\{ \\begin{array}{c l} - \\frac {u}{4 \\sigma} \\exp \\left(\\frac {1}{2} - \\frac {u ^ {2}}{2 \\sigma^ {2}}\\right) & \\text {i f} u < - \\sigma , \\\\ 0. 5 + \\frac {u}{4 \\sigma} \\exp \\left(\\frac {1}{2} - \\frac {u ^ {2}}{2 \\sigma^ {2}}\\right) & \\text {i f} u \\in [ - \\sigma , \\sigma ] \\\\ 1 - \\frac {u}{4 \\sigma} \\exp \\left(\\frac {1}{2} - \\frac {u ^ {2}}{2 \\sigma^ {2}}\\right) & \\text {i f} u > \\sigma . \\end{array} \\right. \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 388, + 483, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "See supplemental Sec. D for a derivation of this result. This is a transcendental equation whose inverse CDF cannot be expressed in closed form via elementary functions [2]. Instead, we 1D-tabulate its values on the range $10\\sigma$ for every $i$ . The inverse function is found by searching in that range, and storing the pre-sorted inverse indices for $O(1)$ access.", + "bbox": [ + 89, + 507, + 482, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The off-diagonals are the product of two partial gradients that we have already derived in Eq. 17. Fortunately, the product of two independent distributions can be sampled by sampling each one independently. Thus, no extra derivation is required here. For sampling in higher dimensions, similar to the gradient sampler in Sec. B, the rest of the dimensions are sampled from a Gaussian distribution.", + "bbox": [ + 89, + 598, + 482, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In all methods, we exploit the symmetry in the Hessian matrix when sampling and estimating.", + "bbox": [ + 89, + 703, + 482, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Hessian-vector product", + "text_level": 1, + "bbox": [ + 89, + 742, + 305, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Operator: As explained in Sec. 2, avoiding to store the full Hessian is possible by using HVPs:", + "bbox": [ + 89, + 760, + 482, + 791 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {D} ^ {\\mathrm {H V P}} = \\nabla^ {2} \\mathbf {v} = \\partial^ {2} / \\partial \\tau_ {i} \\partial \\boldsymbol {\\tau} \\cdot \\partial / \\partial \\boldsymbol {\\tau}. \\tag {22}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 797, + 482, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In essence, a HVP is the directional gradient of the gradient. As there is one direction and the gradient has $n$ dimensions, the HVP is an $n$ -dimensional vector, too. In contrast, a Hessian is the non-directional gradient of the gradient, i.e., the gradient along all $n$ dimensions, and as such in $\\mathbb{R}^{n \\times n}$ .", + "bbox": [ + 89, + 824, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sampling: To sample the smooth HVP, we simply need the directional central differences of an estimator of gradients, which we already have. So the estimator is the difference of two first-order estimators, evaluated at positions shifted from the current solution along the gradient direction [24, 35]:", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {D} ^ {\\mathrm {H V P}} \\kappa (\\boldsymbol {\\tau}) = \\frac {\\mathrm {D} ^ {\\mathrm {G}} \\kappa (\\boldsymbol {\\tau} - \\varepsilon \\mathbf {v}) - \\mathrm {D} ^ {\\mathrm {G}} \\kappa (\\boldsymbol {\\tau} + \\varepsilon \\mathbf {v})}{2 \\varepsilon}. \\tag {23}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 190, + 905, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Aggregate", + "text_level": 1, + "bbox": [ + 513, + 233, + 629, + 248 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We summarize the key property of all samplers - the number of function evaluations required - in Tab. 1. Function evaluations require execution of the black-box rendering engine, the most costly part of the optimization, and hence should be minimized.", + "bbox": [ + 511, + 251, + 905, + 325 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a062daaca0cace5af8289af4c0d81515d7138d70e5eb98d8244349ea8705b7cf.jpg", + "table_caption": [ + "Table 1. Time complexity of different estimator variants. AIS: aggregated importance sampling." + ], + "table_footnote": [], + "table_body": "
no ISISAIS
Fin. Diff.nnn
FR22 [7]111
Our Grad.1n1
Our Hess.1n21
Our HVP1n1
", + "bbox": [ + 514, + 397, + 728, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Classic finite differences take opposing samples in each dimension and hence, for an $n$ -dimensional problem, require $2n$ function evaluations for a single gradient sample. All other methods (second row in Tab. 1 onward) are based on MC, so we can get a gradient estimate using", + "bbox": [ + 743, + 327, + 906, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "a fixed number of M function evaluations. This, however, comes at the cost of variance, which can be reduced with importance sampling, but that is done in all dimensions of the differential quantity independently, so it requires as many evaluations as these have elements (second column in Tab. 1). For a Hessian, this can be substantial, $n^2$ .", + "bbox": [ + 511, + 523, + 905, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As a compromise between low number of function evaluations and low variance, we propose \"aggregate\" importance sampling of the differential quantity (Fig. 4): instead of importance-sampling each dimension optimally but individually, we importance-sample w.r.t. the average of the convolution across all dimensions (combining several proposals by sampling from a mixture is similar to MIS using the balance heuristic [33]). This average convolution is a single function again, and with our combined sampling strategy can again be sampled with one function call per iteration (last column of Tab. 1). We implement this in two simple steps: first we randomly decide which element of the differential representation to choose (e.g., which matrix element out of $n^2$ in the Hessian), and then we compute the convolution sample for all $n^2$ but with the same function value that is only evaluated once. This increases variance, as the average kernel is not identical to the individual kernels but shares many properties, e.g., they are all zero in value at position 0. As a linear combination of unbiased es", + "bbox": [ + 511, + 614, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28202", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6345cca08a39cb8e61a8c87da21843ed9cdfbf2dad76a754eb918008183c771f.jpg", + "image_caption": [ + "Figure 4. Aggregate and non-aggregate sampling for a 2D optimization space and, consequently, a $2 \\times 2$ Hessian: Without aggregation, importance sampling is done for each element of the Hessian independently (four points at different positions in each kernel), leading to four rendering calls and four different mugs. Each value is then weighted by the kernel (blue and red colors denote positive and negative). In aggregate sampling, we importance-sample the average of all four stencils, resulting in a single sample location and, hence, four times the same mug that can be rendered in one call, weighted with four different kernels." + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 480, + 239 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "timators, this is still unbiased. We show in Sec. J that this pays off. Note that Tab. 1 is both time and space complexity, except for the aggregate importance sampling of the HVP, where the space complexity stays $O(n)$ while time complexity is reduced to $O(1)$ .", + "bbox": [ + 89, + 402, + 483, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Evaluation", + "text_level": 1, + "bbox": [ + 89, + 489, + 207, + 505 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Methods", + "text_level": 1, + "bbox": [ + 89, + 510, + 194, + 525 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our evaluation compares the performance of gradients (input to gradient descent), Hessians and non-aggregate as well as aggregate Hessian-vector product (both input to higher-order optimizers).", + "bbox": [ + 89, + 529, + 483, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines An established inverse rendering solution is Mitsuba with gradient descent. FR22 is using gradient descent with Fischer and Ritschel [7] MC gradients. OursG, OursH, OursHVP and OursHVPA are our approach for gradient, Hessians and Hessian-vector products, and aggregates, respectively. OursG is used with gradient descent (GD), the others in combination with conjugate gradient (CG). We have also experimented with LBFGS as detailed in the supplemental, but do not report it here as it almost never converges and never at competitive speed. All methods use the Adam optimizer.", + "bbox": [ + 89, + 609, + 483, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tasks We tackle artificial problems with known analytic Hessians and Hessian-vector products as well as real inverse rendering tasks. As a simple first test case, we optimize the smooth, quadratic potential $(\\mathrm{QUAD}^1)$ function $ax_0^2 +bx_1^2 +cx_0x_1$ in $\\mathbb{R}^2$ , with the fixed variables $a = 5,b = 5,c = 7.5$ .", + "bbox": [ + 89, + 794, + 482, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Second, we optimize the classic plateau-demonstration task from [7]: the 2D position of some boxes is optimized", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to match a reference. This is already a much harder task, as there is a plateau in the cost landscape when the squares do not overlap (almost always the case in the initial configuration), as the image-space error does not change. We study the case of one $(\\mathrm{Box2}^3)$ and five $(\\mathrm{Box10}^5)$ squares, with two and ten dimensions to optimize, respectively.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For real inverse problems, we study the optimization of reflectance, light, and geometry and render using Mitsuba. In the MUG7 task, we optimize the vertical rotation of a coffee cup such that it matches a reference. In the SHADOW9 task, we optimize the position of a sphere that is unobserved in the rendered image, such that the shadow it casts matches a reference shadow. In the BUNNY11 task, the $x$ and $z$ position and the rotation around the $z$ axis of the Stanford bunny are optimized. This task is set up specifically for Mitsuba to converge12, with no plateaus in the loss function.", + "bbox": [ + 511, + 181, + 906, + 332 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To test the scaling with dimensionality, we also optimize the pixels of a $32 \\times 32$ texture (TEXTURE $^{13}$ ), with 1024 parameters. In addition, we have used our approach on a mesh optimization task, where we fit the 3D position of the vertices of a tessellated sphere to images of a target object. For SUZANNE, we fit 2,562 vertices, for BANANA, we fit the 2,737 vertices and their RGB vertex colors, similar to the DTU setup [11]. For trajectories of the optimization and the outcome, see Fig. 5 top and bottom rows, respectively.", + "bbox": [ + 511, + 333, + 905, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, we learn a $\\mathrm{CNN}^{15}$ to predict scene parameters from images using inverse rendering without scene parameter labels. In this case, the optimization is learning in the stricter sense: we tune parameters of a deep architecture instead of scene parameters directly. The CNN takes as input an image of a mug and produces as output the orientation for a mug. As a loss, these parameters are inserted into a renderer, differentiated by different approaches and compared to a target. The CNN has 267745 parameters, learned by differentiating through an image loss and rendering of the scene given the estimated parameters. Another CNN $(\\mathrm{CNN}5^{19})$ is trained to predict the rotation, position, and color of the mug with 268773 parameters.", + "bbox": [ + 511, + 469, + 905, + 665 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Differentials of this function are computed as detailed in supplemental Sec. E, where we derive a grey-box approach that samples only the black-box part (the rendering) and combines its differentials with analytic differentials for the white-box part (the CNN). Importantly, this idea works for gradients, as well as for higher-order differentials.", + "bbox": [ + 511, + 666, + 905, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics Our main measure of success is convergence speed in wall-clock and parameter difference to the true parameters (which we know in all inverse rendering tasks as we created the scenes). A secondary measure of success is the image difference, as during optimization, we consider the parameters hidden. All metrics are evaluated across an ensemble of 20 runs averaged across 10 steps in time. In convergence plots, the ensemble median is shown at every point in time, averaged across 20 time steps.", + "bbox": [ + 511, + 763, + 905, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28203", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f12aa9828642080a6086462d1e0ef531c1774924acb659650dbc30a654c59bb8.jpg", + "table_caption": [ + "Table 2. Quantitative results of different methods on different tasks (rows) and their convergence plots. We report convergence time in wall-clock units, in ratio to the overall best method, OurHVPA. In the numerical columns,.9 and .99 report the time taken to achieve 90 and $99\\%$ error reduction from the initial starting configuration, respectively, while the bar plots graphically show these findings. The line plots report image- and parameter-space convergence in the left and right column, respectively, on a log-log scale." + ], + "table_footnote": [], + "table_body": "
TaskMethodImage errorParameter errorMitsuba FR22 CMA-ES OurG OurH OurHVP OurHVPA
.9.99.999.9.99.999
QUAD1Mitsu2—————ErrorLog error
FR221.802.382.982.203.173.90
CMA-ES0.150.130.140.130.140.23
OurG3.985.136.024.726.427.77
OurH1.171.171.151.181.161.17
OurHVP0.760.620.620.630.590.60
OurHVPA1.001.001.001.001.001.00
BOX23Mitsu——————ErrorLog error
FR223.242.492.572.782.892.70
CMA-ES5.513.453.158.135.514.13
OurG2.882.162.212.862.652.38
OurH0.930.951.080.760.841.84
OurHVP0.810.640.7040.740.720.72
OurHVPA1.001.001.001.001.001.00
BOX105Mitsu——————ErrorLog error
FR221.491.201.711.801.541.29
CMA-ES3.102.271.734.523.232.48
OurG2.181.771.532.532.261.88
OurH5.975.334.685.485.965.45
OurHVP1.511.441.6961.261.501.58
OurHVPA1.001.001.001.001.001.00
MUG7Mitsu——————ErrorLog error
FR225.925.695.463.286.536.29
CMA-ES2.951.971.823.322.382.20
OurG3.993.913.742.364.424.30
OurH1.571.091.531.001.081.20
OurHVP1.151.141.181.140.871.05
OurHVPA81.001.001.001.001.001.00
SHAD9Mitsu——————ErrorLog error
FR221.941.401.551.891.141.71
CMA-ES1.961.151.034.382.201.66
OurG1.371.181.132.381.471.34
OurH1.671.441.662.321.741.46
OurHVP1.140.910.961.371.131.07
OurHVPA1.001.001.001.001.001.00
BUNNY11Mitsu120.901.291.461.871.381.28ErrorLog error
FR220.750.751.120.780.630.71
CMA-ES1.160.950.851.270.910.85
OurG0.760.891.070.690.640.76
OurH1.291.231.291.180.801.07
OurHVP0.680.680.700.630.440.58
OurHVPA1.001.001.001.001.001.00
TEXTURE13Mitsu——————ErrorLog error
FR224.042.271.804.042.271.80
CMA-ES7.084.604.3511.167.477.44
OurG3.032.862.823.032.862.82
OurH440.1239.78—40.1239.78—
OurHVP4.014.004.004.014.004.00
OurHVPA1.001.001.001.001.001.00
CNN15Mitsu——————ErrorLog error
FR222.1116——1.451.322.71
CMA-ES———2.08——
OurG1.14——0.960.832.28
OurH0.770.77—0.740.770.74
OurHVP0.870.9317—0.880.910.87
OurHVPA1.001.00—181.001.001.00
CNN519Mitsu——————ErrorLog error
FR22——————
CMA-ES——————
OurG18.68——3.70——
OurH19.00——5.79——
OurHVP3.64201.46—1.491.58—
OurHVPA1.001.00—1.001.001.001.00
", + "bbox": [ + 93, + 157, + 937, + 890 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": ".9 .99 .999 .9 99 .999", + "bbox": [ + 550, + 892, + 702, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Log time", + "bbox": [ + 759, + 893, + 799, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Log time", + "bbox": [ + 852, + 893, + 892, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28204", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f82b9bb7c56e627efb356317d13685da7bda43592231eca79b1a52ee372358b3.jpg", + "image_caption": [ + "Figure 5. Inverse mesh optimization from renderings. Gradients use FR22, Hessians use OursHVPA, CMA-ES uses the pycma package." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 903, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Results", + "text_level": 1, + "bbox": [ + 89, + 252, + 184, + 266 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The main results of our evaluation are summarize in Tab. 2 where empty cells did not converge or the method is not applicable to that task. On average across tasks and methods, our premiere method oursHVPA speeds up the convergence by a factor of 2.71.", + "bbox": [ + 89, + 272, + 482, + 348 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In general, our methods, in particular OursHVP and OursHVPA, lead the level of error reduction across all time budgets, as seen by comparing the convergence curves vertically where they consistently decrease fastest.", + "bbox": [ + 89, + 349, + 483, + 410 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We benefit most in the artificial QUAD task, but MUG and SHADOW are both real rendering problems, where we are around seven times and $60\\%$ faster, respectively. We see that OurG typically cannot outperform FR22, which uses an approximation of the correct high-dimensional gradient kernel. Doing it \"right\" only pays off when going to higher order. We also note that OursHVP outperforms OursH due to less function evaluations (as the metric is wall-clock - the iteration count for both is close). The log-log plots confirm that the higher-order variants enabled by our approach are much faster, but also converge slightly less stable.", + "bbox": [ + 88, + 411, + 483, + 578 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While tasks like QUAD and Box2 are not applicable to inverse renderers like Mitsuba², our methods treat any loss function as a black-box and retrieve derivatives via sampling. Comparing the performance of OurHVP and OurHVPA across a task with two and ten dimensions, we find the expected relation: OurHVP is faster in low dimensions⁴ than in higher dimensions for a similar task⁶.", + "bbox": [ + 89, + 580, + 482, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We notice that the OurHVPA is not out-performing the OurHVP and OurH in the MUG task. This is because there is a single scene parameter for this task, thus, the higher order derivative is all in order of one. With a higher variance for OurHVPA, it could be a little slower. The effect of using the aggregate starts to show from SHAD, where OurHVPA reaches the 1:1000 convergence the fastest10.", + "bbox": [ + 89, + 686, + 482, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For all the rendering tasks, only the BUNNY task converges for Mitsuba12. Since the other tasks have plateaus in the loss function, the analytical gradient from inverse rendering is effectively zero, so it would be hard to converge within a reasonable time. The log-log plot shows in addition that derivatives from the inverse rendering were slower than the sampled ones. On the contrary, our methods not", + "bbox": [ + 89, + 795, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "only smooth out the plateau, but also improve the speed of the derivative computation.", + "bbox": [ + 511, + 253, + 906, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The drawback of constructing full Hessians is seen in the TEXTURE task, where OurH is more than 40 times slower than the HVP due to the matrix size $^{14}$ . For the most demanding task, CNN and CNN5, and even our best method $^{18}$ is not able to reduce the error below $1:1000$ . At the $1:1000$ -level, the more elaborate differentials $^{17}$ is not faster than the basic one, but faster than first-order $^{16}$ . Finally, the advantage of our ability to compute HVPs is seen when comparing CNN and CNN5: while a CNN producing a single parameter is doable with Hessians $^{17}$ , but if we go to higher dimensions, the aggregate strategy pays off again and is markedly faster $^{20}$ .", + "bbox": [ + 511, + 284, + 906, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In conclusion, albeit convergence varies with problem characteristics and dimensionality, we find that our best method, OurHVPA converges most reliably, beating the other algorithms by an average factor of $2.71 \\times$ in wall-clock units. We show examples of task initialization and outcome, and additional evaluation in the supplemental.", + "bbox": [ + 511, + 467, + 905, + 558 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Limitations and Conclusion", + "text_level": 1, + "bbox": [ + 511, + 573, + 769, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. From the above experiments, it becomes apparent that our method performs well on lower-dimensional examples, but that its performance delta lessens with increasing problem dimensionality (see e.g. TEXTURE). We hypothesize that this is due to sparser sampling in higher dimensions and exacerbated MC noise and acknowledge that further research in this direction is needed. We include an analysis on how our method degrades under MC rendering noise in the supplemental.", + "bbox": [ + 511, + 597, + 905, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusion. The availability of second-order gradient information gives rise to an exciting avenue of future research in inverse rendering. While these algorithms are not new, their use in vision & graphics remains limited due to the convenience and widespread adoption of first-order gradient descent and the increased per-iteration cost that is usually associated with higher-order methods. In this work, our efficient estimators have shown the latter to be negligible in real-world optimization scenarios. We hope that this will inspire future research into unbiased, low-variance estimators of higher-order optimization methods.", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28205", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Albert S Berahas, Richard H Byrd, and Jorge Nocedal. Derivative-free optimization of noisy functions via quasiNewton methods. SIAM Journal on Optimization, 29(2):965-993, 2019. 11", + "[2] John P Boyd. Solving transcendental equations: the Chebyshev polynomial proxy and other numerical rootfinders, perturbation series, and oracles. SIAM, 2014. 5", + "[3] Tony F Chan, Gene H Golub, and Pep Mulet. A nonlinear primal-dual method for total variation-based image restoration. SIAM Scient. Comput., 20(6):1964-1977, 1999. 3", + "[4] Swarat Chaudhuri and Armando Solar-Lezama. Smooth interpretation. ACM Sigplan Notices, 45(6):279-291, 2010. 3", + "[5] Thomas Deliot, Eric Heitz, and Laurent Belcour. Transforming a non-differentiable rasterizer into a differentiable one with stochastic gradient estimation. In ACM i3D, 2024. 1", + "[6] Deshana Desai, Etai Shuchatowitz, Zhongshi Jiang, Teseo Schneider, and Daniele Panozzo. Acorns: An easy-to-use code generator for gradients and hessian. SoftwareX, 17: 100901, 2022. 3", + "[7] Michael Fischer and Tobias Ritschel. Plateau-reduced differentiable path tracing. In CVPR, pages 4285-4294, 2023. 1, 3, 4, 5, 6, 10, 14", + "[8] Michael Fischer and Tobias Ritschel. Zerograds: Learning local surrogates for non-differentiable graphics. ACM Trans. Graph. (Proc. SIGGRAPH), 43(4):1-15, 2024. 1, 3, 13", + "[9] Roger Fletcher and Michael JD Powell. A rapidly convergent descent method for minimization. The computer journal, 6 (2):163-168, 1963. 2", + "[10] Nikolaus Hansen and Andreas Ostermeier. Adapting arbitrary normal mutation distributions in evolution strategies: The covariance matrix adaptation. In Proc. IEEE Evolutionary Computation, pages 312-317, 1996. 3", + "[11] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engil Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In CVPR, pages 406-413, 2014. 6", + "[12] James T Kajiya. The rendering equation. In Proc. SIGGRAPH, 1986. 1, 3", + "[13] Martin Kilian, Niloy J Mitra, and Helmut Pottmann. Geometric modeling in shape space. In SIGGRAPH, pages 64- es, 2007. 3", + "[14] Dilip Krishnan and Rob Fergus. Fast image deconvolution using hyper-Laplacian priors. Proc. NeurIPS, 22, 2009. 3", + "[15] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In Proc. ICCV, 2019. 1", + "[16] Matthew M Loper and Michael J Black. OpenDR: An approximate differentiable renderer. In Proc. ECCV, 2014. 1", + "[17] James Martens. New insights and perspectives on the natural gradient method. J Machine Learning Res, 21(146):1-76, 2020. 11", + "[18] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Comm ACM, 65(1), 2021. 1", + "[19] Baptiste Nicolet, Alec Jacobson, and Wenzel Jakob. Large steps in inverse rendering of geometry. ACM Trans. Graph." + ], + "bbox": [ + 93, + 114, + 483, + 888 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "(Proc. SIGGRAPH Asis), 40(6):1-13, 2021. 3", + "[20] Merlin Nimier-David, Delio Vicini, Tizian Zeltner, and Wenzel Jakob. Mitsuba 2: A retargetable forward and inverse renderer. ACM Trans. Graph., 38(6), 2019. 1", + "[21] Merlin Nimier-David, Sébastien Speierer, Benoit Ruiz, and Wenzel Jakob. Radiative backpropagation: an adjoint method for lightning-fast differentiable rendering. ACM Trans Graph, 39(4), 2020. 3", + "[22] Jorge Nocedal and Stephen J Wright. Numerical optimization. Springer, 1999. 1, 2, 11", + "[23] Art Owen and Yi Zhou. Safe and effective importance sampling. J American Statistical Assoc., 95(449), 2000. 1, 4", + "[24] Barak A Pearlmutter. Fast exact multiplication by the hessian. Neural computation, 6(1):147-160, 1994. 3, 5", + "[25] Michael JD Powell. Algorithms for nonlinear constraints that use lagrangian functions. Math. Prog., 14:224-248, 1978. 11", + "[26] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv:2204.06125, 1(2):3, 2022. 3", + "[27] Sebastian Ruder. An overview of gradient descent optimization algorithms. arXiv:1609.04747, 2016. 1", + "[28] Pedro V Sander, John Snyder, Steven J Gortler, and Hugues Hoppe. Texture mapping progressive meshes. In Proc. SIGGRAPH, pages 409-416, 2001. 3", + "[29] David F Shanno. Conditioning of quasi-newton methods for function minimization. Mathematics of computation, 24 (111):647-656, 1970. 11", + "[30] Jack Sherman. Adjustment of an inverse matrix corresponding to changes in the elements of a given column or row of the original matrix. Annu. Math. Statist., 20:621, 1949. 11", + "[31] Jonathan Richard Shewchuk. An introduction to the conjugate gradient method without the agonizing pain. 1994. 2", + "[32] James C Spall. Multivariate stochastic approximation using a simultaneous perturbation gradient approximation. IEEE Trans Automatic Control, 37(3):332-341, 1992. 3", + "[33] Eric Veach and Leonidas J Guibas. Optimally combining sampling techniques for Monte Carlo rendering. In Proc. SIGGRAPH, pages 419-428, 1995. 4, 5", + "[34] Delio Vicini, Sébastien Speierer, and Wenzel Jakob. Path replay backpropagation: differentiating light paths using constant memory and linear time. ACM Trans Graph, 40(4), 2021. 3", + "[35] Werbos. Backpropagation: Past and future. In IEEE 1988 Int Conf Neural Networks, pages 343-353. IEEE, 1988. 3, 5", + "[36] Manuel Werlberger, Werner Trobin, Thomas Pock, Andreas Wedel, Daniel Cremers, and Horst Bischof. Anisotropic huber-11 optical flow. In BMVC, volume 1, 2009. 3", + "[37] Zhewei Yao, Amir Gholami, Sheng Shen, Mustafa Mustafa, Kurt Keutzer, and Michael Mahoney. Adahessian: An adaptive second order optimizer for machine learning. In Proc. AAAI, volume 35, pages 10665-10673, 2021. 3", + "[38] Yizhou Yu, Paul Debevec, Jitendra Malik, and Tim Hawkins. Inverse global illumination: Recovering reflectance models of real scenes from photographs. In Proc. SIGGRAPH, pages 215-224, 1999. 1", + "[39] Christopher Zach, Thomas Pock, and Horst Bischof. A duality based approach for realtime TV-L1 optical flow. In DAGM Pattern Recognition, pages 214-223, 2007. 3" + ], + "bbox": [ + 516, + 92, + 903, + 893 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28206", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + } +] \ No newline at end of file diff --git a/2025/Stochastic Gradient Estimation for Higher-Order Differentiable Rendering/95d3584a-ba8b-4c05-9f1b-c00fa69cfc88_model.json b/2025/Stochastic Gradient Estimation for Higher-Order Differentiable Rendering/95d3584a-ba8b-4c05-9f1b-c00fa69cfc88_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c3c6b411649d69875b438404a841aa4c708f8916 --- /dev/null +++ b/2025/Stochastic Gradient Estimation for Higher-Order Differentiable Rendering/95d3584a-ba8b-4c05-9f1b-c00fa69cfc88_model.json @@ -0,0 +1,2220 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.124, + 0.131, + 0.877, + 0.155 + ], + "angle": 0, + "content": "Stochastic Gradient Estimation for Higher-order Differentiable Rendering" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.182, + 0.371, + 0.234 + ], + "angle": 0, + "content": "Zican Wang \nUniversity College London \nrobert.wang.19@ucl.ac.uk" + }, + { + "type": "text", + "bbox": [ + 0.413, + 0.182, + 0.586, + 0.234 + ], + "angle": 0, + "content": "Michael Fischer Adobe Research mifischer@adobe.com" + }, + { + "type": "text", + "bbox": [ + 0.629, + 0.182, + 0.846, + 0.234 + ], + "angle": 0, + "content": "Tobias Ritschel \nUniversity College London \nt.ritschel@ucl.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.486, + 0.484 + ], + "angle": 0, + "content": "We derive methods to compute higher-order differentials (Hessians and Hessian-vector products) of the rendering operator. Our approach is based on importance sampling of a convolution that represents the differentials of rendering parameters and shows to be applicable to both rasterization and path tracing. We further suggest an aggregate sampling strategy to importance-sample multiple dimensions of one convolution kernel simultaneously. We demonstrate that this information improves convergence when used in higher-order optimizers such as Newton or Conjugate Gradient relative to a gradient descent baseline in several inverse rendering tasks." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.513, + 0.222, + 0.528 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.536, + 0.484, + 0.687 + ], + "angle": 0, + "content": "Inverse rendering is concerned with optimizing the parameters of a scene to minimize a loss. This can be a useful tool when the true parameters of the scene are unknown or nontrivial to set for a human expert, and therefore need to be inferred from observations or measurements. Examples for such scenarios are multi-view reconstruction [18] or the recovery of illumination and reflectance properties [38]. Differentiable rendering has recently become a popular tool for this optimization, as automatic differentiation (AD) frameworks have become more widespread." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.484, + 0.839 + ], + "angle": 0, + "content": "However, differentiating the rendering process is far from trivial, as the rendering function [12] has zero and/or undefined gradients and step edges. Moreover, as many rendering operations rely on integration, discontinuities in the rendering pipeline cause problems for AD-engines, as we can no longer naively exchange the integral- and derivative-computations. Recent research therefore has designed a plethora of differentiable renderers [16], which compute gradients in various ways, typically incurring more [15, 20] or less [5, 7, 8] implementation and compute effort." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "A surprising insight is that, while these special renderers allow deriving gradients w.r.t. the scene parameters by differentiating (and back-propagating through) the rendering operator, virtually all of them are limited to first-order" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.269, + 0.905, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.402, + 0.907, + 0.458 + ], + "angle": 0, + "content": "Figure 1. Our approach allows sampling the Hessian for inverse rendering, here for the task of rotating the cup around its horizontal \\((x_0)\\) and vertical axes \\((x_{1})\\). The estimated positive and negative gradients are shown in blue and red, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.907, + 0.583 + ], + "angle": 0, + "content": "derivatives, and hence to expressing the gradient solely locally at a point in the parameter space. At the time of writing, no attempts to derive the Hessians required for higher-order inverse rendering methods have been published. However, decades of optimization research have shown the potential of higher-order methods in convergence and robustness [22]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.906, + 0.718 + ], + "angle": 0, + "content": "In this paper, we argue that these benefits are also applicable to inverse rendering scenarios and show that they translate to net-gains in optimization time and performance in differentiable rendering. We tackle the Achilles' heel of higher-order optimizers - their increased per-iteration cost and computational / storage requirements - by developing efficient Monte Carlo (MC) estimators of the required higher-order quantities that can be importance-sampled with established techniques [23]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.906, + 0.78 + ], + "angle": 0, + "content": "Our method leads to speedups of \\(2.71 \\times\\) over previous methods, and net-gains in optimization time, speed and robustness while only assuming the rendering operator to be a black box that can be point-sampled." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.789, + 0.665, + 0.804 + ], + "angle": 0, + "content": "2. Previous Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Gradient-based Optimization, the main workhorse of modern neural network training and inverse rendering, uses the gradient of the objective function to take iterative steps in the parameter space towards an improved solution until a (local) minimum is found. The exact nature of these steps varies with the optimizer that is being used [27]." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28198" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.242 + ], + "angle": 0, + "content": "First-order optimizers are simple to implement and cheap to execute but disregard important higher-order information about the shape of the objective function that could aid optimization, such as the second-order derivative. This usually comes at the expense of higher iteration counts, as many small steps are needed to converge to the solution. In contrast, higher-order optimizers incorporate information about the shape of the objective function, which usually allows them to take bigger steps in parameter space, leading to fewer iterations until convergence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.484, + 0.379 + ], + "angle": 0, + "content": "For the specific case of second-order optimization, this additional information is often provided via the Hessian \\(\\mathsf{H}\\) (or approximations thereof) and the Hessian-vector product (HVP). The Hessian contains the second-order derivatives of the objective function w.r.t. the optimization parameters, and can be interpreted as the curvature of the objective function. As such, it can be used to inform the optimizer about how quickly the current gradient is changing and thus, in turn, about how large the optimization step should be." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.379, + 0.484, + 0.455 + ], + "angle": 0, + "content": "Notation The following will use lowercase boldface to denote vectors and uppercase sans to denote matrices, respectively. Operators, formally defined as functions acting on functions, will be denoted in uppercase Roman lettering in order to avoid confusion with regular functions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.475, + 0.484, + 0.521 + ], + "angle": 0, + "content": "Gradient descent should be familiar to most readers, so we here recall only its terminology: The first-order Taylor expansion of the cost function \\( f \\) at position \\( \\theta \\) is" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.532, + 0.4, + 0.55 + ], + "angle": 0, + "content": "\\[\nf (\\boldsymbol {\\theta}) \\approx f (\\boldsymbol {\\theta} ^ {t}) + g (\\boldsymbol {\\theta} ^ {t}) ^ {\\top} (\\boldsymbol {\\theta} - \\boldsymbol {\\theta} ^ {t}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.562, + 0.483, + 0.592 + ], + "angle": 0, + "content": "where \\(g\\) is the gradient \\(\\nabla f = \\mathrm{d}f / \\mathrm{d}\\theta\\) of \\(f\\). The minimum is where the derivative is zero, which we can solve for as" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.603, + 0.483, + 0.633 + ], + "angle": 0, + "content": "\\[\n\\frac {\\mathrm {d}}{\\mathrm {d} \\boldsymbol {\\theta}} f (\\boldsymbol {\\theta}) = 0 \\approx g \\left(\\boldsymbol {\\theta} ^ {t}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.673 + ], + "angle": 0, + "content": "As \\(f\\) is a linear function only locally, we only make small steps with step size \\(\\gamma\\), by an update direction \\(-g(\\pmb{\\theta}^t)\\), as in" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.684, + 0.359, + 0.703 + ], + "angle": 0, + "content": "\\[\n\\pmb {\\theta} ^ {t + 1} = \\pmb {\\theta} ^ {t} - \\gamma g (\\pmb {\\theta} ^ {t}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.721, + 0.484, + 0.767 + ], + "angle": 0, + "content": "Newton's method is one of the most-used second-order optimizers. This term is derived from the second-order Taylor expansion of the objective \\( f \\) around a point \\( \\theta \\):" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.778, + 0.484, + 0.807 + ], + "angle": 0, + "content": "\\[\nf (\\pmb {\\theta}) \\approx f (\\pmb {\\theta} ^ {t}) + g (\\pmb {\\theta} ^ {t}) ^ {\\top} (\\pmb {\\theta} - \\pmb {\\theta} ^ {t}) + \\frac {1}{2} (\\pmb {\\theta} - \\pmb {\\theta} ^ {t}) ^ {\\top} \\mathsf {H} (\\pmb {\\theta} ^ {t}) (\\pmb {\\theta} - \\pmb {\\theta} ^ {t}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.818, + 0.403, + 0.834 + ], + "angle": 0, + "content": "where \\(\\mathsf{H}\\) is \\(\\nabla^2 f = \\mathrm{d}^2 f / \\mathrm{d}^2\\pmb{\\theta}\\), the Hessian of \\(f\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.835, + 0.483, + 0.865 + ], + "angle": 0, + "content": "Ideally, we would like our update step to take us to an optimum. There, the derivative is necessarily zero:" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.876, + 0.484, + 0.907 + ], + "angle": 0, + "content": "\\[\n\\frac {\\mathrm {d}}{\\mathrm {d} \\boldsymbol {\\theta}} f (\\boldsymbol {\\theta}) = 0 \\approx g \\left(\\boldsymbol {\\theta} ^ {t}\\right) + \\mathrm {H} \\left(\\boldsymbol {\\theta} ^ {t}\\right) \\left(\\boldsymbol {\\theta} - \\boldsymbol {\\theta} ^ {t}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.783, + 0.107 + ], + "angle": 0, + "content": "Solving for \\(\\theta\\) then yields the update rule" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.115, + 0.905, + 0.133 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\theta} ^ {t + 1} = \\boldsymbol {\\theta} ^ {t} - \\gamma \\mathrm {H} ^ {- 1} (\\boldsymbol {\\theta} ^ {t}) g (\\boldsymbol {\\theta} ^ {t}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.142, + 0.903, + 0.158 + ], + "angle": 0, + "content": "where \\(\\mathbf{v} = -\\mathsf{H}^{-1}(\\pmb{\\theta}^t)g(\\pmb{\\theta}^t)\\) is called the Newton direction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.158, + 0.906, + 0.219 + ], + "angle": 0, + "content": "Newton's method requires the computation, storage, and inversion of the full Hessian, which quickly becomes a bottleneck in higher dimensions, as the Hessian for an \\(n\\)-dimensional optimization problem is in \\(\\mathbb{R}^{n\\times n}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.236, + 0.906, + 0.342 + ], + "angle": 0, + "content": "Newton Conjugate Gradient [31], upgrades Newton's method in two ways. First, it solves for Newton's direction iteratively as per the linear equation \\(\\mathsf{H}(\\pmb{\\theta})\\mathbf{v} = -g(\\pmb{\\theta})\\) using conjugate directions [31]. Second, instead of an arbitrary step length \\(\\gamma\\), it also decides the scalar \\(\\alpha\\) by which we move along this direction \\(\\mathbf{v}\\). To derive \\(\\alpha\\), first consider the Taylor expansion:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.35, + 0.905, + 0.381 + ], + "angle": 0, + "content": "\\[\n\\frac {\\mathrm {d}}{\\mathrm {d} \\alpha} f (\\boldsymbol {\\theta} + \\alpha \\mathbf {v}) \\approx g (\\boldsymbol {\\theta}) ^ {\\top} \\mathbf {v} + \\alpha \\mathbf {v} ^ {\\top} H (\\boldsymbol {\\theta}) \\mathbf {v} = 0 \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.388, + 0.704, + 0.403 + ], + "angle": 0, + "content": "which can be re-arranged to" + }, + { + "type": "equation", + "bbox": [ + 0.649, + 0.411, + 0.905, + 0.446 + ], + "angle": 0, + "content": "\\[\n\\alpha = - \\frac {g ^ {\\mathsf {T}} (\\boldsymbol {\\theta}) \\mathbf {v}}{\\mathbf {v} ^ {\\mathsf {T}} \\mathsf {H} (\\boldsymbol {\\theta}) \\mathbf {v}}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.453, + 0.906, + 0.51 + ], + "angle": 0, + "content": "The update rule for the Newton Conjugate Gradient is more involved: First, we maintain the direction \\(\\mathbf{v}\\) and a residual \\(\\mathbf{r}\\), which are initialized to be the gradient at the initial positions:" + }, + { + "type": "equation", + "bbox": [ + 0.644, + 0.512, + 0.905, + 0.53 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} ^ {0} = \\mathbf {r} ^ {0} = - g (\\boldsymbol {\\theta} ^ {t}) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.535, + 0.901, + 0.55 + ], + "angle": 0, + "content": "We then find the \\(\\alpha\\) by equation Eq. 5 and update the point:" + }, + { + "type": "equation", + "bbox": [ + 0.647, + 0.558, + 0.905, + 0.575 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\theta} ^ {t + 1} = \\boldsymbol {\\theta} ^ {t} + \\alpha \\mathbf {v} ^ {t}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.586, + 0.906, + 0.631 + ], + "angle": 0, + "content": "The next direction is chosen by updating the residual \\(\\mathbf{r}\\) and computing a new conjugate search direction \\(\\mathbf{v}\\) with the Fletcher-Reeves formula [9]:" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.639, + 0.788, + 0.71 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {r} ^ {t + 1} = \\mathbf {r} ^ {t} - \\alpha \\mathsf {H} (\\boldsymbol {\\theta} ^ {t}) \\mathbf {v} ^ {t} \\\\ \\beta = \\frac {\\mathbf {r} ^ {t + 1 , \\top} \\mathbf {r} ^ {t + 1}}{\\mathbf {r} ^ {t , \\top} \\mathbf {r} ^ {t}} \\\\ \\mathbf {v} ^ {t + 1} = \\mathbf {r} ^ {t + 1} + \\beta \\mathbf {v} ^ {t}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Since Newton's method is derived via second-order Taylor expansion (approximated by a hyper-parabola), the derived Hessian may lead to either a maximum or a minimum depending on the local curvature of the function being approximated. This is determined by the positive-definiteness of the Hessian, and a negative-definite Hessian will diverge from the gradient's descent direction. To avoid this, one can ensure that the Hessian is always positive semi-definite using the Hessian modification technique and ensure step truncation to a trust region [22], \\(\\alpha = \\min (\\alpha ,\\Delta \\| \\mathbf{v}\\|^{-1})\\), where \\(\\Delta\\) is the trust region radius (see Supplemental Sec. H)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28199" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.155 + ], + "angle": 0, + "content": "Hessian-vector product approximations go one step further by entirely avoiding to compute \\(\\mathsf{H}\\) when producing Hv. Pearlmutter [24] and Werbos [35] discuss different options to do so, but a simple option is central differences" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.168, + 0.484, + 0.201 + ], + "angle": 0, + "content": "\\[\n\\mathsf {H} (\\pmb {\\theta}) \\mathbf {v} \\approx \\lim _ {\\varepsilon \\to 0} \\frac {g (\\pmb {\\theta} + \\varepsilon \\mathbf {v}) - g (\\pmb {\\theta} - \\varepsilon \\mathbf {v})}{2 \\varepsilon}. \\qquad (8)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.231, + 0.485, + 0.442 + ], + "angle": 0, + "content": "Hessians in vision and graphics have been applied to several optimization tasks, for instance, for total-variation denoising [3], texture parameterization and surface mappings with approximated Hessians [28], 3D shape manipulation [13], and image deconvolution tasks via (Hessian-free) Newton methods [14]. For optical flow estimation, both Zach et al. [39] and Werlberger et al. [36] use second-order information. Additionally, several works have explored Hessian-based algorithms in machine learning for vision tasks. Yao et al. [37] introduces a Hessian-based pruning method for CNNs for image classification, while Ramesh et al. [26] utilizes Hessians to improve the training of image generation models and Desai et al. [6] introduces an algorithm that derives Hessians for C code." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.445, + 0.484, + 0.657 + ], + "angle": 0, + "content": "Hessians in inverse rendering, however, have received surprisingly little attention, potentially due to their additional implementation overhead and computational complexity. In addition to the fact that conventional AD systems are mostly designed for first-order gradient computations, calculating the second-order information requires storing the whole forward- and first-derivative-graph in memory, which can lead to exponential memory growth. Nicolet et al. [19] are the closest to our work by approximating second-order steps for mesh reconstruction. However, they set the Hessian to the identity to avoid computational expense and instead work with a Laplacian regularizer, which works for their formulation and the case of mesh optimization, but is unclear how to translate to general problems." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.901 + ], + "angle": 0, + "content": "For derivative-free gradient estimators, it is equally unclear how second-order information would be computed. The zeroth-order estimators simultaneous perturbation stochastic approximation (SPSA) [32] and finite differences (FD) estimate a first-order gradient, whose second-order derivative naturally is zero. Extending these estimators to second-order information requires prohibitive amounts of function evaluations. ZeroGrads [8], which learns a neural network that fits the cost landscape, uses ReLU non-linearities, whose second-order derivative equally decays to zero. Covariance adaptation evolution strategy (CMA-ES) is often used when typical second-order optimization fails to converge [10]. It ensures that the Hessian is always positive definite, but if second-order derivative-based methods are successful, they are usually faster than CMA-ES." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.659, + 0.108 + ], + "angle": 0, + "content": "3. Our approach" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.113, + 0.907, + 0.175 + ], + "angle": 0, + "content": "We first describe the computation of gradients using importance sampling of a combined gradient-smoothing operator from previous work [4, 7] which we then extend to Hessianians, and in a next step to Hessian-vector products." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.184, + 0.646, + 0.2 + ], + "angle": 0, + "content": "3.1. Background" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.203, + 0.907, + 0.248 + ], + "angle": 0, + "content": "Rendering equation The rendering equation (RE) [12] describes the radiance \\( L \\) leaving a point \\( \\mathbf{x} \\) in the scene into the direction \\( \\omega_{\\mathrm{o}} \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.259, + 0.907, + 0.303 + ], + "angle": 0, + "content": "\\[\nL (\\mathbf {x}, \\omega_ {\\mathrm {o}}; \\boldsymbol {\\theta}) = \\int_ {\\Omega} \\underbrace {f _ {\\mathrm {r}} \\left(\\omega_ {\\mathrm {i}} , \\omega_ {\\mathrm {o}}\\right) L \\left(\\mathbf {y} , \\omega_ {\\mathrm {i}} ; \\boldsymbol {\\theta}\\right)} _ {R \\left(\\omega_ {\\mathrm {i}}; \\boldsymbol {\\theta}\\right)} \\mathrm {d} \\omega_ {\\mathrm {i}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.314, + 0.907, + 0.451 + ], + "angle": 0, + "content": "where \\(\\theta\\) are the parameters of the scene we would like to optimize, such as object geometry, reflectance, or light emission. This integral over \\(\\Omega\\), i.e., all incoming \\(\\omega_{\\mathrm{i}}\\) directions that multiply the radiance field arriving from that direction from the closest other point \\(\\mathbf{y}\\) in direction \\(\\omega_{\\mathrm{i}}\\) with the bidirectional reflectance distribution function (BRDF) \\(f_{\\mathrm{r}}\\), has no analytical closed-form solution, and hence usually is approximated - both in forward and inverse rendering - via MC methods. We will shorthand the entire integrand as \\(R\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.471, + 0.906, + 0.502 + ], + "angle": 0, + "content": "Problem statement We would now like to apply a differential operator \\( \\mathrm{D} \\) to the RE, as in" + }, + { + "type": "equation", + "bbox": [ + 0.661, + 0.514, + 0.907, + 0.532 + ], + "angle": 0, + "content": "\\[\n\\mathrm {D} L (\\mathbf {x}, \\omega_ {\\mathrm {i}}; \\boldsymbol {\\theta}). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.544, + 0.906, + 0.589 + ], + "angle": 0, + "content": "If \\( \\mathrm{D} \\) was the gradient operator \\( \\frac{\\partial L}{\\partial \\theta} \\), this would be differentiable rendering, for other operators, this becomes higher-order differentiable rendering." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.59, + 0.907, + 0.65 + ], + "angle": 0, + "content": "The trouble is that we cannot move D, be it gradient or higher-order, into the integral, as in many cases (e.g., for BRDF or spatial derivatives), the integrand is discontinuous in \\(\\theta\\), so" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.661, + 0.907, + 0.695 + ], + "angle": 0, + "content": "\\[\n\\mathrm {D} L (\\mathbf {x}, \\omega_ {\\mathrm {i}}; \\boldsymbol {\\theta}) \\neq \\int_ {\\Omega} \\mathrm {D} R (\\omega_ {\\mathrm {i}}; \\boldsymbol {\\theta}) \\mathrm {d} \\omega_ {\\mathrm {i}}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.704, + 0.907, + 0.765 + ], + "angle": 0, + "content": "However, the right-hand side of the above expression is exactly the quantity that naively-applied AD computes [21, 34], leading to wrong gradients in (any-order) differentiable rendering." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.786, + 0.907, + 0.862 + ], + "angle": 0, + "content": "Solution The idea is to enforce the property that prevents differentiation - smoothness -, so that we actually can differentiate. To that end, assume a further linear operator \\( S \\) that is smoothing any function in \\( \\theta \\). This provides a smooth rendering equation \\( \\bar{L} \\)" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.873, + 0.907, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\bar {L} (\\mathbf {x}, \\omega_ {\\mathrm {o}}; \\boldsymbol {\\theta}) = \\int_ {\\Omega} \\mathrm {S} R \\left(\\omega_ {\\mathrm {i}}; \\boldsymbol {\\theta}\\right) \\mathrm {d} \\omega_ {\\mathrm {i}}. \\tag {12}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "28200" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.912, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.203, + 0.908, + 0.273 + ], + "angle": 0, + "content": "Figure 2. Comparison of classic (yellow) and smooth (red) gradients, as well as our Hessians (blue) on an inverse rendering problem to change the initial parameter \\(\\theta\\) so that the left triangle overlaps the right one. Classic gradients are zero almost everywhere (plateaus) except where the triangles already overlap. These methods do not converge. Smooth gradients point into the right direction, but make steps far from the optimum (dotted line). An update taking into account the curvature of the loss landscape lands at a point very close to the target. We incorporate this curvature information via our Hessians." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.293, + 0.483, + 0.323 + ], + "angle": 0, + "content": "Smoothing can be achieved by convolution, so for any function \\( f \\) in any dimension" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.327, + 0.482, + 0.358 + ], + "angle": 0, + "content": "\\[\n\\mathrm {S} f (\\boldsymbol {\\theta}) = \\int_ {\\Theta} \\kappa (\\tau) f (\\boldsymbol {\\theta} - \\tau) \\mathrm {d} \\tau , \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.363, + 0.484, + 0.424 + ], + "angle": 0, + "content": "where \\(\\kappa\\) is a smoothing kernel, such as a Gaussian, which we use in this work. This convolved integrand is now smooth, which, according to Leibniz' rule, allows us to move the differential operator into the integral" + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.428, + 0.483, + 0.461 + ], + "angle": 0, + "content": "\\[\n\\mathrm {D} \\bar {L} (\\mathbf {x}, \\omega_ {\\mathrm {o}}; \\boldsymbol {\\theta}) = \\int_ {\\Omega} \\mathrm {D S} R \\left(\\omega_ {\\mathrm {i}}; \\boldsymbol {\\theta}\\right) \\mathrm {d} \\omega_ {\\mathrm {i}}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.464, + 0.484, + 0.495 + ], + "angle": 0, + "content": "which, after rearranging and expanding, yields an integral that can be approximated via MC:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.499, + 0.483, + 0.531 + ], + "angle": 0, + "content": "\\[\n\\mathrm {D} \\bar {L} (\\mathbf {x}, \\omega_ {\\mathrm {o}}; \\boldsymbol {\\theta}) = \\int_ {\\Omega} \\int_ {\\Theta} \\mathrm {D} \\kappa (\\boldsymbol {\\tau}) R \\left(\\omega_ {\\mathrm {i}}; \\boldsymbol {\\theta} - \\boldsymbol {\\tau}\\right) \\mathrm {d} \\boldsymbol {\\tau} \\mathrm {d} \\omega_ {\\mathrm {i}}. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.484, + 0.701 + ], + "angle": 0, + "content": "MC here means to take random samples from the product space of light paths and scene parameters. This works best if we can importance-sample for the integrand. The integrand here is a product of four terms. Sampling for the incoming radiance and BRDF terms has been investigated in the rendering community [33] and is not our consideration here, so we simply adopt these strategies. Sampling for the application of the differential operator to the smoothing kernel is the essence of the problem at hand. Depending on the choice of differential operator, we will derive three sampling strategies for the three resulting estimators next." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.718, + 0.484, + 0.824 + ], + "angle": 0, + "content": "Conclusion In conclusion, to perform efficient and practical any-order differentiation of the RE, we would need to implement two functions: first, a convolution kernel that combines smoothing and the desired differentiation, and second, a function to sample from that kernel for importance sampling. We will now do so for the gradient (Sec. 3.2), Hessian (Sec. 3.3) and HVP (Sec. 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.831, + 0.206, + 0.845 + ], + "angle": 0, + "content": "3.2. Gradients" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.849, + 0.483, + 0.88 + ], + "angle": 0, + "content": "Operator: For first-order gradient descent, Fischer and Ritschel [7] have differentiated using the gradient operator" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.884, + 0.483, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathrm {D} ^ {\\mathrm {G}} = \\nabla = \\partial / \\partial \\tau_ {i} \\in (\\mathbb {R} ^ {n} \\rightarrow \\mathbb {R}) \\rightarrow (\\mathbb {R} ^ {n} \\rightarrow \\mathbb {R} ^ {n}), \\tag {16}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.291, + 0.905, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.473, + 0.907, + 0.556 + ], + "angle": 0, + "content": "Figure 3. Derivation of smooth differentiation by convolution for three differential operators (rows) involves three steps (columns): Defining the operator to combine smoothing and differentiation (1st col.), positivization and normalization to become a probability density function (PDF) (2nd col.), creating an inverse mapping (3rd col.), which finally allows sampling (4th col.)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.572, + 0.906, + 0.601 + ], + "angle": 0, + "content": "which maps a scalar function of \\( n \\) dimensions to an \\( n \\)-dimensional gradient vector field." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.602, + 0.88, + 0.617 + ], + "angle": 0, + "content": "The combination of smoothing and differentiation is" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.623, + 0.905, + 0.65 + ], + "angle": 0, + "content": "\\[\n\\mathrm {D} ^ {\\mathrm {G}} \\kappa (\\boldsymbol {\\tau}) = \\nabla_ {i} \\mathcal {N} (\\boldsymbol {\\tau}, \\sigma) = - \\frac {\\tau_ {i}}{\\sigma^ {2}} \\cdot \\mathcal {N} (\\boldsymbol {\\tau}, \\sigma). \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.655, + 0.886, + 0.669 + ], + "angle": 0, + "content": "For a derivation, please see the supplemental, Sec. A." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.67, + 0.906, + 0.759 + ], + "angle": 0, + "content": "Sampling: For sampling a one-dimensional Gaussian gradient, we can use inverse transform sampling via the Smirnov transform [7]. To this end, the integrand has to be a PDF, i.e., positive and integrating-to-1 [23]. Subsequently, we compute the integral of a positivized version of \\( p^{\\mathrm{G}} \\), the cumulative distribution function (CDF) \\( P^{\\mathrm{G}} \\), and invert it as" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.779, + 0.905, + 0.82 + ], + "angle": 0, + "content": "\\[\nP ^ {\\mathrm {G}, - 1} (\\xi) = \\left\\{ \\begin{array}{l l} - \\sqrt {2 \\sigma^ {2} \\log (2 \\xi)} & \\text {i f} \\xi \\leq 0. 5, \\\\ + \\sqrt {2 \\sigma^ {2} \\log (1 - \\xi)} & \\text {e l s e .} \\end{array} \\right. \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "This derivation applies to the dimension which is being differentiated. The separability of multi-dimensional Gaussians ensures that the other dimensions remain a Gaussian distribution, and we can sample these dimensions independently. For a derivation of this, see Suppl. Sec. B." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28201" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.195, + 0.105 + ], + "angle": 0, + "content": "3.3. Hessians" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.11, + 0.448, + 0.125 + ], + "angle": 0, + "content": "Operator: The differential operator for Hessian is" + }, + { + "type": "equation", + "bbox": [ + 0.096, + 0.132, + 0.482, + 0.164 + ], + "angle": 0, + "content": "\\[\n\\mathrm {D} ^ {\\mathrm {H}} = \\nabla^ {2} = \\partial^ {2} / \\partial \\tau_ {i} \\partial \\tau_ {j} \\in (\\mathbb {R} ^ {n} \\rightarrow \\mathbb {R}) \\rightarrow (\\mathbb {R} ^ {n} \\rightarrow \\mathbb {R} ^ {n \\times n}), \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.165, + 0.483, + 0.21 + ], + "angle": 0, + "content": "which maps a scalar function in \\( n \\) dimensions to its \\( n \\times n \\)-element Hessian field. The combination of second-order derivatives and Gaussian smoothing is" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.217, + 0.482, + 0.291 + ], + "angle": 0, + "content": "\\[\n\\mathrm {D} ^ {\\mathrm {H}} \\kappa_ {i, j} (\\boldsymbol {\\tau}) = \\left\\{ \\begin{array}{l l} \\left(- \\frac {1}{\\sigma^ {2}} + \\frac {\\tau_ {i} ^ {2}}{\\sigma^ {4}}\\right) \\cdot \\mathcal {N} (\\boldsymbol {\\tau}, \\sigma) & \\text {i f} i = j, \\\\ \\frac {\\tau_ {i} \\tau_ {j}}{\\sigma^ {4}} \\cdot \\mathcal {N} (\\boldsymbol {\\tau}, \\sigma) & \\text {e l s e .} \\end{array} \\right. \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.292, + 0.37, + 0.306 + ], + "angle": 0, + "content": "For a derivation, see supplemental Sec. C." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.307, + 0.484, + 0.382 + ], + "angle": 0, + "content": "Sampling: We first need to positivize as the function is signed, then scale the function so that it is a valid distribution. This is done differently for diagonal and off-diagonal elements. For the diagonal case, we construct the CDF of a second-order derivative \\( P_{ii}^{\\mathrm{H}} \\) of the 2D Gaussian:" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.39, + 0.484, + 0.508 + ], + "angle": 0, + "content": "\\[\nP _ {i i} ^ {\\mathrm {H}} \\left(\\tau_ {i} = u\\right) = \\left\\{ \\begin{array}{c l} - \\frac {u}{4 \\sigma} \\exp \\left(\\frac {1}{2} - \\frac {u ^ {2}}{2 \\sigma^ {2}}\\right) & \\text {i f} u < - \\sigma , \\\\ 0. 5 + \\frac {u}{4 \\sigma} \\exp \\left(\\frac {1}{2} - \\frac {u ^ {2}}{2 \\sigma^ {2}}\\right) & \\text {i f} u \\in [ - \\sigma , \\sigma ] \\\\ 1 - \\frac {u}{4 \\sigma} \\exp \\left(\\frac {1}{2} - \\frac {u ^ {2}}{2 \\sigma^ {2}}\\right) & \\text {i f} u > \\sigma . \\end{array} \\right. \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.483, + 0.599 + ], + "angle": 0, + "content": "See supplemental Sec. D for a derivation of this result. This is a transcendental equation whose inverse CDF cannot be expressed in closed form via elementary functions [2]. Instead, we 1D-tabulate its values on the range \\(10\\sigma\\) for every \\(i\\). The inverse function is found by searching in that range, and storing the pre-sorted inverse indices for \\(O(1)\\) access." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.704 + ], + "angle": 0, + "content": "The off-diagonals are the product of two partial gradients that we have already derived in Eq. 17. Fortunately, the product of two independent distributions can be sampled by sampling each one independently. Thus, no extra derivation is required here. For sampling in higher dimensions, similar to the gradient sampler in Sec. B, the rest of the dimensions are sampled from a Gaussian distribution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.736 + ], + "angle": 0, + "content": "In all methods, we exploit the symmetry in the Hessian matrix when sampling and estimating." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.307, + 0.759 + ], + "angle": 0, + "content": "3.4. Hessian-vector product" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.761, + 0.483, + 0.792 + ], + "angle": 0, + "content": "Operator: As explained in Sec. 2, avoiding to store the full Hessian is possible by using HVPs:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.798, + 0.483, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\mathrm {D} ^ {\\mathrm {H V P}} = \\nabla^ {2} \\mathbf {v} = \\partial^ {2} / \\partial \\tau_ {i} \\partial \\boldsymbol {\\tau} \\cdot \\partial / \\partial \\boldsymbol {\\tau}. \\tag {22}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In essence, a HVP is the directional gradient of the gradient. As there is one direction and the gradient has \\( n \\) dimensions, the HVP is an \\( n \\)-dimensional vector, too. In contrast, a Hessian is the non-directional gradient of the gradient, i.e., the gradient along all \\( n \\) dimensions, and as such in \\( \\mathbb{R}^{n \\times n} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.182 + ], + "angle": 0, + "content": "Sampling: To sample the smooth HVP, we simply need the directional central differences of an estimator of gradients, which we already have. So the estimator is the difference of two first-order estimators, evaluated at positions shifted from the current solution along the gradient direction [24, 35]:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.191, + 0.906, + 0.226 + ], + "angle": 0, + "content": "\\[\n\\mathrm {D} ^ {\\mathrm {H V P}} \\kappa (\\boldsymbol {\\tau}) = \\frac {\\mathrm {D} ^ {\\mathrm {G}} \\kappa (\\boldsymbol {\\tau} - \\varepsilon \\mathbf {v}) - \\mathrm {D} ^ {\\mathrm {G}} \\kappa (\\boldsymbol {\\tau} + \\varepsilon \\mathbf {v})}{2 \\varepsilon}. \\tag {23}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.234, + 0.63, + 0.249 + ], + "angle": 0, + "content": "3.5. Aggregate" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.252, + 0.906, + 0.327 + ], + "angle": 0, + "content": "We summarize the key property of all samplers - the number of function evaluations required - in Tab. 1. Function evaluations require execution of the black-box rendering engine, the most costly part of the optimization, and hence should be minimized." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.345, + 0.716, + 0.388 + ], + "angle": 0, + "content": "Table 1. Time complexity of different estimator variants. AIS: aggregated importance sampling." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.398, + 0.729, + 0.505 + ], + "angle": 0, + "content": "
no ISISAIS
Fin. Diff.nnn
FR22 [7]111
Our Grad.1n1
Our Hess.1n21
Our HVP1n1
" + }, + { + "type": "text", + "bbox": [ + 0.744, + 0.328, + 0.907, + 0.524 + ], + "angle": 0, + "content": "Classic finite differences take opposing samples in each dimension and hence, for an \\(n\\)-dimensional problem, require \\(2n\\) function evaluations for a single gradient sample. All other methods (second row in Tab. 1 onward) are based on MC, so we can get a gradient estimate using" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.906, + 0.614 + ], + "angle": 0, + "content": "a fixed number of M function evaluations. This, however, comes at the cost of variance, which can be reduced with importance sampling, but that is done in all dimensions of the differential quantity independently, so it requires as many evaluations as these have elements (second column in Tab. 1). For a Hessian, this can be substantial, \\( n^2 \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.615, + 0.907, + 0.901 + ], + "angle": 0, + "content": "As a compromise between low number of function evaluations and low variance, we propose \"aggregate\" importance sampling of the differential quantity (Fig. 4): instead of importance-sampling each dimension optimally but individually, we importance-sample w.r.t. the average of the convolution across all dimensions (combining several proposals by sampling from a mixture is similar to MIS using the balance heuristic [33]). This average convolution is a single function again, and with our combined sampling strategy can again be sampled with one function call per iteration (last column of Tab. 1). We implement this in two simple steps: first we randomly decide which element of the differential representation to choose (e.g., which matrix element out of \\( n^2 \\) in the Hessian), and then we compute the convolution sample for all \\( n^2 \\) but with the same function value that is only evaluated once. This increases variance, as the average kernel is not identical to the individual kernels but shares many properties, e.g., they are all zero in value at position 0. As a linear combination of unbiased es" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28202" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.092, + 0.482, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.26, + 0.483, + 0.399 + ], + "angle": 0, + "content": "Figure 4. Aggregate and non-aggregate sampling for a 2D optimization space and, consequently, a \\(2 \\times 2\\) Hessian: Without aggregation, importance sampling is done for each element of the Hessian independently (four points at different positions in each kernel), leading to four rendering calls and four different mugs. Each value is then weighted by the kernel (blue and red colors denote positive and negative). In aggregate sampling, we importance-sample the average of all four stencils, resulting in a single sample location and, hence, four times the same mug that can be rendered in one call, weighted with four different kernels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.404, + 0.484, + 0.482 + ], + "angle": 0, + "content": "timators, this is still unbiased. We show in Sec. J that this pays off. Note that Tab. 1 is both time and space complexity, except for the aggregate importance sampling of the HVP, where the space complexity stays \\( O(n) \\) while time complexity is reduced to \\( O(1) \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.491, + 0.208, + 0.506 + ], + "angle": 0, + "content": "4. Evaluation" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.511, + 0.195, + 0.526 + ], + "angle": 0, + "content": "4.1. Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.484, + 0.592 + ], + "angle": 0, + "content": "Our evaluation compares the performance of gradients (input to gradient descent), Hessians and non-aggregate as well as aggregate Hessian-vector product (both input to higher-order optimizers)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.61, + 0.484, + 0.777 + ], + "angle": 0, + "content": "Baselines An established inverse rendering solution is Mitsuba with gradient descent. FR22 is using gradient descent with Fischer and Ritschel [7] MC gradients. OursG, OursH, OursHVP and OursHVPA are our approach for gradient, Hessians and Hessian-vector products, and aggregates, respectively. OursG is used with gradient descent (GD), the others in combination with conjugate gradient (CG). We have also experimented with LBFGS as detailed in the supplemental, but do not report it here as it almost never converges and never at competitive speed. All methods use the Adam optimizer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Tasks We tackle artificial problems with known analytic Hessians and Hessian-vector products as well as real inverse rendering tasks. As a simple first test case, we optimize the smooth, quadratic potential \\((\\mathrm{QUAD}^1)\\) function \\(ax_0^2 +bx_1^2 +cx_0x_1\\) in \\(\\mathbb{R}^2\\), with the fixed variables \\(a = 5,b = 5,c = 7.5\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Second, we optimize the classic plateau-demonstration task from [7]: the 2D position of some boxes is optimized" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "to match a reference. This is already a much harder task, as there is a plateau in the cost landscape when the squares do not overlap (almost always the case in the initial configuration), as the image-space error does not change. We study the case of one \\((\\mathrm{Box2}^3)\\) and five \\((\\mathrm{Box10}^5)\\) squares, with two and ten dimensions to optimize, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.907, + 0.333 + ], + "angle": 0, + "content": "For real inverse problems, we study the optimization of reflectance, light, and geometry and render using Mitsuba. In the MUG7 task, we optimize the vertical rotation of a coffee cup such that it matches a reference. In the SHADOW9 task, we optimize the position of a sphere that is unobserved in the rendered image, such that the shadow it casts matches a reference shadow. In the BUNNY11 task, the \\(x\\) and \\(z\\) position and the rotation around the \\(z\\) axis of the Stanford bunny are optimized. This task is set up specifically for Mitsuba to converge12, with no plateaus in the loss function." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.334, + 0.906, + 0.47 + ], + "angle": 0, + "content": "To test the scaling with dimensionality, we also optimize the pixels of a \\(32 \\times 32\\) texture (TEXTURE\\(^{13}\\)), with 1024 parameters. In addition, we have used our approach on a mesh optimization task, where we fit the 3D position of the vertices of a tessellated sphere to images of a target object. For SUZANNE, we fit 2,562 vertices, for BANANA, we fit the 2,737 vertices and their RGB vertex colors, similar to the DTU setup [11]. For trajectories of the optimization and the outcome, see Fig. 5 top and bottom rows, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.906, + 0.666 + ], + "angle": 0, + "content": "Finally, we learn a \\(\\mathrm{CNN}^{15}\\) to predict scene parameters from images using inverse rendering without scene parameter labels. In this case, the optimization is learning in the stricter sense: we tune parameters of a deep architecture instead of scene parameters directly. The CNN takes as input an image of a mug and produces as output the orientation for a mug. As a loss, these parameters are inserted into a renderer, differentiated by different approaches and compared to a target. The CNN has 267745 parameters, learned by differentiating through an image loss and rendering of the scene given the estimated parameters. Another CNN \\((\\mathrm{CNN}5^{19})\\) is trained to predict the rotation, position, and color of the mug with 268773 parameters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.667, + 0.906, + 0.756 + ], + "angle": 0, + "content": "Differentials of this function are computed as detailed in supplemental Sec. E, where we derive a grey-box approach that samples only the black-box part (the rendering) and combines its differentials with analytic differentials for the white-box part (the CNN). Importantly, this idea works for gradients, as well as for higher-order differentials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Metrics Our main measure of success is convergence speed in wall-clock and parameter difference to the true parameters (which we know in all inverse rendering tasks as we created the scenes). A secondary measure of success is the image difference, as during optimization, we consider the parameters hidden. All metrics are evaluated across an ensemble of 20 runs averaged across 10 steps in time. In convergence plots, the ensemble median is shown at every point in time, averaged across 20 time steps." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28203" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.907, + 0.147 + ], + "angle": 0, + "content": "Table 2. Quantitative results of different methods on different tasks (rows) and their convergence plots. We report convergence time in wall-clock units, in ratio to the overall best method, OurHVPA. In the numerical columns,.9 and .99 report the time taken to achieve 90 and \\(99\\%\\) error reduction from the initial starting configuration, respectively, while the bar plots graphically show these findings. The line plots report image- and parameter-space convergence in the left and right column, respectively, on a log-log scale." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.158, + 0.938, + 0.891 + ], + "angle": 0, + "content": "
TaskMethodImage errorParameter errorMitsuba FR22 CMA-ES OurG OurH OurHVP OurHVPA
.9.99.999.9.99.999
QUAD1Mitsu2—————ErrorLog error
FR221.802.382.982.203.173.90
CMA-ES0.150.130.140.130.140.23
OurG3.985.136.024.726.427.77
OurH1.171.171.151.181.161.17
OurHVP0.760.620.620.630.590.60
OurHVPA1.001.001.001.001.001.00
BOX23Mitsu——————ErrorLog error
FR223.242.492.572.782.892.70
CMA-ES5.513.453.158.135.514.13
OurG2.882.162.212.862.652.38
OurH0.930.951.080.760.841.84
OurHVP0.810.640.7040.740.720.72
OurHVPA1.001.001.001.001.001.00
BOX105Mitsu——————ErrorLog error
FR221.491.201.711.801.541.29
CMA-ES3.102.271.734.523.232.48
OurG2.181.771.532.532.261.88
OurH5.975.334.685.485.965.45
OurHVP1.511.441.6961.261.501.58
OurHVPA1.001.001.001.001.001.00
MUG7Mitsu——————ErrorLog error
FR225.925.695.463.286.536.29
CMA-ES2.951.971.823.322.382.20
OurG3.993.913.742.364.424.30
OurH1.571.091.531.001.081.20
OurHVP1.151.141.181.140.871.05
OurHVPA81.001.001.001.001.001.00
SHAD9Mitsu——————ErrorLog error
FR221.941.401.551.891.141.71
CMA-ES1.961.151.034.382.201.66
OurG1.371.181.132.381.471.34
OurH1.671.441.662.321.741.46
OurHVP1.140.910.961.371.131.07
OurHVPA1.001.001.001.001.001.00
BUNNY11Mitsu120.901.291.461.871.381.28ErrorLog error
FR220.750.751.120.780.630.71
CMA-ES1.160.950.851.270.910.85
OurG0.760.891.070.690.640.76
OurH1.291.231.291.180.801.07
OurHVP0.680.680.700.630.440.58
OurHVPA1.001.001.001.001.001.00
TEXTURE13Mitsu——————ErrorLog error
FR224.042.271.804.042.271.80
CMA-ES7.084.604.3511.167.477.44
OurG3.032.862.823.032.862.82
OurH440.1239.78—40.1239.78—
OurHVP4.014.004.004.014.004.00
OurHVPA1.001.001.001.001.001.00
CNN15Mitsu——————ErrorLog error
FR222.1116——1.451.322.71
CMA-ES———2.08——
OurG1.14——0.960.832.28
OurH0.770.77—0.740.770.74
OurHVP0.870.9317—0.880.910.87
OurHVPA1.001.00—181.001.001.00
CNN519Mitsu——————ErrorLog error
FR22——————
CMA-ES——————
OurG18.68——3.70——
OurH19.00——5.79——
OurHVP3.64201.46—1.491.58—
OurHVPA1.001.00—1.001.001.001.00
" + }, + { + "type": "text", + "bbox": [ + 0.551, + 0.893, + 0.704, + 0.902 + ], + "angle": 0, + "content": ".9 .99 .999 .9 99 .999" + }, + { + "type": "text", + "bbox": [ + 0.76, + 0.894, + 0.8, + 0.903 + ], + "angle": 0, + "content": "Log time" + }, + { + "type": "text", + "bbox": [ + 0.854, + 0.894, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Log time" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28204" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.905, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.23, + 0.903, + 0.245 + ], + "angle": 0, + "content": "Figure 5. Inverse mesh optimization from renderings. Gradients use FR22, Hessians use OursHVPA, CMA-ES uses the pycma package." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.253, + 0.185, + 0.267 + ], + "angle": 0, + "content": "4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.483, + 0.349 + ], + "angle": 0, + "content": "The main results of our evaluation are summarize in Tab. 2 where empty cells did not converge or the method is not applicable to that task. On average across tasks and methods, our premiere method oursHVPA speeds up the convergence by a factor of 2.71." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.351, + 0.484, + 0.411 + ], + "angle": 0, + "content": "In general, our methods, in particular OursHVP and OursHVPA, lead the level of error reduction across all time budgets, as seen by comparing the convergence curves vertically where they consistently decrease fastest." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.412, + 0.484, + 0.579 + ], + "angle": 0, + "content": "We benefit most in the artificial QUAD task, but MUG and SHADOW are both real rendering problems, where we are around seven times and \\(60\\%\\) faster, respectively. We see that OurG typically cannot outperform FR22, which uses an approximation of the correct high-dimensional gradient kernel. Doing it \"right\" only pays off when going to higher order. We also note that OursHVP outperforms OursH due to less function evaluations (as the metric is wall-clock - the iteration count for both is close). The log-log plots confirm that the higher-order variants enabled by our approach are much faster, but also converge slightly less stable." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.483, + 0.687 + ], + "angle": 0, + "content": "While tasks like QUAD and Box2 are not applicable to inverse renderers like Mitsuba², our methods treat any loss function as a black-box and retrieve derivatives via sampling. Comparing the performance of OurHVP and OurHVPA across a task with two and ten dimensions, we find the expected relation: OurHVP is faster in low dimensions⁴ than in higher dimensions for a similar task⁶." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.794 + ], + "angle": 0, + "content": "We notice that the OurHVPA is not out-performing the OurHVP and OurH in the MUG task. This is because there is a single scene parameter for this task, thus, the higher order derivative is all in order of one. With a higher variance for OurHVPA, it could be a little slower. The effect of using the aggregate starts to show from SHAD, where OurHVPA reaches the 1:1000 convergence the fastest10." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.901 + ], + "angle": 0, + "content": "For all the rendering tasks, only the BUNNY task converges for Mitsuba12. Since the other tasks have plateaus in the loss function, the analytical gradient from inverse rendering is effectively zero, so it would be hard to converge within a reasonable time. The log-log plot shows in addition that derivatives from the inverse rendering were slower than the sampled ones. On the contrary, our methods not" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.254, + 0.907, + 0.283 + ], + "angle": 0, + "content": "only smooth out the plateau, but also improve the speed of the derivative computation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.285, + 0.907, + 0.466 + ], + "angle": 0, + "content": "The drawback of constructing full Hessians is seen in the TEXTURE task, where OurH is more than 40 times slower than the HVP due to the matrix size\\(^{14}\\). For the most demanding task, CNN and CNN5, and even our best method\\(^{18}\\) is not able to reduce the error below \\(1:1000\\). At the \\(1:1000\\)-level, the more elaborate differentials\\(^{17}\\) is not faster than the basic one, but faster than first-order\\(^{16}\\). Finally, the advantage of our ability to compute HVPs is seen when comparing CNN and CNN5: while a CNN producing a single parameter is doable with Hessians\\(^{17}\\), but if we go to higher dimensions, the aggregate strategy pays off again and is markedly faster\\(^{20}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.468, + 0.906, + 0.559 + ], + "angle": 0, + "content": "In conclusion, albeit convergence varies with problem characteristics and dimensionality, we find that our best method, OurHVPA converges most reliably, beating the other algorithms by an average factor of \\(2.71 \\times\\) in wall-clock units. We show examples of task initialization and outcome, and additional evaluation in the supplemental." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.574, + 0.771, + 0.59 + ], + "angle": 0, + "content": "5. Limitations and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.906, + 0.734 + ], + "angle": 0, + "content": "Limitations. From the above experiments, it becomes apparent that our method performs well on lower-dimensional examples, but that its performance delta lessens with increasing problem dimensionality (see e.g. TEXTURE). We hypothesize that this is due to sparser sampling in higher dimensions and exacerbated MC noise and acknowledge that further research in this direction is needed. We include an analysis on how our method degrades under MC rendering noise in the supplemental." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Conclusion. The availability of second-order gradient information gives rise to an exciting avenue of future research in inverse rendering. While these algorithms are not new, their use in vision & graphics remains limited due to the convenience and widespread adoption of first-order gradient descent and the increased per-iteration cost that is usually associated with higher-order methods. In this work, our efficient estimators have shown the latter to be negligible in real-world optimization scenarios. We hope that this will inspire future research into unbiased, low-variance estimators of higher-order optimization methods." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28205" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.169 + ], + "angle": 0, + "content": "[1] Albert S Berahas, Richard H Byrd, and Jorge Nocedal. Derivative-free optimization of noisy functions via quasiNewton methods. SIAM Journal on Optimization, 29(2):965-993, 2019. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.171, + 0.484, + 0.211 + ], + "angle": 0, + "content": "[2] John P Boyd. Solving transcendental equations: the Chebyshev polynomial proxy and other numerical rootfinders, perturbation series, and oracles. SIAM, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.213, + 0.484, + 0.253 + ], + "angle": 0, + "content": "[3] Tony F Chan, Gene H Golub, and Pep Mulet. A nonlinear primal-dual method for total variation-based image restoration. SIAM Scient. Comput., 20(6):1964-1977, 1999. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.255, + 0.482, + 0.281 + ], + "angle": 0, + "content": "[4] Swarat Chaudhuri and Armando Solar-Lezama. Smooth interpretation. ACM Sigplan Notices, 45(6):279-291, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.283, + 0.482, + 0.322 + ], + "angle": 0, + "content": "[5] Thomas Deliot, Eric Heitz, and Laurent Belcour. Transforming a non-differentiable rasterizer into a differentiable one with stochastic gradient estimation. In ACM i3D, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.324, + 0.482, + 0.377 + ], + "angle": 0, + "content": "[6] Deshana Desai, Etai Shuchatowitz, Zhongshi Jiang, Teseo Schneider, and Daniele Panozzo. Acorns: An easy-to-use code generator for gradients and hessian. SoftwareX, 17: 100901, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.379, + 0.482, + 0.419 + ], + "angle": 0, + "content": "[7] Michael Fischer and Tobias Ritschel. Plateau-reduced differentiable path tracing. In CVPR, pages 4285-4294, 2023. 1, 3, 4, 5, 6, 10, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.421, + 0.482, + 0.461 + ], + "angle": 0, + "content": "[8] Michael Fischer and Tobias Ritschel. Zerograds: Learning local surrogates for non-differentiable graphics. ACM Trans. Graph. (Proc. SIGGRAPH), 43(4):1-15, 2024. 1, 3, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.463, + 0.482, + 0.502 + ], + "angle": 0, + "content": "[9] Roger Fletcher and Michael JD Powell. A rapidly convergent descent method for minimization. The computer journal, 6 (2):163-168, 1963. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.504, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[10] Nikolaus Hansen and Andreas Ostermeier. Adapting arbitrary normal mutation distributions in evolution strategies: The covariance matrix adaptation. In Proc. IEEE Evolutionary Computation, pages 312-317, 1996. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.56, + 0.482, + 0.599 + ], + "angle": 0, + "content": "[11] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engil Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In CVPR, pages 406-413, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.601, + 0.482, + 0.626 + ], + "angle": 0, + "content": "[12] James T Kajiya. The rendering equation. In Proc. SIGGRAPH, 1986. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.628, + 0.482, + 0.667 + ], + "angle": 0, + "content": "[13] Martin Kilian, Niloy J Mitra, and Helmut Pottmann. Geometric modeling in shape space. In SIGGRAPH, pages 64- es, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.669, + 0.482, + 0.696 + ], + "angle": 0, + "content": "[14] Dilip Krishnan and Rob Fergus. Fast image deconvolution using hyper-Laplacian priors. Proc. NeurIPS, 22, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.698, + 0.482, + 0.737 + ], + "angle": 0, + "content": "[15] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In Proc. ICCV, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.739, + 0.482, + 0.765 + ], + "angle": 0, + "content": "[16] Matthew M Loper and Michael J Black. OpenDR: An approximate differentiable renderer. In Proc. ECCV, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.767, + 0.482, + 0.805 + ], + "angle": 0, + "content": "[17] James Martens. New insights and perspectives on the natural gradient method. J Machine Learning Res, 21(146):1-76, 2020. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.808, + 0.482, + 0.861 + ], + "angle": 0, + "content": "[18] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Comm ACM, 65(1), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.863, + 0.482, + 0.89 + ], + "angle": 0, + "content": "[19] Baptiste Nicolet, Alec Jacobson, and Wenzel Jakob. Large steps in inverse rendering of geometry. ACM Trans. Graph." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.821, + 0.106 + ], + "angle": 0, + "content": "(Proc. SIGGRAPH Asis), 40(6):1-13, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.108, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[20] Merlin Nimier-David, Delio Vicini, Tizian Zeltner, and Wenzel Jakob. Mitsuba 2: A retargetable forward and inverse renderer. ACM Trans. Graph., 38(6), 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[21] Merlin Nimier-David, Sébastien Speierer, Benoit Ruiz, and Wenzel Jakob. Radiative backpropagation: an adjoint method for lightning-fast differentiable rendering. ACM Trans Graph, 39(4), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[22] Jorge Nocedal and Stephen J Wright. Numerical optimization. Springer, 1999. 1, 2, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[23] Art Owen and Yi Zhou. Safe and effective importance sampling. J American Statistical Assoc., 95(449), 2000. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.286 + ], + "angle": 0, + "content": "[24] Barak A Pearlmutter. Fast exact multiplication by the hessian. Neural computation, 6(1):147-160, 1994. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.314 + ], + "angle": 0, + "content": "[25] Michael JD Powell. Algorithms for nonlinear constraints that use lagrangian functions. Math. Prog., 14:224-248, 1978. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.355 + ], + "angle": 0, + "content": "[26] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv:2204.06125, 1(2):3, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.357, + 0.905, + 0.382 + ], + "angle": 0, + "content": "[27] Sebastian Ruder. An overview of gradient descent optimization algorithms. arXiv:1609.04747, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.384, + 0.905, + 0.424 + ], + "angle": 0, + "content": "[28] Pedro V Sander, John Snyder, Steven J Gortler, and Hugues Hoppe. Texture mapping progressive meshes. In Proc. SIGGRAPH, pages 409-416, 2001. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.426, + 0.905, + 0.465 + ], + "angle": 0, + "content": "[29] David F Shanno. Conditioning of quasi-newton methods for function minimization. Mathematics of computation, 24 (111):647-656, 1970. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.467, + 0.905, + 0.507 + ], + "angle": 0, + "content": "[30] Jack Sherman. Adjustment of an inverse matrix corresponding to changes in the elements of a given column or row of the original matrix. Annu. Math. Statist., 20:621, 1949. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.509, + 0.905, + 0.535 + ], + "angle": 0, + "content": "[31] Jonathan Richard Shewchuk. An introduction to the conjugate gradient method without the agonizing pain. 1994. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.537, + 0.905, + 0.576 + ], + "angle": 0, + "content": "[32] James C Spall. Multivariate stochastic approximation using a simultaneous perturbation gradient approximation. IEEE Trans Automatic Control, 37(3):332-341, 1992. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.578, + 0.905, + 0.618 + ], + "angle": 0, + "content": "[33] Eric Veach and Leonidas J Guibas. Optimally combining sampling techniques for Monte Carlo rendering. In Proc. SIGGRAPH, pages 419-428, 1995. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.62, + 0.905, + 0.672 + ], + "angle": 0, + "content": "[34] Delio Vicini, Sébastien Speierer, and Wenzel Jakob. Path replay backpropagation: differentiating light paths using constant memory and linear time. ACM Trans Graph, 40(4), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[35] Werbos. Backpropagation: Past and future. In IEEE 1988 Int Conf Neural Networks, pages 343-353. IEEE, 1988. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.703, + 0.905, + 0.742 + ], + "angle": 0, + "content": "[36] Manuel Werlberger, Werner Trobin, Thomas Pock, Andreas Wedel, Daniel Cremers, and Horst Bischof. Anisotropic huber-11 optical flow. In BMVC, volume 1, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.744, + 0.905, + 0.798 + ], + "angle": 0, + "content": "[37] Zhewei Yao, Amir Gholami, Sheng Shen, Mustafa Mustafa, Kurt Keutzer, and Michael Mahoney. Adahessian: An adaptive second order optimizer for machine learning. In Proc. AAAI, volume 35, pages 10665-10673, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.8, + 0.905, + 0.853 + ], + "angle": 0, + "content": "[38] Yizhou Yu, Paul Debevec, Jitendra Malik, and Tim Hawkins. Inverse global illumination: Recovering reflectance models of real scenes from photographs. In Proc. SIGGRAPH, pages 215-224, 1999. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.855, + 0.905, + 0.895 + ], + "angle": 0, + "content": "[39] Christopher Zach, Thomas Pock, and Horst Bischof. A duality based approach for realtime TV-L1 optical flow. In DAGM Pattern Recognition, pages 214-223, 2007. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.895 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28206" + } + ] +] \ No newline at end of file diff --git a/2025/Stochastic Gradient Estimation for Higher-Order Differentiable Rendering/95d3584a-ba8b-4c05-9f1b-c00fa69cfc88_origin.pdf b/2025/Stochastic Gradient Estimation for Higher-Order Differentiable Rendering/95d3584a-ba8b-4c05-9f1b-c00fa69cfc88_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..32206c150107fe4bdbb00536c0ed3a7b7dfc68cc --- /dev/null +++ b/2025/Stochastic Gradient Estimation for Higher-Order Differentiable Rendering/95d3584a-ba8b-4c05-9f1b-c00fa69cfc88_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:46c2bc7cd1f89009fd5d69f4774b36944421b554c05b228882c648499464e0e5 +size 1997249 diff --git a/2025/Stochastic Gradient Estimation for Higher-Order Differentiable Rendering/full.md b/2025/Stochastic Gradient Estimation for Higher-Order Differentiable Rendering/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3d18709846f70c13864f3474abe85b99156de817 --- /dev/null +++ b/2025/Stochastic Gradient Estimation for Higher-Order Differentiable Rendering/full.md @@ -0,0 +1,388 @@ +# Stochastic Gradient Estimation for Higher-order Differentiable Rendering + +Zican Wang +University College London +robert.wang.19@ucl.ac.uk + +Michael Fischer Adobe Research mifischer@adobe.com + +Tobias Ritschel +University College London +t.ritschel@ucl.ac.uk + +# Abstract + +We derive methods to compute higher-order differentials (Hessians and Hessian-vector products) of the rendering operator. Our approach is based on importance sampling of a convolution that represents the differentials of rendering parameters and shows to be applicable to both rasterization and path tracing. We further suggest an aggregate sampling strategy to importance-sample multiple dimensions of one convolution kernel simultaneously. We demonstrate that this information improves convergence when used in higher-order optimizers such as Newton or Conjugate Gradient relative to a gradient descent baseline in several inverse rendering tasks. + +# 1. Introduction + +Inverse rendering is concerned with optimizing the parameters of a scene to minimize a loss. This can be a useful tool when the true parameters of the scene are unknown or nontrivial to set for a human expert, and therefore need to be inferred from observations or measurements. Examples for such scenarios are multi-view reconstruction [18] or the recovery of illumination and reflectance properties [38]. Differentiable rendering has recently become a popular tool for this optimization, as automatic differentiation (AD) frameworks have become more widespread. + +However, differentiating the rendering process is far from trivial, as the rendering function [12] has zero and/or undefined gradients and step edges. Moreover, as many rendering operations rely on integration, discontinuities in the rendering pipeline cause problems for AD-engines, as we can no longer naively exchange the integral- and derivative-computations. Recent research therefore has designed a plethora of differentiable renderers [16], which compute gradients in various ways, typically incurring more [15, 20] or less [5, 7, 8] implementation and compute effort. + +A surprising insight is that, while these special renderers allow deriving gradients w.r.t. the scene parameters by differentiating (and back-propagating through) the rendering operator, virtually all of them are limited to first-order + +![](images/844af3533c6928d02b7c6fcd9d6770b066ec87fcdf554887207bf9d89748cdb0.jpg) +Figure 1. Our approach allows sampling the Hessian for inverse rendering, here for the task of rotating the cup around its horizontal $(x_0)$ and vertical axes $(x_{1})$ . The estimated positive and negative gradients are shown in blue and red, respectively. + +derivatives, and hence to expressing the gradient solely locally at a point in the parameter space. At the time of writing, no attempts to derive the Hessians required for higher-order inverse rendering methods have been published. However, decades of optimization research have shown the potential of higher-order methods in convergence and robustness [22]. + +In this paper, we argue that these benefits are also applicable to inverse rendering scenarios and show that they translate to net-gains in optimization time and performance in differentiable rendering. We tackle the Achilles' heel of higher-order optimizers - their increased per-iteration cost and computational / storage requirements - by developing efficient Monte Carlo (MC) estimators of the required higher-order quantities that can be importance-sampled with established techniques [23]. + +Our method leads to speedups of $2.71 \times$ over previous methods, and net-gains in optimization time, speed and robustness while only assuming the rendering operator to be a black box that can be point-sampled. + +# 2. Previous Work + +Gradient-based Optimization, the main workhorse of modern neural network training and inverse rendering, uses the gradient of the objective function to take iterative steps in the parameter space towards an improved solution until a (local) minimum is found. The exact nature of these steps varies with the optimizer that is being used [27]. + +First-order optimizers are simple to implement and cheap to execute but disregard important higher-order information about the shape of the objective function that could aid optimization, such as the second-order derivative. This usually comes at the expense of higher iteration counts, as many small steps are needed to converge to the solution. In contrast, higher-order optimizers incorporate information about the shape of the objective function, which usually allows them to take bigger steps in parameter space, leading to fewer iterations until convergence. + +For the specific case of second-order optimization, this additional information is often provided via the Hessian $\mathsf{H}$ (or approximations thereof) and the Hessian-vector product (HVP). The Hessian contains the second-order derivatives of the objective function w.r.t. the optimization parameters, and can be interpreted as the curvature of the objective function. As such, it can be used to inform the optimizer about how quickly the current gradient is changing and thus, in turn, about how large the optimization step should be. + +Notation The following will use lowercase boldface to denote vectors and uppercase sans to denote matrices, respectively. Operators, formally defined as functions acting on functions, will be denoted in uppercase Roman lettering in order to avoid confusion with regular functions. + +Gradient descent should be familiar to most readers, so we here recall only its terminology: The first-order Taylor expansion of the cost function $f$ at position $\theta$ is + +$$ +f (\boldsymbol {\theta}) \approx f (\boldsymbol {\theta} ^ {t}) + g (\boldsymbol {\theta} ^ {t}) ^ {\top} (\boldsymbol {\theta} - \boldsymbol {\theta} ^ {t}), +$$ + +where $g$ is the gradient $\nabla f = \mathrm{d}f / \mathrm{d}\theta$ of $f$ . The minimum is where the derivative is zero, which we can solve for as + +$$ +\frac {\mathrm {d}}{\mathrm {d} \boldsymbol {\theta}} f (\boldsymbol {\theta}) = 0 \approx g \left(\boldsymbol {\theta} ^ {t}\right). \tag {1} +$$ + +As $f$ is a linear function only locally, we only make small steps with step size $\gamma$ , by an update direction $-g(\pmb{\theta}^t)$ , as in + +$$ +\pmb {\theta} ^ {t + 1} = \pmb {\theta} ^ {t} - \gamma g (\pmb {\theta} ^ {t}). +$$ + +Newton's method is one of the most-used second-order optimizers. This term is derived from the second-order Taylor expansion of the objective $f$ around a point $\theta$ : + +$$ +f (\pmb {\theta}) \approx f (\pmb {\theta} ^ {t}) + g (\pmb {\theta} ^ {t}) ^ {\top} (\pmb {\theta} - \pmb {\theta} ^ {t}) + \frac {1}{2} (\pmb {\theta} - \pmb {\theta} ^ {t}) ^ {\top} \mathsf {H} (\pmb {\theta} ^ {t}) (\pmb {\theta} - \pmb {\theta} ^ {t}), +$$ + +where $\mathsf{H}$ is $\nabla^2 f = \mathrm{d}^2 f / \mathrm{d}^2\pmb{\theta}$ , the Hessian of $f$ . + +Ideally, we would like our update step to take us to an optimum. There, the derivative is necessarily zero: + +$$ +\frac {\mathrm {d}}{\mathrm {d} \boldsymbol {\theta}} f (\boldsymbol {\theta}) = 0 \approx g \left(\boldsymbol {\theta} ^ {t}\right) + \mathrm {H} \left(\boldsymbol {\theta} ^ {t}\right) \left(\boldsymbol {\theta} - \boldsymbol {\theta} ^ {t}\right). \tag {2} +$$ + +Solving for $\theta$ then yields the update rule + +$$ +\boldsymbol {\theta} ^ {t + 1} = \boldsymbol {\theta} ^ {t} - \gamma \mathrm {H} ^ {- 1} (\boldsymbol {\theta} ^ {t}) g (\boldsymbol {\theta} ^ {t}), \tag {3} +$$ + +where $\mathbf{v} = -\mathsf{H}^{-1}(\pmb{\theta}^t)g(\pmb{\theta}^t)$ is called the Newton direction. + +Newton's method requires the computation, storage, and inversion of the full Hessian, which quickly becomes a bottleneck in higher dimensions, as the Hessian for an $n$ -dimensional optimization problem is in $\mathbb{R}^{n\times n}$ . + +Newton Conjugate Gradient [31], upgrades Newton's method in two ways. First, it solves for Newton's direction iteratively as per the linear equation $\mathsf{H}(\pmb{\theta})\mathbf{v} = -g(\pmb{\theta})$ using conjugate directions [31]. Second, instead of an arbitrary step length $\gamma$ , it also decides the scalar $\alpha$ by which we move along this direction $\mathbf{v}$ . To derive $\alpha$ , first consider the Taylor expansion: + +$$ +\frac {\mathrm {d}}{\mathrm {d} \alpha} f (\boldsymbol {\theta} + \alpha \mathbf {v}) \approx g (\boldsymbol {\theta}) ^ {\top} \mathbf {v} + \alpha \mathbf {v} ^ {\top} H (\boldsymbol {\theta}) \mathbf {v} = 0 \tag {4} +$$ + +which can be re-arranged to + +$$ +\alpha = - \frac {g ^ {\mathsf {T}} (\boldsymbol {\theta}) \mathbf {v}}{\mathbf {v} ^ {\mathsf {T}} \mathsf {H} (\boldsymbol {\theta}) \mathbf {v}}. \tag {5} +$$ + +The update rule for the Newton Conjugate Gradient is more involved: First, we maintain the direction $\mathbf{v}$ and a residual $\mathbf{r}$ , which are initialized to be the gradient at the initial positions: + +$$ +\mathbf {v} ^ {0} = \mathbf {r} ^ {0} = - g (\boldsymbol {\theta} ^ {t}) \tag {6} +$$ + +We then find the $\alpha$ by equation Eq. 5 and update the point: + +$$ +\boldsymbol {\theta} ^ {t + 1} = \boldsymbol {\theta} ^ {t} + \alpha \mathbf {v} ^ {t}. \tag {7} +$$ + +The next direction is chosen by updating the residual $\mathbf{r}$ and computing a new conjugate search direction $\mathbf{v}$ with the Fletcher-Reeves formula [9]: + +$$ +\begin{array}{l} \mathbf {r} ^ {t + 1} = \mathbf {r} ^ {t} - \alpha \mathsf {H} (\boldsymbol {\theta} ^ {t}) \mathbf {v} ^ {t} \\ \beta = \frac {\mathbf {r} ^ {t + 1 , \top} \mathbf {r} ^ {t + 1}}{\mathbf {r} ^ {t , \top} \mathbf {r} ^ {t}} \\ \mathbf {v} ^ {t + 1} = \mathbf {r} ^ {t + 1} + \beta \mathbf {v} ^ {t}. \\ \end{array} +$$ + +Since Newton's method is derived via second-order Taylor expansion (approximated by a hyper-parabola), the derived Hessian may lead to either a maximum or a minimum depending on the local curvature of the function being approximated. This is determined by the positive-definiteness of the Hessian, and a negative-definite Hessian will diverge from the gradient's descent direction. To avoid this, one can ensure that the Hessian is always positive semi-definite using the Hessian modification technique and ensure step truncation to a trust region [22], $\alpha = \min (\alpha ,\Delta \| \mathbf{v}\|^{-1})$ , where $\Delta$ is the trust region radius (see Supplemental Sec. H). + +Hessian-vector product approximations go one step further by entirely avoiding to compute $\mathsf{H}$ when producing Hv. Pearlmutter [24] and Werbos [35] discuss different options to do so, but a simple option is central differences + +$$ +\mathsf {H} (\pmb {\theta}) \mathbf {v} \approx \lim _ {\varepsilon \to 0} \frac {g (\pmb {\theta} + \varepsilon \mathbf {v}) - g (\pmb {\theta} - \varepsilon \mathbf {v})}{2 \varepsilon}. \qquad (8) +$$ + +Hessians in vision and graphics have been applied to several optimization tasks, for instance, for total-variation denoising [3], texture parameterization and surface mappings with approximated Hessians [28], 3D shape manipulation [13], and image deconvolution tasks via (Hessian-free) Newton methods [14]. For optical flow estimation, both Zach et al. [39] and Werlberger et al. [36] use second-order information. Additionally, several works have explored Hessian-based algorithms in machine learning for vision tasks. Yao et al. [37] introduces a Hessian-based pruning method for CNNs for image classification, while Ramesh et al. [26] utilizes Hessians to improve the training of image generation models and Desai et al. [6] introduces an algorithm that derives Hessians for C code. + +Hessians in inverse rendering, however, have received surprisingly little attention, potentially due to their additional implementation overhead and computational complexity. In addition to the fact that conventional AD systems are mostly designed for first-order gradient computations, calculating the second-order information requires storing the whole forward- and first-derivative-graph in memory, which can lead to exponential memory growth. Nicolet et al. [19] are the closest to our work by approximating second-order steps for mesh reconstruction. However, they set the Hessian to the identity to avoid computational expense and instead work with a Laplacian regularizer, which works for their formulation and the case of mesh optimization, but is unclear how to translate to general problems. + +For derivative-free gradient estimators, it is equally unclear how second-order information would be computed. The zeroth-order estimators simultaneous perturbation stochastic approximation (SPSA) [32] and finite differences (FD) estimate a first-order gradient, whose second-order derivative naturally is zero. Extending these estimators to second-order information requires prohibitive amounts of function evaluations. ZeroGrads [8], which learns a neural network that fits the cost landscape, uses ReLU non-linearities, whose second-order derivative equally decays to zero. Covariance adaptation evolution strategy (CMA-ES) is often used when typical second-order optimization fails to converge [10]. It ensures that the Hessian is always positive definite, but if second-order derivative-based methods are successful, they are usually faster than CMA-ES. + +# 3. Our approach + +We first describe the computation of gradients using importance sampling of a combined gradient-smoothing operator from previous work [4, 7] which we then extend to Hessianians, and in a next step to Hessian-vector products. + +# 3.1. Background + +Rendering equation The rendering equation (RE) [12] describes the radiance $L$ leaving a point $\mathbf{x}$ in the scene into the direction $\omega_{\mathrm{o}}$ as + +$$ +L (\mathbf {x}, \omega_ {\mathrm {o}}; \boldsymbol {\theta}) = \int_ {\Omega} \underbrace {f _ {\mathrm {r}} \left(\omega_ {\mathrm {i}} , \omega_ {\mathrm {o}}\right) L \left(\mathbf {y} , \omega_ {\mathrm {i}} ; \boldsymbol {\theta}\right)} _ {R \left(\omega_ {\mathrm {i}}; \boldsymbol {\theta}\right)} \mathrm {d} \omega_ {\mathrm {i}}, \tag {9} +$$ + +where $\theta$ are the parameters of the scene we would like to optimize, such as object geometry, reflectance, or light emission. This integral over $\Omega$ , i.e., all incoming $\omega_{\mathrm{i}}$ directions that multiply the radiance field arriving from that direction from the closest other point $\mathbf{y}$ in direction $\omega_{\mathrm{i}}$ with the bidirectional reflectance distribution function (BRDF) $f_{\mathrm{r}}$ , has no analytical closed-form solution, and hence usually is approximated - both in forward and inverse rendering - via MC methods. We will shorthand the entire integrand as $R$ . + +Problem statement We would now like to apply a differential operator $\mathrm{D}$ to the RE, as in + +$$ +\mathrm {D} L (\mathbf {x}, \omega_ {\mathrm {i}}; \boldsymbol {\theta}). \tag {10} +$$ + +If $\mathrm{D}$ was the gradient operator $\frac{\partial L}{\partial \theta}$ , this would be differentiable rendering, for other operators, this becomes higher-order differentiable rendering. + +The trouble is that we cannot move D, be it gradient or higher-order, into the integral, as in many cases (e.g., for BRDF or spatial derivatives), the integrand is discontinuous in $\theta$ , so + +$$ +\mathrm {D} L (\mathbf {x}, \omega_ {\mathrm {i}}; \boldsymbol {\theta}) \neq \int_ {\Omega} \mathrm {D} R (\omega_ {\mathrm {i}}; \boldsymbol {\theta}) \mathrm {d} \omega_ {\mathrm {i}}. \tag {11} +$$ + +However, the right-hand side of the above expression is exactly the quantity that naively-applied AD computes [21, 34], leading to wrong gradients in (any-order) differentiable rendering. + +Solution The idea is to enforce the property that prevents differentiation - smoothness -, so that we actually can differentiate. To that end, assume a further linear operator $S$ that is smoothing any function in $\theta$ . This provides a smooth rendering equation $\bar{L}$ + +$$ +\bar {L} (\mathbf {x}, \omega_ {\mathrm {o}}; \boldsymbol {\theta}) = \int_ {\Omega} \mathrm {S} R \left(\omega_ {\mathrm {i}}; \boldsymbol {\theta}\right) \mathrm {d} \omega_ {\mathrm {i}}. \tag {12} +$$ + +![](images/a9ec933250838921e9b5c001d23d67037c74810a16b7660bd6c2dc79676bc363.jpg) +Figure 2. Comparison of classic (yellow) and smooth (red) gradients, as well as our Hessians (blue) on an inverse rendering problem to change the initial parameter $\theta$ so that the left triangle overlaps the right one. Classic gradients are zero almost everywhere (plateaus) except where the triangles already overlap. These methods do not converge. Smooth gradients point into the right direction, but make steps far from the optimum (dotted line). An update taking into account the curvature of the loss landscape lands at a point very close to the target. We incorporate this curvature information via our Hessians. + +Smoothing can be achieved by convolution, so for any function $f$ in any dimension + +$$ +\mathrm {S} f (\boldsymbol {\theta}) = \int_ {\Theta} \kappa (\tau) f (\boldsymbol {\theta} - \tau) \mathrm {d} \tau , \tag {13} +$$ + +where $\kappa$ is a smoothing kernel, such as a Gaussian, which we use in this work. This convolved integrand is now smooth, which, according to Leibniz' rule, allows us to move the differential operator into the integral + +$$ +\mathrm {D} \bar {L} (\mathbf {x}, \omega_ {\mathrm {o}}; \boldsymbol {\theta}) = \int_ {\Omega} \mathrm {D S} R \left(\omega_ {\mathrm {i}}; \boldsymbol {\theta}\right) \mathrm {d} \omega_ {\mathrm {i}}, \tag {14} +$$ + +which, after rearranging and expanding, yields an integral that can be approximated via MC: + +$$ +\mathrm {D} \bar {L} (\mathbf {x}, \omega_ {\mathrm {o}}; \boldsymbol {\theta}) = \int_ {\Omega} \int_ {\Theta} \mathrm {D} \kappa (\boldsymbol {\tau}) R \left(\omega_ {\mathrm {i}}; \boldsymbol {\theta} - \boldsymbol {\tau}\right) \mathrm {d} \boldsymbol {\tau} \mathrm {d} \omega_ {\mathrm {i}}. \tag {15} +$$ + +MC here means to take random samples from the product space of light paths and scene parameters. This works best if we can importance-sample for the integrand. The integrand here is a product of four terms. Sampling for the incoming radiance and BRDF terms has been investigated in the rendering community [33] and is not our consideration here, so we simply adopt these strategies. Sampling for the application of the differential operator to the smoothing kernel is the essence of the problem at hand. Depending on the choice of differential operator, we will derive three sampling strategies for the three resulting estimators next. + +Conclusion In conclusion, to perform efficient and practical any-order differentiation of the RE, we would need to implement two functions: first, a convolution kernel that combines smoothing and the desired differentiation, and second, a function to sample from that kernel for importance sampling. We will now do so for the gradient (Sec. 3.2), Hessian (Sec. 3.3) and HVP (Sec. 3.4). + +# 3.2. Gradients + +Operator: For first-order gradient descent, Fischer and Ritschel [7] have differentiated using the gradient operator + +$$ +\mathrm {D} ^ {\mathrm {G}} = \nabla = \partial / \partial \tau_ {i} \in (\mathbb {R} ^ {n} \rightarrow \mathbb {R}) \rightarrow (\mathbb {R} ^ {n} \rightarrow \mathbb {R} ^ {n}), \tag {16} +$$ + +![](images/06540bd316991f2321951872ad779f1dceafe2acf26d1847f48d812f8fd2268e.jpg) +Figure 3. Derivation of smooth differentiation by convolution for three differential operators (rows) involves three steps (columns): Defining the operator to combine smoothing and differentiation (1st col.), positivization and normalization to become a probability density function (PDF) (2nd col.), creating an inverse mapping (3rd col.), which finally allows sampling (4th col.). + +which maps a scalar function of $n$ dimensions to an $n$ -dimensional gradient vector field. + +The combination of smoothing and differentiation is + +$$ +\mathrm {D} ^ {\mathrm {G}} \kappa (\boldsymbol {\tau}) = \nabla_ {i} \mathcal {N} (\boldsymbol {\tau}, \sigma) = - \frac {\tau_ {i}}{\sigma^ {2}} \cdot \mathcal {N} (\boldsymbol {\tau}, \sigma). \tag {17} +$$ + +For a derivation, please see the supplemental, Sec. A. + +Sampling: For sampling a one-dimensional Gaussian gradient, we can use inverse transform sampling via the Smirnov transform [7]. To this end, the integrand has to be a PDF, i.e., positive and integrating-to-1 [23]. Subsequently, we compute the integral of a positivized version of $p^{\mathrm{G}}$ , the cumulative distribution function (CDF) $P^{\mathrm{G}}$ , and invert it as + +$$ +P ^ {\mathrm {G}, - 1} (\xi) = \left\{ \begin{array}{l l} - \sqrt {2 \sigma^ {2} \log (2 \xi)} & \text {i f} \xi \leq 0. 5, \\ + \sqrt {2 \sigma^ {2} \log (1 - \xi)} & \text {e l s e .} \end{array} \right. \tag {18} +$$ + +This derivation applies to the dimension which is being differentiated. The separability of multi-dimensional Gaussians ensures that the other dimensions remain a Gaussian distribution, and we can sample these dimensions independently. For a derivation of this, see Suppl. Sec. B. + +# 3.3. Hessians + +Operator: The differential operator for Hessian is + +$$ +\mathrm {D} ^ {\mathrm {H}} = \nabla^ {2} = \partial^ {2} / \partial \tau_ {i} \partial \tau_ {j} \in (\mathbb {R} ^ {n} \rightarrow \mathbb {R}) \rightarrow (\mathbb {R} ^ {n} \rightarrow \mathbb {R} ^ {n \times n}), \tag {19} +$$ + +which maps a scalar function in $n$ dimensions to its $n \times n$ -element Hessian field. The combination of second-order derivatives and Gaussian smoothing is + +$$ +\mathrm {D} ^ {\mathrm {H}} \kappa_ {i, j} (\boldsymbol {\tau}) = \left\{ \begin{array}{l l} \left(- \frac {1}{\sigma^ {2}} + \frac {\tau_ {i} ^ {2}}{\sigma^ {4}}\right) \cdot \mathcal {N} (\boldsymbol {\tau}, \sigma) & \text {i f} i = j, \\ \frac {\tau_ {i} \tau_ {j}}{\sigma^ {4}} \cdot \mathcal {N} (\boldsymbol {\tau}, \sigma) & \text {e l s e .} \end{array} \right. \tag {20} +$$ + +For a derivation, see supplemental Sec. C. + +Sampling: We first need to positivize as the function is signed, then scale the function so that it is a valid distribution. This is done differently for diagonal and off-diagonal elements. For the diagonal case, we construct the CDF of a second-order derivative $P_{ii}^{\mathrm{H}}$ of the 2D Gaussian: + +$$ +P _ {i i} ^ {\mathrm {H}} \left(\tau_ {i} = u\right) = \left\{ \begin{array}{c l} - \frac {u}{4 \sigma} \exp \left(\frac {1}{2} - \frac {u ^ {2}}{2 \sigma^ {2}}\right) & \text {i f} u < - \sigma , \\ 0. 5 + \frac {u}{4 \sigma} \exp \left(\frac {1}{2} - \frac {u ^ {2}}{2 \sigma^ {2}}\right) & \text {i f} u \in [ - \sigma , \sigma ] \\ 1 - \frac {u}{4 \sigma} \exp \left(\frac {1}{2} - \frac {u ^ {2}}{2 \sigma^ {2}}\right) & \text {i f} u > \sigma . \end{array} \right. \tag {21} +$$ + +See supplemental Sec. D for a derivation of this result. This is a transcendental equation whose inverse CDF cannot be expressed in closed form via elementary functions [2]. Instead, we 1D-tabulate its values on the range $10\sigma$ for every $i$ . The inverse function is found by searching in that range, and storing the pre-sorted inverse indices for $O(1)$ access. + +The off-diagonals are the product of two partial gradients that we have already derived in Eq. 17. Fortunately, the product of two independent distributions can be sampled by sampling each one independently. Thus, no extra derivation is required here. For sampling in higher dimensions, similar to the gradient sampler in Sec. B, the rest of the dimensions are sampled from a Gaussian distribution. + +In all methods, we exploit the symmetry in the Hessian matrix when sampling and estimating. + +# 3.4. Hessian-vector product + +Operator: As explained in Sec. 2, avoiding to store the full Hessian is possible by using HVPs: + +$$ +\mathrm {D} ^ {\mathrm {H V P}} = \nabla^ {2} \mathbf {v} = \partial^ {2} / \partial \tau_ {i} \partial \boldsymbol {\tau} \cdot \partial / \partial \boldsymbol {\tau}. \tag {22} +$$ + +In essence, a HVP is the directional gradient of the gradient. As there is one direction and the gradient has $n$ dimensions, the HVP is an $n$ -dimensional vector, too. In contrast, a Hessian is the non-directional gradient of the gradient, i.e., the gradient along all $n$ dimensions, and as such in $\mathbb{R}^{n \times n}$ . + +Sampling: To sample the smooth HVP, we simply need the directional central differences of an estimator of gradients, which we already have. So the estimator is the difference of two first-order estimators, evaluated at positions shifted from the current solution along the gradient direction [24, 35]: + +$$ +\mathrm {D} ^ {\mathrm {H V P}} \kappa (\boldsymbol {\tau}) = \frac {\mathrm {D} ^ {\mathrm {G}} \kappa (\boldsymbol {\tau} - \varepsilon \mathbf {v}) - \mathrm {D} ^ {\mathrm {G}} \kappa (\boldsymbol {\tau} + \varepsilon \mathbf {v})}{2 \varepsilon}. \tag {23} +$$ + +# 3.5. Aggregate + +We summarize the key property of all samplers - the number of function evaluations required - in Tab. 1. Function evaluations require execution of the black-box rendering engine, the most costly part of the optimization, and hence should be minimized. + +Table 1. Time complexity of different estimator variants. AIS: aggregated importance sampling. + +
no ISISAIS
Fin. Diff.nnn
FR22 [7]111
Our Grad.1n1
Our Hess.1n21
Our HVP1n1
+ +Classic finite differences take opposing samples in each dimension and hence, for an $n$ -dimensional problem, require $2n$ function evaluations for a single gradient sample. All other methods (second row in Tab. 1 onward) are based on MC, so we can get a gradient estimate using + +a fixed number of M function evaluations. This, however, comes at the cost of variance, which can be reduced with importance sampling, but that is done in all dimensions of the differential quantity independently, so it requires as many evaluations as these have elements (second column in Tab. 1). For a Hessian, this can be substantial, $n^2$ . + +As a compromise between low number of function evaluations and low variance, we propose "aggregate" importance sampling of the differential quantity (Fig. 4): instead of importance-sampling each dimension optimally but individually, we importance-sample w.r.t. the average of the convolution across all dimensions (combining several proposals by sampling from a mixture is similar to MIS using the balance heuristic [33]). This average convolution is a single function again, and with our combined sampling strategy can again be sampled with one function call per iteration (last column of Tab. 1). We implement this in two simple steps: first we randomly decide which element of the differential representation to choose (e.g., which matrix element out of $n^2$ in the Hessian), and then we compute the convolution sample for all $n^2$ but with the same function value that is only evaluated once. This increases variance, as the average kernel is not identical to the individual kernels but shares many properties, e.g., they are all zero in value at position 0. As a linear combination of unbiased es + +![](images/6345cca08a39cb8e61a8c87da21843ed9cdfbf2dad76a754eb918008183c771f.jpg) +Figure 4. Aggregate and non-aggregate sampling for a 2D optimization space and, consequently, a $2 \times 2$ Hessian: Without aggregation, importance sampling is done for each element of the Hessian independently (four points at different positions in each kernel), leading to four rendering calls and four different mugs. Each value is then weighted by the kernel (blue and red colors denote positive and negative). In aggregate sampling, we importance-sample the average of all four stencils, resulting in a single sample location and, hence, four times the same mug that can be rendered in one call, weighted with four different kernels. + +timators, this is still unbiased. We show in Sec. J that this pays off. Note that Tab. 1 is both time and space complexity, except for the aggregate importance sampling of the HVP, where the space complexity stays $O(n)$ while time complexity is reduced to $O(1)$ . + +# 4. Evaluation + +# 4.1. Methods + +Our evaluation compares the performance of gradients (input to gradient descent), Hessians and non-aggregate as well as aggregate Hessian-vector product (both input to higher-order optimizers). + +Baselines An established inverse rendering solution is Mitsuba with gradient descent. FR22 is using gradient descent with Fischer and Ritschel [7] MC gradients. OursG, OursH, OursHVP and OursHVPA are our approach for gradient, Hessians and Hessian-vector products, and aggregates, respectively. OursG is used with gradient descent (GD), the others in combination with conjugate gradient (CG). We have also experimented with LBFGS as detailed in the supplemental, but do not report it here as it almost never converges and never at competitive speed. All methods use the Adam optimizer. + +Tasks We tackle artificial problems with known analytic Hessians and Hessian-vector products as well as real inverse rendering tasks. As a simple first test case, we optimize the smooth, quadratic potential $(\mathrm{QUAD}^1)$ function $ax_0^2 +bx_1^2 +cx_0x_1$ in $\mathbb{R}^2$ , with the fixed variables $a = 5,b = 5,c = 7.5$ . + +Second, we optimize the classic plateau-demonstration task from [7]: the 2D position of some boxes is optimized + +to match a reference. This is already a much harder task, as there is a plateau in the cost landscape when the squares do not overlap (almost always the case in the initial configuration), as the image-space error does not change. We study the case of one $(\mathrm{Box2}^3)$ and five $(\mathrm{Box10}^5)$ squares, with two and ten dimensions to optimize, respectively. + +For real inverse problems, we study the optimization of reflectance, light, and geometry and render using Mitsuba. In the MUG7 task, we optimize the vertical rotation of a coffee cup such that it matches a reference. In the SHADOW9 task, we optimize the position of a sphere that is unobserved in the rendered image, such that the shadow it casts matches a reference shadow. In the BUNNY11 task, the $x$ and $z$ position and the rotation around the $z$ axis of the Stanford bunny are optimized. This task is set up specifically for Mitsuba to converge12, with no plateaus in the loss function. + +To test the scaling with dimensionality, we also optimize the pixels of a $32 \times 32$ texture (TEXTURE $^{13}$ ), with 1024 parameters. In addition, we have used our approach on a mesh optimization task, where we fit the 3D position of the vertices of a tessellated sphere to images of a target object. For SUZANNE, we fit 2,562 vertices, for BANANA, we fit the 2,737 vertices and their RGB vertex colors, similar to the DTU setup [11]. For trajectories of the optimization and the outcome, see Fig. 5 top and bottom rows, respectively. + +Finally, we learn a $\mathrm{CNN}^{15}$ to predict scene parameters from images using inverse rendering without scene parameter labels. In this case, the optimization is learning in the stricter sense: we tune parameters of a deep architecture instead of scene parameters directly. The CNN takes as input an image of a mug and produces as output the orientation for a mug. As a loss, these parameters are inserted into a renderer, differentiated by different approaches and compared to a target. The CNN has 267745 parameters, learned by differentiating through an image loss and rendering of the scene given the estimated parameters. Another CNN $(\mathrm{CNN}5^{19})$ is trained to predict the rotation, position, and color of the mug with 268773 parameters. + +Differentials of this function are computed as detailed in supplemental Sec. E, where we derive a grey-box approach that samples only the black-box part (the rendering) and combines its differentials with analytic differentials for the white-box part (the CNN). Importantly, this idea works for gradients, as well as for higher-order differentials. + +Metrics Our main measure of success is convergence speed in wall-clock and parameter difference to the true parameters (which we know in all inverse rendering tasks as we created the scenes). A secondary measure of success is the image difference, as during optimization, we consider the parameters hidden. All metrics are evaluated across an ensemble of 20 runs averaged across 10 steps in time. In convergence plots, the ensemble median is shown at every point in time, averaged across 20 time steps. + +Table 2. Quantitative results of different methods on different tasks (rows) and their convergence plots. We report convergence time in wall-clock units, in ratio to the overall best method, OurHVPA. In the numerical columns,.9 and .99 report the time taken to achieve 90 and $99\%$ error reduction from the initial starting configuration, respectively, while the bar plots graphically show these findings. The line plots report image- and parameter-space convergence in the left and right column, respectively, on a log-log scale. + +
TaskMethodImage errorParameter errorMitsuba FR22 CMA-ES OurG OurH OurHVP OurHVPA
.9.99.999.9.99.999
QUAD1Mitsu2—————ErrorLog error
FR221.802.382.982.203.173.90
CMA-ES0.150.130.140.130.140.23
OurG3.985.136.024.726.427.77
OurH1.171.171.151.181.161.17
OurHVP0.760.620.620.630.590.60
OurHVPA1.001.001.001.001.001.00
BOX23Mitsu——————ErrorLog error
FR223.242.492.572.782.892.70
CMA-ES5.513.453.158.135.514.13
OurG2.882.162.212.862.652.38
OurH0.930.951.080.760.841.84
OurHVP0.810.640.7040.740.720.72
OurHVPA1.001.001.001.001.001.00
BOX105Mitsu——————ErrorLog error
FR221.491.201.711.801.541.29
CMA-ES3.102.271.734.523.232.48
OurG2.181.771.532.532.261.88
OurH5.975.334.685.485.965.45
OurHVP1.511.441.6961.261.501.58
OurHVPA1.001.001.001.001.001.00
MUG7Mitsu——————ErrorLog error
FR225.925.695.463.286.536.29
CMA-ES2.951.971.823.322.382.20
OurG3.993.913.742.364.424.30
OurH1.571.091.531.001.081.20
OurHVP1.151.141.181.140.871.05
OurHVPA81.001.001.001.001.001.00
SHAD9Mitsu——————ErrorLog error
FR221.941.401.551.891.141.71
CMA-ES1.961.151.034.382.201.66
OurG1.371.181.132.381.471.34
OurH1.671.441.662.321.741.46
OurHVP1.140.910.961.371.131.07
OurHVPA1.001.001.001.001.001.00
BUNNY11Mitsu120.901.291.461.871.381.28ErrorLog error
FR220.750.751.120.780.630.71
CMA-ES1.160.950.851.270.910.85
OurG0.760.891.070.690.640.76
OurH1.291.231.291.180.801.07
OurHVP0.680.680.700.630.440.58
OurHVPA1.001.001.001.001.001.00
TEXTURE13Mitsu——————ErrorLog error
FR224.042.271.804.042.271.80
CMA-ES7.084.604.3511.167.477.44
OurG3.032.862.823.032.862.82
OurH440.1239.78—40.1239.78—
OurHVP4.014.004.004.014.004.00
OurHVPA1.001.001.001.001.001.00
CNN15Mitsu——————ErrorLog error
FR222.1116——1.451.322.71
CMA-ES———2.08——
OurG1.14——0.960.832.28
OurH0.770.77—0.740.770.74
OurHVP0.870.9317—0.880.910.87
OurHVPA1.001.00—181.001.001.00
CNN519Mitsu——————ErrorLog error
FR22——————
CMA-ES——————
OurG18.68——3.70——
OurH19.00——5.79——
OurHVP3.64201.46—1.491.58—
OurHVPA1.001.00—1.001.001.001.00
+ +.9 .99 .999 .9 99 .999 + +Log time + +Log time + +![](images/f82b9bb7c56e627efb356317d13685da7bda43592231eca79b1a52ee372358b3.jpg) +Figure 5. Inverse mesh optimization from renderings. Gradients use FR22, Hessians use OursHVPA, CMA-ES uses the pycma package. + +# 4.2. Results + +The main results of our evaluation are summarize in Tab. 2 where empty cells did not converge or the method is not applicable to that task. On average across tasks and methods, our premiere method oursHVPA speeds up the convergence by a factor of 2.71. + +In general, our methods, in particular OursHVP and OursHVPA, lead the level of error reduction across all time budgets, as seen by comparing the convergence curves vertically where they consistently decrease fastest. + +We benefit most in the artificial QUAD task, but MUG and SHADOW are both real rendering problems, where we are around seven times and $60\%$ faster, respectively. We see that OurG typically cannot outperform FR22, which uses an approximation of the correct high-dimensional gradient kernel. Doing it "right" only pays off when going to higher order. We also note that OursHVP outperforms OursH due to less function evaluations (as the metric is wall-clock - the iteration count for both is close). The log-log plots confirm that the higher-order variants enabled by our approach are much faster, but also converge slightly less stable. + +While tasks like QUAD and Box2 are not applicable to inverse renderers like Mitsuba², our methods treat any loss function as a black-box and retrieve derivatives via sampling. Comparing the performance of OurHVP and OurHVPA across a task with two and ten dimensions, we find the expected relation: OurHVP is faster in low dimensions⁴ than in higher dimensions for a similar task⁶. + +We notice that the OurHVPA is not out-performing the OurHVP and OurH in the MUG task. This is because there is a single scene parameter for this task, thus, the higher order derivative is all in order of one. With a higher variance for OurHVPA, it could be a little slower. The effect of using the aggregate starts to show from SHAD, where OurHVPA reaches the 1:1000 convergence the fastest10. + +For all the rendering tasks, only the BUNNY task converges for Mitsuba12. Since the other tasks have plateaus in the loss function, the analytical gradient from inverse rendering is effectively zero, so it would be hard to converge within a reasonable time. The log-log plot shows in addition that derivatives from the inverse rendering were slower than the sampled ones. On the contrary, our methods not + +only smooth out the plateau, but also improve the speed of the derivative computation. + +The drawback of constructing full Hessians is seen in the TEXTURE task, where OurH is more than 40 times slower than the HVP due to the matrix size $^{14}$ . For the most demanding task, CNN and CNN5, and even our best method $^{18}$ is not able to reduce the error below $1:1000$ . At the $1:1000$ -level, the more elaborate differentials $^{17}$ is not faster than the basic one, but faster than first-order $^{16}$ . Finally, the advantage of our ability to compute HVPs is seen when comparing CNN and CNN5: while a CNN producing a single parameter is doable with Hessians $^{17}$ , but if we go to higher dimensions, the aggregate strategy pays off again and is markedly faster $^{20}$ . + +In conclusion, albeit convergence varies with problem characteristics and dimensionality, we find that our best method, OurHVPA converges most reliably, beating the other algorithms by an average factor of $2.71 \times$ in wall-clock units. We show examples of task initialization and outcome, and additional evaluation in the supplemental. + +# 5. Limitations and Conclusion + +Limitations. From the above experiments, it becomes apparent that our method performs well on lower-dimensional examples, but that its performance delta lessens with increasing problem dimensionality (see e.g. TEXTURE). We hypothesize that this is due to sparser sampling in higher dimensions and exacerbated MC noise and acknowledge that further research in this direction is needed. We include an analysis on how our method degrades under MC rendering noise in the supplemental. + +Conclusion. The availability of second-order gradient information gives rise to an exciting avenue of future research in inverse rendering. While these algorithms are not new, their use in vision & graphics remains limited due to the convenience and widespread adoption of first-order gradient descent and the increased per-iteration cost that is usually associated with higher-order methods. In this work, our efficient estimators have shown the latter to be negligible in real-world optimization scenarios. We hope that this will inspire future research into unbiased, low-variance estimators of higher-order optimization methods. + +# References + +[1] Albert S Berahas, Richard H Byrd, and Jorge Nocedal. Derivative-free optimization of noisy functions via quasiNewton methods. SIAM Journal on Optimization, 29(2):965-993, 2019. 11 +[2] John P Boyd. Solving transcendental equations: the Chebyshev polynomial proxy and other numerical rootfinders, perturbation series, and oracles. SIAM, 2014. 5 +[3] Tony F Chan, Gene H Golub, and Pep Mulet. A nonlinear primal-dual method for total variation-based image restoration. SIAM Scient. Comput., 20(6):1964-1977, 1999. 3 +[4] Swarat Chaudhuri and Armando Solar-Lezama. Smooth interpretation. ACM Sigplan Notices, 45(6):279-291, 2010. 3 +[5] Thomas Deliot, Eric Heitz, and Laurent Belcour. Transforming a non-differentiable rasterizer into a differentiable one with stochastic gradient estimation. In ACM i3D, 2024. 1 +[6] Deshana Desai, Etai Shuchatowitz, Zhongshi Jiang, Teseo Schneider, and Daniele Panozzo. Acorns: An easy-to-use code generator for gradients and hessian. SoftwareX, 17: 100901, 2022. 3 +[7] Michael Fischer and Tobias Ritschel. Plateau-reduced differentiable path tracing. In CVPR, pages 4285-4294, 2023. 1, 3, 4, 5, 6, 10, 14 +[8] Michael Fischer and Tobias Ritschel. Zerograds: Learning local surrogates for non-differentiable graphics. ACM Trans. Graph. (Proc. SIGGRAPH), 43(4):1-15, 2024. 1, 3, 13 +[9] Roger Fletcher and Michael JD Powell. A rapidly convergent descent method for minimization. The computer journal, 6 (2):163-168, 1963. 2 +[10] Nikolaus Hansen and Andreas Ostermeier. Adapting arbitrary normal mutation distributions in evolution strategies: The covariance matrix adaptation. In Proc. IEEE Evolutionary Computation, pages 312-317, 1996. 3 +[11] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engil Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In CVPR, pages 406-413, 2014. 6 +[12] James T Kajiya. The rendering equation. In Proc. SIGGRAPH, 1986. 1, 3 +[13] Martin Kilian, Niloy J Mitra, and Helmut Pottmann. Geometric modeling in shape space. In SIGGRAPH, pages 64- es, 2007. 3 +[14] Dilip Krishnan and Rob Fergus. Fast image deconvolution using hyper-Laplacian priors. Proc. NeurIPS, 22, 2009. 3 +[15] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In Proc. ICCV, 2019. 1 +[16] Matthew M Loper and Michael J Black. OpenDR: An approximate differentiable renderer. In Proc. ECCV, 2014. 1 +[17] James Martens. New insights and perspectives on the natural gradient method. J Machine Learning Res, 21(146):1-76, 2020. 11 +[18] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Comm ACM, 65(1), 2021. 1 +[19] Baptiste Nicolet, Alec Jacobson, and Wenzel Jakob. Large steps in inverse rendering of geometry. ACM Trans. Graph. + +(Proc. SIGGRAPH Asis), 40(6):1-13, 2021. 3 +[20] Merlin Nimier-David, Delio Vicini, Tizian Zeltner, and Wenzel Jakob. Mitsuba 2: A retargetable forward and inverse renderer. ACM Trans. Graph., 38(6), 2019. 1 +[21] Merlin Nimier-David, Sébastien Speierer, Benoit Ruiz, and Wenzel Jakob. Radiative backpropagation: an adjoint method for lightning-fast differentiable rendering. ACM Trans Graph, 39(4), 2020. 3 +[22] Jorge Nocedal and Stephen J Wright. Numerical optimization. Springer, 1999. 1, 2, 11 +[23] Art Owen and Yi Zhou. Safe and effective importance sampling. J American Statistical Assoc., 95(449), 2000. 1, 4 +[24] Barak A Pearlmutter. Fast exact multiplication by the hessian. Neural computation, 6(1):147-160, 1994. 3, 5 +[25] Michael JD Powell. Algorithms for nonlinear constraints that use lagrangian functions. Math. Prog., 14:224-248, 1978. 11 +[26] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv:2204.06125, 1(2):3, 2022. 3 +[27] Sebastian Ruder. An overview of gradient descent optimization algorithms. arXiv:1609.04747, 2016. 1 +[28] Pedro V Sander, John Snyder, Steven J Gortler, and Hugues Hoppe. Texture mapping progressive meshes. In Proc. SIGGRAPH, pages 409-416, 2001. 3 +[29] David F Shanno. Conditioning of quasi-newton methods for function minimization. Mathematics of computation, 24 (111):647-656, 1970. 11 +[30] Jack Sherman. Adjustment of an inverse matrix corresponding to changes in the elements of a given column or row of the original matrix. Annu. Math. Statist., 20:621, 1949. 11 +[31] Jonathan Richard Shewchuk. An introduction to the conjugate gradient method without the agonizing pain. 1994. 2 +[32] James C Spall. Multivariate stochastic approximation using a simultaneous perturbation gradient approximation. IEEE Trans Automatic Control, 37(3):332-341, 1992. 3 +[33] Eric Veach and Leonidas J Guibas. Optimally combining sampling techniques for Monte Carlo rendering. In Proc. SIGGRAPH, pages 419-428, 1995. 4, 5 +[34] Delio Vicini, Sébastien Speierer, and Wenzel Jakob. Path replay backpropagation: differentiating light paths using constant memory and linear time. ACM Trans Graph, 40(4), 2021. 3 +[35] Werbos. Backpropagation: Past and future. In IEEE 1988 Int Conf Neural Networks, pages 343-353. IEEE, 1988. 3, 5 +[36] Manuel Werlberger, Werner Trobin, Thomas Pock, Andreas Wedel, Daniel Cremers, and Horst Bischof. Anisotropic huber-11 optical flow. In BMVC, volume 1, 2009. 3 +[37] Zhewei Yao, Amir Gholami, Sheng Shen, Mustafa Mustafa, Kurt Keutzer, and Michael Mahoney. Adahessian: An adaptive second order optimizer for machine learning. In Proc. AAAI, volume 35, pages 10665-10673, 2021. 3 +[38] Yizhou Yu, Paul Debevec, Jitendra Malik, and Tim Hawkins. Inverse global illumination: Recovering reflectance models of real scenes from photographs. In Proc. SIGGRAPH, pages 215-224, 1999. 1 +[39] Christopher Zach, Thomas Pock, and Horst Bischof. A duality based approach for realtime TV-L1 optical flow. In DAGM Pattern Recognition, pages 214-223, 2007. 3 \ No newline at end of file diff --git a/2025/Stochastic Gradient Estimation for Higher-Order Differentiable Rendering/images.zip b/2025/Stochastic Gradient Estimation for Higher-Order Differentiable Rendering/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d92fc28041bc28e415630116924a4086f8ebd0e3 --- /dev/null +++ b/2025/Stochastic Gradient Estimation for Higher-Order Differentiable Rendering/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:019fbacae8655ebd0112a5fd2ca0de5de8cd6922af071a1bc24187763c4dac63 +size 624999 diff --git a/2025/Stochastic Gradient Estimation for Higher-Order Differentiable Rendering/layout.json b/2025/Stochastic Gradient Estimation for Higher-Order Differentiable Rendering/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3070d4df7cc3f78c922832e122fc6591714d31e5 --- /dev/null +++ b/2025/Stochastic Gradient Estimation for Higher-Order Differentiable Rendering/layout.json @@ -0,0 +1,8770 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 75, + 103, + 536, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 103, + 536, + 122 + ], + "spans": [ + { + "bbox": [ + 75, + 103, + 536, + 122 + ], + "type": "text", + "content": "Stochastic Gradient Estimation for Higher-order Differentiable Rendering" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 93, + 144, + 227, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 144, + 227, + 185 + ], + "spans": [ + { + "bbox": [ + 93, + 144, + 227, + 185 + ], + "type": "text", + "content": "Zican Wang \nUniversity College London \nrobert.wang.19@ucl.ac.uk" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 252, + 144, + 358, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 144, + 358, + 185 + ], + "spans": [ + { + "bbox": [ + 252, + 144, + 358, + 185 + ], + "type": "text", + "content": "Michael Fischer Adobe Research mifischer@adobe.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 384, + 144, + 517, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 144, + 517, + 185 + ], + "spans": [ + { + "bbox": [ + 384, + 144, + 517, + 185 + ], + "type": "text", + "content": "Tobias Ritschel \nUniversity College London \nt.ritschel@ucl.ac.uk" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 239, + 297, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 297, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 297, + 383 + ], + "type": "text", + "content": "We derive methods to compute higher-order differentials (Hessians and Hessian-vector products) of the rendering operator. Our approach is based on importance sampling of a convolution that represents the differentials of rendering parameters and shows to be applicable to both rasterization and path tracing. We further suggest an aggregate sampling strategy to importance-sample multiple dimensions of one convolution kernel simultaneously. We demonstrate that this information improves convergence when used in higher-order optimizers such as Newton or Conjugate Gradient relative to a gradient descent baseline in several inverse rendering tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 406, + 135, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 406, + 135, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 406, + 135, + 418 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 424, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 424, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 296, + 544 + ], + "type": "text", + "content": "Inverse rendering is concerned with optimizing the parameters of a scene to minimize a loss. This can be a useful tool when the true parameters of the scene are unknown or nontrivial to set for a human expert, and therefore need to be inferred from observations or measurements. Examples for such scenarios are multi-view reconstruction [18] or the recovery of illumination and reflectance properties [38]. Differentiable rendering has recently become a popular tool for this optimization, as automatic differentiation (AD) frameworks have become more widespread." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 544, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 296, + 664 + ], + "type": "text", + "content": "However, differentiating the rendering process is far from trivial, as the rendering function [12] has zero and/or undefined gradients and step edges. Moreover, as many rendering operations rely on integration, discontinuities in the rendering pipeline cause problems for AD-engines, as we can no longer naively exchange the integral- and derivative-computations. Recent research therefore has designed a plethora of differentiable renderers [16], which compute gradients in various ways, typically incurring more [15, 20] or less [5, 7, 8] implementation and compute effort." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "A surprising insight is that, while these special renderers allow deriving gradients w.r.t. the scene parameters by differentiating (and back-propagating through) the rendering operator, virtually all of them are limited to first-order" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 316, + 213, + 553, + 315 + ], + "blocks": [ + { + "bbox": [ + 316, + 213, + 553, + 315 + ], + "lines": [ + { + "bbox": [ + 316, + 213, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 213, + 553, + 315 + ], + "type": "image", + "image_path": "844af3533c6928d02b7c6fcd9d6770b066ec87fcdf554887207bf9d89748cdb0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 318, + 555, + 362 + ], + "lines": [ + { + "bbox": [ + 313, + 318, + 555, + 362 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 555, + 362 + ], + "type": "text", + "content": "Figure 1. Our approach allows sampling the Hessian for inverse rendering, here for the task of rotating the cup around its horizontal " + }, + { + "bbox": [ + 313, + 318, + 555, + 362 + ], + "type": "inline_equation", + "content": "(x_0)" + }, + { + "bbox": [ + 313, + 318, + 555, + 362 + ], + "type": "text", + "content": " and vertical axes " + }, + { + "bbox": [ + 313, + 318, + 555, + 362 + ], + "type": "inline_equation", + "content": "(x_{1})" + }, + { + "bbox": [ + 313, + 318, + 555, + 362 + ], + "type": "text", + "content": ". The estimated positive and negative gradients are shown in blue and red, respectively." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 377, + 555, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 555, + 461 + ], + "type": "text", + "content": "derivatives, and hence to expressing the gradient solely locally at a point in the parameter space. At the time of writing, no attempts to derive the Hessians required for higher-order inverse rendering methods have been published. However, decades of optimization research have shown the potential of higher-order methods in convergence and robustness [22]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 461, + 554, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 554, + 568 + ], + "type": "text", + "content": "In this paper, we argue that these benefits are also applicable to inverse rendering scenarios and show that they translate to net-gains in optimization time and performance in differentiable rendering. We tackle the Achilles' heel of higher-order optimizers - their increased per-iteration cost and computational / storage requirements - by developing efficient Monte Carlo (MC) estimators of the required higher-order quantities that can be importance-sampled with established techniques [23]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 568, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 554, + 617 + ], + "type": "text", + "content": "Our method leads to speedups of " + }, + { + "bbox": [ + 313, + 568, + 554, + 617 + ], + "type": "inline_equation", + "content": "2.71 \\times" + }, + { + "bbox": [ + 313, + 568, + 554, + 617 + ], + "type": "text", + "content": " over previous methods, and net-gains in optimization time, speed and robustness while only assuming the rendering operator to be a black box that can be point-sampled." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 624, + 406, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 406, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 406, + 636 + ], + "type": "text", + "content": "2. Previous Work" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "content": "Gradient-based Optimization, the main workhorse of modern neural network training and inverse rendering, uses the gradient of the objective function to take iterative steps in the parameter space towards an improved solution until a (local) minimum is found. The exact nature of these steps varies with the optimizer that is being used [27]." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28198" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 191 + ], + "type": "text", + "content": "First-order optimizers are simple to implement and cheap to execute but disregard important higher-order information about the shape of the objective function that could aid optimization, such as the second-order derivative. This usually comes at the expense of higher iteration counts, as many small steps are needed to converge to the solution. In contrast, higher-order optimizers incorporate information about the shape of the objective function, which usually allows them to take bigger steps in parameter space, leading to fewer iterations until convergence." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 192, + 296, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 296, + 300 + ], + "type": "text", + "content": "For the specific case of second-order optimization, this additional information is often provided via the Hessian " + }, + { + "bbox": [ + 55, + 192, + 296, + 300 + ], + "type": "inline_equation", + "content": "\\mathsf{H}" + }, + { + "bbox": [ + 55, + 192, + 296, + 300 + ], + "type": "text", + "content": " (or approximations thereof) and the Hessian-vector product (HVP). The Hessian contains the second-order derivatives of the objective function w.r.t. the optimization parameters, and can be interpreted as the curvature of the objective function. As such, it can be used to inform the optimizer about how quickly the current gradient is changing and thus, in turn, about how large the optimization step should be." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 300, + 296, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 360 + ], + "type": "text", + "content": "Notation The following will use lowercase boldface to denote vectors and uppercase sans to denote matrices, respectively. Operators, formally defined as functions acting on functions, will be denoted in uppercase Roman lettering in order to avoid confusion with regular functions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 376, + 296, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 296, + 412 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 296, + 412 + ], + "type": "text", + "content": "Gradient descent should be familiar to most readers, so we here recall only its terminology: The first-order Taylor expansion of the cost function " + }, + { + "bbox": [ + 55, + 376, + 296, + 412 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 376, + 296, + 412 + ], + "type": "text", + "content": " at position " + }, + { + "bbox": [ + 55, + 376, + 296, + 412 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 376, + 296, + 412 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 421, + 244, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 421, + 244, + 435 + ], + "spans": [ + { + "bbox": [ + 107, + 421, + 244, + 435 + ], + "type": "interline_equation", + "content": "f (\\boldsymbol {\\theta}) \\approx f (\\boldsymbol {\\theta} ^ {t}) + g (\\boldsymbol {\\theta} ^ {t}) ^ {\\top} (\\boldsymbol {\\theta} - \\boldsymbol {\\theta} ^ {t}),", + "image_path": "2ef5f30f5610ecc19b36086e8b05f0130e1016982c32a87d22106913d03b9c03.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 445, + 295, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 295, + 468 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 445, + 295, + 468 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 445, + 295, + 468 + ], + "type": "text", + "content": " is the gradient " + }, + { + "bbox": [ + 55, + 445, + 295, + 468 + ], + "type": "inline_equation", + "content": "\\nabla f = \\mathrm{d}f / \\mathrm{d}\\theta" + }, + { + "bbox": [ + 55, + 445, + 295, + 468 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 445, + 295, + 468 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 445, + 295, + 468 + ], + "type": "text", + "content": ". The minimum is where the derivative is zero, which we can solve for as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 129, + 477, + 295, + 501 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 477, + 295, + 501 + ], + "spans": [ + { + "bbox": [ + 129, + 477, + 295, + 501 + ], + "type": "interline_equation", + "content": "\\frac {\\mathrm {d}}{\\mathrm {d} \\boldsymbol {\\theta}} f (\\boldsymbol {\\theta}) = 0 \\approx g \\left(\\boldsymbol {\\theta} ^ {t}\\right). \\tag {1}", + "image_path": "5eedcf593fcec0c5c5b0671840ec203b17110519e2a6177cd7338cb7c69649e0.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 509, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 533 + ], + "type": "text", + "content": "As " + }, + { + "bbox": [ + 55, + 509, + 295, + 533 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 509, + 295, + 533 + ], + "type": "text", + "content": " is a linear function only locally, we only make small steps with step size " + }, + { + "bbox": [ + 55, + 509, + 295, + 533 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 509, + 295, + 533 + ], + "type": "text", + "content": ", by an update direction " + }, + { + "bbox": [ + 55, + 509, + 295, + 533 + ], + "type": "inline_equation", + "content": "-g(\\pmb{\\theta}^t)" + }, + { + "bbox": [ + 55, + 509, + 295, + 533 + ], + "type": "text", + "content": ", as in" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 541, + 219, + 556 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 541, + 219, + 556 + ], + "spans": [ + { + "bbox": [ + 131, + 541, + 219, + 556 + ], + "type": "interline_equation", + "content": "\\pmb {\\theta} ^ {t + 1} = \\pmb {\\theta} ^ {t} - \\gamma g (\\pmb {\\theta} ^ {t}).", + "image_path": "e42f1c3dfb58fa305104bb75d7c4b3339ec24f5dc60efb48457c96730c7cb54c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 571, + 296, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 296, + 607 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 296, + 607 + ], + "type": "text", + "content": "Newton's method is one of the most-used second-order optimizers. This term is derived from the second-order Taylor expansion of the objective " + }, + { + "bbox": [ + 55, + 571, + 296, + 607 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 571, + 296, + 607 + ], + "type": "text", + "content": " around a point " + }, + { + "bbox": [ + 55, + 571, + 296, + 607 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 571, + 296, + 607 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 616, + 296, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 296, + 639 + ], + "type": "interline_equation", + "content": "f (\\pmb {\\theta}) \\approx f (\\pmb {\\theta} ^ {t}) + g (\\pmb {\\theta} ^ {t}) ^ {\\top} (\\pmb {\\theta} - \\pmb {\\theta} ^ {t}) + \\frac {1}{2} (\\pmb {\\theta} - \\pmb {\\theta} ^ {t}) ^ {\\top} \\mathsf {H} (\\pmb {\\theta} ^ {t}) (\\pmb {\\theta} - \\pmb {\\theta} ^ {t}),", + "image_path": "4989e41ef4b8a7fb9c4ebe55d9a4b519314df49b880669294c2fcdba8d51a837.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 647, + 246, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 246, + 660 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 246, + 660 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 647, + 246, + 660 + ], + "type": "inline_equation", + "content": "\\mathsf{H}" + }, + { + "bbox": [ + 55, + 647, + 246, + 660 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 55, + 647, + 246, + 660 + ], + "type": "inline_equation", + "content": "\\nabla^2 f = \\mathrm{d}^2 f / \\mathrm{d}^2\\pmb{\\theta}" + }, + { + "bbox": [ + 55, + 647, + 246, + 660 + ], + "type": "text", + "content": ", the Hessian of " + }, + { + "bbox": [ + 55, + 647, + 246, + 660 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 647, + 246, + 660 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 661, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 661, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 661, + 295, + 685 + ], + "type": "text", + "content": "Ideally, we would like our update step to take us to an optimum. There, the derivative is necessarily zero:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 93, + 693, + 296, + 718 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 693, + 296, + 718 + ], + "spans": [ + { + "bbox": [ + 93, + 693, + 296, + 718 + ], + "type": "interline_equation", + "content": "\\frac {\\mathrm {d}}{\\mathrm {d} \\boldsymbol {\\theta}} f (\\boldsymbol {\\theta}) = 0 \\approx g \\left(\\boldsymbol {\\theta} ^ {t}\\right) + \\mathrm {H} \\left(\\boldsymbol {\\theta} ^ {t}\\right) \\left(\\boldsymbol {\\theta} - \\boldsymbol {\\theta} ^ {t}\\right). \\tag {2}", + "image_path": "aad4bd2ff5bd0731b9c95b58e2c7aaffd188d67f5cb30c84ddb59a33bd8d4d3c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 72, + 479, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 479, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 479, + 84 + ], + "type": "text", + "content": "Solving for " + }, + { + "bbox": [ + 314, + 72, + 479, + 84 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 314, + 72, + 479, + 84 + ], + "type": "text", + "content": " then yields the update rule" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 372, + 91, + 553, + 105 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 91, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 372, + 91, + 553, + 105 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\theta} ^ {t + 1} = \\boldsymbol {\\theta} ^ {t} - \\gamma \\mathrm {H} ^ {- 1} (\\boldsymbol {\\theta} ^ {t}) g (\\boldsymbol {\\theta} ^ {t}), \\tag {3}", + "image_path": "8c5abb00c1b9dc509c40b5a8426194cc940c05a6e2f52f0d9811a806acf65180.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 112, + 552, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 112, + 552, + 125 + ], + "spans": [ + { + "bbox": [ + 313, + 112, + 552, + 125 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 112, + 552, + 125 + ], + "type": "inline_equation", + "content": "\\mathbf{v} = -\\mathsf{H}^{-1}(\\pmb{\\theta}^t)g(\\pmb{\\theta}^t)" + }, + { + "bbox": [ + 313, + 112, + 552, + 125 + ], + "type": "text", + "content": " is called the Newton direction." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 125, + 554, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 125, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 313, + 125, + 554, + 173 + ], + "type": "text", + "content": "Newton's method requires the computation, storage, and inversion of the full Hessian, which quickly becomes a bottleneck in higher dimensions, as the Hessian for an " + }, + { + "bbox": [ + 313, + 125, + 554, + 173 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 125, + 554, + 173 + ], + "type": "text", + "content": "-dimensional optimization problem is in " + }, + { + "bbox": [ + 313, + 125, + 554, + 173 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{n\\times n}" + }, + { + "bbox": [ + 313, + 125, + 554, + 173 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 186, + 554, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 554, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 554, + 270 + ], + "type": "text", + "content": "Newton Conjugate Gradient [31], upgrades Newton's method in two ways. First, it solves for Newton's direction iteratively as per the linear equation " + }, + { + "bbox": [ + 313, + 186, + 554, + 270 + ], + "type": "inline_equation", + "content": "\\mathsf{H}(\\pmb{\\theta})\\mathbf{v} = -g(\\pmb{\\theta})" + }, + { + "bbox": [ + 313, + 186, + 554, + 270 + ], + "type": "text", + "content": " using conjugate directions [31]. Second, instead of an arbitrary step length " + }, + { + "bbox": [ + 313, + 186, + 554, + 270 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 186, + 554, + 270 + ], + "type": "text", + "content": ", it also decides the scalar " + }, + { + "bbox": [ + 313, + 186, + 554, + 270 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 186, + 554, + 270 + ], + "type": "text", + "content": " by which we move along this direction " + }, + { + "bbox": [ + 313, + 186, + 554, + 270 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 313, + 186, + 554, + 270 + ], + "type": "text", + "content": ". To derive " + }, + { + "bbox": [ + 313, + 186, + 554, + 270 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 186, + 554, + 270 + ], + "type": "text", + "content": ", first consider the Taylor expansion:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 344, + 277, + 553, + 301 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 277, + 553, + 301 + ], + "spans": [ + { + "bbox": [ + 344, + 277, + 553, + 301 + ], + "type": "interline_equation", + "content": "\\frac {\\mathrm {d}}{\\mathrm {d} \\alpha} f (\\boldsymbol {\\theta} + \\alpha \\mathbf {v}) \\approx g (\\boldsymbol {\\theta}) ^ {\\top} \\mathbf {v} + \\alpha \\mathbf {v} ^ {\\top} H (\\boldsymbol {\\theta}) \\mathbf {v} = 0 \\tag {4}", + "image_path": "0b37b0b2e4d6d9b79eb7e1a475451da867dd5a65cd17bd5c403cdc16c2b01ebd.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 307, + 430, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 307, + 430, + 319 + ], + "spans": [ + { + "bbox": [ + 314, + 307, + 430, + 319 + ], + "type": "text", + "content": "which can be re-arranged to" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 397, + 325, + 553, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 325, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 397, + 325, + 553, + 353 + ], + "type": "interline_equation", + "content": "\\alpha = - \\frac {g ^ {\\mathsf {T}} (\\boldsymbol {\\theta}) \\mathbf {v}}{\\mathbf {v} ^ {\\mathsf {T}} \\mathsf {H} (\\boldsymbol {\\theta}) \\mathbf {v}}. \\tag {5}", + "image_path": "9b51330a6a218d497841c948e3a4ee39b7efe9dd8e6ac80889241d0701f6375b.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 358, + 554, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 554, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 554, + 403 + ], + "type": "text", + "content": "The update rule for the Newton Conjugate Gradient is more involved: First, we maintain the direction " + }, + { + "bbox": [ + 313, + 358, + 554, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 313, + 358, + 554, + 403 + ], + "type": "text", + "content": " and a residual " + }, + { + "bbox": [ + 313, + 358, + 554, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 313, + 358, + 554, + 403 + ], + "type": "text", + "content": ", which are initialized to be the gradient at the initial positions:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 394, + 405, + 553, + 419 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 405, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 394, + 405, + 553, + 419 + ], + "type": "interline_equation", + "content": "\\mathbf {v} ^ {0} = \\mathbf {r} ^ {0} = - g (\\boldsymbol {\\theta} ^ {t}) \\tag {6}", + "image_path": "d03c4c5abb05262782ca8cdfdd9ca0ed6a5bb90c001945e2fa87ad3726e3c26c.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 314, + 423, + 551, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 423, + 551, + 435 + ], + "spans": [ + { + "bbox": [ + 314, + 423, + 551, + 435 + ], + "type": "text", + "content": "We then find the " + }, + { + "bbox": [ + 314, + 423, + 551, + 435 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 314, + 423, + 551, + 435 + ], + "type": "text", + "content": " by equation Eq. 5 and update the point:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 395, + 441, + 553, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 441, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 395, + 441, + 553, + 455 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\theta} ^ {t + 1} = \\boldsymbol {\\theta} ^ {t} + \\alpha \\mathbf {v} ^ {t}. \\tag {7}", + "image_path": "c82ca49d3a44762ed5750f57b22b3303251dfd29587648acea68ee39dd1ea20c.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 464, + 554, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 554, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 554, + 499 + ], + "type": "text", + "content": "The next direction is chosen by updating the residual " + }, + { + "bbox": [ + 313, + 464, + 554, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 313, + 464, + 554, + 499 + ], + "type": "text", + "content": " and computing a new conjugate search direction " + }, + { + "bbox": [ + 313, + 464, + 554, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 313, + 464, + 554, + 499 + ], + "type": "text", + "content": " with the Fletcher-Reeves formula [9]:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 386, + 506, + 482, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 506, + 482, + 562 + ], + "spans": [ + { + "bbox": [ + 386, + 506, + 482, + 562 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {r} ^ {t + 1} = \\mathbf {r} ^ {t} - \\alpha \\mathsf {H} (\\boldsymbol {\\theta} ^ {t}) \\mathbf {v} ^ {t} \\\\ \\beta = \\frac {\\mathbf {r} ^ {t + 1 , \\top} \\mathbf {r} ^ {t + 1}}{\\mathbf {r} ^ {t , \\top} \\mathbf {r} ^ {t}} \\\\ \\mathbf {v} ^ {t + 1} = \\mathbf {r} ^ {t + 1} + \\beta \\mathbf {v} ^ {t}. \\\\ \\end{array}", + "image_path": "d3f6446a97759959ff14595b8fa0a4497f40c4567ea517475c20858fefb4ad7a.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "Since Newton's method is derived via second-order Taylor expansion (approximated by a hyper-parabola), the derived Hessian may lead to either a maximum or a minimum depending on the local curvature of the function being approximated. This is determined by the positive-definiteness of the Hessian, and a negative-definite Hessian will diverge from the gradient's descent direction. To avoid this, one can ensure that the Hessian is always positive semi-definite using the Hessian modification technique and ensure step truncation to a trust region [22], " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\alpha = \\min (\\alpha ,\\Delta \\| \\mathbf{v}\\|^{-1})" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " is the trust region radius (see Supplemental Sec. H)." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28199" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": "Hessian-vector product approximations go one step further by entirely avoiding to compute " + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "inline_equation", + "content": "\\mathsf{H}" + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": " when producing Hv. Pearlmutter [24] and Werbos [35] discuss different options to do so, but a simple option is central differences" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 94, + 133, + 296, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 133, + 296, + 159 + ], + "spans": [ + { + "bbox": [ + 94, + 133, + 296, + 159 + ], + "type": "interline_equation", + "content": "\\mathsf {H} (\\pmb {\\theta}) \\mathbf {v} \\approx \\lim _ {\\varepsilon \\to 0} \\frac {g (\\pmb {\\theta} + \\varepsilon \\mathbf {v}) - g (\\pmb {\\theta} - \\varepsilon \\mathbf {v})}{2 \\varepsilon}. \\qquad (8)", + "image_path": "4d68ebb49ee4904c00025d3660a24e0fa585c4ddbc9a7b5bf0b7e93b5bb79bbb.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 182, + 296, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 296, + 350 + ], + "type": "text", + "content": "Hessians in vision and graphics have been applied to several optimization tasks, for instance, for total-variation denoising [3], texture parameterization and surface mappings with approximated Hessians [28], 3D shape manipulation [13], and image deconvolution tasks via (Hessian-free) Newton methods [14]. For optical flow estimation, both Zach et al. [39] and Werlberger et al. [36] use second-order information. Additionally, several works have explored Hessian-based algorithms in machine learning for vision tasks. Yao et al. [37] introduces a Hessian-based pruning method for CNNs for image classification, while Ramesh et al. [26] utilizes Hessians to improve the training of image generation models and Desai et al. [6] introduces an algorithm that derives Hessians for C code." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 352, + 296, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 296, + 520 + ], + "type": "text", + "content": "Hessians in inverse rendering, however, have received surprisingly little attention, potentially due to their additional implementation overhead and computational complexity. In addition to the fact that conventional AD systems are mostly designed for first-order gradient computations, calculating the second-order information requires storing the whole forward- and first-derivative-graph in memory, which can lead to exponential memory growth. Nicolet et al. [19] are the closest to our work by approximating second-order steps for mesh reconstruction. However, they set the Hessian to the identity to avoid computational expense and instead work with a Laplacian regularizer, which works for their formulation and the case of mesh optimization, but is unclear how to translate to general problems." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "content": "For derivative-free gradient estimators, it is equally unclear how second-order information would be computed. The zeroth-order estimators simultaneous perturbation stochastic approximation (SPSA) [32] and finite differences (FD) estimate a first-order gradient, whose second-order derivative naturally is zero. Extending these estimators to second-order information requires prohibitive amounts of function evaluations. ZeroGrads [8], which learns a neural network that fits the cost landscape, uses ReLU non-linearities, whose second-order derivative equally decays to zero. Covariance adaptation evolution strategy (CMA-ES) is often used when typical second-order optimization fails to converge [10]. It ensures that the Hessian is always positive definite, but if second-order derivative-based methods are successful, they are usually faster than CMA-ES." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 72, + 403, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 403, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 403, + 85 + ], + "type": "text", + "content": "3. Our approach" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 89, + 555, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 138 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 138 + ], + "type": "text", + "content": "We first describe the computation of gradients using importance sampling of a combined gradient-smoothing operator from previous work [4, 7] which we then extend to Hessianians, and in a next step to Hessian-vector products." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 145, + 395, + 158 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 145, + 395, + 158 + ], + "spans": [ + { + "bbox": [ + 314, + 145, + 395, + 158 + ], + "type": "text", + "content": "3.1. Background" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 160, + 555, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 160, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 313, + 160, + 555, + 196 + ], + "type": "text", + "content": "Rendering equation The rendering equation (RE) [12] describes the radiance " + }, + { + "bbox": [ + 313, + 160, + 555, + 196 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 160, + 555, + 196 + ], + "type": "text", + "content": " leaving a point " + }, + { + "bbox": [ + 313, + 160, + 555, + 196 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 160, + 555, + 196 + ], + "type": "text", + "content": " in the scene into the direction " + }, + { + "bbox": [ + 313, + 160, + 555, + 196 + ], + "type": "inline_equation", + "content": "\\omega_{\\mathrm{o}}" + }, + { + "bbox": [ + 313, + 160, + 555, + 196 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 344, + 205, + 555, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 205, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 344, + 205, + 555, + 239 + ], + "type": "interline_equation", + "content": "L (\\mathbf {x}, \\omega_ {\\mathrm {o}}; \\boldsymbol {\\theta}) = \\int_ {\\Omega} \\underbrace {f _ {\\mathrm {r}} \\left(\\omega_ {\\mathrm {i}} , \\omega_ {\\mathrm {o}}\\right) L \\left(\\mathbf {y} , \\omega_ {\\mathrm {i}} ; \\boldsymbol {\\theta}\\right)} _ {R \\left(\\omega_ {\\mathrm {i}}; \\boldsymbol {\\theta}\\right)} \\mathrm {d} \\omega_ {\\mathrm {i}}, \\tag {9}", + "image_path": "f832746401fbe105cc39f684ee1725b3f10aa52dc447dc769226385983afe5f7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 248, + 555, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 248, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 248, + 555, + 357 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 248, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 248, + 555, + 357 + ], + "type": "text", + "content": " are the parameters of the scene we would like to optimize, such as object geometry, reflectance, or light emission. This integral over " + }, + { + "bbox": [ + 313, + 248, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 248, + 555, + 357 + ], + "type": "text", + "content": ", i.e., all incoming " + }, + { + "bbox": [ + 313, + 248, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\omega_{\\mathrm{i}}" + }, + { + "bbox": [ + 313, + 248, + 555, + 357 + ], + "type": "text", + "content": " directions that multiply the radiance field arriving from that direction from the closest other point " + }, + { + "bbox": [ + 313, + 248, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 313, + 248, + 555, + 357 + ], + "type": "text", + "content": " in direction " + }, + { + "bbox": [ + 313, + 248, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\omega_{\\mathrm{i}}" + }, + { + "bbox": [ + 313, + 248, + 555, + 357 + ], + "type": "text", + "content": " with the bidirectional reflectance distribution function (BRDF) " + }, + { + "bbox": [ + 313, + 248, + 555, + 357 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{r}}" + }, + { + "bbox": [ + 313, + 248, + 555, + 357 + ], + "type": "text", + "content": ", has no analytical closed-form solution, and hence usually is approximated - both in forward and inverse rendering - via MC methods. We will shorthand the entire integrand as " + }, + { + "bbox": [ + 313, + 248, + 555, + 357 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 248, + 555, + 357 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 373, + 554, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 554, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 554, + 397 + ], + "type": "text", + "content": "Problem statement We would now like to apply a differential operator " + }, + { + "bbox": [ + 313, + 373, + 554, + 397 + ], + "type": "inline_equation", + "content": "\\mathrm{D}" + }, + { + "bbox": [ + 313, + 373, + 554, + 397 + ], + "type": "text", + "content": " to the RE, as in" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 404, + 407, + 555, + 421 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 404, + 407, + 555, + 421 + ], + "spans": [ + { + "bbox": [ + 404, + 407, + 555, + 421 + ], + "type": "interline_equation", + "content": "\\mathrm {D} L (\\mathbf {x}, \\omega_ {\\mathrm {i}}; \\boldsymbol {\\theta}). \\tag {10}", + "image_path": "e6a55d4df549e2395bba403c672d78dafcc49bc965189cf17fd5d67b4978e255.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 430, + 554, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 554, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 554, + 466 + ], + "type": "text", + "content": "If " + }, + { + "bbox": [ + 313, + 430, + 554, + 466 + ], + "type": "inline_equation", + "content": "\\mathrm{D}" + }, + { + "bbox": [ + 313, + 430, + 554, + 466 + ], + "type": "text", + "content": " was the gradient operator " + }, + { + "bbox": [ + 313, + 430, + 554, + 466 + ], + "type": "inline_equation", + "content": "\\frac{\\partial L}{\\partial \\theta}" + }, + { + "bbox": [ + 313, + 430, + 554, + 466 + ], + "type": "text", + "content": ", this would be differentiable rendering, for other operators, this becomes higher-order differentiable rendering." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 467, + 555, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 555, + 514 + ], + "type": "text", + "content": "The trouble is that we cannot move D, be it gradient or higher-order, into the integral, as in many cases (e.g., for BRDF or spatial derivatives), the integrand is discontinuous in " + }, + { + "bbox": [ + 313, + 467, + 555, + 514 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 467, + 555, + 514 + ], + "type": "text", + "content": ", so" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 361, + 523, + 555, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 523, + 555, + 550 + ], + "spans": [ + { + "bbox": [ + 361, + 523, + 555, + 550 + ], + "type": "interline_equation", + "content": "\\mathrm {D} L (\\mathbf {x}, \\omega_ {\\mathrm {i}}; \\boldsymbol {\\theta}) \\neq \\int_ {\\Omega} \\mathrm {D} R (\\omega_ {\\mathrm {i}}; \\boldsymbol {\\theta}) \\mathrm {d} \\omega_ {\\mathrm {i}}. \\tag {11}", + "image_path": "969a45abffc95f4402de431fd0303e5cad59ae5175646f2e5f1e7100b684636d.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 557, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 605 + ], + "type": "text", + "content": "However, the right-hand side of the above expression is exactly the quantity that naively-applied AD computes [21, 34], leading to wrong gradients in (any-order) differentiable rendering." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 622, + 555, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 622, + 555, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 622, + 555, + 682 + ], + "type": "text", + "content": "Solution The idea is to enforce the property that prevents differentiation - smoothness -, so that we actually can differentiate. To that end, assume a further linear operator " + }, + { + "bbox": [ + 313, + 622, + 555, + 682 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 622, + 555, + 682 + ], + "type": "text", + "content": " that is smoothing any function in " + }, + { + "bbox": [ + 313, + 622, + 555, + 682 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 622, + 555, + 682 + ], + "type": "text", + "content": ". This provides a smooth rendering equation " + }, + { + "bbox": [ + 313, + 622, + 555, + 682 + ], + "type": "inline_equation", + "content": "\\bar{L}" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 366, + 691, + 555, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 691, + 555, + 717 + ], + "spans": [ + { + "bbox": [ + 366, + 691, + 555, + 717 + ], + "type": "interline_equation", + "content": "\\bar {L} (\\mathbf {x}, \\omega_ {\\mathrm {o}}; \\boldsymbol {\\theta}) = \\int_ {\\Omega} \\mathrm {S} R \\left(\\omega_ {\\mathrm {i}}; \\boldsymbol {\\theta}\\right) \\mathrm {d} \\omega_ {\\mathrm {i}}. \\tag {12}", + "image_path": "92dbe5ffb5c37dcb7aace7b17633693a73f8e0f37249dd6ad356dba014c69301.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28200" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 558, + 153 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 558, + 153 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 558, + 153 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 558, + 153 + ], + "type": "image", + "image_path": "a9ec933250838921e9b5c001d23d67037c74810a16b7660bd6c2dc79676bc363.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 160, + 555, + 216 + ], + "lines": [ + { + "bbox": [ + 55, + 160, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 160, + 555, + 216 + ], + "type": "text", + "content": "Figure 2. Comparison of classic (yellow) and smooth (red) gradients, as well as our Hessians (blue) on an inverse rendering problem to change the initial parameter " + }, + { + "bbox": [ + 55, + 160, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 160, + 555, + 216 + ], + "type": "text", + "content": " so that the left triangle overlaps the right one. Classic gradients are zero almost everywhere (plateaus) except where the triangles already overlap. These methods do not converge. Smooth gradients point into the right direction, but make steps far from the optimum (dotted line). An update taking into account the curvature of the loss landscape lands at a point very close to the target. We incorporate this curvature information via our Hessians." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 232, + 295, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 295, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 295, + 255 + ], + "type": "text", + "content": "Smoothing can be achieved by convolution, so for any function " + }, + { + "bbox": [ + 55, + 232, + 295, + 255 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 232, + 295, + 255 + ], + "type": "text", + "content": " in any dimension" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 110, + 258, + 294, + 283 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 258, + 294, + 283 + ], + "spans": [ + { + "bbox": [ + 110, + 258, + 294, + 283 + ], + "type": "interline_equation", + "content": "\\mathrm {S} f (\\boldsymbol {\\theta}) = \\int_ {\\Theta} \\kappa (\\tau) f (\\boldsymbol {\\theta} - \\tau) \\mathrm {d} \\tau , \\tag {13}", + "image_path": "d706f126c7e7c594878089b9167b6e9152aafc5d3e868667adefa25fa861e657.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 287, + 296, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 287, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 296, + 335 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 287, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 55, + 287, + 296, + 335 + ], + "type": "text", + "content": " is a smoothing kernel, such as a Gaussian, which we use in this work. This convolved integrand is now smooth, which, according to Leibniz' rule, allows us to move the differential operator into the integral" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 98, + 338, + 295, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 338, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 98, + 338, + 295, + 365 + ], + "type": "interline_equation", + "content": "\\mathrm {D} \\bar {L} (\\mathbf {x}, \\omega_ {\\mathrm {o}}; \\boldsymbol {\\theta}) = \\int_ {\\Omega} \\mathrm {D S} R \\left(\\omega_ {\\mathrm {i}}; \\boldsymbol {\\theta}\\right) \\mathrm {d} \\omega_ {\\mathrm {i}}, \\tag {14}", + "image_path": "cb2f6caf41c328de1e84f0ff4a09be579f8fa04b5811d2432f667f7c20e78e21.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 367, + 296, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 296, + 392 + ], + "type": "text", + "content": "which, after rearranging and expanding, yields an integral that can be approximated via MC:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 395, + 295, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 395, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 61, + 395, + 295, + 420 + ], + "type": "interline_equation", + "content": "\\mathrm {D} \\bar {L} (\\mathbf {x}, \\omega_ {\\mathrm {o}}; \\boldsymbol {\\theta}) = \\int_ {\\Omega} \\int_ {\\Theta} \\mathrm {D} \\kappa (\\boldsymbol {\\tau}) R \\left(\\omega_ {\\mathrm {i}}; \\boldsymbol {\\theta} - \\boldsymbol {\\tau}\\right) \\mathrm {d} \\boldsymbol {\\tau} \\mathrm {d} \\omega_ {\\mathrm {i}}. \\tag {15}", + "image_path": "5504b0e7e36839d65062b8afa9bf825242621eb9c6284302435adccb89740cf8.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 423, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 296, + 555 + ], + "type": "text", + "content": "MC here means to take random samples from the product space of light paths and scene parameters. This works best if we can importance-sample for the integrand. The integrand here is a product of four terms. Sampling for the incoming radiance and BRDF terms has been investigated in the rendering community [33] and is not our consideration here, so we simply adopt these strategies. Sampling for the application of the differential operator to the smoothing kernel is the essence of the problem at hand. Depending on the choice of differential operator, we will derive three sampling strategies for the three resulting estimators next." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 568, + 296, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 296, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 296, + 652 + ], + "type": "text", + "content": "Conclusion In conclusion, to perform efficient and practical any-order differentiation of the RE, we would need to implement two functions: first, a convolution kernel that combines smoothing and the desired differentiation, and second, a function to sample from that kernel for importance sampling. We will now do so for the gradient (Sec. 3.2), Hessian (Sec. 3.3) and HVP (Sec. 3.4)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 658, + 126, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 126, + 669 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 126, + 669 + ], + "type": "text", + "content": "3.2. Gradients" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 672, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 295, + 696 + ], + "type": "text", + "content": "Operator: For first-order gradient descent, Fischer and Ritschel [7] have differentiated using the gradient operator" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 700, + 295, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 700, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 700, + 295, + 714 + ], + "type": "interline_equation", + "content": "\\mathrm {D} ^ {\\mathrm {G}} = \\nabla = \\partial / \\partial \\tau_ {i} \\in (\\mathbb {R} ^ {n} \\rightarrow \\mathbb {R}) \\rightarrow (\\mathbb {R} ^ {n} \\rightarrow \\mathbb {R} ^ {n}), \\tag {16}", + "image_path": "2af59f596fc0ff4b76b523736b0bd0074d2dbf3891cb16b24e01d1749d807f56.jpg" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 230, + 553, + 371 + ], + "blocks": [ + { + "bbox": [ + 317, + 230, + 553, + 371 + ], + "lines": [ + { + "bbox": [ + 317, + 230, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 317, + 230, + 553, + 371 + ], + "type": "image", + "image_path": "06540bd316991f2321951872ad779f1dceafe2acf26d1847f48d812f8fd2268e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 374, + 555, + 440 + ], + "lines": [ + { + "bbox": [ + 313, + 374, + 555, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 555, + 440 + ], + "type": "text", + "content": "Figure 3. Derivation of smooth differentiation by convolution for three differential operators (rows) involves three steps (columns): Defining the operator to combine smoothing and differentiation (1st col.), positivization and normalization to become a probability density function (PDF) (2nd col.), creating an inverse mapping (3rd col.), which finally allows sampling (4th col.)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 453, + 554, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 554, + 475 + ], + "type": "text", + "content": "which maps a scalar function of " + }, + { + "bbox": [ + 313, + 453, + 554, + 475 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 453, + 554, + 475 + ], + "type": "text", + "content": " dimensions to an " + }, + { + "bbox": [ + 313, + 453, + 554, + 475 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 453, + 554, + 475 + ], + "type": "text", + "content": "-dimensional gradient vector field." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 326, + 476, + 538, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 476, + 538, + 488 + ], + "spans": [ + { + "bbox": [ + 326, + 476, + 538, + 488 + ], + "type": "text", + "content": "The combination of smoothing and differentiation is" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 337, + 493, + 553, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 493, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 337, + 493, + 553, + 514 + ], + "type": "interline_equation", + "content": "\\mathrm {D} ^ {\\mathrm {G}} \\kappa (\\boldsymbol {\\tau}) = \\nabla_ {i} \\mathcal {N} (\\boldsymbol {\\tau}, \\sigma) = - \\frac {\\tau_ {i}}{\\sigma^ {2}} \\cdot \\mathcal {N} (\\boldsymbol {\\tau}, \\sigma). \\tag {17}", + "image_path": "2912554943fa17fa87009bcb27111ff6f78b4abcd03abea622f6a5b4afd8b336.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 326, + 518, + 542, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 518, + 542, + 529 + ], + "spans": [ + { + "bbox": [ + 326, + 518, + 542, + 529 + ], + "type": "text", + "content": "For a derivation, please see the supplemental, Sec. A." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 530, + 554, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 554, + 601 + ], + "type": "text", + "content": "Sampling: For sampling a one-dimensional Gaussian gradient, we can use inverse transform sampling via the Smirnov transform [7]. To this end, the integrand has to be a PDF, i.e., positive and integrating-to-1 [23]. Subsequently, we compute the integral of a positivized version of " + }, + { + "bbox": [ + 313, + 530, + 554, + 601 + ], + "type": "inline_equation", + "content": "p^{\\mathrm{G}}" + }, + { + "bbox": [ + 313, + 530, + 554, + 601 + ], + "type": "text", + "content": ", the cumulative distribution function (CDF) " + }, + { + "bbox": [ + 313, + 530, + 554, + 601 + ], + "type": "inline_equation", + "content": "P^{\\mathrm{G}}" + }, + { + "bbox": [ + 313, + 530, + 554, + 601 + ], + "type": "text", + "content": ", and invert it as" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 328, + 616, + 553, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 616, + 553, + 649 + ], + "spans": [ + { + "bbox": [ + 328, + 616, + 553, + 649 + ], + "type": "interline_equation", + "content": "P ^ {\\mathrm {G}, - 1} (\\xi) = \\left\\{ \\begin{array}{l l} - \\sqrt {2 \\sigma^ {2} \\log (2 \\xi)} & \\text {i f} \\xi \\leq 0. 5, \\\\ + \\sqrt {2 \\sigma^ {2} \\log (1 - \\xi)} & \\text {e l s e .} \\end{array} \\right. \\tag {18}", + "image_path": "f613e68928e5bf1d6fa42c75e269f79c37faf4da9d64fb58498f21dae5b20f42.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "This derivation applies to the dimension which is being differentiated. The separability of multi-dimensional Gaussians ensures that the other dimensions remain a Gaussian distribution, and we can sample these dimensions independently. For a derivation of this, see Suppl. Sec. B." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28201" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 119, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 119, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 119, + 83 + ], + "type": "text", + "content": "3.3. Hessians" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 87, + 274, + 99 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 87, + 274, + 99 + ], + "spans": [ + { + "bbox": [ + 55, + 87, + 274, + 99 + ], + "type": "text", + "content": "Operator: The differential operator for Hessian is" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 104, + 294, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 104, + 294, + 129 + ], + "spans": [ + { + "bbox": [ + 58, + 104, + 294, + 129 + ], + "type": "interline_equation", + "content": "\\mathrm {D} ^ {\\mathrm {H}} = \\nabla^ {2} = \\partial^ {2} / \\partial \\tau_ {i} \\partial \\tau_ {j} \\in (\\mathbb {R} ^ {n} \\rightarrow \\mathbb {R}) \\rightarrow (\\mathbb {R} ^ {n} \\rightarrow \\mathbb {R} ^ {n \\times n}), \\tag {19}", + "image_path": "44901e3a8c0e885113f10ca8fbf644e01bd0e0baaf02079b4641aa6c0c0374ca.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 130, + 295, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 130, + 295, + 166 + ], + "spans": [ + { + "bbox": [ + 55, + 130, + 295, + 166 + ], + "type": "text", + "content": "which maps a scalar function in " + }, + { + "bbox": [ + 55, + 130, + 295, + 166 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 130, + 295, + 166 + ], + "type": "text", + "content": " dimensions to its " + }, + { + "bbox": [ + 55, + 130, + 295, + 166 + ], + "type": "inline_equation", + "content": "n \\times n" + }, + { + "bbox": [ + 55, + 130, + 295, + 166 + ], + "type": "text", + "content": "-element Hessian field. The combination of second-order derivatives and Gaussian smoothing is" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 171, + 294, + 230 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 171, + 294, + 230 + ], + "spans": [ + { + "bbox": [ + 67, + 171, + 294, + 230 + ], + "type": "interline_equation", + "content": "\\mathrm {D} ^ {\\mathrm {H}} \\kappa_ {i, j} (\\boldsymbol {\\tau}) = \\left\\{ \\begin{array}{l l} \\left(- \\frac {1}{\\sigma^ {2}} + \\frac {\\tau_ {i} ^ {2}}{\\sigma^ {4}}\\right) \\cdot \\mathcal {N} (\\boldsymbol {\\tau}, \\sigma) & \\text {i f} i = j, \\\\ \\frac {\\tau_ {i} \\tau_ {j}}{\\sigma^ {4}} \\cdot \\mathcal {N} (\\boldsymbol {\\tau}, \\sigma) & \\text {e l s e .} \\end{array} \\right. \\tag {20}", + "image_path": "29914d784ccd7e4bef884d38e9d40a4d0da5561c470fab001767d6349b4371b7.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 231, + 226, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 231, + 226, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 226, + 242 + ], + "type": "text", + "content": "For a derivation, see supplemental Sec. C." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 243, + 296, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 296, + 302 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 296, + 302 + ], + "type": "text", + "content": "Sampling: We first need to positivize as the function is signed, then scale the function so that it is a valid distribution. This is done differently for diagonal and off-diagonal elements. For the diagonal case, we construct the CDF of a second-order derivative " + }, + { + "bbox": [ + 55, + 243, + 296, + 302 + ], + "type": "inline_equation", + "content": "P_{ii}^{\\mathrm{H}}" + }, + { + "bbox": [ + 55, + 243, + 296, + 302 + ], + "type": "text", + "content": " of the 2D Gaussian:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 308, + 296, + 402 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 296, + 402 + ], + "type": "interline_equation", + "content": "P _ {i i} ^ {\\mathrm {H}} \\left(\\tau_ {i} = u\\right) = \\left\\{ \\begin{array}{c l} - \\frac {u}{4 \\sigma} \\exp \\left(\\frac {1}{2} - \\frac {u ^ {2}}{2 \\sigma^ {2}}\\right) & \\text {i f} u < - \\sigma , \\\\ 0. 5 + \\frac {u}{4 \\sigma} \\exp \\left(\\frac {1}{2} - \\frac {u ^ {2}}{2 \\sigma^ {2}}\\right) & \\text {i f} u \\in [ - \\sigma , \\sigma ] \\\\ 1 - \\frac {u}{4 \\sigma} \\exp \\left(\\frac {1}{2} - \\frac {u ^ {2}}{2 \\sigma^ {2}}\\right) & \\text {i f} u > \\sigma . \\end{array} \\right. \\tag {21}", + "image_path": "077eb8437a0a10f6c2aeaac1068fd2f85e3a05de83704edf972de452adc78e22.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 402, + 295, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 295, + 474 + ], + "type": "text", + "content": "See supplemental Sec. D for a derivation of this result. This is a transcendental equation whose inverse CDF cannot be expressed in closed form via elementary functions [2]. Instead, we 1D-tabulate its values on the range " + }, + { + "bbox": [ + 55, + 402, + 295, + 474 + ], + "type": "inline_equation", + "content": "10\\sigma" + }, + { + "bbox": [ + 55, + 402, + 295, + 474 + ], + "type": "text", + "content": " for every " + }, + { + "bbox": [ + 55, + 402, + 295, + 474 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 402, + 295, + 474 + ], + "type": "text", + "content": ". The inverse function is found by searching in that range, and storing the pre-sorted inverse indices for " + }, + { + "bbox": [ + 55, + 402, + 295, + 474 + ], + "type": "inline_equation", + "content": "O(1)" + }, + { + "bbox": [ + 55, + 402, + 295, + 474 + ], + "type": "text", + "content": " access." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 474, + 295, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 557 + ], + "type": "text", + "content": "The off-diagonals are the product of two partial gradients that we have already derived in Eq. 17. Fortunately, the product of two independent distributions can be sampled by sampling each one independently. Thus, no extra derivation is required here. For sampling in higher dimensions, similar to the gradient sampler in Sec. B, the rest of the dimensions are sampled from a Gaussian distribution." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 557, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 582 + ], + "type": "text", + "content": "In all methods, we exploit the symmetry in the Hessian matrix when sampling and estimating." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 588, + 187, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 187, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 187, + 601 + ], + "type": "text", + "content": "3.4. Hessian-vector product" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 602, + 295, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 295, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 295, + 627 + ], + "type": "text", + "content": "Operator: As explained in Sec. 2, avoiding to store the full Hessian is possible by using HVPs:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 100, + 632, + 295, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 632, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 100, + 632, + 295, + 647 + ], + "type": "interline_equation", + "content": "\\mathrm {D} ^ {\\mathrm {H V P}} = \\nabla^ {2} \\mathbf {v} = \\partial^ {2} / \\partial \\tau_ {i} \\partial \\boldsymbol {\\tau} \\cdot \\partial / \\partial \\boldsymbol {\\tau}. \\tag {22}", + "image_path": "52a9f142699a07385fd4d0d3bbf12f06ba2bf7d1c0625edb884d7234e5b4815e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "content": "In essence, a HVP is the directional gradient of the gradient. As there is one direction and the gradient has " + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "content": " dimensions, the HVP is an " + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "content": "-dimensional vector, too. In contrast, a Hessian is the non-directional gradient of the gradient, i.e., the gradient along all " + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "content": " dimensions, and as such in " + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{n \\times n}" + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "Sampling: To sample the smooth HVP, we simply need the directional central differences of an estimator of gradients, which we already have. So the estimator is the difference of two first-order estimators, evaluated at positions shifted from the current solution along the gradient direction [24, 35]:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 325, + 151, + 554, + 178 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 151, + 554, + 178 + ], + "spans": [ + { + "bbox": [ + 325, + 151, + 554, + 178 + ], + "type": "interline_equation", + "content": "\\mathrm {D} ^ {\\mathrm {H V P}} \\kappa (\\boldsymbol {\\tau}) = \\frac {\\mathrm {D} ^ {\\mathrm {G}} \\kappa (\\boldsymbol {\\tau} - \\varepsilon \\mathbf {v}) - \\mathrm {D} ^ {\\mathrm {G}} \\kappa (\\boldsymbol {\\tau} + \\varepsilon \\mathbf {v})}{2 \\varepsilon}. \\tag {23}", + "image_path": "591ab599450690df00e8dfc16979fcee9987941bcf7cdc5971e21452290acc3d.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 185, + 385, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 185, + 385, + 197 + ], + "spans": [ + { + "bbox": [ + 314, + 185, + 385, + 197 + ], + "type": "text", + "content": "3.5. Aggregate" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 199, + 554, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 199, + 554, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 554, + 258 + ], + "type": "text", + "content": "We summarize the key property of all samplers - the number of function evaluations required - in Tab. 1. Function evaluations require execution of the black-box rendering engine, the most costly part of the optimization, and hence should be minimized." + } + ] + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 315, + 315, + 446, + 399 + ], + "blocks": [ + { + "bbox": [ + 313, + 273, + 438, + 307 + ], + "lines": [ + { + "bbox": [ + 313, + 273, + 438, + 307 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 438, + 307 + ], + "type": "text", + "content": "Table 1. Time complexity of different estimator variants. AIS: aggregated importance sampling." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 315, + 446, + 399 + ], + "lines": [ + { + "bbox": [ + 315, + 315, + 446, + 399 + ], + "spans": [ + { + "bbox": [ + 315, + 315, + 446, + 399 + ], + "type": "table", + "html": "
no ISISAIS
Fin. Diff.nnn
FR22 [7]111
Our Grad.1n1
Our Hess.1n21
Our HVP1n1
", + "image_path": "a062daaca0cace5af8289af4c0d81515d7138d70e5eb98d8244349ea8705b7cf.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 455, + 259, + 555, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 455, + 259, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 455, + 259, + 555, + 415 + ], + "type": "text", + "content": "Classic finite differences take opposing samples in each dimension and hence, for an " + }, + { + "bbox": [ + 455, + 259, + 555, + 415 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 455, + 259, + 555, + 415 + ], + "type": "text", + "content": "-dimensional problem, require " + }, + { + "bbox": [ + 455, + 259, + 555, + 415 + ], + "type": "inline_equation", + "content": "2n" + }, + { + "bbox": [ + 455, + 259, + 555, + 415 + ], + "type": "text", + "content": " function evaluations for a single gradient sample. All other methods (second row in Tab. 1 onward) are based on MC, so we can get a gradient estimate using" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 415, + 554, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 486 + ], + "type": "text", + "content": "a fixed number of M function evaluations. This, however, comes at the cost of variance, which can be reduced with importance sampling, but that is done in all dimensions of the differential quantity independently, so it requires as many evaluations as these have elements (second column in Tab. 1). For a Hessian, this can be substantial, " + }, + { + "bbox": [ + 313, + 415, + 554, + 486 + ], + "type": "inline_equation", + "content": "n^2" + }, + { + "bbox": [ + 313, + 415, + 554, + 486 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 487, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 555, + 713 + ], + "type": "text", + "content": "As a compromise between low number of function evaluations and low variance, we propose \"aggregate\" importance sampling of the differential quantity (Fig. 4): instead of importance-sampling each dimension optimally but individually, we importance-sample w.r.t. the average of the convolution across all dimensions (combining several proposals by sampling from a mixture is similar to MIS using the balance heuristic [33]). This average convolution is a single function again, and with our combined sampling strategy can again be sampled with one function call per iteration (last column of Tab. 1). We implement this in two simple steps: first we randomly decide which element of the differential representation to choose (e.g., which matrix element out of " + }, + { + "bbox": [ + 313, + 487, + 555, + 713 + ], + "type": "inline_equation", + "content": "n^2" + }, + { + "bbox": [ + 313, + 487, + 555, + 713 + ], + "type": "text", + "content": " in the Hessian), and then we compute the convolution sample for all " + }, + { + "bbox": [ + 313, + 487, + 555, + 713 + ], + "type": "inline_equation", + "content": "n^2" + }, + { + "bbox": [ + 313, + 487, + 555, + 713 + ], + "type": "text", + "content": " but with the same function value that is only evaluated once. This increases variance, as the average kernel is not identical to the individual kernels but shares many properties, e.g., they are all zero in value at position 0. As a linear combination of unbiased es" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28202" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 294, + 190 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 294, + 190 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 294, + 190 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 294, + 190 + ], + "type": "image", + "image_path": "6345cca08a39cb8e61a8c87da21843ed9cdfbf2dad76a754eb918008183c771f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 205, + 295, + 316 + ], + "lines": [ + { + "bbox": [ + 55, + 205, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 295, + 316 + ], + "type": "text", + "content": "Figure 4. Aggregate and non-aggregate sampling for a 2D optimization space and, consequently, a " + }, + { + "bbox": [ + 55, + 205, + 295, + 316 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 55, + 205, + 295, + 316 + ], + "type": "text", + "content": " Hessian: Without aggregation, importance sampling is done for each element of the Hessian independently (four points at different positions in each kernel), leading to four rendering calls and four different mugs. Each value is then weighted by the kernel (blue and red colors denote positive and negative). In aggregate sampling, we importance-sample the average of all four stencils, resulting in a single sample location and, hence, four times the same mug that can be rendered in one call, weighted with four different kernels." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 319, + 296, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 381 + ], + "type": "text", + "content": "timators, this is still unbiased. We show in Sec. J that this pays off. Note that Tab. 1 is both time and space complexity, except for the aggregate importance sampling of the HVP, where the space complexity stays " + }, + { + "bbox": [ + 55, + 319, + 296, + 381 + ], + "type": "inline_equation", + "content": "O(n)" + }, + { + "bbox": [ + 55, + 319, + 296, + 381 + ], + "type": "text", + "content": " while time complexity is reduced to " + }, + { + "bbox": [ + 55, + 319, + 296, + 381 + ], + "type": "inline_equation", + "content": "O(1)" + }, + { + "bbox": [ + 55, + 319, + 296, + 381 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 388, + 127, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 127, + 400 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 127, + 400 + ], + "type": "text", + "content": "4. Evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 404, + 119, + 416 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 404, + 119, + 416 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 119, + 416 + ], + "type": "text", + "content": "4.1. Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 419, + 296, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 296, + 468 + ], + "type": "text", + "content": "Our evaluation compares the performance of gradients (input to gradient descent), Hessians and non-aggregate as well as aggregate Hessian-vector product (both input to higher-order optimizers)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 483, + 296, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 615 + ], + "type": "text", + "content": "Baselines An established inverse rendering solution is Mitsuba with gradient descent. FR22 is using gradient descent with Fischer and Ritschel [7] MC gradients. OursG, OursH, OursHVP and OursHVPA are our approach for gradient, Hessians and Hessian-vector products, and aggregates, respectively. OursG is used with gradient descent (GD), the others in combination with conjugate gradient (CG). We have also experimented with LBFGS as detailed in the supplemental, but do not report it here as it almost never converges and never at competitive speed. All methods use the Adam optimizer." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "type": "text", + "content": "Tasks We tackle artificial problems with known analytic Hessians and Hessian-vector products as well as real inverse rendering tasks. As a simple first test case, we optimize the smooth, quadratic potential " + }, + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "type": "inline_equation", + "content": "(\\mathrm{QUAD}^1)" + }, + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "type": "text", + "content": " function " + }, + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "type": "inline_equation", + "content": "ax_0^2 +bx_1^2 +cx_0x_1" + }, + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2" + }, + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "type": "text", + "content": ", with the fixed variables " + }, + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "type": "inline_equation", + "content": "a = 5,b = 5,c = 7.5" + }, + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Second, we optimize the classic plateau-demonstration task from [7]: the 2D position of some boxes is optimized" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "to match a reference. This is already a much harder task, as there is a plateau in the cost landscape when the squares do not overlap (almost always the case in the initial configuration), as the image-space error does not change. We study the case of one " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "(\\mathrm{Box2}^3)" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " and five " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "(\\mathrm{Box10}^5)" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " squares, with two and ten dimensions to optimize, respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 144, + 555, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 263 + ], + "type": "text", + "content": "For real inverse problems, we study the optimization of reflectance, light, and geometry and render using Mitsuba. In the MUG7 task, we optimize the vertical rotation of a coffee cup such that it matches a reference. In the SHADOW9 task, we optimize the position of a sphere that is unobserved in the rendered image, such that the shadow it casts matches a reference shadow. In the BUNNY11 task, the " + }, + { + "bbox": [ + 313, + 144, + 555, + 263 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 144, + 555, + 263 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 144, + 555, + 263 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 144, + 555, + 263 + ], + "type": "text", + "content": " position and the rotation around the " + }, + { + "bbox": [ + 313, + 144, + 555, + 263 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 144, + 555, + 263 + ], + "type": "text", + "content": " axis of the Stanford bunny are optimized. This task is set up specifically for Mitsuba to converge12, with no plateaus in the loss function." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 264, + 554, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 554, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 554, + 372 + ], + "type": "text", + "content": "To test the scaling with dimensionality, we also optimize the pixels of a " + }, + { + "bbox": [ + 313, + 264, + 554, + 372 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 313, + 264, + 554, + 372 + ], + "type": "text", + "content": " texture (TEXTURE" + }, + { + "bbox": [ + 313, + 264, + 554, + 372 + ], + "type": "inline_equation", + "content": "^{13}" + }, + { + "bbox": [ + 313, + 264, + 554, + 372 + ], + "type": "text", + "content": "), with 1024 parameters. In addition, we have used our approach on a mesh optimization task, where we fit the 3D position of the vertices of a tessellated sphere to images of a target object. For SUZANNE, we fit 2,562 vertices, for BANANA, we fit the 2,737 vertices and their RGB vertex colors, similar to the DTU setup [11]. For trajectories of the optimization and the outcome, see Fig. 5 top and bottom rows, respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 372, + 554, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 554, + 527 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 554, + 527 + ], + "type": "text", + "content": "Finally, we learn a " + }, + { + "bbox": [ + 313, + 372, + 554, + 527 + ], + "type": "inline_equation", + "content": "\\mathrm{CNN}^{15}" + }, + { + "bbox": [ + 313, + 372, + 554, + 527 + ], + "type": "text", + "content": " to predict scene parameters from images using inverse rendering without scene parameter labels. In this case, the optimization is learning in the stricter sense: we tune parameters of a deep architecture instead of scene parameters directly. The CNN takes as input an image of a mug and produces as output the orientation for a mug. As a loss, these parameters are inserted into a renderer, differentiated by different approaches and compared to a target. The CNN has 267745 parameters, learned by differentiating through an image loss and rendering of the scene given the estimated parameters. Another CNN " + }, + { + "bbox": [ + 313, + 372, + 554, + 527 + ], + "type": "inline_equation", + "content": "(\\mathrm{CNN}5^{19})" + }, + { + "bbox": [ + 313, + 372, + 554, + 527 + ], + "type": "text", + "content": " is trained to predict the rotation, position, and color of the mug with 268773 parameters." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 528, + 554, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 554, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 554, + 598 + ], + "type": "text", + "content": "Differentials of this function are computed as detailed in supplemental Sec. E, where we derive a grey-box approach that samples only the black-box part (the rendering) and combines its differentials with analytic differentials for the white-box part (the CNN). Importantly, this idea works for gradients, as well as for higher-order differentials." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 715 + ], + "type": "text", + "content": "Metrics Our main measure of success is convergence speed in wall-clock and parameter difference to the true parameters (which we know in all inverse rendering tasks as we created the scenes). A secondary measure of success is the image difference, as during optimization, we consider the parameters hidden. All metrics are evaluated across an ensemble of 20 runs averaged across 10 steps in time. In convergence plots, the ensemble median is shown at every point in time, averaged across 20 time steps." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28203" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 125, + 574, + 705 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 116 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 116 + ], + "type": "text", + "content": "Table 2. Quantitative results of different methods on different tasks (rows) and their convergence plots. We report convergence time in wall-clock units, in ratio to the overall best method, OurHVPA. In the numerical columns,.9 and .99 report the time taken to achieve 90 and " + }, + { + "bbox": [ + 55, + 71, + 555, + 116 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 55, + 71, + 555, + 116 + ], + "type": "text", + "content": " error reduction from the initial starting configuration, respectively, while the bar plots graphically show these findings. The line plots report image- and parameter-space convergence in the left and right column, respectively, on a log-log scale." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 125, + 574, + 705 + ], + "lines": [ + { + "bbox": [ + 57, + 125, + 574, + 705 + ], + "spans": [ + { + "bbox": [ + 57, + 125, + 574, + 705 + ], + "type": "table", + "html": "
TaskMethodImage errorParameter errorMitsuba FR22 CMA-ES OurG OurH OurHVP OurHVPA
.9.99.999.9.99.999
QUAD1Mitsu2—————ErrorLog error
FR221.802.382.982.203.173.90
CMA-ES0.150.130.140.130.140.23
OurG3.985.136.024.726.427.77
OurH1.171.171.151.181.161.17
OurHVP0.760.620.620.630.590.60
OurHVPA1.001.001.001.001.001.00
BOX23Mitsu——————ErrorLog error
FR223.242.492.572.782.892.70
CMA-ES5.513.453.158.135.514.13
OurG2.882.162.212.862.652.38
OurH0.930.951.080.760.841.84
OurHVP0.810.640.7040.740.720.72
OurHVPA1.001.001.001.001.001.00
BOX105Mitsu——————ErrorLog error
FR221.491.201.711.801.541.29
CMA-ES3.102.271.734.523.232.48
OurG2.181.771.532.532.261.88
OurH5.975.334.685.485.965.45
OurHVP1.511.441.6961.261.501.58
OurHVPA1.001.001.001.001.001.00
MUG7Mitsu——————ErrorLog error
FR225.925.695.463.286.536.29
CMA-ES2.951.971.823.322.382.20
OurG3.993.913.742.364.424.30
OurH1.571.091.531.001.081.20
OurHVP1.151.141.181.140.871.05
OurHVPA81.001.001.001.001.001.00
SHAD9Mitsu——————ErrorLog error
FR221.941.401.551.891.141.71
CMA-ES1.961.151.034.382.201.66
OurG1.371.181.132.381.471.34
OurH1.671.441.662.321.741.46
OurHVP1.140.910.961.371.131.07
OurHVPA1.001.001.001.001.001.00
BUNNY11Mitsu120.901.291.461.871.381.28ErrorLog error
FR220.750.751.120.780.630.71
CMA-ES1.160.950.851.270.910.85
OurG0.760.891.070.690.640.76
OurH1.291.231.291.180.801.07
OurHVP0.680.680.700.630.440.58
OurHVPA1.001.001.001.001.001.00
TEXTURE13Mitsu——————ErrorLog error
FR224.042.271.804.042.271.80
CMA-ES7.084.604.3511.167.477.44
OurG3.032.862.823.032.862.82
OurH440.1239.78—40.1239.78—
OurHVP4.014.004.004.014.004.00
OurHVPA1.001.001.001.001.001.00
CNN15Mitsu——————ErrorLog error
FR222.1116——1.451.322.71
CMA-ES———2.08——
OurG1.14——0.960.832.28
OurH0.770.77—0.740.770.74
OurHVP0.870.9317—0.880.910.87
OurHVPA1.001.00—181.001.001.00
CNN519Mitsu——————ErrorLog error
FR22——————
CMA-ES——————
OurG18.68——3.70——
OurH19.00——5.79——
OurHVP3.64201.46—1.491.58—
OurHVPA1.001.00—1.001.001.001.00
", + "image_path": "f12aa9828642080a6086462d1e0ef531c1774924acb659650dbc30a654c59bb8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 337, + 707, + 430, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 707, + 430, + 714 + ], + "spans": [ + { + "bbox": [ + 337, + 707, + 430, + 714 + ], + "type": "text", + "content": ".9 .99 .999 .9 99 .999" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 465, + 708, + 489, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 465, + 708, + 489, + 715 + ], + "spans": [ + { + "bbox": [ + 465, + 708, + 489, + 715 + ], + "type": "text", + "content": "Log time" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 522, + 708, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 522, + 708, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 522, + 708, + 546, + 715 + ], + "type": "text", + "content": "Log time" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28204" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 553, + 178 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 553, + 178 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 553, + 178 + ], + "type": "image", + "image_path": "f82b9bb7c56e627efb356317d13685da7bda43592231eca79b1a52ee372358b3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 182, + 552, + 194 + ], + "lines": [ + { + "bbox": [ + 56, + 182, + 552, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 182, + 552, + 194 + ], + "type": "text", + "content": "Figure 5. Inverse mesh optimization from renderings. Gradients use FR22, Hessians use OursHVPA, CMA-ES uses the pycma package." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 200, + 113, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 200, + 113, + 211 + ], + "spans": [ + { + "bbox": [ + 55, + 200, + 113, + 211 + ], + "type": "text", + "content": "4.2. Results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 216, + 295, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 295, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 295, + 276 + ], + "type": "text", + "content": "The main results of our evaluation are summarize in Tab. 2 where empty cells did not converge or the method is not applicable to that task. On average across tasks and methods, our premiere method oursHVPA speeds up the convergence by a factor of 2.71." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 277, + 296, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 296, + 325 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 296, + 325 + ], + "type": "text", + "content": "In general, our methods, in particular OursHVP and OursHVPA, lead the level of error reduction across all time budgets, as seen by comparing the convergence curves vertically where they consistently decrease fastest." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 326, + 296, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 326, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 54, + 326, + 296, + 458 + ], + "type": "text", + "content": "We benefit most in the artificial QUAD task, but MUG and SHADOW are both real rendering problems, where we are around seven times and " + }, + { + "bbox": [ + 54, + 326, + 296, + 458 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 54, + 326, + 296, + 458 + ], + "type": "text", + "content": " faster, respectively. We see that OurG typically cannot outperform FR22, which uses an approximation of the correct high-dimensional gradient kernel. Doing it \"right\" only pays off when going to higher order. We also note that OursHVP outperforms OursH due to less function evaluations (as the metric is wall-clock - the iteration count for both is close). The log-log plots confirm that the higher-order variants enabled by our approach are much faster, but also converge slightly less stable." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "text", + "content": "While tasks like QUAD and Box2 are not applicable to inverse renderers like Mitsuba², our methods treat any loss function as a black-box and retrieve derivatives via sampling. Comparing the performance of OurHVP and OurHVPA across a task with two and ten dimensions, we find the expected relation: OurHVP is faster in low dimensions⁴ than in higher dimensions for a similar task⁶." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 544, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 628 + ], + "type": "text", + "content": "We notice that the OurHVPA is not out-performing the OurHVP and OurH in the MUG task. This is because there is a single scene parameter for this task, thus, the higher order derivative is all in order of one. With a higher variance for OurHVPA, it could be a little slower. The effect of using the aggregate starts to show from SHAD, where OurHVPA reaches the 1:1000 convergence the fastest10." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": "For all the rendering tasks, only the BUNNY task converges for Mitsuba12. Since the other tasks have plateaus in the loss function, the analytical gradient from inverse rendering is effectively zero, so it would be hard to converge within a reasonable time. The log-log plot shows in addition that derivatives from the inverse rendering were slower than the sampled ones. On the contrary, our methods not" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 201, + 555, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 555, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 555, + 224 + ], + "type": "text", + "content": "only smooth out the plateau, but also improve the speed of the derivative computation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 225, + 555, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 225, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 225, + 555, + 369 + ], + "type": "text", + "content": "The drawback of constructing full Hessians is seen in the TEXTURE task, where OurH is more than 40 times slower than the HVP due to the matrix size" + }, + { + "bbox": [ + 313, + 225, + 555, + 369 + ], + "type": "inline_equation", + "content": "^{14}" + }, + { + "bbox": [ + 313, + 225, + 555, + 369 + ], + "type": "text", + "content": ". For the most demanding task, CNN and CNN5, and even our best method" + }, + { + "bbox": [ + 313, + 225, + 555, + 369 + ], + "type": "inline_equation", + "content": "^{18}" + }, + { + "bbox": [ + 313, + 225, + 555, + 369 + ], + "type": "text", + "content": " is not able to reduce the error below " + }, + { + "bbox": [ + 313, + 225, + 555, + 369 + ], + "type": "inline_equation", + "content": "1:1000" + }, + { + "bbox": [ + 313, + 225, + 555, + 369 + ], + "type": "text", + "content": ". At the " + }, + { + "bbox": [ + 313, + 225, + 555, + 369 + ], + "type": "inline_equation", + "content": "1:1000" + }, + { + "bbox": [ + 313, + 225, + 555, + 369 + ], + "type": "text", + "content": "-level, the more elaborate differentials" + }, + { + "bbox": [ + 313, + 225, + 555, + 369 + ], + "type": "inline_equation", + "content": "^{17}" + }, + { + "bbox": [ + 313, + 225, + 555, + 369 + ], + "type": "text", + "content": " is not faster than the basic one, but faster than first-order" + }, + { + "bbox": [ + 313, + 225, + 555, + 369 + ], + "type": "inline_equation", + "content": "^{16}" + }, + { + "bbox": [ + 313, + 225, + 555, + 369 + ], + "type": "text", + "content": ". Finally, the advantage of our ability to compute HVPs is seen when comparing CNN and CNN5: while a CNN producing a single parameter is doable with Hessians" + }, + { + "bbox": [ + 313, + 225, + 555, + 369 + ], + "type": "inline_equation", + "content": "^{17}" + }, + { + "bbox": [ + 313, + 225, + 555, + 369 + ], + "type": "text", + "content": ", but if we go to higher dimensions, the aggregate strategy pays off again and is markedly faster" + }, + { + "bbox": [ + 313, + 225, + 555, + 369 + ], + "type": "inline_equation", + "content": "^{20}" + }, + { + "bbox": [ + 313, + 225, + 555, + 369 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 370, + 554, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 554, + 442 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 554, + 442 + ], + "type": "text", + "content": "In conclusion, albeit convergence varies with problem characteristics and dimensionality, we find that our best method, OurHVPA converges most reliably, beating the other algorithms by an average factor of " + }, + { + "bbox": [ + 313, + 370, + 554, + 442 + ], + "type": "inline_equation", + "content": "2.71 \\times" + }, + { + "bbox": [ + 313, + 370, + 554, + 442 + ], + "type": "text", + "content": " in wall-clock units. We show examples of task initialization and outcome, and additional evaluation in the supplemental." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 454, + 471, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 471, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 471, + 467 + ], + "type": "text", + "content": "5. Limitations and Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 473, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 554, + 581 + ], + "type": "text", + "content": "Limitations. From the above experiments, it becomes apparent that our method performs well on lower-dimensional examples, but that its performance delta lessens with increasing problem dimensionality (see e.g. TEXTURE). We hypothesize that this is due to sparser sampling in higher dimensions and exacerbated MC noise and acknowledge that further research in this direction is needed. We include an analysis on how our method degrades under MC rendering noise in the supplemental." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "Conclusion. The availability of second-order gradient information gives rise to an exciting avenue of future research in inverse rendering. While these algorithms are not new, their use in vision & graphics remains limited due to the convenience and widespread adoption of first-order gradient descent and the increased per-iteration cost that is usually associated with higher-order methods. In this work, our efficient estimators have shown the latter to be negligible in real-world optimization scenarios. We hope that this will inspire future research into unbiased, low-variance estimators of higher-order optimization methods." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28205" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 704 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "text", + "content": "[1] Albert S Berahas, Richard H Byrd, and Jorge Nocedal. Derivative-free optimization of noisy functions via quasiNewton methods. SIAM Journal on Optimization, 29(2):965-993, 2019. 11" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 135, + 296, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 135, + 296, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 135, + 296, + 167 + ], + "type": "text", + "content": "[2] John P Boyd. Solving transcendental equations: the Chebyshev polynomial proxy and other numerical rootfinders, perturbation series, and oracles. SIAM, 2014. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 168, + 296, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 168, + 296, + 200 + ], + "spans": [ + { + "bbox": [ + 62, + 168, + 296, + 200 + ], + "type": "text", + "content": "[3] Tony F Chan, Gene H Golub, and Pep Mulet. A nonlinear primal-dual method for total variation-based image restoration. SIAM Scient. Comput., 20(6):1964-1977, 1999. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 201, + 294, + 222 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 201, + 294, + 222 + ], + "spans": [ + { + "bbox": [ + 62, + 201, + 294, + 222 + ], + "type": "text", + "content": "[4] Swarat Chaudhuri and Armando Solar-Lezama. Smooth interpretation. ACM Sigplan Notices, 45(6):279-291, 2010. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 224, + 294, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 224, + 294, + 255 + ], + "spans": [ + { + "bbox": [ + 62, + 224, + 294, + 255 + ], + "type": "text", + "content": "[5] Thomas Deliot, Eric Heitz, and Laurent Belcour. Transforming a non-differentiable rasterizer into a differentiable one with stochastic gradient estimation. In ACM i3D, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 256, + 294, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 256, + 294, + 298 + ], + "spans": [ + { + "bbox": [ + 62, + 256, + 294, + 298 + ], + "type": "text", + "content": "[6] Deshana Desai, Etai Shuchatowitz, Zhongshi Jiang, Teseo Schneider, and Daniele Panozzo. Acorns: An easy-to-use code generator for gradients and hessian. SoftwareX, 17: 100901, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 300, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 300, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 62, + 300, + 294, + 331 + ], + "type": "text", + "content": "[7] Michael Fischer and Tobias Ritschel. Plateau-reduced differentiable path tracing. In CVPR, pages 4285-4294, 2023. 1, 3, 4, 5, 6, 10, 14" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 333, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 333, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 62, + 333, + 294, + 365 + ], + "type": "text", + "content": "[8] Michael Fischer and Tobias Ritschel. Zerograds: Learning local surrogates for non-differentiable graphics. ACM Trans. Graph. (Proc. SIGGRAPH), 43(4):1-15, 2024. 1, 3, 13" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 366, + 294, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 366, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 62, + 366, + 294, + 397 + ], + "type": "text", + "content": "[9] Roger Fletcher and Michael JD Powell. A rapidly convergent descent method for minimization. The computer journal, 6 (2):163-168, 1963. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 399, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 399, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 57, + 399, + 294, + 441 + ], + "type": "text", + "content": "[10] Nikolaus Hansen and Andreas Ostermeier. Adapting arbitrary normal mutation distributions in evolution strategies: The covariance matrix adaptation. In Proc. IEEE Evolutionary Computation, pages 312-317, 1996. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 443, + 294, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 443, + 294, + 474 + ], + "spans": [ + { + "bbox": [ + 57, + 443, + 294, + 474 + ], + "type": "text", + "content": "[11] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engil Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In CVPR, pages 406-413, 2014. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 475, + 294, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 475, + 294, + 495 + ], + "spans": [ + { + "bbox": [ + 57, + 475, + 294, + 495 + ], + "type": "text", + "content": "[12] James T Kajiya. The rendering equation. In Proc. SIGGRAPH, 1986. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 497, + 294, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 497, + 294, + 528 + ], + "spans": [ + { + "bbox": [ + 57, + 497, + 294, + 528 + ], + "type": "text", + "content": "[13] Martin Kilian, Niloy J Mitra, and Helmut Pottmann. Geometric modeling in shape space. In SIGGRAPH, pages 64- es, 2007. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 529, + 294, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 529, + 294, + 551 + ], + "spans": [ + { + "bbox": [ + 57, + 529, + 294, + 551 + ], + "type": "text", + "content": "[14] Dilip Krishnan and Rob Fergus. Fast image deconvolution using hyper-Laplacian priors. Proc. NeurIPS, 22, 2009. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 552, + 294, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 552, + 294, + 583 + ], + "spans": [ + { + "bbox": [ + 57, + 552, + 294, + 583 + ], + "type": "text", + "content": "[15] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In Proc. ICCV, 2019. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 585, + 294, + 605 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 585, + 294, + 605 + ], + "spans": [ + { + "bbox": [ + 57, + 585, + 294, + 605 + ], + "type": "text", + "content": "[16] Matthew M Loper and Michael J Black. OpenDR: An approximate differentiable renderer. In Proc. ECCV, 2014. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 607, + 294, + 637 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 607, + 294, + 637 + ], + "spans": [ + { + "bbox": [ + 57, + 607, + 294, + 637 + ], + "type": "text", + "content": "[17] James Martens. New insights and perspectives on the natural gradient method. J Machine Learning Res, 21(146):1-76, 2020. 11" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 57, + 639, + 294, + 681 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 639, + 294, + 681 + ], + "spans": [ + { + "bbox": [ + 57, + 639, + 294, + 681 + ], + "type": "text", + "content": "[18] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Comm ACM, 65(1), 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 57, + 683, + 294, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 683, + 294, + 704 + ], + "spans": [ + { + "bbox": [ + 57, + 683, + 294, + 704 + ], + "type": "text", + "content": "[19] Baptiste Nicolet, Alec Jacobson, and Wenzel Jakob. Large steps in inverse rendering of geometry. ACM Trans. Graph." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 708 + ], + "type": "list", + "angle": 0, + "index": 42, + "blocks": [ + { + "bbox": [ + 335, + 73, + 502, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 502, + 83 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 502, + 83 + ], + "type": "text", + "content": "(Proc. SIGGRAPH Asis), 40(6):1-13, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 85, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 85, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 85, + 553, + 116 + ], + "type": "text", + "content": "[20] Merlin Nimier-David, Delio Vicini, Tizian Zeltner, and Wenzel Jakob. Mitsuba 2: A retargetable forward and inverse renderer. ACM Trans. Graph., 38(6), 2019. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "type": "text", + "content": "[21] Merlin Nimier-David, Sébastien Speierer, Benoit Ruiz, and Wenzel Jakob. Radiative backpropagation: an adjoint method for lightning-fast differentiable rendering. ACM Trans Graph, 39(4), 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 162, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 182 + ], + "type": "text", + "content": "[22] Jorge Nocedal and Stephen J Wright. Numerical optimization. Springer, 1999. 1, 2, 11" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 184, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 205 + ], + "type": "text", + "content": "[23] Art Owen and Yi Zhou. Safe and effective importance sampling. J American Statistical Assoc., 95(449), 2000. 1, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 206, + 553, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 226 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 226 + ], + "type": "text", + "content": "[24] Barak A Pearlmutter. Fast exact multiplication by the hessian. Neural computation, 6(1):147-160, 1994. 3, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 228, + 553, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 248 + ], + "type": "text", + "content": "[25] Michael JD Powell. Algorithms for nonlinear constraints that use lagrangian functions. Math. Prog., 14:224-248, 1978. 11" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 250, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 281 + ], + "type": "text", + "content": "[26] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv:2204.06125, 1(2):3, 2022. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 282, + 553, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 282, + 553, + 302 + ], + "spans": [ + { + "bbox": [ + 316, + 282, + 553, + 302 + ], + "type": "text", + "content": "[27] Sebastian Ruder. An overview of gradient descent optimization algorithms. arXiv:1609.04747, 2016. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 304, + 553, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 553, + 335 + ], + "type": "text", + "content": "[28] Pedro V Sander, John Snyder, Steven J Gortler, and Hugues Hoppe. Texture mapping progressive meshes. In Proc. SIGGRAPH, pages 409-416, 2001. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 337, + 553, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 337, + 553, + 368 + ], + "spans": [ + { + "bbox": [ + 316, + 337, + 553, + 368 + ], + "type": "text", + "content": "[29] David F Shanno. Conditioning of quasi-newton methods for function minimization. Mathematics of computation, 24 (111):647-656, 1970. 11" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 369, + 553, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 369, + 553, + 401 + ], + "spans": [ + { + "bbox": [ + 316, + 369, + 553, + 401 + ], + "type": "text", + "content": "[30] Jack Sherman. Adjustment of an inverse matrix corresponding to changes in the elements of a given column or row of the original matrix. Annu. Math. Statist., 20:621, 1949. 11" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 403, + 553, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 403, + 553, + 423 + ], + "spans": [ + { + "bbox": [ + 316, + 403, + 553, + 423 + ], + "type": "text", + "content": "[31] Jonathan Richard Shewchuk. An introduction to the conjugate gradient method without the agonizing pain. 1994. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 425, + 553, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 425, + 553, + 456 + ], + "spans": [ + { + "bbox": [ + 316, + 425, + 553, + 456 + ], + "type": "text", + "content": "[32] James C Spall. Multivariate stochastic approximation using a simultaneous perturbation gradient approximation. IEEE Trans Automatic Control, 37(3):332-341, 1992. 3" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 457, + 553, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 457, + 553, + 489 + ], + "spans": [ + { + "bbox": [ + 316, + 457, + 553, + 489 + ], + "type": "text", + "content": "[33] Eric Veach and Leonidas J Guibas. Optimally combining sampling techniques for Monte Carlo rendering. In Proc. SIGGRAPH, pages 419-428, 1995. 4, 5" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 316, + 491, + 553, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 491, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 316, + 491, + 553, + 532 + ], + "type": "text", + "content": "[34] Delio Vicini, Sébastien Speierer, and Wenzel Jakob. Path replay backpropagation: differentiating light paths using constant memory and linear time. ACM Trans Graph, 40(4), 2021. 3" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 316, + 534, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 553, + 555 + ], + "type": "text", + "content": "[35] Werbos. Backpropagation: Past and future. In IEEE 1988 Int Conf Neural Networks, pages 343-353. IEEE, 1988. 3, 5" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 316, + 556, + 553, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 556, + 553, + 587 + ], + "spans": [ + { + "bbox": [ + 316, + 556, + 553, + 587 + ], + "type": "text", + "content": "[36] Manuel Werlberger, Werner Trobin, Thomas Pock, Andreas Wedel, Daniel Cremers, and Horst Bischof. Anisotropic huber-11 optical flow. In BMVC, volume 1, 2009. 3" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 316, + 589, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 589, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 589, + 553, + 632 + ], + "type": "text", + "content": "[37] Zhewei Yao, Amir Gholami, Sheng Shen, Mustafa Mustafa, Kurt Keutzer, and Michael Mahoney. Adahessian: An adaptive second order optimizer for machine learning. In Proc. AAAI, volume 35, pages 10665-10673, 2021. 3" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 316, + 633, + 553, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 633, + 553, + 675 + ], + "spans": [ + { + "bbox": [ + 316, + 633, + 553, + 675 + ], + "type": "text", + "content": "[38] Yizhou Yu, Paul Debevec, Jitendra Malik, and Tim Hawkins. Inverse global illumination: Recovering reflectance models of real scenes from photographs. In Proc. SIGGRAPH, pages 215-224, 1999. 1" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 316, + 677, + 553, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 677, + 553, + 708 + ], + "spans": [ + { + "bbox": [ + 316, + 677, + 553, + 708 + ], + "type": "text", + "content": "[39] Christopher Zach, Thomas Pock, and Horst Bischof. A duality based approach for realtime TV-L1 optical flow. In DAGM Pattern Recognition, pages 214-223, 2007. 3" + } + ] + } + ], + "index": 41 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "28206" + } + ] + } + ], + "index": 43 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Stochastic Interpolants for Revealing Stylistic Flows across the History of Art/16c23549-7e7f-4e3a-8944-314b6bb00228_content_list.json b/2025/Stochastic Interpolants for Revealing Stylistic Flows across the History of Art/16c23549-7e7f-4e3a-8944-314b6bb00228_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fbf1b76886af69cad2bd04481940c15742bde7ad --- /dev/null +++ b/2025/Stochastic Interpolants for Revealing Stylistic Flows across the History of Art/16c23549-7e7f-4e3a-8944-314b6bb00228_content_list.json @@ -0,0 +1,1582 @@ +[ + { + "type": "text", + "text": "Stochastic Interpolants for Revealing Stylistic Flows across the History of Art", + "text_level": 1, + "bbox": [ + 109, + 130, + 888, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pingchuan Ma* Ming Gui* Johannes Schusterbauer \nXiaopei Yang Olga Grebenkova Vincent Tao Hu Björn Ommer", + "bbox": [ + 207, + 180, + 789, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "CompVis @ LMU Munich Munich Center for Machine Learning (MCML)", + "bbox": [ + 174, + 233, + 820, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generative models have made rapid progress in content creation, particularly in synthesizing artworks and capturing stylistic variation. However, most methods operate at the level of individual images, limiting their ability to reveal broader stylistic trends and temporal transitions. We address this by introducing a framework that models stylistic evolution as an optimal transport problem in a learned style space, using stochastic interpolants and dual diffusion implicit bridges to align artistic distributions across time without requiring paired data. A central contribution is a diverse dataset of over 650,000 artworks spanning 500 years, curated with metadata across multiple genres. Together, our method and dataset enable tracing long-range stylistic transitions and plausible futures of individual artworks, supporting fine-grained temporal analysis. This offers a new tool for modeling historical patterns in visual culture and opens up promising directions in visual understanding. Code and dataset: https://github.com/CompVis/Art-fm", + "bbox": [ + 89, + 318, + 485, + 590 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 621, + 220, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in computer vision have sparked growing interest in applications in interdisciplinary research domains. Such applications not only try to offer novel perspectives, but also address challenges that were impossible before, such as those seen in medical imaging [18, 53], novel detection in material science [50], and simulation of climate [56]. Despite these successes, there remains a noticeable gap in their use for cultural studies [4], particularly in art history.", + "bbox": [ + 89, + 647, + 483, + 767 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the rise of generative models such as Stable Diffusion [66], DALL-E [64], and others [51, 58, 69], there has been increasing use of these tools for artwork synthesis and content creation. However, their primary focus remains on image generation and editing, rather than analyzing the broader narrative of art history [10, 19, 99]. As a result, current models are not designed to capture the complex evo", + "bbox": [ + 89, + 768, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lution of artistic movements and styles over time.", + "bbox": [ + 511, + 287, + 839, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we ask: How would a piece of art evolve? What might it look like if created centuries earlier or later? Naturally, no ground-truth pairs exist for such a task, and the artistic style of a creator is inherently dynamic and historically constrained. To address this challenge, we aim to analyze the flow of artistic styles over time. Given a single artwork, we try to determine its closest \"projection\" onto the manifold of artistic expressions corresponding to a given historical period. While optimal transport provides a natural formulation for such a task, considering only a single isolated artwork is insufficient. Instead, we reformulate the problem as a distribution matching task, where we study the evolution of the underlying distribution of art over time.", + "bbox": [ + 511, + 303, + 906, + 500 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in diffusion bridges such as DDIB [83], DSB [13, 85], and DDBMs [100] have shown that domain translation can be performed implicitly, without the need for paired supervision. We build on this idea in the context of art history by incorporating stochastic interpolants [1], enabling us to model the temporal evolution of artistic styles without requiring explicitly aligned training data. Unlike traditional style transfer methods that transform images based on predefined and discrete styles, our approach projects styles onto their corresponding distributions across time. Operating in a compact latent space allows us to abstract away pixel-level distraction and focus on meaningful stylistic shifts.", + "bbox": [ + 511, + 501, + 908, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To support this temporal modeling, we require a dataset that spans centuries and reflects a wide variety of artistic styles. While datasets like WikiArt [71] offer breadth in genre and metadata, they lack the scale and historical depth needed to model long-term stylistic progression. Others, such as BAM [93] and BAM-FG [70], are either restricted in access or skewed toward modern art. To address this, we curate a new unified dataset of over 650,000 artworks annotated with creation year, location, artist, medium, and other metadata (Fig. 1). This dataset enables the study of artistic evolution at scale and over time. In summary, our key contributions are as follows:", + "bbox": [ + 511, + 684, + 908, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "- We propose the first generative framework that explicitly models the temporal progression of artistic styles across", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution", + "bbox": [ + 109, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5867", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/23df6f6359b304d32dce1aa8383cb1aebd9ab9e8467ea9ad03ba6e3d2502e6f6.jpg", + "image_caption": [ + "Figure 1. Overview of the proposed dataset. We have curated and unified our large-scale dataset, including 650,000 artworks from different creation periods (ranging from 1500 to 2021) and styles with other meta-information, e.g., year, title, artist, media, etc." + ], + "image_footnote": [], + "bbox": [ + 122, + 90, + 880, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "historical periods, introduced in Sec. 3.2 and Sec. 3.3.", + "bbox": [ + 102, + 517, + 460, + 531 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To support this, we construct a large-scale, unified dataset of 650,000 artworks annotated with creation year and other metadata, covering five centuries of diverse artistic styles (Sec. 3.4). An overview is shown in Fig. 1.", + "- Our method models temporal transitions without requiring paired data. We outperform competitive baselines in capturing long-range stylistic shifts (Sec. 4.2)." + ], + "bbox": [ + 89, + 532, + 483, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 655, + 240, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Stylistic representation", + "text_level": 1, + "bbox": [ + 89, + 681, + 303, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The pioneering work by Gatys et al. [21] marked the beginning of the style transfer era, which defines style as Gram Matrices of pre-trained network features. Various methods have been proposed to improve different aspects of style representation, ranging from efficiency-focused approaches like [32, 42, 92] to quality-centric ones such as [36, 89, 98, 101]. Moreover, efforts have been made to enhance the generality and diversity of style transfer methods [30, 90, 91, 95]. Despite these achievements, most existing approaches [6, 41, 61, 86, 98] focus on image editing or the disentanglement of style and content through explicit style/content representations or user-defined prompts. Data-driven methods as [15] defined the representativity of art", + "bbox": [ + 89, + 704, + 485, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "works as a deep feature fusion incorporating not only images, but also painting styles, creation time, and authorship information. On the other hand, ContextNet [19] aims to represent and explore artistic attributes in different modalities and their relationships to improve the retrieval of artworks in automatic art analysis. ArtGCN [99] uses a graph CNN to construct stylistic representation based on artistic comments. More recent works, such as GDA [88] and CSD [79], train style descriptors by fine-tuning on top of the existing feature space of CLIP [62] through contrastive learning with synthetic or curated data.", + "bbox": [ + 511, + 517, + 906, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These studies typically concentrate on rendering or differentiating individual images with specific styles and often overlook treating styles as distributions. Chen et al. [5] tried to match portrait images through time with a family of models for each decade and re-synthesize them. However, it is limited to simple face data. In our work, we focus on analyzing artistic style shifts in artworks depicting more complicated scenes from different eras and using a curated dataset with newly proposed flow-based models, which have not been explored before in similar analyses.", + "bbox": [ + 511, + 684, + 908, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Art datasets", + "text_level": 1, + "bbox": [ + 511, + 847, + 643, + 862 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Various datasets focus on different domains [2, 14, 43, 82, 97]. However, in the domain of art, particularly paintings", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5868", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a9e5bf9510c512ae517c3e63a7030a432de87e69ec8217e43ae397a2f5b26e06.jpg", + "image_caption": [ + "Figure 2. Implicit bridge with 2D example. Two unconditional models are trained between Gaussian noise and data distributions $\\mathcal{A}$ (squares) and $\\mathcal{B}$ (circles), without access to mode labels (colors). Despite this, structure is preserved in noise space. By mapping samples from $\\mathcal{A}$ to noise via $v^{(a)}$ and then back to $\\mathcal{B}$ via $v^{(b)}$ , we show that correspondences can emerge without paired supervision." + ], + "image_footnote": [], + "bbox": [ + 218, + 103, + 781, + 193 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "concerning artistic style, there is a notable absence of large-scale datasets tailored for comprehensive analysis. Existing art datasets exhibit several limitations. WikiArt [71] provides a diverse collection of styles and includes year information in some of the filenames, but is limited in the number of samples to facilitate a comprehensive analysis of artistic evolution. BAM [93] and BAM-FG [70] are restricted in accessibility as they are not publicly available and primarily consist of contemporary artworks, thereby lacking a historical perspective essential for studying stylistic progression. LAION-Styles [79] has curated a large style attribution dataset from LAION [74]. However, it lacks chronological information and does not differentiate between styles and artists, which is crucial for analyzing the evolution of styles.", + "bbox": [ + 88, + 268, + 485, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To the best of our knowledge, no previous work has curated a large-scale dataset of artwork with detailed metadata, including style, artist, year of creation, and medium. Our curated dataset (see Sec. 3.4) unifies these aspects, enabling scalable studies beyond existing approaches.", + "bbox": [ + 89, + 484, + 486, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Flow Matching and Stochastic Interpolants", + "text_level": 1, + "bbox": [ + 89, + 585, + 459, + 603 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Flow Matching [44, 46, 54] has established Continuous Normalizing Flow (CNF) as a notable competitor to diffusion models [26, 78]. It enables simulation-free training of CNFs using an optimal transport conditional probability path with Ordinary Differential Equations (ODE), providing a more efficient alternative to traditional diffusion training and sampling methods [34, 81]. The versatility of Flow Matching has been showcased across various domains, including image [27, 29, 37, 44, 75], depth map [22], video [12], human motion [28], point cloud [94], and manifold data [7]. A more recent family of frameworks - Stochastic Interpolants [1, 47] has been proposed for unifying Flows and Diffusions. They suggested that Stochastic Interpolants can be generalized to align with various generative models, including Normalizing Flow [8], Diffusion Model [66, 80], Flow Matching [22, 29, 75], and Schrödinger Bridge [45, 85]. In this work, we develop our method within the stochastic interpolant framework and use the learned optimal transport paths (flows) to study the progression of styles across eras.", + "bbox": [ + 91, + 613, + 486, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 266, + 604, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We aim to enable the matching of artworks across different time periods, despite the absence of ground-truth correspondences. Analyzing the temporal evolution of individual paintings is inherently infeasible, as artworks are static and finalized entities. Future stylistic transformations are highly unpredictable, even when considering subsequent works by the same artist, given that many artists undergo significant stylistic changes throughout their careers (e.g., Van Gogh or Pablo Picasso). Due to this absence of explicit ground-truth data pairs, we propose an alternative formulation of the problem. Specifically, we approach it as a distribution matching task, where the evolution of artistic styles over time is understood in terms of underlying probability distributions rather than isolated instances.", + "bbox": [ + 511, + 292, + 908, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By modelling on the global distribution of artistic styles across time, we can establish meaningful correspondences spanning different centuries [77]. A principled framework for matching distributions is Optimal Transport (OT), which minimizes the transportation costs between distributions [83, 100]. To achieve this, we leverage implicit bridges and flow matching, which, upon training, inherently define an Ordinary Differential Equation (ODE) that aligns with the principles of optimal transport. This not only captures high-level stylistic transitions over time but also facilitates fine-grained alignment of individual artworks based on their placement within these temporally evolving distributions.", + "bbox": [ + 511, + 505, + 908, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Theoretical background", + "text_level": 1, + "bbox": [ + 511, + 698, + 736, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Stochastic Interpolants In recent years, generative probabilistic models that transform noise into data have gained significant attention. These models are primarily represented by diffusion-based models [17, 66, 80, 81] and flow-based models [26, 44]. This transformation can be defined [1] by a time-dependent forward process with $t \\in [0,1]$ as", + "bbox": [ + 511, + 720, + 908, + 811 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} = \\alpha_ {t} x _ {0} + \\sigma_ {t} x _ {1}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 830, + 906, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $x_0$ corresponds to data sampled from the data distribution $p_0(x)$ for which we have empirical observations available and $x_1 \\sim \\mathcal{N}(0, \\mathbf{I})$ is a simple prior distribution.", + "bbox": [ + 511, + 854, + 911, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5869", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Transforming artworks from source age distribution to target age distribution" + ], + "code_body": "Input: $x_0^{(a)}\\sim p_A(x)$ - data samples from age A $y_{a}$ - original chronological conditioning $y_{b}$ -chronological conditioning for desired age $y_{b} = y_{a} + \\Delta y$ $v_{\\theta}^{(a)}$ - model corresponds to age A, instantiated by $v_{\\theta}(x_t,t,y_a)$ $v_{\\theta}^{(b)}$ - target model corresponds to age B, instantiated by $v_{\\theta}(x_t,t,y_b)$ $t_\\mathrm{th}\\in (0,1]$ - the integration threshold \nOutput: $x_0^{(b)}$ , the result in the target age data distribution B. 1: for each element $x_0^{(a)}\\sim p_A(x)$ do 2: $\\begin{array}{rlr}{x_{\\mathrm{th}} = \\mathrm{ODESolve}(x_0^{(a)};v_{\\theta}^{(a)},0,t_{\\mathrm{th}})} & \\triangleright & \\mathrm{obtain~the~structured~noise~from}\\\\ {\\mathrm{source~age~data}} & {} & \\\\ {x_0^{(b)} = \\mathrm{ODESolve}(x_{\\mathrm{th}};v_{\\theta}^{(b)},t_{\\mathrm{th}},0)} & \\triangleright & \\mathrm{obtain~target~age~data~from~the}\\\\ {\\mathrm{structured~noise}} & {} & \\\\ {\\mathrm{return} x_0^{(b)}} & {} & \\end{array}$ 3:", + "bbox": [ + 89, + 122, + 480, + 281 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The forward process is characterized by the hyperparameters $\\alpha_{t}$ and $\\sigma_{t}$ , with $\\alpha_{t}$ decreasing and $\\sigma_{t}$ increasing as time $t\\in [0,1]$ progresses, interpolating between data and noise. Furthermore, the boundary conditions are normally constrained as $\\alpha_0 = \\sigma_1 = 1$ and $\\alpha_{1} = \\sigma_{0} = 0$ , so that for $t = 0$ we have pure data and $t = 1$ pure noise. In particular, flow-based models, represented by SiT [47, 75], have attracted significant interest due to their faster training and inference compared to diffusion-based models. With the above-defined forward process, the velocity is defined as", + "bbox": [ + 89, + 308, + 486, + 460 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nv (x, t) = \\mathbb {E} [ \\dot {x} _ {t} | x _ {t} = x ], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 470, + 483, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which generates the marginal probability distribution $p_t(x)$ of $x_{t}$ at time $t$ [1, 47]. This aligns with the distributions of the probability flow ODE [81] so that during inference we can generate samples by solving the Ordinary Differential Equation (ODE) backward in time starting from pure noise $\\mathcal{N}(0,\\mathbf{I})$ to arrive at a sample $x\\sim p_0(x)$ . Ma et al. [47] demonstrates that we can train a neural network $v_{\\theta}(x_{t},t)$ to parametrize the velocity using the training objective:", + "bbox": [ + 89, + 500, + 483, + 621 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {v}} (\\theta) = \\int_ {0} ^ {T} \\mathbb {E} [ \\| v _ {\\theta} (x _ {t}, t) - \\dot {\\alpha} _ {t} x _ {0} - \\dot {\\sigma} _ {t} x _ {1} \\| ^ {2} ] \\mathrm {d} t, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 643, + 483, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with $\\dot{\\alpha}_{t}$ and $\\dot{\\sigma}_t$ representing the time derivative of $\\alpha_{t}$ and $\\sigma_t$ , respectively. For simplicity and due to its relatively straight trajectories, we adopt the Linear Interpolant from [46, 47] with $\\alpha_{t} = 1 - t$ and $\\sigma_t = t$ .", + "bbox": [ + 89, + 688, + 485, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Bridging artworks over time", + "text_level": 1, + "bbox": [ + 89, + 757, + 346, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We aim to establish meaningful correspondences between artworks originating from different historical periods. By leveraging and adapting the recent advances in generative modelling, we develop a framework that learns to transport dynamics across time-conditioned data distributions without requiring paired supervision. We begin by introducing the foundations of implicit diffusion bridges, before detailing how they can be applied to match artworks across time.", + "bbox": [ + 89, + 780, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $x^{(a)}$ represent a data sample from data distribution $\\mathcal{A}$ . The goal is to identify the corresponding sample $x^{(b)}$ from a different data distribution $\\mathcal{B}$ . This problem is inherently ill-posed due to the absence of ground-truth data pairs. To address this, we seek to establish optimal transport between the underlying manifolds and construct Schrödinger bridges between the respective distributions [9, 40].", + "bbox": [ + 511, + 90, + 906, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dual Diffusion Implicit Bridges (DDIB) [83] demonstrate that such bridges can be effectively constructed using a pair of pre-trained image diffusion models. Notably, the Ordinary Differential Equations (ODEs) derived from trained diffusion models inherently exhibit optimal transport between the Gaussian noise prior and the corresponding data distributions [85]. Let ODESolve denote the mapping from an initial noise $x_{1}$ to the data $x_0^{(a)}$ with $v_{\\theta}^{(a)}$ defined in Eq. (3):", + "bbox": [ + 511, + 200, + 908, + 325 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {O D E S o l v e} \\left(x _ {t}; v _ {\\theta} ^ {(a)}, 1, 0\\right) = x _ {1} + \\int_ {1} ^ {0} v _ {\\theta} ^ {(a)} \\left(x _ {t}, t\\right) d t. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 344, + 906, + 380 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Analogously, we can train ODESolve $(x_{t};v_{\\theta}^{(b)},1,0)$ for mapping data between noise and $\\mathcal{B}$ . Given these frameworks, the learned mappings between noise and data distributions, along with their corresponding reverse processes ODESolve $(x_{t};v_{\\theta},0,1)$ , inherently constitute optimal transport pairs. As illustrated in the example in Fig. 2, we demonstrate this ability to transfer data between distributions $\\mathcal{A}$ and $\\mathcal{B}$ despite having no prior information about their inherent relationship. First, we use the trained ODE to map data distribution $\\mathcal{A}$ to the standard Gaussian noise. A clear structure related to $\\mathcal{A}$ emerges, highlighting the optimal noise-to-data pairs achieved through the ODE model's transport. Subsequently, another trained ODE transports this structured noise [67] to the distinctly different distribution $\\mathcal{B}$ . In this setup, the mapping between distributions $\\mathcal{A}$ and $\\mathcal{B}$ remains largely preserved, even when using two independently trained ODEs that share no information. This indicates that the noise space might not be pure noise, but rather contains a certain structure as suggested in [67].", + "bbox": [ + 511, + 404, + 908, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The independent $v_{\\theta}$ can be achieved by training the same model with different conditions, namely", + "bbox": [ + 511, + 696, + 906, + 728 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nv _ {\\theta} ^ {(a)} = v _ {\\theta} (\\cdot , t, y _ {a}), v _ {\\theta} ^ {(b)} = v _ {\\theta} (\\cdot , t, y _ {b}). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 747, + 906, + 767 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Under our problem setup, the data $x_0^{(a)}$ resembles a piece of artwork, and $y_{a}$ represents the corresponding chronological information (creation year of the artwork). This parameterization allows a single model $v_{\\theta}$ to learn mappings between arbitrary years $y_{a} \\rightarrow y_{b}$ through conditioning, effectively replacing the need to train separate models for each temporal pair, as illustrated in Fig. 3.", + "bbox": [ + 511, + 792, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5870", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/09cafccceb785dd5d8517a98e3fc4cc6c389f503ed50f671e1ca8d74e35e0fca.jpg", + "image_caption": [ + "(a) Training" + ], + "image_footnote": [], + "bbox": [ + 143, + 98, + 433, + 166 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/558d5811cd82feb59d0b1a9e314e5510c152e825cd36dceb9f6f3e4ea5df7ba4.jpg", + "image_caption": [ + "(b) Inference" + ], + "image_footnote": [], + "bbox": [ + 143, + 191, + 433, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training objective. Similar to Eq. (3), we define our final objective to optimize $v_{\\theta}$ with the year condition $y$ as:", + "bbox": [ + 89, + 401, + 483, + 433 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\theta) = \\int_ {0} ^ {T} \\mathbb {E} \\left[ \\| v _ {\\theta} \\left(x _ {t}, t, y\\right) - \\dot {\\alpha} _ {t} x _ {0} - \\dot {\\sigma} _ {t} x _ {1} \\| ^ {2} \\right] \\mathrm {d} t, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 441, + 483, + 478 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\theta$ is the backbone parameterized by a 1D transformer. We detail the architecture of the model in the Appendix A. We operate in a compact 1D-style space, as pixel space retains excessively low-level details that contain redundancy.", + "bbox": [ + 89, + 487, + 485, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Inference", + "text_level": 1, + "bbox": [ + 89, + 556, + 200, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With a trained model that can generate the style latents sufficiently well, we can start the matching process. The model defines a vector field $v_{\\theta}(\\cdot ,t,y)$ at any given $t$ and $y$ . As explained in Algorithm 1, we first transform the style representation from $x_0^{(a)}$ to the noisy representation $x_{t_{\\mathrm{th}}}$ using the learned model $v_{\\theta}$ . The parameter $t_\\mathrm{th} > 0$ denotes the integration threshold; in practice, we typically set $t_\\mathrm{th} = 1$ , corresponding to a full mapping from data space to the structured noise space. Next, we push forward the $x_{t_\\mathrm{th}}$ back to $x_0^{(b)}$ using the vector field produced by the same model but with a different chronological condition $v_{\\theta}(x_t,t,y_b)$ .", + "bbox": [ + 89, + 579, + 483, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 4 provides a conceptual illustration of our method using a 2D synthetic toy example, where trajectories depict the temporal evolution of data points. We construct temporally coherent toy data pairs to enable access to ground-truth data pairs. A single model is trained to perform conditional mappings across five distinct input conditions (indicated by the x-axis). Although attributes (indicated by color) gradually blend or evolve, the underlying optimal transport plan is preserved. Notably, the generated samples consistently align with the correct target modes, despite the absence of", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/88e743fee0b07984407265cad180dfda0c28dddeb321d05ba974f66515708f8f.jpg", + "image_caption": [ + "Figure 3. Approach overview. $a)$ . Training a single $v_{\\theta}$ between style embeddings and structured noise conditioned on $y$ . $b)$ . During inference, we can guide the ODE forward and backward direction with different chronological conditions $y$ between the style and latent noise space. By using different $y$ , the style embedding can be transformed to a different style manifold w.r.t. to historical patterns.", + "Figure 4. Illustrational example with ground truth correspondences. Dots sharing the same color are permuted versions along the x-axis. In the figure, colored round dots represent the ground truth samples. purple stars (starting from the 2nd column) indicate the warped samples originating from cyan stars (leftmost)." + ], + "image_footnote": [], + "bbox": [ + 563, + 99, + 851, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "explicit mode supervision, demonstrating robust structural consistency across varying style dynamics.", + "bbox": [ + 511, + 344, + 906, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Dataset Construction", + "text_level": 1, + "bbox": [ + 511, + 386, + 714, + 401 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The curated dataset is a large-scale dataset comprising over 650,000 artworks spanning a wide range of artistic styles and mediums. As discussed in Sec. 2.2, existing datasets for art lack unified meta-information, have limited access, or are small in size. In contrast, our dataset provides standardized and cleaned metadata aggregated from multiple sources. For example, while datasets like WikiArt [84] may include creation years within filenames, such information is often unstructured. Hence, we normalize these metadata to facilitate chronological modeling and future research.", + "bbox": [ + 511, + 409, + 906, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The available metadata includes title, artist, style, year of creation (Most of the artworks originate from years ranging from 1500 to 2021), media (e.g., oil on canvas, etching, acrylic), geographical location, and image URLs. The images are provided as URLs to public museums and public art gallery websites or indexed to public datasets, where they are under the original license and copyright. An overview of the dataset is provided in Fig. 1. For experimental purposes, we used a $6/2/2$ split for training, validation, and testing, randomly sampled from centuries without selecting styles or artists explicitly to maintain the underlying data distribution.", + "bbox": [ + 511, + 561, + 908, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Deduplication. Due to the diverse sources of art pieces, duplicate samples are inevitable. While this redundancy provides additional meta-information from different sources, it results in variations of potential duplicates. To address this, we deduplicated our dataset following Somepalli et al. [79], using SSCD [59] embeddings. Entries with a cosine similarity above 0.85 were merged, consolidating their metadata, e.g., retaining the English version when multiple languages were available. This process reduced the dataset from approximately 950,000 to 650,000 art pieces, encompassing", + "bbox": [ + 511, + 750, + 910, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5871", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/477f9da8baafc2a11620e86b973189994f7e1b72fddfa7b05379cc9cb23380ac.jpg", + "image_caption": [ + "Figure 5. Stylistic flows over time. We visualize distribution shifts for $\\Delta y = 0, 20, 40, 80$ years. Blue quivers indicate the dominant movements of art pieces into the next time frame and visualize how style distribution flows across time." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 478, + 286 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5,053 artists, 128 styles, 7,908 media, and 1,442 geographical locations. Due to the space limitation, we defer the more details of the dataset in the Appendix D.", + "bbox": [ + 89, + 367, + 483, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 89, + 426, + 215, + 444 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Choice of Style Representations", + "text_level": 1, + "bbox": [ + 89, + 452, + 370, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The concept of style is inherently complex and multifaceted, encompassing aspects such as color, material, atmosphere, design, and structure. From a human perspective [39, 72], defining style is subjective and often varies depending on context. Similarly, in computer vision, style is represented through diverse latent spaces depending on the task. For instance, works on style classification and retrieval [11, 33] focus on different representations than those used in style transfer [20, 95] or in diffusion-based generation [16, 61, 87]. In both domains, style remains fluid and context-sensitive.", + "bbox": [ + 89, + 474, + 483, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To accommodate this ambiguity, our model is designed to be agnostic to the specific choice of style space. In our experiments, we primarily focus on two feature spaces: CLIP [62] and CSD [79]. As shown in Table S3, CSD excels in style-based retrieval, whereas CLIP performs slightly better in year-based retrieval and can be decoded to images with unCLIP [65] (see Figure S15). We refer readers to Appendix A for a detailed comparison of all candidates.", + "bbox": [ + 89, + 626, + 483, + 746 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experimental Results", + "text_level": 1, + "bbox": [ + 89, + 757, + 292, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tracing stylistic flows across time. We visualize the temporal flows of style distributions in Figure 5, showing how our model progressively aligns embeddings of early-20th-century artworks with those from later periods. Each subplot displays kernel density estimates (KDEs) [57] in the CSD space for three sets of samples: the original early-century inputs transported by our model (green), the real artworks from the corresponding future time window (orange), and", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the fixed target distribution of late-20th-century art (blue).", + "bbox": [ + 511, + 90, + 895, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "At $\\Delta t = 0$ , the transformed distribution (green) coincides with the original input, as expected. As $\\Delta t$ increases to 20, 40, and 80 years, the green distribution increasingly aligns with the orange ground-truth distribution for each target decade, despite the absence of sample-level correspondences. This alignment is not exact, but meaningful: the transformed samples retain certain semantic traces of their origin, resulting in a distribution that blends temporal stylistic adaptation with content continuity. Such divergence offers insight into how artistic style may evolve from a given source, even in ways not fully captured by existing data. Eventually, the green distribution approaches the latecentury manifold (blue), indicating that our model learns a smooth and coherent temporal flow (denoted by the blue quivers) from a distribution perspective.", + "bbox": [ + 511, + 107, + 906, + 332 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To bridge population-level dynamics with sample-level behavior, we track a single artwork—Berlin Street Scene (1913)—through time. Its unCLIP renderings reflect plausible stylistic changes in line with historical trends, while preserving core semantics. This illustrates the model's ability to capture stylistic evolution across distributions and within individual samples, which we explore next.", + "bbox": [ + 511, + 333, + 905, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Transforming individual artworks through time. We present visual examples demonstrating how individual art pieces evolve across time by transferring their semantic embedding in the CLIP [62] space and visualizing using the un-CLIP model [66]. In Figure 6, we show the temporal trajectories of canonical works such as Les Demoiselles d'Avignon (1907, Cubism) and Landscape near Chatou (1904, Fauvism). This task is particularly challenging as these pieces are less realistic, making semantic capture more difficult. Despite this, we preserve artistic intent and content over time by operating in a compact semantic space.", + "bbox": [ + 511, + 458, + 906, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, we compare our approach with image editing methods in pixel space [49, 96] in Figure 7. These methods fail to achieve faithful semantic transformations tied to specific periods; instead, they apply superficial style transfers that ignore historical context. In contrast, our method enables long-range interpolation through time. We further analyze the style distributions of both unconditionally generated embeddings and mapped embeddings in Appendix C.2, showing that both closely follow the stylistic distributions.", + "bbox": [ + 511, + 625, + 908, + 761 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Content preservation. Our model captures how artistic style changes over time and how such stylistic evolution influences content, including shifts in scene composition and object substitution, e.g., horses versus motorcycles in Figure 7. Because style and content are often deeply entangled, meaningful temporal transformation requires flexibility. When the threshold parameter $t_{\\mathrm{th}}$ is set too low (Section 3.3), the model is overly constrained to the original input, limit", + "bbox": [ + 511, + 779, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5872", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6b45fb4c8f449c4767de1be949786f7469ea46bda4c95f919999d59315d9b852.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison of editing methods across different historical styles. Each row shows results from a different method. The 1st and 5th columns show original artworks (Les Demoiselles d'Avignon - 1907 and Landscape near Chatou - 1904)." + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 857, + 301 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/eacee91bd1a4d9c733979f3e0448a9ca3a9d45fe10d043964830bc7534a5edbc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodRococoImpressionismRealism
R@1R@5δ↓τ(%)↑R@1R@5δ↓τ(%)↑R@1R@5δ↓τ(%)↑
RF-inversion [68]5.69.30.69083.710.526.30.65876.82.46.50.74966.3
LEDDS ++ [3]12.623.00.52879.415.430.40.58569.69.720.50.50964.9
SDEdit [49]6.315.80.57245.65.313.40.60534.53.712.40.61245.6
SDEdit (fine-tuned)37.942.80.26564.734.147.30.27247.227.538.40.23657.3
Ours72.585.10.02484.865.076.30.02475.858.569.30.02568.6
", + "bbox": [ + 215, + 345, + 782, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quality of the time matching when transferring artworks over time. We sampled a compact subset of a hundred samples for each style to calculate the metrics. We transfer clusters of artworks to randomly selected years within the range of $[-100, 100]$ and compute the mean metrics. For $\\tau$ , we used the nearest $nns[:25]$ neighbors as positive samples and further $nns[50:75]$ as the negative samples for each chosen anchor, where $nns$ denotes its nearest neighbors. See extended table in Table S4.", + "bbox": [ + 88, + 438, + 906, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing stylistic adaptation. This results in hybrid outputs, such as a Steam Horse locomotive, that blend past and future characteristics without fully transitioning.", + "bbox": [ + 88, + 508, + 480, + 554 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Content deviation is a broader challenge across models. Most methods face a trade-off between low-level fidelity and high-level semantic consistency. Pixel-focused techniques [3, 49, 67] often struggle to produce meaningful temporal shifts even with adjusted guidance scales, sometimes leading to failures like buildings morphing into people, as seen in Figure 6. In contrast, our model preserves conceptual content while enabling coherent, temporally grounded transformations, striking a balance between semantic stability and stylistic progression.", + "bbox": [ + 88, + 555, + 482, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative evaluation of time travel. Assessing the quality of temporal transitions in artworks is challenging due to the absence of ground-truth correspondences. However, semi-ground-truth evidence exists: artworks created close in time and close in style are more likely to evolve together rather than disperse. Based on this intuition, we utilized three key metrics:", + "bbox": [ + 88, + 733, + 482, + 838 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Recall@k assesses alignment with the target era by checking whether time-traveled samples retrieve temporally appropriate neighbors in style space.", + "- Compactness $\\delta$ measures how coherent the transformed" + ], + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "samples are by comparing their variance to that of random samples in the target year: $\\frac{\\sigma_{\\text{samples}}^2}{\\sigma_{\\text{target}}^2}$ . A good transformation should preserve this coherence, as source samples are initially similar in style and time, resulting in lower $\\delta$ .", + "bbox": [ + 524, + 508, + 903, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- Triplet consistency $\\tau$ is adapted from triplet loss [23, 73]. It evaluates whether the relative relation among anchor, positive, and negative samples is preserved posttransformation, reflecting the model's ability to maintain local structure.", + "bbox": [ + 513, + 575, + 906, + 650 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate on subsets of 100 samples by selecting localized neighborhoods of artwork within several style labels from the test split. These subsets serve as the starting samples for time mapping. Due to space constraints, we present results for the few most prominent artistic styles in Table 1, though our analysis extends across a much broader range. Since most editing baselines operate directly in pixel space, we perform an additional step of encoding their generated outputs back into the style space to enable a fair comparison, which allows us to compute nearest neighbor recall and compactness in a shared representation space.", + "bbox": [ + 509, + 655, + 908, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Table 1 (we offer more analysis in Table S4 in the Appendix), conventional pixel-based image-editing methods [3, 49, 67] fail to achieve both objectives: they neither align generated samples with the appropriate data manifold, reflected by the lower Recall@ $k$ , nor preserve", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5873", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/797afe895a26f3af82ffb8e307eaddb62531a04aa20e3119afe6d0d11f9e2f06.jpg", + "image_caption": [ + "Figure 7. Flow preserves content semantics across time. We transform a motorcyclist to the year 1800 with increasing $t_{\\mathrm{th}}$ (left to right, top row). Low $t_{\\mathrm{th}}$ leads to incorrect mappings due to limited editing flexibility, while higher values retain semantic identity (\"man on a horse\") with correct stylistic traits." + ], + "image_footnote": [], + "bbox": [ + 125, + 97, + 444, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "local structural coherence, indicated by the higher $\\delta$ ). This is mainly due to their reliance on pixel-level cues, which are insufficient for modeling more profound, semantic style shifts. For instance, RF-Inversion achieves high fidelity editing [67] but leads to minimal pixel shifts, evidenced by the low recall. We address these issues by operating in a compact latent space and focusing on semantic information to better align with historical stylistic flows.", + "bbox": [ + 88, + 378, + 482, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since our method is trained exclusively on the curated dataset, a distributional mismatch may arise when comparing against baseline models such as Stable Diffusion, which are pretrained on broader, general-domain datasets. For a fair comparison, we also fine-tuned the base Stable Diffusion model on our dataset until convergence. Specifically, we adapt the SD 1.5 model using chronologically conditioned prompts of the form \"A painting in {year} style\". This finetuning procedure improves both the synthesis FID in Table 2 and retrieval-based matching performance in Table 1, confirming our dataset's value. However, applying SDEdit remains suboptimal for our task, as the method struggles to perform temporally faithful transformations. This performance gap persists across all evaluated artistic styles, where our method consistently yields more coherent and historically aligned style transitions.", + "bbox": [ + 89, + 501, + 485, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison to synthesis baselines. As previously discussed, for visualization purposes, we project samples into the CLIP [62] embedding space and render them using the unCLIP model [66]. This enables quantitative evaluation using the standard Frechet Inception Distance (FID) [25], which assesses both the realism of the generated outputs and their fidelity to the target temporal style distribution.", + "bbox": [ + 89, + 763, + 482, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compared our method against three Latent Diffusion Models (LDM) variants: Stable Diffusion 1.5 [66], Stable", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ae278c436f6f3fe54e75e246ef8907cc6739900017a4492ca3e3db4f3de4e629.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelFID↓conditioned on century
16th17th18th19th20th
SD 1.5 [66]131.1▼40.4118.9▼37.1103.5▼17.2149.9▼77.9180.4▼101.5
SD 2.1214.4▼123.7147.1▼65.3134.2▼47.9170.2▼98.2167.1▼88.2
SDXL [60]186.6▼95.9142.4▼60.6143.4▼57.1114.8▼42.8146.1▼67.2
FLUX.1-schnell [38]199.2▼108.5142.2▼60.4137.3▼51.0122.8▼50.8126.1▼47.2
SD 1.5 (fine-tuned)111.7▼2196.4▼14.688.8▼2.596.4▼24.4150.7▼71.8
Ours90.781.886.372.078.9
", + "bbox": [ + 517, + 89, + 903, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Feature synthesis across centuries. Evaluation metrics are computed on 1,000 samples per century. While Stable Diffusion and FLUX rely on text prompts for year conditioning, our model conditions directly on this information. Our method exhibits stronger alignment with ground truth distributions.", + "bbox": [ + 511, + 191, + 906, + 262 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Diffusion 2.1 [66], and SDXL [60], as well as the flow-matching based FLUX model [38]. We also compare against the fine-tuned Stable Diffusion 1.5, with results summarized in Table 2. SDXL [60] has demonstrated strong artistic understanding and has become the foundation for many recent stylistic image generation methods [16, 24, 76, 87]. While SDXL performs well on certain styles, such as Renaissance (16th century) and Cubism (20th century) in the Appendix Figure S13, it struggles to capture stylistic distributions comprehensively. FLUX [38], despite its training scale and exhibiting strong performance across text-to-image tasks, also struggles to capture temporal style distributions effectively (see Figure S14). In contrast, our model, trained on the proposed dataset, better captures the diversity of artistic styles. Our approach achieves significantly lower FID scores than standard text-conditioned models, demonstrating better alignment with the underlying feature distribution of different art periods. Furthermore, it provides more consistent stylistic alignment across centuries. Further rendering results for all models are available in the Appendix.", + "bbox": [ + 511, + 271, + 908, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 588, + 633, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Studying the evolution of art is challenging due to the lack of ground-truth correspondences between artworks from different periods. We present a framework that models stylistic progression using stochastic interpolants and dual diffusion implicit bridges, enabling distribution-level alignment without paired data. By formulating art evolution as optimal transport in a learned style space, our method captures continuous temporal transitions in artistic style.", + "bbox": [ + 511, + 613, + 906, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A central contribution is a dataset of over 650,000 artworks spanning 500 years, which we use to demonstrate that our model captures long-range stylistic shifts. Beyond modeling distributions, our method also enables tracing the stylistic origins and plausible futures of individual artworks, allowing for fine-grained longitudinal analysis. Our work explores a less-studied direction in computer vision: using generative models to investigate the evolution of art and support cultural and historical analysis. We believe our approach offers a foundation for future interdisciplinary work at the intersection of machine learning and cultural analytics.", + "bbox": [ + 511, + 734, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5874", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This project was supported by Bayer AG, the Federal Ministry for Economic Affairs and Energy within the project \"NXT GEN AI METHODS - Generative Methoden für Perzeption, Prädiktion und Planung\", the project \"Genius-Robot\" (01IS24083) funded by the Federal Ministry of Research, Technology and Space (BMFTR), and the bidt project KLIMA-MEMES. The authors gratefully acknowledge the Gauss Center for Supercomputing for providing compute through the NIC on JUWELS/JUPITER at JSC and the HPC resources supplied by the NHR @FAU Erlangen.", + "bbox": [ + 89, + 114, + 485, + 267 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 280, + 187, + 296 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Michael S Albergo, Nicholas M Boffi, and Eric Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions. arXiv, 2023. 1, 3, 4", + "[2] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In ICCV, 2019. 2", + "[3] Manuel Brack, Felix Friedrich, Katharia Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. Ledits++: Limitless image editing using text-to-image models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8861-8870, 2024. 7", + "[4] Levin Brinkmann, Deniz Gezerli, KV Kleist, Thomas F Müller, Iyad Rahwan, and Niccolò Pescetelli. Hybrid social learning in human-algorithm cultural transmission. Philosophical Transactions of the Royal Society A, 380(2227), 2022. 1", + "[5] Eric Ming Chen, Jin Sun, Apoorv Khandelwal, Dani Lischinski, Noah Snavely, and Hadar Averbuch-Elor. What's in a decade? transforming faces through time. In Computer Graphics Forum, 2023. 2", + "[6] Haibo Chen, Lei Zhao, Zhizhong Wang, Huiming Zhang, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Artistic style transfer with internal-external learning and contrastive learning. Advances in Neural Information Processing Systems (NeurIPS), 34, 2021. 2", + "[7] Ricky TQ Chen and Yaron Lipman. Riemannian flow matching on general geometries. arXiv, 2023. 3", + "[8] Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. NeurIPS, 2018. 3", + "[9] Yongxin Chen, Tryphon T Georgiou, and Michele Pavon. On the relation between optimal transport and schrödinger bridges: A stochastic control viewpoint. Journal of Optimization Theory and Applications, 169:671-691, 2016. 4", + "[10] Ceren Comert, Murat Özbayoglu, and Cosku Kasnakoğlu. Painter prediction from artworks with transfer learning. In 2021 7th International Conference on Mechatronics and Robotics Engineering (ICMRE). IEEE, 2021. 1", + "[11] Elliot Crowley and Andrew Zisserman. The state of the art: Object retrieval in paintings using discriminative regions. In" + ], + "bbox": [ + 101, + 304, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the British Machine Vision Conference 2014. British Machine Vision Association, 2014. 6", + "[12] Aram Davtyan, Sepehr Sameni, and Paolo Favaro. Randomized conditional flow matching for video prediction. arXiv, 2022. 3", + "[13] Valentin De Bortoli, James Thornton, Jeremy Heng, and Arnaud Doucet. Diffusion schrödinger bridge with applications to score-based generative modeling. Advances in Neural Information Processing Systems, 2021. 1", + "[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, 2009. 2", + "[15] Yingying Deng, Fan Tang, Weiming Dong, Chongyang Ma, Feiyue Huang, Oliver Deussen, and Changsheng Xu. Exploring the representativity of art paintings. IEEE Transactions on Multimedia, 2020. 2", + "[16] Yarden Frenkel, Yael Vinker, Ariel Shamir, and Daniel Cohen-Or. Implicit style-content separation using b-lora. In ECCV, 2024. 6, 8", + "[17] Michael Fuest, Pingchuan Ma, Ming Gui, Johannes S Fischer, Vincent Tao Hu, and Bjorn Ommer. Diffusion models and representation learning: A survey. arXiv preprint arXiv:2407.00783, 2024. 3", + "[18] Shangqi Gao, Hangqi Zhou, Yibo Gao, and Xiahai Zhuang. Bayeseg: Bayesian modeling for medical image segmentation with interpretable generalizability. Medical Image Analysis, 89, 2023. 1", + "[19] Noa Garcia, Benjamin Renoust, and Yuta Nakashima. Contextnet: representation and exploration for painting classification and retrieval in context. International Journal of Multimedia Information Retrieval, 9(1), 2020. 1, 2", + "[20] Leon Gatys, Alexander S Ecker, and Matthias Bethge. Texture synthesis using convolutional neural networks. In Advances in Neural Information Processing Systems (NeurIPS), 2015. 6", + "[21] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2", + "[22] Ming Gui, Johannes Schusterbauer, Ulrich Prestel, Pingchuan Ma, Dmytro Kotovenko, Olga Grebenkova, Stefan Andreas Baumann, Vincent Tao Hu, and Björn Ommer. Depthfm: Fast monocular depth estimation with flow matching. AAAI, 2025. 3", + "[23] Alexander Hermans, Lucas Beyer, and Bastian Leibe. In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737, 2017. 7", + "[24] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. In CVPR, 2024. 8", + "[25] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 8", + "[26] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems (NeurIPS), 33, 2020. 3" + ], + "bbox": [ + 524, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5875", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Vincent Tao Hu, Yunlu Chen, Mathilde Caron, Yuki M. Asano, Cees G.M. Snoek, and Björn Ommer. Guided diffusion from self-supervised diffusion features. In ARXIV, 2023. 3", + "[28] Vincent Tao Hu, Wenzhe Yin, Pingchuan Ma, Yunlu Chen, Basura Fernando, Yuki M. Asano, Efstratos Gavves, Pascal Mettes, Björn Ommer, and Cees G.M. Snoek. Motion flow matching for human motion synthesis and editing. In ARXIV, 2023. 3", + "[29] Vincent Tao Hu, Stefan Andreas Baumann, Ming Gui, Olga Grebenkova, Pingchuan Ma, Johannes Fischer, and Björn Ommer. *Zigma: A dit-style zigzag mamba diffusion model*. In European conference on computer vision, pages 148-166. Springer, 2024. 3", + "[30] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2017. 2", + "[31] Herve Jegou, Matthijs Douze, and Cordelia Schmid. Product quantization for nearest neighbor search. IEEE transactions on pattern analysis and machine intelligence, 33(1):117-128, 2010. 1", + "[32] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Proceedings of the European Conference on Computer Vision (ECCV). Springer, 2016. 2", + "[33] Sergey Karayev, Matthew Trentacoste, Helen Han, Aseem Agarwala, Trevor Darrell, Aaron Hertzmann, and Holger Winnemoeller. Recognizing image style. arXiv preprint arXiv:1311.3715, 2013. 6", + "[34] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, 2022. 3", + "[35] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 1", + "[36] Dmytro Kotovenko, Matthias Wright, Arthur Heimbrecht, and Bjorn Ommer. Rethinking style transfer: From pixels to parameterized brushstrokes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[37] Felix Krause, Timy Phan, Ming Gui, Stefan Andreas Baumann, Vincent Tao Hu, and Björn Ommer. Tread: Token routing for efficient architecture-agnostic diffusion training. arXiv preprint arXiv:2501.04765, 2025. 3", + "[38] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024.8, 2", + "[39] Berel Lang. The concept of style. Cornell University Press, 1987. 6", + "[40] Christian Léonard. A survey of the schr\\'' odinger problem and some of its connections with optimal transport. arXiv preprint arXiv:1308.0215, 2013. 4", + "[41] Wen Li, Muyuan Fang, Cheng Zou, Biao Gong, Ruobing Zheng, Meng Wang, Jingdong Chen, and Ming Yang. Style-tokensizer: Defining image style by a single instance for controlling diffusion models. In European Conference on Computer Vision, pages 110-126. Springer, 2024. 2" + ], + "bbox": [ + 99, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Xueting Li, Sifei Liu, Jan Kautz, and Ming-Hsuan Yang. Learning linear transformations for fast image and video style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[43] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2", + "[44] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. *ICLR*, 2023. 3", + "[45] Guan-Horng Liu, Tianrong Chen, Oswin So, and Evangelos Theodorou. Deep generalized schrödinger bridge. NeurIPS, 2022. 3", + "[46] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. *ICLR*, 2023. 3, 4", + "[47] Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. arXiv, 2024. 3, 4", + "[48] L. McInnes, J. Healy, and J. Melville. UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction. ArXiv, 2018. 1", + "[49] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 6, 7, 3", + "[50] Amir Merchant, Simon Batzner, Samuel S Schoenholz, Muratahan Aykol, Gowoon Cheon, and Ekin Dogus Cubuk. Scaling deep learning for materials discovery. Nature, 624 (7990), 2023. 1", + "[51] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, 2024. 1", + "[52] Kevin Musgrave, Serge Belongie, and Ser-Nam Lim. A metric learning reality check. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16, pages 681-699. Springer, 2020. 1", + "[53] Ivona Najdenkoska, Xiantong Zhen, Marcel Worring, and Ling Shao. Uncertainty-aware report generation for chest x-rays by variational topic inference. Medical Image Analysis, 82, 2022. 1", + "[54] Kirill Neklyudov, Daniel Severo, and Alireza Makhzani. Action matching: A variational method for learning stochastic dynamics from samples. arXiv, 2022. 3", + "[55] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bo" + ], + "bbox": [ + 522, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5876", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "janowski. Dinov2: Learning robust visual features without supervision, 2023. 3", + "[56] Norihiro Oyama, Noriko N Ishizaki, Satoshi Koide, and Hiroaki Yoshida. Deep generative model super-resolves spatially correlated multiregional climate data. Scientific Reports, 13(1), 2023. 1", + "[57] Emanuel Parzen. On estimation of a probability density function and mode. The annals of mathematical statistics, 33(3):1065-1076, 1962. 6", + "[58] Pablo Pernias, Dominic Rampas, Mats Leon Richter, Christopher Pal, and Marc Aubreville. Würstchen: An efficient architecture for large-scale text-to-image diffusion models. In The Twelfth International Conference on Learning Representations, 2023. 1", + "[59] Ed Pizzi, Sreya Dutta Roy, Sugosh Nagavara Ravindra, Priya Goyal, and Matthijs Douze. A self-supervised descriptor for image copy detection. In CVPR, 2022. 5", + "[60] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 8, 2", + "[61] Tianhao Qi, Shancheng Fang, Yanze Wu, Hongtao Xie, Jiawei Liu, Lang Chen, Qian He, and Yongdong Zhang. Deadiff: An efficient stylization diffusion model with disentangled representations. arXiv preprint arXiv:2403.06951, 2024. 2, 6", + "[62] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning. PMLR, 2021. 2, 6, 8, 1", + "[63] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML). PMLR, 2021. 3", + "[64] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 1", + "[65] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 6, 1, 2", + "[66] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022. 1, 3, 6, 8, 2", + "[67] L Rout, Y Chen, N Ruiz, C Caramanis, S Shakkottai, and W Chu. Semantic image inversion and editing using rectified stochastic differential equations. 2024. 4, 7, 8", + "[68] Litu Rout, Yujia Chen, Nataniel Ruiz, Constantine Caramanis, Sanjay Shakkottai, and Wen-Sheng Chu. Semantic image inversion and editing using rectified stochastic differential equations. In ICLR, 2025. 7" + ], + "bbox": [ + 99, + 90, + 483, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[69] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1", + "[70] Dan Ruta, Saeid Motiian, Baldo Faieta, Zhe Lin, Hailin Jin, Alex Filipkowski, Andrew Gilbert, and John Collomosse. Aladin: All layer adaptive instance normalization for fine-grained style similarity. In ICCV, 2021. 1, 3", + "[71] Babak Saleh and Ahmed Elgammal. Large-scale classification of fine-art paintings: Learning the right metric on the right feature. arXiv preprint arXiv:1505.00855, 2015. 1, 3", + "[72] Meyer Schapiro. Theory and philosophy of art: Style, artist, and society. 1994. 6", + "[73] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 7", + "[74] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 3, 2", + "[75] Johannes Schusterbauer, Ming Gui, Pingchuan Ma, Nick Stracke, Stefan A Baumann, and Björn Ommer. Boosting latent diffusion with flow matching. ECCV, 2024. 3, 4", + "[76] Viraj Shah, Nataniel Ruiz, Forrester Cole, Erika Lu, Svetlana Lazebnik, Yuanzhen Li, and Varun Jampani. Ziplora: Any subject in any style by effectively merging loras. In ECCV, 2024. 8", + "[77] Higor YD Sigaki, Matjaž Perc, and Haroldo V Ribeiro. History of art paintings through the lens of entropy and complexity. Proceedings of the National Academy of Sciences, 115(37):E8585-E8594, 2018. 3", + "[78] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 3", + "[79] Gowthami Somepalli, Anubhav Gupta, Kamal Gupta, Shramay Palta, Micah Goldblum, Jonas Geiping, Abhinav Shrivastava, and Tom Goldstein. Measuring style similarity in diffusion models. ECCV, 2024. 2, 3, 5, 6, 1", + "[80] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3", + "[81] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 3, 4", + "[82] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 2", + "[83] Xuan Su, Jiaming Song, Chenlin Meng, and Stefano Ermon. Dual diffusion implicit bridges for image-to-image translation. *ICLR*, 2023. 1, 3, 4", + "[84] Wei Ren Tan, Chee Seng Chan, Hernan Aguirre, and Kiyoshi Tanaka. Improved artgan for conditional synthesis of natural image and artwork. IEEE Transactions on Image Processing, 28(1):394-409, 2019. 5" + ], + "bbox": [ + 522, + 92, + 906, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "5877", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[85] Zhicong Tang, Tiankai Hang, Shuyang Gu, Dong Chen, and Baining Guo. Simplified diffusion schrödinger bridge. arXiv preprint arXiv:2403.14623, 2024. 1, 3, 4", + "[86] Gemma Canet Tarrés, Dan Ruta, Tu Bui, and John Collomosse. Parasol: Parametric style control for diffusion image synthesis. arXiv preprint arXiv:2303.06464, 2023. 2", + "[87] Haofan Wang, Matteo Spinelli, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv preprint arXiv:2404.02733, 2024. 6, 8", + "[88] Sheng-Yu Wang, Alexei A Efros, Jun-Yan Zhu, and Richard Zhang. Evaluating data attribution for text-to-image models. In ICCV, 2023. 2", + "[89] Zhizhong Wang, Lei Zhao, Sihuan Lin, Qihang Mo, Huiming Zhang, Wei Xing, and Dongming Lu. Glstylenet: exquisite style transfer combining global and local pyramid features. IET Computer Vision, 14(8), 2020. 2", + "[90] Zhizhong Wang, Lei Zhao, Haibo Chen, Ailin Li, Zhiwen Zuo, Wei Xing, and Dongming Lu. Texture reformer: Towards fast and universal interactive texture transfer. In Proceedings of the AAAI Conference on Artificial Intelligence, 2022. 2", + "[91] Zhizhong Wang, Lei Zhao, Haibo Chen, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. DivSwapper: Towards diversified patch-based arbitrary style transfer. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence (IJCAI), 2022. 2", + "[92] Zhizhong Wang, Lei Zhao, Zhiwen Zuo, Ailin Li, Haibo Chen, Wei Xing, and Dongming Lu. Microast: Towards super-fast ultra-resolution arbitrary style transfer. In Proceedings of the AAAI Conference on Artificial Intelligence, 2023. 2", + "[93] Michael J Wilber, Chen Fang, Hailin Jin, Aaron Hertzmann, John Collomosse, and Serge Belongie. Bam! the behave artistic media dataset for recognition beyond photography. In ICCV, 2017. 1, 3", + "[94] Lemeng Wu, Dilin Wang, Chengyue Gong, Xingchao Liu, Yunyang Xiong, Rakesh Ranjan, Raghuraman Krishnamoorthi, Vikas Chandra, and Qiang Liu. Fast point cloud generation with straight flows. arXiv, 2022. 3", + "[95] Zijie Wu, Zhen Zhu, Junping Du, and Xiang Bai. Ccpl: Contrastive coherence preserving loss for versatile style transfer. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 6", + "[96] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 6", + "[97] Sheng Zhang, Yanbo Xu, Naoto Usuyama, Hanwen Xu, Jaspreet Bagga, Robert Tinn, Sam Preston, Rajesh Rao, Mu Wei, Naveen Valluri, et al. Biomedclip: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs. arXiv preprint arXiv:2303.00915, 2023. 2", + "[98] Yuxin Zhang, Fan Tang, Weiming Dong, Haibin Huang, Chongyang Ma, Tong-Yee Lee, and Changsheng Xu. Domain enhanced arbitrary image style transfer via contrastive learning. In ACM SIGGRAPH 2022 Conference Proceedings, 2022. 2" + ], + "bbox": [ + 99, + 92, + 483, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[99] Wentao Zhao, Dalin Zhou, Xinguo Qiu, and Wei Jiang. How to represent paintings: A painting classification using artistic comments. Sensors, 21(6), 2021. 1, 2", + "[100] Linqi Zhou, Aaron Lou, Samar Khanna, and Stefano Ermon. Denoising diffusion bridge models. *ICLR*, 2024. 1, 3", + "[101] Zhiwen Zuo, Lei Zhao, Shuobin Lian, Haibo Chen, Zhizhong Wang, Ailin Li, Wei Xing, and Dongming Lu. Style fader generative adversarial networks for style degree controllable artistic style transfer. In Proc. Int. Joint Conf. on Artif. Intell.(IJCAI), 2022. 2" + ], + "bbox": [ + 516, + 92, + 906, + 232 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "5878", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Stochastic Interpolants for Revealing Stylistic Flows across the History of Art/16c23549-7e7f-4e3a-8944-314b6bb00228_model.json b/2025/Stochastic Interpolants for Revealing Stylistic Flows across the History of Art/16c23549-7e7f-4e3a-8944-314b6bb00228_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b9f6baf04e29be4772bbba7975ade8f1886704d1 --- /dev/null +++ b/2025/Stochastic Interpolants for Revealing Stylistic Flows across the History of Art/16c23549-7e7f-4e3a-8944-314b6bb00228_model.json @@ -0,0 +1,2699 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.131, + 0.89, + 0.153 + ], + "angle": 0, + "content": "Stochastic Interpolants for Revealing Stylistic Flows across the History of Art" + }, + { + "type": "text", + "bbox": [ + 0.208, + 0.181, + 0.79, + 0.218 + ], + "angle": 0, + "content": "Pingchuan Ma* Ming Gui* Johannes Schusterbauer \nXiaopei Yang Olga Grebenkova Vincent Tao Hu Björn Ommer" + }, + { + "type": "text", + "bbox": [ + 0.176, + 0.234, + 0.821, + 0.253 + ], + "angle": 0, + "content": "CompVis @ LMU Munich Munich Center for Machine Learning (MCML)" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.319, + 0.486, + 0.591 + ], + "angle": 0, + "content": "Generative models have made rapid progress in content creation, particularly in synthesizing artworks and capturing stylistic variation. However, most methods operate at the level of individual images, limiting their ability to reveal broader stylistic trends and temporal transitions. We address this by introducing a framework that models stylistic evolution as an optimal transport problem in a learned style space, using stochastic interpolants and dual diffusion implicit bridges to align artistic distributions across time without requiring paired data. A central contribution is a diverse dataset of over 650,000 artworks spanning 500 years, curated with metadata across multiple genres. Together, our method and dataset enable tracing long-range stylistic transitions and plausible futures of individual artworks, supporting fine-grained temporal analysis. This offers a new tool for modeling historical patterns in visual culture and opens up promising directions in visual understanding. Code and dataset: https://github.com/CompVis/Art-fm" + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.622, + 0.222, + 0.637 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.648, + 0.484, + 0.768 + ], + "angle": 0, + "content": "Recent advances in computer vision have sparked growing interest in applications in interdisciplinary research domains. Such applications not only try to offer novel perspectives, but also address challenges that were impossible before, such as those seen in medical imaging [18, 53], novel detection in material science [50], and simulation of climate [56]. Despite these successes, there remains a noticeable gap in their use for cultural studies [4], particularly in art history." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.77, + 0.484, + 0.875 + ], + "angle": 0, + "content": "With the rise of generative models such as Stable Diffusion [66], DALL-E [64], and others [51, 58, 69], there has been increasing use of these tools for artwork synthesis and content creation. However, their primary focus remains on image generation and editing, rather than analyzing the broader narrative of art history [10, 19, 99]. As a result, current models are not designed to capture the complex evo" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.289, + 0.84, + 0.303 + ], + "angle": 0, + "content": "lution of artistic movements and styles over time." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.304, + 0.908, + 0.501 + ], + "angle": 0, + "content": "In this paper, we ask: How would a piece of art evolve? What might it look like if created centuries earlier or later? Naturally, no ground-truth pairs exist for such a task, and the artistic style of a creator is inherently dynamic and historically constrained. To address this challenge, we aim to analyze the flow of artistic styles over time. Given a single artwork, we try to determine its closest \"projection\" onto the manifold of artistic expressions corresponding to a given historical period. While optimal transport provides a natural formulation for such a task, considering only a single isolated artwork is insufficient. Instead, we reformulate the problem as a distribution matching task, where we study the evolution of the underlying distribution of art over time." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.502, + 0.909, + 0.683 + ], + "angle": 0, + "content": "Recent advances in diffusion bridges such as DDIB [83], DSB [13, 85], and DDBMs [100] have shown that domain translation can be performed implicitly, without the need for paired supervision. We build on this idea in the context of art history by incorporating stochastic interpolants [1], enabling us to model the temporal evolution of artistic styles without requiring explicitly aligned training data. Unlike traditional style transfer methods that transform images based on predefined and discrete styles, our approach projects styles onto their corresponding distributions across time. Operating in a compact latent space allows us to abstract away pixel-level distraction and focus on meaningful stylistic shifts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.685, + 0.909, + 0.865 + ], + "angle": 0, + "content": "To support this temporal modeling, we require a dataset that spans centuries and reflects a wide variety of artistic styles. While datasets like WikiArt [71] offer breadth in genre and metadata, they lack the scale and historical depth needed to model long-term stylistic progression. Others, such as BAM [93] and BAM-FG [70], are either restricted in access or skewed toward modern art. To address this, we curate a new unified dataset of over 650,000 artworks annotated with creation year, location, artist, medium, and other metadata (Fig. 1). This dataset enables the study of artistic evolution at scale and over time. In summary, our key contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "- We propose the first generative framework that explicitly models the temporal progression of artistic styles across" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.221, + 0.9 + ], + "angle": 0, + "content": "*Equal Contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5867" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.092, + 0.881, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.463, + 0.908, + 0.491 + ], + "angle": 0, + "content": "Figure 1. Overview of the proposed dataset. We have curated and unified our large-scale dataset, including 650,000 artworks from different creation periods (ranging from 1500 to 2021) and styles with other meta-information, e.g., year, title, artist, media, etc." + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.518, + 0.461, + 0.532 + ], + "angle": 0, + "content": "historical periods, introduced in Sec. 3.2 and Sec. 3.3." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.533, + 0.483, + 0.594 + ], + "angle": 0, + "content": "- To support this, we construct a large-scale, unified dataset of 650,000 artworks annotated with creation year and other metadata, covering five centuries of diverse artistic styles (Sec. 3.4). An overview is shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.594, + 0.484, + 0.639 + ], + "angle": 0, + "content": "- Our method models temporal transitions without requiring paired data. We outperform competitive baselines in capturing long-range stylistic shifts (Sec. 4.2)." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.533, + 0.484, + 0.639 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.656, + 0.241, + 0.672 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.304, + 0.698 + ], + "angle": 0, + "content": "2.1. Stylistic representation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.486, + 0.901 + ], + "angle": 0, + "content": "The pioneering work by Gatys et al. [21] marked the beginning of the style transfer era, which defines style as Gram Matrices of pre-trained network features. Various methods have been proposed to improve different aspects of style representation, ranging from efficiency-focused approaches like [32, 42, 92] to quality-centric ones such as [36, 89, 98, 101]. Moreover, efforts have been made to enhance the generality and diversity of style transfer methods [30, 90, 91, 95]. Despite these achievements, most existing approaches [6, 41, 61, 86, 98] focus on image editing or the disentanglement of style and content through explicit style/content representations or user-defined prompts. Data-driven methods as [15] defined the representativity of art" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.518, + 0.908, + 0.683 + ], + "angle": 0, + "content": "works as a deep feature fusion incorporating not only images, but also painting styles, creation time, and authorship information. On the other hand, ContextNet [19] aims to represent and explore artistic attributes in different modalities and their relationships to improve the retrieval of artworks in automatic art analysis. ArtGCN [99] uses a graph CNN to construct stylistic representation based on artistic comments. More recent works, such as GDA [88] and CSD [79], train style descriptors by fine-tuning on top of the existing feature space of CLIP [62] through contrastive learning with synthetic or curated data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.685, + 0.909, + 0.837 + ], + "angle": 0, + "content": "These studies typically concentrate on rendering or differentiating individual images with specific styles and often overlook treating styles as distributions. Chen et al. [5] tried to match portrait images through time with a family of models for each decade and re-synthesize them. However, it is limited to simple face data. In our work, we focus on analyzing artistic style shifts in artworks depicting more complicated scenes from different eras and using a curated dataset with newly proposed flow-based models, which have not been explored before in similar analyses." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.644, + 0.863 + ], + "angle": 0, + "content": "2.2. Art datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Various datasets focus on different domains [2, 14, 43, 82, 97]. However, in the domain of art, particularly paintings" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5868" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.219, + 0.104, + 0.782, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.207, + 0.908, + 0.251 + ], + "angle": 0, + "content": "Figure 2. Implicit bridge with 2D example. Two unconditional models are trained between Gaussian noise and data distributions \\(\\mathcal{A}\\) (squares) and \\(\\mathcal{B}\\) (circles), without access to mode labels (colors). Despite this, structure is preserved in noise space. By mapping samples from \\(\\mathcal{A}\\) to noise via \\(v^{(a)}\\) and then back to \\(\\mathcal{B}\\) via \\(v^{(b)}\\), we show that correspondences can emerge without paired supervision." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.269, + 0.486, + 0.48 + ], + "angle": 0, + "content": "concerning artistic style, there is a notable absence of large-scale datasets tailored for comprehensive analysis. Existing art datasets exhibit several limitations. WikiArt [71] provides a diverse collection of styles and includes year information in some of the filenames, but is limited in the number of samples to facilitate a comprehensive analysis of artistic evolution. BAM [93] and BAM-FG [70] are restricted in accessibility as they are not publicly available and primarily consist of contemporary artworks, thereby lacking a historical perspective essential for studying stylistic progression. LAION-Styles [79] has curated a large style attribution dataset from LAION [74]. However, it lacks chronological information and does not differentiate between styles and artists, which is crucial for analyzing the evolution of styles." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.485, + 0.487, + 0.562 + ], + "angle": 0, + "content": "To the best of our knowledge, no previous work has curated a large-scale dataset of artwork with detailed metadata, including style, artist, year of creation, and medium. Our curated dataset (see Sec. 3.4) unifies these aspects, enabling scalable studies beyond existing approaches." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.587, + 0.46, + 0.604 + ], + "angle": 0, + "content": "2.3. Flow Matching and Stochastic Interpolants" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.614, + 0.487, + 0.902 + ], + "angle": 0, + "content": "Flow Matching [44, 46, 54] has established Continuous Normalizing Flow (CNF) as a notable competitor to diffusion models [26, 78]. It enables simulation-free training of CNFs using an optimal transport conditional probability path with Ordinary Differential Equations (ODE), providing a more efficient alternative to traditional diffusion training and sampling methods [34, 81]. The versatility of Flow Matching has been showcased across various domains, including image [27, 29, 37, 44, 75], depth map [22], video [12], human motion [28], point cloud [94], and manifold data [7]. A more recent family of frameworks - Stochastic Interpolants [1, 47] has been proposed for unifying Flows and Diffusions. They suggested that Stochastic Interpolants can be generalized to align with various generative models, including Normalizing Flow [8], Diffusion Model [66, 80], Flow Matching [22, 29, 75], and Schrödinger Bridge [45, 85]. In this work, we develop our method within the stochastic interpolant framework and use the learned optimal transport paths (flows) to study the progression of styles across eras." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.267, + 0.606, + 0.283 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.909, + 0.504 + ], + "angle": 0, + "content": "We aim to enable the matching of artworks across different time periods, despite the absence of ground-truth correspondences. Analyzing the temporal evolution of individual paintings is inherently infeasible, as artworks are static and finalized entities. Future stylistic transformations are highly unpredictable, even when considering subsequent works by the same artist, given that many artists undergo significant stylistic changes throughout their careers (e.g., Van Gogh or Pablo Picasso). Due to this absence of explicit ground-truth data pairs, we propose an alternative formulation of the problem. Specifically, we approach it as a distribution matching task, where the evolution of artistic styles over time is understood in terms of underlying probability distributions rather than isolated instances." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.909, + 0.688 + ], + "angle": 0, + "content": "By modelling on the global distribution of artistic styles across time, we can establish meaningful correspondences spanning different centuries [77]. A principled framework for matching distributions is Optimal Transport (OT), which minimizes the transportation costs between distributions [83, 100]. To achieve this, we leverage implicit bridges and flow matching, which, upon training, inherently define an Ordinary Differential Equation (ODE) that aligns with the principles of optimal transport. This not only captures high-level stylistic transitions over time but also facilitates fine-grained alignment of individual artworks based on their placement within these temporally evolving distributions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.699, + 0.738, + 0.715 + ], + "angle": 0, + "content": "3.1. Theoretical background" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.722, + 0.909, + 0.813 + ], + "angle": 0, + "content": "Stochastic Interpolants In recent years, generative probabilistic models that transform noise into data have gained significant attention. These models are primarily represented by diffusion-based models [17, 66, 80, 81] and flow-based models [26, 44]. This transformation can be defined [1] by a time-dependent forward process with \\( t \\in [0,1] \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.832, + 0.908, + 0.846 + ], + "angle": 0, + "content": "\\[\nx _ {t} = \\alpha_ {t} x _ {0} + \\sigma_ {t} x _ {1}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.912, + 0.903 + ], + "angle": 0, + "content": "where \\( x_0 \\) corresponds to data sampled from the data distribution \\( p_0(x) \\) for which we have empirical observations available and \\( x_1 \\sim \\mathcal{N}(0, \\mathbf{I}) \\) is a simple prior distribution." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5869" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.091, + 0.091, + 0.486, + 0.122 + ], + "angle": 0, + "content": "Algorithm 1 Transforming artworks from source age distribution to target age distribution" + }, + { + "type": "algorithm", + "bbox": [ + 0.09, + 0.123, + 0.482, + 0.282 + ], + "angle": 0, + "content": "Input: \\(x_0^{(a)}\\sim p_A(x)\\) - data samples from age A \\(y_{a}\\) - original chronological conditioning \\(y_{b}\\) -chronological conditioning for desired age \\(y_{b} = y_{a} + \\Delta y\\) \\(v_{\\theta}^{(a)}\\) - model corresponds to age A, instantiated by \\(v_{\\theta}(x_t,t,y_a)\\) \\(v_{\\theta}^{(b)}\\) - target model corresponds to age B, instantiated by \\(v_{\\theta}(x_t,t,y_b)\\) \\(t_\\mathrm{th}\\in (0,1]\\) - the integration threshold \nOutput: \\(x_0^{(b)}\\) , the result in the target age data distribution B. 1: for each element \\(x_0^{(a)}\\sim p_A(x)\\) do 2: \\(\\begin{array}{rlr}{x_{\\mathrm{th}} = \\mathrm{ODESolve}(x_0^{(a)};v_{\\theta}^{(a)},0,t_{\\mathrm{th}})} & \\triangleright & \\mathrm{obtain~the~structured~noise~from}\\\\ {\\mathrm{source~age~data}} & {} & \\\\ {x_0^{(b)} = \\mathrm{ODESolve}(x_{\\mathrm{th}};v_{\\theta}^{(b)},t_{\\mathrm{th}},0)} & \\triangleright & \\mathrm{obtain~target~age~data~from~the}\\\\ {\\mathrm{structured~noise}} & {} & \\\\ {\\mathrm{return} x_0^{(b)}} & {} & \\end{array}\\) 3:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.309, + 0.487, + 0.462 + ], + "angle": 0, + "content": "The forward process is characterized by the hyperparameters \\(\\alpha_{t}\\) and \\(\\sigma_{t}\\), with \\(\\alpha_{t}\\) decreasing and \\(\\sigma_{t}\\) increasing as time \\(t\\in [0,1]\\) progresses, interpolating between data and noise. Furthermore, the boundary conditions are normally constrained as \\(\\alpha_0 = \\sigma_1 = 1\\) and \\(\\alpha_{1} = \\sigma_{0} = 0\\), so that for \\(t = 0\\) we have pure data and \\(t = 1\\) pure noise. In particular, flow-based models, represented by SiT [47, 75], have attracted significant interest due to their faster training and inference compared to diffusion-based models. With the above-defined forward process, the velocity is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.207, + 0.472, + 0.484, + 0.49 + ], + "angle": 0, + "content": "\\[\nv (x, t) = \\mathbb {E} [ \\dot {x} _ {t} | x _ {t} = x ], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.5, + 0.485, + 0.622 + ], + "angle": 0, + "content": "which generates the marginal probability distribution \\( p_t(x) \\) of \\( x_{t} \\) at time \\( t \\) [1, 47]. This aligns with the distributions of the probability flow ODE [81] so that during inference we can generate samples by solving the Ordinary Differential Equation (ODE) backward in time starting from pure noise \\( \\mathcal{N}(0,\\mathbf{I}) \\) to arrive at a sample \\( x\\sim p_0(x) \\). Ma et al. [47] demonstrates that we can train a neural network \\( v_{\\theta}(x_{t},t) \\) to parametrize the velocity using the training objective:" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.644, + 0.484, + 0.68 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {v}} (\\theta) = \\int_ {0} ^ {T} \\mathbb {E} [ \\| v _ {\\theta} (x _ {t}, t) - \\dot {\\alpha} _ {t} x _ {0} - \\dot {\\sigma} _ {t} x _ {1} \\| ^ {2} ] \\mathrm {d} t, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.486, + 0.749 + ], + "angle": 0, + "content": "with \\(\\dot{\\alpha}_{t}\\) and \\(\\dot{\\sigma}_t\\) representing the time derivative of \\(\\alpha_{t}\\) and \\(\\sigma_t\\), respectively. For simplicity and due to its relatively straight trajectories, we adopt the Linear Interpolant from [46, 47] with \\(\\alpha_{t} = 1 - t\\) and \\(\\sigma_t = t\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.347, + 0.774 + ], + "angle": 0, + "content": "3.2. Bridging artworks over time" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.485, + 0.901 + ], + "angle": 0, + "content": "We aim to establish meaningful correspondences between artworks originating from different historical periods. By leveraging and adapting the recent advances in generative modelling, we develop a framework that learns to transport dynamics across time-conditioned data distributions without requiring paired supervision. We begin by introducing the foundations of implicit diffusion bridges, before detailing how they can be applied to match artworks across time." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.907, + 0.198 + ], + "angle": 0, + "content": "Let \\( x^{(a)} \\) represent a data sample from data distribution \\( \\mathcal{A} \\). The goal is to identify the corresponding sample \\( x^{(b)} \\) from a different data distribution \\( \\mathcal{B} \\). This problem is inherently ill-posed due to the absence of ground-truth data pairs. To address this, we seek to establish optimal transport between the underlying manifolds and construct Schrödinger bridges between the respective distributions [9, 40]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.201, + 0.909, + 0.327 + ], + "angle": 0, + "content": "Dual Diffusion Implicit Bridges (DDIB) [83] demonstrate that such bridges can be effectively constructed using a pair of pre-trained image diffusion models. Notably, the Ordinary Differential Equations (ODEs) derived from trained diffusion models inherently exhibit optimal transport between the Gaussian noise prior and the corresponding data distributions [85]. Let ODESolve denote the mapping from an initial noise \\( x_{1} \\) to the data \\( x_0^{(a)} \\) with \\( v_{\\theta}^{(a)} \\) defined in Eq. (3):" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.345, + 0.907, + 0.381 + ], + "angle": 0, + "content": "\\[\n\\operatorname {O D E S o l v e} \\left(x _ {t}; v _ {\\theta} ^ {(a)}, 1, 0\\right) = x _ {1} + \\int_ {1} ^ {0} v _ {\\theta} ^ {(a)} \\left(x _ {t}, t\\right) d t. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.405, + 0.909, + 0.695 + ], + "angle": 0, + "content": "Analogously, we can train ODESolve \\((x_{t};v_{\\theta}^{(b)},1,0)\\) for mapping data between noise and \\(\\mathcal{B}\\). Given these frameworks, the learned mappings between noise and data distributions, along with their corresponding reverse processes ODESolve \\((x_{t};v_{\\theta},0,1)\\), inherently constitute optimal transport pairs. As illustrated in the example in Fig. 2, we demonstrate this ability to transfer data between distributions \\(\\mathcal{A}\\) and \\(\\mathcal{B}\\) despite having no prior information about their inherent relationship. First, we use the trained ODE to map data distribution \\(\\mathcal{A}\\) to the standard Gaussian noise. A clear structure related to \\(\\mathcal{A}\\) emerges, highlighting the optimal noise-to-data pairs achieved through the ODE model's transport. Subsequently, another trained ODE transports this structured noise [67] to the distinctly different distribution \\(\\mathcal{B}\\). In this setup, the mapping between distributions \\(\\mathcal{A}\\) and \\(\\mathcal{B}\\) remains largely preserved, even when using two independently trained ODEs that share no information. This indicates that the noise space might not be pure noise, but rather contains a certain structure as suggested in [67]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.698, + 0.908, + 0.729 + ], + "angle": 0, + "content": "The independent \\( v_{\\theta} \\) can be achieved by training the same model with different conditions, namely" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.748, + 0.907, + 0.768 + ], + "angle": 0, + "content": "\\[\nv _ {\\theta} ^ {(a)} = v _ {\\theta} (\\cdot , t, y _ {a}), v _ {\\theta} ^ {(b)} = v _ {\\theta} (\\cdot , t, y _ {b}). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.794, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Under our problem setup, the data \\( x_0^{(a)} \\) resembles a piece of artwork, and \\( y_{a} \\) represents the corresponding chronological information (creation year of the artwork). This parameterization allows a single model \\( v_{\\theta} \\) to learn mappings between arbitrary years \\( y_{a} \\rightarrow y_{b} \\) through conditioning, effectively replacing the need to train separate models for each temporal pair, as illustrated in Fig. 3." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5870" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.145, + 0.099, + 0.434, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.174, + 0.321, + 0.187 + ], + "angle": 0, + "content": "(a) Training" + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.192, + 0.434, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.253, + 0.267, + 0.323, + 0.279 + ], + "angle": 0, + "content": "(b) Inference" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.292, + 0.484, + 0.377 + ], + "angle": 0, + "content": "Figure 3. Approach overview. \\( a) \\). Training a single \\( v_{\\theta} \\) between style embeddings and structured noise conditioned on \\( y \\). \\( b) \\). During inference, we can guide the ODE forward and backward direction with different chronological conditions \\( y \\) between the style and latent noise space. By using different \\( y \\), the style embedding can be transformed to a different style manifold w.r.t. to historical patterns." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.402, + 0.484, + 0.434 + ], + "angle": 0, + "content": "Training objective. Similar to Eq. (3), we define our final objective to optimize \\( v_{\\theta} \\) with the year condition \\( y \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.443, + 0.484, + 0.479 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\theta) = \\int_ {0} ^ {T} \\mathbb {E} \\left[ \\| v _ {\\theta} \\left(x _ {t}, t, y\\right) - \\dot {\\alpha} _ {t} x _ {0} - \\dot {\\sigma} _ {t} x _ {1} \\| ^ {2} \\right] \\mathrm {d} t, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.488, + 0.486, + 0.549 + ], + "angle": 0, + "content": "where \\(\\theta\\) is the backbone parameterized by a 1D transformer. We detail the architecture of the model in the Appendix A. We operate in a compact 1D-style space, as pixel space retains excessively low-level details that contain redundancy." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.558, + 0.202, + 0.572 + ], + "angle": 0, + "content": "3.3. Inference" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.484, + 0.75 + ], + "angle": 0, + "content": "With a trained model that can generate the style latents sufficiently well, we can start the matching process. The model defines a vector field \\( v_{\\theta}(\\cdot ,t,y) \\) at any given \\( t \\) and \\( y \\). As explained in Algorithm 1, we first transform the style representation from \\( x_0^{(a)} \\) to the noisy representation \\( x_{t_{\\mathrm{th}}} \\) using the learned model \\( v_{\\theta} \\). The parameter \\( t_\\mathrm{th} > 0 \\) denotes the integration threshold; in practice, we typically set \\( t_\\mathrm{th} = 1 \\), corresponding to a full mapping from data space to the structured noise space. Next, we push forward the \\( x_{t_\\mathrm{th}} \\) back to \\( x_0^{(b)} \\) using the vector field produced by the same model but with a different chronological condition \\( v_{\\theta}(x_t,t,y_b) \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.751, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Fig. 4 provides a conceptual illustration of our method using a 2D synthetic toy example, where trajectories depict the temporal evolution of data points. We construct temporally coherent toy data pairs to enable access to ground-truth data pairs. A single model is trained to perform conditional mappings across five distinct input conditions (indicated by the x-axis). Although attributes (indicated by color) gradually blend or evolve, the underlying optimal transport plan is preserved. Notably, the generated samples consistently align with the correct target modes, despite the absence of" + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.101, + 0.852, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.261, + 0.908, + 0.331 + ], + "angle": 0, + "content": "Figure 4. Illustrational example with ground truth correspondences. Dots sharing the same color are permuted versions along the x-axis. In the figure, colored round dots represent the ground truth samples. purple stars (starting from the 2nd column) indicate the warped samples originating from cyan stars (leftmost)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.345, + 0.907, + 0.377 + ], + "angle": 0, + "content": "explicit mode supervision, demonstrating robust structural consistency across varying style dynamics." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.387, + 0.715, + 0.402 + ], + "angle": 0, + "content": "3.4. Dataset Construction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.907, + 0.56 + ], + "angle": 0, + "content": "The curated dataset is a large-scale dataset comprising over 650,000 artworks spanning a wide range of artistic styles and mediums. As discussed in Sec. 2.2, existing datasets for art lack unified meta-information, have limited access, or are small in size. In contrast, our dataset provides standardized and cleaned metadata aggregated from multiple sources. For example, while datasets like WikiArt [84] may include creation years within filenames, such information is often unstructured. Hence, we normalize these metadata to facilitate chronological modeling and future research." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.562, + 0.909, + 0.73 + ], + "angle": 0, + "content": "The available metadata includes title, artist, style, year of creation (Most of the artworks originate from years ranging from 1500 to 2021), media (e.g., oil on canvas, etching, acrylic), geographical location, and image URLs. The images are provided as URLs to public museums and public art gallery websites or indexed to public datasets, where they are under the original license and copyright. An overview of the dataset is provided in Fig. 1. For experimental purposes, we used a \\(6/2/2\\) split for training, validation, and testing, randomly sampled from centuries without selecting styles or artists explicitly to maintain the underlying data distribution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Deduplication. Due to the diverse sources of art pieces, duplicate samples are inevitable. While this redundancy provides additional meta-information from different sources, it results in variations of potential duplicates. To address this, we deduplicated our dataset following Somepalli et al. [79], using SSCD [59] embeddings. Entries with a cosine similarity above 0.85 were merged, consolidating their metadata, e.g., retaining the English version when multiple languages were available. This process reduced the dataset from approximately 950,000 to 650,000 art pieces, encompassing" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5871" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.48, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.297, + 0.484, + 0.355 + ], + "angle": 0, + "content": "Figure 5. Stylistic flows over time. We visualize distribution shifts for \\(\\Delta y = 0, 20, 40, 80\\) years. Blue quivers indicate the dominant movements of art pieces into the next time frame and visualize how style distribution flows across time." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.368, + 0.484, + 0.414 + ], + "angle": 0, + "content": "5,053 artists, 128 styles, 7,908 media, and 1,442 geographical locations. Due to the space limitation, we defer the more details of the dataset in the Appendix D." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.428, + 0.216, + 0.445 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.453, + 0.371, + 0.469 + ], + "angle": 0, + "content": "4.1. Choice of Style Representations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.475, + 0.484, + 0.626 + ], + "angle": 0, + "content": "The concept of style is inherently complex and multifaceted, encompassing aspects such as color, material, atmosphere, design, and structure. From a human perspective [39, 72], defining style is subjective and often varies depending on context. Similarly, in computer vision, style is represented through diverse latent spaces depending on the task. For instance, works on style classification and retrieval [11, 33] focus on different representations than those used in style transfer [20, 95] or in diffusion-based generation [16, 61, 87]. In both domains, style remains fluid and context-sensitive." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.484, + 0.747 + ], + "angle": 0, + "content": "To accommodate this ambiguity, our model is designed to be agnostic to the specific choice of style space. In our experiments, we primarily focus on two feature spaces: CLIP [62] and CSD [79]. As shown in Table S3, CSD excels in style-based retrieval, whereas CLIP performs slightly better in year-based retrieval and can be decoded to images with unCLIP [65] (see Figure S15). We refer readers to Appendix A for a detailed comparison of all candidates." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.294, + 0.774 + ], + "angle": 0, + "content": "4.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Tracing stylistic flows across time. We visualize the temporal flows of style distributions in Figure 5, showing how our model progressively aligns embeddings of early-20th-century artworks with those from later periods. Each subplot displays kernel density estimates (KDEs) [57] in the CSD space for three sets of samples: the original early-century inputs transported by our model (green), the real artworks from the corresponding future time window (orange), and" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.896, + 0.107 + ], + "angle": 0, + "content": "the fixed target distribution of late-20th-century art (blue)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.108, + 0.907, + 0.333 + ], + "angle": 0, + "content": "At \\(\\Delta t = 0\\), the transformed distribution (green) coincides with the original input, as expected. As \\(\\Delta t\\) increases to 20, 40, and 80 years, the green distribution increasingly aligns with the orange ground-truth distribution for each target decade, despite the absence of sample-level correspondences. This alignment is not exact, but meaningful: the transformed samples retain certain semantic traces of their origin, resulting in a distribution that blends temporal stylistic adaptation with content continuity. Such divergence offers insight into how artistic style may evolve from a given source, even in ways not fully captured by existing data. Eventually, the green distribution approaches the latecentury manifold (blue), indicating that our model learns a smooth and coherent temporal flow (denoted by the blue quivers) from a distribution perspective." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.334, + 0.906, + 0.44 + ], + "angle": 0, + "content": "To bridge population-level dynamics with sample-level behavior, we track a single artwork—Berlin Street Scene (1913)—through time. Its unCLIP renderings reflect plausible stylistic changes in line with historical trends, while preserving core semantics. This illustrates the model's ability to capture stylistic evolution across distributions and within individual samples, which we explore next." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.907, + 0.625 + ], + "angle": 0, + "content": "Transforming individual artworks through time. We present visual examples demonstrating how individual art pieces evolve across time by transferring their semantic embedding in the CLIP [62] space and visualizing using the un-CLIP model [66]. In Figure 6, we show the temporal trajectories of canonical works such as Les Demoiselles d'Avignon (1907, Cubism) and Landscape near Chatou (1904, Fauvism). This task is particularly challenging as these pieces are less realistic, making semantic capture more difficult. Despite this, we preserve artistic intent and content over time by operating in a compact semantic space." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.625, + 0.909, + 0.762 + ], + "angle": 0, + "content": "Additionally, we compare our approach with image editing methods in pixel space [49, 96] in Figure 7. These methods fail to achieve faithful semantic transformations tied to specific periods; instead, they apply superficial style transfers that ignore historical context. In contrast, our method enables long-range interpolation through time. We further analyze the style distributions of both unconditionally generated embeddings and mapped embeddings in Appendix C.2, showing that both closely follow the stylistic distributions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Content preservation. Our model captures how artistic style changes over time and how such stylistic evolution influences content, including shifts in scene composition and object substitution, e.g., horses versus motorcycles in Figure 7. Because style and content are often deeply entangled, meaningful temporal transformation requires flexibility. When the threshold parameter \\( t_{\\mathrm{th}} \\) is set too low (Section 3.3), the model is overly constrained to the original input, limit" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5872" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.135, + 0.089, + 0.859, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.304, + 0.908, + 0.334 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison of editing methods across different historical styles. Each row shows results from a different method. The 1st and 5th columns show original artworks (Les Demoiselles d'Avignon - 1907 and Landscape near Chatou - 1904)." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.346, + 0.784, + 0.429 + ], + "angle": 0, + "content": "
MethodRococoImpressionismRealism
R@1R@5δ↓τ(%)↑R@1R@5δ↓τ(%)↑R@1R@5δ↓τ(%)↑
RF-inversion [68]5.69.30.69083.710.526.30.65876.82.46.50.74966.3
LEDDS ++ [3]12.623.00.52879.415.430.40.58569.69.720.50.50964.9
SDEdit [49]6.315.80.57245.65.313.40.60534.53.712.40.61245.6
SDEdit (fine-tuned)37.942.80.26564.734.147.30.27247.227.538.40.23657.3
Ours72.585.10.02484.865.076.30.02475.858.569.30.02568.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.439, + 0.908, + 0.496 + ], + "angle": 0, + "content": "Table 1. Quality of the time matching when transferring artworks over time. We sampled a compact subset of a hundred samples for each style to calculate the metrics. We transfer clusters of artworks to randomly selected years within the range of \\([-100, 100]\\) and compute the mean metrics. For \\(\\tau\\), we used the nearest \\(nns[:25]\\) neighbors as positive samples and further \\(nns[50:75]\\) as the negative samples for each chosen anchor, where \\(nns\\) denotes its nearest neighbors. See extended table in Table S4." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.509, + 0.482, + 0.555 + ], + "angle": 0, + "content": "ing stylistic adaptation. This results in hybrid outputs, such as a Steam Horse locomotive, that blend past and future characteristics without fully transitioning." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.556, + 0.483, + 0.707 + ], + "angle": 0, + "content": "Content deviation is a broader challenge across models. Most methods face a trade-off between low-level fidelity and high-level semantic consistency. Pixel-focused techniques [3, 49, 67] often struggle to produce meaningful temporal shifts even with adjusted guidance scales, sometimes leading to failures like buildings morphing into people, as seen in Figure 6. In contrast, our model preserves conceptual content while enabling coherent, temporally grounded transformations, striking a balance between semantic stability and stylistic progression." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.734, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Quantitative evaluation of time travel. Assessing the quality of temporal transitions in artworks is challenging due to the absence of ground-truth correspondences. However, semi-ground-truth evidence exists: artworks created close in time and close in style are more likely to evolve together rather than disperse. Based on this intuition, we utilized three key metrics:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.84, + 0.484, + 0.886 + ], + "angle": 0, + "content": "- Recall@k assesses alignment with the target era by checking whether time-traveled samples retrieve temporally appropriate neighbors in style space." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.484, + 0.901 + ], + "angle": 0, + "content": "- Compactness \\(\\delta\\) measures how coherent the transformed" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.509, + 0.905, + 0.577 + ], + "angle": 0, + "content": "samples are by comparing their variance to that of random samples in the target year: \\(\\frac{\\sigma_{\\text{samples}}^2}{\\sigma_{\\text{target}}^2}\\). A good transformation should preserve this coherence, as source samples are initially similar in style and time, resulting in lower \\(\\delta\\)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.577, + 0.908, + 0.651 + ], + "angle": 0, + "content": "- Triplet consistency \\(\\tau\\) is adapted from triplet loss [23, 73]. It evaluates whether the relative relation among anchor, positive, and negative samples is preserved posttransformation, reflecting the model's ability to maintain local structure." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.656, + 0.909, + 0.822 + ], + "angle": 0, + "content": "We evaluate on subsets of 100 samples by selecting localized neighborhoods of artwork within several style labels from the test split. These subsets serve as the starting samples for time mapping. Due to space constraints, we present results for the few most prominent artistic styles in Table 1, though our analysis extends across a much broader range. Since most editing baselines operate directly in pixel space, we perform an additional step of encoding their generated outputs back into the style space to enable a fair comparison, which allows us to compute nearest neighbor recall and compactness in a shared representation space." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "As shown in Table 1 (we offer more analysis in Table S4 in the Appendix), conventional pixel-based image-editing methods [3, 49, 67] fail to achieve both objectives: they neither align generated samples with the appropriate data manifold, reflected by the lower Recall@ \\( k \\), nor preserve" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5873" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.127, + 0.098, + 0.446, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.295, + 0.486, + 0.366 + ], + "angle": 0, + "content": "Figure 7. Flow preserves content semantics across time. We transform a motorcyclist to the year 1800 with increasing \\( t_{\\mathrm{th}} \\) (left to right, top row). Low \\( t_{\\mathrm{th}} \\) leads to incorrect mappings due to limited editing flexibility, while higher values retain semantic identity (\"man on a horse\") with correct stylistic traits." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.38, + 0.483, + 0.501 + ], + "angle": 0, + "content": "local structural coherence, indicated by the higher \\(\\delta\\)). This is mainly due to their reliance on pixel-level cues, which are insufficient for modeling more profound, semantic style shifts. For instance, RF-Inversion achieves high fidelity editing [67] but leads to minimal pixel shifts, evidenced by the low recall. We address these issues by operating in a compact latent space and focusing on semantic information to better align with historical stylistic flows." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.502, + 0.486, + 0.743 + ], + "angle": 0, + "content": "Since our method is trained exclusively on the curated dataset, a distributional mismatch may arise when comparing against baseline models such as Stable Diffusion, which are pretrained on broader, general-domain datasets. For a fair comparison, we also fine-tuned the base Stable Diffusion model on our dataset until convergence. Specifically, we adapt the SD 1.5 model using chronologically conditioned prompts of the form \"A painting in {year} style\". This finetuning procedure improves both the synthesis FID in Table 2 and retrieval-based matching performance in Table 1, confirming our dataset's value. However, applying SDEdit remains suboptimal for our task, as the method struggles to perform temporally faithful transformations. This performance gap persists across all evaluated artistic styles, where our method consistently yields more coherent and historically aligned style transitions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Comparison to synthesis baselines. As previously discussed, for visualization purposes, we project samples into the CLIP [62] embedding space and render them using the unCLIP model [66]. This enables quantitative evaluation using the standard Frechet Inception Distance (FID) [25], which assesses both the realism of the generated outputs and their fidelity to the target temporal style distribution." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We compared our method against three Latent Diffusion Models (LDM) variants: Stable Diffusion 1.5 [66], Stable" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.09, + 0.905, + 0.183 + ], + "angle": 0, + "content": "
ModelFID↓conditioned on century
16th17th18th19th20th
SD 1.5 [66]131.1▼40.4118.9▼37.1103.5▼17.2149.9▼77.9180.4▼101.5
SD 2.1214.4▼123.7147.1▼65.3134.2▼47.9170.2▼98.2167.1▼88.2
SDXL [60]186.6▼95.9142.4▼60.6143.4▼57.1114.8▼42.8146.1▼67.2
FLUX.1-schnell [38]199.2▼108.5142.2▼60.4137.3▼51.0122.8▼50.8126.1▼47.2
SD 1.5 (fine-tuned)111.7▼2196.4▼14.688.8▼2.596.4▼24.4150.7▼71.8
Ours90.781.886.372.078.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.192, + 0.908, + 0.263 + ], + "angle": 0, + "content": "Table 2. Feature synthesis across centuries. Evaluation metrics are computed on 1,000 samples per century. While Stable Diffusion and FLUX rely on text prompts for year conditioning, our model conditions directly on this information. Our method exhibits stronger alignment with ground truth distributions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.272, + 0.909, + 0.577 + ], + "angle": 0, + "content": "Diffusion 2.1 [66], and SDXL [60], as well as the flow-matching based FLUX model [38]. We also compare against the fine-tuned Stable Diffusion 1.5, with results summarized in Table 2. SDXL [60] has demonstrated strong artistic understanding and has become the foundation for many recent stylistic image generation methods [16, 24, 76, 87]. While SDXL performs well on certain styles, such as Renaissance (16th century) and Cubism (20th century) in the Appendix Figure S13, it struggles to capture stylistic distributions comprehensively. FLUX [38], despite its training scale and exhibiting strong performance across text-to-image tasks, also struggles to capture temporal style distributions effectively (see Figure S14). In contrast, our model, trained on the proposed dataset, better captures the diversity of artistic styles. Our approach achieves significantly lower FID scores than standard text-conditioned models, demonstrating better alignment with the underlying feature distribution of different art periods. Furthermore, it provides more consistent stylistic alignment across centuries. Further rendering results for all models are available in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.589, + 0.634, + 0.605 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.908, + 0.735 + ], + "angle": 0, + "content": "Studying the evolution of art is challenging due to the lack of ground-truth correspondences between artworks from different periods. We present a framework that models stylistic progression using stochastic interpolants and dual diffusion implicit bridges, enabling distribution-level alignment without paired data. By formulating art evolution as optimal transport in a learned style space, our method captures continuous temporal transitions in artistic style." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.901 + ], + "angle": 0, + "content": "A central contribution is a dataset of over 650,000 artworks spanning 500 years, which we use to demonstrate that our model captures long-range stylistic shifts. Beyond modeling distributions, our method also enables tracing the stylistic origins and plausible futures of individual artworks, allowing for fine-grained longitudinal analysis. Our work explores a less-studied direction in computer vision: using generative models to investigate the evolution of art and support cultural and historical analysis. We believe our approach offers a foundation for future interdisciplinary work at the intersection of machine learning and cultural analytics." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5874" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.268 + ], + "angle": 0, + "content": "This project was supported by Bayer AG, the Federal Ministry for Economic Affairs and Energy within the project \"NXT GEN AI METHODS - Generative Methoden für Perzeption, Prädiktion und Planung\", the project \"Genius-Robot\" (01IS24083) funded by the Federal Ministry of Research, Technology and Space (BMFTR), and the bidt project KLIMA-MEMES. The authors gratefully acknowledge the Gauss Center for Supercomputing for providing compute through the NIC on JUWELS/JUPITER at JSC and the HPC resources supplied by the NHR @FAU Erlangen." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.281, + 0.188, + 0.297 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.305, + 0.484, + 0.347 + ], + "angle": 0, + "content": "[1] Michael S Albergo, Nicholas M Boffi, and Eric Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions. arXiv, 2023. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.349, + 0.484, + 0.404 + ], + "angle": 0, + "content": "[2] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.406, + 0.484, + 0.488 + ], + "angle": 0, + "content": "[3] Manuel Brack, Felix Friedrich, Katharia Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. Ledits++: Limitless image editing using text-to-image models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8861-8870, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.49, + 0.485, + 0.558 + ], + "angle": 0, + "content": "[4] Levin Brinkmann, Deniz Gezerli, KV Kleist, Thomas F Müller, Iyad Rahwan, and Niccolò Pescetelli. Hybrid social learning in human-algorithm cultural transmission. Philosophical Transactions of the Royal Society A, 380(2227), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.561, + 0.484, + 0.615 + ], + "angle": 0, + "content": "[5] Eric Ming Chen, Jin Sun, Apoorv Khandelwal, Dani Lischinski, Noah Snavely, and Hadar Averbuch-Elor. What's in a decade? transforming faces through time. In Computer Graphics Forum, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.617, + 0.484, + 0.686 + ], + "angle": 0, + "content": "[6] Haibo Chen, Lei Zhao, Zhizhong Wang, Huiming Zhang, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Artistic style transfer with internal-external learning and contrastive learning. Advances in Neural Information Processing Systems (NeurIPS), 34, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.688, + 0.484, + 0.716 + ], + "angle": 0, + "content": "[7] Ricky TQ Chen and Yaron Lipman. Riemannian flow matching on general geometries. arXiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.718, + 0.484, + 0.758 + ], + "angle": 0, + "content": "[8] Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. NeurIPS, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.76, + 0.484, + 0.815 + ], + "angle": 0, + "content": "[9] Yongxin Chen, Tryphon T Georgiou, and Michele Pavon. On the relation between optimal transport and schrödinger bridges: A stochastic control viewpoint. Journal of Optimization Theory and Applications, 169:671-691, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.817, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[10] Ceren Comert, Murat Özbayoglu, and Cosku Kasnakoğlu. Painter prediction from artworks with transfer learning. In 2021 7th International Conference on Mechatronics and Robotics Engineering (ICMRE). IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[11] Elliot Crowley and Andrew Zisserman. The state of the art: Object retrieval in paintings using discriminative regions. In" + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.305, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.908, + 0.12 + ], + "angle": 0, + "content": "Proceedings of the British Machine Vision Conference 2014. British Machine Vision Association, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.122, + 0.908, + 0.16 + ], + "angle": 0, + "content": "[12] Aram Davtyan, Sepehr Sameni, and Paolo Favaro. Randomized conditional flow matching for video prediction. arXiv, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.164, + 0.907, + 0.218 + ], + "angle": 0, + "content": "[13] Valentin De Bortoli, James Thornton, Jeremy Heng, and Arnaud Doucet. Diffusion schrödinger bridge with applications to score-based generative modeling. Advances in Neural Information Processing Systems, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.219, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.261, + 0.907, + 0.314 + ], + "angle": 0, + "content": "[15] Yingying Deng, Fan Tang, Weiming Dong, Chongyang Ma, Feiyue Huang, Oliver Deussen, and Changsheng Xu. Exploring the representativity of art paintings. IEEE Transactions on Multimedia, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.317, + 0.907, + 0.357 + ], + "angle": 0, + "content": "[16] Yarden Frenkel, Yael Vinker, Ariel Shamir, and Daniel Cohen-Or. Implicit style-content separation using b-lora. In ECCV, 2024. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.359, + 0.907, + 0.412 + ], + "angle": 0, + "content": "[17] Michael Fuest, Pingchuan Ma, Ming Gui, Johannes S Fischer, Vincent Tao Hu, and Bjorn Ommer. Diffusion models and representation learning: A survey. arXiv preprint arXiv:2407.00783, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.414, + 0.907, + 0.468 + ], + "angle": 0, + "content": "[18] Shangqi Gao, Hangqi Zhou, Yibo Gao, and Xiahai Zhuang. Bayeseg: Bayesian modeling for medical image segmentation with interpretable generalizability. Medical Image Analysis, 89, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.47, + 0.907, + 0.524 + ], + "angle": 0, + "content": "[19] Noa Garcia, Benjamin Renoust, and Yuta Nakashima. Contextnet: representation and exploration for painting classification and retrieval in context. International Journal of Multimedia Information Retrieval, 9(1), 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.526, + 0.907, + 0.579 + ], + "angle": 0, + "content": "[20] Leon Gatys, Alexander S Ecker, and Matthias Bethge. Texture synthesis using convolutional neural networks. In Advances in Neural Information Processing Systems (NeurIPS), 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.581, + 0.907, + 0.636 + ], + "angle": 0, + "content": "[21] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.637, + 0.907, + 0.705 + ], + "angle": 0, + "content": "[22] Ming Gui, Johannes Schusterbauer, Ulrich Prestel, Pingchuan Ma, Dmytro Kotovenko, Olga Grebenkova, Stefan Andreas Baumann, Vincent Tao Hu, and Björn Ommer. Depthfm: Fast monocular depth estimation with flow matching. AAAI, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.707, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[23] Alexander Hermans, Lucas Beyer, and Bastian Leibe. In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.749, + 0.907, + 0.788 + ], + "angle": 0, + "content": "[24] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. In CVPR, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.79, + 0.907, + 0.857 + ], + "angle": 0, + "content": "[25] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.86, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[26] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems (NeurIPS), 33, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5875" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.092, + 0.486, + 0.146 + ], + "angle": 0, + "content": "[27] Vincent Tao Hu, Yunlu Chen, Mathilde Caron, Yuki M. Asano, Cees G.M. Snoek, and Björn Ommer. Guided diffusion from self-supervised diffusion features. In ARXIV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.149, + 0.485, + 0.218 + ], + "angle": 0, + "content": "[28] Vincent Tao Hu, Wenzhe Yin, Pingchuan Ma, Yunlu Chen, Basura Fernando, Yuki M. Asano, Efstratos Gavves, Pascal Mettes, Björn Ommer, and Cees G.M. Snoek. Motion flow matching for human motion synthesis and editing. In ARXIV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.221, + 0.484, + 0.289 + ], + "angle": 0, + "content": "[29] Vincent Tao Hu, Stefan Andreas Baumann, Ming Gui, Olga Grebenkova, Pingchuan Ma, Johannes Fischer, and Björn Ommer. *Zigma: A dit-style zigzag mamba diffusion model*. In European conference on computer vision, pages 148-166. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.291, + 0.484, + 0.346 + ], + "angle": 0, + "content": "[30] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.348, + 0.484, + 0.401 + ], + "angle": 0, + "content": "[31] Herve Jegou, Matthijs Douze, and Cordelia Schmid. Product quantization for nearest neighbor search. IEEE transactions on pattern analysis and machine intelligence, 33(1):117-128, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.404, + 0.484, + 0.459 + ], + "angle": 0, + "content": "[32] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Proceedings of the European Conference on Computer Vision (ECCV). Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.461, + 0.484, + 0.515 + ], + "angle": 0, + "content": "[33] Sergey Karayev, Matthew Trentacoste, Helen Han, Aseem Agarwala, Trevor Darrell, Aaron Hertzmann, and Holger Winnemoeller. Recognizing image style. arXiv preprint arXiv:1311.3715, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.518, + 0.484, + 0.558 + ], + "angle": 0, + "content": "[34] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.561, + 0.484, + 0.602 + ], + "angle": 0, + "content": "[35] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.604, + 0.484, + 0.672 + ], + "angle": 0, + "content": "[36] Dmytro Kotovenko, Matthias Wright, Arthur Heimbrecht, and Bjorn Ommer. Rethinking style transfer: From pixels to parameterized brushstrokes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.674, + 0.484, + 0.729 + ], + "angle": 0, + "content": "[37] Felix Krause, Timy Phan, Ming Gui, Stefan Andreas Baumann, Vincent Tao Hu, and Björn Ommer. Tread: Token routing for efficient architecture-agnostic diffusion training. arXiv preprint arXiv:2501.04765, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.731, + 0.484, + 0.757 + ], + "angle": 0, + "content": "[38] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024.8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.76, + 0.484, + 0.786 + ], + "angle": 0, + "content": "[39] Berel Lang. The concept of style. Cornell University Press, 1987. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.789, + 0.484, + 0.83 + ], + "angle": 0, + "content": "[40] Christian Léonard. A survey of the schr\\'' odinger problem and some of its connections with optimal transport. arXiv preprint arXiv:1308.0215, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[41] Wen Li, Muyuan Fang, Cheng Zou, Biao Gong, Ruobing Zheng, Meng Wang, Jingdong Chen, and Ming Yang. Style-tokensizer: Defining image style by a single instance for controlling diffusion models. In European Conference on Computer Vision, pages 110-126. Springer, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.908, + 0.16 + ], + "angle": 0, + "content": "[42] Xueting Li, Sifei Liu, Jan Kautz, and Ming-Hsuan Yang. Learning linear transformations for fast image and video style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.164, + 0.907, + 0.217 + ], + "angle": 0, + "content": "[43] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.221, + 0.907, + 0.261 + ], + "angle": 0, + "content": "[44] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. *ICLR*, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.264, + 0.907, + 0.303 + ], + "angle": 0, + "content": "[45] Guan-Horng Liu, Tianrong Chen, Oswin So, and Evangelos Theodorou. Deep generalized schrödinger bridge. NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.307, + 0.906, + 0.347 + ], + "angle": 0, + "content": "[46] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. *ICLR*, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.35, + 0.906, + 0.404 + ], + "angle": 0, + "content": "[47] Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. arXiv, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.407, + 0.907, + 0.447 + ], + "angle": 0, + "content": "[48] L. McInnes, J. Healy, and J. Melville. UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction. ArXiv, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.45, + 0.907, + 0.504 + ], + "angle": 0, + "content": "[49] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 6, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.507, + 0.907, + 0.56 + ], + "angle": 0, + "content": "[50] Amir Merchant, Simon Batzner, Samuel S Schoenholz, Muratahan Aykol, Gowoon Cheon, and Ekin Dogus Cubuk. Scaling deep learning for materials discovery. Nature, 624 (7990), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.564, + 0.907, + 0.632 + ], + "angle": 0, + "content": "[51] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.634, + 0.907, + 0.702 + ], + "angle": 0, + "content": "[52] Kevin Musgrave, Serge Belongie, and Ser-Nam Lim. A metric learning reality check. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16, pages 681-699. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.705, + 0.907, + 0.758 + ], + "angle": 0, + "content": "[53] Ivona Najdenkoska, Xiantong Zhen, Marcel Worring, and Ling Shao. Uncertainty-aware report generation for chest x-rays by variational topic inference. Medical Image Analysis, 82, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.762, + 0.907, + 0.802 + ], + "angle": 0, + "content": "[54] Kirill Neklyudov, Daniel Severo, and Alireza Makhzani. Action matching: A variational method for learning stochastic dynamics from samples. arXiv, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.804, + 0.908, + 0.9 + ], + "angle": 0, + "content": "[55] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bo" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.093, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5876" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.131, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "janowski. Dinov2: Learning robust visual features without supervision, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.122, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[56] Norihiro Oyama, Noriko N Ishizaki, Satoshi Koide, and Hiroaki Yoshida. Deep generative model super-resolves spatially correlated multiregional climate data. Scientific Reports, 13(1), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.179, + 0.483, + 0.22 + ], + "angle": 0, + "content": "[57] Emanuel Parzen. On estimation of a probability density function and mode. The annals of mathematical statistics, 33(3):1065-1076, 1962. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.222, + 0.484, + 0.291 + ], + "angle": 0, + "content": "[58] Pablo Pernias, Dominic Rampas, Mats Leon Richter, Christopher Pal, and Marc Aubreville. Würstchen: An efficient architecture for large-scale text-to-image diffusion models. In The Twelfth International Conference on Learning Representations, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.293, + 0.483, + 0.333 + ], + "angle": 0, + "content": "[59] Ed Pizzi, Sreya Dutta Roy, Sugosh Nagavara Ravindra, Priya Goyal, and Matthijs Douze. A self-supervised descriptor for image copy detection. In CVPR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.336, + 0.482, + 0.391 + ], + "angle": 0, + "content": "[60] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.393, + 0.483, + 0.46 + ], + "angle": 0, + "content": "[61] Tianhao Qi, Shancheng Fang, Yanze Wu, Hongtao Xie, Jiawei Liu, Lang Chen, Qian He, and Yongdong Zhang. Deadiff: An efficient stylization diffusion model with disentangled representations. arXiv preprint arXiv:2403.06951, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.463, + 0.483, + 0.545 + ], + "angle": 0, + "content": "[62] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning. PMLR, 2021. 2, 6, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.548, + 0.483, + 0.63 + ], + "angle": 0, + "content": "[63] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML). PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.633, + 0.483, + 0.673 + ], + "angle": 0, + "content": "[64] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.676, + 0.483, + 0.73 + ], + "angle": 0, + "content": "[65] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 6, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.733, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[66] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022. 1, 3, 6, 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.803, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[67] L Rout, Y Chen, N Ruiz, C Caramanis, S Shakkottai, and W Chu. Semantic image inversion and editing using rectified stochastic differential equations. 2024. 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[68] Litu Rout, Yujia Chen, Nataniel Ruiz, Constantine Caramanis, Sanjay Shakkottai, and Wen-Sheng Chu. Semantic image inversion and editing using rectified stochastic differential equations. In ICLR, 2025. 7" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.092, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.093, + 0.907, + 0.162 + ], + "angle": 0, + "content": "[69] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.164, + 0.907, + 0.219 + ], + "angle": 0, + "content": "[70] Dan Ruta, Saeid Motiian, Baldo Faieta, Zhe Lin, Hailin Jin, Alex Filipkowski, Andrew Gilbert, and John Collomosse. Aladin: All layer adaptive instance normalization for fine-grained style similarity. In ICCV, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.221, + 0.907, + 0.261 + ], + "angle": 0, + "content": "[71] Babak Saleh and Ahmed Elgammal. Large-scale classification of fine-art paintings: Learning the right metric on the right feature. arXiv preprint arXiv:1505.00855, 2015. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.264, + 0.907, + 0.29 + ], + "angle": 0, + "content": "[72] Meyer Schapiro. Theory and philosophy of art: Style, artist, and society. 1994. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.292, + 0.907, + 0.347 + ], + "angle": 0, + "content": "[73] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.349, + 0.907, + 0.417 + ], + "angle": 0, + "content": "[74] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 3, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.42, + 0.906, + 0.46 + ], + "angle": 0, + "content": "[75] Johannes Schusterbauer, Ming Gui, Pingchuan Ma, Nick Stracke, Stefan A Baumann, and Björn Ommer. Boosting latent diffusion with flow matching. ECCV, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.462, + 0.907, + 0.515 + ], + "angle": 0, + "content": "[76] Viraj Shah, Nataniel Ruiz, Forrester Cole, Erika Lu, Svetlana Lazebnik, Yuanzhen Li, and Varun Jampani. Ziplora: Any subject in any style by effectively merging loras. In ECCV, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.519, + 0.907, + 0.572 + ], + "angle": 0, + "content": "[77] Higor YD Sigaki, Matjaž Perc, and Haroldo V Ribeiro. History of art paintings through the lens of entropy and complexity. Proceedings of the National Academy of Sciences, 115(37):E8585-E8594, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.576, + 0.906, + 0.616 + ], + "angle": 0, + "content": "[78] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.618, + 0.907, + 0.672 + ], + "angle": 0, + "content": "[79] Gowthami Somepalli, Anubhav Gupta, Kamal Gupta, Shramay Palta, Micah Goldblum, Jonas Geiping, Abhinav Shrivastava, and Tom Goldstein. Measuring style similarity in diffusion models. ECCV, 2024. 2, 3, 5, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.675, + 0.907, + 0.701 + ], + "angle": 0, + "content": "[80] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.704, + 0.907, + 0.758 + ], + "angle": 0, + "content": "[81] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.761, + 0.906, + 0.801 + ], + "angle": 0, + "content": "[82] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.804, + 0.906, + 0.843 + ], + "angle": 0, + "content": "[83] Xuan Su, Jiaming Song, Chenlin Meng, and Stefano Ermon. Dual diffusion implicit bridges for image-to-image translation. *ICLR*, 2023. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.846, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[84] Wei Ren Tan, Chee Seng Chan, Hernan Aguirre, and Kiyoshi Tanaka. Improved artgan for conditional synthesis of natural image and artwork. IEEE Transactions on Image Processing, 28(1):394-409, 2019. 5" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5877" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[85] Zhicong Tang, Tiankai Hang, Shuyang Gu, Dong Chen, and Baining Guo. Simplified diffusion schrödinger bridge. arXiv preprint arXiv:2403.14623, 2024. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.134, + 0.484, + 0.176 + ], + "angle": 0, + "content": "[86] Gemma Canet Tarrés, Dan Ruta, Tu Bui, and John Collomosse. Parasol: Parametric style control for diffusion image synthesis. arXiv preprint arXiv:2303.06464, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.177, + 0.483, + 0.232 + ], + "angle": 0, + "content": "[87] Haofan Wang, Matteo Spinelli, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv preprint arXiv:2404.02733, 2024. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.233, + 0.483, + 0.273 + ], + "angle": 0, + "content": "[88] Sheng-Yu Wang, Alexei A Efros, Jun-Yan Zhu, and Richard Zhang. Evaluating data attribution for text-to-image models. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.274, + 0.483, + 0.328 + ], + "angle": 0, + "content": "[89] Zhizhong Wang, Lei Zhao, Sihuan Lin, Qihang Mo, Huiming Zhang, Wei Xing, and Dongming Lu. Glstylenet: exquisite style transfer combining global and local pyramid features. IET Computer Vision, 14(8), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.484, + 0.398 + ], + "angle": 0, + "content": "[90] Zhizhong Wang, Lei Zhao, Haibo Chen, Ailin Li, Zhiwen Zuo, Wei Xing, and Dongming Lu. Texture reformer: Towards fast and universal interactive texture transfer. In Proceedings of the AAAI Conference on Artificial Intelligence, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.4, + 0.483, + 0.468 + ], + "angle": 0, + "content": "[91] Zhizhong Wang, Lei Zhao, Haibo Chen, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. DivSwapper: Towards diversified patch-based arbitrary style transfer. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence (IJCAI), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.469, + 0.483, + 0.537 + ], + "angle": 0, + "content": "[92] Zhizhong Wang, Lei Zhao, Zhiwen Zuo, Ailin Li, Haibo Chen, Wei Xing, and Dongming Lu. Microast: Towards super-fast ultra-resolution arbitrary style transfer. In Proceedings of the AAAI Conference on Artificial Intelligence, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.539, + 0.483, + 0.594 + ], + "angle": 0, + "content": "[93] Michael J Wilber, Chen Fang, Hailin Jin, Aaron Hertzmann, John Collomosse, and Serge Belongie. Bam! the behave artistic media dataset for recognition beyond photography. In ICCV, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.595, + 0.483, + 0.65 + ], + "angle": 0, + "content": "[94] Lemeng Wu, Dilin Wang, Chengyue Gong, Xingchao Liu, Yunyang Xiong, Rakesh Ranjan, Raghuraman Krishnamoorthi, Vikas Chandra, and Qiang Liu. Fast point cloud generation with straight flows. arXiv, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.651, + 0.483, + 0.705 + ], + "angle": 0, + "content": "[95] Zijie Wu, Zhen Zhu, Junping Du, and Xiang Bai. Ccpl: Contrastive coherence preserving loss for versatile style transfer. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.707, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[96] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.748, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[97] Sheng Zhang, Yanbo Xu, Naoto Usuyama, Hanwen Xu, Jaspreet Bagga, Robert Tinn, Sam Preston, Rajesh Rao, Mu Wei, Naveen Valluri, et al. Biomedclip: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs. arXiv preprint arXiv:2303.00915, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[98] Yuxin Zhang, Fan Tang, Weiming Dong, Haibin Huang, Chongyang Ma, Tong-Yee Lee, and Changsheng Xu. Domain enhanced arbitrary image style transfer via contrastive learning. In ACM SIGGRAPH 2022 Conference Proceedings, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[99] Wentao Zhao, Dalin Zhou, Xinguo Qiu, and Wei Jiang. How to represent paintings: A painting classification using artistic comments. Sensors, 21(6), 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.907, + 0.163 + ], + "angle": 0, + "content": "[100] Linqi Zhou, Aaron Lou, Samar Khanna, and Stefano Ermon. Denoising diffusion bridge models. *ICLR*, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.907, + 0.233 + ], + "angle": 0, + "content": "[101] Zhiwen Zuo, Lei Zhao, Shuobin Lian, Haibo Chen, Zhizhong Wang, Ailin Li, Wei Xing, and Dongming Lu. Style fader generative adversarial networks for style degree controllable artistic style transfer. In Proc. Int. Joint Conf. on Artif. Intell.(IJCAI), 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5878" + } + ] +] \ No newline at end of file diff --git a/2025/Stochastic Interpolants for Revealing Stylistic Flows across the History of Art/16c23549-7e7f-4e3a-8944-314b6bb00228_origin.pdf b/2025/Stochastic Interpolants for Revealing Stylistic Flows across the History of Art/16c23549-7e7f-4e3a-8944-314b6bb00228_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f5af10928bec6a3f6a2b3a5cefc6f2172ca946ad --- /dev/null +++ b/2025/Stochastic Interpolants for Revealing Stylistic Flows across the History of Art/16c23549-7e7f-4e3a-8944-314b6bb00228_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:651cbf8551478996e0ae9f4c85dd4f26240c9ab2bcb5d735abc2f12c71a53590 +size 18006536 diff --git a/2025/Stochastic Interpolants for Revealing Stylistic Flows across the History of Art/full.md b/2025/Stochastic Interpolants for Revealing Stylistic Flows across the History of Art/full.md new file mode 100644 index 0000000000000000000000000000000000000000..528aadfe9b84ca026d21ad41f6e12fd4310be451 --- /dev/null +++ b/2025/Stochastic Interpolants for Revealing Stylistic Flows across the History of Art/full.md @@ -0,0 +1,342 @@ +# Stochastic Interpolants for Revealing Stylistic Flows across the History of Art + +Pingchuan Ma* Ming Gui* Johannes Schusterbauer +Xiaopei Yang Olga Grebenkova Vincent Tao Hu Björn Ommer + +CompVis @ LMU Munich Munich Center for Machine Learning (MCML) + +# Abstract + +Generative models have made rapid progress in content creation, particularly in synthesizing artworks and capturing stylistic variation. However, most methods operate at the level of individual images, limiting their ability to reveal broader stylistic trends and temporal transitions. We address this by introducing a framework that models stylistic evolution as an optimal transport problem in a learned style space, using stochastic interpolants and dual diffusion implicit bridges to align artistic distributions across time without requiring paired data. A central contribution is a diverse dataset of over 650,000 artworks spanning 500 years, curated with metadata across multiple genres. Together, our method and dataset enable tracing long-range stylistic transitions and plausible futures of individual artworks, supporting fine-grained temporal analysis. This offers a new tool for modeling historical patterns in visual culture and opens up promising directions in visual understanding. Code and dataset: https://github.com/CompVis/Art-fm + +# 1. Introduction + +Recent advances in computer vision have sparked growing interest in applications in interdisciplinary research domains. Such applications not only try to offer novel perspectives, but also address challenges that were impossible before, such as those seen in medical imaging [18, 53], novel detection in material science [50], and simulation of climate [56]. Despite these successes, there remains a noticeable gap in their use for cultural studies [4], particularly in art history. + +With the rise of generative models such as Stable Diffusion [66], DALL-E [64], and others [51, 58, 69], there has been increasing use of these tools for artwork synthesis and content creation. However, their primary focus remains on image generation and editing, rather than analyzing the broader narrative of art history [10, 19, 99]. As a result, current models are not designed to capture the complex evo + +lution of artistic movements and styles over time. + +In this paper, we ask: How would a piece of art evolve? What might it look like if created centuries earlier or later? Naturally, no ground-truth pairs exist for such a task, and the artistic style of a creator is inherently dynamic and historically constrained. To address this challenge, we aim to analyze the flow of artistic styles over time. Given a single artwork, we try to determine its closest "projection" onto the manifold of artistic expressions corresponding to a given historical period. While optimal transport provides a natural formulation for such a task, considering only a single isolated artwork is insufficient. Instead, we reformulate the problem as a distribution matching task, where we study the evolution of the underlying distribution of art over time. + +Recent advances in diffusion bridges such as DDIB [83], DSB [13, 85], and DDBMs [100] have shown that domain translation can be performed implicitly, without the need for paired supervision. We build on this idea in the context of art history by incorporating stochastic interpolants [1], enabling us to model the temporal evolution of artistic styles without requiring explicitly aligned training data. Unlike traditional style transfer methods that transform images based on predefined and discrete styles, our approach projects styles onto their corresponding distributions across time. Operating in a compact latent space allows us to abstract away pixel-level distraction and focus on meaningful stylistic shifts. + +To support this temporal modeling, we require a dataset that spans centuries and reflects a wide variety of artistic styles. While datasets like WikiArt [71] offer breadth in genre and metadata, they lack the scale and historical depth needed to model long-term stylistic progression. Others, such as BAM [93] and BAM-FG [70], are either restricted in access or skewed toward modern art. To address this, we curate a new unified dataset of over 650,000 artworks annotated with creation year, location, artist, medium, and other metadata (Fig. 1). This dataset enables the study of artistic evolution at scale and over time. In summary, our key contributions are as follows: + +- We propose the first generative framework that explicitly models the temporal progression of artistic styles across + +![](images/23df6f6359b304d32dce1aa8383cb1aebd9ab9e8467ea9ad03ba6e3d2502e6f6.jpg) +Figure 1. Overview of the proposed dataset. We have curated and unified our large-scale dataset, including 650,000 artworks from different creation periods (ranging from 1500 to 2021) and styles with other meta-information, e.g., year, title, artist, media, etc. + +historical periods, introduced in Sec. 3.2 and Sec. 3.3. + +- To support this, we construct a large-scale, unified dataset of 650,000 artworks annotated with creation year and other metadata, covering five centuries of diverse artistic styles (Sec. 3.4). An overview is shown in Fig. 1. +- Our method models temporal transitions without requiring paired data. We outperform competitive baselines in capturing long-range stylistic shifts (Sec. 4.2). + +# 2. Related Works + +# 2.1. Stylistic representation + +The pioneering work by Gatys et al. [21] marked the beginning of the style transfer era, which defines style as Gram Matrices of pre-trained network features. Various methods have been proposed to improve different aspects of style representation, ranging from efficiency-focused approaches like [32, 42, 92] to quality-centric ones such as [36, 89, 98, 101]. Moreover, efforts have been made to enhance the generality and diversity of style transfer methods [30, 90, 91, 95]. Despite these achievements, most existing approaches [6, 41, 61, 86, 98] focus on image editing or the disentanglement of style and content through explicit style/content representations or user-defined prompts. Data-driven methods as [15] defined the representativity of art + +works as a deep feature fusion incorporating not only images, but also painting styles, creation time, and authorship information. On the other hand, ContextNet [19] aims to represent and explore artistic attributes in different modalities and their relationships to improve the retrieval of artworks in automatic art analysis. ArtGCN [99] uses a graph CNN to construct stylistic representation based on artistic comments. More recent works, such as GDA [88] and CSD [79], train style descriptors by fine-tuning on top of the existing feature space of CLIP [62] through contrastive learning with synthetic or curated data. + +These studies typically concentrate on rendering or differentiating individual images with specific styles and often overlook treating styles as distributions. Chen et al. [5] tried to match portrait images through time with a family of models for each decade and re-synthesize them. However, it is limited to simple face data. In our work, we focus on analyzing artistic style shifts in artworks depicting more complicated scenes from different eras and using a curated dataset with newly proposed flow-based models, which have not been explored before in similar analyses. + +# 2.2. Art datasets + +Various datasets focus on different domains [2, 14, 43, 82, 97]. However, in the domain of art, particularly paintings + +![](images/a9e5bf9510c512ae517c3e63a7030a432de87e69ec8217e43ae397a2f5b26e06.jpg) +Figure 2. Implicit bridge with 2D example. Two unconditional models are trained between Gaussian noise and data distributions $\mathcal{A}$ (squares) and $\mathcal{B}$ (circles), without access to mode labels (colors). Despite this, structure is preserved in noise space. By mapping samples from $\mathcal{A}$ to noise via $v^{(a)}$ and then back to $\mathcal{B}$ via $v^{(b)}$ , we show that correspondences can emerge without paired supervision. + +concerning artistic style, there is a notable absence of large-scale datasets tailored for comprehensive analysis. Existing art datasets exhibit several limitations. WikiArt [71] provides a diverse collection of styles and includes year information in some of the filenames, but is limited in the number of samples to facilitate a comprehensive analysis of artistic evolution. BAM [93] and BAM-FG [70] are restricted in accessibility as they are not publicly available and primarily consist of contemporary artworks, thereby lacking a historical perspective essential for studying stylistic progression. LAION-Styles [79] has curated a large style attribution dataset from LAION [74]. However, it lacks chronological information and does not differentiate between styles and artists, which is crucial for analyzing the evolution of styles. + +To the best of our knowledge, no previous work has curated a large-scale dataset of artwork with detailed metadata, including style, artist, year of creation, and medium. Our curated dataset (see Sec. 3.4) unifies these aspects, enabling scalable studies beyond existing approaches. + +# 2.3. Flow Matching and Stochastic Interpolants + +Flow Matching [44, 46, 54] has established Continuous Normalizing Flow (CNF) as a notable competitor to diffusion models [26, 78]. It enables simulation-free training of CNFs using an optimal transport conditional probability path with Ordinary Differential Equations (ODE), providing a more efficient alternative to traditional diffusion training and sampling methods [34, 81]. The versatility of Flow Matching has been showcased across various domains, including image [27, 29, 37, 44, 75], depth map [22], video [12], human motion [28], point cloud [94], and manifold data [7]. A more recent family of frameworks - Stochastic Interpolants [1, 47] has been proposed for unifying Flows and Diffusions. They suggested that Stochastic Interpolants can be generalized to align with various generative models, including Normalizing Flow [8], Diffusion Model [66, 80], Flow Matching [22, 29, 75], and Schrödinger Bridge [45, 85]. In this work, we develop our method within the stochastic interpolant framework and use the learned optimal transport paths (flows) to study the progression of styles across eras. + +# 3. Method + +We aim to enable the matching of artworks across different time periods, despite the absence of ground-truth correspondences. Analyzing the temporal evolution of individual paintings is inherently infeasible, as artworks are static and finalized entities. Future stylistic transformations are highly unpredictable, even when considering subsequent works by the same artist, given that many artists undergo significant stylistic changes throughout their careers (e.g., Van Gogh or Pablo Picasso). Due to this absence of explicit ground-truth data pairs, we propose an alternative formulation of the problem. Specifically, we approach it as a distribution matching task, where the evolution of artistic styles over time is understood in terms of underlying probability distributions rather than isolated instances. + +By modelling on the global distribution of artistic styles across time, we can establish meaningful correspondences spanning different centuries [77]. A principled framework for matching distributions is Optimal Transport (OT), which minimizes the transportation costs between distributions [83, 100]. To achieve this, we leverage implicit bridges and flow matching, which, upon training, inherently define an Ordinary Differential Equation (ODE) that aligns with the principles of optimal transport. This not only captures high-level stylistic transitions over time but also facilitates fine-grained alignment of individual artworks based on their placement within these temporally evolving distributions. + +# 3.1. Theoretical background + +Stochastic Interpolants In recent years, generative probabilistic models that transform noise into data have gained significant attention. These models are primarily represented by diffusion-based models [17, 66, 80, 81] and flow-based models [26, 44]. This transformation can be defined [1] by a time-dependent forward process with $t \in [0,1]$ as + +$$ +x _ {t} = \alpha_ {t} x _ {0} + \sigma_ {t} x _ {1}, \tag {1} +$$ + +where $x_0$ corresponds to data sampled from the data distribution $p_0(x)$ for which we have empirical observations available and $x_1 \sim \mathcal{N}(0, \mathbf{I})$ is a simple prior distribution. + +Algorithm 1 Transforming artworks from source age distribution to target age distribution +Input: $x_0^{(a)}\sim p_A(x)$ - data samples from age A $y_{a}$ - original chronological conditioning $y_{b}$ -chronological conditioning for desired age $y_{b} = y_{a} + \Delta y$ $v_{\theta}^{(a)}$ - model corresponds to age A, instantiated by $v_{\theta}(x_t,t,y_a)$ $v_{\theta}^{(b)}$ - target model corresponds to age B, instantiated by $v_{\theta}(x_t,t,y_b)$ $t_\mathrm{th}\in (0,1]$ - the integration threshold +Output: $x_0^{(b)}$ , the result in the target age data distribution B. 1: for each element $x_0^{(a)}\sim p_A(x)$ do 2: $\begin{array}{rlr}{x_{\mathrm{th}} = \mathrm{ODESolve}(x_0^{(a)};v_{\theta}^{(a)},0,t_{\mathrm{th}})} & \triangleright & \mathrm{obtain~the~structured~noise~from}\\ {\mathrm{source~age~data}} & {} & \\ {x_0^{(b)} = \mathrm{ODESolve}(x_{\mathrm{th}};v_{\theta}^{(b)},t_{\mathrm{th}},0)} & \triangleright & \mathrm{obtain~target~age~data~from~the}\\ {\mathrm{structured~noise}} & {} & \\ {\mathrm{return} x_0^{(b)}} & {} & \end{array}$ 3: + +The forward process is characterized by the hyperparameters $\alpha_{t}$ and $\sigma_{t}$ , with $\alpha_{t}$ decreasing and $\sigma_{t}$ increasing as time $t\in [0,1]$ progresses, interpolating between data and noise. Furthermore, the boundary conditions are normally constrained as $\alpha_0 = \sigma_1 = 1$ and $\alpha_{1} = \sigma_{0} = 0$ , so that for $t = 0$ we have pure data and $t = 1$ pure noise. In particular, flow-based models, represented by SiT [47, 75], have attracted significant interest due to their faster training and inference compared to diffusion-based models. With the above-defined forward process, the velocity is defined as + +$$ +v (x, t) = \mathbb {E} [ \dot {x} _ {t} | x _ {t} = x ], \tag {2} +$$ + +which generates the marginal probability distribution $p_t(x)$ of $x_{t}$ at time $t$ [1, 47]. This aligns with the distributions of the probability flow ODE [81] so that during inference we can generate samples by solving the Ordinary Differential Equation (ODE) backward in time starting from pure noise $\mathcal{N}(0,\mathbf{I})$ to arrive at a sample $x\sim p_0(x)$ . Ma et al. [47] demonstrates that we can train a neural network $v_{\theta}(x_{t},t)$ to parametrize the velocity using the training objective: + +$$ +\mathcal {L} _ {\mathrm {v}} (\theta) = \int_ {0} ^ {T} \mathbb {E} [ \| v _ {\theta} (x _ {t}, t) - \dot {\alpha} _ {t} x _ {0} - \dot {\sigma} _ {t} x _ {1} \| ^ {2} ] \mathrm {d} t, \tag {3} +$$ + +with $\dot{\alpha}_{t}$ and $\dot{\sigma}_t$ representing the time derivative of $\alpha_{t}$ and $\sigma_t$ , respectively. For simplicity and due to its relatively straight trajectories, we adopt the Linear Interpolant from [46, 47] with $\alpha_{t} = 1 - t$ and $\sigma_t = t$ . + +# 3.2. Bridging artworks over time + +We aim to establish meaningful correspondences between artworks originating from different historical periods. By leveraging and adapting the recent advances in generative modelling, we develop a framework that learns to transport dynamics across time-conditioned data distributions without requiring paired supervision. We begin by introducing the foundations of implicit diffusion bridges, before detailing how they can be applied to match artworks across time. + +Let $x^{(a)}$ represent a data sample from data distribution $\mathcal{A}$ . The goal is to identify the corresponding sample $x^{(b)}$ from a different data distribution $\mathcal{B}$ . This problem is inherently ill-posed due to the absence of ground-truth data pairs. To address this, we seek to establish optimal transport between the underlying manifolds and construct Schrödinger bridges between the respective distributions [9, 40]. + +Dual Diffusion Implicit Bridges (DDIB) [83] demonstrate that such bridges can be effectively constructed using a pair of pre-trained image diffusion models. Notably, the Ordinary Differential Equations (ODEs) derived from trained diffusion models inherently exhibit optimal transport between the Gaussian noise prior and the corresponding data distributions [85]. Let ODESolve denote the mapping from an initial noise $x_{1}$ to the data $x_0^{(a)}$ with $v_{\theta}^{(a)}$ defined in Eq. (3): + +$$ +\operatorname {O D E S o l v e} \left(x _ {t}; v _ {\theta} ^ {(a)}, 1, 0\right) = x _ {1} + \int_ {1} ^ {0} v _ {\theta} ^ {(a)} \left(x _ {t}, t\right) d t. \tag {4} +$$ + +Analogously, we can train ODESolve $(x_{t};v_{\theta}^{(b)},1,0)$ for mapping data between noise and $\mathcal{B}$ . Given these frameworks, the learned mappings between noise and data distributions, along with their corresponding reverse processes ODESolve $(x_{t};v_{\theta},0,1)$ , inherently constitute optimal transport pairs. As illustrated in the example in Fig. 2, we demonstrate this ability to transfer data between distributions $\mathcal{A}$ and $\mathcal{B}$ despite having no prior information about their inherent relationship. First, we use the trained ODE to map data distribution $\mathcal{A}$ to the standard Gaussian noise. A clear structure related to $\mathcal{A}$ emerges, highlighting the optimal noise-to-data pairs achieved through the ODE model's transport. Subsequently, another trained ODE transports this structured noise [67] to the distinctly different distribution $\mathcal{B}$ . In this setup, the mapping between distributions $\mathcal{A}$ and $\mathcal{B}$ remains largely preserved, even when using two independently trained ODEs that share no information. This indicates that the noise space might not be pure noise, but rather contains a certain structure as suggested in [67]. + +The independent $v_{\theta}$ can be achieved by training the same model with different conditions, namely + +$$ +v _ {\theta} ^ {(a)} = v _ {\theta} (\cdot , t, y _ {a}), v _ {\theta} ^ {(b)} = v _ {\theta} (\cdot , t, y _ {b}). \tag {5} +$$ + +Under our problem setup, the data $x_0^{(a)}$ resembles a piece of artwork, and $y_{a}$ represents the corresponding chronological information (creation year of the artwork). This parameterization allows a single model $v_{\theta}$ to learn mappings between arbitrary years $y_{a} \rightarrow y_{b}$ through conditioning, effectively replacing the need to train separate models for each temporal pair, as illustrated in Fig. 3. + +![](images/09cafccceb785dd5d8517a98e3fc4cc6c389f503ed50f671e1ca8d74e35e0fca.jpg) +(a) Training + +![](images/558d5811cd82feb59d0b1a9e314e5510c152e825cd36dceb9f6f3e4ea5df7ba4.jpg) +(b) Inference + +Training objective. Similar to Eq. (3), we define our final objective to optimize $v_{\theta}$ with the year condition $y$ as: + +$$ +\mathcal {L} (\theta) = \int_ {0} ^ {T} \mathbb {E} \left[ \| v _ {\theta} \left(x _ {t}, t, y\right) - \dot {\alpha} _ {t} x _ {0} - \dot {\sigma} _ {t} x _ {1} \| ^ {2} \right] \mathrm {d} t, \tag {6} +$$ + +where $\theta$ is the backbone parameterized by a 1D transformer. We detail the architecture of the model in the Appendix A. We operate in a compact 1D-style space, as pixel space retains excessively low-level details that contain redundancy. + +# 3.3. Inference + +With a trained model that can generate the style latents sufficiently well, we can start the matching process. The model defines a vector field $v_{\theta}(\cdot ,t,y)$ at any given $t$ and $y$ . As explained in Algorithm 1, we first transform the style representation from $x_0^{(a)}$ to the noisy representation $x_{t_{\mathrm{th}}}$ using the learned model $v_{\theta}$ . The parameter $t_\mathrm{th} > 0$ denotes the integration threshold; in practice, we typically set $t_\mathrm{th} = 1$ , corresponding to a full mapping from data space to the structured noise space. Next, we push forward the $x_{t_\mathrm{th}}$ back to $x_0^{(b)}$ using the vector field produced by the same model but with a different chronological condition $v_{\theta}(x_t,t,y_b)$ . + +Fig. 4 provides a conceptual illustration of our method using a 2D synthetic toy example, where trajectories depict the temporal evolution of data points. We construct temporally coherent toy data pairs to enable access to ground-truth data pairs. A single model is trained to perform conditional mappings across five distinct input conditions (indicated by the x-axis). Although attributes (indicated by color) gradually blend or evolve, the underlying optimal transport plan is preserved. Notably, the generated samples consistently align with the correct target modes, despite the absence of + +![](images/88e743fee0b07984407265cad180dfda0c28dddeb321d05ba974f66515708f8f.jpg) +Figure 3. Approach overview. $a)$ . Training a single $v_{\theta}$ between style embeddings and structured noise conditioned on $y$ . $b)$ . During inference, we can guide the ODE forward and backward direction with different chronological conditions $y$ between the style and latent noise space. By using different $y$ , the style embedding can be transformed to a different style manifold w.r.t. to historical patterns. +Figure 4. Illustrational example with ground truth correspondences. Dots sharing the same color are permuted versions along the x-axis. In the figure, colored round dots represent the ground truth samples. purple stars (starting from the 2nd column) indicate the warped samples originating from cyan stars (leftmost). + +explicit mode supervision, demonstrating robust structural consistency across varying style dynamics. + +# 3.4. Dataset Construction + +The curated dataset is a large-scale dataset comprising over 650,000 artworks spanning a wide range of artistic styles and mediums. As discussed in Sec. 2.2, existing datasets for art lack unified meta-information, have limited access, or are small in size. In contrast, our dataset provides standardized and cleaned metadata aggregated from multiple sources. For example, while datasets like WikiArt [84] may include creation years within filenames, such information is often unstructured. Hence, we normalize these metadata to facilitate chronological modeling and future research. + +The available metadata includes title, artist, style, year of creation (Most of the artworks originate from years ranging from 1500 to 2021), media (e.g., oil on canvas, etching, acrylic), geographical location, and image URLs. The images are provided as URLs to public museums and public art gallery websites or indexed to public datasets, where they are under the original license and copyright. An overview of the dataset is provided in Fig. 1. For experimental purposes, we used a $6/2/2$ split for training, validation, and testing, randomly sampled from centuries without selecting styles or artists explicitly to maintain the underlying data distribution. + +Deduplication. Due to the diverse sources of art pieces, duplicate samples are inevitable. While this redundancy provides additional meta-information from different sources, it results in variations of potential duplicates. To address this, we deduplicated our dataset following Somepalli et al. [79], using SSCD [59] embeddings. Entries with a cosine similarity above 0.85 were merged, consolidating their metadata, e.g., retaining the English version when multiple languages were available. This process reduced the dataset from approximately 950,000 to 650,000 art pieces, encompassing + +![](images/477f9da8baafc2a11620e86b973189994f7e1b72fddfa7b05379cc9cb23380ac.jpg) +Figure 5. Stylistic flows over time. We visualize distribution shifts for $\Delta y = 0, 20, 40, 80$ years. Blue quivers indicate the dominant movements of art pieces into the next time frame and visualize how style distribution flows across time. + +5,053 artists, 128 styles, 7,908 media, and 1,442 geographical locations. Due to the space limitation, we defer the more details of the dataset in the Appendix D. + +# 4. Experiment + +# 4.1. Choice of Style Representations + +The concept of style is inherently complex and multifaceted, encompassing aspects such as color, material, atmosphere, design, and structure. From a human perspective [39, 72], defining style is subjective and often varies depending on context. Similarly, in computer vision, style is represented through diverse latent spaces depending on the task. For instance, works on style classification and retrieval [11, 33] focus on different representations than those used in style transfer [20, 95] or in diffusion-based generation [16, 61, 87]. In both domains, style remains fluid and context-sensitive. + +To accommodate this ambiguity, our model is designed to be agnostic to the specific choice of style space. In our experiments, we primarily focus on two feature spaces: CLIP [62] and CSD [79]. As shown in Table S3, CSD excels in style-based retrieval, whereas CLIP performs slightly better in year-based retrieval and can be decoded to images with unCLIP [65] (see Figure S15). We refer readers to Appendix A for a detailed comparison of all candidates. + +# 4.2. Experimental Results + +Tracing stylistic flows across time. We visualize the temporal flows of style distributions in Figure 5, showing how our model progressively aligns embeddings of early-20th-century artworks with those from later periods. Each subplot displays kernel density estimates (KDEs) [57] in the CSD space for three sets of samples: the original early-century inputs transported by our model (green), the real artworks from the corresponding future time window (orange), and + +the fixed target distribution of late-20th-century art (blue). + +At $\Delta t = 0$ , the transformed distribution (green) coincides with the original input, as expected. As $\Delta t$ increases to 20, 40, and 80 years, the green distribution increasingly aligns with the orange ground-truth distribution for each target decade, despite the absence of sample-level correspondences. This alignment is not exact, but meaningful: the transformed samples retain certain semantic traces of their origin, resulting in a distribution that blends temporal stylistic adaptation with content continuity. Such divergence offers insight into how artistic style may evolve from a given source, even in ways not fully captured by existing data. Eventually, the green distribution approaches the latecentury manifold (blue), indicating that our model learns a smooth and coherent temporal flow (denoted by the blue quivers) from a distribution perspective. + +To bridge population-level dynamics with sample-level behavior, we track a single artwork—Berlin Street Scene (1913)—through time. Its unCLIP renderings reflect plausible stylistic changes in line with historical trends, while preserving core semantics. This illustrates the model's ability to capture stylistic evolution across distributions and within individual samples, which we explore next. + +Transforming individual artworks through time. We present visual examples demonstrating how individual art pieces evolve across time by transferring their semantic embedding in the CLIP [62] space and visualizing using the un-CLIP model [66]. In Figure 6, we show the temporal trajectories of canonical works such as Les Demoiselles d'Avignon (1907, Cubism) and Landscape near Chatou (1904, Fauvism). This task is particularly challenging as these pieces are less realistic, making semantic capture more difficult. Despite this, we preserve artistic intent and content over time by operating in a compact semantic space. + +Additionally, we compare our approach with image editing methods in pixel space [49, 96] in Figure 7. These methods fail to achieve faithful semantic transformations tied to specific periods; instead, they apply superficial style transfers that ignore historical context. In contrast, our method enables long-range interpolation through time. We further analyze the style distributions of both unconditionally generated embeddings and mapped embeddings in Appendix C.2, showing that both closely follow the stylistic distributions. + +Content preservation. Our model captures how artistic style changes over time and how such stylistic evolution influences content, including shifts in scene composition and object substitution, e.g., horses versus motorcycles in Figure 7. Because style and content are often deeply entangled, meaningful temporal transformation requires flexibility. When the threshold parameter $t_{\mathrm{th}}$ is set too low (Section 3.3), the model is overly constrained to the original input, limit + +![](images/6b45fb4c8f449c4767de1be949786f7469ea46bda4c95f919999d59315d9b852.jpg) +Figure 6. Qualitative comparison of editing methods across different historical styles. Each row shows results from a different method. The 1st and 5th columns show original artworks (Les Demoiselles d'Avignon - 1907 and Landscape near Chatou - 1904). + +
MethodRococoImpressionismRealism
R@1R@5δ↓τ(%)↑R@1R@5δ↓τ(%)↑R@1R@5δ↓τ(%)↑
RF-inversion [68]5.69.30.69083.710.526.30.65876.82.46.50.74966.3
LEDDS ++ [3]12.623.00.52879.415.430.40.58569.69.720.50.50964.9
SDEdit [49]6.315.80.57245.65.313.40.60534.53.712.40.61245.6
SDEdit (fine-tuned)37.942.80.26564.734.147.30.27247.227.538.40.23657.3
Ours72.585.10.02484.865.076.30.02475.858.569.30.02568.6
+ +Table 1. Quality of the time matching when transferring artworks over time. We sampled a compact subset of a hundred samples for each style to calculate the metrics. We transfer clusters of artworks to randomly selected years within the range of $[-100, 100]$ and compute the mean metrics. For $\tau$ , we used the nearest $nns[:25]$ neighbors as positive samples and further $nns[50:75]$ as the negative samples for each chosen anchor, where $nns$ denotes its nearest neighbors. See extended table in Table S4. + +ing stylistic adaptation. This results in hybrid outputs, such as a Steam Horse locomotive, that blend past and future characteristics without fully transitioning. + +Content deviation is a broader challenge across models. Most methods face a trade-off between low-level fidelity and high-level semantic consistency. Pixel-focused techniques [3, 49, 67] often struggle to produce meaningful temporal shifts even with adjusted guidance scales, sometimes leading to failures like buildings morphing into people, as seen in Figure 6. In contrast, our model preserves conceptual content while enabling coherent, temporally grounded transformations, striking a balance between semantic stability and stylistic progression. + +Quantitative evaluation of time travel. Assessing the quality of temporal transitions in artworks is challenging due to the absence of ground-truth correspondences. However, semi-ground-truth evidence exists: artworks created close in time and close in style are more likely to evolve together rather than disperse. Based on this intuition, we utilized three key metrics: + +- Recall@k assesses alignment with the target era by checking whether time-traveled samples retrieve temporally appropriate neighbors in style space. +- Compactness $\delta$ measures how coherent the transformed + +samples are by comparing their variance to that of random samples in the target year: $\frac{\sigma_{\text{samples}}^2}{\sigma_{\text{target}}^2}$ . A good transformation should preserve this coherence, as source samples are initially similar in style and time, resulting in lower $\delta$ . + +- Triplet consistency $\tau$ is adapted from triplet loss [23, 73]. It evaluates whether the relative relation among anchor, positive, and negative samples is preserved posttransformation, reflecting the model's ability to maintain local structure. + +We evaluate on subsets of 100 samples by selecting localized neighborhoods of artwork within several style labels from the test split. These subsets serve as the starting samples for time mapping. Due to space constraints, we present results for the few most prominent artistic styles in Table 1, though our analysis extends across a much broader range. Since most editing baselines operate directly in pixel space, we perform an additional step of encoding their generated outputs back into the style space to enable a fair comparison, which allows us to compute nearest neighbor recall and compactness in a shared representation space. + +As shown in Table 1 (we offer more analysis in Table S4 in the Appendix), conventional pixel-based image-editing methods [3, 49, 67] fail to achieve both objectives: they neither align generated samples with the appropriate data manifold, reflected by the lower Recall@ $k$ , nor preserve + +![](images/797afe895a26f3af82ffb8e307eaddb62531a04aa20e3119afe6d0d11f9e2f06.jpg) +Figure 7. Flow preserves content semantics across time. We transform a motorcyclist to the year 1800 with increasing $t_{\mathrm{th}}$ (left to right, top row). Low $t_{\mathrm{th}}$ leads to incorrect mappings due to limited editing flexibility, while higher values retain semantic identity ("man on a horse") with correct stylistic traits. + +local structural coherence, indicated by the higher $\delta$ ). This is mainly due to their reliance on pixel-level cues, which are insufficient for modeling more profound, semantic style shifts. For instance, RF-Inversion achieves high fidelity editing [67] but leads to minimal pixel shifts, evidenced by the low recall. We address these issues by operating in a compact latent space and focusing on semantic information to better align with historical stylistic flows. + +Since our method is trained exclusively on the curated dataset, a distributional mismatch may arise when comparing against baseline models such as Stable Diffusion, which are pretrained on broader, general-domain datasets. For a fair comparison, we also fine-tuned the base Stable Diffusion model on our dataset until convergence. Specifically, we adapt the SD 1.5 model using chronologically conditioned prompts of the form "A painting in {year} style". This finetuning procedure improves both the synthesis FID in Table 2 and retrieval-based matching performance in Table 1, confirming our dataset's value. However, applying SDEdit remains suboptimal for our task, as the method struggles to perform temporally faithful transformations. This performance gap persists across all evaluated artistic styles, where our method consistently yields more coherent and historically aligned style transitions. + +Comparison to synthesis baselines. As previously discussed, for visualization purposes, we project samples into the CLIP [62] embedding space and render them using the unCLIP model [66]. This enables quantitative evaluation using the standard Frechet Inception Distance (FID) [25], which assesses both the realism of the generated outputs and their fidelity to the target temporal style distribution. + +We compared our method against three Latent Diffusion Models (LDM) variants: Stable Diffusion 1.5 [66], Stable + +
ModelFID↓conditioned on century
16th17th18th19th20th
SD 1.5 [66]131.1▼40.4118.9▼37.1103.5▼17.2149.9▼77.9180.4▼101.5
SD 2.1214.4▼123.7147.1▼65.3134.2▼47.9170.2▼98.2167.1▼88.2
SDXL [60]186.6▼95.9142.4▼60.6143.4▼57.1114.8▼42.8146.1▼67.2
FLUX.1-schnell [38]199.2▼108.5142.2▼60.4137.3▼51.0122.8▼50.8126.1▼47.2
SD 1.5 (fine-tuned)111.7▼2196.4▼14.688.8▼2.596.4▼24.4150.7▼71.8
Ours90.781.886.372.078.9
+ +Table 2. Feature synthesis across centuries. Evaluation metrics are computed on 1,000 samples per century. While Stable Diffusion and FLUX rely on text prompts for year conditioning, our model conditions directly on this information. Our method exhibits stronger alignment with ground truth distributions. + +Diffusion 2.1 [66], and SDXL [60], as well as the flow-matching based FLUX model [38]. We also compare against the fine-tuned Stable Diffusion 1.5, with results summarized in Table 2. SDXL [60] has demonstrated strong artistic understanding and has become the foundation for many recent stylistic image generation methods [16, 24, 76, 87]. While SDXL performs well on certain styles, such as Renaissance (16th century) and Cubism (20th century) in the Appendix Figure S13, it struggles to capture stylistic distributions comprehensively. FLUX [38], despite its training scale and exhibiting strong performance across text-to-image tasks, also struggles to capture temporal style distributions effectively (see Figure S14). In contrast, our model, trained on the proposed dataset, better captures the diversity of artistic styles. Our approach achieves significantly lower FID scores than standard text-conditioned models, demonstrating better alignment with the underlying feature distribution of different art periods. Furthermore, it provides more consistent stylistic alignment across centuries. Further rendering results for all models are available in the Appendix. + +# 5. Conclusion + +Studying the evolution of art is challenging due to the lack of ground-truth correspondences between artworks from different periods. We present a framework that models stylistic progression using stochastic interpolants and dual diffusion implicit bridges, enabling distribution-level alignment without paired data. By formulating art evolution as optimal transport in a learned style space, our method captures continuous temporal transitions in artistic style. + +A central contribution is a dataset of over 650,000 artworks spanning 500 years, which we use to demonstrate that our model captures long-range stylistic shifts. Beyond modeling distributions, our method also enables tracing the stylistic origins and plausible futures of individual artworks, allowing for fine-grained longitudinal analysis. Our work explores a less-studied direction in computer vision: using generative models to investigate the evolution of art and support cultural and historical analysis. We believe our approach offers a foundation for future interdisciplinary work at the intersection of machine learning and cultural analytics. + +# Acknowledgement + +This project was supported by Bayer AG, the Federal Ministry for Economic Affairs and Energy within the project "NXT GEN AI METHODS - Generative Methoden für Perzeption, Prädiktion und Planung", the project "Genius-Robot" (01IS24083) funded by the Federal Ministry of Research, Technology and Space (BMFTR), and the bidt project KLIMA-MEMES. The authors gratefully acknowledge the Gauss Center for Supercomputing for providing compute through the NIC on JUWELS/JUPITER at JSC and the HPC resources supplied by the NHR @FAU Erlangen. + +# References + +[1] Michael S Albergo, Nicholas M Boffi, and Eric Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions. arXiv, 2023. 1, 3, 4 +[2] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In ICCV, 2019. 2 +[3] Manuel Brack, Felix Friedrich, Katharia Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. Ledits++: Limitless image editing using text-to-image models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8861-8870, 2024. 7 +[4] Levin Brinkmann, Deniz Gezerli, KV Kleist, Thomas F Müller, Iyad Rahwan, and Niccolò Pescetelli. Hybrid social learning in human-algorithm cultural transmission. Philosophical Transactions of the Royal Society A, 380(2227), 2022. 1 +[5] Eric Ming Chen, Jin Sun, Apoorv Khandelwal, Dani Lischinski, Noah Snavely, and Hadar Averbuch-Elor. What's in a decade? transforming faces through time. In Computer Graphics Forum, 2023. 2 +[6] Haibo Chen, Lei Zhao, Zhizhong Wang, Huiming Zhang, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Artistic style transfer with internal-external learning and contrastive learning. Advances in Neural Information Processing Systems (NeurIPS), 34, 2021. 2 +[7] Ricky TQ Chen and Yaron Lipman. Riemannian flow matching on general geometries. arXiv, 2023. 3 +[8] Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. NeurIPS, 2018. 3 +[9] Yongxin Chen, Tryphon T Georgiou, and Michele Pavon. On the relation between optimal transport and schrödinger bridges: A stochastic control viewpoint. Journal of Optimization Theory and Applications, 169:671-691, 2016. 4 +[10] Ceren Comert, Murat Özbayoglu, and Cosku Kasnakoğlu. Painter prediction from artworks with transfer learning. In 2021 7th International Conference on Mechatronics and Robotics Engineering (ICMRE). IEEE, 2021. 1 +[11] Elliot Crowley and Andrew Zisserman. The state of the art: Object retrieval in paintings using discriminative regions. In + +Proceedings of the British Machine Vision Conference 2014. British Machine Vision Association, 2014. 6 +[12] Aram Davtyan, Sepehr Sameni, and Paolo Favaro. Randomized conditional flow matching for video prediction. arXiv, 2022. 3 +[13] Valentin De Bortoli, James Thornton, Jeremy Heng, and Arnaud Doucet. Diffusion schrödinger bridge with applications to score-based generative modeling. Advances in Neural Information Processing Systems, 2021. 1 +[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, 2009. 2 +[15] Yingying Deng, Fan Tang, Weiming Dong, Chongyang Ma, Feiyue Huang, Oliver Deussen, and Changsheng Xu. Exploring the representativity of art paintings. IEEE Transactions on Multimedia, 2020. 2 +[16] Yarden Frenkel, Yael Vinker, Ariel Shamir, and Daniel Cohen-Or. Implicit style-content separation using b-lora. In ECCV, 2024. 6, 8 +[17] Michael Fuest, Pingchuan Ma, Ming Gui, Johannes S Fischer, Vincent Tao Hu, and Bjorn Ommer. Diffusion models and representation learning: A survey. arXiv preprint arXiv:2407.00783, 2024. 3 +[18] Shangqi Gao, Hangqi Zhou, Yibo Gao, and Xiahai Zhuang. Bayeseg: Bayesian modeling for medical image segmentation with interpretable generalizability. Medical Image Analysis, 89, 2023. 1 +[19] Noa Garcia, Benjamin Renoust, and Yuta Nakashima. Contextnet: representation and exploration for painting classification and retrieval in context. International Journal of Multimedia Information Retrieval, 9(1), 2020. 1, 2 +[20] Leon Gatys, Alexander S Ecker, and Matthias Bethge. Texture synthesis using convolutional neural networks. In Advances in Neural Information Processing Systems (NeurIPS), 2015. 6 +[21] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2 +[22] Ming Gui, Johannes Schusterbauer, Ulrich Prestel, Pingchuan Ma, Dmytro Kotovenko, Olga Grebenkova, Stefan Andreas Baumann, Vincent Tao Hu, and Björn Ommer. Depthfm: Fast monocular depth estimation with flow matching. AAAI, 2025. 3 +[23] Alexander Hermans, Lucas Beyer, and Bastian Leibe. In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737, 2017. 7 +[24] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. In CVPR, 2024. 8 +[25] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 8 +[26] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems (NeurIPS), 33, 2020. 3 + +[27] Vincent Tao Hu, Yunlu Chen, Mathilde Caron, Yuki M. Asano, Cees G.M. Snoek, and Björn Ommer. Guided diffusion from self-supervised diffusion features. In ARXIV, 2023. 3 +[28] Vincent Tao Hu, Wenzhe Yin, Pingchuan Ma, Yunlu Chen, Basura Fernando, Yuki M. Asano, Efstratos Gavves, Pascal Mettes, Björn Ommer, and Cees G.M. Snoek. Motion flow matching for human motion synthesis and editing. In ARXIV, 2023. 3 +[29] Vincent Tao Hu, Stefan Andreas Baumann, Ming Gui, Olga Grebenkova, Pingchuan Ma, Johannes Fischer, and Björn Ommer. *Zigma: A dit-style zigzag mamba diffusion model*. In European conference on computer vision, pages 148-166. Springer, 2024. 3 +[30] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2017. 2 +[31] Herve Jegou, Matthijs Douze, and Cordelia Schmid. Product quantization for nearest neighbor search. IEEE transactions on pattern analysis and machine intelligence, 33(1):117-128, 2010. 1 +[32] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Proceedings of the European Conference on Computer Vision (ECCV). Springer, 2016. 2 +[33] Sergey Karayev, Matthew Trentacoste, Helen Han, Aseem Agarwala, Trevor Darrell, Aaron Hertzmann, and Holger Winnemoeller. Recognizing image style. arXiv preprint arXiv:1311.3715, 2013. 6 +[34] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, 2022. 3 +[35] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 1 +[36] Dmytro Kotovenko, Matthias Wright, Arthur Heimbrecht, and Bjorn Ommer. Rethinking style transfer: From pixels to parameterized brushstrokes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[37] Felix Krause, Timy Phan, Ming Gui, Stefan Andreas Baumann, Vincent Tao Hu, and Björn Ommer. Tread: Token routing for efficient architecture-agnostic diffusion training. arXiv preprint arXiv:2501.04765, 2025. 3 +[38] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024.8, 2 +[39] Berel Lang. The concept of style. Cornell University Press, 1987. 6 +[40] Christian Léonard. A survey of the schr\'' odinger problem and some of its connections with optimal transport. arXiv preprint arXiv:1308.0215, 2013. 4 +[41] Wen Li, Muyuan Fang, Cheng Zou, Biao Gong, Ruobing Zheng, Meng Wang, Jingdong Chen, and Ming Yang. Style-tokensizer: Defining image style by a single instance for controlling diffusion models. In European Conference on Computer Vision, pages 110-126. Springer, 2024. 2 + +[42] Xueting Li, Sifei Liu, Jan Kautz, and Ming-Hsuan Yang. Learning linear transformations for fast image and video style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[43] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2 +[44] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. *ICLR*, 2023. 3 +[45] Guan-Horng Liu, Tianrong Chen, Oswin So, and Evangelos Theodorou. Deep generalized schrödinger bridge. NeurIPS, 2022. 3 +[46] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. *ICLR*, 2023. 3, 4 +[47] Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. arXiv, 2024. 3, 4 +[48] L. McInnes, J. Healy, and J. Melville. UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction. ArXiv, 2018. 1 +[49] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 6, 7, 3 +[50] Amir Merchant, Simon Batzner, Samuel S Schoenholz, Muratahan Aykol, Gowoon Cheon, and Ekin Dogus Cubuk. Scaling deep learning for materials discovery. Nature, 624 (7990), 2023. 1 +[51] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, 2024. 1 +[52] Kevin Musgrave, Serge Belongie, and Ser-Nam Lim. A metric learning reality check. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16, pages 681-699. Springer, 2020. 1 +[53] Ivona Najdenkoska, Xiantong Zhen, Marcel Worring, and Ling Shao. Uncertainty-aware report generation for chest x-rays by variational topic inference. Medical Image Analysis, 82, 2022. 1 +[54] Kirill Neklyudov, Daniel Severo, and Alireza Makhzani. Action matching: A variational method for learning stochastic dynamics from samples. arXiv, 2022. 3 +[55] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bo + +janowski. Dinov2: Learning robust visual features without supervision, 2023. 3 +[56] Norihiro Oyama, Noriko N Ishizaki, Satoshi Koide, and Hiroaki Yoshida. Deep generative model super-resolves spatially correlated multiregional climate data. Scientific Reports, 13(1), 2023. 1 +[57] Emanuel Parzen. On estimation of a probability density function and mode. The annals of mathematical statistics, 33(3):1065-1076, 1962. 6 +[58] Pablo Pernias, Dominic Rampas, Mats Leon Richter, Christopher Pal, and Marc Aubreville. Würstchen: An efficient architecture for large-scale text-to-image diffusion models. In The Twelfth International Conference on Learning Representations, 2023. 1 +[59] Ed Pizzi, Sreya Dutta Roy, Sugosh Nagavara Ravindra, Priya Goyal, and Matthijs Douze. A self-supervised descriptor for image copy detection. In CVPR, 2022. 5 +[60] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 8, 2 +[61] Tianhao Qi, Shancheng Fang, Yanze Wu, Hongtao Xie, Jiawei Liu, Lang Chen, Qian He, and Yongdong Zhang. Deadiff: An efficient stylization diffusion model with disentangled representations. arXiv preprint arXiv:2403.06951, 2024. 2, 6 +[62] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning. PMLR, 2021. 2, 6, 8, 1 +[63] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML). PMLR, 2021. 3 +[64] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 1 +[65] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 6, 1, 2 +[66] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022. 1, 3, 6, 8, 2 +[67] L Rout, Y Chen, N Ruiz, C Caramanis, S Shakkottai, and W Chu. Semantic image inversion and editing using rectified stochastic differential equations. 2024. 4, 7, 8 +[68] Litu Rout, Yujia Chen, Nataniel Ruiz, Constantine Caramanis, Sanjay Shakkottai, and Wen-Sheng Chu. Semantic image inversion and editing using rectified stochastic differential equations. In ICLR, 2025. 7 + +[69] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1 +[70] Dan Ruta, Saeid Motiian, Baldo Faieta, Zhe Lin, Hailin Jin, Alex Filipkowski, Andrew Gilbert, and John Collomosse. Aladin: All layer adaptive instance normalization for fine-grained style similarity. In ICCV, 2021. 1, 3 +[71] Babak Saleh and Ahmed Elgammal. Large-scale classification of fine-art paintings: Learning the right metric on the right feature. arXiv preprint arXiv:1505.00855, 2015. 1, 3 +[72] Meyer Schapiro. Theory and philosophy of art: Style, artist, and society. 1994. 6 +[73] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 7 +[74] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 3, 2 +[75] Johannes Schusterbauer, Ming Gui, Pingchuan Ma, Nick Stracke, Stefan A Baumann, and Björn Ommer. Boosting latent diffusion with flow matching. ECCV, 2024. 3, 4 +[76] Viraj Shah, Nataniel Ruiz, Forrester Cole, Erika Lu, Svetlana Lazebnik, Yuanzhen Li, and Varun Jampani. Ziplora: Any subject in any style by effectively merging loras. In ECCV, 2024. 8 +[77] Higor YD Sigaki, Matjaž Perc, and Haroldo V Ribeiro. History of art paintings through the lens of entropy and complexity. Proceedings of the National Academy of Sciences, 115(37):E8585-E8594, 2018. 3 +[78] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 3 +[79] Gowthami Somepalli, Anubhav Gupta, Kamal Gupta, Shramay Palta, Micah Goldblum, Jonas Geiping, Abhinav Shrivastava, and Tom Goldstein. Measuring style similarity in diffusion models. ECCV, 2024. 2, 3, 5, 6, 1 +[80] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3 +[81] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 3, 4 +[82] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 2 +[83] Xuan Su, Jiaming Song, Chenlin Meng, and Stefano Ermon. Dual diffusion implicit bridges for image-to-image translation. *ICLR*, 2023. 1, 3, 4 +[84] Wei Ren Tan, Chee Seng Chan, Hernan Aguirre, and Kiyoshi Tanaka. Improved artgan for conditional synthesis of natural image and artwork. IEEE Transactions on Image Processing, 28(1):394-409, 2019. 5 + +[85] Zhicong Tang, Tiankai Hang, Shuyang Gu, Dong Chen, and Baining Guo. Simplified diffusion schrödinger bridge. arXiv preprint arXiv:2403.14623, 2024. 1, 3, 4 +[86] Gemma Canet Tarrés, Dan Ruta, Tu Bui, and John Collomosse. Parasol: Parametric style control for diffusion image synthesis. arXiv preprint arXiv:2303.06464, 2023. 2 +[87] Haofan Wang, Matteo Spinelli, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv preprint arXiv:2404.02733, 2024. 6, 8 +[88] Sheng-Yu Wang, Alexei A Efros, Jun-Yan Zhu, and Richard Zhang. Evaluating data attribution for text-to-image models. In ICCV, 2023. 2 +[89] Zhizhong Wang, Lei Zhao, Sihuan Lin, Qihang Mo, Huiming Zhang, Wei Xing, and Dongming Lu. Glstylenet: exquisite style transfer combining global and local pyramid features. IET Computer Vision, 14(8), 2020. 2 +[90] Zhizhong Wang, Lei Zhao, Haibo Chen, Ailin Li, Zhiwen Zuo, Wei Xing, and Dongming Lu. Texture reformer: Towards fast and universal interactive texture transfer. In Proceedings of the AAAI Conference on Artificial Intelligence, 2022. 2 +[91] Zhizhong Wang, Lei Zhao, Haibo Chen, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. DivSwapper: Towards diversified patch-based arbitrary style transfer. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence (IJCAI), 2022. 2 +[92] Zhizhong Wang, Lei Zhao, Zhiwen Zuo, Ailin Li, Haibo Chen, Wei Xing, and Dongming Lu. Microast: Towards super-fast ultra-resolution arbitrary style transfer. In Proceedings of the AAAI Conference on Artificial Intelligence, 2023. 2 +[93] Michael J Wilber, Chen Fang, Hailin Jin, Aaron Hertzmann, John Collomosse, and Serge Belongie. Bam! the behave artistic media dataset for recognition beyond photography. In ICCV, 2017. 1, 3 +[94] Lemeng Wu, Dilin Wang, Chengyue Gong, Xingchao Liu, Yunyang Xiong, Rakesh Ranjan, Raghuraman Krishnamoorthi, Vikas Chandra, and Qiang Liu. Fast point cloud generation with straight flows. arXiv, 2022. 3 +[95] Zijie Wu, Zhen Zhu, Junping Du, and Xiang Bai. Ccpl: Contrastive coherence preserving loss for versatile style transfer. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 6 +[96] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 6 +[97] Sheng Zhang, Yanbo Xu, Naoto Usuyama, Hanwen Xu, Jaspreet Bagga, Robert Tinn, Sam Preston, Rajesh Rao, Mu Wei, Naveen Valluri, et al. Biomedclip: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs. arXiv preprint arXiv:2303.00915, 2023. 2 +[98] Yuxin Zhang, Fan Tang, Weiming Dong, Haibin Huang, Chongyang Ma, Tong-Yee Lee, and Changsheng Xu. Domain enhanced arbitrary image style transfer via contrastive learning. In ACM SIGGRAPH 2022 Conference Proceedings, 2022. 2 + +[99] Wentao Zhao, Dalin Zhou, Xinguo Qiu, and Wei Jiang. How to represent paintings: A painting classification using artistic comments. Sensors, 21(6), 2021. 1, 2 +[100] Linqi Zhou, Aaron Lou, Samar Khanna, and Stefano Ermon. Denoising diffusion bridge models. *ICLR*, 2024. 1, 3 +[101] Zhiwen Zuo, Lei Zhao, Shuobin Lian, Haibo Chen, Zhizhong Wang, Ailin Li, Wei Xing, and Dongming Lu. Style fader generative adversarial networks for style degree controllable artistic style transfer. In Proc. Int. Joint Conf. on Artif. Intell.(IJCAI), 2022. 2 \ No newline at end of file diff --git a/2025/Stochastic Interpolants for Revealing Stylistic Flows across the History of Art/images.zip b/2025/Stochastic Interpolants for Revealing Stylistic Flows across the History of Art/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..62882d31987a002e46143976177626df813edaa4 --- /dev/null +++ b/2025/Stochastic Interpolants for Revealing Stylistic Flows across the History of Art/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:04c3c1ccc7e18f305231c6dce4ab088a6670b3e2b9019fa87392bf3582613232 +size 523995 diff --git a/2025/Stochastic Interpolants for Revealing Stylistic Flows across the History of Art/layout.json b/2025/Stochastic Interpolants for Revealing Stylistic Flows across the History of Art/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9aad64ae6881a15a25b41ba5191760b884ed98d5 --- /dev/null +++ b/2025/Stochastic Interpolants for Revealing Stylistic Flows across the History of Art/layout.json @@ -0,0 +1,10138 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 103, + 544, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 103, + 544, + 121 + ], + "spans": [ + { + "bbox": [ + 67, + 103, + 544, + 121 + ], + "type": "text", + "content": "Stochastic Interpolants for Revealing Stylistic Flows across the History of Art" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 143, + 483, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 143, + 483, + 172 + ], + "spans": [ + { + "bbox": [ + 127, + 143, + 483, + 172 + ], + "type": "text", + "content": "Pingchuan Ma* Ming Gui* Johannes Schusterbauer \nXiaopei Yang Olga Grebenkova Vincent Tao Hu Björn Ommer" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 185, + 502, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 185, + 502, + 200 + ], + "spans": [ + { + "bbox": [ + 107, + 185, + 502, + 200 + ], + "type": "text", + "content": "CompVis @ LMU Munich Munich Center for Machine Learning (MCML)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 252, + 297, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 297, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 297, + 468 + ], + "type": "text", + "content": "Generative models have made rapid progress in content creation, particularly in synthesizing artworks and capturing stylistic variation. However, most methods operate at the level of individual images, limiting their ability to reveal broader stylistic trends and temporal transitions. We address this by introducing a framework that models stylistic evolution as an optimal transport problem in a learned style space, using stochastic interpolants and dual diffusion implicit bridges to align artistic distributions across time without requiring paired data. A central contribution is a diverse dataset of over 650,000 artworks spanning 500 years, curated with metadata across multiple genres. Together, our method and dataset enable tracing long-range stylistic transitions and plausible futures of individual artworks, supporting fine-grained temporal analysis. This offers a new tool for modeling historical patterns in visual culture and opens up promising directions in visual understanding. Code and dataset: https://github.com/CompVis/Art-fm" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 492, + 135, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 135, + 504 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 135, + 504 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 513, + 296, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 513, + 296, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 513, + 296, + 608 + ], + "type": "text", + "content": "Recent advances in computer vision have sparked growing interest in applications in interdisciplinary research domains. Such applications not only try to offer novel perspectives, but also address challenges that were impossible before, such as those seen in medical imaging [18, 53], novel detection in material science [50], and simulation of climate [56]. Despite these successes, there remains a noticeable gap in their use for cultural studies [4], particularly in art history." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 609, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 296, + 693 + ], + "type": "text", + "content": "With the rise of generative models such as Stable Diffusion [66], DALL-E [64], and others [51, 58, 69], there has been increasing use of these tools for artwork synthesis and content creation. However, their primary focus remains on image generation and editing, rather than analyzing the broader narrative of art history [10, 19, 99]. As a result, current models are not designed to capture the complex evo" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 228, + 514, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 514, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 514, + 239 + ], + "type": "text", + "content": "lution of artistic movements and styles over time." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 240, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 555, + 396 + ], + "type": "text", + "content": "In this paper, we ask: How would a piece of art evolve? What might it look like if created centuries earlier or later? Naturally, no ground-truth pairs exist for such a task, and the artistic style of a creator is inherently dynamic and historically constrained. To address this challenge, we aim to analyze the flow of artistic styles over time. Given a single artwork, we try to determine its closest \"projection\" onto the manifold of artistic expressions corresponding to a given historical period. While optimal transport provides a natural formulation for such a task, considering only a single isolated artwork is insufficient. Instead, we reformulate the problem as a distribution matching task, where we study the evolution of the underlying distribution of art over time." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 397, + 556, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 556, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 556, + 540 + ], + "type": "text", + "content": "Recent advances in diffusion bridges such as DDIB [83], DSB [13, 85], and DDBMs [100] have shown that domain translation can be performed implicitly, without the need for paired supervision. We build on this idea in the context of art history by incorporating stochastic interpolants [1], enabling us to model the temporal evolution of artistic styles without requiring explicitly aligned training data. Unlike traditional style transfer methods that transform images based on predefined and discrete styles, our approach projects styles onto their corresponding distributions across time. Operating in a compact latent space allows us to abstract away pixel-level distraction and focus on meaningful stylistic shifts." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 542, + 556, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 556, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 556, + 685 + ], + "type": "text", + "content": "To support this temporal modeling, we require a dataset that spans centuries and reflects a wide variety of artistic styles. While datasets like WikiArt [71] offer breadth in genre and metadata, they lack the scale and historical depth needed to model long-term stylistic progression. Others, such as BAM [93] and BAM-FG [70], are either restricted in access or skewed toward modern art. To address this, we curate a new unified dataset of over 650,000 artworks annotated with creation year, location, artist, medium, and other metadata (Fig. 1). This dataset enables the study of artistic evolution at scale and over time. In summary, our key contributions are as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "- We propose the first generative framework that explicitly models the temporal progression of artistic styles across" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Equal Contribution" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5867" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 72, + 539, + 357 + ], + "blocks": [ + { + "bbox": [ + 75, + 72, + 539, + 357 + ], + "lines": [ + { + "bbox": [ + 75, + 72, + 539, + 357 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 539, + 357 + ], + "type": "image", + "image_path": "23df6f6359b304d32dce1aa8383cb1aebd9ab9e8467ea9ad03ba6e3d2502e6f6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 366, + 555, + 388 + ], + "lines": [ + { + "bbox": [ + 54, + 366, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 54, + 366, + 555, + 388 + ], + "type": "text", + "content": "Figure 1. Overview of the proposed dataset. We have curated and unified our large-scale dataset, including 650,000 artworks from different creation periods (ranging from 1500 to 2021) and styles with other meta-information, e.g., year, title, artist, media, etc." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 410, + 282, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 410, + 282, + 421 + ], + "spans": [ + { + "bbox": [ + 63, + 410, + 282, + 421 + ], + "type": "text", + "content": "historical periods, introduced in Sec. 3.2 and Sec. 3.3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 422, + 296, + 506 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 422, + 295, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 295, + 470 + ], + "type": "text", + "content": "- To support this, we construct a large-scale, unified dataset of 650,000 artworks annotated with creation year and other metadata, covering five centuries of diverse artistic styles (Sec. 3.4). An overview is shown in Fig. 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 470, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 470, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 296, + 506 + ], + "type": "text", + "content": "- Our method models temporal transitions without requiring paired data. We outperform competitive baselines in capturing long-range stylistic shifts (Sec. 4.2)." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 519, + 147, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 147, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 147, + 532 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 540, + 186, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 186, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 186, + 552 + ], + "type": "text", + "content": "2.1. Stylistic representation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 558, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 297, + 713 + ], + "type": "text", + "content": "The pioneering work by Gatys et al. [21] marked the beginning of the style transfer era, which defines style as Gram Matrices of pre-trained network features. Various methods have been proposed to improve different aspects of style representation, ranging from efficiency-focused approaches like [32, 42, 92] to quality-centric ones such as [36, 89, 98, 101]. Moreover, efforts have been made to enhance the generality and diversity of style transfer methods [30, 90, 91, 95]. Despite these achievements, most existing approaches [6, 41, 61, 86, 98] focus on image editing or the disentanglement of style and content through explicit style/content representations or user-defined prompts. Data-driven methods as [15] defined the representativity of art" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 410, + 555, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 555, + 540 + ], + "type": "text", + "content": "works as a deep feature fusion incorporating not only images, but also painting styles, creation time, and authorship information. On the other hand, ContextNet [19] aims to represent and explore artistic attributes in different modalities and their relationships to improve the retrieval of artworks in automatic art analysis. ArtGCN [99] uses a graph CNN to construct stylistic representation based on artistic comments. More recent works, such as GDA [88] and CSD [79], train style descriptors by fine-tuning on top of the existing feature space of CLIP [62] through contrastive learning with synthetic or curated data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 542, + 556, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 556, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 556, + 662 + ], + "type": "text", + "content": "These studies typically concentrate on rendering or differentiating individual images with specific styles and often overlook treating styles as distributions. Chen et al. [5] tried to match portrait images through time with a family of models for each decade and re-synthesize them. However, it is limited to simple face data. In our work, we focus on analyzing artistic style shifts in artworks depicting more complicated scenes from different eras and using a curated dataset with newly proposed flow-based models, which have not been explored before in similar analyses." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 671, + 394, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 394, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 394, + 683 + ], + "type": "text", + "content": "2.2. Art datasets" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "Various datasets focus on different domains [2, 14, 43, 82, 97]. However, in the domain of art, particularly paintings" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5868" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 82, + 478, + 153 + ], + "blocks": [ + { + "bbox": [ + 134, + 82, + 478, + 153 + ], + "lines": [ + { + "bbox": [ + 134, + 82, + 478, + 153 + ], + "spans": [ + { + "bbox": [ + 134, + 82, + 478, + 153 + ], + "type": "image", + "image_path": "a9e5bf9510c512ae517c3e63a7030a432de87e69ec8217e43ae397a2f5b26e06.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 163, + 555, + 198 + ], + "lines": [ + { + "bbox": [ + 54, + 163, + 555, + 198 + ], + "spans": [ + { + "bbox": [ + 54, + 163, + 555, + 198 + ], + "type": "text", + "content": "Figure 2. Implicit bridge with 2D example. Two unconditional models are trained between Gaussian noise and data distributions " + }, + { + "bbox": [ + 54, + 163, + 555, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 54, + 163, + 555, + 198 + ], + "type": "text", + "content": " (squares) and " + }, + { + "bbox": [ + 54, + 163, + 555, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 54, + 163, + 555, + 198 + ], + "type": "text", + "content": " (circles), without access to mode labels (colors). Despite this, structure is preserved in noise space. By mapping samples from " + }, + { + "bbox": [ + 54, + 163, + 555, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 54, + 163, + 555, + 198 + ], + "type": "text", + "content": " to noise via " + }, + { + "bbox": [ + 54, + 163, + 555, + 198 + ], + "type": "inline_equation", + "content": "v^{(a)}" + }, + { + "bbox": [ + 54, + 163, + 555, + 198 + ], + "type": "text", + "content": " and then back to " + }, + { + "bbox": [ + 54, + 163, + 555, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 54, + 163, + 555, + 198 + ], + "type": "text", + "content": " via " + }, + { + "bbox": [ + 54, + 163, + 555, + 198 + ], + "type": "inline_equation", + "content": "v^{(b)}" + }, + { + "bbox": [ + 54, + 163, + 555, + 198 + ], + "type": "text", + "content": ", we show that correspondences can emerge without paired supervision." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 213, + 297, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 213, + 297, + 380 + ], + "spans": [ + { + "bbox": [ + 54, + 213, + 297, + 380 + ], + "type": "text", + "content": "concerning artistic style, there is a notable absence of large-scale datasets tailored for comprehensive analysis. Existing art datasets exhibit several limitations. WikiArt [71] provides a diverse collection of styles and includes year information in some of the filenames, but is limited in the number of samples to facilitate a comprehensive analysis of artistic evolution. BAM [93] and BAM-FG [70] are restricted in accessibility as they are not publicly available and primarily consist of contemporary artworks, thereby lacking a historical perspective essential for studying stylistic progression. LAION-Styles [79] has curated a large style attribution dataset from LAION [74]. However, it lacks chronological information and does not differentiate between styles and artists, which is crucial for analyzing the evolution of styles." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 384, + 298, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 298, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 298, + 445 + ], + "type": "text", + "content": "To the best of our knowledge, no previous work has curated a large-scale dataset of artwork with detailed metadata, including style, artist, year of creation, and medium. Our curated dataset (see Sec. 3.4) unifies these aspects, enabling scalable studies beyond existing approaches." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 464, + 281, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 281, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 281, + 478 + ], + "type": "text", + "content": "2.3. Flow Matching and Stochastic Interpolants" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 486, + 298, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 486, + 298, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 486, + 298, + 714 + ], + "type": "text", + "content": "Flow Matching [44, 46, 54] has established Continuous Normalizing Flow (CNF) as a notable competitor to diffusion models [26, 78]. It enables simulation-free training of CNFs using an optimal transport conditional probability path with Ordinary Differential Equations (ODE), providing a more efficient alternative to traditional diffusion training and sampling methods [34, 81]. The versatility of Flow Matching has been showcased across various domains, including image [27, 29, 37, 44, 75], depth map [22], video [12], human motion [28], point cloud [94], and manifold data [7]. A more recent family of frameworks - Stochastic Interpolants [1, 47] has been proposed for unifying Flows and Diffusions. They suggested that Stochastic Interpolants can be generalized to align with various generative models, including Normalizing Flow [8], Diffusion Model [66, 80], Flow Matching [22, 29, 75], and Schrödinger Bridge [45, 85]. In this work, we develop our method within the stochastic interpolant framework and use the learned optimal transport paths (flows) to study the progression of styles across eras." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 211, + 370, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 211, + 370, + 224 + ], + "spans": [ + { + "bbox": [ + 314, + 211, + 370, + 224 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 232, + 556, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 556, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 556, + 399 + ], + "type": "text", + "content": "We aim to enable the matching of artworks across different time periods, despite the absence of ground-truth correspondences. Analyzing the temporal evolution of individual paintings is inherently infeasible, as artworks are static and finalized entities. Future stylistic transformations are highly unpredictable, even when considering subsequent works by the same artist, given that many artists undergo significant stylistic changes throughout their careers (e.g., Van Gogh or Pablo Picasso). Due to this absence of explicit ground-truth data pairs, we propose an alternative formulation of the problem. Specifically, we approach it as a distribution matching task, where the evolution of artistic styles over time is understood in terms of underlying probability distributions rather than isolated instances." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 400, + 556, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 556, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 556, + 544 + ], + "type": "text", + "content": "By modelling on the global distribution of artistic styles across time, we can establish meaningful correspondences spanning different centuries [77]. A principled framework for matching distributions is Optimal Transport (OT), which minimizes the transportation costs between distributions [83, 100]. To achieve this, we leverage implicit bridges and flow matching, which, upon training, inherently define an Ordinary Differential Equation (ODE) that aligns with the principles of optimal transport. This not only captures high-level stylistic transitions over time but also facilitates fine-grained alignment of individual artworks based on their placement within these temporally evolving distributions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 553, + 451, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 553, + 451, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 553, + 451, + 566 + ], + "type": "text", + "content": "3.1. Theoretical background" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 571, + 556, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 571, + 556, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 571, + 556, + 643 + ], + "type": "text", + "content": "Stochastic Interpolants In recent years, generative probabilistic models that transform noise into data have gained significant attention. These models are primarily represented by diffusion-based models [17, 66, 80, 81] and flow-based models [26, 44]. This transformation can be defined [1] by a time-dependent forward process with " + }, + { + "bbox": [ + 313, + 571, + 556, + 643 + ], + "type": "inline_equation", + "content": "t \\in [0,1]" + }, + { + "bbox": [ + 313, + 571, + 556, + 643 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 393, + 658, + 555, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 658, + 555, + 670 + ], + "spans": [ + { + "bbox": [ + 393, + 658, + 555, + 670 + ], + "type": "interline_equation", + "content": "x _ {t} = \\alpha_ {t} x _ {0} + \\sigma_ {t} x _ {1}, \\tag {1}", + "image_path": "c78f42d978c0d13dedde9448c0ffe87c1a6efcabebaa938c46fffcc4bcf7a0c5.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 558, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 558, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 558, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 558, + 715 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 677, + 558, + 715 + ], + "type": "text", + "content": " corresponds to data sampled from the data distribution " + }, + { + "bbox": [ + 313, + 677, + 558, + 715 + ], + "type": "inline_equation", + "content": "p_0(x)" + }, + { + "bbox": [ + 313, + 677, + 558, + 715 + ], + "type": "text", + "content": " for which we have empirical observations available and " + }, + { + "bbox": [ + 313, + 677, + 558, + 715 + ], + "type": "inline_equation", + "content": "x_1 \\sim \\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 313, + 677, + 558, + 715 + ], + "type": "text", + "content": " is a simple prior distribution." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5869" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 55, + 97, + 294, + 223 + ], + "blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "type": "text", + "content": "Algorithm 1 Transforming artworks from source age distribution to target age distribution" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 55, + 97, + 294, + 223 + ], + "lines": [ + { + "bbox": [ + 55, + 97, + 294, + 223 + ], + "spans": [ + { + "bbox": [ + 55, + 97, + 294, + 223 + ], + "type": "text", + "content": "Input: " + }, + { + "bbox": [ + 55, + 97, + 294, + 223 + ], + "type": "inline_equation", + "content": "x_0^{(a)}\\sim p_A(x)" + }, + { + "bbox": [ + 55, + 97, + 294, + 223 + ], + "type": "text", + "content": " - data samples from age A " + }, + { + "bbox": [ + 55, + 97, + 294, + 223 + ], + "type": "inline_equation", + "content": "y_{a}" + }, + { + "bbox": [ + 55, + 97, + 294, + 223 + ], + "type": "text", + "content": " - original chronological conditioning " + }, + { + "bbox": [ + 55, + 97, + 294, + 223 + ], + "type": "inline_equation", + "content": "y_{b}" + }, + { + "bbox": [ + 55, + 97, + 294, + 223 + ], + "type": "text", + "content": " -chronological conditioning for desired age " + }, + { + "bbox": [ + 55, + 97, + 294, + 223 + ], + "type": "inline_equation", + "content": "y_{b} = y_{a} + \\Delta y" + }, + { + "bbox": [ + 55, + 97, + 294, + 223 + ], + "type": "inline_equation", + "content": "v_{\\theta}^{(a)}" + }, + { + "bbox": [ + 55, + 97, + 294, + 223 + ], + "type": "text", + "content": " - model corresponds to age A, instantiated by " + }, + { + "bbox": [ + 55, + 97, + 294, + 223 + ], + "type": "inline_equation", + "content": "v_{\\theta}(x_t,t,y_a)" + }, + { + "bbox": [ + 55, + 97, + 294, + 223 + ], + "type": "inline_equation", + "content": "v_{\\theta}^{(b)}" + }, + { + "bbox": [ + 55, + 97, + 294, + 223 + ], + "type": "text", + "content": " - target model corresponds to age B, instantiated by " + }, + { + "bbox": [ + 55, + 97, + 294, + 223 + ], + "type": "inline_equation", + "content": "v_{\\theta}(x_t,t,y_b)" + }, + { + "bbox": [ + 55, + 97, + 294, + 223 + ], + "type": "inline_equation", + "content": "t_\\mathrm{th}\\in (0,1]" + }, + { + "bbox": [ + 55, + 97, + 294, + 223 + ], + "type": "text", + "content": " - the integration threshold \nOutput: " + }, + { + "bbox": [ + 55, + 97, + 294, + 223 + ], + "type": "inline_equation", + "content": "x_0^{(b)}" + }, + { + "bbox": [ + 55, + 97, + 294, + 223 + ], + "type": "text", + "content": " , the result in the target age data distribution B. 1: for each element " + }, + { + "bbox": [ + 55, + 97, + 294, + 223 + ], + "type": "inline_equation", + "content": "x_0^{(a)}\\sim p_A(x)" + }, + { + "bbox": [ + 55, + 97, + 294, + 223 + ], + "type": "text", + "content": " do 2: " + }, + { + "bbox": [ + 55, + 97, + 294, + 223 + ], + "type": "inline_equation", + "content": "\\begin{array}{rlr}{x_{\\mathrm{th}} = \\mathrm{ODESolve}(x_0^{(a)};v_{\\theta}^{(a)},0,t_{\\mathrm{th}})} & \\triangleright & \\mathrm{obtain~the~structured~noise~from}\\\\ {\\mathrm{source~age~data}} & {} & \\\\ {x_0^{(b)} = \\mathrm{ODESolve}(x_{\\mathrm{th}};v_{\\theta}^{(b)},t_{\\mathrm{th}},0)} & \\triangleright & \\mathrm{obtain~target~age~data~from~the}\\\\ {\\mathrm{structured~noise}} & {} & \\\\ {\\mathrm{return} x_0^{(b)}} & {} & \\end{array}" + }, + { + "bbox": [ + 55, + 97, + 294, + 223 + ], + "type": "text", + "content": " 3:" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 244, + 298, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 244, + 298, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 244, + 298, + 365 + ], + "type": "text", + "content": "The forward process is characterized by the hyperparameters " + }, + { + "bbox": [ + 55, + 244, + 298, + 365 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 55, + 244, + 298, + 365 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 244, + 298, + 365 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 55, + 244, + 298, + 365 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 55, + 244, + 298, + 365 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 55, + 244, + 298, + 365 + ], + "type": "text", + "content": " decreasing and " + }, + { + "bbox": [ + 55, + 244, + 298, + 365 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 55, + 244, + 298, + 365 + ], + "type": "text", + "content": " increasing as time " + }, + { + "bbox": [ + 55, + 244, + 298, + 365 + ], + "type": "inline_equation", + "content": "t\\in [0,1]" + }, + { + "bbox": [ + 55, + 244, + 298, + 365 + ], + "type": "text", + "content": " progresses, interpolating between data and noise. Furthermore, the boundary conditions are normally constrained as " + }, + { + "bbox": [ + 55, + 244, + 298, + 365 + ], + "type": "inline_equation", + "content": "\\alpha_0 = \\sigma_1 = 1" + }, + { + "bbox": [ + 55, + 244, + 298, + 365 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 244, + 298, + 365 + ], + "type": "inline_equation", + "content": "\\alpha_{1} = \\sigma_{0} = 0" + }, + { + "bbox": [ + 55, + 244, + 298, + 365 + ], + "type": "text", + "content": ", so that for " + }, + { + "bbox": [ + 55, + 244, + 298, + 365 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 55, + 244, + 298, + 365 + ], + "type": "text", + "content": " we have pure data and " + }, + { + "bbox": [ + 55, + 244, + 298, + 365 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 55, + 244, + 298, + 365 + ], + "type": "text", + "content": " pure noise. In particular, flow-based models, represented by SiT [47, 75], have attracted significant interest due to their faster training and inference compared to diffusion-based models. With the above-defined forward process, the velocity is defined as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 373, + 296, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 373, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 126, + 373, + 296, + 388 + ], + "type": "interline_equation", + "content": "v (x, t) = \\mathbb {E} [ \\dot {x} _ {t} | x _ {t} = x ], \\tag {2}", + "image_path": "f928e0ca428fd582faf8067e0c3f0884b36b5a3435a6be8cf8865bebcd21d987.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 396, + 296, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 492 + ], + "type": "text", + "content": "which generates the marginal probability distribution " + }, + { + "bbox": [ + 55, + 396, + 296, + 492 + ], + "type": "inline_equation", + "content": "p_t(x)" + }, + { + "bbox": [ + 55, + 396, + 296, + 492 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 396, + 296, + 492 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 55, + 396, + 296, + 492 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 55, + 396, + 296, + 492 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 396, + 296, + 492 + ], + "type": "text", + "content": " [1, 47]. This aligns with the distributions of the probability flow ODE [81] so that during inference we can generate samples by solving the Ordinary Differential Equation (ODE) backward in time starting from pure noise " + }, + { + "bbox": [ + 55, + 396, + 296, + 492 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 55, + 396, + 296, + 492 + ], + "type": "text", + "content": " to arrive at a sample " + }, + { + "bbox": [ + 55, + 396, + 296, + 492 + ], + "type": "inline_equation", + "content": "x\\sim p_0(x)" + }, + { + "bbox": [ + 55, + 396, + 296, + 492 + ], + "type": "text", + "content": ". Ma et al. [47] demonstrates that we can train a neural network " + }, + { + "bbox": [ + 55, + 396, + 296, + 492 + ], + "type": "inline_equation", + "content": "v_{\\theta}(x_{t},t)" + }, + { + "bbox": [ + 55, + 396, + 296, + 492 + ], + "type": "text", + "content": " to parametrize the velocity using the training objective:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 510, + 296, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 510, + 296, + 538 + ], + "spans": [ + { + "bbox": [ + 77, + 510, + 296, + 538 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {v}} (\\theta) = \\int_ {0} ^ {T} \\mathbb {E} [ \\| v _ {\\theta} (x _ {t}, t) - \\dot {\\alpha} _ {t} x _ {0} - \\dot {\\sigma} _ {t} x _ {1} \\| ^ {2} ] \\mathrm {d} t, \\tag {3}", + "image_path": "205f96e34fd157b436194a391be3c623abf74fb978085e26a3094777d69c6e46.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 545, + 297, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 297, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 297, + 593 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 55, + 545, + 297, + 593 + ], + "type": "inline_equation", + "content": "\\dot{\\alpha}_{t}" + }, + { + "bbox": [ + 55, + 545, + 297, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 545, + 297, + 593 + ], + "type": "inline_equation", + "content": "\\dot{\\sigma}_t" + }, + { + "bbox": [ + 55, + 545, + 297, + 593 + ], + "type": "text", + "content": " representing the time derivative of " + }, + { + "bbox": [ + 55, + 545, + 297, + 593 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 55, + 545, + 297, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 545, + 297, + 593 + ], + "type": "inline_equation", + "content": "\\sigma_t" + }, + { + "bbox": [ + 55, + 545, + 297, + 593 + ], + "type": "text", + "content": ", respectively. For simplicity and due to its relatively straight trajectories, we adopt the Linear Interpolant from [46, 47] with " + }, + { + "bbox": [ + 55, + 545, + 297, + 593 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = 1 - t" + }, + { + "bbox": [ + 55, + 545, + 297, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 545, + 297, + 593 + ], + "type": "inline_equation", + "content": "\\sigma_t = t" + }, + { + "bbox": [ + 55, + 545, + 297, + 593 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 600, + 212, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 212, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 212, + 613 + ], + "type": "text", + "content": "3.2. Bridging artworks over time" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": "We aim to establish meaningful correspondences between artworks originating from different historical periods. By leveraging and adapting the recent advances in generative modelling, we develop a framework that learns to transport dynamics across time-conditioned data distributions without requiring paired supervision. We begin by introducing the foundations of implicit diffusion bridges, before detailing how they can be applied to match artworks across time." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "x^{(a)}" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": " represent a data sample from data distribution " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": ". The goal is to identify the corresponding sample " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "x^{(b)}" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": " from a different data distribution " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": ". This problem is inherently ill-posed due to the absence of ground-truth data pairs. To address this, we seek to establish optimal transport between the underlying manifolds and construct Schrödinger bridges between the respective distributions [9, 40]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 159, + 556, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 159, + 556, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 159, + 556, + 258 + ], + "type": "text", + "content": "Dual Diffusion Implicit Bridges (DDIB) [83] demonstrate that such bridges can be effectively constructed using a pair of pre-trained image diffusion models. Notably, the Ordinary Differential Equations (ODEs) derived from trained diffusion models inherently exhibit optimal transport between the Gaussian noise prior and the corresponding data distributions [85]. Let ODESolve denote the mapping from an initial noise " + }, + { + "bbox": [ + 313, + 159, + 556, + 258 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 313, + 159, + 556, + 258 + ], + "type": "text", + "content": " to the data " + }, + { + "bbox": [ + 313, + 159, + 556, + 258 + ], + "type": "inline_equation", + "content": "x_0^{(a)}" + }, + { + "bbox": [ + 313, + 159, + 556, + 258 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 159, + 556, + 258 + ], + "type": "inline_equation", + "content": "v_{\\theta}^{(a)}" + }, + { + "bbox": [ + 313, + 159, + 556, + 258 + ], + "type": "text", + "content": " defined in Eq. (3):" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 325, + 273, + 555, + 301 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 273, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 325, + 273, + 555, + 301 + ], + "type": "interline_equation", + "content": "\\operatorname {O D E S o l v e} \\left(x _ {t}; v _ {\\theta} ^ {(a)}, 1, 0\\right) = x _ {1} + \\int_ {1} ^ {0} v _ {\\theta} ^ {(a)} \\left(x _ {t}, t\\right) d t. \\tag {4}", + "image_path": "ac14ee5fe0b3ac0077d32bf00216775a06e772d7e17e9d6d0317deaccfe2ed86.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 320, + 556, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 556, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 556, + 550 + ], + "type": "text", + "content": "Analogously, we can train ODESolve " + }, + { + "bbox": [ + 313, + 320, + 556, + 550 + ], + "type": "inline_equation", + "content": "(x_{t};v_{\\theta}^{(b)},1,0)" + }, + { + "bbox": [ + 313, + 320, + 556, + 550 + ], + "type": "text", + "content": " for mapping data between noise and " + }, + { + "bbox": [ + 313, + 320, + 556, + 550 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 320, + 556, + 550 + ], + "type": "text", + "content": ". Given these frameworks, the learned mappings between noise and data distributions, along with their corresponding reverse processes ODESolve " + }, + { + "bbox": [ + 313, + 320, + 556, + 550 + ], + "type": "inline_equation", + "content": "(x_{t};v_{\\theta},0,1)" + }, + { + "bbox": [ + 313, + 320, + 556, + 550 + ], + "type": "text", + "content": ", inherently constitute optimal transport pairs. As illustrated in the example in Fig. 2, we demonstrate this ability to transfer data between distributions " + }, + { + "bbox": [ + 313, + 320, + 556, + 550 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 320, + 556, + 550 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 320, + 556, + 550 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 320, + 556, + 550 + ], + "type": "text", + "content": " despite having no prior information about their inherent relationship. First, we use the trained ODE to map data distribution " + }, + { + "bbox": [ + 313, + 320, + 556, + 550 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 320, + 556, + 550 + ], + "type": "text", + "content": " to the standard Gaussian noise. A clear structure related to " + }, + { + "bbox": [ + 313, + 320, + 556, + 550 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 320, + 556, + 550 + ], + "type": "text", + "content": " emerges, highlighting the optimal noise-to-data pairs achieved through the ODE model's transport. Subsequently, another trained ODE transports this structured noise [67] to the distinctly different distribution " + }, + { + "bbox": [ + 313, + 320, + 556, + 550 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 320, + 556, + 550 + ], + "type": "text", + "content": ". In this setup, the mapping between distributions " + }, + { + "bbox": [ + 313, + 320, + 556, + 550 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 320, + 556, + 550 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 320, + 556, + 550 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 320, + 556, + 550 + ], + "type": "text", + "content": " remains largely preserved, even when using two independently trained ODEs that share no information. This indicates that the noise space might not be pure noise, but rather contains a certain structure as suggested in [67]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 552, + 555, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 555, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 555, + 577 + ], + "type": "text", + "content": "The independent " + }, + { + "bbox": [ + 313, + 552, + 555, + 577 + ], + "type": "inline_equation", + "content": "v_{\\theta}" + }, + { + "bbox": [ + 313, + 552, + 555, + 577 + ], + "type": "text", + "content": " can be achieved by training the same model with different conditions, namely" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 357, + 592, + 555, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 592, + 555, + 608 + ], + "spans": [ + { + "bbox": [ + 357, + 592, + 555, + 608 + ], + "type": "interline_equation", + "content": "v _ {\\theta} ^ {(a)} = v _ {\\theta} (\\cdot , t, y _ {a}), v _ {\\theta} ^ {(b)} = v _ {\\theta} (\\cdot , t, y _ {b}). \\tag {5}", + "image_path": "639d31cbb08e3c36b8dec4384cf97387b16883c4a1e1e9c41418e9cb3bb1108d.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "text", + "content": "Under our problem setup, the data " + }, + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "inline_equation", + "content": "x_0^{(a)}" + }, + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "text", + "content": " resembles a piece of artwork, and " + }, + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "inline_equation", + "content": "y_{a}" + }, + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "text", + "content": " represents the corresponding chronological information (creation year of the artwork). This parameterization allows a single model " + }, + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "inline_equation", + "content": "v_{\\theta}" + }, + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "text", + "content": " to learn mappings between arbitrary years " + }, + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "inline_equation", + "content": "y_{a} \\rightarrow y_{b}" + }, + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "text", + "content": " through conditioning, effectively replacing the need to train separate models for each temporal pair, as illustrated in Fig. 3." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5870" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 88, + 78, + 265, + 132 + ], + "blocks": [ + { + "bbox": [ + 88, + 78, + 265, + 132 + ], + "lines": [ + { + "bbox": [ + 88, + 78, + 265, + 132 + ], + "spans": [ + { + "bbox": [ + 88, + 78, + 265, + 132 + ], + "type": "image", + "image_path": "09cafccceb785dd5d8517a98e3fc4cc6c389f503ed50f671e1ca8d74e35e0fca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 137, + 196, + 148 + ], + "lines": [ + { + "bbox": [ + 156, + 137, + 196, + 148 + ], + "spans": [ + { + "bbox": [ + 156, + 137, + 196, + 148 + ], + "type": "text", + "content": "(a) Training" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 88, + 152, + 265, + 204 + ], + "blocks": [ + { + "bbox": [ + 88, + 152, + 265, + 204 + ], + "lines": [ + { + "bbox": [ + 88, + 152, + 265, + 204 + ], + "spans": [ + { + "bbox": [ + 88, + 152, + 265, + 204 + ], + "type": "image", + "image_path": "558d5811cd82feb59d0b1a9e314e5510c152e825cd36dceb9f6f3e4ea5df7ba4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 211, + 197, + 220 + ], + "lines": [ + { + "bbox": [ + 154, + 211, + 197, + 220 + ], + "spans": [ + { + "bbox": [ + 154, + 211, + 197, + 220 + ], + "type": "text", + "content": "(b) Inference" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 318, + 296, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 296, + 343 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 296, + 343 + ], + "type": "text", + "content": "Training objective. Similar to Eq. (3), we define our final objective to optimize " + }, + { + "bbox": [ + 55, + 318, + 296, + 343 + ], + "type": "inline_equation", + "content": "v_{\\theta}" + }, + { + "bbox": [ + 55, + 318, + 296, + 343 + ], + "type": "text", + "content": " with the year condition " + }, + { + "bbox": [ + 55, + 318, + 296, + 343 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 318, + 296, + 343 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 350, + 296, + 379 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 350, + 296, + 379 + ], + "spans": [ + { + "bbox": [ + 69, + 350, + 296, + 379 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\theta) = \\int_ {0} ^ {T} \\mathbb {E} \\left[ \\| v _ {\\theta} \\left(x _ {t}, t, y\\right) - \\dot {\\alpha} _ {t} x _ {0} - \\dot {\\sigma} _ {t} x _ {1} \\| ^ {2} \\right] \\mathrm {d} t, \\tag {6}", + "image_path": "da725dc7db2d0e7f45a78ad3548c23fe36bd63cf4998ac84abe45cf13139b128.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 386, + 297, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 386, + 297, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 386, + 297, + 434 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 386, + 297, + 434 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 386, + 297, + 434 + ], + "type": "text", + "content": " is the backbone parameterized by a 1D transformer. We detail the architecture of the model in the Appendix A. We operate in a compact 1D-style space, as pixel space retains excessively low-level details that contain redundancy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 441, + 123, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 123, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 123, + 453 + ], + "type": "text", + "content": "3.3. Inference" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 459, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 296, + 594 + ], + "type": "text", + "content": "With a trained model that can generate the style latents sufficiently well, we can start the matching process. The model defines a vector field " + }, + { + "bbox": [ + 55, + 459, + 296, + 594 + ], + "type": "inline_equation", + "content": "v_{\\theta}(\\cdot ,t,y)" + }, + { + "bbox": [ + 55, + 459, + 296, + 594 + ], + "type": "text", + "content": " at any given " + }, + { + "bbox": [ + 55, + 459, + 296, + 594 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 459, + 296, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 459, + 296, + 594 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 459, + 296, + 594 + ], + "type": "text", + "content": ". As explained in Algorithm 1, we first transform the style representation from " + }, + { + "bbox": [ + 55, + 459, + 296, + 594 + ], + "type": "inline_equation", + "content": "x_0^{(a)}" + }, + { + "bbox": [ + 55, + 459, + 296, + 594 + ], + "type": "text", + "content": " to the noisy representation " + }, + { + "bbox": [ + 55, + 459, + 296, + 594 + ], + "type": "inline_equation", + "content": "x_{t_{\\mathrm{th}}}" + }, + { + "bbox": [ + 55, + 459, + 296, + 594 + ], + "type": "text", + "content": " using the learned model " + }, + { + "bbox": [ + 55, + 459, + 296, + 594 + ], + "type": "inline_equation", + "content": "v_{\\theta}" + }, + { + "bbox": [ + 55, + 459, + 296, + 594 + ], + "type": "text", + "content": ". The parameter " + }, + { + "bbox": [ + 55, + 459, + 296, + 594 + ], + "type": "inline_equation", + "content": "t_\\mathrm{th} > 0" + }, + { + "bbox": [ + 55, + 459, + 296, + 594 + ], + "type": "text", + "content": " denotes the integration threshold; in practice, we typically set " + }, + { + "bbox": [ + 55, + 459, + 296, + 594 + ], + "type": "inline_equation", + "content": "t_\\mathrm{th} = 1" + }, + { + "bbox": [ + 55, + 459, + 296, + 594 + ], + "type": "text", + "content": ", corresponding to a full mapping from data space to the structured noise space. Next, we push forward the " + }, + { + "bbox": [ + 55, + 459, + 296, + 594 + ], + "type": "inline_equation", + "content": "x_{t_\\mathrm{th}}" + }, + { + "bbox": [ + 55, + 459, + 296, + 594 + ], + "type": "text", + "content": " back to " + }, + { + "bbox": [ + 55, + 459, + 296, + 594 + ], + "type": "inline_equation", + "content": "x_0^{(b)}" + }, + { + "bbox": [ + 55, + 459, + 296, + 594 + ], + "type": "text", + "content": " using the vector field produced by the same model but with a different chronological condition " + }, + { + "bbox": [ + 55, + 459, + 296, + 594 + ], + "type": "inline_equation", + "content": "v_{\\theta}(x_t,t,y_b)" + }, + { + "bbox": [ + 55, + 459, + 296, + 594 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "Fig. 4 provides a conceptual illustration of our method using a 2D synthetic toy example, where trajectories depict the temporal evolution of data points. We construct temporally coherent toy data pairs to enable access to ground-truth data pairs. A single model is trained to perform conditional mappings across five distinct input conditions (indicated by the x-axis). Although attributes (indicated by color) gradually blend or evolve, the underlying optimal transport plan is preserved. Notably, the generated samples consistently align with the correct target modes, despite the absence of" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 345, + 79, + 521, + 194 + ], + "blocks": [ + { + "bbox": [ + 55, + 231, + 296, + 298 + ], + "lines": [ + { + "bbox": [ + 55, + 231, + 296, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 296, + 298 + ], + "type": "text", + "content": "Figure 3. Approach overview. " + }, + { + "bbox": [ + 55, + 231, + 296, + 298 + ], + "type": "inline_equation", + "content": "a)" + }, + { + "bbox": [ + 55, + 231, + 296, + 298 + ], + "type": "text", + "content": ". Training a single " + }, + { + "bbox": [ + 55, + 231, + 296, + 298 + ], + "type": "inline_equation", + "content": "v_{\\theta}" + }, + { + "bbox": [ + 55, + 231, + 296, + 298 + ], + "type": "text", + "content": " between style embeddings and structured noise conditioned on " + }, + { + "bbox": [ + 55, + 231, + 296, + 298 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 231, + 296, + 298 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 231, + 296, + 298 + ], + "type": "inline_equation", + "content": "b)" + }, + { + "bbox": [ + 55, + 231, + 296, + 298 + ], + "type": "text", + "content": ". During inference, we can guide the ODE forward and backward direction with different chronological conditions " + }, + { + "bbox": [ + 55, + 231, + 296, + 298 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 231, + 296, + 298 + ], + "type": "text", + "content": " between the style and latent noise space. By using different " + }, + { + "bbox": [ + 55, + 231, + 296, + 298 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 231, + 296, + 298 + ], + "type": "text", + "content": ", the style embedding can be transformed to a different style manifold w.r.t. to historical patterns." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 345, + 79, + 521, + 194 + ], + "lines": [ + { + "bbox": [ + 345, + 79, + 521, + 194 + ], + "spans": [ + { + "bbox": [ + 345, + 79, + 521, + 194 + ], + "type": "image", + "image_path": "88e743fee0b07984407265cad180dfda0c28dddeb321d05ba974f66515708f8f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 206, + 555, + 262 + ], + "lines": [ + { + "bbox": [ + 313, + 206, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 206, + 555, + 262 + ], + "type": "text", + "content": "Figure 4. Illustrational example with ground truth correspondences. Dots sharing the same color are permuted versions along the x-axis. In the figure, colored round dots represent the ground truth samples. purple stars (starting from the 2nd column) indicate the warped samples originating from cyan stars (leftmost)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 273, + 555, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 555, + 298 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 555, + 298 + ], + "type": "text", + "content": "explicit mode supervision, demonstrating robust structural consistency across varying style dynamics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 306, + 437, + 318 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 437, + 318 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 437, + 318 + ], + "type": "text", + "content": "3.4. Dataset Construction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 324, + 555, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 555, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 555, + 443 + ], + "type": "text", + "content": "The curated dataset is a large-scale dataset comprising over 650,000 artworks spanning a wide range of artistic styles and mediums. As discussed in Sec. 2.2, existing datasets for art lack unified meta-information, have limited access, or are small in size. In contrast, our dataset provides standardized and cleaned metadata aggregated from multiple sources. For example, while datasets like WikiArt [84] may include creation years within filenames, such information is often unstructured. Hence, we normalize these metadata to facilitate chronological modeling and future research." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 445, + 556, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 556, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 556, + 578 + ], + "type": "text", + "content": "The available metadata includes title, artist, style, year of creation (Most of the artworks originate from years ranging from 1500 to 2021), media (e.g., oil on canvas, etching, acrylic), geographical location, and image URLs. The images are provided as URLs to public museums and public art gallery websites or indexed to public datasets, where they are under the original license and copyright. An overview of the dataset is provided in Fig. 1. For experimental purposes, we used a " + }, + { + "bbox": [ + 313, + 445, + 556, + 578 + ], + "type": "inline_equation", + "content": "6/2/2" + }, + { + "bbox": [ + 313, + 445, + 556, + 578 + ], + "type": "text", + "content": " split for training, validation, and testing, randomly sampled from centuries without selecting styles or artists explicitly to maintain the underlying data distribution." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "content": "Deduplication. Due to the diverse sources of art pieces, duplicate samples are inevitable. While this redundancy provides additional meta-information from different sources, it results in variations of potential duplicates. To address this, we deduplicated our dataset following Somepalli et al. [79], using SSCD [59] embeddings. Entries with a cosine similarity above 0.85 were merged, consolidating their metadata, e.g., retaining the English version when multiple languages were available. This process reduced the dataset from approximately 950,000 to 650,000 art pieces, encompassing" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5871" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 293, + 227 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 293, + 227 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 293, + 227 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 293, + 227 + ], + "type": "image", + "image_path": "477f9da8baafc2a11620e86b973189994f7e1b72fddfa7b05379cc9cb23380ac.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 235, + 296, + 281 + ], + "lines": [ + { + "bbox": [ + 55, + 235, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 235, + 296, + 281 + ], + "type": "text", + "content": "Figure 5. Stylistic flows over time. We visualize distribution shifts for " + }, + { + "bbox": [ + 55, + 235, + 296, + 281 + ], + "type": "inline_equation", + "content": "\\Delta y = 0, 20, 40, 80" + }, + { + "bbox": [ + 55, + 235, + 296, + 281 + ], + "type": "text", + "content": " years. Blue quivers indicate the dominant movements of art pieces into the next time frame and visualize how style distribution flows across time." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 291, + 296, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 291, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 296, + 327 + ], + "type": "text", + "content": "5,053 artists, 128 styles, 7,908 media, and 1,442 geographical locations. Due to the space limitation, we defer the more details of the dataset in the Appendix D." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 338, + 132, + 352 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 132, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 132, + 352 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 358, + 227, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 227, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 227, + 371 + ], + "type": "text", + "content": "4.1. Choice of Style Representations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 376, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 296, + 495 + ], + "type": "text", + "content": "The concept of style is inherently complex and multifaceted, encompassing aspects such as color, material, atmosphere, design, and structure. From a human perspective [39, 72], defining style is subjective and often varies depending on context. Similarly, in computer vision, style is represented through diverse latent spaces depending on the task. For instance, works on style classification and retrieval [11, 33] focus on different representations than those used in style transfer [20, 95] or in diffusion-based generation [16, 61, 87]. In both domains, style remains fluid and context-sensitive." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 496, + 296, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 296, + 591 + ], + "type": "text", + "content": "To accommodate this ambiguity, our model is designed to be agnostic to the specific choice of style space. In our experiments, we primarily focus on two feature spaces: CLIP [62] and CSD [79]. As shown in Table S3, CSD excels in style-based retrieval, whereas CLIP performs slightly better in year-based retrieval and can be decoded to images with unCLIP [65] (see Figure S15). We refer readers to Appendix A for a detailed comparison of all candidates." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 600, + 179, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 179, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 179, + 613 + ], + "type": "text", + "content": "4.2. Experimental Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "Tracing stylistic flows across time. We visualize the temporal flows of style distributions in Figure 5, showing how our model progressively aligns embeddings of early-20th-century artworks with those from later periods. Each subplot displays kernel density estimates (KDEs) [57] in the CSD space for three sets of samples: the original early-century inputs transported by our model (green), the real artworks from the corresponding future time window (orange), and" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 548, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 548, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 548, + 84 + ], + "type": "text", + "content": "the fixed target distribution of late-20th-century art (blue)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 85, + 555, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 85, + 555, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 85, + 555, + 263 + ], + "type": "text", + "content": "At " + }, + { + "bbox": [ + 313, + 85, + 555, + 263 + ], + "type": "inline_equation", + "content": "\\Delta t = 0" + }, + { + "bbox": [ + 313, + 85, + 555, + 263 + ], + "type": "text", + "content": ", the transformed distribution (green) coincides with the original input, as expected. As " + }, + { + "bbox": [ + 313, + 85, + 555, + 263 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 313, + 85, + 555, + 263 + ], + "type": "text", + "content": " increases to 20, 40, and 80 years, the green distribution increasingly aligns with the orange ground-truth distribution for each target decade, despite the absence of sample-level correspondences. This alignment is not exact, but meaningful: the transformed samples retain certain semantic traces of their origin, resulting in a distribution that blends temporal stylistic adaptation with content continuity. Such divergence offers insight into how artistic style may evolve from a given source, even in ways not fully captured by existing data. Eventually, the green distribution approaches the latecentury manifold (blue), indicating that our model learns a smooth and coherent temporal flow (denoted by the blue quivers) from a distribution perspective." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 264, + 554, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 554, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 554, + 348 + ], + "type": "text", + "content": "To bridge population-level dynamics with sample-level behavior, we track a single artwork—Berlin Street Scene (1913)—through time. Its unCLIP renderings reflect plausible stylistic changes in line with historical trends, while preserving core semantics. This illustrates the model's ability to capture stylistic evolution across distributions and within individual samples, which we explore next." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 363, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 555, + 495 + ], + "type": "text", + "content": "Transforming individual artworks through time. We present visual examples demonstrating how individual art pieces evolve across time by transferring their semantic embedding in the CLIP [62] space and visualizing using the un-CLIP model [66]. In Figure 6, we show the temporal trajectories of canonical works such as Les Demoiselles d'Avignon (1907, Cubism) and Landscape near Chatou (1904, Fauvism). This task is particularly challenging as these pieces are less realistic, making semantic capture more difficult. Despite this, we preserve artistic intent and content over time by operating in a compact semantic space." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 495, + 556, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 556, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 556, + 603 + ], + "type": "text", + "content": "Additionally, we compare our approach with image editing methods in pixel space [49, 96] in Figure 7. These methods fail to achieve faithful semantic transformations tied to specific periods; instead, they apply superficial style transfers that ignore historical context. In contrast, our method enables long-range interpolation through time. We further analyze the style distributions of both unconditionally generated embeddings and mapped embeddings in Appendix C.2, showing that both closely follow the stylistic distributions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": "Content preservation. Our model captures how artistic style changes over time and how such stylistic evolution influences content, including shifts in scene composition and object substitution, e.g., horses versus motorcycles in Figure 7. Because style and content are often deeply entangled, meaningful temporal transformation requires flexibility. When the threshold parameter " + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{th}}" + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": " is set too low (Section 3.3), the model is overly constrained to the original input, limit" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5872" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 525, + 239 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 525, + 239 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 525, + 239 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 525, + 239 + ], + "type": "image", + "image_path": "6b45fb4c8f449c4767de1be949786f7469ea46bda4c95f919999d59315d9b852.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 240, + 555, + 264 + ], + "lines": [ + { + "bbox": [ + 55, + 240, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 555, + 264 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison of editing methods across different historical styles. Each row shows results from a different method. The 1st and 5th columns show original artworks (Les Demoiselles d'Avignon - 1907 and Landscape near Chatou - 1904)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 132, + 274, + 479, + 339 + ], + "blocks": [ + { + "bbox": [ + 132, + 274, + 479, + 339 + ], + "lines": [ + { + "bbox": [ + 132, + 274, + 479, + 339 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 479, + 339 + ], + "type": "table", + "html": "
MethodRococoImpressionismRealism
R@1R@5δ↓τ(%)↑R@1R@5δ↓τ(%)↑R@1R@5δ↓τ(%)↑
RF-inversion [68]5.69.30.69083.710.526.30.65876.82.46.50.74966.3
LEDDS ++ [3]12.623.00.52879.415.430.40.58569.69.720.50.50964.9
SDEdit [49]6.315.80.57245.65.313.40.60534.53.712.40.61245.6
SDEdit (fine-tuned)37.942.80.26564.734.147.30.27247.227.538.40.23657.3
Ours72.585.10.02484.865.076.30.02475.858.569.30.02568.6
", + "image_path": "eacee91bd1a4d9c733979f3e0448a9ca3a9d45fe10d043964830bc7534a5edbc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 347, + 555, + 392 + ], + "lines": [ + { + "bbox": [ + 54, + 347, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 54, + 347, + 555, + 392 + ], + "type": "text", + "content": "Table 1. Quality of the time matching when transferring artworks over time. We sampled a compact subset of a hundred samples for each style to calculate the metrics. We transfer clusters of artworks to randomly selected years within the range of " + }, + { + "bbox": [ + 54, + 347, + 555, + 392 + ], + "type": "inline_equation", + "content": "[-100, 100]" + }, + { + "bbox": [ + 54, + 347, + 555, + 392 + ], + "type": "text", + "content": " and compute the mean metrics. For " + }, + { + "bbox": [ + 54, + 347, + 555, + 392 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 54, + 347, + 555, + 392 + ], + "type": "text", + "content": ", we used the nearest " + }, + { + "bbox": [ + 54, + 347, + 555, + 392 + ], + "type": "inline_equation", + "content": "nns[:25]" + }, + { + "bbox": [ + 54, + 347, + 555, + 392 + ], + "type": "text", + "content": " neighbors as positive samples and further " + }, + { + "bbox": [ + 54, + 347, + 555, + 392 + ], + "type": "inline_equation", + "content": "nns[50:75]" + }, + { + "bbox": [ + 54, + 347, + 555, + 392 + ], + "type": "text", + "content": " as the negative samples for each chosen anchor, where " + }, + { + "bbox": [ + 54, + 347, + 555, + 392 + ], + "type": "inline_equation", + "content": "nns" + }, + { + "bbox": [ + 54, + 347, + 555, + 392 + ], + "type": "text", + "content": " denotes its nearest neighbors. See extended table in Table S4." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 403, + 294, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 403, + 294, + 439 + ], + "spans": [ + { + "bbox": [ + 54, + 403, + 294, + 439 + ], + "type": "text", + "content": "ing stylistic adaptation. This results in hybrid outputs, such as a Steam Horse locomotive, that blend past and future characteristics without fully transitioning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 440, + 295, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 440, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 54, + 440, + 295, + 559 + ], + "type": "text", + "content": "Content deviation is a broader challenge across models. Most methods face a trade-off between low-level fidelity and high-level semantic consistency. Pixel-focused techniques [3, 49, 67] often struggle to produce meaningful temporal shifts even with adjusted guidance scales, sometimes leading to failures like buildings morphing into people, as seen in Figure 6. In contrast, our model preserves conceptual content while enabling coherent, temporally grounded transformations, striking a balance between semantic stability and stylistic progression." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 581, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 581, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 54, + 581, + 295, + 664 + ], + "type": "text", + "content": "Quantitative evaluation of time travel. Assessing the quality of temporal transitions in artworks is challenging due to the absence of ground-truth correspondences. However, semi-ground-truth evidence exists: artworks created close in time and close in style are more likely to evolve together rather than disperse. Based on this intuition, we utilized three key metrics:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 55, + 665, + 296, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 701 + ], + "type": "text", + "content": "- Recall@k assesses alignment with the target era by checking whether time-traveled samples retrieve temporally appropriate neighbors in style space." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "type": "text", + "content": "- Compactness " + }, + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "type": "text", + "content": " measures how coherent the transformed" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 321, + 403, + 553, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 403, + 553, + 456 + ], + "spans": [ + { + "bbox": [ + 321, + 403, + 553, + 456 + ], + "type": "text", + "content": "samples are by comparing their variance to that of random samples in the target year: " + }, + { + "bbox": [ + 321, + 403, + 553, + 456 + ], + "type": "inline_equation", + "content": "\\frac{\\sigma_{\\text{samples}}^2}{\\sigma_{\\text{target}}^2}" + }, + { + "bbox": [ + 321, + 403, + 553, + 456 + ], + "type": "text", + "content": ". A good transformation should preserve this coherence, as source samples are initially similar in style and time, resulting in lower " + }, + { + "bbox": [ + 321, + 403, + 553, + 456 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 321, + 403, + 553, + 456 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 456, + 555, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 456, + 555, + 515 + ], + "spans": [ + { + "bbox": [ + 314, + 456, + 555, + 515 + ], + "type": "text", + "content": "- Triplet consistency " + }, + { + "bbox": [ + 314, + 456, + 555, + 515 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 314, + 456, + 555, + 515 + ], + "type": "text", + "content": " is adapted from triplet loss [23, 73]. It evaluates whether the relative relation among anchor, positive, and negative samples is preserved posttransformation, reflecting the model's ability to maintain local structure." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 519, + 556, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 519, + 556, + 651 + ], + "spans": [ + { + "bbox": [ + 312, + 519, + 556, + 651 + ], + "type": "text", + "content": "We evaluate on subsets of 100 samples by selecting localized neighborhoods of artwork within several style labels from the test split. These subsets serve as the starting samples for time mapping. Due to space constraints, we present results for the few most prominent artistic styles in Table 1, though our analysis extends across a much broader range. Since most editing baselines operate directly in pixel space, we perform an additional step of encoding their generated outputs back into the style space to enable a fair comparison, which allows us to compute nearest neighbor recall and compactness in a shared representation space." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "As shown in Table 1 (we offer more analysis in Table S4 in the Appendix), conventional pixel-based image-editing methods [3, 49, 67] fail to achieve both objectives: they neither align generated samples with the appropriate data manifold, reflected by the lower Recall@ " + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": ", nor preserve" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5873" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 77, + 272, + 228 + ], + "blocks": [ + { + "bbox": [ + 77, + 77, + 272, + 228 + ], + "lines": [ + { + "bbox": [ + 77, + 77, + 272, + 228 + ], + "spans": [ + { + "bbox": [ + 77, + 77, + 272, + 228 + ], + "type": "image", + "image_path": "797afe895a26f3af82ffb8e307eaddb62531a04aa20e3119afe6d0d11f9e2f06.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 233, + 297, + 289 + ], + "lines": [ + { + "bbox": [ + 55, + 233, + 297, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 233, + 297, + 289 + ], + "type": "text", + "content": "Figure 7. Flow preserves content semantics across time. We transform a motorcyclist to the year 1800 with increasing " + }, + { + "bbox": [ + 55, + 233, + 297, + 289 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{th}}" + }, + { + "bbox": [ + 55, + 233, + 297, + 289 + ], + "type": "text", + "content": " (left to right, top row). Low " + }, + { + "bbox": [ + 55, + 233, + 297, + 289 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{th}}" + }, + { + "bbox": [ + 55, + 233, + 297, + 289 + ], + "type": "text", + "content": " leads to incorrect mappings due to limited editing flexibility, while higher values retain semantic identity (\"man on a horse\") with correct stylistic traits." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 300, + 295, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 300, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 54, + 300, + 295, + 396 + ], + "type": "text", + "content": "local structural coherence, indicated by the higher " + }, + { + "bbox": [ + 54, + 300, + 295, + 396 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 54, + 300, + 295, + 396 + ], + "type": "text", + "content": "). This is mainly due to their reliance on pixel-level cues, which are insufficient for modeling more profound, semantic style shifts. For instance, RF-Inversion achieves high fidelity editing [67] but leads to minimal pixel shifts, evidenced by the low recall. We address these issues by operating in a compact latent space and focusing on semantic information to better align with historical stylistic flows." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 397, + 297, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 397, + 297, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 297, + 588 + ], + "type": "text", + "content": "Since our method is trained exclusively on the curated dataset, a distributional mismatch may arise when comparing against baseline models such as Stable Diffusion, which are pretrained on broader, general-domain datasets. For a fair comparison, we also fine-tuned the base Stable Diffusion model on our dataset until convergence. Specifically, we adapt the SD 1.5 model using chronologically conditioned prompts of the form \"A painting in {year} style\". This finetuning procedure improves both the synthesis FID in Table 2 and retrieval-based matching performance in Table 1, confirming our dataset's value. However, applying SDEdit remains suboptimal for our task, as the method struggles to perform temporally faithful transformations. This performance gap persists across all evaluated artistic styles, where our method consistently yields more coherent and historically aligned style transitions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "content": "Comparison to synthesis baselines. As previously discussed, for visualization purposes, we project samples into the CLIP [62] embedding space and render them using the unCLIP model [66]. This enables quantitative evaluation using the standard Frechet Inception Distance (FID) [25], which assesses both the realism of the generated outputs and their fidelity to the target temporal style distribution." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "We compared our method against three Latent Diffusion Models (LDM) variants: Stable Diffusion 1.5 [66], Stable" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 71, + 553, + 144 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 553, + 144 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 553, + 144 + ], + "type": "table", + "html": "
ModelFID↓conditioned on century
16th17th18th19th20th
SD 1.5 [66]131.1▼40.4118.9▼37.1103.5▼17.2149.9▼77.9180.4▼101.5
SD 2.1214.4▼123.7147.1▼65.3134.2▼47.9170.2▼98.2167.1▼88.2
SDXL [60]186.6▼95.9142.4▼60.6143.4▼57.1114.8▼42.8146.1▼67.2
FLUX.1-schnell [38]199.2▼108.5142.2▼60.4137.3▼51.0122.8▼50.8126.1▼47.2
SD 1.5 (fine-tuned)111.7▼2196.4▼14.688.8▼2.596.4▼24.4150.7▼71.8
Ours90.781.886.372.078.9
", + "image_path": "ae278c436f6f3fe54e75e246ef8907cc6739900017a4492ca3e3db4f3de4e629.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 152, + 555, + 208 + ], + "lines": [ + { + "bbox": [ + 313, + 152, + 555, + 208 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 555, + 208 + ], + "type": "text", + "content": "Table 2. Feature synthesis across centuries. Evaluation metrics are computed on 1,000 samples per century. While Stable Diffusion and FLUX rely on text prompts for year conditioning, our model conditions directly on this information. Our method exhibits stronger alignment with ground truth distributions." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 215, + 556, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 215, + 556, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 215, + 556, + 456 + ], + "type": "text", + "content": "Diffusion 2.1 [66], and SDXL [60], as well as the flow-matching based FLUX model [38]. We also compare against the fine-tuned Stable Diffusion 1.5, with results summarized in Table 2. SDXL [60] has demonstrated strong artistic understanding and has become the foundation for many recent stylistic image generation methods [16, 24, 76, 87]. While SDXL performs well on certain styles, such as Renaissance (16th century) and Cubism (20th century) in the Appendix Figure S13, it struggles to capture stylistic distributions comprehensively. FLUX [38], despite its training scale and exhibiting strong performance across text-to-image tasks, also struggles to capture temporal style distributions effectively (see Figure S14). In contrast, our model, trained on the proposed dataset, better captures the diversity of artistic styles. Our approach achieves significantly lower FID scores than standard text-conditioned models, demonstrating better alignment with the underlying feature distribution of different art periods. Furthermore, it provides more consistent stylistic alignment across centuries. Further rendering results for all models are available in the Appendix." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 466, + 388, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 388, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 388, + 479 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 486, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 582 + ], + "type": "text", + "content": "Studying the evolution of art is challenging due to the lack of ground-truth correspondences between artworks from different periods. We present a framework that models stylistic progression using stochastic interpolants and dual diffusion implicit bridges, enabling distribution-level alignment without paired data. By formulating art evolution as optimal transport in a learned style space, our method captures continuous temporal transitions in artistic style." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": "A central contribution is a dataset of over 650,000 artworks spanning 500 years, which we use to demonstrate that our model captures long-range stylistic shifts. Beyond modeling distributions, our method also enables tracing the stylistic origins and plausible futures of individual artworks, allowing for fine-grained longitudinal analysis. Our work explores a less-studied direction in computer vision: using generative models to investigate the evolution of art and support cultural and historical analysis. We believe our approach offers a foundation for future interdisciplinary work at the intersection of machine learning and cultural analytics." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5874" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 212 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 212 + ], + "type": "text", + "content": "This project was supported by Bayer AG, the Federal Ministry for Economic Affairs and Energy within the project \"NXT GEN AI METHODS - Generative Methoden für Perzeption, Prädiktion und Planung\", the project \"Genius-Robot\" (01IS24083) funded by the Federal Ministry of Research, Technology and Space (BMFTR), and the bidt project KLIMA-MEMES. The authors gratefully acknowledge the Gauss Center for Supercomputing for providing compute through the NIC on JUWELS/JUPITER at JSC and the HPC resources supplied by the NHR @FAU Erlangen." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 222, + 115, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 222, + 115, + 235 + ], + "spans": [ + { + "bbox": [ + 56, + 222, + 115, + 235 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 241, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 241, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 241, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 66, + 241, + 296, + 274 + ], + "type": "text", + "content": "[1] Michael S Albergo, Nicholas M Boffi, and Eric Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions. arXiv, 2023. 1, 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 276, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 276, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 66, + 276, + 296, + 319 + ], + "type": "text", + "content": "[2] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In ICCV, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 321, + 296, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 321, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 66, + 321, + 296, + 386 + ], + "type": "text", + "content": "[3] Manuel Brack, Felix Friedrich, Katharia Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. Ledits++: Limitless image editing using text-to-image models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8861-8870, 2024. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 388, + 296, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 388, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 66, + 388, + 296, + 441 + ], + "type": "text", + "content": "[4] Levin Brinkmann, Deniz Gezerli, KV Kleist, Thomas F Müller, Iyad Rahwan, and Niccolò Pescetelli. Hybrid social learning in human-algorithm cultural transmission. Philosophical Transactions of the Royal Society A, 380(2227), 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 444, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 444, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 66, + 444, + 296, + 487 + ], + "type": "text", + "content": "[5] Eric Ming Chen, Jin Sun, Apoorv Khandelwal, Dani Lischinski, Noah Snavely, and Hadar Averbuch-Elor. What's in a decade? transforming faces through time. In Computer Graphics Forum, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 488, + 296, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 488, + 296, + 543 + ], + "spans": [ + { + "bbox": [ + 66, + 488, + 296, + 543 + ], + "type": "text", + "content": "[6] Haibo Chen, Lei Zhao, Zhizhong Wang, Huiming Zhang, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Artistic style transfer with internal-external learning and contrastive learning. Advances in Neural Information Processing Systems (NeurIPS), 34, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 544, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 544, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 66, + 544, + 296, + 567 + ], + "type": "text", + "content": "[7] Ricky TQ Chen and Yaron Lipman. Riemannian flow matching on general geometries. arXiv, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 66, + 568, + 296, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 568, + 296, + 600 + ], + "spans": [ + { + "bbox": [ + 66, + 568, + 296, + 600 + ], + "type": "text", + "content": "[8] Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. NeurIPS, 2018. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 66, + 601, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 601, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 66, + 601, + 296, + 645 + ], + "type": "text", + "content": "[9] Yongxin Chen, Tryphon T Georgiou, and Michele Pavon. On the relation between optimal transport and schrödinger bridges: A stochastic control viewpoint. Journal of Optimization Theory and Applications, 169:671-691, 2016. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 647, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 296, + 689 + ], + "type": "text", + "content": "[10] Ceren Comert, Murat Özbayoglu, and Cosku Kasnakoğlu. Painter prediction from artworks with transfer learning. In 2021 7th International Conference on Mechatronics and Robotics Engineering (ICMRE). IEEE, 2021. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 691, + 296, + 713 + ], + "type": "text", + "content": "[11] Elliot Crowley and Andrew Zisserman. The state of the art: Object retrieval in paintings using discriminative regions. In" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 339, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 555, + 95 + ], + "type": "text", + "content": "Proceedings of the British Machine Vision Conference 2014. British Machine Vision Association, 2014. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 96, + 555, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 96, + 555, + 126 + ], + "spans": [ + { + "bbox": [ + 321, + 96, + 555, + 126 + ], + "type": "text", + "content": "[12] Aram Davtyan, Sepehr Sameni, and Paolo Favaro. Randomized conditional flow matching for video prediction. arXiv, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 129, + 555, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 129, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 321, + 129, + 555, + 172 + ], + "type": "text", + "content": "[13] Valentin De Bortoli, James Thornton, Jeremy Heng, and Arnaud Doucet. Diffusion schrödinger bridge with applications to score-based generative modeling. Advances in Neural Information Processing Systems, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 173, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 173, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 321, + 173, + 553, + 205 + ], + "type": "text", + "content": "[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, 2009. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 206, + 555, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 206, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 321, + 206, + 555, + 248 + ], + "type": "text", + "content": "[15] Yingying Deng, Fan Tang, Weiming Dong, Chongyang Ma, Feiyue Huang, Oliver Deussen, and Changsheng Xu. Exploring the representativity of art paintings. IEEE Transactions on Multimedia, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 251, + 555, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 251, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 321, + 251, + 555, + 282 + ], + "type": "text", + "content": "[16] Yarden Frenkel, Yael Vinker, Ariel Shamir, and Daniel Cohen-Or. Implicit style-content separation using b-lora. In ECCV, 2024. 6, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 284, + 555, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 284, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 321, + 284, + 555, + 326 + ], + "type": "text", + "content": "[17] Michael Fuest, Pingchuan Ma, Ming Gui, Johannes S Fischer, Vincent Tao Hu, and Bjorn Ommer. Diffusion models and representation learning: A survey. arXiv preprint arXiv:2407.00783, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 327, + 555, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 327, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 321, + 327, + 555, + 370 + ], + "type": "text", + "content": "[18] Shangqi Gao, Hangqi Zhou, Yibo Gao, and Xiahai Zhuang. Bayeseg: Bayesian modeling for medical image segmentation with interpretable generalizability. Medical Image Analysis, 89, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 372, + 555, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 372, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 321, + 372, + 555, + 415 + ], + "type": "text", + "content": "[19] Noa Garcia, Benjamin Renoust, and Yuta Nakashima. Contextnet: representation and exploration for painting classification and retrieval in context. International Journal of Multimedia Information Retrieval, 9(1), 2020. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 416, + 555, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 416, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 321, + 416, + 555, + 458 + ], + "type": "text", + "content": "[20] Leon Gatys, Alexander S Ecker, and Matthias Bethge. Texture synthesis using convolutional neural networks. In Advances in Neural Information Processing Systems (NeurIPS), 2015. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 460, + 555, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 460, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 321, + 460, + 555, + 503 + ], + "type": "text", + "content": "[21] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 504, + 555, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 504, + 555, + 558 + ], + "spans": [ + { + "bbox": [ + 321, + 504, + 555, + 558 + ], + "type": "text", + "content": "[22] Ming Gui, Johannes Schusterbauer, Ulrich Prestel, Pingchuan Ma, Dmytro Kotovenko, Olga Grebenkova, Stefan Andreas Baumann, Vincent Tao Hu, and Björn Ommer. Depthfm: Fast monocular depth estimation with flow matching. AAAI, 2025. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 559, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 559, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 321, + 559, + 553, + 592 + ], + "type": "text", + "content": "[23] Alexander Hermans, Lucas Beyer, and Bastian Leibe. In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737, 2017. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 321, + 593, + 555, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 593, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 321, + 593, + 555, + 624 + ], + "type": "text", + "content": "[24] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. In CVPR, 2024. 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 321, + 625, + 555, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 625, + 555, + 678 + ], + "spans": [ + { + "bbox": [ + 321, + 625, + 555, + 678 + ], + "type": "text", + "content": "[25] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 321, + 681, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 681, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 321, + 681, + 555, + 713 + ], + "type": "text", + "content": "[26] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems (NeurIPS), 33, 2020. 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5875" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 72, + 297, + 115 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 297, + 115 + ], + "type": "text", + "content": "[27] Vincent Tao Hu, Yunlu Chen, Mathilde Caron, Yuki M. Asano, Cees G.M. Snoek, and Björn Ommer. Guided diffusion from self-supervised diffusion features. In ARXIV, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 118, + 296, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 296, + 172 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 296, + 172 + ], + "type": "text", + "content": "[28] Vincent Tao Hu, Wenzhe Yin, Pingchuan Ma, Yunlu Chen, Basura Fernando, Yuki M. Asano, Efstratos Gavves, Pascal Mettes, Björn Ommer, and Cees G.M. Snoek. Motion flow matching for human motion synthesis and editing. In ARXIV, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 175, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 175, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 62, + 175, + 296, + 228 + ], + "type": "text", + "content": "[29] Vincent Tao Hu, Stefan Andreas Baumann, Ming Gui, Olga Grebenkova, Pingchuan Ma, Johannes Fischer, and Björn Ommer. *Zigma: A dit-style zigzag mamba diffusion model*. In European conference on computer vision, pages 148-166. Springer, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 230, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 230, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 62, + 230, + 296, + 274 + ], + "type": "text", + "content": "[30] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 275, + 296, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 275, + 296, + 317 + ], + "spans": [ + { + "bbox": [ + 62, + 275, + 296, + 317 + ], + "type": "text", + "content": "[31] Herve Jegou, Matthijs Douze, and Cordelia Schmid. Product quantization for nearest neighbor search. IEEE transactions on pattern analysis and machine intelligence, 33(1):117-128, 2010. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 319, + 296, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 296, + 363 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 296, + 363 + ], + "type": "text", + "content": "[32] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Proceedings of the European Conference on Computer Vision (ECCV). Springer, 2016. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 365, + 296, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 365, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 365, + 296, + 407 + ], + "type": "text", + "content": "[33] Sergey Karayev, Matthew Trentacoste, Helen Han, Aseem Agarwala, Trevor Darrell, Aaron Hertzmann, and Holger Winnemoeller. Recognizing image style. arXiv preprint arXiv:1311.3715, 2013. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 410, + 296, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 410, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 410, + 296, + 441 + ], + "type": "text", + "content": "[34] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 444, + 296, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 444, + 296, + 476 + ], + "spans": [ + { + "bbox": [ + 62, + 444, + 296, + 476 + ], + "type": "text", + "content": "[35] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 478, + 296, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 478, + 296, + 532 + ], + "spans": [ + { + "bbox": [ + 62, + 478, + 296, + 532 + ], + "type": "text", + "content": "[36] Dmytro Kotovenko, Matthias Wright, Arthur Heimbrecht, and Bjorn Ommer. Rethinking style transfer: From pixels to parameterized brushstrokes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 533, + 296, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 296, + 577 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 296, + 577 + ], + "type": "text", + "content": "[37] Felix Krause, Timy Phan, Ming Gui, Stefan Andreas Baumann, Vincent Tao Hu, and Björn Ommer. Tread: Token routing for efficient architecture-agnostic diffusion training. arXiv preprint arXiv:2501.04765, 2025. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 578, + 296, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 578, + 296, + 599 + ], + "spans": [ + { + "bbox": [ + 62, + 578, + 296, + 599 + ], + "type": "text", + "content": "[38] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024.8, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 601, + 296, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 601, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 62, + 601, + 296, + 622 + ], + "type": "text", + "content": "[39] Berel Lang. The concept of style. Cornell University Press, 1987. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 624, + 296, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 296, + 657 + ], + "type": "text", + "content": "[40] Christian Léonard. A survey of the schr\\'' odinger problem and some of its connections with optimal transport. arXiv preprint arXiv:1308.0215, 2013. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 658, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 296, + 713 + ], + "type": "text", + "content": "[41] Wen Li, Muyuan Fang, Cheng Zou, Biao Gong, Ruobing Zheng, Meng Wang, Jingdong Chen, and Ming Yang. Style-tokensizer: Defining image style by a single instance for controlling diffusion models. In European Conference on Computer Vision, pages 110-126. Springer, 2024. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 320, + 73, + 555, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 555, + 126 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 555, + 126 + ], + "type": "text", + "content": "[42] Xueting Li, Sifei Liu, Jan Kautz, and Ming-Hsuan Yang. Learning linear transformations for fast image and video style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 129, + 555, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 129, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 320, + 129, + 555, + 171 + ], + "type": "text", + "content": "[43] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 175, + 555, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 175, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 320, + 175, + 555, + 206 + ], + "type": "text", + "content": "[44] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. *ICLR*, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 209, + 555, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 209, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 320, + 209, + 555, + 239 + ], + "type": "text", + "content": "[45] Guan-Horng Liu, Tianrong Chen, Oswin So, and Evangelos Theodorou. Deep generalized schrödinger bridge. NeurIPS, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 243, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 243, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 320, + 243, + 554, + 274 + ], + "type": "text", + "content": "[46] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. *ICLR*, 2023. 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 277, + 554, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 277, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 320, + 277, + 554, + 319 + ], + "type": "text", + "content": "[47] Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. arXiv, 2024. 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 322, + 555, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 322, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 320, + 322, + 555, + 354 + ], + "type": "text", + "content": "[48] L. McInnes, J. Healy, and J. Melville. UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction. ArXiv, 2018. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 356, + 555, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 356, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 320, + 356, + 555, + 399 + ], + "type": "text", + "content": "[49] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 6, 7, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 401, + 555, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 401, + 555, + 443 + ], + "spans": [ + { + "bbox": [ + 320, + 401, + 555, + 443 + ], + "type": "text", + "content": "[50] Amir Merchant, Simon Batzner, Samuel S Schoenholz, Muratahan Aykol, Gowoon Cheon, and Ekin Dogus Cubuk. Scaling deep learning for materials discovery. Nature, 624 (7990), 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 446, + 555, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 446, + 555, + 500 + ], + "spans": [ + { + "bbox": [ + 320, + 446, + 555, + 500 + ], + "type": "text", + "content": "[51] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, 2024. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 502, + 555, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 502, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 320, + 502, + 555, + 555 + ], + "type": "text", + "content": "[52] Kevin Musgrave, Serge Belongie, and Ser-Nam Lim. A metric learning reality check. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16, pages 681-699. Springer, 2020. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 558, + 555, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 558, + 555, + 600 + ], + "spans": [ + { + "bbox": [ + 320, + 558, + 555, + 600 + ], + "type": "text", + "content": "[53] Ivona Najdenkoska, Xiantong Zhen, Marcel Worring, and Ling Shao. Uncertainty-aware report generation for chest x-rays by variational topic inference. Medical Image Analysis, 82, 2022. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 603, + 555, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 603, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 320, + 603, + 555, + 635 + ], + "type": "text", + "content": "[54] Kirill Neklyudov, Daniel Severo, and Alireza Makhzani. Action matching: A variational method for learning stochastic dynamics from samples. arXiv, 2022. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 636, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 636, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 320, + 636, + 555, + 712 + ], + "type": "text", + "content": "[55] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bo" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5876" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 80, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 294, + 95 + ], + "type": "text", + "content": "janowski. Dinov2: Learning robust visual features without supervision, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 61, + 96, + 295, + 140 + ], + "type": "text", + "content": "[56] Norihiro Oyama, Noriko N Ishizaki, Satoshi Koide, and Hiroaki Yoshida. Deep generative model super-resolves spatially correlated multiregional climate data. Scientific Reports, 13(1), 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 141, + 295, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 141, + 295, + 174 + ], + "spans": [ + { + "bbox": [ + 61, + 141, + 295, + 174 + ], + "type": "text", + "content": "[57] Emanuel Parzen. On estimation of a probability density function and mode. The annals of mathematical statistics, 33(3):1065-1076, 1962. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 175, + 296, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 175, + 296, + 230 + ], + "spans": [ + { + "bbox": [ + 62, + 175, + 296, + 230 + ], + "type": "text", + "content": "[58] Pablo Pernias, Dominic Rampas, Mats Leon Richter, Christopher Pal, and Marc Aubreville. Würstchen: An efficient architecture for large-scale text-to-image diffusion models. In The Twelfth International Conference on Learning Representations, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 232, + 295, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 232, + 295, + 263 + ], + "spans": [ + { + "bbox": [ + 62, + 232, + 295, + 263 + ], + "type": "text", + "content": "[59] Ed Pizzi, Sreya Dutta Roy, Sugosh Nagavara Ravindra, Priya Goyal, and Matthijs Douze. A self-supervised descriptor for image copy detection. In CVPR, 2022. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 266, + 294, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 266, + 294, + 309 + ], + "spans": [ + { + "bbox": [ + 62, + 266, + 294, + 309 + ], + "type": "text", + "content": "[60] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 8, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 311, + 295, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 311, + 295, + 364 + ], + "spans": [ + { + "bbox": [ + 62, + 311, + 295, + 364 + ], + "type": "text", + "content": "[61] Tianhao Qi, Shancheng Fang, Yanze Wu, Hongtao Xie, Jiawei Liu, Lang Chen, Qian He, and Yongdong Zhang. Deadiff: An efficient stylization diffusion model with disentangled representations. arXiv preprint arXiv:2403.06951, 2024. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 366, + 295, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 366, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 62, + 366, + 295, + 431 + ], + "type": "text", + "content": "[62] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning. PMLR, 2021. 2, 6, 8, 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 434, + 295, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 434, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 295, + 498 + ], + "type": "text", + "content": "[63] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML). PMLR, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 501, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 501, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 501, + 295, + 533 + ], + "type": "text", + "content": "[64] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 535, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 535, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 535, + 295, + 578 + ], + "type": "text", + "content": "[65] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 6, 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 580, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 580, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 62, + 580, + 295, + 633 + ], + "type": "text", + "content": "[66] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022. 1, 3, 6, 8, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 635, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 295, + 667 + ], + "type": "text", + "content": "[67] L Rout, Y Chen, N Ruiz, C Caramanis, S Shakkottai, and W Chu. Semantic image inversion and editing using rectified stochastic differential equations. 2024. 4, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 295, + 712 + ], + "type": "text", + "content": "[68] Litu Rout, Yujia Chen, Nataniel Ruiz, Constantine Caramanis, Sanjay Shakkottai, and Wen-Sheng Chu. Semantic image inversion and editing using rectified stochastic differential equations. In ICLR, 2025. 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 321, + 73, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 73, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 555, + 128 + ], + "type": "text", + "content": "[69] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 129, + 555, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 129, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 320, + 129, + 555, + 173 + ], + "type": "text", + "content": "[70] Dan Ruta, Saeid Motiian, Baldo Faieta, Zhe Lin, Hailin Jin, Alex Filipkowski, Andrew Gilbert, and John Collomosse. Aladin: All layer adaptive instance normalization for fine-grained style similarity. In ICCV, 2021. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 175, + 555, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 175, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 320, + 175, + 555, + 206 + ], + "type": "text", + "content": "[71] Babak Saleh and Ahmed Elgammal. Large-scale classification of fine-art paintings: Learning the right metric on the right feature. arXiv preprint arXiv:1505.00855, 2015. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 209, + 555, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 209, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 320, + 209, + 555, + 229 + ], + "type": "text", + "content": "[72] Meyer Schapiro. Theory and philosophy of art: Style, artist, and society. 1994. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 231, + 555, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 231, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 320, + 231, + 555, + 274 + ], + "type": "text", + "content": "[73] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 276, + 555, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 276, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 320, + 276, + 555, + 330 + ], + "type": "text", + "content": "[74] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 3, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 332, + 554, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 332, + 554, + 364 + ], + "spans": [ + { + "bbox": [ + 320, + 332, + 554, + 364 + ], + "type": "text", + "content": "[75] Johannes Schusterbauer, Ming Gui, Pingchuan Ma, Nick Stracke, Stefan A Baumann, and Björn Ommer. Boosting latent diffusion with flow matching. ECCV, 2024. 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 365, + 555, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 365, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 320, + 365, + 555, + 407 + ], + "type": "text", + "content": "[76] Viraj Shah, Nataniel Ruiz, Forrester Cole, Erika Lu, Svetlana Lazebnik, Yuanzhen Li, and Varun Jampani. Ziplora: Any subject in any style by effectively merging loras. In ECCV, 2024. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 411, + 555, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 411, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 320, + 411, + 555, + 453 + ], + "type": "text", + "content": "[77] Higor YD Sigaki, Matjaž Perc, and Haroldo V Ribeiro. History of art paintings through the lens of entropy and complexity. Proceedings of the National Academy of Sciences, 115(37):E8585-E8594, 2018. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 456, + 554, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 456, + 554, + 487 + ], + "spans": [ + { + "bbox": [ + 320, + 456, + 554, + 487 + ], + "type": "text", + "content": "[78] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 489, + 555, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 489, + 555, + 532 + ], + "spans": [ + { + "bbox": [ + 320, + 489, + 555, + 532 + ], + "type": "text", + "content": "[79] Gowthami Somepalli, Anubhav Gupta, Kamal Gupta, Shramay Palta, Micah Goldblum, Jonas Geiping, Abhinav Shrivastava, and Tom Goldstein. Measuring style similarity in diffusion models. ECCV, 2024. 2, 3, 5, 6, 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 534, + 555, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 534, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 320, + 534, + 555, + 555 + ], + "type": "text", + "content": "[80] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 557, + 555, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 557, + 555, + 600 + ], + "spans": [ + { + "bbox": [ + 320, + 557, + 555, + 600 + ], + "type": "text", + "content": "[81] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 3, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 602, + 554, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 602, + 554, + 634 + ], + "spans": [ + { + "bbox": [ + 320, + 602, + 554, + 634 + ], + "type": "text", + "content": "[82] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 636, + 554, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 636, + 554, + 667 + ], + "spans": [ + { + "bbox": [ + 320, + 636, + 554, + 667 + ], + "type": "text", + "content": "[83] Xuan Su, Jiaming Song, Chenlin Meng, and Stefano Ermon. Dual diffusion implicit bridges for image-to-image translation. *ICLR*, 2023. 1, 3, 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 320, + 670, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 670, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 320, + 670, + 554, + 712 + ], + "type": "text", + "content": "[84] Wei Ren Tan, Chee Seng Chan, Hernan Aguirre, and Kiyoshi Tanaka. Improved artgan for conditional synthesis of natural image and artwork. IEEE Transactions on Image Processing, 28(1):394-409, 2019. 5" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5877" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 106 + ], + "type": "text", + "content": "[85] Zhicong Tang, Tiankai Hang, Shuyang Gu, Dong Chen, and Baining Guo. Simplified diffusion schrödinger bridge. arXiv preprint arXiv:2403.14623, 2024. 1, 3, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 106, + 296, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 106, + 296, + 139 + ], + "spans": [ + { + "bbox": [ + 61, + 106, + 296, + 139 + ], + "type": "text", + "content": "[86] Gemma Canet Tarrés, Dan Ruta, Tu Bui, and John Collomosse. Parasol: Parametric style control for diffusion image synthesis. arXiv preprint arXiv:2303.06464, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 140, + 295, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 140, + 295, + 183 + ], + "spans": [ + { + "bbox": [ + 62, + 140, + 295, + 183 + ], + "type": "text", + "content": "[87] Haofan Wang, Matteo Spinelli, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv preprint arXiv:2404.02733, 2024. 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 184, + 295, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 184, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 62, + 184, + 295, + 216 + ], + "type": "text", + "content": "[88] Sheng-Yu Wang, Alexei A Efros, Jun-Yan Zhu, and Richard Zhang. Evaluating data attribution for text-to-image models. In ICCV, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 217, + 295, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 217, + 295, + 259 + ], + "spans": [ + { + "bbox": [ + 62, + 217, + 295, + 259 + ], + "type": "text", + "content": "[89] Zhizhong Wang, Lei Zhao, Sihuan Lin, Qihang Mo, Huiming Zhang, Wei Xing, and Dongming Lu. Glstylenet: exquisite style transfer combining global and local pyramid features. IET Computer Vision, 14(8), 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 261, + 296, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 296, + 315 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 296, + 315 + ], + "type": "text", + "content": "[90] Zhizhong Wang, Lei Zhao, Haibo Chen, Ailin Li, Zhiwen Zuo, Wei Xing, and Dongming Lu. Texture reformer: Towards fast and universal interactive texture transfer. In Proceedings of the AAAI Conference on Artificial Intelligence, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 316, + 295, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 316, + 295, + 370 + ], + "spans": [ + { + "bbox": [ + 62, + 316, + 295, + 370 + ], + "type": "text", + "content": "[91] Zhizhong Wang, Lei Zhao, Haibo Chen, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. DivSwapper: Towards diversified patch-based arbitrary style transfer. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence (IJCAI), 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 371, + 295, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 371, + 295, + 425 + ], + "spans": [ + { + "bbox": [ + 62, + 371, + 295, + 425 + ], + "type": "text", + "content": "[92] Zhizhong Wang, Lei Zhao, Zhiwen Zuo, Ailin Li, Haibo Chen, Wei Xing, and Dongming Lu. Microast: Towards super-fast ultra-resolution arbitrary style transfer. In Proceedings of the AAAI Conference on Artificial Intelligence, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 426, + 295, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 426, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 62, + 426, + 295, + 470 + ], + "type": "text", + "content": "[93] Michael J Wilber, Chen Fang, Hailin Jin, Aaron Hertzmann, John Collomosse, and Serge Belongie. Bam! the behave artistic media dataset for recognition beyond photography. In ICCV, 2017. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 471, + 295, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 471, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 62, + 471, + 295, + 514 + ], + "type": "text", + "content": "[94] Lemeng Wu, Dilin Wang, Chengyue Gong, Xingchao Liu, Yunyang Xiong, Rakesh Ranjan, Raghuraman Krishnamoorthi, Vikas Chandra, and Qiang Liu. Fast point cloud generation with straight flows. arXiv, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 515, + 295, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 515, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 62, + 515, + 295, + 558 + ], + "type": "text", + "content": "[95] Zijie Wu, Zhen Zhu, Junping Du, and Xiang Bai. Ccpl: Contrastive coherence preserving loss for versatile style transfer. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 559, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 559, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 559, + 295, + 590 + ], + "type": "text", + "content": "[96] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 592, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 592, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 62, + 592, + 295, + 656 + ], + "type": "text", + "content": "[97] Sheng Zhang, Yanbo Xu, Naoto Usuyama, Hanwen Xu, Jaspreet Bagga, Robert Tinn, Sam Preston, Rajesh Rao, Mu Wei, Naveen Valluri, et al. Biomedclip: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs. arXiv preprint arXiv:2303.00915, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 295, + 712 + ], + "type": "text", + "content": "[98] Yuxin Zhang, Fan Tang, Weiming Dong, Haibin Huang, Chongyang Ma, Tong-Yee Lee, and Changsheng Xu. Domain enhanced arbitrary image style transfer via contrastive learning. In ACM SIGGRAPH 2022 Conference Proceedings, 2022. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 184 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 320, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 553, + 106 + ], + "type": "text", + "content": "[99] Wentao Zhao, Dalin Zhou, Xinguo Qiu, and Wei Jiang. How to represent paintings: A painting classification using artistic comments. Sensors, 21(6), 2021. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 107, + 555, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 555, + 129 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 555, + 129 + ], + "type": "text", + "content": "[100] Linqi Zhou, Aaron Lou, Samar Khanna, and Stefano Ermon. Denoising diffusion bridge models. *ICLR*, 2024. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 129, + 555, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 184 + ], + "type": "text", + "content": "[101] Zhiwen Zuo, Lei Zhao, Shuobin Lian, Haibo Chen, Zhizhong Wang, Ailin Li, Wei Xing, and Dongming Lu. Style fader generative adversarial networks for style degree controllable artistic style transfer. In Proc. Int. Joint Conf. on Artif. Intell.(IJCAI), 2022. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5878" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/StochasticSplats_ Stochastic Rasterization for Sorting-Free 3D Gaussian Splatting/d963f664-7a0a-43e1-95cb-e52f6169cd94_content_list.json b/2025/StochasticSplats_ Stochastic Rasterization for Sorting-Free 3D Gaussian Splatting/d963f664-7a0a-43e1-95cb-e52f6169cd94_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1cc19c59d8525b9e765233992dcdc0b331755892 --- /dev/null +++ b/2025/StochasticSplats_ Stochastic Rasterization for Sorting-Free 3D Gaussian Splatting/d963f664-7a0a-43e1-95cb-e52f6169cd94_content_list.json @@ -0,0 +1,2082 @@ +[ + { + "type": "text", + "text": "Stochastic Splats: Stochastic Rasterization for Sorting-Free 3D Gaussian Splatting", + "text_level": 1, + "bbox": [ + 174, + 130, + 821, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shakiba Kheradmand\\*1,2, Delio Vicini\\*3, George Kopanas\\*3,4, Dmitry Lagun $^{1}$ , Kwang Moo Yi $^{2}$ , Mark Matthews $^{1}$ , Andrea Tagliasacchi $^{1,5,6}$", + "bbox": [ + 196, + 202, + 797, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Google DeepMind, 2University of British Columbia, 3Google, 4Runway ML, 5Simon Fraser University, 6University of Toronto, *equal contribution,†work done at Google", + "bbox": [ + 191, + 252, + 805, + 306 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/22c9091e33f6ad2f0daf8f624a3b04b3a4e0536521347343605c06cee4877ce6.jpg", + "image_caption": [ + "Figure 1. Teaser - We render a 3DGS radiance field using an unbiased Monte Carlo estimator of the volume rendering equation. Unlike prior radiance field approaches, our method does not require sorted back-to-front rendering. This results in fast, portable, and pop-free rendering that is easy to implement. The number of Monte Carlo samples per pixel allows a trade-off between interactivity and visual quality, similar to physically-based rendering. From left to right, we increase the number of samples per pixel, which increases visual fidelity at a predictable increase in rendering cost. We report the mean square error (MSE) and render time on an NVIDIA RTX 4090." + ], + "image_footnote": [], + "bbox": [ + 106, + 323, + 895, + 593 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 681, + 328, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D Gaussian splatting (3DGS) is a popular radiance field method, with many application-specific extensions. Most variants rely on the same core algorithm: depth-sorting of Gaussian splats then rasterizing in primitive order. This ensures correct alpha compositing, but can cause rendering artifacts due to built-in approximations. Moreover, for a fixed representation, sorted rendering offers little control over render cost and visual fidelity. For example, and counter-intuitively, rendering a lower-resolution image is not necessarily faster. In this work, we address the above limitations by combining 3D Gaussian splatting with stochastic rasterization. Concretely, we leverage an unbi", + "bbox": [ + 89, + 712, + 485, + 893 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ased Monte Carlo estimator of the volume rendering equation. This removes the need for sorting, and allows for accurate 3D blending of overlapping Gaussians. The number of Monte Carlo samples further imbues 3DGS with a way to trade off computation time and quality. We implement our method using OpenGL shaders, enabling efficient rendering on modern GPU hardware. At a reasonable visual quality, our method renders more than four times faster than sorted rasterization. The project page and source code are publicly available at https://ubc-vision.github.io/stochasticsplats/.", + "bbox": [ + 511, + 681, + 908, + 849 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26326", + "bbox": [ + 478, + 940, + 519, + 952 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 89, + 222, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advances in novel view synthesis have been driven by the success of neural radiance fields (NeRF) [29] and 3D Gaussian splatting (3DGS) [18]. The former produced high-quality visuals at a low rendering speed, while the latter enabled efficient rendering with comparable visual quality. 3DGS represents a scene as a set of anisotropic 3D Gaussians that can be \"splatted\" [57] in rasterization pipelines, enabling real-time viewing and efficient training.", + "bbox": [ + 89, + 114, + 483, + 234 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In 3DGS, 3D Gaussians are flattened onto camera-oriented billboards that not only prevent accurate volumetric intermixing [25], but must also be depthwise sorted for rasterization [58]. This global sorting operation is computationally expensive and introduces popping artifacts [37], as the sort order is view-dependent: small camera movements can result in temporal discontinuities in the rendered images. Further, there is no straightforward way to balance rendering cost vs. visual fidelity. One might assume that reducing the rendering resolution would lead to faster rasterization... yet, we observe the opposite in our experiments. Overall, this limits 3DGS's use on low-end hardware and in latency-critical applications, such as robotics and autonomous driving. Finally, most implementation use custom software rasterizers or GPGPU frameworks such as CUDA. Software rasterizers are notoriously hard to optimize, and often cannot match the performance of the hardware stack [23]. Additionally, many GPGPU frameworks are vendor-specific, reducing portability.", + "bbox": [ + 89, + 236, + 483, + 522 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several previous works address popping artifacts, rendering performance, and portability, but typically improve only one, at the expense of the others. Popping can be alleviated by per-pixel depth sorting, but this introduces extra computational demands. While this cost can be amortized using clever spatial biases [37], per-pixel sorting cannot fully solve volumetric intermixing of 3D Gaussians. Constant density ellipsoids allow analytical intermixing [25], but are slow to render because they replace sorting with expensive ray tracing. Other methods [15] reduce the rendering cost and the popping artifacts by using a mix of per-pixel sorted lists and weighted order-independent transparency, but this comes at a cost in terms of portability since they need to implement complicated GPGPU kernels hindering cross-platform adoption. An orthogonal route to reduce render cost is to reduce the number of Gaussians [20]. Finally, packages such as Splatapult [42] implement an OpenGL-based rasterization pipeline, but do not outperform the original CUDA 3DGS implementation.", + "bbox": [ + 89, + 523, + 483, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We instead propose to render 3D Gaussians using stochastic transparency (ST) [8]; see Figure 2. Stochastic transparency replaces sorted alpha blending by an unbiased Monte Carlo estimator. Practically, transparency is rendered by probabilistically selecting individual, opaque, samples that are averaged together with multi-sampling [6].", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9531edc4d98741eb90c0260fb004ea3699ba48c5be4f799d62a6c8ff45a09cca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 89, + 700, + 170 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f3fec56a804b5faa63d6b81524b26bde9bf0894ad8fe7483f8bfb9080afe9c80.jpg", + "image_caption": [ + "Alpha blending" + ], + "image_footnote": [], + "bbox": [ + 521, + 172, + 696, + 276 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2916131be94fdd51ba5dd1850a3e775d00870fad31c8c6d82ae22013f6deef4b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 90, + 901, + 171 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/262c2b18e7dddd00bb2a34489cbb216e299a392d22f4308758f9ec02edcfd51d.jpg", + "image_caption": [ + "Stochastic transparency (4 SPP)", + "Figure 2. Blending vs. stochastic - We visualize a 1D and 3D example of the different ways to compute transparency. Two foreground Gaussians are composited on a white background." + ], + "image_footnote": [], + "bbox": [ + 727, + 176, + 901, + 276 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Individual samples are rendered using order-independent Z-buffering, without sorting. We provide unbiased Monte-Carlo estimators of both the forward alpha blending and the gradient computation. The reverse-mode gradient estimator also permits sorting-free training and is largely agnostic to the primitive type and could be used for other semitransparent representations (e.g., triangle soups). We also do an early analysis on proper volumetric intermixing of 3D Gaussians by extending the theory of stochastic transparency. Critically, all of the above is naturally compatible with the hardware rasterization pipeline, allowing us a portable OpenGL implementation of our method.", + "bbox": [ + 511, + 348, + 906, + 529 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "At one sample per pixel (SPP), our OpenGL implementation renders up to $4 \\times$ faster than the original CUDA implementation, enabling latency-critical applications, such as open-vocabulary 3D object localization [19], which we demonstrate. Higher SPP increase quality and render time, approaching the PSNR of alpha-blended 3DGS. One key advantage is that downstream applications can dynamically adjust sampling at rendering time, allowing them to potentially trade-off latency against quality with careful engineering; see Figure 1.", + "bbox": [ + 511, + 530, + 908, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 693, + 653, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Order-independent transparency. While current 3DGS methods rely on sorted alpha blending, there is a rich literature on order-independent transparency for rasterization-based real-time rendering. Prior work can be classified into: (1) stochastic methods [8, 21, 49], that probabilistically discard fragments based on their opacity to create Monte Carlo estimators of the alpha blending equation and (2) non-stochastic methods that avoid sorting by using per-pixel linked lists [50], depth peeling [9], weighted blending [27], moments [30], or deep learning [44] to (approximately) composite transparent fragments.", + "bbox": [ + 511, + 719, + 906, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given the success of order-independent transparency in", + "bbox": [ + 531, + 885, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26327", + "bbox": [ + 478, + 940, + 517, + 952 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "real-time graphics, it becomes desirable to incorporate these techniques with 3DGS. Recent pre-prints [15, 17] combine 3DGS with non-stochastic order-independent transparency, but stochastic variants remain unexplored. We bridge this gap by extensively studying stochastic transparency with 3DGS, both in terms of rendering and training, enabled by our novel stochastic gradient estimator.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Popping artifacts. Real-time graphics eschew sorting primitives due to both the computational cost and the inexistence of a consistent global ordering of non-trivial primitives. In 3DGS this is overlooked, because most primitives have a relatively small spatial extent. This introduces artifacts often dubbed as \"popping\", where a small motion of the camera forces a change in the order of the primitives introducing a sharp discontinuity in the alpha blending equation. These artifacts are particularly noticeable in immersive contexts like virtual reality. Radl et al. [37] address this using per-pixel sorting based on the maximal Gaussian response. While this eliminates popping artifacts, it introduces prohibitive computational costs, necessitating complex approximate sorting through a multi-stage hierarchical renderer. EVER [25] addresses popping artifacts by ray tracing constant density ellipsoids to compute the full volume rendering equation, rather than approximate alpha-blending. However, this approach is significantly slower than rasterization. In contrast, stochastic transparency uniquely eliminates popping by trading sorting for some noise in the final image.", + "bbox": [ + 91, + 199, + 482, + 516 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fast 3DGS rendering and training. While 3DGS [18] offered a significant speedup over NeRF-based methods, subsequent works have further improved performance. Numerous improvements have been made to the implementation of training and rendering routines [14, 15, 26, 51]. Further speedups can be made by reducing the number of Gaussians and compressing their attributes. This is covered in an extensive literature review [3]. Any approach that reduces the number of primitives readily complements our work and is likely to improve rendering performance.", + "bbox": [ + 89, + 518, + 482, + 670 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Differentiable Monte Carlo methods. We introduce a novel Monte Carlo gradient estimator for stochastic transparency. Such differentiable Monte Carlo estimators are commonly used in physically-based inverse rendering [12, 22, 32, 46, 54, 56]. Various works propose differentiable Monte Carlo methods for neural radiance fields [24, 45] or surface reconstruction [16], and Condor et al. [5] differentiate physically-based Monte Carlo ray tracing of 3D Gaussian primitives. Deliot et al. [7] applied stochastic simultaneous perturbation gradients to real-time rendering algorithms, including standard 3DGS. Beyond rendering, differentiable Monte Carlo methods find application in nuclear engineering [28], finance [11], atmospheric physics [43] and inverse partial differential equation problems [52].", + "bbox": [ + 89, + 672, + 482, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 89, + 604, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Kerbl et al. [18] represents a 3D scene in 3DGS as a collection of anisotropic 3D Gaussians $\\mathcal{G} = \\{g_i\\}$ . Given a set of posed images $\\{\\mathbf{I}_n\\}$ , we optimize the parameters of the 3D Gaussians $\\mathcal{G}$ to photometrically reconstruct the scene. Each $g_i$ is parameterized as $\\theta = \\{\\mu, \\Sigma, \\alpha, \\mathbf{c}\\}$ , where $\\mu$ is the mean, $\\Sigma$ is the covariance matrix, $\\alpha$ is the opacity, and $\\mathbf{c}$ is the view-dependent color modeled using spherical harmonics [38]. We start by reviewing the classical alpha-blending based rendering of 3DGS in Section 3.1, and then introduce our stochastic transparency algorithm in Section 3.2.", + "bbox": [ + 511, + 114, + 906, + 267 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Sorted alpha blending", + "text_level": 1, + "bbox": [ + 511, + 276, + 720, + 292 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To calculate the radiance $C$ at a given pixel, we aggregate the radiances emitted by $L$ primitives that intersect the associated ray, blending them in depth $(z)$ order [35]:", + "bbox": [ + 511, + 297, + 906, + 345 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nC = \\sum_ {i = 1} ^ {L} c _ {i} \\alpha_ {i} \\prod_ {z _ {k} < z _ {i}} (1 - \\alpha_ {k}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 354, + 906, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $z_{k} < z_{i}$ denotes all primitives in front of primitive $i$ . While it can be shown that this expression derives from volume rendering [41], the product of terms in (1) requires sorting the primitives along $z$ , then blend radiances in either back-to-front or front-to-back order. As sorting is independent of how many pixels one needs to render, we simply lose the ability to trade off accuracy with speed.", + "bbox": [ + 511, + 409, + 905, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sorting is done with respect to the $z$ coordinates of the Gaussian means $\\mu_{k}$ remapped to the coordinate frame of the current camera, and the covariance is projected via a linear approximation of the non-linear projection process to the camera's view point, as proposed by Zwicker et al. [58]. This effectively treats each Gaussian as a billboard, flattening it parallel to the camera's imaging plane.", + "bbox": [ + 511, + 515, + 906, + 621 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Stochastic transparency", + "text_level": 1, + "bbox": [ + 511, + 631, + 733, + 646 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Stochastic transparency (ST) [8] avoids sorting by stochastically estimating alpha blending in (1). Avoiding sorting when rendering transparent elements is critical, because (i) primitives often do not have a global unique sorting order across all pixels, (ii) sorting can be computationally expensive, and (iii) in the context of volume rendering the 3D extent of primitives influences the intermixing of colors.", + "bbox": [ + 511, + 652, + 905, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Stochastic transparency approximates transparency by randomly assigning a number of fully opaque samples per pixel proportional to a fragment's opacity, and averaging these samples to produce the final pixel color without sorting. In our context, the opacity of a fragment is proportional to the opacity of the Gaussian $\\alpha$ , and to its distance from the mean $\\mu$ , as measured by the projected covariance", + "bbox": [ + 511, + 758, + 906, + 864 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1Note that as we do not use multi-sampling anti-aliasing, making fragments synonymous with pixels.", + "bbox": [ + 511, + 875, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26328", + "bbox": [ + 478, + 940, + 519, + 952 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "$\\boldsymbol {z}\\leftarrow \\infty$ $\\mathbf{C}\\gets c_{\\mathrm{background}}$ \nfor each Gaussian $i$ in unsorted order do Compute $\\alpha_{i}$ and depth $z_{i}$ Sample $u\\sim \\mathcal{U}(0,1)$ if $u < \\alpha_{i}$ and $z_{i} < z$ then $\\mathbf{C}\\gets c_i$ $z\\gets z_{i}$ \nend if \nend for", + "guess_lang": "latex", + "bbox": [ + 116, + 90, + 387, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Listing 1. Pseudocode for stochastic transparency.", + "bbox": [ + 135, + 233, + 434, + 247 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "matrix $\\Sigma$ . Stochastic transparency can also be interpreted as a randomized extension of classical screen-door transparency [10, Chapter 16], where, instead of a randomized selection, a fixed pattern is employed to discard fragments.", + "bbox": [ + 89, + 258, + 482, + 320 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Simple example. As a simple example, consider a single 3D Gaussian $g$ covering pixel $p$ with opacity $\\alpha = 0.5$ ; in this case, if we sample $x \\sim \\mathcal{U}(0,1)$ and accept $g$ only when $x > \\alpha$ , approximately $50\\%$ of the pixel's samples will randomly select this Gaussian. We next formalize this idea to generalize to rendering an arbitrary number of Gaussians.", + "bbox": [ + 89, + 321, + 483, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formalization. Formally, stochastic transparency is a Monte Carlo estimator of (1). It estimates the radiance $C$ by randomly sampling a summand $i$ (i.e., the $i$ -th Gaussian):", + "bbox": [ + 89, + 416, + 483, + 460 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC \\approx \\frac {1}{P (i)} c _ {i} \\alpha_ {i} \\prod_ {z _ {k} < z _ {i}} \\left(1 - \\alpha_ {k}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 465, + 482, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $P(i)$ is the probability mass function of the random index $i$ . This estimator is unbiased, as long as $P(i) > 0$ for any non-zero summand. The standard approach is to use:", + "bbox": [ + 89, + 508, + 483, + 555 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP (i) = \\alpha_ {i} \\prod_ {z _ {k} < z _ {i}} \\left(1 - \\alpha_ {k}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 560, + 482, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The division by this probability mass function cancels out part of the sample weight, leading to a simple estimator:", + "bbox": [ + 89, + 598, + 482, + 628 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC \\approx c _ {i} \\quad \\text {f o r} \\quad i \\sim P. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 633, + 480, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thus, a pixel has the probability $P(C = c_i) = P(i)$ to pick color $c_i$ . In practice, we average multiple independent samples per pixel to reduce estimator noise.", + "bbox": [ + 89, + 657, + 482, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pseudocode. Listing 1 provides pseudocode for the above estimator. By testing that the depth of the current sample is smaller than the previously selected one, we sample exactly according to $P(i)$ [8]. In the fragment Shader, this is implemented by using standard depth testing and calling discard on samples where $u \\geq \\alpha_i$ . This only requires access to the current Gaussian's opacity and works for any primitive ordering. Figure 2 further shows a toy example of using this estimator on two Gaussians.", + "bbox": [ + 89, + 704, + 483, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Differentiable stochastic transparency", + "text_level": 1, + "bbox": [ + 89, + 848, + 418, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Applying stochastic transparency to inverse rendering problems requires computing the gradient of the photometric", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1e9b9805511e55f6a1a5f4fd1bf0aad3954762598c8306adf83ce74766bee03b.jpg", + "image_caption": [ + "Alpha blending" + ], + "image_footnote": [], + "bbox": [ + 517, + 89, + 645, + 232 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c0b58e11d23e59ed55ed7f101145c3efd946c524547d08231490947011fa13f9.jpg", + "image_caption": [ + "Ours (128 SPP)" + ], + "image_footnote": [], + "bbox": [ + 645, + 90, + 772, + 232 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bcb4c2a7b38bca3cb03c430e94d3044530be14dc7f017ed93e371194356c72c3.jpg", + "image_caption": [ + "Ours (512 SPP)", + "Figure 3. Gradient validation - We compare alpha blending gradients to our stochastic estimator with 128 and 512 SPP. Red and blue colors encode positive and negative values, respectively. Our estimator accurately approximates the alpha blending gradients." + ], + "image_footnote": [], + "bbox": [ + 774, + 90, + 901, + 232 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "loss $\\mathcal{L}$ with respect to the primitive parameters $\\theta$ . Primarily, we require gradients with respect to color $\\partial \\mathcal{L} / \\partial c$ , and gradients with respect to opacity $\\partial \\mathcal{L} / \\partial \\alpha$ .", + "bbox": [ + 511, + 318, + 905, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Gradient estimator. We take inspiration from the differentiable rendering literature [46], and use a detached gradient estimator [46, 53] to correctly differentiate the binary sampling decision in Listing 1. We derive this estimator by applying the gradient operator to the alpha blending equation:", + "bbox": [ + 511, + 366, + 905, + 441 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\boldsymbol {\\theta}} C = \\sum_ {i = 1} ^ {L} \\nabla_ {\\boldsymbol {\\theta}} \\left[ c _ {i} \\alpha_ {i} \\prod_ {z _ {k} < z _ {i}} (1 - \\alpha_ {k}) \\right]. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 450, + 903, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then form an unbiased Monte Carlo gradient estimator by again sampling a summand according to $P(i)$ . Crucially, we explicitly do not differentiate $P(i)$ itself and obtain:", + "bbox": [ + 511, + 500, + 905, + 546 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\boldsymbol {\\theta}} C \\approx \\frac {1}{P (i)} \\nabla_ {\\boldsymbol {\\theta}} \\left[ c _ {i} \\alpha_ {i} \\prod_ {z _ {k} < z _ {i}} \\left(1 - \\alpha_ {k}\\right) \\right]. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 554, + 903, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Substituting the definition of $P(i)$ from (3) and differentiating $\\mathcal{L}$ with respect to $c$ and $\\alpha$ , we get:", + "bbox": [ + 511, + 604, + 903, + 635 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\mathcal {L}}{\\partial c _ {i}} = \\frac {\\partial \\mathcal {L}}{\\partial C} \\quad \\frac {\\partial \\mathcal {L}}{\\partial c _ {z _ {k} < z _ {i}}} = 0 \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 642, + 903, + 675 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\mathcal {L}}{\\partial \\alpha_ {i}} = \\frac {\\partial \\mathcal {L}}{\\partial C} \\frac {c _ {i}}{\\alpha_ {i}} \\quad \\frac {\\partial \\mathcal {L}}{\\partial \\alpha_ {z _ {k} < z _ {i}}} = \\frac {\\partial \\mathcal {L}}{\\partial C} \\frac {- c _ {i}}{1 - \\alpha_ {z _ {k} < z _ {i}}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 676, + 903, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where subscript $i$ denotes the sampled primitive, and $\\partial \\mathcal{L} / \\partial C$ is the photometric loss gradient, dependent upon the loss formulation (L1, SSIM, ...). Note how in (7) only the $i$ -th Gaussian receives gradients with respect to its color, while the others have a zero gradient. Meanwhile, (8) shows how the $i$ -th Gaussian, and ones closer to the camera receive a gradient with respect to their opacity. The Gaussians behind the sampled one do not receive any gradient. Finally, our derivation is not specific to 3DGS, and applies to any representation rendered with stochastic transparency.", + "bbox": [ + 511, + 717, + 905, + 867 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Decorrelated loss gradient. A well-known challenge for differentiable Monte Carlo methods is that the derivative of", + "bbox": [ + 511, + 869, + 905, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26329", + "bbox": [ + 478, + 940, + 517, + 952 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8db56bfdffccd852f2a8fa615395d868b543caf740b4394f6914c8b2ada454a1.jpg", + "image_caption": [ + "3DGS" + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 285, + 193 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8626203bc7b47627c1cf4230a6699b27458a03e6f97f28f423ad7ed9c440aeee.jpg", + "image_caption": [ + "StopThePop [37] vs. ours", + "Figure 4. Popping in 3DGS - (left) As sorting in 3DGS is done with respect to the $z$ distance of the Gaussian mean from the camera, a small camera rotation can cause visible \"popping\" artifacts (i.e., sudden pixel color changes). (right) StopThePop [37] corrects for this behavior by associating a surface with each Gaussian, and determining the depth $z$ per-fragment (dashed line), rather than per-Gaussian. Our solution leads to visually comparable results, but approximates this surface linearly (solid line), so that per-fragment depth $z$ can be computed efficiently in hardware." + ], + "image_footnote": [], + "bbox": [ + 287, + 87, + 480, + 194 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the loss function is a product of two terms that are corrupted by noise (i.e., they are both random variables):", + "bbox": [ + 89, + 356, + 483, + 387 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\mathcal {L}}{\\partial \\theta} = \\frac {\\partial \\mathcal {L}}{\\partial C} \\frac {\\partial C}{\\partial \\theta}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 232, + 392, + 482, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\theta$ are the geometry parameters. If the noise in both terms is correlated, we get biased gradients, which harm optimization. Following [2, 12], we use different random samples to evaluate the two factors above. If $\\mathcal{L}$ is the L2 loss, the decorrelated evaluation results in unbiased gradients. For other loss functions, it at least reduces bias. Unbiased gradient estimation for arbitrary loss functions applied to a Monte Carlo estimator is an open problem.", + "bbox": [ + 89, + 428, + 483, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Practical algorithm. The presence of $c_{i}$ in (8) implies that we need access to the result of the forward pass to compute gradients. We therefore follow path replay backpropagation [46] and compute gradients in three passes:", + "bbox": [ + 89, + 551, + 483, + 611 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. In a first pass, we render an image and evaluate $\\partial \\mathcal{L} / \\partial C$ .", + "2. We then render a decorrelated image using a different random seed, giving us $c_{i}$", + "3. Finally, we replay the second pass by re-rendering using the same random seed and using the stored $c_{i}$ to compute parameter gradients." + ], + "bbox": [ + 91, + 611, + 482, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Validation. We validate our estimator by comparing to the gradients of the sorted alpha blending implementation. For easier visual interpretation, we use our reverse-mode gradient implementation to compute a gradient image. We do this by backpropagating an L1 loss separately for each pixel, and visualizing the sum of the x-component of the positional gradients of all Gaussians affecting a given pixel. As shown in Figure 3, our method produces noisy gradients that closely match those of alpha blending in structure.", + "bbox": [ + 89, + 704, + 482, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Removing popping artifacts", + "text_level": 1, + "bbox": [ + 89, + 848, + 341, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A naive implementation of stochastic transparency re-uses the billboard geometry of 3DGS, and consequently, to com", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/65811c20896c8e54a0c56d773b57252a7216c16c7f8009af2b902840b30886a3.jpg", + "image_caption": [ + "Alpha blending", + "Figure 5. Fully volumetric intermixing - While alpha blending (left) suffers from popping and \"pop-free\" variants (middle) show discontinuities, our full volumetric intermixing (right) accurately composes overlapping Gaussians." + ], + "image_footnote": [], + "bbox": [ + 522, + 92, + 645, + 174 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b61a6b8bfe2048be453292bc26317187a4b27dbb47d4755edecedd740b26db76.jpg", + "image_caption": [ + "Pop-free" + ], + "image_footnote": [], + "bbox": [ + 648, + 92, + 771, + 174 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a3050a1698f8d6dac1529063e9f1f8241f9a529355e01c833688993095e1d32a.jpg", + "image_caption": [ + "Our full intermixing" + ], + "image_footnote": [], + "bbox": [ + 774, + 92, + 898, + 174 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "pute the depths $z_{i}$ , the means of the Gaussians. As we will see in Section 4.2, at low sample count this provides high rendering efficiency, as we do not require a global sort. However, popping artifacts still exist, as the computation of $z_{i}$ is analogous to the one in 3DGS, small changes in the camera rotation can therefore change the order of primitives, and cause visible video artifacts; see Figure 4.", + "bbox": [ + 511, + 253, + 903, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The stop-the-pop way. Radl et al. [37] overcome the popping problem by computing the per-fragment (vs. per-Gaussian) depth. Their core idea is to compute the location of the peak of the Gaussian density along the ray. When coupled with a (computationally expensive) per-pixel sorting of Gaussians, this can provide a pop-free rendering. This can be interpreted as finding \"maximum density surface\" of the Gaussian, visualized in 2D in Figure 4, and using it to define fragment depths. However, per-fragment updates of depth are not particularly efficient, as early-discard mechanisms implemented in hardware [13] can only execute for fragments whose depth has not been modified by a shader. While in our setting we can forego sorting, the inability to early-discard fragments still affects us, which lead us to propose a different approach, described next.", + "bbox": [ + 511, + 362, + 905, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our simplified way. We propose to avoid this inefficiency by foregoing per-fragment depth changes. Instead, we orient our billboards to linearly approximate the maximum density surface. If the ray associated with a pixel has origin o, the plane can be defined as:", + "bbox": [ + 511, + 590, + 905, + 666 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {n} ^ {\\top} (\\mathbf {x} - \\boldsymbol {\\mu}) = 0, \\quad \\text {w h e r e} \\quad \\mathbf {n} = \\Sigma^ {- 1} (\\boldsymbol {\\mu} - \\mathbf {o}) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 671, + 903, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By approximating this surface with a plane, we can easily use the interpolation kernels within the hardware rasterizer to compute the splat's depth. This approach adds no extra time to the render process and gives us popping-free rendering with a similar PSNR to [37] at higher sample counts.", + "bbox": [ + 511, + 696, + 905, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fully volumetric intermixing. If maximum rendering fidelity is desired and the cost of per-fragment depth modification is acceptable, stochastic transparency can naturally support full volumetric intermixing. The key insight is that this can be achieved by replacing the fixed depth of a Gaussian by a sampled free-flight distance. For this, we first express the total contribution of a single Gaussian using the Beer-Lambert law: $\\int_0^\\infty \\sigma_i(\\mathbf{x}_t)\\exp \\left(-\\int_0^t\\sigma_i(\\mathbf{x}_s)\\mathrm{d}s\\right)c_i\\mathrm{d}t,$", + "bbox": [ + 511, + 773, + 905, + 904 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26330", + "bbox": [ + 478, + 940, + 519, + 952 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the opacity of a Gaussian is reparametrized using a spatially-varying extinction function $\\sigma_{i}$ and $\\mathbf{x}_a\\coloneqq \\mathbf{o} + a\\cdot \\mathbf{d}$ for a ray $(\\mathbf{o},\\mathbf{d})$ . We follow [5] and sample a free-flight distance $t$ proportional to $p(t) = \\sigma_i(\\mathbf{x}_t)\\exp (-\\int_0^t\\sigma_i(\\mathbf{x}_s)\\mathrm{d}s)$ . To handle overlapping Gaussians, we adapt decomposition tracking [33]: If two Gaussians overlap, we sample a free-flight distance according to each and take the minimum. This accurately resolves overlap of an arbitrary number of Gaussians without additional data structures, sorting, or ray tracing, and readily fits into the stochastic transparency algorithm. We simply need to replace the computation of $z_{i}$ by a probabilistic sampling $z_{i}\\sim p(t)$ . We show an example rendering of two overlapping Gaussians in Fig. 5. See the supplemental for the full free-flight sampling routine.", + "bbox": [ + 89, + 90, + 483, + 305 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 310, + 222, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate rendering performance on the real-world scenes from the MipNerf360 [4] dataset. For demonstrating semantic object localization with single-sample rendering we use the LERF dataset [19] - it consists of complex in-the-wild scenes, designed for 3D object localization.", + "bbox": [ + 89, + 334, + 483, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. For rendering baselines, we compare the performance of our method against the original 3DGS implementation [18] that uses CUDA, as well as Splatapult [42], an equivalent OpenGL implementation. We also compare with StopThePop [37], a pop-free method for 3DGS rendering.", + "bbox": [ + 89, + 412, + 483, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. We compute the standard metrics of peak signal-to-noise ratio (PSNR), structural similarity index metric (SSIM) [47], and learned perceptual image patch similarity (LPIPS) [55]. We report the average per-frame processing time for all methods on the MipNerf360 scenes.", + "bbox": [ + 89, + 489, + 483, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. To demonstrate portability, we implement our method in both CUDA and OpenGL. Our CUDA implementation builds on the existing 3DGS framework by Kerbl et al. [18] and our OpenGL implementation builds upon Splatapult [42]. We evaluate on a Windows laptop equipped with an NVIDIA Quadro T1000 Max-Q GPU, and Linux desktops with an NVIDIA RTX 3090 and RTX 4090. Rather than training from scratch, we finetune a trained 3DGS scene with the CUDA implementation of our method for 1000 iterations. Since we are performing fine-tuning, we do not apply densification or pruning. All learning rates for the non-position parameters are kept unchanged, while the positions are optimized using a learning rate of $5 \\times 10^{-5}$ with 128 samples. We keep all other hyperparameters the same. On average, the finetuning process requires approximately 14 minutes on an NVIDIA RTX 3090 GPU for scenes from MipNerf360.", + "bbox": [ + 89, + 568, + 483, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Rendering quality", + "text_level": 1, + "bbox": [ + 89, + 833, + 266, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative results - Figs. 6 and 7. We show two scenes (Room and Kitchen) in Figure 7 from the MipN-erf360 dataset. As shown, our method allows a trade-off", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5e63d90586df3c2513d981957964f45da53f3c9ddab16779fdbf1fd524675829.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 90, + 645, + 200 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/79035a8d0371cc0b2520044c7d1b61b318b7219148d6b198bb49b2af58dc3285.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 648, + 90, + 771, + 200 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/847e750c81fb2077154732fde3cf7e1f03ed371f01bd0c9a794f6aaebedf3bba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 776, + 90, + 898, + 200 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/33d21d8180d064514545a36e8f40dc68c628d8d3a3ab57d3fda7731c7811c224.jpg", + "image_caption": [ + "Frame $t$" + ], + "image_footnote": [], + "bbox": [ + 521, + 202, + 643, + 310 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/49a1ec46c55484282d557be2f0d68965465ea41ea0c786836cec8d8248dd3ab2.jpg", + "image_caption": [ + "Frame $t + 1$" + ], + "image_footnote": [], + "bbox": [ + 648, + 202, + 771, + 310 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c6ffa01c056fc62f02908df7d4a72ff50fabba06c7bd66405098288c01bac93a.jpg", + "image_caption": [ + "FLIP error", + "Figure 6. Popping artifact - The top row displays two consecutive frames, fine-tuned and rendered using our StochasticSplats, along with their corresponding fLIP error map [1]. The bottom row presents the same frames processed with standard 3DGS. The foreground error indicates popping while the background errors are due to the camera motion." + ], + "image_footnote": [], + "bbox": [ + 776, + 202, + 898, + 310 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b07b6e83845d80bde5303e7123a2225806d2e51afd65689f2f471d09782625c7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
3DGSStopThePop14Ours (SPP)
16642561024
PSNR28.9928.7917.9422.7526.1127.6128.1728.33
SSIM0.8690.8700.2840.5080.7050.8030.8400.852
LPIPS0.1850.1810.6030.4810.3390.2340.1870.186
", + "bbox": [ + 517, + 416, + 906, + 476 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Novel view synthesis - quantitative results of different methods on MipNerf360 scenes. StochasticSplats is finetuned on StopThePop-trained scenes using 128 SPP, and rendered using varying number of samples.", + "bbox": [ + 511, + 481, + 906, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "between quality and rendering time. Figure 6 further highlights that popping is handled properly. Note that, due to the sorting-free nature of our method, any popping free formulations can be easily integrated into our method.", + "bbox": [ + 511, + 545, + 905, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative results - Tab. 1. We report the PSNR values of our fine-tuned scenes rendered with varying numbers of samples, alongside comparisons to 3DGS and StopThePop. We use our pop compensation from Section 3.4. By increasing the number of samples per pixel, our method approaches the quality of alpha blending. While the Monte Carlo noise affects metric, i.i.d. noise is visually often less distracting than structured artifacts; recent research suggests that the human visual system tolerates some unbiased noise, especially at interactive frame rates [31]. Additionally, as we show in Section 4.4 and Section 4.4, this can be mitigated via temporal anti-aliasing, and even without any mitigation, does not harm downstream perception tasks.", + "bbox": [ + 511, + 608, + 906, + 805 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Computation time", + "text_level": 1, + "bbox": [ + 511, + 815, + 691, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Forward pass timings - Tab. 2. We compare the forward rendering time of our method implemented in OpenGL, against 3DGS on both CUDA and OpenGL, as well as StopThePop [37] on CUDA. We report timings with a vary-", + "bbox": [ + 511, + 839, + 905, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26331", + "bbox": [ + 478, + 940, + 517, + 952 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/192b0f54fc66ad55ba371f2a06a25e87d44ae430868fef44380b6ae0ab01f42f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 95, + 256, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/33e226995a4655fa49dfded435141adbb50643b5b87defd0b5202eb13ee9dc04.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 261, + 95, + 418, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9edb61efb938a063d47fe0b119b30edcdcf6c4daf3f4889818297a648eaf8e17.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 95, + 576, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5bd3908a0499e508975188293c03a2bc78e407c00b4f724229643efebbbfdf3d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 95, + 736, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1c8d299a115e624295e75492b396776f3c85973cd6aa4fda6fd4bc755cf737ee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 740, + 95, + 897, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/343eef2f232e9f5e73e16fbf6b4ee86fb1aff2bd9b74c7240a56e7b0876d1bb4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 181, + 256, + 303 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f125978960767cb7dc86f0f294d0bba3901a5b997f05c27defb21ebb58bd849b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 261, + 181, + 416, + 303 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eb21155a9d8ef7c25de658c93d9a8ebdb16a77c5494905175ef27b422f72fd61.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 181, + 576, + 303 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3d800f0cf180713551722432461845476304e8daa6d475c26de5af787f5faf78.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 181, + 736, + 303 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7270906033fa29db918943972b913e7bf88974470ed0ffb5fcfb2cc468c0d4c0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 740, + 181, + 898, + 303 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8a0281dc118751ff326ef287ff584686bf9b316009f68c13469bd54bbbca535d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 313, + 256, + 395 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/56fbe39da250b90346dd92696bcc253abcbf994a867387578a627b2b36e46f4a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 261, + 313, + 416, + 395 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c3292f860425a4825b58b6df4d9fb490dbf6d0928ac69d94b8708556ab28cd9f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 313, + 576, + 395 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b8ef27816aae8b110ce614fc292bc9162057b49370c69c02d983f3cb104f689a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 313, + 736, + 395 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7a3db570cd5a7f3853e3148ffe739d168119b6e4b5ce247497ae885079e23ca3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 740, + 313, + 898, + 395 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8a20bf59b5b4caa85d52272db8d9e38cf4581e511d94684edb0f500dc4f89dbf.jpg", + "image_caption": [ + "1 SPP" + ], + "image_footnote": [], + "bbox": [ + 98, + 398, + 256, + 521 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6479be1972017f0f3dd3fb18f9b565674c343edf20a730afe88707472a437ff9.jpg", + "image_caption": [ + "4 SPP" + ], + "image_footnote": [], + "bbox": [ + 259, + 398, + 416, + 521 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a8c432404dae02891d5b9007a1ed3441e13c8c9a690410499c55f3e1e20d5aba.jpg", + "image_caption": [ + "16 SPP" + ], + "image_footnote": [], + "bbox": [ + 419, + 398, + 576, + 521 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/481fd1a276bc5f03a09f1f71bcc63411c31d0daeb1aec6ed8cacefa56126913a.jpg", + "image_caption": [ + "128 SPP", + "Figure 7. Qualitative results - for the Room and Kitchen scenes from the MipNerf360 dataset are shown in Fig. 7. We compare StochasticSplats using 1, 4, and 16 samples per pixel against the standard alpha blending approach. Even with just 1 sample per pixel, our method produces reasonable renderings, and with 16 samples, the results closely match the quality of alpha blending." + ], + "image_footnote": [], + "bbox": [ + 581, + 398, + 736, + 521 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c5ccdb50b4e768f9159400504113d6ad01ce558c43d60bb5cc45824e9d37882b.jpg", + "image_caption": [ + "Alpha blending" + ], + "image_footnote": [], + "bbox": [ + 740, + 398, + 898, + 521 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a0c23bab304ec5e573d5d949fd0212333d0868b51f8b099e43e0d2d38b08c893.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
3DGS-CUDA3DGS-OpenGLStopThePop1Ours-OpenGL (SPP)
24816
T100065.08100.2875.9116.2318.9524.7137.5161.25
RTX30908.1432.159.483.254.186.4215.3118.48
RTX40905.6020.70-1.852.052.866.718.00
", + "bbox": [ + 94, + 593, + 480, + 650 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/396e3193c37f4a595c0a8d093d7371084a8edf72f6c7c49d638b9f82a5e46f95.jpg", + "table_caption": [ + "Table 2. Forward pass time - in milliseconds (ms) for different techniques on MipNerf360-trained scenes on three different GPUs. The values are averaged over all scenes and all views." + ], + "table_footnote": [], + "table_body": "
3DGSStopThePop1Ours (SPP)
8128
23.0625.7718.1033.19478.93
", + "bbox": [ + 94, + 696, + 480, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Backward pass time - in ms of the CUDA code for StochasticSplats (as we vary SPP), 3DGS [18], StopThePop [37], averaged over test views of the MipNerf360 scenes on RTX 3090.", + "bbox": [ + 89, + 744, + 482, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing number of samples per pixel, exercising our method's ability to trade off performance and quality. With a single sample, our method achieves $2 \\times$ to $4 \\times$ lower latency compared to alpha-blended 3DGS, depending on the hardware; see Table 2. For higher sample counts, we employ supersampling: render at a higher resolution and then downsample to the target display resolution.", + "bbox": [ + 89, + 794, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our method is unique at being fast, portable and pop-free at the same time. The portability of our method is only comparable with the 3DGS-OpenGL variant, that has popping artifacts and is relatively slow. StopThePop cannot be easily ported to standard graphics frameworks, and 3DGS-CUDA slows down significantly when ported to OpenGL. This is because the CUDA radix sort routines utilize platform specific optimizations [34].", + "bbox": [ + 511, + 595, + 906, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Backward pass timings - Tab. 3. We further report the computation time for a single backward pass, this time with the CUDA implementation for all methods, as training typically is done on CUDA devices.", + "bbox": [ + 511, + 719, + 906, + 780 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Lowering resolution does not enhance speed", + "text_level": 1, + "bbox": [ + 511, + 787, + 885, + 804 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our method inherently offers a performance vs. quality trade-off. Attempting to naively integrate such a trade-off into 3DGS, such as reducing rendering resolution, does not work well. Simply lowering the rendering resolution does not speed up rasterization – the same number of Gaussians must be processed, and the view frustum for each pixel in", + "bbox": [ + 511, + 809, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26332", + "bbox": [ + 478, + 940, + 517, + 952 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1792147d97dde17498b2f881fab67930bcbdb60ea35becefa6f5cb75c8aa60bf.jpg", + "image_caption": [ + "Figure 8. Timings at lower resolution - Render timings (milliseconds) at different downsampled resolutions for our method (OpenGL) and alpha blending (CUDA) with different tile sizes on an NVIDIA RTX 3090." + ], + "image_footnote": [], + "bbox": [ + 109, + 75, + 468, + 280 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f0f2b084fc8229ecb843ceb1f5cd796f67111e6afcd5bea700918d2a2b9fc468.jpg", + "image_caption": [ + "Figure 9. Temporal anti-aliasing - StochasticSplats at 1 SPP, along with FLIP error, without (left) and with (right) TAA after moving along a trajectory for 70 frames. TAA introduces a negligible latency increase. For videos, please refer to our website." + ], + "image_footnote": [], + "bbox": [ + 102, + 340, + 284, + 462 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d5361622742efbae84724c3c7feaecf56f7223a98e662d8a39ada39bc54b496b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 340, + 467, + 462 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "creases. In short, more Gaussians contribute to each pixel, increasing the computation time per pixel. Figure 8 reports rendering times for our method (OpenGL) and alpha blending (AB-CUDA) with various tile sizes. While AB render time decreases to some degree, it starts to increase after a point, as the Gaussian count per-tile increases. No configuration can reach the rendering speed of our method (1 SPP).", + "bbox": [ + 89, + 532, + 483, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Making the most out of a single sample", + "text_level": 1, + "bbox": [ + 89, + 648, + 429, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We now practically demonstrate that even a single-sample rendering strategy can be useful.", + "bbox": [ + 89, + 672, + 482, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Temporal anti-aliasing - Fig. 9. In interactive applications, multiple frames are rendered in rapid succession. A common strategy to further reduce variance is to average pixel values across frames by reusing information from previous renderings [40]. We employ temporal anti-aliasing (TAA) to not only mitigate aliasing artifacts, but also accumulate samples over time, achieving an effectively higher sample count at minimal cost. Frame-to-frame consistency is maintained by warping previously rendered frames according to per-pixel depth whenever the camera moves. As shown in Figure 9, this greatly reduces rendering noise, while maintaining low latency. We provide implementation details in the appendix, and our supplemental", + "bbox": [ + 89, + 704, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c7e624f212b59042de5480de38f0c349a6f82f1718c4ea6d6a88cdee73b1c168.jpg", + "image_caption": [ + "(a) Green apple (AB)" + ], + "image_footnote": [], + "bbox": [ + 532, + 88, + 709, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5c3439c8a5b405cb0b083a2d132341d29b3c771eb2574080751ec52527110482.jpg", + "image_caption": [ + "(b) Green apple (ours)" + ], + "image_footnote": [], + "bbox": [ + 709, + 89, + 887, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/256f2849db8bfcfc9da827861d9aea1c5060e082ec79c139b290477426cc15e2.jpg", + "image_caption": [ + "(c) Chopsticks (AB)" + ], + "image_footnote": [], + "bbox": [ + 531, + 205, + 707, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/abed106002d109459632ac80b5f89dd4c7af83d0066d103b404243e163ae901f.jpg", + "image_caption": [ + "(d) Chopsticks (ours)" + ], + "image_footnote": [], + "bbox": [ + 709, + 205, + 885, + 306 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/900d4e6d5d2d7c0d80ecd3ae6dd975866f957254b792cc880081b40211e0702f.jpg", + "table_caption": [ + "Figure 10. Open vocabulary localization - Qualitative comparison of open-vocabulary 3D object localization on the LERF dataset using alpha blending vs. StochasticSplats at 1 SPP." + ], + "table_footnote": [], + "table_body": "
TeatimeRamenWaldo-KitchenFigurinesMean
AB88.173.295.580.484.3
Ours88.173.286.480.482.0
", + "bbox": [ + 516, + 367, + 903, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Open vocabulary localization - accuracy on the LERF dataset using alpha blending (AB) and StochasticSplats at 1 SPP.", + "bbox": [ + 511, + 417, + 906, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "videos demonstrate TAA applied to 1-SPP frames.", + "bbox": [ + 511, + 453, + 852, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Semantic localization - Fig. 10 and Tab. 4. We further show that while the single-sample render has low PSNR, it is still useful. Specifically, we use LangSplat [36] and replace their alpha blending renderer with our method, and follow their pipeline to localize objects. As shown in Figure 10, while our renders are noisy, it still properly renders semantics such that the localizer can correctly identify its location. There is a moderate reduction in accuracy (Tab. 4), but we achieve $4 \\times$ faster rendering speed.", + "bbox": [ + 511, + 470, + 906, + 608 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 613, + 633, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We proposed a sorting-free stochastic rendering method for 3D Gaussian splatting, that is, technically, compatible with any semi-transparent representation. We derive both the forward (rendering) and backward (gradient) passes for stochastic rendering, allowing a quality vs. cost trade off, and renders without popping artifacts. Our OpenGL implementation renders $2 \\times -4 \\times$ faster with reasonable quality, demonstrating high portability.", + "bbox": [ + 511, + 638, + 906, + 760 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and future works. While we obtain reasonable quality when rendering with low sample counts, our results are noisy. Future work could focus on reducing the variance of both forward and backward estimators, for example using importance sampling, stratified sampling [21] or control variates. Moreover, quality could further be improved by leveraging lightweight (neural) denoising algorithms, similar to how diffusion models are used in conjunction with rough 3D renderings [39, 48].", + "bbox": [ + 511, + 761, + 908, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26333", + "bbox": [ + 478, + 940, + 519, + 952 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Pontus Andersson, Jim Nilsson, Tomas Akenine-Moller, Magnus Oskarsson, Kalle Åström, and Mark D. Fairchild. FLIP: A Difference Evaluator for Alternating Images. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 3(2):15:1-15:23, 2020. 6", + "[2] Dejan Azinovic, Tzu-Mao Li, Anton Kaplanyan, and Matthias Nießner. Inverse path tracing for joint material and lighting estimation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2442-2451, 2019. 5", + "[3] Milena T Bagdasarian, Paul Knoll, Yi-Hsin Li, Florian Barthel, Anna Hilsmann, Peter Eisert, and Wieland Morgenstern. 3dgs. zip: A survey on 3d gaussian splatting compression methods. arXiv preprint arXiv:2407.09510, 2024. 3", + "[4] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 6", + "[5] Jorge Condor, Sebastien Speierer, Lukas Bode, Aljaz Bozic, Simon Green, Piotr Didyk, and Adrian Jarabo. Don't splat your gaussians: Volumetric ray-traced primitives for modeling and rendering scattering and emissive media. ACM Trans. Graph., 44(1), 2025. 3, 6", + "[6] Franklin C. Crow. The aliasing problem in computer-generated shaded images. Commun. ACM, 20(11):799-805, 1977. 2", + "[7] Thomas Deliot, Eric Heitz, and Laurent Belcour. Transforming a non-differentiable rasterizer into a differentiable one with stochastic gradient estimation. Proc. Symposium on Interactive 3D Graphics and Games (I3D), 2024. 3", + "[8] Eric Enderton, Erik Sintorn, Peter Shirley, and David Luebke. Stochastic transparency. In Proc. Symposium on Interactive 3D Graphics and Games (I3D), pages 157-164, 2010. 2, 3, 4", + "[9] Cass Everitt. Interactive order-independent transparency. White paper, NVIDIA, 2(6):7, 2001. 2", + "[10] James D Foley. Computer graphics: principles and practice. Addison-Wesley Professional, 1996. 4", + "[11] Michael B. Giles. Monte carlo evaluation of sensitives in computational finance. Technical report, Oxford-Man Institute of Quantitative Finance, 2007. 3", + "[12] Ioannis Gkioulekas, Anat Levin, and Todd Zickler. An evaluation of computational imaging techniques for heterogeneous inverse scattering. In European Conference on Computer Vision (ECCV), pages 685-701. Springer International Publishing, 2016. 3, 5", + "[13] Khronos Group. Early Fragment Test. https://www.khronos.org/opengl/wiki/Early_Fragment_Test, 2025.5", + "[14] Hao Gui, Lin Hu, Rui Chen, Mingxiao Huang, Yuxin Yin, Jin Yang, and Yong Wu. Balanced 3DGS: Gaussian-wise parallelism rendering with fine-grained tiling. arXiv preprint arXiv:2412.17378, 2024. 3", + "[15] Florian Hahlbohm, Fabian Friederichs, Tim Weyrich, Linus Franke, Moritz Kappel, Susana Castillo, Marc Stamminger, Martin Eisemann, and Marcus Magnor. Effi" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "cient perspective-correct 3D Gaussian splatting using hybrid transparency. arXiv preprint arXiv:2410.08129, 2024. 2, 3", + "[16] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, Light, and Material Decomposition from Images using Monte Carlo Rendering and Denoising. Advances in Neural Information Processing Systems (NeurIPS), 2022. 3", + "[17] Qiqi Hou, Randall Rauwendaal, Zifeng Li, Hoang Le, Farzad Farhadzadeh, Fatih Porikli, Alexei Bourd, and Amir Said. Sort-free gaussian splatting via weighted sum rendering. arXiv preprint arXiv:2410.18931, 2024. 3", + "[18] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D gaussian splatting for real-time radiance field rendering. ACM Trans. Graph. (Proc. SIGGRAPH), 42(4), 2023. 2, 3, 6, 7", + "[19] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Leref: Language embedded radiance fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 19729-19739, 2023. 2, 6", + "[20] Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Weiwei Sun, Yang-Che Tseng, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. 3D gaussian splatting as markov chain monte carlo. In Advances in Neural Information Processing Systems (NeurIPS), 2024. Spotlight Presentation. 2", + "[21] Samuli Laine and Tero Karras. Stratified sampling for stochastic transparency. In Comp. Graph. Forum (Proc. EGSR), pages 1197-1204, 2011. 2, 8", + "[22] Tzu-Mao Li, Miika Aittala, Frédo Durand, and Jaakko Lehtinen. Differentiable monte carlo ray tracing through edge sampling. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 37 (6):222:1-222:11, 2018. 3", + "[23] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft Rasterizer: A Differentiable Renderer for Image-Based 3D Reasoning. In IEEE International Conference on Computer Vision (ICCV), pages 7708-7717, 2019. 2", + "[24] Alexander Mai, Dor Verbin, Falko Kuester, and Sara Fridovich-Keil. Neural microfacet fields for inverse rendering. In IEEE International Conference on Computer Vision (ICCV), 2023. 3", + "[25] Alexander Mai, Peter Hedman, George Kopanas, Dor Verbin, David Futschik, Qiangeng Xu, Falko Kuester, Jonathan T Barron, and Yinda Zhang. EVER: Exact volumetric ellipsoid rendering for real-time view synthesis. arXiv preprint arXiv:2410.01804, 2024. 2, 3", + "[26] Saswat Subhajyoti Mallick, Rahul Goel, Bernhard Kerbl, Markus Steinberger, Francisco Vicente Carrasco, and Fernando De La Torre. Taming 3dgs: High-quality radiance fields with limited resources. In SIGGRAPH Asia Conference Papers, pages 1-11, 2024. 3", + "[27] Morgan McGuire and Louis Bavoil. Weighted blended order-independent transparency. Journal of Computer Graphics Techniques (JCGT), 2(2):122-141, 2013. 2", + "[28] Guennady A. Mikhailov. On the calculation of nuclear reactor disturbances by the monte carlo method. USSR Computational Mathematics and Mathematical Physics, 6(2):268-273, 1966. 3" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26334", + "bbox": [ + 478, + 941, + 519, + 952 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Ben Mildenhall, Pratul Srinivasan, Matthew Tancik, Jonathan Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In European Conference on Computer Vision (ECCV), pages 405-421. Springer, 2020. 2", + "[30] Cedrick Münstermann, Stefan Krumpen, Reinhard Klein, and Christoph Peters. Moment-based order-independent transparency. Proc. Symposium on Interactive 3D Graphics and Games (I3D), 1(1):1-20, 2018. 2", + "[31] Vasiliki Myrodia. Psychophysical studies on Monte Carlo rendering-noise visual perception. PhD thesis, Université de Lille, 2021. 6", + "[32] Merlin Nimier-David, Delio Vicini, Tizian Zeltner, and Wenzel Jakob. Mitsuba 2: A retargetable forward and inverse renderer. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 38 (6), 2019. 3", + "[33] Jan Novák, Andrew Selle, and Wojciech Jarosz. Residual ratio tracking for estimating attenuation in participating media. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 33(6):179:1-179:11, 2014. 6", + "[34] Lukas Polok, Viorela Ila, and Pavel Smrz. Fast radix sort for sparse linear algebra on gpu. In *SpringSim (HPS)*, page 11, 2014. 7", + "[35] Thomas Porter and Tom Duff. Compositing digital images. In Computer Graphics (Proc. SIGGRAPH), page 253-259. ACM, 1984. 3", + "[36] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 20051-20060, 2024. 8", + "[37] Lukas Radl, Michael Steiner, Mathias Parger, Alexander Weinrauch, Bernhard Kerbl, and Markus Steinberger. Stophepop: Sorted gaussian splatting for view-consistent real-time rendering. ACM Trans. Graph. (Proc. SIGGRAPH), 43(4):1-17, 2024. 2, 3, 5, 6, 7", + "[38] Ravi Ramamoorthi and Pat Hanrahan. An efficient representation for irradiance environment maps. In ACM Trans. Graph. (Proc. SIGGRAPH), 2001. 3", + "[39] Xuanchi Ren, Tianchang Shen, Jiahui Huang, Huan Ling, Yifan Lu, Merlin Nimier-David, Thomas Müller, Alexander Keller, Sanja Fidler, and Jun Gao. GEN3C: 3d-informed world-consistent video generation with precise camera control. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 8", + "[40] Daniel Scherzer, Lei Yang, Oliver Mattausch, Diego Nehab, Pedro V. Sander, Michael Wimmer, and Elmar Eisemann. Temporal Coherence Methods in Real-Time Rendering. Comp. Graph. Forum, 31(8):2378-2408, 2012. 8", + "[41] Andrea Tagliasacchi and Ben Mildenhall. Volume rendering digest (for nef). arXiv preprint arXiv:2209.02417, 2022. 3", + "[42] Anthony J. Thibault. Splatapult. https://github.com/hyperlogic/splatapult, 2023. Accessed January 2, 2025. 2, 6", + "[43] JM Tregan, S Blanco, J Dauchet, M Hafi, R Fournier, L Ibarrart, P Lapeyre, and N Villefranque. Convergence issues in derivatives of Monte Carlo null-collision integral formulations: a solution. arXiv preprint arXiv:1903.06508, 2019. 3" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Grigoris Tsopouridis, Andreas A Vasilakis, and Ioannis Fudos. Deep and fast approximate order independent transparency. In Comp. Graph. Forum, 2024. 2", + "[45] Dor Verbin, Ben Mildenhall, Peter Hedman, Jonathan T. Barron, Todd Zickler, and Pratul P. Srinivasan. Eclipse: Disambiguating illumination and materials using unintended shadows. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3", + "[46] Delio Vicini, Sébastien Speierer, and Wenzel Jakob. Path replay backpropagation: Differentiating light paths using constant memory and linear time. ACM Trans. Graph. (Proc. SIGGRAPH), 40(4):108:1-108:14, 2021. 3, 4, 5", + "[47] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Proc., 13(4), 2004. 6", + "[48] Jay Zhangjie Wu, Yuxuan Zhang, Haithem Turki, Xuanchi Ren, Jun Gao, Mike Zheng Shou, Sanja Fidler, Zan Gojcic, and Huan Ling. Difix3D+: Improving 3d reconstructions with single-step diffusion models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 8", + "[49] Chris Wyman. Stochastic layered alpha blending. In ACM SIGGRAPH 2016 Talks, pages 1-2. 2016. 2", + "[50] Jason C Yang, Justin Hensley, Holger Grün, and Nicolas Thibierz. Real-time concurrent linked list construction on thegpu. In Comp. Graph. Forum (Proc. EGSR), 2010. 2", + "[51] Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jeffrey Hu, Matthew Tancik, and Angjoo Kanazawa. gsplat: An open-source library for Gaussian splatting. arXiv preprint arXiv:2409.06765, 2024.3", + "[52] Ekrem Fatih Yilmazer, Delio Vicini, and Wenzel Jakob. Solving inverse pde problems using monte carlo estimators. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 43, 2024. 3", + "[53] Tizian Zeltner, Sébastien Speierer, Iliyan Georgiev, and Wenzel Jakob. Monte carlo estimators for differential light transport. Transactions on Graphics (Proceedings of SIGGRAPH), 40(4), 2021. 4", + "[54] Cheng Zhang, Lifan Wu, Changxi Zheng, Ioannis Gkioulekas, Ravi Ramamoorthi, and Shuang Zhao. A differential theory of radiative transfer. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 38(6), 2019. 3", + "[55] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6", + "[56] Shuang Zhao, Lifan Wu, Frédo Durand, and Ravi Ramamoorthi. Downsampling scattering parameters for rendering anisotropic media. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 35(6), 2016. 3", + "[57] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Surface splatting. In ACM Trans. Graph. (Proc. SIGGRAPH), 2001. 2", + "[58] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In Proceedings Visualization (VIS), pages 29-538. IEEE, 2001. 2, 3" + ], + "bbox": [ + 516, + 92, + 903, + 871 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26335", + "bbox": [ + 478, + 940, + 517, + 952 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/StochasticSplats_ Stochastic Rasterization for Sorting-Free 3D Gaussian Splatting/d963f664-7a0a-43e1-95cb-e52f6169cd94_model.json b/2025/StochasticSplats_ Stochastic Rasterization for Sorting-Free 3D Gaussian Splatting/d963f664-7a0a-43e1-95cb-e52f6169cd94_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a9abbc24cc3a1766124da4098b7284f2bfe63e10 --- /dev/null +++ b/2025/StochasticSplats_ Stochastic Rasterization for Sorting-Free 3D Gaussian Splatting/d963f664-7a0a-43e1-95cb-e52f6169cd94_model.json @@ -0,0 +1,2882 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.176, + 0.131, + 0.823, + 0.177 + ], + "angle": 0, + "content": "Stochastic Splats: Stochastic Rasterization for Sorting-Free 3D Gaussian Splatting" + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.203, + 0.799, + 0.241 + ], + "angle": 0, + "content": "Shakiba Kheradmand\\*1,2, Delio Vicini\\*3, George Kopanas\\*3,4, Dmitry Lagun\\(^{1}\\), Kwang Moo Yi\\(^{2}\\), Mark Matthews\\(^{1}\\), Andrea Tagliasacchi\\(^{1,5,6}\\)" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.253, + 0.807, + 0.308 + ], + "angle": 0, + "content": "1Google DeepMind, 2University of British Columbia, 3Google, 4Runway ML, 5Simon Fraser University, 6University of Toronto, *equal contribution,†work done at Google" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.324, + 0.896, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.601, + 0.908, + 0.673 + ], + "angle": 0, + "content": "Figure 1. Teaser - We render a 3DGS radiance field using an unbiased Monte Carlo estimator of the volume rendering equation. Unlike prior radiance field approaches, our method does not require sorted back-to-front rendering. This results in fast, portable, and pop-free rendering that is easy to implement. The number of Monte Carlo samples per pixel allows a trade-off between interactivity and visual quality, similar to physically-based rendering. From left to right, we increase the number of samples per pixel, which increases visual fidelity at a predictable increase in rendering cost. We report the mean square error (MSE) and render time on an NVIDIA RTX 4090." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.682, + 0.329, + 0.699 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.713, + 0.486, + 0.895 + ], + "angle": 0, + "content": "3D Gaussian splatting (3DGS) is a popular radiance field method, with many application-specific extensions. Most variants rely on the same core algorithm: depth-sorting of Gaussian splats then rasterizing in primitive order. This ensures correct alpha compositing, but can cause rendering artifacts due to built-in approximations. Moreover, for a fixed representation, sorted rendering offers little control over render cost and visual fidelity. For example, and counter-intuitively, rendering a lower-resolution image is not necessarily faster. In this work, we address the above limitations by combining 3D Gaussian splatting with stochastic rasterization. Concretely, we leverage an unbi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.683, + 0.909, + 0.851 + ], + "angle": 0, + "content": "ased Monte Carlo estimator of the volume rendering equation. This removes the need for sorting, and allows for accurate 3D blending of overlapping Gaussians. The number of Monte Carlo samples further imbues 3DGS with a way to trade off computation time and quality. We implement our method using OpenGL shaders, enabling efficient rendering on modern GPU hardware. At a reasonable visual quality, our method renders more than four times faster than sorted rasterization. The project page and source code are publicly available at https://ubc-vision.github.io/stochasticsplats/." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.941, + 0.521, + 0.953 + ], + "angle": 0, + "content": "26326" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.09, + 0.223, + 0.106 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.236 + ], + "angle": 0, + "content": "Recent advances in novel view synthesis have been driven by the success of neural radiance fields (NeRF) [29] and 3D Gaussian splatting (3DGS) [18]. The former produced high-quality visuals at a low rendering speed, while the latter enabled efficient rendering with comparable visual quality. 3DGS represents a scene as a set of anisotropic 3D Gaussians that can be \"splatted\" [57] in rasterization pipelines, enabling real-time viewing and efficient training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.237, + 0.484, + 0.523 + ], + "angle": 0, + "content": "In 3DGS, 3D Gaussians are flattened onto camera-oriented billboards that not only prevent accurate volumetric intermixing [25], but must also be depthwise sorted for rasterization [58]. This global sorting operation is computationally expensive and introduces popping artifacts [37], as the sort order is view-dependent: small camera movements can result in temporal discontinuities in the rendered images. Further, there is no straightforward way to balance rendering cost vs. visual fidelity. One might assume that reducing the rendering resolution would lead to faster rasterization... yet, we observe the opposite in our experiments. Overall, this limits 3DGS's use on low-end hardware and in latency-critical applications, such as robotics and autonomous driving. Finally, most implementation use custom software rasterizers or GPGPU frameworks such as CUDA. Software rasterizers are notoriously hard to optimize, and often cannot match the performance of the hardware stack [23]. Additionally, many GPGPU frameworks are vendor-specific, reducing portability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.484, + 0.81 + ], + "angle": 0, + "content": "Several previous works address popping artifacts, rendering performance, and portability, but typically improve only one, at the expense of the others. Popping can be alleviated by per-pixel depth sorting, but this introduces extra computational demands. While this cost can be amortized using clever spatial biases [37], per-pixel sorting cannot fully solve volumetric intermixing of 3D Gaussians. Constant density ellipsoids allow analytical intermixing [25], but are slow to render because they replace sorting with expensive ray tracing. Other methods [15] reduce the rendering cost and the popping artifacts by using a mix of per-pixel sorted lists and weighted order-independent transparency, but this comes at a cost in terms of portability since they need to implement complicated GPGPU kernels hindering cross-platform adoption. An orthogonal route to reduce render cost is to reduce the number of Gaussians [20]. Finally, packages such as Splatapult [42] implement an OpenGL-based rasterization pipeline, but do not outperform the original CUDA 3DGS implementation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We instead propose to render 3D Gaussians using stochastic transparency (ST) [8]; see Figure 2. Stochastic transparency replaces sorted alpha blending by an unbiased Monte Carlo estimator. Practically, transparency is rendered by probabilistically selecting individual, opaque, samples that are averaged together with multi-sampling [6]." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.701, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.174, + 0.697, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.278, + 0.663, + 0.291 + ], + "angle": 0, + "content": "Alpha blending" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.092, + 0.902, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.177, + 0.902, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.719, + 0.278, + 0.908, + 0.291 + ], + "angle": 0, + "content": "Stochastic transparency (4 SPP)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.296, + 0.908, + 0.339 + ], + "angle": 0, + "content": "Figure 2. Blending vs. stochastic - We visualize a 1D and 3D example of the different ways to compute transparency. Two foreground Gaussians are composited on a white background." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.349, + 0.907, + 0.53 + ], + "angle": 0, + "content": "Individual samples are rendered using order-independent Z-buffering, without sorting. We provide unbiased Monte-Carlo estimators of both the forward alpha blending and the gradient computation. The reverse-mode gradient estimator also permits sorting-free training and is largely agnostic to the primitive type and could be used for other semitransparent representations (e.g., triangle soups). We also do an early analysis on proper volumetric intermixing of 3D Gaussians by extending the theory of stochastic transparency. Critically, all of the above is naturally compatible with the hardware rasterization pipeline, allowing us a portable OpenGL implementation of our method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.531, + 0.909, + 0.682 + ], + "angle": 0, + "content": "At one sample per pixel (SPP), our OpenGL implementation renders up to \\(4 \\times\\) faster than the original CUDA implementation, enabling latency-critical applications, such as open-vocabulary 3D object localization [19], which we demonstrate. Higher SPP increase quality and render time, approaching the PSNR of alpha-blended 3DGS. One key advantage is that downstream applications can dynamically adjust sampling at rendering time, allowing them to potentially trade-off latency against quality with careful engineering; see Figure 1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.694, + 0.655, + 0.71 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.885 + ], + "angle": 0, + "content": "Order-independent transparency. While current 3DGS methods rely on sorted alpha blending, there is a rich literature on order-independent transparency for rasterization-based real-time rendering. Prior work can be classified into: (1) stochastic methods [8, 21, 49], that probabilistically discard fragments based on their opacity to create Monte Carlo estimators of the alpha blending equation and (2) non-stochastic methods that avoid sorting by using per-pixel linked lists [50], depth peeling [9], weighted blending [27], moments [30], or deep learning [44] to (approximately) composite transparent fragments." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Given the success of order-independent transparency in" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.941, + 0.519, + 0.953 + ], + "angle": 0, + "content": "26327" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "real-time graphics, it becomes desirable to incorporate these techniques with 3DGS. Recent pre-prints [15, 17] combine 3DGS with non-stochastic order-independent transparency, but stochastic variants remain unexplored. We bridge this gap by extensively studying stochastic transparency with 3DGS, both in terms of rendering and training, enabled by our novel stochastic gradient estimator." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.2, + 0.483, + 0.517 + ], + "angle": 0, + "content": "Popping artifacts. Real-time graphics eschew sorting primitives due to both the computational cost and the inexistence of a consistent global ordering of non-trivial primitives. In 3DGS this is overlooked, because most primitives have a relatively small spatial extent. This introduces artifacts often dubbed as \"popping\", where a small motion of the camera forces a change in the order of the primitives introducing a sharp discontinuity in the alpha blending equation. These artifacts are particularly noticeable in immersive contexts like virtual reality. Radl et al. [37] address this using per-pixel sorting based on the maximal Gaussian response. While this eliminates popping artifacts, it introduces prohibitive computational costs, necessitating complex approximate sorting through a multi-stage hierarchical renderer. EVER [25] addresses popping artifacts by ray tracing constant density ellipsoids to compute the full volume rendering equation, rather than approximate alpha-blending. However, this approach is significantly slower than rasterization. In contrast, stochastic transparency uniquely eliminates popping by trading sorting for some noise in the final image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.483, + 0.671 + ], + "angle": 0, + "content": "Fast 3DGS rendering and training. While 3DGS [18] offered a significant speedup over NeRF-based methods, subsequent works have further improved performance. Numerous improvements have been made to the implementation of training and rendering routines [14, 15, 26, 51]. Further speedups can be made by reducing the number of Gaussians and compressing their attributes. This is covered in an extensive literature review [3]. Any approach that reduces the number of primitives readily complements our work and is likely to improve rendering performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.885 + ], + "angle": 0, + "content": "Differentiable Monte Carlo methods. We introduce a novel Monte Carlo gradient estimator for stochastic transparency. Such differentiable Monte Carlo estimators are commonly used in physically-based inverse rendering [12, 22, 32, 46, 54, 56]. Various works propose differentiable Monte Carlo methods for neural radiance fields [24, 45] or surface reconstruction [16], and Condor et al. [5] differentiate physically-based Monte Carlo ray tracing of 3D Gaussian primitives. Deliot et al. [7] applied stochastic simultaneous perturbation gradients to real-time rendering algorithms, including standard 3DGS. Beyond rendering, differentiable Monte Carlo methods find application in nuclear engineering [28], finance [11], atmospheric physics [43] and inverse partial differential equation problems [52]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.605, + 0.105 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.907, + 0.268 + ], + "angle": 0, + "content": "Kerbl et al. [18] represents a 3D scene in 3DGS as a collection of anisotropic 3D Gaussians \\(\\mathcal{G} = \\{g_i\\}\\). Given a set of posed images \\(\\{\\mathbf{I}_n\\}\\), we optimize the parameters of the 3D Gaussians \\(\\mathcal{G}\\) to photometrically reconstruct the scene. Each \\(g_i\\) is parameterized as \\(\\theta = \\{\\mu, \\Sigma, \\alpha, \\mathbf{c}\\}\\), where \\(\\mu\\) is the mean, \\(\\Sigma\\) is the covariance matrix, \\(\\alpha\\) is the opacity, and \\(\\mathbf{c}\\) is the view-dependent color modeled using spherical harmonics [38]. We start by reviewing the classical alpha-blending based rendering of 3DGS in Section 3.1, and then introduce our stochastic transparency algorithm in Section 3.2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.277, + 0.722, + 0.293 + ], + "angle": 0, + "content": "3.1. Sorted alpha blending" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.299, + 0.907, + 0.346 + ], + "angle": 0, + "content": "To calculate the radiance \\(C\\) at a given pixel, we aggregate the radiances emitted by \\(L\\) primitives that intersect the associated ray, blending them in depth \\((z)\\) order [35]:" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.356, + 0.907, + 0.4 + ], + "angle": 0, + "content": "\\[\nC = \\sum_ {i = 1} ^ {L} c _ {i} \\alpha_ {i} \\prod_ {z _ {k} < z _ {i}} (1 - \\alpha_ {k}) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.906, + 0.516 + ], + "angle": 0, + "content": "where \\( z_{k} < z_{i} \\) denotes all primitives in front of primitive \\( i \\). While it can be shown that this expression derives from volume rendering [41], the product of terms in (1) requires sorting the primitives along \\( z \\), then blend radiances in either back-to-front or front-to-back order. As sorting is independent of how many pixels one needs to render, we simply lose the ability to trade off accuracy with speed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.907, + 0.622 + ], + "angle": 0, + "content": "Sorting is done with respect to the \\(z\\) coordinates of the Gaussian means \\(\\mu_{k}\\) remapped to the coordinate frame of the current camera, and the covariance is projected via a linear approximation of the non-linear projection process to the camera's view point, as proposed by Zwicker et al. [58]. This effectively treats each Gaussian as a billboard, flattening it parallel to the camera's imaging plane." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.632, + 0.734, + 0.647 + ], + "angle": 0, + "content": "3.2. Stochastic transparency" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.653, + 0.906, + 0.759 + ], + "angle": 0, + "content": "Stochastic transparency (ST) [8] avoids sorting by stochastically estimating alpha blending in (1). Avoiding sorting when rendering transparent elements is critical, because (i) primitives often do not have a global unique sorting order across all pixels, (ii) sorting can be computationally expensive, and (iii) in the context of volume rendering the 3D extent of primitives influences the intermixing of colors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.76, + 0.907, + 0.865 + ], + "angle": 0, + "content": "Stochastic transparency approximates transparency by randomly assigning a number of fully opaque samples per pixel proportional to a fragment's opacity, and averaging these samples to produce the final pixel color without sorting. In our context, the opacity of a fragment is proportional to the opacity of the Gaussian \\(\\alpha\\), and to its distance from the mean \\(\\mu\\), as measured by the projected covariance" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.906, + 0.901 + ], + "angle": 0, + "content": "1Note that as we do not use multi-sampling anti-aliasing, making fragments synonymous with pixels." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.941, + 0.521, + 0.953 + ], + "angle": 0, + "content": "26328" + } + ], + [ + { + "type": "code", + "bbox": [ + 0.117, + 0.092, + 0.388, + 0.228 + ], + "angle": 0, + "content": "\\(\\boldsymbol {z}\\leftarrow \\infty\\) \\(\\mathbf{C}\\gets c_{\\mathrm{background}}\\) \nfor each Gaussian \\(i\\) in unsorted order do Compute \\(\\alpha_{i}\\) and depth \\(z_{i}\\) Sample \\(u\\sim \\mathcal{U}(0,1)\\) if \\(u < \\alpha_{i}\\) and \\(z_{i} < z\\) then \\(\\mathbf{C}\\gets c_i\\) \\(z\\gets z_{i}\\) \nend if \nend for" + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.234, + 0.436, + 0.248 + ], + "angle": 0, + "content": "Listing 1. Pseudocode for stochastic transparency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.26, + 0.483, + 0.321 + ], + "angle": 0, + "content": "matrix \\(\\Sigma\\). Stochastic transparency can also be interpreted as a randomized extension of classical screen-door transparency [10, Chapter 16], where, instead of a randomized selection, a fixed pattern is employed to discard fragments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.323, + 0.484, + 0.414 + ], + "angle": 0, + "content": "Simple example. As a simple example, consider a single 3D Gaussian \\( g \\) covering pixel \\( p \\) with opacity \\( \\alpha = 0.5 \\); in this case, if we sample \\( x \\sim \\mathcal{U}(0,1) \\) and accept \\( g \\) only when \\( x > \\alpha \\), approximately \\( 50\\% \\) of the pixel's samples will randomly select this Gaussian. We next formalize this idea to generalize to rendering an arbitrary number of Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.417, + 0.484, + 0.462 + ], + "angle": 0, + "content": "Formalization. Formally, stochastic transparency is a Monte Carlo estimator of (1). It estimates the radiance \\( C \\) by randomly sampling a summand \\( i \\) (i.e., the \\( i \\)-th Gaussian):" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.467, + 0.483, + 0.502 + ], + "angle": 0, + "content": "\\[\nC \\approx \\frac {1}{P (i)} c _ {i} \\alpha_ {i} \\prod_ {z _ {k} < z _ {i}} \\left(1 - \\alpha_ {k}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.484, + 0.556 + ], + "angle": 0, + "content": "where \\( P(i) \\) is the probability mass function of the random index \\( i \\). This estimator is unbiased, as long as \\( P(i) > 0 \\) for any non-zero summand. The standard approach is to use:" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.561, + 0.483, + 0.592 + ], + "angle": 0, + "content": "\\[\nP (i) = \\alpha_ {i} \\prod_ {z _ {k} < z _ {i}} \\left(1 - \\alpha_ {k}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.629 + ], + "angle": 0, + "content": "The division by this probability mass function cancels out part of the sample weight, leading to a simple estimator:" + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.635, + 0.482, + 0.651 + ], + "angle": 0, + "content": "\\[\nC \\approx c _ {i} \\quad \\text {f o r} \\quad i \\sim P. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.703 + ], + "angle": 0, + "content": "Thus, a pixel has the probability \\( P(C = c_i) = P(i) \\) to pick color \\( c_i \\). In practice, we average multiple independent samples per pixel to reduce estimator noise." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.841 + ], + "angle": 0, + "content": "Pseudocode. Listing 1 provides pseudocode for the above estimator. By testing that the depth of the current sample is smaller than the previously selected one, we sample exactly according to \\( P(i) \\) [8]. In the fragment Shader, this is implemented by using standard depth testing and calling discard on samples where \\( u \\geq \\alpha_i \\). This only requires access to the current Gaussian's opacity and works for any primitive ordering. Figure 2 further shows a toy example of using this estimator on two Gaussians." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.419, + 0.866 + ], + "angle": 0, + "content": "3.3. Differentiable stochastic transparency" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Applying stochastic transparency to inverse rendering problems requires computing the gradient of the photometric" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.646, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.237, + 0.633, + 0.251 + ], + "angle": 0, + "content": "Alpha blending" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.091, + 0.774, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.237, + 0.76, + 0.251 + ], + "angle": 0, + "content": "Ours (128 SPP)" + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.091, + 0.903, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.792, + 0.237, + 0.887, + 0.251 + ], + "angle": 0, + "content": "Ours (512 SPP)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.256, + 0.906, + 0.313 + ], + "angle": 0, + "content": "Figure 3. Gradient validation - We compare alpha blending gradients to our stochastic estimator with 128 and 512 SPP. Red and blue colors encode positive and negative values, respectively. Our estimator accurately approximates the alpha blending gradients." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.319, + 0.906, + 0.365 + ], + "angle": 0, + "content": "loss \\(\\mathcal{L}\\) with respect to the primitive parameters \\(\\theta\\). Primarily, we require gradients with respect to color \\(\\partial \\mathcal{L} / \\partial c\\), and gradients with respect to opacity \\(\\partial \\mathcal{L} / \\partial \\alpha\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.367, + 0.906, + 0.442 + ], + "angle": 0, + "content": "Gradient estimator. We take inspiration from the differentiable rendering literature [46], and use a detached gradient estimator [46, 53] to correctly differentiate the binary sampling decision in Listing 1. We derive this estimator by applying the gradient operator to the alpha blending equation:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.451, + 0.905, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\boldsymbol {\\theta}} C = \\sum_ {i = 1} ^ {L} \\nabla_ {\\boldsymbol {\\theta}} \\left[ c _ {i} \\alpha_ {i} \\prod_ {z _ {k} < z _ {i}} (1 - \\alpha_ {k}) \\right]. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.906, + 0.547 + ], + "angle": 0, + "content": "We then form an unbiased Monte Carlo gradient estimator by again sampling a summand according to \\( P(i) \\). Crucially, we explicitly do not differentiate \\( P(i) \\) itself and obtain:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.555, + 0.905, + 0.597 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\boldsymbol {\\theta}} C \\approx \\frac {1}{P (i)} \\nabla_ {\\boldsymbol {\\theta}} \\left[ c _ {i} \\alpha_ {i} \\prod_ {z _ {k} < z _ {i}} \\left(1 - \\alpha_ {k}\\right) \\right]. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.605, + 0.904, + 0.636 + ], + "angle": 0, + "content": "Substituting the definition of \\(P(i)\\) from (3) and differentiating \\(\\mathcal{L}\\) with respect to \\(c\\) and \\(\\alpha\\), we get:" + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.643, + 0.905, + 0.676 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\mathcal {L}}{\\partial c _ {i}} = \\frac {\\partial \\mathcal {L}}{\\partial C} \\quad \\frac {\\partial \\mathcal {L}}{\\partial c _ {z _ {k} < z _ {i}}} = 0 \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.677, + 0.905, + 0.711 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\mathcal {L}}{\\partial \\alpha_ {i}} = \\frac {\\partial \\mathcal {L}}{\\partial C} \\frac {c _ {i}}{\\alpha_ {i}} \\quad \\frac {\\partial \\mathcal {L}}{\\partial \\alpha_ {z _ {k} < z _ {i}}} = \\frac {\\partial \\mathcal {L}}{\\partial C} \\frac {- c _ {i}}{1 - \\alpha_ {z _ {k} < z _ {i}}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.906, + 0.868 + ], + "angle": 0, + "content": "where subscript \\(i\\) denotes the sampled primitive, and \\(\\partial \\mathcal{L} / \\partial C\\) is the photometric loss gradient, dependent upon the loss formulation (L1, SSIM, ...). Note how in (7) only the \\(i\\)-th Gaussian receives gradients with respect to its color, while the others have a zero gradient. Meanwhile, (8) shows how the \\(i\\)-th Gaussian, and ones closer to the camera receive a gradient with respect to their opacity. The Gaussians behind the sampled one do not receive any gradient. Finally, our derivation is not specific to 3DGS, and applies to any representation rendered with stochastic transparency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.9 + ], + "angle": 0, + "content": "Decorrelated loss gradient. A well-known challenge for differentiable Monte Carlo methods is that the derivative of" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.941, + 0.519, + 0.953 + ], + "angle": 0, + "content": "26329" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.287, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.174, + 0.2, + 0.214, + 0.213 + ], + "angle": 0, + "content": "3DGS" + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.088, + 0.481, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.308, + 0.2, + 0.465, + 0.214 + ], + "angle": 0, + "content": "StopThePop [37] vs. ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.219, + 0.484, + 0.346 + ], + "angle": 0, + "content": "Figure 4. Popping in 3DGS - (left) As sorting in 3DGS is done with respect to the \\( z \\) distance of the Gaussian mean from the camera, a small camera rotation can cause visible \"popping\" artifacts (i.e., sudden pixel color changes). (right) StopThePop [37] corrects for this behavior by associating a surface with each Gaussian, and determining the depth \\( z \\) per-fragment (dashed line), rather than per-Gaussian. Our solution leads to visually comparable results, but approximates this surface linearly (solid line), so that per-fragment depth \\( z \\) can be computed efficiently in hardware." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.357, + 0.484, + 0.388 + ], + "angle": 0, + "content": "the loss function is a product of two terms that are corrupted by noise (i.e., they are both random variables):" + }, + { + "type": "equation", + "bbox": [ + 0.233, + 0.393, + 0.483, + 0.424 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\mathcal {L}}{\\partial \\theta} = \\frac {\\partial \\mathcal {L}}{\\partial C} \\frac {\\partial C}{\\partial \\theta}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.429, + 0.484, + 0.549 + ], + "angle": 0, + "content": "where \\(\\theta\\) are the geometry parameters. If the noise in both terms is correlated, we get biased gradients, which harm optimization. Following [2, 12], we use different random samples to evaluate the two factors above. If \\(\\mathcal{L}\\) is the L2 loss, the decorrelated evaluation results in unbiased gradients. For other loss functions, it at least reduces bias. Unbiased gradient estimation for arbitrary loss functions applied to a Monte Carlo estimator is an open problem." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.552, + 0.484, + 0.612 + ], + "angle": 0, + "content": "Practical algorithm. The presence of \\( c_{i} \\) in (8) implies that we need access to the result of the forward pass to compute gradients. We therefore follow path replay backpropagation [46] and compute gradients in three passes:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.612, + 0.483, + 0.627 + ], + "angle": 0, + "content": "1. In a first pass, we render an image and evaluate \\(\\partial \\mathcal{L} / \\partial C\\)." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.627, + 0.483, + 0.657 + ], + "angle": 0, + "content": "2. We then render a decorrelated image using a different random seed, giving us \\( c_{i} \\)" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.658, + 0.483, + 0.703 + ], + "angle": 0, + "content": "3. Finally, we replay the second pass by re-rendering using the same random seed and using the stored \\( c_{i} \\) to compute parameter gradients." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.612, + 0.483, + 0.703 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.841 + ], + "angle": 0, + "content": "Validation. We validate our estimator by comparing to the gradients of the sorted alpha blending implementation. For easier visual interpretation, we use our reverse-mode gradient implementation to compute a gradient image. We do this by backpropagating an L1 loss separately for each pixel, and visualizing the sum of the x-component of the positional gradients of all Gaussians affecting a given pixel. As shown in Figure 3, our method produces noisy gradients that closely match those of alpha blending in structure." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.342, + 0.866 + ], + "angle": 0, + "content": "3.4. Removing popping artifacts" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "A naive implementation of stochastic transparency re-uses the billboard geometry of 3DGS, and consequently, to com" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.093, + 0.646, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.178, + 0.629, + 0.191 + ], + "angle": 0, + "content": "Alpha blending" + }, + { + "type": "image", + "bbox": [ + 0.649, + 0.093, + 0.772, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.681, + 0.178, + 0.736, + 0.19 + ], + "angle": 0, + "content": "Pop-free" + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.093, + 0.899, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.775, + 0.178, + 0.897, + 0.191 + ], + "angle": 0, + "content": "Our full intermixing" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.193, + 0.905, + 0.248 + ], + "angle": 0, + "content": "Figure 5. Fully volumetric intermixing - While alpha blending (left) suffers from popping and \"pop-free\" variants (middle) show discontinuities, our full volumetric intermixing (right) accurately composes overlapping Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.255, + 0.905, + 0.36 + ], + "angle": 0, + "content": "pute the depths \\( z_{i} \\), the means of the Gaussians. As we will see in Section 4.2, at low sample count this provides high rendering efficiency, as we do not require a global sort. However, popping artifacts still exist, as the computation of \\( z_{i} \\) is analogous to the one in 3DGS, small changes in the camera rotation can therefore change the order of primitives, and cause visible video artifacts; see Figure 4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.363, + 0.906, + 0.59 + ], + "angle": 0, + "content": "The stop-the-pop way. Radl et al. [37] overcome the popping problem by computing the per-fragment (vs. per-Gaussian) depth. Their core idea is to compute the location of the peak of the Gaussian density along the ray. When coupled with a (computationally expensive) per-pixel sorting of Gaussians, this can provide a pop-free rendering. This can be interpreted as finding \"maximum density surface\" of the Gaussian, visualized in 2D in Figure 4, and using it to define fragment depths. However, per-fragment updates of depth are not particularly efficient, as early-discard mechanisms implemented in hardware [13] can only execute for fragments whose depth has not been modified by a shader. While in our setting we can forego sorting, the inability to early-discard fragments still affects us, which lead us to propose a different approach, described next." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.592, + 0.906, + 0.667 + ], + "angle": 0, + "content": "Our simplified way. We propose to avoid this inefficiency by foregoing per-fragment depth changes. Instead, we orient our billboards to linearly approximate the maximum density surface. If the ray associated with a pixel has origin o, the plane can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.672, + 0.905, + 0.693 + ], + "angle": 0, + "content": "\\[\n\\mathbf {n} ^ {\\top} (\\mathbf {x} - \\boldsymbol {\\mu}) = 0, \\quad \\text {w h e r e} \\quad \\mathbf {n} = \\Sigma^ {- 1} (\\boldsymbol {\\mu} - \\mathbf {o}) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.697, + 0.906, + 0.773 + ], + "angle": 0, + "content": "By approximating this surface with a plane, we can easily use the interpolation kernels within the hardware rasterizer to compute the splat's depth. This approach adds no extra time to the render process and gives us popping-free rendering with a similar PSNR to [37] at higher sample counts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.775, + 0.906, + 0.905 + ], + "angle": 0, + "content": "Fully volumetric intermixing. If maximum rendering fidelity is desired and the cost of per-fragment depth modification is acceptable, stochastic transparency can naturally support full volumetric intermixing. The key insight is that this can be achieved by replacing the fixed depth of a Gaussian by a sampled free-flight distance. For this, we first express the total contribution of a single Gaussian using the Beer-Lambert law: \\(\\int_0^\\infty \\sigma_i(\\mathbf{x}_t)\\exp \\left(-\\int_0^t\\sigma_i(\\mathbf{x}_s)\\mathrm{d}s\\right)c_i\\mathrm{d}t,\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.941, + 0.521, + 0.953 + ], + "angle": 0, + "content": "26330" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.306 + ], + "angle": 0, + "content": "where the opacity of a Gaussian is reparametrized using a spatially-varying extinction function \\(\\sigma_{i}\\) and \\(\\mathbf{x}_a\\coloneqq \\mathbf{o} + a\\cdot \\mathbf{d}\\) for a ray \\((\\mathbf{o},\\mathbf{d})\\). We follow [5] and sample a free-flight distance \\(t\\) proportional to \\(p(t) = \\sigma_i(\\mathbf{x}_t)\\exp (-\\int_0^t\\sigma_i(\\mathbf{x}_s)\\mathrm{d}s)\\). To handle overlapping Gaussians, we adapt decomposition tracking [33]: If two Gaussians overlap, we sample a free-flight distance according to each and take the minimum. This accurately resolves overlap of an arbitrary number of Gaussians without additional data structures, sorting, or ray tracing, and readily fits into the stochastic transparency algorithm. We simply need to replace the computation of \\(z_{i}\\) by a probabilistic sampling \\(z_{i}\\sim p(t)\\). We show an example rendering of two overlapping Gaussians in Fig. 5. See the supplemental for the full free-flight sampling routine." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.311, + 0.223, + 0.327 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.335, + 0.484, + 0.412 + ], + "angle": 0, + "content": "We evaluate rendering performance on the real-world scenes from the MipNerf360 [4] dataset. For demonstrating semantic object localization with single-sample rendering we use the LERF dataset [19] - it consists of complex in-the-wild scenes, designed for 3D object localization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.484, + 0.49 + ], + "angle": 0, + "content": "Baselines. For rendering baselines, we compare the performance of our method against the original 3DGS implementation [18] that uses CUDA, as well as Splatapult [42], an equivalent OpenGL implementation. We also compare with StopThePop [37], a pop-free method for 3DGS rendering." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.491, + 0.484, + 0.567 + ], + "angle": 0, + "content": "Metrics. We compute the standard metrics of peak signal-to-noise ratio (PSNR), structural similarity index metric (SSIM) [47], and learned perceptual image patch similarity (LPIPS) [55]. We report the average per-frame processing time for all methods on the MipNerf360 scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.484, + 0.827 + ], + "angle": 0, + "content": "Implementation details. To demonstrate portability, we implement our method in both CUDA and OpenGL. Our CUDA implementation builds on the existing 3DGS framework by Kerbl et al. [18] and our OpenGL implementation builds upon Splatapult [42]. We evaluate on a Windows laptop equipped with an NVIDIA Quadro T1000 Max-Q GPU, and Linux desktops with an NVIDIA RTX 3090 and RTX 4090. Rather than training from scratch, we finetune a trained 3DGS scene with the CUDA implementation of our method for 1000 iterations. Since we are performing fine-tuning, we do not apply densification or pruning. All learning rates for the non-position parameters are kept unchanged, while the positions are optimized using a learning rate of \\(5 \\times 10^{-5}\\) with 128 samples. We keep all other hyperparameters the same. On average, the finetuning process requires approximately 14 minutes on an NVIDIA RTX 3090 GPU for scenes from MipNerf360." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.267, + 0.85 + ], + "angle": 0, + "content": "4.1. Rendering quality" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Qualitative results - Figs. 6 and 7. We show two scenes (Room and Kitchen) in Figure 7 from the MipN-erf360 dataset. As shown, our method allows a trade-off" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.092, + 0.646, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.65, + 0.092, + 0.772, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.092, + 0.899, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.203, + 0.645, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.314, + 0.619, + 0.326 + ], + "angle": 0, + "content": "Frame \\(t\\)" + }, + { + "type": "image", + "bbox": [ + 0.65, + 0.203, + 0.772, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.314, + 0.744, + 0.326 + ], + "angle": 0, + "content": "Frame \\(t + 1\\)" + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.203, + 0.899, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.794, + 0.314, + 0.859, + 0.326 + ], + "angle": 0, + "content": "FLIP error" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.331, + 0.907, + 0.414 + ], + "angle": 0, + "content": "Figure 6. Popping artifact - The top row displays two consecutive frames, fine-tuned and rendered using our StochasticSplats, along with their corresponding fLIP error map [1]. The bottom row presents the same frames processed with standard 3DGS. The foreground error indicates popping while the background errors are due to the camera motion." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.417, + 0.907, + 0.477 + ], + "angle": 0, + "content": "
3DGSStopThePop14Ours (SPP)
16642561024
PSNR28.9928.7917.9422.7526.1127.6128.1728.33
SSIM0.8690.8700.2840.5080.7050.8030.8400.852
LPIPS0.1850.1810.6030.4810.3390.2340.1870.186
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.482, + 0.907, + 0.539 + ], + "angle": 0, + "content": "Table 1. Novel view synthesis - quantitative results of different methods on MipNerf360 scenes. StochasticSplats is finetuned on StopThePop-trained scenes using 128 SPP, and rendered using varying number of samples." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.546, + 0.906, + 0.606 + ], + "angle": 0, + "content": "between quality and rendering time. Figure 6 further highlights that popping is handled properly. Note that, due to the sorting-free nature of our method, any popping free formulations can be easily integrated into our method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.907, + 0.806 + ], + "angle": 0, + "content": "Quantitative results - Tab. 1. We report the PSNR values of our fine-tuned scenes rendered with varying numbers of samples, alongside comparisons to 3DGS and StopThePop. We use our pop compensation from Section 3.4. By increasing the number of samples per pixel, our method approaches the quality of alpha blending. While the Monte Carlo noise affects metric, i.i.d. noise is visually often less distracting than structured artifacts; recent research suggests that the human visual system tolerates some unbiased noise, especially at interactive frame rates [31]. Additionally, as we show in Section 4.4 and Section 4.4, this can be mitigated via temporal anti-aliasing, and even without any mitigation, does not harm downstream perception tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.816, + 0.692, + 0.832 + ], + "angle": 0, + "content": "4.2. Computation time" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Forward pass timings - Tab. 2. We compare the forward rendering time of our method implemented in OpenGL, against 3DGS on both CUDA and OpenGL, as well as StopThePop [37] on CUDA. We report timings with a vary-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.941, + 0.518, + 0.953 + ], + "angle": 0, + "content": "26331" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.096, + 0.258, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.096, + 0.419, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.096, + 0.578, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.096, + 0.738, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.096, + 0.898, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.182, + 0.258, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.182, + 0.418, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.182, + 0.578, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.182, + 0.738, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.182, + 0.899, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.314, + 0.257, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.314, + 0.417, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.421, + 0.314, + 0.577, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.314, + 0.738, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.314, + 0.899, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.399, + 0.257, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.162, + 0.524, + 0.199, + 0.535 + ], + "angle": 0, + "content": "1 SPP" + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.399, + 0.417, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.319, + 0.524, + 0.358, + 0.536 + ], + "angle": 0, + "content": "4 SPP" + }, + { + "type": "image", + "bbox": [ + 0.421, + 0.399, + 0.577, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.476, + 0.524, + 0.52, + 0.536 + ], + "angle": 0, + "content": "16 SPP" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.399, + 0.738, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.524, + 0.682, + 0.536 + ], + "angle": 0, + "content": "128 SPP" + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.399, + 0.899, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.524, + 0.861, + 0.538 + ], + "angle": 0, + "content": "Alpha blending" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.54, + 0.907, + 0.582 + ], + "angle": 0, + "content": "Figure 7. Qualitative results - for the Room and Kitchen scenes from the MipNerf360 dataset are shown in Fig. 7. We compare StochasticSplats using 1, 4, and 16 samples per pixel against the standard alpha blending approach. Even with just 1 sample per pixel, our method produces reasonable renderings, and with 16 samples, the results closely match the quality of alpha blending." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.594, + 0.481, + 0.651 + ], + "angle": 0, + "content": "
3DGS-CUDA3DGS-OpenGLStopThePop1Ours-OpenGL (SPP)
24816
T100065.08100.2875.9116.2318.9524.7137.5161.25
RTX30908.1432.159.483.254.186.4215.3118.48
RTX40905.6020.70-1.852.052.866.718.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.656, + 0.483, + 0.697 + ], + "angle": 0, + "content": "Table 2. Forward pass time - in milliseconds (ms) for different techniques on MipNerf360-trained scenes on three different GPUs. The values are averaged over all scenes and all views." + }, + { + "type": "table", + "bbox": [ + 0.096, + 0.698, + 0.481, + 0.74 + ], + "angle": 0, + "content": "
3DGSStopThePop1Ours (SPP)
8128
23.0625.7718.1033.19478.93
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.745, + 0.483, + 0.788 + ], + "angle": 0, + "content": "Table 3. Backward pass time - in ms of the CUDA code for StochasticSplats (as we vary SPP), 3DGS [18], StopThePop [37], averaged over test views of the MipNerf360 scenes on RTX 3090." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.902 + ], + "angle": 0, + "content": "ing number of samples per pixel, exercising our method's ability to trade off performance and quality. With a single sample, our method achieves \\(2 \\times\\) to \\(4 \\times\\) lower latency compared to alpha-blended 3DGS, depending on the hardware; see Table 2. For higher sample counts, we employ supersampling: render at a higher resolution and then downsample to the target display resolution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.907, + 0.718 + ], + "angle": 0, + "content": "Our method is unique at being fast, portable and pop-free at the same time. The portability of our method is only comparable with the 3DGS-OpenGL variant, that has popping artifacts and is relatively slow. StopThePop cannot be easily ported to standard graphics frameworks, and 3DGS-CUDA slows down significantly when ported to OpenGL. This is because the CUDA radix sort routines utilize platform specific optimizations [34]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.781 + ], + "angle": 0, + "content": "Backward pass timings - Tab. 3. We further report the computation time for a single backward pass, this time with the CUDA implementation for all methods, as training typically is done on CUDA devices." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.789, + 0.887, + 0.805 + ], + "angle": 0, + "content": "4.3. Lowering resolution does not enhance speed" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Our method inherently offers a performance vs. quality trade-off. Attempting to naively integrate such a trade-off into 3DGS, such as reducing rendering resolution, does not work well. Simply lowering the rendering resolution does not speed up rasterization – the same number of Gaussians must be processed, and the view frustum for each pixel in" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.941, + 0.519, + 0.953 + ], + "angle": 0, + "content": "26332" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.077, + 0.469, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.284, + 0.487, + 0.34 + ], + "angle": 0, + "content": "Figure 8. Timings at lower resolution - Render timings (milliseconds) at different downsampled resolutions for our method (OpenGL) and alpha blending (CUDA) with different tile sizes on an NVIDIA RTX 3090." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.342, + 0.285, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.342, + 0.468, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.466, + 0.485, + 0.523 + ], + "angle": 0, + "content": "Figure 9. Temporal anti-aliasing - StochasticSplats at 1 SPP, along with FLIP error, without (left) and with (right) TAA after moving along a trajectory for 70 frames. TAA introduces a negligible latency increase. For videos, please refer to our website." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.533, + 0.484, + 0.64 + ], + "angle": 0, + "content": "creases. In short, more Gaussians contribute to each pixel, increasing the computation time per pixel. Figure 8 reports rendering times for our method (OpenGL) and alpha blending (AB-CUDA) with various tile sizes. While AB render time decreases to some degree, it starts to increase after a point, as the Gaussian count per-tile increases. No configuration can reach the rendering speed of our method (1 SPP)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.65, + 0.43, + 0.667 + ], + "angle": 0, + "content": "4.4. Making the most out of a single sample" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.703 + ], + "angle": 0, + "content": "We now practically demonstrate that even a single-sample rendering strategy can be useful." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Temporal anti-aliasing - Fig. 9. In interactive applications, multiple frames are rendered in rapid succession. A common strategy to further reduce variance is to average pixel values across frames by reusing information from previous renderings [40]. We employ temporal anti-aliasing (TAA) to not only mitigate aliasing artifacts, but also accumulate samples over time, achieving an effectively higher sample count at minimal cost. Frame-to-frame consistency is maintained by warping previously rendered frames according to per-pixel depth whenever the camera moves. As shown in Figure 9, this greatly reduces rendering noise, while maintaining low latency. We provide implementation details in the appendix, and our supplemental" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.089, + 0.71, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.193, + 0.68, + 0.205 + ], + "angle": 0, + "content": "(a) Green apple (AB)" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.09, + 0.888, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.74, + 0.193, + 0.859, + 0.205 + ], + "angle": 0, + "content": "(b) Green apple (ours)" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.206, + 0.709, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.309, + 0.675, + 0.321 + ], + "angle": 0, + "content": "(c) Chopsticks (AB)" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.206, + 0.887, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.309, + 0.855, + 0.321 + ], + "angle": 0, + "content": "(d) Chopsticks (ours)" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.327, + 0.907, + 0.368 + ], + "angle": 0, + "content": "Figure 10. Open vocabulary localization - Qualitative comparison of open-vocabulary 3D object localization on the LERF dataset using alpha blending vs. StochasticSplats at 1 SPP." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.368, + 0.905, + 0.413 + ], + "angle": 0, + "content": "
TeatimeRamenWaldo-KitchenFigurinesMean
AB88.173.295.580.484.3
Ours88.173.286.480.482.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.418, + 0.907, + 0.448 + ], + "angle": 0, + "content": "Table 4. Open vocabulary localization - accuracy on the LERF dataset using alpha blending (AB) and StochasticSplats at 1 SPP." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.454, + 0.854, + 0.469 + ], + "angle": 0, + "content": "videos demonstrate TAA applied to 1-SPP frames." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.471, + 0.907, + 0.609 + ], + "angle": 0, + "content": "Semantic localization - Fig. 10 and Tab. 4. We further show that while the single-sample render has low PSNR, it is still useful. Specifically, we use LangSplat [36] and replace their alpha blending renderer with our method, and follow their pipeline to localize objects. As shown in Figure 10, while our renders are noisy, it still properly renders semantics such that the localizer can correctly identify its location. There is a moderate reduction in accuracy (Tab. 4), but we achieve \\(4 \\times\\) faster rendering speed." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.614, + 0.634, + 0.63 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.907, + 0.761 + ], + "angle": 0, + "content": "We proposed a sorting-free stochastic rendering method for 3D Gaussian splatting, that is, technically, compatible with any semi-transparent representation. We derive both the forward (rendering) and backward (gradient) passes for stochastic rendering, allowing a quality vs. cost trade off, and renders without popping artifacts. Our OpenGL implementation renders \\(2 \\times -4 \\times\\) faster with reasonable quality, demonstrating high portability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.762, + 0.909, + 0.889 + ], + "angle": 0, + "content": "Limitations and future works. While we obtain reasonable quality when rendering with low sample counts, our results are noisy. Future work could focus on reducing the variance of both forward and backward estimators, for example using importance sampling, stratified sampling [21] or control variates. Moreover, quality could further be improved by leveraging lightweight (neural) denoising algorithms, similar to how diffusion models are used in conjunction with rough 3D renderings [39, 48]." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.941, + 0.521, + 0.953 + ], + "angle": 0, + "content": "26333" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Pontus Andersson, Jim Nilsson, Tomas Akenine-Moller, Magnus Oskarsson, Kalle Åström, and Mark D. Fairchild. FLIP: A Difference Evaluator for Alternating Images. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 3(2):15:1-15:23, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.242 + ], + "angle": 0, + "content": "[2] Dejan Azinovic, Tzu-Mao Li, Anton Kaplanyan, and Matthias Nießner. Inverse path tracing for joint material and lighting estimation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2442-2451, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.242, + 0.483, + 0.297 + ], + "angle": 0, + "content": "[3] Milena T Bagdasarian, Paul Knoll, Yi-Hsin Li, Florian Barthel, Anna Hilsmann, Peter Eisert, and Wieland Morgenstern. 3dgs. zip: A survey on 3d gaussian splatting compression methods. arXiv preprint arXiv:2407.09510, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.483, + 0.366 + ], + "angle": 0, + "content": "[4] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.483, + 0.437 + ], + "angle": 0, + "content": "[5] Jorge Condor, Sebastien Speierer, Lukas Bode, Aljaz Bozic, Simon Green, Piotr Didyk, and Adrian Jarabo. Don't splat your gaussians: Volumetric ray-traced primitives for modeling and rendering scattering and emissive media. ACM Trans. Graph., 44(1), 2025. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.483, + 0.478 + ], + "angle": 0, + "content": "[6] Franklin C. Crow. The aliasing problem in computer-generated shaded images. Commun. ACM, 20(11):799-805, 1977. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.48, + 0.483, + 0.535 + ], + "angle": 0, + "content": "[7] Thomas Deliot, Eric Heitz, and Laurent Belcour. Transforming a non-differentiable rasterizer into a differentiable one with stochastic gradient estimation. Proc. Symposium on Interactive 3D Graphics and Games (I3D), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.536, + 0.483, + 0.59 + ], + "angle": 0, + "content": "[8] Eric Enderton, Erik Sintorn, Peter Shirley, and David Luebke. Stochastic transparency. In Proc. Symposium on Interactive 3D Graphics and Games (I3D), pages 157-164, 2010. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.593, + 0.483, + 0.62 + ], + "angle": 0, + "content": "[9] Cass Everitt. Interactive order-independent transparency. White paper, NVIDIA, 2(6):7, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.621, + 0.482, + 0.648 + ], + "angle": 0, + "content": "[10] James D Foley. Computer graphics: principles and practice. Addison-Wesley Professional, 1996. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.649, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[11] Michael B. Giles. Monte carlo evaluation of sensitives in computational finance. Technical report, Oxford-Man Institute of Quantitative Finance, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[12] Ioannis Gkioulekas, Anat Levin, and Todd Zickler. An evaluation of computational imaging techniques for heterogeneous inverse scattering. In European Conference on Computer Vision (ECCV), pages 685-701. Springer International Publishing, 2016. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[13] Khronos Group. Early Fragment Test. https://www.khronos.org/opengl/wiki/Early_Fragment_Test, 2025.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[14] Hao Gui, Lin Hu, Rui Chen, Mingxiao Huang, Yuxin Yin, Jin Yang, and Yong Wu. Balanced 3DGS: Gaussian-wise parallelism rendering with fine-grained tiling. arXiv preprint arXiv:2412.17378, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[15] Florian Hahlbohm, Fabian Friederichs, Tim Weyrich, Linus Franke, Moritz Kappel, Susana Castillo, Marc Stamminger, Martin Eisemann, and Marcus Magnor. Effi" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "cient perspective-correct 3D Gaussian splatting using hybrid transparency. arXiv preprint arXiv:2410.08129, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[16] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, Light, and Material Decomposition from Images using Monte Carlo Rendering and Denoising. Advances in Neural Information Processing Systems (NeurIPS), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[17] Qiqi Hou, Randall Rauwendaal, Zifeng Li, Hoang Le, Farzad Farhadzadeh, Fatih Porikli, Alexei Bourd, and Amir Said. Sort-free gaussian splatting via weighted sum rendering. arXiv preprint arXiv:2410.18931, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[18] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D gaussian splatting for real-time radiance field rendering. ACM Trans. Graph. (Proc. SIGGRAPH), 42(4), 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.293, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[19] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Leref: Language embedded radiance fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 19729-19739, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[20] Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Weiwei Sun, Yang-Che Tseng, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. 3D gaussian splatting as markov chain monte carlo. In Advances in Neural Information Processing Systems (NeurIPS), 2024. Spotlight Presentation. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.49 + ], + "angle": 0, + "content": "[21] Samuli Laine and Tero Karras. Stratified sampling for stochastic transparency. In Comp. Graph. Forum (Proc. EGSR), pages 1197-1204, 2011. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[22] Tzu-Mao Li, Miika Aittala, Frédo Durand, and Jaakko Lehtinen. Differentiable monte carlo ray tracing through edge sampling. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 37 (6):222:1-222:11, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[23] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft Rasterizer: A Differentiable Renderer for Image-Based 3D Reasoning. In IEEE International Conference on Computer Vision (ICCV), pages 7708-7717, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[24] Alexander Mai, Dor Verbin, Falko Kuester, and Sara Fridovich-Keil. Neural microfacet fields for inverse rendering. In IEEE International Conference on Computer Vision (ICCV), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[25] Alexander Mai, Peter Hedman, George Kopanas, Dor Verbin, David Futschik, Qiangeng Xu, Falko Kuester, Jonathan T Barron, and Yinda Zhang. EVER: Exact volumetric ellipsoid rendering for real-time view synthesis. arXiv preprint arXiv:2410.01804, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[26] Saswat Subhajyoti Mallick, Rahul Goel, Bernhard Kerbl, Markus Steinberger, Francisco Vicente Carrasco, and Fernando De La Torre. Taming 3dgs: High-quality radiance fields with limited resources. In SIGGRAPH Asia Conference Papers, pages 1-11, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[27] Morgan McGuire and Louis Bavoil. Weighted blended order-independent transparency. Journal of Computer Graphics Techniques (JCGT), 2(2):122-141, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[28] Guennady A. Mikhailov. On the calculation of nuclear reactor disturbances by the monte carlo method. USSR Computational Mathematics and Mathematical Physics, 6(2):268-273, 1966. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.942, + 0.52, + 0.953 + ], + "angle": 0, + "content": "26334" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[29] Ben Mildenhall, Pratul Srinivasan, Matthew Tancik, Jonathan Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In European Conference on Computer Vision (ECCV), pages 405-421. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[30] Cedrick Münstermann, Stefan Krumpen, Reinhard Klein, and Christoph Peters. Moment-based order-independent transparency. Proc. Symposium on Interactive 3D Graphics and Games (I3D), 1(1):1-20, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[31] Vasiliki Myrodia. Psychophysical studies on Monte Carlo rendering-noise visual perception. PhD thesis, Université de Lille, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[32] Merlin Nimier-David, Delio Vicini, Tizian Zeltner, and Wenzel Jakob. Mitsuba 2: A retargetable forward and inverse renderer. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 38 (6), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[33] Jan Novák, Andrew Selle, and Wojciech Jarosz. Residual ratio tracking for estimating attenuation in participating media. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 33(6):179:1-179:11, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.371, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[34] Lukas Polok, Viorela Ila, and Pavel Smrz. Fast radix sort for sparse linear algebra on gpu. In *SpringSim (HPS)*, page 11, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[35] Thomas Porter and Tom Duff. Compositing digital images. In Computer Graphics (Proc. SIGGRAPH), page 253-259. ACM, 1984. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.511 + ], + "angle": 0, + "content": "[36] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 20051-20060, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[37] Lukas Radl, Michael Steiner, Mathias Parger, Alexander Weinrauch, Bernhard Kerbl, and Markus Steinberger. Stophepop: Sorted gaussian splatting for view-consistent real-time rendering. ACM Trans. Graph. (Proc. SIGGRAPH), 43(4):1-17, 2024. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.58, + 0.482, + 0.622 + ], + "angle": 0, + "content": "[38] Ravi Ramamoorthi and Pat Hanrahan. An efficient representation for irradiance environment maps. In ACM Trans. Graph. (Proc. SIGGRAPH), 2001. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.482, + 0.705 + ], + "angle": 0, + "content": "[39] Xuanchi Ren, Tianchang Shen, Jiahui Huang, Huan Ling, Yifan Lu, Merlin Nimier-David, Thomas Müller, Alexander Keller, Sanja Fidler, and Jun Gao. GEN3C: 3d-informed world-consistent video generation with precise camera control. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.706, + 0.482, + 0.762 + ], + "angle": 0, + "content": "[40] Daniel Scherzer, Lei Yang, Oliver Mattausch, Diego Nehab, Pedro V. Sander, Michael Wimmer, and Elmar Eisemann. Temporal Coherence Methods in Real-Time Rendering. Comp. Graph. Forum, 31(8):2378-2408, 2012. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.79 + ], + "angle": 0, + "content": "[41] Andrea Tagliasacchi and Ben Mildenhall. Volume rendering digest (for nef). arXiv preprint arXiv:2209.02417, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[42] Anthony J. Thibault. Splatapult. https://github.com/hyperlogic/splatapult, 2023. Accessed January 2, 2025. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[43] JM Tregan, S Blanco, J Dauchet, M Hafi, R Fournier, L Ibarrart, P Lapeyre, and N Villefranque. Convergence issues in derivatives of Monte Carlo null-collision integral formulations: a solution. arXiv preprint arXiv:1903.06508, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[44] Grigoris Tsopouridis, Andreas A Vasilakis, and Ioannis Fudos. Deep and fast approximate order independent transparency. In Comp. Graph. Forum, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[45] Dor Verbin, Ben Mildenhall, Peter Hedman, Jonathan T. Barron, Todd Zickler, and Pratul P. Srinivasan. Eclipse: Disambiguating illumination and materials using unintended shadows. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[46] Delio Vicini, Sébastien Speierer, and Wenzel Jakob. Path replay backpropagation: Differentiating light paths using constant memory and linear time. ACM Trans. Graph. (Proc. SIGGRAPH), 40(4):108:1-108:14, 2021. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[47] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Proc., 13(4), 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[48] Jay Zhangjie Wu, Yuxuan Zhang, Haithem Turki, Xuanchi Ren, Jun Gao, Mike Zheng Shou, Sanja Fidler, Zan Gojcic, and Huan Ling. Difix3D+: Improving 3d reconstructions with single-step diffusion models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[49] Chris Wyman. Stochastic layered alpha blending. In ACM SIGGRAPH 2016 Talks, pages 1-2. 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[50] Jason C Yang, Justin Hensley, Holger Grün, and Nicolas Thibierz. Real-time concurrent linked list construction on thegpu. In Comp. Graph. Forum (Proc. EGSR), 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[51] Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jeffrey Hu, Matthew Tancik, and Angjoo Kanazawa. gsplat: An open-source library for Gaussian splatting. arXiv preprint arXiv:2409.06765, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.517, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[52] Ekrem Fatih Yilmazer, Delio Vicini, and Wenzel Jakob. Solving inverse pde problems using monte carlo estimators. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 43, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.56, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[53] Tizian Zeltner, Sébastien Speierer, Iliyan Georgiev, and Wenzel Jakob. Monte carlo estimators for differential light transport. Transactions on Graphics (Proceedings of SIGGRAPH), 40(4), 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.617, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[54] Cheng Zhang, Lifan Wu, Changxi Zheng, Ioannis Gkioulekas, Ravi Ramamoorthi, and Shuang Zhao. A differential theory of radiative transfer. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 38(6), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.673, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[55] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.73, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[56] Shuang Zhao, Lifan Wu, Frédo Durand, and Ravi Ramamoorthi. Downsampling scattering parameters for rendering anisotropic media. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 35(6), 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.786, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[57] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Surface splatting. In ACM Trans. Graph. (Proc. SIGGRAPH), 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.83, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[58] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In Proceedings Visualization (VIS), pages 29-538. IEEE, 2001. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.872 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.941, + 0.519, + 0.953 + ], + "angle": 0, + "content": "26335" + } + ] +] \ No newline at end of file diff --git a/2025/StochasticSplats_ Stochastic Rasterization for Sorting-Free 3D Gaussian Splatting/d963f664-7a0a-43e1-95cb-e52f6169cd94_origin.pdf b/2025/StochasticSplats_ Stochastic Rasterization for Sorting-Free 3D Gaussian Splatting/d963f664-7a0a-43e1-95cb-e52f6169cd94_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f03e2e0f1e57bcc41b7eaa03e6bbb9d06d65040a --- /dev/null +++ b/2025/StochasticSplats_ Stochastic Rasterization for Sorting-Free 3D Gaussian Splatting/d963f664-7a0a-43e1-95cb-e52f6169cd94_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1423fd853c69d8c6ca8de7670297eb70182912f4501f81c1f1d0ac91a149a2bf +size 25509117 diff --git a/2025/StochasticSplats_ Stochastic Rasterization for Sorting-Free 3D Gaussian Splatting/full.md b/2025/StochasticSplats_ Stochastic Rasterization for Sorting-Free 3D Gaussian Splatting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..779839b42171cb3bde9459270fcc1ebe2d2b1df6 --- /dev/null +++ b/2025/StochasticSplats_ Stochastic Rasterization for Sorting-Free 3D Gaussian Splatting/full.md @@ -0,0 +1,420 @@ +# Stochastic Splats: Stochastic Rasterization for Sorting-Free 3D Gaussian Splatting + +Shakiba Kheradmand\*1,2, Delio Vicini\*3, George Kopanas\*3,4, Dmitry Lagun $^{1}$ , Kwang Moo Yi $^{2}$ , Mark Matthews $^{1}$ , Andrea Tagliasacchi $^{1,5,6}$ + +1Google DeepMind, 2University of British Columbia, 3Google, 4Runway ML, 5Simon Fraser University, 6University of Toronto, *equal contribution,†work done at Google + +![](images/22c9091e33f6ad2f0daf8f624a3b04b3a4e0536521347343605c06cee4877ce6.jpg) +Figure 1. Teaser - We render a 3DGS radiance field using an unbiased Monte Carlo estimator of the volume rendering equation. Unlike prior radiance field approaches, our method does not require sorted back-to-front rendering. This results in fast, portable, and pop-free rendering that is easy to implement. The number of Monte Carlo samples per pixel allows a trade-off between interactivity and visual quality, similar to physically-based rendering. From left to right, we increase the number of samples per pixel, which increases visual fidelity at a predictable increase in rendering cost. We report the mean square error (MSE) and render time on an NVIDIA RTX 4090. + +# Abstract + +3D Gaussian splatting (3DGS) is a popular radiance field method, with many application-specific extensions. Most variants rely on the same core algorithm: depth-sorting of Gaussian splats then rasterizing in primitive order. This ensures correct alpha compositing, but can cause rendering artifacts due to built-in approximations. Moreover, for a fixed representation, sorted rendering offers little control over render cost and visual fidelity. For example, and counter-intuitively, rendering a lower-resolution image is not necessarily faster. In this work, we address the above limitations by combining 3D Gaussian splatting with stochastic rasterization. Concretely, we leverage an unbi + +ased Monte Carlo estimator of the volume rendering equation. This removes the need for sorting, and allows for accurate 3D blending of overlapping Gaussians. The number of Monte Carlo samples further imbues 3DGS with a way to trade off computation time and quality. We implement our method using OpenGL shaders, enabling efficient rendering on modern GPU hardware. At a reasonable visual quality, our method renders more than four times faster than sorted rasterization. The project page and source code are publicly available at https://ubc-vision.github.io/stochasticsplats/. + +# 1. Introduction + +Recent advances in novel view synthesis have been driven by the success of neural radiance fields (NeRF) [29] and 3D Gaussian splatting (3DGS) [18]. The former produced high-quality visuals at a low rendering speed, while the latter enabled efficient rendering with comparable visual quality. 3DGS represents a scene as a set of anisotropic 3D Gaussians that can be "splatted" [57] in rasterization pipelines, enabling real-time viewing and efficient training. + +In 3DGS, 3D Gaussians are flattened onto camera-oriented billboards that not only prevent accurate volumetric intermixing [25], but must also be depthwise sorted for rasterization [58]. This global sorting operation is computationally expensive and introduces popping artifacts [37], as the sort order is view-dependent: small camera movements can result in temporal discontinuities in the rendered images. Further, there is no straightforward way to balance rendering cost vs. visual fidelity. One might assume that reducing the rendering resolution would lead to faster rasterization... yet, we observe the opposite in our experiments. Overall, this limits 3DGS's use on low-end hardware and in latency-critical applications, such as robotics and autonomous driving. Finally, most implementation use custom software rasterizers or GPGPU frameworks such as CUDA. Software rasterizers are notoriously hard to optimize, and often cannot match the performance of the hardware stack [23]. Additionally, many GPGPU frameworks are vendor-specific, reducing portability. + +Several previous works address popping artifacts, rendering performance, and portability, but typically improve only one, at the expense of the others. Popping can be alleviated by per-pixel depth sorting, but this introduces extra computational demands. While this cost can be amortized using clever spatial biases [37], per-pixel sorting cannot fully solve volumetric intermixing of 3D Gaussians. Constant density ellipsoids allow analytical intermixing [25], but are slow to render because they replace sorting with expensive ray tracing. Other methods [15] reduce the rendering cost and the popping artifacts by using a mix of per-pixel sorted lists and weighted order-independent transparency, but this comes at a cost in terms of portability since they need to implement complicated GPGPU kernels hindering cross-platform adoption. An orthogonal route to reduce render cost is to reduce the number of Gaussians [20]. Finally, packages such as Splatapult [42] implement an OpenGL-based rasterization pipeline, but do not outperform the original CUDA 3DGS implementation. + +We instead propose to render 3D Gaussians using stochastic transparency (ST) [8]; see Figure 2. Stochastic transparency replaces sorted alpha blending by an unbiased Monte Carlo estimator. Practically, transparency is rendered by probabilistically selecting individual, opaque, samples that are averaged together with multi-sampling [6]. + +![](images/9531edc4d98741eb90c0260fb004ea3699ba48c5be4f799d62a6c8ff45a09cca.jpg) + +![](images/f3fec56a804b5faa63d6b81524b26bde9bf0894ad8fe7483f8bfb9080afe9c80.jpg) +Alpha blending + +![](images/2916131be94fdd51ba5dd1850a3e775d00870fad31c8c6d82ae22013f6deef4b.jpg) + +![](images/262c2b18e7dddd00bb2a34489cbb216e299a392d22f4308758f9ec02edcfd51d.jpg) +Stochastic transparency (4 SPP) +Figure 2. Blending vs. stochastic - We visualize a 1D and 3D example of the different ways to compute transparency. Two foreground Gaussians are composited on a white background. + +Individual samples are rendered using order-independent Z-buffering, without sorting. We provide unbiased Monte-Carlo estimators of both the forward alpha blending and the gradient computation. The reverse-mode gradient estimator also permits sorting-free training and is largely agnostic to the primitive type and could be used for other semitransparent representations (e.g., triangle soups). We also do an early analysis on proper volumetric intermixing of 3D Gaussians by extending the theory of stochastic transparency. Critically, all of the above is naturally compatible with the hardware rasterization pipeline, allowing us a portable OpenGL implementation of our method. + +At one sample per pixel (SPP), our OpenGL implementation renders up to $4 \times$ faster than the original CUDA implementation, enabling latency-critical applications, such as open-vocabulary 3D object localization [19], which we demonstrate. Higher SPP increase quality and render time, approaching the PSNR of alpha-blended 3DGS. One key advantage is that downstream applications can dynamically adjust sampling at rendering time, allowing them to potentially trade-off latency against quality with careful engineering; see Figure 1. + +# 2. Related Work + +Order-independent transparency. While current 3DGS methods rely on sorted alpha blending, there is a rich literature on order-independent transparency for rasterization-based real-time rendering. Prior work can be classified into: (1) stochastic methods [8, 21, 49], that probabilistically discard fragments based on their opacity to create Monte Carlo estimators of the alpha blending equation and (2) non-stochastic methods that avoid sorting by using per-pixel linked lists [50], depth peeling [9], weighted blending [27], moments [30], or deep learning [44] to (approximately) composite transparent fragments. + +Given the success of order-independent transparency in + +real-time graphics, it becomes desirable to incorporate these techniques with 3DGS. Recent pre-prints [15, 17] combine 3DGS with non-stochastic order-independent transparency, but stochastic variants remain unexplored. We bridge this gap by extensively studying stochastic transparency with 3DGS, both in terms of rendering and training, enabled by our novel stochastic gradient estimator. + +Popping artifacts. Real-time graphics eschew sorting primitives due to both the computational cost and the inexistence of a consistent global ordering of non-trivial primitives. In 3DGS this is overlooked, because most primitives have a relatively small spatial extent. This introduces artifacts often dubbed as "popping", where a small motion of the camera forces a change in the order of the primitives introducing a sharp discontinuity in the alpha blending equation. These artifacts are particularly noticeable in immersive contexts like virtual reality. Radl et al. [37] address this using per-pixel sorting based on the maximal Gaussian response. While this eliminates popping artifacts, it introduces prohibitive computational costs, necessitating complex approximate sorting through a multi-stage hierarchical renderer. EVER [25] addresses popping artifacts by ray tracing constant density ellipsoids to compute the full volume rendering equation, rather than approximate alpha-blending. However, this approach is significantly slower than rasterization. In contrast, stochastic transparency uniquely eliminates popping by trading sorting for some noise in the final image. + +Fast 3DGS rendering and training. While 3DGS [18] offered a significant speedup over NeRF-based methods, subsequent works have further improved performance. Numerous improvements have been made to the implementation of training and rendering routines [14, 15, 26, 51]. Further speedups can be made by reducing the number of Gaussians and compressing their attributes. This is covered in an extensive literature review [3]. Any approach that reduces the number of primitives readily complements our work and is likely to improve rendering performance. + +Differentiable Monte Carlo methods. We introduce a novel Monte Carlo gradient estimator for stochastic transparency. Such differentiable Monte Carlo estimators are commonly used in physically-based inverse rendering [12, 22, 32, 46, 54, 56]. Various works propose differentiable Monte Carlo methods for neural radiance fields [24, 45] or surface reconstruction [16], and Condor et al. [5] differentiate physically-based Monte Carlo ray tracing of 3D Gaussian primitives. Deliot et al. [7] applied stochastic simultaneous perturbation gradients to real-time rendering algorithms, including standard 3DGS. Beyond rendering, differentiable Monte Carlo methods find application in nuclear engineering [28], finance [11], atmospheric physics [43] and inverse partial differential equation problems [52]. + +# 3. Method + +Kerbl et al. [18] represents a 3D scene in 3DGS as a collection of anisotropic 3D Gaussians $\mathcal{G} = \{g_i\}$ . Given a set of posed images $\{\mathbf{I}_n\}$ , we optimize the parameters of the 3D Gaussians $\mathcal{G}$ to photometrically reconstruct the scene. Each $g_i$ is parameterized as $\theta = \{\mu, \Sigma, \alpha, \mathbf{c}\}$ , where $\mu$ is the mean, $\Sigma$ is the covariance matrix, $\alpha$ is the opacity, and $\mathbf{c}$ is the view-dependent color modeled using spherical harmonics [38]. We start by reviewing the classical alpha-blending based rendering of 3DGS in Section 3.1, and then introduce our stochastic transparency algorithm in Section 3.2. + +# 3.1. Sorted alpha blending + +To calculate the radiance $C$ at a given pixel, we aggregate the radiances emitted by $L$ primitives that intersect the associated ray, blending them in depth $(z)$ order [35]: + +$$ +C = \sum_ {i = 1} ^ {L} c _ {i} \alpha_ {i} \prod_ {z _ {k} < z _ {i}} (1 - \alpha_ {k}) \tag {1} +$$ + +where $z_{k} < z_{i}$ denotes all primitives in front of primitive $i$ . While it can be shown that this expression derives from volume rendering [41], the product of terms in (1) requires sorting the primitives along $z$ , then blend radiances in either back-to-front or front-to-back order. As sorting is independent of how many pixels one needs to render, we simply lose the ability to trade off accuracy with speed. + +Sorting is done with respect to the $z$ coordinates of the Gaussian means $\mu_{k}$ remapped to the coordinate frame of the current camera, and the covariance is projected via a linear approximation of the non-linear projection process to the camera's view point, as proposed by Zwicker et al. [58]. This effectively treats each Gaussian as a billboard, flattening it parallel to the camera's imaging plane. + +# 3.2. Stochastic transparency + +Stochastic transparency (ST) [8] avoids sorting by stochastically estimating alpha blending in (1). Avoiding sorting when rendering transparent elements is critical, because (i) primitives often do not have a global unique sorting order across all pixels, (ii) sorting can be computationally expensive, and (iii) in the context of volume rendering the 3D extent of primitives influences the intermixing of colors. + +Stochastic transparency approximates transparency by randomly assigning a number of fully opaque samples per pixel proportional to a fragment's opacity, and averaging these samples to produce the final pixel color without sorting. In our context, the opacity of a fragment is proportional to the opacity of the Gaussian $\alpha$ , and to its distance from the mean $\mu$ , as measured by the projected covariance + +```latex +$\boldsymbol {z}\leftarrow \infty$ $\mathbf{C}\gets c_{\mathrm{background}}$ +for each Gaussian $i$ in unsorted order do Compute $\alpha_{i}$ and depth $z_{i}$ Sample $u\sim \mathcal{U}(0,1)$ if $u < \alpha_{i}$ and $z_{i} < z$ then $\mathbf{C}\gets c_i$ $z\gets z_{i}$ +end if +end for +``` + +Listing 1. Pseudocode for stochastic transparency. + +matrix $\Sigma$ . Stochastic transparency can also be interpreted as a randomized extension of classical screen-door transparency [10, Chapter 16], where, instead of a randomized selection, a fixed pattern is employed to discard fragments. + +Simple example. As a simple example, consider a single 3D Gaussian $g$ covering pixel $p$ with opacity $\alpha = 0.5$ ; in this case, if we sample $x \sim \mathcal{U}(0,1)$ and accept $g$ only when $x > \alpha$ , approximately $50\%$ of the pixel's samples will randomly select this Gaussian. We next formalize this idea to generalize to rendering an arbitrary number of Gaussians. + +Formalization. Formally, stochastic transparency is a Monte Carlo estimator of (1). It estimates the radiance $C$ by randomly sampling a summand $i$ (i.e., the $i$ -th Gaussian): + +$$ +C \approx \frac {1}{P (i)} c _ {i} \alpha_ {i} \prod_ {z _ {k} < z _ {i}} \left(1 - \alpha_ {k}\right), \tag {2} +$$ + +where $P(i)$ is the probability mass function of the random index $i$ . This estimator is unbiased, as long as $P(i) > 0$ for any non-zero summand. The standard approach is to use: + +$$ +P (i) = \alpha_ {i} \prod_ {z _ {k} < z _ {i}} \left(1 - \alpha_ {k}\right). \tag {3} +$$ + +The division by this probability mass function cancels out part of the sample weight, leading to a simple estimator: + +$$ +C \approx c _ {i} \quad \text {f o r} \quad i \sim P. \tag {4} +$$ + +Thus, a pixel has the probability $P(C = c_i) = P(i)$ to pick color $c_i$ . In practice, we average multiple independent samples per pixel to reduce estimator noise. + +Pseudocode. Listing 1 provides pseudocode for the above estimator. By testing that the depth of the current sample is smaller than the previously selected one, we sample exactly according to $P(i)$ [8]. In the fragment Shader, this is implemented by using standard depth testing and calling discard on samples where $u \geq \alpha_i$ . This only requires access to the current Gaussian's opacity and works for any primitive ordering. Figure 2 further shows a toy example of using this estimator on two Gaussians. + +# 3.3. Differentiable stochastic transparency + +Applying stochastic transparency to inverse rendering problems requires computing the gradient of the photometric + +![](images/1e9b9805511e55f6a1a5f4fd1bf0aad3954762598c8306adf83ce74766bee03b.jpg) +Alpha blending + +![](images/c0b58e11d23e59ed55ed7f101145c3efd946c524547d08231490947011fa13f9.jpg) +Ours (128 SPP) + +![](images/bcb4c2a7b38bca3cb03c430e94d3044530be14dc7f017ed93e371194356c72c3.jpg) +Ours (512 SPP) +Figure 3. Gradient validation - We compare alpha blending gradients to our stochastic estimator with 128 and 512 SPP. Red and blue colors encode positive and negative values, respectively. Our estimator accurately approximates the alpha blending gradients. + +loss $\mathcal{L}$ with respect to the primitive parameters $\theta$ . Primarily, we require gradients with respect to color $\partial \mathcal{L} / \partial c$ , and gradients with respect to opacity $\partial \mathcal{L} / \partial \alpha$ . + +Gradient estimator. We take inspiration from the differentiable rendering literature [46], and use a detached gradient estimator [46, 53] to correctly differentiate the binary sampling decision in Listing 1. We derive this estimator by applying the gradient operator to the alpha blending equation: + +$$ +\nabla_ {\boldsymbol {\theta}} C = \sum_ {i = 1} ^ {L} \nabla_ {\boldsymbol {\theta}} \left[ c _ {i} \alpha_ {i} \prod_ {z _ {k} < z _ {i}} (1 - \alpha_ {k}) \right]. \tag {5} +$$ + +We then form an unbiased Monte Carlo gradient estimator by again sampling a summand according to $P(i)$ . Crucially, we explicitly do not differentiate $P(i)$ itself and obtain: + +$$ +\nabla_ {\boldsymbol {\theta}} C \approx \frac {1}{P (i)} \nabla_ {\boldsymbol {\theta}} \left[ c _ {i} \alpha_ {i} \prod_ {z _ {k} < z _ {i}} \left(1 - \alpha_ {k}\right) \right]. \tag {6} +$$ + +Substituting the definition of $P(i)$ from (3) and differentiating $\mathcal{L}$ with respect to $c$ and $\alpha$ , we get: + +$$ +\frac {\partial \mathcal {L}}{\partial c _ {i}} = \frac {\partial \mathcal {L}}{\partial C} \quad \frac {\partial \mathcal {L}}{\partial c _ {z _ {k} < z _ {i}}} = 0 \tag {7} +$$ + +$$ +\frac {\partial \mathcal {L}}{\partial \alpha_ {i}} = \frac {\partial \mathcal {L}}{\partial C} \frac {c _ {i}}{\alpha_ {i}} \quad \frac {\partial \mathcal {L}}{\partial \alpha_ {z _ {k} < z _ {i}}} = \frac {\partial \mathcal {L}}{\partial C} \frac {- c _ {i}}{1 - \alpha_ {z _ {k} < z _ {i}}}, \tag {8} +$$ + +where subscript $i$ denotes the sampled primitive, and $\partial \mathcal{L} / \partial C$ is the photometric loss gradient, dependent upon the loss formulation (L1, SSIM, ...). Note how in (7) only the $i$ -th Gaussian receives gradients with respect to its color, while the others have a zero gradient. Meanwhile, (8) shows how the $i$ -th Gaussian, and ones closer to the camera receive a gradient with respect to their opacity. The Gaussians behind the sampled one do not receive any gradient. Finally, our derivation is not specific to 3DGS, and applies to any representation rendered with stochastic transparency. + +Decorrelated loss gradient. A well-known challenge for differentiable Monte Carlo methods is that the derivative of + +![](images/8db56bfdffccd852f2a8fa615395d868b543caf740b4394f6914c8b2ada454a1.jpg) +3DGS + +![](images/8626203bc7b47627c1cf4230a6699b27458a03e6f97f28f423ad7ed9c440aeee.jpg) +StopThePop [37] vs. ours +Figure 4. Popping in 3DGS - (left) As sorting in 3DGS is done with respect to the $z$ distance of the Gaussian mean from the camera, a small camera rotation can cause visible "popping" artifacts (i.e., sudden pixel color changes). (right) StopThePop [37] corrects for this behavior by associating a surface with each Gaussian, and determining the depth $z$ per-fragment (dashed line), rather than per-Gaussian. Our solution leads to visually comparable results, but approximates this surface linearly (solid line), so that per-fragment depth $z$ can be computed efficiently in hardware. + +the loss function is a product of two terms that are corrupted by noise (i.e., they are both random variables): + +$$ +\frac {\partial \mathcal {L}}{\partial \theta} = \frac {\partial \mathcal {L}}{\partial C} \frac {\partial C}{\partial \theta}. \tag {9} +$$ + +where $\theta$ are the geometry parameters. If the noise in both terms is correlated, we get biased gradients, which harm optimization. Following [2, 12], we use different random samples to evaluate the two factors above. If $\mathcal{L}$ is the L2 loss, the decorrelated evaluation results in unbiased gradients. For other loss functions, it at least reduces bias. Unbiased gradient estimation for arbitrary loss functions applied to a Monte Carlo estimator is an open problem. + +Practical algorithm. The presence of $c_{i}$ in (8) implies that we need access to the result of the forward pass to compute gradients. We therefore follow path replay backpropagation [46] and compute gradients in three passes: + +1. In a first pass, we render an image and evaluate $\partial \mathcal{L} / \partial C$ . +2. We then render a decorrelated image using a different random seed, giving us $c_{i}$ +3. Finally, we replay the second pass by re-rendering using the same random seed and using the stored $c_{i}$ to compute parameter gradients. + +Validation. We validate our estimator by comparing to the gradients of the sorted alpha blending implementation. For easier visual interpretation, we use our reverse-mode gradient implementation to compute a gradient image. We do this by backpropagating an L1 loss separately for each pixel, and visualizing the sum of the x-component of the positional gradients of all Gaussians affecting a given pixel. As shown in Figure 3, our method produces noisy gradients that closely match those of alpha blending in structure. + +# 3.4. Removing popping artifacts + +A naive implementation of stochastic transparency re-uses the billboard geometry of 3DGS, and consequently, to com + +![](images/65811c20896c8e54a0c56d773b57252a7216c16c7f8009af2b902840b30886a3.jpg) +Alpha blending +Figure 5. Fully volumetric intermixing - While alpha blending (left) suffers from popping and "pop-free" variants (middle) show discontinuities, our full volumetric intermixing (right) accurately composes overlapping Gaussians. + +![](images/b61a6b8bfe2048be453292bc26317187a4b27dbb47d4755edecedd740b26db76.jpg) +Pop-free + +![](images/a3050a1698f8d6dac1529063e9f1f8241f9a529355e01c833688993095e1d32a.jpg) +Our full intermixing + +pute the depths $z_{i}$ , the means of the Gaussians. As we will see in Section 4.2, at low sample count this provides high rendering efficiency, as we do not require a global sort. However, popping artifacts still exist, as the computation of $z_{i}$ is analogous to the one in 3DGS, small changes in the camera rotation can therefore change the order of primitives, and cause visible video artifacts; see Figure 4. + +The stop-the-pop way. Radl et al. [37] overcome the popping problem by computing the per-fragment (vs. per-Gaussian) depth. Their core idea is to compute the location of the peak of the Gaussian density along the ray. When coupled with a (computationally expensive) per-pixel sorting of Gaussians, this can provide a pop-free rendering. This can be interpreted as finding "maximum density surface" of the Gaussian, visualized in 2D in Figure 4, and using it to define fragment depths. However, per-fragment updates of depth are not particularly efficient, as early-discard mechanisms implemented in hardware [13] can only execute for fragments whose depth has not been modified by a shader. While in our setting we can forego sorting, the inability to early-discard fragments still affects us, which lead us to propose a different approach, described next. + +Our simplified way. We propose to avoid this inefficiency by foregoing per-fragment depth changes. Instead, we orient our billboards to linearly approximate the maximum density surface. If the ray associated with a pixel has origin o, the plane can be defined as: + +$$ +\mathbf {n} ^ {\top} (\mathbf {x} - \boldsymbol {\mu}) = 0, \quad \text {w h e r e} \quad \mathbf {n} = \Sigma^ {- 1} (\boldsymbol {\mu} - \mathbf {o}) \tag {10} +$$ + +By approximating this surface with a plane, we can easily use the interpolation kernels within the hardware rasterizer to compute the splat's depth. This approach adds no extra time to the render process and gives us popping-free rendering with a similar PSNR to [37] at higher sample counts. + +Fully volumetric intermixing. If maximum rendering fidelity is desired and the cost of per-fragment depth modification is acceptable, stochastic transparency can naturally support full volumetric intermixing. The key insight is that this can be achieved by replacing the fixed depth of a Gaussian by a sampled free-flight distance. For this, we first express the total contribution of a single Gaussian using the Beer-Lambert law: $\int_0^\infty \sigma_i(\mathbf{x}_t)\exp \left(-\int_0^t\sigma_i(\mathbf{x}_s)\mathrm{d}s\right)c_i\mathrm{d}t,$ + +where the opacity of a Gaussian is reparametrized using a spatially-varying extinction function $\sigma_{i}$ and $\mathbf{x}_a\coloneqq \mathbf{o} + a\cdot \mathbf{d}$ for a ray $(\mathbf{o},\mathbf{d})$ . We follow [5] and sample a free-flight distance $t$ proportional to $p(t) = \sigma_i(\mathbf{x}_t)\exp (-\int_0^t\sigma_i(\mathbf{x}_s)\mathrm{d}s)$ . To handle overlapping Gaussians, we adapt decomposition tracking [33]: If two Gaussians overlap, we sample a free-flight distance according to each and take the minimum. This accurately resolves overlap of an arbitrary number of Gaussians without additional data structures, sorting, or ray tracing, and readily fits into the stochastic transparency algorithm. We simply need to replace the computation of $z_{i}$ by a probabilistic sampling $z_{i}\sim p(t)$ . We show an example rendering of two overlapping Gaussians in Fig. 5. See the supplemental for the full free-flight sampling routine. + +# 4. Experiments + +We evaluate rendering performance on the real-world scenes from the MipNerf360 [4] dataset. For demonstrating semantic object localization with single-sample rendering we use the LERF dataset [19] - it consists of complex in-the-wild scenes, designed for 3D object localization. + +Baselines. For rendering baselines, we compare the performance of our method against the original 3DGS implementation [18] that uses CUDA, as well as Splatapult [42], an equivalent OpenGL implementation. We also compare with StopThePop [37], a pop-free method for 3DGS rendering. + +Metrics. We compute the standard metrics of peak signal-to-noise ratio (PSNR), structural similarity index metric (SSIM) [47], and learned perceptual image patch similarity (LPIPS) [55]. We report the average per-frame processing time for all methods on the MipNerf360 scenes. + +Implementation details. To demonstrate portability, we implement our method in both CUDA and OpenGL. Our CUDA implementation builds on the existing 3DGS framework by Kerbl et al. [18] and our OpenGL implementation builds upon Splatapult [42]. We evaluate on a Windows laptop equipped with an NVIDIA Quadro T1000 Max-Q GPU, and Linux desktops with an NVIDIA RTX 3090 and RTX 4090. Rather than training from scratch, we finetune a trained 3DGS scene with the CUDA implementation of our method for 1000 iterations. Since we are performing fine-tuning, we do not apply densification or pruning. All learning rates for the non-position parameters are kept unchanged, while the positions are optimized using a learning rate of $5 \times 10^{-5}$ with 128 samples. We keep all other hyperparameters the same. On average, the finetuning process requires approximately 14 minutes on an NVIDIA RTX 3090 GPU for scenes from MipNerf360. + +# 4.1. Rendering quality + +Qualitative results - Figs. 6 and 7. We show two scenes (Room and Kitchen) in Figure 7 from the MipN-erf360 dataset. As shown, our method allows a trade-off + +![](images/5e63d90586df3c2513d981957964f45da53f3c9ddab16779fdbf1fd524675829.jpg) + +![](images/79035a8d0371cc0b2520044c7d1b61b318b7219148d6b198bb49b2af58dc3285.jpg) + +![](images/847e750c81fb2077154732fde3cf7e1f03ed371f01bd0c9a794f6aaebedf3bba.jpg) + +![](images/33d21d8180d064514545a36e8f40dc68c628d8d3a3ab57d3fda7731c7811c224.jpg) +Frame $t$ + +![](images/49a1ec46c55484282d557be2f0d68965465ea41ea0c786836cec8d8248dd3ab2.jpg) +Frame $t + 1$ + +![](images/c6ffa01c056fc62f02908df7d4a72ff50fabba06c7bd66405098288c01bac93a.jpg) +FLIP error +Figure 6. Popping artifact - The top row displays two consecutive frames, fine-tuned and rendered using our StochasticSplats, along with their corresponding fLIP error map [1]. The bottom row presents the same frames processed with standard 3DGS. The foreground error indicates popping while the background errors are due to the camera motion. + +
3DGSStopThePop14Ours (SPP)
16642561024
PSNR28.9928.7917.9422.7526.1127.6128.1728.33
SSIM0.8690.8700.2840.5080.7050.8030.8400.852
LPIPS0.1850.1810.6030.4810.3390.2340.1870.186
+ +Table 1. Novel view synthesis - quantitative results of different methods on MipNerf360 scenes. StochasticSplats is finetuned on StopThePop-trained scenes using 128 SPP, and rendered using varying number of samples. + +between quality and rendering time. Figure 6 further highlights that popping is handled properly. Note that, due to the sorting-free nature of our method, any popping free formulations can be easily integrated into our method. + +Quantitative results - Tab. 1. We report the PSNR values of our fine-tuned scenes rendered with varying numbers of samples, alongside comparisons to 3DGS and StopThePop. We use our pop compensation from Section 3.4. By increasing the number of samples per pixel, our method approaches the quality of alpha blending. While the Monte Carlo noise affects metric, i.i.d. noise is visually often less distracting than structured artifacts; recent research suggests that the human visual system tolerates some unbiased noise, especially at interactive frame rates [31]. Additionally, as we show in Section 4.4 and Section 4.4, this can be mitigated via temporal anti-aliasing, and even without any mitigation, does not harm downstream perception tasks. + +# 4.2. Computation time + +Forward pass timings - Tab. 2. We compare the forward rendering time of our method implemented in OpenGL, against 3DGS on both CUDA and OpenGL, as well as StopThePop [37] on CUDA. We report timings with a vary- + +![](images/192b0f54fc66ad55ba371f2a06a25e87d44ae430868fef44380b6ae0ab01f42f.jpg) + +![](images/33e226995a4655fa49dfded435141adbb50643b5b87defd0b5202eb13ee9dc04.jpg) + +![](images/9edb61efb938a063d47fe0b119b30edcdcf6c4daf3f4889818297a648eaf8e17.jpg) + +![](images/5bd3908a0499e508975188293c03a2bc78e407c00b4f724229643efebbbfdf3d.jpg) + +![](images/1c8d299a115e624295e75492b396776f3c85973cd6aa4fda6fd4bc755cf737ee.jpg) + +![](images/343eef2f232e9f5e73e16fbf6b4ee86fb1aff2bd9b74c7240a56e7b0876d1bb4.jpg) + +![](images/f125978960767cb7dc86f0f294d0bba3901a5b997f05c27defb21ebb58bd849b.jpg) + +![](images/eb21155a9d8ef7c25de658c93d9a8ebdb16a77c5494905175ef27b422f72fd61.jpg) + +![](images/3d800f0cf180713551722432461845476304e8daa6d475c26de5af787f5faf78.jpg) + +![](images/7270906033fa29db918943972b913e7bf88974470ed0ffb5fcfb2cc468c0d4c0.jpg) + +![](images/8a0281dc118751ff326ef287ff584686bf9b316009f68c13469bd54bbbca535d.jpg) + +![](images/56fbe39da250b90346dd92696bcc253abcbf994a867387578a627b2b36e46f4a.jpg) + +![](images/c3292f860425a4825b58b6df4d9fb490dbf6d0928ac69d94b8708556ab28cd9f.jpg) + +![](images/b8ef27816aae8b110ce614fc292bc9162057b49370c69c02d983f3cb104f689a.jpg) + +![](images/7a3db570cd5a7f3853e3148ffe739d168119b6e4b5ce247497ae885079e23ca3.jpg) + +![](images/8a20bf59b5b4caa85d52272db8d9e38cf4581e511d94684edb0f500dc4f89dbf.jpg) +1 SPP + +![](images/6479be1972017f0f3dd3fb18f9b565674c343edf20a730afe88707472a437ff9.jpg) +4 SPP + +![](images/a8c432404dae02891d5b9007a1ed3441e13c8c9a690410499c55f3e1e20d5aba.jpg) +16 SPP + +![](images/481fd1a276bc5f03a09f1f71bcc63411c31d0daeb1aec6ed8cacefa56126913a.jpg) +128 SPP +Figure 7. Qualitative results - for the Room and Kitchen scenes from the MipNerf360 dataset are shown in Fig. 7. We compare StochasticSplats using 1, 4, and 16 samples per pixel against the standard alpha blending approach. Even with just 1 sample per pixel, our method produces reasonable renderings, and with 16 samples, the results closely match the quality of alpha blending. + +![](images/c5ccdb50b4e768f9159400504113d6ad01ce558c43d60bb5cc45824e9d37882b.jpg) +Alpha blending + +
3DGS-CUDA3DGS-OpenGLStopThePop1Ours-OpenGL (SPP)
24816
T100065.08100.2875.9116.2318.9524.7137.5161.25
RTX30908.1432.159.483.254.186.4215.3118.48
RTX40905.6020.70-1.852.052.866.718.00
+ +Table 2. Forward pass time - in milliseconds (ms) for different techniques on MipNerf360-trained scenes on three different GPUs. The values are averaged over all scenes and all views. + +
3DGSStopThePop1Ours (SPP)
8128
23.0625.7718.1033.19478.93
+ +Table 3. Backward pass time - in ms of the CUDA code for StochasticSplats (as we vary SPP), 3DGS [18], StopThePop [37], averaged over test views of the MipNerf360 scenes on RTX 3090. + +ing number of samples per pixel, exercising our method's ability to trade off performance and quality. With a single sample, our method achieves $2 \times$ to $4 \times$ lower latency compared to alpha-blended 3DGS, depending on the hardware; see Table 2. For higher sample counts, we employ supersampling: render at a higher resolution and then downsample to the target display resolution. + +Our method is unique at being fast, portable and pop-free at the same time. The portability of our method is only comparable with the 3DGS-OpenGL variant, that has popping artifacts and is relatively slow. StopThePop cannot be easily ported to standard graphics frameworks, and 3DGS-CUDA slows down significantly when ported to OpenGL. This is because the CUDA radix sort routines utilize platform specific optimizations [34]. + +Backward pass timings - Tab. 3. We further report the computation time for a single backward pass, this time with the CUDA implementation for all methods, as training typically is done on CUDA devices. + +# 4.3. Lowering resolution does not enhance speed + +Our method inherently offers a performance vs. quality trade-off. Attempting to naively integrate such a trade-off into 3DGS, such as reducing rendering resolution, does not work well. Simply lowering the rendering resolution does not speed up rasterization – the same number of Gaussians must be processed, and the view frustum for each pixel in + +![](images/1792147d97dde17498b2f881fab67930bcbdb60ea35becefa6f5cb75c8aa60bf.jpg) +Figure 8. Timings at lower resolution - Render timings (milliseconds) at different downsampled resolutions for our method (OpenGL) and alpha blending (CUDA) with different tile sizes on an NVIDIA RTX 3090. + +![](images/f0f2b084fc8229ecb843ceb1f5cd796f67111e6afcd5bea700918d2a2b9fc468.jpg) +Figure 9. Temporal anti-aliasing - StochasticSplats at 1 SPP, along with FLIP error, without (left) and with (right) TAA after moving along a trajectory for 70 frames. TAA introduces a negligible latency increase. For videos, please refer to our website. + +![](images/d5361622742efbae84724c3c7feaecf56f7223a98e662d8a39ada39bc54b496b.jpg) + +creases. In short, more Gaussians contribute to each pixel, increasing the computation time per pixel. Figure 8 reports rendering times for our method (OpenGL) and alpha blending (AB-CUDA) with various tile sizes. While AB render time decreases to some degree, it starts to increase after a point, as the Gaussian count per-tile increases. No configuration can reach the rendering speed of our method (1 SPP). + +# 4.4. Making the most out of a single sample + +We now practically demonstrate that even a single-sample rendering strategy can be useful. + +Temporal anti-aliasing - Fig. 9. In interactive applications, multiple frames are rendered in rapid succession. A common strategy to further reduce variance is to average pixel values across frames by reusing information from previous renderings [40]. We employ temporal anti-aliasing (TAA) to not only mitigate aliasing artifacts, but also accumulate samples over time, achieving an effectively higher sample count at minimal cost. Frame-to-frame consistency is maintained by warping previously rendered frames according to per-pixel depth whenever the camera moves. As shown in Figure 9, this greatly reduces rendering noise, while maintaining low latency. We provide implementation details in the appendix, and our supplemental + +![](images/c7e624f212b59042de5480de38f0c349a6f82f1718c4ea6d6a88cdee73b1c168.jpg) +(a) Green apple (AB) + +![](images/5c3439c8a5b405cb0b083a2d132341d29b3c771eb2574080751ec52527110482.jpg) +(b) Green apple (ours) + +![](images/256f2849db8bfcfc9da827861d9aea1c5060e082ec79c139b290477426cc15e2.jpg) +(c) Chopsticks (AB) + +![](images/abed106002d109459632ac80b5f89dd4c7af83d0066d103b404243e163ae901f.jpg) +(d) Chopsticks (ours) + +Figure 10. Open vocabulary localization - Qualitative comparison of open-vocabulary 3D object localization on the LERF dataset using alpha blending vs. StochasticSplats at 1 SPP. + +
TeatimeRamenWaldo-KitchenFigurinesMean
AB88.173.295.580.484.3
Ours88.173.286.480.482.0
+ +Table 4. Open vocabulary localization - accuracy on the LERF dataset using alpha blending (AB) and StochasticSplats at 1 SPP. + +videos demonstrate TAA applied to 1-SPP frames. + +Semantic localization - Fig. 10 and Tab. 4. We further show that while the single-sample render has low PSNR, it is still useful. Specifically, we use LangSplat [36] and replace their alpha blending renderer with our method, and follow their pipeline to localize objects. As shown in Figure 10, while our renders are noisy, it still properly renders semantics such that the localizer can correctly identify its location. There is a moderate reduction in accuracy (Tab. 4), but we achieve $4 \times$ faster rendering speed. + +# 5. Conclusion + +We proposed a sorting-free stochastic rendering method for 3D Gaussian splatting, that is, technically, compatible with any semi-transparent representation. We derive both the forward (rendering) and backward (gradient) passes for stochastic rendering, allowing a quality vs. cost trade off, and renders without popping artifacts. Our OpenGL implementation renders $2 \times -4 \times$ faster with reasonable quality, demonstrating high portability. + +Limitations and future works. While we obtain reasonable quality when rendering with low sample counts, our results are noisy. Future work could focus on reducing the variance of both forward and backward estimators, for example using importance sampling, stratified sampling [21] or control variates. Moreover, quality could further be improved by leveraging lightweight (neural) denoising algorithms, similar to how diffusion models are used in conjunction with rough 3D renderings [39, 48]. + +# References + +[1] Pontus Andersson, Jim Nilsson, Tomas Akenine-Moller, Magnus Oskarsson, Kalle Åström, and Mark D. Fairchild. FLIP: A Difference Evaluator for Alternating Images. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 3(2):15:1-15:23, 2020. 6 +[2] Dejan Azinovic, Tzu-Mao Li, Anton Kaplanyan, and Matthias Nießner. Inverse path tracing for joint material and lighting estimation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2442-2451, 2019. 5 +[3] Milena T Bagdasarian, Paul Knoll, Yi-Hsin Li, Florian Barthel, Anna Hilsmann, Peter Eisert, and Wieland Morgenstern. 3dgs. zip: A survey on 3d gaussian splatting compression methods. arXiv preprint arXiv:2407.09510, 2024. 3 +[4] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 6 +[5] Jorge Condor, Sebastien Speierer, Lukas Bode, Aljaz Bozic, Simon Green, Piotr Didyk, and Adrian Jarabo. Don't splat your gaussians: Volumetric ray-traced primitives for modeling and rendering scattering and emissive media. ACM Trans. Graph., 44(1), 2025. 3, 6 +[6] Franklin C. Crow. The aliasing problem in computer-generated shaded images. Commun. ACM, 20(11):799-805, 1977. 2 +[7] Thomas Deliot, Eric Heitz, and Laurent Belcour. Transforming a non-differentiable rasterizer into a differentiable one with stochastic gradient estimation. Proc. Symposium on Interactive 3D Graphics and Games (I3D), 2024. 3 +[8] Eric Enderton, Erik Sintorn, Peter Shirley, and David Luebke. Stochastic transparency. In Proc. Symposium on Interactive 3D Graphics and Games (I3D), pages 157-164, 2010. 2, 3, 4 +[9] Cass Everitt. Interactive order-independent transparency. White paper, NVIDIA, 2(6):7, 2001. 2 +[10] James D Foley. Computer graphics: principles and practice. Addison-Wesley Professional, 1996. 4 +[11] Michael B. Giles. Monte carlo evaluation of sensitives in computational finance. Technical report, Oxford-Man Institute of Quantitative Finance, 2007. 3 +[12] Ioannis Gkioulekas, Anat Levin, and Todd Zickler. An evaluation of computational imaging techniques for heterogeneous inverse scattering. In European Conference on Computer Vision (ECCV), pages 685-701. Springer International Publishing, 2016. 3, 5 +[13] Khronos Group. Early Fragment Test. https://www.khronos.org/opengl/wiki/Early_Fragment_Test, 2025.5 +[14] Hao Gui, Lin Hu, Rui Chen, Mingxiao Huang, Yuxin Yin, Jin Yang, and Yong Wu. Balanced 3DGS: Gaussian-wise parallelism rendering with fine-grained tiling. arXiv preprint arXiv:2412.17378, 2024. 3 +[15] Florian Hahlbohm, Fabian Friederichs, Tim Weyrich, Linus Franke, Moritz Kappel, Susana Castillo, Marc Stamminger, Martin Eisemann, and Marcus Magnor. Effi + +cient perspective-correct 3D Gaussian splatting using hybrid transparency. arXiv preprint arXiv:2410.08129, 2024. 2, 3 +[16] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, Light, and Material Decomposition from Images using Monte Carlo Rendering and Denoising. Advances in Neural Information Processing Systems (NeurIPS), 2022. 3 +[17] Qiqi Hou, Randall Rauwendaal, Zifeng Li, Hoang Le, Farzad Farhadzadeh, Fatih Porikli, Alexei Bourd, and Amir Said. Sort-free gaussian splatting via weighted sum rendering. arXiv preprint arXiv:2410.18931, 2024. 3 +[18] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D gaussian splatting for real-time radiance field rendering. ACM Trans. Graph. (Proc. SIGGRAPH), 42(4), 2023. 2, 3, 6, 7 +[19] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Leref: Language embedded radiance fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 19729-19739, 2023. 2, 6 +[20] Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Weiwei Sun, Yang-Che Tseng, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. 3D gaussian splatting as markov chain monte carlo. In Advances in Neural Information Processing Systems (NeurIPS), 2024. Spotlight Presentation. 2 +[21] Samuli Laine and Tero Karras. Stratified sampling for stochastic transparency. In Comp. Graph. Forum (Proc. EGSR), pages 1197-1204, 2011. 2, 8 +[22] Tzu-Mao Li, Miika Aittala, Frédo Durand, and Jaakko Lehtinen. Differentiable monte carlo ray tracing through edge sampling. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 37 (6):222:1-222:11, 2018. 3 +[23] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft Rasterizer: A Differentiable Renderer for Image-Based 3D Reasoning. In IEEE International Conference on Computer Vision (ICCV), pages 7708-7717, 2019. 2 +[24] Alexander Mai, Dor Verbin, Falko Kuester, and Sara Fridovich-Keil. Neural microfacet fields for inverse rendering. In IEEE International Conference on Computer Vision (ICCV), 2023. 3 +[25] Alexander Mai, Peter Hedman, George Kopanas, Dor Verbin, David Futschik, Qiangeng Xu, Falko Kuester, Jonathan T Barron, and Yinda Zhang. EVER: Exact volumetric ellipsoid rendering for real-time view synthesis. arXiv preprint arXiv:2410.01804, 2024. 2, 3 +[26] Saswat Subhajyoti Mallick, Rahul Goel, Bernhard Kerbl, Markus Steinberger, Francisco Vicente Carrasco, and Fernando De La Torre. Taming 3dgs: High-quality radiance fields with limited resources. In SIGGRAPH Asia Conference Papers, pages 1-11, 2024. 3 +[27] Morgan McGuire and Louis Bavoil. Weighted blended order-independent transparency. Journal of Computer Graphics Techniques (JCGT), 2(2):122-141, 2013. 2 +[28] Guennady A. Mikhailov. On the calculation of nuclear reactor disturbances by the monte carlo method. USSR Computational Mathematics and Mathematical Physics, 6(2):268-273, 1966. 3 + +[29] Ben Mildenhall, Pratul Srinivasan, Matthew Tancik, Jonathan Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In European Conference on Computer Vision (ECCV), pages 405-421. Springer, 2020. 2 +[30] Cedrick Münstermann, Stefan Krumpen, Reinhard Klein, and Christoph Peters. Moment-based order-independent transparency. Proc. Symposium on Interactive 3D Graphics and Games (I3D), 1(1):1-20, 2018. 2 +[31] Vasiliki Myrodia. Psychophysical studies on Monte Carlo rendering-noise visual perception. PhD thesis, Université de Lille, 2021. 6 +[32] Merlin Nimier-David, Delio Vicini, Tizian Zeltner, and Wenzel Jakob. Mitsuba 2: A retargetable forward and inverse renderer. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 38 (6), 2019. 3 +[33] Jan Novák, Andrew Selle, and Wojciech Jarosz. Residual ratio tracking for estimating attenuation in participating media. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 33(6):179:1-179:11, 2014. 6 +[34] Lukas Polok, Viorela Ila, and Pavel Smrz. Fast radix sort for sparse linear algebra on gpu. In *SpringSim (HPS)*, page 11, 2014. 7 +[35] Thomas Porter and Tom Duff. Compositing digital images. In Computer Graphics (Proc. SIGGRAPH), page 253-259. ACM, 1984. 3 +[36] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 20051-20060, 2024. 8 +[37] Lukas Radl, Michael Steiner, Mathias Parger, Alexander Weinrauch, Bernhard Kerbl, and Markus Steinberger. Stophepop: Sorted gaussian splatting for view-consistent real-time rendering. ACM Trans. Graph. (Proc. SIGGRAPH), 43(4):1-17, 2024. 2, 3, 5, 6, 7 +[38] Ravi Ramamoorthi and Pat Hanrahan. An efficient representation for irradiance environment maps. In ACM Trans. Graph. (Proc. SIGGRAPH), 2001. 3 +[39] Xuanchi Ren, Tianchang Shen, Jiahui Huang, Huan Ling, Yifan Lu, Merlin Nimier-David, Thomas Müller, Alexander Keller, Sanja Fidler, and Jun Gao. GEN3C: 3d-informed world-consistent video generation with precise camera control. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 8 +[40] Daniel Scherzer, Lei Yang, Oliver Mattausch, Diego Nehab, Pedro V. Sander, Michael Wimmer, and Elmar Eisemann. Temporal Coherence Methods in Real-Time Rendering. Comp. Graph. Forum, 31(8):2378-2408, 2012. 8 +[41] Andrea Tagliasacchi and Ben Mildenhall. Volume rendering digest (for nef). arXiv preprint arXiv:2209.02417, 2022. 3 +[42] Anthony J. Thibault. Splatapult. https://github.com/hyperlogic/splatapult, 2023. Accessed January 2, 2025. 2, 6 +[43] JM Tregan, S Blanco, J Dauchet, M Hafi, R Fournier, L Ibarrart, P Lapeyre, and N Villefranque. Convergence issues in derivatives of Monte Carlo null-collision integral formulations: a solution. arXiv preprint arXiv:1903.06508, 2019. 3 + +[44] Grigoris Tsopouridis, Andreas A Vasilakis, and Ioannis Fudos. Deep and fast approximate order independent transparency. In Comp. Graph. Forum, 2024. 2 +[45] Dor Verbin, Ben Mildenhall, Peter Hedman, Jonathan T. Barron, Todd Zickler, and Pratul P. Srinivasan. Eclipse: Disambiguating illumination and materials using unintended shadows. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3 +[46] Delio Vicini, Sébastien Speierer, and Wenzel Jakob. Path replay backpropagation: Differentiating light paths using constant memory and linear time. ACM Trans. Graph. (Proc. SIGGRAPH), 40(4):108:1-108:14, 2021. 3, 4, 5 +[47] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Proc., 13(4), 2004. 6 +[48] Jay Zhangjie Wu, Yuxuan Zhang, Haithem Turki, Xuanchi Ren, Jun Gao, Mike Zheng Shou, Sanja Fidler, Zan Gojcic, and Huan Ling. Difix3D+: Improving 3d reconstructions with single-step diffusion models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 8 +[49] Chris Wyman. Stochastic layered alpha blending. In ACM SIGGRAPH 2016 Talks, pages 1-2. 2016. 2 +[50] Jason C Yang, Justin Hensley, Holger Grün, and Nicolas Thibierz. Real-time concurrent linked list construction on thegpu. In Comp. Graph. Forum (Proc. EGSR), 2010. 2 +[51] Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jeffrey Hu, Matthew Tancik, and Angjoo Kanazawa. gsplat: An open-source library for Gaussian splatting. arXiv preprint arXiv:2409.06765, 2024.3 +[52] Ekrem Fatih Yilmazer, Delio Vicini, and Wenzel Jakob. Solving inverse pde problems using monte carlo estimators. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 43, 2024. 3 +[53] Tizian Zeltner, Sébastien Speierer, Iliyan Georgiev, and Wenzel Jakob. Monte carlo estimators for differential light transport. Transactions on Graphics (Proceedings of SIGGRAPH), 40(4), 2021. 4 +[54] Cheng Zhang, Lifan Wu, Changxi Zheng, Ioannis Gkioulekas, Ravi Ramamoorthi, and Shuang Zhao. A differential theory of radiative transfer. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 38(6), 2019. 3 +[55] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6 +[56] Shuang Zhao, Lifan Wu, Frédo Durand, and Ravi Ramamoorthi. Downsampling scattering parameters for rendering anisotropic media. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 35(6), 2016. 3 +[57] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Surface splatting. In ACM Trans. Graph. (Proc. SIGGRAPH), 2001. 2 +[58] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In Proceedings Visualization (VIS), pages 29-538. IEEE, 2001. 2, 3 \ No newline at end of file diff --git a/2025/StochasticSplats_ Stochastic Rasterization for Sorting-Free 3D Gaussian Splatting/images.zip b/2025/StochasticSplats_ Stochastic Rasterization for Sorting-Free 3D Gaussian Splatting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e9bfaf8a4c2052a54f90e08e775ce1a9942246f4 --- /dev/null +++ b/2025/StochasticSplats_ Stochastic Rasterization for Sorting-Free 3D Gaussian Splatting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:78f63368d0b425445c4db1d7a1a7bcfe7b771a6317f9cd14bd752ef01cd309c8 +size 713248 diff --git a/2025/StochasticSplats_ Stochastic Rasterization for Sorting-Free 3D Gaussian Splatting/layout.json b/2025/StochasticSplats_ Stochastic Rasterization for Sorting-Free 3D Gaussian Splatting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5c90dbf8c5ffd9368ede91ac0f477a4e8e8e3d87 --- /dev/null +++ b/2025/StochasticSplats_ Stochastic Rasterization for Sorting-Free 3D Gaussian Splatting/layout.json @@ -0,0 +1,10989 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 107, + 103, + 503, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 103, + 503, + 140 + ], + "spans": [ + { + "bbox": [ + 107, + 103, + 503, + 140 + ], + "type": "text", + "content": "Stochastic Splats: Stochastic Rasterization for Sorting-Free 3D Gaussian Splatting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 120, + 160, + 488, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 160, + 488, + 190 + ], + "spans": [ + { + "bbox": [ + 120, + 160, + 488, + 190 + ], + "type": "text", + "content": "Shakiba Kheradmand\\*1,2, Delio Vicini\\*3, George Kopanas\\*3,4, Dmitry Lagun" + }, + { + "bbox": [ + 120, + 160, + 488, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 120, + 160, + 488, + 190 + ], + "type": "text", + "content": ", Kwang Moo Yi" + }, + { + "bbox": [ + 120, + 160, + 488, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 120, + 160, + 488, + 190 + ], + "type": "text", + "content": ", Mark Matthews" + }, + { + "bbox": [ + 120, + 160, + 488, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 120, + 160, + 488, + 190 + ], + "type": "text", + "content": ", Andrea Tagliasacchi" + }, + { + "bbox": [ + 120, + 160, + 488, + 190 + ], + "type": "inline_equation", + "content": "^{1,5,6}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 117, + 200, + 493, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 200, + 493, + 243 + ], + "spans": [ + { + "bbox": [ + 117, + 200, + 493, + 243 + ], + "type": "text", + "content": "1Google DeepMind, 2University of British Columbia, 3Google, 4Runway ML, 5Simon Fraser University, 6University of Toronto, *equal contribution,†work done at Google" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 65, + 256, + 548, + 470 + ], + "blocks": [ + { + "bbox": [ + 65, + 256, + 548, + 470 + ], + "lines": [ + { + "bbox": [ + 65, + 256, + 548, + 470 + ], + "spans": [ + { + "bbox": [ + 65, + 256, + 548, + 470 + ], + "type": "image", + "image_path": "22c9091e33f6ad2f0daf8f624a3b04b3a4e0536521347343605c06cee4877ce6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 475, + 555, + 533 + ], + "lines": [ + { + "bbox": [ + 54, + 475, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 54, + 475, + 555, + 533 + ], + "type": "text", + "content": "Figure 1. Teaser - We render a 3DGS radiance field using an unbiased Monte Carlo estimator of the volume rendering equation. Unlike prior radiance field approaches, our method does not require sorted back-to-front rendering. This results in fast, portable, and pop-free rendering that is easy to implement. The number of Monte Carlo samples per pixel allows a trade-off between interactivity and visual quality, similar to physically-based rendering. From left to right, we increase the number of samples per pixel, which increases visual fidelity at a predictable increase in rendering cost. We report the mean square error (MSE) and render time on an NVIDIA RTX 4090." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 540, + 201, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 540, + 201, + 553 + ], + "spans": [ + { + "bbox": [ + 151, + 540, + 201, + 553 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 564, + 297, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 297, + 708 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 297, + 708 + ], + "type": "text", + "content": "3D Gaussian splatting (3DGS) is a popular radiance field method, with many application-specific extensions. Most variants rely on the same core algorithm: depth-sorting of Gaussian splats then rasterizing in primitive order. This ensures correct alpha compositing, but can cause rendering artifacts due to built-in approximations. Moreover, for a fixed representation, sorted rendering offers little control over render cost and visual fidelity. For example, and counter-intuitively, rendering a lower-resolution image is not necessarily faster. In this work, we address the above limitations by combining 3D Gaussian splatting with stochastic rasterization. Concretely, we leverage an unbi" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 540, + 556, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 556, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 556, + 673 + ], + "type": "text", + "content": "ased Monte Carlo estimator of the volume rendering equation. This removes the need for sorting, and allows for accurate 3D blending of overlapping Gaussians. The number of Monte Carlo samples further imbues 3DGS with a way to trade off computation time and quality. We implement our method using OpenGL shaders, enabling efficient rendering on modern GPU hardware. At a reasonable visual quality, our method renders more than four times faster than sorted rasterization. The project page and source code are publicly available at https://ubc-vision.github.io/stochasticsplats/." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 745, + 318, + 754 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 745, + 318, + 754 + ], + "spans": [ + { + "bbox": [ + 293, + 745, + 318, + 754 + ], + "type": "text", + "content": "26326" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 186 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 186 + ], + "type": "text", + "content": "Recent advances in novel view synthesis have been driven by the success of neural radiance fields (NeRF) [29] and 3D Gaussian splatting (3DGS) [18]. The former produced high-quality visuals at a low rendering speed, while the latter enabled efficient rendering with comparable visual quality. 3DGS represents a scene as a set of anisotropic 3D Gaussians that can be \"splatted\" [57] in rasterization pipelines, enabling real-time viewing and efficient training." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 187, + 296, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 187, + 296, + 414 + ], + "spans": [ + { + "bbox": [ + 55, + 187, + 296, + 414 + ], + "type": "text", + "content": "In 3DGS, 3D Gaussians are flattened onto camera-oriented billboards that not only prevent accurate volumetric intermixing [25], but must also be depthwise sorted for rasterization [58]. This global sorting operation is computationally expensive and introduces popping artifacts [37], as the sort order is view-dependent: small camera movements can result in temporal discontinuities in the rendered images. Further, there is no straightforward way to balance rendering cost vs. visual fidelity. One might assume that reducing the rendering resolution would lead to faster rasterization... yet, we observe the opposite in our experiments. Overall, this limits 3DGS's use on low-end hardware and in latency-critical applications, such as robotics and autonomous driving. Finally, most implementation use custom software rasterizers or GPGPU frameworks such as CUDA. Software rasterizers are notoriously hard to optimize, and often cannot match the performance of the hardware stack [23]. Additionally, many GPGPU frameworks are vendor-specific, reducing portability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 415, + 296, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 641 + ], + "type": "text", + "content": "Several previous works address popping artifacts, rendering performance, and portability, but typically improve only one, at the expense of the others. Popping can be alleviated by per-pixel depth sorting, but this introduces extra computational demands. While this cost can be amortized using clever spatial biases [37], per-pixel sorting cannot fully solve volumetric intermixing of 3D Gaussians. Constant density ellipsoids allow analytical intermixing [25], but are slow to render because they replace sorting with expensive ray tracing. Other methods [15] reduce the rendering cost and the popping artifacts by using a mix of per-pixel sorted lists and weighted order-independent transparency, but this comes at a cost in terms of portability since they need to implement complicated GPGPU kernels hindering cross-platform adoption. An orthogonal route to reduce render cost is to reduce the number of Gaussians [20]. Finally, packages such as Splatapult [42] implement an OpenGL-based rasterization pipeline, but do not outperform the original CUDA 3DGS implementation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "We instead propose to render 3D Gaussians using stochastic transparency (ST) [8]; see Figure 2. Stochastic transparency replaces sorted alpha blending by an unbiased Monte Carlo estimator. Practically, transparency is rendered by probabilistically selecting individual, opaque, samples that are averaged together with multi-sampling [6]." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 429, + 135 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 429, + 135 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 429, + 135 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 429, + 135 + ], + "type": "image", + "image_path": "9531edc4d98741eb90c0260fb004ea3699ba48c5be4f799d62a6c8ff45a09cca.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 319, + 137, + 426, + 219 + ], + "blocks": [ + { + "bbox": [ + 319, + 137, + 426, + 219 + ], + "lines": [ + { + "bbox": [ + 319, + 137, + 426, + 219 + ], + "spans": [ + { + "bbox": [ + 319, + 137, + 426, + 219 + ], + "type": "image", + "image_path": "f3fec56a804b5faa63d6b81524b26bde9bf0894ad8fe7483f8bfb9080afe9c80.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 220, + 405, + 230 + ], + "lines": [ + { + "bbox": [ + 348, + 220, + 405, + 230 + ], + "spans": [ + { + "bbox": [ + 348, + 220, + 405, + 230 + ], + "type": "text", + "content": "Alpha blending" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 440, + 72, + 552, + 136 + ], + "blocks": [ + { + "bbox": [ + 440, + 72, + 552, + 136 + ], + "lines": [ + { + "bbox": [ + 440, + 72, + 552, + 136 + ], + "spans": [ + { + "bbox": [ + 440, + 72, + 552, + 136 + ], + "type": "image", + "image_path": "2916131be94fdd51ba5dd1850a3e775d00870fad31c8c6d82ae22013f6deef4b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 445, + 140, + 552, + 219 + ], + "blocks": [ + { + "bbox": [ + 445, + 140, + 552, + 219 + ], + "lines": [ + { + "bbox": [ + 445, + 140, + 552, + 219 + ], + "spans": [ + { + "bbox": [ + 445, + 140, + 552, + 219 + ], + "type": "image", + "image_path": "262c2b18e7dddd00bb2a34489cbb216e299a392d22f4308758f9ec02edcfd51d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 220, + 555, + 230 + ], + "lines": [ + { + "bbox": [ + 440, + 220, + 555, + 230 + ], + "spans": [ + { + "bbox": [ + 440, + 220, + 555, + 230 + ], + "type": "text", + "content": "Stochastic transparency (4 SPP)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 234, + 555, + 268 + ], + "lines": [ + { + "bbox": [ + 313, + 234, + 555, + 268 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 555, + 268 + ], + "type": "text", + "content": "Figure 2. Blending vs. stochastic - We visualize a 1D and 3D example of the different ways to compute transparency. Two foreground Gaussians are composited on a white background." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 276, + 555, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 555, + 419 + ], + "type": "text", + "content": "Individual samples are rendered using order-independent Z-buffering, without sorting. We provide unbiased Monte-Carlo estimators of both the forward alpha blending and the gradient computation. The reverse-mode gradient estimator also permits sorting-free training and is largely agnostic to the primitive type and could be used for other semitransparent representations (e.g., triangle soups). We also do an early analysis on proper volumetric intermixing of 3D Gaussians by extending the theory of stochastic transparency. Critically, all of the above is naturally compatible with the hardware rasterization pipeline, allowing us a portable OpenGL implementation of our method." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 420, + 556, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 556, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 556, + 540 + ], + "type": "text", + "content": "At one sample per pixel (SPP), our OpenGL implementation renders up to " + }, + { + "bbox": [ + 313, + 420, + 556, + 540 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 313, + 420, + 556, + 540 + ], + "type": "text", + "content": " faster than the original CUDA implementation, enabling latency-critical applications, such as open-vocabulary 3D object localization [19], which we demonstrate. Higher SPP increase quality and render time, approaching the PSNR of alpha-blended 3DGS. One key advantage is that downstream applications can dynamically adjust sampling at rendering time, allowing them to potentially trade-off latency against quality with careful engineering; see Figure 1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 549, + 400, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 549, + 400, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 549, + 400, + 562 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 570, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 700 + ], + "type": "text", + "content": "Order-independent transparency. While current 3DGS methods rely on sorted alpha blending, there is a rich literature on order-independent transparency for rasterization-based real-time rendering. Prior work can be classified into: (1) stochastic methods [8, 21, 49], that probabilistically discard fragments based on their opacity to create Monte Carlo estimators of the alpha blending equation and (2) non-stochastic methods that avoid sorting by using per-pixel linked lists [50], depth peeling [9], weighted blending [27], moments [30], or deep learning [44] to (approximately) composite transparent fragments." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "text", + "content": "Given the success of order-independent transparency in" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 745, + 317, + 754 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 745, + 317, + 754 + ], + "spans": [ + { + "bbox": [ + 293, + 745, + 317, + 754 + ], + "type": "text", + "content": "26327" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "real-time graphics, it becomes desirable to incorporate these techniques with 3DGS. Recent pre-prints [15, 17] combine 3DGS with non-stochastic order-independent transparency, but stochastic variants remain unexplored. We bridge this gap by extensively studying stochastic transparency with 3DGS, both in terms of rendering and training, enabled by our novel stochastic gradient estimator." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 158, + 295, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 158, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 158, + 295, + 409 + ], + "type": "text", + "content": "Popping artifacts. Real-time graphics eschew sorting primitives due to both the computational cost and the inexistence of a consistent global ordering of non-trivial primitives. In 3DGS this is overlooked, because most primitives have a relatively small spatial extent. This introduces artifacts often dubbed as \"popping\", where a small motion of the camera forces a change in the order of the primitives introducing a sharp discontinuity in the alpha blending equation. These artifacts are particularly noticeable in immersive contexts like virtual reality. Radl et al. [37] address this using per-pixel sorting based on the maximal Gaussian response. While this eliminates popping artifacts, it introduces prohibitive computational costs, necessitating complex approximate sorting through a multi-stage hierarchical renderer. EVER [25] addresses popping artifacts by ray tracing constant density ellipsoids to compute the full volume rendering equation, rather than approximate alpha-blending. However, this approach is significantly slower than rasterization. In contrast, stochastic transparency uniquely eliminates popping by trading sorting for some noise in the final image." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 411, + 295, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 295, + 531 + ], + "type": "text", + "content": "Fast 3DGS rendering and training. While 3DGS [18] offered a significant speedup over NeRF-based methods, subsequent works have further improved performance. Numerous improvements have been made to the implementation of training and rendering routines [14, 15, 26, 51]. Further speedups can be made by reducing the number of Gaussians and compressing their attributes. This is covered in an extensive literature review [3]. Any approach that reduces the number of primitives readily complements our work and is likely to improve rendering performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 533, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 700 + ], + "type": "text", + "content": "Differentiable Monte Carlo methods. We introduce a novel Monte Carlo gradient estimator for stochastic transparency. Such differentiable Monte Carlo estimators are commonly used in physically-based inverse rendering [12, 22, 32, 46, 54, 56]. Various works propose differentiable Monte Carlo methods for neural radiance fields [24, 45] or surface reconstruction [16], and Condor et al. [5] differentiate physically-based Monte Carlo ray tracing of 3D Gaussian primitives. Deliot et al. [7] applied stochastic simultaneous perturbation gradients to real-time rendering algorithms, including standard 3DGS. Beyond rendering, differentiable Monte Carlo methods find application in nuclear engineering [28], finance [11], atmospheric physics [43] and inverse partial differential equation problems [52]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 71, + 370, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 370, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 370, + 83 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "text", + "content": "Kerbl et al. [18] represents a 3D scene in 3DGS as a collection of anisotropic 3D Gaussians " + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = \\{g_i\\}" + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "text", + "content": ". Given a set of posed images " + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_n\\}" + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "text", + "content": ", we optimize the parameters of the 3D Gaussians " + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "text", + "content": " to photometrically reconstruct the scene. Each " + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "inline_equation", + "content": "g_i" + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "text", + "content": " is parameterized as " + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\theta = \\{\\mu, \\Sigma, \\alpha, \\mathbf{c}\\}" + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "text", + "content": " is the mean, " + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "text", + "content": " is the covariance matrix, " + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "text", + "content": " is the opacity, and " + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "text", + "content": " is the view-dependent color modeled using spherical harmonics [38]. We start by reviewing the classical alpha-blending based rendering of 3DGS in Section 3.1, and then introduce our stochastic transparency algorithm in Section 3.2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 219, + 441, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 219, + 441, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 219, + 441, + 232 + ], + "type": "text", + "content": "3.1. Sorted alpha blending" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 236, + 555, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 274 + ], + "type": "text", + "content": "To calculate the radiance " + }, + { + "bbox": [ + 313, + 236, + 555, + 274 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 236, + 555, + 274 + ], + "type": "text", + "content": " at a given pixel, we aggregate the radiances emitted by " + }, + { + "bbox": [ + 313, + 236, + 555, + 274 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 236, + 555, + 274 + ], + "type": "text", + "content": " primitives that intersect the associated ray, blending them in depth " + }, + { + "bbox": [ + 313, + 236, + 555, + 274 + ], + "type": "inline_equation", + "content": "(z)" + }, + { + "bbox": [ + 313, + 236, + 555, + 274 + ], + "type": "text", + "content": " order [35]:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 375, + 281, + 555, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 281, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 375, + 281, + 555, + 316 + ], + "type": "interline_equation", + "content": "C = \\sum_ {i = 1} ^ {L} c _ {i} \\alpha_ {i} \\prod_ {z _ {k} < z _ {i}} (1 - \\alpha_ {k}) \\tag {1}", + "image_path": "1fc9a922bd99a27d2f04d2a4a54207cf7873e7824c5dba52300134e27fb3db99.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 324, + 554, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 554, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 554, + 408 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 324, + 554, + 408 + ], + "type": "inline_equation", + "content": "z_{k} < z_{i}" + }, + { + "bbox": [ + 313, + 324, + 554, + 408 + ], + "type": "text", + "content": " denotes all primitives in front of primitive " + }, + { + "bbox": [ + 313, + 324, + 554, + 408 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 324, + 554, + 408 + ], + "type": "text", + "content": ". While it can be shown that this expression derives from volume rendering [41], the product of terms in (1) requires sorting the primitives along " + }, + { + "bbox": [ + 313, + 324, + 554, + 408 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 324, + 554, + 408 + ], + "type": "text", + "content": ", then blend radiances in either back-to-front or front-to-back order. As sorting is independent of how many pixels one needs to render, we simply lose the ability to trade off accuracy with speed." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 408, + 555, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 555, + 492 + ], + "type": "text", + "content": "Sorting is done with respect to the " + }, + { + "bbox": [ + 313, + 408, + 555, + 492 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 408, + 555, + 492 + ], + "type": "text", + "content": " coordinates of the Gaussian means " + }, + { + "bbox": [ + 313, + 408, + 555, + 492 + ], + "type": "inline_equation", + "content": "\\mu_{k}" + }, + { + "bbox": [ + 313, + 408, + 555, + 492 + ], + "type": "text", + "content": " remapped to the coordinate frame of the current camera, and the covariance is projected via a linear approximation of the non-linear projection process to the camera's view point, as proposed by Zwicker et al. [58]. This effectively treats each Gaussian as a billboard, flattening it parallel to the camera's imaging plane." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 500, + 449, + 512 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 500, + 449, + 512 + ], + "spans": [ + { + "bbox": [ + 313, + 500, + 449, + 512 + ], + "type": "text", + "content": "3.2. Stochastic transparency" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 517, + 554, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 554, + 601 + ], + "type": "text", + "content": "Stochastic transparency (ST) [8] avoids sorting by stochastically estimating alpha blending in (1). Avoiding sorting when rendering transparent elements is critical, because (i) primitives often do not have a global unique sorting order across all pixels, (ii) sorting can be computationally expensive, and (iii) in the context of volume rendering the 3D extent of primitives influences the intermixing of colors." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 601, + 555, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 555, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 555, + 685 + ], + "type": "text", + "content": "Stochastic transparency approximates transparency by randomly assigning a number of fully opaque samples per pixel proportional to a fragment's opacity, and averaging these samples to produce the final pixel color without sorting. In our context, the opacity of a fragment is proportional to the opacity of the Gaussian " + }, + { + "bbox": [ + 313, + 601, + 555, + 685 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 601, + 555, + 685 + ], + "type": "text", + "content": ", and to its distance from the mean " + }, + { + "bbox": [ + 313, + 601, + 555, + 685 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 601, + 555, + 685 + ], + "type": "text", + "content": ", as measured by the projected covariance" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 554, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 554, + 713 + ], + "type": "text", + "content": "1Note that as we do not use multi-sampling anti-aliasing, making fragments synonymous with pixels." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 745, + 318, + 754 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 745, + 318, + 754 + ], + "spans": [ + { + "bbox": [ + 293, + 745, + 318, + 754 + ], + "type": "text", + "content": "26328" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 71, + 72, + 237, + 180 + ], + "blocks": [ + { + "bbox": [ + 71, + 72, + 237, + 180 + ], + "lines": [ + { + "bbox": [ + 71, + 72, + 237, + 180 + ], + "spans": [ + { + "bbox": [ + 71, + 72, + 237, + 180 + ], + "type": "inline_equation", + "content": "\\boldsymbol {z}\\leftarrow \\infty" + }, + { + "bbox": [ + 71, + 72, + 237, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{C}\\gets c_{\\mathrm{background}}" + }, + { + "bbox": [ + 71, + 72, + 237, + 180 + ], + "type": "text", + "content": " \nfor each Gaussian " + }, + { + "bbox": [ + 71, + 72, + 237, + 180 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 71, + 72, + 237, + 180 + ], + "type": "text", + "content": " in unsorted order do Compute " + }, + { + "bbox": [ + 71, + 72, + 237, + 180 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 71, + 72, + 237, + 180 + ], + "type": "text", + "content": " and depth " + }, + { + "bbox": [ + 71, + 72, + 237, + 180 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 71, + 72, + 237, + 180 + ], + "type": "text", + "content": " Sample " + }, + { + "bbox": [ + 71, + 72, + 237, + 180 + ], + "type": "inline_equation", + "content": "u\\sim \\mathcal{U}(0,1)" + }, + { + "bbox": [ + 71, + 72, + 237, + 180 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 71, + 72, + 237, + 180 + ], + "type": "inline_equation", + "content": "u < \\alpha_{i}" + }, + { + "bbox": [ + 71, + 72, + 237, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 71, + 72, + 237, + 180 + ], + "type": "inline_equation", + "content": "z_{i} < z" + }, + { + "bbox": [ + 71, + 72, + 237, + 180 + ], + "type": "text", + "content": " then " + }, + { + "bbox": [ + 71, + 72, + 237, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{C}\\gets c_i" + }, + { + "bbox": [ + 71, + 72, + 237, + 180 + ], + "type": "inline_equation", + "content": "z\\gets z_{i}" + }, + { + "bbox": [ + 71, + 72, + 237, + 180 + ], + "type": "text", + "content": " \nend if \nend for" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_body" + } + ], + "index": 0, + "sub_type": "code", + "guess_lang": "latex" + }, + { + "bbox": [ + 83, + 185, + 266, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 185, + 266, + 196 + ], + "spans": [ + { + "bbox": [ + 83, + 185, + 266, + 196 + ], + "type": "text", + "content": "Listing 1. Pseudocode for stochastic transparency." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 205, + 295, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 295, + 254 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 295, + 254 + ], + "type": "text", + "content": "matrix " + }, + { + "bbox": [ + 55, + 205, + 295, + 254 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 55, + 205, + 295, + 254 + ], + "type": "text", + "content": ". Stochastic transparency can also be interpreted as a randomized extension of classical screen-door transparency [10, Chapter 16], where, instead of a randomized selection, a fixed pattern is employed to discard fragments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 255, + 296, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 327 + ], + "type": "text", + "content": "Simple example. As a simple example, consider a single 3D Gaussian " + }, + { + "bbox": [ + 55, + 255, + 296, + 327 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 255, + 296, + 327 + ], + "type": "text", + "content": " covering pixel " + }, + { + "bbox": [ + 55, + 255, + 296, + 327 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 255, + 296, + 327 + ], + "type": "text", + "content": " with opacity " + }, + { + "bbox": [ + 55, + 255, + 296, + 327 + ], + "type": "inline_equation", + "content": "\\alpha = 0.5" + }, + { + "bbox": [ + 55, + 255, + 296, + 327 + ], + "type": "text", + "content": "; in this case, if we sample " + }, + { + "bbox": [ + 55, + 255, + 296, + 327 + ], + "type": "inline_equation", + "content": "x \\sim \\mathcal{U}(0,1)" + }, + { + "bbox": [ + 55, + 255, + 296, + 327 + ], + "type": "text", + "content": " and accept " + }, + { + "bbox": [ + 55, + 255, + 296, + 327 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 255, + 296, + 327 + ], + "type": "text", + "content": " only when " + }, + { + "bbox": [ + 55, + 255, + 296, + 327 + ], + "type": "inline_equation", + "content": "x > \\alpha" + }, + { + "bbox": [ + 55, + 255, + 296, + 327 + ], + "type": "text", + "content": ", approximately " + }, + { + "bbox": [ + 55, + 255, + 296, + 327 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 255, + 296, + 327 + ], + "type": "text", + "content": " of the pixel's samples will randomly select this Gaussian. We next formalize this idea to generalize to rendering an arbitrary number of Gaussians." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 330, + 296, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 330, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 296, + 365 + ], + "type": "text", + "content": "Formalization. Formally, stochastic transparency is a Monte Carlo estimator of (1). It estimates the radiance " + }, + { + "bbox": [ + 55, + 330, + 296, + 365 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 330, + 296, + 365 + ], + "type": "text", + "content": " by randomly sampling a summand " + }, + { + "bbox": [ + 55, + 330, + 296, + 365 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 330, + 296, + 365 + ], + "type": "text", + "content": " (i.e., the " + }, + { + "bbox": [ + 55, + 330, + 296, + 365 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 330, + 296, + 365 + ], + "type": "text", + "content": "-th Gaussian):" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 111, + 369, + 295, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 369, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 111, + 369, + 295, + 397 + ], + "type": "interline_equation", + "content": "C \\approx \\frac {1}{P (i)} c _ {i} \\alpha_ {i} \\prod_ {z _ {k} < z _ {i}} \\left(1 - \\alpha_ {k}\\right), \\tag {2}", + "image_path": "bba64f32c1a345c16918b8471497c06241d2c6977511e3029e81c83d1337674e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 403, + 296, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 296, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 296, + 440 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 403, + 296, + 440 + ], + "type": "inline_equation", + "content": "P(i)" + }, + { + "bbox": [ + 55, + 403, + 296, + 440 + ], + "type": "text", + "content": " is the probability mass function of the random index " + }, + { + "bbox": [ + 55, + 403, + 296, + 440 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 403, + 296, + 440 + ], + "type": "text", + "content": ". This estimator is unbiased, as long as " + }, + { + "bbox": [ + 55, + 403, + 296, + 440 + ], + "type": "inline_equation", + "content": "P(i) > 0" + }, + { + "bbox": [ + 55, + 403, + 296, + 440 + ], + "type": "text", + "content": " for any non-zero summand. The standard approach is to use:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 121, + 444, + 295, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 444, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 121, + 444, + 295, + 468 + ], + "type": "interline_equation", + "content": "P (i) = \\alpha_ {i} \\prod_ {z _ {k} < z _ {i}} \\left(1 - \\alpha_ {k}\\right). \\tag {3}", + "image_path": "5c8ea204c6e2272c5287d8b8b803b0c2eca9a8c62730dbd57ea878a0910645d6.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 474, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 498 + ], + "type": "text", + "content": "The division by this probability mass function cancels out part of the sample weight, leading to a simple estimator:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 502, + 294, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 294, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 294, + 515 + ], + "type": "interline_equation", + "content": "C \\approx c _ {i} \\quad \\text {f o r} \\quad i \\sim P. \\tag {4}", + "image_path": "c343c0ee4a921a6bbc1b3ddab12d9b5639e3c19e75d34014b07de24177370d9d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 521, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 556 + ], + "type": "text", + "content": "Thus, a pixel has the probability " + }, + { + "bbox": [ + 55, + 521, + 295, + 556 + ], + "type": "inline_equation", + "content": "P(C = c_i) = P(i)" + }, + { + "bbox": [ + 55, + 521, + 295, + 556 + ], + "type": "text", + "content": " to pick color " + }, + { + "bbox": [ + 55, + 521, + 295, + 556 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 55, + 521, + 295, + 556 + ], + "type": "text", + "content": ". In practice, we average multiple independent samples per pixel to reduce estimator noise." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 558, + 296, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 666 + ], + "type": "text", + "content": "Pseudocode. Listing 1 provides pseudocode for the above estimator. By testing that the depth of the current sample is smaller than the previously selected one, we sample exactly according to " + }, + { + "bbox": [ + 55, + 558, + 296, + 666 + ], + "type": "inline_equation", + "content": "P(i)" + }, + { + "bbox": [ + 55, + 558, + 296, + 666 + ], + "type": "text", + "content": " [8]. In the fragment Shader, this is implemented by using standard depth testing and calling discard on samples where " + }, + { + "bbox": [ + 55, + 558, + 296, + 666 + ], + "type": "inline_equation", + "content": "u \\geq \\alpha_i" + }, + { + "bbox": [ + 55, + 558, + 296, + 666 + ], + "type": "text", + "content": ". This only requires access to the current Gaussian's opacity and works for any primitive ordering. Figure 2 further shows a toy example of using this estimator on two Gaussians." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 672, + 256, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 256, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 256, + 685 + ], + "type": "text", + "content": "3.3. Differentiable stochastic transparency" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Applying stochastic transparency to inverse rendering problems requires computing the gradient of the photometric" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 395, + 184 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 395, + 184 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 395, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 395, + 184 + ], + "type": "image", + "image_path": "1e9b9805511e55f6a1a5f4fd1bf0aad3954762598c8306adf83ce74766bee03b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 187, + 387, + 198 + ], + "lines": [ + { + "bbox": [ + 329, + 187, + 387, + 198 + ], + "spans": [ + { + "bbox": [ + 329, + 187, + 387, + 198 + ], + "type": "text", + "content": "Alpha blending" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 395, + 72, + 473, + 184 + ], + "blocks": [ + { + "bbox": [ + 395, + 72, + 473, + 184 + ], + "lines": [ + { + "bbox": [ + 395, + 72, + 473, + 184 + ], + "spans": [ + { + "bbox": [ + 395, + 72, + 473, + 184 + ], + "type": "image", + "image_path": "c0b58e11d23e59ed55ed7f101145c3efd946c524547d08231490947011fa13f9.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 187, + 465, + 198 + ], + "lines": [ + { + "bbox": [ + 406, + 187, + 465, + 198 + ], + "spans": [ + { + "bbox": [ + 406, + 187, + 465, + 198 + ], + "type": "text", + "content": "Ours (128 SPP)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 474, + 72, + 552, + 184 + ], + "blocks": [ + { + "bbox": [ + 474, + 72, + 552, + 184 + ], + "lines": [ + { + "bbox": [ + 474, + 72, + 552, + 184 + ], + "spans": [ + { + "bbox": [ + 474, + 72, + 552, + 184 + ], + "type": "image", + "image_path": "bcb4c2a7b38bca3cb03c430e94d3044530be14dc7f017ed93e371194356c72c3.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 484, + 187, + 542, + 198 + ], + "lines": [ + { + "bbox": [ + 484, + 187, + 542, + 198 + ], + "spans": [ + { + "bbox": [ + 484, + 187, + 542, + 198 + ], + "type": "text", + "content": "Ours (512 SPP)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 202, + 554, + 247 + ], + "lines": [ + { + "bbox": [ + 313, + 202, + 554, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 202, + 554, + 247 + ], + "type": "text", + "content": "Figure 3. Gradient validation - We compare alpha blending gradients to our stochastic estimator with 128 and 512 SPP. Red and blue colors encode positive and negative values, respectively. Our estimator accurately approximates the alpha blending gradients." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 252, + 554, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 252, + 554, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 252, + 554, + 289 + ], + "type": "text", + "content": "loss " + }, + { + "bbox": [ + 313, + 252, + 554, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 252, + 554, + 289 + ], + "type": "text", + "content": " with respect to the primitive parameters " + }, + { + "bbox": [ + 313, + 252, + 554, + 289 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 252, + 554, + 289 + ], + "type": "text", + "content": ". Primarily, we require gradients with respect to color " + }, + { + "bbox": [ + 313, + 252, + 554, + 289 + ], + "type": "inline_equation", + "content": "\\partial \\mathcal{L} / \\partial c" + }, + { + "bbox": [ + 313, + 252, + 554, + 289 + ], + "type": "text", + "content": ", and gradients with respect to opacity " + }, + { + "bbox": [ + 313, + 252, + 554, + 289 + ], + "type": "inline_equation", + "content": "\\partial \\mathcal{L} / \\partial \\alpha" + }, + { + "bbox": [ + 313, + 252, + 554, + 289 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 290, + 554, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 554, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 554, + 350 + ], + "type": "text", + "content": "Gradient estimator. We take inspiration from the differentiable rendering literature [46], and use a detached gradient estimator [46, 53] to correctly differentiate the binary sampling decision in Listing 1. We derive this estimator by applying the gradient operator to the alpha blending equation:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 353, + 357, + 553, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 357, + 553, + 389 + ], + "spans": [ + { + "bbox": [ + 353, + 357, + 553, + 389 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\boldsymbol {\\theta}} C = \\sum_ {i = 1} ^ {L} \\nabla_ {\\boldsymbol {\\theta}} \\left[ c _ {i} \\alpha_ {i} \\prod_ {z _ {k} < z _ {i}} (1 - \\alpha_ {k}) \\right]. \\tag {5}", + "image_path": "2112ac32edb74970b120310471ab5345cd13c29e98ecc7cdc98359d7170206d8.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 396, + 554, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 554, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 554, + 433 + ], + "type": "text", + "content": "We then form an unbiased Monte Carlo gradient estimator by again sampling a summand according to " + }, + { + "bbox": [ + 313, + 396, + 554, + 433 + ], + "type": "inline_equation", + "content": "P(i)" + }, + { + "bbox": [ + 313, + 396, + 554, + 433 + ], + "type": "text", + "content": ". Crucially, we explicitly do not differentiate " + }, + { + "bbox": [ + 313, + 396, + 554, + 433 + ], + "type": "inline_equation", + "content": "P(i)" + }, + { + "bbox": [ + 313, + 396, + 554, + 433 + ], + "type": "text", + "content": " itself and obtain:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 350, + 439, + 553, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 439, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 350, + 439, + 553, + 472 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\boldsymbol {\\theta}} C \\approx \\frac {1}{P (i)} \\nabla_ {\\boldsymbol {\\theta}} \\left[ c _ {i} \\alpha_ {i} \\prod_ {z _ {k} < z _ {i}} \\left(1 - \\alpha_ {k}\\right) \\right]. \\tag {6}", + "image_path": "abd655d40f7d740adc739e9d078c82e43b43cbc40814838767234fccea1629b4.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 479, + 553, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 553, + 503 + ], + "type": "text", + "content": "Substituting the definition of " + }, + { + "bbox": [ + 313, + 479, + 553, + 503 + ], + "type": "inline_equation", + "content": "P(i)" + }, + { + "bbox": [ + 313, + 479, + 553, + 503 + ], + "type": "text", + "content": " from (3) and differentiating " + }, + { + "bbox": [ + 313, + 479, + 553, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 479, + 553, + 503 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 313, + 479, + 553, + 503 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 479, + 553, + 503 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 479, + 553, + 503 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 479, + 553, + 503 + ], + "type": "text", + "content": ", we get:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 332, + 509, + 553, + 535 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 509, + 553, + 535 + ], + "spans": [ + { + "bbox": [ + 332, + 509, + 553, + 535 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\mathcal {L}}{\\partial c _ {i}} = \\frac {\\partial \\mathcal {L}}{\\partial C} \\quad \\frac {\\partial \\mathcal {L}}{\\partial c _ {z _ {k} < z _ {i}}} = 0 \\tag {7}", + "image_path": "189b8cfc7166cbd4d9d635aa90284541ad34b7ab4cd585af4108e24b4646ee75.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 331, + 536, + 553, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 536, + 553, + 563 + ], + "spans": [ + { + "bbox": [ + 331, + 536, + 553, + 563 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\mathcal {L}}{\\partial \\alpha_ {i}} = \\frac {\\partial \\mathcal {L}}{\\partial C} \\frac {c _ {i}}{\\alpha_ {i}} \\quad \\frac {\\partial \\mathcal {L}}{\\partial \\alpha_ {z _ {k} < z _ {i}}} = \\frac {\\partial \\mathcal {L}}{\\partial C} \\frac {- c _ {i}}{1 - \\alpha_ {z _ {k} < z _ {i}}}, \\tag {8}", + "image_path": "e24fc84d7560e8254718946308f4242deaeb8ef5dcd8852df8152b8d57a3cc74.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 568, + 554, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 554, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 554, + 687 + ], + "type": "text", + "content": "where subscript " + }, + { + "bbox": [ + 313, + 568, + 554, + 687 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 568, + 554, + 687 + ], + "type": "text", + "content": " denotes the sampled primitive, and " + }, + { + "bbox": [ + 313, + 568, + 554, + 687 + ], + "type": "inline_equation", + "content": "\\partial \\mathcal{L} / \\partial C" + }, + { + "bbox": [ + 313, + 568, + 554, + 687 + ], + "type": "text", + "content": " is the photometric loss gradient, dependent upon the loss formulation (L1, SSIM, ...). Note how in (7) only the " + }, + { + "bbox": [ + 313, + 568, + 554, + 687 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 568, + 554, + 687 + ], + "type": "text", + "content": "-th Gaussian receives gradients with respect to its color, while the others have a zero gradient. Meanwhile, (8) shows how the " + }, + { + "bbox": [ + 313, + 568, + 554, + 687 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 568, + 554, + 687 + ], + "type": "text", + "content": "-th Gaussian, and ones closer to the camera receive a gradient with respect to their opacity. The Gaussians behind the sampled one do not receive any gradient. Finally, our derivation is not specific to 3DGS, and applies to any representation rendered with stochastic transparency." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 689, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 712 + ], + "type": "text", + "content": "Decorrelated loss gradient. A well-known challenge for differentiable Monte Carlo methods is that the derivative of" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 745, + 317, + 754 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 745, + 317, + 754 + ], + "spans": [ + { + "bbox": [ + 293, + 745, + 317, + 754 + ], + "type": "text", + "content": "26329" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 175, + 153 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 175, + 153 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 175, + 153 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 175, + 153 + ], + "type": "image", + "image_path": "8db56bfdffccd852f2a8fa615395d868b543caf740b4394f6914c8b2ada454a1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 106, + 158, + 130, + 168 + ], + "lines": [ + { + "bbox": [ + 106, + 158, + 130, + 168 + ], + "spans": [ + { + "bbox": [ + 106, + 158, + 130, + 168 + ], + "type": "text", + "content": "3DGS" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 176, + 69, + 294, + 154 + ], + "blocks": [ + { + "bbox": [ + 176, + 69, + 294, + 154 + ], + "lines": [ + { + "bbox": [ + 176, + 69, + 294, + 154 + ], + "spans": [ + { + "bbox": [ + 176, + 69, + 294, + 154 + ], + "type": "image", + "image_path": "8626203bc7b47627c1cf4230a6699b27458a03e6f97f28f423ad7ed9c440aeee.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 188, + 158, + 284, + 169 + ], + "lines": [ + { + "bbox": [ + 188, + 158, + 284, + 169 + ], + "spans": [ + { + "bbox": [ + 188, + 158, + 284, + 169 + ], + "type": "text", + "content": "StopThePop [37] vs. ours" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 173, + 296, + 274 + ], + "lines": [ + { + "bbox": [ + 55, + 173, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 173, + 296, + 274 + ], + "type": "text", + "content": "Figure 4. Popping in 3DGS - (left) As sorting in 3DGS is done with respect to the " + }, + { + "bbox": [ + 55, + 173, + 296, + 274 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 173, + 296, + 274 + ], + "type": "text", + "content": " distance of the Gaussian mean from the camera, a small camera rotation can cause visible \"popping\" artifacts (i.e., sudden pixel color changes). (right) StopThePop [37] corrects for this behavior by associating a surface with each Gaussian, and determining the depth " + }, + { + "bbox": [ + 55, + 173, + 296, + 274 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 173, + 296, + 274 + ], + "type": "text", + "content": " per-fragment (dashed line), rather than per-Gaussian. Our solution leads to visually comparable results, but approximates this surface linearly (solid line), so that per-fragment depth " + }, + { + "bbox": [ + 55, + 173, + 296, + 274 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 173, + 296, + 274 + ], + "type": "text", + "content": " can be computed efficiently in hardware." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 282, + 296, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 282, + 296, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 296, + 307 + ], + "type": "text", + "content": "the loss function is a product of two terms that are corrupted by noise (i.e., they are both random variables):" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 142, + 311, + 295, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 311, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 142, + 311, + 295, + 335 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\mathcal {L}}{\\partial \\theta} = \\frac {\\partial \\mathcal {L}}{\\partial C} \\frac {\\partial C}{\\partial \\theta}. \\tag {9}", + "image_path": "13eb9a0371957973b9af45a94cf273ea7ba64ecdc0cbe1e8d93c969685750ba1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 339, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 296, + 434 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 339, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 339, + 296, + 434 + ], + "type": "text", + "content": " are the geometry parameters. If the noise in both terms is correlated, we get biased gradients, which harm optimization. Following [2, 12], we use different random samples to evaluate the two factors above. If " + }, + { + "bbox": [ + 55, + 339, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 339, + 296, + 434 + ], + "type": "text", + "content": " is the L2 loss, the decorrelated evaluation results in unbiased gradients. For other loss functions, it at least reduces bias. Unbiased gradient estimation for arbitrary loss functions applied to a Monte Carlo estimator is an open problem." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 437, + 296, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 296, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 296, + 484 + ], + "type": "text", + "content": "Practical algorithm. The presence of " + }, + { + "bbox": [ + 55, + 437, + 296, + 484 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 55, + 437, + 296, + 484 + ], + "type": "text", + "content": " in (8) implies that we need access to the result of the forward pass to compute gradients. We therefore follow path replay backpropagation [46] and compute gradients in three passes:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 484, + 295, + 556 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 484, + 295, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 484, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 56, + 484, + 295, + 496 + ], + "type": "text", + "content": "1. In a first pass, we render an image and evaluate " + }, + { + "bbox": [ + 56, + 484, + 295, + 496 + ], + "type": "inline_equation", + "content": "\\partial \\mathcal{L} / \\partial C" + }, + { + "bbox": [ + 56, + 484, + 295, + 496 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 496, + 295, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 496, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 496, + 295, + 520 + ], + "type": "text", + "content": "2. We then render a decorrelated image using a different random seed, giving us " + }, + { + "bbox": [ + 56, + 496, + 295, + 520 + ], + "type": "inline_equation", + "content": "c_{i}" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 521, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 521, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 56, + 521, + 295, + 556 + ], + "type": "text", + "content": "3. Finally, we replay the second pass by re-rendering using the same random seed and using the stored " + }, + { + "bbox": [ + 56, + 521, + 295, + 556 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 56, + 521, + 295, + 556 + ], + "type": "text", + "content": " to compute parameter gradients." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 558, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 666 + ], + "type": "text", + "content": "Validation. We validate our estimator by comparing to the gradients of the sorted alpha blending implementation. For easier visual interpretation, we use our reverse-mode gradient implementation to compute a gradient image. We do this by backpropagating an L1 loss separately for each pixel, and visualizing the sum of the x-component of the positional gradients of all Gaussians affecting a given pixel. As shown in Figure 3, our method produces noisy gradients that closely match those of alpha blending in structure." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 672, + 209, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 209, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 209, + 685 + ], + "type": "text", + "content": "3.4. Removing popping artifacts" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "A naive implementation of stochastic transparency re-uses the billboard geometry of 3DGS, and consequently, to com" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 320, + 73, + 395, + 138 + ], + "blocks": [ + { + "bbox": [ + 320, + 73, + 395, + 138 + ], + "lines": [ + { + "bbox": [ + 320, + 73, + 395, + 138 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 395, + 138 + ], + "type": "image", + "image_path": "65811c20896c8e54a0c56d773b57252a7216c16c7f8009af2b902840b30886a3.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 327, + 140, + 384, + 151 + ], + "lines": [ + { + "bbox": [ + 327, + 140, + 384, + 151 + ], + "spans": [ + { + "bbox": [ + 327, + 140, + 384, + 151 + ], + "type": "text", + "content": "Alpha blending" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 152, + 553, + 196 + ], + "lines": [ + { + "bbox": [ + 313, + 152, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 553, + 196 + ], + "type": "text", + "content": "Figure 5. Fully volumetric intermixing - While alpha blending (left) suffers from popping and \"pop-free\" variants (middle) show discontinuities, our full volumetric intermixing (right) accurately composes overlapping Gaussians." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 397, + 73, + 472, + 138 + ], + "blocks": [ + { + "bbox": [ + 397, + 73, + 472, + 138 + ], + "lines": [ + { + "bbox": [ + 397, + 73, + 472, + 138 + ], + "spans": [ + { + "bbox": [ + 397, + 73, + 472, + 138 + ], + "type": "image", + "image_path": "b61a6b8bfe2048be453292bc26317187a4b27dbb47d4755edecedd740b26db76.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 416, + 140, + 450, + 150 + ], + "lines": [ + { + "bbox": [ + 416, + 140, + 450, + 150 + ], + "spans": [ + { + "bbox": [ + 416, + 140, + 450, + 150 + ], + "type": "text", + "content": "Pop-free" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 474, + 73, + 550, + 138 + ], + "blocks": [ + { + "bbox": [ + 474, + 73, + 550, + 138 + ], + "lines": [ + { + "bbox": [ + 474, + 73, + 550, + 138 + ], + "spans": [ + { + "bbox": [ + 474, + 73, + 550, + 138 + ], + "type": "image", + "image_path": "a3050a1698f8d6dac1529063e9f1f8241f9a529355e01c833688993095e1d32a.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 474, + 140, + 548, + 151 + ], + "lines": [ + { + "bbox": [ + 474, + 140, + 548, + 151 + ], + "spans": [ + { + "bbox": [ + 474, + 140, + 548, + 151 + ], + "type": "text", + "content": "Our full intermixing" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 201, + 553, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 553, + 285 + ], + "type": "text", + "content": "pute the depths " + }, + { + "bbox": [ + 313, + 201, + 553, + 285 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 313, + 201, + 553, + 285 + ], + "type": "text", + "content": ", the means of the Gaussians. As we will see in Section 4.2, at low sample count this provides high rendering efficiency, as we do not require a global sort. However, popping artifacts still exist, as the computation of " + }, + { + "bbox": [ + 313, + 201, + 553, + 285 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 313, + 201, + 553, + 285 + ], + "type": "text", + "content": " is analogous to the one in 3DGS, small changes in the camera rotation can therefore change the order of primitives, and cause visible video artifacts; see Figure 4." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 287, + 554, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 554, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 554, + 467 + ], + "type": "text", + "content": "The stop-the-pop way. Radl et al. [37] overcome the popping problem by computing the per-fragment (vs. per-Gaussian) depth. Their core idea is to compute the location of the peak of the Gaussian density along the ray. When coupled with a (computationally expensive) per-pixel sorting of Gaussians, this can provide a pop-free rendering. This can be interpreted as finding \"maximum density surface\" of the Gaussian, visualized in 2D in Figure 4, and using it to define fragment depths. However, per-fragment updates of depth are not particularly efficient, as early-discard mechanisms implemented in hardware [13] can only execute for fragments whose depth has not been modified by a shader. While in our setting we can forego sorting, the inability to early-discard fragments still affects us, which lead us to propose a different approach, described next." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 468, + 554, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 554, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 554, + 528 + ], + "type": "text", + "content": "Our simplified way. We propose to avoid this inefficiency by foregoing per-fragment depth changes. Instead, we orient our billboards to linearly approximate the maximum density surface. If the ray associated with a pixel has origin o, the plane can be defined as:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 325, + 532, + 553, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 532, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 325, + 532, + 553, + 548 + ], + "type": "interline_equation", + "content": "\\mathbf {n} ^ {\\top} (\\mathbf {x} - \\boldsymbol {\\mu}) = 0, \\quad \\text {w h e r e} \\quad \\mathbf {n} = \\Sigma^ {- 1} (\\boldsymbol {\\mu} - \\mathbf {o}) \\tag {10}", + "image_path": "f1cdff90eea57407de1fc1afb76b8cf48c898dcb7df74496a84bf71936c67b5f.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 552, + 554, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 554, + 612 + ], + "type": "text", + "content": "By approximating this surface with a plane, we can easily use the interpolation kernels within the hardware rasterizer to compute the splat's depth. This approach adds no extra time to the render process and gives us popping-free rendering with a similar PSNR to [37] at higher sample counts." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 613, + 554, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 554, + 716 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 554, + 716 + ], + "type": "text", + "content": "Fully volumetric intermixing. If maximum rendering fidelity is desired and the cost of per-fragment depth modification is acceptable, stochastic transparency can naturally support full volumetric intermixing. The key insight is that this can be achieved by replacing the fixed depth of a Gaussian by a sampled free-flight distance. For this, we first express the total contribution of a single Gaussian using the Beer-Lambert law: " + }, + { + "bbox": [ + 313, + 613, + 554, + 716 + ], + "type": "inline_equation", + "content": "\\int_0^\\infty \\sigma_i(\\mathbf{x}_t)\\exp \\left(-\\int_0^t\\sigma_i(\\mathbf{x}_s)\\mathrm{d}s\\right)c_i\\mathrm{d}t," + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 745, + 318, + 754 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 745, + 318, + 754 + ], + "spans": [ + { + "bbox": [ + 293, + 745, + 318, + 754 + ], + "type": "text", + "content": "26330" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "content": "where the opacity of a Gaussian is reparametrized using a spatially-varying extinction function " + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_a\\coloneqq \\mathbf{o} + a\\cdot \\mathbf{d}" + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "content": " for a ray " + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "inline_equation", + "content": "(\\mathbf{o},\\mathbf{d})" + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "content": ". We follow [5] and sample a free-flight distance " + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "content": " proportional to " + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "inline_equation", + "content": "p(t) = \\sigma_i(\\mathbf{x}_t)\\exp (-\\int_0^t\\sigma_i(\\mathbf{x}_s)\\mathrm{d}s)" + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "content": ". To handle overlapping Gaussians, we adapt decomposition tracking [33]: If two Gaussians overlap, we sample a free-flight distance according to each and take the minimum. This accurately resolves overlap of an arbitrary number of Gaussians without additional data structures, sorting, or ray tracing, and readily fits into the stochastic transparency algorithm. We simply need to replace the computation of " + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "content": " by a probabilistic sampling " + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "inline_equation", + "content": "z_{i}\\sim p(t)" + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "content": ". We show an example rendering of two overlapping Gaussians in Fig. 5. See the supplemental for the full free-flight sampling routine." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 246, + 136, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 246, + 136, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 246, + 136, + 258 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 265, + 296, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 296, + 326 + ], + "type": "text", + "content": "We evaluate rendering performance on the real-world scenes from the MipNerf360 [4] dataset. For demonstrating semantic object localization with single-sample rendering we use the LERF dataset [19] - it consists of complex in-the-wild scenes, designed for 3D object localization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 327, + 296, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 388 + ], + "type": "text", + "content": "Baselines. For rendering baselines, we compare the performance of our method against the original 3DGS implementation [18] that uses CUDA, as well as Splatapult [42], an equivalent OpenGL implementation. We also compare with StopThePop [37], a pop-free method for 3DGS rendering." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 388, + 296, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 296, + 449 + ], + "type": "text", + "content": "Metrics. We compute the standard metrics of peak signal-to-noise ratio (PSNR), structural similarity index metric (SSIM) [47], and learned perceptual image patch similarity (LPIPS) [55]. We report the average per-frame processing time for all methods on the MipNerf360 scenes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 450, + 296, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 296, + 654 + ], + "type": "text", + "content": "Implementation details. To demonstrate portability, we implement our method in both CUDA and OpenGL. Our CUDA implementation builds on the existing 3DGS framework by Kerbl et al. [18] and our OpenGL implementation builds upon Splatapult [42]. We evaluate on a Windows laptop equipped with an NVIDIA Quadro T1000 Max-Q GPU, and Linux desktops with an NVIDIA RTX 3090 and RTX 4090. Rather than training from scratch, we finetune a trained 3DGS scene with the CUDA implementation of our method for 1000 iterations. Since we are performing fine-tuning, we do not apply densification or pruning. All learning rates for the non-position parameters are kept unchanged, while the positions are optimized using a learning rate of " + }, + { + "bbox": [ + 55, + 450, + 296, + 654 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-5}" + }, + { + "bbox": [ + 55, + 450, + 296, + 654 + ], + "type": "text", + "content": " with 128 samples. We keep all other hyperparameters the same. On average, the finetuning process requires approximately 14 minutes on an NVIDIA RTX 3090 GPU for scenes from MipNerf360." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 660, + 163, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 163, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 163, + 673 + ], + "type": "text", + "content": "4.1. Rendering quality" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Qualitative results - Figs. 6 and 7. We show two scenes (Room and Kitchen) in Figure 7 from the MipN-erf360 dataset. As shown, our method allows a trade-off" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 320, + 72, + 395, + 159 + ], + "blocks": [ + { + "bbox": [ + 320, + 72, + 395, + 159 + ], + "lines": [ + { + "bbox": [ + 320, + 72, + 395, + 159 + ], + "spans": [ + { + "bbox": [ + 320, + 72, + 395, + 159 + ], + "type": "image", + "image_path": "5e63d90586df3c2513d981957964f45da53f3c9ddab16779fdbf1fd524675829.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 397, + 72, + 472, + 159 + ], + "blocks": [ + { + "bbox": [ + 397, + 72, + 472, + 159 + ], + "lines": [ + { + "bbox": [ + 397, + 72, + 472, + 159 + ], + "spans": [ + { + "bbox": [ + 397, + 72, + 472, + 159 + ], + "type": "image", + "image_path": "79035a8d0371cc0b2520044c7d1b61b318b7219148d6b198bb49b2af58dc3285.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 475, + 72, + 550, + 159 + ], + "blocks": [ + { + "bbox": [ + 475, + 72, + 550, + 159 + ], + "lines": [ + { + "bbox": [ + 475, + 72, + 550, + 159 + ], + "spans": [ + { + "bbox": [ + 475, + 72, + 550, + 159 + ], + "type": "image", + "image_path": "847e750c81fb2077154732fde3cf7e1f03ed371f01bd0c9a794f6aaebedf3bba.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 319, + 160, + 394, + 246 + ], + "blocks": [ + { + "bbox": [ + 319, + 160, + 394, + 246 + ], + "lines": [ + { + "bbox": [ + 319, + 160, + 394, + 246 + ], + "spans": [ + { + "bbox": [ + 319, + 160, + 394, + 246 + ], + "type": "image", + "image_path": "33d21d8180d064514545a36e8f40dc68c628d8d3a3ab57d3fda7731c7811c224.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 248, + 378, + 258 + ], + "lines": [ + { + "bbox": [ + 348, + 248, + 378, + 258 + ], + "spans": [ + { + "bbox": [ + 348, + 248, + 378, + 258 + ], + "type": "text", + "content": "Frame " + }, + { + "bbox": [ + 348, + 248, + 378, + 258 + ], + "type": "inline_equation", + "content": "t" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 397, + 160, + 472, + 246 + ], + "blocks": [ + { + "bbox": [ + 397, + 160, + 472, + 246 + ], + "lines": [ + { + "bbox": [ + 397, + 160, + 472, + 246 + ], + "spans": [ + { + "bbox": [ + 397, + 160, + 472, + 246 + ], + "type": "image", + "image_path": "49a1ec46c55484282d557be2f0d68965465ea41ea0c786836cec8d8248dd3ab2.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 248, + 455, + 258 + ], + "lines": [ + { + "bbox": [ + 413, + 248, + 455, + 258 + ], + "spans": [ + { + "bbox": [ + 413, + 248, + 455, + 258 + ], + "type": "text", + "content": "Frame " + }, + { + "bbox": [ + 413, + 248, + 455, + 258 + ], + "type": "inline_equation", + "content": "t + 1" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 475, + 160, + 550, + 246 + ], + "blocks": [ + { + "bbox": [ + 475, + 160, + 550, + 246 + ], + "lines": [ + { + "bbox": [ + 475, + 160, + 550, + 246 + ], + "spans": [ + { + "bbox": [ + 475, + 160, + 550, + 246 + ], + "type": "image", + "image_path": "c6ffa01c056fc62f02908df7d4a72ff50fabba06c7bd66405098288c01bac93a.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 485, + 248, + 525, + 258 + ], + "lines": [ + { + "bbox": [ + 485, + 248, + 525, + 258 + ], + "spans": [ + { + "bbox": [ + 485, + 248, + 525, + 258 + ], + "type": "text", + "content": "FLIP error" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 262, + 555, + 327 + ], + "lines": [ + { + "bbox": [ + 313, + 262, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 555, + 327 + ], + "type": "text", + "content": "Figure 6. Popping artifact - The top row displays two consecutive frames, fine-tuned and rendered using our StochasticSplats, along with their corresponding fLIP error map [1]. The bottom row presents the same frames processed with standard 3DGS. The foreground error indicates popping while the background errors are due to the camera motion." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 317, + 330, + 555, + 377 + ], + "blocks": [ + { + "bbox": [ + 317, + 330, + 555, + 377 + ], + "lines": [ + { + "bbox": [ + 317, + 330, + 555, + 377 + ], + "spans": [ + { + "bbox": [ + 317, + 330, + 555, + 377 + ], + "type": "table", + "html": "
3DGSStopThePop14Ours (SPP)
16642561024
PSNR28.9928.7917.9422.7526.1127.6128.1728.33
SSIM0.8690.8700.2840.5080.7050.8030.8400.852
LPIPS0.1850.1810.6030.4810.3390.2340.1870.186
", + "image_path": "b07b6e83845d80bde5303e7123a2225806d2e51afd65689f2f471d09782625c7.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 381, + 555, + 426 + ], + "lines": [ + { + "bbox": [ + 313, + 381, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 555, + 426 + ], + "type": "text", + "content": "Table 1. Novel view synthesis - quantitative results of different methods on MipNerf360 scenes. StochasticSplats is finetuned on StopThePop-trained scenes using 128 SPP, and rendered using varying number of samples." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 432, + 554, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 554, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 554, + 479 + ], + "type": "text", + "content": "between quality and rendering time. Figure 6 further highlights that popping is handled properly. Note that, due to the sorting-free nature of our method, any popping free formulations can be easily integrated into our method." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 482, + 555, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 555, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 555, + 638 + ], + "type": "text", + "content": "Quantitative results - Tab. 1. We report the PSNR values of our fine-tuned scenes rendered with varying numbers of samples, alongside comparisons to 3DGS and StopThePop. We use our pop compensation from Section 3.4. By increasing the number of samples per pixel, our method approaches the quality of alpha blending. While the Monte Carlo noise affects metric, i.i.d. noise is visually often less distracting than structured artifacts; recent research suggests that the human visual system tolerates some unbiased noise, especially at interactive frame rates [31]. Additionally, as we show in Section 4.4 and Section 4.4, this can be mitigated via temporal anti-aliasing, and even without any mitigation, does not harm downstream perception tasks." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 646, + 423, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 646, + 423, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 646, + 423, + 658 + ], + "type": "text", + "content": "4.2. Computation time" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "content": "Forward pass timings - Tab. 2. We compare the forward rendering time of our method implemented in OpenGL, against 3DGS on both CUDA and OpenGL, as well as StopThePop [37] on CUDA. We report timings with a vary-" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 745, + 317, + 754 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 745, + 317, + 754 + ], + "spans": [ + { + "bbox": [ + 293, + 745, + 317, + 754 + ], + "type": "text", + "content": "26331" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 76, + 157, + 140 + ], + "blocks": [ + { + "bbox": [ + 61, + 76, + 157, + 140 + ], + "lines": [ + { + "bbox": [ + 61, + 76, + 157, + 140 + ], + "spans": [ + { + "bbox": [ + 61, + 76, + 157, + 140 + ], + "type": "image", + "image_path": "192b0f54fc66ad55ba371f2a06a25e87d44ae430868fef44380b6ae0ab01f42f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 160, + 76, + 256, + 140 + ], + "blocks": [ + { + "bbox": [ + 160, + 76, + 256, + 140 + ], + "lines": [ + { + "bbox": [ + 160, + 76, + 256, + 140 + ], + "spans": [ + { + "bbox": [ + 160, + 76, + 256, + 140 + ], + "type": "image", + "image_path": "33e226995a4655fa49dfded435141adbb50643b5b87defd0b5202eb13ee9dc04.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 258, + 76, + 353, + 140 + ], + "blocks": [ + { + "bbox": [ + 258, + 76, + 353, + 140 + ], + "lines": [ + { + "bbox": [ + 258, + 76, + 353, + 140 + ], + "spans": [ + { + "bbox": [ + 258, + 76, + 353, + 140 + ], + "type": "image", + "image_path": "9edb61efb938a063d47fe0b119b30edcdcf6c4daf3f4889818297a648eaf8e17.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 356, + 76, + 451, + 140 + ], + "blocks": [ + { + "bbox": [ + 356, + 76, + 451, + 140 + ], + "lines": [ + { + "bbox": [ + 356, + 76, + 451, + 140 + ], + "spans": [ + { + "bbox": [ + 356, + 76, + 451, + 140 + ], + "type": "image", + "image_path": "5bd3908a0499e508975188293c03a2bc78e407c00b4f724229643efebbbfdf3d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 453, + 76, + 549, + 140 + ], + "blocks": [ + { + "bbox": [ + 453, + 76, + 549, + 140 + ], + "lines": [ + { + "bbox": [ + 453, + 76, + 549, + 140 + ], + "spans": [ + { + "bbox": [ + 453, + 76, + 549, + 140 + ], + "type": "image", + "image_path": "1c8d299a115e624295e75492b396776f3c85973cd6aa4fda6fd4bc755cf737ee.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 61, + 144, + 157, + 240 + ], + "blocks": [ + { + "bbox": [ + 61, + 144, + 157, + 240 + ], + "lines": [ + { + "bbox": [ + 61, + 144, + 157, + 240 + ], + "spans": [ + { + "bbox": [ + 61, + 144, + 157, + 240 + ], + "type": "image", + "image_path": "343eef2f232e9f5e73e16fbf6b4ee86fb1aff2bd9b74c7240a56e7b0876d1bb4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 160, + 144, + 255, + 240 + ], + "blocks": [ + { + "bbox": [ + 160, + 144, + 255, + 240 + ], + "lines": [ + { + "bbox": [ + 160, + 144, + 255, + 240 + ], + "spans": [ + { + "bbox": [ + 160, + 144, + 255, + 240 + ], + "type": "image", + "image_path": "f125978960767cb7dc86f0f294d0bba3901a5b997f05c27defb21ebb58bd849b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 258, + 144, + 353, + 240 + ], + "blocks": [ + { + "bbox": [ + 258, + 144, + 353, + 240 + ], + "lines": [ + { + "bbox": [ + 258, + 144, + 353, + 240 + ], + "spans": [ + { + "bbox": [ + 258, + 144, + 353, + 240 + ], + "type": "image", + "image_path": "eb21155a9d8ef7c25de658c93d9a8ebdb16a77c5494905175ef27b422f72fd61.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 356, + 144, + 451, + 240 + ], + "blocks": [ + { + "bbox": [ + 356, + 144, + 451, + 240 + ], + "lines": [ + { + "bbox": [ + 356, + 144, + 451, + 240 + ], + "spans": [ + { + "bbox": [ + 356, + 144, + 451, + 240 + ], + "type": "image", + "image_path": "3d800f0cf180713551722432461845476304e8daa6d475c26de5af787f5faf78.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 453, + 144, + 550, + 240 + ], + "blocks": [ + { + "bbox": [ + 453, + 144, + 550, + 240 + ], + "lines": [ + { + "bbox": [ + 453, + 144, + 550, + 240 + ], + "spans": [ + { + "bbox": [ + 453, + 144, + 550, + 240 + ], + "type": "image", + "image_path": "7270906033fa29db918943972b913e7bf88974470ed0ffb5fcfb2cc468c0d4c0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 61, + 248, + 157, + 313 + ], + "blocks": [ + { + "bbox": [ + 61, + 248, + 157, + 313 + ], + "lines": [ + { + "bbox": [ + 61, + 248, + 157, + 313 + ], + "spans": [ + { + "bbox": [ + 61, + 248, + 157, + 313 + ], + "type": "image", + "image_path": "8a0281dc118751ff326ef287ff584686bf9b316009f68c13469bd54bbbca535d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 160, + 248, + 255, + 313 + ], + "blocks": [ + { + "bbox": [ + 160, + 248, + 255, + 313 + ], + "lines": [ + { + "bbox": [ + 160, + 248, + 255, + 313 + ], + "spans": [ + { + "bbox": [ + 160, + 248, + 255, + 313 + ], + "type": "image", + "image_path": "56fbe39da250b90346dd92696bcc253abcbf994a867387578a627b2b36e46f4a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 257, + 248, + 353, + 313 + ], + "blocks": [ + { + "bbox": [ + 257, + 248, + 353, + 313 + ], + "lines": [ + { + "bbox": [ + 257, + 248, + 353, + 313 + ], + "spans": [ + { + "bbox": [ + 257, + 248, + 353, + 313 + ], + "type": "image", + "image_path": "c3292f860425a4825b58b6df4d9fb490dbf6d0928ac69d94b8708556ab28cd9f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 356, + 248, + 451, + 313 + ], + "blocks": [ + { + "bbox": [ + 356, + 248, + 451, + 313 + ], + "lines": [ + { + "bbox": [ + 356, + 248, + 451, + 313 + ], + "spans": [ + { + "bbox": [ + 356, + 248, + 451, + 313 + ], + "type": "image", + "image_path": "b8ef27816aae8b110ce614fc292bc9162057b49370c69c02d983f3cb104f689a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 453, + 248, + 550, + 313 + ], + "blocks": [ + { + "bbox": [ + 453, + 248, + 550, + 313 + ], + "lines": [ + { + "bbox": [ + 453, + 248, + 550, + 313 + ], + "spans": [ + { + "bbox": [ + 453, + 248, + 550, + 313 + ], + "type": "image", + "image_path": "7a3db570cd5a7f3853e3148ffe739d168119b6e4b5ce247497ae885079e23ca3.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 60, + 316, + 157, + 413 + ], + "blocks": [ + { + "bbox": [ + 60, + 316, + 157, + 413 + ], + "lines": [ + { + "bbox": [ + 60, + 316, + 157, + 413 + ], + "spans": [ + { + "bbox": [ + 60, + 316, + 157, + 413 + ], + "type": "image", + "image_path": "8a20bf59b5b4caa85d52272db8d9e38cf4581e511d94684edb0f500dc4f89dbf.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 415, + 121, + 423 + ], + "lines": [ + { + "bbox": [ + 99, + 415, + 121, + 423 + ], + "spans": [ + { + "bbox": [ + 99, + 415, + 121, + 423 + ], + "type": "text", + "content": "1 SPP" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 159, + 316, + 255, + 413 + ], + "blocks": [ + { + "bbox": [ + 159, + 316, + 255, + 413 + ], + "lines": [ + { + "bbox": [ + 159, + 316, + 255, + 413 + ], + "spans": [ + { + "bbox": [ + 159, + 316, + 255, + 413 + ], + "type": "image", + "image_path": "6479be1972017f0f3dd3fb18f9b565674c343edf20a730afe88707472a437ff9.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 415, + 219, + 424 + ], + "lines": [ + { + "bbox": [ + 195, + 415, + 219, + 424 + ], + "spans": [ + { + "bbox": [ + 195, + 415, + 219, + 424 + ], + "type": "text", + "content": "4 SPP" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 257, + 316, + 353, + 413 + ], + "blocks": [ + { + "bbox": [ + 257, + 316, + 353, + 413 + ], + "lines": [ + { + "bbox": [ + 257, + 316, + 353, + 413 + ], + "spans": [ + { + "bbox": [ + 257, + 316, + 353, + 413 + ], + "type": "image", + "image_path": "a8c432404dae02891d5b9007a1ed3441e13c8c9a690410499c55f3e1e20d5aba.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 291, + 415, + 318, + 424 + ], + "lines": [ + { + "bbox": [ + 291, + 415, + 318, + 424 + ], + "spans": [ + { + "bbox": [ + 291, + 415, + 318, + 424 + ], + "type": "text", + "content": "16 SPP" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 356, + 316, + 451, + 413 + ], + "blocks": [ + { + "bbox": [ + 356, + 316, + 451, + 413 + ], + "lines": [ + { + "bbox": [ + 356, + 316, + 451, + 413 + ], + "spans": [ + { + "bbox": [ + 356, + 316, + 451, + 413 + ], + "type": "image", + "image_path": "481fd1a276bc5f03a09f1f71bcc63411c31d0daeb1aec6ed8cacefa56126913a.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 415, + 417, + 424 + ], + "lines": [ + { + "bbox": [ + 386, + 415, + 417, + 424 + ], + "spans": [ + { + "bbox": [ + 386, + 415, + 417, + 424 + ], + "type": "text", + "content": "128 SPP" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 427, + 555, + 460 + ], + "lines": [ + { + "bbox": [ + 55, + 427, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 427, + 555, + 460 + ], + "type": "text", + "content": "Figure 7. Qualitative results - for the Room and Kitchen scenes from the MipNerf360 dataset are shown in Fig. 7. We compare StochasticSplats using 1, 4, and 16 samples per pixel against the standard alpha blending approach. Even with just 1 sample per pixel, our method produces reasonable renderings, and with 16 samples, the results closely match the quality of alpha blending." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 453, + 316, + 550, + 413 + ], + "blocks": [ + { + "bbox": [ + 453, + 316, + 550, + 413 + ], + "lines": [ + { + "bbox": [ + 453, + 316, + 550, + 413 + ], + "spans": [ + { + "bbox": [ + 453, + 316, + 550, + 413 + ], + "type": "image", + "image_path": "c5ccdb50b4e768f9159400504113d6ad01ce558c43d60bb5cc45824e9d37882b.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 469, + 415, + 526, + 426 + ], + "lines": [ + { + "bbox": [ + 469, + 415, + 526, + 426 + ], + "spans": [ + { + "bbox": [ + 469, + 415, + 526, + 426 + ], + "type": "text", + "content": "Alpha blending" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "table", + "bbox": [ + 58, + 470, + 294, + 515 + ], + "blocks": [ + { + "bbox": [ + 58, + 470, + 294, + 515 + ], + "lines": [ + { + "bbox": [ + 58, + 470, + 294, + 515 + ], + "spans": [ + { + "bbox": [ + 58, + 470, + 294, + 515 + ], + "type": "table", + "html": "
3DGS-CUDA3DGS-OpenGLStopThePop1Ours-OpenGL (SPP)
24816
T100065.08100.2875.9116.2318.9524.7137.5161.25
RTX30908.1432.159.483.254.186.4215.3118.48
RTX40905.6020.70-1.852.052.866.718.00
", + "image_path": "a0c23bab304ec5e573d5d949fd0212333d0868b51f8b099e43e0d2d38b08c893.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_body" + } + ], + "index": 26 + }, + { + "type": "table", + "bbox": [ + 58, + 552, + 294, + 586 + ], + "blocks": [ + { + "bbox": [ + 55, + 519, + 295, + 552 + ], + "lines": [ + { + "bbox": [ + 55, + 519, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 295, + 552 + ], + "type": "text", + "content": "Table 2. Forward pass time - in milliseconds (ms) for different techniques on MipNerf360-trained scenes on three different GPUs. The values are averaged over all scenes and all views." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 552, + 294, + 586 + ], + "lines": [ + { + "bbox": [ + 58, + 552, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 58, + 552, + 294, + 586 + ], + "type": "table", + "html": "
3DGSStopThePop1Ours (SPP)
8128
23.0625.7718.1033.19478.93
", + "image_path": "396e3193c37f4a595c0a8d093d7371084a8edf72f6c7c49d638b9f82a5e46f95.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "table_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 55, + 590, + 295, + 624 + ], + "lines": [ + { + "bbox": [ + 55, + 590, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 295, + 624 + ], + "type": "text", + "content": "Table 3. Backward pass time - in ms of the CUDA code for StochasticSplats (as we vary SPP), 3DGS [18], StopThePop [37], averaged over test views of the MipNerf360 scenes on RTX 3090." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": "ing number of samples per pixel, exercising our method's ability to trade off performance and quality. With a single sample, our method achieves " + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": " lower latency compared to alpha-blended 3DGS, depending on the hardware; see Table 2. For higher sample counts, we employ supersampling: render at a higher resolution and then downsample to the target display resolution." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 472, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 555, + 568 + ], + "type": "text", + "content": "Our method is unique at being fast, portable and pop-free at the same time. The portability of our method is only comparable with the 3DGS-OpenGL variant, that has popping artifacts and is relatively slow. StopThePop cannot be easily ported to standard graphics frameworks, and 3DGS-CUDA slows down significantly when ported to OpenGL. This is because the CUDA radix sort routines utilize platform specific optimizations [34]." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 570, + 555, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 618 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 618 + ], + "type": "text", + "content": "Backward pass timings - Tab. 3. We further report the computation time for a single backward pass, this time with the CUDA implementation for all methods, as training typically is done on CUDA devices." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 624, + 542, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 542, + 637 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 542, + 637 + ], + "type": "text", + "content": "4.3. Lowering resolution does not enhance speed" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "content": "Our method inherently offers a performance vs. quality trade-off. Attempting to naively integrate such a trade-off into 3DGS, such as reducing rendering resolution, does not work well. Simply lowering the rendering resolution does not speed up rasterization – the same number of Gaussians must be processed, and the view frustum for each pixel in" + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 745, + 317, + 754 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 745, + 317, + 754 + ], + "spans": [ + { + "bbox": [ + 293, + 745, + 317, + 754 + ], + "type": "text", + "content": "26332" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 60, + 287, + 222 + ], + "blocks": [ + { + "bbox": [ + 67, + 60, + 287, + 222 + ], + "lines": [ + { + "bbox": [ + 67, + 60, + 287, + 222 + ], + "spans": [ + { + "bbox": [ + 67, + 60, + 287, + 222 + ], + "type": "image", + "image_path": "1792147d97dde17498b2f881fab67930bcbdb60ea35becefa6f5cb75c8aa60bf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 224, + 298, + 269 + ], + "lines": [ + { + "bbox": [ + 55, + 224, + 298, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 298, + 269 + ], + "type": "text", + "content": "Figure 8. Timings at lower resolution - Render timings (milliseconds) at different downsampled resolutions for our method (OpenGL) and alpha blending (CUDA) with different tile sizes on an NVIDIA RTX 3090." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 63, + 270, + 174, + 366 + ], + "blocks": [ + { + "bbox": [ + 63, + 270, + 174, + 366 + ], + "lines": [ + { + "bbox": [ + 63, + 270, + 174, + 366 + ], + "spans": [ + { + "bbox": [ + 63, + 270, + 174, + 366 + ], + "type": "image", + "image_path": "f0f2b084fc8229ecb843ceb1f5cd796f67111e6afcd5bea700918d2a2b9fc468.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 369, + 296, + 414 + ], + "lines": [ + { + "bbox": [ + 55, + 369, + 296, + 414 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 296, + 414 + ], + "type": "text", + "content": "Figure 9. Temporal anti-aliasing - StochasticSplats at 1 SPP, along with FLIP error, without (left) and with (right) TAA after moving along a trajectory for 70 frames. TAA introduces a negligible latency increase. For videos, please refer to our website." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 175, + 270, + 286, + 366 + ], + "blocks": [ + { + "bbox": [ + 175, + 270, + 286, + 366 + ], + "lines": [ + { + "bbox": [ + 175, + 270, + 286, + 366 + ], + "spans": [ + { + "bbox": [ + 175, + 270, + 286, + 366 + ], + "type": "image", + "image_path": "d5361622742efbae84724c3c7feaecf56f7223a98e662d8a39ada39bc54b496b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 422, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 506 + ], + "type": "text", + "content": "creases. In short, more Gaussians contribute to each pixel, increasing the computation time per pixel. Figure 8 reports rendering times for our method (OpenGL) and alpha blending (AB-CUDA) with various tile sizes. While AB render time decreases to some degree, it starts to increase after a point, as the Gaussian count per-tile increases. No configuration can reach the rendering speed of our method (1 SPP)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 514, + 263, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 263, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 263, + 528 + ], + "type": "text", + "content": "4.4. Making the most out of a single sample" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 533, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 556 + ], + "type": "text", + "content": "We now practically demonstrate that even a single-sample rendering strategy can be useful." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": "Temporal anti-aliasing - Fig. 9. In interactive applications, multiple frames are rendered in rapid succession. A common strategy to further reduce variance is to average pixel values across frames by reusing information from previous renderings [40]. We employ temporal anti-aliasing (TAA) to not only mitigate aliasing artifacts, but also accumulate samples over time, achieving an effectively higher sample count at minimal cost. Frame-to-frame consistency is maintained by warping previously rendered frames according to per-pixel depth whenever the camera moves. As shown in Figure 9, this greatly reduces rendering noise, while maintaining low latency. We provide implementation details in the appendix, and our supplemental" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 326, + 70, + 434, + 150 + ], + "blocks": [ + { + "bbox": [ + 326, + 70, + 434, + 150 + ], + "lines": [ + { + "bbox": [ + 326, + 70, + 434, + 150 + ], + "spans": [ + { + "bbox": [ + 326, + 70, + 434, + 150 + ], + "type": "image", + "image_path": "c7e624f212b59042de5480de38f0c349a6f82f1718c4ea6d6a88cdee73b1c168.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 152, + 416, + 162 + ], + "lines": [ + { + "bbox": [ + 345, + 152, + 416, + 162 + ], + "spans": [ + { + "bbox": [ + 345, + 152, + 416, + 162 + ], + "type": "text", + "content": "(a) Green apple (AB)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 434, + 71, + 543, + 150 + ], + "blocks": [ + { + "bbox": [ + 434, + 71, + 543, + 150 + ], + "lines": [ + { + "bbox": [ + 434, + 71, + 543, + 150 + ], + "spans": [ + { + "bbox": [ + 434, + 71, + 543, + 150 + ], + "type": "image", + "image_path": "5c3439c8a5b405cb0b083a2d132341d29b3c771eb2574080751ec52527110482.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 452, + 152, + 525, + 162 + ], + "lines": [ + { + "bbox": [ + 452, + 152, + 525, + 162 + ], + "spans": [ + { + "bbox": [ + 452, + 152, + 525, + 162 + ], + "type": "text", + "content": "(b) Green apple (ours)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 325, + 163, + 433, + 243 + ], + "blocks": [ + { + "bbox": [ + 325, + 163, + 433, + 243 + ], + "lines": [ + { + "bbox": [ + 325, + 163, + 433, + 243 + ], + "spans": [ + { + "bbox": [ + 325, + 163, + 433, + 243 + ], + "type": "image", + "image_path": "256f2849db8bfcfc9da827861d9aea1c5060e082ec79c139b290477426cc15e2.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 244, + 413, + 254 + ], + "lines": [ + { + "bbox": [ + 345, + 244, + 413, + 254 + ], + "spans": [ + { + "bbox": [ + 345, + 244, + 413, + 254 + ], + "type": "text", + "content": "(c) Chopsticks (AB)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 434, + 163, + 542, + 243 + ], + "blocks": [ + { + "bbox": [ + 434, + 163, + 542, + 243 + ], + "lines": [ + { + "bbox": [ + 434, + 163, + 542, + 243 + ], + "spans": [ + { + "bbox": [ + 434, + 163, + 542, + 243 + ], + "type": "image", + "image_path": "abed106002d109459632ac80b5f89dd4c7af83d0066d103b404243e163ae901f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 453, + 244, + 523, + 254 + ], + "lines": [ + { + "bbox": [ + 453, + 244, + 523, + 254 + ], + "spans": [ + { + "bbox": [ + 453, + 244, + 523, + 254 + ], + "type": "text", + "content": "(d) Chopsticks (ours)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 316, + 291, + 553, + 327 + ], + "blocks": [ + { + "bbox": [ + 313, + 258, + 555, + 291 + ], + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 291 + ], + "type": "text", + "content": "Figure 10. Open vocabulary localization - Qualitative comparison of open-vocabulary 3D object localization on the LERF dataset using alpha blending vs. StochasticSplats at 1 SPP." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 291, + 553, + 327 + ], + "lines": [ + { + "bbox": [ + 316, + 291, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 291, + 553, + 327 + ], + "type": "table", + "html": "
TeatimeRamenWaldo-KitchenFigurinesMean
AB88.173.295.580.484.3
Ours88.173.286.480.482.0
", + "image_path": "900d4e6d5d2d7c0d80ecd3ae6dd975866f957254b792cc880081b40211e0702f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 331, + 555, + 354 + ], + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 354 + ], + "type": "text", + "content": "Table 4. Open vocabulary localization - accuracy on the LERF dataset using alpha blending (AB) and StochasticSplats at 1 SPP." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 359, + 522, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 522, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 522, + 371 + ], + "type": "text", + "content": "videos demonstrate TAA applied to 1-SPP frames." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 373, + 555, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 482 + ], + "type": "text", + "content": "Semantic localization - Fig. 10 and Tab. 4. We further show that while the single-sample render has low PSNR, it is still useful. Specifically, we use LangSplat [36] and replace their alpha blending renderer with our method, and follow their pipeline to localize objects. As shown in Figure 10, while our renders are noisy, it still properly renders semantics such that the localizer can correctly identify its location. There is a moderate reduction in accuracy (Tab. 4), but we achieve " + }, + { + "bbox": [ + 313, + 373, + 555, + 482 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 313, + 373, + 555, + 482 + ], + "type": "text", + "content": " faster rendering speed." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 486, + 388, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 388, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 388, + 498 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 506, + 555, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 602 + ], + "type": "text", + "content": "We proposed a sorting-free stochastic rendering method for 3D Gaussian splatting, that is, technically, compatible with any semi-transparent representation. We derive both the forward (rendering) and backward (gradient) passes for stochastic rendering, allowing a quality vs. cost trade off, and renders without popping artifacts. Our OpenGL implementation renders " + }, + { + "bbox": [ + 313, + 506, + 555, + 602 + ], + "type": "inline_equation", + "content": "2 \\times -4 \\times" + }, + { + "bbox": [ + 313, + 506, + 555, + 602 + ], + "type": "text", + "content": " faster with reasonable quality, demonstrating high portability." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 603, + 556, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 603, + 556, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 603, + 556, + 704 + ], + "type": "text", + "content": "Limitations and future works. While we obtain reasonable quality when rendering with low sample counts, our results are noisy. Future work could focus on reducing the variance of both forward and backward estimators, for example using importance sampling, stratified sampling [21] or control variates. Moreover, quality could further be improved by leveraging lightweight (neural) denoising algorithms, similar to how diffusion models are used in conjunction with rough 3D renderings [39, 48]." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 745, + 318, + 754 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 745, + 318, + 754 + ], + "spans": [ + { + "bbox": [ + 293, + 745, + 318, + 754 + ], + "type": "text", + "content": "26333" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Pontus Andersson, Jim Nilsson, Tomas Akenine-Moller, Magnus Oskarsson, Kalle Åström, and Mark D. Fairchild. FLIP: A Difference Evaluator for Alternating Images. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 3(2):15:1-15:23, 2020. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 191 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 191 + ], + "type": "text", + "content": "[2] Dejan Azinovic, Tzu-Mao Li, Anton Kaplanyan, and Matthias Nießner. Inverse path tracing for joint material and lighting estimation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2442-2451, 2019. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 191, + 295, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 295, + 235 + ], + "type": "text", + "content": "[3] Milena T Bagdasarian, Paul Knoll, Yi-Hsin Li, Florian Barthel, Anna Hilsmann, Peter Eisert, and Wieland Morgenstern. 3dgs. zip: A survey on 3d gaussian splatting compression methods. arXiv preprint arXiv:2407.09510, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 236, + 295, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 295, + 289 + ], + "type": "text", + "content": "[4] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 291, + 295, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 295, + 346 + ], + "type": "text", + "content": "[5] Jorge Condor, Sebastien Speierer, Lukas Bode, Aljaz Bozic, Simon Green, Piotr Didyk, and Adrian Jarabo. Don't splat your gaussians: Volumetric ray-traced primitives for modeling and rendering scattering and emissive media. ACM Trans. Graph., 44(1), 2025. 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 346, + 295, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 295, + 378 + ], + "type": "text", + "content": "[6] Franklin C. Crow. The aliasing problem in computer-generated shaded images. Commun. ACM, 20(11):799-805, 1977. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "type": "text", + "content": "[7] Thomas Deliot, Eric Heitz, and Laurent Belcour. Transforming a non-differentiable rasterizer into a differentiable one with stochastic gradient estimation. Proc. Symposium on Interactive 3D Graphics and Games (I3D), 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 424, + 295, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 424, + 295, + 467 + ], + "spans": [ + { + "bbox": [ + 62, + 424, + 295, + 467 + ], + "type": "text", + "content": "[8] Eric Enderton, Erik Sintorn, Peter Shirley, and David Luebke. Stochastic transparency. In Proc. Symposium on Interactive 3D Graphics and Games (I3D), pages 157-164, 2010. 2, 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 469, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 469, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 469, + 295, + 491 + ], + "type": "text", + "content": "[9] Cass Everitt. Interactive order-independent transparency. White paper, NVIDIA, 2(6):7, 2001. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 491, + 294, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 491, + 294, + 513 + ], + "spans": [ + { + "bbox": [ + 57, + 491, + 294, + 513 + ], + "type": "text", + "content": "[10] James D Foley. Computer graphics: principles and practice. Addison-Wesley Professional, 1996. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 514, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 294, + 546 + ], + "type": "text", + "content": "[11] Michael B. Giles. Monte carlo evaluation of sensitives in computational finance. Technical report, Oxford-Man Institute of Quantitative Finance, 2007. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 547, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 294, + 602 + ], + "type": "text", + "content": "[12] Ioannis Gkioulekas, Anat Levin, and Todd Zickler. An evaluation of computational imaging techniques for heterogeneous inverse scattering. In European Conference on Computer Vision (ECCV), pages 685-701. Springer International Publishing, 2016. 3, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 602, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 635 + ], + "type": "text", + "content": "[13] Khronos Group. Early Fragment Test. https://www.khronos.org/opengl/wiki/Early_Fragment_Test, 2025.5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 636, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 679 + ], + "type": "text", + "content": "[14] Hao Gui, Lin Hu, Rui Chen, Mingxiao Huang, Yuxin Yin, Jin Yang, and Yong Wu. Balanced 3DGS: Gaussian-wise parallelism rendering with fine-grained tiling. arXiv preprint arXiv:2412.17378, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[15] Florian Hahlbohm, Fabian Friederichs, Tim Weyrich, Linus Franke, Moritz Kappel, Susana Castillo, Marc Stamminger, Martin Eisemann, and Marcus Magnor. Effi" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "cient perspective-correct 3D Gaussian splatting using hybrid transparency. arXiv preprint arXiv:2410.08129, 2024. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[16] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, Light, and Material Decomposition from Images using Monte Carlo Rendering and Denoising. Advances in Neural Information Processing Systems (NeurIPS), 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "text", + "content": "[17] Qiqi Hou, Randall Rauwendaal, Zifeng Li, Hoang Le, Farzad Farhadzadeh, Fatih Porikli, Alexei Bourd, and Amir Said. Sort-free gaussian splatting via weighted sum rendering. arXiv preprint arXiv:2410.18931, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "type": "text", + "content": "[18] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D gaussian splatting for real-time radiance field rendering. ACM Trans. Graph. (Proc. SIGGRAPH), 42(4), 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 232, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 553, + 285 + ], + "type": "text", + "content": "[19] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Leref: Language embedded radiance fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 19729-19739, 2023. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 286, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 353 + ], + "type": "text", + "content": "[20] Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Weiwei Sun, Yang-Che Tseng, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. 3D gaussian splatting as markov chain monte carlo. In Advances in Neural Information Processing Systems (NeurIPS), 2024. Spotlight Presentation. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 354, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 388 + ], + "type": "text", + "content": "[21] Samuli Laine and Tero Karras. Stratified sampling for stochastic transparency. In Comp. Graph. Forum (Proc. EGSR), pages 1197-1204, 2011. 2, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "type": "text", + "content": "[22] Tzu-Mao Li, Miika Aittala, Frédo Durand, and Jaakko Lehtinen. Differentiable monte carlo ray tracing through edge sampling. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 37 (6):222:1-222:11, 2018. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 433, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 477 + ], + "type": "text", + "content": "[23] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft Rasterizer: A Differentiable Renderer for Image-Based 3D Reasoning. In IEEE International Conference on Computer Vision (ICCV), pages 7708-7717, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "type": "text", + "content": "[24] Alexander Mai, Dor Verbin, Falko Kuester, and Sara Fridovich-Keil. Neural microfacet fields for inverse rendering. In IEEE International Conference on Computer Vision (ICCV), 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 523, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 578 + ], + "type": "text", + "content": "[25] Alexander Mai, Peter Hedman, George Kopanas, Dor Verbin, David Futschik, Qiangeng Xu, Falko Kuester, Jonathan T Barron, and Yinda Zhang. EVER: Exact volumetric ellipsoid rendering for real-time view synthesis. arXiv preprint arXiv:2410.01804, 2024. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 579, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 635 + ], + "type": "text", + "content": "[26] Saswat Subhajyoti Mallick, Rahul Goel, Bernhard Kerbl, Markus Steinberger, Francisco Vicente Carrasco, and Fernando De La Torre. Taming 3dgs: High-quality radiance fields with limited resources. In SIGGRAPH Asia Conference Papers, pages 1-11, 2024. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 635, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 669 + ], + "type": "text", + "content": "[27] Morgan McGuire and Louis Bavoil. Weighted blended order-independent transparency. Journal of Computer Graphics Techniques (JCGT), 2(2):122-141, 2013. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[28] Guennady A. Mikhailov. On the calculation of nuclear reactor disturbances by the monte carlo method. USSR Computational Mathematics and Mathematical Physics, 6(2):268-273, 1966. 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 746, + 318, + 754 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 746, + 318, + 754 + ], + "spans": [ + { + "bbox": [ + 293, + 746, + 318, + 754 + ], + "type": "text", + "content": "26334" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[29] Ben Mildenhall, Pratul Srinivasan, Matthew Tancik, Jonathan Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In European Conference on Computer Vision (ECCV), pages 405-421. Springer, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "type": "text", + "content": "[30] Cedrick Münstermann, Stefan Krumpen, Reinhard Klein, and Christoph Peters. Moment-based order-independent transparency. Proc. Symposium on Interactive 3D Graphics and Games (I3D), 1(1):1-20, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 294, + 205 + ], + "type": "text", + "content": "[31] Vasiliki Myrodia. Psychophysical studies on Monte Carlo rendering-noise visual perception. PhD thesis, Université de Lille, 2021. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 205, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 294, + 249 + ], + "type": "text", + "content": "[32] Merlin Nimier-David, Delio Vicini, Tizian Zeltner, and Wenzel Jakob. Mitsuba 2: A retargetable forward and inverse renderer. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 38 (6), 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "text", + "content": "[33] Jan Novák, Andrew Selle, and Wojciech Jarosz. Residual ratio tracking for estimating attenuation in participating media. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 33(6):179:1-179:11, 2014. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 293, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 294, + 326 + ], + "type": "text", + "content": "[34] Lukas Polok, Viorela Ila, and Pavel Smrz. Fast radix sort for sparse linear algebra on gpu. In *SpringSim (HPS)*, page 11, 2014. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 358 + ], + "type": "text", + "content": "[35] Thomas Porter and Tom Duff. Compositing digital images. In Computer Graphics (Proc. SIGGRAPH), page 253-259. ACM, 1984. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 360, + 294, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 404 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 404 + ], + "type": "text", + "content": "[36] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 20051-20060, 2024. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 404, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 459 + ], + "type": "text", + "content": "[37] Lukas Radl, Michael Steiner, Mathias Parger, Alexander Weinrauch, Bernhard Kerbl, and Markus Steinberger. Stophepop: Sorted gaussian splatting for view-consistent real-time rendering. ACM Trans. Graph. (Proc. SIGGRAPH), 43(4):1-17, 2024. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 459, + 294, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 459, + 294, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 459, + 294, + 492 + ], + "type": "text", + "content": "[38] Ravi Ramamoorthi and Pat Hanrahan. An efficient representation for irradiance environment maps. In ACM Trans. Graph. (Proc. SIGGRAPH), 2001. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 492, + 294, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 294, + 558 + ], + "type": "text", + "content": "[39] Xuanchi Ren, Tianchang Shen, Jiahui Huang, Huan Ling, Yifan Lu, Merlin Nimier-David, Thomas Müller, Alexander Keller, Sanja Fidler, and Jun Gao. GEN3C: 3d-informed world-consistent video generation with precise camera control. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 559, + 294, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 603 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 603 + ], + "type": "text", + "content": "[40] Daniel Scherzer, Lei Yang, Oliver Mattausch, Diego Nehab, Pedro V. Sander, Michael Wimmer, and Elmar Eisemann. Temporal Coherence Methods in Real-Time Rendering. Comp. Graph. Forum, 31(8):2378-2408, 2012. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 603, + 294, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 625 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 625 + ], + "type": "text", + "content": "[41] Andrea Tagliasacchi and Ben Mildenhall. Volume rendering digest (for nef). arXiv preprint arXiv:2209.02417, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 625, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 658 + ], + "type": "text", + "content": "[42] Anthony J. Thibault. Splatapult. https://github.com/hyperlogic/splatapult, 2023. Accessed January 2, 2025. 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[43] JM Tregan, S Blanco, J Dauchet, M Hafi, R Fournier, L Ibarrart, P Lapeyre, and N Villefranque. Convergence issues in derivatives of Monte Carlo null-collision integral formulations: a solution. arXiv preprint arXiv:1903.06508, 2019. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 690 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[44] Grigoris Tsopouridis, Andreas A Vasilakis, and Ioannis Fudos. Deep and fast approximate order independent transparency. In Comp. Graph. Forum, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "text", + "content": "[45] Dor Verbin, Ben Mildenhall, Peter Hedman, Jonathan T. Barron, Todd Zickler, and Pratul P. Srinivasan. Eclipse: Disambiguating illumination and materials using unintended shadows. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "text", + "content": "[46] Delio Vicini, Sébastien Speierer, and Wenzel Jakob. Path replay backpropagation: Differentiating light paths using constant memory and linear time. ACM Trans. Graph. (Proc. SIGGRAPH), 40(4):108:1-108:14, 2021. 3, 4, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 207, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 553, + 240 + ], + "type": "text", + "content": "[47] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Proc., 13(4), 2004. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 241, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 297 + ], + "type": "text", + "content": "[48] Jay Zhangjie Wu, Yuxuan Zhang, Haithem Turki, Xuanchi Ren, Jun Gao, Mike Zheng Shou, Sanja Fidler, Zan Gojcic, and Huan Ling. Difix3D+: Improving 3d reconstructions with single-step diffusion models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 297, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 319 + ], + "type": "text", + "content": "[49] Chris Wyman. Stochastic layered alpha blending. In ACM SIGGRAPH 2016 Talks, pages 1-2. 2016. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 320, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 354 + ], + "type": "text", + "content": "[50] Jason C Yang, Justin Hensley, Holger Grün, and Nicolas Thibierz. Real-time concurrent linked list construction on thegpu. In Comp. Graph. Forum (Proc. EGSR), 2010. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 354, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 408 + ], + "type": "text", + "content": "[51] Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jeffrey Hu, Matthew Tancik, and Angjoo Kanazawa. gsplat: An open-source library for Gaussian splatting. arXiv preprint arXiv:2409.06765, 2024.3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 409, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 409, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 409, + 553, + 443 + ], + "type": "text", + "content": "[52] Ekrem Fatih Yilmazer, Delio Vicini, and Wenzel Jakob. Solving inverse pde problems using monte carlo estimators. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 43, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 443, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 443, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 443, + 553, + 487 + ], + "type": "text", + "content": "[53] Tizian Zeltner, Sébastien Speierer, Iliyan Georgiev, and Wenzel Jakob. Monte carlo estimators for differential light transport. Transactions on Graphics (Proceedings of SIGGRAPH), 40(4), 2021. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 488, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 488, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 488, + 553, + 533 + ], + "type": "text", + "content": "[54] Cheng Zhang, Lifan Wu, Changxi Zheng, Ioannis Gkioulekas, Ravi Ramamoorthi, and Shuang Zhao. A differential theory of radiative transfer. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 38(6), 2019. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 533, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 553, + 578 + ], + "type": "text", + "content": "[55] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 578, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 553, + 622 + ], + "type": "text", + "content": "[56] Shuang Zhao, Lifan Wu, Frédo Durand, and Ravi Ramamoorthi. Downsampling scattering parameters for rendering anisotropic media. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 35(6), 2016. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 622, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 622, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 622, + 553, + 656 + ], + "type": "text", + "content": "[57] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Surface splatting. In ACM Trans. Graph. (Proc. SIGGRAPH), 2001. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 657, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 657, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 657, + 553, + 690 + ], + "type": "text", + "content": "[58] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In Proceedings Visualization (VIS), pages 29-538. IEEE, 2001. 2, 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 745, + 317, + 754 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 745, + 317, + 754 + ], + "spans": [ + { + "bbox": [ + 293, + 745, + 317, + 754 + ], + "type": "text", + "content": "26335" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/StolenLoRA_ Exploring LoRA Extraction Attacks via Synthetic Data/f4ddd460-9cf1-429e-89c2-3d861bf9c661_content_list.json b/2025/StolenLoRA_ Exploring LoRA Extraction Attacks via Synthetic Data/f4ddd460-9cf1-429e-89c2-3d861bf9c661_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..340c53f2f0d76798816dcd2d072cf227d6807629 --- /dev/null +++ b/2025/StolenLoRA_ Exploring LoRA Extraction Attacks via Synthetic Data/f4ddd460-9cf1-429e-89c2-3d861bf9c661_content_list.json @@ -0,0 +1,1497 @@ +[ + { + "type": "text", + "text": "StolenLoRA: Exploring LoRA Extraction Attacks via Synthetic Data", + "text_level": 1, + "bbox": [ + 150, + 130, + 846, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yixu Wang $^{1,2\\dagger}$ , Yan Teng $^{2*}$ , Yingchun Wang $^{2}$ , Xingjun Ma $^{1}$", + "bbox": [ + 261, + 179, + 733, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Fudan University $^{2}$ Shanghai Artificial Intelligence Laboratory", + "bbox": [ + 238, + 204, + 759, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 258, + 326, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Parameter-Efficient Fine-Tuning (PEFT) methods like LoRA have transformed vision model adaptation, enabling the rapid deployment of customized models. However, the compactness of LoRA adaptations introduces new safety concerns, particularly their vulnerability to model extraction attacks. This paper introduces a new focus of model extraction attacks named LoRA extraction that extracts LoRA-adaptive models based on a public pre-trained model. We then propose a novel extraction method called StolenLoRA which trains a substitute model to extract the functionality of a LoRA-adapted model using synthetic data. StolenLoRA leverages a Large Language Model to craft effective prompts for data generation, and it incorporates a Disagreement-based Semi-supervised Learning (DSL) strategy to maximize information gain from limited queries. Our experiments demonstrate the effectiveness of StolenLoRA, achieving up to a $96.60\\%$ attack success rate with only 10k queries, even in cross-backbone scenarios where the attacker and victim models utilize different pretrained backbones. These findings reveal the specific vulnerability of LoRA-adapted models to this type of extraction and underscore the urgent need for robust defense mechanisms tailored to PEFT methods. We also explore a preliminary defense strategy based on diversified LoRA deployments, highlighting its potential to mitigate such attacks.", + "bbox": [ + 89, + 290, + 485, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 695, + 220, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adapting large-scale pre-trained foundation models via Parameter-Efficient Fine-Tuning (PEFT) such as Low-Rank Adaptation (LoRA) [11, 15] has become a popular approach for obtaining high-performance models for diverse downstream tasks. LoRA enables efficient adaptation of large models to a new task by fine-tuning only low-rank matrices added to specific layers, reducing computational cost and memory usage while preserving the pre-trained model's", + "bbox": [ + 89, + 720, + 483, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "core knowledge. Despite the wide adoption of LoRA, the lightweight and compactness of LoRA parameters raise a new safety concern: they might be more vulnerable to model extraction (ME) attacks. An ME attack extracts the functionality [17, 22, 28, 39] of a victim model by training a substitute model based on the outputs of the victim model for a certain number of specially designed queries. While ME attacks have been extensively studied for traditional models [29, 40, 42, 43, 46, 47], the vulnerability of LoRA adaptations to ME attacks remains largely unexplored[25]. This gap is significant, as the public availability of a large number of pre-trained models, coupled with the compactness of LoRA parameters, makes it easier to replicate LoRA-adapted models and compromise intellectual property.", + "bbox": [ + 511, + 258, + 906, + 472 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address this gap, we introduce the concept of LoRA Extraction, a novel direction of ME attacks specifically targeting LoRA-adapted models (i.e., ViTs [4, 12, 20]), based on a publicly available pre-trained model. Unlike traditional ME which focuses on replicating the entire model's functionality, LoRA extraction centers on stealing the efficient adaptations encoded within the compact LoRA parameters. The adversary's objective is to reconstruct the victim's LoRA-adapted model by training their own LoRA-adapted substitute that achieves similar downstream performance using a pre-trained foundation model that is publicly available. This attack can manifest in two scenarios: the identical-backbone scenario, where the attacker uses the same pre-trained ViT model as the victim, and the more challenging cross-backbone scenario, where the attacker uses a different pre-trained ViT.", + "bbox": [ + 511, + 474, + 908, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing traditional ME methods largely depend on sample selection from existing datasets or synthetic data generation. Sample selection methods [28, 29, 42, 46] typically require searching extensive datasets for in-distribution samples, making them computationally expensive and less practical for LoRA extraction. This impracticality arises because 1) inference in LoRA-adapted models is slower due to the large-scale parameters of LoRA-adapted models, and 2) finding appropriate samples is challenging, particularly for domain-specific data fine-tuned with LoRA. Alternatively, synthetic data generation approaches [18, 23, 40, 47] often rely on Generative Adversarial Networks (GANs) [8]", + "bbox": [ + 511, + 719, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "\\*Corresponding authors: < tengyan@pjlab.org.cn, xingjunma@fudan.edu.cn > T Work done during internship at Shanghai Artificial Intelligence Laboratory.", + "bbox": [ + 89, + 848, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "263", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to produce in-distribution data. However, GANs frequently face difficulties in generating high-quality, diverse samples, especially for high-dimensional data. As Zhao et al. [46] noted, creating high-dimensional samples (e.g., $224 \\times 224$ images) for effective ME can require millions of queries and is prone to failure due to GAN mode collapse [1, 6, 21]. These limitations make GAN-based approaches less effective for LoRA extraction, where the tuning often involves high-dimensional image data.", + "bbox": [ + 89, + 90, + 480, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a novel LoRA extraction method named StolenLoRA, which employs synthetic data to train the substitute model. To obtain effective in-distribution synthetic data, we leverage Large Language Models (LLMs) to generate diverse textual descriptions based on the target class names and prompt a pre-trained Stable Diffusion model [34, 36] to generate high-quality images. This allows us to tailor the generated data to the specific downstream task the victim's LoRA model is fine-tuned for. Moreover, StolenLoRA introduces a distinctive strategy to improve attack efficiency, i.e., the Disagreement-based Semi-supervised Learning (DSL). DSL uses class information from the data synthesis process as pseudo-labels for part of the generated dataset and focuses on queries where the substitute model disagrees with these pseudo-labels. DSL effectively targets areas of uncertainty, guiding the substitute model's learning process and refining its alignment with the victim's behavior. To further enhance DSL, we iteratively refine the pseudo-labels based on evolving predictions from the substitute model, improving the quality of synthetic data labels and the effectiveness of the extraction.", + "bbox": [ + 91, + 227, + 486, + 556 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 107, + 559, + 410, + 573 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We explore the vulnerability of LoRA and introduce a novel focus of ME attacks called LoRA Extraction which aims to extract the functionality of LoRA-adapted models based on a publicly available pre-trained model.", + "- We present StolenLoRA, a novel LoRA extraction method that leverages LLM-driven Stable Diffusion to generate high-quality synthetic data, circumventing the need to search large-scale datasets or rely on less dependable GAN-based generation.", + "- Through comprehensive experiments on five widely used datasets, we demonstrate the effectiveness of Stolen-LoRA in both identical- and cross-backbone scenarios, highlighting critical vulnerabilities in LoRA-adapted models. We also explore a preliminary defense mechanism based on diversified LoRA deployments." + ], + "bbox": [ + 91, + 574, + 482, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 814, + 232, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Low-Rank Adaptation (LoRA). LoRA [15] is a type of popular Parameter-Efficient Fine-Tuning (PEFT) method [10, 14] for adapting large pre-trained models for downstream tasks. LoRA [15] introduces small rank-", + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "decomposition matrices as trainable parameters into the pre-trained model. The number of trainable parameters is significantly reduced by freezing the original model weights and updating only these lightweight LoRA modules. LoRA inspires numerous variants and extensions, such as Mixture-of-LoRA Experts (X-LoRA)[3], Low-Rank Hadamard Product (LoHa)[16], and Low-Rank Kronecker Product (LoKr)[44], further enhancing its efficiency and flexibility for various applications. However, the compact nature of these LoRA updates raises concerns about their vulnerability to model extraction attacks. This paper investigates the potential for adversaries to extract the efficient adaptations encoded within these compact parameters, specifically focusing on the standard LoRA method [15].", + "bbox": [ + 511, + 90, + 903, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Model Extraction Attacks. Model Extraction (ME) attacks [18, 22, 23, 28, 29, 40, 42, 46, 47] aim to construct a substitute model that mimics the functionality of a victim model by querying its API. Existing ME methods can be broadly categorized based on their query generation strategies: Data-Driven ME leverage existing datasets [28, 29, 42, 46] to select queries, often employing active learning[32] or other search strategies[28] to identify informative samples. However, this type can incur computational overhead as it necessitates traversing the large-scale dataset to find suitable in-distribution samples. Data-Free ME [18, 23, 40, 46, 47] utilize generative models (e.g., GANs[8]) to synthesize data for querying the victim model. While promising, generating effective queries with high-dimensional inputs can require tens of millions of queries and is prone to failure due to GAN mode collapse [1, 6, 21].", + "bbox": [ + 511, + 335, + 906, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning from Synthetic Data. The use of synthetic data for training models gains significant traction due to its potential to address data scarcity. While traditional synthetic data often lacked the complexity and representativeness of real data, recent advancements in generative models, particularly diffusion models like Stable Diffusion [34, 36], enable the synthesis of high-quality and diverse images. However, scaling the quantity of synthetic data does not necessarily translate to improved performance for training supervised image classifiers. As shown by Fan et al.[7], performance with synthetic data can lag significantly behind training with real data. This disparity stems from the limitations of current text-to-image models in accurately generating diverse representations of certain concepts at scale[30]. Efforts to mitigate this gap, such as the distribution-matching framework[45], focus on aligning the distributions of synthetic and real data. However, leveraging such high-quality synthetic data for model extraction attacks remains an unexplored area.", + "bbox": [ + 511, + 613, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "264", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ee8f6f3384583ba07b2dbe2a5d69d0ed57a8c92aea61e38c2c949168b526c2d6.jpg", + "image_caption": [ + "Figure 1. Overview of StolenLoRA. It collects target victim model information (dashed red box), synthesizes training data guided by an LLM-driven prompting (dashed blue box), and then efficiently queries the victim API to train a substitute model (dashed green box)." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 898, + 318 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 361, + 225, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 89, + 386, + 294, + 401 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LoRA Extraction. Given a victim model $F \\colon [0,1]^d \\mapsto \\mathbb{R}^N$ , which is a pre-trained model $F_{base}$ (e.g., ViTs) adapted using LoRA. The parameters of $F^1$ can be represented as $\\theta = \\{\\theta_{base}, \\Delta \\theta\\}$ where $\\theta_{base}$ are the parameters of $F_{base}$ and $\\Delta \\theta$ represents the LoRA updates, which are typically low-rank matrices. The objective of a LoRA extraction attack is to train a substitute model $F'$ with parameters $\\theta' = \\{\\theta_{base}', \\Delta \\theta'\\}$ that mimics the functionality of $F$ . This is formulated as the following optimization problem:", + "bbox": [ + 89, + 407, + 483, + 544 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\arg \\min _ {\\Delta \\theta^ {\\prime}} \\mathbb {E} _ {\\boldsymbol {x}} \\mathcal {D} (F (\\boldsymbol {x}), F ^ {\\prime} (\\boldsymbol {x})), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 551, + 480, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\pmb{x} \\in \\mathbb{R}^d$ represents an input sample from the victim model task distribution, $\\mathcal{D}(\\cdot, \\cdot)$ represents a measure of functional similarity between the two models.", + "bbox": [ + 89, + 583, + 483, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Depending on whether the attacker has knowledge of $F_{base}$ , we distinguish between two scenarios: identical backbone and cross-backbone LoRA extractions. The substitute model $F'$ can be defined as follows:", + "bbox": [ + 89, + 628, + 483, + 686 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nF ^ {\\prime} (\\boldsymbol {x}) = \\left\\{ \\begin{array}{l l} F _ {\\text {b a s e}} (\\boldsymbol {x}) + \\Delta F ^ {\\prime} (\\boldsymbol {x}), & \\text {i d e n t i c a l - b a c k b o n e} \\\\ G _ {\\text {b a s e}} (\\boldsymbol {x}) + \\Delta F ^ {\\prime} (\\boldsymbol {x}), & \\text {c r o s s - b a c k b o n e} \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 694, + 480, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the former uses the same $F_{base}$ , the latter uses different base models $G_{base}$ , and the $+$ symbol denotes the combination of the base model and the LoRA updates.", + "bbox": [ + 89, + 747, + 482, + 791 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Distinction from Traditional Extraction. Traditional model extraction trains a substitute from scratch, while LoRA extraction leverages a pre-trained base model, inheriting its knowledge and biases. Consequently, we hypothesize that a perfectly extracted LoRA substitute might", + "bbox": [ + 89, + 792, + 483, + 868 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "achieve functional equivalence on In-Distribution (ID) data but not on Out-of-Distribution (OOD) data, particularly in the cross-backbone setting, and we experimentally verify this in the experimental section. This motivates the use of ID data for effective LoRA extraction.", + "bbox": [ + 511, + 363, + 906, + 438 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. StolenLoRA", + "text_level": 1, + "bbox": [ + 511, + 446, + 643, + 459 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. An overview of our proposed LLM-driven generative attack for extracting LoRA, StolenLoRA, is illustrated in Fig. 1. It comprises two key stages: 1) Data Synthesis, where an LLM guides the generation of synthetic data tailored to the victim's task, and 2) Efficient Querying and Training, which strategically selects queries and refines labels to maximize information gain from the victim model.", + "bbox": [ + 511, + 468, + 905, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1. Data Synthesis", + "text_level": 1, + "bbox": [ + 511, + 580, + 661, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The above discussion motivates us to use ID data for extraction. However, acquiring such datasets in real-world scenarios is often challenging because the victim model's training data is proprietary. While the attacker might not have access to the original training data, they often have access to the model's deployment context, which usually includes a description of the model's functionality and the names of the target classes. This information can be leveraged to semantically expand the search space for potential training data. Inspired by this, we propose to utilize LLM-driven Stable Diffusion to generate synthetic data based on the class names, effectively bridging the gap between semantic information and visual representations.", + "bbox": [ + 511, + 598, + 906, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LLM-Driven Prompting. Given a set of target class names $\\mathbb{C} = \\{c_1, c_2, \\dots, c_n\\}$ , we employ an LLM to generate detailed prompts for image synthesis. We structure the prompts using a predefined template $T$ comprising key visual elements: $T = \\text{\"[Subject, Background, Angle/ Pose, Lighting, Style]\"}$ . For each class $c_i$ , the LLM generates $m$ variations of the", + "bbox": [ + 511, + 795, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1We omit the parameter for simplicity when there is no ambiguity, e.g., using $F(\\pmb{x})$ instead of $F(\\pmb{x};\\theta)$ .", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "265", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "prompt, denoted as $p_{i,j}$ , where $j \\in \\{1, \\dots, m\\}$ . This diversification ensures a richer representation of the target class. The prompt generation process is formalized as follows:", + "bbox": [ + 89, + 90, + 482, + 137 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {i, j} = \\operatorname {L L M} \\left(c _ {i}, T, \\omega_ {j}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 147, + 482, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\omega_{j}$ represents a random seed or instruction to control the variation generated by the LLM.", + "bbox": [ + 89, + 172, + 482, + 203 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image Synthesis. We employ a public pre-trained Stable Diffusion model, denoted as $\\mathrm{SD}(\\cdot)$ , to generate images corresponding to the crafted prompts. For each prompt $p_{i,j}$ we synthesize one image, resulting in a set of synthetic images $\\mathbb{X}_i = \\{\\pmb{x}_{i,j}\\}$ , where $\\pmb{x}_{i,j} = \\mathrm{SD}(p_{i,j})$ . The complete synthetic dataset is then $\\mathbb{X} = \\bigcup_{i}\\mathbb{X}_{i}$ .", + "bbox": [ + 89, + 203, + 483, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. Efficient Querying and Training", + "text_level": 1, + "bbox": [ + 89, + 301, + 361, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This stage focuses on strategically querying the victim model to refine the substitute model's LoRA parameters while minimizing the number of queries. We propose two attack strategies: random learning and DSL.", + "bbox": [ + 89, + 319, + 483, + 380 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Random Learning (StolenLoRA-Rand). This straightforward approach directly uses the synthesized dataset $\\mathbb{X}$ to query the victim model $F$ . Then, the substitute model's LoRA parameters $\\Delta \\theta^{\\prime}$ are trained by minimizing the following objective function:", + "bbox": [ + 89, + 380, + 483, + 455 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\prime} (\\Delta \\theta^ {\\prime}) = \\mathbb {E} _ {\\boldsymbol {x} \\in \\mathbb {X}} \\mathcal {L} \\left(F ^ {\\prime} (\\boldsymbol {x}), F (\\boldsymbol {x})\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 465, + 482, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}$ denotes a loss function (e.g., cross-entropy loss). This provides a baseline for evaluating the benefits of more sophisticated attack strategies.", + "bbox": [ + 89, + 493, + 482, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Disagreement-based Semi-supervised Learning (StolenLoRA-DSL). While directly using synthesized data for querying and training is feasible, we propose DSL to significantly enhance both the effectiveness and efficiency of the attack. DSL improves query efficiency by selectively querying the victim model based on the substitute model's uncertainty. Furthermore, it iteratively refines both the pseudo-labels of the synthetic data and the labels obtained from querying the victim, mitigating the impact of noisy pseudo-labels and bridging the gap between the substitute and victim model, especially in cross-backbone scenarios.", + "bbox": [ + 89, + 537, + 483, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The process begins by generating an initial synthetic dataset $\\mathbb{X}^0 = \\bigcup_i\\mathbb{X}_i^0$ and assigning pseudo-labels based on their generating prompts without querying the victim model. These pseudo-labels, denoted as $c(x)$ , provide a starting point for training the initial substitute model $F_0'$ . Then, DSL proceeds iteratively, refining both the synthetic data for querying and the substitute model. In each iteration $t$ , a new set $\\mathbb{X}_{cand}^{t} = \\bigcup_{i}\\mathbb{X}_{cand,i}^{t}$ with $\\beta * b_{t}$ samples is generated using the LLM-driven Stable Diffusion process, where $\\beta$ is a scaling factor. These candidates are then subjected to a disagreement-based filtering based on the current substitute model's predictions. For each candidate $x \\in \\mathbb{X}_{cand}^{t}$ ,", + "bbox": [ + 89, + 719, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 StolenLoRA-DSL", + "bbox": [ + 514, + 90, + 727, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input: Victim model $F$ , Class names $\\mathbb{C}$ , Query budget $b_{t}$ , Scaling factor $\\beta$ , Threshold $\\tau$ , Initial sample size $N$ .", + "bbox": [ + 511, + 107, + 906, + 137 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Output: Substitute LoRA model $F^{\\prime}$", + "bbox": [ + 514, + 137, + 759, + 152 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: $\\mathbb{X}_0\\gets \\bigcup_i\\mathrm{SynIMG}(\\mathbb{C},N)$", + "2: $F_0^{\\prime}\\gets \\mathrm{Update}(\\mathbb{X}_0,\\mathcal{L}_{LR})$ 1 Initial substitute model", + "3: $t\\gets 0,B_t\\gets B$", + "4: while $B_{t} > 0$ do" + ], + "bbox": [ + 522, + 155, + 903, + 215 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "5: $\\mathbb{X}_{\\text{cand}}^t \\gets \\text{SynIMG}(\\mathbb{C}, \\beta b_t)$ Generate candidate", + "6: $\\mathbb{X}_{conf}^t\\gets \\emptyset ,\\mathbb{X}_{uncer}^t\\gets \\emptyset$", + "7: for $\\pmb{x} \\in \\mathbb{X}_{\\text{cand}}^t$ do", + "8: $\\hat{c},\\hat{p}\\gets \\operatorname {Predict}(F_t'(x))$", + "9: if $\\hat{c} = c(\\pmb{x})$ and $\\hat{p} \\geq \\tau$ then Disagreement filtering", + "10: $\\mathbb{X}_{conf}^t\\gets \\mathbb{X}_{conf}^t\\cup \\{(x,c(x))\\}$", + "11: else", + "12: $\\mathbb{X}_{uncer}^t\\gets \\mathbb{X}_{uncer}^t\\cup \\{\\pmb {x}\\}$", + "13: end if", + "14: end for", + "15: $\\mathbb{X}_{query}^t\\gets \\mathrm{Query}(F,\\mathbb{X}_{uncer}^t,b_t)$ $\\triangleright$ Selective Query", + "16: $\\mathbb{X}_{train}^{t}\\leftarrow \\mathbb{X}_{conf}^{t}\\cup \\mathbb{X}_{query}^{t}$", + "17: $F_{t + 1}^{\\prime}\\gets \\mathrm{Update}(\\mathbb{X}_{train}^{t},\\mathcal{L}_{LR})$", + "18: $B_{t}\\gets B_{t} - |\\mathbb{X}_{query}^{t}|,t\\gets t + 1$", + "19: end while", + "20: return $F_{t}^{\\prime}$" + ], + "bbox": [ + 519, + 155, + 903, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the substitute model $F_{t}^{\\prime}$ predicts a class $\\hat{c}$ and associated confidence $\\hat{p}$ . If the predicted class $\\hat{c}$ matches the prompt-based pseudo-label $c(\\boldsymbol{x})$ and the confidence $\\hat{p}$ exceeds a predefined threshold $\\tau$ , the pseudo-label is considered reliable, and the sample is added to a confidently labeled set $\\mathbb{X}_{conf}^{t} = \\{(\\boldsymbol{x}, c(\\boldsymbol{x}))\\}$ . Conversely, if the prediction disagrees with the pseudo-label or the confidence is low, the sample is added to an uncertain set $\\mathbb{X}_{uncer}^{t}$ .", + "bbox": [ + 511, + 493, + 906, + 616 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "From the uncertain set $\\mathbb{X}_{uncer}^{t}$ , a subset of $b_{t}$ samples with the lowest confidence scores are selected for querying the victim model $F$ . These queried samples, along with their true labels obtained from $F$ , form the query set $\\mathbb{X}_{query}^{t} = \\{(\\boldsymbol{x}, F(\\boldsymbol{x}))\\}$ . Combine the confidently pseudolabeled samples $\\mathbb{X}_{conf}^{t}$ and the queried samples $\\mathbb{X}_{query}^{t}$ to form the training set for this iteration: $\\mathbb{X}_{train}^{t} = \\mathbb{X}_{conf}^{t} \\cup \\mathbb{X}_{query}^{t}$ . Crucially, this combined set undergoes label refining during training to further address potential inaccuracies in pseudo-labels and mitigate the distribution shift.", + "bbox": [ + 511, + 616, + 905, + 770 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we employ a label refining strategy. Let $z$ be the logits produced by the substitute model for a given input $x$ . The label refining training loss is calculated as:", + "bbox": [ + 511, + 771, + 905, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {L R} (\\boldsymbol {z}, \\boldsymbol {q}) = - \\sum_ {i = 1} ^ {C} q _ {i} \\log \\left(\\frac {\\exp (z _ {i})}{\\sum_ {j = 1} ^ {C} \\exp (z _ {j})}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 830, + 903, + 873 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{q}$ are the soft labels. They are updated using an ex", + "bbox": [ + 513, + 885, + 903, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "266", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/71c64f2064c1250215e8ac81899d306936db061aef670caa615b1441548e2ba1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ScenarioMethodCUBS200Caltech256Indoor67Food101Flowers102
AccASRAccASRAccASRAccASRAccASR
Identical BackboneKnockoffNets70.9580.5485.6990.7179.1892.5982.5390.6276.2477.28
ActiveThief72.3382.1186.9292.0177.4690.5780.3488.2277.5178.56
DFME0.560.640.480.512.713.171.621.782.812.85
E371.9481.6782.3687.1881.4395.2279.6587.4680.6781.77
StolenLoRA-Rand75.3585.5487.6292.7582.3896.3379.0086.7593.7495.01
StolenLoRA-DSL73.2383.1389.3094.5382.6196.6080.5788.4787.4688.65
Cross BackboneKnockoffNets6.777.6941.3443.7641.7948.8714.4715.8914.1614.35
ActiveThief15.4217.5042.8945.4036.0442.1415.1116.5929.0929.49
DFME0.510.580.500.531.331.560.890.981.461.48
E317.5519.9248.1750.9955.8565.3140.5844.5648.2848.94
StolenLoRA-Rand45.7051.8851.7554.7859.1869.2043.1647.3959.2960.10
StolenLoRA-DSL50.1456.9265.0168.8265.0776.0944.5348.9061.1661.99
", + "bbox": [ + 94, + 89, + 903, + 329 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Effectiveness of StolenLoRA: the Acc (%) and ASR (%) of extracted substitute model by different model extraction attacks under 10k queries. (Boldface: the best value, Underline: the second-best value.)", + "bbox": [ + 89, + 339, + 906, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ponential moving average of the predicted probabilities $\\pmb{p}$ :", + "bbox": [ + 89, + 377, + 478, + 393 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {q} ^ {(i + 1)} = \\mu \\boldsymbol {q} ^ {(i)} + (1 - \\mu) \\boldsymbol {p} ^ {(i + 1)}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 402, + 482, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mu$ is the momentum parameter. This iterative refinement allows the model to learn from both the initial labels and its own evolving predictions, progressively improving the quality of the training data. The updated substitute model $F_{t + 1}^{\\prime}$ is thus trained on $\\mathbb{X}_{train}^{t}$ by minimizing:", + "bbox": [ + 89, + 433, + 483, + 511 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\prime} \\left(\\Delta \\theta_ {t + 1} ^ {\\prime}\\right) = \\mathbb {E} _ {\\boldsymbol {x} \\in \\mathbb {X} _ {t r a i n} ^ {t}} \\mathcal {L} _ {L R} \\left(F _ {t + 1} ^ {\\prime} (\\boldsymbol {x}), \\boldsymbol {q} (\\boldsymbol {x})\\right). \\quad (7)\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 521, + 482, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This iterative process continues until the query budget is exhausted, resulting in a final substitute model $F'$ trained on a combination of high-confidence pseudo-labeled data and a smaller set of strategically queried data. DSL allows for efficient use of the query budget by focusing on the most informative samples, leading to a more accurate substitute LoRA. We present the algorithm detail of it in Alg. 1.", + "bbox": [ + 89, + 551, + 483, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 672, + 225, + 690 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 696, + 282, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Victim Models. We use a ViT-Base model pre-trained on ImageNet-21k [33] with additional augmentations and regularization [38]. This base model is then fine-tuned with LoRA ( $r = 4$ ) on five commonly used datasets: CUBS200 [41], Caltech256 [9], Indoor67 [31], Food101 [2], and Flowers102 [26]. The resulting victim models achieve test accuracies of $88.09\\%$ , $94.47\\%$ , $85.52\\%$ , $91.07\\%$ , and $98.66\\%$ , respectively.", + "bbox": [ + 89, + 719, + 482, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Attack Settings. We evaluate LoRA extraction attacks in Identical-Backbone (IB) and Cross-Backbone (XB) scenarios. In the IB setting, we use the same base model as the substitute base model. In the XB setting, we use", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the ViT-base model pre-trained on ImageNet-1k[35] with self-supervised masked autoencoder (MAE) method[12] as the substitute base model. Both substitute models utilize LoRA with $r = 4$ and are trained for 20 epochs using the Adam optimizer[19] with a cosine annealing learning rate schedule[24] and a base learning rate of 0.01.", + "bbox": [ + 511, + 377, + 906, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use GPT-4o mini[27] as the LLM in StolenLoRA as it has similar performance at a lower price and can effectively reduce the attack cost. For the SD model, we use an open-source SDXL-Turbo model[36]. It only takes 4 sampling steps (compared to SDXL 1.0 which usually requires 40 steps) to synthesize high-quality images, significantly reducing the computational cost. For the StolenLoRA-DSL method, initial per-category sample size $N$ is set to 10, the scaling factor $\\beta$ is 1.5 to not increase too much generation overhead, and the confidence threshold $\\tau$ is 0.95. Hyperparameter analysis is provided in later sections.", + "bbox": [ + 511, + 494, + 908, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines and Evaluation Metric. We compare Stolen-LoRA against KnockoffNets [28], ActiveThief [29], DFME [40], and $E^3$ [48] using a query budget of 10k. For methods such as KnockoffNets, ActiveThief, and $E^3$ that require real data, we use 3M images in CC3M[37] as the attack dataset. For $E^3$ , which originally selects samples based on semantic similarity between dataset category names and target class names, we improve its effectiveness by calculating semantic similarity between image captions (available in CC3M) and target class names for a more fine-grained sample selection. Following prior work [28, 40], we report Test Accuracy (Acc) and Attack Success Rate (ASR), defined as the ratio of the substitute model's accuracy to the victim model's accuracy.", + "bbox": [ + 511, + 688, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "267", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fb642a10e02c21f2b5a1ad15c81bfb17c288d6bbef4cd1df8f80fd51e84d35a6.jpg", + "image_caption": [ + "Figure 2. Curves of the test accuracy versus the number of queries in IB (solid line) and XB (dashed line) scenarios respectively. The gray dashed line represents the victim models' test accuracy." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 911, + 233 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experimental Results", + "text_level": 1, + "bbox": [ + 89, + 276, + 294, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness of StolenLoRA. Tab. 1 presents the effectiveness of StolenLoRA, comparing both the random strategy (StolenLoRA-Rand) and DSL-based (StolenLoRADSL) variants against existing extraction attacks. Across both identical- and cross-backbone scenarios, StolenLoRA consistently achieves state-of-the-art performance. In the IB setting, StolenLoRA-Rand shows superior performance on CUBS200 (75.35%) and Flowers102 (93.74%), while StolenLoRA-DSL excels on Caltech256 (89.30%) and Indoor67 (82.61%), maintaining competitive results on other datasets. The XB setting, more challenging due to base model differences, highlights StolenLoRA's significant advantage over baselines, with StolenLoRA-DSL achieving the highest accuracy across all five datasets. This demonstrates DSL's effectiveness in refining synthetic data and improving attack efficiency. Fig. 2 further analyzes query efficiency, showing StolenLoRA-DSL's rapid convergence in the IB setting and consistent progress in the XB setting, outperforming other methods in both scenarios. This efficiency stems from DSL's ability to prioritize informative queries, focusing on uncertain samples, thereby maximizing information gain from each query.", + "bbox": [ + 91, + 305, + 483, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness under Hard-Label Scenario. We also evaluate StolenLoRA's performance when only hard labels (one-hot encoded predictions) are available, as is the case with some real-world APIs. Fig. 3 presents the results for both StolenLoRA-Rand and StolenLoRA-DSL in the IB and XB settings. In the IB setting, using hard labels leads to a minor performance decrease. For example, StolenLoRA-Rand's accuracy on CUBS200 drops from $75.35\\%$ with soft labels to $67.00\\%$ with hard labels. A similar trend is observed for other datasets and StolenLoRA-DSL. This suggests that the additional information provided by soft labels is beneficial in the IB setting. In the XB setting, the impact of hard labels is inconsistent. While a slight decrease is observed on some datasets, others, like Caltech256 and Indoor67, show a marginal increase in accuracy (from $51.75\\%$ to $56.69\\%$ ). This unexpected improvement might be due to the hard labels acting as a regularizer, preventing the substi", + "bbox": [ + 89, + 643, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/018d8e09783e3741ce179587f870e0b434b43444fe5b8d91a1d1c75460e10ece.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 513, + 273, + 906, + 436 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ddd5145e5d8f3a70ae5c14228176d60b1e8b30e63fa16790db2aee0a4bc1dd4e.jpg", + "image_caption": [ + "Figure 3. Comparison of StolenLoRA performance using hard labels (one-hot) versus soft labels (probabilities) from the victim model. Results are shown for both Random and DSL in identical- and cross-backbone scenarios across five datasets." + ], + "image_footnote": [], + "bbox": [ + 511, + 436, + 906, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tute from overfitting to the synthetic data, which is more prone to distributional shifts in the XB setting. Overall, StolenLoRA exhibits robustness to hard labels in both IB and XB scenarios, demonstrating its practical applicability.", + "bbox": [ + 511, + 641, + 908, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ablation Study. We conduct an ablation study to analyze the contributions of the key components of StolenLoRA: the LLM-driven prompting and the DSL strategy, including the label refining loss. Tab. 2 presents the results on CUBS200 and Indoor67 datasets under both IB and XB settings with a 10k query budget. Removing the structured prompt template leads to a performance drop in both IB and XB settings, demonstrating the importance of providing structured visual information to the LLM. Removing the LLM entirely results in a more significant performance decrease, confirming the crucial role of the LLM in generating effective prompts for diverse and representative synthetic data. Furthermore, we evaluate different LLMs, observing that using", + "bbox": [ + 511, + 703, + 910, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "268", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7a84661705af436a534a2ab11f13cf6115c8c4da1d1e856cb19773235e44ad4a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodIBXB
CUBS200Indoor67CUBS200Indoor67
Random75.3582.3845.7059.18
- Template72.3380.9442.2154.49
- LLM70.6674.7039.2047.09
+ Llama-3.1-8B73.9681.0446.7458.43
+ GPT-4o76.3083.1349.1963.58
DSL73.2382.6150.1465.07
- LLR70.1780.9748.2164.10
", + "bbox": [ + 94, + 89, + 478, + 220 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f91a2cc6d0bfa574c916752d1b9b14b0888bc2ba8351aa6e5545b9ad0e8d3a6c.jpg", + "table_caption": [ + "Table 2. An ablation experiment showing the effectiveness of the two modules we designed on CUBS200 and Indoor67 datasets under both IB and XB settings with 10k queries." + ], + "table_footnote": [], + "table_body": "
τIBXB
CUBS200Indoor67CUBS200Indoor67
0.570.6878.8148.4360.90
0.771.3780.0048.3663.51
0.974.0481.1948.5263.06
0.9573.2382.6150.1465.07
0.9973.3181.1949.8661.72
", + "bbox": [ + 96, + 277, + 478, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "a more powerful LLM like GPT-4o generally yields better performance than Llama-3.1-8B[5], particularly in the more challenging XB setting. This highlights the benefit of utilizing a stronger LLM for prompt generation.", + "bbox": [ + 89, + 460, + 482, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Disabling DSL and reverting to random querying results in a noticeable performance reduction, especially in the XB setting. This underscores the effectiveness of DSL in selectively querying informative samples. Similarly, removing SAT from the DSL framework also leads to a performance decrease, albeit less pronounced than removing DSL altogether. This indicates the benefit of label refining in mitigating the distribution shift between synthetic and real data.", + "bbox": [ + 89, + 522, + 482, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also analyze the impact of the confidence threshold $\\tau$ within DSL in Tab. 3. The results show that performance generally improves as $\\tau$ increases from 0.5 to 0.95. This indicates that focusing queries on samples where the substitute model is less confident leads to a more efficient use of the query budget. However, increasing $\\tau$ further to 0.99 results in a slight performance decrease, suggesting that an overly stringent threshold can exclude valuable samples from being queried. We therefore select $\\tau = 0.95$ as the optimal value for our experiments.", + "bbox": [ + 89, + 643, + 482, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fidelity and Distribution Analysis of Synthetic Data. StolenLoRA's effectiveness relies on the quality of its synthetic training data. We evaluate this quality both visually and quantitatively. Visually, as shown in Fig. 4, synthetic images generated by StolenLoRA demonstrate a striking resemblance to real images from the target categories, capturing key features and background context crucial for train-", + "bbox": [ + 89, + 795, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e3667a820246873db8cca7a1f6f2627b73ecfc3481137a2156d4a902a93990f6.jpg", + "image_caption": [ + "(a) Visualization on CUBS200" + ], + "image_footnote": [], + "bbox": [ + 526, + 88, + 897, + 220 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dcc1be32d6ec2ff6484cb54d57bcbe9b990544315cb4489d539c0b297009618f.jpg", + "image_caption": [ + "(b) Visualization on Indoor67", + "Figure 4. Visualization of StolenLoRA's synthetic images and corresponding real images of the target category." + ], + "image_footnote": [], + "bbox": [ + 524, + 236, + 897, + 369 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/701e56bcc741396b44a45f0cda81d2fc6e31e764ccbfe198acc0adf01fb33c73.jpg", + "table_caption": [ + "Table 3. Performance of StolenLoRA-DSL with varying confidence thresholds $\\tau$ . Results are reported on CUBS200 and Indoor67 under both IB and XB settings with 10k queries." + ], + "table_footnote": [], + "table_body": "
MethodCUBS200Caltech256Indoor67Food101Flowers102
KnockoffNets51.448.9422.0150.0455.77
ActiveThief49.336.0115.7247.8450.19
DFME220.53179.45225.84188.29204.79
E341.574.296.6825.5134.82
StolenLoRA2.145.596.9319.5416.63
", + "bbox": [ + 517, + 419, + 901, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. FID scores $(\\downarrow)$ between the datasets used by each attack method and the victim model's training data distribution.", + "bbox": [ + 511, + 497, + 903, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing. This visual fidelity is corroborated by significantly lower Fréchet Inception Distance (FID) scores[13] (in Tab. 4) compared to other attack methods. For example, StolenLoRA achieves an FID of 2.14 compared to 51.44 for KnockoffNets on CUBS200, highlighting a much closer distributional match to the target domain. This demonstrates that StolenLoRA's LLM-driven synthesis effectively generates high-fidelity, domain-specific data, even without access to the original training set, which is key to its superior extraction performance.", + "bbox": [ + 511, + 537, + 906, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Verifying the Distinction Hypothesis. As mentioned in Sec. 3.1, we hypothesize that the performance of a perfectly extracted surrogate model will be consistent across In-Distribution (ID) data with the victim model, but not necessarily Out-of-Distribution (OOD) data, motivating us to perform extraction attacks using ID data. To empirically verify this conjecture, we train a substitute model using the XB setting on the victim model's training dataset, aiming to achieve a near-perfect extraction of the LoRA adaptations. We then evaluate both the victim model and the extracted substitute on two sets of data: ID data, represented by the victim model's held-out test set, and OOD data, sampled from the CC3M dataset. For each sample, we calculate the cross-entropy between the victim model's predictions and", + "bbox": [ + 511, + 689, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "269", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a8effbcfe67e86eaaf5ffbcb4f61507356f7e1869bd230eb26787cc60792d0c0.jpg", + "image_caption": [ + "Figure 5. Cross-entropy difference between the victim model and a near-perfectly substitute model on In-Distribution (ID) and Out-of-Distribution (OOD) data. The substitute model is trained using the cross-backbone setting on the victim's training dataset." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 480, + 234 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the substitute model's predictions, providing a measure of their functional divergence. Fig. 5 presents the distribution of these differences. On both CUBS200 and Indoor67, the difference is smaller for the ID data compared to the OOD data. This observation supports our conjecture that even near-perfect LoRA extraction yields a substitute model that exhibits functional equivalence primarily within the ID domain. The divergence in OOD data likely stems from the inherent differences between the pre-trained backbone used by the victim and the substitute. This result underscores the importance of utilizing ID data for effective LoRA extraction, as the attacker primarily seeks to replicate the victim's specialized adaptations rather than the general capabilities.", + "bbox": [ + 88, + 308, + 483, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Defending Against LoRA Extraction", + "text_level": 1, + "bbox": [ + 89, + 518, + 419, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While this paper primarily focuses on the feasibility of LoRA extraction, we also explore a potential defense mechanism based on deploying multiple, diverse LoRA adapters to increase attacker uncertainty.", + "bbox": [ + 89, + 542, + 482, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dual LoRA with Diversified Predictions. Our proposed defense involves training two LoRA adapters, denoted as $\\mathrm{L}_A$ and $\\mathrm{L}_B$ , to maximize the difference in their output distributions while maintaining comparable performance on the target task. Instead of minimizing accuracy as initially conceived, we found that maintaining high accuracy while maximizing divergence is crucial for practical deployment. We achieve this by optimizing the following loss function:", + "bbox": [ + 89, + 603, + 482, + 724 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\prime} = \\mathcal {L} (\\mathrm {L} _ {A}) + \\mathcal {L} (\\mathrm {L} _ {B}) - \\lambda \\times \\mathrm {K L} (\\mathrm {L} _ {A} | | \\mathrm {L} _ {B}) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 736, + 482, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{L}$ represents the cross-entropy loss for the target task, KL represents the Kullback-Leibler divergence between the output distributions of the two LoRAs, and $\\lambda$ is a hyperparameter controlling the balance between task performance and divergence. At deployment, we randomly select either $L_{A}$ or $L_{B}$ for each incoming query to generate the prediction. This random selection introduces uncertainty and makes it challenging for the attacker to extract the underlying LoRA function consistently.", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ac1b89e4554fa8412367296f397452b438629a1081ed55ee1b1f02be72bd8252.jpg", + "image_caption": [ + "Figure 6. Defneder Acc versus Substitute model Acc trade-off for defenses evaluated against StolenLoRA and KnockoffNets on the CUBS200 dataset with 10k queries." + ], + "image_footnote": [], + "bbox": [ + 513, + 87, + 903, + 234 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Experimental Evaluation. We evaluate the proposed defense mechanism under both IB and XB settings with 10k queries. Fig. 6 illustrates the trade-off between defender Acc and substitute model Acc. The defense consistently degrades the substitute model's performance. KnockoffNets experiences substantial accuracy drops (from $70.90\\%$ to $20.78\\%$ in IB and from $6.77\\%$ to $1.21\\%$ in XB), and StolenLoRA-Rand shows similar vulnerability (decreasing $22.95\\%$ in IB and $15.63\\%$ in XB). However, StolenLoRADSL demonstrates greater resilience, with comparatively smaller accuracy reductions ( $9.63\\%$ in IB and $14.34\\%$ in XB). This resilience is attributed to DSL's efficient query selection and label refining, focusing on informative samples and mitigating the uncertainty introduced by the defense. While this defense effectively hinders less sophisticated attacks, it underscores the need for more robust mechanisms to counter advanced techniques like DSL that can learn effectively from limited queries.", + "bbox": [ + 511, + 285, + 906, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 571, + 633, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces LoRA Extraction, a novel model extraction attack targeting the widespread practice of adapting large vision models with LoRA. We present StolenLoRA, a highly effective method leveraging LLM-driven Stable Diffusion to generate task-specific, high-fidelity synthetic training data, bypassing the limitations of traditional extraction techniques reliant on large-scale real datasets or unreliable GANs. Furthermore, StolenLoRA incorporates Disagreement-based Semi-supervised Learning (DSL) to efficiently query the victim model, focusing on uncertain predictions and iteratively refining labels, enabling successful extraction with limited queries. Extensive experiments demonstrate StolenLoRA's strong performance across diverse datasets and in challenging cross-backbone scenarios. We also explore a preliminary defense based on diversified LoRA deployments, showing promise in mitigating these attacks. This work represents a crucial first step towards understanding and addressing the safety risks associated with LoRA adaptation, paving the way for more secure deployments of efficient fine-tuning methods.", + "bbox": [ + 511, + 598, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "270", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is in part supported by National Key R&D Program of China (Grant No. 2022ZD0160103) and National Natural Science Foundation of China (Grant No. 62276067), and Shanghai Artificial Intelligence Laboratory.", + "bbox": [ + 89, + 113, + 485, + 172 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 196, + 187, + 213 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Martin Arjovsky and Léon Bottou. Towards principled methods for training generative adversarial networks. arXiv preprint arXiv:1701.04862, 2017. 2", + "[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In ECCV, 2014. 5", + "[3] Eric L Buehler and Markus J Buehler. X-lora: Mixture of low-rank adapter experts, a flexible framework for large language models with applications in protein mechanics and molecular design. APL Machine Learning, 2024. 2", + "[4] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1", + "[5] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 7", + "[6] Ricard Durall, Avraam Chatzimichailidis, Peter Labus, and Janis Keuper. Combating mode collapse in gan training: An empirical analysis using hessian eigenvalues. arXiv preprint arXiv:2012.09673, 2020. 2", + "[7] Lijie Fan, Kaifeng Chen, Dilip Krishnan, Dina Katabi, Phillip Isola, and Yonglong Tian. Scaling laws of synthetic images for model training... for now. In CVPR, 2024. 2", + "[8] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NIPS, 2014. 1, 2", + "[9] Gregory Griffin, Alex Holub, and Pietro Perona. Caltech-256 object category dataset. 2007. 5", + "[10] Zeyu Han, Chao Gao, Jinyang Liu, Sai Qian Zhang, et al. Parameter-efficient fine-tuning for large models: A comprehensive survey. arXiv preprint arXiv:2403.14608, 2024. 2", + "[11] Soufiane Hayou, Nikhil Ghosh, and Bin Yu. Lora+: Efficient low rank adaptation of large models. arXiv preprint arXiv:2402.12354, 2024. 1", + "[12] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 1, 5", + "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 7", + "[14] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In ICML, 2019. 2" + ], + "bbox": [ + 93, + 222, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1, 2", + "[16] Nam Hyeon-Woo, Moon Ye-Bin, and Tae-Hyun Oh. Fedpara: Low-rank hadamard product for communication-efficient federated learning. arXiv preprint arXiv:2108.06098, 2021. 2", + "[17] Matthew Jagielski, Nicholas Carlini, David Berthelot, Alex Kurakin, and Nicolas Papernot. High accuracy and high fidelity extraction of neural networks. In *Usenix Security*, 2020. 1", + "[18] Sanjay Kariyappa, Atul Prakash, and Moinuddin Qureshi. Maze: Data-free model stealing attack using zeroth-order gradient estimation. In CVPR, 2021. 1, 2", + "[19] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5", + "[20] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In CVPR, 2023. 1", + "[21] Youssef Kossale, Mohammed Airaj, and Aziz Darouichi. Mode collapse in generative adversarial networks: An overview. In ICOA, 2022. 2", + "[22] Zi Liang, Qingqing Ye, Yanyun Wang, Sen Zhang, Yaxin Xiao, Ronghua Li, Jianliang Xu, and Haibo Hu. Alignment-aware model extraction attacks on large language models. arXiv preprint arXiv:2409.02718, 2024. 1, 2", + "[23] Zijun Lin, Ke Xu, Chengfang Fang, Huadi Zheng, Aneez Ahmed Jaheezuddin, and Jie Shi. Quda: Query-limited data-free model extraction. In ACM Asia CCS, 2023. 1, 2", + "[24] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In *ICLR*, 2017. 5", + "[25] Xingjun Ma, Yifeng Gao, Yixu Wang, Ruofan Wang, Xin Wang, Ye Sun, Yifan Ding, Hengyuan Xu, Yunhao Chen, Yunhan Zhao, et al. Safety at scale: A comprehensive survey of large model safety. arXiv preprint arXiv:2502.05206, 2025.1", + "[26] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian conference on computer vision, graphics & image processing, 2008. 5", + "[27] OpenAI. Gpt-4o mini: Advancing cost-efficient intelligence. https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence/, 2024.5", + "[28] Tribhuvanesh Orekondy, Bernt Schiele, and Mario Fritz. Knockoff nets: Stealing functionality of black-box models. In CVPR, 2019. 1, 2, 5", + "[29] Soham Pal, Yash Gupta, Aditya Shukla, Aditya Kanade, Shirish Shevade, and Vinod Ganapathy. Activethief: Model extraction using active learning and unannotated public data. In AAAI, 2020. 1, 2, 5", + "[30] Jie Qin, Jie Wu, Weifeng Chen, Yuxi Ren, Huixia Li, Hefeng Wu, Xuefeng Xiao, Rui Wang, and Shilei Wen. Diffusiongpt: Llm-driven text-to-image generation system. arXiv preprint arXiv:2401.10061, 2024. 2" + ], + "bbox": [ + 516, + 92, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "271", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Ariadna Quattoni and Antonio Torralba. Recognizing indoor scenes. In CVPR, 2009. 5", + "[32] Pengzhen Ren, Yun Xiao, Xiaojun Chang, Po-Yao Huang, Zhihui Li, Brij B Gupta, Xiaojiang Chen, and Xin Wang. A survey of deep active learning. ACM computing surveys, 2021. 2", + "[33] Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik-Manor. Imagenet-21k pretraining for the masses. arXiv preprint arXiv:2104.10972, 2021. 5", + "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2", + "[35] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 2015. 5", + "[36] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 2, 5", + "[37] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 5", + "[38] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. arXiv preprint arXiv:2106.10270, 2021. 5", + "[39] Florian Tramér, Fan Zhang, Ari Juels, Michael K. Reiter, and Thomas Ristenpart. Stealing machine learning models via prediction apis. In USENIX Security Symposium, 2016. 1", + "[40] Jean-Baptiste Truong, Pratyush Maini, Robert J Walls, and Nicolas Papernot. Data-free model extraction. In CVPR, 2021. 1, 2, 5", + "[41] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5", + "[42] Yixu Wang, Jie Li, Hong Liu, Yan Wang, Yongjian Wu, Feiyue Huang, and Rongrong Ji. Black-box dissector: Towards erasing-based hard-label model stealing attack. In ECCV, 2022. 1, 2", + "[43] Yixu Wang, Tianle Gu, Yan Teng, Yingchun Wang, and Xingjun Ma. Honeypotnet: Backdoor attacks against model extraction. In AAAI, 2025. 1", + "[44] Shih-Ying Yeh, Yu-Guan Hsieh, Zhidong Gao, Bernard BW Yang, Giyeong Oh, and Yanmin Gong. Navigating text-to-image customization: From lycoris fine-tuning to model evaluation. In ICLR, 2023. 2", + "[45] Jianhao Yuan, Jie Zhang, Shuyang Sun, Philip Torr, and Bo Zhao. Real-fake: Effective training data synthesis through distribution matching. In ICLR, 2024. 2", + "[46] Yunlong Zhao, Xiaoheng Deng, Yijing Liu, Xinjun Pei, Ji-azhi Xia, and Wei Chen. Fully exploiting every real sample: Superpixel sample gradient model stealing. In CVPR, 2024. 1, 2", + "[47] Mingyi Zhou, Jing Wu, Yipeng Liu, Shuaicheng Liu, and Ce Zhu. Dast: Data-free substitute training for adversarial attacks. In CVPR, 2020. 1, 2" + ], + "bbox": [ + 89, + 92, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "[48] Hongyu Zhu, Wentao Hu, Sichu Liang, Fangqi Li, Wenwen Wang, and Shilin Wang. Efficient and effective model extraction. arXiv preprint arXiv:2409.14122, 2024. 5", + "bbox": [ + 514, + 92, + 905, + 133 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "272", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/StolenLoRA_ Exploring LoRA Extraction Attacks via Synthetic Data/f4ddd460-9cf1-429e-89c2-3d861bf9c661_model.json b/2025/StolenLoRA_ Exploring LoRA Extraction Attacks via Synthetic Data/f4ddd460-9cf1-429e-89c2-3d861bf9c661_model.json new file mode 100644 index 0000000000000000000000000000000000000000..39065915ef478290b78592986234caba9783e79b --- /dev/null +++ b/2025/StolenLoRA_ Exploring LoRA Extraction Attacks via Synthetic Data/f4ddd460-9cf1-429e-89c2-3d861bf9c661_model.json @@ -0,0 +1,2244 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.151, + 0.131, + 0.848, + 0.154 + ], + "angle": 0, + "content": "StolenLoRA: Exploring LoRA Extraction Attacks via Synthetic Data" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.18, + 0.734, + 0.201 + ], + "angle": 0, + "content": "Yixu Wang\\(^{1,2\\dagger}\\), Yan Teng\\(^{2*}\\), Yingchun Wang\\(^{2}\\), Xingjun Ma\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.205, + 0.76, + 0.225 + ], + "angle": 0, + "content": "\\(^{1}\\)Fudan University \\(^{2}\\)Shanghai Artificial Intelligence Laboratory" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.259, + 0.327, + 0.275 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.291, + 0.486, + 0.669 + ], + "angle": 0, + "content": "Parameter-Efficient Fine-Tuning (PEFT) methods like LoRA have transformed vision model adaptation, enabling the rapid deployment of customized models. However, the compactness of LoRA adaptations introduces new safety concerns, particularly their vulnerability to model extraction attacks. This paper introduces a new focus of model extraction attacks named LoRA extraction that extracts LoRA-adaptive models based on a public pre-trained model. We then propose a novel extraction method called StolenLoRA which trains a substitute model to extract the functionality of a LoRA-adapted model using synthetic data. StolenLoRA leverages a Large Language Model to craft effective prompts for data generation, and it incorporates a Disagreement-based Semi-supervised Learning (DSL) strategy to maximize information gain from limited queries. Our experiments demonstrate the effectiveness of StolenLoRA, achieving up to a \\(96.60\\%\\) attack success rate with only 10k queries, even in cross-backbone scenarios where the attacker and victim models utilize different pretrained backbones. These findings reveal the specific vulnerability of LoRA-adapted models to this type of extraction and underscore the urgent need for robust defense mechanisms tailored to PEFT methods. We also explore a preliminary defense strategy based on diversified LoRA deployments, highlighting its potential to mitigate such attacks." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.696, + 0.222, + 0.71 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.721, + 0.485, + 0.842 + ], + "angle": 0, + "content": "Adapting large-scale pre-trained foundation models via Parameter-Efficient Fine-Tuning (PEFT) such as Low-Rank Adaptation (LoRA) [11, 15] has become a popular approach for obtaining high-performance models for diverse downstream tasks. LoRA enables efficient adaptation of large models to a new task by fine-tuning only low-rank matrices added to specific layers, reducing computational cost and memory usage while preserving the pre-trained model's" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.26, + 0.908, + 0.473 + ], + "angle": 0, + "content": "core knowledge. Despite the wide adoption of LoRA, the lightweight and compactness of LoRA parameters raise a new safety concern: they might be more vulnerable to model extraction (ME) attacks. An ME attack extracts the functionality [17, 22, 28, 39] of a victim model by training a substitute model based on the outputs of the victim model for a certain number of specially designed queries. While ME attacks have been extensively studied for traditional models [29, 40, 42, 43, 46, 47], the vulnerability of LoRA adaptations to ME attacks remains largely unexplored[25]. This gap is significant, as the public availability of a large number of pre-trained models, coupled with the compactness of LoRA parameters, makes it easier to replicate LoRA-adapted models and compromise intellectual property." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.475, + 0.909, + 0.717 + ], + "angle": 0, + "content": "To address this gap, we introduce the concept of LoRA Extraction, a novel direction of ME attacks specifically targeting LoRA-adapted models (i.e., ViTs [4, 12, 20]), based on a publicly available pre-trained model. Unlike traditional ME which focuses on replicating the entire model's functionality, LoRA extraction centers on stealing the efficient adaptations encoded within the compact LoRA parameters. The adversary's objective is to reconstruct the victim's LoRA-adapted model by training their own LoRA-adapted substitute that achieves similar downstream performance using a pre-trained foundation model that is publicly available. This attack can manifest in two scenarios: the identical-backbone scenario, where the attacker uses the same pre-trained ViT model as the victim, and the more challenging cross-backbone scenario, where the attacker uses a different pre-trained ViT." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Existing traditional ME methods largely depend on sample selection from existing datasets or synthetic data generation. Sample selection methods [28, 29, 42, 46] typically require searching extensive datasets for in-distribution samples, making them computationally expensive and less practical for LoRA extraction. This impracticality arises because 1) inference in LoRA-adapted models is slower due to the large-scale parameters of LoRA-adapted models, and 2) finding appropriate samples is challenging, particularly for domain-specific data fine-tuned with LoRA. Alternatively, synthetic data generation approaches [18, 23, 40, 47] often rely on Generative Adversarial Networks (GANs) [8]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.849, + 0.485, + 0.902 + ], + "angle": 0, + "content": "\\*Corresponding authors: < tengyan@pjlab.org.cn, xingjunma@fudan.edu.cn > T Work done during internship at Shanghai Artificial Intelligence Laboratory." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "263" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.227 + ], + "angle": 0, + "content": "to produce in-distribution data. However, GANs frequently face difficulties in generating high-quality, diverse samples, especially for high-dimensional data. As Zhao et al. [46] noted, creating high-dimensional samples (e.g., \\(224 \\times 224\\) images) for effective ME can require millions of queries and is prone to failure due to GAN mode collapse [1, 6, 21]. These limitations make GAN-based approaches less effective for LoRA extraction, where the tuning often involves high-dimensional image data." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.228, + 0.487, + 0.558 + ], + "angle": 0, + "content": "In this work, we propose a novel LoRA extraction method named StolenLoRA, which employs synthetic data to train the substitute model. To obtain effective in-distribution synthetic data, we leverage Large Language Models (LLMs) to generate diverse textual descriptions based on the target class names and prompt a pre-trained Stable Diffusion model [34, 36] to generate high-quality images. This allows us to tailor the generated data to the specific downstream task the victim's LoRA model is fine-tuned for. Moreover, StolenLoRA introduces a distinctive strategy to improve attack efficiency, i.e., the Disagreement-based Semi-supervised Learning (DSL). DSL uses class information from the data synthesis process as pseudo-labels for part of the generated dataset and focuses on queries where the substitute model disagrees with these pseudo-labels. DSL effectively targets areas of uncertainty, guiding the substitute model's learning process and refining its alignment with the victim's behavior. To further enhance DSL, we iteratively refine the pseudo-labels based on evolving predictions from the substitute model, improving the quality of synthetic data labels and the effectiveness of the extraction." + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.56, + 0.411, + 0.574 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.575, + 0.483, + 0.635 + ], + "angle": 0, + "content": "- We explore the vulnerability of LoRA and introduce a novel focus of ME attacks called LoRA Extraction which aims to extract the functionality of LoRA-adapted models based on a publicly available pre-trained model." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.636, + 0.483, + 0.71 + ], + "angle": 0, + "content": "- We present StolenLoRA, a novel LoRA extraction method that leverages LLM-driven Stable Diffusion to generate high-quality synthetic data, circumventing the need to search large-scale datasets or rely on less dependable GAN-based generation." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.711, + 0.483, + 0.802 + ], + "angle": 0, + "content": "- Through comprehensive experiments on five widely used datasets, we demonstrate the effectiveness of Stolen-LoRA in both identical- and cross-backbone scenarios, highlighting critical vulnerabilities in LoRA-adapted models. We also explore a preliminary defense mechanism based on diversified LoRA deployments." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.575, + 0.483, + 0.802 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.815, + 0.233, + 0.831 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Low-Rank Adaptation (LoRA). LoRA [15] is a type of popular Parameter-Efficient Fine-Tuning (PEFT) method [10, 14] for adapting large pre-trained models for downstream tasks. LoRA [15] introduces small rank-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.304 + ], + "angle": 0, + "content": "decomposition matrices as trainable parameters into the pre-trained model. The number of trainable parameters is significantly reduced by freezing the original model weights and updating only these lightweight LoRA modules. LoRA inspires numerous variants and extensions, such as Mixture-of-LoRA Experts (X-LoRA)[3], Low-Rank Hadamard Product (LoHa)[16], and Low-Rank Kronecker Product (LoKr)[44], further enhancing its efficiency and flexibility for various applications. However, the compact nature of these LoRA updates raises concerns about their vulnerability to model extraction attacks. This paper investigates the potential for adversaries to extract the efficient adaptations encoded within these compact parameters, specifically focusing on the standard LoRA method [15]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.337, + 0.907, + 0.581 + ], + "angle": 0, + "content": "Model Extraction Attacks. Model Extraction (ME) attacks [18, 22, 23, 28, 29, 40, 42, 46, 47] aim to construct a substitute model that mimics the functionality of a victim model by querying its API. Existing ME methods can be broadly categorized based on their query generation strategies: Data-Driven ME leverage existing datasets [28, 29, 42, 46] to select queries, often employing active learning[32] or other search strategies[28] to identify informative samples. However, this type can incur computational overhead as it necessitates traversing the large-scale dataset to find suitable in-distribution samples. Data-Free ME [18, 23, 40, 46, 47] utilize generative models (e.g., GANs[8]) to synthesize data for querying the victim model. While promising, generating effective queries with high-dimensional inputs can require tens of millions of queries and is prone to failure due to GAN mode collapse [1, 6, 21]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Learning from Synthetic Data. The use of synthetic data for training models gains significant traction due to its potential to address data scarcity. While traditional synthetic data often lacked the complexity and representativeness of real data, recent advancements in generative models, particularly diffusion models like Stable Diffusion [34, 36], enable the synthesis of high-quality and diverse images. However, scaling the quantity of synthetic data does not necessarily translate to improved performance for training supervised image classifiers. As shown by Fan et al.[7], performance with synthetic data can lag significantly behind training with real data. This disparity stems from the limitations of current text-to-image models in accurately generating diverse representations of certain concepts at scale[30]. Efforts to mitigate this gap, such as the distribution-matching framework[45], focus on aligning the distributions of synthetic and real data. However, leveraging such high-quality synthetic data for model extraction attacks remains an unexplored area." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "264" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.9, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.327, + 0.908, + 0.357 + ], + "angle": 0, + "content": "Figure 1. Overview of StolenLoRA. It collects target victim model information (dashed red box), synthesizes training data guided by an LLM-driven prompting (dashed blue box), and then efficiently queries the victim API to train a substitute model (dashed green box)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.362, + 0.226, + 0.38 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.387, + 0.295, + 0.402 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.408, + 0.484, + 0.545 + ], + "angle": 0, + "content": "LoRA Extraction. Given a victim model \\( F \\colon [0,1]^d \\mapsto \\mathbb{R}^N \\), which is a pre-trained model \\( F_{base} \\) (e.g., ViTs) adapted using LoRA. The parameters of \\( F^1 \\) can be represented as \\( \\theta = \\{\\theta_{base}, \\Delta \\theta\\} \\) where \\( \\theta_{base} \\) are the parameters of \\( F_{base} \\) and \\( \\Delta \\theta \\) represents the LoRA updates, which are typically low-rank matrices. The objective of a LoRA extraction attack is to train a substitute model \\( F' \\) with parameters \\( \\theta' = \\{\\theta_{base}', \\Delta \\theta'\\} \\) that mimics the functionality of \\( F \\). This is formulated as the following optimization problem:" + }, + { + "type": "equation", + "bbox": [ + 0.189, + 0.552, + 0.482, + 0.575 + ], + "angle": 0, + "content": "\\[\n\\arg \\min _ {\\Delta \\theta^ {\\prime}} \\mathbb {E} _ {\\boldsymbol {x}} \\mathcal {D} (F (\\boldsymbol {x}), F ^ {\\prime} (\\boldsymbol {x})), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.484, + 0.629 + ], + "angle": 0, + "content": "where \\( \\pmb{x} \\in \\mathbb{R}^d \\) represents an input sample from the victim model task distribution, \\( \\mathcal{D}(\\cdot, \\cdot) \\) represents a measure of functional similarity between the two models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.484, + 0.688 + ], + "angle": 0, + "content": "Depending on whether the attacker has knowledge of \\( F_{base} \\), we distinguish between two scenarios: identical backbone and cross-backbone LoRA extractions. The substitute model \\( F' \\) can be defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.695, + 0.482, + 0.746 + ], + "angle": 0, + "content": "\\[\nF ^ {\\prime} (\\boldsymbol {x}) = \\left\\{ \\begin{array}{l l} F _ {\\text {b a s e}} (\\boldsymbol {x}) + \\Delta F ^ {\\prime} (\\boldsymbol {x}), & \\text {i d e n t i c a l - b a c k b o n e} \\\\ G _ {\\text {b a s e}} (\\boldsymbol {x}) + \\Delta F ^ {\\prime} (\\boldsymbol {x}), & \\text {c r o s s - b a c k b o n e} \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.748, + 0.483, + 0.792 + ], + "angle": 0, + "content": "where the former uses the same \\( F_{base} \\), the latter uses different base models \\( G_{base} \\), and the \\( + \\) symbol denotes the combination of the base model and the LoRA updates." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.793, + 0.484, + 0.869 + ], + "angle": 0, + "content": "Distinction from Traditional Extraction. Traditional model extraction trains a substitute from scratch, while LoRA extraction leverages a pre-trained base model, inheriting its knowledge and biases. Consequently, we hypothesize that a perfectly extracted LoRA substitute might" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.907, + 0.439 + ], + "angle": 0, + "content": "achieve functional equivalence on In-Distribution (ID) data but not on Out-of-Distribution (OOD) data, particularly in the cross-backbone setting, and we experimentally verify this in the experimental section. This motivates the use of ID data for effective LoRA extraction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.447, + 0.644, + 0.46 + ], + "angle": 0, + "content": "3.2. StolenLoRA" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.906, + 0.575 + ], + "angle": 0, + "content": "Overview. An overview of our proposed LLM-driven generative attack for extracting LoRA, StolenLoRA, is illustrated in Fig. 1. It comprises two key stages: 1) Data Synthesis, where an LLM guides the generation of synthetic data tailored to the victim's task, and 2) Efficient Querying and Training, which strategically selects queries and refines labels to maximize information gain from the victim model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.581, + 0.663, + 0.595 + ], + "angle": 0, + "content": "3.2.1. Data Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.795 + ], + "angle": 0, + "content": "The above discussion motivates us to use ID data for extraction. However, acquiring such datasets in real-world scenarios is often challenging because the victim model's training data is proprietary. While the attacker might not have access to the original training data, they often have access to the model's deployment context, which usually includes a description of the model's functionality and the names of the target classes. This information can be leveraged to semantically expand the search space for potential training data. Inspired by this, we propose to utilize LLM-driven Stable Diffusion to generate synthetic data based on the class names, effectively bridging the gap between semantic information and visual representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.901 + ], + "angle": 0, + "content": "LLM-Driven Prompting. Given a set of target class names \\(\\mathbb{C} = \\{c_1, c_2, \\dots, c_n\\}\\), we employ an LLM to generate detailed prompts for image synthesis. We structure the prompts using a predefined template \\(T\\) comprising key visual elements: \\(T = \\text{\"[Subject, Background, Angle/ Pose, Lighting, Style]\"}\\). For each class \\(c_i\\), the LLM generates \\(m\\) variations of the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1We omit the parameter for simplicity when there is no ambiguity, e.g., using \\( F(\\pmb{x}) \\) instead of \\( F(\\pmb{x};\\theta) \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "265" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.138 + ], + "angle": 0, + "content": "prompt, denoted as \\( p_{i,j} \\), where \\( j \\in \\{1, \\dots, m\\} \\). This diversification ensures a richer representation of the target class. The prompt generation process is formalized as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.148, + 0.483, + 0.166 + ], + "angle": 0, + "content": "\\[\np _ {i, j} = \\operatorname {L L M} \\left(c _ {i}, T, \\omega_ {j}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.174, + 0.483, + 0.204 + ], + "angle": 0, + "content": "where \\(\\omega_{j}\\) represents a random seed or instruction to control the variation generated by the LLM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.204, + 0.484, + 0.298 + ], + "angle": 0, + "content": "Image Synthesis. We employ a public pre-trained Stable Diffusion model, denoted as \\(\\mathrm{SD}(\\cdot)\\), to generate images corresponding to the crafted prompts. For each prompt \\(p_{i,j}\\) we synthesize one image, resulting in a set of synthetic images \\(\\mathbb{X}_i = \\{\\pmb{x}_{i,j}\\}\\), where \\(\\pmb{x}_{i,j} = \\mathrm{SD}(p_{i,j})\\). The complete synthetic dataset is then \\(\\mathbb{X} = \\bigcup_{i}\\mathbb{X}_{i}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.302, + 0.362, + 0.318 + ], + "angle": 0, + "content": "3.2.2. Efficient Querying and Training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.32, + 0.484, + 0.381 + ], + "angle": 0, + "content": "This stage focuses on strategically querying the victim model to refine the substitute model's LoRA parameters while minimizing the number of queries. We propose two attack strategies: random learning and DSL." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.381, + 0.484, + 0.457 + ], + "angle": 0, + "content": "Random Learning (StolenLoRA-Rand). This straightforward approach directly uses the synthesized dataset \\(\\mathbb{X}\\) to query the victim model \\(F\\). Then, the substitute model's LoRA parameters \\(\\Delta \\theta^{\\prime}\\) are trained by minimizing the following objective function:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.467, + 0.483, + 0.485 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\prime} (\\Delta \\theta^ {\\prime}) = \\mathbb {E} _ {\\boldsymbol {x} \\in \\mathbb {X}} \\mathcal {L} \\left(F ^ {\\prime} (\\boldsymbol {x}), F (\\boldsymbol {x})\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.494, + 0.483, + 0.539 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}\\) denotes a loss function (e.g., cross-entropy loss). This provides a baseline for evaluating the benefits of more sophisticated attack strategies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.484, + 0.718 + ], + "angle": 0, + "content": "Disagreement-based Semi-supervised Learning (StolenLoRA-DSL). While directly using synthesized data for querying and training is feasible, we propose DSL to significantly enhance both the effectiveness and efficiency of the attack. DSL improves query efficiency by selectively querying the victim model based on the substitute model's uncertainty. Furthermore, it iteratively refines both the pseudo-labels of the synthetic data and the labels obtained from querying the victim, mitigating the impact of noisy pseudo-labels and bridging the gap between the substitute and victim model, especially in cross-backbone scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.904 + ], + "angle": 0, + "content": "The process begins by generating an initial synthetic dataset \\(\\mathbb{X}^0 = \\bigcup_i\\mathbb{X}_i^0\\) and assigning pseudo-labels based on their generating prompts without querying the victim model. These pseudo-labels, denoted as \\(c(x)\\), provide a starting point for training the initial substitute model \\(F_0'\\). Then, DSL proceeds iteratively, refining both the synthetic data for querying and the substitute model. In each iteration \\(t\\), a new set \\(\\mathbb{X}_{cand}^{t} = \\bigcup_{i}\\mathbb{X}_{cand,i}^{t}\\) with \\(\\beta * b_{t}\\) samples is generated using the LLM-driven Stable Diffusion process, where \\(\\beta\\) is a scaling factor. These candidates are then subjected to a disagreement-based filtering based on the current substitute model's predictions. For each candidate \\(x \\in \\mathbb{X}_{cand}^{t}\\)," + }, + { + "type": "code_caption", + "bbox": [ + 0.516, + 0.092, + 0.728, + 0.106 + ], + "angle": 0, + "content": "Algorithm 1 StolenLoRA-DSL" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.108, + 0.907, + 0.138 + ], + "angle": 0, + "content": "Input: Victim model \\(F\\), Class names \\(\\mathbb{C}\\), Query budget \\(b_{t}\\), Scaling factor \\(\\beta\\), Threshold \\(\\tau\\), Initial sample size \\(N\\)." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.138, + 0.76, + 0.153 + ], + "angle": 0, + "content": "Output: Substitute LoRA model \\(F^{\\prime}\\)" + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.156, + 0.716, + 0.171 + ], + "angle": 0, + "content": "1: \\(\\mathbb{X}_0\\gets \\bigcup_i\\mathrm{SynIMG}(\\mathbb{C},N)\\)" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.171, + 0.905, + 0.186 + ], + "angle": 0, + "content": "2: \\(F_0^{\\prime}\\gets \\mathrm{Update}(\\mathbb{X}_0,\\mathcal{L}_{LR})\\) 1 Initial substitute model" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.187, + 0.652, + 0.2 + ], + "angle": 0, + "content": "3: \\(t\\gets 0,B_t\\gets B\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.202, + 0.657, + 0.216 + ], + "angle": 0, + "content": "4: while \\( B_{t} > 0 \\) do" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.156, + 0.905, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.217, + 0.905, + 0.231 + ], + "angle": 0, + "content": "5: \\(\\mathbb{X}_{\\text{cand}}^t \\gets \\text{SynIMG}(\\mathbb{C}, \\beta b_t)\\) Generate candidate" + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.232, + 0.734, + 0.248 + ], + "angle": 0, + "content": "6: \\(\\mathbb{X}_{conf}^t\\gets \\emptyset ,\\mathbb{X}_{uncer}^t\\gets \\emptyset\\)" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.248, + 0.689, + 0.264 + ], + "angle": 0, + "content": "7: for \\(\\pmb{x} \\in \\mathbb{X}_{\\text{cand}}^t\\) do" + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.264, + 0.742, + 0.279 + ], + "angle": 0, + "content": "8: \\(\\hat{c},\\hat{p}\\gets \\operatorname {Predict}(F_t'(x))\\)" + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.279, + 0.905, + 0.293 + ], + "angle": 0, + "content": "9: if \\(\\hat{c} = c(\\pmb{x})\\) and \\(\\hat{p} \\geq \\tau\\) then Disagreement filtering" + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.293, + 0.821, + 0.311 + ], + "angle": 0, + "content": "10: \\(\\mathbb{X}_{conf}^t\\gets \\mathbb{X}_{conf}^t\\cup \\{(x,c(x))\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.311, + 0.621, + 0.321 + ], + "angle": 0, + "content": "11: else" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.323, + 0.782, + 0.339 + ], + "angle": 0, + "content": "12: \\(\\mathbb{X}_{uncer}^t\\gets \\mathbb{X}_{uncer}^t\\cup \\{\\pmb {x}\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.34, + 0.636, + 0.353 + ], + "angle": 0, + "content": "13: end if" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.355, + 0.621, + 0.367 + ], + "angle": 0, + "content": "14: end for" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.368, + 0.905, + 0.385 + ], + "angle": 0, + "content": "15: \\(\\mathbb{X}_{query}^t\\gets \\mathrm{Query}(F,\\mathbb{X}_{uncer}^t,b_t)\\) \\(\\triangleright\\) Selective Query" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.385, + 0.744, + 0.401 + ], + "angle": 0, + "content": "16: \\(\\mathbb{X}_{train}^{t}\\leftarrow \\mathbb{X}_{conf}^{t}\\cup \\mathbb{X}_{query}^{t}\\)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.401, + 0.77, + 0.417 + ], + "angle": 0, + "content": "17: \\(F_{t + 1}^{\\prime}\\gets \\mathrm{Update}(\\mathbb{X}_{train}^{t},\\mathcal{L}_{LR})\\)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.417, + 0.78, + 0.433 + ], + "angle": 0, + "content": "18: \\(B_{t}\\gets B_{t} - |\\mathbb{X}_{query}^{t}|,t\\gets t + 1\\)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.433, + 0.613, + 0.444 + ], + "angle": 0, + "content": "19: end while" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.446, + 0.612, + 0.461 + ], + "angle": 0, + "content": "20: return \\( F_{t}^{\\prime} \\)" + }, + { + "type": "list", + "bbox": [ + 0.521, + 0.156, + 0.905, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.907, + 0.617 + ], + "angle": 0, + "content": "the substitute model \\(F_{t}^{\\prime}\\) predicts a class \\(\\hat{c}\\) and associated confidence \\(\\hat{p}\\). If the predicted class \\(\\hat{c}\\) matches the prompt-based pseudo-label \\(c(\\boldsymbol{x})\\) and the confidence \\(\\hat{p}\\) exceeds a predefined threshold \\(\\tau\\), the pseudo-label is considered reliable, and the sample is added to a confidently labeled set \\(\\mathbb{X}_{conf}^{t} = \\{(\\boldsymbol{x}, c(\\boldsymbol{x}))\\}\\). Conversely, if the prediction disagrees with the pseudo-label or the confidence is low, the sample is added to an uncertain set \\(\\mathbb{X}_{uncer}^{t}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.617, + 0.906, + 0.771 + ], + "angle": 0, + "content": "From the uncertain set \\(\\mathbb{X}_{uncer}^{t}\\), a subset of \\(b_{t}\\) samples with the lowest confidence scores are selected for querying the victim model \\(F\\). These queried samples, along with their true labels obtained from \\(F\\), form the query set \\(\\mathbb{X}_{query}^{t} = \\{(\\boldsymbol{x}, F(\\boldsymbol{x}))\\}\\). Combine the confidently pseudolabeled samples \\(\\mathbb{X}_{conf}^{t}\\) and the queried samples \\(\\mathbb{X}_{query}^{t}\\) to form the training set for this iteration: \\(\\mathbb{X}_{train}^{t} = \\mathbb{X}_{conf}^{t} \\cup \\mathbb{X}_{query}^{t}\\). Crucially, this combined set undergoes label refining during training to further address potential inaccuracies in pseudo-labels and mitigate the distribution shift." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.772, + 0.906, + 0.818 + ], + "angle": 0, + "content": "Specifically, we employ a label refining strategy. Let \\( z \\) be the logits produced by the substitute model for a given input \\( x \\). The label refining training loss is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.832, + 0.905, + 0.874 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {L R} (\\boldsymbol {z}, \\boldsymbol {q}) = - \\sum_ {i = 1} ^ {C} q _ {i} \\log \\left(\\frac {\\exp (z _ {i})}{\\sum_ {j = 1} ^ {C} \\exp (z _ {j})}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "where \\( \\pmb{q} \\) are the soft labels. They are updated using an ex" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "266" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.09, + 0.904, + 0.33 + ], + "angle": 0, + "content": "
ScenarioMethodCUBS200Caltech256Indoor67Food101Flowers102
AccASRAccASRAccASRAccASRAccASR
Identical BackboneKnockoffNets70.9580.5485.6990.7179.1892.5982.5390.6276.2477.28
ActiveThief72.3382.1186.9292.0177.4690.5780.3488.2277.5178.56
DFME0.560.640.480.512.713.171.621.782.812.85
E371.9481.6782.3687.1881.4395.2279.6587.4680.6781.77
StolenLoRA-Rand75.3585.5487.6292.7582.3896.3379.0086.7593.7495.01
StolenLoRA-DSL73.2383.1389.3094.5382.6196.6080.5788.4787.4688.65
Cross BackboneKnockoffNets6.777.6941.3443.7641.7948.8714.4715.8914.1614.35
ActiveThief15.4217.5042.8945.4036.0442.1415.1116.5929.0929.49
DFME0.510.580.500.531.331.560.890.981.461.48
E317.5519.9248.1750.9955.8565.3140.5844.5648.2848.94
StolenLoRA-Rand45.7051.8851.7554.7859.1869.2043.1647.3959.2960.10
StolenLoRA-DSL50.1456.9265.0168.8265.0776.0944.5348.9061.1661.99
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.34, + 0.907, + 0.371 + ], + "angle": 0, + "content": "Table 1. Effectiveness of StolenLoRA: the Acc (%) and ASR (%) of extracted substitute model by different model extraction attacks under 10k queries. (Boldface: the best value, Underline: the second-best value.)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.378, + 0.479, + 0.394 + ], + "angle": 0, + "content": "ponential moving average of the predicted probabilities \\( \\pmb{p} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.403, + 0.483, + 0.423 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {q} ^ {(i + 1)} = \\mu \\boldsymbol {q} ^ {(i)} + (1 - \\mu) \\boldsymbol {p} ^ {(i + 1)}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.434, + 0.485, + 0.512 + ], + "angle": 0, + "content": "where \\(\\mu\\) is the momentum parameter. This iterative refinement allows the model to learn from both the initial labels and its own evolving predictions, progressively improving the quality of the training data. The updated substitute model \\(F_{t + 1}^{\\prime}\\) is thus trained on \\(\\mathbb{X}_{train}^{t}\\) by minimizing:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.522, + 0.483, + 0.543 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\prime} \\left(\\Delta \\theta_ {t + 1} ^ {\\prime}\\right) = \\mathbb {E} _ {\\boldsymbol {x} \\in \\mathbb {X} _ {t r a i n} ^ {t}} \\mathcal {L} _ {L R} \\left(F _ {t + 1} ^ {\\prime} (\\boldsymbol {x}), \\boldsymbol {q} (\\boldsymbol {x})\\right). \\quad (7)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.484, + 0.66 + ], + "angle": 0, + "content": "This iterative process continues until the query budget is exhausted, resulting in a final substitute model \\( F' \\) trained on a combination of high-confidence pseudo-labeled data and a smaller set of strategically queried data. DSL allows for efficient use of the query budget by focusing on the most informative samples, leading to a more accurate substitute LoRA. We present the algorithm detail of it in Alg. 1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.673, + 0.226, + 0.691 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.697, + 0.283, + 0.714 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Victim Models. We use a ViT-Base model pre-trained on ImageNet-21k [33] with additional augmentations and regularization [38]. This base model is then fine-tuned with LoRA (\\(r = 4\\)) on five commonly used datasets: CUBS200 [41], Caltech256 [9], Indoor67 [31], Food101 [2], and Flowers102 [26]. The resulting victim models achieve test accuracies of \\(88.09\\%\\), \\(94.47\\%\\), \\(85.52\\%\\), \\(91.07\\%\\), and \\(98.66\\%\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Attack Settings. We evaluate LoRA extraction attacks in Identical-Backbone (IB) and Cross-Backbone (XB) scenarios. In the IB setting, we use the same base model as the substitute base model. In the XB setting, we use" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.378, + 0.907, + 0.47 + ], + "angle": 0, + "content": "the ViT-base model pre-trained on ImageNet-1k[35] with self-supervised masked autoencoder (MAE) method[12] as the substitute base model. Both substitute models utilize LoRA with \\( r = 4 \\) and are trained for 20 epochs using the Adam optimizer[19] with a cosine annealing learning rate schedule[24] and a base learning rate of 0.01." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.496, + 0.909, + 0.664 + ], + "angle": 0, + "content": "We use GPT-4o mini[27] as the LLM in StolenLoRA as it has similar performance at a lower price and can effectively reduce the attack cost. For the SD model, we use an open-source SDXL-Turbo model[36]. It only takes 4 sampling steps (compared to SDXL 1.0 which usually requires 40 steps) to synthesize high-quality images, significantly reducing the computational cost. For the StolenLoRA-DSL method, initial per-category sample size \\(N\\) is set to 10, the scaling factor \\(\\beta\\) is 1.5 to not increase too much generation overhead, and the confidence threshold \\(\\tau\\) is 0.95. Hyperparameter analysis is provided in later sections." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Baselines and Evaluation Metric. We compare Stolen-LoRA against KnockoffNets [28], ActiveThief [29], DFME [40], and \\( E^3 \\) [48] using a query budget of 10k. For methods such as KnockoffNets, ActiveThief, and \\( E^3 \\) that require real data, we use 3M images in CC3M[37] as the attack dataset. For \\( E^3 \\), which originally selects samples based on semantic similarity between dataset category names and target class names, we improve its effectiveness by calculating semantic similarity between image captions (available in CC3M) and target class names for a more fine-grained sample selection. Following prior work [28, 40], we report Test Accuracy (Acc) and Attack Success Rate (ASR), defined as the ratio of the substitute model's accuracy to the victim model's accuracy." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "267" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.912, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.241, + 0.908, + 0.27 + ], + "angle": 0, + "content": "Figure 2. Curves of the test accuracy versus the number of queries in IB (solid line) and XB (dashed line) scenarios respectively. The gray dashed line represents the victim models' test accuracy." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.277, + 0.295, + 0.294 + ], + "angle": 0, + "content": "4.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.306, + 0.485, + 0.639 + ], + "angle": 0, + "content": "Effectiveness of StolenLoRA. Tab. 1 presents the effectiveness of StolenLoRA, comparing both the random strategy (StolenLoRA-Rand) and DSL-based (StolenLoRADSL) variants against existing extraction attacks. Across both identical- and cross-backbone scenarios, StolenLoRA consistently achieves state-of-the-art performance. In the IB setting, StolenLoRA-Rand shows superior performance on CUBS200 (75.35%) and Flowers102 (93.74%), while StolenLoRA-DSL excels on Caltech256 (89.30%) and Indoor67 (82.61%), maintaining competitive results on other datasets. The XB setting, more challenging due to base model differences, highlights StolenLoRA's significant advantage over baselines, with StolenLoRA-DSL achieving the highest accuracy across all five datasets. This demonstrates DSL's effectiveness in refining synthetic data and improving attack efficiency. Fig. 2 further analyzes query efficiency, showing StolenLoRA-DSL's rapid convergence in the IB setting and consistent progress in the XB setting, outperforming other methods in both scenarios. This efficiency stems from DSL's ability to prioritize informative queries, focusing on uncertain samples, thereby maximizing information gain from each query." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Effectiveness under Hard-Label Scenario. We also evaluate StolenLoRA's performance when only hard labels (one-hot encoded predictions) are available, as is the case with some real-world APIs. Fig. 3 presents the results for both StolenLoRA-Rand and StolenLoRA-DSL in the IB and XB settings. In the IB setting, using hard labels leads to a minor performance decrease. For example, StolenLoRA-Rand's accuracy on CUBS200 drops from \\(75.35\\%\\) with soft labels to \\(67.00\\%\\) with hard labels. A similar trend is observed for other datasets and StolenLoRA-DSL. This suggests that the additional information provided by soft labels is beneficial in the IB setting. In the XB setting, the impact of hard labels is inconsistent. While a slight decrease is observed on some datasets, others, like Caltech256 and Indoor67, show a marginal increase in accuracy (from \\(51.75\\%\\) to \\(56.69\\%\\)). This unexpected improvement might be due to the hard labels acting as a regularizer, preventing the substi" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.275, + 0.907, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.438, + 0.907, + 0.575 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.583, + 0.907, + 0.638 + ], + "angle": 0, + "content": "Figure 3. Comparison of StolenLoRA performance using hard labels (one-hot) versus soft labels (probabilities) from the victim model. Results are shown for both Random and DSL in identical- and cross-backbone scenarios across five datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.909, + 0.703 + ], + "angle": 0, + "content": "tute from overfitting to the synthetic data, which is more prone to distributional shifts in the XB setting. Overall, StolenLoRA exhibits robustness to hard labels in both IB and XB scenarios, demonstrating its practical applicability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Ablation Study. We conduct an ablation study to analyze the contributions of the key components of StolenLoRA: the LLM-driven prompting and the DSL strategy, including the label refining loss. Tab. 2 presents the results on CUBS200 and Indoor67 datasets under both IB and XB settings with a 10k query budget. Removing the structured prompt template leads to a performance drop in both IB and XB settings, demonstrating the importance of providing structured visual information to the LLM. Removing the LLM entirely results in a more significant performance decrease, confirming the crucial role of the LLM in generating effective prompts for diverse and representative synthetic data. Furthermore, we evaluate different LLMs, observing that using" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "268" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.09, + 0.48, + 0.221 + ], + "angle": 0, + "content": "
MethodIBXB
CUBS200Indoor67CUBS200Indoor67
Random75.3582.3845.7059.18
- Template72.3380.9442.2154.49
- LLM70.6674.7039.2047.09
+ Llama-3.1-8B73.9681.0446.7458.43
+ GPT-4o76.3083.1349.1963.58
DSL73.2382.6150.1465.07
- LLR70.1780.9748.2164.10
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.232, + 0.483, + 0.274 + ], + "angle": 0, + "content": "Table 2. An ablation experiment showing the effectiveness of the two modules we designed on CUBS200 and Indoor67 datasets under both IB and XB settings with 10k queries." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.279, + 0.48, + 0.4 + ], + "angle": 0, + "content": "
τIBXB
CUBS200Indoor67CUBS200Indoor67
0.570.6878.8148.4360.90
0.771.3780.0048.3663.51
0.974.0481.1948.5263.06
0.9573.2382.6150.1465.07
0.9973.3181.1949.8661.72
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.41, + 0.483, + 0.453 + ], + "angle": 0, + "content": "Table 3. Performance of StolenLoRA-DSL with varying confidence thresholds \\( \\tau \\) . Results are reported on CUBS200 and Indoor67 under both IB and XB settings with 10k queries." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.483, + 0.522 + ], + "angle": 0, + "content": "a more powerful LLM like GPT-4o generally yields better performance than Llama-3.1-8B[5], particularly in the more challenging XB setting. This highlights the benefit of utilizing a stronger LLM for prompt generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.523, + 0.483, + 0.644 + ], + "angle": 0, + "content": "Disabling DSL and reverting to random querying results in a noticeable performance reduction, especially in the XB setting. This underscores the effectiveness of DSL in selectively querying informative samples. Similarly, removing SAT from the DSL framework also leads to a performance decrease, albeit less pronounced than removing DSL altogether. This indicates the benefit of label refining in mitigating the distribution shift between synthetic and real data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.795 + ], + "angle": 0, + "content": "We also analyze the impact of the confidence threshold \\(\\tau\\) within DSL in Tab. 3. The results show that performance generally improves as \\(\\tau\\) increases from 0.5 to 0.95. This indicates that focusing queries on samples where the substitute model is less confident leads to a more efficient use of the query budget. However, increasing \\(\\tau\\) further to 0.99 results in a slight performance decrease, suggesting that an overly stringent threshold can exclude valuable samples from being queried. We therefore select \\(\\tau = 0.95\\) as the optimal value for our experiments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Fidelity and Distribution Analysis of Synthetic Data. StolenLoRA's effectiveness relies on the quality of its synthetic training data. We evaluate this quality both visually and quantitatively. Visually, as shown in Fig. 4, synthetic images generated by StolenLoRA demonstrate a striking resemblance to real images from the target categories, capturing key features and background context crucial for train-" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.089, + 0.898, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.63, + 0.224, + 0.792, + 0.236 + ], + "angle": 0, + "content": "(a) Visualization on CUBS200" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.237, + 0.898, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.373, + 0.789, + 0.384 + ], + "angle": 0, + "content": "(b) Visualization on Indoor67" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.387, + 0.905, + 0.415 + ], + "angle": 0, + "content": "Figure 4. Visualization of StolenLoRA's synthetic images and corresponding real images of the target category." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.42, + 0.903, + 0.496 + ], + "angle": 0, + "content": "
MethodCUBS200Caltech256Indoor67Food101Flowers102
KnockoffNets51.448.9422.0150.0455.77
ActiveThief49.336.0115.7247.8450.19
DFME220.53179.45225.84188.29204.79
E341.574.296.6825.5134.82
StolenLoRA2.145.596.9319.5416.63
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.498, + 0.905, + 0.527 + ], + "angle": 0, + "content": "Table 4. FID scores \\((\\downarrow)\\) between the datasets used by each attack method and the victim model's training data distribution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.907, + 0.688 + ], + "angle": 0, + "content": "ing. This visual fidelity is corroborated by significantly lower Fréchet Inception Distance (FID) scores[13] (in Tab. 4) compared to other attack methods. For example, StolenLoRA achieves an FID of 2.14 compared to 51.44 for KnockoffNets on CUBS200, highlighting a much closer distributional match to the target domain. This demonstrates that StolenLoRA's LLM-driven synthesis effectively generates high-fidelity, domain-specific data, even without access to the original training set, which is key to its superior extraction performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Verifying the Distinction Hypothesis. As mentioned in Sec. 3.1, we hypothesize that the performance of a perfectly extracted surrogate model will be consistent across In-Distribution (ID) data with the victim model, but not necessarily Out-of-Distribution (OOD) data, motivating us to perform extraction attacks using ID data. To empirically verify this conjecture, we train a substitute model using the XB setting on the victim model's training dataset, aiming to achieve a near-perfect extraction of the LoRA adaptations. We then evaluate both the victim model and the extracted substitute on two sets of data: ID data, represented by the victim model's held-out test set, and OOD data, sampled from the CC3M dataset. For each sample, we calculate the cross-entropy between the victim model's predictions and" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "269" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.087, + 0.482, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.244, + 0.485, + 0.302 + ], + "angle": 0, + "content": "Figure 5. Cross-entropy difference between the victim model and a near-perfectly substitute model on In-Distribution (ID) and Out-of-Distribution (OOD) data. The substitute model is trained using the cross-backbone setting on the victim's training dataset." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.309, + 0.485, + 0.506 + ], + "angle": 0, + "content": "the substitute model's predictions, providing a measure of their functional divergence. Fig. 5 presents the distribution of these differences. On both CUBS200 and Indoor67, the difference is smaller for the ID data compared to the OOD data. This observation supports our conjecture that even near-perfect LoRA extraction yields a substitute model that exhibits functional equivalence primarily within the ID domain. The divergence in OOD data likely stems from the inherent differences between the pre-trained backbone used by the victim and the substitute. This result underscores the importance of utilizing ID data for effective LoRA extraction, as the attacker primarily seeks to replicate the victim's specialized adaptations rather than the general capabilities." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.519, + 0.421, + 0.536 + ], + "angle": 0, + "content": "5. Defending Against LoRA Extraction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.543, + 0.483, + 0.603 + ], + "angle": 0, + "content": "While this paper primarily focuses on the feasibility of LoRA extraction, we also explore a potential defense mechanism based on deploying multiple, diverse LoRA adapters to increase attacker uncertainty." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.604, + 0.483, + 0.726 + ], + "angle": 0, + "content": "Dual LoRA with Diversified Predictions. Our proposed defense involves training two LoRA adapters, denoted as \\(\\mathrm{L}_A\\) and \\(\\mathrm{L}_B\\), to maximize the difference in their output distributions while maintaining comparable performance on the target task. Instead of minimizing accuracy as initially conceived, we found that maintaining high accuracy while maximizing divergence is crucial for practical deployment. We achieve this by optimizing the following loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.737, + 0.483, + 0.755 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\prime} = \\mathcal {L} (\\mathrm {L} _ {A}) + \\mathcal {L} (\\mathrm {L} _ {B}) - \\lambda \\times \\mathrm {K L} (\\mathrm {L} _ {A} | | \\mathrm {L} _ {B}) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}\\) represents the cross-entropy loss for the target task, KL represents the Kullback-Leibler divergence between the output distributions of the two LoRAs, and \\(\\lambda\\) is a hyperparameter controlling the balance between task performance and divergence. At deployment, we randomly select either \\(L_{A}\\) or \\(L_{B}\\) for each incoming query to generate the prediction. This random selection introduces uncertainty and makes it challenging for the attacker to extract the underlying LoRA function consistently." + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.088, + 0.905, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.244, + 0.905, + 0.285 + ], + "angle": 0, + "content": "Figure 6. Defneder Acc versus Substitute model Acc trade-off for defenses evaluated against StolenLoRA and KnockoffNets on the CUBS200 dataset with 10k queries." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.286, + 0.907, + 0.558 + ], + "angle": 0, + "content": "Experimental Evaluation. We evaluate the proposed defense mechanism under both IB and XB settings with 10k queries. Fig. 6 illustrates the trade-off between defender Acc and substitute model Acc. The defense consistently degrades the substitute model's performance. KnockoffNets experiences substantial accuracy drops (from \\(70.90\\%\\) to \\(20.78\\%\\) in IB and from \\(6.77\\%\\) to \\(1.21\\%\\) in XB), and StolenLoRA-Rand shows similar vulnerability (decreasing \\(22.95\\%\\) in IB and \\(15.63\\%\\) in XB). However, StolenLoRADSL demonstrates greater resilience, with comparatively smaller accuracy reductions (\\(9.63\\%\\) in IB and \\(14.34\\%\\) in XB). This resilience is attributed to DSL's efficient query selection and label refining, focusing on informative samples and mitigating the uncertainty introduced by the defense. While this defense effectively hinders less sophisticated attacks, it underscores the need for more robust mechanisms to counter advanced techniques like DSL that can learn effectively from limited queries." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.573, + 0.634, + 0.589 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This paper introduces LoRA Extraction, a novel model extraction attack targeting the widespread practice of adapting large vision models with LoRA. We present StolenLoRA, a highly effective method leveraging LLM-driven Stable Diffusion to generate task-specific, high-fidelity synthetic training data, bypassing the limitations of traditional extraction techniques reliant on large-scale real datasets or unreliable GANs. Furthermore, StolenLoRA incorporates Disagreement-based Semi-supervised Learning (DSL) to efficiently query the victim model, focusing on uncertain predictions and iteratively refining labels, enabling successful extraction with limited queries. Extensive experiments demonstrate StolenLoRA's strong performance across diverse datasets and in challenging cross-backbone scenarios. We also explore a preliminary defense based on diversified LoRA deployments, showing promise in mitigating these attacks. This work represents a crucial first step towards understanding and addressing the safety risks associated with LoRA adaptation, paving the way for more secure deployments of efficient fine-tuning methods." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "270" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.114, + 0.486, + 0.173 + ], + "angle": 0, + "content": "This work is in part supported by National Key R&D Program of China (Grant No. 2022ZD0160103) and National Natural Science Foundation of China (Grant No. 62276067), and Shanghai Artificial Intelligence Laboratory." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.198, + 0.188, + 0.214 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.223, + 0.484, + 0.265 + ], + "angle": 0, + "content": "[1] Martin Arjovsky and Léon Bottou. Towards principled methods for training generative adversarial networks. arXiv preprint arXiv:1701.04862, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.266, + 0.484, + 0.307 + ], + "angle": 0, + "content": "[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In ECCV, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.309, + 0.484, + 0.363 + ], + "angle": 0, + "content": "[3] Eric L Buehler and Markus J Buehler. X-lora: Mixture of low-rank adapter experts, a flexible framework for large language models with applications in protein mechanics and molecular design. APL Machine Learning, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.365, + 0.483, + 0.405 + ], + "angle": 0, + "content": "[4] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.407, + 0.484, + 0.475 + ], + "angle": 0, + "content": "[5] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.478, + 0.484, + 0.532 + ], + "angle": 0, + "content": "[6] Ricard Durall, Avraam Chatzimichailidis, Peter Labus, and Janis Keuper. Combating mode collapse in gan training: An empirical analysis using hessian eigenvalues. arXiv preprint arXiv:2012.09673, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.534, + 0.484, + 0.576 + ], + "angle": 0, + "content": "[7] Lijie Fan, Kaifeng Chen, Dilip Krishnan, Dina Katabi, Phillip Isola, and Yonglong Tian. Scaling laws of synthetic images for model training... for now. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.577, + 0.484, + 0.63 + ], + "angle": 0, + "content": "[8] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NIPS, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.633, + 0.484, + 0.66 + ], + "angle": 0, + "content": "[9] Gregory Griffin, Alex Holub, and Pietro Perona. Caltech-256 object category dataset. 2007. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.662, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[10] Zeyu Han, Chao Gao, Jinyang Liu, Sai Qian Zhang, et al. Parameter-efficient fine-tuning for large models: A comprehensive survey. arXiv preprint arXiv:2403.14608, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.704, + 0.484, + 0.745 + ], + "angle": 0, + "content": "[11] Soufiane Hayou, Nikhil Ghosh, and Bin Yu. Lora+: Efficient low rank adaptation of large models. arXiv preprint arXiv:2402.12354, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[12] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In ICML, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.223, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.148 + ], + "angle": 0, + "content": "[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.906, + 0.204 + ], + "angle": 0, + "content": "[16] Nam Hyeon-Woo, Moon Ye-Bin, and Tae-Hyun Oh. Fedpara: Low-rank hadamard product for communication-efficient federated learning. arXiv preprint arXiv:2108.06098, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.906, + 0.26 + ], + "angle": 0, + "content": "[17] Matthew Jagielski, Nicholas Carlini, David Berthelot, Alex Kurakin, and Nicolas Papernot. High accuracy and high fidelity extraction of neural networks. In *Usenix Security*, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[18] Sanjay Kariyappa, Atul Prakash, and Moinuddin Qureshi. Maze: Data-free model stealing attack using zeroth-order gradient estimation. In CVPR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.333 + ], + "angle": 0, + "content": "[19] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.335, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[20] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[21] Youssef Kossale, Mohammed Airaj, and Aziz Darouichi. Mode collapse in generative adversarial networks: An overview. In ICOA, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[22] Zi Liang, Qingqing Ye, Yanyun Wang, Sen Zhang, Yaxin Xiao, Ronghua Li, Jianliang Xu, and Haibo Hu. Alignment-aware model extraction attacks on large language models. arXiv preprint arXiv:2409.02718, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[23] Zijun Lin, Ke Xu, Chengfang Fang, Huadi Zheng, Aneez Ahmed Jaheezuddin, and Jie Shi. Quda: Query-limited data-free model extraction. In ACM Asia CCS, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.534, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[24] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In *ICLR*, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.563, + 0.906, + 0.63 + ], + "angle": 0, + "content": "[25] Xingjun Ma, Yifeng Gao, Yixu Wang, Ruofan Wang, Xin Wang, Ye Sun, Yifan Ding, Hengyuan Xu, Yunhao Chen, Yunhan Zhao, et al. Safety at scale: A comprehensive survey of large model safety. arXiv preprint arXiv:2502.05206, 2025.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.906, + 0.688 + ], + "angle": 0, + "content": "[26] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian conference on computer vision, graphics & image processing, 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.69, + 0.906, + 0.744 + ], + "angle": 0, + "content": "[27] OpenAI. Gpt-4o mini: Advancing cost-efficient intelligence. https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence/, 2024.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.747, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[28] Tribhuvanesh Orekondy, Bernt Schiele, and Mario Fritz. Knockoff nets: Stealing functionality of black-box models. In CVPR, 2019. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.906, + 0.844 + ], + "angle": 0, + "content": "[29] Soham Pal, Yash Gupta, Aditya Shukla, Aditya Kanade, Shirish Shevade, and Vinod Ganapathy. Activethief: Model extraction using active learning and unannotated public data. In AAAI, 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[30] Jie Qin, Jie Wu, Weifeng Chen, Yuxi Ren, Huixia Li, Hefeng Wu, Xuefeng Xiao, Rui Wang, and Shilei Wen. Diffusiongpt: Llm-driven text-to-image generation system. arXiv preprint arXiv:2401.10061, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "271" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.484, + 0.12 + ], + "angle": 0, + "content": "[31] Ariadna Quattoni and Antonio Torralba. Recognizing indoor scenes. In CVPR, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.122, + 0.484, + 0.176 + ], + "angle": 0, + "content": "[32] Pengzhen Ren, Yun Xiao, Xiaojun Chang, Po-Yao Huang, Zhihui Li, Brij B Gupta, Xiaojiang Chen, and Xin Wang. A survey of deep active learning. ACM computing surveys, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.179, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[33] Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik-Manor. Imagenet-21k pretraining for the masses. arXiv preprint arXiv:2104.10972, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.222, + 0.483, + 0.262 + ], + "angle": 0, + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.264, + 0.483, + 0.319 + ], + "angle": 0, + "content": "[35] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.321, + 0.483, + 0.36 + ], + "angle": 0, + "content": "[36] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.363, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[37] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.42, + 0.483, + 0.475 + ], + "angle": 0, + "content": "[38] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. arXiv preprint arXiv:2106.10270, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.477, + 0.483, + 0.518 + ], + "angle": 0, + "content": "[39] Florian Tramér, Fan Zhang, Ari Juels, Michael K. Reiter, and Thomas Ristenpart. Stealing machine learning models via prediction apis. In USENIX Security Symposium, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.52, + 0.483, + 0.559 + ], + "angle": 0, + "content": "[40] Jean-Baptiste Truong, Pratyush Maini, Robert J Walls, and Nicolas Papernot. Data-free model extraction. In CVPR, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.562, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[41] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.604, + 0.483, + 0.658 + ], + "angle": 0, + "content": "[42] Yixu Wang, Jie Li, Hong Liu, Yan Wang, Yongjian Wu, Feiyue Huang, and Rongrong Ji. Black-box dissector: Towards erasing-based hard-label model stealing attack. In ECCV, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.661, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[43] Yixu Wang, Tianle Gu, Yan Teng, Yingchun Wang, and Xingjun Ma. Honeypotnet: Backdoor attacks against model extraction. In AAAI, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.704, + 0.483, + 0.758 + ], + "angle": 0, + "content": "[44] Shih-Ying Yeh, Yu-Guan Hsieh, Zhidong Gao, Bernard BW Yang, Giyeong Oh, and Yanmin Gong. Navigating text-to-image customization: From lycoris fine-tuning to model evaluation. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[45] Jianhao Yuan, Jie Zhang, Shuyang Sun, Philip Torr, and Bo Zhao. Real-fake: Effective training data synthesis through distribution matching. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[46] Yunlong Zhao, Xiaoheng Deng, Yijing Liu, Xinjun Pei, Ji-azhi Xia, and Wei Chen. Fully exploiting every real sample: Superpixel sample gradient model stealing. In CVPR, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[47] Mingyi Zhou, Jing Wu, Yipeng Liu, Shuaicheng Liu, and Ce Zhu. Dast: Data-free substitute training for adversarial attacks. In CVPR, 2020. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.906, + 0.135 + ], + "angle": 0, + "content": "[48] Hongyu Zhu, Wentao Hu, Sichu Liang, Fangqi Li, Wenwen Wang, and Shilin Wang. Efficient and effective model extraction. arXiv preprint arXiv:2409.14122, 2024. 5" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "272" + } + ] +] \ No newline at end of file diff --git a/2025/StolenLoRA_ Exploring LoRA Extraction Attacks via Synthetic Data/f4ddd460-9cf1-429e-89c2-3d861bf9c661_origin.pdf b/2025/StolenLoRA_ Exploring LoRA Extraction Attacks via Synthetic Data/f4ddd460-9cf1-429e-89c2-3d861bf9c661_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6496ad91a073db2b49c36348284b369346657d21 --- /dev/null +++ b/2025/StolenLoRA_ Exploring LoRA Extraction Attacks via Synthetic Data/f4ddd460-9cf1-429e-89c2-3d861bf9c661_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5747da3dc179a6db258417ceb575a0d2875885183cdd74b8f22f7f7d3e2c1a6d +size 1529920 diff --git a/2025/StolenLoRA_ Exploring LoRA Extraction Attacks via Synthetic Data/full.md b/2025/StolenLoRA_ Exploring LoRA Extraction Attacks via Synthetic Data/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9bba43f6961b318ddfc90af009312756b5bde4e3 --- /dev/null +++ b/2025/StolenLoRA_ Exploring LoRA Extraction Attacks via Synthetic Data/full.md @@ -0,0 +1,307 @@ +# StolenLoRA: Exploring LoRA Extraction Attacks via Synthetic Data + +Yixu Wang $^{1,2\dagger}$ , Yan Teng $^{2*}$ , Yingchun Wang $^{2}$ , Xingjun Ma $^{1}$ + +$^{1}$ Fudan University $^{2}$ Shanghai Artificial Intelligence Laboratory + +# Abstract + +Parameter-Efficient Fine-Tuning (PEFT) methods like LoRA have transformed vision model adaptation, enabling the rapid deployment of customized models. However, the compactness of LoRA adaptations introduces new safety concerns, particularly their vulnerability to model extraction attacks. This paper introduces a new focus of model extraction attacks named LoRA extraction that extracts LoRA-adaptive models based on a public pre-trained model. We then propose a novel extraction method called StolenLoRA which trains a substitute model to extract the functionality of a LoRA-adapted model using synthetic data. StolenLoRA leverages a Large Language Model to craft effective prompts for data generation, and it incorporates a Disagreement-based Semi-supervised Learning (DSL) strategy to maximize information gain from limited queries. Our experiments demonstrate the effectiveness of StolenLoRA, achieving up to a $96.60\%$ attack success rate with only 10k queries, even in cross-backbone scenarios where the attacker and victim models utilize different pretrained backbones. These findings reveal the specific vulnerability of LoRA-adapted models to this type of extraction and underscore the urgent need for robust defense mechanisms tailored to PEFT methods. We also explore a preliminary defense strategy based on diversified LoRA deployments, highlighting its potential to mitigate such attacks. + +# 1. Introduction + +Adapting large-scale pre-trained foundation models via Parameter-Efficient Fine-Tuning (PEFT) such as Low-Rank Adaptation (LoRA) [11, 15] has become a popular approach for obtaining high-performance models for diverse downstream tasks. LoRA enables efficient adaptation of large models to a new task by fine-tuning only low-rank matrices added to specific layers, reducing computational cost and memory usage while preserving the pre-trained model's + +core knowledge. Despite the wide adoption of LoRA, the lightweight and compactness of LoRA parameters raise a new safety concern: they might be more vulnerable to model extraction (ME) attacks. An ME attack extracts the functionality [17, 22, 28, 39] of a victim model by training a substitute model based on the outputs of the victim model for a certain number of specially designed queries. While ME attacks have been extensively studied for traditional models [29, 40, 42, 43, 46, 47], the vulnerability of LoRA adaptations to ME attacks remains largely unexplored[25]. This gap is significant, as the public availability of a large number of pre-trained models, coupled with the compactness of LoRA parameters, makes it easier to replicate LoRA-adapted models and compromise intellectual property. + +To address this gap, we introduce the concept of LoRA Extraction, a novel direction of ME attacks specifically targeting LoRA-adapted models (i.e., ViTs [4, 12, 20]), based on a publicly available pre-trained model. Unlike traditional ME which focuses on replicating the entire model's functionality, LoRA extraction centers on stealing the efficient adaptations encoded within the compact LoRA parameters. The adversary's objective is to reconstruct the victim's LoRA-adapted model by training their own LoRA-adapted substitute that achieves similar downstream performance using a pre-trained foundation model that is publicly available. This attack can manifest in two scenarios: the identical-backbone scenario, where the attacker uses the same pre-trained ViT model as the victim, and the more challenging cross-backbone scenario, where the attacker uses a different pre-trained ViT. + +Existing traditional ME methods largely depend on sample selection from existing datasets or synthetic data generation. Sample selection methods [28, 29, 42, 46] typically require searching extensive datasets for in-distribution samples, making them computationally expensive and less practical for LoRA extraction. This impracticality arises because 1) inference in LoRA-adapted models is slower due to the large-scale parameters of LoRA-adapted models, and 2) finding appropriate samples is challenging, particularly for domain-specific data fine-tuned with LoRA. Alternatively, synthetic data generation approaches [18, 23, 40, 47] often rely on Generative Adversarial Networks (GANs) [8] + +to produce in-distribution data. However, GANs frequently face difficulties in generating high-quality, diverse samples, especially for high-dimensional data. As Zhao et al. [46] noted, creating high-dimensional samples (e.g., $224 \times 224$ images) for effective ME can require millions of queries and is prone to failure due to GAN mode collapse [1, 6, 21]. These limitations make GAN-based approaches less effective for LoRA extraction, where the tuning often involves high-dimensional image data. + +In this work, we propose a novel LoRA extraction method named StolenLoRA, which employs synthetic data to train the substitute model. To obtain effective in-distribution synthetic data, we leverage Large Language Models (LLMs) to generate diverse textual descriptions based on the target class names and prompt a pre-trained Stable Diffusion model [34, 36] to generate high-quality images. This allows us to tailor the generated data to the specific downstream task the victim's LoRA model is fine-tuned for. Moreover, StolenLoRA introduces a distinctive strategy to improve attack efficiency, i.e., the Disagreement-based Semi-supervised Learning (DSL). DSL uses class information from the data synthesis process as pseudo-labels for part of the generated dataset and focuses on queries where the substitute model disagrees with these pseudo-labels. DSL effectively targets areas of uncertainty, guiding the substitute model's learning process and refining its alignment with the victim's behavior. To further enhance DSL, we iteratively refine the pseudo-labels based on evolving predictions from the substitute model, improving the quality of synthetic data labels and the effectiveness of the extraction. + +In summary, our contributions are as follows: + +- We explore the vulnerability of LoRA and introduce a novel focus of ME attacks called LoRA Extraction which aims to extract the functionality of LoRA-adapted models based on a publicly available pre-trained model. +- We present StolenLoRA, a novel LoRA extraction method that leverages LLM-driven Stable Diffusion to generate high-quality synthetic data, circumventing the need to search large-scale datasets or rely on less dependable GAN-based generation. +- Through comprehensive experiments on five widely used datasets, we demonstrate the effectiveness of Stolen-LoRA in both identical- and cross-backbone scenarios, highlighting critical vulnerabilities in LoRA-adapted models. We also explore a preliminary defense mechanism based on diversified LoRA deployments. + +# 2. Related Work + +Low-Rank Adaptation (LoRA). LoRA [15] is a type of popular Parameter-Efficient Fine-Tuning (PEFT) method [10, 14] for adapting large pre-trained models for downstream tasks. LoRA [15] introduces small rank- + +decomposition matrices as trainable parameters into the pre-trained model. The number of trainable parameters is significantly reduced by freezing the original model weights and updating only these lightweight LoRA modules. LoRA inspires numerous variants and extensions, such as Mixture-of-LoRA Experts (X-LoRA)[3], Low-Rank Hadamard Product (LoHa)[16], and Low-Rank Kronecker Product (LoKr)[44], further enhancing its efficiency and flexibility for various applications. However, the compact nature of these LoRA updates raises concerns about their vulnerability to model extraction attacks. This paper investigates the potential for adversaries to extract the efficient adaptations encoded within these compact parameters, specifically focusing on the standard LoRA method [15]. + +Model Extraction Attacks. Model Extraction (ME) attacks [18, 22, 23, 28, 29, 40, 42, 46, 47] aim to construct a substitute model that mimics the functionality of a victim model by querying its API. Existing ME methods can be broadly categorized based on their query generation strategies: Data-Driven ME leverage existing datasets [28, 29, 42, 46] to select queries, often employing active learning[32] or other search strategies[28] to identify informative samples. However, this type can incur computational overhead as it necessitates traversing the large-scale dataset to find suitable in-distribution samples. Data-Free ME [18, 23, 40, 46, 47] utilize generative models (e.g., GANs[8]) to synthesize data for querying the victim model. While promising, generating effective queries with high-dimensional inputs can require tens of millions of queries and is prone to failure due to GAN mode collapse [1, 6, 21]. + +Learning from Synthetic Data. The use of synthetic data for training models gains significant traction due to its potential to address data scarcity. While traditional synthetic data often lacked the complexity and representativeness of real data, recent advancements in generative models, particularly diffusion models like Stable Diffusion [34, 36], enable the synthesis of high-quality and diverse images. However, scaling the quantity of synthetic data does not necessarily translate to improved performance for training supervised image classifiers. As shown by Fan et al.[7], performance with synthetic data can lag significantly behind training with real data. This disparity stems from the limitations of current text-to-image models in accurately generating diverse representations of certain concepts at scale[30]. Efforts to mitigate this gap, such as the distribution-matching framework[45], focus on aligning the distributions of synthetic and real data. However, leveraging such high-quality synthetic data for model extraction attacks remains an unexplored area. + +![](images/ee8f6f3384583ba07b2dbe2a5d69d0ed57a8c92aea61e38c2c949168b526c2d6.jpg) +Figure 1. Overview of StolenLoRA. It collects target victim model information (dashed red box), synthesizes training data guided by an LLM-driven prompting (dashed blue box), and then efficiently queries the victim API to train a substitute model (dashed green box). + +# 3. Methodology + +# 3.1. Problem Formulation + +LoRA Extraction. Given a victim model $F \colon [0,1]^d \mapsto \mathbb{R}^N$ , which is a pre-trained model $F_{base}$ (e.g., ViTs) adapted using LoRA. The parameters of $F^1$ can be represented as $\theta = \{\theta_{base}, \Delta \theta\}$ where $\theta_{base}$ are the parameters of $F_{base}$ and $\Delta \theta$ represents the LoRA updates, which are typically low-rank matrices. The objective of a LoRA extraction attack is to train a substitute model $F'$ with parameters $\theta' = \{\theta_{base}', \Delta \theta'\}$ that mimics the functionality of $F$ . This is formulated as the following optimization problem: + +$$ +\arg \min _ {\Delta \theta^ {\prime}} \mathbb {E} _ {\boldsymbol {x}} \mathcal {D} (F (\boldsymbol {x}), F ^ {\prime} (\boldsymbol {x})), \tag {1} +$$ + +where $\pmb{x} \in \mathbb{R}^d$ represents an input sample from the victim model task distribution, $\mathcal{D}(\cdot, \cdot)$ represents a measure of functional similarity between the two models. + +Depending on whether the attacker has knowledge of $F_{base}$ , we distinguish between two scenarios: identical backbone and cross-backbone LoRA extractions. The substitute model $F'$ can be defined as follows: + +$$ +F ^ {\prime} (\boldsymbol {x}) = \left\{ \begin{array}{l l} F _ {\text {b a s e}} (\boldsymbol {x}) + \Delta F ^ {\prime} (\boldsymbol {x}), & \text {i d e n t i c a l - b a c k b o n e} \\ G _ {\text {b a s e}} (\boldsymbol {x}) + \Delta F ^ {\prime} (\boldsymbol {x}), & \text {c r o s s - b a c k b o n e} \end{array} \right. \tag {2} +$$ + +where the former uses the same $F_{base}$ , the latter uses different base models $G_{base}$ , and the $+$ symbol denotes the combination of the base model and the LoRA updates. + +Distinction from Traditional Extraction. Traditional model extraction trains a substitute from scratch, while LoRA extraction leverages a pre-trained base model, inheriting its knowledge and biases. Consequently, we hypothesize that a perfectly extracted LoRA substitute might + +achieve functional equivalence on In-Distribution (ID) data but not on Out-of-Distribution (OOD) data, particularly in the cross-backbone setting, and we experimentally verify this in the experimental section. This motivates the use of ID data for effective LoRA extraction. + +# 3.2. StolenLoRA + +Overview. An overview of our proposed LLM-driven generative attack for extracting LoRA, StolenLoRA, is illustrated in Fig. 1. It comprises two key stages: 1) Data Synthesis, where an LLM guides the generation of synthetic data tailored to the victim's task, and 2) Efficient Querying and Training, which strategically selects queries and refines labels to maximize information gain from the victim model. + +# 3.2.1. Data Synthesis + +The above discussion motivates us to use ID data for extraction. However, acquiring such datasets in real-world scenarios is often challenging because the victim model's training data is proprietary. While the attacker might not have access to the original training data, they often have access to the model's deployment context, which usually includes a description of the model's functionality and the names of the target classes. This information can be leveraged to semantically expand the search space for potential training data. Inspired by this, we propose to utilize LLM-driven Stable Diffusion to generate synthetic data based on the class names, effectively bridging the gap between semantic information and visual representations. + +LLM-Driven Prompting. Given a set of target class names $\mathbb{C} = \{c_1, c_2, \dots, c_n\}$ , we employ an LLM to generate detailed prompts for image synthesis. We structure the prompts using a predefined template $T$ comprising key visual elements: $T = \text{"[Subject, Background, Angle/ Pose, Lighting, Style]"}$ . For each class $c_i$ , the LLM generates $m$ variations of the + +prompt, denoted as $p_{i,j}$ , where $j \in \{1, \dots, m\}$ . This diversification ensures a richer representation of the target class. The prompt generation process is formalized as follows: + +$$ +p _ {i, j} = \operatorname {L L M} \left(c _ {i}, T, \omega_ {j}\right), \tag {3} +$$ + +where $\omega_{j}$ represents a random seed or instruction to control the variation generated by the LLM. + +Image Synthesis. We employ a public pre-trained Stable Diffusion model, denoted as $\mathrm{SD}(\cdot)$ , to generate images corresponding to the crafted prompts. For each prompt $p_{i,j}$ we synthesize one image, resulting in a set of synthetic images $\mathbb{X}_i = \{\pmb{x}_{i,j}\}$ , where $\pmb{x}_{i,j} = \mathrm{SD}(p_{i,j})$ . The complete synthetic dataset is then $\mathbb{X} = \bigcup_{i}\mathbb{X}_{i}$ . + +# 3.2.2. Efficient Querying and Training + +This stage focuses on strategically querying the victim model to refine the substitute model's LoRA parameters while minimizing the number of queries. We propose two attack strategies: random learning and DSL. + +Random Learning (StolenLoRA-Rand). This straightforward approach directly uses the synthesized dataset $\mathbb{X}$ to query the victim model $F$ . Then, the substitute model's LoRA parameters $\Delta \theta^{\prime}$ are trained by minimizing the following objective function: + +$$ +\mathcal {L} ^ {\prime} (\Delta \theta^ {\prime}) = \mathbb {E} _ {\boldsymbol {x} \in \mathbb {X}} \mathcal {L} \left(F ^ {\prime} (\boldsymbol {x}), F (\boldsymbol {x})\right), \tag {4} +$$ + +where $\mathcal{L}$ denotes a loss function (e.g., cross-entropy loss). This provides a baseline for evaluating the benefits of more sophisticated attack strategies. + +Disagreement-based Semi-supervised Learning (StolenLoRA-DSL). While directly using synthesized data for querying and training is feasible, we propose DSL to significantly enhance both the effectiveness and efficiency of the attack. DSL improves query efficiency by selectively querying the victim model based on the substitute model's uncertainty. Furthermore, it iteratively refines both the pseudo-labels of the synthetic data and the labels obtained from querying the victim, mitigating the impact of noisy pseudo-labels and bridging the gap between the substitute and victim model, especially in cross-backbone scenarios. + +The process begins by generating an initial synthetic dataset $\mathbb{X}^0 = \bigcup_i\mathbb{X}_i^0$ and assigning pseudo-labels based on their generating prompts without querying the victim model. These pseudo-labels, denoted as $c(x)$ , provide a starting point for training the initial substitute model $F_0'$ . Then, DSL proceeds iteratively, refining both the synthetic data for querying and the substitute model. In each iteration $t$ , a new set $\mathbb{X}_{cand}^{t} = \bigcup_{i}\mathbb{X}_{cand,i}^{t}$ with $\beta * b_{t}$ samples is generated using the LLM-driven Stable Diffusion process, where $\beta$ is a scaling factor. These candidates are then subjected to a disagreement-based filtering based on the current substitute model's predictions. For each candidate $x \in \mathbb{X}_{cand}^{t}$ , + +Algorithm 1 StolenLoRA-DSL + +Input: Victim model $F$ , Class names $\mathbb{C}$ , Query budget $b_{t}$ , Scaling factor $\beta$ , Threshold $\tau$ , Initial sample size $N$ . + +Output: Substitute LoRA model $F^{\prime}$ + +1: $\mathbb{X}_0\gets \bigcup_i\mathrm{SynIMG}(\mathbb{C},N)$ +2: $F_0^{\prime}\gets \mathrm{Update}(\mathbb{X}_0,\mathcal{L}_{LR})$ 1 Initial substitute model +3: $t\gets 0,B_t\gets B$ +4: while $B_{t} > 0$ do + +5: $\mathbb{X}_{\text{cand}}^t \gets \text{SynIMG}(\mathbb{C}, \beta b_t)$ Generate candidate +6: $\mathbb{X}_{conf}^t\gets \emptyset ,\mathbb{X}_{uncer}^t\gets \emptyset$ +7: for $\pmb{x} \in \mathbb{X}_{\text{cand}}^t$ do +8: $\hat{c},\hat{p}\gets \operatorname {Predict}(F_t'(x))$ +9: if $\hat{c} = c(\pmb{x})$ and $\hat{p} \geq \tau$ then Disagreement filtering +10: $\mathbb{X}_{conf}^t\gets \mathbb{X}_{conf}^t\cup \{(x,c(x))\}$ +11: else +12: $\mathbb{X}_{uncer}^t\gets \mathbb{X}_{uncer}^t\cup \{\pmb {x}\}$ +13: end if +14: end for +15: $\mathbb{X}_{query}^t\gets \mathrm{Query}(F,\mathbb{X}_{uncer}^t,b_t)$ $\triangleright$ Selective Query +16: $\mathbb{X}_{train}^{t}\leftarrow \mathbb{X}_{conf}^{t}\cup \mathbb{X}_{query}^{t}$ +17: $F_{t + 1}^{\prime}\gets \mathrm{Update}(\mathbb{X}_{train}^{t},\mathcal{L}_{LR})$ +18: $B_{t}\gets B_{t} - |\mathbb{X}_{query}^{t}|,t\gets t + 1$ +19: end while +20: return $F_{t}^{\prime}$ + +the substitute model $F_{t}^{\prime}$ predicts a class $\hat{c}$ and associated confidence $\hat{p}$ . If the predicted class $\hat{c}$ matches the prompt-based pseudo-label $c(\boldsymbol{x})$ and the confidence $\hat{p}$ exceeds a predefined threshold $\tau$ , the pseudo-label is considered reliable, and the sample is added to a confidently labeled set $\mathbb{X}_{conf}^{t} = \{(\boldsymbol{x}, c(\boldsymbol{x}))\}$ . Conversely, if the prediction disagrees with the pseudo-label or the confidence is low, the sample is added to an uncertain set $\mathbb{X}_{uncer}^{t}$ . + +From the uncertain set $\mathbb{X}_{uncer}^{t}$ , a subset of $b_{t}$ samples with the lowest confidence scores are selected for querying the victim model $F$ . These queried samples, along with their true labels obtained from $F$ , form the query set $\mathbb{X}_{query}^{t} = \{(\boldsymbol{x}, F(\boldsymbol{x}))\}$ . Combine the confidently pseudolabeled samples $\mathbb{X}_{conf}^{t}$ and the queried samples $\mathbb{X}_{query}^{t}$ to form the training set for this iteration: $\mathbb{X}_{train}^{t} = \mathbb{X}_{conf}^{t} \cup \mathbb{X}_{query}^{t}$ . Crucially, this combined set undergoes label refining during training to further address potential inaccuracies in pseudo-labels and mitigate the distribution shift. + +Specifically, we employ a label refining strategy. Let $z$ be the logits produced by the substitute model for a given input $x$ . The label refining training loss is calculated as: + +$$ +\mathcal {L} _ {L R} (\boldsymbol {z}, \boldsymbol {q}) = - \sum_ {i = 1} ^ {C} q _ {i} \log \left(\frac {\exp (z _ {i})}{\sum_ {j = 1} ^ {C} \exp (z _ {j})}\right), \tag {5} +$$ + +where $\pmb{q}$ are the soft labels. They are updated using an ex + +
ScenarioMethodCUBS200Caltech256Indoor67Food101Flowers102
AccASRAccASRAccASRAccASRAccASR
Identical BackboneKnockoffNets70.9580.5485.6990.7179.1892.5982.5390.6276.2477.28
ActiveThief72.3382.1186.9292.0177.4690.5780.3488.2277.5178.56
DFME0.560.640.480.512.713.171.621.782.812.85
E371.9481.6782.3687.1881.4395.2279.6587.4680.6781.77
StolenLoRA-Rand75.3585.5487.6292.7582.3896.3379.0086.7593.7495.01
StolenLoRA-DSL73.2383.1389.3094.5382.6196.6080.5788.4787.4688.65
Cross BackboneKnockoffNets6.777.6941.3443.7641.7948.8714.4715.8914.1614.35
ActiveThief15.4217.5042.8945.4036.0442.1415.1116.5929.0929.49
DFME0.510.580.500.531.331.560.890.981.461.48
E317.5519.9248.1750.9955.8565.3140.5844.5648.2848.94
StolenLoRA-Rand45.7051.8851.7554.7859.1869.2043.1647.3959.2960.10
StolenLoRA-DSL50.1456.9265.0168.8265.0776.0944.5348.9061.1661.99
+ +Table 1. Effectiveness of StolenLoRA: the Acc (%) and ASR (%) of extracted substitute model by different model extraction attacks under 10k queries. (Boldface: the best value, Underline: the second-best value.) + +ponential moving average of the predicted probabilities $\pmb{p}$ : + +$$ +\boldsymbol {q} ^ {(i + 1)} = \mu \boldsymbol {q} ^ {(i)} + (1 - \mu) \boldsymbol {p} ^ {(i + 1)}, \tag {6} +$$ + +where $\mu$ is the momentum parameter. This iterative refinement allows the model to learn from both the initial labels and its own evolving predictions, progressively improving the quality of the training data. The updated substitute model $F_{t + 1}^{\prime}$ is thus trained on $\mathbb{X}_{train}^{t}$ by minimizing: + +$$ +\mathcal {L} ^ {\prime} \left(\Delta \theta_ {t + 1} ^ {\prime}\right) = \mathbb {E} _ {\boldsymbol {x} \in \mathbb {X} _ {t r a i n} ^ {t}} \mathcal {L} _ {L R} \left(F _ {t + 1} ^ {\prime} (\boldsymbol {x}), \boldsymbol {q} (\boldsymbol {x})\right). \quad (7) +$$ + +This iterative process continues until the query budget is exhausted, resulting in a final substitute model $F'$ trained on a combination of high-confidence pseudo-labeled data and a smaller set of strategically queried data. DSL allows for efficient use of the query budget by focusing on the most informative samples, leading to a more accurate substitute LoRA. We present the algorithm detail of it in Alg. 1. + +# 4. Experiments + +# 4.1. Experimental Setup + +Victim Models. We use a ViT-Base model pre-trained on ImageNet-21k [33] with additional augmentations and regularization [38]. This base model is then fine-tuned with LoRA ( $r = 4$ ) on five commonly used datasets: CUBS200 [41], Caltech256 [9], Indoor67 [31], Food101 [2], and Flowers102 [26]. The resulting victim models achieve test accuracies of $88.09\%$ , $94.47\%$ , $85.52\%$ , $91.07\%$ , and $98.66\%$ , respectively. + +Attack Settings. We evaluate LoRA extraction attacks in Identical-Backbone (IB) and Cross-Backbone (XB) scenarios. In the IB setting, we use the same base model as the substitute base model. In the XB setting, we use + +the ViT-base model pre-trained on ImageNet-1k[35] with self-supervised masked autoencoder (MAE) method[12] as the substitute base model. Both substitute models utilize LoRA with $r = 4$ and are trained for 20 epochs using the Adam optimizer[19] with a cosine annealing learning rate schedule[24] and a base learning rate of 0.01. + +We use GPT-4o mini[27] as the LLM in StolenLoRA as it has similar performance at a lower price and can effectively reduce the attack cost. For the SD model, we use an open-source SDXL-Turbo model[36]. It only takes 4 sampling steps (compared to SDXL 1.0 which usually requires 40 steps) to synthesize high-quality images, significantly reducing the computational cost. For the StolenLoRA-DSL method, initial per-category sample size $N$ is set to 10, the scaling factor $\beta$ is 1.5 to not increase too much generation overhead, and the confidence threshold $\tau$ is 0.95. Hyperparameter analysis is provided in later sections. + +Baselines and Evaluation Metric. We compare Stolen-LoRA against KnockoffNets [28], ActiveThief [29], DFME [40], and $E^3$ [48] using a query budget of 10k. For methods such as KnockoffNets, ActiveThief, and $E^3$ that require real data, we use 3M images in CC3M[37] as the attack dataset. For $E^3$ , which originally selects samples based on semantic similarity between dataset category names and target class names, we improve its effectiveness by calculating semantic similarity between image captions (available in CC3M) and target class names for a more fine-grained sample selection. Following prior work [28, 40], we report Test Accuracy (Acc) and Attack Success Rate (ASR), defined as the ratio of the substitute model's accuracy to the victim model's accuracy. + +![](images/fb642a10e02c21f2b5a1ad15c81bfb17c288d6bbef4cd1df8f80fd51e84d35a6.jpg) +Figure 2. Curves of the test accuracy versus the number of queries in IB (solid line) and XB (dashed line) scenarios respectively. The gray dashed line represents the victim models' test accuracy. + +# 4.2. Experimental Results + +Effectiveness of StolenLoRA. Tab. 1 presents the effectiveness of StolenLoRA, comparing both the random strategy (StolenLoRA-Rand) and DSL-based (StolenLoRADSL) variants against existing extraction attacks. Across both identical- and cross-backbone scenarios, StolenLoRA consistently achieves state-of-the-art performance. In the IB setting, StolenLoRA-Rand shows superior performance on CUBS200 (75.35%) and Flowers102 (93.74%), while StolenLoRA-DSL excels on Caltech256 (89.30%) and Indoor67 (82.61%), maintaining competitive results on other datasets. The XB setting, more challenging due to base model differences, highlights StolenLoRA's significant advantage over baselines, with StolenLoRA-DSL achieving the highest accuracy across all five datasets. This demonstrates DSL's effectiveness in refining synthetic data and improving attack efficiency. Fig. 2 further analyzes query efficiency, showing StolenLoRA-DSL's rapid convergence in the IB setting and consistent progress in the XB setting, outperforming other methods in both scenarios. This efficiency stems from DSL's ability to prioritize informative queries, focusing on uncertain samples, thereby maximizing information gain from each query. + +Effectiveness under Hard-Label Scenario. We also evaluate StolenLoRA's performance when only hard labels (one-hot encoded predictions) are available, as is the case with some real-world APIs. Fig. 3 presents the results for both StolenLoRA-Rand and StolenLoRA-DSL in the IB and XB settings. In the IB setting, using hard labels leads to a minor performance decrease. For example, StolenLoRA-Rand's accuracy on CUBS200 drops from $75.35\%$ with soft labels to $67.00\%$ with hard labels. A similar trend is observed for other datasets and StolenLoRA-DSL. This suggests that the additional information provided by soft labels is beneficial in the IB setting. In the XB setting, the impact of hard labels is inconsistent. While a slight decrease is observed on some datasets, others, like Caltech256 and Indoor67, show a marginal increase in accuracy (from $51.75\%$ to $56.69\%$ ). This unexpected improvement might be due to the hard labels acting as a regularizer, preventing the substi + +![](images/018d8e09783e3741ce179587f870e0b434b43444fe5b8d91a1d1c75460e10ece.jpg) + +![](images/ddd5145e5d8f3a70ae5c14228176d60b1e8b30e63fa16790db2aee0a4bc1dd4e.jpg) +Figure 3. Comparison of StolenLoRA performance using hard labels (one-hot) versus soft labels (probabilities) from the victim model. Results are shown for both Random and DSL in identical- and cross-backbone scenarios across five datasets. + +tute from overfitting to the synthetic data, which is more prone to distributional shifts in the XB setting. Overall, StolenLoRA exhibits robustness to hard labels in both IB and XB scenarios, demonstrating its practical applicability. + +Ablation Study. We conduct an ablation study to analyze the contributions of the key components of StolenLoRA: the LLM-driven prompting and the DSL strategy, including the label refining loss. Tab. 2 presents the results on CUBS200 and Indoor67 datasets under both IB and XB settings with a 10k query budget. Removing the structured prompt template leads to a performance drop in both IB and XB settings, demonstrating the importance of providing structured visual information to the LLM. Removing the LLM entirely results in a more significant performance decrease, confirming the crucial role of the LLM in generating effective prompts for diverse and representative synthetic data. Furthermore, we evaluate different LLMs, observing that using + +
MethodIBXB
CUBS200Indoor67CUBS200Indoor67
Random75.3582.3845.7059.18
- Template72.3380.9442.2154.49
- LLM70.6674.7039.2047.09
+ Llama-3.1-8B73.9681.0446.7458.43
+ GPT-4o76.3083.1349.1963.58
DSL73.2382.6150.1465.07
- LLR70.1780.9748.2164.10
+ +Table 2. An ablation experiment showing the effectiveness of the two modules we designed on CUBS200 and Indoor67 datasets under both IB and XB settings with 10k queries. + +
τIBXB
CUBS200Indoor67CUBS200Indoor67
0.570.6878.8148.4360.90
0.771.3780.0048.3663.51
0.974.0481.1948.5263.06
0.9573.2382.6150.1465.07
0.9973.3181.1949.8661.72
+ +a more powerful LLM like GPT-4o generally yields better performance than Llama-3.1-8B[5], particularly in the more challenging XB setting. This highlights the benefit of utilizing a stronger LLM for prompt generation. + +Disabling DSL and reverting to random querying results in a noticeable performance reduction, especially in the XB setting. This underscores the effectiveness of DSL in selectively querying informative samples. Similarly, removing SAT from the DSL framework also leads to a performance decrease, albeit less pronounced than removing DSL altogether. This indicates the benefit of label refining in mitigating the distribution shift between synthetic and real data. + +We also analyze the impact of the confidence threshold $\tau$ within DSL in Tab. 3. The results show that performance generally improves as $\tau$ increases from 0.5 to 0.95. This indicates that focusing queries on samples where the substitute model is less confident leads to a more efficient use of the query budget. However, increasing $\tau$ further to 0.99 results in a slight performance decrease, suggesting that an overly stringent threshold can exclude valuable samples from being queried. We therefore select $\tau = 0.95$ as the optimal value for our experiments. + +Fidelity and Distribution Analysis of Synthetic Data. StolenLoRA's effectiveness relies on the quality of its synthetic training data. We evaluate this quality both visually and quantitatively. Visually, as shown in Fig. 4, synthetic images generated by StolenLoRA demonstrate a striking resemblance to real images from the target categories, capturing key features and background context crucial for train- + +![](images/e3667a820246873db8cca7a1f6f2627b73ecfc3481137a2156d4a902a93990f6.jpg) +(a) Visualization on CUBS200 + +![](images/dcc1be32d6ec2ff6484cb54d57bcbe9b990544315cb4489d539c0b297009618f.jpg) +(b) Visualization on Indoor67 +Figure 4. Visualization of StolenLoRA's synthetic images and corresponding real images of the target category. + +Table 3. Performance of StolenLoRA-DSL with varying confidence thresholds $\tau$ . Results are reported on CUBS200 and Indoor67 under both IB and XB settings with 10k queries. + +
MethodCUBS200Caltech256Indoor67Food101Flowers102
KnockoffNets51.448.9422.0150.0455.77
ActiveThief49.336.0115.7247.8450.19
DFME220.53179.45225.84188.29204.79
E341.574.296.6825.5134.82
StolenLoRA2.145.596.9319.5416.63
+ +Table 4. FID scores $(\downarrow)$ between the datasets used by each attack method and the victim model's training data distribution. + +ing. This visual fidelity is corroborated by significantly lower Fréchet Inception Distance (FID) scores[13] (in Tab. 4) compared to other attack methods. For example, StolenLoRA achieves an FID of 2.14 compared to 51.44 for KnockoffNets on CUBS200, highlighting a much closer distributional match to the target domain. This demonstrates that StolenLoRA's LLM-driven synthesis effectively generates high-fidelity, domain-specific data, even without access to the original training set, which is key to its superior extraction performance. + +Verifying the Distinction Hypothesis. As mentioned in Sec. 3.1, we hypothesize that the performance of a perfectly extracted surrogate model will be consistent across In-Distribution (ID) data with the victim model, but not necessarily Out-of-Distribution (OOD) data, motivating us to perform extraction attacks using ID data. To empirically verify this conjecture, we train a substitute model using the XB setting on the victim model's training dataset, aiming to achieve a near-perfect extraction of the LoRA adaptations. We then evaluate both the victim model and the extracted substitute on two sets of data: ID data, represented by the victim model's held-out test set, and OOD data, sampled from the CC3M dataset. For each sample, we calculate the cross-entropy between the victim model's predictions and + +![](images/a8effbcfe67e86eaaf5ffbcb4f61507356f7e1869bd230eb26787cc60792d0c0.jpg) +Figure 5. Cross-entropy difference between the victim model and a near-perfectly substitute model on In-Distribution (ID) and Out-of-Distribution (OOD) data. The substitute model is trained using the cross-backbone setting on the victim's training dataset. + +the substitute model's predictions, providing a measure of their functional divergence. Fig. 5 presents the distribution of these differences. On both CUBS200 and Indoor67, the difference is smaller for the ID data compared to the OOD data. This observation supports our conjecture that even near-perfect LoRA extraction yields a substitute model that exhibits functional equivalence primarily within the ID domain. The divergence in OOD data likely stems from the inherent differences between the pre-trained backbone used by the victim and the substitute. This result underscores the importance of utilizing ID data for effective LoRA extraction, as the attacker primarily seeks to replicate the victim's specialized adaptations rather than the general capabilities. + +# 5. Defending Against LoRA Extraction + +While this paper primarily focuses on the feasibility of LoRA extraction, we also explore a potential defense mechanism based on deploying multiple, diverse LoRA adapters to increase attacker uncertainty. + +Dual LoRA with Diversified Predictions. Our proposed defense involves training two LoRA adapters, denoted as $\mathrm{L}_A$ and $\mathrm{L}_B$ , to maximize the difference in their output distributions while maintaining comparable performance on the target task. Instead of minimizing accuracy as initially conceived, we found that maintaining high accuracy while maximizing divergence is crucial for practical deployment. We achieve this by optimizing the following loss function: + +$$ +\mathcal {L} ^ {\prime} = \mathcal {L} (\mathrm {L} _ {A}) + \mathcal {L} (\mathrm {L} _ {B}) - \lambda \times \mathrm {K L} (\mathrm {L} _ {A} | | \mathrm {L} _ {B}) \tag {8} +$$ + +where $\mathcal{L}$ represents the cross-entropy loss for the target task, KL represents the Kullback-Leibler divergence between the output distributions of the two LoRAs, and $\lambda$ is a hyperparameter controlling the balance between task performance and divergence. At deployment, we randomly select either $L_{A}$ or $L_{B}$ for each incoming query to generate the prediction. This random selection introduces uncertainty and makes it challenging for the attacker to extract the underlying LoRA function consistently. + +![](images/ac1b89e4554fa8412367296f397452b438629a1081ed55ee1b1f02be72bd8252.jpg) +Figure 6. Defneder Acc versus Substitute model Acc trade-off for defenses evaluated against StolenLoRA and KnockoffNets on the CUBS200 dataset with 10k queries. + +Experimental Evaluation. We evaluate the proposed defense mechanism under both IB and XB settings with 10k queries. Fig. 6 illustrates the trade-off between defender Acc and substitute model Acc. The defense consistently degrades the substitute model's performance. KnockoffNets experiences substantial accuracy drops (from $70.90\%$ to $20.78\%$ in IB and from $6.77\%$ to $1.21\%$ in XB), and StolenLoRA-Rand shows similar vulnerability (decreasing $22.95\%$ in IB and $15.63\%$ in XB). However, StolenLoRADSL demonstrates greater resilience, with comparatively smaller accuracy reductions ( $9.63\%$ in IB and $14.34\%$ in XB). This resilience is attributed to DSL's efficient query selection and label refining, focusing on informative samples and mitigating the uncertainty introduced by the defense. While this defense effectively hinders less sophisticated attacks, it underscores the need for more robust mechanisms to counter advanced techniques like DSL that can learn effectively from limited queries. + +# 6. Conclusion + +This paper introduces LoRA Extraction, a novel model extraction attack targeting the widespread practice of adapting large vision models with LoRA. We present StolenLoRA, a highly effective method leveraging LLM-driven Stable Diffusion to generate task-specific, high-fidelity synthetic training data, bypassing the limitations of traditional extraction techniques reliant on large-scale real datasets or unreliable GANs. Furthermore, StolenLoRA incorporates Disagreement-based Semi-supervised Learning (DSL) to efficiently query the victim model, focusing on uncertain predictions and iteratively refining labels, enabling successful extraction with limited queries. Extensive experiments demonstrate StolenLoRA's strong performance across diverse datasets and in challenging cross-backbone scenarios. We also explore a preliminary defense based on diversified LoRA deployments, showing promise in mitigating these attacks. This work represents a crucial first step towards understanding and addressing the safety risks associated with LoRA adaptation, paving the way for more secure deployments of efficient fine-tuning methods. + +# Acknowledgments + +This work is in part supported by National Key R&D Program of China (Grant No. 2022ZD0160103) and National Natural Science Foundation of China (Grant No. 62276067), and Shanghai Artificial Intelligence Laboratory. + +# References + +[1] Martin Arjovsky and Léon Bottou. Towards principled methods for training generative adversarial networks. arXiv preprint arXiv:1701.04862, 2017. 2 +[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In ECCV, 2014. 5 +[3] Eric L Buehler and Markus J Buehler. X-lora: Mixture of low-rank adapter experts, a flexible framework for large language models with applications in protein mechanics and molecular design. APL Machine Learning, 2024. 2 +[4] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1 +[5] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 7 +[6] Ricard Durall, Avraam Chatzimichailidis, Peter Labus, and Janis Keuper. Combating mode collapse in gan training: An empirical analysis using hessian eigenvalues. arXiv preprint arXiv:2012.09673, 2020. 2 +[7] Lijie Fan, Kaifeng Chen, Dilip Krishnan, Dina Katabi, Phillip Isola, and Yonglong Tian. Scaling laws of synthetic images for model training... for now. In CVPR, 2024. 2 +[8] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NIPS, 2014. 1, 2 +[9] Gregory Griffin, Alex Holub, and Pietro Perona. Caltech-256 object category dataset. 2007. 5 +[10] Zeyu Han, Chao Gao, Jinyang Liu, Sai Qian Zhang, et al. Parameter-efficient fine-tuning for large models: A comprehensive survey. arXiv preprint arXiv:2403.14608, 2024. 2 +[11] Soufiane Hayou, Nikhil Ghosh, and Bin Yu. Lora+: Efficient low rank adaptation of large models. arXiv preprint arXiv:2402.12354, 2024. 1 +[12] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 1, 5 +[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 7 +[14] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In ICML, 2019. 2 + +[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1, 2 +[16] Nam Hyeon-Woo, Moon Ye-Bin, and Tae-Hyun Oh. Fedpara: Low-rank hadamard product for communication-efficient federated learning. arXiv preprint arXiv:2108.06098, 2021. 2 +[17] Matthew Jagielski, Nicholas Carlini, David Berthelot, Alex Kurakin, and Nicolas Papernot. High accuracy and high fidelity extraction of neural networks. In *Usenix Security*, 2020. 1 +[18] Sanjay Kariyappa, Atul Prakash, and Moinuddin Qureshi. Maze: Data-free model stealing attack using zeroth-order gradient estimation. In CVPR, 2021. 1, 2 +[19] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5 +[20] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In CVPR, 2023. 1 +[21] Youssef Kossale, Mohammed Airaj, and Aziz Darouichi. Mode collapse in generative adversarial networks: An overview. In ICOA, 2022. 2 +[22] Zi Liang, Qingqing Ye, Yanyun Wang, Sen Zhang, Yaxin Xiao, Ronghua Li, Jianliang Xu, and Haibo Hu. Alignment-aware model extraction attacks on large language models. arXiv preprint arXiv:2409.02718, 2024. 1, 2 +[23] Zijun Lin, Ke Xu, Chengfang Fang, Huadi Zheng, Aneez Ahmed Jaheezuddin, and Jie Shi. Quda: Query-limited data-free model extraction. In ACM Asia CCS, 2023. 1, 2 +[24] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In *ICLR*, 2017. 5 +[25] Xingjun Ma, Yifeng Gao, Yixu Wang, Ruofan Wang, Xin Wang, Ye Sun, Yifan Ding, Hengyuan Xu, Yunhao Chen, Yunhan Zhao, et al. Safety at scale: A comprehensive survey of large model safety. arXiv preprint arXiv:2502.05206, 2025.1 +[26] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian conference on computer vision, graphics & image processing, 2008. 5 +[27] OpenAI. Gpt-4o mini: Advancing cost-efficient intelligence. https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence/, 2024.5 +[28] Tribhuvanesh Orekondy, Bernt Schiele, and Mario Fritz. Knockoff nets: Stealing functionality of black-box models. In CVPR, 2019. 1, 2, 5 +[29] Soham Pal, Yash Gupta, Aditya Shukla, Aditya Kanade, Shirish Shevade, and Vinod Ganapathy. Activethief: Model extraction using active learning and unannotated public data. In AAAI, 2020. 1, 2, 5 +[30] Jie Qin, Jie Wu, Weifeng Chen, Yuxi Ren, Huixia Li, Hefeng Wu, Xuefeng Xiao, Rui Wang, and Shilei Wen. Diffusiongpt: Llm-driven text-to-image generation system. arXiv preprint arXiv:2401.10061, 2024. 2 + +[31] Ariadna Quattoni and Antonio Torralba. Recognizing indoor scenes. In CVPR, 2009. 5 +[32] Pengzhen Ren, Yun Xiao, Xiaojun Chang, Po-Yao Huang, Zhihui Li, Brij B Gupta, Xiaojiang Chen, and Xin Wang. A survey of deep active learning. ACM computing surveys, 2021. 2 +[33] Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik-Manor. Imagenet-21k pretraining for the masses. arXiv preprint arXiv:2104.10972, 2021. 5 +[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2 +[35] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 2015. 5 +[36] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 2, 5 +[37] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 5 +[38] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. arXiv preprint arXiv:2106.10270, 2021. 5 +[39] Florian Tramér, Fan Zhang, Ari Juels, Michael K. Reiter, and Thomas Ristenpart. Stealing machine learning models via prediction apis. In USENIX Security Symposium, 2016. 1 +[40] Jean-Baptiste Truong, Pratyush Maini, Robert J Walls, and Nicolas Papernot. Data-free model extraction. In CVPR, 2021. 1, 2, 5 +[41] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5 +[42] Yixu Wang, Jie Li, Hong Liu, Yan Wang, Yongjian Wu, Feiyue Huang, and Rongrong Ji. Black-box dissector: Towards erasing-based hard-label model stealing attack. In ECCV, 2022. 1, 2 +[43] Yixu Wang, Tianle Gu, Yan Teng, Yingchun Wang, and Xingjun Ma. Honeypotnet: Backdoor attacks against model extraction. In AAAI, 2025. 1 +[44] Shih-Ying Yeh, Yu-Guan Hsieh, Zhidong Gao, Bernard BW Yang, Giyeong Oh, and Yanmin Gong. Navigating text-to-image customization: From lycoris fine-tuning to model evaluation. In ICLR, 2023. 2 +[45] Jianhao Yuan, Jie Zhang, Shuyang Sun, Philip Torr, and Bo Zhao. Real-fake: Effective training data synthesis through distribution matching. In ICLR, 2024. 2 +[46] Yunlong Zhao, Xiaoheng Deng, Yijing Liu, Xinjun Pei, Ji-azhi Xia, and Wei Chen. Fully exploiting every real sample: Superpixel sample gradient model stealing. In CVPR, 2024. 1, 2 +[47] Mingyi Zhou, Jing Wu, Yipeng Liu, Shuaicheng Liu, and Ce Zhu. Dast: Data-free substitute training for adversarial attacks. In CVPR, 2020. 1, 2 + +[48] Hongyu Zhu, Wentao Hu, Sichu Liang, Fangqi Li, Wenwen Wang, and Shilin Wang. Efficient and effective model extraction. arXiv preprint arXiv:2409.14122, 2024. 5 \ No newline at end of file diff --git a/2025/StolenLoRA_ Exploring LoRA Extraction Attacks via Synthetic Data/images.zip b/2025/StolenLoRA_ Exploring LoRA Extraction Attacks via Synthetic Data/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d4675fe1f087a4d2b18c19db61659b5db288fec4 --- /dev/null +++ b/2025/StolenLoRA_ Exploring LoRA Extraction Attacks via Synthetic Data/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aad6de440ba2a4728f9d7ef0bb8153c8f7a40350d7bdd8f179c68aebd3fd97a0 +size 671285 diff --git a/2025/StolenLoRA_ Exploring LoRA Extraction Attacks via Synthetic Data/layout.json b/2025/StolenLoRA_ Exploring LoRA Extraction Attacks via Synthetic Data/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..77bed6604a6c49b601b19f2705fddfa6aba67ee8 --- /dev/null +++ b/2025/StolenLoRA_ Exploring LoRA Extraction Attacks via Synthetic Data/layout.json @@ -0,0 +1,9464 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 92, + 103, + 518, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 103, + 518, + 121 + ], + "spans": [ + { + "bbox": [ + 92, + 103, + 518, + 121 + ], + "type": "text", + "content": "StolenLoRA: Exploring LoRA Extraction Attacks via Synthetic Data" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 142, + 449, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 142, + 449, + 159 + ], + "spans": [ + { + "bbox": [ + 160, + 142, + 449, + 159 + ], + "type": "text", + "content": "Yixu Wang" + }, + { + "bbox": [ + 160, + 142, + 449, + 159 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 160, + 142, + 449, + 159 + ], + "type": "text", + "content": ", Yan Teng" + }, + { + "bbox": [ + 160, + 142, + 449, + 159 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 160, + 142, + 449, + 159 + ], + "type": "text", + "content": ", Yingchun Wang" + }, + { + "bbox": [ + 160, + 142, + 449, + 159 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 160, + 142, + 449, + 159 + ], + "type": "text", + "content": ", Xingjun Ma" + }, + { + "bbox": [ + 160, + 142, + 449, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 162, + 465, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 162, + 465, + 178 + ], + "spans": [ + { + "bbox": [ + 146, + 162, + 465, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 162, + 465, + 178 + ], + "type": "text", + "content": "Fudan University " + }, + { + "bbox": [ + 146, + 162, + 465, + 178 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 146, + 162, + 465, + 178 + ], + "type": "text", + "content": "Shanghai Artificial Intelligence Laboratory" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 205, + 200, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 205, + 200, + 217 + ], + "spans": [ + { + "bbox": [ + 151, + 205, + 200, + 217 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 230, + 297, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 230, + 297, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 230, + 297, + 529 + ], + "type": "text", + "content": "Parameter-Efficient Fine-Tuning (PEFT) methods like LoRA have transformed vision model adaptation, enabling the rapid deployment of customized models. However, the compactness of LoRA adaptations introduces new safety concerns, particularly their vulnerability to model extraction attacks. This paper introduces a new focus of model extraction attacks named LoRA extraction that extracts LoRA-adaptive models based on a public pre-trained model. We then propose a novel extraction method called StolenLoRA which trains a substitute model to extract the functionality of a LoRA-adapted model using synthetic data. StolenLoRA leverages a Large Language Model to craft effective prompts for data generation, and it incorporates a Disagreement-based Semi-supervised Learning (DSL) strategy to maximize information gain from limited queries. Our experiments demonstrate the effectiveness of StolenLoRA, achieving up to a " + }, + { + "bbox": [ + 55, + 230, + 297, + 529 + ], + "type": "inline_equation", + "content": "96.60\\%" + }, + { + "bbox": [ + 55, + 230, + 297, + 529 + ], + "type": "text", + "content": " attack success rate with only 10k queries, even in cross-backbone scenarios where the attacker and victim models utilize different pretrained backbones. These findings reveal the specific vulnerability of LoRA-adapted models to this type of extraction and underscore the urgent need for robust defense mechanisms tailored to PEFT methods. We also explore a preliminary defense strategy based on diversified LoRA deployments, highlighting its potential to mitigate such attacks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 551, + 135, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 551, + 135, + 562 + ], + "spans": [ + { + "bbox": [ + 56, + 551, + 135, + 562 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 571, + 296, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 296, + 666 + ], + "type": "text", + "content": "Adapting large-scale pre-trained foundation models via Parameter-Efficient Fine-Tuning (PEFT) such as Low-Rank Adaptation (LoRA) [11, 15] has become a popular approach for obtaining high-performance models for diverse downstream tasks. LoRA enables efficient adaptation of large models to a new task by fine-tuning only low-rank matrices added to specific layers, reducing computational cost and memory usage while preserving the pre-trained model's" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 205, + 555, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 555, + 374 + ], + "type": "text", + "content": "core knowledge. Despite the wide adoption of LoRA, the lightweight and compactness of LoRA parameters raise a new safety concern: they might be more vulnerable to model extraction (ME) attacks. An ME attack extracts the functionality [17, 22, 28, 39] of a victim model by training a substitute model based on the outputs of the victim model for a certain number of specially designed queries. While ME attacks have been extensively studied for traditional models [29, 40, 42, 43, 46, 47], the vulnerability of LoRA adaptations to ME attacks remains largely unexplored[25]. This gap is significant, as the public availability of a large number of pre-trained models, coupled with the compactness of LoRA parameters, makes it easier to replicate LoRA-adapted models and compromise intellectual property." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 376, + 556, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 556, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 556, + 567 + ], + "type": "text", + "content": "To address this gap, we introduce the concept of LoRA Extraction, a novel direction of ME attacks specifically targeting LoRA-adapted models (i.e., ViTs [4, 12, 20]), based on a publicly available pre-trained model. Unlike traditional ME which focuses on replicating the entire model's functionality, LoRA extraction centers on stealing the efficient adaptations encoded within the compact LoRA parameters. The adversary's objective is to reconstruct the victim's LoRA-adapted model by training their own LoRA-adapted substitute that achieves similar downstream performance using a pre-trained foundation model that is publicly available. This attack can manifest in two scenarios: the identical-backbone scenario, where the attacker uses the same pre-trained ViT model as the victim, and the more challenging cross-backbone scenario, where the attacker uses a different pre-trained ViT." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "type": "text", + "content": "Existing traditional ME methods largely depend on sample selection from existing datasets or synthetic data generation. Sample selection methods [28, 29, 42, 46] typically require searching extensive datasets for in-distribution samples, making them computationally expensive and less practical for LoRA extraction. This impracticality arises because 1) inference in LoRA-adapted models is slower due to the large-scale parameters of LoRA-adapted models, and 2) finding appropriate samples is challenging, particularly for domain-specific data fine-tuned with LoRA. Alternatively, synthetic data generation approaches [18, 23, 40, 47] often rely on Generative Adversarial Networks (GANs) [8]" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 672, + 296, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 296, + 714 + ], + "type": "text", + "content": "\\*Corresponding authors: < tengyan@pjlab.org.cn, xingjunma@fudan.edu.cn > T Work done during internship at Shanghai Artificial Intelligence Laboratory." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "263" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "to produce in-distribution data. However, GANs frequently face difficulties in generating high-quality, diverse samples, especially for high-dimensional data. As Zhao et al. [46] noted, creating high-dimensional samples (e.g., " + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": " images) for effective ME can require millions of queries and is prone to failure due to GAN mode collapse [1, 6, 21]. These limitations make GAN-based approaches less effective for LoRA extraction, where the tuning often involves high-dimensional image data." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 180, + 298, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 180, + 298, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 180, + 298, + 441 + ], + "type": "text", + "content": "In this work, we propose a novel LoRA extraction method named StolenLoRA, which employs synthetic data to train the substitute model. To obtain effective in-distribution synthetic data, we leverage Large Language Models (LLMs) to generate diverse textual descriptions based on the target class names and prompt a pre-trained Stable Diffusion model [34, 36] to generate high-quality images. This allows us to tailor the generated data to the specific downstream task the victim's LoRA model is fine-tuned for. Moreover, StolenLoRA introduces a distinctive strategy to improve attack efficiency, i.e., the Disagreement-based Semi-supervised Learning (DSL). DSL uses class information from the data synthesis process as pseudo-labels for part of the generated dataset and focuses on queries where the substitute model disagrees with these pseudo-labels. DSL effectively targets areas of uncertainty, guiding the substitute model's learning process and refining its alignment with the victim's behavior. To further enhance DSL, we iteratively refine the pseudo-labels based on evolving predictions from the substitute model, improving the quality of synthetic data labels and the effectiveness of the extraction." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 443, + 251, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 443, + 251, + 454 + ], + "spans": [ + { + "bbox": [ + 66, + 443, + 251, + 454 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 455, + 295, + 635 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 455, + 295, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 295, + 502 + ], + "type": "text", + "content": "- We explore the vulnerability of LoRA and introduce a novel focus of ME attacks called LoRA Extraction which aims to extract the functionality of LoRA-adapted models based on a publicly available pre-trained model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 503, + 295, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 295, + 562 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 295, + 562 + ], + "type": "text", + "content": "- We present StolenLoRA, a novel LoRA extraction method that leverages LLM-driven Stable Diffusion to generate high-quality synthetic data, circumventing the need to search large-scale datasets or rely on less dependable GAN-based generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 563, + 295, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 563, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 563, + 295, + 635 + ], + "type": "text", + "content": "- Through comprehensive experiments on five widely used datasets, we demonstrate the effectiveness of Stolen-LoRA in both identical- and cross-backbone scenarios, highlighting critical vulnerabilities in LoRA-adapted models. We also explore a preliminary defense mechanism based on diversified LoRA deployments." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 645, + 142, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 142, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 142, + 658 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": "Low-Rank Adaptation (LoRA). LoRA [15] is a type of popular Parameter-Efficient Fine-Tuning (PEFT) method [10, 14] for adapting large pre-trained models for downstream tasks. LoRA [15] introduces small rank-" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "content": "decomposition matrices as trainable parameters into the pre-trained model. The number of trainable parameters is significantly reduced by freezing the original model weights and updating only these lightweight LoRA modules. LoRA inspires numerous variants and extensions, such as Mixture-of-LoRA Experts (X-LoRA)[3], Low-Rank Hadamard Product (LoHa)[16], and Low-Rank Kronecker Product (LoKr)[44], further enhancing its efficiency and flexibility for various applications. However, the compact nature of these LoRA updates raises concerns about their vulnerability to model extraction attacks. This paper investigates the potential for adversaries to extract the efficient adaptations encoded within these compact parameters, specifically focusing on the standard LoRA method [15]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 266, + 555, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 555, + 460 + ], + "type": "text", + "content": "Model Extraction Attacks. Model Extraction (ME) attacks [18, 22, 23, 28, 29, 40, 42, 46, 47] aim to construct a substitute model that mimics the functionality of a victim model by querying its API. Existing ME methods can be broadly categorized based on their query generation strategies: Data-Driven ME leverage existing datasets [28, 29, 42, 46] to select queries, often employing active learning[32] or other search strategies[28] to identify informative samples. However, this type can incur computational overhead as it necessitates traversing the large-scale dataset to find suitable in-distribution samples. Data-Free ME [18, 23, 40, 46, 47] utilize generative models (e.g., GANs[8]) to synthesize data for querying the victim model. While promising, generating effective queries with high-dimensional inputs can require tens of millions of queries and is prone to failure due to GAN mode collapse [1, 6, 21]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": "Learning from Synthetic Data. The use of synthetic data for training models gains significant traction due to its potential to address data scarcity. While traditional synthetic data often lacked the complexity and representativeness of real data, recent advancements in generative models, particularly diffusion models like Stable Diffusion [34, 36], enable the synthesis of high-quality and diverse images. However, scaling the quantity of synthetic data does not necessarily translate to improved performance for training supervised image classifiers. As shown by Fan et al.[7], performance with synthetic data can lag significantly behind training with real data. This disparity stems from the limitations of current text-to-image models in accurately generating diverse representations of certain concepts at scale[30]. Efforts to mitigate this gap, such as the distribution-matching framework[45], focus on aligning the distributions of synthetic and real data. However, leveraging such high-quality synthetic data for model extraction attacks remains an unexplored area." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "264" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 550, + 252 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 550, + 252 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 550, + 252 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 550, + 252 + ], + "type": "image", + "image_path": "ee8f6f3384583ba07b2dbe2a5d69d0ed57a8c92aea61e38c2c949168b526c2d6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 258, + 555, + 282 + ], + "lines": [ + { + "bbox": [ + 55, + 258, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 555, + 282 + ], + "type": "text", + "content": "Figure 1. Overview of StolenLoRA. It collects target victim model information (dashed red box), synthesizes training data guided by an LLM-driven prompting (dashed blue box), and then efficiently queries the victim API to train a substitute model (dashed green box)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 286, + 138, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 286, + 138, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 138, + 300 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 306, + 180, + 318 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 306, + 180, + 318 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 180, + 318 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 323, + 296, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 431 + ], + "type": "text", + "content": "LoRA Extraction. Given a victim model " + }, + { + "bbox": [ + 55, + 323, + 296, + 431 + ], + "type": "inline_equation", + "content": "F \\colon [0,1]^d \\mapsto \\mathbb{R}^N" + }, + { + "bbox": [ + 55, + 323, + 296, + 431 + ], + "type": "text", + "content": ", which is a pre-trained model " + }, + { + "bbox": [ + 55, + 323, + 296, + 431 + ], + "type": "inline_equation", + "content": "F_{base}" + }, + { + "bbox": [ + 55, + 323, + 296, + 431 + ], + "type": "text", + "content": " (e.g., ViTs) adapted using LoRA. The parameters of " + }, + { + "bbox": [ + 55, + 323, + 296, + 431 + ], + "type": "inline_equation", + "content": "F^1" + }, + { + "bbox": [ + 55, + 323, + 296, + 431 + ], + "type": "text", + "content": " can be represented as " + }, + { + "bbox": [ + 55, + 323, + 296, + 431 + ], + "type": "inline_equation", + "content": "\\theta = \\{\\theta_{base}, \\Delta \\theta\\}" + }, + { + "bbox": [ + 55, + 323, + 296, + 431 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 323, + 296, + 431 + ], + "type": "inline_equation", + "content": "\\theta_{base}" + }, + { + "bbox": [ + 55, + 323, + 296, + 431 + ], + "type": "text", + "content": " are the parameters of " + }, + { + "bbox": [ + 55, + 323, + 296, + 431 + ], + "type": "inline_equation", + "content": "F_{base}" + }, + { + "bbox": [ + 55, + 323, + 296, + 431 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 323, + 296, + 431 + ], + "type": "inline_equation", + "content": "\\Delta \\theta" + }, + { + "bbox": [ + 55, + 323, + 296, + 431 + ], + "type": "text", + "content": " represents the LoRA updates, which are typically low-rank matrices. The objective of a LoRA extraction attack is to train a substitute model " + }, + { + "bbox": [ + 55, + 323, + 296, + 431 + ], + "type": "inline_equation", + "content": "F'" + }, + { + "bbox": [ + 55, + 323, + 296, + 431 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 55, + 323, + 296, + 431 + ], + "type": "inline_equation", + "content": "\\theta' = \\{\\theta_{base}', \\Delta \\theta'\\}" + }, + { + "bbox": [ + 55, + 323, + 296, + 431 + ], + "type": "text", + "content": " that mimics the functionality of " + }, + { + "bbox": [ + 55, + 323, + 296, + 431 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 323, + 296, + 431 + ], + "type": "text", + "content": ". This is formulated as the following optimization problem:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 437, + 294, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 437, + 294, + 455 + ], + "spans": [ + { + "bbox": [ + 115, + 437, + 294, + 455 + ], + "type": "interline_equation", + "content": "\\arg \\min _ {\\Delta \\theta^ {\\prime}} \\mathbb {E} _ {\\boldsymbol {x}} \\mathcal {D} (F (\\boldsymbol {x}), F ^ {\\prime} (\\boldsymbol {x})), \\tag {1}", + "image_path": "2a17f5673a4449531459dfe1e0fcbe70b76b579faad4c29a445d184a3cda3940.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 462, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 296, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 462, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 462, + 296, + 498 + ], + "type": "text", + "content": " represents an input sample from the victim model task distribution, " + }, + { + "bbox": [ + 55, + 462, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 55, + 462, + 296, + 498 + ], + "type": "text", + "content": " represents a measure of functional similarity between the two models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 498, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 544 + ], + "type": "text", + "content": "Depending on whether the attacker has knowledge of " + }, + { + "bbox": [ + 55, + 498, + 296, + 544 + ], + "type": "inline_equation", + "content": "F_{base}" + }, + { + "bbox": [ + 55, + 498, + 296, + 544 + ], + "type": "text", + "content": ", we distinguish between two scenarios: identical backbone and cross-backbone LoRA extractions. The substitute model " + }, + { + "bbox": [ + 55, + 498, + 296, + 544 + ], + "type": "inline_equation", + "content": "F'" + }, + { + "bbox": [ + 55, + 498, + 296, + 544 + ], + "type": "text", + "content": " can be defined as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 64, + 550, + 294, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 550, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 64, + 550, + 294, + 590 + ], + "type": "interline_equation", + "content": "F ^ {\\prime} (\\boldsymbol {x}) = \\left\\{ \\begin{array}{l l} F _ {\\text {b a s e}} (\\boldsymbol {x}) + \\Delta F ^ {\\prime} (\\boldsymbol {x}), & \\text {i d e n t i c a l - b a c k b o n e} \\\\ G _ {\\text {b a s e}} (\\boldsymbol {x}) + \\Delta F ^ {\\prime} (\\boldsymbol {x}), & \\text {c r o s s - b a c k b o n e} \\end{array} \\right. \\tag {2}", + "image_path": "34744987dea7095d5bc7661e4885ada49174c73bd5cacc1b90cf957614d72a5e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 592, + 295, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 295, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 295, + 627 + ], + "type": "text", + "content": "where the former uses the same " + }, + { + "bbox": [ + 55, + 592, + 295, + 627 + ], + "type": "inline_equation", + "content": "F_{base}" + }, + { + "bbox": [ + 55, + 592, + 295, + 627 + ], + "type": "text", + "content": ", the latter uses different base models " + }, + { + "bbox": [ + 55, + 592, + 295, + 627 + ], + "type": "inline_equation", + "content": "G_{base}" + }, + { + "bbox": [ + 55, + 592, + 295, + 627 + ], + "type": "text", + "content": ", and the " + }, + { + "bbox": [ + 55, + 592, + 295, + 627 + ], + "type": "inline_equation", + "content": "+" + }, + { + "bbox": [ + 55, + 592, + 295, + 627 + ], + "type": "text", + "content": " symbol denotes the combination of the base model and the LoRA updates." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 628, + 296, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 296, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 296, + 688 + ], + "type": "text", + "content": "Distinction from Traditional Extraction. Traditional model extraction trains a substitute from scratch, while LoRA extraction leverages a pre-trained base model, inheriting its knowledge and biases. Consequently, we hypothesize that a perfectly extracted LoRA substitute might" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 288, + 555, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 555, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 555, + 347 + ], + "type": "text", + "content": "achieve functional equivalence on In-Distribution (ID) data but not on Out-of-Distribution (OOD) data, particularly in the cross-backbone setting, and we experimentally verify this in the experimental section. This motivates the use of ID data for effective LoRA extraction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 354, + 394, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 394, + 364 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 394, + 364 + ], + "type": "text", + "content": "3.2. StolenLoRA" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 371, + 554, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 554, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 554, + 455 + ], + "type": "text", + "content": "Overview. An overview of our proposed LLM-driven generative attack for extracting LoRA, StolenLoRA, is illustrated in Fig. 1. It comprises two key stages: 1) Data Synthesis, where an LLM guides the generation of synthetic data tailored to the victim's task, and 2) Efficient Querying and Training, which strategically selects queries and refines labels to maximize information gain from the victim model." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 460, + 405, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 405, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 405, + 471 + ], + "type": "text", + "content": "3.2.1. Data Synthesis" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "text", + "content": "The above discussion motivates us to use ID data for extraction. However, acquiring such datasets in real-world scenarios is often challenging because the victim model's training data is proprietary. While the attacker might not have access to the original training data, they often have access to the model's deployment context, which usually includes a description of the model's functionality and the names of the target classes. This information can be leveraged to semantically expand the search space for potential training data. Inspired by this, we propose to utilize LLM-driven Stable Diffusion to generate synthetic data based on the class names, effectively bridging the gap between semantic information and visual representations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "content": "LLM-Driven Prompting. Given a set of target class names " + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{C} = \\{c_1, c_2, \\dots, c_n\\}" + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "content": ", we employ an LLM to generate detailed prompts for image synthesis. We structure the prompts using a predefined template " + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "content": " comprising key visual elements: " + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "inline_equation", + "content": "T = \\text{\"[Subject, Background, Angle/ Pose, Lighting, Style]\"}" + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "content": ". For each class " + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "content": ", the LLM generates " + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "content": " variations of the" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "1We omit the parameter for simplicity when there is no ambiguity, e.g., using " + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "inline_equation", + "content": "F(\\pmb{x})" + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "inline_equation", + "content": "F(\\pmb{x};\\theta)" + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "265" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": "prompt, denoted as " + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "inline_equation", + "content": "p_{i,j}" + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "inline_equation", + "content": "j \\in \\{1, \\dots, m\\}" + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": ". This diversification ensures a richer representation of the target class. The prompt generation process is formalized as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 127, + 117, + 295, + 131 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 117, + 295, + 131 + ], + "spans": [ + { + "bbox": [ + 127, + 117, + 295, + 131 + ], + "type": "interline_equation", + "content": "p _ {i, j} = \\operatorname {L L M} \\left(c _ {i}, T, \\omega_ {j}\\right), \\tag {3}", + "image_path": "675563f52a0464e1050df15832c9d6fbc7886274a1ce9e342be1434258a337c2.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 137, + 295, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 137, + 295, + 161 + ], + "spans": [ + { + "bbox": [ + 55, + 137, + 295, + 161 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 137, + 295, + 161 + ], + "type": "inline_equation", + "content": "\\omega_{j}" + }, + { + "bbox": [ + 55, + 137, + 295, + 161 + ], + "type": "text", + "content": " represents a random seed or instruction to control the variation generated by the LLM." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 161, + 296, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 161, + 296, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 161, + 296, + 236 + ], + "type": "text", + "content": "Image Synthesis. We employ a public pre-trained Stable Diffusion model, denoted as " + }, + { + "bbox": [ + 55, + 161, + 296, + 236 + ], + "type": "inline_equation", + "content": "\\mathrm{SD}(\\cdot)" + }, + { + "bbox": [ + 55, + 161, + 296, + 236 + ], + "type": "text", + "content": ", to generate images corresponding to the crafted prompts. For each prompt " + }, + { + "bbox": [ + 55, + 161, + 296, + 236 + ], + "type": "inline_equation", + "content": "p_{i,j}" + }, + { + "bbox": [ + 55, + 161, + 296, + 236 + ], + "type": "text", + "content": " we synthesize one image, resulting in a set of synthetic images " + }, + { + "bbox": [ + 55, + 161, + 296, + 236 + ], + "type": "inline_equation", + "content": "\\mathbb{X}_i = \\{\\pmb{x}_{i,j}\\}" + }, + { + "bbox": [ + 55, + 161, + 296, + 236 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 161, + 296, + 236 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{i,j} = \\mathrm{SD}(p_{i,j})" + }, + { + "bbox": [ + 55, + 161, + 296, + 236 + ], + "type": "text", + "content": ". The complete synthetic dataset is then " + }, + { + "bbox": [ + 55, + 161, + 296, + 236 + ], + "type": "inline_equation", + "content": "\\mathbb{X} = \\bigcup_{i}\\mathbb{X}_{i}" + }, + { + "bbox": [ + 55, + 161, + 296, + 236 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 239, + 221, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 221, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 221, + 251 + ], + "type": "text", + "content": "3.2.2. Efficient Querying and Training" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 253, + 296, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 253, + 296, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 296, + 301 + ], + "type": "text", + "content": "This stage focuses on strategically querying the victim model to refine the substitute model's LoRA parameters while minimizing the number of queries. We propose two attack strategies: random learning and DSL." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 301, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 296, + 361 + ], + "type": "text", + "content": "Random Learning (StolenLoRA-Rand). This straightforward approach directly uses the synthesized dataset " + }, + { + "bbox": [ + 55, + 301, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathbb{X}" + }, + { + "bbox": [ + 55, + 301, + 296, + 361 + ], + "type": "text", + "content": " to query the victim model " + }, + { + "bbox": [ + 55, + 301, + 296, + 361 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 301, + 296, + 361 + ], + "type": "text", + "content": ". Then, the substitute model's LoRA parameters " + }, + { + "bbox": [ + 55, + 301, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\Delta \\theta^{\\prime}" + }, + { + "bbox": [ + 55, + 301, + 296, + 361 + ], + "type": "text", + "content": " are trained by minimizing the following objective function:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 369, + 295, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 369, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 104, + 369, + 295, + 384 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\prime} (\\Delta \\theta^ {\\prime}) = \\mathbb {E} _ {\\boldsymbol {x} \\in \\mathbb {X}} \\mathcal {L} \\left(F ^ {\\prime} (\\boldsymbol {x}), F (\\boldsymbol {x})\\right), \\tag {4}", + "image_path": "bc4752d5a15b38d3099b34d02f21b5a60da3a86d38b0faa59a3200afa9c0c873.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 391, + 295, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 295, + 426 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 391, + 295, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 391, + 295, + 426 + ], + "type": "text", + "content": " denotes a loss function (e.g., cross-entropy loss). This provides a baseline for evaluating the benefits of more sophisticated attack strategies." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 426, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 568 + ], + "type": "text", + "content": "Disagreement-based Semi-supervised Learning (StolenLoRA-DSL). While directly using synthesized data for querying and training is feasible, we propose DSL to significantly enhance both the effectiveness and efficiency of the attack. DSL improves query efficiency by selectively querying the victim model based on the substitute model's uncertainty. Furthermore, it iteratively refines both the pseudo-labels of the synthetic data and the labels obtained from querying the victim, mitigating the impact of noisy pseudo-labels and bridging the gap between the substitute and victim model, especially in cross-backbone scenarios." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": "The process begins by generating an initial synthetic dataset " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{X}^0 = \\bigcup_i\\mathbb{X}_i^0" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": " and assigning pseudo-labels based on their generating prompts without querying the victim model. These pseudo-labels, denoted as " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "c(x)" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": ", provide a starting point for training the initial substitute model " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "F_0'" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": ". Then, DSL proceeds iteratively, refining both the synthetic data for querying and the substitute model. In each iteration " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": ", a new set " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{X}_{cand}^{t} = \\bigcup_{i}\\mathbb{X}_{cand,i}^{t}" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\beta * b_{t}" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": " samples is generated using the LLM-driven Stable Diffusion process, where " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": " is a scaling factor. These candidates are then subjected to a disagreement-based filtering based on the current substitute model's predictions. For each candidate " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{X}_{cand}^{t}" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 72, + 445, + 83 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 445, + 83 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 445, + 83 + ], + "type": "text", + "content": "Algorithm 1 StolenLoRA-DSL" + } + ] + } + ], + "index": 11, + "type": "text" + }, + { + "bbox": [ + 313, + 85, + 555, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 85, + 555, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 85, + 555, + 109 + ], + "type": "text", + "content": "Input: Victim model " + }, + { + "bbox": [ + 313, + 85, + 555, + 109 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 85, + 555, + 109 + ], + "type": "text", + "content": ", Class names " + }, + { + "bbox": [ + 313, + 85, + 555, + 109 + ], + "type": "inline_equation", + "content": "\\mathbb{C}" + }, + { + "bbox": [ + 313, + 85, + 555, + 109 + ], + "type": "text", + "content": ", Query budget " + }, + { + "bbox": [ + 313, + 85, + 555, + 109 + ], + "type": "inline_equation", + "content": "b_{t}" + }, + { + "bbox": [ + 313, + 85, + 555, + 109 + ], + "type": "text", + "content": ", Scaling factor " + }, + { + "bbox": [ + 313, + 85, + 555, + 109 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 85, + 555, + 109 + ], + "type": "text", + "content": ", Threshold " + }, + { + "bbox": [ + 313, + 85, + 555, + 109 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 85, + 555, + 109 + ], + "type": "text", + "content": ", Initial sample size " + }, + { + "bbox": [ + 313, + 85, + 555, + 109 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 85, + 555, + 109 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 109, + 465, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 109, + 465, + 121 + ], + "spans": [ + { + "bbox": [ + 315, + 109, + 465, + 121 + ], + "type": "text", + "content": "Output: Substitute LoRA model " + }, + { + "bbox": [ + 315, + 109, + 465, + 121 + ], + "type": "inline_equation", + "content": "F^{\\prime}" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 320, + 123, + 553, + 171 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 320, + 123, + 438, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 123, + 438, + 135 + ], + "spans": [ + { + "bbox": [ + 320, + 123, + 438, + 135 + ], + "type": "text", + "content": "1: " + }, + { + "bbox": [ + 320, + 123, + 438, + 135 + ], + "type": "inline_equation", + "content": "\\mathbb{X}_0\\gets \\bigcup_i\\mathrm{SynIMG}(\\mathbb{C},N)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 135, + 553, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 135, + 553, + 147 + ], + "spans": [ + { + "bbox": [ + 321, + 135, + 553, + 147 + ], + "type": "text", + "content": "2: " + }, + { + "bbox": [ + 321, + 135, + 553, + 147 + ], + "type": "inline_equation", + "content": "F_0^{\\prime}\\gets \\mathrm{Update}(\\mathbb{X}_0,\\mathcal{L}_{LR})" + }, + { + "bbox": [ + 321, + 135, + 553, + 147 + ], + "type": "text", + "content": " 1 Initial substitute model" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 148, + 399, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 148, + 399, + 158 + ], + "spans": [ + { + "bbox": [ + 321, + 148, + 399, + 158 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 321, + 148, + 399, + 158 + ], + "type": "inline_equation", + "content": "t\\gets 0,B_t\\gets B" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 159, + 402, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 159, + 402, + 171 + ], + "spans": [ + { + "bbox": [ + 321, + 159, + 402, + 171 + ], + "type": "text", + "content": "4: while " + }, + { + "bbox": [ + 321, + 159, + 402, + 171 + ], + "type": "inline_equation", + "content": "B_{t} > 0" + }, + { + "bbox": [ + 321, + 159, + 402, + 171 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 318, + 123, + 553, + 365 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 320, + 171, + 553, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 171, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 320, + 171, + 553, + 182 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 320, + 171, + 553, + 182 + ], + "type": "inline_equation", + "content": "\\mathbb{X}_{\\text{cand}}^t \\gets \\text{SynIMG}(\\mathbb{C}, \\beta b_t)" + }, + { + "bbox": [ + 320, + 171, + 553, + 182 + ], + "type": "text", + "content": " Generate candidate" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 183, + 449, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 183, + 449, + 196 + ], + "spans": [ + { + "bbox": [ + 320, + 183, + 449, + 196 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 320, + 183, + 449, + 196 + ], + "type": "inline_equation", + "content": "\\mathbb{X}_{conf}^t\\gets \\emptyset ,\\mathbb{X}_{uncer}^t\\gets \\emptyset" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 196, + 421, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 196, + 421, + 209 + ], + "spans": [ + { + "bbox": [ + 321, + 196, + 421, + 209 + ], + "type": "text", + "content": "7: for " + }, + { + "bbox": [ + 321, + 196, + 421, + 209 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathbb{X}_{\\text{cand}}^t" + }, + { + "bbox": [ + 321, + 196, + 421, + 209 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 209, + 454, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 209, + 454, + 220 + ], + "spans": [ + { + "bbox": [ + 320, + 209, + 454, + 220 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 320, + 209, + 454, + 220 + ], + "type": "inline_equation", + "content": "\\hat{c},\\hat{p}\\gets \\operatorname {Predict}(F_t'(x))" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 220, + 553, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 220, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 320, + 220, + 553, + 232 + ], + "type": "text", + "content": "9: if " + }, + { + "bbox": [ + 320, + 220, + 553, + 232 + ], + "type": "inline_equation", + "content": "\\hat{c} = c(\\pmb{x})" + }, + { + "bbox": [ + 320, + 220, + 553, + 232 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 320, + 220, + 553, + 232 + ], + "type": "inline_equation", + "content": "\\hat{p} \\geq \\tau" + }, + { + "bbox": [ + 320, + 220, + 553, + 232 + ], + "type": "text", + "content": " then Disagreement filtering" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 232, + 502, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 232, + 502, + 246 + ], + "spans": [ + { + "bbox": [ + 320, + 232, + 502, + 246 + ], + "type": "text", + "content": "10: " + }, + { + "bbox": [ + 320, + 232, + 502, + 246 + ], + "type": "inline_equation", + "content": "\\mathbb{X}_{conf}^t\\gets \\mathbb{X}_{conf}^t\\cup \\{(x,c(x))\\}" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 246, + 380, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 246, + 380, + 254 + ], + "spans": [ + { + "bbox": [ + 320, + 246, + 380, + 254 + ], + "type": "text", + "content": "11: else" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 319, + 255, + 478, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 255, + 478, + 268 + ], + "spans": [ + { + "bbox": [ + 319, + 255, + 478, + 268 + ], + "type": "text", + "content": "12: " + }, + { + "bbox": [ + 319, + 255, + 478, + 268 + ], + "type": "inline_equation", + "content": "\\mathbb{X}_{uncer}^t\\gets \\mathbb{X}_{uncer}^t\\cup \\{\\pmb {x}\\}" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 319, + 269, + 389, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 269, + 389, + 279 + ], + "spans": [ + { + "bbox": [ + 319, + 269, + 389, + 279 + ], + "type": "text", + "content": "13: end if" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 319, + 281, + 380, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 281, + 380, + 290 + ], + "spans": [ + { + "bbox": [ + 319, + 281, + 380, + 290 + ], + "type": "text", + "content": "14: end for" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 319, + 291, + 553, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 291, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 319, + 291, + 553, + 304 + ], + "type": "text", + "content": "15: " + }, + { + "bbox": [ + 319, + 291, + 553, + 304 + ], + "type": "inline_equation", + "content": "\\mathbb{X}_{query}^t\\gets \\mathrm{Query}(F,\\mathbb{X}_{uncer}^t,b_t)" + }, + { + "bbox": [ + 319, + 291, + 553, + 304 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 319, + 291, + 553, + 304 + ], + "type": "text", + "content": " Selective Query" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 319, + 304, + 455, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 304, + 455, + 317 + ], + "spans": [ + { + "bbox": [ + 319, + 304, + 455, + 317 + ], + "type": "text", + "content": "16: " + }, + { + "bbox": [ + 319, + 304, + 455, + 317 + ], + "type": "inline_equation", + "content": "\\mathbb{X}_{train}^{t}\\leftarrow \\mathbb{X}_{conf}^{t}\\cup \\mathbb{X}_{query}^{t}" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 319, + 317, + 471, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 317, + 471, + 330 + ], + "spans": [ + { + "bbox": [ + 319, + 317, + 471, + 330 + ], + "type": "text", + "content": "17: " + }, + { + "bbox": [ + 319, + 317, + 471, + 330 + ], + "type": "inline_equation", + "content": "F_{t + 1}^{\\prime}\\gets \\mathrm{Update}(\\mathbb{X}_{train}^{t},\\mathcal{L}_{LR})" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 319, + 330, + 477, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 330, + 477, + 342 + ], + "spans": [ + { + "bbox": [ + 319, + 330, + 477, + 342 + ], + "type": "text", + "content": "18: " + }, + { + "bbox": [ + 319, + 330, + 477, + 342 + ], + "type": "inline_equation", + "content": "B_{t}\\gets B_{t} - |\\mathbb{X}_{query}^{t}|,t\\gets t + 1" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 319, + 342, + 375, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 342, + 375, + 351 + ], + "spans": [ + { + "bbox": [ + 319, + 342, + 375, + 351 + ], + "type": "text", + "content": "19: end while" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 318, + 353, + 374, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 353, + 374, + 365 + ], + "spans": [ + { + "bbox": [ + 318, + 353, + 374, + 365 + ], + "type": "text", + "content": "20: return " + }, + { + "bbox": [ + 318, + 353, + 374, + 365 + ], + "type": "inline_equation", + "content": "F_{t}^{\\prime}" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 391, + 555, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 555, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 555, + 488 + ], + "type": "text", + "content": "the substitute model " + }, + { + "bbox": [ + 313, + 391, + 555, + 488 + ], + "type": "inline_equation", + "content": "F_{t}^{\\prime}" + }, + { + "bbox": [ + 313, + 391, + 555, + 488 + ], + "type": "text", + "content": " predicts a class " + }, + { + "bbox": [ + 313, + 391, + 555, + 488 + ], + "type": "inline_equation", + "content": "\\hat{c}" + }, + { + "bbox": [ + 313, + 391, + 555, + 488 + ], + "type": "text", + "content": " and associated confidence " + }, + { + "bbox": [ + 313, + 391, + 555, + 488 + ], + "type": "inline_equation", + "content": "\\hat{p}" + }, + { + "bbox": [ + 313, + 391, + 555, + 488 + ], + "type": "text", + "content": ". If the predicted class " + }, + { + "bbox": [ + 313, + 391, + 555, + 488 + ], + "type": "inline_equation", + "content": "\\hat{c}" + }, + { + "bbox": [ + 313, + 391, + 555, + 488 + ], + "type": "text", + "content": " matches the prompt-based pseudo-label " + }, + { + "bbox": [ + 313, + 391, + 555, + 488 + ], + "type": "inline_equation", + "content": "c(\\boldsymbol{x})" + }, + { + "bbox": [ + 313, + 391, + 555, + 488 + ], + "type": "text", + "content": " and the confidence " + }, + { + "bbox": [ + 313, + 391, + 555, + 488 + ], + "type": "inline_equation", + "content": "\\hat{p}" + }, + { + "bbox": [ + 313, + 391, + 555, + 488 + ], + "type": "text", + "content": " exceeds a predefined threshold " + }, + { + "bbox": [ + 313, + 391, + 555, + 488 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 391, + 555, + 488 + ], + "type": "text", + "content": ", the pseudo-label is considered reliable, and the sample is added to a confidently labeled set " + }, + { + "bbox": [ + 313, + 391, + 555, + 488 + ], + "type": "inline_equation", + "content": "\\mathbb{X}_{conf}^{t} = \\{(\\boldsymbol{x}, c(\\boldsymbol{x}))\\}" + }, + { + "bbox": [ + 313, + 391, + 555, + 488 + ], + "type": "text", + "content": ". Conversely, if the prediction disagrees with the pseudo-label or the confidence is low, the sample is added to an uncertain set " + }, + { + "bbox": [ + 313, + 391, + 555, + 488 + ], + "type": "inline_equation", + "content": "\\mathbb{X}_{uncer}^{t}" + }, + { + "bbox": [ + 313, + 391, + 555, + 488 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 488, + 554, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 488, + 554, + 610 + ], + "spans": [ + { + "bbox": [ + 313, + 488, + 554, + 610 + ], + "type": "text", + "content": "From the uncertain set " + }, + { + "bbox": [ + 313, + 488, + 554, + 610 + ], + "type": "inline_equation", + "content": "\\mathbb{X}_{uncer}^{t}" + }, + { + "bbox": [ + 313, + 488, + 554, + 610 + ], + "type": "text", + "content": ", a subset of " + }, + { + "bbox": [ + 313, + 488, + 554, + 610 + ], + "type": "inline_equation", + "content": "b_{t}" + }, + { + "bbox": [ + 313, + 488, + 554, + 610 + ], + "type": "text", + "content": " samples with the lowest confidence scores are selected for querying the victim model " + }, + { + "bbox": [ + 313, + 488, + 554, + 610 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 488, + 554, + 610 + ], + "type": "text", + "content": ". These queried samples, along with their true labels obtained from " + }, + { + "bbox": [ + 313, + 488, + 554, + 610 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 488, + 554, + 610 + ], + "type": "text", + "content": ", form the query set " + }, + { + "bbox": [ + 313, + 488, + 554, + 610 + ], + "type": "inline_equation", + "content": "\\mathbb{X}_{query}^{t} = \\{(\\boldsymbol{x}, F(\\boldsymbol{x}))\\}" + }, + { + "bbox": [ + 313, + 488, + 554, + 610 + ], + "type": "text", + "content": ". Combine the confidently pseudolabeled samples " + }, + { + "bbox": [ + 313, + 488, + 554, + 610 + ], + "type": "inline_equation", + "content": "\\mathbb{X}_{conf}^{t}" + }, + { + "bbox": [ + 313, + 488, + 554, + 610 + ], + "type": "text", + "content": " and the queried samples " + }, + { + "bbox": [ + 313, + 488, + 554, + 610 + ], + "type": "inline_equation", + "content": "\\mathbb{X}_{query}^{t}" + }, + { + "bbox": [ + 313, + 488, + 554, + 610 + ], + "type": "text", + "content": " to form the training set for this iteration: " + }, + { + "bbox": [ + 313, + 488, + 554, + 610 + ], + "type": "inline_equation", + "content": "\\mathbb{X}_{train}^{t} = \\mathbb{X}_{conf}^{t} \\cup \\mathbb{X}_{query}^{t}" + }, + { + "bbox": [ + 313, + 488, + 554, + 610 + ], + "type": "text", + "content": ". Crucially, this combined set undergoes label refining during training to further address potential inaccuracies in pseudo-labels and mitigate the distribution shift." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 611, + 554, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 554, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 554, + 647 + ], + "type": "text", + "content": "Specifically, we employ a label refining strategy. Let " + }, + { + "bbox": [ + 313, + 611, + 554, + 647 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 611, + 554, + 647 + ], + "type": "text", + "content": " be the logits produced by the substitute model for a given input " + }, + { + "bbox": [ + 313, + 611, + 554, + 647 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 611, + 554, + 647 + ], + "type": "text", + "content": ". The label refining training loss is calculated as:" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 339, + 658, + 553, + 692 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 658, + 553, + 692 + ], + "spans": [ + { + "bbox": [ + 339, + 658, + 553, + 692 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {L R} (\\boldsymbol {z}, \\boldsymbol {q}) = - \\sum_ {i = 1} ^ {C} q _ {i} \\log \\left(\\frac {\\exp (z _ {i})}{\\sum_ {j = 1} ^ {C} \\exp (z _ {j})}\\right), \\tag {5}", + "image_path": "96e68903e80d332820454113adfde6e2549366c87b9b137793b653ddb880c016.jpg" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 314, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 314, + 701, + 553, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 314, + 701, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{q}" + }, + { + "bbox": [ + 314, + 701, + 553, + 714 + ], + "type": "text", + "content": " are the soft labels. They are updated using an ex" + } + ] + } + ], + "index": 40 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "266" + } + ] + } + ], + "index": 41 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 71, + 553, + 261 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 553, + 261 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 553, + 261 + ], + "type": "table", + "html": "
ScenarioMethodCUBS200Caltech256Indoor67Food101Flowers102
AccASRAccASRAccASRAccASRAccASR
Identical BackboneKnockoffNets70.9580.5485.6990.7179.1892.5982.5390.6276.2477.28
ActiveThief72.3382.1186.9292.0177.4690.5780.3488.2277.5178.56
DFME0.560.640.480.512.713.171.621.782.812.85
E371.9481.6782.3687.1881.4395.2279.6587.4680.6781.77
StolenLoRA-Rand75.3585.5487.6292.7582.3896.3379.0086.7593.7495.01
StolenLoRA-DSL73.2383.1389.3094.5382.6196.6080.5788.4787.4688.65
Cross BackboneKnockoffNets6.777.6941.3443.7641.7948.8714.4715.8914.1614.35
ActiveThief15.4217.5042.8945.4036.0442.1415.1116.5929.0929.49
DFME0.510.580.500.531.331.560.890.981.461.48
E317.5519.9248.1750.9955.8565.3140.5844.5648.2848.94
StolenLoRA-Rand45.7051.8851.7554.7859.1869.2043.1647.3959.2960.10
StolenLoRA-DSL50.1456.9265.0168.8265.0776.0944.5348.9061.1661.99
", + "image_path": "71c64f2064c1250215e8ac81899d306936db061aef670caa615b1441548e2ba1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 269, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 55, + 269, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 555, + 293 + ], + "type": "text", + "content": "Table 1. Effectiveness of StolenLoRA: the Acc (%) and ASR (%) of extracted substitute model by different model extraction attacks under 10k queries. (Boldface: the best value, Underline: the second-best value.)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 299, + 293, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 293, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 293, + 312 + ], + "type": "text", + "content": "ponential moving average of the predicted probabilities " + }, + { + "bbox": [ + 55, + 299, + 293, + 312 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 55, + 299, + 293, + 312 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 319, + 295, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 319, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 107, + 319, + 295, + 335 + ], + "type": "interline_equation", + "content": "\\boldsymbol {q} ^ {(i + 1)} = \\mu \\boldsymbol {q} ^ {(i)} + (1 - \\mu) \\boldsymbol {p} ^ {(i + 1)}, \\tag {6}", + "image_path": "aa4d2f2aba0a5e68db3fc8f9f787f09627deec9d5bf15888dc4caa965c06a481.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 343, + 296, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 296, + 405 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 296, + 405 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 343, + 296, + 405 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 343, + 296, + 405 + ], + "type": "text", + "content": " is the momentum parameter. This iterative refinement allows the model to learn from both the initial labels and its own evolving predictions, progressively improving the quality of the training data. The updated substitute model " + }, + { + "bbox": [ + 55, + 343, + 296, + 405 + ], + "type": "inline_equation", + "content": "F_{t + 1}^{\\prime}" + }, + { + "bbox": [ + 55, + 343, + 296, + 405 + ], + "type": "text", + "content": " is thus trained on " + }, + { + "bbox": [ + 55, + 343, + 296, + 405 + ], + "type": "inline_equation", + "content": "\\mathbb{X}_{train}^{t}" + }, + { + "bbox": [ + 55, + 343, + 296, + 405 + ], + "type": "text", + "content": " by minimizing:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 80, + 413, + 295, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 413, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 80, + 413, + 295, + 430 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\prime} \\left(\\Delta \\theta_ {t + 1} ^ {\\prime}\\right) = \\mathbb {E} _ {\\boldsymbol {x} \\in \\mathbb {X} _ {t r a i n} ^ {t}} \\mathcal {L} _ {L R} \\left(F _ {t + 1} ^ {\\prime} (\\boldsymbol {x}), \\boldsymbol {q} (\\boldsymbol {x})\\right). \\quad (7)", + "image_path": "5eabe1a8f668fd1a24de4e54e9a0c65d8609e829172e9e5f611a6e12bd45f7e1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 437, + 296, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 296, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 296, + 522 + ], + "type": "text", + "content": "This iterative process continues until the query budget is exhausted, resulting in a final substitute model " + }, + { + "bbox": [ + 55, + 437, + 296, + 522 + ], + "type": "inline_equation", + "content": "F'" + }, + { + "bbox": [ + 55, + 437, + 296, + 522 + ], + "type": "text", + "content": " trained on a combination of high-confidence pseudo-labeled data and a smaller set of strategically queried data. DSL allows for efficient use of the query budget by focusing on the most informative samples, leading to a more accurate substitute LoRA. We present the algorithm detail of it in Alg. 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 533, + 138, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 138, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 138, + 547 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 552, + 173, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 173, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 173, + 565 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "type": "text", + "content": "Victim Models. We use a ViT-Base model pre-trained on ImageNet-21k [33] with additional augmentations and regularization [38]. This base model is then fine-tuned with LoRA (" + }, + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "type": "inline_equation", + "content": "r = 4" + }, + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "type": "text", + "content": ") on five commonly used datasets: CUBS200 [41], Caltech256 [9], Indoor67 [31], Food101 [2], and Flowers102 [26]. The resulting victim models achieve test accuracies of " + }, + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "type": "inline_equation", + "content": "88.09\\%" + }, + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "type": "inline_equation", + "content": "94.47\\%" + }, + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "type": "inline_equation", + "content": "85.52\\%" + }, + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "type": "inline_equation", + "content": "91.07\\%" + }, + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "type": "inline_equation", + "content": "98.66\\%" + }, + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Attack Settings. We evaluate LoRA extraction attacks in Identical-Backbone (IB) and Cross-Backbone (XB) scenarios. In the IB setting, we use the same base model as the substitute base model. In the XB setting, we use" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 299, + 555, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 299, + 555, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 555, + 372 + ], + "type": "text", + "content": "the ViT-base model pre-trained on ImageNet-1k[35] with self-supervised masked autoencoder (MAE) method[12] as the substitute base model. Both substitute models utilize LoRA with " + }, + { + "bbox": [ + 313, + 299, + 555, + 372 + ], + "type": "inline_equation", + "content": "r = 4" + }, + { + "bbox": [ + 313, + 299, + 555, + 372 + ], + "type": "text", + "content": " and are trained for 20 epochs using the Adam optimizer[19] with a cosine annealing learning rate schedule[24] and a base learning rate of 0.01." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 392, + 556, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 556, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 556, + 525 + ], + "type": "text", + "content": "We use GPT-4o mini[27] as the LLM in StolenLoRA as it has similar performance at a lower price and can effectively reduce the attack cost. For the SD model, we use an open-source SDXL-Turbo model[36]. It only takes 4 sampling steps (compared to SDXL 1.0 which usually requires 40 steps) to synthesize high-quality images, significantly reducing the computational cost. For the StolenLoRA-DSL method, initial per-category sample size " + }, + { + "bbox": [ + 313, + 392, + 556, + 525 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 392, + 556, + 525 + ], + "type": "text", + "content": " is set to 10, the scaling factor " + }, + { + "bbox": [ + 313, + 392, + 556, + 525 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 392, + 556, + 525 + ], + "type": "text", + "content": " is 1.5 to not increase too much generation overhead, and the confidence threshold " + }, + { + "bbox": [ + 313, + 392, + 556, + 525 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 392, + 556, + 525 + ], + "type": "text", + "content": " is 0.95. Hyperparameter analysis is provided in later sections." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": "Baselines and Evaluation Metric. We compare Stolen-LoRA against KnockoffNets [28], ActiveThief [29], DFME [40], and " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "E^3" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": " [48] using a query budget of 10k. For methods such as KnockoffNets, ActiveThief, and " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "E^3" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": " that require real data, we use 3M images in CC3M[37] as the attack dataset. For " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "E^3" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": ", which originally selects samples based on semantic similarity between dataset category names and target class names, we improve its effectiveness by calculating semantic similarity between image captions (available in CC3M) and target class names for a more fine-grained sample selection. Following prior work [28, 40], we report Test Accuracy (Acc) and Attack Success Rate (ASR), defined as the ratio of the substitute model's accuracy to the victim model's accuracy." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "267" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 558, + 185 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 558, + 185 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 558, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 558, + 185 + ], + "type": "image", + "image_path": "fb642a10e02c21f2b5a1ad15c81bfb17c288d6bbef4cd1df8f80fd51e84d35a6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 190, + 555, + 213 + ], + "lines": [ + { + "bbox": [ + 54, + 190, + 555, + 213 + ], + "spans": [ + { + "bbox": [ + 54, + 190, + 555, + 213 + ], + "type": "text", + "content": "Figure 2. Curves of the test accuracy versus the number of queries in IB (solid line) and XB (dashed line) scenarios respectively. The gray dashed line represents the victim models' test accuracy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 219, + 180, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 219, + 180, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 219, + 180, + 232 + ], + "type": "text", + "content": "4.2. Experimental Results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 242, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 242, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 56, + 242, + 296, + 506 + ], + "type": "text", + "content": "Effectiveness of StolenLoRA. Tab. 1 presents the effectiveness of StolenLoRA, comparing both the random strategy (StolenLoRA-Rand) and DSL-based (StolenLoRADSL) variants against existing extraction attacks. Across both identical- and cross-backbone scenarios, StolenLoRA consistently achieves state-of-the-art performance. In the IB setting, StolenLoRA-Rand shows superior performance on CUBS200 (75.35%) and Flowers102 (93.74%), while StolenLoRA-DSL excels on Caltech256 (89.30%) and Indoor67 (82.61%), maintaining competitive results on other datasets. The XB setting, more challenging due to base model differences, highlights StolenLoRA's significant advantage over baselines, with StolenLoRA-DSL achieving the highest accuracy across all five datasets. This demonstrates DSL's effectiveness in refining synthetic data and improving attack efficiency. Fig. 2 further analyzes query efficiency, showing StolenLoRA-DSL's rapid convergence in the IB setting and consistent progress in the XB setting, outperforming other methods in both scenarios. This efficiency stems from DSL's ability to prioritize informative queries, focusing on uncertain samples, thereby maximizing information gain from each query." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": "Effectiveness under Hard-Label Scenario. We also evaluate StolenLoRA's performance when only hard labels (one-hot encoded predictions) are available, as is the case with some real-world APIs. Fig. 3 presents the results for both StolenLoRA-Rand and StolenLoRA-DSL in the IB and XB settings. In the IB setting, using hard labels leads to a minor performance decrease. For example, StolenLoRA-Rand's accuracy on CUBS200 drops from " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "75.35\\%" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " with soft labels to " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "67.00\\%" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " with hard labels. A similar trend is observed for other datasets and StolenLoRA-DSL. This suggests that the additional information provided by soft labels is beneficial in the IB setting. In the XB setting, the impact of hard labels is inconsistent. While a slight decrease is observed on some datasets, others, like Caltech256 and Indoor67, show a marginal increase in accuracy (from " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "51.75\\%" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "56.69\\%" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": "). This unexpected improvement might be due to the hard labels acting as a regularizer, preventing the substi" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 314, + 217, + 555, + 346 + ], + "blocks": [ + { + "bbox": [ + 314, + 217, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 314, + 217, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 314, + 217, + 555, + 346 + ], + "type": "image", + "image_path": "018d8e09783e3741ce179587f870e0b434b43444fe5b8d91a1d1c75460e10ece.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 313, + 346, + 555, + 455 + ], + "blocks": [ + { + "bbox": [ + 313, + 346, + 555, + 455 + ], + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 455 + ], + "type": "image", + "image_path": "ddd5145e5d8f3a70ae5c14228176d60b1e8b30e63fa16790db2aee0a4bc1dd4e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 461, + 555, + 505 + ], + "lines": [ + { + "bbox": [ + 313, + 461, + 555, + 505 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 555, + 505 + ], + "type": "text", + "content": "Figure 3. Comparison of StolenLoRA performance using hard labels (one-hot) versus soft labels (probabilities) from the victim model. Results are shown for both Random and DSL in identical- and cross-backbone scenarios across five datasets." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 508, + 556, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 556, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 556, + 556 + ], + "type": "text", + "content": "tute from overfitting to the synthetic data, which is more prone to distributional shifts in the XB setting. Overall, StolenLoRA exhibits robustness to hard labels in both IB and XB scenarios, demonstrating its practical applicability." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 557, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 557, + 714 + ], + "type": "text", + "content": "Ablation Study. We conduct an ablation study to analyze the contributions of the key components of StolenLoRA: the LLM-driven prompting and the DSL strategy, including the label refining loss. Tab. 2 presents the results on CUBS200 and Indoor67 datasets under both IB and XB settings with a 10k query budget. Removing the structured prompt template leads to a performance drop in both IB and XB settings, demonstrating the importance of providing structured visual information to the LLM. Removing the LLM entirely results in a more significant performance decrease, confirming the crucial role of the LLM in generating effective prompts for diverse and representative synthetic data. Furthermore, we evaluate different LLMs, observing that using" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "268" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 71, + 293, + 175 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 293, + 175 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 293, + 175 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 293, + 175 + ], + "type": "table", + "html": "
MethodIBXB
CUBS200Indoor67CUBS200Indoor67
Random75.3582.3845.7059.18
- Template72.3380.9442.2154.49
- LLM70.6674.7039.2047.09
+ Llama-3.1-8B73.9681.0446.7458.43
+ GPT-4o76.3083.1349.1963.58
DSL73.2382.6150.1465.07
- LLR70.1780.9748.2164.10
", + "image_path": "7a84661705af436a534a2ab11f13cf6115c8c4da1d1e856cb19773235e44ad4a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 220, + 293, + 316 + ], + "blocks": [ + { + "bbox": [ + 55, + 183, + 295, + 217 + ], + "lines": [ + { + "bbox": [ + 55, + 183, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 183, + 295, + 217 + ], + "type": "text", + "content": "Table 2. An ablation experiment showing the effectiveness of the two modules we designed on CUBS200 and Indoor67 datasets under both IB and XB settings with 10k queries." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 220, + 293, + 316 + ], + "lines": [ + { + "bbox": [ + 59, + 220, + 293, + 316 + ], + "spans": [ + { + "bbox": [ + 59, + 220, + 293, + 316 + ], + "type": "table", + "html": "
τIBXB
CUBS200Indoor67CUBS200Indoor67
0.570.6878.8148.4360.90
0.771.3780.0048.3663.51
0.974.0481.1948.5263.06
0.9573.2382.6150.1465.07
0.9973.3181.1949.8661.72
", + "image_path": "f91a2cc6d0bfa574c916752d1b9b14b0888bc2ba8351aa6e5545b9ad0e8d3a6c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 365, + 295, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 413 + ], + "type": "text", + "content": "a more powerful LLM like GPT-4o generally yields better performance than Llama-3.1-8B[5], particularly in the more challenging XB setting. This highlights the benefit of utilizing a stronger LLM for prompt generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 414, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 414, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 414, + 295, + 510 + ], + "type": "text", + "content": "Disabling DSL and reverting to random querying results in a noticeable performance reduction, especially in the XB setting. This underscores the effectiveness of DSL in selectively querying informative samples. Similarly, removing SAT from the DSL framework also leads to a performance decrease, albeit less pronounced than removing DSL altogether. This indicates the benefit of label refining in mitigating the distribution shift between synthetic and real data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 510, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 629 + ], + "type": "text", + "content": "We also analyze the impact of the confidence threshold " + }, + { + "bbox": [ + 55, + 510, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 510, + 295, + 629 + ], + "type": "text", + "content": " within DSL in Tab. 3. The results show that performance generally improves as " + }, + { + "bbox": [ + 55, + 510, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 510, + 295, + 629 + ], + "type": "text", + "content": " increases from 0.5 to 0.95. This indicates that focusing queries on samples where the substitute model is less confident leads to a more efficient use of the query budget. However, increasing " + }, + { + "bbox": [ + 55, + 510, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 510, + 295, + 629 + ], + "type": "text", + "content": " further to 0.99 results in a slight performance decrease, suggesting that an overly stringent threshold can exclude valuable samples from being queried. We therefore select " + }, + { + "bbox": [ + 55, + 510, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\tau = 0.95" + }, + { + "bbox": [ + 55, + 510, + 295, + 629 + ], + "type": "text", + "content": " as the optimal value for our experiments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": "Fidelity and Distribution Analysis of Synthetic Data. StolenLoRA's effectiveness relies on the quality of its synthetic training data. We evaluate this quality both visually and quantitatively. Visually, as shown in Fig. 4, synthetic images generated by StolenLoRA demonstrate a striking resemblance to real images from the target categories, capturing key features and background context crucial for train-" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 322, + 70, + 549, + 175 + ], + "blocks": [ + { + "bbox": [ + 322, + 70, + 549, + 175 + ], + "lines": [ + { + "bbox": [ + 322, + 70, + 549, + 175 + ], + "spans": [ + { + "bbox": [ + 322, + 70, + 549, + 175 + ], + "type": "image", + "image_path": "e3667a820246873db8cca7a1f6f2627b73ecfc3481137a2156d4a902a93990f6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 385, + 177, + 484, + 186 + ], + "lines": [ + { + "bbox": [ + 385, + 177, + 484, + 186 + ], + "spans": [ + { + "bbox": [ + 385, + 177, + 484, + 186 + ], + "type": "text", + "content": "(a) Visualization on CUBS200" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 321, + 187, + 549, + 293 + ], + "blocks": [ + { + "bbox": [ + 321, + 187, + 549, + 293 + ], + "lines": [ + { + "bbox": [ + 321, + 187, + 549, + 293 + ], + "spans": [ + { + "bbox": [ + 321, + 187, + 549, + 293 + ], + "type": "image", + "image_path": "dcc1be32d6ec2ff6484cb54d57bcbe9b990544315cb4489d539c0b297009618f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 295, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 386, + 295, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 386, + 295, + 482, + 304 + ], + "type": "text", + "content": "(b) Visualization on Indoor67" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 306, + 553, + 328 + ], + "lines": [ + { + "bbox": [ + 313, + 306, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 553, + 328 + ], + "type": "text", + "content": "Figure 4. Visualization of StolenLoRA's synthetic images and corresponding real images of the target category." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 317, + 332, + 552, + 392 + ], + "blocks": [ + { + "bbox": [ + 55, + 324, + 295, + 358 + ], + "lines": [ + { + "bbox": [ + 55, + 324, + 295, + 358 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 295, + 358 + ], + "type": "text", + "content": "Table 3. Performance of StolenLoRA-DSL with varying confidence thresholds " + }, + { + "bbox": [ + 55, + 324, + 295, + 358 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 324, + 295, + 358 + ], + "type": "text", + "content": " . Results are reported on CUBS200 and Indoor67 under both IB and XB settings with 10k queries." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 332, + 552, + 392 + ], + "lines": [ + { + "bbox": [ + 317, + 332, + 552, + 392 + ], + "spans": [ + { + "bbox": [ + 317, + 332, + 552, + 392 + ], + "type": "table", + "html": "
MethodCUBS200Caltech256Indoor67Food101Flowers102
KnockoffNets51.448.9422.0150.0455.77
ActiveThief49.336.0115.7247.8450.19
DFME220.53179.45225.84188.29204.79
E341.574.296.6825.5134.82
StolenLoRA2.145.596.9319.5416.63
", + "image_path": "701e56bcc741396b44a45f0cda81d2fc6e31e764ccbfe198acc0adf01fb33c73.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 394, + 553, + 417 + ], + "lines": [ + { + "bbox": [ + 313, + 394, + 553, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 553, + 417 + ], + "type": "text", + "content": "Table 4. FID scores " + }, + { + "bbox": [ + 313, + 394, + 553, + 417 + ], + "type": "inline_equation", + "content": "(\\downarrow)" + }, + { + "bbox": [ + 313, + 394, + 553, + 417 + ], + "type": "text", + "content": " between the datasets used by each attack method and the victim model's training data distribution." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 426, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 544 + ], + "type": "text", + "content": "ing. This visual fidelity is corroborated by significantly lower Fréchet Inception Distance (FID) scores[13] (in Tab. 4) compared to other attack methods. For example, StolenLoRA achieves an FID of 2.14 compared to 51.44 for KnockoffNets on CUBS200, highlighting a much closer distributional match to the target domain. This demonstrates that StolenLoRA's LLM-driven synthesis effectively generates high-fidelity, domain-specific data, even without access to the original training set, which is key to its superior extraction performance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": "Verifying the Distinction Hypothesis. As mentioned in Sec. 3.1, we hypothesize that the performance of a perfectly extracted surrogate model will be consistent across In-Distribution (ID) data with the victim model, but not necessarily Out-of-Distribution (OOD) data, motivating us to perform extraction attacks using ID data. To empirically verify this conjecture, we train a substitute model using the XB setting on the victim model's training dataset, aiming to achieve a near-perfect extraction of the LoRA adaptations. We then evaluate both the victim model and the extracted substitute on two sets of data: ID data, represented by the victim model's held-out test set, and OOD data, sampled from the CC3M dataset. For each sample, we calculate the cross-entropy between the victim model's predictions and" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "269" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 294, + 186 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 294, + 186 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 294, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 294, + 186 + ], + "type": "image", + "image_path": "a8effbcfe67e86eaaf5ffbcb4f61507356f7e1869bd230eb26787cc60792d0c0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 193, + 296, + 239 + ], + "lines": [ + { + "bbox": [ + 55, + 193, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 296, + 239 + ], + "type": "text", + "content": "Figure 5. Cross-entropy difference between the victim model and a near-perfectly substitute model on In-Distribution (ID) and Out-of-Distribution (OOD) data. The substitute model is trained using the cross-backbone setting on the victim's training dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 244, + 296, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 244, + 296, + 400 + ], + "spans": [ + { + "bbox": [ + 54, + 244, + 296, + 400 + ], + "type": "text", + "content": "the substitute model's predictions, providing a measure of their functional divergence. Fig. 5 presents the distribution of these differences. On both CUBS200 and Indoor67, the difference is smaller for the ID data compared to the OOD data. This observation supports our conjecture that even near-perfect LoRA extraction yields a substitute model that exhibits functional equivalence primarily within the ID domain. The divergence in OOD data likely stems from the inherent differences between the pre-trained backbone used by the victim and the substitute. This result underscores the importance of utilizing ID data for effective LoRA extraction, as the attacker primarily seeks to replicate the victim's specialized adaptations rather than the general capabilities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 411, + 257, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 257, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 257, + 424 + ], + "type": "text", + "content": "5. Defending Against LoRA Extraction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 430, + 295, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 295, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 295, + 477 + ], + "type": "text", + "content": "While this paper primarily focuses on the feasibility of LoRA extraction, we also explore a potential defense mechanism based on deploying multiple, diverse LoRA adapters to increase attacker uncertainty." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 478, + 295, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 295, + 574 + ], + "type": "text", + "content": "Dual LoRA with Diversified Predictions. Our proposed defense involves training two LoRA adapters, denoted as " + }, + { + "bbox": [ + 55, + 478, + 295, + 574 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_A" + }, + { + "bbox": [ + 55, + 478, + 295, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 478, + 295, + 574 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_B" + }, + { + "bbox": [ + 55, + 478, + 295, + 574 + ], + "type": "text", + "content": ", to maximize the difference in their output distributions while maintaining comparable performance on the target task. Instead of minimizing accuracy as initially conceived, we found that maintaining high accuracy while maximizing divergence is crucial for practical deployment. We achieve this by optimizing the following loss function:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 87, + 583, + 295, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 583, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 87, + 583, + 295, + 597 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\prime} = \\mathcal {L} (\\mathrm {L} _ {A}) + \\mathcal {L} (\\mathrm {L} _ {B}) - \\lambda \\times \\mathrm {K L} (\\mathrm {L} _ {A} | | \\mathrm {L} _ {B}) \\tag {8}", + "image_path": "8a6e774c44519a8e6ed579d6bbfb8f68912a40582ea5e4f070d2d5122a93e840.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " represents the cross-entropy loss for the target task, KL represents the Kullback-Leibler divergence between the output distributions of the two LoRAs, and " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " is a hyperparameter controlling the balance between task performance and divergence. At deployment, we randomly select either " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "L_{A}" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "L_{B}" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " for each incoming query to generate the prediction. This random selection introduces uncertainty and makes it challenging for the attacker to extract the underlying LoRA function consistently." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 314, + 69, + 553, + 186 + ], + "blocks": [ + { + "bbox": [ + 314, + 69, + 553, + 186 + ], + "lines": [ + { + "bbox": [ + 314, + 69, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 314, + 69, + 553, + 186 + ], + "type": "image", + "image_path": "ac1b89e4554fa8412367296f397452b438629a1081ed55ee1b1f02be72bd8252.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 193, + 553, + 225 + ], + "lines": [ + { + "bbox": [ + 313, + 193, + 553, + 225 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 553, + 225 + ], + "type": "text", + "content": "Figure 6. Defneder Acc versus Substitute model Acc trade-off for defenses evaluated against StolenLoRA and KnockoffNets on the CUBS200 dataset with 10k queries." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 226, + 555, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 226, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 226, + 555, + 441 + ], + "type": "text", + "content": "Experimental Evaluation. We evaluate the proposed defense mechanism under both IB and XB settings with 10k queries. Fig. 6 illustrates the trade-off between defender Acc and substitute model Acc. The defense consistently degrades the substitute model's performance. KnockoffNets experiences substantial accuracy drops (from " + }, + { + "bbox": [ + 313, + 226, + 555, + 441 + ], + "type": "inline_equation", + "content": "70.90\\%" + }, + { + "bbox": [ + 313, + 226, + 555, + 441 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 226, + 555, + 441 + ], + "type": "inline_equation", + "content": "20.78\\%" + }, + { + "bbox": [ + 313, + 226, + 555, + 441 + ], + "type": "text", + "content": " in IB and from " + }, + { + "bbox": [ + 313, + 226, + 555, + 441 + ], + "type": "inline_equation", + "content": "6.77\\%" + }, + { + "bbox": [ + 313, + 226, + 555, + 441 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 226, + 555, + 441 + ], + "type": "inline_equation", + "content": "1.21\\%" + }, + { + "bbox": [ + 313, + 226, + 555, + 441 + ], + "type": "text", + "content": " in XB), and StolenLoRA-Rand shows similar vulnerability (decreasing " + }, + { + "bbox": [ + 313, + 226, + 555, + 441 + ], + "type": "inline_equation", + "content": "22.95\\%" + }, + { + "bbox": [ + 313, + 226, + 555, + 441 + ], + "type": "text", + "content": " in IB and " + }, + { + "bbox": [ + 313, + 226, + 555, + 441 + ], + "type": "inline_equation", + "content": "15.63\\%" + }, + { + "bbox": [ + 313, + 226, + 555, + 441 + ], + "type": "text", + "content": " in XB). However, StolenLoRADSL demonstrates greater resilience, with comparatively smaller accuracy reductions (" + }, + { + "bbox": [ + 313, + 226, + 555, + 441 + ], + "type": "inline_equation", + "content": "9.63\\%" + }, + { + "bbox": [ + 313, + 226, + 555, + 441 + ], + "type": "text", + "content": " in IB and " + }, + { + "bbox": [ + 313, + 226, + 555, + 441 + ], + "type": "inline_equation", + "content": "14.34\\%" + }, + { + "bbox": [ + 313, + 226, + 555, + 441 + ], + "type": "text", + "content": " in XB). This resilience is attributed to DSL's efficient query selection and label refining, focusing on informative samples and mitigating the uncertainty introduced by the defense. While this defense effectively hinders less sophisticated attacks, it underscores the need for more robust mechanisms to counter advanced techniques like DSL that can learn effectively from limited queries." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 453, + 388, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 453, + 388, + 466 + ], + "spans": [ + { + "bbox": [ + 314, + 453, + 388, + 466 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "type": "text", + "content": "This paper introduces LoRA Extraction, a novel model extraction attack targeting the widespread practice of adapting large vision models with LoRA. We present StolenLoRA, a highly effective method leveraging LLM-driven Stable Diffusion to generate task-specific, high-fidelity synthetic training data, bypassing the limitations of traditional extraction techniques reliant on large-scale real datasets or unreliable GANs. Furthermore, StolenLoRA incorporates Disagreement-based Semi-supervised Learning (DSL) to efficiently query the victim model, focusing on uncertain predictions and iteratively refining labels, enabling successful extraction with limited queries. Extensive experiments demonstrate StolenLoRA's strong performance across diverse datasets and in challenging cross-backbone scenarios. We also explore a preliminary defense based on diversified LoRA deployments, showing promise in mitigating these attacks. This work represents a crucial first step towards understanding and addressing the safety risks associated with LoRA adaptation, paving the way for more secure deployments of efficient fine-tuning methods." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "270" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 137 + ], + "type": "text", + "content": "This work is in part supported by National Key R&D Program of China (Grant No. 2022ZD0160103) and National Natural Science Foundation of China (Grant No. 62276067), and Shanghai Artificial Intelligence Laboratory." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 156, + 115, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 156, + 115, + 169 + ], + "spans": [ + { + "bbox": [ + 56, + 156, + 115, + 169 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 176, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 176, + 296, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 176, + 296, + 209 + ], + "spans": [ + { + "bbox": [ + 61, + 176, + 296, + 209 + ], + "type": "text", + "content": "[1] Martin Arjovsky and Léon Bottou. Towards principled methods for training generative adversarial networks. arXiv preprint arXiv:1701.04862, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 210, + 296, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 210, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 61, + 210, + 296, + 243 + ], + "type": "text", + "content": "[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In ECCV, 2014. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 244, + 296, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 244, + 296, + 287 + ], + "spans": [ + { + "bbox": [ + 61, + 244, + 296, + 287 + ], + "type": "text", + "content": "[3] Eric L Buehler and Markus J Buehler. X-lora: Mixture of low-rank adapter experts, a flexible framework for large language models with applications in protein mechanics and molecular design. APL Machine Learning, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 289, + 295, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 289, + 295, + 320 + ], + "spans": [ + { + "bbox": [ + 61, + 289, + 295, + 320 + ], + "type": "text", + "content": "[4] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 322, + 296, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 322, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 61, + 322, + 296, + 376 + ], + "type": "text", + "content": "[5] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 378, + 296, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 378, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 61, + 378, + 296, + 421 + ], + "type": "text", + "content": "[6] Ricard Durall, Avraam Chatzimichailidis, Peter Labus, and Janis Keuper. Combating mode collapse in gan training: An empirical analysis using hessian eigenvalues. arXiv preprint arXiv:2012.09673, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 422, + 296, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 422, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 422, + 296, + 456 + ], + "type": "text", + "content": "[7] Lijie Fan, Kaifeng Chen, Dilip Krishnan, Dina Katabi, Phillip Isola, and Yonglong Tian. Scaling laws of synthetic images for model training... for now. In CVPR, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 456, + 296, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 456, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 62, + 456, + 296, + 498 + ], + "type": "text", + "content": "[8] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NIPS, 2014. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 501, + 296, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 501, + 296, + 522 + ], + "spans": [ + { + "bbox": [ + 62, + 501, + 296, + 522 + ], + "type": "text", + "content": "[9] Gregory Griffin, Alex Holub, and Pietro Perona. Caltech-256 object category dataset. 2007. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 524, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 524, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 57, + 524, + 295, + 555 + ], + "type": "text", + "content": "[10] Zeyu Han, Chao Gao, Jinyang Liu, Sai Qian Zhang, et al. Parameter-efficient fine-tuning for large models: A comprehensive survey. arXiv preprint arXiv:2403.14608, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 557, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 557, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 557, + 296, + 590 + ], + "type": "text", + "content": "[11] Soufiane Hayou, Nikhil Ghosh, and Bin Yu. Lora+: Efficient low rank adaptation of large models. arXiv preprint arXiv:2402.12354, 2024. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 591, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 624 + ], + "type": "text", + "content": "[12] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 1, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 625, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 668 + ], + "type": "text", + "content": "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[14] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In ICML, 2019. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "text", + "content": "[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "type": "text", + "content": "[16] Nam Hyeon-Woo, Moon Ye-Bin, and Tae-Hyun Oh. Fedpara: Low-rank hadamard product for communication-efficient federated learning. arXiv preprint arXiv:2108.06098, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 163, + 554, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 554, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 554, + 205 + ], + "type": "text", + "content": "[17] Matthew Jagielski, Nicholas Carlini, David Berthelot, Alex Kurakin, and Nicolas Papernot. High accuracy and high fidelity extraction of neural networks. In *Usenix Security*, 2020. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 207, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 553, + 240 + ], + "type": "text", + "content": "[18] Sanjay Kariyappa, Atul Prakash, and Moinuddin Qureshi. Maze: Data-free model stealing attack using zeroth-order gradient estimation. In CVPR, 2021. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 241, + 553, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 263 + ], + "type": "text", + "content": "[19] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 265, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 265, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 265, + 553, + 308 + ], + "type": "text", + "content": "[20] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In CVPR, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "type": "text", + "content": "[21] Youssef Kossale, Mohammed Airaj, and Aziz Darouichi. Mode collapse in generative adversarial networks: An overview. In ICOA, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "type": "text", + "content": "[22] Zi Liang, Qingqing Ye, Yanyun Wang, Sen Zhang, Yaxin Xiao, Ronghua Li, Jianliang Xu, and Haibo Hu. Alignment-aware model extraction attacks on large language models. arXiv preprint arXiv:2409.02718, 2024. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 388, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 421 + ], + "type": "text", + "content": "[23] Zijun Lin, Ke Xu, Chengfang Fang, Huadi Zheng, Aneez Ahmed Jaheezuddin, and Jie Shi. Quda: Query-limited data-free model extraction. In ACM Asia CCS, 2023. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 422, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 443 + ], + "type": "text", + "content": "[24] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In *ICLR*, 2017. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 445, + 554, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 554, + 498 + ], + "type": "text", + "content": "[25] Xingjun Ma, Yifeng Gao, Yixu Wang, Ruofan Wang, Xin Wang, Ye Sun, Yifan Ding, Hengyuan Xu, Yunhao Chen, Yunhan Zhao, et al. Safety at scale: A comprehensive survey of large model safety. arXiv preprint arXiv:2502.05206, 2025.1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 501, + 554, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 554, + 544 + ], + "type": "text", + "content": "[26] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian conference on computer vision, graphics & image processing, 2008. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 546, + 554, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 546, + 554, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 546, + 554, + 589 + ], + "type": "text", + "content": "[27] OpenAI. Gpt-4o mini: Advancing cost-efficient intelligence. https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence/, 2024.5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 591, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 591, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 553, + 623 + ], + "type": "text", + "content": "[28] Tribhuvanesh Orekondy, Bernt Schiele, and Mario Fritz. Knockoff nets: Stealing functionality of black-box models. In CVPR, 2019. 1, 2, 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 624, + 554, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 554, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 554, + 668 + ], + "type": "text", + "content": "[29] Soham Pal, Yash Gupta, Aditya Shukla, Aditya Kanade, Shirish Shevade, and Vinod Ganapathy. Activethief: Model extraction using active learning and unannotated public data. In AAAI, 2020. 1, 2, 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "type": "text", + "content": "[30] Jie Qin, Jie Wu, Weifeng Chen, Yuxi Ren, Huixia Li, Hefeng Wu, Xuefeng Xiao, Rui Wang, and Shilei Wen. Diffusiongpt: Llm-driven text-to-image generation system. arXiv preprint arXiv:2401.10061, 2024. 2" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "271" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 95 + ], + "type": "text", + "content": "[31] Ariadna Quattoni and Antonio Torralba. Recognizing indoor scenes. In CVPR, 2009. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 139 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 139 + ], + "type": "text", + "content": "[32] Pengzhen Ren, Yun Xiao, Xiaojun Chang, Po-Yao Huang, Zhihui Li, Brij B Gupta, Xiaojiang Chen, and Xin Wang. A survey of deep active learning. ACM computing surveys, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 141, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 141, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 57, + 141, + 294, + 174 + ], + "type": "text", + "content": "[33] Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik-Manor. Imagenet-21k pretraining for the masses. arXiv preprint arXiv:2104.10972, 2021. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 175, + 295, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 175, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 57, + 175, + 295, + 207 + ], + "type": "text", + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 209, + 295, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 209, + 295, + 252 + ], + "spans": [ + { + "bbox": [ + 57, + 209, + 295, + 252 + ], + "type": "text", + "content": "[35] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 2015. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 254, + 295, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 254, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 57, + 254, + 295, + 285 + ], + "type": "text", + "content": "[36] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 287, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 287, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 57, + 287, + 294, + 330 + ], + "type": "text", + "content": "[37] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 332, + 295, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 332, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 57, + 332, + 295, + 376 + ], + "type": "text", + "content": "[38] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. arXiv preprint arXiv:2106.10270, 2021. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 377, + 295, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 377, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 57, + 377, + 295, + 410 + ], + "type": "text", + "content": "[39] Florian Tramér, Fan Zhang, Ari Juels, Michael K. Reiter, and Thomas Ristenpart. Stealing machine learning models via prediction apis. In USENIX Security Symposium, 2016. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 411, + 295, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 411, + 295, + 442 + ], + "spans": [ + { + "bbox": [ + 57, + 411, + 295, + 442 + ], + "type": "text", + "content": "[40] Jean-Baptiste Truong, Pratyush Maini, Robert J Walls, and Nicolas Papernot. Data-free model extraction. In CVPR, 2021. 1, 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 445, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 445, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 57, + 445, + 294, + 476 + ], + "type": "text", + "content": "[41] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 478, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 478, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 57, + 478, + 295, + 521 + ], + "type": "text", + "content": "[42] Yixu Wang, Jie Li, Hong Liu, Yan Wang, Yongjian Wu, Feiyue Huang, and Rongrong Ji. Black-box dissector: Towards erasing-based hard-label model stealing attack. In ECCV, 2022. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 523, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 523, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 57, + 523, + 295, + 555 + ], + "type": "text", + "content": "[43] Yixu Wang, Tianle Gu, Yan Teng, Yingchun Wang, and Xingjun Ma. Honeypotnet: Backdoor attacks against model extraction. In AAAI, 2025. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 557, + 295, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 557, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 57, + 557, + 295, + 600 + ], + "type": "text", + "content": "[44] Shih-Ying Yeh, Yu-Guan Hsieh, Zhidong Gao, Bernard BW Yang, Giyeong Oh, and Yanmin Gong. Navigating text-to-image customization: From lycoris fine-tuning to model evaluation. In ICLR, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 602, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 635 + ], + "type": "text", + "content": "[45] Jianhao Yuan, Jie Zhang, Shuyang Sun, Philip Torr, and Bo Zhao. Real-fake: Effective training data synthesis through distribution matching. In ICLR, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 636, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 678 + ], + "type": "text", + "content": "[46] Yunlong Zhao, Xiaoheng Deng, Yijing Liu, Xinjun Pei, Ji-azhi Xia, and Wei Chen. Fully exploiting every real sample: Superpixel sample gradient model stealing. In CVPR, 2024. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "type": "text", + "content": "[47] Mingyi Zhou, Jing Wu, Yipeng Liu, Shuaicheng Liu, and Ce Zhu. Dast: Data-free substitute training for adversarial attacks. In CVPR, 2020. 1, 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 554, + 106 + ], + "type": "text", + "content": "[48] Hongyu Zhu, Wentao Hu, Sichu Liang, Fangqi Li, Wenwen Wang, and Shilin Wang. Efficient and effective model extraction. arXiv preprint arXiv:2409.14122, 2024. 5" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "272" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Straighten Viscous Rectified Flow via Noise Optimization/7bcafc85-3856-49c2-a845-9dc4fd667b05_content_list.json b/2025/Straighten Viscous Rectified Flow via Noise Optimization/7bcafc85-3856-49c2-a845-9dc4fd667b05_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b7e1668297c47005584188febd41f4e99f7c64c3 --- /dev/null +++ b/2025/Straighten Viscous Rectified Flow via Noise Optimization/7bcafc85-3856-49c2-a845-9dc4fd667b05_content_list.json @@ -0,0 +1,1628 @@ +[ + { + "type": "text", + "text": "Straighten Viscous Rectified Flow via Noise Optimization", + "text_level": 1, + "bbox": [ + 207, + 130, + 789, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jimin Dai1, Jiexi Yan2, Jian Yang1, Lei Luo1 \n1PCA Lab, Nanjing University of Science and Technology, 2Xidian University {jimindai, csjyang}@njust.edu.cn, {jxyan1995, luoleipitt}@gmail.com", + "bbox": [ + 189, + 179, + 807, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 285 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The Reflow operation aims to straighten the inference trajectories of the rectified flow during training by constructing deterministic couplings between noises and images, thereby improving the quality of generated images in single-step or few-step generation. However, we identify critical limitations in Reflow, particularly its inability to rapidly generate high-quality images due to a distribution gap between images in its constructed deterministic couplings and real images. To address these shortcomings, we propose a novel alternative called Straighten Viscous Rectified Flow via Noise Optimization (VRFNO), which is a joint training framework integrating an encoder and a neural velocity field. VRFNO introduces two key innovations: (1) a historical velocity term that enhances trajectory distinction, enabling the model to more accurately predict the velocity of the current trajectory, and (2) the noise optimization through reparameterization to form optimized couplings with real images which are then utilized for training, effectively mitigating errors caused by Reflow's limitations. Comprehensive experiments on synthetic data and real datasets with varying resolutions show that VRFNO significantly mitigates the limitations of Reflow, achieving state-of-the-art performance in both one-step and few-step generation tasks.", + "bbox": [ + 91, + 301, + 483, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 678, + 220, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models (DMs) [9, 23, 25] have attracted significant attention due to their remarkable generative capabilities and stable optimization process. DMs have outperformed traditional models such as generative adversarial networks (GANs) [1, 4] and Variational autoencoders (VAEs) [30] in various generation tasks. DMs work by designing a forward noise-adding scheme, which progressively transforms the target distribution into the initial distribution (usually a Gaussian distribution), thereby constructing a Probability Flow (PF) from complex to simple distributions. DMs learn the inverse process of PF, and then they can sample Gaussian noises and generate samples that approximate the target distribution through multi", + "bbox": [ + 89, + 703, + 482, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ple inference steps. Currently, various forward noise-adding schemes have been developed for DMs. Among them, the most straightforward approach is the Rectified Flow (RF) [14, 15, 22]. The core idea of RF is to add noise to the data in a linear manner, constructing a deterministic PF from the Gaussian distribution to the target distribution. The model only needs to learn the velocity field of the straight-line interpolation trajectories between the noises and the samples. Once trained, the model can efficiently generate samples that approximate the target distribution.", + "bbox": [ + 511, + 270, + 906, + 421 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generating samples with DMs typically requires multiple inference steps that consume substantial computational resources and time. RF constructs straight-line interpolation trajectories between noises and samples as the reference, and it theoretically learns a straight flow from the Gaussian distribution to the target distribution. The straight flow implies that RF learns a constant velocity field, which enables few-step or even single-step sampling in an ideal scenario. However, during actual training, the RF struggles to learn a constant velocity field, and the resulting PF trajectories remain curved. Therefore, RF still relies on multiple steps to generate the desired images. RF attributes the curvature to the crossing of reference trajectories during training and employs a Reflow operation to mitigate this issue.", + "bbox": [ + 511, + 424, + 908, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we observe that during the actual training process, the probability of crossing between the straight-line interpolation trajectories of randomly matched noises and images at a given time step $t$ is $P \\sim O(e^{-c(n \\times n)})$ , which is very small in high-dimensional data space (Theorem 1). Therefore, we are more inclined to believe that the curvature arises because the neural network model is not an exact solver, intermediate states along different trajectories may become \"approximately crossing\" due to their high similarity in statistical property, thereby interfering with the model predictions. Furthermore, we conducted an in-depth analysis of the factors that contribute to the success of Reflow, as well as its limitations, detailed in Section 2.2.", + "bbox": [ + 511, + 638, + 908, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To retain the advantages of Reflow while mitigating its drawbacks, we propose a new method called Straighten Viscous Rectified Flow via Noise Optimization (VRFNO). VRFNO introduces a historical velocity term (HVT) as in", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15005", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "put to the neural velocity field, allowing the model to distinguish the direction of the RF in the presence of approximate crossing. By incorporating historical velocity information, the model can more accurately predict the velocity. Moreover, VRFNO directly uses the original dataset to avoid the distribution gap and achieves more appropriate matches between the noises and the images (we refer to it as optimized coupling) by noise optimization. Specifically, we propose a joint training framework for an encoder and a neural velocity field. The encoder first encodes images from randomly matched noise-image pairs and then outputs the corresponding mean and variance matrices. The mean and variance matrices are combined with the noise using the reparameterization technique to generate a new noise, which is then fed into the neural velocity field for training. This joint training approach enables the encoder to adjust the mean and variance of the noise based on the image, thereby optimizing the matches between the noises and the images, which in turn achieves optimized couplings. By using these optimized couplings to train the velocity field, the model can more efficiently straighten the inference trajectory during training, enabling one-step or few-step generation.", + "bbox": [ + 89, + 90, + 483, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 109, + 424, + 410, + 436 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a new method VRFNO to straighten the inference trajectories for higher quality few-step and single-step generation.", + "- We introduce an HVT as the auxiliary information to enhance the model's accuracy in velocity prediction when the flow approximates crossing.", + "- We construct optimized couplings by optimizing the noise to train the neural velocity field. This avoids relying on deterministic couplings used in Reflow, which will constrain the model's generation quality." + ], + "bbox": [ + 89, + 438, + 483, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 602, + 217, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Rectified Flow", + "text_level": 1, + "bbox": [ + 89, + 627, + 236, + 642 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "RF is an ordinary differential equations (ODE)-based DM that aims to learn the mapping between two distributions, $\\pi_0$ and $\\pi_{1}$ . In image generation tasks, $\\pi_0$ typically represents a standard Gaussian distribution, while $\\pi_{1}$ corresponds to the target image distribution. For empirical observations $X_0\\sim$ $\\pi_0$ and $X_{1}\\sim \\pi_{1}$ on time $t\\in [0,1]$ , RF is defined as:", + "bbox": [ + 89, + 648, + 483, + 739 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nd X _ {t} = v \\left(X _ {t}, t\\right) d t, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 752, + 482, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where intermediate state $X_{t} = tX_{1} + (1 - t)X_{0}$ represents a time-differentiable interpolation between $X_{0}$ and $X_{1}$ , $v: \\mathbb{R}^{n\\times n}\\times [0,1]\\to \\mathbb{R}^{n\\times n}$ denotes the velocity field defined on the data-time domain. Since RF uses simple straight-line interpolation to connect $X_{0}$ and $X_{1}$ , its trajectory has a constant velocity field $v_{ref} = X_1 - X_0$ . Thus, the training process optimizes the model by solving a least-squares regression problem, i.e., fitting $v_{ref}$ to neural ve", + "bbox": [ + 89, + 779, + 483, + 902 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d51084fa122bfcd920a78e0cdf4d9fc6dffce3275c9706661f3913e1e0e76166.jpg", + "image_caption": [ + "Figure 1. Comparison between Reflow and Vanilla training mode. The top illustrates the Reflow training mode, where noises generate images via a pre-trained model, forming deterministic couplings for training. These couplings are reused during training by randomly sampling different intermediate states along the trajectories. The bottom describes the vanilla training mode, where noises and images are randomly sampled to form arbitrary couplings for training, without reusing data during training." + ], + "image_footnote": [], + "bbox": [ + 517, + 89, + 903, + 296 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "locity field $v_{\\theta}$", + "bbox": [ + 511, + 434, + 612, + 450 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\min \\mathbb {E} _ {X _ {0}, X _ {1} \\sim \\gamma , t \\in p (t)} \\left[ \\| v _ {\\text {r e f}} - v _ {\\theta} \\left(X _ {t}, t\\right) \\| ^ {2} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 459, + 903, + 486 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\gamma$ represents the coupling of $(X_0, X_1)$ , $p(t)$ denotes the time distribution defined on $[0, 1]$ . During the inference process, the ODE usually needs to be discretized and solved via the Euler method, expressed as:", + "bbox": [ + 511, + 494, + 906, + 556 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nX _ {t + \\Delta t} = X _ {t} + \\Delta t \\cdot v _ {\\theta} \\left(X _ {t}, t\\right), t \\in \\{0, \\Delta t, \\dots , (N - 1) \\Delta t \\}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 566, + 903, + 597 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\Delta t = \\frac{1}{N}$ , $N$ represents the total number of steps. Generally, a larger $N$ results in higher-quality generated images, but it requires more computational resources for sampling. The reference trajectory of RF is a straight flow. If RF can converge to the straight flow through training, it can significantly minimize numerical errors in the ODE solver, enabling few-step or single-step generation: as the state moves along the trajectory with uniform linear motion, the velocity at a single time step equals the average velocity over the entire trajectory. As long as the cumulative sum of all time increments throughout the motion equals unity, the final endpoint of the trajectory will remain consistent.", + "bbox": [ + 511, + 598, + 906, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Analysis of Reflow", + "text_level": 1, + "bbox": [ + 511, + 787, + 691, + 803 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The inference trajectory of RF exhibits a curved pattern when trained using the vanilla method. While RF attributes this phenomenon to the crossing of reference trajectories [13] and proposes Reflow as a solution (with implementation details provided in Appendix C), our experiments (Fig.2) and analysis suggest that the effectiveness", + "bbox": [ + 511, + 809, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15006", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/df502e8514c154ff40587f2cef1d98da81f090d9336cf65457bedd50a594772d.jpg", + "image_caption": [ + "Figure 2. Impact of coupling type and data reuse on RF's performance. Training RF from scratch with different coupling types, we control the data volume to construct different data reuse scenarios and observe the impact of coupling type and data reuse on model performance under the same training iterations. Each iteration samples 500 data pairs, ensuring that the total data volume seen by the model remains the same across different reuse scenarios under the same training iterations." + ], + "image_footnote": [], + "bbox": [ + 98, + 90, + 475, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of Reflow cannot be solely explained by its ability to mitigate trajectory-crossing issues. This is particularly evident given that trajectory crossings are extremely rare in high-dimensional spaces, as we will elaborate in Section 3.1.", + "bbox": [ + 89, + 460, + 482, + 522 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Through experimental and analytical investigations, the success of Reflow can also be attributed to the following two factors: (1) the images in the training data for Reflow are generated by the pre-trained model, i.e., deterministic couplings. This implies the existence of a deterministic trajectory between a coupling that can be learned and inferred by the model, as shown in the first row of Fig.1. Training/finetuning with Reflow is analogous to learning this trajectory during training and progressively aligning it with a straight-line trajectory. In contrast, randomly matched data pairs (arbitrary couplings) lack explicit trajectory relationships, requiring the model to infer unknown trajectories from noises and constrain them to straight lines, making the training more complex and unstable. As demonstrated in Fig.2, under identical data volume and training iterations, RF trained with deterministic couplings achieves superior image quality (green line vs. orange line). (2) Reflow reuses data by sampling intermediate states from different time steps of the same trajectory for training, as depicted in the first row of Fig.1. This training strategy resembles the multi-time-scale optimization methods employed in the distillation, both of which aim to enhance model performance by leveraging multi-level information. Fig.2 demonstrates that more data reuse iterations lead to better performance under deterministic couplings (orange line vs. blue line).", + "bbox": [ + 91, + 523, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "More details and analyses can be found in Appendix B.", + "bbox": [ + 511, + 90, + 879, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Additionally, the drawbacks of Reflow are evident. As shown in Fig.1, there is a distribution gap between the generated images and the real images. Each iteration of training utilizes images generated by the previous model as training data, leading to the accumulation of errors over time. Consequently, the quality of the images generated by the subsequent model tends to degrade compared to the previous one. To balance high image quality with computational efficiency, it is typically recommended to limit the iterative training to 2 or 3 cycles. Furthermore, while data reuse is effective, its effectiveness is influenced by the volume of data. If the volume of reused data is small, it may not adequately represent the true probabilistic distribution of the data, thereby affecting the model's learning of the PF [14]. On the other hand, excessive reused data will cause significant storage pressure, which increases the consumption of computational resources and reduces training efficiency.", + "bbox": [ + 511, + 107, + 906, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 375, + 604, + 390 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We aim to develop an advanced training framework for RF as a superior alternative to Reflow, further improving the quality of generated images in single-step and few-step generation tasks. To achieve this, we propose a novel method called VRFNO and the overview of VRFNO is shown in Fig.3. Specifically, VRFNO introduces an HVT to effectively improves the model's capability to differentiate inference trajectories, thereby substantially mitigating prediction disturbances arising from intermediate state similarities. Furthermore, to circumvent the use of pre-trained model-generated images as training data during the training process, as employed in Reflow, and to ensure that randomly matched noise-image data pairs $(X_0,X_1)$ exhibit desirable characteristics akin to deterministic coupling. We define a new concept \"optimized coupling\" and propose to achieve optimized coupling by optimizing noise.", + "bbox": [ + 511, + 401, + 906, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 1. $(X_0, X_1)$ is called an optimized coupling if it satisfies the following condition:", + "bbox": [ + 511, + 651, + 906, + 681 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| v _ {\\theta} \\left(t X _ {1} + (1 - t) X _ {0}\\right) - \\left(X _ {1} - X _ {0}\\right) \\right\\| \\leq \\varepsilon ,\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 691, + 854, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\varepsilon > 0$ is a small positive constant.", + "bbox": [ + 511, + 719, + 782, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Viscous Rectified Flow", + "text_level": 1, + "bbox": [ + 511, + 742, + 722, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "RF attributes the curvature of inference trajectories to the frequent crossings of reference trajectories. Specifically, when noise-image pairs are randomly matched, multiple reference trajectories may cross at the same point $X_{t}$ at time step $t$ . During training, this causes different reference trajectories to provide identical inputs to the model at time step $t$ , confusing the model and degrading its prediction accuracy. However, our calculations (Theorem 1) show that the probability of trajectory crossings occurring during actual", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15007", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8407e436529b4479a37ca8bb6f696ed88a276bfb966c2447c5097d11e28299b7.jpg", + "image_caption": [ + "Optimized coupling $(X_0, X_1)$ is obtained by optimizing the noise $\\epsilon$ through the encoder using the reparameterization technique.", + "Figure 3. Overview of VRFNO. The encoder and the neural velocity field form a joint training framework: randomly matched noise $\\epsilon$ and image $X_{1}$ (called arbitrary coupling) are optimized by the encoder to obtain optimized coupling $(X_0,X_1)$ , which are then used to train the neural velocity field. The introduction of the HVT $v_{history}$ enhances the distinction of the trajectories. Both components work together during training to straighten the inference trajectories, thereby improving the model's predictive accuracy." + ], + "image_footnote": [], + "bbox": [ + 98, + 108, + 535, + 253 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3f7f17ed349f21aeac53746a098fb6ab3513f7455c013d3acbf04adf00d87627.jpg", + "image_caption": [ + "The historical velocity term $\\nu_{\\text{history}}$ is introduced into the neural velocity field, and its result is obtained by the prediction method shown on the dotted line." + ], + "image_footnote": [], + "bbox": [ + 535, + 109, + 895, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "training is $P \\sim O(e^{-c(n \\times n)})$ , which is extremely rare in high-dimensional space. More commonly, since the initial states are independently sampled from a standard Gaussian distribution, their statistical properties are very similar. In the early stages of the interpolation trajectory, the intermediate states contain limited image information, causing the differences between the intermediate states to remain small. Since RF learns the mapping relationship between distributions, this high similarity in statistical properties may hinder the model's ability to distinguish between states, leading to blurred predictions. Ideally, RF has a constant velocity field, which means that the intermediate state $X_{t}$ moves in a straight line from noise $X_{0}$ to image $X_{1}$ at uniform velocity $v_{ref}$ , forming a straight inference trajectory.", + "bbox": [ + 88, + 339, + 483, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1. In $(n\\times n)$ -dimensional space, for straight-line interpolation trajectories $X^{(\\cdot)} = \\left\\{X_{t}^{(\\cdot)}:t\\in [0,1]\\right\\}$ , the probability of $X^{(i)}$ and $X^{(j)}$ crossing at the point $X_{t}$ at time step $t$ is $P\\sim O(e^{-c(n\\times n)}),c > 0$ .", + "bbox": [ + 89, + 563, + 483, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Theorem 2, the velocity difference $\\Delta(v_{ref}^{(i)}, v_{ref}^{(j)})$ between each $X_t$ is more significant than their state difference $\\Delta(X_t^{(i)}, X_t^{(j)})$ , and this more significant difference enables the model to better identify the currently inferred trajectory, thus improving prediction accuracy. Therefore, we introduce an HVT $v_{history}$ as an approximation of the constant velocity, which is fed into the model at each time step, resulting in the Viscous Rectified Flow (VRF):", + "bbox": [ + 89, + 645, + 483, + 782 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd X _ {t} = v \\left(X _ {t}, t, v _ {\\text {h i s t o r y}}\\right) d t, \\quad t \\in [ \\Delta t, 1 ], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 795, + 482, + 813 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Delta t$ is the time interval between the current state and the adjacent history state. During the inference process, the HVT input to the model at the first time step is 0. In subsequent sampling iterations, each iteration inputs the velocity predicted by the model at the previous time step into the", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 Training of VRFNO.", + "text_level": 1, + "bbox": [ + 514, + 339, + 741, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input: image $X_{1}$ , noise $\\epsilon$ , learning rate $\\eta$ , time interval $\\Delta t$ , hyperparameter $\\alpha$ , encoder $E_{\\phi}$ , velocity field $v_{\\theta}$ .", + "bbox": [ + 527, + 358, + 905, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "repeat", + "text_level": 1, + "bbox": [ + 531, + 391, + 578, + 402 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mu , \\sigma^ {2} = E _ {\\phi} (X _ {1})\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 404, + 658, + 419 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX _ {0} = \\epsilon \\cdot \\sigma^ {2} + \\mu\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 419, + 658, + 434 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nv _ {r e f} = X _ {1} - X _ {0}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 435, + 661, + 450 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {S a m p l e} t \\sim U (\\Delta t, 1)\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 450, + 691, + 464 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX _ {t} = t X _ {1} + (1 - t) X _ {0}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 465, + 705, + 479 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX _ {t - \\Delta t} = (t - \\Delta t) X _ {1} + (1 - t + \\Delta t) X _ {0}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 479, + 821, + 494 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nv _ {\\text {h i s t o r y}} = \\operatorname {s t o p g r a d} \\left(v _ {\\theta} \\left(X _ {t - \\Delta t}, t - \\Delta t, 0\\right)\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 494, + 839, + 510 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nv _ {t} = v _ {\\theta} \\left(X _ {t}, t, v _ {\\text {h i s t o r y}}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 510, + 705, + 525 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL (\\theta , \\phi) \\leftarrow\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 525, + 620, + 539 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} \\left[ d \\left(v _ {r e f}, v _ {t}\\right) \\right] + \\frac {\\alpha}{2} \\left(\\sigma^ {2} + \\mu^ {2} - 1 - \\log \\left(\\sigma^ {2}\\right)\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 539, + 856, + 556 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta \\leftarrow \\theta - \\eta \\nabla_ {\\theta} L (\\theta , \\phi)\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 556, + 689, + 571 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\phi \\leftarrow \\phi - \\eta \\nabla_ {\\phi} L (\\theta , \\phi)\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 571, + 694, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "until convergence", + "bbox": [ + 531, + 585, + 653, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "model as an HVT, providing the model with auxiliary information on the flow direction and enhancing the model's ability to make correct predictions.", + "bbox": [ + 511, + 630, + 906, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 2. For each intermediate state $X_{t}$ along a PF trajectory, the velocity difference between $X_{t}$ are greater than their state difference:", + "bbox": [ + 511, + 685, + 906, + 731 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta (v _ {r e f} ^ {(i)}, v _ {r e f} ^ {(j)}) \\geq \\Delta (X _ {t} ^ {(i)}, X _ {t} ^ {(j)}),\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 738, + 839, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Delta (\\cdot ,\\cdot) = \\mathbb{E}[\\| \\cdot -\\| _F^2 ]$ and $\\mathbb{E}[\\cdot ]$ is the expectation.", + "bbox": [ + 511, + 768, + 879, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Noise Optimization for Optimized Coupling", + "text_level": 1, + "bbox": [ + 511, + 801, + 885, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose using noise reparameterization technique to optimize the noise representation. Specifically, we design a joint training framework for an encoder and a neural velocity field (as shown in Fig.3), where the encoder optimizes the noises to achieve optimized couplings and uses", + "bbox": [ + 511, + 825, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15008", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "these optimized data pairs to train the neural velocity field. The complete training process is shown in Algorithm 1. We adopt the working mode of the encoder in the VAE architecture [8]: the encoder $E_{\\phi}$ takes the image $X_{1}$ as input and outputs the corresponding mean matrix $\\mu$ and variance matrix $\\sigma^2$ , then generates the optimized noise $X_{0}$ by combining $\\mu$ and $\\sigma^2$ with randomly matched noise $\\epsilon$ using the reparameterization technique:", + "bbox": [ + 89, + 90, + 483, + 212 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nX _ {0} = \\epsilon \\cdot \\sigma^ {2} + \\mu . \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 218, + 482, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the limited number of images in the dataset, to prevent the model from developing a memory effect and ensure diversity in generated images, we introduce random perturbation $\\tau \\sim N(0,\\mathrm{I})$ at the intermediate layers of the encoder, as shown in Fig.3. This ensures that the reparameterized mean and variance matrices still exhibit differences even when multiple Gaussian noises are matched to the same image. Furthermore, after these Gaussian noises matched the same image are reparameterized, they tend to concentrate in a smaller subspace. During training, each image is matched with noise from its corresponding subspace and participates in the training. This approach is similar to the data reuse strategy in Reflow, and it avoids the issue of limiting the model's generalization ability due to the reuse of limited data, thereby helping to improve the performance of the model. After obtaining the optimized noise $X_0$ , we perform straight-line interpolation between it and the corresponding image $X_{1}$ to obtain $X_{t}$ and $X_{t - \\Delta t}$ :", + "bbox": [ + 89, + 242, + 483, + 515 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{c} X _ {t} = t X _ {1} + (1 - t) X _ {0} \\\\ X _ {t - \\Delta t} = (t - \\Delta t) X _ {1} + (1 - t + \\Delta t) X _ {0} \\end{array} , t \\in [ \\Delta t, 1 ]. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 521, + 482, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use the stopgrad operator to calculate the HVT $v_{history} = \\text{stopgrad}(v_{\\theta}(X_{t - \\Delta t}, t - \\Delta t, 0))$ . The neural velocity field $v_{\\theta}$ takes the intermediate state $X_{t}$ , time step $t$ , and HVT as inputs, and learns a constant velocity field by fitting the reference velocity $v_{ref}$ :", + "bbox": [ + 89, + 566, + 483, + 642 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nV C L (\\theta , \\phi) = \\mathbb {E} _ {t \\in p (t)} \\left[ d \\left(v _ {\\text {r e f}}, v _ {\\theta} \\left(X _ {t}, t, v _ {\\text {h i s t o r y}}\\right)\\right) \\right], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 648, + 482, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d(\\cdot, \\cdot)$ indicates distance metric. This process simultaneously optimizes the encoder $E_{\\phi}$ to ensure that the reparameterized noises and the corresponding images satisfy the property of optimized couplings. Meanwhile, to prevent the encoder from overfitting during training, which could cause the reparameterized noises to rely too heavily on image information from datasets and thus constrain the diversity of generated images, we introduce KL divergence regularization to constrain the mean and variance of the encoder's output near the standard Gaussian distribution:", + "bbox": [ + 89, + 672, + 483, + 821 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nK L L (\\phi) = \\frac {1}{2} \\left(\\sigma^ {2} + \\mu^ {2} - 1 - \\log \\left(\\sigma^ {2}\\right)\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 828, + 482, + 858 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, the total loss of our joint training framework is:", + "bbox": [ + 89, + 862, + 478, + 878 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL (\\theta , \\phi) = V C L (\\theta , \\phi) + \\alpha K L L (\\phi), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 883, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 2 Sampling of VRFNO." + ], + "code_body": "Input: image in dataset $X_{1}$ sampling steps $N$ encoder $E_{\\phi}$ velocity field $v_{\\theta}$ \nOutput: generated image $\\tilde{X}_1$ \nSample $\\epsilon \\sim N(0,\\mathrm{I})$ $\\mu ,\\sigma^2 = E_\\phi (X_1)$ $X_0 = \\epsilon \\cdot \\sigma^2 +\\mu$ $\\Delta t = 1 / N$ $v_{history} = 0$ \nfor $i = 0$ to $N - 1$ do \n $v_{pred} = v_{\\theta}(X_{i\\Delta t},i\\Delta t,v_{history})$ $v_{history} = v_{pred}$ $X_{(i + 1)\\Delta t} = X_{i\\Delta t} + \\Delta t\\cdot v_{pred}$ \nend for \n $\\tilde{X}_1 = X_N\\Delta t$", + "bbox": [ + 516, + 109, + 903, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ is a hyperparameter that controls the strength of regularization. By using the optimized couplings optimized through the encoder to train the neural velocity field, it can gradually learn how to infer nearly straight trajectories, thereby enabling few-step and single-step sampling. Furthermore, our encoder-neural velocity field joint generation framework still satisfies the marginal preserving property (Theorem 3), with a detailed definition and proof provided in the Appendix A.", + "bbox": [ + 511, + 352, + 906, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 3. Assume $X$ is rectifiable and $Z$ is its viscous rectified flow. The marginal law of $Z_{t}$ equals that of $X_{t}$ at every time $t$ , i.e., $\\text{Law}(Z_t) = \\text{Law}(X_t), \\forall t \\in [0,1]$ .", + "bbox": [ + 511, + 498, + 906, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Noise optimization. In previous studies, noise optimization typically treats noise as a learnable parameter and performs small-scale iterative updates over multiple iterations using gradient information provided by the model, until it converges to an approximate optimal solution [6, 35]. While it can generate high-quality images, its limitations are also evident: it requires personalized optimization for each noise and relies on pre-trained models. In contrast, our method utilizes the reparameterization technique to optimize the noise through linear transformations directly. This approach allows for significant adjustments to the noise in a single optimization step, enabling rapid and efficient approximation to the optimal solution, significantly reducing the number of iterations and computational overhead, while also eliminating the dependence on pre-trained models.", + "bbox": [ + 511, + 553, + 906, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Two-stage training. In the training process, we adopt a two-stage strategy based on empirical observations. In the first stage, we set $d(\\cdot, \\cdot)$ to the MSE loss, and once convergence is achieved, we proceed to the second stage. In this stage, we incorporate the LPIPS loss [32] for joint training on top of the first-stage setup, continuing until convergence is reached again. Notably, our method does not rely on distillation or adversarial training mechanisms.", + "bbox": [ + 511, + 780, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15009", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8ca191f01a1a07686a891e55b467551102bc71a197e279f871cf4e064a76fe2d.jpg", + "image_caption": [ + "Figure 4. Visualization of inference trajectories on synthetic data. The first column on the left shows the inference trajectories of 1-RF with $N$ steps, and the three columns on the right show the inference trajectories of VRFNO with $N$ steps." + ], + "image_footnote": [], + "bbox": [ + 94, + 90, + 189, + 223 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/53074f9c9151eb4d260a6afa1f58bfdc0f055c7c4a947b38776283878a876a04.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 194, + 99, + 287, + 222 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a43cdd7a4976f71a1c40ee88633529d2a5642fb386730ff332c8d22c3ba00fa9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 101, + 382, + 222 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/82c941c05efd8db3cf68e03b1d98013f6ee4dc5efb85aba578749eb8225a5e05.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 101, + 480, + 222 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sampling. After training the encoder and the neural velocity field, we can generate images using VRF ODE introduced in Eq.(4), the discrete sampling process is shown in Algorithm 2. For single-step generation, the HVT is set to 0 during sampling, while in few-step generation, the HVT is updated at each step. Additionally, we integrate the dataset into the sampler, meaning that the sampled noise is also reparameterized during the generation process. Since we introduced perturbations in the encoder, this will not affect the diversity of the generated images but can further enhance the quality of the generated images.", + "bbox": [ + 89, + 311, + 483, + 478 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 89, + 494, + 215, + 512 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the performance of VRFNO in various scenarios, including synthetic data and real data. In Section 4.1, we compare the performance of RF and VRFNO by studying the inference trajectories between two 2D Gaussian distributions, which clearly demonstrates the effectiveness of our method. In Section 4.2, we extend the experiments to real image datasets, CIFAR-10 [12] and AFHQ [3], at different resolutions to validate the performance of VRFNO in one-step and few-step generation. Additionally, we conducted further analysis on VRFNO, including ablation studies, as well as quantitative evaluations of its trajectory straightness and inference efficiency.", + "bbox": [ + 89, + 521, + 483, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines and evaluation. We evaluated VRFNO and compared it with state-of-the-art diffusion models [10, 18, 24, 27, 28] and rectified flow models (RF [14], Constant Acceleration Flow (CAF) [21], TraFlow [29], Shortcut Model (SM) [7]) in single-step and few-step generation. Among them, RF serves as the baseline for our method, while CAF is the most similar approach to ours. Therefore, these two methods are the primary focus of our comparison. We use Fréchet Inception Distance (FID) and Kernel Inception Distance (KID) to assess the quality of generated images, and use Inception Score (IS) to measure diversity. Lower FID and KID indicate higher generation quality, while a higher IS suggests greater diversity in the generated images.", + "bbox": [ + 89, + 704, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2846db380ce81c7667787afa019a3f124d906ca890b843f607c9aeb55a1f2a81.jpg", + "table_caption": [ + "Table 1. Quantitative results on CIFAR-10." + ], + "table_footnote": [], + "table_body": "
MethodsNFEIS (↑)FID (↓)KID(×10-3)(↓)
Diffusion/Consistency Models
VP ODE (+distill)11.20 (8.73)451 16.23)-
sub-VP ODE (+distill)11.21 (8.80)451 (14.32)-
DDIM distillation18.369.36-
Progressive1-9.12-
CT18.498.71-
EDM5-37.75-
Rectified Flow Models
1-RF11.13379428
2-RF18.1511.978.66
CAF18.324.81-
TraFlow1-4.50-
SM1-4.93-
VRFNO (Ours)19.594.502.73
1-RF57.1234.8132.26
2-RF59.014.362.25
CAF58.674.03-
VRFNO (Ours)59.424.032.13
1-RF108.4412.7011.50
2-RF109.133.831.63
CAF109.123.77-
VRFNO (Ours)109.513.361.31
", + "bbox": [ + 516, + 104, + 936, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Synthetic Experiments", + "text_level": 1, + "bbox": [ + 511, + 530, + 723, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We demonstrate the trajectory straightening effect of VRFNO through 2D synthetic data experiments. For the neural velocity field, we use a simple multi-layer perceptron (MLP) architecture with three hidden layers, each containing 64 units. For the encoder, we adopt the encoder architecture from the VAE framework, consisting of two hidden layers, each with 16 units. Additionally, a fully connected layer is used to produce the mean and variance separately.", + "bbox": [ + 511, + 551, + 906, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The visualization results are shown in Fig.4, where we compared VRFNO with the 1st generation RF (1-RF). Additionally, in Appendix B, we provide the inference trajectories of 1-RF and the 2nd generation RF (2-RF) for different numbers of steps. In the multi-step generation, the inference trajectory of VRFNO is closer to a straight line compared to 1-RF, which provides a guarantee for accurate single-step and few-step generation. The single-step and few-step inference trajectories of VRFNO are visually straight, indicating that the neural velocity field of VRFNO better converges to a constant velocity field, making its predictions of the trajectory direction more stable. In addition, the error between the endpoints of the single-step and the few-step inference trajectories and the endpoints of the multi-step inference trajectories is small, indicating that VRFNO is able", + "bbox": [ + 511, + 674, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15010", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/82c338e1a24126ac5881cc66c4dc65cfcff261d85b7a6d1a6907fe6f2c75fe7e.jpg", + "image_caption": [ + "Figure 5. Qualitative results on CIFAR-10. Visualization of one-step generation of VRFNO" + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 480, + 388 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to accurately map samples from the initial distribution to the target distribution in both single-step and few-step generation. In contrast, when 1-RF is generated in a single step, the endpoint of the trajectory is usually concentrated at the mean point of the target distribution, while the samples near this point are usually meaningless in practical applications, as shown in Fig.II (a) in Appendix B.", + "bbox": [ + 89, + 445, + 483, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Real-data Experiments", + "text_level": 1, + "bbox": [ + 89, + 561, + 305, + 577 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further validate the effectiveness of our method, we evaluate VRFNO on real image datasets, namely CIFAR-10 with a resolution of $32 \\times 32$ and AFHQ with different resolutions (including $64 \\times 64$ , $128 \\times 128$ , $256 \\times 256$ ). For the neural velocity field, we adopt the UNet architecture from DDPM++ [27] to train on CIFAR-10 and the NCSNv2 [26] to train on AFHQ; for the encoder, we refer to the encoder architecture from [8], and its details can be found in Appendix E. Specifically, the encoder requires very few parameters, taking the train of CIFAR-10 as an example, the number of parameters in the encoder is less than 1/20 of those in the neural velocity field.", + "bbox": [ + 89, + 583, + 482, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab.1 presents the quantitative evaluation of VRFNO on CIFAR-10. Compared to other diffusion models, our approach achieves superior performance in single-step generation, with FID = 4.53, KID = 2.73, and IS = 10.59. Compared to previous rectified flow models, VRFNO attains state-of-the-art performance in both single-step and few-step generation (our training does not involve distillation). Fig.5 illustrates the qualitative evaluation of VRFNO in single-step generation on CIFAR-10. Further compara", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1c73188d9b91c1933f51e52ec6c4598cf24418fd24f753ef45955b1aa909bcf3.jpg", + "image_caption": [ + "Figure 6. Qualitative results on AFHQ. Visualization of one-step generation of VRFNO at different resolutions" + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 903, + 388 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/175605a0a1121589df8c9dbaa667de7de4df62ff4e87dc3c089e1a0d02a3051a.jpg", + "table_caption": [ + "Table 2. Quantitative results on AFHQ with different resolutions." + ], + "table_footnote": [], + "table_body": "
MethodsDatasetsNFEResolutionFID(↓)
2-RF (+distill)AFHQ -CAT164 × 64181.93(31.57)
128 × 128172.66(29.81)
256 × 256171.84(29.33)
AFHQ -DOG164 × 64200.77(33.64)
128 × 128192.30(32.70)
256 × 256189.82(30.27)
VRFNO (Ours)AFHQ -CAT164 × 6428.69
128 × 12827.56
256 × 25627.04
AFHQ -DOG164 × 6444.64
128 × 12827.21
256 × 25627.37
", + "bbox": [ + 514, + 446, + 923, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tive qualitative evaluations between VRFNO and RF in N-step generation can be found in Appendix B. Tab.2 provides the quantitative evaluation of VRFNO's generated images on the AFHQ cat and dog datasets, focusing on the image quality of single-step generation at different resolutions. Our method is designed as a better alternative to Reflow, so the comparison primarily highlights the performance difference with 2-RF. From the tab.2, it is evident that VRFNO outperforms 2-RF in image generation quality across different resolutions. The qualitative evaluation of AFHQ at different resolutions can be found in Fig.6.", + "bbox": [ + 509, + 672, + 906, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Further Analysis", + "text_level": 1, + "bbox": [ + 511, + 848, + 681, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Study. We conduct an ablation study to evaluate the effectiveness of each component in our framework", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15011", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/db5f4cecae70497432be3371c2cdd5924934ecb2ad22fc54d9a40b95d05849ca.jpg", + "table_caption": [ + "Table 3. Ablation study of HVT and noise optimization." + ], + "table_footnote": [], + "table_body": "
ConfigNFEHVTNoise optimizationFID(↓)
A1XX379
B1✓X332
C1X✓4.72
D1✓✓4.53
E5XX34.81
F5✓X32.50
G5X✓4.28
H5✓✓4.03
I10XX12.70
J10✓X9.34
K10X✓4.75
L10✓✓3.40
", + "bbox": [ + 119, + 104, + 455, + 320 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/95ac0d56f11b2d87ef419987bc8f2d5ee8a728dd07ba8020da10f05959d8c5e0.jpg", + "table_caption": [ + "Table 4. Comparison of flow straightness." + ], + "table_footnote": [], + "table_body": "
Dataset2-RF3-RFCAFVRFNO(Ours)
2D0.0670.0530.0580.054
CIFAR-100.0580.0560.0350.026
", + "bbox": [ + 99, + 349, + 475, + 401 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/40ba63157b8698125892afc44468c2e2b0714b11492b8c9d9f58a13260232c61.jpg", + "table_caption": [ + "Table 5. Comparison of the inference time (sec) in $N$ -step." + ], + "table_footnote": [], + "table_body": "
NFERFCAFVRFNO(Ours)
10.1720.1810.305
101.4041.4151.646
", + "bbox": [ + 145, + 430, + 429, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "for single-step and few-step generation. Specifically, we investigate the improvements brought to RF by (1) the introduction of the HVT and (2) the noise optimization operation. The configurations and results are shown in Tab.3. The two components have both contributed to the improvement of model performance. Regarding the introduction of the HVT, even with one-step sampling, the model performance improved, indicating that it facilitates trajectory rectification during the training process. As for the noise optimization operation, it results in a significant performance boost, demonstrating the effectiveness of training the neural velocity field with our defined optimized coupling.", + "bbox": [ + 89, + 506, + 482, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Flow Straightness. To evaluate the straightness of the inference trajectories, we introduce the Normalized Flow Straightness Score (NFSS) [14, 15]. A smaller NFSS indicates that the inference trajectory is closer to a straight line. We compare VRFNO with RF and CAF on synthetic datasets and CIFAR-10, with the experimental results shown in Tab.4. Our method generates inference trajectories that are straighter compared to the other methods.", + "bbox": [ + 89, + 688, + 482, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Time Efficiency. We also compare the inference time of our method with RF and CAF. Tab.5 presents the time required to generate 512 images in both one-step and ten-step settings. Our method incurs a slightly higher time cost than RF due to the additional encoder. In contrast, CAF consists of two neural velocity fields, it necessitates an additional", + "bbox": [ + 89, + 810, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "computation (its total NFE equals $N + 1$ ), resulting in significantly higher time consumption.", + "bbox": [ + 511, + 90, + 903, + 122 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Related Work", + "text_level": 1, + "bbox": [ + 511, + 137, + 655, + 152 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The inference process of DMs can be viewed as an iterative solution of ODE [25] or stochastic differential equations [2, 9], and ODE-based methods [5, 31] are more effective in few-step sampling due to their determinism. A widely used first-order ODE solver is the Euler solver, but its large local truncation error necessitates more iteration steps for high-quality image generation. To reduce iteration steps, higher order ODE solvers such as DEIS [31] and DPM solvers [16] exploit the semilinear structure of diffusion ODEs, deriving exact solutions for high-quality image generation with fewer steps. DPM-solver++ [17] enhances DPM-solver by improving stability in higher-order solvers. UniPC [33] builds a unified predictor-corrector framework for DPM solver, enhancing the quality of sampling in a few steps. DC-solver [34] uses dynamic compensation to correct misalignment in UniPC, thus improving image quality.", + "bbox": [ + 511, + 162, + 906, + 405 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Another mainstream method for accelerated sampling is distillation. Progressive distillation [24] iteratively halves the steps of a DDIM until it enables image generation in just 4 steps. Diff-Instruct [20] accelerates sampling by minimizing Integra Kullback-Leibler divergence to merge knowledge from multiple time steps of a pre-trained model. Consistency Model (CM) [28] enables direct mapping from noise to data by minimizing the differences between the final states when adjacent ODE trajectory points are used as inputs. Latent CM [19] treats the guided reverse diffusion process as solving an augmented PF ODE, directly predicting the ODE solution in the latent space, enabling fast and high-fidelity sampling. Consistency Trajectory Model [11] further optimizes CM by allowing the model to learn mappings between arbitrary initial and final times along the ODE trajectory during the diffusion process.", + "bbox": [ + 511, + 406, + 908, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 662, + 633, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper proposes a novel method VRFNO, which is a joint training framework combining an encoder (for noise optimization) and a neural velocity field (for velocity prediction). VRFNO utilizes the encoder to optimize noises to transform randomly matched noises and images into optimized couplings, which are then used to train the neural velocity field for precise prediction of straight inference trajectories. To further straighten the inference trajectories during training, we introduce the HVT into the neural velocity field to enhance its prediction accuracy. Extensive experiments conducted on both synthetic and real world datasets demonstrate the effectiveness and scalability of VRFNO, and the results show that it achieves state-of-the-art performance across multiple datasets and varying resolutions.", + "bbox": [ + 511, + 688, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15012", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Science Fund of China (No.62276135, U24A20330 and 62361166670).", + "bbox": [ + 89, + 113, + 485, + 156 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 167, + 187, + 184 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jonas Adler and Sebastian Lunz. Banach wasserstein gan. Advances in neural information processing systems, 31, 2018. 1", + "[2] Fan Bao, Chongxuan Li, Jun Zhu, and Bo Zhang. Analytic-dpm: an analytic estimate of the optimal reverse variance in diffusion probabilistic models. arXiv preprint arXiv:2201.06503, 2022.8", + "[3] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8188-8197, 2020. 6", + "[4] Tianyi Chu, Wei Xing, Jiafu Chen, Zhizhong Wang, Jiakai Sun, Lei Zhao, Haibo Chen, and Huaizhong Lin. Attack deterministic conditional image generative models for diverse and controllable generation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1362-1370, 2024. 1", + "[5] Tim Dockhorn, Arash Vahdat, and Karsten Kreis. Genie: Higher-order denoising diffusion solvers. Advances in Neural Information Processing Systems, 35:30150-30166, 2022. 8", + "[6] Luca Eyring, Shyamgopal Karthik, Karsten Roth, Alexey Dosovitskiy, and Zeynep Akata. Reno: Enhancing one-step text-to-image models through reward-based noise optimization. Advances in Neural Information Processing Systems, 37:125487-125519, 2025. 5", + "[7] Kevin Frans, Danijar Hafner, Sergey Levine, and Pieter Abbeel. One step diffusion via shortcut models. arXiv preprint arXiv:2410.12557, 2024.6", + "[8] Irina Higgins, Loic Matthew, Arka Pal, Christopher P Burgess, Xavier Glorot, Matthew M Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-vae: Learning basic visual concepts with a constrained variational framework. ICLR (Poster), 3, 2017. 5, 7", + "[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 8", + "[10] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35:26565-26577, 2022. 6", + "[11] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ode trajectory of diffusion. arXiv preprint arXiv:2310.02279, 2023. 8", + "[12] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6", + "[13] Qiang Liu. Rectified flow: A marginal preserving approach to optimal transport. arXiv preprint arXiv:2209.14577, 2022. 2" + ], + "bbox": [ + 93, + 193, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2023. 1, 3, 6, 8", + "[15] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In The Twelfth International Conference on Learning Representations, 2024. 1, 8", + "[16] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 8", + "[17] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 8", + "[18] Eric Luhman and Troy Luhman. Knowledge distillation in iterative generative models for improved sampling speed. arXiv preprint arXiv:2101.02388, 2021. 6", + "[19] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 8", + "[20] Weijian Luo, Tianyang Hu, Shifeng Zhang, Jiacheng Sun, Zhenguo Li, and Zhihua Zhang. Diff-instruct: A universal approach for transferring knowledge from pre-trained diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 8", + "[21] Dogyun Park, Sojin Lee, Sihyeon Kim, Taehoon Lee, Youngjoon Hong, and Hyunwoo J Kim. Constant acceleration flow. arXiv preprint arXiv:2411.00322, 2024. 6", + "[22] Saptarshi Roy, Vansh Bansal, Purnamrita Sarkar, and Alessandro Rinaldo. 2-rectifications are enough for straight flows: A theoretical insight into wasserstein convergence. arXiv e-prints, pages arXiv-2410, 2024. 1", + "[23] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 conference proceedings, pages 1-10, 2022. 1", + "[24] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 6, 8", + "[25] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 8", + "[26] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. Advances in neural information processing systems, 33:12438-12448, 2020. 7", + "[27] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 6, 7", + "[28] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023. 6, 8" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15013", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Zhangkai Wu, Xuhui Fan, Hongyu Wu, and Longbing Cao. Traflow: Trajectory distillation on pre-trained rectified flow, 2025. 6", + "[30] YoungJoon Yoo and Jongwon Choi. Topic-vq-vae: Leveraging latent codebooks for flexible topic-guided document generation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 19422–19430, 2024. 1", + "[31] Qinsheng Zhang and Yongxin Chen. Fast sampling of diffusion models with exponential integrator. arXiv preprint arXiv:2204.13902, 2022. 8", + "[32] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5", + "[33] Wenliang Zhao, Lujia Bai, Yongming Rao, Jie Zhou, and Jiwen Lu. Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 8", + "[34] Wenliang Zhao, Haolin Wang, Jie Zhou, and Jiwen Lu. Dcsolver: Improving predictor-corrector diffusion sampler via dynamic compensation. arXiv preprint arXiv:2409.03755, 2024.8", + "[35]Zikai Zhou,Shitong Shao,Lichen Bai,Zhiqiang Xu,Bo Han,and Zeke Xie. Golden noise for diffusion models: A learning framework.arXiv preprint arXiv:2411.09502, 2024.5" + ], + "bbox": [ + 89, + 92, + 482, + 472 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15014", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Straighten Viscous Rectified Flow via Noise Optimization/7bcafc85-3856-49c2-a845-9dc4fd667b05_model.json b/2025/Straighten Viscous Rectified Flow via Noise Optimization/7bcafc85-3856-49c2-a845-9dc4fd667b05_model.json new file mode 100644 index 0000000000000000000000000000000000000000..68b964eac9ebeb6ec591cb2b2b4dc4b0a476e085 --- /dev/null +++ b/2025/Straighten Viscous Rectified Flow via Noise Optimization/7bcafc85-3856-49c2-a845-9dc4fd667b05_model.json @@ -0,0 +1,2068 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.209, + 0.131, + 0.79, + 0.154 + ], + "angle": 0, + "content": "Straighten Viscous Rectified Flow via Noise Optimization" + }, + { + "type": "text", + "bbox": [ + 0.191, + 0.18, + 0.808, + 0.235 + ], + "angle": 0, + "content": "Jimin Dai1, Jiexi Yan2, Jian Yang1, Lei Luo1 \n1PCA Lab, Nanjing University of Science and Technology, 2Xidian University {jimindai, csjyang}@njust.edu.cn, {jxyan1995, luoleipitt}@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.286 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.302, + 0.485, + 0.65 + ], + "angle": 0, + "content": "The Reflow operation aims to straighten the inference trajectories of the rectified flow during training by constructing deterministic couplings between noises and images, thereby improving the quality of generated images in single-step or few-step generation. However, we identify critical limitations in Reflow, particularly its inability to rapidly generate high-quality images due to a distribution gap between images in its constructed deterministic couplings and real images. To address these shortcomings, we propose a novel alternative called Straighten Viscous Rectified Flow via Noise Optimization (VRFNO), which is a joint training framework integrating an encoder and a neural velocity field. VRFNO introduces two key innovations: (1) a historical velocity term that enhances trajectory distinction, enabling the model to more accurately predict the velocity of the current trajectory, and (2) the noise optimization through reparameterization to form optimized couplings with real images which are then utilized for training, effectively mitigating errors caused by Reflow's limitations. Comprehensive experiments on synthetic data and real datasets with varying resolutions show that VRFNO significantly mitigates the limitations of Reflow, achieving state-of-the-art performance in both one-step and few-step generation tasks." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.679, + 0.222, + 0.695 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Diffusion models (DMs) [9, 23, 25] have attracted significant attention due to their remarkable generative capabilities and stable optimization process. DMs have outperformed traditional models such as generative adversarial networks (GANs) [1, 4] and Variational autoencoders (VAEs) [30] in various generation tasks. DMs work by designing a forward noise-adding scheme, which progressively transforms the target distribution into the initial distribution (usually a Gaussian distribution), thereby constructing a Probability Flow (PF) from complex to simple distributions. DMs learn the inverse process of PF, and then they can sample Gaussian noises and generate samples that approximate the target distribution through multi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.271, + 0.907, + 0.422 + ], + "angle": 0, + "content": "ple inference steps. Currently, various forward noise-adding schemes have been developed for DMs. Among them, the most straightforward approach is the Rectified Flow (RF) [14, 15, 22]. The core idea of RF is to add noise to the data in a linear manner, constructing a deterministic PF from the Gaussian distribution to the target distribution. The model only needs to learn the velocity field of the straight-line interpolation trajectories between the noises and the samples. Once trained, the model can efficiently generate samples that approximate the target distribution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.425, + 0.909, + 0.637 + ], + "angle": 0, + "content": "Generating samples with DMs typically requires multiple inference steps that consume substantial computational resources and time. RF constructs straight-line interpolation trajectories between noises and samples as the reference, and it theoretically learns a straight flow from the Gaussian distribution to the target distribution. The straight flow implies that RF learns a constant velocity field, which enables few-step or even single-step sampling in an ideal scenario. However, during actual training, the RF struggles to learn a constant velocity field, and the resulting PF trajectories remain curved. Therefore, RF still relies on multiple steps to generate the desired images. RF attributes the curvature to the crossing of reference trajectories during training and employs a Reflow operation to mitigate this issue." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.909, + 0.837 + ], + "angle": 0, + "content": "In this paper, we observe that during the actual training process, the probability of crossing between the straight-line interpolation trajectories of randomly matched noises and images at a given time step \\( t \\) is \\( P \\sim O(e^{-c(n \\times n)}) \\), which is very small in high-dimensional data space (Theorem 1). Therefore, we are more inclined to believe that the curvature arises because the neural network model is not an exact solver, intermediate states along different trajectories may become \"approximately crossing\" due to their high similarity in statistical property, thereby interfering with the model predictions. Furthermore, we conducted an in-depth analysis of the factors that contribute to the success of Reflow, as well as its limitations, detailed in Section 2.2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To retain the advantages of Reflow while mitigating its drawbacks, we propose a new method called Straighten Viscous Rectified Flow via Noise Optimization (VRFNO). VRFNO introduces a historical velocity term (HVT) as in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15005" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.424 + ], + "angle": 0, + "content": "put to the neural velocity field, allowing the model to distinguish the direction of the RF in the presence of approximate crossing. By incorporating historical velocity information, the model can more accurately predict the velocity. Moreover, VRFNO directly uses the original dataset to avoid the distribution gap and achieves more appropriate matches between the noises and the images (we refer to it as optimized coupling) by noise optimization. Specifically, we propose a joint training framework for an encoder and a neural velocity field. The encoder first encodes images from randomly matched noise-image pairs and then outputs the corresponding mean and variance matrices. The mean and variance matrices are combined with the noise using the reparameterization technique to generate a new noise, which is then fed into the neural velocity field for training. This joint training approach enables the encoder to adjust the mean and variance of the noise based on the image, thereby optimizing the matches between the noises and the images, which in turn achieves optimized couplings. By using these optimized couplings to train the velocity field, the model can more efficiently straighten the inference trajectory during training, enabling one-step or few-step generation." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.425, + 0.411, + 0.438 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.439, + 0.483, + 0.484 + ], + "angle": 0, + "content": "- We propose a new method VRFNO to straighten the inference trajectories for higher quality few-step and single-step generation." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.484, + 0.483, + 0.53 + ], + "angle": 0, + "content": "- We introduce an HVT as the auxiliary information to enhance the model's accuracy in velocity prediction when the flow approximates crossing." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.53, + 0.484, + 0.59 + ], + "angle": 0, + "content": "- We construct optimized couplings by optimizing the noise to train the neural velocity field. This avoids relying on deterministic couplings used in Reflow, which will constrain the model's generation quality." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.439, + 0.484, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.603, + 0.218, + 0.62 + ], + "angle": 0, + "content": "2. Preliminary" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.628, + 0.238, + 0.643 + ], + "angle": 0, + "content": "2.1. Rectified Flow" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.65, + 0.484, + 0.741 + ], + "angle": 0, + "content": "RF is an ordinary differential equations (ODE)-based DM that aims to learn the mapping between two distributions, \\(\\pi_0\\) and \\(\\pi_{1}\\). In image generation tasks, \\(\\pi_0\\) typically represents a standard Gaussian distribution, while \\(\\pi_{1}\\) corresponds to the target image distribution. For empirical observations \\(X_0\\sim\\) \\(\\pi_0\\) and \\(X_{1}\\sim \\pi_{1}\\) on time \\(t\\in [0,1]\\), RF is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.753, + 0.483, + 0.769 + ], + "angle": 0, + "content": "\\[\nd X _ {t} = v \\left(X _ {t}, t\\right) d t, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.903 + ], + "angle": 0, + "content": "where intermediate state \\( X_{t} = tX_{1} + (1 - t)X_{0} \\) represents a time-differentiable interpolation between \\( X_{0} \\) and \\( X_{1} \\), \\( v: \\mathbb{R}^{n\\times n}\\times [0,1]\\to \\mathbb{R}^{n\\times n} \\) denotes the velocity field defined on the data-time domain. Since RF uses simple straight-line interpolation to connect \\( X_{0} \\) and \\( X_{1} \\), its trajectory has a constant velocity field \\( v_{ref} = X_1 - X_0 \\). Thus, the training process optimizes the model by solving a least-squares regression problem, i.e., fitting \\( v_{ref} \\) to neural ve" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.905, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.299, + 0.908, + 0.411 + ], + "angle": 0, + "content": "Figure 1. Comparison between Reflow and Vanilla training mode. The top illustrates the Reflow training mode, where noises generate images via a pre-trained model, forming deterministic couplings for training. These couplings are reused during training by randomly sampling different intermediate states along the trajectories. The bottom describes the vanilla training mode, where noises and images are randomly sampled to form arbitrary couplings for training, without reusing data during training." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.435, + 0.614, + 0.451 + ], + "angle": 0, + "content": "locity field \\(v_{\\theta}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.559, + 0.46, + 0.905, + 0.487 + ], + "angle": 0, + "content": "\\[\n\\min \\mathbb {E} _ {X _ {0}, X _ {1} \\sim \\gamma , t \\in p (t)} \\left[ \\| v _ {\\text {r e f}} - v _ {\\theta} \\left(X _ {t}, t\\right) \\| ^ {2} \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.496, + 0.907, + 0.557 + ], + "angle": 0, + "content": "where \\(\\gamma\\) represents the coupling of \\((X_0, X_1)\\), \\(p(t)\\) denotes the time distribution defined on \\([0, 1]\\). During the inference process, the ODE usually needs to be discretized and solved via the Euler method, expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.567, + 0.905, + 0.598 + ], + "angle": 0, + "content": "\\[\nX _ {t + \\Delta t} = X _ {t} + \\Delta t \\cdot v _ {\\theta} \\left(X _ {t}, t\\right), t \\in \\{0, \\Delta t, \\dots , (N - 1) \\Delta t \\}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.78 + ], + "angle": 0, + "content": "where \\(\\Delta t = \\frac{1}{N}\\), \\(N\\) represents the total number of steps. Generally, a larger \\(N\\) results in higher-quality generated images, but it requires more computational resources for sampling. The reference trajectory of RF is a straight flow. If RF can converge to the straight flow through training, it can significantly minimize numerical errors in the ODE solver, enabling few-step or single-step generation: as the state moves along the trajectory with uniform linear motion, the velocity at a single time step equals the average velocity over the entire trajectory. As long as the cumulative sum of all time increments throughout the motion equals unity, the final endpoint of the trajectory will remain consistent." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.789, + 0.692, + 0.804 + ], + "angle": 0, + "content": "2.2. Analysis of Reflow" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.902 + ], + "angle": 0, + "content": "The inference trajectory of RF exhibits a curved pattern when trained using the vanilla method. While RF attributes this phenomenon to the crossing of reference trajectories [13] and proposes Reflow as a solution (with implementation details provided in Appendix C), our experiments (Fig.2) and analysis suggest that the effectiveness" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15006" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.092, + 0.476, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.323, + 0.483, + 0.435 + ], + "angle": 0, + "content": "Figure 2. Impact of coupling type and data reuse on RF's performance. Training RF from scratch with different coupling types, we control the data volume to construct different data reuse scenarios and observe the impact of coupling type and data reuse on model performance under the same training iterations. Each iteration samples 500 data pairs, ensuring that the total data volume seen by the model remains the same across different reuse scenarios under the same training iterations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.483, + 0.523 + ], + "angle": 0, + "content": "of Reflow cannot be solely explained by its ability to mitigate trajectory-crossing issues. This is particularly evident given that trajectory crossings are extremely rare in high-dimensional spaces, as we will elaborate in Section 3.1." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.524, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Through experimental and analytical investigations, the success of Reflow can also be attributed to the following two factors: (1) the images in the training data for Reflow are generated by the pre-trained model, i.e., deterministic couplings. This implies the existence of a deterministic trajectory between a coupling that can be learned and inferred by the model, as shown in the first row of Fig.1. Training/finetuning with Reflow is analogous to learning this trajectory during training and progressively aligning it with a straight-line trajectory. In contrast, randomly matched data pairs (arbitrary couplings) lack explicit trajectory relationships, requiring the model to infer unknown trajectories from noises and constrain them to straight lines, making the training more complex and unstable. As demonstrated in Fig.2, under identical data volume and training iterations, RF trained with deterministic couplings achieves superior image quality (green line vs. orange line). (2) Reflow reuses data by sampling intermediate states from different time steps of the same trajectory for training, as depicted in the first row of Fig.1. This training strategy resembles the multi-time-scale optimization methods employed in the distillation, both of which aim to enhance model performance by leveraging multi-level information. Fig.2 demonstrates that more data reuse iterations lead to better performance under deterministic couplings (orange line vs. blue line)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.88, + 0.107 + ], + "angle": 0, + "content": "More details and analyses can be found in Appendix B." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.108, + 0.907, + 0.365 + ], + "angle": 0, + "content": "Additionally, the drawbacks of Reflow are evident. As shown in Fig.1, there is a distribution gap between the generated images and the real images. Each iteration of training utilizes images generated by the previous model as training data, leading to the accumulation of errors over time. Consequently, the quality of the images generated by the subsequent model tends to degrade compared to the previous one. To balance high image quality with computational efficiency, it is typically recommended to limit the iterative training to 2 or 3 cycles. Furthermore, while data reuse is effective, its effectiveness is influenced by the volume of data. If the volume of reused data is small, it may not adequately represent the true probabilistic distribution of the data, thereby affecting the model's learning of the PF [14]. On the other hand, excessive reused data will cause significant storage pressure, which increases the consumption of computational resources and reduces training efficiency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.376, + 0.605, + 0.391 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.907, + 0.645 + ], + "angle": 0, + "content": "We aim to develop an advanced training framework for RF as a superior alternative to Reflow, further improving the quality of generated images in single-step and few-step generation tasks. To achieve this, we propose a novel method called VRFNO and the overview of VRFNO is shown in Fig.3. Specifically, VRFNO introduces an HVT to effectively improves the model's capability to differentiate inference trajectories, thereby substantially mitigating prediction disturbances arising from intermediate state similarities. Furthermore, to circumvent the use of pre-trained model-generated images as training data during the training process, as employed in Reflow, and to ensure that randomly matched noise-image data pairs \\((X_0,X_1)\\) exhibit desirable characteristics akin to deterministic coupling. We define a new concept \"optimized coupling\" and propose to achieve optimized coupling by optimizing noise." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.652, + 0.907, + 0.683 + ], + "angle": 0, + "content": "Definition 1. \\((X_0, X_1)\\) is called an optimized coupling if it satisfies the following condition:" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.693, + 0.855, + 0.71 + ], + "angle": 0, + "content": "\\[\n\\left\\| v _ {\\theta} \\left(t X _ {1} + (1 - t) X _ {0}\\right) - \\left(X _ {1} - X _ {0}\\right) \\right\\| \\leq \\varepsilon ,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.72, + 0.783, + 0.735 + ], + "angle": 0, + "content": "where \\(\\varepsilon > 0\\) is a small positive constant." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.723, + 0.758 + ], + "angle": 0, + "content": "3.1. Viscous Rectified Flow" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "RF attributes the curvature of inference trajectories to the frequent crossings of reference trajectories. Specifically, when noise-image pairs are randomly matched, multiple reference trajectories may cross at the same point \\( X_{t} \\) at time step \\( t \\). During training, this causes different reference trajectories to provide identical inputs to the model at time step \\( t \\), confusing the model and degrading its prediction accuracy. However, our calculations (Theorem 1) show that the probability of trajectory crossings occurring during actual" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15007" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.142, + 0.089, + 0.418, + 0.11 + ], + "angle": 0, + "content": "Optimized coupling \\((X_0, X_1)\\) is obtained by optimizing the noise \\(\\epsilon\\) through the encoder using the reparameterization technique." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.109, + 0.536, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.553, + 0.089, + 0.88, + 0.11 + ], + "angle": 0, + "content": "The historical velocity term \\( \\nu_{\\text{history}} \\) is introduced into the neural velocity field, and its result is obtained by the prediction method shown on the dotted line." + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.111, + 0.896, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.258, + 0.908, + 0.316 + ], + "angle": 0, + "content": "Figure 3. Overview of VRFNO. The encoder and the neural velocity field form a joint training framework: randomly matched noise \\(\\epsilon\\) and image \\(X_{1}\\) (called arbitrary coupling) are optimized by the encoder to obtain optimized coupling \\((X_0,X_1)\\), which are then used to train the neural velocity field. The introduction of the HVT \\(v_{history}\\) enhances the distinction of the trajectories. Both components work together during training to straighten the inference trajectories, thereby improving the model's predictive accuracy." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.34, + 0.485, + 0.552 + ], + "angle": 0, + "content": "training is \\( P \\sim O(e^{-c(n \\times n)}) \\), which is extremely rare in high-dimensional space. More commonly, since the initial states are independently sampled from a standard Gaussian distribution, their statistical properties are very similar. In the early stages of the interpolation trajectory, the intermediate states contain limited image information, causing the differences between the intermediate states to remain small. Since RF learns the mapping relationship between distributions, this high similarity in statistical properties may hinder the model's ability to distinguish between states, leading to blurred predictions. Ideally, RF has a constant velocity field, which means that the intermediate state \\( X_{t} \\) moves in a straight line from noise \\( X_{0} \\) to image \\( X_{1} \\) at uniform velocity \\( v_{ref} \\), forming a straight inference trajectory." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.564, + 0.484, + 0.634 + ], + "angle": 0, + "content": "Theorem 1. In \\((n\\times n)\\)-dimensional space, for straight-line interpolation trajectories \\(X^{(\\cdot)} = \\left\\{X_{t}^{(\\cdot)}:t\\in [0,1]\\right\\}\\), the probability of \\(X^{(i)}\\) and \\(X^{(j)}\\) crossing at the point \\(X_{t}\\) at time step \\(t\\) is \\(P\\sim O(e^{-c(n\\times n)}),c > 0\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.646, + 0.484, + 0.783 + ], + "angle": 0, + "content": "As shown in Theorem 2, the velocity difference \\(\\Delta(v_{ref}^{(i)}, v_{ref}^{(j)})\\) between each \\(X_t\\) is more significant than their state difference \\(\\Delta(X_t^{(i)}, X_t^{(j)})\\), and this more significant difference enables the model to better identify the currently inferred trajectory, thus improving prediction accuracy. Therefore, we introduce an HVT \\(v_{history}\\) as an approximation of the constant velocity, which is fed into the model at each time step, resulting in the Viscous Rectified Flow (VRF):" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.796, + 0.483, + 0.814 + ], + "angle": 0, + "content": "\\[\nd X _ {t} = v \\left(X _ {t}, t, v _ {\\text {h i s t o r y}}\\right) d t, \\quad t \\in [ \\Delta t, 1 ], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\Delta t\\) is the time interval between the current state and the adjacent history state. During the inference process, the HVT input to the model at the first time step is 0. In subsequent sampling iterations, each iteration inputs the velocity predicted by the model at the previous time step into the" + }, + { + "type": "title", + "bbox": [ + 0.516, + 0.34, + 0.743, + 0.355 + ], + "angle": 0, + "content": "Algorithm 1 Training of VRFNO." + }, + { + "type": "text", + "bbox": [ + 0.529, + 0.359, + 0.906, + 0.39 + ], + "angle": 0, + "content": "Input: image \\(X_{1}\\), noise \\(\\epsilon\\), learning rate \\(\\eta\\), time interval \\(\\Delta t\\), hyperparameter \\(\\alpha\\), encoder \\(E_{\\phi}\\), velocity field \\(v_{\\theta}\\)." + }, + { + "type": "title", + "bbox": [ + 0.532, + 0.392, + 0.58, + 0.404 + ], + "angle": 0, + "content": "repeat" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.405, + 0.66, + 0.42 + ], + "angle": 0, + "content": "\\[\n\\mu , \\sigma^ {2} = E _ {\\phi} (X _ {1})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.42, + 0.66, + 0.435 + ], + "angle": 0, + "content": "\\[\nX _ {0} = \\epsilon \\cdot \\sigma^ {2} + \\mu\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.436, + 0.663, + 0.451 + ], + "angle": 0, + "content": "\\[\nv _ {r e f} = X _ {1} - X _ {0}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.451, + 0.692, + 0.465 + ], + "angle": 0, + "content": "\\[\n\\text {S a m p l e} t \\sim U (\\Delta t, 1)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.466, + 0.706, + 0.481 + ], + "angle": 0, + "content": "\\[\nX _ {t} = t X _ {1} + (1 - t) X _ {0}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.481, + 0.822, + 0.495 + ], + "angle": 0, + "content": "\\[\nX _ {t - \\Delta t} = (t - \\Delta t) X _ {1} + (1 - t + \\Delta t) X _ {0}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.496, + 0.841, + 0.511 + ], + "angle": 0, + "content": "\\[\nv _ {\\text {h i s t o r y}} = \\operatorname {s t o p g r a d} \\left(v _ {\\theta} \\left(X _ {t - \\Delta t}, t - \\Delta t, 0\\right)\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.511, + 0.706, + 0.526 + ], + "angle": 0, + "content": "\\[\nv _ {t} = v _ {\\theta} \\left(X _ {t}, t, v _ {\\text {h i s t o r y}}\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.526, + 0.621, + 0.54 + ], + "angle": 0, + "content": "\\[\nL (\\theta , \\phi) \\leftarrow\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.54, + 0.857, + 0.557 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} \\left[ d \\left(v _ {r e f}, v _ {t}\\right) \\right] + \\frac {\\alpha}{2} \\left(\\sigma^ {2} + \\mu^ {2} - 1 - \\log \\left(\\sigma^ {2}\\right)\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.557, + 0.69, + 0.572 + ], + "angle": 0, + "content": "\\[\n\\theta \\leftarrow \\theta - \\eta \\nabla_ {\\theta} L (\\theta , \\phi)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.572, + 0.695, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\phi \\leftarrow \\phi - \\eta \\nabla_ {\\phi} L (\\theta , \\phi)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.587, + 0.654, + 0.601 + ], + "angle": 0, + "content": "until convergence" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.631, + 0.907, + 0.677 + ], + "angle": 0, + "content": "model as an HVT, providing the model with auxiliary information on the flow direction and enhancing the model's ability to make correct predictions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.686, + 0.908, + 0.732 + ], + "angle": 0, + "content": "Theorem 2. For each intermediate state \\( X_{t} \\) along a PF trajectory, the velocity difference between \\( X_{t} \\) are greater than their state difference:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.739, + 0.84, + 0.759 + ], + "angle": 0, + "content": "\\[\n\\Delta (v _ {r e f} ^ {(i)}, v _ {r e f} ^ {(j)}) \\geq \\Delta (X _ {t} ^ {(i)}, X _ {t} ^ {(j)}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.769, + 0.88, + 0.787 + ], + "angle": 0, + "content": "where \\(\\Delta (\\cdot ,\\cdot) = \\mathbb{E}[\\| \\cdot -\\| _F^2 ]\\) and \\(\\mathbb{E}[\\cdot ]\\) is the expectation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.887, + 0.82 + ], + "angle": 0, + "content": "3.2. Noise Optimization for Optimized Coupling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We propose using noise reparameterization technique to optimize the noise representation. Specifically, we design a joint training framework for an encoder and a neural velocity field (as shown in Fig.3), where the encoder optimizes the noises to achieve optimized couplings and uses" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15008" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.213 + ], + "angle": 0, + "content": "these optimized data pairs to train the neural velocity field. The complete training process is shown in Algorithm 1. We adopt the working mode of the encoder in the VAE architecture [8]: the encoder \\( E_{\\phi} \\) takes the image \\( X_{1} \\) as input and outputs the corresponding mean matrix \\( \\mu \\) and variance matrix \\( \\sigma^2 \\), then generates the optimized noise \\( X_{0} \\) by combining \\( \\mu \\) and \\( \\sigma^2 \\) with randomly matched noise \\( \\epsilon \\) using the reparameterization technique:" + }, + { + "type": "equation", + "bbox": [ + 0.229, + 0.219, + 0.483, + 0.236 + ], + "angle": 0, + "content": "\\[\nX _ {0} = \\epsilon \\cdot \\sigma^ {2} + \\mu . \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.485, + 0.516 + ], + "angle": 0, + "content": "Given the limited number of images in the dataset, to prevent the model from developing a memory effect and ensure diversity in generated images, we introduce random perturbation \\(\\tau \\sim N(0,\\mathrm{I})\\) at the intermediate layers of the encoder, as shown in Fig.3. This ensures that the reparameterized mean and variance matrices still exhibit differences even when multiple Gaussian noises are matched to the same image. Furthermore, after these Gaussian noises matched the same image are reparameterized, they tend to concentrate in a smaller subspace. During training, each image is matched with noise from its corresponding subspace and participates in the training. This approach is similar to the data reuse strategy in Reflow, and it avoids the issue of limiting the model's generalization ability due to the reuse of limited data, thereby helping to improve the performance of the model. After obtaining the optimized noise \\(X_0\\), we perform straight-line interpolation between it and the corresponding image \\(X_{1}\\) to obtain \\(X_{t}\\) and \\(X_{t - \\Delta t}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.522, + 0.483, + 0.567 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{c} X _ {t} = t X _ {1} + (1 - t) X _ {0} \\\\ X _ {t - \\Delta t} = (t - \\Delta t) X _ {1} + (1 - t + \\Delta t) X _ {0} \\end{array} , t \\in [ \\Delta t, 1 ]. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.484, + 0.643 + ], + "angle": 0, + "content": "We use the stopgrad operator to calculate the HVT \\( v_{history} = \\text{stopgrad}(v_{\\theta}(X_{t - \\Delta t}, t - \\Delta t, 0)) \\). The neural velocity field \\( v_{\\theta} \\) takes the intermediate state \\( X_{t} \\), time step \\( t \\), and HVT as inputs, and learns a constant velocity field by fitting the reference velocity \\( v_{ref} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.65, + 0.483, + 0.668 + ], + "angle": 0, + "content": "\\[\nV C L (\\theta , \\phi) = \\mathbb {E} _ {t \\in p (t)} \\left[ d \\left(v _ {\\text {r e f}}, v _ {\\theta} \\left(X _ {t}, t, v _ {\\text {h i s t o r y}}\\right)\\right) \\right], \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.484, + 0.822 + ], + "angle": 0, + "content": "where \\(d(\\cdot, \\cdot)\\) indicates distance metric. This process simultaneously optimizes the encoder \\(E_{\\phi}\\) to ensure that the reparameterized noises and the corresponding images satisfy the property of optimized couplings. Meanwhile, to prevent the encoder from overfitting during training, which could cause the reparameterized noises to rely too heavily on image information from datasets and thus constrain the diversity of generated images, we introduce KL divergence regularization to constrain the mean and variance of the encoder's output near the standard Gaussian distribution:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.829, + 0.483, + 0.859 + ], + "angle": 0, + "content": "\\[\nK L L (\\phi) = \\frac {1}{2} \\left(\\sigma^ {2} + \\mu^ {2} - 1 - \\log \\left(\\sigma^ {2}\\right)\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.863, + 0.479, + 0.879 + ], + "angle": 0, + "content": "Therefore, the total loss of our joint training framework is:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.885, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\nL (\\theta , \\phi) = V C L (\\theta , \\phi) + \\alpha K L L (\\phi), \\tag {9}\n\\]" + }, + { + "type": "code_caption", + "bbox": [ + 0.516, + 0.092, + 0.75, + 0.107 + ], + "angle": 0, + "content": "Algorithm 2 Sampling of VRFNO." + }, + { + "type": "algorithm", + "bbox": [ + 0.517, + 0.11, + 0.905, + 0.327 + ], + "angle": 0, + "content": "Input: image in dataset \\(X_{1}\\) sampling steps \\(N\\) encoder \\(E_{\\phi}\\) velocity field \\(v_{\\theta}\\) \nOutput: generated image \\(\\tilde{X}_1\\) \nSample \\(\\epsilon \\sim N(0,\\mathrm{I})\\) \n\\(\\mu ,\\sigma^2 = E_\\phi (X_1)\\) \n\\(X_0 = \\epsilon \\cdot \\sigma^2 +\\mu\\) \n\\(\\Delta t = 1 / N\\) \n\\(v_{history} = 0\\) \nfor \\(i = 0\\) to \\(N - 1\\) do \n\\(v_{pred} = v_{\\theta}(X_{i\\Delta t},i\\Delta t,v_{history})\\) \n\\(v_{history} = v_{pred}\\) \n\\(X_{(i + 1)\\Delta t} = X_{i\\Delta t} + \\Delta t\\cdot v_{pred}\\) \nend for \n\\(\\tilde{X}_1 = X_N\\Delta t\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.907, + 0.49 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a hyperparameter that controls the strength of regularization. By using the optimized couplings optimized through the encoder to train the neural velocity field, it can gradually learn how to infer nearly straight trajectories, thereby enabling few-step and single-step sampling. Furthermore, our encoder-neural velocity field joint generation framework still satisfies the marginal preserving property (Theorem 3), with a detailed definition and proof provided in the Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.499, + 0.907, + 0.545 + ], + "angle": 0, + "content": "Theorem 3. Assume \\( X \\) is rectifiable and \\( Z \\) is its viscous rectified flow. The marginal law of \\( Z_{t} \\) equals that of \\( X_{t} \\) at every time \\( t \\), i.e., \\( \\text{Law}(Z_t) = \\text{Law}(X_t), \\forall t \\in [0,1] \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.907, + 0.78 + ], + "angle": 0, + "content": "Noise optimization. In previous studies, noise optimization typically treats noise as a learnable parameter and performs small-scale iterative updates over multiple iterations using gradient information provided by the model, until it converges to an approximate optimal solution [6, 35]. While it can generate high-quality images, its limitations are also evident: it requires personalized optimization for each noise and relies on pre-trained models. In contrast, our method utilizes the reparameterization technique to optimize the noise through linear transformations directly. This approach allows for significant adjustments to the noise in a single optimization step, enabling rapid and efficient approximation to the optimal solution, significantly reducing the number of iterations and computational overhead, while also eliminating the dependence on pre-trained models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Two-stage training. In the training process, we adopt a two-stage strategy based on empirical observations. In the first stage, we set \\( d(\\cdot, \\cdot) \\) to the MSE loss, and once convergence is achieved, we proceed to the second stage. In this stage, we incorporate the LPIPS loss [32] for joint training on top of the first-stage setup, continuing until convergence is reached again. Notably, our method does not rely on distillation or adversarial training mechanisms." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15009" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.092, + 0.191, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.196, + 0.101, + 0.288, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.102, + 0.383, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.102, + 0.481, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.227, + 0.483, + 0.283 + ], + "angle": 0, + "content": "Figure 4. Visualization of inference trajectories on synthetic data. The first column on the left shows the inference trajectories of 1-RF with \\( N \\) steps, and the three columns on the right show the inference trajectories of VRFNO with \\( N \\) steps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.313, + 0.484, + 0.479 + ], + "angle": 0, + "content": "Sampling. After training the encoder and the neural velocity field, we can generate images using VRF ODE introduced in Eq.(4), the discrete sampling process is shown in Algorithm 2. For single-step generation, the HVT is set to 0 during sampling, while in few-step generation, the HVT is updated at each step. Additionally, we integrate the dataset into the sampler, meaning that the sampled noise is also reparameterized during the generation process. Since we introduced perturbations in the encoder, this will not affect the diversity of the generated images but can further enhance the quality of the generated images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.496, + 0.216, + 0.513 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.484, + 0.703 + ], + "angle": 0, + "content": "We evaluate the performance of VRFNO in various scenarios, including synthetic data and real data. In Section 4.1, we compare the performance of RF and VRFNO by studying the inference trajectories between two 2D Gaussian distributions, which clearly demonstrates the effectiveness of our method. In Section 4.2, we extend the experiments to real image datasets, CIFAR-10 [12] and AFHQ [3], at different resolutions to validate the performance of VRFNO in one-step and few-step generation. Additionally, we conducted further analysis on VRFNO, including ablation studies, as well as quantitative evaluations of its trajectory straightness and inference efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Baselines and evaluation. We evaluated VRFNO and compared it with state-of-the-art diffusion models [10, 18, 24, 27, 28] and rectified flow models (RF [14], Constant Acceleration Flow (CAF) [21], TraFlow [29], Shortcut Model (SM) [7]) in single-step and few-step generation. Among them, RF serves as the baseline for our method, while CAF is the most similar approach to ours. Therefore, these two methods are the primary focus of our comparison. We use Fréchet Inception Distance (FID) and Kernel Inception Distance (KID) to assess the quality of generated images, and use Inception Score (IS) to measure diversity. Lower FID and KID indicate higher generation quality, while a higher IS suggests greater diversity in the generated images." + }, + { + "type": "table_caption", + "bbox": [ + 0.581, + 0.09, + 0.838, + 0.103 + ], + "angle": 0, + "content": "Table 1. Quantitative results on CIFAR-10." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.106, + 0.937, + 0.504 + ], + "angle": 0, + "content": "
MethodsNFEIS (↑)FID (↓)KID(×10-3)(↓)
Diffusion/Consistency Models
VP ODE (+distill)11.20 (8.73)451 16.23)-
sub-VP ODE (+distill)11.21 (8.80)451 (14.32)-
DDIM distillation18.369.36-
Progressive1-9.12-
CT18.498.71-
EDM5-37.75-
Rectified Flow Models
1-RF11.13379428
2-RF18.1511.978.66
CAF18.324.81-
TraFlow1-4.50-
SM1-4.93-
VRFNO (Ours)19.594.502.73
1-RF57.1234.8132.26
2-RF59.014.362.25
CAF58.674.03-
VRFNO (Ours)59.424.032.13
1-RF108.4412.7011.50
2-RF109.133.831.63
CAF109.123.77-
VRFNO (Ours)109.513.361.31
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.531, + 0.725, + 0.547 + ], + "angle": 0, + "content": "4.1. Synthetic Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.674 + ], + "angle": 0, + "content": "We demonstrate the trajectory straightening effect of VRFNO through 2D synthetic data experiments. For the neural velocity field, we use a simple multi-layer perceptron (MLP) architecture with three hidden layers, each containing 64 units. For the encoder, we adopt the encoder architecture from the VAE framework, consisting of two hidden layers, each with 16 units. Additionally, a fully connected layer is used to produce the mean and variance separately." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.901 + ], + "angle": 0, + "content": "The visualization results are shown in Fig.4, where we compared VRFNO with the 1st generation RF (1-RF). Additionally, in Appendix B, we provide the inference trajectories of 1-RF and the 2nd generation RF (2-RF) for different numbers of steps. In the multi-step generation, the inference trajectory of VRFNO is closer to a straight line compared to 1-RF, which provides a guarantee for accurate single-step and few-step generation. The single-step and few-step inference trajectories of VRFNO are visually straight, indicating that the neural velocity field of VRFNO better converges to a constant velocity field, making its predictions of the trajectory direction more stable. In addition, the error between the endpoints of the single-step and the few-step inference trajectories and the endpoints of the multi-step inference trajectories is small, indicating that VRFNO is able" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15010" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.481, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.391, + 0.483, + 0.418 + ], + "angle": 0, + "content": "Figure 5. Qualitative results on CIFAR-10. Visualization of one-step generation of VRFNO" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.446, + 0.484, + 0.552 + ], + "angle": 0, + "content": "to accurately map samples from the initial distribution to the target distribution in both single-step and few-step generation. In contrast, when 1-RF is generated in a single step, the endpoint of the trajectory is usually concentrated at the mean point of the target distribution, while the samples near this point are usually meaningless in practical applications, as shown in Fig.II (a) in Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.562, + 0.307, + 0.578 + ], + "angle": 0, + "content": "4.2. Real-data Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.765 + ], + "angle": 0, + "content": "To further validate the effectiveness of our method, we evaluate VRFNO on real image datasets, namely CIFAR-10 with a resolution of \\(32 \\times 32\\) and AFHQ with different resolutions (including \\(64 \\times 64\\), \\(128 \\times 128\\), \\(256 \\times 256\\)). For the neural velocity field, we adopt the UNet architecture from DDPM++ [27] to train on CIFAR-10 and the NCSNv2 [26] to train on AFHQ; for the encoder, we refer to the encoder architecture from [8], and its details can be found in Appendix E. Specifically, the encoder requires very few parameters, taking the train of CIFAR-10 as an example, the number of parameters in the encoder is less than 1/20 of those in the neural velocity field." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Tab.1 presents the quantitative evaluation of VRFNO on CIFAR-10. Compared to other diffusion models, our approach achieves superior performance in single-step generation, with FID = 4.53, KID = 2.73, and IS = 10.59. Compared to previous rectified flow models, VRFNO attains state-of-the-art performance in both single-step and few-step generation (our training does not involve distillation). Fig.5 illustrates the qualitative evaluation of VRFNO in single-step generation on CIFAR-10. Further compara" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.905, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.391, + 0.907, + 0.418 + ], + "angle": 0, + "content": "Figure 6. Qualitative results on AFHQ. Visualization of one-step generation of VRFNO at different resolutions" + }, + { + "type": "table_caption", + "bbox": [ + 0.516, + 0.433, + 0.902, + 0.447 + ], + "angle": 0, + "content": "Table 2. Quantitative results on AFHQ with different resolutions." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.448, + 0.924, + 0.649 + ], + "angle": 0, + "content": "
MethodsDatasetsNFEResolutionFID(↓)
2-RF (+distill)AFHQ -CAT164 × 64181.93(31.57)
128 × 128172.66(29.81)
256 × 256171.84(29.33)
AFHQ -DOG164 × 64200.77(33.64)
128 × 128192.30(32.70)
256 × 256189.82(30.27)
VRFNO (Ours)AFHQ -CAT164 × 6428.69
128 × 12827.56
256 × 25627.04
AFHQ -DOG164 × 6444.64
128 × 12827.21
256 × 25627.37
" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.673, + 0.907, + 0.84 + ], + "angle": 0, + "content": "tive qualitative evaluations between VRFNO and RF in N-step generation can be found in Appendix B. Tab.2 provides the quantitative evaluation of VRFNO's generated images on the AFHQ cat and dog datasets, focusing on the image quality of single-step generation at different resolutions. Our method is designed as a better alternative to Reflow, so the comparison primarily highlights the performance difference with 2-RF. From the tab.2, it is evident that VRFNO outperforms 2-RF in image generation quality across different resolutions. The qualitative evaluation of AFHQ at different resolutions can be found in Fig.6." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.682, + 0.865 + ], + "angle": 0, + "content": "4.3. Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Ablation Study. We conduct an ablation study to evaluate the effectiveness of each component in our framework" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15011" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.12, + 0.09, + 0.454, + 0.104 + ], + "angle": 0, + "content": "Table 3. Ablation study of HVT and noise optimization." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.106, + 0.457, + 0.321 + ], + "angle": 0, + "content": "
ConfigNFEHVTNoise optimizationFID(↓)
A1XX379
B1✓X332
C1X✓4.72
D1✓✓4.53
E5XX34.81
F5✓X32.50
G5X✓4.28
H5✓✓4.03
I10XX12.70
J10✓X9.34
K10X✓4.75
L10✓✓3.40
" + }, + { + "type": "table_caption", + "bbox": [ + 0.162, + 0.335, + 0.411, + 0.349 + ], + "angle": 0, + "content": "Table 4. Comparison of flow straightness." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.351, + 0.476, + 0.402 + ], + "angle": 0, + "content": "
Dataset2-RF3-RFCAFVRFNO(Ours)
2D0.0670.0530.0580.054
CIFAR-100.0580.0560.0350.026
" + }, + { + "type": "table_caption", + "bbox": [ + 0.112, + 0.415, + 0.462, + 0.43 + ], + "angle": 0, + "content": "Table 5. Comparison of the inference time (sec) in \\(N\\) -step." + }, + { + "type": "table", + "bbox": [ + 0.147, + 0.431, + 0.43, + 0.48 + ], + "angle": 0, + "content": "
NFERFCAFVRFNO(Ours)
10.1720.1810.305
101.4041.4151.646
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.483, + 0.688 + ], + "angle": 0, + "content": "for single-step and few-step generation. Specifically, we investigate the improvements brought to RF by (1) the introduction of the HVT and (2) the noise optimization operation. The configurations and results are shown in Tab.3. The two components have both contributed to the improvement of model performance. Regarding the introduction of the HVT, even with one-step sampling, the model performance improved, indicating that it facilitates trajectory rectification during the training process. As for the noise optimization operation, it results in a significant performance boost, demonstrating the effectiveness of training the neural velocity field with our defined optimized coupling." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Flow Straightness. To evaluate the straightness of the inference trajectories, we introduce the Normalized Flow Straightness Score (NFSS) [14, 15]. A smaller NFSS indicates that the inference trajectory is closer to a straight line. We compare VRFNO with RF and CAF on synthetic datasets and CIFAR-10, with the experimental results shown in Tab.4. Our method generates inference trajectories that are straighter compared to the other methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Time Efficiency. We also compare the inference time of our method with RF and CAF. Tab.5 presents the time required to generate 512 images in both one-step and ten-step settings. Our method incurs a slightly higher time cost than RF due to the additional encoder. In contrast, CAF consists of two neural velocity fields, it necessitates an additional" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.123 + ], + "angle": 0, + "content": "computation (its total NFE equals \\(N + 1\\)), resulting in significantly higher time consumption." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.138, + 0.656, + 0.154 + ], + "angle": 0, + "content": "5. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.164, + 0.907, + 0.406 + ], + "angle": 0, + "content": "The inference process of DMs can be viewed as an iterative solution of ODE [25] or stochastic differential equations [2, 9], and ODE-based methods [5, 31] are more effective in few-step sampling due to their determinism. A widely used first-order ODE solver is the Euler solver, but its large local truncation error necessitates more iteration steps for high-quality image generation. To reduce iteration steps, higher order ODE solvers such as DEIS [31] and DPM solvers [16] exploit the semilinear structure of diffusion ODEs, deriving exact solutions for high-quality image generation with fewer steps. DPM-solver++ [17] enhances DPM-solver by improving stability in higher-order solvers. UniPC [33] builds a unified predictor-corrector framework for DPM solver, enhancing the quality of sampling in a few steps. DC-solver [34] uses dynamic compensation to correct misalignment in UniPC, thus improving image quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.407, + 0.909, + 0.649 + ], + "angle": 0, + "content": "Another mainstream method for accelerated sampling is distillation. Progressive distillation [24] iteratively halves the steps of a DDIM until it enables image generation in just 4 steps. Diff-Instruct [20] accelerates sampling by minimizing Integra Kullback-Leibler divergence to merge knowledge from multiple time steps of a pre-trained model. Consistency Model (CM) [28] enables direct mapping from noise to data by minimizing the differences between the final states when adjacent ODE trajectory points are used as inputs. Latent CM [19] treats the guided reverse diffusion process as solving an augmented PF ODE, directly predicting the ODE solution in the latent space, enabling fast and high-fidelity sampling. Consistency Trajectory Model [11] further optimizes CM by allowing the model to learn mappings between arbitrary initial and final times along the ODE trajectory during the diffusion process." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.664, + 0.634, + 0.68 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.901 + ], + "angle": 0, + "content": "This paper proposes a novel method VRFNO, which is a joint training framework combining an encoder (for noise optimization) and a neural velocity field (for velocity prediction). VRFNO utilizes the encoder to optimize noises to transform randomly matched noises and images into optimized couplings, which are then used to train the neural velocity field for precise prediction of straight inference trajectories. To further straighten the inference trajectories during training, we introduce the HVT into the neural velocity field to enhance its prediction accuracy. Extensive experiments conducted on both synthetic and real world datasets demonstrate the effectiveness and scalability of VRFNO, and the results show that it achieves state-of-the-art performance across multiple datasets and varying resolutions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15012" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.486, + 0.157 + ], + "angle": 0, + "content": "This work was supported in part by the National Science Fund of China (No.62276135, U24A20330 and 62361166670)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.169, + 0.188, + 0.185 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.194, + 0.484, + 0.234 + ], + "angle": 0, + "content": "[1] Jonas Adler and Sebastian Lunz. Banach wasserstein gan. Advances in neural information processing systems, 31, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.236, + 0.484, + 0.29 + ], + "angle": 0, + "content": "[2] Fan Bao, Chongxuan Li, Jun Zhu, and Bo Zhang. Analytic-dpm: an analytic estimate of the optimal reverse variance in diffusion probabilistic models. arXiv preprint arXiv:2201.06503, 2022.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.291, + 0.484, + 0.347 + ], + "angle": 0, + "content": "[3] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8188-8197, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.347, + 0.484, + 0.429 + ], + "angle": 0, + "content": "[4] Tianyi Chu, Wei Xing, Jiafu Chen, Zhizhong Wang, Jiakai Sun, Lei Zhao, Haibo Chen, and Huaizhong Lin. Attack deterministic conditional image generative models for diverse and controllable generation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1362-1370, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.43, + 0.484, + 0.484 + ], + "angle": 0, + "content": "[5] Tim Dockhorn, Arash Vahdat, and Karsten Kreis. Genie: Higher-order denoising diffusion solvers. Advances in Neural Information Processing Systems, 35:30150-30166, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.486, + 0.484, + 0.554 + ], + "angle": 0, + "content": "[6] Luca Eyring, Shyamgopal Karthik, Karsten Roth, Alexey Dosovitskiy, and Zeynep Akata. Reno: Enhancing one-step text-to-image models through reward-based noise optimization. Advances in Neural Information Processing Systems, 37:125487-125519, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.555, + 0.484, + 0.596 + ], + "angle": 0, + "content": "[7] Kevin Frans, Danijar Hafner, Sergey Levine, and Pieter Abbeel. One step diffusion via shortcut models. arXiv preprint arXiv:2410.12557, 2024.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.597, + 0.484, + 0.664 + ], + "angle": 0, + "content": "[8] Irina Higgins, Loic Matthew, Arka Pal, Christopher P Burgess, Xavier Glorot, Matthew M Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-vae: Learning basic visual concepts with a constrained variational framework. ICLR (Poster), 3, 2017. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.665, + 0.484, + 0.707 + ], + "angle": 0, + "content": "[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.708, + 0.484, + 0.762 + ], + "angle": 0, + "content": "[10] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35:26565-26577, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.763, + 0.484, + 0.831 + ], + "angle": 0, + "content": "[11] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ode trajectory of diffusion. arXiv preprint arXiv:2310.02279, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.859 + ], + "angle": 0, + "content": "[12] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.484, + 0.899 + ], + "angle": 0, + "content": "[13] Qiang Liu. Rectified flow: A marginal preserving approach to optimal transport. arXiv preprint arXiv:2209.14577, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.194, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[14] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2023. 1, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.906, + 0.205 + ], + "angle": 0, + "content": "[15] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In The Twelfth International Conference on Learning Representations, 2024. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.906, + 0.274 + ], + "angle": 0, + "content": "[16] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.906, + 0.331 + ], + "angle": 0, + "content": "[17] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.906, + 0.375 + ], + "angle": 0, + "content": "[18] Eric Luhman and Troy Luhman. Knowledge distillation in iterative generative models for improved sampling speed. arXiv preprint arXiv:2101.02388, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.906, + 0.431 + ], + "angle": 0, + "content": "[19] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.906, + 0.502 + ], + "angle": 0, + "content": "[20] Weijian Luo, Tianyang Hu, Shifeng Zhang, Jiacheng Sun, Zhenguo Li, and Zhihua Zhang. Diff-instruct: A universal approach for transferring knowledge from pre-trained diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.906, + 0.546 + ], + "angle": 0, + "content": "[21] Dogyun Park, Sojin Lee, Sihyeon Kim, Taehoon Lee, Youngjoon Hong, and Hyunwoo J Kim. Constant acceleration flow. arXiv preprint arXiv:2411.00322, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.906, + 0.602 + ], + "angle": 0, + "content": "[22] Saptarshi Roy, Vansh Bansal, Purnamrita Sarkar, and Alessandro Rinaldo. 2-rectifications are enough for straight flows: A theoretical insight into wasserstein convergence. arXiv e-prints, pages arXiv-2410, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.906, + 0.671 + ], + "angle": 0, + "content": "[23] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 conference proceedings, pages 1-10, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.674, + 0.906, + 0.715 + ], + "angle": 0, + "content": "[24] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.717, + 0.906, + 0.758 + ], + "angle": 0, + "content": "[25] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.906, + 0.801 + ], + "angle": 0, + "content": "[26] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. Advances in neural information processing systems, 33:12438-12448, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.906, + 0.858 + ], + "angle": 0, + "content": "[27] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[28] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023. 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15013" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[29] Zhangkai Wu, Xuhui Fan, Hongyu Wu, and Longbing Cao. Traflow: Trajectory distillation on pre-trained rectified flow, 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.136, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[30] YoungJoon Yoo and Jongwon Choi. Topic-vq-vae: Leveraging latent codebooks for flexible topic-guided document generation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 19422–19430, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.192, + 0.483, + 0.233 + ], + "angle": 0, + "content": "[31] Qinsheng Zhang and Yongxin Chen. Fast sampling of diffusion models with exponential integrator. arXiv preprint arXiv:2204.13902, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.235, + 0.483, + 0.304 + ], + "angle": 0, + "content": "[32] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.305, + 0.483, + 0.36 + ], + "angle": 0, + "content": "[33] Wenliang Zhao, Lujia Bai, Yongming Rao, Jie Zhou, and Jiwen Lu. Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.362, + 0.482, + 0.415 + ], + "angle": 0, + "content": "[34] Wenliang Zhao, Haolin Wang, Jie Zhou, and Jiwen Lu. Dcsolver: Improving predictor-corrector diffusion sampler via dynamic compensation. arXiv preprint arXiv:2409.03755, 2024.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.418, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[35]Zikai Zhou,Shitong Shao,Lichen Bai,Zhiqiang Xu,Bo Han,and Zeke Xie. Golden noise for diffusion models: A learning framework.arXiv preprint arXiv:2411.09502, 2024.5" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.483, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "15014" + } + ] +] \ No newline at end of file diff --git a/2025/Straighten Viscous Rectified Flow via Noise Optimization/7bcafc85-3856-49c2-a845-9dc4fd667b05_origin.pdf b/2025/Straighten Viscous Rectified Flow via Noise Optimization/7bcafc85-3856-49c2-a845-9dc4fd667b05_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b792893b094f205d56b5451136efea77a412a554 --- /dev/null +++ b/2025/Straighten Viscous Rectified Flow via Noise Optimization/7bcafc85-3856-49c2-a845-9dc4fd667b05_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0aa218eb3d4f22af9f67961ea23f7018f3e45542a2d58321bdd94b870b879aac +size 7057331 diff --git a/2025/Straighten Viscous Rectified Flow via Noise Optimization/full.md b/2025/Straighten Viscous Rectified Flow via Noise Optimization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..019060b5d65e226784b8eb939b1a87666d0631f4 --- /dev/null +++ b/2025/Straighten Viscous Rectified Flow via Noise Optimization/full.md @@ -0,0 +1,349 @@ +# Straighten Viscous Rectified Flow via Noise Optimization + +Jimin Dai1, Jiexi Yan2, Jian Yang1, Lei Luo1 +1PCA Lab, Nanjing University of Science and Technology, 2Xidian University {jimindai, csjyang}@njust.edu.cn, {jxyan1995, luoleipitt}@gmail.com + +# Abstract + +The Reflow operation aims to straighten the inference trajectories of the rectified flow during training by constructing deterministic couplings between noises and images, thereby improving the quality of generated images in single-step or few-step generation. However, we identify critical limitations in Reflow, particularly its inability to rapidly generate high-quality images due to a distribution gap between images in its constructed deterministic couplings and real images. To address these shortcomings, we propose a novel alternative called Straighten Viscous Rectified Flow via Noise Optimization (VRFNO), which is a joint training framework integrating an encoder and a neural velocity field. VRFNO introduces two key innovations: (1) a historical velocity term that enhances trajectory distinction, enabling the model to more accurately predict the velocity of the current trajectory, and (2) the noise optimization through reparameterization to form optimized couplings with real images which are then utilized for training, effectively mitigating errors caused by Reflow's limitations. Comprehensive experiments on synthetic data and real datasets with varying resolutions show that VRFNO significantly mitigates the limitations of Reflow, achieving state-of-the-art performance in both one-step and few-step generation tasks. + +# 1. Introduction + +Diffusion models (DMs) [9, 23, 25] have attracted significant attention due to their remarkable generative capabilities and stable optimization process. DMs have outperformed traditional models such as generative adversarial networks (GANs) [1, 4] and Variational autoencoders (VAEs) [30] in various generation tasks. DMs work by designing a forward noise-adding scheme, which progressively transforms the target distribution into the initial distribution (usually a Gaussian distribution), thereby constructing a Probability Flow (PF) from complex to simple distributions. DMs learn the inverse process of PF, and then they can sample Gaussian noises and generate samples that approximate the target distribution through multi + +ple inference steps. Currently, various forward noise-adding schemes have been developed for DMs. Among them, the most straightforward approach is the Rectified Flow (RF) [14, 15, 22]. The core idea of RF is to add noise to the data in a linear manner, constructing a deterministic PF from the Gaussian distribution to the target distribution. The model only needs to learn the velocity field of the straight-line interpolation trajectories between the noises and the samples. Once trained, the model can efficiently generate samples that approximate the target distribution. + +Generating samples with DMs typically requires multiple inference steps that consume substantial computational resources and time. RF constructs straight-line interpolation trajectories between noises and samples as the reference, and it theoretically learns a straight flow from the Gaussian distribution to the target distribution. The straight flow implies that RF learns a constant velocity field, which enables few-step or even single-step sampling in an ideal scenario. However, during actual training, the RF struggles to learn a constant velocity field, and the resulting PF trajectories remain curved. Therefore, RF still relies on multiple steps to generate the desired images. RF attributes the curvature to the crossing of reference trajectories during training and employs a Reflow operation to mitigate this issue. + +In this paper, we observe that during the actual training process, the probability of crossing between the straight-line interpolation trajectories of randomly matched noises and images at a given time step $t$ is $P \sim O(e^{-c(n \times n)})$ , which is very small in high-dimensional data space (Theorem 1). Therefore, we are more inclined to believe that the curvature arises because the neural network model is not an exact solver, intermediate states along different trajectories may become "approximately crossing" due to their high similarity in statistical property, thereby interfering with the model predictions. Furthermore, we conducted an in-depth analysis of the factors that contribute to the success of Reflow, as well as its limitations, detailed in Section 2.2. + +To retain the advantages of Reflow while mitigating its drawbacks, we propose a new method called Straighten Viscous Rectified Flow via Noise Optimization (VRFNO). VRFNO introduces a historical velocity term (HVT) as in + +put to the neural velocity field, allowing the model to distinguish the direction of the RF in the presence of approximate crossing. By incorporating historical velocity information, the model can more accurately predict the velocity. Moreover, VRFNO directly uses the original dataset to avoid the distribution gap and achieves more appropriate matches between the noises and the images (we refer to it as optimized coupling) by noise optimization. Specifically, we propose a joint training framework for an encoder and a neural velocity field. The encoder first encodes images from randomly matched noise-image pairs and then outputs the corresponding mean and variance matrices. The mean and variance matrices are combined with the noise using the reparameterization technique to generate a new noise, which is then fed into the neural velocity field for training. This joint training approach enables the encoder to adjust the mean and variance of the noise based on the image, thereby optimizing the matches between the noises and the images, which in turn achieves optimized couplings. By using these optimized couplings to train the velocity field, the model can more efficiently straighten the inference trajectory during training, enabling one-step or few-step generation. + +In summary, our contributions are as follows: + +- We propose a new method VRFNO to straighten the inference trajectories for higher quality few-step and single-step generation. +- We introduce an HVT as the auxiliary information to enhance the model's accuracy in velocity prediction when the flow approximates crossing. +- We construct optimized couplings by optimizing the noise to train the neural velocity field. This avoids relying on deterministic couplings used in Reflow, which will constrain the model's generation quality. + +# 2. Preliminary + +# 2.1. Rectified Flow + +RF is an ordinary differential equations (ODE)-based DM that aims to learn the mapping between two distributions, $\pi_0$ and $\pi_{1}$ . In image generation tasks, $\pi_0$ typically represents a standard Gaussian distribution, while $\pi_{1}$ corresponds to the target image distribution. For empirical observations $X_0\sim$ $\pi_0$ and $X_{1}\sim \pi_{1}$ on time $t\in [0,1]$ , RF is defined as: + +$$ +d X _ {t} = v \left(X _ {t}, t\right) d t, \tag {1} +$$ + +where intermediate state $X_{t} = tX_{1} + (1 - t)X_{0}$ represents a time-differentiable interpolation between $X_{0}$ and $X_{1}$ , $v: \mathbb{R}^{n\times n}\times [0,1]\to \mathbb{R}^{n\times n}$ denotes the velocity field defined on the data-time domain. Since RF uses simple straight-line interpolation to connect $X_{0}$ and $X_{1}$ , its trajectory has a constant velocity field $v_{ref} = X_1 - X_0$ . Thus, the training process optimizes the model by solving a least-squares regression problem, i.e., fitting $v_{ref}$ to neural ve + +![](images/d51084fa122bfcd920a78e0cdf4d9fc6dffce3275c9706661f3913e1e0e76166.jpg) +Figure 1. Comparison between Reflow and Vanilla training mode. The top illustrates the Reflow training mode, where noises generate images via a pre-trained model, forming deterministic couplings for training. These couplings are reused during training by randomly sampling different intermediate states along the trajectories. The bottom describes the vanilla training mode, where noises and images are randomly sampled to form arbitrary couplings for training, without reusing data during training. + +locity field $v_{\theta}$ + +$$ +\min \mathbb {E} _ {X _ {0}, X _ {1} \sim \gamma , t \in p (t)} \left[ \| v _ {\text {r e f}} - v _ {\theta} \left(X _ {t}, t\right) \| ^ {2} \right], \tag {2} +$$ + +where $\gamma$ represents the coupling of $(X_0, X_1)$ , $p(t)$ denotes the time distribution defined on $[0, 1]$ . During the inference process, the ODE usually needs to be discretized and solved via the Euler method, expressed as: + +$$ +X _ {t + \Delta t} = X _ {t} + \Delta t \cdot v _ {\theta} \left(X _ {t}, t\right), t \in \{0, \Delta t, \dots , (N - 1) \Delta t \}, \tag {3} +$$ + +where $\Delta t = \frac{1}{N}$ , $N$ represents the total number of steps. Generally, a larger $N$ results in higher-quality generated images, but it requires more computational resources for sampling. The reference trajectory of RF is a straight flow. If RF can converge to the straight flow through training, it can significantly minimize numerical errors in the ODE solver, enabling few-step or single-step generation: as the state moves along the trajectory with uniform linear motion, the velocity at a single time step equals the average velocity over the entire trajectory. As long as the cumulative sum of all time increments throughout the motion equals unity, the final endpoint of the trajectory will remain consistent. + +# 2.2. Analysis of Reflow + +The inference trajectory of RF exhibits a curved pattern when trained using the vanilla method. While RF attributes this phenomenon to the crossing of reference trajectories [13] and proposes Reflow as a solution (with implementation details provided in Appendix C), our experiments (Fig.2) and analysis suggest that the effectiveness + +![](images/df502e8514c154ff40587f2cef1d98da81f090d9336cf65457bedd50a594772d.jpg) +Figure 2. Impact of coupling type and data reuse on RF's performance. Training RF from scratch with different coupling types, we control the data volume to construct different data reuse scenarios and observe the impact of coupling type and data reuse on model performance under the same training iterations. Each iteration samples 500 data pairs, ensuring that the total data volume seen by the model remains the same across different reuse scenarios under the same training iterations. + +of Reflow cannot be solely explained by its ability to mitigate trajectory-crossing issues. This is particularly evident given that trajectory crossings are extremely rare in high-dimensional spaces, as we will elaborate in Section 3.1. + +Through experimental and analytical investigations, the success of Reflow can also be attributed to the following two factors: (1) the images in the training data for Reflow are generated by the pre-trained model, i.e., deterministic couplings. This implies the existence of a deterministic trajectory between a coupling that can be learned and inferred by the model, as shown in the first row of Fig.1. Training/finetuning with Reflow is analogous to learning this trajectory during training and progressively aligning it with a straight-line trajectory. In contrast, randomly matched data pairs (arbitrary couplings) lack explicit trajectory relationships, requiring the model to infer unknown trajectories from noises and constrain them to straight lines, making the training more complex and unstable. As demonstrated in Fig.2, under identical data volume and training iterations, RF trained with deterministic couplings achieves superior image quality (green line vs. orange line). (2) Reflow reuses data by sampling intermediate states from different time steps of the same trajectory for training, as depicted in the first row of Fig.1. This training strategy resembles the multi-time-scale optimization methods employed in the distillation, both of which aim to enhance model performance by leveraging multi-level information. Fig.2 demonstrates that more data reuse iterations lead to better performance under deterministic couplings (orange line vs. blue line). + +More details and analyses can be found in Appendix B. + +Additionally, the drawbacks of Reflow are evident. As shown in Fig.1, there is a distribution gap between the generated images and the real images. Each iteration of training utilizes images generated by the previous model as training data, leading to the accumulation of errors over time. Consequently, the quality of the images generated by the subsequent model tends to degrade compared to the previous one. To balance high image quality with computational efficiency, it is typically recommended to limit the iterative training to 2 or 3 cycles. Furthermore, while data reuse is effective, its effectiveness is influenced by the volume of data. If the volume of reused data is small, it may not adequately represent the true probabilistic distribution of the data, thereby affecting the model's learning of the PF [14]. On the other hand, excessive reused data will cause significant storage pressure, which increases the consumption of computational resources and reduces training efficiency. + +# 3. Method + +We aim to develop an advanced training framework for RF as a superior alternative to Reflow, further improving the quality of generated images in single-step and few-step generation tasks. To achieve this, we propose a novel method called VRFNO and the overview of VRFNO is shown in Fig.3. Specifically, VRFNO introduces an HVT to effectively improves the model's capability to differentiate inference trajectories, thereby substantially mitigating prediction disturbances arising from intermediate state similarities. Furthermore, to circumvent the use of pre-trained model-generated images as training data during the training process, as employed in Reflow, and to ensure that randomly matched noise-image data pairs $(X_0,X_1)$ exhibit desirable characteristics akin to deterministic coupling. We define a new concept "optimized coupling" and propose to achieve optimized coupling by optimizing noise. + +Definition 1. $(X_0, X_1)$ is called an optimized coupling if it satisfies the following condition: + +$$ +\left\| v _ {\theta} \left(t X _ {1} + (1 - t) X _ {0}\right) - \left(X _ {1} - X _ {0}\right) \right\| \leq \varepsilon , +$$ + +where $\varepsilon > 0$ is a small positive constant. + +# 3.1. Viscous Rectified Flow + +RF attributes the curvature of inference trajectories to the frequent crossings of reference trajectories. Specifically, when noise-image pairs are randomly matched, multiple reference trajectories may cross at the same point $X_{t}$ at time step $t$ . During training, this causes different reference trajectories to provide identical inputs to the model at time step $t$ , confusing the model and degrading its prediction accuracy. However, our calculations (Theorem 1) show that the probability of trajectory crossings occurring during actual + +![](images/8407e436529b4479a37ca8bb6f696ed88a276bfb966c2447c5097d11e28299b7.jpg) +Optimized coupling $(X_0, X_1)$ is obtained by optimizing the noise $\epsilon$ through the encoder using the reparameterization technique. +Figure 3. Overview of VRFNO. The encoder and the neural velocity field form a joint training framework: randomly matched noise $\epsilon$ and image $X_{1}$ (called arbitrary coupling) are optimized by the encoder to obtain optimized coupling $(X_0,X_1)$ , which are then used to train the neural velocity field. The introduction of the HVT $v_{history}$ enhances the distinction of the trajectories. Both components work together during training to straighten the inference trajectories, thereby improving the model's predictive accuracy. + +![](images/3f7f17ed349f21aeac53746a098fb6ab3513f7455c013d3acbf04adf00d87627.jpg) +The historical velocity term $\nu_{\text{history}}$ is introduced into the neural velocity field, and its result is obtained by the prediction method shown on the dotted line. + +training is $P \sim O(e^{-c(n \times n)})$ , which is extremely rare in high-dimensional space. More commonly, since the initial states are independently sampled from a standard Gaussian distribution, their statistical properties are very similar. In the early stages of the interpolation trajectory, the intermediate states contain limited image information, causing the differences between the intermediate states to remain small. Since RF learns the mapping relationship between distributions, this high similarity in statistical properties may hinder the model's ability to distinguish between states, leading to blurred predictions. Ideally, RF has a constant velocity field, which means that the intermediate state $X_{t}$ moves in a straight line from noise $X_{0}$ to image $X_{1}$ at uniform velocity $v_{ref}$ , forming a straight inference trajectory. + +Theorem 1. In $(n\times n)$ -dimensional space, for straight-line interpolation trajectories $X^{(\cdot)} = \left\{X_{t}^{(\cdot)}:t\in [0,1]\right\}$ , the probability of $X^{(i)}$ and $X^{(j)}$ crossing at the point $X_{t}$ at time step $t$ is $P\sim O(e^{-c(n\times n)}),c > 0$ . + +As shown in Theorem 2, the velocity difference $\Delta(v_{ref}^{(i)}, v_{ref}^{(j)})$ between each $X_t$ is more significant than their state difference $\Delta(X_t^{(i)}, X_t^{(j)})$ , and this more significant difference enables the model to better identify the currently inferred trajectory, thus improving prediction accuracy. Therefore, we introduce an HVT $v_{history}$ as an approximation of the constant velocity, which is fed into the model at each time step, resulting in the Viscous Rectified Flow (VRF): + +$$ +d X _ {t} = v \left(X _ {t}, t, v _ {\text {h i s t o r y}}\right) d t, \quad t \in [ \Delta t, 1 ], \tag {4} +$$ + +where $\Delta t$ is the time interval between the current state and the adjacent history state. During the inference process, the HVT input to the model at the first time step is 0. In subsequent sampling iterations, each iteration inputs the velocity predicted by the model at the previous time step into the + +# Algorithm 1 Training of VRFNO. + +Input: image $X_{1}$ , noise $\epsilon$ , learning rate $\eta$ , time interval $\Delta t$ , hyperparameter $\alpha$ , encoder $E_{\phi}$ , velocity field $v_{\theta}$ . + +# repeat + +$$ +\mu , \sigma^ {2} = E _ {\phi} (X _ {1}) +$$ + +$$ +X _ {0} = \epsilon \cdot \sigma^ {2} + \mu +$$ + +$$ +v _ {r e f} = X _ {1} - X _ {0} +$$ + +$$ +\text {S a m p l e} t \sim U (\Delta t, 1) +$$ + +$$ +X _ {t} = t X _ {1} + (1 - t) X _ {0} +$$ + +$$ +X _ {t - \Delta t} = (t - \Delta t) X _ {1} + (1 - t + \Delta t) X _ {0} +$$ + +$$ +v _ {\text {h i s t o r y}} = \operatorname {s t o p g r a d} \left(v _ {\theta} \left(X _ {t - \Delta t}, t - \Delta t, 0\right)\right) +$$ + +$$ +v _ {t} = v _ {\theta} \left(X _ {t}, t, v _ {\text {h i s t o r y}}\right) +$$ + +$$ +L (\theta , \phi) \leftarrow +$$ + +$$ +\mathbb {E} \left[ d \left(v _ {r e f}, v _ {t}\right) \right] + \frac {\alpha}{2} \left(\sigma^ {2} + \mu^ {2} - 1 - \log \left(\sigma^ {2}\right)\right) +$$ + +$$ +\theta \leftarrow \theta - \eta \nabla_ {\theta} L (\theta , \phi) +$$ + +$$ +\phi \leftarrow \phi - \eta \nabla_ {\phi} L (\theta , \phi) +$$ + +until convergence + +model as an HVT, providing the model with auxiliary information on the flow direction and enhancing the model's ability to make correct predictions. + +Theorem 2. For each intermediate state $X_{t}$ along a PF trajectory, the velocity difference between $X_{t}$ are greater than their state difference: + +$$ +\Delta (v _ {r e f} ^ {(i)}, v _ {r e f} ^ {(j)}) \geq \Delta (X _ {t} ^ {(i)}, X _ {t} ^ {(j)}), +$$ + +where $\Delta (\cdot ,\cdot) = \mathbb{E}[\| \cdot -\| _F^2 ]$ and $\mathbb{E}[\cdot ]$ is the expectation. + +# 3.2. Noise Optimization for Optimized Coupling + +We propose using noise reparameterization technique to optimize the noise representation. Specifically, we design a joint training framework for an encoder and a neural velocity field (as shown in Fig.3), where the encoder optimizes the noises to achieve optimized couplings and uses + +these optimized data pairs to train the neural velocity field. The complete training process is shown in Algorithm 1. We adopt the working mode of the encoder in the VAE architecture [8]: the encoder $E_{\phi}$ takes the image $X_{1}$ as input and outputs the corresponding mean matrix $\mu$ and variance matrix $\sigma^2$ , then generates the optimized noise $X_{0}$ by combining $\mu$ and $\sigma^2$ with randomly matched noise $\epsilon$ using the reparameterization technique: + +$$ +X _ {0} = \epsilon \cdot \sigma^ {2} + \mu . \tag {5} +$$ + +Given the limited number of images in the dataset, to prevent the model from developing a memory effect and ensure diversity in generated images, we introduce random perturbation $\tau \sim N(0,\mathrm{I})$ at the intermediate layers of the encoder, as shown in Fig.3. This ensures that the reparameterized mean and variance matrices still exhibit differences even when multiple Gaussian noises are matched to the same image. Furthermore, after these Gaussian noises matched the same image are reparameterized, they tend to concentrate in a smaller subspace. During training, each image is matched with noise from its corresponding subspace and participates in the training. This approach is similar to the data reuse strategy in Reflow, and it avoids the issue of limiting the model's generalization ability due to the reuse of limited data, thereby helping to improve the performance of the model. After obtaining the optimized noise $X_0$ , we perform straight-line interpolation between it and the corresponding image $X_{1}$ to obtain $X_{t}$ and $X_{t - \Delta t}$ : + +$$ +\begin{array}{c} X _ {t} = t X _ {1} + (1 - t) X _ {0} \\ X _ {t - \Delta t} = (t - \Delta t) X _ {1} + (1 - t + \Delta t) X _ {0} \end{array} , t \in [ \Delta t, 1 ]. \tag {6} +$$ + +We use the stopgrad operator to calculate the HVT $v_{history} = \text{stopgrad}(v_{\theta}(X_{t - \Delta t}, t - \Delta t, 0))$ . The neural velocity field $v_{\theta}$ takes the intermediate state $X_{t}$ , time step $t$ , and HVT as inputs, and learns a constant velocity field by fitting the reference velocity $v_{ref}$ : + +$$ +V C L (\theta , \phi) = \mathbb {E} _ {t \in p (t)} \left[ d \left(v _ {\text {r e f}}, v _ {\theta} \left(X _ {t}, t, v _ {\text {h i s t o r y}}\right)\right) \right], \tag {7} +$$ + +where $d(\cdot, \cdot)$ indicates distance metric. This process simultaneously optimizes the encoder $E_{\phi}$ to ensure that the reparameterized noises and the corresponding images satisfy the property of optimized couplings. Meanwhile, to prevent the encoder from overfitting during training, which could cause the reparameterized noises to rely too heavily on image information from datasets and thus constrain the diversity of generated images, we introduce KL divergence regularization to constrain the mean and variance of the encoder's output near the standard Gaussian distribution: + +$$ +K L L (\phi) = \frac {1}{2} \left(\sigma^ {2} + \mu^ {2} - 1 - \log \left(\sigma^ {2}\right)\right). \tag {8} +$$ + +Therefore, the total loss of our joint training framework is: + +$$ +L (\theta , \phi) = V C L (\theta , \phi) + \alpha K L L (\phi), \tag {9} +$$ + +Algorithm 2 Sampling of VRFNO. +Input: image in dataset $X_{1}$ sampling steps $N$ encoder $E_{\phi}$ velocity field $v_{\theta}$ +Output: generated image $\tilde{X}_1$ +Sample $\epsilon \sim N(0,\mathrm{I})$ $\mu ,\sigma^2 = E_\phi (X_1)$ $X_0 = \epsilon \cdot \sigma^2 +\mu$ $\Delta t = 1 / N$ $v_{history} = 0$ +for $i = 0$ to $N - 1$ do + $v_{pred} = v_{\theta}(X_{i\Delta t},i\Delta t,v_{history})$ $v_{history} = v_{pred}$ $X_{(i + 1)\Delta t} = X_{i\Delta t} + \Delta t\cdot v_{pred}$ +end for + $\tilde{X}_1 = X_N\Delta t$ + +where $\alpha$ is a hyperparameter that controls the strength of regularization. By using the optimized couplings optimized through the encoder to train the neural velocity field, it can gradually learn how to infer nearly straight trajectories, thereby enabling few-step and single-step sampling. Furthermore, our encoder-neural velocity field joint generation framework still satisfies the marginal preserving property (Theorem 3), with a detailed definition and proof provided in the Appendix A. + +Theorem 3. Assume $X$ is rectifiable and $Z$ is its viscous rectified flow. The marginal law of $Z_{t}$ equals that of $X_{t}$ at every time $t$ , i.e., $\text{Law}(Z_t) = \text{Law}(X_t), \forall t \in [0,1]$ . + +Noise optimization. In previous studies, noise optimization typically treats noise as a learnable parameter and performs small-scale iterative updates over multiple iterations using gradient information provided by the model, until it converges to an approximate optimal solution [6, 35]. While it can generate high-quality images, its limitations are also evident: it requires personalized optimization for each noise and relies on pre-trained models. In contrast, our method utilizes the reparameterization technique to optimize the noise through linear transformations directly. This approach allows for significant adjustments to the noise in a single optimization step, enabling rapid and efficient approximation to the optimal solution, significantly reducing the number of iterations and computational overhead, while also eliminating the dependence on pre-trained models. + +Two-stage training. In the training process, we adopt a two-stage strategy based on empirical observations. In the first stage, we set $d(\cdot, \cdot)$ to the MSE loss, and once convergence is achieved, we proceed to the second stage. In this stage, we incorporate the LPIPS loss [32] for joint training on top of the first-stage setup, continuing until convergence is reached again. Notably, our method does not rely on distillation or adversarial training mechanisms. + +![](images/8ca191f01a1a07686a891e55b467551102bc71a197e279f871cf4e064a76fe2d.jpg) +Figure 4. Visualization of inference trajectories on synthetic data. The first column on the left shows the inference trajectories of 1-RF with $N$ steps, and the three columns on the right show the inference trajectories of VRFNO with $N$ steps. + +![](images/53074f9c9151eb4d260a6afa1f58bfdc0f055c7c4a947b38776283878a876a04.jpg) + +![](images/a43cdd7a4976f71a1c40ee88633529d2a5642fb386730ff332c8d22c3ba00fa9.jpg) + +![](images/82c941c05efd8db3cf68e03b1d98013f6ee4dc5efb85aba578749eb8225a5e05.jpg) + +Sampling. After training the encoder and the neural velocity field, we can generate images using VRF ODE introduced in Eq.(4), the discrete sampling process is shown in Algorithm 2. For single-step generation, the HVT is set to 0 during sampling, while in few-step generation, the HVT is updated at each step. Additionally, we integrate the dataset into the sampler, meaning that the sampled noise is also reparameterized during the generation process. Since we introduced perturbations in the encoder, this will not affect the diversity of the generated images but can further enhance the quality of the generated images. + +# 4. Experiment + +We evaluate the performance of VRFNO in various scenarios, including synthetic data and real data. In Section 4.1, we compare the performance of RF and VRFNO by studying the inference trajectories between two 2D Gaussian distributions, which clearly demonstrates the effectiveness of our method. In Section 4.2, we extend the experiments to real image datasets, CIFAR-10 [12] and AFHQ [3], at different resolutions to validate the performance of VRFNO in one-step and few-step generation. Additionally, we conducted further analysis on VRFNO, including ablation studies, as well as quantitative evaluations of its trajectory straightness and inference efficiency. + +Baselines and evaluation. We evaluated VRFNO and compared it with state-of-the-art diffusion models [10, 18, 24, 27, 28] and rectified flow models (RF [14], Constant Acceleration Flow (CAF) [21], TraFlow [29], Shortcut Model (SM) [7]) in single-step and few-step generation. Among them, RF serves as the baseline for our method, while CAF is the most similar approach to ours. Therefore, these two methods are the primary focus of our comparison. We use Fréchet Inception Distance (FID) and Kernel Inception Distance (KID) to assess the quality of generated images, and use Inception Score (IS) to measure diversity. Lower FID and KID indicate higher generation quality, while a higher IS suggests greater diversity in the generated images. + +Table 1. Quantitative results on CIFAR-10. + +
MethodsNFEIS (↑)FID (↓)KID(×10-3)(↓)
Diffusion/Consistency Models
VP ODE (+distill)11.20 (8.73)451 16.23)-
sub-VP ODE (+distill)11.21 (8.80)451 (14.32)-
DDIM distillation18.369.36-
Progressive1-9.12-
CT18.498.71-
EDM5-37.75-
Rectified Flow Models
1-RF11.13379428
2-RF18.1511.978.66
CAF18.324.81-
TraFlow1-4.50-
SM1-4.93-
VRFNO (Ours)19.594.502.73
1-RF57.1234.8132.26
2-RF59.014.362.25
CAF58.674.03-
VRFNO (Ours)59.424.032.13
1-RF108.4412.7011.50
2-RF109.133.831.63
CAF109.123.77-
VRFNO (Ours)109.513.361.31
+ +# 4.1. Synthetic Experiments + +We demonstrate the trajectory straightening effect of VRFNO through 2D synthetic data experiments. For the neural velocity field, we use a simple multi-layer perceptron (MLP) architecture with three hidden layers, each containing 64 units. For the encoder, we adopt the encoder architecture from the VAE framework, consisting of two hidden layers, each with 16 units. Additionally, a fully connected layer is used to produce the mean and variance separately. + +The visualization results are shown in Fig.4, where we compared VRFNO with the 1st generation RF (1-RF). Additionally, in Appendix B, we provide the inference trajectories of 1-RF and the 2nd generation RF (2-RF) for different numbers of steps. In the multi-step generation, the inference trajectory of VRFNO is closer to a straight line compared to 1-RF, which provides a guarantee for accurate single-step and few-step generation. The single-step and few-step inference trajectories of VRFNO are visually straight, indicating that the neural velocity field of VRFNO better converges to a constant velocity field, making its predictions of the trajectory direction more stable. In addition, the error between the endpoints of the single-step and the few-step inference trajectories and the endpoints of the multi-step inference trajectories is small, indicating that VRFNO is able + +![](images/82c338e1a24126ac5881cc66c4dc65cfcff261d85b7a6d1a6907fe6f2c75fe7e.jpg) +Figure 5. Qualitative results on CIFAR-10. Visualization of one-step generation of VRFNO + +to accurately map samples from the initial distribution to the target distribution in both single-step and few-step generation. In contrast, when 1-RF is generated in a single step, the endpoint of the trajectory is usually concentrated at the mean point of the target distribution, while the samples near this point are usually meaningless in practical applications, as shown in Fig.II (a) in Appendix B. + +# 4.2. Real-data Experiments + +To further validate the effectiveness of our method, we evaluate VRFNO on real image datasets, namely CIFAR-10 with a resolution of $32 \times 32$ and AFHQ with different resolutions (including $64 \times 64$ , $128 \times 128$ , $256 \times 256$ ). For the neural velocity field, we adopt the UNet architecture from DDPM++ [27] to train on CIFAR-10 and the NCSNv2 [26] to train on AFHQ; for the encoder, we refer to the encoder architecture from [8], and its details can be found in Appendix E. Specifically, the encoder requires very few parameters, taking the train of CIFAR-10 as an example, the number of parameters in the encoder is less than 1/20 of those in the neural velocity field. + +Tab.1 presents the quantitative evaluation of VRFNO on CIFAR-10. Compared to other diffusion models, our approach achieves superior performance in single-step generation, with FID = 4.53, KID = 2.73, and IS = 10.59. Compared to previous rectified flow models, VRFNO attains state-of-the-art performance in both single-step and few-step generation (our training does not involve distillation). Fig.5 illustrates the qualitative evaluation of VRFNO in single-step generation on CIFAR-10. Further compara + +![](images/1c73188d9b91c1933f51e52ec6c4598cf24418fd24f753ef45955b1aa909bcf3.jpg) +Figure 6. Qualitative results on AFHQ. Visualization of one-step generation of VRFNO at different resolutions + +Table 2. Quantitative results on AFHQ with different resolutions. + +
MethodsDatasetsNFEResolutionFID(↓)
2-RF (+distill)AFHQ -CAT164 × 64181.93(31.57)
128 × 128172.66(29.81)
256 × 256171.84(29.33)
AFHQ -DOG164 × 64200.77(33.64)
128 × 128192.30(32.70)
256 × 256189.82(30.27)
VRFNO (Ours)AFHQ -CAT164 × 6428.69
128 × 12827.56
256 × 25627.04
AFHQ -DOG164 × 6444.64
128 × 12827.21
256 × 25627.37
+ +tive qualitative evaluations between VRFNO and RF in N-step generation can be found in Appendix B. Tab.2 provides the quantitative evaluation of VRFNO's generated images on the AFHQ cat and dog datasets, focusing on the image quality of single-step generation at different resolutions. Our method is designed as a better alternative to Reflow, so the comparison primarily highlights the performance difference with 2-RF. From the tab.2, it is evident that VRFNO outperforms 2-RF in image generation quality across different resolutions. The qualitative evaluation of AFHQ at different resolutions can be found in Fig.6. + +# 4.3. Further Analysis + +Ablation Study. We conduct an ablation study to evaluate the effectiveness of each component in our framework + +Table 3. Ablation study of HVT and noise optimization. + +
ConfigNFEHVTNoise optimizationFID(↓)
A1XX379
B1✓X332
C1X✓4.72
D1✓✓4.53
E5XX34.81
F5✓X32.50
G5X✓4.28
H5✓✓4.03
I10XX12.70
J10✓X9.34
K10X✓4.75
L10✓✓3.40
+ +Table 4. Comparison of flow straightness. + +
Dataset2-RF3-RFCAFVRFNO(Ours)
2D0.0670.0530.0580.054
CIFAR-100.0580.0560.0350.026
+ +Table 5. Comparison of the inference time (sec) in $N$ -step. + +
NFERFCAFVRFNO(Ours)
10.1720.1810.305
101.4041.4151.646
+ +for single-step and few-step generation. Specifically, we investigate the improvements brought to RF by (1) the introduction of the HVT and (2) the noise optimization operation. The configurations and results are shown in Tab.3. The two components have both contributed to the improvement of model performance. Regarding the introduction of the HVT, even with one-step sampling, the model performance improved, indicating that it facilitates trajectory rectification during the training process. As for the noise optimization operation, it results in a significant performance boost, demonstrating the effectiveness of training the neural velocity field with our defined optimized coupling. + +Flow Straightness. To evaluate the straightness of the inference trajectories, we introduce the Normalized Flow Straightness Score (NFSS) [14, 15]. A smaller NFSS indicates that the inference trajectory is closer to a straight line. We compare VRFNO with RF and CAF on synthetic datasets and CIFAR-10, with the experimental results shown in Tab.4. Our method generates inference trajectories that are straighter compared to the other methods. + +Time Efficiency. We also compare the inference time of our method with RF and CAF. Tab.5 presents the time required to generate 512 images in both one-step and ten-step settings. Our method incurs a slightly higher time cost than RF due to the additional encoder. In contrast, CAF consists of two neural velocity fields, it necessitates an additional + +computation (its total NFE equals $N + 1$ ), resulting in significantly higher time consumption. + +# 5. Related Work + +The inference process of DMs can be viewed as an iterative solution of ODE [25] or stochastic differential equations [2, 9], and ODE-based methods [5, 31] are more effective in few-step sampling due to their determinism. A widely used first-order ODE solver is the Euler solver, but its large local truncation error necessitates more iteration steps for high-quality image generation. To reduce iteration steps, higher order ODE solvers such as DEIS [31] and DPM solvers [16] exploit the semilinear structure of diffusion ODEs, deriving exact solutions for high-quality image generation with fewer steps. DPM-solver++ [17] enhances DPM-solver by improving stability in higher-order solvers. UniPC [33] builds a unified predictor-corrector framework for DPM solver, enhancing the quality of sampling in a few steps. DC-solver [34] uses dynamic compensation to correct misalignment in UniPC, thus improving image quality. + +Another mainstream method for accelerated sampling is distillation. Progressive distillation [24] iteratively halves the steps of a DDIM until it enables image generation in just 4 steps. Diff-Instruct [20] accelerates sampling by minimizing Integra Kullback-Leibler divergence to merge knowledge from multiple time steps of a pre-trained model. Consistency Model (CM) [28] enables direct mapping from noise to data by minimizing the differences between the final states when adjacent ODE trajectory points are used as inputs. Latent CM [19] treats the guided reverse diffusion process as solving an augmented PF ODE, directly predicting the ODE solution in the latent space, enabling fast and high-fidelity sampling. Consistency Trajectory Model [11] further optimizes CM by allowing the model to learn mappings between arbitrary initial and final times along the ODE trajectory during the diffusion process. + +# 6. Conclusion + +This paper proposes a novel method VRFNO, which is a joint training framework combining an encoder (for noise optimization) and a neural velocity field (for velocity prediction). VRFNO utilizes the encoder to optimize noises to transform randomly matched noises and images into optimized couplings, which are then used to train the neural velocity field for precise prediction of straight inference trajectories. To further straighten the inference trajectories during training, we introduce the HVT into the neural velocity field to enhance its prediction accuracy. Extensive experiments conducted on both synthetic and real world datasets demonstrate the effectiveness and scalability of VRFNO, and the results show that it achieves state-of-the-art performance across multiple datasets and varying resolutions. + +# Acknowledgments + +This work was supported in part by the National Science Fund of China (No.62276135, U24A20330 and 62361166670). + +# References + +[1] Jonas Adler and Sebastian Lunz. Banach wasserstein gan. Advances in neural information processing systems, 31, 2018. 1 +[2] Fan Bao, Chongxuan Li, Jun Zhu, and Bo Zhang. Analytic-dpm: an analytic estimate of the optimal reverse variance in diffusion probabilistic models. arXiv preprint arXiv:2201.06503, 2022.8 +[3] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8188-8197, 2020. 6 +[4] Tianyi Chu, Wei Xing, Jiafu Chen, Zhizhong Wang, Jiakai Sun, Lei Zhao, Haibo Chen, and Huaizhong Lin. Attack deterministic conditional image generative models for diverse and controllable generation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1362-1370, 2024. 1 +[5] Tim Dockhorn, Arash Vahdat, and Karsten Kreis. Genie: Higher-order denoising diffusion solvers. Advances in Neural Information Processing Systems, 35:30150-30166, 2022. 8 +[6] Luca Eyring, Shyamgopal Karthik, Karsten Roth, Alexey Dosovitskiy, and Zeynep Akata. Reno: Enhancing one-step text-to-image models through reward-based noise optimization. Advances in Neural Information Processing Systems, 37:125487-125519, 2025. 5 +[7] Kevin Frans, Danijar Hafner, Sergey Levine, and Pieter Abbeel. One step diffusion via shortcut models. arXiv preprint arXiv:2410.12557, 2024.6 +[8] Irina Higgins, Loic Matthew, Arka Pal, Christopher P Burgess, Xavier Glorot, Matthew M Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-vae: Learning basic visual concepts with a constrained variational framework. ICLR (Poster), 3, 2017. 5, 7 +[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 8 +[10] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35:26565-26577, 2022. 6 +[11] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ode trajectory of diffusion. arXiv preprint arXiv:2310.02279, 2023. 8 +[12] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6 +[13] Qiang Liu. Rectified flow: A marginal preserving approach to optimal transport. arXiv preprint arXiv:2209.14577, 2022. 2 + +[14] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2023. 1, 3, 6, 8 +[15] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In The Twelfth International Conference on Learning Representations, 2024. 1, 8 +[16] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 8 +[17] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 8 +[18] Eric Luhman and Troy Luhman. Knowledge distillation in iterative generative models for improved sampling speed. arXiv preprint arXiv:2101.02388, 2021. 6 +[19] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 8 +[20] Weijian Luo, Tianyang Hu, Shifeng Zhang, Jiacheng Sun, Zhenguo Li, and Zhihua Zhang. Diff-instruct: A universal approach for transferring knowledge from pre-trained diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 8 +[21] Dogyun Park, Sojin Lee, Sihyeon Kim, Taehoon Lee, Youngjoon Hong, and Hyunwoo J Kim. Constant acceleration flow. arXiv preprint arXiv:2411.00322, 2024. 6 +[22] Saptarshi Roy, Vansh Bansal, Purnamrita Sarkar, and Alessandro Rinaldo. 2-rectifications are enough for straight flows: A theoretical insight into wasserstein convergence. arXiv e-prints, pages arXiv-2410, 2024. 1 +[23] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 conference proceedings, pages 1-10, 2022. 1 +[24] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 6, 8 +[25] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 8 +[26] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. Advances in neural information processing systems, 33:12438-12448, 2020. 7 +[27] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 6, 7 +[28] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023. 6, 8 + +[29] Zhangkai Wu, Xuhui Fan, Hongyu Wu, and Longbing Cao. Traflow: Trajectory distillation on pre-trained rectified flow, 2025. 6 +[30] YoungJoon Yoo and Jongwon Choi. Topic-vq-vae: Leveraging latent codebooks for flexible topic-guided document generation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 19422–19430, 2024. 1 +[31] Qinsheng Zhang and Yongxin Chen. Fast sampling of diffusion models with exponential integrator. arXiv preprint arXiv:2204.13902, 2022. 8 +[32] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5 +[33] Wenliang Zhao, Lujia Bai, Yongming Rao, Jie Zhou, and Jiwen Lu. Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 8 +[34] Wenliang Zhao, Haolin Wang, Jie Zhou, and Jiwen Lu. Dcsolver: Improving predictor-corrector diffusion sampler via dynamic compensation. arXiv preprint arXiv:2409.03755, 2024.8 +[35]Zikai Zhou,Shitong Shao,Lichen Bai,Zhiqiang Xu,Bo Han,and Zeke Xie. Golden noise for diffusion models: A learning framework.arXiv preprint arXiv:2411.09502, 2024.5 \ No newline at end of file diff --git a/2025/Straighten Viscous Rectified Flow via Noise Optimization/images.zip b/2025/Straighten Viscous Rectified Flow via Noise Optimization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c937130a6aeedc2322dbef94d267d386e95b72d6 --- /dev/null +++ b/2025/Straighten Viscous Rectified Flow via Noise Optimization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8ca6d4732a43fe9185cbac255770a856459e8630ad76374f314dab1bcc08dab7 +size 634399 diff --git a/2025/Straighten Viscous Rectified Flow via Noise Optimization/layout.json b/2025/Straighten Viscous Rectified Flow via Noise Optimization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5974ea198b00fff3adbb79d564c2efb0181df2c7 --- /dev/null +++ b/2025/Straighten Viscous Rectified Flow via Noise Optimization/layout.json @@ -0,0 +1,8689 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 127, + 103, + 483, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 103, + 483, + 121 + ], + "spans": [ + { + "bbox": [ + 127, + 103, + 483, + 121 + ], + "type": "text", + "content": "Straighten Viscous Rectified Flow via Noise Optimization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 142, + 494, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 142, + 494, + 186 + ], + "spans": [ + { + "bbox": [ + 116, + 142, + 494, + 186 + ], + "type": "text", + "content": "Jimin Dai1, Jiexi Yan2, Jian Yang1, Lei Luo1 \n1PCA Lab, Nanjing University of Science and Technology, 2Xidian University {jimindai, csjyang}@njust.edu.cn, {jxyan1995, luoleipitt}@gmail.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 213, + 200, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 226 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 226 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 239, + 296, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 296, + 514 + ], + "type": "text", + "content": "The Reflow operation aims to straighten the inference trajectories of the rectified flow during training by constructing deterministic couplings between noises and images, thereby improving the quality of generated images in single-step or few-step generation. However, we identify critical limitations in Reflow, particularly its inability to rapidly generate high-quality images due to a distribution gap between images in its constructed deterministic couplings and real images. To address these shortcomings, we propose a novel alternative called Straighten Viscous Rectified Flow via Noise Optimization (VRFNO), which is a joint training framework integrating an encoder and a neural velocity field. VRFNO introduces two key innovations: (1) a historical velocity term that enhances trajectory distinction, enabling the model to more accurately predict the velocity of the current trajectory, and (2) the noise optimization through reparameterization to form optimized couplings with real images which are then utilized for training, effectively mitigating errors caused by Reflow's limitations. Comprehensive experiments on synthetic data and real datasets with varying resolutions show that VRFNO significantly mitigates the limitations of Reflow, achieving state-of-the-art performance in both one-step and few-step generation tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 537, + 135, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 135, + 550 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 135, + 550 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "content": "Diffusion models (DMs) [9, 23, 25] have attracted significant attention due to their remarkable generative capabilities and stable optimization process. DMs have outperformed traditional models such as generative adversarial networks (GANs) [1, 4] and Variational autoencoders (VAEs) [30] in various generation tasks. DMs work by designing a forward noise-adding scheme, which progressively transforms the target distribution into the initial distribution (usually a Gaussian distribution), thereby constructing a Probability Flow (PF) from complex to simple distributions. DMs learn the inverse process of PF, and then they can sample Gaussian noises and generate samples that approximate the target distribution through multi" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 214, + 555, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 214, + 555, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 214, + 555, + 334 + ], + "type": "text", + "content": "ple inference steps. Currently, various forward noise-adding schemes have been developed for DMs. Among them, the most straightforward approach is the Rectified Flow (RF) [14, 15, 22]. The core idea of RF is to add noise to the data in a linear manner, constructing a deterministic PF from the Gaussian distribution to the target distribution. The model only needs to learn the velocity field of the straight-line interpolation trajectories between the noises and the samples. Once trained, the model can efficiently generate samples that approximate the target distribution." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 336, + 556, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 556, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 556, + 504 + ], + "type": "text", + "content": "Generating samples with DMs typically requires multiple inference steps that consume substantial computational resources and time. RF constructs straight-line interpolation trajectories between noises and samples as the reference, and it theoretically learns a straight flow from the Gaussian distribution to the target distribution. The straight flow implies that RF learns a constant velocity field, which enables few-step or even single-step sampling in an ideal scenario. However, during actual training, the RF struggles to learn a constant velocity field, and the resulting PF trajectories remain curved. Therefore, RF still relies on multiple steps to generate the desired images. RF attributes the curvature to the crossing of reference trajectories during training and employs a Reflow operation to mitigate this issue." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 506, + 556, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 556, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 556, + 662 + ], + "type": "text", + "content": "In this paper, we observe that during the actual training process, the probability of crossing between the straight-line interpolation trajectories of randomly matched noises and images at a given time step " + }, + { + "bbox": [ + 313, + 506, + 556, + 662 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 506, + 556, + 662 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 313, + 506, + 556, + 662 + ], + "type": "inline_equation", + "content": "P \\sim O(e^{-c(n \\times n)})" + }, + { + "bbox": [ + 313, + 506, + 556, + 662 + ], + "type": "text", + "content": ", which is very small in high-dimensional data space (Theorem 1). Therefore, we are more inclined to believe that the curvature arises because the neural network model is not an exact solver, intermediate states along different trajectories may become \"approximately crossing\" due to their high similarity in statistical property, thereby interfering with the model predictions. Furthermore, we conducted an in-depth analysis of the factors that contribute to the success of Reflow, as well as its limitations, detailed in Section 2.2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "To retain the advantages of Reflow while mitigating its drawbacks, we propose a new method called Straighten Viscous Rectified Flow via Noise Optimization (VRFNO). VRFNO introduces a historical velocity term (HVT) as in" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15005" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 335 + ], + "type": "text", + "content": "put to the neural velocity field, allowing the model to distinguish the direction of the RF in the presence of approximate crossing. By incorporating historical velocity information, the model can more accurately predict the velocity. Moreover, VRFNO directly uses the original dataset to avoid the distribution gap and achieves more appropriate matches between the noises and the images (we refer to it as optimized coupling) by noise optimization. Specifically, we propose a joint training framework for an encoder and a neural velocity field. The encoder first encodes images from randomly matched noise-image pairs and then outputs the corresponding mean and variance matrices. The mean and variance matrices are combined with the noise using the reparameterization technique to generate a new noise, which is then fed into the neural velocity field for training. This joint training approach enables the encoder to adjust the mean and variance of the noise based on the image, thereby optimizing the matches between the noises and the images, which in turn achieves optimized couplings. By using these optimized couplings to train the velocity field, the model can more efficiently straighten the inference trajectory during training, enabling one-step or few-step generation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 336, + 251, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 336, + 251, + 346 + ], + "spans": [ + { + "bbox": [ + 67, + 336, + 251, + 346 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 347, + 296, + 467 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 347, + 295, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 295, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 295, + 383 + ], + "type": "text", + "content": "- We propose a new method VRFNO to straighten the inference trajectories for higher quality few-step and single-step generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 383, + 295, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 383, + 295, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 383, + 295, + 419 + ], + "type": "text", + "content": "- We introduce an HVT as the auxiliary information to enhance the model's accuracy in velocity prediction when the flow approximates crossing." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 419, + 296, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 296, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 296, + 467 + ], + "type": "text", + "content": "- We construct optimized couplings by optimizing the noise to train the neural velocity field. This avoids relying on deterministic couplings used in Reflow, which will constrain the model's generation quality." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 477, + 133, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 477, + 133, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 477, + 133, + 491 + ], + "type": "text", + "content": "2. Preliminary" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 497, + 145, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 145, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 145, + 509 + ], + "type": "text", + "content": "2.1. Rectified Flow" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 514, + 296, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 296, + 586 + ], + "type": "text", + "content": "RF is an ordinary differential equations (ODE)-based DM that aims to learn the mapping between two distributions, " + }, + { + "bbox": [ + 55, + 514, + 296, + 586 + ], + "type": "inline_equation", + "content": "\\pi_0" + }, + { + "bbox": [ + 55, + 514, + 296, + 586 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 514, + 296, + 586 + ], + "type": "inline_equation", + "content": "\\pi_{1}" + }, + { + "bbox": [ + 55, + 514, + 296, + 586 + ], + "type": "text", + "content": ". In image generation tasks, " + }, + { + "bbox": [ + 55, + 514, + 296, + 586 + ], + "type": "inline_equation", + "content": "\\pi_0" + }, + { + "bbox": [ + 55, + 514, + 296, + 586 + ], + "type": "text", + "content": " typically represents a standard Gaussian distribution, while " + }, + { + "bbox": [ + 55, + 514, + 296, + 586 + ], + "type": "inline_equation", + "content": "\\pi_{1}" + }, + { + "bbox": [ + 55, + 514, + 296, + 586 + ], + "type": "text", + "content": " corresponds to the target image distribution. For empirical observations " + }, + { + "bbox": [ + 55, + 514, + 296, + 586 + ], + "type": "inline_equation", + "content": "X_0\\sim" + }, + { + "bbox": [ + 55, + 514, + 296, + 586 + ], + "type": "inline_equation", + "content": "\\pi_0" + }, + { + "bbox": [ + 55, + 514, + 296, + 586 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 514, + 296, + 586 + ], + "type": "inline_equation", + "content": "X_{1}\\sim \\pi_{1}" + }, + { + "bbox": [ + 55, + 514, + 296, + 586 + ], + "type": "text", + "content": " on time " + }, + { + "bbox": [ + 55, + 514, + 296, + 586 + ], + "type": "inline_equation", + "content": "t\\in [0,1]" + }, + { + "bbox": [ + 55, + 514, + 296, + 586 + ], + "type": "text", + "content": ", RF is defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 136, + 596, + 295, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 596, + 295, + 609 + ], + "spans": [ + { + "bbox": [ + 136, + 596, + 295, + 609 + ], + "type": "interline_equation", + "content": "d X _ {t} = v \\left(X _ {t}, t\\right) d t, \\tag {1}", + "image_path": "a56204df965b0e2a73a352f3834d5ac37688a704cd8fa285aeba5a4f1a15de10.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "content": "where intermediate state " + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "inline_equation", + "content": "X_{t} = tX_{1} + (1 - t)X_{0}" + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "content": " represents a time-differentiable interpolation between " + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "inline_equation", + "content": "X_{0}" + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "inline_equation", + "content": "X_{1}" + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "inline_equation", + "content": "v: \\mathbb{R}^{n\\times n}\\times [0,1]\\to \\mathbb{R}^{n\\times n}" + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "content": " denotes the velocity field defined on the data-time domain. Since RF uses simple straight-line interpolation to connect " + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "inline_equation", + "content": "X_{0}" + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "inline_equation", + "content": "X_{1}" + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "content": ", its trajectory has a constant velocity field " + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "inline_equation", + "content": "v_{ref} = X_1 - X_0" + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "content": ". Thus, the training process optimizes the model by solving a least-squares regression problem, i.e., fitting " + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "inline_equation", + "content": "v_{ref}" + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "content": " to neural ve" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 553, + 235 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 553, + 235 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 553, + 235 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 553, + 235 + ], + "type": "image", + "image_path": "d51084fa122bfcd920a78e0cdf4d9fc6dffce3275c9706661f3913e1e0e76166.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 236, + 555, + 325 + ], + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 325 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 325 + ], + "type": "text", + "content": "Figure 1. Comparison between Reflow and Vanilla training mode. The top illustrates the Reflow training mode, where noises generate images via a pre-trained model, forming deterministic couplings for training. These couplings are reused during training by randomly sampling different intermediate states along the trajectories. The bottom describes the vanilla training mode, where noises and images are randomly sampled to form arbitrary couplings for training, without reusing data during training." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 344, + 375, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 344, + 375, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 344, + 375, + 357 + ], + "type": "text", + "content": "locity field " + }, + { + "bbox": [ + 313, + 344, + 375, + 357 + ], + "type": "inline_equation", + "content": "v_{\\theta}" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 342, + 364, + 553, + 385 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 364, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 342, + 364, + 553, + 385 + ], + "type": "interline_equation", + "content": "\\min \\mathbb {E} _ {X _ {0}, X _ {1} \\sim \\gamma , t \\in p (t)} \\left[ \\| v _ {\\text {r e f}} - v _ {\\theta} \\left(X _ {t}, t\\right) \\| ^ {2} \\right], \\tag {2}", + "image_path": "717dd27e3ed60d1d38c4b878b03cad56d6885f9243bf21729ab1c9bd5f4118c8.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 392, + 555, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 555, + 441 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 392, + 555, + 441 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 392, + 555, + 441 + ], + "type": "text", + "content": " represents the coupling of " + }, + { + "bbox": [ + 313, + 392, + 555, + 441 + ], + "type": "inline_equation", + "content": "(X_0, X_1)" + }, + { + "bbox": [ + 313, + 392, + 555, + 441 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 392, + 555, + 441 + ], + "type": "inline_equation", + "content": "p(t)" + }, + { + "bbox": [ + 313, + 392, + 555, + 441 + ], + "type": "text", + "content": " denotes the time distribution defined on " + }, + { + "bbox": [ + 313, + 392, + 555, + 441 + ], + "type": "inline_equation", + "content": "[0, 1]" + }, + { + "bbox": [ + 313, + 392, + 555, + 441 + ], + "type": "text", + "content": ". During the inference process, the ODE usually needs to be discretized and solved via the Euler method, expressed as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 449, + 553, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 449, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 315, + 449, + 553, + 473 + ], + "type": "interline_equation", + "content": "X _ {t + \\Delta t} = X _ {t} + \\Delta t \\cdot v _ {\\theta} \\left(X _ {t}, t\\right), t \\in \\{0, \\Delta t, \\dots , (N - 1) \\Delta t \\}, \\tag {3}", + "image_path": "90b564f1c3c0cc9fba64653af0b04c3b26f6e9a502e1c867f5ef5d8dd7929608.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 474, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 617 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 474, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\Delta t = \\frac{1}{N}" + }, + { + "bbox": [ + 313, + 474, + 555, + 617 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 474, + 555, + 617 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 474, + 555, + 617 + ], + "type": "text", + "content": " represents the total number of steps. Generally, a larger " + }, + { + "bbox": [ + 313, + 474, + 555, + 617 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 474, + 555, + 617 + ], + "type": "text", + "content": " results in higher-quality generated images, but it requires more computational resources for sampling. The reference trajectory of RF is a straight flow. If RF can converge to the straight flow through training, it can significantly minimize numerical errors in the ODE solver, enabling few-step or single-step generation: as the state moves along the trajectory with uniform linear motion, the velocity at a single time step equals the average velocity over the entire trajectory. As long as the cumulative sum of all time increments throughout the motion equals unity, the final endpoint of the trajectory will remain consistent." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 624, + 423, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 423, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 423, + 636 + ], + "type": "text", + "content": "2.2. Analysis of Reflow" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "content": "The inference trajectory of RF exhibits a curved pattern when trained using the vanilla method. While RF attributes this phenomenon to the crossing of reference trajectories [13] and proposes Reflow as a solution (with implementation details provided in Appendix C), our experiments (Fig.2) and analysis suggest that the effectiveness" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15006" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 72, + 291, + 251 + ], + "blocks": [ + { + "bbox": [ + 60, + 72, + 291, + 251 + ], + "lines": [ + { + "bbox": [ + 60, + 72, + 291, + 251 + ], + "spans": [ + { + "bbox": [ + 60, + 72, + 291, + 251 + ], + "type": "image", + "image_path": "df502e8514c154ff40587f2cef1d98da81f090d9336cf65457bedd50a594772d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 255, + 295, + 344 + ], + "lines": [ + { + "bbox": [ + 55, + 255, + 295, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 295, + 344 + ], + "type": "text", + "content": "Figure 2. Impact of coupling type and data reuse on RF's performance. Training RF from scratch with different coupling types, we control the data volume to construct different data reuse scenarios and observe the impact of coupling type and data reuse on model performance under the same training iterations. Each iteration samples 500 data pairs, ensuring that the total data volume seen by the model remains the same across different reuse scenarios under the same training iterations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 365, + 295, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 414 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 414 + ], + "type": "text", + "content": "of Reflow cannot be solely explained by its ability to mitigate trajectory-crossing issues. This is particularly evident given that trajectory crossings are extremely rare in high-dimensional spaces, as we will elaborate in Section 3.1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 415, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 295, + 714 + ], + "type": "text", + "content": "Through experimental and analytical investigations, the success of Reflow can also be attributed to the following two factors: (1) the images in the training data for Reflow are generated by the pre-trained model, i.e., deterministic couplings. This implies the existence of a deterministic trajectory between a coupling that can be learned and inferred by the model, as shown in the first row of Fig.1. Training/finetuning with Reflow is analogous to learning this trajectory during training and progressively aligning it with a straight-line trajectory. In contrast, randomly matched data pairs (arbitrary couplings) lack explicit trajectory relationships, requiring the model to infer unknown trajectories from noises and constrain them to straight lines, making the training more complex and unstable. As demonstrated in Fig.2, under identical data volume and training iterations, RF trained with deterministic couplings achieves superior image quality (green line vs. orange line). (2) Reflow reuses data by sampling intermediate states from different time steps of the same trajectory for training, as depicted in the first row of Fig.1. This training strategy resembles the multi-time-scale optimization methods employed in the distillation, both of which aim to enhance model performance by leveraging multi-level information. Fig.2 demonstrates that more data reuse iterations lead to better performance under deterministic couplings (orange line vs. blue line)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 538, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 538, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 538, + 84 + ], + "type": "text", + "content": "More details and analyses can be found in Appendix B." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 85, + 555, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 85, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 85, + 555, + 289 + ], + "type": "text", + "content": "Additionally, the drawbacks of Reflow are evident. As shown in Fig.1, there is a distribution gap between the generated images and the real images. Each iteration of training utilizes images generated by the previous model as training data, leading to the accumulation of errors over time. Consequently, the quality of the images generated by the subsequent model tends to degrade compared to the previous one. To balance high image quality with computational efficiency, it is typically recommended to limit the iterative training to 2 or 3 cycles. Furthermore, while data reuse is effective, its effectiveness is influenced by the volume of data. If the volume of reused data is small, it may not adequately represent the true probabilistic distribution of the data, thereby affecting the model's learning of the PF [14]. On the other hand, excessive reused data will cause significant storage pressure, which increases the consumption of computational resources and reduces training efficiency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 297, + 370, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 370, + 309 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 370, + 309 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 318, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 555, + 510 + ], + "type": "text", + "content": "We aim to develop an advanced training framework for RF as a superior alternative to Reflow, further improving the quality of generated images in single-step and few-step generation tasks. To achieve this, we propose a novel method called VRFNO and the overview of VRFNO is shown in Fig.3. Specifically, VRFNO introduces an HVT to effectively improves the model's capability to differentiate inference trajectories, thereby substantially mitigating prediction disturbances arising from intermediate state similarities. Furthermore, to circumvent the use of pre-trained model-generated images as training data during the training process, as employed in Reflow, and to ensure that randomly matched noise-image data pairs " + }, + { + "bbox": [ + 313, + 318, + 555, + 510 + ], + "type": "inline_equation", + "content": "(X_0,X_1)" + }, + { + "bbox": [ + 313, + 318, + 555, + 510 + ], + "type": "text", + "content": " exhibit desirable characteristics akin to deterministic coupling. We define a new concept \"optimized coupling\" and propose to achieve optimized coupling by optimizing noise." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 516, + 555, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 555, + 540 + ], + "type": "text", + "content": "Definition 1. " + }, + { + "bbox": [ + 313, + 516, + 555, + 540 + ], + "type": "inline_equation", + "content": "(X_0, X_1)" + }, + { + "bbox": [ + 313, + 516, + 555, + 540 + ], + "type": "text", + "content": " is called an optimized coupling if it satisfies the following condition:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 345, + 548, + 523, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 548, + 523, + 562 + ], + "spans": [ + { + "bbox": [ + 345, + 548, + 523, + 562 + ], + "type": "interline_equation", + "content": "\\left\\| v _ {\\theta} \\left(t X _ {1} + (1 - t) X _ {0}\\right) - \\left(X _ {1} - X _ {0}\\right) \\right\\| \\leq \\varepsilon ,", + "image_path": "d1ef5e237c33fd326a65ad6538fe0ed176625f911c6d03a90212eb16bd790b66.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 570, + 479, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 479, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 479, + 582 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 570, + 479, + 582 + ], + "type": "inline_equation", + "content": "\\varepsilon > 0" + }, + { + "bbox": [ + 313, + 570, + 479, + 582 + ], + "type": "text", + "content": " is a small positive constant." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 588, + 442, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 442, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 442, + 600 + ], + "type": "text", + "content": "3.1. Viscous Rectified Flow" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "RF attributes the curvature of inference trajectories to the frequent crossings of reference trajectories. Specifically, when noise-image pairs are randomly matched, multiple reference trajectories may cross at the same point " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "X_{t}" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " at time step " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": ". During training, this causes different reference trajectories to provide identical inputs to the model at time step " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": ", confusing the model and degrading its prediction accuracy. However, our calculations (Theorem 1) show that the probability of trajectory crossings occurring during actual" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15007" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 86, + 328, + 201 + ], + "blocks": [ + { + "bbox": [ + 86, + 70, + 255, + 87 + ], + "lines": [ + { + "bbox": [ + 86, + 70, + 255, + 87 + ], + "spans": [ + { + "bbox": [ + 86, + 70, + 255, + 87 + ], + "type": "text", + "content": "Optimized coupling " + }, + { + "bbox": [ + 86, + 70, + 255, + 87 + ], + "type": "inline_equation", + "content": "(X_0, X_1)" + }, + { + "bbox": [ + 86, + 70, + 255, + 87 + ], + "type": "text", + "content": " is obtained by optimizing the noise " + }, + { + "bbox": [ + 86, + 70, + 255, + 87 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 86, + 70, + 255, + 87 + ], + "type": "text", + "content": " through the encoder using the reparameterization technique." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 86, + 328, + 201 + ], + "lines": [ + { + "bbox": [ + 60, + 86, + 328, + 201 + ], + "spans": [ + { + "bbox": [ + 60, + 86, + 328, + 201 + ], + "type": "image", + "image_path": "8407e436529b4479a37ca8bb6f696ed88a276bfb966c2447c5097d11e28299b7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 204, + 555, + 250 + ], + "lines": [ + { + "bbox": [ + 54, + 204, + 555, + 250 + ], + "spans": [ + { + "bbox": [ + 54, + 204, + 555, + 250 + ], + "type": "text", + "content": "Figure 3. Overview of VRFNO. The encoder and the neural velocity field form a joint training framework: randomly matched noise " + }, + { + "bbox": [ + 54, + 204, + 555, + 250 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 54, + 204, + 555, + 250 + ], + "type": "text", + "content": " and image " + }, + { + "bbox": [ + 54, + 204, + 555, + 250 + ], + "type": "inline_equation", + "content": "X_{1}" + }, + { + "bbox": [ + 54, + 204, + 555, + 250 + ], + "type": "text", + "content": " (called arbitrary coupling) are optimized by the encoder to obtain optimized coupling " + }, + { + "bbox": [ + 54, + 204, + 555, + 250 + ], + "type": "inline_equation", + "content": "(X_0,X_1)" + }, + { + "bbox": [ + 54, + 204, + 555, + 250 + ], + "type": "text", + "content": ", which are then used to train the neural velocity field. The introduction of the HVT " + }, + { + "bbox": [ + 54, + 204, + 555, + 250 + ], + "type": "inline_equation", + "content": "v_{history}" + }, + { + "bbox": [ + 54, + 204, + 555, + 250 + ], + "type": "text", + "content": " enhances the distinction of the trajectories. Both components work together during training to straighten the inference trajectories, thereby improving the model's predictive accuracy." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 328, + 87, + 548, + 201 + ], + "blocks": [ + { + "bbox": [ + 338, + 70, + 538, + 87 + ], + "lines": [ + { + "bbox": [ + 338, + 70, + 538, + 87 + ], + "spans": [ + { + "bbox": [ + 338, + 70, + 538, + 87 + ], + "type": "text", + "content": "The historical velocity term " + }, + { + "bbox": [ + 338, + 70, + 538, + 87 + ], + "type": "inline_equation", + "content": "\\nu_{\\text{history}}" + }, + { + "bbox": [ + 338, + 70, + 538, + 87 + ], + "type": "text", + "content": " is introduced into the neural velocity field, and its result is obtained by the prediction method shown on the dotted line." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 328, + 87, + 548, + 201 + ], + "lines": [ + { + "bbox": [ + 328, + 87, + 548, + 201 + ], + "spans": [ + { + "bbox": [ + 328, + 87, + 548, + 201 + ], + "type": "image", + "image_path": "3f7f17ed349f21aeac53746a098fb6ab3513f7455c013d3acbf04adf00d87627.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 269, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 269, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 54, + 269, + 296, + 437 + ], + "type": "text", + "content": "training is " + }, + { + "bbox": [ + 54, + 269, + 296, + 437 + ], + "type": "inline_equation", + "content": "P \\sim O(e^{-c(n \\times n)})" + }, + { + "bbox": [ + 54, + 269, + 296, + 437 + ], + "type": "text", + "content": ", which is extremely rare in high-dimensional space. More commonly, since the initial states are independently sampled from a standard Gaussian distribution, their statistical properties are very similar. In the early stages of the interpolation trajectory, the intermediate states contain limited image information, causing the differences between the intermediate states to remain small. Since RF learns the mapping relationship between distributions, this high similarity in statistical properties may hinder the model's ability to distinguish between states, leading to blurred predictions. Ideally, RF has a constant velocity field, which means that the intermediate state " + }, + { + "bbox": [ + 54, + 269, + 296, + 437 + ], + "type": "inline_equation", + "content": "X_{t}" + }, + { + "bbox": [ + 54, + 269, + 296, + 437 + ], + "type": "text", + "content": " moves in a straight line from noise " + }, + { + "bbox": [ + 54, + 269, + 296, + 437 + ], + "type": "inline_equation", + "content": "X_{0}" + }, + { + "bbox": [ + 54, + 269, + 296, + 437 + ], + "type": "text", + "content": " to image " + }, + { + "bbox": [ + 54, + 269, + 296, + 437 + ], + "type": "inline_equation", + "content": "X_{1}" + }, + { + "bbox": [ + 54, + 269, + 296, + 437 + ], + "type": "text", + "content": " at uniform velocity " + }, + { + "bbox": [ + 54, + 269, + 296, + 437 + ], + "type": "inline_equation", + "content": "v_{ref}" + }, + { + "bbox": [ + 54, + 269, + 296, + 437 + ], + "type": "text", + "content": ", forming a straight inference trajectory." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 446, + 296, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 296, + 502 + ], + "type": "text", + "content": "Theorem 1. In " + }, + { + "bbox": [ + 55, + 446, + 296, + 502 + ], + "type": "inline_equation", + "content": "(n\\times n)" + }, + { + "bbox": [ + 55, + 446, + 296, + 502 + ], + "type": "text", + "content": "-dimensional space, for straight-line interpolation trajectories " + }, + { + "bbox": [ + 55, + 446, + 296, + 502 + ], + "type": "inline_equation", + "content": "X^{(\\cdot)} = \\left\\{X_{t}^{(\\cdot)}:t\\in [0,1]\\right\\}" + }, + { + "bbox": [ + 55, + 446, + 296, + 502 + ], + "type": "text", + "content": ", the probability of " + }, + { + "bbox": [ + 55, + 446, + 296, + 502 + ], + "type": "inline_equation", + "content": "X^{(i)}" + }, + { + "bbox": [ + 55, + 446, + 296, + 502 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 446, + 296, + 502 + ], + "type": "inline_equation", + "content": "X^{(j)}" + }, + { + "bbox": [ + 55, + 446, + 296, + 502 + ], + "type": "text", + "content": " crossing at the point " + }, + { + "bbox": [ + 55, + 446, + 296, + 502 + ], + "type": "inline_equation", + "content": "X_{t}" + }, + { + "bbox": [ + 55, + 446, + 296, + 502 + ], + "type": "text", + "content": " at time step " + }, + { + "bbox": [ + 55, + 446, + 296, + 502 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 446, + 296, + 502 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 55, + 446, + 296, + 502 + ], + "type": "inline_equation", + "content": "P\\sim O(e^{-c(n\\times n)}),c > 0" + }, + { + "bbox": [ + 55, + 446, + 296, + 502 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 511, + 296, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 511, + 296, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 511, + 296, + 620 + ], + "type": "text", + "content": "As shown in Theorem 2, the velocity difference " + }, + { + "bbox": [ + 55, + 511, + 296, + 620 + ], + "type": "inline_equation", + "content": "\\Delta(v_{ref}^{(i)}, v_{ref}^{(j)})" + }, + { + "bbox": [ + 55, + 511, + 296, + 620 + ], + "type": "text", + "content": " between each " + }, + { + "bbox": [ + 55, + 511, + 296, + 620 + ], + "type": "inline_equation", + "content": "X_t" + }, + { + "bbox": [ + 55, + 511, + 296, + 620 + ], + "type": "text", + "content": " is more significant than their state difference " + }, + { + "bbox": [ + 55, + 511, + 296, + 620 + ], + "type": "inline_equation", + "content": "\\Delta(X_t^{(i)}, X_t^{(j)})" + }, + { + "bbox": [ + 55, + 511, + 296, + 620 + ], + "type": "text", + "content": ", and this more significant difference enables the model to better identify the currently inferred trajectory, thus improving prediction accuracy. Therefore, we introduce an HVT " + }, + { + "bbox": [ + 55, + 511, + 296, + 620 + ], + "type": "inline_equation", + "content": "v_{history}" + }, + { + "bbox": [ + 55, + 511, + 296, + 620 + ], + "type": "text", + "content": " as an approximation of the constant velocity, which is fed into the model at each time step, resulting in the Viscous Rectified Flow (VRF):" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 92, + 630, + 295, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 630, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 92, + 630, + 295, + 644 + ], + "type": "interline_equation", + "content": "d X _ {t} = v \\left(X _ {t}, t, v _ {\\text {h i s t o r y}}\\right) d t, \\quad t \\in [ \\Delta t, 1 ], \\tag {4}", + "image_path": "e7eb7e34337118bf15f02fd4d031ef319bc33633964fa49510203183c40db160.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " is the time interval between the current state and the adjacent history state. During the inference process, the HVT input to the model at the first time step is 0. In subsequent sampling iterations, each iteration inputs the velocity predicted by the model at the previous time step into the" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 269, + 454, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 269, + 454, + 281 + ], + "spans": [ + { + "bbox": [ + 315, + 269, + 454, + 281 + ], + "type": "text", + "content": "Algorithm 1 Training of VRFNO." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 323, + 284, + 554, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 284, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 323, + 284, + 554, + 308 + ], + "type": "text", + "content": "Input: image " + }, + { + "bbox": [ + 323, + 284, + 554, + 308 + ], + "type": "inline_equation", + "content": "X_{1}" + }, + { + "bbox": [ + 323, + 284, + 554, + 308 + ], + "type": "text", + "content": ", noise " + }, + { + "bbox": [ + 323, + 284, + 554, + 308 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 323, + 284, + 554, + 308 + ], + "type": "text", + "content": ", learning rate " + }, + { + "bbox": [ + 323, + 284, + 554, + 308 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 323, + 284, + 554, + 308 + ], + "type": "text", + "content": ", time interval " + }, + { + "bbox": [ + 323, + 284, + 554, + 308 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 323, + 284, + 554, + 308 + ], + "type": "text", + "content": ", hyperparameter " + }, + { + "bbox": [ + 323, + 284, + 554, + 308 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 323, + 284, + 554, + 308 + ], + "type": "text", + "content": ", encoder " + }, + { + "bbox": [ + 323, + 284, + 554, + 308 + ], + "type": "inline_equation", + "content": "E_{\\phi}" + }, + { + "bbox": [ + 323, + 284, + 554, + 308 + ], + "type": "text", + "content": ", velocity field " + }, + { + "bbox": [ + 323, + 284, + 554, + 308 + ], + "type": "inline_equation", + "content": "v_{\\theta}" + }, + { + "bbox": [ + 323, + 284, + 554, + 308 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 325, + 310, + 354, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 310, + 354, + 319 + ], + "spans": [ + { + "bbox": [ + 325, + 310, + 354, + 319 + ], + "type": "text", + "content": "repeat" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 334, + 320, + 403, + 332 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 320, + 403, + 332 + ], + "spans": [ + { + "bbox": [ + 334, + 320, + 403, + 332 + ], + "type": "interline_equation", + "content": "\\mu , \\sigma^ {2} = E _ {\\phi} (X _ {1})", + "image_path": "8a1039884dbbf8b676827c132201583a4ab66e9cb9877eafc720aa2732f0c637.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 335, + 332, + 403, + 344 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 332, + 403, + 344 + ], + "spans": [ + { + "bbox": [ + 335, + 332, + 403, + 344 + ], + "type": "interline_equation", + "content": "X _ {0} = \\epsilon \\cdot \\sigma^ {2} + \\mu", + "image_path": "adfe929a5d4a9e75fd445dff7d14976145f2f557f8b8d58445d1fdf17dec46bd.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 335, + 345, + 405, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 345, + 405, + 357 + ], + "spans": [ + { + "bbox": [ + 335, + 345, + 405, + 357 + ], + "type": "interline_equation", + "content": "v _ {r e f} = X _ {1} - X _ {0}", + "image_path": "d25e45355cd83c3cce0a6b9eeda80bf3552f977de0c66355338d58d6ee46dc23.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 335, + 357, + 423, + 368 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 357, + 423, + 368 + ], + "spans": [ + { + "bbox": [ + 335, + 357, + 423, + 368 + ], + "type": "interline_equation", + "content": "\\text {S a m p l e} t \\sim U (\\Delta t, 1)", + "image_path": "992af043f90e7f4d86832548aed389c0be3af139e37c709ba08781b0d118a70a.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 335, + 369, + 432, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 369, + 432, + 380 + ], + "spans": [ + { + "bbox": [ + 335, + 369, + 432, + 380 + ], + "type": "interline_equation", + "content": "X _ {t} = t X _ {1} + (1 - t) X _ {0}", + "image_path": "1759df9adef3467ce41928140a3c20930fca66c1acd6213d32c93a00a1999a75.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 335, + 380, + 503, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 380, + 503, + 392 + ], + "spans": [ + { + "bbox": [ + 335, + 380, + 503, + 392 + ], + "type": "interline_equation", + "content": "X _ {t - \\Delta t} = (t - \\Delta t) X _ {1} + (1 - t + \\Delta t) X _ {0}", + "image_path": "ab9a632bfe1aa749994b73bd2e3e596a7ed4cdc44d7fab2bee6dac2162503a9e.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 335, + 392, + 514, + 404 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 392, + 514, + 404 + ], + "spans": [ + { + "bbox": [ + 335, + 392, + 514, + 404 + ], + "type": "interline_equation", + "content": "v _ {\\text {h i s t o r y}} = \\operatorname {s t o p g r a d} \\left(v _ {\\theta} \\left(X _ {t - \\Delta t}, t - \\Delta t, 0\\right)\\right)", + "image_path": "bab5d053fd848deae27eff83ebfe204c2d562a1996b2de3f96246414f89579ab.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 335, + 404, + 432, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 404, + 432, + 416 + ], + "spans": [ + { + "bbox": [ + 335, + 404, + 432, + 416 + ], + "type": "interline_equation", + "content": "v _ {t} = v _ {\\theta} \\left(X _ {t}, t, v _ {\\text {h i s t o r y}}\\right)", + "image_path": "bd4bc5e74896fd478a194db43164abe62551b0ddb4e52ef5f43752c69a038d29.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 335, + 416, + 380, + 427 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 416, + 380, + 427 + ], + "spans": [ + { + "bbox": [ + 335, + 416, + 380, + 427 + ], + "type": "interline_equation", + "content": "L (\\theta , \\phi) \\leftarrow", + "image_path": "6f0f9056690f729fae58cc634535c9679dcad77cbcc68eb1ad3672866913d3da.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 345, + 427, + 524, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 427, + 524, + 441 + ], + "spans": [ + { + "bbox": [ + 345, + 427, + 524, + 441 + ], + "type": "interline_equation", + "content": "\\mathbb {E} \\left[ d \\left(v _ {r e f}, v _ {t}\\right) \\right] + \\frac {\\alpha}{2} \\left(\\sigma^ {2} + \\mu^ {2} - 1 - \\log \\left(\\sigma^ {2}\\right)\\right)", + "image_path": "921cfa208efb08678083ac2b7ef70fdd84efa065db15e2e49712e27713120733.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 335, + 441, + 422, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 441, + 422, + 453 + ], + "spans": [ + { + "bbox": [ + 335, + 441, + 422, + 453 + ], + "type": "interline_equation", + "content": "\\theta \\leftarrow \\theta - \\eta \\nabla_ {\\theta} L (\\theta , \\phi)", + "image_path": "e50cc6b6158dd7e14744cb7abd0b3c7dab38f2baabc351116cb2aa03fb761093.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 335, + 453, + 425, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 453, + 425, + 464 + ], + "spans": [ + { + "bbox": [ + 335, + 453, + 425, + 464 + ], + "type": "interline_equation", + "content": "\\phi \\leftarrow \\phi - \\eta \\nabla_ {\\phi} L (\\theta , \\phi)", + "image_path": "b8a755e5ff8456f6fb0536a2565789ef2914524befaab90943990de0a8a9e6cd.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 325, + 464, + 400, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 464, + 400, + 475 + ], + "spans": [ + { + "bbox": [ + 325, + 464, + 400, + 475 + ], + "type": "text", + "content": "until convergence" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 499, + 555, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 499, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 499, + 555, + 536 + ], + "type": "text", + "content": "model as an HVT, providing the model with auxiliary information on the flow direction and enhancing the model's ability to make correct predictions." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 543, + 555, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 555, + 579 + ], + "type": "text", + "content": "Theorem 2. For each intermediate state " + }, + { + "bbox": [ + 313, + 543, + 555, + 579 + ], + "type": "inline_equation", + "content": "X_{t}" + }, + { + "bbox": [ + 313, + 543, + 555, + 579 + ], + "type": "text", + "content": " along a PF trajectory, the velocity difference between " + }, + { + "bbox": [ + 313, + 543, + 555, + 579 + ], + "type": "inline_equation", + "content": "X_{t}" + }, + { + "bbox": [ + 313, + 543, + 555, + 579 + ], + "type": "text", + "content": " are greater than their state difference:" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 354, + 585, + 514, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 585, + 514, + 601 + ], + "spans": [ + { + "bbox": [ + 354, + 585, + 514, + 601 + ], + "type": "interline_equation", + "content": "\\Delta (v _ {r e f} ^ {(i)}, v _ {r e f} ^ {(j)}) \\geq \\Delta (X _ {t} ^ {(i)}, X _ {t} ^ {(j)}),", + "image_path": "11a045421c27e9f2f8855492a40fe2d65cdc26a0e9f8bf130500b5ad6ee7ab2c.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 609, + 538, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 538, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 538, + 623 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 609, + 538, + 623 + ], + "type": "inline_equation", + "content": "\\Delta (\\cdot ,\\cdot) = \\mathbb{E}[\\| \\cdot -\\| _F^2 ]" + }, + { + "bbox": [ + 313, + 609, + 538, + 623 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 609, + 538, + 623 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[\\cdot ]" + }, + { + "bbox": [ + 313, + 609, + 538, + 623 + ], + "type": "text", + "content": " is the expectation." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 635, + 542, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 542, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 542, + 649 + ], + "type": "text", + "content": "3.2. Noise Optimization for Optimized Coupling" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": "We propose using noise reparameterization technique to optimize the noise representation. Specifically, we design a joint training framework for an encoder and a neural velocity field (as shown in Fig.3), where the encoder optimizes the noises to achieve optimized couplings and uses" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15008" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": "these optimized data pairs to train the neural velocity field. The complete training process is shown in Algorithm 1. We adopt the working mode of the encoder in the VAE architecture [8]: the encoder " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "E_{\\phi}" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " takes the image " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "X_{1}" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " as input and outputs the corresponding mean matrix " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " and variance matrix " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\sigma^2" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": ", then generates the optimized noise " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "X_{0}" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " by combining " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\sigma^2" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " with randomly matched noise " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " using the reparameterization technique:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 140, + 173, + 295, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 173, + 295, + 186 + ], + "spans": [ + { + "bbox": [ + 140, + 173, + 295, + 186 + ], + "type": "interline_equation", + "content": "X _ {0} = \\epsilon \\cdot \\sigma^ {2} + \\mu . \\tag {5}", + "image_path": "a22cdc6a73769c46c107a2e4ecb089e628896a37345067ed3c3dd78a01905ff4.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 192, + 296, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 296, + 408 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 296, + 408 + ], + "type": "text", + "content": "Given the limited number of images in the dataset, to prevent the model from developing a memory effect and ensure diversity in generated images, we introduce random perturbation " + }, + { + "bbox": [ + 55, + 192, + 296, + 408 + ], + "type": "inline_equation", + "content": "\\tau \\sim N(0,\\mathrm{I})" + }, + { + "bbox": [ + 55, + 192, + 296, + 408 + ], + "type": "text", + "content": " at the intermediate layers of the encoder, as shown in Fig.3. This ensures that the reparameterized mean and variance matrices still exhibit differences even when multiple Gaussian noises are matched to the same image. Furthermore, after these Gaussian noises matched the same image are reparameterized, they tend to concentrate in a smaller subspace. During training, each image is matched with noise from its corresponding subspace and participates in the training. This approach is similar to the data reuse strategy in Reflow, and it avoids the issue of limiting the model's generalization ability due to the reuse of limited data, thereby helping to improve the performance of the model. After obtaining the optimized noise " + }, + { + "bbox": [ + 55, + 192, + 296, + 408 + ], + "type": "inline_equation", + "content": "X_0" + }, + { + "bbox": [ + 55, + 192, + 296, + 408 + ], + "type": "text", + "content": ", we perform straight-line interpolation between it and the corresponding image " + }, + { + "bbox": [ + 55, + 192, + 296, + 408 + ], + "type": "inline_equation", + "content": "X_{1}" + }, + { + "bbox": [ + 55, + 192, + 296, + 408 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 55, + 192, + 296, + 408 + ], + "type": "inline_equation", + "content": "X_{t}" + }, + { + "bbox": [ + 55, + 192, + 296, + 408 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 192, + 296, + 408 + ], + "type": "inline_equation", + "content": "X_{t - \\Delta t}" + }, + { + "bbox": [ + 55, + 192, + 296, + 408 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 413, + 295, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 413, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 63, + 413, + 295, + 449 + ], + "type": "interline_equation", + "content": "\\begin{array}{c} X _ {t} = t X _ {1} + (1 - t) X _ {0} \\\\ X _ {t - \\Delta t} = (t - \\Delta t) X _ {1} + (1 - t + \\Delta t) X _ {0} \\end{array} , t \\in [ \\Delta t, 1 ]. \\tag {6}", + "image_path": "dcd948198f26d443aa9c751528a8e84dd18f34a8886657dea7ab3b7c2c669f72.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 449, + 296, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 509 + ], + "type": "text", + "content": "We use the stopgrad operator to calculate the HVT " + }, + { + "bbox": [ + 55, + 449, + 296, + 509 + ], + "type": "inline_equation", + "content": "v_{history} = \\text{stopgrad}(v_{\\theta}(X_{t - \\Delta t}, t - \\Delta t, 0))" + }, + { + "bbox": [ + 55, + 449, + 296, + 509 + ], + "type": "text", + "content": ". The neural velocity field " + }, + { + "bbox": [ + 55, + 449, + 296, + 509 + ], + "type": "inline_equation", + "content": "v_{\\theta}" + }, + { + "bbox": [ + 55, + 449, + 296, + 509 + ], + "type": "text", + "content": " takes the intermediate state " + }, + { + "bbox": [ + 55, + 449, + 296, + 509 + ], + "type": "inline_equation", + "content": "X_{t}" + }, + { + "bbox": [ + 55, + 449, + 296, + 509 + ], + "type": "text", + "content": ", time step " + }, + { + "bbox": [ + 55, + 449, + 296, + 509 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 449, + 296, + 509 + ], + "type": "text", + "content": ", and HVT as inputs, and learns a constant velocity field by fitting the reference velocity " + }, + { + "bbox": [ + 55, + 449, + 296, + 509 + ], + "type": "inline_equation", + "content": "v_{ref}" + }, + { + "bbox": [ + 55, + 449, + 296, + 509 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 64, + 514, + 295, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 514, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 64, + 514, + 295, + 529 + ], + "type": "interline_equation", + "content": "V C L (\\theta , \\phi) = \\mathbb {E} _ {t \\in p (t)} \\left[ d \\left(v _ {\\text {r e f}}, v _ {\\theta} \\left(X _ {t}, t, v _ {\\text {h i s t o r y}}\\right)\\right) \\right], \\tag {7}", + "image_path": "49cb0d26d41431f373ae0541c955bdaf297839fd2f56245be20037c13b014138.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 533, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 296, + 651 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 533, + 296, + 651 + ], + "type": "inline_equation", + "content": "d(\\cdot, \\cdot)" + }, + { + "bbox": [ + 55, + 533, + 296, + 651 + ], + "type": "text", + "content": " indicates distance metric. This process simultaneously optimizes the encoder " + }, + { + "bbox": [ + 55, + 533, + 296, + 651 + ], + "type": "inline_equation", + "content": "E_{\\phi}" + }, + { + "bbox": [ + 55, + 533, + 296, + 651 + ], + "type": "text", + "content": " to ensure that the reparameterized noises and the corresponding images satisfy the property of optimized couplings. Meanwhile, to prevent the encoder from overfitting during training, which could cause the reparameterized noises to rely too heavily on image information from datasets and thus constrain the diversity of generated images, we introduce KL divergence regularization to constrain the mean and variance of the encoder's output near the standard Gaussian distribution:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 93, + 656, + 295, + 680 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 656, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 93, + 656, + 295, + 680 + ], + "type": "interline_equation", + "content": "K L L (\\phi) = \\frac {1}{2} \\left(\\sigma^ {2} + \\mu^ {2} - 1 - \\log \\left(\\sigma^ {2}\\right)\\right). \\tag {8}", + "image_path": "00e069485b17058f1cc5d71c34e7a6764e7b0dae3a52ecb66f7a5f95eb6a0526.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 683, + 293, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 293, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 293, + 696 + ], + "type": "text", + "content": "Therefore, the total loss of our joint training framework is:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 100, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 100, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "L (\\theta , \\phi) = V C L (\\theta , \\phi) + \\alpha K L L (\\phi), \\tag {9}", + "image_path": "2a1dfd5983f7d0338dd332b0510808dafe170b905b258f8d31dc218789025272.jpg" + } + ] + } + ], + "index": 9 + }, + { + "type": "code", + "bbox": [ + 316, + 87, + 553, + 258 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 459, + 84 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 459, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 459, + 84 + ], + "type": "text", + "content": "Algorithm 2 Sampling of VRFNO." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 316, + 87, + 553, + 258 + ], + "lines": [ + { + "bbox": [ + 316, + 87, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 316, + 87, + 553, + 258 + ], + "type": "text", + "content": "Input: image in dataset " + }, + { + "bbox": [ + 316, + 87, + 553, + 258 + ], + "type": "inline_equation", + "content": "X_{1}" + }, + { + "bbox": [ + 316, + 87, + 553, + 258 + ], + "type": "text", + "content": " sampling steps " + }, + { + "bbox": [ + 316, + 87, + 553, + 258 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 316, + 87, + 553, + 258 + ], + "type": "text", + "content": " encoder " + }, + { + "bbox": [ + 316, + 87, + 553, + 258 + ], + "type": "inline_equation", + "content": "E_{\\phi}" + }, + { + "bbox": [ + 316, + 87, + 553, + 258 + ], + "type": "text", + "content": " velocity field " + }, + { + "bbox": [ + 316, + 87, + 553, + 258 + ], + "type": "inline_equation", + "content": "v_{\\theta}" + }, + { + "bbox": [ + 316, + 87, + 553, + 258 + ], + "type": "text", + "content": " \nOutput: generated image " + }, + { + "bbox": [ + 316, + 87, + 553, + 258 + ], + "type": "inline_equation", + "content": "\\tilde{X}_1" + }, + { + "bbox": [ + 316, + 87, + 553, + 258 + ], + "type": "text", + "content": " \nSample " + }, + { + "bbox": [ + 316, + 87, + 553, + 258 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim N(0,\\mathrm{I})" + }, + { + "bbox": [ + 316, + 87, + 553, + 258 + ], + "type": "inline_equation", + "content": "\\mu ,\\sigma^2 = E_\\phi (X_1)" + }, + { + "bbox": [ + 316, + 87, + 553, + 258 + ], + "type": "inline_equation", + "content": "X_0 = \\epsilon \\cdot \\sigma^2 +\\mu" + }, + { + "bbox": [ + 316, + 87, + 553, + 258 + ], + "type": "inline_equation", + "content": "\\Delta t = 1 / N" + }, + { + "bbox": [ + 316, + 87, + 553, + 258 + ], + "type": "inline_equation", + "content": "v_{history} = 0" + }, + { + "bbox": [ + 316, + 87, + 553, + 258 + ], + "type": "text", + "content": " \nfor " + }, + { + "bbox": [ + 316, + 87, + 553, + 258 + ], + "type": "inline_equation", + "content": "i = 0" + }, + { + "bbox": [ + 316, + 87, + 553, + 258 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 316, + 87, + 553, + 258 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 316, + 87, + 553, + 258 + ], + "type": "text", + "content": " do \n" + }, + { + "bbox": [ + 316, + 87, + 553, + 258 + ], + "type": "inline_equation", + "content": "v_{pred} = v_{\\theta}(X_{i\\Delta t},i\\Delta t,v_{history})" + }, + { + "bbox": [ + 316, + 87, + 553, + 258 + ], + "type": "inline_equation", + "content": "v_{history} = v_{pred}" + }, + { + "bbox": [ + 316, + 87, + 553, + 258 + ], + "type": "inline_equation", + "content": "X_{(i + 1)\\Delta t} = X_{i\\Delta t} + \\Delta t\\cdot v_{pred}" + }, + { + "bbox": [ + 316, + 87, + 553, + 258 + ], + "type": "text", + "content": " \nend for \n" + }, + { + "bbox": [ + 316, + 87, + 553, + 258 + ], + "type": "inline_equation", + "content": "\\tilde{X}_1 = X_N\\Delta t" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "code_body" + } + ], + "index": 11, + "sub_type": "algorithm" + }, + { + "bbox": [ + 313, + 279, + 555, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 555, + 388 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 279, + 555, + 388 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 279, + 555, + 388 + ], + "type": "text", + "content": " is a hyperparameter that controls the strength of regularization. By using the optimized couplings optimized through the encoder to train the neural velocity field, it can gradually learn how to infer nearly straight trajectories, thereby enabling few-step and single-step sampling. Furthermore, our encoder-neural velocity field joint generation framework still satisfies the marginal preserving property (Theorem 3), with a detailed definition and proof provided in the Appendix A." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 395, + 555, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 555, + 431 + ], + "type": "text", + "content": "Theorem 3. Assume " + }, + { + "bbox": [ + 313, + 395, + 555, + 431 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 395, + 555, + 431 + ], + "type": "text", + "content": " is rectifiable and " + }, + { + "bbox": [ + 313, + 395, + 555, + 431 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 395, + 555, + 431 + ], + "type": "text", + "content": " is its viscous rectified flow. The marginal law of " + }, + { + "bbox": [ + 313, + 395, + 555, + 431 + ], + "type": "inline_equation", + "content": "Z_{t}" + }, + { + "bbox": [ + 313, + 395, + 555, + 431 + ], + "type": "text", + "content": " equals that of " + }, + { + "bbox": [ + 313, + 395, + 555, + 431 + ], + "type": "inline_equation", + "content": "X_{t}" + }, + { + "bbox": [ + 313, + 395, + 555, + 431 + ], + "type": "text", + "content": " at every time " + }, + { + "bbox": [ + 313, + 395, + 555, + 431 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 395, + 555, + 431 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 395, + 555, + 431 + ], + "type": "inline_equation", + "content": "\\text{Law}(Z_t) = \\text{Law}(X_t), \\forall t \\in [0,1]" + }, + { + "bbox": [ + 313, + 395, + 555, + 431 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 438, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 555, + 617 + ], + "type": "text", + "content": "Noise optimization. In previous studies, noise optimization typically treats noise as a learnable parameter and performs small-scale iterative updates over multiple iterations using gradient information provided by the model, until it converges to an approximate optimal solution [6, 35]. While it can generate high-quality images, its limitations are also evident: it requires personalized optimization for each noise and relies on pre-trained models. In contrast, our method utilizes the reparameterization technique to optimize the noise through linear transformations directly. This approach allows for significant adjustments to the noise in a single optimization step, enabling rapid and efficient approximation to the optimal solution, significantly reducing the number of iterations and computational overhead, while also eliminating the dependence on pre-trained models." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "text", + "content": "Two-stage training. In the training process, we adopt a two-stage strategy based on empirical observations. In the first stage, we set " + }, + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "inline_equation", + "content": "d(\\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "text", + "content": " to the MSE loss, and once convergence is achieved, we proceed to the second stage. In this stage, we incorporate the LPIPS loss [32] for joint training on top of the first-stage setup, continuing until convergence is reached again. Notably, our method does not rely on distillation or adversarial training mechanisms." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15009" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 72, + 116, + 177 + ], + "blocks": [ + { + "bbox": [ + 58, + 72, + 116, + 177 + ], + "lines": [ + { + "bbox": [ + 58, + 72, + 116, + 177 + ], + "spans": [ + { + "bbox": [ + 58, + 72, + 116, + 177 + ], + "type": "image", + "image_path": "8ca191f01a1a07686a891e55b467551102bc71a197e279f871cf4e064a76fe2d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 179, + 295, + 224 + ], + "lines": [ + { + "bbox": [ + 55, + 179, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 179, + 295, + 224 + ], + "type": "text", + "content": "Figure 4. Visualization of inference trajectories on synthetic data. The first column on the left shows the inference trajectories of 1-RF with " + }, + { + "bbox": [ + 55, + 179, + 295, + 224 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 179, + 295, + 224 + ], + "type": "text", + "content": " steps, and the three columns on the right show the inference trajectories of VRFNO with " + }, + { + "bbox": [ + 55, + 179, + 295, + 224 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 179, + 295, + 224 + ], + "type": "text", + "content": " steps." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 119, + 79, + 176, + 176 + ], + "blocks": [ + { + "bbox": [ + 119, + 79, + 176, + 176 + ], + "lines": [ + { + "bbox": [ + 119, + 79, + 176, + 176 + ], + "spans": [ + { + "bbox": [ + 119, + 79, + 176, + 176 + ], + "type": "image", + "image_path": "53074f9c9151eb4d260a6afa1f58bfdc0f055c7c4a947b38776283878a876a04.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 179, + 80, + 234, + 176 + ], + "blocks": [ + { + "bbox": [ + 179, + 80, + 234, + 176 + ], + "lines": [ + { + "bbox": [ + 179, + 80, + 234, + 176 + ], + "spans": [ + { + "bbox": [ + 179, + 80, + 234, + 176 + ], + "type": "image", + "image_path": "a43cdd7a4976f71a1c40ee88633529d2a5642fb386730ff332c8d22c3ba00fa9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 238, + 80, + 294, + 176 + ], + "blocks": [ + { + "bbox": [ + 238, + 80, + 294, + 176 + ], + "lines": [ + { + "bbox": [ + 238, + 80, + 294, + 176 + ], + "spans": [ + { + "bbox": [ + 238, + 80, + 294, + 176 + ], + "type": "image", + "image_path": "82c941c05efd8db3cf68e03b1d98013f6ee4dc5efb85aba578749eb8225a5e05.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 247, + 296, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 296, + 379 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 296, + 379 + ], + "type": "text", + "content": "Sampling. After training the encoder and the neural velocity field, we can generate images using VRF ODE introduced in Eq.(4), the discrete sampling process is shown in Algorithm 2. For single-step generation, the HVT is set to 0 during sampling, while in few-step generation, the HVT is updated at each step. Additionally, we integrate the dataset into the sampler, meaning that the sampled noise is also reparameterized during the generation process. Since we introduced perturbations in the encoder, this will not affect the diversity of the generated images but can further enhance the quality of the generated images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 392, + 132, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 132, + 406 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 132, + 406 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 413, + 296, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 296, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 296, + 556 + ], + "type": "text", + "content": "We evaluate the performance of VRFNO in various scenarios, including synthetic data and real data. In Section 4.1, we compare the performance of RF and VRFNO by studying the inference trajectories between two 2D Gaussian distributions, which clearly demonstrates the effectiveness of our method. In Section 4.2, we extend the experiments to real image datasets, CIFAR-10 [12] and AFHQ [3], at different resolutions to validate the performance of VRFNO in one-step and few-step generation. Additionally, we conducted further analysis on VRFNO, including ablation studies, as well as quantitative evaluations of its trajectory straightness and inference efficiency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": "Baselines and evaluation. We evaluated VRFNO and compared it with state-of-the-art diffusion models [10, 18, 24, 27, 28] and rectified flow models (RF [14], Constant Acceleration Flow (CAF) [21], TraFlow [29], Shortcut Model (SM) [7]) in single-step and few-step generation. Among them, RF serves as the baseline for our method, while CAF is the most similar approach to ours. Therefore, these two methods are the primary focus of our comparison. We use Fréchet Inception Distance (FID) and Kernel Inception Distance (KID) to assess the quality of generated images, and use Inception Score (IS) to measure diversity. Lower FID and KID indicate higher generation quality, while a higher IS suggests greater diversity in the generated images." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 316, + 83, + 573, + 399 + ], + "blocks": [ + { + "bbox": [ + 355, + 71, + 512, + 81 + ], + "lines": [ + { + "bbox": [ + 355, + 71, + 512, + 81 + ], + "spans": [ + { + "bbox": [ + 355, + 71, + 512, + 81 + ], + "type": "text", + "content": "Table 1. Quantitative results on CIFAR-10." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 83, + 573, + 399 + ], + "lines": [ + { + "bbox": [ + 316, + 83, + 573, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 83, + 573, + 399 + ], + "type": "table", + "html": "
MethodsNFEIS (↑)FID (↓)KID(×10-3)(↓)
Diffusion/Consistency Models
VP ODE (+distill)11.20 (8.73)451 16.23)-
sub-VP ODE (+distill)11.21 (8.80)451 (14.32)-
DDIM distillation18.369.36-
Progressive1-9.12-
CT18.498.71-
EDM5-37.75-
Rectified Flow Models
1-RF11.13379428
2-RF18.1511.978.66
CAF18.324.81-
TraFlow1-4.50-
SM1-4.93-
VRFNO (Ours)19.594.502.73
1-RF57.1234.8132.26
2-RF59.014.362.25
CAF58.674.03-
VRFNO (Ours)59.424.032.13
1-RF108.4412.7011.50
2-RF109.133.831.63
CAF109.123.77-
VRFNO (Ours)109.513.361.31
", + "image_path": "2846db380ce81c7667787afa019a3f124d906ca890b843f607c9aeb55a1f2a81.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 420, + 443, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 443, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 443, + 433 + ], + "type": "text", + "content": "4.1. Synthetic Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 437, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 533 + ], + "type": "text", + "content": "We demonstrate the trajectory straightening effect of VRFNO through 2D synthetic data experiments. For the neural velocity field, we use a simple multi-layer perceptron (MLP) architecture with three hidden layers, each containing 64 units. For the encoder, we adopt the encoder architecture from the VAE framework, consisting of two hidden layers, each with 16 units. Additionally, a fully connected layer is used to produce the mean and variance separately." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "content": "The visualization results are shown in Fig.4, where we compared VRFNO with the 1st generation RF (1-RF). Additionally, in Appendix B, we provide the inference trajectories of 1-RF and the 2nd generation RF (2-RF) for different numbers of steps. In the multi-step generation, the inference trajectory of VRFNO is closer to a straight line compared to 1-RF, which provides a guarantee for accurate single-step and few-step generation. The single-step and few-step inference trajectories of VRFNO are visually straight, indicating that the neural velocity field of VRFNO better converges to a constant velocity field, making its predictions of the trajectory direction more stable. In addition, the error between the endpoints of the single-step and the few-step inference trajectories and the endpoints of the multi-step inference trajectories is small, indicating that VRFNO is able" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15010" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 294, + 308 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 308 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 308 + ], + "type": "image", + "image_path": "82c338e1a24126ac5881cc66c4dc65cfcff261d85b7a6d1a6907fe6f2c75fe7e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 309, + 295, + 331 + ], + "lines": [ + { + "bbox": [ + 55, + 309, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 295, + 331 + ], + "type": "text", + "content": "Figure 5. Qualitative results on CIFAR-10. Visualization of one-step generation of VRFNO" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 353, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 353, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 353, + 296, + 437 + ], + "type": "text", + "content": "to accurately map samples from the initial distribution to the target distribution in both single-step and few-step generation. In contrast, when 1-RF is generated in a single step, the endpoint of the trajectory is usually concentrated at the mean point of the target distribution, while the samples near this point are usually meaningless in practical applications, as shown in Fig.II (a) in Appendix B." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 445, + 187, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 187, + 457 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 187, + 457 + ], + "type": "text", + "content": "4.2. Real-data Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 462, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 605 + ], + "type": "text", + "content": "To further validate the effectiveness of our method, we evaluate VRFNO on real image datasets, namely CIFAR-10 with a resolution of " + }, + { + "bbox": [ + 55, + 462, + 295, + 605 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 55, + 462, + 295, + 605 + ], + "type": "text", + "content": " and AFHQ with different resolutions (including " + }, + { + "bbox": [ + 55, + 462, + 295, + 605 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 55, + 462, + 295, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 462, + 295, + 605 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 55, + 462, + 295, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 462, + 295, + 605 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 462, + 295, + 605 + ], + "type": "text", + "content": "). For the neural velocity field, we adopt the UNet architecture from DDPM++ [27] to train on CIFAR-10 and the NCSNv2 [26] to train on AFHQ; for the encoder, we refer to the encoder architecture from [8], and its details can be found in Appendix E. Specifically, the encoder requires very few parameters, taking the train of CIFAR-10 as an example, the number of parameters in the encoder is less than 1/20 of those in the neural velocity field." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "Tab.1 presents the quantitative evaluation of VRFNO on CIFAR-10. Compared to other diffusion models, our approach achieves superior performance in single-step generation, with FID = 4.53, KID = 2.73, and IS = 10.59. Compared to previous rectified flow models, VRFNO attains state-of-the-art performance in both single-step and few-step generation (our training does not involve distillation). Fig.5 illustrates the qualitative evaluation of VRFNO in single-step generation on CIFAR-10. Further compara" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 553, + 308 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 553, + 308 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 553, + 308 + ], + "type": "image", + "image_path": "1c73188d9b91c1933f51e52ec6c4598cf24418fd24f753ef45955b1aa909bcf3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 309, + 555, + 331 + ], + "lines": [ + { + "bbox": [ + 314, + 309, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 314, + 309, + 555, + 331 + ], + "type": "text", + "content": "Figure 6. Qualitative results on AFHQ. Visualization of one-step generation of VRFNO at different resolutions" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 354, + 565, + 514 + ], + "blocks": [ + { + "bbox": [ + 315, + 342, + 552, + 354 + ], + "lines": [ + { + "bbox": [ + 315, + 342, + 552, + 354 + ], + "spans": [ + { + "bbox": [ + 315, + 342, + 552, + 354 + ], + "type": "text", + "content": "Table 2. Quantitative results on AFHQ with different resolutions." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 354, + 565, + 514 + ], + "lines": [ + { + "bbox": [ + 315, + 354, + 565, + 514 + ], + "spans": [ + { + "bbox": [ + 315, + 354, + 565, + 514 + ], + "type": "table", + "html": "
MethodsDatasetsNFEResolutionFID(↓)
2-RF (+distill)AFHQ -CAT164 × 64181.93(31.57)
128 × 128172.66(29.81)
256 × 256171.84(29.33)
AFHQ -DOG164 × 64200.77(33.64)
128 × 128192.30(32.70)
256 × 256189.82(30.27)
VRFNO (Ours)AFHQ -CAT164 × 6428.69
128 × 12827.56
256 × 25627.04
AFHQ -DOG164 × 6444.64
128 × 12827.21
256 × 25627.37
", + "image_path": "175605a0a1121589df8c9dbaa667de7de4df62ff4e87dc3c089e1a0d02a3051a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 533, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 533, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 312, + 533, + 555, + 665 + ], + "type": "text", + "content": "tive qualitative evaluations between VRFNO and RF in N-step generation can be found in Appendix B. Tab.2 provides the quantitative evaluation of VRFNO's generated images on the AFHQ cat and dog datasets, focusing on the image quality of single-step generation at different resolutions. Our method is designed as a better alternative to Reflow, so the comparison primarily highlights the performance difference with 2-RF. From the tab.2, it is evident that VRFNO outperforms 2-RF in image generation quality across different resolutions. The qualitative evaluation of AFHQ at different resolutions can be found in Fig.6." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 672, + 417, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 417, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 417, + 685 + ], + "type": "text", + "content": "4.3. Further Analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Ablation Study. We conduct an ablation study to evaluate the effectiveness of each component in our framework" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15011" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 83, + 279, + 254 + ], + "blocks": [ + { + "bbox": [ + 73, + 71, + 277, + 82 + ], + "lines": [ + { + "bbox": [ + 73, + 71, + 277, + 82 + ], + "spans": [ + { + "bbox": [ + 73, + 71, + 277, + 82 + ], + "type": "text", + "content": "Table 3. Ablation study of HVT and noise optimization." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 73, + 83, + 279, + 254 + ], + "lines": [ + { + "bbox": [ + 73, + 83, + 279, + 254 + ], + "spans": [ + { + "bbox": [ + 73, + 83, + 279, + 254 + ], + "type": "table", + "html": "
ConfigNFEHVTNoise optimizationFID(↓)
A1XX379
B1✓X332
C1X✓4.72
D1✓✓4.53
E5XX34.81
F5✓X32.50
G5X✓4.28
H5✓✓4.03
I10XX12.70
J10✓X9.34
K10X✓4.75
L10✓✓3.40
", + "image_path": "db5f4cecae70497432be3371c2cdd5924934ecb2ad22fc54d9a40b95d05849ca.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 61, + 277, + 291, + 318 + ], + "blocks": [ + { + "bbox": [ + 99, + 265, + 251, + 276 + ], + "lines": [ + { + "bbox": [ + 99, + 265, + 251, + 276 + ], + "spans": [ + { + "bbox": [ + 99, + 265, + 251, + 276 + ], + "type": "text", + "content": "Table 4. Comparison of flow straightness." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 277, + 291, + 318 + ], + "lines": [ + { + "bbox": [ + 61, + 277, + 291, + 318 + ], + "spans": [ + { + "bbox": [ + 61, + 277, + 291, + 318 + ], + "type": "table", + "html": "
Dataset2-RF3-RFCAFVRFNO(Ours)
2D0.0670.0530.0580.054
CIFAR-100.0580.0560.0350.026
", + "image_path": "95ac0d56f11b2d87ef419987bc8f2d5ee8a728dd07ba8020da10f05959d8c5e0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 89, + 341, + 263, + 380 + ], + "blocks": [ + { + "bbox": [ + 68, + 328, + 282, + 340 + ], + "lines": [ + { + "bbox": [ + 68, + 328, + 282, + 340 + ], + "spans": [ + { + "bbox": [ + 68, + 328, + 282, + 340 + ], + "type": "text", + "content": "Table 5. Comparison of the inference time (sec) in " + }, + { + "bbox": [ + 68, + 328, + 282, + 340 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 68, + 328, + 282, + 340 + ], + "type": "text", + "content": " -step." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 89, + 341, + 263, + 380 + ], + "lines": [ + { + "bbox": [ + 89, + 341, + 263, + 380 + ], + "spans": [ + { + "bbox": [ + 89, + 341, + 263, + 380 + ], + "type": "table", + "html": "
NFERFCAFVRFNO(Ours)
10.1720.1810.305
101.4041.4151.646
", + "image_path": "40ba63157b8698125892afc44468c2e2b0714b11492b8c9d9f58a13260232c61.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 401, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 295, + 544 + ], + "type": "text", + "content": "for single-step and few-step generation. Specifically, we investigate the improvements brought to RF by (1) the introduction of the HVT and (2) the noise optimization operation. The configurations and results are shown in Tab.3. The two components have both contributed to the improvement of model performance. Regarding the introduction of the HVT, even with one-step sampling, the model performance improved, indicating that it facilitates trajectory rectification during the training process. As for the noise optimization operation, it results in a significant performance boost, demonstrating the effectiveness of training the neural velocity field with our defined optimized coupling." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 545, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 640 + ], + "type": "text", + "content": "Flow Straightness. To evaluate the straightness of the inference trajectories, we introduce the Normalized Flow Straightness Score (NFSS) [14, 15]. A smaller NFSS indicates that the inference trajectory is closer to a straight line. We compare VRFNO with RF and CAF on synthetic datasets and CIFAR-10, with the experimental results shown in Tab.4. Our method generates inference trajectories that are straighter compared to the other methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": "Time Efficiency. We also compare the inference time of our method with RF and CAF. Tab.5 presents the time required to generate 512 images in both one-step and ten-step settings. Our method incurs a slightly higher time cost than RF due to the additional encoder. In contrast, CAF consists of two neural velocity fields, it necessitates an additional" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "text", + "content": "computation (its total NFE equals " + }, + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "inline_equation", + "content": "N + 1" + }, + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "text", + "content": "), resulting in significantly higher time consumption." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 109, + 401, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 401, + 121 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 401, + 121 + ], + "type": "text", + "content": "5. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 129, + 555, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 129, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 129, + 555, + 321 + ], + "type": "text", + "content": "The inference process of DMs can be viewed as an iterative solution of ODE [25] or stochastic differential equations [2, 9], and ODE-based methods [5, 31] are more effective in few-step sampling due to their determinism. A widely used first-order ODE solver is the Euler solver, but its large local truncation error necessitates more iteration steps for high-quality image generation. To reduce iteration steps, higher order ODE solvers such as DEIS [31] and DPM solvers [16] exploit the semilinear structure of diffusion ODEs, deriving exact solutions for high-quality image generation with fewer steps. DPM-solver++ [17] enhances DPM-solver by improving stability in higher-order solvers. UniPC [33] builds a unified predictor-corrector framework for DPM solver, enhancing the quality of sampling in a few steps. DC-solver [34] uses dynamic compensation to correct misalignment in UniPC, thus improving image quality." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 322, + 556, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 556, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 556, + 514 + ], + "type": "text", + "content": "Another mainstream method for accelerated sampling is distillation. Progressive distillation [24] iteratively halves the steps of a DDIM until it enables image generation in just 4 steps. Diff-Instruct [20] accelerates sampling by minimizing Integra Kullback-Leibler divergence to merge knowledge from multiple time steps of a pre-trained model. Consistency Model (CM) [28] enables direct mapping from noise to data by minimizing the differences between the final states when adjacent ODE trajectory points are used as inputs. Latent CM [19] treats the guided reverse diffusion process as solving an augmented PF ODE, directly predicting the ODE solution in the latent space, enabling fast and high-fidelity sampling. Consistency Trajectory Model [11] further optimizes CM by allowing the model to learn mappings between arbitrary initial and final times along the ODE trajectory during the diffusion process." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "content": "This paper proposes a novel method VRFNO, which is a joint training framework combining an encoder (for noise optimization) and a neural velocity field (for velocity prediction). VRFNO utilizes the encoder to optimize noises to transform randomly matched noises and images into optimized couplings, which are then used to train the neural velocity field for precise prediction of straight inference trajectories. To further straighten the inference trajectories during training, we introduce the HVT into the neural velocity field to enhance its prediction accuracy. Extensive experiments conducted on both synthetic and real world datasets demonstrate the effectiveness and scalability of VRFNO, and the results show that it achieves state-of-the-art performance across multiple datasets and varying resolutions." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15012" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 124 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 124 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 124 + ], + "type": "text", + "content": "This work was supported in part by the National Science Fund of China (No.62276135, U24A20330 and 62361166670)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 133, + 115, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 133, + 115, + 146 + ], + "spans": [ + { + "bbox": [ + 56, + 133, + 115, + 146 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 153, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 153, + 296, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 153, + 296, + 185 + ], + "spans": [ + { + "bbox": [ + 61, + 153, + 296, + 185 + ], + "type": "text", + "content": "[1] Jonas Adler and Sebastian Lunz. Banach wasserstein gan. Advances in neural information processing systems, 31, 2018. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 186, + 296, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 186, + 296, + 229 + ], + "spans": [ + { + "bbox": [ + 61, + 186, + 296, + 229 + ], + "type": "text", + "content": "[2] Fan Bao, Chongxuan Li, Jun Zhu, and Bo Zhang. Analytic-dpm: an analytic estimate of the optimal reverse variance in diffusion probabilistic models. arXiv preprint arXiv:2201.06503, 2022.8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 230, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 230, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 61, + 230, + 296, + 274 + ], + "type": "text", + "content": "[3] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8188-8197, 2020. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 274, + 296, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 274, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 61, + 274, + 296, + 339 + ], + "type": "text", + "content": "[4] Tianyi Chu, Wei Xing, Jiafu Chen, Zhizhong Wang, Jiakai Sun, Lei Zhao, Haibo Chen, and Huaizhong Lin. Attack deterministic conditional image generative models for diverse and controllable generation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1362-1370, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 340, + 296, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 340, + 296, + 383 + ], + "spans": [ + { + "bbox": [ + 62, + 340, + 296, + 383 + ], + "type": "text", + "content": "[5] Tim Dockhorn, Arash Vahdat, and Karsten Kreis. Genie: Higher-order denoising diffusion solvers. Advances in Neural Information Processing Systems, 35:30150-30166, 2022. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 384, + 296, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 384, + 296, + 438 + ], + "spans": [ + { + "bbox": [ + 62, + 384, + 296, + 438 + ], + "type": "text", + "content": "[6] Luca Eyring, Shyamgopal Karthik, Karsten Roth, Alexey Dosovitskiy, and Zeynep Akata. Reno: Enhancing one-step text-to-image models through reward-based noise optimization. Advances in Neural Information Processing Systems, 37:125487-125519, 2025. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 439, + 296, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 439, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 62, + 439, + 296, + 472 + ], + "type": "text", + "content": "[7] Kevin Frans, Danijar Hafner, Sergey Levine, and Pieter Abbeel. One step diffusion via shortcut models. arXiv preprint arXiv:2410.12557, 2024.6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 472, + 296, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 472, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 62, + 472, + 296, + 525 + ], + "type": "text", + "content": "[8] Irina Higgins, Loic Matthew, Arka Pal, Christopher P Burgess, Xavier Glorot, Matthew M Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-vae: Learning basic visual concepts with a constrained variational framework. ICLR (Poster), 3, 2017. 5, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 526, + 296, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 526, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 62, + 526, + 296, + 559 + ], + "type": "text", + "content": "[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 560, + 296, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 560, + 296, + 603 + ], + "spans": [ + { + "bbox": [ + 57, + 560, + 296, + 603 + ], + "type": "text", + "content": "[10] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35:26565-26577, 2022. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 604, + 296, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 604, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 57, + 604, + 296, + 658 + ], + "type": "text", + "content": "[11] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ode trajectory of diffusion. arXiv preprint arXiv:2310.02279, 2023. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 658, + 296, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 680 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 680 + ], + "type": "text", + "content": "[12] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 681, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 296, + 712 + ], + "type": "text", + "content": "[13] Qiang Liu. Rectified flow: A marginal preserving approach to optimal transport. arXiv preprint arXiv:2209.14577, 2022. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[14] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2023. 1, 3, 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 118, + 554, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 162 + ], + "type": "text", + "content": "[15] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In The Twelfth International Conference on Learning Representations, 2024. 1, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 163, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 554, + 217 + ], + "type": "text", + "content": "[16] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 219, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 554, + 262 + ], + "type": "text", + "content": "[17] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 263, + 554, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 554, + 297 + ], + "type": "text", + "content": "[18] Eric Luhman and Troy Luhman. Knowledge distillation in iterative generative models for improved sampling speed. arXiv preprint arXiv:2101.02388, 2021. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 298, + 554, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 554, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 554, + 341 + ], + "type": "text", + "content": "[19] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 342, + 554, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 554, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 554, + 397 + ], + "type": "text", + "content": "[20] Weijian Luo, Tianyang Hu, Shifeng Zhang, Jiacheng Sun, Zhenguo Li, and Zhihua Zhang. Diff-instruct: A universal approach for transferring knowledge from pre-trained diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 399, + 554, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 554, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 554, + 432 + ], + "type": "text", + "content": "[21] Dogyun Park, Sojin Lee, Sihyeon Kim, Taehoon Lee, Youngjoon Hong, and Hyunwoo J Kim. Constant acceleration flow. arXiv preprint arXiv:2411.00322, 2024. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 433, + 554, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 554, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 554, + 476 + ], + "type": "text", + "content": "[22] Saptarshi Roy, Vansh Bansal, Purnamrita Sarkar, and Alessandro Rinaldo. 2-rectifications are enough for straight flows: A theoretical insight into wasserstein convergence. arXiv e-prints, pages arXiv-2410, 2024. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 478, + 554, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 554, + 531 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 554, + 531 + ], + "type": "text", + "content": "[23] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 conference proceedings, pages 1-10, 2022. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 533, + 554, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 554, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 554, + 566 + ], + "type": "text", + "content": "[24] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 6, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 567, + 554, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 554, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 554, + 600 + ], + "type": "text", + "content": "[25] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 601, + 554, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 554, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 554, + 634 + ], + "type": "text", + "content": "[26] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. Advances in neural information processing systems, 33:12438-12448, 2020. 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 635, + 554, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 554, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 554, + 679 + ], + "type": "text", + "content": "[27] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 6, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "text", + "content": "[28] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023. 6, 8" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15013" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 295, + 374 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "text", + "content": "[29] Zhangkai Wu, Xuhui Fan, Hongyu Wu, and Longbing Cao. Traflow: Trajectory distillation on pre-trained rectified flow, 2025. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 107, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 107, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 55, + 107, + 295, + 150 + ], + "type": "text", + "content": "[30] YoungJoon Yoo and Jongwon Choi. Topic-vq-vae: Leveraging latent codebooks for flexible topic-guided document generation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 19422–19430, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 152, + 295, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 152, + 295, + 184 + ], + "spans": [ + { + "bbox": [ + 57, + 152, + 295, + 184 + ], + "type": "text", + "content": "[31] Qinsheng Zhang and Yongxin Chen. Fast sampling of diffusion models with exponential integrator. arXiv preprint arXiv:2204.13902, 2022. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 186, + 295, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 186, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 57, + 186, + 295, + 240 + ], + "type": "text", + "content": "[32] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 241, + 295, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 241, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 57, + 241, + 295, + 285 + ], + "type": "text", + "content": "[33] Wenliang Zhao, Lujia Bai, Yongming Rao, Jie Zhou, and Jiwen Lu. Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 286, + 294, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 286, + 294, + 328 + ], + "spans": [ + { + "bbox": [ + 57, + 286, + 294, + 328 + ], + "type": "text", + "content": "[34] Wenliang Zhao, Haolin Wang, Jie Zhou, and Jiwen Lu. Dcsolver: Improving predictor-corrector diffusion sampler via dynamic compensation. arXiv preprint arXiv:2409.03755, 2024.8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 331, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 331, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 57, + 331, + 294, + 374 + ], + "type": "text", + "content": "[35]Zikai Zhou,Shitong Shao,Lichen Bai,Zhiqiang Xu,Bo Han,and Zeke Xie. Golden noise for diffusion models: A learning framework.arXiv preprint arXiv:2411.09502, 2024.5" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15014" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/StrandHead_ Text to Hair-Disentangled 3D Head Avatars Using Human-Centric Priors/5c673a06-82fc-4d7f-915c-54daefec29d8_content_list.json b/2025/StrandHead_ Text to Hair-Disentangled 3D Head Avatars Using Human-Centric Priors/5c673a06-82fc-4d7f-915c-54daefec29d8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..09c6a1290cfe3a372e7131b477b37ffc506d7227 --- /dev/null +++ b/2025/StrandHead_ Text to Hair-Disentangled 3D Head Avatars Using Human-Centric Priors/5c673a06-82fc-4d7f-915c-54daefec29d8_content_list.json @@ -0,0 +1,2149 @@ +[ + { + "type": "text", + "text": "StrandHead: Text to Hair-Disentangled 3D Head Avatars Using Human-Centric Priors", + "text_level": 1, + "bbox": [ + 209, + 130, + 789, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaokun Sun, Zeyu Cai, Ying Tai, Jian Yang, Zhenyu Zhang* Nanjing University *Corresponding Author", + "bbox": [ + 259, + 203, + 743, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "xiaokun.sun@smail.nju.edu.cn, caizeyu010612@gmail.com, {yingtai, csjyang}@nju.edu.cn, zhangjesse@foxmail.com", + "bbox": [ + 259, + 241, + 733, + 273 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cc9a9c1cbcd3b6e4a09d1752acb3c3373d9480d276a6e7d6590e8282c25f2f96.jpg", + "image_caption": [ + "(a) Input Prompts" + ], + "image_footnote": [], + "bbox": [ + 158, + 273, + 834, + 450 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d6828f5fab38e6849ba9abba90adf049d66e30e46521a378a46c183696929be5.jpg", + "image_caption": [ + "Figure 1. Given (a) input prompts, StrandHead generates (b) realistic 3D head avatars featuring strand-level attributes and (c) 3D hair strands by utilizing human-specific 2D generative priors and 3D hair strand geometric priors. By precisely capturing the internal geometry of hair strands, our approach enables seamless and flexible (d) hairstyle transfer and (e) editing, as well as (f) physics-based simulation." + ], + "image_footnote": [], + "bbox": [ + 308, + 452, + 831, + 539 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 593, + 326, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While haircut indicates distinct personality, existing avatar generation methods fail to model practical hair due to the data limitation or entangled representation. We propose StrandHead, a novel text-driven method capable of generating 3D hair strands and disentangled head avatars with strand-level attributes. Instead of using large-scale hair-text paired data for supervision, we demonstrate that realistic hair strands can be generated from prompts by distilling 2D generative models pre-trained on human mesh data. To this end, we propose a meshing approach guided by strand geometry to guarantee the gradient flow from the distillation objective to the neural strand representation. The optimization is then regularized by statistically significant haircut features, leading to stable updating of strands against unreasonable drifting. These employed 2D/3D human-centric priors contribute to text-aligned and realistic 3D strand generation. Extensive experiments show that StrandHead achieves the state-of-the-art performance", + "bbox": [ + 89, + 628, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "on text to strand generation and disentangled 3D head avatar modeling. The generated 3D hair can be applied on avatars for strand-level editing, as well as implemented in the graphics engine for physical simulation or other applications. Project page: https://xiaokunsun.github.io/StrandHead.github.io/.", + "bbox": [ + 511, + 595, + 908, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 694, + 645, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Creating 3D head avatars is crucial for many applications including digital telepresence, gaming, movies, and AR/VR. Hairstyles reflecting personal characteristics greatly impact the fidelity and realism of digital humans. Traditional methods rely heavily on manual effort, making them time-consuming and labor-intensive. Recent advances [52, 75, 79, 88] achieve automatic 3D head generation with the research paradigm based on supervised learning. However, these methods require costly 3D training data, which significantly limits their generalization capabilities and restricts potential applications.", + "bbox": [ + 509, + 719, + 906, + 885 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the rapid development of text-driven generation", + "bbox": [ + 532, + 885, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13393", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "methods [47, 54], creating 3D head avatars from given prompts without requiring 3D-text datasets becomes possible. Recent studies [1, 12, 84, 94] unlock the potential of 2D diffusion models in modeling 3D head avatars by integrating them with 3D head priors [4, 26, 77]. HumanNorm [17] further improves domain-specific fidelity by fine-tuning diffusion models on high-quality 3D human meshes. While focusing on facial geometry and texture modeling, these methods utilize holistic meshes or NeRF [38] to represent 3D haircuts, failing to capture the internal geometric structure of hair strands, i.e., 3D curves. This limitation not only significantly reduces the realism of the generated avatars but also makes them incompatible with strand-based applications and simulation systems [2, 9, 10].", + "bbox": [ + 89, + 90, + 480, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To accurately model 3D haircuts, recent efforts achieve strand-level hairstyle reconstruction [55, 64, 89] or generation [13, 92] using VAEs or parametric models. Despite their impressive performance, these methods require constrained multi-view images or manual latent space searching, preventing them from freely creating 3D hair based on user-friendly text. The most recent work, HAAR [65] pioneers text-to-strand by training a text-conditioned hair map diffusion model. However, due to its reliance on large-scale and costly paired data including 9825 haircuts and descriptions, HAAR faces challenges in generating diverse hairstyles outside the training set. Additionally, HAAR overlooks hair texture and geometry that adapts to specific head shapes, further limiting its practical applications.", + "bbox": [ + 89, + 305, + 482, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "So can we use powerful human-specific 2D generative priors to create realistic and diverse strand-based hair from text? We address this meaningful and challenging problem by developing StrandHead, a novel framework that generates high-fidelity 3D head avatars with strand-accurate hair from prompts. As illustrated in Fig. 1, the generated head avatars feature diverse and realistic hair strands, enabling seamless and flexible strand-level transfer, editing, and physics-based simulation. Instead of using large-scale hair-prompt paired data, we achieve text-aligned 3D hair strand generation by effectively utilizing human-specific 2D generative priors and 3D hair geometric priors.", + "bbox": [ + 89, + 520, + 482, + 702 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we propose a novel differentiable prismatization algorithm inspired by the cylindrical structure of hair strands. This algorithm can efficiently and differentiably convert strands into watertight prismatic meshes, which enables smooth backpropagation of gradients from 2D diffusion models pre-trained on human mesh data to 3D hair strands using mesh-based differentiable renderers, thereby making it possible to use 2D generative priors for modeling 3D hairstyles. Then we propose two effective losses inspired by the distribution patterns of 3D hair geometric features to further regularize the hair shape. These rich 2D/3D human-centric priors work together to achieve text-aligned and realistic 3D hair strand generation. Extensive experi", + "bbox": [ + 89, + 704, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ments demonstrate that StrandHead outperforms the state-of-the-art (SOTA) methods in both head and hair generation tasks, and supports flexible haircut transfer and editing, as well as physical-based rendering and simulation.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are summarized as follows:", + "bbox": [ + 532, + 152, + 872, + 165 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose StrandHead, a novel framework for generating realistic 3D head avatars with strand-level attributes. To the best of our knowledge, StrandHead is the first work to generate 3D hair strands by distilling human-specific 2D diffusion models.", + "- We propose a differentiable prismatization algorithm that converts hair strands into watertight prismatic meshes. This ensures stable gradient flow from 2D generative priors to 3D hair strand representation, thereby achieving reliable end-to-end strand-based hair optimization.", + "- Inspired by statistical 3D hair geometric features, we introduce two simple but solid regularization losses to supervise both local and global hair shapes, enabling reasonable and realistic hairstyle generation." + ], + "bbox": [ + 511, + 166, + 903, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 392, + 653, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-3D General Object Generation. Inspired by the success of text-to-image (T2I) generation [42, 50, 51, 54, 56, 86], many studies have explored using pre-trained vision-language models [49, 54] to achieve text-guided 3D content generation without large-scale 3D-text paired data. Early methods [18, 37, 39, 58, 69] employ the CLIP model [49] to supervise the alignment of 3D representations with prompts. DreamFusion [47] introduces the Score Distillation Sampling (SDS) loss, significantly enhancing the fidelity of generated 3D content by leveraging more powerful pre-trained diffusion models [54]. Subsequent works further advance text-driven 3D generation by improving 3D representations [5, 6, 80], optimization strategies [29, 68, 81], SDS loss [28, 71], and diffusion models [27, 48, 62]. Despite these advancements, current approaches focused on general content generation do not fully leverage the extensive prior knowledge of human heads and hair, limiting their ability to generate realistic, high-quality 3D head avatars with strand-based hair.", + "bbox": [ + 511, + 416, + 903, + 702 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-3D Head Avatar Generation. CLIPFace [1], T2P [87] and Describe3D [74] pioneer zero-shot text-driven 3D head generation by combining CLIP [49] with 3D parametric head models [4, 8, 26, 77]. With the introduction of SDS loss [47], DreamFace [85] and FaceG2E [76] leverage pre-trained T2I diffusion models to greatly improve generation quality within specific domains. HeadEvolver [70] enhances the expressiveness of head mesh deformations by introducing vector fields, while HeadSculpt [12], and HeadArtist [30] employ DMTet [60] instead of traditional meshes to capture geometric details. HeadStudio [94] incorporates 3DGS [20] to produce realistic and animatable 3D heads. HumanNorm [17], further", + "bbox": [ + 511, + 704, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13394", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/cca5e3cb3c15a0fbd9f5d0a3cf79b2163df40126bddf5474c3ba3cc7278996f8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TaskMethodHead-Hair-DecoupledStrand-Based HairGeometry & TextureNo Large-Scale 3D-Text Paired Data
Text-to-Head[12, 17, 30, 94]XX✓✓
[84]✓X✓✓
Text-to-Hair[28, 48, 62, 80]XX✓✓
[65]X✓XX
Text-to-Head-HairOurs✓✓✓✓
", + "bbox": [ + 130, + 88, + 441, + 142 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Comparison with current related methods.", + "bbox": [ + 135, + 143, + 437, + 156 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "enhances human-specific fidelity by fine-tuning 2D diffusion models on high-quality human mesh data. Despite these advancements, these approaches treat the head and hair as a holistic model, limiting support for downstream applications such as hairstyle transfer and editing. To enable flexible head-hair-disentangled generation, TECA [84] independently represents the head with mesh and the hair with NeRF [38]. However, these methods focus only on the realistic external appearance of hair without modeling its internal geometric structure, restricting their application for strand-level editing and physics-based simulation.", + "bbox": [ + 89, + 161, + 482, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Strand-Based Hair Creation. Automatic strand-based hair modeling has garnered significant attention from both industry [2, 9, 10] and academia [3, 14, 33, 82], as it enables downstream physics-based rendering and simulation. Such methods can be broadly categorized into two types: (1) reconstructing hair from input images or videos and (2) generating hair based on control conditions. Works on the former one [15, 21, 32, 35, 41, 55, 57, 61, 64, 67, 72, 73, 78, 83, 89-91, 93] relies on classical 3D reconstruction frameworks [20, 36, 38, 45, 59] and incorporates hair-specific features [44] to realize accurate 3D strand reconstruction. However, these methods depend on constrained multi-view images, making hair creation costly and challenging in less controlled settings. Recently, 3D hair generation approaches [13, 92] have emerged using GANs or parametric models, but they struggle with controllable generation. The most relevant work to ours is HAAR [65], which pioneers text-to-strand by training a text-conditioned hair map diffusion model. However, HAAR relies on a large-scale paired dataset including 9825 haircuts and prompts with limited diversity, restricting its ability to generate novel hairstyles beyond the training data. Moreover, HAAR overlooks hair texture and geometry that adapts to specific head shapes, further limiting its application range.", + "bbox": [ + 91, + 327, + 483, + 689 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Compared to previous methods, StrandHead generates 3D heads featuring strand-level hairstyles. Instead of requiring large-scale hair-text paired data, we achieve realistic and diverse 3D hair strand generation by using 2D/3D human-centric priors. We summarize the main differences between our methods and related works in Tab. 1.", + "bbox": [ + 89, + 689, + 482, + 780 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 792, + 181, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 816, + 230, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "FLAME [26] is a 3D parametric head model that represents the human head shape, pose, and expression using a compact set of parameters. Given a set of shape parameters $\\beta$ , pose parameters $\\theta_{\\mathrm{pose}}$ , and expression parameters $\\psi_{\\mathrm{exp}}$", + "bbox": [ + 89, + 839, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "FLAME can create a 3D head mesh $M$ .", + "bbox": [ + 511, + 90, + 777, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "HumanNorm [17] is a text-driven, high-quality 3D head generation method. In specific, it models the geometry and texture of the human head using DMTet [60] and a texture field. The alignment between the text $y$ and the optimized representation $\\theta$ is achieved using the following SDS loss [47] with T2I diffusion models $\\phi$ :", + "bbox": [ + 511, + 106, + 903, + 196 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\theta} \\mathcal {L} _ {S D S} = \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\epsilon_ {\\phi} \\left(x _ {t}; y, t\\right) - \\epsilon\\right) \\frac {\\partial x}{\\partial \\theta} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 196, + 903, + 229 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $x = g(\\theta)$ is the image rendered from $\\theta$ by a differentiable renderer $g$ , $t$ is the time step, $x_{t} = x + \\epsilon$ is a noised version of $x$ , $\\epsilon_{\\phi}(x_{t};y,t)$ is the denoised image, and $w(t)$ is a weighting function. HumanNorm employs human-specific diffusion models fine-tuned on high-quality human mesh data to replace the general diffusion model, enhancing the fidelity of the generated human heads.", + "bbox": [ + 511, + 229, + 903, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural Scalp Textures (NST) [55] is an efficient hair representation where each pixel stores a feature vector conveying the shape information of a single strand at the corresponding scalp location. Using a pre-trained hair strand generator [55] $G$ , the low-dimensional 2D neural scalp texture $T$ can be decoded into high-dimensional 3D strand polylines $S = \\{s^i\\}_{i=1}^{N_s}$ , where $N_s$ is the number of strands and $s^i = \\{p_j^i\\}_{j=1}^{N_p}$ consists of $N_p$ 3D points. This process is formulated as follows: $S = G(T)$ .", + "bbox": [ + 511, + 335, + 905, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overview", + "text_level": 1, + "bbox": [ + 513, + 483, + 624, + 497 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a text prompt, StrandHead aims to create a realistic 3D head avatar with strand-based hair without relying on large-scale 3D-text paired data. Fig. 2 provides an overview of our three-stage pipeline. First, we generate a FLAME-aligned 3D bald head for accurate hair initialization using human-specific diffusion models (Sec. 3.3). In the subsequent stages, we model the reasonable geometry and realistic texture of 3D hair strands by leveraging rich 2D/3D human-centric priors (Sec. 3.4 and Sec. 3.5).", + "bbox": [ + 511, + 505, + 905, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Bald Head Generation", + "text_level": 1, + "bbox": [ + 513, + 650, + 723, + 665 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve accurate hair initialization and optimize 3D hair strands in subsequent steps using human-specific 2D generative priors, we first need to obtain a reasonable and semantic-aligned 3D bald head from the text (Fig. 2-(a)). To achieve this goal, we enhance HumanNorm [17] by incorporating an evolving FLAME [26] model while optimizing the bald head DMTet $\\theta_{\\mathrm{h}}$ . This provides accurate semantic information and prevents unnatural geometry. Please refer to the Supp. Mat. for details on bald head generation.", + "bbox": [ + 511, + 672, + 905, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.4. Hair Geometry Generation", + "text_level": 1, + "bbox": [ + 511, + 818, + 758, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With a FLAME-aligned bald head, we can generate 3D hair in the scalp area. Specifically, we first initialize the haircut based on hairstyle descriptions, and then further sculpt the 3D hair shape utilizing 2D/3D human-centric priors.", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13395", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ad99dc90aabd00317097bf4905e68ae7922bcf2888ea155b69f46dcbb32984e5.jpg", + "image_caption": [ + "Figure 2. StrandHead includes three stages: (a) We first create a FLAME-aligned 3D bald head using the improved HumanNorm [17]. (b) Next, we introduce a differentiable prismatization algorithm to enable human-specific geometry-aware 2D diffusion models to supervise hair shape modeling. Additionally, two losses inspired by 3D hair geometric priors are applied to further regularize the hair geometry. (c) Finally, we use a human-specific normal-conditioned 2D diffusion model to generate lifelike hair textures." + ], + "image_footnote": [], + "bbox": [ + 133, + 89, + 867, + 323 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hair Initialization. To achieve reasonable and diverse hair initialization, we first utilize ChatGPT [43] to select the 20 most representative hairstyles from the USC-HairSalon Dataset [15]. We then optimize neural scalp textures to fit the selected hairstyle using the following loss function:", + "bbox": [ + 89, + 385, + 483, + 460 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {f i t}} = \\sum_ {i = 1} ^ {N _ {s}} \\sum_ {j = 1} ^ {N _ {p}} \\| \\hat {p} _ {j} ^ {i} - p _ {j} ^ {i} \\| _ {2} + \\lambda_ {\\mathrm {o r i}} \\left(1 - \\hat {o} _ {j} ^ {i} \\cdot o _ {j} ^ {i}\\right) + \\lambda_ {\\mathrm {c u r}} \\| \\hat {c} _ {j} ^ {i} - c _ {j} ^ {i} \\| _ {1}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 462, + 483, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{p}_j^i$ and $p_j^i$ are the position of the $j$ -th point on the $i$ -th polyline of the GT and generated hair, respectively, and $\\hat{o}_j^i$ , $o_j^i$ , $\\hat{c}_j^i$ , and $c_j^i$ denote their orientation and curvature, respectively. Here, orientation $o_j^i = (p_{j+1}^i - p_j^i) / \\|p_{j+1}^i - p_j^i\\|_2$ represents the direction of change in strand position, and curvature $c_j^i = \\|o_j^i - o_{j-1}^i\\|_2$ denote the rate of change in strand orientation. Given a hair description, we select an optimal pre-trained NST as initialization and further optimize it in subsequent stages.", + "bbox": [ + 89, + 518, + 483, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Discussion on Strand-Based Differentiable Rendering. With reasonable hair initialization, the next step is to leverage human-specific 2D diffusion models to further sculpt the hair geometry. However, due to the lack of a stable strand-based differentiable renderer, it is very challenging to optimize hair shape using 2D generative priors like other SDS-based methods. Besides, we expect hair strands to produce mesh-style smooth geometry and texture maps to fully exploit the powerful generative capabilities of 2D diffusion models pre-trained on human mesh data. A feasible alternative is to first differentiably convert the strands into quad meshes [82], similar to NeuralHaircut [64], and then leverage SDS-based mesh optimization frameworks to model hair strands. However, this non-watertight stripe-like mesh easily produces ambiguous normal maps or excessively thin sides (see the zoom image of Fig. 3-(c)), which significantly", + "bbox": [ + 89, + 659, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ed9e26760e95466fa25acc4d1f5bb5d0d729d50c0068141ef3c4704f96d3870e.jpg", + "image_caption": [ + "(b) A strand-to-mesh conversion example" + ], + "image_footnote": [], + "bbox": [ + 527, + 383, + 890, + 489 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9e422028a461b3b809ddd48b3bde11da94fc0d77b134209cc6ebed939ac9f01c.jpg", + "image_caption": [ + "(c) Differentiable prismatization's advantages" + ], + "image_footnote": [], + "bbox": [ + 526, + 500, + 879, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 3. The differentiable prismatization algorithm's (a) gradient backpropagation process, (b) a strand-to-mesh conversion example and (c) advantages over NeuralHaircut [64]. Non-watertight quad meshes can easily produce ambiguous normal maps, which significantly reduce the stability of hair shape modeling (see the drifting hair highlighted by the oval dotted box in (c)).", + "bbox": [ + 511, + 625, + 905, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "reduce the optimization stability (see the drifting hair highlighted by the oval dotted box in Fig. 3-(c)). Therefore, a differentiable strand-to-mesh conversion method with superior optimization properties is urgently needed.", + "bbox": [ + 511, + 717, + 905, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Differentiable Prismatization (DP). To this end, we propose a novel differentiable prismatization algorithm inspired by the internal structure of 3D hair (i.e., hair fiber is a dielectric cylinder covered with tilted scales and with a pigmented interior [19, 34]). This algorithm can efficiently and differentiably convert strands into watertight prismatic meshes with arbitrary thickness and lateral edges to approximate the cylindrical structure of the hair. This enables", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13396", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "smooth backpropagation of gradients from the SDS loss to the 3D strand representation using mesh-based differentiable renderers [22], thereby making it possible to fully exploit 2D generative priors distilled from high-fidelity human mesh data for modeling realistic hair strands. Fig. 3-(a) visualizes the complete gradient backpropagation process.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In specific, given a hair strand $s$ , our DP converts it into a watertight prismatic mesh with $K$ lateral edges and radius $R$ through the following five steps: (1) Compute the Initial Normal Vector. (2) Generate $K$ Rotated Normals. (3) Translate to Form Lateral Edges. (4) Construct Lateral Faces. (5) Construct Top and Bottom Faces. An example of this conversion is shown in Fig. 3-(b). Unlike quad meshes [64], our DP effectively avoids ambiguous normals, enabling smooth gradient backpropagation and hair strand optimization, as shown in Fig. 3-(c). Further algorithm details are provided in the Supp. Mat.", + "bbox": [ + 89, + 184, + 482, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the help of our proposed differentiable prismatization algorithm, we optimize the neural scalp texture $T$ by utilizing rich human-specific 2D generative priors through the following SDS losses (Fig. 2-(b)):", + "bbox": [ + 89, + 352, + 483, + 412 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {T} \\mathcal {L} _ {S D S} ^ {\\mathrm {h n}} = \\mathbb {E} _ {t, \\epsilon} \\left[ \\left(\\epsilon_ {\\phi_ {\\mathrm {h n}}} \\left(n _ {t} ^ {\\mathrm {h + s}}; y _ {\\mathrm {h + s}}, t\\right) - \\epsilon\\right) \\frac {\\partial n ^ {\\mathrm {h + s}}}{\\partial T} \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 417, + 482, + 450 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {T} \\mathcal {L} _ {S D S} ^ {\\mathrm {h d}} = \\mathbb {E} _ {t, \\epsilon} \\left[ \\left(\\epsilon_ {\\phi_ {\\mathrm {h d}}} \\left(d _ {t} ^ {\\mathrm {h + s}}; y _ {\\mathrm {h + s}}, t\\right) - \\epsilon\\right) \\frac {\\partial d ^ {\\mathrm {h + s}}}{\\partial T} \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 454, + 482, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\phi_{\\mathrm{hn}}$ and $\\phi_{\\mathrm{hd}}$ are the human-specific normal-adapted and depth-adapted diffusion models from Human-Norm [17], $n^{\\mathrm{h + s}}$ and $d^{\\mathrm{h + s}}$ are the rendered normal and depth maps of the head DMTet $\\theta_{\\mathrm{h}}$ with hair prismatic mesh $M_p = DP(G(T))$ , respectively. Here, $DP(\\cdot)$ denotes the differentiable prismatization operation function, $y_{\\mathrm{h + s}}$ is the full text description of the head with hair.", + "bbox": [ + 89, + 503, + 482, + 607 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, due to the excessive flexibility of hair strands, relying solely on the SDS loss leads to unnatural strand orientations and unrealistic hairstyles (Fig. 9). Furthermore, the lack of constrained multi-view images prevents us from leveraging hair-specific features [44] to provide powerful supervision as in previous works [55].", + "bbox": [ + 89, + 611, + 482, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Prior-Driven Losses. To address this issue, we propose two straightforward yet robust prior-driven losses based on observations of the 3D hair strand orientation and curvature to ensure the rationality of the generated hair strands. Specifically, we observe two geometric properties of hair: (1) neighboring strand orientations are highly consistent, and (2) strand curvature is strongly and positively correlated with the curliness of the hairstyle. To validate the properties, we calculate the cosine similarity of adjacent hair strand orientations $CS_{\\mathrm{ori}}$ and the average curvature $C_{\\mathrm{mean}}$ of 343 hairstyles in the USC-HairSalon dataset [15] using the following equations and visualize their distributions and some examples in Fig. 4:", + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/64cf141a3d757c21a2f49eb33cc63169c2a91e50e6d29ccfdefa63c0c7c6de40.jpg", + "image_caption": [ + "(a) Distribution of Haircuts' $CS_{\\mathrm{ori}}$" + ], + "image_footnote": [], + "bbox": [ + 524, + 88, + 699, + 189 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/21154d866dc276c48d57f1652882bea263b5d835d6f8129359cd98fba729d170.jpg", + "image_caption": [ + "(b) Distribution of Haircuts' $C_{\\text{mean}}$" + ], + "image_footnote": [], + "bbox": [ + 705, + 88, + 882, + 189 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/106b55f2d6e435ee3533a0ea1403203377d017baff4e427d3afbf7a6b0c05f6c.jpg", + "image_caption": [ + "$CS_{\\mathrm{ori}} = 0.9949$ $C_{\\mathrm{mean}} = 0.0$" + ], + "image_footnote": [], + "bbox": [ + 526, + 202, + 622, + 263 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c3e4fac1e2faadf5565ed83ca0979ce893b44cfd999b0c7f1e8b22108c3415ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 203, + 736, + 262 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/80b8a04d78694e2f8701df09132eeaba3f0bb6c953461c285b4c8b9630905698.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 738, + 202, + 815, + 262 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/664f8c1138d4a077c29e5290468cdf5f593cd7acce4383e23c069f731331aacd.jpg", + "image_caption": [ + "$CS_{\\mathrm{ori}} = 0.9265$", + "(c) Some hairstyles from the USC-HairSalon dataset", + "Figure 4. Observation of hair geometric features: (1) neighboring strand orientations are highly consistent. (2) strand curvature is strongly and positively related to the haircut curliness." + ], + "image_footnote": [], + "bbox": [ + 531, + 271, + 596, + 330 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b66a1339d299f12c2485cbce97515ef4a80a0b3dedaf282728378f2025e593cf.jpg", + "image_caption": [ + "1 $CS_{\\mathrm{ori}} = 0.9868$ $C_{\\mathrm{mean}} = 0.0290$", + "$C_{\\text{mean}} = 0.0598$" + ], + "image_footnote": [], + "bbox": [ + 598, + 271, + 656, + 330 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ea27960902ded73167b12cc263ffcdc40ac65d118713789938f41b2818a15c78.jpg", + "image_caption": [ + "$CS_{\\mathrm{ori}} = 0.9719$ $C_{\\mathrm{mean}}$" + ], + "image_footnote": [], + "bbox": [ + 666, + 271, + 736, + 330 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5b03f838af5eb63107b0f2ea2b8e1fe975a06e775c8e315c3ad5609b6db2c663.jpg", + "image_caption": [ + "$C S_{\\mathrm{ori}} = 0.9605$ $C_{\\mathrm{mean}} = 0.0493$", + "$= 0.0870$ $CS_{\\mathrm{ori}} = 0.9297$ $C_{\\mathrm{mean}} = 0.1059$" + ], + "image_footnote": [], + "bbox": [ + 738, + 271, + 826, + 330 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC S _ {\\text {o r i}} = \\frac {1}{N _ {s} N _ {p}} \\sum_ {i = 1} ^ {N _ {s}} \\sum_ {j = 1} ^ {N _ {p}} \\sum_ {k \\in A (i)} \\frac {o _ {j} ^ {i} \\cdot o _ {j} ^ {k}}{| A (i) |}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 398, + 905, + 441 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC _ {\\text {m e a n}} = \\frac {1}{N _ {s} N _ {p}} \\sum_ {i = 1} ^ {N _ {s}} \\sum_ {j = 1} ^ {N _ {p}} c _ {j} ^ {i}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 446, + 903, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $A(i)$ is the set of adjacent strands for the $i$ -th strand, and $|A(i)|$ is the number of neighboring strands.", + "bbox": [ + 511, + 498, + 903, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As illustrated, over $95\\%$ of hairstyles exhibit an orientation similarity above 0.9, and the average curvature is significantly positively related to curliness, which successfully validates the properties. Based on these observations, we introduce orientation consistency loss and curvature regularization to guide the local and global strand shapes, formulated as follows:", + "bbox": [ + 511, + 529, + 903, + 632 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {o r i}} = 1 - C S _ {\\mathrm {o r i}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 636, + 903, + 651 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c u r}} = \\left\\| C _ {\\text {m e a n}} - C _ {\\text {t a r g e t}} \\right\\| _ {1}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 655, + 903, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $C_{\\mathrm{target}}$ represents the target average curvature set according to the input hair description. These two losses, inspired by 3D hair geometry priors, regularize the hair shape by supervising the consistency of orientations between adjacent strands and the overall hairstyle's curvature.", + "bbox": [ + 511, + 680, + 903, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Additionally, we introduce a series of losses $\\mathcal{L}_{bbox}$ , $\\mathcal{L}_{face}$ , and $\\mathcal{L}_{colli}$ , which prevent the hair from exceeding the bounding box, obscuring the face, and colliding with the head, respectively. These losses can further improve the rationality of hair geometry. Their details are provided in the Supp. Mat. The final loss function for optimizing hair shape is expressed as follows:", + "bbox": [ + 511, + 755, + 905, + 859 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {h a i r - g e o}} = \\mathcal {L} _ {S D S} ^ {\\mathrm {h d}} + \\lambda_ {S D S} ^ {\\mathrm {h n}} \\mathcal {L} _ {S D S} ^ {\\mathrm {h n}} + \\lambda_ {\\mathrm {o r i}} \\mathcal {L} _ {\\mathrm {o r i}} + \\lambda_ {\\mathrm {c u r}} \\mathcal {L} _ {\\mathrm {c u r}} \\\\ + \\lambda_ {\\text {b b o x}} \\mathcal {L} _ {\\text {b b o x}} + \\lambda_ {\\text {f a c e}} \\mathcal {L} _ {\\text {f a c e}} + \\lambda_ {\\text {c o l l i}} \\mathcal {L} _ {\\text {c o l l i}}. \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 864, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13397", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fdec83993177d362470f63f08009e56af763091211d580eb5e9011ee5930ad6b.jpg", + "image_caption": [ + "\"An elderly Caucasian grandmother with thin lips and a straight, short silver bob.\"" + ], + "image_footnote": [], + "bbox": [ + 147, + 88, + 251, + 148 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f0999092568345f52319c6c3be356adc2e521d4df0e83b4437659fef89e95a97.jpg", + "image_caption": [ + "\"Brad Pitt with medium-length, tousled blonde hair.\"" + ], + "image_footnote": [], + "bbox": [ + 261, + 90, + 380, + 152 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4f01c4af11edfbae20afbbc640ebc6970c6166a17dbfe3e576cdbc3b8b239b1d.jpg", + "image_caption": [ + "\"Chris Hemsworth with side-swept sandy blonde hair.\"" + ], + "image_footnote": [], + "bbox": [ + 395, + 90, + 488, + 151 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/88c9f5b6d8cb5c835936a4e7e4e415d56d48bb0226780cc84d8e42af8ac9700a.jpg", + "image_caption": [ + "Lionel Messi with a blonde mohawk hairstyle." + ], + "image_footnote": [], + "bbox": [ + 514, + 90, + 602, + 152 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c3b02bf01fe6f15c8616ba0df153e2511b817b0a6717681e2ea54610dc443013.jpg", + "image_caption": [ + "\"Scarlet Johansson with a textured blonde bob.\"" + ], + "image_footnote": [], + "bbox": [ + 617, + 90, + 725, + 151 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b924b600bcb8f03d1405a30c0d4184493a189046da16156995f84023d7cd28c6.jpg", + "image_caption": [ + "\"A mature European woman with a straight nose and medium-length, straight auburn hair.\"" + ], + "image_footnote": [], + "bbox": [ + 750, + 89, + 851, + 148 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3959cf6a4eaf806e371795c51ced39ec83533c857fe1979a4db0ba07b6471e92.jpg", + "image_caption": [ + "Beyonce with voluminous, curly honey blonde hair." + ], + "image_footnote": [], + "bbox": [ + 143, + 178, + 245, + 239 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b361d81b71aafa0b986da759404cebd173fe277e1f1cc456848f81b0f846a061.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 246, + 178, + 297, + 239 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/eeb6efaff330d0842e09289795f92d1bac4a814c2663def2daf0b5188767d178.jpg", + "image_caption": [ + "Dwayne Johnson with a short black Afro." + ], + "image_footnote": [], + "bbox": [ + 313, + 178, + 406, + 238 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0babe7ef1075c71af234f22c026dbfab50140b9b2fa0fe572d50c4278ea85b32.jpg", + "image_caption": [ + "\"A beautiful Caucasian girl with delicate features and long silky golden blonde hair.\"" + ], + "image_footnote": [], + "bbox": [ + 426, + 178, + 498, + 238 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/14363b7e59c3001c0e63fd3818c5dd4bd9c5d33926b2eb92a660b2799a4da647.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 178, + 573, + 238 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/732b9149abd68a2c3f88c8f0bd50ddfa89bc475126481a47da2b6e1417bf5a73.jpg", + "image_caption": [ + "A confident Hispanic man with a beard and a curly black afro." + ], + "image_footnote": [], + "bbox": [ + 589, + 178, + 638, + 238 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c9ef585d10485663d0345f37f687314a02524d06220df44c67a5955b2d8a1bd8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 176, + 779, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d4891a97efda9acd967d9c5e35acf933ab01b5591e45bd31f37464ab5aeb5aeb.jpg", + "image_caption": [ + "\"A charming Caucasian woman with a straight nose and shoulder-length curly sleek blonde hair.\"" + ], + "image_footnote": [], + "bbox": [ + 784, + 176, + 851, + 236 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e1e8939fe535d0fe1adf32a97bd040dc539acf88867d9957f540a3ebc3bde40e.jpg", + "image_caption": [ + "\"A curly dark blonde haircut.\"" + ], + "image_footnote": [], + "bbox": [ + 147, + 270, + 254, + 414 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/029876642d1f015224e2a3893c2bdf656708b8fe171773eb14295d044be18679.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 271, + 305, + 414 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b4f13ccfae3ccbdf592bf6e1eb5a9e65c05a8d7ff026e4e92a85e52ba1b0995f.jpg", + "image_caption": [ + "\"Medium-length, straight auburn hair.\"" + ], + "image_footnote": [], + "bbox": [ + 307, + 270, + 403, + 412 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a8fda1589724736c4684dc24a17abfc5ea15c59dea5dd4c3efa194dc2d776adc.jpg", + "image_caption": [ + "\"Long way chestnut brown hair.\"" + ], + "image_footnote": [], + "bbox": [ + 426, + 270, + 495, + 328 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/86166d947a7aa6c3c593494c57a6c28cc8b0edbb88b9dafeacc5eb959f259bdc.jpg", + "image_caption": [ + "\"A short, wavy brown bob.\"" + ], + "image_footnote": [], + "bbox": [ + 522, + 270, + 576, + 328 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f9c10bbfcef42bb59080e090bbbac45031de02e22695d7aaee51304e3b3bc21a.jpg", + "image_caption": [ + "\"Long silky golden blonde hair.\"" + ], + "image_footnote": [], + "bbox": [ + 604, + 270, + 660, + 328 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e920cbd347fc90a1af3bef2827fddfc0a91f9bc2632e770d14ac46a044d3e633.jpg", + "image_caption": [ + "\"A light brown mohawk hairstyle.\"" + ], + "image_footnote": [], + "bbox": [ + 689, + 268, + 736, + 328 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/02dc237ab140ad69005e7cb4d61271fe98c6b8856f817429e8b0393323b0a7d9.jpg", + "image_caption": [ + "\"Side-swept sandy blonde hair.\"" + ], + "image_footnote": [], + "bbox": [ + 776, + 268, + 828, + 328 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a74c041fd1bc22d8cc1a7854ab1bd4ce119d9c256766ab717b7c261af25b8597.jpg", + "image_caption": [ + "\"A curly black afro.\"", + "Figure 5. Examples of high-fidelity and diverse 3D heads and strand-accurate haircuts generated by our method. The upper visualization includes rendered color and normal maps of the head and hair prismatic meshes. The lower visualization shows the physics-based hair strand rendering result using Blender [10]. For better strand-based visualization, we interpolate generated hair to approximately 10,000 strands and apply a consistent appearance. Please zoom in for detailed views, and refer to the Supp. Mat. for video demonstrations." + ], + "image_footnote": [], + "bbox": [ + 431, + 351, + 488, + 415 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d50f87087e83a7104ec99d896989ed0b2f7836635e64d8fc838332214644ada5.jpg", + "image_caption": [ + "Medium-length wavy caramel-colored hair." + ], + "image_footnote": [], + "bbox": [ + 524, + 353, + 576, + 412 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6d7a97f106771f43c430ecda7c3ee2e9337dbc512685888ade40b6bcefa77ffd.jpg", + "image_caption": [ + "\"A side-swept silver haircut.\"" + ], + "image_footnote": [], + "bbox": [ + 606, + 352, + 656, + 414 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7c2d2122a1904144ffdeb28709a0d673f30f4c055c630a82fe47bbf9b706ee53.jpg", + "image_caption": [ + "\"A sleek long black hairstyle.\"" + ], + "image_footnote": [], + "bbox": [ + 681, + 353, + 743, + 412 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/70d595b5a43aac4182a6b019945dfce1aa92e7ca1eafd73f4a620e0745f3be11.jpg", + "image_caption": [ + "\"Medium-length, tousled blonde hair.\"" + ], + "image_footnote": [], + "bbox": [ + 774, + 354, + 831, + 414 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Hair Texture Generation", + "text_level": 1, + "bbox": [ + 89, + 494, + 316, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next, we fix the optimized hair strand geometry and model the realistic hair strand texture under the supervision of the SDS loss as follows (Fig. 2-(c)):", + "bbox": [ + 89, + 516, + 483, + 561 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\psi_ {s}} \\mathcal {L} _ {S D S} ^ {\\mathrm {h c}} = \\mathbb {E} _ {t, \\epsilon} \\left(\\epsilon_ {\\phi_ {\\mathrm {h c}}} \\left(c _ {t} ^ {\\mathrm {h + s}}; n ^ {\\mathrm {h + s}}, y _ {\\mathrm {h + s}}, t\\right) - \\epsilon\\right) \\frac {\\partial c ^ {\\mathrm {h + s}}}{\\partial \\psi_ {\\mathrm {s}}}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 569, + 482, + 603 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\phi_{\\mathrm{hc}}$ is the human-specific normal-conditioned diffusion model from HumanNorm [17], $c^{\\mathrm{h + s}}$ is the rendered color map of the head texture field $\\psi_{\\mathrm{h}}$ and hair strand texture field $\\psi_{\\mathrm{s}}$ . Since the vanilla SDS loss often leads to color oversaturation, we replace it with the following MSDS loss [17] to further enhance the texture's realism in later iterations of texture optimization:", + "bbox": [ + 89, + 609, + 483, + 715 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\psi_ {s}} \\mathcal {L} _ {M S D S} ^ {\\mathrm {h c}} = \\mathbb {E} _ {t, \\epsilon} \\left(H \\left(c _ {t} ^ {\\mathrm {h + s}}; n ^ {\\mathrm {h + s}}, y _ {\\mathrm {h + s}}, t\\right) - \\epsilon\\right) \\frac {\\partial c ^ {\\mathrm {h + s}}}{\\partial \\psi_ {s}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 722, + 482, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $H(\\cdot)$ denotes the multi-step operation function. Additionally, we propose a strand-aware texture field that models orientation-dependent texture to better generate high-frequency color variations. Please refer to the Supp. Mat. for more their details.", + "bbox": [ + 89, + 763, + 483, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.6. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 848, + 307, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method is implemented using PyTorch [46] and ThreeStudio [11]. Experiments are conducted on an Ubuntu", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "server equipped with A6000 GPUs. Starting with a prompt such as \"A man with brown hair\", we first generate a bald head using \"A bald man\". Then, we generate the hair based on \"A man with brown hair\". Further implementation details are available in the Supp. Mat.", + "bbox": [ + 511, + 494, + 906, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 585, + 645, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Some examples of 3D head models with strand-based hair generated by StrandHead are shown in Fig. 5. Due to space limitations, only the key experimental settings and results are presented here. For more results, evaluations, and details, please refer to our Supp. Mat.", + "bbox": [ + 511, + 609, + 905, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 511, + 696, + 720, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare StrandHead with the SOTA methods for head avatar and haircut generation. The baseline methods for head generation include text-to-holistic-head works (HeadArtist [30], HumanNorm [17], and HeadStudio [94]) and text-to-decoupled-head work (TECA [84]). For text-to-hair generation, we consider general text-to-3D methods (MVDream [62], GaussianDreamer [80], LucidDreamer [28], and RichDreamer [48]), text-to-strand method (HAAR [65]), and TECA [84].", + "bbox": [ + 511, + 719, + 906, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset Construction. To create the dataset, we use ChatGPT[43] to randomly generate 30 text descriptions of heads with hair. These include 30 descriptions for assessing", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13398", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/726ab651220bc3ba911020f5f5816018680420bf204b6a632deee9dcd53356b0.jpg", + "image_caption": [ + "Figure 6. Qualitative comparisons with the SOTA methods. Since TECA [84] uses the vanilla NeRF to represent hair, rendering normals is not supported. HAAR [65] generates only the geometry of hair strands, so we first convert the strands into prismatic meshes using differentiable prismatization and then utilize TEXTure [53] to generate texture for visualization and comparison." + ], + "image_footnote": [], + "bbox": [ + 130, + 88, + 864, + 296 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e1a3d8f8618aa06101d317578e6207b453c1e9d061269bd34e9ae3766a7c1ab5.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison with HAAR [65]. HAAR, which does not model heads, often produces unreasonable hair-head collisions (highlighted in the black box)." + ], + "image_footnote": [], + "bbox": [ + 125, + 345, + 441, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "head generation and 30 for evaluating hair generation. The full list of prompts can be found in the Supp. Mat.", + "bbox": [ + 89, + 595, + 482, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Metrics. Current text-to-3D methods typically use CLIP-based metrics to assess text-image alignment and output quality. However, research [16, 31] shows that these metrics struggle to capture fine-grained alignment between 3D content and input prompts, as validated by our experiments in the Supp. Mat. To address this limitation, we draw inspiration from Progressive3D [7] and Barbie [66], using BLIP-VQA [23, 24] and BLIP2-VQA [23, 25] for evaluation. Specifically, we convert each prompt into two questions to separately verify head and hair. We then feed the rendered image of the generated 3D content into the VQA model, asking each question in sequence and using the probability of a “yes” response as our evaluation metric. For example, the head prompt “A man with black hair” is transformed into “Is the person in the picture a man?” and “Does the person in the picture have black hair?” The hair prompt “Black hair” becomes “Is the object in the picture black hair?” Finally, we conduct a user study by randomly", + "bbox": [ + 89, + 628, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d5b34825cbf36f128af514b3600ae54b787d474ac956baf7034b65cfc110127f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBLIP-VQA ↑BLIP2-VQA ↑GQP ↑TAP ↑
HeadArtist0.76670.96671.002.33
HeadStudio0.78330.88333.333.67
HumanNorm0.70000.95007.677.67
TECA0.73330.950034.3328.33
StrandHead (Ours)0.85000.966753.6758.00
MVDream0.90000.833324.6720.00
GaussianDreamer0.33330.36675.333.00
LucidDreamer0.80000.93335.335.00
RichDreamer0.83330.76674.005.67
TECA0.70000.76671.673.67
HAAR0.63330.20001.332.33
StrandHead (Ours)0.90000.900057.6760.33
", + "bbox": [ + 555, + 345, + 864, + 481 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Quantitative comparisons with the SOTA methods. The best and second-best results are highlighted in **bolded** and **underlined**, respectively. GQP: generation quality preference (\\%), TAP: text-image alignment preference (\\%).", + "bbox": [ + 511, + 483, + 906, + 539 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "selecting 10 generated examples and asking 30 volunteers to assess (1) generation quality and (2) text-image alignment, and select the best methods.", + "bbox": [ + 511, + 545, + 906, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparisons of Head Generation", + "text_level": 1, + "bbox": [ + 511, + 602, + 808, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As illustrated in the upper section of Tab. 2, StrandHead outperforms all comparison methods across every evaluation metric. The qualitative results on the left side of Fig. 6 further highlight the advantages of our method. Compared to other head generation methods, the head avatars generated by StrandHead exhibit not only more refined facial geometry and texture details but also strand-accurate hair with a plausible appearance that integrates seamlessly with physics-based simulation systems. To the best of our knowledge, StrandHead is the first head generation framework that realizes strand-level hair modeling, a capability that holds substantial potential for advancing human-centric 3D AIGC applications in the industry.", + "bbox": [ + 511, + 625, + 906, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparisons of Hair Generation", + "text_level": 1, + "bbox": [ + 511, + 832, + 803, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in the lower part of Tab. 2, our approach surpasses other methods across most evaluation metrics, and is only slightly lower than LucidDreamer (the latest open-source", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13399", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c40766afd0a49292a565959bd32f41365bfa422c41ada07c83ffd878d3bc0e45.jpg", + "image_caption": [ + "(b) Generated strand-based hair under different bald heads", + "Figure 8. Analysis of human-specific 2D diffusion models." + ], + "image_footnote": [], + "bbox": [ + 125, + 89, + 449, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "text-to-3D method) on BLIP2-VQA, ranking the second. The qualitative comparisons on the right side of Fig. 6 display sample generation results. Compared to general text-to-3D methods, our approach more precisely captures the internal structure of hair strands, resulting in realistic hair-cut geometry and appearance without incorporating incorrect content such as parts of the human head.", + "bbox": [ + 89, + 393, + 482, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Compared with HAAR [65], StrandHead relies on robust 2D/3D human-centric priors rather than large-scale paired datasets to supervise hair generation. This not only allows it to generate hairstyles that are uncommon in the dataset (see slicked-back and side-swept hairstyles in Fig. 7), but also models hair texture or geometry that adapts to specific head shapes (Fig. 8-(b)), thereby avoiding unnatural hairhead collisions (see the black box in Fig. 7).", + "bbox": [ + 89, + 501, + 483, + 622 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 635, + 243, + 651 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Human-Specific 2D Generative Priors. We demonstrate the importance of human-specific 2D generative priors from two aspects: (1) As shown in Fig. 8-(a), ignoring 2D supervision fails to generate meaningful hair. While a general 2D diffusion model can slightly improve modeling performance, it still falls short of producing high-quality results. Only by incorporating human-specific 2D generative priors can one achieve realistic textures and reasonable shapes for hairstyles. (2) Fig. 8-(b) illustrates generated hair under varying bald head conditions while maintaining a fixed initial hairstyle and hair prompt. The hairstyles created under the supervision of human-specific 2D diffusion models exhibit clear geometric and textural variations that adapt to specific bald heads. This strongly demonstrates the necessity of considering the bald head for hair strand generation and highlights the advantages of", + "bbox": [ + 89, + 659, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/caa771c0a1343c0ff68f158b02de5e524c828527c22e7cb49e6e0eaea649f817.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 544, + 88, + 691, + 195 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f5fff9e96c6ad00b9b22679b3c6b79fabbe20f13be35b581ca93b3517290221c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 95, + 870, + 195 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e0b563cded1f058245363216826416541b97eaacb3379cab272a282b58689713.jpg", + "image_caption": [ + "Figure 9. Ablation study on (a) orientation consistency loss $\\mathcal{L}_{\\mathrm{ori}}$ and (b) curvature regularization loss $\\mathcal{L}_{\\mathrm{cur}}$ .", + "Figure 10. Failure cases." + ], + "image_footnote": [], + "bbox": [ + 527, + 223, + 694, + 268 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b601b5b6baea13c04a08fba7c1b13ee2a6eb9e0540e7ae07225ecd5eea0b057c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 223, + 890, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "StrandHead over HAAR [65].", + "bbox": [ + 511, + 287, + 712, + 301 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Prior-Driven Losses. Fig. 9-(a) demonstrates that incorporating $\\mathcal{L}_{\\mathrm{ori}}$ provides strong guidance for generating natural hair by supervising the consistency of orientations between adjacent strands. As shown in Fig. 9-(b), our $\\mathcal{L}_{\\mathrm{cur}}$ preserves the desired curliness by constraining the overall hairstyle's curvature. In summary, these two losses, inspired by 3D hair geometry priors, effectively enhance the realism and rationality of the generated haircut.", + "bbox": [ + 511, + 301, + 906, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 436, + 633, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, a novel framework StrandHead is proposed to generate realistic hair-disentangled 3D head avatars using text prompts. With a series of 2D generative models pre-trained on human-centric data, StrandHead generates strand-based hair requiring no large-scale text-hair paired data for supervision. This goal is achieved mainly by two approaches: (1) proposing a novel differentiable prismatization method that transforms the hair strands into prismatic meshes, and (2) using the prior-driven losses inspired by the observation of real-world haircut and the statistical features of human-crafted 3D hair data. The former one produces watertight geometry more similar to the original hair strands, boosting a stable optimization with significantly less normal ambiguity. Besides, the latter one well constrains the optimization with reliable guidance to generate reasonable 3D haircuts. Extensive experiments demonstrate that StrandHead obtains the state-of-the-art performance on 3D head avatars and haircuts generation. The generated results are easily implemented in industrial software to produce physical simulation and high-fidelity rendering.", + "bbox": [ + 511, + 462, + 906, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Future Work. Due to the limited representation capabilities of the strand generator, StrandHead is unable to create highly complex 3D hairstyles (e.g., dreadlocks and ponytails in Fig. 10). Additionally, although the SDS-based method effectively leverages 2D generative priors, its high computational cost restricts practical application. In future work, we aim to address these challenges by incorporating richer hairstyle datasets and exploring deeper hair priors.", + "bbox": [ + 511, + 765, + 908, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13400", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported by the National Science Foundation of China (62376121).", + "bbox": [ + 89, + 90, + 485, + 121 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 151, + 187, + 167 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shivangi Aneja, Justus Thies, Angela Dai, and Matthias Nießner. Clipface: Text-guided editing of textured 3d morphable models. ACM SIGGRAIPH, 2022. 2", + "[2] Inc. Autodesk. Autodesk maya - 3d animation and modeling software. https://www.autodesk.com/products/maya, 2024.2, 3", + "[3] Gaurav Bhokare, Eisen Montalvo, Elie Diaz, and Cem Yuksel. Real-time hair rendering with hair meshes. In ACM SIGGRAIPH, 2024. 3", + "[4] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, 2023. 2", + "[5] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Int. Conf. Comput. Vis., 2023. 2", + "[6] Zilong Chen, Feng Wang, and Huaping Liu. Text-to-3d using gaussian splatting. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2", + "[7] Xinhua Cheng, Tianyu Yang, Jianan Wang, Yu Li, Lei Zhang, Jian Zhang, and Li Yuan. Progressive3d: Progressively local editing for text-to-3d content creation with complex semantic prompts. In Int. Conf. Learn. Represent., 2023. 7, 5", + "[8] Hang Dai, Nick Pears, William Smith, and Christian Duncan. Statistical modeling of craniofacial shape and texture. Int. J. Comput. Vis., 2020. 2", + "[9] Inc. Epic Games. Unreal engine - real-time 3d creation tool. https://www.unrealengine.com/, 2024.2, 3", + "[10] Blender Foundation. Blender - a 3d modelling and rendering software. https://www.blender.org/, 2024.2,3,6, 7", + "[11] Yuan-Chen Guo, Ying-Tian Liu, Ruizhi Shao, Christian Laforte, Vikram Voleti, Guan Luo, Chia-Hao Chen, Zixin Zou, Chen Wang, Yan-Pei Cao, and Song-Hai Zhang. threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio, 2023.6", + "[12] Xiaoping Han, Yukang Cao, K. Han, Xiatian Zhu, Jiankang Deng, Yi-Zhe Song, Tao Xiang, and Kwan-Yee K. Wong. Headsculpt: Crafting 3d head avatars with text. In Adv. Neural Inform. Process. Syst., 2023. 2, 3", + "[13] Chengan He, Xin Sun, Zhixin Shu, Fujun Luan, Soren Pirk, Jorge Alejandro Amador Herrera, Dominik L Michels, Tuanfeng Y Wang, Meng Zhang, Holly Rushmeier, and Yi Zhou. Perm: A parametric representation for multi-style 3d hair modeling. arXiv preprint arXiv:2407.19451, 2024. 2, 3", + "[14] Jerry Hsu, Tongtong Wang, Zherong Pan, Xifeng Gao, Cem Yuksel, and Kui Wu. Real-time physically guided hair interpolation. ACM Trans. Graph., 2024. 3" + ], + "bbox": [ + 93, + 176, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Liwen Hu, Chongyang Ma, Linjie Luo, and Hao Li. Single-view hair modeling using a hairstyle database. ACM Trans. Graph., 2015. 3, 4, 5, 2", + "[16] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. In Adv. Neural Inform. Process. Syst., 2023. 7, 5", + "[17] Xin Huang, Ruizhi Shao, Qi Zhang, Hongwen Zhang, Ying Feng, Yebin Liu, and Qing Wang. Humannorm: Learning normal diffusion model for high-quality and realistic 3d human generation. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3, 4, 5, 6, 1", + "[18] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2", + "[19] James T Kajiya and Timothy L Kay. Rendering fur with three dimensional textures. ACM Siggraph Computer Graphics, 1989. 4", + "[20] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 2023. 2, 3", + "[21] Zhiyi Kuang, Yiyang Chen, Hongbo Fu, Kun Zhou, and Youyi Zheng. Deepmvshair: Deep hair modeling from sparse views. In ACM SIGGRAPH Asia, 2022. 3", + "[22] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Trans. Graph., 2020. 5", + "[23] Dongxu Li, Junnan Li, Hung Le, Guangsen Wang, Silvio Savarese, and Steven C.H. Hoi. LAVIS: A one-stop library for language-vision intelligence. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), 2023. 7, 5", + "[24] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, 2022. 7, 5", + "[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International Conference on Machine Learning, 2023. 7, 5", + "[26] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 2017. 2, 3", + "[27] Weiyu Li, Rui Chen, Xuelin Chen, and Ping Tan. Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. arXiv preprint arXiv:2310.02596, 2023. 2", + "[28] Yixun Liang, Xin Yang, Jiantao Lin, Haodong Li, Xiaogang Xu, and Yingcong Chen. Luciddreamer: Towards high-fidelity text-to-3d generation via interval score matching. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3, 6", + "[29] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13401", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Hongyu Liu, Xuan Wang, Ziyu Wan, Yujun Shen, Yibing Song, Jing Liao, and Qifeng Chen. Headartist: Text-conditioned 3d head generation with self score distillation. In ACM SIGGRAIPH, 2024. 2, 3, 6, 1", + "[31] Yujie Lu, Xianjun Yang, Xiujun Li, Xin Eric Wang, and William Yang Wang. Llmscore: Unveiling the power of large language models in text-to-image synthesis evaluation. In Adv. Neural Inform. Process. Syst., 2024. 7, 5", + "[32] Haimin Luo, Min Ouyang, Zijun Zhao, Suyi Jiang, Longwen Zhang, Qixuan Zhang, Wei Yang, Lan Xu, and Jingyi Yu. Gaussianhair: Hair modeling and rendering with light-aware gaussians. arXiv preprint arXiv:2402.10483, 2024. 3", + "[33] Ryota Maeda, Kenshi Takayama, and Takafumi Taketomi. Refinement of hair geometry by strand integration. In Comput. Graph. Forum, 2023. 3", + "[34] Stephen R Marschner, Henrik Wann Jensen, Mike Cammarano, Steve Worley, and Pat Hanrahan. Light scattering from human hair fibers. ACM Trans. Graph., 2003. 4", + "[35] Givi Meishvili, James Clemoes, Charlie Hewitt, Zafiirah Hosenie, Xian Xiao, Martin de La Gorce, Tibor Takacs, Tadas Baltrusaitis, Antonio Criminisi, Chyna McRae, et al. Hairmetry: Fairness-aware hairstyle classification. arXiv preprint arXiv:2410.11528, 2024. 3", + "[36] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 3", + "[37] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2", + "[38] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Eur. Conf. Comput. Vis., 2020. 2, 3", + "[39] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In ACM SIGGRAPH Asia, 2022. 2", + "[40] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 2022. 1", + "[41] Giljoo Nam, Chenglei Wu, Min H Kim, and Yaser Sheikh. Strand-accurate multi-view hair capture. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 3", + "[42] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2", + "[43] OpenAI. Chatgpt. https://openai.com/, 2024.4, 6, 2", + "[44] Sylvain Paris, Hector M Briceno, and François X Sillion. Capture of hair geometry from multiple images. ACM Trans. Graph., 2004. 3, 5", + "[45] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning con" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tinuous signed distance functions for shape representation. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 3", + "[46] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Adv. Neural Inform. Process. Syst., 2019. 6", + "[47] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In Int. Conf. Learn. Represent., 2022. 2, 3", + "[48] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng Bo, and Xiaoguang Han. Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3, 6", + "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 2", + "[50] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, 2021. 2", + "[51] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2", + "[52] Anurag Ranjan, Timo Bolkart, Soubhik Sanyal, and Michael J Black. Generating 3d faces using convolutional mesh autoencoders. In Eur. Conf. Comput. Vis., 2018. 1", + "[53] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In ACM SIGGRAIPH, 2023. 7", + "[54] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2", + "[55] Radu Alexandru Rosu, Shunsuke Saito, Ziyan Wang, Chenglei Wu, Sven Behnke, and Giljoo Nam. Neural strands: Learning hair geometry and appearance from multi-view images. In Eur. Conf. Comput. Vis., 2022. 2, 3, 5", + "[56] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In Adv. Neural Inform. Process. Syst., 2022. 2", + "[57] Shunsuke Saito, Liwen Hu, Chongyang Ma, Hikaru Ibayashi, Linjie Luo, and Hao Li. 3d hair synthesis using volumetric variational autoencoders. ACM Trans. Graph., 2018. 3", + "[58] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13402", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[59] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In Eur. Conf. Comput. Vis., 2016. 3", + "[60] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In Adv. Neural Inform. Process. Syst., 2021. 2, 3, 1", + "[61] Yuefan Shen, Shunsuke Saito, Ziyan Wang, Olivier Maury, Chenglei Wu, Jessica Hodgins, Youyi Zheng, and Giljoo Nam. Ct2hair: High-fidelity 3d hair modeling using computed tomography. ACM Trans. Graph., 2023. 3", + "[62] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. In Int. Conf. Learn. Represent., 2023. 2, 3, 6", + "[63] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In Int. Conf. Learn. Represent., 2014. 2", + "[64] Vanessa Sklyarova, Jenya Chelishev, Andreea Dogaru, Igor Medvedev, Victor Lempitsky, and Egor Zakharov. Neural haircut: Prior-guided strand-based hair reconstruction. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 2, 3, 4, 5", + "[65] Vanessa Sklyarova, Egor Zakharov, Otmar Hilliges, Michael J Black, and Justus Thies. Text-conditioned generative model of 3d strand-based human hairstyles. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3, 6, 7, 8", + "[66] Xiaokun Sun, Zhenyu Zhang, Ying Tai, Qian Wang, Hao Tang, Zili Yi, and Jian Yang. Barbie: Text to barbie-style 3d avatars. arXiv preprint arXiv:2408.09126, 2024. 7, 1", + "[67] Yusuke Takimoto, Hikari Takehara, Hiroyuki Sato, Zihao Zhu, and Bo Zheng. Dr. hair: Reconstructing scalp-connected hair strands without pre-training via differentiable rendering of line segments. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 3", + "[68] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splattering for efficient 3d content creation. In Int. Conf. Learn. Represent., 2023. 2", + "[69] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2", + "[70] Duotun Wang, Hengyu Meng, Zeyu Cai, Zhijing Shao, Qianxi Liu, Lin Wang, Mingming Fan, Xiaohang Zhan, and Zeyu Wang. Headevolver: Text to head avatars via expressive and attribute-preserving mesh deformation. arXiv preprint arXiv:2403.09326, 2024. 2", + "[71] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In Adv. Neural Inform. Process. Syst., 2024. 2", + "[72] Keyu Wu, Yifan Ye, Lingchen Yang, Hongbo Fu, Kun Zhou, and Youyi Zheng. Neuralhdhair: Automatic high-fidelity hair modeling from a single image using implicit neural representations. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 3", + "[73] Keyu Wu, Lingchen Yang, Zhiyi Kuang, Yao Feng, Xutao Han, Yuefan Shen, Hongbo Fu, Kun Zhou, and Youyi Zheng." + ], + "bbox": [ + 91, + 90, + 485, + 902 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Monohair: High-fidelity hair modeling from a monocular video. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 3", + "[74] Menghua Wu, Hao Zhu, Linjiang Huang, Yi Zhuang, Yuanxun Lu, and Xun Cao. High-fidelity 3d face generation from natural language descriptions. IEEE Conf. Comput. Vis. Pattern Recog., 2023. 2", + "[75] Sijing Wu, Yichao Yan, Yunhao Li, Yuhao Cheng, Wenhan Zhu, Ke Gao, Xiaobo Li, and Guangtao Zhai. Ganhead: Towards generative animatable neural head avatars. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 1", + "[76] Yunjie Wu, Yapeng Meng, Zhipeng Hu, Lincheng Li, Haoqian Wu, Kun Zhou, Weiwei Xu, and Xin Yu. Text-guided 3d face synthesis-from generation to editing. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2", + "[77] Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu Shen, Ruigang Yang, and Xun Cao. Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction. In IEEE Conf. Comput. Vis. Pattern Recog., 2020. 2", + "[78] Lingchen Yang, Zefeng Shi, Youyi Zheng, and Kun Zhou. Dynamic hair modeling from monocular videos using deep neural networks. ACM Trans. Graph., 2019. 3", + "[79] Tarun Yenamandra, Ayush Tewari, Florian Bernard, Hans-Peter Seidel, Mohamed Elgharib, Daniel Cremers, and Christian Theobalt. i3dmm: Deep implicit 3d morphable model of human heads. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 1", + "[80] Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3, 6", + "[81] Taoran Yi, Jiemin Fang, Zanwei Zhou, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Xinggang Wang, and Qi Tian. Gaussiandreamerpro: Text to manipulable 3d gaussians with highly enhanced quality. arXiv preprint arXiv:2406.18462, 2024. 2", + "[82] Cem Yuksel, Scott Schaefer, and John Keyser. Hair meshes. ACM Trans. Graph., 2009. 3, 4", + "[83] Egor Zakharov, Vanessa Sklyarova, Michael Black, Giljoo Nam, Justus Thies, and Otmar Hilliges. Human hair reconstruction with strand-aligned 3d gaussians. In Eur. Conf. Comput. Vis., 2024. 3", + "[84] Hao Zhang, Yao Feng, Peter Kulits, Yandong Wen, Justus Thies, and Michael J. Black. Teca: Text-guided generation and editing of compositional 3d avatars. In International Conference on 3D Vision, 2024. 2, 3, 6, 7", + "[85] Longwen Zhang, Qiwei Qiu, Hongyang Lin, Qixuan Zhang, Cheng Shi, Wei Yang, Ye Shi, Sibei Yang, Lan Xu, and Jingyi Yu. Dreamface: Progressive generation of animatable 3d faces under text guidance. ACM Trans. Graph., 2023. 2", + "[86] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Int. Conf. Comput. Vis., 2023. 2", + "[87] Rui Zhao, Wei Li, Zhipeng Hu, Lincheng Li, Zhengxia Zou, Zhen Xia Shi, and Changjie Fan. Zero-shot text-to-parameter" + ], + "bbox": [ + 516, + 92, + 903, + 902 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13403", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "translation for game character auto-creation. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 2", + "[88] Mingwu Zheng, Hongyu Yang, Di Huang, and Liming Chen. Imface: A nonlinear 3d morphable face model with implicit neural representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022. 1", + "[89] Yujuan Zheng, Zirong Jin, Moran Li, Haibin Huang, Chongyang Ma, Shuguang Cui, and Xiaoguang Han. Hirstep: Transfer synthetic to real using strand and depth maps for single-view 3d hair modeling. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 2, 3", + "[90] Yujian Zheng, Yuda Qiu, Leyang Jin, Chongyang Ma, Haibin Huang, Di Zhang, Pengfei Wan, and Xiaoguang Han. Towards unified 3d hair reconstruction from single-view portraits. arXiv preprint arXiv:2409.16863, 2024.", + "[91] Yi Zhou, Liwen Hu, Jun Xing, Weikai Chen, Han-Wei Kung, Xin Tong, and Hao Li. Hairnet: Single-view hair reconstruction using convolutional neural networks. In *Eur. Conf. Comput. Vis.*, 2018. 3", + "[92] Yuxiao Zhou, Menglei Chai, Alessandro Pepe, Markus Gross, and Thabo Beeler. Groomen: A high-quality generative hair model using hierarchical latent representations. ACM Trans. Graph., 2023. 2, 3", + "[93] Yuxiao Zhou, Menglei Chai, Daoye Wang, Sebastian Winberg, Erroll Wood, Kripasindhu Sarkar, Markus Gross, and Thabo Beeler. Groomcap: High-fidelity prior-free hair capture. arXiv preprint arXiv:2409.00831, 2024. 3", + "[94] Zhen Zhou, Fan Ma, Hehe Fan, and Yi Yang. Headstudio: Text to animatable head avatars with 3d gaussian splatting. In Eur. Conf. Comput. Vis., 2024. 2, 3, 6" + ], + "bbox": [ + 91, + 90, + 482, + 516 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "13404", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/StrandHead_ Text to Hair-Disentangled 3D Head Avatars Using Human-Centric Priors/5c673a06-82fc-4d7f-915c-54daefec29d8_model.json b/2025/StrandHead_ Text to Hair-Disentangled 3D Head Avatars Using Human-Centric Priors/5c673a06-82fc-4d7f-915c-54daefec29d8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..98f19d04184f40fe3010f8dbce83768a0fea857f --- /dev/null +++ b/2025/StrandHead_ Text to Hair-Disentangled 3D Head Avatars Using Human-Centric Priors/5c673a06-82fc-4d7f-915c-54daefec29d8_model.json @@ -0,0 +1,3447 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.21, + 0.131, + 0.79, + 0.177 + ], + "angle": 0, + "content": "StrandHead: Text to Hair-Disentangled 3D Head Avatars Using Human-Centric Priors" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.204, + 0.744, + 0.24 + ], + "angle": 0, + "content": "Xiaokun Sun, Zeyu Cai, Ying Tai, Jian Yang, Zhenyu Zhang* Nanjing University *Corresponding Author" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.242, + 0.734, + 0.274 + ], + "angle": 0, + "content": "xiaokun.sun@smail.nju.edu.cn, caizeyu010612@gmail.com, {yingtai, csjyang}@nju.edu.cn, zhangjesse@foxmail.com" + }, + { + "type": "image", + "bbox": [ + 0.16, + 0.275, + 0.836, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.187, + 0.454, + 0.279, + 0.466 + ], + "angle": 0, + "content": "(a) Input Prompts" + }, + { + "type": "image", + "bbox": [ + 0.309, + 0.453, + 0.832, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.54, + 0.908, + 0.583 + ], + "angle": 0, + "content": "Figure 1. Given (a) input prompts, StrandHead generates (b) realistic 3D head avatars featuring strand-level attributes and (c) 3D hair strands by utilizing human-specific 2D generative priors and 3D hair strand geometric priors. By precisely capturing the internal geometry of hair strands, our approach enables seamless and flexible (d) hairstyle transfer and (e) editing, as well as (f) physics-based simulation." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.594, + 0.327, + 0.61 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.486, + 0.902 + ], + "angle": 0, + "content": "While haircut indicates distinct personality, existing avatar generation methods fail to model practical hair due to the data limitation or entangled representation. We propose StrandHead, a novel text-driven method capable of generating 3D hair strands and disentangled head avatars with strand-level attributes. Instead of using large-scale hair-text paired data for supervision, we demonstrate that realistic hair strands can be generated from prompts by distilling 2D generative models pre-trained on human mesh data. To this end, we propose a meshing approach guided by strand geometry to guarantee the gradient flow from the distillation objective to the neural strand representation. The optimization is then regularized by statistically significant haircut features, leading to stable updating of strands against unreasonable drifting. These employed 2D/3D human-centric priors contribute to text-aligned and realistic 3D strand generation. Extensive experiments show that StrandHead achieves the state-of-the-art performance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.909, + 0.687 + ], + "angle": 0, + "content": "on text to strand generation and disentangled 3D head avatar modeling. The generated 3D hair can be applied on avatars for strand-level editing, as well as implemented in the graphics engine for physical simulation or other applications. Project page: https://xiaokunsun.github.io/StrandHead.github.io/." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.695, + 0.646, + 0.71 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.72, + 0.907, + 0.886 + ], + "angle": 0, + "content": "Creating 3D head avatars is crucial for many applications including digital telepresence, gaming, movies, and AR/VR. Hairstyles reflecting personal characteristics greatly impact the fidelity and realism of digital humans. Traditional methods rely heavily on manual effort, making them time-consuming and labor-intensive. Recent advances [52, 75, 79, 88] achieve automatic 3D head generation with the research paradigm based on supervised learning. However, these methods require costly 3D training data, which significantly limits their generalization capabilities and restricts potential applications." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "With the rapid development of text-driven generation" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13393" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.304 + ], + "angle": 0, + "content": "methods [47, 54], creating 3D head avatars from given prompts without requiring 3D-text datasets becomes possible. Recent studies [1, 12, 84, 94] unlock the potential of 2D diffusion models in modeling 3D head avatars by integrating them with 3D head priors [4, 26, 77]. HumanNorm [17] further improves domain-specific fidelity by fine-tuning diffusion models on high-quality 3D human meshes. While focusing on facial geometry and texture modeling, these methods utilize holistic meshes or NeRF [38] to represent 3D haircuts, failing to capture the internal geometric structure of hair strands, i.e., 3D curves. This limitation not only significantly reduces the realism of the generated avatars but also makes them incompatible with strand-based applications and simulation systems [2, 9, 10]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.306, + 0.483, + 0.518 + ], + "angle": 0, + "content": "To accurately model 3D haircuts, recent efforts achieve strand-level hairstyle reconstruction [55, 64, 89] or generation [13, 92] using VAEs or parametric models. Despite their impressive performance, these methods require constrained multi-view images or manual latent space searching, preventing them from freely creating 3D hair based on user-friendly text. The most recent work, HAAR [65] pioneers text-to-strand by training a text-conditioned hair map diffusion model. However, due to its reliance on large-scale and costly paired data including 9825 haircuts and descriptions, HAAR faces challenges in generating diverse hairstyles outside the training set. Additionally, HAAR overlooks hair texture and geometry that adapts to specific head shapes, further limiting its practical applications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.521, + 0.483, + 0.703 + ], + "angle": 0, + "content": "So can we use powerful human-specific 2D generative priors to create realistic and diverse strand-based hair from text? We address this meaningful and challenging problem by developing StrandHead, a novel framework that generates high-fidelity 3D head avatars with strand-accurate hair from prompts. As illustrated in Fig. 1, the generated head avatars feature diverse and realistic hair strands, enabling seamless and flexible strand-level transfer, editing, and physics-based simulation. Instead of using large-scale hair-prompt paired data, we achieve text-aligned 3D hair strand generation by effectively utilizing human-specific 2D generative priors and 3D hair geometric priors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Specifically, we propose a novel differentiable prismatization algorithm inspired by the cylindrical structure of hair strands. This algorithm can efficiently and differentiably convert strands into watertight prismatic meshes, which enables smooth backpropagation of gradients from 2D diffusion models pre-trained on human mesh data to 3D hair strands using mesh-based differentiable renderers, thereby making it possible to use 2D generative priors for modeling 3D hairstyles. Then we propose two effective losses inspired by the distribution patterns of 3D hair geometric features to further regularize the hair shape. These rich 2D/3D human-centric priors work together to achieve text-aligned and realistic 3D hair strand generation. Extensive experi" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "ments demonstrate that StrandHead outperforms the state-of-the-art (SOTA) methods in both head and hair generation tasks, and supports flexible haircut transfer and editing, as well as physical-based rendering and simulation." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.153, + 0.874, + 0.166 + ], + "angle": 0, + "content": "Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.167, + 0.905, + 0.242 + ], + "angle": 0, + "content": "- We propose StrandHead, a novel framework for generating realistic 3D head avatars with strand-level attributes. To the best of our knowledge, StrandHead is the first work to generate 3D hair strands by distilling human-specific 2D diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.243, + 0.905, + 0.318 + ], + "angle": 0, + "content": "- We propose a differentiable prismatization algorithm that converts hair strands into watertight prismatic meshes. This ensures stable gradient flow from 2D generative priors to 3D hair strand representation, thereby achieving reliable end-to-end strand-based hair optimization." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.318, + 0.905, + 0.379 + ], + "angle": 0, + "content": "- Inspired by statistical 3D hair geometric features, we introduce two simple but solid regularization losses to supervise both local and global hair shapes, enabling reasonable and realistic hairstyle generation." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.167, + 0.905, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.393, + 0.655, + 0.408 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.417, + 0.905, + 0.703 + ], + "angle": 0, + "content": "Text-to-3D General Object Generation. Inspired by the success of text-to-image (T2I) generation [42, 50, 51, 54, 56, 86], many studies have explored using pre-trained vision-language models [49, 54] to achieve text-guided 3D content generation without large-scale 3D-text paired data. Early methods [18, 37, 39, 58, 69] employ the CLIP model [49] to supervise the alignment of 3D representations with prompts. DreamFusion [47] introduces the Score Distillation Sampling (SDS) loss, significantly enhancing the fidelity of generated 3D content by leveraging more powerful pre-trained diffusion models [54]. Subsequent works further advance text-driven 3D generation by improving 3D representations [5, 6, 80], optimization strategies [29, 68, 81], SDS loss [28, 71], and diffusion models [27, 48, 62]. Despite these advancements, current approaches focused on general content generation do not fully leverage the extensive prior knowledge of human heads and hair, limiting their ability to generate realistic, high-quality 3D head avatars with strand-based hair." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.705, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Text-to-3D Head Avatar Generation. CLIPFace [1], T2P [87] and Describe3D [74] pioneer zero-shot text-driven 3D head generation by combining CLIP [49] with 3D parametric head models [4, 8, 26, 77]. With the introduction of SDS loss [47], DreamFace [85] and FaceG2E [76] leverage pre-trained T2I diffusion models to greatly improve generation quality within specific domains. HeadEvolver [70] enhances the expressiveness of head mesh deformations by introducing vector fields, while HeadSculpt [12], and HeadArtist [30] employ DMTet [60] instead of traditional meshes to capture geometric details. HeadStudio [94] incorporates 3DGS [20] to produce realistic and animatable 3D heads. HumanNorm [17], further" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13394" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.131, + 0.089, + 0.442, + 0.143 + ], + "angle": 0, + "content": "
TaskMethodHead-Hair-DecoupledStrand-Based HairGeometry & TextureNo Large-Scale 3D-Text Paired Data
Text-to-Head[12, 17, 30, 94]XX✓✓
[84]✓X✓✓
Text-to-Hair[28, 48, 62, 80]XX✓✓
[65]X✓XX
Text-to-Head-HairOurs✓✓✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.136, + 0.144, + 0.438, + 0.157 + ], + "angle": 0, + "content": "Table 1. Comparison with current related methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.162, + 0.483, + 0.327 + ], + "angle": 0, + "content": "enhances human-specific fidelity by fine-tuning 2D diffusion models on high-quality human mesh data. Despite these advancements, these approaches treat the head and hair as a holistic model, limiting support for downstream applications such as hairstyle transfer and editing. To enable flexible head-hair-disentangled generation, TECA [84] independently represents the head with mesh and the hair with NeRF [38]. However, these methods focus only on the realistic external appearance of hair without modeling its internal geometric structure, restricting their application for strand-level editing and physics-based simulation." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.328, + 0.485, + 0.69 + ], + "angle": 0, + "content": "Strand-Based Hair Creation. Automatic strand-based hair modeling has garnered significant attention from both industry [2, 9, 10] and academia [3, 14, 33, 82], as it enables downstream physics-based rendering and simulation. Such methods can be broadly categorized into two types: (1) reconstructing hair from input images or videos and (2) generating hair based on control conditions. Works on the former one [15, 21, 32, 35, 41, 55, 57, 61, 64, 67, 72, 73, 78, 83, 89-91, 93] relies on classical 3D reconstruction frameworks [20, 36, 38, 45, 59] and incorporates hair-specific features [44] to realize accurate 3D strand reconstruction. However, these methods depend on constrained multi-view images, making hair creation costly and challenging in less controlled settings. Recently, 3D hair generation approaches [13, 92] have emerged using GANs or parametric models, but they struggle with controllable generation. The most relevant work to ours is HAAR [65], which pioneers text-to-strand by training a text-conditioned hair map diffusion model. However, HAAR relies on a large-scale paired dataset including 9825 haircuts and prompts with limited diversity, restricting its ability to generate novel hairstyles beyond the training data. Moreover, HAAR overlooks hair texture and geometry that adapts to specific head shapes, further limiting its application range." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.781 + ], + "angle": 0, + "content": "Compared to previous methods, StrandHead generates 3D heads featuring strand-level hairstyles. Instead of requiring large-scale hair-text paired data, we achieve realistic and diverse 3D hair strand generation by using 2D/3D human-centric priors. We summarize the main differences between our methods and related works in Tab. 1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.794, + 0.182, + 0.809 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.232, + 0.833 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.903 + ], + "angle": 0, + "content": "FLAME [26] is a 3D parametric head model that represents the human head shape, pose, and expression using a compact set of parameters. Given a set of shape parameters \\(\\beta\\), pose parameters \\(\\theta_{\\mathrm{pose}}\\), and expression parameters \\(\\psi_{\\mathrm{exp}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.778, + 0.106 + ], + "angle": 0, + "content": "FLAME can create a 3D head mesh \\(M\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.107, + 0.905, + 0.197 + ], + "angle": 0, + "content": "HumanNorm [17] is a text-driven, high-quality 3D head generation method. In specific, it models the geometry and texture of the human head using DMTet [60] and a texture field. The alignment between the text \\( y \\) and the optimized representation \\( \\theta \\) is achieved using the following SDS loss [47] with T2I diffusion models \\( \\phi \\):" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.198, + 0.905, + 0.231 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\theta} \\mathcal {L} _ {S D S} = \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\epsilon_ {\\phi} \\left(x _ {t}; y, t\\right) - \\epsilon\\right) \\frac {\\partial x}{\\partial \\theta} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.231, + 0.905, + 0.336 + ], + "angle": 0, + "content": "where \\(x = g(\\theta)\\) is the image rendered from \\(\\theta\\) by a differentiable renderer \\(g\\), \\(t\\) is the time step, \\(x_{t} = x + \\epsilon\\) is a noised version of \\(x\\), \\(\\epsilon_{\\phi}(x_{t};y,t)\\) is the denoised image, and \\(w(t)\\) is a weighting function. HumanNorm employs human-specific diffusion models fine-tuned on high-quality human mesh data to replace the general diffusion model, enhancing the fidelity of the generated human heads." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.337, + 0.906, + 0.475 + ], + "angle": 0, + "content": "Neural Scalp Textures (NST) [55] is an efficient hair representation where each pixel stores a feature vector conveying the shape information of a single strand at the corresponding scalp location. Using a pre-trained hair strand generator [55] \\( G \\), the low-dimensional 2D neural scalp texture \\( T \\) can be decoded into high-dimensional 3D strand polylines \\( S = \\{s^i\\}_{i=1}^{N_s} \\), where \\( N_s \\) is the number of strands and \\( s^i = \\{p_j^i\\}_{j=1}^{N_p} \\) consists of \\( N_p \\) 3D points. This process is formulated as follows: \\( S = G(T) \\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.484, + 0.625, + 0.498 + ], + "angle": 0, + "content": "3.2. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.906, + 0.643 + ], + "angle": 0, + "content": "Given a text prompt, StrandHead aims to create a realistic 3D head avatar with strand-based hair without relying on large-scale 3D-text paired data. Fig. 2 provides an overview of our three-stage pipeline. First, we generate a FLAME-aligned 3D bald head for accurate hair initialization using human-specific diffusion models (Sec. 3.3). In the subsequent stages, we model the reasonable geometry and realistic texture of 3D hair strands by leveraging rich 2D/3D human-centric priors (Sec. 3.4 and Sec. 3.5)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.651, + 0.724, + 0.666 + ], + "angle": 0, + "content": "3.3. Bald Head Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.906, + 0.81 + ], + "angle": 0, + "content": "To achieve accurate hair initialization and optimize 3D hair strands in subsequent steps using human-specific 2D generative priors, we first need to obtain a reasonable and semantic-aligned 3D bald head from the text (Fig. 2-(a)). To achieve this goal, we enhance HumanNorm [17] by incorporating an evolving FLAME [26] model while optimizing the bald head DMTet \\(\\theta_{\\mathrm{h}}\\). This provides accurate semantic information and prevents unnatural geometry. Please refer to the Supp. Mat. for details on bald head generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.759, + 0.835 + ], + "angle": 0, + "content": "3.4. Hair Geometry Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "With a FLAME-aligned bald head, we can generate 3D hair in the scalp area. Specifically, we first initialize the haircut based on hairstyle descriptions, and then further sculpt the 3D hair shape utilizing 2D/3D human-centric priors." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13395" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.135, + 0.09, + 0.868, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.325, + 0.907, + 0.382 + ], + "angle": 0, + "content": "Figure 2. StrandHead includes three stages: (a) We first create a FLAME-aligned 3D bald head using the improved HumanNorm [17]. (b) Next, we introduce a differentiable prismatization algorithm to enable human-specific geometry-aware 2D diffusion models to supervise hair shape modeling. Additionally, two losses inspired by 3D hair geometric priors are applied to further regularize the hair geometry. (c) Finally, we use a human-specific normal-conditioned 2D diffusion model to generate lifelike hair textures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.386, + 0.484, + 0.462 + ], + "angle": 0, + "content": "Hair Initialization. To achieve reasonable and diverse hair initialization, we first utilize ChatGPT [43] to select the 20 most representative hairstyles from the USC-HairSalon Dataset [15]. We then optimize neural scalp textures to fit the selected hairstyle using the following loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.463, + 0.484, + 0.519 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {f i t}} = \\sum_ {i = 1} ^ {N _ {s}} \\sum_ {j = 1} ^ {N _ {p}} \\| \\hat {p} _ {j} ^ {i} - p _ {j} ^ {i} \\| _ {2} + \\lambda_ {\\mathrm {o r i}} \\left(1 - \\hat {o} _ {j} ^ {i} \\cdot o _ {j} ^ {i}\\right) + \\lambda_ {\\mathrm {c u r}} \\| \\hat {c} _ {j} ^ {i} - c _ {j} ^ {i} \\| _ {1}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.519, + 0.484, + 0.659 + ], + "angle": 0, + "content": "where \\(\\hat{p}_j^i\\) and \\(p_j^i\\) are the position of the \\(j\\)-th point on the \\(i\\)-th polyline of the GT and generated hair, respectively, and \\(\\hat{o}_j^i\\), \\(o_j^i\\), \\(\\hat{c}_j^i\\), and \\(c_j^i\\) denote their orientation and curvature, respectively. Here, orientation \\(o_j^i = (p_{j+1}^i - p_j^i) / \\|p_{j+1}^i - p_j^i\\|_2\\) represents the direction of change in strand position, and curvature \\(c_j^i = \\|o_j^i - o_{j-1}^i\\|_2\\) denote the rate of change in strand orientation. Given a hair description, we select an optimal pre-trained NST as initialization and further optimize it in subsequent stages." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Discussion on Strand-Based Differentiable Rendering. With reasonable hair initialization, the next step is to leverage human-specific 2D diffusion models to further sculpt the hair geometry. However, due to the lack of a stable strand-based differentiable renderer, it is very challenging to optimize hair shape using 2D generative priors like other SDS-based methods. Besides, we expect hair strands to produce mesh-style smooth geometry and texture maps to fully exploit the powerful generative capabilities of 2D diffusion models pre-trained on human mesh data. A feasible alternative is to first differentiably convert the strands into quad meshes [82], similar to NeuralHaircut [64], and then leverage SDS-based mesh optimization frameworks to model hair strands. However, this non-watertight stripe-like mesh easily produces ambiguous normal maps or excessively thin sides (see the zoom image of Fig. 3-(c)), which significantly" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.385, + 0.891, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.608, + 0.491, + 0.813, + 0.5 + ], + "angle": 0, + "content": "(b) A strand-to-mesh conversion example" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.5, + 0.88, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.614, + 0.823, + 0.625 + ], + "angle": 0, + "content": "(c) Differentiable prismatization's advantages" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.626, + 0.906, + 0.709 + ], + "angle": 0, + "content": "Figure 3. The differentiable prismatization algorithm's (a) gradient backpropagation process, (b) a strand-to-mesh conversion example and (c) advantages over NeuralHaircut [64]. Non-watertight quad meshes can easily produce ambiguous normal maps, which significantly reduce the stability of hair shape modeling (see the drifting hair highlighted by the oval dotted box in (c))." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.906, + 0.779 + ], + "angle": 0, + "content": "reduce the optimization stability (see the drifting hair highlighted by the oval dotted box in Fig. 3-(c)). Therefore, a differentiable strand-to-mesh conversion method with superior optimization properties is urgently needed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Differentiable Prismatization (DP). To this end, we propose a novel differentiable prismatization algorithm inspired by the internal structure of 3D hair (i.e., hair fiber is a dielectric cylinder covered with tilted scales and with a pigmented interior [19, 34]). This algorithm can efficiently and differentiably convert strands into watertight prismatic meshes with arbitrary thickness and lateral edges to approximate the cylindrical structure of the hair. This enables" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13396" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.183 + ], + "angle": 0, + "content": "smooth backpropagation of gradients from the SDS loss to the 3D strand representation using mesh-based differentiable renderers [22], thereby making it possible to fully exploit 2D generative priors distilled from high-fidelity human mesh data for modeling realistic hair strands. Fig. 3-(a) visualizes the complete gradient backpropagation process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.185, + 0.483, + 0.35 + ], + "angle": 0, + "content": "In specific, given a hair strand \\(s\\), our DP converts it into a watertight prismatic mesh with \\(K\\) lateral edges and radius \\(R\\) through the following five steps: (1) Compute the Initial Normal Vector. (2) Generate \\(K\\) Rotated Normals. (3) Translate to Form Lateral Edges. (4) Construct Lateral Faces. (5) Construct Top and Bottom Faces. An example of this conversion is shown in Fig. 3-(b). Unlike quad meshes [64], our DP effectively avoids ambiguous normals, enabling smooth gradient backpropagation and hair strand optimization, as shown in Fig. 3-(c). Further algorithm details are provided in the Supp. Mat." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.353, + 0.484, + 0.414 + ], + "angle": 0, + "content": "With the help of our proposed differentiable prismatization algorithm, we optimize the neural scalp texture \\( T \\) by utilizing rich human-specific 2D generative priors through the following SDS losses (Fig. 2-(b)):" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.419, + 0.483, + 0.452 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {T} \\mathcal {L} _ {S D S} ^ {\\mathrm {h n}} = \\mathbb {E} _ {t, \\epsilon} \\left[ \\left(\\epsilon_ {\\phi_ {\\mathrm {h n}}} \\left(n _ {t} ^ {\\mathrm {h + s}}; y _ {\\mathrm {h + s}}, t\\right) - \\epsilon\\right) \\frac {\\partial n ^ {\\mathrm {h + s}}}{\\partial T} \\right], \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.455, + 0.483, + 0.489 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {T} \\mathcal {L} _ {S D S} ^ {\\mathrm {h d}} = \\mathbb {E} _ {t, \\epsilon} \\left[ \\left(\\epsilon_ {\\phi_ {\\mathrm {h d}}} \\left(d _ {t} ^ {\\mathrm {h + s}}; y _ {\\mathrm {h + s}}, t\\right) - \\epsilon\\right) \\frac {\\partial d ^ {\\mathrm {h + s}}}{\\partial T} \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.504, + 0.483, + 0.608 + ], + "angle": 0, + "content": "where \\(\\phi_{\\mathrm{hn}}\\) and \\(\\phi_{\\mathrm{hd}}\\) are the human-specific normal-adapted and depth-adapted diffusion models from Human-Norm [17], \\(n^{\\mathrm{h + s}}\\) and \\(d^{\\mathrm{h + s}}\\) are the rendered normal and depth maps of the head DMTet \\(\\theta_{\\mathrm{h}}\\) with hair prismatic mesh \\(M_p = DP(G(T))\\), respectively. Here, \\(DP(\\cdot)\\) denotes the differentiable prismatization operation function, \\(y_{\\mathrm{h + s}}\\) is the full text description of the head with hair." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.483, + 0.702 + ], + "angle": 0, + "content": "However, due to the excessive flexibility of hair strands, relying solely on the SDS loss leads to unnatural strand orientations and unrealistic hairstyles (Fig. 9). Furthermore, the lack of constrained multi-view images prevents us from leveraging hair-specific features [44] to provide powerful supervision as in previous works [55]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Prior-Driven Losses. To address this issue, we propose two straightforward yet robust prior-driven losses based on observations of the 3D hair strand orientation and curvature to ensure the rationality of the generated hair strands. Specifically, we observe two geometric properties of hair: (1) neighboring strand orientations are highly consistent, and (2) strand curvature is strongly and positively correlated with the curliness of the hairstyle. To validate the properties, we calculate the cosine similarity of adjacent hair strand orientations \\( CS_{\\mathrm{ori}} \\) and the average curvature \\( C_{\\mathrm{mean}} \\) of 343 hairstyles in the USC-HairSalon dataset [15] using the following equations and visualize their distributions and some examples in Fig. 4:" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.089, + 0.7, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.191, + 0.704, + 0.201 + ], + "angle": 0, + "content": "(a) Distribution of Haircuts' \\( CS_{\\mathrm{ori}} \\)" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.089, + 0.883, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.725, + 0.191, + 0.889, + 0.2 + ], + "angle": 0, + "content": "(b) Distribution of Haircuts' \\( C_{\\text{mean}} \\)" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.203, + 0.623, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.264, + 0.623, + 0.271 + ], + "angle": 0, + "content": "\\(CS_{\\mathrm{ori}} = 0.9949\\) \\(C_{\\mathrm{mean}} = 0.0\\)" + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.204, + 0.738, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.264, + 0.748, + 0.272 + ], + "angle": 0, + "content": "1 \\(CS_{\\mathrm{ori}} = 0.9868\\) \\(C_{\\mathrm{mean}} = 0.0290\\)" + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.203, + 0.816, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.745, + 0.264, + 0.885, + 0.272 + ], + "angle": 0, + "content": "\\(C S_{\\mathrm{ori}} = 0.9605\\) \\(C_{\\mathrm{mean}} = 0.0493\\)" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.272, + 0.597, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.333, + 0.584, + 0.339 + ], + "angle": 0, + "content": "\\(CS_{\\mathrm{ori}} = 0.9265\\)" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.272, + 0.657, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.6, + 0.333, + 0.651, + 0.339 + ], + "angle": 0, + "content": "\\(C_{\\text{mean}} = 0.0598\\)" + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.272, + 0.738, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.666, + 0.333, + 0.738, + 0.339 + ], + "angle": 0, + "content": "\\(CS_{\\mathrm{ori}} = 0.9719\\) \\(C_{\\mathrm{mean}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.272, + 0.828, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.74, + 0.333, + 0.885, + 0.34 + ], + "angle": 0, + "content": "\\(= 0.0870\\) \\(CS_{\\mathrm{ori}} = 0.9297\\) \\(C_{\\mathrm{mean}} = 0.1059\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.577, + 0.342, + 0.827, + 0.352 + ], + "angle": 0, + "content": "(c) Some hairstyles from the USC-HairSalon dataset" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.353, + 0.906, + 0.394 + ], + "angle": 0, + "content": "Figure 4. Observation of hair geometric features: (1) neighboring strand orientations are highly consistent. (2) strand curvature is strongly and positively related to the haircut curliness." + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.399, + 0.906, + 0.443 + ], + "angle": 0, + "content": "\\[\nC S _ {\\text {o r i}} = \\frac {1}{N _ {s} N _ {p}} \\sum_ {i = 1} ^ {N _ {s}} \\sum_ {j = 1} ^ {N _ {p}} \\sum_ {k \\in A (i)} \\frac {o _ {j} ^ {i} \\cdot o _ {j} ^ {k}}{| A (i) |}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.448, + 0.905, + 0.49 + ], + "angle": 0, + "content": "\\[\nC _ {\\text {m e a n}} = \\frac {1}{N _ {s} N _ {p}} \\sum_ {i = 1} ^ {N _ {s}} \\sum_ {j = 1} ^ {N _ {p}} c _ {j} ^ {i}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.499, + 0.904, + 0.529 + ], + "angle": 0, + "content": "where \\(A(i)\\) is the set of adjacent strands for the \\(i\\)-th strand, and \\(|A(i)|\\) is the number of neighboring strands." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.905, + 0.633 + ], + "angle": 0, + "content": "As illustrated, over \\(95\\%\\) of hairstyles exhibit an orientation similarity above 0.9, and the average curvature is significantly positively related to curliness, which successfully validates the properties. Based on these observations, we introduce orientation consistency loss and curvature regularization to guide the local and global strand shapes, formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.637, + 0.905, + 0.652 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {o r i}} = 1 - C S _ {\\mathrm {o r i}}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.656, + 0.905, + 0.672 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c u r}} = \\left\\| C _ {\\text {m e a n}} - C _ {\\text {t a r g e t}} \\right\\| _ {1}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.681, + 0.905, + 0.755 + ], + "angle": 0, + "content": "where \\( C_{\\mathrm{target}} \\) represents the target average curvature set according to the input hair description. These two losses, inspired by 3D hair geometry priors, regularize the hair shape by supervising the consistency of orientations between adjacent strands and the overall hairstyle's curvature." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.756, + 0.906, + 0.861 + ], + "angle": 0, + "content": "Additionally, we introduce a series of losses \\(\\mathcal{L}_{bbox}\\), \\(\\mathcal{L}_{face}\\), and \\(\\mathcal{L}_{colli}\\), which prevent the hair from exceeding the bounding box, obscuring the face, and colliding with the head, respectively. These losses can further improve the rationality of hair geometry. Their details are provided in the Supp. Mat. The final loss function for optimizing hair shape is expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.866, + 0.905, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {h a i r - g e o}} = \\mathcal {L} _ {S D S} ^ {\\mathrm {h d}} + \\lambda_ {S D S} ^ {\\mathrm {h n}} \\mathcal {L} _ {S D S} ^ {\\mathrm {h n}} + \\lambda_ {\\mathrm {o r i}} \\mathcal {L} _ {\\mathrm {o r i}} + \\lambda_ {\\mathrm {c u r}} \\mathcal {L} _ {\\mathrm {c u r}} \\\\ + \\lambda_ {\\text {b b o x}} \\mathcal {L} _ {\\text {b b o x}} + \\lambda_ {\\text {f a c e}} \\mathcal {L} _ {\\text {f a c e}} + \\lambda_ {\\text {c o l l i}} \\mathcal {L} _ {\\text {c o l l i}}. \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13397" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.148, + 0.089, + 0.252, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.139, + 0.15, + 0.25, + 0.174 + ], + "angle": 0, + "content": "\"An elderly Caucasian grandmother with thin lips and a straight, short silver bob.\"" + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.091, + 0.382, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.273, + 0.153, + 0.38, + 0.169 + ], + "angle": 0, + "content": "\"Brad Pitt with medium-length, tousled blonde hair.\"" + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.091, + 0.49, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.394, + 0.153, + 0.495, + 0.169 + ], + "angle": 0, + "content": "\"Chris Hemsworth with side-swept sandy blonde hair.\"" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.091, + 0.603, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.153, + 0.612, + 0.169 + ], + "angle": 0, + "content": "Lionel Messi with a blonde mohawk hairstyle." + }, + { + "type": "image", + "bbox": [ + 0.618, + 0.091, + 0.727, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.626, + 0.153, + 0.721, + 0.169 + ], + "angle": 0, + "content": "\"Scarlet Johansson with a textured blonde bob.\"" + }, + { + "type": "image", + "bbox": [ + 0.751, + 0.09, + 0.852, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.74, + 0.15, + 0.855, + 0.174 + ], + "angle": 0, + "content": "\"A mature European woman with a straight nose and medium-length, straight auburn hair.\"" + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.179, + 0.246, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.174, + 0.241, + 0.267, + 0.256 + ], + "angle": 0, + "content": "Beyonce with voluminous, curly honey blonde hair." + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.179, + 0.298, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.179, + 0.407, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.321, + 0.241, + 0.403, + 0.256 + ], + "angle": 0, + "content": "Dwayne Johnson with a short black Afro." + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.179, + 0.499, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.424, + 0.241, + 0.576, + 0.256 + ], + "angle": 0, + "content": "\"A beautiful Caucasian girl with delicate features and long silky golden blonde hair.\"" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.179, + 0.574, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.179, + 0.64, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.587, + 0.241, + 0.702, + 0.256 + ], + "angle": 0, + "content": "A confident Hispanic man with a beard and a curly black afro." + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.177, + 0.781, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.785, + 0.177, + 0.852, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.723, + 0.237, + 0.842, + 0.261 + ], + "angle": 0, + "content": "\"A charming Caucasian woman with a straight nose and shoulder-length curly sleek blonde hair.\"" + }, + { + "type": "image", + "bbox": [ + 0.148, + 0.271, + 0.256, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.15, + 0.419, + 0.256, + 0.428 + ], + "angle": 0, + "content": "\"A curly dark blonde haircut.\"" + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.272, + 0.307, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.308, + 0.271, + 0.405, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.319, + 0.415, + 0.396, + 0.431 + ], + "angle": 0, + "content": "\"Medium-length, straight auburn hair.\"" + }, + { + "type": "image", + "bbox": [ + 0.428, + 0.271, + 0.496, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.427, + 0.33, + 0.5, + 0.345 + ], + "angle": 0, + "content": "\"Long way chestnut brown hair.\"" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.271, + 0.578, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.33, + 0.577, + 0.344 + ], + "angle": 0, + "content": "\"A short, wavy brown bob.\"" + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.271, + 0.661, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.601, + 0.33, + 0.666, + 0.345 + ], + "angle": 0, + "content": "\"Long silky golden blonde hair.\"" + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.27, + 0.738, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.68, + 0.33, + 0.75, + 0.345 + ], + "angle": 0, + "content": "\"A light brown mohawk hairstyle.\"" + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.269, + 0.829, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.771, + 0.33, + 0.836, + 0.345 + ], + "angle": 0, + "content": "\"Side-swept sandy blonde hair.\"" + }, + { + "type": "image", + "bbox": [ + 0.433, + 0.352, + 0.49, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.427, + 0.418, + 0.5, + 0.428 + ], + "angle": 0, + "content": "\"A curly black afro.\"" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.354, + 0.577, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.416, + 0.591, + 0.431 + ], + "angle": 0, + "content": "Medium-length wavy caramel-colored hair." + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.353, + 0.657, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.416, + 0.658, + 0.431 + ], + "angle": 0, + "content": "\"A side-swept silver haircut.\"" + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.354, + 0.744, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.685, + 0.415, + 0.746, + 0.431 + ], + "angle": 0, + "content": "\"A sleek long black hairstyle.\"" + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.355, + 0.833, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.766, + 0.416, + 0.841, + 0.431 + ], + "angle": 0, + "content": "\"Medium-length, tousled blonde hair.\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.433, + 0.907, + 0.488 + ], + "angle": 0, + "content": "Figure 5. Examples of high-fidelity and diverse 3D heads and strand-accurate haircuts generated by our method. The upper visualization includes rendered color and normal maps of the head and hair prismatic meshes. The lower visualization shows the physics-based hair strand rendering result using Blender [10]. For better strand-based visualization, we interpolate generated hair to approximately 10,000 strands and apply a consistent appearance. Please zoom in for detailed views, and refer to the Supp. Mat. for video demonstrations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.495, + 0.318, + 0.509 + ], + "angle": 0, + "content": "3.5. Hair Texture Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.517, + 0.484, + 0.563 + ], + "angle": 0, + "content": "Next, we fix the optimized hair strand geometry and model the realistic hair strand texture under the supervision of the SDS loss as follows (Fig. 2-(c)):" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.57, + 0.483, + 0.604 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\psi_ {s}} \\mathcal {L} _ {S D S} ^ {\\mathrm {h c}} = \\mathbb {E} _ {t, \\epsilon} \\left(\\epsilon_ {\\phi_ {\\mathrm {h c}}} \\left(c _ {t} ^ {\\mathrm {h + s}}; n ^ {\\mathrm {h + s}}, y _ {\\mathrm {h + s}}, t\\right) - \\epsilon\\right) \\frac {\\partial c ^ {\\mathrm {h + s}}}{\\partial \\psi_ {\\mathrm {s}}}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.61, + 0.484, + 0.716 + ], + "angle": 0, + "content": "where \\(\\phi_{\\mathrm{hc}}\\) is the human-specific normal-conditioned diffusion model from HumanNorm [17], \\(c^{\\mathrm{h + s}}\\) is the rendered color map of the head texture field \\(\\psi_{\\mathrm{h}}\\) and hair strand texture field \\(\\psi_{\\mathrm{s}}\\). Since the vanilla SDS loss often leads to color oversaturation, we replace it with the following MSDS loss [17] to further enhance the texture's realism in later iterations of texture optimization:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.723, + 0.483, + 0.756 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\psi_ {s}} \\mathcal {L} _ {M S D S} ^ {\\mathrm {h c}} = \\mathbb {E} _ {t, \\epsilon} \\left(H \\left(c _ {t} ^ {\\mathrm {h + s}}; n ^ {\\mathrm {h + s}}, y _ {\\mathrm {h + s}}, t\\right) - \\epsilon\\right) \\frac {\\partial c ^ {\\mathrm {h + s}}}{\\partial \\psi_ {s}}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.764, + 0.484, + 0.839 + ], + "angle": 0, + "content": "where \\( H(\\cdot) \\) denotes the multi-step operation function. Additionally, we propose a strand-aware texture field that models orientation-dependent texture to better generate high-frequency color variations. Please refer to the Supp. Mat. for more their details." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.308, + 0.866 + ], + "angle": 0, + "content": "3.6. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Our method is implemented using PyTorch [46] and ThreeStudio [11]. Experiments are conducted on an Ubuntu" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.496, + 0.907, + 0.572 + ], + "angle": 0, + "content": "server equipped with A6000 GPUs. Starting with a prompt such as \"A man with brown hair\", we first generate a bald head using \"A bald man\". Then, we generate the hair based on \"A man with brown hair\". Further implementation details are available in the Supp. Mat." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.586, + 0.646, + 0.603 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.906, + 0.687 + ], + "angle": 0, + "content": "Some examples of 3D head models with strand-based hair generated by StrandHead are shown in Fig. 5. Due to space limitations, only the key experimental settings and results are presented here. For more results, evaluations, and details, please refer to our Supp. Mat." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.697, + 0.722, + 0.714 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.856 + ], + "angle": 0, + "content": "Baselines. We compare StrandHead with the SOTA methods for head avatar and haircut generation. The baseline methods for head generation include text-to-holistic-head works (HeadArtist [30], HumanNorm [17], and HeadStudio [94]) and text-to-decoupled-head work (TECA [84]). For text-to-hair generation, we consider general text-to-3D methods (MVDream [62], GaussianDreamer [80], LucidDreamer [28], and RichDreamer [48]), text-to-strand method (HAAR [65]), and TECA [84]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Dataset Construction. To create the dataset, we use ChatGPT[43] to randomly generate 30 text descriptions of heads with hair. These include 30 descriptions for assessing" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13398" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.131, + 0.089, + 0.865, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.3, + 0.907, + 0.342 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparisons with the SOTA methods. Since TECA [84] uses the vanilla NeRF to represent hair, rendering normals is not supported. HAAR [65] generates only the geometry of hair strands, so we first convert the strands into prismatic meshes using differentiable prismatization and then utilize TEXTure [53] to generate texture for visualization and comparison." + }, + { + "type": "image", + "bbox": [ + 0.126, + 0.346, + 0.442, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.537, + 0.483, + 0.58 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison with HAAR [65]. HAAR, which does not model heads, often produces unreasonable hair-head collisions (highlighted in the black box)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.596, + 0.483, + 0.626 + ], + "angle": 0, + "content": "head generation and 30 for evaluating hair generation. The full list of prompts can be found in the Supp. Mat." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Evaluation Metrics. Current text-to-3D methods typically use CLIP-based metrics to assess text-image alignment and output quality. However, research [16, 31] shows that these metrics struggle to capture fine-grained alignment between 3D content and input prompts, as validated by our experiments in the Supp. Mat. To address this limitation, we draw inspiration from Progressive3D [7] and Barbie [66], using BLIP-VQA [23, 24] and BLIP2-VQA [23, 25] for evaluation. Specifically, we convert each prompt into two questions to separately verify head and hair. We then feed the rendered image of the generated 3D content into the VQA model, asking each question in sequence and using the probability of a “yes” response as our evaluation metric. For example, the head prompt “A man with black hair” is transformed into “Is the person in the picture a man?” and “Does the person in the picture have black hair?” The hair prompt “Black hair” becomes “Is the object in the picture black hair?” Finally, we conduct a user study by randomly" + }, + { + "type": "table", + "bbox": [ + 0.556, + 0.347, + 0.865, + 0.482 + ], + "angle": 0, + "content": "
MethodBLIP-VQA ↑BLIP2-VQA ↑GQP ↑TAP ↑
HeadArtist0.76670.96671.002.33
HeadStudio0.78330.88333.333.67
HumanNorm0.70000.95007.677.67
TECA0.73330.950034.3328.33
StrandHead (Ours)0.85000.966753.6758.00
MVDream0.90000.833324.6720.00
GaussianDreamer0.33330.36675.333.00
LucidDreamer0.80000.93335.335.00
RichDreamer0.83330.76674.005.67
TECA0.70000.76671.673.67
HAAR0.63330.20001.332.33
StrandHead (Ours)0.90000.900057.6760.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.484, + 0.907, + 0.54 + ], + "angle": 0, + "content": "Table 2. Quantitative comparisons with the SOTA methods. The best and second-best results are highlighted in **bolded** and **underlined**, respectively. GQP: generation quality preference (\\%), TAP: text-image alignment preference (\\%)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.546, + 0.907, + 0.591 + ], + "angle": 0, + "content": "selecting 10 generated examples and asking 30 volunteers to assess (1) generation quality and (2) text-image alignment, and select the best methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.603, + 0.809, + 0.619 + ], + "angle": 0, + "content": "4.2. Comparisons of Head Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.625, + 0.907, + 0.822 + ], + "angle": 0, + "content": "As illustrated in the upper section of Tab. 2, StrandHead outperforms all comparison methods across every evaluation metric. The qualitative results on the left side of Fig. 6 further highlight the advantages of our method. Compared to other head generation methods, the head avatars generated by StrandHead exhibit not only more refined facial geometry and texture details but also strand-accurate hair with a plausible appearance that integrates seamlessly with physics-based simulation systems. To the best of our knowledge, StrandHead is the first head generation framework that realizes strand-level hair modeling, a capability that holds substantial potential for advancing human-centric 3D AIGC applications in the industry." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.804, + 0.849 + ], + "angle": 0, + "content": "4.3. Comparisons of Hair Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "As shown in the lower part of Tab. 2, our approach surpasses other methods across most evaluation metrics, and is only slightly lower than LucidDreamer (the latest open-source" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13399" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.127, + 0.09, + 0.45, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.357, + 0.416, + 0.367 + ], + "angle": 0, + "content": "(b) Generated strand-based hair under different bald heads" + }, + { + "type": "image_caption", + "bbox": [ + 0.112, + 0.369, + 0.461, + 0.383 + ], + "angle": 0, + "content": "Figure 8. Analysis of human-specific 2D diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.394, + 0.483, + 0.5 + ], + "angle": 0, + "content": "text-to-3D method) on BLIP2-VQA, ranking the second. The qualitative comparisons on the right side of Fig. 6 display sample generation results. Compared to general text-to-3D methods, our approach more precisely captures the internal structure of hair strands, resulting in realistic hair-cut geometry and appearance without incorporating incorrect content such as parts of the human head." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.502, + 0.484, + 0.623 + ], + "angle": 0, + "content": "Compared with HAAR [65], StrandHead relies on robust 2D/3D human-centric priors rather than large-scale paired datasets to supervise hair generation. This not only allows it to generate hairstyles that are uncommon in the dataset (see slicked-back and side-swept hairstyles in Fig. 7), but also models hair texture or geometry that adapts to specific head shapes (Fig. 8-(b)), thereby avoiding unnatural hairhead collisions (see the black box in Fig. 7)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.636, + 0.245, + 0.652 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Effectiveness of Human-Specific 2D Generative Priors. We demonstrate the importance of human-specific 2D generative priors from two aspects: (1) As shown in Fig. 8-(a), ignoring 2D supervision fails to generate meaningful hair. While a general 2D diffusion model can slightly improve modeling performance, it still falls short of producing high-quality results. Only by incorporating human-specific 2D generative priors can one achieve realistic textures and reasonable shapes for hairstyles. (2) Fig. 8-(b) illustrates generated hair under varying bald head conditions while maintaining a fixed initial hairstyle and hair prompt. The hairstyles created under the supervision of human-specific 2D diffusion models exhibit clear geometric and textural variations that adapt to specific bald heads. This strongly demonstrates the necessity of considering the bald head for hair strand generation and highlights the advantages of" + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.089, + 0.692, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.096, + 0.871, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.198, + 0.905, + 0.224 + ], + "angle": 0, + "content": "Figure 9. Ablation study on (a) orientation consistency loss \\(\\mathcal{L}_{\\mathrm{ori}}\\) and (b) curvature regularization loss \\(\\mathcal{L}_{\\mathrm{cur}}\\)." + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.224, + 0.696, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.224, + 0.892, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.272, + 0.783, + 0.284 + ], + "angle": 0, + "content": "Figure 10. Failure cases." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.288, + 0.714, + 0.302 + ], + "angle": 0, + "content": "StrandHead over HAAR [65]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.302, + 0.907, + 0.424 + ], + "angle": 0, + "content": "Effectiveness of Prior-Driven Losses. Fig. 9-(a) demonstrates that incorporating \\(\\mathcal{L}_{\\mathrm{ori}}\\) provides strong guidance for generating natural hair by supervising the consistency of orientations between adjacent strands. As shown in Fig. 9-(b), our \\(\\mathcal{L}_{\\mathrm{cur}}\\) preserves the desired curliness by constraining the overall hairstyle's curvature. In summary, these two losses, inspired by 3D hair geometry priors, effectively enhance the realism and rationality of the generated haircut." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.438, + 0.634, + 0.453 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.907, + 0.765 + ], + "angle": 0, + "content": "In this paper, a novel framework StrandHead is proposed to generate realistic hair-disentangled 3D head avatars using text prompts. With a series of 2D generative models pre-trained on human-centric data, StrandHead generates strand-based hair requiring no large-scale text-hair paired data for supervision. This goal is achieved mainly by two approaches: (1) proposing a novel differentiable prismatization method that transforms the hair strands into prismatic meshes, and (2) using the prior-driven losses inspired by the observation of real-world haircut and the statistical features of human-crafted 3D hair data. The former one produces watertight geometry more similar to the original hair strands, boosting a stable optimization with significantly less normal ambiguity. Besides, the latter one well constrains the optimization with reliable guidance to generate reasonable 3D haircuts. Extensive experiments demonstrate that StrandHead obtains the state-of-the-art performance on 3D head avatars and haircuts generation. The generated results are easily implemented in industrial software to produce physical simulation and high-fidelity rendering." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Limitations and Future Work. Due to the limited representation capabilities of the strand generator, StrandHead is unable to create highly complex 3D hairstyles (e.g., dreadlocks and ponytails in Fig. 10). Additionally, although the SDS-based method effectively leverages 2D generative priors, its high computational cost restricts practical application. In future work, we aim to address these challenges by incorporating richer hairstyle datasets and exploring deeper hair priors." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13400" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.122 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported by the National Science Foundation of China (62376121)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.152, + 0.188, + 0.168 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.177, + 0.484, + 0.219 + ], + "angle": 0, + "content": "[1] Shivangi Aneja, Justus Thies, Angela Dai, and Matthias Nießner. Clipface: Text-guided editing of textured 3d morphable models. ACM SIGGRAIPH, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.221, + 0.484, + 0.262 + ], + "angle": 0, + "content": "[2] Inc. Autodesk. Autodesk maya - 3d animation and modeling software. https://www.autodesk.com/products/maya, 2024.2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.264, + 0.484, + 0.305 + ], + "angle": 0, + "content": "[3] Gaurav Bhokare, Eisen Montalvo, Elie Diaz, and Cem Yuksel. Real-time hair rendering with hair meshes. In ACM SIGGRAIPH, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.309, + 0.483, + 0.35 + ], + "angle": 0, + "content": "[4] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.353, + 0.483, + 0.407 + ], + "angle": 0, + "content": "[5] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Int. Conf. Comput. Vis., 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.41, + 0.484, + 0.452 + ], + "angle": 0, + "content": "[6] Zilong Chen, Feng Wang, and Huaping Liu. Text-to-3d using gaussian splatting. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.455, + 0.484, + 0.523 + ], + "angle": 0, + "content": "[7] Xinhua Cheng, Tianyu Yang, Jianan Wang, Yu Li, Lei Zhang, Jian Zhang, and Li Yuan. Progressive3d: Progressively local editing for text-to-3d content creation with complex semantic prompts. In Int. Conf. Learn. Represent., 2023. 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.526, + 0.483, + 0.568 + ], + "angle": 0, + "content": "[8] Hang Dai, Nick Pears, William Smith, and Christian Duncan. Statistical modeling of craniofacial shape and texture. Int. J. Comput. Vis., 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.571, + 0.483, + 0.598 + ], + "angle": 0, + "content": "[9] Inc. Epic Games. Unreal engine - real-time 3d creation tool. https://www.unrealengine.com/, 2024.2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.601, + 0.483, + 0.64 + ], + "angle": 0, + "content": "[10] Blender Foundation. Blender - a 3d modelling and rendering software. https://www.blender.org/, 2024.2,3,6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.644, + 0.483, + 0.727 + ], + "angle": 0, + "content": "[11] Yuan-Chen Guo, Ying-Tian Liu, Ruizhi Shao, Christian Laforte, Vikram Voleti, Guan Luo, Chia-Hao Chen, Zixin Zou, Chen Wang, Yan-Pei Cao, and Song-Hai Zhang. threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio, 2023.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.73, + 0.483, + 0.785 + ], + "angle": 0, + "content": "[12] Xiaoping Han, Yukang Cao, K. Han, Xiatian Zhu, Jiankang Deng, Yi-Zhe Song, Tao Xiang, and Kwan-Yee K. Wong. Headsculpt: Crafting 3d head avatars with text. In Adv. Neural Inform. Process. Syst., 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.788, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[13] Chengan He, Xin Sun, Zhixin Shu, Fujun Luan, Soren Pirk, Jorge Alejandro Amador Herrera, Dominik L Michels, Tuanfeng Y Wang, Meng Zhang, Holly Rushmeier, and Yi Zhou. Perm: A parametric representation for multi-style 3d hair modeling. arXiv preprint arXiv:2407.19451, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Jerry Hsu, Tongtong Wang, Zherong Pan, Xifeng Gao, Cem Yuksel, and Kui Wu. Real-time physically guided hair interpolation. ACM Trans. Graph., 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.177, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[15] Liwen Hu, Chongyang Ma, Linjie Luo, and Hao Li. Single-view hair modeling using a hairstyle database. ACM Trans. Graph., 2015. 3, 4, 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[16] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. In Adv. Neural Inform. Process. Syst., 2023. 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[17] Xin Huang, Ruizhi Shao, Qi Zhang, Hongwen Zhang, Ying Feng, Yebin Liu, and Qing Wang. Humannorm: Learning normal diffusion model for high-quality and realistic 3d human generation. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3, 4, 5, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[18] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.355 + ], + "angle": 0, + "content": "[19] James T Kajiya and Timothy L Kay. Rendering fur with three dimensional textures. ACM Siggraph Computer Graphics, 1989. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.357, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[20] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.401, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[21] Zhiyi Kuang, Yiyang Chen, Hongbo Fu, Kun Zhou, and Youyi Zheng. Deepmvshair: Deep hair modeling from sparse views. In ACM SIGGRAPH Asia, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[22] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Trans. Graph., 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[23] Dongxu Li, Junnan Li, Hung Le, Guangsen Wang, Silvio Savarese, and Steven C.H. Hoi. LAVIS: A one-stop library for language-vision intelligence. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), 2023. 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.905, + 0.622 + ], + "angle": 0, + "content": "[24] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, 2022. 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.624, + 0.905, + 0.678 + ], + "angle": 0, + "content": "[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International Conference on Machine Learning, 2023. 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[26] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[27] Weiyu Li, Rui Chen, Xuelin Chen, and Ping Tan. Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. arXiv preprint arXiv:2310.02596, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[28] Yixun Liang, Xin Yang, Jiantao Lin, Haodong Li, Xiaogang Xu, and Yingcong Chen. Luciddreamer: Towards high-fidelity text-to-3d generation via interval score matching. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[29] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13401" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[30] Hongyu Liu, Xuan Wang, Ziyu Wan, Yujun Shen, Yibing Song, Jing Liao, and Qifeng Chen. Headartist: Text-conditioned 3d head generation with self score distillation. In ACM SIGGRAIPH, 2024. 2, 3, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[31] Yujie Lu, Xianjun Yang, Xiujun Li, Xin Eric Wang, and William Yang Wang. Llmscore: Unveiling the power of large language models in text-to-image synthesis evaluation. In Adv. Neural Inform. Process. Syst., 2024. 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[32] Haimin Luo, Min Ouyang, Zijun Zhao, Suyi Jiang, Longwen Zhang, Qixuan Zhang, Wei Yang, Lan Xu, and Jingyi Yu. Gaussianhair: Hair modeling and rendering with light-aware gaussians. arXiv preprint arXiv:2402.10483, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.304 + ], + "angle": 0, + "content": "[33] Ryota Maeda, Kenshi Takayama, and Takafumi Taketomi. Refinement of hair geometry by strand integration. In Comput. Graph. Forum, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[34] Stephen R Marschner, Henrik Wann Jensen, Mike Cammarano, Steve Worley, and Pat Hanrahan. Light scattering from human hair fibers. ACM Trans. Graph., 2003. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[35] Givi Meishvili, James Clemoes, Charlie Hewitt, Zafiirah Hosenie, Xian Xiao, Martin de La Gorce, Tibor Takacs, Tadas Baltrusaitis, Antonio Criminisi, Chyna McRae, et al. Hairmetry: Fairness-aware hairstyle classification. arXiv preprint arXiv:2410.11528, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[36] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[37] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.532, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[38] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Eur. Conf. Comput. Vis., 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.589, + 0.482, + 0.643 + ], + "angle": 0, + "content": "[39] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In ACM SIGGRAPH Asia, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.645, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[40] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.688, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[41] Giljoo Nam, Chenglei Wu, Min H Kim, and Yaser Sheikh. Strand-accurate multi-view hair capture. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.799 + ], + "angle": 0, + "content": "[42] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.828 + ], + "angle": 0, + "content": "[43] OpenAI. Chatgpt. https://openai.com/, 2024.4, 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.83, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[44] Sylvain Paris, Hector M Briceno, and François X Sillion. Capture of hair geometry from multiple images. ACM Trans. Graph., 2004. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[45] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning con" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "tinuous signed distance functions for shape representation. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[46] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Adv. Neural Inform. Process. Syst., 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[47] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In Int. Conf. Learn. Represent., 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.313 + ], + "angle": 0, + "content": "[48] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng Bo, and Xiaoguang Han. Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.315, + 0.905, + 0.397 + ], + "angle": 0, + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.399, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[50] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.905, + 0.509 + ], + "angle": 0, + "content": "[51] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[52] Anurag Ranjan, Timo Bolkart, Soubhik Sanyal, and Michael J Black. Generating 3d faces using convolutional mesh autoencoders. In Eur. Conf. Comput. Vis., 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[53] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In ACM SIGGRAIPH, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[54] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.706 + ], + "angle": 0, + "content": "[55] Radu Alexandru Rosu, Shunsuke Saito, Ziyan Wang, Chenglei Wu, Sven Behnke, and Giljoo Nam. Neural strands: Learning hair geometry and appearance from multi-view images. In Eur. Conf. Comput. Vis., 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[56] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In Adv. Neural Inform. Process. Syst., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[57] Shunsuke Saito, Liwen Hu, Chongyang Ma, Hikaru Ibayashi, Linjie Luo, and Hao Li. 3d hair synthesis using volumetric variational autoencoders. ACM Trans. Graph., 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[58] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13402" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.146 + ], + "angle": 0, + "content": "[59] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In Eur. Conf. Comput. Vis., 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.148, + 0.483, + 0.204 + ], + "angle": 0, + "content": "[60] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In Adv. Neural Inform. Process. Syst., 2021. 2, 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.483, + 0.26 + ], + "angle": 0, + "content": "[61] Yuefan Shen, Shunsuke Saito, Ziyan Wang, Olivier Maury, Chenglei Wu, Jessica Hodgins, Youyi Zheng, and Giljoo Nam. Ct2hair: High-fidelity 3d hair modeling using computed tomography. ACM Trans. Graph., 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.483, + 0.302 + ], + "angle": 0, + "content": "[62] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. In Int. Conf. Learn. Represent., 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.483, + 0.344 + ], + "angle": 0, + "content": "[63] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In Int. Conf. Learn. Represent., 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.483, + 0.399 + ], + "angle": 0, + "content": "[64] Vanessa Sklyarova, Jenya Chelishev, Andreea Dogaru, Igor Medvedev, Victor Lempitsky, and Egor Zakharov. Neural haircut: Prior-guided strand-based hair reconstruction. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.401, + 0.483, + 0.455 + ], + "angle": 0, + "content": "[65] Vanessa Sklyarova, Egor Zakharov, Otmar Hilliges, Michael J Black, and Justus Thies. Text-conditioned generative model of 3d strand-based human hairstyles. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.483, + 0.497 + ], + "angle": 0, + "content": "[66] Xiaokun Sun, Zhenyu Zhang, Ying Tai, Qian Wang, Hao Tang, Zili Yi, and Jian Yang. Barbie: Text to barbie-style 3d avatars. arXiv preprint arXiv:2408.09126, 2024. 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.483, + 0.566 + ], + "angle": 0, + "content": "[67] Yusuke Takimoto, Hikari Takehara, Hiroyuki Sato, Zihao Zhu, and Bo Zheng. Dr. hair: Reconstructing scalp-connected hair strands without pre-training via differentiable rendering of line segments. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.568, + 0.483, + 0.621 + ], + "angle": 0, + "content": "[68] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splattering for efficient 3d content creation. In Int. Conf. Learn. Represent., 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.483, + 0.678 + ], + "angle": 0, + "content": "[69] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.483, + 0.747 + ], + "angle": 0, + "content": "[70] Duotun Wang, Hengyu Meng, Zeyu Cai, Zhijing Shao, Qianxi Liu, Lin Wang, Mingming Fan, Xiaohang Zhan, and Zeyu Wang. Headevolver: Text to head avatars via expressive and attribute-preserving mesh deformation. arXiv preprint arXiv:2403.09326, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.483, + 0.803 + ], + "angle": 0, + "content": "[71] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In Adv. Neural Inform. Process. Syst., 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[72] Keyu Wu, Yifan Ye, Lingchen Yang, Hongbo Fu, Kun Zhou, and Youyi Zheng. Neuralhdhair: Automatic high-fidelity hair modeling from a single image using implicit neural representations. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.903 + ], + "angle": 0, + "content": "[73] Keyu Wu, Lingchen Yang, Zhiyi Kuang, Yao Feng, Xutao Han, Yuefan Shen, Hongbo Fu, Kun Zhou, and Youyi Zheng." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "Monohair: High-fidelity hair modeling from a monocular video. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[74] Menghua Wu, Hao Zhu, Linjiang Huang, Yi Zhuang, Yuanxun Lu, and Xun Cao. High-fidelity 3d face generation from natural language descriptions. IEEE Conf. Comput. Vis. Pattern Recog., 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[75] Sijing Wu, Yichao Yan, Yunhao Li, Yuhao Cheng, Wenhan Zhu, Ke Gao, Xiaobo Li, and Guangtao Zhai. Ganhead: Towards generative animatable neural head avatars. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[76] Yunjie Wu, Yapeng Meng, Zhipeng Hu, Lincheng Li, Haoqian Wu, Kun Zhou, Weiwei Xu, and Xin Yu. Text-guided 3d face synthesis-from generation to editing. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.306, + 0.905, + 0.373 + ], + "angle": 0, + "content": "[77] Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu Shen, Ruigang Yang, and Xun Cao. Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction. In IEEE Conf. Comput. Vis. Pattern Recog., 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[78] Lingchen Yang, Zefeng Shi, Youyi Zheng, and Kun Zhou. Dynamic hair modeling from monocular videos using deep neural networks. ACM Trans. Graph., 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[79] Tarun Yenamandra, Ayush Tewari, Florian Bernard, Hans-Peter Seidel, Mohamed Elgharib, Daniel Cremers, and Christian Theobalt. i3dmm: Deep implicit 3d morphable model of human heads. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[80] Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[81] Taoran Yi, Jiemin Fang, Zanwei Zhou, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Xinggang Wang, and Qi Tian. Gaussiandreamerpro: Text to manipulable 3d gaussians with highly enhanced quality. arXiv preprint arXiv:2406.18462, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.631, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[82] Cem Yuksel, Scott Schaefer, and John Keyser. Hair meshes. ACM Trans. Graph., 2009. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[83] Egor Zakharov, Vanessa Sklyarova, Michael Black, Giljoo Nam, Justus Thies, and Otmar Hilliges. Human hair reconstruction with strand-aligned 3d gaussians. In Eur. Conf. Comput. Vis., 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.717, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[84] Hao Zhang, Yao Feng, Peter Kulits, Yandong Wen, Justus Thies, and Michael J. Black. Teca: Text-guided generation and editing of compositional 3d avatars. In International Conference on 3D Vision, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.773, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[85] Longwen Zhang, Qiwei Qiu, Hongyang Lin, Qixuan Zhang, Cheng Shi, Wei Yang, Ye Shi, Sibei Yang, Lan Xu, and Jingyi Yu. Dreamface: Progressive generation of animatable 3d faces under text guidance. ACM Trans. Graph., 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[86] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Int. Conf. Comput. Vis., 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.903 + ], + "angle": 0, + "content": "[87] Rui Zhao, Wei Li, Zhipeng Hu, Lincheng Li, Zhengxia Zou, Zhen Xia Shi, and Changjie Fan. Zero-shot text-to-parameter" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13403" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "translation for game character auto-creation. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[88] Mingwu Zheng, Hongyu Yang, Di Huang, and Liming Chen. Imface: A nonlinear 3d morphable face model with implicit neural representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[89] Yujuan Zheng, Zirong Jin, Moran Li, Haibin Huang, Chongyang Ma, Shuguang Cui, and Xiaoguang Han. Hirstep: Transfer synthetic to real using strand and depth maps for single-view 3d hair modeling. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.304 + ], + "angle": 0, + "content": "[90] Yujian Zheng, Yuda Qiu, Leyang Jin, Chongyang Ma, Haibin Huang, Di Zhang, Pengfei Wan, and Xiaoguang Han. Towards unified 3d hair reconstruction from single-view portraits. arXiv preprint arXiv:2409.16863, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[91] Yi Zhou, Liwen Hu, Jun Xing, Weikai Chen, Han-Wei Kung, Xin Tong, and Hao Li. Hairnet: Single-view hair reconstruction using convolutional neural networks. In *Eur. Conf. Comput. Vis.*, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[92] Yuxiao Zhou, Menglei Chai, Alessandro Pepe, Markus Gross, and Thabo Beeler. Groomen: A high-quality generative hair model using hierarchical latent representations. ACM Trans. Graph., 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[93] Yuxiao Zhou, Menglei Chai, Daoye Wang, Sebastian Winberg, Erroll Wood, Kripasindhu Sarkar, Markus Gross, and Thabo Beeler. Groomcap: High-fidelity prior-free hair capture. arXiv preprint arXiv:2409.00831, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.475, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[94] Zhen Zhou, Fan Ma, Hehe Fan, and Yi Yang. Headstudio: Text to animatable head avatars with 3d gaussian splatting. In Eur. Conf. Comput. Vis., 2024. 2, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "13404" + } + ] +] \ No newline at end of file diff --git a/2025/StrandHead_ Text to Hair-Disentangled 3D Head Avatars Using Human-Centric Priors/5c673a06-82fc-4d7f-915c-54daefec29d8_origin.pdf b/2025/StrandHead_ Text to Hair-Disentangled 3D Head Avatars Using Human-Centric Priors/5c673a06-82fc-4d7f-915c-54daefec29d8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0ee0d75eabb74ff854834c218bf413c530e3d40e --- /dev/null +++ b/2025/StrandHead_ Text to Hair-Disentangled 3D Head Avatars Using Human-Centric Priors/5c673a06-82fc-4d7f-915c-54daefec29d8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1fd426a9169d1cd169bb2d0876f4845bb63b39f79bdbb94365315d8566d51bb1 +size 18496830 diff --git a/2025/StrandHead_ Text to Hair-Disentangled 3D Head Avatars Using Human-Centric Priors/full.md b/2025/StrandHead_ Text to Hair-Disentangled 3D Head Avatars Using Human-Centric Priors/full.md new file mode 100644 index 0000000000000000000000000000000000000000..146b63b09b4ed07adf93d954d20129b6dbb7d39a --- /dev/null +++ b/2025/StrandHead_ Text to Hair-Disentangled 3D Head Avatars Using Human-Centric Priors/full.md @@ -0,0 +1,461 @@ +# StrandHead: Text to Hair-Disentangled 3D Head Avatars Using Human-Centric Priors + +Xiaokun Sun, Zeyu Cai, Ying Tai, Jian Yang, Zhenyu Zhang* Nanjing University *Corresponding Author + +xiaokun.sun@smail.nju.edu.cn, caizeyu010612@gmail.com, {yingtai, csjyang}@nju.edu.cn, zhangjesse@foxmail.com + +![](images/cc9a9c1cbcd3b6e4a09d1752acb3c3373d9480d276a6e7d6590e8282c25f2f96.jpg) +(a) Input Prompts + +![](images/d6828f5fab38e6849ba9abba90adf049d66e30e46521a378a46c183696929be5.jpg) +Figure 1. Given (a) input prompts, StrandHead generates (b) realistic 3D head avatars featuring strand-level attributes and (c) 3D hair strands by utilizing human-specific 2D generative priors and 3D hair strand geometric priors. By precisely capturing the internal geometry of hair strands, our approach enables seamless and flexible (d) hairstyle transfer and (e) editing, as well as (f) physics-based simulation. + +# Abstract + +While haircut indicates distinct personality, existing avatar generation methods fail to model practical hair due to the data limitation or entangled representation. We propose StrandHead, a novel text-driven method capable of generating 3D hair strands and disentangled head avatars with strand-level attributes. Instead of using large-scale hair-text paired data for supervision, we demonstrate that realistic hair strands can be generated from prompts by distilling 2D generative models pre-trained on human mesh data. To this end, we propose a meshing approach guided by strand geometry to guarantee the gradient flow from the distillation objective to the neural strand representation. The optimization is then regularized by statistically significant haircut features, leading to stable updating of strands against unreasonable drifting. These employed 2D/3D human-centric priors contribute to text-aligned and realistic 3D strand generation. Extensive experiments show that StrandHead achieves the state-of-the-art performance + +on text to strand generation and disentangled 3D head avatar modeling. The generated 3D hair can be applied on avatars for strand-level editing, as well as implemented in the graphics engine for physical simulation or other applications. Project page: https://xiaokunsun.github.io/StrandHead.github.io/. + +# 1. Introduction + +Creating 3D head avatars is crucial for many applications including digital telepresence, gaming, movies, and AR/VR. Hairstyles reflecting personal characteristics greatly impact the fidelity and realism of digital humans. Traditional methods rely heavily on manual effort, making them time-consuming and labor-intensive. Recent advances [52, 75, 79, 88] achieve automatic 3D head generation with the research paradigm based on supervised learning. However, these methods require costly 3D training data, which significantly limits their generalization capabilities and restricts potential applications. + +With the rapid development of text-driven generation + +methods [47, 54], creating 3D head avatars from given prompts without requiring 3D-text datasets becomes possible. Recent studies [1, 12, 84, 94] unlock the potential of 2D diffusion models in modeling 3D head avatars by integrating them with 3D head priors [4, 26, 77]. HumanNorm [17] further improves domain-specific fidelity by fine-tuning diffusion models on high-quality 3D human meshes. While focusing on facial geometry and texture modeling, these methods utilize holistic meshes or NeRF [38] to represent 3D haircuts, failing to capture the internal geometric structure of hair strands, i.e., 3D curves. This limitation not only significantly reduces the realism of the generated avatars but also makes them incompatible with strand-based applications and simulation systems [2, 9, 10]. + +To accurately model 3D haircuts, recent efforts achieve strand-level hairstyle reconstruction [55, 64, 89] or generation [13, 92] using VAEs or parametric models. Despite their impressive performance, these methods require constrained multi-view images or manual latent space searching, preventing them from freely creating 3D hair based on user-friendly text. The most recent work, HAAR [65] pioneers text-to-strand by training a text-conditioned hair map diffusion model. However, due to its reliance on large-scale and costly paired data including 9825 haircuts and descriptions, HAAR faces challenges in generating diverse hairstyles outside the training set. Additionally, HAAR overlooks hair texture and geometry that adapts to specific head shapes, further limiting its practical applications. + +So can we use powerful human-specific 2D generative priors to create realistic and diverse strand-based hair from text? We address this meaningful and challenging problem by developing StrandHead, a novel framework that generates high-fidelity 3D head avatars with strand-accurate hair from prompts. As illustrated in Fig. 1, the generated head avatars feature diverse and realistic hair strands, enabling seamless and flexible strand-level transfer, editing, and physics-based simulation. Instead of using large-scale hair-prompt paired data, we achieve text-aligned 3D hair strand generation by effectively utilizing human-specific 2D generative priors and 3D hair geometric priors. + +Specifically, we propose a novel differentiable prismatization algorithm inspired by the cylindrical structure of hair strands. This algorithm can efficiently and differentiably convert strands into watertight prismatic meshes, which enables smooth backpropagation of gradients from 2D diffusion models pre-trained on human mesh data to 3D hair strands using mesh-based differentiable renderers, thereby making it possible to use 2D generative priors for modeling 3D hairstyles. Then we propose two effective losses inspired by the distribution patterns of 3D hair geometric features to further regularize the hair shape. These rich 2D/3D human-centric priors work together to achieve text-aligned and realistic 3D hair strand generation. Extensive experi + +ments demonstrate that StrandHead outperforms the state-of-the-art (SOTA) methods in both head and hair generation tasks, and supports flexible haircut transfer and editing, as well as physical-based rendering and simulation. + +Our main contributions are summarized as follows: + +- We propose StrandHead, a novel framework for generating realistic 3D head avatars with strand-level attributes. To the best of our knowledge, StrandHead is the first work to generate 3D hair strands by distilling human-specific 2D diffusion models. +- We propose a differentiable prismatization algorithm that converts hair strands into watertight prismatic meshes. This ensures stable gradient flow from 2D generative priors to 3D hair strand representation, thereby achieving reliable end-to-end strand-based hair optimization. +- Inspired by statistical 3D hair geometric features, we introduce two simple but solid regularization losses to supervise both local and global hair shapes, enabling reasonable and realistic hairstyle generation. + +# 2. Related Work + +Text-to-3D General Object Generation. Inspired by the success of text-to-image (T2I) generation [42, 50, 51, 54, 56, 86], many studies have explored using pre-trained vision-language models [49, 54] to achieve text-guided 3D content generation without large-scale 3D-text paired data. Early methods [18, 37, 39, 58, 69] employ the CLIP model [49] to supervise the alignment of 3D representations with prompts. DreamFusion [47] introduces the Score Distillation Sampling (SDS) loss, significantly enhancing the fidelity of generated 3D content by leveraging more powerful pre-trained diffusion models [54]. Subsequent works further advance text-driven 3D generation by improving 3D representations [5, 6, 80], optimization strategies [29, 68, 81], SDS loss [28, 71], and diffusion models [27, 48, 62]. Despite these advancements, current approaches focused on general content generation do not fully leverage the extensive prior knowledge of human heads and hair, limiting their ability to generate realistic, high-quality 3D head avatars with strand-based hair. + +Text-to-3D Head Avatar Generation. CLIPFace [1], T2P [87] and Describe3D [74] pioneer zero-shot text-driven 3D head generation by combining CLIP [49] with 3D parametric head models [4, 8, 26, 77]. With the introduction of SDS loss [47], DreamFace [85] and FaceG2E [76] leverage pre-trained T2I diffusion models to greatly improve generation quality within specific domains. HeadEvolver [70] enhances the expressiveness of head mesh deformations by introducing vector fields, while HeadSculpt [12], and HeadArtist [30] employ DMTet [60] instead of traditional meshes to capture geometric details. HeadStudio [94] incorporates 3DGS [20] to produce realistic and animatable 3D heads. HumanNorm [17], further + +
TaskMethodHead-Hair-DecoupledStrand-Based HairGeometry & TextureNo Large-Scale 3D-Text Paired Data
Text-to-Head[12, 17, 30, 94]XX✓✓
[84]✓X✓✓
Text-to-Hair[28, 48, 62, 80]XX✓✓
[65]X✓XX
Text-to-Head-HairOurs✓✓✓✓
+ +Table 1. Comparison with current related methods. + +enhances human-specific fidelity by fine-tuning 2D diffusion models on high-quality human mesh data. Despite these advancements, these approaches treat the head and hair as a holistic model, limiting support for downstream applications such as hairstyle transfer and editing. To enable flexible head-hair-disentangled generation, TECA [84] independently represents the head with mesh and the hair with NeRF [38]. However, these methods focus only on the realistic external appearance of hair without modeling its internal geometric structure, restricting their application for strand-level editing and physics-based simulation. + +Strand-Based Hair Creation. Automatic strand-based hair modeling has garnered significant attention from both industry [2, 9, 10] and academia [3, 14, 33, 82], as it enables downstream physics-based rendering and simulation. Such methods can be broadly categorized into two types: (1) reconstructing hair from input images or videos and (2) generating hair based on control conditions. Works on the former one [15, 21, 32, 35, 41, 55, 57, 61, 64, 67, 72, 73, 78, 83, 89-91, 93] relies on classical 3D reconstruction frameworks [20, 36, 38, 45, 59] and incorporates hair-specific features [44] to realize accurate 3D strand reconstruction. However, these methods depend on constrained multi-view images, making hair creation costly and challenging in less controlled settings. Recently, 3D hair generation approaches [13, 92] have emerged using GANs or parametric models, but they struggle with controllable generation. The most relevant work to ours is HAAR [65], which pioneers text-to-strand by training a text-conditioned hair map diffusion model. However, HAAR relies on a large-scale paired dataset including 9825 haircuts and prompts with limited diversity, restricting its ability to generate novel hairstyles beyond the training data. Moreover, HAAR overlooks hair texture and geometry that adapts to specific head shapes, further limiting its application range. + +Compared to previous methods, StrandHead generates 3D heads featuring strand-level hairstyles. Instead of requiring large-scale hair-text paired data, we achieve realistic and diverse 3D hair strand generation by using 2D/3D human-centric priors. We summarize the main differences between our methods and related works in Tab. 1. + +# 3. Method + +# 3.1. Preliminaries + +FLAME [26] is a 3D parametric head model that represents the human head shape, pose, and expression using a compact set of parameters. Given a set of shape parameters $\beta$ , pose parameters $\theta_{\mathrm{pose}}$ , and expression parameters $\psi_{\mathrm{exp}}$ + +FLAME can create a 3D head mesh $M$ . + +HumanNorm [17] is a text-driven, high-quality 3D head generation method. In specific, it models the geometry and texture of the human head using DMTet [60] and a texture field. The alignment between the text $y$ and the optimized representation $\theta$ is achieved using the following SDS loss [47] with T2I diffusion models $\phi$ : + +$$ +\nabla_ {\theta} \mathcal {L} _ {S D S} = \mathbb {E} _ {t, \epsilon} \left[ w (t) \left(\epsilon_ {\phi} \left(x _ {t}; y, t\right) - \epsilon\right) \frac {\partial x}{\partial \theta} \right], \tag {1} +$$ + +where $x = g(\theta)$ is the image rendered from $\theta$ by a differentiable renderer $g$ , $t$ is the time step, $x_{t} = x + \epsilon$ is a noised version of $x$ , $\epsilon_{\phi}(x_{t};y,t)$ is the denoised image, and $w(t)$ is a weighting function. HumanNorm employs human-specific diffusion models fine-tuned on high-quality human mesh data to replace the general diffusion model, enhancing the fidelity of the generated human heads. + +Neural Scalp Textures (NST) [55] is an efficient hair representation where each pixel stores a feature vector conveying the shape information of a single strand at the corresponding scalp location. Using a pre-trained hair strand generator [55] $G$ , the low-dimensional 2D neural scalp texture $T$ can be decoded into high-dimensional 3D strand polylines $S = \{s^i\}_{i=1}^{N_s}$ , where $N_s$ is the number of strands and $s^i = \{p_j^i\}_{j=1}^{N_p}$ consists of $N_p$ 3D points. This process is formulated as follows: $S = G(T)$ . + +# 3.2. Overview + +Given a text prompt, StrandHead aims to create a realistic 3D head avatar with strand-based hair without relying on large-scale 3D-text paired data. Fig. 2 provides an overview of our three-stage pipeline. First, we generate a FLAME-aligned 3D bald head for accurate hair initialization using human-specific diffusion models (Sec. 3.3). In the subsequent stages, we model the reasonable geometry and realistic texture of 3D hair strands by leveraging rich 2D/3D human-centric priors (Sec. 3.4 and Sec. 3.5). + +# 3.3. Bald Head Generation + +To achieve accurate hair initialization and optimize 3D hair strands in subsequent steps using human-specific 2D generative priors, we first need to obtain a reasonable and semantic-aligned 3D bald head from the text (Fig. 2-(a)). To achieve this goal, we enhance HumanNorm [17] by incorporating an evolving FLAME [26] model while optimizing the bald head DMTet $\theta_{\mathrm{h}}$ . This provides accurate semantic information and prevents unnatural geometry. Please refer to the Supp. Mat. for details on bald head generation. + +# 3.4. Hair Geometry Generation + +With a FLAME-aligned bald head, we can generate 3D hair in the scalp area. Specifically, we first initialize the haircut based on hairstyle descriptions, and then further sculpt the 3D hair shape utilizing 2D/3D human-centric priors. + +![](images/ad99dc90aabd00317097bf4905e68ae7922bcf2888ea155b69f46dcbb32984e5.jpg) +Figure 2. StrandHead includes three stages: (a) We first create a FLAME-aligned 3D bald head using the improved HumanNorm [17]. (b) Next, we introduce a differentiable prismatization algorithm to enable human-specific geometry-aware 2D diffusion models to supervise hair shape modeling. Additionally, two losses inspired by 3D hair geometric priors are applied to further regularize the hair geometry. (c) Finally, we use a human-specific normal-conditioned 2D diffusion model to generate lifelike hair textures. + +Hair Initialization. To achieve reasonable and diverse hair initialization, we first utilize ChatGPT [43] to select the 20 most representative hairstyles from the USC-HairSalon Dataset [15]. We then optimize neural scalp textures to fit the selected hairstyle using the following loss function: + +$$ +\mathcal {L} _ {\mathrm {f i t}} = \sum_ {i = 1} ^ {N _ {s}} \sum_ {j = 1} ^ {N _ {p}} \| \hat {p} _ {j} ^ {i} - p _ {j} ^ {i} \| _ {2} + \lambda_ {\mathrm {o r i}} \left(1 - \hat {o} _ {j} ^ {i} \cdot o _ {j} ^ {i}\right) + \lambda_ {\mathrm {c u r}} \| \hat {c} _ {j} ^ {i} - c _ {j} ^ {i} \| _ {1}, \tag {2} +$$ + +where $\hat{p}_j^i$ and $p_j^i$ are the position of the $j$ -th point on the $i$ -th polyline of the GT and generated hair, respectively, and $\hat{o}_j^i$ , $o_j^i$ , $\hat{c}_j^i$ , and $c_j^i$ denote their orientation and curvature, respectively. Here, orientation $o_j^i = (p_{j+1}^i - p_j^i) / \|p_{j+1}^i - p_j^i\|_2$ represents the direction of change in strand position, and curvature $c_j^i = \|o_j^i - o_{j-1}^i\|_2$ denote the rate of change in strand orientation. Given a hair description, we select an optimal pre-trained NST as initialization and further optimize it in subsequent stages. + +Discussion on Strand-Based Differentiable Rendering. With reasonable hair initialization, the next step is to leverage human-specific 2D diffusion models to further sculpt the hair geometry. However, due to the lack of a stable strand-based differentiable renderer, it is very challenging to optimize hair shape using 2D generative priors like other SDS-based methods. Besides, we expect hair strands to produce mesh-style smooth geometry and texture maps to fully exploit the powerful generative capabilities of 2D diffusion models pre-trained on human mesh data. A feasible alternative is to first differentiably convert the strands into quad meshes [82], similar to NeuralHaircut [64], and then leverage SDS-based mesh optimization frameworks to model hair strands. However, this non-watertight stripe-like mesh easily produces ambiguous normal maps or excessively thin sides (see the zoom image of Fig. 3-(c)), which significantly + +![](images/ed9e26760e95466fa25acc4d1f5bb5d0d729d50c0068141ef3c4704f96d3870e.jpg) +(b) A strand-to-mesh conversion example + +![](images/9e422028a461b3b809ddd48b3bde11da94fc0d77b134209cc6ebed939ac9f01c.jpg) +(c) Differentiable prismatization's advantages + +Figure 3. The differentiable prismatization algorithm's (a) gradient backpropagation process, (b) a strand-to-mesh conversion example and (c) advantages over NeuralHaircut [64]. Non-watertight quad meshes can easily produce ambiguous normal maps, which significantly reduce the stability of hair shape modeling (see the drifting hair highlighted by the oval dotted box in (c)). + +reduce the optimization stability (see the drifting hair highlighted by the oval dotted box in Fig. 3-(c)). Therefore, a differentiable strand-to-mesh conversion method with superior optimization properties is urgently needed. + +Differentiable Prismatization (DP). To this end, we propose a novel differentiable prismatization algorithm inspired by the internal structure of 3D hair (i.e., hair fiber is a dielectric cylinder covered with tilted scales and with a pigmented interior [19, 34]). This algorithm can efficiently and differentiably convert strands into watertight prismatic meshes with arbitrary thickness and lateral edges to approximate the cylindrical structure of the hair. This enables + +smooth backpropagation of gradients from the SDS loss to the 3D strand representation using mesh-based differentiable renderers [22], thereby making it possible to fully exploit 2D generative priors distilled from high-fidelity human mesh data for modeling realistic hair strands. Fig. 3-(a) visualizes the complete gradient backpropagation process. + +In specific, given a hair strand $s$ , our DP converts it into a watertight prismatic mesh with $K$ lateral edges and radius $R$ through the following five steps: (1) Compute the Initial Normal Vector. (2) Generate $K$ Rotated Normals. (3) Translate to Form Lateral Edges. (4) Construct Lateral Faces. (5) Construct Top and Bottom Faces. An example of this conversion is shown in Fig. 3-(b). Unlike quad meshes [64], our DP effectively avoids ambiguous normals, enabling smooth gradient backpropagation and hair strand optimization, as shown in Fig. 3-(c). Further algorithm details are provided in the Supp. Mat. + +With the help of our proposed differentiable prismatization algorithm, we optimize the neural scalp texture $T$ by utilizing rich human-specific 2D generative priors through the following SDS losses (Fig. 2-(b)): + +$$ +\nabla_ {T} \mathcal {L} _ {S D S} ^ {\mathrm {h n}} = \mathbb {E} _ {t, \epsilon} \left[ \left(\epsilon_ {\phi_ {\mathrm {h n}}} \left(n _ {t} ^ {\mathrm {h + s}}; y _ {\mathrm {h + s}}, t\right) - \epsilon\right) \frac {\partial n ^ {\mathrm {h + s}}}{\partial T} \right], \tag {3} +$$ + +$$ +\nabla_ {T} \mathcal {L} _ {S D S} ^ {\mathrm {h d}} = \mathbb {E} _ {t, \epsilon} \left[ \left(\epsilon_ {\phi_ {\mathrm {h d}}} \left(d _ {t} ^ {\mathrm {h + s}}; y _ {\mathrm {h + s}}, t\right) - \epsilon\right) \frac {\partial d ^ {\mathrm {h + s}}}{\partial T} \right], \tag {4} +$$ + +where $\phi_{\mathrm{hn}}$ and $\phi_{\mathrm{hd}}$ are the human-specific normal-adapted and depth-adapted diffusion models from Human-Norm [17], $n^{\mathrm{h + s}}$ and $d^{\mathrm{h + s}}$ are the rendered normal and depth maps of the head DMTet $\theta_{\mathrm{h}}$ with hair prismatic mesh $M_p = DP(G(T))$ , respectively. Here, $DP(\cdot)$ denotes the differentiable prismatization operation function, $y_{\mathrm{h + s}}$ is the full text description of the head with hair. + +However, due to the excessive flexibility of hair strands, relying solely on the SDS loss leads to unnatural strand orientations and unrealistic hairstyles (Fig. 9). Furthermore, the lack of constrained multi-view images prevents us from leveraging hair-specific features [44] to provide powerful supervision as in previous works [55]. + +Prior-Driven Losses. To address this issue, we propose two straightforward yet robust prior-driven losses based on observations of the 3D hair strand orientation and curvature to ensure the rationality of the generated hair strands. Specifically, we observe two geometric properties of hair: (1) neighboring strand orientations are highly consistent, and (2) strand curvature is strongly and positively correlated with the curliness of the hairstyle. To validate the properties, we calculate the cosine similarity of adjacent hair strand orientations $CS_{\mathrm{ori}}$ and the average curvature $C_{\mathrm{mean}}$ of 343 hairstyles in the USC-HairSalon dataset [15] using the following equations and visualize their distributions and some examples in Fig. 4: + +![](images/64cf141a3d757c21a2f49eb33cc63169c2a91e50e6d29ccfdefa63c0c7c6de40.jpg) +(a) Distribution of Haircuts' $CS_{\mathrm{ori}}$ + +![](images/21154d866dc276c48d57f1652882bea263b5d835d6f8129359cd98fba729d170.jpg) +(b) Distribution of Haircuts' $C_{\text{mean}}$ + +![](images/106b55f2d6e435ee3533a0ea1403203377d017baff4e427d3afbf7a6b0c05f6c.jpg) +$CS_{\mathrm{ori}} = 0.9949$ $C_{\mathrm{mean}} = 0.0$ + +![](images/c3e4fac1e2faadf5565ed83ca0979ce893b44cfd999b0c7f1e8b22108c3415ff.jpg) + +![](images/80b8a04d78694e2f8701df09132eeaba3f0bb6c953461c285b4c8b9630905698.jpg) + +![](images/664f8c1138d4a077c29e5290468cdf5f593cd7acce4383e23c069f731331aacd.jpg) +$CS_{\mathrm{ori}} = 0.9265$ +(c) Some hairstyles from the USC-HairSalon dataset +Figure 4. Observation of hair geometric features: (1) neighboring strand orientations are highly consistent. (2) strand curvature is strongly and positively related to the haircut curliness. + +![](images/b66a1339d299f12c2485cbce97515ef4a80a0b3dedaf282728378f2025e593cf.jpg) +1 $CS_{\mathrm{ori}} = 0.9868$ $C_{\mathrm{mean}} = 0.0290$ +$C_{\text{mean}} = 0.0598$ + +![](images/ea27960902ded73167b12cc263ffcdc40ac65d118713789938f41b2818a15c78.jpg) +$CS_{\mathrm{ori}} = 0.9719$ $C_{\mathrm{mean}}$ + +![](images/5b03f838af5eb63107b0f2ea2b8e1fe975a06e775c8e315c3ad5609b6db2c663.jpg) +$C S_{\mathrm{ori}} = 0.9605$ $C_{\mathrm{mean}} = 0.0493$ +$= 0.0870$ $CS_{\mathrm{ori}} = 0.9297$ $C_{\mathrm{mean}} = 0.1059$ + +$$ +C S _ {\text {o r i}} = \frac {1}{N _ {s} N _ {p}} \sum_ {i = 1} ^ {N _ {s}} \sum_ {j = 1} ^ {N _ {p}} \sum_ {k \in A (i)} \frac {o _ {j} ^ {i} \cdot o _ {j} ^ {k}}{| A (i) |}, \tag {5} +$$ + +$$ +C _ {\text {m e a n}} = \frac {1}{N _ {s} N _ {p}} \sum_ {i = 1} ^ {N _ {s}} \sum_ {j = 1} ^ {N _ {p}} c _ {j} ^ {i}, \tag {6} +$$ + +where $A(i)$ is the set of adjacent strands for the $i$ -th strand, and $|A(i)|$ is the number of neighboring strands. + +As illustrated, over $95\%$ of hairstyles exhibit an orientation similarity above 0.9, and the average curvature is significantly positively related to curliness, which successfully validates the properties. Based on these observations, we introduce orientation consistency loss and curvature regularization to guide the local and global strand shapes, formulated as follows: + +$$ +\mathcal {L} _ {\mathrm {o r i}} = 1 - C S _ {\mathrm {o r i}}, \tag {7} +$$ + +$$ +\mathcal {L} _ {\text {c u r}} = \left\| C _ {\text {m e a n}} - C _ {\text {t a r g e t}} \right\| _ {1}, \tag {8} +$$ + +where $C_{\mathrm{target}}$ represents the target average curvature set according to the input hair description. These two losses, inspired by 3D hair geometry priors, regularize the hair shape by supervising the consistency of orientations between adjacent strands and the overall hairstyle's curvature. + +Additionally, we introduce a series of losses $\mathcal{L}_{bbox}$ , $\mathcal{L}_{face}$ , and $\mathcal{L}_{colli}$ , which prevent the hair from exceeding the bounding box, obscuring the face, and colliding with the head, respectively. These losses can further improve the rationality of hair geometry. Their details are provided in the Supp. Mat. The final loss function for optimizing hair shape is expressed as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {h a i r - g e o}} = \mathcal {L} _ {S D S} ^ {\mathrm {h d}} + \lambda_ {S D S} ^ {\mathrm {h n}} \mathcal {L} _ {S D S} ^ {\mathrm {h n}} + \lambda_ {\mathrm {o r i}} \mathcal {L} _ {\mathrm {o r i}} + \lambda_ {\mathrm {c u r}} \mathcal {L} _ {\mathrm {c u r}} \\ + \lambda_ {\text {b b o x}} \mathcal {L} _ {\text {b b o x}} + \lambda_ {\text {f a c e}} \mathcal {L} _ {\text {f a c e}} + \lambda_ {\text {c o l l i}} \mathcal {L} _ {\text {c o l l i}}. \tag {9} \\ \end{array} +$$ + +![](images/fdec83993177d362470f63f08009e56af763091211d580eb5e9011ee5930ad6b.jpg) +"An elderly Caucasian grandmother with thin lips and a straight, short silver bob." + +![](images/f0999092568345f52319c6c3be356adc2e521d4df0e83b4437659fef89e95a97.jpg) +"Brad Pitt with medium-length, tousled blonde hair." + +![](images/4f01c4af11edfbae20afbbc640ebc6970c6166a17dbfe3e576cdbc3b8b239b1d.jpg) +"Chris Hemsworth with side-swept sandy blonde hair." + +![](images/88c9f5b6d8cb5c835936a4e7e4e415d56d48bb0226780cc84d8e42af8ac9700a.jpg) +Lionel Messi with a blonde mohawk hairstyle. + +![](images/c3b02bf01fe6f15c8616ba0df153e2511b817b0a6717681e2ea54610dc443013.jpg) +"Scarlet Johansson with a textured blonde bob." + +![](images/b924b600bcb8f03d1405a30c0d4184493a189046da16156995f84023d7cd28c6.jpg) +"A mature European woman with a straight nose and medium-length, straight auburn hair." + +![](images/3959cf6a4eaf806e371795c51ced39ec83533c857fe1979a4db0ba07b6471e92.jpg) +Beyonce with voluminous, curly honey blonde hair. + +![](images/b361d81b71aafa0b986da759404cebd173fe277e1f1cc456848f81b0f846a061.jpg) + +![](images/eeb6efaff330d0842e09289795f92d1bac4a814c2663def2daf0b5188767d178.jpg) +Dwayne Johnson with a short black Afro. + +![](images/0babe7ef1075c71af234f22c026dbfab50140b9b2fa0fe572d50c4278ea85b32.jpg) +"A beautiful Caucasian girl with delicate features and long silky golden blonde hair." + +![](images/14363b7e59c3001c0e63fd3818c5dd4bd9c5d33926b2eb92a660b2799a4da647.jpg) + +![](images/732b9149abd68a2c3f88c8f0bd50ddfa89bc475126481a47da2b6e1417bf5a73.jpg) +A confident Hispanic man with a beard and a curly black afro. + +![](images/c9ef585d10485663d0345f37f687314a02524d06220df44c67a5955b2d8a1bd8.jpg) + +![](images/d4891a97efda9acd967d9c5e35acf933ab01b5591e45bd31f37464ab5aeb5aeb.jpg) +"A charming Caucasian woman with a straight nose and shoulder-length curly sleek blonde hair." + +![](images/e1e8939fe535d0fe1adf32a97bd040dc539acf88867d9957f540a3ebc3bde40e.jpg) +"A curly dark blonde haircut." + +![](images/029876642d1f015224e2a3893c2bdf656708b8fe171773eb14295d044be18679.jpg) + +![](images/b4f13ccfae3ccbdf592bf6e1eb5a9e65c05a8d7ff026e4e92a85e52ba1b0995f.jpg) +"Medium-length, straight auburn hair." + +![](images/a8fda1589724736c4684dc24a17abfc5ea15c59dea5dd4c3efa194dc2d776adc.jpg) +"Long way chestnut brown hair." + +![](images/86166d947a7aa6c3c593494c57a6c28cc8b0edbb88b9dafeacc5eb959f259bdc.jpg) +"A short, wavy brown bob." + +![](images/f9c10bbfcef42bb59080e090bbbac45031de02e22695d7aaee51304e3b3bc21a.jpg) +"Long silky golden blonde hair." + +![](images/e920cbd347fc90a1af3bef2827fddfc0a91f9bc2632e770d14ac46a044d3e633.jpg) +"A light brown mohawk hairstyle." + +![](images/02dc237ab140ad69005e7cb4d61271fe98c6b8856f817429e8b0393323b0a7d9.jpg) +"Side-swept sandy blonde hair." + +![](images/a74c041fd1bc22d8cc1a7854ab1bd4ce119d9c256766ab717b7c261af25b8597.jpg) +"A curly black afro." +Figure 5. Examples of high-fidelity and diverse 3D heads and strand-accurate haircuts generated by our method. The upper visualization includes rendered color and normal maps of the head and hair prismatic meshes. The lower visualization shows the physics-based hair strand rendering result using Blender [10]. For better strand-based visualization, we interpolate generated hair to approximately 10,000 strands and apply a consistent appearance. Please zoom in for detailed views, and refer to the Supp. Mat. for video demonstrations. + +![](images/d50f87087e83a7104ec99d896989ed0b2f7836635e64d8fc838332214644ada5.jpg) +Medium-length wavy caramel-colored hair. + +![](images/6d7a97f106771f43c430ecda7c3ee2e9337dbc512685888ade40b6bcefa77ffd.jpg) +"A side-swept silver haircut." + +![](images/7c2d2122a1904144ffdeb28709a0d673f30f4c055c630a82fe47bbf9b706ee53.jpg) +"A sleek long black hairstyle." + +![](images/70d595b5a43aac4182a6b019945dfce1aa92e7ca1eafd73f4a620e0745f3be11.jpg) +"Medium-length, tousled blonde hair." + +# 3.5. Hair Texture Generation + +Next, we fix the optimized hair strand geometry and model the realistic hair strand texture under the supervision of the SDS loss as follows (Fig. 2-(c)): + +$$ +\nabla_ {\psi_ {s}} \mathcal {L} _ {S D S} ^ {\mathrm {h c}} = \mathbb {E} _ {t, \epsilon} \left(\epsilon_ {\phi_ {\mathrm {h c}}} \left(c _ {t} ^ {\mathrm {h + s}}; n ^ {\mathrm {h + s}}, y _ {\mathrm {h + s}}, t\right) - \epsilon\right) \frac {\partial c ^ {\mathrm {h + s}}}{\partial \psi_ {\mathrm {s}}}, \tag {10} +$$ + +where $\phi_{\mathrm{hc}}$ is the human-specific normal-conditioned diffusion model from HumanNorm [17], $c^{\mathrm{h + s}}$ is the rendered color map of the head texture field $\psi_{\mathrm{h}}$ and hair strand texture field $\psi_{\mathrm{s}}$ . Since the vanilla SDS loss often leads to color oversaturation, we replace it with the following MSDS loss [17] to further enhance the texture's realism in later iterations of texture optimization: + +$$ +\nabla_ {\psi_ {s}} \mathcal {L} _ {M S D S} ^ {\mathrm {h c}} = \mathbb {E} _ {t, \epsilon} \left(H \left(c _ {t} ^ {\mathrm {h + s}}; n ^ {\mathrm {h + s}}, y _ {\mathrm {h + s}}, t\right) - \epsilon\right) \frac {\partial c ^ {\mathrm {h + s}}}{\partial \psi_ {s}}, \tag {11} +$$ + +where $H(\cdot)$ denotes the multi-step operation function. Additionally, we propose a strand-aware texture field that models orientation-dependent texture to better generate high-frequency color variations. Please refer to the Supp. Mat. for more their details. + +# 3.6. Implementation Details + +Our method is implemented using PyTorch [46] and ThreeStudio [11]. Experiments are conducted on an Ubuntu + +server equipped with A6000 GPUs. Starting with a prompt such as "A man with brown hair", we first generate a bald head using "A bald man". Then, we generate the hair based on "A man with brown hair". Further implementation details are available in the Supp. Mat. + +# 4. Experiments + +Some examples of 3D head models with strand-based hair generated by StrandHead are shown in Fig. 5. Due to space limitations, only the key experimental settings and results are presented here. For more results, evaluations, and details, please refer to our Supp. Mat. + +# 4.1. Experimental Settings + +Baselines. We compare StrandHead with the SOTA methods for head avatar and haircut generation. The baseline methods for head generation include text-to-holistic-head works (HeadArtist [30], HumanNorm [17], and HeadStudio [94]) and text-to-decoupled-head work (TECA [84]). For text-to-hair generation, we consider general text-to-3D methods (MVDream [62], GaussianDreamer [80], LucidDreamer [28], and RichDreamer [48]), text-to-strand method (HAAR [65]), and TECA [84]. + +Dataset Construction. To create the dataset, we use ChatGPT[43] to randomly generate 30 text descriptions of heads with hair. These include 30 descriptions for assessing + +![](images/726ab651220bc3ba911020f5f5816018680420bf204b6a632deee9dcd53356b0.jpg) +Figure 6. Qualitative comparisons with the SOTA methods. Since TECA [84] uses the vanilla NeRF to represent hair, rendering normals is not supported. HAAR [65] generates only the geometry of hair strands, so we first convert the strands into prismatic meshes using differentiable prismatization and then utilize TEXTure [53] to generate texture for visualization and comparison. + +![](images/e1a3d8f8618aa06101d317578e6207b453c1e9d061269bd34e9ae3766a7c1ab5.jpg) +Figure 7. Qualitative comparison with HAAR [65]. HAAR, which does not model heads, often produces unreasonable hair-head collisions (highlighted in the black box). + +head generation and 30 for evaluating hair generation. The full list of prompts can be found in the Supp. Mat. + +Evaluation Metrics. Current text-to-3D methods typically use CLIP-based metrics to assess text-image alignment and output quality. However, research [16, 31] shows that these metrics struggle to capture fine-grained alignment between 3D content and input prompts, as validated by our experiments in the Supp. Mat. To address this limitation, we draw inspiration from Progressive3D [7] and Barbie [66], using BLIP-VQA [23, 24] and BLIP2-VQA [23, 25] for evaluation. Specifically, we convert each prompt into two questions to separately verify head and hair. We then feed the rendered image of the generated 3D content into the VQA model, asking each question in sequence and using the probability of a “yes” response as our evaluation metric. For example, the head prompt “A man with black hair” is transformed into “Is the person in the picture a man?” and “Does the person in the picture have black hair?” The hair prompt “Black hair” becomes “Is the object in the picture black hair?” Finally, we conduct a user study by randomly + +
MethodBLIP-VQA ↑BLIP2-VQA ↑GQP ↑TAP ↑
HeadArtist0.76670.96671.002.33
HeadStudio0.78330.88333.333.67
HumanNorm0.70000.95007.677.67
TECA0.73330.950034.3328.33
StrandHead (Ours)0.85000.966753.6758.00
MVDream0.90000.833324.6720.00
GaussianDreamer0.33330.36675.333.00
LucidDreamer0.80000.93335.335.00
RichDreamer0.83330.76674.005.67
TECA0.70000.76671.673.67
HAAR0.63330.20001.332.33
StrandHead (Ours)0.90000.900057.6760.33
+ +Table 2. Quantitative comparisons with the SOTA methods. The best and second-best results are highlighted in **bolded** and **underlined**, respectively. GQP: generation quality preference (\%), TAP: text-image alignment preference (\%). + +selecting 10 generated examples and asking 30 volunteers to assess (1) generation quality and (2) text-image alignment, and select the best methods. + +# 4.2. Comparisons of Head Generation + +As illustrated in the upper section of Tab. 2, StrandHead outperforms all comparison methods across every evaluation metric. The qualitative results on the left side of Fig. 6 further highlight the advantages of our method. Compared to other head generation methods, the head avatars generated by StrandHead exhibit not only more refined facial geometry and texture details but also strand-accurate hair with a plausible appearance that integrates seamlessly with physics-based simulation systems. To the best of our knowledge, StrandHead is the first head generation framework that realizes strand-level hair modeling, a capability that holds substantial potential for advancing human-centric 3D AIGC applications in the industry. + +# 4.3. Comparisons of Hair Generation + +As shown in the lower part of Tab. 2, our approach surpasses other methods across most evaluation metrics, and is only slightly lower than LucidDreamer (the latest open-source + +![](images/c40766afd0a49292a565959bd32f41365bfa422c41ada07c83ffd878d3bc0e45.jpg) +(b) Generated strand-based hair under different bald heads +Figure 8. Analysis of human-specific 2D diffusion models. + +text-to-3D method) on BLIP2-VQA, ranking the second. The qualitative comparisons on the right side of Fig. 6 display sample generation results. Compared to general text-to-3D methods, our approach more precisely captures the internal structure of hair strands, resulting in realistic hair-cut geometry and appearance without incorporating incorrect content such as parts of the human head. + +Compared with HAAR [65], StrandHead relies on robust 2D/3D human-centric priors rather than large-scale paired datasets to supervise hair generation. This not only allows it to generate hairstyles that are uncommon in the dataset (see slicked-back and side-swept hairstyles in Fig. 7), but also models hair texture or geometry that adapts to specific head shapes (Fig. 8-(b)), thereby avoiding unnatural hairhead collisions (see the black box in Fig. 7). + +# 4.4. Ablation Study + +Effectiveness of Human-Specific 2D Generative Priors. We demonstrate the importance of human-specific 2D generative priors from two aspects: (1) As shown in Fig. 8-(a), ignoring 2D supervision fails to generate meaningful hair. While a general 2D diffusion model can slightly improve modeling performance, it still falls short of producing high-quality results. Only by incorporating human-specific 2D generative priors can one achieve realistic textures and reasonable shapes for hairstyles. (2) Fig. 8-(b) illustrates generated hair under varying bald head conditions while maintaining a fixed initial hairstyle and hair prompt. The hairstyles created under the supervision of human-specific 2D diffusion models exhibit clear geometric and textural variations that adapt to specific bald heads. This strongly demonstrates the necessity of considering the bald head for hair strand generation and highlights the advantages of + +![](images/caa771c0a1343c0ff68f158b02de5e524c828527c22e7cb49e6e0eaea649f817.jpg) + +![](images/f5fff9e96c6ad00b9b22679b3c6b79fabbe20f13be35b581ca93b3517290221c.jpg) + +![](images/e0b563cded1f058245363216826416541b97eaacb3379cab272a282b58689713.jpg) +Figure 9. Ablation study on (a) orientation consistency loss $\mathcal{L}_{\mathrm{ori}}$ and (b) curvature regularization loss $\mathcal{L}_{\mathrm{cur}}$ . +Figure 10. Failure cases. + +![](images/b601b5b6baea13c04a08fba7c1b13ee2a6eb9e0540e7ae07225ecd5eea0b057c.jpg) + +StrandHead over HAAR [65]. + +Effectiveness of Prior-Driven Losses. Fig. 9-(a) demonstrates that incorporating $\mathcal{L}_{\mathrm{ori}}$ provides strong guidance for generating natural hair by supervising the consistency of orientations between adjacent strands. As shown in Fig. 9-(b), our $\mathcal{L}_{\mathrm{cur}}$ preserves the desired curliness by constraining the overall hairstyle's curvature. In summary, these two losses, inspired by 3D hair geometry priors, effectively enhance the realism and rationality of the generated haircut. + +# 5. Conclusion + +In this paper, a novel framework StrandHead is proposed to generate realistic hair-disentangled 3D head avatars using text prompts. With a series of 2D generative models pre-trained on human-centric data, StrandHead generates strand-based hair requiring no large-scale text-hair paired data for supervision. This goal is achieved mainly by two approaches: (1) proposing a novel differentiable prismatization method that transforms the hair strands into prismatic meshes, and (2) using the prior-driven losses inspired by the observation of real-world haircut and the statistical features of human-crafted 3D hair data. The former one produces watertight geometry more similar to the original hair strands, boosting a stable optimization with significantly less normal ambiguity. Besides, the latter one well constrains the optimization with reliable guidance to generate reasonable 3D haircuts. Extensive experiments demonstrate that StrandHead obtains the state-of-the-art performance on 3D head avatars and haircuts generation. The generated results are easily implemented in industrial software to produce physical simulation and high-fidelity rendering. + +Limitations and Future Work. Due to the limited representation capabilities of the strand generator, StrandHead is unable to create highly complex 3D hairstyles (e.g., dreadlocks and ponytails in Fig. 10). Additionally, although the SDS-based method effectively leverages 2D generative priors, its high computational cost restricts practical application. In future work, we aim to address these challenges by incorporating richer hairstyle datasets and exploring deeper hair priors. + +Acknowledgement. This work was supported by the National Science Foundation of China (62376121). + +# References + +[1] Shivangi Aneja, Justus Thies, Angela Dai, and Matthias Nießner. Clipface: Text-guided editing of textured 3d morphable models. ACM SIGGRAIPH, 2022. 2 +[2] Inc. Autodesk. Autodesk maya - 3d animation and modeling software. https://www.autodesk.com/products/maya, 2024.2, 3 +[3] Gaurav Bhokare, Eisen Montalvo, Elie Diaz, and Cem Yuksel. Real-time hair rendering with hair meshes. In ACM SIGGRAIPH, 2024. 3 +[4] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, 2023. 2 +[5] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Int. Conf. Comput. Vis., 2023. 2 +[6] Zilong Chen, Feng Wang, and Huaping Liu. Text-to-3d using gaussian splatting. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2 +[7] Xinhua Cheng, Tianyu Yang, Jianan Wang, Yu Li, Lei Zhang, Jian Zhang, and Li Yuan. Progressive3d: Progressively local editing for text-to-3d content creation with complex semantic prompts. In Int. Conf. Learn. Represent., 2023. 7, 5 +[8] Hang Dai, Nick Pears, William Smith, and Christian Duncan. Statistical modeling of craniofacial shape and texture. Int. J. Comput. Vis., 2020. 2 +[9] Inc. Epic Games. Unreal engine - real-time 3d creation tool. https://www.unrealengine.com/, 2024.2, 3 +[10] Blender Foundation. Blender - a 3d modelling and rendering software. https://www.blender.org/, 2024.2,3,6, 7 +[11] Yuan-Chen Guo, Ying-Tian Liu, Ruizhi Shao, Christian Laforte, Vikram Voleti, Guan Luo, Chia-Hao Chen, Zixin Zou, Chen Wang, Yan-Pei Cao, and Song-Hai Zhang. threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio, 2023.6 +[12] Xiaoping Han, Yukang Cao, K. Han, Xiatian Zhu, Jiankang Deng, Yi-Zhe Song, Tao Xiang, and Kwan-Yee K. Wong. Headsculpt: Crafting 3d head avatars with text. In Adv. Neural Inform. Process. Syst., 2023. 2, 3 +[13] Chengan He, Xin Sun, Zhixin Shu, Fujun Luan, Soren Pirk, Jorge Alejandro Amador Herrera, Dominik L Michels, Tuanfeng Y Wang, Meng Zhang, Holly Rushmeier, and Yi Zhou. Perm: A parametric representation for multi-style 3d hair modeling. arXiv preprint arXiv:2407.19451, 2024. 2, 3 +[14] Jerry Hsu, Tongtong Wang, Zherong Pan, Xifeng Gao, Cem Yuksel, and Kui Wu. Real-time physically guided hair interpolation. ACM Trans. Graph., 2024. 3 + +[15] Liwen Hu, Chongyang Ma, Linjie Luo, and Hao Li. Single-view hair modeling using a hairstyle database. ACM Trans. Graph., 2015. 3, 4, 5, 2 +[16] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. In Adv. Neural Inform. Process. Syst., 2023. 7, 5 +[17] Xin Huang, Ruizhi Shao, Qi Zhang, Hongwen Zhang, Ying Feng, Yebin Liu, and Qing Wang. Humannorm: Learning normal diffusion model for high-quality and realistic 3d human generation. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3, 4, 5, 6, 1 +[18] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2 +[19] James T Kajiya and Timothy L Kay. Rendering fur with three dimensional textures. ACM Siggraph Computer Graphics, 1989. 4 +[20] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 2023. 2, 3 +[21] Zhiyi Kuang, Yiyang Chen, Hongbo Fu, Kun Zhou, and Youyi Zheng. Deepmvshair: Deep hair modeling from sparse views. In ACM SIGGRAPH Asia, 2022. 3 +[22] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Trans. Graph., 2020. 5 +[23] Dongxu Li, Junnan Li, Hung Le, Guangsen Wang, Silvio Savarese, and Steven C.H. Hoi. LAVIS: A one-stop library for language-vision intelligence. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), 2023. 7, 5 +[24] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, 2022. 7, 5 +[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International Conference on Machine Learning, 2023. 7, 5 +[26] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 2017. 2, 3 +[27] Weiyu Li, Rui Chen, Xuelin Chen, and Ping Tan. Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. arXiv preprint arXiv:2310.02596, 2023. 2 +[28] Yixun Liang, Xin Yang, Jiantao Lin, Haodong Li, Xiaogang Xu, and Yingcong Chen. Luciddreamer: Towards high-fidelity text-to-3d generation via interval score matching. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3, 6 +[29] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 2 + +[30] Hongyu Liu, Xuan Wang, Ziyu Wan, Yujun Shen, Yibing Song, Jing Liao, and Qifeng Chen. Headartist: Text-conditioned 3d head generation with self score distillation. In ACM SIGGRAIPH, 2024. 2, 3, 6, 1 +[31] Yujie Lu, Xianjun Yang, Xiujun Li, Xin Eric Wang, and William Yang Wang. Llmscore: Unveiling the power of large language models in text-to-image synthesis evaluation. In Adv. Neural Inform. Process. Syst., 2024. 7, 5 +[32] Haimin Luo, Min Ouyang, Zijun Zhao, Suyi Jiang, Longwen Zhang, Qixuan Zhang, Wei Yang, Lan Xu, and Jingyi Yu. Gaussianhair: Hair modeling and rendering with light-aware gaussians. arXiv preprint arXiv:2402.10483, 2024. 3 +[33] Ryota Maeda, Kenshi Takayama, and Takafumi Taketomi. Refinement of hair geometry by strand integration. In Comput. Graph. Forum, 2023. 3 +[34] Stephen R Marschner, Henrik Wann Jensen, Mike Cammarano, Steve Worley, and Pat Hanrahan. Light scattering from human hair fibers. ACM Trans. Graph., 2003. 4 +[35] Givi Meishvili, James Clemoes, Charlie Hewitt, Zafiirah Hosenie, Xian Xiao, Martin de La Gorce, Tibor Takacs, Tadas Baltrusaitis, Antonio Criminisi, Chyna McRae, et al. Hairmetry: Fairness-aware hairstyle classification. arXiv preprint arXiv:2410.11528, 2024. 3 +[36] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 3 +[37] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2 +[38] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Eur. Conf. Comput. Vis., 2020. 2, 3 +[39] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In ACM SIGGRAPH Asia, 2022. 2 +[40] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 2022. 1 +[41] Giljoo Nam, Chenglei Wu, Min H Kim, and Yaser Sheikh. Strand-accurate multi-view hair capture. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 3 +[42] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2 +[43] OpenAI. Chatgpt. https://openai.com/, 2024.4, 6, 2 +[44] Sylvain Paris, Hector M Briceno, and François X Sillion. Capture of hair geometry from multiple images. ACM Trans. Graph., 2004. 3, 5 +[45] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning con + +tinuous signed distance functions for shape representation. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 3 +[46] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Adv. Neural Inform. Process. Syst., 2019. 6 +[47] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In Int. Conf. Learn. Represent., 2022. 2, 3 +[48] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng Bo, and Xiaoguang Han. Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3, 6 +[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 2 +[50] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, 2021. 2 +[51] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2 +[52] Anurag Ranjan, Timo Bolkart, Soubhik Sanyal, and Michael J Black. Generating 3d faces using convolutional mesh autoencoders. In Eur. Conf. Comput. Vis., 2018. 1 +[53] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In ACM SIGGRAIPH, 2023. 7 +[54] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2 +[55] Radu Alexandru Rosu, Shunsuke Saito, Ziyan Wang, Chenglei Wu, Sven Behnke, and Giljoo Nam. Neural strands: Learning hair geometry and appearance from multi-view images. In Eur. Conf. Comput. Vis., 2022. 2, 3, 5 +[56] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In Adv. Neural Inform. Process. Syst., 2022. 2 +[57] Shunsuke Saito, Liwen Hu, Chongyang Ma, Hikaru Ibayashi, Linjie Luo, and Hao Li. 3d hair synthesis using volumetric variational autoencoders. ACM Trans. Graph., 2018. 3 +[58] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2 + +[59] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In Eur. Conf. Comput. Vis., 2016. 3 +[60] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In Adv. Neural Inform. Process. Syst., 2021. 2, 3, 1 +[61] Yuefan Shen, Shunsuke Saito, Ziyan Wang, Olivier Maury, Chenglei Wu, Jessica Hodgins, Youyi Zheng, and Giljoo Nam. Ct2hair: High-fidelity 3d hair modeling using computed tomography. ACM Trans. Graph., 2023. 3 +[62] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. In Int. Conf. Learn. Represent., 2023. 2, 3, 6 +[63] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In Int. Conf. Learn. Represent., 2014. 2 +[64] Vanessa Sklyarova, Jenya Chelishev, Andreea Dogaru, Igor Medvedev, Victor Lempitsky, and Egor Zakharov. Neural haircut: Prior-guided strand-based hair reconstruction. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 2, 3, 4, 5 +[65] Vanessa Sklyarova, Egor Zakharov, Otmar Hilliges, Michael J Black, and Justus Thies. Text-conditioned generative model of 3d strand-based human hairstyles. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3, 6, 7, 8 +[66] Xiaokun Sun, Zhenyu Zhang, Ying Tai, Qian Wang, Hao Tang, Zili Yi, and Jian Yang. Barbie: Text to barbie-style 3d avatars. arXiv preprint arXiv:2408.09126, 2024. 7, 1 +[67] Yusuke Takimoto, Hikari Takehara, Hiroyuki Sato, Zihao Zhu, and Bo Zheng. Dr. hair: Reconstructing scalp-connected hair strands without pre-training via differentiable rendering of line segments. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 3 +[68] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splattering for efficient 3d content creation. In Int. Conf. Learn. Represent., 2023. 2 +[69] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2 +[70] Duotun Wang, Hengyu Meng, Zeyu Cai, Zhijing Shao, Qianxi Liu, Lin Wang, Mingming Fan, Xiaohang Zhan, and Zeyu Wang. Headevolver: Text to head avatars via expressive and attribute-preserving mesh deformation. arXiv preprint arXiv:2403.09326, 2024. 2 +[71] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In Adv. Neural Inform. Process. Syst., 2024. 2 +[72] Keyu Wu, Yifan Ye, Lingchen Yang, Hongbo Fu, Kun Zhou, and Youyi Zheng. Neuralhdhair: Automatic high-fidelity hair modeling from a single image using implicit neural representations. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 3 +[73] Keyu Wu, Lingchen Yang, Zhiyi Kuang, Yao Feng, Xutao Han, Yuefan Shen, Hongbo Fu, Kun Zhou, and Youyi Zheng. + +Monohair: High-fidelity hair modeling from a monocular video. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 3 +[74] Menghua Wu, Hao Zhu, Linjiang Huang, Yi Zhuang, Yuanxun Lu, and Xun Cao. High-fidelity 3d face generation from natural language descriptions. IEEE Conf. Comput. Vis. Pattern Recog., 2023. 2 +[75] Sijing Wu, Yichao Yan, Yunhao Li, Yuhao Cheng, Wenhan Zhu, Ke Gao, Xiaobo Li, and Guangtao Zhai. Ganhead: Towards generative animatable neural head avatars. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 1 +[76] Yunjie Wu, Yapeng Meng, Zhipeng Hu, Lincheng Li, Haoqian Wu, Kun Zhou, Weiwei Xu, and Xin Yu. Text-guided 3d face synthesis-from generation to editing. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2 +[77] Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu Shen, Ruigang Yang, and Xun Cao. Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction. In IEEE Conf. Comput. Vis. Pattern Recog., 2020. 2 +[78] Lingchen Yang, Zefeng Shi, Youyi Zheng, and Kun Zhou. Dynamic hair modeling from monocular videos using deep neural networks. ACM Trans. Graph., 2019. 3 +[79] Tarun Yenamandra, Ayush Tewari, Florian Bernard, Hans-Peter Seidel, Mohamed Elgharib, Daniel Cremers, and Christian Theobalt. i3dmm: Deep implicit 3d morphable model of human heads. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 1 +[80] Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3, 6 +[81] Taoran Yi, Jiemin Fang, Zanwei Zhou, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Xinggang Wang, and Qi Tian. Gaussiandreamerpro: Text to manipulable 3d gaussians with highly enhanced quality. arXiv preprint arXiv:2406.18462, 2024. 2 +[82] Cem Yuksel, Scott Schaefer, and John Keyser. Hair meshes. ACM Trans. Graph., 2009. 3, 4 +[83] Egor Zakharov, Vanessa Sklyarova, Michael Black, Giljoo Nam, Justus Thies, and Otmar Hilliges. Human hair reconstruction with strand-aligned 3d gaussians. In Eur. Conf. Comput. Vis., 2024. 3 +[84] Hao Zhang, Yao Feng, Peter Kulits, Yandong Wen, Justus Thies, and Michael J. Black. Teca: Text-guided generation and editing of compositional 3d avatars. In International Conference on 3D Vision, 2024. 2, 3, 6, 7 +[85] Longwen Zhang, Qiwei Qiu, Hongyang Lin, Qixuan Zhang, Cheng Shi, Wei Yang, Ye Shi, Sibei Yang, Lan Xu, and Jingyi Yu. Dreamface: Progressive generation of animatable 3d faces under text guidance. ACM Trans. Graph., 2023. 2 +[86] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Int. Conf. Comput. Vis., 2023. 2 +[87] Rui Zhao, Wei Li, Zhipeng Hu, Lincheng Li, Zhengxia Zou, Zhen Xia Shi, and Changjie Fan. Zero-shot text-to-parameter + +translation for game character auto-creation. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 2 +[88] Mingwu Zheng, Hongyu Yang, Di Huang, and Liming Chen. Imface: A nonlinear 3d morphable face model with implicit neural representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022. 1 +[89] Yujuan Zheng, Zirong Jin, Moran Li, Haibin Huang, Chongyang Ma, Shuguang Cui, and Xiaoguang Han. Hirstep: Transfer synthetic to real using strand and depth maps for single-view 3d hair modeling. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 2, 3 +[90] Yujian Zheng, Yuda Qiu, Leyang Jin, Chongyang Ma, Haibin Huang, Di Zhang, Pengfei Wan, and Xiaoguang Han. Towards unified 3d hair reconstruction from single-view portraits. arXiv preprint arXiv:2409.16863, 2024. +[91] Yi Zhou, Liwen Hu, Jun Xing, Weikai Chen, Han-Wei Kung, Xin Tong, and Hao Li. Hairnet: Single-view hair reconstruction using convolutional neural networks. In *Eur. Conf. Comput. Vis.*, 2018. 3 +[92] Yuxiao Zhou, Menglei Chai, Alessandro Pepe, Markus Gross, and Thabo Beeler. Groomen: A high-quality generative hair model using hierarchical latent representations. ACM Trans. Graph., 2023. 2, 3 +[93] Yuxiao Zhou, Menglei Chai, Daoye Wang, Sebastian Winberg, Erroll Wood, Kripasindhu Sarkar, Markus Gross, and Thabo Beeler. Groomcap: High-fidelity prior-free hair capture. arXiv preprint arXiv:2409.00831, 2024. 3 +[94] Zhen Zhou, Fan Ma, Hehe Fan, and Yi Yang. Headstudio: Text to animatable head avatars with 3d gaussian splatting. In Eur. Conf. Comput. Vis., 2024. 2, 3, 6 \ No newline at end of file diff --git a/2025/StrandHead_ Text to Hair-Disentangled 3D Head Avatars Using Human-Centric Priors/images.zip b/2025/StrandHead_ Text to Hair-Disentangled 3D Head Avatars Using Human-Centric Priors/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0c621941def43cc5297d6b12cc6f7160b51250eb --- /dev/null +++ b/2025/StrandHead_ Text to Hair-Disentangled 3D Head Avatars Using Human-Centric Priors/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2ca6b17d3975b8a666840b1b246e41f212249ee468a51c093216caa83b00a04e +size 837761 diff --git a/2025/StrandHead_ Text to Hair-Disentangled 3D Head Avatars Using Human-Centric Priors/layout.json b/2025/StrandHead_ Text to Hair-Disentangled 3D Head Avatars Using Human-Centric Priors/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..21c444b93e5ae14c56b6773acb744d61c1f3a468 --- /dev/null +++ b/2025/StrandHead_ Text to Hair-Disentangled 3D Head Avatars Using Human-Centric Priors/layout.json @@ -0,0 +1,12242 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 128, + 103, + 483, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 103, + 483, + 140 + ], + "spans": [ + { + "bbox": [ + 128, + 103, + 483, + 140 + ], + "type": "text", + "content": "StrandHead: Text to Hair-Disentangled 3D Head Avatars Using Human-Centric Priors" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 161, + 455, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 161, + 455, + 190 + ], + "spans": [ + { + "bbox": [ + 159, + 161, + 455, + 190 + ], + "type": "text", + "content": "Xiaokun Sun, Zeyu Cai, Ying Tai, Jian Yang, Zhenyu Zhang* Nanjing University *Corresponding Author" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 191, + 449, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 191, + 449, + 217 + ], + "spans": [ + { + "bbox": [ + 159, + 191, + 449, + 217 + ], + "type": "text", + "content": "xiaokun.sun@smail.nju.edu.cn, caizeyu010612@gmail.com, {yingtai, csjyang}@nju.edu.cn, zhangjesse@foxmail.com" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 97, + 217, + 511, + 357 + ], + "blocks": [ + { + "bbox": [ + 97, + 217, + 511, + 357 + ], + "lines": [ + { + "bbox": [ + 97, + 217, + 511, + 357 + ], + "spans": [ + { + "bbox": [ + 97, + 217, + 511, + 357 + ], + "type": "image", + "image_path": "cc9a9c1cbcd3b6e4a09d1752acb3c3373d9480d276a6e7d6590e8282c25f2f96.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 114, + 359, + 170, + 369 + ], + "lines": [ + { + "bbox": [ + 114, + 359, + 170, + 369 + ], + "spans": [ + { + "bbox": [ + 114, + 359, + 170, + 369 + ], + "type": "text", + "content": "(a) Input Prompts" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 189, + 358, + 509, + 427 + ], + "blocks": [ + { + "bbox": [ + 189, + 358, + 509, + 427 + ], + "lines": [ + { + "bbox": [ + 189, + 358, + 509, + 427 + ], + "spans": [ + { + "bbox": [ + 189, + 358, + 509, + 427 + ], + "type": "image", + "image_path": "d6828f5fab38e6849ba9abba90adf049d66e30e46521a378a46c183696929be5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 427, + 555, + 461 + ], + "lines": [ + { + "bbox": [ + 54, + 427, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 54, + 427, + 555, + 461 + ], + "type": "text", + "content": "Figure 1. Given (a) input prompts, StrandHead generates (b) realistic 3D head avatars featuring strand-level attributes and (c) 3D hair strands by utilizing human-specific 2D generative priors and 3D hair strand geometric priors. By precisely capturing the internal geometry of hair strands, our approach enables seamless and flexible (d) hairstyle transfer and (e) editing, as well as (f) physics-based simulation." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 152, + 470, + 200, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 470, + 200, + 483 + ], + "spans": [ + { + "bbox": [ + 152, + 470, + 200, + 483 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "text", + "content": "While haircut indicates distinct personality, existing avatar generation methods fail to model practical hair due to the data limitation or entangled representation. We propose StrandHead, a novel text-driven method capable of generating 3D hair strands and disentangled head avatars with strand-level attributes. Instead of using large-scale hair-text paired data for supervision, we demonstrate that realistic hair strands can be generated from prompts by distilling 2D generative models pre-trained on human mesh data. To this end, we propose a meshing approach guided by strand geometry to guarantee the gradient flow from the distillation objective to the neural strand representation. The optimization is then regularized by statistically significant haircut features, leading to stable updating of strands against unreasonable drifting. These employed 2D/3D human-centric priors contribute to text-aligned and realistic 3D strand generation. Extensive experiments show that StrandHead achieves the state-of-the-art performance" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 472, + 556, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 556, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 556, + 544 + ], + "type": "text", + "content": "on text to strand generation and disentangled 3D head avatar modeling. The generated 3D hair can be applied on avatars for strand-level editing, as well as implemented in the graphics engine for physical simulation or other applications. Project page: https://xiaokunsun.github.io/StrandHead.github.io/." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 550, + 395, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 550, + 395, + 562 + ], + "spans": [ + { + "bbox": [ + 314, + 550, + 395, + 562 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 570, + 555, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 570, + 555, + 701 + ], + "spans": [ + { + "bbox": [ + 312, + 570, + 555, + 701 + ], + "type": "text", + "content": "Creating 3D head avatars is crucial for many applications including digital telepresence, gaming, movies, and AR/VR. Hairstyles reflecting personal characteristics greatly impact the fidelity and realism of digital humans. Traditional methods rely heavily on manual effort, making them time-consuming and labor-intensive. Recent advances [52, 75, 79, 88] achieve automatic 3D head generation with the research paradigm based on supervised learning. However, these methods require costly 3D training data, which significantly limits their generalization capabilities and restricts potential applications." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "content": "With the rapid development of text-driven generation" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13393" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "content": "methods [47, 54], creating 3D head avatars from given prompts without requiring 3D-text datasets becomes possible. Recent studies [1, 12, 84, 94] unlock the potential of 2D diffusion models in modeling 3D head avatars by integrating them with 3D head priors [4, 26, 77]. HumanNorm [17] further improves domain-specific fidelity by fine-tuning diffusion models on high-quality 3D human meshes. While focusing on facial geometry and texture modeling, these methods utilize holistic meshes or NeRF [38] to represent 3D haircuts, failing to capture the internal geometric structure of hair strands, i.e., 3D curves. This limitation not only significantly reduces the realism of the generated avatars but also makes them incompatible with strand-based applications and simulation systems [2, 9, 10]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 242, + 295, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 242, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 295, + 410 + ], + "type": "text", + "content": "To accurately model 3D haircuts, recent efforts achieve strand-level hairstyle reconstruction [55, 64, 89] or generation [13, 92] using VAEs or parametric models. Despite their impressive performance, these methods require constrained multi-view images or manual latent space searching, preventing them from freely creating 3D hair based on user-friendly text. The most recent work, HAAR [65] pioneers text-to-strand by training a text-conditioned hair map diffusion model. However, due to its reliance on large-scale and costly paired data including 9825 haircuts and descriptions, HAAR faces challenges in generating diverse hairstyles outside the training set. Additionally, HAAR overlooks hair texture and geometry that adapts to specific head shapes, further limiting its practical applications." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 412, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 295, + 556 + ], + "type": "text", + "content": "So can we use powerful human-specific 2D generative priors to create realistic and diverse strand-based hair from text? We address this meaningful and challenging problem by developing StrandHead, a novel framework that generates high-fidelity 3D head avatars with strand-accurate hair from prompts. As illustrated in Fig. 1, the generated head avatars feature diverse and realistic hair strands, enabling seamless and flexible strand-level transfer, editing, and physics-based simulation. Instead of using large-scale hair-prompt paired data, we achieve text-aligned 3D hair strand generation by effectively utilizing human-specific 2D generative priors and 3D hair geometric priors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": "Specifically, we propose a novel differentiable prismatization algorithm inspired by the cylindrical structure of hair strands. This algorithm can efficiently and differentiably convert strands into watertight prismatic meshes, which enables smooth backpropagation of gradients from 2D diffusion models pre-trained on human mesh data to 3D hair strands using mesh-based differentiable renderers, thereby making it possible to use 2D generative priors for modeling 3D hairstyles. Then we propose two effective losses inspired by the distribution patterns of 3D hair geometric features to further regularize the hair shape. These rich 2D/3D human-centric priors work together to achieve text-aligned and realistic 3D hair strand generation. Extensive experi" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "ments demonstrate that StrandHead outperforms the state-of-the-art (SOTA) methods in both head and hair generation tasks, and supports flexible haircut transfer and editing, as well as physical-based rendering and simulation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 326, + 121, + 534, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 121, + 534, + 131 + ], + "spans": [ + { + "bbox": [ + 326, + 121, + 534, + 131 + ], + "type": "text", + "content": "Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 132, + 553, + 300 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 132, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 132, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 314, + 132, + 553, + 191 + ], + "type": "text", + "content": "- We propose StrandHead, a novel framework for generating realistic 3D head avatars with strand-level attributes. To the best of our knowledge, StrandHead is the first work to generate 3D hair strands by distilling human-specific 2D diffusion models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 192, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 553, + 251 + ], + "type": "text", + "content": "- We propose a differentiable prismatization algorithm that converts hair strands into watertight prismatic meshes. This ensures stable gradient flow from 2D generative priors to 3D hair strand representation, thereby achieving reliable end-to-end strand-based hair optimization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 251, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 553, + 300 + ], + "type": "text", + "content": "- Inspired by statistical 3D hair geometric features, we introduce two simple but solid regularization losses to supervise both local and global hair shapes, enabling reasonable and realistic hairstyle generation." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 311, + 400, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 311, + 400, + 323 + ], + "spans": [ + { + "bbox": [ + 314, + 311, + 400, + 323 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 330, + 553, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 553, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 553, + 556 + ], + "type": "text", + "content": "Text-to-3D General Object Generation. Inspired by the success of text-to-image (T2I) generation [42, 50, 51, 54, 56, 86], many studies have explored using pre-trained vision-language models [49, 54] to achieve text-guided 3D content generation without large-scale 3D-text paired data. Early methods [18, 37, 39, 58, 69] employ the CLIP model [49] to supervise the alignment of 3D representations with prompts. DreamFusion [47] introduces the Score Distillation Sampling (SDS) loss, significantly enhancing the fidelity of generated 3D content by leveraging more powerful pre-trained diffusion models [54]. Subsequent works further advance text-driven 3D generation by improving 3D representations [5, 6, 80], optimization strategies [29, 68, 81], SDS loss [28, 71], and diffusion models [27, 48, 62]. Despite these advancements, current approaches focused on general content generation do not fully leverage the extensive prior knowledge of human heads and hair, limiting their ability to generate realistic, high-quality 3D head avatars with strand-based hair." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "text", + "content": "Text-to-3D Head Avatar Generation. CLIPFace [1], T2P [87] and Describe3D [74] pioneer zero-shot text-driven 3D head generation by combining CLIP [49] with 3D parametric head models [4, 8, 26, 77]. With the introduction of SDS loss [47], DreamFace [85] and FaceG2E [76] leverage pre-trained T2I diffusion models to greatly improve generation quality within specific domains. HeadEvolver [70] enhances the expressiveness of head mesh deformations by introducing vector fields, while HeadSculpt [12], and HeadArtist [30] employ DMTet [60] instead of traditional meshes to capture geometric details. HeadStudio [94] incorporates 3DGS [20] to produce realistic and animatable 3D heads. HumanNorm [17], further" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13394" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 80, + 70, + 270, + 113 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 270, + 113 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 270, + 113 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 270, + 113 + ], + "type": "table", + "html": "
TaskMethodHead-Hair-DecoupledStrand-Based HairGeometry & TextureNo Large-Scale 3D-Text Paired Data
Text-to-Head[12, 17, 30, 94]XX✓✓
[84]✓X✓✓
Text-to-Hair[28, 48, 62, 80]XX✓✓
[65]X✓XX
Text-to-Head-HairOurs✓✓✓✓
", + "image_path": "cca5e3cb3c15a0fbd9f5d0a3cf79b2163df40126bddf5474c3ba3cc7278996f8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 83, + 114, + 268, + 124 + ], + "lines": [ + { + "bbox": [ + 83, + 114, + 268, + 124 + ], + "spans": [ + { + "bbox": [ + 83, + 114, + 268, + 124 + ], + "type": "text", + "content": "Table 1. Comparison with current related methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 128, + 295, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 128, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 128, + 295, + 258 + ], + "type": "text", + "content": "enhances human-specific fidelity by fine-tuning 2D diffusion models on high-quality human mesh data. Despite these advancements, these approaches treat the head and hair as a holistic model, limiting support for downstream applications such as hairstyle transfer and editing. To enable flexible head-hair-disentangled generation, TECA [84] independently represents the head with mesh and the hair with NeRF [38]. However, these methods focus only on the realistic external appearance of hair without modeling its internal geometric structure, restricting their application for strand-level editing and physics-based simulation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 259, + 296, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 259, + 296, + 546 + ], + "spans": [ + { + "bbox": [ + 56, + 259, + 296, + 546 + ], + "type": "text", + "content": "Strand-Based Hair Creation. Automatic strand-based hair modeling has garnered significant attention from both industry [2, 9, 10] and academia [3, 14, 33, 82], as it enables downstream physics-based rendering and simulation. Such methods can be broadly categorized into two types: (1) reconstructing hair from input images or videos and (2) generating hair based on control conditions. Works on the former one [15, 21, 32, 35, 41, 55, 57, 61, 64, 67, 72, 73, 78, 83, 89-91, 93] relies on classical 3D reconstruction frameworks [20, 36, 38, 45, 59] and incorporates hair-specific features [44] to realize accurate 3D strand reconstruction. However, these methods depend on constrained multi-view images, making hair creation costly and challenging in less controlled settings. Recently, 3D hair generation approaches [13, 92] have emerged using GANs or parametric models, but they struggle with controllable generation. The most relevant work to ours is HAAR [65], which pioneers text-to-strand by training a text-conditioned hair map diffusion model. However, HAAR relies on a large-scale paired dataset including 9825 haircuts and prompts with limited diversity, restricting its ability to generate novel hairstyles beyond the training data. Moreover, HAAR overlooks hair texture and geometry that adapts to specific head shapes, further limiting its application range." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 546, + 295, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 618 + ], + "type": "text", + "content": "Compared to previous methods, StrandHead generates 3D heads featuring strand-level hairstyles. Instead of requiring large-scale hair-text paired data, we achieve realistic and diverse 3D hair strand generation by using 2D/3D human-centric priors. We summarize the main differences between our methods and related works in Tab. 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 628, + 111, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 111, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 111, + 640 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 647, + 141, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 141, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 141, + 659 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "text", + "content": "FLAME [26] is a 3D parametric head model that represents the human head shape, pose, and expression using a compact set of parameters. Given a set of shape parameters " + }, + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "text", + "content": ", pose parameters " + }, + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{pose}}" + }, + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "text", + "content": ", and expression parameters " + }, + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\psi_{\\mathrm{exp}}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 476, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 476, + 83 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 476, + 83 + ], + "type": "text", + "content": "FLAME can create a 3D head mesh " + }, + { + "bbox": [ + 313, + 72, + 476, + 83 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 72, + 476, + 83 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 84, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 84, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 84, + 553, + 156 + ], + "type": "text", + "content": "HumanNorm [17] is a text-driven, high-quality 3D head generation method. In specific, it models the geometry and texture of the human head using DMTet [60] and a texture field. The alignment between the text " + }, + { + "bbox": [ + 313, + 84, + 553, + 156 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 84, + 553, + 156 + ], + "type": "text", + "content": " and the optimized representation " + }, + { + "bbox": [ + 313, + 84, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 84, + 553, + 156 + ], + "type": "text", + "content": " is achieved using the following SDS loss [47] with T2I diffusion models " + }, + { + "bbox": [ + 313, + 84, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 84, + 553, + 156 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 340, + 156, + 553, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 156, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 340, + 156, + 553, + 182 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\theta} \\mathcal {L} _ {S D S} = \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\epsilon_ {\\phi} \\left(x _ {t}; y, t\\right) - \\epsilon\\right) \\frac {\\partial x}{\\partial \\theta} \\right], \\tag {1}", + "image_path": "6ff0d54d54ff6c1dc4dd5dcfd040b50f518f25acac581b5522196cfe8ab21fb2.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 182, + 553, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 553, + 266 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 182, + 553, + 266 + ], + "type": "inline_equation", + "content": "x = g(\\theta)" + }, + { + "bbox": [ + 313, + 182, + 553, + 266 + ], + "type": "text", + "content": " is the image rendered from " + }, + { + "bbox": [ + 313, + 182, + 553, + 266 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 182, + 553, + 266 + ], + "type": "text", + "content": " by a differentiable renderer " + }, + { + "bbox": [ + 313, + 182, + 553, + 266 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 182, + 553, + 266 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 182, + 553, + 266 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 182, + 553, + 266 + ], + "type": "text", + "content": " is the time step, " + }, + { + "bbox": [ + 313, + 182, + 553, + 266 + ], + "type": "inline_equation", + "content": "x_{t} = x + \\epsilon" + }, + { + "bbox": [ + 313, + 182, + 553, + 266 + ], + "type": "text", + "content": " is a noised version of " + }, + { + "bbox": [ + 313, + 182, + 553, + 266 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 182, + 553, + 266 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 182, + 553, + 266 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}(x_{t};y,t)" + }, + { + "bbox": [ + 313, + 182, + 553, + 266 + ], + "type": "text", + "content": " is the denoised image, and " + }, + { + "bbox": [ + 313, + 182, + 553, + 266 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 313, + 182, + 553, + 266 + ], + "type": "text", + "content": " is a weighting function. HumanNorm employs human-specific diffusion models fine-tuned on high-quality human mesh data to replace the general diffusion model, enhancing the fidelity of the generated human heads." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 266, + 554, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 554, + 376 + ], + "type": "text", + "content": "Neural Scalp Textures (NST) [55] is an efficient hair representation where each pixel stores a feature vector conveying the shape information of a single strand at the corresponding scalp location. Using a pre-trained hair strand generator [55] " + }, + { + "bbox": [ + 313, + 266, + 554, + 376 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 266, + 554, + 376 + ], + "type": "text", + "content": ", the low-dimensional 2D neural scalp texture " + }, + { + "bbox": [ + 313, + 266, + 554, + 376 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 266, + 554, + 376 + ], + "type": "text", + "content": " can be decoded into high-dimensional 3D strand polylines " + }, + { + "bbox": [ + 313, + 266, + 554, + 376 + ], + "type": "inline_equation", + "content": "S = \\{s^i\\}_{i=1}^{N_s}" + }, + { + "bbox": [ + 313, + 266, + 554, + 376 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 266, + 554, + 376 + ], + "type": "inline_equation", + "content": "N_s" + }, + { + "bbox": [ + 313, + 266, + 554, + 376 + ], + "type": "text", + "content": " is the number of strands and " + }, + { + "bbox": [ + 313, + 266, + 554, + 376 + ], + "type": "inline_equation", + "content": "s^i = \\{p_j^i\\}_{j=1}^{N_p}" + }, + { + "bbox": [ + 313, + 266, + 554, + 376 + ], + "type": "text", + "content": " consists of " + }, + { + "bbox": [ + 313, + 266, + 554, + 376 + ], + "type": "inline_equation", + "content": "N_p" + }, + { + "bbox": [ + 313, + 266, + 554, + 376 + ], + "type": "text", + "content": " 3D points. This process is formulated as follows: " + }, + { + "bbox": [ + 313, + 266, + 554, + 376 + ], + "type": "inline_equation", + "content": "S = G(T)" + }, + { + "bbox": [ + 313, + 266, + 554, + 376 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 383, + 382, + 394 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 383, + 382, + 394 + ], + "spans": [ + { + "bbox": [ + 314, + 383, + 382, + 394 + ], + "type": "text", + "content": "3.2. Overview" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 400, + 554, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 554, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 554, + 509 + ], + "type": "text", + "content": "Given a text prompt, StrandHead aims to create a realistic 3D head avatar with strand-based hair without relying on large-scale 3D-text paired data. Fig. 2 provides an overview of our three-stage pipeline. First, we generate a FLAME-aligned 3D bald head for accurate hair initialization using human-specific diffusion models (Sec. 3.3). In the subsequent stages, we model the reasonable geometry and realistic texture of 3D hair strands by leveraging rich 2D/3D human-centric priors (Sec. 3.4 and Sec. 3.5)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 515, + 443, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 515, + 443, + 527 + ], + "spans": [ + { + "bbox": [ + 314, + 515, + 443, + 527 + ], + "type": "text", + "content": "3.3. Bald Head Generation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 533, + 554, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 554, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 554, + 641 + ], + "type": "text", + "content": "To achieve accurate hair initialization and optimize 3D hair strands in subsequent steps using human-specific 2D generative priors, we first need to obtain a reasonable and semantic-aligned 3D bald head from the text (Fig. 2-(a)). To achieve this goal, we enhance HumanNorm [17] by incorporating an evolving FLAME [26] model while optimizing the bald head DMTet " + }, + { + "bbox": [ + 313, + 533, + 554, + 641 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{h}}" + }, + { + "bbox": [ + 313, + 533, + 554, + 641 + ], + "type": "text", + "content": ". This provides accurate semantic information and prevents unnatural geometry. Please refer to the Supp. Mat. for details on bald head generation." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 648, + 464, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 464, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 464, + 661 + ], + "type": "text", + "content": "3.4. Hair Geometry Generation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "With a FLAME-aligned bald head, we can generate 3D hair in the scalp area. Specifically, we first initialize the haircut based on hairstyle descriptions, and then further sculpt the 3D hair shape utilizing 2D/3D human-centric priors." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13395" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 71, + 531, + 256 + ], + "blocks": [ + { + "bbox": [ + 82, + 71, + 531, + 256 + ], + "lines": [ + { + "bbox": [ + 82, + 71, + 531, + 256 + ], + "spans": [ + { + "bbox": [ + 82, + 71, + 531, + 256 + ], + "type": "image", + "image_path": "ad99dc90aabd00317097bf4905e68ae7922bcf2888ea155b69f46dcbb32984e5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 257, + 555, + 302 + ], + "lines": [ + { + "bbox": [ + 55, + 257, + 555, + 302 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 555, + 302 + ], + "type": "text", + "content": "Figure 2. StrandHead includes three stages: (a) We first create a FLAME-aligned 3D bald head using the improved HumanNorm [17]. (b) Next, we introduce a differentiable prismatization algorithm to enable human-specific geometry-aware 2D diffusion models to supervise hair shape modeling. Additionally, two losses inspired by 3D hair geometric priors are applied to further regularize the hair geometry. (c) Finally, we use a human-specific normal-conditioned 2D diffusion model to generate lifelike hair textures." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 305, + 296, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 305, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 305, + 296, + 365 + ], + "type": "text", + "content": "Hair Initialization. To achieve reasonable and diverse hair initialization, we first utilize ChatGPT [43] to select the 20 most representative hairstyles from the USC-HairSalon Dataset [15]. We then optimize neural scalp textures to fit the selected hairstyle using the following loss function:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 366, + 296, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 296, + 411 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {f i t}} = \\sum_ {i = 1} ^ {N _ {s}} \\sum_ {j = 1} ^ {N _ {p}} \\| \\hat {p} _ {j} ^ {i} - p _ {j} ^ {i} \\| _ {2} + \\lambda_ {\\mathrm {o r i}} \\left(1 - \\hat {o} _ {j} ^ {i} \\cdot o _ {j} ^ {i}\\right) + \\lambda_ {\\mathrm {c u r}} \\| \\hat {c} _ {j} ^ {i} - c _ {j} ^ {i} \\| _ {1}, \\tag {2}", + "image_path": "00aa8324717f0ff668ada97bf5c879fe795b6e66a4b3708b54432c0d6f530c3b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 411, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 296, + 521 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 411, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\hat{p}_j^i" + }, + { + "bbox": [ + 55, + 411, + 296, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 411, + 296, + 521 + ], + "type": "inline_equation", + "content": "p_j^i" + }, + { + "bbox": [ + 55, + 411, + 296, + 521 + ], + "type": "text", + "content": " are the position of the " + }, + { + "bbox": [ + 55, + 411, + 296, + 521 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 411, + 296, + 521 + ], + "type": "text", + "content": "-th point on the " + }, + { + "bbox": [ + 55, + 411, + 296, + 521 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 411, + 296, + 521 + ], + "type": "text", + "content": "-th polyline of the GT and generated hair, respectively, and " + }, + { + "bbox": [ + 55, + 411, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\hat{o}_j^i" + }, + { + "bbox": [ + 55, + 411, + 296, + 521 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 411, + 296, + 521 + ], + "type": "inline_equation", + "content": "o_j^i" + }, + { + "bbox": [ + 55, + 411, + 296, + 521 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 411, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\hat{c}_j^i" + }, + { + "bbox": [ + 55, + 411, + 296, + 521 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 411, + 296, + 521 + ], + "type": "inline_equation", + "content": "c_j^i" + }, + { + "bbox": [ + 55, + 411, + 296, + 521 + ], + "type": "text", + "content": " denote their orientation and curvature, respectively. Here, orientation " + }, + { + "bbox": [ + 55, + 411, + 296, + 521 + ], + "type": "inline_equation", + "content": "o_j^i = (p_{j+1}^i - p_j^i) / \\|p_{j+1}^i - p_j^i\\|_2" + }, + { + "bbox": [ + 55, + 411, + 296, + 521 + ], + "type": "text", + "content": " represents the direction of change in strand position, and curvature " + }, + { + "bbox": [ + 55, + 411, + 296, + 521 + ], + "type": "inline_equation", + "content": "c_j^i = \\|o_j^i - o_{j-1}^i\\|_2" + }, + { + "bbox": [ + 55, + 411, + 296, + 521 + ], + "type": "text", + "content": " denote the rate of change in strand orientation. Given a hair description, we select an optimal pre-trained NST as initialization and further optimize it in subsequent stages." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 522, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 714 + ], + "type": "text", + "content": "Discussion on Strand-Based Differentiable Rendering. With reasonable hair initialization, the next step is to leverage human-specific 2D diffusion models to further sculpt the hair geometry. However, due to the lack of a stable strand-based differentiable renderer, it is very challenging to optimize hair shape using 2D generative priors like other SDS-based methods. Besides, we expect hair strands to produce mesh-style smooth geometry and texture maps to fully exploit the powerful generative capabilities of 2D diffusion models pre-trained on human mesh data. A feasible alternative is to first differentiably convert the strands into quad meshes [82], similar to NeuralHaircut [64], and then leverage SDS-based mesh optimization frameworks to model hair strands. However, this non-watertight stripe-like mesh easily produces ambiguous normal maps or excessively thin sides (see the zoom image of Fig. 3-(c)), which significantly" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 323, + 304, + 545, + 388 + ], + "blocks": [ + { + "bbox": [ + 323, + 304, + 545, + 388 + ], + "lines": [ + { + "bbox": [ + 323, + 304, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 323, + 304, + 545, + 388 + ], + "type": "image", + "image_path": "ed9e26760e95466fa25acc4d1f5bb5d0d729d50c0068141ef3c4704f96d3870e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 388, + 497, + 396 + ], + "lines": [ + { + "bbox": [ + 372, + 388, + 497, + 396 + ], + "spans": [ + { + "bbox": [ + 372, + 388, + 497, + 396 + ], + "type": "text", + "content": "(b) A strand-to-mesh conversion example" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 322, + 396, + 538, + 486 + ], + "blocks": [ + { + "bbox": [ + 322, + 396, + 538, + 486 + ], + "lines": [ + { + "bbox": [ + 322, + 396, + 538, + 486 + ], + "spans": [ + { + "bbox": [ + 322, + 396, + 538, + 486 + ], + "type": "image", + "image_path": "9e422028a461b3b809ddd48b3bde11da94fc0d77b134209cc6ebed939ac9f01c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 486, + 503, + 495 + ], + "lines": [ + { + "bbox": [ + 365, + 486, + 503, + 495 + ], + "spans": [ + { + "bbox": [ + 365, + 486, + 503, + 495 + ], + "type": "text", + "content": "(c) Differentiable prismatization's advantages" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 495, + 554, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 554, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 554, + 561 + ], + "type": "text", + "content": "Figure 3. The differentiable prismatization algorithm's (a) gradient backpropagation process, (b) a strand-to-mesh conversion example and (c) advantages over NeuralHaircut [64]. Non-watertight quad meshes can easily produce ambiguous normal maps, which significantly reduce the stability of hair shape modeling (see the drifting hair highlighted by the oval dotted box in (c))." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 568, + 554, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 554, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 554, + 616 + ], + "type": "text", + "content": "reduce the optimization stability (see the drifting hair highlighted by the oval dotted box in Fig. 3-(c)). Therefore, a differentiable strand-to-mesh conversion method with superior optimization properties is urgently needed." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "Differentiable Prismatization (DP). To this end, we propose a novel differentiable prismatization algorithm inspired by the internal structure of 3D hair (i.e., hair fiber is a dielectric cylinder covered with tilted scales and with a pigmented interior [19, 34]). This algorithm can efficiently and differentiably convert strands into watertight prismatic meshes with arbitrary thickness and lateral edges to approximate the cylindrical structure of the hair. This enables" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13396" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "smooth backpropagation of gradients from the SDS loss to the 3D strand representation using mesh-based differentiable renderers [22], thereby making it possible to fully exploit 2D generative priors distilled from high-fidelity human mesh data for modeling realistic hair strands. Fig. 3-(a) visualizes the complete gradient backpropagation process." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 146, + 295, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 146, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 146, + 295, + 277 + ], + "type": "text", + "content": "In specific, given a hair strand " + }, + { + "bbox": [ + 55, + 146, + 295, + 277 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 146, + 295, + 277 + ], + "type": "text", + "content": ", our DP converts it into a watertight prismatic mesh with " + }, + { + "bbox": [ + 55, + 146, + 295, + 277 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 146, + 295, + 277 + ], + "type": "text", + "content": " lateral edges and radius " + }, + { + "bbox": [ + 55, + 146, + 295, + 277 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 146, + 295, + 277 + ], + "type": "text", + "content": " through the following five steps: (1) Compute the Initial Normal Vector. (2) Generate " + }, + { + "bbox": [ + 55, + 146, + 295, + 277 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 146, + 295, + 277 + ], + "type": "text", + "content": " Rotated Normals. (3) Translate to Form Lateral Edges. (4) Construct Lateral Faces. (5) Construct Top and Bottom Faces. An example of this conversion is shown in Fig. 3-(b). Unlike quad meshes [64], our DP effectively avoids ambiguous normals, enabling smooth gradient backpropagation and hair strand optimization, as shown in Fig. 3-(c). Further algorithm details are provided in the Supp. Mat." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 279, + 296, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 296, + 327 + ], + "type": "text", + "content": "With the help of our proposed differentiable prismatization algorithm, we optimize the neural scalp texture " + }, + { + "bbox": [ + 55, + 279, + 296, + 327 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 279, + 296, + 327 + ], + "type": "text", + "content": " by utilizing rich human-specific 2D generative priors through the following SDS losses (Fig. 2-(b)):" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 331, + 295, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 331, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 67, + 331, + 295, + 357 + ], + "type": "interline_equation", + "content": "\\nabla_ {T} \\mathcal {L} _ {S D S} ^ {\\mathrm {h n}} = \\mathbb {E} _ {t, \\epsilon} \\left[ \\left(\\epsilon_ {\\phi_ {\\mathrm {h n}}} \\left(n _ {t} ^ {\\mathrm {h + s}}; y _ {\\mathrm {h + s}}, t\\right) - \\epsilon\\right) \\frac {\\partial n ^ {\\mathrm {h + s}}}{\\partial T} \\right], \\tag {3}", + "image_path": "6a8af109849a266ccece38cbdeb6f7efe544364674ffc8df12d0e7e320256c2d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 360, + 295, + 387 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 360, + 295, + 387 + ], + "spans": [ + { + "bbox": [ + 67, + 360, + 295, + 387 + ], + "type": "interline_equation", + "content": "\\nabla_ {T} \\mathcal {L} _ {S D S} ^ {\\mathrm {h d}} = \\mathbb {E} _ {t, \\epsilon} \\left[ \\left(\\epsilon_ {\\phi_ {\\mathrm {h d}}} \\left(d _ {t} ^ {\\mathrm {h + s}}; y _ {\\mathrm {h + s}}, t\\right) - \\epsilon\\right) \\frac {\\partial d ^ {\\mathrm {h + s}}}{\\partial T} \\right], \\tag {4}", + "image_path": "31315ca45913235298f982ec6aa44c77140b58c90c18f62b64608d9beb46d938.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 399, + 295, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 295, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 295, + 481 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 399, + 295, + 481 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{hn}}" + }, + { + "bbox": [ + 55, + 399, + 295, + 481 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 399, + 295, + 481 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{hd}}" + }, + { + "bbox": [ + 55, + 399, + 295, + 481 + ], + "type": "text", + "content": " are the human-specific normal-adapted and depth-adapted diffusion models from Human-Norm [17], " + }, + { + "bbox": [ + 55, + 399, + 295, + 481 + ], + "type": "inline_equation", + "content": "n^{\\mathrm{h + s}}" + }, + { + "bbox": [ + 55, + 399, + 295, + 481 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 399, + 295, + 481 + ], + "type": "inline_equation", + "content": "d^{\\mathrm{h + s}}" + }, + { + "bbox": [ + 55, + 399, + 295, + 481 + ], + "type": "text", + "content": " are the rendered normal and depth maps of the head DMTet " + }, + { + "bbox": [ + 55, + 399, + 295, + 481 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{h}}" + }, + { + "bbox": [ + 55, + 399, + 295, + 481 + ], + "type": "text", + "content": " with hair prismatic mesh " + }, + { + "bbox": [ + 55, + 399, + 295, + 481 + ], + "type": "inline_equation", + "content": "M_p = DP(G(T))" + }, + { + "bbox": [ + 55, + 399, + 295, + 481 + ], + "type": "text", + "content": ", respectively. Here, " + }, + { + "bbox": [ + 55, + 399, + 295, + 481 + ], + "type": "inline_equation", + "content": "DP(\\cdot)" + }, + { + "bbox": [ + 55, + 399, + 295, + 481 + ], + "type": "text", + "content": " denotes the differentiable prismatization operation function, " + }, + { + "bbox": [ + 55, + 399, + 295, + 481 + ], + "type": "inline_equation", + "content": "y_{\\mathrm{h + s}}" + }, + { + "bbox": [ + 55, + 399, + 295, + 481 + ], + "type": "text", + "content": " is the full text description of the head with hair." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 484, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 295, + 555 + ], + "type": "text", + "content": "However, due to the excessive flexibility of hair strands, relying solely on the SDS loss leads to unnatural strand orientations and unrealistic hairstyles (Fig. 9). Furthermore, the lack of constrained multi-view images prevents us from leveraging hair-specific features [44] to provide powerful supervision as in previous works [55]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": "Prior-Driven Losses. To address this issue, we propose two straightforward yet robust prior-driven losses based on observations of the 3D hair strand orientation and curvature to ensure the rationality of the generated hair strands. Specifically, we observe two geometric properties of hair: (1) neighboring strand orientations are highly consistent, and (2) strand curvature is strongly and positively correlated with the curliness of the hairstyle. To validate the properties, we calculate the cosine similarity of adjacent hair strand orientations " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "CS_{\\mathrm{ori}}" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": " and the average curvature " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{mean}}" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": " of 343 hairstyles in the USC-HairSalon dataset [15] using the following equations and visualize their distributions and some examples in Fig. 4:" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 321, + 70, + 428, + 150 + ], + "blocks": [ + { + "bbox": [ + 321, + 70, + 428, + 150 + ], + "lines": [ + { + "bbox": [ + 321, + 70, + 428, + 150 + ], + "spans": [ + { + "bbox": [ + 321, + 70, + 428, + 150 + ], + "type": "image", + "image_path": "64cf141a3d757c21a2f49eb33cc63169c2a91e50e6d29ccfdefa63c0c7c6de40.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 151, + 430, + 159 + ], + "lines": [ + { + "bbox": [ + 333, + 151, + 430, + 159 + ], + "spans": [ + { + "bbox": [ + 333, + 151, + 430, + 159 + ], + "type": "text", + "content": "(a) Distribution of Haircuts' " + }, + { + "bbox": [ + 333, + 151, + 430, + 159 + ], + "type": "inline_equation", + "content": "CS_{\\mathrm{ori}}" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 432, + 70, + 540, + 150 + ], + "blocks": [ + { + "bbox": [ + 432, + 70, + 540, + 150 + ], + "lines": [ + { + "bbox": [ + 432, + 70, + 540, + 150 + ], + "spans": [ + { + "bbox": [ + 432, + 70, + 540, + 150 + ], + "type": "image", + "image_path": "21154d866dc276c48d57f1652882bea263b5d835d6f8129359cd98fba729d170.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 151, + 544, + 158 + ], + "lines": [ + { + "bbox": [ + 443, + 151, + 544, + 158 + ], + "spans": [ + { + "bbox": [ + 443, + 151, + 544, + 158 + ], + "type": "text", + "content": "(b) Distribution of Haircuts' " + }, + { + "bbox": [ + 443, + 151, + 544, + 158 + ], + "type": "inline_equation", + "content": "C_{\\text{mean}}" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 322, + 160, + 381, + 209 + ], + "blocks": [ + { + "bbox": [ + 322, + 160, + 381, + 209 + ], + "lines": [ + { + "bbox": [ + 322, + 160, + 381, + 209 + ], + "spans": [ + { + "bbox": [ + 322, + 160, + 381, + 209 + ], + "type": "image", + "image_path": "106b55f2d6e435ee3533a0ea1403203377d017baff4e427d3afbf7a6b0c05f6c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 209, + 381, + 214 + ], + "lines": [ + { + "bbox": [ + 322, + 209, + 381, + 214 + ], + "spans": [ + { + "bbox": [ + 322, + 209, + 381, + 214 + ], + "type": "inline_equation", + "content": "CS_{\\mathrm{ori}} = 0.9949" + }, + { + "bbox": [ + 322, + 209, + 381, + 214 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{mean}} = 0.0" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 386, + 161, + 451, + 208 + ], + "blocks": [ + { + "bbox": [ + 386, + 161, + 451, + 208 + ], + "lines": [ + { + "bbox": [ + 386, + 161, + 451, + 208 + ], + "spans": [ + { + "bbox": [ + 386, + 161, + 451, + 208 + ], + "type": "image", + "image_path": "c3e4fac1e2faadf5565ed83ca0979ce893b44cfd999b0c7f1e8b22108c3415ff.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 452, + 160, + 499, + 208 + ], + "blocks": [ + { + "bbox": [ + 452, + 160, + 499, + 208 + ], + "lines": [ + { + "bbox": [ + 452, + 160, + 499, + 208 + ], + "spans": [ + { + "bbox": [ + 452, + 160, + 499, + 208 + ], + "type": "image", + "image_path": "80b8a04d78694e2f8701df09132eeaba3f0bb6c953461c285b4c8b9630905698.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 325, + 215, + 365, + 262 + ], + "blocks": [ + { + "bbox": [ + 325, + 215, + 365, + 262 + ], + "lines": [ + { + "bbox": [ + 325, + 215, + 365, + 262 + ], + "spans": [ + { + "bbox": [ + 325, + 215, + 365, + 262 + ], + "type": "image", + "image_path": "664f8c1138d4a077c29e5290468cdf5f593cd7acce4383e23c069f731331aacd.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 327, + 263, + 357, + 268 + ], + "lines": [ + { + "bbox": [ + 327, + 263, + 357, + 268 + ], + "spans": [ + { + "bbox": [ + 327, + 263, + 357, + 268 + ], + "type": "inline_equation", + "content": "CS_{\\mathrm{ori}} = 0.9265" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 353, + 270, + 506, + 278 + ], + "lines": [ + { + "bbox": [ + 353, + 270, + 506, + 278 + ], + "spans": [ + { + "bbox": [ + 353, + 270, + 506, + 278 + ], + "type": "text", + "content": "(c) Some hairstyles from the USC-HairSalon dataset" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 279, + 554, + 312 + ], + "lines": [ + { + "bbox": [ + 313, + 279, + 554, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 554, + 312 + ], + "type": "text", + "content": "Figure 4. Observation of hair geometric features: (1) neighboring strand orientations are highly consistent. (2) strand curvature is strongly and positively related to the haircut curliness." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 366, + 215, + 402, + 262 + ], + "blocks": [ + { + "bbox": [ + 386, + 209, + 457, + 215 + ], + "lines": [ + { + "bbox": [ + 386, + 209, + 457, + 215 + ], + "spans": [ + { + "bbox": [ + 386, + 209, + 457, + 215 + ], + "type": "text", + "content": "1 " + }, + { + "bbox": [ + 386, + 209, + 457, + 215 + ], + "type": "inline_equation", + "content": "CS_{\\mathrm{ori}} = 0.9868" + }, + { + "bbox": [ + 386, + 209, + 457, + 215 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{mean}} = 0.0290" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 366, + 215, + 402, + 262 + ], + "lines": [ + { + "bbox": [ + 366, + 215, + 402, + 262 + ], + "spans": [ + { + "bbox": [ + 366, + 215, + 402, + 262 + ], + "type": "image", + "image_path": "b66a1339d299f12c2485cbce97515ef4a80a0b3dedaf282728378f2025e593cf.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 263, + 398, + 268 + ], + "lines": [ + { + "bbox": [ + 367, + 263, + 398, + 268 + ], + "spans": [ + { + "bbox": [ + 367, + 263, + 398, + 268 + ], + "type": "inline_equation", + "content": "C_{\\text{mean}} = 0.0598" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 408, + 215, + 451, + 262 + ], + "blocks": [ + { + "bbox": [ + 408, + 215, + 451, + 262 + ], + "lines": [ + { + "bbox": [ + 408, + 215, + 451, + 262 + ], + "spans": [ + { + "bbox": [ + 408, + 215, + 451, + 262 + ], + "type": "image", + "image_path": "ea27960902ded73167b12cc263ffcdc40ac65d118713789938f41b2818a15c78.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 407, + 263, + 451, + 268 + ], + "lines": [ + { + "bbox": [ + 407, + 263, + 451, + 268 + ], + "spans": [ + { + "bbox": [ + 407, + 263, + 451, + 268 + ], + "type": "inline_equation", + "content": "CS_{\\mathrm{ori}} = 0.9719" + }, + { + "bbox": [ + 407, + 263, + 451, + 268 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{mean}}" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 452, + 215, + 506, + 262 + ], + "blocks": [ + { + "bbox": [ + 455, + 209, + 541, + 215 + ], + "lines": [ + { + "bbox": [ + 455, + 209, + 541, + 215 + ], + "spans": [ + { + "bbox": [ + 455, + 209, + 541, + 215 + ], + "type": "inline_equation", + "content": "C S_{\\mathrm{ori}} = 0.9605" + }, + { + "bbox": [ + 455, + 209, + 541, + 215 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{mean}} = 0.0493" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 452, + 215, + 506, + 262 + ], + "lines": [ + { + "bbox": [ + 452, + 215, + 506, + 262 + ], + "spans": [ + { + "bbox": [ + 452, + 215, + 506, + 262 + ], + "type": "image", + "image_path": "5b03f838af5eb63107b0f2ea2b8e1fe975a06e775c8e315c3ad5609b6db2c663.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 452, + 263, + 541, + 269 + ], + "lines": [ + { + "bbox": [ + 452, + 263, + 541, + 269 + ], + "spans": [ + { + "bbox": [ + 452, + 263, + 541, + 269 + ], + "type": "inline_equation", + "content": "= 0.0870" + }, + { + "bbox": [ + 452, + 263, + 541, + 269 + ], + "type": "inline_equation", + "content": "CS_{\\mathrm{ori}} = 0.9297" + }, + { + "bbox": [ + 452, + 263, + 541, + 269 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{mean}} = 0.1059" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 355, + 316, + 554, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 316, + 554, + 350 + ], + "spans": [ + { + "bbox": [ + 355, + 316, + 554, + 350 + ], + "type": "interline_equation", + "content": "C S _ {\\text {o r i}} = \\frac {1}{N _ {s} N _ {p}} \\sum_ {i = 1} ^ {N _ {s}} \\sum_ {j = 1} ^ {N _ {p}} \\sum_ {k \\in A (i)} \\frac {o _ {j} ^ {i} \\cdot o _ {j} ^ {k}}{| A (i) |}, \\tag {5}", + "image_path": "e54cee25a61eecb5a6d3dc1fff3d9cd45093b8d6630f30f0acf9d806e64492bb.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 355, + 354, + 553, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 354, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 355, + 354, + 553, + 388 + ], + "type": "interline_equation", + "content": "C _ {\\text {m e a n}} = \\frac {1}{N _ {s} N _ {p}} \\sum_ {i = 1} ^ {N _ {s}} \\sum_ {j = 1} ^ {N _ {p}} c _ {j} ^ {i}, \\tag {6}", + "image_path": "c8ecee0487842d6bc5ce488bb4910d7d233d59efc350ba18208e1cf378169513.jpg" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 395, + 553, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 553, + 418 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 395, + 553, + 418 + ], + "type": "inline_equation", + "content": "A(i)" + }, + { + "bbox": [ + 313, + 395, + 553, + 418 + ], + "type": "text", + "content": " is the set of adjacent strands for the " + }, + { + "bbox": [ + 313, + 395, + 553, + 418 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 395, + 553, + 418 + ], + "type": "text", + "content": "-th strand, and " + }, + { + "bbox": [ + 313, + 395, + 553, + 418 + ], + "type": "inline_equation", + "content": "|A(i)|" + }, + { + "bbox": [ + 313, + 395, + 553, + 418 + ], + "type": "text", + "content": " is the number of neighboring strands." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 419, + 553, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 553, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 553, + 501 + ], + "type": "text", + "content": "As illustrated, over " + }, + { + "bbox": [ + 313, + 419, + 553, + 501 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 313, + 419, + 553, + 501 + ], + "type": "text", + "content": " of hairstyles exhibit an orientation similarity above 0.9, and the average curvature is significantly positively related to curliness, which successfully validates the properties. Based on these observations, we introduce orientation consistency loss and curvature regularization to guide the local and global strand shapes, formulated as follows:" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 381, + 504, + 553, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 504, + 553, + 516 + ], + "spans": [ + { + "bbox": [ + 381, + 504, + 553, + 516 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {o r i}} = 1 - C S _ {\\mathrm {o r i}}, \\tag {7}", + "image_path": "111f411c2437b9d54accad0544f8d1dfc700fd50c78ccff76e0ea1f1618e14ce.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 380, + 519, + 553, + 532 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 519, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 380, + 519, + 553, + 532 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c u r}} = \\left\\| C _ {\\text {m e a n}} - C _ {\\text {t a r g e t}} \\right\\| _ {1}, \\tag {8}", + "image_path": "b72c761b0e6cca934734eda54235acdf223bf848d84f6e28479d7eefdb8c67df.jpg" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 539, + 553, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 553, + 597 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 539, + 553, + 597 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{target}}" + }, + { + "bbox": [ + 313, + 539, + 553, + 597 + ], + "type": "text", + "content": " represents the target average curvature set according to the input hair description. These two losses, inspired by 3D hair geometry priors, regularize the hair shape by supervising the consistency of orientations between adjacent strands and the overall hairstyle's curvature." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 598, + 554, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 598, + 554, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 598, + 554, + 681 + ], + "type": "text", + "content": "Additionally, we introduce a series of losses " + }, + { + "bbox": [ + 313, + 598, + 554, + 681 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bbox}" + }, + { + "bbox": [ + 313, + 598, + 554, + 681 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 598, + 554, + 681 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{face}" + }, + { + "bbox": [ + 313, + 598, + 554, + 681 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 598, + 554, + 681 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{colli}" + }, + { + "bbox": [ + 313, + 598, + 554, + 681 + ], + "type": "text", + "content": ", which prevent the hair from exceeding the bounding box, obscuring the face, and colliding with the head, respectively. These losses can further improve the rationality of hair geometry. Their details are provided in the Supp. Mat. The final loss function for optimizing hair shape is expressed as follows:" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 325, + 685, + 553, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 685, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 685, + 553, + 714 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {h a i r - g e o}} = \\mathcal {L} _ {S D S} ^ {\\mathrm {h d}} + \\lambda_ {S D S} ^ {\\mathrm {h n}} \\mathcal {L} _ {S D S} ^ {\\mathrm {h n}} + \\lambda_ {\\mathrm {o r i}} \\mathcal {L} _ {\\mathrm {o r i}} + \\lambda_ {\\mathrm {c u r}} \\mathcal {L} _ {\\mathrm {c u r}} \\\\ + \\lambda_ {\\text {b b o x}} \\mathcal {L} _ {\\text {b b o x}} + \\lambda_ {\\text {f a c e}} \\mathcal {L} _ {\\text {f a c e}} + \\lambda_ {\\text {c o l l i}} \\mathcal {L} _ {\\text {c o l l i}}. \\tag {9} \\\\ \\end{array}", + "image_path": "8221ddd798f515916f2a172ddc897d35927cf2bfdfa3033a5a83d7466f975f0b.jpg" + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "13397" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 90, + 70, + 154, + 118 + ], + "blocks": [ + { + "bbox": [ + 90, + 70, + 154, + 118 + ], + "lines": [ + { + "bbox": [ + 90, + 70, + 154, + 118 + ], + "spans": [ + { + "bbox": [ + 90, + 70, + 154, + 118 + ], + "type": "image", + "image_path": "fdec83993177d362470f63f08009e56af763091211d580eb5e9011ee5930ad6b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 118, + 153, + 137 + ], + "lines": [ + { + "bbox": [ + 85, + 118, + 153, + 137 + ], + "spans": [ + { + "bbox": [ + 85, + 118, + 153, + 137 + ], + "type": "text", + "content": "\"An elderly Caucasian grandmother with thin lips and a straight, short silver bob.\"" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 160, + 72, + 233, + 121 + ], + "blocks": [ + { + "bbox": [ + 160, + 72, + 233, + 121 + ], + "lines": [ + { + "bbox": [ + 160, + 72, + 233, + 121 + ], + "spans": [ + { + "bbox": [ + 160, + 72, + 233, + 121 + ], + "type": "image", + "image_path": "f0999092568345f52319c6c3be356adc2e521d4df0e83b4437659fef89e95a97.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 121, + 232, + 133 + ], + "lines": [ + { + "bbox": [ + 167, + 121, + 232, + 133 + ], + "spans": [ + { + "bbox": [ + 167, + 121, + 232, + 133 + ], + "type": "text", + "content": "\"Brad Pitt with medium-length, tousled blonde hair.\"" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 242, + 72, + 299, + 120 + ], + "blocks": [ + { + "bbox": [ + 242, + 72, + 299, + 120 + ], + "lines": [ + { + "bbox": [ + 242, + 72, + 299, + 120 + ], + "spans": [ + { + "bbox": [ + 242, + 72, + 299, + 120 + ], + "type": "image", + "image_path": "4f01c4af11edfbae20afbbc640ebc6970c6166a17dbfe3e576cdbc3b8b239b1d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 121, + 302, + 133 + ], + "lines": [ + { + "bbox": [ + 241, + 121, + 302, + 133 + ], + "spans": [ + { + "bbox": [ + 241, + 121, + 302, + 133 + ], + "type": "text", + "content": "\"Chris Hemsworth with side-swept sandy blonde hair.\"" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 315, + 72, + 369, + 121 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 369, + 121 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 369, + 121 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 369, + 121 + ], + "type": "image", + "image_path": "88c9f5b6d8cb5c835936a4e7e4e415d56d48bb0226780cc84d8e42af8ac9700a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 121, + 374, + 133 + ], + "lines": [ + { + "bbox": [ + 315, + 121, + 374, + 133 + ], + "spans": [ + { + "bbox": [ + 315, + 121, + 374, + 133 + ], + "type": "text", + "content": "Lionel Messi with a blonde mohawk hairstyle." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 378, + 72, + 444, + 120 + ], + "blocks": [ + { + "bbox": [ + 378, + 72, + 444, + 120 + ], + "lines": [ + { + "bbox": [ + 378, + 72, + 444, + 120 + ], + "spans": [ + { + "bbox": [ + 378, + 72, + 444, + 120 + ], + "type": "image", + "image_path": "c3b02bf01fe6f15c8616ba0df153e2511b817b0a6717681e2ea54610dc443013.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 121, + 441, + 133 + ], + "lines": [ + { + "bbox": [ + 383, + 121, + 441, + 133 + ], + "spans": [ + { + "bbox": [ + 383, + 121, + 441, + 133 + ], + "type": "text", + "content": "\"Scarlet Johansson with a textured blonde bob.\"" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 459, + 71, + 521, + 118 + ], + "blocks": [ + { + "bbox": [ + 459, + 71, + 521, + 118 + ], + "lines": [ + { + "bbox": [ + 459, + 71, + 521, + 118 + ], + "spans": [ + { + "bbox": [ + 459, + 71, + 521, + 118 + ], + "type": "image", + "image_path": "b924b600bcb8f03d1405a30c0d4184493a189046da16156995f84023d7cd28c6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 452, + 118, + 523, + 137 + ], + "lines": [ + { + "bbox": [ + 452, + 118, + 523, + 137 + ], + "spans": [ + { + "bbox": [ + 452, + 118, + 523, + 137 + ], + "type": "text", + "content": "\"A mature European woman with a straight nose and medium-length, straight auburn hair.\"" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 88, + 141, + 150, + 190 + ], + "blocks": [ + { + "bbox": [ + 88, + 141, + 150, + 190 + ], + "lines": [ + { + "bbox": [ + 88, + 141, + 150, + 190 + ], + "spans": [ + { + "bbox": [ + 88, + 141, + 150, + 190 + ], + "type": "image", + "image_path": "3959cf6a4eaf806e371795c51ced39ec83533c857fe1979a4db0ba07b6471e92.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 106, + 190, + 163, + 202 + ], + "lines": [ + { + "bbox": [ + 106, + 190, + 163, + 202 + ], + "spans": [ + { + "bbox": [ + 106, + 190, + 163, + 202 + ], + "type": "text", + "content": "Beyonce with voluminous, curly honey blonde hair." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 151, + 141, + 182, + 190 + ], + "blocks": [ + { + "bbox": [ + 151, + 141, + 182, + 190 + ], + "lines": [ + { + "bbox": [ + 151, + 141, + 182, + 190 + ], + "spans": [ + { + "bbox": [ + 151, + 141, + 182, + 190 + ], + "type": "image", + "image_path": "b361d81b71aafa0b986da759404cebd173fe277e1f1cc456848f81b0f846a061.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 192, + 141, + 249, + 189 + ], + "blocks": [ + { + "bbox": [ + 192, + 141, + 249, + 189 + ], + "lines": [ + { + "bbox": [ + 192, + 141, + 249, + 189 + ], + "spans": [ + { + "bbox": [ + 192, + 141, + 249, + 189 + ], + "type": "image", + "image_path": "eeb6efaff330d0842e09289795f92d1bac4a814c2663def2daf0b5188767d178.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 196, + 190, + 246, + 202 + ], + "lines": [ + { + "bbox": [ + 196, + 190, + 246, + 202 + ], + "spans": [ + { + "bbox": [ + 196, + 190, + 246, + 202 + ], + "type": "text", + "content": "Dwayne Johnson with a short black Afro." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 261, + 141, + 305, + 189 + ], + "blocks": [ + { + "bbox": [ + 261, + 141, + 305, + 189 + ], + "lines": [ + { + "bbox": [ + 261, + 141, + 305, + 189 + ], + "spans": [ + { + "bbox": [ + 261, + 141, + 305, + 189 + ], + "type": "image", + "image_path": "0babe7ef1075c71af234f22c026dbfab50140b9b2fa0fe572d50c4278ea85b32.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 259, + 190, + 352, + 202 + ], + "lines": [ + { + "bbox": [ + 259, + 190, + 352, + 202 + ], + "spans": [ + { + "bbox": [ + 259, + 190, + 352, + 202 + ], + "type": "text", + "content": "\"A beautiful Caucasian girl with delicate features and long silky golden blonde hair.\"" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 307, + 141, + 351, + 189 + ], + "blocks": [ + { + "bbox": [ + 307, + 141, + 351, + 189 + ], + "lines": [ + { + "bbox": [ + 307, + 141, + 351, + 189 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 351, + 189 + ], + "type": "image", + "image_path": "14363b7e59c3001c0e63fd3818c5dd4bd9c5d33926b2eb92a660b2799a4da647.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 361, + 141, + 391, + 189 + ], + "blocks": [ + { + "bbox": [ + 361, + 141, + 391, + 189 + ], + "lines": [ + { + "bbox": [ + 361, + 141, + 391, + 189 + ], + "spans": [ + { + "bbox": [ + 361, + 141, + 391, + 189 + ], + "type": "image", + "image_path": "732b9149abd68a2c3f88c8f0bd50ddfa89bc475126481a47da2b6e1417bf5a73.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 190, + 429, + 202 + ], + "lines": [ + { + "bbox": [ + 359, + 190, + 429, + 202 + ], + "spans": [ + { + "bbox": [ + 359, + 190, + 429, + 202 + ], + "type": "text", + "content": "A confident Hispanic man with a beard and a curly black afro." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 438, + 140, + 477, + 186 + ], + "blocks": [ + { + "bbox": [ + 438, + 140, + 477, + 186 + ], + "lines": [ + { + "bbox": [ + 438, + 140, + 477, + 186 + ], + "spans": [ + { + "bbox": [ + 438, + 140, + 477, + 186 + ], + "type": "image", + "image_path": "c9ef585d10485663d0345f37f687314a02524d06220df44c67a5955b2d8a1bd8.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 480, + 140, + 521, + 187 + ], + "blocks": [ + { + "bbox": [ + 480, + 140, + 521, + 187 + ], + "lines": [ + { + "bbox": [ + 480, + 140, + 521, + 187 + ], + "spans": [ + { + "bbox": [ + 480, + 140, + 521, + 187 + ], + "type": "image", + "image_path": "d4891a97efda9acd967d9c5e35acf933ab01b5591e45bd31f37464ab5aeb5aeb.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 442, + 187, + 515, + 206 + ], + "lines": [ + { + "bbox": [ + 442, + 187, + 515, + 206 + ], + "spans": [ + { + "bbox": [ + 442, + 187, + 515, + 206 + ], + "type": "text", + "content": "\"A charming Caucasian woman with a straight nose and shoulder-length curly sleek blonde hair.\"" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 90, + 214, + 156, + 328 + ], + "blocks": [ + { + "bbox": [ + 90, + 214, + 156, + 328 + ], + "lines": [ + { + "bbox": [ + 90, + 214, + 156, + 328 + ], + "spans": [ + { + "bbox": [ + 90, + 214, + 156, + 328 + ], + "type": "image", + "image_path": "e1e8939fe535d0fe1adf32a97bd040dc539acf88867d9957f540a3ebc3bde40e.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 331, + 156, + 338 + ], + "lines": [ + { + "bbox": [ + 91, + 331, + 156, + 338 + ], + "spans": [ + { + "bbox": [ + 91, + 331, + 156, + 338 + ], + "type": "text", + "content": "\"A curly dark blonde haircut.\"" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 156, + 215, + 187, + 328 + ], + "blocks": [ + { + "bbox": [ + 156, + 215, + 187, + 328 + ], + "lines": [ + { + "bbox": [ + 156, + 215, + 187, + 328 + ], + "spans": [ + { + "bbox": [ + 156, + 215, + 187, + 328 + ], + "type": "image", + "image_path": "029876642d1f015224e2a3893c2bdf656708b8fe171773eb14295d044be18679.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 188, + 214, + 247, + 327 + ], + "blocks": [ + { + "bbox": [ + 188, + 214, + 247, + 327 + ], + "lines": [ + { + "bbox": [ + 188, + 214, + 247, + 327 + ], + "spans": [ + { + "bbox": [ + 188, + 214, + 247, + 327 + ], + "type": "image", + "image_path": "b4f13ccfae3ccbdf592bf6e1eb5a9e65c05a8d7ff026e4e92a85e52ba1b0995f.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 328, + 242, + 341 + ], + "lines": [ + { + "bbox": [ + 195, + 328, + 242, + 341 + ], + "spans": [ + { + "bbox": [ + 195, + 328, + 242, + 341 + ], + "type": "text", + "content": "\"Medium-length, straight auburn hair.\"" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 261, + 214, + 303, + 260 + ], + "blocks": [ + { + "bbox": [ + 261, + 214, + 303, + 260 + ], + "lines": [ + { + "bbox": [ + 261, + 214, + 303, + 260 + ], + "spans": [ + { + "bbox": [ + 261, + 214, + 303, + 260 + ], + "type": "image", + "image_path": "a8fda1589724736c4684dc24a17abfc5ea15c59dea5dd4c3efa194dc2d776adc.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 261, + 261, + 306, + 273 + ], + "lines": [ + { + "bbox": [ + 261, + 261, + 306, + 273 + ], + "spans": [ + { + "bbox": [ + 261, + 261, + 306, + 273 + ], + "type": "text", + "content": "\"Long way chestnut brown hair.\"" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 320, + 214, + 353, + 260 + ], + "blocks": [ + { + "bbox": [ + 320, + 214, + 353, + 260 + ], + "lines": [ + { + "bbox": [ + 320, + 214, + 353, + 260 + ], + "spans": [ + { + "bbox": [ + 320, + 214, + 353, + 260 + ], + "type": "image", + "image_path": "86166d947a7aa6c3c593494c57a6c28cc8b0edbb88b9dafeacc5eb959f259bdc.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 261, + 353, + 272 + ], + "lines": [ + { + "bbox": [ + 323, + 261, + 353, + 272 + ], + "spans": [ + { + "bbox": [ + 323, + 261, + 353, + 272 + ], + "type": "text", + "content": "\"A short, wavy brown bob.\"" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 370, + 214, + 404, + 260 + ], + "blocks": [ + { + "bbox": [ + 370, + 214, + 404, + 260 + ], + "lines": [ + { + "bbox": [ + 370, + 214, + 404, + 260 + ], + "spans": [ + { + "bbox": [ + 370, + 214, + 404, + 260 + ], + "type": "image", + "image_path": "f9c10bbfcef42bb59080e090bbbac45031de02e22695d7aaee51304e3b3bc21a.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 261, + 407, + 273 + ], + "lines": [ + { + "bbox": [ + 367, + 261, + 407, + 273 + ], + "spans": [ + { + "bbox": [ + 367, + 261, + 407, + 273 + ], + "type": "text", + "content": "\"Long silky golden blonde hair.\"" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 422, + 213, + 451, + 260 + ], + "blocks": [ + { + "bbox": [ + 422, + 213, + 451, + 260 + ], + "lines": [ + { + "bbox": [ + 422, + 213, + 451, + 260 + ], + "spans": [ + { + "bbox": [ + 422, + 213, + 451, + 260 + ], + "type": "image", + "image_path": "e920cbd347fc90a1af3bef2827fddfc0a91f9bc2632e770d14ac46a044d3e633.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 416, + 261, + 459, + 273 + ], + "lines": [ + { + "bbox": [ + 416, + 261, + 459, + 273 + ], + "spans": [ + { + "bbox": [ + 416, + 261, + 459, + 273 + ], + "type": "text", + "content": "\"A light brown mohawk hairstyle.\"" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 475, + 213, + 507, + 260 + ], + "blocks": [ + { + "bbox": [ + 475, + 213, + 507, + 260 + ], + "lines": [ + { + "bbox": [ + 475, + 213, + 507, + 260 + ], + "spans": [ + { + "bbox": [ + 475, + 213, + 507, + 260 + ], + "type": "image", + "image_path": "02dc237ab140ad69005e7cb4d61271fe98c6b8856f817429e8b0393323b0a7d9.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 261, + 511, + 273 + ], + "lines": [ + { + "bbox": [ + 471, + 261, + 511, + 273 + ], + "spans": [ + { + "bbox": [ + 471, + 261, + 511, + 273 + ], + "type": "text", + "content": "\"Side-swept sandy blonde hair.\"" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 264, + 278, + 299, + 329 + ], + "blocks": [ + { + "bbox": [ + 264, + 278, + 299, + 329 + ], + "lines": [ + { + "bbox": [ + 264, + 278, + 299, + 329 + ], + "spans": [ + { + "bbox": [ + 264, + 278, + 299, + 329 + ], + "type": "image", + "image_path": "a74c041fd1bc22d8cc1a7854ab1bd4ce119d9c256766ab717b7c261af25b8597.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 261, + 331, + 306, + 338 + ], + "lines": [ + { + "bbox": [ + 261, + 331, + 306, + 338 + ], + "spans": [ + { + "bbox": [ + 261, + 331, + 306, + 338 + ], + "type": "text", + "content": "\"A curly black afro.\"" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 342, + 555, + 386 + ], + "lines": [ + { + "bbox": [ + 55, + 342, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 555, + 386 + ], + "type": "text", + "content": "Figure 5. Examples of high-fidelity and diverse 3D heads and strand-accurate haircuts generated by our method. The upper visualization includes rendered color and normal maps of the head and hair prismatic meshes. The lower visualization shows the physics-based hair strand rendering result using Blender [10]. For better strand-based visualization, we interpolate generated hair to approximately 10,000 strands and apply a consistent appearance. Please zoom in for detailed views, and refer to the Supp. Mat. for video demonstrations." + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 321, + 280, + 353, + 327 + ], + "blocks": [ + { + "bbox": [ + 321, + 280, + 353, + 327 + ], + "lines": [ + { + "bbox": [ + 321, + 280, + 353, + 327 + ], + "spans": [ + { + "bbox": [ + 321, + 280, + 353, + 327 + ], + "type": "image", + "image_path": "d50f87087e83a7104ec99d896989ed0b2f7836635e64d8fc838332214644ada5.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 329, + 361, + 341 + ], + "lines": [ + { + "bbox": [ + 312, + 329, + 361, + 341 + ], + "spans": [ + { + "bbox": [ + 312, + 329, + 361, + 341 + ], + "type": "text", + "content": "Medium-length wavy caramel-colored hair." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 371, + 279, + 402, + 328 + ], + "blocks": [ + { + "bbox": [ + 371, + 279, + 402, + 328 + ], + "lines": [ + { + "bbox": [ + 371, + 279, + 402, + 328 + ], + "spans": [ + { + "bbox": [ + 371, + 279, + 402, + 328 + ], + "type": "image", + "image_path": "6d7a97f106771f43c430ecda7c3ee2e9337dbc512685888ade40b6bcefa77ffd.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 329, + 402, + 341 + ], + "lines": [ + { + "bbox": [ + 371, + 329, + 402, + 341 + ], + "spans": [ + { + "bbox": [ + 371, + 329, + 402, + 341 + ], + "type": "text", + "content": "\"A side-swept silver haircut.\"" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 417, + 280, + 455, + 327 + ], + "blocks": [ + { + "bbox": [ + 417, + 280, + 455, + 327 + ], + "lines": [ + { + "bbox": [ + 417, + 280, + 455, + 327 + ], + "spans": [ + { + "bbox": [ + 417, + 280, + 455, + 327 + ], + "type": "image", + "image_path": "7c2d2122a1904144ffdeb28709a0d673f30f4c055c630a82fe47bbf9b706ee53.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 328, + 456, + 341 + ], + "lines": [ + { + "bbox": [ + 419, + 328, + 456, + 341 + ], + "spans": [ + { + "bbox": [ + 419, + 328, + 456, + 341 + ], + "type": "text", + "content": "\"A sleek long black hairstyle.\"" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 474, + 281, + 509, + 328 + ], + "blocks": [ + { + "bbox": [ + 474, + 281, + 509, + 328 + ], + "lines": [ + { + "bbox": [ + 474, + 281, + 509, + 328 + ], + "spans": [ + { + "bbox": [ + 474, + 281, + 509, + 328 + ], + "type": "image", + "image_path": "70d595b5a43aac4182a6b019945dfce1aa92e7ca1eafd73f4a620e0745f3be11.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 329, + 514, + 341 + ], + "lines": [ + { + "bbox": [ + 468, + 329, + 514, + 341 + ], + "spans": [ + { + "bbox": [ + 468, + 329, + 514, + 341 + ], + "type": "text", + "content": "\"Medium-length, tousled blonde hair.\"" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 48 + }, + { + "bbox": [ + 55, + 392, + 194, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 194, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 194, + 403 + ], + "type": "text", + "content": "3.5. Hair Texture Generation" + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 55, + 409, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 296, + 445 + ], + "type": "text", + "content": "Next, we fix the optimized hair strand geometry and model the realistic hair strand texture under the supervision of the SDS loss as follows (Fig. 2-(c)):" + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 62, + 451, + 295, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 451, + 295, + 478 + ], + "spans": [ + { + "bbox": [ + 62, + 451, + 295, + 478 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\psi_ {s}} \\mathcal {L} _ {S D S} ^ {\\mathrm {h c}} = \\mathbb {E} _ {t, \\epsilon} \\left(\\epsilon_ {\\phi_ {\\mathrm {h c}}} \\left(c _ {t} ^ {\\mathrm {h + s}}; n ^ {\\mathrm {h + s}}, y _ {\\mathrm {h + s}}, t\\right) - \\epsilon\\right) \\frac {\\partial c ^ {\\mathrm {h + s}}}{\\partial \\psi_ {\\mathrm {s}}}, \\tag {10}", + "image_path": "b044523c4f049aa5b2ce74def103a09f1fd9ae0d4f4278c4e384dab367c0967e.jpg" + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 55, + 483, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 483, + 296, + 567 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{hc}}" + }, + { + "bbox": [ + 55, + 483, + 296, + 567 + ], + "type": "text", + "content": " is the human-specific normal-conditioned diffusion model from HumanNorm [17], " + }, + { + "bbox": [ + 55, + 483, + 296, + 567 + ], + "type": "inline_equation", + "content": "c^{\\mathrm{h + s}}" + }, + { + "bbox": [ + 55, + 483, + 296, + 567 + ], + "type": "text", + "content": " is the rendered color map of the head texture field " + }, + { + "bbox": [ + 55, + 483, + 296, + 567 + ], + "type": "inline_equation", + "content": "\\psi_{\\mathrm{h}}" + }, + { + "bbox": [ + 55, + 483, + 296, + 567 + ], + "type": "text", + "content": " and hair strand texture field " + }, + { + "bbox": [ + 55, + 483, + 296, + 567 + ], + "type": "inline_equation", + "content": "\\psi_{\\mathrm{s}}" + }, + { + "bbox": [ + 55, + 483, + 296, + 567 + ], + "type": "text", + "content": ". Since the vanilla SDS loss often leads to color oversaturation, we replace it with the following MSDS loss [17] to further enhance the texture's realism in later iterations of texture optimization:" + } + ] + } + ], + "index": 54 + }, + { + "bbox": [ + 61, + 572, + 295, + 598 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 572, + 295, + 598 + ], + "spans": [ + { + "bbox": [ + 61, + 572, + 295, + 598 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\psi_ {s}} \\mathcal {L} _ {M S D S} ^ {\\mathrm {h c}} = \\mathbb {E} _ {t, \\epsilon} \\left(H \\left(c _ {t} ^ {\\mathrm {h + s}}; n ^ {\\mathrm {h + s}}, y _ {\\mathrm {h + s}}, t\\right) - \\epsilon\\right) \\frac {\\partial c ^ {\\mathrm {h + s}}}{\\partial \\psi_ {s}}, \\tag {11}", + "image_path": "41d03c36c4dc964bb912fc7a78f6cf79d328ae5d1ab68de84042eed81aedaf72.jpg" + } + ] + } + ], + "index": 55 + }, + { + "bbox": [ + 55, + 605, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 664 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 605, + 296, + 664 + ], + "type": "inline_equation", + "content": "H(\\cdot)" + }, + { + "bbox": [ + 55, + 605, + 296, + 664 + ], + "type": "text", + "content": " denotes the multi-step operation function. Additionally, we propose a strand-aware texture field that models orientation-dependent texture to better generate high-frequency color variations. Please refer to the Supp. Mat. for more their details." + } + ] + } + ], + "index": 56 + }, + { + "bbox": [ + 55, + 672, + 188, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 188, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 188, + 685 + ], + "type": "text", + "content": "3.6. Implementation Details" + } + ] + } + ], + "index": 57 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Our method is implemented using PyTorch [46] and ThreeStudio [11]. Experiments are conducted on an Ubuntu" + } + ] + } + ], + "index": 58 + }, + { + "bbox": [ + 313, + 392, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 555, + 453 + ], + "type": "text", + "content": "server equipped with A6000 GPUs. Starting with a prompt such as \"A man with brown hair\", we first generate a bald head using \"A bald man\". Then, we generate the hair based on \"A man with brown hair\". Further implementation details are available in the Supp. Mat." + } + ] + } + ], + "index": 59 + }, + { + "bbox": [ + 313, + 464, + 395, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 395, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 395, + 477 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 60 + }, + { + "bbox": [ + 313, + 483, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 554, + 544 + ], + "type": "text", + "content": "Some examples of 3D head models with strand-based hair generated by StrandHead are shown in Fig. 5. Due to space limitations, only the key experimental settings and results are presented here. For more results, evaluations, and details, please refer to our Supp. Mat." + } + ] + } + ], + "index": 61 + }, + { + "bbox": [ + 313, + 552, + 441, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 441, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 441, + 565 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 62 + }, + { + "bbox": [ + 313, + 570, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 677 + ], + "type": "text", + "content": "Baselines. We compare StrandHead with the SOTA methods for head avatar and haircut generation. The baseline methods for head generation include text-to-holistic-head works (HeadArtist [30], HumanNorm [17], and HeadStudio [94]) and text-to-decoupled-head work (TECA [84]). For text-to-hair generation, we consider general text-to-3D methods (MVDream [62], GaussianDreamer [80], LucidDreamer [28], and RichDreamer [48]), text-to-strand method (HAAR [65]), and TECA [84]." + } + ] + } + ], + "index": 63 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Dataset Construction. To create the dataset, we use ChatGPT[43] to randomly generate 30 text descriptions of heads with hair. These include 30 descriptions for assessing" + } + ] + } + ], + "index": 64 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13398" + } + ] + } + ], + "index": 65 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 70, + 529, + 235 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 529, + 235 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 529, + 235 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 529, + 235 + ], + "type": "image", + "image_path": "726ab651220bc3ba911020f5f5816018680420bf204b6a632deee9dcd53356b0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 237, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 237, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 555, + 270 + ], + "type": "text", + "content": "Figure 6. Qualitative comparisons with the SOTA methods. Since TECA [84] uses the vanilla NeRF to represent hair, rendering normals is not supported. HAAR [65] generates only the geometry of hair strands, so we first convert the strands into prismatic meshes using differentiable prismatization and then utilize TEXTure [53] to generate texture for visualization and comparison." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 77, + 274, + 270, + 422 + ], + "blocks": [ + { + "bbox": [ + 77, + 274, + 270, + 422 + ], + "lines": [ + { + "bbox": [ + 77, + 274, + 270, + 422 + ], + "spans": [ + { + "bbox": [ + 77, + 274, + 270, + 422 + ], + "type": "image", + "image_path": "e1a3d8f8618aa06101d317578e6207b453c1e9d061269bd34e9ae3766a7c1ab5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 425, + 295, + 459 + ], + "lines": [ + { + "bbox": [ + 55, + 425, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 295, + 459 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison with HAAR [65]. HAAR, which does not model heads, often produces unreasonable hair-head collisions (highlighted in the black box)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 472, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 495 + ], + "type": "text", + "content": "head generation and 30 for evaluating hair generation. The full list of prompts can be found in the Supp. Mat." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "content": "Evaluation Metrics. Current text-to-3D methods typically use CLIP-based metrics to assess text-image alignment and output quality. However, research [16, 31] shows that these metrics struggle to capture fine-grained alignment between 3D content and input prompts, as validated by our experiments in the Supp. Mat. To address this limitation, we draw inspiration from Progressive3D [7] and Barbie [66], using BLIP-VQA [23, 24] and BLIP2-VQA [23, 25] for evaluation. Specifically, we convert each prompt into two questions to separately verify head and hair. We then feed the rendered image of the generated 3D content into the VQA model, asking each question in sequence and using the probability of a “yes” response as our evaluation metric. For example, the head prompt “A man with black hair” is transformed into “Is the person in the picture a man?” and “Does the person in the picture have black hair?” The hair prompt “Black hair” becomes “Is the object in the picture black hair?” Finally, we conduct a user study by randomly" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 340, + 274, + 529, + 381 + ], + "blocks": [ + { + "bbox": [ + 340, + 274, + 529, + 381 + ], + "lines": [ + { + "bbox": [ + 340, + 274, + 529, + 381 + ], + "spans": [ + { + "bbox": [ + 340, + 274, + 529, + 381 + ], + "type": "table", + "html": "
MethodBLIP-VQA ↑BLIP2-VQA ↑GQP ↑TAP ↑
HeadArtist0.76670.96671.002.33
HeadStudio0.78330.88333.333.67
HumanNorm0.70000.95007.677.67
TECA0.73330.950034.3328.33
StrandHead (Ours)0.85000.966753.6758.00
MVDream0.90000.833324.6720.00
GaussianDreamer0.33330.36675.333.00
LucidDreamer0.80000.93335.335.00
RichDreamer0.83330.76674.005.67
TECA0.70000.76671.673.67
HAAR0.63330.20001.332.33
StrandHead (Ours)0.90000.900057.6760.33
", + "image_path": "d5b34825cbf36f128af514b3600ae54b787d474ac956baf7034b65cfc110127f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 383, + 555, + 427 + ], + "lines": [ + { + "bbox": [ + 313, + 383, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 555, + 427 + ], + "type": "text", + "content": "Table 2. Quantitative comparisons with the SOTA methods. The best and second-best results are highlighted in **bolded** and **underlined**, respectively. GQP: generation quality preference (\\%), TAP: text-image alignment preference (\\%)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 432, + 555, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 555, + 468 + ], + "type": "text", + "content": "selecting 10 generated examples and asking 30 volunteers to assess (1) generation quality and (2) text-image alignment, and select the best methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 477, + 495, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 477, + 495, + 490 + ], + "spans": [ + { + "bbox": [ + 313, + 477, + 495, + 490 + ], + "type": "text", + "content": "4.2. Comparisons of Head Generation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 495, + 555, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 555, + 651 + ], + "type": "text", + "content": "As illustrated in the upper section of Tab. 2, StrandHead outperforms all comparison methods across every evaluation metric. The qualitative results on the left side of Fig. 6 further highlight the advantages of our method. Compared to other head generation methods, the head avatars generated by StrandHead exhibit not only more refined facial geometry and texture details but also strand-accurate hair with a plausible appearance that integrates seamlessly with physics-based simulation systems. To the best of our knowledge, StrandHead is the first head generation framework that realizes strand-level hair modeling, a capability that holds substantial potential for advancing human-centric 3D AIGC applications in the industry." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 659, + 492, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 492, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 492, + 672 + ], + "type": "text", + "content": "4.3. Comparisons of Hair Generation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "As shown in the lower part of Tab. 2, our approach surpasses other methods across most evaluation metrics, and is only slightly lower than LucidDreamer (the latest open-source" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13399" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 71, + 275, + 275 + ], + "blocks": [ + { + "bbox": [ + 77, + 71, + 275, + 275 + ], + "lines": [ + { + "bbox": [ + 77, + 71, + 275, + 275 + ], + "spans": [ + { + "bbox": [ + 77, + 71, + 275, + 275 + ], + "type": "image", + "image_path": "c40766afd0a49292a565959bd32f41365bfa422c41ada07c83ffd878d3bc0e45.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 93, + 282, + 254, + 290 + ], + "lines": [ + { + "bbox": [ + 93, + 282, + 254, + 290 + ], + "spans": [ + { + "bbox": [ + 93, + 282, + 254, + 290 + ], + "type": "text", + "content": "(b) Generated strand-based hair under different bald heads" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 68, + 292, + 282, + 303 + ], + "lines": [ + { + "bbox": [ + 68, + 292, + 282, + 303 + ], + "spans": [ + { + "bbox": [ + 68, + 292, + 282, + 303 + ], + "type": "text", + "content": "Figure 8. Analysis of human-specific 2D diffusion models." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 312, + 295, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 295, + 396 + ], + "type": "text", + "content": "text-to-3D method) on BLIP2-VQA, ranking the second. The qualitative comparisons on the right side of Fig. 6 display sample generation results. Compared to general text-to-3D methods, our approach more precisely captures the internal structure of hair strands, resulting in realistic hair-cut geometry and appearance without incorporating incorrect content such as parts of the human head." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 397, + 296, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 397, + 296, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 296, + 493 + ], + "type": "text", + "content": "Compared with HAAR [65], StrandHead relies on robust 2D/3D human-centric priors rather than large-scale paired datasets to supervise hair generation. This not only allows it to generate hairstyles that are uncommon in the dataset (see slicked-back and side-swept hairstyles in Fig. 7), but also models hair texture or geometry that adapts to specific head shapes (Fig. 8-(b)), thereby avoiding unnatural hairhead collisions (see the black box in Fig. 7)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 503, + 149, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 149, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 149, + 516 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 522, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 715 + ], + "type": "text", + "content": "Effectiveness of Human-Specific 2D Generative Priors. We demonstrate the importance of human-specific 2D generative priors from two aspects: (1) As shown in Fig. 8-(a), ignoring 2D supervision fails to generate meaningful hair. While a general 2D diffusion model can slightly improve modeling performance, it still falls short of producing high-quality results. Only by incorporating human-specific 2D generative priors can one achieve realistic textures and reasonable shapes for hairstyles. (2) Fig. 8-(b) illustrates generated hair under varying bald head conditions while maintaining a fixed initial hairstyle and hair prompt. The hairstyles created under the supervision of human-specific 2D diffusion models exhibit clear geometric and textural variations that adapt to specific bald heads. This strongly demonstrates the necessity of considering the bald head for hair strand generation and highlights the advantages of" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 333, + 70, + 423, + 155 + ], + "blocks": [ + { + "bbox": [ + 333, + 70, + 423, + 155 + ], + "lines": [ + { + "bbox": [ + 333, + 70, + 423, + 155 + ], + "spans": [ + { + "bbox": [ + 333, + 70, + 423, + 155 + ], + "type": "image", + "image_path": "caa771c0a1343c0ff68f158b02de5e524c828527c22e7cb49e6e0eaea649f817.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 432, + 76, + 533, + 155 + ], + "blocks": [ + { + "bbox": [ + 432, + 76, + 533, + 155 + ], + "lines": [ + { + "bbox": [ + 432, + 76, + 533, + 155 + ], + "spans": [ + { + "bbox": [ + 432, + 76, + 533, + 155 + ], + "type": "image", + "image_path": "f5fff9e96c6ad00b9b22679b3c6b79fabbe20f13be35b581ca93b3517290221c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 323, + 177, + 425, + 213 + ], + "blocks": [ + { + "bbox": [ + 313, + 156, + 553, + 177 + ], + "lines": [ + { + "bbox": [ + 313, + 156, + 553, + 177 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 553, + 177 + ], + "type": "text", + "content": "Figure 9. Ablation study on (a) orientation consistency loss " + }, + { + "bbox": [ + 313, + 156, + 553, + 177 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ori}}" + }, + { + "bbox": [ + 313, + 156, + 553, + 177 + ], + "type": "text", + "content": " and (b) curvature regularization loss " + }, + { + "bbox": [ + 313, + 156, + 553, + 177 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cur}}" + }, + { + "bbox": [ + 313, + 156, + 553, + 177 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 323, + 177, + 425, + 213 + ], + "lines": [ + { + "bbox": [ + 323, + 177, + 425, + 213 + ], + "spans": [ + { + "bbox": [ + 323, + 177, + 425, + 213 + ], + "type": "image", + "image_path": "e0b563cded1f058245363216826416541b97eaacb3379cab272a282b58689713.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 215, + 479, + 224 + ], + "lines": [ + { + "bbox": [ + 389, + 215, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 389, + 215, + 479, + 224 + ], + "type": "text", + "content": "Figure 10. Failure cases." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 428, + 177, + 545, + 213 + ], + "blocks": [ + { + "bbox": [ + 428, + 177, + 545, + 213 + ], + "lines": [ + { + "bbox": [ + 428, + 177, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 428, + 177, + 545, + 213 + ], + "type": "image", + "image_path": "b601b5b6baea13c04a08fba7c1b13ee2a6eb9e0540e7ae07225ecd5eea0b057c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 228, + 436, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 436, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 436, + 239 + ], + "type": "text", + "content": "StrandHead over HAAR [65]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "text", + "content": "Effectiveness of Prior-Driven Losses. Fig. 9-(a) demonstrates that incorporating " + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ori}}" + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "text", + "content": " provides strong guidance for generating natural hair by supervising the consistency of orientations between adjacent strands. As shown in Fig. 9-(b), our " + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cur}}" + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "text", + "content": " preserves the desired curliness by constraining the overall hairstyle's curvature. In summary, these two losses, inspired by 3D hair geometry priors, effectively enhance the realism and rationality of the generated haircut." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 346, + 388, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 388, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 388, + 358 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 366, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 555, + 605 + ], + "type": "text", + "content": "In this paper, a novel framework StrandHead is proposed to generate realistic hair-disentangled 3D head avatars using text prompts. With a series of 2D generative models pre-trained on human-centric data, StrandHead generates strand-based hair requiring no large-scale text-hair paired data for supervision. This goal is achieved mainly by two approaches: (1) proposing a novel differentiable prismatization method that transforms the hair strands into prismatic meshes, and (2) using the prior-driven losses inspired by the observation of real-world haircut and the statistical features of human-crafted 3D hair data. The former one produces watertight geometry more similar to the original hair strands, boosting a stable optimization with significantly less normal ambiguity. Besides, the latter one well constrains the optimization with reliable guidance to generate reasonable 3D haircuts. Extensive experiments demonstrate that StrandHead obtains the state-of-the-art performance on 3D head avatars and haircuts generation. The generated results are easily implemented in industrial software to produce physical simulation and high-fidelity rendering." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "text", + "content": "Limitations and Future Work. Due to the limited representation capabilities of the strand generator, StrandHead is unable to create highly complex 3D hairstyles (e.g., dreadlocks and ponytails in Fig. 10). Additionally, although the SDS-based method effectively leverages 2D generative priors, its high computational cost restricts practical application. In future work, we aim to address these challenges by incorporating richer hairstyle datasets and exploring deeper hair priors." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13400" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "type": "text", + "content": "Acknowledgement. This work was supported by the National Science Foundation of China (62376121)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 120, + 115, + 133 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 120, + 115, + 133 + ], + "spans": [ + { + "bbox": [ + 56, + 120, + 115, + 133 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 140, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 140, + 296, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 140, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 61, + 140, + 296, + 173 + ], + "type": "text", + "content": "[1] Shivangi Aneja, Justus Thies, Angela Dai, and Matthias Nießner. Clipface: Text-guided editing of textured 3d morphable models. ACM SIGGRAIPH, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 175, + 296, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 175, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 61, + 175, + 296, + 207 + ], + "type": "text", + "content": "[2] Inc. Autodesk. Autodesk maya - 3d animation and modeling software. https://www.autodesk.com/products/maya, 2024.2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 209, + 296, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 209, + 296, + 241 + ], + "spans": [ + { + "bbox": [ + 61, + 209, + 296, + 241 + ], + "type": "text", + "content": "[3] Gaurav Bhokare, Eisen Montalvo, Elie Diaz, and Cem Yuksel. Real-time hair rendering with hair meshes. In ACM SIGGRAIPH, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 244, + 295, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 244, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 61, + 244, + 295, + 277 + ], + "type": "text", + "content": "[4] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 279, + 295, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 279, + 295, + 322 + ], + "spans": [ + { + "bbox": [ + 61, + 279, + 295, + 322 + ], + "type": "text", + "content": "[5] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Int. Conf. Comput. Vis., 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 324, + 296, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 324, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 61, + 324, + 296, + 357 + ], + "type": "text", + "content": "[6] Zilong Chen, Feng Wang, and Huaping Liu. Text-to-3d using gaussian splatting. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 360, + 296, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 360, + 296, + 414 + ], + "spans": [ + { + "bbox": [ + 61, + 360, + 296, + 414 + ], + "type": "text", + "content": "[7] Xinhua Cheng, Tianyu Yang, Jianan Wang, Yu Li, Lei Zhang, Jian Zhang, and Li Yuan. Progressive3d: Progressively local editing for text-to-3d content creation with complex semantic prompts. In Int. Conf. Learn. Represent., 2023. 7, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 416, + 295, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 416, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 61, + 416, + 295, + 449 + ], + "type": "text", + "content": "[8] Hang Dai, Nick Pears, William Smith, and Christian Duncan. Statistical modeling of craniofacial shape and texture. Int. J. Comput. Vis., 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 452, + 295, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 452, + 295, + 473 + ], + "spans": [ + { + "bbox": [ + 61, + 452, + 295, + 473 + ], + "type": "text", + "content": "[9] Inc. Epic Games. Unreal engine - real-time 3d creation tool. https://www.unrealengine.com/, 2024.2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 475, + 295, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 475, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 57, + 475, + 295, + 506 + ], + "type": "text", + "content": "[10] Blender Foundation. Blender - a 3d modelling and rendering software. https://www.blender.org/, 2024.2,3,6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 510, + 295, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 510, + 295, + 575 + ], + "spans": [ + { + "bbox": [ + 57, + 510, + 295, + 575 + ], + "type": "text", + "content": "[11] Yuan-Chen Guo, Ying-Tian Liu, Ruizhi Shao, Christian Laforte, Vikram Voleti, Guan Luo, Chia-Hao Chen, Zixin Zou, Chen Wang, Yan-Pei Cao, and Song-Hai Zhang. threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio, 2023.6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 578, + 295, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 295, + 621 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 295, + 621 + ], + "type": "text", + "content": "[12] Xiaoping Han, Yukang Cao, K. Han, Xiatian Zhu, Jiankang Deng, Yi-Zhe Song, Tao Xiang, and Kwan-Yee K. Wong. Headsculpt: Crafting 3d head avatars with text. In Adv. Neural Inform. Process. Syst., 2023. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 624, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 678 + ], + "type": "text", + "content": "[13] Chengan He, Xin Sun, Zhixin Shu, Fujun Luan, Soren Pirk, Jorge Alejandro Amador Herrera, Dominik L Michels, Tuanfeng Y Wang, Meng Zhang, Holly Rushmeier, and Yi Zhou. Perm: A parametric representation for multi-style 3d hair modeling. arXiv preprint arXiv:2407.19451, 2024. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "text", + "content": "[14] Jerry Hsu, Tongtong Wang, Zherong Pan, Xifeng Gao, Cem Yuksel, and Kui Wu. Real-time physically guided hair interpolation. ACM Trans. Graph., 2024. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[15] Liwen Hu, Chongyang Ma, Linjie Luo, and Hao Li. Single-view hair modeling using a hairstyle database. ACM Trans. Graph., 2015. 3, 4, 5, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 107, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 149 + ], + "type": "text", + "content": "[16] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. In Adv. Neural Inform. Process. Syst., 2023. 7, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 151, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 553, + 205 + ], + "type": "text", + "content": "[17] Xin Huang, Ruizhi Shao, Qi Zhang, Hongwen Zhang, Ying Feng, Yebin Liu, and Qing Wang. Humannorm: Learning normal diffusion model for high-quality and realistic 3d human generation. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3, 4, 5, 6, 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 206, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 249 + ], + "type": "text", + "content": "[18] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 251, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 281 + ], + "type": "text", + "content": "[19] James T Kajiya and Timothy L Kay. Rendering fur with three dimensional textures. ACM Siggraph Computer Graphics, 1989. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 282, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 282, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 282, + 553, + 316 + ], + "type": "text", + "content": "[20] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 2023. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 317, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 317, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 317, + 553, + 348 + ], + "type": "text", + "content": "[21] Zhiyi Kuang, Yiyang Chen, Hongbo Fu, Kun Zhou, and Youyi Zheng. Deepmvshair: Deep hair modeling from sparse views. In ACM SIGGRAPH Asia, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "text", + "content": "[22] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Trans. Graph., 2020. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 394, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 448 + ], + "type": "text", + "content": "[23] Dongxu Li, Junnan Li, Hung Le, Guangsen Wang, Silvio Savarese, and Steven C.H. Hoi. LAVIS: A one-stop library for language-vision intelligence. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), 2023. 7, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "type": "text", + "content": "[24] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, 2022. 7, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 494, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 494, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 494, + 553, + 536 + ], + "type": "text", + "content": "[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International Conference on Machine Learning, 2023. 7, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 537, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 569 + ], + "type": "text", + "content": "[26] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 2017. 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 571, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 553, + 612 + ], + "type": "text", + "content": "[27] Weiyu Li, Rui Chen, Xuelin Chen, and Ping Tan. Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. arXiv preprint arXiv:2310.02596, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "type": "text", + "content": "[28] Yixun Liang, Xin Yang, Jiantao Lin, Haodong Li, Xiaogang Xu, and Yingcong Chen. Luciddreamer: Towards high-fidelity text-to-3d generation via interval score matching. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3, 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[29] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13401" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[30] Hongyu Liu, Xuan Wang, Ziyu Wan, Yujun Shen, Yibing Song, Jing Liao, and Qifeng Chen. Headartist: Text-conditioned 3d head generation with self score distillation. In ACM SIGGRAIPH, 2024. 2, 3, 6, 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[31] Yujie Lu, Xianjun Yang, Xiujun Li, Xin Eric Wang, and William Yang Wang. Llmscore: Unveiling the power of large language models in text-to-image synthesis evaluation. In Adv. Neural Inform. Process. Syst., 2024. 7, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "text", + "content": "[32] Haimin Luo, Min Ouyang, Zijun Zhao, Suyi Jiang, Longwen Zhang, Qixuan Zhang, Wei Yang, Lan Xu, and Jingyi Yu. Gaussianhair: Hair modeling and rendering with light-aware gaussians. arXiv preprint arXiv:2402.10483, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 240 + ], + "type": "text", + "content": "[33] Ryota Maeda, Kenshi Takayama, and Takafumi Taketomi. Refinement of hair geometry by strand integration. In Comput. Graph. Forum, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "type": "text", + "content": "[34] Stephen R Marschner, Henrik Wann Jensen, Mike Cammarano, Steve Worley, and Pat Hanrahan. Light scattering from human hair fibers. ACM Trans. Graph., 2003. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "type": "text", + "content": "[35] Givi Meishvili, James Clemoes, Charlie Hewitt, Zafiirah Hosenie, Xian Xiao, Martin de La Gorce, Tibor Takacs, Tadas Baltrusaitis, Antonio Criminisi, Chyna McRae, et al. Hairmetry: Fairness-aware hairstyle classification. arXiv preprint arXiv:2410.11528, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 331, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 374 + ], + "type": "text", + "content": "[36] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 376, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 419 + ], + "type": "text", + "content": "[37] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 421, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 294, + 464 + ], + "type": "text", + "content": "[38] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Eur. Conf. Comput. Vis., 2020. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 466, + 294, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 294, + 509 + ], + "type": "text", + "content": "[39] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In ACM SIGGRAPH Asia, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 510, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 294, + 544 + ], + "type": "text", + "content": "[40] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 544, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 294, + 578 + ], + "type": "text", + "content": "[41] Giljoo Nam, Chenglei Wu, Min H Kim, and Yaser Sheikh. Strand-accurate multi-view hair capture. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 579, + 294, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 632 + ], + "type": "text", + "content": "[42] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 294, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 655 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 655 + ], + "type": "text", + "content": "[43] OpenAI. Chatgpt. https://openai.com/, 2024.4, 6, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 657, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 657, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 657, + 294, + 689 + ], + "type": "text", + "content": "[44] Sylvain Paris, Hector M Briceno, and François X Sillion. Capture of hair geometry from multiple images. ACM Trans. Graph., 2004. 3, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[45] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning con" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "tinuous signed distance functions for shape representation. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 95, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 149 + ], + "type": "text", + "content": "[46] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Adv. Neural Inform. Process. Syst., 2019. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 151, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 183 + ], + "type": "text", + "content": "[47] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In Int. Conf. Learn. Represent., 2022. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 184, + 553, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 247 + ], + "type": "text", + "content": "[48] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng Bo, and Xiaoguang Han. Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 249, + 553, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 249, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 316, + 249, + 553, + 314 + ], + "type": "text", + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 316, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 360 + ], + "type": "text", + "content": "[50] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 361, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 403 + ], + "type": "text", + "content": "[51] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 404, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 437 + ], + "type": "text", + "content": "[52] Anurag Ranjan, Timo Bolkart, Soubhik Sanyal, and Michael J Black. Generating 3d faces using convolutional mesh autoencoders. In Eur. Conf. Comput. Vis., 2018. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 438, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 470 + ], + "type": "text", + "content": "[53] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In ACM SIGGRAIPH, 2023. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "type": "text", + "content": "[54] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 515, + 553, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 559 + ], + "type": "text", + "content": "[55] Radu Alexandru Rosu, Shunsuke Saito, Ziyan Wang, Chenglei Wu, Sven Behnke, and Giljoo Nam. Neural strands: Learning hair geometry and appearance from multi-view images. In Eur. Conf. Comput. Vis., 2022. 2, 3, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 559, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 624 + ], + "type": "text", + "content": "[56] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In Adv. Neural Inform. Process. Syst., 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 625, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 667 + ], + "type": "text", + "content": "[57] Shunsuke Saito, Liwen Hu, Chongyang Ma, Hikaru Ibayashi, Linjie Luo, and Hao Li. 3d hair synthesis using volumetric variational autoencoders. ACM Trans. Graph., 2018. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "text", + "content": "[58] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13402" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 715 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 115 + ], + "type": "text", + "content": "[59] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In Eur. Conf. Comput. Vis., 2016. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 117, + 295, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 117, + 295, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 117, + 295, + 161 + ], + "type": "text", + "content": "[60] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In Adv. Neural Inform. Process. Syst., 2021. 2, 3, 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 295, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 295, + 205 + ], + "type": "text", + "content": "[61] Yuefan Shen, Shunsuke Saito, Ziyan Wang, Olivier Maury, Chenglei Wu, Jessica Hodgins, Youyi Zheng, and Giljoo Nam. Ct2hair: High-fidelity 3d hair modeling using computed tomography. ACM Trans. Graph., 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 295, + 239 + ], + "type": "text", + "content": "[62] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. In Int. Conf. Learn. Represent., 2023. 2, 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 295, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 295, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 295, + 272 + ], + "type": "text", + "content": "[63] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In Int. Conf. Learn. Represent., 2014. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 272, + 295, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 295, + 316 + ], + "type": "text", + "content": "[64] Vanessa Sklyarova, Jenya Chelishev, Andreea Dogaru, Igor Medvedev, Victor Lempitsky, and Egor Zakharov. Neural haircut: Prior-guided strand-based hair reconstruction. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 2, 3, 4, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 317, + 295, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 317, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 295, + 360 + ], + "type": "text", + "content": "[65] Vanessa Sklyarova, Egor Zakharov, Otmar Hilliges, Michael J Black, and Justus Thies. Text-conditioned generative model of 3d strand-based human hairstyles. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 361, + 295, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 295, + 393 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 295, + 393 + ], + "type": "text", + "content": "[66] Xiaokun Sun, Zhenyu Zhang, Ying Tai, Qian Wang, Hao Tang, Zili Yi, and Jian Yang. Barbie: Text to barbie-style 3d avatars. arXiv preprint arXiv:2408.09126, 2024. 7, 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 393, + 295, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 295, + 448 + ], + "type": "text", + "content": "[67] Yusuke Takimoto, Hikari Takehara, Hiroyuki Sato, Zihao Zhu, and Bo Zheng. Dr. hair: Reconstructing scalp-connected hair strands without pre-training via differentiable rendering of line segments. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 449, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 295, + 491 + ], + "type": "text", + "content": "[68] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splattering for efficient 3d content creation. In Int. Conf. Learn. Represent., 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 492, + 295, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 295, + 536 + ], + "type": "text", + "content": "[69] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 537, + 295, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 295, + 591 + ], + "type": "text", + "content": "[70] Duotun Wang, Hengyu Meng, Zeyu Cai, Zhijing Shao, Qianxi Liu, Lin Wang, Mingming Fan, Xiaohang Zhan, and Zeyu Wang. Headevolver: Text to head avatars via expressive and attribute-preserving mesh deformation. arXiv preprint arXiv:2403.09326, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 592, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 295, + 635 + ], + "type": "text", + "content": "[71] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In Adv. Neural Inform. Process. Syst., 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 636, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 295, + 689 + ], + "type": "text", + "content": "[72] Keyu Wu, Yifan Ye, Lingchen Yang, Hongbo Fu, Kun Zhou, and Youyi Zheng. Neuralhdhair: Automatic high-fidelity hair modeling from a single image using implicit neural representations. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 295, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 715 + ], + "type": "text", + "content": "[73] Keyu Wu, Lingchen Yang, Zhiyi Kuang, Yao Feng, Xutao Han, Yuefan Shen, Hongbo Fu, Kun Zhou, and Youyi Zheng." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 715 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "text", + "content": "Monohair: High-fidelity hair modeling from a monocular video. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[74] Menghua Wu, Hao Zhu, Linjiang Huang, Yi Zhuang, Yuanxun Lu, and Xun Cao. High-fidelity 3d face generation from natural language descriptions. IEEE Conf. Comput. Vis. Pattern Recog., 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "text", + "content": "[75] Sijing Wu, Yichao Yan, Yunhao Li, Yuhao Cheng, Wenhan Zhu, Ke Gao, Xiaobo Li, and Guangtao Zhai. Ganhead: Towards generative animatable neural head avatars. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 197, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 241 + ], + "type": "text", + "content": "[76] Yunjie Wu, Yapeng Meng, Zhipeng Hu, Lincheng Li, Haoqian Wu, Kun Zhou, Weiwei Xu, and Xin Yu. Text-guided 3d face synthesis-from generation to editing. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 242, + 553, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 242, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 242, + 553, + 295 + ], + "type": "text", + "content": "[77] Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu Shen, Ruigang Yang, and Xun Cao. Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction. In IEEE Conf. Comput. Vis. Pattern Recog., 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 297, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 331 + ], + "type": "text", + "content": "[78] Lingchen Yang, Zefeng Shi, Youyi Zheng, and Kun Zhou. Dynamic hair modeling from monocular videos using deep neural networks. ACM Trans. Graph., 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 332, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 387 + ], + "type": "text", + "content": "[79] Tarun Yenamandra, Ayush Tewari, Florian Bernard, Hans-Peter Seidel, Mohamed Elgharib, Daniel Cremers, and Christian Theobalt. i3dmm: Deep implicit 3d morphable model of human heads. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "type": "text", + "content": "[80] Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 444, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 498 + ], + "type": "text", + "content": "[81] Taoran Yi, Jiemin Fang, Zanwei Zhou, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Xinggang Wang, and Qi Tian. Gaussiandreamerpro: Text to manipulable 3d gaussians with highly enhanced quality. arXiv preprint arXiv:2406.18462, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 499, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 499, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 499, + 553, + 521 + ], + "type": "text", + "content": "[82] Cem Yuksel, Scott Schaefer, and John Keyser. Hair meshes. ACM Trans. Graph., 2009. 3, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 522, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 567 + ], + "type": "text", + "content": "[83] Egor Zakharov, Vanessa Sklyarova, Michael Black, Giljoo Nam, Justus Thies, and Otmar Hilliges. Human hair reconstruction with strand-aligned 3d gaussians. In Eur. Conf. Comput. Vis., 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 567, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 611 + ], + "type": "text", + "content": "[84] Hao Zhang, Yao Feng, Peter Kulits, Yandong Wen, Justus Thies, and Michael J. Black. Teca: Text-guided generation and editing of compositional 3d avatars. In International Conference on 3D Vision, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 612, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 612, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 612, + 553, + 656 + ], + "type": "text", + "content": "[85] Longwen Zhang, Qiwei Qiu, Hongyang Lin, Qixuan Zhang, Cheng Shi, Wei Yang, Ye Shi, Sibei Yang, Lan Xu, and Jingyi Yu. Dreamface: Progressive generation of animatable 3d faces under text guidance. ACM Trans. Graph., 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "text", + "content": "[86] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Int. Conf. Comput. Vis., 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "type": "text", + "content": "[87] Rui Zhao, Wei Li, Zhipeng Hu, Lincheng Li, Zhengxia Zou, Zhen Xia Shi, and Changjie Fan. Zero-shot text-to-parameter" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13403" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 409 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "translation for game character auto-creation. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[88] Mingwu Zheng, Hongyu Yang, Di Huang, and Liming Chen. Imface: A nonlinear 3d morphable face model with implicit neural representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "type": "text", + "content": "[89] Yujuan Zheng, Zirong Jin, Moran Li, Haibin Huang, Chongyang Ma, Shuguang Cui, and Xiaoguang Han. Hirstep: Transfer synthetic to real using strand and depth maps for single-view 3d hair modeling. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 240 + ], + "type": "text", + "content": "[90] Yujian Zheng, Yuda Qiu, Leyang Jin, Chongyang Ma, Haibin Huang, Di Zhang, Pengfei Wan, and Xiaoguang Han. Towards unified 3d hair reconstruction from single-view portraits. arXiv preprint arXiv:2409.16863, 2024." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 285 + ], + "type": "text", + "content": "[91] Yi Zhou, Liwen Hu, Jun Xing, Weikai Chen, Han-Wei Kung, Xin Tong, and Hao Li. Hairnet: Single-view hair reconstruction using convolutional neural networks. In *Eur. Conf. Comput. Vis.*, 2018. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "text", + "content": "[92] Yuxiao Zhou, Menglei Chai, Alessandro Pepe, Markus Gross, and Thabo Beeler. Groomen: A high-quality generative hair model using hierarchical latent representations. ACM Trans. Graph., 2023. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 331, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 374 + ], + "type": "text", + "content": "[93] Yuxiao Zhou, Menglei Chai, Daoye Wang, Sebastian Winberg, Erroll Wood, Kripasindhu Sarkar, Markus Gross, and Thabo Beeler. Groomcap: High-fidelity prior-free hair capture. arXiv preprint arXiv:2409.00831, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 376, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 409 + ], + "type": "text", + "content": "[94] Zhen Zhou, Fan Ma, Hehe Fan, and Yi Yang. Headstudio: Text to animatable head avatars with 3d gaussian splatting. In Eur. Conf. Comput. Vis., 2024. 2, 3, 6" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13404" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/StreamDiffusion_ A Pipeline-level Solution for Real-Time Interactive Generation/522f37ba-5716-44cf-af98-f93d9030fe6d_content_list.json b/2025/StreamDiffusion_ A Pipeline-level Solution for Real-Time Interactive Generation/522f37ba-5716-44cf-af98-f93d9030fe6d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d179694778a522445ce42dd8b6a8207c6fb2c857 --- /dev/null +++ b/2025/StreamDiffusion_ A Pipeline-level Solution for Real-Time Interactive Generation/522f37ba-5716-44cf-af98-f93d9030fe6d_content_list.json @@ -0,0 +1,1459 @@ +[ + { + "type": "text", + "text": "StreamDiffusion: A Pipeline-level Solution for Real-time Interactive Generation", + "text_level": 1, + "bbox": [ + 94, + 130, + 898, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Akio Kodaira1* Chenfeng Xu1,* Toshiki Hazama1,* Takanori Yoshimoto2 Kohei Ohno3 Shogo Mitsuhori4 Soichi Sugano5 Hanying Cho6 Zhijian Liu7 Masayoshi Tomizuka1 Kurt Keutzer1", + "bbox": [ + 99, + 179, + 911, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ UC Berkeley $^{2}$ University of Tsukuba $^{3}$ International Christian University $^{4}$ Toyo University $^{5}$ Tokyo Institute of Technology $^{6}$ Tohoku University $^{7}$ MIT https://github.com/cumulo-autumn/StreamDiffusion {akio.kodaira, xuchenfeng} @berkeley.edu", + "bbox": [ + 199, + 232, + 810, + 309 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/500c3f3d1a4914d3e3b8bde2cebd872209cca1259e7a1a2f5baca5f1838c85bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 321, + 254, + 696 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c9161a56f1584030121bb00e4a259cbbd9e1d0b20f60c49afddfeac54db00eef.jpg", + "image_caption": [ + "Figure 1. Interactive generation results with StreamDiffusion." + ], + "image_footnote": [], + "bbox": [ + 254, + 321, + 418, + 696 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bad1c61184cd583945b70eee9af4f1125a13963a99e5105ede4ba6d54c5acb38.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 321, + 580, + 696 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/66a3129c80545bfb81ab3b97a3876729b5138f3d581458bc4e403c0a98fbcaa8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 583, + 321, + 743, + 696 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9ac85bbec7f903bca8fd5782bab9c08b0550d8e135a751d44004bbaa07f12de4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 321, + 906, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 738, + 326, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce StreamDiffusion, a real-time diffusion pipeline designed for streaming image generation. Existing diffusion models are adept at creating images from text or image prompts, yet they often fall short in real-time interaction. This limitation becomes particularly evident in scenarios involving continuous input, such as augmented/virtual reality,", + "bbox": [ + 89, + 771, + 485, + 863 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "video game graphics rendering, live video streaming, and broadcasting, where high throughput is imperative. StreamDiffusion tackles this challenge through a novel pipeline-level system design. It employs unique strategies like batching the denoising process (Stream Batch), residual classifier-free guidance (R-CFG), and stochastic similarity filtering (SSF). Additionally, it seamlessly integrates advanced acceleration technologies for maximum efficiency. Specifically, Stream Batch reformulates the denoising process by eliminating the traditional wait-and-execute approach and utilizing", + "bbox": [ + 511, + 739, + 908, + 891 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* denotes equal contribution \nThis work was done when Toshiki was a remote intern at UC Berkeley", + "bbox": [ + 107, + 875, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12371", + "bbox": [ + 480, + 943, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a batching denoising approach, facilitating fluid and high-throughput streams. This results in $1.5x$ higher throughput compared to the conventional sequential denoising approach. R-CFG significantly addresses inefficiencies caused by repetitive computations during denoising. It optimizes the process to require minimal or no additional computations, leading to speed improvements of up to $2.05x$ compared to previous classifier-free methods. Besides, our stochastic similarity filtering dramatically lowers GPU activation frequency by halting computations for static image flows, achieving a remarkable reduction in computational consumption—2.39 times on an RTX 3060 GPU and 1.99 times on an RTX 4090 GPU, respectively. The synergy of our proposed strategies with established acceleration technologies enables image generation to reach speeds of up to 91.07 fps on a single RTX 4090 GPU, outperforming the throughput of AutoPipeline, developed by Diffusers, by more than 59.56x.", + "bbox": [ + 89, + 90, + 485, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 376, + 222, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, there has been a growing trend in the commercialization of diffusion models [3, 28, 32, 34] for applications within the entertainment industry such as Metaverse, online video streaming, broadcasting, and even the robotic field [7]. A pertinent example is the use of diffusion models to create virtual YouTubers. These digital personas should be capable of reacting in a fluid and responsive manner to user input. These areas require diffusion pipelines that offer high throughput and low latency to ensure the efficient interactive streaming generation.", + "bbox": [ + 89, + 402, + 483, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To advance the efficiency, current efforts primarily focus on reducing the number of denoising steps, such as decreasing from 50 denoise steps to just a few [24, 25] or even one [21, 42]. The strategy includes distilling the multi-step diffusion models into a few steps [36, 40] or re-framing the diffusion process with neural Ordinary Differential Equations (ODE) [22, 23]. Quantization has also been applied to diffusion models [14, 18] to improve efficiency. These methods share the common goal of approximating either the diffusion process itself or the original model weights to achieve efficiency gains. In this paper, we aim at an orthogonal direction and introduce StreamDiffusion, a pipeline-level solution that enables streaming image generation with high throughput. We highlight that existing model design efforts can still be integrated with our pipeline. Our approach enables the use of N-step denoising diffusion models while keeping high throughput and offers users more flexibility in choosing their preferred models.", + "bbox": [ + 89, + 553, + 483, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, StreamDiffusion seamlessly incorporates a suite of novel strategies. Among these, we propose a simple yet novel approach termed Stream Batch. This method differs from the traditional sequential denoising mode, instead of batching the denoising steps. This subtle modifi", + "bbox": [ + 89, + 825, + 483, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cation enhances efficiency without sacrificing the quality of image generation. We highlight Stream Batch enables a new capability of generating images conditioned on future frames in a streaming mode, which is something impossible in previous works. Via injecting the future frames, Stream Batch significantly improves the temporal consistency with few additional overhead. Furthermore, the key novelty of Stream Batch lies not only in its GPU parallelization; rather, it serves as a practical realization of a broader Stream Denoising framework. By denoising inputs at diagonally-offset timesteps within a streaming queue structure, diffusion models naturally achieve continuous, autoregressive generation—emitting one output frame per each newly sampled frame—while benefiting from parallel computation. Crucially, this design generalizes directly to sequential tasks such as video, audio, or robotic action-sequence generation, enabling interactive, unbounded-length synthesis.", + "bbox": [ + 511, + 90, + 906, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Besides, we point out that it is time-consuming for existing diffusion pipelines to use classifier-free guidance for emphasizing the prompts during generation, due to the repetitive and redundant computations for negative conditions. To address this issue, we introduce an innovative approach termed as residual classifier-free guidance (R-CFG). R-CFG approximates the negative condition with a virtual residual noise, which allows us to calculate the negative condition noise only during the initial step of the process. We also indicate that using the original input image latent as the residual term effectively generates results that diverge from the original input image according to the magnitude of the guidance scale, which is a special case of our R-CFG and does not require any computations for the negative condition.", + "bbox": [ + 511, + 351, + 908, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, in real applications such as virtual youtu and AR/VR cases, maintaining the diffusion models always in an active mode is energy-consuming as it keeps hitting GPU. To reduce the energy, we further apply a stochastic similarity filtering (SSF) strategy. In the pipeline, we compute the similarities between continuous inputs and determine whether the diffusion model should process the images based on the probability of an activated similarity. This enables both energy efficiency and visual fluency. In order to further improve the efficiency to cater to the real applications, we apply simple yet effective engineering implementations such as Input-Output Queue (IO-Queue), pre-computing caching, and TensorRT. The overview of the StreamDiffusion pipeline is shown in Fig. 2.", + "bbox": [ + 511, + 565, + 908, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experiments demonstrate that StreamDiffusion can achieve up to 91.07fps for image generation on one RTX4090 GPU, surpassing the diffusion Autopipeline from Diffusers [15] by up to 59.6x. Besides, our stochastic similarity filtering strategy significantly reduces the power usage by 2.39x on one RTX 3090GPU and by 1.99x on one RTX 4090GPU. Our proposed StreamDiffusion is a new diffusion pipeline that is not only efficient but also energy-saving.", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12372", + "bbox": [ + 480, + 943, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3a68d090762dee15ec1178a4deedd8b83d7943bc33592f66b1d28a0c73fab071.jpg", + "image_caption": [ + "Figure 2. The overview of StreamDiffusion. StreamDiffusion combines several key components: (1) Stream Batch efficiently processes the denoising steps in batches. (2) Residual Classifier-Free Guidance approximates the negative condition term to reduces the unnecessary repetitive calculations in the UNet. (3) Stochastic Similarity Filter controls the pass of the image stream by calculating the similarity between frames to eliminate the redundant hit onto GPUs. Furthermore, StreamDiffusion leverages techniques like input-output queues for smooth data flow, cache management for faster processing through pre-calculated embedding, and a tiny VAE model to further contribute to overall efficiency. This synergistic combination allows StreamDiffusion to generate images at high throughputs while consuming minimal energy." + ], + "image_footnote": [], + "bbox": [ + 91, + 99, + 911, + 305 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 89, + 417, + 228, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Efficient Diffusion Models Diffusion models [13, 29, 32, 39] have sparked considerable interest in the commercial sector due to their high-quality image/video generation capabilities. These models have been progressively adapted for various applications, including text-to-image generation [2, 30, 31], image editing [1, 33], video generation [4, 5] and even perception [16, 19, 41]. However, diffusion models are currently limited by their slow speed in generating outputs. In response to this challenge, a variety of strategies have been proposed. One of the mainstreams is to approximate the SDE-based diffusion process [38, 39] through an ordinary differentiable equation (ODE) framework. For example, DPM and $\\mathrm{DPM}++$ [22, 23] introduce ODE-based samplers, which significantly reduce the hundreds of denoising steps to between 15 and 20. Building upon the ODE formulation, InstaFlow [21] advances the reduction of denoising steps to a single instance through the novel strategy of rectified flow [20], while achieving performance close to that of Stable Diffusion [32]. Additionally, distillation from pre-trained diffusion models has also been explored as a method to facilitate few-step denoising. For instance, the consistency model [40] leverages the principle of self-consistency between noise at different denoising steps and uses pre-trained diffusion models [32] to guide the learning of a few-step denoising model, thereby enabling the generation of images within a minimal number of steps. In a notable extension of this concept, LCM [25, 26] applies the idea to the latent space rather than the pixel space. The use of distillation methods [35, 36, 42] to enhance the efficiency of the original Stable Diffusion model presents promising", + "bbox": [ + 91, + 446, + 485, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "results. Besides improving efficiency through the lens of reducing denoising steps, quantization methods [14, 18] are proposed to make the model run in the regime of low float points, albeit with the potential trade-off of fidelity and efficiency. Moreover, parallel sampling [37] tries to utilize the approximate-parallel denoising strategy to improve the latency of the diffusion model. We emphasize that our work focuses on enhancing throughput and our work is orthogonal to [37]. Notably, our method is optimized for single-GPU, which are common among users. In contrast, the parallel sampling method reduces throughput on a single GPU.", + "bbox": [ + 511, + 417, + 908, + 584 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our proposed StreamDiffusion is significantly different from the approaches mentioned previously. While earlier methods primarily focus on the low latency of their individual model designs, our approach takes a different route. We introduce a pipeline-level solution specifically tailored for high throughput. Our pipeline can seamlessly integrate the low-latency diffusion models discussed above. Our proposed Stream Batch, residual classifier-free guidance, and the integration of other efficiency-enhancement methods focus on improving the efficiency of the whole pipeline instead of a single diffusion model.", + "bbox": [ + 511, + 585, + 910, + 752 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Classifier-free Guidance Classifier-free guidance [12] is widely used for conditional generation due to the simplicity, efficiency, and stability compared to classifier-guidance [9]. It leverages negative prompts [8, 11, 32] and essentially operates the vector arithmetic shift in latent space, i.e., we take a step of size (usually set by the guidance scale) away from the unconditional vector or negatively conditioned vector in the direction toward the conditioning manifold [12]. In", + "bbox": [ + 511, + 779, + 911, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12373", + "bbox": [ + 480, + 943, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9396a76707ed26a7c24187047be046b4ca5d318c8ba0c8d50d39a412d3e466fa.jpg", + "image_caption": [ + "Figure 3. The concept of Stream Batch. In our approach, instead of waiting for a single image to be fully denoised before processing the next input image, we accept the next input image after each denoising step. This creates a denoising batch where the denoising steps are staggered for each image. By concatenating these staggered denoising steps into a batch, we can efficiently process continuous inputs using a U-Net for batch processing. The input image encoded at timestep $t$ is generated and decoded at timestep $t + n$ , where $n$ is the number of the denoising steps." + ], + "image_footnote": [], + "bbox": [ + 93, + 85, + 486, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the practical implementation, the classifier-free guidance is conducted by sharing the same UNet for both the conditional term and unconditional (or negative) term and subtracting the effect of the unconditional term from the conditioned one. We point out that the way of multiple denoising processes leads to unnecessary computations for the unconditional (or negative) term. To get rid of these redundant computations, we propose a novel residual-classifier-free guidance, termed as R-CFG, which approximates the conditional noise prediction with only requiring one or even zero-time computation for the negatively conditioned noise prediction for the UNet. We note that $R$ -CFG is especially designed for the SDEdit method [27], as we mainly focus on the applications of translating streaming image flow. Our proposed R-CFG significantly improves the latency of the conditional image-to-image generation.", + "bbox": [ + 88, + 424, + 486, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. StreamDiffusion", + "text_level": 1, + "bbox": [ + 89, + 678, + 253, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "StreamDiffusion is a new diffusion pipeline aiming for high throughput. It comprises three key components: Stream Batch strategy, Residual Classifier-Free Guidance (R-CFG), and Stochastic Similarity Filter. Besides, we also incorporate other acceleration methods like a novel input-output queue designed by us, the pre-computation procedure, the tinyautoencoder, and model acceleration tools such as TensorRT. We elaborate on the details below.", + "bbox": [ + 89, + 703, + 485, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Stream Batch: Batching the Denoise Step", + "text_level": 1, + "bbox": [ + 89, + 833, + 444, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In diffusion models, denoising steps are performed sequentially, resulting in a proportional increase in the processing time of U-Net relative to the number of steps. However,", + "bbox": [ + 89, + 854, + 485, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to generate high-fidelity images, it is necessary to increase the number of steps. To resolve this problem in interactive diffusion, we propose a method called Stream Batch.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Stream Batch technique restructures sequential denoising operations into batched processes, wherein each batch corresponds to a predetermined number of denoising steps, as depicted in Fig. 3. The size of each batch is determined by the number of these denoising steps. This approach allows for each batch element to advance one step further in the denoising sequence via a single pass through U-Net. By iteratively applying this method, it is possible to effectively transform input images encoded at timestep $t$ into their corresponding image-to-image results at timestep $t + n$ , thereby streamlining the denoising procedure.", + "bbox": [ + 511, + 136, + 906, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stream Batch significantly reduces the need for multiple U-Net inferences. The processing time does not escalate linearly with the number of steps. This technique effectively shifts the trade-off from balancing processing time and generation quality to balancing VRAM capacity and generation quality. With adequate VRAM scaling, this method enables the production of high-quality images within the span of a single U-Net processing cycle, effectively overcoming the constraints imposed by increasing denoising steps.", + "bbox": [ + 511, + 301, + 908, + 438 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Waiting and Batching can also increase the throughput of the diffusion pipeline. However, with naive Waiting and Batching (WB), denoising cannot begin immediately on the first input frame, leading to higher latency compared to Stream Batch. We mainly aim for smooth streaming applications. Yet achieving a smooth frame rate with WB requires additional engineering, such as precise inference speed estimation and input-output frame synchronization, and minor timing errors must be carefully managed. In contrast, Stream Batch automatically ensures a consistent interval between input and output frames, providing the advantage of lower latency while dynamically reaching the optimal throughput.", + "bbox": [ + 511, + 438, + 908, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Improve Time Consistency by Stream Batch", + "text_level": 1, + "bbox": [ + 511, + 643, + 885, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Maintaining temporal consistency in video generation is challenging. Many approaches ensure frame coherence by referencing past frames, often through cross-frame attention. However, our Stream Batch method uniquely enables temporal consistency using information from future frames. As shown in Fig. 3, Stream Batch allows simultaneous denoising of multiple frames, passing information from future frames to the current frame. This supports real-time image translation that adapts to sudden changes in input while preserving consistency. In Stream Batch with $n$ denoising steps, keys and values for each frame at each time step form the following batches:", + "bbox": [ + 511, + 665, + 908, + 847 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nK _ {\\text {B a t c h}} = \\left[ K _ {t + i, 0}, \\dots , K _ {t, i}, \\dots , K _ {t - (n - 1 - i), n - 1} \\right]\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 856, + 874, + 876 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {\\mathrm {B a t c h}} = \\left[ V _ {t + i, 0}, \\dots , V _ {t, i}, \\dots , V _ {t - (n - 1 - i), n - 1} \\right]\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 882, + 867, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12374", + "bbox": [ + 478, + 943, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/944cdd937aa8e28e4add4912d4f0c5b19275100ebb59a74d75c0ea6559784250.jpg", + "image_caption": [ + "Figure 4. Virtual residual noise vectors: The orange vectors depict the virtual residual noise that starts from PF ODE trajectory and points to the original input latent $x_0$" + ], + "image_footnote": [], + "bbox": [ + 109, + 87, + 540, + 265 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f3d56de7abc686ed7dda90ad0cc7dafb9ef629e2adab1525a2e356275051e5b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 88, + 885, + 262 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "These key and value batches incorporate information across different time frames and denoising steps. For example, if the frame at time step $t$ has reached the $i$ -th denoising step, the batch includes $i$ future denoising steps for different frames and $n - 1 - i$ past frames. In Stream Batch Cross-frame Attention, rather than using the typical $K_{t,i}$ and $V_{t,i}$ , we employ $K_{\\mathrm{Batch}}$ and $V_{\\mathrm{Batch}}$ , which integrate past and future frame information for the attention computation:", + "bbox": [ + 89, + 330, + 483, + 452 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t n} \\left(Q _ {t, i}, K _ {\\text {B a t c h}}, V _ {\\text {B a t c h}}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {Q _ {t , i} \\cdot K _ {\\text {B a t c h}} ^ {T}}{\\sqrt {d}}\\right) V _ {\\text {B a t c h}}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 474, + 483, + 510 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This approach enables the generation process to account for information from both past and future frames with across denoising stages, thus enhancing temporal consistency.", + "bbox": [ + 89, + 520, + 483, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Residual Classifier-Free Guidance", + "text_level": 1, + "bbox": [ + 89, + 575, + 390, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Firstly, SDEdit based method [27] adds perturbation to the input image $x_0$ and transfers it to the noise distribution $x_{\\tau_0}$", + "bbox": [ + 89, + 597, + 483, + 628 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nx _ {\\tau_ {0}} = \\sqrt {\\alpha_ {\\tau_ {0}}} x _ {0} + \\sqrt {\\beta_ {\\tau_ {0}}} \\epsilon_ {0}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 641, + 482, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha_{\\tau_0}$ and $\\beta_{\\tau_0}$ are values determined by a noise scheduler and $\\epsilon_0$ is a sampled noise from a Gaussian $\\mathcal{N}(0,I)$ . When using consistency models for conditional image editing, $x_{\\tau_0}$ can be considered as a point on the PF ODE trajectory, which leads to the conditioning manifold. To intensify the conditioning by Classifier-Free Guidance (CFG)[10], it is imperative to compute a noise for a negative condition PF ODE trajectory, which is used in vector arithmetic shift for the guidance (Eq. 2).", + "bbox": [ + 89, + 666, + 485, + 801 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {\\tau_ {i}, \\mathrm {c f g}} = \\epsilon_ {\\tau_ {i}, \\bar {c}} + \\gamma \\left(\\epsilon_ {\\tau_ {i}, c} - \\epsilon_ {\\tau_ {i}, \\bar {c}}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 816, + 483, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This requirement introduces additional computational overhead at each denoising step. To reduce this computational overhead, R-CFG utilizes the fact that the original input image $x_0$ is referable at any stage of denoising steps.", + "bbox": [ + 89, + 840, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For any latent $x_{\\tau_i}$ at the denoising step $\\tau_i$ , we can assume the existence of the virtual negative condition $\\vec{c}_{\\tau_i}'$ , that satisfies the self-consistency described as Eq. 3 This implies that $x_{\\tau_i}$ is on the PF ODE trajectory going back to the input image $x_0$ .", + "bbox": [ + 511, + 330, + 906, + 406 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nx _ {0} \\approx \\hat {x} _ {0, \\tau_ {i}, \\bar {c} _ {\\tau_ {i}} ^ {\\prime}} = f _ {\\theta} \\left(x _ {\\tau_ {i}}, \\tau_ {i}, \\bar {c} _ {\\tau_ {i}} ^ {\\prime}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 422, + 906, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following the LCM model parameterization [25] and our approximation for the inference time skip connections $(c_{\\mathrm{skip}}(\\tau) = 0, c_{\\mathrm{out}}(\\tau) = 1$ at $\\tau \\neq 0$ ), the self-consistency equation (Eq. 3) can be expressed as follows,", + "bbox": [ + 511, + 446, + 906, + 507 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nx _ {0} \\approx \\frac {x _ {\\tau_ {i}} - \\sqrt {\\beta_ {\\tau_ {i}}} \\epsilon_ {\\tau_ {i} , \\tilde {c} _ {\\tau_ {i}} ^ {\\prime}}}{\\sqrt {\\alpha_ {\\tau_ {i}}}} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 518, + 906, + 558 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the initial value $x_0$ , and the subsequent values of $x_{\\tau_i}$ obtained sequentially through the iterative denoising, the virtual noise vector $\\epsilon_{\\tau_i,\\vec{c}'}$ in the direction toward the input image can be analytically determined by employing these values with the Eq. 4:", + "bbox": [ + 511, + 564, + 908, + 640 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {\\tau_ {i}, \\bar {c} ^ {\\prime}} = \\frac {x _ {\\tau_ {i}} - \\sqrt {\\alpha_ {\\tau_ {i}}} x _ {0}}{\\sqrt {\\beta_ {\\tau_ {i}}}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 651, + 906, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the virtual noise $\\epsilon_{\\tau_i,\\overline{c'}}$ obtained from Eq. 5, we formulate R-CFG by:", + "bbox": [ + 511, + 691, + 906, + 722 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {\\tau_ {i}, \\text {c f g}} = \\delta \\epsilon_ {\\tau_ {i}, \\bar {c} ^ {\\prime}} + \\gamma \\left(\\epsilon_ {\\tau_ {i}, c} - \\delta \\epsilon_ {\\tau_ {i}, \\bar {c} ^ {\\prime}}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 736, + 906, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\delta$ is a magnitude moderation coefficient for the virtual residual noise that softens the effect and the approximation error of the virtual residual noise.", + "bbox": [ + 511, + 763, + 905, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "R-CFG that uses the original input image latent $x_0$ as the residual term can effectively generate results that diverge from the original input image according to the magnitude of the guidance scale $\\gamma$ , thereby enhancing the effect of conditioning without the need for additional U-Net computations. We call this method Self-Negative R-CFG.", + "bbox": [ + 511, + 810, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12375", + "bbox": [ + 480, + 943, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Not only to deviate from the original input image $x_0$ , but also to diverge from any negative condition, we can find the desired reference point $\\hat{x}_{0,\\tau_0,\\bar{c}}$ that reflects the negative condition $\\bar{c}$ using the same self-consistency formulation:", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} _ {0, \\tau_ {0}, \\bar {c}} = \\frac {x _ {\\tau_ {0}} - \\sqrt {\\beta_ {\\tau_ {0}}} \\epsilon_ {\\tau_ {0} , \\bar {c}}}{\\sqrt {\\alpha_ {\\tau_ {0}}}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 162, + 483, + 200 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We can obtain $\\hat{x}_{0,\\tau_0,\\bar{c}}$ by computing the actual negative conditioned noise $\\epsilon_{\\tau_0,\\bar{c}}$ using U-Net only one time for the first denoising step.", + "bbox": [ + 89, + 205, + 483, + 251 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Eq. 5, instead of $x_0$ , using $\\hat{x}_{0,\\tau_0,\\bar{c}}$ , we can obtain the virtual negative conditioned noise $\\epsilon_{\\tau_{i + 1},\\bar{c}'}$ that can effectively diverge the generation results from the controllable negative conditioning $\\bar{c}$ . We name this method Onetime-Negative R-CFG. In contrast to the conventional CFG, which requires $2n$ computations of U-Net, the Self-Negative RCFG and Onetime-Negative RCFG necessitate only $n$ and $n + 1$ computations of U-Net, respectively, where $n$ is the number of the denoising steps.", + "bbox": [ + 88, + 252, + 483, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Stochastic Similarity Filter", + "text_level": 1, + "bbox": [ + 89, + 397, + 334, + 414 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When images remain unchanged or show minimal changes, particularly in scenarios without active user interaction or static environment, nearly identical input images are often repeatedly fed into the VAE and U-Net. This leads to the generation of identical or nearly identical images and unnecessary consumption of GPU resources. In contexts involving continuous inputs, such instances of unmodified input images can occasionally occur. To tackle this issue and minimize unnecessary computational load, we propose a strategy termed stochastic similarity filter (SSF), as shown in Fig. 2.", + "bbox": [ + 89, + 420, + 483, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We calculate the cosine similarity between the current input image $I_{t}$ and the past reference frame image $I_{\\mathrm{ref}}$ .", + "bbox": [ + 89, + 571, + 483, + 602 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nS _ {C} \\left(I _ {t}, I _ {\\text {r e f}}\\right) = \\frac {I _ {t} \\cdot I _ {\\text {r e f}}}{\\| I _ {t} \\| \\| I _ {\\text {r e f}} \\|} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 612, + 482, + 646 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Based on this cosine similarity, we calculate the probability of skipping the subsequent VAE and U-Net processes. It is given by", + "bbox": [ + 89, + 651, + 483, + 696 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} (\\operatorname {s k i p} \\mid I _ {t}, I _ {\\text {r e f}}) = \\max \\left\\{0, \\frac {S _ {C} \\left(I _ {t} , I _ {\\text {r e f}}\\right) - \\eta}{1 - \\eta} \\right\\}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 720, + 483, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\eta$ is the similarity threshold. This probability decides whether subsequent processes like VAE Encoding, U-Net, and VAE Decoding should be skipped or not. If not skipped, the input image at that time is saved and updated as the reference image $I_{ref}$ for future use. This probabilistic skipping mechanism allows the network to operate fully in dynamic scenes with low inter-frame similarity, while in static scenes with high inter-frame similarity, the network's operational rate decreases, conserving computational resources. The", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5a47214f6b866ccb49e15649e48244f706232c02eb1b3646da105c33aa4555c9.jpg", + "image_caption": [ + "Figure 5. Average inference time comparison between StreamBatch and normal sequential denoising." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 903, + 238 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "GPU usage is modulated seamlessly based on the similarity of the input images, enabling smooth adaptation to scenes with varying dynamics.", + "bbox": [ + 511, + 306, + 903, + 352 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note: We emphasize that compared to determining whether we skip the compute via a hard threshold, the proposed probability-sampling-based similarity filtering strategy leads to smoother video generation. Because the hard threshold is prone to making the video stuck, which hurts the impression of watching video streaming, while the sampling-based method significantly improves the smoothness. For the other efficiency improvement methods, we illustrate them in the supplementary material.", + "bbox": [ + 511, + 352, + 906, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 502, + 645, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We implement StreamDiffusion pipeline upon LCM, LCM-LoRA [25, 26] and SD-turbo [36]. As a model accelerator, we use TensorRT and for the lightweight efficient VAE, we use TAESD [17]. Our pipeline is compatible with the customer-level GPU. We test our pipeline on NVIDIA RTX4090 GPU, Intel Core i9-13900K CPU, Ubuntu 22.04.3 LTS, and NVIDIA RTX3060 GPU, Intel Core i7-12700K, Windows 11 for image generation. We note that we evaluate the throughput mainly via the average inference time per image through processing 100 images.", + "bbox": [ + 511, + 527, + 908, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 511, + 688, + 735, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method with the AutoPipelineForImage2Image, which is a pipeline developed by Huggingface diffusers1. The average inference time comparison is presented in Table. 1. Our pipeline demonstrates a substantial speed increase. When we use TensorRT, StreamDiffusion achieves a minimum speed-up of 13.0 times when running the 10 denoising steps, and reaching up to 59.6 times in scenarios involving a single denoising step. Even though without TensorRT, StreamDiffusion achieves a 29.7 times speed up compared to AutoPipeline when using one step denoising, and an 8.3 times speedup at 10 step denoising.", + "bbox": [ + 511, + 710, + 908, + 878 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1https://github.com/huggingface/diffusers", + "bbox": [ + 529, + 886, + 854, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12376", + "bbox": [ + 480, + 943, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2db1e4dbc923c7c9993b3c9d2320305e94f77688c34b4888438406b129c8301e.jpg", + "table_caption": [ + "Table 1. Average inference time (ms) at different denoising steps. AutoPipeline refers to Diffusers' implementation [15]." + ], + "table_footnote": [], + "table_body": "
StepStreamDiffusionw/o TRTAutoPipeline
110.65 (59.6x)21.34 (29.7x)634.40 (1x)
216.74 (39.3x)30.61 (21.3x)652.66 (1x)
426.93 (25.8x)48.15 (14.4x)695.20 (1x)
1062.00 (13.0x)96.94 (8.3x)803.23 (1x)
", + "bbox": [ + 94, + 128, + 480, + 215 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/61e4b6d93e4b598924d9285137dedfe47e56cd650f9f593b5521e25da319df3e.jpg", + "table_caption": [ + "Table 2. Average inference time (ms) at different denoising steps across CFG methods" + ], + "table_footnote": [], + "table_body": "
StepSelf-Neg. R-CFGOne-Neg. R-CFGCFG
111.04 (1.52x)16.55 (1.01x)16.74 (1x)
216.61 (1.64x)20.64 (1.32x)27.18 (1x)
320.64 (1.74x)27.25 (1.32x)35.91 (1x)
426.19 (1.90x)31.65 (1.57x)49.71 (1x)
531.47 (2.05x)36.04 (1.79x)64.64 (1x)
", + "bbox": [ + 94, + 268, + 483, + 367 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Efficiency comparison regarding Stream Batch. The efficiency comparison between Stream Batch and the original sequential U-Net loop is shown in Fig. 5. When implementing a denoising batch strategy, we observe a significant improvement in processing time. It achieves a reduction by half when compared to a conventional U-Net loop at sequential denoising steps.", + "bbox": [ + 89, + 393, + 485, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Efficiency comparison regarding R-CFG. Table 2 presents a comparison of the inference times for StreamDiffusion pipelines with R-CFG and conventional CFG. The additional computations required to apply Self-Negative R-CFG are merely lightweight vector operations, resulting in negligible changes in inference time compared to when Self-Negative is not used. When employing Onetime-Negative R-CFG, additional UNet computations are necessary for the first step of the denoising process. Therefore, One-time-negative R-CFG and conventional CFG have almost identical inference times for a single denoising step case. However, as the number of denoising steps increases, the difference in inference time from conventional CFG to both Self-Negative and Onetime-Negative R-CFG becomes more pronounced. At denoising step 5, a speed improvement of $2.05\\mathrm{x}$ is observed with Self-Negative R-CFG and $1.79\\mathrm{x}$ with Onetime-Negative R-CFG, compared to conventional CFG.", + "bbox": [ + 89, + 520, + 485, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Energy Consumption", + "text_level": 1, + "bbox": [ + 89, + 787, + 290, + 804 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We then conduct a comprehensive evaluation of the energy consumption associated with our proposed stochastic similarity filter (SSF), as depicted in Figure. 11 and Figure. 12. These figures provide the GPU utilization patterns when SSF (Threshold $\\eta$ set at 0.98) is applied to input videos containing scenes with periodic static characteristics. The comparative", + "bbox": [ + 89, + 810, + 485, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "analysis reveals that the incorporation of SSF significantly mitigates GPU usage in instances where the input images are predominantly static or with a high degree of similarity.", + "bbox": [ + 511, + 90, + 906, + 136 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure. 11 delineates the results derived from a meticulously executed two-denoise-step img2img experiment. This experiment was conducted on a 20-frame video sequence, employing NVIDIA RTX3060 graphics processing units with or without the integration of SSF. The experiment results indicate a substantial decrease in average power consumption from 85.96w to 35.91w on one RTX3060 GPU. Using the same static scene input video with one NVIDIA RTX4090GPU, the power consumption was reduced from 238.68w to 119.77w.", + "bbox": [ + 511, + 137, + 908, + 287 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, Figure. 12 expounds on the findings from a similar two-denoise-step img2img experiment using one RTX4090GPU. This time the evaluation of energy consumption is performed on a 1000-frame video featuring dynamic scenes. Remarkably, even under drastically dynamic conditions, the SSF efficiently extracted several frames exhibiting similarity from the dynamic sequence. This process results in a noteworthy reduction in average power consumption, from 236.13w to 199.38w. These findings underscore the efficacy of the Stochastic Similarity Filter in enhancing energy efficiency, particularly in scenarios involving static or minimally varying visual content.", + "bbox": [ + 511, + 289, + 910, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation study", + "text_level": 1, + "bbox": [ + 513, + 482, + 663, + 500 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6a23082892a1ecbe77c3aef8b58969b4ab462c0bc315df88c90cb19ca7d60c27.jpg", + "image_caption": [ + "Figure 6. Ablation study on different components." + ], + "image_footnote": [], + "bbox": [ + 516, + 518, + 901, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In our ablation study, as summarized in Fig. 6, we evaluate the average inference time of our proposed method under various configurations to understand the contribution of each component. Our proposed StreamDiffusion achieves an average inference time of $10.98 / 9.42\\mathrm{ms}$ and $26.93 / 26.30\\mathrm{ms}$ for denoising steps 1 and 4 on image-to-image/text-to-image generation, respectively. When the R-CFG is not used, we observe this results in the second largest efficiency drop. This demonstrates that R-CFG is one of the most critical", + "bbox": [ + 511, + 763, + 910, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12377", + "bbox": [ + 480, + 943, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "components in our pipeline. When the Stream Batch processing is removed ('w/o stream batch'), we observe a large time consumption increase, especially at 4 denoising steps. We also evaluate the impact on the inference time of our pipeline regarding SSF. We observe that SSF plays a significant role in enabling energy saving and does not introduce extra time cost.", + "bbox": [ + 89, + 90, + 480, + 195 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Besides, the absence of TensorRT ('w/o TRT') leads to a large increase in time cost. The removal of pre-computation also results in increased time cost but not much. We attribute the reason to the limited number of key-value computations in Stable Diffusion. Besides, the exclusion of input-output queue ('w/o IO queue') also demonstrates an impact on average inference time, which mainly aims to optimize the parallelization issue resulting from pre- and post-processing. In the AutoPipelineImage2Image's adding noise function, the precision of tensors is converted from fp32 to fp16 for each request, leading to a decrease in speed. In contrast, the StreamDiffusion pipeline standardizes the precision of variables and computational devices beforehand. It does not perform tensor precision conversion or computation device transfers during inference. Consequently, even without any optimization ('w/o any optimization'), our pipeline significantly outperforms the AutoPipelineImage2Image in terms of throughputs.", + "bbox": [ + 91, + 196, + 486, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Quantitative Evaluation for the Image Quality.", + "text_level": 1, + "bbox": [ + 89, + 479, + 480, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct a quantitative evaluation on the image quality. Specifically, we first evaluate the FID and CLIP scores on the text-to-image generation. We use the same dataset as [6] for the evaluation. We use LCM as our main baseline for the comparison. Note that our method is never trained; our method still improves LCM by a large margin in terms of the FID (29.69 vs. 26.79) and maintains a similar CLIP score (24.95 vs. 24.99). This demonstrates the effectiveness of our proposed method.", + "bbox": [ + 89, + 502, + 482, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "User study. We conduct a user study to validate the visual quality of different components of our StreamDiffusion. The results are shown in the Table. 3. Our Stream Batch with future-frame attention significantly enhances time consistency compared to its absence. Additionally, our SSF method addresses the crucial yet rarely explored issue of energy efficiency. While SSF is simple, it is far from trivial: a common approach might apply a hard threshold on similarity to regulate streaming flow. However, as noted in the main text, we introduce a novel approach—using probability sampling to achieve superior streaming quality. As Table 3 illustrates, this approach is preferred by more users over the hard-threshold method for visual quality. Moreover, compared to the vanilla CFG method, both our self-Negative R-CFG and one-time R-CFG are more preferred by the users, demonstrating our method not only improves efficiency but", + "bbox": [ + 89, + 659, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "also visual quality.", + "bbox": [ + 513, + 90, + 640, + 106 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/46fadac35959cc0a3713ea2d7888a0800f26f84406b456f295e471980257b6c2.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. User study results: Future-frame attention consistency (486 users), SSF quality imperceptibility (144 users), streaming quality between hard threshold filter and SSF (45 users), comparison of with/without CFG and self-negative/one-time R-CFG (257 users)." + ], + "table_body": "
# UsersWithout Future-Frame (%)With Future-Frame (%)
4869.6790.33
# UsersWithout SSF (%)With SSF (%)
14448.6151.39
# UsersHard Threshold (%)SSF (%)
4524.4475.56
# UsersWith / Without CFG (%)Self-Neg R-CFG / One-Time R-CFG (%)
25722 / 2035 / 23
", + "bbox": [ + 516, + 125, + 903, + 239 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This pipeline enables image generation with very low throughput from input images received in real-time from cameras or screen capture devices. At the same time, it is capable of producing high-quality images that effectively align to the specified prompt conditions. These capabilities demonstrate the applicability of our pipeline in various real-time applications, such as real-time game graphic rendering, generative camera effect filters, real-time face conversion, and AI-assisted drawing.", + "bbox": [ + 511, + 321, + 906, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 479, + 632, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose StreamDiffusion, a pipeline-level solution for interactive diffusion generation. StreamDiffusion consists of several optimization strategies for both throughput and GPU usage, including Stream Batch with cross-attention, residual classifier-free guidance (R-CFG), IO-queue for parallelization, stochastic similarity filter, pre-computation, Tiny AutoEncoder, and the use of the model acceleration tool. The synergistic combination of these elements results in a marked improvement in efficiency. Specifically, StreamDiffusion achieves up to 91.07 frames per second (fps) on a standard consumer-grade GPU for image generation tasks. This performance level is particularly beneficial for a variety of applications, including but not limited to the Metaverse, online video streaming, and broadcasting sectors. Furthermore, StreamDiffusion demonstrates a significant reduction in GPU power consumption, achieving at least a $1.99\\mathrm{x}$ decrease. This notable efficiency gain underscores StreamDiffusion's potential for commercial application, offering a compelling solution for energy-conscious, high-performance computing environments.", + "bbox": [ + 511, + 507, + 908, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 830, + 692, + 847 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We sincerely thank Taku Fujimoto and Huggingface team for their invaluable feedback, courteous support, and insightful discussions.", + "bbox": [ + 511, + 857, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12378", + "bbox": [ + 480, + 943, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[2] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM Trans. Graph., 2023. 3", + "[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, Wesam Manassra, Prafulla Dhariwal, Casey Chu, Yunxin Jiao, and Aditya Ramesh. Improving image generation with better captions. 2", + "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3", + "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3", + "[6] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 8", + "[7] Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. arXiv preprint arXiv:2303.04137, 2023. 2", + "[8] Katherine Crowson, Stella Biderman, Daniel Kornis, Dashiell Stander, Eric Hallahan, Louis Castricato, and Edward Raff. Vqgan-clip: Open domain image generation and editing with natural language guidance. In European Conference on Computer Vision, pages 88–105. Springer, 2022. 3", + "[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 3", + "[10] Yilun Du, Shuang Li, and Igor Mordatch. Compositional visual generation with energy based models. In Advances in Neural Information Processing Systems, pages 6637-6647. Curran Associates, Inc., 2020. 5", + "[11] Yilun Du, Shuang Li, and Igor Mordatch. Compositional visual generation with energy based models. Advances in Neural Information Processing Systems, 33:6637-6647, 2020. 3", + "[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3", + "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3", + "[14] Yushi Huang, Ruihao Gong, Jing Liu, Tianlong Chen, and Xianglong Liu. Tfmq-dm: Temporal feature maintenance quantization for diffusion models. arXiv preprint arXiv:2311.16503, 2023. 2, 3", + "[15] Huggingface. https://huggingface.co/docs/diffusers/, 2024.2, 7" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Aliasghar Khani, Saeid Asgari, Aditya Sanghi, Ali Mahdavi Amiri, and Ghassan Hamarneh. Slime: Segment like me. In The Twelfth International Conference on Learning Representations, 2024. 3", + "[17] Diederik P Kingma and Max Welling. Auto-encoding variational bayes, 2022. 6, 13", + "[18] Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-diffusion: Quantizing diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 17535-17545, 2023. 2, 3", + "[19] Ziyi Li, Qinye Zhou, Xiaoyun Zhang, Ya Zhang, Yanfeng Wang, and Weidi Xie. Open-vocabulary object segmentation with diffusion models. 2023. 3", + "[20] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 3", + "[21] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, and Qiang Liu. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. arXiv preprint arXiv:2309.06380, 2023. 2, 3", + "[22] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 2, 3", + "[23] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 2, 3", + "[24] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference, 2023. 2", + "[25] Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module. arXiv preprint arXiv:2311.05556, 2023. 2, 3, 5, 6, 11", + "[26] Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module, 2023. 3, 6", + "[27] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 4, 5", + "[28] OpenAI. https://openai.com/sora, 2024.2", + "[29] William Peebles and Saining Xie. Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748, 2022. 3", + "[30] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2): 3, 2022. 3", + "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 3, 13" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12379", + "bbox": [ + 480, + 943, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3", + "[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. 2022. 3", + "[34] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho†, David Fleet, and Mohammad Norouzi. Imagen: unprecedented photorealism × deep level of language understanding. 2022. 2", + "[35] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 3", + "[36] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 2, 3, 6", + "[37] Andy Shih, Suneel Belkhale, Stefano Ermon, Dorsa Sadigh, and Nima Anari. Parallel sampling of diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[38] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 3", + "[39] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3", + "[40] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023. 2, 3", + "[41] Chenfeng Xu, Huan Ling, Sanja Fidler, and Or Litany. 3diff- tection: 3d object detection with geometry-aware diffusion features, 2023. 3", + "[42] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Frédo Durand, William T. Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. arXiv, 2023. 2, 3" + ], + "bbox": [ + 91, + 92, + 485, + 684 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12380", + "bbox": [ + 480, + 943, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/StreamDiffusion_ A Pipeline-level Solution for Real-Time Interactive Generation/522f37ba-5716-44cf-af98-f93d9030fe6d_model.json b/2025/StreamDiffusion_ A Pipeline-level Solution for Real-Time Interactive Generation/522f37ba-5716-44cf-af98-f93d9030fe6d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cd8a7e121a1d3ff756fadf2db7b51653535f87d4 --- /dev/null +++ b/2025/StreamDiffusion_ A Pipeline-level Solution for Real-Time Interactive Generation/522f37ba-5716-44cf-af98-f93d9030fe6d_model.json @@ -0,0 +1,1914 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.096, + 0.131, + 0.9, + 0.152 + ], + "angle": 0, + "content": "StreamDiffusion: A Pipeline-level Solution for Real-time Interactive Generation" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.18, + 0.912, + 0.218 + ], + "angle": 0, + "content": "Akio Kodaira1* Chenfeng Xu1,* Toshiki Hazama1,* Takanori Yoshimoto2 Kohei Ohno3 Shogo Mitsuhori4 Soichi Sugano5 Hanying Cho6 Zhijian Liu7 Masayoshi Tomizuka1 Kurt Keutzer1" + }, + { + "type": "text", + "bbox": [ + 0.2, + 0.233, + 0.812, + 0.31 + ], + "angle": 0, + "content": "\\(^{1}\\)UC Berkeley \\(^{2}\\)University of Tsukuba \\(^{3}\\)International Christian University \\(^{4}\\)Toyo University \\(^{5}\\)Tokyo Institute of Technology \\(^{6}\\)Tohoku University \\(^{7}\\)MIT https://github.com/cumulo-autumn/StreamDiffusion {akio.kodaira, xuchenfeng} @berkeley.edu" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.322, + 0.255, + 0.697 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.322, + 0.419, + 0.697 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.42, + 0.322, + 0.581, + 0.697 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.322, + 0.744, + 0.697 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.322, + 0.907, + 0.697 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.314, + 0.709, + 0.68, + 0.723 + ], + "angle": 0, + "content": "Figure 1. Interactive generation results with StreamDiffusion." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.739, + 0.327, + 0.755 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.772, + 0.486, + 0.864 + ], + "angle": 0, + "content": "We introduce StreamDiffusion, a real-time diffusion pipeline designed for streaming image generation. Existing diffusion models are adept at creating images from text or image prompts, yet they often fall short in real-time interaction. This limitation becomes particularly evident in scenarios involving continuous input, such as augmented/virtual reality," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.741, + 0.909, + 0.892 + ], + "angle": 0, + "content": "video game graphics rendering, live video streaming, and broadcasting, where high throughput is imperative. StreamDiffusion tackles this challenge through a novel pipeline-level system design. It employs unique strategies like batching the denoising process (Stream Batch), residual classifier-free guidance (R-CFG), and stochastic similarity filtering (SSF). Additionally, it seamlessly integrates advanced acceleration technologies for maximum efficiency. Specifically, Stream Batch reformulates the denoising process by eliminating the traditional wait-and-execute approach and utilizing" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.876, + 0.483, + 0.901 + ], + "angle": 0, + "content": "* denotes equal contribution \nThis work was done when Toshiki was a remote intern at UC Berkeley" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.518, + 0.956 + ], + "angle": 0, + "content": "12371" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.351 + ], + "angle": 0, + "content": "a batching denoising approach, facilitating fluid and high-throughput streams. This results in \\(1.5x\\) higher throughput compared to the conventional sequential denoising approach. R-CFG significantly addresses inefficiencies caused by repetitive computations during denoising. It optimizes the process to require minimal or no additional computations, leading to speed improvements of up to \\(2.05x\\) compared to previous classifier-free methods. Besides, our stochastic similarity filtering dramatically lowers GPU activation frequency by halting computations for static image flows, achieving a remarkable reduction in computational consumption—2.39 times on an RTX 3060 GPU and 1.99 times on an RTX 4090 GPU, respectively. The synergy of our proposed strategies with established acceleration technologies enables image generation to reach speeds of up to 91.07 fps on a single RTX 4090 GPU, outperforming the throughput of AutoPipeline, developed by Diffusers, by more than 59.56x." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.377, + 0.223, + 0.394 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.485, + 0.553 + ], + "angle": 0, + "content": "Recently, there has been a growing trend in the commercialization of diffusion models [3, 28, 32, 34] for applications within the entertainment industry such as Metaverse, online video streaming, broadcasting, and even the robotic field [7]. A pertinent example is the use of diffusion models to create virtual YouTubers. These digital personas should be capable of reacting in a fluid and responsive manner to user input. These areas require diffusion pipelines that offer high throughput and low latency to ensure the efficient interactive streaming generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.485, + 0.825 + ], + "angle": 0, + "content": "To advance the efficiency, current efforts primarily focus on reducing the number of denoising steps, such as decreasing from 50 denoise steps to just a few [24, 25] or even one [21, 42]. The strategy includes distilling the multi-step diffusion models into a few steps [36, 40] or re-framing the diffusion process with neural Ordinary Differential Equations (ODE) [22, 23]. Quantization has also been applied to diffusion models [14, 18] to improve efficiency. These methods share the common goal of approximating either the diffusion process itself or the original model weights to achieve efficiency gains. In this paper, we aim at an orthogonal direction and introduce StreamDiffusion, a pipeline-level solution that enables streaming image generation with high throughput. We highlight that existing model design efforts can still be integrated with our pipeline. Our approach enables the use of N-step denoising diffusion models while keeping high throughput and offers users more flexibility in choosing their preferred models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Specifically, StreamDiffusion seamlessly incorporates a suite of novel strategies. Among these, we propose a simple yet novel approach termed Stream Batch. This method differs from the traditional sequential denoising mode, instead of batching the denoising steps. This subtle modifi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.349 + ], + "angle": 0, + "content": "cation enhances efficiency without sacrificing the quality of image generation. We highlight Stream Batch enables a new capability of generating images conditioned on future frames in a streaming mode, which is something impossible in previous works. Via injecting the future frames, Stream Batch significantly improves the temporal consistency with few additional overhead. Furthermore, the key novelty of Stream Batch lies not only in its GPU parallelization; rather, it serves as a practical realization of a broader Stream Denoising framework. By denoising inputs at diagonally-offset timesteps within a streaming queue structure, diffusion models naturally achieve continuous, autoregressive generation—emitting one output frame per each newly sampled frame—while benefiting from parallel computation. Crucially, this design generalizes directly to sequential tasks such as video, audio, or robotic action-sequence generation, enabling interactive, unbounded-length synthesis." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.352, + 0.909, + 0.563 + ], + "angle": 0, + "content": "Besides, we point out that it is time-consuming for existing diffusion pipelines to use classifier-free guidance for emphasizing the prompts during generation, due to the repetitive and redundant computations for negative conditions. To address this issue, we introduce an innovative approach termed as residual classifier-free guidance (R-CFG). R-CFG approximates the negative condition with a virtual residual noise, which allows us to calculate the negative condition noise only during the initial step of the process. We also indicate that using the original input image latent as the residual term effectively generates results that diverge from the original input image according to the magnitude of the guidance scale, which is a special case of our R-CFG and does not require any computations for the negative condition." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.566, + 0.909, + 0.778 + ], + "angle": 0, + "content": "Furthermore, in real applications such as virtual youtu and AR/VR cases, maintaining the diffusion models always in an active mode is energy-consuming as it keeps hitting GPU. To reduce the energy, we further apply a stochastic similarity filtering (SSF) strategy. In the pipeline, we compute the similarities between continuous inputs and determine whether the diffusion model should process the images based on the probability of an activated similarity. This enables both energy efficiency and visual fluency. In order to further improve the efficiency to cater to the real applications, we apply simple yet effective engineering implementations such as Input-Output Queue (IO-Queue), pre-computing caching, and TensorRT. The overview of the StreamDiffusion pipeline is shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Experiments demonstrate that StreamDiffusion can achieve up to 91.07fps for image generation on one RTX4090 GPU, surpassing the diffusion Autopipeline from Diffusers [15] by up to 59.6x. Besides, our stochastic similarity filtering strategy significantly reduces the power usage by 2.39x on one RTX 3090GPU and by 1.99x on one RTX 4090GPU. Our proposed StreamDiffusion is a new diffusion pipeline that is not only efficient but also energy-saving." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.52, + 0.956 + ], + "angle": 0, + "content": "12372" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.101, + 0.912, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.327, + 0.908, + 0.412 + ], + "angle": 0, + "content": "Figure 2. The overview of StreamDiffusion. StreamDiffusion combines several key components: (1) Stream Batch efficiently processes the denoising steps in batches. (2) Residual Classifier-Free Guidance approximates the negative condition term to reduces the unnecessary repetitive calculations in the UNet. (3) Stochastic Similarity Filter controls the pass of the image stream by calculating the similarity between frames to eliminate the redundant hit onto GPUs. Furthermore, StreamDiffusion leverages techniques like input-output queues for smooth data flow, cache management for faster processing through pre-calculated embedding, and a tiny VAE model to further contribute to overall efficiency. This synergistic combination allows StreamDiffusion to generate images at high throughputs while consuming minimal energy." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.418, + 0.229, + 0.433 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.448, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Efficient Diffusion Models Diffusion models [13, 29, 32, 39] have sparked considerable interest in the commercial sector due to their high-quality image/video generation capabilities. These models have been progressively adapted for various applications, including text-to-image generation [2, 30, 31], image editing [1, 33], video generation [4, 5] and even perception [16, 19, 41]. However, diffusion models are currently limited by their slow speed in generating outputs. In response to this challenge, a variety of strategies have been proposed. One of the mainstreams is to approximate the SDE-based diffusion process [38, 39] through an ordinary differentiable equation (ODE) framework. For example, DPM and \\(\\mathrm{DPM}++\\) [22, 23] introduce ODE-based samplers, which significantly reduce the hundreds of denoising steps to between 15 and 20. Building upon the ODE formulation, InstaFlow [21] advances the reduction of denoising steps to a single instance through the novel strategy of rectified flow [20], while achieving performance close to that of Stable Diffusion [32]. Additionally, distillation from pre-trained diffusion models has also been explored as a method to facilitate few-step denoising. For instance, the consistency model [40] leverages the principle of self-consistency between noise at different denoising steps and uses pre-trained diffusion models [32] to guide the learning of a few-step denoising model, thereby enabling the generation of images within a minimal number of steps. In a notable extension of this concept, LCM [25, 26] applies the idea to the latent space rather than the pixel space. The use of distillation methods [35, 36, 42] to enhance the efficiency of the original Stable Diffusion model presents promising" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.419, + 0.909, + 0.585 + ], + "angle": 0, + "content": "results. Besides improving efficiency through the lens of reducing denoising steps, quantization methods [14, 18] are proposed to make the model run in the regime of low float points, albeit with the potential trade-off of fidelity and efficiency. Moreover, parallel sampling [37] tries to utilize the approximate-parallel denoising strategy to improve the latency of the diffusion model. We emphasize that our work focuses on enhancing throughput and our work is orthogonal to [37]. Notably, our method is optimized for single-GPU, which are common among users. In contrast, the parallel sampling method reduces throughput on a single GPU." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.587, + 0.911, + 0.753 + ], + "angle": 0, + "content": "Our proposed StreamDiffusion is significantly different from the approaches mentioned previously. While earlier methods primarily focus on the low latency of their individual model designs, our approach takes a different route. We introduce a pipeline-level solution specifically tailored for high throughput. Our pipeline can seamlessly integrate the low-latency diffusion models discussed above. Our proposed Stream Batch, residual classifier-free guidance, and the integration of other efficiency-enhancement methods focus on improving the efficiency of the whole pipeline instead of a single diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.912, + 0.903 + ], + "angle": 0, + "content": "Classifier-free Guidance Classifier-free guidance [12] is widely used for conditional generation due to the simplicity, efficiency, and stability compared to classifier-guidance [9]. It leverages negative prompts [8, 11, 32] and essentially operates the vector arithmetic shift in latent space, i.e., we take a step of size (usually set by the guidance scale) away from the unconditional vector or negatively conditioned vector in the direction toward the conditioning manifold [12]. In" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.519, + 0.956 + ], + "angle": 0, + "content": "12373" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.087, + 0.488, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.275, + 0.486, + 0.404 + ], + "angle": 0, + "content": "Figure 3. The concept of Stream Batch. In our approach, instead of waiting for a single image to be fully denoised before processing the next input image, we accept the next input image after each denoising step. This creates a denoising batch where the denoising steps are staggered for each image. By concatenating these staggered denoising steps into a batch, we can efficiently process continuous inputs using a U-Net for batch processing. The input image encoded at timestep \\( t \\) is generated and decoded at timestep \\( t + n \\), where \\( n \\) is the number of the denoising steps." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.425, + 0.487, + 0.669 + ], + "angle": 0, + "content": "the practical implementation, the classifier-free guidance is conducted by sharing the same UNet for both the conditional term and unconditional (or negative) term and subtracting the effect of the unconditional term from the conditioned one. We point out that the way of multiple denoising processes leads to unnecessary computations for the unconditional (or negative) term. To get rid of these redundant computations, we propose a novel residual-classifier-free guidance, termed as R-CFG, which approximates the conditional noise prediction with only requiring one or even zero-time computation for the negatively conditioned noise prediction for the UNet. We note that \\(R\\)-CFG is especially designed for the SDEdit method [27], as we mainly focus on the applications of translating streaming image flow. Our proposed R-CFG significantly improves the latency of the conditional image-to-image generation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.679, + 0.254, + 0.695 + ], + "angle": 0, + "content": "3. StreamDiffusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.486, + 0.826 + ], + "angle": 0, + "content": "StreamDiffusion is a new diffusion pipeline aiming for high throughput. It comprises three key components: Stream Batch strategy, Residual Classifier-Free Guidance (R-CFG), and Stochastic Similarity Filter. Besides, we also incorporate other acceleration methods like a novel input-output queue designed by us, the pre-computation procedure, the tinyautoencoder, and model acceleration tools such as TensorRT. We elaborate on the details below." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.446, + 0.851 + ], + "angle": 0, + "content": "3.1. Stream Batch: Batching the Denoise Step" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.486, + 0.903 + ], + "angle": 0, + "content": "In diffusion models, denoising steps are performed sequentially, resulting in a proportional increase in the processing time of U-Net relative to the number of steps. However," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "to generate high-fidelity images, it is necessary to increase the number of steps. To resolve this problem in interactive diffusion, we propose a method called Stream Batch." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.137, + 0.907, + 0.304 + ], + "angle": 0, + "content": "The Stream Batch technique restructures sequential denoising operations into batched processes, wherein each batch corresponds to a predetermined number of denoising steps, as depicted in Fig. 3. The size of each batch is determined by the number of these denoising steps. This approach allows for each batch element to advance one step further in the denoising sequence via a single pass through U-Net. By iteratively applying this method, it is possible to effectively transform input images encoded at timestep \\( t \\) into their corresponding image-to-image results at timestep \\( t + n \\), thereby streamlining the denoising procedure." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.303, + 0.909, + 0.439 + ], + "angle": 0, + "content": "Stream Batch significantly reduces the need for multiple U-Net inferences. The processing time does not escalate linearly with the number of steps. This technique effectively shifts the trade-off from balancing processing time and generation quality to balancing VRAM capacity and generation quality. With adequate VRAM scaling, this method enables the production of high-quality images within the span of a single U-Net processing cycle, effectively overcoming the constraints imposed by increasing denoising steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.439, + 0.909, + 0.637 + ], + "angle": 0, + "content": "Waiting and Batching can also increase the throughput of the diffusion pipeline. However, with naive Waiting and Batching (WB), denoising cannot begin immediately on the first input frame, leading to higher latency compared to Stream Batch. We mainly aim for smooth streaming applications. Yet achieving a smooth frame rate with WB requires additional engineering, such as precise inference speed estimation and input-output frame synchronization, and minor timing errors must be carefully managed. In contrast, Stream Batch automatically ensures a consistent interval between input and output frames, providing the advantage of lower latency while dynamically reaching the optimal throughput." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.644, + 0.887, + 0.661 + ], + "angle": 0, + "content": "3.2. Improve Time Consistency by Stream Batch" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.666, + 0.909, + 0.848 + ], + "angle": 0, + "content": "Maintaining temporal consistency in video generation is challenging. Many approaches ensure frame coherence by referencing past frames, often through cross-frame attention. However, our Stream Batch method uniquely enables temporal consistency using information from future frames. As shown in Fig. 3, Stream Batch allows simultaneous denoising of multiple frames, passing information from future frames to the current frame. This supports real-time image translation that adapts to sudden changes in input while preserving consistency. In Stream Batch with \\( n \\) denoising steps, keys and values for each frame at each time step form the following batches:" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.857, + 0.875, + 0.877 + ], + "angle": 0, + "content": "\\[\nK _ {\\text {B a t c h}} = \\left[ K _ {t + i, 0}, \\dots , K _ {t, i}, \\dots , K _ {t - (n - 1 - i), n - 1} \\right]\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.883, + 0.868, + 0.906 + ], + "angle": 0, + "content": "\\[\nV _ {\\mathrm {B a t c h}} = \\left[ V _ {t + i, 0}, \\dots , V _ {t, i}, \\dots , V _ {t - (n - 1 - i), n - 1} \\right]\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.944, + 0.52, + 0.956 + ], + "angle": 0, + "content": "12374" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.088, + 0.541, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.089, + 0.887, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.277, + 0.908, + 0.307 + ], + "angle": 0, + "content": "Figure 4. Virtual residual noise vectors: The orange vectors depict the virtual residual noise that starts from PF ODE trajectory and points to the original input latent \\( x_0 \\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.332, + 0.485, + 0.453 + ], + "angle": 0, + "content": "These key and value batches incorporate information across different time frames and denoising steps. For example, if the frame at time step \\( t \\) has reached the \\( i \\)-th denoising step, the batch includes \\( i \\) future denoising steps for different frames and \\( n - 1 - i \\) past frames. In Stream Batch Cross-frame Attention, rather than using the typical \\( K_{t,i} \\) and \\( V_{t,i} \\), we employ \\( K_{\\mathrm{Batch}} \\) and \\( V_{\\mathrm{Batch}} \\), which integrate past and future frame information for the attention computation:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.476, + 0.484, + 0.511 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t n} \\left(Q _ {t, i}, K _ {\\text {B a t c h}}, V _ {\\text {B a t c h}}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {Q _ {t , i} \\cdot K _ {\\text {B a t c h}} ^ {T}}{\\sqrt {d}}\\right) V _ {\\text {B a t c h}}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.521, + 0.484, + 0.567 + ], + "angle": 0, + "content": "This approach enables the generation process to account for information from both past and future frames with across denoising stages, thus enhancing temporal consistency." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.576, + 0.391, + 0.59 + ], + "angle": 0, + "content": "3.3. Residual Classifier-Free Guidance" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.484, + 0.63 + ], + "angle": 0, + "content": "Firstly, SDEdit based method [27] adds perturbation to the input image \\(x_0\\) and transfers it to the noise distribution \\(x_{\\tau_0}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.642, + 0.483, + 0.66 + ], + "angle": 0, + "content": "\\[\nx _ {\\tau_ {0}} = \\sqrt {\\alpha_ {\\tau_ {0}}} x _ {0} + \\sqrt {\\beta_ {\\tau_ {0}}} \\epsilon_ {0}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.667, + 0.486, + 0.803 + ], + "angle": 0, + "content": "where \\(\\alpha_{\\tau_0}\\) and \\(\\beta_{\\tau_0}\\) are values determined by a noise scheduler and \\(\\epsilon_0\\) is a sampled noise from a Gaussian \\(\\mathcal{N}(0,I)\\). When using consistency models for conditional image editing, \\(x_{\\tau_0}\\) can be considered as a point on the PF ODE trajectory, which leads to the conditioning manifold. To intensify the conditioning by Classifier-Free Guidance (CFG)[10], it is imperative to compute a noise for a negative condition PF ODE trajectory, which is used in vector arithmetic shift for the guidance (Eq. 2)." + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.818, + 0.484, + 0.835 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {\\tau_ {i}, \\mathrm {c f g}} = \\epsilon_ {\\tau_ {i}, \\bar {c}} + \\gamma \\left(\\epsilon_ {\\tau_ {i}, c} - \\epsilon_ {\\tau_ {i}, \\bar {c}}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.902 + ], + "angle": 0, + "content": "This requirement introduces additional computational overhead at each denoising step. To reduce this computational overhead, R-CFG utilizes the fact that the original input image \\( x_0 \\) is referable at any stage of denoising steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.332, + 0.907, + 0.407 + ], + "angle": 0, + "content": "For any latent \\( x_{\\tau_i} \\) at the denoising step \\( \\tau_i \\), we can assume the existence of the virtual negative condition \\( \\vec{c}_{\\tau_i}' \\), that satisfies the self-consistency described as Eq. 3 This implies that \\( x_{\\tau_i} \\) is on the PF ODE trajectory going back to the input image \\( x_0 \\)." + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.423, + 0.907, + 0.442 + ], + "angle": 0, + "content": "\\[\nx _ {0} \\approx \\hat {x} _ {0, \\tau_ {i}, \\bar {c} _ {\\tau_ {i}} ^ {\\prime}} = f _ {\\theta} \\left(x _ {\\tau_ {i}}, \\tau_ {i}, \\bar {c} _ {\\tau_ {i}} ^ {\\prime}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.447, + 0.907, + 0.508 + ], + "angle": 0, + "content": "Following the LCM model parameterization [25] and our approximation for the inference time skip connections \\((c_{\\mathrm{skip}}(\\tau) = 0, c_{\\mathrm{out}}(\\tau) = 1\\) at \\(\\tau \\neq 0\\)), the self-consistency equation (Eq. 3) can be expressed as follows," + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.52, + 0.907, + 0.559 + ], + "angle": 0, + "content": "\\[\nx _ {0} \\approx \\frac {x _ {\\tau_ {i}} - \\sqrt {\\beta_ {\\tau_ {i}}} \\epsilon_ {\\tau_ {i} , \\tilde {c} _ {\\tau_ {i}} ^ {\\prime}}}{\\sqrt {\\alpha_ {\\tau_ {i}}}} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.909, + 0.641 + ], + "angle": 0, + "content": "Given the initial value \\( x_0 \\), and the subsequent values of \\( x_{\\tau_i} \\) obtained sequentially through the iterative denoising, the virtual noise vector \\( \\epsilon_{\\tau_i,\\vec{c}'} \\) in the direction toward the input image can be analytically determined by employing these values with the Eq. 4:" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.652, + 0.907, + 0.687 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {\\tau_ {i}, \\bar {c} ^ {\\prime}} = \\frac {x _ {\\tau_ {i}} - \\sqrt {\\alpha_ {\\tau_ {i}}} x _ {0}}{\\sqrt {\\beta_ {\\tau_ {i}}}} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.693, + 0.908, + 0.723 + ], + "angle": 0, + "content": "With the virtual noise \\(\\epsilon_{\\tau_i,\\overline{c'}}\\) obtained from Eq. 5, we formulate R-CFG by:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.737, + 0.907, + 0.754 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {\\tau_ {i}, \\text {c f g}} = \\delta \\epsilon_ {\\tau_ {i}, \\bar {c} ^ {\\prime}} + \\gamma \\left(\\epsilon_ {\\tau_ {i}, c} - \\delta \\epsilon_ {\\tau_ {i}, \\bar {c} ^ {\\prime}}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.809 + ], + "angle": 0, + "content": "where \\(\\delta\\) is a magnitude moderation coefficient for the virtual residual noise that softens the effect and the approximation error of the virtual residual noise." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.901 + ], + "angle": 0, + "content": "R-CFG that uses the original input image latent \\( x_0 \\) as the residual term can effectively generate results that diverge from the original input image according to the magnitude of the guidance scale \\( \\gamma \\), thereby enhancing the effect of conditioning without the need for additional U-Net computations. We call this method Self-Negative R-CFG." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.519, + 0.956 + ], + "angle": 0, + "content": "12375" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "Not only to deviate from the original input image \\( x_0 \\), but also to diverge from any negative condition, we can find the desired reference point \\( \\hat{x}_{0,\\tau_0,\\bar{c}} \\) that reflects the negative condition \\( \\bar{c} \\) using the same self-consistency formulation:" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.164, + 0.484, + 0.201 + ], + "angle": 0, + "content": "\\[\n\\hat {x} _ {0, \\tau_ {0}, \\bar {c}} = \\frac {x _ {\\tau_ {0}} - \\sqrt {\\beta_ {\\tau_ {0}}} \\epsilon_ {\\tau_ {0} , \\bar {c}}}{\\sqrt {\\alpha_ {\\tau_ {0}}}} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.207, + 0.484, + 0.252 + ], + "angle": 0, + "content": "We can obtain \\(\\hat{x}_{0,\\tau_0,\\bar{c}}\\) by computing the actual negative conditioned noise \\(\\epsilon_{\\tau_0,\\bar{c}}\\) using U-Net only one time for the first denoising step." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.253, + 0.485, + 0.389 + ], + "angle": 0, + "content": "In Eq. 5, instead of \\( x_0 \\), using \\( \\hat{x}_{0,\\tau_0,\\bar{c}} \\), we can obtain the virtual negative conditioned noise \\( \\epsilon_{\\tau_{i + 1},\\bar{c}'} \\) that can effectively diverge the generation results from the controllable negative conditioning \\( \\bar{c} \\). We name this method Onetime-Negative R-CFG. In contrast to the conventional CFG, which requires \\( 2n \\) computations of U-Net, the Self-Negative RCFG and Onetime-Negative RCFG necessitate only \\( n \\) and \\( n + 1 \\) computations of U-Net, respectively, where \\( n \\) is the number of the denoising steps." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.398, + 0.336, + 0.415 + ], + "angle": 0, + "content": "3.4. Stochastic Similarity Filter" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.421, + 0.484, + 0.572 + ], + "angle": 0, + "content": "When images remain unchanged or show minimal changes, particularly in scenarios without active user interaction or static environment, nearly identical input images are often repeatedly fed into the VAE and U-Net. This leads to the generation of identical or nearly identical images and unnecessary consumption of GPU resources. In contexts involving continuous inputs, such instances of unmodified input images can occasionally occur. To tackle this issue and minimize unnecessary computational load, we propose a strategy termed stochastic similarity filter (SSF), as shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.572, + 0.484, + 0.603 + ], + "angle": 0, + "content": "We calculate the cosine similarity between the current input image \\( I_{t} \\) and the past reference frame image \\( I_{\\mathrm{ref}} \\)." + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.613, + 0.483, + 0.647 + ], + "angle": 0, + "content": "\\[\nS _ {C} \\left(I _ {t}, I _ {\\text {r e f}}\\right) = \\frac {I _ {t} \\cdot I _ {\\text {r e f}}}{\\| I _ {t} \\| \\| I _ {\\text {r e f}} \\|} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.652, + 0.485, + 0.698 + ], + "angle": 0, + "content": "Based on this cosine similarity, we calculate the probability of skipping the subsequent VAE and U-Net processes. It is given by" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.721, + 0.484, + 0.755 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} (\\operatorname {s k i p} \\mid I _ {t}, I _ {\\text {r e f}}) = \\max \\left\\{0, \\frac {S _ {C} \\left(I _ {t} , I _ {\\text {r e f}}\\right) - \\eta}{1 - \\eta} \\right\\}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "where \\(\\eta\\) is the similarity threshold. This probability decides whether subsequent processes like VAE Encoding, U-Net, and VAE Decoding should be skipped or not. If not skipped, the input image at that time is saved and updated as the reference image \\(I_{ref}\\) for future use. This probabilistic skipping mechanism allows the network to operate fully in dynamic scenes with low inter-frame similarity, while in static scenes with high inter-frame similarity, the network's operational rate decreases, conserving computational resources. The" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.092, + 0.905, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.252, + 0.907, + 0.281 + ], + "angle": 0, + "content": "Figure 5. Average inference time comparison between StreamBatch and normal sequential denoising." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.307, + 0.905, + 0.353 + ], + "angle": 0, + "content": "GPU usage is modulated seamlessly based on the similarity of the input images, enabling smooth adaptation to scenes with varying dynamics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.907, + 0.489 + ], + "angle": 0, + "content": "Note: We emphasize that compared to determining whether we skip the compute via a hard threshold, the proposed probability-sampling-based similarity filtering strategy leads to smoother video generation. Because the hard threshold is prone to making the video stuck, which hurts the impression of watching video streaming, while the sampling-based method significantly improves the smoothness. For the other efficiency improvement methods, we illustrate them in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.503, + 0.646, + 0.52 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.528, + 0.909, + 0.68 + ], + "angle": 0, + "content": "We implement StreamDiffusion pipeline upon LCM, LCM-LoRA [25, 26] and SD-turbo [36]. As a model accelerator, we use TensorRT and for the lightweight efficient VAE, we use TAESD [17]. Our pipeline is compatible with the customer-level GPU. We test our pipeline on NVIDIA RTX4090 GPU, Intel Core i9-13900K CPU, Ubuntu 22.04.3 LTS, and NVIDIA RTX3060 GPU, Intel Core i7-12700K, Windows 11 for image generation. We note that we evaluate the throughput mainly via the average inference time per image through processing 100 images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.689, + 0.736, + 0.705 + ], + "angle": 0, + "content": "4.1.Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.711, + 0.909, + 0.879 + ], + "angle": 0, + "content": "We compare our method with the AutoPipelineForImage2Image, which is a pipeline developed by Huggingface diffusers1. The average inference time comparison is presented in Table. 1. Our pipeline demonstrates a substantial speed increase. When we use TensorRT, StreamDiffusion achieves a minimum speed-up of 13.0 times when running the 10 denoising steps, and reaching up to 59.6 times in scenarios involving a single denoising step. Even though without TensorRT, StreamDiffusion achieves a 29.7 times speed up compared to AutoPipeline when using one step denoising, and an 8.3 times speedup at 10 step denoising." + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.887, + 0.855, + 0.9 + ], + "angle": 0, + "content": "1https://github.com/huggingface/diffusers" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.52, + 0.956 + ], + "angle": 0, + "content": "12376" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.09, + 0.486, + 0.119 + ], + "angle": 0, + "content": "Table 1. Average inference time (ms) at different denoising steps. AutoPipeline refers to Diffusers' implementation [15]." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.129, + 0.482, + 0.217 + ], + "angle": 0, + "content": "
StepStreamDiffusionw/o TRTAutoPipeline
110.65 (59.6x)21.34 (29.7x)634.40 (1x)
216.74 (39.3x)30.61 (21.3x)652.66 (1x)
426.93 (25.8x)48.15 (14.4x)695.20 (1x)
1062.00 (13.0x)96.94 (8.3x)803.23 (1x)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.231, + 0.483, + 0.259 + ], + "angle": 0, + "content": "Table 2. Average inference time (ms) at different denoising steps across CFG methods" + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.27, + 0.485, + 0.368 + ], + "angle": 0, + "content": "
StepSelf-Neg. R-CFGOne-Neg. R-CFGCFG
111.04 (1.52x)16.55 (1.01x)16.74 (1x)
216.61 (1.64x)20.64 (1.32x)27.18 (1x)
320.64 (1.74x)27.25 (1.32x)35.91 (1x)
426.19 (1.90x)31.65 (1.57x)49.71 (1x)
531.47 (2.05x)36.04 (1.79x)64.64 (1x)
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.394, + 0.486, + 0.501 + ], + "angle": 0, + "content": "Efficiency comparison regarding Stream Batch. The efficiency comparison between Stream Batch and the original sequential U-Net loop is shown in Fig. 5. When implementing a denoising batch strategy, we observe a significant improvement in processing time. It achieves a reduction by half when compared to a conventional U-Net loop at sequential denoising steps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.521, + 0.486, + 0.779 + ], + "angle": 0, + "content": "Efficiency comparison regarding R-CFG. Table 2 presents a comparison of the inference times for StreamDiffusion pipelines with R-CFG and conventional CFG. The additional computations required to apply Self-Negative R-CFG are merely lightweight vector operations, resulting in negligible changes in inference time compared to when Self-Negative is not used. When employing Onetime-Negative R-CFG, additional UNet computations are necessary for the first step of the denoising process. Therefore, One-time-negative R-CFG and conventional CFG have almost identical inference times for a single denoising step case. However, as the number of denoising steps increases, the difference in inference time from conventional CFG to both Self-Negative and Onetime-Negative R-CFG becomes more pronounced. At denoising step 5, a speed improvement of \\(2.05\\mathrm{x}\\) is observed with Self-Negative R-CFG and \\(1.79\\mathrm{x}\\) with Onetime-Negative R-CFG, compared to conventional CFG." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.292, + 0.805 + ], + "angle": 0, + "content": "4.2. Energy Consumption" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.486, + 0.903 + ], + "angle": 0, + "content": "We then conduct a comprehensive evaluation of the energy consumption associated with our proposed stochastic similarity filter (SSF), as depicted in Figure. 11 and Figure. 12. These figures provide the GPU utilization patterns when SSF (Threshold \\(\\eta\\) set at 0.98) is applied to input videos containing scenes with periodic static characteristics. The comparative" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.137 + ], + "angle": 0, + "content": "analysis reveals that the incorporation of SSF significantly mitigates GPU usage in instances where the input images are predominantly static or with a high degree of similarity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.138, + 0.909, + 0.288 + ], + "angle": 0, + "content": "Figure. 11 delineates the results derived from a meticulously executed two-denoise-step img2img experiment. This experiment was conducted on a 20-frame video sequence, employing NVIDIA RTX3060 graphics processing units with or without the integration of SSF. The experiment results indicate a substantial decrease in average power consumption from 85.96w to 35.91w on one RTX3060 GPU. Using the same static scene input video with one NVIDIA RTX4090GPU, the power consumption was reduced from 238.68w to 119.77w." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.29, + 0.911, + 0.472 + ], + "angle": 0, + "content": "Furthermore, Figure. 12 expounds on the findings from a similar two-denoise-step img2img experiment using one RTX4090GPU. This time the evaluation of energy consumption is performed on a 1000-frame video featuring dynamic scenes. Remarkably, even under drastically dynamic conditions, the SSF efficiently extracted several frames exhibiting similarity from the dynamic sequence. This process results in a noteworthy reduction in average power consumption, from 236.13w to 199.38w. These findings underscore the efficacy of the Stochastic Similarity Filter in enhancing energy efficiency, particularly in scenarios involving static or minimally varying visual content." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.483, + 0.665, + 0.5 + ], + "angle": 0, + "content": "4.3. Ablation study" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.52, + 0.903, + 0.722 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.733, + 0.86, + 0.748 + ], + "angle": 0, + "content": "Figure 6. Ablation study on different components." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.911, + 0.903 + ], + "angle": 0, + "content": "In our ablation study, as summarized in Fig. 6, we evaluate the average inference time of our proposed method under various configurations to understand the contribution of each component. Our proposed StreamDiffusion achieves an average inference time of \\(10.98 / 9.42\\mathrm{ms}\\) and \\(26.93 / 26.30\\mathrm{ms}\\) for denoising steps 1 and 4 on image-to-image/text-to-image generation, respectively. When the R-CFG is not used, we observe this results in the second largest efficiency drop. This demonstrates that R-CFG is one of the most critical" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.52, + 0.956 + ], + "angle": 0, + "content": "12377" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.196 + ], + "angle": 0, + "content": "components in our pipeline. When the Stream Batch processing is removed ('w/o stream batch'), we observe a large time consumption increase, especially at 4 denoising steps. We also evaluate the impact on the inference time of our pipeline regarding SSF. We observe that SSF plays a significant role in enabling energy saving and does not introduce extra time cost." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.198, + 0.487, + 0.471 + ], + "angle": 0, + "content": "Besides, the absence of TensorRT ('w/o TRT') leads to a large increase in time cost. The removal of pre-computation also results in increased time cost but not much. We attribute the reason to the limited number of key-value computations in Stable Diffusion. Besides, the exclusion of input-output queue ('w/o IO queue') also demonstrates an impact on average inference time, which mainly aims to optimize the parallelization issue resulting from pre- and post-processing. In the AutoPipelineImage2Image's adding noise function, the precision of tensors is converted from fp32 to fp16 for each request, leading to a decrease in speed. In contrast, the StreamDiffusion pipeline standardizes the precision of variables and computational devices beforehand. It does not perform tensor precision conversion or computation device transfers during inference. Consequently, even without any optimization ('w/o any optimization'), our pipeline significantly outperforms the AutoPipelineImage2Image in terms of throughputs." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.48, + 0.482, + 0.497 + ], + "angle": 0, + "content": "4.4. Quantitative Evaluation for the Image Quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.503, + 0.483, + 0.639 + ], + "angle": 0, + "content": "We conduct a quantitative evaluation on the image quality. Specifically, we first evaluate the FID and CLIP scores on the text-to-image generation. We use the same dataset as [6] for the evaluation. We use LCM as our main baseline for the comparison. Note that our method is never trained; our method still improves LCM by a large margin in terms of the FID (29.69 vs. 26.79) and maintains a similar CLIP score (24.95 vs. 24.99). This demonstrates the effectiveness of our proposed method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.901 + ], + "angle": 0, + "content": "User study. We conduct a user study to validate the visual quality of different components of our StreamDiffusion. The results are shown in the Table. 3. Our Stream Batch with future-frame attention significantly enhances time consistency compared to its absence. Additionally, our SSF method addresses the crucial yet rarely explored issue of energy efficiency. While SSF is simple, it is far from trivial: a common approach might apply a hard threshold on similarity to regulate streaming flow. However, as noted in the main text, we introduce a novel approach—using probability sampling to achieve superior streaming quality. As Table 3 illustrates, this approach is preferred by more users over the hard-threshold method for visual quality. Moreover, compared to the vanilla CFG method, both our self-Negative R-CFG and one-time R-CFG are more preferred by the users, demonstrating our method not only improves efficiency but" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.641, + 0.107 + ], + "angle": 0, + "content": "also visual quality." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.125, + 0.905, + 0.24 + ], + "angle": 0, + "content": "
# UsersWithout Future-Frame (%)With Future-Frame (%)
4869.6790.33
# UsersWithout SSF (%)With SSF (%)
14448.6151.39
# UsersHard Threshold (%)SSF (%)
4524.4475.56
# UsersWith / Without CFG (%)Self-Neg R-CFG / One-Time R-CFG (%)
25722 / 2035 / 23
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.245, + 0.905, + 0.301 + ], + "angle": 0, + "content": "Table 3. User study results: Future-frame attention consistency (486 users), SSF quality imperceptibility (144 users), streaming quality between hard threshold filter and SSF (45 users), comparison of with/without CFG and self-negative/one-time R-CFG (257 users)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.322, + 0.907, + 0.459 + ], + "angle": 0, + "content": "This pipeline enables image generation with very low throughput from input images received in real-time from cameras or screen capture devices. At the same time, it is capable of producing high-quality images that effectively align to the specified prompt conditions. These capabilities demonstrate the applicability of our pipeline in various real-time applications, such as real-time game graphic rendering, generative camera effect filters, real-time face conversion, and AI-assisted drawing." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.48, + 0.633, + 0.495 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.909, + 0.81 + ], + "angle": 0, + "content": "We propose StreamDiffusion, a pipeline-level solution for interactive diffusion generation. StreamDiffusion consists of several optimization strategies for both throughput and GPU usage, including Stream Batch with cross-attention, residual classifier-free guidance (R-CFG), IO-queue for parallelization, stochastic similarity filter, pre-computation, Tiny AutoEncoder, and the use of the model acceleration tool. The synergistic combination of these elements results in a marked improvement in efficiency. Specifically, StreamDiffusion achieves up to 91.07 frames per second (fps) on a standard consumer-grade GPU for image generation tasks. This performance level is particularly beneficial for a variety of applications, including but not limited to the Metaverse, online video streaming, and broadcasting sectors. Furthermore, StreamDiffusion demonstrates a significant reduction in GPU power consumption, achieving at least a \\(1.99\\mathrm{x}\\) decrease. This notable efficiency gain underscores StreamDiffusion's potential for commercial application, offering a compelling solution for energy-conscious, high-performance computing environments." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.832, + 0.693, + 0.848 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.858, + 0.908, + 0.901 + ], + "angle": 0, + "content": "We sincerely thank Taku Fujimoto and Huggingface team for their invaluable feedback, courteous support, and insightful discussions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.519, + 0.956 + ], + "angle": 0, + "content": "12378" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.171, + 0.483, + 0.199 + ], + "angle": 0, + "content": "[2] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM Trans. Graph., 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.2, + 0.484, + 0.269 + ], + "angle": 0, + "content": "[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, Wesam Manassra, Prafulla Dhariwal, Casey Chu, Yunxin Jiao, and Aditya Ramesh. Improving image generation with better captions. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.27, + 0.483, + 0.339 + ], + "angle": 0, + "content": "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.341, + 0.484, + 0.422 + ], + "angle": 0, + "content": "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.424, + 0.484, + 0.465 + ], + "angle": 0, + "content": "[6] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.466, + 0.484, + 0.522 + ], + "angle": 0, + "content": "[7] Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. arXiv preprint arXiv:2303.04137, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.523, + 0.484, + 0.592 + ], + "angle": 0, + "content": "[8] Katherine Crowson, Stella Biderman, Daniel Kornis, Dashiell Stander, Eric Hallahan, Louis Castricato, and Edward Raff. Vqgan-clip: Open domain image generation and editing with natural language guidance. In European Conference on Computer Vision, pages 88–105. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.593, + 0.483, + 0.634 + ], + "angle": 0, + "content": "[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.635, + 0.484, + 0.689 + ], + "angle": 0, + "content": "[10] Yilun Du, Shuang Li, and Igor Mordatch. Compositional visual generation with energy based models. In Advances in Neural Information Processing Systems, pages 6637-6647. Curran Associates, Inc., 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.69, + 0.484, + 0.746 + ], + "angle": 0, + "content": "[11] Yilun Du, Shuang Li, and Igor Mordatch. Compositional visual generation with energy based models. Advances in Neural Information Processing Systems, 33:6637-6647, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.774 + ], + "angle": 0, + "content": "[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.483, + 0.817 + ], + "angle": 0, + "content": "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[14] Yushi Huang, Ruihao Gong, Jing Liu, Tianlong Chen, and Xianglong Liu. Tfmq-dm: Temporal feature maintenance quantization for diffusion models. arXiv preprint arXiv:2311.16503, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.872, + 0.484, + 0.899 + ], + "angle": 0, + "content": "[15] Huggingface. https://huggingface.co/docs/diffusers/, 2024.2, 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.147 + ], + "angle": 0, + "content": "[16] Aliasghar Khani, Saeid Asgari, Aditya Sanghi, Ali Mahdavi Amiri, and Ghassan Hamarneh. Slime: Segment like me. In The Twelfth International Conference on Learning Representations, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.908, + 0.177 + ], + "angle": 0, + "content": "[17] Diederik P Kingma and Max Welling. Auto-encoding variational bayes, 2022. 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.908, + 0.248 + ], + "angle": 0, + "content": "[18] Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-diffusion: Quantizing diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 17535-17545, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.906, + 0.29 + ], + "angle": 0, + "content": "[19] Ziyi Li, Qinye Zhou, Xiaoyun Zhang, Ya Zhang, Yanfeng Wang, and Weidi Xie. Open-vocabulary object segmentation with diffusion models. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.906, + 0.333 + ], + "angle": 0, + "content": "[20] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.906, + 0.389 + ], + "angle": 0, + "content": "[21] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, and Qiang Liu. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. arXiv preprint arXiv:2309.06380, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.906, + 0.46 + ], + "angle": 0, + "content": "[22] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.906, + 0.516 + ], + "angle": 0, + "content": "[23] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.906, + 0.56 + ], + "angle": 0, + "content": "[24] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.561, + 0.906, + 0.629 + ], + "angle": 0, + "content": "[25] Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module. arXiv preprint arXiv:2311.05556, 2023. 2, 3, 5, 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.631, + 0.906, + 0.686 + ], + "angle": 0, + "content": "[26] Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.688, + 0.906, + 0.757 + ], + "angle": 0, + "content": "[27] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.758, + 0.799, + 0.773 + ], + "angle": 0, + "content": "[28] OpenAI. https://openai.com/sora, 2024.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.906, + 0.802 + ], + "angle": 0, + "content": "[29] William Peebles and Saining Xie. Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.906, + 0.857 + ], + "angle": 0, + "content": "[30] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2): 3, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 3, 13" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.52, + 0.956 + ], + "angle": 0, + "content": "12379" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.485, + 0.218 + ], + "angle": 0, + "content": "[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.483, + 0.302 + ], + "angle": 0, + "content": "[34] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho†, David Fleet, and Mohammad Norouzi. Imagen: unprecedented photorealism × deep level of language understanding. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.304, + 0.483, + 0.344 + ], + "angle": 0, + "content": "[35] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.347, + 0.483, + 0.387 + ], + "angle": 0, + "content": "[36] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.485, + 0.442 + ], + "angle": 0, + "content": "[37] Andy Shih, Suneel Belkhale, Stefano Ermon, Dorsa Sadigh, and Nima Anari. Parallel sampling of diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.446, + 0.483, + 0.487 + ], + "angle": 0, + "content": "[38] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.489, + 0.483, + 0.543 + ], + "angle": 0, + "content": "[39] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.546, + 0.483, + 0.585 + ], + "angle": 0, + "content": "[40] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.588, + 0.483, + 0.629 + ], + "angle": 0, + "content": "[41] Chenfeng Xu, Huan Ling, Sanja Fidler, and Or Litany. 3diff- tection: 3d object detection with geometry-aware diffusion features, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.631, + 0.483, + 0.685 + ], + "angle": 0, + "content": "[42] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Frédo Durand, William T. Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. arXiv, 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.685 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.52, + 0.956 + ], + "angle": 0, + "content": "12380" + } + ] +] \ No newline at end of file diff --git a/2025/StreamDiffusion_ A Pipeline-level Solution for Real-Time Interactive Generation/522f37ba-5716-44cf-af98-f93d9030fe6d_origin.pdf b/2025/StreamDiffusion_ A Pipeline-level Solution for Real-Time Interactive Generation/522f37ba-5716-44cf-af98-f93d9030fe6d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6c3686dd7dc659fd3fb2ae50bbcfffe0bff517e4 --- /dev/null +++ b/2025/StreamDiffusion_ A Pipeline-level Solution for Real-Time Interactive Generation/522f37ba-5716-44cf-af98-f93d9030fe6d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:97b706c90e2e0aedce88e2a1206f50435a9f334aaf9cb67554acdc84805bc2c2 +size 5421389 diff --git a/2025/StreamDiffusion_ A Pipeline-level Solution for Real-Time Interactive Generation/full.md b/2025/StreamDiffusion_ A Pipeline-level Solution for Real-Time Interactive Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..01d0e584cb5c2669395432e4162e35d4f7504327 --- /dev/null +++ b/2025/StreamDiffusion_ A Pipeline-level Solution for Real-Time Interactive Generation/full.md @@ -0,0 +1,288 @@ +# StreamDiffusion: A Pipeline-level Solution for Real-time Interactive Generation + +Akio Kodaira1* Chenfeng Xu1,* Toshiki Hazama1,* Takanori Yoshimoto2 Kohei Ohno3 Shogo Mitsuhori4 Soichi Sugano5 Hanying Cho6 Zhijian Liu7 Masayoshi Tomizuka1 Kurt Keutzer1 + +$^{1}$ UC Berkeley $^{2}$ University of Tsukuba $^{3}$ International Christian University $^{4}$ Toyo University $^{5}$ Tokyo Institute of Technology $^{6}$ Tohoku University $^{7}$ MIT https://github.com/cumulo-autumn/StreamDiffusion {akio.kodaira, xuchenfeng} @berkeley.edu + +![](images/500c3f3d1a4914d3e3b8bde2cebd872209cca1259e7a1a2f5baca5f1838c85bc.jpg) + +![](images/c9161a56f1584030121bb00e4a259cbbd9e1d0b20f60c49afddfeac54db00eef.jpg) +Figure 1. Interactive generation results with StreamDiffusion. + +![](images/bad1c61184cd583945b70eee9af4f1125a13963a99e5105ede4ba6d54c5acb38.jpg) + +![](images/66a3129c80545bfb81ab3b97a3876729b5138f3d581458bc4e403c0a98fbcaa8.jpg) + +![](images/9ac85bbec7f903bca8fd5782bab9c08b0550d8e135a751d44004bbaa07f12de4.jpg) + +# Abstract + +We introduce StreamDiffusion, a real-time diffusion pipeline designed for streaming image generation. Existing diffusion models are adept at creating images from text or image prompts, yet they often fall short in real-time interaction. This limitation becomes particularly evident in scenarios involving continuous input, such as augmented/virtual reality, + +video game graphics rendering, live video streaming, and broadcasting, where high throughput is imperative. StreamDiffusion tackles this challenge through a novel pipeline-level system design. It employs unique strategies like batching the denoising process (Stream Batch), residual classifier-free guidance (R-CFG), and stochastic similarity filtering (SSF). Additionally, it seamlessly integrates advanced acceleration technologies for maximum efficiency. Specifically, Stream Batch reformulates the denoising process by eliminating the traditional wait-and-execute approach and utilizing + +a batching denoising approach, facilitating fluid and high-throughput streams. This results in $1.5x$ higher throughput compared to the conventional sequential denoising approach. R-CFG significantly addresses inefficiencies caused by repetitive computations during denoising. It optimizes the process to require minimal or no additional computations, leading to speed improvements of up to $2.05x$ compared to previous classifier-free methods. Besides, our stochastic similarity filtering dramatically lowers GPU activation frequency by halting computations for static image flows, achieving a remarkable reduction in computational consumption—2.39 times on an RTX 3060 GPU and 1.99 times on an RTX 4090 GPU, respectively. The synergy of our proposed strategies with established acceleration technologies enables image generation to reach speeds of up to 91.07 fps on a single RTX 4090 GPU, outperforming the throughput of AutoPipeline, developed by Diffusers, by more than 59.56x. + +# 1. Introduction + +Recently, there has been a growing trend in the commercialization of diffusion models [3, 28, 32, 34] for applications within the entertainment industry such as Metaverse, online video streaming, broadcasting, and even the robotic field [7]. A pertinent example is the use of diffusion models to create virtual YouTubers. These digital personas should be capable of reacting in a fluid and responsive manner to user input. These areas require diffusion pipelines that offer high throughput and low latency to ensure the efficient interactive streaming generation. + +To advance the efficiency, current efforts primarily focus on reducing the number of denoising steps, such as decreasing from 50 denoise steps to just a few [24, 25] or even one [21, 42]. The strategy includes distilling the multi-step diffusion models into a few steps [36, 40] or re-framing the diffusion process with neural Ordinary Differential Equations (ODE) [22, 23]. Quantization has also been applied to diffusion models [14, 18] to improve efficiency. These methods share the common goal of approximating either the diffusion process itself or the original model weights to achieve efficiency gains. In this paper, we aim at an orthogonal direction and introduce StreamDiffusion, a pipeline-level solution that enables streaming image generation with high throughput. We highlight that existing model design efforts can still be integrated with our pipeline. Our approach enables the use of N-step denoising diffusion models while keeping high throughput and offers users more flexibility in choosing their preferred models. + +Specifically, StreamDiffusion seamlessly incorporates a suite of novel strategies. Among these, we propose a simple yet novel approach termed Stream Batch. This method differs from the traditional sequential denoising mode, instead of batching the denoising steps. This subtle modifi + +cation enhances efficiency without sacrificing the quality of image generation. We highlight Stream Batch enables a new capability of generating images conditioned on future frames in a streaming mode, which is something impossible in previous works. Via injecting the future frames, Stream Batch significantly improves the temporal consistency with few additional overhead. Furthermore, the key novelty of Stream Batch lies not only in its GPU parallelization; rather, it serves as a practical realization of a broader Stream Denoising framework. By denoising inputs at diagonally-offset timesteps within a streaming queue structure, diffusion models naturally achieve continuous, autoregressive generation—emitting one output frame per each newly sampled frame—while benefiting from parallel computation. Crucially, this design generalizes directly to sequential tasks such as video, audio, or robotic action-sequence generation, enabling interactive, unbounded-length synthesis. + +Besides, we point out that it is time-consuming for existing diffusion pipelines to use classifier-free guidance for emphasizing the prompts during generation, due to the repetitive and redundant computations for negative conditions. To address this issue, we introduce an innovative approach termed as residual classifier-free guidance (R-CFG). R-CFG approximates the negative condition with a virtual residual noise, which allows us to calculate the negative condition noise only during the initial step of the process. We also indicate that using the original input image latent as the residual term effectively generates results that diverge from the original input image according to the magnitude of the guidance scale, which is a special case of our R-CFG and does not require any computations for the negative condition. + +Furthermore, in real applications such as virtual youtu and AR/VR cases, maintaining the diffusion models always in an active mode is energy-consuming as it keeps hitting GPU. To reduce the energy, we further apply a stochastic similarity filtering (SSF) strategy. In the pipeline, we compute the similarities between continuous inputs and determine whether the diffusion model should process the images based on the probability of an activated similarity. This enables both energy efficiency and visual fluency. In order to further improve the efficiency to cater to the real applications, we apply simple yet effective engineering implementations such as Input-Output Queue (IO-Queue), pre-computing caching, and TensorRT. The overview of the StreamDiffusion pipeline is shown in Fig. 2. + +Experiments demonstrate that StreamDiffusion can achieve up to 91.07fps for image generation on one RTX4090 GPU, surpassing the diffusion Autopipeline from Diffusers [15] by up to 59.6x. Besides, our stochastic similarity filtering strategy significantly reduces the power usage by 2.39x on one RTX 3090GPU and by 1.99x on one RTX 4090GPU. Our proposed StreamDiffusion is a new diffusion pipeline that is not only efficient but also energy-saving. + +![](images/3a68d090762dee15ec1178a4deedd8b83d7943bc33592f66b1d28a0c73fab071.jpg) +Figure 2. The overview of StreamDiffusion. StreamDiffusion combines several key components: (1) Stream Batch efficiently processes the denoising steps in batches. (2) Residual Classifier-Free Guidance approximates the negative condition term to reduces the unnecessary repetitive calculations in the UNet. (3) Stochastic Similarity Filter controls the pass of the image stream by calculating the similarity between frames to eliminate the redundant hit onto GPUs. Furthermore, StreamDiffusion leverages techniques like input-output queues for smooth data flow, cache management for faster processing through pre-calculated embedding, and a tiny VAE model to further contribute to overall efficiency. This synergistic combination allows StreamDiffusion to generate images at high throughputs while consuming minimal energy. + +# 2. Related work + +Efficient Diffusion Models Diffusion models [13, 29, 32, 39] have sparked considerable interest in the commercial sector due to their high-quality image/video generation capabilities. These models have been progressively adapted for various applications, including text-to-image generation [2, 30, 31], image editing [1, 33], video generation [4, 5] and even perception [16, 19, 41]. However, diffusion models are currently limited by their slow speed in generating outputs. In response to this challenge, a variety of strategies have been proposed. One of the mainstreams is to approximate the SDE-based diffusion process [38, 39] through an ordinary differentiable equation (ODE) framework. For example, DPM and $\mathrm{DPM}++$ [22, 23] introduce ODE-based samplers, which significantly reduce the hundreds of denoising steps to between 15 and 20. Building upon the ODE formulation, InstaFlow [21] advances the reduction of denoising steps to a single instance through the novel strategy of rectified flow [20], while achieving performance close to that of Stable Diffusion [32]. Additionally, distillation from pre-trained diffusion models has also been explored as a method to facilitate few-step denoising. For instance, the consistency model [40] leverages the principle of self-consistency between noise at different denoising steps and uses pre-trained diffusion models [32] to guide the learning of a few-step denoising model, thereby enabling the generation of images within a minimal number of steps. In a notable extension of this concept, LCM [25, 26] applies the idea to the latent space rather than the pixel space. The use of distillation methods [35, 36, 42] to enhance the efficiency of the original Stable Diffusion model presents promising + +results. Besides improving efficiency through the lens of reducing denoising steps, quantization methods [14, 18] are proposed to make the model run in the regime of low float points, albeit with the potential trade-off of fidelity and efficiency. Moreover, parallel sampling [37] tries to utilize the approximate-parallel denoising strategy to improve the latency of the diffusion model. We emphasize that our work focuses on enhancing throughput and our work is orthogonal to [37]. Notably, our method is optimized for single-GPU, which are common among users. In contrast, the parallel sampling method reduces throughput on a single GPU. + +Our proposed StreamDiffusion is significantly different from the approaches mentioned previously. While earlier methods primarily focus on the low latency of their individual model designs, our approach takes a different route. We introduce a pipeline-level solution specifically tailored for high throughput. Our pipeline can seamlessly integrate the low-latency diffusion models discussed above. Our proposed Stream Batch, residual classifier-free guidance, and the integration of other efficiency-enhancement methods focus on improving the efficiency of the whole pipeline instead of a single diffusion model. + +Classifier-free Guidance Classifier-free guidance [12] is widely used for conditional generation due to the simplicity, efficiency, and stability compared to classifier-guidance [9]. It leverages negative prompts [8, 11, 32] and essentially operates the vector arithmetic shift in latent space, i.e., we take a step of size (usually set by the guidance scale) away from the unconditional vector or negatively conditioned vector in the direction toward the conditioning manifold [12]. In + +![](images/9396a76707ed26a7c24187047be046b4ca5d318c8ba0c8d50d39a412d3e466fa.jpg) +Figure 3. The concept of Stream Batch. In our approach, instead of waiting for a single image to be fully denoised before processing the next input image, we accept the next input image after each denoising step. This creates a denoising batch where the denoising steps are staggered for each image. By concatenating these staggered denoising steps into a batch, we can efficiently process continuous inputs using a U-Net for batch processing. The input image encoded at timestep $t$ is generated and decoded at timestep $t + n$ , where $n$ is the number of the denoising steps. + +the practical implementation, the classifier-free guidance is conducted by sharing the same UNet for both the conditional term and unconditional (or negative) term and subtracting the effect of the unconditional term from the conditioned one. We point out that the way of multiple denoising processes leads to unnecessary computations for the unconditional (or negative) term. To get rid of these redundant computations, we propose a novel residual-classifier-free guidance, termed as R-CFG, which approximates the conditional noise prediction with only requiring one or even zero-time computation for the negatively conditioned noise prediction for the UNet. We note that $R$ -CFG is especially designed for the SDEdit method [27], as we mainly focus on the applications of translating streaming image flow. Our proposed R-CFG significantly improves the latency of the conditional image-to-image generation. + +# 3. StreamDiffusion + +StreamDiffusion is a new diffusion pipeline aiming for high throughput. It comprises three key components: Stream Batch strategy, Residual Classifier-Free Guidance (R-CFG), and Stochastic Similarity Filter. Besides, we also incorporate other acceleration methods like a novel input-output queue designed by us, the pre-computation procedure, the tinyautoencoder, and model acceleration tools such as TensorRT. We elaborate on the details below. + +# 3.1. Stream Batch: Batching the Denoise Step + +In diffusion models, denoising steps are performed sequentially, resulting in a proportional increase in the processing time of U-Net relative to the number of steps. However, + +to generate high-fidelity images, it is necessary to increase the number of steps. To resolve this problem in interactive diffusion, we propose a method called Stream Batch. + +The Stream Batch technique restructures sequential denoising operations into batched processes, wherein each batch corresponds to a predetermined number of denoising steps, as depicted in Fig. 3. The size of each batch is determined by the number of these denoising steps. This approach allows for each batch element to advance one step further in the denoising sequence via a single pass through U-Net. By iteratively applying this method, it is possible to effectively transform input images encoded at timestep $t$ into their corresponding image-to-image results at timestep $t + n$ , thereby streamlining the denoising procedure. + +Stream Batch significantly reduces the need for multiple U-Net inferences. The processing time does not escalate linearly with the number of steps. This technique effectively shifts the trade-off from balancing processing time and generation quality to balancing VRAM capacity and generation quality. With adequate VRAM scaling, this method enables the production of high-quality images within the span of a single U-Net processing cycle, effectively overcoming the constraints imposed by increasing denoising steps. + +Waiting and Batching can also increase the throughput of the diffusion pipeline. However, with naive Waiting and Batching (WB), denoising cannot begin immediately on the first input frame, leading to higher latency compared to Stream Batch. We mainly aim for smooth streaming applications. Yet achieving a smooth frame rate with WB requires additional engineering, such as precise inference speed estimation and input-output frame synchronization, and minor timing errors must be carefully managed. In contrast, Stream Batch automatically ensures a consistent interval between input and output frames, providing the advantage of lower latency while dynamically reaching the optimal throughput. + +# 3.2. Improve Time Consistency by Stream Batch + +Maintaining temporal consistency in video generation is challenging. Many approaches ensure frame coherence by referencing past frames, often through cross-frame attention. However, our Stream Batch method uniquely enables temporal consistency using information from future frames. As shown in Fig. 3, Stream Batch allows simultaneous denoising of multiple frames, passing information from future frames to the current frame. This supports real-time image translation that adapts to sudden changes in input while preserving consistency. In Stream Batch with $n$ denoising steps, keys and values for each frame at each time step form the following batches: + +$$ +K _ {\text {B a t c h}} = \left[ K _ {t + i, 0}, \dots , K _ {t, i}, \dots , K _ {t - (n - 1 - i), n - 1} \right] +$$ + +$$ +V _ {\mathrm {B a t c h}} = \left[ V _ {t + i, 0}, \dots , V _ {t, i}, \dots , V _ {t - (n - 1 - i), n - 1} \right] +$$ + +![](images/944cdd937aa8e28e4add4912d4f0c5b19275100ebb59a74d75c0ea6559784250.jpg) +Figure 4. Virtual residual noise vectors: The orange vectors depict the virtual residual noise that starts from PF ODE trajectory and points to the original input latent $x_0$ + +![](images/f3d56de7abc686ed7dda90ad0cc7dafb9ef629e2adab1525a2e356275051e5b0.jpg) + +These key and value batches incorporate information across different time frames and denoising steps. For example, if the frame at time step $t$ has reached the $i$ -th denoising step, the batch includes $i$ future denoising steps for different frames and $n - 1 - i$ past frames. In Stream Batch Cross-frame Attention, rather than using the typical $K_{t,i}$ and $V_{t,i}$ , we employ $K_{\mathrm{Batch}}$ and $V_{\mathrm{Batch}}$ , which integrate past and future frame information for the attention computation: + +$$ +\operatorname {A t t n} \left(Q _ {t, i}, K _ {\text {B a t c h}}, V _ {\text {B a t c h}}\right) = \operatorname {S o f t m a x} \left(\frac {Q _ {t , i} \cdot K _ {\text {B a t c h}} ^ {T}}{\sqrt {d}}\right) V _ {\text {B a t c h}} +$$ + +This approach enables the generation process to account for information from both past and future frames with across denoising stages, thus enhancing temporal consistency. + +# 3.3. Residual Classifier-Free Guidance + +Firstly, SDEdit based method [27] adds perturbation to the input image $x_0$ and transfers it to the noise distribution $x_{\tau_0}$ + +$$ +x _ {\tau_ {0}} = \sqrt {\alpha_ {\tau_ {0}}} x _ {0} + \sqrt {\beta_ {\tau_ {0}}} \epsilon_ {0}, \tag {1} +$$ + +where $\alpha_{\tau_0}$ and $\beta_{\tau_0}$ are values determined by a noise scheduler and $\epsilon_0$ is a sampled noise from a Gaussian $\mathcal{N}(0,I)$ . When using consistency models for conditional image editing, $x_{\tau_0}$ can be considered as a point on the PF ODE trajectory, which leads to the conditioning manifold. To intensify the conditioning by Classifier-Free Guidance (CFG)[10], it is imperative to compute a noise for a negative condition PF ODE trajectory, which is used in vector arithmetic shift for the guidance (Eq. 2). + +$$ +\epsilon_ {\tau_ {i}, \mathrm {c f g}} = \epsilon_ {\tau_ {i}, \bar {c}} + \gamma \left(\epsilon_ {\tau_ {i}, c} - \epsilon_ {\tau_ {i}, \bar {c}}\right), \tag {2} +$$ + +This requirement introduces additional computational overhead at each denoising step. To reduce this computational overhead, R-CFG utilizes the fact that the original input image $x_0$ is referable at any stage of denoising steps. + +For any latent $x_{\tau_i}$ at the denoising step $\tau_i$ , we can assume the existence of the virtual negative condition $\vec{c}_{\tau_i}'$ , that satisfies the self-consistency described as Eq. 3 This implies that $x_{\tau_i}$ is on the PF ODE trajectory going back to the input image $x_0$ . + +$$ +x _ {0} \approx \hat {x} _ {0, \tau_ {i}, \bar {c} _ {\tau_ {i}} ^ {\prime}} = f _ {\theta} \left(x _ {\tau_ {i}}, \tau_ {i}, \bar {c} _ {\tau_ {i}} ^ {\prime}\right) \tag {3} +$$ + +Following the LCM model parameterization [25] and our approximation for the inference time skip connections $(c_{\mathrm{skip}}(\tau) = 0, c_{\mathrm{out}}(\tau) = 1$ at $\tau \neq 0$ ), the self-consistency equation (Eq. 3) can be expressed as follows, + +$$ +x _ {0} \approx \frac {x _ {\tau_ {i}} - \sqrt {\beta_ {\tau_ {i}}} \epsilon_ {\tau_ {i} , \tilde {c} _ {\tau_ {i}} ^ {\prime}}}{\sqrt {\alpha_ {\tau_ {i}}}} \tag {4} +$$ + +Given the initial value $x_0$ , and the subsequent values of $x_{\tau_i}$ obtained sequentially through the iterative denoising, the virtual noise vector $\epsilon_{\tau_i,\vec{c}'}$ in the direction toward the input image can be analytically determined by employing these values with the Eq. 4: + +$$ +\epsilon_ {\tau_ {i}, \bar {c} ^ {\prime}} = \frac {x _ {\tau_ {i}} - \sqrt {\alpha_ {\tau_ {i}}} x _ {0}}{\sqrt {\beta_ {\tau_ {i}}}} \tag {5} +$$ + +With the virtual noise $\epsilon_{\tau_i,\overline{c'}}$ obtained from Eq. 5, we formulate R-CFG by: + +$$ +\epsilon_ {\tau_ {i}, \text {c f g}} = \delta \epsilon_ {\tau_ {i}, \bar {c} ^ {\prime}} + \gamma \left(\epsilon_ {\tau_ {i}, c} - \delta \epsilon_ {\tau_ {i}, \bar {c} ^ {\prime}}\right) \tag {6} +$$ + +where $\delta$ is a magnitude moderation coefficient for the virtual residual noise that softens the effect and the approximation error of the virtual residual noise. + +R-CFG that uses the original input image latent $x_0$ as the residual term can effectively generate results that diverge from the original input image according to the magnitude of the guidance scale $\gamma$ , thereby enhancing the effect of conditioning without the need for additional U-Net computations. We call this method Self-Negative R-CFG. + +Not only to deviate from the original input image $x_0$ , but also to diverge from any negative condition, we can find the desired reference point $\hat{x}_{0,\tau_0,\bar{c}}$ that reflects the negative condition $\bar{c}$ using the same self-consistency formulation: + +$$ +\hat {x} _ {0, \tau_ {0}, \bar {c}} = \frac {x _ {\tau_ {0}} - \sqrt {\beta_ {\tau_ {0}}} \epsilon_ {\tau_ {0} , \bar {c}}}{\sqrt {\alpha_ {\tau_ {0}}}} \tag {7} +$$ + +We can obtain $\hat{x}_{0,\tau_0,\bar{c}}$ by computing the actual negative conditioned noise $\epsilon_{\tau_0,\bar{c}}$ using U-Net only one time for the first denoising step. + +In Eq. 5, instead of $x_0$ , using $\hat{x}_{0,\tau_0,\bar{c}}$ , we can obtain the virtual negative conditioned noise $\epsilon_{\tau_{i + 1},\bar{c}'}$ that can effectively diverge the generation results from the controllable negative conditioning $\bar{c}$ . We name this method Onetime-Negative R-CFG. In contrast to the conventional CFG, which requires $2n$ computations of U-Net, the Self-Negative RCFG and Onetime-Negative RCFG necessitate only $n$ and $n + 1$ computations of U-Net, respectively, where $n$ is the number of the denoising steps. + +# 3.4. Stochastic Similarity Filter + +When images remain unchanged or show minimal changes, particularly in scenarios without active user interaction or static environment, nearly identical input images are often repeatedly fed into the VAE and U-Net. This leads to the generation of identical or nearly identical images and unnecessary consumption of GPU resources. In contexts involving continuous inputs, such instances of unmodified input images can occasionally occur. To tackle this issue and minimize unnecessary computational load, we propose a strategy termed stochastic similarity filter (SSF), as shown in Fig. 2. + +We calculate the cosine similarity between the current input image $I_{t}$ and the past reference frame image $I_{\mathrm{ref}}$ . + +$$ +S _ {C} \left(I _ {t}, I _ {\text {r e f}}\right) = \frac {I _ {t} \cdot I _ {\text {r e f}}}{\| I _ {t} \| \| I _ {\text {r e f}} \|} \tag {8} +$$ + +Based on this cosine similarity, we calculate the probability of skipping the subsequent VAE and U-Net processes. It is given by + +$$ +\mathbf {P} (\operatorname {s k i p} \mid I _ {t}, I _ {\text {r e f}}) = \max \left\{0, \frac {S _ {C} \left(I _ {t} , I _ {\text {r e f}}\right) - \eta}{1 - \eta} \right\}, \tag {9} +$$ + +where $\eta$ is the similarity threshold. This probability decides whether subsequent processes like VAE Encoding, U-Net, and VAE Decoding should be skipped or not. If not skipped, the input image at that time is saved and updated as the reference image $I_{ref}$ for future use. This probabilistic skipping mechanism allows the network to operate fully in dynamic scenes with low inter-frame similarity, while in static scenes with high inter-frame similarity, the network's operational rate decreases, conserving computational resources. The + +![](images/5a47214f6b866ccb49e15649e48244f706232c02eb1b3646da105c33aa4555c9.jpg) +Figure 5. Average inference time comparison between StreamBatch and normal sequential denoising. + +GPU usage is modulated seamlessly based on the similarity of the input images, enabling smooth adaptation to scenes with varying dynamics. + +Note: We emphasize that compared to determining whether we skip the compute via a hard threshold, the proposed probability-sampling-based similarity filtering strategy leads to smoother video generation. Because the hard threshold is prone to making the video stuck, which hurts the impression of watching video streaming, while the sampling-based method significantly improves the smoothness. For the other efficiency improvement methods, we illustrate them in the supplementary material. + +# 4. Experiments + +We implement StreamDiffusion pipeline upon LCM, LCM-LoRA [25, 26] and SD-turbo [36]. As a model accelerator, we use TensorRT and for the lightweight efficient VAE, we use TAESD [17]. Our pipeline is compatible with the customer-level GPU. We test our pipeline on NVIDIA RTX4090 GPU, Intel Core i9-13900K CPU, Ubuntu 22.04.3 LTS, and NVIDIA RTX3060 GPU, Intel Core i7-12700K, Windows 11 for image generation. We note that we evaluate the throughput mainly via the average inference time per image through processing 100 images. + +# 4.1.Quantitative Evaluation + +We compare our method with the AutoPipelineForImage2Image, which is a pipeline developed by Huggingface diffusers1. The average inference time comparison is presented in Table. 1. Our pipeline demonstrates a substantial speed increase. When we use TensorRT, StreamDiffusion achieves a minimum speed-up of 13.0 times when running the 10 denoising steps, and reaching up to 59.6 times in scenarios involving a single denoising step. Even though without TensorRT, StreamDiffusion achieves a 29.7 times speed up compared to AutoPipeline when using one step denoising, and an 8.3 times speedup at 10 step denoising. + +Table 1. Average inference time (ms) at different denoising steps. AutoPipeline refers to Diffusers' implementation [15]. + +
StepStreamDiffusionw/o TRTAutoPipeline
110.65 (59.6x)21.34 (29.7x)634.40 (1x)
216.74 (39.3x)30.61 (21.3x)652.66 (1x)
426.93 (25.8x)48.15 (14.4x)695.20 (1x)
1062.00 (13.0x)96.94 (8.3x)803.23 (1x)
+ +Table 2. Average inference time (ms) at different denoising steps across CFG methods + +
StepSelf-Neg. R-CFGOne-Neg. R-CFGCFG
111.04 (1.52x)16.55 (1.01x)16.74 (1x)
216.61 (1.64x)20.64 (1.32x)27.18 (1x)
320.64 (1.74x)27.25 (1.32x)35.91 (1x)
426.19 (1.90x)31.65 (1.57x)49.71 (1x)
531.47 (2.05x)36.04 (1.79x)64.64 (1x)
+ +Efficiency comparison regarding Stream Batch. The efficiency comparison between Stream Batch and the original sequential U-Net loop is shown in Fig. 5. When implementing a denoising batch strategy, we observe a significant improvement in processing time. It achieves a reduction by half when compared to a conventional U-Net loop at sequential denoising steps. + +Efficiency comparison regarding R-CFG. Table 2 presents a comparison of the inference times for StreamDiffusion pipelines with R-CFG and conventional CFG. The additional computations required to apply Self-Negative R-CFG are merely lightweight vector operations, resulting in negligible changes in inference time compared to when Self-Negative is not used. When employing Onetime-Negative R-CFG, additional UNet computations are necessary for the first step of the denoising process. Therefore, One-time-negative R-CFG and conventional CFG have almost identical inference times for a single denoising step case. However, as the number of denoising steps increases, the difference in inference time from conventional CFG to both Self-Negative and Onetime-Negative R-CFG becomes more pronounced. At denoising step 5, a speed improvement of $2.05\mathrm{x}$ is observed with Self-Negative R-CFG and $1.79\mathrm{x}$ with Onetime-Negative R-CFG, compared to conventional CFG. + +# 4.2. Energy Consumption + +We then conduct a comprehensive evaluation of the energy consumption associated with our proposed stochastic similarity filter (SSF), as depicted in Figure. 11 and Figure. 12. These figures provide the GPU utilization patterns when SSF (Threshold $\eta$ set at 0.98) is applied to input videos containing scenes with periodic static characteristics. The comparative + +analysis reveals that the incorporation of SSF significantly mitigates GPU usage in instances where the input images are predominantly static or with a high degree of similarity. + +Figure. 11 delineates the results derived from a meticulously executed two-denoise-step img2img experiment. This experiment was conducted on a 20-frame video sequence, employing NVIDIA RTX3060 graphics processing units with or without the integration of SSF. The experiment results indicate a substantial decrease in average power consumption from 85.96w to 35.91w on one RTX3060 GPU. Using the same static scene input video with one NVIDIA RTX4090GPU, the power consumption was reduced from 238.68w to 119.77w. + +Furthermore, Figure. 12 expounds on the findings from a similar two-denoise-step img2img experiment using one RTX4090GPU. This time the evaluation of energy consumption is performed on a 1000-frame video featuring dynamic scenes. Remarkably, even under drastically dynamic conditions, the SSF efficiently extracted several frames exhibiting similarity from the dynamic sequence. This process results in a noteworthy reduction in average power consumption, from 236.13w to 199.38w. These findings underscore the efficacy of the Stochastic Similarity Filter in enhancing energy efficiency, particularly in scenarios involving static or minimally varying visual content. + +# 4.3. Ablation study + +![](images/6a23082892a1ecbe77c3aef8b58969b4ab462c0bc315df88c90cb19ca7d60c27.jpg) +Figure 6. Ablation study on different components. + +In our ablation study, as summarized in Fig. 6, we evaluate the average inference time of our proposed method under various configurations to understand the contribution of each component. Our proposed StreamDiffusion achieves an average inference time of $10.98 / 9.42\mathrm{ms}$ and $26.93 / 26.30\mathrm{ms}$ for denoising steps 1 and 4 on image-to-image/text-to-image generation, respectively. When the R-CFG is not used, we observe this results in the second largest efficiency drop. This demonstrates that R-CFG is one of the most critical + +components in our pipeline. When the Stream Batch processing is removed ('w/o stream batch'), we observe a large time consumption increase, especially at 4 denoising steps. We also evaluate the impact on the inference time of our pipeline regarding SSF. We observe that SSF plays a significant role in enabling energy saving and does not introduce extra time cost. + +Besides, the absence of TensorRT ('w/o TRT') leads to a large increase in time cost. The removal of pre-computation also results in increased time cost but not much. We attribute the reason to the limited number of key-value computations in Stable Diffusion. Besides, the exclusion of input-output queue ('w/o IO queue') also demonstrates an impact on average inference time, which mainly aims to optimize the parallelization issue resulting from pre- and post-processing. In the AutoPipelineImage2Image's adding noise function, the precision of tensors is converted from fp32 to fp16 for each request, leading to a decrease in speed. In contrast, the StreamDiffusion pipeline standardizes the precision of variables and computational devices beforehand. It does not perform tensor precision conversion or computation device transfers during inference. Consequently, even without any optimization ('w/o any optimization'), our pipeline significantly outperforms the AutoPipelineImage2Image in terms of throughputs. + +# 4.4. Quantitative Evaluation for the Image Quality. + +We conduct a quantitative evaluation on the image quality. Specifically, we first evaluate the FID and CLIP scores on the text-to-image generation. We use the same dataset as [6] for the evaluation. We use LCM as our main baseline for the comparison. Note that our method is never trained; our method still improves LCM by a large margin in terms of the FID (29.69 vs. 26.79) and maintains a similar CLIP score (24.95 vs. 24.99). This demonstrates the effectiveness of our proposed method. + +User study. We conduct a user study to validate the visual quality of different components of our StreamDiffusion. The results are shown in the Table. 3. Our Stream Batch with future-frame attention significantly enhances time consistency compared to its absence. Additionally, our SSF method addresses the crucial yet rarely explored issue of energy efficiency. While SSF is simple, it is far from trivial: a common approach might apply a hard threshold on similarity to regulate streaming flow. However, as noted in the main text, we introduce a novel approach—using probability sampling to achieve superior streaming quality. As Table 3 illustrates, this approach is preferred by more users over the hard-threshold method for visual quality. Moreover, compared to the vanilla CFG method, both our self-Negative R-CFG and one-time R-CFG are more preferred by the users, demonstrating our method not only improves efficiency but + +also visual quality. + +
# UsersWithout Future-Frame (%)With Future-Frame (%)
4869.6790.33
# UsersWithout SSF (%)With SSF (%)
14448.6151.39
# UsersHard Threshold (%)SSF (%)
4524.4475.56
# UsersWith / Without CFG (%)Self-Neg R-CFG / One-Time R-CFG (%)
25722 / 2035 / 23
+ +Table 3. User study results: Future-frame attention consistency (486 users), SSF quality imperceptibility (144 users), streaming quality between hard threshold filter and SSF (45 users), comparison of with/without CFG and self-negative/one-time R-CFG (257 users). + +This pipeline enables image generation with very low throughput from input images received in real-time from cameras or screen capture devices. At the same time, it is capable of producing high-quality images that effectively align to the specified prompt conditions. These capabilities demonstrate the applicability of our pipeline in various real-time applications, such as real-time game graphic rendering, generative camera effect filters, real-time face conversion, and AI-assisted drawing. + +# 5. Conclusion + +We propose StreamDiffusion, a pipeline-level solution for interactive diffusion generation. StreamDiffusion consists of several optimization strategies for both throughput and GPU usage, including Stream Batch with cross-attention, residual classifier-free guidance (R-CFG), IO-queue for parallelization, stochastic similarity filter, pre-computation, Tiny AutoEncoder, and the use of the model acceleration tool. The synergistic combination of these elements results in a marked improvement in efficiency. Specifically, StreamDiffusion achieves up to 91.07 frames per second (fps) on a standard consumer-grade GPU for image generation tasks. This performance level is particularly beneficial for a variety of applications, including but not limited to the Metaverse, online video streaming, and broadcasting sectors. Furthermore, StreamDiffusion demonstrates a significant reduction in GPU power consumption, achieving at least a $1.99\mathrm{x}$ decrease. This notable efficiency gain underscores StreamDiffusion's potential for commercial application, offering a compelling solution for energy-conscious, high-performance computing environments. + +# 6. Acknowledgments + +We sincerely thank Taku Fujimoto and Huggingface team for their invaluable feedback, courteous support, and insightful discussions. + +# References + +[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[2] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM Trans. Graph., 2023. 3 +[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, Wesam Manassra, Prafulla Dhariwal, Casey Chu, Yunxin Jiao, and Aditya Ramesh. Improving image generation with better captions. 2 +[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3 +[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3 +[6] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 8 +[7] Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. arXiv preprint arXiv:2303.04137, 2023. 2 +[8] Katherine Crowson, Stella Biderman, Daniel Kornis, Dashiell Stander, Eric Hallahan, Louis Castricato, and Edward Raff. Vqgan-clip: Open domain image generation and editing with natural language guidance. In European Conference on Computer Vision, pages 88–105. Springer, 2022. 3 +[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 3 +[10] Yilun Du, Shuang Li, and Igor Mordatch. Compositional visual generation with energy based models. In Advances in Neural Information Processing Systems, pages 6637-6647. Curran Associates, Inc., 2020. 5 +[11] Yilun Du, Shuang Li, and Igor Mordatch. Compositional visual generation with energy based models. Advances in Neural Information Processing Systems, 33:6637-6647, 2020. 3 +[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3 +[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3 +[14] Yushi Huang, Ruihao Gong, Jing Liu, Tianlong Chen, and Xianglong Liu. Tfmq-dm: Temporal feature maintenance quantization for diffusion models. arXiv preprint arXiv:2311.16503, 2023. 2, 3 +[15] Huggingface. https://huggingface.co/docs/diffusers/, 2024.2, 7 + +[16] Aliasghar Khani, Saeid Asgari, Aditya Sanghi, Ali Mahdavi Amiri, and Ghassan Hamarneh. Slime: Segment like me. In The Twelfth International Conference on Learning Representations, 2024. 3 +[17] Diederik P Kingma and Max Welling. Auto-encoding variational bayes, 2022. 6, 13 +[18] Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-diffusion: Quantizing diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 17535-17545, 2023. 2, 3 +[19] Ziyi Li, Qinye Zhou, Xiaoyun Zhang, Ya Zhang, Yanfeng Wang, and Weidi Xie. Open-vocabulary object segmentation with diffusion models. 2023. 3 +[20] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 3 +[21] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, and Qiang Liu. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. arXiv preprint arXiv:2309.06380, 2023. 2, 3 +[22] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 2, 3 +[23] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 2, 3 +[24] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference, 2023. 2 +[25] Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module. arXiv preprint arXiv:2311.05556, 2023. 2, 3, 5, 6, 11 +[26] Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module, 2023. 3, 6 +[27] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 4, 5 +[28] OpenAI. https://openai.com/sora, 2024.2 +[29] William Peebles and Saining Xie. Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748, 2022. 3 +[30] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2): 3, 2022. 3 +[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 3, 13 + +[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3 +[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. 2022. 3 +[34] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho†, David Fleet, and Mohammad Norouzi. Imagen: unprecedented photorealism × deep level of language understanding. 2022. 2 +[35] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 3 +[36] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 2, 3, 6 +[37] Andy Shih, Suneel Belkhale, Stefano Ermon, Dorsa Sadigh, and Nima Anari. Parallel sampling of diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 3 +[38] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 3 +[39] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3 +[40] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023. 2, 3 +[41] Chenfeng Xu, Huan Ling, Sanja Fidler, and Or Litany. 3diff- tection: 3d object detection with geometry-aware diffusion features, 2023. 3 +[42] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Frédo Durand, William T. Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. arXiv, 2023. 2, 3 \ No newline at end of file diff --git a/2025/StreamDiffusion_ A Pipeline-level Solution for Real-Time Interactive Generation/images.zip b/2025/StreamDiffusion_ A Pipeline-level Solution for Real-Time Interactive Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..460adb55bdbf16fe1617069c76cb5808da49b65d --- /dev/null +++ b/2025/StreamDiffusion_ A Pipeline-level Solution for Real-Time Interactive Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8cc21ae69f18e49cc4929ad260cae2ded652820ae7a8504b45dde6037ab8708d +size 555246 diff --git a/2025/StreamDiffusion_ A Pipeline-level Solution for Real-Time Interactive Generation/layout.json b/2025/StreamDiffusion_ A Pipeline-level Solution for Real-Time Interactive Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c72e94279feaa8b5be12a26587133468c3b326ab --- /dev/null +++ b/2025/StreamDiffusion_ A Pipeline-level Solution for Real-Time Interactive Generation/layout.json @@ -0,0 +1,7308 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 103, + 550, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 103, + 550, + 120 + ], + "spans": [ + { + "bbox": [ + 58, + 103, + 550, + 120 + ], + "type": "text", + "content": "StreamDiffusion: A Pipeline-level Solution for Real-time Interactive Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 142, + 558, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 142, + 558, + 172 + ], + "spans": [ + { + "bbox": [ + 61, + 142, + 558, + 172 + ], + "type": "text", + "content": "Akio Kodaira1* Chenfeng Xu1,* Toshiki Hazama1,* Takanori Yoshimoto2 Kohei Ohno3 Shogo Mitsuhori4 Soichi Sugano5 Hanying Cho6 Zhijian Liu7 Masayoshi Tomizuka1 Kurt Keutzer1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 122, + 184, + 496, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 184, + 496, + 245 + ], + "spans": [ + { + "bbox": [ + 122, + 184, + 496, + 245 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 122, + 184, + 496, + 245 + ], + "type": "text", + "content": "UC Berkeley " + }, + { + "bbox": [ + 122, + 184, + 496, + 245 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 122, + 184, + 496, + 245 + ], + "type": "text", + "content": "University of Tsukuba " + }, + { + "bbox": [ + 122, + 184, + 496, + 245 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 122, + 184, + 496, + 245 + ], + "type": "text", + "content": "International Christian University " + }, + { + "bbox": [ + 122, + 184, + 496, + 245 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 122, + 184, + 496, + 245 + ], + "type": "text", + "content": "Toyo University " + }, + { + "bbox": [ + 122, + 184, + 496, + 245 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 122, + 184, + 496, + 245 + ], + "type": "text", + "content": "Tokyo Institute of Technology " + }, + { + "bbox": [ + 122, + 184, + 496, + 245 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 122, + 184, + 496, + 245 + ], + "type": "text", + "content": "Tohoku University " + }, + { + "bbox": [ + 122, + 184, + 496, + 245 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 122, + 184, + 496, + 245 + ], + "type": "text", + "content": "MIT https://github.com/cumulo-autumn/StreamDiffusion {akio.kodaira, xuchenfeng} @berkeley.edu" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 56, + 255, + 156, + 552 + ], + "blocks": [ + { + "bbox": [ + 56, + 255, + 156, + 552 + ], + "lines": [ + { + "bbox": [ + 56, + 255, + 156, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 156, + 552 + ], + "type": "image", + "image_path": "500c3f3d1a4914d3e3b8bde2cebd872209cca1259e7a1a2f5baca5f1838c85bc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 156, + 255, + 256, + 552 + ], + "blocks": [ + { + "bbox": [ + 156, + 255, + 256, + 552 + ], + "lines": [ + { + "bbox": [ + 156, + 255, + 256, + 552 + ], + "spans": [ + { + "bbox": [ + 156, + 255, + 256, + 552 + ], + "type": "image", + "image_path": "c9161a56f1584030121bb00e4a259cbbd9e1d0b20f60c49afddfeac54db00eef.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 561, + 416, + 572 + ], + "lines": [ + { + "bbox": [ + 192, + 561, + 416, + 572 + ], + "spans": [ + { + "bbox": [ + 192, + 561, + 416, + 572 + ], + "type": "text", + "content": "Figure 1. Interactive generation results with StreamDiffusion." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 257, + 255, + 355, + 552 + ], + "blocks": [ + { + "bbox": [ + 257, + 255, + 355, + 552 + ], + "lines": [ + { + "bbox": [ + 257, + 255, + 355, + 552 + ], + "spans": [ + { + "bbox": [ + 257, + 255, + 355, + 552 + ], + "type": "image", + "image_path": "bad1c61184cd583945b70eee9af4f1125a13963a99e5105ede4ba6d54c5acb38.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 357, + 255, + 455, + 552 + ], + "blocks": [ + { + "bbox": [ + 357, + 255, + 455, + 552 + ], + "lines": [ + { + "bbox": [ + 357, + 255, + 455, + 552 + ], + "spans": [ + { + "bbox": [ + 357, + 255, + 455, + 552 + ], + "type": "image", + "image_path": "66a3129c80545bfb81ab3b97a3876729b5138f3d581458bc4e403c0a98fbcaa8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 457, + 255, + 555, + 552 + ], + "blocks": [ + { + "bbox": [ + 457, + 255, + 555, + 552 + ], + "lines": [ + { + "bbox": [ + 457, + 255, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 457, + 255, + 555, + 552 + ], + "type": "image", + "image_path": "9ac85bbec7f903bca8fd5782bab9c08b0550d8e135a751d44004bbaa07f12de4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 151, + 585, + 200, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 585, + 200, + 597 + ], + "spans": [ + { + "bbox": [ + 151, + 585, + 200, + 597 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 611, + 297, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 297, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 297, + 684 + ], + "type": "text", + "content": "We introduce StreamDiffusion, a real-time diffusion pipeline designed for streaming image generation. Existing diffusion models are adept at creating images from text or image prompts, yet they often fall short in real-time interaction. This limitation becomes particularly evident in scenarios involving continuous input, such as augmented/virtual reality," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 586, + 556, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 556, + 706 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 556, + 706 + ], + "type": "text", + "content": "video game graphics rendering, live video streaming, and broadcasting, where high throughput is imperative. StreamDiffusion tackles this challenge through a novel pipeline-level system design. It employs unique strategies like batching the denoising process (Stream Batch), residual classifier-free guidance (R-CFG), and stochastic similarity filtering (SSF). Additionally, it seamlessly integrates advanced acceleration technologies for maximum efficiency. Specifically, Stream Batch reformulates the denoising process by eliminating the traditional wait-and-execute approach and utilizing" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 295, + 713 + ], + "type": "text", + "content": "* denotes equal contribution \nThis work was done when Toshiki was a remote intern at UC Berkeley" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "type": "text", + "content": "12371" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 277 + ], + "type": "text", + "content": "a batching denoising approach, facilitating fluid and high-throughput streams. This results in " + }, + { + "bbox": [ + 55, + 72, + 297, + 277 + ], + "type": "inline_equation", + "content": "1.5x" + }, + { + "bbox": [ + 55, + 72, + 297, + 277 + ], + "type": "text", + "content": " higher throughput compared to the conventional sequential denoising approach. R-CFG significantly addresses inefficiencies caused by repetitive computations during denoising. It optimizes the process to require minimal or no additional computations, leading to speed improvements of up to " + }, + { + "bbox": [ + 55, + 72, + 297, + 277 + ], + "type": "inline_equation", + "content": "2.05x" + }, + { + "bbox": [ + 55, + 72, + 297, + 277 + ], + "type": "text", + "content": " compared to previous classifier-free methods. Besides, our stochastic similarity filtering dramatically lowers GPU activation frequency by halting computations for static image flows, achieving a remarkable reduction in computational consumption—2.39 times on an RTX 3060 GPU and 1.99 times on an RTX 4090 GPU, respectively. The synergy of our proposed strategies with established acceleration technologies enables image generation to reach speeds of up to 91.07 fps on a single RTX 4090 GPU, outperforming the throughput of AutoPipeline, developed by Diffusers, by more than 59.56x." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 298, + 136, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 136, + 312 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 136, + 312 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 319, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 437 + ], + "type": "text", + "content": "Recently, there has been a growing trend in the commercialization of diffusion models [3, 28, 32, 34] for applications within the entertainment industry such as Metaverse, online video streaming, broadcasting, and even the robotic field [7]. A pertinent example is the use of diffusion models to create virtual YouTubers. These digital personas should be capable of reacting in a fluid and responsive manner to user input. These areas require diffusion pipelines that offer high throughput and low latency to ensure the efficient interactive streaming generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 438, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 296, + 653 + ], + "type": "text", + "content": "To advance the efficiency, current efforts primarily focus on reducing the number of denoising steps, such as decreasing from 50 denoise steps to just a few [24, 25] or even one [21, 42]. The strategy includes distilling the multi-step diffusion models into a few steps [36, 40] or re-framing the diffusion process with neural Ordinary Differential Equations (ODE) [22, 23]. Quantization has also been applied to diffusion models [14, 18] to improve efficiency. These methods share the common goal of approximating either the diffusion process itself or the original model weights to achieve efficiency gains. In this paper, we aim at an orthogonal direction and introduce StreamDiffusion, a pipeline-level solution that enables streaming image generation with high throughput. We highlight that existing model design efforts can still be integrated with our pipeline. Our approach enables the use of N-step denoising diffusion models while keeping high throughput and offers users more flexibility in choosing their preferred models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": "Specifically, StreamDiffusion seamlessly incorporates a suite of novel strategies. Among these, we propose a simple yet novel approach termed Stream Batch. This method differs from the traditional sequential denoising mode, instead of batching the denoising steps. This subtle modifi" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 276 + ], + "type": "text", + "content": "cation enhances efficiency without sacrificing the quality of image generation. We highlight Stream Batch enables a new capability of generating images conditioned on future frames in a streaming mode, which is something impossible in previous works. Via injecting the future frames, Stream Batch significantly improves the temporal consistency with few additional overhead. Furthermore, the key novelty of Stream Batch lies not only in its GPU parallelization; rather, it serves as a practical realization of a broader Stream Denoising framework. By denoising inputs at diagonally-offset timesteps within a streaming queue structure, diffusion models naturally achieve continuous, autoregressive generation—emitting one output frame per each newly sampled frame—while benefiting from parallel computation. Crucially, this design generalizes directly to sequential tasks such as video, audio, or robotic action-sequence generation, enabling interactive, unbounded-length synthesis." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 278, + 556, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 556, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 556, + 445 + ], + "type": "text", + "content": "Besides, we point out that it is time-consuming for existing diffusion pipelines to use classifier-free guidance for emphasizing the prompts during generation, due to the repetitive and redundant computations for negative conditions. To address this issue, we introduce an innovative approach termed as residual classifier-free guidance (R-CFG). R-CFG approximates the negative condition with a virtual residual noise, which allows us to calculate the negative condition noise only during the initial step of the process. We also indicate that using the original input image latent as the residual term effectively generates results that diverge from the original input image according to the magnitude of the guidance scale, which is a special case of our R-CFG and does not require any computations for the negative condition." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 448, + 556, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 556, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 556, + 616 + ], + "type": "text", + "content": "Furthermore, in real applications such as virtual youtu and AR/VR cases, maintaining the diffusion models always in an active mode is energy-consuming as it keeps hitting GPU. To reduce the energy, we further apply a stochastic similarity filtering (SSF) strategy. In the pipeline, we compute the similarities between continuous inputs and determine whether the diffusion model should process the images based on the probability of an activated similarity. This enables both energy efficiency and visual fluency. In order to further improve the efficiency to cater to the real applications, we apply simple yet effective engineering implementations such as Input-Output Queue (IO-Queue), pre-computing caching, and TensorRT. The overview of the StreamDiffusion pipeline is shown in Fig. 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Experiments demonstrate that StreamDiffusion can achieve up to 91.07fps for image generation on one RTX4090 GPU, surpassing the diffusion Autopipeline from Diffusers [15] by up to 59.6x. Besides, our stochastic similarity filtering strategy significantly reduces the power usage by 2.39x on one RTX 3090GPU and by 1.99x on one RTX 4090GPU. Our proposed StreamDiffusion is a new diffusion pipeline that is not only efficient but also energy-saving." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "text", + "content": "12372" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 79, + 558, + 242 + ], + "blocks": [ + { + "bbox": [ + 56, + 79, + 558, + 242 + ], + "lines": [ + { + "bbox": [ + 56, + 79, + 558, + 242 + ], + "spans": [ + { + "bbox": [ + 56, + 79, + 558, + 242 + ], + "type": "image", + "image_path": "3a68d090762dee15ec1178a4deedd8b83d7943bc33592f66b1d28a0c73fab071.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 258, + 555, + 326 + ], + "lines": [ + { + "bbox": [ + 54, + 258, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 54, + 258, + 555, + 326 + ], + "type": "text", + "content": "Figure 2. The overview of StreamDiffusion. StreamDiffusion combines several key components: (1) Stream Batch efficiently processes the denoising steps in batches. (2) Residual Classifier-Free Guidance approximates the negative condition term to reduces the unnecessary repetitive calculations in the UNet. (3) Stochastic Similarity Filter controls the pass of the image stream by calculating the similarity between frames to eliminate the redundant hit onto GPUs. Furthermore, StreamDiffusion leverages techniques like input-output queues for smooth data flow, cache management for faster processing through pre-calculated embedding, and a tiny VAE model to further contribute to overall efficiency. This synergistic combination allows StreamDiffusion to generate images at high throughputs while consuming minimal energy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 331, + 140, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 140, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 140, + 342 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 354, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 297, + 714 + ], + "type": "text", + "content": "Efficient Diffusion Models Diffusion models [13, 29, 32, 39] have sparked considerable interest in the commercial sector due to their high-quality image/video generation capabilities. These models have been progressively adapted for various applications, including text-to-image generation [2, 30, 31], image editing [1, 33], video generation [4, 5] and even perception [16, 19, 41]. However, diffusion models are currently limited by their slow speed in generating outputs. In response to this challenge, a variety of strategies have been proposed. One of the mainstreams is to approximate the SDE-based diffusion process [38, 39] through an ordinary differentiable equation (ODE) framework. For example, DPM and " + }, + { + "bbox": [ + 56, + 354, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{DPM}++" + }, + { + "bbox": [ + 56, + 354, + 297, + 714 + ], + "type": "text", + "content": " [22, 23] introduce ODE-based samplers, which significantly reduce the hundreds of denoising steps to between 15 and 20. Building upon the ODE formulation, InstaFlow [21] advances the reduction of denoising steps to a single instance through the novel strategy of rectified flow [20], while achieving performance close to that of Stable Diffusion [32]. Additionally, distillation from pre-trained diffusion models has also been explored as a method to facilitate few-step denoising. For instance, the consistency model [40] leverages the principle of self-consistency between noise at different denoising steps and uses pre-trained diffusion models [32] to guide the learning of a few-step denoising model, thereby enabling the generation of images within a minimal number of steps. In a notable extension of this concept, LCM [25, 26] applies the idea to the latent space rather than the pixel space. The use of distillation methods [35, 36, 42] to enhance the efficiency of the original Stable Diffusion model presents promising" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 331, + 556, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 556, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 556, + 463 + ], + "type": "text", + "content": "results. Besides improving efficiency through the lens of reducing denoising steps, quantization methods [14, 18] are proposed to make the model run in the regime of low float points, albeit with the potential trade-off of fidelity and efficiency. Moreover, parallel sampling [37] tries to utilize the approximate-parallel denoising strategy to improve the latency of the diffusion model. We emphasize that our work focuses on enhancing throughput and our work is orthogonal to [37]. Notably, our method is optimized for single-GPU, which are common among users. In contrast, the parallel sampling method reduces throughput on a single GPU." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 464, + 557, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 557, + 596 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 557, + 596 + ], + "type": "text", + "content": "Our proposed StreamDiffusion is significantly different from the approaches mentioned previously. While earlier methods primarily focus on the low latency of their individual model designs, our approach takes a different route. We introduce a pipeline-level solution specifically tailored for high throughput. Our pipeline can seamlessly integrate the low-latency diffusion models discussed above. Our proposed Stream Batch, residual classifier-free guidance, and the integration of other efficiency-enhancement methods focus on improving the efficiency of the whole pipeline instead of a single diffusion model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 617, + 558, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 558, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 558, + 715 + ], + "type": "text", + "content": "Classifier-free Guidance Classifier-free guidance [12] is widely used for conditional generation due to the simplicity, efficiency, and stability compared to classifier-guidance [9]. It leverages negative prompts [8, 11, 32] and essentially operates the vector arithmetic shift in latent space, i.e., we take a step of size (usually set by the guidance scale) away from the unconditional vector or negatively conditioned vector in the direction toward the conditioning manifold [12]. In" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "type": "text", + "content": "12373" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 68, + 298, + 213 + ], + "blocks": [ + { + "bbox": [ + 57, + 68, + 298, + 213 + ], + "lines": [ + { + "bbox": [ + 57, + 68, + 298, + 213 + ], + "spans": [ + { + "bbox": [ + 57, + 68, + 298, + 213 + ], + "type": "image", + "image_path": "9396a76707ed26a7c24187047be046b4ca5d318c8ba0c8d50d39a412d3e466fa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 217, + 297, + 319 + ], + "lines": [ + { + "bbox": [ + 55, + 217, + 297, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 297, + 319 + ], + "type": "text", + "content": "Figure 3. The concept of Stream Batch. In our approach, instead of waiting for a single image to be fully denoised before processing the next input image, we accept the next input image after each denoising step. This creates a denoising batch where the denoising steps are staggered for each image. By concatenating these staggered denoising steps into a batch, we can efficiently process continuous inputs using a U-Net for batch processing. The input image encoded at timestep " + }, + { + "bbox": [ + 55, + 217, + 297, + 319 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 217, + 297, + 319 + ], + "type": "text", + "content": " is generated and decoded at timestep " + }, + { + "bbox": [ + 55, + 217, + 297, + 319 + ], + "type": "inline_equation", + "content": "t + n" + }, + { + "bbox": [ + 55, + 217, + 297, + 319 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 217, + 297, + 319 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 217, + 297, + 319 + ], + "type": "text", + "content": " is the number of the denoising steps." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 336, + 298, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 336, + 298, + 529 + ], + "spans": [ + { + "bbox": [ + 54, + 336, + 298, + 529 + ], + "type": "text", + "content": "the practical implementation, the classifier-free guidance is conducted by sharing the same UNet for both the conditional term and unconditional (or negative) term and subtracting the effect of the unconditional term from the conditioned one. We point out that the way of multiple denoising processes leads to unnecessary computations for the unconditional (or negative) term. To get rid of these redundant computations, we propose a novel residual-classifier-free guidance, termed as R-CFG, which approximates the conditional noise prediction with only requiring one or even zero-time computation for the negatively conditioned noise prediction for the UNet. We note that " + }, + { + "bbox": [ + 54, + 336, + 298, + 529 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 54, + 336, + 298, + 529 + ], + "type": "text", + "content": "-CFG is especially designed for the SDEdit method [27], as we mainly focus on the applications of translating streaming image flow. Our proposed R-CFG significantly improves the latency of the conditional image-to-image generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 537, + 155, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 537, + 155, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 537, + 155, + 550 + ], + "type": "text", + "content": "3. StreamDiffusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 557, + 297, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 297, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 297, + 654 + ], + "type": "text", + "content": "StreamDiffusion is a new diffusion pipeline aiming for high throughput. It comprises three key components: Stream Batch strategy, Residual Classifier-Free Guidance (R-CFG), and Stochastic Similarity Filter. Besides, we also incorporate other acceleration methods like a novel input-output queue designed by us, the pre-computation procedure, the tinyautoencoder, and model acceleration tools such as TensorRT. We elaborate on the details below." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 660, + 272, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 272, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 272, + 673 + ], + "type": "text", + "content": "3.1. Stream Batch: Batching the Denoise Step" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 297, + 715 + ], + "type": "text", + "content": "In diffusion models, denoising steps are performed sequentially, resulting in a proportional increase in the processing time of U-Net relative to the number of steps. However," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "to generate high-fidelity images, it is necessary to increase the number of steps. To resolve this problem in interactive diffusion, we propose a method called Stream Batch." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 108, + 555, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 555, + 240 + ], + "type": "text", + "content": "The Stream Batch technique restructures sequential denoising operations into batched processes, wherein each batch corresponds to a predetermined number of denoising steps, as depicted in Fig. 3. The size of each batch is determined by the number of these denoising steps. This approach allows for each batch element to advance one step further in the denoising sequence via a single pass through U-Net. By iteratively applying this method, it is possible to effectively transform input images encoded at timestep " + }, + { + "bbox": [ + 313, + 108, + 555, + 240 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 108, + 555, + 240 + ], + "type": "text", + "content": " into their corresponding image-to-image results at timestep " + }, + { + "bbox": [ + 313, + 108, + 555, + 240 + ], + "type": "inline_equation", + "content": "t + n" + }, + { + "bbox": [ + 313, + 108, + 555, + 240 + ], + "type": "text", + "content": ", thereby streamlining the denoising procedure." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 239, + 556, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 556, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 556, + 347 + ], + "type": "text", + "content": "Stream Batch significantly reduces the need for multiple U-Net inferences. The processing time does not escalate linearly with the number of steps. This technique effectively shifts the trade-off from balancing processing time and generation quality to balancing VRAM capacity and generation quality. With adequate VRAM scaling, this method enables the production of high-quality images within the span of a single U-Net processing cycle, effectively overcoming the constraints imposed by increasing denoising steps." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 347, + 556, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 556, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 556, + 504 + ], + "type": "text", + "content": "Waiting and Batching can also increase the throughput of the diffusion pipeline. However, with naive Waiting and Batching (WB), denoising cannot begin immediately on the first input frame, leading to higher latency compared to Stream Batch. We mainly aim for smooth streaming applications. Yet achieving a smooth frame rate with WB requires additional engineering, such as precise inference speed estimation and input-output frame synchronization, and minor timing errors must be carefully managed. In contrast, Stream Batch automatically ensures a consistent interval between input and output frames, providing the advantage of lower latency while dynamically reaching the optimal throughput." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 510, + 542, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 542, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 542, + 523 + ], + "type": "text", + "content": "3.2. Improve Time Consistency by Stream Batch" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 527, + 556, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 556, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 556, + 671 + ], + "type": "text", + "content": "Maintaining temporal consistency in video generation is challenging. Many approaches ensure frame coherence by referencing past frames, often through cross-frame attention. However, our Stream Batch method uniquely enables temporal consistency using information from future frames. As shown in Fig. 3, Stream Batch allows simultaneous denoising of multiple frames, passing information from future frames to the current frame. This supports real-time image translation that adapts to sudden changes in input while preserving consistency. In Stream Batch with " + }, + { + "bbox": [ + 313, + 527, + 556, + 671 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 527, + 556, + 671 + ], + "type": "text", + "content": " denoising steps, keys and values for each frame at each time step form the following batches:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 331, + 678, + 535, + 694 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 678, + 535, + 694 + ], + "spans": [ + { + "bbox": [ + 331, + 678, + 535, + 694 + ], + "type": "interline_equation", + "content": "K _ {\\text {B a t c h}} = \\left[ K _ {t + i, 0}, \\dots , K _ {t, i}, \\dots , K _ {t - (n - 1 - i), n - 1} \\right]", + "image_path": "13974b2fb15d109143c56b563159c57ae3802cae7ad99133ca6963d40eb2236f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 337, + 699, + 531, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 699, + 531, + 717 + ], + "spans": [ + { + "bbox": [ + 337, + 699, + 531, + 717 + ], + "type": "interline_equation", + "content": "V _ {\\mathrm {B a t c h}} = \\left[ V _ {t + i, 0}, \\dots , V _ {t, i}, \\dots , V _ {t - (n - 1 - i), n - 1} \\right]", + "image_path": "0f7abab31f01fef601c9a8bfa9aa4eafbfb93763e00ef30ad0d091ea76c80400.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "text", + "content": "12374" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 69, + 331, + 210 + ], + "blocks": [ + { + "bbox": [ + 67, + 69, + 331, + 210 + ], + "lines": [ + { + "bbox": [ + 67, + 69, + 331, + 210 + ], + "spans": [ + { + "bbox": [ + 67, + 69, + 331, + 210 + ], + "type": "image", + "image_path": "944cdd937aa8e28e4add4912d4f0c5b19275100ebb59a74d75c0ea6559784250.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 219, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 219, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 219, + 555, + 243 + ], + "type": "text", + "content": "Figure 4. Virtual residual noise vectors: The orange vectors depict the virtual residual noise that starts from PF ODE trajectory and points to the original input latent " + }, + { + "bbox": [ + 55, + 219, + 555, + 243 + ], + "type": "inline_equation", + "content": "x_0" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 331, + 70, + 542, + 208 + ], + "blocks": [ + { + "bbox": [ + 331, + 70, + 542, + 208 + ], + "lines": [ + { + "bbox": [ + 331, + 70, + 542, + 208 + ], + "spans": [ + { + "bbox": [ + 331, + 70, + 542, + 208 + ], + "type": "image", + "image_path": "f3d56de7abc686ed7dda90ad0cc7dafb9ef629e2adab1525a2e356275051e5b0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 262, + 296, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 296, + 358 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 296, + 358 + ], + "type": "text", + "content": "These key and value batches incorporate information across different time frames and denoising steps. For example, if the frame at time step " + }, + { + "bbox": [ + 55, + 262, + 296, + 358 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 262, + 296, + 358 + ], + "type": "text", + "content": " has reached the " + }, + { + "bbox": [ + 55, + 262, + 296, + 358 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 262, + 296, + 358 + ], + "type": "text", + "content": "-th denoising step, the batch includes " + }, + { + "bbox": [ + 55, + 262, + 296, + 358 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 262, + 296, + 358 + ], + "type": "text", + "content": " future denoising steps for different frames and " + }, + { + "bbox": [ + 55, + 262, + 296, + 358 + ], + "type": "inline_equation", + "content": "n - 1 - i" + }, + { + "bbox": [ + 55, + 262, + 296, + 358 + ], + "type": "text", + "content": " past frames. In Stream Batch Cross-frame Attention, rather than using the typical " + }, + { + "bbox": [ + 55, + 262, + 296, + 358 + ], + "type": "inline_equation", + "content": "K_{t,i}" + }, + { + "bbox": [ + 55, + 262, + 296, + 358 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 262, + 296, + 358 + ], + "type": "inline_equation", + "content": "V_{t,i}" + }, + { + "bbox": [ + 55, + 262, + 296, + 358 + ], + "type": "text", + "content": ", we employ " + }, + { + "bbox": [ + 55, + 262, + 296, + 358 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{Batch}}" + }, + { + "bbox": [ + 55, + 262, + 296, + 358 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 262, + 296, + 358 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{Batch}}" + }, + { + "bbox": [ + 55, + 262, + 296, + 358 + ], + "type": "text", + "content": ", which integrate past and future frame information for the attention computation:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 376, + 296, + 404 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 296, + 404 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 296, + 404 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t n} \\left(Q _ {t, i}, K _ {\\text {B a t c h}}, V _ {\\text {B a t c h}}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {Q _ {t , i} \\cdot K _ {\\text {B a t c h}} ^ {T}}{\\sqrt {d}}\\right) V _ {\\text {B a t c h}}", + "image_path": "adedb1edaa1e768c058a97c70f43d2a3aa8fa8b5d4795cbe7aacd740d0745b6b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 412, + 296, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 296, + 449 + ], + "type": "text", + "content": "This approach enables the generation process to account for information from both past and future frames with across denoising stages, thus enhancing temporal consistency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 456, + 239, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 239, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 239, + 467 + ], + "type": "text", + "content": "3.3. Residual Classifier-Free Guidance" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 473, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 296, + 498 + ], + "type": "text", + "content": "Firstly, SDEdit based method [27] adds perturbation to the input image " + }, + { + "bbox": [ + 55, + 473, + 296, + 498 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 473, + 296, + 498 + ], + "type": "text", + "content": " and transfers it to the noise distribution " + }, + { + "bbox": [ + 55, + 473, + 296, + 498 + ], + "type": "inline_equation", + "content": "x_{\\tau_0}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 119, + 508, + 295, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 508, + 295, + 522 + ], + "spans": [ + { + "bbox": [ + 119, + 508, + 295, + 522 + ], + "type": "interline_equation", + "content": "x _ {\\tau_ {0}} = \\sqrt {\\alpha_ {\\tau_ {0}}} x _ {0} + \\sqrt {\\beta_ {\\tau_ {0}}} \\epsilon_ {0}, \\tag {1}", + "image_path": "9552bbc1a47848a340390c0dc5f0a6eb7a381cb77bb885650a05df01089f66c6.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 528, + 297, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 297, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 297, + 635 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 528, + 297, + 635 + ], + "type": "inline_equation", + "content": "\\alpha_{\\tau_0}" + }, + { + "bbox": [ + 55, + 528, + 297, + 635 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 528, + 297, + 635 + ], + "type": "inline_equation", + "content": "\\beta_{\\tau_0}" + }, + { + "bbox": [ + 55, + 528, + 297, + 635 + ], + "type": "text", + "content": " are values determined by a noise scheduler and " + }, + { + "bbox": [ + 55, + 528, + 297, + 635 + ], + "type": "inline_equation", + "content": "\\epsilon_0" + }, + { + "bbox": [ + 55, + 528, + 297, + 635 + ], + "type": "text", + "content": " is a sampled noise from a Gaussian " + }, + { + "bbox": [ + 55, + 528, + 297, + 635 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,I)" + }, + { + "bbox": [ + 55, + 528, + 297, + 635 + ], + "type": "text", + "content": ". When using consistency models for conditional image editing, " + }, + { + "bbox": [ + 55, + 528, + 297, + 635 + ], + "type": "inline_equation", + "content": "x_{\\tau_0}" + }, + { + "bbox": [ + 55, + 528, + 297, + 635 + ], + "type": "text", + "content": " can be considered as a point on the PF ODE trajectory, which leads to the conditioning manifold. To intensify the conditioning by Classifier-Free Guidance (CFG)[10], it is imperative to compute a noise for a negative condition PF ODE trajectory, which is used in vector arithmetic shift for the guidance (Eq. 2)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 108, + 647, + 296, + 661 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 647, + 296, + 661 + ], + "spans": [ + { + "bbox": [ + 108, + 647, + 296, + 661 + ], + "type": "interline_equation", + "content": "\\epsilon_ {\\tau_ {i}, \\mathrm {c f g}} = \\epsilon_ {\\tau_ {i}, \\bar {c}} + \\gamma \\left(\\epsilon_ {\\tau_ {i}, c} - \\epsilon_ {\\tau_ {i}, \\bar {c}}\\right), \\tag {2}", + "image_path": "7653864d66838e53d5c09d88c39932edd890972694917b31ee34612d077ca9f9.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": "This requirement introduces additional computational overhead at each denoising step. To reduce this computational overhead, R-CFG utilizes the fact that the original input image " + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": " is referable at any stage of denoising steps." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 262, + 555, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 555, + 322 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 555, + 322 + ], + "type": "text", + "content": "For any latent " + }, + { + "bbox": [ + 313, + 262, + 555, + 322 + ], + "type": "inline_equation", + "content": "x_{\\tau_i}" + }, + { + "bbox": [ + 313, + 262, + 555, + 322 + ], + "type": "text", + "content": " at the denoising step " + }, + { + "bbox": [ + 313, + 262, + 555, + 322 + ], + "type": "inline_equation", + "content": "\\tau_i" + }, + { + "bbox": [ + 313, + 262, + 555, + 322 + ], + "type": "text", + "content": ", we can assume the existence of the virtual negative condition " + }, + { + "bbox": [ + 313, + 262, + 555, + 322 + ], + "type": "inline_equation", + "content": "\\vec{c}_{\\tau_i}'" + }, + { + "bbox": [ + 313, + 262, + 555, + 322 + ], + "type": "text", + "content": ", that satisfies the self-consistency described as Eq. 3 This implies that " + }, + { + "bbox": [ + 313, + 262, + 555, + 322 + ], + "type": "inline_equation", + "content": "x_{\\tau_i}" + }, + { + "bbox": [ + 313, + 262, + 555, + 322 + ], + "type": "text", + "content": " is on the PF ODE trajectory going back to the input image " + }, + { + "bbox": [ + 313, + 262, + 555, + 322 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 262, + 555, + 322 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 368, + 335, + 555, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 335, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 368, + 335, + 555, + 350 + ], + "type": "interline_equation", + "content": "x _ {0} \\approx \\hat {x} _ {0, \\tau_ {i}, \\bar {c} _ {\\tau_ {i}} ^ {\\prime}} = f _ {\\theta} \\left(x _ {\\tau_ {i}}, \\tau_ {i}, \\bar {c} _ {\\tau_ {i}} ^ {\\prime}\\right) \\tag {3}", + "image_path": "175ba03a78d62931c2c70965a9ff9a8b2e55248bf01a9a782e45254f4ccd1211.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 354, + 555, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 402 + ], + "type": "text", + "content": "Following the LCM model parameterization [25] and our approximation for the inference time skip connections " + }, + { + "bbox": [ + 313, + 354, + 555, + 402 + ], + "type": "inline_equation", + "content": "(c_{\\mathrm{skip}}(\\tau) = 0, c_{\\mathrm{out}}(\\tau) = 1" + }, + { + "bbox": [ + 313, + 354, + 555, + 402 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 313, + 354, + 555, + 402 + ], + "type": "inline_equation", + "content": "\\tau \\neq 0" + }, + { + "bbox": [ + 313, + 354, + 555, + 402 + ], + "type": "text", + "content": "), the self-consistency equation (Eq. 3) can be expressed as follows," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 382, + 411, + 555, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 411, + 555, + 442 + ], + "spans": [ + { + "bbox": [ + 382, + 411, + 555, + 442 + ], + "type": "interline_equation", + "content": "x _ {0} \\approx \\frac {x _ {\\tau_ {i}} - \\sqrt {\\beta_ {\\tau_ {i}}} \\epsilon_ {\\tau_ {i} , \\tilde {c} _ {\\tau_ {i}} ^ {\\prime}}}{\\sqrt {\\alpha_ {\\tau_ {i}}}} \\tag {4}", + "image_path": "5b26ef6c8d3bf7b62bd2f004364a3e6cfa486aa87c8f1f46dfaf5706b7e9700e.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 447, + 556, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 556, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 556, + 507 + ], + "type": "text", + "content": "Given the initial value " + }, + { + "bbox": [ + 313, + 447, + 556, + 507 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 447, + 556, + 507 + ], + "type": "text", + "content": ", and the subsequent values of " + }, + { + "bbox": [ + 313, + 447, + 556, + 507 + ], + "type": "inline_equation", + "content": "x_{\\tau_i}" + }, + { + "bbox": [ + 313, + 447, + 556, + 507 + ], + "type": "text", + "content": " obtained sequentially through the iterative denoising, the virtual noise vector " + }, + { + "bbox": [ + 313, + 447, + 556, + 507 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\tau_i,\\vec{c}'}" + }, + { + "bbox": [ + 313, + 447, + 556, + 507 + ], + "type": "text", + "content": " in the direction toward the input image can be analytically determined by employing these values with the Eq. 4:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 386, + 516, + 555, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 516, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 386, + 516, + 555, + 544 + ], + "type": "interline_equation", + "content": "\\epsilon_ {\\tau_ {i}, \\bar {c} ^ {\\prime}} = \\frac {x _ {\\tau_ {i}} - \\sqrt {\\alpha_ {\\tau_ {i}}} x _ {0}}{\\sqrt {\\beta_ {\\tau_ {i}}}} \\tag {5}", + "image_path": "cd8bbcd5e7227710dff52fbb52336c18df2d06be581758e4c4963c11fd348cf5.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 548, + 555, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 555, + 572 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 555, + 572 + ], + "type": "text", + "content": "With the virtual noise " + }, + { + "bbox": [ + 313, + 548, + 555, + 572 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\tau_i,\\overline{c'}}" + }, + { + "bbox": [ + 313, + 548, + 555, + 572 + ], + "type": "text", + "content": " obtained from Eq. 5, we formulate R-CFG by:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 361, + 583, + 555, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 583, + 555, + 597 + ], + "spans": [ + { + "bbox": [ + 361, + 583, + 555, + 597 + ], + "type": "interline_equation", + "content": "\\epsilon_ {\\tau_ {i}, \\text {c f g}} = \\delta \\epsilon_ {\\tau_ {i}, \\bar {c} ^ {\\prime}} + \\gamma \\left(\\epsilon_ {\\tau_ {i}, c} - \\delta \\epsilon_ {\\tau_ {i}, \\bar {c} ^ {\\prime}}\\right) \\tag {6}", + "image_path": "eee977c04da839eeae05277a30660d9f27690026dc4d2bd1625ef32eef63dc2e.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 605, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 605, + 554, + 640 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 605, + 554, + 640 + ], + "type": "text", + "content": " is a magnitude moderation coefficient for the virtual residual noise that softens the effect and the approximation error of the virtual residual noise." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "text", + "content": "R-CFG that uses the original input image latent " + }, + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "text", + "content": " as the residual term can effectively generate results that diverge from the original input image according to the magnitude of the guidance scale " + }, + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "text", + "content": ", thereby enhancing the effect of conditioning without the need for additional U-Net computations. We call this method Self-Negative R-CFG." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "type": "text", + "content": "12375" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "Not only to deviate from the original input image " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": ", but also to diverge from any negative condition, we can find the desired reference point " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "\\hat{x}_{0,\\tau_0,\\bar{c}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": " that reflects the negative condition " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "\\bar{c}" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": " using the same self-consistency formulation:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 119, + 129, + 296, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 129, + 296, + 159 + ], + "spans": [ + { + "bbox": [ + 119, + 129, + 296, + 159 + ], + "type": "interline_equation", + "content": "\\hat {x} _ {0, \\tau_ {0}, \\bar {c}} = \\frac {x _ {\\tau_ {0}} - \\sqrt {\\beta_ {\\tau_ {0}}} \\epsilon_ {\\tau_ {0} , \\bar {c}}}{\\sqrt {\\alpha_ {\\tau_ {0}}}} \\tag {7}", + "image_path": "bf40624cc82f7b3f148836b45137b1d014a5641d5a285fc01bc3cc31f9f8afa4.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 163, + 296, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 163, + 296, + 199 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 296, + 199 + ], + "type": "text", + "content": "We can obtain " + }, + { + "bbox": [ + 55, + 163, + 296, + 199 + ], + "type": "inline_equation", + "content": "\\hat{x}_{0,\\tau_0,\\bar{c}}" + }, + { + "bbox": [ + 55, + 163, + 296, + 199 + ], + "type": "text", + "content": " by computing the actual negative conditioned noise " + }, + { + "bbox": [ + 55, + 163, + 296, + 199 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\tau_0,\\bar{c}}" + }, + { + "bbox": [ + 55, + 163, + 296, + 199 + ], + "type": "text", + "content": " using U-Net only one time for the first denoising step." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 200, + 296, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 200, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 54, + 200, + 296, + 308 + ], + "type": "text", + "content": "In Eq. 5, instead of " + }, + { + "bbox": [ + 54, + 200, + 296, + 308 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 54, + 200, + 296, + 308 + ], + "type": "text", + "content": ", using " + }, + { + "bbox": [ + 54, + 200, + 296, + 308 + ], + "type": "inline_equation", + "content": "\\hat{x}_{0,\\tau_0,\\bar{c}}" + }, + { + "bbox": [ + 54, + 200, + 296, + 308 + ], + "type": "text", + "content": ", we can obtain the virtual negative conditioned noise " + }, + { + "bbox": [ + 54, + 200, + 296, + 308 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\tau_{i + 1},\\bar{c}'}" + }, + { + "bbox": [ + 54, + 200, + 296, + 308 + ], + "type": "text", + "content": " that can effectively diverge the generation results from the controllable negative conditioning " + }, + { + "bbox": [ + 54, + 200, + 296, + 308 + ], + "type": "inline_equation", + "content": "\\bar{c}" + }, + { + "bbox": [ + 54, + 200, + 296, + 308 + ], + "type": "text", + "content": ". We name this method Onetime-Negative R-CFG. In contrast to the conventional CFG, which requires " + }, + { + "bbox": [ + 54, + 200, + 296, + 308 + ], + "type": "inline_equation", + "content": "2n" + }, + { + "bbox": [ + 54, + 200, + 296, + 308 + ], + "type": "text", + "content": " computations of U-Net, the Self-Negative RCFG and Onetime-Negative RCFG necessitate only " + }, + { + "bbox": [ + 54, + 200, + 296, + 308 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 54, + 200, + 296, + 308 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 200, + 296, + 308 + ], + "type": "inline_equation", + "content": "n + 1" + }, + { + "bbox": [ + 54, + 200, + 296, + 308 + ], + "type": "text", + "content": " computations of U-Net, respectively, where " + }, + { + "bbox": [ + 54, + 200, + 296, + 308 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 54, + 200, + 296, + 308 + ], + "type": "text", + "content": " is the number of the denoising steps." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 315, + 205, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 205, + 328 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 205, + 328 + ], + "type": "text", + "content": "3.4. Stochastic Similarity Filter" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 333, + 296, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 333, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 333, + 296, + 453 + ], + "type": "text", + "content": "When images remain unchanged or show minimal changes, particularly in scenarios without active user interaction or static environment, nearly identical input images are often repeatedly fed into the VAE and U-Net. This leads to the generation of identical or nearly identical images and unnecessary consumption of GPU resources. In contexts involving continuous inputs, such instances of unmodified input images can occasionally occur. To tackle this issue and minimize unnecessary computational load, we propose a strategy termed stochastic similarity filter (SSF), as shown in Fig. 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 453, + 296, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 296, + 477 + ], + "type": "text", + "content": "We calculate the cosine similarity between the current input image " + }, + { + "bbox": [ + 55, + 453, + 296, + 477 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 55, + 453, + 296, + 477 + ], + "type": "text", + "content": " and the past reference frame image " + }, + { + "bbox": [ + 55, + 453, + 296, + 477 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{ref}}" + }, + { + "bbox": [ + 55, + 453, + 296, + 477 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 121, + 485, + 295, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 485, + 295, + 512 + ], + "spans": [ + { + "bbox": [ + 121, + 485, + 295, + 512 + ], + "type": "interline_equation", + "content": "S _ {C} \\left(I _ {t}, I _ {\\text {r e f}}\\right) = \\frac {I _ {t} \\cdot I _ {\\text {r e f}}}{\\| I _ {t} \\| \\| I _ {\\text {r e f}} \\|} \\tag {8}", + "image_path": "81eacbf48cb540897b4c935aa01ed51f5bba9b8c9b4dd3cb293b2032f63a0fbc.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 516, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 296, + 552 + ], + "type": "text", + "content": "Based on this cosine similarity, we calculate the probability of skipping the subsequent VAE and U-Net processes. It is given by" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 571, + 296, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 571, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 69, + 571, + 296, + 597 + ], + "type": "interline_equation", + "content": "\\mathbf {P} (\\operatorname {s k i p} \\mid I _ {t}, I _ {\\text {r e f}}) = \\max \\left\\{0, \\frac {S _ {C} \\left(I _ {t} , I _ {\\text {r e f}}\\right) - \\eta}{1 - \\eta} \\right\\}, \\tag {9}", + "image_path": "10247dcd56680535d6624a95264ea27f75a5fb98b448270d9385c5c21276c56e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " is the similarity threshold. This probability decides whether subsequent processes like VAE Encoding, U-Net, and VAE Decoding should be skipped or not. If not skipped, the input image at that time is saved and updated as the reference image " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "I_{ref}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " for future use. This probabilistic skipping mechanism allows the network to operate fully in dynamic scenes with low inter-frame similarity, while in static scenes with high inter-frame similarity, the network's operational rate decreases, conserving computational resources. The" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 553, + 189 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 553, + 189 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 553, + 189 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 553, + 189 + ], + "type": "image", + "image_path": "5a47214f6b866ccb49e15649e48244f706232c02eb1b3646da105c33aa4555c9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 199, + 555, + 222 + ], + "lines": [ + { + "bbox": [ + 313, + 199, + 555, + 222 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 555, + 222 + ], + "type": "text", + "content": "Figure 5. Average inference time comparison between StreamBatch and normal sequential denoising." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 243, + 553, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 553, + 279 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 553, + 279 + ], + "type": "text", + "content": "GPU usage is modulated seamlessly based on the similarity of the input images, enabling smooth adaptation to scenes with varying dynamics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 279, + 555, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 555, + 387 + ], + "type": "text", + "content": "Note: We emphasize that compared to determining whether we skip the compute via a hard threshold, the proposed probability-sampling-based similarity filtering strategy leads to smoother video generation. Because the hard threshold is prone to making the video stuck, which hurts the impression of watching video streaming, while the sampling-based method significantly improves the smoothness. For the other efficiency improvement methods, we illustrate them in the supplementary material." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 398, + 395, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 395, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 395, + 411 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 418, + 556, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 556, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 556, + 538 + ], + "type": "text", + "content": "We implement StreamDiffusion pipeline upon LCM, LCM-LoRA [25, 26] and SD-turbo [36]. As a model accelerator, we use TensorRT and for the lightweight efficient VAE, we use TAESD [17]. Our pipeline is compatible with the customer-level GPU. We test our pipeline on NVIDIA RTX4090 GPU, Intel Core i9-13900K CPU, Ubuntu 22.04.3 LTS, and NVIDIA RTX3060 GPU, Intel Core i7-12700K, Windows 11 for image generation. We note that we evaluate the throughput mainly via the average inference time per image through processing 100 images." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 545, + 450, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 450, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 450, + 558 + ], + "type": "text", + "content": "4.1.Quantitative Evaluation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 563, + 556, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 556, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 556, + 696 + ], + "type": "text", + "content": "We compare our method with the AutoPipelineForImage2Image, which is a pipeline developed by Huggingface diffusers1. The average inference time comparison is presented in Table. 1. Our pipeline demonstrates a substantial speed increase. When we use TensorRT, StreamDiffusion achieves a minimum speed-up of 13.0 times when running the 10 denoising steps, and reaching up to 59.6 times in scenarios involving a single denoising step. Even though without TensorRT, StreamDiffusion achieves a 29.7 times speed up compared to AutoPipeline when using one step denoising, and an 8.3 times speedup at 10 step denoising." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 702, + 523, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 702, + 523, + 712 + ], + "spans": [ + { + "bbox": [ + 324, + 702, + 523, + 712 + ], + "type": "text", + "content": "1https://github.com/huggingface/diffusers" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "text", + "content": "12376" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 102, + 294, + 171 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 297, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 297, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 297, + 94 + ], + "type": "text", + "content": "Table 1. Average inference time (ms) at different denoising steps. AutoPipeline refers to Diffusers' implementation [15]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 102, + 294, + 171 + ], + "lines": [ + { + "bbox": [ + 58, + 102, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 58, + 102, + 294, + 171 + ], + "type": "table", + "html": "
StepStreamDiffusionw/o TRTAutoPipeline
110.65 (59.6x)21.34 (29.7x)634.40 (1x)
216.74 (39.3x)30.61 (21.3x)652.66 (1x)
426.93 (25.8x)48.15 (14.4x)695.20 (1x)
1062.00 (13.0x)96.94 (8.3x)803.23 (1x)
", + "image_path": "2db1e4dbc923c7c9993b3c9d2320305e94f77688c34b4888438406b129c8301e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 213, + 296, + 291 + ], + "blocks": [ + { + "bbox": [ + 55, + 182, + 295, + 205 + ], + "lines": [ + { + "bbox": [ + 55, + 182, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 295, + 205 + ], + "type": "text", + "content": "Table 2. Average inference time (ms) at different denoising steps across CFG methods" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 213, + 296, + 291 + ], + "lines": [ + { + "bbox": [ + 58, + 213, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 58, + 213, + 296, + 291 + ], + "type": "table", + "html": "
StepSelf-Neg. R-CFGOne-Neg. R-CFGCFG
111.04 (1.52x)16.55 (1.01x)16.74 (1x)
216.61 (1.64x)20.64 (1.32x)27.18 (1x)
320.64 (1.74x)27.25 (1.32x)35.91 (1x)
426.19 (1.90x)31.65 (1.57x)49.71 (1x)
531.47 (2.05x)36.04 (1.79x)64.64 (1x)
", + "image_path": "61e4b6d93e4b598924d9285137dedfe47e56cd650f9f593b5521e25da319df3e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 312, + 297, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 297, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 297, + 396 + ], + "type": "text", + "content": "Efficiency comparison regarding Stream Batch. The efficiency comparison between Stream Batch and the original sequential U-Net loop is shown in Fig. 5. When implementing a denoising batch strategy, we observe a significant improvement in processing time. It achieves a reduction by half when compared to a conventional U-Net loop at sequential denoising steps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 412, + 297, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 297, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 297, + 616 + ], + "type": "text", + "content": "Efficiency comparison regarding R-CFG. Table 2 presents a comparison of the inference times for StreamDiffusion pipelines with R-CFG and conventional CFG. The additional computations required to apply Self-Negative R-CFG are merely lightweight vector operations, resulting in negligible changes in inference time compared to when Self-Negative is not used. When employing Onetime-Negative R-CFG, additional UNet computations are necessary for the first step of the denoising process. Therefore, One-time-negative R-CFG and conventional CFG have almost identical inference times for a single denoising step case. However, as the number of denoising steps increases, the difference in inference time from conventional CFG to both Self-Negative and Onetime-Negative R-CFG becomes more pronounced. At denoising step 5, a speed improvement of " + }, + { + "bbox": [ + 55, + 412, + 297, + 616 + ], + "type": "inline_equation", + "content": "2.05\\mathrm{x}" + }, + { + "bbox": [ + 55, + 412, + 297, + 616 + ], + "type": "text", + "content": " is observed with Self-Negative R-CFG and " + }, + { + "bbox": [ + 55, + 412, + 297, + 616 + ], + "type": "inline_equation", + "content": "1.79\\mathrm{x}" + }, + { + "bbox": [ + 55, + 412, + 297, + 616 + ], + "type": "text", + "content": " with Onetime-Negative R-CFG, compared to conventional CFG." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 624, + 178, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 178, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 178, + 637 + ], + "type": "text", + "content": "4.2. Energy Consumption" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 297, + 715 + ], + "type": "text", + "content": "We then conduct a comprehensive evaluation of the energy consumption associated with our proposed stochastic similarity filter (SSF), as depicted in Figure. 11 and Figure. 12. These figures provide the GPU utilization patterns when SSF (Threshold " + }, + { + "bbox": [ + 55, + 642, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 642, + 297, + 715 + ], + "type": "text", + "content": " set at 0.98) is applied to input videos containing scenes with periodic static characteristics. The comparative" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "content": "analysis reveals that the incorporation of SSF significantly mitigates GPU usage in instances where the input images are predominantly static or with a high degree of similarity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 109, + 556, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 556, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 556, + 228 + ], + "type": "text", + "content": "Figure. 11 delineates the results derived from a meticulously executed two-denoise-step img2img experiment. This experiment was conducted on a 20-frame video sequence, employing NVIDIA RTX3060 graphics processing units with or without the integration of SSF. The experiment results indicate a substantial decrease in average power consumption from 85.96w to 35.91w on one RTX3060 GPU. Using the same static scene input video with one NVIDIA RTX4090GPU, the power consumption was reduced from 238.68w to 119.77w." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 229, + 557, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 229, + 557, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 557, + 373 + ], + "type": "text", + "content": "Furthermore, Figure. 12 expounds on the findings from a similar two-denoise-step img2img experiment using one RTX4090GPU. This time the evaluation of energy consumption is performed on a 1000-frame video featuring dynamic scenes. Remarkably, even under drastically dynamic conditions, the SSF efficiently extracted several frames exhibiting similarity from the dynamic sequence. This process results in a noteworthy reduction in average power consumption, from 236.13w to 199.38w. These findings underscore the efficacy of the Stochastic Similarity Filter in enhancing energy efficiency, particularly in scenarios involving static or minimally varying visual content." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 382, + 406, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 382, + 406, + 396 + ], + "spans": [ + { + "bbox": [ + 314, + 382, + 406, + 396 + ], + "type": "text", + "content": "4.3. Ablation study" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 316, + 411, + 552, + 571 + ], + "blocks": [ + { + "bbox": [ + 316, + 411, + 552, + 571 + ], + "lines": [ + { + "bbox": [ + 316, + 411, + 552, + 571 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 552, + 571 + ], + "type": "image", + "image_path": "6a23082892a1ecbe77c3aef8b58969b4ab462c0bc315df88c90cb19ca7d60c27.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 580, + 526, + 592 + ], + "lines": [ + { + "bbox": [ + 342, + 580, + 526, + 592 + ], + "spans": [ + { + "bbox": [ + 342, + 580, + 526, + 592 + ], + "type": "text", + "content": "Figure 6. Ablation study on different components." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "text", + "content": "In our ablation study, as summarized in Fig. 6, we evaluate the average inference time of our proposed method under various configurations to understand the contribution of each component. Our proposed StreamDiffusion achieves an average inference time of " + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "inline_equation", + "content": "10.98 / 9.42\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "inline_equation", + "content": "26.93 / 26.30\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "text", + "content": " for denoising steps 1 and 4 on image-to-image/text-to-image generation, respectively. When the R-CFG is not used, we observe this results in the second largest efficiency drop. This demonstrates that R-CFG is one of the most critical" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "text", + "content": "12377" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 155 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 155 + ], + "type": "text", + "content": "components in our pipeline. When the Stream Batch processing is removed ('w/o stream batch'), we observe a large time consumption increase, especially at 4 denoising steps. We also evaluate the impact on the inference time of our pipeline regarding SSF. We observe that SSF plays a significant role in enabling energy saving and does not introduce extra time cost." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 156, + 298, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 156, + 298, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 156, + 298, + 373 + ], + "type": "text", + "content": "Besides, the absence of TensorRT ('w/o TRT') leads to a large increase in time cost. The removal of pre-computation also results in increased time cost but not much. We attribute the reason to the limited number of key-value computations in Stable Diffusion. Besides, the exclusion of input-output queue ('w/o IO queue') also demonstrates an impact on average inference time, which mainly aims to optimize the parallelization issue resulting from pre- and post-processing. In the AutoPipelineImage2Image's adding noise function, the precision of tensors is converted from fp32 to fp16 for each request, leading to a decrease in speed. In contrast, the StreamDiffusion pipeline standardizes the precision of variables and computational devices beforehand. It does not perform tensor precision conversion or computation device transfers during inference. Consequently, even without any optimization ('w/o any optimization'), our pipeline significantly outperforms the AutoPipelineImage2Image in terms of throughputs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 380, + 294, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 294, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 294, + 393 + ], + "type": "text", + "content": "4.4. Quantitative Evaluation for the Image Quality." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 398, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 398, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 295, + 506 + ], + "type": "text", + "content": "We conduct a quantitative evaluation on the image quality. Specifically, we first evaluate the FID and CLIP scores on the text-to-image generation. We use the same dataset as [6] for the evaluation. We use LCM as our main baseline for the comparison. Note that our method is never trained; our method still improves LCM by a large margin in terms of the FID (29.69 vs. 26.79) and maintains a similar CLIP score (24.95 vs. 24.99). This demonstrates the effectiveness of our proposed method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": "User study. We conduct a user study to validate the visual quality of different components of our StreamDiffusion. The results are shown in the Table. 3. Our Stream Batch with future-frame attention significantly enhances time consistency compared to its absence. Additionally, our SSF method addresses the crucial yet rarely explored issue of energy efficiency. While SSF is simple, it is far from trivial: a common approach might apply a hard threshold on similarity to regulate streaming flow. However, as noted in the main text, we introduce a novel approach—using probability sampling to achieve superior streaming quality. As Table 3 illustrates, this approach is preferred by more users over the hard-threshold method for visual quality. Moreover, compared to the vanilla CFG method, both our self-Negative R-CFG and one-time R-CFG are more preferred by the users, demonstrating our method not only improves efficiency but" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 72, + 392, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 392, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 392, + 84 + ], + "type": "text", + "content": "also visual quality." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 316, + 99, + 553, + 190 + ], + "blocks": [ + { + "bbox": [ + 316, + 99, + 553, + 190 + ], + "lines": [ + { + "bbox": [ + 316, + 99, + 553, + 190 + ], + "spans": [ + { + "bbox": [ + 316, + 99, + 553, + 190 + ], + "type": "table", + "html": "
# UsersWithout Future-Frame (%)With Future-Frame (%)
4869.6790.33
# UsersWithout SSF (%)With SSF (%)
14448.6151.39
# UsersHard Threshold (%)SSF (%)
4524.4475.56
# UsersWith / Without CFG (%)Self-Neg R-CFG / One-Time R-CFG (%)
25722 / 2035 / 23
", + "image_path": "46fadac35959cc0a3713ea2d7888a0800f26f84406b456f295e471980257b6c2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 194, + 553, + 238 + ], + "lines": [ + { + "bbox": [ + 313, + 194, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 553, + 238 + ], + "type": "text", + "content": "Table 3. User study results: Future-frame attention consistency (486 users), SSF quality imperceptibility (144 users), streaming quality between hard threshold filter and SSF (45 users), comparison of with/without CFG and self-negative/one-time R-CFG (257 users)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 255, + 555, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 363 + ], + "type": "text", + "content": "This pipeline enables image generation with very low throughput from input images received in real-time from cameras or screen capture devices. At the same time, it is capable of producing high-quality images that effectively align to the specified prompt conditions. These capabilities demonstrate the applicability of our pipeline in various real-time applications, such as real-time game graphic rendering, generative camera effect filters, real-time face conversion, and AI-assisted drawing." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 380, + 387, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 380, + 387, + 392 + ], + "spans": [ + { + "bbox": [ + 314, + 380, + 387, + 392 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 402, + 556, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 556, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 556, + 641 + ], + "type": "text", + "content": "We propose StreamDiffusion, a pipeline-level solution for interactive diffusion generation. StreamDiffusion consists of several optimization strategies for both throughput and GPU usage, including Stream Batch with cross-attention, residual classifier-free guidance (R-CFG), IO-queue for parallelization, stochastic similarity filter, pre-computation, Tiny AutoEncoder, and the use of the model acceleration tool. The synergistic combination of these elements results in a marked improvement in efficiency. Specifically, StreamDiffusion achieves up to 91.07 frames per second (fps) on a standard consumer-grade GPU for image generation tasks. This performance level is particularly beneficial for a variety of applications, including but not limited to the Metaverse, online video streaming, and broadcasting sectors. Furthermore, StreamDiffusion demonstrates a significant reduction in GPU power consumption, achieving at least a " + }, + { + "bbox": [ + 313, + 402, + 556, + 641 + ], + "type": "inline_equation", + "content": "1.99\\mathrm{x}" + }, + { + "bbox": [ + 313, + 402, + 556, + 641 + ], + "type": "text", + "content": " decrease. This notable efficiency gain underscores StreamDiffusion's potential for commercial application, offering a compelling solution for energy-conscious, high-performance computing environments." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 658, + 424, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 424, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 424, + 671 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 679, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 679, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 679, + 555, + 713 + ], + "type": "text", + "content": "We sincerely thank Taku Fujimoto and Huggingface team for their invaluable feedback, courteous support, and insightful discussions." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "type": "text", + "content": "12378" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 135, + 295, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 135, + 295, + 157 + ], + "spans": [ + { + "bbox": [ + 61, + 135, + 295, + 157 + ], + "type": "text", + "content": "[2] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM Trans. Graph., 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 158, + 296, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 158, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 158, + 296, + 213 + ], + "type": "text", + "content": "[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, Wesam Manassra, Prafulla Dhariwal, Casey Chu, Yunxin Jiao, and Aditya Ramesh. Improving image generation with better captions. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 213, + 295, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 213, + 295, + 268 + ], + "spans": [ + { + "bbox": [ + 62, + 213, + 295, + 268 + ], + "type": "text", + "content": "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 270, + 296, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 296, + 334 + ], + "type": "text", + "content": "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 335, + 296, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 296, + 368 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 296, + 368 + ], + "type": "text", + "content": "[6] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 369, + 296, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 296, + 413 + ], + "type": "text", + "content": "[7] Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. arXiv preprint arXiv:2303.04137, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 414, + 296, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 296, + 468 + ], + "type": "text", + "content": "[8] Katherine Crowson, Stella Biderman, Daniel Kornis, Dashiell Stander, Eric Hallahan, Louis Castricato, and Edward Raff. Vqgan-clip: Open domain image generation and editing with natural language guidance. In European Conference on Computer Vision, pages 88–105. Springer, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 469, + 295, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 469, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 62, + 469, + 295, + 502 + ], + "type": "text", + "content": "[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 502, + 296, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 502, + 296, + 545 + ], + "spans": [ + { + "bbox": [ + 57, + 502, + 296, + 545 + ], + "type": "text", + "content": "[10] Yilun Du, Shuang Li, and Igor Mordatch. Compositional visual generation with energy based models. In Advances in Neural Information Processing Systems, pages 6637-6647. Curran Associates, Inc., 2020. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 546, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 546, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 546, + 296, + 590 + ], + "type": "text", + "content": "[11] Yilun Du, Shuang Li, and Igor Mordatch. Compositional visual generation with energy based models. Advances in Neural Information Processing Systems, 33:6637-6647, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 591, + 295, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 613 + ], + "type": "text", + "content": "[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 613, + 295, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 647 + ], + "type": "text", + "content": "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 647, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 689 + ], + "type": "text", + "content": "[14] Yushi Huang, Ruihao Gong, Jing Liu, Tianlong Chen, and Xianglong Liu. Tfmq-dm: Temporal feature maintenance quantization for diffusion models. arXiv preprint arXiv:2311.16503, 2023. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 690, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 690, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 690, + 296, + 712 + ], + "type": "text", + "content": "[15] Huggingface. https://huggingface.co/docs/diffusers/, 2024.2, 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "text", + "content": "[16] Aliasghar Khani, Saeid Asgari, Aditya Sanghi, Ali Mahdavi Amiri, and Ghassan Hamarneh. Slime: Segment like me. In The Twelfth International Conference on Learning Representations, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 118, + 555, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 555, + 140 + ], + "type": "text", + "content": "[17] Diederik P Kingma and Max Welling. Auto-encoding variational bayes, 2022. 6, 13" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 141, + 555, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 555, + 196 + ], + "type": "text", + "content": "[18] Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-diffusion: Quantizing diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 17535-17545, 2023. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 197, + 554, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 554, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 554, + 229 + ], + "type": "text", + "content": "[19] Ziyi Li, Qinye Zhou, Xiaoyun Zhang, Ya Zhang, Yanfeng Wang, and Weidi Xie. Open-vocabulary object segmentation with diffusion models. 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 230, + 554, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 554, + 263 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 554, + 263 + ], + "type": "text", + "content": "[20] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 264, + 554, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 554, + 308 + ], + "type": "text", + "content": "[21] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, and Qiang Liu. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. arXiv preprint arXiv:2309.06380, 2023. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 309, + 554, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 554, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 554, + 364 + ], + "type": "text", + "content": "[22] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 365, + 554, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 554, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 554, + 408 + ], + "type": "text", + "content": "[23] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 410, + 554, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 554, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 554, + 443 + ], + "type": "text", + "content": "[24] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 444, + 554, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 444, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 444, + 554, + 498 + ], + "type": "text", + "content": "[25] Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module. arXiv preprint arXiv:2311.05556, 2023. 2, 3, 5, 6, 11" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 499, + 554, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 499, + 554, + 543 + ], + "spans": [ + { + "bbox": [ + 316, + 499, + 554, + 543 + ], + "type": "text", + "content": "[26] Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module, 2023. 3, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 544, + 554, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 554, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 554, + 599 + ], + "type": "text", + "content": "[27] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 4, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 600, + 488, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 600, + 488, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 600, + 488, + 612 + ], + "type": "text", + "content": "[28] OpenAI. https://openai.com/sora, 2024.2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 613, + 554, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 554, + 635 + ], + "type": "text", + "content": "[29] William Peebles and Saining Xie. Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748, 2022. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 635, + 554, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 554, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 554, + 678 + ], + "type": "text", + "content": "[30] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2): 3, 2022. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "type": "text", + "content": "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 3, 13" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "text", + "content": "12379" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 542 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "type": "text", + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 296, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 296, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 296, + 172 + ], + "type": "text", + "content": "[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 295, + 239 + ], + "type": "text", + "content": "[34] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho†, David Fleet, and Mohammad Norouzi. Imagen: unprecedented photorealism × deep level of language understanding. 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 240, + 295, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 240, + 295, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 295, + 272 + ], + "type": "text", + "content": "[35] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 274, + 295, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 295, + 306 + ], + "type": "text", + "content": "[36] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 2, 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 296, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 296, + 350 + ], + "type": "text", + "content": "[37] Andy Shih, Suneel Belkhale, Stefano Ermon, Dorsa Sadigh, and Nima Anari. Parallel sampling of diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 353, + 295, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 295, + 385 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 295, + 385 + ], + "type": "text", + "content": "[38] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 387, + 295, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 387, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 387, + 295, + 430 + ], + "type": "text", + "content": "[39] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 432, + 295, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 432, + 295, + 463 + ], + "spans": [ + { + "bbox": [ + 56, + 432, + 295, + 463 + ], + "type": "text", + "content": "[40] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 465, + 295, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 465, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 465, + 295, + 498 + ], + "type": "text", + "content": "[41] Chenfeng Xu, Huan Ling, Sanja Fidler, and Or Litany. 3diff- tection: 3d object detection with geometry-aware diffusion features, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 499, + 295, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 499, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 499, + 295, + 542 + ], + "type": "text", + "content": "[42] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Frédo Durand, William T. Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. arXiv, 2023. 2, 3" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "text", + "content": "12380" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/StreamGS_ Online Generalizable Gaussian Splatting Reconstruction for Unposed Image Streams/9d1940fe-237f-4495-9085-9441928f0c47_content_list.json b/2025/StreamGS_ Online Generalizable Gaussian Splatting Reconstruction for Unposed Image Streams/9d1940fe-237f-4495-9085-9441928f0c47_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5657f1dcefae4a73f66332f78f432cfa95b5822c --- /dev/null +++ b/2025/StreamGS_ Online Generalizable Gaussian Splatting Reconstruction for Unposed Image Streams/9d1940fe-237f-4495-9085-9441928f0c47_content_list.json @@ -0,0 +1,1413 @@ +[ + { + "type": "text", + "text": "StreamGS: Online Generalizable Gaussian Splitting Reconstruction for Unposed Image Streams", + "text_level": 1, + "bbox": [ + 153, + 128, + 843, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yang Li $^{1,4,*}$ Jinglu Wang $^{1}$ Lei Chu $^{1}$ Xiao Li $^{1}$ Shiu-Hong Kao $^{1,2,*}$ Ying-Cong Chen $^{2,3}$ Yan Lu $^{1}$ $^{1}$ Microsoft Research Asia ${}^{2}$ HKUST ${}^{3}$ HKUST(GZ) ${}^{4}$ Rutgers University yangliaftermath@gmail.com skao@cse.ust.hk yingcongchen@ust.hk {jinglu.wang, leichu, li.xiao, yanlu}@microsoft.com", + "bbox": [ + 181, + 200, + 815, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The advent of 3D Gaussian Splatting (3DGS) has advanced 3D scene reconstruction and novel view synthesis. With the growing interest of interactive applications that need immediate feedback, online 3DGS reconstruction in real-time is in high demand. However, none of existing methods yet meet the demand due to three main challenges: the absence of predetermined camera parameters, the need for generalizable 3DGS optimization, and the necessity of reducing redundancy. We propose StreamGS, an online generalizable 3DGS reconstruction method for unposed image streams, which progressively transform image streams to 3D Gaussian streams by predicting and aggregating per-frame Gaussians. Our method overcomes the limitation of the initial point reconstruction [27] in tackling out-of-domain (OOD) issues by introducing a content adaptive refinement. The refinement enhances cross-frame consistency by establishing reliable pixel correspondences between adjacent frames. Such correspondences further aid in merging redundant Gaussians through cross-frame feature aggregation. The density of Gaussians is thereby reduced, empowering online reconstruction by significantly lowering computational and memory costs. Extensive experiments on diverse datasets have demonstrated that StreamGS achieves quality on par with optimization-based approaches but does so 150 times faster, and exhibits superior generalizability in handling OOD scenes.", + "bbox": [ + 89, + 359, + 483, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 786, + 220, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The field of 3D Scene reconstruction [10, 30] for novel view synthesis from image streams has gained increasing attention, due to its significance in enabling interactive applications that offer users instant feedback. In this context, the", + "bbox": [ + 89, + 811, + 483, + 873 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a2731129e2b61360219e70938d881361d034a47e89913f3e31c904ee8fdc7188.jpg", + "image_caption": [ + "Figure 1. The proposed StreamGS efficiently transforms image streams into Gaussian streams by progressively reconstructing and aggregating per-frame 3D Gaussians. We show our reconstructed 3DGS (visualized as points) alongside estimated camera poses (in blue), and synthesized novel views." + ], + "image_footnote": [], + "bbox": [ + 524, + 327, + 897, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "advent of 3D Gaussian Splatting (3DGS) [12] marks a major advancement in high-quality, real-time rendering. This progress highlights the urgency for efficient, on-the-fly generation of 3DGS from image streams.", + "bbox": [ + 511, + 779, + 906, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nonetheless, online 3DGS reconstruction faces unique challenges. (1) Unknown camera. The conventional preprocessing using Structure from Motion (SfM) [21] for camera estimation is impractical for real-time streaming.", + "bbox": [ + 511, + 840, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done during the internship at MSRA.", + "bbox": [ + 114, + 886, + 351, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25841", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This is due to the absence of the full image set and the time-consuming computation. Thus, there is a need for methods that can operate without pre-determined cameras. (2) Generalizability. 3DGS reconstruction requires multiteration optimization, impractical for online applications due to the need for all images in advance. This restricts the development of a generalizable method that can process image streams in a feed-forward manner. (3) Redundancy. The significant overlap between frames leads to high redundancy when reconstructing Gaussians individually for each frame, increasing the streaming process's resource intensity.", + "bbox": [ + 89, + 90, + 480, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Scene reconstruction without known cameras has been explored in SLAM-based [30, 34, 35, 40, 41] and NeRF-based [1, 23] methods. Only a few 3DGS-based methods are discussed [9, 20]. These methods primarily consider camera poses as learnable parameters and optimized alongside Gaussians through the iterative optimization. Yet, this optimization-driven approach for aligning each frame considerably increases the reconstruction time, rendering it impractical in the online scenario.", + "bbox": [ + 89, + 260, + 480, + 395 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, generalizable 3DGS reconstruction has been investigated for sparse views [2, 3, 24]. These approaches transform pixels into Gaussians, whose parameters are decoded from images via 2D encoder-decoder networks. They achieve the 3DGS reconstruction for each image in a single feed-forward pass. However, these generalizable approaches are primarily designed for monocular or binocular settings, suited to sparse-view inputs. In addition, multi-view methods [2, 3] typically infer Gaussian centers through stereo matching, highly dependent on known cameras, limiting their applicability for online scenarios with numerous unposed images. Furthermore, they combine multi-view 3DGS sets by simply uniting them, which overlooks cross-view alignment and overlaps, leading to misalignment and redundancy issues within image streams. The recent method [7] performs Gaussian downsampling in 3D space to reduce redundancy, but it necessitates traversing and processing Gaussians in 3D grids, which is also time-consuming and is highly dependent on the grid resolution.", + "bbox": [ + 89, + 398, + 482, + 699 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Emerging models like DUSt3R [27] and MASt3R [15] enable sparse-view geometry reconstruction by simultaneously predicting 3D points and estimating camera parameters. These advances pave the way for more efficient, feedforward, pose-free 3D reconstruction methods. A straightforward approach to generalize 3DGS reconstruction is to add a Gaussian predictor to the DUSt3R-like framework. However, this introduces several challenges. These models usually require datasets with ground truth 3D geometry, which may not always be available. Additionally, applying pretrained models to out-of-domain (OOD) data can result in inaccurate pose and 3D point estimations. Moreover, generating 3D points for each frame individually causes re", + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "dundancy due to overlapping adjacent frames, potentially leading to ghosting artifacts from pose estimation errors.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce StreamGS, a novel pipeline for online, generalizable 3DGS reconstruction from unposed image streams. StreamGS aims to progressively construct and update the 3DGS representation of the scene frame-by-frame, in a feed-forward manner, as illustrated in Fig. 1. We leverage the pretrained DUSt3R to initially predict 3D point for the current frame using the previous frame as a reference. However, this initialization may encounter inaccuracies due to OOD issues. To mitigate this, we capitalize the insight that adjacent frames offer sufficient correspondences to refine the reconstruction. Unlike DUSt3R using predicted 3D points to establish correspondences, we adopt content-adaptive descriptors for more reliable matching, allowing for the adaptive refinement of the reconstruction by enhancing consistency between adjacent views. Furthermore, such correspondences help to prune redundant pixel-aligned Gaussians. Correlated pixel-wise features across frames are effectively aggregated, removing duplicates and achieving adaptive density control. Finally, we decode Gaussians from such aggregated features. StreamGS is adept at predicting and integrating Gaussians for the current frame into the existing Gaussian set seamlessly with a feed-forward pass. In summary, our contribution is three-fold.", + "bbox": [ + 511, + 122, + 903, + 482 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a novel pipeline for the online, generalizable reconstruction of image streams without requiring camera parameters, marking a first in this field.", + "- The proposed adaptive refinement enhance cross-frame consistency of 3DGS reconstruction, and the adaptive density control mechanism minimizes adjacent-view redundancy, thereby highly reducing computational costs in online reconstruction.", + "- Upon evaluation across diverse datasets, our method achieves high novel view synthesis quality comparable to the optimization-based method [9] but with $150\\mathrm{x}$ faster reconstruction speed. Additionally, our method outperforms existing pose-dependent generalizable 3DGS methods in handling OOD scenes, showing superior generalizability." + ], + "bbox": [ + 513, + 483, + 903, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 723, + 663, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generalizable 3D Gaussian Splitting. Many recent studies aim to propose generalizable 3D-GS methods capable of predicting Gaussians within a single feed-forward pass. These works can be classified into two main categories: single-view reconstruction and multi-view reconstruction. Single-view reconstruction does not involve pose estimation as there are no multi-view constraints. Inspired by the insight from LRM [11] that large transformer-based [26] backbone networks can learn 3D priors from large-scale 3D data, the potential of predicting Gaussians from a", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25842", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "single image in a single feed-forward pass has been comprehensively explored. Numerous feed-forward models have been proposed, such as GRM [29], TriplaneGS [42], and GMamba [22]. However, these methods are primarily not applicable to multi-view scenarios as they always assume canonical poses.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Concurrently, many works focusing on multi-view inputs follow a similar paradigm, such as GS-LRM [36], LGM [25], and MVGMamba [31]. They concatenate the input images with camera embeddings like Plücker rays to facilitate the network in learning the proper fusion of Gaussians from multi-view inputs. However, these large 3D backbone networks mostly perform well only on synthetic objects due to the shortage of large-scale scene-level 3D data in the real world. Referring to generalizable NeRF methods like pixelNeRF [32], other research turns to multi-view stereo (MVS) matching to locate or initialize the centers of Gaussians, with other attributes decoded using a lightweight 2D encoder. Representative works with this design include MVSGaussian [18], pixelSplat [2], and MVSplat [3]. However, both camera ray embedding in large transformer-based models and stereo matching rely on known poses and intrinsicities of each input view. Another main limitation of these methods is that they focus on sparse-view inputs. PixelSplat [2] and MVSplat [3] only support up to three views. Therefore, existing generalizable 3D-GS models cannot address the problem of feed-forward reconstruction from endless image streams.", + "bbox": [ + 91, + 183, + 486, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pose-free 3D Gaussian Splatting. Recently, many works have aimed to eliminate the need for Structure-from-Motion (SfM) preprocessing steps using COLMAP [21] software. Following the design of previous pose-free NeRF methods like NoPe-NeRF [1], Lu-NeRF [4], and localRF [19], CF-3DGS [9] introduces depth priors into the optimization of 3D-GS and performs progressive reconstruction. As each new image arrives, CF-3DGS optimizes both the pose and 3D Gaussians of the input image based on its depths, aiming to align the 3DGS from the new image with the preceding reconstruction. However, it still relies on known camera intrinsics. CF-3DGS is also not robust, as the accuracy of depth priors significantly impacts its reconstruction quality, limiting its application to common scenes. Moreover, it depends on thousands of optimization steps for each view, significantly extending the reconstruction time for each scene. Compared to generalizable 3D-GS models, current pose-free methods are so inefficient that they cannot be applied to image streams with a large number of frames.", + "bbox": [ + 89, + 532, + 483, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Online 3D reconstruction of image streams. Online 3D reconstruction has been extensively studied in the field of SLAM [30, 34, 35, 39-41]. However, these methods typically involve additional information and most leverage SDF", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and NeRF as scene representations. NICE-SLAM [39] uses RGB-D streams as input, with the reconstructed scene represented by NeRF. NICER-SLAM [40] relies on geometric priors, including surface normals and depths. Surfel-NeRF [10] focuses on the novel view synthesis quality of online reconstruction with RGB streams, but it requires the poses and intrinsics of frames. Gaussian-SLAM [34] reconstructs the scene using 3D-GS, but it also relies on RGB-D streams.", + "bbox": [ + 511, + 90, + 906, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 513, + 239, + 612, + 253 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a sequence of unposed images over time, our objective is to progressively reconstruct the 3D Gaussian Splating (3DGS) representation in an online manner. Specifically, at each timestamp $t$ , the goal is to derive the 3DGS $\\mathcal{G}^t$ , which encapsulates the 3D scene aggregated from the images $\\mathcal{I}^t = \\{\\mathbf{I}^i\\}_{i=1}^t$ .", + "bbox": [ + 511, + 265, + 906, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 2 illustrates our overall framework. In order to make reconstruction efficient with limited computational resources, we employ an incremental construction strategy. At each time step $t$ , we focus on generating the 3DGS $\\mathbf{G}^t$ of the current frame and merge it with the previous accumulated reconstruction $\\mathcal{G}^{t-1}$ to obtain the full reconstruction $\\mathcal{G}^t$ of the current time step. Specifically, with current frame $\\mathbf{I}^t$ , we use $\\mathbf{I}^{t-1}$ as reference frame and estimate the point maps and cameras of each using an initial reconstruction module. With newly established matches between current and reference views, we further refine the quality of the points and the cameras in the adaptive refinement module. Finally, we generate $\\mathbf{G}^t$ using the refined points and features of $\\mathbf{I}^t$ , and merge it with previous reconstructions according to the established matches, achieving the feedforward adaptive density control (ADC).", + "bbox": [ + 511, + 356, + 908, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 604, + 653, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Gaussian splatting. Previous work [12] represents a scene or object using a set of Gaussian distributions. Specifically each gaussian primitive could be denoted as $G(x;\\mu ,\\Sigma) = e^{-\\frac{1}{2} (x - \\mu)^T\\Sigma^{-1}(x - \\mu)}$ and the covariance $\\Sigma$ is decomposed into the rotation matrix $R$ and scaling matrix $S$ to ensure the positive semi-definiteness during optimization, that is $\\Sigma = RSS^T R^T$ . The view-dependent color of the appearance is represented by a set of spherical harmonics (SH) coefficients and opacity value $\\alpha$ .", + "bbox": [ + 511, + 627, + 908, + 765 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Initial Two-view Reconstruction", + "text_level": 1, + "bbox": [ + 511, + 773, + 797, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the current frame $\\mathbf{I}^t$ and a reference frame $\\mathbf{I}^{t'}$ , we use a coarse predictor $\\phi_{3D}$ and estimate the point map $\\mathbf{X}^{t|t'}$ of the current frame under the local coordinate system of the reference frame together with its corresponding confidence map $\\mathbf{C}^{t|t'}$ , where the superscript $\\cdot |t'$ indicates that the local coordinate system adapts to that of $\\mathbf{I}^{t'}$ . Intuitively, the predicted point map stores the 3D point coordinate that the", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25843", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fe2c56a6263ac031dabbc6ba493b884753d04e3d39a3df14e5e387effe744e8c.jpg", + "image_caption": [ + "Figure 2. Method overview. Our StreamGS progressively reconstruct and aggregate 3D Gaussians from the unposed image stream. Given the adjacent image pair $(\\mathbf{I}^{t - 1},\\mathbf{I}^t)$ , we first perform the initial reconstruction that predicts pixel-wise 3D points with their features and coarse camera poses, using a pretrained coarse predictor. Since the coarse predictions may suffer from OOD issues, we refine both the camera poses and 3D positions by establishing new point-wise correspondences. We aggregate cross-frame image and 3D features by warping and merging to reduce redundancy. Finally we decode the aggregated features to Gaussian primitives." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 903, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "pixel unprojected to in, and the confidence maps measure the certainty of point maps at each pixel, reflecting a prior-based estimation of reconstruction accuracy and difficulty. Formally, we define", + "bbox": [ + 89, + 414, + 483, + 474 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\mathbf {X} ^ {t \\mid t ^ {\\prime}}, \\mathbf {X} ^ {t ^ {\\prime} \\mid t ^ {\\prime}}, \\mathbf {C} ^ {t \\mid t ^ {\\prime}}, \\mathbf {C} ^ {t ^ {\\prime} \\mid t ^ {\\prime}}\\right) = \\phi_ {3 D} \\left(\\mathbf {I} ^ {t}, \\mathbf {I} ^ {t ^ {\\prime}}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 486, + 482, + 507 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to help align the current local-coordinate pointmap with the global coordinate, we also get the output in the local coordinate system of $\\mathbf{I}^t$ by computing $\\phi_{3D}(\\mathbf{I}^{t'},\\mathbf{I}^t)$ . With these predicted point maps, we could further estimate the camera matrix $\\mathbf{P} = \\mathbf{K}[\\mathbf{R}|\\mathbf{t}]$ , composed of its intrinsic parameters $\\mathbf{K}$ , and its extrinsic parameters, the rotation matrix $\\mathbf{R}$ and the translation vector $\\mathbf{t}$ . Specifically, we assume that the principle points $\\mathbf{c}$ are centered and pixels are squares and have", + "bbox": [ + 89, + 511, + 483, + 648 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {f} ^ {t} = \\arg \\min _ {f} \\Sigma_ {\\mathbf {p} \\in \\mathbf {I} ^ {t}} \\| \\mathbf {p} - \\mathbf {c} - f \\frac {\\left(\\mathbf {x} _ {\\mathbf {p}} , \\mathbf {y} _ {\\mathbf {p}}\\right)}{\\mathbf {z} _ {\\mathbf {p}}} \\|, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 657, + 482, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $(\\mathbf{x_p},\\mathbf{y_p},\\mathbf{z_p})\\in \\mathbf{X}^{t|t}$ is the 3D point coordinate that the pixel $\\mathbf{p}$ unprojected to. Moreover, we approximate relative pose $\\mathbf{P}^t = [\\mathbf{R},\\mathbf{t}]$ of $\\mathbf{I}^t$ to $\\mathbf{I}^{t'}$ by solving the following points registration problem:", + "bbox": [ + 89, + 700, + 483, + 763 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\mathbf {R} ^ {t} \\mid \\mathbf {t} ^ {t} \\right] = \\arg \\min _ {s, \\mathbf {R}, \\mathbf {t}} \\sum_ {\\mathbf {p} \\in \\mathbf {I} ^ {t - 1}} \\mathbf {C} ^ {t} (\\mathbf {p}) \\| s \\left(\\mathbf {R} \\mathbf {X} ^ {t ^ {\\prime} | t ^ {\\prime}} (\\mathbf {p}) + \\mathbf {t}\\right) - \\mathbf {X} ^ {t ^ {\\prime} | t} (\\mathbf {p}) \\| ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 771, + 480, + 799 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} ^ {t} = \\mathbf {C} ^ {t ^ {\\prime} | t ^ {\\prime}} \\odot \\mathbf {C} ^ {t ^ {\\prime} | t},\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 800, + 246, + 815 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\odot$ is the Hadamard product and $s$ is the scale factor.", + "bbox": [ + 89, + 825, + 477, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our implementation, we leverage DUSt3R [27] as the coarse predictor due to its effective pretraining on dedicated 3D scene datasets and its efficiency in reconstructing 3D points. For each time step $t$ , we process the pair", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$(\\mathbf{I}^{t - 1},\\mathbf{I}^t)$ to derive the 3D points $(\\mathbf{X}^{t|t - 1},\\mathbf{X}^{t - 1|t - 1})$ corresponding to both frames referenced in the coordinate frame $t - 1$ , the reversed pair $(\\mathbf{I}^t,\\mathbf{I}^{t - 1})$ to obtain the same points $(\\mathbf{X}^{t - 1|t},\\mathbf{X}^{t|t})$ but in the coordinate frame of $t$ , which will be reused at the next timestamp to boost efficiency. For simplicity, and without loss of generality, the following discussion will focus on the reconstruction of the image pair $(\\mathbf{I}^{t - 1},\\mathbf{I}^t)$ . Note that training $\\phi_{3D}$ requires 3D geometric supervision, which may not be available in our monocular video input scenario. Thus, we leverage the pretrained $\\phi_{3D}$ from [27].", + "bbox": [ + 511, + 412, + 906, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Adaptive Refinement", + "text_level": 1, + "bbox": [ + 511, + 590, + 712, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We note that the initial reconstruction quality is compromised due to the OOD challenge as the coarse predictor is frozen. This observation motivates us to enhance reconstruction through content adaptation, with the goal of adaptively refining both poses and 3D reconstructions.", + "bbox": [ + 511, + 612, + 905, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We perform the adaptive refinement based on establishing new robust matches between adjacent frames. We employ a matching head, $\\phi_{\\text{match}}$ , to extract local 3D features, denoted as $(\\mathbf{F}_{3D}^{t-1}, \\mathbf{F}_{3D}^t) \\in \\mathbb{R}^{H \\times W \\times d}$ , from the consecutive image pair $(\\mathbf{I}^{t-1}, \\mathbf{I}^t)$ . The matches between the two images can be established through nearest reciprocal (NN) searching, satisfying the following condition:", + "bbox": [ + 511, + 686, + 905, + 792 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {M} ^ {t - 1, t} = \\left\\{i _ {k} \\leftrightarrow j _ {k} \\mid i _ {k} = \\mathrm {N N} (j _ {k}) \\text {a n d} j _ {k} = \\mathrm {N N} (i _ {k}) \\right\\} _ {k = 1} ^ {N}, \\\\ \\text {s . t .} \\operatorname {N N} \\left(i _ {k}\\right) = \\underset {0 \\leq j _ {k} \\leq H \\times W} {\\arg \\min } | 1 - \\cos < \\mathbf {F} _ {3 D, i _ {k}} ^ {t - 1}, \\mathbf {F} _ {3 D, j _ {k}} ^ {t} > |, \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 803, + 903, + 843 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $i_k, j_k$ are pixel index in image $\\mathbf{I}^{t-1}$ , $\\mathbf{I}^t$ respectively, and the measurement of feature distance is cosine similarity. With the correspondences found, a residual transform", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25844", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\Delta = [\\Delta \\mathbf{R}, \\Delta \\mathbf{t}]$ can be re-estimated following Eq. (3) by only taking matched points into account. Then we apply the residual transform to the pointmap $\\mathbf{X}^{t,t-1}$ to retain refined $\\tilde{\\mathbf{X}}^{t|t-1}$ , and the pose of $\\mathbf{I}^t$ is updated to $\\tilde{\\mathbf{P}}^t$ by performing PnP-RANSAC [8, 14] on 3D-2D correspondences derived from matches.", + "bbox": [ + 89, + 90, + 483, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Gaussian decoding. We directly predict the other parameters of 3D Gaussians at each pixel with a light-weight decoder $\\phi_{GS}$ as:", + "bbox": [ + 89, + 209, + 483, + 255 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {G} ^ {i} = \\left[ \\mathbf {q} ^ {i}, \\mathbf {s} ^ {i}, \\alpha^ {i}, \\mathbf {c} ^ {i} \\right] = \\phi_ {G S} (\\mathbf {F} _ {g s}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 268, + 483, + 292 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {G S} ^ {i} = \\mathbf {F} _ {2 D} ^ {i} \\oplus \\mathbf {X} ^ {i} \\oplus \\mathbf {F} _ {3 D} ^ {i}, \\quad \\mathbf {F} _ {2 D} = \\phi_ {2 D} (\\mathbf {I} ^ {i}),\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 289, + 441, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $i = \\{t - 1,t\\}$ $\\oplus$ denotes the channel-wise concatenation, $\\phi_{2D}$ denotes the 2D image feature extractor, $\\mathbf{q}^i\\in \\mathbb{R}^{H\\times W\\times 4}$ and $\\mathbf{s}^i\\in \\mathbb{R}^{H\\times W\\times 3}$ represent rotation quaternions and scales of pixel-aligned Gaussians at image $\\mathbf{I}^i$ . We incorporate an additional image feature extractor because the coarse predictor is frozen and cannot be trained by our monocular video setting. Extracting new image features is essential for decoding Gaussians, especially for texture-related properties. Experiments show the importance of the image feature extractor. Then covariance matrix is built with $\\Sigma^i = \\mathbf{R}(\\mathbf{q}^i)\\mathbf{ss}^{\\mathrm{T}}\\mathbf{R}(\\mathbf{q}^i)^{\\mathrm{T}}$ Note that $\\bar{\\mathbf{G}}$ is not the final Gaussians since it needs to be merged into the previous Gaussian set following Sec. 3.4.", + "bbox": [ + 89, + 321, + 483, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Feed-Forward ADC", + "text_level": 1, + "bbox": [ + 89, + 532, + 282, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With pixel-aligned Gaussian parameters $\\mathbf{G}^t$ of $T$ images, previous methods [2, 3, 24, 25] always naively take the union of Gaussians in all images as the final prediction, i.e., $\\mathcal{G} = \\bigcup_{t=1}^{T} \\mathbf{G}^i$ and $|\\mathcal{G}| = T \\times H \\times W$ . However, this approach is both memory-intensive and inefficient in rendering, particularly when dealing with the continuous input of video frames during online reconstruction. Our key observation is that the matched Gaussian pairs in neighboring frames are excessive and prunable since they consistently share similar attributes in shape and color, and are closely distributed, which is validated in Tab. 1. It is noted that we have already acquired dense pixel-wise matches between neighboring frames from Eq. (4). Therefore, we propose a novel feed-forward Adaptive Density Control strategy based on revisiting these dense correspondences.", + "bbox": [ + 89, + 555, + 483, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature aggregation. The primary advantage of pixel-based correspondences is that they are able to convert the computationally intensive 3D Gaussian aggregation process into a more efficient 2D pixel-wise one. This significantly enhances computational efficiency. Initially, the feature of Gaussian parameters $\\mathbf{F}_{gs}^{t}$ of the frame $\\mathbf{I}^t$ can be aligned to", + "bbox": [ + 89, + 810, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the previous frame $\\mathbf{I}^{t - 1}$ using the following wrapping:", + "bbox": [ + 511, + 90, + 874, + 107 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {G S} ^ {t \\mid t - 1} (j) = \\left\\{ \\begin{array}{l l} \\mathbf {F} _ {G S} ^ {t} (k) & \\text {i f} (j, k) \\in \\overline {{\\mathcal {M}}} ^ {t - 1, t} \\\\ \\mathbf {F} _ {G S} ^ {t - 1} (j) & \\text {e l s e ,} \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 112, + 903, + 154 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{F}_{GS}^{t|t-1}$ denotes the feature of Gaussian primitives in the frame $\\mathbf{I}^t$ aligned to $\\mathbf{I}^{t-1}$ . Instead of using the raw correspondences set $\\mathcal{M}^{t-1,t}$ in Eq. (4), we use the extended set $\\overline{\\mathcal{M}}^{t-1,t}$ , which includes matches of neighboring pixels such as $(i+1,j+1)$ in addition to the initial $(i,j)$ . It serves as a type of anti-aliasing technique to reduce void pixels within the wrapped feature map. As for the unmatched pixels, we simply replicate the corresponding feature vector within $\\mathbf{F}_{gs}^{t-1}$ . With the aligned feature, the Gaussian feature of two frames can be merged by modifying Eq. (5), taking the form:", + "bbox": [ + 511, + 161, + 906, + 330 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {G}} ^ {t \\mid t - 1} = \\left[ \\hat {\\mathbf {q}} ^ {t}, \\hat {\\mathbf {s}} ^ {t}, \\hat {\\alpha} ^ {t}, \\hat {\\mathbf {c}} ^ {t} \\right] = \\phi_ {M G} \\left(\\mathbf {F} _ {G S} ^ {t \\mid t - 1} \\oplus \\mathbf {F} _ {G S} ^ {t - 1}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 337, + 903, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\phi_{MG}$ denotes the MergeNet that simply consists of two convolutional layers, which merges features and decodes them to Gaussian primitives. In this way, every matched pair of Gaussians between $\\mathbf{I}^t$ and $\\mathbf{I}^{t - 1}$ is aggregated into a single one, highly reducing the number of Gaussians. The final aggregated Gaussian set is $\\mathcal{G}^t = \\mathcal{G}^{t - 1}\\cup \\mathbf{G}^{t|t - 1}$ .", + "bbox": [ + 511, + 364, + 906, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Without gradient back-propagation of rendering loss in the original 3DGS paper [12], our ADC process runs exceptional fast. In terms of the input group, the final prediction of Gaussian primitives $\\mathcal{G}^t$ consists of the merged Gaussians and the unmatched Gaussians at each frame.", + "bbox": [ + 511, + 470, + 905, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Loss Functions", + "text_level": 1, + "bbox": [ + 511, + 551, + 666, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The optimization of the 2D feature extractor $\\phi_{2D}$ , the Gaussian decoder network, $\\phi_{GS}$ involves both a rendering loss function and a reconstruction loss function:", + "bbox": [ + 511, + 574, + 905, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} \\left(\\mathbf {I} ^ {i}, \\hat {\\mathbf {I}} ^ {i}, \\hat {\\mathbf {I}} _ {M} ^ {i}\\right) = \\mathcal {L} _ {\\text {r e n d e r}} \\left(\\mathbf {I} ^ {i}, \\hat {\\mathbf {I}} ^ {i}\\right) + \\mathcal {L} _ {\\text {r e c o n}} \\left(\\hat {\\mathbf {I}} ^ {i}, \\hat {\\mathbf {I}} _ {M} ^ {i}\\right) \\tag {8} \\\\ = \\| \\mathbf {I} ^ {i} - \\hat {\\mathbf {I}} ^ {i} \\| _ {2} + \\lambda \\| \\mathbf {I} ^ {i} - \\hat {\\mathbf {I}} ^ {i} \\| _ {\\mathrm {L P I P S}} + \\| \\hat {\\mathbf {I}} _ {\\mathrm {M}} ^ {i} - \\hat {\\mathbf {I}} ^ {i} \\| _ {2}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 625, + 903, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{I}^i$ is the ground truth frame, $\\hat{\\mathbf{I}}^i$ is the rendered image with full Gaussian primitives before the merge process, and $\\hat{\\mathbf{I}}_{\\mathrm{M}}^{i}$ is the rendered image with merged Gaussians. Due to the lightweight nature of the two networks, the algorithm converges quickly within thousands of steps. The reconstruction loss term facilitates the merge network in fusing the pixel-aligned Gaussian parameters from different frames, aiming to render the same frame as before the merge process but with significantly fewer Gaussian primitives.", + "bbox": [ + 511, + 666, + 906, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 830, + 645, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "StreamGS operates on an image stream of a scene, jointly predicting the corresponding poses and Gaussians in a feedforward manner. To assess its performance, we evaluate our", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25845", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3773a140c5269d1cf04d021af47224f0b9a46b889946cf7a6c9f11769c20b429.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
mean ↓Opa. ↓Rot. ↓Scale ↓SH ↓
Random Pick0.410.103.312.470.27
Matched Pairs0.180.043.061.280.04
", + "bbox": [ + 106, + 88, + 470, + 140 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Similarities of attributes between matched GS across adjacent frames on RE10K. Random Pick refers to randomly picking two GS from two frames separately, while Matched Pairs refers to our matched GS defined in Eq. (4).", + "bbox": [ + 89, + 151, + 482, + 209 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "method on the task of novel view synthesis from monocular videos, as detailed in Sec. 4.1. Additionally, we validate the effectiveness of the proposed alignment module and the efficiency of the Gaussians merge process, as described in Sec. 4.2.", + "bbox": [ + 89, + 215, + 482, + 290 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines and datasets. To the best of our knowledge, StreamGS is the first method to reconstruct unposed videos in a feed-forward manner. Consequently, we compare our method separately with pose-free 3DGS works and generalizable splatting methods, including pixelSplat [2], MVSplat [3], and CF-3DGS [9]. PixelSplat and MVSplat are representative methods of generalizable Gaussian Splatting, but they both rely on known camera poses and intrinsics. In contrast, CF-3DGS is a pose-free method but requires optimization loops to align poses and Gaussians. For a comprehensive comparison, we evaluate the methods on large-scale datasets with diverse scenes, including RE10K [38], ACID [17], ScanNet [5], DL3DV [16], and MVImgNet [33]. Each dataset comprises monocular video sequences with per-frame camera pose annotations.", + "bbox": [ + 89, + 291, + 482, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. Both the 2D image feature extractor $\\phi_{2D}$ and the MergeNet $\\phi_{GS}$ are double-layer convolutional networks. For fairness, StreamGS and other generalizable methods are trained on the identical training split of the RE10K dataset using the Adam [13] optimizer with the same learning rate of $2\\times 10^{-4}$ and a cosine scheduler. The parameter $\\lambda$ in Eq. 8 is set to 0.05. All methods are trained for 30K iterations and tested on a single NVIDIA Tesla A100 80GB GPU. The images are resized to $224\\times 224$ and the batch size is set to 14. For the non-generalizable method CF-3DGS, we follow its original setting [9]. Note that CF-3DGS is evaluated only on a subset of the full test set due to its low reconstruction efficiency, as shown in Fig. 5. Unlike pose-dependent methods, CF-3DGS and our method require poses of novel views for rendering. CF-3DGS freezes the trained Gaussian model and performs additional optimization steps to learn the poses. Our method, as described in Sec. 3.3, carries out pose alignment and refinement processes to estimate the poses. More details can be referred to supplementary materials.", + "bbox": [ + 89, + 517, + 482, + 820 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Novel View Synthesis", + "text_level": 1, + "bbox": [ + 89, + 829, + 290, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.1. Reconstruction Quality", + "text_level": 1, + "bbox": [ + 89, + 851, + 299, + 867 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative comparison. We compare StreamGS with baseline methods on the quality and efficiency of novel view", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "synthesis. Following previous 3D-GS research [12], we report PSNR, SSIM [28], and LPIPS [37] as metrics of reconstruction quality. The quantitative results are shown in Tab. 2. On the source domain RE10K [38], existing state-of-the-art generalizable methods demonstrate competitive scores, with MVSplat even outperforming the optimization-based CF-3DGS. However, their performance degrades significantly on out-of-domain datasets such as DL3DV [16] and MVImgNet [33], as these datasets contain various scene types, including outdoor environments and more complex indoor scenes with different objects and illumination conditions compared to RE10K. Naturally, CF-3DGS, which is based on thousands of optimization steps, performs well on the aforementioned datasets. However, its PSNR slightly decreases on ScanNet due to more irregular camera movements and increased motion blur in the frames, posing a challenge for CF-3DGS in recovering camera poses. Pixel-Splat performs well on ScanNet, thanks to its robust feature extraction backbone trained on ImageNet [6], while MVSplat achieves the lowest score. According to the table, StreamGS consistently achieves scores comparable to all other methods. Its PSNR on MVImgNet and DL3DV is significantly higher than that of PixelSplat and MVSplat, even without given poses and intrinsics, demonstrating our method's superior generalizability on unseen datasets with significant domain gaps.", + "bbox": [ + 511, + 90, + 906, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative comparison. A qualitative comparison with state-of-the-art methods is also presented in Fig. 3 and Fig. 4. While all methods demonstrate high-quality novel view rendering on RE10K [38], our method exhibits superior robustness on out-of-domain datasets. Due to the significant domain gap including texture, illumination and camera motion, both MVSplat and pixelSplat fail to predict accurate depths for the printer shown in the second row, resulting in severe floating artifacts in the rendered image. The third row shows the reconstruction of a plaza from DL3DV [16]. Similarly, since outdoor scenes are much less represented in RE10K [38], MVSplat struggles to extract stereo cues from text-less skies and objects with disparate illumination, deteriorating the rendering quality. The final row shows a bench scene in MVImgNet, on which baseline generalizable methods also fails due to the similar domain gap issues. These cases also demonstrate that the generalizability of pixelSplat surpasses that of MVSplat. As the figure shows, CF-3DGS also does not perform well on some outdoor scenes. In contrast, the visual quality of our method on out-of-domain datasets remains high.", + "bbox": [ + 511, + 493, + 908, + 811 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.2. Reconstruction Efficiency", + "text_level": 1, + "bbox": [ + 511, + 820, + 738, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition to rendering quality, we also compare the reconstruction efficiency of our method with baseline models. Fig. 5 illustrates a plot of reconstruction quality, measured by the average PSNR reported in Tab. 2, versus ef", + "bbox": [ + 511, + 840, + 905, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25846", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1b6fca2ee1ea35e662e47c3319c3b2c4e6f2e86ef6d12dd921e379ee03013178.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PF GRE10K [38] (Source Domain)DL3DV [16]MVIImgNet [33]ScanNet [5]ACID [17]
PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑
pixelSplat [2]×✓22.640.210.7519.430.390.5019.220.490.6027.560.190.8229.930.140.84
MVSplat [3]×✓23.540.130.9017.840.360.4516.290.500.5326.250.190.8228.830.140.86
CF-GS [9]✓×23.460.240.7519.930.310.6226.330.330.8822.880.420.7628.160.190.81
StreamGS✓✓22.420.170.8320.540.240.6425.050.310.7928.430.160.8628.500.150.84
", + "bbox": [ + 93, + 88, + 906, + 186 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/64fe47016b2b937b89ce2ce319e4501ae392d490663133c0dc5c39db9502ff56.jpg", + "table_caption": [ + "Table 2. Quantitative comparison with existing state-of-art methods on Novel View Synthesis of monocular videos. PF indicates whether the method is pose-free. G indicates whether the methods is generalizable. Our method consistently achieves scores comparable to state-of-the-art methods that are either not pose-free or lack generalizability. In the table, the best result is highlighted in bold." + ], + "table_footnote": [], + "table_body": "
Re10K\n(Source Domain)
ScanNet
DL3DV
MVImgNet
Ground TruthStreamGS (Ours)MVSplatpixelSplatCF-3DGS
", + "bbox": [ + 91, + 239, + 903, + 502 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fae2db546b11249465676c712401ba999da91b55c21b3ad23257d30b0228cbda.jpg", + "table_caption": [ + "Figure 3. Qualitative comparison on novel view synthesis. We show the results on both source domain, RE10K [38], and other domains, ScanNet [5], DL3DV [16] and MVImgNet [33]. All generalizable methods are trained only on RE10K and tested on the other datasets. StreamGS outperforms other methods in several challenging scenarios, especially for the out-of-domain data." + ], + "table_footnote": [ + "Table 3. Efficiency metrics of each component." + ], + "table_body": "
Coarse Pred (Sec. 3.2)Ada. Refine (Sec. 3.3)ADC (Sec. 3.4)Total
time/sparam/Mtime/sparam/Mtime/sparam/Mtime/sparam/M
StreamGS0.02656.740.081.830.010.040.11658.61
", + "bbox": [ + 99, + 553, + 473, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ficiency, measured by the processing time per frame (FPS). The figure shows that StreamGS achieves second place with a PSNR of 23.1, only 0.05 lower than CF-3DGS [9]. This indicates that our method achieves nearly the same rendering quality as the best model. However, thanks to the feed-forward design, StreamGS is 150 times faster than CF-3DGS, predicting Gaussian primitives for up to 9 frames within one second. Without known camera information, our method involves additional alignment and pose estimation processes across frames, which limits the inference speed, making it slower than MVSplat [3] and pixelSplat [2]. However, according to the scores reported in Tab. 2 and Fig. 5, our method is more generalizable and achieves better rendering quality than these methods. Tab. 3 shows the efficiency of each component of our method.", + "bbox": [ + 89, + 611, + 482, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 848, + 243, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we discuss the effectiveness of our main design about joint refinement (in Sec. 3.3) and feed-forward", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ADC module (in Sec 3.4). More ablation studies on framework design can be found in supplementary materials.", + "bbox": [ + 511, + 556, + 906, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.1. Effectiveness of Joint Refinement", + "text_level": 1, + "bbox": [ + 513, + 592, + 790, + 606 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7eac9d3a6ab6cfc4fe08b880ff0d86300a5269db0d3d096497df45bc28c0694a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
RE10K [38]ACID [17]
PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑
w/o refine.17.250.320.5817.230.390.48
w/ refine22.420.170.8328.500.150.84
", + "bbox": [ + 527, + 619, + 895, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Evaluation of effectiveness of joint refinement.", + "bbox": [ + 542, + 704, + 874, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Joint refinement of cameras and centers of Gaussians plays a crucial role in the success of our method. To validate the effectiveness of joint refinement, we conduct an ablation study by skipping the refinement process during inference. In other words, the poses and intrinsics are directly estimated by Eq. 2 and 3. Tab. 4 shows the quantitative comparison between the two settings. Without joint refinement, Gaussian primitives are cast from shifted origins of the camera with erroneous orientations, and the poses of novel views are also inaccurate, causing the rendered images to be shifted and distorted. This severely deteriorates the rendering quality, as shown in Tab. 4. The PSNR", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25847", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/14d46c0f9362784cc8aa6c6afe6e215f9542d48572e4c64aa39a760066dbf1d3.jpg", + "image_caption": [ + "Figure 4. Visual comparison of Gaussian reconstruction and novel view synthesis from image streams with ScanNet [5] dataset. Unlike MVSplat [3], which struggles with view aggregation, our results show significantly better visual quality on OOD data. Note that Our StreamGS and MVSplat are both trained with RE10K [38] data, and MVSplat needs predetermined cameras." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 903, + 333 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/73a111e40a52f66e6e9fff9424d2b99821b4549bc2fed74909574c80fc369c22.jpg", + "image_caption": [ + "Figure 5. Reconstruction speed measured by frames processed per second (FPS). The x-axis is log-scaled for the better visualization. of the images rendered with direct estimation decreases by $23.06\\%$ and $38.62\\%$ on the RE10K [38] and ACID [17] datasets, respectively." + ], + "image_footnote": [], + "bbox": [ + 94, + 400, + 477, + 564 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2.2. Effectiveness of Gaussian Merging Process", + "text_level": 1, + "bbox": [ + 89, + 647, + 433, + 664 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/db764eeea26a1c64284b55c6c8460f4015c9743fbe43ad767aa5cc3cbdaf6646.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MVImgNet [33]ACID [17]
Compress. Ratio ↑PSNR ↑Compress. Ratio ↑PSNR ↑
w/o merge.1.0025.701.0029.21
merge all1.5825.051.6828.07
", + "bbox": [ + 104, + 676, + 472, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Evaluation of the efficiency improvements of the Gaussian merging process and its impact on rendering quality.", + "bbox": [ + 89, + 763, + 482, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "During the feed-forward ADC, StreamGS prunes pixel-aligned Gaussians through a merging process. To evaluate the memory efficiency improvement and its impact on rendering quality, we compare the average number of Gaussians per frame and PSNR before and after the merging process. We define the compression ratio of Gaussians during the merging process as the ratio of the average number of", + "bbox": [ + 89, + 795, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Gaussians per frame, i.e., $H \\times W$ , to that after the merging process. The metrics are reported in Tab. 5. The results demonstrate that the merging process can prune Gaussians per frame by $36.71\\%$ and $40.48\\%$ on MVImgNet [33] and ACID [17], respectively. Meanwhile, the PSNR scores after the merging process only slightly decrease by $2.53\\%$ and $3.90\\%$ , respectively. This ablation study demonstrates that the designed Gaussian merging process efficiently reduces memory usage during reconstruction and rendering, with a negligible impact on reconstruction quality.", + "bbox": [ + 511, + 390, + 906, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 553, + 633, + 569 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a novel and holistic generalizable pose-free reconstruction pipeline named StreamGS, dedicated to the online reconstruction of endless unposed image streams, such as monocular videos. To the best of our knowledge, our method is the first generalizable model capable of predicting Gaussians corresponding to the input stream in a feedforward manner, without relying on known poses and intrinsics. Compared to pose-free but optimization-based methods, our method achieves comparable reconstruction quality while reducing the learning time to within several milliseconds, avoiding optimization steps. Compared to other generalizable methods, StreamGS eliminates the dependence on poses and intrinsics and manages to reconstruct more accurate scenes on out-of-domain datasets, demonstrating better domain generalizability.", + "bbox": [ + 511, + 580, + 906, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Although our method runs fast, joint refinement process still includes additional time costs, making it slower than MVSplat [3]. Moreover, our approach encounters common reconstruction challenges, including textureless regions and long sequences.", + "bbox": [ + 511, + 825, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25848", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Wenjing Bian, Zirui Wang, Kejie Li, Jia-Wang Bian, and Victor Adrian Prisacariu. Nope-nerf: Optimising neural radiance field with no pose prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4160-4169, 2023. 2, 3", + "[2] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 2, 3, 5, 6, 7", + "[3] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. arXiv preprint arXiv:2403.14627, 2024. 2, 3, 5, 6, 7, 8", + "[4] Zezhou Cheng, Carlos Esteves, Varun Jampani, Abhishek Kar, Subhransu Maji, and Ameesh Makadia. Lu-nerf: Scene and pose estimation by synchronizing local unposed nerfs. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18312–18321, 2023. 3", + "[5] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 6, 7, 8", + "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6", + "[7] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.20309, 2024. 2", + "[8] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 5", + "[9] Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A. Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20796-20805, 2024. 2, 3, 6, 7", + "[10] Yiming Gao, Yan-Pei Cao, and Ying Shan. Surfelnerf: Neural surfel radiance fields for online photorealistic reconstruction of indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 108-118, 2023. 1, 3", + "[11] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 2", + "[12] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 3, 5, 6" + ], + "bbox": [ + 93, + 114, + 485, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[14] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Ep n p: An accurate o (n) solution to the p np problem. International journal of computer vision, 81:155-166, 2009. 5", + "[15] Vincent Leroy, Yohann Cabon, and Jerome Revaud. Grounding image matching in 3d with mast3r, 2024. 2", + "[16] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. D13dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024. 6, 7", + "[17] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In ICCV, 2021. 6, 7, 8", + "[18] Tianqi Liu, Guangcong Wang, Shoukang Hu, Liao Shen, Xinyi Ye, Yuhang Zang, Zhiguo Cao, Wei Li, and Ziwei Liu. Mvsgaussian: Fast generalizable gaussian splattering reconstruction from multi-view stereo. In European Conference on Computer Vision, pages 37-53. Springer, 2025. 3", + "[19] Andreas Meuleman, Yu-Lun Liu, Chen Gao, Jia-Bin Huang, Changil Kim, Min H Kim, and Johannes Kopf. Progressively optimized local radiance fields for robust view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16539-16548, 2023. 3", + "[20] Jongmin Park, Minh-Quan Viet Bui, Juan Luis Gonzalez Bello, Jaeho Moon, Jihyong Oh, and Munchurl Kim. Splinegs: Robust motion-adaptive spline for real-time dynamic 3d gaussians from monocular video. arXiv preprint arXiv:2412.09982, 2024. 2", + "[21] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 1, 3", + "[22] Qiuhong Shen, Zike Wu, Xuanyu Yi, Pan Zhou, Hanwang Zhang, Shuicheng Yan, and Xinchao Wang. Gamba: Marry gaussian splatting with mamba for single view 3d reconstruction. arXiv preprint arXiv:2403.18795, 2024. 3", + "[23] Liang Song, Guangming Wang, Jiuming Liu, Zhenyang Fu, Yanzi Miao, et al. Sc-nerf: Self-correcting neural radiance field with sparse views. arXiv preprint arXiv:2309.05028, 2023. 2", + "[24] Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10208-10217, 2024. 2, 5", + "[25] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In European Conference on Computer Vision, pages 1-18. Springer, 2025. 3, 5", + "[26] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25849", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 1, 2, 4", + "[28] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6", + "[29] Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. Grm: Large gaussian reconstruction model for efficient 3d reconstruction and generation. arXiv preprint arXiv:2403.14621, 2024. 3", + "[30] Chi Yan, Delin Qu, Dan Xu, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Gs-slam: Dense visual slam with 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19595-19604, 2024. 1, 2, 3", + "[31] Xuanyu Yi, Zike Wu, Qiuhong Shen, Qingshan Xu, Pan Zhou, Joo-Hwee Lim, Shuicheng Yan, Xinchao Wang, and Hanwang Zhang. Mvgamba: Unify 3d content generation as state space sequence modeling. arXiv preprint arXiv:2406.06367, 2024. 3", + "[32] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4578-4587, 2021. 3", + "[33] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Tianyou Liang, Guanying Chen, Shuguang Cui, and Xiaoguang Han. Mvimgnet: A large-scale dataset of multi-view images. In CVPR, 2023. 6, 7, 8", + "[34] Vladimir Yugay, Yue Li, Theo Gevers, and Martin R Oswald. Gaussian-slam: Photo-realistic dense slam with gaussian splatting. arXiv preprint arXiv:2312.10070, 2023. 2, 3", + "[35] Haojun Zhang, Yuan Yao, and Xuefeng Yan. Improved end-to-end multilevel nerf-based dense rgb-d slam. In Chinese Conference on Pattern Recognition and Computer Vision (PRCV), pages 132-146. Springer, 2024. 2, 3", + "[36] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In European Conference on Computer Vision, pages 1-19. Springer, 2025. 3", + "[37] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6", + "[38] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. ACM Trans. Graph., 37, 2018. 6, 7, 8", + "[39] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12786-12796, 2022. 3" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Zihan Zhu, Songyou Peng, Viktor Larsson, Zhaopeng Cui, Martin R Oswald, Andreas Geiger, and Marc Pollefeys. Nicer-slam: Neural implicit scene encoding for rgb slam. In 2024 International Conference on 3D Vision (3DV), pages 42-52. IEEE, 2024. 2, 3", + "[41] Zi-Xin Zou, Shi-Sheng Huang, Yan-Pei Cao, Tai-Jiang Mu, Ying Shan, and Hongbo Fu. Mononeuralfusion: Online monocular neural 3d reconstruction with geometric priors. arXiv preprint arXiv:2209.15153, 2022. 2, 3", + "[42] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10324-10335, 2024. 3" + ], + "bbox": [ + 516, + 92, + 906, + 301 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25850", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/StreamGS_ Online Generalizable Gaussian Splatting Reconstruction for Unposed Image Streams/9d1940fe-237f-4495-9085-9441928f0c47_model.json b/2025/StreamGS_ Online Generalizable Gaussian Splatting Reconstruction for Unposed Image Streams/9d1940fe-237f-4495-9085-9441928f0c47_model.json new file mode 100644 index 0000000000000000000000000000000000000000..869b4e006a788a6aea477be9bf5a41d3ab5a999b --- /dev/null +++ b/2025/StreamGS_ Online Generalizable Gaussian Splatting Reconstruction for Unposed Image Streams/9d1940fe-237f-4495-9085-9441928f0c47_model.json @@ -0,0 +1,1881 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.154, + 0.13, + 0.844, + 0.177 + ], + "angle": 0, + "content": "StreamGS: Online Generalizable Gaussian Splitting Reconstruction for Unposed Image Streams" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.202, + 0.816, + 0.294 + ], + "angle": 0, + "content": "Yang Li\\(^{1,4,*}\\) Jinglu Wang\\(^{1}\\) Lei Chu\\(^{1}\\) Xiao Li\\(^{1}\\) Shiu-Hong Kao\\(^{1,2,*}\\) Ying-Cong Chen\\(^{2,3}\\) Yan Lu\\(^{1}\\) \n\\(^{1}\\)Microsoft Research Asia \\({}^{2}\\)HKUST \\({}^{3}\\)HKUST(GZ) \\({}^{4}\\)Rutgers University yangliaftermath@gmail.com skao@cse.ust.hk yingcongchen@ust.hk {jinglu.wang, leichu, li.xiao, yanlu}@microsoft.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.361, + 0.485, + 0.755 + ], + "angle": 0, + "content": "The advent of 3D Gaussian Splatting (3DGS) has advanced 3D scene reconstruction and novel view synthesis. With the growing interest of interactive applications that need immediate feedback, online 3DGS reconstruction in real-time is in high demand. However, none of existing methods yet meet the demand due to three main challenges: the absence of predetermined camera parameters, the need for generalizable 3DGS optimization, and the necessity of reducing redundancy. We propose StreamGS, an online generalizable 3DGS reconstruction method for unposed image streams, which progressively transform image streams to 3D Gaussian streams by predicting and aggregating per-frame Gaussians. Our method overcomes the limitation of the initial point reconstruction [27] in tackling out-of-domain (OOD) issues by introducing a content adaptive refinement. The refinement enhances cross-frame consistency by establishing reliable pixel correspondences between adjacent frames. Such correspondences further aid in merging redundant Gaussians through cross-frame feature aggregation. The density of Gaussians is thereby reduced, empowering online reconstruction by significantly lowering computational and memory costs. Extensive experiments on diverse datasets have demonstrated that StreamGS achieves quality on par with optimization-based approaches but does so 150 times faster, and exhibits superior generalizability in handling OOD scenes." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.787, + 0.222, + 0.803 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.813, + 0.484, + 0.874 + ], + "angle": 0, + "content": "The field of 3D Scene reconstruction [10, 30] for novel view synthesis from image streams has gained increasing attention, due to its significance in enabling interactive applications that offer users instant feedback. In this context, the" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.328, + 0.898, + 0.673 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.684, + 0.908, + 0.753 + ], + "angle": 0, + "content": "Figure 1. The proposed StreamGS efficiently transforms image streams into Gaussian streams by progressively reconstructing and aggregating per-frame 3D Gaussians. We show our reconstructed 3DGS (visualized as points) alongside estimated camera poses (in blue), and synthesized novel views." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.841 + ], + "angle": 0, + "content": "advent of 3D Gaussian Splatting (3DGS) [12] marks a major advancement in high-quality, real-time rendering. This progress highlights the urgency for efficient, on-the-fly generation of 3DGS from image streams." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Nonetheless, online 3DGS reconstruction faces unique challenges. (1) Unknown camera. The conventional preprocessing using Structure from Motion (SfM) [21] for camera estimation is impractical for real-time streaming." + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.352, + 0.901 + ], + "angle": 0, + "content": "*Work done during the internship at MSRA." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25841" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.258 + ], + "angle": 0, + "content": "This is due to the absence of the full image set and the time-consuming computation. Thus, there is a need for methods that can operate without pre-determined cameras. (2) Generalizability. 3DGS reconstruction requires multiteration optimization, impractical for online applications due to the need for all images in advance. This restricts the development of a generalizable method that can process image streams in a feed-forward manner. (3) Redundancy. The significant overlap between frames leads to high redundancy when reconstructing Gaussians individually for each frame, increasing the streaming process's resource intensity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.261, + 0.482, + 0.396 + ], + "angle": 0, + "content": "Scene reconstruction without known cameras has been explored in SLAM-based [30, 34, 35, 40, 41] and NeRF-based [1, 23] methods. Only a few 3DGS-based methods are discussed [9, 20]. These methods primarily consider camera poses as learnable parameters and optimized alongside Gaussians through the iterative optimization. Yet, this optimization-driven approach for aligning each frame considerably increases the reconstruction time, rendering it impractical in the online scenario." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.4, + 0.483, + 0.7 + ], + "angle": 0, + "content": "Recently, generalizable 3DGS reconstruction has been investigated for sparse views [2, 3, 24]. These approaches transform pixels into Gaussians, whose parameters are decoded from images via 2D encoder-decoder networks. They achieve the 3DGS reconstruction for each image in a single feed-forward pass. However, these generalizable approaches are primarily designed for monocular or binocular settings, suited to sparse-view inputs. In addition, multi-view methods [2, 3] typically infer Gaussian centers through stereo matching, highly dependent on known cameras, limiting their applicability for online scenarios with numerous unposed images. Furthermore, they combine multi-view 3DGS sets by simply uniting them, which overlooks cross-view alignment and overlaps, leading to misalignment and redundancy issues within image streams. The recent method [7] performs Gaussian downsampling in 3D space to reduce redundancy, but it necessitates traversing and processing Gaussians in 3D grids, which is also time-consuming and is highly dependent on the grid resolution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Emerging models like DUSt3R [27] and MASt3R [15] enable sparse-view geometry reconstruction by simultaneously predicting 3D points and estimating camera parameters. These advances pave the way for more efficient, feedforward, pose-free 3D reconstruction methods. A straightforward approach to generalize 3DGS reconstruction is to add a Gaussian predictor to the DUSt3R-like framework. However, this introduces several challenges. These models usually require datasets with ground truth 3D geometry, which may not always be available. Additionally, applying pretrained models to out-of-domain (OOD) data can result in inaccurate pose and 3D point estimations. Moreover, generating 3D points for each frame individually causes re" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "dundancy due to overlapping adjacent frames, potentially leading to ghosting artifacts from pose estimation errors." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.123, + 0.905, + 0.483 + ], + "angle": 0, + "content": "In this paper, we introduce StreamGS, a novel pipeline for online, generalizable 3DGS reconstruction from unposed image streams. StreamGS aims to progressively construct and update the 3DGS representation of the scene frame-by-frame, in a feed-forward manner, as illustrated in Fig. 1. We leverage the pretrained DUSt3R to initially predict 3D point for the current frame using the previous frame as a reference. However, this initialization may encounter inaccuracies due to OOD issues. To mitigate this, we capitalize the insight that adjacent frames offer sufficient correspondences to refine the reconstruction. Unlike DUSt3R using predicted 3D points to establish correspondences, we adopt content-adaptive descriptors for more reliable matching, allowing for the adaptive refinement of the reconstruction by enhancing consistency between adjacent views. Furthermore, such correspondences help to prune redundant pixel-aligned Gaussians. Correlated pixel-wise features across frames are effectively aggregated, removing duplicates and achieving adaptive density control. Finally, we decode Gaussians from such aggregated features. StreamGS is adept at predicting and integrating Gaussians for the current frame into the existing Gaussian set seamlessly with a feed-forward pass. In summary, our contribution is three-fold." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.484, + 0.904, + 0.529 + ], + "angle": 0, + "content": "- We introduce a novel pipeline for the online, generalizable reconstruction of image streams without requiring camera parameters, marking a first in this field." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.53, + 0.905, + 0.604 + ], + "angle": 0, + "content": "- The proposed adaptive refinement enhance cross-frame consistency of 3DGS reconstruction, and the adaptive density control mechanism minimizes adjacent-view redundancy, thereby highly reducing computational costs in online reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.606, + 0.905, + 0.71 + ], + "angle": 0, + "content": "- Upon evaluation across diverse datasets, our method achieves high novel view synthesis quality comparable to the optimization-based method [9] but with \\(150\\mathrm{x}\\) faster reconstruction speed. Additionally, our method outperforms existing pose-dependent generalizable 3DGS methods in handling OOD scenes, showing superior generalizability." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.484, + 0.905, + 0.71 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.724, + 0.665, + 0.74 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Generalizable 3D Gaussian Splitting. Many recent studies aim to propose generalizable 3D-GS methods capable of predicting Gaussians within a single feed-forward pass. These works can be classified into two main categories: single-view reconstruction and multi-view reconstruction. Single-view reconstruction does not involve pose estimation as there are no multi-view constraints. Inspired by the insight from LRM [11] that large transformer-based [26] backbone networks can learn 3D priors from large-scale 3D data, the potential of predicting Gaussians from a" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "25842" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "single image in a single feed-forward pass has been comprehensively explored. Numerous feed-forward models have been proposed, such as GRM [29], TriplaneGS [42], and GMamba [22]. However, these methods are primarily not applicable to multi-view scenarios as they always assume canonical poses." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.184, + 0.487, + 0.515 + ], + "angle": 0, + "content": "Concurrently, many works focusing on multi-view inputs follow a similar paradigm, such as GS-LRM [36], LGM [25], and MVGMamba [31]. They concatenate the input images with camera embeddings like Plücker rays to facilitate the network in learning the proper fusion of Gaussians from multi-view inputs. However, these large 3D backbone networks mostly perform well only on synthetic objects due to the shortage of large-scale scene-level 3D data in the real world. Referring to generalizable NeRF methods like pixelNeRF [32], other research turns to multi-view stereo (MVS) matching to locate or initialize the centers of Gaussians, with other attributes decoded using a lightweight 2D encoder. Representative works with this design include MVSGaussian [18], pixelSplat [2], and MVSplat [3]. However, both camera ray embedding in large transformer-based models and stereo matching rely on known poses and intrinsicities of each input view. Another main limitation of these methods is that they focus on sparse-view inputs. PixelSplat [2] and MVSplat [3] only support up to three views. Therefore, existing generalizable 3D-GS models cannot address the problem of feed-forward reconstruction from endless image streams." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.534, + 0.485, + 0.822 + ], + "angle": 0, + "content": "Pose-free 3D Gaussian Splatting. Recently, many works have aimed to eliminate the need for Structure-from-Motion (SfM) preprocessing steps using COLMAP [21] software. Following the design of previous pose-free NeRF methods like NoPe-NeRF [1], Lu-NeRF [4], and localRF [19], CF-3DGS [9] introduces depth priors into the optimization of 3D-GS and performs progressive reconstruction. As each new image arrives, CF-3DGS optimizes both the pose and 3D Gaussians of the input image based on its depths, aiming to align the 3DGS from the new image with the preceding reconstruction. However, it still relies on known camera intrinsics. CF-3DGS is also not robust, as the accuracy of depth priors significantly impacts its reconstruction quality, limiting its application to common scenes. Moreover, it depends on thousands of optimization steps for each view, significantly extending the reconstruction time for each scene. Compared to generalizable 3D-GS models, current pose-free methods are so inefficient that they cannot be applied to image streams with a large number of frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Online 3D reconstruction of image streams. Online 3D reconstruction has been extensively studied in the field of SLAM [30, 34, 35, 39-41]. However, these methods typically involve additional information and most leverage SDF" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.228 + ], + "angle": 0, + "content": "and NeRF as scene representations. NICE-SLAM [39] uses RGB-D streams as input, with the reconstructed scene represented by NeRF. NICER-SLAM [40] relies on geometric priors, including surface normals and depths. Surfel-NeRF [10] focuses on the novel view synthesis quality of online reconstruction with RGB streams, but it requires the poses and intrinsics of frames. Gaussian-SLAM [34] reconstructs the scene using 3D-GS, but it also relies on RGB-D streams." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.24, + 0.613, + 0.255 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.266, + 0.907, + 0.356 + ], + "angle": 0, + "content": "Given a sequence of unposed images over time, our objective is to progressively reconstruct the 3D Gaussian Splating (3DGS) representation in an online manner. Specifically, at each timestamp \\( t \\), the goal is to derive the 3DGS \\( \\mathcal{G}^t \\), which encapsulates the 3D scene aggregated from the images \\( \\mathcal{I}^t = \\{\\mathbf{I}^i\\}_{i=1}^t \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.909, + 0.599 + ], + "angle": 0, + "content": "Fig. 2 illustrates our overall framework. In order to make reconstruction efficient with limited computational resources, we employ an incremental construction strategy. At each time step \\( t \\), we focus on generating the 3DGS \\( \\mathbf{G}^t \\) of the current frame and merge it with the previous accumulated reconstruction \\( \\mathcal{G}^{t-1} \\) to obtain the full reconstruction \\( \\mathcal{G}^t \\) of the current time step. Specifically, with current frame \\( \\mathbf{I}^t \\), we use \\( \\mathbf{I}^{t-1} \\) as reference frame and estimate the point maps and cameras of each using an initial reconstruction module. With newly established matches between current and reference views, we further refine the quality of the points and the cameras in the adaptive refinement module. Finally, we generate \\( \\mathbf{G}^t \\) using the refined points and features of \\( \\mathbf{I}^t \\), and merge it with previous reconstructions according to the established matches, achieving the feedforward adaptive density control (ADC)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.606, + 0.655, + 0.62 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.909, + 0.766 + ], + "angle": 0, + "content": "Gaussian splatting. Previous work [12] represents a scene or object using a set of Gaussian distributions. Specifically each gaussian primitive could be denoted as \\( G(x;\\mu ,\\Sigma) = e^{-\\frac{1}{2} (x - \\mu)^T\\Sigma^{-1}(x - \\mu)} \\) and the covariance \\( \\Sigma \\) is decomposed into the rotation matrix \\( R \\) and scaling matrix \\( S \\) to ensure the positive semi-definiteness during optimization, that is \\( \\Sigma = RSS^T R^T \\). The view-dependent color of the appearance is represented by a set of spherical harmonics (SH) coefficients and opacity value \\( \\alpha \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.774, + 0.798, + 0.789 + ], + "angle": 0, + "content": "3.2. Initial Two-view Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Given the current frame \\(\\mathbf{I}^t\\) and a reference frame \\(\\mathbf{I}^{t'}\\), we use a coarse predictor \\(\\phi_{3D}\\) and estimate the point map \\(\\mathbf{X}^{t|t'}\\) of the current frame under the local coordinate system of the reference frame together with its corresponding confidence map \\(\\mathbf{C}^{t|t'}\\), where the superscript \\(\\cdot |t'\\) indicates that the local coordinate system adapts to that of \\(\\mathbf{I}^{t'}\\). Intuitively, the predicted point map stores the 3D point coordinate that the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25843" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.905, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.337, + 0.908, + 0.409 + ], + "angle": 0, + "content": "Figure 2. Method overview. Our StreamGS progressively reconstruct and aggregate 3D Gaussians from the unposed image stream. Given the adjacent image pair \\((\\mathbf{I}^{t - 1},\\mathbf{I}^t)\\), we first perform the initial reconstruction that predicts pixel-wise 3D points with their features and coarse camera poses, using a pretrained coarse predictor. Since the coarse predictions may suffer from OOD issues, we refine both the camera poses and 3D positions by establishing new point-wise correspondences. We aggregate cross-frame image and 3D features by warping and merging to reduce redundancy. Finally we decode the aggregated features to Gaussian primitives." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.484, + 0.476 + ], + "angle": 0, + "content": "pixel unprojected to in, and the confidence maps measure the certainty of point maps at each pixel, reflecting a prior-based estimation of reconstruction accuracy and difficulty. Formally, we define" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.487, + 0.483, + 0.508 + ], + "angle": 0, + "content": "\\[\n\\left(\\mathbf {X} ^ {t \\mid t ^ {\\prime}}, \\mathbf {X} ^ {t ^ {\\prime} \\mid t ^ {\\prime}}, \\mathbf {C} ^ {t \\mid t ^ {\\prime}}, \\mathbf {C} ^ {t ^ {\\prime} \\mid t ^ {\\prime}}\\right) = \\phi_ {3 D} \\left(\\mathbf {I} ^ {t}, \\mathbf {I} ^ {t ^ {\\prime}}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.512, + 0.484, + 0.65 + ], + "angle": 0, + "content": "In order to help align the current local-coordinate pointmap with the global coordinate, we also get the output in the local coordinate system of \\(\\mathbf{I}^t\\) by computing \\(\\phi_{3D}(\\mathbf{I}^{t'},\\mathbf{I}^t)\\). With these predicted point maps, we could further estimate the camera matrix \\(\\mathbf{P} = \\mathbf{K}[\\mathbf{R}|\\mathbf{t}]\\), composed of its intrinsic parameters \\(\\mathbf{K}\\), and its extrinsic parameters, the rotation matrix \\(\\mathbf{R}\\) and the translation vector \\(\\mathbf{t}\\). Specifically, we assume that the principle points \\(\\mathbf{c}\\) are centered and pixels are squares and have" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.659, + 0.483, + 0.692 + ], + "angle": 0, + "content": "\\[\n\\hat {f} ^ {t} = \\arg \\min _ {f} \\Sigma_ {\\mathbf {p} \\in \\mathbf {I} ^ {t}} \\| \\mathbf {p} - \\mathbf {c} - f \\frac {\\left(\\mathbf {x} _ {\\mathbf {p}} , \\mathbf {y} _ {\\mathbf {p}}\\right)}{\\mathbf {z} _ {\\mathbf {p}}} \\|, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.702, + 0.484, + 0.764 + ], + "angle": 0, + "content": "where \\((\\mathbf{x_p},\\mathbf{y_p},\\mathbf{z_p})\\in \\mathbf{X}^{t|t}\\) is the 3D point coordinate that the pixel \\(\\mathbf{p}\\) unprojected to. Moreover, we approximate relative pose \\(\\mathbf{P}^t = [\\mathbf{R},\\mathbf{t}]\\) of \\(\\mathbf{I}^t\\) to \\(\\mathbf{I}^{t'}\\) by solving the following points registration problem:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.772, + 0.481, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\left[ \\mathbf {R} ^ {t} \\mid \\mathbf {t} ^ {t} \\right] = \\arg \\min _ {s, \\mathbf {R}, \\mathbf {t}} \\sum_ {\\mathbf {p} \\in \\mathbf {I} ^ {t - 1}} \\mathbf {C} ^ {t} (\\mathbf {p}) \\| s \\left(\\mathbf {R} \\mathbf {X} ^ {t ^ {\\prime} | t ^ {\\prime}} (\\mathbf {p}) + \\mathbf {t}\\right) - \\mathbf {X} ^ {t ^ {\\prime} | t} (\\mathbf {p}) \\| ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.801, + 0.248, + 0.816 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} ^ {t} = \\mathbf {C} ^ {t ^ {\\prime} | t ^ {\\prime}} \\odot \\mathbf {C} ^ {t ^ {\\prime} | t},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.478, + 0.84 + ], + "angle": 0, + "content": "where \\(\\odot\\) is the Hadamard product and \\(s\\) is the scale factor." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In our implementation, we leverage DUSt3R [27] as the coarse predictor due to its effective pretraining on dedicated 3D scene datasets and its efficiency in reconstructing 3D points. For each time step \\( t \\), we process the pair" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.413, + 0.907, + 0.581 + ], + "angle": 0, + "content": "\\((\\mathbf{I}^{t - 1},\\mathbf{I}^t)\\) to derive the 3D points \\((\\mathbf{X}^{t|t - 1},\\mathbf{X}^{t - 1|t - 1})\\) corresponding to both frames referenced in the coordinate frame \\(t - 1\\), the reversed pair \\((\\mathbf{I}^t,\\mathbf{I}^{t - 1})\\) to obtain the same points \\((\\mathbf{X}^{t - 1|t},\\mathbf{X}^{t|t})\\) but in the coordinate frame of \\(t\\), which will be reused at the next timestamp to boost efficiency. For simplicity, and without loss of generality, the following discussion will focus on the reconstruction of the image pair \\((\\mathbf{I}^{t - 1},\\mathbf{I}^t)\\). Note that training \\(\\phi_{3D}\\) requires 3D geometric supervision, which may not be available in our monocular video input scenario. Thus, we leverage the pretrained \\(\\phi_{3D}\\) from [27]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.591, + 0.714, + 0.607 + ], + "angle": 0, + "content": "3.3. Adaptive Refinement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.906, + 0.688 + ], + "angle": 0, + "content": "We note that the initial reconstruction quality is compromised due to the OOD challenge as the coarse predictor is frozen. This observation motivates us to enhance reconstruction through content adaptation, with the goal of adaptively refining both poses and 3D reconstructions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.906, + 0.794 + ], + "angle": 0, + "content": "We perform the adaptive refinement based on establishing new robust matches between adjacent frames. We employ a matching head, \\(\\phi_{\\text{match}}\\), to extract local 3D features, denoted as \\((\\mathbf{F}_{3D}^{t-1}, \\mathbf{F}_{3D}^t) \\in \\mathbb{R}^{H \\times W \\times d}\\), from the consecutive image pair \\((\\mathbf{I}^{t-1}, \\mathbf{I}^t)\\). The matches between the two images can be established through nearest reciprocal (NN) searching, satisfying the following condition:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.804, + 0.905, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {M} ^ {t - 1, t} = \\left\\{i _ {k} \\leftrightarrow j _ {k} \\mid i _ {k} = \\mathrm {N N} (j _ {k}) \\text {a n d} j _ {k} = \\mathrm {N N} (i _ {k}) \\right\\} _ {k = 1} ^ {N}, \\\\ \\text {s . t .} \\operatorname {N N} \\left(i _ {k}\\right) = \\underset {0 \\leq j _ {k} \\leq H \\times W} {\\arg \\min } | 1 - \\cos < \\mathbf {F} _ {3 D, i _ {k}} ^ {t - 1}, \\mathbf {F} _ {3 D, j _ {k}} ^ {t} > |, \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.855, + 0.906, + 0.901 + ], + "angle": 0, + "content": "where \\(i_k, j_k\\) are pixel index in image \\(\\mathbf{I}^{t-1}\\), \\(\\mathbf{I}^t\\) respectively, and the measurement of feature distance is cosine similarity. With the correspondences found, a residual transform" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "25844" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.182 + ], + "angle": 0, + "content": "\\(\\Delta = [\\Delta \\mathbf{R}, \\Delta \\mathbf{t}]\\) can be re-estimated following Eq. (3) by only taking matched points into account. Then we apply the residual transform to the pointmap \\(\\mathbf{X}^{t,t-1}\\) to retain refined \\(\\tilde{\\mathbf{X}}^{t|t-1}\\), and the pose of \\(\\mathbf{I}^t\\) is updated to \\(\\tilde{\\mathbf{P}}^t\\) by performing PnP-RANSAC [8, 14] on 3D-2D correspondences derived from matches." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.21, + 0.485, + 0.256 + ], + "angle": 0, + "content": "Gaussian decoding. We directly predict the other parameters of 3D Gaussians at each pixel with a light-weight decoder \\(\\phi_{GS}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.269, + 0.484, + 0.294 + ], + "angle": 0, + "content": "\\[\n\\mathbf {G} ^ {i} = \\left[ \\mathbf {q} ^ {i}, \\mathbf {s} ^ {i}, \\alpha^ {i}, \\mathbf {c} ^ {i} \\right] = \\phi_ {G S} (\\mathbf {F} _ {g s}), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.29, + 0.442, + 0.307 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {G S} ^ {i} = \\mathbf {F} _ {2 D} ^ {i} \\oplus \\mathbf {X} ^ {i} \\oplus \\mathbf {F} _ {3 D} ^ {i}, \\quad \\mathbf {F} _ {2 D} = \\phi_ {2 D} (\\mathbf {I} ^ {i}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.485, + 0.519 + ], + "angle": 0, + "content": "where \\(i = \\{t - 1,t\\}\\) \\(\\oplus\\) denotes the channel-wise concatenation, \\(\\phi_{2D}\\) denotes the 2D image feature extractor, \\(\\mathbf{q}^i\\in \\mathbb{R}^{H\\times W\\times 4}\\) and \\(\\mathbf{s}^i\\in \\mathbb{R}^{H\\times W\\times 3}\\) represent rotation quaternions and scales of pixel-aligned Gaussians at image \\(\\mathbf{I}^i\\) . We incorporate an additional image feature extractor because the coarse predictor is frozen and cannot be trained by our monocular video setting. Extracting new image features is essential for decoding Gaussians, especially for texture-related properties. Experiments show the importance of the image feature extractor. Then covariance matrix is built with \\(\\Sigma^i = \\mathbf{R}(\\mathbf{q}^i)\\mathbf{ss}^{\\mathrm{T}}\\mathbf{R}(\\mathbf{q}^i)^{\\mathrm{T}}\\) Note that \\(\\bar{\\mathbf{G}}\\) is not the final Gaussians since it needs to be merged into the previous Gaussian set following Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.533, + 0.283, + 0.547 + ], + "angle": 0, + "content": "3.4. Feed-Forward ADC" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.556, + 0.484, + 0.784 + ], + "angle": 0, + "content": "With pixel-aligned Gaussian parameters \\(\\mathbf{G}^t\\) of \\(T\\) images, previous methods [2, 3, 24, 25] always naively take the union of Gaussians in all images as the final prediction, i.e., \\(\\mathcal{G} = \\bigcup_{t=1}^{T} \\mathbf{G}^i\\) and \\(|\\mathcal{G}| = T \\times H \\times W\\). However, this approach is both memory-intensive and inefficient in rendering, particularly when dealing with the continuous input of video frames during online reconstruction. Our key observation is that the matched Gaussian pairs in neighboring frames are excessive and prunable since they consistently share similar attributes in shape and color, and are closely distributed, which is validated in Tab. 1. It is noted that we have already acquired dense pixel-wise matches between neighboring frames from Eq. (4). Therefore, we propose a novel feed-forward Adaptive Density Control strategy based on revisiting these dense correspondences." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.485, + 0.904 + ], + "angle": 0, + "content": "Feature aggregation. The primary advantage of pixel-based correspondences is that they are able to convert the computationally intensive 3D Gaussian aggregation process into a more efficient 2D pixel-wise one. This significantly enhances computational efficiency. Initially, the feature of Gaussian parameters \\(\\mathbf{F}_{gs}^{t}\\) of the frame \\(\\mathbf{I}^t\\) can be aligned to" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.091, + 0.875, + 0.108 + ], + "angle": 0, + "content": "the previous frame \\(\\mathbf{I}^{t - 1}\\) using the following wrapping:" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.113, + 0.905, + 0.155 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {G S} ^ {t \\mid t - 1} (j) = \\left\\{ \\begin{array}{l l} \\mathbf {F} _ {G S} ^ {t} (k) & \\text {i f} (j, k) \\in \\overline {{\\mathcal {M}}} ^ {t - 1, t} \\\\ \\mathbf {F} _ {G S} ^ {t - 1} (j) & \\text {e l s e ,} \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.162, + 0.907, + 0.332 + ], + "angle": 0, + "content": "where \\(\\mathbf{F}_{GS}^{t|t-1}\\) denotes the feature of Gaussian primitives in the frame \\(\\mathbf{I}^t\\) aligned to \\(\\mathbf{I}^{t-1}\\). Instead of using the raw correspondences set \\(\\mathcal{M}^{t-1,t}\\) in Eq. (4), we use the extended set \\(\\overline{\\mathcal{M}}^{t-1,t}\\), which includes matches of neighboring pixels such as \\((i+1,j+1)\\) in addition to the initial \\((i,j)\\). It serves as a type of anti-aliasing technique to reduce void pixels within the wrapped feature map. As for the unmatched pixels, we simply replicate the corresponding feature vector within \\(\\mathbf{F}_{gs}^{t-1}\\). With the aligned feature, the Gaussian feature of two frames can be merged by modifying Eq. (5), taking the form:" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.338, + 0.905, + 0.358 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {G}} ^ {t \\mid t - 1} = \\left[ \\hat {\\mathbf {q}} ^ {t}, \\hat {\\mathbf {s}} ^ {t}, \\hat {\\alpha} ^ {t}, \\hat {\\mathbf {c}} ^ {t} \\right] = \\phi_ {M G} \\left(\\mathbf {F} _ {G S} ^ {t \\mid t - 1} \\oplus \\mathbf {F} _ {G S} ^ {t - 1}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.365, + 0.907, + 0.469 + ], + "angle": 0, + "content": "where \\(\\phi_{MG}\\) denotes the MergeNet that simply consists of two convolutional layers, which merges features and decodes them to Gaussian primitives. In this way, every matched pair of Gaussians between \\(\\mathbf{I}^t\\) and \\(\\mathbf{I}^{t - 1}\\) is aggregated into a single one, highly reducing the number of Gaussians. The final aggregated Gaussian set is \\(\\mathcal{G}^t = \\mathcal{G}^{t - 1}\\cup \\mathbf{G}^{t|t - 1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.471, + 0.906, + 0.545 + ], + "angle": 0, + "content": "Without gradient back-propagation of rendering loss in the original 3DGS paper [12], our ADC process runs exceptional fast. In terms of the input group, the final prediction of Gaussian primitives \\(\\mathcal{G}^t\\) consists of the merged Gaussians and the unmatched Gaussians at each frame." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.553, + 0.667, + 0.568 + ], + "angle": 0, + "content": "3.5. Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.575, + 0.906, + 0.621 + ], + "angle": 0, + "content": "The optimization of the 2D feature extractor \\(\\phi_{2D}\\), the Gaussian decoder network, \\(\\phi_{GS}\\) involves both a rendering loss function and a reconstruction loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.625, + 0.905, + 0.658 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} \\left(\\mathbf {I} ^ {i}, \\hat {\\mathbf {I}} ^ {i}, \\hat {\\mathbf {I}} _ {M} ^ {i}\\right) = \\mathcal {L} _ {\\text {r e n d e r}} \\left(\\mathbf {I} ^ {i}, \\hat {\\mathbf {I}} ^ {i}\\right) + \\mathcal {L} _ {\\text {r e c o n}} \\left(\\hat {\\mathbf {I}} ^ {i}, \\hat {\\mathbf {I}} _ {M} ^ {i}\\right) \\tag {8} \\\\ = \\| \\mathbf {I} ^ {i} - \\hat {\\mathbf {I}} ^ {i} \\| _ {2} + \\lambda \\| \\mathbf {I} ^ {i} - \\hat {\\mathbf {I}} ^ {i} \\| _ {\\mathrm {L P I P S}} + \\| \\hat {\\mathbf {I}} _ {\\mathrm {M}} ^ {i} - \\hat {\\mathbf {I}} ^ {i} \\| _ {2}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.667, + 0.907, + 0.817 + ], + "angle": 0, + "content": "where \\(\\mathbf{I}^i\\) is the ground truth frame, \\(\\hat{\\mathbf{I}}^i\\) is the rendered image with full Gaussian primitives before the merge process, and \\(\\hat{\\mathbf{I}}_{\\mathrm{M}}^{i}\\) is the rendered image with merged Gaussians. Due to the lightweight nature of the two networks, the algorithm converges quickly within thousands of steps. The reconstruction loss term facilitates the merge network in fusing the pixel-aligned Gaussian parameters from different frames, aiming to render the same frame as before the merge process but with significantly fewer Gaussian primitives." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.831, + 0.646, + 0.847 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "StreamGS operates on an image stream of a scene, jointly predicting the corresponding poses and Gaussians in a feedforward manner. To assess its performance, we evaluate our" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25845" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.089, + 0.472, + 0.141 + ], + "angle": 0, + "content": "
mean ↓Opa. ↓Rot. ↓Scale ↓SH ↓
Random Pick0.410.103.312.470.27
Matched Pairs0.180.043.061.280.04
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.152, + 0.483, + 0.21 + ], + "angle": 0, + "content": "Table 1. Similarities of attributes between matched GS across adjacent frames on RE10K. Random Pick refers to randomly picking two GS from two frames separately, while Matched Pairs refers to our matched GS defined in Eq. (4)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.217, + 0.483, + 0.291 + ], + "angle": 0, + "content": "method on the task of novel view synthesis from monocular videos, as detailed in Sec. 4.1. Additionally, we validate the effectiveness of the proposed alignment module and the efficiency of the Gaussians merge process, as described in Sec. 4.2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.292, + 0.483, + 0.518 + ], + "angle": 0, + "content": "Baselines and datasets. To the best of our knowledge, StreamGS is the first method to reconstruct unposed videos in a feed-forward manner. Consequently, we compare our method separately with pose-free 3DGS works and generalizable splatting methods, including pixelSplat [2], MVSplat [3], and CF-3DGS [9]. PixelSplat and MVSplat are representative methods of generalizable Gaussian Splatting, but they both rely on known camera poses and intrinsics. In contrast, CF-3DGS is a pose-free method but requires optimization loops to align poses and Gaussians. For a comprehensive comparison, we evaluate the methods on large-scale datasets with diverse scenes, including RE10K [38], ACID [17], ScanNet [5], DL3DV [16], and MVImgNet [33]. Each dataset comprises monocular video sequences with per-frame camera pose annotations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.483, + 0.821 + ], + "angle": 0, + "content": "Implementation details. Both the 2D image feature extractor \\(\\phi_{2D}\\) and the MergeNet \\(\\phi_{GS}\\) are double-layer convolutional networks. For fairness, StreamGS and other generalizable methods are trained on the identical training split of the RE10K dataset using the Adam [13] optimizer with the same learning rate of \\(2\\times 10^{-4}\\) and a cosine scheduler. The parameter \\(\\lambda\\) in Eq. 8 is set to 0.05. All methods are trained for 30K iterations and tested on a single NVIDIA Tesla A100 80GB GPU. The images are resized to \\(224\\times 224\\) and the batch size is set to 14. For the non-generalizable method CF-3DGS, we follow its original setting [9]. Note that CF-3DGS is evaluated only on a subset of the full test set due to its low reconstruction efficiency, as shown in Fig. 5. Unlike pose-dependent methods, CF-3DGS and our method require poses of novel views for rendering. CF-3DGS freezes the trained Gaussian model and performs additional optimization steps to learn the poses. Our method, as described in Sec. 3.3, carries out pose alignment and refinement processes to estimate the poses. More details can be referred to supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.83, + 0.292, + 0.846 + ], + "angle": 0, + "content": "4.1. Novel View Synthesis" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.852, + 0.3, + 0.868 + ], + "angle": 0, + "content": "4.1.1. Reconstruction Quality" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Quantitative comparison. We compare StreamGS with baseline methods on the quality and efficiency of novel view" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.486 + ], + "angle": 0, + "content": "synthesis. Following previous 3D-GS research [12], we report PSNR, SSIM [28], and LPIPS [37] as metrics of reconstruction quality. The quantitative results are shown in Tab. 2. On the source domain RE10K [38], existing state-of-the-art generalizable methods demonstrate competitive scores, with MVSplat even outperforming the optimization-based CF-3DGS. However, their performance degrades significantly on out-of-domain datasets such as DL3DV [16] and MVImgNet [33], as these datasets contain various scene types, including outdoor environments and more complex indoor scenes with different objects and illumination conditions compared to RE10K. Naturally, CF-3DGS, which is based on thousands of optimization steps, performs well on the aforementioned datasets. However, its PSNR slightly decreases on ScanNet due to more irregular camera movements and increased motion blur in the frames, posing a challenge for CF-3DGS in recovering camera poses. Pixel-Splat performs well on ScanNet, thanks to its robust feature extraction backbone trained on ImageNet [6], while MVSplat achieves the lowest score. According to the table, StreamGS consistently achieves scores comparable to all other methods. Its PSNR on MVImgNet and DL3DV is significantly higher than that of PixelSplat and MVSplat, even without given poses and intrinsics, demonstrating our method's superior generalizability on unseen datasets with significant domain gaps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.909, + 0.812 + ], + "angle": 0, + "content": "Qualitative comparison. A qualitative comparison with state-of-the-art methods is also presented in Fig. 3 and Fig. 4. While all methods demonstrate high-quality novel view rendering on RE10K [38], our method exhibits superior robustness on out-of-domain datasets. Due to the significant domain gap including texture, illumination and camera motion, both MVSplat and pixelSplat fail to predict accurate depths for the printer shown in the second row, resulting in severe floating artifacts in the rendered image. The third row shows the reconstruction of a plaza from DL3DV [16]. Similarly, since outdoor scenes are much less represented in RE10K [38], MVSplat struggles to extract stereo cues from text-less skies and objects with disparate illumination, deteriorating the rendering quality. The final row shows a bench scene in MVImgNet, on which baseline generalizable methods also fails due to the similar domain gap issues. These cases also demonstrate that the generalizability of pixelSplat surpasses that of MVSplat. As the figure shows, CF-3DGS also does not perform well on some outdoor scenes. In contrast, the visual quality of our method on out-of-domain datasets remains high." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.821, + 0.74, + 0.836 + ], + "angle": 0, + "content": "4.1.2. Reconstruction Efficiency" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.903 + ], + "angle": 0, + "content": "In addition to rendering quality, we also compare the reconstruction efficiency of our method with baseline models. Fig. 5 illustrates a plot of reconstruction quality, measured by the average PSNR reported in Tab. 2, versus ef" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "25846" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.908, + 0.187 + ], + "angle": 0, + "content": "
PF GRE10K [38] (Source Domain)DL3DV [16]MVIImgNet [33]ScanNet [5]ACID [17]
PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑
pixelSplat [2]×✓22.640.210.7519.430.390.5019.220.490.6027.560.190.8229.930.140.84
MVSplat [3]×✓23.540.130.9017.840.360.4516.290.500.5326.250.190.8228.830.140.86
CF-GS [9]✓×23.460.240.7519.930.310.6226.330.330.8822.880.420.7628.160.190.81
StreamGS✓✓22.420.170.8320.540.240.6425.050.310.7928.430.160.8628.500.150.84
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.197, + 0.907, + 0.238 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison with existing state-of-art methods on Novel View Synthesis of monocular videos. PF indicates whether the method is pose-free. G indicates whether the methods is generalizable. Our method consistently achieves scores comparable to state-of-the-art methods that are either not pose-free or lack generalizability. In the table, the best result is highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.24, + 0.905, + 0.503 + ], + "angle": 0, + "content": "
Re10K\n(Source Domain)
ScanNet
DL3DV
MVImgNet
Ground TruthStreamGS (Ours)MVSplatpixelSplatCF-3DGS
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.507, + 0.908, + 0.55 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparison on novel view synthesis. We show the results on both source domain, RE10K [38], and other domains, ScanNet [5], DL3DV [16] and MVImgNet [33]. All generalizable methods are trained only on RE10K and tested on the other datasets. StreamGS outperforms other methods in several challenging scenarios, especially for the out-of-domain data." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.554, + 0.475, + 0.6 + ], + "angle": 0, + "content": "
Coarse Pred (Sec. 3.2)Ada. Refine (Sec. 3.3)ADC (Sec. 3.4)Total
time/sparam/Mtime/sparam/Mtime/sparam/Mtime/sparam/M
StreamGS0.02656.740.081.830.010.040.11658.61
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.147, + 0.601, + 0.428, + 0.612 + ], + "angle": 0, + "content": "Table 3. Efficiency metrics of each component." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.483, + 0.837 + ], + "angle": 0, + "content": "ficiency, measured by the processing time per frame (FPS). The figure shows that StreamGS achieves second place with a PSNR of 23.1, only 0.05 lower than CF-3DGS [9]. This indicates that our method achieves nearly the same rendering quality as the best model. However, thanks to the feed-forward design, StreamGS is 150 times faster than CF-3DGS, predicting Gaussian primitives for up to 9 frames within one second. Without known camera information, our method involves additional alignment and pose estimation processes across frames, which limits the inference speed, making it slower than MVSplat [3] and pixelSplat [2]. However, according to the scores reported in Tab. 2 and Fig. 5, our method is more generalizable and achieves better rendering quality than these methods. Tab. 3 shows the efficiency of each component of our method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.245, + 0.864 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this section, we discuss the effectiveness of our main design about joint refinement (in Sec. 3.3) and feed-forward" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.557, + 0.907, + 0.587 + ], + "angle": 0, + "content": "ADC module (in Sec 3.4). More ablation studies on framework design can be found in supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.593, + 0.792, + 0.607 + ], + "angle": 0, + "content": "4.2.1. Effectiveness of Joint Refinement" + }, + { + "type": "table", + "bbox": [ + 0.528, + 0.621, + 0.897, + 0.702 + ], + "angle": 0, + "content": "
RE10K [38]ACID [17]
PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑
w/o refine.17.250.320.5817.230.390.48
w/ refine22.420.170.8328.500.150.84
" + }, + { + "type": "table_caption", + "bbox": [ + 0.543, + 0.705, + 0.875, + 0.718 + ], + "angle": 0, + "content": "Table 4. Evaluation of effectiveness of joint refinement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Joint refinement of cameras and centers of Gaussians plays a crucial role in the success of our method. To validate the effectiveness of joint refinement, we conduct an ablation study by skipping the refinement process during inference. In other words, the poses and intrinsics are directly estimated by Eq. 2 and 3. Tab. 4 shows the quantitative comparison between the two settings. Without joint refinement, Gaussian primitives are cast from shifted origins of the camera with erroneous orientations, and the poses of novel views are also inaccurate, causing the rendered images to be shifted and distorted. This severely deteriorates the rendering quality, as shown in Tab. 4. The PSNR" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25847" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.086, + 0.905, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.342, + 0.908, + 0.385 + ], + "angle": 0, + "content": "Figure 4. Visual comparison of Gaussian reconstruction and novel view synthesis from image streams with ScanNet [5] dataset. Unlike MVSplat [3], which struggles with view aggregation, our results show significantly better visual quality on OOD data. Note that Our StreamGS and MVSplat are both trained with RE10K [38] data, and MVSplat needs predetermined cameras." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.401, + 0.478, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.568, + 0.483, + 0.643 + ], + "angle": 0, + "content": "Figure 5. Reconstruction speed measured by frames processed per second (FPS). The x-axis is log-scaled for the better visualization. of the images rendered with direct estimation decreases by \\(23.06\\%\\) and \\(38.62\\%\\) on the RE10K [38] and ACID [17] datasets, respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.648, + 0.434, + 0.665 + ], + "angle": 0, + "content": "4.2.2. Effectiveness of Gaussian Merging Process" + }, + { + "type": "table", + "bbox": [ + 0.105, + 0.678, + 0.473, + 0.753 + ], + "angle": 0, + "content": "
MVImgNet [33]ACID [17]
Compress. Ratio ↑PSNR ↑Compress. Ratio ↑PSNR ↑
w/o merge.1.0025.701.0029.21
merge all1.5825.051.6828.07
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.764, + 0.483, + 0.793 + ], + "angle": 0, + "content": "Table 5. Evaluation of the efficiency improvements of the Gaussian merging process and its impact on rendering quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.903 + ], + "angle": 0, + "content": "During the feed-forward ADC, StreamGS prunes pixel-aligned Gaussians through a merging process. To evaluate the memory efficiency improvement and its impact on rendering quality, we compare the average number of Gaussians per frame and PSNR before and after the merging process. We define the compression ratio of Gaussians during the merging process as the ratio of the average number of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.391, + 0.907, + 0.543 + ], + "angle": 0, + "content": "Gaussians per frame, i.e., \\( H \\times W \\), to that after the merging process. The metrics are reported in Tab. 5. The results demonstrate that the merging process can prune Gaussians per frame by \\( 36.71\\% \\) and \\( 40.48\\% \\) on MVImgNet [33] and ACID [17], respectively. Meanwhile, the PSNR scores after the merging process only slightly decrease by \\( 2.53\\% \\) and \\( 3.90\\% \\), respectively. This ablation study demonstrates that the designed Gaussian merging process efficiently reduces memory usage during reconstruction and rendering, with a negligible impact on reconstruction quality." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.554, + 0.634, + 0.57 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.907, + 0.809 + ], + "angle": 0, + "content": "We propose a novel and holistic generalizable pose-free reconstruction pipeline named StreamGS, dedicated to the online reconstruction of endless unposed image streams, such as monocular videos. To the best of our knowledge, our method is the first generalizable model capable of predicting Gaussians corresponding to the input stream in a feedforward manner, without relying on known poses and intrinsics. Compared to pose-free but optimization-based methods, our method achieves comparable reconstruction quality while reducing the learning time to within several milliseconds, avoiding optimization steps. Compared to other generalizable methods, StreamGS eliminates the dependence on poses and intrinsics and manages to reconstruct more accurate scenes on out-of-domain datasets, demonstrating better domain generalizability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Limitations. Although our method runs fast, joint refinement process still includes additional time costs, making it slower than MVSplat [3]. Moreover, our approach encounters common reconstruction challenges, including textureless regions and long sequences." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25848" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.486, + 0.184 + ], + "angle": 0, + "content": "[1] Wenjing Bian, Zirui Wang, Kejie Li, Jia-Wang Bian, and Victor Adrian Prisacariu. Nope-nerf: Optimising neural radiance field with no pose prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4160-4169, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.186, + 0.484, + 0.268 + ], + "angle": 0, + "content": "[2] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.482, + 0.338 + ], + "angle": 0, + "content": "[3] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. arXiv preprint arXiv:2403.14627, 2024. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.341, + 0.483, + 0.409 + ], + "angle": 0, + "content": "[4] Zezhou Cheng, Carlos Esteves, Varun Jampani, Abhishek Kar, Subhransu Maji, and Ameesh Makadia. Lu-nerf: Scene and pose estimation by synchronizing local unposed nerfs. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18312–18321, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.411, + 0.483, + 0.464 + ], + "angle": 0, + "content": "[5] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.467, + 0.483, + 0.522 + ], + "angle": 0, + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.524, + 0.483, + 0.591 + ], + "angle": 0, + "content": "[7] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.20309, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.594, + 0.483, + 0.648 + ], + "angle": 0, + "content": "[8] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.65, + 0.483, + 0.717 + ], + "angle": 0, + "content": "[9] Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A. Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20796-20805, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.72, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[10] Yiming Gao, Yan-Pei Cao, and Ying Shan. Surfelnerf: Neural surfel radiance fields for online photorealistic reconstruction of indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 108-118, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[11] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.847, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[12] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 3, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.486, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "[13] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[14] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Ep n p: An accurate o (n) solution to the p np problem. International journal of computer vision, 81:155-166, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.18, + 0.905, + 0.207 + ], + "angle": 0, + "content": "[15] Vincent Leroy, Yohann Cabon, and Jerome Revaud. Grounding image matching in 3d with mast3r, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.209, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[16] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. D13dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.294, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[17] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In ICCV, 2021. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.35, + 0.905, + 0.419 + ], + "angle": 0, + "content": "[18] Tianqi Liu, Guangcong Wang, Shoukang Hu, Liao Shen, Xinyi Ye, Yuhang Zang, Zhiguo Cao, Wei Li, and Ziwei Liu. Mvsgaussian: Fast generalizable gaussian splattering reconstruction from multi-view stereo. In European Conference on Computer Vision, pages 37-53. Springer, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.421, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[19] Andreas Meuleman, Yu-Lun Liu, Chen Gao, Jia-Bin Huang, Changil Kim, Min H Kim, and Johannes Kopf. Progressively optimized local radiance fields for robust view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16539-16548, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.491, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[20] Jongmin Park, Minh-Quan Viet Bui, Juan Luis Gonzalez Bello, Jaeho Moon, Jihyong Oh, and Munchurl Kim. Splinegs: Robust motion-adaptive spline for real-time dynamic 3d gaussians from monocular video. arXiv preprint arXiv:2412.09982, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[21] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.619, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[22] Qiuhong Shen, Zike Wu, Xuanyu Yi, Pan Zhou, Hanwang Zhang, Shuicheng Yan, and Xinchao Wang. Gamba: Marry gaussian splatting with mamba for single view 3d reconstruction. arXiv preprint arXiv:2403.18795, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[23] Liang Song, Guangming Wang, Jiuming Liu, Zhenyang Fu, Yanzi Miao, et al. Sc-nerf: Self-correcting neural radiance field with sparse views. arXiv preprint arXiv:2309.05028, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[24] Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10208-10217, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[25] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In European Conference on Computer Vision, pages 1-18. Springer, 2025. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25849" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[27] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[28] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[29] Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. Grm: Large gaussian reconstruction model for efficient 3d reconstruction and generation. arXiv preprint arXiv:2403.14621, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[30] Chi Yan, Delin Qu, Dan Xu, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Gs-slam: Dense visual slam with 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19595-19604, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[31] Xuanyu Yi, Zike Wu, Qiuhong Shen, Qingshan Xu, Pan Zhou, Joo-Hwee Lim, Shuicheng Yan, Xinchao Wang, and Hanwang Zhang. Mvgamba: Unify 3d content generation as state space sequence modeling. arXiv preprint arXiv:2406.06367, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.461 + ], + "angle": 0, + "content": "[32] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4578-4587, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.531 + ], + "angle": 0, + "content": "[33] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Tianyou Liang, Guanying Chen, Shuguang Cui, and Xiaoguang Han. Mvimgnet: A large-scale dataset of multi-view images. In CVPR, 2023. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[34] Vladimir Yugay, Yue Li, Theo Gevers, and Martin R Oswald. Gaussian-slam: Photo-realistic dense slam with gaussian splatting. arXiv preprint arXiv:2312.10070, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.646 + ], + "angle": 0, + "content": "[35] Haojun Zhang, Yuan Yao, and Xuefeng Yan. Improved end-to-end multilevel nerf-based dense rgb-d slam. In Chinese Conference on Pattern Recognition and Computer Vision (PRCV), pages 132-146. Springer, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[36] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In European Conference on Computer Vision, pages 1-19. Springer, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[37] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[38] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. ACM Trans. Graph., 37, 2018. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[39] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12786-12796, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[40] Zihan Zhu, Songyou Peng, Viktor Larsson, Zhaopeng Cui, Martin R Oswald, Andreas Geiger, and Marc Pollefeys. Nicer-slam: Neural implicit scene encoding for rgb slam. In 2024 International Conference on 3D Vision (3DV), pages 42-52. IEEE, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[41] Zi-Xin Zou, Shi-Sheng Huang, Yan-Pei Cao, Tai-Jiang Mu, Ying Shan, and Hongbo Fu. Mononeuralfusion: Online monocular neural 3d reconstruction with geometric priors. arXiv preprint arXiv:2209.15153, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.907, + 0.303 + ], + "angle": 0, + "content": "[42] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10324-10335, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25850" + } + ] +] \ No newline at end of file diff --git a/2025/StreamGS_ Online Generalizable Gaussian Splatting Reconstruction for Unposed Image Streams/9d1940fe-237f-4495-9085-9441928f0c47_origin.pdf b/2025/StreamGS_ Online Generalizable Gaussian Splatting Reconstruction for Unposed Image Streams/9d1940fe-237f-4495-9085-9441928f0c47_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5ce26495f1fab93fe58f58f3f8758ce4c49710eb --- /dev/null +++ b/2025/StreamGS_ Online Generalizable Gaussian Splatting Reconstruction for Unposed Image Streams/9d1940fe-237f-4495-9085-9441928f0c47_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd003e7812854156f0173d8e3ca4fb156d3750c195e1d587dabed56677b79249 +size 1575854 diff --git a/2025/StreamGS_ Online Generalizable Gaussian Splatting Reconstruction for Unposed Image Streams/full.md b/2025/StreamGS_ Online Generalizable Gaussian Splatting Reconstruction for Unposed Image Streams/full.md new file mode 100644 index 0000000000000000000000000000000000000000..375a7f68d72fce11ec283ae9314f8ec4aa70ca62 --- /dev/null +++ b/2025/StreamGS_ Online Generalizable Gaussian Splatting Reconstruction for Unposed Image Streams/full.md @@ -0,0 +1,278 @@ +# StreamGS: Online Generalizable Gaussian Splitting Reconstruction for Unposed Image Streams + +Yang Li $^{1,4,*}$ Jinglu Wang $^{1}$ Lei Chu $^{1}$ Xiao Li $^{1}$ Shiu-Hong Kao $^{1,2,*}$ Ying-Cong Chen $^{2,3}$ Yan Lu $^{1}$ $^{1}$ Microsoft Research Asia ${}^{2}$ HKUST ${}^{3}$ HKUST(GZ) ${}^{4}$ Rutgers University yangliaftermath@gmail.com skao@cse.ust.hk yingcongchen@ust.hk {jinglu.wang, leichu, li.xiao, yanlu}@microsoft.com + +# Abstract + +The advent of 3D Gaussian Splatting (3DGS) has advanced 3D scene reconstruction and novel view synthesis. With the growing interest of interactive applications that need immediate feedback, online 3DGS reconstruction in real-time is in high demand. However, none of existing methods yet meet the demand due to three main challenges: the absence of predetermined camera parameters, the need for generalizable 3DGS optimization, and the necessity of reducing redundancy. We propose StreamGS, an online generalizable 3DGS reconstruction method for unposed image streams, which progressively transform image streams to 3D Gaussian streams by predicting and aggregating per-frame Gaussians. Our method overcomes the limitation of the initial point reconstruction [27] in tackling out-of-domain (OOD) issues by introducing a content adaptive refinement. The refinement enhances cross-frame consistency by establishing reliable pixel correspondences between adjacent frames. Such correspondences further aid in merging redundant Gaussians through cross-frame feature aggregation. The density of Gaussians is thereby reduced, empowering online reconstruction by significantly lowering computational and memory costs. Extensive experiments on diverse datasets have demonstrated that StreamGS achieves quality on par with optimization-based approaches but does so 150 times faster, and exhibits superior generalizability in handling OOD scenes. + +# 1. Introduction + +The field of 3D Scene reconstruction [10, 30] for novel view synthesis from image streams has gained increasing attention, due to its significance in enabling interactive applications that offer users instant feedback. In this context, the + +![](images/a2731129e2b61360219e70938d881361d034a47e89913f3e31c904ee8fdc7188.jpg) +Figure 1. The proposed StreamGS efficiently transforms image streams into Gaussian streams by progressively reconstructing and aggregating per-frame 3D Gaussians. We show our reconstructed 3DGS (visualized as points) alongside estimated camera poses (in blue), and synthesized novel views. + +advent of 3D Gaussian Splatting (3DGS) [12] marks a major advancement in high-quality, real-time rendering. This progress highlights the urgency for efficient, on-the-fly generation of 3DGS from image streams. + +Nonetheless, online 3DGS reconstruction faces unique challenges. (1) Unknown camera. The conventional preprocessing using Structure from Motion (SfM) [21] for camera estimation is impractical for real-time streaming. + +This is due to the absence of the full image set and the time-consuming computation. Thus, there is a need for methods that can operate without pre-determined cameras. (2) Generalizability. 3DGS reconstruction requires multiteration optimization, impractical for online applications due to the need for all images in advance. This restricts the development of a generalizable method that can process image streams in a feed-forward manner. (3) Redundancy. The significant overlap between frames leads to high redundancy when reconstructing Gaussians individually for each frame, increasing the streaming process's resource intensity. + +Scene reconstruction without known cameras has been explored in SLAM-based [30, 34, 35, 40, 41] and NeRF-based [1, 23] methods. Only a few 3DGS-based methods are discussed [9, 20]. These methods primarily consider camera poses as learnable parameters and optimized alongside Gaussians through the iterative optimization. Yet, this optimization-driven approach for aligning each frame considerably increases the reconstruction time, rendering it impractical in the online scenario. + +Recently, generalizable 3DGS reconstruction has been investigated for sparse views [2, 3, 24]. These approaches transform pixels into Gaussians, whose parameters are decoded from images via 2D encoder-decoder networks. They achieve the 3DGS reconstruction for each image in a single feed-forward pass. However, these generalizable approaches are primarily designed for monocular or binocular settings, suited to sparse-view inputs. In addition, multi-view methods [2, 3] typically infer Gaussian centers through stereo matching, highly dependent on known cameras, limiting their applicability for online scenarios with numerous unposed images. Furthermore, they combine multi-view 3DGS sets by simply uniting them, which overlooks cross-view alignment and overlaps, leading to misalignment and redundancy issues within image streams. The recent method [7] performs Gaussian downsampling in 3D space to reduce redundancy, but it necessitates traversing and processing Gaussians in 3D grids, which is also time-consuming and is highly dependent on the grid resolution. + +Emerging models like DUSt3R [27] and MASt3R [15] enable sparse-view geometry reconstruction by simultaneously predicting 3D points and estimating camera parameters. These advances pave the way for more efficient, feedforward, pose-free 3D reconstruction methods. A straightforward approach to generalize 3DGS reconstruction is to add a Gaussian predictor to the DUSt3R-like framework. However, this introduces several challenges. These models usually require datasets with ground truth 3D geometry, which may not always be available. Additionally, applying pretrained models to out-of-domain (OOD) data can result in inaccurate pose and 3D point estimations. Moreover, generating 3D points for each frame individually causes re + +dundancy due to overlapping adjacent frames, potentially leading to ghosting artifacts from pose estimation errors. + +In this paper, we introduce StreamGS, a novel pipeline for online, generalizable 3DGS reconstruction from unposed image streams. StreamGS aims to progressively construct and update the 3DGS representation of the scene frame-by-frame, in a feed-forward manner, as illustrated in Fig. 1. We leverage the pretrained DUSt3R to initially predict 3D point for the current frame using the previous frame as a reference. However, this initialization may encounter inaccuracies due to OOD issues. To mitigate this, we capitalize the insight that adjacent frames offer sufficient correspondences to refine the reconstruction. Unlike DUSt3R using predicted 3D points to establish correspondences, we adopt content-adaptive descriptors for more reliable matching, allowing for the adaptive refinement of the reconstruction by enhancing consistency between adjacent views. Furthermore, such correspondences help to prune redundant pixel-aligned Gaussians. Correlated pixel-wise features across frames are effectively aggregated, removing duplicates and achieving adaptive density control. Finally, we decode Gaussians from such aggregated features. StreamGS is adept at predicting and integrating Gaussians for the current frame into the existing Gaussian set seamlessly with a feed-forward pass. In summary, our contribution is three-fold. + +- We introduce a novel pipeline for the online, generalizable reconstruction of image streams without requiring camera parameters, marking a first in this field. +- The proposed adaptive refinement enhance cross-frame consistency of 3DGS reconstruction, and the adaptive density control mechanism minimizes adjacent-view redundancy, thereby highly reducing computational costs in online reconstruction. +- Upon evaluation across diverse datasets, our method achieves high novel view synthesis quality comparable to the optimization-based method [9] but with $150\mathrm{x}$ faster reconstruction speed. Additionally, our method outperforms existing pose-dependent generalizable 3DGS methods in handling OOD scenes, showing superior generalizability. + +# 2. Related Works + +Generalizable 3D Gaussian Splitting. Many recent studies aim to propose generalizable 3D-GS methods capable of predicting Gaussians within a single feed-forward pass. These works can be classified into two main categories: single-view reconstruction and multi-view reconstruction. Single-view reconstruction does not involve pose estimation as there are no multi-view constraints. Inspired by the insight from LRM [11] that large transformer-based [26] backbone networks can learn 3D priors from large-scale 3D data, the potential of predicting Gaussians from a + +single image in a single feed-forward pass has been comprehensively explored. Numerous feed-forward models have been proposed, such as GRM [29], TriplaneGS [42], and GMamba [22]. However, these methods are primarily not applicable to multi-view scenarios as they always assume canonical poses. + +Concurrently, many works focusing on multi-view inputs follow a similar paradigm, such as GS-LRM [36], LGM [25], and MVGMamba [31]. They concatenate the input images with camera embeddings like Plücker rays to facilitate the network in learning the proper fusion of Gaussians from multi-view inputs. However, these large 3D backbone networks mostly perform well only on synthetic objects due to the shortage of large-scale scene-level 3D data in the real world. Referring to generalizable NeRF methods like pixelNeRF [32], other research turns to multi-view stereo (MVS) matching to locate or initialize the centers of Gaussians, with other attributes decoded using a lightweight 2D encoder. Representative works with this design include MVSGaussian [18], pixelSplat [2], and MVSplat [3]. However, both camera ray embedding in large transformer-based models and stereo matching rely on known poses and intrinsicities of each input view. Another main limitation of these methods is that they focus on sparse-view inputs. PixelSplat [2] and MVSplat [3] only support up to three views. Therefore, existing generalizable 3D-GS models cannot address the problem of feed-forward reconstruction from endless image streams. + +Pose-free 3D Gaussian Splatting. Recently, many works have aimed to eliminate the need for Structure-from-Motion (SfM) preprocessing steps using COLMAP [21] software. Following the design of previous pose-free NeRF methods like NoPe-NeRF [1], Lu-NeRF [4], and localRF [19], CF-3DGS [9] introduces depth priors into the optimization of 3D-GS and performs progressive reconstruction. As each new image arrives, CF-3DGS optimizes both the pose and 3D Gaussians of the input image based on its depths, aiming to align the 3DGS from the new image with the preceding reconstruction. However, it still relies on known camera intrinsics. CF-3DGS is also not robust, as the accuracy of depth priors significantly impacts its reconstruction quality, limiting its application to common scenes. Moreover, it depends on thousands of optimization steps for each view, significantly extending the reconstruction time for each scene. Compared to generalizable 3D-GS models, current pose-free methods are so inefficient that they cannot be applied to image streams with a large number of frames. + +Online 3D reconstruction of image streams. Online 3D reconstruction has been extensively studied in the field of SLAM [30, 34, 35, 39-41]. However, these methods typically involve additional information and most leverage SDF + +and NeRF as scene representations. NICE-SLAM [39] uses RGB-D streams as input, with the reconstructed scene represented by NeRF. NICER-SLAM [40] relies on geometric priors, including surface normals and depths. Surfel-NeRF [10] focuses on the novel view synthesis quality of online reconstruction with RGB streams, but it requires the poses and intrinsics of frames. Gaussian-SLAM [34] reconstructs the scene using 3D-GS, but it also relies on RGB-D streams. + +# 3. Methods + +Given a sequence of unposed images over time, our objective is to progressively reconstruct the 3D Gaussian Splating (3DGS) representation in an online manner. Specifically, at each timestamp $t$ , the goal is to derive the 3DGS $\mathcal{G}^t$ , which encapsulates the 3D scene aggregated from the images $\mathcal{I}^t = \{\mathbf{I}^i\}_{i=1}^t$ . + +Fig. 2 illustrates our overall framework. In order to make reconstruction efficient with limited computational resources, we employ an incremental construction strategy. At each time step $t$ , we focus on generating the 3DGS $\mathbf{G}^t$ of the current frame and merge it with the previous accumulated reconstruction $\mathcal{G}^{t-1}$ to obtain the full reconstruction $\mathcal{G}^t$ of the current time step. Specifically, with current frame $\mathbf{I}^t$ , we use $\mathbf{I}^{t-1}$ as reference frame and estimate the point maps and cameras of each using an initial reconstruction module. With newly established matches between current and reference views, we further refine the quality of the points and the cameras in the adaptive refinement module. Finally, we generate $\mathbf{G}^t$ using the refined points and features of $\mathbf{I}^t$ , and merge it with previous reconstructions according to the established matches, achieving the feedforward adaptive density control (ADC). + +# 3.1. Preliminaries + +Gaussian splatting. Previous work [12] represents a scene or object using a set of Gaussian distributions. Specifically each gaussian primitive could be denoted as $G(x;\mu ,\Sigma) = e^{-\frac{1}{2} (x - \mu)^T\Sigma^{-1}(x - \mu)}$ and the covariance $\Sigma$ is decomposed into the rotation matrix $R$ and scaling matrix $S$ to ensure the positive semi-definiteness during optimization, that is $\Sigma = RSS^T R^T$ . The view-dependent color of the appearance is represented by a set of spherical harmonics (SH) coefficients and opacity value $\alpha$ . + +# 3.2. Initial Two-view Reconstruction + +Given the current frame $\mathbf{I}^t$ and a reference frame $\mathbf{I}^{t'}$ , we use a coarse predictor $\phi_{3D}$ and estimate the point map $\mathbf{X}^{t|t'}$ of the current frame under the local coordinate system of the reference frame together with its corresponding confidence map $\mathbf{C}^{t|t'}$ , where the superscript $\cdot |t'$ indicates that the local coordinate system adapts to that of $\mathbf{I}^{t'}$ . Intuitively, the predicted point map stores the 3D point coordinate that the + +![](images/fe2c56a6263ac031dabbc6ba493b884753d04e3d39a3df14e5e387effe744e8c.jpg) +Figure 2. Method overview. Our StreamGS progressively reconstruct and aggregate 3D Gaussians from the unposed image stream. Given the adjacent image pair $(\mathbf{I}^{t - 1},\mathbf{I}^t)$ , we first perform the initial reconstruction that predicts pixel-wise 3D points with their features and coarse camera poses, using a pretrained coarse predictor. Since the coarse predictions may suffer from OOD issues, we refine both the camera poses and 3D positions by establishing new point-wise correspondences. We aggregate cross-frame image and 3D features by warping and merging to reduce redundancy. Finally we decode the aggregated features to Gaussian primitives. + +pixel unprojected to in, and the confidence maps measure the certainty of point maps at each pixel, reflecting a prior-based estimation of reconstruction accuracy and difficulty. Formally, we define + +$$ +\left(\mathbf {X} ^ {t \mid t ^ {\prime}}, \mathbf {X} ^ {t ^ {\prime} \mid t ^ {\prime}}, \mathbf {C} ^ {t \mid t ^ {\prime}}, \mathbf {C} ^ {t ^ {\prime} \mid t ^ {\prime}}\right) = \phi_ {3 D} \left(\mathbf {I} ^ {t}, \mathbf {I} ^ {t ^ {\prime}}\right). \tag {1} +$$ + +In order to help align the current local-coordinate pointmap with the global coordinate, we also get the output in the local coordinate system of $\mathbf{I}^t$ by computing $\phi_{3D}(\mathbf{I}^{t'},\mathbf{I}^t)$ . With these predicted point maps, we could further estimate the camera matrix $\mathbf{P} = \mathbf{K}[\mathbf{R}|\mathbf{t}]$ , composed of its intrinsic parameters $\mathbf{K}$ , and its extrinsic parameters, the rotation matrix $\mathbf{R}$ and the translation vector $\mathbf{t}$ . Specifically, we assume that the principle points $\mathbf{c}$ are centered and pixels are squares and have + +$$ +\hat {f} ^ {t} = \arg \min _ {f} \Sigma_ {\mathbf {p} \in \mathbf {I} ^ {t}} \| \mathbf {p} - \mathbf {c} - f \frac {\left(\mathbf {x} _ {\mathbf {p}} , \mathbf {y} _ {\mathbf {p}}\right)}{\mathbf {z} _ {\mathbf {p}}} \|, \tag {2} +$$ + +where $(\mathbf{x_p},\mathbf{y_p},\mathbf{z_p})\in \mathbf{X}^{t|t}$ is the 3D point coordinate that the pixel $\mathbf{p}$ unprojected to. Moreover, we approximate relative pose $\mathbf{P}^t = [\mathbf{R},\mathbf{t}]$ of $\mathbf{I}^t$ to $\mathbf{I}^{t'}$ by solving the following points registration problem: + +$$ +\left[ \mathbf {R} ^ {t} \mid \mathbf {t} ^ {t} \right] = \arg \min _ {s, \mathbf {R}, \mathbf {t}} \sum_ {\mathbf {p} \in \mathbf {I} ^ {t - 1}} \mathbf {C} ^ {t} (\mathbf {p}) \| s \left(\mathbf {R} \mathbf {X} ^ {t ^ {\prime} | t ^ {\prime}} (\mathbf {p}) + \mathbf {t}\right) - \mathbf {X} ^ {t ^ {\prime} | t} (\mathbf {p}) \| ^ {2}, \tag {3} +$$ + +$$ +\mathbf {C} ^ {t} = \mathbf {C} ^ {t ^ {\prime} | t ^ {\prime}} \odot \mathbf {C} ^ {t ^ {\prime} | t}, +$$ + +where $\odot$ is the Hadamard product and $s$ is the scale factor. + +In our implementation, we leverage DUSt3R [27] as the coarse predictor due to its effective pretraining on dedicated 3D scene datasets and its efficiency in reconstructing 3D points. For each time step $t$ , we process the pair + +$(\mathbf{I}^{t - 1},\mathbf{I}^t)$ to derive the 3D points $(\mathbf{X}^{t|t - 1},\mathbf{X}^{t - 1|t - 1})$ corresponding to both frames referenced in the coordinate frame $t - 1$ , the reversed pair $(\mathbf{I}^t,\mathbf{I}^{t - 1})$ to obtain the same points $(\mathbf{X}^{t - 1|t},\mathbf{X}^{t|t})$ but in the coordinate frame of $t$ , which will be reused at the next timestamp to boost efficiency. For simplicity, and without loss of generality, the following discussion will focus on the reconstruction of the image pair $(\mathbf{I}^{t - 1},\mathbf{I}^t)$ . Note that training $\phi_{3D}$ requires 3D geometric supervision, which may not be available in our monocular video input scenario. Thus, we leverage the pretrained $\phi_{3D}$ from [27]. + +# 3.3. Adaptive Refinement + +We note that the initial reconstruction quality is compromised due to the OOD challenge as the coarse predictor is frozen. This observation motivates us to enhance reconstruction through content adaptation, with the goal of adaptively refining both poses and 3D reconstructions. + +We perform the adaptive refinement based on establishing new robust matches between adjacent frames. We employ a matching head, $\phi_{\text{match}}$ , to extract local 3D features, denoted as $(\mathbf{F}_{3D}^{t-1}, \mathbf{F}_{3D}^t) \in \mathbb{R}^{H \times W \times d}$ , from the consecutive image pair $(\mathbf{I}^{t-1}, \mathbf{I}^t)$ . The matches between the two images can be established through nearest reciprocal (NN) searching, satisfying the following condition: + +$$ +\begin{array}{l} \mathcal {M} ^ {t - 1, t} = \left\{i _ {k} \leftrightarrow j _ {k} \mid i _ {k} = \mathrm {N N} (j _ {k}) \text {a n d} j _ {k} = \mathrm {N N} (i _ {k}) \right\} _ {k = 1} ^ {N}, \\ \text {s . t .} \operatorname {N N} \left(i _ {k}\right) = \underset {0 \leq j _ {k} \leq H \times W} {\arg \min } | 1 - \cos < \mathbf {F} _ {3 D, i _ {k}} ^ {t - 1}, \mathbf {F} _ {3 D, j _ {k}} ^ {t} > |, \tag {4} \\ \end{array} +$$ + +where $i_k, j_k$ are pixel index in image $\mathbf{I}^{t-1}$ , $\mathbf{I}^t$ respectively, and the measurement of feature distance is cosine similarity. With the correspondences found, a residual transform + +$\Delta = [\Delta \mathbf{R}, \Delta \mathbf{t}]$ can be re-estimated following Eq. (3) by only taking matched points into account. Then we apply the residual transform to the pointmap $\mathbf{X}^{t,t-1}$ to retain refined $\tilde{\mathbf{X}}^{t|t-1}$ , and the pose of $\mathbf{I}^t$ is updated to $\tilde{\mathbf{P}}^t$ by performing PnP-RANSAC [8, 14] on 3D-2D correspondences derived from matches. + +Gaussian decoding. We directly predict the other parameters of 3D Gaussians at each pixel with a light-weight decoder $\phi_{GS}$ as: + +$$ +\mathbf {G} ^ {i} = \left[ \mathbf {q} ^ {i}, \mathbf {s} ^ {i}, \alpha^ {i}, \mathbf {c} ^ {i} \right] = \phi_ {G S} (\mathbf {F} _ {g s}), \tag {5} +$$ + +$$ +\mathbf {F} _ {G S} ^ {i} = \mathbf {F} _ {2 D} ^ {i} \oplus \mathbf {X} ^ {i} \oplus \mathbf {F} _ {3 D} ^ {i}, \quad \mathbf {F} _ {2 D} = \phi_ {2 D} (\mathbf {I} ^ {i}), +$$ + +where $i = \{t - 1,t\}$ $\oplus$ denotes the channel-wise concatenation, $\phi_{2D}$ denotes the 2D image feature extractor, $\mathbf{q}^i\in \mathbb{R}^{H\times W\times 4}$ and $\mathbf{s}^i\in \mathbb{R}^{H\times W\times 3}$ represent rotation quaternions and scales of pixel-aligned Gaussians at image $\mathbf{I}^i$ . We incorporate an additional image feature extractor because the coarse predictor is frozen and cannot be trained by our monocular video setting. Extracting new image features is essential for decoding Gaussians, especially for texture-related properties. Experiments show the importance of the image feature extractor. Then covariance matrix is built with $\Sigma^i = \mathbf{R}(\mathbf{q}^i)\mathbf{ss}^{\mathrm{T}}\mathbf{R}(\mathbf{q}^i)^{\mathrm{T}}$ Note that $\bar{\mathbf{G}}$ is not the final Gaussians since it needs to be merged into the previous Gaussian set following Sec. 3.4. + +# 3.4. Feed-Forward ADC + +With pixel-aligned Gaussian parameters $\mathbf{G}^t$ of $T$ images, previous methods [2, 3, 24, 25] always naively take the union of Gaussians in all images as the final prediction, i.e., $\mathcal{G} = \bigcup_{t=1}^{T} \mathbf{G}^i$ and $|\mathcal{G}| = T \times H \times W$ . However, this approach is both memory-intensive and inefficient in rendering, particularly when dealing with the continuous input of video frames during online reconstruction. Our key observation is that the matched Gaussian pairs in neighboring frames are excessive and prunable since they consistently share similar attributes in shape and color, and are closely distributed, which is validated in Tab. 1. It is noted that we have already acquired dense pixel-wise matches between neighboring frames from Eq. (4). Therefore, we propose a novel feed-forward Adaptive Density Control strategy based on revisiting these dense correspondences. + +Feature aggregation. The primary advantage of pixel-based correspondences is that they are able to convert the computationally intensive 3D Gaussian aggregation process into a more efficient 2D pixel-wise one. This significantly enhances computational efficiency. Initially, the feature of Gaussian parameters $\mathbf{F}_{gs}^{t}$ of the frame $\mathbf{I}^t$ can be aligned to + +the previous frame $\mathbf{I}^{t - 1}$ using the following wrapping: + +$$ +\mathbf {F} _ {G S} ^ {t \mid t - 1} (j) = \left\{ \begin{array}{l l} \mathbf {F} _ {G S} ^ {t} (k) & \text {i f} (j, k) \in \overline {{\mathcal {M}}} ^ {t - 1, t} \\ \mathbf {F} _ {G S} ^ {t - 1} (j) & \text {e l s e ,} \end{array} \right. \tag {6} +$$ + +where $\mathbf{F}_{GS}^{t|t-1}$ denotes the feature of Gaussian primitives in the frame $\mathbf{I}^t$ aligned to $\mathbf{I}^{t-1}$ . Instead of using the raw correspondences set $\mathcal{M}^{t-1,t}$ in Eq. (4), we use the extended set $\overline{\mathcal{M}}^{t-1,t}$ , which includes matches of neighboring pixels such as $(i+1,j+1)$ in addition to the initial $(i,j)$ . It serves as a type of anti-aliasing technique to reduce void pixels within the wrapped feature map. As for the unmatched pixels, we simply replicate the corresponding feature vector within $\mathbf{F}_{gs}^{t-1}$ . With the aligned feature, the Gaussian feature of two frames can be merged by modifying Eq. (5), taking the form: + +$$ +\hat {\mathbf {G}} ^ {t \mid t - 1} = \left[ \hat {\mathbf {q}} ^ {t}, \hat {\mathbf {s}} ^ {t}, \hat {\alpha} ^ {t}, \hat {\mathbf {c}} ^ {t} \right] = \phi_ {M G} \left(\mathbf {F} _ {G S} ^ {t \mid t - 1} \oplus \mathbf {F} _ {G S} ^ {t - 1}\right), \tag {7} +$$ + +where $\phi_{MG}$ denotes the MergeNet that simply consists of two convolutional layers, which merges features and decodes them to Gaussian primitives. In this way, every matched pair of Gaussians between $\mathbf{I}^t$ and $\mathbf{I}^{t - 1}$ is aggregated into a single one, highly reducing the number of Gaussians. The final aggregated Gaussian set is $\mathcal{G}^t = \mathcal{G}^{t - 1}\cup \mathbf{G}^{t|t - 1}$ . + +Without gradient back-propagation of rendering loss in the original 3DGS paper [12], our ADC process runs exceptional fast. In terms of the input group, the final prediction of Gaussian primitives $\mathcal{G}^t$ consists of the merged Gaussians and the unmatched Gaussians at each frame. + +# 3.5. Loss Functions + +The optimization of the 2D feature extractor $\phi_{2D}$ , the Gaussian decoder network, $\phi_{GS}$ involves both a rendering loss function and a reconstruction loss function: + +$$ +\begin{array}{l} \mathcal {L} \left(\mathbf {I} ^ {i}, \hat {\mathbf {I}} ^ {i}, \hat {\mathbf {I}} _ {M} ^ {i}\right) = \mathcal {L} _ {\text {r e n d e r}} \left(\mathbf {I} ^ {i}, \hat {\mathbf {I}} ^ {i}\right) + \mathcal {L} _ {\text {r e c o n}} \left(\hat {\mathbf {I}} ^ {i}, \hat {\mathbf {I}} _ {M} ^ {i}\right) \tag {8} \\ = \| \mathbf {I} ^ {i} - \hat {\mathbf {I}} ^ {i} \| _ {2} + \lambda \| \mathbf {I} ^ {i} - \hat {\mathbf {I}} ^ {i} \| _ {\mathrm {L P I P S}} + \| \hat {\mathbf {I}} _ {\mathrm {M}} ^ {i} - \hat {\mathbf {I}} ^ {i} \| _ {2}, \\ \end{array} +$$ + +where $\mathbf{I}^i$ is the ground truth frame, $\hat{\mathbf{I}}^i$ is the rendered image with full Gaussian primitives before the merge process, and $\hat{\mathbf{I}}_{\mathrm{M}}^{i}$ is the rendered image with merged Gaussians. Due to the lightweight nature of the two networks, the algorithm converges quickly within thousands of steps. The reconstruction loss term facilitates the merge network in fusing the pixel-aligned Gaussian parameters from different frames, aiming to render the same frame as before the merge process but with significantly fewer Gaussian primitives. + +# 4. Experiments + +StreamGS operates on an image stream of a scene, jointly predicting the corresponding poses and Gaussians in a feedforward manner. To assess its performance, we evaluate our + +
mean ↓Opa. ↓Rot. ↓Scale ↓SH ↓
Random Pick0.410.103.312.470.27
Matched Pairs0.180.043.061.280.04
+ +Table 1. Similarities of attributes between matched GS across adjacent frames on RE10K. Random Pick refers to randomly picking two GS from two frames separately, while Matched Pairs refers to our matched GS defined in Eq. (4). + +method on the task of novel view synthesis from monocular videos, as detailed in Sec. 4.1. Additionally, we validate the effectiveness of the proposed alignment module and the efficiency of the Gaussians merge process, as described in Sec. 4.2. + +Baselines and datasets. To the best of our knowledge, StreamGS is the first method to reconstruct unposed videos in a feed-forward manner. Consequently, we compare our method separately with pose-free 3DGS works and generalizable splatting methods, including pixelSplat [2], MVSplat [3], and CF-3DGS [9]. PixelSplat and MVSplat are representative methods of generalizable Gaussian Splatting, but they both rely on known camera poses and intrinsics. In contrast, CF-3DGS is a pose-free method but requires optimization loops to align poses and Gaussians. For a comprehensive comparison, we evaluate the methods on large-scale datasets with diverse scenes, including RE10K [38], ACID [17], ScanNet [5], DL3DV [16], and MVImgNet [33]. Each dataset comprises monocular video sequences with per-frame camera pose annotations. + +Implementation details. Both the 2D image feature extractor $\phi_{2D}$ and the MergeNet $\phi_{GS}$ are double-layer convolutional networks. For fairness, StreamGS and other generalizable methods are trained on the identical training split of the RE10K dataset using the Adam [13] optimizer with the same learning rate of $2\times 10^{-4}$ and a cosine scheduler. The parameter $\lambda$ in Eq. 8 is set to 0.05. All methods are trained for 30K iterations and tested on a single NVIDIA Tesla A100 80GB GPU. The images are resized to $224\times 224$ and the batch size is set to 14. For the non-generalizable method CF-3DGS, we follow its original setting [9]. Note that CF-3DGS is evaluated only on a subset of the full test set due to its low reconstruction efficiency, as shown in Fig. 5. Unlike pose-dependent methods, CF-3DGS and our method require poses of novel views for rendering. CF-3DGS freezes the trained Gaussian model and performs additional optimization steps to learn the poses. Our method, as described in Sec. 3.3, carries out pose alignment and refinement processes to estimate the poses. More details can be referred to supplementary materials. + +# 4.1. Novel View Synthesis + +# 4.1.1. Reconstruction Quality + +Quantitative comparison. We compare StreamGS with baseline methods on the quality and efficiency of novel view + +synthesis. Following previous 3D-GS research [12], we report PSNR, SSIM [28], and LPIPS [37] as metrics of reconstruction quality. The quantitative results are shown in Tab. 2. On the source domain RE10K [38], existing state-of-the-art generalizable methods demonstrate competitive scores, with MVSplat even outperforming the optimization-based CF-3DGS. However, their performance degrades significantly on out-of-domain datasets such as DL3DV [16] and MVImgNet [33], as these datasets contain various scene types, including outdoor environments and more complex indoor scenes with different objects and illumination conditions compared to RE10K. Naturally, CF-3DGS, which is based on thousands of optimization steps, performs well on the aforementioned datasets. However, its PSNR slightly decreases on ScanNet due to more irregular camera movements and increased motion blur in the frames, posing a challenge for CF-3DGS in recovering camera poses. Pixel-Splat performs well on ScanNet, thanks to its robust feature extraction backbone trained on ImageNet [6], while MVSplat achieves the lowest score. According to the table, StreamGS consistently achieves scores comparable to all other methods. Its PSNR on MVImgNet and DL3DV is significantly higher than that of PixelSplat and MVSplat, even without given poses and intrinsics, demonstrating our method's superior generalizability on unseen datasets with significant domain gaps. + +Qualitative comparison. A qualitative comparison with state-of-the-art methods is also presented in Fig. 3 and Fig. 4. While all methods demonstrate high-quality novel view rendering on RE10K [38], our method exhibits superior robustness on out-of-domain datasets. Due to the significant domain gap including texture, illumination and camera motion, both MVSplat and pixelSplat fail to predict accurate depths for the printer shown in the second row, resulting in severe floating artifacts in the rendered image. The third row shows the reconstruction of a plaza from DL3DV [16]. Similarly, since outdoor scenes are much less represented in RE10K [38], MVSplat struggles to extract stereo cues from text-less skies and objects with disparate illumination, deteriorating the rendering quality. The final row shows a bench scene in MVImgNet, on which baseline generalizable methods also fails due to the similar domain gap issues. These cases also demonstrate that the generalizability of pixelSplat surpasses that of MVSplat. As the figure shows, CF-3DGS also does not perform well on some outdoor scenes. In contrast, the visual quality of our method on out-of-domain datasets remains high. + +# 4.1.2. Reconstruction Efficiency + +In addition to rendering quality, we also compare the reconstruction efficiency of our method with baseline models. Fig. 5 illustrates a plot of reconstruction quality, measured by the average PSNR reported in Tab. 2, versus ef + +
PF GRE10K [38] (Source Domain)DL3DV [16]MVIImgNet [33]ScanNet [5]ACID [17]
PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑
pixelSplat [2]×✓22.640.210.7519.430.390.5019.220.490.6027.560.190.8229.930.140.84
MVSplat [3]×✓23.540.130.9017.840.360.4516.290.500.5326.250.190.8228.830.140.86
CF-GS [9]✓×23.460.240.7519.930.310.6226.330.330.8822.880.420.7628.160.190.81
StreamGS✓✓22.420.170.8320.540.240.6425.050.310.7928.430.160.8628.500.150.84
+ +Table 2. Quantitative comparison with existing state-of-art methods on Novel View Synthesis of monocular videos. PF indicates whether the method is pose-free. G indicates whether the methods is generalizable. Our method consistently achieves scores comparable to state-of-the-art methods that are either not pose-free or lack generalizability. In the table, the best result is highlighted in bold. + +
Re10K +(Source Domain)
ScanNet
DL3DV
MVImgNet
Ground TruthStreamGS (Ours)MVSplatpixelSplatCF-3DGS
+ +Figure 3. Qualitative comparison on novel view synthesis. We show the results on both source domain, RE10K [38], and other domains, ScanNet [5], DL3DV [16] and MVImgNet [33]. All generalizable methods are trained only on RE10K and tested on the other datasets. StreamGS outperforms other methods in several challenging scenarios, especially for the out-of-domain data. + +
Coarse Pred (Sec. 3.2)Ada. Refine (Sec. 3.3)ADC (Sec. 3.4)Total
time/sparam/Mtime/sparam/Mtime/sparam/Mtime/sparam/M
StreamGS0.02656.740.081.830.010.040.11658.61
+ +Table 3. Efficiency metrics of each component. + +ficiency, measured by the processing time per frame (FPS). The figure shows that StreamGS achieves second place with a PSNR of 23.1, only 0.05 lower than CF-3DGS [9]. This indicates that our method achieves nearly the same rendering quality as the best model. However, thanks to the feed-forward design, StreamGS is 150 times faster than CF-3DGS, predicting Gaussian primitives for up to 9 frames within one second. Without known camera information, our method involves additional alignment and pose estimation processes across frames, which limits the inference speed, making it slower than MVSplat [3] and pixelSplat [2]. However, according to the scores reported in Tab. 2 and Fig. 5, our method is more generalizable and achieves better rendering quality than these methods. Tab. 3 shows the efficiency of each component of our method. + +# 4.2. Ablation Study + +In this section, we discuss the effectiveness of our main design about joint refinement (in Sec. 3.3) and feed-forward + +ADC module (in Sec 3.4). More ablation studies on framework design can be found in supplementary materials. + +# 4.2.1. Effectiveness of Joint Refinement + +
RE10K [38]ACID [17]
PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑
w/o refine.17.250.320.5817.230.390.48
w/ refine22.420.170.8328.500.150.84
+ +Table 4. Evaluation of effectiveness of joint refinement. + +Joint refinement of cameras and centers of Gaussians plays a crucial role in the success of our method. To validate the effectiveness of joint refinement, we conduct an ablation study by skipping the refinement process during inference. In other words, the poses and intrinsics are directly estimated by Eq. 2 and 3. Tab. 4 shows the quantitative comparison between the two settings. Without joint refinement, Gaussian primitives are cast from shifted origins of the camera with erroneous orientations, and the poses of novel views are also inaccurate, causing the rendered images to be shifted and distorted. This severely deteriorates the rendering quality, as shown in Tab. 4. The PSNR + +![](images/14d46c0f9362784cc8aa6c6afe6e215f9542d48572e4c64aa39a760066dbf1d3.jpg) +Figure 4. Visual comparison of Gaussian reconstruction and novel view synthesis from image streams with ScanNet [5] dataset. Unlike MVSplat [3], which struggles with view aggregation, our results show significantly better visual quality on OOD data. Note that Our StreamGS and MVSplat are both trained with RE10K [38] data, and MVSplat needs predetermined cameras. + +![](images/73a111e40a52f66e6e9fff9424d2b99821b4549bc2fed74909574c80fc369c22.jpg) +Figure 5. Reconstruction speed measured by frames processed per second (FPS). The x-axis is log-scaled for the better visualization. of the images rendered with direct estimation decreases by $23.06\%$ and $38.62\%$ on the RE10K [38] and ACID [17] datasets, respectively. + +# 4.2.2. Effectiveness of Gaussian Merging Process + +
MVImgNet [33]ACID [17]
Compress. Ratio ↑PSNR ↑Compress. Ratio ↑PSNR ↑
w/o merge.1.0025.701.0029.21
merge all1.5825.051.6828.07
+ +Table 5. Evaluation of the efficiency improvements of the Gaussian merging process and its impact on rendering quality. + +During the feed-forward ADC, StreamGS prunes pixel-aligned Gaussians through a merging process. To evaluate the memory efficiency improvement and its impact on rendering quality, we compare the average number of Gaussians per frame and PSNR before and after the merging process. We define the compression ratio of Gaussians during the merging process as the ratio of the average number of + +Gaussians per frame, i.e., $H \times W$ , to that after the merging process. The metrics are reported in Tab. 5. The results demonstrate that the merging process can prune Gaussians per frame by $36.71\%$ and $40.48\%$ on MVImgNet [33] and ACID [17], respectively. Meanwhile, the PSNR scores after the merging process only slightly decrease by $2.53\%$ and $3.90\%$ , respectively. This ablation study demonstrates that the designed Gaussian merging process efficiently reduces memory usage during reconstruction and rendering, with a negligible impact on reconstruction quality. + +# 5. Conclusion + +We propose a novel and holistic generalizable pose-free reconstruction pipeline named StreamGS, dedicated to the online reconstruction of endless unposed image streams, such as monocular videos. To the best of our knowledge, our method is the first generalizable model capable of predicting Gaussians corresponding to the input stream in a feedforward manner, without relying on known poses and intrinsics. Compared to pose-free but optimization-based methods, our method achieves comparable reconstruction quality while reducing the learning time to within several milliseconds, avoiding optimization steps. Compared to other generalizable methods, StreamGS eliminates the dependence on poses and intrinsics and manages to reconstruct more accurate scenes on out-of-domain datasets, demonstrating better domain generalizability. + +Limitations. Although our method runs fast, joint refinement process still includes additional time costs, making it slower than MVSplat [3]. Moreover, our approach encounters common reconstruction challenges, including textureless regions and long sequences. + +# References + +[1] Wenjing Bian, Zirui Wang, Kejie Li, Jia-Wang Bian, and Victor Adrian Prisacariu. Nope-nerf: Optimising neural radiance field with no pose prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4160-4169, 2023. 2, 3 +[2] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 2, 3, 5, 6, 7 +[3] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. arXiv preprint arXiv:2403.14627, 2024. 2, 3, 5, 6, 7, 8 +[4] Zezhou Cheng, Carlos Esteves, Varun Jampani, Abhishek Kar, Subhransu Maji, and Ameesh Makadia. Lu-nerf: Scene and pose estimation by synchronizing local unposed nerfs. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18312–18321, 2023. 3 +[5] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 6, 7, 8 +[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6 +[7] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.20309, 2024. 2 +[8] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 5 +[9] Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A. Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20796-20805, 2024. 2, 3, 6, 7 +[10] Yiming Gao, Yan-Pei Cao, and Ying Shan. Surfelnerf: Neural surfel radiance fields for online photorealistic reconstruction of indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 108-118, 2023. 1, 3 +[11] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 2 +[12] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 3, 5, 6 + +[13] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[14] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Ep n p: An accurate o (n) solution to the p np problem. International journal of computer vision, 81:155-166, 2009. 5 +[15] Vincent Leroy, Yohann Cabon, and Jerome Revaud. Grounding image matching in 3d with mast3r, 2024. 2 +[16] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. D13dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024. 6, 7 +[17] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In ICCV, 2021. 6, 7, 8 +[18] Tianqi Liu, Guangcong Wang, Shoukang Hu, Liao Shen, Xinyi Ye, Yuhang Zang, Zhiguo Cao, Wei Li, and Ziwei Liu. Mvsgaussian: Fast generalizable gaussian splattering reconstruction from multi-view stereo. In European Conference on Computer Vision, pages 37-53. Springer, 2025. 3 +[19] Andreas Meuleman, Yu-Lun Liu, Chen Gao, Jia-Bin Huang, Changil Kim, Min H Kim, and Johannes Kopf. Progressively optimized local radiance fields for robust view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16539-16548, 2023. 3 +[20] Jongmin Park, Minh-Quan Viet Bui, Juan Luis Gonzalez Bello, Jaeho Moon, Jihyong Oh, and Munchurl Kim. Splinegs: Robust motion-adaptive spline for real-time dynamic 3d gaussians from monocular video. arXiv preprint arXiv:2412.09982, 2024. 2 +[21] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 1, 3 +[22] Qiuhong Shen, Zike Wu, Xuanyu Yi, Pan Zhou, Hanwang Zhang, Shuicheng Yan, and Xinchao Wang. Gamba: Marry gaussian splatting with mamba for single view 3d reconstruction. arXiv preprint arXiv:2403.18795, 2024. 3 +[23] Liang Song, Guangming Wang, Jiuming Liu, Zhenyang Fu, Yanzi Miao, et al. Sc-nerf: Self-correcting neural radiance field with sparse views. arXiv preprint arXiv:2309.05028, 2023. 2 +[24] Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10208-10217, 2024. 2, 5 +[25] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In European Conference on Computer Vision, pages 1-18. Springer, 2025. 3, 5 +[26] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2 + +[27] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 1, 2, 4 +[28] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6 +[29] Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. Grm: Large gaussian reconstruction model for efficient 3d reconstruction and generation. arXiv preprint arXiv:2403.14621, 2024. 3 +[30] Chi Yan, Delin Qu, Dan Xu, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Gs-slam: Dense visual slam with 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19595-19604, 2024. 1, 2, 3 +[31] Xuanyu Yi, Zike Wu, Qiuhong Shen, Qingshan Xu, Pan Zhou, Joo-Hwee Lim, Shuicheng Yan, Xinchao Wang, and Hanwang Zhang. Mvgamba: Unify 3d content generation as state space sequence modeling. arXiv preprint arXiv:2406.06367, 2024. 3 +[32] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4578-4587, 2021. 3 +[33] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Tianyou Liang, Guanying Chen, Shuguang Cui, and Xiaoguang Han. Mvimgnet: A large-scale dataset of multi-view images. In CVPR, 2023. 6, 7, 8 +[34] Vladimir Yugay, Yue Li, Theo Gevers, and Martin R Oswald. Gaussian-slam: Photo-realistic dense slam with gaussian splatting. arXiv preprint arXiv:2312.10070, 2023. 2, 3 +[35] Haojun Zhang, Yuan Yao, and Xuefeng Yan. Improved end-to-end multilevel nerf-based dense rgb-d slam. In Chinese Conference on Pattern Recognition and Computer Vision (PRCV), pages 132-146. Springer, 2024. 2, 3 +[36] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In European Conference on Computer Vision, pages 1-19. Springer, 2025. 3 +[37] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6 +[38] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. ACM Trans. Graph., 37, 2018. 6, 7, 8 +[39] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12786-12796, 2022. 3 + +[40] Zihan Zhu, Songyou Peng, Viktor Larsson, Zhaopeng Cui, Martin R Oswald, Andreas Geiger, and Marc Pollefeys. Nicer-slam: Neural implicit scene encoding for rgb slam. In 2024 International Conference on 3D Vision (3DV), pages 42-52. IEEE, 2024. 2, 3 +[41] Zi-Xin Zou, Shi-Sheng Huang, Yan-Pei Cao, Tai-Jiang Mu, Ying Shan, and Hongbo Fu. Mononeuralfusion: Online monocular neural 3d reconstruction with geometric priors. arXiv preprint arXiv:2209.15153, 2022. 2, 3 +[42] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10324-10335, 2024. 3 \ No newline at end of file diff --git a/2025/StreamGS_ Online Generalizable Gaussian Splatting Reconstruction for Unposed Image Streams/images.zip b/2025/StreamGS_ Online Generalizable Gaussian Splatting Reconstruction for Unposed Image Streams/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..986a6cf0c0a74d7598d15c2f11f57aff48ff65eb --- /dev/null +++ b/2025/StreamGS_ Online Generalizable Gaussian Splatting Reconstruction for Unposed Image Streams/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:58c8ef7c6cd1410bd9b69739f7d7ef477c5391236295d1dda65a16a5d9ba40b0 +size 522406 diff --git a/2025/StreamGS_ Online Generalizable Gaussian Splatting Reconstruction for Unposed Image Streams/layout.json b/2025/StreamGS_ Online Generalizable Gaussian Splatting Reconstruction for Unposed Image Streams/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c08e9df6cd0e4c24953def2619653073aa95af6a --- /dev/null +++ b/2025/StreamGS_ Online Generalizable Gaussian Splatting Reconstruction for Unposed Image Streams/layout.json @@ -0,0 +1,8054 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 94, + 102, + 516, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 102, + 516, + 140 + ], + "spans": [ + { + "bbox": [ + 94, + 102, + 516, + 140 + ], + "type": "text", + "content": "StreamGS: Online Generalizable Gaussian Splitting Reconstruction for Unposed Image Streams" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 159, + 499, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 159, + 499, + 232 + ], + "spans": [ + { + "bbox": [ + 111, + 159, + 499, + 232 + ], + "type": "text", + "content": "Yang Li" + }, + { + "bbox": [ + 111, + 159, + 499, + 232 + ], + "type": "inline_equation", + "content": "^{1,4,*}" + }, + { + "bbox": [ + 111, + 159, + 499, + 232 + ], + "type": "text", + "content": " Jinglu Wang" + }, + { + "bbox": [ + 111, + 159, + 499, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 159, + 499, + 232 + ], + "type": "text", + "content": " Lei Chu" + }, + { + "bbox": [ + 111, + 159, + 499, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 159, + 499, + 232 + ], + "type": "text", + "content": " Xiao Li" + }, + { + "bbox": [ + 111, + 159, + 499, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 159, + 499, + 232 + ], + "type": "text", + "content": " Shiu-Hong Kao" + }, + { + "bbox": [ + 111, + 159, + 499, + 232 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 111, + 159, + 499, + 232 + ], + "type": "text", + "content": " Ying-Cong Chen" + }, + { + "bbox": [ + 111, + 159, + 499, + 232 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 111, + 159, + 499, + 232 + ], + "type": "text", + "content": " Yan Lu" + }, + { + "bbox": [ + 111, + 159, + 499, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 159, + 499, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 159, + 499, + 232 + ], + "type": "text", + "content": "Microsoft Research Asia " + }, + { + "bbox": [ + 111, + 159, + 499, + 232 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 111, + 159, + 499, + 232 + ], + "type": "text", + "content": "HKUST " + }, + { + "bbox": [ + 111, + 159, + 499, + 232 + ], + "type": "inline_equation", + "content": "{}^{3}" + }, + { + "bbox": [ + 111, + 159, + 499, + 232 + ], + "type": "text", + "content": "HKUST(GZ) " + }, + { + "bbox": [ + 111, + 159, + 499, + 232 + ], + "type": "inline_equation", + "content": "{}^{4}" + }, + { + "bbox": [ + 111, + 159, + 499, + 232 + ], + "type": "text", + "content": "Rutgers University yangliaftermath@gmail.com skao@cse.ust.hk yingcongchen@ust.hk {jinglu.wang, leichu, li.xiao, yanlu}@microsoft.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 285, + 296, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 296, + 597 + ], + "type": "text", + "content": "The advent of 3D Gaussian Splatting (3DGS) has advanced 3D scene reconstruction and novel view synthesis. With the growing interest of interactive applications that need immediate feedback, online 3DGS reconstruction in real-time is in high demand. However, none of existing methods yet meet the demand due to three main challenges: the absence of predetermined camera parameters, the need for generalizable 3DGS optimization, and the necessity of reducing redundancy. We propose StreamGS, an online generalizable 3DGS reconstruction method for unposed image streams, which progressively transform image streams to 3D Gaussian streams by predicting and aggregating per-frame Gaussians. Our method overcomes the limitation of the initial point reconstruction [27] in tackling out-of-domain (OOD) issues by introducing a content adaptive refinement. The refinement enhances cross-frame consistency by establishing reliable pixel correspondences between adjacent frames. Such correspondences further aid in merging redundant Gaussians through cross-frame feature aggregation. The density of Gaussians is thereby reduced, empowering online reconstruction by significantly lowering computational and memory costs. Extensive experiments on diverse datasets have demonstrated that StreamGS achieves quality on par with optimization-based approaches but does so 150 times faster, and exhibits superior generalizability in handling OOD scenes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 623, + 135, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 623, + 135, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 623, + 135, + 635 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 643, + 296, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 643, + 296, + 692 + ], + "spans": [ + { + "bbox": [ + 55, + 643, + 296, + 692 + ], + "type": "text", + "content": "The field of 3D Scene reconstruction [10, 30] for novel view synthesis from image streams has gained increasing attention, due to its significance in enabling interactive applications that offer users instant feedback. In this context, the" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 321, + 259, + 549, + 533 + ], + "blocks": [ + { + "bbox": [ + 321, + 259, + 549, + 533 + ], + "lines": [ + { + "bbox": [ + 321, + 259, + 549, + 533 + ], + "spans": [ + { + "bbox": [ + 321, + 259, + 549, + 533 + ], + "type": "image", + "image_path": "a2731129e2b61360219e70938d881361d034a47e89913f3e31c904ee8fdc7188.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 541, + 555, + 596 + ], + "lines": [ + { + "bbox": [ + 313, + 541, + 555, + 596 + ], + "spans": [ + { + "bbox": [ + 313, + 541, + 555, + 596 + ], + "type": "text", + "content": "Figure 1. The proposed StreamGS efficiently transforms image streams into Gaussian streams by progressively reconstructing and aggregating per-frame 3D Gaussians. We show our reconstructed 3DGS (visualized as points) alongside estimated camera poses (in blue), and synthesized novel views." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 617, + 555, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 666 + ], + "type": "text", + "content": "advent of 3D Gaussian Splatting (3DGS) [12] marks a major advancement in high-quality, real-time rendering. This progress highlights the urgency for efficient, on-the-fly generation of 3DGS from image streams." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": "Nonetheless, online 3DGS reconstruction faces unique challenges. (1) Unknown camera. The conventional preprocessing using Structure from Motion (SfM) [21] for camera estimation is impractical for real-time streaming." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 215, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 215, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 215, + 713 + ], + "type": "text", + "content": "*Work done during the internship at MSRA." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25841" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": "This is due to the absence of the full image set and the time-consuming computation. Thus, there is a need for methods that can operate without pre-determined cameras. (2) Generalizability. 3DGS reconstruction requires multiteration optimization, impractical for online applications due to the need for all images in advance. This restricts the development of a generalizable method that can process image streams in a feed-forward manner. (3) Redundancy. The significant overlap between frames leads to high redundancy when reconstructing Gaussians individually for each frame, increasing the streaming process's resource intensity." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 206, + 294, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 206, + 294, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 294, + 313 + ], + "type": "text", + "content": "Scene reconstruction without known cameras has been explored in SLAM-based [30, 34, 35, 40, 41] and NeRF-based [1, 23] methods. Only a few 3DGS-based methods are discussed [9, 20]. These methods primarily consider camera poses as learnable parameters and optimized alongside Gaussians through the iterative optimization. Yet, this optimization-driven approach for aligning each frame considerably increases the reconstruction time, rendering it impractical in the online scenario." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 316, + 295, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 295, + 554 + ], + "type": "text", + "content": "Recently, generalizable 3DGS reconstruction has been investigated for sparse views [2, 3, 24]. These approaches transform pixels into Gaussians, whose parameters are decoded from images via 2D encoder-decoder networks. They achieve the 3DGS reconstruction for each image in a single feed-forward pass. However, these generalizable approaches are primarily designed for monocular or binocular settings, suited to sparse-view inputs. In addition, multi-view methods [2, 3] typically infer Gaussian centers through stereo matching, highly dependent on known cameras, limiting their applicability for online scenarios with numerous unposed images. Furthermore, they combine multi-view 3DGS sets by simply uniting them, which overlooks cross-view alignment and overlaps, leading to misalignment and redundancy issues within image streams. The recent method [7] performs Gaussian downsampling in 3D space to reduce redundancy, but it necessitates traversing and processing Gaussians in 3D grids, which is also time-consuming and is highly dependent on the grid resolution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": "Emerging models like DUSt3R [27] and MASt3R [15] enable sparse-view geometry reconstruction by simultaneously predicting 3D points and estimating camera parameters. These advances pave the way for more efficient, feedforward, pose-free 3D reconstruction methods. A straightforward approach to generalize 3DGS reconstruction is to add a Gaussian predictor to the DUSt3R-like framework. However, this introduces several challenges. These models usually require datasets with ground truth 3D geometry, which may not always be available. Additionally, applying pretrained models to out-of-domain (OOD) data can result in inaccurate pose and 3D point estimations. Moreover, generating 3D points for each frame individually causes re" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "dundancy due to overlapping adjacent frames, potentially leading to ghosting artifacts from pose estimation errors." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 97, + 553, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 97, + 553, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 97, + 553, + 382 + ], + "type": "text", + "content": "In this paper, we introduce StreamGS, a novel pipeline for online, generalizable 3DGS reconstruction from unposed image streams. StreamGS aims to progressively construct and update the 3DGS representation of the scene frame-by-frame, in a feed-forward manner, as illustrated in Fig. 1. We leverage the pretrained DUSt3R to initially predict 3D point for the current frame using the previous frame as a reference. However, this initialization may encounter inaccuracies due to OOD issues. To mitigate this, we capitalize the insight that adjacent frames offer sufficient correspondences to refine the reconstruction. Unlike DUSt3R using predicted 3D points to establish correspondences, we adopt content-adaptive descriptors for more reliable matching, allowing for the adaptive refinement of the reconstruction by enhancing consistency between adjacent views. Furthermore, such correspondences help to prune redundant pixel-aligned Gaussians. Correlated pixel-wise features across frames are effectively aggregated, removing duplicates and achieving adaptive density control. Finally, we decode Gaussians from such aggregated features. StreamGS is adept at predicting and integrating Gaussians for the current frame into the existing Gaussian set seamlessly with a feed-forward pass. In summary, our contribution is three-fold." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 383, + 553, + 562 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 383, + 553, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 383, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 314, + 383, + 553, + 418 + ], + "type": "text", + "content": "- We introduce a novel pipeline for the online, generalizable reconstruction of image streams without requiring camera parameters, marking a first in this field." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 419, + 553, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 419, + 553, + 478 + ], + "spans": [ + { + "bbox": [ + 315, + 419, + 553, + 478 + ], + "type": "text", + "content": "- The proposed adaptive refinement enhance cross-frame consistency of 3DGS reconstruction, and the adaptive density control mechanism minimizes adjacent-view redundancy, thereby highly reducing computational costs in online reconstruction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 479, + 553, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 479, + 553, + 562 + ], + "spans": [ + { + "bbox": [ + 315, + 479, + 553, + 562 + ], + "type": "text", + "content": "- Upon evaluation across diverse datasets, our method achieves high novel view synthesis quality comparable to the optimization-based method [9] but with " + }, + { + "bbox": [ + 315, + 479, + 553, + 562 + ], + "type": "inline_equation", + "content": "150\\mathrm{x}" + }, + { + "bbox": [ + 315, + 479, + 553, + 562 + ], + "type": "text", + "content": " faster reconstruction speed. Additionally, our method outperforms existing pose-dependent generalizable 3DGS methods in handling OOD scenes, showing superior generalizability." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 573, + 406, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 573, + 406, + 586 + ], + "spans": [ + { + "bbox": [ + 314, + 573, + 406, + 586 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "Generalizable 3D Gaussian Splitting. Many recent studies aim to propose generalizable 3D-GS methods capable of predicting Gaussians within a single feed-forward pass. These works can be classified into two main categories: single-view reconstruction and multi-view reconstruction. Single-view reconstruction does not involve pose estimation as there are no multi-view constraints. Inspired by the insight from LRM [11] that large transformer-based [26] backbone networks can learn 3D priors from large-scale 3D data, the potential of predicting Gaussians from a" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25842" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "single image in a single feed-forward pass has been comprehensively explored. Numerous feed-forward models have been proposed, such as GRM [29], TriplaneGS [42], and GMamba [22]. However, these methods are primarily not applicable to multi-view scenarios as they always assume canonical poses." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 145, + 298, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 145, + 298, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 145, + 298, + 407 + ], + "type": "text", + "content": "Concurrently, many works focusing on multi-view inputs follow a similar paradigm, such as GS-LRM [36], LGM [25], and MVGMamba [31]. They concatenate the input images with camera embeddings like Plücker rays to facilitate the network in learning the proper fusion of Gaussians from multi-view inputs. However, these large 3D backbone networks mostly perform well only on synthetic objects due to the shortage of large-scale scene-level 3D data in the real world. Referring to generalizable NeRF methods like pixelNeRF [32], other research turns to multi-view stereo (MVS) matching to locate or initialize the centers of Gaussians, with other attributes decoded using a lightweight 2D encoder. Representative works with this design include MVSGaussian [18], pixelSplat [2], and MVSplat [3]. However, both camera ray embedding in large transformer-based models and stereo matching rely on known poses and intrinsicities of each input view. Another main limitation of these methods is that they focus on sparse-view inputs. PixelSplat [2] and MVSplat [3] only support up to three views. Therefore, existing generalizable 3D-GS models cannot address the problem of feed-forward reconstruction from endless image streams." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 422, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 651 + ], + "type": "text", + "content": "Pose-free 3D Gaussian Splatting. Recently, many works have aimed to eliminate the need for Structure-from-Motion (SfM) preprocessing steps using COLMAP [21] software. Following the design of previous pose-free NeRF methods like NoPe-NeRF [1], Lu-NeRF [4], and localRF [19], CF-3DGS [9] introduces depth priors into the optimization of 3D-GS and performs progressive reconstruction. As each new image arrives, CF-3DGS optimizes both the pose and 3D Gaussians of the input image based on its depths, aiming to align the 3DGS from the new image with the preceding reconstruction. However, it still relies on known camera intrinsics. CF-3DGS is also not robust, as the accuracy of depth priors significantly impacts its reconstruction quality, limiting its application to common scenes. Moreover, it depends on thousands of optimization steps for each view, significantly extending the reconstruction time for each scene. Compared to generalizable 3D-GS models, current pose-free methods are so inefficient that they cannot be applied to image streams with a large number of frames." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "Online 3D reconstruction of image streams. Online 3D reconstruction has been extensively studied in the field of SLAM [30, 34, 35, 39-41]. However, these methods typically involve additional information and most leverage SDF" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": "and NeRF as scene representations. NICE-SLAM [39] uses RGB-D streams as input, with the reconstructed scene represented by NeRF. NICER-SLAM [40] relies on geometric priors, including surface normals and depths. Surfel-NeRF [10] focuses on the novel view synthesis quality of online reconstruction with RGB streams, but it requires the poses and intrinsics of frames. Gaussian-SLAM [34] reconstructs the scene using 3D-GS, but it also relies on RGB-D streams." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 190, + 375, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 190, + 375, + 201 + ], + "spans": [ + { + "bbox": [ + 314, + 190, + 375, + 201 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 210, + 555, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 210, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 210, + 555, + 281 + ], + "type": "text", + "content": "Given a sequence of unposed images over time, our objective is to progressively reconstruct the 3D Gaussian Splating (3DGS) representation in an online manner. Specifically, at each timestamp " + }, + { + "bbox": [ + 313, + 210, + 555, + 281 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 210, + 555, + 281 + ], + "type": "text", + "content": ", the goal is to derive the 3DGS " + }, + { + "bbox": [ + 313, + 210, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^t" + }, + { + "bbox": [ + 313, + 210, + 555, + 281 + ], + "type": "text", + "content": ", which encapsulates the 3D scene aggregated from the images " + }, + { + "bbox": [ + 313, + 210, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^t = \\{\\mathbf{I}^i\\}_{i=1}^t" + }, + { + "bbox": [ + 313, + 210, + 555, + 281 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 282, + 556, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 556, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 556, + 474 + ], + "type": "text", + "content": "Fig. 2 illustrates our overall framework. In order to make reconstruction efficient with limited computational resources, we employ an incremental construction strategy. At each time step " + }, + { + "bbox": [ + 313, + 282, + 556, + 474 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 282, + 556, + 474 + ], + "type": "text", + "content": ", we focus on generating the 3DGS " + }, + { + "bbox": [ + 313, + 282, + 556, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{G}^t" + }, + { + "bbox": [ + 313, + 282, + 556, + 474 + ], + "type": "text", + "content": " of the current frame and merge it with the previous accumulated reconstruction " + }, + { + "bbox": [ + 313, + 282, + 556, + 474 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{t-1}" + }, + { + "bbox": [ + 313, + 282, + 556, + 474 + ], + "type": "text", + "content": " to obtain the full reconstruction " + }, + { + "bbox": [ + 313, + 282, + 556, + 474 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^t" + }, + { + "bbox": [ + 313, + 282, + 556, + 474 + ], + "type": "text", + "content": " of the current time step. Specifically, with current frame " + }, + { + "bbox": [ + 313, + 282, + 556, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^t" + }, + { + "bbox": [ + 313, + 282, + 556, + 474 + ], + "type": "text", + "content": ", we use " + }, + { + "bbox": [ + 313, + 282, + 556, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{t-1}" + }, + { + "bbox": [ + 313, + 282, + 556, + 474 + ], + "type": "text", + "content": " as reference frame and estimate the point maps and cameras of each using an initial reconstruction module. With newly established matches between current and reference views, we further refine the quality of the points and the cameras in the adaptive refinement module. Finally, we generate " + }, + { + "bbox": [ + 313, + 282, + 556, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{G}^t" + }, + { + "bbox": [ + 313, + 282, + 556, + 474 + ], + "type": "text", + "content": " using the refined points and features of " + }, + { + "bbox": [ + 313, + 282, + 556, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^t" + }, + { + "bbox": [ + 313, + 282, + 556, + 474 + ], + "type": "text", + "content": ", and merge it with previous reconstructions according to the established matches, achieving the feedforward adaptive density control (ADC)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 479, + 400, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 400, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 400, + 491 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 497, + 556, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 556, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 556, + 606 + ], + "type": "text", + "content": "Gaussian splatting. Previous work [12] represents a scene or object using a set of Gaussian distributions. Specifically each gaussian primitive could be denoted as " + }, + { + "bbox": [ + 313, + 497, + 556, + 606 + ], + "type": "inline_equation", + "content": "G(x;\\mu ,\\Sigma) = e^{-\\frac{1}{2} (x - \\mu)^T\\Sigma^{-1}(x - \\mu)}" + }, + { + "bbox": [ + 313, + 497, + 556, + 606 + ], + "type": "text", + "content": " and the covariance " + }, + { + "bbox": [ + 313, + 497, + 556, + 606 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 313, + 497, + 556, + 606 + ], + "type": "text", + "content": " is decomposed into the rotation matrix " + }, + { + "bbox": [ + 313, + 497, + 556, + 606 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 497, + 556, + 606 + ], + "type": "text", + "content": " and scaling matrix " + }, + { + "bbox": [ + 313, + 497, + 556, + 606 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 497, + 556, + 606 + ], + "type": "text", + "content": " to ensure the positive semi-definiteness during optimization, that is " + }, + { + "bbox": [ + 313, + 497, + 556, + 606 + ], + "type": "inline_equation", + "content": "\\Sigma = RSS^T R^T" + }, + { + "bbox": [ + 313, + 497, + 556, + 606 + ], + "type": "text", + "content": ". The view-dependent color of the appearance is represented by a set of spherical harmonics (SH) coefficients and opacity value " + }, + { + "bbox": [ + 313, + 497, + 556, + 606 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 497, + 556, + 606 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 613, + 488, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 488, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 488, + 624 + ], + "type": "text", + "content": "3.2. Initial Two-view Reconstruction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "Given the current frame " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^t" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": " and a reference frame " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{t'}" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": ", we use a coarse predictor " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\phi_{3D}" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": " and estimate the point map " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{t|t'}" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": " of the current frame under the local coordinate system of the reference frame together with its corresponding confidence map " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{C}^{t|t'}" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": ", where the superscript " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\cdot |t'" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": " indicates that the local coordinate system adapts to that of " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{t'}" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": ". Intuitively, the predicted point map stores the 3D point coordinate that the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25843" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 553, + 259 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 553, + 259 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 553, + 259 + ], + "type": "image", + "image_path": "fe2c56a6263ac031dabbc6ba493b884753d04e3d39a3df14e5e387effe744e8c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 266, + 555, + 323 + ], + "lines": [ + { + "bbox": [ + 54, + 266, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 555, + 323 + ], + "type": "text", + "content": "Figure 2. Method overview. Our StreamGS progressively reconstruct and aggregate 3D Gaussians from the unposed image stream. Given the adjacent image pair " + }, + { + "bbox": [ + 54, + 266, + 555, + 323 + ], + "type": "inline_equation", + "content": "(\\mathbf{I}^{t - 1},\\mathbf{I}^t)" + }, + { + "bbox": [ + 54, + 266, + 555, + 323 + ], + "type": "text", + "content": ", we first perform the initial reconstruction that predicts pixel-wise 3D points with their features and coarse camera poses, using a pretrained coarse predictor. Since the coarse predictions may suffer from OOD issues, we refine both the camera poses and 3D positions by establishing new point-wise correspondences. We aggregate cross-frame image and 3D features by warping and merging to reduce redundancy. Finally we decode the aggregated features to Gaussian primitives." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 328, + 296, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 296, + 376 + ], + "type": "text", + "content": "pixel unprojected to in, and the confidence maps measure the certainty of point maps at each pixel, reflecting a prior-based estimation of reconstruction accuracy and difficulty. Formally, we define" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 385, + 295, + 402 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 385, + 295, + 402 + ], + "spans": [ + { + "bbox": [ + 88, + 385, + 295, + 402 + ], + "type": "interline_equation", + "content": "\\left(\\mathbf {X} ^ {t \\mid t ^ {\\prime}}, \\mathbf {X} ^ {t ^ {\\prime} \\mid t ^ {\\prime}}, \\mathbf {C} ^ {t \\mid t ^ {\\prime}}, \\mathbf {C} ^ {t ^ {\\prime} \\mid t ^ {\\prime}}\\right) = \\phi_ {3 D} \\left(\\mathbf {I} ^ {t}, \\mathbf {I} ^ {t ^ {\\prime}}\\right). \\tag {1}", + "image_path": "af72cf4532b3273880383d51a0fda723e3ab7ed471f851d0a860e439db1958fa.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 405, + 296, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 405, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 296, + 514 + ], + "type": "text", + "content": "In order to help align the current local-coordinate pointmap with the global coordinate, we also get the output in the local coordinate system of " + }, + { + "bbox": [ + 55, + 405, + 296, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^t" + }, + { + "bbox": [ + 55, + 405, + 296, + 514 + ], + "type": "text", + "content": " by computing " + }, + { + "bbox": [ + 55, + 405, + 296, + 514 + ], + "type": "inline_equation", + "content": "\\phi_{3D}(\\mathbf{I}^{t'},\\mathbf{I}^t)" + }, + { + "bbox": [ + 55, + 405, + 296, + 514 + ], + "type": "text", + "content": ". With these predicted point maps, we could further estimate the camera matrix " + }, + { + "bbox": [ + 55, + 405, + 296, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = \\mathbf{K}[\\mathbf{R}|\\mathbf{t}]" + }, + { + "bbox": [ + 55, + 405, + 296, + 514 + ], + "type": "text", + "content": ", composed of its intrinsic parameters " + }, + { + "bbox": [ + 55, + 405, + 296, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 55, + 405, + 296, + 514 + ], + "type": "text", + "content": ", and its extrinsic parameters, the rotation matrix " + }, + { + "bbox": [ + 55, + 405, + 296, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 55, + 405, + 296, + 514 + ], + "type": "text", + "content": " and the translation vector " + }, + { + "bbox": [ + 55, + 405, + 296, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 55, + 405, + 296, + 514 + ], + "type": "text", + "content": ". Specifically, we assume that the principle points " + }, + { + "bbox": [ + 55, + 405, + 296, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 405, + 296, + 514 + ], + "type": "text", + "content": " are centered and pixels are squares and have" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 82, + 521, + 295, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 521, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 82, + 521, + 295, + 548 + ], + "type": "interline_equation", + "content": "\\hat {f} ^ {t} = \\arg \\min _ {f} \\Sigma_ {\\mathbf {p} \\in \\mathbf {I} ^ {t}} \\| \\mathbf {p} - \\mathbf {c} - f \\frac {\\left(\\mathbf {x} _ {\\mathbf {p}} , \\mathbf {y} _ {\\mathbf {p}}\\right)}{\\mathbf {z} _ {\\mathbf {p}}} \\|, \\tag {2}", + "image_path": "6e29d6fb9f6b20398375196568e63d0776b3eee1bca8b130c3ab15ede7006355.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 555, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 296, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 555, + 296, + 605 + ], + "type": "inline_equation", + "content": "(\\mathbf{x_p},\\mathbf{y_p},\\mathbf{z_p})\\in \\mathbf{X}^{t|t}" + }, + { + "bbox": [ + 55, + 555, + 296, + 605 + ], + "type": "text", + "content": " is the 3D point coordinate that the pixel " + }, + { + "bbox": [ + 55, + 555, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 55, + 555, + 296, + 605 + ], + "type": "text", + "content": " unprojected to. Moreover, we approximate relative pose " + }, + { + "bbox": [ + 55, + 555, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^t = [\\mathbf{R},\\mathbf{t}]" + }, + { + "bbox": [ + 55, + 555, + 296, + 605 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 555, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^t" + }, + { + "bbox": [ + 55, + 555, + 296, + 605 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 555, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{t'}" + }, + { + "bbox": [ + 55, + 555, + 296, + 605 + ], + "type": "text", + "content": " by solving the following points registration problem:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 611, + 294, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 611, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 61, + 611, + 294, + 633 + ], + "type": "interline_equation", + "content": "\\left[ \\mathbf {R} ^ {t} \\mid \\mathbf {t} ^ {t} \\right] = \\arg \\min _ {s, \\mathbf {R}, \\mathbf {t}} \\sum_ {\\mathbf {p} \\in \\mathbf {I} ^ {t - 1}} \\mathbf {C} ^ {t} (\\mathbf {p}) \\| s \\left(\\mathbf {R} \\mathbf {X} ^ {t ^ {\\prime} | t ^ {\\prime}} (\\mathbf {p}) + \\mathbf {t}\\right) - \\mathbf {X} ^ {t ^ {\\prime} | t} (\\mathbf {p}) \\| ^ {2}, \\tag {3}", + "image_path": "585b4106d90b707337a4cc75c034ab1017045ac962f47d39fe28c6eb00693e95.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 85, + 634, + 151, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 634, + 151, + 646 + ], + "spans": [ + { + "bbox": [ + 85, + 634, + 151, + 646 + ], + "type": "interline_equation", + "content": "\\mathbf {C} ^ {t} = \\mathbf {C} ^ {t ^ {\\prime} | t ^ {\\prime}} \\odot \\mathbf {C} ^ {t ^ {\\prime} | t},", + "image_path": "d3f600d50600a8d8e0d4aa9088d5d1877becbd5572a97cd75c67961909e3f39f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 654, + 292, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 292, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 292, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 654, + 292, + 665 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 654, + 292, + 665 + ], + "type": "text", + "content": " is the Hadamard product and " + }, + { + "bbox": [ + 55, + 654, + 292, + 665 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 654, + 292, + 665 + ], + "type": "text", + "content": " is the scale factor." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "In our implementation, we leverage DUSt3R [27] as the coarse predictor due to its effective pretraining on dedicated 3D scene datasets and its efficiency in reconstructing 3D points. For each time step " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", we process the pair" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "inline_equation", + "content": "(\\mathbf{I}^{t - 1},\\mathbf{I}^t)" + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "text", + "content": " to derive the 3D points " + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "inline_equation", + "content": "(\\mathbf{X}^{t|t - 1},\\mathbf{X}^{t - 1|t - 1})" + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "text", + "content": " corresponding to both frames referenced in the coordinate frame " + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "text", + "content": ", the reversed pair " + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "inline_equation", + "content": "(\\mathbf{I}^t,\\mathbf{I}^{t - 1})" + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "text", + "content": " to obtain the same points " + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "inline_equation", + "content": "(\\mathbf{X}^{t - 1|t},\\mathbf{X}^{t|t})" + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "text", + "content": " but in the coordinate frame of " + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "text", + "content": ", which will be reused at the next timestamp to boost efficiency. For simplicity, and without loss of generality, the following discussion will focus on the reconstruction of the image pair " + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "inline_equation", + "content": "(\\mathbf{I}^{t - 1},\\mathbf{I}^t)" + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "text", + "content": ". Note that training " + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "inline_equation", + "content": "\\phi_{3D}" + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "text", + "content": " requires 3D geometric supervision, which may not be available in our monocular video input scenario. Thus, we leverage the pretrained " + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "inline_equation", + "content": "\\phi_{3D}" + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "text", + "content": " from [27]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 468, + 436, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 436, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 436, + 480 + ], + "type": "text", + "content": "3.3. Adaptive Refinement" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 485, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 554, + 544 + ], + "type": "text", + "content": "We note that the initial reconstruction quality is compromised due to the OOD challenge as the coarse predictor is frozen. This observation motivates us to enhance reconstruction through content adaptation, with the goal of adaptively refining both poses and 3D reconstructions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 544, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 628 + ], + "type": "text", + "content": "We perform the adaptive refinement based on establishing new robust matches between adjacent frames. We employ a matching head, " + }, + { + "bbox": [ + 313, + 544, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\phi_{\\text{match}}" + }, + { + "bbox": [ + 313, + 544, + 554, + 628 + ], + "type": "text", + "content": ", to extract local 3D features, denoted as " + }, + { + "bbox": [ + 313, + 544, + 554, + 628 + ], + "type": "inline_equation", + "content": "(\\mathbf{F}_{3D}^{t-1}, \\mathbf{F}_{3D}^t) \\in \\mathbb{R}^{H \\times W \\times d}" + }, + { + "bbox": [ + 313, + 544, + 554, + 628 + ], + "type": "text", + "content": ", from the consecutive image pair " + }, + { + "bbox": [ + 313, + 544, + 554, + 628 + ], + "type": "inline_equation", + "content": "(\\mathbf{I}^{t-1}, \\mathbf{I}^t)" + }, + { + "bbox": [ + 313, + 544, + 554, + 628 + ], + "type": "text", + "content": ". The matches between the two images can be established through nearest reciprocal (NN) searching, satisfying the following condition:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 319, + 636, + 553, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 636, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 319, + 636, + 553, + 668 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {M} ^ {t - 1, t} = \\left\\{i _ {k} \\leftrightarrow j _ {k} \\mid i _ {k} = \\mathrm {N N} (j _ {k}) \\text {a n d} j _ {k} = \\mathrm {N N} (i _ {k}) \\right\\} _ {k = 1} ^ {N}, \\\\ \\text {s . t .} \\operatorname {N N} \\left(i _ {k}\\right) = \\underset {0 \\leq j _ {k} \\leq H \\times W} {\\arg \\min } | 1 - \\cos < \\mathbf {F} _ {3 D, i _ {k}} ^ {t - 1}, \\mathbf {F} _ {3 D, j _ {k}} ^ {t} > |, \\tag {4} \\\\ \\end{array}", + "image_path": "fea7bef30cfe2a00a454f3984b80cd4acac2f32212f3e6e89f22b2f3c3f12948.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "i_k, j_k" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": " are pixel index in image " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{t-1}" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^t" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": " respectively, and the measurement of feature distance is cosine similarity. With the correspondences found, a residual transform" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25844" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "\\Delta = [\\Delta \\mathbf{R}, \\Delta \\mathbf{t}]" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": " can be re-estimated following Eq. (3) by only taking matched points into account. Then we apply the residual transform to the pointmap " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{t,t-1}" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": " to retain refined " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{X}}^{t|t-1}" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": ", and the pose of " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^t" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": " is updated to " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{P}}^t" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": " by performing PnP-RANSAC [8, 14] on 3D-2D correspondences derived from matches." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 166, + 296, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 166, + 296, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 166, + 296, + 202 + ], + "type": "text", + "content": "Gaussian decoding. We directly predict the other parameters of 3D Gaussians at each pixel with a light-weight decoder " + }, + { + "bbox": [ + 55, + 166, + 296, + 202 + ], + "type": "inline_equation", + "content": "\\phi_{GS}" + }, + { + "bbox": [ + 55, + 166, + 296, + 202 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 131, + 213, + 296, + 232 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 213, + 296, + 232 + ], + "spans": [ + { + "bbox": [ + 131, + 213, + 296, + 232 + ], + "type": "interline_equation", + "content": "\\mathbf {G} ^ {i} = \\left[ \\mathbf {q} ^ {i}, \\mathbf {s} ^ {i}, \\alpha^ {i}, \\mathbf {c} ^ {i} \\right] = \\phi_ {G S} (\\mathbf {F} _ {g s}), \\tag {5}", + "image_path": "0d5cdf98847eba436de7472eea2dac8dbf4bc6db685f56b80d56b8a6d7d59f84.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 229, + 270, + 243 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 229, + 270, + 243 + ], + "spans": [ + { + "bbox": [ + 82, + 229, + 270, + 243 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {G S} ^ {i} = \\mathbf {F} _ {2 D} ^ {i} \\oplus \\mathbf {X} ^ {i} \\oplus \\mathbf {F} _ {3 D} ^ {i}, \\quad \\mathbf {F} _ {2 D} = \\phi_ {2 D} (\\mathbf {I} ^ {i}),", + "image_path": "bbf3eb413a7318f4379fc7770c5bc245a52345d37fbaafbc6b553fd6fb2d0b62.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 255, + 296, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 411 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 255, + 296, + 411 + ], + "type": "inline_equation", + "content": "i = \\{t - 1,t\\}" + }, + { + "bbox": [ + 55, + 255, + 296, + 411 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 55, + 255, + 296, + 411 + ], + "type": "text", + "content": " denotes the channel-wise concatenation, " + }, + { + "bbox": [ + 55, + 255, + 296, + 411 + ], + "type": "inline_equation", + "content": "\\phi_{2D}" + }, + { + "bbox": [ + 55, + 255, + 296, + 411 + ], + "type": "text", + "content": " denotes the 2D image feature extractor, " + }, + { + "bbox": [ + 55, + 255, + 296, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{q}^i\\in \\mathbb{R}^{H\\times W\\times 4}" + }, + { + "bbox": [ + 55, + 255, + 296, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 255, + 296, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{s}^i\\in \\mathbb{R}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 55, + 255, + 296, + 411 + ], + "type": "text", + "content": " represent rotation quaternions and scales of pixel-aligned Gaussians at image " + }, + { + "bbox": [ + 55, + 255, + 296, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^i" + }, + { + "bbox": [ + 55, + 255, + 296, + 411 + ], + "type": "text", + "content": " . We incorporate an additional image feature extractor because the coarse predictor is frozen and cannot be trained by our monocular video setting. Extracting new image features is essential for decoding Gaussians, especially for texture-related properties. Experiments show the importance of the image feature extractor. Then covariance matrix is built with " + }, + { + "bbox": [ + 55, + 255, + 296, + 411 + ], + "type": "inline_equation", + "content": "\\Sigma^i = \\mathbf{R}(\\mathbf{q}^i)\\mathbf{ss}^{\\mathrm{T}}\\mathbf{R}(\\mathbf{q}^i)^{\\mathrm{T}}" + }, + { + "bbox": [ + 55, + 255, + 296, + 411 + ], + "type": "text", + "content": " Note that " + }, + { + "bbox": [ + 55, + 255, + 296, + 411 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{G}}" + }, + { + "bbox": [ + 55, + 255, + 296, + 411 + ], + "type": "text", + "content": " is not the final Gaussians since it needs to be merged into the previous Gaussian set following Sec. 3.4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 422, + 173, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 173, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 173, + 433 + ], + "type": "text", + "content": "3.4. Feed-Forward ADC" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 440, + 296, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 440, + 296, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 440, + 296, + 620 + ], + "type": "text", + "content": "With pixel-aligned Gaussian parameters " + }, + { + "bbox": [ + 55, + 440, + 296, + 620 + ], + "type": "inline_equation", + "content": "\\mathbf{G}^t" + }, + { + "bbox": [ + 55, + 440, + 296, + 620 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 440, + 296, + 620 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 440, + 296, + 620 + ], + "type": "text", + "content": " images, previous methods [2, 3, 24, 25] always naively take the union of Gaussians in all images as the final prediction, i.e., " + }, + { + "bbox": [ + 55, + 440, + 296, + 620 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = \\bigcup_{t=1}^{T} \\mathbf{G}^i" + }, + { + "bbox": [ + 55, + 440, + 296, + 620 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 440, + 296, + 620 + ], + "type": "inline_equation", + "content": "|\\mathcal{G}| = T \\times H \\times W" + }, + { + "bbox": [ + 55, + 440, + 296, + 620 + ], + "type": "text", + "content": ". However, this approach is both memory-intensive and inefficient in rendering, particularly when dealing with the continuous input of video frames during online reconstruction. Our key observation is that the matched Gaussian pairs in neighboring frames are excessive and prunable since they consistently share similar attributes in shape and color, and are closely distributed, which is validated in Tab. 1. It is noted that we have already acquired dense pixel-wise matches between neighboring frames from Eq. (4). Therefore, we propose a novel feed-forward Adaptive Density Control strategy based on revisiting these dense correspondences." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": "Feature aggregation. The primary advantage of pixel-based correspondences is that they are able to convert the computationally intensive 3D Gaussian aggregation process into a more efficient 2D pixel-wise one. This significantly enhances computational efficiency. Initially, the feature of Gaussian parameters " + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{gs}^{t}" + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": " of the frame " + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^t" + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": " can be aligned to" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 535, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 535, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 535, + 85 + ], + "type": "text", + "content": "the previous frame " + }, + { + "bbox": [ + 313, + 72, + 535, + 85 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{t - 1}" + }, + { + "bbox": [ + 313, + 72, + 535, + 85 + ], + "type": "text", + "content": " using the following wrapping:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 342, + 89, + 553, + 122 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 89, + 553, + 122 + ], + "spans": [ + { + "bbox": [ + 342, + 89, + 553, + 122 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {G S} ^ {t \\mid t - 1} (j) = \\left\\{ \\begin{array}{l l} \\mathbf {F} _ {G S} ^ {t} (k) & \\text {i f} (j, k) \\in \\overline {{\\mathcal {M}}} ^ {t - 1, t} \\\\ \\mathbf {F} _ {G S} ^ {t - 1} (j) & \\text {e l s e ,} \\end{array} \\right. \\tag {6}", + "image_path": "9268740cc97ec4d77a9755a5507947277e9fe1f2f3e196c93235b5712b0bfb39.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 128, + 555, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 128, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 128, + 555, + 262 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 128, + 555, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{GS}^{t|t-1}" + }, + { + "bbox": [ + 313, + 128, + 555, + 262 + ], + "type": "text", + "content": " denotes the feature of Gaussian primitives in the frame " + }, + { + "bbox": [ + 313, + 128, + 555, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^t" + }, + { + "bbox": [ + 313, + 128, + 555, + 262 + ], + "type": "text", + "content": " aligned to " + }, + { + "bbox": [ + 313, + 128, + 555, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{t-1}" + }, + { + "bbox": [ + 313, + 128, + 555, + 262 + ], + "type": "text", + "content": ". Instead of using the raw correspondences set " + }, + { + "bbox": [ + 313, + 128, + 555, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{t-1,t}" + }, + { + "bbox": [ + 313, + 128, + 555, + 262 + ], + "type": "text", + "content": " in Eq. (4), we use the extended set " + }, + { + "bbox": [ + 313, + 128, + 555, + 262 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{M}}^{t-1,t}" + }, + { + "bbox": [ + 313, + 128, + 555, + 262 + ], + "type": "text", + "content": ", which includes matches of neighboring pixels such as " + }, + { + "bbox": [ + 313, + 128, + 555, + 262 + ], + "type": "inline_equation", + "content": "(i+1,j+1)" + }, + { + "bbox": [ + 313, + 128, + 555, + 262 + ], + "type": "text", + "content": " in addition to the initial " + }, + { + "bbox": [ + 313, + 128, + 555, + 262 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 313, + 128, + 555, + 262 + ], + "type": "text", + "content": ". It serves as a type of anti-aliasing technique to reduce void pixels within the wrapped feature map. As for the unmatched pixels, we simply replicate the corresponding feature vector within " + }, + { + "bbox": [ + 313, + 128, + 555, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{gs}^{t-1}" + }, + { + "bbox": [ + 313, + 128, + 555, + 262 + ], + "type": "text", + "content": ". With the aligned feature, the Gaussian feature of two frames can be merged by modifying Eq. (5), taking the form:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 326, + 267, + 553, + 283 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 267, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 326, + 267, + 553, + 283 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {G}} ^ {t \\mid t - 1} = \\left[ \\hat {\\mathbf {q}} ^ {t}, \\hat {\\mathbf {s}} ^ {t}, \\hat {\\alpha} ^ {t}, \\hat {\\mathbf {c}} ^ {t} \\right] = \\phi_ {M G} \\left(\\mathbf {F} _ {G S} ^ {t \\mid t - 1} \\oplus \\mathbf {F} _ {G S} ^ {t - 1}\\right), \\tag {7}", + "image_path": "2f74b34ace2368edb43b33b36b7ab41d9f5d54f12444562045759ed1da0be7b3.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 289, + 555, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 555, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 555, + 371 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 289, + 555, + 371 + ], + "type": "inline_equation", + "content": "\\phi_{MG}" + }, + { + "bbox": [ + 313, + 289, + 555, + 371 + ], + "type": "text", + "content": " denotes the MergeNet that simply consists of two convolutional layers, which merges features and decodes them to Gaussian primitives. In this way, every matched pair of Gaussians between " + }, + { + "bbox": [ + 313, + 289, + 555, + 371 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^t" + }, + { + "bbox": [ + 313, + 289, + 555, + 371 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 289, + 555, + 371 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{t - 1}" + }, + { + "bbox": [ + 313, + 289, + 555, + 371 + ], + "type": "text", + "content": " is aggregated into a single one, highly reducing the number of Gaussians. The final aggregated Gaussian set is " + }, + { + "bbox": [ + 313, + 289, + 555, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^t = \\mathcal{G}^{t - 1}\\cup \\mathbf{G}^{t|t - 1}" + }, + { + "bbox": [ + 313, + 289, + 555, + 371 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 373, + 554, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 554, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 554, + 431 + ], + "type": "text", + "content": "Without gradient back-propagation of rendering loss in the original 3DGS paper [12], our ADC process runs exceptional fast. In terms of the input group, the final prediction of Gaussian primitives " + }, + { + "bbox": [ + 313, + 373, + 554, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^t" + }, + { + "bbox": [ + 313, + 373, + 554, + 431 + ], + "type": "text", + "content": " consists of the merged Gaussians and the unmatched Gaussians at each frame." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 437, + 408, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 408, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 408, + 449 + ], + "type": "text", + "content": "3.5. Loss Functions" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 455, + 554, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 554, + 491 + ], + "type": "text", + "content": "The optimization of the 2D feature extractor " + }, + { + "bbox": [ + 313, + 455, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\phi_{2D}" + }, + { + "bbox": [ + 313, + 455, + 554, + 491 + ], + "type": "text", + "content": ", the Gaussian decoder network, " + }, + { + "bbox": [ + 313, + 455, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\phi_{GS}" + }, + { + "bbox": [ + 313, + 455, + 554, + 491 + ], + "type": "text", + "content": " involves both a rendering loss function and a reconstruction loss function:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 495, + 553, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 495, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 320, + 495, + 553, + 521 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} \\left(\\mathbf {I} ^ {i}, \\hat {\\mathbf {I}} ^ {i}, \\hat {\\mathbf {I}} _ {M} ^ {i}\\right) = \\mathcal {L} _ {\\text {r e n d e r}} \\left(\\mathbf {I} ^ {i}, \\hat {\\mathbf {I}} ^ {i}\\right) + \\mathcal {L} _ {\\text {r e c o n}} \\left(\\hat {\\mathbf {I}} ^ {i}, \\hat {\\mathbf {I}} _ {M} ^ {i}\\right) \\tag {8} \\\\ = \\| \\mathbf {I} ^ {i} - \\hat {\\mathbf {I}} ^ {i} \\| _ {2} + \\lambda \\| \\mathbf {I} ^ {i} - \\hat {\\mathbf {I}} ^ {i} \\| _ {\\mathrm {L P I P S}} + \\| \\hat {\\mathbf {I}} _ {\\mathrm {M}} ^ {i} - \\hat {\\mathbf {I}} ^ {i} \\| _ {2}, \\\\ \\end{array}", + "image_path": "d12db7a6ef7f2cb52229352b8512a82a7c358f24e01e29203b3b7ef9bbe8f62f.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 528, + 555, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 555, + 647 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 528, + 555, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^i" + }, + { + "bbox": [ + 313, + 528, + 555, + 647 + ], + "type": "text", + "content": " is the ground truth frame, " + }, + { + "bbox": [ + 313, + 528, + 555, + 647 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}^i" + }, + { + "bbox": [ + 313, + 528, + 555, + 647 + ], + "type": "text", + "content": " is the rendered image with full Gaussian primitives before the merge process, and " + }, + { + "bbox": [ + 313, + 528, + 555, + 647 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_{\\mathrm{M}}^{i}" + }, + { + "bbox": [ + 313, + 528, + 555, + 647 + ], + "type": "text", + "content": " is the rendered image with merged Gaussians. Due to the lightweight nature of the two networks, the algorithm converges quickly within thousands of steps. The reconstruction loss term facilitates the merge network in fusing the pixel-aligned Gaussian parameters from different frames, aiming to render the same frame as before the merge process but with significantly fewer Gaussian primitives." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 658, + 395, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 395, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 395, + 670 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "StreamGS operates on an image stream of a scene, jointly predicting the corresponding poses and Gaussians in a feedforward manner. To assess its performance, we evaluate our" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25845" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 70, + 288, + 111 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 288, + 111 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 288, + 111 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 288, + 111 + ], + "type": "table", + "html": "
mean ↓Opa. ↓Rot. ↓Scale ↓SH ↓
Random Pick0.410.103.312.470.27
Matched Pairs0.180.043.061.280.04
", + "image_path": "3773a140c5269d1cf04d021af47224f0b9a46b889946cf7a6c9f11769c20b429.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 120, + 295, + 166 + ], + "lines": [ + { + "bbox": [ + 55, + 120, + 295, + 166 + ], + "spans": [ + { + "bbox": [ + 55, + 120, + 295, + 166 + ], + "type": "text", + "content": "Table 1. Similarities of attributes between matched GS across adjacent frames on RE10K. Random Pick refers to randomly picking two GS from two frames separately, while Matched Pairs refers to our matched GS defined in Eq. (4)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 171, + 295, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 295, + 230 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 295, + 230 + ], + "type": "text", + "content": "method on the task of novel view synthesis from monocular videos, as detailed in Sec. 4.1. Additionally, we validate the effectiveness of the proposed alignment module and the efficiency of the Gaussians merge process, as described in Sec. 4.2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 231, + 295, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 231, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 295, + 410 + ], + "type": "text", + "content": "Baselines and datasets. To the best of our knowledge, StreamGS is the first method to reconstruct unposed videos in a feed-forward manner. Consequently, we compare our method separately with pose-free 3DGS works and generalizable splatting methods, including pixelSplat [2], MVSplat [3], and CF-3DGS [9]. PixelSplat and MVSplat are representative methods of generalizable Gaussian Splatting, but they both rely on known camera poses and intrinsics. In contrast, CF-3DGS is a pose-free method but requires optimization loops to align poses and Gaussians. For a comprehensive comparison, we evaluate the methods on large-scale datasets with diverse scenes, including RE10K [38], ACID [17], ScanNet [5], DL3DV [16], and MVImgNet [33]. Each dataset comprises monocular video sequences with per-frame camera pose annotations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 410, + 295, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 295, + 650 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 295, + 650 + ], + "type": "text", + "content": "Implementation details. Both the 2D image feature extractor " + }, + { + "bbox": [ + 55, + 410, + 295, + 650 + ], + "type": "inline_equation", + "content": "\\phi_{2D}" + }, + { + "bbox": [ + 55, + 410, + 295, + 650 + ], + "type": "text", + "content": " and the MergeNet " + }, + { + "bbox": [ + 55, + 410, + 295, + 650 + ], + "type": "inline_equation", + "content": "\\phi_{GS}" + }, + { + "bbox": [ + 55, + 410, + 295, + 650 + ], + "type": "text", + "content": " are double-layer convolutional networks. For fairness, StreamGS and other generalizable methods are trained on the identical training split of the RE10K dataset using the Adam [13] optimizer with the same learning rate of " + }, + { + "bbox": [ + 55, + 410, + 295, + 650 + ], + "type": "inline_equation", + "content": "2\\times 10^{-4}" + }, + { + "bbox": [ + 55, + 410, + 295, + 650 + ], + "type": "text", + "content": " and a cosine scheduler. The parameter " + }, + { + "bbox": [ + 55, + 410, + 295, + 650 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 410, + 295, + 650 + ], + "type": "text", + "content": " in Eq. 8 is set to 0.05. All methods are trained for 30K iterations and tested on a single NVIDIA Tesla A100 80GB GPU. The images are resized to " + }, + { + "bbox": [ + 55, + 410, + 295, + 650 + ], + "type": "inline_equation", + "content": "224\\times 224" + }, + { + "bbox": [ + 55, + 410, + 295, + 650 + ], + "type": "text", + "content": " and the batch size is set to 14. For the non-generalizable method CF-3DGS, we follow its original setting [9]. Note that CF-3DGS is evaluated only on a subset of the full test set due to its low reconstruction efficiency, as shown in Fig. 5. Unlike pose-dependent methods, CF-3DGS and our method require poses of novel views for rendering. CF-3DGS freezes the trained Gaussian model and performs additional optimization steps to learn the poses. Our method, as described in Sec. 3.3, carries out pose alignment and refinement processes to estimate the poses. More details can be referred to supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 657, + 178, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 178, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 178, + 670 + ], + "type": "text", + "content": "4.1. Novel View Synthesis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 674, + 183, + 687 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 674, + 183, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 674, + 183, + 687 + ], + "type": "text", + "content": "4.1.1. Reconstruction Quality" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Quantitative comparison. We compare StreamGS with baseline methods on the quality and efficiency of novel view" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 384 + ], + "type": "text", + "content": "synthesis. Following previous 3D-GS research [12], we report PSNR, SSIM [28], and LPIPS [37] as metrics of reconstruction quality. The quantitative results are shown in Tab. 2. On the source domain RE10K [38], existing state-of-the-art generalizable methods demonstrate competitive scores, with MVSplat even outperforming the optimization-based CF-3DGS. However, their performance degrades significantly on out-of-domain datasets such as DL3DV [16] and MVImgNet [33], as these datasets contain various scene types, including outdoor environments and more complex indoor scenes with different objects and illumination conditions compared to RE10K. Naturally, CF-3DGS, which is based on thousands of optimization steps, performs well on the aforementioned datasets. However, its PSNR slightly decreases on ScanNet due to more irregular camera movements and increased motion blur in the frames, posing a challenge for CF-3DGS in recovering camera poses. Pixel-Splat performs well on ScanNet, thanks to its robust feature extraction backbone trained on ImageNet [6], while MVSplat achieves the lowest score. According to the table, StreamGS consistently achieves scores comparable to all other methods. Its PSNR on MVImgNet and DL3DV is significantly higher than that of PixelSplat and MVSplat, even without given poses and intrinsics, demonstrating our method's superior generalizability on unseen datasets with significant domain gaps." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 391, + 556, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 556, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 556, + 643 + ], + "type": "text", + "content": "Qualitative comparison. A qualitative comparison with state-of-the-art methods is also presented in Fig. 3 and Fig. 4. While all methods demonstrate high-quality novel view rendering on RE10K [38], our method exhibits superior robustness on out-of-domain datasets. Due to the significant domain gap including texture, illumination and camera motion, both MVSplat and pixelSplat fail to predict accurate depths for the printer shown in the second row, resulting in severe floating artifacts in the rendered image. The third row shows the reconstruction of a plaza from DL3DV [16]. Similarly, since outdoor scenes are much less represented in RE10K [38], MVSplat struggles to extract stereo cues from text-less skies and objects with disparate illumination, deteriorating the rendering quality. The final row shows a bench scene in MVImgNet, on which baseline generalizable methods also fails due to the similar domain gap issues. These cases also demonstrate that the generalizability of pixelSplat surpasses that of MVSplat. As the figure shows, CF-3DGS also does not perform well on some outdoor scenes. In contrast, the visual quality of our method on out-of-domain datasets remains high." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 650, + 452, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 650, + 452, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 650, + 452, + 662 + ], + "type": "text", + "content": "4.1.2. Reconstruction Efficiency" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 666, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 715 + ], + "type": "text", + "content": "In addition to rendering quality, we also compare the reconstruction efficiency of our method with baseline models. Fig. 5 illustrates a plot of reconstruction quality, measured by the average PSNR reported in Tab. 2, versus ef" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "25846" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 555, + 148 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 555, + 148 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 555, + 148 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 555, + 148 + ], + "type": "table", + "html": "
PF GRE10K [38] (Source Domain)DL3DV [16]MVIImgNet [33]ScanNet [5]ACID [17]
PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑
pixelSplat [2]×✓22.640.210.7519.430.390.5019.220.490.6027.560.190.8229.930.140.84
MVSplat [3]×✓23.540.130.9017.840.360.4516.290.500.5326.250.190.8228.830.140.86
CF-GS [9]✓×23.460.240.7519.930.310.6226.330.330.8822.880.420.7628.160.190.81
StreamGS✓✓22.420.170.8320.540.240.6425.050.310.7928.430.160.8628.500.150.84
", + "image_path": "1b6fca2ee1ea35e662e47c3319c3b2c4e6f2e86ef6d12dd921e379ee03013178.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 190, + 553, + 398 + ], + "blocks": [ + { + "bbox": [ + 55, + 156, + 555, + 188 + ], + "lines": [ + { + "bbox": [ + 55, + 156, + 555, + 188 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 555, + 188 + ], + "type": "text", + "content": "Table 2. Quantitative comparison with existing state-of-art methods on Novel View Synthesis of monocular videos. PF indicates whether the method is pose-free. G indicates whether the methods is generalizable. Our method consistently achieves scores comparable to state-of-the-art methods that are either not pose-free or lack generalizability. In the table, the best result is highlighted in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 190, + 553, + 398 + ], + "lines": [ + { + "bbox": [ + 56, + 190, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 190, + 553, + 398 + ], + "type": "table", + "html": "
Re10K\n(Source Domain)
ScanNet
DL3DV
MVImgNet
Ground TruthStreamGS (Ours)MVSplatpixelSplatCF-3DGS
", + "image_path": "64fe47016b2b937b89ce2ce319e4501ae392d490663133c0dc5c39db9502ff56.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 61, + 438, + 290, + 475 + ], + "blocks": [ + { + "bbox": [ + 55, + 401, + 555, + 435 + ], + "lines": [ + { + "bbox": [ + 55, + 401, + 555, + 435 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 555, + 435 + ], + "type": "text", + "content": "Figure 3. Qualitative comparison on novel view synthesis. We show the results on both source domain, RE10K [38], and other domains, ScanNet [5], DL3DV [16] and MVImgNet [33]. All generalizable methods are trained only on RE10K and tested on the other datasets. StreamGS outperforms other methods in several challenging scenarios, especially for the out-of-domain data." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 438, + 290, + 475 + ], + "lines": [ + { + "bbox": [ + 61, + 438, + 290, + 475 + ], + "spans": [ + { + "bbox": [ + 61, + 438, + 290, + 475 + ], + "type": "table", + "html": "
Coarse Pred (Sec. 3.2)Ada. Refine (Sec. 3.3)ADC (Sec. 3.4)Total
time/sparam/Mtime/sparam/Mtime/sparam/Mtime/sparam/M
StreamGS0.02656.740.081.830.010.040.11658.61
", + "image_path": "fae2db546b11249465676c712401ba999da91b55c21b3ad23257d30b0228cbda.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 89, + 475, + 261, + 484 + ], + "lines": [ + { + "bbox": [ + 89, + 475, + 261, + 484 + ], + "spans": [ + { + "bbox": [ + 89, + 475, + 261, + 484 + ], + "type": "text", + "content": "Table 3. Efficiency metrics of each component." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 484, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 295, + 662 + ], + "type": "text", + "content": "ficiency, measured by the processing time per frame (FPS). The figure shows that StreamGS achieves second place with a PSNR of 23.1, only 0.05 lower than CF-3DGS [9]. This indicates that our method achieves nearly the same rendering quality as the best model. However, thanks to the feed-forward design, StreamGS is 150 times faster than CF-3DGS, predicting Gaussian primitives for up to 9 frames within one second. Without known camera information, our method involves additional alignment and pose estimation processes across frames, which limits the inference speed, making it slower than MVSplat [3] and pixelSplat [2]. However, according to the scores reported in Tab. 2 and Fig. 5, our method is more generalizable and achieves better rendering quality than these methods. Tab. 3 shows the efficiency of each component of our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 672, + 149, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 149, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 149, + 684 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "In this section, we discuss the effectiveness of our main design about joint refinement (in Sec. 3.3) and feed-forward" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 441, + 555, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 555, + 464 + ], + "type": "text", + "content": "ADC module (in Sec 3.4). More ablation studies on framework design can be found in supplementary materials." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 469, + 484, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 469, + 484, + 480 + ], + "spans": [ + { + "bbox": [ + 314, + 469, + 484, + 480 + ], + "type": "text", + "content": "4.2.1. Effectiveness of Joint Refinement" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 323, + 491, + 548, + 555 + ], + "blocks": [ + { + "bbox": [ + 323, + 491, + 548, + 555 + ], + "lines": [ + { + "bbox": [ + 323, + 491, + 548, + 555 + ], + "spans": [ + { + "bbox": [ + 323, + 491, + 548, + 555 + ], + "type": "table", + "html": "
RE10K [38]ACID [17]
PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑
w/o refine.17.250.320.5817.230.390.48
w/ refine22.420.170.8328.500.150.84
", + "image_path": "7eac9d3a6ab6cfc4fe08b880ff0d86300a5269db0d3d096497df45bc28c0694a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 332, + 558, + 535, + 568 + ], + "lines": [ + { + "bbox": [ + 332, + 558, + 535, + 568 + ], + "spans": [ + { + "bbox": [ + 332, + 558, + 535, + 568 + ], + "type": "text", + "content": "Table 4. Evaluation of effectiveness of joint refinement." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "Joint refinement of cameras and centers of Gaussians plays a crucial role in the success of our method. To validate the effectiveness of joint refinement, we conduct an ablation study by skipping the refinement process during inference. In other words, the poses and intrinsics are directly estimated by Eq. 2 and 3. Tab. 4 shows the quantitative comparison between the two settings. Without joint refinement, Gaussian primitives are cast from shifted origins of the camera with erroneous orientations, and the poses of novel views are also inaccurate, causing the rendered images to be shifted and distorted. This severely deteriorates the rendering quality, as shown in Tab. 4. The PSNR" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25847" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 553, + 264 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 553, + 264 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 553, + 264 + ], + "type": "image", + "image_path": "14d46c0f9362784cc8aa6c6afe6e215f9542d48572e4c64aa39a760066dbf1d3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 270, + 555, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 270, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 555, + 304 + ], + "type": "text", + "content": "Figure 4. Visual comparison of Gaussian reconstruction and novel view synthesis from image streams with ScanNet [5] dataset. Unlike MVSplat [3], which struggles with view aggregation, our results show significantly better visual quality on OOD data. Note that Our StreamGS and MVSplat are both trained with RE10K [38] data, and MVSplat needs predetermined cameras." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 317, + 292, + 447 + ], + "blocks": [ + { + "bbox": [ + 58, + 317, + 292, + 447 + ], + "lines": [ + { + "bbox": [ + 58, + 317, + 292, + 447 + ], + "spans": [ + { + "bbox": [ + 58, + 317, + 292, + 447 + ], + "type": "image", + "image_path": "73a111e40a52f66e6e9fff9424d2b99821b4549bc2fed74909574c80fc369c22.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 449, + 295, + 509 + ], + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 509 + ], + "type": "text", + "content": "Figure 5. Reconstruction speed measured by frames processed per second (FPS). The x-axis is log-scaled for the better visualization. of the images rendered with direct estimation decreases by " + }, + { + "bbox": [ + 55, + 449, + 295, + 509 + ], + "type": "inline_equation", + "content": "23.06\\%" + }, + { + "bbox": [ + 55, + 449, + 295, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 449, + 295, + 509 + ], + "type": "inline_equation", + "content": "38.62\\%" + }, + { + "bbox": [ + 55, + 449, + 295, + 509 + ], + "type": "text", + "content": " on the RE10K [38] and ACID [17] datasets, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 513, + 265, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 513, + 265, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 513, + 265, + 526 + ], + "type": "text", + "content": "4.2.2. Effectiveness of Gaussian Merging Process" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 64, + 536, + 289, + 596 + ], + "blocks": [ + { + "bbox": [ + 64, + 536, + 289, + 596 + ], + "lines": [ + { + "bbox": [ + 64, + 536, + 289, + 596 + ], + "spans": [ + { + "bbox": [ + 64, + 536, + 289, + 596 + ], + "type": "table", + "html": "
MVImgNet [33]ACID [17]
Compress. Ratio ↑PSNR ↑Compress. Ratio ↑PSNR ↑
w/o merge.1.0025.701.0029.21
merge all1.5825.051.6828.07
", + "image_path": "db764eeea26a1c64284b55c6c8460f4015c9743fbe43ad767aa5cc3cbdaf6646.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 295, + 628 + ], + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 628 + ], + "type": "text", + "content": "Table 5. Evaluation of the efficiency improvements of the Gaussian merging process and its impact on rendering quality." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": "During the feed-forward ADC, StreamGS prunes pixel-aligned Gaussians through a merging process. To evaluate the memory efficiency improvement and its impact on rendering quality, we compare the average number of Gaussians per frame and PSNR before and after the merging process. We define the compression ratio of Gaussians during the merging process as the ratio of the average number of" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 309, + 555, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 309, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 555, + 430 + ], + "type": "text", + "content": "Gaussians per frame, i.e., " + }, + { + "bbox": [ + 313, + 309, + 555, + 430 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 313, + 309, + 555, + 430 + ], + "type": "text", + "content": ", to that after the merging process. The metrics are reported in Tab. 5. The results demonstrate that the merging process can prune Gaussians per frame by " + }, + { + "bbox": [ + 313, + 309, + 555, + 430 + ], + "type": "inline_equation", + "content": "36.71\\%" + }, + { + "bbox": [ + 313, + 309, + 555, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 309, + 555, + 430 + ], + "type": "inline_equation", + "content": "40.48\\%" + }, + { + "bbox": [ + 313, + 309, + 555, + 430 + ], + "type": "text", + "content": " on MVImgNet [33] and ACID [17], respectively. Meanwhile, the PSNR scores after the merging process only slightly decrease by " + }, + { + "bbox": [ + 313, + 309, + 555, + 430 + ], + "type": "inline_equation", + "content": "2.53\\%" + }, + { + "bbox": [ + 313, + 309, + 555, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 309, + 555, + 430 + ], + "type": "inline_equation", + "content": "3.90\\%" + }, + { + "bbox": [ + 313, + 309, + 555, + 430 + ], + "type": "text", + "content": ", respectively. This ablation study demonstrates that the designed Gaussian merging process efficiently reduces memory usage during reconstruction and rendering, with a negligible impact on reconstruction quality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 438, + 388, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 438, + 388, + 451 + ], + "spans": [ + { + "bbox": [ + 314, + 438, + 388, + 451 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 460, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 640 + ], + "type": "text", + "content": "We propose a novel and holistic generalizable pose-free reconstruction pipeline named StreamGS, dedicated to the online reconstruction of endless unposed image streams, such as monocular videos. To the best of our knowledge, our method is the first generalizable model capable of predicting Gaussians corresponding to the input stream in a feedforward manner, without relying on known poses and intrinsics. Compared to pose-free but optimization-based methods, our method achieves comparable reconstruction quality while reducing the learning time to within several milliseconds, avoiding optimization steps. Compared to other generalizable methods, StreamGS eliminates the dependence on poses and intrinsics and manages to reconstruct more accurate scenes on out-of-domain datasets, demonstrating better domain generalizability." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": "Limitations. Although our method runs fast, joint refinement process still includes additional time costs, making it slower than MVSplat [3]. Moreover, our approach encounters common reconstruction challenges, including textureless regions and long sequences." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25848" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 145 + ], + "type": "text", + "content": "[1] Wenjing Bian, Zirui Wang, Kejie Li, Jia-Wang Bian, and Victor Adrian Prisacariu. Nope-nerf: Optimising neural radiance field with no pose prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4160-4169, 2023. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 147, + 296, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 147, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 61, + 147, + 296, + 212 + ], + "type": "text", + "content": "[2] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 214, + 294, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 294, + 267 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 294, + 267 + ], + "type": "text", + "content": "[3] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. arXiv preprint arXiv:2403.14627, 2024. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 270, + 295, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 295, + 323 + ], + "type": "text", + "content": "[4] Zezhou Cheng, Carlos Esteves, Varun Jampani, Abhishek Kar, Subhransu Maji, and Ameesh Makadia. Lu-nerf: Scene and pose estimation by synchronizing local unposed nerfs. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18312–18321, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 325, + 295, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 325, + 295, + 367 + ], + "spans": [ + { + "bbox": [ + 62, + 325, + 295, + 367 + ], + "type": "text", + "content": "[5] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 369, + 295, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 295, + 413 + ], + "type": "text", + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 415, + 295, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 295, + 468 + ], + "type": "text", + "content": "[7] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.20309, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 470, + 295, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 470, + 295, + 513 + ], + "spans": [ + { + "bbox": [ + 62, + 470, + 295, + 513 + ], + "type": "text", + "content": "[8] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 514, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 295, + 567 + ], + "type": "text", + "content": "[9] Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A. Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20796-20805, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 570, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 570, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 57, + 570, + 295, + 623 + ], + "type": "text", + "content": "[10] Yiming Gao, Yan-Pei Cao, and Ying Shan. Surfelnerf: Neural surfel radiance fields for online photorealistic reconstruction of indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 108-118, 2023. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 625, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 668 + ], + "type": "text", + "content": "[11] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[12] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 3, 5, 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "text", + "content": "[13] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 97, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 139 + ], + "type": "text", + "content": "[14] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Ep n p: An accurate o (n) solution to the p np problem. International journal of computer vision, 81:155-166, 2009. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 142, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 142, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 142, + 553, + 163 + ], + "type": "text", + "content": "[15] Vincent Leroy, Yohann Cabon, and Jerome Revaud. Grounding image matching in 3d with mast3r, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 165, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 165, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 317, + 165, + 553, + 230 + ], + "type": "text", + "content": "[16] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. D13dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024. 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "type": "text", + "content": "[17] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In ICCV, 2021. 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 277, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 277, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 317, + 277, + 553, + 331 + ], + "type": "text", + "content": "[18] Tianqi Liu, Guangcong Wang, Shoukang Hu, Liao Shen, Xinyi Ye, Yuhang Zang, Zhiguo Cao, Wei Li, and Ziwei Liu. Mvsgaussian: Fast generalizable gaussian splattering reconstruction from multi-view stereo. In European Conference on Computer Vision, pages 37-53. Springer, 2025. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 333, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 333, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 317, + 333, + 553, + 387 + ], + "type": "text", + "content": "[19] Andreas Meuleman, Yu-Lun Liu, Chen Gao, Jia-Bin Huang, Changil Kim, Min H Kim, and Johannes Kopf. Progressively optimized local radiance fields for robust view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16539-16548, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 388, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 553, + 442 + ], + "type": "text", + "content": "[20] Jongmin Park, Minh-Quan Viet Bui, Juan Luis Gonzalez Bello, Jaeho Moon, Jihyong Oh, and Munchurl Kim. Splinegs: Robust motion-adaptive spline for real-time dynamic 3d gaussians from monocular video. arXiv preprint arXiv:2412.09982, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 445, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 487 + ], + "type": "text", + "content": "[21] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 490, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 490, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 490, + 553, + 533 + ], + "type": "text", + "content": "[22] Qiuhong Shen, Zike Wu, Xuanyu Yi, Pan Zhou, Hanwang Zhang, Shuicheng Yan, and Xinchao Wang. Gamba: Marry gaussian splatting with mamba for single view 3d reconstruction. arXiv preprint arXiv:2403.18795, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 535, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 578 + ], + "type": "text", + "content": "[23] Liang Song, Guangming Wang, Jiuming Liu, Zhenyang Fu, Yanzi Miao, et al. Sc-nerf: Self-correcting neural radiance field with sparse views. arXiv preprint arXiv:2309.05028, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 580, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 553, + 633 + ], + "type": "text", + "content": "[24] Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10208-10217, 2024. 2, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "text", + "content": "[25] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In European Conference on Computer Vision, pages 1-18. Springer, 2025. 3, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[26] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25849" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[27] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 1, 2, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "text", + "content": "[28] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "text", + "content": "[29] Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. Grm: Large gaussian reconstruction model for efficient 3d reconstruction and generation. arXiv preprint arXiv:2403.14621, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "type": "text", + "content": "[30] Chi Yan, Delin Qu, Dan Xu, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Gs-slam: Dense visual slam with 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19595-19604, 2024. 1, 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "text", + "content": "[31] Xuanyu Yi, Zike Wu, Qiuhong Shen, Qingshan Xu, Pan Zhou, Joo-Hwee Lim, Shuicheng Yan, Xinchao Wang, and Hanwang Zhang. Mvgamba: Unify 3d content generation as state space sequence modeling. arXiv preprint arXiv:2406.06367, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 365 + ], + "type": "text", + "content": "[32] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4578-4587, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 420 + ], + "type": "text", + "content": "[33] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Tianyou Liang, Guanying Chen, Shuguang Cui, and Xiaoguang Han. Mvimgnet: A large-scale dataset of multi-view images. In CVPR, 2023. 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "type": "text", + "content": "[34] Vladimir Yugay, Yue Li, Theo Gevers, and Martin R Oswald. Gaussian-slam: Photo-realistic dense slam with gaussian splatting. arXiv preprint arXiv:2312.10070, 2023. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 294, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 511 + ], + "type": "text", + "content": "[35] Haojun Zhang, Yuan Yao, and Xuefeng Yan. Improved end-to-end multilevel nerf-based dense rgb-d slam. In Chinese Conference on Pattern Recognition and Computer Vision (PRCV), pages 132-146. Springer, 2024. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 555 + ], + "type": "text", + "content": "[36] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In European Conference on Computer Vision, pages 1-19. Springer, 2025. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 557, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 612 + ], + "type": "text", + "content": "[37] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 613, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 656 + ], + "type": "text", + "content": "[38] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. ACM Trans. Graph., 37, 2018. 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[39] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12786-12796, 2022. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 239 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[40] Zihan Zhu, Songyou Peng, Viktor Larsson, Zhaopeng Cui, Martin R Oswald, Andreas Geiger, and Marc Pollefeys. Nicer-slam: Neural implicit scene encoding for rgb slam. In 2024 International Conference on 3D Vision (3DV), pages 42-52. IEEE, 2024. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "text", + "content": "[41] Zi-Xin Zou, Shi-Sheng Huang, Yan-Pei Cao, Tai-Jiang Mu, Ying Shan, and Hongbo Fu. Mononeuralfusion: Online monocular neural 3d reconstruction with geometric priors. arXiv preprint arXiv:2209.15153, 2022. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 555, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 555, + 239 + ], + "type": "text", + "content": "[42] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10324-10335, 2024. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25850" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/StreamMind_ Unlocking Full Frame Rate Streaming Video Dialogue through Event-Gated Cognition/61e82f55-a3d1-4244-b3cc-22632c8cab6f_content_list.json b/2025/StreamMind_ Unlocking Full Frame Rate Streaming Video Dialogue through Event-Gated Cognition/61e82f55-a3d1-4244-b3cc-22632c8cab6f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..334dd4cdadd9d9de842c6fa48acbddd08466483f --- /dev/null +++ b/2025/StreamMind_ Unlocking Full Frame Rate Streaming Video Dialogue through Event-Gated Cognition/61e82f55-a3d1-4244-b3cc-22632c8cab6f_content_list.json @@ -0,0 +1,1759 @@ +[ + { + "type": "text", + "text": "STREAMMIND: Unlocking Full Frame Rate Streaming Video Dialogue through Event-Gated Cognition", + "text_level": 1, + "bbox": [ + 98, + 128, + 898, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xin Ding $^{1}$ * Hao Wu $^{3}$ † Yifan Yang $^{2}$ Shiqi Jiang $^{2}$ Qianxi Zhang $^{2}$ Donglin Bai $^{2}$ Zhibo Chen $^{1}$ Ting Cao $^{4}$ †", + "bbox": [ + 184, + 200, + 813, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Science and Technology of China 2Microsoft Research 3Nanjing University Institute for AI Industry Research (AIR), Tsinghua University", + "bbox": [ + 120, + 239, + 874, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the rise of real-world human-AI interaction applications, such as AI assistants, the need for Streaming Video Dialogue is critical. To address this need, we introduce STREAMMIND, a video LLM framework that achieves ultra-FPS streaming video processing (100fps on a single A100) and enables proactive, always-on responses in real time, without explicit user intervention.", + "bbox": [ + 89, + 340, + 483, + 446 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To solve the key challenge of the contradiction between linear video streaming speed and quadratic transformer computation cost, we propose a novel perception-cognition interleaving paradigm named \"event-gated LLM invocation\", in contrast to the existing per-time-step LLM invocation. By introducing a Cognition Gate network between the video encoder and the LLM, LLM is only invoked when relevant events occur. To realize the event feature extraction with constant cost, we propose Event-Preserving Feature Extractor (EPFE) based on state-space method, generating a single perception token for spatiotemporal features. These techniques enable the video LLM with full-FPS perception and real-time cognition response.", + "bbox": [ + 88, + 449, + 483, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Experiments on Ego4D and SoccerNet streaming tasks, as well as standard offline benchmarks, demonstrate state-of-the-art performance in both model capability and real-time efficiency, paving the way for ultra-high-FPS applications, such as Game AI Copilot and interactive media. The code and data is available at https://aka.ms/StreamMind.", + "bbox": [ + 89, + 646, + 483, + 737 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 762, + 220, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The advancement of large foundation models is driving an increasing number of real-world human-AI interaction applications, such as AI home companions [3, 41] and human-robot collaboration [55, 64]. These applications", + "bbox": [ + 89, + 787, + 482, + 849 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3adaf8afee610aaa13e2cd3ca26d5864a019fdf770a722f6606ba269a692111a.jpg", + "image_caption": [ + "Existing Works 10~15 FPS on A100 GPU", + "Figure 1. The paradigm of existing per-step LLM invocation (left) versus our event-gated LLM invocation (right) for streaming video dialogue task. Our paradigm boosts frame rate up-to 100 fps on A100. $(R_{i}$ : response, $f_{i}$ : frame, t: time)" + ], + "image_footnote": [], + "bbox": [ + 516, + 306, + 717, + 407 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/35b51e882842b88a1cdd620d9ce8e208cbaae50c948226330f8044f9927ddb7d.jpg", + "image_caption": [ + "Ours Up to 100 FPS on A100 GPU" + ], + "image_footnote": [], + "bbox": [ + 720, + 308, + 921, + 407 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "heavily rely on streaming video understanding capabilities [14, 26, 45, 62, 72], specifically streaming video dialogue [8, 60]. Streaming video dialogue (StreamingVD) aims to continuously perceive incoming video streams and, based on user queries, provide proactive, real-time, and always-on responses without human intervention. For example, an AI companion could respond to user queries like \"Guide me to fix this bicycle\" or \"Let's watch and discuss the soccer game together\".", + "bbox": [ + 511, + 496, + 906, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Streaming video dialogue can be viewed as a generalization of standard offline video dialogue tasks [30, 32, 39, 40, 49, 66, 67], extending video LLMs to broader real-world scenarios. The fundamental challenge of StreamingVD compared to its offline counterpart lies in the requirement for timing alignment between events and responses. That is, as the video progresses, the model is required to (1) proactively decide when to respond and (2) respond in real time before the next event occurs according to user queries, as illustrated in Fig. 2. However, with current techniques, these two requirements are contradictory and cannot be achieved simultaneously.", + "bbox": [ + 511, + 638, + 908, + 820 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The seminal VideoLLM-Online [8] and VideoLLM-MoD [60] introduce a per-step LLM invocation paradigm for StreamingVD, shown in Fig. 1. At each time step, it inputs all past frames and the user query to the LLM, to decide when to respond or remain silent. However,", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was done during Xin Ding's internship at Microsoft Research.", + "bbox": [ + 89, + 862, + 482, + 886 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding Author.", + "bbox": [ + 116, + 887, + 245, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13448", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the $O(n^3)^1$ computational complexity and limited context window are mismatched with the $O(n)$ frame streaming, making real-time response difficult to achieve. More works [14, 45, 62, 72] focus on enhancing the efficiency of offline video dialogue aiming to meet real-time requirements. However, similar to offline approaches, they depend on user queries to trigger responses rather than the model to do so proactively.", + "bbox": [ + 88, + 90, + 480, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To empower StreamingVD, this paper proposes a new perception-cognition interleaving paradigm, named event-gated LLM invocation, shown in Fig. 1. It is inspired by the event perception mechanism of human brain [69], which states that events are key components of human's perception, attention, and memory $^2$ . We therefore propose an event-based selective cognition process. Rather than invoking the LLM at every time step, we introduce a Cognition Gate between the video encoder and LLM. The video encoder continuously perceives visual signals of each frame, and only when query-related events occur, the gate will open to invoke the LLM. Analogous to how a text prompt triggers an LLM, streaming visual signals lack a clear stopping point, unlike a finite text sequence. The cognition gate bridges this gap by acting as a mechanism to identify \"physical world prompts\" for the LLM.", + "bbox": [ + 93, + 214, + 485, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This new paradigm raises two major technical concerns, to achieve both high capability and efficiency. Firstly, existing video encoders in VideoLLMs (e.g., STC in VideoLaMA) only capture local spatio-temporal features. This results in a linear increase in the number of feature tokens for streaming frames and subsequently a quadratic computation for the processing, failing to meet real-time requirements. Secondly, the Cognition Gate must decide to invoke the LLM or not, based on visual features and the user query. However, only considering efficiency and simply attending visual and query tokens (e.g., Cross Attention in Qformer [28]) limits the gate capability to tasks like feature-matching [57, 58] and retrieval [1, 16, 75], struggle to make decisions with deeper semantic understanding.", + "bbox": [ + 89, + 455, + 482, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By addressing the challenges, we realize STREAM-MIND, a novel framework for streamingVD, with the processing speed matching frame rate for the first time. It integrates two techniques. For the video encoder, inspired by the exceptional ability of the state-assisted module [22] in modeling continuous physical signals with constant cost, we propose the Event-Preserving Feature Extractor (EPFE) based on the state-assisted module. It generates a single perception token of event features to the gate, resulting in a constant perception cost. For the Cognition Gate, to leverage the world knowledge of LLM while meeting the real-", + "bbox": [ + 89, + 669, + 482, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "time requirement, we propose a Shallow Layer Transfer method. The Gate is designed to reuse the shallow layers of the LLM and its autoregressive training to maximize the probabilities of “< response/silence >” tokens during training.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To evaluate the performance of the STREAMMIND framework, we adopt evaluation metrics from VideoLLM-Online and a wide range of text generation evaluation metrics. Additionally, we design two novel metrics to comprehensively assess the model's temporal responsiveness. Experiments on real-time Ego4D [21] and SoccerNet [20] demonstrate superior performance across all metrics. Furthermore, our model achieves state-of-the-art results on multiple offline benchmarks, including short- and long-term activity recognition and forecasting on the COIN [53] and Ego4D LTA [24] datasets. Notably, our framework breaks the ultra-FPS streaming video processing bottleneck, enabling streamingVD at up to 100 fps on a single A100 GPU, laying a robust foundation for real-world human-AI interaction applications.", + "bbox": [ + 511, + 167, + 906, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Method", + "text_level": 1, + "bbox": [ + 513, + 400, + 604, + 416 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Task Discussion and Definition", + "text_level": 1, + "bbox": [ + 511, + 425, + 785, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Problem formulation of StreamingVD Given a video stream $\\mathcal{V}^T \\coloneqq [v_1, v_2, \\dots, v_T]$ consisting of $T$ frames and a set of queries $\\mathcal{Q} \\coloneqq \\{q_1, q_2, \\dots, q_N\\}$ . Traditional video dialogue, shown in Fig. 2, provides a response triggered by a user query. For a query $q_i$ given at time step $t_s$ , it responds based on past frames up to time $t_s (1 < t_s \\leq T)$ , regardless of an online or offline scenario. On the other hand, StreamingVD proactively generates responses to a query given at $t_s$ at any appropriate time steps until the query termination at $t_e (t_s < t_e \\leq T)$ . At each time step $t_i (t_s \\leq t_i \\leq t_e)$ , StreamingVD needs to decide whether to generate a response according to the past frames and query, and then generate the response, formulated as follows.", + "bbox": [ + 511, + 446, + 906, + 643 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\max P \\left(\\left[ \\operatorname {R e s} ^ {t _ {i}} \\right] \\mid \\left[ \\operatorname {C t x} ^ {< t _ {i}} \\right], \\left[ F ^ {t _ {i}} \\right]\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 648, + 903, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $[\\mathrm{Res}^{t_i}] \\in \\{[\\mathrm{Txt}^{t_i}], [\\mathrm{EOS}^{t_i}]\\}$ means generate a response at a time step $t_i$ , or remain silence, i.e., [EOS]. $[\\mathrm{Ctx}^{< t_i}]$ denotes the contextual tokens accumulated from past frames and queries before $t_i$ . $[\\mathrm{F}^{t_i}]$ represents the visual features extracted from the video frame at $t_i$ .", + "bbox": [ + 511, + 671, + 906, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Challenges of StreamingVD tasks Firstly, increasing number of frames versus real-time requirement. Different from the offline tasks[17, 19, 30, 63, 68], streaming tasks[8, 14, 26, 45, 62, 72] require time alignment between event and response. Besides only accuracy constraint of offline task, streaming video tasks need to consider both accuracy and timeliness constraints. Offline tasks can trade resources and latency for better quality, while streaming task cannot.", + "bbox": [ + 511, + 763, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1Invoking a quadratic-complexity LLM n times results in cubic complexity.", + "bbox": [ + 89, + 849, + 482, + 875 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "2A formal definition of an event [69]: \"A segment of time at a given location that is conceived by an observer to have a beginning and an end.\"", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13449", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2b7857db5e85331170fbe38d99a83bb0d8bf03235c857757b86617c572cdf93a.jpg", + "image_caption": [ + "Figure 2. Comparison between streaming dialogue (bottom) and traditional dialogue tasks (top). Streaming dialogue proactively generates responses at appropriate time based on user queries and perception, whereas traditional dialogue requires human triggering for each response." + ], + "image_footnote": [], + "bbox": [ + 174, + 93, + 816, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Secondly, proactive response versus real-time requirement. Traditional video dialogue [13, 14, 45, 61, 62, 65] are reactive, responding only at the time of user query. However, real-world dialogue is often without human triggering. The user may lack the specific knowledge to formulate questions. Similarly, embodied AI should anticipate user needs rather than merely reacting to commands. However, proactive response means at every frame the model needs to judge, inceasing the model cost by $\\mathrm{O}(\\mathfrak{n})$ times.", + "bbox": [ + 88, + 390, + 483, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing works Conventional offline Video LLMs [10, 32, 33, 35, 73, 74] typically employ a visual encoder [18, 46, 70] and a projection module [28, 39] to process uniformly sampled frames. The output is concatenated with the tokenized question to form a sequence, which is then passed to an LLM decoder to predict a response. However, this approach is impractical for streaming tasks due to the high computational cost associated with processing a large number of frames $(T)$ . Therefore, works such as [14, 62] have implemented a sliding-window attention mechanism and a KV-cache system to reduce computational overhead, or incorporated a parallel sampling strategy to enhance processing speed and reduce latency. However, these methods do not proactively trigger response when needed. In contrast, Videollm-Online [8] and MOD [60] utilize the LLM decoder to process every frame of streaming videos, resulting in a cubic computation cost. STREAMMIND, on the other hand, has the potential to process streaming videos at 100 fps, which is crucial for enabling interactive video analysis applications.", + "bbox": [ + 91, + 537, + 483, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. The Overview of STREAMMIND", + "text_level": 1, + "bbox": [ + 89, + 848, + 362, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For a streaming video sequence $\\left[v^{t_s} \\leq t_i \\leq t_e\\right]$ , STREAMMIND models the following process as shown in Fig. 3:", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "1. Perception Phase: With continuous streaming video, each incoming frame $v$ is processed through the CLIP [46] encoder to extract spatial features. The features are then input to our proposed EPFE. EPFE updates its hidden state to correlate the frame's features across temporal dimensions and reduce redundant spatial information. This process generates a new perception token $[\\mathsf{F}_{per}^{t_i}]$ and saves this token in the Perception Memory $\\mathcal{M}_{\\mathrm{per}}^{t_i}$ . The perception phase can be formulated as follows:", + "bbox": [ + 511, + 388, + 906, + 525 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\mathrm {F} _ {p e r} ^ {t _ {i}} \\right] = \\operatorname {E P F E} \\left(\\operatorname {C L I P} \\left(v ^ {t _ {i}}\\right), \\mathrm {H} ^ {t _ {i} - 1}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 534, + 906, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbb{H}^{t - 1}$ denotes the hidden state from the previous time.", + "bbox": [ + 511, + 561, + 903, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Cognition Gate Judgment: We introduce Cognition Gate $\\mathcal{G}$ between perception and cognition. The input are user query [Prompt] and the generated perception token $[\\mathsf{F}_{per}^{t_i}]$ at each time step $t_i$ . The Cognition Gate learns to determine whether to trigger the cognition phase at $t_i$ :", + "bbox": [ + 511, + 577, + 905, + 654 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\max P \\left(\\left[ \\operatorname {R e s} ^ {t _ {i}} \\right] \\mid \\mathcal {G} \\left(\\left[ \\text {P r o m p t} \\right], \\left[ F _ {\\text {p e r}} ^ {t _ {i}} \\right]\\right)\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 661, + 903, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $[\\mathrm{Res}^{t_i}] \\in \\{< /\\text{response}>, < /\\text{silence}>\\}$ determines whether to trigger a response or remain silent.", + "bbox": [ + 511, + 688, + 905, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Cognition Phase Upon receiving the trigger signal from the Cognition Gate, STREAMMIND samples perception tokens from the Perception Memory $\\mathcal{M}_{\\mathrm{per}}^{t_i}$ to the Cognition Pooling as the input to LLM to generate response.", + "bbox": [ + 511, + 719, + 905, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Perception Phase", + "text_level": 1, + "bbox": [ + 511, + 787, + 681, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motivation: To match the constant frame rate of streaming videos, ideally the perception phase should be constant computation cost for processing each frame. However, current video encoders generally extract local (dozens of frames) spatiotemporal features with constant cost, and input all the generated perception tokens to a transformer", + "bbox": [ + 511, + 810, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13450", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1eb6a4808df4538648aaa78b2ebe560f1b83446fe1685610cc8a0d9b28a29c2d.jpg", + "image_caption": [ + "Figure 3. STREAMMIND workflow. For each video frame, the perception stage goes through Clip for spatial features and the proposed EPFE for spatiotemporal features, to generate a single perception token saved in perception memory. The Cognition Gate determines whether to invoke LLM only based on the current perception token. If yes (a query-related event occurs), the tokens in perception memory are sampled to cognition pooling as the input for LLM for cognition response. The gate decouples perception and cognition to enable full-FPS perception and event-based cognition." + ], + "image_footnote": [], + "bbox": [ + 112, + 87, + 885, + 265 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "model to extract the long-term temporal relationships. The computational cost is $O(n^{3})$ , obviously the lagger behind the streaming rate.", + "bbox": [ + 88, + 354, + 482, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To solve the issue, we propose Event-Preserving Feature Extractor (EPFE) to replace the projection module used in existing video LLMs (e.g., cross-attention and Q-Former). EPFE is based on the SOTA state-space model, i.e., Selective State Space model [22]. As shown in Fig. 3 (blue box) and Eq. 4, for each frame, EPFE dynamically correlates the spatial features and the past internal states to generate a single perception token $\\left[\\mathbb{F}_{per}^{t_i}\\right]$ as well as update its state. This enables a matched speed of streaming video and perception processing, for full FPS StreamingVD.", + "bbox": [ + 88, + 398, + 482, + 551 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} _ {t + 1} = \\mathbf {A h} _ {t} + \\mathbf {B x} _ {t}, \\quad \\mathbf {y} _ {t} = \\mathbf {C h} _ {t} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 559, + 482, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{A},\\mathbf{B}$ , and $\\mathbf{C}$ are learnable state-space matrices for state, input, and output, respectively. $\\mathbf{x_t}$ is the input frame of the current time step. $\\mathbf{h_t}$ is the internal hidden state. $\\mathbf{y_t}$ is the output, i.e., perception token.", + "bbox": [ + 88, + 583, + 482, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Benefit from the strong temporal modeling capability of Selective SSM, our EPFE with only $56\\mathrm{M}$ parameters can learn long-term event-level spatiotemporal features with constant cost. As Appendix Fig. 1 illustrates, perception tokens generated by EPFE can distinguish different events and noise frames, while maintaining a state of the event. Even after unrelated noise frames, the perception tokens are capable of refocusing on the event.", + "bbox": [ + 88, + 643, + 482, + 765 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.4. Cognition Gate", + "text_level": 1, + "bbox": [ + 89, + 772, + 243, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motivation: Cognition Gate needs to determine whether to invoke LLM in real-time, based on the current perception token and user query. For example, during the soccer match, some user query may only care LLM response for scoring moments, while others may need detailed tactical analysis.", + "bbox": [ + 88, + 794, + 482, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, we find it highly challenging. For real-time consideration, we first employ a simple Transformer block", + "bbox": [ + 88, + 869, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with cross-attention to associate the user query with the visual encoder's output for trigger decisions. However, it struggles to make accurate judgments at the right time step, as shown in Tab.5. We identify two key reasons for this: 1) The Cognition Gate only has access to information up to the current step, lacking a global temporal perspective for decision-making. 2) User queries may not directly correspond to specific video frames in a retrieval-like manner (e.g., guide me to repair a bicycle or cooking).", + "bbox": [ + 511, + 354, + 903, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, the Cognition Gate requires the world knowledge of LLMs. However, due to the size of LLMs, even processing a single perception token as input incurs a high computational cost. Notably, the Gate only needs to generate a response/silence token for a single iteration.", + "bbox": [ + 511, + 494, + 903, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To achieve it, We propose the Shallow Layer Transfer method, which initializes the Gate using the early layers of the LLM and then fine-tunes it in a supervised autoregressive manner with the StreamVD dataset. As shown in Sec. 4.2, we conducted extensive experiments to analyze its impact. The results indicate that the first few layers exhibit slight variations in behavior, allowing for scalability based on requirements. In this work, we use four layers.", + "bbox": [ + 511, + 575, + 906, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.5. The Training Strategy of STREAMMIND", + "text_level": 1, + "bbox": [ + 511, + 722, + 844, + 738 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Model Architecture As illustrated in Figure 3, the STREAMMIND video LLM integrates CLIP [46], EPFE, Cognition Gate, and the LLM. CLIP is the pre-trained spatial feature extractor. EPFE then takes the spatial features as input as well as its internal state to extract spatiotemporal features and generate a perception token. Cognition Gate then evaluates the current perception token and user query to determine whether to invoke LLM. To achieve this, it is needed to train EPFE, Cognition Gate, and also align the representation space of EPFE and LLM, as shown in Fig. 4.", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13451", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Streaming Video Dataset Construction" + ], + "code_body": "Require: Offline video dataset $C = \\{c_i,t_i\\}_{i = 1}^n (c_i$ caption at time $t_i)$ \nEnsure: Dataset $D$ with captions meeting StreamVD requirements and Cognition Gate labels. \n1: Step 1: Preprocessing \n2: $D\\gets []$ , cap $\\leftarrow c_{1}$ ,time $\\leftarrow t_{s}$ \n3: for $i = 2$ to n do \n4: if $c_{i}\\neq$ cap then \n5: Append (cap, time) to $D$ \n6: cap $\\leftarrow c_{i}$ ,time $\\leftarrow t_{i}$ \n7: end if \n8: end for \n9: Append (cap, time) to $D$ \n10: Step 2: Silence-Response Labeling \n11: for each frame $f$ between $d_{i}$ and $d_{i + 1}$ do \n12: Label $f$ withif captioned, else \n13: end for \n14: return $D$", + "bbox": [ + 109, + 125, + 468, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data Preparation Streaming datasets are constructed by processing existing offline datasets. StreamingVD tasks require the model to response as early as possible when relevant events occur while avoid redundant responses. To achieve this, we eliminate repetitive captions from the offline datasets and transform them into streaming datasets using the following methods, detailed in Algorithm 1.", + "bbox": [ + 89, + 444, + 482, + 549 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Preprocessing: Adjacent identical captions are merged, with the timestamp of the first occurrence being recorded as the annotation point for the caption.", + "- Silence-Response Labeling: To train the Cognition Gate, we insert $< / \\text{silence}>$ tokens into video frames between two adjacent captions. These tokens act as indicators to guide the Cognition Gate in generating the $< / \\text{silence}>$ . The initial frame corresponding to captioned events is labeled with the $< / \\text{response}>$ token." + ], + "bbox": [ + 89, + 550, + 482, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Strategy We adopt a two-stage training strategy, as shown in Fig. 4. In contrast to previous typical two-stage training approaches [29, 40, 71], our method uses a dedicated stage for Cognition Gate. In the first stage, the LLM and EPFE are jointly trained using video frames and captions from the streaming dataset, ensuring spatiotemporal feature alignment with the LLM. The second stage trains only the Cognition Gate, which is initialized from the shallow layers of the LLM. This training also follows the LLM autoregressive manner to generate $< / \\text{silence} >$ or $< / \\text{response} >$ tokens, determining whether to invoke the LLM. Since the response label is much sparser than silence (e.g. 310:1), we introduce a balancing weight into", + "bbox": [ + 89, + 704, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the standard Cross Entropy loss during training to balance the labels. Refer to Sec. 3.1 and Appendix Sec. 1 for more training details.", + "bbox": [ + 511, + 90, + 905, + 137 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/739ae2abf9e281cfbd7ecf9e6f650725ee50be11cd644d0abd9ed25a0efaee6a.jpg", + "image_caption": [ + "Figure 4. The two-stage training process of STREAMMIND. In the first stage, the EPFE spatiotemporal extractor and the LLM are jointly trained to align their representations. In the second stage, the Cognition module is trained to generate response/silence tokens. Both stages follow a supervised autoregressive training approach for language modeling." + ], + "image_footnote": [], + "bbox": [ + 516, + 148, + 709, + 244 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1e4c820793b4f84127f077e2a6c5b6bc3df4bdcb771bd1d8e826e59ec4f93c33.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 148, + 905, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3. Experiment", + "text_level": 1, + "bbox": [ + 513, + 363, + 638, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 511, + 387, + 720, + 404 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baseline and Datasets Since existing works focusing on streamingVD [8, 26, 60], except for VideoLLM-Online [8], do not provide open-source data or training methods, to demonstrate our outstanding performance, we re-implemented all previous works (VideoLLM-Online [8] and VideoLLM-MOD [60]) in the StreamingVD domain as baselines, using consistent settings across all experiments.", + "bbox": [ + 511, + 409, + 905, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For dataset selection, since the Ego4D dataset [21] is inherently collected in a streaming manner, containing various queries, real-time responses, and narrations, and has been used as an main evaluation dataset in works [8, 31, 60], we selected it as one of our evaluation datasets. Furthermore, to further demonstrate the effectiveness of our approach on long videos, we also conducted experiments on the SoccerNet-Caption dataset [20]. This dataset includes 471 complete soccer match videos with a total duration of 715 hours, where each video is approximately 45 minutes long and provides commentary at appropriate timestamps.", + "bbox": [ + 511, + 515, + 905, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details: Model training was conducted on videos sampled at 2 fps, using 8 NVIDIA A100 GPUs. Each stage of the training process was conducted for one epoch. To optimize the learning rate, we applied the CosineAnnealingLR scheduler, with learning rates set to $2e - 5$ and $2e - 6$ respectively. Additionally, specific techniques were applied to the design of the Cognition Gate and methods to address the imbalance in the Silence-Response ratio, which are detailed in Section 4 and the supplementary materials.", + "bbox": [ + 511, + 688, + 905, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Evaluate Metrics", + "text_level": 1, + "bbox": [ + 511, + 848, + 681, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of STREAMMIND, we focus on assessing its Timing Alignment Capability and Lan", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13452", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/35e34eefc2f14d46ffb7295708eb3bf6b262d424061ca271ffc5de384f6ae975.jpg", + "table_caption": [ + "Table 1. Online Experiments Results." + ], + "table_footnote": [], + "table_body": "
DataSetMethodTiming Alignment ↑Language Modeling Capability ↑Training Cost
TriggerAccTimValBLUE-1BLUE-4METEORROUGE-L
Ego4dVideoLLM-Online32.34%29.66%66.0135.2531.1263.0624h
VideoLLM-MoD32.36%29.65%65.3435.2130.6563.0214h
Ours43.34%39.73%67.1239.2631.6065.71(11+8)h
SoccerVideoLLM-Online31.25%28.34%75.3664.2350.9281.5712h
VideoLLM-MoD31.24%28.12%74.9664.1850.2481.597h
Ours52.18%47.36%82.7866.7051.4382.04(5+3)h
", + "bbox": [ + 150, + 109, + 848, + 247 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b180a5df8550470273b4296b8853f80373ffcc582cc9bc09be9705d069f4b18e.jpg", + "table_caption": [ + "Table 2. Online Experiments Results. LION-FS was released on arXiv just two days before our submission, but its code was not released. Therefore, we could only obtain the Ego4D results based on its paper." + ], + "table_footnote": [], + "table_body": "
DataSetMethodTimeDiff ↓Fluency ↑PPL ↓Correctiness ↑
Ego4dVideoLLM-Online2.0445.0%2.4348.1%
VideoLLM-MOD2.0445.2%2.4148.9%
LION-FS2.1546.1%2.0952.4%
Ours1.8960.2%2.0277.3%
SoccerVideoLLM-Online15.6246.35%1.7953.5%
VideoLLM-MOD15.6845.34%1.8053.3%
Ours14.0270.35%1.5989.2%
", + "bbox": [ + 96, + 330, + 478, + 444 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ea0093b07efddd34fdf0971ff3ac94448925eec2e04be713771e33f18110cc8c.jpg", + "table_caption": [ + "Table 3. Results on COIN benchmarks(left to right): step recognition, task recognition, next forecasting, procedure forecasting, procedure forecasting with a goal." + ], + "table_footnote": [], + "table_body": "
MethodNot use HT 100MCOINBenchmark Top-1 Accuracy ↑
StepTaskNextProcProg.+
ClipBERT✓30.865.4---
TimeFormer×46.585.334.017.040.1
Paprika×51.085.843.2--
DistantSup×54.190.039.4--
VideoTF×56.591.042.440.246.4
ProcedureVRL×56.990.846.8--
VideoTaskGraph×57.290.540.2--
VideoLLM-online-7B-v1✓59.892.148.147.952.9
VideoLLM-online -8B-v1+✓63.192.749.149.854.1
VideoLLM-MOD✓63.492.849.749.853.3
Ours✓63.793.249.949.854.2
", + "bbox": [ + 94, + 503, + 480, + 645 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1b07ab548ab481fb902ac93d597432b2d85ee43db52883e282ea0325391f49fa.jpg", + "table_caption": [ + "Table 4. Results on Ego4DLTA benchmark, evaluated on public server. ED@Z=20 denotes editing distance for future 20 actions." + ], + "table_footnote": [], + "table_body": "
MethodNot use Ego VLPEnd-to -End?Ego4D LTA ED@Z=20
Verb ↓Noun ↓Action ↓
CLIP✓✓0.7390.7690.941
EgoT2✓✓0.7220.7640.935
I-CVAE✓✓0.7530.7490.931
HierVL✓✓0.7240.7350.928
VideoLLM×✓0.7210.7250.921
VideoLLM-online-7B-v1✓✓0.6970.6980.897
VideoLLM-online-8B-v1+✓✓0.6890.6710.884
VideoLLM-MOD✓✓0.6890.6760.884
Ours✓✓0.6890.6550.881
", + "bbox": [ + 98, + 693, + 478, + 814 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "guage Modeling Capability. Existing evaluation metrics include LM-PPL (Language Modeling Perplexity) and LM-Correctness, which evaluate language modeling accuracy at specific timestamps. Additionally, the Time Difference", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(TimeDiff) metric is used to measure the temporal alignment of responses in an streaming video LLM, while the Fluency metric aims to provide a holistic evaluation of both linguistic quality and temporal consistency.", + "bbox": [ + 511, + 271, + 906, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, these metrics have several limitations:", + "bbox": [ + 529, + 333, + 854, + 348 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- LM-PPL and LM-Correctness focus solely on the exact alignment between model outputs and reference narrations, ignoring aspects such as synonym substitutions, sentence fluency, and the ability to capture critical information.", + "- TimeDiff, as a metric for timing alignment Capability, evaluates temporal alignment only within a single dialogue turn. It does not assess the Cognition Gate's overall decision-making across the entire video stream." + ], + "bbox": [ + 513, + 349, + 906, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To address these limitations, we introduce the following metrics:", + "bbox": [ + 511, + 486, + 905, + 516 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Trigger Accuracy (TriggerAcc): To evaluate whether the streaming video LLM responds at the correct time steps during entire streaming video.", + "- Timing Validity (TimVal): To comprehensively evaluate whether the streaming video LLM consistently makes the correct decisions throughout the streaming video—speaking when necessary and remaining silent when appropriate.", + "- BLUE [43], METEOR [2], ROUGE-L [34]: Widely used evaluation metrics in the field of text generation [6, 9, 38] and image narration [7, 44, 48], which fully consider factors such as synonym matching, stemming, word order, and key semantic expressions. These metrics enable a comprehensive assessment of dialogues in the StreamingVD process." + ], + "bbox": [ + 513, + 517, + 905, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Comprehensive Evaluation on Streaming and Offline Video", + "text_level": 1, + "bbox": [ + 511, + 756, + 905, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Online Experiments: We compared our model with all existing streaming dialogue models [8, 31, 60] in Tab 1 and Tab 2. Under comparable training costs, experiments on real-time Ego4D and SoccerNet demonstrate superior performance across all metrics. Additionally, STREAMMIND can process video streams at a $10 \\times$ higher FPS. This improvement is attributed to the event-level modeling capability.", + "bbox": [ + 511, + 794, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13453", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ity of EPFE and the strong timing alignment capability of the cognition gate in the perception stage, which together provide the LLM in the cognition phase with a concise yet comprehensive representation.", + "bbox": [ + 89, + 90, + 482, + 151 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Offline Experiments: We also leveraged the COIN [53] and Ego4D LTA [24] benchmark to demonstrate STREAM-MIND effectiveness in traditional offline video scenarios, covering six common benchmarks: step recognition, step forecasting, task summarization, procedure forecasting, and procedure forecasting with a goal. As shown in Tab 3 and Tab 4, our approach achieves state-of-the-art performance across all tasks.", + "bbox": [ + 89, + 157, + 482, + 280 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4. Real-time Inference Efficiency", + "text_level": 1, + "bbox": [ + 89, + 287, + 359, + 305 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Standard frame rates for films are typically 24 FPS, which is considered the minimum requirement for smooth motion perception. In contrast, games often operate at much higher frame rates, ranging from 30 FPS to over 100 FPS, to ensure a immersive experience. To validate the real-time performance of STREAMMIND, we conducted inference efficiency evaluations on NVIDIA A100 and H100 GPUs.", + "bbox": [ + 89, + 309, + 482, + 414 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, we compared VideoLLM-Online, VideoLLM-MOD, and STREAMMIND by measuring their actual processing time for 1-second video across input frame rates from 5 to 100 FPS. As shown in Figure 5, a processing time under 1 second indicates real-time performance without disrupting playback.", + "bbox": [ + 89, + 415, + 482, + 505 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results reveal that VideoLLM-Online and VideoLLM-MOD struggle to process video streams beyond 10 FPS, leading to potential latency issues in real-world applications. In contrast, our method achieves true real-time performance, efficiently handling both standard film/TV frame rates and highRefresh gaming scenarios up to 100 FPS, ensuring real-time responses.", + "bbox": [ + 89, + 506, + 482, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. Ablation", + "text_level": 1, + "bbox": [ + 89, + 625, + 191, + 640 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Impact of Silence-Response Sample Imbalance", + "text_level": 1, + "bbox": [ + 89, + 648, + 482, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we explored the imbalance in the Silence-Response ratio within the dataset and derived an empirical balancing formula through experiments. For detailed experimental results, please refer to the supplementary material.", + "bbox": [ + 89, + 670, + 482, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. The Ablation of Cognition Gate", + "text_level": 1, + "bbox": [ + 89, + 739, + 370, + 757 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The Cognition Gate plays a crucial role throughout the perception-cognition interleaving process. We conducted detailed ablation experiments to analyze its impact.", + "bbox": [ + 89, + 761, + 482, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The Architecture of Cognition Gate The structure of the gate directly determines its capabilities. To this end, we constructed multiple classification networks using MLP projectors, Cross-Attention, and Transformers. [54] Additionally, we built a set of autoregressive networks using an", + "bbox": [ + 89, + 824, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/177b0ddf4d7b4d1cbcdbb3dd92ae00c24c8c053d74c13cd4ee0f00aa452170ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 90, + 870, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1d9a738ef34a21e23a4bf4748ba50df552804677ae8aa33db725f0e165ea4397.jpg", + "image_caption": [ + "Figure 5. Comparison of running time for processing 1 second of streaming video frames using different methods on A100 (Top) and H100 (Bottom)." + ], + "image_footnote": [], + "bbox": [ + 549, + 248, + 870, + 404 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LLM block. The results are shown in Tab.5. It is evident that the LLM-like structure achieved better performance. For a detailed analysis, please refer to Section 2.4.", + "bbox": [ + 511, + 476, + 906, + 522 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6c308948da08b3d3b907b08d44232da795d8a1428cd2728eb2f17606e06a7f3a.jpg", + "table_caption": [ + "Table 5. Ablation Study on the architecture of cognition gate" + ], + "table_footnote": [], + "table_body": "
ArchitectureTimeDiff ↓TriggerAcc ↑TimVal ↑
Linear Layer5.0620.13%17.65%
MLP projector + Linear4.3421.75%18.33%
Transformers + Linear3.7821.58%18.97%
Cross-Attention + Linear3.6424.34%20.36%
Single LLM block2.2235.35%32.11%
", + "bbox": [ + 519, + 551, + 901, + 631 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Initialization Strategies We investigate three initialization strategies for the Cognition Gate to enhance its effectiveness: (i) Random Initialization: All layers are initialized from a standard random distribution without utilizing prior knowledge from the LLM. (ii) SkipBlock Initialization: The Cognition Gate is initialized by uniformly selecting layers from the LLM, with the number of skipped layers determined by the ratio of the Cognition Gate's depth to the total number of LLM layers. (iii) EarlyBlock Initialization: The first few layers of the LLM are directly used to initialize the Cognition Gate, leveraging the LLM's early-stage processing capabilities.", + "bbox": [ + 511, + 654, + 906, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Number of Blocks The Cognition Gate consists of multiple layers, and its depth plays a crucial role in balancing computational efficiency and perceptual accuracy.", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13454", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5f8b654b184b558949a740fadf5b44477c964e21bb7f0a448cecca09e66889f0.jpg", + "table_caption": [ + "Table 6. Ablation Study on the Strategy of Initialization" + ], + "table_footnote": [], + "table_body": "
Strategy of InitializationTimeDiff ↓TriggerAcc ↑TimVal ↑
Random2.1039.65%37.35%
SkipBlock1.8940.23%37.67%
EarlyBlock1.8943.34%39.73%
", + "bbox": [ + 91, + 109, + 483, + 167 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f0f618264d24e09550e7330fd876a227b88e5387b19f4b750ffad76dba1e3c56.jpg", + "table_caption": [ + "Table 7. Ablation Study on the Strategy of Initialization" + ], + "table_footnote": [], + "table_body": "
#Num of BlockTimeDiff ↓TriggerAcc ↑TimVal ↑
21.9241.53%37.65%
31.9042.35%38.34%
41.8943.34%39.73%
51.8842.67%38.56%
61.8942.56%38.34%
", + "bbox": [ + 94, + 198, + 480, + 279 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results in Tab.6 and Tab.7 show that fine-tuning the shallow layers of the LLM yields the best performance. We attribute this to the fact that gate operates within a limited decision space (binary: Yes/No), making deeper layers—primarily responsible for nuanced language generation—unnecessary. Thus, we propose the Shallow Layer Transfer method to construct the Cognition Gate, as detailed in Section 2.4.", + "bbox": [ + 89, + 303, + 483, + 424 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. The Performance of Event-preserving Feature extractor", + "text_level": 1, + "bbox": [ + 89, + 435, + 483, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To evaluate the role of EPFE in the perception phase, we compare it with commonly used extractors in VideoLLMs, including: 1) The Q-Former [28]. 2) The Spatial-Temporal Convolution Connector (STC Connector) [12], which has demonstrated outstanding performance in videolm and has been widely adopted [29, 59, 71]. Notably, STC employs 3D convolution for spatial-temporal aggregation. To preserve local visual patterns during spatial compression, it follows work [5] by incorporating a RegStage block [47] before and after the 3D convolution, which has been shown to enhance spatial understanding. Experimental results indicate that EPFE achieves superior performance.", + "bbox": [ + 89, + 470, + 483, + 654 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/64d227cedc0229c59fcafdf4e1f53875aa35e7f97d1a843f7423d174b07db96d.jpg", + "table_caption": [ + "Table 8. Perceptive Judgement Capability" + ], + "table_footnote": [], + "table_body": "
MethodTimeDiff ↓TriggerAcc ↑TimVal ↑
Q-former3.7826.65 %25.31%
STC3.5627,54%26.87%
EPFE1.8943.34%39.73%
", + "bbox": [ + 112, + 675, + 464, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Related Work", + "text_level": 1, + "bbox": [ + 89, + 762, + 233, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1. Offline VideoLLMs.", + "text_level": 1, + "bbox": [ + 89, + 787, + 281, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In recent years, there has been a surge of interest in leveraging Large Language Models (LLMs) for video understanding, leading to the development of offline videoLLMs. The biggest challenge for it is managing memory requirements [4, 11, 15, 23, 50, 51] and compressing redundant frame features [12, 25, 27, 52, 76]. LaViLa [37] addresses", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "this challenge by transforming LLMs into narrators that generate detailed descriptions of long videos from visual inputs. Other models, such as ChatVideo [56] and MMVID [36], convert videos into text to improve comprehension. MovieChat [50] combines all frame features using an averaging strategy. However, in the context of online streaming, there has been limited exploration into how VideoLLMs can meet the demands of temporal alignment and real-time processing for streaming video inputs.", + "bbox": [ + 511, + 90, + 906, + 227 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Online VideoLLMs.", + "text_level": 1, + "bbox": [ + 511, + 247, + 702, + 262 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Recent studies have introduced VideoLLMs specifically designed for online stream understanding, while most existing works [14, 42, 45, 62, 72, 77] focus on improving offline video efficiency to meet real-time requirements. However, they fail to proactively decide when to respond based on user requirements. For example, ReKV [14] implemented a sliding-window attention mechanism and a KV-cache system to reduce computational overhead. To address this, VideoLLM-Online [8] introduces per-step LLM invocation for online video dialogue, as shown in Fig.1. However, its performance is constrained by limited per-frame processing capability due to the absence of effective frame compression strategies. VideoLLM-MOD [60] mitigates this issue by incorporating a mixture-of-depth approach for efficient visual token computation, enabling higher visual input resolutions. Building on MOD, LION-FS [31] proposes Token Aggregation Router and Token DroppingRouter as the Fast Path, adaptively aggregating distinct features while discarding redundant ones. However, they remain limited to processing videos at only 10 FPS.", + "bbox": [ + 511, + 270, + 906, + 573 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Overall, prior work remains restricted by the contradiction between sequential full-frame processing and the extremely high efficiency requirements of streaming video, lacking breakthrough paradigm designs explicitly suited for streaming frame inputs. To address these challenges, our approach introduces a novel perception-cognition interleaving paradigm and proposes the STREAMMIND framework, enabling VideoLLM to achieve proactive responses and real-time streaming video understanding, with a maximum throughput of up to 100 FPS.", + "bbox": [ + 511, + 577, + 908, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 753, + 633, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce a novel perception-cognition interleaving paradigm called \"event-gated LLM invocation\" and propose the STREAMMIND framework, addressing the contradiction between sequential full-frame processing and the extremely high efficiency requirements of streaming video. As a result, STREAMMIND can process video at up to 100 FPS, achieving the SOTA performance across all existing metrics and benchmarks.", + "bbox": [ + 511, + 780, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13455", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 272, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We gratefully acknowledge the support from the following foundations and institutions. Ting Cao was supported in part by the Tsinghua University (AIR)-AsiaInfo Technologies (China), Inc. Joint Research Center for 6G Network and Intelligent Computing. Hao Wu was supported in part by the National Natural Science Foundation of China (NSFC) under Grants 62432004 and 62302207, and by the Science Foundation for Youths of Jiangsu Province under Grant BK20220772. Xin Ding and Zhibo Chen were supported in part by the NSFC under Grants 62371434 and 62021001.", + "bbox": [ + 89, + 113, + 485, + 270 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 295, + 187, + 311 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Roberto Amoroso, Gengyuan Zhang, Rajat Koner, Lorenzo Baraldi, Rita Cucchiara, and Volker Tresp. Perceive, query & reason: Enhancing video qa with question-guided temporal queries. arXiv preprint arXiv:2412.19304, 2024. 2", + "[2] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 6", + "[3] Pietro Battistoni, Andrea Antonio Cantone, Mariarosaria Esposito, Rita Francese, Francesca Pia Perillo, Marco Romano, Monica Sebillo, and Giuliana Vitiello. Using artificial intelligence and companion robots to improve home healthcare for the elderly. In International Conference on Human-Computer Interaction, pages 3–17. Springer, 2023. 1", + "[4] Jiarui Cai, Mingze Xu, Wei Li, Yuanjun Xiong, Wei Xia, Zhuowen Tu, and Stefano Soatto. Memot: Multi-object tracking with memory. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8090-8100, 2022. 8", + "[5] Junbum Cha, Wooyoung Kang, Jonghwan Mun, and Byungseok Roh. Honeybee: Locality-enhanced projector for multimodal llm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13817-13827, 2024. 8", + "[6] Yupeng Chang, Xu Wang, Jindong Wang, Yuan Wu, Linyi Yang, Kaijie Zhu, Hao Chen, Xiaoyuan Yi, Cunxiang Wang, Yidong Wang, et al. A survey on evaluation of large language models. ACM transactions on intelligent systems and technology, 15(3):1-45, 2024. 6", + "[7] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3558-3568, 2021. 6", + "[8] Joya Chen, Zhaoyang Lv, Shiwei Wu, Kevin Qinghong Lin, Chenan Song, Difei Gao, Jia-Wei Liu, Ziteng Gao, Dongxing Mao, and Mike Zheng Shou. Videollm-online: Online video large language model for streaming video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18407-18418, 2024. 1, 2, 3, 5, 6, 8" + ], + "bbox": [ + 99, + 320, + 480, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[9] Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji, and Quanquan Gu. Self-play fine-tuning converts weak language models to strong language models. arXiv preprint arXiv:2401.01335, 2024. 6", + "[10] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24185–24198, 2024. 3", + "[11] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In European Conference on Computer Vision, pages 640-658. Springer, 2022. 8", + "[12] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 8", + "[13] Shangzhe Di and Weidi Xie. Grounded question-answering in long egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12934–12943, 2024. 3", + "[14] Shangzhe Di, Zhelun Yu, Guanghao Zhang, Haoyuan Li, Hao Cheng, Bolin Li, Wanggui He, Fangxun Shu, Hao Jiang, et al. Streaming video question-answering with in-context video kv-cache retrieval. In The Thirteenth International Conference on Learning Representations. 1, 2, 3, 8", + "[15] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. In European Conference on Computer Vision, pages 75-92. Springer, 2024. 8", + "[16] Feiteng Fang, Yuelin Bai, Shiwen Ni, Min Yang, Xiaojun Chen, and Ruifeng Xu. Enhancing noise robustness of retrieval-augmented language models with adaptive adversarial training. arXiv preprint arXiv:2405.20978, 2024. 2", + "[17] Xinyu Fang, Kangrui Mao, Haodong Duan, Xiangyu Zhao, Yining Li, Dahua Lin, and Kai Chen. Mmbench-video: A long-form multi-shot benchmark for holistic video understanding. Advances in Neural Information Processing Systems, 37:89098-89124, 2025. 2", + "[18] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19358-19369, 2023. 3", + "[19] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 2", + "[20] Silvio Giancola, Mohieddine Amine, Tarek Dghaily, and Bernard Ghanem. Soccernet: A scalable dataset for action spotting in soccer videos. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 1711-1721, 2018. 2, 5" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13456", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[21] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18995-19012, 2022. 2, 5", + "[22] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. CoRR, abs/2312.00752, 2023. 2, 4", + "[23] Bo He, Hengduo Li, Young Kyun Jang, Menglin Jia, Xuefei Cao, Ashish Shah, Abhinav Shrivastava, and Ser-Nam Lim. Ma-lmm: Memory-augmented large multimodal model for long-term video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13504-13514, 2024. 8", + "[24] Daoji Huang, Otmar Hilliges, Luc Van Gool, and Xi Wang. Palm: Predicting actions through language models@ ego4d long-term action anticipation challenge 2023. arXiv preprint arXiv:2306.16545, 2023. 2, 7", + "[25] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In European Conference on Computer Vision, pages 202-218. Springer, 2024. 8", + "[26] Zhenpeng Huang, Xinhao Li, Jiaqi Li, Jing Wang, Xi-angyu Zeng, Cheng Liang, Tao Wu, Xi Chen, Liang Li, and Limin Wang. Online video understanding: A comprehensive benchmark and memory-augmented method. arXiv preprint arXiv:2501.00584, 2024. 1, 2, 5", + "[27] Yang Jin, Zhicheng Sun, Kun Xu, Liwei Chen, Hao Jiang, Quzhe Huang, Chengru Song, Yuliang Liu, Di Zhang, Yang Song, et al. Video-lavit: Unified video-language pre-training with decoupled visual-motional tokenization. arXiv preprint arXiv:2402.03161, 2024. 8", + "[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2, 3, 8", + "[29] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 5, 8", + "[30] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 1, 2", + "[31] Wei Li, Bing Hu, Rui Shao, Leyang Shen, and Liqiang Nie. Lion-fs: Fast & slow video-language thinker as online video assistant. arXiv preprint arXiv:2503.03663, 2025. 5, 6, 8", + "[32] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2024. 1, 3" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 3", + "[34] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 6", + "[35] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 26689-26699, 2024. 3", + "[36] Kevin Lin, Faisal Ahmed, Linjie Li, Chung-Ching Lin, Ehsan Azarnasab, Zhengyuan Yang, Jianfeng Wang, Lin Liang, Zicheng Liu, Yumao Lu, et al. Mm-vid: Advancing video understanding with gpt-4v (isdiction). arXiv preprint arXiv:2310.19773, 2023. 8", + "[37] Xingyu Liu, Joon-Young Lee, and Hailin Jin. Learning video representations from correspondence proposals. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 4273–4281, 2019. 8", + "[38] Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. G-eval: Nlg evaluation using gpt-4 with better human alignment. arXiv preprint arXiv:2303.16634, 2023. 6", + "[39] Ruipu Luo, Ziwang Zhao, Min Yang, Junwei Dong, Da Li, Pengcheng Lu, Tao Wang, Linmei Hu, Minghui Qiu, and Zhongyu Wei. Valley: Video assistant with large language model enhanced ability. arXiv preprint arXiv:2306.07207, 2023. 1, 3", + "[40] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 1, 5", + "[41] Kelly Merrill Jr, Jihyun Kim, and Chad Collins. Ai companions for lonely individuals and the role of social presence. Communication Research Reports, 39(2):93-103, 2022. 1", + "[42] Sunny Panchal, Apratim Bhattacharyya, Guillaume Berger, Antoine Mercier, Cornelius Böhm, Florian Dietrichkeit, Reza Pourreza, Xuanlin Li, Pulkit Madan, Mingu Lee, et al. What to say and when to say it: Live fitness coaching as a testbed for situated interaction. Advances in Neural Information Processing Systems, 37:75853-75882, 2024. 8", + "[43] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 6", + "[44] Jordi Pont-Tuset, Jasper Uijlings, Soravit Changpinyo, Radu Soricut, and Vittorio Ferrari. Connecting vision and language with localized narratives. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 647-664. Springer, 2020. 6", + "[45] Rui Qian, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Shuanggrui Ding, Dahua Lin, and Jiaqi Wang. Streaming long video understanding with large language models. Advances in Neu" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13457", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ral Information Processing Systems, 37:119336-119360, 2025. 1, 2, 3, 8", + "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 3, 4", + "[47] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólár. Designing network design spaces. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10428-10436, 2020. 8", + "[48] Jiayuan Rao, Haoning Wu, Chang Liu, Yanfeng Wang, and Weidi Xie. Matchtime: Towards automatic soccer game commentary generation. arXiv preprint arXiv:2406.18530, 2024. 6", + "[49] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 1", + "[50] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 8", + "[51] Larry R Squire, Lisa Genzel, John T Wixted, and Richard G Morris. Memory consolidation. Cold Spring Harbor perspectives in biology, 7(8):a021766, 2015. 8", + "[52] Guangzhi Sun, Wenyi Yu, Changli Tang, Xianzhao Chen, Tian Tan, Wei Li, Lu Lu, Zejun Ma, and Chao Zhang. Fine-grained audio-visual joint representations for multimodal large language models. arXiv preprint arXiv:2310.05863, 2023. 8", + "[53] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1207-1216, 2019. 2, 7", + "[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 7", + "[55] Ales Vysocky and Petr Novak. Human-robot collaboration in industry. MM Science Journal, 9(2):903–906, 2016. 1", + "[56] Junke Wang, Dongdong Chen, Chong Luo, Xiyang Dai, Lu Yuan, Zuxuan Wu, and Yu-Gang Jiang. Chatvideo: A tracklet-centric multimodal and versatile video understanding system. arXiv preprint arXiv:2304.14407, 2023. 8", + "[57] Qing Wang, Jiaming Zhang, Kailun Yang, Kunyu Peng, and Rainer Stiefelhagen. Matchformer: Interleaving attention in transformers for feature matching. In Proceedings of the Asian Conference on Computer Vision, pages 2746-2762, 2022. 2" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[58] Xunguang Wang, Zhenlan Ji, Pingchuan Ma, Zongjie Li, and Shuai Wang. Instructta: Instruction-tuned targeted attack for large vision-language models. arXiv preprint arXiv:2312.01886, 2023. 2", + "[59] Yuxuan Wang, Yueqian Wang, Pengfei Wu, Jianxin Liang, Dongyan Zhao, and Zilong Zheng. Lstp: Language-guided spatial-temporal prompt learning for long-form video-text understanding. arXiv e-prints, pages arXiv-2402, 2024. 8", + "[60] Shiwei Wu, Joya Chen, Kevin Qinghong Lin, Qimeng Wang, Yan Gao, Qianli Xu, Tong Xu, Yao Hu, Enhong Chen, and Mike Zheng Shou. Videollm-mod: Efficient video-language streaming with mixture-of-depths vision computation. Advances in Neural Information Processing Systems, 37:109922-109947, 2025. 1, 3, 5, 6, 8", + "[61] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9777–9786, 2021. 3", + "[62] Haomiao Xiong, Zongxin Yang, Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Jiawen Zhu, and Huchuan Lu. Streaming video understanding and multi-round interaction with memory-enhanced knowledge. arXiv preprint arXiv:2501.13468, 2025. 1, 2, 3, 8", + "[63] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the 25th ACM international conference on Multimedia, pages 1645-1653, 2017. 2", + "[64] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 1", + "[65] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1686-1697, 2021. 3", + "[66] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. Advances in Neural Information Processing Systems, 35:124-141, 2022. 1", + "[67] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10714-10726, 2023. 1", + "[68] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9127–9134, 2019. 2", + "[69] Jeffrey M Zacks, Nicole K Speer, Khena M Swallow, Todd S Braver, and Jeremy R Reynolds. Event perception: a mind-brain perspective. Psychological bulletin, 133(2):273, 2007. 2" + ], + "bbox": [ + 516, + 90, + 906, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13458", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[70] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975-11986, 2023. 3", + "[71] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 5, 8", + "[72] Haoji Zhang, Yiqin Wang, Yansong Tang, Yong Liu, Jiashi Feng, Jifeng Dai, and Xiaojie Jin. Flash-vstream: Memory-based real-time understanding for long video streams. arXiv preprint arXiv:2406.08085, 2024. 1, 2, 8", + "[73] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 3", + "[74] Ruohong Zhang, Liangke Gui, Zhiqing Sun, Yihao Feng, Keyang Xu, Yuanhan Zhang, Di Fu, Chunyuan Li, Alexander Hauptmann, Yonatan Bisk, et al. Direct preference optimization of video large multimodal models from language model reward. arXiv preprint arXiv:2404.01258, 2024. 3", + "[75] Ruoyu Zhang, Lulu Wang, Yi He, Tongling Pan, Zhengtao Yu, and Yingna Li. Tpcap: Unlocking zero-shot image captioning with trigger-augmented and multi-modal purification modules. arXiv preprint arXiv:2502.11024, 2025. 2", + "[76] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6586-6597, 2023. 8", + "[77] Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. Streaming dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18243-18252, 2024. 8" + ], + "bbox": [ + 91, + 92, + 480, + 584 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "13459", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/StreamMind_ Unlocking Full Frame Rate Streaming Video Dialogue through Event-Gated Cognition/61e82f55-a3d1-4244-b3cc-22632c8cab6f_model.json b/2025/StreamMind_ Unlocking Full Frame Rate Streaming Video Dialogue through Event-Gated Cognition/61e82f55-a3d1-4244-b3cc-22632c8cab6f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e39af217b4560818888235c9a65b50658f2fc9ae --- /dev/null +++ b/2025/StreamMind_ Unlocking Full Frame Rate Streaming Video Dialogue through Event-Gated Cognition/61e82f55-a3d1-4244-b3cc-22632c8cab6f_model.json @@ -0,0 +1,2710 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.099, + 0.13, + 0.9, + 0.178 + ], + "angle": 0, + "content": "STREAMMIND: Unlocking Full Frame Rate Streaming Video Dialogue through Event-Gated Cognition" + }, + { + "type": "text", + "bbox": [ + 0.185, + 0.202, + 0.814, + 0.24 + ], + "angle": 0, + "content": "Xin Ding\\(^{1}\\) * Hao Wu\\(^{3}\\) † Yifan Yang\\(^{2}\\) Shiqi Jiang\\(^{2}\\) Qianxi Zhang\\(^{2}\\) Donglin Bai\\(^{2}\\) Zhibo Chen\\(^{1}\\) Ting Cao\\(^{4}\\) †" + }, + { + "type": "text", + "bbox": [ + 0.122, + 0.24, + 0.875, + 0.276 + ], + "angle": 0, + "content": "1University of Science and Technology of China 2Microsoft Research 3Nanjing University Institute for AI Industry Research (AIR), Tsinghua University" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.484, + 0.448 + ], + "angle": 0, + "content": "With the rise of real-world human-AI interaction applications, such as AI assistants, the need for Streaming Video Dialogue is critical. To address this need, we introduce STREAMMIND, a video LLM framework that achieves ultra-FPS streaming video processing (100fps on a single A100) and enables proactive, always-on responses in real time, without explicit user intervention." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.45, + 0.484, + 0.646 + ], + "angle": 0, + "content": "To solve the key challenge of the contradiction between linear video streaming speed and quadratic transformer computation cost, we propose a novel perception-cognition interleaving paradigm named \"event-gated LLM invocation\", in contrast to the existing per-time-step LLM invocation. By introducing a Cognition Gate network between the video encoder and the LLM, LLM is only invoked when relevant events occur. To realize the event feature extraction with constant cost, we propose Event-Preserving Feature Extractor (EPFE) based on state-space method, generating a single perception token for spatiotemporal features. These techniques enable the video LLM with full-FPS perception and real-time cognition response." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.647, + 0.484, + 0.738 + ], + "angle": 0, + "content": "Experiments on Ego4D and SoccerNet streaming tasks, as well as standard offline benchmarks, demonstrate state-of-the-art performance in both model capability and real-time efficiency, paving the way for ultra-high-FPS applications, such as Game AI Copilot and interactive media. The code and data is available at https://aka.ms/StreamMind." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.763, + 0.222, + 0.779 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.789, + 0.483, + 0.851 + ], + "angle": 0, + "content": "The advancement of large foundation models is driving an increasing number of real-world human-AI interaction applications, such as AI home companions [3, 41] and human-robot collaboration [55, 64]. These applications" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.308, + 0.718, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.408, + 0.721, + 0.42 + ], + "angle": 0, + "content": "Existing Works 10~15 FPS on A100 GPU" + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.309, + 0.923, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.734, + 0.408, + 0.907, + 0.42 + ], + "angle": 0, + "content": "Ours Up to 100 FPS on A100 GPU" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.425, + 0.907, + 0.482 + ], + "angle": 0, + "content": "Figure 1. The paradigm of existing per-step LLM invocation (left) versus our event-gated LLM invocation (right) for streaming video dialogue task. Our paradigm boosts frame rate up-to 100 fps on A100. \\((R_{i}\\): response, \\(f_{i}\\): frame, t: time)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.497, + 0.907, + 0.634 + ], + "angle": 0, + "content": "heavily rely on streaming video understanding capabilities [14, 26, 45, 62, 72], specifically streaming video dialogue [8, 60]. Streaming video dialogue (StreamingVD) aims to continuously perceive incoming video streams and, based on user queries, provide proactive, real-time, and always-on responses without human intervention. For example, an AI companion could respond to user queries like \"Guide me to fix this bicycle\" or \"Let's watch and discuss the soccer game together\"." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.639, + 0.909, + 0.821 + ], + "angle": 0, + "content": "Streaming video dialogue can be viewed as a generalization of standard offline video dialogue tasks [30, 32, 39, 40, 49, 66, 67], extending video LLMs to broader real-world scenarios. The fundamental challenge of StreamingVD compared to its offline counterpart lies in the requirement for timing alignment between events and responses. That is, as the video progresses, the model is required to (1) proactively decide when to respond and (2) respond in real time before the next event occurs according to user queries, as illustrated in Fig. 2. However, with current techniques, these two requirements are contradictory and cannot be achieved simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The seminal VideoLLM-Online [8] and VideoLLM-MoD [60] introduce a per-step LLM invocation paradigm for StreamingVD, shown in Fig. 1. At each time step, it inputs all past frames and the user query to the LLM, to decide when to respond or remain silent. However," + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.483, + 0.887 + ], + "angle": 0, + "content": "*This work was done during Xin Ding's internship at Microsoft Research." + }, + { + "type": "page_footnote", + "bbox": [ + 0.117, + 0.888, + 0.246, + 0.9 + ], + "angle": 0, + "content": "†Corresponding Author." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.863, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13448" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.091, + 0.482, + 0.212 + ], + "angle": 0, + "content": "the \\(O(n^3)^1\\) computational complexity and limited context window are mismatched with the \\(O(n)\\) frame streaming, making real-time response difficult to achieve. More works [14, 45, 62, 72] focus on enhancing the efficiency of offline video dialogue aiming to meet real-time requirements. However, similar to offline approaches, they depend on user queries to trigger responses rather than the model to do so proactively." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.215, + 0.486, + 0.455 + ], + "angle": 0, + "content": "To empower StreamingVD, this paper proposes a new perception-cognition interleaving paradigm, named event-gated LLM invocation, shown in Fig. 1. It is inspired by the event perception mechanism of human brain [69], which states that events are key components of human's perception, attention, and memory\\(^2\\). We therefore propose an event-based selective cognition process. Rather than invoking the LLM at every time step, we introduce a Cognition Gate between the video encoder and LLM. The video encoder continuously perceives visual signals of each frame, and only when query-related events occur, the gate will open to invoke the LLM. Analogous to how a text prompt triggers an LLM, streaming visual signals lack a clear stopping point, unlike a finite text sequence. The cognition gate bridges this gap by acting as a mechanism to identify \"physical world prompts\" for the LLM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.483, + 0.669 + ], + "angle": 0, + "content": "This new paradigm raises two major technical concerns, to achieve both high capability and efficiency. Firstly, existing video encoders in VideoLLMs (e.g., STC in VideoLaMA) only capture local spatio-temporal features. This results in a linear increase in the number of feature tokens for streaming frames and subsequently a quadratic computation for the processing, failing to meet real-time requirements. Secondly, the Cognition Gate must decide to invoke the LLM or not, based on visual features and the user query. However, only considering efficiency and simply attending visual and query tokens (e.g., Cross Attention in Qformer [28]) limits the gate capability to tasks like feature-matching [57, 58] and retrieval [1, 16, 75], struggle to make decisions with deeper semantic understanding." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.67, + 0.483, + 0.837 + ], + "angle": 0, + "content": "By addressing the challenges, we realize STREAM-MIND, a novel framework for streamingVD, with the processing speed matching frame rate for the first time. It integrates two techniques. For the video encoder, inspired by the exceptional ability of the state-assisted module [22] in modeling continuous physical signals with constant cost, we propose the Event-Preserving Feature Extractor (EPFE) based on the state-assisted module. It generates a single perception token of event features to the gate, resulting in a constant perception cost. For the Cognition Gate, to leverage the world knowledge of LLM while meeting the real-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "time requirement, we propose a Shallow Layer Transfer method. The Gate is designed to reuse the shallow layers of the LLM and its autoregressive training to maximize the probabilities of “< response/silence >” tokens during training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.907, + 0.394 + ], + "angle": 0, + "content": "To evaluate the performance of the STREAMMIND framework, we adopt evaluation metrics from VideoLLM-Online and a wide range of text generation evaluation metrics. Additionally, we design two novel metrics to comprehensively assess the model's temporal responsiveness. Experiments on real-time Ego4D [21] and SoccerNet [20] demonstrate superior performance across all metrics. Furthermore, our model achieves state-of-the-art results on multiple offline benchmarks, including short- and long-term activity recognition and forecasting on the COIN [53] and Ego4D LTA [24] datasets. Notably, our framework breaks the ultra-FPS streaming video processing bottleneck, enabling streamingVD at up to 100 fps on a single A100 GPU, laying a robust foundation for real-world human-AI interaction applications." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.401, + 0.605, + 0.417 + ], + "angle": 0, + "content": "2. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.426, + 0.786, + 0.44 + ], + "angle": 0, + "content": "2.1. Task Discussion and Definition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.907, + 0.645 + ], + "angle": 0, + "content": "Problem formulation of StreamingVD Given a video stream \\(\\mathcal{V}^T \\coloneqq [v_1, v_2, \\dots, v_T]\\) consisting of \\(T\\) frames and a set of queries \\(\\mathcal{Q} \\coloneqq \\{q_1, q_2, \\dots, q_N\\}\\). Traditional video dialogue, shown in Fig. 2, provides a response triggered by a user query. For a query \\(q_i\\) given at time step \\(t_s\\), it responds based on past frames up to time \\(t_s (1 < t_s \\leq T)\\), regardless of an online or offline scenario. On the other hand, StreamingVD proactively generates responses to a query given at \\(t_s\\) at any appropriate time steps until the query termination at \\(t_e (t_s < t_e \\leq T)\\). At each time step \\(t_i (t_s \\leq t_i \\leq t_e)\\), StreamingVD needs to decide whether to generate a response according to the past frames and query, and then generate the response, formulated as follows." + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.649, + 0.905, + 0.668 + ], + "angle": 0, + "content": "\\[\n\\max P \\left(\\left[ \\operatorname {R e s} ^ {t _ {i}} \\right] \\mid \\left[ \\operatorname {C t x} ^ {< t _ {i}} \\right], \\left[ F ^ {t _ {i}} \\right]\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.907, + 0.749 + ], + "angle": 0, + "content": "where \\([\\mathrm{Res}^{t_i}] \\in \\{[\\mathrm{Txt}^{t_i}], [\\mathrm{EOS}^{t_i}]\\}\\) means generate a response at a time step \\(t_i\\), or remain silence, i.e., [EOS]. \\([\\mathrm{Ctx}^{< t_i}]\\) denotes the contextual tokens accumulated from past frames and queries before \\(t_i\\). \\([\\mathrm{F}^{t_i}]\\) represents the visual features extracted from the video frame at \\(t_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Challenges of StreamingVD tasks Firstly, increasing number of frames versus real-time requirement. Different from the offline tasks[17, 19, 30, 63, 68], streaming tasks[8, 14, 26, 45, 62, 72] require time alignment between event and response. Besides only accuracy constraint of offline task, streaming video tasks need to consider both accuracy and timeliness constraints. Offline tasks can trade resources and latency for better quality, while streaming task cannot." + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.851, + 0.483, + 0.875 + ], + "angle": 0, + "content": "1Invoking a quadratic-complexity LLM n times results in cubic complexity." + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "2A formal definition of an event [69]: \"A segment of time at a given location that is conceived by an observer to have a beginning and an end.\"" + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.851, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13449" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.176, + 0.094, + 0.817, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.336, + 0.908, + 0.38 + ], + "angle": 0, + "content": "Figure 2. Comparison between streaming dialogue (bottom) and traditional dialogue tasks (top). Streaming dialogue proactively generates responses at appropriate time based on user queries and perception, whereas traditional dialogue requires human triggering for each response." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.391, + 0.484, + 0.527 + ], + "angle": 0, + "content": "Secondly, proactive response versus real-time requirement. Traditional video dialogue [13, 14, 45, 61, 62, 65] are reactive, responding only at the time of user query. However, real-world dialogue is often without human triggering. The user may lack the specific knowledge to formulate questions. Similarly, embodied AI should anticipate user needs rather than merely reacting to commands. However, proactive response means at every frame the model needs to judge, inceasing the model cost by \\( \\mathrm{O}(\\mathfrak{n}) \\) times." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.538, + 0.484, + 0.84 + ], + "angle": 0, + "content": "Existing works Conventional offline Video LLMs [10, 32, 33, 35, 73, 74] typically employ a visual encoder [18, 46, 70] and a projection module [28, 39] to process uniformly sampled frames. The output is concatenated with the tokenized question to form a sequence, which is then passed to an LLM decoder to predict a response. However, this approach is impractical for streaming tasks due to the high computational cost associated with processing a large number of frames \\((T)\\). Therefore, works such as [14, 62] have implemented a sliding-window attention mechanism and a KV-cache system to reduce computational overhead, or incorporated a parallel sampling strategy to enhance processing speed and reduce latency. However, these methods do not proactively trigger response when needed. In contrast, Videollm-Online [8] and MOD [60] utilize the LLM decoder to process every frame of streaming videos, resulting in a cubic computation cost. STREAMMIND, on the other hand, has the potential to process streaming videos at 100 fps, which is crucial for enabling interactive video analysis applications." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.363, + 0.864 + ], + "angle": 0, + "content": "2.2. The Overview of STREAMMIND" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.87, + 0.483, + 0.903 + ], + "angle": 0, + "content": "For a streaming video sequence \\(\\left[v^{t_s} \\leq t_i \\leq t_e\\right]\\), STREAMMIND models the following process as shown in Fig. 3:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.39, + 0.907, + 0.526 + ], + "angle": 0, + "content": "1. Perception Phase: With continuous streaming video, each incoming frame \\( v \\) is processed through the CLIP [46] encoder to extract spatial features. The features are then input to our proposed EPFE. EPFE updates its hidden state to correlate the frame's features across temporal dimensions and reduce redundant spatial information. This process generates a new perception token \\( [\\mathsf{F}_{per}^{t_i}] \\) and saves this token in the Perception Memory \\( \\mathcal{M}_{\\mathrm{per}}^{t_i} \\). The perception phase can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.535, + 0.907, + 0.557 + ], + "angle": 0, + "content": "\\[\n\\left[ \\mathrm {F} _ {p e r} ^ {t _ {i}} \\right] = \\operatorname {E P F E} \\left(\\operatorname {C L I P} \\left(v ^ {t _ {i}}\\right), \\mathrm {H} ^ {t _ {i} - 1}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.562, + 0.904, + 0.577 + ], + "angle": 0, + "content": "where \\(\\mathbb{H}^{t - 1}\\) denotes the hidden state from the previous time." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.578, + 0.906, + 0.655 + ], + "angle": 0, + "content": "2. Cognition Gate Judgment: We introduce Cognition Gate \\(\\mathcal{G}\\) between perception and cognition. The input are user query [Prompt] and the generated perception token \\([\\mathsf{F}_{per}^{t_i}]\\) at each time step \\(t_i\\). The Cognition Gate learns to determine whether to trigger the cognition phase at \\(t_i\\):" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.662, + 0.905, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\max P \\left(\\left[ \\operatorname {R e s} ^ {t _ {i}} \\right] \\mid \\mathcal {G} \\left(\\left[ \\text {P r o m p t} \\right], \\left[ F _ {\\text {p e r}} ^ {t _ {i}} \\right]\\right)\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.719 + ], + "angle": 0, + "content": "where \\([\\mathrm{Res}^{t_i}] \\in \\{< /\\text{response}>, < /\\text{silence}>\\}\\) determines whether to trigger a response or remain silent." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.78 + ], + "angle": 0, + "content": "3. Cognition Phase Upon receiving the trigger signal from the Cognition Gate, STREAMMIND samples perception tokens from the Perception Memory \\(\\mathcal{M}_{\\mathrm{per}}^{t_i}\\) to the Cognition Pooling as the input to LLM to generate response." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.789, + 0.683, + 0.804 + ], + "angle": 0, + "content": "2.3. Perception Phase" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Motivation: To match the constant frame rate of streaming videos, ideally the perception phase should be constant computation cost for processing each frame. However, current video encoders generally extract local (dozens of frames) spatiotemporal features with constant cost, and input all the generated perception tokens to a transformer" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13450" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.088, + 0.887, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.27, + 0.908, + 0.34 + ], + "angle": 0, + "content": "Figure 3. STREAMMIND workflow. For each video frame, the perception stage goes through Clip for spatial features and the proposed EPFE for spatiotemporal features, to generate a single perception token saved in perception memory. The Cognition Gate determines whether to invoke LLM only based on the current perception token. If yes (a query-related event occurs), the tokens in perception memory are sampled to cognition pooling as the input for LLM for cognition response. The gate decouples perception and cognition to enable full-FPS perception and event-based cognition." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.355, + 0.483, + 0.4 + ], + "angle": 0, + "content": "model to extract the long-term temporal relationships. The computational cost is \\( O(n^{3}) \\), obviously the lagger behind the streaming rate." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.4, + 0.483, + 0.552 + ], + "angle": 0, + "content": "To solve the issue, we propose Event-Preserving Feature Extractor (EPFE) to replace the projection module used in existing video LLMs (e.g., cross-attention and Q-Former). EPFE is based on the SOTA state-space model, i.e., Selective State Space model [22]. As shown in Fig. 3 (blue box) and Eq. 4, for each frame, EPFE dynamically correlates the spatial features and the past internal states to generate a single perception token \\(\\left[\\mathbb{F}_{per}^{t_i}\\right]\\) as well as update its state. This enables a matched speed of streaming video and perception processing, for full FPS StreamingVD." + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.56, + 0.483, + 0.577 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} _ {t + 1} = \\mathbf {A h} _ {t} + \\mathbf {B x} _ {t}, \\quad \\mathbf {y} _ {t} = \\mathbf {C h} _ {t} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.584, + 0.483, + 0.645 + ], + "angle": 0, + "content": "where \\(\\mathbf{A},\\mathbf{B}\\), and \\(\\mathbf{C}\\) are learnable state-space matrices for state, input, and output, respectively. \\(\\mathbf{x_t}\\) is the input frame of the current time step. \\(\\mathbf{h_t}\\) is the internal hidden state. \\(\\mathbf{y_t}\\) is the output, i.e., perception token." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.645, + 0.483, + 0.766 + ], + "angle": 0, + "content": "Benefit from the strong temporal modeling capability of Selective SSM, our EPFE with only \\(56\\mathrm{M}\\) parameters can learn long-term event-level spatiotemporal features with constant cost. As Appendix Fig. 1 illustrates, perception tokens generated by EPFE can distinguish different events and noise frames, while maintaining a state of the event. Even after unrelated noise frames, the perception tokens are capable of refocusing on the event." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.245, + 0.789 + ], + "angle": 0, + "content": "2.4. Cognition Gate" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.795, + 0.483, + 0.871 + ], + "angle": 0, + "content": "Motivation: Cognition Gate needs to determine whether to invoke LLM in real-time, based on the current perception token and user query. For example, during the soccer match, some user query may only care LLM response for scoring moments, while others may need detailed tactical analysis." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "However, we find it highly challenging. For real-time consideration, we first employ a simple Transformer block" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.355, + 0.905, + 0.492 + ], + "angle": 0, + "content": "with cross-attention to associate the user query with the visual encoder's output for trigger decisions. However, it struggles to make accurate judgments at the right time step, as shown in Tab.5. We identify two key reasons for this: 1) The Cognition Gate only has access to information up to the current step, lacking a global temporal perspective for decision-making. 2) User queries may not directly correspond to specific video frames in a retrieval-like manner (e.g., guide me to repair a bicycle or cooking)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.496, + 0.905, + 0.572 + ], + "angle": 0, + "content": "Therefore, the Cognition Gate requires the world knowledge of LLMs. However, due to the size of LLMs, even processing a single perception token as input incurs a high computational cost. Notably, the Gate only needs to generate a response/silence token for a single iteration." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.577, + 0.907, + 0.7 + ], + "angle": 0, + "content": "To achieve it, We propose the Shallow Layer Transfer method, which initializes the Gate using the early layers of the LLM and then fine-tunes it in a supervised autoregressive manner with the StreamVD dataset. As shown in Sec. 4.2, we conducted extensive experiments to analyze its impact. The results indicate that the first few layers exhibit slight variations in behavior, allowing for scalability based on requirements. In this work, we use four layers." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.723, + 0.845, + 0.739 + ], + "angle": 0, + "content": "2.5. The Training Strategy of STREAMMIND" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Model Architecture As illustrated in Figure 3, the STREAMMIND video LLM integrates CLIP [46], EPFE, Cognition Gate, and the LLM. CLIP is the pre-trained spatial feature extractor. EPFE then takes the spatial features as input as well as its internal state to extract spatiotemporal features and generate a perception token. Cognition Gate then evaluates the current perception token and user query to determine whether to invoke LLM. To achieve this, it is needed to train EPFE, Cognition Gate, and also align the representation space of EPFE and LLM, as shown in Fig. 4." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13451" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.101, + 0.096, + 0.438, + 0.123 + ], + "angle": 0, + "content": "Algorithm 1: Streaming Video Dataset Construction" + }, + { + "type": "algorithm", + "bbox": [ + 0.11, + 0.125, + 0.47, + 0.416 + ], + "angle": 0, + "content": "Require: Offline video dataset \\(C = \\{c_i,t_i\\}_{i = 1}^n (c_i\\) caption at time \\(t_i)\\) \nEnsure: Dataset \\(D\\) with captions meeting StreamVD requirements and Cognition Gate labels. \n1: Step 1: Preprocessing \n2: \\(D\\gets []\\) , cap \\(\\leftarrow c_{1}\\) ,time \\(\\leftarrow t_{s}\\) \n3: for \\(i = 2\\) to n do \n4: if \\(c_{i}\\neq\\) cap then \n5: Append (cap, time) to \\(D\\) \n6: cap \\(\\leftarrow c_{i}\\) ,time \\(\\leftarrow t_{i}\\) \n7: end if \n8: end for \n9: Append (cap, time) to \\(D\\) \n10: Step 2: Silence-Response Labeling \n11: for each frame \\(f\\) between \\(d_{i}\\) and \\(d_{i + 1}\\) do \n12: Label \\(f\\) withif captioned, else \n13: end for \n14: return \\(D\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.445, + 0.483, + 0.55 + ], + "angle": 0, + "content": "Data Preparation Streaming datasets are constructed by processing existing offline datasets. StreamingVD tasks require the model to response as early as possible when relevant events occur while avoid redundant responses. To achieve this, we eliminate repetitive captions from the offline datasets and transform them into streaming datasets using the following methods, detailed in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.551, + 0.483, + 0.596 + ], + "angle": 0, + "content": "- Preprocessing: Adjacent identical captions are merged, with the timestamp of the first occurrence being recorded as the annotation point for the caption." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.596, + 0.483, + 0.687 + ], + "angle": 0, + "content": "- Silence-Response Labeling: To train the Cognition Gate, we insert \\(< / \\text{silence}>\\) tokens into video frames between two adjacent captions. These tokens act as indicators to guide the Cognition Gate in generating the \\(< / \\text{silence}>\\). The initial frame corresponding to captioned events is labeled with the \\(< / \\text{response}>\\) token." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.551, + 0.483, + 0.687 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Training Strategy We adopt a two-stage training strategy, as shown in Fig. 4. In contrast to previous typical two-stage training approaches [29, 40, 71], our method uses a dedicated stage for Cognition Gate. In the first stage, the LLM and EPFE are jointly trained using video frames and captions from the streaming dataset, ensuring spatiotemporal feature alignment with the LLM. The second stage trains only the Cognition Gate, which is initialized from the shallow layers of the LLM. This training also follows the LLM autoregressive manner to generate \\( < / \\text{silence} > \\) or \\( < / \\text{response} > \\) tokens, determining whether to invoke the LLM. Since the response label is much sparser than silence (e.g. 310:1), we introduce a balancing weight into" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.138 + ], + "angle": 0, + "content": "the standard Cross Entropy loss during training to balance the labels. Refer to Sec. 3.1 and Appendix Sec. 1 for more training details." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.149, + 0.71, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.149, + 0.906, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.252, + 0.906, + 0.336 + ], + "angle": 0, + "content": "Figure 4. The two-stage training process of STREAMMIND. In the first stage, the EPFE spatiotemporal extractor and the LLM are jointly trained to align their representations. In the second stage, the Cognition module is trained to generate response/silence tokens. Both stages follow a supervised autoregressive training approach for language modeling." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.364, + 0.639, + 0.381 + ], + "angle": 0, + "content": "3. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.388, + 0.722, + 0.405 + ], + "angle": 0, + "content": "3.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.906, + 0.516 + ], + "angle": 0, + "content": "Baseline and Datasets Since existing works focusing on streamingVD [8, 26, 60], except for VideoLLM-Online [8], do not provide open-source data or training methods, to demonstrate our outstanding performance, we re-implemented all previous works (VideoLLM-Online [8] and VideoLLM-MOD [60]) in the StreamingVD domain as baselines, using consistent settings across all experiments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.906, + 0.683 + ], + "angle": 0, + "content": "For dataset selection, since the Ego4D dataset [21] is inherently collected in a streaming manner, containing various queries, real-time responses, and narrations, and has been used as an main evaluation dataset in works [8, 31, 60], we selected it as one of our evaluation datasets. Furthermore, to further demonstrate the effectiveness of our approach on long videos, we also conducted experiments on the SoccerNet-Caption dataset [20]. This dataset includes 471 complete soccer match videos with a total duration of 715 hours, where each video is approximately 45 minutes long and provides commentary at appropriate timestamps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.839 + ], + "angle": 0, + "content": "Implementation Details: Model training was conducted on videos sampled at 2 fps, using 8 NVIDIA A100 GPUs. Each stage of the training process was conducted for one epoch. To optimize the learning rate, we applied the CosineAnnealingLR scheduler, with learning rates set to \\(2e - 5\\) and \\(2e - 6\\) respectively. Additionally, specific techniques were applied to the design of the Cognition Gate and methods to address the imbalance in the Silence-Response ratio, which are detailed in Section 4 and the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.682, + 0.864 + ], + "angle": 0, + "content": "3.2. Evaluate Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "To evaluate the effectiveness of STREAMMIND, we focus on assessing its Timing Alignment Capability and Lan" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13452" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.386, + 0.09, + 0.61, + 0.104 + ], + "angle": 0, + "content": "Table 1. Online Experiments Results." + }, + { + "type": "table", + "bbox": [ + 0.151, + 0.11, + 0.849, + 0.248 + ], + "angle": 0, + "content": "
DataSetMethodTiming Alignment ↑Language Modeling Capability ↑Training Cost
TriggerAccTimValBLUE-1BLUE-4METEORROUGE-L
Ego4dVideoLLM-Online32.34%29.66%66.0135.2531.1263.0624h
VideoLLM-MoD32.36%29.65%65.3435.2130.6563.0214h
Ours43.34%39.73%67.1239.2631.6065.71(11+8)h
SoccerVideoLLM-Online31.25%28.34%75.3664.2350.9281.5712h
VideoLLM-MoD31.24%28.12%74.9664.1850.2481.597h
Ours52.18%47.36%82.7866.7051.4382.04(5+3)h
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.27, + 0.483, + 0.327 + ], + "angle": 0, + "content": "Table 2. Online Experiments Results. LION-FS was released on arXiv just two days before our submission, but its code was not released. Therefore, we could only obtain the Ego4D results based on its paper." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.332, + 0.48, + 0.445 + ], + "angle": 0, + "content": "
DataSetMethodTimeDiff ↓Fluency ↑PPL ↓Correctiness ↑
Ego4dVideoLLM-Online2.0445.0%2.4348.1%
VideoLLM-MOD2.0445.2%2.4148.9%
LION-FS2.1546.1%2.0952.4%
Ours1.8960.2%2.0277.3%
SoccerVideoLLM-Online15.6246.35%1.7953.5%
VideoLLM-MOD15.6845.34%1.8053.3%
Ours14.0270.35%1.5989.2%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.457, + 0.483, + 0.501 + ], + "angle": 0, + "content": "Table 3. Results on COIN benchmarks(left to right): step recognition, task recognition, next forecasting, procedure forecasting, procedure forecasting with a goal." + }, + { + "type": "table", + "bbox": [ + 0.096, + 0.505, + 0.481, + 0.646 + ], + "angle": 0, + "content": "
MethodNot use HT 100MCOINBenchmark Top-1 Accuracy ↑
StepTaskNextProcProg.+
ClipBERT✓30.865.4---
TimeFormer×46.585.334.017.040.1
Paprika×51.085.843.2--
DistantSup×54.190.039.4--
VideoTF×56.591.042.440.246.4
ProcedureVRL×56.990.846.8--
VideoTaskGraph×57.290.540.2--
VideoLLM-online-7B-v1✓59.892.148.147.952.9
VideoLLM-online -8B-v1+✓63.192.749.149.854.1
VideoLLM-MOD✓63.492.849.749.853.3
Ours✓63.793.249.949.854.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.689 + ], + "angle": 0, + "content": "Table 4. Results on Ego4DLTA benchmark, evaluated on public server. ED@Z=20 denotes editing distance for future 20 actions." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.694, + 0.479, + 0.815 + ], + "angle": 0, + "content": "
MethodNot use Ego VLPEnd-to -End?Ego4D LTA ED@Z=20
Verb ↓Noun ↓Action ↓
CLIP✓✓0.7390.7690.941
EgoT2✓✓0.7220.7640.935
I-CVAE✓✓0.7530.7490.931
HierVL✓✓0.7240.7350.928
VideoLLM×✓0.7210.7250.921
VideoLLM-online-7B-v1✓✓0.6970.6980.897
VideoLLM-online-8B-v1+✓✓0.6890.6710.884
VideoLLM-MOD✓✓0.6890.6760.884
Ours✓✓0.6890.6550.881
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "guage Modeling Capability. Existing evaluation metrics include LM-PPL (Language Modeling Perplexity) and LM-Correctness, which evaluate language modeling accuracy at specific timestamps. Additionally, the Time Difference" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.272, + 0.907, + 0.334 + ], + "angle": 0, + "content": "(TimeDiff) metric is used to measure the temporal alignment of responses in an streaming video LLM, while the Fluency metric aims to provide a holistic evaluation of both linguistic quality and temporal consistency." + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.334, + 0.855, + 0.349 + ], + "angle": 0, + "content": "However, these metrics have several limitations:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.35, + 0.907, + 0.425 + ], + "angle": 0, + "content": "- LM-PPL and LM-Correctness focus solely on the exact alignment between model outputs and reference narrations, ignoring aspects such as synonym substitutions, sentence fluency, and the ability to capture critical information." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.426, + 0.906, + 0.486 + ], + "angle": 0, + "content": "- TimeDiff, as a metric for timing alignment Capability, evaluates temporal alignment only within a single dialogue turn. It does not assess the Cognition Gate's overall decision-making across the entire video stream." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.35, + 0.907, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.487, + 0.906, + 0.517 + ], + "angle": 0, + "content": "To address these limitations, we introduce the following metrics:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.518, + 0.906, + 0.564 + ], + "angle": 0, + "content": "- Trigger Accuracy (TriggerAcc): To evaluate whether the streaming video LLM responds at the correct time steps during entire streaming video." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.564, + 0.906, + 0.639 + ], + "angle": 0, + "content": "- Timing Validity (TimVal): To comprehensively evaluate whether the streaming video LLM consistently makes the correct decisions throughout the streaming video—speaking when necessary and remaining silent when appropriate." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.639, + 0.906, + 0.745 + ], + "angle": 0, + "content": "- BLUE [43], METEOR [2], ROUGE-L [34]: Widely used evaluation metrics in the field of text generation [6, 9, 38] and image narration [7, 44, 48], which fully consider factors such as synonym matching, stemming, word order, and key semantic expressions. These metrics enable a comprehensive assessment of dialogues in the StreamingVD process." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.518, + 0.906, + 0.745 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.757, + 0.906, + 0.788 + ], + "angle": 0, + "content": "3.3. Comprehensive Evaluation on Streaming and Offline Video" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Online Experiments: We compared our model with all existing streaming dialogue models [8, 31, 60] in Tab 1 and Tab 2. Under comparable training costs, experiments on real-time Ego4D and SoccerNet demonstrate superior performance across all metrics. Additionally, STREAMMIND can process video streams at a \\(10 \\times\\) higher FPS. This improvement is attributed to the event-level modeling capability." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13453" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.152 + ], + "angle": 0, + "content": "ity of EPFE and the strong timing alignment capability of the cognition gate in the perception stage, which together provide the LLM in the cognition phase with a concise yet comprehensive representation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.159, + 0.483, + 0.281 + ], + "angle": 0, + "content": "Offline Experiments: We also leveraged the COIN [53] and Ego4D LTA [24] benchmark to demonstrate STREAM-MIND effectiveness in traditional offline video scenarios, covering six common benchmarks: step recognition, step forecasting, task summarization, procedure forecasting, and procedure forecasting with a goal. As shown in Tab 3 and Tab 4, our approach achieves state-of-the-art performance across all tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.289, + 0.36, + 0.306 + ], + "angle": 0, + "content": "3.4. Real-time Inference Efficiency" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.31, + 0.483, + 0.415 + ], + "angle": 0, + "content": "Standard frame rates for films are typically 24 FPS, which is considered the minimum requirement for smooth motion perception. In contrast, games often operate at much higher frame rates, ranging from 30 FPS to over 100 FPS, to ensure a immersive experience. To validate the real-time performance of STREAMMIND, we conducted inference efficiency evaluations on NVIDIA A100 and H100 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.416, + 0.483, + 0.506 + ], + "angle": 0, + "content": "Specifically, we compared VideoLLM-Online, VideoLLM-MOD, and STREAMMIND by measuring their actual processing time for 1-second video across input frame rates from 5 to 100 FPS. As shown in Figure 5, a processing time under 1 second indicates real-time performance without disrupting playback." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.483, + 0.613 + ], + "angle": 0, + "content": "The results reveal that VideoLLM-Online and VideoLLM-MOD struggle to process video streams beyond 10 FPS, leading to potential latency issues in real-world applications. In contrast, our method achieves true real-time performance, efficiently handling both standard film/TV frame rates and highRefresh gaming scenarios up to 100 FPS, ensuring real-time responses." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.625, + 0.192, + 0.641 + ], + "angle": 0, + "content": "4. Ablation" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.649, + 0.483, + 0.667 + ], + "angle": 0, + "content": "4.1. Impact of Silence-Response Sample Imbalance" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.671, + 0.483, + 0.734 + ], + "angle": 0, + "content": "In this section, we explored the imbalance in the Silence-Response ratio within the dataset and derived an empirical balancing formula through experiments. For detailed experimental results, please refer to the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.74, + 0.371, + 0.758 + ], + "angle": 0, + "content": "4.2. The Ablation of Cognition Gate" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.762, + 0.483, + 0.809 + ], + "angle": 0, + "content": "The Cognition Gate plays a crucial role throughout the perception-cognition interleaving process. We conducted detailed ablation experiments to analyze its impact." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.903 + ], + "angle": 0, + "content": "The Architecture of Cognition Gate The structure of the gate directly determines its capabilities. To this end, we constructed multiple classification networks using MLP projectors, Cross-Attention, and Transformers. [54] Additionally, we built a set of autoregressive networks using an" + }, + { + "type": "image", + "bbox": [ + 0.551, + 0.091, + 0.872, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.55, + 0.249, + 0.872, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.412, + 0.907, + 0.455 + ], + "angle": 0, + "content": "Figure 5. Comparison of running time for processing 1 second of streaming video frames using different methods on A100 (Top) and H100 (Bottom)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.477, + 0.907, + 0.523 + ], + "angle": 0, + "content": "LLM block. The results are shown in Tab.5. It is evident that the LLM-like structure achieved better performance. For a detailed analysis, please refer to Section 2.4." + }, + { + "type": "table_caption", + "bbox": [ + 0.529, + 0.534, + 0.892, + 0.549 + ], + "angle": 0, + "content": "Table 5. Ablation Study on the architecture of cognition gate" + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.552, + 0.902, + 0.632 + ], + "angle": 0, + "content": "
ArchitectureTimeDiff ↓TriggerAcc ↑TimVal ↑
Linear Layer5.0620.13%17.65%
MLP projector + Linear4.3421.75%18.33%
Transformers + Linear3.7821.58%18.97%
Cross-Attention + Linear3.6424.34%20.36%
Single LLM block2.2235.35%32.11%
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.655, + 0.907, + 0.837 + ], + "angle": 0, + "content": "Initialization Strategies We investigate three initialization strategies for the Cognition Gate to enhance its effectiveness: (i) Random Initialization: All layers are initialized from a standard random distribution without utilizing prior knowledge from the LLM. (ii) SkipBlock Initialization: The Cognition Gate is initialized by uniformly selecting layers from the LLM, with the number of skipped layers determined by the ratio of the Cognition Gate's depth to the total number of LLM layers. (iii) EarlyBlock Initialization: The first few layers of the LLM are directly used to initialize the Cognition Gate, leveraging the LLM's early-stage processing capabilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Number of Blocks The Cognition Gate consists of multiple layers, and its depth plays a crucial role in balancing computational efficiency and perceptual accuracy." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13454" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.12, + 0.09, + 0.454, + 0.104 + ], + "angle": 0, + "content": "Table 6. Ablation Study on the Strategy of Initialization" + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.11, + 0.485, + 0.168 + ], + "angle": 0, + "content": "
Strategy of InitializationTimeDiff ↓TriggerAcc ↑TimVal ↑
Random2.1039.65%37.35%
SkipBlock1.8940.23%37.67%
EarlyBlock1.8943.34%39.73%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.12, + 0.18, + 0.454, + 0.194 + ], + "angle": 0, + "content": "Table 7. Ablation Study on the Strategy of Initialization" + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.199, + 0.481, + 0.28 + ], + "angle": 0, + "content": "
#Num of BlockTimeDiff ↓TriggerAcc ↑TimVal ↑
21.9241.53%37.65%
31.9042.35%38.34%
41.8943.34%39.73%
51.8842.67%38.56%
61.8942.56%38.34%
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.304, + 0.484, + 0.425 + ], + "angle": 0, + "content": "The results in Tab.6 and Tab.7 show that fine-tuning the shallow layers of the LLM yields the best performance. We attribute this to the fact that gate operates within a limited decision space (binary: Yes/No), making deeper layers—primarily responsible for nuanced language generation—unnecessary. Thus, we propose the Shallow Layer Transfer method to construct the Cognition Gate, as detailed in Section 2.4." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.436, + 0.484, + 0.466 + ], + "angle": 0, + "content": "4.3. The Performance of Event-preserving Feature extractor" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.472, + 0.484, + 0.655 + ], + "angle": 0, + "content": "To evaluate the role of EPFE in the perception phase, we compare it with commonly used extractors in VideoLLMs, including: 1) The Q-Former [28]. 2) The Spatial-Temporal Convolution Connector (STC Connector) [12], which has demonstrated outstanding performance in videolm and has been widely adopted [29, 59, 71]. Notably, STC employs 3D convolution for spatial-temporal aggregation. To preserve local visual patterns during spatial compression, it follows work [5] by incorporating a RegStage block [47] before and after the 3D convolution, which has been shown to enhance spatial understanding. Experimental results indicate that EPFE achieves superior performance." + }, + { + "type": "table_caption", + "bbox": [ + 0.162, + 0.657, + 0.412, + 0.671 + ], + "angle": 0, + "content": "Table 8. Perceptive Judgement Capability" + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.676, + 0.465, + 0.735 + ], + "angle": 0, + "content": "
MethodTimeDiff ↓TriggerAcc ↑TimVal ↑
Q-former3.7826.65 %25.31%
STC3.5627,54%26.87%
EPFE1.8943.34%39.73%
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.763, + 0.234, + 0.779 + ], + "angle": 0, + "content": "5. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.282, + 0.803 + ], + "angle": 0, + "content": "5.1. Offline VideoLLMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In recent years, there has been a surge of interest in leveraging Large Language Models (LLMs) for video understanding, leading to the development of offline videoLLMs. The biggest challenge for it is managing memory requirements [4, 11, 15, 23, 50, 51] and compressing redundant frame features [12, 25, 27, 52, 76]. LaViLa [37] addresses" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.228 + ], + "angle": 0, + "content": "this challenge by transforming LLMs into narrators that generate detailed descriptions of long videos from visual inputs. Other models, such as ChatVideo [56] and MMVID [36], convert videos into text to improve comprehension. MovieChat [50] combines all frame features using an averaging strategy. However, in the context of online streaming, there has been limited exploration into how VideoLLMs can meet the demands of temporal alignment and real-time processing for streaming video inputs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.248, + 0.704, + 0.263 + ], + "angle": 0, + "content": "5.2. Online VideoLLMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.271, + 0.908, + 0.574 + ], + "angle": 0, + "content": "Recent studies have introduced VideoLLMs specifically designed for online stream understanding, while most existing works [14, 42, 45, 62, 72, 77] focus on improving offline video efficiency to meet real-time requirements. However, they fail to proactively decide when to respond based on user requirements. For example, ReKV [14] implemented a sliding-window attention mechanism and a KV-cache system to reduce computational overhead. To address this, VideoLLM-Online [8] introduces per-step LLM invocation for online video dialogue, as shown in Fig.1. However, its performance is constrained by limited per-frame processing capability due to the absence of effective frame compression strategies. VideoLLM-MOD [60] mitigates this issue by incorporating a mixture-of-depth approach for efficient visual token computation, enabling higher visual input resolutions. Building on MOD, LION-FS [31] proposes Token Aggregation Router and Token DroppingRouter as the Fast Path, adaptively aggregating distinct features while discarding redundant ones. However, they remain limited to processing videos at only 10 FPS." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.578, + 0.909, + 0.73 + ], + "angle": 0, + "content": "Overall, prior work remains restricted by the contradiction between sequential full-frame processing and the extremely high efficiency requirements of streaming video, lacking breakthrough paradigm designs explicitly suited for streaming frame inputs. To address these challenges, our approach introduces a novel perception-cognition interleaving paradigm and proposes the STREAMMIND framework, enabling VideoLLM to achieve proactive responses and real-time streaming video understanding, with a maximum throughput of up to 100 FPS." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.754, + 0.634, + 0.769 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this paper, we introduce a novel perception-cognition interleaving paradigm called \"event-gated LLM invocation\" and propose the STREAMMIND framework, addressing the contradiction between sequential full-frame processing and the extremely high efficiency requirements of streaming video. As a result, STREAMMIND can process video at up to 100 FPS, achieving the SOTA performance across all existing metrics and benchmarks." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13455" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.273, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.486, + 0.271 + ], + "angle": 0, + "content": "We gratefully acknowledge the support from the following foundations and institutions. Ting Cao was supported in part by the Tsinghua University (AIR)-AsiaInfo Technologies (China), Inc. Joint Research Center for 6G Network and Intelligent Computing. Hao Wu was supported in part by the National Natural Science Foundation of China (NSFC) under Grants 62432004 and 62302207, and by the Science Foundation for Youths of Jiangsu Province under Grant BK20220772. Xin Ding and Zhibo Chen were supported in part by the NSFC under Grants 62371434 and 62021001." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.296, + 0.188, + 0.312 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.321, + 0.482, + 0.378 + ], + "angle": 0, + "content": "[1] Roberto Amoroso, Gengyuan Zhang, Rajat Koner, Lorenzo Baraldi, Rita Cucchiara, and Volker Tresp. Perceive, query & reason: Enhancing video qa with question-guided temporal queries. arXiv preprint arXiv:2412.19304, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.379, + 0.482, + 0.448 + ], + "angle": 0, + "content": "[2] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.45, + 0.482, + 0.533 + ], + "angle": 0, + "content": "[3] Pietro Battistoni, Andrea Antonio Cantone, Mariarosaria Esposito, Rita Francese, Francesca Pia Perillo, Marco Romano, Monica Sebillo, and Giuliana Vitiello. Using artificial intelligence and companion robots to improve home healthcare for the elderly. In International Conference on Human-Computer Interaction, pages 3–17. Springer, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.535, + 0.482, + 0.603 + ], + "angle": 0, + "content": "[4] Jiarui Cai, Mingze Xu, Wei Li, Yuanjun Xiong, Wei Xia, Zhuowen Tu, and Stefano Soatto. Memot: Multi-object tracking with memory. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8090-8100, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.605, + 0.482, + 0.674 + ], + "angle": 0, + "content": "[5] Junbum Cha, Wooyoung Kang, Jonghwan Mun, and Byungseok Roh. Honeybee: Locality-enhanced projector for multimodal llm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13817-13827, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.676, + 0.482, + 0.745 + ], + "angle": 0, + "content": "[6] Yupeng Chang, Xu Wang, Jindong Wang, Yuan Wu, Linyi Yang, Kaijie Zhu, Hao Chen, Xiaoyuan Yi, Cunxiang Wang, Yidong Wang, et al. A survey on evaluation of large language models. ACM transactions on intelligent systems and technology, 15(3):1-45, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.747, + 0.482, + 0.817 + ], + "angle": 0, + "content": "[7] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3558-3568, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.818, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[8] Joya Chen, Zhaoyang Lv, Shiwei Wu, Kevin Qinghong Lin, Chenan Song, Difei Gao, Jia-Wei Liu, Ziteng Gao, Dongxing Mao, and Mike Zheng Shou. Videollm-online: Online video large language model for streaming video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18407-18418, 2024. 1, 2, 3, 5, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.321, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[9] Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji, and Quanquan Gu. Self-play fine-tuning converts weak language models to strong language models. arXiv preprint arXiv:2401.01335, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[10] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24185–24198, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.904, + 0.288 + ], + "angle": 0, + "content": "[11] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In European Conference on Computer Vision, pages 640-658. Springer, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.289, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[12] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[13] Shangzhe Di and Weidi Xie. Grounded question-answering in long egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12934–12943, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.414, + 0.905, + 0.483 + ], + "angle": 0, + "content": "[14] Shangzhe Di, Zhelun Yu, Guanghao Zhang, Haoyuan Li, Hao Cheng, Bolin Li, Wanggui He, Fangxun Shu, Hao Jiang, et al. Streaming video question-answering with in-context video kv-cache retrieval. In The Thirteenth International Conference on Learning Representations. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.905, + 0.55 + ], + "angle": 0, + "content": "[15] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. In European Conference on Computer Vision, pages 75-92. Springer, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.553, + 0.905, + 0.608 + ], + "angle": 0, + "content": "[16] Feiteng Fang, Yuelin Bai, Shiwen Ni, Min Yang, Xiaojun Chen, and Ruifeng Xu. Enhancing noise robustness of retrieval-augmented language models with adaptive adversarial training. arXiv preprint arXiv:2405.20978, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[17] Xinyu Fang, Kangrui Mao, Haodong Duan, Xiangyu Zhao, Yining Li, Dahua Lin, and Kai Chen. Mmbench-video: A long-form multi-shot benchmark for holistic video understanding. Advances in Neural Information Processing Systems, 37:89098-89124, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[18] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19358-19369, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.762, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[19] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[20] Silvio Giancola, Mohieddine Amine, Tarek Dghaily, and Bernard Ghanem. Soccernet: A scalable dataset for action spotting in soccer videos. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 1711-1721, 2018. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13456" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.188 + ], + "angle": 0, + "content": "[21] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18995-19012, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[22] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. CoRR, abs/2312.00752, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[23] Bo He, Hengduo Li, Young Kyun Jang, Menglin Jia, Xuefei Cao, Ashish Shah, Abhinav Shrivastava, and Ser-Nam Lim. Ma-lmm: Memory-augmented large multimodal model for long-term video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13504-13514, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.319, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[24] Daoji Huang, Otmar Hilliges, Luc Van Gool, and Xi Wang. Palm: Predicting actions through language models@ ego4d long-term action anticipation challenge 2023. arXiv preprint arXiv:2306.16545, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[25] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In European Conference on Computer Vision, pages 202-218. Springer, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[26] Zhenpeng Huang, Xinhao Li, Jiaqi Li, Jing Wang, Xi-angyu Zeng, Cheng Liang, Tao Wu, Xi Chen, Liang Li, and Limin Wang. Online video understanding: A comprehensive benchmark and memory-augmented method. arXiv preprint arXiv:2501.00584, 2024. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.518, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[27] Yang Jin, Zhicheng Sun, Kun Xu, Liwei Chen, Hao Jiang, Quzhe Huang, Chengru Song, Yuliang Liu, Di Zhang, Yang Song, et al. Video-lavit: Unified video-language pre-training with decoupled visual-motional tokenization. arXiv preprint arXiv:2402.03161, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.589, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.66, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[29] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.717, + 0.482, + 0.799 + ], + "angle": 0, + "content": "[30] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[31] Wei Li, Bing Hu, Rui Shao, Leyang Shen, and Liqiang Nie. Lion-fs: Fast & slow video-language thinker as online video assistant. arXiv preprint arXiv:2503.03663, 2025. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[32] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2024. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[33] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[34] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[35] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 26689-26699, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[36] Kevin Lin, Faisal Ahmed, Linjie Li, Chung-Ching Lin, Ehsan Azarnasab, Zhengyuan Yang, Jianfeng Wang, Lin Liang, Zicheng Liu, Yumao Lu, et al. Mm-vid: Advancing video understanding with gpt-4v (isdiction). arXiv preprint arXiv:2310.19773, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[37] Xingyu Liu, Joon-Young Lee, and Hailin Jin. Learning video representations from correspondence proposals. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 4273–4281, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[38] Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. G-eval: Nlg evaluation using gpt-4 with better human alignment. arXiv preprint arXiv:2303.16634, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[39] Ruipu Luo, Ziwang Zhao, Min Yang, Junwei Dong, Da Li, Pengcheng Lu, Tao Wang, Linmei Hu, Minghui Qiu, and Zhongyu Wei. Valley: Video assistant with large language model enhanced ability. arXiv preprint arXiv:2306.07207, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[40] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[41] Kelly Merrill Jr, Jihyun Kim, and Chad Collins. Ai companions for lonely individuals and the role of social presence. Communication Research Reports, 39(2):93-103, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[42] Sunny Panchal, Apratim Bhattacharyya, Guillaume Berger, Antoine Mercier, Cornelius Böhm, Florian Dietrichkeit, Reza Pourreza, Xuanlin Li, Pulkit Madan, Mingu Lee, et al. What to say and when to say it: Live fitness coaching as a testbed for situated interaction. Advances in Neural Information Processing Systems, 37:75853-75882, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[43] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.905, + 0.856 + ], + "angle": 0, + "content": "[44] Jordi Pont-Tuset, Jasper Uijlings, Soravit Changpinyo, Radu Soricut, and Vittorio Ferrari. Connecting vision and language with localized narratives. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 647-664. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[45] Rui Qian, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Shuanggrui Ding, Dahua Lin, and Jiaqi Wang. Streaming long video understanding with large language models. Advances in Neu" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13457" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "ral Information Processing Systems, 37:119336-119360, 2025. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[47] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólár. Designing network design spaces. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10428-10436, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[48] Jiayuan Rao, Haoning Wu, Chang Liu, Yanfeng Wang, and Weidi Xie. Matchtime: Towards automatic soccer game commentary generation. arXiv preprint arXiv:2406.18530, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[49] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[50] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[51] Larry R Squire, Lisa Genzel, John T Wixted, and Richard G Morris. Memory consolidation. Cold Spring Harbor perspectives in biology, 7(8):a021766, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.601 + ], + "angle": 0, + "content": "[52] Guangzhi Sun, Wenyi Yu, Changli Tang, Xianzhao Chen, Tian Tan, Wei Li, Lu Lu, Zejun Ma, and Chao Zhang. Fine-grained audio-visual joint representations for multimodal large language models. arXiv preprint arXiv:2310.05863, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.603, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[53] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1207-1216, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.688, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.745, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[55] Ales Vysocky and Petr Novak. Human-robot collaboration in industry. MM Science Journal, 9(2):903–906, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[56] Junke Wang, Dongdong Chen, Chong Luo, Xiyang Dai, Lu Yuan, Zuxuan Wu, and Yu-Gang Jiang. Chatvideo: A tracklet-centric multimodal and versatile video understanding system. arXiv preprint arXiv:2304.14407, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[57] Qing Wang, Jiaming Zhang, Kailun Yang, Kunyu Peng, and Rainer Stiefelhagen. Matchformer: Interleaving attention in transformers for feature matching. In Proceedings of the Asian Conference on Computer Vision, pages 2746-2762, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[58] Xunguang Wang, Zhenlan Ji, Pingchuan Ma, Zongjie Li, and Shuai Wang. Instructta: Instruction-tuned targeted attack for large vision-language models. arXiv preprint arXiv:2312.01886, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.207 + ], + "angle": 0, + "content": "[59] Yuxuan Wang, Yueqian Wang, Pengfei Wu, Jianxin Liang, Dongyan Zhao, and Zilong Zheng. Lstp: Language-guided spatial-temporal prompt learning for long-form video-text understanding. arXiv e-prints, pages arXiv-2402, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[60] Shiwei Wu, Joya Chen, Kevin Qinghong Lin, Qimeng Wang, Yan Gao, Qianli Xu, Tong Xu, Yao Hu, Enhong Chen, and Mike Zheng Shou. Videollm-mod: Efficient video-language streaming with mixture-of-depths vision computation. Advances in Neural Information Processing Systems, 37:109922-109947, 2025. 1, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[61] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9777–9786, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[62] Haomiao Xiong, Zongxin Yang, Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Jiawen Zhu, and Huchuan Lu. Streaming video understanding and multi-round interaction with memory-enhanced knowledge. arXiv preprint arXiv:2501.13468, 2025. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[63] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the 25th ACM international conference on Multimedia, pages 1645-1653, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[64] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.907, + 0.631 + ], + "angle": 0, + "content": "[65] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1686-1697, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.688 + ], + "angle": 0, + "content": "[66] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. Advances in Neural Information Processing Systems, 35:124-141, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[67] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10714-10726, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[68] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9127–9134, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[69] Jeffrey M Zacks, Nicole K Speer, Khena M Swallow, Todd S Braver, and Jeremy R Reynolds. Event perception: a mind-brain perspective. Psychological bulletin, 133(2):273, 2007. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13458" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[70] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975-11986, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[71] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[72] Haoji Zhang, Yiqin Wang, Yansong Tang, Yong Liu, Jiashi Feng, Jifeng Dai, and Xiaojie Jin. Flash-vstream: Memory-based real-time understanding for long video streams. arXiv preprint arXiv:2406.08085, 2024. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.316 + ], + "angle": 0, + "content": "[73] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.319, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[74] Ruohong Zhang, Liangke Gui, Zhiqing Sun, Yihao Feng, Keyang Xu, Yuanhan Zhang, Di Fu, Chunyuan Li, Alexander Hauptmann, Yonatan Bisk, et al. Direct preference optimization of video large multimodal models from language model reward. arXiv preprint arXiv:2404.01258, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.482, + 0.444 + ], + "angle": 0, + "content": "[75] Ruoyu Zhang, Lulu Wang, Yi He, Tongling Pan, Zhengtao Yu, and Yingna Li. Tpcap: Unlocking zero-shot image captioning with trigger-augmented and multi-modal purification modules. arXiv preprint arXiv:2502.11024, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.446, + 0.482, + 0.514 + ], + "angle": 0, + "content": "[76] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6586-6597, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.517, + 0.482, + 0.585 + ], + "angle": 0, + "content": "[77] Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. Streaming dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18243-18252, 2024. 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "13459" + } + ] +] \ No newline at end of file diff --git a/2025/StreamMind_ Unlocking Full Frame Rate Streaming Video Dialogue through Event-Gated Cognition/61e82f55-a3d1-4244-b3cc-22632c8cab6f_origin.pdf b/2025/StreamMind_ Unlocking Full Frame Rate Streaming Video Dialogue through Event-Gated Cognition/61e82f55-a3d1-4244-b3cc-22632c8cab6f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f65751082bcc09555010688d199a7938089aacee --- /dev/null +++ b/2025/StreamMind_ Unlocking Full Frame Rate Streaming Video Dialogue through Event-Gated Cognition/61e82f55-a3d1-4244-b3cc-22632c8cab6f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:937d9f77f591f717fbd9fe557d6b217627d743fac60039abbd7683f3bb69bba3 +size 929819 diff --git a/2025/StreamMind_ Unlocking Full Frame Rate Streaming Video Dialogue through Event-Gated Cognition/full.md b/2025/StreamMind_ Unlocking Full Frame Rate Streaming Video Dialogue through Event-Gated Cognition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..380c57ed8c2dfffdcbbf975a08ea3ec6f6c65cb4 --- /dev/null +++ b/2025/StreamMind_ Unlocking Full Frame Rate Streaming Video Dialogue through Event-Gated Cognition/full.md @@ -0,0 +1,367 @@ +# STREAMMIND: Unlocking Full Frame Rate Streaming Video Dialogue through Event-Gated Cognition + +Xin Ding $^{1}$ * Hao Wu $^{3}$ † Yifan Yang $^{2}$ Shiqi Jiang $^{2}$ Qianxi Zhang $^{2}$ Donglin Bai $^{2}$ Zhibo Chen $^{1}$ Ting Cao $^{4}$ † + +1University of Science and Technology of China 2Microsoft Research 3Nanjing University Institute for AI Industry Research (AIR), Tsinghua University + +# Abstract + +With the rise of real-world human-AI interaction applications, such as AI assistants, the need for Streaming Video Dialogue is critical. To address this need, we introduce STREAMMIND, a video LLM framework that achieves ultra-FPS streaming video processing (100fps on a single A100) and enables proactive, always-on responses in real time, without explicit user intervention. + +To solve the key challenge of the contradiction between linear video streaming speed and quadratic transformer computation cost, we propose a novel perception-cognition interleaving paradigm named "event-gated LLM invocation", in contrast to the existing per-time-step LLM invocation. By introducing a Cognition Gate network between the video encoder and the LLM, LLM is only invoked when relevant events occur. To realize the event feature extraction with constant cost, we propose Event-Preserving Feature Extractor (EPFE) based on state-space method, generating a single perception token for spatiotemporal features. These techniques enable the video LLM with full-FPS perception and real-time cognition response. + +Experiments on Ego4D and SoccerNet streaming tasks, as well as standard offline benchmarks, demonstrate state-of-the-art performance in both model capability and real-time efficiency, paving the way for ultra-high-FPS applications, such as Game AI Copilot and interactive media. The code and data is available at https://aka.ms/StreamMind. + +# 1. Introduction + +The advancement of large foundation models is driving an increasing number of real-world human-AI interaction applications, such as AI home companions [3, 41] and human-robot collaboration [55, 64]. These applications + +![](images/3adaf8afee610aaa13e2cd3ca26d5864a019fdf770a722f6606ba269a692111a.jpg) +Existing Works 10~15 FPS on A100 GPU +Figure 1. The paradigm of existing per-step LLM invocation (left) versus our event-gated LLM invocation (right) for streaming video dialogue task. Our paradigm boosts frame rate up-to 100 fps on A100. $(R_{i}$ : response, $f_{i}$ : frame, t: time) + +![](images/35b51e882842b88a1cdd620d9ce8e208cbaae50c948226330f8044f9927ddb7d.jpg) +Ours Up to 100 FPS on A100 GPU + +heavily rely on streaming video understanding capabilities [14, 26, 45, 62, 72], specifically streaming video dialogue [8, 60]. Streaming video dialogue (StreamingVD) aims to continuously perceive incoming video streams and, based on user queries, provide proactive, real-time, and always-on responses without human intervention. For example, an AI companion could respond to user queries like "Guide me to fix this bicycle" or "Let's watch and discuss the soccer game together". + +Streaming video dialogue can be viewed as a generalization of standard offline video dialogue tasks [30, 32, 39, 40, 49, 66, 67], extending video LLMs to broader real-world scenarios. The fundamental challenge of StreamingVD compared to its offline counterpart lies in the requirement for timing alignment between events and responses. That is, as the video progresses, the model is required to (1) proactively decide when to respond and (2) respond in real time before the next event occurs according to user queries, as illustrated in Fig. 2. However, with current techniques, these two requirements are contradictory and cannot be achieved simultaneously. + +The seminal VideoLLM-Online [8] and VideoLLM-MoD [60] introduce a per-step LLM invocation paradigm for StreamingVD, shown in Fig. 1. At each time step, it inputs all past frames and the user query to the LLM, to decide when to respond or remain silent. However, + +the $O(n^3)^1$ computational complexity and limited context window are mismatched with the $O(n)$ frame streaming, making real-time response difficult to achieve. More works [14, 45, 62, 72] focus on enhancing the efficiency of offline video dialogue aiming to meet real-time requirements. However, similar to offline approaches, they depend on user queries to trigger responses rather than the model to do so proactively. + +To empower StreamingVD, this paper proposes a new perception-cognition interleaving paradigm, named event-gated LLM invocation, shown in Fig. 1. It is inspired by the event perception mechanism of human brain [69], which states that events are key components of human's perception, attention, and memory $^2$ . We therefore propose an event-based selective cognition process. Rather than invoking the LLM at every time step, we introduce a Cognition Gate between the video encoder and LLM. The video encoder continuously perceives visual signals of each frame, and only when query-related events occur, the gate will open to invoke the LLM. Analogous to how a text prompt triggers an LLM, streaming visual signals lack a clear stopping point, unlike a finite text sequence. The cognition gate bridges this gap by acting as a mechanism to identify "physical world prompts" for the LLM. + +This new paradigm raises two major technical concerns, to achieve both high capability and efficiency. Firstly, existing video encoders in VideoLLMs (e.g., STC in VideoLaMA) only capture local spatio-temporal features. This results in a linear increase in the number of feature tokens for streaming frames and subsequently a quadratic computation for the processing, failing to meet real-time requirements. Secondly, the Cognition Gate must decide to invoke the LLM or not, based on visual features and the user query. However, only considering efficiency and simply attending visual and query tokens (e.g., Cross Attention in Qformer [28]) limits the gate capability to tasks like feature-matching [57, 58] and retrieval [1, 16, 75], struggle to make decisions with deeper semantic understanding. + +By addressing the challenges, we realize STREAM-MIND, a novel framework for streamingVD, with the processing speed matching frame rate for the first time. It integrates two techniques. For the video encoder, inspired by the exceptional ability of the state-assisted module [22] in modeling continuous physical signals with constant cost, we propose the Event-Preserving Feature Extractor (EPFE) based on the state-assisted module. It generates a single perception token of event features to the gate, resulting in a constant perception cost. For the Cognition Gate, to leverage the world knowledge of LLM while meeting the real- + +time requirement, we propose a Shallow Layer Transfer method. The Gate is designed to reuse the shallow layers of the LLM and its autoregressive training to maximize the probabilities of “< response/silence >” tokens during training. + +To evaluate the performance of the STREAMMIND framework, we adopt evaluation metrics from VideoLLM-Online and a wide range of text generation evaluation metrics. Additionally, we design two novel metrics to comprehensively assess the model's temporal responsiveness. Experiments on real-time Ego4D [21] and SoccerNet [20] demonstrate superior performance across all metrics. Furthermore, our model achieves state-of-the-art results on multiple offline benchmarks, including short- and long-term activity recognition and forecasting on the COIN [53] and Ego4D LTA [24] datasets. Notably, our framework breaks the ultra-FPS streaming video processing bottleneck, enabling streamingVD at up to 100 fps on a single A100 GPU, laying a robust foundation for real-world human-AI interaction applications. + +# 2. Method + +# 2.1. Task Discussion and Definition + +Problem formulation of StreamingVD Given a video stream $\mathcal{V}^T \coloneqq [v_1, v_2, \dots, v_T]$ consisting of $T$ frames and a set of queries $\mathcal{Q} \coloneqq \{q_1, q_2, \dots, q_N\}$ . Traditional video dialogue, shown in Fig. 2, provides a response triggered by a user query. For a query $q_i$ given at time step $t_s$ , it responds based on past frames up to time $t_s (1 < t_s \leq T)$ , regardless of an online or offline scenario. On the other hand, StreamingVD proactively generates responses to a query given at $t_s$ at any appropriate time steps until the query termination at $t_e (t_s < t_e \leq T)$ . At each time step $t_i (t_s \leq t_i \leq t_e)$ , StreamingVD needs to decide whether to generate a response according to the past frames and query, and then generate the response, formulated as follows. + +$$ +\max P \left(\left[ \operatorname {R e s} ^ {t _ {i}} \right] \mid \left[ \operatorname {C t x} ^ {< t _ {i}} \right], \left[ F ^ {t _ {i}} \right]\right) \tag {1} +$$ + +where $[\mathrm{Res}^{t_i}] \in \{[\mathrm{Txt}^{t_i}], [\mathrm{EOS}^{t_i}]\}$ means generate a response at a time step $t_i$ , or remain silence, i.e., [EOS]. $[\mathrm{Ctx}^{< t_i}]$ denotes the contextual tokens accumulated from past frames and queries before $t_i$ . $[\mathrm{F}^{t_i}]$ represents the visual features extracted from the video frame at $t_i$ . + +Challenges of StreamingVD tasks Firstly, increasing number of frames versus real-time requirement. Different from the offline tasks[17, 19, 30, 63, 68], streaming tasks[8, 14, 26, 45, 62, 72] require time alignment between event and response. Besides only accuracy constraint of offline task, streaming video tasks need to consider both accuracy and timeliness constraints. Offline tasks can trade resources and latency for better quality, while streaming task cannot. + +![](images/2b7857db5e85331170fbe38d99a83bb0d8bf03235c857757b86617c572cdf93a.jpg) +Figure 2. Comparison between streaming dialogue (bottom) and traditional dialogue tasks (top). Streaming dialogue proactively generates responses at appropriate time based on user queries and perception, whereas traditional dialogue requires human triggering for each response. + +Secondly, proactive response versus real-time requirement. Traditional video dialogue [13, 14, 45, 61, 62, 65] are reactive, responding only at the time of user query. However, real-world dialogue is often without human triggering. The user may lack the specific knowledge to formulate questions. Similarly, embodied AI should anticipate user needs rather than merely reacting to commands. However, proactive response means at every frame the model needs to judge, inceasing the model cost by $\mathrm{O}(\mathfrak{n})$ times. + +Existing works Conventional offline Video LLMs [10, 32, 33, 35, 73, 74] typically employ a visual encoder [18, 46, 70] and a projection module [28, 39] to process uniformly sampled frames. The output is concatenated with the tokenized question to form a sequence, which is then passed to an LLM decoder to predict a response. However, this approach is impractical for streaming tasks due to the high computational cost associated with processing a large number of frames $(T)$ . Therefore, works such as [14, 62] have implemented a sliding-window attention mechanism and a KV-cache system to reduce computational overhead, or incorporated a parallel sampling strategy to enhance processing speed and reduce latency. However, these methods do not proactively trigger response when needed. In contrast, Videollm-Online [8] and MOD [60] utilize the LLM decoder to process every frame of streaming videos, resulting in a cubic computation cost. STREAMMIND, on the other hand, has the potential to process streaming videos at 100 fps, which is crucial for enabling interactive video analysis applications. + +# 2.2. The Overview of STREAMMIND + +For a streaming video sequence $\left[v^{t_s} \leq t_i \leq t_e\right]$ , STREAMMIND models the following process as shown in Fig. 3: + +1. Perception Phase: With continuous streaming video, each incoming frame $v$ is processed through the CLIP [46] encoder to extract spatial features. The features are then input to our proposed EPFE. EPFE updates its hidden state to correlate the frame's features across temporal dimensions and reduce redundant spatial information. This process generates a new perception token $[\mathsf{F}_{per}^{t_i}]$ and saves this token in the Perception Memory $\mathcal{M}_{\mathrm{per}}^{t_i}$ . The perception phase can be formulated as follows: + +$$ +\left[ \mathrm {F} _ {p e r} ^ {t _ {i}} \right] = \operatorname {E P F E} \left(\operatorname {C L I P} \left(v ^ {t _ {i}}\right), \mathrm {H} ^ {t _ {i} - 1}\right) \tag {2} +$$ + +where $\mathbb{H}^{t - 1}$ denotes the hidden state from the previous time. + +2. Cognition Gate Judgment: We introduce Cognition Gate $\mathcal{G}$ between perception and cognition. The input are user query [Prompt] and the generated perception token $[\mathsf{F}_{per}^{t_i}]$ at each time step $t_i$ . The Cognition Gate learns to determine whether to trigger the cognition phase at $t_i$ : + +$$ +\max P \left(\left[ \operatorname {R e s} ^ {t _ {i}} \right] \mid \mathcal {G} \left(\left[ \text {P r o m p t} \right], \left[ F _ {\text {p e r}} ^ {t _ {i}} \right]\right)\right) \tag {3} +$$ + +where $[\mathrm{Res}^{t_i}] \in \{< /\text{response}>, < /\text{silence}>\}$ determines whether to trigger a response or remain silent. + +3. Cognition Phase Upon receiving the trigger signal from the Cognition Gate, STREAMMIND samples perception tokens from the Perception Memory $\mathcal{M}_{\mathrm{per}}^{t_i}$ to the Cognition Pooling as the input to LLM to generate response. + +# 2.3. Perception Phase + +Motivation: To match the constant frame rate of streaming videos, ideally the perception phase should be constant computation cost for processing each frame. However, current video encoders generally extract local (dozens of frames) spatiotemporal features with constant cost, and input all the generated perception tokens to a transformer + +![](images/1eb6a4808df4538648aaa78b2ebe560f1b83446fe1685610cc8a0d9b28a29c2d.jpg) +Figure 3. STREAMMIND workflow. For each video frame, the perception stage goes through Clip for spatial features and the proposed EPFE for spatiotemporal features, to generate a single perception token saved in perception memory. The Cognition Gate determines whether to invoke LLM only based on the current perception token. If yes (a query-related event occurs), the tokens in perception memory are sampled to cognition pooling as the input for LLM for cognition response. The gate decouples perception and cognition to enable full-FPS perception and event-based cognition. + +model to extract the long-term temporal relationships. The computational cost is $O(n^{3})$ , obviously the lagger behind the streaming rate. + +To solve the issue, we propose Event-Preserving Feature Extractor (EPFE) to replace the projection module used in existing video LLMs (e.g., cross-attention and Q-Former). EPFE is based on the SOTA state-space model, i.e., Selective State Space model [22]. As shown in Fig. 3 (blue box) and Eq. 4, for each frame, EPFE dynamically correlates the spatial features and the past internal states to generate a single perception token $\left[\mathbb{F}_{per}^{t_i}\right]$ as well as update its state. This enables a matched speed of streaming video and perception processing, for full FPS StreamingVD. + +$$ +\mathbf {h} _ {t + 1} = \mathbf {A h} _ {t} + \mathbf {B x} _ {t}, \quad \mathbf {y} _ {t} = \mathbf {C h} _ {t} \tag {4} +$$ + +where $\mathbf{A},\mathbf{B}$ , and $\mathbf{C}$ are learnable state-space matrices for state, input, and output, respectively. $\mathbf{x_t}$ is the input frame of the current time step. $\mathbf{h_t}$ is the internal hidden state. $\mathbf{y_t}$ is the output, i.e., perception token. + +Benefit from the strong temporal modeling capability of Selective SSM, our EPFE with only $56\mathrm{M}$ parameters can learn long-term event-level spatiotemporal features with constant cost. As Appendix Fig. 1 illustrates, perception tokens generated by EPFE can distinguish different events and noise frames, while maintaining a state of the event. Even after unrelated noise frames, the perception tokens are capable of refocusing on the event. + +# 2.4. Cognition Gate + +Motivation: Cognition Gate needs to determine whether to invoke LLM in real-time, based on the current perception token and user query. For example, during the soccer match, some user query may only care LLM response for scoring moments, while others may need detailed tactical analysis. + +However, we find it highly challenging. For real-time consideration, we first employ a simple Transformer block + +with cross-attention to associate the user query with the visual encoder's output for trigger decisions. However, it struggles to make accurate judgments at the right time step, as shown in Tab.5. We identify two key reasons for this: 1) The Cognition Gate only has access to information up to the current step, lacking a global temporal perspective for decision-making. 2) User queries may not directly correspond to specific video frames in a retrieval-like manner (e.g., guide me to repair a bicycle or cooking). + +Therefore, the Cognition Gate requires the world knowledge of LLMs. However, due to the size of LLMs, even processing a single perception token as input incurs a high computational cost. Notably, the Gate only needs to generate a response/silence token for a single iteration. + +To achieve it, We propose the Shallow Layer Transfer method, which initializes the Gate using the early layers of the LLM and then fine-tunes it in a supervised autoregressive manner with the StreamVD dataset. As shown in Sec. 4.2, we conducted extensive experiments to analyze its impact. The results indicate that the first few layers exhibit slight variations in behavior, allowing for scalability based on requirements. In this work, we use four layers. + +# 2.5. The Training Strategy of STREAMMIND + +Model Architecture As illustrated in Figure 3, the STREAMMIND video LLM integrates CLIP [46], EPFE, Cognition Gate, and the LLM. CLIP is the pre-trained spatial feature extractor. EPFE then takes the spatial features as input as well as its internal state to extract spatiotemporal features and generate a perception token. Cognition Gate then evaluates the current perception token and user query to determine whether to invoke LLM. To achieve this, it is needed to train EPFE, Cognition Gate, and also align the representation space of EPFE and LLM, as shown in Fig. 4. + +Algorithm 1: Streaming Video Dataset Construction +Require: Offline video dataset $C = \{c_i,t_i\}_{i = 1}^n (c_i$ caption at time $t_i)$ +Ensure: Dataset $D$ with captions meeting StreamVD requirements and Cognition Gate labels. +1: Step 1: Preprocessing +2: $D\gets []$ , cap $\leftarrow c_{1}$ ,time $\leftarrow t_{s}$ +3: for $i = 2$ to n do +4: if $c_{i}\neq$ cap then +5: Append (cap, time) to $D$ +6: cap $\leftarrow c_{i}$ ,time $\leftarrow t_{i}$ +7: end if +8: end for +9: Append (cap, time) to $D$ +10: Step 2: Silence-Response Labeling +11: for each frame $f$ between $d_{i}$ and $d_{i + 1}$ do +12: Label $f$ withif captioned, else +13: end for +14: return $D$ + +Data Preparation Streaming datasets are constructed by processing existing offline datasets. StreamingVD tasks require the model to response as early as possible when relevant events occur while avoid redundant responses. To achieve this, we eliminate repetitive captions from the offline datasets and transform them into streaming datasets using the following methods, detailed in Algorithm 1. + +- Preprocessing: Adjacent identical captions are merged, with the timestamp of the first occurrence being recorded as the annotation point for the caption. +- Silence-Response Labeling: To train the Cognition Gate, we insert $< / \text{silence}>$ tokens into video frames between two adjacent captions. These tokens act as indicators to guide the Cognition Gate in generating the $< / \text{silence}>$ . The initial frame corresponding to captioned events is labeled with the $< / \text{response}>$ token. + +Training Strategy We adopt a two-stage training strategy, as shown in Fig. 4. In contrast to previous typical two-stage training approaches [29, 40, 71], our method uses a dedicated stage for Cognition Gate. In the first stage, the LLM and EPFE are jointly trained using video frames and captions from the streaming dataset, ensuring spatiotemporal feature alignment with the LLM. The second stage trains only the Cognition Gate, which is initialized from the shallow layers of the LLM. This training also follows the LLM autoregressive manner to generate $< / \text{silence} >$ or $< / \text{response} >$ tokens, determining whether to invoke the LLM. Since the response label is much sparser than silence (e.g. 310:1), we introduce a balancing weight into + +the standard Cross Entropy loss during training to balance the labels. Refer to Sec. 3.1 and Appendix Sec. 1 for more training details. + +![](images/739ae2abf9e281cfbd7ecf9e6f650725ee50be11cd644d0abd9ed25a0efaee6a.jpg) +Figure 4. The two-stage training process of STREAMMIND. In the first stage, the EPFE spatiotemporal extractor and the LLM are jointly trained to align their representations. In the second stage, the Cognition module is trained to generate response/silence tokens. Both stages follow a supervised autoregressive training approach for language modeling. + +![](images/1e4c820793b4f84127f077e2a6c5b6bc3df4bdcb771bd1d8e826e59ec4f93c33.jpg) + +# 3. Experiment + +# 3.1. Experimental Settings + +Baseline and Datasets Since existing works focusing on streamingVD [8, 26, 60], except for VideoLLM-Online [8], do not provide open-source data or training methods, to demonstrate our outstanding performance, we re-implemented all previous works (VideoLLM-Online [8] and VideoLLM-MOD [60]) in the StreamingVD domain as baselines, using consistent settings across all experiments. + +For dataset selection, since the Ego4D dataset [21] is inherently collected in a streaming manner, containing various queries, real-time responses, and narrations, and has been used as an main evaluation dataset in works [8, 31, 60], we selected it as one of our evaluation datasets. Furthermore, to further demonstrate the effectiveness of our approach on long videos, we also conducted experiments on the SoccerNet-Caption dataset [20]. This dataset includes 471 complete soccer match videos with a total duration of 715 hours, where each video is approximately 45 minutes long and provides commentary at appropriate timestamps. + +Implementation Details: Model training was conducted on videos sampled at 2 fps, using 8 NVIDIA A100 GPUs. Each stage of the training process was conducted for one epoch. To optimize the learning rate, we applied the CosineAnnealingLR scheduler, with learning rates set to $2e - 5$ and $2e - 6$ respectively. Additionally, specific techniques were applied to the design of the Cognition Gate and methods to address the imbalance in the Silence-Response ratio, which are detailed in Section 4 and the supplementary materials. + +# 3.2. Evaluate Metrics + +To evaluate the effectiveness of STREAMMIND, we focus on assessing its Timing Alignment Capability and Lan + +Table 1. Online Experiments Results. + +
DataSetMethodTiming Alignment ↑Language Modeling Capability ↑Training Cost
TriggerAccTimValBLUE-1BLUE-4METEORROUGE-L
Ego4dVideoLLM-Online32.34%29.66%66.0135.2531.1263.0624h
VideoLLM-MoD32.36%29.65%65.3435.2130.6563.0214h
Ours43.34%39.73%67.1239.2631.6065.71(11+8)h
SoccerVideoLLM-Online31.25%28.34%75.3664.2350.9281.5712h
VideoLLM-MoD31.24%28.12%74.9664.1850.2481.597h
Ours52.18%47.36%82.7866.7051.4382.04(5+3)h
+ +Table 2. Online Experiments Results. LION-FS was released on arXiv just two days before our submission, but its code was not released. Therefore, we could only obtain the Ego4D results based on its paper. + +
DataSetMethodTimeDiff ↓Fluency ↑PPL ↓Correctiness ↑
Ego4dVideoLLM-Online2.0445.0%2.4348.1%
VideoLLM-MOD2.0445.2%2.4148.9%
LION-FS2.1546.1%2.0952.4%
Ours1.8960.2%2.0277.3%
SoccerVideoLLM-Online15.6246.35%1.7953.5%
VideoLLM-MOD15.6845.34%1.8053.3%
Ours14.0270.35%1.5989.2%
+ +Table 3. Results on COIN benchmarks(left to right): step recognition, task recognition, next forecasting, procedure forecasting, procedure forecasting with a goal. + +
MethodNot use HT 100MCOINBenchmark Top-1 Accuracy ↑
StepTaskNextProcProg.+
ClipBERT✓30.865.4---
TimeFormer×46.585.334.017.040.1
Paprika×51.085.843.2--
DistantSup×54.190.039.4--
VideoTF×56.591.042.440.246.4
ProcedureVRL×56.990.846.8--
VideoTaskGraph×57.290.540.2--
VideoLLM-online-7B-v1✓59.892.148.147.952.9
VideoLLM-online -8B-v1+✓63.192.749.149.854.1
VideoLLM-MOD✓63.492.849.749.853.3
Ours✓63.793.249.949.854.2
+ +Table 4. Results on Ego4DLTA benchmark, evaluated on public server. ED@Z=20 denotes editing distance for future 20 actions. + +
MethodNot use Ego VLPEnd-to -End?Ego4D LTA ED@Z=20
Verb ↓Noun ↓Action ↓
CLIP✓✓0.7390.7690.941
EgoT2✓✓0.7220.7640.935
I-CVAE✓✓0.7530.7490.931
HierVL✓✓0.7240.7350.928
VideoLLM×✓0.7210.7250.921
VideoLLM-online-7B-v1✓✓0.6970.6980.897
VideoLLM-online-8B-v1+✓✓0.6890.6710.884
VideoLLM-MOD✓✓0.6890.6760.884
Ours✓✓0.6890.6550.881
+ +guage Modeling Capability. Existing evaluation metrics include LM-PPL (Language Modeling Perplexity) and LM-Correctness, which evaluate language modeling accuracy at specific timestamps. Additionally, the Time Difference + +(TimeDiff) metric is used to measure the temporal alignment of responses in an streaming video LLM, while the Fluency metric aims to provide a holistic evaluation of both linguistic quality and temporal consistency. + +However, these metrics have several limitations: + +- LM-PPL and LM-Correctness focus solely on the exact alignment between model outputs and reference narrations, ignoring aspects such as synonym substitutions, sentence fluency, and the ability to capture critical information. +- TimeDiff, as a metric for timing alignment Capability, evaluates temporal alignment only within a single dialogue turn. It does not assess the Cognition Gate's overall decision-making across the entire video stream. + +To address these limitations, we introduce the following metrics: + +- Trigger Accuracy (TriggerAcc): To evaluate whether the streaming video LLM responds at the correct time steps during entire streaming video. +- Timing Validity (TimVal): To comprehensively evaluate whether the streaming video LLM consistently makes the correct decisions throughout the streaming video—speaking when necessary and remaining silent when appropriate. +- BLUE [43], METEOR [2], ROUGE-L [34]: Widely used evaluation metrics in the field of text generation [6, 9, 38] and image narration [7, 44, 48], which fully consider factors such as synonym matching, stemming, word order, and key semantic expressions. These metrics enable a comprehensive assessment of dialogues in the StreamingVD process. + +# 3.3. Comprehensive Evaluation on Streaming and Offline Video + +Online Experiments: We compared our model with all existing streaming dialogue models [8, 31, 60] in Tab 1 and Tab 2. Under comparable training costs, experiments on real-time Ego4D and SoccerNet demonstrate superior performance across all metrics. Additionally, STREAMMIND can process video streams at a $10 \times$ higher FPS. This improvement is attributed to the event-level modeling capability. + +ity of EPFE and the strong timing alignment capability of the cognition gate in the perception stage, which together provide the LLM in the cognition phase with a concise yet comprehensive representation. + +Offline Experiments: We also leveraged the COIN [53] and Ego4D LTA [24] benchmark to demonstrate STREAM-MIND effectiveness in traditional offline video scenarios, covering six common benchmarks: step recognition, step forecasting, task summarization, procedure forecasting, and procedure forecasting with a goal. As shown in Tab 3 and Tab 4, our approach achieves state-of-the-art performance across all tasks. + +# 3.4. Real-time Inference Efficiency + +Standard frame rates for films are typically 24 FPS, which is considered the minimum requirement for smooth motion perception. In contrast, games often operate at much higher frame rates, ranging from 30 FPS to over 100 FPS, to ensure a immersive experience. To validate the real-time performance of STREAMMIND, we conducted inference efficiency evaluations on NVIDIA A100 and H100 GPUs. + +Specifically, we compared VideoLLM-Online, VideoLLM-MOD, and STREAMMIND by measuring their actual processing time for 1-second video across input frame rates from 5 to 100 FPS. As shown in Figure 5, a processing time under 1 second indicates real-time performance without disrupting playback. + +The results reveal that VideoLLM-Online and VideoLLM-MOD struggle to process video streams beyond 10 FPS, leading to potential latency issues in real-world applications. In contrast, our method achieves true real-time performance, efficiently handling both standard film/TV frame rates and highRefresh gaming scenarios up to 100 FPS, ensuring real-time responses. + +# 4. Ablation + +# 4.1. Impact of Silence-Response Sample Imbalance + +In this section, we explored the imbalance in the Silence-Response ratio within the dataset and derived an empirical balancing formula through experiments. For detailed experimental results, please refer to the supplementary material. + +# 4.2. The Ablation of Cognition Gate + +The Cognition Gate plays a crucial role throughout the perception-cognition interleaving process. We conducted detailed ablation experiments to analyze its impact. + +The Architecture of Cognition Gate The structure of the gate directly determines its capabilities. To this end, we constructed multiple classification networks using MLP projectors, Cross-Attention, and Transformers. [54] Additionally, we built a set of autoregressive networks using an + +![](images/177b0ddf4d7b4d1cbcdbb3dd92ae00c24c8c053d74c13cd4ee0f00aa452170ce.jpg) + +![](images/1d9a738ef34a21e23a4bf4748ba50df552804677ae8aa33db725f0e165ea4397.jpg) +Figure 5. Comparison of running time for processing 1 second of streaming video frames using different methods on A100 (Top) and H100 (Bottom). + +LLM block. The results are shown in Tab.5. It is evident that the LLM-like structure achieved better performance. For a detailed analysis, please refer to Section 2.4. + +Table 5. Ablation Study on the architecture of cognition gate + +
ArchitectureTimeDiff ↓TriggerAcc ↑TimVal ↑
Linear Layer5.0620.13%17.65%
MLP projector + Linear4.3421.75%18.33%
Transformers + Linear3.7821.58%18.97%
Cross-Attention + Linear3.6424.34%20.36%
Single LLM block2.2235.35%32.11%
+ +Initialization Strategies We investigate three initialization strategies for the Cognition Gate to enhance its effectiveness: (i) Random Initialization: All layers are initialized from a standard random distribution without utilizing prior knowledge from the LLM. (ii) SkipBlock Initialization: The Cognition Gate is initialized by uniformly selecting layers from the LLM, with the number of skipped layers determined by the ratio of the Cognition Gate's depth to the total number of LLM layers. (iii) EarlyBlock Initialization: The first few layers of the LLM are directly used to initialize the Cognition Gate, leveraging the LLM's early-stage processing capabilities. + +Number of Blocks The Cognition Gate consists of multiple layers, and its depth plays a crucial role in balancing computational efficiency and perceptual accuracy. + +Table 6. Ablation Study on the Strategy of Initialization + +
Strategy of InitializationTimeDiff ↓TriggerAcc ↑TimVal ↑
Random2.1039.65%37.35%
SkipBlock1.8940.23%37.67%
EarlyBlock1.8943.34%39.73%
+ +Table 7. Ablation Study on the Strategy of Initialization + +
#Num of BlockTimeDiff ↓TriggerAcc ↑TimVal ↑
21.9241.53%37.65%
31.9042.35%38.34%
41.8943.34%39.73%
51.8842.67%38.56%
61.8942.56%38.34%
+ +The results in Tab.6 and Tab.7 show that fine-tuning the shallow layers of the LLM yields the best performance. We attribute this to the fact that gate operates within a limited decision space (binary: Yes/No), making deeper layers—primarily responsible for nuanced language generation—unnecessary. Thus, we propose the Shallow Layer Transfer method to construct the Cognition Gate, as detailed in Section 2.4. + +# 4.3. The Performance of Event-preserving Feature extractor + +To evaluate the role of EPFE in the perception phase, we compare it with commonly used extractors in VideoLLMs, including: 1) The Q-Former [28]. 2) The Spatial-Temporal Convolution Connector (STC Connector) [12], which has demonstrated outstanding performance in videolm and has been widely adopted [29, 59, 71]. Notably, STC employs 3D convolution for spatial-temporal aggregation. To preserve local visual patterns during spatial compression, it follows work [5] by incorporating a RegStage block [47] before and after the 3D convolution, which has been shown to enhance spatial understanding. Experimental results indicate that EPFE achieves superior performance. + +Table 8. Perceptive Judgement Capability + +
MethodTimeDiff ↓TriggerAcc ↑TimVal ↑
Q-former3.7826.65 %25.31%
STC3.5627,54%26.87%
EPFE1.8943.34%39.73%
+ +# 5. Related Work + +# 5.1. Offline VideoLLMs. + +In recent years, there has been a surge of interest in leveraging Large Language Models (LLMs) for video understanding, leading to the development of offline videoLLMs. The biggest challenge for it is managing memory requirements [4, 11, 15, 23, 50, 51] and compressing redundant frame features [12, 25, 27, 52, 76]. LaViLa [37] addresses + +this challenge by transforming LLMs into narrators that generate detailed descriptions of long videos from visual inputs. Other models, such as ChatVideo [56] and MMVID [36], convert videos into text to improve comprehension. MovieChat [50] combines all frame features using an averaging strategy. However, in the context of online streaming, there has been limited exploration into how VideoLLMs can meet the demands of temporal alignment and real-time processing for streaming video inputs. + +# 5.2. Online VideoLLMs. + +Recent studies have introduced VideoLLMs specifically designed for online stream understanding, while most existing works [14, 42, 45, 62, 72, 77] focus on improving offline video efficiency to meet real-time requirements. However, they fail to proactively decide when to respond based on user requirements. For example, ReKV [14] implemented a sliding-window attention mechanism and a KV-cache system to reduce computational overhead. To address this, VideoLLM-Online [8] introduces per-step LLM invocation for online video dialogue, as shown in Fig.1. However, its performance is constrained by limited per-frame processing capability due to the absence of effective frame compression strategies. VideoLLM-MOD [60] mitigates this issue by incorporating a mixture-of-depth approach for efficient visual token computation, enabling higher visual input resolutions. Building on MOD, LION-FS [31] proposes Token Aggregation Router and Token DroppingRouter as the Fast Path, adaptively aggregating distinct features while discarding redundant ones. However, they remain limited to processing videos at only 10 FPS. + +Overall, prior work remains restricted by the contradiction between sequential full-frame processing and the extremely high efficiency requirements of streaming video, lacking breakthrough paradigm designs explicitly suited for streaming frame inputs. To address these challenges, our approach introduces a novel perception-cognition interleaving paradigm and proposes the STREAMMIND framework, enabling VideoLLM to achieve proactive responses and real-time streaming video understanding, with a maximum throughput of up to 100 FPS. + +# 6. Conclusion + +In this paper, we introduce a novel perception-cognition interleaving paradigm called "event-gated LLM invocation" and propose the STREAMMIND framework, addressing the contradiction between sequential full-frame processing and the extremely high efficiency requirements of streaming video. As a result, STREAMMIND can process video at up to 100 FPS, achieving the SOTA performance across all existing metrics and benchmarks. + +# 7. Acknowledgement + +We gratefully acknowledge the support from the following foundations and institutions. Ting Cao was supported in part by the Tsinghua University (AIR)-AsiaInfo Technologies (China), Inc. Joint Research Center for 6G Network and Intelligent Computing. Hao Wu was supported in part by the National Natural Science Foundation of China (NSFC) under Grants 62432004 and 62302207, and by the Science Foundation for Youths of Jiangsu Province under Grant BK20220772. Xin Ding and Zhibo Chen were supported in part by the NSFC under Grants 62371434 and 62021001. + +# References + +[1] Roberto Amoroso, Gengyuan Zhang, Rajat Koner, Lorenzo Baraldi, Rita Cucchiara, and Volker Tresp. Perceive, query & reason: Enhancing video qa with question-guided temporal queries. arXiv preprint arXiv:2412.19304, 2024. 2 +[2] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 6 +[3] Pietro Battistoni, Andrea Antonio Cantone, Mariarosaria Esposito, Rita Francese, Francesca Pia Perillo, Marco Romano, Monica Sebillo, and Giuliana Vitiello. Using artificial intelligence and companion robots to improve home healthcare for the elderly. In International Conference on Human-Computer Interaction, pages 3–17. Springer, 2023. 1 +[4] Jiarui Cai, Mingze Xu, Wei Li, Yuanjun Xiong, Wei Xia, Zhuowen Tu, and Stefano Soatto. Memot: Multi-object tracking with memory. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8090-8100, 2022. 8 +[5] Junbum Cha, Wooyoung Kang, Jonghwan Mun, and Byungseok Roh. Honeybee: Locality-enhanced projector for multimodal llm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13817-13827, 2024. 8 +[6] Yupeng Chang, Xu Wang, Jindong Wang, Yuan Wu, Linyi Yang, Kaijie Zhu, Hao Chen, Xiaoyuan Yi, Cunxiang Wang, Yidong Wang, et al. A survey on evaluation of large language models. ACM transactions on intelligent systems and technology, 15(3):1-45, 2024. 6 +[7] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3558-3568, 2021. 6 +[8] Joya Chen, Zhaoyang Lv, Shiwei Wu, Kevin Qinghong Lin, Chenan Song, Difei Gao, Jia-Wei Liu, Ziteng Gao, Dongxing Mao, and Mike Zheng Shou. Videollm-online: Online video large language model for streaming video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18407-18418, 2024. 1, 2, 3, 5, 6, 8 + +[9] Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji, and Quanquan Gu. Self-play fine-tuning converts weak language models to strong language models. arXiv preprint arXiv:2401.01335, 2024. 6 +[10] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24185–24198, 2024. 3 +[11] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In European Conference on Computer Vision, pages 640-658. Springer, 2022. 8 +[12] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 8 +[13] Shangzhe Di and Weidi Xie. Grounded question-answering in long egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12934–12943, 2024. 3 +[14] Shangzhe Di, Zhelun Yu, Guanghao Zhang, Haoyuan Li, Hao Cheng, Bolin Li, Wanggui He, Fangxun Shu, Hao Jiang, et al. Streaming video question-answering with in-context video kv-cache retrieval. In The Thirteenth International Conference on Learning Representations. 1, 2, 3, 8 +[15] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. In European Conference on Computer Vision, pages 75-92. Springer, 2024. 8 +[16] Feiteng Fang, Yuelin Bai, Shiwen Ni, Min Yang, Xiaojun Chen, and Ruifeng Xu. Enhancing noise robustness of retrieval-augmented language models with adaptive adversarial training. arXiv preprint arXiv:2405.20978, 2024. 2 +[17] Xinyu Fang, Kangrui Mao, Haodong Duan, Xiangyu Zhao, Yining Li, Dahua Lin, and Kai Chen. Mmbench-video: A long-form multi-shot benchmark for holistic video understanding. Advances in Neural Information Processing Systems, 37:89098-89124, 2025. 2 +[18] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19358-19369, 2023. 3 +[19] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 2 +[20] Silvio Giancola, Mohieddine Amine, Tarek Dghaily, and Bernard Ghanem. Soccernet: A scalable dataset for action spotting in soccer videos. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 1711-1721, 2018. 2, 5 + +[21] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18995-19012, 2022. 2, 5 +[22] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. CoRR, abs/2312.00752, 2023. 2, 4 +[23] Bo He, Hengduo Li, Young Kyun Jang, Menglin Jia, Xuefei Cao, Ashish Shah, Abhinav Shrivastava, and Ser-Nam Lim. Ma-lmm: Memory-augmented large multimodal model for long-term video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13504-13514, 2024. 8 +[24] Daoji Huang, Otmar Hilliges, Luc Van Gool, and Xi Wang. Palm: Predicting actions through language models@ ego4d long-term action anticipation challenge 2023. arXiv preprint arXiv:2306.16545, 2023. 2, 7 +[25] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In European Conference on Computer Vision, pages 202-218. Springer, 2024. 8 +[26] Zhenpeng Huang, Xinhao Li, Jiaqi Li, Jing Wang, Xi-angyu Zeng, Cheng Liang, Tao Wu, Xi Chen, Liang Li, and Limin Wang. Online video understanding: A comprehensive benchmark and memory-augmented method. arXiv preprint arXiv:2501.00584, 2024. 1, 2, 5 +[27] Yang Jin, Zhicheng Sun, Kun Xu, Liwei Chen, Hao Jiang, Quzhe Huang, Chengru Song, Yuliang Liu, Di Zhang, Yang Song, et al. Video-lavit: Unified video-language pre-training with decoupled visual-motional tokenization. arXiv preprint arXiv:2402.03161, 2024. 8 +[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2, 3, 8 +[29] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 5, 8 +[30] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 1, 2 +[31] Wei Li, Bing Hu, Rui Shao, Leyang Shen, and Liqiang Nie. Lion-fs: Fast & slow video-language thinker as online video assistant. arXiv preprint arXiv:2503.03663, 2025. 5, 6, 8 +[32] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2024. 1, 3 + +[33] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 3 +[34] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 6 +[35] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 26689-26699, 2024. 3 +[36] Kevin Lin, Faisal Ahmed, Linjie Li, Chung-Ching Lin, Ehsan Azarnasab, Zhengyuan Yang, Jianfeng Wang, Lin Liang, Zicheng Liu, Yumao Lu, et al. Mm-vid: Advancing video understanding with gpt-4v (isdiction). arXiv preprint arXiv:2310.19773, 2023. 8 +[37] Xingyu Liu, Joon-Young Lee, and Hailin Jin. Learning video representations from correspondence proposals. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 4273–4281, 2019. 8 +[38] Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. G-eval: Nlg evaluation using gpt-4 with better human alignment. arXiv preprint arXiv:2303.16634, 2023. 6 +[39] Ruipu Luo, Ziwang Zhao, Min Yang, Junwei Dong, Da Li, Pengcheng Lu, Tao Wang, Linmei Hu, Minghui Qiu, and Zhongyu Wei. Valley: Video assistant with large language model enhanced ability. arXiv preprint arXiv:2306.07207, 2023. 1, 3 +[40] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 1, 5 +[41] Kelly Merrill Jr, Jihyun Kim, and Chad Collins. Ai companions for lonely individuals and the role of social presence. Communication Research Reports, 39(2):93-103, 2022. 1 +[42] Sunny Panchal, Apratim Bhattacharyya, Guillaume Berger, Antoine Mercier, Cornelius Böhm, Florian Dietrichkeit, Reza Pourreza, Xuanlin Li, Pulkit Madan, Mingu Lee, et al. What to say and when to say it: Live fitness coaching as a testbed for situated interaction. Advances in Neural Information Processing Systems, 37:75853-75882, 2024. 8 +[43] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 6 +[44] Jordi Pont-Tuset, Jasper Uijlings, Soravit Changpinyo, Radu Soricut, and Vittorio Ferrari. Connecting vision and language with localized narratives. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 647-664. Springer, 2020. 6 +[45] Rui Qian, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Shuanggrui Ding, Dahua Lin, and Jiaqi Wang. Streaming long video understanding with large language models. Advances in Neu + +ral Information Processing Systems, 37:119336-119360, 2025. 1, 2, 3, 8 +[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 3, 4 +[47] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólár. Designing network design spaces. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10428-10436, 2020. 8 +[48] Jiayuan Rao, Haoning Wu, Chang Liu, Yanfeng Wang, and Weidi Xie. Matchtime: Towards automatic soccer game commentary generation. arXiv preprint arXiv:2406.18530, 2024. 6 +[49] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 1 +[50] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 8 +[51] Larry R Squire, Lisa Genzel, John T Wixted, and Richard G Morris. Memory consolidation. Cold Spring Harbor perspectives in biology, 7(8):a021766, 2015. 8 +[52] Guangzhi Sun, Wenyi Yu, Changli Tang, Xianzhao Chen, Tian Tan, Wei Li, Lu Lu, Zejun Ma, and Chao Zhang. Fine-grained audio-visual joint representations for multimodal large language models. arXiv preprint arXiv:2310.05863, 2023. 8 +[53] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1207-1216, 2019. 2, 7 +[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 7 +[55] Ales Vysocky and Petr Novak. Human-robot collaboration in industry. MM Science Journal, 9(2):903–906, 2016. 1 +[56] Junke Wang, Dongdong Chen, Chong Luo, Xiyang Dai, Lu Yuan, Zuxuan Wu, and Yu-Gang Jiang. Chatvideo: A tracklet-centric multimodal and versatile video understanding system. arXiv preprint arXiv:2304.14407, 2023. 8 +[57] Qing Wang, Jiaming Zhang, Kailun Yang, Kunyu Peng, and Rainer Stiefelhagen. Matchformer: Interleaving attention in transformers for feature matching. In Proceedings of the Asian Conference on Computer Vision, pages 2746-2762, 2022. 2 + +[58] Xunguang Wang, Zhenlan Ji, Pingchuan Ma, Zongjie Li, and Shuai Wang. Instructta: Instruction-tuned targeted attack for large vision-language models. arXiv preprint arXiv:2312.01886, 2023. 2 +[59] Yuxuan Wang, Yueqian Wang, Pengfei Wu, Jianxin Liang, Dongyan Zhao, and Zilong Zheng. Lstp: Language-guided spatial-temporal prompt learning for long-form video-text understanding. arXiv e-prints, pages arXiv-2402, 2024. 8 +[60] Shiwei Wu, Joya Chen, Kevin Qinghong Lin, Qimeng Wang, Yan Gao, Qianli Xu, Tong Xu, Yao Hu, Enhong Chen, and Mike Zheng Shou. Videollm-mod: Efficient video-language streaming with mixture-of-depths vision computation. Advances in Neural Information Processing Systems, 37:109922-109947, 2025. 1, 3, 5, 6, 8 +[61] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9777–9786, 2021. 3 +[62] Haomiao Xiong, Zongxin Yang, Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Jiawen Zhu, and Huchuan Lu. Streaming video understanding and multi-round interaction with memory-enhanced knowledge. arXiv preprint arXiv:2501.13468, 2025. 1, 2, 3, 8 +[63] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the 25th ACM international conference on Multimedia, pages 1645-1653, 2017. 2 +[64] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 1 +[65] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1686-1697, 2021. 3 +[66] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. Advances in Neural Information Processing Systems, 35:124-141, 2022. 1 +[67] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10714-10726, 2023. 1 +[68] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9127–9134, 2019. 2 +[69] Jeffrey M Zacks, Nicole K Speer, Khena M Swallow, Todd S Braver, and Jeremy R Reynolds. Event perception: a mind-brain perspective. Psychological bulletin, 133(2):273, 2007. 2 + +[70] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975-11986, 2023. 3 +[71] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 5, 8 +[72] Haoji Zhang, Yiqin Wang, Yansong Tang, Yong Liu, Jiashi Feng, Jifeng Dai, and Xiaojie Jin. Flash-vstream: Memory-based real-time understanding for long video streams. arXiv preprint arXiv:2406.08085, 2024. 1, 2, 8 +[73] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 3 +[74] Ruohong Zhang, Liangke Gui, Zhiqing Sun, Yihao Feng, Keyang Xu, Yuanhan Zhang, Di Fu, Chunyuan Li, Alexander Hauptmann, Yonatan Bisk, et al. Direct preference optimization of video large multimodal models from language model reward. arXiv preprint arXiv:2404.01258, 2024. 3 +[75] Ruoyu Zhang, Lulu Wang, Yi He, Tongling Pan, Zhengtao Yu, and Yingna Li. Tpcap: Unlocking zero-shot image captioning with trigger-augmented and multi-modal purification modules. arXiv preprint arXiv:2502.11024, 2025. 2 +[76] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6586-6597, 2023. 8 +[77] Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. Streaming dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18243-18252, 2024. 8 \ No newline at end of file diff --git a/2025/StreamMind_ Unlocking Full Frame Rate Streaming Video Dialogue through Event-Gated Cognition/images.zip b/2025/StreamMind_ Unlocking Full Frame Rate Streaming Video Dialogue through Event-Gated Cognition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b167ab3e6eb096c845fdb9ca873e219378f05a36 --- /dev/null +++ b/2025/StreamMind_ Unlocking Full Frame Rate Streaming Video Dialogue through Event-Gated Cognition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8f41704498fa8765e5d4b6ae521bdaaa3763f58327a3f6c15ac028c5665b117d +size 532228 diff --git a/2025/StreamMind_ Unlocking Full Frame Rate Streaming Video Dialogue through Event-Gated Cognition/layout.json b/2025/StreamMind_ Unlocking Full Frame Rate Streaming Video Dialogue through Event-Gated Cognition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5740528c75648a5e6a49df40ead1599521f4105d --- /dev/null +++ b/2025/StreamMind_ Unlocking Full Frame Rate Streaming Video Dialogue through Event-Gated Cognition/layout.json @@ -0,0 +1,9621 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 60, + 102, + 550, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 102, + 550, + 140 + ], + "spans": [ + { + "bbox": [ + 60, + 102, + 550, + 140 + ], + "type": "text", + "content": "STREAMMIND: Unlocking Full Frame Rate Streaming Video Dialogue through Event-Gated Cognition" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 159, + 498, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 159, + 498, + 190 + ], + "spans": [ + { + "bbox": [ + 113, + 159, + 498, + 190 + ], + "type": "text", + "content": "Xin Ding" + }, + { + "bbox": [ + 113, + 159, + 498, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 159, + 498, + 190 + ], + "type": "text", + "content": " * Hao Wu" + }, + { + "bbox": [ + 113, + 159, + 498, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 113, + 159, + 498, + 190 + ], + "type": "text", + "content": " † Yifan Yang" + }, + { + "bbox": [ + 113, + 159, + 498, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 159, + 498, + 190 + ], + "type": "text", + "content": " Shiqi Jiang" + }, + { + "bbox": [ + 113, + 159, + 498, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 159, + 498, + 190 + ], + "type": "text", + "content": " Qianxi Zhang" + }, + { + "bbox": [ + 113, + 159, + 498, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 159, + 498, + 190 + ], + "type": "text", + "content": " Donglin Bai" + }, + { + "bbox": [ + 113, + 159, + 498, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 159, + 498, + 190 + ], + "type": "text", + "content": " Zhibo Chen" + }, + { + "bbox": [ + 113, + 159, + 498, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 159, + 498, + 190 + ], + "type": "text", + "content": " Ting Cao" + }, + { + "bbox": [ + 113, + 159, + 498, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 113, + 159, + 498, + 190 + ], + "type": "text", + "content": " †" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 74, + 190, + 535, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 190, + 535, + 218 + ], + "spans": [ + { + "bbox": [ + 74, + 190, + 535, + 218 + ], + "type": "text", + "content": "1University of Science and Technology of China 2Microsoft Research 3Nanjing University Institute for AI Industry Research (AIR), Tsinghua University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 270, + 296, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 354 + ], + "type": "text", + "content": "With the rise of real-world human-AI interaction applications, such as AI assistants, the need for Streaming Video Dialogue is critical. To address this need, we introduce STREAMMIND, a video LLM framework that achieves ultra-FPS streaming video processing (100fps on a single A100) and enables proactive, always-on responses in real time, without explicit user intervention." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 356, + 296, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 356, + 296, + 511 + ], + "spans": [ + { + "bbox": [ + 54, + 356, + 296, + 511 + ], + "type": "text", + "content": "To solve the key challenge of the contradiction between linear video streaming speed and quadratic transformer computation cost, we propose a novel perception-cognition interleaving paradigm named \"event-gated LLM invocation\", in contrast to the existing per-time-step LLM invocation. By introducing a Cognition Gate network between the video encoder and the LLM, LLM is only invoked when relevant events occur. To realize the event feature extraction with constant cost, we propose Event-Preserving Feature Extractor (EPFE) based on state-space method, generating a single perception token for spatiotemporal features. These techniques enable the video LLM with full-FPS perception and real-time cognition response." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 512, + 296, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 512, + 296, + 584 + ], + "spans": [ + { + "bbox": [ + 55, + 512, + 296, + 584 + ], + "type": "text", + "content": "Experiments on Ego4D and SoccerNet streaming tasks, as well as standard offline benchmarks, demonstrate state-of-the-art performance in both model capability and real-time efficiency, paving the way for ultra-high-FPS applications, such as Game AI Copilot and interactive media. The code and data is available at https://aka.ms/StreamMind." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 604, + 135, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 135, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 135, + 616 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 624, + 295, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 295, + 673 + ], + "type": "text", + "content": "The advancement of large foundation models is driving an increasing number of real-world human-AI interaction applications, such as AI home companions [3, 41] and human-robot collaboration [55, 64]. These applications" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 243, + 439, + 323 + ], + "blocks": [ + { + "bbox": [ + 316, + 243, + 439, + 323 + ], + "lines": [ + { + "bbox": [ + 316, + 243, + 439, + 323 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 439, + 323 + ], + "type": "image", + "image_path": "3adaf8afee610aaa13e2cd3ca26d5864a019fdf770a722f6606ba269a692111a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 323, + 441, + 332 + ], + "lines": [ + { + "bbox": [ + 315, + 323, + 441, + 332 + ], + "spans": [ + { + "bbox": [ + 315, + 323, + 441, + 332 + ], + "type": "text", + "content": "Existing Works 10~15 FPS on A100 GPU" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 336, + 555, + 381 + ], + "lines": [ + { + "bbox": [ + 313, + 336, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 555, + 381 + ], + "type": "text", + "content": "Figure 1. The paradigm of existing per-step LLM invocation (left) versus our event-gated LLM invocation (right) for streaming video dialogue task. Our paradigm boosts frame rate up-to 100 fps on A100. " + }, + { + "bbox": [ + 313, + 336, + 555, + 381 + ], + "type": "inline_equation", + "content": "(R_{i}" + }, + { + "bbox": [ + 313, + 336, + 555, + 381 + ], + "type": "text", + "content": ": response, " + }, + { + "bbox": [ + 313, + 336, + 555, + 381 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 313, + 336, + 555, + 381 + ], + "type": "text", + "content": ": frame, t: time)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 441, + 244, + 564, + 323 + ], + "blocks": [ + { + "bbox": [ + 441, + 244, + 564, + 323 + ], + "lines": [ + { + "bbox": [ + 441, + 244, + 564, + 323 + ], + "spans": [ + { + "bbox": [ + 441, + 244, + 564, + 323 + ], + "type": "image", + "image_path": "35b51e882842b88a1cdd620d9ce8e208cbaae50c948226330f8044f9927ddb7d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 323, + 555, + 332 + ], + "lines": [ + { + "bbox": [ + 449, + 323, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 449, + 323, + 555, + 332 + ], + "type": "text", + "content": "Ours Up to 100 FPS on A100 GPU" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 393, + 555, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 555, + 502 + ], + "type": "text", + "content": "heavily rely on streaming video understanding capabilities [14, 26, 45, 62, 72], specifically streaming video dialogue [8, 60]. Streaming video dialogue (StreamingVD) aims to continuously perceive incoming video streams and, based on user queries, provide proactive, real-time, and always-on responses without human intervention. For example, an AI companion could respond to user queries like \"Guide me to fix this bicycle\" or \"Let's watch and discuss the soccer game together\"." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 506, + 556, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 556, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 556, + 650 + ], + "type": "text", + "content": "Streaming video dialogue can be viewed as a generalization of standard offline video dialogue tasks [30, 32, 39, 40, 49, 66, 67], extending video LLMs to broader real-world scenarios. The fundamental challenge of StreamingVD compared to its offline counterpart lies in the requirement for timing alignment between events and responses. That is, as the video progresses, the model is required to (1) proactively decide when to respond and (2) respond in real time before the next event occurs according to user queries, as illustrated in Fig. 2. However, with current techniques, these two requirements are contradictory and cannot be achieved simultaneously." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "The seminal VideoLLM-Online [8] and VideoLLM-MoD [60] introduce a per-step LLM invocation paradigm for StreamingVD, shown in Fig. 1. At each time step, it inputs all past frames and the user query to the LLM, to decide when to respond or remain silent. However," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 683, + 295, + 702 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 295, + 702 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 295, + 702 + ], + "type": "text", + "content": "*This work was done during Xin Ding's internship at Microsoft Research." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 71, + 703, + 150, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 703, + 150, + 712 + ], + "spans": [ + { + "bbox": [ + 71, + 703, + 150, + 712 + ], + "type": "text", + "content": "†Corresponding Author." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13448" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "text", + "content": "the " + }, + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "O(n^3)^1" + }, + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "text", + "content": " computational complexity and limited context window are mismatched with the " + }, + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "O(n)" + }, + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "text", + "content": " frame streaming, making real-time response difficult to achieve. More works [14, 45, 62, 72] focus on enhancing the efficiency of offline video dialogue aiming to meet real-time requirements. However, similar to offline approaches, they depend on user queries to trigger responses rather than the model to do so proactively." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 170, + 297, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 170, + 297, + 360 + ], + "spans": [ + { + "bbox": [ + 57, + 170, + 297, + 360 + ], + "type": "text", + "content": "To empower StreamingVD, this paper proposes a new perception-cognition interleaving paradigm, named event-gated LLM invocation, shown in Fig. 1. It is inspired by the event perception mechanism of human brain [69], which states that events are key components of human's perception, attention, and memory" + }, + { + "bbox": [ + 57, + 170, + 297, + 360 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 57, + 170, + 297, + 360 + ], + "type": "text", + "content": ". We therefore propose an event-based selective cognition process. Rather than invoking the LLM at every time step, we introduce a Cognition Gate between the video encoder and LLM. The video encoder continuously perceives visual signals of each frame, and only when query-related events occur, the gate will open to invoke the LLM. Analogous to how a text prompt triggers an LLM, streaming visual signals lack a clear stopping point, unlike a finite text sequence. The cognition gate bridges this gap by acting as a mechanism to identify \"physical world prompts\" for the LLM." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 361, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 529 + ], + "type": "text", + "content": "This new paradigm raises two major technical concerns, to achieve both high capability and efficiency. Firstly, existing video encoders in VideoLLMs (e.g., STC in VideoLaMA) only capture local spatio-temporal features. This results in a linear increase in the number of feature tokens for streaming frames and subsequently a quadratic computation for the processing, failing to meet real-time requirements. Secondly, the Cognition Gate must decide to invoke the LLM or not, based on visual features and the user query. However, only considering efficiency and simply attending visual and query tokens (e.g., Cross Attention in Qformer [28]) limits the gate capability to tasks like feature-matching [57, 58] and retrieval [1, 16, 75], struggle to make decisions with deeper semantic understanding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 530, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 530, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 295, + 662 + ], + "type": "text", + "content": "By addressing the challenges, we realize STREAM-MIND, a novel framework for streamingVD, with the processing speed matching frame rate for the first time. It integrates two techniques. For the video encoder, inspired by the exceptional ability of the state-assisted module [22] in modeling continuous physical signals with constant cost, we propose the Event-Preserving Feature Extractor (EPFE) based on the state-assisted module. It generates a single perception token of event features to the gate, resulting in a constant perception cost. For the Cognition Gate, to leverage the world knowledge of LLM while meeting the real-" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "time requirement, we propose a Shallow Layer Transfer method. The Gate is designed to reuse the shallow layers of the LLM and its autoregressive training to maximize the probabilities of “< response/silence >” tokens during training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 133, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 312 + ], + "type": "text", + "content": "To evaluate the performance of the STREAMMIND framework, we adopt evaluation metrics from VideoLLM-Online and a wide range of text generation evaluation metrics. Additionally, we design two novel metrics to comprehensively assess the model's temporal responsiveness. Experiments on real-time Ego4D [21] and SoccerNet [20] demonstrate superior performance across all metrics. Furthermore, our model achieves state-of-the-art results on multiple offline benchmarks, including short- and long-term activity recognition and forecasting on the COIN [53] and Ego4D LTA [24] datasets. Notably, our framework breaks the ultra-FPS streaming video processing bottleneck, enabling streamingVD at up to 100 fps on a single A100 GPU, laying a robust foundation for real-world human-AI interaction applications." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 317, + 370, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 317, + 370, + 330 + ], + "spans": [ + { + "bbox": [ + 314, + 317, + 370, + 330 + ], + "type": "text", + "content": "2. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 337, + 481, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 481, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 481, + 348 + ], + "type": "text", + "content": "2.1. Task Discussion and Definition" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 354, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 510 + ], + "type": "text", + "content": "Problem formulation of StreamingVD Given a video stream " + }, + { + "bbox": [ + 313, + 354, + 555, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^T \\coloneqq [v_1, v_2, \\dots, v_T]" + }, + { + "bbox": [ + 313, + 354, + 555, + 510 + ], + "type": "text", + "content": " consisting of " + }, + { + "bbox": [ + 313, + 354, + 555, + 510 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 354, + 555, + 510 + ], + "type": "text", + "content": " frames and a set of queries " + }, + { + "bbox": [ + 313, + 354, + 555, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{Q} \\coloneqq \\{q_1, q_2, \\dots, q_N\\}" + }, + { + "bbox": [ + 313, + 354, + 555, + 510 + ], + "type": "text", + "content": ". Traditional video dialogue, shown in Fig. 2, provides a response triggered by a user query. For a query " + }, + { + "bbox": [ + 313, + 354, + 555, + 510 + ], + "type": "inline_equation", + "content": "q_i" + }, + { + "bbox": [ + 313, + 354, + 555, + 510 + ], + "type": "text", + "content": " given at time step " + }, + { + "bbox": [ + 313, + 354, + 555, + 510 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 313, + 354, + 555, + 510 + ], + "type": "text", + "content": ", it responds based on past frames up to time " + }, + { + "bbox": [ + 313, + 354, + 555, + 510 + ], + "type": "inline_equation", + "content": "t_s (1 < t_s \\leq T)" + }, + { + "bbox": [ + 313, + 354, + 555, + 510 + ], + "type": "text", + "content": ", regardless of an online or offline scenario. On the other hand, StreamingVD proactively generates responses to a query given at " + }, + { + "bbox": [ + 313, + 354, + 555, + 510 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 313, + 354, + 555, + 510 + ], + "type": "text", + "content": " at any appropriate time steps until the query termination at " + }, + { + "bbox": [ + 313, + 354, + 555, + 510 + ], + "type": "inline_equation", + "content": "t_e (t_s < t_e \\leq T)" + }, + { + "bbox": [ + 313, + 354, + 555, + 510 + ], + "type": "text", + "content": ". At each time step " + }, + { + "bbox": [ + 313, + 354, + 555, + 510 + ], + "type": "inline_equation", + "content": "t_i (t_s \\leq t_i \\leq t_e)" + }, + { + "bbox": [ + 313, + 354, + 555, + 510 + ], + "type": "text", + "content": ", StreamingVD needs to decide whether to generate a response according to the past frames and query, and then generate the response, formulated as follows." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 356, + 514, + 553, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 514, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 356, + 514, + 553, + 529 + ], + "type": "interline_equation", + "content": "\\max P \\left(\\left[ \\operatorname {R e s} ^ {t _ {i}} \\right] \\mid \\left[ \\operatorname {C t x} ^ {< t _ {i}} \\right], \\left[ F ^ {t _ {i}} \\right]\\right) \\tag {1}", + "image_path": "6bf8cdcb11d84067de07c199ad119075c82bc473db3ba21c1fc8196f1da7b05e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 532, + 555, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 555, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 555, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 532, + 555, + 593 + ], + "type": "inline_equation", + "content": "[\\mathrm{Res}^{t_i}] \\in \\{[\\mathrm{Txt}^{t_i}], [\\mathrm{EOS}^{t_i}]\\}" + }, + { + "bbox": [ + 313, + 532, + 555, + 593 + ], + "type": "text", + "content": " means generate a response at a time step " + }, + { + "bbox": [ + 313, + 532, + 555, + 593 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 313, + 532, + 555, + 593 + ], + "type": "text", + "content": ", or remain silence, i.e., [EOS]. " + }, + { + "bbox": [ + 313, + 532, + 555, + 593 + ], + "type": "inline_equation", + "content": "[\\mathrm{Ctx}^{< t_i}]" + }, + { + "bbox": [ + 313, + 532, + 555, + 593 + ], + "type": "text", + "content": " denotes the contextual tokens accumulated from past frames and queries before " + }, + { + "bbox": [ + 313, + 532, + 555, + 593 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 313, + 532, + 555, + 593 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 532, + 555, + 593 + ], + "type": "inline_equation", + "content": "[\\mathrm{F}^{t_i}]" + }, + { + "bbox": [ + 313, + 532, + 555, + 593 + ], + "type": "text", + "content": " represents the visual features extracted from the video frame at " + }, + { + "bbox": [ + 313, + 532, + 555, + 593 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 313, + 532, + 555, + 593 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": "Challenges of StreamingVD tasks Firstly, increasing number of frames versus real-time requirement. Different from the offline tasks[17, 19, 30, 63, 68], streaming tasks[8, 14, 26, 45, 62, 72] require time alignment between event and response. Besides only accuracy constraint of offline task, streaming video tasks need to consider both accuracy and timeliness constraints. Offline tasks can trade resources and latency for better quality, while streaming task cannot." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 673, + 295, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 673, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 673, + 295, + 693 + ], + "type": "text", + "content": "1Invoking a quadratic-complexity LLM n times results in cubic complexity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "2A formal definition of an event [69]: \"A segment of time at a given location that is conceived by an observer to have a beginning and an end.\"" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13449" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 74, + 500, + 259 + ], + "blocks": [ + { + "bbox": [ + 107, + 74, + 500, + 259 + ], + "lines": [ + { + "bbox": [ + 107, + 74, + 500, + 259 + ], + "spans": [ + { + "bbox": [ + 107, + 74, + 500, + 259 + ], + "type": "image", + "image_path": "2b7857db5e85331170fbe38d99a83bb0d8bf03235c857757b86617c572cdf93a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 266, + 555, + 300 + ], + "lines": [ + { + "bbox": [ + 54, + 266, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 555, + 300 + ], + "type": "text", + "content": "Figure 2. Comparison between streaming dialogue (bottom) and traditional dialogue tasks (top). Streaming dialogue proactively generates responses at appropriate time based on user queries and perception, whereas traditional dialogue requires human triggering for each response." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 309, + 296, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 309, + 296, + 417 + ], + "spans": [ + { + "bbox": [ + 54, + 309, + 296, + 417 + ], + "type": "text", + "content": "Secondly, proactive response versus real-time requirement. Traditional video dialogue [13, 14, 45, 61, 62, 65] are reactive, responding only at the time of user query. However, real-world dialogue is often without human triggering. The user may lack the specific knowledge to formulate questions. Similarly, embodied AI should anticipate user needs rather than merely reacting to commands. However, proactive response means at every frame the model needs to judge, inceasing the model cost by " + }, + { + "bbox": [ + 54, + 309, + 296, + 417 + ], + "type": "inline_equation", + "content": "\\mathrm{O}(\\mathfrak{n})" + }, + { + "bbox": [ + 54, + 309, + 296, + 417 + ], + "type": "text", + "content": " times." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 426, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 296, + 665 + ], + "type": "text", + "content": "Existing works Conventional offline Video LLMs [10, 32, 33, 35, 73, 74] typically employ a visual encoder [18, 46, 70] and a projection module [28, 39] to process uniformly sampled frames. The output is concatenated with the tokenized question to form a sequence, which is then passed to an LLM decoder to predict a response. However, this approach is impractical for streaming tasks due to the high computational cost associated with processing a large number of frames " + }, + { + "bbox": [ + 56, + 426, + 296, + 665 + ], + "type": "inline_equation", + "content": "(T)" + }, + { + "bbox": [ + 56, + 426, + 296, + 665 + ], + "type": "text", + "content": ". Therefore, works such as [14, 62] have implemented a sliding-window attention mechanism and a KV-cache system to reduce computational overhead, or incorporated a parallel sampling strategy to enhance processing speed and reduce latency. However, these methods do not proactively trigger response when needed. In contrast, Videollm-Online [8] and MOD [60] utilize the LLM decoder to process every frame of streaming videos, resulting in a cubic computation cost. STREAMMIND, on the other hand, has the potential to process streaming videos at 100 fps, which is crucial for enabling interactive video analysis applications." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 672, + 222, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 222, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 222, + 684 + ], + "type": "text", + "content": "2.2. The Overview of STREAMMIND" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "For a streaming video sequence " + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\left[v^{t_s} \\leq t_i \\leq t_e\\right]" + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": ", STREAMMIND models the following process as shown in Fig. 3:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 308, + 555, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 416 + ], + "type": "text", + "content": "1. Perception Phase: With continuous streaming video, each incoming frame " + }, + { + "bbox": [ + 313, + 308, + 555, + 416 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 308, + 555, + 416 + ], + "type": "text", + "content": " is processed through the CLIP [46] encoder to extract spatial features. The features are then input to our proposed EPFE. EPFE updates its hidden state to correlate the frame's features across temporal dimensions and reduce redundant spatial information. This process generates a new perception token " + }, + { + "bbox": [ + 313, + 308, + 555, + 416 + ], + "type": "inline_equation", + "content": "[\\mathsf{F}_{per}^{t_i}]" + }, + { + "bbox": [ + 313, + 308, + 555, + 416 + ], + "type": "text", + "content": " and saves this token in the Perception Memory " + }, + { + "bbox": [ + 313, + 308, + 555, + 416 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathrm{per}}^{t_i}" + }, + { + "bbox": [ + 313, + 308, + 555, + 416 + ], + "type": "text", + "content": ". The perception phase can be formulated as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 361, + 423, + 555, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 423, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 361, + 423, + 555, + 441 + ], + "type": "interline_equation", + "content": "\\left[ \\mathrm {F} _ {p e r} ^ {t _ {i}} \\right] = \\operatorname {E P F E} \\left(\\operatorname {C L I P} \\left(v ^ {t _ {i}}\\right), \\mathrm {H} ^ {t _ {i} - 1}\\right) \\tag {2}", + "image_path": "8f0574fa493fba619ccf4ba20aa1dac181543917e3709dd77e6adffa8e7d4e6d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 445, + 553, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 553, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 553, + 456 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 445, + 553, + 456 + ], + "type": "inline_equation", + "content": "\\mathbb{H}^{t - 1}" + }, + { + "bbox": [ + 313, + 445, + 553, + 456 + ], + "type": "text", + "content": " denotes the hidden state from the previous time." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 457, + 554, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 554, + 518 + ], + "type": "text", + "content": "2. Cognition Gate Judgment: We introduce Cognition Gate " + }, + { + "bbox": [ + 313, + 457, + 554, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 457, + 554, + 518 + ], + "type": "text", + "content": " between perception and cognition. The input are user query [Prompt] and the generated perception token " + }, + { + "bbox": [ + 313, + 457, + 554, + 518 + ], + "type": "inline_equation", + "content": "[\\mathsf{F}_{per}^{t_i}]" + }, + { + "bbox": [ + 313, + 457, + 554, + 518 + ], + "type": "text", + "content": " at each time step " + }, + { + "bbox": [ + 313, + 457, + 554, + 518 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 313, + 457, + 554, + 518 + ], + "type": "text", + "content": ". The Cognition Gate learns to determine whether to trigger the cognition phase at " + }, + { + "bbox": [ + 313, + 457, + 554, + 518 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 313, + 457, + 554, + 518 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 343, + 524, + 553, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 524, + 553, + 539 + ], + "spans": [ + { + "bbox": [ + 343, + 524, + 553, + 539 + ], + "type": "interline_equation", + "content": "\\max P \\left(\\left[ \\operatorname {R e s} ^ {t _ {i}} \\right] \\mid \\mathcal {G} \\left(\\left[ \\text {P r o m p t} \\right], \\left[ F _ {\\text {p e r}} ^ {t _ {i}} \\right]\\right)\\right) \\tag {3}", + "image_path": "561bb90e8592edb1e32c4a4aefe985aeae03edd52d5d2bad456fdff9c4c4a059.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 545, + 554, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 569 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 545, + 554, + 569 + ], + "type": "inline_equation", + "content": "[\\mathrm{Res}^{t_i}] \\in \\{< /\\text{response}>, < /\\text{silence}>\\}" + }, + { + "bbox": [ + 313, + 545, + 554, + 569 + ], + "type": "text", + "content": " determines whether to trigger a response or remain silent." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 617 + ], + "type": "text", + "content": "3. Cognition Phase Upon receiving the trigger signal from the Cognition Gate, STREAMMIND samples perception tokens from the Perception Memory " + }, + { + "bbox": [ + 313, + 570, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathrm{per}}^{t_i}" + }, + { + "bbox": [ + 313, + 570, + 554, + 617 + ], + "type": "text", + "content": " to the Cognition Pooling as the input to LLM to generate response." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 624, + 417, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 417, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 417, + 636 + ], + "type": "text", + "content": "2.3. Perception Phase" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": "Motivation: To match the constant frame rate of streaming videos, ideally the perception phase should be constant computation cost for processing each frame. However, current video encoders generally extract local (dozens of frames) spatiotemporal features with constant cost, and input all the generated perception tokens to a transformer" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13450" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 69, + 542, + 210 + ], + "blocks": [ + { + "bbox": [ + 69, + 69, + 542, + 210 + ], + "lines": [ + { + "bbox": [ + 69, + 69, + 542, + 210 + ], + "spans": [ + { + "bbox": [ + 69, + 69, + 542, + 210 + ], + "type": "image", + "image_path": "1eb6a4808df4538648aaa78b2ebe560f1b83446fe1685610cc8a0d9b28a29c2d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 213, + 555, + 269 + ], + "lines": [ + { + "bbox": [ + 54, + 213, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 54, + 213, + 555, + 269 + ], + "type": "text", + "content": "Figure 3. STREAMMIND workflow. For each video frame, the perception stage goes through Clip for spatial features and the proposed EPFE for spatiotemporal features, to generate a single perception token saved in perception memory. The Cognition Gate determines whether to invoke LLM only based on the current perception token. If yes (a query-related event occurs), the tokens in perception memory are sampled to cognition pooling as the input for LLM for cognition response. The gate decouples perception and cognition to enable full-FPS perception and event-based cognition." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 281, + 295, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 281, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 54, + 281, + 295, + 316 + ], + "type": "text", + "content": "model to extract the long-term temporal relationships. The computational cost is " + }, + { + "bbox": [ + 54, + 281, + 295, + 316 + ], + "type": "inline_equation", + "content": "O(n^{3})" + }, + { + "bbox": [ + 54, + 281, + 295, + 316 + ], + "type": "text", + "content": ", obviously the lagger behind the streaming rate." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 316, + 295, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 316, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 54, + 316, + 295, + 437 + ], + "type": "text", + "content": "To solve the issue, we propose Event-Preserving Feature Extractor (EPFE) to replace the projection module used in existing video LLMs (e.g., cross-attention and Q-Former). EPFE is based on the SOTA state-space model, i.e., Selective State Space model [22]. As shown in Fig. 3 (blue box) and Eq. 4, for each frame, EPFE dynamically correlates the spatial features and the past internal states to generate a single perception token " + }, + { + "bbox": [ + 54, + 316, + 295, + 437 + ], + "type": "inline_equation", + "content": "\\left[\\mathbb{F}_{per}^{t_i}\\right]" + }, + { + "bbox": [ + 54, + 316, + 295, + 437 + ], + "type": "text", + "content": " as well as update its state. This enables a matched speed of streaming video and perception processing, for full FPS StreamingVD." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 443, + 295, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 443, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 104, + 443, + 295, + 456 + ], + "type": "interline_equation", + "content": "\\mathbf {h} _ {t + 1} = \\mathbf {A h} _ {t} + \\mathbf {B x} _ {t}, \\quad \\mathbf {y} _ {t} = \\mathbf {C h} _ {t} \\tag {4}", + "image_path": "582e5bae6f868d88bda6603ac6a36d2f7ded42fa018df17ac5b1a20cac7a5e4b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 462, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 462, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 54, + 462, + 295, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 54, + 462, + 295, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{A},\\mathbf{B}" + }, + { + "bbox": [ + 54, + 462, + 295, + 510 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 462, + 295, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 54, + 462, + 295, + 510 + ], + "type": "text", + "content": " are learnable state-space matrices for state, input, and output, respectively. " + }, + { + "bbox": [ + 54, + 462, + 295, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{x_t}" + }, + { + "bbox": [ + 54, + 462, + 295, + 510 + ], + "type": "text", + "content": " is the input frame of the current time step. " + }, + { + "bbox": [ + 54, + 462, + 295, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{h_t}" + }, + { + "bbox": [ + 54, + 462, + 295, + 510 + ], + "type": "text", + "content": " is the internal hidden state. " + }, + { + "bbox": [ + 54, + 462, + 295, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{y_t}" + }, + { + "bbox": [ + 54, + 462, + 295, + 510 + ], + "type": "text", + "content": " is the output, i.e., perception token." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 510, + 295, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 510, + 295, + 606 + ], + "spans": [ + { + "bbox": [ + 54, + 510, + 295, + 606 + ], + "type": "text", + "content": "Benefit from the strong temporal modeling capability of Selective SSM, our EPFE with only " + }, + { + "bbox": [ + 54, + 510, + 295, + 606 + ], + "type": "inline_equation", + "content": "56\\mathrm{M}" + }, + { + "bbox": [ + 54, + 510, + 295, + 606 + ], + "type": "text", + "content": " parameters can learn long-term event-level spatiotemporal features with constant cost. As Appendix Fig. 1 illustrates, perception tokens generated by EPFE can distinguish different events and noise frames, while maintaining a state of the event. Even after unrelated noise frames, the perception tokens are capable of refocusing on the event." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 612, + 149, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 149, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 149, + 624 + ], + "type": "text", + "content": "2.4. Cognition Gate" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 629, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 629, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 54, + 629, + 295, + 689 + ], + "type": "text", + "content": "Motivation: Cognition Gate needs to determine whether to invoke LLM in real-time, based on the current perception token and user query. For example, during the soccer match, some user query may only care LLM response for scoring moments, while others may need detailed tactical analysis." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 689, + 295, + 713 + ], + "type": "text", + "content": "However, we find it highly challenging. For real-time consideration, we first employ a simple Transformer block" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 281, + 553, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 553, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 553, + 389 + ], + "type": "text", + "content": "with cross-attention to associate the user query with the visual encoder's output for trigger decisions. However, it struggles to make accurate judgments at the right time step, as shown in Tab.5. We identify two key reasons for this: 1) The Cognition Gate only has access to information up to the current step, lacking a global temporal perspective for decision-making. 2) User queries may not directly correspond to specific video frames in a retrieval-like manner (e.g., guide me to repair a bicycle or cooking)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 392, + 553, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 553, + 453 + ], + "type": "text", + "content": "Therefore, the Cognition Gate requires the world knowledge of LLMs. However, due to the size of LLMs, even processing a single perception token as input incurs a high computational cost. Notably, the Gate only needs to generate a response/silence token for a single iteration." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 456, + 555, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 555, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 555, + 554 + ], + "type": "text", + "content": "To achieve it, We propose the Shallow Layer Transfer method, which initializes the Gate using the early layers of the LLM and then fine-tunes it in a supervised autoregressive manner with the StreamVD dataset. As shown in Sec. 4.2, we conducted extensive experiments to analyze its impact. The results indicate that the first few layers exhibit slight variations in behavior, allowing for scalability based on requirements. In this work, we use four layers." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 572, + 517, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 572, + 517, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 572, + 517, + 585 + ], + "type": "text", + "content": "2.5. The Training Strategy of STREAMMIND" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "Model Architecture As illustrated in Figure 3, the STREAMMIND video LLM integrates CLIP [46], EPFE, Cognition Gate, and the LLM. CLIP is the pre-trained spatial feature extractor. EPFE then takes the spatial features as input as well as its internal state to extract spatiotemporal features and generate a perception token. Cognition Gate then evaluates the current perception token and user query to determine whether to invoke LLM. To achieve this, it is needed to train EPFE, Cognition Gate, and also align the representation space of EPFE and LLM, as shown in Fig. 4." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13451" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 67, + 99, + 287, + 329 + ], + "blocks": [ + { + "bbox": [ + 61, + 76, + 268, + 97 + ], + "lines": [ + { + "bbox": [ + 61, + 76, + 268, + 97 + ], + "spans": [ + { + "bbox": [ + 61, + 76, + 268, + 97 + ], + "type": "text", + "content": "Algorithm 1: Streaming Video Dataset Construction" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "lines": [ + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "type": "text", + "content": "Require: Offline video dataset " + }, + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "type": "inline_equation", + "content": "C = \\{c_i,t_i\\}_{i = 1}^n (c_i" + }, + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "type": "text", + "content": " caption at time " + }, + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "type": "inline_equation", + "content": "t_i)" + }, + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "type": "text", + "content": " \nEnsure: Dataset " + }, + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "type": "text", + "content": " with captions meeting StreamVD requirements and Cognition Gate labels. \n1: Step 1: Preprocessing \n2: " + }, + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "type": "inline_equation", + "content": "D\\gets []" + }, + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "type": "text", + "content": " , cap " + }, + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "type": "inline_equation", + "content": "\\leftarrow c_{1}" + }, + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "type": "text", + "content": " ,time " + }, + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "type": "inline_equation", + "content": "\\leftarrow t_{s}" + }, + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "type": "text", + "content": " \n3: for " + }, + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "type": "inline_equation", + "content": "i = 2" + }, + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "type": "text", + "content": " to n do \n4: if " + }, + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "type": "inline_equation", + "content": "c_{i}\\neq" + }, + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "type": "text", + "content": " cap then \n5: Append (cap, time) to " + }, + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "type": "text", + "content": " \n6: cap " + }, + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "type": "inline_equation", + "content": "\\leftarrow c_{i}" + }, + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "type": "text", + "content": " ,time " + }, + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "type": "inline_equation", + "content": "\\leftarrow t_{i}" + }, + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "type": "text", + "content": " \n7: end if \n8: end for \n9: Append (cap, time) to " + }, + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "type": "text", + "content": " \n10: Step 2: Silence-Response Labeling \n11: for each frame " + }, + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "type": "inline_equation", + "content": "d_{i + 1}" + }, + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "type": "text", + "content": " do \n12: Label " + }, + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "type": "text", + "content": " withif captioned, else \n13: end for \n14: return " + }, + { + "bbox": [ + 67, + 99, + 287, + 329 + ], + "type": "inline_equation", + "content": "D" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 352, + 295, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 295, + 435 + ], + "type": "text", + "content": "Data Preparation Streaming datasets are constructed by processing existing offline datasets. StreamingVD tasks require the model to response as early as possible when relevant events occur while avoid redundant responses. To achieve this, we eliminate repetitive captions from the offline datasets and transform them into streaming datasets using the following methods, detailed in Algorithm 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 436, + 295, + 544 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 436, + 295, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 295, + 472 + ], + "type": "text", + "content": "- Preprocessing: Adjacent identical captions are merged, with the timestamp of the first occurrence being recorded as the annotation point for the caption." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "text", + "content": "- Silence-Response Labeling: To train the Cognition Gate, we insert " + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "inline_equation", + "content": "< / \\text{silence}>" + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "text", + "content": " tokens into video frames between two adjacent captions. These tokens act as indicators to guide the Cognition Gate in generating the " + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "inline_equation", + "content": "< / \\text{silence}>" + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "text", + "content": ". The initial frame corresponding to captioned events is labeled with the " + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "inline_equation", + "content": "< / \\text{response}>" + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "text", + "content": " token." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": "Training Strategy We adopt a two-stage training strategy, as shown in Fig. 4. In contrast to previous typical two-stage training approaches [29, 40, 71], our method uses a dedicated stage for Cognition Gate. In the first stage, the LLM and EPFE are jointly trained using video frames and captions from the streaming dataset, ensuring spatiotemporal feature alignment with the LLM. The second stage trains only the Cognition Gate, which is initialized from the shallow layers of the LLM. This training also follows the LLM autoregressive manner to generate " + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "inline_equation", + "content": "< / \\text{silence} >" + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "inline_equation", + "content": "< / \\text{response} >" + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": " tokens, determining whether to invoke the LLM. Since the response label is much sparser than silence (e.g. 310:1), we introduce a balancing weight into" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": "the standard Cross Entropy loss during training to balance the labels. Refer to Sec. 3.1 and Appendix Sec. 1 for more training details." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 118, + 434, + 194 + ], + "blocks": [ + { + "bbox": [ + 316, + 118, + 434, + 194 + ], + "lines": [ + { + "bbox": [ + 316, + 118, + 434, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 434, + 194 + ], + "type": "image", + "image_path": "739ae2abf9e281cfbd7ecf9e6f650725ee50be11cd644d0abd9ed25a0efaee6a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 199, + 554, + 266 + ], + "lines": [ + { + "bbox": [ + 313, + 199, + 554, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 554, + 266 + ], + "type": "text", + "content": "Figure 4. The two-stage training process of STREAMMIND. In the first stage, the EPFE spatiotemporal extractor and the LLM are jointly trained to align their representations. In the second stage, the Cognition module is trained to generate response/silence tokens. Both stages follow a supervised autoregressive training approach for language modeling." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 436, + 118, + 554, + 194 + ], + "blocks": [ + { + "bbox": [ + 436, + 118, + 554, + 194 + ], + "lines": [ + { + "bbox": [ + 436, + 118, + 554, + 194 + ], + "spans": [ + { + "bbox": [ + 436, + 118, + 554, + 194 + ], + "type": "image", + "image_path": "1e4c820793b4f84127f077e2a6c5b6bc3df4bdcb771bd1d8e826e59ec4f93c33.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 288, + 391, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 288, + 391, + 301 + ], + "spans": [ + { + "bbox": [ + 314, + 288, + 391, + 301 + ], + "type": "text", + "content": "3. Experiment" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 307, + 441, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 441, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 441, + 320 + ], + "type": "text", + "content": "3.1. Experimental Settings" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 324, + 554, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 554, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 554, + 408 + ], + "type": "text", + "content": "Baseline and Datasets Since existing works focusing on streamingVD [8, 26, 60], except for VideoLLM-Online [8], do not provide open-source data or training methods, to demonstrate our outstanding performance, we re-implemented all previous works (VideoLLM-Online [8] and VideoLLM-MOD [60]) in the StreamingVD domain as baselines, using consistent settings across all experiments." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 408, + 554, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 554, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 554, + 540 + ], + "type": "text", + "content": "For dataset selection, since the Ego4D dataset [21] is inherently collected in a streaming manner, containing various queries, real-time responses, and narrations, and has been used as an main evaluation dataset in works [8, 31, 60], we selected it as one of our evaluation datasets. Furthermore, to further demonstrate the effectiveness of our approach on long videos, we also conducted experiments on the SoccerNet-Caption dataset [20]. This dataset includes 471 complete soccer match videos with a total duration of 715 hours, where each video is approximately 45 minutes long and provides commentary at appropriate timestamps." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 545, + 554, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 664 + ], + "type": "text", + "content": "Implementation Details: Model training was conducted on videos sampled at 2 fps, using 8 NVIDIA A100 GPUs. Each stage of the training process was conducted for one epoch. To optimize the learning rate, we applied the CosineAnnealingLR scheduler, with learning rates set to " + }, + { + "bbox": [ + 313, + 545, + 554, + 664 + ], + "type": "inline_equation", + "content": "2e - 5" + }, + { + "bbox": [ + 313, + 545, + 554, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 545, + 554, + 664 + ], + "type": "inline_equation", + "content": "2e - 6" + }, + { + "bbox": [ + 313, + 545, + 554, + 664 + ], + "type": "text", + "content": " respectively. Additionally, specific techniques were applied to the design of the Cognition Gate and methods to address the imbalance in the Silence-Response ratio, which are detailed in Section 4 and the supplementary materials." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 672, + 417, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 417, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 417, + 684 + ], + "type": "text", + "content": "3.2. Evaluate Metrics" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "To evaluate the effectiveness of STREAMMIND, we focus on assessing its Timing Alignment Capability and Lan" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13452" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 92, + 87, + 519, + 196 + ], + "blocks": [ + { + "bbox": [ + 236, + 71, + 373, + 82 + ], + "lines": [ + { + "bbox": [ + 236, + 71, + 373, + 82 + ], + "spans": [ + { + "bbox": [ + 236, + 71, + 373, + 82 + ], + "type": "text", + "content": "Table 1. Online Experiments Results." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 92, + 87, + 519, + 196 + ], + "lines": [ + { + "bbox": [ + 92, + 87, + 519, + 196 + ], + "spans": [ + { + "bbox": [ + 92, + 87, + 519, + 196 + ], + "type": "table", + "html": "
DataSetMethodTiming Alignment ↑Language Modeling Capability ↑Training Cost
TriggerAccTimValBLUE-1BLUE-4METEORROUGE-L
Ego4dVideoLLM-Online32.34%29.66%66.0135.2531.1263.0624h
VideoLLM-MoD32.36%29.65%65.3435.2130.6563.0214h
Ours43.34%39.73%67.1239.2631.6065.71(11+8)h
SoccerVideoLLM-Online31.25%28.34%75.3664.2350.9281.5712h
VideoLLM-MoD31.24%28.12%74.9664.1850.2481.597h
Ours52.18%47.36%82.7866.7051.4382.04(5+3)h
", + "image_path": "35e34eefc2f14d46ffb7295708eb3bf6b262d424061ca271ffc5de384f6ae975.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 59, + 262, + 293, + 352 + ], + "blocks": [ + { + "bbox": [ + 55, + 213, + 295, + 258 + ], + "lines": [ + { + "bbox": [ + 55, + 213, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 295, + 258 + ], + "type": "text", + "content": "Table 2. Online Experiments Results. LION-FS was released on arXiv just two days before our submission, but its code was not released. Therefore, we could only obtain the Ego4D results based on its paper." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 262, + 293, + 352 + ], + "lines": [ + { + "bbox": [ + 59, + 262, + 293, + 352 + ], + "spans": [ + { + "bbox": [ + 59, + 262, + 293, + 352 + ], + "type": "table", + "html": "
DataSetMethodTimeDiff ↓Fluency ↑PPL ↓Correctiness ↑
Ego4dVideoLLM-Online2.0445.0%2.4348.1%
VideoLLM-MOD2.0445.2%2.4148.9%
LION-FS2.1546.1%2.0952.4%
Ours1.8960.2%2.0277.3%
SoccerVideoLLM-Online15.6246.35%1.7953.5%
VideoLLM-MOD15.6845.34%1.8053.3%
Ours14.0270.35%1.5989.2%
", + "image_path": "b180a5df8550470273b4296b8853f80373ffcc582cc9bc09be9705d069f4b18e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 58, + 399, + 294, + 511 + ], + "blocks": [ + { + "bbox": [ + 55, + 361, + 295, + 396 + ], + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 396 + ], + "type": "text", + "content": "Table 3. Results on COIN benchmarks(left to right): step recognition, task recognition, next forecasting, procedure forecasting, procedure forecasting with a goal." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 399, + 294, + 511 + ], + "lines": [ + { + "bbox": [ + 58, + 399, + 294, + 511 + ], + "spans": [ + { + "bbox": [ + 58, + 399, + 294, + 511 + ], + "type": "table", + "html": "
MethodNot use HT 100MCOINBenchmark Top-1 Accuracy ↑
StepTaskNextProcProg.+
ClipBERT✓30.865.4---
TimeFormer×46.585.334.017.040.1
Paprika×51.085.843.2--
DistantSup×54.190.039.4--
VideoTF×56.591.042.440.246.4
ProcedureVRL×56.990.846.8--
VideoTaskGraph×57.290.540.2--
VideoLLM-online-7B-v1✓59.892.148.147.952.9
VideoLLM-online -8B-v1+✓63.192.749.149.854.1
VideoLLM-MOD✓63.492.849.749.853.3
Ours✓63.793.249.949.854.2
", + "image_path": "ea0093b07efddd34fdf0971ff3ac94448925eec2e04be713771e33f18110cc8c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 60, + 549, + 293, + 645 + ], + "blocks": [ + { + "bbox": [ + 55, + 522, + 295, + 545 + ], + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 545 + ], + "type": "text", + "content": "Table 4. Results on Ego4DLTA benchmark, evaluated on public server. ED@Z=20 denotes editing distance for future 20 actions." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 549, + 293, + 645 + ], + "lines": [ + { + "bbox": [ + 60, + 549, + 293, + 645 + ], + "spans": [ + { + "bbox": [ + 60, + 549, + 293, + 645 + ], + "type": "table", + "html": "
MethodNot use Ego VLPEnd-to -End?Ego4D LTA ED@Z=20
Verb ↓Noun ↓Action ↓
CLIP✓✓0.7390.7690.941
EgoT2✓✓0.7220.7640.935
I-CVAE✓✓0.7530.7490.931
HierVL✓✓0.7240.7350.928
VideoLLM×✓0.7210.7250.921
VideoLLM-online-7B-v1✓✓0.6970.6980.897
VideoLLM-online-8B-v1+✓✓0.6890.6710.884
VideoLLM-MOD✓✓0.6890.6760.884
Ours✓✓0.6890.6550.881
", + "image_path": "1b07ab548ab481fb902ac93d597432b2d85ee43db52883e282ea0325391f49fa.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "guage Modeling Capability. Existing evaluation metrics include LM-PPL (Language Modeling Perplexity) and LM-Correctness, which evaluate language modeling accuracy at specific timestamps. Additionally, the Time Difference" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 215, + 555, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 215, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 215, + 555, + 264 + ], + "type": "text", + "content": "(TimeDiff) metric is used to measure the temporal alignment of responses in an streaming video LLM, while the Fluency metric aims to provide a holistic evaluation of both linguistic quality and temporal consistency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 324, + 264, + 523, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 264, + 523, + 276 + ], + "spans": [ + { + "bbox": [ + 324, + 264, + 523, + 276 + ], + "type": "text", + "content": "However, these metrics have several limitations:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 277, + 555, + 384 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 314, + 277, + 555, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 277, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 314, + 277, + 555, + 336 + ], + "type": "text", + "content": "- LM-PPL and LM-Correctness focus solely on the exact alignment between model outputs and reference narrations, ignoring aspects such as synonym substitutions, sentence fluency, and the ability to capture critical information." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 337, + 554, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 337, + 554, + 384 + ], + "spans": [ + { + "bbox": [ + 314, + 337, + 554, + 384 + ], + "type": "text", + "content": "- TimeDiff, as a metric for timing alignment Capability, evaluates temporal alignment only within a single dialogue turn. It does not assess the Cognition Gate's overall decision-making across the entire video stream." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 385, + 554, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 554, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 554, + 409 + ], + "type": "text", + "content": "To address these limitations, we introduce the following metrics:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 410, + 554, + 590 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 314, + 410, + 554, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 410, + 554, + 446 + ], + "spans": [ + { + "bbox": [ + 314, + 410, + 554, + 446 + ], + "type": "text", + "content": "- Trigger Accuracy (TriggerAcc): To evaluate whether the streaming video LLM responds at the correct time steps during entire streaming video." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 446, + 554, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 446, + 554, + 506 + ], + "spans": [ + { + "bbox": [ + 314, + 446, + 554, + 506 + ], + "type": "text", + "content": "- Timing Validity (TimVal): To comprehensively evaluate whether the streaming video LLM consistently makes the correct decisions throughout the streaming video—speaking when necessary and remaining silent when appropriate." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 506, + 554, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 506, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 314, + 506, + 554, + 590 + ], + "type": "text", + "content": "- BLUE [43], METEOR [2], ROUGE-L [34]: Widely used evaluation metrics in the field of text generation [6, 9, 38] and image narration [7, 44, 48], which fully consider factors such as synonym matching, stemming, word order, and key semantic expressions. These metrics enable a comprehensive assessment of dialogues in the StreamingVD process." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 599, + 554, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 554, + 624 + ], + "type": "text", + "content": "3.3. Comprehensive Evaluation on Streaming and Offline Video" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": "Online Experiments: We compared our model with all existing streaming dialogue models [8, 31, 60] in Tab 1 and Tab 2. Under comparable training costs, experiments on real-time Ego4D and SoccerNet demonstrate superior performance across all metrics. Additionally, STREAMMIND can process video streams at a " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": " higher FPS. This improvement is attributed to the event-level modeling capability." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13453" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "content": "ity of EPFE and the strong timing alignment capability of the cognition gate in the perception stage, which together provide the LLM in the cognition phase with a concise yet comprehensive representation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 125, + 295, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 125, + 295, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 125, + 295, + 222 + ], + "type": "text", + "content": "Offline Experiments: We also leveraged the COIN [53] and Ego4D LTA [24] benchmark to demonstrate STREAM-MIND effectiveness in traditional offline video scenarios, covering six common benchmarks: step recognition, step forecasting, task summarization, procedure forecasting, and procedure forecasting with a goal. As shown in Tab 3 and Tab 4, our approach achieves state-of-the-art performance across all tasks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 228, + 220, + 242 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 220, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 220, + 242 + ], + "type": "text", + "content": "3.4. Real-time Inference Efficiency" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 245, + 295, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 245, + 295, + 328 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 295, + 328 + ], + "type": "text", + "content": "Standard frame rates for films are typically 24 FPS, which is considered the minimum requirement for smooth motion perception. In contrast, games often operate at much higher frame rates, ranging from 30 FPS to over 100 FPS, to ensure a immersive experience. To validate the real-time performance of STREAMMIND, we conducted inference efficiency evaluations on NVIDIA A100 and H100 GPUs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 329, + 295, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 329, + 295, + 400 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 295, + 400 + ], + "type": "text", + "content": "Specifically, we compared VideoLLM-Online, VideoLLM-MOD, and STREAMMIND by measuring their actual processing time for 1-second video across input frame rates from 5 to 100 FPS. As shown in Figure 5, a processing time under 1 second indicates real-time performance without disrupting playback." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 401, + 295, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 295, + 485 + ], + "type": "text", + "content": "The results reveal that VideoLLM-Online and VideoLLM-MOD struggle to process video streams beyond 10 FPS, leading to potential latency issues in real-world applications. In contrast, our method achieves true real-time performance, efficiently handling both standard film/TV frame rates and highRefresh gaming scenarios up to 100 FPS, ensuring real-time responses." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 495, + 117, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 117, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 117, + 507 + ], + "type": "text", + "content": "4. Ablation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 514, + 295, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 295, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 295, + 528 + ], + "type": "text", + "content": "4.1. Impact of Silence-Response Sample Imbalance" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 531, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 295, + 581 + ], + "type": "text", + "content": "In this section, we explored the imbalance in the Silence-Response ratio within the dataset and derived an empirical balancing formula through experiments. For detailed experimental results, please refer to the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 586, + 227, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 227, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 227, + 600 + ], + "type": "text", + "content": "4.2. The Ablation of Cognition Gate" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 603, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 295, + 640 + ], + "type": "text", + "content": "The Cognition Gate plays a crucial role throughout the perception-cognition interleaving process. We conducted detailed ablation experiments to analyze its impact." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "type": "text", + "content": "The Architecture of Cognition Gate The structure of the gate directly determines its capabilities. To this end, we constructed multiple classification networks using MLP projectors, Cross-Attention, and Transformers. [54] Additionally, we built a set of autoregressive networks using an" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 337, + 72, + 533, + 194 + ], + "blocks": [ + { + "bbox": [ + 337, + 72, + 533, + 194 + ], + "lines": [ + { + "bbox": [ + 337, + 72, + 533, + 194 + ], + "spans": [ + { + "bbox": [ + 337, + 72, + 533, + 194 + ], + "type": "image", + "image_path": "177b0ddf4d7b4d1cbcdbb3dd92ae00c24c8c053d74c13cd4ee0f00aa452170ce.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 336, + 197, + 533, + 320 + ], + "blocks": [ + { + "bbox": [ + 336, + 197, + 533, + 320 + ], + "lines": [ + { + "bbox": [ + 336, + 197, + 533, + 320 + ], + "spans": [ + { + "bbox": [ + 336, + 197, + 533, + 320 + ], + "type": "image", + "image_path": "1d9a738ef34a21e23a4bf4748ba50df552804677ae8aa33db725f0e165ea4397.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 326, + 555, + 360 + ], + "lines": [ + { + "bbox": [ + 313, + 326, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 555, + 360 + ], + "type": "text", + "content": "Figure 5. Comparison of running time for processing 1 second of streaming video frames using different methods on A100 (Top) and H100 (Bottom)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 377, + 555, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 555, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 555, + 414 + ], + "type": "text", + "content": "LLM block. The results are shown in Tab.5. It is evident that the LLM-like structure achieved better performance. For a detailed analysis, please refer to Section 2.4." + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 318, + 437, + 552, + 500 + ], + "blocks": [ + { + "bbox": [ + 323, + 422, + 545, + 434 + ], + "lines": [ + { + "bbox": [ + 323, + 422, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 323, + 422, + 545, + 434 + ], + "type": "text", + "content": "Table 5. Ablation Study on the architecture of cognition gate" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 437, + 552, + 500 + ], + "lines": [ + { + "bbox": [ + 318, + 437, + 552, + 500 + ], + "spans": [ + { + "bbox": [ + 318, + 437, + 552, + 500 + ], + "type": "table", + "html": "
ArchitectureTimeDiff ↓TriggerAcc ↑TimVal ↑
Linear Layer5.0620.13%17.65%
MLP projector + Linear4.3421.75%18.33%
Transformers + Linear3.7821.58%18.97%
Cross-Attention + Linear3.6424.34%20.36%
Single LLM block2.2235.35%32.11%
", + "image_path": "6c308948da08b3d3b907b08d44232da795d8a1428cd2728eb2f17606e06a7f3a.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 518, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 555, + 662 + ], + "type": "text", + "content": "Initialization Strategies We investigate three initialization strategies for the Cognition Gate to enhance its effectiveness: (i) Random Initialization: All layers are initialized from a standard random distribution without utilizing prior knowledge from the LLM. (ii) SkipBlock Initialization: The Cognition Gate is initialized by uniformly selecting layers from the LLM, with the number of skipped layers determined by the ratio of the Cognition Gate's depth to the total number of LLM layers. (iii) EarlyBlock Initialization: The first few layers of the LLM are directly used to initialize the Cognition Gate, leveraging the LLM's early-stage processing capabilities." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "Number of Blocks The Cognition Gate consists of multiple layers, and its depth plays a crucial role in balancing computational efficiency and perceptual accuracy." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13454" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 87, + 296, + 133 + ], + "blocks": [ + { + "bbox": [ + 73, + 71, + 277, + 82 + ], + "lines": [ + { + "bbox": [ + 73, + 71, + 277, + 82 + ], + "spans": [ + { + "bbox": [ + 73, + 71, + 277, + 82 + ], + "type": "text", + "content": "Table 6. Ablation Study on the Strategy of Initialization" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 87, + 296, + 133 + ], + "lines": [ + { + "bbox": [ + 56, + 87, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 56, + 87, + 296, + 133 + ], + "type": "table", + "html": "
Strategy of InitializationTimeDiff ↓TriggerAcc ↑TimVal ↑
Random2.1039.65%37.35%
SkipBlock1.8940.23%37.67%
EarlyBlock1.8943.34%39.73%
", + "image_path": "5f8b654b184b558949a740fadf5b44477c964e21bb7f0a448cecca09e66889f0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 157, + 294, + 221 + ], + "blocks": [ + { + "bbox": [ + 73, + 142, + 277, + 153 + ], + "lines": [ + { + "bbox": [ + 73, + 142, + 277, + 153 + ], + "spans": [ + { + "bbox": [ + 73, + 142, + 277, + 153 + ], + "type": "text", + "content": "Table 7. Ablation Study on the Strategy of Initialization" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 157, + 294, + 221 + ], + "lines": [ + { + "bbox": [ + 58, + 157, + 294, + 221 + ], + "spans": [ + { + "bbox": [ + 58, + 157, + 294, + 221 + ], + "type": "table", + "html": "
#Num of BlockTimeDiff ↓TriggerAcc ↑TimVal ↑
21.9241.53%37.65%
31.9042.35%38.34%
41.8943.34%39.73%
51.8842.67%38.56%
61.8942.56%38.34%
", + "image_path": "f0f618264d24e09550e7330fd876a227b88e5387b19f4b750ffad76dba1e3c56.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 240, + 296, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 296, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 296, + 336 + ], + "type": "text", + "content": "The results in Tab.6 and Tab.7 show that fine-tuning the shallow layers of the LLM yields the best performance. We attribute this to the fact that gate operates within a limited decision space (binary: Yes/No), making deeper layers—primarily responsible for nuanced language generation—unnecessary. Thus, we propose the Shallow Layer Transfer method to construct the Cognition Gate, as detailed in Section 2.4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 345, + 296, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 345, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 296, + 369 + ], + "type": "text", + "content": "4.3. The Performance of Event-preserving Feature extractor" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 373, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 518 + ], + "type": "text", + "content": "To evaluate the role of EPFE in the perception phase, we compare it with commonly used extractors in VideoLLMs, including: 1) The Q-Former [28]. 2) The Spatial-Temporal Convolution Connector (STC Connector) [12], which has demonstrated outstanding performance in videolm and has been widely adopted [29, 59, 71]. Notably, STC employs 3D convolution for spatial-temporal aggregation. To preserve local visual patterns during spatial compression, it follows work [5] by incorporating a RegStage block [47] before and after the 3D convolution, which has been shown to enhance spatial understanding. Experimental results indicate that EPFE achieves superior performance." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 69, + 535, + 284, + 582 + ], + "blocks": [ + { + "bbox": [ + 99, + 520, + 252, + 531 + ], + "lines": [ + { + "bbox": [ + 99, + 520, + 252, + 531 + ], + "spans": [ + { + "bbox": [ + 99, + 520, + 252, + 531 + ], + "type": "text", + "content": "Table 8. Perceptive Judgement Capability" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 535, + 284, + 582 + ], + "lines": [ + { + "bbox": [ + 69, + 535, + 284, + 582 + ], + "spans": [ + { + "bbox": [ + 69, + 535, + 284, + 582 + ], + "type": "table", + "html": "
MethodTimeDiff ↓TriggerAcc ↑TimVal ↑
Q-former3.7826.65 %25.31%
STC3.5627,54%26.87%
EPFE1.8943.34%39.73%
", + "image_path": "64d227cedc0229c59fcafdf4e1f53875aa35e7f97d1a843f7423d174b07db96d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 604, + 143, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 143, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 143, + 616 + ], + "type": "text", + "content": "5. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 624, + 172, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 172, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 172, + 635 + ], + "type": "text", + "content": "5.1. Offline VideoLLMs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "In recent years, there has been a surge of interest in leveraging Large Language Models (LLMs) for video understanding, leading to the development of offline videoLLMs. The biggest challenge for it is managing memory requirements [4, 11, 15, 23, 50, 51] and compressing redundant frame features [12, 25, 27, 52, 76]. LaViLa [37] addresses" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": "this challenge by transforming LLMs into narrators that generate detailed descriptions of long videos from visual inputs. Other models, such as ChatVideo [56] and MMVID [36], convert videos into text to improve comprehension. MovieChat [50] combines all frame features using an averaging strategy. However, in the context of online streaming, there has been limited exploration into how VideoLLMs can meet the demands of temporal alignment and real-time processing for streaming video inputs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 196, + 430, + 208 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 196, + 430, + 208 + ], + "spans": [ + { + "bbox": [ + 313, + 196, + 430, + 208 + ], + "type": "text", + "content": "5.2. Online VideoLLMs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 214, + 555, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 214, + 555, + 454 + ], + "spans": [ + { + "bbox": [ + 313, + 214, + 555, + 454 + ], + "type": "text", + "content": "Recent studies have introduced VideoLLMs specifically designed for online stream understanding, while most existing works [14, 42, 45, 62, 72, 77] focus on improving offline video efficiency to meet real-time requirements. However, they fail to proactively decide when to respond based on user requirements. For example, ReKV [14] implemented a sliding-window attention mechanism and a KV-cache system to reduce computational overhead. To address this, VideoLLM-Online [8] introduces per-step LLM invocation for online video dialogue, as shown in Fig.1. However, its performance is constrained by limited per-frame processing capability due to the absence of effective frame compression strategies. VideoLLM-MOD [60] mitigates this issue by incorporating a mixture-of-depth approach for efficient visual token computation, enabling higher visual input resolutions. Building on MOD, LION-FS [31] proposes Token Aggregation Router and Token DroppingRouter as the Fast Path, adaptively aggregating distinct features while discarding redundant ones. However, they remain limited to processing videos at only 10 FPS." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 457, + 556, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 556, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 556, + 578 + ], + "type": "text", + "content": "Overall, prior work remains restricted by the contradiction between sequential full-frame processing and the extremely high efficiency requirements of streaming video, lacking breakthrough paradigm designs explicitly suited for streaming frame inputs. To address these challenges, our approach introduces a novel perception-cognition interleaving paradigm and proposes the STREAMMIND framework, enabling VideoLLM to achieve proactive responses and real-time streaming video understanding, with a maximum throughput of up to 100 FPS." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 597, + 388, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 388, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 388, + 609 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "content": "In this paper, we introduce a novel perception-cognition interleaving paradigm called \"event-gated LLM invocation\" and propose the STREAMMIND framework, addressing the contradiction between sequential full-frame processing and the extremely high efficiency requirements of streaming video. As a result, STREAMMIND can process video at up to 100 FPS, achieving the SOTA performance across all existing metrics and benchmarks." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13455" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 167, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 167, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 167, + 85 + ], + "type": "text", + "content": "7. Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 214 + ], + "type": "text", + "content": "We gratefully acknowledge the support from the following foundations and institutions. Ting Cao was supported in part by the Tsinghua University (AIR)-AsiaInfo Technologies (China), Inc. Joint Research Center for 6G Network and Intelligent Computing. Hao Wu was supported in part by the National Natural Science Foundation of China (NSFC) under Grants 62432004 and 62302207, and by the Science Foundation for Youths of Jiangsu Province under Grant BK20220772. Xin Ding and Zhibo Chen were supported in part by the NSFC under Grants 62371434 and 62021001." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 234, + 115, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 234, + 115, + 247 + ], + "spans": [ + { + "bbox": [ + 56, + 234, + 115, + 247 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 254, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 61, + 254, + 294, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 254, + 294, + 299 + ], + "spans": [ + { + "bbox": [ + 61, + 254, + 294, + 299 + ], + "type": "text", + "content": "[1] Roberto Amoroso, Gengyuan Zhang, Rajat Koner, Lorenzo Baraldi, Rita Cucchiara, and Volker Tresp. Perceive, query & reason: Enhancing video qa with question-guided temporal queries. arXiv preprint arXiv:2412.19304, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 300, + 294, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 300, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 61, + 300, + 294, + 354 + ], + "type": "text", + "content": "[2] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 356, + 294, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 356, + 294, + 422 + ], + "spans": [ + { + "bbox": [ + 61, + 356, + 294, + 422 + ], + "type": "text", + "content": "[3] Pietro Battistoni, Andrea Antonio Cantone, Mariarosaria Esposito, Rita Francese, Francesca Pia Perillo, Marco Romano, Monica Sebillo, and Giuliana Vitiello. Using artificial intelligence and companion robots to improve home healthcare for the elderly. In International Conference on Human-Computer Interaction, pages 3–17. Springer, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 423, + 294, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 423, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 62, + 423, + 294, + 477 + ], + "type": "text", + "content": "[4] Jiarui Cai, Mingze Xu, Wei Li, Yuanjun Xiong, Wei Xia, Zhuowen Tu, and Stefano Soatto. Memot: Multi-object tracking with memory. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8090-8100, 2022. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 479, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 479, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 479, + 294, + 533 + ], + "type": "text", + "content": "[5] Junbum Cha, Wooyoung Kang, Jonghwan Mun, and Byungseok Roh. Honeybee: Locality-enhanced projector for multimodal llm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13817-13827, 2024. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 535, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 535, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 535, + 294, + 590 + ], + "type": "text", + "content": "[6] Yupeng Chang, Xu Wang, Jindong Wang, Yuan Wu, Linyi Yang, Kaijie Zhu, Hao Chen, Xiaoyuan Yi, Cunxiang Wang, Yidong Wang, et al. A survey on evaluation of large language models. ACM transactions on intelligent systems and technology, 15(3):1-45, 2024. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 591, + 294, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 591, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 62, + 591, + 294, + 647 + ], + "type": "text", + "content": "[7] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3558-3568, 2021. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 647, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 294, + 712 + ], + "type": "text", + "content": "[8] Joya Chen, Zhaoyang Lv, Shiwei Wu, Kevin Qinghong Lin, Chenan Song, Difei Gao, Jia-Wei Liu, Ziteng Gao, Dongxing Mao, and Mike Zheng Shou. Videollm-online: Online video large language model for streaming video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18407-18418, 2024. 1, 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 320, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 553, + 116 + ], + "type": "text", + "content": "[9] Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji, and Quanquan Gu. Self-play fine-tuning converts weak language models to strong language models. arXiv preprint arXiv:2401.01335, 2024. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 118, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 183 + ], + "type": "text", + "content": "[10] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24185–24198, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 184, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 553, + 228 + ], + "type": "text", + "content": "[11] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In European Conference on Computer Vision, pages 640-658. Springer, 2022. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 228, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 553, + 282 + ], + "type": "text", + "content": "[12] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "type": "text", + "content": "[13] Shangzhe Di and Weidi Xie. Grounded question-answering in long egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12934–12943, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 327, + 553, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 327, + 553, + 382 + ], + "spans": [ + { + "bbox": [ + 317, + 327, + 553, + 382 + ], + "type": "text", + "content": "[14] Shangzhe Di, Zhelun Yu, Guanghao Zhang, Haoyuan Li, Hao Cheng, Bolin Li, Wanggui He, Fangxun Shu, Hao Jiang, et al. Streaming video question-answering with in-context video kv-cache retrieval. In The Thirteenth International Conference on Learning Representations. 1, 2, 3, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 383, + 553, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 553, + 435 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 553, + 435 + ], + "type": "text", + "content": "[15] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. In European Conference on Computer Vision, pages 75-92. Springer, 2024. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 437, + 553, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 437, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 317, + 437, + 553, + 481 + ], + "type": "text", + "content": "[16] Feiteng Fang, Yuelin Bai, Shiwen Ni, Min Yang, Xiaojun Chen, and Ruifeng Xu. Enhancing noise robustness of retrieval-augmented language models with adaptive adversarial training. arXiv preprint arXiv:2405.20978, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 482, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 553, + 536 + ], + "type": "text", + "content": "[17] Xinyu Fang, Kangrui Mao, Haodong Duan, Xiangyu Zhao, Yining Li, Dahua Lin, and Kai Chen. Mmbench-video: A long-form multi-shot benchmark for holistic video understanding. Advances in Neural Information Processing Systems, 37:89098-89124, 2025. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 537, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 553, + 602 + ], + "type": "text", + "content": "[18] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19358-19369, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 603, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 603, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 317, + 603, + 553, + 658 + ], + "type": "text", + "content": "[19] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "text", + "content": "[20] Silvio Giancola, Mohieddine Amine, Tarek Dghaily, and Bernard Ghanem. Soccernet: A scalable dataset for action spotting in soccer videos. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 1711-1721, 2018. 2, 5" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13456" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 148 + ], + "type": "text", + "content": "[21] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18995-19012, 2022. 2, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "text", + "content": "[22] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. CoRR, abs/2312.00752, 2023. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 251 + ], + "type": "text", + "content": "[23] Bo He, Hengduo Li, Young Kyun Jang, Menglin Jia, Xuefei Cao, Ashish Shah, Abhinav Shrivastava, and Ser-Nam Lim. Ma-lmm: Memory-augmented large multimodal model for long-term video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13504-13514, 2024. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 252, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 252, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 252, + 294, + 296 + ], + "type": "text", + "content": "[24] Daoji Huang, Otmar Hilliges, Luc Van Gool, and Xi Wang. Palm: Predicting actions through language models@ ego4d long-term action anticipation challenge 2023. arXiv preprint arXiv:2306.16545, 2023. 2, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 297, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 353 + ], + "type": "text", + "content": "[25] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In European Conference on Computer Vision, pages 202-218. Springer, 2024. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "type": "text", + "content": "[26] Zhenpeng Huang, Xinhao Li, Jiaqi Li, Jing Wang, Xi-angyu Zeng, Cheng Liang, Tao Wu, Xi Chen, Liang Li, and Limin Wang. Online video understanding: A comprehensive benchmark and memory-augmented method. arXiv preprint arXiv:2501.00584, 2024. 1, 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 410, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 410, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 294, + 464 + ], + "type": "text", + "content": "[27] Yang Jin, Zhicheng Sun, Kun Xu, Liwei Chen, Hao Jiang, Quzhe Huang, Chengru Song, Yuliang Liu, Di Zhang, Yang Song, et al. Video-lavit: Unified video-language pre-training with decoupled visual-motional tokenization. arXiv preprint arXiv:2402.03161, 2024. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 466, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 294, + 520 + ], + "type": "text", + "content": "[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2, 3, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 522, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 522, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 522, + 294, + 566 + ], + "type": "text", + "content": "[29] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 5, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 567, + 294, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 294, + 632 + ], + "type": "text", + "content": "[30] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 635, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 667 + ], + "type": "text", + "content": "[31] Wei Li, Bing Hu, Rui Shao, Leyang Shen, and Liqiang Nie. Lion-fs: Fast & slow video-language thinker as online video assistant. arXiv preprint arXiv:2503.03663, 2025. 5, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "type": "text", + "content": "[32] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2024. 1, 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "text", + "content": "[33] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "type": "text", + "content": "[34] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "text", + "content": "[35] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 26689-26699, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "text", + "content": "[36] Kevin Lin, Faisal Ahmed, Linjie Li, Chung-Ching Lin, Ehsan Azarnasab, Zhengyuan Yang, Jianfeng Wang, Lin Liang, Zicheng Liu, Yumao Lu, et al. Mm-vid: Advancing video understanding with gpt-4v (isdiction). arXiv preprint arXiv:2310.19773, 2023. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "text", + "content": "[37] Xingyu Liu, Joon-Young Lee, and Hailin Jin. Learning video representations from correspondence proposals. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 4273–4281, 2019. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 310, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 353 + ], + "type": "text", + "content": "[38] Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. G-eval: Nlg evaluation using gpt-4 with better human alignment. arXiv preprint arXiv:2303.16634, 2023. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 354, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 408 + ], + "type": "text", + "content": "[39] Ruipu Luo, Ziwang Zhao, Min Yang, Junwei Dong, Da Li, Pengcheng Lu, Tao Wang, Linmei Hu, Minghui Qiu, and Zhongyu Wei. Valley: Video assistant with large language model enhanced ability. arXiv preprint arXiv:2306.07207, 2023. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "type": "text", + "content": "[40] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 1, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 456, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 488 + ], + "type": "text", + "content": "[41] Kelly Merrill Jr, Jihyun Kim, and Chad Collins. Ai companions for lonely individuals and the role of social presence. Communication Research Reports, 39(2):93-103, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 490, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 553, + 555 + ], + "type": "text", + "content": "[42] Sunny Panchal, Apratim Bhattacharyya, Guillaume Berger, Antoine Mercier, Cornelius Böhm, Florian Dietrichkeit, Reza Pourreza, Xuanlin Li, Pulkit Madan, Mingu Lee, et al. What to say and when to say it: Live fitness coaching as a testbed for situated interaction. Advances in Neural Information Processing Systems, 37:75853-75882, 2024. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 557, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 610 + ], + "type": "text", + "content": "[43] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 613, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 677 + ], + "type": "text", + "content": "[44] Jordi Pont-Tuset, Jasper Uijlings, Soravit Changpinyo, Radu Soricut, and Vittorio Ferrari. Connecting vision and language with localized narratives. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 647-664. Springer, 2020. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[45] Rui Qian, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Shuanggrui Ding, Dahua Lin, and Jiaqi Wang. Streaming long video understanding with large language models. Advances in Neu" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13457" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "ral Information Processing Systems, 37:119336-119360, 2025. 1, 2, 3, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "text", + "content": "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "text", + "content": "[47] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólár. Designing network design spaces. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10428-10436, 2020. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "type": "text", + "content": "[48] Jiayuan Rao, Haoning Wu, Chang Liu, Yanfeng Wang, and Weidi Xie. Matchtime: Towards automatic soccer game commentary generation. arXiv preprint arXiv:2406.18530, 2024. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "text", + "content": "[49] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 387 + ], + "type": "text", + "content": "[50] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 421 + ], + "type": "text", + "content": "[51] Larry R Squire, Lisa Genzel, John T Wixted, and Richard G Morris. Memory consolidation. Cold Spring Harbor perspectives in biology, 7(8):a021766, 2015. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 475 + ], + "type": "text", + "content": "[52] Guangzhi Sun, Wenyi Yu, Changli Tang, Xianzhao Chen, Tian Tan, Wei Li, Lu Lu, Zejun Ma, and Chao Zhang. Fine-grained audio-visual joint representations for multimodal large language models. arXiv preprint arXiv:2310.05863, 2023. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 477, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 477, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 477, + 294, + 544 + ], + "type": "text", + "content": "[53] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1207-1216, 2019. 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 544, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 294, + 589 + ], + "type": "text", + "content": "[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 590, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 612 + ], + "type": "text", + "content": "[55] Ales Vysocky and Petr Novak. Human-robot collaboration in industry. MM Science Journal, 9(2):903–906, 2016. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 613, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 658 + ], + "type": "text", + "content": "[56] Junke Wang, Dongdong Chen, Chong Luo, Xiyang Dai, Lu Yuan, Zuxuan Wu, and Yu-Gang Jiang. Chatvideo: A tracklet-centric multimodal and versatile video understanding system. arXiv preprint arXiv:2304.14407, 2023. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[57] Qing Wang, Jiaming Zhang, Kailun Yang, Kunyu Peng, and Rainer Stiefelhagen. Matchformer: Interleaving attention in transformers for feature matching. In Proceedings of the Asian Conference on Computer Vision, pages 2746-2762, 2022. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "text", + "content": "[58] Xunguang Wang, Zhenlan Ji, Pingchuan Ma, Zongjie Li, and Shuai Wang. Instructta: Instruction-tuned targeted attack for large vision-language models. arXiv preprint arXiv:2312.01886, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 163 + ], + "type": "text", + "content": "[59] Yuxuan Wang, Yueqian Wang, Pengfei Wu, Jianxin Liang, Dongyan Zhao, and Zilong Zheng. Lstp: Language-guided spatial-temporal prompt learning for long-form video-text understanding. arXiv e-prints, pages arXiv-2402, 2024. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 164, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 553, + 229 + ], + "type": "text", + "content": "[60] Shiwei Wu, Joya Chen, Kevin Qinghong Lin, Qimeng Wang, Yan Gao, Qianli Xu, Tong Xu, Yao Hu, Enhong Chen, and Mike Zheng Shou. Videollm-mod: Efficient video-language streaming with mixture-of-depths vision computation. Advances in Neural Information Processing Systems, 37:109922-109947, 2025. 1, 3, 5, 6, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 231, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 285 + ], + "type": "text", + "content": "[61] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9777–9786, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 286, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 342 + ], + "type": "text", + "content": "[62] Haomiao Xiong, Zongxin Yang, Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Jiawen Zhu, and Huchuan Lu. Streaming video understanding and multi-round interaction with memory-enhanced knowledge. arXiv preprint arXiv:2501.13468, 2025. 1, 2, 3, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 343, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 398 + ], + "type": "text", + "content": "[63] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the 25th ACM international conference on Multimedia, pages 1645-1653, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "type": "text", + "content": "[64] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 445, + 555, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 555, + 499 + ], + "type": "text", + "content": "[65] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1686-1697, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "type": "text", + "content": "[66] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. Advances in Neural Information Processing Systems, 35:124-141, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 545, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 612 + ], + "type": "text", + "content": "[67] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10714-10726, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 613, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 669 + ], + "type": "text", + "content": "[68] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9127–9134, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[69] Jeffrey M Zacks, Nicole K Speer, Khena M Swallow, Todd S Braver, and Jeremy R Reynolds. Event perception: a mind-brain perspective. Psychological bulletin, 133(2):273, 2007. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13458" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 463 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[70] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975-11986, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "text", + "content": "[71] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 5, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[72] Haoji Zhang, Yiqin Wang, Yansong Tang, Yong Liu, Jiashi Feng, Jifeng Dai, and Xiaojie Jin. Flash-vstream: Memory-based real-time understanding for long video streams. arXiv preprint arXiv:2406.08085, 2024. 1, 2, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 250 + ], + "type": "text", + "content": "[73] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 252, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 252, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 252, + 294, + 306 + ], + "type": "text", + "content": "[74] Ruohong Zhang, Liangke Gui, Zhiqing Sun, Yihao Feng, Keyang Xu, Yuanhan Zhang, Di Fu, Chunyuan Li, Alexander Hauptmann, Yonatan Bisk, et al. Direct preference optimization of video large multimodal models from language model reward. arXiv preprint arXiv:2404.01258, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 294, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 351 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 351 + ], + "type": "text", + "content": "[75] Ruoyu Zhang, Lulu Wang, Yi He, Tongling Pan, Zhengtao Yu, and Yingna Li. Tpcap: Unlocking zero-shot image captioning with trigger-augmented and multi-modal purification modules. arXiv preprint arXiv:2502.11024, 2025. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 353, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 294, + 407 + ], + "type": "text", + "content": "[76] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6586-6597, 2023. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 409, + 294, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 294, + 463 + ], + "type": "text", + "content": "[77] Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. Streaming dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18243-18252, 2024. 8" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13459" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Streaming VideoLLMs for Real-Time Procedural Video Understanding/e3ae8cdd-650d-48b8-9247-fe7363ebedec_content_list.json b/2025/Streaming VideoLLMs for Real-Time Procedural Video Understanding/e3ae8cdd-650d-48b8-9247-fe7363ebedec_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..70dea634b94ce87741f64a32acfc22a99b92526c --- /dev/null +++ b/2025/Streaming VideoLLMs for Real-Time Procedural Video Understanding/e3ae8cdd-650d-48b8-9247-fe7363ebedec_content_list.json @@ -0,0 +1,1652 @@ +[ + { + "type": "text", + "text": "Streaming VideoLLMs for Real-time Procedural Video Understanding", + "text_level": 1, + "bbox": [ + 142, + 130, + 854, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dibyadip Chatterjee $^{1,3*}$ Edoardo Remelli $^{1}$ Yale Song $^{2}$ Bugra Tekin $^{1}$ Abhay Mittal $^{1}$ Bharat Bhatnagar $^{1}$ Necati Cihan Camgoz $^{1}$ Shreyas Hampali $^{1}$ Eric Sauser $^{1}$ Shugao Ma $^{1}$ Angela Yao $^{3}$ Fadime Sener $^{1}$", + "bbox": [ + 99, + 179, + 915, + 218 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Meta Reality Labs $^{2}$ FAIR, Meta $^{3}$ National University of Singapore", + "bbox": [ + 214, + 223, + 803, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dibschat.github.io/ProVideLLM", + "bbox": [ + 377, + 244, + 637, + 258 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9109d7903435817505e944fca36d9aec8c09ad0ece6a264152e2040549ee34c6.jpg", + "image_caption": [ + "Figure 1. Overview of ProVideLLM. A streaming video large language model for real-time procedural video tasks with a low memory footprint. It features a multimodal interleaved cache that stores textual descriptions of long-term observations (spanning several minutes) and visual tokens representing short-term observations (spanning a few seconds). We also introduce a new DETR-QFormer connector for better fine-grained tokenization of the short-term. ProVideLLM is capable of handling multiple procedural tasks within a single model." + ], + "image_footnote": [], + "bbox": [ + 116, + 305, + 883, + 488 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 577, + 326, + 593 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce ProVideLLM, an end-to-end framework for real-time procedural video understanding. ProVideLLM integrates a multimodal cache configured to store two types of tokens – verbalized text tokens, which provide compressed textual summaries of long-term observations, and visual tokens, encoded with DETR-QFormer to capture fine-grained details from short-term observations. This design reduces token count by $22 \\times$ over existing methods when representing one hour of long-term observations while effectively encoding fine-granularity of the present. By interleaving these tokens in our multimodal cache, ProVideLLM achieves sublinear scaling of memory and compute with video length, ensuring per-frame streaming inference at 10 FPS and streaming dialogue at 25 FPS, with a minimal 2GB GPU memory footprint. ProVideLLM also sets new state-of-the-art results on six procedural tasks across four datasets.", + "bbox": [ + 88, + 614, + 485, + 858 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 570, + 643, + 585 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Procedural videos, spanning several minutes to hours, capture multi-step tasks [82] such as cooking [65, 101], toy assembly [60, 67], and furniture building [7, 8]. With the growing adoption of AR assistants, streaming video understanding is essential for real-time procedural assistance. This requires real-time, online inference, as frames appear in a live video, without access to future observations. Despite its importance, streaming video understanding remains largely unexplored in existing procedural research [19, 36, 100].", + "bbox": [ + 511, + 597, + 908, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing works on streaming video understanding [17, 70] focus on online action detection [53, 79, 86, 97] or live conversations [14, 83]. These use-cases target a small set of unrelated actions [17, 34] and do not require long-term temporal modeling. In contrast, procedural videos are inherently long and causal, where each step depends on prior steps. Applying existing approaches to such videos significantly increases token count and memory overhead, leading to long-context degradation [31, 46] and making them unsuitable for streaming, even with excessive frame dropping [29]. The semantic inter-connectivity of activity steps introduces re", + "bbox": [ + 511, + 734, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done during an internship at Meta.", + "bbox": [ + 114, + 886, + 336, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22586", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dundancy when long-term observations are stored as raw frame tokens [79, 86]. Furthermore, while short-term observations are crucial for understanding the present, not all frame tokens are equally relevant, since procedural tasks often center around hand-object interactions [28, 67, 72].", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose an end-to-end framework for streaming procedural video understanding (Fig. 1), addressing the above mentioned challenges of processing long-form videos. Specifically, we introduce ProVideLLM, a Procedural VideoLLM that comprises a visual (frame) encoder, a vision-language connector, and a pre-trained large language model (LLM) [1, 20, 58, 77] as the decoder. Our key contribution is a vision-language multimodal cache that efficiently manages streaming video data by balancing compression and representation quality of past observations.", + "bbox": [ + 88, + 169, + 482, + 319 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We factorize the streaming past observations into long-term semantic history and short-term visual context. They are encoded as text and visual tokens, respectively, and interleaved in our multimodal cache. First, for the long-term history spanning several minutes, ProVideLLM decoder predicts step labels per frame, groups similar steps into verbalized sequences (e.g. get $\\rightarrow$ crack $\\rightarrow$ whisk eggs), and stores them in the cache. This semantic compression significantly reduces token count and also enhances current step prediction (e.g. $\\rightarrow$ pour mixture). Second, to capture short-term visual context, typically spanning a few seconds, understanding the fine-grained changes across frames is crucial. However, standard VideoLLM encoders [59, 89] face two key challenges in such modeling: (i) temporal collapse, where low temporal variance conflates distinct actions and (ii) background distractors across frames [5, 13, 61, 90]. We mitigate the former by using visual encoders trained with dense spatial alignment objectives [11, 30, 52]. We address the latter with a novel vision-text connector DETR-QFormer, pretrained to explicitly focus on hand-object interaction regions, thereby compressing them into a smaller set of tokens.", + "bbox": [ + 91, + 321, + 485, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, in existing methods [86], both long- and short-term tokens are visual and stored separately. Applying such separate caching to our multimodal tokens would require redundant attention recomputation when verbalizing short-term visual tokens into long-term text. To overcome this, we interleave text and visual tokens within our multimodal cache. This enables prefilling the cache, which ensures that computation scales linearly with cache size and sub-linearly with the video length, thereby significantly reducing latency for real-time performance.", + "bbox": [ + 89, + 642, + 482, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are: (i) a first-of-its-kind streaming multimodal cache with interleaved text and visual tokens, enabling per-frame inference [72] at 10 FPS and streaming dialogue [14] at 25 FPS; (ii) verbalizing long-term observations into compressed textual summaries and (iii) modeling hand-object interactions for short-term observations using DETR-QFormer; all combined, it achieves a", + "bbox": [ + 89, + 795, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "$22 \\times$ reduction in token count compared to existing methods. (iv) state-of-the-art results on six tasks across four datasets, making ProVideLLM the first real-time framework unifying recognition, anticipation, planning for procedural guidance.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 157, + 663, + 172 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Procedural Video Understanding have primarily focused on instructional YouTube videos [48, 74, 101, 103]. Recent large-scale egocentric datasets [27, 28, 67, 72] have advanced research by providing longer videos with more activity steps. Existing works have focused on offline tasks [19, 43, 100], where the entire video is available for prediction, as well as some online tasks, like procedural planning [12, 33, 95], forecasting [23, 26, 49, 65, 66, 96], and mistake detection [18, 22, 37, 67]. However, deploying these models for streaming inference results in high computational complexity where it grows quadratically with the video length. In contrast, ProVideLLM incurs a sub-linear cost per update, both in terms of memory and runtime, enabling efficient streaming inference.", + "bbox": [ + 511, + 181, + 906, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Large Language Models. Multimodal large language models (MLLMs) have been popularized by BLIP-2-like [15, 39, 102] and LLaVA-like [38, 44, 45] models where an LLM is trained to understand vision-text data. Extending these to video have led to the development of Video Large Language Models (VideoLLMs) [40, 41, 47, 83, 94, 98]. However, most of these models are limited to offline tasks [62, 71, 73, 80, 81] where the entire video is queried to answer a question. As such, architectural elements such as VideoQ-Former [92] and temporal pooling layers [40], developed without causal requirements, limit streaming applications. Models like MA-LMM [29] and VideoLLM online [14] address video understanding online, but are too computationally intensive to be deployed for efficient streaming inference on long-form videos. ProVideLLM advances these models by introducing a multimodal interleaved cache for faster inference and proposing DETR-QFormer for modeling fine-grained visual changes, improving accuracy.", + "bbox": [ + 511, + 393, + 908, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Streaming Video Understanding has been explored for online action detection [17, 25, 70] where dominant approaches [53, 79, 86, 97] split streaming input into long- and short-term visual histories within an encoder-decoder architecture. However, storing both as visual tokens is redundant and computationally prohibitive for long-form procedural videos, where multiple actions are temporally connected. Instead, ProVideLLM verbalizes long-term observations into text which drastically decreases token count required to represent hour long videos, thus improving memory efficiency.", + "bbox": [ + 511, + 666, + 908, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. ProVideLLM", + "text_level": 1, + "bbox": [ + 513, + 829, + 651, + 844 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ProVideLLM encodes long- and short-term streaming observations into text and visual tokens and interleaves them in a multimodal cache. We first present the preliminaries", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22587", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "in Sec. 3.1, followed by long- and short-term token generation in Sec. 3.2 and Sec. 3.3. Finally, we present our multimodal cache in Sec. 3.4 and training objectives in Sec. 3.5.", + "bbox": [ + 89, + 90, + 485, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 142, + 230, + 157 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ProVideLLM follows a LLaVA-style [44] architecture, consisting of a visual encoder $(E_V)$ , a language decoder $(D_L)$ and a vision-language connector $(C_{VL})$ . Given a video $V = \\{v_1, v_2, \\ldots, v_T\\}$ , the model's output $(o_t)$ at any time step $t \\in [1, T]$ is", + "bbox": [ + 89, + 162, + 483, + 238 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\no _ {t} = \\operatorname {P r o V i d e L L M} \\left(v _ {1: t}\\right) = D _ {L} \\left(\\left\\| _ {i = 1} ^ {t} C _ {V L} \\left(E _ {V} \\left(v _ {i}\\right)\\right)\\right) \\quad (1) \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 243, + 483, + 267 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\parallel$ denotes concatenation. While LSTR-inspired long short-term memory approaches [79, 97] also use an encoder-decoder architecture, they rely exclusively on visual tokens (Fig. 2a.1). In contrast, ProVideLLM incorporates a language decoder, which enables more efficient long-term compression, improves generalization, and supports multiple tasks within a single model.", + "bbox": [ + 89, + 271, + 483, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Tokenizing long-term observations", + "text_level": 1, + "bbox": [ + 89, + 383, + 392, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Procedural videos are inherently long-form, ranging from tens of minutes [67] to several hours [28, 72]. While modern LLMs [20, 75, 87] offer massive context lengths and, in theory, can process an hour-long raw video at 24 fps, performance degrades with larger context sizes [31, 46], and training such models is highly memory-intensive [29, 93]. Accurately recognizing ongoing steps require effectively encoding long-term context. Our key insight is that procedural steps in a video are temporally connected, i.e. a sequence of step labels can serve as a strong predictor of ongoing steps. For instance, when \"preparing an omelet\", the step \"pour mixture into pan\" depends on earlier steps \"get eggs from fridge, crack eggs into bowl, whisk eggs\". Prior works on flow graphs [21, 64] and planning [3, 54, 100] also show that step sequences, often extracted from external sources [35], serve as strong baselines for procedural understanding.", + "bbox": [ + 88, + 402, + 483, + 645 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instead of constructing computationally expensive graphs, we propose online verbalization (Fig. 2a.2). Given a streaming video, ProVideLLM decoder continuously generates action predictions (summaries) of the present observation on the fly, and groups semantically similar actions to verbalize the past (Fig. 1). This approach efficiently compresses past visual frame tokens into information-dense language tokens. For example, in Ego4D Goal-Step [72], storing one hour of long-term past requires only 630 verbalized text tokens on average. Compared to long short-term approaches [79, 86, 97], our verbalization reduces token count by $22 \\times$ at the same sampling rate, significantly reducing the memory overhead.", + "bbox": [ + 89, + 645, + 485, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Tokenizing short-term observations", + "text_level": 1, + "bbox": [ + 89, + 833, + 398, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to recognize procedural activity steps in a streaming manner. This requires understanding \"what is happening\" in the present while tracking \"what has happened in", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/549ec69cc1878c662a86f32b066748037c4486fa185624f420f11daa967e6b21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 89, + 898, + 247 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0a552e73e5bf57d1c56da3a067243342e80adc27a808b023a7c3c5b2f0132174.jpg", + "image_caption": [ + "Figure 2. Converting short-term tokens to long-term under various caching types. (a) Factorizing the streaming observations into long-term and short-term and ordering them progressively. (a.1) Conversion done at $O(1)$ cost but is unable to represent long-form videos due to massive token count and memory needs. (a.2) Online verbalization can represent long-form videos at a manageable memory but is not suitable for streaming due to $O(N_S^2 + N_S N_L)$ conversion cost. (b) Interleaving them reduces the conversion cost to $O(N)$ , allowing streaming inference on long-form videos." + ], + "image_footnote": [], + "bbox": [ + 596, + 258, + 905, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the past\". Language is a strong indicator of the semantic structure of step sequences, but determining if and when the step is performed requires fine-grained visual information. Therefore, we allocate more compute for understanding the short-term past, typically spanning 8-16 seconds, by keeping its visual tokens uncompressed. Standard VideoLLM encoders [38, 44] face two key challenges in capturing fine-grained short-term observations: $(i)$ they generate temporally collapsed video features, a limitation we identify and address in Sec. 3.3.1, and $(ii)$ they get biased towards high-level scene and background distractors, which we address in Sec. 3.3.2.", + "bbox": [ + 511, + 507, + 908, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3.1. Choosing the visual encoder", + "text_level": 1, + "bbox": [ + 511, + 684, + 756, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The predominant visual encoders for MLLMs are CLIP [59] and SigLIP [2, 89]. These encoders are pre-trained with only a language-alignment objective for integration with LLMs. However, because they are trained on high-level and sparse web-scale captions such as \"cooking in the kitchen\", they struggle with fine-grained reasoning, particularly in answering the \"what?\" and \"how to?\" aspects of a video [84, 85]. Consequently, distinct images may yield high similarity scores in their feature space [76], leading to low feature diversity across frames within a step. This results in relatively static frame features, a phenomenon we term temporal collapse. For certain fine-grained procedural steps, such as \"pick up\" vs. \"put down\", temporally collapsed", + "bbox": [ + 511, + 703, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22588", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/04a5d58bb747d1f827032af5c311f84394c786979c079a378006d3e83aee0eea.jpg", + "image_caption": [ + "a)", + "b)", + "Figure 3. a) Per-class mean temporal variance on EgoExo4D & Ego4D Goal-Step. Highly temporally varying classes on the left and low-variance classes on the right. b) Visualization of the top-3 PCA components of patch tokens computed across video frames, thresholded by first component. The primary factors of variation for DINOv2 [52] are hands, and are unfocused and scattered across the image for language-aligned encoders, CLIP [59] and SigLIP [89]." + ], + "image_footnote": [], + "bbox": [ + 119, + 89, + 473, + 284 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "features make it nearly impossible for the downstream LLM to decode the correct step.", + "bbox": [ + 89, + 412, + 482, + 443 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 3a shows the impact of temporal collapse across three visual encoders – two language-aligned ones [59, 89] and one self-supervised1 [52]. We plot temporal variance per class in descending order, with high-variance classes on the left and low-variance classes on the right (see Suppl.Sec.1 for temporal variance computation). At both extremes, the trends are consistent across all encoders. High-variance classes, which often involve verbs such as “get”, “fetch” or “put” correspond to longer videos with substantial scene and environmental changes. Conversely, low-variance classes involving adverbs such as “slowly” or “boil” or repetitive actions like “rotate” or “peel”, exhibit gradual state changes. Among the three models, DINOv2 consistently captures temporal variations more effectively, covering a broader range of temporal variance across different classes. Hence, we choose DINOv2 as our visual encoder.", + "bbox": [ + 88, + 443, + 483, + 684 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2. Aligning hand-object transformations with text", + "text_level": 1, + "bbox": [ + 89, + 693, + 465, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Procedural datasets predominantly capture hand-centric actions. For instance, nearly $99\\%$ of actions in [28, 72] involve hand interactions2. Moreover, video frames often contain distracting background objects, while hand locations serve as strong cues for recognizing the ongoing actions [68, 91]. Thus, a vision encoder focused on hand regions can enhance performance by mitigating background distractions and capturing fine-grained transformations more precisely.", + "bbox": [ + 89, + 712, + 483, + 833 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Another contribution of our work is a vision-language connector, DETR-QFormer, pretrained to attend to salient", + "bbox": [ + 89, + 834, + 482, + 864 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/08271fe2ace3d397040a590658c1cce39a5bd4e7e03de589699a49401f6a6c19.jpg", + "image_caption": [ + "Figure 4. DETR-QFormer architecture & Stage-1 pre-training. Trained to detect hands and objects-in-contact from pseudo ground truths bboxes generated by [69], DETR-QFormer queries DINOv2 patch tokens to extract hand-object-focused activations." + ], + "image_footnote": [], + "bbox": [ + 517, + 97, + 906, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "objects. In procedural videos, these are typically in contact with the hand. Interestingly, our choice of DINOv2 as the visual encoder provides an advantage as its patch tokens capture hands as primary factors of variation Fig. 3b. Leveraging this, we design DETR-QFormer to query DINOv2 patch tokens, extracting hand-object activations and suppressing background distractions. DETR-QFormer explicitly attends to hand-object activations using a DETR loss [10] and compresses salient regions into fewer tokens via QFormer's [39] cross-attention, enhancing representation quality.", + "bbox": [ + 511, + 393, + 908, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Architecture. Our connector integrates Q-Former's compression [39] and DETR's bbox loss into a single crossattention block. DETR-QFormer, illustrated in Fig. 4, is pretrained to detect the hands and objects-in-contact (Sec. 3.5). The connector $(C_{VL})$ is a cross-attention transformer decoder with a set of $m$ learnable visual queries $Q_{v} = (q_{v_{1}}, q_{v_{2}}, \\ldots, q_{v_{m}})$ . Key-value pairs $(K, V)$ are generated from the flattened $n^{2}$ DINOv2 patch tokens using two projection heads $(W_{K}, W_{V})$ respectively. To explicitly model hand-object interactions, we append two sets of special queries - two hand queries $Q_{h} = (q_{h_{1}}, q_{h_{2}})$ for the left and right hands, and $k$ object queries $Q_{o} = (q_{o_{1}}, q_{o_{2}}, \\ldots, q_{o_{k}})$ corresponding to objects-in-contact. The role of visual queries $Q_{v}$ is to compress the $n^{2} \\times D$ patch tokens into $m \\times d$ tokens, where $m << n^{2}$ and $d << D$ . Whereas, outputs corresponding to the learnable hand-object queries $(Q_{h}^{\\prime}, Q_{o}^{\\prime})$ are passed through an MLP to make bounding box predictions, $\\hat{B} = \\{\\hat{b}_{h_{1}}, \\hat{b}_{h_{2}}, \\hat{b}_{o_{1}}, \\ldots, \\hat{b}_{o_{k}}\\}$ .", + "bbox": [ + 511, + 546, + 908, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Multimodal Interleaved Cache", + "text_level": 1, + "bbox": [ + 511, + 832, + 784, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The effectiveness of a procedural video model is best assessed when deployed for real-time inference. Naive long-and short-term approaches [79, 86] maintain separate caches", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1We chose DINOV2 due to its comparable scale to CLIP and SigLIP. \n2Remaining 1% are visually undetectable actions \"call for help, waiting.\"", + "bbox": [ + 107, + 875, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22589", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Multimodal Interleaved Cache" + ], + "code_body": "1: class MultimodalInterleavedCache: \n2: def __init__(self, N.S, N.L): \n3: self_tokens = [] \n4: self.N.S = N.S \n5: self.N.L = N.L \n6: # entry function: adds a new token to cache \n7: def entry(self, token): \n8: self_tokens.append(token) \n9: # exit function for short-term tokens \n10: def exit_short(self): \n11: if self_tokens.count() > self.N.S: \n12: self_tokens.pop(self_tokens.index()) \n13: return True \n14: # exit function for long-term tokens \n15: def exit_long(self): \n16: if self_tokens.count() > self.N.L: \n17: self_tokens.pop(self_tokens.index()", + "bbox": [ + 91, + 108, + 483, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for the two token types, arranged progressively with a single entry point (see Fig. 2a). With multimodal tokens, such setups require two distinct entry points, forcing attention recomputation over the short-term cache whenever a short-term token is verbalized into a long-term one. In the worst case, this leads to quadratic scaling of compute. To efficiently leverage online verbalization, we propose a multimodal interleaved cache that interleaves long-term text and short-term visual tokens from Sec. 3.2 and Sec. 3.3.", + "bbox": [ + 88, + 319, + 483, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Interleaving long- and short-term tokens. We store the running short-term visual tokens and long-term text tokens in a single multimodal cache operating as a FIFO (First In, First Out) queue. It has one entry point and two distinct exits for short- and long-term tokens. Our entry and exit functions are detailed in Algorithm 1, where $N_{S}$ denotes the short-term cache length (in number of frames) and $N_{L}$ the long-term cache length (in number of observed steps). As frame tokens $(v_{i})$ exit the short-term span, the ProVideLLM decoder verbalizes the oldest token $(v_{-N_s})$ . The verbalized text is then added to the cache unless already cached within the past $\\tau$ frames. To distinguish the two token types within the same cache, we prepend a special token (represented as $<\\mathbb{L}>$ ) to the long-term text tokens. ProVideLLM is instruction-tuned to recognize tokens associated with $<\\mathbb{L}>$ as belonging to the long-term observations (Sec. 3.5). Since both long- and short-term tokens share a single entry point, visual and textual tokens are interleaved in the cache at any given time.", + "bbox": [ + 89, + 458, + 483, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Streaming inference. Every time a new token is added to the cache (Algorithm 2), instead of re-computing attention across all tokens, we use the causal decoder's K-V cache [55] to compute attention only for the new token over the previous $N - 1$ tokens. This ensures that the computational complexity for processing a new frame scales linearly with cache size $(O(N))$ . Additionally, maintaining a single interleaved cache enables direct application of optimized attention implementations [6, 81], though such techniques remain orthogonal to our approach. More importantly, for a fixed $N_S$ , compression via verbalization ensures that $N_L$", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 2 Streaming Inference" + ], + "code_body": "1: def inference(V, N_S, N_L, $\\tau$ ): \n2: cache = MultimodalInterleavedCache(N_S, N_L, $\\tau$ ) \n3: out = [] \n4: for $v_i$ in V: \n5: cache_entry $(v_i)$ # add frame tokens to cache \n6: $t_i = \\text{ProVideLLM}(\\text{cache.tokens})$ # current prediction \n7: out.append $(t_i)$ # buffer predictions \n8: flag = cache_exit_short() # short-term overflow \n9: if flag and out[-N_S] not in out[- $\\tau$ ): \n10: cache_entry(out[-N_S]) # verbalize & interleave \n11: cache_exit_long() # long-term overflow", + "bbox": [ + 516, + 108, + 903, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and consequently $N = (N_{L} + N_{S})$ scales sub-linearly with the number of video frames. Hence, compute also scales sublinearly with the number of streaming video frames. This makes our model well-suited for low-latency applications.", + "bbox": [ + 511, + 252, + 906, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Training ProVideLLM", + "text_level": 1, + "bbox": [ + 511, + 321, + 725, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We adopt a two-stage training approach similar to LLaVA - (1) alignment pre-training to align the encoder's visual tokens to the decoder's language space, and (2) instruction tuning of the model for specific tasks.", + "bbox": [ + 511, + 344, + 906, + 405 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stage-1: Alignment Pre-training. In this stage, our connector DETR-QFormer $(C_{VL})$ is trained while keeping the visual encoder and language decoder frozen (Fig. 4). The objective is to caption short video clips featuring hand-centric actions using the standard language modeling loss [57]:", + "bbox": [ + 511, + 405, + 908, + 481 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {L M}} = - \\frac {1}{| W |} \\sum_ {i = 1} ^ {| W |} \\log P (w _ {i} \\mid w _ {< i}, V; C _ {V L} ^ {\\Theta}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 503, + 906, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $V$ is the input video clip, $W = \\{w_{1}, w_{2}, \\dots, w_{|C|}\\}$ is the ground-truth caption, and $C_{VL}^{\\Theta}$ is our DETR-QFormer parameterized by $\\Theta$ . For detecting hands and corresponding objects-in-contact, we use a DETR [10]-style Generalized IoU [63] loss:", + "bbox": [ + 511, + 559, + 906, + 635 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {H O}} \\left(b _ {i}, \\hat {b} _ {\\sigma (i)}\\right) = \\mathcal {L} _ {\\mathrm {i o u}} \\left(b _ {i}, \\hat {b} _ {\\sigma (i)}\\right) + \\left\\| b _ {i} - \\hat {b} _ {\\sigma (i)} \\right\\| _ {1} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 645, + 906, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $b_{i}$ 's are the pseudo ground-truth bounding boxes of hands and objects-in-contact generated by an off-the-shelf detector [69] and $\\hat{b}_{\\sigma(i)} \\in \\hat{B}$ is the bounding box matched from the predicted set $\\hat{B}$ . For simplicity, we keep our loss category-agnostic. The final loss for Stage-1 is thus:", + "bbox": [ + 511, + 675, + 908, + 753 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {L M}} + \\lambda_ {1} \\mathcal {L} _ {\\mathrm {H O}} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 766, + 906, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ is a hyper-parameter. In addition to vision-language alignment, Stage-1 also enables our model to attend to hand-object interactions.", + "bbox": [ + 511, + 794, + 906, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stage-2: Instruction Tuning. This stage focuses on fine-tuning our model with task-specific instructions such as \"Provide the goal.\", \"Provide what I am doing now.\", \"Provide what I need to do next.\", or \"Provide the", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22590", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/756e247997a4ec65a49a922d93ae3303dffa8cb6a015736a0156aed65e7b652b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Modelper-frame mAP
ValTest
LSTR [86] (only short-term)8.8-
LSTR [86]8.98.1
EgoOnly [78]10.310.9
ProVideLLM-1B/5 (no verbalization)12.112.2
ProVideLLM-1B/5 (verbalization & interleaving)13.012.9
", + "bbox": [ + 99, + 88, + 475, + 179 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/529a713b1c92f4e136c8b3b0559b4d45a8a17912991a932bb243e461548128fd.jpg", + "table_caption": [ + "Table 1. Ego4D [72] GoalStep Online Step Detection. For fair comparison with LSTR, we consider 16s of short-term past and an avg. of 128s of long-term past compressed into language tokens." + ], + "table_footnote": [], + "table_body": "
ModelViewsAction
TempAgg [66]12 views8.5
+ Goal [88]12 views12.0
ProVideLLM-8B/11 (+long-term verbalization)v413.8
", + "bbox": [ + 99, + 246, + 473, + 309 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "next 10 steps I need to do.\" For instruction tuning, we fine-tune DETR-QFormer and the language decoder using $\\mathcal{L}_{\\mathrm{LM}}$ while keeping the visual encoder frozen. For streaming tasks, we also train our model with interleaved special $<\\mathbb{L}>$ tokens to represent long-term cache entries, distinguishing them from standard prompts. See Suppl.Sec.5 for further details.", + "bbox": [ + 89, + 375, + 483, + 467 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 481, + 223, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct extensive evaluations of ProVideLLM on six tasks across four datasets covering streaming (Sec. 4.1.1), long-term (Sec. 4.1.2) and fine-grained (Sec. 4.1.3) procedural understanding. We further assess ProVideLLM's performance in multi-task settings and its generalization to new datasets in Sec. 4.2. Finally, we provide ablations in Sec. 4.3 and ProVideLLM's runtime analysis in Sec. 4.4.", + "bbox": [ + 89, + 507, + 483, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. Depending on the choice of vision encoder and language decoder, we have two model variants. ProVideLLM-1B/5 is our efficiency-focused model designed for real-time streaming. It uses DINOv2 (ViT-S) with Llama-3.2-1B-Instruct [20] and requires 5 tokens per frame. ProVideLLM-8B/11 is our more performance-focused model using DINOv2 (ViT-g) and Llama-3.1-8B-Instruct and requires 11 tokens per frame. Furthermore, for benchmarking on youtube videos [74], we introduce an additional model ProVideLLM-8B/11+, which uses SigLIP's [CLS] token along with DINOv2 patch tokens. Unless otherwise mentioned we use $N_{S} = 64$ , $N_{L} = 5$ i.e. storing a maximum of 64 frames in the short-term cache and a maximum of 5 observed steps in the long-term cache. DETR-QFormer is pre-trained (Stage-1) on EPIC-KITCHENS [16] with pseudo ground-truth bounding boxes from 100DOH [69]. For Stage-2, the LLM (decoder) is fine-tuned using LoRA [32] with $r = 128$ and $\\alpha = 256$ . Additional implementation details are provided in Suppl.Sec.5.", + "bbox": [ + 89, + 613, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2168c69feb0267a6780bc07419107c6fd69c91e0040864930d1783fe1cbeaaa0.jpg", + "table_caption": [ + "Table 2. Action Anticipation Results on Assembly101 validation set. Performance measured by class-mean Top-5 recall." + ], + "table_footnote": [], + "table_body": "
MethodTop-1 Acc. (%)
TimeSformer [9]34.0
DistantSup39.4
VideoTF42.4
ProcedureVRL46.8
VEDIT [42]51.8
VideoLLM-online [14]49.1
VideoLLM-MoD [83]49.7
ProVideoLLM-8B/11+52.1
ProVideoLLM-8B/11+ (+long-term verbalization)53.6
", + "bbox": [ + 524, + 88, + 895, + 229 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Next Step Forecasting in COIN. The last four rows including ProVideLLM are VideoLLMs.", + "bbox": [ + 511, + 239, + 903, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate our model's performance on four large-scale procedural datasets - EgoExo4D [28], Ego4D-GoalStep [72], COIN [74], and Assembly101 [67]. As shown in Sec. 4.1, ProVideLLM sets state-of-the-art on both egocentric and YouTube benchmarks, demonstrating robustness across diverse video domains.", + "bbox": [ + 511, + 295, + 906, + 386 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Main results", + "text_level": 1, + "bbox": [ + 511, + 395, + 648, + 410 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.1. Online Step Detection", + "text_level": 1, + "bbox": [ + 511, + 417, + 712, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We report results in Tab. 1, using per-frame mAP on action classes (steps + substeps) at 4fps, a standard metric in online action detection [79, 97]. We evaluate ProVideLLM-1B/5 with and without long-term verbalization in rows four and five, respectively. Results for LSTR [86] and EgoOnly [78] are taken from [72]. Accordingly, LSTR spans $128s$ in long-term memory and $16s$ in short-term memory. In [72], each step spans on average $32s$ with around 5.7 tokens representing each step. Our long-term cache can encode $128s$ of past using only 22 tokens at $N_{L} = 4$ steps. To measure improvements from our long-term design, we set the cache length $N = 86$ (spanning LSTR's temporal context), split into $(N_{S} = 86, N_{L} = 0)$ and $(N_{S} = 64, N_{L} = 4)$ in rows four and five, respectively. Our model surpasses both baselines, outperforming LSTR by over $4\\%$ at similar temporal span. ProVideLLM also achieves greater gains from long-term context compared to LSTR, highlighting the effectiveness of verbalizing the long-term past.", + "bbox": [ + 511, + 436, + 908, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.2. Online Step Forecasting", + "text_level": 1, + "bbox": [ + 511, + 715, + 728, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Step forecasting requires modeling long-term dependencies, making it an ideal task to evaluate our contributions. We evaluate ProVideoLLM for step forecasting on COIN and Assembly101 measured by Top-1 accuracy and Top-5 Recall [24], respectively. Our model achieves state-of-the-art performance on both datasets as shown in Tab. 3 and Tab. 2. On Assembly101, we surpass [88] by $1.8\\%$ using only a single view. On COIN, we use Llama-3-8B-Instruct to ensure fair comparisons with VideoLLMs [14, 83] and achieve a performance boost of around $4\\%$ . Moreover, while state-of-the-art models [14, 83] use approximately 600 to 6000 visual", + "bbox": [ + 511, + 734, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22591", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/685a07c5e829e6d6941fd443b32352218831a5414cfa2d31acb37091b1a178b1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelTraining ViewsEgo Acc. (%) Val / Test
TimeSFformer [9]ego35.13 / 35.93
EgoVLPv2 [56]ego+exo39.10 / 38.76
Viewpoint Distillationego+exo38.19 / 39.49
View Invariant Encoder [51]ego+exo40.34 / 41.53
VideoLLM-MoD† [83]ego42.62 / —
ProVideLLM-8B/11ego44.36 / 50.74
", + "bbox": [ + 106, + 88, + 470, + 213 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/842ca40dc5205173594da70fb3ad48eb65bf7316e65b7cbc9ad6f159ac273e1a.jpg", + "table_caption": [ + "Table 4. EgoExo4D Fine-grained Keystep Recognition. Test results obtained from EgoExo4D's Fine-grained Keystep Recognition Challenge on EvalAI. † Uses 2fps sampling per segment." + ], + "table_footnote": [], + "table_body": "
MethodTop-1 Accuracy (%)
TimeSformer [9]34.0
Paprika [100]51.0
DistantSup [43]54.1
ProcedureVRL [99]56.9
VideoTaskGraph [4]57.2
VEDIT [42]62.7
VideoLLM-online [14]63.1
VideoLLM-MoD [83]63.4
ProVideLLM-8B/11+67.3
", + "bbox": [ + 104, + 280, + 470, + 419 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tokens to model past context, our approach spans a larger temporal context with only 200 tokens through verbalization.", + "bbox": [ + 89, + 460, + 483, + 491 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1.3. Step Recognition", + "text_level": 1, + "bbox": [ + 89, + 500, + 256, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Tab. 4 and Tab. 5, we evaluate ProVideLLM on fine-grained step recognition on EgoExo4D and COIN. We uniformly sample 16 frames from each segment to construct our short-term cache $(N_S)$ and use the prediction from the last frame as our output. The closest model, VideoLLM-MoD [83], uses 10 tokens per frame, but densely samples frames at 2fps, resulting in a larger token count. In both cases, we outperform VideoLLM-MoD by $1.74\\%$ and $3.9\\%$ respectively. Moreover, at the time of our submission, our method, using only egocentric video, achieved the best results on the EgoExo4D Fine-grained Recognition Challenge, with a $9.2\\%$ improvement over the second-ranked model.", + "bbox": [ + 88, + 518, + 485, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Generalized Procedural Understanding", + "text_level": 1, + "bbox": [ + 89, + 710, + 429, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we explore additional capabilities of ProVideLLM, leveraging the advantages of using a pretrained LLM as the language decoder. This section provides insights into developing a \"generalist\" model that advance procedural video understanding within a unified framework. We first evaluate ProVideLLM's ability to multi-task and then present our benchmark for cross dataset generalization. Multi-task learning. In Tab. 6, we present our results for multi-task learning, an underexplored aspect of procedural understanding. Existing models are trained only for isolated tasks, such as step recognition and forecasting. We compare", + "bbox": [ + 89, + 734, + 485, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fde3747cd30d1adc6607b4a3ebc01776d38be6d4a36841888e357254ab848971.jpg", + "table_caption": [ + "Table 5. COIN Step Recognition. Last three rows are VideoLLMs." + ], + "table_footnote": [], + "table_body": "
ModelCOINBenchmark Top-1 Accuracy (%)
StepTaskNextProc.Proc.+
VideoTF† [50]56.591.042.440.246.4
VideoLLM-online [14]63.192.749.149.854.1
VideoLLM-MoD [83]63.492.849.749.853.3
ProVideLLM-8B/11+66.995.050.551.055.9
", + "bbox": [ + 527, + 88, + 893, + 176 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ca672561f95c906bbbc940c4839e43453601a5f125e358d20c200dff660f072a.jpg", + "table_caption": [ + "Table 6. Multi-task learning on COIN. Step - step recognition, Task - task recognition, Next - next step forecasting, Proc. - long-term forecasting and $\\mathrm{Proc + }$ - long-term forecasting with task. $^{\\dagger}$ VideoTF [50] is trained separately for each task." + ], + "table_footnote": [], + "table_body": "
ModelEgo Acc. (%)
VideoLLaVA [41] + Llama-3.1-8B [20]3.6
LLaVA-OneVision [38] + Llama-3.1-8B [20]8.3
ProVideLLM-8B/1112.4
", + "bbox": [ + 527, + 255, + 893, + 321 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. Cross-dataset generalization. Evaluated on EgoExo4D KeyStep recognition. Our model, trained on Ego4D Goal-Step is compared with a mixture of vision + language experts.", + "bbox": [ + 511, + 332, + 906, + 375 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "our model with recent VideoLLMs, also trained in a multitask setting, on 5 tasks. Despite using the same LLM as [14, 83], our model outperforms these baselines across all tasks. Cross-Dataset Generalization. In Tab. 7, we present a cross-dataset generalization benchmark for procedural videos using the EgoExo4D fine-grained keystep recognition task. We compare two recent VideoLLMs [38, 41] which were primarily trained for video captioning and use Llama-3.1-8B-Instruct to convert the captions into EgoExo4D classes. Our baseline models act as a combination of visual and language experts. ProVideLLM, using the same LLM and trained on the much smaller Ego4D Goal-Step dataset, significantly outperforms these baselines. Despite these advances, performance remains modest, highlighting the need for further research in procedural generalization. Additional details are provided in the Suppl. Material.", + "bbox": [ + 511, + 401, + 908, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablations", + "text_level": 1, + "bbox": [ + 511, + 652, + 625, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visual Encoders & DETR-QFormer. In Tab. 8, we ablate visual encoders-SigLIP and DINOv2-and connectorsMLP, QFormer, and our DETR-QFormer-on EgoExo4D keystep recognition. First, comparing encoders across connectors, SigLIP's CLS tokens, being language-aligned, outperform DINOv2; however, with more patch tokens, DINOv2 surpasses SigLIP, aligning with our observations in Fig. 3. Second, comparing connectors across encoders, Q-Former boosts DINOv2's performance $(+1.2\\%, +1.9\\%)$ but has minimal impact on SigLIP $(+0.8\\%, -0.1\\%)$ , highlighting DINOv2's need for a stronger connector. Finally, DETR-QFormer further improves DINOv2 $(+1.1\\%, +2.6\\%)$ , with gains increasing at higher resolutions but degrading SigLIP's performance due to weak hand-object activations in its patch tokens. Notably, DETR-QFormer remains highly", + "bbox": [ + 511, + 672, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22592", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6e6e34a8305e7b2c824986f4238cab6e903f0518efd064c15a0273bda013fadf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
[CLS] TokenPatch Tokens#Tokens /frameEgo Val./ Acc. (%)
SigLIPDINOv2
w/ 2-layer MLP [45]
✓X131.8±0.428.7±0.2
✓16×16 → 2×2532.1±0.632.4±0.5
✓16×16 → 4×41733.3±0.636.6±0.5
w/ QFormer [39]
✓16×16 ⇒ 4532.9±0.633.6±0.5
✓16×16 ⇒ 161733.2±0.438.1±0.7
w/ DETR-QFormer (2 hands, 2 objects)
✓16×16 ⇒ 0530.9±0.334.7±0.5
✓16×16 ⇒ 121733.6±0.340.7±0.6
", + "bbox": [ + 101, + 88, + 475, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "token-efficient with DINOv2, even when compressing to only hand-object queries, reinforcing its advantage for low-latency streaming applications.", + "bbox": [ + 89, + 363, + 485, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Verbalization & Interleaving. In Fig. 5, we analyze our verbalization and interleaving strategies using ProVideLLM-1B/5 on Ego4D Goal-Step validation set, selecting a video containing over 15 steps and spanning $>30$ minutes. At 4fps sampling, we ablate on a fixed memory budget of 8GB on an A6000 GPU. Short-term cache length is set as $N_{S} = 64$ (16 seconds), and no limit is imposed on the long-term cache. Using only visual tokens (red), memory scales linearly with video length, reaching 40K tokens to represent 20 minutes of video before exceeding 8GB memory. Such massive context also degrades performance. Verbalization (orange) compresses 20 minutes of video into 2.5K tokens, reducing memory usage to 3GB and achieving sub-linear scaling. However, it introduces runtime instability, with frequent spikes in per-frame prediction time due to attention recomputation when converting short-term visual tokens into long-term text. Interleaving (green) in our multimodal cache mitigates these spikes, improving runtime stability while maintaining sub-linear scaling in both memory and compute.", + "bbox": [ + 89, + 410, + 485, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Runtime Analysis", + "text_level": 1, + "bbox": [ + 89, + 710, + 264, + 727 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Tab. 9, we report the per-frame streaming inference and streaming dialogue [14] performance of ProVideLLM variants on a single A6000 GPU. Our lightest variant, ProVideLLM-1B/5, achieves over 10 FPS with a minimal 2GB memory footprint, making it ideal for low-latency applications. Meanwhile, ProVideLLM-8B/11 offer higher performance but require more compute. For most videos in this dataset, ProVideLLM-1B/5 consistently maintains over 10 FPS, sometimes even reaching 20 FPS when the video has many no-action frames. For per-frame predictions, a key bottleneck for VideoLLMs is their autoregressive text gener", + "bbox": [ + 89, + 734, + 485, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/df08352eb241b1f2ad870d76d27acff468840a35df3fb639ed4836948392ae95.jpg", + "table_caption": [ + "Table 8. DETR-QFormer ablations on EgoExo4D fine-grained keysteps. Llama-3.2-1B-Instruct is used as the language decoder. Results reported over 3 runs. $(\\rightarrow$ means avg-pooling, $\\Rightarrow$ means Q-Former compression.)" + ], + "table_footnote": [], + "table_body": "
Streaming PredictionsModelFPS (↑)Memory (GB) (↓)
Vision Enc.Connector + LLMFull Model
Per-frameProVideLLM-1B/574.810.49.12.0
ProVideLLM-8B/1138.04.23.516.2
Dialogue [14]ProVideLLM-1B/574.834.224.62.2
ProVideLLM-8B/1138.023.717.216.9
", + "bbox": [ + 517, + 88, + 903, + 183 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 9. Runtime Analysis on a single A6000 GPU for the validation set of Ego4D-GoalStep online step detection for per-frame streaming and Ego4D-GoalStep Narration [83].", + "bbox": [ + 511, + 193, + 906, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e84445eb7c0fe0969400e2c46db1304bd111e84bd51298edacaa12083b12bf59.jpg", + "image_caption": [ + "Figure 5. Scaling of memory, runtime and context length under various caching strategies. Results reported on a Ego4D-GoalStep video at a memory budget of 8GB on a single A6000 GPU." + ], + "image_footnote": [], + "bbox": [ + 514, + 247, + 903, + 338 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ation, adding around 6 extra iterations, corresponding to 5.7 tokens/(activity step) on average for Ego4D-GoalStep. This issue is mitigated with streaming dialogue [14], where the model is trained to remain silent and respond to user queries only when necessary. In this setting, ProVideLLM-1B/5 delivers real-time performance at 24.6 FPS.", + "bbox": [ + 511, + 393, + 906, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Limitations", + "text_level": 1, + "bbox": [ + 511, + 497, + 769, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we presented ProVideLLM for real-time understanding of procedural videos by interleaving two types of tokens in a vision-text multimodal cache - verbalized text tokens, which provide compressed summaries of long-term observations, and visual tokens, which capture fine-grained details from short-term observations. We show that ProVideLLM reduces the token count for representing long-form procedural videos while effectively capturing critical spatio-temporal dependencies. We demonstrated state-of-the-art results on six tasks across four datasets, showing per-frame inference at 10 FPS and streaming dialogue at 25 FPS. ProVideLLM is the first real-time framework unifying recognition, anticipation, and planning for procedural assistance. Our multimodal interleaved cache is also generic, applicable to non-procedural videos, and unlocking a new area of compressing long videos on the fly while reducing redundancy and boosting performance.", + "bbox": [ + 511, + 523, + 908, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. DETR-QFormer, trained on pseudo ground-truth bounding boxes, generalizes well but could benefit from GT data to reduce biases from noisy annotations. Additionally ProVideLLM focuses on detection but lacks real-time feedback on the \"how\" of task execution. Finally, as discussed in Sec. 4.2, ProVideLLM's generalization performance remains modest, presenting a potential direction for future work.", + "bbox": [ + 511, + 779, + 910, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22593", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2", + "[2] Ibrahim M Alabdulmohsin, Xiaohua Zhai, Alexander Kolesnikov, and Lucas Beyer. Getting vit in shape: Scaling laws for compute-optimal model design. Advances in Neural Information Processing Systems, 36, 2023. 3", + "[3] Kumar Ashutosh, Santhosh Kumar Ramakrishnan, Triantafyllos Afouras, and Kristen Grauman. Video-mined task graphs for keystep recognition in instructional videos. Advances in Neural Information Processing Systems, 36: 67833-67846, 2023. 3", + "[4] Kumar Ashutosh, Santhosh Kumar Ramakrishnan, Triantafyllos Afouras, and Kristen Grauman. Video-mined task graphs for keystep recognition in instructional videos. Advances in Neural Information Processing Systems, 36, 2024. 7", + "[5] Fabien Baradel, Natalia Neverova, Christian Wolf, Julien Mille, and Greg Mori. Object level visual reasoning in videos. In Proceedings of the European Conference on Computer Vision (ECCV), pages 105-121, 2018. 2", + "[6] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020. 5", + "[7] Yizhak Ben-Shabat, Xin Yu, Fatemeh Saleh, Dylan Campbell, Cristian Rodriguez-Opazo, Hongdong Li, and Stephen Gould. The ikaa asm dataset: Understanding people assembling furniture through actions, objects and pose. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 847–859, 2021. 1", + "[8] Yizhak Ben-Shabat, Jonathan Paul, Eviatar Segev, Oren Shrout, and Stephen Gould. Ika ego 3d dataset: Understanding furniture assembly actions from ego-view 3d point clouds. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4355-4364, 2024. 1", + "[9] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. 6, 7", + "[10] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 4, 5", + "[11] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2", + "[12] Chien-Yi Chang, De-An Huang, Danfei Xu, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. Procedure planning in instructional videos. In European Conference on Computer Vision, pages 334–350. Springer, 2020. 2" + ], + "bbox": [ + 101, + 114, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Dibyadip Chatterjee, Fadime Sener, Shugao Ma, and Angela Yao. Opening the vocabulary of egocentric actions. Advances in Neural Information Processing Systems, 36: 33174-33187, 2023. 2", + "[14] Joya Chen, Zhaoyang Lv, Shiwei Wu, Kevin Qinghong Lin, Chenan Song, Difei Gao, Jia-Wei Liu, Ziteng Gao, Dongxing Mao, and Mike Zheng Shou. Videollm-online: Online video large language model for streaming video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18407-18418, 2024. 1, 2, 6, 7, 8", + "[15] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2", + "[16] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision, pages 1-23, 2022. 6", + "[17] Roeland De Geest, Efstratios Gavves, Amir Ghodrati, Zhenyang Li, Cees Snoek, and Tinne Tuytelaars. Online action detection. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14, pages 269-284. Springer, 2016. 1, 2", + "[18] Guodong Ding, Fadime Sener, Shugao Ma, and Angela Yao. Every mistake counts in assembly. arXiv preprint arXiv:2307.16453, 2023. 2", + "[19] Guodong Ding, Fadime Sener, and Angela Yao. Temporal action segmentation: An analysis of modern techniques. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1, 2", + "[20] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2, 3, 6, 7", + "[21] Nikita Dvornik, Isma Hadji, Hai Pham, Dhaivat Bhatt, Brais Martinez, Afsaneh Fazly, and Allan D Jepson. Flow graph to video grounding for weakly-supervised multi-step localization. In European Conference on Computer Vision, pages 319-335. Springer, 2022. 3", + "[22] Alessandro Flaborea, Guido Maria D'Amely di Melendugno, Leonardo Plini, Luca Scofano, Edoardo De Matteis, Antonio Furnari, Giovanni Maria Farinella, and Fabio Galasso. Prego: online mistake detection in procedural egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18483-18492, 2024. 2", + "[23] Antonino Furnari and Giovanni Maria Farinella. What would you expect? anticipating egocentric actions with rolling-unrolling lstms and modality attention. In Proceedings of the IEEE/CVF International conference on computer vision, pages 6252-6261, 2019. 2" + ], + "bbox": [ + 522, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22594", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Antonino Furnari, Sebastiano Battiato, and Giovanni Maria Farinella. Leveraging uncertainty to rethink loss functions and evaluation measures for egocentric action anticipation. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 6", + "[25] Mingfei Gao, Mingze Xu, Larry S Davis, Richard Socher, and Caiming Xiong. Startnet: Online detection of action start in untrimmed videos. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5542-5551, 2019. 2", + "[26] Rohit Girdhar and Kristen Grauman. Anticipative video transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13505-13515, 2021. 2", + "[27] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 2", + "[28] Kristen Grauman, Andrew Westbury, Lorenzo Torresani, Kris Kitani, Jitendra Malik, Triantafyllos Afouras, Kumar Ashutosh, Vijay Baiyya, Siddhant Bansal, Bikram Boote, et al. Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19383-19400, 2024. 2, 3, 4, 6", + "[29] Bo He, Hengduo Li, Young Kyun Jang, Menglin Jia, Xuefei Cao, Ashish Shah, Abhinav Shrivastava, and Ser-Nam Lim. Ma-lmm: Memory-augmented large multimodal model for long-term video understanding. CVPR, 2024. 1, 2, 3", + "[30] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2", + "[31] Cheng-Ping Hsieh, Simeng Sun, Samuel Kriman, Shantanu Acharya, Dima Rekesh, Fei Jia, Yang Zhang, and Boris Ginsburg. Ruler: What's the real context size of your long-context language models? arXiv preprint arXiv:2404.06654, 2024. 1, 3", + "[32] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 6", + "[33] Md Mohaiminul Islam, Tushar Nagarajan, Huiyu Wang, Fu-Jen Chu, Kris Kitani, Gedas Bertasius, and Xitong Yang. Propose, assess, search: Harnessing llms for goal-oriented planning in instructional videos. arXiv preprint arXiv:2409.20557, 2024. 2", + "[34] Y.-G. Jiang, J. Liu, A. Roshan Zamir, G. Toderici, I. Laptev, M. Shah, and R. Sukthankar. THUMOS challenge: Action recognition with a large number of classes. http://crcv.ucf.edu/THUMOS14/, 2014.1", + "[35] Mahnaz Koupae and William Yang Wang. Wikihow: A large scale text summarization dataset. arXiv preprint arXiv:1810.09305, 2018. 3" + ], + "bbox": [ + 99, + 92, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Bolin Lai, Sam Toyer, Tushar Nagarajan, Rohit Girdhar, Shengxin Zha, James M Rehg, Kris Kitani, Kristen Grauman, Ruta Desai, and Miao Liu. Human action anticipation: A survey. arXiv preprint arXiv:2410.14045, 2024. 1", + "[37] Shih-Po Lee, Zijia Lu, Zekun Zhang, Minh Hoai, and Ehsan Elhamifar. Error detection in egocentric procedural task videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18655-18666, 2024. 2", + "[38] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 2, 3, 7", + "[39] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2, 4, 8", + "[40] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 2", + "[41] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 2, 7", + "[42] Han Lin, Tushar Nagarajan, Nicolas Ballas, Mido Assran, Mojtaba Komeili, Mohit Bansal, and Koustuv Sinha. Vedit: Latent prediction architecture for procedural video representation learning. arXiv preprint arXiv:2410.03478, 2024. 6, 7", + "[43] Xudong Lin, Fabio Petroni, Gedas Bertasius, Marcus Rohrbach, Shih-Fu Chang, and Lorenzo Torresani. Learning to recognize procedural activities with distant supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13853-13863, 2022. 2, 7", + "[44] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2023. 2, 3", + "[45] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 2, 8", + "[46] Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. Lost in the middle: How language models use long contexts. Transactions of the Association for Computational Linguistics, 12: 157-173, 2024. 1, 3", + "[47] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 2", + "[48] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of" + ], + "bbox": [ + 522, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22595", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF international conference on computer vision, pages 2630-2640, 2019. 2", + "[49] Medhini Narasimhan, Arsha Nagrani, Chen Sun, Michael Rubinstein, Trevor Darrell, Anna Rohrbach, and Cordelia Schmid. Tl; dw? summarizing instructional videos with task relevance and cross-modal saliency. In European Conference on Computer Vision, pages 540-557. Springer, 2022. 2", + "[50] Medhini Narasimhan, Licheng Yu, Sean Bell, Ning Zhang, and Trevor Darrell. Learning and verification of task structure in instructional videos. arXiv preprint arXiv:2303.13519, 2023. 7", + "[51] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 7", + "[52] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 4", + "[53] Zhanzhong Pang, Fadime Sener, and Angela Yao. Context-enhanced memory-refined transformer for online action detection. arXiv preprint arXiv:2503.18359, 2025. 1, 2", + "[54] Dhruvesh Patel, Hamid Eghbalzadeh, Nitin Kamra, Michael Louis Iuzzolino, Unnat Jain, and Ruta Desai. Pretrained language models as visual planners for human assistance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15302-15314, 2023. 3", + "[55] Reiner Pope, Sholto Douglas, Aakanksha Chowdhery, Jacob Devlin, James Bradbury, Jonathan Heek, Kefan Xiao, Shivani Agrawal, and Jeff Dean. Efficiently scaling transformer inference. Proceedings of Machine Learning and Systems, 5:606-624, 2023. 5", + "[56] Shraman Pramanick, Yale Song, Sayan Nag, Kevin Qinghong Lin, Hardik Shah, Mike Zheng Shou, Rama Chellappa, and Pengchuan Zhang. Egovlpv2: Egocentric video-language pre-training with fusion in the backbone. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5285-5297, 2023. 7", + "[57] Alec Radford. Improving language understanding by generative pre-training. Openai, 2018. 5", + "[58] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2", + "[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 4", + "[60] Francesco Ragusa, Antonino Furnari, Salvatore Livatino, and Giovanni Maria Farinella. The meccano dataset: Understanding human-object interactions from egocentric videos in an industrial-like domain. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1569-1578, 2021. 1" + ], + "bbox": [ + 99, + 92, + 483, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[61] Abhinav Rai, Fadime Sener, and Angela Yao. Transformed rois for capturing visual transformations in videos. Computer Vision and Image Understanding, 224:103558, 2022. 2", + "[62] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 2", + "[63] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 5", + "[64] Luigi Seminara, Giovanni Maria Farinella, and Antonino Furnari. Differentiable task graph learning: Procedural activity representation and online mistake detection from egocentric videos. In Advances in Neural Information Processing Systems, pages 59373-59407. Curran Associates, Inc., 2024. 3", + "[65] Fadime Sener and Angela Yao. Zero-shot anticipation for instructional activities. In Proceedings of the IEEE/CVF international conference on computer vision, pages 862-871, 2019. 1, 2", + "[66] Fadime Sener, Dipika Singhania, and Angela Yao. Temporal aggregate representations for long-range video understanding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16, pages 154-171. Springer, 2020. 2, 6", + "[67] Fadime Sener, Dibyadip Chatterjee, Daniel Shelepov, Kun He, Dipika Singhania, Robert Wang, and Angela Yao. Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21096-21106, 2022. 1, 2, 3, 6", + "[68] Md Salman Shamil, Dibyadip Chatterjee, Fadime Sener, Shugao Ma, and Angela Yao. On the utility of 3d hand poses for action recognition. In European Conference on Computer Vision, pages 436-454. Springer, 2024. 4", + "[69] Dandan Shan, Jiaqi Geng, Michelle Shu, and David F Fouhey. Understanding human hands in contact at internet scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9869-9878, 2020. 4, 5, 6", + "[70] Zheng Shou, Junting Pan, Jonathan Chan, Kazuyuki Miyazawa, Hassan Mansour, Anthony Vetro, Xavier Giro-i Nieto, and Shih-Fu Chang. Online detection of action start in untrimmed, streaming videos. In Proceedings of the European conference on computer vision (ECCV), pages 534-551, 2018. 1, 2", + "[71] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 2" + ], + "bbox": [ + 522, + 92, + 906, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22596", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[72] Yale Song, Eugene Byrne, Tushar Nagarajan, Huiyu Wang, Miguel Martin, and Lorenzo Torresani. Ego4d goal-step: Toward hierarchical understanding of procedural activities. Advances in Neural Information Processing Systems, 36, 2023. 2, 3, 4, 6", + "[73] Reuben Tan, Xineng Sun, Ping Hu, Jui-hsien Wang, Hanieh Deilamsalehy, Bryan A Plummer, Bryan Russell, and Kate Saenko. Koala: Key frame-conditioned long video-llm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13581–13591, 2024. 2", + "[74] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1207-1216, 2019. 2, 6", + "[75] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 3", + "[76] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568-9578, 2024. 3", + "[77] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2", + "[78] Huiyu Wang, Mitesh Kumar Singh, and Lorenzo Torresani. Ego-only: Egocentric action detection without exocentric transferring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5250-5261, 2023. 6", + "[79] Jiahao Wang, Guo Chen, Yifei Huang, Limin Wang, and Tong Lu. Memory-and-anticipation transformer for online action understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13824–13835, 2023. 1, 2, 3, 4, 6", + "[80] Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video understanding with large language model as agent. arXiv preprint arXiv:2403.10517, 2024. 2", + "[81] Yuetian Weng, Mingfei Han, Haoyu He, Xiaojun Chang, and Bohan Zhuang. Longvlm: Efficient long video understanding via large language models. arXiv preprint arXiv:2404.03384, 2024. 2, 5", + "[82] Wikipedia. How to do anything. http://www.wikipedia.com/, 2005.1", + "[83] Shiwei Wu, Joya Chen, Kevin Qinghong Lin, Qimeng Wang, Yan Gao, Qianli Xu, Tong Xu, Yao Hu, Enhong Chen, and Mike Zheng Shou. Videollm-mod: Efficient video-language streaming with mixture-of-depths vision computation. arXiv preprint arXiv:2408.16730, 2024. 1, 2, 6, 7, 8", + "[84] Junbin Xiao, Angela Yao, Yicong Li, and Tat-Seng Chua. Can i trust your answer? visually grounded video question" + ], + "bbox": [ + 99, + 92, + 485, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13204-13214, 2024. 3", + "[85] Junbin Xiao, Nanxin Huang, Hangyu Qin, Dongyang Li, Yicong Li, Fengbin Zhu, Zhulin Tao, Jianxing Yu, Liang Lin, Tat-Seng Chua, et al. Videoqa in the era of llms: An empirical study. International Journal of Computer Vision, pages 1-24, 2025. 3", + "[86] Mingze Xu, Yuanjun Xiong, Hao Chen, Xinyu Li, Wei Xia, Zhuowen Tu, and Stefano Soatto. Long short-term transformer for online action detection. Advances in Neural Information Processing Systems, 34:1086-1099, 2021. 1, 2, 3, 4, 6", + "[87] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 3", + "[88] Olga Zatsarynna and Juergen Gall. Action anticipation with goal consistency. In 2023 IEEE International Conference on Image Processing (ICIP), pages 1630-1634. IEEE, 2023. 6", + "[89] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 2, 3, 4", + "[90] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Is an object-centric video representation beneficial for transfer? In Proceedings of the Asian Conference on Computer Vision, pages 1976-1994, 2022. 2", + "[91] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Helping hands: An object-aware ego-centric video recognition model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13901-13912, 2023. 4", + "[92] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 543–553, 2023. 2", + "[93] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 3", + "[94] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 2", + "[95] He Zhao, Isma Hadji, Nikita Dvornik, Konstantinos G Derpanis, Richard P Wildes, and Allan D Jepson. P3iv: Probabilistic procedure planning from instructional videos with weak supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2938-2948, 2022. 2", + "[96] Qi Zhao, Shijie Wang, Ce Zhang, Changcheng Fu, Minh Quan Do, Nakul Agarwal, Kwonjoon Lee, and Chen Sun. AntGPT: Can large language models help long-term action anticipation from videos? In The Twelfth International Conference on Learning Representations, 2024. 2" + ], + "bbox": [ + 524, + 92, + 906, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "22597", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[97] Yue Zhao and Philipp Krahenbuhl. Real-time online video detection with temporal smoothing transformers. In European Conference on Computer Vision, pages 485-502. Springer, 2022. 1, 2, 3, 6", + "[98] Yue Zhao, Long Zhao, Xingyi Zhou, Jialin Wu, Chun-Te Chu, Hui Miao, Florian Schroff, Hartwig Adam, Ting Liu, Boqing Gong, et al. Distilling vision-language models on millions of videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13106-13116, 2024. 2", + "[99] Yiwu Zhong, Licheng Yu, Yang Bai, Shangwen Li, Xueting Yan, and Yin Li. Learning procedure-aware video representation from instructional videos and their narrations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14825-14835, 2023. 7", + "[100] Honglu Zhou, Roberto Martin-Martín, Mubbasir Kapadia, Silvio Savarese, and Juan Carlos Niebles. Procedure-aware pretraining for instructional video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10727-10738, 2023. 1, 2, 3, 7", + "[101] Luowei Zhou, Chenliang Xu, and Jason Corso. Towards automatic learning of procedures from web instructional videos. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 1, 2", + "[102] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2", + "[103] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramazan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Crosstask weakly supervised learning from instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3537-3545, 2019. 2" + ], + "bbox": [ + 93, + 92, + 485, + 583 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "22598", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2025/Streaming VideoLLMs for Real-Time Procedural Video Understanding/e3ae8cdd-650d-48b8-9247-fe7363ebedec_model.json b/2025/Streaming VideoLLMs for Real-Time Procedural Video Understanding/e3ae8cdd-650d-48b8-9247-fe7363ebedec_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0ffd1cb0dbf454677cf404e32292700453f68dd6 --- /dev/null +++ b/2025/Streaming VideoLLMs for Real-Time Procedural Video Understanding/e3ae8cdd-650d-48b8-9247-fe7363ebedec_model.json @@ -0,0 +1,2778 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.143, + 0.131, + 0.856, + 0.154 + ], + "angle": 0, + "content": "Streaming VideoLLMs for Real-time Procedural Video Understanding" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.18, + 0.916, + 0.219 + ], + "angle": 0, + "content": "Dibyadip Chatterjee\\(^{1,3*}\\) Edoardo Remelli\\(^{1}\\) Yale Song\\(^{2}\\) Bugra Tekin\\(^{1}\\) Abhay Mittal\\(^{1}\\) Bharat Bhatnagar\\(^{1}\\) Necati Cihan Camgoz\\(^{1}\\) Shreyas Hampali\\(^{1}\\) Eric Sauser\\(^{1}\\) Shugao Ma\\(^{1}\\) Angela Yao\\(^{3}\\) Fadime Sener\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.224, + 0.805, + 0.245 + ], + "angle": 0, + "content": "\\(^{1}\\)Meta Reality Labs \\(^{2}\\)FAIR, Meta \\(^{3}\\)National University of Singapore" + }, + { + "type": "text", + "bbox": [ + 0.378, + 0.246, + 0.638, + 0.26 + ], + "angle": 0, + "content": "dibschat.github.io/ProVideLLM" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.306, + 0.885, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.496, + 0.908, + 0.555 + ], + "angle": 0, + "content": "Figure 1. Overview of ProVideLLM. A streaming video large language model for real-time procedural video tasks with a low memory footprint. It features a multimodal interleaved cache that stores textual descriptions of long-term observations (spanning several minutes) and visual tokens representing short-term observations (spanning a few seconds). We also introduce a new DETR-QFormer connector for better fine-grained tokenization of the short-term. ProVideLLM is capable of handling multiple procedural tasks within a single model." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.578, + 0.327, + 0.594 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.616, + 0.486, + 0.859 + ], + "angle": 0, + "content": "We introduce ProVideLLM, an end-to-end framework for real-time procedural video understanding. ProVideLLM integrates a multimodal cache configured to store two types of tokens – verbalized text tokens, which provide compressed textual summaries of long-term observations, and visual tokens, encoded with DETR-QFormer to capture fine-grained details from short-term observations. This design reduces token count by \\(22 \\times\\) over existing methods when representing one hour of long-term observations while effectively encoding fine-granularity of the present. By interleaving these tokens in our multimodal cache, ProVideLLM achieves sublinear scaling of memory and compute with video length, ensuring per-frame streaming inference at 10 FPS and streaming dialogue at 25 FPS, with a minimal 2GB GPU memory footprint. ProVideLLM also sets new state-of-the-art results on six procedural tasks across four datasets." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.571, + 0.645, + 0.587 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.909, + 0.734 + ], + "angle": 0, + "content": "Procedural videos, spanning several minutes to hours, capture multi-step tasks [82] such as cooking [65, 101], toy assembly [60, 67], and furniture building [7, 8]. With the growing adoption of AR assistants, streaming video understanding is essential for real-time procedural assistance. This requires real-time, online inference, as frames appear in a live video, without access to future observations. Despite its importance, streaming video understanding remains largely unexplored in existing procedural research [19, 36, 100]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Existing works on streaming video understanding [17, 70] focus on online action detection [53, 79, 86, 97] or live conversations [14, 83]. These use-cases target a small set of unrelated actions [17, 34] and do not require long-term temporal modeling. In contrast, procedural videos are inherently long and causal, where each step depends on prior steps. Applying existing approaches to such videos significantly increases token count and memory overhead, leading to long-context degradation [31, 46] and making them unsuitable for streaming, even with excessive frame dropping [29]. The semantic inter-connectivity of activity steps introduces re" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.338, + 0.901 + ], + "angle": 0, + "content": "*Work done during an internship at Meta." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22586" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "dundancy when long-term observations are stored as raw frame tokens [79, 86]. Furthermore, while short-term observations are crucial for understanding the present, not all frame tokens are equally relevant, since procedural tasks often center around hand-object interactions [28, 67, 72]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.17, + 0.483, + 0.32 + ], + "angle": 0, + "content": "In this paper, we propose an end-to-end framework for streaming procedural video understanding (Fig. 1), addressing the above mentioned challenges of processing long-form videos. Specifically, we introduce ProVideLLM, a Procedural VideoLLM that comprises a visual (frame) encoder, a vision-language connector, and a pre-trained large language model (LLM) [1, 20, 58, 77] as the decoder. Our key contribution is a vision-language multimodal cache that efficiently manages streaming video data by balancing compression and representation quality of past observations." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.323, + 0.486, + 0.64 + ], + "angle": 0, + "content": "We factorize the streaming past observations into long-term semantic history and short-term visual context. They are encoded as text and visual tokens, respectively, and interleaved in our multimodal cache. First, for the long-term history spanning several minutes, ProVideLLM decoder predicts step labels per frame, groups similar steps into verbalized sequences (e.g. get \\(\\rightarrow\\) crack \\(\\rightarrow\\) whisk eggs), and stores them in the cache. This semantic compression significantly reduces token count and also enhances current step prediction (e.g. \\(\\rightarrow\\) pour mixture). Second, to capture short-term visual context, typically spanning a few seconds, understanding the fine-grained changes across frames is crucial. However, standard VideoLLM encoders [59, 89] face two key challenges in such modeling: (i) temporal collapse, where low temporal variance conflates distinct actions and (ii) background distractors across frames [5, 13, 61, 90]. We mitigate the former by using visual encoders trained with dense spatial alignment objectives [11, 30, 52]. We address the latter with a novel vision-text connector DETR-QFormer, pretrained to explicitly focus on hand-object interaction regions, thereby compressing them into a smaller set of tokens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.793 + ], + "angle": 0, + "content": "Finally, in existing methods [86], both long- and short-term tokens are visual and stored separately. Applying such separate caching to our multimodal tokens would require redundant attention recomputation when verbalizing short-term visual tokens into long-term text. To overcome this, we interleave text and visual tokens within our multimodal cache. This enables prefilling the cache, which ensures that computation scales linearly with cache size and sub-linearly with the video length, thereby significantly reducing latency for real-time performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In summary, our contributions are: (i) a first-of-its-kind streaming multimodal cache with interleaved text and visual tokens, enabling per-frame inference [72] at 10 FPS and streaming dialogue [14] at 25 FPS; (ii) verbalizing long-term observations into compressed textual summaries and (iii) modeling hand-object interactions for short-term observations using DETR-QFormer; all combined, it achieves a" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "\\(22 \\times\\) reduction in token count compared to existing methods. (iv) state-of-the-art results on six tasks across four datasets, making ProVideLLM the first real-time framework unifying recognition, anticipation, planning for procedural guidance." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.158, + 0.664, + 0.173 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.183, + 0.907, + 0.394 + ], + "angle": 0, + "content": "Procedural Video Understanding have primarily focused on instructional YouTube videos [48, 74, 101, 103]. Recent large-scale egocentric datasets [27, 28, 67, 72] have advanced research by providing longer videos with more activity steps. Existing works have focused on offline tasks [19, 43, 100], where the entire video is available for prediction, as well as some online tasks, like procedural planning [12, 33, 95], forecasting [23, 26, 49, 65, 66, 96], and mistake detection [18, 22, 37, 67]. However, deploying these models for streaming inference results in high computational complexity where it grows quadratically with the video length. In contrast, ProVideLLM incurs a sub-linear cost per update, both in terms of memory and runtime, enabling efficient streaming inference." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.395, + 0.909, + 0.667 + ], + "angle": 0, + "content": "Video Large Language Models. Multimodal large language models (MLLMs) have been popularized by BLIP-2-like [15, 39, 102] and LLaVA-like [38, 44, 45] models where an LLM is trained to understand vision-text data. Extending these to video have led to the development of Video Large Language Models (VideoLLMs) [40, 41, 47, 83, 94, 98]. However, most of these models are limited to offline tasks [62, 71, 73, 80, 81] where the entire video is queried to answer a question. As such, architectural elements such as VideoQ-Former [92] and temporal pooling layers [40], developed without causal requirements, limit streaming applications. Models like MA-LMM [29] and VideoLLM online [14] address video understanding online, but are too computationally intensive to be deployed for efficient streaming inference on long-form videos. ProVideLLM advances these models by introducing a multimodal interleaved cache for faster inference and proposing DETR-QFormer for modeling fine-grained visual changes, improving accuracy." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.667, + 0.909, + 0.818 + ], + "angle": 0, + "content": "Streaming Video Understanding has been explored for online action detection [17, 25, 70] where dominant approaches [53, 79, 86, 97] split streaming input into long- and short-term visual histories within an encoder-decoder architecture. However, storing both as visual tokens is redundant and computationally prohibitive for long-form procedural videos, where multiple actions are temporally connected. Instead, ProVideLLM verbalizes long-term observations into text which drastically decreases token count required to represent hour long videos, thus improving memory efficiency." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.83, + 0.652, + 0.845 + ], + "angle": 0, + "content": "3. ProVideLLM" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "ProVideLLM encodes long- and short-term streaming observations into text and visual tokens and interleaves them in a multimodal cache. We first present the preliminaries" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22587" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.138 + ], + "angle": 0, + "content": "in Sec. 3.1, followed by long- and short-term token generation in Sec. 3.2 and Sec. 3.3. Finally, we present our multimodal cache in Sec. 3.4 and training objectives in Sec. 3.5." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.143, + 0.232, + 0.158 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.163, + 0.485, + 0.239 + ], + "angle": 0, + "content": "ProVideLLM follows a LLaVA-style [44] architecture, consisting of a visual encoder \\((E_V)\\), a language decoder \\((D_L)\\) and a vision-language connector \\((C_{VL})\\). Given a video \\(V = \\{v_1, v_2, \\ldots, v_T\\}\\), the model's output \\((o_t)\\) at any time step \\(t \\in [1, T]\\) is" + }, + { + "type": "equation", + "bbox": [ + 0.122, + 0.244, + 0.484, + 0.268 + ], + "angle": 0, + "content": "\\[\no _ {t} = \\operatorname {P r o V i d e L L M} \\left(v _ {1: t}\\right) = D _ {L} \\left(\\left\\| _ {i = 1} ^ {t} C _ {V L} \\left(E _ {V} \\left(v _ {i}\\right)\\right)\\right) \\quad (1) \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.272, + 0.485, + 0.379 + ], + "angle": 0, + "content": "where \\( \\parallel \\) denotes concatenation. While LSTR-inspired long short-term memory approaches [79, 97] also use an encoder-decoder architecture, they rely exclusively on visual tokens (Fig. 2a.1). In contrast, ProVideLLM incorporates a language decoder, which enables more efficient long-term compression, improves generalization, and supports multiple tasks within a single model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.384, + 0.393, + 0.399 + ], + "angle": 0, + "content": "3.2. Tokenizing long-term observations" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.403, + 0.485, + 0.646 + ], + "angle": 0, + "content": "Procedural videos are inherently long-form, ranging from tens of minutes [67] to several hours [28, 72]. While modern LLMs [20, 75, 87] offer massive context lengths and, in theory, can process an hour-long raw video at 24 fps, performance degrades with larger context sizes [31, 46], and training such models is highly memory-intensive [29, 93]. Accurately recognizing ongoing steps require effectively encoding long-term context. Our key insight is that procedural steps in a video are temporally connected, i.e. a sequence of step labels can serve as a strong predictor of ongoing steps. For instance, when \"preparing an omelet\", the step \"pour mixture into pan\" depends on earlier steps \"get eggs from fridge, crack eggs into bowl, whisk eggs\". Prior works on flow graphs [21, 64] and planning [3, 54, 100] also show that step sequences, often extracted from external sources [35], serve as strong baselines for procedural understanding." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.646, + 0.486, + 0.827 + ], + "angle": 0, + "content": "Instead of constructing computationally expensive graphs, we propose online verbalization (Fig. 2a.2). Given a streaming video, ProVideLLM decoder continuously generates action predictions (summaries) of the present observation on the fly, and groups semantically similar actions to verbalize the past (Fig. 1). This approach efficiently compresses past visual frame tokens into information-dense language tokens. For example, in Ego4D Goal-Step [72], storing one hour of long-term past requires only 630 verbalized text tokens on average. Compared to long short-term approaches [79, 86, 97], our verbalization reduces token count by \\(22 \\times\\) at the same sampling rate, significantly reducing the memory overhead." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.4, + 0.85 + ], + "angle": 0, + "content": "3.3. Tokenizing short-term observations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Our goal is to recognize procedural activity steps in a streaming manner. This requires understanding \"what is happening\" in the present while tracking \"what has happened in" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.9, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.26, + 0.906, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.361, + 0.908, + 0.487 + ], + "angle": 0, + "content": "Figure 2. Converting short-term tokens to long-term under various caching types. (a) Factorizing the streaming observations into long-term and short-term and ordering them progressively. (a.1) Conversion done at \\( O(1) \\) cost but is unable to represent long-form videos due to massive token count and memory needs. (a.2) Online verbalization can represent long-form videos at a manageable memory but is not suitable for streaming due to \\( O(N_S^2 + N_S N_L) \\) conversion cost. (b) Interleaving them reduces the conversion cost to \\( O(N) \\), allowing streaming inference on long-form videos." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.909, + 0.675 + ], + "angle": 0, + "content": "the past\". Language is a strong indicator of the semantic structure of step sequences, but determining if and when the step is performed requires fine-grained visual information. Therefore, we allocate more compute for understanding the short-term past, typically spanning 8-16 seconds, by keeping its visual tokens uncompressed. Standard VideoLLM encoders [38, 44] face two key challenges in capturing fine-grained short-term observations: \\((i)\\) they generate temporally collapsed video features, a limitation we identify and address in Sec. 3.3.1, and \\((ii)\\) they get biased towards high-level scene and background distractors, which we address in Sec. 3.3.2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.685, + 0.757, + 0.699 + ], + "angle": 0, + "content": "3.3.1. Choosing the visual encoder" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.909, + 0.902 + ], + "angle": 0, + "content": "The predominant visual encoders for MLLMs are CLIP [59] and SigLIP [2, 89]. These encoders are pre-trained with only a language-alignment objective for integration with LLMs. However, because they are trained on high-level and sparse web-scale captions such as \"cooking in the kitchen\", they struggle with fine-grained reasoning, particularly in answering the \"what?\" and \"how to?\" aspects of a video [84, 85]. Consequently, distinct images may yield high similarity scores in their feature space [76], leading to low feature diversity across frames within a step. This results in relatively static frame features, a phenomenon we term temporal collapse. For certain fine-grained procedural steps, such as \"pick up\" vs. \"put down\", temporally collapsed" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22588" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.132, + 0.114, + 0.145 + ], + "angle": 0, + "content": "a)" + }, + { + "type": "image", + "bbox": [ + 0.12, + 0.09, + 0.474, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.226, + 0.114, + 0.24 + ], + "angle": 0, + "content": "b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.296, + 0.484, + 0.395 + ], + "angle": 0, + "content": "Figure 3. a) Per-class mean temporal variance on EgoExo4D & Ego4D Goal-Step. Highly temporally varying classes on the left and low-variance classes on the right. b) Visualization of the top-3 PCA components of patch tokens computed across video frames, thresholded by first component. The primary factors of variation for DINOv2 [52] are hands, and are unfocused and scattered across the image for language-aligned encoders, CLIP [59] and SigLIP [89]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.483, + 0.444 + ], + "angle": 0, + "content": "features make it nearly impossible for the downstream LLM to decode the correct step." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.444, + 0.485, + 0.685 + ], + "angle": 0, + "content": "Fig. 3a shows the impact of temporal collapse across three visual encoders – two language-aligned ones [59, 89] and one self-supervised1 [52]. We plot temporal variance per class in descending order, with high-variance classes on the left and low-variance classes on the right (see Suppl.Sec.1 for temporal variance computation). At both extremes, the trends are consistent across all encoders. High-variance classes, which often involve verbs such as “get”, “fetch” or “put” correspond to longer videos with substantial scene and environmental changes. Conversely, low-variance classes involving adverbs such as “slowly” or “boil” or repetitive actions like “rotate” or “peel”, exhibit gradual state changes. Among the three models, DINOv2 consistently captures temporal variations more effectively, covering a broader range of temporal variance across different classes. Hence, we choose DINOv2 as our visual encoder." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.694, + 0.466, + 0.709 + ], + "angle": 0, + "content": "3.3.2. Aligning hand-object transformations with text" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.713, + 0.484, + 0.834 + ], + "angle": 0, + "content": "Procedural datasets predominantly capture hand-centric actions. For instance, nearly \\(99\\%\\) of actions in [28, 72] involve hand interactions2. Moreover, video frames often contain distracting background objects, while hand locations serve as strong cues for recognizing the ongoing actions [68, 91]. Thus, a vision encoder focused on hand regions can enhance performance by mitigating background distractions and capturing fine-grained transformations more precisely." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.835, + 0.483, + 0.865 + ], + "angle": 0, + "content": "Another contribution of our work is a vision-language connector, DETR-QFormer, pretrained to attend to salient" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.098, + 0.907, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.328, + 0.908, + 0.385 + ], + "angle": 0, + "content": "Figure 4. DETR-QFormer architecture & Stage-1 pre-training. Trained to detect hands and objects-in-contact from pseudo ground truths bboxes generated by [69], DETR-QFormer queries DINOv2 patch tokens to extract hand-object-focused activations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.909, + 0.546 + ], + "angle": 0, + "content": "objects. In procedural videos, these are typically in contact with the hand. Interestingly, our choice of DINOv2 as the visual encoder provides an advantage as its patch tokens capture hands as primary factors of variation Fig. 3b. Leveraging this, we design DETR-QFormer to query DINOv2 patch tokens, extracting hand-object activations and suppressing background distractions. DETR-QFormer explicitly attends to hand-object activations using a DETR loss [10] and compresses salient regions into fewer tokens via QFormer's [39] cross-attention, enhancing representation quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.547, + 0.909, + 0.822 + ], + "angle": 0, + "content": "Architecture. Our connector integrates Q-Former's compression [39] and DETR's bbox loss into a single crossattention block. DETR-QFormer, illustrated in Fig. 4, is pretrained to detect the hands and objects-in-contact (Sec. 3.5). The connector \\((C_{VL})\\) is a cross-attention transformer decoder with a set of \\(m\\) learnable visual queries \\(Q_{v} = (q_{v_{1}}, q_{v_{2}}, \\ldots, q_{v_{m}})\\). Key-value pairs \\((K, V)\\) are generated from the flattened \\(n^{2}\\) DINOv2 patch tokens using two projection heads \\((W_{K}, W_{V})\\) respectively. To explicitly model hand-object interactions, we append two sets of special queries - two hand queries \\(Q_{h} = (q_{h_{1}}, q_{h_{2}})\\) for the left and right hands, and \\(k\\) object queries \\(Q_{o} = (q_{o_{1}}, q_{o_{2}}, \\ldots, q_{o_{k}})\\) corresponding to objects-in-contact. The role of visual queries \\(Q_{v}\\) is to compress the \\(n^{2} \\times D\\) patch tokens into \\(m \\times d\\) tokens, where \\(m << n^{2}\\) and \\(d << D\\). Whereas, outputs corresponding to the learnable hand-object queries \\((Q_{h}^{\\prime}, Q_{o}^{\\prime})\\) are passed through an MLP to make bounding box predictions, \\(\\hat{B} = \\{\\hat{b}_{h_{1}}, \\hat{b}_{h_{2}}, \\hat{b}_{o_{1}}, \\ldots, \\hat{b}_{o_{k}}\\}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.785, + 0.847 + ], + "angle": 0, + "content": "3.4. Multimodal Interleaved Cache" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "The effectiveness of a procedural video model is best assessed when deployed for real-time inference. Naive long-and short-term approaches [79, 86] maintain separate caches" + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1We chose DINOV2 due to its comparable scale to CLIP and SigLIP. \n2Remaining 1% are visually undetectable actions \"call for help, waiting.\"" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22589" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.093, + 0.092, + 0.383, + 0.106 + ], + "angle": 0, + "content": "Algorithm 1 Multimodal Interleaved Cache" + }, + { + "type": "algorithm", + "bbox": [ + 0.093, + 0.109, + 0.484, + 0.287 + ], + "angle": 0, + "content": "1: class MultimodalInterleavedCache: \n2: def __init__(self, N.S, N.L): \n3: self_tokens = [] \n4: self.N.S = N.S \n5: self.N.L = N.L \n6: # entry function: adds a new token to cache \n7: def entry(self, token): \n8: self_tokens.append(token) \n9: # exit function for short-term tokens \n10: def exit_short(self): \n11: if self_tokens.count() > self.N.S: \n12: self_tokens.pop(self_tokens.index()) \n13: return True \n14: # exit function for long-term tokens \n15: def exit_long(self): \n16: if self_tokens.count() > self.N.L: \n17: self_tokens.pop(self_tokens.index()" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.32, + 0.485, + 0.456 + ], + "angle": 0, + "content": "for the two token types, arranged progressively with a single entry point (see Fig. 2a). With multimodal tokens, such setups require two distinct entry points, forcing attention recomputation over the short-term cache whenever a short-term token is verbalized into a long-term one. In the worst case, this leads to quadratic scaling of compute. To efficiently leverage online verbalization, we propose a multimodal interleaved cache that interleaves long-term text and short-term visual tokens from Sec. 3.2 and Sec. 3.3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.484, + 0.732 + ], + "angle": 0, + "content": "Interleaving long- and short-term tokens. We store the running short-term visual tokens and long-term text tokens in a single multimodal cache operating as a FIFO (First In, First Out) queue. It has one entry point and two distinct exits for short- and long-term tokens. Our entry and exit functions are detailed in Algorithm 1, where \\( N_{S} \\) denotes the short-term cache length (in number of frames) and \\( N_{L} \\) the long-term cache length (in number of observed steps). As frame tokens \\( (v_{i}) \\) exit the short-term span, the ProVideLLM decoder verbalizes the oldest token \\( (v_{-N_s}) \\). The verbalized text is then added to the cache unless already cached within the past \\( \\tau \\) frames. To distinguish the two token types within the same cache, we prepend a special token (represented as \\( <\\mathbb{L}> \\)) to the long-term text tokens. ProVideLLM is instruction-tuned to recognize tokens associated with \\( <\\mathbb{L}> \\) as belonging to the long-term observations (Sec. 3.5). Since both long- and short-term tokens share a single entry point, visual and textual tokens are interleaved in the cache at any given time." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Streaming inference. Every time a new token is added to the cache (Algorithm 2), instead of re-computing attention across all tokens, we use the causal decoder's K-V cache [55] to compute attention only for the new token over the previous \\( N - 1 \\) tokens. This ensures that the computational complexity for processing a new frame scales linearly with cache size \\( (O(N)) \\). Additionally, maintaining a single interleaved cache enables direct application of optimized attention implementations [6, 81], though such techniques remain orthogonal to our approach. More importantly, for a fixed \\( N_S \\), compression via verbalization ensures that \\( N_L \\)" + }, + { + "type": "code_caption", + "bbox": [ + 0.515, + 0.092, + 0.741, + 0.106 + ], + "angle": 0, + "content": "Algorithm 2 Streaming Inference" + }, + { + "type": "algorithm", + "bbox": [ + 0.517, + 0.109, + 0.905, + 0.226 + ], + "angle": 0, + "content": "1: def inference(V, N_S, N_L, \\(\\tau\\)): \n2: cache = MultimodalInterleavedCache(N_S, N_L, \\(\\tau\\)) \n3: out = [] \n4: for \\(v_i\\) in V: \n5: cache_entry \\((v_i)\\) # add frame tokens to cache \n6: \\(t_i = \\text{ProVideLLM}(\\text{cache.tokens})\\) # current prediction \n7: out.append \\((t_i)\\) # buffer predictions \n8: flag = cache_exit_short() # short-term overflow \n9: if flag and out[-N_S] not in out[-\\(\\tau\\)): \n10: cache_entry(out[-N_S]) # verbalize & interleave \n11: cache_exit_long() # long-term overflow" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.253, + 0.908, + 0.314 + ], + "angle": 0, + "content": "and consequently \\(N = (N_{L} + N_{S})\\) scales sub-linearly with the number of video frames. Hence, compute also scales sublinearly with the number of streaming video frames. This makes our model well-suited for low-latency applications." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.323, + 0.726, + 0.34 + ], + "angle": 0, + "content": "3.5. Training ProVideLLM" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.345, + 0.907, + 0.406 + ], + "angle": 0, + "content": "We adopt a two-stage training approach similar to LLaVA - (1) alignment pre-training to align the encoder's visual tokens to the decoder's language space, and (2) instruction tuning of the model for specific tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.406, + 0.909, + 0.482 + ], + "angle": 0, + "content": "Stage-1: Alignment Pre-training. In this stage, our connector DETR-QFormer \\((C_{VL})\\) is trained while keeping the visual encoder and language decoder frozen (Fig. 4). The objective is to caption short video clips featuring hand-centric actions using the standard language modeling loss [57]:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.505, + 0.907, + 0.548 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {L M}} = - \\frac {1}{| W |} \\sum_ {i = 1} ^ {| W |} \\log P (w _ {i} \\mid w _ {< i}, V; C _ {V L} ^ {\\Theta}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.56, + 0.907, + 0.636 + ], + "angle": 0, + "content": "where \\(V\\) is the input video clip, \\(W = \\{w_{1}, w_{2}, \\dots, w_{|C|}\\}\\) is the ground-truth caption, and \\(C_{VL}^{\\Theta}\\) is our DETR-QFormer parameterized by \\(\\Theta\\). For detecting hands and corresponding objects-in-contact, we use a DETR [10]-style Generalized IoU [63] loss:" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.646, + 0.907, + 0.667 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {H O}} \\left(b _ {i}, \\hat {b} _ {\\sigma (i)}\\right) = \\mathcal {L} _ {\\mathrm {i o u}} \\left(b _ {i}, \\hat {b} _ {\\sigma (i)}\\right) + \\left\\| b _ {i} - \\hat {b} _ {\\sigma (i)} \\right\\| _ {1} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.676, + 0.909, + 0.755 + ], + "angle": 0, + "content": "where \\( b_{i} \\)'s are the pseudo ground-truth bounding boxes of hands and objects-in-contact generated by an off-the-shelf detector [69] and \\( \\hat{b}_{\\sigma(i)} \\in \\hat{B} \\) is the bounding box matched from the predicted set \\( \\hat{B} \\). For simplicity, we keep our loss category-agnostic. The final loss for Stage-1 is thus:" + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.767, + 0.907, + 0.783 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {L M}} + \\lambda_ {1} \\mathcal {L} _ {\\mathrm {H O}} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.908, + 0.84 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\) is a hyper-parameter. In addition to vision-language alignment, Stage-1 also enables our model to attend to hand-object interactions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Stage-2: Instruction Tuning. This stage focuses on fine-tuning our model with task-specific instructions such as \"Provide the goal.\", \"Provide what I am doing now.\", \"Provide what I need to do next.\", or \"Provide the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22590" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.477, + 0.18 + ], + "angle": 0, + "content": "
Modelper-frame mAP
ValTest
LSTR [86] (only short-term)8.8-
LSTR [86]8.98.1
EgoOnly [78]10.310.9
ProVideLLM-1B/5 (no verbalization)12.112.2
ProVideLLM-1B/5 (verbalization & interleaving)13.012.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.191, + 0.483, + 0.234 + ], + "angle": 0, + "content": "Table 1. Ego4D [72] GoalStep Online Step Detection. For fair comparison with LSTR, we consider 16s of short-term past and an avg. of 128s of long-term past compressed into language tokens." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.247, + 0.475, + 0.31 + ], + "angle": 0, + "content": "
ModelViewsAction
TempAgg [66]12 views8.5
+ Goal [88]12 views12.0
ProVideLLM-8B/11 (+long-term verbalization)v413.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.32, + 0.483, + 0.35 + ], + "angle": 0, + "content": "Table 2. Action Anticipation Results on Assembly101 validation set. Performance measured by class-mean Top-5 recall." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.376, + 0.484, + 0.468 + ], + "angle": 0, + "content": "next 10 steps I need to do.\" For instruction tuning, we fine-tune DETR-QFormer and the language decoder using \\(\\mathcal{L}_{\\mathrm{LM}}\\) while keeping the visual encoder frozen. For streaming tasks, we also train our model with interleaved special \\(<\\mathbb{L}>\\) tokens to represent long-term cache entries, distinguishing them from standard prompts. See Suppl.Sec.5 for further details." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.482, + 0.224, + 0.5 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.484, + 0.613 + ], + "angle": 0, + "content": "We conduct extensive evaluations of ProVideLLM on six tasks across four datasets covering streaming (Sec. 4.1.1), long-term (Sec. 4.1.2) and fine-grained (Sec. 4.1.3) procedural understanding. We further assess ProVideLLM's performance in multi-task settings and its generalization to new datasets in Sec. 4.2. Finally, we provide ablations in Sec. 4.3 and ProVideLLM's runtime analysis in Sec. 4.4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. Depending on the choice of vision encoder and language decoder, we have two model variants. ProVideLLM-1B/5 is our efficiency-focused model designed for real-time streaming. It uses DINOv2 (ViT-S) with Llama-3.2-1B-Instruct [20] and requires 5 tokens per frame. ProVideLLM-8B/11 is our more performance-focused model using DINOv2 (ViT-g) and Llama-3.1-8B-Instruct and requires 11 tokens per frame. Furthermore, for benchmarking on youtube videos [74], we introduce an additional model ProVideLLM-8B/11+, which uses SigLIP's [CLS] token along with DINOv2 patch tokens. Unless otherwise mentioned we use \\( N_{S} = 64 \\), \\( N_{L} = 5 \\) i.e. storing a maximum of 64 frames in the short-term cache and a maximum of 5 observed steps in the long-term cache. DETR-QFormer is pre-trained (Stage-1) on EPIC-KITCHENS [16] with pseudo ground-truth bounding boxes from 100DOH [69]. For Stage-2, the LLM (decoder) is fine-tuned using LoRA [32] with \\( r = 128 \\) and \\( \\alpha = 256 \\). Additional implementation details are provided in Suppl.Sec.5." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.089, + 0.897, + 0.231 + ], + "angle": 0, + "content": "
MethodTop-1 Acc. (%)
TimeSformer [9]34.0
DistantSup39.4
VideoTF42.4
ProcedureVRL46.8
VEDIT [42]51.8
VideoLLM-online [14]49.1
VideoLLM-MoD [83]49.7
ProVideoLLM-8B/11+52.1
ProVideoLLM-8B/11+ (+long-term verbalization)53.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.241, + 0.905, + 0.269 + ], + "angle": 0, + "content": "Table 3. Next Step Forecasting in COIN. The last four rows including ProVideLLM are VideoLLMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.296, + 0.908, + 0.387 + ], + "angle": 0, + "content": "Datasets. We evaluate our model's performance on four large-scale procedural datasets - EgoExo4D [28], Ego4D-GoalStep [72], COIN [74], and Assembly101 [67]. As shown in Sec. 4.1, ProVideLLM sets state-of-the-art on both egocentric and YouTube benchmarks, demonstrating robustness across diverse video domains." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.396, + 0.649, + 0.411 + ], + "angle": 0, + "content": "4.1. Main results" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.418, + 0.713, + 0.433 + ], + "angle": 0, + "content": "4.1.1. Online Step Detection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.437, + 0.909, + 0.71 + ], + "angle": 0, + "content": "We report results in Tab. 1, using per-frame mAP on action classes (steps + substeps) at 4fps, a standard metric in online action detection [79, 97]. We evaluate ProVideLLM-1B/5 with and without long-term verbalization in rows four and five, respectively. Results for LSTR [86] and EgoOnly [78] are taken from [72]. Accordingly, LSTR spans \\(128s\\) in long-term memory and \\(16s\\) in short-term memory. In [72], each step spans on average \\(32s\\) with around 5.7 tokens representing each step. Our long-term cache can encode \\(128s\\) of past using only 22 tokens at \\(N_{L} = 4\\) steps. To measure improvements from our long-term design, we set the cache length \\(N = 86\\) (spanning LSTR's temporal context), split into \\((N_{S} = 86, N_{L} = 0)\\) and \\((N_{S} = 64, N_{L} = 4)\\) in rows four and five, respectively. Our model surpasses both baselines, outperforming LSTR by over \\(4\\%\\) at similar temporal span. ProVideLLM also achieves greater gains from long-term context compared to LSTR, highlighting the effectiveness of verbalizing the long-term past." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.716, + 0.729, + 0.731 + ], + "angle": 0, + "content": "4.1.2. Online Step Forecasting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Step forecasting requires modeling long-term dependencies, making it an ideal task to evaluate our contributions. We evaluate ProVideoLLM for step forecasting on COIN and Assembly101 measured by Top-1 accuracy and Top-5 Recall [24], respectively. Our model achieves state-of-the-art performance on both datasets as shown in Tab. 3 and Tab. 2. On Assembly101, we surpass [88] by \\(1.8\\%\\) using only a single view. On COIN, we use Llama-3-8B-Instruct to ensure fair comparisons with VideoLLMs [14, 83] and achieve a performance boost of around \\(4\\%\\). Moreover, while state-of-the-art models [14, 83] use approximately 600 to 6000 visual" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22591" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.089, + 0.472, + 0.214 + ], + "angle": 0, + "content": "
ModelTraining ViewsEgo Acc. (%) Val / Test
TimeSFformer [9]ego35.13 / 35.93
EgoVLPv2 [56]ego+exo39.10 / 38.76
Viewpoint Distillationego+exo38.19 / 39.49
View Invariant Encoder [51]ego+exo40.34 / 41.53
VideoLLM-MoD† [83]ego42.62 / —
ProVideLLM-8B/11ego44.36 / 50.74
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.225, + 0.486, + 0.268 + ], + "angle": 0, + "content": "Table 4. EgoExo4D Fine-grained Keystep Recognition. Test results obtained from EgoExo4D's Fine-grained Keystep Recognition Challenge on EvalAI. † Uses 2fps sampling per segment." + }, + { + "type": "table", + "bbox": [ + 0.106, + 0.281, + 0.472, + 0.42 + ], + "angle": 0, + "content": "
MethodTop-1 Accuracy (%)
TimeSformer [9]34.0
Paprika [100]51.0
DistantSup [43]54.1
ProcedureVRL [99]56.9
VideoTaskGraph [4]57.2
VEDIT [42]62.7
VideoLLM-online [14]63.1
VideoLLM-MoD [83]63.4
ProVideLLM-8B/11+67.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.43, + 0.483, + 0.445 + ], + "angle": 0, + "content": "Table 5. COIN Step Recognition. Last three rows are VideoLLMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.484, + 0.492 + ], + "angle": 0, + "content": "tokens to model past context, our approach spans a larger temporal context with only 200 tokens through verbalization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.501, + 0.258, + 0.515 + ], + "angle": 0, + "content": "4.1.3. Step Recognition" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.52, + 0.486, + 0.701 + ], + "angle": 0, + "content": "In Tab. 4 and Tab. 5, we evaluate ProVideLLM on fine-grained step recognition on EgoExo4D and COIN. We uniformly sample 16 frames from each segment to construct our short-term cache \\((N_S)\\) and use the prediction from the last frame as our output. The closest model, VideoLLM-MoD [83], uses 10 tokens per frame, but densely samples frames at 2fps, resulting in a larger token count. In both cases, we outperform VideoLLM-MoD by \\(1.74\\%\\) and \\(3.9\\%\\) respectively. Moreover, at the time of our submission, our method, using only egocentric video, achieved the best results on the EgoExo4D Fine-grained Recognition Challenge, with a \\(9.2\\%\\) improvement over the second-ranked model." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.712, + 0.431, + 0.729 + ], + "angle": 0, + "content": "4.2. Generalized Procedural Understanding" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.901 + ], + "angle": 0, + "content": "In this section, we explore additional capabilities of ProVideLLM, leveraging the advantages of using a pretrained LLM as the language decoder. This section provides insights into developing a \"generalist\" model that advance procedural video understanding within a unified framework. We first evaluate ProVideLLM's ability to multi-task and then present our benchmark for cross dataset generalization. Multi-task learning. In Tab. 6, we present our results for multi-task learning, an underexplored aspect of procedural understanding. Existing models are trained only for isolated tasks, such as step recognition and forecasting. We compare" + }, + { + "type": "table", + "bbox": [ + 0.528, + 0.089, + 0.895, + 0.178 + ], + "angle": 0, + "content": "
ModelCOINBenchmark Top-1 Accuracy (%)
StepTaskNextProc.Proc.+
VideoTF† [50]56.591.042.440.246.4
VideoLLM-online [14]63.192.749.149.854.1
VideoLLM-MoD [83]63.492.849.749.853.3
ProVideLLM-8B/11+66.995.050.551.055.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.187, + 0.908, + 0.243 + ], + "angle": 0, + "content": "Table 6. Multi-task learning on COIN. Step - step recognition, Task - task recognition, Next - next step forecasting, Proc. - long-term forecasting and \\(\\mathrm{Proc + }\\) - long-term forecasting with task. \\(^{\\dagger}\\) VideoTF [50] is trained separately for each task." + }, + { + "type": "table", + "bbox": [ + 0.528, + 0.256, + 0.895, + 0.323 + ], + "angle": 0, + "content": "
ModelEgo Acc. (%)
VideoLLaVA [41] + Llama-3.1-8B [20]3.6
LLaVA-OneVision [38] + Llama-3.1-8B [20]8.3
ProVideLLM-8B/1112.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.333, + 0.907, + 0.376 + ], + "angle": 0, + "content": "Table 7. Cross-dataset generalization. Evaluated on EgoExo4D KeyStep recognition. Our model, trained on Ego4D Goal-Step is compared with a mixture of vision + language experts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.909, + 0.644 + ], + "angle": 0, + "content": "our model with recent VideoLLMs, also trained in a multitask setting, on 5 tasks. Despite using the same LLM as [14, 83], our model outperforms these baselines across all tasks. Cross-Dataset Generalization. In Tab. 7, we present a cross-dataset generalization benchmark for procedural videos using the EgoExo4D fine-grained keystep recognition task. We compare two recent VideoLLMs [38, 41] which were primarily trained for video captioning and use Llama-3.1-8B-Instruct to convert the captions into EgoExo4D classes. Our baseline models act as a combination of visual and language experts. ProVideLLM, using the same LLM and trained on the much smaller Ego4D Goal-Step dataset, significantly outperforms these baselines. Despite these advances, performance remains modest, highlighting the need for further research in procedural generalization. Additional details are provided in the Suppl. Material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.653, + 0.627, + 0.667 + ], + "angle": 0, + "content": "4.3. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Visual Encoders & DETR-QFormer. In Tab. 8, we ablate visual encoders-SigLIP and DINOv2-and connectorsMLP, QFormer, and our DETR-QFormer-on EgoExo4D keystep recognition. First, comparing encoders across connectors, SigLIP's CLS tokens, being language-aligned, outperform DINOv2; however, with more patch tokens, DINOv2 surpasses SigLIP, aligning with our observations in Fig. 3. Second, comparing connectors across encoders, Q-Former boosts DINOv2's performance \\((+1.2\\%, +1.9\\%)\\) but has minimal impact on SigLIP \\((+0.8\\%, -0.1\\%)\\), highlighting DINOv2's need for a stronger connector. Finally, DETR-QFormer further improves DINOv2 \\((+1.1\\%, +2.6\\%)\\), with gains increasing at higher resolutions but degrading SigLIP's performance due to weak hand-object activations in its patch tokens. Notably, DETR-QFormer remains highly" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22592" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.089, + 0.477, + 0.273 + ], + "angle": 0, + "content": "
[CLS] TokenPatch Tokens#Tokens /frameEgo Val./ Acc. (%)
SigLIPDINOv2
w/ 2-layer MLP [45]
✓X131.8±0.428.7±0.2
✓16×16 → 2×2532.1±0.632.4±0.5
✓16×16 → 4×41733.3±0.636.6±0.5
w/ QFormer [39]
✓16×16 ⇒ 4532.9±0.633.6±0.5
✓16×16 ⇒ 161733.2±0.438.1±0.7
w/ DETR-QFormer (2 hands, 2 objects)
✓16×16 ⇒ 0530.9±0.334.7±0.5
✓16×16 ⇒ 121733.6±0.340.7±0.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.282, + 0.486, + 0.34 + ], + "angle": 0, + "content": "Table 8. DETR-QFormer ablations on EgoExo4D fine-grained keysteps. Llama-3.2-1B-Instruct is used as the language decoder. Results reported over 3 runs. \\((\\rightarrow\\) means avg-pooling, \\(\\Rightarrow\\) means Q-Former compression.)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.486, + 0.41 + ], + "angle": 0, + "content": "token-efficient with DINOv2, even when compressing to only hand-object queries, reinforcing its advantage for low-latency streaming applications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.411, + 0.486, + 0.699 + ], + "angle": 0, + "content": "Verbalization & Interleaving. In Fig. 5, we analyze our verbalization and interleaving strategies using ProVideLLM-1B/5 on Ego4D Goal-Step validation set, selecting a video containing over 15 steps and spanning \\(>30\\) minutes. At 4fps sampling, we ablate on a fixed memory budget of 8GB on an A6000 GPU. Short-term cache length is set as \\(N_{S} = 64\\) (16 seconds), and no limit is imposed on the long-term cache. Using only visual tokens (red), memory scales linearly with video length, reaching 40K tokens to represent 20 minutes of video before exceeding 8GB memory. Such massive context also degrades performance. Verbalization (orange) compresses 20 minutes of video into 2.5K tokens, reducing memory usage to 3GB and achieving sub-linear scaling. However, it introduces runtime instability, with frequent spikes in per-frame prediction time due to attention recomputation when converting short-term visual tokens into long-term text. Interleaving (green) in our multimodal cache mitigates these spikes, improving runtime stability while maintaining sub-linear scaling in both memory and compute." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.712, + 0.265, + 0.728 + ], + "angle": 0, + "content": "4.4. Runtime Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.901 + ], + "angle": 0, + "content": "In Tab. 9, we report the per-frame streaming inference and streaming dialogue [14] performance of ProVideLLM variants on a single A6000 GPU. Our lightest variant, ProVideLLM-1B/5, achieves over 10 FPS with a minimal 2GB memory footprint, making it ideal for low-latency applications. Meanwhile, ProVideLLM-8B/11 offer higher performance but require more compute. For most videos in this dataset, ProVideLLM-1B/5 consistently maintains over 10 FPS, sometimes even reaching 20 FPS when the video has many no-action frames. For per-frame predictions, a key bottleneck for VideoLLMs is their autoregressive text gener" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.184 + ], + "angle": 0, + "content": "
Streaming PredictionsModelFPS (↑)Memory (GB) (↓)
Vision Enc.Connector + LLMFull Model
Per-frameProVideLLM-1B/574.810.49.12.0
ProVideLLM-8B/1138.04.23.516.2
Dialogue [14]ProVideLLM-1B/574.834.224.62.2
ProVideLLM-8B/1138.023.717.216.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.194, + 0.908, + 0.236 + ], + "angle": 0, + "content": "Table 9. Runtime Analysis on a single A6000 GPU for the validation set of Ego4D-GoalStep online step detection for per-frame streaming and Ego4D-GoalStep Narration [83]." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.248, + 0.905, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.349, + 0.907, + 0.391 + ], + "angle": 0, + "content": "Figure 5. Scaling of memory, runtime and context length under various caching strategies. Results reported on a Ego4D-GoalStep video at a memory budget of 8GB on a single A6000 GPU." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.907, + 0.485 + ], + "angle": 0, + "content": "ation, adding around 6 extra iterations, corresponding to 5.7 tokens/(activity step) on average for Ego4D-GoalStep. This issue is mitigated with streaming dialogue [14], where the model is trained to remain silent and respond to user queries only when necessary. In this setting, ProVideLLM-1B/5 delivers real-time performance at 24.6 FPS." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.498, + 0.771, + 0.514 + ], + "angle": 0, + "content": "5. Conclusion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.909, + 0.78 + ], + "angle": 0, + "content": "In this paper, we presented ProVideLLM for real-time understanding of procedural videos by interleaving two types of tokens in a vision-text multimodal cache - verbalized text tokens, which provide compressed summaries of long-term observations, and visual tokens, which capture fine-grained details from short-term observations. We show that ProVideLLM reduces the token count for representing long-form procedural videos while effectively capturing critical spatio-temporal dependencies. We demonstrated state-of-the-art results on six tasks across four datasets, showing per-frame inference at 10 FPS and streaming dialogue at 25 FPS. ProVideLLM is the first real-time framework unifying recognition, anticipation, and planning for procedural assistance. Our multimodal interleaved cache is also generic, applicable to non-procedural videos, and unlocking a new area of compressing long videos on the fly while reducing redundancy and boosting performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.911, + 0.901 + ], + "angle": 0, + "content": "Limitations. DETR-QFormer, trained on pseudo ground-truth bounding boxes, generalizes well but could benefit from GT data to reduce biases from noisy annotations. Additionally ProVideLLM focuses on detection but lacks real-time feedback on the \"how\" of task execution. Finally, as discussed in Sec. 4.2, ProVideLLM's generalization performance remains modest, presenting a potential direction for future work." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22593" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.115, + 0.486, + 0.184 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.187, + 0.484, + 0.243 + ], + "angle": 0, + "content": "[2] Ibrahim M Alabdulmohsin, Xiaohua Zhai, Alexander Kolesnikov, and Lucas Beyer. Getting vit in shape: Scaling laws for compute-optimal model design. Advances in Neural Information Processing Systems, 36, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.245, + 0.484, + 0.313 + ], + "angle": 0, + "content": "[3] Kumar Ashutosh, Santhosh Kumar Ramakrishnan, Triantafyllos Afouras, and Kristen Grauman. Video-mined task graphs for keystep recognition in instructional videos. Advances in Neural Information Processing Systems, 36: 67833-67846, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.316, + 0.484, + 0.383 + ], + "angle": 0, + "content": "[4] Kumar Ashutosh, Santhosh Kumar Ramakrishnan, Triantafyllos Afouras, and Kristen Grauman. Video-mined task graphs for keystep recognition in instructional videos. Advances in Neural Information Processing Systems, 36, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.387, + 0.482, + 0.443 + ], + "angle": 0, + "content": "[5] Fabien Baradel, Natalia Neverova, Christian Wolf, Julien Mille, and Greg Mori. Object level visual reasoning in videos. In Proceedings of the European Conference on Computer Vision (ECCV), pages 105-121, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.445, + 0.484, + 0.486 + ], + "angle": 0, + "content": "[6] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.489, + 0.484, + 0.572 + ], + "angle": 0, + "content": "[7] Yizhak Ben-Shabat, Xin Yu, Fatemeh Saleh, Dylan Campbell, Cristian Rodriguez-Opazo, Hongdong Li, and Stephen Gould. The ikaa asm dataset: Understanding people assembling furniture through actions, objects and pose. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 847–859, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.574, + 0.484, + 0.655 + ], + "angle": 0, + "content": "[8] Yizhak Ben-Shabat, Jonathan Paul, Eviatar Segev, Oren Shrout, and Stephen Gould. Ika ego 3d dataset: Understanding furniture assembly actions from ego-view 3d point clouds. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4355-4364, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.659, + 0.482, + 0.7 + ], + "angle": 0, + "content": "[9] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.703, + 0.484, + 0.771 + ], + "angle": 0, + "content": "[10] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.774, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[11] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[12] Chien-Yi Chang, De-An Huang, Danfei Xu, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. Procedure planning in instructional videos. In European Conference on Computer Vision, pages 334–350. Springer, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.115, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.908, + 0.147 + ], + "angle": 0, + "content": "[13] Dibyadip Chatterjee, Fadime Sener, Shugao Ma, and Angela Yao. Opening the vocabulary of egocentric actions. Advances in Neural Information Processing Systems, 36: 33174-33187, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.151, + 0.908, + 0.234 + ], + "angle": 0, + "content": "[14] Joya Chen, Zhaoyang Lv, Shiwei Wu, Kevin Qinghong Lin, Chenan Song, Difei Gao, Jia-Wei Liu, Ziteng Gao, Dongxing Mao, and Mike Zheng Shou. Videollm-online: Online video large language model for streaming video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18407-18418, 2024. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.236, + 0.908, + 0.317 + ], + "angle": 0, + "content": "[15] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.321, + 0.908, + 0.403 + ], + "angle": 0, + "content": "[16] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision, pages 1-23, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.406, + 0.908, + 0.488 + ], + "angle": 0, + "content": "[17] Roeland De Geest, Efstratios Gavves, Amir Ghodrati, Zhenyang Li, Cees Snoek, and Tinne Tuytelaars. Online action detection. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14, pages 269-284. Springer, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.491, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[18] Guodong Ding, Fadime Sener, Shugao Ma, and Angela Yao. Every mistake counts in assembly. arXiv preprint arXiv:2307.16453, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.534, + 0.908, + 0.588 + ], + "angle": 0, + "content": "[19] Guodong Ding, Fadime Sener, and Angela Yao. Temporal action segmentation: An analysis of modern techniques. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.591, + 0.908, + 0.659 + ], + "angle": 0, + "content": "[20] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.662, + 0.907, + 0.731 + ], + "angle": 0, + "content": "[21] Nikita Dvornik, Isma Hadji, Hai Pham, Dhaivat Bhatt, Brais Martinez, Afsaneh Fazly, and Allan D Jepson. Flow graph to video grounding for weakly-supervised multi-step localization. In European Conference on Computer Vision, pages 319-335. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.734, + 0.908, + 0.829 + ], + "angle": 0, + "content": "[22] Alessandro Flaborea, Guido Maria D'Amely di Melendugno, Leonardo Plini, Luca Scofano, Edoardo De Matteis, Antonio Furnari, Giovanni Maria Farinella, and Fabio Galasso. Prego: online mistake detection in procedural egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18483-18492, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.832, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[23] Antonino Furnari and Giovanni Maria Farinella. What would you expect? anticipating egocentric actions with rolling-unrolling lstms and modality attention. In Proceedings of the IEEE/CVF International conference on computer vision, pages 6252-6261, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22594" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.484, + 0.161 + ], + "angle": 0, + "content": "[24] Antonino Furnari, Sebastiano Battiato, and Giovanni Maria Farinella. Leveraging uncertainty to rethink loss functions and evaluation measures for egocentric action anticipation. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.162, + 0.484, + 0.232 + ], + "angle": 0, + "content": "[25] Mingfei Gao, Mingze Xu, Larry S Davis, Richard Socher, and Caiming Xiong. Startnet: Online detection of action start in untrimmed videos. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5542-5551, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.233, + 0.484, + 0.286 + ], + "angle": 0, + "content": "[26] Rohit Girdhar and Kristen Grauman. Anticipative video transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13505-13515, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.288, + 0.484, + 0.37 + ], + "angle": 0, + "content": "[27] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.372, + 0.484, + 0.468 + ], + "angle": 0, + "content": "[28] Kristen Grauman, Andrew Westbury, Lorenzo Torresani, Kris Kitani, Jitendra Malik, Triantafyllos Afouras, Kumar Ashutosh, Vijay Baiyya, Siddhant Bansal, Bikram Boote, et al. Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19383-19400, 2024. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.469, + 0.484, + 0.524 + ], + "angle": 0, + "content": "[29] Bo He, Hengduo Li, Young Kyun Jang, Menglin Jia, Xuefei Cao, Ashish Shah, Abhinav Shrivastava, and Ser-Nam Lim. Ma-lmm: Memory-augmented large multimodal model for long-term video understanding. CVPR, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.525, + 0.484, + 0.593 + ], + "angle": 0, + "content": "[30] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.595, + 0.484, + 0.663 + ], + "angle": 0, + "content": "[31] Cheng-Ping Hsieh, Simeng Sun, Samuel Kriman, Shantanu Acharya, Dima Rekesh, Fei Jia, Yang Zhang, and Boris Ginsburg. Ruler: What's the real context size of your long-context language models? arXiv preprint arXiv:2404.06654, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.665, + 0.484, + 0.732 + ], + "angle": 0, + "content": "[32] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.734, + 0.484, + 0.803 + ], + "angle": 0, + "content": "[33] Md Mohaiminul Islam, Tushar Nagarajan, Huiyu Wang, Fu-Jen Chu, Kris Kitani, Gedas Bertasius, and Xitong Yang. Propose, assess, search: Harnessing llms for goal-oriented planning in instructional videos. arXiv preprint arXiv:2409.20557, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.804, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[34] Y.-G. Jiang, J. Liu, A. Roshan Zamir, G. Toderici, I. Laptev, M. Shah, and R. Sukthankar. THUMOS challenge: Action recognition with a large number of classes. http://crcv.ucf.edu/THUMOS14/, 2014.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.86, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[35] Mahnaz Koupae and William Yang Wang. Wikihow: A large scale text summarization dataset. arXiv preprint arXiv:1810.09305, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.093, + 0.908, + 0.148 + ], + "angle": 0, + "content": "[36] Bolin Lai, Sam Toyer, Tushar Nagarajan, Rohit Girdhar, Shengxin Zha, James M Rehg, Kris Kitani, Kristen Grauman, Ruta Desai, and Miao Liu. Human action anticipation: A survey. arXiv preprint arXiv:2410.14045, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.15, + 0.908, + 0.218 + ], + "angle": 0, + "content": "[37] Shih-Po Lee, Zijia Lu, Zekun Zhang, Minh Hoai, and Ehsan Elhamifar. Error detection in egocentric procedural task videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18655-18666, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.221, + 0.908, + 0.275 + ], + "angle": 0, + "content": "[38] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.278, + 0.908, + 0.346 + ], + "angle": 0, + "content": "[39] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.349, + 0.908, + 0.403 + ], + "angle": 0, + "content": "[40] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.406, + 0.908, + 0.46 + ], + "angle": 0, + "content": "[41] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.462, + 0.908, + 0.53 + ], + "angle": 0, + "content": "[42] Han Lin, Tushar Nagarajan, Nicolas Ballas, Mido Assran, Mojtaba Komeili, Mohit Bansal, and Koustuv Sinha. Vedit: Latent prediction architecture for procedural video representation learning. arXiv preprint arXiv:2410.03478, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.533, + 0.908, + 0.614 + ], + "angle": 0, + "content": "[43] Xudong Lin, Fabio Petroni, Gedas Bertasius, Marcus Rohrbach, Shih-Fu Chang, and Lorenzo Torresani. Learning to recognize procedural activities with distant supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13853-13863, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.618, + 0.908, + 0.659 + ], + "angle": 0, + "content": "[44] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.661, + 0.908, + 0.716 + ], + "angle": 0, + "content": "[45] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.718, + 0.908, + 0.785 + ], + "angle": 0, + "content": "[46] Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. Lost in the middle: How language models use long contexts. Transactions of the Association for Computational Linguistics, 12: 157-173, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.789, + 0.908, + 0.844 + ], + "angle": 0, + "content": "[47] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.846, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[48] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22595" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.484, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF international conference on computer vision, pages 2630-2640, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.122, + 0.484, + 0.191 + ], + "angle": 0, + "content": "[49] Medhini Narasimhan, Arsha Nagrani, Chen Sun, Michael Rubinstein, Trevor Darrell, Anna Rohrbach, and Cordelia Schmid. Tl; dw? summarizing instructional videos with task relevance and cross-modal saliency. In European Conference on Computer Vision, pages 540-557. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.193, + 0.483, + 0.247 + ], + "angle": 0, + "content": "[50] Medhini Narasimhan, Licheng Yu, Sean Bell, Ning Zhang, and Trevor Darrell. Learning and verification of task structure in instructional videos. arXiv preprint arXiv:2303.13519, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.25, + 0.484, + 0.291 + ], + "angle": 0, + "content": "[51] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.293, + 0.484, + 0.361 + ], + "angle": 0, + "content": "[52] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.364, + 0.484, + 0.405 + ], + "angle": 0, + "content": "[53] Zhanzhong Pang, Fadime Sener, and Angela Yao. Context-enhanced memory-refined transformer for online action detection. arXiv preprint arXiv:2503.18359, 2025. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.407, + 0.484, + 0.488 + ], + "angle": 0, + "content": "[54] Dhruvesh Patel, Hamid Eghbalzadeh, Nitin Kamra, Michael Louis Iuzzolino, Unnat Jain, and Ruta Desai. Pretrained language models as visual planners for human assistance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15302-15314, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.492, + 0.484, + 0.56 + ], + "angle": 0, + "content": "[55] Reiner Pope, Sholto Douglas, Aakanksha Chowdhery, Jacob Devlin, James Bradbury, Jonathan Heek, Kefan Xiao, Shivani Agrawal, and Jeff Dean. Efficiently scaling transformer inference. Proceedings of Machine Learning and Systems, 5:606-624, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.563, + 0.484, + 0.645 + ], + "angle": 0, + "content": "[56] Shraman Pramanick, Yale Song, Sayan Nag, Kevin Qinghong Lin, Hardik Shah, Mike Zheng Shou, Rama Chellappa, and Pengchuan Zhang. Egovlpv2: Egocentric video-language pre-training with fusion in the backbone. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5285-5297, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.648, + 0.484, + 0.674 + ], + "angle": 0, + "content": "[57] Alec Radford. Improving language understanding by generative pre-training. Openai, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.677, + 0.484, + 0.731 + ], + "angle": 0, + "content": "[58] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.734, + 0.484, + 0.816 + ], + "angle": 0, + "content": "[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.818, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[60] Francesco Ragusa, Antonino Furnari, Salvatore Livatino, and Giovanni Maria Farinella. The meccano dataset: Understanding human-object interactions from egocentric videos in an industrial-like domain. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1569-1578, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.093, + 0.908, + 0.146 + ], + "angle": 0, + "content": "[61] Abhinav Rai, Fadime Sener, and Angela Yao. Transformed rois for capturing visual transformations in videos. Computer Vision and Image Understanding, 224:103558, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.15, + 0.908, + 0.219 + ], + "angle": 0, + "content": "[62] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.221, + 0.907, + 0.303 + ], + "angle": 0, + "content": "[63] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.307, + 0.908, + 0.387 + ], + "angle": 0, + "content": "[64] Luigi Seminara, Giovanni Maria Farinella, and Antonino Furnari. Differentiable task graph learning: Procedural activity representation and online mistake detection from egocentric videos. In Advances in Neural Information Processing Systems, pages 59373-59407. Curran Associates, Inc., 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.391, + 0.908, + 0.445 + ], + "angle": 0, + "content": "[65] Fadime Sener and Angela Yao. Zero-shot anticipation for instructional activities. In Proceedings of the IEEE/CVF international conference on computer vision, pages 862-871, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.449, + 0.908, + 0.518 + ], + "angle": 0, + "content": "[66] Fadime Sener, Dipika Singhania, and Angela Yao. Temporal aggregate representations for long-range video understanding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16, pages 154-171. Springer, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.52, + 0.908, + 0.603 + ], + "angle": 0, + "content": "[67] Fadime Sener, Dibyadip Chatterjee, Daniel Shelepov, Kun He, Dipika Singhania, Robert Wang, and Angela Yao. Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21096-21106, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.605, + 0.907, + 0.66 + ], + "angle": 0, + "content": "[68] Md Salman Shamil, Dibyadip Chatterjee, Fadime Sener, Shugao Ma, and Angela Yao. On the utility of 3d hand poses for action recognition. In European Conference on Computer Vision, pages 436-454. Springer, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.662, + 0.908, + 0.731 + ], + "angle": 0, + "content": "[69] Dandan Shan, Jiaqi Geng, Michelle Shu, and David F Fouhey. Understanding human hands in contact at internet scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9869-9878, 2020. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.733, + 0.908, + 0.815 + ], + "angle": 0, + "content": "[70] Zheng Shou, Junting Pan, Jonathan Chan, Kazuyuki Miyazawa, Hassan Mansour, Anthony Vetro, Xavier Giro-i Nieto, and Shih-Fu Chang. Online detection of action start in untrimmed, streaming videos. In Proceedings of the European conference on computer vision (ECCV), pages 534-551, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.818, + 0.908, + 0.9 + ], + "angle": 0, + "content": "[71] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22596" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[72] Yale Song, Eugene Byrne, Tushar Nagarajan, Huiyu Wang, Miguel Martin, and Lorenzo Torresani. Ego4d goal-step: Toward hierarchical understanding of procedural activities. Advances in Neural Information Processing Systems, 36, 2023. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.162, + 0.485, + 0.244 + ], + "angle": 0, + "content": "[73] Reuben Tan, Xineng Sun, Ping Hu, Jui-hsien Wang, Hanieh Deilamsalehy, Bryan A Plummer, Bryan Russell, and Kate Saenko. Koala: Key frame-conditioned long video-llm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13581–13591, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.246, + 0.484, + 0.328 + ], + "angle": 0, + "content": "[74] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1207-1216, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.484, + 0.398 + ], + "angle": 0, + "content": "[75] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.4, + 0.484, + 0.468 + ], + "angle": 0, + "content": "[76] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568-9578, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.469, + 0.484, + 0.537 + ], + "angle": 0, + "content": "[77] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.539, + 0.484, + 0.594 + ], + "angle": 0, + "content": "[78] Huiyu Wang, Mitesh Kumar Singh, and Lorenzo Torresani. Ego-only: Egocentric action detection without exocentric transferring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5250-5261, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.595, + 0.484, + 0.662 + ], + "angle": 0, + "content": "[79] Jiahao Wang, Guo Chen, Yifei Huang, Limin Wang, and Tong Lu. Memory-and-anticipation transformer for online action understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13824–13835, 2023. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.664, + 0.484, + 0.718 + ], + "angle": 0, + "content": "[80] Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video understanding with large language model as agent. arXiv preprint arXiv:2403.10517, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.72, + 0.484, + 0.774 + ], + "angle": 0, + "content": "[81] Yuetian Weng, Mingfei Han, Haoyu He, Xiaojun Chang, and Bohan Zhuang. Longvlm: Efficient long video understanding via large language models. arXiv preprint arXiv:2404.03384, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.776, + 0.484, + 0.803 + ], + "angle": 0, + "content": "[82] Wikipedia. How to do anything. http://www.wikipedia.com/, 2005.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.805, + 0.484, + 0.872 + ], + "angle": 0, + "content": "[83] Shiwei Wu, Joya Chen, Kevin Qinghong Lin, Qimeng Wang, Yan Gao, Qianli Xu, Tong Xu, Yao Hu, Enhong Chen, and Mike Zheng Shou. Videollm-mod: Efficient video-language streaming with mixture-of-depths vision computation. arXiv preprint arXiv:2408.16730, 2024. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.874, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[84] Junbin Xiao, Angela Yao, Yicong Li, and Tat-Seng Chua. Can i trust your answer? visually grounded video question" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.093, + 0.908, + 0.134 + ], + "angle": 0, + "content": "answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13204-13214, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.136, + 0.908, + 0.204 + ], + "angle": 0, + "content": "[85] Junbin Xiao, Nanxin Huang, Hangyu Qin, Dongyang Li, Yicong Li, Fengbin Zhu, Zhulin Tao, Jianxing Yu, Liang Lin, Tat-Seng Chua, et al. Videoqa in the era of llms: An empirical study. International Journal of Computer Vision, pages 1-24, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.205, + 0.908, + 0.272 + ], + "angle": 0, + "content": "[86] Mingze Xu, Yuanjun Xiong, Hao Chen, Xinyu Li, Wei Xia, Zhuowen Tu, and Stefano Soatto. Long short-term transformer for online action detection. Advances in Neural Information Processing Systems, 34:1086-1099, 2021. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.274, + 0.907, + 0.328 + ], + "angle": 0, + "content": "[87] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.33, + 0.907, + 0.383 + ], + "angle": 0, + "content": "[88] Olga Zatsarynna and Juergen Gall. Action anticipation with goal consistency. In 2023 IEEE International Conference on Image Processing (ICIP), pages 1630-1634. IEEE, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.385, + 0.907, + 0.441 + ], + "angle": 0, + "content": "[89] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.442, + 0.907, + 0.496 + ], + "angle": 0, + "content": "[90] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Is an object-centric video representation beneficial for transfer? In Proceedings of the Asian Conference on Computer Vision, pages 1976-1994, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.498, + 0.908, + 0.564 + ], + "angle": 0, + "content": "[91] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Helping hands: An object-aware ego-centric video recognition model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13901-13912, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.567, + 0.906, + 0.636 + ], + "angle": 0, + "content": "[92] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 543–553, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.637, + 0.908, + 0.704 + ], + "angle": 0, + "content": "[93] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.706, + 0.907, + 0.748 + ], + "angle": 0, + "content": "[94] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.749, + 0.908, + 0.83 + ], + "angle": 0, + "content": "[95] He Zhao, Isma Hadji, Nikita Dvornik, Konstantinos G Derpanis, Richard P Wildes, and Allan D Jepson. P3iv: Probabilistic procedure planning from instructional videos with weak supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2938-2948, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.832, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[96] Qi Zhao, Shijie Wang, Ce Zhang, Changcheng Fu, Minh Quan Do, Nakul Agarwal, Kwonjoon Lee, and Chen Sun. AntGPT: Can large language models help long-term action anticipation from videos? In The Twelfth International Conference on Learning Representations, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22597" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[97] Yue Zhao and Philipp Krahenbuhl. Real-time online video detection with temporal smoothing transformers. In European Conference on Computer Vision, pages 485-502. Springer, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.149, + 0.483, + 0.231 + ], + "angle": 0, + "content": "[98] Yue Zhao, Long Zhao, Xingyi Zhou, Jialin Wu, Chun-Te Chu, Hui Miao, Florian Schroff, Hartwig Adam, Ting Liu, Boqing Gong, et al. Distilling vision-language models on millions of videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13106-13116, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.234, + 0.484, + 0.314 + ], + "angle": 0, + "content": "[99] Yiwu Zhong, Licheng Yu, Yang Bai, Shangwen Li, Xueting Yan, and Yin Li. Learning procedure-aware video representation from instructional videos and their narrations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14825-14835, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.318, + 0.484, + 0.399 + ], + "angle": 0, + "content": "[100] Honglu Zhou, Roberto Martin-Martín, Mubbasir Kapadia, Silvio Savarese, and Juan Carlos Niebles. Procedure-aware pretraining for instructional video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10727-10738, 2023. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.402, + 0.482, + 0.457 + ], + "angle": 0, + "content": "[101] Luowei Zhou, Chenliang Xu, and Jason Corso. Towards automatic learning of procedures from web instructional videos. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.459, + 0.483, + 0.514 + ], + "angle": 0, + "content": "[102] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.516, + 0.483, + 0.584 + ], + "angle": 0, + "content": "[103] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramazan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Crosstask weakly supervised learning from instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3537-3545, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.486, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "22598" + } + ] +] \ No newline at end of file diff --git a/2025/Streaming VideoLLMs for Real-Time Procedural Video Understanding/e3ae8cdd-650d-48b8-9247-fe7363ebedec_origin.pdf b/2025/Streaming VideoLLMs for Real-Time Procedural Video Understanding/e3ae8cdd-650d-48b8-9247-fe7363ebedec_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d1f2bf67b36dd6d3463f5d50ac1b85b7f3a87128 --- /dev/null +++ b/2025/Streaming VideoLLMs for Real-Time Procedural Video Understanding/e3ae8cdd-650d-48b8-9247-fe7363ebedec_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:192ca99aed1d3ff745d40c1c39df5397188f5bf0c6bf1dc6a2ce9b377c47fea0 +size 3608857 diff --git a/2025/Streaming VideoLLMs for Real-Time Procedural Video Understanding/full.md b/2025/Streaming VideoLLMs for Real-Time Procedural Video Understanding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c56fe4dc0c4daa5200a1edfc064aa4660ab77350 --- /dev/null +++ b/2025/Streaming VideoLLMs for Real-Time Procedural Video Understanding/full.md @@ -0,0 +1,376 @@ +# Streaming VideoLLMs for Real-time Procedural Video Understanding + +Dibyadip Chatterjee $^{1,3*}$ Edoardo Remelli $^{1}$ Yale Song $^{2}$ Bugra Tekin $^{1}$ Abhay Mittal $^{1}$ Bharat Bhatnagar $^{1}$ Necati Cihan Camgoz $^{1}$ Shreyas Hampali $^{1}$ Eric Sauser $^{1}$ Shugao Ma $^{1}$ Angela Yao $^{3}$ Fadime Sener $^{1}$ + +$^{1}$ Meta Reality Labs $^{2}$ FAIR, Meta $^{3}$ National University of Singapore + +dibschat.github.io/ProVideLLM + +![](images/9109d7903435817505e944fca36d9aec8c09ad0ece6a264152e2040549ee34c6.jpg) +Figure 1. Overview of ProVideLLM. A streaming video large language model for real-time procedural video tasks with a low memory footprint. It features a multimodal interleaved cache that stores textual descriptions of long-term observations (spanning several minutes) and visual tokens representing short-term observations (spanning a few seconds). We also introduce a new DETR-QFormer connector for better fine-grained tokenization of the short-term. ProVideLLM is capable of handling multiple procedural tasks within a single model. + +# Abstract + +We introduce ProVideLLM, an end-to-end framework for real-time procedural video understanding. ProVideLLM integrates a multimodal cache configured to store two types of tokens – verbalized text tokens, which provide compressed textual summaries of long-term observations, and visual tokens, encoded with DETR-QFormer to capture fine-grained details from short-term observations. This design reduces token count by $22 \times$ over existing methods when representing one hour of long-term observations while effectively encoding fine-granularity of the present. By interleaving these tokens in our multimodal cache, ProVideLLM achieves sublinear scaling of memory and compute with video length, ensuring per-frame streaming inference at 10 FPS and streaming dialogue at 25 FPS, with a minimal 2GB GPU memory footprint. ProVideLLM also sets new state-of-the-art results on six procedural tasks across four datasets. + +# 1. Introduction + +Procedural videos, spanning several minutes to hours, capture multi-step tasks [82] such as cooking [65, 101], toy assembly [60, 67], and furniture building [7, 8]. With the growing adoption of AR assistants, streaming video understanding is essential for real-time procedural assistance. This requires real-time, online inference, as frames appear in a live video, without access to future observations. Despite its importance, streaming video understanding remains largely unexplored in existing procedural research [19, 36, 100]. + +Existing works on streaming video understanding [17, 70] focus on online action detection [53, 79, 86, 97] or live conversations [14, 83]. These use-cases target a small set of unrelated actions [17, 34] and do not require long-term temporal modeling. In contrast, procedural videos are inherently long and causal, where each step depends on prior steps. Applying existing approaches to such videos significantly increases token count and memory overhead, leading to long-context degradation [31, 46] and making them unsuitable for streaming, even with excessive frame dropping [29]. The semantic inter-connectivity of activity steps introduces re + +dundancy when long-term observations are stored as raw frame tokens [79, 86]. Furthermore, while short-term observations are crucial for understanding the present, not all frame tokens are equally relevant, since procedural tasks often center around hand-object interactions [28, 67, 72]. + +In this paper, we propose an end-to-end framework for streaming procedural video understanding (Fig. 1), addressing the above mentioned challenges of processing long-form videos. Specifically, we introduce ProVideLLM, a Procedural VideoLLM that comprises a visual (frame) encoder, a vision-language connector, and a pre-trained large language model (LLM) [1, 20, 58, 77] as the decoder. Our key contribution is a vision-language multimodal cache that efficiently manages streaming video data by balancing compression and representation quality of past observations. + +We factorize the streaming past observations into long-term semantic history and short-term visual context. They are encoded as text and visual tokens, respectively, and interleaved in our multimodal cache. First, for the long-term history spanning several minutes, ProVideLLM decoder predicts step labels per frame, groups similar steps into verbalized sequences (e.g. get $\rightarrow$ crack $\rightarrow$ whisk eggs), and stores them in the cache. This semantic compression significantly reduces token count and also enhances current step prediction (e.g. $\rightarrow$ pour mixture). Second, to capture short-term visual context, typically spanning a few seconds, understanding the fine-grained changes across frames is crucial. However, standard VideoLLM encoders [59, 89] face two key challenges in such modeling: (i) temporal collapse, where low temporal variance conflates distinct actions and (ii) background distractors across frames [5, 13, 61, 90]. We mitigate the former by using visual encoders trained with dense spatial alignment objectives [11, 30, 52]. We address the latter with a novel vision-text connector DETR-QFormer, pretrained to explicitly focus on hand-object interaction regions, thereby compressing them into a smaller set of tokens. + +Finally, in existing methods [86], both long- and short-term tokens are visual and stored separately. Applying such separate caching to our multimodal tokens would require redundant attention recomputation when verbalizing short-term visual tokens into long-term text. To overcome this, we interleave text and visual tokens within our multimodal cache. This enables prefilling the cache, which ensures that computation scales linearly with cache size and sub-linearly with the video length, thereby significantly reducing latency for real-time performance. + +In summary, our contributions are: (i) a first-of-its-kind streaming multimodal cache with interleaved text and visual tokens, enabling per-frame inference [72] at 10 FPS and streaming dialogue [14] at 25 FPS; (ii) verbalizing long-term observations into compressed textual summaries and (iii) modeling hand-object interactions for short-term observations using DETR-QFormer; all combined, it achieves a + +$22 \times$ reduction in token count compared to existing methods. (iv) state-of-the-art results on six tasks across four datasets, making ProVideLLM the first real-time framework unifying recognition, anticipation, planning for procedural guidance. + +# 2. Related Works + +Procedural Video Understanding have primarily focused on instructional YouTube videos [48, 74, 101, 103]. Recent large-scale egocentric datasets [27, 28, 67, 72] have advanced research by providing longer videos with more activity steps. Existing works have focused on offline tasks [19, 43, 100], where the entire video is available for prediction, as well as some online tasks, like procedural planning [12, 33, 95], forecasting [23, 26, 49, 65, 66, 96], and mistake detection [18, 22, 37, 67]. However, deploying these models for streaming inference results in high computational complexity where it grows quadratically with the video length. In contrast, ProVideLLM incurs a sub-linear cost per update, both in terms of memory and runtime, enabling efficient streaming inference. + +Video Large Language Models. Multimodal large language models (MLLMs) have been popularized by BLIP-2-like [15, 39, 102] and LLaVA-like [38, 44, 45] models where an LLM is trained to understand vision-text data. Extending these to video have led to the development of Video Large Language Models (VideoLLMs) [40, 41, 47, 83, 94, 98]. However, most of these models are limited to offline tasks [62, 71, 73, 80, 81] where the entire video is queried to answer a question. As such, architectural elements such as VideoQ-Former [92] and temporal pooling layers [40], developed without causal requirements, limit streaming applications. Models like MA-LMM [29] and VideoLLM online [14] address video understanding online, but are too computationally intensive to be deployed for efficient streaming inference on long-form videos. ProVideLLM advances these models by introducing a multimodal interleaved cache for faster inference and proposing DETR-QFormer for modeling fine-grained visual changes, improving accuracy. + +Streaming Video Understanding has been explored for online action detection [17, 25, 70] where dominant approaches [53, 79, 86, 97] split streaming input into long- and short-term visual histories within an encoder-decoder architecture. However, storing both as visual tokens is redundant and computationally prohibitive for long-form procedural videos, where multiple actions are temporally connected. Instead, ProVideLLM verbalizes long-term observations into text which drastically decreases token count required to represent hour long videos, thus improving memory efficiency. + +# 3. ProVideLLM + +ProVideLLM encodes long- and short-term streaming observations into text and visual tokens and interleaves them in a multimodal cache. We first present the preliminaries + +in Sec. 3.1, followed by long- and short-term token generation in Sec. 3.2 and Sec. 3.3. Finally, we present our multimodal cache in Sec. 3.4 and training objectives in Sec. 3.5. + +# 3.1. Preliminaries + +ProVideLLM follows a LLaVA-style [44] architecture, consisting of a visual encoder $(E_V)$ , a language decoder $(D_L)$ and a vision-language connector $(C_{VL})$ . Given a video $V = \{v_1, v_2, \ldots, v_T\}$ , the model's output $(o_t)$ at any time step $t \in [1, T]$ is + +$$ +o _ {t} = \operatorname {P r o V i d e L L M} \left(v _ {1: t}\right) = D _ {L} \left(\left\| _ {i = 1} ^ {t} C _ {V L} \left(E _ {V} \left(v _ {i}\right)\right)\right) \quad (1) \right. +$$ + +where $\parallel$ denotes concatenation. While LSTR-inspired long short-term memory approaches [79, 97] also use an encoder-decoder architecture, they rely exclusively on visual tokens (Fig. 2a.1). In contrast, ProVideLLM incorporates a language decoder, which enables more efficient long-term compression, improves generalization, and supports multiple tasks within a single model. + +# 3.2. Tokenizing long-term observations + +Procedural videos are inherently long-form, ranging from tens of minutes [67] to several hours [28, 72]. While modern LLMs [20, 75, 87] offer massive context lengths and, in theory, can process an hour-long raw video at 24 fps, performance degrades with larger context sizes [31, 46], and training such models is highly memory-intensive [29, 93]. Accurately recognizing ongoing steps require effectively encoding long-term context. Our key insight is that procedural steps in a video are temporally connected, i.e. a sequence of step labels can serve as a strong predictor of ongoing steps. For instance, when "preparing an omelet", the step "pour mixture into pan" depends on earlier steps "get eggs from fridge, crack eggs into bowl, whisk eggs". Prior works on flow graphs [21, 64] and planning [3, 54, 100] also show that step sequences, often extracted from external sources [35], serve as strong baselines for procedural understanding. + +Instead of constructing computationally expensive graphs, we propose online verbalization (Fig. 2a.2). Given a streaming video, ProVideLLM decoder continuously generates action predictions (summaries) of the present observation on the fly, and groups semantically similar actions to verbalize the past (Fig. 1). This approach efficiently compresses past visual frame tokens into information-dense language tokens. For example, in Ego4D Goal-Step [72], storing one hour of long-term past requires only 630 verbalized text tokens on average. Compared to long short-term approaches [79, 86, 97], our verbalization reduces token count by $22 \times$ at the same sampling rate, significantly reducing the memory overhead. + +# 3.3. Tokenizing short-term observations + +Our goal is to recognize procedural activity steps in a streaming manner. This requires understanding "what is happening" in the present while tracking "what has happened in + +![](images/549ec69cc1878c662a86f32b066748037c4486fa185624f420f11daa967e6b21.jpg) + +![](images/0a552e73e5bf57d1c56da3a067243342e80adc27a808b023a7c3c5b2f0132174.jpg) +Figure 2. Converting short-term tokens to long-term under various caching types. (a) Factorizing the streaming observations into long-term and short-term and ordering them progressively. (a.1) Conversion done at $O(1)$ cost but is unable to represent long-form videos due to massive token count and memory needs. (a.2) Online verbalization can represent long-form videos at a manageable memory but is not suitable for streaming due to $O(N_S^2 + N_S N_L)$ conversion cost. (b) Interleaving them reduces the conversion cost to $O(N)$ , allowing streaming inference on long-form videos. + +the past". Language is a strong indicator of the semantic structure of step sequences, but determining if and when the step is performed requires fine-grained visual information. Therefore, we allocate more compute for understanding the short-term past, typically spanning 8-16 seconds, by keeping its visual tokens uncompressed. Standard VideoLLM encoders [38, 44] face two key challenges in capturing fine-grained short-term observations: $(i)$ they generate temporally collapsed video features, a limitation we identify and address in Sec. 3.3.1, and $(ii)$ they get biased towards high-level scene and background distractors, which we address in Sec. 3.3.2. + +# 3.3.1. Choosing the visual encoder + +The predominant visual encoders for MLLMs are CLIP [59] and SigLIP [2, 89]. These encoders are pre-trained with only a language-alignment objective for integration with LLMs. However, because they are trained on high-level and sparse web-scale captions such as "cooking in the kitchen", they struggle with fine-grained reasoning, particularly in answering the "what?" and "how to?" aspects of a video [84, 85]. Consequently, distinct images may yield high similarity scores in their feature space [76], leading to low feature diversity across frames within a step. This results in relatively static frame features, a phenomenon we term temporal collapse. For certain fine-grained procedural steps, such as "pick up" vs. "put down", temporally collapsed + +![](images/04a5d58bb747d1f827032af5c311f84394c786979c079a378006d3e83aee0eea.jpg) +a) +b) +Figure 3. a) Per-class mean temporal variance on EgoExo4D & Ego4D Goal-Step. Highly temporally varying classes on the left and low-variance classes on the right. b) Visualization of the top-3 PCA components of patch tokens computed across video frames, thresholded by first component. The primary factors of variation for DINOv2 [52] are hands, and are unfocused and scattered across the image for language-aligned encoders, CLIP [59] and SigLIP [89]. + +features make it nearly impossible for the downstream LLM to decode the correct step. + +Fig. 3a shows the impact of temporal collapse across three visual encoders – two language-aligned ones [59, 89] and one self-supervised1 [52]. We plot temporal variance per class in descending order, with high-variance classes on the left and low-variance classes on the right (see Suppl.Sec.1 for temporal variance computation). At both extremes, the trends are consistent across all encoders. High-variance classes, which often involve verbs such as “get”, “fetch” or “put” correspond to longer videos with substantial scene and environmental changes. Conversely, low-variance classes involving adverbs such as “slowly” or “boil” or repetitive actions like “rotate” or “peel”, exhibit gradual state changes. Among the three models, DINOv2 consistently captures temporal variations more effectively, covering a broader range of temporal variance across different classes. Hence, we choose DINOv2 as our visual encoder. + +# 3.3.2. Aligning hand-object transformations with text + +Procedural datasets predominantly capture hand-centric actions. For instance, nearly $99\%$ of actions in [28, 72] involve hand interactions2. Moreover, video frames often contain distracting background objects, while hand locations serve as strong cues for recognizing the ongoing actions [68, 91]. Thus, a vision encoder focused on hand regions can enhance performance by mitigating background distractions and capturing fine-grained transformations more precisely. + +Another contribution of our work is a vision-language connector, DETR-QFormer, pretrained to attend to salient + +![](images/08271fe2ace3d397040a590658c1cce39a5bd4e7e03de589699a49401f6a6c19.jpg) +Figure 4. DETR-QFormer architecture & Stage-1 pre-training. Trained to detect hands and objects-in-contact from pseudo ground truths bboxes generated by [69], DETR-QFormer queries DINOv2 patch tokens to extract hand-object-focused activations. + +objects. In procedural videos, these are typically in contact with the hand. Interestingly, our choice of DINOv2 as the visual encoder provides an advantage as its patch tokens capture hands as primary factors of variation Fig. 3b. Leveraging this, we design DETR-QFormer to query DINOv2 patch tokens, extracting hand-object activations and suppressing background distractions. DETR-QFormer explicitly attends to hand-object activations using a DETR loss [10] and compresses salient regions into fewer tokens via QFormer's [39] cross-attention, enhancing representation quality. + +Architecture. Our connector integrates Q-Former's compression [39] and DETR's bbox loss into a single crossattention block. DETR-QFormer, illustrated in Fig. 4, is pretrained to detect the hands and objects-in-contact (Sec. 3.5). The connector $(C_{VL})$ is a cross-attention transformer decoder with a set of $m$ learnable visual queries $Q_{v} = (q_{v_{1}}, q_{v_{2}}, \ldots, q_{v_{m}})$ . Key-value pairs $(K, V)$ are generated from the flattened $n^{2}$ DINOv2 patch tokens using two projection heads $(W_{K}, W_{V})$ respectively. To explicitly model hand-object interactions, we append two sets of special queries - two hand queries $Q_{h} = (q_{h_{1}}, q_{h_{2}})$ for the left and right hands, and $k$ object queries $Q_{o} = (q_{o_{1}}, q_{o_{2}}, \ldots, q_{o_{k}})$ corresponding to objects-in-contact. The role of visual queries $Q_{v}$ is to compress the $n^{2} \times D$ patch tokens into $m \times d$ tokens, where $m << n^{2}$ and $d << D$ . Whereas, outputs corresponding to the learnable hand-object queries $(Q_{h}^{\prime}, Q_{o}^{\prime})$ are passed through an MLP to make bounding box predictions, $\hat{B} = \{\hat{b}_{h_{1}}, \hat{b}_{h_{2}}, \hat{b}_{o_{1}}, \ldots, \hat{b}_{o_{k}}\}$ . + +# 3.4. Multimodal Interleaved Cache + +The effectiveness of a procedural video model is best assessed when deployed for real-time inference. Naive long-and short-term approaches [79, 86] maintain separate caches + +Algorithm 1 Multimodal Interleaved Cache +1: class MultimodalInterleavedCache: +2: def __init__(self, N.S, N.L): +3: self_tokens = [] +4: self.N.S = N.S +5: self.N.L = N.L +6: # entry function: adds a new token to cache +7: def entry(self, token): +8: self_tokens.append(token) +9: # exit function for short-term tokens +10: def exit_short(self): +11: if self_tokens.count() > self.N.S: +12: self_tokens.pop(self_tokens.index()) +13: return True +14: # exit function for long-term tokens +15: def exit_long(self): +16: if self_tokens.count() > self.N.L: +17: self_tokens.pop(self_tokens.index() + +for the two token types, arranged progressively with a single entry point (see Fig. 2a). With multimodal tokens, such setups require two distinct entry points, forcing attention recomputation over the short-term cache whenever a short-term token is verbalized into a long-term one. In the worst case, this leads to quadratic scaling of compute. To efficiently leverage online verbalization, we propose a multimodal interleaved cache that interleaves long-term text and short-term visual tokens from Sec. 3.2 and Sec. 3.3. + +Interleaving long- and short-term tokens. We store the running short-term visual tokens and long-term text tokens in a single multimodal cache operating as a FIFO (First In, First Out) queue. It has one entry point and two distinct exits for short- and long-term tokens. Our entry and exit functions are detailed in Algorithm 1, where $N_{S}$ denotes the short-term cache length (in number of frames) and $N_{L}$ the long-term cache length (in number of observed steps). As frame tokens $(v_{i})$ exit the short-term span, the ProVideLLM decoder verbalizes the oldest token $(v_{-N_s})$ . The verbalized text is then added to the cache unless already cached within the past $\tau$ frames. To distinguish the two token types within the same cache, we prepend a special token (represented as $<\mathbb{L}>$ ) to the long-term text tokens. ProVideLLM is instruction-tuned to recognize tokens associated with $<\mathbb{L}>$ as belonging to the long-term observations (Sec. 3.5). Since both long- and short-term tokens share a single entry point, visual and textual tokens are interleaved in the cache at any given time. + +Streaming inference. Every time a new token is added to the cache (Algorithm 2), instead of re-computing attention across all tokens, we use the causal decoder's K-V cache [55] to compute attention only for the new token over the previous $N - 1$ tokens. This ensures that the computational complexity for processing a new frame scales linearly with cache size $(O(N))$ . Additionally, maintaining a single interleaved cache enables direct application of optimized attention implementations [6, 81], though such techniques remain orthogonal to our approach. More importantly, for a fixed $N_S$ , compression via verbalization ensures that $N_L$ + +Algorithm 2 Streaming Inference +1: def inference(V, N_S, N_L, $\tau$ ): +2: cache = MultimodalInterleavedCache(N_S, N_L, $\tau$ ) +3: out = [] +4: for $v_i$ in V: +5: cache_entry $(v_i)$ # add frame tokens to cache +6: $t_i = \text{ProVideLLM}(\text{cache.tokens})$ # current prediction +7: out.append $(t_i)$ # buffer predictions +8: flag = cache_exit_short() # short-term overflow +9: if flag and out[-N_S] not in out[- $\tau$ ): +10: cache_entry(out[-N_S]) # verbalize & interleave +11: cache_exit_long() # long-term overflow + +and consequently $N = (N_{L} + N_{S})$ scales sub-linearly with the number of video frames. Hence, compute also scales sublinearly with the number of streaming video frames. This makes our model well-suited for low-latency applications. + +# 3.5. Training ProVideLLM + +We adopt a two-stage training approach similar to LLaVA - (1) alignment pre-training to align the encoder's visual tokens to the decoder's language space, and (2) instruction tuning of the model for specific tasks. + +Stage-1: Alignment Pre-training. In this stage, our connector DETR-QFormer $(C_{VL})$ is trained while keeping the visual encoder and language decoder frozen (Fig. 4). The objective is to caption short video clips featuring hand-centric actions using the standard language modeling loss [57]: + +$$ +\mathcal {L} _ {\mathrm {L M}} = - \frac {1}{| W |} \sum_ {i = 1} ^ {| W |} \log P (w _ {i} \mid w _ {< i}, V; C _ {V L} ^ {\Theta}), \tag {2} +$$ + +where $V$ is the input video clip, $W = \{w_{1}, w_{2}, \dots, w_{|C|}\}$ is the ground-truth caption, and $C_{VL}^{\Theta}$ is our DETR-QFormer parameterized by $\Theta$ . For detecting hands and corresponding objects-in-contact, we use a DETR [10]-style Generalized IoU [63] loss: + +$$ +\mathcal {L} _ {\mathrm {H O}} \left(b _ {i}, \hat {b} _ {\sigma (i)}\right) = \mathcal {L} _ {\mathrm {i o u}} \left(b _ {i}, \hat {b} _ {\sigma (i)}\right) + \left\| b _ {i} - \hat {b} _ {\sigma (i)} \right\| _ {1} \tag {3} +$$ + +where $b_{i}$ 's are the pseudo ground-truth bounding boxes of hands and objects-in-contact generated by an off-the-shelf detector [69] and $\hat{b}_{\sigma(i)} \in \hat{B}$ is the bounding box matched from the predicted set $\hat{B}$ . For simplicity, we keep our loss category-agnostic. The final loss for Stage-1 is thus: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {L M}} + \lambda_ {1} \mathcal {L} _ {\mathrm {H O}} \tag {4} +$$ + +where $\lambda_{1}$ is a hyper-parameter. In addition to vision-language alignment, Stage-1 also enables our model to attend to hand-object interactions. + +Stage-2: Instruction Tuning. This stage focuses on fine-tuning our model with task-specific instructions such as "Provide the goal.", "Provide what I am doing now.", "Provide what I need to do next.", or "Provide the + +
Modelper-frame mAP
ValTest
LSTR [86] (only short-term)8.8-
LSTR [86]8.98.1
EgoOnly [78]10.310.9
ProVideLLM-1B/5 (no verbalization)12.112.2
ProVideLLM-1B/5 (verbalization & interleaving)13.012.9
+ +Table 1. Ego4D [72] GoalStep Online Step Detection. For fair comparison with LSTR, we consider 16s of short-term past and an avg. of 128s of long-term past compressed into language tokens. + +
ModelViewsAction
TempAgg [66]12 views8.5
+ Goal [88]12 views12.0
ProVideLLM-8B/11 (+long-term verbalization)v413.8
+ +next 10 steps I need to do." For instruction tuning, we fine-tune DETR-QFormer and the language decoder using $\mathcal{L}_{\mathrm{LM}}$ while keeping the visual encoder frozen. For streaming tasks, we also train our model with interleaved special $<\mathbb{L}>$ tokens to represent long-term cache entries, distinguishing them from standard prompts. See Suppl.Sec.5 for further details. + +# 4. Experiments + +We conduct extensive evaluations of ProVideLLM on six tasks across four datasets covering streaming (Sec. 4.1.1), long-term (Sec. 4.1.2) and fine-grained (Sec. 4.1.3) procedural understanding. We further assess ProVideLLM's performance in multi-task settings and its generalization to new datasets in Sec. 4.2. Finally, we provide ablations in Sec. 4.3 and ProVideLLM's runtime analysis in Sec. 4.4. + +Implementation Details. Depending on the choice of vision encoder and language decoder, we have two model variants. ProVideLLM-1B/5 is our efficiency-focused model designed for real-time streaming. It uses DINOv2 (ViT-S) with Llama-3.2-1B-Instruct [20] and requires 5 tokens per frame. ProVideLLM-8B/11 is our more performance-focused model using DINOv2 (ViT-g) and Llama-3.1-8B-Instruct and requires 11 tokens per frame. Furthermore, for benchmarking on youtube videos [74], we introduce an additional model ProVideLLM-8B/11+, which uses SigLIP's [CLS] token along with DINOv2 patch tokens. Unless otherwise mentioned we use $N_{S} = 64$ , $N_{L} = 5$ i.e. storing a maximum of 64 frames in the short-term cache and a maximum of 5 observed steps in the long-term cache. DETR-QFormer is pre-trained (Stage-1) on EPIC-KITCHENS [16] with pseudo ground-truth bounding boxes from 100DOH [69]. For Stage-2, the LLM (decoder) is fine-tuned using LoRA [32] with $r = 128$ and $\alpha = 256$ . Additional implementation details are provided in Suppl.Sec.5. + +Table 2. Action Anticipation Results on Assembly101 validation set. Performance measured by class-mean Top-5 recall. + +
MethodTop-1 Acc. (%)
TimeSformer [9]34.0
DistantSup39.4
VideoTF42.4
ProcedureVRL46.8
VEDIT [42]51.8
VideoLLM-online [14]49.1
VideoLLM-MoD [83]49.7
ProVideoLLM-8B/11+52.1
ProVideoLLM-8B/11+ (+long-term verbalization)53.6
+ +Table 3. Next Step Forecasting in COIN. The last four rows including ProVideLLM are VideoLLMs. + +Datasets. We evaluate our model's performance on four large-scale procedural datasets - EgoExo4D [28], Ego4D-GoalStep [72], COIN [74], and Assembly101 [67]. As shown in Sec. 4.1, ProVideLLM sets state-of-the-art on both egocentric and YouTube benchmarks, demonstrating robustness across diverse video domains. + +# 4.1. Main results + +# 4.1.1. Online Step Detection + +We report results in Tab. 1, using per-frame mAP on action classes (steps + substeps) at 4fps, a standard metric in online action detection [79, 97]. We evaluate ProVideLLM-1B/5 with and without long-term verbalization in rows four and five, respectively. Results for LSTR [86] and EgoOnly [78] are taken from [72]. Accordingly, LSTR spans $128s$ in long-term memory and $16s$ in short-term memory. In [72], each step spans on average $32s$ with around 5.7 tokens representing each step. Our long-term cache can encode $128s$ of past using only 22 tokens at $N_{L} = 4$ steps. To measure improvements from our long-term design, we set the cache length $N = 86$ (spanning LSTR's temporal context), split into $(N_{S} = 86, N_{L} = 0)$ and $(N_{S} = 64, N_{L} = 4)$ in rows four and five, respectively. Our model surpasses both baselines, outperforming LSTR by over $4\%$ at similar temporal span. ProVideLLM also achieves greater gains from long-term context compared to LSTR, highlighting the effectiveness of verbalizing the long-term past. + +# 4.1.2. Online Step Forecasting + +Step forecasting requires modeling long-term dependencies, making it an ideal task to evaluate our contributions. We evaluate ProVideoLLM for step forecasting on COIN and Assembly101 measured by Top-1 accuracy and Top-5 Recall [24], respectively. Our model achieves state-of-the-art performance on both datasets as shown in Tab. 3 and Tab. 2. On Assembly101, we surpass [88] by $1.8\%$ using only a single view. On COIN, we use Llama-3-8B-Instruct to ensure fair comparisons with VideoLLMs [14, 83] and achieve a performance boost of around $4\%$ . Moreover, while state-of-the-art models [14, 83] use approximately 600 to 6000 visual + +
ModelTraining ViewsEgo Acc. (%) Val / Test
TimeSFformer [9]ego35.13 / 35.93
EgoVLPv2 [56]ego+exo39.10 / 38.76
Viewpoint Distillationego+exo38.19 / 39.49
View Invariant Encoder [51]ego+exo40.34 / 41.53
VideoLLM-MoD† [83]ego42.62 / —
ProVideLLM-8B/11ego44.36 / 50.74
+ +Table 4. EgoExo4D Fine-grained Keystep Recognition. Test results obtained from EgoExo4D's Fine-grained Keystep Recognition Challenge on EvalAI. † Uses 2fps sampling per segment. + +
MethodTop-1 Accuracy (%)
TimeSformer [9]34.0
Paprika [100]51.0
DistantSup [43]54.1
ProcedureVRL [99]56.9
VideoTaskGraph [4]57.2
VEDIT [42]62.7
VideoLLM-online [14]63.1
VideoLLM-MoD [83]63.4
ProVideLLM-8B/11+67.3
+ +tokens to model past context, our approach spans a larger temporal context with only 200 tokens through verbalization. + +# 4.1.3. Step Recognition + +In Tab. 4 and Tab. 5, we evaluate ProVideLLM on fine-grained step recognition on EgoExo4D and COIN. We uniformly sample 16 frames from each segment to construct our short-term cache $(N_S)$ and use the prediction from the last frame as our output. The closest model, VideoLLM-MoD [83], uses 10 tokens per frame, but densely samples frames at 2fps, resulting in a larger token count. In both cases, we outperform VideoLLM-MoD by $1.74\%$ and $3.9\%$ respectively. Moreover, at the time of our submission, our method, using only egocentric video, achieved the best results on the EgoExo4D Fine-grained Recognition Challenge, with a $9.2\%$ improvement over the second-ranked model. + +# 4.2. Generalized Procedural Understanding + +In this section, we explore additional capabilities of ProVideLLM, leveraging the advantages of using a pretrained LLM as the language decoder. This section provides insights into developing a "generalist" model that advance procedural video understanding within a unified framework. We first evaluate ProVideLLM's ability to multi-task and then present our benchmark for cross dataset generalization. Multi-task learning. In Tab. 6, we present our results for multi-task learning, an underexplored aspect of procedural understanding. Existing models are trained only for isolated tasks, such as step recognition and forecasting. We compare + +Table 5. COIN Step Recognition. Last three rows are VideoLLMs. + +
ModelCOINBenchmark Top-1 Accuracy (%)
StepTaskNextProc.Proc.+
VideoTF† [50]56.591.042.440.246.4
VideoLLM-online [14]63.192.749.149.854.1
VideoLLM-MoD [83]63.492.849.749.853.3
ProVideLLM-8B/11+66.995.050.551.055.9
+ +Table 6. Multi-task learning on COIN. Step - step recognition, Task - task recognition, Next - next step forecasting, Proc. - long-term forecasting and $\mathrm{Proc + }$ - long-term forecasting with task. $^{\dagger}$ VideoTF [50] is trained separately for each task. + +
ModelEgo Acc. (%)
VideoLLaVA [41] + Llama-3.1-8B [20]3.6
LLaVA-OneVision [38] + Llama-3.1-8B [20]8.3
ProVideLLM-8B/1112.4
+ +Table 7. Cross-dataset generalization. Evaluated on EgoExo4D KeyStep recognition. Our model, trained on Ego4D Goal-Step is compared with a mixture of vision + language experts. + +our model with recent VideoLLMs, also trained in a multitask setting, on 5 tasks. Despite using the same LLM as [14, 83], our model outperforms these baselines across all tasks. Cross-Dataset Generalization. In Tab. 7, we present a cross-dataset generalization benchmark for procedural videos using the EgoExo4D fine-grained keystep recognition task. We compare two recent VideoLLMs [38, 41] which were primarily trained for video captioning and use Llama-3.1-8B-Instruct to convert the captions into EgoExo4D classes. Our baseline models act as a combination of visual and language experts. ProVideLLM, using the same LLM and trained on the much smaller Ego4D Goal-Step dataset, significantly outperforms these baselines. Despite these advances, performance remains modest, highlighting the need for further research in procedural generalization. Additional details are provided in the Suppl. Material. + +# 4.3. Ablations + +Visual Encoders & DETR-QFormer. In Tab. 8, we ablate visual encoders-SigLIP and DINOv2-and connectorsMLP, QFormer, and our DETR-QFormer-on EgoExo4D keystep recognition. First, comparing encoders across connectors, SigLIP's CLS tokens, being language-aligned, outperform DINOv2; however, with more patch tokens, DINOv2 surpasses SigLIP, aligning with our observations in Fig. 3. Second, comparing connectors across encoders, Q-Former boosts DINOv2's performance $(+1.2\%, +1.9\%)$ but has minimal impact on SigLIP $(+0.8\%, -0.1\%)$ , highlighting DINOv2's need for a stronger connector. Finally, DETR-QFormer further improves DINOv2 $(+1.1\%, +2.6\%)$ , with gains increasing at higher resolutions but degrading SigLIP's performance due to weak hand-object activations in its patch tokens. Notably, DETR-QFormer remains highly + +
[CLS] TokenPatch Tokens#Tokens /frameEgo Val./ Acc. (%)
SigLIPDINOv2
w/ 2-layer MLP [45]
✓X131.8±0.428.7±0.2
✓16×16 → 2×2532.1±0.632.4±0.5
✓16×16 → 4×41733.3±0.636.6±0.5
w/ QFormer [39]
✓16×16 ⇒ 4532.9±0.633.6±0.5
✓16×16 ⇒ 161733.2±0.438.1±0.7
w/ DETR-QFormer (2 hands, 2 objects)
✓16×16 ⇒ 0530.9±0.334.7±0.5
✓16×16 ⇒ 121733.6±0.340.7±0.6
+ +token-efficient with DINOv2, even when compressing to only hand-object queries, reinforcing its advantage for low-latency streaming applications. + +Verbalization & Interleaving. In Fig. 5, we analyze our verbalization and interleaving strategies using ProVideLLM-1B/5 on Ego4D Goal-Step validation set, selecting a video containing over 15 steps and spanning $>30$ minutes. At 4fps sampling, we ablate on a fixed memory budget of 8GB on an A6000 GPU. Short-term cache length is set as $N_{S} = 64$ (16 seconds), and no limit is imposed on the long-term cache. Using only visual tokens (red), memory scales linearly with video length, reaching 40K tokens to represent 20 minutes of video before exceeding 8GB memory. Such massive context also degrades performance. Verbalization (orange) compresses 20 minutes of video into 2.5K tokens, reducing memory usage to 3GB and achieving sub-linear scaling. However, it introduces runtime instability, with frequent spikes in per-frame prediction time due to attention recomputation when converting short-term visual tokens into long-term text. Interleaving (green) in our multimodal cache mitigates these spikes, improving runtime stability while maintaining sub-linear scaling in both memory and compute. + +# 4.4. Runtime Analysis + +In Tab. 9, we report the per-frame streaming inference and streaming dialogue [14] performance of ProVideLLM variants on a single A6000 GPU. Our lightest variant, ProVideLLM-1B/5, achieves over 10 FPS with a minimal 2GB memory footprint, making it ideal for low-latency applications. Meanwhile, ProVideLLM-8B/11 offer higher performance but require more compute. For most videos in this dataset, ProVideLLM-1B/5 consistently maintains over 10 FPS, sometimes even reaching 20 FPS when the video has many no-action frames. For per-frame predictions, a key bottleneck for VideoLLMs is their autoregressive text gener + +Table 8. DETR-QFormer ablations on EgoExo4D fine-grained keysteps. Llama-3.2-1B-Instruct is used as the language decoder. Results reported over 3 runs. $(\rightarrow$ means avg-pooling, $\Rightarrow$ means Q-Former compression.) + +
Streaming PredictionsModelFPS (↑)Memory (GB) (↓)
Vision Enc.Connector + LLMFull Model
Per-frameProVideLLM-1B/574.810.49.12.0
ProVideLLM-8B/1138.04.23.516.2
Dialogue [14]ProVideLLM-1B/574.834.224.62.2
ProVideLLM-8B/1138.023.717.216.9
+ +Table 9. Runtime Analysis on a single A6000 GPU for the validation set of Ego4D-GoalStep online step detection for per-frame streaming and Ego4D-GoalStep Narration [83]. + +![](images/e84445eb7c0fe0969400e2c46db1304bd111e84bd51298edacaa12083b12bf59.jpg) +Figure 5. Scaling of memory, runtime and context length under various caching strategies. Results reported on a Ego4D-GoalStep video at a memory budget of 8GB on a single A6000 GPU. + +ation, adding around 6 extra iterations, corresponding to 5.7 tokens/(activity step) on average for Ego4D-GoalStep. This issue is mitigated with streaming dialogue [14], where the model is trained to remain silent and respond to user queries only when necessary. In this setting, ProVideLLM-1B/5 delivers real-time performance at 24.6 FPS. + +# 5. Conclusion and Limitations + +In this paper, we presented ProVideLLM for real-time understanding of procedural videos by interleaving two types of tokens in a vision-text multimodal cache - verbalized text tokens, which provide compressed summaries of long-term observations, and visual tokens, which capture fine-grained details from short-term observations. We show that ProVideLLM reduces the token count for representing long-form procedural videos while effectively capturing critical spatio-temporal dependencies. We demonstrated state-of-the-art results on six tasks across four datasets, showing per-frame inference at 10 FPS and streaming dialogue at 25 FPS. ProVideLLM is the first real-time framework unifying recognition, anticipation, and planning for procedural assistance. Our multimodal interleaved cache is also generic, applicable to non-procedural videos, and unlocking a new area of compressing long videos on the fly while reducing redundancy and boosting performance. + +Limitations. DETR-QFormer, trained on pseudo ground-truth bounding boxes, generalizes well but could benefit from GT data to reduce biases from noisy annotations. Additionally ProVideLLM focuses on detection but lacks real-time feedback on the "how" of task execution. Finally, as discussed in Sec. 4.2, ProVideLLM's generalization performance remains modest, presenting a potential direction for future work. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2 +[2] Ibrahim M Alabdulmohsin, Xiaohua Zhai, Alexander Kolesnikov, and Lucas Beyer. Getting vit in shape: Scaling laws for compute-optimal model design. Advances in Neural Information Processing Systems, 36, 2023. 3 +[3] Kumar Ashutosh, Santhosh Kumar Ramakrishnan, Triantafyllos Afouras, and Kristen Grauman. Video-mined task graphs for keystep recognition in instructional videos. Advances in Neural Information Processing Systems, 36: 67833-67846, 2023. 3 +[4] Kumar Ashutosh, Santhosh Kumar Ramakrishnan, Triantafyllos Afouras, and Kristen Grauman. Video-mined task graphs for keystep recognition in instructional videos. Advances in Neural Information Processing Systems, 36, 2024. 7 +[5] Fabien Baradel, Natalia Neverova, Christian Wolf, Julien Mille, and Greg Mori. Object level visual reasoning in videos. In Proceedings of the European Conference on Computer Vision (ECCV), pages 105-121, 2018. 2 +[6] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020. 5 +[7] Yizhak Ben-Shabat, Xin Yu, Fatemeh Saleh, Dylan Campbell, Cristian Rodriguez-Opazo, Hongdong Li, and Stephen Gould. The ikaa asm dataset: Understanding people assembling furniture through actions, objects and pose. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 847–859, 2021. 1 +[8] Yizhak Ben-Shabat, Jonathan Paul, Eviatar Segev, Oren Shrout, and Stephen Gould. Ika ego 3d dataset: Understanding furniture assembly actions from ego-view 3d point clouds. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4355-4364, 2024. 1 +[9] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. 6, 7 +[10] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 4, 5 +[11] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2 +[12] Chien-Yi Chang, De-An Huang, Danfei Xu, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. Procedure planning in instructional videos. In European Conference on Computer Vision, pages 334–350. Springer, 2020. 2 + +[13] Dibyadip Chatterjee, Fadime Sener, Shugao Ma, and Angela Yao. Opening the vocabulary of egocentric actions. Advances in Neural Information Processing Systems, 36: 33174-33187, 2023. 2 +[14] Joya Chen, Zhaoyang Lv, Shiwei Wu, Kevin Qinghong Lin, Chenan Song, Difei Gao, Jia-Wei Liu, Ziteng Gao, Dongxing Mao, and Mike Zheng Shou. Videollm-online: Online video large language model for streaming video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18407-18418, 2024. 1, 2, 6, 7, 8 +[15] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2 +[16] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision, pages 1-23, 2022. 6 +[17] Roeland De Geest, Efstratios Gavves, Amir Ghodrati, Zhenyang Li, Cees Snoek, and Tinne Tuytelaars. Online action detection. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14, pages 269-284. Springer, 2016. 1, 2 +[18] Guodong Ding, Fadime Sener, Shugao Ma, and Angela Yao. Every mistake counts in assembly. arXiv preprint arXiv:2307.16453, 2023. 2 +[19] Guodong Ding, Fadime Sener, and Angela Yao. Temporal action segmentation: An analysis of modern techniques. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1, 2 +[20] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2, 3, 6, 7 +[21] Nikita Dvornik, Isma Hadji, Hai Pham, Dhaivat Bhatt, Brais Martinez, Afsaneh Fazly, and Allan D Jepson. Flow graph to video grounding for weakly-supervised multi-step localization. In European Conference on Computer Vision, pages 319-335. Springer, 2022. 3 +[22] Alessandro Flaborea, Guido Maria D'Amely di Melendugno, Leonardo Plini, Luca Scofano, Edoardo De Matteis, Antonio Furnari, Giovanni Maria Farinella, and Fabio Galasso. Prego: online mistake detection in procedural egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18483-18492, 2024. 2 +[23] Antonino Furnari and Giovanni Maria Farinella. What would you expect? anticipating egocentric actions with rolling-unrolling lstms and modality attention. In Proceedings of the IEEE/CVF International conference on computer vision, pages 6252-6261, 2019. 2 + +[24] Antonino Furnari, Sebastiano Battiato, and Giovanni Maria Farinella. Leveraging uncertainty to rethink loss functions and evaluation measures for egocentric action anticipation. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 6 +[25] Mingfei Gao, Mingze Xu, Larry S Davis, Richard Socher, and Caiming Xiong. Startnet: Online detection of action start in untrimmed videos. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5542-5551, 2019. 2 +[26] Rohit Girdhar and Kristen Grauman. Anticipative video transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13505-13515, 2021. 2 +[27] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 2 +[28] Kristen Grauman, Andrew Westbury, Lorenzo Torresani, Kris Kitani, Jitendra Malik, Triantafyllos Afouras, Kumar Ashutosh, Vijay Baiyya, Siddhant Bansal, Bikram Boote, et al. Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19383-19400, 2024. 2, 3, 4, 6 +[29] Bo He, Hengduo Li, Young Kyun Jang, Menglin Jia, Xuefei Cao, Ashish Shah, Abhinav Shrivastava, and Ser-Nam Lim. Ma-lmm: Memory-augmented large multimodal model for long-term video understanding. CVPR, 2024. 1, 2, 3 +[30] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2 +[31] Cheng-Ping Hsieh, Simeng Sun, Samuel Kriman, Shantanu Acharya, Dima Rekesh, Fei Jia, Yang Zhang, and Boris Ginsburg. Ruler: What's the real context size of your long-context language models? arXiv preprint arXiv:2404.06654, 2024. 1, 3 +[32] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 6 +[33] Md Mohaiminul Islam, Tushar Nagarajan, Huiyu Wang, Fu-Jen Chu, Kris Kitani, Gedas Bertasius, and Xitong Yang. Propose, assess, search: Harnessing llms for goal-oriented planning in instructional videos. arXiv preprint arXiv:2409.20557, 2024. 2 +[34] Y.-G. Jiang, J. Liu, A. Roshan Zamir, G. Toderici, I. Laptev, M. Shah, and R. Sukthankar. THUMOS challenge: Action recognition with a large number of classes. http://crcv.ucf.edu/THUMOS14/, 2014.1 +[35] Mahnaz Koupae and William Yang Wang. Wikihow: A large scale text summarization dataset. arXiv preprint arXiv:1810.09305, 2018. 3 + +[36] Bolin Lai, Sam Toyer, Tushar Nagarajan, Rohit Girdhar, Shengxin Zha, James M Rehg, Kris Kitani, Kristen Grauman, Ruta Desai, and Miao Liu. Human action anticipation: A survey. arXiv preprint arXiv:2410.14045, 2024. 1 +[37] Shih-Po Lee, Zijia Lu, Zekun Zhang, Minh Hoai, and Ehsan Elhamifar. Error detection in egocentric procedural task videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18655-18666, 2024. 2 +[38] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 2, 3, 7 +[39] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2, 4, 8 +[40] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 2 +[41] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 2, 7 +[42] Han Lin, Tushar Nagarajan, Nicolas Ballas, Mido Assran, Mojtaba Komeili, Mohit Bansal, and Koustuv Sinha. Vedit: Latent prediction architecture for procedural video representation learning. arXiv preprint arXiv:2410.03478, 2024. 6, 7 +[43] Xudong Lin, Fabio Petroni, Gedas Bertasius, Marcus Rohrbach, Shih-Fu Chang, and Lorenzo Torresani. Learning to recognize procedural activities with distant supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13853-13863, 2022. 2, 7 +[44] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2023. 2, 3 +[45] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 2, 8 +[46] Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. Lost in the middle: How language models use long contexts. Transactions of the Association for Computational Linguistics, 12: 157-173, 2024. 1, 3 +[47] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 2 +[48] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of + +the IEEE/CVF international conference on computer vision, pages 2630-2640, 2019. 2 +[49] Medhini Narasimhan, Arsha Nagrani, Chen Sun, Michael Rubinstein, Trevor Darrell, Anna Rohrbach, and Cordelia Schmid. Tl; dw? summarizing instructional videos with task relevance and cross-modal saliency. In European Conference on Computer Vision, pages 540-557. Springer, 2022. 2 +[50] Medhini Narasimhan, Licheng Yu, Sean Bell, Ning Zhang, and Trevor Darrell. Learning and verification of task structure in instructional videos. arXiv preprint arXiv:2303.13519, 2023. 7 +[51] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 7 +[52] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 4 +[53] Zhanzhong Pang, Fadime Sener, and Angela Yao. Context-enhanced memory-refined transformer for online action detection. arXiv preprint arXiv:2503.18359, 2025. 1, 2 +[54] Dhruvesh Patel, Hamid Eghbalzadeh, Nitin Kamra, Michael Louis Iuzzolino, Unnat Jain, and Ruta Desai. Pretrained language models as visual planners for human assistance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15302-15314, 2023. 3 +[55] Reiner Pope, Sholto Douglas, Aakanksha Chowdhery, Jacob Devlin, James Bradbury, Jonathan Heek, Kefan Xiao, Shivani Agrawal, and Jeff Dean. Efficiently scaling transformer inference. Proceedings of Machine Learning and Systems, 5:606-624, 2023. 5 +[56] Shraman Pramanick, Yale Song, Sayan Nag, Kevin Qinghong Lin, Hardik Shah, Mike Zheng Shou, Rama Chellappa, and Pengchuan Zhang. Egovlpv2: Egocentric video-language pre-training with fusion in the backbone. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5285-5297, 2023. 7 +[57] Alec Radford. Improving language understanding by generative pre-training. Openai, 2018. 5 +[58] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2 +[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 4 +[60] Francesco Ragusa, Antonino Furnari, Salvatore Livatino, and Giovanni Maria Farinella. The meccano dataset: Understanding human-object interactions from egocentric videos in an industrial-like domain. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1569-1578, 2021. 1 + +[61] Abhinav Rai, Fadime Sener, and Angela Yao. Transformed rois for capturing visual transformations in videos. Computer Vision and Image Understanding, 224:103558, 2022. 2 +[62] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 2 +[63] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 5 +[64] Luigi Seminara, Giovanni Maria Farinella, and Antonino Furnari. Differentiable task graph learning: Procedural activity representation and online mistake detection from egocentric videos. In Advances in Neural Information Processing Systems, pages 59373-59407. Curran Associates, Inc., 2024. 3 +[65] Fadime Sener and Angela Yao. Zero-shot anticipation for instructional activities. In Proceedings of the IEEE/CVF international conference on computer vision, pages 862-871, 2019. 1, 2 +[66] Fadime Sener, Dipika Singhania, and Angela Yao. Temporal aggregate representations for long-range video understanding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16, pages 154-171. Springer, 2020. 2, 6 +[67] Fadime Sener, Dibyadip Chatterjee, Daniel Shelepov, Kun He, Dipika Singhania, Robert Wang, and Angela Yao. Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21096-21106, 2022. 1, 2, 3, 6 +[68] Md Salman Shamil, Dibyadip Chatterjee, Fadime Sener, Shugao Ma, and Angela Yao. On the utility of 3d hand poses for action recognition. In European Conference on Computer Vision, pages 436-454. Springer, 2024. 4 +[69] Dandan Shan, Jiaqi Geng, Michelle Shu, and David F Fouhey. Understanding human hands in contact at internet scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9869-9878, 2020. 4, 5, 6 +[70] Zheng Shou, Junting Pan, Jonathan Chan, Kazuyuki Miyazawa, Hassan Mansour, Anthony Vetro, Xavier Giro-i Nieto, and Shih-Fu Chang. Online detection of action start in untrimmed, streaming videos. In Proceedings of the European conference on computer vision (ECCV), pages 534-551, 2018. 1, 2 +[71] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 2 + +[72] Yale Song, Eugene Byrne, Tushar Nagarajan, Huiyu Wang, Miguel Martin, and Lorenzo Torresani. Ego4d goal-step: Toward hierarchical understanding of procedural activities. Advances in Neural Information Processing Systems, 36, 2023. 2, 3, 4, 6 +[73] Reuben Tan, Xineng Sun, Ping Hu, Jui-hsien Wang, Hanieh Deilamsalehy, Bryan A Plummer, Bryan Russell, and Kate Saenko. Koala: Key frame-conditioned long video-llm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13581–13591, 2024. 2 +[74] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1207-1216, 2019. 2, 6 +[75] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 3 +[76] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568-9578, 2024. 3 +[77] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2 +[78] Huiyu Wang, Mitesh Kumar Singh, and Lorenzo Torresani. Ego-only: Egocentric action detection without exocentric transferring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5250-5261, 2023. 6 +[79] Jiahao Wang, Guo Chen, Yifei Huang, Limin Wang, and Tong Lu. Memory-and-anticipation transformer for online action understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13824–13835, 2023. 1, 2, 3, 4, 6 +[80] Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video understanding with large language model as agent. arXiv preprint arXiv:2403.10517, 2024. 2 +[81] Yuetian Weng, Mingfei Han, Haoyu He, Xiaojun Chang, and Bohan Zhuang. Longvlm: Efficient long video understanding via large language models. arXiv preprint arXiv:2404.03384, 2024. 2, 5 +[82] Wikipedia. How to do anything. http://www.wikipedia.com/, 2005.1 +[83] Shiwei Wu, Joya Chen, Kevin Qinghong Lin, Qimeng Wang, Yan Gao, Qianli Xu, Tong Xu, Yao Hu, Enhong Chen, and Mike Zheng Shou. Videollm-mod: Efficient video-language streaming with mixture-of-depths vision computation. arXiv preprint arXiv:2408.16730, 2024. 1, 2, 6, 7, 8 +[84] Junbin Xiao, Angela Yao, Yicong Li, and Tat-Seng Chua. Can i trust your answer? visually grounded video question + +answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13204-13214, 2024. 3 +[85] Junbin Xiao, Nanxin Huang, Hangyu Qin, Dongyang Li, Yicong Li, Fengbin Zhu, Zhulin Tao, Jianxing Yu, Liang Lin, Tat-Seng Chua, et al. Videoqa in the era of llms: An empirical study. International Journal of Computer Vision, pages 1-24, 2025. 3 +[86] Mingze Xu, Yuanjun Xiong, Hao Chen, Xinyu Li, Wei Xia, Zhuowen Tu, and Stefano Soatto. Long short-term transformer for online action detection. Advances in Neural Information Processing Systems, 34:1086-1099, 2021. 1, 2, 3, 4, 6 +[87] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 3 +[88] Olga Zatsarynna and Juergen Gall. Action anticipation with goal consistency. In 2023 IEEE International Conference on Image Processing (ICIP), pages 1630-1634. IEEE, 2023. 6 +[89] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 2, 3, 4 +[90] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Is an object-centric video representation beneficial for transfer? In Proceedings of the Asian Conference on Computer Vision, pages 1976-1994, 2022. 2 +[91] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Helping hands: An object-aware ego-centric video recognition model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13901-13912, 2023. 4 +[92] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 543–553, 2023. 2 +[93] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 3 +[94] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 2 +[95] He Zhao, Isma Hadji, Nikita Dvornik, Konstantinos G Derpanis, Richard P Wildes, and Allan D Jepson. P3iv: Probabilistic procedure planning from instructional videos with weak supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2938-2948, 2022. 2 +[96] Qi Zhao, Shijie Wang, Ce Zhang, Changcheng Fu, Minh Quan Do, Nakul Agarwal, Kwonjoon Lee, and Chen Sun. AntGPT: Can large language models help long-term action anticipation from videos? In The Twelfth International Conference on Learning Representations, 2024. 2 + +[97] Yue Zhao and Philipp Krahenbuhl. Real-time online video detection with temporal smoothing transformers. In European Conference on Computer Vision, pages 485-502. Springer, 2022. 1, 2, 3, 6 +[98] Yue Zhao, Long Zhao, Xingyi Zhou, Jialin Wu, Chun-Te Chu, Hui Miao, Florian Schroff, Hartwig Adam, Ting Liu, Boqing Gong, et al. Distilling vision-language models on millions of videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13106-13116, 2024. 2 +[99] Yiwu Zhong, Licheng Yu, Yang Bai, Shangwen Li, Xueting Yan, and Yin Li. Learning procedure-aware video representation from instructional videos and their narrations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14825-14835, 2023. 7 +[100] Honglu Zhou, Roberto Martin-Martín, Mubbasir Kapadia, Silvio Savarese, and Juan Carlos Niebles. Procedure-aware pretraining for instructional video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10727-10738, 2023. 1, 2, 3, 7 +[101] Luowei Zhou, Chenliang Xu, and Jason Corso. Towards automatic learning of procedures from web instructional videos. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 1, 2 +[102] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2 +[103] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramazan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Crosstask weakly supervised learning from instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3537-3545, 2019. 2 \ No newline at end of file diff --git a/2025/Streaming VideoLLMs for Real-Time Procedural Video Understanding/images.zip b/2025/Streaming VideoLLMs for Real-Time Procedural Video Understanding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f6a234397602b0aa10ed0f071f8f239be3a1e80e --- /dev/null +++ b/2025/Streaming VideoLLMs for Real-Time Procedural Video Understanding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d1fcad7a7e507749a7e91b3855bfbb6349b2e0c41655d9e038524192356687b6 +size 457855 diff --git a/2025/Streaming VideoLLMs for Real-Time Procedural Video Understanding/layout.json b/2025/Streaming VideoLLMs for Real-Time Procedural Video Understanding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d8a4f77609a52d36ef9d88d2e3a406a6f0a5cf73 --- /dev/null +++ b/2025/Streaming VideoLLMs for Real-Time Procedural Video Understanding/layout.json @@ -0,0 +1,10567 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 87, + 103, + 523, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 103, + 523, + 121 + ], + "spans": [ + { + "bbox": [ + 87, + 103, + 523, + 121 + ], + "type": "text", + "content": "Streaming VideoLLMs for Real-time Procedural Video Understanding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 142, + 560, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 142, + 560, + 173 + ], + "spans": [ + { + "bbox": [ + 61, + 142, + 560, + 173 + ], + "type": "text", + "content": "Dibyadip Chatterjee" + }, + { + "bbox": [ + 61, + 142, + 560, + 173 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 61, + 142, + 560, + 173 + ], + "type": "text", + "content": " Edoardo Remelli" + }, + { + "bbox": [ + 61, + 142, + 560, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 560, + 173 + ], + "type": "text", + "content": " Yale Song" + }, + { + "bbox": [ + 61, + 142, + 560, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 142, + 560, + 173 + ], + "type": "text", + "content": " Bugra Tekin" + }, + { + "bbox": [ + 61, + 142, + 560, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 560, + 173 + ], + "type": "text", + "content": " Abhay Mittal" + }, + { + "bbox": [ + 61, + 142, + 560, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 560, + 173 + ], + "type": "text", + "content": " Bharat Bhatnagar" + }, + { + "bbox": [ + 61, + 142, + 560, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 560, + 173 + ], + "type": "text", + "content": " Necati Cihan Camgoz" + }, + { + "bbox": [ + 61, + 142, + 560, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 560, + 173 + ], + "type": "text", + "content": " Shreyas Hampali" + }, + { + "bbox": [ + 61, + 142, + 560, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 560, + 173 + ], + "type": "text", + "content": " Eric Sauser" + }, + { + "bbox": [ + 61, + 142, + 560, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 560, + 173 + ], + "type": "text", + "content": " Shugao Ma" + }, + { + "bbox": [ + 61, + 142, + 560, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 560, + 173 + ], + "type": "text", + "content": " Angela Yao" + }, + { + "bbox": [ + 61, + 142, + 560, + 173 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 142, + 560, + 173 + ], + "type": "text", + "content": " Fadime Sener" + }, + { + "bbox": [ + 61, + 142, + 560, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 177, + 492, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 177, + 492, + 194 + ], + "spans": [ + { + "bbox": [ + 131, + 177, + 492, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 131, + 177, + 492, + 194 + ], + "type": "text", + "content": "Meta Reality Labs " + }, + { + "bbox": [ + 131, + 177, + 492, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 131, + 177, + 492, + 194 + ], + "type": "text", + "content": "FAIR, Meta " + }, + { + "bbox": [ + 131, + 177, + 492, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 131, + 177, + 492, + 194 + ], + "type": "text", + "content": "National University of Singapore" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 231, + 194, + 390, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 194, + 390, + 205 + ], + "spans": [ + { + "bbox": [ + 231, + 194, + 390, + 205 + ], + "type": "text", + "content": "dibschat.github.io/ProVideLLM" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 71, + 242, + 541, + 387 + ], + "blocks": [ + { + "bbox": [ + 71, + 242, + 541, + 387 + ], + "lines": [ + { + "bbox": [ + 71, + 242, + 541, + 387 + ], + "spans": [ + { + "bbox": [ + 71, + 242, + 541, + 387 + ], + "type": "image", + "image_path": "9109d7903435817505e944fca36d9aec8c09ad0ece6a264152e2040549ee34c6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 392, + 555, + 439 + ], + "lines": [ + { + "bbox": [ + 54, + 392, + 555, + 439 + ], + "spans": [ + { + "bbox": [ + 54, + 392, + 555, + 439 + ], + "type": "text", + "content": "Figure 1. Overview of ProVideLLM. A streaming video large language model for real-time procedural video tasks with a low memory footprint. It features a multimodal interleaved cache that stores textual descriptions of long-term observations (spanning several minutes) and visual tokens representing short-term observations (spanning a few seconds). We also introduce a new DETR-QFormer connector for better fine-grained tokenization of the short-term. ProVideLLM is capable of handling multiple procedural tasks within a single model." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 457, + 200, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 457, + 200, + 470 + ], + "spans": [ + { + "bbox": [ + 151, + 457, + 200, + 470 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 487, + 297, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 487, + 297, + 680 + ], + "spans": [ + { + "bbox": [ + 54, + 487, + 297, + 680 + ], + "type": "text", + "content": "We introduce ProVideLLM, an end-to-end framework for real-time procedural video understanding. ProVideLLM integrates a multimodal cache configured to store two types of tokens – verbalized text tokens, which provide compressed textual summaries of long-term observations, and visual tokens, encoded with DETR-QFormer to capture fine-grained details from short-term observations. This design reduces token count by " + }, + { + "bbox": [ + 54, + 487, + 297, + 680 + ], + "type": "inline_equation", + "content": "22 \\times" + }, + { + "bbox": [ + 54, + 487, + 297, + 680 + ], + "type": "text", + "content": " over existing methods when representing one hour of long-term observations while effectively encoding fine-granularity of the present. By interleaving these tokens in our multimodal cache, ProVideLLM achieves sublinear scaling of memory and compute with video length, ensuring per-frame streaming inference at 10 FPS and streaming dialogue at 25 FPS, with a minimal 2GB GPU memory footprint. ProVideLLM also sets new state-of-the-art results on six procedural tasks across four datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 452, + 394, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 452, + 394, + 464 + ], + "spans": [ + { + "bbox": [ + 315, + 452, + 394, + 464 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 473, + 556, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 556, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 556, + 581 + ], + "type": "text", + "content": "Procedural videos, spanning several minutes to hours, capture multi-step tasks [82] such as cooking [65, 101], toy assembly [60, 67], and furniture building [7, 8]. With the growing adoption of AR assistants, streaming video understanding is essential for real-time procedural assistance. This requires real-time, online inference, as frames appear in a live video, without access to future observations. Despite its importance, streaming video understanding remains largely unexplored in existing procedural research [19, 36, 100]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "content": "Existing works on streaming video understanding [17, 70] focus on online action detection [53, 79, 86, 97] or live conversations [14, 83]. These use-cases target a small set of unrelated actions [17, 34] and do not require long-term temporal modeling. In contrast, procedural videos are inherently long and causal, where each step depends on prior steps. Applying existing approaches to such videos significantly increases token count and memory overhead, leading to long-context degradation [31, 46] and making them unsuitable for streaming, even with excessive frame dropping [29]. The semantic inter-connectivity of activity steps introduces re" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 206, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 206, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 206, + 713 + ], + "type": "text", + "content": "*Work done during an internship at Meta." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22586" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "dundancy when long-term observations are stored as raw frame tokens [79, 86]. Furthermore, while short-term observations are crucial for understanding the present, not all frame tokens are equally relevant, since procedural tasks often center around hand-object interactions [28, 67, 72]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 134, + 295, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 134, + 295, + 253 + ], + "spans": [ + { + "bbox": [ + 54, + 134, + 295, + 253 + ], + "type": "text", + "content": "In this paper, we propose an end-to-end framework for streaming procedural video understanding (Fig. 1), addressing the above mentioned challenges of processing long-form videos. Specifically, we introduce ProVideLLM, a Procedural VideoLLM that comprises a visual (frame) encoder, a vision-language connector, and a pre-trained large language model (LLM) [1, 20, 58, 77] as the decoder. Our key contribution is a vision-language multimodal cache that efficiently manages streaming video data by balancing compression and representation quality of past observations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 255, + 297, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 297, + 506 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 297, + 506 + ], + "type": "text", + "content": "We factorize the streaming past observations into long-term semantic history and short-term visual context. They are encoded as text and visual tokens, respectively, and interleaved in our multimodal cache. First, for the long-term history spanning several minutes, ProVideLLM decoder predicts step labels per frame, groups similar steps into verbalized sequences (e.g. get " + }, + { + "bbox": [ + 56, + 255, + 297, + 506 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 56, + 255, + 297, + 506 + ], + "type": "text", + "content": " crack " + }, + { + "bbox": [ + 56, + 255, + 297, + 506 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 56, + 255, + 297, + 506 + ], + "type": "text", + "content": " whisk eggs), and stores them in the cache. This semantic compression significantly reduces token count and also enhances current step prediction (e.g. " + }, + { + "bbox": [ + 56, + 255, + 297, + 506 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 56, + 255, + 297, + 506 + ], + "type": "text", + "content": " pour mixture). Second, to capture short-term visual context, typically spanning a few seconds, understanding the fine-grained changes across frames is crucial. However, standard VideoLLM encoders [59, 89] face two key challenges in such modeling: (i) temporal collapse, where low temporal variance conflates distinct actions and (ii) background distractors across frames [5, 13, 61, 90]. We mitigate the former by using visual encoders trained with dense spatial alignment objectives [11, 30, 52]. We address the latter with a novel vision-text connector DETR-QFormer, pretrained to explicitly focus on hand-object interaction regions, thereby compressing them into a smaller set of tokens." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 509, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 628 + ], + "type": "text", + "content": "Finally, in existing methods [86], both long- and short-term tokens are visual and stored separately. Applying such separate caching to our multimodal tokens would require redundant attention recomputation when verbalizing short-term visual tokens into long-term text. To overcome this, we interleave text and visual tokens within our multimodal cache. This enables prefilling the cache, which ensures that computation scales linearly with cache size and sub-linearly with the video length, thereby significantly reducing latency for real-time performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": "In summary, our contributions are: (i) a first-of-its-kind streaming multimodal cache with interleaved text and visual tokens, enabling per-frame inference [72] at 10 FPS and streaming dialogue [14] at 25 FPS; (ii) verbalizing long-term observations into compressed textual summaries and (iii) modeling hand-object interactions for short-term observations using DETR-QFormer; all combined, it achieves a" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "22 \\times" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " reduction in token count compared to existing methods. (iv) state-of-the-art results on six tasks across four datasets, making ProVideLLM the first real-time framework unifying recognition, anticipation, planning for procedural guidance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 125, + 406, + 137 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 125, + 406, + 137 + ], + "spans": [ + { + "bbox": [ + 314, + 125, + 406, + 137 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 144, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 312 + ], + "type": "text", + "content": "Procedural Video Understanding have primarily focused on instructional YouTube videos [48, 74, 101, 103]. Recent large-scale egocentric datasets [27, 28, 67, 72] have advanced research by providing longer videos with more activity steps. Existing works have focused on offline tasks [19, 43, 100], where the entire video is available for prediction, as well as some online tasks, like procedural planning [12, 33, 95], forecasting [23, 26, 49, 65, 66, 96], and mistake detection [18, 22, 37, 67]. However, deploying these models for streaming inference results in high computational complexity where it grows quadratically with the video length. In contrast, ProVideLLM incurs a sub-linear cost per update, both in terms of memory and runtime, enabling efficient streaming inference." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 312, + 556, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 556, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 556, + 528 + ], + "type": "text", + "content": "Video Large Language Models. Multimodal large language models (MLLMs) have been popularized by BLIP-2-like [15, 39, 102] and LLaVA-like [38, 44, 45] models where an LLM is trained to understand vision-text data. Extending these to video have led to the development of Video Large Language Models (VideoLLMs) [40, 41, 47, 83, 94, 98]. However, most of these models are limited to offline tasks [62, 71, 73, 80, 81] where the entire video is queried to answer a question. As such, architectural elements such as VideoQ-Former [92] and temporal pooling layers [40], developed without causal requirements, limit streaming applications. Models like MA-LMM [29] and VideoLLM online [14] address video understanding online, but are too computationally intensive to be deployed for efficient streaming inference on long-form videos. ProVideLLM advances these models by introducing a multimodal interleaved cache for faster inference and proposing DETR-QFormer for modeling fine-grained visual changes, improving accuracy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 528, + 556, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 556, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 556, + 647 + ], + "type": "text", + "content": "Streaming Video Understanding has been explored for online action detection [17, 25, 70] where dominant approaches [53, 79, 86, 97] split streaming input into long- and short-term visual histories within an encoder-decoder architecture. However, storing both as visual tokens is redundant and computationally prohibitive for long-form procedural videos, where multiple actions are temporally connected. Instead, ProVideLLM verbalizes long-term observations into text which drastically decreases token count required to represent hour long videos, thus improving memory efficiency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 657, + 399, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 657, + 399, + 669 + ], + "spans": [ + { + "bbox": [ + 314, + 657, + 399, + 669 + ], + "type": "text", + "content": "3. ProVideLLM" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "ProVideLLM encodes long- and short-term streaming observations into text and visual tokens and interleaves them in a multimodal cache. We first present the preliminaries" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22587" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": "in Sec. 3.1, followed by long- and short-term token generation in Sec. 3.2 and Sec. 3.3. Finally, we present our multimodal cache in Sec. 3.4 and training objectives in Sec. 3.5." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 113, + 141, + 125 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 113, + 141, + 125 + ], + "spans": [ + { + "bbox": [ + 55, + 113, + 141, + 125 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 129, + 296, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 129, + 296, + 189 + ], + "spans": [ + { + "bbox": [ + 55, + 129, + 296, + 189 + ], + "type": "text", + "content": "ProVideLLM follows a LLaVA-style [44] architecture, consisting of a visual encoder " + }, + { + "bbox": [ + 55, + 129, + 296, + 189 + ], + "type": "inline_equation", + "content": "(E_V)" + }, + { + "bbox": [ + 55, + 129, + 296, + 189 + ], + "type": "text", + "content": ", a language decoder " + }, + { + "bbox": [ + 55, + 129, + 296, + 189 + ], + "type": "inline_equation", + "content": "(D_L)" + }, + { + "bbox": [ + 55, + 129, + 296, + 189 + ], + "type": "text", + "content": " and a vision-language connector " + }, + { + "bbox": [ + 55, + 129, + 296, + 189 + ], + "type": "inline_equation", + "content": "(C_{VL})" + }, + { + "bbox": [ + 55, + 129, + 296, + 189 + ], + "type": "text", + "content": ". Given a video " + }, + { + "bbox": [ + 55, + 129, + 296, + 189 + ], + "type": "inline_equation", + "content": "V = \\{v_1, v_2, \\ldots, v_T\\}" + }, + { + "bbox": [ + 55, + 129, + 296, + 189 + ], + "type": "text", + "content": ", the model's output " + }, + { + "bbox": [ + 55, + 129, + 296, + 189 + ], + "type": "inline_equation", + "content": "(o_t)" + }, + { + "bbox": [ + 55, + 129, + 296, + 189 + ], + "type": "text", + "content": " at any time step " + }, + { + "bbox": [ + 55, + 129, + 296, + 189 + ], + "type": "inline_equation", + "content": "t \\in [1, T]" + }, + { + "bbox": [ + 55, + 129, + 296, + 189 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 74, + 193, + 296, + 212 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 193, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 74, + 193, + 296, + 212 + ], + "type": "interline_equation", + "content": "o _ {t} = \\operatorname {P r o V i d e L L M} \\left(v _ {1: t}\\right) = D _ {L} \\left(\\left\\| _ {i = 1} ^ {t} C _ {V L} \\left(E _ {V} \\left(v _ {i}\\right)\\right)\\right) \\quad (1) \\right.", + "image_path": "ba7624bc96e4136db9af2f8196c6cd4aca2132fdeb6a31ffbccf56379127072b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 215, + 296, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 215, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 296, + 300 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 215, + 296, + 300 + ], + "type": "inline_equation", + "content": "\\parallel" + }, + { + "bbox": [ + 55, + 215, + 296, + 300 + ], + "type": "text", + "content": " denotes concatenation. While LSTR-inspired long short-term memory approaches [79, 97] also use an encoder-decoder architecture, they rely exclusively on visual tokens (Fig. 2a.1). In contrast, ProVideLLM incorporates a language decoder, which enables more efficient long-term compression, improves generalization, and supports multiple tasks within a single model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 304, + 240, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 240, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 240, + 316 + ], + "type": "text", + "content": "3.2. Tokenizing long-term observations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 319, + 296, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 319, + 296, + 511 + ], + "spans": [ + { + "bbox": [ + 54, + 319, + 296, + 511 + ], + "type": "text", + "content": "Procedural videos are inherently long-form, ranging from tens of minutes [67] to several hours [28, 72]. While modern LLMs [20, 75, 87] offer massive context lengths and, in theory, can process an hour-long raw video at 24 fps, performance degrades with larger context sizes [31, 46], and training such models is highly memory-intensive [29, 93]. Accurately recognizing ongoing steps require effectively encoding long-term context. Our key insight is that procedural steps in a video are temporally connected, i.e. a sequence of step labels can serve as a strong predictor of ongoing steps. For instance, when \"preparing an omelet\", the step \"pour mixture into pan\" depends on earlier steps \"get eggs from fridge, crack eggs into bowl, whisk eggs\". Prior works on flow graphs [21, 64] and planning [3, 54, 100] also show that step sequences, often extracted from external sources [35], serve as strong baselines for procedural understanding." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 511, + 297, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 511, + 297, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 511, + 297, + 654 + ], + "type": "text", + "content": "Instead of constructing computationally expensive graphs, we propose online verbalization (Fig. 2a.2). Given a streaming video, ProVideLLM decoder continuously generates action predictions (summaries) of the present observation on the fly, and groups semantically similar actions to verbalize the past (Fig. 1). This approach efficiently compresses past visual frame tokens into information-dense language tokens. For example, in Ego4D Goal-Step [72], storing one hour of long-term past requires only 630 verbalized text tokens on average. Compared to long short-term approaches [79, 86, 97], our verbalization reduces token count by " + }, + { + "bbox": [ + 55, + 511, + 297, + 654 + ], + "type": "inline_equation", + "content": "22 \\times" + }, + { + "bbox": [ + 55, + 511, + 297, + 654 + ], + "type": "text", + "content": " at the same sampling rate, significantly reducing the memory overhead." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 660, + 244, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 244, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 244, + 673 + ], + "type": "text", + "content": "3.3. Tokenizing short-term observations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Our goal is to recognize procedural activity steps in a streaming manner. This requires understanding \"what is happening\" in the present while tracking \"what has happened in" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 550, + 196 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 550, + 196 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 550, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 550, + 196 + ], + "type": "image", + "image_path": "549ec69cc1878c662a86f32b066748037c4486fa185624f420f11daa967e6b21.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 365, + 205, + 554, + 277 + ], + "blocks": [ + { + "bbox": [ + 365, + 205, + 554, + 277 + ], + "lines": [ + { + "bbox": [ + 365, + 205, + 554, + 277 + ], + "spans": [ + { + "bbox": [ + 365, + 205, + 554, + 277 + ], + "type": "image", + "image_path": "0a552e73e5bf57d1c56da3a067243342e80adc27a808b023a7c3c5b2f0132174.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 285, + 555, + 385 + ], + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 385 + ], + "type": "text", + "content": "Figure 2. Converting short-term tokens to long-term under various caching types. (a) Factorizing the streaming observations into long-term and short-term and ordering them progressively. (a.1) Conversion done at " + }, + { + "bbox": [ + 313, + 285, + 555, + 385 + ], + "type": "inline_equation", + "content": "O(1)" + }, + { + "bbox": [ + 313, + 285, + 555, + 385 + ], + "type": "text", + "content": " cost but is unable to represent long-form videos due to massive token count and memory needs. (a.2) Online verbalization can represent long-form videos at a manageable memory but is not suitable for streaming due to " + }, + { + "bbox": [ + 313, + 285, + 555, + 385 + ], + "type": "inline_equation", + "content": "O(N_S^2 + N_S N_L)" + }, + { + "bbox": [ + 313, + 285, + 555, + 385 + ], + "type": "text", + "content": " conversion cost. (b) Interleaving them reduces the conversion cost to " + }, + { + "bbox": [ + 313, + 285, + 555, + 385 + ], + "type": "inline_equation", + "content": "O(N)" + }, + { + "bbox": [ + 313, + 285, + 555, + 385 + ], + "type": "text", + "content": ", allowing streaming inference on long-form videos." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 402, + 556, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 556, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 556, + 534 + ], + "type": "text", + "content": "the past\". Language is a strong indicator of the semantic structure of step sequences, but determining if and when the step is performed requires fine-grained visual information. Therefore, we allocate more compute for understanding the short-term past, typically spanning 8-16 seconds, by keeping its visual tokens uncompressed. Standard VideoLLM encoders [38, 44] face two key challenges in capturing fine-grained short-term observations: " + }, + { + "bbox": [ + 313, + 402, + 556, + 534 + ], + "type": "inline_equation", + "content": "(i)" + }, + { + "bbox": [ + 313, + 402, + 556, + 534 + ], + "type": "text", + "content": " they generate temporally collapsed video features, a limitation we identify and address in Sec. 3.3.1, and " + }, + { + "bbox": [ + 313, + 402, + 556, + 534 + ], + "type": "inline_equation", + "content": "(ii)" + }, + { + "bbox": [ + 313, + 402, + 556, + 534 + ], + "type": "text", + "content": " they get biased towards high-level scene and background distractors, which we address in Sec. 3.3.2." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 542, + 463, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 463, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 463, + 553 + ], + "type": "text", + "content": "3.3.1. Choosing the visual encoder" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 557, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 556, + 714 + ], + "type": "text", + "content": "The predominant visual encoders for MLLMs are CLIP [59] and SigLIP [2, 89]. These encoders are pre-trained with only a language-alignment objective for integration with LLMs. However, because they are trained on high-level and sparse web-scale captions such as \"cooking in the kitchen\", they struggle with fine-grained reasoning, particularly in answering the \"what?\" and \"how to?\" aspects of a video [84, 85]. Consequently, distinct images may yield high similarity scores in their feature space [76], leading to low feature diversity across frames within a step. This results in relatively static frame features, a phenomenon we term temporal collapse. For certain fine-grained procedural steps, such as \"pick up\" vs. \"put down\", temporally collapsed" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22588" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 71, + 290, + 225 + ], + "blocks": [ + { + "bbox": [ + 58, + 104, + 69, + 114 + ], + "lines": [ + { + "bbox": [ + 58, + 104, + 69, + 114 + ], + "spans": [ + { + "bbox": [ + 58, + 104, + 69, + 114 + ], + "type": "text", + "content": "a)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 73, + 71, + 290, + 225 + ], + "lines": [ + { + "bbox": [ + 73, + 71, + 290, + 225 + ], + "spans": [ + { + "bbox": [ + 73, + 71, + 290, + 225 + ], + "type": "image", + "image_path": "04a5d58bb747d1f827032af5c311f84394c786979c079a378006d3e83aee0eea.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 178, + 69, + 190 + ], + "lines": [ + { + "bbox": [ + 55, + 178, + 69, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 69, + 190 + ], + "type": "text", + "content": "b)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 234, + 296, + 312 + ], + "lines": [ + { + "bbox": [ + 55, + 234, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 296, + 312 + ], + "type": "text", + "content": "Figure 3. a) Per-class mean temporal variance on EgoExo4D & Ego4D Goal-Step. Highly temporally varying classes on the left and low-variance classes on the right. b) Visualization of the top-3 PCA components of patch tokens computed across video frames, thresholded by first component. The primary factors of variation for DINOv2 [52] are hands, and are unfocused and scattered across the image for language-aligned encoders, CLIP [59] and SigLIP [89]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 327, + 295, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 295, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 295, + 351 + ], + "type": "text", + "content": "features make it nearly impossible for the downstream LLM to decode the correct step." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 351, + 296, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 351, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 54, + 351, + 296, + 542 + ], + "type": "text", + "content": "Fig. 3a shows the impact of temporal collapse across three visual encoders – two language-aligned ones [59, 89] and one self-supervised1 [52]. We plot temporal variance per class in descending order, with high-variance classes on the left and low-variance classes on the right (see Suppl.Sec.1 for temporal variance computation). At both extremes, the trends are consistent across all encoders. High-variance classes, which often involve verbs such as “get”, “fetch” or “put” correspond to longer videos with substantial scene and environmental changes. Conversely, low-variance classes involving adverbs such as “slowly” or “boil” or repetitive actions like “rotate” or “peel”, exhibit gradual state changes. Among the three models, DINOv2 consistently captures temporal variations more effectively, covering a broader range of temporal variance across different classes. Hence, we choose DINOv2 as our visual encoder." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 549, + 285, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 549, + 285, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 549, + 285, + 561 + ], + "type": "text", + "content": "3.3.2. Aligning hand-object transformations with text" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 564, + 296, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 296, + 660 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 296, + 660 + ], + "type": "text", + "content": "Procedural datasets predominantly capture hand-centric actions. For instance, nearly " + }, + { + "bbox": [ + 55, + 564, + 296, + 660 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 55, + 564, + 296, + 660 + ], + "type": "text", + "content": " of actions in [28, 72] involve hand interactions2. Moreover, video frames often contain distracting background objects, while hand locations serve as strong cues for recognizing the ongoing actions [68, 91]. Thus, a vision encoder focused on hand regions can enhance performance by mitigating background distractions and capturing fine-grained transformations more precisely." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 661, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 661, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 661, + 295, + 685 + ], + "type": "text", + "content": "Another contribution of our work is a vision-language connector, DETR-QFormer, pretrained to attend to salient" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 77, + 555, + 251 + ], + "blocks": [ + { + "bbox": [ + 317, + 77, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 317, + 77, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 77, + 555, + 251 + ], + "type": "image", + "image_path": "08271fe2ace3d397040a590658c1cce39a5bd4e7e03de589699a49401f6a6c19.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 259, + 555, + 304 + ], + "lines": [ + { + "bbox": [ + 313, + 259, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 555, + 304 + ], + "type": "text", + "content": "Figure 4. DETR-QFormer architecture & Stage-1 pre-training. Trained to detect hands and objects-in-contact from pseudo ground truths bboxes generated by [69], DETR-QFormer queries DINOv2 patch tokens to extract hand-object-focused activations." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 312, + 556, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 556, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 556, + 432 + ], + "type": "text", + "content": "objects. In procedural videos, these are typically in contact with the hand. Interestingly, our choice of DINOv2 as the visual encoder provides an advantage as its patch tokens capture hands as primary factors of variation Fig. 3b. Leveraging this, we design DETR-QFormer to query DINOv2 patch tokens, extracting hand-object activations and suppressing background distractions. DETR-QFormer explicitly attends to hand-object activations using a DETR loss [10] and compresses salient regions into fewer tokens via QFormer's [39] cross-attention, enhancing representation quality." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "type": "text", + "content": "Architecture. Our connector integrates Q-Former's compression [39] and DETR's bbox loss into a single crossattention block. DETR-QFormer, illustrated in Fig. 4, is pretrained to detect the hands and objects-in-contact (Sec. 3.5). The connector " + }, + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "type": "inline_equation", + "content": "(C_{VL})" + }, + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "type": "text", + "content": " is a cross-attention transformer decoder with a set of " + }, + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "type": "text", + "content": " learnable visual queries " + }, + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "type": "inline_equation", + "content": "Q_{v} = (q_{v_{1}}, q_{v_{2}}, \\ldots, q_{v_{m}})" + }, + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "type": "text", + "content": ". Key-value pairs " + }, + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "type": "inline_equation", + "content": "(K, V)" + }, + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "type": "text", + "content": " are generated from the flattened " + }, + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "type": "inline_equation", + "content": "n^{2}" + }, + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "type": "text", + "content": " DINOv2 patch tokens using two projection heads " + }, + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "type": "inline_equation", + "content": "(W_{K}, W_{V})" + }, + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "type": "text", + "content": " respectively. To explicitly model hand-object interactions, we append two sets of special queries - two hand queries " + }, + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "type": "inline_equation", + "content": "Q_{h} = (q_{h_{1}}, q_{h_{2}})" + }, + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "type": "text", + "content": " for the left and right hands, and " + }, + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "type": "text", + "content": " object queries " + }, + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "type": "inline_equation", + "content": "Q_{o} = (q_{o_{1}}, q_{o_{2}}, \\ldots, q_{o_{k}})" + }, + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "type": "text", + "content": " corresponding to objects-in-contact. The role of visual queries " + }, + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "type": "inline_equation", + "content": "Q_{v}" + }, + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "type": "text", + "content": " is to compress the " + }, + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "type": "inline_equation", + "content": "n^{2} \\times D" + }, + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "type": "text", + "content": " patch tokens into " + }, + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "type": "inline_equation", + "content": "m \\times d" + }, + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "type": "text", + "content": " tokens, where " + }, + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "type": "inline_equation", + "content": "m << n^{2}" + }, + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "type": "inline_equation", + "content": "d << D" + }, + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "type": "text", + "content": ". Whereas, outputs corresponding to the learnable hand-object queries " + }, + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "type": "inline_equation", + "content": "(Q_{h}^{\\prime}, Q_{o}^{\\prime})" + }, + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "type": "text", + "content": " are passed through an MLP to make bounding box predictions, " + }, + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "type": "inline_equation", + "content": "\\hat{B} = \\{\\hat{b}_{h_{1}}, \\hat{b}_{h_{2}}, \\hat{b}_{o_{1}}, \\ldots, \\hat{b}_{o_{k}}\\}" + }, + { + "bbox": [ + 313, + 433, + 556, + 651 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 659, + 480, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 480, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 480, + 670 + ], + "type": "text", + "content": "3.4. Multimodal Interleaved Cache" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "The effectiveness of a procedural video model is best assessed when deployed for real-time inference. Naive long-and short-term approaches [79, 86] maintain separate caches" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 295, + 713 + ], + "type": "text", + "content": "1We chose DINOV2 due to its comparable scale to CLIP and SigLIP. \n2Remaining 1% are visually undetectable actions \"call for help, waiting.\"" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22589" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 56, + 86, + 296, + 227 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 234, + 83 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 234, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 234, + 83 + ], + "type": "text", + "content": "Algorithm 1 Multimodal Interleaved Cache" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 56, + 86, + 296, + 227 + ], + "lines": [ + { + "bbox": [ + 56, + 86, + 296, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 86, + 296, + 227 + ], + "type": "text", + "content": "1: class MultimodalInterleavedCache: \n2: def __init__(self, N.S, N.L): \n3: self_tokens = [] \n4: self.N.S = N.S \n5: self.N.L = N.L \n6: # entry function: adds a new token to cache \n7: def entry(self, token): \n8: self_tokens.append(token) \n9: # exit function for short-term tokens \n10: def exit_short(self): \n11: if self_tokens.count() > self.N.S: \n12: self_tokens.pop(self_tokens.index()) \n13: return True \n14: # exit function for long-term tokens \n15: def exit_long(self): \n16: if self_tokens.count() > self.N.L: \n17: self_tokens.pop(self_tokens.index()" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 54, + 253, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 253, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 54, + 253, + 296, + 361 + ], + "type": "text", + "content": "for the two token types, arranged progressively with a single entry point (see Fig. 2a). With multimodal tokens, such setups require two distinct entry points, forcing attention recomputation over the short-term cache whenever a short-term token is verbalized into a long-term one. In the worst case, this leads to quadratic scaling of compute. To efficiently leverage online verbalization, we propose a multimodal interleaved cache that interleaves long-term text and short-term visual tokens from Sec. 3.2 and Sec. 3.3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 363, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 296, + 579 + ], + "type": "text", + "content": "Interleaving long- and short-term tokens. We store the running short-term visual tokens and long-term text tokens in a single multimodal cache operating as a FIFO (First In, First Out) queue. It has one entry point and two distinct exits for short- and long-term tokens. Our entry and exit functions are detailed in Algorithm 1, where " + }, + { + "bbox": [ + 55, + 363, + 296, + 579 + ], + "type": "inline_equation", + "content": "N_{S}" + }, + { + "bbox": [ + 55, + 363, + 296, + 579 + ], + "type": "text", + "content": " denotes the short-term cache length (in number of frames) and " + }, + { + "bbox": [ + 55, + 363, + 296, + 579 + ], + "type": "inline_equation", + "content": "N_{L}" + }, + { + "bbox": [ + 55, + 363, + 296, + 579 + ], + "type": "text", + "content": " the long-term cache length (in number of observed steps). As frame tokens " + }, + { + "bbox": [ + 55, + 363, + 296, + 579 + ], + "type": "inline_equation", + "content": "(v_{i})" + }, + { + "bbox": [ + 55, + 363, + 296, + 579 + ], + "type": "text", + "content": " exit the short-term span, the ProVideLLM decoder verbalizes the oldest token " + }, + { + "bbox": [ + 55, + 363, + 296, + 579 + ], + "type": "inline_equation", + "content": "(v_{-N_s})" + }, + { + "bbox": [ + 55, + 363, + 296, + 579 + ], + "type": "text", + "content": ". The verbalized text is then added to the cache unless already cached within the past " + }, + { + "bbox": [ + 55, + 363, + 296, + 579 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 363, + 296, + 579 + ], + "type": "text", + "content": " frames. To distinguish the two token types within the same cache, we prepend a special token (represented as " + }, + { + "bbox": [ + 55, + 363, + 296, + 579 + ], + "type": "inline_equation", + "content": "<\\mathbb{L}>" + }, + { + "bbox": [ + 55, + 363, + 296, + 579 + ], + "type": "text", + "content": ") to the long-term text tokens. ProVideLLM is instruction-tuned to recognize tokens associated with " + }, + { + "bbox": [ + 55, + 363, + 296, + 579 + ], + "type": "inline_equation", + "content": "<\\mathbb{L}>" + }, + { + "bbox": [ + 55, + 363, + 296, + 579 + ], + "type": "text", + "content": " as belonging to the long-term observations (Sec. 3.5). Since both long- and short-term tokens share a single entry point, visual and textual tokens are interleaved in the cache at any given time." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "Streaming inference. Every time a new token is added to the cache (Algorithm 2), instead of re-computing attention across all tokens, we use the causal decoder's K-V cache [55] to compute attention only for the new token over the previous " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " tokens. This ensures that the computational complexity for processing a new frame scales linearly with cache size " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "(O(N))" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": ". Additionally, maintaining a single interleaved cache enables direct application of optimized attention implementations [6, 81], though such techniques remain orthogonal to our approach. More importantly, for a fixed " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "N_S" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": ", compression via verbalization ensures that " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "N_L" + } + ] + } + ], + "index": 4 + }, + { + "type": "code", + "bbox": [ + 316, + 86, + 553, + 178 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 453, + 83 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 453, + 83 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 453, + 83 + ], + "type": "text", + "content": "Algorithm 2 Streaming Inference" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 316, + 86, + 553, + 178 + ], + "lines": [ + { + "bbox": [ + 316, + 86, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 316, + 86, + 553, + 178 + ], + "type": "text", + "content": "1: def inference(V, N_S, N_L, " + }, + { + "bbox": [ + 316, + 86, + 553, + 178 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 316, + 86, + 553, + 178 + ], + "type": "text", + "content": "): \n2: cache = MultimodalInterleavedCache(N_S, N_L, " + }, + { + "bbox": [ + 316, + 86, + 553, + 178 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 316, + 86, + 553, + 178 + ], + "type": "text", + "content": ") \n3: out = [] \n4: for " + }, + { + "bbox": [ + 316, + 86, + 553, + 178 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 316, + 86, + 553, + 178 + ], + "type": "text", + "content": " in V: \n5: cache_entry " + }, + { + "bbox": [ + 316, + 86, + 553, + 178 + ], + "type": "inline_equation", + "content": "(v_i)" + }, + { + "bbox": [ + 316, + 86, + 553, + 178 + ], + "type": "text", + "content": " # add frame tokens to cache \n6: " + }, + { + "bbox": [ + 316, + 86, + 553, + 178 + ], + "type": "inline_equation", + "content": "t_i = \\text{ProVideLLM}(\\text{cache.tokens})" + }, + { + "bbox": [ + 316, + 86, + 553, + 178 + ], + "type": "text", + "content": " # current prediction \n7: out.append " + }, + { + "bbox": [ + 316, + 86, + 553, + 178 + ], + "type": "inline_equation", + "content": "(t_i)" + }, + { + "bbox": [ + 316, + 86, + 553, + 178 + ], + "type": "text", + "content": " # buffer predictions \n8: flag = cache_exit_short() # short-term overflow \n9: if flag and out[-N_S] not in out[-" + }, + { + "bbox": [ + 316, + 86, + 553, + 178 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 316, + 86, + 553, + 178 + ], + "type": "text", + "content": "): \n10: cache_entry(out[-N_S]) # verbalize & interleave \n11: cache_exit_long() # long-term overflow" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "code_body" + } + ], + "index": 6, + "sub_type": "algorithm" + }, + { + "bbox": [ + 313, + 200, + 555, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 200, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 200, + 555, + 248 + ], + "type": "text", + "content": "and consequently " + }, + { + "bbox": [ + 313, + 200, + 555, + 248 + ], + "type": "inline_equation", + "content": "N = (N_{L} + N_{S})" + }, + { + "bbox": [ + 313, + 200, + 555, + 248 + ], + "type": "text", + "content": " scales sub-linearly with the number of video frames. Hence, compute also scales sublinearly with the number of streaming video frames. This makes our model well-suited for low-latency applications." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 255, + 444, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 444, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 444, + 269 + ], + "type": "text", + "content": "3.5. Training ProVideLLM" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 273, + 555, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 555, + 321 + ], + "type": "text", + "content": "We adopt a two-stage training approach similar to LLaVA - (1) alignment pre-training to align the encoder's visual tokens to the decoder's language space, and (2) instruction tuning of the model for specific tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 321, + 556, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 321, + 556, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 556, + 381 + ], + "type": "text", + "content": "Stage-1: Alignment Pre-training. In this stage, our connector DETR-QFormer " + }, + { + "bbox": [ + 313, + 321, + 556, + 381 + ], + "type": "inline_equation", + "content": "(C_{VL})" + }, + { + "bbox": [ + 313, + 321, + 556, + 381 + ], + "type": "text", + "content": " is trained while keeping the visual encoder and language decoder frozen (Fig. 4). The objective is to caption short video clips featuring hand-centric actions using the standard language modeling loss [57]:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 339, + 399, + 555, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 399, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 339, + 399, + 555, + 434 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {L M}} = - \\frac {1}{| W |} \\sum_ {i = 1} ^ {| W |} \\log P (w _ {i} \\mid w _ {< i}, V; C _ {V L} ^ {\\Theta}), \\tag {2}", + "image_path": "2287ecc8f4347f18d1e3da0a6a57bfc95f0f2d6a1b112a06750b7e1340f8940c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 443, + 555, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 555, + 503 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 443, + 555, + 503 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 443, + 555, + 503 + ], + "type": "text", + "content": " is the input video clip, " + }, + { + "bbox": [ + 313, + 443, + 555, + 503 + ], + "type": "inline_equation", + "content": "W = \\{w_{1}, w_{2}, \\dots, w_{|C|}\\}" + }, + { + "bbox": [ + 313, + 443, + 555, + 503 + ], + "type": "text", + "content": " is the ground-truth caption, and " + }, + { + "bbox": [ + 313, + 443, + 555, + 503 + ], + "type": "inline_equation", + "content": "C_{VL}^{\\Theta}" + }, + { + "bbox": [ + 313, + 443, + 555, + 503 + ], + "type": "text", + "content": " is our DETR-QFormer parameterized by " + }, + { + "bbox": [ + 313, + 443, + 555, + 503 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 313, + 443, + 555, + 503 + ], + "type": "text", + "content": ". For detecting hands and corresponding objects-in-contact, we use a DETR [10]-style Generalized IoU [63] loss:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 337, + 511, + 555, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 511, + 555, + 528 + ], + "spans": [ + { + "bbox": [ + 337, + 511, + 555, + 528 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {H O}} \\left(b _ {i}, \\hat {b} _ {\\sigma (i)}\\right) = \\mathcal {L} _ {\\mathrm {i o u}} \\left(b _ {i}, \\hat {b} _ {\\sigma (i)}\\right) + \\left\\| b _ {i} - \\hat {b} _ {\\sigma (i)} \\right\\| _ {1} \\tag {3}", + "image_path": "98c86032360bf13b11ad274fbf070106e99287405d0d8d7997678fb25fec8c2f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 535, + 556, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 535, + 556, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 535, + 556, + 597 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 535, + 556, + 597 + ], + "type": "inline_equation", + "content": "b_{i}" + }, + { + "bbox": [ + 313, + 535, + 556, + 597 + ], + "type": "text", + "content": "'s are the pseudo ground-truth bounding boxes of hands and objects-in-contact generated by an off-the-shelf detector [69] and " + }, + { + "bbox": [ + 313, + 535, + 556, + 597 + ], + "type": "inline_equation", + "content": "\\hat{b}_{\\sigma(i)} \\in \\hat{B}" + }, + { + "bbox": [ + 313, + 535, + 556, + 597 + ], + "type": "text", + "content": " is the bounding box matched from the predicted set " + }, + { + "bbox": [ + 313, + 535, + 556, + 597 + ], + "type": "inline_equation", + "content": "\\hat{B}" + }, + { + "bbox": [ + 313, + 535, + 556, + 597 + ], + "type": "text", + "content": ". For simplicity, we keep our loss category-agnostic. The final loss for Stage-1 is thus:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 393, + 607, + 555, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 607, + 555, + 620 + ], + "spans": [ + { + "bbox": [ + 393, + 607, + 555, + 620 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {L M}} + \\lambda_ {1} \\mathcal {L} _ {\\mathrm {H O}} \\tag {4}", + "image_path": "ad249813513697bc5513ee7b1f5a1b0047f0be98bb3728f5f16edf70693b0a56.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 629, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 629, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 313, + 629, + 555, + 665 + ], + "type": "text", + "content": " is a hyper-parameter. In addition to vision-language alignment, Stage-1 also enables our model to attend to hand-object interactions." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "Stage-2: Instruction Tuning. This stage focuses on fine-tuning our model with task-specific instructions such as \"Provide the goal.\", \"Provide what I am doing now.\", \"Provide what I need to do next.\", or \"Provide the" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22590" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 291, + 142 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 291, + 142 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 291, + 142 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 291, + 142 + ], + "type": "table", + "html": "
Modelper-frame mAP
ValTest
LSTR [86] (only short-term)8.8-
LSTR [86]8.98.1
EgoOnly [78]10.310.9
ProVideLLM-1B/5 (no verbalization)12.112.2
ProVideLLM-1B/5 (verbalization & interleaving)13.012.9
", + "image_path": "756e247997a4ec65a49a922d93ae3303dffa8cb6a015736a0156aed65e7b652b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 61, + 195, + 290, + 245 + ], + "blocks": [ + { + "bbox": [ + 55, + 151, + 295, + 185 + ], + "lines": [ + { + "bbox": [ + 55, + 151, + 295, + 185 + ], + "spans": [ + { + "bbox": [ + 55, + 151, + 295, + 185 + ], + "type": "text", + "content": "Table 1. Ego4D [72] GoalStep Online Step Detection. For fair comparison with LSTR, we consider 16s of short-term past and an avg. of 128s of long-term past compressed into language tokens." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 195, + 290, + 245 + ], + "lines": [ + { + "bbox": [ + 61, + 195, + 290, + 245 + ], + "spans": [ + { + "bbox": [ + 61, + 195, + 290, + 245 + ], + "type": "table", + "html": "
ModelViewsAction
TempAgg [66]12 views8.5
+ Goal [88]12 views12.0
ProVideLLM-8B/11 (+long-term verbalization)v413.8
", + "image_path": "529a713b1c92f4e136c8b3b0559b4d45a8a17912991a932bb243e461548128fd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 297, + 296, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 296, + 370 + ], + "type": "text", + "content": "next 10 steps I need to do.\" For instruction tuning, we fine-tune DETR-QFormer and the language decoder using " + }, + { + "bbox": [ + 55, + 297, + 296, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{LM}}" + }, + { + "bbox": [ + 55, + 297, + 296, + 370 + ], + "type": "text", + "content": " while keeping the visual encoder frozen. For streaming tasks, we also train our model with interleaved special " + }, + { + "bbox": [ + 55, + 297, + 296, + 370 + ], + "type": "inline_equation", + "content": "<\\mathbb{L}>" + }, + { + "bbox": [ + 55, + 297, + 296, + 370 + ], + "type": "text", + "content": " tokens to represent long-term cache entries, distinguishing them from standard prompts. See Suppl.Sec.5 for further details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 381, + 137, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 381, + 137, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 381, + 137, + 396 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 402, + 296, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 296, + 485 + ], + "type": "text", + "content": "We conduct extensive evaluations of ProVideLLM on six tasks across four datasets covering streaming (Sec. 4.1.1), long-term (Sec. 4.1.2) and fine-grained (Sec. 4.1.3) procedural understanding. We further assess ProVideLLM's performance in multi-task settings and its generalization to new datasets in Sec. 4.2. Finally, we provide ablations in Sec. 4.3 and ProVideLLM's runtime analysis in Sec. 4.4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "content": "Implementation Details. Depending on the choice of vision encoder and language decoder, we have two model variants. ProVideLLM-1B/5 is our efficiency-focused model designed for real-time streaming. It uses DINOv2 (ViT-S) with Llama-3.2-1B-Instruct [20] and requires 5 tokens per frame. ProVideLLM-8B/11 is our more performance-focused model using DINOv2 (ViT-g) and Llama-3.1-8B-Instruct and requires 11 tokens per frame. Furthermore, for benchmarking on youtube videos [74], we introduce an additional model ProVideLLM-8B/11+, which uses SigLIP's [CLS] token along with DINOv2 patch tokens. Unless otherwise mentioned we use " + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "inline_equation", + "content": "N_{S} = 64" + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "inline_equation", + "content": "N_{L} = 5" + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "content": " i.e. storing a maximum of 64 frames in the short-term cache and a maximum of 5 observed steps in the long-term cache. DETR-QFormer is pre-trained (Stage-1) on EPIC-KITCHENS [16] with pseudo ground-truth bounding boxes from 100DOH [69]. For Stage-2, the LLM (decoder) is fine-tuned using LoRA [32] with " + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "inline_equation", + "content": "r = 128" + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\alpha = 256" + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "content": ". Additional implementation details are provided in Suppl.Sec.5." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 321, + 70, + 548, + 182 + ], + "blocks": [ + { + "bbox": [ + 55, + 253, + 295, + 277 + ], + "lines": [ + { + "bbox": [ + 55, + 253, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 295, + 277 + ], + "type": "text", + "content": "Table 2. Action Anticipation Results on Assembly101 validation set. Performance measured by class-mean Top-5 recall." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 70, + 548, + 182 + ], + "lines": [ + { + "bbox": [ + 321, + 70, + 548, + 182 + ], + "spans": [ + { + "bbox": [ + 321, + 70, + 548, + 182 + ], + "type": "table", + "html": "
MethodTop-1 Acc. (%)
TimeSformer [9]34.0
DistantSup39.4
VideoTF42.4
ProcedureVRL46.8
VEDIT [42]51.8
VideoLLM-online [14]49.1
VideoLLM-MoD [83]49.7
ProVideoLLM-8B/11+52.1
ProVideoLLM-8B/11+ (+long-term verbalization)53.6
", + "image_path": "2168c69feb0267a6780bc07419107c6fd69c91e0040864930d1783fe1cbeaaa0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 190, + 553, + 213 + ], + "lines": [ + { + "bbox": [ + 313, + 190, + 553, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 190, + 553, + 213 + ], + "type": "text", + "content": "Table 3. Next Step Forecasting in COIN. The last four rows including ProVideLLM are VideoLLMs." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 234, + 555, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 555, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 555, + 306 + ], + "type": "text", + "content": "Datasets. We evaluate our model's performance on four large-scale procedural datasets - EgoExo4D [28], Ego4D-GoalStep [72], COIN [74], and Assembly101 [67]. As shown in Sec. 4.1, ProVideLLM sets state-of-the-art on both egocentric and YouTube benchmarks, demonstrating robustness across diverse video domains." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 313, + 397, + 325 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 313, + 397, + 325 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 397, + 325 + ], + "type": "text", + "content": "4.1. Main results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 331, + 436, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 436, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 436, + 342 + ], + "type": "text", + "content": "4.1.1. Online Step Detection" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 346, + 556, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 556, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 556, + 562 + ], + "type": "text", + "content": "We report results in Tab. 1, using per-frame mAP on action classes (steps + substeps) at 4fps, a standard metric in online action detection [79, 97]. We evaluate ProVideLLM-1B/5 with and without long-term verbalization in rows four and five, respectively. Results for LSTR [86] and EgoOnly [78] are taken from [72]. Accordingly, LSTR spans " + }, + { + "bbox": [ + 313, + 346, + 556, + 562 + ], + "type": "inline_equation", + "content": "128s" + }, + { + "bbox": [ + 313, + 346, + 556, + 562 + ], + "type": "text", + "content": " in long-term memory and " + }, + { + "bbox": [ + 313, + 346, + 556, + 562 + ], + "type": "inline_equation", + "content": "16s" + }, + { + "bbox": [ + 313, + 346, + 556, + 562 + ], + "type": "text", + "content": " in short-term memory. In [72], each step spans on average " + }, + { + "bbox": [ + 313, + 346, + 556, + 562 + ], + "type": "inline_equation", + "content": "32s" + }, + { + "bbox": [ + 313, + 346, + 556, + 562 + ], + "type": "text", + "content": " with around 5.7 tokens representing each step. Our long-term cache can encode " + }, + { + "bbox": [ + 313, + 346, + 556, + 562 + ], + "type": "inline_equation", + "content": "128s" + }, + { + "bbox": [ + 313, + 346, + 556, + 562 + ], + "type": "text", + "content": " of past using only 22 tokens at " + }, + { + "bbox": [ + 313, + 346, + 556, + 562 + ], + "type": "inline_equation", + "content": "N_{L} = 4" + }, + { + "bbox": [ + 313, + 346, + 556, + 562 + ], + "type": "text", + "content": " steps. To measure improvements from our long-term design, we set the cache length " + }, + { + "bbox": [ + 313, + 346, + 556, + 562 + ], + "type": "inline_equation", + "content": "N = 86" + }, + { + "bbox": [ + 313, + 346, + 556, + 562 + ], + "type": "text", + "content": " (spanning LSTR's temporal context), split into " + }, + { + "bbox": [ + 313, + 346, + 556, + 562 + ], + "type": "inline_equation", + "content": "(N_{S} = 86, N_{L} = 0)" + }, + { + "bbox": [ + 313, + 346, + 556, + 562 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 346, + 556, + 562 + ], + "type": "inline_equation", + "content": "(N_{S} = 64, N_{L} = 4)" + }, + { + "bbox": [ + 313, + 346, + 556, + 562 + ], + "type": "text", + "content": " in rows four and five, respectively. Our model surpasses both baselines, outperforming LSTR by over " + }, + { + "bbox": [ + 313, + 346, + 556, + 562 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 313, + 346, + 556, + 562 + ], + "type": "text", + "content": " at similar temporal span. ProVideLLM also achieves greater gains from long-term context compared to LSTR, highlighting the effectiveness of verbalizing the long-term past." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 567, + 446, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 446, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 446, + 578 + ], + "type": "text", + "content": "4.1.2. Online Step Forecasting" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": "Step forecasting requires modeling long-term dependencies, making it an ideal task to evaluate our contributions. We evaluate ProVideoLLM for step forecasting on COIN and Assembly101 measured by Top-1 accuracy and Top-5 Recall [24], respectively. Our model achieves state-of-the-art performance on both datasets as shown in Tab. 3 and Tab. 2. On Assembly101, we surpass [88] by " + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": " using only a single view. On COIN, we use Llama-3-8B-Instruct to ensure fair comparisons with VideoLLMs [14, 83] and achieve a performance boost of around " + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": ". Moreover, while state-of-the-art models [14, 83] use approximately 600 to 6000 visual" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22591" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 70, + 288, + 169 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 288, + 169 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 288, + 169 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 288, + 169 + ], + "type": "table", + "html": "
ModelTraining ViewsEgo Acc. (%) Val / Test
TimeSFformer [9]ego35.13 / 35.93
EgoVLPv2 [56]ego+exo39.10 / 38.76
Viewpoint Distillationego+exo38.19 / 39.49
View Invariant Encoder [51]ego+exo40.34 / 41.53
VideoLLM-MoD† [83]ego42.62 / —
ProVideLLM-8B/11ego44.36 / 50.74
", + "image_path": "685a07c5e829e6d6941fd443b32352218831a5414cfa2d31acb37091b1a178b1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 64, + 222, + 288, + 332 + ], + "blocks": [ + { + "bbox": [ + 55, + 178, + 297, + 212 + ], + "lines": [ + { + "bbox": [ + 55, + 178, + 297, + 212 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 297, + 212 + ], + "type": "text", + "content": "Table 4. EgoExo4D Fine-grained Keystep Recognition. Test results obtained from EgoExo4D's Fine-grained Keystep Recognition Challenge on EvalAI. † Uses 2fps sampling per segment." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 64, + 222, + 288, + 332 + ], + "lines": [ + { + "bbox": [ + 64, + 222, + 288, + 332 + ], + "spans": [ + { + "bbox": [ + 64, + 222, + 288, + 332 + ], + "type": "table", + "html": "
MethodTop-1 Accuracy (%)
TimeSformer [9]34.0
Paprika [100]51.0
DistantSup [43]54.1
ProcedureVRL [99]56.9
VideoTaskGraph [4]57.2
VEDIT [42]62.7
VideoLLM-online [14]63.1
VideoLLM-MoD [83]63.4
ProVideLLM-8B/11+67.3
", + "image_path": "842ca40dc5205173594da70fb3ad48eb65bf7316e65b7cbc9ad6f159ac273e1a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 365, + 296, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 296, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 296, + 389 + ], + "type": "text", + "content": "tokens to model past context, our approach spans a larger temporal context with only 200 tokens through verbalization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 396, + 157, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 157, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 157, + 407 + ], + "type": "text", + "content": "4.1.3. Step Recognition" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 411, + 297, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 411, + 297, + 555 + ], + "spans": [ + { + "bbox": [ + 54, + 411, + 297, + 555 + ], + "type": "text", + "content": "In Tab. 4 and Tab. 5, we evaluate ProVideLLM on fine-grained step recognition on EgoExo4D and COIN. We uniformly sample 16 frames from each segment to construct our short-term cache " + }, + { + "bbox": [ + 54, + 411, + 297, + 555 + ], + "type": "inline_equation", + "content": "(N_S)" + }, + { + "bbox": [ + 54, + 411, + 297, + 555 + ], + "type": "text", + "content": " and use the prediction from the last frame as our output. The closest model, VideoLLM-MoD [83], uses 10 tokens per frame, but densely samples frames at 2fps, resulting in a larger token count. In both cases, we outperform VideoLLM-MoD by " + }, + { + "bbox": [ + 54, + 411, + 297, + 555 + ], + "type": "inline_equation", + "content": "1.74\\%" + }, + { + "bbox": [ + 54, + 411, + 297, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 411, + 297, + 555 + ], + "type": "inline_equation", + "content": "3.9\\%" + }, + { + "bbox": [ + 54, + 411, + 297, + 555 + ], + "type": "text", + "content": " respectively. Moreover, at the time of our submission, our method, using only egocentric video, achieved the best results on the EgoExo4D Fine-grained Recognition Challenge, with a " + }, + { + "bbox": [ + 54, + 411, + 297, + 555 + ], + "type": "inline_equation", + "content": "9.2\\%" + }, + { + "bbox": [ + 54, + 411, + 297, + 555 + ], + "type": "text", + "content": " improvement over the second-ranked model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 563, + 263, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 263, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 263, + 577 + ], + "type": "text", + "content": "4.2. Generalized Procedural Understanding" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "type": "text", + "content": "In this section, we explore additional capabilities of ProVideLLM, leveraging the advantages of using a pretrained LLM as the language decoder. This section provides insights into developing a \"generalist\" model that advance procedural video understanding within a unified framework. We first evaluate ProVideLLM's ability to multi-task and then present our benchmark for cross dataset generalization. Multi-task learning. In Tab. 6, we present our results for multi-task learning, an underexplored aspect of procedural understanding. Existing models are trained only for isolated tasks, such as step recognition and forecasting. We compare" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 323, + 70, + 547, + 140 + ], + "blocks": [ + { + "bbox": [ + 55, + 340, + 295, + 352 + ], + "lines": [ + { + "bbox": [ + 55, + 340, + 295, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 295, + 352 + ], + "type": "text", + "content": "Table 5. COIN Step Recognition. Last three rows are VideoLLMs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 70, + 547, + 140 + ], + "lines": [ + { + "bbox": [ + 323, + 70, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 323, + 70, + 547, + 140 + ], + "type": "table", + "html": "
ModelCOINBenchmark Top-1 Accuracy (%)
StepTaskNextProc.Proc.+
VideoTF† [50]56.591.042.440.246.4
VideoLLM-online [14]63.192.749.149.854.1
VideoLLM-MoD [83]63.492.849.749.853.3
ProVideLLM-8B/11+66.995.050.551.055.9
", + "image_path": "fde3747cd30d1adc6607b4a3ebc01776d38be6d4a36841888e357254ab848971.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 323, + 202, + 547, + 255 + ], + "blocks": [ + { + "bbox": [ + 313, + 148, + 555, + 192 + ], + "lines": [ + { + "bbox": [ + 313, + 148, + 555, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 148, + 555, + 192 + ], + "type": "text", + "content": "Table 6. Multi-task learning on COIN. Step - step recognition, Task - task recognition, Next - next step forecasting, Proc. - long-term forecasting and " + }, + { + "bbox": [ + 313, + 148, + 555, + 192 + ], + "type": "inline_equation", + "content": "\\mathrm{Proc + }" + }, + { + "bbox": [ + 313, + 148, + 555, + 192 + ], + "type": "text", + "content": " - long-term forecasting with task. " + }, + { + "bbox": [ + 313, + 148, + 555, + 192 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 313, + 148, + 555, + 192 + ], + "type": "text", + "content": " VideoTF [50] is trained separately for each task." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 202, + 547, + 255 + ], + "lines": [ + { + "bbox": [ + 323, + 202, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 323, + 202, + 547, + 255 + ], + "type": "table", + "html": "
ModelEgo Acc. (%)
VideoLLaVA [41] + Llama-3.1-8B [20]3.6
LLaVA-OneVision [38] + Llama-3.1-8B [20]8.3
ProVideLLM-8B/1112.4
", + "image_path": "ca672561f95c906bbbc940c4839e43453601a5f125e358d20c200dff660f072a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 263, + 555, + 297 + ], + "lines": [ + { + "bbox": [ + 313, + 263, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 555, + 297 + ], + "type": "text", + "content": "Table 7. Cross-dataset generalization. Evaluated on EgoExo4D KeyStep recognition. Our model, trained on Ego4D Goal-Step is compared with a mixture of vision + language experts." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 318, + 556, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 556, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 556, + 510 + ], + "type": "text", + "content": "our model with recent VideoLLMs, also trained in a multitask setting, on 5 tasks. Despite using the same LLM as [14, 83], our model outperforms these baselines across all tasks. Cross-Dataset Generalization. In Tab. 7, we present a cross-dataset generalization benchmark for procedural videos using the EgoExo4D fine-grained keystep recognition task. We compare two recent VideoLLMs [38, 41] which were primarily trained for video captioning and use Llama-3.1-8B-Instruct to convert the captions into EgoExo4D classes. Our baseline models act as a combination of visual and language experts. ProVideLLM, using the same LLM and trained on the much smaller Ego4D Goal-Step dataset, significantly outperforms these baselines. Despite these advances, performance remains modest, highlighting the need for further research in procedural generalization. Additional details are provided in the Suppl. Material." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 517, + 383, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 383, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 383, + 528 + ], + "type": "text", + "content": "4.3. Ablations" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "text", + "content": "Visual Encoders & DETR-QFormer. In Tab. 8, we ablate visual encoders-SigLIP and DINOv2-and connectorsMLP, QFormer, and our DETR-QFormer-on EgoExo4D keystep recognition. First, comparing encoders across connectors, SigLIP's CLS tokens, being language-aligned, outperform DINOv2; however, with more patch tokens, DINOv2 surpasses SigLIP, aligning with our observations in Fig. 3. Second, comparing connectors across encoders, Q-Former boosts DINOv2's performance " + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "inline_equation", + "content": "(+1.2\\%, +1.9\\%)" + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "text", + "content": " but has minimal impact on SigLIP " + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "inline_equation", + "content": "(+0.8\\%, -0.1\\%)" + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "text", + "content": ", highlighting DINOv2's need for a stronger connector. Finally, DETR-QFormer further improves DINOv2 " + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "inline_equation", + "content": "(+1.1\\%, +2.6\\%)" + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "text", + "content": ", with gains increasing at higher resolutions but degrading SigLIP's performance due to weak hand-object activations in its patch tokens. Notably, DETR-QFormer remains highly" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22592" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 70, + 291, + 216 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 291, + 216 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 291, + 216 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 291, + 216 + ], + "type": "table", + "html": "
[CLS] TokenPatch Tokens#Tokens /frameEgo Val./ Acc. (%)
SigLIPDINOv2
w/ 2-layer MLP [45]
✓X131.8±0.428.7±0.2
✓16×16 → 2×2532.1±0.632.4±0.5
✓16×16 → 4×41733.3±0.636.6±0.5
w/ QFormer [39]
✓16×16 ⇒ 4532.9±0.633.6±0.5
✓16×16 ⇒ 161733.2±0.438.1±0.7
w/ DETR-QFormer (2 hands, 2 objects)
✓16×16 ⇒ 0530.9±0.334.7±0.5
✓16×16 ⇒ 121733.6±0.340.7±0.6
", + "image_path": "6e6e34a8305e7b2c824986f4238cab6e903f0518efd064c15a0273bda013fadf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 288, + 297, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 297, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 297, + 324 + ], + "type": "text", + "content": "token-efficient with DINOv2, even when compressing to only hand-object queries, reinforcing its advantage for low-latency streaming applications." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 325, + 297, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 325, + 297, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 325, + 297, + 553 + ], + "type": "text", + "content": "Verbalization & Interleaving. In Fig. 5, we analyze our verbalization and interleaving strategies using ProVideLLM-1B/5 on Ego4D Goal-Step validation set, selecting a video containing over 15 steps and spanning " + }, + { + "bbox": [ + 55, + 325, + 297, + 553 + ], + "type": "inline_equation", + "content": ">30" + }, + { + "bbox": [ + 55, + 325, + 297, + 553 + ], + "type": "text", + "content": " minutes. At 4fps sampling, we ablate on a fixed memory budget of 8GB on an A6000 GPU. Short-term cache length is set as " + }, + { + "bbox": [ + 55, + 325, + 297, + 553 + ], + "type": "inline_equation", + "content": "N_{S} = 64" + }, + { + "bbox": [ + 55, + 325, + 297, + 553 + ], + "type": "text", + "content": " (16 seconds), and no limit is imposed on the long-term cache. Using only visual tokens (red), memory scales linearly with video length, reaching 40K tokens to represent 20 minutes of video before exceeding 8GB memory. Such massive context also degrades performance. Verbalization (orange) compresses 20 minutes of video into 2.5K tokens, reducing memory usage to 3GB and achieving sub-linear scaling. However, it introduces runtime instability, with frequent spikes in per-frame prediction time due to attention recomputation when converting short-term visual tokens into long-term text. Interleaving (green) in our multimodal cache mitigates these spikes, improving runtime stability while maintaining sub-linear scaling in both memory and compute." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 563, + 162, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 162, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 162, + 576 + ], + "type": "text", + "content": "4.4. Runtime Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "type": "text", + "content": "In Tab. 9, we report the per-frame streaming inference and streaming dialogue [14] performance of ProVideLLM variants on a single A6000 GPU. Our lightest variant, ProVideLLM-1B/5, achieves over 10 FPS with a minimal 2GB memory footprint, making it ideal for low-latency applications. Meanwhile, ProVideLLM-8B/11 offer higher performance but require more compute. For most videos in this dataset, ProVideLLM-1B/5 consistently maintains over 10 FPS, sometimes even reaching 20 FPS when the video has many no-action frames. For per-frame predictions, a key bottleneck for VideoLLMs is their autoregressive text gener" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 553, + 145 + ], + "blocks": [ + { + "bbox": [ + 55, + 223, + 297, + 269 + ], + "lines": [ + { + "bbox": [ + 55, + 223, + 297, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 297, + 269 + ], + "type": "text", + "content": "Table 8. DETR-QFormer ablations on EgoExo4D fine-grained keysteps. Llama-3.2-1B-Instruct is used as the language decoder. Results reported over 3 runs. " + }, + { + "bbox": [ + 55, + 223, + 297, + 269 + ], + "type": "inline_equation", + "content": "(\\rightarrow" + }, + { + "bbox": [ + 55, + 223, + 297, + 269 + ], + "type": "text", + "content": " means avg-pooling, " + }, + { + "bbox": [ + 55, + 223, + 297, + 269 + ], + "type": "inline_equation", + "content": "\\Rightarrow" + }, + { + "bbox": [ + 55, + 223, + 297, + 269 + ], + "type": "text", + "content": " means Q-Former compression.)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 70, + 553, + 145 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 145 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 145 + ], + "type": "table", + "html": "
Streaming PredictionsModelFPS (↑)Memory (GB) (↓)
Vision Enc.Connector + LLMFull Model
Per-frameProVideLLM-1B/574.810.49.12.0
ProVideLLM-8B/1138.04.23.516.2
Dialogue [14]ProVideLLM-1B/574.834.224.62.2
ProVideLLM-8B/1138.023.717.216.9
", + "image_path": "df08352eb241b1f2ad870d76d27acff468840a35df3fb639ed4836948392ae95.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 153, + 555, + 186 + ], + "lines": [ + { + "bbox": [ + 313, + 153, + 555, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 153, + 555, + 186 + ], + "type": "text", + "content": "Table 9. Runtime Analysis on a single A6000 GPU for the validation set of Ego4D-GoalStep online step detection for per-frame streaming and Ego4D-GoalStep Narration [83]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 315, + 196, + 553, + 268 + ], + "blocks": [ + { + "bbox": [ + 315, + 196, + 553, + 268 + ], + "lines": [ + { + "bbox": [ + 315, + 196, + 553, + 268 + ], + "spans": [ + { + "bbox": [ + 315, + 196, + 553, + 268 + ], + "type": "image", + "image_path": "e84445eb7c0fe0969400e2c46db1304bd111e84bd51298edacaa12083b12bf59.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 276, + 555, + 309 + ], + "lines": [ + { + "bbox": [ + 313, + 276, + 555, + 309 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 555, + 309 + ], + "type": "text", + "content": "Figure 5. Scaling of memory, runtime and context length under various caching strategies. Results reported on a Ego4D-GoalStep video at a memory budget of 8GB on a single A6000 GPU." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 312, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 384 + ], + "type": "text", + "content": "ation, adding around 6 extra iterations, corresponding to 5.7 tokens/(activity step) on average for Ego4D-GoalStep. This issue is mitigated with streaming dialogue [14], where the model is trained to remain silent and respond to user queries only when necessary. In this setting, ProVideLLM-1B/5 delivers real-time performance at 24.6 FPS." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 394, + 471, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 471, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 471, + 407 + ], + "type": "text", + "content": "5. Conclusion and Limitations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 415, + 556, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 556, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 556, + 617 + ], + "type": "text", + "content": "In this paper, we presented ProVideLLM for real-time understanding of procedural videos by interleaving two types of tokens in a vision-text multimodal cache - verbalized text tokens, which provide compressed summaries of long-term observations, and visual tokens, which capture fine-grained details from short-term observations. We show that ProVideLLM reduces the token count for representing long-form procedural videos while effectively capturing critical spatio-temporal dependencies. We demonstrated state-of-the-art results on six tasks across four datasets, showing per-frame inference at 10 FPS and streaming dialogue at 25 FPS. ProVideLLM is the first real-time framework unifying recognition, anticipation, and planning for procedural assistance. Our multimodal interleaved cache is also generic, applicable to non-procedural videos, and unlocking a new area of compressing long videos on the fly while reducing redundancy and boosting performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 557, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 557, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 557, + 713 + ], + "type": "text", + "content": "Limitations. DETR-QFormer, trained on pseudo ground-truth bounding boxes, generalizes well but could benefit from GT data to reduce biases from noisy annotations. Additionally ProVideLLM focuses on detection but lacks real-time feedback on the \"how\" of task execution. Finally, as discussed in Sec. 4.2, ProVideLLM's generalization performance remains modest, presenting a potential direction for future work." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22593" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 62, + 91, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 91, + 297, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 91, + 297, + 145 + ], + "spans": [ + { + "bbox": [ + 66, + 91, + 297, + 145 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 148, + 296, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 148, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 66, + 148, + 296, + 192 + ], + "type": "text", + "content": "[2] Ibrahim M Alabdulmohsin, Xiaohua Zhai, Alexander Kolesnikov, and Lucas Beyer. Getting vit in shape: Scaling laws for compute-optimal model design. Advances in Neural Information Processing Systems, 36, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 194, + 296, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 194, + 296, + 247 + ], + "spans": [ + { + "bbox": [ + 66, + 194, + 296, + 247 + ], + "type": "text", + "content": "[3] Kumar Ashutosh, Santhosh Kumar Ramakrishnan, Triantafyllos Afouras, and Kristen Grauman. Video-mined task graphs for keystep recognition in instructional videos. Advances in Neural Information Processing Systems, 36: 67833-67846, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 250, + 296, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 250, + 296, + 303 + ], + "spans": [ + { + "bbox": [ + 66, + 250, + 296, + 303 + ], + "type": "text", + "content": "[4] Kumar Ashutosh, Santhosh Kumar Ramakrishnan, Triantafyllos Afouras, and Kristen Grauman. Video-mined task graphs for keystep recognition in instructional videos. Advances in Neural Information Processing Systems, 36, 2024. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 306, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 306, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 66, + 306, + 294, + 350 + ], + "type": "text", + "content": "[5] Fabien Baradel, Natalia Neverova, Christian Wolf, Julien Mille, and Greg Mori. Object level visual reasoning in videos. In Proceedings of the European Conference on Computer Vision (ECCV), pages 105-121, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 352, + 296, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 352, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 66, + 352, + 296, + 384 + ], + "type": "text", + "content": "[6] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 387, + 296, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 387, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 66, + 387, + 296, + 453 + ], + "type": "text", + "content": "[7] Yizhak Ben-Shabat, Xin Yu, Fatemeh Saleh, Dylan Campbell, Cristian Rodriguez-Opazo, Hongdong Li, and Stephen Gould. The ikaa asm dataset: Understanding people assembling furniture through actions, objects and pose. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 847–859, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 454, + 296, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 454, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 66, + 454, + 296, + 518 + ], + "type": "text", + "content": "[8] Yizhak Ben-Shabat, Jonathan Paul, Eviatar Segev, Oren Shrout, and Stephen Gould. Ika ego 3d dataset: Understanding furniture assembly actions from ego-view 3d point clouds. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4355-4364, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 521, + 294, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 521, + 294, + 554 + ], + "spans": [ + { + "bbox": [ + 66, + 521, + 294, + 554 + ], + "type": "text", + "content": "[9] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 556, + 296, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 556, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 62, + 556, + 296, + 610 + ], + "type": "text", + "content": "[10] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 4, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 613, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 613, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 62, + 613, + 296, + 667 + ], + "type": "text", + "content": "[11] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 296, + 713 + ], + "type": "text", + "content": "[12] Chien-Yi Chang, De-An Huang, Danfei Xu, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. Procedure planning in instructional videos. In European Conference on Computer Vision, pages 334–350. Springer, 2020. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 320, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 555, + 116 + ], + "type": "text", + "content": "[13] Dibyadip Chatterjee, Fadime Sener, Shugao Ma, and Angela Yao. Opening the vocabulary of egocentric actions. Advances in Neural Information Processing Systems, 36: 33174-33187, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 119, + 555, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 119, + 555, + 185 + ], + "spans": [ + { + "bbox": [ + 321, + 119, + 555, + 185 + ], + "type": "text", + "content": "[14] Joya Chen, Zhaoyang Lv, Shiwei Wu, Kevin Qinghong Lin, Chenan Song, Difei Gao, Jia-Wei Liu, Ziteng Gao, Dongxing Mao, and Mike Zheng Shou. Videollm-online: Online video large language model for streaming video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18407-18418, 2024. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 186, + 555, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 186, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 321, + 186, + 555, + 251 + ], + "type": "text", + "content": "[15] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 254, + 555, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 254, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 321, + 254, + 555, + 319 + ], + "type": "text", + "content": "[16] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision, pages 1-23, 2022. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 321, + 555, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 321, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 321, + 321, + 555, + 386 + ], + "type": "text", + "content": "[17] Roeland De Geest, Efstratios Gavves, Amir Ghodrati, Zhenyang Li, Cees Snoek, and Tinne Tuytelaars. Online action detection. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14, pages 269-284. Springer, 2016. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 388, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 388, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 321, + 388, + 553, + 420 + ], + "type": "text", + "content": "[18] Guodong Ding, Fadime Sener, Shugao Ma, and Angela Yao. Every mistake counts in assembly. arXiv preprint arXiv:2307.16453, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 422, + 555, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 422, + 555, + 465 + ], + "spans": [ + { + "bbox": [ + 321, + 422, + 555, + 465 + ], + "type": "text", + "content": "[19] Guodong Ding, Fadime Sener, and Angela Yao. Temporal action segmentation: An analysis of modern techniques. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 468, + 555, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 468, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 321, + 468, + 555, + 521 + ], + "type": "text", + "content": "[20] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 524, + 555, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 524, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 321, + 524, + 555, + 578 + ], + "type": "text", + "content": "[21] Nikita Dvornik, Isma Hadji, Hai Pham, Dhaivat Bhatt, Brais Martinez, Afsaneh Fazly, and Allan D Jepson. Flow graph to video grounding for weakly-supervised multi-step localization. In European Conference on Computer Vision, pages 319-335. Springer, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 581, + 555, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 581, + 555, + 656 + ], + "spans": [ + { + "bbox": [ + 321, + 581, + 555, + 656 + ], + "type": "text", + "content": "[22] Alessandro Flaborea, Guido Maria D'Amely di Melendugno, Leonardo Plini, Luca Scofano, Edoardo De Matteis, Antonio Furnari, Giovanni Maria Farinella, and Fabio Galasso. Prego: online mistake detection in procedural egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18483-18492, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 658, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 658, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 321, + 658, + 555, + 713 + ], + "type": "text", + "content": "[23] Antonino Furnari and Giovanni Maria Farinella. What would you expect? anticipating egocentric actions with rolling-unrolling lstms and modality attention. In Proceedings of the IEEE/CVF International conference on computer vision, pages 6252-6261, 2019. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22594" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 73, + 296, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 296, + 127 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 296, + 127 + ], + "type": "text", + "content": "[24] Antonino Furnari, Sebastiano Battiato, and Giovanni Maria Farinella. Leveraging uncertainty to rethink loss functions and evaluation measures for egocentric action anticipation. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 128, + 296, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 128, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 61, + 128, + 296, + 183 + ], + "type": "text", + "content": "[25] Mingfei Gao, Mingze Xu, Larry S Davis, Richard Socher, and Caiming Xiong. Startnet: Online detection of action start in untrimmed videos. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5542-5551, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 184, + 296, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 184, + 296, + 226 + ], + "spans": [ + { + "bbox": [ + 62, + 184, + 296, + 226 + ], + "type": "text", + "content": "[26] Rohit Girdhar and Kristen Grauman. Anticipative video transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13505-13515, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 228, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 228, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 228, + 296, + 293 + ], + "type": "text", + "content": "[27] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 294, + 296, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 294, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 62, + 294, + 296, + 370 + ], + "type": "text", + "content": "[28] Kristen Grauman, Andrew Westbury, Lorenzo Torresani, Kris Kitani, Jitendra Malik, Triantafyllos Afouras, Kumar Ashutosh, Vijay Baiyya, Siddhant Bansal, Bikram Boote, et al. Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19383-19400, 2024. 2, 3, 4, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 371, + 296, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 371, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 62, + 371, + 296, + 415 + ], + "type": "text", + "content": "[29] Bo He, Hengduo Li, Young Kyun Jang, Menglin Jia, Xuefei Cao, Ashish Shah, Abhinav Shrivastava, and Ser-Nam Lim. Ma-lmm: Memory-augmented large multimodal model for long-term video understanding. CVPR, 2024. 1, 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 415, + 296, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 296, + 469 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 296, + 469 + ], + "type": "text", + "content": "[30] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 471, + 296, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 471, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 62, + 471, + 296, + 525 + ], + "type": "text", + "content": "[31] Cheng-Ping Hsieh, Simeng Sun, Samuel Kriman, Shantanu Acharya, Dima Rekesh, Fei Jia, Yang Zhang, and Boris Ginsburg. Ruler: What's the real context size of your long-context language models? arXiv preprint arXiv:2404.06654, 2024. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 526, + 296, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 526, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 62, + 526, + 296, + 579 + ], + "type": "text", + "content": "[32] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 581, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 581, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 581, + 296, + 635 + ], + "type": "text", + "content": "[33] Md Mohaiminul Islam, Tushar Nagarajan, Huiyu Wang, Fu-Jen Chu, Kris Kitani, Gedas Bertasius, and Xitong Yang. Propose, assess, search: Harnessing llms for goal-oriented planning in instructional videos. arXiv preprint arXiv:2409.20557, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 636, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 636, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 62, + 636, + 296, + 679 + ], + "type": "text", + "content": "[34] Y.-G. Jiang, J. Liu, A. Roshan Zamir, G. Toderici, I. Laptev, M. Shah, and R. Sukthankar. THUMOS challenge: Action recognition with a large number of classes. http://crcv.ucf.edu/THUMOS14/, 2014.1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 681, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 681, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 681, + 296, + 712 + ], + "type": "text", + "content": "[35] Mahnaz Koupae and William Yang Wang. Wikihow: A large scale text summarization dataset. arXiv preprint arXiv:1810.09305, 2018. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 320, + 73, + 555, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 555, + 117 + ], + "type": "text", + "content": "[36] Bolin Lai, Sam Toyer, Tushar Nagarajan, Rohit Girdhar, Shengxin Zha, James M Rehg, Kris Kitani, Kristen Grauman, Ruta Desai, and Miao Liu. Human action anticipation: A survey. arXiv preprint arXiv:2410.14045, 2024. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 320, + 118, + 555, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 118, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 320, + 118, + 555, + 172 + ], + "type": "text", + "content": "[37] Shih-Po Lee, Zijia Lu, Zekun Zhang, Minh Hoai, and Ehsan Elhamifar. Error detection in egocentric procedural task videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18655-18666, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 175, + 555, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 175, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 320, + 175, + 555, + 217 + ], + "type": "text", + "content": "[38] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 2, 3, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 220, + 555, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 220, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 320, + 220, + 555, + 274 + ], + "type": "text", + "content": "[39] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2, 4, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 276, + 555, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 276, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 320, + 276, + 555, + 319 + ], + "type": "text", + "content": "[40] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 321, + 555, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 321, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 320, + 321, + 555, + 364 + ], + "type": "text", + "content": "[41] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 365, + 555, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 365, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 320, + 365, + 555, + 419 + ], + "type": "text", + "content": "[42] Han Lin, Tushar Nagarajan, Nicolas Ballas, Mido Assran, Mojtaba Komeili, Mohit Bansal, and Koustuv Sinha. Vedit: Latent prediction architecture for procedural video representation learning. arXiv preprint arXiv:2410.03478, 2024. 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 422, + 555, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 422, + 555, + 486 + ], + "spans": [ + { + "bbox": [ + 320, + 422, + 555, + 486 + ], + "type": "text", + "content": "[43] Xudong Lin, Fabio Petroni, Gedas Bertasius, Marcus Rohrbach, Shih-Fu Chang, and Lorenzo Torresani. Learning to recognize procedural activities with distant supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13853-13863, 2022. 2, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 489, + 555, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 489, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 320, + 489, + 555, + 521 + ], + "type": "text", + "content": "[44] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2023. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 523, + 555, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 523, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 320, + 523, + 555, + 567 + ], + "type": "text", + "content": "[45] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 2, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 568, + 555, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 568, + 555, + 621 + ], + "spans": [ + { + "bbox": [ + 320, + 568, + 555, + 621 + ], + "type": "text", + "content": "[46] Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. Lost in the middle: How language models use long contexts. Transactions of the Association for Computational Linguistics, 12: 157-173, 2024. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 624, + 555, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 624, + 555, + 668 + ], + "spans": [ + { + "bbox": [ + 320, + 624, + 555, + 668 + ], + "type": "text", + "content": "[47] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 670, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 670, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 670, + 555, + 713 + ], + "type": "text", + "content": "[48] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22595" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 81, + 73, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 296, + 95 + ], + "type": "text", + "content": "the IEEE/CVF international conference on computer vision, pages 2630-2640, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 96, + 296, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 96, + 296, + 151 + ], + "spans": [ + { + "bbox": [ + 61, + 96, + 296, + 151 + ], + "type": "text", + "content": "[49] Medhini Narasimhan, Arsha Nagrani, Chen Sun, Michael Rubinstein, Trevor Darrell, Anna Rohrbach, and Cordelia Schmid. Tl; dw? summarizing instructional videos with task relevance and cross-modal saliency. In European Conference on Computer Vision, pages 540-557. Springer, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 152, + 295, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 152, + 295, + 195 + ], + "spans": [ + { + "bbox": [ + 61, + 152, + 295, + 195 + ], + "type": "text", + "content": "[50] Medhini Narasimhan, Licheng Yu, Sean Bell, Ning Zhang, and Trevor Darrell. Learning and verification of task structure in instructional videos. arXiv preprint arXiv:2303.13519, 2023. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 198, + 296, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 198, + 296, + 230 + ], + "spans": [ + { + "bbox": [ + 62, + 198, + 296, + 230 + ], + "type": "text", + "content": "[51] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 232, + 296, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 232, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 62, + 232, + 296, + 285 + ], + "type": "text", + "content": "[52] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 288, + 296, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 288, + 296, + 320 + ], + "spans": [ + { + "bbox": [ + 62, + 288, + 296, + 320 + ], + "type": "text", + "content": "[53] Zhanzhong Pang, Fadime Sener, and Angela Yao. Context-enhanced memory-refined transformer for online action detection. arXiv preprint arXiv:2503.18359, 2025. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 322, + 296, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 322, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 62, + 322, + 296, + 386 + ], + "type": "text", + "content": "[54] Dhruvesh Patel, Hamid Eghbalzadeh, Nitin Kamra, Michael Louis Iuzzolino, Unnat Jain, and Ruta Desai. Pretrained language models as visual planners for human assistance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15302-15314, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 389, + 296, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 389, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 62, + 389, + 296, + 443 + ], + "type": "text", + "content": "[55] Reiner Pope, Sholto Douglas, Aakanksha Chowdhery, Jacob Devlin, James Bradbury, Jonathan Heek, Kefan Xiao, Shivani Agrawal, and Jeff Dean. Efficiently scaling transformer inference. Proceedings of Machine Learning and Systems, 5:606-624, 2023. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 445, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 296, + 510 + ], + "type": "text", + "content": "[56] Shraman Pramanick, Yale Song, Sayan Nag, Kevin Qinghong Lin, Hardik Shah, Mike Zheng Shou, Rama Chellappa, and Pengchuan Zhang. Egovlpv2: Egocentric video-language pre-training with fusion in the backbone. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5285-5297, 2023. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 513, + 296, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 513, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 513, + 296, + 533 + ], + "type": "text", + "content": "[57] Alec Radford. Improving language understanding by generative pre-training. Openai, 2018. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 536, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 296, + 578 + ], + "type": "text", + "content": "[58] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 581, + 296, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 581, + 296, + 646 + ], + "spans": [ + { + "bbox": [ + 62, + 581, + 296, + 646 + ], + "type": "text", + "content": "[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 647, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 296, + 712 + ], + "type": "text", + "content": "[60] Francesco Ragusa, Antonino Furnari, Salvatore Livatino, and Giovanni Maria Farinella. The meccano dataset: Understanding human-object interactions from egocentric videos in an industrial-like domain. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1569-1578, 2021. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 320, + 73, + 555, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 555, + 115 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 555, + 115 + ], + "type": "text", + "content": "[61] Abhinav Rai, Fadime Sener, and Angela Yao. Transformed rois for capturing visual transformations in videos. Computer Vision and Image Understanding, 224:103558, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 118, + 555, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 118, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 320, + 118, + 555, + 173 + ], + "type": "text", + "content": "[62] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 175, + 555, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 175, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 320, + 175, + 555, + 239 + ], + "type": "text", + "content": "[63] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 243, + 555, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 243, + 555, + 306 + ], + "spans": [ + { + "bbox": [ + 320, + 243, + 555, + 306 + ], + "type": "text", + "content": "[64] Luigi Seminara, Giovanni Maria Farinella, and Antonino Furnari. Differentiable task graph learning: Procedural activity representation and online mistake detection from egocentric videos. In Advances in Neural Information Processing Systems, pages 59373-59407. Curran Associates, Inc., 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 309, + 555, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 309, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 320, + 309, + 555, + 352 + ], + "type": "text", + "content": "[65] Fadime Sener and Angela Yao. Zero-shot anticipation for instructional activities. In Proceedings of the IEEE/CVF international conference on computer vision, pages 862-871, 2019. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 355, + 555, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 355, + 555, + 410 + ], + "spans": [ + { + "bbox": [ + 320, + 355, + 555, + 410 + ], + "type": "text", + "content": "[66] Fadime Sener, Dipika Singhania, and Angela Yao. Temporal aggregate representations for long-range video understanding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16, pages 154-171. Springer, 2020. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 411, + 555, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 411, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 320, + 411, + 555, + 477 + ], + "type": "text", + "content": "[67] Fadime Sener, Dibyadip Chatterjee, Daniel Shelepov, Kun He, Dipika Singhania, Robert Wang, and Angela Yao. Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21096-21106, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 479, + 555, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 479, + 555, + 522 + ], + "spans": [ + { + "bbox": [ + 320, + 479, + 555, + 522 + ], + "type": "text", + "content": "[68] Md Salman Shamil, Dibyadip Chatterjee, Fadime Sener, Shugao Ma, and Angela Yao. On the utility of 3d hand poses for action recognition. In European Conference on Computer Vision, pages 436-454. Springer, 2024. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 524, + 555, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 524, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 320, + 524, + 555, + 578 + ], + "type": "text", + "content": "[69] Dandan Shan, Jiaqi Geng, Michelle Shu, and David F Fouhey. Understanding human hands in contact at internet scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9869-9878, 2020. 4, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 580, + 555, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 580, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 320, + 580, + 555, + 645 + ], + "type": "text", + "content": "[70] Zheng Shou, Junting Pan, Jonathan Chan, Kazuyuki Miyazawa, Hassan Mansour, Anthony Vetro, Xavier Giro-i Nieto, and Shih-Fu Chang. Online detection of action start in untrimmed, streaming videos. In Proceedings of the European conference on computer vision (ECCV), pages 534-551, 2018. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 647, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 647, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 320, + 647, + 555, + 712 + ], + "type": "text", + "content": "[71] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22596" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 73, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 297, + 127 + ], + "type": "text", + "content": "[72] Yale Song, Eugene Byrne, Tushar Nagarajan, Huiyu Wang, Miguel Martin, and Lorenzo Torresani. Ego4d goal-step: Toward hierarchical understanding of procedural activities. Advances in Neural Information Processing Systems, 36, 2023. 2, 3, 4, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 128, + 296, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 128, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 61, + 128, + 296, + 193 + ], + "type": "text", + "content": "[73] Reuben Tan, Xineng Sun, Ping Hu, Jui-hsien Wang, Hanieh Deilamsalehy, Bryan A Plummer, Bryan Russell, and Kate Saenko. Koala: Key frame-conditioned long video-llm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13581–13591, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 194, + 296, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 296, + 259 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 296, + 259 + ], + "type": "text", + "content": "[74] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1207-1216, 2019. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 261, + 296, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 296, + 315 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 296, + 315 + ], + "type": "text", + "content": "[75] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 316, + 296, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 316, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 62, + 316, + 296, + 370 + ], + "type": "text", + "content": "[76] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568-9578, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 371, + 296, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 371, + 296, + 425 + ], + "spans": [ + { + "bbox": [ + 62, + 371, + 296, + 425 + ], + "type": "text", + "content": "[77] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 426, + 296, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 426, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 62, + 426, + 296, + 470 + ], + "type": "text", + "content": "[78] Huiyu Wang, Mitesh Kumar Singh, and Lorenzo Torresani. Ego-only: Egocentric action detection without exocentric transferring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5250-5261, 2023. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 471, + 296, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 471, + 296, + 524 + ], + "spans": [ + { + "bbox": [ + 62, + 471, + 296, + 524 + ], + "type": "text", + "content": "[79] Jiahao Wang, Guo Chen, Yifei Huang, Limin Wang, and Tong Lu. Memory-and-anticipation transformer for online action understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13824–13835, 2023. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 525, + 296, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 525, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 62, + 525, + 296, + 568 + ], + "type": "text", + "content": "[80] Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video understanding with large language model as agent. arXiv preprint arXiv:2403.10517, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 570, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 570, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 62, + 570, + 296, + 613 + ], + "type": "text", + "content": "[81] Yuetian Weng, Mingfei Han, Haoyu He, Xiaojun Chang, and Bohan Zhuang. Longvlm: Efficient long video understanding via large language models. arXiv preprint arXiv:2404.03384, 2024. 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 614, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 614, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 614, + 296, + 635 + ], + "type": "text", + "content": "[82] Wikipedia. How to do anything. http://www.wikipedia.com/, 2005.1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 637, + 296, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 637, + 296, + 690 + ], + "spans": [ + { + "bbox": [ + 62, + 637, + 296, + 690 + ], + "type": "text", + "content": "[83] Shiwei Wu, Joya Chen, Kevin Qinghong Lin, Qimeng Wang, Yan Gao, Qianli Xu, Tong Xu, Yao Hu, Enhong Chen, and Mike Zheng Shou. Videollm-mod: Efficient video-language streaming with mixture-of-depths vision computation. arXiv preprint arXiv:2408.16730, 2024. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 692, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 692, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 692, + 296, + 713 + ], + "type": "text", + "content": "[84] Junbin Xiao, Angela Yao, Yicong Li, and Tat-Seng Chua. Can i trust your answer? visually grounded video question" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 338, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 338, + 73, + 555, + 106 + ], + "type": "text", + "content": "answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13204-13214, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 107, + 555, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 107, + 555, + 161 + ], + "spans": [ + { + "bbox": [ + 321, + 107, + 555, + 161 + ], + "type": "text", + "content": "[85] Junbin Xiao, Nanxin Huang, Hangyu Qin, Dongyang Li, Yicong Li, Fengbin Zhu, Zhulin Tao, Jianxing Yu, Liang Lin, Tat-Seng Chua, et al. Videoqa in the era of llms: An empirical study. International Journal of Computer Vision, pages 1-24, 2025. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 162, + 555, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 162, + 555, + 215 + ], + "spans": [ + { + "bbox": [ + 321, + 162, + 555, + 215 + ], + "type": "text", + "content": "[86] Mingze Xu, Yuanjun Xiong, Hao Chen, Xinyu Li, Wei Xia, Zhuowen Tu, and Stefano Soatto. Long short-term transformer for online action detection. Advances in Neural Information Processing Systems, 34:1086-1099, 2021. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 217, + 555, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 217, + 555, + 259 + ], + "spans": [ + { + "bbox": [ + 321, + 217, + 555, + 259 + ], + "type": "text", + "content": "[87] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 261, + 555, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 261, + 555, + 303 + ], + "spans": [ + { + "bbox": [ + 321, + 261, + 555, + 303 + ], + "type": "text", + "content": "[88] Olga Zatsarynna and Juergen Gall. Action anticipation with goal consistency. In 2023 IEEE International Conference on Image Processing (ICIP), pages 1630-1634. IEEE, 2023. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 304, + 555, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 304, + 555, + 349 + ], + "spans": [ + { + "bbox": [ + 321, + 304, + 555, + 349 + ], + "type": "text", + "content": "[89] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 2, 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 350, + 555, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 350, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 321, + 350, + 555, + 392 + ], + "type": "text", + "content": "[90] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Is an object-centric video representation beneficial for transfer? In Proceedings of the Asian Conference on Computer Vision, pages 1976-1994, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 394, + 555, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 394, + 555, + 446 + ], + "spans": [ + { + "bbox": [ + 321, + 394, + 555, + 446 + ], + "type": "text", + "content": "[91] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Helping hands: An object-aware ego-centric video recognition model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13901-13912, 2023. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 449, + 554, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 449, + 554, + 503 + ], + "spans": [ + { + "bbox": [ + 321, + 449, + 554, + 503 + ], + "type": "text", + "content": "[92] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 543–553, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 504, + 555, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 504, + 555, + 557 + ], + "spans": [ + { + "bbox": [ + 321, + 504, + 555, + 557 + ], + "type": "text", + "content": "[93] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 559, + 555, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 559, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 321, + 559, + 555, + 592 + ], + "type": "text", + "content": "[94] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 593, + 555, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 593, + 555, + 657 + ], + "spans": [ + { + "bbox": [ + 321, + 593, + 555, + 657 + ], + "type": "text", + "content": "[95] He Zhao, Isma Hadji, Nikita Dvornik, Konstantinos G Derpanis, Richard P Wildes, and Allan D Jepson. P3iv: Probabilistic procedure planning from instructional videos with weak supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2938-2948, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 658, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 658, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 321, + 658, + 555, + 713 + ], + "type": "text", + "content": "[96] Qi Zhao, Shijie Wang, Ce Zhang, Changcheng Fu, Minh Quan Do, Nakul Agarwal, Kwonjoon Lee, and Chen Sun. AntGPT: Can large language models help long-term action anticipation from videos? In The Twelfth International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22597" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 73, + 297, + 462 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 61, + 73, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 297, + 116 + ], + "type": "text", + "content": "[97] Yue Zhao and Philipp Krahenbuhl. Real-time online video detection with temporal smoothing transformers. In European Conference on Computer Vision, pages 485-502. Springer, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 118, + 295, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 295, + 182 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 295, + 182 + ], + "type": "text", + "content": "[98] Yue Zhao, Long Zhao, Xingyi Zhou, Jialin Wu, Chun-Te Chu, Hui Miao, Florian Schroff, Hartwig Adam, Ting Liu, Boqing Gong, et al. Distilling vision-language models on millions of videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13106-13116, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 185, + 296, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 185, + 296, + 248 + ], + "spans": [ + { + "bbox": [ + 62, + 185, + 296, + 248 + ], + "type": "text", + "content": "[99] Yiwu Zhong, Licheng Yu, Yang Bai, Shangwen Li, Xueting Yan, and Yin Li. Learning procedure-aware video representation from instructional videos and their narrations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14825-14835, 2023. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 251, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 251, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 57, + 251, + 296, + 316 + ], + "type": "text", + "content": "[100] Honglu Zhou, Roberto Martin-Martín, Mubbasir Kapadia, Silvio Savarese, and Juan Carlos Niebles. Procedure-aware pretraining for instructional video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10727-10738, 2023. 1, 2, 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 318, + 294, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 318, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 57, + 318, + 294, + 361 + ], + "type": "text", + "content": "[101] Luowei Zhou, Chenliang Xu, and Jason Corso. Towards automatic learning of procedures from web instructional videos. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 363, + 295, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 363, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 57, + 363, + 295, + 407 + ], + "type": "text", + "content": "[102] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 408, + 295, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 408, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 57, + 408, + 295, + 462 + ], + "type": "text", + "content": "[103] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramazan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Crosstask weakly supervised learning from instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3537-3545, 2019. 2" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22598" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Streamlining Image Editing with Layered Diffusion Brushes/6116442a-236c-4dff-8b3c-c87a9184cffb_content_list.json b/2025/Streamlining Image Editing with Layered Diffusion Brushes/6116442a-236c-4dff-8b3c-c87a9184cffb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..74f75b4d6da3e913d6eb675dd96d4925bf340d2b --- /dev/null +++ b/2025/Streamlining Image Editing with Layered Diffusion Brushes/6116442a-236c-4dff-8b3c-c87a9184cffb_content_list.json @@ -0,0 +1,1966 @@ +[ + { + "type": "text", + "text": "Streamlining Image Editing with Layered Diffusion Brushes", + "text_level": 1, + "bbox": [ + 194, + 122, + 803, + 143 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Peyman Gholami Robert Xiao University of British Columbia", + "bbox": [ + 357, + 169, + 640, + 204 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{peymang, brx}@cs.ubc.ca", + "bbox": [ + 393, + 207, + 604, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://layered-diffusion-brushes.github.io/", + "bbox": [ + 277, + 232, + 712, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Edit prompt (type)", + "bbox": [ + 189, + 267, + 243, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sunglasses (add objects)", + "bbox": [ + 282, + 270, + 351, + 294 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hat, Takashi Murakami style (change style)", + "bbox": [ + 382, + 268, + 467, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "starry night, van gogh style (fix problems, maintaining style", + "bbox": [ + 486, + 268, + 576, + 318 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sky, Leonid Afremov style (mixing styles)", + "bbox": [ + 601, + 268, + 681, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ornate frame (change attribute / texture)", + "bbox": [ + 699, + 268, + 799, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Layer Mask", + "bbox": [ + 189, + 327, + 233, + 353 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/acaab80982df157db185097f8b40b63ad9574f638f96c248dbbfefb5407b33bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 259, + 319, + 367, + 400 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/77f70798e125df1a37ec37aa708ce20fd04b231e2691b8fade439235a5556149.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 319, + 477, + 401 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f60328ab3bcc6d1830ed940bce16320b379ad75844848dc0d8e8ead7821fffe9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 319, + 586, + 401 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/94fd2bed5093f2db8214436672bb3cd81aa40a48d03d1109eae55ba28246a858.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 319, + 694, + 401 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7ba782e1c1776a26f3e515ef1ba6e68a809a2ff81daaebe84a6c6f132290b958.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 319, + 805, + 401 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Edited image", + "bbox": [ + 189, + 425, + 236, + 453 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e89c61fa9b418e83a4dea980178f168cd36a02b6096d7c6d759e5709e5945996.jpg", + "image_caption": [ + "Layer 1" + ], + "image_footnote": [], + "bbox": [ + 259, + 402, + 367, + 484 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f2804203a71914757c150a574dced80015846053d91cc90137f4e70db1547e19.jpg", + "image_caption": [ + "Layer 2", + "Figure 1. Hierarchical image editing with Layered Diffusion Brushes: LDB is capable of creating and stacking a wide range of independent edits, including object addition, removal, or replacement, colour and style changes/combining, and object attribute modification. Each edit is performed independently, and users are able to switch between the edits seamlessly." + ], + "image_footnote": [], + "bbox": [ + 370, + 402, + 477, + 484 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/82ebbcc202c3e13a464807958caa0d049cb69d123e40903a3ab8fea09ab174b4.jpg", + "image_caption": [ + "Layer 3" + ], + "image_footnote": [], + "bbox": [ + 480, + 402, + 586, + 483 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a9d5c136e1562341f197982dc749cde3300a9cfe67094fc2efb66be60ee26e57.jpg", + "image_caption": [ + "Layer 4" + ], + "image_footnote": [], + "bbox": [ + 589, + 402, + 694, + 483 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/860e7207bd9ab2e80530566d5be95c5c12f3721773ecee93a98f3ae369729cf9.jpg", + "image_caption": [ + "Layer 5" + ], + "image_footnote": [], + "bbox": [ + 697, + 402, + 805, + 483 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 585, + 326, + 602 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Denoising diffusion models have emerged as powerful tools for image manipulation, yet interactive, localized editing workflows remain underdeveloped. We introduce Layered Diffusion Brushes (LDB), a novel training-free framework that enables interactive, layer-based editing using standard diffusion models. LDB defines each \"layer\" as a self-contained set of parameters guiding the generative process, enabling independent, non-destructive, and fine-grained prompt-guided edits, even in overlapping regions. LDB leverages a unique intermediate latent caching approach to reduce each edit to only a few denoising steps, achieving 140 ms per edit on consumer GPUs. An editor implementing LDB, incorporating familiar layer concepts, was evaluated via user study and quantitative metrics. Results demonstrate LDB's superior speed alongside comparable or improved image quality, background preservation, and edit fidelity relative to state-of-the-art methods across various sequential image manipulation tasks. The findings highlight LDB's ability to significantly enhance creative", + "bbox": [ + 88, + 613, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "workflows by providing an intuitive and efficient approach to diffusion-based image editing and its potential for expansion into related subdomains, such as video editing.", + "bbox": [ + 511, + 588, + 906, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 662, + 645, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image editing has undergone transformative advancements with the rise of text-to-image (T2I) generative models, enabling unprecedented creative expression through textual guidance. These models, including Generative Adversarial Networks (GANs) [20], Variational Autoencoders (VAEs), and Denoising Diffusion Models (DMs) [23], have redefined image synthesis and manipulation. Among these, DMs [56] have emerged as the state of the art due to their training stability, high-fidelity outputs, and versatility across tasks like inpainting [37], super-resolution [53], and style transfer [22]. However, despite their capabilities, a critical gap remains: enabling real-time, localized, and iterative edits that align with professional workflows, where artists demand precise control over specific regions without", + "bbox": [ + 509, + 688, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17368", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "disrupting the global composition.", + "bbox": [ + 89, + 90, + 318, + 107 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing DM-based editing methods face several core challenges. First, their stochastic nature often necessitates numerous generations to achieve desired results [5]. Second, they lack intuitive mechanisms for layered, nondestructive editing—a cornerstone of tools like Adobe Photoshop [28]—where edits can be independently adjusted, stacked, or removed. Third, while mask-guided approaches enable regional control, they struggle with seamless blending, artifact-free transitions, and real-time feedback. These limitations restrict their adoption in creative pipelines, where rapid iteration and granular control are critical.", + "bbox": [ + 89, + 113, + 483, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose Layered Diffusion Brushes (LDB), a novel framework based on Latent Diffusion Models (LDM) [50] that integrates mask-guided diffusion with a non-destructive layered editing paradigm.", + "bbox": [ + 88, + 287, + 483, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "At its core, LDB introduces new noise patterns into the image latents during diffusion process, guided by both the user-specified mask and the edit prompt. This preserves the original context while seamlessly integrating localized edits. We implement an intuitive user interface (UI) with a layering system to support consecutive edits (Fig. 1). Specifically, as key contributions, LDB introduces:", + "bbox": [ + 89, + 356, + 483, + 464 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Latent Caching for Real-Time Edits: By reusing intermediate denoising states from initial generation, edits bypass redundant computations and achieve as low as 140 ms per edit on $512 \\times 512$ images ( $53 \\times$ faster than BrushNet [31] using the same consumer GPUs).", + "- Non-destructive Layered Editing: LDB introduces an order-agnostic layering mechanism by defining the concept of layers for DMs, enabling:", + "- Region-targeted adjustments with background preservation, using mask-prompt pairs,", + "- Stacking, toggling, or deleting layers without cross-interference—even in overlapping regions,", + "- Post-hoc revision of edits while preserving underlying content.", + "- Seed-Driven Exploration: Our UI provides familiar \"brush\" and \"scroll\" gestures to enable instant exploration of variations by modulating noise seeds, bridging stochastic generation with deterministic refinement and instant feedback." + ], + "bbox": [ + 89, + 469, + 482, + 755 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We validate LDB through extensive experiments and a user study with graphic designers. Quantitatively, LDB outperforms state-of-the-art methods in terms of speed and image quality and is comparable in terms of edit fidelity. The user study revealed superior usability and creativity support in iterative design. Additionally, LDB is a plug-and-play, training-free system adaptable to existing models and applications, and we demonstrate this by applying LDB to the task of video editing.", + "bbox": [ + 89, + 763, + 483, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 89, + 653, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. DM-based Image Editing", + "text_level": 1, + "bbox": [ + 513, + 114, + 743, + 132 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image editing is the task of modifying existing images in terms of appearance, structure, or composition, ranging from subtle adjustments to major transformations. Unlike GAN-based approaches [1, 35, 44], which are prone to limitations in inversion stability [49] and localized control [6], diffusion-based methods harness the power of controllable, high-quality DMs in various image-editing tasks, including text and image-driven image manipulation studies [14, 26, 34, 37].", + "bbox": [ + 511, + 138, + 906, + 275 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Instruction-based text editing methods [9, 18, 19, 21, 64] typically train DMs on instruction-image pairs. For example, InstructPix2Pix [9] is trained using synthetic pairs from Stable Diffusion [50] and Prompt-to-Prompt [22]. However, expressing nuanced edits solely through text instructions remains challenging, particularly for object-specific style or color changes.", + "bbox": [ + 511, + 276, + 905, + 382 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Mask-based methods [4, 5, 14, 63] sample within specified regions. While effective for localized edits, they can introduce unintended global changes, especially problematic in sequential editing, and may struggle with complex edits. For instance, Blended Latent Diffusion's lossy VAE latent space hinders perfect reconstruction even before noise addition [5]. Though a background reconstruction strategy is included, it increases computation and may still yield incoherent results for complex edits. Conversely, our method directly modifies the original latent space, enhancing context preservation and natural blending.", + "bbox": [ + 511, + 383, + 905, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Attention-based editing manipulates cross-attention maps to guide the image generation process toward the desired modifications [22, 45]. These methods generally face challenges in achieving fine-grained edits without unwanted global modifications. Yang et al. [61] attribute unintended changes to inaccurate attention maps and propose attention focusing. Inversion-based methods like ILVR [12], Textual Inversion [16], and DreamBooth [51] focus on context modification while preserving subjects. DDIM inversion converts images to noisy latents, and sampling generates edited results based on prompts. We employ Direct Inversion [30] for efficient real image latent inversion.", + "bbox": [ + 511, + 551, + 905, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image inpainting involves replacing or restoring the missing regions while maintaining global coherency [60]. Many inpainting works [39, 50, 62, 69] require using a fine-tuned DM specifically designed for inpainting tasks, limiting their applicability. Some, including SmartBrush, which uses object-mask prediction guidance [59], offer more flexibility. PowerPaint [69] introduces learnable task embeddings for improved control. While these models effectively generate new content, they are generally unsuitable for making small, targeted adjustments [4, 37, 52]. Inspired by ControlNet [67], BrushNet [31] builds a decomposed", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17369", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "plug-and-play dual-branch DM, but struggles with real-time interaction due to its computational overhead. In Sec. 4.1 we compare LDB with several inpainting techniques.", + "bbox": [ + 89, + 90, + 480, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Layered and Sequential Image editing", + "text_level": 1, + "bbox": [ + 89, + 148, + 419, + 165 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Layer-based image editing is fundamental in computer graphics [46], and recent works integrate this concept into AI methodologies [6, 54]. Layered representations enable dynamic manipulation of image components, transforming single images into multi-layered structures.", + "bbox": [ + 89, + 171, + 482, + 247 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LayeringDiff [33] decomposes images into foreground and background. ParallelEdits [27] uses attention for efficient multi-aspect text edits. MAG-Edit [40] employs a two-layer process with attention injection to a single edit from background. Joseph et al. [29] highlight error accumulation in sequential editing, where artifacts compound across edits. Collage Diffusion [54], built on modified Blended Latent Diffusion [5], employs alpha masks to guide cross-attention and generate harmonized images while respecting scene composition. However, it assumes prelayered inputs and synthesizes scenes from scratch. In contrast, LDB is training-free, operates directly on existing images, and supports fully independent layers—unlike methods such as [6] that require per-image training.", + "bbox": [ + 89, + 248, + 483, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Accelerated Generation using Caching", + "text_level": 1, + "bbox": [ + 89, + 470, + 423, + 488 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Caching and reusing intermediate features has proven effective for accelerating DM inference through reducing redundant computations. Several works have utilized caching in diffusion transformers (DiTs) for video generation. DeepCache [38] reuses high-level U-Net features in video generation, while AdaCache [32] dynamically adjusts cached residuals based on temporal content. Cache Me If You Can [57] employs block caching by reusing outputs from layer blocks of previous steps during inference. For image generation, Approximate Caching [2] reuses intermediate latents created during prior image generation processes for similar prompts. We employ a similar strategy through caching key latent representations and adapt it specifically for interactive image editing, enabling the real-time feedback that is crucial for creative workflows.", + "bbox": [ + 89, + 494, + 483, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 737, + 181, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use an LDM-based variant of image generative models and make intermediate adjustments to the latent space, similar to [5, 37]. Therefore, LDB requires no additional training or fine-tuning of the underlying LDM; all modifications are applied during the reverse diffusion process.", + "bbox": [ + 89, + 763, + 482, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We adopt the standard LDM formulation, where image generation begins with a sample from a Gaussian distribution, $Z_0 \\sim \\mathcal{N}(0, \\sigma_{max}^2 I)$ and is iteratively denoised through a sequence of steps $N$ , resulting in a series of la", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tents $Z_{i}$ corresponding to decreasing noise levels $\\sigma_{i}$ , where $\\sigma_0 = \\sigma_{max} > \\sigma_1 > \\dots >\\sigma_N\\approx 0$", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As demonstrated in Fig. 2, the overall LDB pipeline comprises three key stages: initial image generation (or inversion), latent caching, and iterative layered editing.", + "bbox": [ + 511, + 121, + 903, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For DM-generated images, we first initialize the sample $Z_{0} = \\epsilon_{0}$ and noise level $\\sigma_0$ ( $i = 0$ ). For real images, the initial noise latent is obtained using inversion. We use Direct Inversion [30] for its high speed and comparable performance to other inversion methods, including Null-Text Inversion [43] and Negative-Prompt Inversion [42]. The noisy sample then undergoes the diffusion process, caching certain intermediate latents to facilitate editing.", + "bbox": [ + 511, + 167, + 906, + 289 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Latent Caching", + "text_level": 1, + "bbox": [ + 513, + 300, + 669, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enable rapid, interactive editing with instant exploration and feedback, we employ latent caching to reuse intermediate representations in subsequent steps, minimizing redundant computations. We store two key intermediate latents:", + "bbox": [ + 511, + 321, + 903, + 382 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Regeneration Latent $\\mathbf{Z}_{\\mathbf{r}}$ : At diffusion step $r = N - n$ , where $N$ is the total number of diffusion steps for initial image generation and $n$ is the number of editing steps, we cache the latent $Z_{r}$ , which serves as the starting point for all subsequent edits. By reusing $Z_{r}$ , we avoid recomputing the initial denoising steps for each new edit, significantly speeding up the editing process (from $N$ denoising steps to $n$ ). Effectively, $Z_{r}$ represents a partially denoised latent state that retains the global image structure but is still malleable enough to accommodate localized edits.", + "- Blending Latent $Z_{\\mathrm{b}}$ : We cache the latent at diffusion step $b$ which is specifically used for the layer merging process (Algorithm 1, line 7). We set $b = N - 2$ for maximum background preservation (as discussed in Sec. 4.3). $Z_{b}$ represents a more denoised latent compared to $Z_{r}$ , capturing more refined image details while still allowing for seamless blending of new edits into the existing image context. Utilizing this cached blending latent ensures smoother integration of edits and reduces visual artifacts at layer boundaries during the merging process." + ], + "bbox": [ + 513, + 383, + 906, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Layered Diffusion Brushes Editing", + "text_level": 1, + "bbox": [ + 513, + 696, + 816, + 713 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To initiate an edit, the algorithm begins by generating a new noise pattern $Z_0' = \\epsilon_k'$ , sampled from $\\mathcal{N}(0, \\sigma^2 I)$ using a different seed $S'$ , and scaling it to match the variance of the cached latent $Z_r$ . This ensures that the additive noise stays in a reasonable range from the latent for editing, preventing visual artifacts. $Z_0'$ is then added to the regeneration latent $Z_r$ , controlled by the mask $m$ and strength $\\alpha$ .", + "bbox": [ + 511, + 719, + 905, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the editing stage, at step $b$ , a new noisy sample is merged with the cached blending latent using the strength control and the mask, resulting in $Z_b'$ . Subsequently, the new latent is progressively denoised from steps $b$ through $N$ and processed through the VAE to output edited image $I'$ .", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17370", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/489a9b6fe8d363f442658601222ddb107028169dae9d03f2560069fd87c9171c.jpg", + "image_caption": [ + "Figure 2. Overview of the Proposed Method: The top box shows standard DM-based image generation from noisy latent $Z_{0}$ and prompt $\\mathcal{P}$ . The middle section depicts the latent caching module, storing and retrieving intermediate latents for different layers. The bottom box illustrates the editing process: a new noise sample $S'$ merges with the original latent at step $r$ using mask $m$ and strength control $\\alpha$ . Diffusion continues until step $b$ , where modified and cached latents blend to generate the final edited image." + ], + "image_footnote": [], + "bbox": [ + 178, + 88, + 823, + 373 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 presents the pseudocode for the editing process for a single layer (for simplicity):", + "bbox": [ + 89, + 465, + 482, + 498 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: LDB editing process (single layer)" + ], + "code_body": "Input: Edit prompt $\\mathcal{P}'$ , Mask $m \\in [0,1]^{H \\times W}$ , Random seed $S'$ , Strength $\\alpha$ , Number of edit steps $n$ , Regeneration latent $Z_r$ , Blending latent $Z_b$ \nOutput: Edited latent $Z_N'$ \n1 $Z_0' \\gets \\epsilon_{nk}' \\sim \\mathcal{N}(0, \\sigma^2 I) //$ sampled using seed $S'$ \n2 $Z_0' \\gets \\sqrt{\\text{Var}(Z_r)} \\cdot Z_0' //$ scale new sample \n3 $Z_0' \\gets Z_r + \\alpha \\cdot (Z_0' \\odot m) //$ noise injection \n4 for $i = 0,1,\\ldots,n$ do \n5 $Z_{i+1}' \\gets DM(Z_i', \\mathcal{P}', i, S')$ \n6 if $i == b$ then \n7 $Z_b' \\gets Z_b' \\odot m + Z_b \\odot (1 - m) //$ blending \n8 end \n9 Return $Z_N'$", + "bbox": [ + 94, + 529, + 483, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Layer Formulation", + "text_level": 1, + "bbox": [ + 89, + 776, + 274, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike prior works that rely on transparent decomposable layers [66] or explicit object segmentation [54], we redefine a layer as a self-contained set of reproducible parameters that govern localized edits. For layer $\\mathcal{L}_k$ , we formalize this as a generalized version of parameters in Algorithm 1:", + "bbox": [ + 89, + 799, + 482, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {(k)} = \\left(\\mathbf {S} ^ {\\prime (k)}, \\mathbf {m} ^ {(k)}, \\mathbf {v} ^ {(k)}, \\mathbf {Z} _ {r} ^ {(k)}, \\mathbf {Z} _ {b} ^ {(k)}, \\alpha^ {(k)}, n ^ {(k)}, \\mathcal {P} ^ {\\prime (k)}, j\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 880, + 482, + 905 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- $\\mathbf{S}'^{(k)} \\in \\mathbb{Z}^+$ : Seed space for stochastic variations", + "- $\\mathbf{m}^{(k)}\\in [0,1]^{H\\times W}$ : Edit mask", + "- $\\mathbf{v}^{(k)}\\in \\{0,1\\}$ : Visibility state", + "- $\\mathbf{Z}_r^{(k)},\\mathbf{Z}_b^{(k)}\\in \\mathbb{R}^{C\\times H\\times W}$ : Regeneration/blending latents", + "- $\\alpha^{(k)}\\in [0,1]$ : Layer strength value", + "- $\\mathbf{n}^{(k)}\\in [0,N]$ Number of denoising steps", + "$\\mathcal{P}^{\\prime (k)}$ : Edit prompt", + "- $j \\in {\\mathbb{Z}}^{ + }$ : Index of last layer index." + ], + "bbox": [ + 513, + 463, + 890, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Notably, within a given layer $\\mathcal{L}_k$ with previous layer $\\mathcal{L}_j$ , the cached latents $\\mathbf{Z}_r^{(j)}$ and $\\mathbf{Z}_b^{(j)}$ inherently incorporate the cumulative edits from all preceding layers. This is because edits to layer $\\mathcal{L}_k$ , are applied to the already edited output of layer $\\mathcal{L}_j$ which serves as the input to the diffusion process and the algorithm always keeps the last layer updated. Therefore, any modification in a previous layer automatically propagates through the subsequent layers. By defining $\\Phi$ as a single-layer latent generation and caching step as:", + "bbox": [ + 511, + 595, + 906, + 734 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left(Z _ {r} ^ {(k)}, Z _ {b} ^ {(k)}\\right) = \\Phi \\left(\\mathcal {L} ^ {(k)}, \\mathcal {L} ^ {(j)}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 746, + 906, + 767 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in essence, if a given layer $\\mathcal{L}^{(i)}$ (where $i < k$ ) is removed or its visibility $\\mathbf{v}^{(i)}$ is toggled, the operator $\\Phi$ will be recursively invoked to recreate all latents for layers from $\\mathcal{L}^{(i)}$ to $\\mathcal{L}^{(k)}$ . This recomputation, accelerated by latent caching, is automatically triggered and typically completes within milliseconds to a few seconds, depending on the number of layers. This design allows edits to remain independent yet seamlessly integrated into the final composition.", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17371", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1. Overlapping Regions", + "text_level": 1, + "bbox": [ + 89, + 90, + 284, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A key advantage of layered editing in LDB is the ability to create overlapping edits, where one layer can partially or fully modify areas affected by earlier layers. This requires careful handling of each layer's regeneration latent, $Z_{r}$ , to ensure that changes in visibility or content from higher layers are accurately reflected in subsequent layers, even in overlapping regions.", + "bbox": [ + 89, + 109, + 483, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By default, all layers use the initial image's latent $(Z_{r})$ as their regeneration latent. However, this approach fails to account for overlapping edits from preceding layers. To address this, when processing a layer $k$ , we compute its regeneration latent by inverting the output image of the previous layer $(I^{\\prime (k)})$ as shown using the feedback arrow on Fig. 2. This inversion yields $Z_0^{(k)}$ , which is then sent through the generation stage in LDB. Both $Z_{r}^{(k)}$ and $Z_{b}^{(k)}$ are cached for efficient processing (as shown in Fig. 3).", + "bbox": [ + 89, + 217, + 483, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This mechanism enables precise control and seamless integration of edits across overlapping regions. Changes to any layer propagate correctly without introducing artifacts, offering flexibility and fine-grained control.", + "bbox": [ + 89, + 359, + 482, + 420 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. User-Interface and Interaction Design", + "text_level": 1, + "bbox": [ + 89, + 431, + 418, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To develop a practical tool for artists and designers, we designed an custom UI that balances control and simplicity. The UI allows users to generate, upload, and edit images, manage layers, and adjust parameters seamlessly. Two interaction modes streamline edits (Fig. 4):", + "bbox": [ + 89, + 454, + 483, + 530 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Box Mode: Users can click or drag on the image to move a resizeable square mask around. This option enables a quick and interactive exploration of how various parts of the image will change in response to a given set of editing settings (prompt and strength), simply by moving the cursor.", + "bbox": [ + 89, + 531, + 483, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Custom Mask Mode: Users can draw free-form masks over the desired around and navigate between new generation samples by scrolling the mouse up or down while hovering over the image, allowing them to rapidly explore variations on their edit.", + "bbox": [ + 89, + 607, + 482, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose a workflow where users first position edits spatially using Box Mode, then refine mask geometry and appearance details via Custom Mask Mode.", + "bbox": [ + 89, + 684, + 482, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Layering capabilities include stacking, visibility toggling, and deletion. Each layer is independently modifiable. Detailed information on the UI design user interactions and a demo video can be found in supplementary material.", + "bbox": [ + 89, + 729, + 482, + 790 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 806, + 220, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. User Study", + "text_level": 1, + "bbox": [ + 89, + 832, + 212, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conducted a user study in order to evaluate the effectiveness of LDB for providing targeted image fine-tuning, using two other well-known existing image editing tools,", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "InstructPix2Pix (IP2P) [9] and Stable Diffusion Inpainting (SDI) [50] as baselines for comparison.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We recruited a cohort of seven expert participants with extensive experience in using image editing software. As part of our selection criteria, we ensured that all had at least a basic level of familiarity with AI image generation techniques [3, 41] and were regular users of editing software, such as Adobe Photoshop [28] for creating visual art.", + "bbox": [ + 511, + 122, + 903, + 213 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.1. Study Procedure and Task Description", + "text_level": 1, + "bbox": [ + 511, + 224, + 828, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Each user engaged in two types of tasks: free-form tasks where users generated an image for editing using a fixed prompt and seed (type 1), and pre-determined tasks where the user worked with existing real images from the MagicBrush dataset [65] (type 2).", + "bbox": [ + 511, + 244, + 903, + 319 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the type 1 tasks, we selected specific types of edits that showcase various functionalities and capabilities of the system, including:", + "bbox": [ + 511, + 321, + 903, + 366 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Stack layers and create sequential edits (draw with LDB)", + "2. Modify attributes and features of objects", + "3. Correct image imperfections and errors", + "4. Enhance discernibility of similar objects", + "5. Target specific regions for style transfer, refine aesthetics Type 2 tasks were more structured, with the mask, edit prompt, and input images provided by the dataset. The dataset provides manually annotated masks and instructions for each edit. We selected a subset of 35 input images, each containing up to three layers of edits. Users refined masks/parameters if necessary and completed editing tasks." + ], + "bbox": [ + 511, + 367, + 903, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 5, second row, shows example edits generated by the participants. Additional examples are provided in the supplementary material. As shown, LDB produces targeted edits that integrate seamlessly with the images.", + "bbox": [ + 511, + 536, + 903, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.2. Evaluation Survey Results", + "text_level": 1, + "bbox": [ + 511, + 607, + 743, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The participants completed a three stage evaluation survey following the image editing tasks. The first part included a System Usability Scale (SUS) form to rate the usability, ease of use, design, and performance of each method. SUS is a standard usability evaluation survey widely used in user-experience literature [8]. Overall, participants indicated that they are more likely to use LDB compared to IP2P and SDI, and that they find it the easiest tool to use. LDB obtained a SUS score of $80.35\\%$ , while IP2P and SDI achieved a SUS of $38.21\\%$ and $37.5\\%$ respectively.", + "bbox": [ + 511, + 627, + 905, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The SUS survey was followed by a Creativity Support Index [11] survey to evaluate the system's degree of creative work support. Participants expressed positivity towards LDB, indicating that it enhanced their enjoyment, exploration, expressiveness, and immersion, while also deeming the results worth their effort. Lastly, the survey was followed by a semi-structured interview where participants appreciated the intuitiveness, ease of use and versatility of", + "bbox": [ + 511, + 780, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17372", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/060d835aaaea3bc13692762088787ea8c8c2c26fbd33e4e373a7edc2376e50bd.jpg", + "image_caption": [ + "Figure 3. Overlapping edit regions in LDB: overlapping edits enable complex, interacting modifications. For example, one layer can adjust color while another changes shape, with the final result combining both." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 598, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/73023f7987fc7bf9424f1c29630395e60cfea9a1ee9cc89fe094d4aee4cb2c41.jpg", + "image_caption": [ + "Figure 4. Box and Custom Mask Options: In box mode, users click the target region's center to generate edits within the specified area and can drag the box to explore variations instantly. In custom mask mode, users draw a mask over the desired region and adjust the seed using the mouse wheel or scrolling gestures to generate new variations." + ], + "image_footnote": [], + "bbox": [ + 656, + 90, + 905, + 205 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dcfae83d646a828f0af4cf222fd1a45bb71b1c7192f708ddd5eaee96764e3307.jpg", + "image_caption": [ + "Figure 5. Qualitative editing results on PIE-Bench (top) and MagicBrush (bottom) benchmarks using different methods. Edit prompts are presented on top of each row. More examples available in supplementary material." + ], + "image_footnote": [], + "bbox": [ + 112, + 357, + 888, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LDB. Further details about the study, interview, results, and discussion can be found in supplemental material.", + "bbox": [ + 89, + 599, + 482, + 630 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative Analysis", + "text_level": 1, + "bbox": [ + 89, + 637, + 294, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To quantitatively evaluate the performance of LDB, we employed a comprehensive suite of metrics, aligning with established practices in image editing evaluation.", + "bbox": [ + 89, + 659, + 482, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, for text-image alignment, we used CLIP Score (CS) [47] for global alignment, CS-L for masked-region alignment, and CS-D [17] for consistency between image and caption changes in CLIP space.", + "bbox": [ + 89, + 704, + 482, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We adopted Learned Perceptual Image Patch Similarity (LPIPS) [68] and Peak Signal-to-Noise Ratio (PSNR) [24] for evaluating content preservation and pixel-level fidelity in unmasked regions. Furthermore, to gauge overall image quality and aesthetic appeal, we incorporated Aesthetic Score (AS) [55], Image Reward (IR), and Human Preference Score V2 (HPS) [58], the latter two reflecting human-aligned preferences.", + "bbox": [ + 89, + 763, + 482, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compared LDB against a diverse set of state-of", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the-art editing and inpainting methods, including InstructPix2Pix (IP2P) [9], Stable Diffusion Inpainting (SDI) [50], HD-Painter (HDP) [39], BrushNet (BN) [31], and Blended Latent Diffusion (BLD) [5], on two benchmarks: MagicBrush [65] and PIE-Bench [30]. For MagicBrush, we also report results on the provided ground truth (GT) images.", + "bbox": [ + 511, + 599, + 906, + 689 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative results are summarized in Tab. 1. All methods were evaluated using their default editing settings, except for LDB, IP2P, and SDI on the MagicBrush benchmark, where we used user-edited images from our user study for consecutive edits. Inference times denote average per-edit durations, measured on a single NVIDIA RTX 4090 GPU with $N = 25$ diffusion steps for baseline methods and $n = 8$ for LDB.", + "bbox": [ + 511, + 690, + 908, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 818, + 666, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We perform three ablation studies for two main components of the LDB caching mechanism, i.e. the caching timesteps for the regeneration latent $(r)$ , and the blending latent $(b)$ . We also ablate and discuss the effect of strength control $\\alpha$", + "bbox": [ + 511, + 840, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17373", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/caffedebdb7b4568dec8eb535457d3cceeab7dc4171d1c1f89408faf1fba04c9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BenchmarkMethodImage QualityMasked Region PreservationText AlignmentTime (s)
IR ×10↑HPS ×102↑AS↑PSNR↑LPIPS ×102↓CS↑CS-L↑CS-D ×102↑(per edit) ↓
MagicBrushIP2P-62.8321.165.297.2815.0729.3922.016.641.72
SDI-39.2120.885.4812.208.7030.0822.154.111.84
HDP-20.6923.275.4412.056.1331.0122.069.8912.85
BN-0.0422.575.7311.558.7531.1622.1712.927.49
BLD-24.1022.805.4812.646.9430.6421.9910.051.41
GT-1.9322.625.3617.642.3030.7522.149.78NA
Ours7.7422.655.7412.857.0531.0422.079.540.26
PIE-BenchIP2P-40.7323.125.76172.1815.4730.0022.7914.271.83
SDI43.4625.776.00181.583.8931.2422.7114.833.36
HDP39.0225.926.01178.844.6231.0822.7316.2013.44
BN72.7726.666.17177.078.6731.5022.8016.887.51
BLD50.6826.366.11180.854.1931.3522.7817.221.47
Ours86.0226.606.51184.571.9131.6622.7616.740.25
", + "bbox": [ + 135, + 88, + 859, + 327 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative results on MagicBrush and PIE-Bench. Metrics are grouped into Image Quality, Masked Region Preservation, and Text Alignment. $\\uparrow$ indicates higher is better; $\\downarrow$ indicates lower is better. The best and second-best scores are highlighted.", + "bbox": [ + 89, + 329, + 906, + 359 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and its relationship with $n$ in supplementary material.", + "bbox": [ + 89, + 385, + 444, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.1. Ablation on Regeneration Latent Step", + "text_level": 1, + "bbox": [ + 89, + 411, + 401, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The timestep $r$ for caching the regeneration latent is critical, as it dictates the extent of possible modifications during the regeneration process. We performed an ablation study by varying $r$ while holding the total diffusion steps $N$ constant. This variation in $r$ implicitly changes the number of regeneration steps $(n)$ and necessitates adjustments to the strength parameter accordingly. Qualitatively, as shown in Fig. 6, excessively small $r$ values lead to incoherent edits and noticeable artifacts due to insufficient blending with the original image. Conversely, large $r$ values limit the model's ability to modify the masked region, resulting in minimal changes and preserving the original content.", + "bbox": [ + 88, + 430, + 482, + 611 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitatively, we observe that smaller $r$ steps (e.g. $r = 2$ ) yield higher LPIPS (0.04) and low PSNR (27.03), indicating poor image quality and fidelity. Edit fidelity scores such as CS-L also confirm that larger $r$ steps result in lower scores (22.98), suggesting ineffective edits within the masked region. The HPS index demonstrates a higher score for mid-range steps (0.33, $r = 12$ ) compared to both ends of the spectrum (0.29, $r = 23$ ), highlighting a performance sweet spot for intermediate $r$ values. Detailed metric graphs are available in the supplemental material.", + "bbox": [ + 88, + 613, + 482, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.2. Ablation on Blending Latent Step", + "text_level": 1, + "bbox": [ + 89, + 773, + 370, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The blending latent step, controlled by the parameter $b$ , determines when the cached regeneration latent is blended back into the diffusion process and is crucial for seamless integration of the edited region with the original image and preserving background. We conduct an ablation study by varying $b$ while keeping $r$ and $N$ fixed. Fig. 7 qualitatively demonstrates the effect of different $b$ values.", + "bbox": [ + 88, + 794, + 482, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "When $b$ is small, the blending process starts prematurely, causing the edit to bleed into the background and distorting the original image context. Conversely, larger $b$ values, representing late blending, effectively preserve the background integrity while still allowing for meaningful edits within the masked region.", + "bbox": [ + 511, + 385, + 906, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitatively, smaller $b$ values ( $b = n$ ) lead to higher LPIPS (0.17) and lower PSNR (11.61), indicating worse background preservation. Edit fidelity scores (CS-L) within the masked remained stable across the spectrum while CS-D improves at larger $b$ (0.32 at $b = N - 1$ ), reflecting better edit alignment. These findings indicate that later blending is preferable, leading us to select $b = N - 2$ in the LDB algorithm to prioritize background preservation while maintaining effective localized editing. Further details and metric plots are available in the supplementary material.", + "bbox": [ + 511, + 477, + 908, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 511, + 645, + 627, + 660 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our experiments demonstrate that LDB establishes new benchmarks for speed and workflow adaptability in diffusion-based image editing. Key findings include:", + "bbox": [ + 511, + 671, + 905, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Enhanced Control via Layering: LDB's layered design enables creating non-destructive refinements as well as iterative complex compositions. Participants highlighted how this mirrors professional editing tools like Photoshop [28].", + "bbox": [ + 511, + 718, + 903, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Speed and Efficiency: LDB achieves remarkable speed, $53 \\times$ faster than BrushNet (evaluated on the same hardware), crucial for interactive editing. We observe that reducing diffusion steps to as few as $n = 4$ maintains reasonable quality (HPS: 0.34, CS-D: 0.35), yielding a latency of 140ms per edit. User studies confirm instant feedback as a key advantage, enabling rapid iteration (tens of variations per minute vs. 1-2 for baselines). This speed results from", + "bbox": [ + 511, + 779, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17374", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/789119496e1d05996b56b8c934887f6c7de7171ece7879b66ebc2ab86caa27ad.jpg", + "image_caption": [ + "Figure 6. Ablation study on regeneration latent step $r$ (increasing left to right). Small $r$ results in strong prompt adherence (\"cat\") but introduces artifacts. Large $r$ (near $N$ ) leads to insufficient modification, retaining the original \"dog\". An intermediate $r$ achieves the best balance of edit fidelity and background preservation." + ], + "image_footnote": [], + "bbox": [ + 137, + 85, + 864, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/688c66e8f6c106ca331aa7ab564e1530b9de38be41a3e9aad7276ef8f200bb23.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 168, + 250, + 261, + 323 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fe9623824347b79d8b3eae85c8e346bcfcf671b7f0433662285e409b6a31c709.jpg", + "image_caption": [ + "a sushi plate", + "$b = n$" + ], + "image_footnote": [], + "bbox": [ + 274, + 250, + 370, + 321 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3a709338daa81b755e8ccf9ea69cb47bfab24041cf42358f7cf6bba6ebb8c551.jpg", + "image_caption": [ + "#", + "Figure 7. Ablation study on blending latent step $b$ (increasing left to right). The prompt \"steak\" is applied to an image of \"sushi plate\" while increasing $b$ from left to right. At $b = n$ (left), the edit disrupts the original structure, affecting unmasked regions. As $b$ approaches $N$ (right), background preservation improves, and edits blend more seamlessly." + ], + "image_footnote": [], + "bbox": [ + 372, + 251, + 465, + 321 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/41509a13eaf4ae7e5924523f34d3aa009441bd40bd1e269c25688865ac0417fd.jpg", + "image_caption": [ + "edit prompt: steak" + ], + "image_footnote": [], + "bbox": [ + 467, + 251, + 558, + 321 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7950237fa58273d2802c1df99d299bb20d0e95f5b8211e9f594dd5925a130c47.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 562, + 251, + 653, + 321 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f59e2dd7dabb7bc3493e2fb362f36deb7799fc6835a0692d418ddb1022c0d4b4.jpg", + "image_caption": [ + "" + ], + "image_footnote": [], + "bbox": [ + 656, + 251, + 750, + 321 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8eeae6152bc0fc228e844a472e57441ab2e86833f4da74cf47384b2e91bb1875.jpg", + "image_caption": [ + "$b = N - 1$" + ], + "image_footnote": [], + "bbox": [ + 751, + 251, + 843, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "efficient latent caching (Sec. 3.1), minimizing computation and memory overhead ( $\\sim 1.25$ MB for 10 layers).", + "bbox": [ + 88, + 436, + 482, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quantitative Performance: LDB demonstrates a superior combination of speed, image quality, and edit fidelity across both benchmarks. On the PIE-Bench dataset, LDB achieves the best performance in six key metrics, excelling in human preference (HPS = 86.02), background preservation (LPIPS = 1.91), and text alignment (CS = 31.66), while also being the fastest method by a significant margin. This highlights its ability to generalize across a diverse set of editing tasks while maintaining high speed. Similarly, on the MagicBrush benchmark, LDB delivers strong performance with highest score in crucial metrics such as IR, AS, and PSNR. While BrushNet shows a slight advantage in some text alignment metrics, its practical usability is hindered by substantially slower runtime.", + "bbox": [ + 88, + 472, + 482, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1. Limitations and Future Work", + "text_level": 1, + "bbox": [ + 89, + 707, + 354, + 722 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Brush strength $(\\alpha)$ and diffusion step count $(n)$ coupling (Fig. 12) still requires minor user tuning across scenarios. Although preset profiles partially address this, future work could explore adaptive parameter tuning mechanisms to further improve usability. Moreover, semantically implausible edits (e.g. placing a boat in the sky) remain challenging due to inherent biases within diffusion models. Integrating techniques like semantic guidance could expand plausible edit ranges. Finally, responsible deployment necessitates robust watermarking [15] and provenance tracking to mitigate misuse and ensure transparency.", + "bbox": [ + 88, + 734, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Broader Applications", + "text_level": 1, + "bbox": [ + 511, + 436, + 714, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "LDB's training-free design only requires a standard iterative denoising process, which allows seamless integration into diverse diffusion models and applications requiring rapid editing. We validated this by integrating LDB to other commonly used methods, including DiT-based text-to-image (e.g., PixArt- $\\alpha$ [10]) and video generation models [7] without any model-specific tuning.", + "bbox": [ + 511, + 463, + 905, + 569 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Traditional diffusion-based video editing typically propagates edits from the first frame using additional supervision (e.g. optical flow [36]), risking temporal inconsistencies. LDB's high fidelity background preservation and efficiency naturally address these issues.", + "bbox": [ + 511, + 573, + 905, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We demonstrate preliminary success integrating LDB with Stable Video Diffusion (SVD) [7], editing the first frame and applying LDB's latent caching across frames for fast consecutive edits (see supplementary material, Fig. 15). This approach opens avenues for accelerated video manipulation, 3D asset editing, and collaborative design platforms.", + "bbox": [ + 511, + 654, + 905, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 767, + 635, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "LDB reimagines diffusion-based editing through latent caching and non-destructive layering, achieving unmatched speed and control. Quantitative results and user study show superior performance in image preference, edit fidelity, time, and usability. By bridging interactive editing with high-fidelity generative models, LDB can empower artists to iterate fluidly while maintaining artistic intent.", + "bbox": [ + 511, + 794, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17375", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the Natural Sciences and Engineering Research Council of Canada (NSERC). The authors would also like to thank Professors Leonid Signal and Kwang Moo Yi for their guidance and support throughout this project.", + "bbox": [ + 89, + 114, + 483, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 203, + 187, + 218 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 2", + "[2] Shubham Agarwal, Subrata Mitra, Sarthak Chakraborty, Srikrishna Karanam, Koyel Mukherjee, and Shiv Kumar Saini. Approximate caching for efficiently serving {Text-to-Image} diffusion models. In 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI 24), pages 1173-1189, 2024. 3", + "[3] Ideogram AI. Ideogram: Text-to-image generation platform. https://ideogram.ai, 2025.5", + "[4] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. 2", + "[5] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM Transactions on Graphics (TOG), 42 (4):1-11, 2023. 2, 3, 6", + "[6] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 707-723. Springer, 2022. 2, 3", + "[7] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 8, 2", + "[8] John Brooke. SUS-A quick and dirty usability scale. Usability evaluation in industry, 189(194), 1996. 5, 6", + "[9] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 2, 5, 6", + "[10] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 8", + "[11] Erin Cherry and Celine Latulipe. Quantifying the creativity support of digital tools through the creativity support index. ACM Transactions on Computer-Human Interaction (TOCHI), 21(4):1-25, 2014. 5, 9", + "[12] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. Ilvr: Conditioning method for denoising diffusion probabilistic models. arXiv preprint arXiv:2108.02938, 2021. 2" + ], + "bbox": [ + 93, + 227, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Civitai. Dreamshaper - 7 — stable diffusion checkpoint, 2024. Accessed on Feb 19, 2024. 4", + "[14] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 2", + "[15] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22466-22477, 2023. 8", + "[16] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2", + "[17] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 6", + "[18] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Han Hu, Dong Chen, et al. Instructdiffusion: A generalist modeling interface for vision tasks. arXiv preprint arXiv:2309.03895, 2023. 2", + "[19] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, et al. Instructdiffusion: A generalist modeling interface for vision tasks. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 12709-12720, 2024. 2", + "[20] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1", + "[21] Qin Guo and Tianwei Lin. Focus on your instruction: Fine-grained and multi-instruction image editing by attention modulation. arXiv preprint arXiv:2312.10113, 2023. 2", + "[22] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1, 2", + "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1", + "[24] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssm. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 6", + "[25] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1", + "[26] Minghui Hu, Yujie Wang, Tat-Jen Cham, Jianfei Yang, and Ponnuthurai N Suganthan. Global context with discrete diffusion in vector quantised modelling for image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11502-11511, 2022." + ], + "bbox": [ + 516, + 92, + 903, + 887 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17376", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Mingzhen Huang, Jialing Cai, Shan Jia, Vishnu Suresh Lokhande, and Siwei Lyu. Paralleledits: Efficient multi-aspect text-driven image editing with attention grouping. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3", + "[28] Adobe Inc. Adobe photoshop (2024 version). https://www.adobe.com/products/photoshop.htm1, 2024. 2, 5, 7, 9", + "[29] K. J. Joseph, Prateksha Udhayanan, Tripti Shukla, Aishwarya Agarwal, Srikrishna Karanam, Koustava Goswami, and Balaji Vasan Srinivasan. Iterative multi-granular image editing using diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 8107-8116, 2024. 3", + "[30] Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. Direct inversion: Boosting diffusion-based editing with 3 lines of code. arXiv preprint arXiv:2310.01506, 2023. 2, 3, 6", + "[31] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. In European Conference on Computer Vision, pages 150–168. Springer, 2024. 2, 6", + "[32] Kumara Kahatapitiya, Haozhe Liu, Sen He, Ding Liu, Menglin Jia, Chenyang Zhang, Michael S Ryoo, and Tian Xie. Adaptive caching for faster video generation with diffusion transformers. arXiv preprint arXiv:2411.02397, 2024. 3", + "[33] Kyoungkook Kang, Gyujin Sim, Geonung Kim, Donguk Kim, Seungho Nam, and Sunghyun Cho. Layeringdiff: Layered image synthesis via generation, then disassembly with generative knowledge. arXiv preprint arXiv:2501.01197, 2025. 3", + "[34] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2426-2435, 2022. 2", + "[35] Oran Lang, Yossi Gandelsman, Michal Yarom, Yoav Wald, Gal Elidan, Avinatan Hassidim, William T Freeman, Phillip Isola, Amir Globerson, Michal Irani, et al. Explaining in style: Training a gan to explain a classifier in stylespace. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 693-702, 2021. 2", + "[36] Feng Liang, Bichen Wu, Jialiang Wang, Licheng Yu, Kunpeng Li, Yinan Zhao, Ishan Misra, Jia-Bin Huang, Peizhao Zhang, Peter Vajda, et al. Flowvid: Taming imperfect optical flows for consistent video-to-video synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8207-8216, 2024. 8", + "[37] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 1, 2, 3", + "[38] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Deepcache: Accelerating diffusion models for free. In Proceedings of" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF conference on computer vision and pattern recognition, pages 15762-15772, 2024. 3", + "[39] Hayk Manukyan, Andranik Sargsyan, Barsegh Atanyan, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Hd-painter: High-resolution and prompt-faithful text-guided image inpainting with diffusion models. arXiv preprint arXiv:2312.14091, 2023. 2, 6", + "[40] Qi Mao, Lan Chen, Yuchao Gu, Zhen Fang, and Mike Zheng Shou. Mag-edit: Localized image editing in complex scenarios via mask-based attention-adjusted guidance. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6842-6850, 2024. 3", + "[41] Inc. Midjourney. Midjourney: Ai image generation. https://www.midjourney.com, 2025.5", + "[42] Daiki Miyake, Akihiro Iohara, Yu Saito, and Toshiyuki Tanaka. Negative-prompt inversion: Fast image inversion for editing with text-guided diffusion models. arXiv preprint arXiv:2305.16807, 2023. 3", + "[43] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 3", + "[44] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag your gan: Interactive point-based manipulation on the generative image manifold. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2", + "[45] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, New York, NY, USA, 2023. Association for Computing Machinery. 2", + "[46] Thomas Porter and Tom Duff. Compositing digital images. In Proceedings of the 11th annual conference on Computer graphics and interactive techniques, pages 253-259, 1984. 3", + "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 6", + "[48] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 6", + "[49] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2287–2296, 2021. 2", + "[50] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 5, 6", + "[51] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17377", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tuning text-to-image diffusion models for subject-driven generation, 2023. 2", + "[52] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2", + "[53] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1", + "[54] Vishnu Sarukkai, Linden Li, Arden Ma, Christopher Ré, and Kayvon Fatahalian. Collage diffusion. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 4208-4217, 2024. 3, 4", + "[55] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in neural information processing systems, 35:25278-25294, 2022. 6", + "[56] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 1", + "[57] Felix Wimbauer, Bichen Wu, Edgar Schoenfeld, Xiaoliang Dai, Ji Hou, Zijian He, Artsiom Sanakoyeu, Peizhao Zhang, Sam Tsai, Jonas Kohler, et al. Cache me if you can: Accelerating diffusion models through block caching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6211-6220, 2024. 3", + "[58] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 6", + "[59] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22428-22437, 2023. 2", + "[60] Zishan Xu, Xiaofeng Zhang, Wei Chen, Minda Yao, Jueting Liu, Tingting Xu, and Zehua Wang. A review of image inpainting methods based on deep learning. Applied Sciences, 13(20):11189, 2023. 2", + "[61] Fei Yang, Shiqi Yang, Muhammad Atif Butt, Joost van de Weijer, et al. Dynamic prompt learning: Addressing cross-attention leakage for text-based image editing. Advances in Neural Information Processing Systems, 36:26291–26303, 2023. 2", + "[62] Shiyuan Yang, Xiaodong Chen, and Jing Liao. Uni-paint: A unified framework for multimodal image inpainting with pretrained diffusion model. In Proceedings of the 31st ACM International Conference on Multimedia, pages 3190-3199, 2023. 2", + "[63] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything:" + ], + "bbox": [ + 91, + 92, + 482, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. 2", + "[64] Ziyun Zeng, Hang Hua, Jianlong Fu, Jiebo Luo, et al. Promptfix: You prompt and we fix the photo. Advances in Neural Information Processing Systems, 37:40000-40031, 2025. 2", + "[65] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 36, 2024. 5, 6", + "[66] Lvmin Zhang and Maneesh Agrawala. Transparent image layer diffusion using latent transparency. arXiv preprint arXiv:2402.17113, 2024. 4", + "[67] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3836-3847, 2023. 2", + "[68] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6", + "[69] Junhao Zhuang, Yanhong Zeng, Wenran Liu, Chun Yuan, and Kai Chen. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. In European Conference on Computer Vision, pages 195-211. Springer, 2024. 2" + ], + "bbox": [ + 516, + 92, + 903, + 473 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "17378", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Streamlining Image Editing with Layered Diffusion Brushes/6116442a-236c-4dff-8b3c-c87a9184cffb_model.json b/2025/Streamlining Image Editing with Layered Diffusion Brushes/6116442a-236c-4dff-8b3c-c87a9184cffb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..91cd33bfc82c37e8d88370aa0705d6f3e85cb8e1 --- /dev/null +++ b/2025/Streamlining Image Editing with Layered Diffusion Brushes/6116442a-236c-4dff-8b3c-c87a9184cffb_model.json @@ -0,0 +1,2994 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.195, + 0.123, + 0.804, + 0.145 + ], + "angle": 0, + "content": "Streamlining Image Editing with Layered Diffusion Brushes" + }, + { + "type": "text", + "bbox": [ + 0.359, + 0.17, + 0.641, + 0.205 + ], + "angle": 0, + "content": "Peyman Gholami Robert Xiao University of British Columbia" + }, + { + "type": "text", + "bbox": [ + 0.394, + 0.208, + 0.605, + 0.222 + ], + "angle": 0, + "content": "{peymang, brx}@cs.ubc.ca" + }, + { + "type": "text", + "bbox": [ + 0.279, + 0.233, + 0.713, + 0.248 + ], + "angle": 0, + "content": "https://layered-diffusion-brushes.github.io/" + }, + { + "type": "text", + "bbox": [ + 0.191, + 0.268, + 0.245, + 0.311 + ], + "angle": 0, + "content": "Edit prompt (type)" + }, + { + "type": "text", + "bbox": [ + 0.284, + 0.271, + 0.352, + 0.295 + ], + "angle": 0, + "content": "sunglasses (add objects)" + }, + { + "type": "text", + "bbox": [ + 0.383, + 0.27, + 0.468, + 0.308 + ], + "angle": 0, + "content": "hat, Takashi Murakami style (change style)" + }, + { + "type": "text", + "bbox": [ + 0.487, + 0.27, + 0.578, + 0.319 + ], + "angle": 0, + "content": "starry night, van gogh style (fix problems, maintaining style" + }, + { + "type": "text", + "bbox": [ + 0.602, + 0.27, + 0.683, + 0.308 + ], + "angle": 0, + "content": "sky, Leonid Afremov style (mixing styles)" + }, + { + "type": "text", + "bbox": [ + 0.7, + 0.27, + 0.8, + 0.308 + ], + "angle": 0, + "content": "ornate frame (change attribute / texture)" + }, + { + "type": "text", + "bbox": [ + 0.191, + 0.328, + 0.234, + 0.354 + ], + "angle": 0, + "content": "Layer Mask" + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.32, + 0.368, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.32, + 0.478, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.32, + 0.587, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.32, + 0.696, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.32, + 0.807, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.191, + 0.426, + 0.237, + 0.454 + ], + "angle": 0, + "content": "Edited image" + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.404, + 0.368, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.29, + 0.497, + 0.339, + 0.51 + ], + "angle": 0, + "content": "Layer 1" + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.404, + 0.478, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.4, + 0.497, + 0.449, + 0.51 + ], + "angle": 0, + "content": "Layer 2" + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.404, + 0.588, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.497, + 0.558, + 0.51 + ], + "angle": 0, + "content": "Layer 3" + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.404, + 0.696, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.619, + 0.497, + 0.668, + 0.51 + ], + "angle": 0, + "content": "Layer 4" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.404, + 0.806, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.728, + 0.497, + 0.776, + 0.51 + ], + "angle": 0, + "content": "Layer 5" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.522, + 0.908, + 0.566 + ], + "angle": 0, + "content": "Figure 1. Hierarchical image editing with Layered Diffusion Brushes: LDB is capable of creating and stacking a wide range of independent edits, including object addition, removal, or replacement, colour and style changes/combining, and object attribute modification. Each edit is performed independently, and users are able to switch between the edits seamlessly." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.587, + 0.327, + 0.603 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.614, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Denoising diffusion models have emerged as powerful tools for image manipulation, yet interactive, localized editing workflows remain underdeveloped. We introduce Layered Diffusion Brushes (LDB), a novel training-free framework that enables interactive, layer-based editing using standard diffusion models. LDB defines each \"layer\" as a self-contained set of parameters guiding the generative process, enabling independent, non-destructive, and fine-grained prompt-guided edits, even in overlapping regions. LDB leverages a unique intermediate latent caching approach to reduce each edit to only a few denoising steps, achieving 140 ms per edit on consumer GPUs. An editor implementing LDB, incorporating familiar layer concepts, was evaluated via user study and quantitative metrics. Results demonstrate LDB's superior speed alongside comparable or improved image quality, background preservation, and edit fidelity relative to state-of-the-art methods across various sequential image manipulation tasks. The findings highlight LDB's ability to significantly enhance creative" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.589, + 0.907, + 0.635 + ], + "angle": 0, + "content": "workflows by providing an intuitive and efficient approach to diffusion-based image editing and its potential for expansion into related subdomains, such as video editing." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.663, + 0.646, + 0.679 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.689, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Image editing has undergone transformative advancements with the rise of text-to-image (T2I) generative models, enabling unprecedented creative expression through textual guidance. These models, including Generative Adversarial Networks (GANs) [20], Variational Autoencoders (VAEs), and Denoising Diffusion Models (DMs) [23], have redefined image synthesis and manipulation. Among these, DMs [56] have emerged as the state of the art due to their training stability, high-fidelity outputs, and versatility across tasks like inpainting [37], super-resolution [53], and style transfer [22]. However, despite their capabilities, a critical gap remains: enabling real-time, localized, and iterative edits that align with professional workflows, where artists demand precise control over specific regions without" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17368" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.32, + 0.108 + ], + "angle": 0, + "content": "disrupting the global composition." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.283 + ], + "angle": 0, + "content": "Existing DM-based editing methods face several core challenges. First, their stochastic nature often necessitates numerous generations to achieve desired results [5]. Second, they lack intuitive mechanisms for layered, nondestructive editing—a cornerstone of tools like Adobe Photoshop [28]—where edits can be independently adjusted, stacked, or removed. Third, while mask-guided approaches enable regional control, they struggle with seamless blending, artifact-free transitions, and real-time feedback. These limitations restrict their adoption in creative pipelines, where rapid iteration and granular control are critical." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.289, + 0.484, + 0.351 + ], + "angle": 0, + "content": "To address these challenges, we propose Layered Diffusion Brushes (LDB), a novel framework based on Latent Diffusion Models (LDM) [50] that integrates mask-guided diffusion with a non-destructive layered editing paradigm." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.357, + 0.484, + 0.465 + ], + "angle": 0, + "content": "At its core, LDB introduces new noise patterns into the image latents during diffusion process, guided by both the user-specified mask and the edit prompt. This preserves the original context while seamlessly integrating localized edits. We implement an intuitive user interface (UI) with a layering system to support consecutive edits (Fig. 1). Specifically, as key contributions, LDB introduces:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.47, + 0.483, + 0.546 + ], + "angle": 0, + "content": "- Latent Caching for Real-Time Edits: By reusing intermediate denoising states from initial generation, edits bypass redundant computations and achieve as low as 140 ms per edit on \\(512 \\times 512\\) images (\\(53 \\times\\) faster than BrushNet [31] using the same consumer GPUs)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.547, + 0.483, + 0.59 + ], + "angle": 0, + "content": "- Non-destructive Layered Editing: LDB introduces an order-agnostic layering mechanism by defining the concept of layers for DMs, enabling:" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.592, + 0.481, + 0.621 + ], + "angle": 0, + "content": "- Region-targeted adjustments with background preservation, using mask-prompt pairs," + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.622, + 0.481, + 0.651 + ], + "angle": 0, + "content": "- Stacking, toggling, or deleting layers without cross-interference—even in overlapping regions," + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.652, + 0.481, + 0.68 + ], + "angle": 0, + "content": "- Post-hoc revision of edits while preserving underlying content." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.682, + 0.483, + 0.756 + ], + "angle": 0, + "content": "- Seed-Driven Exploration: Our UI provides familiar \"brush\" and \"scroll\" gestures to enable instant exploration of variations by modulating noise seeds, bridging stochastic generation with deterministic refinement and instant feedback." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.47, + 0.483, + 0.756 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We validate LDB through extensive experiments and a user study with graphic designers. Quantitatively, LDB outperforms state-of-the-art methods in terms of speed and image quality and is comparable in terms of edit fidelity. The user study revealed superior usability and creativity support in iterative design. Additionally, LDB is a plug-and-play, training-free system adaptable to existing models and applications, and we demonstrate this by applying LDB to the task of video editing." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.655, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.116, + 0.745, + 0.133 + ], + "angle": 0, + "content": "2.1. DM-based Image Editing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.14, + 0.907, + 0.276 + ], + "angle": 0, + "content": "Image editing is the task of modifying existing images in terms of appearance, structure, or composition, ranging from subtle adjustments to major transformations. Unlike GAN-based approaches [1, 35, 44], which are prone to limitations in inversion stability [49] and localized control [6], diffusion-based methods harness the power of controllable, high-quality DMs in various image-editing tasks, including text and image-driven image manipulation studies [14, 26, 34, 37]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.277, + 0.906, + 0.383 + ], + "angle": 0, + "content": "Instruction-based text editing methods [9, 18, 19, 21, 64] typically train DMs on instruction-image pairs. For example, InstructPix2Pix [9] is trained using synthetic pairs from Stable Diffusion [50] and Prompt-to-Prompt [22]. However, expressing nuanced edits solely through text instructions remains challenging, particularly for object-specific style or color changes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.385, + 0.906, + 0.552 + ], + "angle": 0, + "content": "Mask-based methods [4, 5, 14, 63] sample within specified regions. While effective for localized edits, they can introduce unintended global changes, especially problematic in sequential editing, and may struggle with complex edits. For instance, Blended Latent Diffusion's lossy VAE latent space hinders perfect reconstruction even before noise addition [5]. Though a background reconstruction strategy is included, it increases computation and may still yield incoherent results for complex edits. Conversely, our method directly modifies the original latent space, enhancing context preservation and natural blending." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.734 + ], + "angle": 0, + "content": "Attention-based editing manipulates cross-attention maps to guide the image generation process toward the desired modifications [22, 45]. These methods generally face challenges in achieving fine-grained edits without unwanted global modifications. Yang et al. [61] attribute unintended changes to inaccurate attention maps and propose attention focusing. Inversion-based methods like ILVR [12], Textual Inversion [16], and DreamBooth [51] focus on context modification while preserving subjects. DDIM inversion converts images to noisy latents, and sampling generates edited results based on prompts. We employ Direct Inversion [30] for efficient real image latent inversion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Image inpainting involves replacing or restoring the missing regions while maintaining global coherency [60]. Many inpainting works [39, 50, 62, 69] require using a fine-tuned DM specifically designed for inpainting tasks, limiting their applicability. Some, including SmartBrush, which uses object-mask prediction guidance [59], offer more flexibility. PowerPaint [69] introduces learnable task embeddings for improved control. While these models effectively generate new content, they are generally unsuitable for making small, targeted adjustments [4, 37, 52]. Inspired by ControlNet [67], BrushNet [31] builds a decomposed" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17369" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.138 + ], + "angle": 0, + "content": "plug-and-play dual-branch DM, but struggles with real-time interaction due to its computational overhead. In Sec. 4.1 we compare LDB with several inpainting techniques." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.149, + 0.421, + 0.166 + ], + "angle": 0, + "content": "2.2. Layered and Sequential Image editing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.172, + 0.483, + 0.248 + ], + "angle": 0, + "content": "Layer-based image editing is fundamental in computer graphics [46], and recent works integrate this concept into AI methodologies [6, 54]. Layered representations enable dynamic manipulation of image components, transforming single images into multi-layered structures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.249, + 0.484, + 0.462 + ], + "angle": 0, + "content": "LayeringDiff [33] decomposes images into foreground and background. ParallelEdits [27] uses attention for efficient multi-aspect text edits. MAG-Edit [40] employs a two-layer process with attention injection to a single edit from background. Joseph et al. [29] highlight error accumulation in sequential editing, where artifacts compound across edits. Collage Diffusion [54], built on modified Blended Latent Diffusion [5], employs alpha masks to guide cross-attention and generate harmonized images while respecting scene composition. However, it assumes prelayered inputs and synthesizes scenes from scratch. In contrast, LDB is training-free, operates directly on existing images, and supports fully independent layers—unlike methods such as [6] that require per-image training." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.472, + 0.424, + 0.489 + ], + "angle": 0, + "content": "2.3. Accelerated Generation using Caching" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.495, + 0.484, + 0.721 + ], + "angle": 0, + "content": "Caching and reusing intermediate features has proven effective for accelerating DM inference through reducing redundant computations. Several works have utilized caching in diffusion transformers (DiTs) for video generation. DeepCache [38] reuses high-level U-Net features in video generation, while AdaCache [32] dynamically adjusts cached residuals based on temporal content. Cache Me If You Can [57] employs block caching by reusing outputs from layer blocks of previous steps during inference. For image generation, Approximate Caching [2] reuses intermediate latents created during prior image generation processes for similar prompts. We employ a similar strategy through caching key latent representations and adapt it specifically for interactive image editing, enabling the real-time feedback that is crucial for creative workflows." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.738, + 0.182, + 0.754 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.764, + 0.483, + 0.84 + ], + "angle": 0, + "content": "We use an LDM-based variant of image generative models and make intermediate adjustments to the latent space, similar to [5, 37]. Therefore, LDB requires no additional training or fine-tuning of the underlying LDM; all modifications are applied during the reverse diffusion process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We adopt the standard LDM formulation, where image generation begins with a sample from a Gaussian distribution, \\( Z_0 \\sim \\mathcal{N}(0, \\sigma_{max}^2 I) \\) and is iteratively denoised through a sequence of steps \\( N \\), resulting in a series of la" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "tents \\(Z_{i}\\) corresponding to decreasing noise levels \\(\\sigma_{i}\\), where \\(\\sigma_0 = \\sigma_{max} > \\sigma_1 > \\dots >\\sigma_N\\approx 0\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.122, + 0.905, + 0.168 + ], + "angle": 0, + "content": "As demonstrated in Fig. 2, the overall LDB pipeline comprises three key stages: initial image generation (or inversion), latent caching, and iterative layered editing." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.169, + 0.907, + 0.29 + ], + "angle": 0, + "content": "For DM-generated images, we first initialize the sample \\( Z_{0} = \\epsilon_{0} \\) and noise level \\( \\sigma_0 \\) (\\( i = 0 \\)). For real images, the initial noise latent is obtained using inversion. We use Direct Inversion [30] for its high speed and comparable performance to other inversion methods, including Null-Text Inversion [43] and Negative-Prompt Inversion [42]. The noisy sample then undergoes the diffusion process, caching certain intermediate latents to facilitate editing." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.301, + 0.671, + 0.317 + ], + "angle": 0, + "content": "3.1. Latent Caching" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.323, + 0.905, + 0.383 + ], + "angle": 0, + "content": "To enable rapid, interactive editing with instant exploration and feedback, we employ latent caching to reuse intermediate representations in subsequent steps, minimizing redundant computations. We store two key intermediate latents:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.384, + 0.906, + 0.534 + ], + "angle": 0, + "content": "- Regeneration Latent \\( \\mathbf{Z}_{\\mathbf{r}} \\): At diffusion step \\( r = N - n \\), where \\( N \\) is the total number of diffusion steps for initial image generation and \\( n \\) is the number of editing steps, we cache the latent \\( Z_{r} \\), which serves as the starting point for all subsequent edits. By reusing \\( Z_{r} \\), we avoid recomputing the initial denoising steps for each new edit, significantly speeding up the editing process (from \\( N \\) denoising steps to \\( n \\)). Effectively, \\( Z_{r} \\) represents a partially denoised latent state that retains the global image structure but is still malleable enough to accommodate localized edits." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.535, + 0.907, + 0.687 + ], + "angle": 0, + "content": "- Blending Latent \\( Z_{\\mathrm{b}} \\): We cache the latent at diffusion step \\( b \\) which is specifically used for the layer merging process (Algorithm 1, line 7). We set \\( b = N - 2 \\) for maximum background preservation (as discussed in Sec. 4.3). \\( Z_{b} \\) represents a more denoised latent compared to \\( Z_{r} \\), capturing more refined image details while still allowing for seamless blending of new edits into the existing image context. Utilizing this cached blending latent ensures smoother integration of edits and reduces visual artifacts at layer boundaries during the merging process." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.384, + 0.907, + 0.687 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.697, + 0.818, + 0.714 + ], + "angle": 0, + "content": "3.2. Layered Diffusion Brushes Editing" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.72, + 0.906, + 0.825 + ], + "angle": 0, + "content": "To initiate an edit, the algorithm begins by generating a new noise pattern \\( Z_0' = \\epsilon_k' \\), sampled from \\( \\mathcal{N}(0, \\sigma^2 I) \\) using a different seed \\( S' \\), and scaling it to match the variance of the cached latent \\( Z_r \\). This ensures that the additive noise stays in a reasonable range from the latent for editing, preventing visual artifacts. \\( Z_0' \\) is then added to the regeneration latent \\( Z_r \\), controlled by the mask \\( m \\) and strength \\( \\alpha \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In the editing stage, at step \\( b \\), a new noisy sample is merged with the cached blending latent using the strength control and the mask, resulting in \\( Z_b' \\). Subsequently, the new latent is progressively denoised from steps \\( b \\) through \\( N \\) and processed through the VAE to output edited image \\( I' \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17370" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.179, + 0.089, + 0.824, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.384, + 0.907, + 0.442 + ], + "angle": 0, + "content": "Figure 2. Overview of the Proposed Method: The top box shows standard DM-based image generation from noisy latent \\( Z_{0} \\) and prompt \\( \\mathcal{P} \\). The middle section depicts the latent caching module, storing and retrieving intermediate latents for different layers. The bottom box illustrates the editing process: a new noise sample \\( S' \\) merges with the original latent at step \\( r \\) using mask \\( m \\) and strength control \\( \\alpha \\). Diffusion continues until step \\( b \\), where modified and cached latents blend to generate the final edited image." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.466, + 0.483, + 0.499 + ], + "angle": 0, + "content": "Algorithm 1 presents the pseudocode for the editing process for a single layer (for simplicity):" + }, + { + "type": "code_caption", + "bbox": [ + 0.103, + 0.513, + 0.426, + 0.529 + ], + "angle": 0, + "content": "Algorithm 1: LDB editing process (single layer)" + }, + { + "type": "algorithm", + "bbox": [ + 0.095, + 0.53, + 0.485, + 0.751 + ], + "angle": 0, + "content": "Input: Edit prompt \\(\\mathcal{P}'\\), Mask \\(m \\in [0,1]^{H \\times W}\\), Random seed \\(S'\\), Strength \\(\\alpha\\), Number of edit steps \\(n\\), Regeneration latent \\(Z_r\\), Blending latent \\(Z_b\\) \nOutput: Edited latent \\(Z_N'\\) \n1 \\(Z_0' \\gets \\epsilon_{nk}' \\sim \\mathcal{N}(0, \\sigma^2 I) //\\) sampled using seed \\(S'\\) \n2 \\(Z_0' \\gets \\sqrt{\\text{Var}(Z_r)} \\cdot Z_0' //\\) scale new sample \n3 \\(Z_0' \\gets Z_r + \\alpha \\cdot (Z_0' \\odot m) //\\) noise injection \n4 for \\(i = 0,1,\\ldots,n\\) do \n5 \\(Z_{i+1}' \\gets DM(Z_i', \\mathcal{P}', i, S')\\) \n6 if \\(i == b\\) then \n7 \\(Z_b' \\gets Z_b' \\odot m + Z_b \\odot (1 - m) //\\) blending \n8 end \n9 Return \\(Z_N'\\)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.777, + 0.275, + 0.793 + ], + "angle": 0, + "content": "3.3. Layer Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.8, + 0.483, + 0.876 + ], + "angle": 0, + "content": "Unlike prior works that rely on transparent decomposable layers [66] or explicit object segmentation [54], we redefine a layer as a self-contained set of reproducible parameters that govern localized edits. For layer \\(\\mathcal{L}_k\\), we formalize this as a generalized version of parameters in Algorithm 1:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.881, + 0.483, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {(k)} = \\left(\\mathbf {S} ^ {\\prime (k)}, \\mathbf {m} ^ {(k)}, \\mathbf {v} ^ {(k)}, \\mathbf {Z} _ {r} ^ {(k)}, \\mathbf {Z} _ {b} ^ {(k)}, \\alpha^ {(k)}, n ^ {(k)}, \\mathcal {P} ^ {\\prime (k)}, j\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.464, + 0.848, + 0.481 + ], + "angle": 0, + "content": "- \\(\\mathbf{S}'^{(k)} \\in \\mathbb{Z}^+\\): Seed space for stochastic variations" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.481, + 0.732, + 0.496 + ], + "angle": 0, + "content": "- \\(\\mathbf{m}^{(k)}\\in [0,1]^{H\\times W}\\) : Edit mask" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.496, + 0.725, + 0.512 + ], + "angle": 0, + "content": "- \\(\\mathbf{v}^{(k)}\\in \\{0,1\\}\\): Visibility state" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.512, + 0.891, + 0.529 + ], + "angle": 0, + "content": "- \\(\\mathbf{Z}_r^{(k)},\\mathbf{Z}_b^{(k)}\\in \\mathbb{R}^{C\\times H\\times W}\\) : Regeneration/blending latents" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.529, + 0.758, + 0.544 + ], + "angle": 0, + "content": "- \\(\\alpha^{(k)}\\in [0,1]\\) : Layer strength value" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.544, + 0.799, + 0.56 + ], + "angle": 0, + "content": "- \\(\\mathbf{n}^{(k)}\\in [0,N]\\) Number of denoising steps" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.56, + 0.657, + 0.577 + ], + "angle": 0, + "content": "\\(\\mathcal{P}^{\\prime (k)}\\) : Edit prompt" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.577, + 0.752, + 0.592 + ], + "angle": 0, + "content": "- \\( j \\in {\\mathbb{Z}}^{ + } \\) : Index of last layer index." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.464, + 0.891, + 0.592 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.907, + 0.736 + ], + "angle": 0, + "content": "Notably, within a given layer \\(\\mathcal{L}_k\\) with previous layer \\(\\mathcal{L}_j\\), the cached latents \\(\\mathbf{Z}_r^{(j)}\\) and \\(\\mathbf{Z}_b^{(j)}\\) inherently incorporate the cumulative edits from all preceding layers. This is because edits to layer \\(\\mathcal{L}_k\\), are applied to the already edited output of layer \\(\\mathcal{L}_j\\) which serves as the input to the diffusion process and the algorithm always keeps the last layer updated. Therefore, any modification in a previous layer automatically propagates through the subsequent layers. By defining \\(\\Phi\\) as a single-layer latent generation and caching step as:" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.747, + 0.907, + 0.768 + ], + "angle": 0, + "content": "\\[\n\\left(Z _ {r} ^ {(k)}, Z _ {b} ^ {(k)}\\right) = \\Phi \\left(\\mathcal {L} ^ {(k)}, \\mathcal {L} ^ {(j)}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "in essence, if a given layer \\(\\mathcal{L}^{(i)}\\) (where \\(i < k\\)) is removed or its visibility \\(\\mathbf{v}^{(i)}\\) is toggled, the operator \\(\\Phi\\) will be recursively invoked to recreate all latents for layers from \\(\\mathcal{L}^{(i)}\\) to \\(\\mathcal{L}^{(k)}\\). This recomputation, accelerated by latent caching, is automatically triggered and typically completes within milliseconds to a few seconds, depending on the number of layers. This design allows edits to remain independent yet seamlessly integrated into the final composition." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17371" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.092, + 0.285, + 0.108 + ], + "angle": 0, + "content": "3.3.1. Overlapping Regions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.111, + 0.484, + 0.217 + ], + "angle": 0, + "content": "A key advantage of layered editing in LDB is the ability to create overlapping edits, where one layer can partially or fully modify areas affected by earlier layers. This requires careful handling of each layer's regeneration latent, \\( Z_{r} \\), to ensure that changes in visibility or content from higher layers are accurately reflected in subsequent layers, even in overlapping regions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.218, + 0.484, + 0.359 + ], + "angle": 0, + "content": "By default, all layers use the initial image's latent \\((Z_{r})\\) as their regeneration latent. However, this approach fails to account for overlapping edits from preceding layers. To address this, when processing a layer \\(k\\), we compute its regeneration latent by inverting the output image of the previous layer \\((I^{\\prime (k)})\\) as shown using the feedback arrow on Fig. 2. This inversion yields \\(Z_0^{(k)}\\), which is then sent through the generation stage in LDB. Both \\(Z_{r}^{(k)}\\) and \\(Z_{b}^{(k)}\\) are cached for efficient processing (as shown in Fig. 3)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.36, + 0.483, + 0.421 + ], + "angle": 0, + "content": "This mechanism enables precise control and seamless integration of edits across overlapping regions. Changes to any layer propagate correctly without introducing artifacts, offering flexibility and fine-grained control." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.432, + 0.419, + 0.449 + ], + "angle": 0, + "content": "3.4. User-Interface and Interaction Design" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.455, + 0.484, + 0.531 + ], + "angle": 0, + "content": "To develop a practical tool for artists and designers, we designed an custom UI that balances control and simplicity. The UI allows users to generate, upload, and edit images, manage layers, and adjust parameters seamlessly. Two interaction modes streamline edits (Fig. 4):" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.532, + 0.484, + 0.607 + ], + "angle": 0, + "content": "Box Mode: Users can click or drag on the image to move a resizeable square mask around. This option enables a quick and interactive exploration of how various parts of the image will change in response to a given set of editing settings (prompt and strength), simply by moving the cursor." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.608, + 0.483, + 0.682 + ], + "angle": 0, + "content": "Custom Mask Mode: Users can draw free-form masks over the desired around and navigate between new generation samples by scrolling the mouse up or down while hovering over the image, allowing them to rapidly explore variations on their edit." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.685, + 0.483, + 0.729 + ], + "angle": 0, + "content": "We propose a workflow where users first position edits spatially using Box Mode, then refine mask geometry and appearance details via Custom Mask Mode." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.731, + 0.483, + 0.791 + ], + "angle": 0, + "content": "Layering capabilities include stacking, visibility toggling, and deletion. Each layer is independently modifiable. Detailed information on the UI design user interactions and a demo video can be found in supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.807, + 0.222, + 0.824 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.214, + 0.849 + ], + "angle": 0, + "content": "4.1. User Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We conducted a user study in order to evaluate the effectiveness of LDB for providing targeted image fine-tuning, using two other well-known existing image editing tools," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "InstructPix2Pix (IP2P) [9] and Stable Diffusion Inpainting (SDI) [50] as baselines for comparison." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.123, + 0.905, + 0.214 + ], + "angle": 0, + "content": "We recruited a cohort of seven expert participants with extensive experience in using image editing software. As part of our selection criteria, we ensured that all had at least a basic level of familiarity with AI image generation techniques [3, 41] and were regular users of editing software, such as Adobe Photoshop [28] for creating visual art." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.225, + 0.829, + 0.24 + ], + "angle": 0, + "content": "4.1.1. Study Procedure and Task Description" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.245, + 0.905, + 0.32 + ], + "angle": 0, + "content": "Each user engaged in two types of tasks: free-form tasks where users generated an image for editing using a fixed prompt and seed (type 1), and pre-determined tasks where the user worked with existing real images from the MagicBrush dataset [65] (type 2)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.322, + 0.905, + 0.367 + ], + "angle": 0, + "content": "For the type 1 tasks, we selected specific types of edits that showcase various functionalities and capabilities of the system, including:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.368, + 0.905, + 0.383 + ], + "angle": 0, + "content": "1. Stack layers and create sequential edits (draw with LDB)" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.384, + 0.804, + 0.398 + ], + "angle": 0, + "content": "2. Modify attributes and features of objects" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.399, + 0.797, + 0.413 + ], + "angle": 0, + "content": "3. Correct image imperfections and errors" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.414, + 0.803, + 0.428 + ], + "angle": 0, + "content": "4. Enhance discernibility of similar objects" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.429, + 0.905, + 0.535 + ], + "angle": 0, + "content": "5. Target specific regions for style transfer, refine aesthetics Type 2 tasks were more structured, with the mask, edit prompt, and input images provided by the dataset. The dataset provides manually annotated masks and instructions for each edit. We selected a subset of 35 input images, each containing up to three layers of edits. Users refined masks/parameters if necessary and completed editing tasks." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.368, + 0.905, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.905, + 0.597 + ], + "angle": 0, + "content": "Figure 5, second row, shows example edits generated by the participants. Additional examples are provided in the supplementary material. As shown, LDB produces targeted edits that integrate seamlessly with the images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.608, + 0.744, + 0.623 + ], + "angle": 0, + "content": "4.1.2. Evaluation Survey Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.906, + 0.78 + ], + "angle": 0, + "content": "The participants completed a three stage evaluation survey following the image editing tasks. The first part included a System Usability Scale (SUS) form to rate the usability, ease of use, design, and performance of each method. SUS is a standard usability evaluation survey widely used in user-experience literature [8]. Overall, participants indicated that they are more likely to use LDB compared to IP2P and SDI, and that they find it the easiest tool to use. LDB obtained a SUS score of \\(80.35\\%\\), while IP2P and SDI achieved a SUS of \\(38.21\\%\\) and \\(37.5\\%\\) respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.906, + 0.901 + ], + "angle": 0, + "content": "The SUS survey was followed by a Creativity Support Index [11] survey to evaluate the system's degree of creative work support. Participants expressed positivity towards LDB, indicating that it enhanced their enjoyment, exploration, expressiveness, and immersion, while also deeming the results worth their effort. Lastly, the survey was followed by a semi-structured interview where participants appreciated the intuitiveness, ease of use and versatility of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17372" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.599, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.303, + 0.601, + 0.346 + ], + "angle": 0, + "content": "Figure 3. Overlapping edit regions in LDB: overlapping edits enable complex, interacting modifications. For example, one layer can adjust color while another changes shape, with the final result combining both." + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.092, + 0.906, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.656, + 0.217, + 0.907, + 0.342 + ], + "angle": 0, + "content": "Figure 4. Box and Custom Mask Options: In box mode, users click the target region's center to generate edits within the specified area and can drag the box to explore variations instantly. In custom mask mode, users draw a mask over the desired region and adjust the seed using the mouse wheel or scrolling gestures to generate new variations." + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.358, + 0.889, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.545, + 0.908, + 0.575 + ], + "angle": 0, + "content": "Figure 5. Qualitative editing results on PIE-Bench (top) and MagicBrush (bottom) benchmarks using different methods. Edit prompts are presented on top of each row. More examples available in supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.483, + 0.631 + ], + "angle": 0, + "content": "LDB. Further details about the study, interview, results, and discussion can be found in supplemental material." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.638, + 0.295, + 0.654 + ], + "angle": 0, + "content": "4.2. Quantitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.704 + ], + "angle": 0, + "content": "To quantitatively evaluate the performance of LDB, we employed a comprehensive suite of metrics, aligning with established practices in image editing evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.765 + ], + "angle": 0, + "content": "Specifically, for text-image alignment, we used CLIP Score (CS) [47] for global alignment, CS-L for masked-region alignment, and CS-D [17] for consistency between image and caption changes in CLIP space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.885 + ], + "angle": 0, + "content": "We adopted Learned Perceptual Image Patch Similarity (LPIPS) [68] and Peak Signal-to-Noise Ratio (PSNR) [24] for evaluating content preservation and pixel-level fidelity in unmasked regions. Furthermore, to gauge overall image quality and aesthetic appeal, we incorporated Aesthetic Score (AS) [55], Image Reward (IR), and Human Preference Score V2 (HPS) [58], the latter two reflecting human-aligned preferences." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We compared LDB against a diverse set of state-of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.907, + 0.69 + ], + "angle": 0, + "content": "the-art editing and inpainting methods, including InstructPix2Pix (IP2P) [9], Stable Diffusion Inpainting (SDI) [50], HD-Painter (HDP) [39], BrushNet (BN) [31], and Blended Latent Diffusion (BLD) [5], on two benchmarks: MagicBrush [65] and PIE-Bench [30]. For MagicBrush, we also report results on the provided ground truth (GT) images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.691, + 0.909, + 0.81 + ], + "angle": 0, + "content": "Quantitative results are summarized in Tab. 1. All methods were evaluated using their default editing settings, except for LDB, IP2P, and SDI on the MagicBrush benchmark, where we used user-edited images from our user study for consecutive edits. Inference times denote average per-edit durations, measured on a single NVIDIA RTX 4090 GPU with \\( N = 25 \\) diffusion steps for baseline methods and \\( n = 8 \\) for LDB." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.667, + 0.835 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We perform three ablation studies for two main components of the LDB caching mechanism, i.e. the caching timesteps for the regeneration latent \\((r)\\), and the blending latent \\((b)\\). We also ablate and discuss the effect of strength control \\(\\alpha\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17373" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.137, + 0.089, + 0.86, + 0.328 + ], + "angle": 0, + "content": "
BenchmarkMethodImage QualityMasked Region PreservationText AlignmentTime (s)
IR ×10↑HPS ×102↑AS↑PSNR↑LPIPS ×102↓CS↑CS-L↑CS-D ×102↑(per edit) ↓
MagicBrushIP2P-62.8321.165.297.2815.0729.3922.016.641.72
SDI-39.2120.885.4812.208.7030.0822.154.111.84
HDP-20.6923.275.4412.056.1331.0122.069.8912.85
BN-0.0422.575.7311.558.7531.1622.1712.927.49
BLD-24.1022.805.4812.646.9430.6421.9910.051.41
GT-1.9322.625.3617.642.3030.7522.149.78NA
Ours7.7422.655.7412.857.0531.0422.079.540.26
PIE-BenchIP2P-40.7323.125.76172.1815.4730.0022.7914.271.83
SDI43.4625.776.00181.583.8931.2422.7114.833.36
HDP39.0225.926.01178.844.6231.0822.7316.2013.44
BN72.7726.666.17177.078.6731.5022.8016.887.51
BLD50.6826.366.11180.854.1931.3522.7817.221.47
Ours86.0226.606.51184.571.9131.6622.7616.740.25
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.33, + 0.907, + 0.361 + ], + "angle": 0, + "content": "Table 1. Quantitative results on MagicBrush and PIE-Bench. Metrics are grouped into Image Quality, Masked Region Preservation, and Text Alignment. \\(\\uparrow\\) indicates higher is better; \\(\\downarrow\\) indicates lower is better. The best and second-best scores are highlighted." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.386, + 0.445, + 0.403 + ], + "angle": 0, + "content": "and its relationship with \\( n \\) in supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.412, + 0.402, + 0.427 + ], + "angle": 0, + "content": "4.3.1. Ablation on Regeneration Latent Step" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.431, + 0.483, + 0.612 + ], + "angle": 0, + "content": "The timestep \\( r \\) for caching the regeneration latent is critical, as it dictates the extent of possible modifications during the regeneration process. We performed an ablation study by varying \\( r \\) while holding the total diffusion steps \\( N \\) constant. This variation in \\( r \\) implicitly changes the number of regeneration steps \\( (n) \\) and necessitates adjustments to the strength parameter accordingly. Qualitatively, as shown in Fig. 6, excessively small \\( r \\) values lead to incoherent edits and noticeable artifacts due to insufficient blending with the original image. Conversely, large \\( r \\) values limit the model's ability to modify the masked region, resulting in minimal changes and preserving the original content." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.614, + 0.483, + 0.765 + ], + "angle": 0, + "content": "Quantitatively, we observe that smaller \\( r \\) steps (e.g. \\( r = 2 \\)) yield higher LPIPS (0.04) and low PSNR (27.03), indicating poor image quality and fidelity. Edit fidelity scores such as CS-L also confirm that larger \\( r \\) steps result in lower scores (22.98), suggesting ineffective edits within the masked region. The HPS index demonstrates a higher score for mid-range steps (0.33, \\( r = 12 \\)) compared to both ends of the spectrum (0.29, \\( r = 23 \\)), highlighting a performance sweet spot for intermediate \\( r \\) values. Detailed metric graphs are available in the supplemental material." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.775, + 0.371, + 0.79 + ], + "angle": 0, + "content": "4.3.2. Ablation on Blending Latent Step" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.795, + 0.483, + 0.9 + ], + "angle": 0, + "content": "The blending latent step, controlled by the parameter \\( b \\), determines when the cached regeneration latent is blended back into the diffusion process and is crucial for seamless integration of the edited region with the original image and preserving background. We conduct an ablation study by varying \\( b \\) while keeping \\( r \\) and \\( N \\) fixed. Fig. 7 qualitatively demonstrates the effect of different \\( b \\) values." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.386, + 0.907, + 0.476 + ], + "angle": 0, + "content": "When \\( b \\) is small, the blending process starts prematurely, causing the edit to bleed into the background and distorting the original image context. Conversely, larger \\( b \\) values, representing late blending, effectively preserve the background integrity while still allowing for meaningful edits within the masked region." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.909, + 0.63 + ], + "angle": 0, + "content": "Quantitatively, smaller \\( b \\) values (\\( b = n \\)) lead to higher LPIPS (0.17) and lower PSNR (11.61), indicating worse background preservation. Edit fidelity scores (CS-L) within the masked remained stable across the spectrum while CS-D improves at larger \\( b \\) (0.32 at \\( b = N - 1 \\)), reflecting better edit alignment. These findings indicate that later blending is preferable, leading us to select \\( b = N - 2 \\) in the LDB algorithm to prioritize background preservation while maintaining effective localized editing. Further details and metric plots are available in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.646, + 0.628, + 0.661 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.906, + 0.717 + ], + "angle": 0, + "content": "Our experiments demonstrate that LDB establishes new benchmarks for speed and workflow adaptability in diffusion-based image editing. Key findings include:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.905, + 0.779 + ], + "angle": 0, + "content": "Enhanced Control via Layering: LDB's layered design enables creating non-destructive refinements as well as iterative complex compositions. Participants highlighted how this mirrors professional editing tools like Photoshop [28]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Speed and Efficiency: LDB achieves remarkable speed, \\(53 \\times\\) faster than BrushNet (evaluated on the same hardware), crucial for interactive editing. We observe that reducing diffusion steps to as few as \\(n = 4\\) maintains reasonable quality (HPS: 0.34, CS-D: 0.35), yielding a latency of 140ms per edit. User studies confirm instant feedback as a key advantage, enabling rapid iteration (tens of variations per minute vs. 1-2 for baselines). This speed results from" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17374" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.138, + 0.087, + 0.865, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.196, + 0.908, + 0.239 + ], + "angle": 0, + "content": "Figure 6. Ablation study on regeneration latent step \\( r \\) (increasing left to right). Small \\( r \\) results in strong prompt adherence (\"cat\") but introduces artifacts. Large \\( r \\) (near \\( N \\)) leads to insufficient modification, retaining the original \"dog\". An intermediate \\( r \\) achieves the best balance of edit fidelity and background preservation." + }, + { + "type": "image", + "bbox": [ + 0.169, + 0.251, + 0.262, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.176, + 0.338, + 0.245, + 0.35 + ], + "angle": 0, + "content": "a sushi plate" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.251, + 0.371, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.326, + 0.37, + 0.35 + ], + "angle": 0, + "content": "\\(b = n\\)" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.252, + 0.466, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.374, + 0.325, + 0.466, + 0.332 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.252, + 0.559, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.563, + 0.252, + 0.655, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.337, + 0.61, + 0.348 + ], + "angle": 0, + "content": "edit prompt: steak" + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.252, + 0.75, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.658, + 0.334, + 0.749, + 0.347 + ], + "angle": 0, + "content": "" + }, + { + "type": "image", + "bbox": [ + 0.753, + 0.252, + 0.844, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.335, + 0.829, + 0.347 + ], + "angle": 0, + "content": "\\(b = N - 1\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.369, + 0.908, + 0.412 + ], + "angle": 0, + "content": "Figure 7. Ablation study on blending latent step \\( b \\) (increasing left to right). The prompt \"steak\" is applied to an image of \"sushi plate\" while increasing \\( b \\) from left to right. At \\( b = n \\) (left), the edit disrupts the original structure, affecting unmasked regions. As \\( b \\) approaches \\( N \\) (right), background preservation improves, and edits blend more seamlessly." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.438, + 0.483, + 0.468 + ], + "angle": 0, + "content": "efficient latent caching (Sec. 3.1), minimizing computation and memory overhead (\\(\\sim 1.25\\) MB for 10 layers)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.473, + 0.483, + 0.685 + ], + "angle": 0, + "content": "Quantitative Performance: LDB demonstrates a superior combination of speed, image quality, and edit fidelity across both benchmarks. On the PIE-Bench dataset, LDB achieves the best performance in six key metrics, excelling in human preference (HPS = 86.02), background preservation (LPIPS = 1.91), and text alignment (CS = 31.66), while also being the fastest method by a significant margin. This highlights its ability to generalize across a diverse set of editing tasks while maintaining high speed. Similarly, on the MagicBrush benchmark, LDB delivers strong performance with highest score in crucial metrics such as IR, AS, and PSNR. While BrushNet shows a slight advantage in some text alignment metrics, its practical usability is hindered by substantially slower runtime." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.708, + 0.355, + 0.723 + ], + "angle": 0, + "content": "5.1. Limitations and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Brush strength \\((\\alpha)\\) and diffusion step count \\((n)\\) coupling (Fig. 12) still requires minor user tuning across scenarios. Although preset profiles partially address this, future work could explore adaptive parameter tuning mechanisms to further improve usability. Moreover, semantically implausible edits (e.g. placing a boat in the sky) remain challenging due to inherent biases within diffusion models. Integrating techniques like semantic guidance could expand plausible edit ranges. Finally, responsible deployment necessitates robust watermarking [15] and provenance tracking to mitigate misuse and ensure transparency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.437, + 0.715, + 0.453 + ], + "angle": 0, + "content": "5.2. Broader Applications" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.464, + 0.906, + 0.57 + ], + "angle": 0, + "content": "LDB's training-free design only requires a standard iterative denoising process, which allows seamless integration into diverse diffusion models and applications requiring rapid editing. We validated this by integrating LDB to other commonly used methods, including DiT-based text-to-image (e.g., PixArt-\\(\\alpha\\) [10]) and video generation models [7] without any model-specific tuning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.574, + 0.906, + 0.65 + ], + "angle": 0, + "content": "Traditional diffusion-based video editing typically propagates edits from the first frame using additional supervision (e.g. optical flow [36]), risking temporal inconsistencies. LDB's high fidelity background preservation and efficiency naturally address these issues." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.655, + 0.906, + 0.746 + ], + "angle": 0, + "content": "We demonstrate preliminary success integrating LDB with Stable Video Diffusion (SVD) [7], editing the first frame and applying LDB's latent caching across frames for fast consecutive edits (see supplementary material, Fig. 15). This approach opens avenues for accelerated video manipulation, 3D asset editing, and collaborative design platforms." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.768, + 0.637, + 0.783 + ], + "angle": 0, + "content": "5.3. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.901 + ], + "angle": 0, + "content": "LDB reimagines diffusion-based editing through latent caching and non-destructive layering, achieving unmatched speed and control. Quantitative results and user study show superior performance in image preference, edit fidelity, time, and usability. By bridging interactive editing with high-fidelity generative models, LDB can empower artists to iterate fluidly while maintaining artistic intent." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17375" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.192 + ], + "angle": 0, + "content": "This work was supported by the Natural Sciences and Engineering Research Council of Canada (NSERC). The authors would also like to thank Professors Leonid Signal and Kwang Moo Yi for their guidance and support throughout this project." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.204, + 0.188, + 0.219 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.228, + 0.482, + 0.296 + ], + "angle": 0, + "content": "[1] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.299, + 0.484, + 0.382 + ], + "angle": 0, + "content": "[2] Shubham Agarwal, Subrata Mitra, Sarthak Chakraborty, Srikrishna Karanam, Koyel Mukherjee, and Shiv Kumar Saini. Approximate caching for efficiently serving {Text-to-Image} diffusion models. In 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI 24), pages 1173-1189, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.383, + 0.482, + 0.41 + ], + "angle": 0, + "content": "[3] Ideogram AI. Ideogram: Text-to-image generation platform. https://ideogram.ai, 2025.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.411, + 0.483, + 0.466 + ], + "angle": 0, + "content": "[4] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.468, + 0.483, + 0.508 + ], + "angle": 0, + "content": "[5] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM Transactions on Graphics (TOG), 42 (4):1-11, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.51, + 0.483, + 0.579 + ], + "angle": 0, + "content": "[6] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 707-723. Springer, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.58, + 0.482, + 0.648 + ], + "angle": 0, + "content": "[7] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.649, + 0.482, + 0.677 + ], + "angle": 0, + "content": "[8] John Brooke. SUS-A quick and dirty usability scale. Usability evaluation in industry, 189(194), 1996. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.678, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[9] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.72, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[10] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[11] Erin Cherry and Celine Latulipe. Quantifying the creativity support of digital tools through the creativity support index. ACM Transactions on Computer-Human Interaction (TOCHI), 21(4):1-25, 2014. 5, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[12] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. Ilvr: Conditioning method for denoising diffusion probabilistic models. arXiv preprint arXiv:2108.02938, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.228, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "[13] Civitai. Dreamshaper - 7 — stable diffusion checkpoint, 2024. Accessed on Feb 19, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.121, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[14] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.246 + ], + "angle": 0, + "content": "[15] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22466-22477, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[16] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.372 + ], + "angle": 0, + "content": "[17] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.372, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[18] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Han Hu, Dong Chen, et al. Instructdiffusion: A generalist modeling interface for vision tasks. arXiv preprint arXiv:2309.03895, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[19] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, et al. Instructdiffusion: A generalist modeling interface for vision tasks. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 12709-12720, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[20] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.904, + 0.622 + ], + "angle": 0, + "content": "[21] Qin Guo and Tianwei Lin. Focus on your instruction: Fine-grained and multi-instruction image editing by attention modulation. arXiv preprint arXiv:2312.10113, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[22] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.905, + 0.762 + ], + "angle": 0, + "content": "[24] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssm. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[25] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.888 + ], + "angle": 0, + "content": "[26] Minghui Hu, Yujie Wang, Tat-Jen Cham, Jianfei Yang, and Ponnuthurai N Suganthan. Global context with discrete diffusion in vector quantised modelling for image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11502-11511, 2022." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.888 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "17376" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[27] Mingzhen Huang, Jialing Cai, Shan Jia, Vishnu Suresh Lokhande, and Siwei Lyu. Paralleledits: Efficient multi-aspect text-driven image editing with attention grouping. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.207 + ], + "angle": 0, + "content": "[28] Adobe Inc. Adobe photoshop (2024 version). https://www.adobe.com/products/photoshop.htm1, 2024. 2, 5, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[29] K. J. Joseph, Prateksha Udhayanan, Tripti Shukla, Aishwarya Agarwal, Srikrishna Karanam, Koustava Goswami, and Balaji Vasan Srinivasan. Iterative multi-granular image editing using diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 8107-8116, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[30] Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. Direct inversion: Boosting diffusion-based editing with 3 lines of code. arXiv preprint arXiv:2310.01506, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[31] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. In European Conference on Computer Vision, pages 150–168. Springer, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[32] Kumara Kahatapitiya, Haozhe Liu, Sen He, Ding Liu, Menglin Jia, Chenyang Zhang, Michael S Ryoo, and Tian Xie. Adaptive caching for faster video generation with diffusion transformers. arXiv preprint arXiv:2411.02397, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[33] Kyoungkook Kang, Gyujin Sim, Geonung Kim, Donguk Kim, Seungho Nam, and Sunghyun Cho. Layeringdiff: Layered image synthesis via generation, then disassembly with generative knowledge. arXiv preprint arXiv:2501.01197, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[34] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2426-2435, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[35] Oran Lang, Yossi Gandelsman, Michal Yarom, Yoav Wald, Gal Elidan, Avinatan Hassidim, William T Freeman, Phillip Isola, Amir Globerson, Michal Irani, et al. Explaining in style: Training a gan to explain a classifier in stylespace. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 693-702, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[36] Feng Liang, Bichen Wu, Jialiang Wang, Licheng Yu, Kunpeng Li, Yinan Zhao, Ishan Misra, Jia-Bin Huang, Peizhao Zhang, Peter Vajda, et al. Flowvid: Taming imperfect optical flows for consistent video-to-video synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8207-8216, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[37] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[38] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Deepcache: Accelerating diffusion models for free. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF conference on computer vision and pattern recognition, pages 15762-15772, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[39] Hayk Manukyan, Andranik Sargsyan, Barsegh Atanyan, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Hd-painter: High-resolution and prompt-faithful text-guided image inpainting with diffusion models. arXiv preprint arXiv:2312.14091, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[40] Qi Mao, Lan Chen, Yuchao Gu, Zhen Fang, and Mike Zheng Shou. Mag-edit: Localized image editing in complex scenarios via mask-based attention-adjusted guidance. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6842-6850, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[41] Inc. Midjourney. Midjourney: Ai image generation. https://www.midjourney.com, 2025.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[42] Daiki Miyake, Akihiro Iohara, Yu Saito, and Toshiyuki Tanaka. Negative-prompt inversion: Fast image inversion for editing with text-guided diffusion models. arXiv preprint arXiv:2305.16807, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.345, + 0.907, + 0.413 + ], + "angle": 0, + "content": "[43] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[44] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag your gan: Interactive point-based manipulation on the generative image manifold. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[45] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, New York, NY, USA, 2023. Association for Computing Machinery. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.595 + ], + "angle": 0, + "content": "[46] Thomas Porter and Tom Duff. Compositing digital images. In Proceedings of the 11th annual conference on Computer graphics and interactive techniques, pages 253-259, 1984. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.596, + 0.905, + 0.676 + ], + "angle": 0, + "content": "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[48] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[49] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2287–2296, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.907, + 0.872 + ], + "angle": 0, + "content": "[50] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[51] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17377" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "tuning text-to-image diffusion models for subject-driven generation, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[52] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[53] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.305 + ], + "angle": 0, + "content": "[54] Vishnu Sarukkai, Linden Li, Arden Ma, Christopher Ré, and Kayvon Fatahalian. Collage diffusion. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 4208-4217, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.39 + ], + "angle": 0, + "content": "[55] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in neural information processing systems, 35:25278-25294, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.482, + 0.447 + ], + "angle": 0, + "content": "[56] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[57] Felix Wimbauer, Bichen Wu, Edgar Schoenfeld, Xiaoliang Dai, Ji Hou, Zijian He, Artsiom Sanakoyeu, Peizhao Zhang, Sam Tsai, Jonas Kohler, et al. Cache me if you can: Accelerating diffusion models through block caching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6211-6220, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.601 + ], + "angle": 0, + "content": "[58] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[59] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22428-22437, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[60] Zishan Xu, Xiaofeng Zhang, Wei Chen, Minda Yao, Jueting Liu, Tingting Xu, and Zehua Wang. A review of image inpainting methods based on deep learning. Applied Sciences, 13(20):11189, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[61] Fei Yang, Shiqi Yang, Muhammad Atif Butt, Joost van de Weijer, et al. Dynamic prompt learning: Addressing cross-attention leakage for text-based image editing. Advances in Neural Information Processing Systems, 36:26291–26303, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[62] Shiyuan Yang, Xiaodong Chen, and Jing Liao. Uni-paint: A unified framework for multimodal image inpainting with pretrained diffusion model. In Proceedings of the 31st ACM International Conference on Multimedia, pages 3190-3199, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[63] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything:" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[64] Ziyun Zeng, Hang Hua, Jianlong Fu, Jiebo Luo, et al. Promptfix: You prompt and we fix the photo. Advances in Neural Information Processing Systems, 37:40000-40031, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[65] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 36, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[66] Lvmin Zhang and Maneesh Agrawala. Transparent image layer diffusion using latent transparency. arXiv preprint arXiv:2402.17113, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.333 + ], + "angle": 0, + "content": "[67] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3836-3847, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.335, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[68] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.474 + ], + "angle": 0, + "content": "[69] Junhao Zhuang, Yanhong Zeng, Wenran Liu, Chun Yuan, and Kai Chen. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. In European Conference on Computer Vision, pages 195-211. Springer, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17378" + } + ] +] \ No newline at end of file diff --git a/2025/Streamlining Image Editing with Layered Diffusion Brushes/6116442a-236c-4dff-8b3c-c87a9184cffb_origin.pdf b/2025/Streamlining Image Editing with Layered Diffusion Brushes/6116442a-236c-4dff-8b3c-c87a9184cffb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a49beb0758f0efcccbc55d874e6a777b89f69755 --- /dev/null +++ b/2025/Streamlining Image Editing with Layered Diffusion Brushes/6116442a-236c-4dff-8b3c-c87a9184cffb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d2f7090a89385923bdca2efe9d2994f3544bc3c8a7e5fa3ef50edd98ef242af0 +size 29535711 diff --git a/2025/Streamlining Image Editing with Layered Diffusion Brushes/full.md b/2025/Streamlining Image Editing with Layered Diffusion Brushes/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ea611636dd4f7161b15cf0520356342a3d6fca15 --- /dev/null +++ b/2025/Streamlining Image Editing with Layered Diffusion Brushes/full.md @@ -0,0 +1,404 @@ +# Streamlining Image Editing with Layered Diffusion Brushes + +Peyman Gholami Robert Xiao University of British Columbia + +{peymang, brx}@cs.ubc.ca + +https://layered-diffusion-brushes.github.io/ + +Edit prompt (type) + +sunglasses (add objects) + +hat, Takashi Murakami style (change style) + +starry night, van gogh style (fix problems, maintaining style + +sky, Leonid Afremov style (mixing styles) + +ornate frame (change attribute / texture) + +Layer Mask + +![](images/acaab80982df157db185097f8b40b63ad9574f638f96c248dbbfefb5407b33bd.jpg) + +![](images/77f70798e125df1a37ec37aa708ce20fd04b231e2691b8fade439235a5556149.jpg) + +![](images/f60328ab3bcc6d1830ed940bce16320b379ad75844848dc0d8e8ead7821fffe9.jpg) + +![](images/94fd2bed5093f2db8214436672bb3cd81aa40a48d03d1109eae55ba28246a858.jpg) + +![](images/7ba782e1c1776a26f3e515ef1ba6e68a809a2ff81daaebe84a6c6f132290b958.jpg) + +Edited image + +![](images/e89c61fa9b418e83a4dea980178f168cd36a02b6096d7c6d759e5709e5945996.jpg) +Layer 1 + +![](images/f2804203a71914757c150a574dced80015846053d91cc90137f4e70db1547e19.jpg) +Layer 2 +Figure 1. Hierarchical image editing with Layered Diffusion Brushes: LDB is capable of creating and stacking a wide range of independent edits, including object addition, removal, or replacement, colour and style changes/combining, and object attribute modification. Each edit is performed independently, and users are able to switch between the edits seamlessly. + +![](images/82ebbcc202c3e13a464807958caa0d049cb69d123e40903a3ab8fea09ab174b4.jpg) +Layer 3 + +![](images/a9d5c136e1562341f197982dc749cde3300a9cfe67094fc2efb66be60ee26e57.jpg) +Layer 4 + +![](images/860e7207bd9ab2e80530566d5be95c5c12f3721773ecee93a98f3ae369729cf9.jpg) +Layer 5 + +# Abstract + +Denoising diffusion models have emerged as powerful tools for image manipulation, yet interactive, localized editing workflows remain underdeveloped. We introduce Layered Diffusion Brushes (LDB), a novel training-free framework that enables interactive, layer-based editing using standard diffusion models. LDB defines each "layer" as a self-contained set of parameters guiding the generative process, enabling independent, non-destructive, and fine-grained prompt-guided edits, even in overlapping regions. LDB leverages a unique intermediate latent caching approach to reduce each edit to only a few denoising steps, achieving 140 ms per edit on consumer GPUs. An editor implementing LDB, incorporating familiar layer concepts, was evaluated via user study and quantitative metrics. Results demonstrate LDB's superior speed alongside comparable or improved image quality, background preservation, and edit fidelity relative to state-of-the-art methods across various sequential image manipulation tasks. The findings highlight LDB's ability to significantly enhance creative + +workflows by providing an intuitive and efficient approach to diffusion-based image editing and its potential for expansion into related subdomains, such as video editing. + +# 1. Introduction + +Image editing has undergone transformative advancements with the rise of text-to-image (T2I) generative models, enabling unprecedented creative expression through textual guidance. These models, including Generative Adversarial Networks (GANs) [20], Variational Autoencoders (VAEs), and Denoising Diffusion Models (DMs) [23], have redefined image synthesis and manipulation. Among these, DMs [56] have emerged as the state of the art due to their training stability, high-fidelity outputs, and versatility across tasks like inpainting [37], super-resolution [53], and style transfer [22]. However, despite their capabilities, a critical gap remains: enabling real-time, localized, and iterative edits that align with professional workflows, where artists demand precise control over specific regions without + +disrupting the global composition. + +Existing DM-based editing methods face several core challenges. First, their stochastic nature often necessitates numerous generations to achieve desired results [5]. Second, they lack intuitive mechanisms for layered, nondestructive editing—a cornerstone of tools like Adobe Photoshop [28]—where edits can be independently adjusted, stacked, or removed. Third, while mask-guided approaches enable regional control, they struggle with seamless blending, artifact-free transitions, and real-time feedback. These limitations restrict their adoption in creative pipelines, where rapid iteration and granular control are critical. + +To address these challenges, we propose Layered Diffusion Brushes (LDB), a novel framework based on Latent Diffusion Models (LDM) [50] that integrates mask-guided diffusion with a non-destructive layered editing paradigm. + +At its core, LDB introduces new noise patterns into the image latents during diffusion process, guided by both the user-specified mask and the edit prompt. This preserves the original context while seamlessly integrating localized edits. We implement an intuitive user interface (UI) with a layering system to support consecutive edits (Fig. 1). Specifically, as key contributions, LDB introduces: + +- Latent Caching for Real-Time Edits: By reusing intermediate denoising states from initial generation, edits bypass redundant computations and achieve as low as 140 ms per edit on $512 \times 512$ images ( $53 \times$ faster than BrushNet [31] using the same consumer GPUs). +- Non-destructive Layered Editing: LDB introduces an order-agnostic layering mechanism by defining the concept of layers for DMs, enabling: +- Region-targeted adjustments with background preservation, using mask-prompt pairs, +- Stacking, toggling, or deleting layers without cross-interference—even in overlapping regions, +- Post-hoc revision of edits while preserving underlying content. +- Seed-Driven Exploration: Our UI provides familiar "brush" and "scroll" gestures to enable instant exploration of variations by modulating noise seeds, bridging stochastic generation with deterministic refinement and instant feedback. + +We validate LDB through extensive experiments and a user study with graphic designers. Quantitatively, LDB outperforms state-of-the-art methods in terms of speed and image quality and is comparable in terms of edit fidelity. The user study revealed superior usability and creativity support in iterative design. Additionally, LDB is a plug-and-play, training-free system adaptable to existing models and applications, and we demonstrate this by applying LDB to the task of video editing. + +# 2. Related Work + +# 2.1. DM-based Image Editing + +Image editing is the task of modifying existing images in terms of appearance, structure, or composition, ranging from subtle adjustments to major transformations. Unlike GAN-based approaches [1, 35, 44], which are prone to limitations in inversion stability [49] and localized control [6], diffusion-based methods harness the power of controllable, high-quality DMs in various image-editing tasks, including text and image-driven image manipulation studies [14, 26, 34, 37]. + +Instruction-based text editing methods [9, 18, 19, 21, 64] typically train DMs on instruction-image pairs. For example, InstructPix2Pix [9] is trained using synthetic pairs from Stable Diffusion [50] and Prompt-to-Prompt [22]. However, expressing nuanced edits solely through text instructions remains challenging, particularly for object-specific style or color changes. + +Mask-based methods [4, 5, 14, 63] sample within specified regions. While effective for localized edits, they can introduce unintended global changes, especially problematic in sequential editing, and may struggle with complex edits. For instance, Blended Latent Diffusion's lossy VAE latent space hinders perfect reconstruction even before noise addition [5]. Though a background reconstruction strategy is included, it increases computation and may still yield incoherent results for complex edits. Conversely, our method directly modifies the original latent space, enhancing context preservation and natural blending. + +Attention-based editing manipulates cross-attention maps to guide the image generation process toward the desired modifications [22, 45]. These methods generally face challenges in achieving fine-grained edits without unwanted global modifications. Yang et al. [61] attribute unintended changes to inaccurate attention maps and propose attention focusing. Inversion-based methods like ILVR [12], Textual Inversion [16], and DreamBooth [51] focus on context modification while preserving subjects. DDIM inversion converts images to noisy latents, and sampling generates edited results based on prompts. We employ Direct Inversion [30] for efficient real image latent inversion. + +Image inpainting involves replacing or restoring the missing regions while maintaining global coherency [60]. Many inpainting works [39, 50, 62, 69] require using a fine-tuned DM specifically designed for inpainting tasks, limiting their applicability. Some, including SmartBrush, which uses object-mask prediction guidance [59], offer more flexibility. PowerPaint [69] introduces learnable task embeddings for improved control. While these models effectively generate new content, they are generally unsuitable for making small, targeted adjustments [4, 37, 52]. Inspired by ControlNet [67], BrushNet [31] builds a decomposed + +plug-and-play dual-branch DM, but struggles with real-time interaction due to its computational overhead. In Sec. 4.1 we compare LDB with several inpainting techniques. + +# 2.2. Layered and Sequential Image editing + +Layer-based image editing is fundamental in computer graphics [46], and recent works integrate this concept into AI methodologies [6, 54]. Layered representations enable dynamic manipulation of image components, transforming single images into multi-layered structures. + +LayeringDiff [33] decomposes images into foreground and background. ParallelEdits [27] uses attention for efficient multi-aspect text edits. MAG-Edit [40] employs a two-layer process with attention injection to a single edit from background. Joseph et al. [29] highlight error accumulation in sequential editing, where artifacts compound across edits. Collage Diffusion [54], built on modified Blended Latent Diffusion [5], employs alpha masks to guide cross-attention and generate harmonized images while respecting scene composition. However, it assumes prelayered inputs and synthesizes scenes from scratch. In contrast, LDB is training-free, operates directly on existing images, and supports fully independent layers—unlike methods such as [6] that require per-image training. + +# 2.3. Accelerated Generation using Caching + +Caching and reusing intermediate features has proven effective for accelerating DM inference through reducing redundant computations. Several works have utilized caching in diffusion transformers (DiTs) for video generation. DeepCache [38] reuses high-level U-Net features in video generation, while AdaCache [32] dynamically adjusts cached residuals based on temporal content. Cache Me If You Can [57] employs block caching by reusing outputs from layer blocks of previous steps during inference. For image generation, Approximate Caching [2] reuses intermediate latents created during prior image generation processes for similar prompts. We employ a similar strategy through caching key latent representations and adapt it specifically for interactive image editing, enabling the real-time feedback that is crucial for creative workflows. + +# 3. Method + +We use an LDM-based variant of image generative models and make intermediate adjustments to the latent space, similar to [5, 37]. Therefore, LDB requires no additional training or fine-tuning of the underlying LDM; all modifications are applied during the reverse diffusion process. + +We adopt the standard LDM formulation, where image generation begins with a sample from a Gaussian distribution, $Z_0 \sim \mathcal{N}(0, \sigma_{max}^2 I)$ and is iteratively denoised through a sequence of steps $N$ , resulting in a series of la + +tents $Z_{i}$ corresponding to decreasing noise levels $\sigma_{i}$ , where $\sigma_0 = \sigma_{max} > \sigma_1 > \dots >\sigma_N\approx 0$ + +As demonstrated in Fig. 2, the overall LDB pipeline comprises three key stages: initial image generation (or inversion), latent caching, and iterative layered editing. + +For DM-generated images, we first initialize the sample $Z_{0} = \epsilon_{0}$ and noise level $\sigma_0$ ( $i = 0$ ). For real images, the initial noise latent is obtained using inversion. We use Direct Inversion [30] for its high speed and comparable performance to other inversion methods, including Null-Text Inversion [43] and Negative-Prompt Inversion [42]. The noisy sample then undergoes the diffusion process, caching certain intermediate latents to facilitate editing. + +# 3.1. Latent Caching + +To enable rapid, interactive editing with instant exploration and feedback, we employ latent caching to reuse intermediate representations in subsequent steps, minimizing redundant computations. We store two key intermediate latents: + +- Regeneration Latent $\mathbf{Z}_{\mathbf{r}}$ : At diffusion step $r = N - n$ , where $N$ is the total number of diffusion steps for initial image generation and $n$ is the number of editing steps, we cache the latent $Z_{r}$ , which serves as the starting point for all subsequent edits. By reusing $Z_{r}$ , we avoid recomputing the initial denoising steps for each new edit, significantly speeding up the editing process (from $N$ denoising steps to $n$ ). Effectively, $Z_{r}$ represents a partially denoised latent state that retains the global image structure but is still malleable enough to accommodate localized edits. +- Blending Latent $Z_{\mathrm{b}}$ : We cache the latent at diffusion step $b$ which is specifically used for the layer merging process (Algorithm 1, line 7). We set $b = N - 2$ for maximum background preservation (as discussed in Sec. 4.3). $Z_{b}$ represents a more denoised latent compared to $Z_{r}$ , capturing more refined image details while still allowing for seamless blending of new edits into the existing image context. Utilizing this cached blending latent ensures smoother integration of edits and reduces visual artifacts at layer boundaries during the merging process. + +# 3.2. Layered Diffusion Brushes Editing + +To initiate an edit, the algorithm begins by generating a new noise pattern $Z_0' = \epsilon_k'$ , sampled from $\mathcal{N}(0, \sigma^2 I)$ using a different seed $S'$ , and scaling it to match the variance of the cached latent $Z_r$ . This ensures that the additive noise stays in a reasonable range from the latent for editing, preventing visual artifacts. $Z_0'$ is then added to the regeneration latent $Z_r$ , controlled by the mask $m$ and strength $\alpha$ . + +In the editing stage, at step $b$ , a new noisy sample is merged with the cached blending latent using the strength control and the mask, resulting in $Z_b'$ . Subsequently, the new latent is progressively denoised from steps $b$ through $N$ and processed through the VAE to output edited image $I'$ . + +![](images/489a9b6fe8d363f442658601222ddb107028169dae9d03f2560069fd87c9171c.jpg) +Figure 2. Overview of the Proposed Method: The top box shows standard DM-based image generation from noisy latent $Z_{0}$ and prompt $\mathcal{P}$ . The middle section depicts the latent caching module, storing and retrieving intermediate latents for different layers. The bottom box illustrates the editing process: a new noise sample $S'$ merges with the original latent at step $r$ using mask $m$ and strength control $\alpha$ . Diffusion continues until step $b$ , where modified and cached latents blend to generate the final edited image. + +Algorithm 1 presents the pseudocode for the editing process for a single layer (for simplicity): + +Algorithm 1: LDB editing process (single layer) +Input: Edit prompt $\mathcal{P}'$ , Mask $m \in [0,1]^{H \times W}$ , Random seed $S'$ , Strength $\alpha$ , Number of edit steps $n$ , Regeneration latent $Z_r$ , Blending latent $Z_b$ +Output: Edited latent $Z_N'$ +1 $Z_0' \gets \epsilon_{nk}' \sim \mathcal{N}(0, \sigma^2 I) //$ sampled using seed $S'$ +2 $Z_0' \gets \sqrt{\text{Var}(Z_r)} \cdot Z_0' //$ scale new sample +3 $Z_0' \gets Z_r + \alpha \cdot (Z_0' \odot m) //$ noise injection +4 for $i = 0,1,\ldots,n$ do +5 $Z_{i+1}' \gets DM(Z_i', \mathcal{P}', i, S')$ +6 if $i == b$ then +7 $Z_b' \gets Z_b' \odot m + Z_b \odot (1 - m) //$ blending +8 end +9 Return $Z_N'$ + +# 3.3. Layer Formulation + +Unlike prior works that rely on transparent decomposable layers [66] or explicit object segmentation [54], we redefine a layer as a self-contained set of reproducible parameters that govern localized edits. For layer $\mathcal{L}_k$ , we formalize this as a generalized version of parameters in Algorithm 1: + +$$ +\mathcal {L} ^ {(k)} = \left(\mathbf {S} ^ {\prime (k)}, \mathbf {m} ^ {(k)}, \mathbf {v} ^ {(k)}, \mathbf {Z} _ {r} ^ {(k)}, \mathbf {Z} _ {b} ^ {(k)}, \alpha^ {(k)}, n ^ {(k)}, \mathcal {P} ^ {\prime (k)}, j\right) \tag {1} +$$ + +- $\mathbf{S}'^{(k)} \in \mathbb{Z}^+$ : Seed space for stochastic variations +- $\mathbf{m}^{(k)}\in [0,1]^{H\times W}$ : Edit mask +- $\mathbf{v}^{(k)}\in \{0,1\}$ : Visibility state +- $\mathbf{Z}_r^{(k)},\mathbf{Z}_b^{(k)}\in \mathbb{R}^{C\times H\times W}$ : Regeneration/blending latents +- $\alpha^{(k)}\in [0,1]$ : Layer strength value +- $\mathbf{n}^{(k)}\in [0,N]$ Number of denoising steps +$\mathcal{P}^{\prime (k)}$ : Edit prompt +- $j \in {\mathbb{Z}}^{ + }$ : Index of last layer index. + +Notably, within a given layer $\mathcal{L}_k$ with previous layer $\mathcal{L}_j$ , the cached latents $\mathbf{Z}_r^{(j)}$ and $\mathbf{Z}_b^{(j)}$ inherently incorporate the cumulative edits from all preceding layers. This is because edits to layer $\mathcal{L}_k$ , are applied to the already edited output of layer $\mathcal{L}_j$ which serves as the input to the diffusion process and the algorithm always keeps the last layer updated. Therefore, any modification in a previous layer automatically propagates through the subsequent layers. By defining $\Phi$ as a single-layer latent generation and caching step as: + +$$ +\left(Z _ {r} ^ {(k)}, Z _ {b} ^ {(k)}\right) = \Phi \left(\mathcal {L} ^ {(k)}, \mathcal {L} ^ {(j)}\right) \tag {2} +$$ + +in essence, if a given layer $\mathcal{L}^{(i)}$ (where $i < k$ ) is removed or its visibility $\mathbf{v}^{(i)}$ is toggled, the operator $\Phi$ will be recursively invoked to recreate all latents for layers from $\mathcal{L}^{(i)}$ to $\mathcal{L}^{(k)}$ . This recomputation, accelerated by latent caching, is automatically triggered and typically completes within milliseconds to a few seconds, depending on the number of layers. This design allows edits to remain independent yet seamlessly integrated into the final composition. + +# 3.3.1. Overlapping Regions + +A key advantage of layered editing in LDB is the ability to create overlapping edits, where one layer can partially or fully modify areas affected by earlier layers. This requires careful handling of each layer's regeneration latent, $Z_{r}$ , to ensure that changes in visibility or content from higher layers are accurately reflected in subsequent layers, even in overlapping regions. + +By default, all layers use the initial image's latent $(Z_{r})$ as their regeneration latent. However, this approach fails to account for overlapping edits from preceding layers. To address this, when processing a layer $k$ , we compute its regeneration latent by inverting the output image of the previous layer $(I^{\prime (k)})$ as shown using the feedback arrow on Fig. 2. This inversion yields $Z_0^{(k)}$ , which is then sent through the generation stage in LDB. Both $Z_{r}^{(k)}$ and $Z_{b}^{(k)}$ are cached for efficient processing (as shown in Fig. 3). + +This mechanism enables precise control and seamless integration of edits across overlapping regions. Changes to any layer propagate correctly without introducing artifacts, offering flexibility and fine-grained control. + +# 3.4. User-Interface and Interaction Design + +To develop a practical tool for artists and designers, we designed an custom UI that balances control and simplicity. The UI allows users to generate, upload, and edit images, manage layers, and adjust parameters seamlessly. Two interaction modes streamline edits (Fig. 4): + +Box Mode: Users can click or drag on the image to move a resizeable square mask around. This option enables a quick and interactive exploration of how various parts of the image will change in response to a given set of editing settings (prompt and strength), simply by moving the cursor. + +Custom Mask Mode: Users can draw free-form masks over the desired around and navigate between new generation samples by scrolling the mouse up or down while hovering over the image, allowing them to rapidly explore variations on their edit. + +We propose a workflow where users first position edits spatially using Box Mode, then refine mask geometry and appearance details via Custom Mask Mode. + +Layering capabilities include stacking, visibility toggling, and deletion. Each layer is independently modifiable. Detailed information on the UI design user interactions and a demo video can be found in supplementary material. + +# 4. Experiments + +# 4.1. User Study + +We conducted a user study in order to evaluate the effectiveness of LDB for providing targeted image fine-tuning, using two other well-known existing image editing tools, + +InstructPix2Pix (IP2P) [9] and Stable Diffusion Inpainting (SDI) [50] as baselines for comparison. + +We recruited a cohort of seven expert participants with extensive experience in using image editing software. As part of our selection criteria, we ensured that all had at least a basic level of familiarity with AI image generation techniques [3, 41] and were regular users of editing software, such as Adobe Photoshop [28] for creating visual art. + +# 4.1.1. Study Procedure and Task Description + +Each user engaged in two types of tasks: free-form tasks where users generated an image for editing using a fixed prompt and seed (type 1), and pre-determined tasks where the user worked with existing real images from the MagicBrush dataset [65] (type 2). + +For the type 1 tasks, we selected specific types of edits that showcase various functionalities and capabilities of the system, including: + +1. Stack layers and create sequential edits (draw with LDB) +2. Modify attributes and features of objects +3. Correct image imperfections and errors +4. Enhance discernibility of similar objects +5. Target specific regions for style transfer, refine aesthetics Type 2 tasks were more structured, with the mask, edit prompt, and input images provided by the dataset. The dataset provides manually annotated masks and instructions for each edit. We selected a subset of 35 input images, each containing up to three layers of edits. Users refined masks/parameters if necessary and completed editing tasks. + +Figure 5, second row, shows example edits generated by the participants. Additional examples are provided in the supplementary material. As shown, LDB produces targeted edits that integrate seamlessly with the images. + +# 4.1.2. Evaluation Survey Results + +The participants completed a three stage evaluation survey following the image editing tasks. The first part included a System Usability Scale (SUS) form to rate the usability, ease of use, design, and performance of each method. SUS is a standard usability evaluation survey widely used in user-experience literature [8]. Overall, participants indicated that they are more likely to use LDB compared to IP2P and SDI, and that they find it the easiest tool to use. LDB obtained a SUS score of $80.35\%$ , while IP2P and SDI achieved a SUS of $38.21\%$ and $37.5\%$ respectively. + +The SUS survey was followed by a Creativity Support Index [11] survey to evaluate the system's degree of creative work support. Participants expressed positivity towards LDB, indicating that it enhanced their enjoyment, exploration, expressiveness, and immersion, while also deeming the results worth their effort. Lastly, the survey was followed by a semi-structured interview where participants appreciated the intuitiveness, ease of use and versatility of + +![](images/060d835aaaea3bc13692762088787ea8c8c2c26fbd33e4e373a7edc2376e50bd.jpg) +Figure 3. Overlapping edit regions in LDB: overlapping edits enable complex, interacting modifications. For example, one layer can adjust color while another changes shape, with the final result combining both. + +![](images/73023f7987fc7bf9424f1c29630395e60cfea9a1ee9cc89fe094d4aee4cb2c41.jpg) +Figure 4. Box and Custom Mask Options: In box mode, users click the target region's center to generate edits within the specified area and can drag the box to explore variations instantly. In custom mask mode, users draw a mask over the desired region and adjust the seed using the mouse wheel or scrolling gestures to generate new variations. + +![](images/dcfae83d646a828f0af4cf222fd1a45bb71b1c7192f708ddd5eaee96764e3307.jpg) +Figure 5. Qualitative editing results on PIE-Bench (top) and MagicBrush (bottom) benchmarks using different methods. Edit prompts are presented on top of each row. More examples available in supplementary material. + +LDB. Further details about the study, interview, results, and discussion can be found in supplemental material. + +# 4.2. Quantitative Analysis + +To quantitatively evaluate the performance of LDB, we employed a comprehensive suite of metrics, aligning with established practices in image editing evaluation. + +Specifically, for text-image alignment, we used CLIP Score (CS) [47] for global alignment, CS-L for masked-region alignment, and CS-D [17] for consistency between image and caption changes in CLIP space. + +We adopted Learned Perceptual Image Patch Similarity (LPIPS) [68] and Peak Signal-to-Noise Ratio (PSNR) [24] for evaluating content preservation and pixel-level fidelity in unmasked regions. Furthermore, to gauge overall image quality and aesthetic appeal, we incorporated Aesthetic Score (AS) [55], Image Reward (IR), and Human Preference Score V2 (HPS) [58], the latter two reflecting human-aligned preferences. + +We compared LDB against a diverse set of state-of + +the-art editing and inpainting methods, including InstructPix2Pix (IP2P) [9], Stable Diffusion Inpainting (SDI) [50], HD-Painter (HDP) [39], BrushNet (BN) [31], and Blended Latent Diffusion (BLD) [5], on two benchmarks: MagicBrush [65] and PIE-Bench [30]. For MagicBrush, we also report results on the provided ground truth (GT) images. + +Quantitative results are summarized in Tab. 1. All methods were evaluated using their default editing settings, except for LDB, IP2P, and SDI on the MagicBrush benchmark, where we used user-edited images from our user study for consecutive edits. Inference times denote average per-edit durations, measured on a single NVIDIA RTX 4090 GPU with $N = 25$ diffusion steps for baseline methods and $n = 8$ for LDB. + +# 4.3. Ablation Study + +We perform three ablation studies for two main components of the LDB caching mechanism, i.e. the caching timesteps for the regeneration latent $(r)$ , and the blending latent $(b)$ . We also ablate and discuss the effect of strength control $\alpha$ + +
BenchmarkMethodImage QualityMasked Region PreservationText AlignmentTime (s)
IR ×10↑HPS ×102↑AS↑PSNR↑LPIPS ×102↓CS↑CS-L↑CS-D ×102↑(per edit) ↓
MagicBrushIP2P-62.8321.165.297.2815.0729.3922.016.641.72
SDI-39.2120.885.4812.208.7030.0822.154.111.84
HDP-20.6923.275.4412.056.1331.0122.069.8912.85
BN-0.0422.575.7311.558.7531.1622.1712.927.49
BLD-24.1022.805.4812.646.9430.6421.9910.051.41
GT-1.9322.625.3617.642.3030.7522.149.78NA
Ours7.7422.655.7412.857.0531.0422.079.540.26
PIE-BenchIP2P-40.7323.125.76172.1815.4730.0022.7914.271.83
SDI43.4625.776.00181.583.8931.2422.7114.833.36
HDP39.0225.926.01178.844.6231.0822.7316.2013.44
BN72.7726.666.17177.078.6731.5022.8016.887.51
BLD50.6826.366.11180.854.1931.3522.7817.221.47
Ours86.0226.606.51184.571.9131.6622.7616.740.25
+ +Table 1. Quantitative results on MagicBrush and PIE-Bench. Metrics are grouped into Image Quality, Masked Region Preservation, and Text Alignment. $\uparrow$ indicates higher is better; $\downarrow$ indicates lower is better. The best and second-best scores are highlighted. + +and its relationship with $n$ in supplementary material. + +# 4.3.1. Ablation on Regeneration Latent Step + +The timestep $r$ for caching the regeneration latent is critical, as it dictates the extent of possible modifications during the regeneration process. We performed an ablation study by varying $r$ while holding the total diffusion steps $N$ constant. This variation in $r$ implicitly changes the number of regeneration steps $(n)$ and necessitates adjustments to the strength parameter accordingly. Qualitatively, as shown in Fig. 6, excessively small $r$ values lead to incoherent edits and noticeable artifacts due to insufficient blending with the original image. Conversely, large $r$ values limit the model's ability to modify the masked region, resulting in minimal changes and preserving the original content. + +Quantitatively, we observe that smaller $r$ steps (e.g. $r = 2$ ) yield higher LPIPS (0.04) and low PSNR (27.03), indicating poor image quality and fidelity. Edit fidelity scores such as CS-L also confirm that larger $r$ steps result in lower scores (22.98), suggesting ineffective edits within the masked region. The HPS index demonstrates a higher score for mid-range steps (0.33, $r = 12$ ) compared to both ends of the spectrum (0.29, $r = 23$ ), highlighting a performance sweet spot for intermediate $r$ values. Detailed metric graphs are available in the supplemental material. + +# 4.3.2. Ablation on Blending Latent Step + +The blending latent step, controlled by the parameter $b$ , determines when the cached regeneration latent is blended back into the diffusion process and is crucial for seamless integration of the edited region with the original image and preserving background. We conduct an ablation study by varying $b$ while keeping $r$ and $N$ fixed. Fig. 7 qualitatively demonstrates the effect of different $b$ values. + +When $b$ is small, the blending process starts prematurely, causing the edit to bleed into the background and distorting the original image context. Conversely, larger $b$ values, representing late blending, effectively preserve the background integrity while still allowing for meaningful edits within the masked region. + +Quantitatively, smaller $b$ values ( $b = n$ ) lead to higher LPIPS (0.17) and lower PSNR (11.61), indicating worse background preservation. Edit fidelity scores (CS-L) within the masked remained stable across the spectrum while CS-D improves at larger $b$ (0.32 at $b = N - 1$ ), reflecting better edit alignment. These findings indicate that later blending is preferable, leading us to select $b = N - 2$ in the LDB algorithm to prioritize background preservation while maintaining effective localized editing. Further details and metric plots are available in the supplementary material. + +# 5. Discussion + +Our experiments demonstrate that LDB establishes new benchmarks for speed and workflow adaptability in diffusion-based image editing. Key findings include: + +Enhanced Control via Layering: LDB's layered design enables creating non-destructive refinements as well as iterative complex compositions. Participants highlighted how this mirrors professional editing tools like Photoshop [28]. + +Speed and Efficiency: LDB achieves remarkable speed, $53 \times$ faster than BrushNet (evaluated on the same hardware), crucial for interactive editing. We observe that reducing diffusion steps to as few as $n = 4$ maintains reasonable quality (HPS: 0.34, CS-D: 0.35), yielding a latency of 140ms per edit. User studies confirm instant feedback as a key advantage, enabling rapid iteration (tens of variations per minute vs. 1-2 for baselines). This speed results from + +![](images/789119496e1d05996b56b8c934887f6c7de7171ece7879b66ebc2ab86caa27ad.jpg) +Figure 6. Ablation study on regeneration latent step $r$ (increasing left to right). Small $r$ results in strong prompt adherence ("cat") but introduces artifacts. Large $r$ (near $N$ ) leads to insufficient modification, retaining the original "dog". An intermediate $r$ achieves the best balance of edit fidelity and background preservation. + +![](images/688c66e8f6c106ca331aa7ab564e1530b9de38be41a3e9aad7276ef8f200bb23.jpg) + +![](images/fe9623824347b79d8b3eae85c8e346bcfcf671b7f0433662285e409b6a31c709.jpg) +a sushi plate +$b = n$ + +![](images/3a709338daa81b755e8ccf9ea69cb47bfab24041cf42358f7cf6bba6ebb8c551.jpg) +# +Figure 7. Ablation study on blending latent step $b$ (increasing left to right). The prompt "steak" is applied to an image of "sushi plate" while increasing $b$ from left to right. At $b = n$ (left), the edit disrupts the original structure, affecting unmasked regions. As $b$ approaches $N$ (right), background preservation improves, and edits blend more seamlessly. + +![](images/41509a13eaf4ae7e5924523f34d3aa009441bd40bd1e269c25688865ac0417fd.jpg) +edit prompt: steak + +![](images/7950237fa58273d2802c1df99d299bb20d0e95f5b8211e9f594dd5925a130c47.jpg) + +![](images/f59e2dd7dabb7bc3493e2fb362f36deb7799fc6835a0692d418ddb1022c0d4b4.jpg) + +![](images/8eeae6152bc0fc228e844a472e57441ab2e86833f4da74cf47384b2e91bb1875.jpg) +$b = N - 1$ + +efficient latent caching (Sec. 3.1), minimizing computation and memory overhead ( $\sim 1.25$ MB for 10 layers). + +Quantitative Performance: LDB demonstrates a superior combination of speed, image quality, and edit fidelity across both benchmarks. On the PIE-Bench dataset, LDB achieves the best performance in six key metrics, excelling in human preference (HPS = 86.02), background preservation (LPIPS = 1.91), and text alignment (CS = 31.66), while also being the fastest method by a significant margin. This highlights its ability to generalize across a diverse set of editing tasks while maintaining high speed. Similarly, on the MagicBrush benchmark, LDB delivers strong performance with highest score in crucial metrics such as IR, AS, and PSNR. While BrushNet shows a slight advantage in some text alignment metrics, its practical usability is hindered by substantially slower runtime. + +# 5.1. Limitations and Future Work + +Brush strength $(\alpha)$ and diffusion step count $(n)$ coupling (Fig. 12) still requires minor user tuning across scenarios. Although preset profiles partially address this, future work could explore adaptive parameter tuning mechanisms to further improve usability. Moreover, semantically implausible edits (e.g. placing a boat in the sky) remain challenging due to inherent biases within diffusion models. Integrating techniques like semantic guidance could expand plausible edit ranges. Finally, responsible deployment necessitates robust watermarking [15] and provenance tracking to mitigate misuse and ensure transparency. + +# 5.2. Broader Applications + +LDB's training-free design only requires a standard iterative denoising process, which allows seamless integration into diverse diffusion models and applications requiring rapid editing. We validated this by integrating LDB to other commonly used methods, including DiT-based text-to-image (e.g., PixArt- $\alpha$ [10]) and video generation models [7] without any model-specific tuning. + +Traditional diffusion-based video editing typically propagates edits from the first frame using additional supervision (e.g. optical flow [36]), risking temporal inconsistencies. LDB's high fidelity background preservation and efficiency naturally address these issues. + +We demonstrate preliminary success integrating LDB with Stable Video Diffusion (SVD) [7], editing the first frame and applying LDB's latent caching across frames for fast consecutive edits (see supplementary material, Fig. 15). This approach opens avenues for accelerated video manipulation, 3D asset editing, and collaborative design platforms. + +# 5.3. Conclusion + +LDB reimagines diffusion-based editing through latent caching and non-destructive layering, achieving unmatched speed and control. Quantitative results and user study show superior performance in image preference, edit fidelity, time, and usability. By bridging interactive editing with high-fidelity generative models, LDB can empower artists to iterate fluidly while maintaining artistic intent. + +# Acknowledgment + +This work was supported by the Natural Sciences and Engineering Research Council of Canada (NSERC). The authors would also like to thank Professors Leonid Signal and Kwang Moo Yi for their guidance and support throughout this project. + +# References + +[1] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 2 +[2] Shubham Agarwal, Subrata Mitra, Sarthak Chakraborty, Srikrishna Karanam, Koyel Mukherjee, and Shiv Kumar Saini. Approximate caching for efficiently serving {Text-to-Image} diffusion models. In 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI 24), pages 1173-1189, 2024. 3 +[3] Ideogram AI. Ideogram: Text-to-image generation platform. https://ideogram.ai, 2025.5 +[4] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. 2 +[5] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM Transactions on Graphics (TOG), 42 (4):1-11, 2023. 2, 3, 6 +[6] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 707-723. Springer, 2022. 2, 3 +[7] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 8, 2 +[8] John Brooke. SUS-A quick and dirty usability scale. Usability evaluation in industry, 189(194), 1996. 5, 6 +[9] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 2, 5, 6 +[10] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 8 +[11] Erin Cherry and Celine Latulipe. Quantifying the creativity support of digital tools through the creativity support index. ACM Transactions on Computer-Human Interaction (TOCHI), 21(4):1-25, 2014. 5, 9 +[12] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. Ilvr: Conditioning method for denoising diffusion probabilistic models. arXiv preprint arXiv:2108.02938, 2021. 2 + +[13] Civitai. Dreamshaper - 7 — stable diffusion checkpoint, 2024. Accessed on Feb 19, 2024. 4 +[14] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 2 +[15] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22466-22477, 2023. 8 +[16] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2 +[17] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 6 +[18] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Han Hu, Dong Chen, et al. Instructdiffusion: A generalist modeling interface for vision tasks. arXiv preprint arXiv:2309.03895, 2023. 2 +[19] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, et al. Instructdiffusion: A generalist modeling interface for vision tasks. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 12709-12720, 2024. 2 +[20] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1 +[21] Qin Guo and Tianwei Lin. Focus on your instruction: Fine-grained and multi-instruction image editing by attention modulation. arXiv preprint arXiv:2312.10113, 2023. 2 +[22] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1, 2 +[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1 +[24] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssm. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 6 +[25] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1 +[26] Minghui Hu, Yujie Wang, Tat-Jen Cham, Jianfei Yang, and Ponnuthurai N Suganthan. Global context with discrete diffusion in vector quantised modelling for image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11502-11511, 2022. + +[27] Mingzhen Huang, Jialing Cai, Shan Jia, Vishnu Suresh Lokhande, and Siwei Lyu. Paralleledits: Efficient multi-aspect text-driven image editing with attention grouping. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3 +[28] Adobe Inc. Adobe photoshop (2024 version). https://www.adobe.com/products/photoshop.htm1, 2024. 2, 5, 7, 9 +[29] K. J. Joseph, Prateksha Udhayanan, Tripti Shukla, Aishwarya Agarwal, Srikrishna Karanam, Koustava Goswami, and Balaji Vasan Srinivasan. Iterative multi-granular image editing using diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 8107-8116, 2024. 3 +[30] Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. Direct inversion: Boosting diffusion-based editing with 3 lines of code. arXiv preprint arXiv:2310.01506, 2023. 2, 3, 6 +[31] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. In European Conference on Computer Vision, pages 150–168. Springer, 2024. 2, 6 +[32] Kumara Kahatapitiya, Haozhe Liu, Sen He, Ding Liu, Menglin Jia, Chenyang Zhang, Michael S Ryoo, and Tian Xie. Adaptive caching for faster video generation with diffusion transformers. arXiv preprint arXiv:2411.02397, 2024. 3 +[33] Kyoungkook Kang, Gyujin Sim, Geonung Kim, Donguk Kim, Seungho Nam, and Sunghyun Cho. Layeringdiff: Layered image synthesis via generation, then disassembly with generative knowledge. arXiv preprint arXiv:2501.01197, 2025. 3 +[34] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2426-2435, 2022. 2 +[35] Oran Lang, Yossi Gandelsman, Michal Yarom, Yoav Wald, Gal Elidan, Avinatan Hassidim, William T Freeman, Phillip Isola, Amir Globerson, Michal Irani, et al. Explaining in style: Training a gan to explain a classifier in stylespace. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 693-702, 2021. 2 +[36] Feng Liang, Bichen Wu, Jialiang Wang, Licheng Yu, Kunpeng Li, Yinan Zhao, Ishan Misra, Jia-Bin Huang, Peizhao Zhang, Peter Vajda, et al. Flowvid: Taming imperfect optical flows for consistent video-to-video synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8207-8216, 2024. 8 +[37] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 1, 2, 3 +[38] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Deepcache: Accelerating diffusion models for free. In Proceedings of + +the IEEE/CVF conference on computer vision and pattern recognition, pages 15762-15772, 2024. 3 +[39] Hayk Manukyan, Andranik Sargsyan, Barsegh Atanyan, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Hd-painter: High-resolution and prompt-faithful text-guided image inpainting with diffusion models. arXiv preprint arXiv:2312.14091, 2023. 2, 6 +[40] Qi Mao, Lan Chen, Yuchao Gu, Zhen Fang, and Mike Zheng Shou. Mag-edit: Localized image editing in complex scenarios via mask-based attention-adjusted guidance. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6842-6850, 2024. 3 +[41] Inc. Midjourney. Midjourney: Ai image generation. https://www.midjourney.com, 2025.5 +[42] Daiki Miyake, Akihiro Iohara, Yu Saito, and Toshiyuki Tanaka. Negative-prompt inversion: Fast image inversion for editing with text-guided diffusion models. arXiv preprint arXiv:2305.16807, 2023. 3 +[43] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 3 +[44] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag your gan: Interactive point-based manipulation on the generative image manifold. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2 +[45] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, New York, NY, USA, 2023. Association for Computing Machinery. 2 +[46] Thomas Porter and Tom Duff. Compositing digital images. In Proceedings of the 11th annual conference on Computer graphics and interactive techniques, pages 253-259, 1984. 3 +[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 6 +[48] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 6 +[49] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2287–2296, 2021. 2 +[50] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 5, 6 +[51] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine + +tuning text-to-image diffusion models for subject-driven generation, 2023. 2 +[52] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2 +[53] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1 +[54] Vishnu Sarukkai, Linden Li, Arden Ma, Christopher Ré, and Kayvon Fatahalian. Collage diffusion. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 4208-4217, 2024. 3, 4 +[55] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in neural information processing systems, 35:25278-25294, 2022. 6 +[56] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 1 +[57] Felix Wimbauer, Bichen Wu, Edgar Schoenfeld, Xiaoliang Dai, Ji Hou, Zijian He, Artsiom Sanakoyeu, Peizhao Zhang, Sam Tsai, Jonas Kohler, et al. Cache me if you can: Accelerating diffusion models through block caching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6211-6220, 2024. 3 +[58] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 6 +[59] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22428-22437, 2023. 2 +[60] Zishan Xu, Xiaofeng Zhang, Wei Chen, Minda Yao, Jueting Liu, Tingting Xu, and Zehua Wang. A review of image inpainting methods based on deep learning. Applied Sciences, 13(20):11189, 2023. 2 +[61] Fei Yang, Shiqi Yang, Muhammad Atif Butt, Joost van de Weijer, et al. Dynamic prompt learning: Addressing cross-attention leakage for text-based image editing. Advances in Neural Information Processing Systems, 36:26291–26303, 2023. 2 +[62] Shiyuan Yang, Xiaodong Chen, and Jing Liao. Uni-paint: A unified framework for multimodal image inpainting with pretrained diffusion model. In Proceedings of the 31st ACM International Conference on Multimedia, pages 3190-3199, 2023. 2 +[63] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: + +Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. 2 +[64] Ziyun Zeng, Hang Hua, Jianlong Fu, Jiebo Luo, et al. Promptfix: You prompt and we fix the photo. Advances in Neural Information Processing Systems, 37:40000-40031, 2025. 2 +[65] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 36, 2024. 5, 6 +[66] Lvmin Zhang and Maneesh Agrawala. Transparent image layer diffusion using latent transparency. arXiv preprint arXiv:2402.17113, 2024. 4 +[67] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3836-3847, 2023. 2 +[68] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6 +[69] Junhao Zhuang, Yanhong Zeng, Wenran Liu, Chun Yuan, and Kai Chen. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. In European Conference on Computer Vision, pages 195-211. Springer, 2024. 2 \ No newline at end of file diff --git a/2025/Streamlining Image Editing with Layered Diffusion Brushes/images.zip b/2025/Streamlining Image Editing with Layered Diffusion Brushes/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..127c038a89d7a518f93f860e38345f36b1f18b2e --- /dev/null +++ b/2025/Streamlining Image Editing with Layered Diffusion Brushes/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:30c36c21f8fb3b8ffa2cf6245eb718e788d83089c1366d4a462d0b8f68fc6f91 +size 523411 diff --git a/2025/Streamlining Image Editing with Layered Diffusion Brushes/layout.json b/2025/Streamlining Image Editing with Layered Diffusion Brushes/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f9cd526c3ae4334b057bdf8ef321a3968929304c --- /dev/null +++ b/2025/Streamlining Image Editing with Layered Diffusion Brushes/layout.json @@ -0,0 +1,11939 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 119, + 97, + 492, + 114 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 97, + 492, + 114 + ], + "spans": [ + { + "bbox": [ + 119, + 97, + 492, + 114 + ], + "type": "text", + "content": "Streamlining Image Editing with Layered Diffusion Brushes" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 219, + 134, + 392, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 134, + 392, + 162 + ], + "spans": [ + { + "bbox": [ + 219, + 134, + 392, + 162 + ], + "type": "text", + "content": "Peyman Gholami Robert Xiao University of British Columbia" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 241, + 164, + 370, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 164, + 370, + 175 + ], + "spans": [ + { + "bbox": [ + 241, + 164, + 370, + 175 + ], + "type": "text", + "content": "{peymang, brx}@cs.ubc.ca" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 170, + 184, + 436, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 184, + 436, + 196 + ], + "spans": [ + { + "bbox": [ + 170, + 184, + 436, + 196 + ], + "type": "text", + "content": "https://layered-diffusion-brushes.github.io/" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 212, + 149, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 212, + 149, + 246 + ], + "spans": [ + { + "bbox": [ + 116, + 212, + 149, + 246 + ], + "type": "text", + "content": "Edit prompt (type)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 173, + 214, + 215, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 214, + 215, + 233 + ], + "spans": [ + { + "bbox": [ + 173, + 214, + 215, + 233 + ], + "type": "text", + "content": "sunglasses (add objects)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 234, + 213, + 286, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 213, + 286, + 243 + ], + "spans": [ + { + "bbox": [ + 234, + 213, + 286, + 243 + ], + "type": "text", + "content": "hat, Takashi Murakami style (change style)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 298, + 213, + 353, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 213, + 353, + 252 + ], + "spans": [ + { + "bbox": [ + 298, + 213, + 353, + 252 + ], + "type": "text", + "content": "starry night, van gogh style (fix problems, maintaining style" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 368, + 213, + 417, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 213, + 417, + 243 + ], + "spans": [ + { + "bbox": [ + 368, + 213, + 417, + 243 + ], + "type": "text", + "content": "sky, Leonid Afremov style (mixing styles)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 428, + 213, + 489, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 428, + 213, + 489, + 243 + ], + "spans": [ + { + "bbox": [ + 428, + 213, + 489, + 243 + ], + "type": "text", + "content": "ornate frame (change attribute / texture)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 259, + 143, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 259, + 143, + 280 + ], + "spans": [ + { + "bbox": [ + 116, + 259, + 143, + 280 + ], + "type": "text", + "content": "Layer Mask" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 159, + 253, + 225, + 317 + ], + "blocks": [ + { + "bbox": [ + 159, + 253, + 225, + 317 + ], + "lines": [ + { + "bbox": [ + 159, + 253, + 225, + 317 + ], + "spans": [ + { + "bbox": [ + 159, + 253, + 225, + 317 + ], + "type": "image", + "image_path": "acaab80982df157db185097f8b40b63ad9574f638f96c248dbbfefb5407b33bd.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 227, + 253, + 292, + 318 + ], + "blocks": [ + { + "bbox": [ + 227, + 253, + 292, + 318 + ], + "lines": [ + { + "bbox": [ + 227, + 253, + 292, + 318 + ], + "spans": [ + { + "bbox": [ + 227, + 253, + 292, + 318 + ], + "type": "image", + "image_path": "77f70798e125df1a37ec37aa708ce20fd04b231e2691b8fade439235a5556149.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 294, + 253, + 359, + 318 + ], + "blocks": [ + { + "bbox": [ + 294, + 253, + 359, + 318 + ], + "lines": [ + { + "bbox": [ + 294, + 253, + 359, + 318 + ], + "spans": [ + { + "bbox": [ + 294, + 253, + 359, + 318 + ], + "type": "image", + "image_path": "f60328ab3bcc6d1830ed940bce16320b379ad75844848dc0d8e8ead7821fffe9.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 361, + 253, + 425, + 318 + ], + "blocks": [ + { + "bbox": [ + 361, + 253, + 425, + 318 + ], + "lines": [ + { + "bbox": [ + 361, + 253, + 425, + 318 + ], + "spans": [ + { + "bbox": [ + 361, + 253, + 425, + 318 + ], + "type": "image", + "image_path": "94fd2bed5093f2db8214436672bb3cd81aa40a48d03d1109eae55ba28246a858.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 428, + 253, + 493, + 318 + ], + "blocks": [ + { + "bbox": [ + 428, + 253, + 493, + 318 + ], + "lines": [ + { + "bbox": [ + 428, + 253, + 493, + 318 + ], + "spans": [ + { + "bbox": [ + 428, + 253, + 493, + 318 + ], + "type": "image", + "image_path": "7ba782e1c1776a26f3e515ef1ba6e68a809a2ff81daaebe84a6c6f132290b958.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 337, + 145, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 337, + 145, + 359 + ], + "spans": [ + { + "bbox": [ + 116, + 337, + 145, + 359 + ], + "type": "text", + "content": "Edited image" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 159, + 319, + 225, + 384 + ], + "blocks": [ + { + "bbox": [ + 159, + 319, + 225, + 384 + ], + "lines": [ + { + "bbox": [ + 159, + 319, + 225, + 384 + ], + "spans": [ + { + "bbox": [ + 159, + 319, + 225, + 384 + ], + "type": "image", + "image_path": "e89c61fa9b418e83a4dea980178f168cd36a02b6096d7c6d759e5709e5945996.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 177, + 393, + 207, + 403 + ], + "lines": [ + { + "bbox": [ + 177, + 393, + 207, + 403 + ], + "spans": [ + { + "bbox": [ + 177, + 393, + 207, + 403 + ], + "type": "text", + "content": "Layer 1" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 227, + 319, + 292, + 384 + ], + "blocks": [ + { + "bbox": [ + 227, + 319, + 292, + 384 + ], + "lines": [ + { + "bbox": [ + 227, + 319, + 292, + 384 + ], + "spans": [ + { + "bbox": [ + 227, + 319, + 292, + 384 + ], + "type": "image", + "image_path": "f2804203a71914757c150a574dced80015846053d91cc90137f4e70db1547e19.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 393, + 274, + 403 + ], + "lines": [ + { + "bbox": [ + 244, + 393, + 274, + 403 + ], + "spans": [ + { + "bbox": [ + 244, + 393, + 274, + 403 + ], + "type": "text", + "content": "Layer 2" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 413, + 555, + 448 + ], + "lines": [ + { + "bbox": [ + 54, + 413, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 54, + 413, + 555, + 448 + ], + "type": "text", + "content": "Figure 1. Hierarchical image editing with Layered Diffusion Brushes: LDB is capable of creating and stacking a wide range of independent edits, including object addition, removal, or replacement, colour and style changes/combining, and object attribute modification. Each edit is performed independently, and users are able to switch between the edits seamlessly." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 294, + 319, + 359, + 383 + ], + "blocks": [ + { + "bbox": [ + 294, + 319, + 359, + 383 + ], + "lines": [ + { + "bbox": [ + 294, + 319, + 359, + 383 + ], + "spans": [ + { + "bbox": [ + 294, + 319, + 359, + 383 + ], + "type": "image", + "image_path": "82ebbcc202c3e13a464807958caa0d049cb69d123e40903a3ab8fea09ab174b4.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 393, + 341, + 403 + ], + "lines": [ + { + "bbox": [ + 311, + 393, + 341, + 403 + ], + "spans": [ + { + "bbox": [ + 311, + 393, + 341, + 403 + ], + "type": "text", + "content": "Layer 3" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 361, + 319, + 425, + 383 + ], + "blocks": [ + { + "bbox": [ + 361, + 319, + 425, + 383 + ], + "lines": [ + { + "bbox": [ + 361, + 319, + 425, + 383 + ], + "spans": [ + { + "bbox": [ + 361, + 319, + 425, + 383 + ], + "type": "image", + "image_path": "a9d5c136e1562341f197982dc749cde3300a9cfe67094fc2efb66be60ee26e57.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 393, + 408, + 403 + ], + "lines": [ + { + "bbox": [ + 378, + 393, + 408, + 403 + ], + "spans": [ + { + "bbox": [ + 378, + 393, + 408, + 403 + ], + "type": "text", + "content": "Layer 4" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 427, + 319, + 493, + 383 + ], + "blocks": [ + { + "bbox": [ + 427, + 319, + 493, + 383 + ], + "lines": [ + { + "bbox": [ + 427, + 319, + 493, + 383 + ], + "spans": [ + { + "bbox": [ + 427, + 319, + 493, + 383 + ], + "type": "image", + "image_path": "860e7207bd9ab2e80530566d5be95c5c12f3721773ecee93a98f3ae369729cf9.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 445, + 393, + 474, + 403 + ], + "lines": [ + { + "bbox": [ + 445, + 393, + 474, + 403 + ], + "spans": [ + { + "bbox": [ + 445, + 393, + 474, + 403 + ], + "type": "text", + "content": "Layer 5" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "bbox": [ + 151, + 464, + 200, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 464, + 200, + 477 + ], + "spans": [ + { + "bbox": [ + 151, + 464, + 200, + 477 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 54, + 486, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 486, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 486, + 297, + 714 + ], + "type": "text", + "content": "Denoising diffusion models have emerged as powerful tools for image manipulation, yet interactive, localized editing workflows remain underdeveloped. We introduce Layered Diffusion Brushes (LDB), a novel training-free framework that enables interactive, layer-based editing using standard diffusion models. LDB defines each \"layer\" as a self-contained set of parameters guiding the generative process, enabling independent, non-destructive, and fine-grained prompt-guided edits, even in overlapping regions. LDB leverages a unique intermediate latent caching approach to reduce each edit to only a few denoising steps, achieving 140 ms per edit on consumer GPUs. An editor implementing LDB, incorporating familiar layer concepts, was evaluated via user study and quantitative metrics. Results demonstrate LDB's superior speed alongside comparable or improved image quality, background preservation, and edit fidelity relative to state-of-the-art methods across various sequential image manipulation tasks. The findings highlight LDB's ability to significantly enhance creative" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 466, + 555, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 555, + 502 + ], + "type": "text", + "content": "workflows by providing an intuitive and efficient approach to diffusion-based image editing and its potential for expansion into related subdomains, such as video editing." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 314, + 525, + 395, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 525, + 395, + 537 + ], + "spans": [ + { + "bbox": [ + 314, + 525, + 395, + 537 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 312, + 545, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 545, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 545, + 556, + 715 + ], + "type": "text", + "content": "Image editing has undergone transformative advancements with the rise of text-to-image (T2I) generative models, enabling unprecedented creative expression through textual guidance. These models, including Generative Adversarial Networks (GANs) [20], Variational Autoencoders (VAEs), and Denoising Diffusion Models (DMs) [23], have redefined image synthesis and manipulation. Among these, DMs [56] have emerged as the state of the art due to their training stability, high-fidelity outputs, and versatility across tasks like inpainting [37], super-resolution [53], and style transfer [22]. However, despite their capabilities, a critical gap remains: enabling real-time, localized, and iterative edits that align with professional workflows, where artists demand precise control over specific regions without" + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17368" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 195, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 195, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 195, + 85 + ], + "type": "text", + "content": "disrupting the global composition." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 224 + ], + "type": "text", + "content": "Existing DM-based editing methods face several core challenges. First, their stochastic nature often necessitates numerous generations to achieve desired results [5]. Second, they lack intuitive mechanisms for layered, nondestructive editing—a cornerstone of tools like Adobe Photoshop [28]—where edits can be independently adjusted, stacked, or removed. Third, while mask-guided approaches enable regional control, they struggle with seamless blending, artifact-free transitions, and real-time feedback. These limitations restrict their adoption in creative pipelines, where rapid iteration and granular control are critical." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 228, + 296, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 228, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 54, + 228, + 296, + 277 + ], + "type": "text", + "content": "To address these challenges, we propose Layered Diffusion Brushes (LDB), a novel framework based on Latent Diffusion Models (LDM) [50] that integrates mask-guided diffusion with a non-destructive layered editing paradigm." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 282, + 296, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 282, + 296, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 296, + 368 + ], + "type": "text", + "content": "At its core, LDB introduces new noise patterns into the image latents during diffusion process, guided by both the user-specified mask and the edit prompt. This preserves the original context while seamlessly integrating localized edits. We implement an intuitive user interface (UI) with a layering system to support consecutive edits (Fig. 1). Specifically, as key contributions, LDB introduces:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 372, + 295, + 598 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 55, + 372, + 295, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 295, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 295, + 432 + ], + "type": "text", + "content": "- Latent Caching for Real-Time Edits: By reusing intermediate denoising states from initial generation, edits bypass redundant computations and achieve as low as 140 ms per edit on " + }, + { + "bbox": [ + 55, + 372, + 295, + 432 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 372, + 295, + 432 + ], + "type": "text", + "content": " images (" + }, + { + "bbox": [ + 55, + 372, + 295, + 432 + ], + "type": "inline_equation", + "content": "53 \\times" + }, + { + "bbox": [ + 55, + 372, + 295, + 432 + ], + "type": "text", + "content": " faster than BrushNet [31] using the same consumer GPUs)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 433, + 295, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 295, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 295, + 467 + ], + "type": "text", + "content": "- Non-destructive Layered Editing: LDB introduces an order-agnostic layering mechanism by defining the concept of layers for DMs, enabling:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 468, + 294, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 468, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 63, + 468, + 294, + 491 + ], + "type": "text", + "content": "- Region-targeted adjustments with background preservation, using mask-prompt pairs," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 63, + 492, + 294, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 492, + 294, + 515 + ], + "spans": [ + { + "bbox": [ + 63, + 492, + 294, + 515 + ], + "type": "text", + "content": "- Stacking, toggling, or deleting layers without cross-interference—even in overlapping regions," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 63, + 516, + 294, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 516, + 294, + 538 + ], + "spans": [ + { + "bbox": [ + 63, + 516, + 294, + 538 + ], + "type": "text", + "content": "- Post-hoc revision of edits while preserving underlying content." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 540, + 295, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 295, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 295, + 598 + ], + "type": "text", + "content": "- Seed-Driven Exploration: Our UI provides familiar \"brush\" and \"scroll\" gestures to enable instant exploration of variations by modulating noise seeds, bridging stochastic generation with deterministic refinement and instant feedback." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": "We validate LDB through extensive experiments and a user study with graphic designers. Quantitatively, LDB outperforms state-of-the-art methods in terms of speed and image quality and is comparable in terms of edit fidelity. The user study revealed superior usability and creativity support in iterative design. Additionally, LDB is a plug-and-play, training-free system adaptable to existing models and applications, and we demonstrate this by applying LDB to the task of video editing." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 91, + 455, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 91, + 455, + 105 + ], + "spans": [ + { + "bbox": [ + 314, + 91, + 455, + 105 + ], + "type": "text", + "content": "2.1. DM-based Image Editing" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 110, + 555, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 555, + 218 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 555, + 218 + ], + "type": "text", + "content": "Image editing is the task of modifying existing images in terms of appearance, structure, or composition, ranging from subtle adjustments to major transformations. Unlike GAN-based approaches [1, 35, 44], which are prone to limitations in inversion stability [49] and localized control [6], diffusion-based methods harness the power of controllable, high-quality DMs in various image-editing tasks, including text and image-driven image manipulation studies [14, 26, 34, 37]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 219, + 554, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 219, + 554, + 303 + ], + "spans": [ + { + "bbox": [ + 313, + 219, + 554, + 303 + ], + "type": "text", + "content": "Instruction-based text editing methods [9, 18, 19, 21, 64] typically train DMs on instruction-image pairs. For example, InstructPix2Pix [9] is trained using synthetic pairs from Stable Diffusion [50] and Prompt-to-Prompt [22]. However, expressing nuanced edits solely through text instructions remains challenging, particularly for object-specific style or color changes." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 304, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 554, + 437 + ], + "type": "text", + "content": "Mask-based methods [4, 5, 14, 63] sample within specified regions. While effective for localized edits, they can introduce unintended global changes, especially problematic in sequential editing, and may struggle with complex edits. For instance, Blended Latent Diffusion's lossy VAE latent space hinders perfect reconstruction even before noise addition [5]. Though a background reconstruction strategy is included, it increases computation and may still yield incoherent results for complex edits. Conversely, our method directly modifies the original latent space, enhancing context preservation and natural blending." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 437, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 581 + ], + "type": "text", + "content": "Attention-based editing manipulates cross-attention maps to guide the image generation process toward the desired modifications [22, 45]. These methods generally face challenges in achieving fine-grained edits without unwanted global modifications. Yang et al. [61] attribute unintended changes to inaccurate attention maps and propose attention focusing. Inversion-based methods like ILVR [12], Textual Inversion [16], and DreamBooth [51] focus on context modification while preserving subjects. DDIM inversion converts images to noisy latents, and sampling generates edited results based on prompts. We employ Direct Inversion [30] for efficient real image latent inversion." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "Image inpainting involves replacing or restoring the missing regions while maintaining global coherency [60]. Many inpainting works [39, 50, 62, 69] require using a fine-tuned DM specifically designed for inpainting tasks, limiting their applicability. Some, including SmartBrush, which uses object-mask prediction guidance [59], offer more flexibility. PowerPaint [69] introduces learnable task embeddings for improved control. While these models effectively generate new content, they are generally unsuitable for making small, targeted adjustments [4, 37, 52]. Inspired by ControlNet [67], BrushNet [31] builds a decomposed" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17369" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 109 + ], + "type": "text", + "content": "plug-and-play dual-branch DM, but struggles with real-time interaction due to its computational overhead. In Sec. 4.1 we compare LDB with several inpainting techniques." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 118, + 257, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 118, + 257, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 257, + 131 + ], + "type": "text", + "content": "2.2. Layered and Sequential Image editing" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 136, + 295, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 136, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 55, + 136, + 295, + 196 + ], + "type": "text", + "content": "Layer-based image editing is fundamental in computer graphics [46], and recent works integrate this concept into AI methodologies [6, 54]. Layered representations enable dynamic manipulation of image components, transforming single images into multi-layered structures." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 197, + 296, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 197, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 197, + 296, + 365 + ], + "type": "text", + "content": "LayeringDiff [33] decomposes images into foreground and background. ParallelEdits [27] uses attention for efficient multi-aspect text edits. MAG-Edit [40] employs a two-layer process with attention injection to a single edit from background. Joseph et al. [29] highlight error accumulation in sequential editing, where artifacts compound across edits. Collage Diffusion [54], built on modified Blended Latent Diffusion [5], employs alpha masks to guide cross-attention and generate harmonized images while respecting scene composition. However, it assumes prelayered inputs and synthesizes scenes from scratch. In contrast, LDB is training-free, operates directly on existing images, and supports fully independent layers—unlike methods such as [6] that require per-image training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 373, + 259, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 259, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 259, + 387 + ], + "type": "text", + "content": "2.3. Accelerated Generation using Caching" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 392, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 296, + 571 + ], + "type": "text", + "content": "Caching and reusing intermediate features has proven effective for accelerating DM inference through reducing redundant computations. Several works have utilized caching in diffusion transformers (DiTs) for video generation. DeepCache [38] reuses high-level U-Net features in video generation, while AdaCache [32] dynamically adjusts cached residuals based on temporal content. Cache Me If You Can [57] employs block caching by reusing outputs from layer blocks of previous steps during inference. For image generation, Approximate Caching [2] reuses intermediate latents created during prior image generation processes for similar prompts. We employ a similar strategy through caching key latent representations and adapt it specifically for interactive image editing, enabling the real-time feedback that is crucial for creative workflows." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 584, + 111, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 584, + 111, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 584, + 111, + 597 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "text", + "content": "We use an LDM-based variant of image generative models and make intermediate adjustments to the latent space, similar to [5, 37]. Therefore, LDB requires no additional training or fine-tuning of the underlying LDM; all modifications are applied during the reverse diffusion process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "We adopt the standard LDM formulation, where image generation begins with a sample from a Gaussian distribution, " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "Z_0 \\sim \\mathcal{N}(0, \\sigma_{max}^2 I)" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " and is iteratively denoised through a sequence of steps " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", resulting in a series of la" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "tents " + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "inline_equation", + "content": "Z_{i}" + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": " corresponding to decreasing noise levels " + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "inline_equation", + "content": "\\sigma_0 = \\sigma_{max} > \\sigma_1 > \\dots >\\sigma_N\\approx 0" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 96, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 553, + 133 + ], + "type": "text", + "content": "As demonstrated in Fig. 2, the overall LDB pipeline comprises three key stages: initial image generation (or inversion), latent caching, and iterative layered editing." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 133, + 555, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 229 + ], + "type": "text", + "content": "For DM-generated images, we first initialize the sample " + }, + { + "bbox": [ + 313, + 133, + 555, + 229 + ], + "type": "inline_equation", + "content": "Z_{0} = \\epsilon_{0}" + }, + { + "bbox": [ + 313, + 133, + 555, + 229 + ], + "type": "text", + "content": " and noise level " + }, + { + "bbox": [ + 313, + 133, + 555, + 229 + ], + "type": "inline_equation", + "content": "\\sigma_0" + }, + { + "bbox": [ + 313, + 133, + 555, + 229 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 133, + 555, + 229 + ], + "type": "inline_equation", + "content": "i = 0" + }, + { + "bbox": [ + 313, + 133, + 555, + 229 + ], + "type": "text", + "content": "). For real images, the initial noise latent is obtained using inversion. We use Direct Inversion [30] for its high speed and comparable performance to other inversion methods, including Null-Text Inversion [43] and Negative-Prompt Inversion [42]. The noisy sample then undergoes the diffusion process, caching certain intermediate latents to facilitate editing." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 238, + 410, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 238, + 410, + 251 + ], + "spans": [ + { + "bbox": [ + 314, + 238, + 410, + 251 + ], + "type": "text", + "content": "3.1. Latent Caching" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 255, + 553, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 553, + 303 + ], + "type": "text", + "content": "To enable rapid, interactive editing with instant exploration and feedback, we employ latent caching to reuse intermediate representations in subsequent steps, minimizing redundant computations. We store two key intermediate latents:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 304, + 555, + 544 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 314, + 304, + 554, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 304, + 554, + 422 + ], + "spans": [ + { + "bbox": [ + 314, + 304, + 554, + 422 + ], + "type": "text", + "content": "- Regeneration Latent " + }, + { + "bbox": [ + 314, + 304, + 554, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{\\mathbf{r}}" + }, + { + "bbox": [ + 314, + 304, + 554, + 422 + ], + "type": "text", + "content": ": At diffusion step " + }, + { + "bbox": [ + 314, + 304, + 554, + 422 + ], + "type": "inline_equation", + "content": "r = N - n" + }, + { + "bbox": [ + 314, + 304, + 554, + 422 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 314, + 304, + 554, + 422 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 314, + 304, + 554, + 422 + ], + "type": "text", + "content": " is the total number of diffusion steps for initial image generation and " + }, + { + "bbox": [ + 314, + 304, + 554, + 422 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 314, + 304, + 554, + 422 + ], + "type": "text", + "content": " is the number of editing steps, we cache the latent " + }, + { + "bbox": [ + 314, + 304, + 554, + 422 + ], + "type": "inline_equation", + "content": "Z_{r}" + }, + { + "bbox": [ + 314, + 304, + 554, + 422 + ], + "type": "text", + "content": ", which serves as the starting point for all subsequent edits. By reusing " + }, + { + "bbox": [ + 314, + 304, + 554, + 422 + ], + "type": "inline_equation", + "content": "Z_{r}" + }, + { + "bbox": [ + 314, + 304, + 554, + 422 + ], + "type": "text", + "content": ", we avoid recomputing the initial denoising steps for each new edit, significantly speeding up the editing process (from " + }, + { + "bbox": [ + 314, + 304, + 554, + 422 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 314, + 304, + 554, + 422 + ], + "type": "text", + "content": " denoising steps to " + }, + { + "bbox": [ + 314, + 304, + 554, + 422 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 314, + 304, + 554, + 422 + ], + "type": "text", + "content": "). Effectively, " + }, + { + "bbox": [ + 314, + 304, + 554, + 422 + ], + "type": "inline_equation", + "content": "Z_{r}" + }, + { + "bbox": [ + 314, + 304, + 554, + 422 + ], + "type": "text", + "content": " represents a partially denoised latent state that retains the global image structure but is still malleable enough to accommodate localized edits." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 423, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 423, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 315, + 423, + 555, + 544 + ], + "type": "text", + "content": "- Blending Latent " + }, + { + "bbox": [ + 315, + 423, + 555, + 544 + ], + "type": "inline_equation", + "content": "Z_{\\mathrm{b}}" + }, + { + "bbox": [ + 315, + 423, + 555, + 544 + ], + "type": "text", + "content": ": We cache the latent at diffusion step " + }, + { + "bbox": [ + 315, + 423, + 555, + 544 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 315, + 423, + 555, + 544 + ], + "type": "text", + "content": " which is specifically used for the layer merging process (Algorithm 1, line 7). We set " + }, + { + "bbox": [ + 315, + 423, + 555, + 544 + ], + "type": "inline_equation", + "content": "b = N - 2" + }, + { + "bbox": [ + 315, + 423, + 555, + 544 + ], + "type": "text", + "content": " for maximum background preservation (as discussed in Sec. 4.3). " + }, + { + "bbox": [ + 315, + 423, + 555, + 544 + ], + "type": "inline_equation", + "content": "Z_{b}" + }, + { + "bbox": [ + 315, + 423, + 555, + 544 + ], + "type": "text", + "content": " represents a more denoised latent compared to " + }, + { + "bbox": [ + 315, + 423, + 555, + 544 + ], + "type": "inline_equation", + "content": "Z_{r}" + }, + { + "bbox": [ + 315, + 423, + 555, + 544 + ], + "type": "text", + "content": ", capturing more refined image details while still allowing for seamless blending of new edits into the existing image context. Utilizing this cached blending latent ensures smoother integration of edits and reduces visual artifacts at layer boundaries during the merging process." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 552, + 500, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 552, + 500, + 565 + ], + "spans": [ + { + "bbox": [ + 314, + 552, + 500, + 565 + ], + "type": "text", + "content": "3.2. Layered Diffusion Brushes Editing" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 570, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 653 + ], + "type": "text", + "content": "To initiate an edit, the algorithm begins by generating a new noise pattern " + }, + { + "bbox": [ + 313, + 570, + 554, + 653 + ], + "type": "inline_equation", + "content": "Z_0' = \\epsilon_k'" + }, + { + "bbox": [ + 313, + 570, + 554, + 653 + ], + "type": "text", + "content": ", sampled from " + }, + { + "bbox": [ + 313, + 570, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0, \\sigma^2 I)" + }, + { + "bbox": [ + 313, + 570, + 554, + 653 + ], + "type": "text", + "content": " using a different seed " + }, + { + "bbox": [ + 313, + 570, + 554, + 653 + ], + "type": "inline_equation", + "content": "S'" + }, + { + "bbox": [ + 313, + 570, + 554, + 653 + ], + "type": "text", + "content": ", and scaling it to match the variance of the cached latent " + }, + { + "bbox": [ + 313, + 570, + 554, + 653 + ], + "type": "inline_equation", + "content": "Z_r" + }, + { + "bbox": [ + 313, + 570, + 554, + 653 + ], + "type": "text", + "content": ". This ensures that the additive noise stays in a reasonable range from the latent for editing, preventing visual artifacts. " + }, + { + "bbox": [ + 313, + 570, + 554, + 653 + ], + "type": "inline_equation", + "content": "Z_0'" + }, + { + "bbox": [ + 313, + 570, + 554, + 653 + ], + "type": "text", + "content": " is then added to the regeneration latent " + }, + { + "bbox": [ + 313, + 570, + 554, + 653 + ], + "type": "inline_equation", + "content": "Z_r" + }, + { + "bbox": [ + 313, + 570, + 554, + 653 + ], + "type": "text", + "content": ", controlled by the mask " + }, + { + "bbox": [ + 313, + 570, + 554, + 653 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 570, + 554, + 653 + ], + "type": "text", + "content": " and strength " + }, + { + "bbox": [ + 313, + 570, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 570, + 554, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "In the editing stage, at step " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": ", a new noisy sample is merged with the cached blending latent using the strength control and the mask, resulting in " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "Z_b'" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": ". Subsequently, the new latent is progressively denoised from steps " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": " and processed through the VAE to output edited image " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17370" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 70, + 504, + 296 + ], + "blocks": [ + { + "bbox": [ + 109, + 70, + 504, + 296 + ], + "lines": [ + { + "bbox": [ + 109, + 70, + 504, + 296 + ], + "spans": [ + { + "bbox": [ + 109, + 70, + 504, + 296 + ], + "type": "image", + "image_path": "489a9b6fe8d363f442658601222ddb107028169dae9d03f2560069fd87c9171c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 304, + 555, + 350 + ], + "lines": [ + { + "bbox": [ + 55, + 304, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 555, + 350 + ], + "type": "text", + "content": "Figure 2. Overview of the Proposed Method: The top box shows standard DM-based image generation from noisy latent " + }, + { + "bbox": [ + 55, + 304, + 555, + 350 + ], + "type": "inline_equation", + "content": "Z_{0}" + }, + { + "bbox": [ + 55, + 304, + 555, + 350 + ], + "type": "text", + "content": " and prompt " + }, + { + "bbox": [ + 55, + 304, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 304, + 555, + 350 + ], + "type": "text", + "content": ". The middle section depicts the latent caching module, storing and retrieving intermediate latents for different layers. The bottom box illustrates the editing process: a new noise sample " + }, + { + "bbox": [ + 55, + 304, + 555, + 350 + ], + "type": "inline_equation", + "content": "S'" + }, + { + "bbox": [ + 55, + 304, + 555, + 350 + ], + "type": "text", + "content": " merges with the original latent at step " + }, + { + "bbox": [ + 55, + 304, + 555, + 350 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 304, + 555, + 350 + ], + "type": "text", + "content": " using mask " + }, + { + "bbox": [ + 55, + 304, + 555, + 350 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 304, + 555, + 350 + ], + "type": "text", + "content": " and strength control " + }, + { + "bbox": [ + 55, + 304, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 304, + 555, + 350 + ], + "type": "text", + "content": ". Diffusion continues until step " + }, + { + "bbox": [ + 55, + 304, + 555, + 350 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 304, + 555, + 350 + ], + "type": "text", + "content": ", where modified and cached latents blend to generate the final edited image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 369, + 295, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 295, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 295, + 395 + ], + "type": "text", + "content": "Algorithm 1 presents the pseudocode for the editing process for a single layer (for simplicity):" + } + ] + } + ], + "index": 2 + }, + { + "type": "code", + "bbox": [ + 58, + 419, + 296, + 594 + ], + "blocks": [ + { + "bbox": [ + 63, + 406, + 260, + 418 + ], + "lines": [ + { + "bbox": [ + 63, + 406, + 260, + 418 + ], + "spans": [ + { + "bbox": [ + 63, + 406, + 260, + 418 + ], + "type": "text", + "content": "Algorithm 1: LDB editing process (single layer)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "lines": [ + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "type": "text", + "content": "Input: Edit prompt " + }, + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{P}'" + }, + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "type": "text", + "content": ", Mask " + }, + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "type": "inline_equation", + "content": "m \\in [0,1]^{H \\times W}" + }, + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "type": "text", + "content": ", Random seed " + }, + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "type": "inline_equation", + "content": "S'" + }, + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "type": "text", + "content": ", Strength " + }, + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "type": "text", + "content": ", Number of edit steps " + }, + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "type": "text", + "content": ", Regeneration latent " + }, + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "type": "inline_equation", + "content": "Z_r" + }, + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "type": "text", + "content": ", Blending latent " + }, + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "type": "inline_equation", + "content": "Z_b" + }, + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "type": "text", + "content": " \nOutput: Edited latent " + }, + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "type": "inline_equation", + "content": "Z_N'" + }, + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "type": "text", + "content": " \n1 " + }, + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "type": "inline_equation", + "content": "Z_0' \\gets \\epsilon_{nk}' \\sim \\mathcal{N}(0, \\sigma^2 I) //" + }, + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "type": "text", + "content": " sampled using seed " + }, + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "type": "inline_equation", + "content": "S'" + }, + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "type": "text", + "content": " \n2 " + }, + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "type": "inline_equation", + "content": "Z_0' \\gets \\sqrt{\\text{Var}(Z_r)} \\cdot Z_0' //" + }, + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "type": "text", + "content": " scale new sample \n3 " + }, + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "type": "inline_equation", + "content": "Z_0' \\gets Z_r + \\alpha \\cdot (Z_0' \\odot m) //" + }, + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "type": "text", + "content": " noise injection \n4 for " + }, + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "type": "inline_equation", + "content": "i = 0,1,\\ldots,n" + }, + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "type": "text", + "content": " do \n5 " + }, + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "type": "inline_equation", + "content": "Z_{i+1}' \\gets DM(Z_i', \\mathcal{P}', i, S')" + }, + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "type": "text", + "content": " \n6 if " + }, + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "type": "inline_equation", + "content": "i == b" + }, + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "type": "text", + "content": " then \n7 " + }, + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "type": "inline_equation", + "content": "Z_b' \\gets Z_b' \\odot m + Z_b \\odot (1 - m) //" + }, + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "type": "text", + "content": " blending \n8 end \n9 Return " + }, + { + "bbox": [ + 58, + 419, + 296, + 594 + ], + "type": "inline_equation", + "content": "Z_N'" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "code_body" + } + ], + "index": 4, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 615, + 168, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 615, + 168, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 615, + 168, + 628 + ], + "type": "text", + "content": "3.3. Layer Formulation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 633, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 295, + 693 + ], + "type": "text", + "content": "Unlike prior works that rely on transparent decomposable layers [66] or explicit object segmentation [54], we redefine a layer as a self-contained set of reproducible parameters that govern localized edits. For layer " + }, + { + "bbox": [ + 55, + 633, + 295, + 693 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_k" + }, + { + "bbox": [ + 55, + 633, + 295, + 693 + ], + "type": "text", + "content": ", we formalize this as a generalized version of parameters in Algorithm 1:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 697, + 295, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 697, + 295, + 717 + ], + "spans": [ + { + "bbox": [ + 61, + 697, + 295, + 717 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {(k)} = \\left(\\mathbf {S} ^ {\\prime (k)}, \\mathbf {m} ^ {(k)}, \\mathbf {v} ^ {(k)}, \\mathbf {Z} _ {r} ^ {(k)}, \\mathbf {Z} _ {b} ^ {(k)}, \\alpha^ {(k)}, n ^ {(k)}, \\mathcal {P} ^ {\\prime (k)}, j\\right) \\tag {1}", + "image_path": "daf370e5b4d7876a6ee7993cfffe7d73e6d2499be5dcaae6691c720609afa6e2.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 367, + 545, + 468 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 314, + 367, + 518, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 367, + 518, + 380 + ], + "spans": [ + { + "bbox": [ + 314, + 367, + 518, + 380 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 314, + 367, + 518, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{S}'^{(k)} \\in \\mathbb{Z}^+" + }, + { + "bbox": [ + 314, + 367, + 518, + 380 + ], + "type": "text", + "content": ": Seed space for stochastic variations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 380, + 447, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 380, + 447, + 392 + ], + "spans": [ + { + "bbox": [ + 315, + 380, + 447, + 392 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 315, + 380, + 447, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^{(k)}\\in [0,1]^{H\\times W}" + }, + { + "bbox": [ + 315, + 380, + 447, + 392 + ], + "type": "text", + "content": " : Edit mask" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 392, + 443, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 392, + 443, + 405 + ], + "spans": [ + { + "bbox": [ + 315, + 392, + 443, + 405 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 315, + 392, + 443, + 405 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^{(k)}\\in \\{0,1\\}" + }, + { + "bbox": [ + 315, + 392, + 443, + 405 + ], + "type": "text", + "content": ": Visibility state" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 405, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 405, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 315, + 405, + 545, + 418 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 315, + 405, + 545, + 418 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_r^{(k)},\\mathbf{Z}_b^{(k)}\\in \\mathbb{R}^{C\\times H\\times W}" + }, + { + "bbox": [ + 315, + 405, + 545, + 418 + ], + "type": "text", + "content": " : Regeneration/blending latents" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 418, + 463, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 418, + 463, + 430 + ], + "spans": [ + { + "bbox": [ + 315, + 418, + 463, + 430 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 315, + 418, + 463, + 430 + ], + "type": "inline_equation", + "content": "\\alpha^{(k)}\\in [0,1]" + }, + { + "bbox": [ + 315, + 418, + 463, + 430 + ], + "type": "text", + "content": " : Layer strength value" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 430, + 488, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 430, + 488, + 443 + ], + "spans": [ + { + "bbox": [ + 315, + 430, + 488, + 443 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 315, + 430, + 488, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{n}^{(k)}\\in [0,N]" + }, + { + "bbox": [ + 315, + 430, + 488, + 443 + ], + "type": "text", + "content": " Number of denoising steps" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 443, + 402, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 443, + 402, + 456 + ], + "spans": [ + { + "bbox": [ + 315, + 443, + 402, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{\\prime (k)}" + }, + { + "bbox": [ + 315, + 443, + 402, + 456 + ], + "type": "text", + "content": " : Edit prompt" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 456, + 460, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 456, + 460, + 468 + ], + "spans": [ + { + "bbox": [ + 315, + 456, + 460, + 468 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 315, + 456, + 460, + 468 + ], + "type": "inline_equation", + "content": "j \\in {\\mathbb{Z}}^{ + }" + }, + { + "bbox": [ + 315, + 456, + 460, + 468 + ], + "type": "text", + "content": " : Index of last layer index." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 472, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 555, + 582 + ], + "type": "text", + "content": "Notably, within a given layer " + }, + { + "bbox": [ + 313, + 472, + 555, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_k" + }, + { + "bbox": [ + 313, + 472, + 555, + 582 + ], + "type": "text", + "content": " with previous layer " + }, + { + "bbox": [ + 313, + 472, + 555, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_j" + }, + { + "bbox": [ + 313, + 472, + 555, + 582 + ], + "type": "text", + "content": ", the cached latents " + }, + { + "bbox": [ + 313, + 472, + 555, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_r^{(j)}" + }, + { + "bbox": [ + 313, + 472, + 555, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 472, + 555, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_b^{(j)}" + }, + { + "bbox": [ + 313, + 472, + 555, + 582 + ], + "type": "text", + "content": " inherently incorporate the cumulative edits from all preceding layers. This is because edits to layer " + }, + { + "bbox": [ + 313, + 472, + 555, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_k" + }, + { + "bbox": [ + 313, + 472, + 555, + 582 + ], + "type": "text", + "content": ", are applied to the already edited output of layer " + }, + { + "bbox": [ + 313, + 472, + 555, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_j" + }, + { + "bbox": [ + 313, + 472, + 555, + 582 + ], + "type": "text", + "content": " which serves as the input to the diffusion process and the algorithm always keeps the last layer updated. Therefore, any modification in a previous layer automatically propagates through the subsequent layers. By defining " + }, + { + "bbox": [ + 313, + 472, + 555, + 582 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 313, + 472, + 555, + 582 + ], + "type": "text", + "content": " as a single-layer latent generation and caching step as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 374, + 591, + 555, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 591, + 555, + 608 + ], + "spans": [ + { + "bbox": [ + 374, + 591, + 555, + 608 + ], + "type": "interline_equation", + "content": "\\left(Z _ {r} ^ {(k)}, Z _ {b} ^ {(k)}\\right) = \\Phi \\left(\\mathcal {L} ^ {(k)}, \\mathcal {L} ^ {(j)}\\right) \\tag {2}", + "image_path": "84ef112934d8b1ae4f511927ae717baf532619c9b4bc6fdf4289d2edcf174c15.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "in essence, if a given layer " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{(i)}" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "i < k" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": ") is removed or its visibility " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^{(i)}" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": " is toggled, the operator " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": " will be recursively invoked to recreate all latents for layers from " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{(i)}" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{(k)}" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": ". This recomputation, accelerated by latent caching, is automatically triggered and typically completes within milliseconds to a few seconds, depending on the number of layers. This design allows edits to remain independent yet seamlessly integrated into the final composition." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "17371" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 174, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 174, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 174, + 85 + ], + "type": "text", + "content": "3.3.1. Overlapping Regions" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 87, + 296, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 87, + 296, + 171 + ], + "spans": [ + { + "bbox": [ + 55, + 87, + 296, + 171 + ], + "type": "text", + "content": "A key advantage of layered editing in LDB is the ability to create overlapping edits, where one layer can partially or fully modify areas affected by earlier layers. This requires careful handling of each layer's regeneration latent, " + }, + { + "bbox": [ + 55, + 87, + 296, + 171 + ], + "type": "inline_equation", + "content": "Z_{r}" + }, + { + "bbox": [ + 55, + 87, + 296, + 171 + ], + "type": "text", + "content": ", to ensure that changes in visibility or content from higher layers are accurately reflected in subsequent layers, even in overlapping regions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 172, + 296, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 172, + 296, + 284 + ], + "spans": [ + { + "bbox": [ + 55, + 172, + 296, + 284 + ], + "type": "text", + "content": "By default, all layers use the initial image's latent " + }, + { + "bbox": [ + 55, + 172, + 296, + 284 + ], + "type": "inline_equation", + "content": "(Z_{r})" + }, + { + "bbox": [ + 55, + 172, + 296, + 284 + ], + "type": "text", + "content": " as their regeneration latent. However, this approach fails to account for overlapping edits from preceding layers. To address this, when processing a layer " + }, + { + "bbox": [ + 55, + 172, + 296, + 284 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 172, + 296, + 284 + ], + "type": "text", + "content": ", we compute its regeneration latent by inverting the output image of the previous layer " + }, + { + "bbox": [ + 55, + 172, + 296, + 284 + ], + "type": "inline_equation", + "content": "(I^{\\prime (k)})" + }, + { + "bbox": [ + 55, + 172, + 296, + 284 + ], + "type": "text", + "content": " as shown using the feedback arrow on Fig. 2. This inversion yields " + }, + { + "bbox": [ + 55, + 172, + 296, + 284 + ], + "type": "inline_equation", + "content": "Z_0^{(k)}" + }, + { + "bbox": [ + 55, + 172, + 296, + 284 + ], + "type": "text", + "content": ", which is then sent through the generation stage in LDB. Both " + }, + { + "bbox": [ + 55, + 172, + 296, + 284 + ], + "type": "inline_equation", + "content": "Z_{r}^{(k)}" + }, + { + "bbox": [ + 55, + 172, + 296, + 284 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 172, + 296, + 284 + ], + "type": "inline_equation", + "content": "Z_{b}^{(k)}" + }, + { + "bbox": [ + 55, + 172, + 296, + 284 + ], + "type": "text", + "content": " are cached for efficient processing (as shown in Fig. 3)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 285, + 295, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 295, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 295, + 333 + ], + "type": "text", + "content": "This mechanism enables precise control and seamless integration of edits across overlapping regions. Changes to any layer propagate correctly without introducing artifacts, offering flexibility and fine-grained control." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 342, + 256, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 256, + 355 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 256, + 355 + ], + "type": "text", + "content": "3.4. User-Interface and Interaction Design" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 360, + 296, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 296, + 420 + ], + "type": "text", + "content": "To develop a practical tool for artists and designers, we designed an custom UI that balances control and simplicity. The UI allows users to generate, upload, and edit images, manage layers, and adjust parameters seamlessly. Two interaction modes streamline edits (Fig. 4):" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 421, + 296, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 296, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 296, + 480 + ], + "type": "text", + "content": "Box Mode: Users can click or drag on the image to move a resizeable square mask around. This option enables a quick and interactive exploration of how various parts of the image will change in response to a given set of editing settings (prompt and strength), simply by moving the cursor." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 481, + 295, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 481, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 295, + 540 + ], + "type": "text", + "content": "Custom Mask Mode: Users can draw free-form masks over the desired around and navigate between new generation samples by scrolling the mouse up or down while hovering over the image, allowing them to rapidly explore variations on their edit." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 542, + 295, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 295, + 577 + ], + "type": "text", + "content": "We propose a workflow where users first position edits spatially using Box Mode, then refine mask geometry and appearance details via Custom Mask Mode." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 578, + 295, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 295, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 295, + 626 + ], + "type": "text", + "content": "Layering capabilities include stacking, visibility toggling, and deletion. Each layer is independently modifiable. Detailed information on the UI design user interactions and a demo video can be found in supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 639, + 135, + 652 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 135, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 135, + 652 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 659, + 130, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 130, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 130, + 672 + ], + "type": "text", + "content": "4.1. User Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "We conducted a user study in order to evaluate the effectiveness of LDB for providing targeted image fine-tuning, using two other well-known existing image editing tools," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "InstructPix2Pix (IP2P) [9] and Stable Diffusion Inpainting (SDI) [50] as baselines for comparison." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 97, + 553, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 97, + 553, + 169 + ], + "spans": [ + { + "bbox": [ + 313, + 97, + 553, + 169 + ], + "type": "text", + "content": "We recruited a cohort of seven expert participants with extensive experience in using image editing software. As part of our selection criteria, we ensured that all had at least a basic level of familiarity with AI image generation techniques [3, 41] and were regular users of editing software, such as Adobe Photoshop [28] for creating visual art." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 178, + 507, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 507, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 507, + 190 + ], + "type": "text", + "content": "4.1.1. Study Procedure and Task Description" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 194, + 553, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 553, + 253 + ], + "type": "text", + "content": "Each user engaged in two types of tasks: free-form tasks where users generated an image for editing using a fixed prompt and seed (type 1), and pre-determined tasks where the user worked with existing real images from the MagicBrush dataset [65] (type 2)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 255, + 553, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 553, + 290 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 553, + 290 + ], + "type": "text", + "content": "For the type 1 tasks, we selected specific types of edits that showcase various functionalities and capabilities of the system, including:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 291, + 553, + 423 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 315, + 291, + 553, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 291, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 315, + 291, + 553, + 303 + ], + "type": "text", + "content": "1. Stack layers and create sequential edits (draw with LDB)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 304, + 492, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 304, + 492, + 315 + ], + "spans": [ + { + "bbox": [ + 315, + 304, + 492, + 315 + ], + "type": "text", + "content": "2. Modify attributes and features of objects" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 316, + 487, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 316, + 487, + 327 + ], + "spans": [ + { + "bbox": [ + 315, + 316, + 487, + 327 + ], + "type": "text", + "content": "3. Correct image imperfections and errors" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 327, + 491, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 327, + 491, + 338 + ], + "spans": [ + { + "bbox": [ + 315, + 327, + 491, + 338 + ], + "type": "text", + "content": "4. Enhance discernibility of similar objects" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 339, + 553, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 553, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 553, + 423 + ], + "type": "text", + "content": "5. Target specific regions for style transfer, refine aesthetics Type 2 tasks were more structured, with the mask, edit prompt, and input images provided by the dataset. The dataset provides manually annotated masks and instructions for each edit. We selected a subset of 35 input images, each containing up to three layers of edits. Users refined masks/parameters if necessary and completed editing tasks." + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 425, + 553, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 553, + 472 + ], + "type": "text", + "content": "Figure 5, second row, shows example edits generated by the participants. Additional examples are provided in the supplementary material. As shown, LDB produces targeted edits that integrate seamlessly with the images." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 481, + 455, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 481, + 455, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 481, + 455, + 493 + ], + "type": "text", + "content": "4.1.2. Evaluation Survey Results" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 497, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 554, + 617 + ], + "type": "text", + "content": "The participants completed a three stage evaluation survey following the image editing tasks. The first part included a System Usability Scale (SUS) form to rate the usability, ease of use, design, and performance of each method. SUS is a standard usability evaluation survey widely used in user-experience literature [8]. Overall, participants indicated that they are more likely to use LDB compared to IP2P and SDI, and that they find it the easiest tool to use. LDB obtained a SUS score of " + }, + { + "bbox": [ + 313, + 497, + 554, + 617 + ], + "type": "inline_equation", + "content": "80.35\\%" + }, + { + "bbox": [ + 313, + 497, + 554, + 617 + ], + "type": "text", + "content": ", while IP2P and SDI achieved a SUS of " + }, + { + "bbox": [ + 313, + 497, + 554, + 617 + ], + "type": "inline_equation", + "content": "38.21\\%" + }, + { + "bbox": [ + 313, + 497, + 554, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 497, + 554, + 617 + ], + "type": "inline_equation", + "content": "37.5\\%" + }, + { + "bbox": [ + 313, + 497, + 554, + 617 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "text", + "content": "The SUS survey was followed by a Creativity Support Index [11] survey to evaluate the system's degree of creative work support. Participants expressed positivity towards LDB, indicating that it enhanced their enjoyment, exploration, expressiveness, and immersion, while also deeming the results worth their effort. Lastly, the survey was followed by a semi-structured interview where participants appreciated the intuitiveness, ease of use and versatility of" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17372" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 366, + 232 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 366, + 232 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 366, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 366, + 232 + ], + "type": "image", + "image_path": "060d835aaaea3bc13692762088787ea8c8c2c26fbd33e4e373a7edc2376e50bd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 239, + 367, + 274 + ], + "lines": [ + { + "bbox": [ + 55, + 239, + 367, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 367, + 274 + ], + "type": "text", + "content": "Figure 3. Overlapping edit regions in LDB: overlapping edits enable complex, interacting modifications. For example, one layer can adjust color while another changes shape, with the final result combining both." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 402, + 72, + 554, + 163 + ], + "blocks": [ + { + "bbox": [ + 402, + 72, + 554, + 163 + ], + "lines": [ + { + "bbox": [ + 402, + 72, + 554, + 163 + ], + "spans": [ + { + "bbox": [ + 402, + 72, + 554, + 163 + ], + "type": "image", + "image_path": "73023f7987fc7bf9424f1c29630395e60cfea9a1ee9cc89fe094d4aee4cb2c41.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 401, + 171, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 401, + 171, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 401, + 171, + 555, + 270 + ], + "type": "text", + "content": "Figure 4. Box and Custom Mask Options: In box mode, users click the target region's center to generate edits within the specified area and can drag the box to explore variations instantly. In custom mask mode, users draw a mask over the desired region and adjust the seed using the mouse wheel or scrolling gestures to generate new variations." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 69, + 283, + 544, + 423 + ], + "blocks": [ + { + "bbox": [ + 69, + 283, + 544, + 423 + ], + "lines": [ + { + "bbox": [ + 69, + 283, + 544, + 423 + ], + "spans": [ + { + "bbox": [ + 69, + 283, + 544, + 423 + ], + "type": "image", + "image_path": "dcfae83d646a828f0af4cf222fd1a45bb71b1c7192f708ddd5eaee96764e3307.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 431, + 555, + 455 + ], + "lines": [ + { + "bbox": [ + 54, + 431, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 54, + 431, + 555, + 455 + ], + "type": "text", + "content": "Figure 5. Qualitative editing results on PIE-Bench (top) and MagicBrush (bottom) benchmarks using different methods. Edit prompts are presented on top of each row. More examples available in supplementary material." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 475, + 295, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 295, + 499 + ], + "type": "text", + "content": "LDB. Further details about the study, interview, results, and discussion can be found in supplemental material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 505, + 180, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 180, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 180, + 517 + ], + "type": "text", + "content": "4.2. Quantitative Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 522, + 295, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 557 + ], + "type": "text", + "content": "To quantitatively evaluate the performance of LDB, we employed a comprehensive suite of metrics, aligning with established practices in image editing evaluation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 558, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 605 + ], + "type": "text", + "content": "Specifically, for text-image alignment, we used CLIP Score (CS) [47] for global alignment, CS-L for masked-region alignment, and CS-D [17] for consistency between image and caption changes in CLIP space." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "text", + "content": "We adopted Learned Perceptual Image Patch Similarity (LPIPS) [68] and Peak Signal-to-Noise Ratio (PSNR) [24] for evaluating content preservation and pixel-level fidelity in unmasked regions. Furthermore, to gauge overall image quality and aesthetic appeal, we incorporated Aesthetic Score (AS) [55], Image Reward (IR), and Human Preference Score V2 (HPS) [58], the latter two reflecting human-aligned preferences." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "We compared LDB against a diverse set of state-of" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 475, + 555, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 555, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 555, + 546 + ], + "type": "text", + "content": "the-art editing and inpainting methods, including InstructPix2Pix (IP2P) [9], Stable Diffusion Inpainting (SDI) [50], HD-Painter (HDP) [39], BrushNet (BN) [31], and Blended Latent Diffusion (BLD) [5], on two benchmarks: MagicBrush [65] and PIE-Bench [30]. For MagicBrush, we also report results on the provided ground truth (GT) images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 547, + 556, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 547, + 556, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 547, + 556, + 641 + ], + "type": "text", + "content": "Quantitative results are summarized in Tab. 1. All methods were evaluated using their default editing settings, except for LDB, IP2P, and SDI on the MagicBrush benchmark, where we used user-edited images from our user study for consecutive edits. Inference times denote average per-edit durations, measured on a single NVIDIA RTX 4090 GPU with " + }, + { + "bbox": [ + 313, + 547, + 556, + 641 + ], + "type": "inline_equation", + "content": "N = 25" + }, + { + "bbox": [ + 313, + 547, + 556, + 641 + ], + "type": "text", + "content": " diffusion steps for baseline methods and " + }, + { + "bbox": [ + 313, + 547, + 556, + 641 + ], + "type": "inline_equation", + "content": "n = 8" + }, + { + "bbox": [ + 313, + 547, + 556, + 641 + ], + "type": "text", + "content": " for LDB." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 648, + 408, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 408, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 408, + 661 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "content": "We perform three ablation studies for two main components of the LDB caching mechanism, i.e. the caching timesteps for the regeneration latent " + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "inline_equation", + "content": "(r)" + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "content": ", and the blending latent " + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "inline_equation", + "content": "(b)" + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "content": ". We also ablate and discuss the effect of strength control " + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17373" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 70, + 526, + 259 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 526, + 259 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 526, + 259 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 526, + 259 + ], + "type": "table", + "html": "
BenchmarkMethodImage QualityMasked Region PreservationText AlignmentTime (s)
IR ×10↑HPS ×102↑AS↑PSNR↑LPIPS ×102↓CS↑CS-L↑CS-D ×102↑(per edit) ↓
MagicBrushIP2P-62.8321.165.297.2815.0729.3922.016.641.72
SDI-39.2120.885.4812.208.7030.0822.154.111.84
HDP-20.6923.275.4412.056.1331.0122.069.8912.85
BN-0.0422.575.7311.558.7531.1622.1712.927.49
BLD-24.1022.805.4812.646.9430.6421.9910.051.41
GT-1.9322.625.3617.642.3030.7522.149.78NA
Ours7.7422.655.7412.857.0531.0422.079.540.26
PIE-BenchIP2P-40.7323.125.76172.1815.4730.0022.7914.271.83
SDI43.4625.776.00181.583.8931.2422.7114.833.36
HDP39.0225.926.01178.844.6231.0822.7316.2013.44
BN72.7726.666.17177.078.6731.5022.8016.887.51
BLD50.6826.366.11180.854.1931.3522.7817.221.47
Ours86.0226.606.51184.571.9131.6622.7616.740.25
", + "image_path": "caffedebdb7b4568dec8eb535457d3cceeab7dc4171d1c1f89408faf1fba04c9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 261, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 261, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 261, + 555, + 285 + ], + "type": "text", + "content": "Table 1. Quantitative results on MagicBrush and PIE-Bench. Metrics are grouped into Image Quality, Masked Region Preservation, and Text Alignment. " + }, + { + "bbox": [ + 55, + 261, + 555, + 285 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 55, + 261, + 555, + 285 + ], + "type": "text", + "content": " indicates higher is better; " + }, + { + "bbox": [ + 55, + 261, + 555, + 285 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 55, + 261, + 555, + 285 + ], + "type": "text", + "content": " indicates lower is better. The best and second-best scores are highlighted." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 305, + 272, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 305, + 272, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 305, + 272, + 319 + ], + "type": "text", + "content": "and its relationship with " + }, + { + "bbox": [ + 55, + 305, + 272, + 319 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 305, + 272, + 319 + ], + "type": "text", + "content": " in supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 326, + 246, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 246, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 246, + 338 + ], + "type": "text", + "content": "4.3.1. Ablation on Regeneration Latent Step" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 341, + 295, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 341, + 295, + 484 + ], + "spans": [ + { + "bbox": [ + 54, + 341, + 295, + 484 + ], + "type": "text", + "content": "The timestep " + }, + { + "bbox": [ + 54, + 341, + 295, + 484 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 54, + 341, + 295, + 484 + ], + "type": "text", + "content": " for caching the regeneration latent is critical, as it dictates the extent of possible modifications during the regeneration process. We performed an ablation study by varying " + }, + { + "bbox": [ + 54, + 341, + 295, + 484 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 54, + 341, + 295, + 484 + ], + "type": "text", + "content": " while holding the total diffusion steps " + }, + { + "bbox": [ + 54, + 341, + 295, + 484 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 341, + 295, + 484 + ], + "type": "text", + "content": " constant. This variation in " + }, + { + "bbox": [ + 54, + 341, + 295, + 484 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 54, + 341, + 295, + 484 + ], + "type": "text", + "content": " implicitly changes the number of regeneration steps " + }, + { + "bbox": [ + 54, + 341, + 295, + 484 + ], + "type": "inline_equation", + "content": "(n)" + }, + { + "bbox": [ + 54, + 341, + 295, + 484 + ], + "type": "text", + "content": " and necessitates adjustments to the strength parameter accordingly. Qualitatively, as shown in Fig. 6, excessively small " + }, + { + "bbox": [ + 54, + 341, + 295, + 484 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 54, + 341, + 295, + 484 + ], + "type": "text", + "content": " values lead to incoherent edits and noticeable artifacts due to insufficient blending with the original image. Conversely, large " + }, + { + "bbox": [ + 54, + 341, + 295, + 484 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 54, + 341, + 295, + 484 + ], + "type": "text", + "content": " values limit the model's ability to modify the masked region, resulting in minimal changes and preserving the original content." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 486, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 486, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 54, + 486, + 295, + 605 + ], + "type": "text", + "content": "Quantitatively, we observe that smaller " + }, + { + "bbox": [ + 54, + 486, + 295, + 605 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 54, + 486, + 295, + 605 + ], + "type": "text", + "content": " steps (e.g. " + }, + { + "bbox": [ + 54, + 486, + 295, + 605 + ], + "type": "inline_equation", + "content": "r = 2" + }, + { + "bbox": [ + 54, + 486, + 295, + 605 + ], + "type": "text", + "content": ") yield higher LPIPS (0.04) and low PSNR (27.03), indicating poor image quality and fidelity. Edit fidelity scores such as CS-L also confirm that larger " + }, + { + "bbox": [ + 54, + 486, + 295, + 605 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 54, + 486, + 295, + 605 + ], + "type": "text", + "content": " steps result in lower scores (22.98), suggesting ineffective edits within the masked region. The HPS index demonstrates a higher score for mid-range steps (0.33, " + }, + { + "bbox": [ + 54, + 486, + 295, + 605 + ], + "type": "inline_equation", + "content": "r = 12" + }, + { + "bbox": [ + 54, + 486, + 295, + 605 + ], + "type": "text", + "content": ") compared to both ends of the spectrum (0.29, " + }, + { + "bbox": [ + 54, + 486, + 295, + 605 + ], + "type": "inline_equation", + "content": "r = 23" + }, + { + "bbox": [ + 54, + 486, + 295, + 605 + ], + "type": "text", + "content": "), highlighting a performance sweet spot for intermediate " + }, + { + "bbox": [ + 54, + 486, + 295, + 605 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 54, + 486, + 295, + 605 + ], + "type": "text", + "content": " values. Detailed metric graphs are available in the supplemental material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 613, + 227, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 227, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 227, + 625 + ], + "type": "text", + "content": "4.3.2. Ablation on Blending Latent Step" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 629, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 629, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 54, + 629, + 295, + 712 + ], + "type": "text", + "content": "The blending latent step, controlled by the parameter " + }, + { + "bbox": [ + 54, + 629, + 295, + 712 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 54, + 629, + 295, + 712 + ], + "type": "text", + "content": ", determines when the cached regeneration latent is blended back into the diffusion process and is crucial for seamless integration of the edited region with the original image and preserving background. We conduct an ablation study by varying " + }, + { + "bbox": [ + 54, + 629, + 295, + 712 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 54, + 629, + 295, + 712 + ], + "type": "text", + "content": " while keeping " + }, + { + "bbox": [ + 54, + 629, + 295, + 712 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 54, + 629, + 295, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 629, + 295, + 712 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 629, + 295, + 712 + ], + "type": "text", + "content": " fixed. Fig. 7 qualitatively demonstrates the effect of different " + }, + { + "bbox": [ + 54, + 629, + 295, + 712 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 54, + 629, + 295, + 712 + ], + "type": "text", + "content": " values." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 305, + 555, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 305, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 555, + 376 + ], + "type": "text", + "content": "When " + }, + { + "bbox": [ + 313, + 305, + 555, + 376 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 305, + 555, + 376 + ], + "type": "text", + "content": " is small, the blending process starts prematurely, causing the edit to bleed into the background and distorting the original image context. Conversely, larger " + }, + { + "bbox": [ + 313, + 305, + 555, + 376 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 305, + 555, + 376 + ], + "type": "text", + "content": " values, representing late blending, effectively preserve the background integrity while still allowing for meaningful edits within the masked region." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 378, + 556, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 556, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 556, + 498 + ], + "type": "text", + "content": "Quantitatively, smaller " + }, + { + "bbox": [ + 313, + 378, + 556, + 498 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 378, + 556, + 498 + ], + "type": "text", + "content": " values (" + }, + { + "bbox": [ + 313, + 378, + 556, + 498 + ], + "type": "inline_equation", + "content": "b = n" + }, + { + "bbox": [ + 313, + 378, + 556, + 498 + ], + "type": "text", + "content": ") lead to higher LPIPS (0.17) and lower PSNR (11.61), indicating worse background preservation. Edit fidelity scores (CS-L) within the masked remained stable across the spectrum while CS-D improves at larger " + }, + { + "bbox": [ + 313, + 378, + 556, + 498 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 378, + 556, + 498 + ], + "type": "text", + "content": " (0.32 at " + }, + { + "bbox": [ + 313, + 378, + 556, + 498 + ], + "type": "inline_equation", + "content": "b = N - 1" + }, + { + "bbox": [ + 313, + 378, + 556, + 498 + ], + "type": "text", + "content": "), reflecting better edit alignment. These findings indicate that later blending is preferable, leading us to select " + }, + { + "bbox": [ + 313, + 378, + 556, + 498 + ], + "type": "inline_equation", + "content": "b = N - 2" + }, + { + "bbox": [ + 313, + 378, + 556, + 498 + ], + "type": "text", + "content": " in the LDB algorithm to prioritize background preservation while maintaining effective localized editing. Further details and metric plots are available in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 511, + 384, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 511, + 384, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 384, + 523 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 532, + 554, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 554, + 567 + ], + "type": "text", + "content": "Our experiments demonstrate that LDB establishes new benchmarks for speed and workflow adaptability in diffusion-based image editing. Key findings include:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 569, + 553, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 553, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 553, + 616 + ], + "type": "text", + "content": "Enhanced Control via Layering: LDB's layered design enables creating non-destructive refinements as well as iterative complex compositions. Participants highlighted how this mirrors professional editing tools like Photoshop [28]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": "Speed and Efficiency: LDB achieves remarkable speed, " + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "inline_equation", + "content": "53 \\times" + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": " faster than BrushNet (evaluated on the same hardware), crucial for interactive editing. We observe that reducing diffusion steps to as few as " + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "inline_equation", + "content": "n = 4" + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": " maintains reasonable quality (HPS: 0.34, CS-D: 0.35), yielding a latency of 140ms per edit. User studies confirm instant feedback as a key advantage, enabling rapid iteration (tens of variations per minute vs. 1-2 for baselines). This speed results from" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17374" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 84, + 68, + 529, + 148 + ], + "blocks": [ + { + "bbox": [ + 84, + 68, + 529, + 148 + ], + "lines": [ + { + "bbox": [ + 84, + 68, + 529, + 148 + ], + "spans": [ + { + "bbox": [ + 84, + 68, + 529, + 148 + ], + "type": "image", + "image_path": "789119496e1d05996b56b8c934887f6c7de7171ece7879b66ebc2ab86caa27ad.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 155, + 555, + 189 + ], + "lines": [ + { + "bbox": [ + 54, + 155, + 555, + 189 + ], + "spans": [ + { + "bbox": [ + 54, + 155, + 555, + 189 + ], + "type": "text", + "content": "Figure 6. Ablation study on regeneration latent step " + }, + { + "bbox": [ + 54, + 155, + 555, + 189 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 54, + 155, + 555, + 189 + ], + "type": "text", + "content": " (increasing left to right). Small " + }, + { + "bbox": [ + 54, + 155, + 555, + 189 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 54, + 155, + 555, + 189 + ], + "type": "text", + "content": " results in strong prompt adherence (\"cat\") but introduces artifacts. Large " + }, + { + "bbox": [ + 54, + 155, + 555, + 189 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 54, + 155, + 555, + 189 + ], + "type": "text", + "content": " (near " + }, + { + "bbox": [ + 54, + 155, + 555, + 189 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 155, + 555, + 189 + ], + "type": "text", + "content": ") leads to insufficient modification, retaining the original \"dog\". An intermediate " + }, + { + "bbox": [ + 54, + 155, + 555, + 189 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 54, + 155, + 555, + 189 + ], + "type": "text", + "content": " achieves the best balance of edit fidelity and background preservation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 103, + 198, + 160, + 256 + ], + "blocks": [ + { + "bbox": [ + 103, + 198, + 160, + 256 + ], + "lines": [ + { + "bbox": [ + 103, + 198, + 160, + 256 + ], + "spans": [ + { + "bbox": [ + 103, + 198, + 160, + 256 + ], + "type": "image", + "image_path": "688c66e8f6c106ca331aa7ab564e1530b9de38be41a3e9aad7276ef8f200bb23.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 168, + 198, + 227, + 255 + ], + "blocks": [ + { + "bbox": [ + 107, + 267, + 149, + 277 + ], + "lines": [ + { + "bbox": [ + 107, + 267, + 149, + 277 + ], + "spans": [ + { + "bbox": [ + 107, + 267, + 149, + 277 + ], + "type": "text", + "content": "a sushi plate" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 168, + 198, + 227, + 255 + ], + "lines": [ + { + "bbox": [ + 168, + 198, + 227, + 255 + ], + "spans": [ + { + "bbox": [ + 168, + 198, + 227, + 255 + ], + "type": "image", + "image_path": "fe9623824347b79d8b3eae85c8e346bcfcf671b7f0433662285e409b6a31c709.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 169, + 258, + 226, + 277 + ], + "lines": [ + { + "bbox": [ + 169, + 258, + 226, + 277 + ], + "spans": [ + { + "bbox": [ + 169, + 258, + 226, + 277 + ], + "type": "inline_equation", + "content": "b = n" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 228, + 199, + 285, + 255 + ], + "blocks": [ + { + "bbox": [ + 228, + 199, + 285, + 255 + ], + "lines": [ + { + "bbox": [ + 228, + 199, + 285, + 255 + ], + "spans": [ + { + "bbox": [ + 228, + 199, + 285, + 255 + ], + "type": "image", + "image_path": "3a709338daa81b755e8ccf9ea69cb47bfab24041cf42358f7cf6bba6ebb8c551.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 257, + 285, + 262 + ], + "lines": [ + { + "bbox": [ + 228, + 257, + 285, + 262 + ], + "spans": [ + { + "bbox": [ + 228, + 257, + 285, + 262 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 292, + 555, + 326 + ], + "lines": [ + { + "bbox": [ + 54, + 292, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 54, + 292, + 555, + 326 + ], + "type": "text", + "content": "Figure 7. Ablation study on blending latent step " + }, + { + "bbox": [ + 54, + 292, + 555, + 326 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 54, + 292, + 555, + 326 + ], + "type": "text", + "content": " (increasing left to right). The prompt \"steak\" is applied to an image of \"sushi plate\" while increasing " + }, + { + "bbox": [ + 54, + 292, + 555, + 326 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 54, + 292, + 555, + 326 + ], + "type": "text", + "content": " from left to right. At " + }, + { + "bbox": [ + 54, + 292, + 555, + 326 + ], + "type": "inline_equation", + "content": "b = n" + }, + { + "bbox": [ + 54, + 292, + 555, + 326 + ], + "type": "text", + "content": " (left), the edit disrupts the original structure, affecting unmasked regions. As " + }, + { + "bbox": [ + 54, + 292, + 555, + 326 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 54, + 292, + 555, + 326 + ], + "type": "text", + "content": " approaches " + }, + { + "bbox": [ + 54, + 292, + 555, + 326 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 292, + 555, + 326 + ], + "type": "text", + "content": " (right), background preservation improves, and edits blend more seamlessly." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 286, + 199, + 342, + 255 + ], + "blocks": [ + { + "bbox": [ + 286, + 199, + 342, + 255 + ], + "lines": [ + { + "bbox": [ + 286, + 199, + 342, + 255 + ], + "spans": [ + { + "bbox": [ + 286, + 199, + 342, + 255 + ], + "type": "image", + "image_path": "41509a13eaf4ae7e5924523f34d3aa009441bd40bd1e269c25688865ac0417fd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 266, + 373, + 275 + ], + "lines": [ + { + "bbox": [ + 314, + 266, + 373, + 275 + ], + "spans": [ + { + "bbox": [ + 314, + 266, + 373, + 275 + ], + "type": "text", + "content": "edit prompt: steak" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 344, + 199, + 400, + 255 + ], + "blocks": [ + { + "bbox": [ + 344, + 199, + 400, + 255 + ], + "lines": [ + { + "bbox": [ + 344, + 199, + 400, + 255 + ], + "spans": [ + { + "bbox": [ + 344, + 199, + 400, + 255 + ], + "type": "image", + "image_path": "7950237fa58273d2802c1df99d299bb20d0e95f5b8211e9f594dd5925a130c47.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 402, + 199, + 459, + 255 + ], + "blocks": [ + { + "bbox": [ + 402, + 199, + 459, + 255 + ], + "lines": [ + { + "bbox": [ + 402, + 199, + 459, + 255 + ], + "spans": [ + { + "bbox": [ + 402, + 199, + 459, + 255 + ], + "type": "image", + "image_path": "f59e2dd7dabb7bc3493e2fb362f36deb7799fc6835a0692d418ddb1022c0d4b4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 264, + 458, + 274 + ], + "lines": [ + { + "bbox": [ + 402, + 264, + 458, + 274 + ], + "spans": [ + { + "bbox": [ + 402, + 264, + 458, + 274 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 460, + 199, + 516, + 255 + ], + "blocks": [ + { + "bbox": [ + 460, + 199, + 516, + 255 + ], + "lines": [ + { + "bbox": [ + 460, + 199, + 516, + 255 + ], + "spans": [ + { + "bbox": [ + 460, + 199, + 516, + 255 + ], + "type": "image", + "image_path": "8eeae6152bc0fc228e844a472e57441ab2e86833f4da74cf47384b2e91bb1875.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 469, + 265, + 507, + 274 + ], + "lines": [ + { + "bbox": [ + 469, + 265, + 507, + 274 + ], + "spans": [ + { + "bbox": [ + 469, + 265, + 507, + 274 + ], + "type": "inline_equation", + "content": "b = N - 1" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 54, + 346, + 295, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 346, + 295, + 370 + ], + "spans": [ + { + "bbox": [ + 54, + 346, + 295, + 370 + ], + "type": "text", + "content": "efficient latent caching (Sec. 3.1), minimizing computation and memory overhead (" + }, + { + "bbox": [ + 54, + 346, + 295, + 370 + ], + "type": "inline_equation", + "content": "\\sim 1.25" + }, + { + "bbox": [ + 54, + 346, + 295, + 370 + ], + "type": "text", + "content": " MB for 10 layers)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 54, + 374, + 295, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 374, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 54, + 374, + 295, + 542 + ], + "type": "text", + "content": "Quantitative Performance: LDB demonstrates a superior combination of speed, image quality, and edit fidelity across both benchmarks. On the PIE-Bench dataset, LDB achieves the best performance in six key metrics, excelling in human preference (HPS = 86.02), background preservation (LPIPS = 1.91), and text alignment (CS = 31.66), while also being the fastest method by a significant margin. This highlights its ability to generalize across a diverse set of editing tasks while maintaining high speed. Similarly, on the MagicBrush benchmark, LDB delivers strong performance with highest score in crucial metrics such as IR, AS, and PSNR. While BrushNet shows a slight advantage in some text alignment metrics, its practical usability is hindered by substantially slower runtime." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 560, + 217, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 217, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 217, + 572 + ], + "type": "text", + "content": "5.1. Limitations and Future Work" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "content": "Brush strength " + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "(\\alpha)" + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "content": " and diffusion step count " + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "(n)" + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "content": " coupling (Fig. 12) still requires minor user tuning across scenarios. Although preset profiles partially address this, future work could explore adaptive parameter tuning mechanisms to further improve usability. Moreover, semantically implausible edits (e.g. placing a boat in the sky) remain challenging due to inherent biases within diffusion models. Integrating techniques like semantic guidance could expand plausible edit ranges. Finally, responsible deployment necessitates robust watermarking [15] and provenance tracking to mitigate misuse and ensure transparency." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 346, + 437, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 437, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 437, + 358 + ], + "type": "text", + "content": "5.2. Broader Applications" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 367, + 554, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 554, + 451 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 554, + 451 + ], + "type": "text", + "content": "LDB's training-free design only requires a standard iterative denoising process, which allows seamless integration into diverse diffusion models and applications requiring rapid editing. We validated this by integrating LDB to other commonly used methods, including DiT-based text-to-image (e.g., PixArt-" + }, + { + "bbox": [ + 313, + 367, + 554, + 451 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 367, + 554, + 451 + ], + "type": "text", + "content": " [10]) and video generation models [7] without any model-specific tuning." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 454, + 554, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 554, + 514 + ], + "type": "text", + "content": "Traditional diffusion-based video editing typically propagates edits from the first frame using additional supervision (e.g. optical flow [36]), risking temporal inconsistencies. LDB's high fidelity background preservation and efficiency naturally address these issues." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 518, + 554, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 554, + 590 + ], + "type": "text", + "content": "We demonstrate preliminary success integrating LDB with Stable Video Diffusion (SVD) [7], editing the first frame and applying LDB's latent caching across frames for fast consecutive edits (see supplementary material, Fig. 15). This approach opens avenues for accelerated video manipulation, 3D asset editing, and collaborative design platforms." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 608, + 389, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 608, + 389, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 608, + 389, + 620 + ], + "type": "text", + "content": "5.3. Conclusion" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": "LDB reimagines diffusion-based editing through latent caching and non-destructive layering, achieving unmatched speed and control. Quantitative results and user study show superior performance in image preference, edit fidelity, time, and usability. By bridging interactive editing with high-fidelity generative models, LDB can empower artists to iterate fluidly while maintaining artistic intent." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17375" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "content": "This work was supported by the Natural Sciences and Engineering Research Council of Canada (NSERC). The authors would also like to thank Professors Leonid Signal and Kwang Moo Yi for their guidance and support throughout this project." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 180, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 180, + 294, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 180, + 294, + 234 + ], + "spans": [ + { + "bbox": [ + 61, + 180, + 294, + 234 + ], + "type": "text", + "content": "[1] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 236, + 296, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 236, + 296, + 302 + ], + "spans": [ + { + "bbox": [ + 61, + 236, + 296, + 302 + ], + "type": "text", + "content": "[2] Shubham Agarwal, Subrata Mitra, Sarthak Chakraborty, Srikrishna Karanam, Koyel Mukherjee, and Shiv Kumar Saini. Approximate caching for efficiently serving {Text-to-Image} diffusion models. In 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI 24), pages 1173-1189, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 303, + 294, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 303, + 294, + 324 + ], + "spans": [ + { + "bbox": [ + 61, + 303, + 294, + 324 + ], + "type": "text", + "content": "[3] Ideogram AI. Ideogram: Text-to-image generation platform. https://ideogram.ai, 2025.5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 325, + 295, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 325, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 62, + 325, + 295, + 369 + ], + "type": "text", + "content": "[4] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 370, + 295, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 370, + 295, + 402 + ], + "spans": [ + { + "bbox": [ + 62, + 370, + 295, + 402 + ], + "type": "text", + "content": "[5] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM Transactions on Graphics (TOG), 42 (4):1-11, 2023. 2, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 403, + 295, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 295, + 458 + ], + "type": "text", + "content": "[6] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 707-723. Springer, 2022. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 459, + 294, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 459, + 294, + 513 + ], + "spans": [ + { + "bbox": [ + 62, + 459, + 294, + 513 + ], + "type": "text", + "content": "[7] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 8, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 514, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 294, + 536 + ], + "type": "text", + "content": "[8] John Brooke. SUS-A quick and dirty usability scale. Usability evaluation in industry, 189(194), 1996. 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 536, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 294, + 568 + ], + "type": "text", + "content": "[9] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 2, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 570, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 570, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 570, + 295, + 624 + ], + "type": "text", + "content": "[10] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 625, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 669 + ], + "type": "text", + "content": "[11] Erin Cherry and Celine Latulipe. Quantifying the creativity support of digital tools through the creativity support index. ACM Transactions on Computer-Human Interaction (TOCHI), 21(4):1-25, 2014. 5, 9" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[12] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. Ilvr: Conditioning method for denoising diffusion probabilistic models. arXiv preprint arXiv:2108.02938, 2021. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 703 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 94 + ], + "type": "text", + "content": "[13] Civitai. Dreamshaper - 7 — stable diffusion checkpoint, 2024. Accessed on Feb 19, 2024. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 95, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 95, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 317, + 95, + 553, + 138 + ], + "type": "text", + "content": "[14] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "type": "text", + "content": "[15] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22466-22477, 2023. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 195, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 249 + ], + "type": "text", + "content": "[16] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 251, + 553, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 294 + ], + "type": "text", + "content": "[17] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 294, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 294, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 317, + 294, + 553, + 348 + ], + "type": "text", + "content": "[18] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Han Hu, Dong Chen, et al. Instructdiffusion: A generalist modeling interface for vision tasks. arXiv preprint arXiv:2309.03895, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 350, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 415 + ], + "type": "text", + "content": "[19] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, et al. Instructdiffusion: A generalist modeling interface for vision tasks. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 12709-12720, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "type": "text", + "content": "[20] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 460, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 492 + ], + "type": "text", + "content": "[21] Qin Guo and Tianwei Lin. Focus on your instruction: Fine-grained and multi-instruction image editing by attention modulation. arXiv preprint arXiv:2312.10113, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 493, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 536 + ], + "type": "text", + "content": "[22] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 537, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 569 + ], + "type": "text", + "content": "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 571, + 553, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 553, + 603 + ], + "type": "text", + "content": "[24] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssm. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 604, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 553, + 647 + ], + "type": "text", + "content": "[25] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 647, + 553, + 703 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 703 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 703 + ], + "type": "text", + "content": "[26] Minghui Hu, Yujie Wang, Tat-Jen Cham, Jianfei Yang, and Ponnuthurai N Suganthan. Global context with discrete diffusion in vector quantised modelling for image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11502-11511, 2022." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "17376" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[27] Mingzhen Huang, Jialing Cai, Shan Jia, Vishnu Suresh Lokhande, and Siwei Lyu. Paralleledits: Efficient multi-aspect text-driven image editing with attention grouping. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 163 + ], + "type": "text", + "content": "[28] Adobe Inc. Adobe photoshop (2024 version). https://www.adobe.com/products/photoshop.htm1, 2024. 2, 5, 7, 9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 294, + 229 + ], + "type": "text", + "content": "[29] K. J. Joseph, Prateksha Udhayanan, Tripti Shukla, Aishwarya Agarwal, Srikrishna Karanam, Koustava Goswami, and Balaji Vasan Srinivasan. Iterative multi-granular image editing using diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 8107-8116, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "type": "text", + "content": "[30] Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. Direct inversion: Boosting diffusion-based editing with 3 lines of code. arXiv preprint arXiv:2310.01506, 2023. 2, 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "type": "text", + "content": "[31] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. In European Conference on Computer Vision, pages 150–168. Springer, 2024. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 386 + ], + "type": "text", + "content": "[32] Kumara Kahatapitiya, Haozhe Liu, Sen He, Ding Liu, Menglin Jia, Chenyang Zhang, Michael S Ryoo, and Tian Xie. Adaptive caching for faster video generation with diffusion transformers. arXiv preprint arXiv:2411.02397, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 442 + ], + "type": "text", + "content": "[33] Kyoungkook Kang, Gyujin Sim, Geonung Kim, Donguk Kim, Seungho Nam, and Sunghyun Cho. Layeringdiff: Layered image synthesis via generation, then disassembly with generative knowledge. arXiv preprint arXiv:2501.01197, 2025. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "type": "text", + "content": "[34] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2426-2435, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 501, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 567 + ], + "type": "text", + "content": "[35] Oran Lang, Yossi Gandelsman, Michal Yarom, Yoav Wald, Gal Elidan, Avinatan Hassidim, William T Freeman, Phillip Isola, Amir Globerson, Michal Irani, et al. Explaining in style: Training a gan to explain a classifier in stylespace. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 693-702, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 568, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 634 + ], + "type": "text", + "content": "[36] Feng Liang, Bichen Wu, Jialiang Wang, Licheng Yu, Kunpeng Li, Yinan Zhao, Ishan Misra, Jia-Bin Huang, Peizhao Zhang, Peter Vajda, et al. Flowvid: Taming imperfect optical flows for consistent video-to-video synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8207-8216, 2024. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 635, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 690 + ], + "type": "text", + "content": "[37] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 1, 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[38] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Deepcache: Accelerating diffusion models for free. In Proceedings of" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "the IEEE/CVF conference on computer vision and pattern recognition, pages 15762-15772, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "text", + "content": "[39] Hayk Manukyan, Andranik Sargsyan, Barsegh Atanyan, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Hd-painter: High-resolution and prompt-faithful text-guided image inpainting with diffusion models. arXiv preprint arXiv:2312.14091, 2023. 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 151, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 205 + ], + "type": "text", + "content": "[40] Qi Mao, Lan Chen, Yuchao Gu, Zhen Fang, and Mike Zheng Shou. Mag-edit: Localized image editing in complex scenarios via mask-based attention-adjusted guidance. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6842-6850, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 206, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 227 + ], + "type": "text", + "content": "[41] Inc. Midjourney. Midjourney: Ai image generation. https://www.midjourney.com, 2025.5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "type": "text", + "content": "[42] Daiki Miyake, Akihiro Iohara, Yu Saito, and Toshiyuki Tanaka. Negative-prompt inversion: Fast image inversion for editing with text-guided diffusion models. arXiv preprint arXiv:2305.16807, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 273, + 555, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 273, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 273, + 555, + 327 + ], + "type": "text", + "content": "[43] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "type": "text", + "content": "[44] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag your gan: Interactive point-based manipulation on the generative image manifold. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 383, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 437 + ], + "type": "text", + "content": "[45] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, New York, NY, USA, 2023. Association for Computing Machinery. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 438, + 553, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 471 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 471 + ], + "type": "text", + "content": "[46] Thomas Porter and Tom Duff. Compositing digital images. In Proceedings of the 11th annual conference on Computer graphics and interactive techniques, pages 253-259, 1984. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 472, + 553, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 553, + 535 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 553, + 535 + ], + "type": "text", + "content": "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 537, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 581 + ], + "type": "text", + "content": "[48] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "type": "text", + "content": "[49] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2287–2296, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 636, + 555, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 555, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 555, + 690 + ], + "type": "text", + "content": "[50] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[51] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17377" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "tuning text-to-image diffusion models for subject-driven generation, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "type": "text", + "content": "[52] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "text", + "content": "[53] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "type": "text", + "content": "[54] Vishnu Sarukkai, Linden Li, Arden Ma, Christopher Ré, and Kayvon Fatahalian. Collage diffusion. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 4208-4217, 2024. 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 308 + ], + "type": "text", + "content": "[55] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in neural information processing systems, 35:25278-25294, 2022. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 310, + 294, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 294, + 354 + ], + "type": "text", + "content": "[56] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 355, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 294, + 421 + ], + "type": "text", + "content": "[57] Felix Wimbauer, Bichen Wu, Edgar Schoenfeld, Xiaoliang Dai, Ji Hou, Zijian He, Artsiom Sanakoyeu, Peizhao Zhang, Sam Tsai, Jonas Kohler, et al. Cache me if you can: Accelerating diffusion models through block caching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6211-6220, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 475 + ], + "type": "text", + "content": "[58] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 478, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 533 + ], + "type": "text", + "content": "[59] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22428-22437, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 534, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 578 + ], + "type": "text", + "content": "[60] Zishan Xu, Xiaofeng Zhang, Wei Chen, Minda Yao, Jueting Liu, Tingting Xu, and Zehua Wang. A review of image inpainting methods based on deep learning. Applied Sciences, 13(20):11189, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "text", + "content": "[61] Fei Yang, Shiqi Yang, Muhammad Atif Butt, Joost van de Weijer, et al. Dynamic prompt learning: Addressing cross-attention leakage for text-based image editing. Advances in Neural Information Processing Systems, 36:26291–26303, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[62] Shiyuan Yang, Xiaodong Chen, and Jing Liao. Uni-paint: A unified framework for multimodal image inpainting with pretrained diffusion model. In Proceedings of the 31st ACM International Conference on Multimedia, pages 3190-3199, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "text", + "content": "[63] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything:" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 375 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[64] Ziyun Zeng, Hang Hua, Jianlong Fu, Jiebo Luo, et al. Promptfix: You prompt and we fix the photo. Advances in Neural Information Processing Systems, 37:40000-40031, 2025. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "text", + "content": "[65] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 36, 2024. 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 186, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 217 + ], + "type": "text", + "content": "[66] Lvmin Zhang and Maneesh Agrawala. Transparent image layer diffusion using latent transparency. arXiv preprint arXiv:2402.17113, 2024. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 220, + 553, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 263 + ], + "type": "text", + "content": "[67] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3836-3847, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 265, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 265, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 265, + 553, + 319 + ], + "type": "text", + "content": "[68] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 320, + 553, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 375 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 375 + ], + "type": "text", + "content": "[69] Junhao Zhuang, Yanhong Zeng, Wenran Liu, Chun Yuan, and Kai Chen. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. In European Conference on Computer Vision, pages 195-211. Springer, 2024. 2" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17378" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Street Gaussians without 3D Object Tracker/ceea65cc-c6ed-4695-b497-f6fc77f3a4c6_content_list.json b/2025/Street Gaussians without 3D Object Tracker/ceea65cc-c6ed-4695-b497-f6fc77f3a4c6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..baa5551ccd0197fd61199042459bd247e0670c27 --- /dev/null +++ b/2025/Street Gaussians without 3D Object Tracker/ceea65cc-c6ed-4695-b497-f6fc77f3a4c6_content_list.json @@ -0,0 +1,1855 @@ +[ + { + "type": "text", + "text": "Street Gaussians without 3D Object Tracker", + "text_level": 1, + "bbox": [ + 272, + 130, + 723, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ruida Zhang $^{1,2}$ , Chengxi Li $^{1}$ , Chenyangguang Zhang $^{1}$ , Xingyu Liu $^{1}$ , Haili Yuan $^{1}$ , Yanyan Li $^{2}$ , Xiangyang Ji $^{1}$ , Gim Hee Lee $^{2}$ $^{1}$ Tsinghua University, $^{2}$ National University of Singapore", + "bbox": [ + 176, + 179, + 818, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhangrd23@mails, xyji@}.tsinghua.edu.cn, gimhee.lee@nus.edu.sg", + "bbox": [ + 217, + 236, + 774, + 252 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5c0906e1d8a489a3a426abffcc9c2fb05a484304038c38d4a87724cee7c557e9.jpg", + "image_caption": [ + "Figure 1. Comparison of 3D tracker-based Street Gaussians [89] (left) and our approach (right). Existing methods heavily rely on object poses, but 3D trackers struggle with limited generalization [21, 73, 97], leading to flaws in novel view synthesis. In contrast, 2D foundation models show better generalization [84, 88]. Our approach leverages a 2D foundation model [84] for object tracking and learns point motion within an implicit feature space to autonomously correct tracking errors, improving robustness across diverse scenes." + ], + "image_footnote": [], + "bbox": [ + 96, + 263, + 491, + 467 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7593f0549658f8187461f1f91ffff7fd5f5e85cbf312b35d5e34ce8d5fa0f1ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 263, + 901, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 547, + 326, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Realistic scene reconstruction in driving scenarios poses significant challenges due to fast-moving objects. Most existing methods rely on labor-intensive manual labeling of object poses to reconstruct dynamic objects in canonical space and move them based on these poses during rendering. While some approaches attempt to use 3D object trackers to replace manual annotations, the limited generalization of 3D trackers – caused by the scarcity of large-scale 3D datasets – results in inferior reconstructions in real-world settings. In contrast, 2D foundation models demonstrate strong generalization capabilities. To eliminate the reliance on 3D trackers and enhance robustness across diverse environments, we propose a stable object tracking module by leveraging associations from 2D deep trackers within a 3D object fusion strategy. We address inevitable tracking errors by further introducing a motion learning strategy in an implicit feature space that autonomously corrects trajectory errors and recovers missed detections. Experimental results on Waymo-NOTR and KITTI show that our method outperforms existing approaches. Our code will be made publicly available.", + "bbox": [ + 88, + 579, + 485, + 898 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 547, + 643, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Modeling dynamic 3D street scenes underpins modern autonomous driving by enabling realistic, controllable simulations for tasks such as perception [11, 42, 56, 94, 105], prediction [25, 33, 47], and motion planning [13, 14, 16]. With the rise of end-to-end autonomous systems that require real-time sensor feedback [29, 31, 35], high-quality scene reconstructions have become essential for closed-loop evaluations [46, 96], particularly to simulate critical corner cases safely and cost-effectively.", + "bbox": [ + 511, + 577, + 906, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite extensive efforts in achieving photo-realistic reconstruction of small-scale scenes, the unique challenges posed by the large-scale and highly dynamic nature of driving scenarios complicate effective 3D scene modeling. To address these challenges, most existing methods [22, 38, 65, 85, 106] rely on ground truth vehicle poses to differentiate between static background and moving vehicles. Typically, vehicles are reconstructed in canonical space and subsequently positioned based on known poses during rendering. However, collecting ground-truth poses is labor-intensive, limiting the applicability of these methods to scenes beyond the existing datasets.", + "bbox": [ + 509, + 719, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25722", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To eliminate reliance of the systems on ground truth object poses, Street Gaussians [89] instead uses poses generated by 3D object trackers [82, 83]. However, by defining vehicle motion solely through these tracking poses, rendering quality becomes highly dependent on pose accuracy. While Street Gaussians optimizes object poses during training, it struggles with detection failures and large pose errors, as shown in Fig. 1. Unfortunately, 3D trackers often struggle to generalize effectively across different scenarios, primarily due to the scarcity of open-source largescale 3D datasets [79, 103]. Collecting ground-truth poses for 3D tracking is time-consuming and costly, resulting in relatively few open-source datasets for autonomous driving. For example, even widely-used datasets such as Waymo [76], Nuscenes [6], KITTI [23, 49], Pandaset [86] collectively contain fewer than 4,000 annotated scenes, limiting the diversity and scale needed for robust 3D tracking. In contrast, 2D data is far easier to collect and annotate, leading to an abundance of 2D datasets. The BDD100K dataset [95] alone includes annotations for 100,000 driving scenes, and additional large-scale 2D perception datasets, such as COCO [50], LVIS [28], and OpenImages [41], bring the total number of annotated scenes into the millions. This wealth of 2D data has enabled 2D foundation models to achieve strong generalization across a variety of tasks, including embedding extraction [64, 68], detection [8, 62], segmentation [12, 40], and tracking [84]. Nonetheless, how to use these accurate and robust 2D trackers to effectively model dynamic objects in 3D is still an open challenge for street scene reconstruction.", + "bbox": [ + 93, + 90, + 480, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, a novel architecture is designed to achieve high-fidelity novel view synthesis performance in street scenarios for autonomous driving applications. Unlike previous approaches such as Driving Gaussian [106] and Street Gaussian [89] that rely on ground truth vehicle poses or poses predicted by 3D trackers, the first contribution of this work is a stable object tracking module. This module leverages associations from 2D deep trackers within a 3D object fusion framework to enhance robustness and accuracy. Specifically, we integrate 2D tracking outputs with LiDAR data to trace the trajectory of each vehicle in 3D. We then incrementally reconstruct the point cloud of each vehicle frame-by-frame in the canonical space, estimating its pose by aligning successive frames with this canonical model. This approach eliminates the reliance on 3D trackers and enhances robustness across diverse environments.", + "bbox": [ + 93, + 549, + 480, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although 2D tracking models demonstrate stronger generalization ability, tracking errors still can exist, especially under adverse conditions such as severe occlusions or distant objects. Since moving the object points solely based on the tracked object pose directly exposes any tracking errors, we aim to go beyond a straightforward reliance on tracked trajectories. Instead, we propose to learn point mo", + "bbox": [ + 93, + 795, + 480, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion from the predicted trajectory, equipping the model to autonomously identify and correct tracking inaccuracies, recover missed detections, and infer motion in new time steps. An implicit representation is essential for this purpose: 1) It enables the model to refine trajectories without being constrained by bounding box tracks, facilitating smoother and more continuous motion that can adaptively respond to changes. 2) An implicit feature space offers versatility to move each point of an object in a different way, recognizing that vehicles are not strictly rigid objects — for example, doors can open or close. This approach makes it possible to capture subtle, dynamic changes within objects, ultimately producing more accurate reconstructions and enhancing the robustness of novel view synthesis in challenging scenarios.", + "bbox": [ + 516, + 90, + 903, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we leverage HexPlane representation [7] following 4DGS [81]. HexPlane stores motion-related features by decomposing the 4D spatial-temporal space into six 2D learnable feature planes. A decoder then utilizes these features to predict deformation offsets, dynamically adjusting 3D Gaussians over time. However, 4DGS relies solely on image reconstruction loss for supervision, which is insufficient for handling rapid object motion. When a Gaussian's initial position is far from its current location, its projection on the image may fall outside the object's actual area, preventing gradient propagation and hindering optimization. This issue is evident in S3Gaussian [32], which struggles with moving cars due to the lack of explicit motion supervision. To address this, we introduce a training strategy that supervises learned point motion using the predicted trajectory, providing explicit 3D supervision to guide HexPlane in capturing motion dynamics more accurately.", + "bbox": [ + 516, + 319, + 903, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are summarized as follows:", + "bbox": [ + 535, + 577, + 870, + 590 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We eliminate the need of 3D tracker for street scene reconstruction by introducing a robust object tracking module which leverages 2D foundation model, achieving superior generalization ability across diverse scenarios.", + "- We introduce a motion learning framework to learn from the predicted trajectory in an implicit feature space, enabling it to automatically correct pose errors and infer motion for novel time steps.", + "- We outperform existing methods on Waymo-NOTR and KITTI datasets without relying on ground truth annotations." + ], + "bbox": [ + 513, + 593, + 903, + 756 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 516, + 776, + 661, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. 3D Gaussian Splatting for Dynamic Scene", + "text_level": 1, + "bbox": [ + 516, + 801, + 867, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Gaussian Splatting (3DGS) [37, 44] has advanced scene reconstruction by enabling high-quality, real-time rendering with 3D Gaussians and efficient splat-based rasterization, reducing computation and parameters compared to NeRF-based methods [1-3, 61, 63] and other representa", + "bbox": [ + 516, + 825, + 903, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25723", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tions [15, 26, 27, 30, 45, 75, 102]. While originally designed for static scenes, 3DGS has been adapted for dynamic scenes [48, 53, 57, 74, 81, 92, 93]. Dynamic3DGS [57] directly stores information for each 3D Gaussian at every timestamp, and Yang et al. [92] approximate the spatiotemporal 4D volume by optimizing 4D Gaussian primitives. Deformable-3DGS [93] and GauFRe [48] employ deformation fields to model motions, while 4DGS [81] introduces HexPlane representation [7] to store spatial-temporal features efficiently. Gaussian Marbles [74] leverages isotropic \"marbles\" and a divide-and-conquer trajectory learning algorithm. Despite these advances, challenges persist in high-speed driving scenarios with rapid object motion. To tackle this challenge, we leverage a robust object tracking module to guide the motion learning process.", + "bbox": [ + 89, + 90, + 480, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Street Scene Reconstruction", + "text_level": 1, + "bbox": [ + 89, + 329, + 341, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing autonomous driving simulation engines [20, 43, 52, 71] face high manual effort in creating virtual environments and a lack of realism in the generated data. The creation of high-fidelity simulations from driving logs is therefore essential for advancing closed-loop training and testing. Recent works [65, 72, 77, 78, 85, 87, 90, 91] have continuously made improvements to NeRF [61] to model street scenes dynamically. Despite the progress, NeRF-based methods remain computationally expensive and require densely overlapping views. Building on the effective 3DGS approach [37] for scene reconstruction, several 3DGS-based methods [10, 67, 89, 106] have emerged. PVG [10] model dynamic scenarios by using periodic vibration-based temporal dynamics. Driving Gaussian [106] introduces incremental static Gaussians and composite dynamic Gaussian graph. Street Gaussians [89] equips Gaussians with semantic logits, and optimize dynamic parts using tracked poses from 3D tracker. AutoSplat [38] enforces geometric constraints in road and sky regions for multi-view consistency. Among these approaches, most of them requires ground truth object pose [10, 38, 106] or the 3D tracker [89]. However, manual annotation is laborious and the 3D tracker is lack in generalization ability, which limits their applications in diverse scenarios. In contrast, S3Gaussian [32] models dynamics in a self-supervised manner using dynamic Gaussians from 4DGS [81], despite struggling with dynamic object modeling due to the lack of explicit motion supervision. We address this challenge by introducing a robust object tracking strategy based on a 2D foundation model [84] and apply motion supervision from the predicted object trajectory.", + "bbox": [ + 89, + 352, + 483, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. 2D and 3D trackers", + "text_level": 1, + "bbox": [ + 89, + 832, + 277, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multiple Object Tracking (MOT) aims at locating multiple objects in each frame, and establishing correspondences between them across frames within input videos", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "[4, 5, 36, 58, 98]. The reliance of most MOT methods on object detection [4, 19, 58, 66, 98, 104] makes detection accuracy crucial for MOT performance. Recent advances in 3D object detection are promising [11, 18, 34, 42, 51, 54, 56, 60, 80, 94, 99-101], but these models often generalize poorly due to limited and biased datasets [21, 55, 73, 97]. Existing open-source autonomous driving datasets lack scale and show regional or environmental biases, such as vehicle density and weather conditions [6, 23, 59, 76]. The main reason is that building large-scale, well-annotated multimodal datasets is costly [6, 49, 59, 69]. In contrast, 2D image data is easier to capture and annotate, enabling extensive dataset collections [9, 17, 41, 95]. Visual foundation models trained on large-scale 2D data have shown strong generalization [12, 40, 64, 68, 84]. We thus opt to use 2D object tracker [84] to locate the dynamic objects.", + "bbox": [ + 511, + 90, + 903, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Our Method", + "text_level": 1, + "bbox": [ + 513, + 359, + 643, + 376 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Prerequisites: 3D Gaussian Splatting", + "text_level": 1, + "bbox": [ + 511, + 385, + 834, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Gaussian Splatting (3DGS) [37] represents a 3D scene explicitly as a collection of 3D Gaussian primitives. Each Gaussian is defined by its center position $\\mathcal{X} \\in \\mathbb{R}^3$ and covariance $\\Sigma \\in \\mathbb{R}^{3 \\times 3}$ . To enforce semi-positive definiteness, the covariance matrix is further factorized into a scaling vector $S \\in \\mathbb{R}^3$ and a rotation matrix $R \\in SO(3)$ by:", + "bbox": [ + 511, + 407, + 905, + 500 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Sigma = R S S ^ {\\top} R ^ {\\top} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 506, + 903, + 525 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Additionally, each Gaussian is described with its opacity $o \\in \\mathbb{R}$ and view-dependent color defined by spherical harmonic (SH) coefficient $\\mathcal{C} \\in \\mathbb{R}^k$ , where $k$ represents numbers of SH functions.", + "bbox": [ + 511, + 534, + 905, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During rendering of novel views, differential splatting is applied to 3D Gaussians within the camera planes. The blending of $N$ ordered points that overlap a pixel is given by the formula:", + "bbox": [ + 511, + 595, + 905, + 655 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nC = \\sum_ {i \\in N} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 665, + 903, + 708 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{i}$ and $c_{i}$ represents the opacity and color of the ith splatted Gaussian, which is computed from per-point opacity and SH coefficients (see [37] for details).", + "bbox": [ + 511, + 715, + 905, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overview", + "text_level": 1, + "bbox": [ + 511, + 770, + 624, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method takes as input multi-view images $\\mathbf{I}_t^{(j)}$ from multiple cameras placed around the vehicle, each indexed by time step $t$ and camera index $j$ , along with intrinsic $\\mathbf{K}^{(j)}$ and extrinsic $\\mathbf{E}_t^{(j)}$ matrices for each view. Additionally, a top-mounted LiDAR provides 3D point clouds $\\mathbf{L}_t$ for each frame. Using this multi-sensor data, our approach reconstructs the 3D scene and synthesizes novel views from any", + "bbox": [ + 511, + 790, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25724", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bb66c5e49d0b54cdbedde8e690b18e2232ffc8dcb586ecf24b921d2354e23ad2.jpg", + "image_caption": [ + "Figure 2. Overview of our method. To overcome the limited generalization of 3D object trackers, we introduce a robust object tracking module based on a 2D object tracker [84]. We integrate 2D tracking with LiDAR data to segment the object's point cloud and incrementally reconstruct the object model in canonical space frame-by-frame. The canonical model is used to estimate object pose $\\mathbf{T}_t$ and serves as the initialization for Gaussian optimization. We model dynamic objects with isotropic Gaussian marble [74] to simplify optimization. Our approach learns point motion within the HexPlane [7] feature space, enabling tracking error correction and missed detection recovery. Decoders $D_{\\mathcal{X}}, D_{\\mathcal{C}}$ predict point motion $\\Delta \\mathcal{X}_t$ and color change $\\Delta \\mathcal{C}_t$ based on the HexPlane features $f(x,y,z,t)$ , with the predicted trajectory providing supervision. Finally, novel view synthesis is performed by splatting the deformed Gaussian onto the image plane." + ], + "image_footnote": [], + "bbox": [ + 101, + 85, + 901, + 286 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "desired camera pose and time frame, without reliance on 3D object trackers or ground truth object trajectories.", + "bbox": [ + 88, + 417, + 480, + 448 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We present an overview of our method in Fig. 2. Visual images and point clouds are obtained from the sensor setup. For images, static and dynamic components are analyzed first. We employ Mask2Former [12] to segment the scene into static and dynamic parts. Specifically, the dynamic part includes humans (pedestrians, cyclists, etc.) and vehicles (cars, trucks, etc.), and all the other objects are categorized as static. We model the static part as in 3DGS [37] and focus on modeling dynamic objects in the following sections. To enhance robustness, we propose an object tracking strategy by leveraging a robust 2D tracker [84] and the LiDAR point cloud, providing accurate initial point clouds and object trajectories for subsequent Gaussian optimization (cf. Sec 3.3). We leverage an implicit representation HexPlane [7] following 4DGS [81] to learn a continuous and smooth per-point motion based on the object trajectory (cf. Sec 3.4). The optimization objective is described in Sec 3.5.", + "bbox": [ + 91, + 449, + 482, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Tracking Cars in 3D with Robust 2D Tracker", + "text_level": 1, + "bbox": [ + 89, + 712, + 473, + 728 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To improve the robustness of novel view synthesis for diverse scenarios, the first challenge is how to track and build vehicles in 3D based on the 2D tracking results.", + "bbox": [ + 89, + 734, + 482, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We employ GLEE [84] to track all vehicles within the 2D image plane, obtaining 2D object trajectories in each camera view. Each trajectory includes the 2D segmentation masks of the object over a time period. To lift 2D trajectories into 3D, we first re-project the LiDAR point clouds onto the image plane and assign points to objects based on their presence within the corresponding 2D segmentation masks $\\mathcal{M}$ . This yields a segmented object point clouds for each", + "bbox": [ + 89, + 779, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "camera view via the following function:", + "bbox": [ + 511, + 417, + 777, + 434 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {O} _ {t} ^ {(j)} = \\left\\{\\mathbf {P} \\mid \\Pi^ {- 1} \\left(\\mathbf {P}, \\mathbf {K} ^ {(j)}, \\mathbf {E} _ {t} ^ {(j)}\\right) \\in \\mathcal {M}, \\mathbf {P} \\in \\mathbf {L} _ {t} \\right\\}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 445, + 903, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $j$ is the camera index, $t$ is the time index, $\\mathbf{P}$ is a LiDAR point and $\\Pi(\\cdot)$ is the re-projection function. Given the slight misalignment between the LiDAR point cloud and RGB image, points near the edges of segmentation masks often fall outside the object boundaries. To address this issue, we perform outlier removal to improve stability in subsequent reconstruction based on the point distance to the point cloud center. We then associate the same objects across different camera views. Two point clouds in different viewpoints are considered belonging to the same object if they share more than 50 points in one time step. This association process results in a set of associated partial object point cloud in different time steps.", + "bbox": [ + 511, + 478, + 906, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After obtaining the 3D partial objects in different time steps, the next step is to associate these partial point cloud into a unified and complete model. We initialize the 3D reconstruction of each object from the first frame which it appears. For simplicity, we assume an object is visible between frames $0$ and $T$ and denote the object point cloud at time $t$ as $\\mathcal{O}_t$ . Our goal is to obtain a temporally consistent reconstruction $\\mathcal{O}$ and the object pose in each time frame. This reconstruction is done frame-by-frame incrementally.", + "bbox": [ + 511, + 674, + 906, + 810 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Starting with the initial frame, we add $\\mathcal{O}_0$ to $\\mathcal{O}$ . For each subsequent frame, we apply Iterative Closest Point (ICP) for the alignment of $\\mathcal{O}_t$ to $\\mathcal{O}$ by extracting the relative pose $\\mathbf{T}_t$ . The overlap between $\\mathcal{O}_t$ and $\\mathcal{O}$ is given by:", + "bbox": [ + 511, + 811, + 905, + 872 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tau_ {o v e r l a p} = \\mathcal {O} \\cap \\mathbf {T} _ {t} ^ {- 1} \\mathcal {O} _ {t}, \\quad \\mathbf {T} _ {t} = \\mathrm {I C P} (\\mathcal {O} _ {t}, \\mathcal {O}). \\quad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 883, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25725", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "If ICP reveals an overlap $\\tau_{\\text{overlap}}$ of more than $30\\%$ with $\\mathcal{O}$ , we record the object pose $\\mathbf{T}_t$ and update $\\mathcal{O}$ by integrating $\\mathbf{T}_t^{-1}\\mathcal{O}_t$ . If $\\tau_{\\text{overlap}}$ is between $10\\%$ and $30\\%$ , we only record the object pose $P_t$ without updating $\\mathcal{O}$ . If $\\tau_{\\text{overlap}}$ is lower than $10\\%$ , we assume a 2D tracking failure for that frame and discard it to filter out tracking errors. This process is repeated for each frame using the last recorded pose as the initial guess for ICP to accelerate convergence. Through this approach, we reconstruct all objects and retrieve their poses in each frame, providing a robust initialization for subsequent 3D Gaussian optimization.", + "bbox": [ + 89, + 90, + 483, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Learning Point Motion", + "text_level": 1, + "bbox": [ + 89, + 266, + 303, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A common approach to model object motion in prior works is directly transforming the object points using the given object pose. However, despite the strong generalization capabilities of 2D trackers, detection failures are inevitable in challenging scenarios, such as when objects are heavily occluded or located at great distances. Relying solely on object poses derived from these erroneous or missing detections can easily lead to rendering failures. Additionally, treating cars as rigid objects does not adequately address corner cases, such as when a car door is open or closed. Furthermore, these methods lack the ability to infer an object motion at arbitrary time stamps.", + "bbox": [ + 89, + 287, + 483, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address these limitations, we seek for a motion modeling approach to enhance robustness and flexibility. Instead of explicitly using object pose as rigid transformation, we learn the per-point motion in a pre-defined feature space. In this feature space, the object motion can be optimized through both explicit guidance and photometric loss. Moreover, the object motion can be interpolated through time and space in this feature space to compensate for missing detections. To this end, we leverage HexPlane representation [7] as in 4DGS [81] to efficiently capture spatial and temporal information by decomposing the 4D feature voxels into six learnable feature planes. This representation satisfies all our requirements and is also memory efficient.", + "bbox": [ + 89, + 469, + 483, + 665 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use three planes: $P_{xy}$ , $P_{yz}$ , and $P_{xz}$ for the spatial dimensions, and another three planes $P_{xt}$ , $P_{yt}$ , and $P_{zt}$ for the spatial-temporal features. Additionally, the Hexplane includes multiple resolution levels which is formulated as:", + "bbox": [ + 89, + 666, + 483, + 726 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{P _ {i j} ^ {\\rho} \\in \\mathbb {R} ^ {d \\times \\rho r _ {i} \\times \\rho r _ {j}} | (i, j) \\in \\mathcal {P}, \\rho \\in \\{1, 2 \\} \\right\\}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 738, + 482, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{P} = \\{(x,y),(x,z),(y,z),(x,t),(y,t),(z,t)\\}$ , $d$ is the feature dimension, $\\rho$ denotes the upsampling scale, and $r$ is the base resolution.", + "bbox": [ + 89, + 768, + 482, + 813 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the center position of a Gaussian $(x,y,z)$ and the time step $t$ , features in all six planes are queried and combined via a small MLP $\\phi_d$ as follows:", + "bbox": [ + 89, + 813, + 483, + 859 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf (x, y, z, t) = \\phi_ {d} \\left(\\bigcup_ {\\rho} \\prod_ {(i, j) \\in \\mathcal {P}} \\pi \\left(P _ {i j} ^ {\\rho}, \\psi_ {i j} ^ {\\rho} (x, y, z, t)\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 869, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\psi_{ij}^{\\rho}(x,y,z,t)$ projects the 4D coordinate onto the respective plane, and $\\pi$ performs bilinear interpolation on the voxel features at each point. The features of all planes are combined using the Hadamard product.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, two MLP decoders $D_{\\mathcal{X}}, D_{c}$ are utilized to predict the point motion $\\Delta \\mathcal{X}_t$ and color change $\\Delta \\mathcal{C}_t$ as $\\Delta \\mathcal{X}_t = D_{\\mathcal{X}}(f(x,y,z,t))$ and $\\Delta \\mathcal{C}_t = D_{\\mathcal{C}}(f(x,y,z,t))$ , respectively. The deformed 4D Gaussians are formulated as: $\\mathcal{G}_t = \\{\\mathcal{X} + \\Delta \\mathcal{X}_t, \\mathcal{C} + \\Delta \\mathcal{C}_t, S, R, o\\}$ .", + "bbox": [ + 511, + 152, + 903, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To allow the HexPlane feature to learn point motion from the predicted pose, we define the motion loss as:", + "bbox": [ + 511, + 228, + 903, + 258 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m o t i o n}} = \\operatorname {a v g} _ {\\mathcal {X} \\in \\mathcal {O}} | \\Delta \\mathcal {X} _ {t} - (\\mathbf {T} _ {t} \\mathcal {X} - \\mathcal {X}) |, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 272, + 903, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{X}$ is the center position of a Gaussian in object $\\mathcal{O}$ and avg is the average operator. This loss encourages the predicted point motion $\\Delta \\mathcal{X}_t$ to align with the predicted pose. We apply this loss only for the first $40\\%$ iterations to provide a strong initial motion prior. The loss is subsequently removed, allowing the network to adjust and potentially compensate for pose errors and detection failures.", + "bbox": [ + 511, + 301, + 903, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following [74], we adopt isotropic Gaussian marbles for dynamic points to reduce degrees of freedom and simplify optimization. In this setup, the rotation of each Gaussian is represented by the identity matrix with identical scales across all three dimensions. The spherical harmonics coefficients are also limited to three dimensions, providing view-consistent color. This approach ensures that all deformations are captured purely by point motion and color changes, enforcing strong constraints and enhancing robustness in novel view synthesis.", + "bbox": [ + 511, + 407, + 905, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ours vs. S3Gaussian. Although S3Gaussian [32] also uses Hexplane representation [7] and the deformation network from 4DGS [81], our design serves a different purpose: S3Gaussian learns point motion without extra supervision, while we use 4DGS to refine motion and address detection failures. Consequently, S3Gaussian becomes susceptible to unsatisfactory results with rapid car motion due to the lack of explicit guidance. In contrast, our approach enables high-quality reconstruction of dynamic objects.", + "bbox": [ + 511, + 564, + 905, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Optimization Objective", + "text_level": 1, + "bbox": [ + 511, + 710, + 730, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Besides the motion loss introduced above, the loss function comprises five components to collectively optimize the scene representation and the point motion. The primary loss $\\mathcal{L}_{\\mathrm{rgb}}$ is an L1 loss measuring the photometric difference between rendered and ground truth images while $\\mathcal{L}_{\\mathrm{ssim}}$ evaluates their structural similarity. The L1 loss between the rendered depth map and the depth estimated from the LiDAR point cloud $\\mathcal{L}_{\\mathrm{depth}}$ is used to supervise the Gaussian positions. Following K-Planes [70], a grid-based total variation loss $\\mathcal{L}_{\\mathrm{tv}}$ is introduced. Recognizing that the color of most dynamic points in the scene are unchanged, a L1 regulariza", + "bbox": [ + 511, + 734, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25726", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/82a6f2241af8d1eb636ed0cee16d502821d994a96a88d232adf10d04ccad6802.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TaskScene ReconstructionNovel View Synthesis
MethodExtra InputPSNR↑SSIM↑LPIPS↓DPSNR↑DSSIM↑PSNR↑SSIM↑LPIPS↓DPSNR↑DSSIM↑
3DGS [37]N/A25.770.8330.16020.260.60424.460.8020.17018.120.521
EmerNeRF [90]N/A28.160.8060.22824.320.68225.140.7470.31323.490.660
S3Gaussian [32]N/A31.350.9110.10626.020.78327.440.8570.13722.920.680
MARS [85]GT pose28.240.8660.25223.370.70126.610.7960.30522.210.697
StreetGS [89]3D tracker29.170.8730.13827.780.81826.980.8380.14924.620.742
Ours2D tracker32.560.9360.05928.510.86828.850.8670.08825.580.779
", + "bbox": [ + 91, + 88, + 906, + 244 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison with state-of-the-art methods on Waymo-NOTR dataset. StreetGS represents Street Gaussian [89]. The best and the second best results are denoted by pink and blue.", + "bbox": [ + 89, + 257, + 906, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tion loss $\\mathcal{L}_{\\mathrm{color - reg}}$ is applied to the deformation network by minimizing $\\Delta C$ .", + "bbox": [ + 89, + 314, + 480, + 344 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The total loss function is thus defined as:", + "bbox": [ + 109, + 345, + 382, + 359 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\lambda_ {\\mathrm {r g b}} \\mathcal {L} _ {\\mathrm {r g b}} + \\lambda_ {\\text {d e p t h}} \\mathcal {L} _ {\\text {d e p t h}} + \\lambda_ {\\text {s s i m}} \\mathcal {L} _ {\\text {s s i m}} + \\lambda_ {\\mathrm {t v}} \\mathcal {L} _ {\\mathrm {t v}} \\tag {8} \\\\ + \\lambda_ {\\text {c o l o r - r e g}} \\mathcal {L} _ {\\text {c o l o r - r e g}} + \\lambda_ {\\text {m o t i o n}} \\mathcal {L} _ {\\text {m o t i o n}}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 367, + 480, + 404 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the weights are assigned as follows: $\\lambda_{\\mathrm{rgb}} = 1.0$ $\\lambda_{\\mathrm{depth}} = 1.0$ $\\lambda_{\\mathrm{ssim}} = 0.1$ $\\lambda_{\\mathrm{tv}} = 0.1$ $\\lambda_{\\mathrm{color-reg}} = 0.01$ and $\\lambda_{\\mathrm{motion}} = 1$", + "bbox": [ + 89, + 414, + 482, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The optimization process is divided into two stages. In the first stage, we render using only static Gaussians and supervise only the static regions of the images over 20,000 iterations. This phase focuses on reconstructing the static background and stabilizes the training process. In the second stage, we train all Gaussians on the whole images for 50,000 iterations.", + "bbox": [ + 89, + 460, + 483, + 565 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 578, + 223, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 603, + 279, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We leverage Waymo-NOTR dataset [76, 90] and KITTI benchmark [23, 24]. The NeRF On-The-Road (NOTR) dataset, introduced by EmerNeRF [90], is a subset of the Waymo Open dataset [76] and includes diverse challenging driving scenarios, such as high-speed, exposure mismatch, and various weather conditions. We use the dynamic32 subset, consisting of 32 dynamic scenes, for evaluating dynamic reconstruction. Following EmerNeRF's setup [90], we use three frontal camera images resized to $960 \\times 640$ . For scene reconstruction, all image frames are used for training and evaluation; for novel view synthesis, every 10th time step is excluded for evaluation [90]. For the KITTI dataset, we follow the setup of MARS [85] to use $75\\%$ or $50\\%$ of the images for training with every 4th or every 2nd frame held out for testing, respectively.", + "bbox": [ + 89, + 625, + 482, + 851 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We use the Adam optimizer [39] with the same learning rate schedule as in 3DGS [37]. For long-sequence scene reconstruction, we follow S3Gaussian", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[32] and segment the scene into multiple clips. The multi-resolution HexPlane encoder has a base resolution of 64, upsampled by factors of 2 and 4 as in [81], and other hyperparameters match those in 3DGS [37]. As LiDAR points lack data for the sky region, we add a plane of points above the maximum height of the scene to represent it following [38]. All experiments are run on a single NVIDIA RTX 3090 GPU, with training taking about 2 hours per video clip and inference speed at 100 FPS at $960 \\times 640$ resolution.", + "bbox": [ + 511, + 314, + 906, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline Methods. We compare our approach against state-of-the-art methods, including NeRF-based methods (MARS [85], NSG [65], EmerNeRF [90]) and Gaussian-based methods (3DGS [37], StreetGaussian [89], S3Gaussian [32]). To ensure fair comparisons, we apply LiDAR point cloud initialization and add depth regularization to 3DGS. For the Waymo-NOTR dataset, we borrow results of MARS [85] and EmerNeRF [90] from S3Gaussian [32]. We evaluate the 3D-tracker-based method, Street Gaussians [89], using detection results from VoxelNext [11] and associate the detections with SimpleTrack [66]. Our goal is to assess the generalization ability of 3D-tracker-based methods in scenarios without ground truth tracking labels. To emulate a real-world scenario where a 3D tracker trained on a public benchmark dataset is applied to a novel environment, we use pretrained weights from nuScenes [6]—one of the largest and most diverse 3D datasets—and evaluate the model on the Waymo-NOTR dataset. Since NOTR is part of the training set for the Waymo perception task, using Waymo pretrained weights would introduce domain overlap and bias the evaluation, making the nuScenes weights a fairer choice. For a fair comparison, we use pretrained 2D tracker weights from [84] for our method, which have not been trained on the Waymo or KITTI datasets. Please refer to the supplementary material for discussion of the choice of 3D trackers. For the KITTI dataset, we borrow the results of all other methods from Street Gaussians. Please note that Street Gaussians leverages GT tracking annotations on KITTI dataset.", + "bbox": [ + 511, + 463, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25727", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/adb78bd5aae696541ac6b36c87113c0ef7399b603f4f1c40e6ccadaf9b589bde.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 114, + 107, + 302, + 204 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d2c05fd3d944350e55146f6f7b856b26cdfbb091f6e06300cbe67c6318694edc.jpg", + "image_caption": [ + "S3Gaussian" + ], + "image_footnote": [], + "bbox": [ + 305, + 107, + 493, + 204 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d59d369509dd33f47f225170a24d9ed605933171dcfcb29c6e367571db7f671d.jpg", + "image_caption": [ + "Street Gaussian" + ], + "image_footnote": [], + "bbox": [ + 500, + 107, + 687, + 204 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/81b8cdadfe611141909cc8417c57d85ba2b0f10ab5df4d0b89712c8784c8cdbf.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 694, + 107, + 883, + 204 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5df529ecf0004a6a96c2ad0dfa62667653d39dc362b50ac96b5a6b16a04e5888.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 114, + 205, + 302, + 304 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4924e0c1f01051557eaa8e9143c81d0219a7f3c10f565c3bf2e65a5bcea0b94e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 305, + 205, + 493, + 304 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8b0e8a9311ede3f05844cd7ca7aa62a2b0e7b1fdcdd35ca07286b0b96f9dabac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 205, + 687, + 304 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6939d243ac553a629ef06e0efd4763649494891d49911fe4b09ce46bafca7ac1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 205, + 883, + 304 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9605e61606317d5c9f7d250fbda8f37b42cd7fed48d3857f46ce00870f1cd617.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 114, + 306, + 302, + 404 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cc6dfeb6b8c089369b408ccbd18a17f91dd2c1b2debcf9db41116ea2d8885c93.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 305, + 306, + 493, + 404 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/279869d5a6455c6ef8c58ab0617acf743196665c4b9bc249e21ab5dfbeaf002b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 306, + 687, + 404 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9888d28bf48e41917c575bec7dba41b21413cc1366248ce6b1becb2f13ea4ff2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 306, + 883, + 404 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b066907af73fbc81b4b4cc5c84def1915148ce1ad7efc96ad6d647164c72ac6c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 114, + 406, + 302, + 505 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d7551168474300e162ab28a1904cefcbe6d0d01ccf6f5e50f087df9ae7f777fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 305, + 406, + 493, + 505 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c47fc0aea547c28ac8c75c279e4c25ac091b2c7511cde40fe6144e8d3948ce3c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 406, + 687, + 505 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6e9ac1194a1144189f517d5e3b60d1bbaaa1995f0e562474fa51be0620ef1556.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 406, + 883, + 505 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/da6ba43bf89505272eb6fe1987d545187cce8a47bf812aa092b1e31ad4152e97.jpg", + "image_caption": [ + "Figure 3. Qualitative comparison of novel view synthesis on NOTR dataset. Best viewed with zoom." + ], + "image_footnote": [], + "bbox": [ + 114, + 506, + 302, + 604 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b82259fa799bcb919099bbae0b7f78f794afc6264b72b788c38f760334efa06a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 305, + 506, + 493, + 604 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/030639fa14ec1fd5122fbc19acabdc8bb5d4284dc864e704b73b195493025a7b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 506, + 687, + 604 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/90d5cd84366e64bfa75e381301ea0d16a1cb7b1b6cf05cfc903ef4f635ef1eb8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 506, + 883, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics. We leverage peak signal-to-noise ratio (PSNR), structural similarity index (SSIM) and learned perceptual image patch similarity (LPIPS) to evaluate the rendering quality. Additionally, following EmerNeRF [90], we apply DPSNR and DSSIM metrics to dynamic objects by projecting their ground truth 3D bounding boxes onto the 2D image plane and calculating pixel loss within these regions.", + "bbox": [ + 89, + 659, + 482, + 765 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparisons with the State-of-the-art", + "text_level": 1, + "bbox": [ + 89, + 772, + 419, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On the NOTR dataset, our method outperforms all competitors across every metric, as shown in Table 1. Our approach sets a new state-of-the-art in both the scene reconstruction and novel view synthesis tasks. Specifically, our method outperforms S3Gaussian [32] by $2.66\\mathrm{dB}$ in DPSNR and 0.099 in DSSIM for novel view synthesis. It also surpasses Street Gaussians [89] by $1.87\\mathrm{dB}$ in PSNR, mainly due", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to the limited generalization of 3D trackers used in Street Gaussians. Despite optimizing object poses, Street Gaussians struggles with large pose errors and detection failures, leading to inferior performance. Additionally, we outperform NeRF-based methods like EmerNeRF and MARS.", + "bbox": [ + 511, + 659, + 906, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present qualitative comparison with S3Gaussian [32] and Street Gaussians [89] in Fig. 3. Results from S3Gaussian show that using 4DGS without explicit motion guidance results in weaker performance when handling moving vehicles (e.g. row 2, 3, 5). Street Gaussians suffers from tracking errors of 3D trackers, leading to erroneous reconstructions (e.g. row 1, 3) or entirely missed objects (e.g. rows 2, 4, and 5). In contrast, our approach performs robustly across diverse scenarios owing to our tracking strategy leveraging 2D foundation model and robust motion", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25728", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2983810f71fb79df77a48f2f5e2c62dc996667575291f1351fc81ef745cd26d9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
KITTI-75%KITTI-50%
MethodExtra InputPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
3DGS [37]N/A19.190.7370.17219.230.7390.174
NSG [65]GT pose21.530.6730.25421.260.6590.266
MARS [85]GT pose24.230.8450.16024.000.8010.164
Street Gussians [89]GT pose25.790.8440.08125.520.8410.084
Ours2D tracker [84]25.490.8890.06325.110.8770.067
", + "bbox": [ + 151, + 88, + 846, + 227 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0215615f9d01b72de2778309eab04801b5c80074336643a07c779232d00dcc60.jpg", + "table_caption": [ + "Table 2. Comparison with state-of-the-art methods on KITTI dataset. The best and the second best results are denoted by pink and blue." + ], + "table_footnote": [], + "table_body": "
MethodPSNRSSIMDPSNRDSSIM
Aw/o track25.250.76121.710.662
Bw/o Lmotion28.390.84324.540.719
Cw/o iso.28.580.84725.350.765
D3D tracker28.020.83524.610.746
ED w/o Lmotion27.540.82123.530.703
FGT pose28.980.87325.710.787
GUNINEXT28.760.86425.430.772
HOurs28.850.86725.580.779
", + "bbox": [ + 96, + 289, + 478, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation study of Novel View Synthesis on NOTR dataset.", + "bbox": [ + 89, + 474, + 482, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "learning module. Please refer to the supplementary material for additional video results and editing examples.", + "bbox": [ + 89, + 530, + 482, + 560 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results on the KITTI dataset are presented in Table 2. Our method outperforms the NeRF-based NSG and MARS across all metrics. In the KITTI-75% setting, our approach achieves a 0.045 higher SSIM and a 0.018 lower LPIPS compared to Street Gaussians, although its PSNR is 0.3 dB lower. A similar trend is observed in the KITTI-50% setting. This lower PSNR is primarily because Street Gaussians leverages labor-intensive ground truth tracking annotations, whereas our method uses a generalized 2D tracker—yet still attains comparable performance.", + "bbox": [ + 88, + 560, + 482, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 720, + 254, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct an ablation study on the NOTR dataset and the results are presented in Tab. 3.", + "bbox": [ + 89, + 742, + 482, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of object tracking module. In Tab. 3 (A), we model dynamic points using 4DGS [81] without performing any form of tracking, and observe significant drops across all metrics. This experiment highlights that the sole use of 4DGS does not provide sufficient accuracy for effective motion modeling in autonomous driving scenarios.", + "bbox": [ + 89, + 776, + 482, + 867 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of motion learning strategy. In Tab. 3 (B), we use the reconstruction of the tracking module for initialization", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "but remove the motion loss introduced in Eq. 7, resulting in a $1.04\\mathrm{dB}$ drop in DPSNR. This result highlights the importance of our motion learning strategy. A comparison with rigid-transformation-based motion modeling [89] is provided in the supplementary material.", + "bbox": [ + 511, + 292, + 903, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of isotropic Gaussian marbles. In Tab. 3 (C), we substitute isotropic Gaussian marbles with the original anisotropic Gaussian ellipsoids for dynamic points, which leads to a decrease in DPSNR and DSSIM.", + "bbox": [ + 511, + 373, + 903, + 434 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3D tracker / GT pose vs. Our object tracking module. In Tab. 3 (D), we use the 3D tracker [11, 66] to reconstruct dynamic objects and supervise motion learning, resulting in a 0.97 dB DPSNR decrease. Removing motion loss leads to a further 1.08 dB drop (Tab. 3 (E)), showing that our motion learning strategy compensates for tracking errors, though these errors still impact novel view synthesis, emphasizing the need for a more generalizable 2D tracker. Using ground-truth trajectories (Tab. 3 (F)) shows only a small gap, confirming the effectiveness of our 2D tracker in motion learning. See supplementary for tracking error analysis.", + "bbox": [ + 511, + 439, + 906, + 606 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Choice of 2D tracker. To demonstrate the robustness of our method with respect to the choice of 2D tracker, we employ UNINEXT [88] for 2D tracking. As shown in Table 3 (G), our method maintains stable performance.", + "bbox": [ + 511, + 611, + 903, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 691, + 633, + 707 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce a novel framework for robust dynamic 3D street scene reconstruction that eliminates the reliance on 3D object trackers. Addressing the generalization limitations of 3D trackers, we propose a robust object tracking strategy based on a 2D foundation model. Our framework also features a motion learning module within an implicit feature space to handle inevitable tracking errors by autonomously refining pose inaccuracies and recovering missed detections. Experiments on the Waymo-NOTR and KITTI datasets demonstrate its adaptability and superior performance. Limitations and failure cases are discussed in the supplementary material.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25729", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Key R&D Program of China under Grant 2018AAA0102801, PCL-CMCC Foundation for Science and Innovation Grant No. 2024ZY1C0040, and the National Research Foundation, Singapore, under its NRF-Investigatorship Programme (Award ID. NRF-NRFI09-0008).", + "bbox": [ + 89, + 114, + 483, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 220, + 187, + 237 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5855–5864, 2021. 2", + "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470-5479, 2022.", + "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 2", + "[4] Mk Bashar, Samia Islam, Kashifa Kawaakib Hussain, Md Bakhtiar Hasan, ABM Rahman, and Md Hasanul Kabir. Multiple object tracking in recent times: A literature review. arXiv preprint arXiv:2209.04796, 2022. 3", + "[5] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In 2016 IEEE international conference on image processing (ICIP), pages 3464-3468. IEEE, 2016. 3", + "[6] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 3, 6", + "[7] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. 2, 3, 4, 5", + "[8] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2", + "[9] Zhengping Che, Guangyu Li, Tracy Li, Bo Jiang, Xuefeng Shi, Xinsheng Zhang, Ying Lu, Guobin Wu, Yan Liu, and Jieping Ye. D2-city: a large-scale dashcam video dataset of diverse traffic scenarios. arXiv preprint arXiv:1904.01975, 2019.3", + "[10] Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene" + ], + "bbox": [ + 99, + 247, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "reconstruction and real-time rendering. arXiv preprint arXiv:2311.18561, 2023. 3", + "[11] Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21674-21683, 2023. 1, 3, 6, 8", + "[12] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 2, 3, 4", + "[13] Jie Cheng, Yingbing Chen, Qingwen Zhang, Lu Gan, Chengju Liu, and Ming Liu. Real-time trajectory planning for autonomous driving with gaussian process and incremental refinement. In 2022 International Conference on Robotics and Automation (ICRA), pages 8999-9005. IEEE, 2022. 1", + "[14] Jie Cheng, Xiaodong Mei, and Ming Liu. Forecast-mae: Self-supervised pre-training for motion forecasting with masked autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8679-8689, 2023. 1", + "[15] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. ACM Transactions on Graphics (ToG), 34(4):1-13, 2015. 3", + "[16] Daniel Dauner, Marcel Hallgarten, Andreas Geiger, and Kashyap Chitta. Parting with misconceptions about learning-based vehicle motion planning. In Conference on Robot Learning, pages 1268-1281. PMLR, 2023. 1", + "[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 3", + "[18] Yan Di, Ruida Zhang, Zhiqiang Lou, Fabian Manhardt, Xiangyang Ji, Nassir Navab, and Federico Tombari. Gpvpose: Category-level object pose estimation via geometry-guided point-wise voting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6781-6791, 2022. 3", + "[19] Shuxiao Ding, Eike Rehder, Lukas Schneider, Marius Cordts, and Juergen Gall. 3dmotformer: Graph transformer for online 3d multi-object tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9784-9794, 2023. 3", + "[20] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 3", + "[21] George Eskandar. An empirical study of the generalization ability of lidar 3d object detectors to unseen domains. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23815-23825, 2024. 1, 3", + "[22] Tobias Fischer, Jonas Kulhanek, Samuel Rota Bulo, Lorenzo Porzi, Marc Pollefeys, and Peter Kontschieder." + ], + "bbox": [ + 522, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25730", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Dynamic 3d gaussian fields for urban areas. arXiv preprint arXiv:2406.03175, 2024. 1", + "[23] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 2, 3, 6", + "[24] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 6", + "[25] Junru Gu, Chenxu Hu, Tianyuan Zhang, Xuanyao Chen, Yilun Wang, Yue Wang, and Hang Zhao. Vip3d: End-to-end visual trajectory prediction via 3d agent queries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5496-5506, 2023. 1", + "[26] Kaiwen Guo, Feng Xu, Yangang Wang, Yebin Liu, and Qionghai Dai. Robust non-rigid motion tracking and surface reconstruction using 10 regularization. In Proceedings of the IEEE International Conference on Computer Vision, pages 3083-3091, 2015. 3", + "[27] Kaiwen Guo, Peter Lincoln, Philip Davidson, Jay Busch, Xueming Yu, Matt Whalen, Geoff Harvey, Sergio Orts-Escolano, Rohit Pandey, Jason Dourgarian, et al. The relightables: Volumetric performance capture of humans with realistic relighting. ACM Transactions on Graphics (ToG), 38(6):1-19, 2019. 3", + "[28] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 2", + "[29] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 1", + "[30] Tao Hu, Tao Yu, Zerong Zheng, He Zhang, Yebin Liu, and Matthias Zwicker. Hvtr: Hybrid volumetric-textural rendering for human avatars. In 2022 International Conference on 3D Vision (3DV), pages 197-208. IEEE, 2022. 3", + "[31] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1", + "[32] Nan Huang, Xiaobao Wei, Wenzhao Zheng, Pengju An, Ming Lu, Wei Zhan, Masayoshi Tomizuka, Kurt Keutzer, and Shanghang Zhang. S3gaussian: Self-supervised street gaussians for autonomous driving. arXiv preprint arXiv:2405.20323, 2024. 2, 3, 5, 6, 7", + "[33] Yanjun Huang, Jiatong Du, Ziru Yang, Zewei Zhou, Lin Zhang, and Hong Chen. A survey on trajectory-prediction methods for autonomous driving. IEEE Transactions on Intelligent Vehicles, 7(3):652-674, 2022. 1", + "[34] Ziqin Huang, Gu Wang, Chenyangguang Zhang, Ruida Zhang, Xiu Li, and Xiangyang Ji. Givepose: Gradual intraclass variation elimination for rgb-based category-level object pose estimation. In Proceedings of the Computer Vision" + ], + "bbox": [ + 99, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Pattern Recognition Conference, pages 22055-22066, 2025.3", + "[35] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8350, 2023. 1", + "[36] Lesole Kalake, Wanggen Wan, and Li Hou. Analysis based on recent deep learning approaches applied in real-time multi-object tracking: a review. IEEE Access, 9:32650-32671, 2021. 3", + "[37] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 3, 4, 6, 8", + "[38] Mustafa Khan, Hamidreza Fazlali, Dhruv Sharma, Tongtong Cao, Dongfeng Bai, Yuan Ren, and Bingbing Liu. Autosplat: Constrained gaussian splatting for autonomous driving scene reconstruction. arXiv preprint arXiv:2407.02598, 2024. 1, 3, 6", + "[39] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[40] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 2, 3", + "[41] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision, 128(7): 1956-1981, 2020. 2, 3", + "[42] Jinyu Li, Chenxu Luo, and Xiaodong Yang. Pillarnext: Rethinking network designs for 3d object detection in lidar point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17567-17576, 2023. 1, 3", + "[43] Wei Li, CW Pan, Rong Zhang, JP Ren, YX Ma, Jin Fang, FL Yan, QC Geng, XY Huang, HJ Gong, et al. Aads: Augmented autonomous driving simulation using data-driven algorithms. Science robotics, 4(28):eaaw0863, 2019. 3", + "[44] Yanyan Li, Chenyu Lyu, Yan Di, Guangyao Zhai, Gim Hee Lee, and Federico Tombari. Geogaussian: Geometry-aware gaussian splatting for scene rendering. In European Conference on Computer Vision, pages 441-457. Springer, 2024. 2", + "[45] Zhong Li, Yu Ji, Wei Yang, Jinwei Ye, and Jingyi Yu. Robust 3d human motion reconstruction via dynamic template construction. In 2017 International Conference on 3D Vision (3DV), pages 496-505. IEEE, 2017. 3", + "[46] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In Proceedings of" + ], + "bbox": [ + 522, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25731", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14864-14873, 2024. 1", + "[47] Ming Liang, Bin Yang, Wenyuan Zeng, Yun Chen, Rui Hu, Sergio Casas, and Raquel Urtasun. Pnpnet: End-to-end perception and prediction with tracking in the loop. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11553-11562, 2020. 1", + "[48] Yiqing Liang, Numair Khan, Zhengqin Li, Thu Nguyen-Phuoc, Douglas Lanman, James Tompkin, and Lei Xiao. Gaufre: Gaussian deformation fields for real-time dynamic novel view synthesis. arXiv preprint arXiv:2312.11458, 2023. 3", + "[49] Yiyi Liao, Jun Xie, and Andreas Geiger. Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3292-3310, 2022. 2, 3", + "[50] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 2", + "[51] Xingyu Liu, Gu Wang, Yi Li, and Xiangyang Ji. CATRE: iterative point clouds alignment for category-level object pose refinement. In European Conference on Computer Vision, 2022. 3", + "[52] Xingyu Liu, Chenyangguang Zhang, Gu Wang, Ruida Zhang, and Xiangyang Ji. Rasim: A range-aware high-fidelity rgb-d data simulation pipeline for real-world applications. In 2024 IEEE international conference on robotics and automation (ICRA), pages 17057-17064. IEEE, 2024. 3", + "[53] Xingyu Liu, Yingyue Li, Chengxi Li, Gu Wang, Chenyangguang Zhang, Ziqin Huang, and Xiangyang Ji. Gfreedet: Exploiting gaussian splatting and foundation models for model-free unseen object detection in the bop challenge 2024. IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop, 2025. 3", + "[54] Xingyu Liu, Gu Wang, Ruida Zhang, Chenyangguang Zhang, Federico Tombari, and Xiangyang Ji. UNOPose: Unseen object pose estimation with an unposed rgb-d reference image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 3", + "[55] Xingyu Liu, Ruida Zhang, Chenyangguang Zhang, Gu Wang, Jiwen Tang, Zhigang Li, and Xiangyang Ji. Gdnpp: A geometry-guided and fully learning-based object pose estimator. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 3", + "[56] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 1, 3", + "[57] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 3" + ], + "bbox": [ + 99, + 92, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[58] Wenhan Luo, Junliang Xing, Anton Milan, Xiaoqin Zhang, Wei Liu, and Tae-Kyun Kim. Multiple object tracking: A literature review. Artificial intelligence, 293:103448, 2021. 3", + "[59] Jiageng Mao, Minzhe Niu, Chenhan Jiang, Hanxue Liang, Jingheng Chen, Xiaodan Liang, Yamin Li, Chaoqiang Ye, Wei Zhang, Zhenguo Li, et al. One million scenes for autonomous driving: Once dataset. arXiv preprint arXiv:2106.11037, 2021.3", + "[60] Jiageng Mao, Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. 3d object detection for autonomous driving: A comprehensive survey. International Journal of Computer Vision, 131(8):1909-1963, 2023. 3", + "[61] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3", + "[62] M Minderer, A Gritsenko, A Stone, M Neumann, D Weissenborn, A Dosovitskiy, A Mahendran, A Arnab, M Dehghani, Z Shen, et al. Simple open-vocabulary object detection with vision transformers. arxiv 2022. arXiv preprint arXiv:2205.06230, 2, 2022. 2", + "[63] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 2", + "[64] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 3", + "[65] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2856-2865, 2021. 1, 3, 6, 8", + "[66] Ziqi Pang, Zhichao Li, and Naiyan Wang. Simpletrack: Understanding and rethinking 3d multi-object tracking. In European Conference on Computer Vision, pages 680-696. Springer, 2022. 3, 6, 8", + "[67] Chensheng Peng, Chengwei Zhang, Yixiao Wang, Chenfeng Xu, Yichen Xie, Wenzhao Zheng, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. Desire-gs: 4d street gaussians for static-dynamic decomposition and surface reconstruction for urban driving scenes. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 6782-6791, 2025. 3", + "[68] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3", + "[69] Jules Sanchez, Jean-Emmanuel Deschaud, and François Goulette. Domain generalization of 3d semantic segmentation in autonomous driving. In Proceedings of the" + ], + "bbox": [ + 522, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "25732", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF International Conference on Computer Vision, pages 18077-18087, 2023. 3", + "[70] Sara Fridovich-Keil and Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In CVPR, 2023. 5", + "[71] Shital Shah, Debadeepta Dey, Chris Lovett, and Ashish Kapoor. Airsim: High-fidelity visual and physical simulation for autonomous vehicles. In Field and Service Robotics: Results of the 11th International Conference, pages 621-635. Springer, 2018. 3", + "[72] Yeji Song, Chaerin Kong, Seoyoung Lee, Nojun Kwak, and Joonseok Lee. Towards efficient neural scene graphs by learning consistency fields. arXiv preprint arXiv:2210.04127, 2022. 3", + "[73] Louis Soum-Fontez, Jean-Emmanuel Deschaud, and François Goulette. Mdt3d: Multi-dataset training for lidar 3d object detection generalization. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5765-5772. IEEE, 2023. 1, 3", + "[74] Colton Stearns, Adam Harley, Mikaela Uy, Florian Dubost, Federico Tombari, Gordon Wetzstein, and Leonidas Guibas. Dynamic gaussian marbles for novel view synthesis of casual monocular videos. arXiv preprint arXiv:2406.18717, 2024. 3, 4, 5", + "[75] Zhuo Su, Lan Xu, Zerong Zheng, Tao Yu, Yebin Liu, and Lu Fang. Robustfusion: Human volumetric capture with data-driven visual cues using a rgbd camera. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 246-264. Springer, 2020. 3", + "[76] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 2, 3, 6", + "[77] Shanlin Sun, Bingbing Zhuang, Ziyu Jiang, Buyu Liu, Xiaohui Xie, and Manmohan Chandraker. Lidarf: Delving into lidar for neural radiance field on street scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19563-19572, 2024. 3", + "[78] Haithem Turki, Jason Y Zhang, Francesco Ferroni, and Deva Ramanan. Suds: Scalable urban dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12375-12385, 2023. 3", + "[79] Yan Wang, Xiangyu Chen, Yurong You, Li Erran Li, Bharath Hariharan, Mark Campbell, Kilian Q Weinberger, and Wei-Lun Chao. Train in germany, test in the usa: Making 3d object detectors generalize. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11713-11723, 2020. 2", + "[80] Yu Wang, Shaohua Wang, Yicheng Li, and Mingchun Liu. A comprehensive review of 3d object detection in autonomous driving: Technological advances and future directions. arXiv preprint arXiv:2408.16530, 2024. 3" + ], + "bbox": [ + 99, + 92, + 482, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[81] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20310-20320, 2024. 2, 3, 4, 5, 6, 8", + "[82] Hai Wu, Wenkai Han, Chenglu Wen, Xin Li, and Cheng Wang. 3d multi-object tracking in point clouds based on prediction confidence-guided data association. IEEE Transactions on Intelligent Transportation Systems, 23(6):5668-5677, 2021. 2", + "[83] Hai Wu, Jinhao Deng, Chenglu Wen, Xin Li, Cheng Wang, and Jonathan Li. Casa: A cascade attention network for 3-d object detection from lidar point clouds. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. 2", + "[84] Junfeng Wu, Yi Jiang, Qihao Liu, Zehuan Yuan, Xiang Bai, and Song Bai. General object foundation model for images and videos at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3783-3795, 2024. 1, 2, 3, 4, 6, 8", + "[85] Zirui Wu, Tianyu Liu, Liyi Luo, Zhide Zhong, Jianteng Chen, Hongmin Xiao, Chao Hou, Haozhe Lou, Yuantao Chen, Runyi Yang, et al. Mars: An instance-aware, modular and realistic simulator for autonomous driving. In CAAI International Conference on Artificial Intelligence, pages 3-15. Springer, 2023. 1, 3, 6, 8", + "[86] Pengchuan Xiao, Zhenlei Shao, Steven Hao, Zishuo Zhang, Xiaolin Chai, Judy Jiao, Zesong Li, Jian Wu, Kai Sun, Kun Jiang, et al. Pandaset: Advanced sensor suite dataset for autonomous driving. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 3095-3101. IEEE, 2021. 2", + "[87] Ziyang Xie, Junge Zhang, Wenye Li, Feihu Zhang, and Li Zhang. S-nerf: Neural radiance fields for street views. arXiv preprint arXiv:2303.00749, 2023. 3", + "[88] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15325-15336, 2023. 1, 8", + "[89] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians: Modeling dynamic urban scenes with gaussian splatting. In ECCV, 2024. 1, 2, 3, 6, 7, 8", + "[90] Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Seung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, et al. Emernerf: Emergent spatial-temporal scene decomposition via self-supervision. arXiv preprint arXiv:2311.02077, 2023. 3, 6, 7", + "[91] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1389-1399, 2023. 3", + "[92] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time photorealistic dynamic scene representation and" + ], + "bbox": [ + 522, + 92, + 906, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "25733", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023.3", + "[93] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20331-20341, 2024. 3", + "[94] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 1, 3", + "[95] Fisher Yu, Wenqi Xian, Yingying Chen, Fangchen Liu, Mike Liao, Vashisht Madhavan, Trevor Darrell, et al. Bdd100k: A diverse driving video database with scalable annotation tooling. arXiv preprint arXiv:1805.04687, 2(5): 6, 2018. 2, 3", + "[96] Jiang-Tian Zhai, Ze Feng, Jinhao Du, Yongqiang Mao, Jiang-Jiang Liu, Zichang Tan, Yifu Zhang, Xiaqing Ye, and Jingdong Wang. Rethinking the open-loop evaluation of end-to-end autonomous driving in nuscenes. arXiv preprint arXiv:2305.10430, 2023. 1", + "[97] Bo Zhang, Jiakang Yuan, Botian Shi, Tao Chen, Yikang Li, and Yu Qiao. Uni3d: A unified baseline for multi-dataset 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9253-9262, 2023. 1, 3", + "[98] Peng Zhang, Xin Li, Liang He, and Xin Lin. 3d multiple object tracking on autonomous driving: A literature review. arXiv preprint arXiv:2309.15411, 2023. 3", + "[99] Ruida Zhang, Yan Di, Zhiqiang Lou, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. Rbp-pose: Residual bounding box projection for category-level pose estimation. In European conference on computer vision, pages 655–672. Springer, 2022. 3", + "[100] Ruida Zhang, Yan Di, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. Ssp-pose: Symmetry-aware shape prior deformation for direct category-level object pose estimation. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7452–7459. IEEE, 2022.", + "[101] Ruida Zhang, Ziqin Huang, Gu Wang, Chenyangguang Zhang, Yan Di, Xingxing Zuo, Jiwen Tang, and Xiangyang Ji. Lapose: Laplacian mixture shape modeling for rgb-based category-level object pose estimation. In European Conference on Computer Vision, pages 467-484. Springer, 2024. 3", + "[102] Ruida Zhang, Chenyangguang Zhang, Yan Di, Fabian Manhardt, Xingyu Liu, Federico Tombari, and Xiangyang Ji. Kp-red: Exploiting semantic keypoints for joint 3d shape retrieval and deformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20540-20550, 2024. 3", + "[103] Xingxuan Zhang, Zekai Xu, Renzhe Xu, Jiashuo Liu, Peng Cui, Weitao Wan, Chong Sun, and Chen Li. Towards domain generalization in object detection. arXiv preprint arXiv:2203.14387, 2022. 2" + ], + "bbox": [ + 93, + 92, + 482, + 898 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[104] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. International journal of computer vision, 129:3069–3087, 2021. 3", + "[105] Hongyu Zhou, Jiahao Shao, Lu Xu, Dongfeng Bai, Weichao Qiu, Bingbing Liu, Yue Wang, Andreas Geiger, and Yiyi Liao. Hugs: Holistic urban 3d scene understanding via gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21336-21345, 2024. 1", + "[106] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21634-21643, 2024. 1, 2, 3" + ], + "bbox": [ + 516, + 92, + 903, + 316 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "25734", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2025/Street Gaussians without 3D Object Tracker/ceea65cc-c6ed-4695-b497-f6fc77f3a4c6_model.json b/2025/Street Gaussians without 3D Object Tracker/ceea65cc-c6ed-4695-b497-f6fc77f3a4c6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4e08d9a108d0c050d064c44ced0f0bf8bf21cfab --- /dev/null +++ b/2025/Street Gaussians without 3D Object Tracker/ceea65cc-c6ed-4695-b497-f6fc77f3a4c6_model.json @@ -0,0 +1,2998 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.274, + 0.131, + 0.724, + 0.154 + ], + "angle": 0, + "content": "Street Gaussians without 3D Object Tracker" + }, + { + "type": "text", + "bbox": [ + 0.177, + 0.18, + 0.82, + 0.235 + ], + "angle": 0, + "content": "Ruida Zhang\\(^{1,2}\\), Chengxi Li\\(^{1}\\), Chenyangguang Zhang\\(^{1}\\), Xingyu Liu\\(^{1}\\), Haili Yuan\\(^{1}\\), Yanyan Li\\(^{2}\\), Xiangyang Ji\\(^{1}\\), Gim Hee Lee\\(^{2}\\) \n\\(^{1}\\)Tsinghua University, \\(^{2}\\)National University of Singapore" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.237, + 0.775, + 0.253 + ], + "angle": 0, + "content": "{zhangrd23@mails, xyji@}.tsinghua.edu.cn, gimhee.lee@nus.edu.sg" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.264, + 0.492, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.264, + 0.902, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.477, + 0.908, + 0.536 + ], + "angle": 0, + "content": "Figure 1. Comparison of 3D tracker-based Street Gaussians [89] (left) and our approach (right). Existing methods heavily rely on object poses, but 3D trackers struggle with limited generalization [21, 73, 97], leading to flaws in novel view synthesis. In contrast, 2D foundation models show better generalization [84, 88]. Our approach leverages a 2D foundation model [84] for object tracking and learns point motion within an implicit feature space to autonomously correct tracking errors, improving robustness across diverse scenes." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.548, + 0.327, + 0.563 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.58, + 0.486, + 0.899 + ], + "angle": 0, + "content": "Realistic scene reconstruction in driving scenarios poses significant challenges due to fast-moving objects. Most existing methods rely on labor-intensive manual labeling of object poses to reconstruct dynamic objects in canonical space and move them based on these poses during rendering. While some approaches attempt to use 3D object trackers to replace manual annotations, the limited generalization of 3D trackers – caused by the scarcity of large-scale 3D datasets – results in inferior reconstructions in real-world settings. In contrast, 2D foundation models demonstrate strong generalization capabilities. To eliminate the reliance on 3D trackers and enhance robustness across diverse environments, we propose a stable object tracking module by leveraging associations from 2D deep trackers within a 3D object fusion strategy. We address inevitable tracking errors by further introducing a motion learning strategy in an implicit feature space that autonomously corrects trajectory errors and recovers missed detections. Experimental results on Waymo-NOTR and KITTI show that our method outperforms existing approaches. Our code will be made publicly available." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.548, + 0.645, + 0.563 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.578, + 0.907, + 0.716 + ], + "angle": 0, + "content": "Modeling dynamic 3D street scenes underpins modern autonomous driving by enabling realistic, controllable simulations for tasks such as perception [11, 42, 56, 94, 105], prediction [25, 33, 47], and motion planning [13, 14, 16]. With the rise of end-to-end autonomous systems that require real-time sensor feedback [29, 31, 35], high-quality scene reconstructions have become essential for closed-loop evaluations [46, 96], particularly to simulate critical corner cases safely and cost-effectively." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Despite extensive efforts in achieving photo-realistic reconstruction of small-scale scenes, the unique challenges posed by the large-scale and highly dynamic nature of driving scenarios complicate effective 3D scene modeling. To address these challenges, most existing methods [22, 38, 65, 85, 106] rely on ground truth vehicle poses to differentiate between static background and moving vehicles. Typically, vehicles are reconstructed in canonical space and subsequently positioned based on known poses during rendering. However, collecting ground-truth poses is labor-intensive, limiting the applicability of these methods to scenes beyond the existing datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25722" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.482, + 0.543 + ], + "angle": 0, + "content": "To eliminate reliance of the systems on ground truth object poses, Street Gaussians [89] instead uses poses generated by 3D object trackers [82, 83]. However, by defining vehicle motion solely through these tracking poses, rendering quality becomes highly dependent on pose accuracy. While Street Gaussians optimizes object poses during training, it struggles with detection failures and large pose errors, as shown in Fig. 1. Unfortunately, 3D trackers often struggle to generalize effectively across different scenarios, primarily due to the scarcity of open-source largescale 3D datasets [79, 103]. Collecting ground-truth poses for 3D tracking is time-consuming and costly, resulting in relatively few open-source datasets for autonomous driving. For example, even widely-used datasets such as Waymo [76], Nuscenes [6], KITTI [23, 49], Pandaset [86] collectively contain fewer than 4,000 annotated scenes, limiting the diversity and scale needed for robust 3D tracking. In contrast, 2D data is far easier to collect and annotate, leading to an abundance of 2D datasets. The BDD100K dataset [95] alone includes annotations for 100,000 driving scenes, and additional large-scale 2D perception datasets, such as COCO [50], LVIS [28], and OpenImages [41], bring the total number of annotated scenes into the millions. This wealth of 2D data has enabled 2D foundation models to achieve strong generalization across a variety of tasks, including embedding extraction [64, 68], detection [8, 62], segmentation [12, 40], and tracking [84]. Nonetheless, how to use these accurate and robust 2D trackers to effectively model dynamic objects in 3D is still an open challenge for street scene reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.55, + 0.482, + 0.789 + ], + "angle": 0, + "content": "In this paper, a novel architecture is designed to achieve high-fidelity novel view synthesis performance in street scenarios for autonomous driving applications. Unlike previous approaches such as Driving Gaussian [106] and Street Gaussian [89] that rely on ground truth vehicle poses or poses predicted by 3D trackers, the first contribution of this work is a stable object tracking module. This module leverages associations from 2D deep trackers within a 3D object fusion framework to enhance robustness and accuracy. Specifically, we integrate 2D tracking outputs with LiDAR data to trace the trajectory of each vehicle in 3D. We then incrementally reconstruct the point cloud of each vehicle frame-by-frame in the canonical space, estimating its pose by aligning successive frames with this canonical model. This approach eliminates the reliance on 3D trackers and enhances robustness across diverse environments." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.796, + 0.482, + 0.9 + ], + "angle": 0, + "content": "Although 2D tracking models demonstrate stronger generalization ability, tracking errors still can exist, especially under adverse conditions such as severe occlusions or distant objects. Since moving the object points solely based on the tracked object pose directly exposes any tracking errors, we aim to go beyond a straightforward reliance on tracked trajectories. Instead, we propose to learn point mo" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.318 + ], + "angle": 0, + "content": "tion from the predicted trajectory, equipping the model to autonomously identify and correct tracking inaccuracies, recover missed detections, and infer motion in new time steps. An implicit representation is essential for this purpose: 1) It enables the model to refine trajectories without being constrained by bounding box tracks, facilitating smoother and more continuous motion that can adaptively respond to changes. 2) An implicit feature space offers versatility to move each point of an object in a different way, recognizing that vehicles are not strictly rigid objects — for example, doors can open or close. This approach makes it possible to capture subtle, dynamic changes within objects, ultimately producing more accurate reconstructions and enhancing the robustness of novel view synthesis in challenging scenarios." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.576 + ], + "angle": 0, + "content": "To this end, we leverage HexPlane representation [7] following 4DGS [81]. HexPlane stores motion-related features by decomposing the 4D spatial-temporal space into six 2D learnable feature planes. A decoder then utilizes these features to predict deformation offsets, dynamically adjusting 3D Gaussians over time. However, 4DGS relies solely on image reconstruction loss for supervision, which is insufficient for handling rapid object motion. When a Gaussian's initial position is far from its current location, its projection on the image may fall outside the object's actual area, preventing gradient propagation and hindering optimization. This issue is evident in S3Gaussian [32], which struggles with moving cars due to the lack of explicit motion supervision. To address this, we introduce a training strategy that supervises learned point motion using the predicted trajectory, providing explicit 3D supervision to guide HexPlane in capturing motion dynamics more accurately." + }, + { + "type": "text", + "bbox": [ + 0.536, + 0.578, + 0.872, + 0.591 + ], + "angle": 0, + "content": "Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.594, + 0.905, + 0.653 + ], + "angle": 0, + "content": "- We eliminate the need of 3D tracker for street scene reconstruction by introducing a robust object tracking module which leverages 2D foundation model, achieving superior generalization ability across diverse scenarios." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.654, + 0.905, + 0.713 + ], + "angle": 0, + "content": "- We introduce a motion learning framework to learn from the predicted trajectory in an implicit feature space, enabling it to automatically correct pose errors and infer motion for novel time steps." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.715, + 0.905, + 0.757 + ], + "angle": 0, + "content": "- We outperform existing methods on Waymo-NOTR and KITTI datasets without relying on ground truth annotations." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.594, + 0.905, + 0.757 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.777, + 0.662, + 0.791 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.803, + 0.869, + 0.818 + ], + "angle": 0, + "content": "2.1. 3D Gaussian Splatting for Dynamic Scene" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.826, + 0.904, + 0.9 + ], + "angle": 0, + "content": "3D Gaussian Splatting (3DGS) [37, 44] has advanced scene reconstruction by enabling high-quality, real-time rendering with 3D Gaussians and efficient splat-based rasterization, reducing computation and parameters compared to NeRF-based methods [1-3, 61, 63] and other representa" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25723" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.318 + ], + "angle": 0, + "content": "tions [15, 26, 27, 30, 45, 75, 102]. While originally designed for static scenes, 3DGS has been adapted for dynamic scenes [48, 53, 57, 74, 81, 92, 93]. Dynamic3DGS [57] directly stores information for each 3D Gaussian at every timestamp, and Yang et al. [92] approximate the spatiotemporal 4D volume by optimizing 4D Gaussian primitives. Deformable-3DGS [93] and GauFRe [48] employ deformation fields to model motions, while 4DGS [81] introduces HexPlane representation [7] to store spatial-temporal features efficiently. Gaussian Marbles [74] leverages isotropic \"marbles\" and a divide-and-conquer trajectory learning algorithm. Despite these advances, challenges persist in high-speed driving scenarios with rapid object motion. To tackle this challenge, we leverage a robust object tracking module to guide the motion learning process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.33, + 0.342, + 0.344 + ], + "angle": 0, + "content": "2.2. Street Scene Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.353, + 0.485, + 0.822 + ], + "angle": 0, + "content": "Existing autonomous driving simulation engines [20, 43, 52, 71] face high manual effort in creating virtual environments and a lack of realism in the generated data. The creation of high-fidelity simulations from driving logs is therefore essential for advancing closed-loop training and testing. Recent works [65, 72, 77, 78, 85, 87, 90, 91] have continuously made improvements to NeRF [61] to model street scenes dynamically. Despite the progress, NeRF-based methods remain computationally expensive and require densely overlapping views. Building on the effective 3DGS approach [37] for scene reconstruction, several 3DGS-based methods [10, 67, 89, 106] have emerged. PVG [10] model dynamic scenarios by using periodic vibration-based temporal dynamics. Driving Gaussian [106] introduces incremental static Gaussians and composite dynamic Gaussian graph. Street Gaussians [89] equips Gaussians with semantic logits, and optimize dynamic parts using tracked poses from 3D tracker. AutoSplat [38] enforces geometric constraints in road and sky regions for multi-view consistency. Among these approaches, most of them requires ground truth object pose [10, 38, 106] or the 3D tracker [89]. However, manual annotation is laborious and the 3D tracker is lack in generalization ability, which limits their applications in diverse scenarios. In contrast, S3Gaussian [32] models dynamics in a self-supervised manner using dynamic Gaussians from 4DGS [81], despite struggling with dynamic object modeling due to the lack of explicit motion supervision. We address this challenge by introducing a robust object tracking strategy based on a 2D foundation model [84] and apply motion supervision from the predicted object trajectory." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.279, + 0.848 + ], + "angle": 0, + "content": "2.3. 2D and 3D trackers" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Multiple Object Tracking (MOT) aims at locating multiple objects in each frame, and establishing correspondences between them across frames within input videos" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.905, + 0.349 + ], + "angle": 0, + "content": "[4, 5, 36, 58, 98]. The reliance of most MOT methods on object detection [4, 19, 58, 66, 98, 104] makes detection accuracy crucial for MOT performance. Recent advances in 3D object detection are promising [11, 18, 34, 42, 51, 54, 56, 60, 80, 94, 99-101], but these models often generalize poorly due to limited and biased datasets [21, 55, 73, 97]. Existing open-source autonomous driving datasets lack scale and show regional or environmental biases, such as vehicle density and weather conditions [6, 23, 59, 76]. The main reason is that building large-scale, well-annotated multimodal datasets is costly [6, 49, 59, 69]. In contrast, 2D image data is easier to capture and annotate, enabling extensive dataset collections [9, 17, 41, 95]. Visual foundation models trained on large-scale 2D data have shown strong generalization [12, 40, 64, 68, 84]. We thus opt to use 2D object tracker [84] to locate the dynamic objects." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.361, + 0.644, + 0.377 + ], + "angle": 0, + "content": "3. Our Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.386, + 0.835, + 0.403 + ], + "angle": 0, + "content": "3.1. Prerequisites: 3D Gaussian Splatting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.408, + 0.906, + 0.5 + ], + "angle": 0, + "content": "3D Gaussian Splatting (3DGS) [37] represents a 3D scene explicitly as a collection of 3D Gaussian primitives. Each Gaussian is defined by its center position \\(\\mathcal{X} \\in \\mathbb{R}^3\\) and covariance \\(\\Sigma \\in \\mathbb{R}^{3 \\times 3}\\). To enforce semi-positive definiteness, the covariance matrix is further factorized into a scaling vector \\(S \\in \\mathbb{R}^3\\) and a rotation matrix \\(R \\in SO(3)\\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.655, + 0.507, + 0.905, + 0.526 + ], + "angle": 0, + "content": "\\[\n\\Sigma = R S S ^ {\\top} R ^ {\\top} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.906, + 0.595 + ], + "angle": 0, + "content": "Additionally, each Gaussian is described with its opacity \\(o \\in \\mathbb{R}\\) and view-dependent color defined by spherical harmonic (SH) coefficient \\(\\mathcal{C} \\in \\mathbb{R}^k\\), where \\(k\\) represents numbers of SH functions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.906, + 0.656 + ], + "angle": 0, + "content": "During rendering of novel views, differential splatting is applied to 3D Gaussians within the camera planes. The blending of \\( N \\) ordered points that overlap a pixel is given by the formula:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.666, + 0.905, + 0.709 + ], + "angle": 0, + "content": "\\[\nC = \\sum_ {i \\in N} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.717, + 0.906, + 0.763 + ], + "angle": 0, + "content": "where \\(\\alpha_{i}\\) and \\(c_{i}\\) represents the opacity and color of the ith splatted Gaussian, which is computed from per-point opacity and SH coefficients (see [37] for details)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.771, + 0.625, + 0.786 + ], + "angle": 0, + "content": "3.2. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.791, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Our method takes as input multi-view images \\(\\mathbf{I}_t^{(j)}\\) from multiple cameras placed around the vehicle, each indexed by time step \\(t\\) and camera index \\(j\\), along with intrinsic \\(\\mathbf{K}^{(j)}\\) and extrinsic \\(\\mathbf{E}_t^{(j)}\\) matrices for each view. Additionally, a top-mounted LiDAR provides 3D point clouds \\(\\mathbf{L}_t\\) for each frame. Using this multi-sensor data, our approach reconstructs the 3D scene and synthesizes novel views from any" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25724" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.087, + 0.903, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.295, + 0.907, + 0.396 + ], + "angle": 0, + "content": "Figure 2. Overview of our method. To overcome the limited generalization of 3D object trackers, we introduce a robust object tracking module based on a 2D object tracker [84]. We integrate 2D tracking with LiDAR data to segment the object's point cloud and incrementally reconstruct the object model in canonical space frame-by-frame. The canonical model is used to estimate object pose \\(\\mathbf{T}_t\\) and serves as the initialization for Gaussian optimization. We model dynamic objects with isotropic Gaussian marble [74] to simplify optimization. Our approach learns point motion within the HexPlane [7] feature space, enabling tracking error correction and missed detection recovery. Decoders \\(D_{\\mathcal{X}}, D_{\\mathcal{C}}\\) predict point motion \\(\\Delta \\mathcal{X}_t\\) and color change \\(\\Delta \\mathcal{C}_t\\) based on the HexPlane features \\(f(x,y,z,t)\\), with the predicted trajectory providing supervision. Finally, novel view synthesis is performed by splatting the deformed Gaussian onto the image plane." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.419, + 0.482, + 0.449 + ], + "angle": 0, + "content": "desired camera pose and time frame, without reliance on 3D object trackers or ground truth object trajectories." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.45, + 0.483, + 0.706 + ], + "angle": 0, + "content": "We present an overview of our method in Fig. 2. Visual images and point clouds are obtained from the sensor setup. For images, static and dynamic components are analyzed first. We employ Mask2Former [12] to segment the scene into static and dynamic parts. Specifically, the dynamic part includes humans (pedestrians, cyclists, etc.) and vehicles (cars, trucks, etc.), and all the other objects are categorized as static. We model the static part as in 3DGS [37] and focus on modeling dynamic objects in the following sections. To enhance robustness, we propose an object tracking strategy by leveraging a robust 2D tracker [84] and the LiDAR point cloud, providing accurate initial point clouds and object trajectories for subsequent Gaussian optimization (cf. Sec 3.3). We leverage an implicit representation HexPlane [7] following 4DGS [81] to learn a continuous and smooth per-point motion based on the object trajectory (cf. Sec 3.4). The optimization objective is described in Sec 3.5." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.713, + 0.475, + 0.729 + ], + "angle": 0, + "content": "3.3. Tracking Cars in 3D with Robust 2D Tracker" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.78 + ], + "angle": 0, + "content": "To improve the robustness of novel view synthesis for diverse scenarios, the first challenge is how to track and build vehicles in 3D based on the 2D tracking results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We employ GLEE [84] to track all vehicles within the 2D image plane, obtaining 2D object trajectories in each camera view. Each trajectory includes the 2D segmentation masks of the object over a time period. To lift 2D trajectories into 3D, we first re-project the LiDAR point clouds onto the image plane and assign points to objects based on their presence within the corresponding 2D segmentation masks \\(\\mathcal{M}\\). This yields a segmented object point clouds for each" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.419, + 0.779, + 0.435 + ], + "angle": 0, + "content": "camera view via the following function:" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.446, + 0.905, + 0.467 + ], + "angle": 0, + "content": "\\[\n\\mathcal {O} _ {t} ^ {(j)} = \\left\\{\\mathbf {P} \\mid \\Pi^ {- 1} \\left(\\mathbf {P}, \\mathbf {K} ^ {(j)}, \\mathbf {E} _ {t} ^ {(j)}\\right) \\in \\mathcal {M}, \\mathbf {P} \\in \\mathbf {L} _ {t} \\right\\}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.479, + 0.907, + 0.675 + ], + "angle": 0, + "content": "where \\( j \\) is the camera index, \\( t \\) is the time index, \\( \\mathbf{P} \\) is a LiDAR point and \\( \\Pi(\\cdot) \\) is the re-projection function. Given the slight misalignment between the LiDAR point cloud and RGB image, points near the edges of segmentation masks often fall outside the object boundaries. To address this issue, we perform outlier removal to improve stability in subsequent reconstruction based on the point distance to the point cloud center. We then associate the same objects across different camera views. Two point clouds in different viewpoints are considered belonging to the same object if they share more than 50 points in one time step. This association process results in a set of associated partial object point cloud in different time steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.811 + ], + "angle": 0, + "content": "After obtaining the 3D partial objects in different time steps, the next step is to associate these partial point cloud into a unified and complete model. We initialize the 3D reconstruction of each object from the first frame which it appears. For simplicity, we assume an object is visible between frames \\(0\\) and \\(T\\) and denote the object point cloud at time \\(t\\) as \\(\\mathcal{O}_t\\). Our goal is to obtain a temporally consistent reconstruction \\(\\mathcal{O}\\) and the object pose in each time frame. This reconstruction is done frame-by-frame incrementally." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.812, + 0.906, + 0.873 + ], + "angle": 0, + "content": "Starting with the initial frame, we add \\(\\mathcal{O}_0\\) to \\(\\mathcal{O}\\). For each subsequent frame, we apply Iterative Closest Point (ICP) for the alignment of \\(\\mathcal{O}_t\\) to \\(\\mathcal{O}\\) by extracting the relative pose \\(\\mathbf{T}_t\\). The overlap between \\(\\mathcal{O}_t\\) and \\(\\mathcal{O}\\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.884, + 0.907, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\tau_ {o v e r l a p} = \\mathcal {O} \\cap \\mathbf {T} _ {t} ^ {- 1} \\mathcal {O} _ {t}, \\quad \\mathbf {T} _ {t} = \\mathrm {I C P} (\\mathcal {O} _ {t}, \\mathcal {O}). \\quad (4)\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "25725" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.258 + ], + "angle": 0, + "content": "If ICP reveals an overlap \\(\\tau_{\\text{overlap}}\\) of more than \\(30\\%\\) with \\(\\mathcal{O}\\), we record the object pose \\(\\mathbf{T}_t\\) and update \\(\\mathcal{O}\\) by integrating \\(\\mathbf{T}_t^{-1}\\mathcal{O}_t\\). If \\(\\tau_{\\text{overlap}}\\) is between \\(10\\%\\) and \\(30\\%\\), we only record the object pose \\(P_t\\) without updating \\(\\mathcal{O}\\). If \\(\\tau_{\\text{overlap}}\\) is lower than \\(10\\%\\), we assume a 2D tracking failure for that frame and discard it to filter out tracking errors. This process is repeated for each frame using the last recorded pose as the initial guess for ICP to accelerate convergence. Through this approach, we reconstruct all objects and retrieve their poses in each frame, providing a robust initialization for subsequent 3D Gaussian optimization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.267, + 0.305, + 0.283 + ], + "angle": 0, + "content": "3.4. Learning Point Motion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.289, + 0.484, + 0.47 + ], + "angle": 0, + "content": "A common approach to model object motion in prior works is directly transforming the object points using the given object pose. However, despite the strong generalization capabilities of 2D trackers, detection failures are inevitable in challenging scenarios, such as when objects are heavily occluded or located at great distances. Relying solely on object poses derived from these erroneous or missing detections can easily lead to rendering failures. Additionally, treating cars as rigid objects does not adequately address corner cases, such as when a car door is open or closed. Furthermore, these methods lack the ability to infer an object motion at arbitrary time stamps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.47, + 0.484, + 0.666 + ], + "angle": 0, + "content": "To address these limitations, we seek for a motion modeling approach to enhance robustness and flexibility. Instead of explicitly using object pose as rigid transformation, we learn the per-point motion in a pre-defined feature space. In this feature space, the object motion can be optimized through both explicit guidance and photometric loss. Moreover, the object motion can be interpolated through time and space in this feature space to compensate for missing detections. To this end, we leverage HexPlane representation [7] as in 4DGS [81] to efficiently capture spatial and temporal information by decomposing the 4D feature voxels into six learnable feature planes. This representation satisfies all our requirements and is also memory efficient." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.667, + 0.484, + 0.727 + ], + "angle": 0, + "content": "We use three planes: \\( P_{xy} \\), \\( P_{yz} \\), and \\( P_{xz} \\) for the spatial dimensions, and another three planes \\( P_{xt} \\), \\( P_{yt} \\), and \\( P_{zt} \\) for the spatial-temporal features. Additionally, the Hexplane includes multiple resolution levels which is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.739, + 0.483, + 0.758 + ], + "angle": 0, + "content": "\\[\n\\left\\{P _ {i j} ^ {\\rho} \\in \\mathbb {R} ^ {d \\times \\rho r _ {i} \\times \\rho r _ {j}} | (i, j) \\in \\mathcal {P}, \\rho \\in \\{1, 2 \\} \\right\\}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.769, + 0.483, + 0.814 + ], + "angle": 0, + "content": "where \\(\\mathcal{P} = \\{(x,y),(x,z),(y,z),(x,t),(y,t),(z,t)\\}\\), \\(d\\) is the feature dimension, \\(\\rho\\) denotes the upsampling scale, and \\(r\\) is the base resolution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.814, + 0.484, + 0.86 + ], + "angle": 0, + "content": "Given the center position of a Gaussian \\((x,y,z)\\) and the time step \\(t\\), features in all six planes are queried and combined via a small MLP \\(\\phi_d\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.87, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\nf (x, y, z, t) = \\phi_ {d} \\left(\\bigcup_ {\\rho} \\prod_ {(i, j) \\in \\mathcal {P}} \\pi \\left(P _ {i j} ^ {\\rho}, \\psi_ {i j} ^ {\\rho} (x, y, z, t)\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "where \\(\\psi_{ij}^{\\rho}(x,y,z,t)\\) projects the 4D coordinate onto the respective plane, and \\(\\pi\\) performs bilinear interpolation on the voxel features at each point. The features of all planes are combined using the Hadamard product." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.153, + 0.905, + 0.228 + ], + "angle": 0, + "content": "Finally, two MLP decoders \\(D_{\\mathcal{X}}, D_{c}\\) are utilized to predict the point motion \\(\\Delta \\mathcal{X}_t\\) and color change \\(\\Delta \\mathcal{C}_t\\) as \\(\\Delta \\mathcal{X}_t = D_{\\mathcal{X}}(f(x,y,z,t))\\) and \\(\\Delta \\mathcal{C}_t = D_{\\mathcal{C}}(f(x,y,z,t))\\), respectively. The deformed 4D Gaussians are formulated as: \\(\\mathcal{G}_t = \\{\\mathcal{X} + \\Delta \\mathcal{X}_t, \\mathcal{C} + \\Delta \\mathcal{C}_t, S, R, o\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.229, + 0.905, + 0.26 + ], + "angle": 0, + "content": "To allow the HexPlane feature to learn point motion from the predicted pose, we define the motion loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.273, + 0.905, + 0.291 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m o t i o n}} = \\operatorname {a v g} _ {\\mathcal {X} \\in \\mathcal {O}} | \\Delta \\mathcal {X} _ {t} - (\\mathbf {T} _ {t} \\mathcal {X} - \\mathcal {X}) |, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.303, + 0.905, + 0.408 + ], + "angle": 0, + "content": "where \\(\\mathcal{X}\\) is the center position of a Gaussian in object \\(\\mathcal{O}\\) and avg is the average operator. This loss encourages the predicted point motion \\(\\Delta \\mathcal{X}_t\\) to align with the predicted pose. We apply this loss only for the first \\(40\\%\\) iterations to provide a strong initial motion prior. The loss is subsequently removed, allowing the network to adjust and potentially compensate for pose errors and detection failures." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.409, + 0.906, + 0.56 + ], + "angle": 0, + "content": "Following [74], we adopt isotropic Gaussian marbles for dynamic points to reduce degrees of freedom and simplify optimization. In this setup, the rotation of each Gaussian is represented by the identity matrix with identical scales across all three dimensions. The spherical harmonics coefficients are also limited to three dimensions, providing view-consistent color. This approach ensures that all deformations are captured purely by point motion and color changes, enforcing strong constraints and enhancing robustness in novel view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.906, + 0.702 + ], + "angle": 0, + "content": "Ours vs. S3Gaussian. Although S3Gaussian [32] also uses Hexplane representation [7] and the deformation network from 4DGS [81], our design serves a different purpose: S3Gaussian learns point motion without extra supervision, while we use 4DGS to refine motion and address detection failures. Consequently, S3Gaussian becomes susceptible to unsatisfactory results with rapid car motion due to the lack of explicit guidance. In contrast, our approach enables high-quality reconstruction of dynamic objects." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.731, + 0.729 + ], + "angle": 0, + "content": "3.5. Optimization Objective" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Besides the motion loss introduced above, the loss function comprises five components to collectively optimize the scene representation and the point motion. The primary loss \\(\\mathcal{L}_{\\mathrm{rgb}}\\) is an L1 loss measuring the photometric difference between rendered and ground truth images while \\(\\mathcal{L}_{\\mathrm{ssim}}\\) evaluates their structural similarity. The L1 loss between the rendered depth map and the depth estimated from the LiDAR point cloud \\(\\mathcal{L}_{\\mathrm{depth}}\\) is used to supervise the Gaussian positions. Following K-Planes [70], a grid-based total variation loss \\(\\mathcal{L}_{\\mathrm{tv}}\\) is introduced. Recognizing that the color of most dynamic points in the scene are unchanged, a L1 regulariza" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25726" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.907, + 0.245 + ], + "angle": 0, + "content": "
TaskScene ReconstructionNovel View Synthesis
MethodExtra InputPSNR↑SSIM↑LPIPS↓DPSNR↑DSSIM↑PSNR↑SSIM↑LPIPS↓DPSNR↑DSSIM↑
3DGS [37]N/A25.770.8330.16020.260.60424.460.8020.17018.120.521
EmerNeRF [90]N/A28.160.8060.22824.320.68225.140.7470.31323.490.660
S3Gaussian [32]N/A31.350.9110.10626.020.78327.440.8570.13722.920.680
MARS [85]GT pose28.240.8660.25223.370.70126.610.7960.30522.210.697
StreetGS [89]3D tracker29.170.8730.13827.780.81826.980.8380.14924.620.742
Ours2D tracker32.560.9360.05928.510.86828.850.8670.08825.580.779
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.258, + 0.907, + 0.289 + ], + "angle": 0, + "content": "Table 1. Comparison with state-of-the-art methods on Waymo-NOTR dataset. StreetGS represents Street Gaussian [89]. The best and the second best results are denoted by pink and blue." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.315, + 0.482, + 0.345 + ], + "angle": 0, + "content": "tion loss \\(\\mathcal{L}_{\\mathrm{color - reg}}\\) is applied to the deformation network by minimizing \\(\\Delta C\\)." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.346, + 0.383, + 0.36 + ], + "angle": 0, + "content": "The total loss function is thus defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.368, + 0.482, + 0.405 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\lambda_ {\\mathrm {r g b}} \\mathcal {L} _ {\\mathrm {r g b}} + \\lambda_ {\\text {d e p t h}} \\mathcal {L} _ {\\text {d e p t h}} + \\lambda_ {\\text {s s i m}} \\mathcal {L} _ {\\text {s s i m}} + \\lambda_ {\\mathrm {t v}} \\mathcal {L} _ {\\mathrm {t v}} \\tag {8} \\\\ + \\lambda_ {\\text {c o l o r - r e g}} \\mathcal {L} _ {\\text {c o l o r - r e g}} + \\lambda_ {\\text {m o t i o n}} \\mathcal {L} _ {\\text {m o t i o n}}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.483, + 0.459 + ], + "angle": 0, + "content": "where the weights are assigned as follows: \\(\\lambda_{\\mathrm{rgb}} = 1.0\\) \\(\\lambda_{\\mathrm{depth}} = 1.0\\) \\(\\lambda_{\\mathrm{ssim}} = 0.1\\) \\(\\lambda_{\\mathrm{tv}} = 0.1\\) \\(\\lambda_{\\mathrm{color-reg}} = 0.01\\) and \\(\\lambda_{\\mathrm{motion}} = 1\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.461, + 0.484, + 0.566 + ], + "angle": 0, + "content": "The optimization process is divided into two stages. In the first stage, we render using only static Gaussians and supervise only the static regions of the images over 20,000 iterations. This phase focuses on reconstructing the static background and stabilizes the training process. In the second stage, we train all Gaussians on the whole images for 50,000 iterations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.579, + 0.224, + 0.596 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.604, + 0.281, + 0.621 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.626, + 0.483, + 0.852 + ], + "angle": 0, + "content": "Datasets. We leverage Waymo-NOTR dataset [76, 90] and KITTI benchmark [23, 24]. The NeRF On-The-Road (NOTR) dataset, introduced by EmerNeRF [90], is a subset of the Waymo Open dataset [76] and includes diverse challenging driving scenarios, such as high-speed, exposure mismatch, and various weather conditions. We use the dynamic32 subset, consisting of 32 dynamic scenes, for evaluating dynamic reconstruction. Following EmerNeRF's setup [90], we use three frontal camera images resized to \\(960 \\times 640\\). For scene reconstruction, all image frames are used for training and evaluation; for novel view synthesis, every 10th time step is excluded for evaluation [90]. For the KITTI dataset, we follow the setup of MARS [85] to use \\(75\\%\\) or \\(50\\%\\) of the images for training with every 4th or every 2nd frame held out for testing, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. We use the Adam optimizer [39] with the same learning rate schedule as in 3DGS [37]. For long-sequence scene reconstruction, we follow S3Gaussian" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.315, + 0.907, + 0.452 + ], + "angle": 0, + "content": "[32] and segment the scene into multiple clips. The multi-resolution HexPlane encoder has a base resolution of 64, upsampled by factors of 2 and 4 as in [81], and other hyperparameters match those in 3DGS [37]. As LiDAR points lack data for the sky region, we add a plane of points above the maximum height of the scene to represent it following [38]. All experiments are run on a single NVIDIA RTX 3090 GPU, with training taking about 2 hours per video clip and inference speed at 100 FPS at \\(960 \\times 640\\) resolution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.464, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Baseline Methods. We compare our approach against state-of-the-art methods, including NeRF-based methods (MARS [85], NSG [65], EmerNeRF [90]) and Gaussian-based methods (3DGS [37], StreetGaussian [89], S3Gaussian [32]). To ensure fair comparisons, we apply LiDAR point cloud initialization and add depth regularization to 3DGS. For the Waymo-NOTR dataset, we borrow results of MARS [85] and EmerNeRF [90] from S3Gaussian [32]. We evaluate the 3D-tracker-based method, Street Gaussians [89], using detection results from VoxelNext [11] and associate the detections with SimpleTrack [66]. Our goal is to assess the generalization ability of 3D-tracker-based methods in scenarios without ground truth tracking labels. To emulate a real-world scenario where a 3D tracker trained on a public benchmark dataset is applied to a novel environment, we use pretrained weights from nuScenes [6]—one of the largest and most diverse 3D datasets—and evaluate the model on the Waymo-NOTR dataset. Since NOTR is part of the training set for the Waymo perception task, using Waymo pretrained weights would introduce domain overlap and bias the evaluation, making the nuScenes weights a fairer choice. For a fair comparison, we use pretrained 2D tracker weights from [84] for our method, which have not been trained on the Waymo or KITTI datasets. Please refer to the supplementary material for discussion of the choice of 3D trackers. For the KITTI dataset, we borrow the results of all other methods from Street Gaussians. Please note that Street Gaussians leverages GT tracking annotations on KITTI dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25727" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.195, + 0.089, + 0.223, + 0.102 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.356, + 0.089, + 0.444, + 0.102 + ], + "angle": 0, + "content": "S3Gaussian" + }, + { + "type": "image_caption", + "bbox": [ + 0.537, + 0.089, + 0.653, + 0.102 + ], + "angle": 0, + "content": "Street Gaussian" + }, + { + "type": "image_caption", + "bbox": [ + 0.769, + 0.089, + 0.809, + 0.102 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.108, + 0.303, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.108, + 0.495, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.108, + 0.689, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.108, + 0.884, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.207, + 0.303, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.207, + 0.495, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.207, + 0.689, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.207, + 0.884, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.307, + 0.303, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.307, + 0.495, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.307, + 0.689, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.307, + 0.884, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.407, + 0.303, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.407, + 0.495, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.407, + 0.689, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.407, + 0.884, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.507, + 0.303, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.507, + 0.495, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.507, + 0.689, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.507, + 0.884, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.2, + 0.619, + 0.796, + 0.633 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparison of novel view synthesis on NOTR dataset. Best viewed with zoom." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.766 + ], + "angle": 0, + "content": "Metrics. We leverage peak signal-to-noise ratio (PSNR), structural similarity index (SSIM) and learned perceptual image patch similarity (LPIPS) to evaluate the rendering quality. Additionally, following EmerNeRF [90], we apply DPSNR and DSSIM metrics to dynamic objects by projecting their ground truth 3D bounding boxes onto the 2D image plane and calculating pixel loss within these regions." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.773, + 0.421, + 0.789 + ], + "angle": 0, + "content": "4.2. Comparisons with the State-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "On the NOTR dataset, our method outperforms all competitors across every metric, as shown in Table 1. Our approach sets a new state-of-the-art in both the scene reconstruction and novel view synthesis tasks. Specifically, our method outperforms S3Gaussian [32] by \\(2.66\\mathrm{dB}\\) in DPSNR and 0.099 in DSSIM for novel view synthesis. It also surpasses Street Gaussians [89] by \\(1.87\\mathrm{dB}\\) in PSNR, mainly due" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.907, + 0.735 + ], + "angle": 0, + "content": "to the limited generalization of 3D trackers used in Street Gaussians. Despite optimizing object poses, Street Gaussians struggles with large pose errors and detection failures, leading to inferior performance. Additionally, we outperform NeRF-based methods like EmerNeRF and MARS." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We present qualitative comparison with S3Gaussian [32] and Street Gaussians [89] in Fig. 3. Results from S3Gaussian show that using 4DGS without explicit motion guidance results in weaker performance when handling moving vehicles (e.g. row 2, 3, 5). Street Gaussians suffers from tracking errors of 3D trackers, leading to erroneous reconstructions (e.g. row 1, 3) or entirely missed objects (e.g. rows 2, 4, and 5). In contrast, our approach performs robustly across diverse scenarios owing to our tracking strategy leveraging 2D foundation model and robust motion" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25728" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.153, + 0.089, + 0.847, + 0.228 + ], + "angle": 0, + "content": "
KITTI-75%KITTI-50%
MethodExtra InputPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
3DGS [37]N/A19.190.7370.17219.230.7390.174
NSG [65]GT pose21.530.6730.25421.260.6590.266
MARS [85]GT pose24.230.8450.16024.000.8010.164
Street Gussians [89]GT pose25.790.8440.08125.520.8410.084
Ours2D tracker [84]25.490.8890.06325.110.8770.067
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.238, + 0.905, + 0.253 + ], + "angle": 0, + "content": "Table 2. Comparison with state-of-the-art methods on KITTI dataset. The best and the second best results are denoted by pink and blue." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.29, + 0.48, + 0.468 + ], + "angle": 0, + "content": "
MethodPSNRSSIMDPSNRDSSIM
Aw/o track25.250.76121.710.662
Bw/o Lmotion28.390.84324.540.719
Cw/o iso.28.580.84725.350.765
D3D tracker28.020.83524.610.746
ED w/o Lmotion27.540.82123.530.703
FGT pose28.980.87325.710.787
GUNINEXT28.760.86425.430.772
HOurs28.850.86725.580.779
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.476, + 0.483, + 0.505 + ], + "angle": 0, + "content": "Table 3. Ablation study of Novel View Synthesis on NOTR dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.531, + 0.483, + 0.561 + ], + "angle": 0, + "content": "learning module. Please refer to the supplementary material for additional video results and editing examples." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.561, + 0.483, + 0.713 + ], + "angle": 0, + "content": "The results on the KITTI dataset are presented in Table 2. Our method outperforms the NeRF-based NSG and MARS across all metrics. In the KITTI-75% setting, our approach achieves a 0.045 higher SSIM and a 0.018 lower LPIPS compared to Street Gaussians, although its PSNR is 0.3 dB lower. A similar trend is observed in the KITTI-50% setting. This lower PSNR is primarily because Street Gaussians leverages labor-intensive ground truth tracking annotations, whereas our method uses a generalized 2D tracker—yet still attains comparable performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.721, + 0.255, + 0.735 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.743, + 0.483, + 0.773 + ], + "angle": 0, + "content": "We conduct an ablation study on the NOTR dataset and the results are presented in Tab. 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.777, + 0.483, + 0.868 + ], + "angle": 0, + "content": "Effect of object tracking module. In Tab. 3 (A), we model dynamic points using 4DGS [81] without performing any form of tracking, and observe significant drops across all metrics. This experiment highlights that the sole use of 4DGS does not provide sufficient accuracy for effective motion modeling in autonomous driving scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Effect of motion learning strategy. In Tab. 3 (B), we use the reconstruction of the tracking module for initialization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.294, + 0.905, + 0.37 + ], + "angle": 0, + "content": "but remove the motion loss introduced in Eq. 7, resulting in a \\(1.04\\mathrm{dB}\\) drop in DPSNR. This result highlights the importance of our motion learning strategy. A comparison with rigid-transformation-based motion modeling [89] is provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.374, + 0.905, + 0.435 + ], + "angle": 0, + "content": "Effect of isotropic Gaussian marbles. In Tab. 3 (C), we substitute isotropic Gaussian marbles with the original anisotropic Gaussian ellipsoids for dynamic points, which leads to a decrease in DPSNR and DSSIM." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.44, + 0.907, + 0.607 + ], + "angle": 0, + "content": "3D tracker / GT pose vs. Our object tracking module. In Tab. 3 (D), we use the 3D tracker [11, 66] to reconstruct dynamic objects and supervise motion learning, resulting in a 0.97 dB DPSNR decrease. Removing motion loss leads to a further 1.08 dB drop (Tab. 3 (E)), showing that our motion learning strategy compensates for tracking errors, though these errors still impact novel view synthesis, emphasizing the need for a more generalizable 2D tracker. Using ground-truth trajectories (Tab. 3 (F)) shows only a small gap, confirming the effectiveness of our 2D tracker in motion learning. See supplementary for tracking error analysis." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.905, + 0.673 + ], + "angle": 0, + "content": "Choice of 2D tracker. To demonstrate the robustness of our method with respect to the choice of 2D tracker, we employ UNINEXT [88] for 2D tracking. As shown in Table 3 (G), our method maintains stable performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.693, + 0.634, + 0.708 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduce a novel framework for robust dynamic 3D street scene reconstruction that eliminates the reliance on 3D object trackers. Addressing the generalization limitations of 3D trackers, we propose a robust object tracking strategy based on a 2D foundation model. Our framework also features a motion learning module within an implicit feature space to handle inevitable tracking errors by autonomously refining pose inaccuracies and recovering missed detections. Experiments on the Waymo-NOTR and KITTI datasets demonstrate its adaptability and superior performance. Limitations and failure cases are discussed in the supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25729" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.207 + ], + "angle": 0, + "content": "This work was supported by the National Key R&D Program of China under Grant 2018AAA0102801, PCL-CMCC Foundation for Science and Innovation Grant No. 2024ZY1C0040, and the National Research Foundation, Singapore, under its NRF-Investigatorship Programme (Award ID. NRF-NRFI09-0008)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.222, + 0.188, + 0.238 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.248, + 0.484, + 0.329 + ], + "angle": 0, + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5855–5864, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.332, + 0.484, + 0.401 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470-5479, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.403, + 0.483, + 0.473 + ], + "angle": 0, + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.475, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[4] Mk Bashar, Samia Islam, Kashifa Kawaakib Hussain, Md Bakhtiar Hasan, ABM Rahman, and Md Hasanul Kabir. Multiple object tracking in recent times: A literature review. arXiv preprint arXiv:2209.04796, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.532, + 0.483, + 0.587 + ], + "angle": 0, + "content": "[5] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In 2016 IEEE international conference on image processing (ICIP), pages 3464-3468. IEEE, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.589, + 0.485, + 0.672 + ], + "angle": 0, + "content": "[6] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.674, + 0.483, + 0.73 + ], + "angle": 0, + "content": "[7] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.732, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[8] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.802, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[9] Zhengping Che, Guangyu Li, Tracy Li, Bo Jiang, Xuefeng Shi, Xinsheng Zhang, Ying Lu, Guobin Wu, Yan Liu, and Jieping Ye. D2-city: a large-scale dashcam video dataset of diverse traffic scenarios. arXiv preprint arXiv:1904.01975, 2019.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[10] Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.248, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.907, + 0.12 + ], + "angle": 0, + "content": "reconstruction and real-time rendering. arXiv preprint arXiv:2311.18561, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.122, + 0.906, + 0.191 + ], + "angle": 0, + "content": "[11] Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21674-21683, 2023. 1, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.193, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[12] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.25, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[13] Jie Cheng, Yingbing Chen, Qingwen Zhang, Lu Gan, Chengju Liu, and Ming Liu. Real-time trajectory planning for autonomous driving with gaussian process and incremental refinement. In 2022 International Conference on Robotics and Automation (ICRA), pages 8999-9005. IEEE, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.334, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[14] Jie Cheng, Xiaodong Mei, and Ming Liu. Forecast-mae: Self-supervised pre-training for motion forecasting with masked autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8679-8689, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.406, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[15] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. ACM Transactions on Graphics (ToG), 34(4):1-13, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.476, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[16] Daniel Dauner, Marcel Hallgarten, Andreas Geiger, and Kashyap Chitta. Parting with misconceptions about learning-based vehicle motion planning. In Conference on Robot Learning, pages 1268-1281. PMLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.533, + 0.905, + 0.589 + ], + "angle": 0, + "content": "[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.591, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[18] Yan Di, Ruida Zhang, Zhiqiang Lou, Fabian Manhardt, Xiangyang Ji, Nassir Navab, and Federico Tombari. Gpvpose: Category-level object pose estimation via geometry-guided point-wise voting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6781-6791, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.675, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[19] Shuxiao Ding, Eike Rehder, Lukas Schneider, Marius Cordts, and Juergen Gall. 3dmotformer: Graph transformer for online 3d multi-object tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9784-9794, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.746, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[20] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.802, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[21] George Eskandar. An empirical study of the generalization ability of lidar 3d object detectors to unseen domains. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23815-23825, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[22] Tobias Fischer, Jonas Kulhanek, Samuel Rota Bulo, Lorenzo Porzi, Marc Pollefeys, and Peter Kontschieder." + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25730" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.483, + 0.12 + ], + "angle": 0, + "content": "Dynamic 3d gaussian fields for urban areas. arXiv preprint arXiv:2406.03175, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.122, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[23] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.179, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[24] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.236, + 0.483, + 0.305 + ], + "angle": 0, + "content": "[25] Junru Gu, Chenxu Hu, Tianyuan Zhang, Xuanyao Chen, Yilun Wang, Yue Wang, and Hang Zhao. Vip3d: End-to-end visual trajectory prediction via 3d agent queries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5496-5506, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.307, + 0.483, + 0.376 + ], + "angle": 0, + "content": "[26] Kaiwen Guo, Feng Xu, Yangang Wang, Yebin Liu, and Qionghai Dai. Robust non-rigid motion tracking and surface reconstruction using 10 regularization. In Proceedings of the IEEE International Conference on Computer Vision, pages 3083-3091, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.378, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[27] Kaiwen Guo, Peter Lincoln, Philip Davidson, Jay Busch, Xueming Yu, Matt Whalen, Geoff Harvey, Sergio Orts-Escolano, Rohit Pandey, Jason Dourgarian, et al. The relightables: Volumetric performance capture of humans with realistic relighting. ACM Transactions on Graphics (ToG), 38(6):1-19, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.462, + 0.482, + 0.518 + ], + "angle": 0, + "content": "[28] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.52, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[29] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.591, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[30] Tao Hu, Tao Yu, Zerong Zheng, He Zhang, Yebin Liu, and Matthias Zwicker. Hvtr: Hybrid volumetric-textural rendering for human avatars. In 2022 International Conference on 3D Vision (3DV), pages 197-208. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.647, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[31] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.718, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[32] Nan Huang, Xiaobao Wei, Wenzhao Zheng, Pengju An, Ming Lu, Wei Zhan, Masayoshi Tomizuka, Kurt Keutzer, and Shanghang Zhang. S3gaussian: Self-supervised street gaussians for autonomous driving. arXiv preprint arXiv:2405.20323, 2024. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.789, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[33] Yanjun Huang, Jiatong Du, Ziru Yang, Zewei Zhou, Lin Zhang, and Hong Chen. A survey on trajectory-prediction methods for autonomous driving. IEEE Transactions on Intelligent Vehicles, 7(3):652-674, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[34] Ziqin Huang, Gu Wang, Chenyangguang Zhang, Ruida Zhang, Xiu Li, and Xiangyang Ji. Givepose: Gradual intraclass variation elimination for rgb-based category-level object pose estimation. In Proceedings of the Computer Vision" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "and Pattern Recognition Conference, pages 22055-22066, 2025.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.122, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[35] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8350, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.207, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[36] Lesole Kalake, Wanggen Wan, and Li Hou. Analysis based on recent deep learning approaches applied in real-time multi-object tracking: a review. IEEE Access, 9:32650-32671, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.264, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[37] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 3, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.321, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[38] Mustafa Khan, Hamidreza Fazlali, Dhruv Sharma, Tongtong Cao, Dongfeng Bai, Yuan Ren, and Bingbing Liu. Autosplat: Constrained gaussian splatting for autonomous driving scene reconstruction. arXiv preprint arXiv:2407.02598, 2024. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.391, + 0.904, + 0.419 + ], + "angle": 0, + "content": "[39] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.421, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[40] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.505, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[41] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision, 128(7): 1956-1981, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.604, + 0.905, + 0.672 + ], + "angle": 0, + "content": "[42] Jinyu Li, Chenxu Luo, and Xiaodong Yang. Pillarnext: Rethinking network designs for 3d object detection in lidar point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17567-17576, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.675, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[43] Wei Li, CW Pan, Rong Zhang, JP Ren, YX Ma, Jin Fang, FL Yan, QC Geng, XY Huang, HJ Gong, et al. Aads: Augmented autonomous driving simulation using data-driven algorithms. Science robotics, 4(28):eaaw0863, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.732, + 0.905, + 0.799 + ], + "angle": 0, + "content": "[44] Yanyan Li, Chenyu Lyu, Yan Di, Guangyao Zhai, Gim Hee Lee, and Federico Tombari. Geogaussian: Geometry-aware gaussian splatting for scene rendering. In European Conference on Computer Vision, pages 441-457. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.802, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[45] Zhong Li, Yu Ji, Wei Yang, Jinwei Ye, and Jingyi Yu. Robust 3d human motion reconstruction via dynamic template construction. In 2017 International Conference on 3D Vision (3DV), pages 496-505. IEEE, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.86, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[46] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25731" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14864-14873, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.122, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[47] Ming Liang, Bin Yang, Wenyuan Zeng, Yun Chen, Rui Hu, Sergio Casas, and Raquel Urtasun. Pnpnet: End-to-end perception and prediction with tracking in the loop. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11553-11562, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.191, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[48] Yiqing Liang, Numair Khan, Zhengqin Li, Thu Nguyen-Phuoc, Douglas Lanman, James Tompkin, and Lei Xiao. Gaufre: Gaussian deformation fields for real-time dynamic novel view synthesis. arXiv preprint arXiv:2312.11458, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.261, + 0.483, + 0.316 + ], + "angle": 0, + "content": "[49] Yiyi Liao, Jun Xie, and Andreas Geiger. Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3292-3310, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.317, + 0.483, + 0.399 + ], + "angle": 0, + "content": "[50] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.401, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[51] Xingyu Liu, Gu Wang, Yi Li, and Xiangyang Ji. CATRE: iterative point clouds alignment for category-level object pose refinement. In European Conference on Computer Vision, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.457, + 0.482, + 0.537 + ], + "angle": 0, + "content": "[52] Xingyu Liu, Chenyangguang Zhang, Gu Wang, Ruida Zhang, and Xiangyang Ji. Rasim: A range-aware high-fidelity rgb-d data simulation pipeline for real-world applications. In 2024 IEEE international conference on robotics and automation (ICRA), pages 17057-17064. IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.54, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[53] Xingyu Liu, Yingyue Li, Chengxi Li, Gu Wang, Chenyangguang Zhang, Ziqin Huang, and Xiangyang Ji. Gfreedet: Exploiting gaussian splatting and foundation models for model-free unseen object detection in the bop challenge 2024. IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.623, + 0.482, + 0.691 + ], + "angle": 0, + "content": "[54] Xingyu Liu, Gu Wang, Ruida Zhang, Chenyangguang Zhang, Federico Tombari, and Xiangyang Ji. UNOPose: Unseen object pose estimation with an unposed rgb-d reference image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.693, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[55] Xingyu Liu, Ruida Zhang, Chenyangguang Zhang, Gu Wang, Jiwen Tang, Zhigang Li, and Xiangyang Ji. Gdnpp: A geometry-guided and fully learning-based object pose estimator. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.763, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[56] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[57] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[58] Wenhan Luo, Junliang Xing, Anton Milan, Xiaoqin Zhang, Wei Liu, and Tae-Kyun Kim. Multiple object tracking: A literature review. Artificial intelligence, 293:103448, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.15, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[59] Jiageng Mao, Minzhe Niu, Chenhan Jiang, Hanxue Liang, Jingheng Chen, Xiaodan Liang, Yamin Li, Chaoqiang Ye, Wei Zhang, Zhenguo Li, et al. One million scenes for autonomous driving: Once dataset. arXiv preprint arXiv:2106.11037, 2021.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.221, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[60] Jiageng Mao, Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. 3d object detection for autonomous driving: A comprehensive survey. International Journal of Computer Vision, 131(8):1909-1963, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.278, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[61] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.349, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[62] M Minderer, A Gritsenko, A Stone, M Neumann, D Weissenborn, A Dosovitskiy, A Mahendran, A Arnab, M Dehghani, Z Shen, et al. Simple open-vocabulary object detection with vision transformers. arxiv 2022. arXiv preprint arXiv:2205.06230, 2, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.42, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[63] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.477, + 0.905, + 0.546 + ], + "angle": 0, + "content": "[64] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.548, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[65] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2856-2865, 2021. 1, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.619, + 0.905, + 0.674 + ], + "angle": 0, + "content": "[66] Ziqi Pang, Zhichao Li, and Naiyan Wang. Simpletrack: Understanding and rethinking 3d multi-object tracking. In European Conference on Computer Vision, pages 680-696. Springer, 2022. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.676, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[67] Chensheng Peng, Chengwei Zhang, Yixiao Wang, Chenfeng Xu, Yichen Xie, Wenzhao Zheng, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. Desire-gs: 4d street gaussians for static-dynamic decomposition and surface reconstruction for urban driving scenes. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 6782-6791, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.775, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[68] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[69] Jules Sanchez, Jean-Emmanuel Deschaud, and François Goulette. Domain generalization of 3d semantic segmentation in autonomous driving. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25732" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "IEEE/CVF International Conference on Computer Vision, pages 18077-18087, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.122, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[70] Sara Fridovich-Keil and Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In CVPR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.178, + 0.482, + 0.246 + ], + "angle": 0, + "content": "[71] Shital Shah, Debadeepta Dey, Chris Lovett, and Ashish Kapoor. Airsim: High-fidelity visual and physical simulation for autonomous vehicles. In Field and Service Robotics: Results of the 11th International Conference, pages 621-635. Springer, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.247, + 0.483, + 0.302 + ], + "angle": 0, + "content": "[72] Yeji Song, Chaerin Kong, Seoyoung Lee, Nojun Kwak, and Joonseok Lee. Towards efficient neural scene graphs by learning consistency fields. arXiv preprint arXiv:2210.04127, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.303, + 0.483, + 0.371 + ], + "angle": 0, + "content": "[73] Louis Soum-Fontez, Jean-Emmanuel Deschaud, and François Goulette. Mdt3d: Multi-dataset training for lidar 3d object detection generalization. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5765-5772. IEEE, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.373, + 0.483, + 0.44 + ], + "angle": 0, + "content": "[74] Colton Stearns, Adam Harley, Mikaela Uy, Florian Dubost, Federico Tombari, Gordon Wetzstein, and Leonidas Guibas. Dynamic gaussian marbles for novel view synthesis of casual monocular videos. arXiv preprint arXiv:2406.18717, 2024. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.442, + 0.483, + 0.525 + ], + "angle": 0, + "content": "[75] Zhuo Su, Lan Xu, Zerong Zheng, Tao Yu, Yebin Liu, and Lu Fang. Robustfusion: Human volumetric capture with data-driven visual cues using a rgbd camera. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 246-264. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.526, + 0.483, + 0.621 + ], + "angle": 0, + "content": "[76] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.623, + 0.483, + 0.692 + ], + "angle": 0, + "content": "[77] Shanlin Sun, Bingbing Zhuang, Ziyu Jiang, Buyu Liu, Xiaohui Xie, and Manmohan Chandraker. Lidarf: Delving into lidar for neural radiance field on street scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19563-19572, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.693, + 0.483, + 0.76 + ], + "angle": 0, + "content": "[78] Haithem Turki, Jason Y Zhang, Francesco Ferroni, and Deva Ramanan. Suds: Scalable urban dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12375-12385, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.762, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[79] Yan Wang, Xiangyu Chen, Yurong You, Li Erran Li, Bharath Hariharan, Mark Campbell, Kilian Q Weinberger, and Wei-Lun Chao. Train in germany, test in the usa: Making 3d object detectors generalize. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11713-11723, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[80] Yu Wang, Shaohua Wang, Yicheng Li, and Mingchun Liu. A comprehensive review of 3d object detection in autonomous driving: Technological advances and future directions. arXiv preprint arXiv:2408.16530, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[81] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20310-20320, 2024. 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.179, + 0.905, + 0.246 + ], + "angle": 0, + "content": "[82] Hai Wu, Wenkai Han, Chenglu Wen, Xin Li, and Cheng Wang. 3d multi-object tracking in point clouds based on prediction confidence-guided data association. IEEE Transactions on Intelligent Transportation Systems, 23(6):5668-5677, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.249, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[83] Hai Wu, Jinhao Deng, Chenglu Wen, Xin Li, Cheng Wang, and Jonathan Li. Casa: A cascade attention network for 3-d object detection from lidar point clouds. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.307, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[84] Junfeng Wu, Yi Jiang, Qihao Liu, Zehuan Yuan, Xiang Bai, and Song Bai. General object foundation model for images and videos at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3783-3795, 2024. 1, 2, 3, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.377, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[85] Zirui Wu, Tianyu Liu, Liyi Luo, Zhide Zhong, Jianteng Chen, Hongmin Xiao, Chao Hou, Haozhe Lou, Yuantao Chen, Runyi Yang, et al. Mars: An instance-aware, modular and realistic simulator for autonomous driving. In CAAI International Conference on Artificial Intelligence, pages 3-15. Springer, 2023. 1, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.462, + 0.905, + 0.543 + ], + "angle": 0, + "content": "[86] Pengchuan Xiao, Zhenlei Shao, Steven Hao, Zishuo Zhang, Xiaolin Chai, Judy Jiao, Zesong Li, Jian Wu, Kai Sun, Kun Jiang, et al. Pandaset: Advanced sensor suite dataset for autonomous driving. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 3095-3101. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.547, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[87] Ziyang Xie, Junge Zhang, Wenye Li, Feihu Zhang, and Li Zhang. S-nerf: Neural radiance fields for street views. arXiv preprint arXiv:2303.00749, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.59, + 0.907, + 0.658 + ], + "angle": 0, + "content": "[88] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15325-15336, 2023. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.661, + 0.905, + 0.728 + ], + "angle": 0, + "content": "[89] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians: Modeling dynamic urban scenes with gaussian splatting. In ECCV, 2024. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.732, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[90] Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Seung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, et al. Emernerf: Emergent spatial-temporal scene decomposition via self-supervision. arXiv preprint arXiv:2311.02077, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.802, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[91] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1389-1399, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[92] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time photorealistic dynamic scene representation and" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25733" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.483, + 0.12 + ], + "angle": 0, + "content": "rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.122, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[93] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20331-20341, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.207, + 0.483, + 0.262 + ], + "angle": 0, + "content": "[94] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.264, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[95] Fisher Yu, Wenqi Xian, Yingying Chen, Fangchen Liu, Mike Liao, Vashisht Madhavan, Trevor Darrell, et al. Bdd100k: A diverse driving video database with scalable annotation tooling. arXiv preprint arXiv:1805.04687, 2(5): 6, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.334, + 0.483, + 0.404 + ], + "angle": 0, + "content": "[96] Jiang-Tian Zhai, Ze Feng, Jinhao Du, Yongqiang Mao, Jiang-Jiang Liu, Zichang Tan, Yifu Zhang, Xiaqing Ye, and Jingdong Wang. Rethinking the open-loop evaluation of end-to-end autonomous driving in nuscenes. arXiv preprint arXiv:2305.10430, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.406, + 0.483, + 0.474 + ], + "angle": 0, + "content": "[97] Bo Zhang, Jiakang Yuan, Botian Shi, Tao Chen, Yikang Li, and Yu Qiao. Uni3d: A unified baseline for multi-dataset 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9253-9262, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.477, + 0.483, + 0.518 + ], + "angle": 0, + "content": "[98] Peng Zhang, Xin Li, Liang He, and Xin Lin. 3d multiple object tracking on autonomous driving: A literature review. arXiv preprint arXiv:2309.15411, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.52, + 0.483, + 0.589 + ], + "angle": 0, + "content": "[99] Ruida Zhang, Yan Di, Zhiqiang Lou, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. Rbp-pose: Residual bounding box projection for category-level pose estimation. In European conference on computer vision, pages 655–672. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.591, + 0.483, + 0.673 + ], + "angle": 0, + "content": "[100] Ruida Zhang, Yan Di, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. Ssp-pose: Symmetry-aware shape prior deformation for direct category-level object pose estimation. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7452–7459. IEEE, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.676, + 0.483, + 0.758 + ], + "angle": 0, + "content": "[101] Ruida Zhang, Ziqin Huang, Gu Wang, Chenyangguang Zhang, Yan Di, Xingxing Zuo, Jiwen Tang, and Xiangyang Ji. Lapose: Laplacian mixture shape modeling for rgb-based category-level object pose estimation. In European Conference on Computer Vision, pages 467-484. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[102] Ruida Zhang, Chenyangguang Zhang, Yan Di, Fabian Manhardt, Xingyu Liu, Federico Tombari, and Xiangyang Ji. Kp-red: Exploiting semantic keypoints for joint 3d shape retrieval and deformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20540-20550, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.845, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[103] Xingxuan Zhang, Zekai Xu, Renzhe Xu, Jiashuo Liu, Peng Cui, Weitao Wan, Chong Sun, and Chen Li. Towards domain generalization in object detection. arXiv preprint arXiv:2203.14387, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[104] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. International journal of computer vision, 129:3069–3087, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[105] Hongyu Zhou, Jiahao Shao, Lu Xu, Dongfeng Bai, Weichao Qiu, Bingbing Liu, Yue Wang, Andreas Geiger, and Yiyi Liao. Hugs: Holistic urban 3d scene understanding via gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21336-21345, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[106] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21634-21643, 2024. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25734" + } + ] +] \ No newline at end of file diff --git a/2025/Street Gaussians without 3D Object Tracker/ceea65cc-c6ed-4695-b497-f6fc77f3a4c6_origin.pdf b/2025/Street Gaussians without 3D Object Tracker/ceea65cc-c6ed-4695-b497-f6fc77f3a4c6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3b087cbf3e7625308bb5ba99c32a775ec0d22ed4 --- /dev/null +++ b/2025/Street Gaussians without 3D Object Tracker/ceea65cc-c6ed-4695-b497-f6fc77f3a4c6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f9c8286f4b930b3a65d13926f384014227cf9df760c503ef2e988e32981716b7 +size 2836349 diff --git a/2025/Street Gaussians without 3D Object Tracker/full.md b/2025/Street Gaussians without 3D Object Tracker/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4a855bd8977eda483576ec9acde2013f66563547 --- /dev/null +++ b/2025/Street Gaussians without 3D Object Tracker/full.md @@ -0,0 +1,392 @@ +# Street Gaussians without 3D Object Tracker + +Ruida Zhang $^{1,2}$ , Chengxi Li $^{1}$ , Chenyangguang Zhang $^{1}$ , Xingyu Liu $^{1}$ , Haili Yuan $^{1}$ , Yanyan Li $^{2}$ , Xiangyang Ji $^{1}$ , Gim Hee Lee $^{2}$ $^{1}$ Tsinghua University, $^{2}$ National University of Singapore + +{zhangrd23@mails, xyji@}.tsinghua.edu.cn, gimhee.lee@nus.edu.sg + +![](images/5c0906e1d8a489a3a426abffcc9c2fb05a484304038c38d4a87724cee7c557e9.jpg) +Figure 1. Comparison of 3D tracker-based Street Gaussians [89] (left) and our approach (right). Existing methods heavily rely on object poses, but 3D trackers struggle with limited generalization [21, 73, 97], leading to flaws in novel view synthesis. In contrast, 2D foundation models show better generalization [84, 88]. Our approach leverages a 2D foundation model [84] for object tracking and learns point motion within an implicit feature space to autonomously correct tracking errors, improving robustness across diverse scenes. + +![](images/7593f0549658f8187461f1f91ffff7fd5f5e85cbf312b35d5e34ce8d5fa0f1ba.jpg) + +# Abstract + +Realistic scene reconstruction in driving scenarios poses significant challenges due to fast-moving objects. Most existing methods rely on labor-intensive manual labeling of object poses to reconstruct dynamic objects in canonical space and move them based on these poses during rendering. While some approaches attempt to use 3D object trackers to replace manual annotations, the limited generalization of 3D trackers – caused by the scarcity of large-scale 3D datasets – results in inferior reconstructions in real-world settings. In contrast, 2D foundation models demonstrate strong generalization capabilities. To eliminate the reliance on 3D trackers and enhance robustness across diverse environments, we propose a stable object tracking module by leveraging associations from 2D deep trackers within a 3D object fusion strategy. We address inevitable tracking errors by further introducing a motion learning strategy in an implicit feature space that autonomously corrects trajectory errors and recovers missed detections. Experimental results on Waymo-NOTR and KITTI show that our method outperforms existing approaches. Our code will be made publicly available. + +# 1. Introduction + +Modeling dynamic 3D street scenes underpins modern autonomous driving by enabling realistic, controllable simulations for tasks such as perception [11, 42, 56, 94, 105], prediction [25, 33, 47], and motion planning [13, 14, 16]. With the rise of end-to-end autonomous systems that require real-time sensor feedback [29, 31, 35], high-quality scene reconstructions have become essential for closed-loop evaluations [46, 96], particularly to simulate critical corner cases safely and cost-effectively. + +Despite extensive efforts in achieving photo-realistic reconstruction of small-scale scenes, the unique challenges posed by the large-scale and highly dynamic nature of driving scenarios complicate effective 3D scene modeling. To address these challenges, most existing methods [22, 38, 65, 85, 106] rely on ground truth vehicle poses to differentiate between static background and moving vehicles. Typically, vehicles are reconstructed in canonical space and subsequently positioned based on known poses during rendering. However, collecting ground-truth poses is labor-intensive, limiting the applicability of these methods to scenes beyond the existing datasets. + +To eliminate reliance of the systems on ground truth object poses, Street Gaussians [89] instead uses poses generated by 3D object trackers [82, 83]. However, by defining vehicle motion solely through these tracking poses, rendering quality becomes highly dependent on pose accuracy. While Street Gaussians optimizes object poses during training, it struggles with detection failures and large pose errors, as shown in Fig. 1. Unfortunately, 3D trackers often struggle to generalize effectively across different scenarios, primarily due to the scarcity of open-source largescale 3D datasets [79, 103]. Collecting ground-truth poses for 3D tracking is time-consuming and costly, resulting in relatively few open-source datasets for autonomous driving. For example, even widely-used datasets such as Waymo [76], Nuscenes [6], KITTI [23, 49], Pandaset [86] collectively contain fewer than 4,000 annotated scenes, limiting the diversity and scale needed for robust 3D tracking. In contrast, 2D data is far easier to collect and annotate, leading to an abundance of 2D datasets. The BDD100K dataset [95] alone includes annotations for 100,000 driving scenes, and additional large-scale 2D perception datasets, such as COCO [50], LVIS [28], and OpenImages [41], bring the total number of annotated scenes into the millions. This wealth of 2D data has enabled 2D foundation models to achieve strong generalization across a variety of tasks, including embedding extraction [64, 68], detection [8, 62], segmentation [12, 40], and tracking [84]. Nonetheless, how to use these accurate and robust 2D trackers to effectively model dynamic objects in 3D is still an open challenge for street scene reconstruction. + +In this paper, a novel architecture is designed to achieve high-fidelity novel view synthesis performance in street scenarios for autonomous driving applications. Unlike previous approaches such as Driving Gaussian [106] and Street Gaussian [89] that rely on ground truth vehicle poses or poses predicted by 3D trackers, the first contribution of this work is a stable object tracking module. This module leverages associations from 2D deep trackers within a 3D object fusion framework to enhance robustness and accuracy. Specifically, we integrate 2D tracking outputs with LiDAR data to trace the trajectory of each vehicle in 3D. We then incrementally reconstruct the point cloud of each vehicle frame-by-frame in the canonical space, estimating its pose by aligning successive frames with this canonical model. This approach eliminates the reliance on 3D trackers and enhances robustness across diverse environments. + +Although 2D tracking models demonstrate stronger generalization ability, tracking errors still can exist, especially under adverse conditions such as severe occlusions or distant objects. Since moving the object points solely based on the tracked object pose directly exposes any tracking errors, we aim to go beyond a straightforward reliance on tracked trajectories. Instead, we propose to learn point mo + +tion from the predicted trajectory, equipping the model to autonomously identify and correct tracking inaccuracies, recover missed detections, and infer motion in new time steps. An implicit representation is essential for this purpose: 1) It enables the model to refine trajectories without being constrained by bounding box tracks, facilitating smoother and more continuous motion that can adaptively respond to changes. 2) An implicit feature space offers versatility to move each point of an object in a different way, recognizing that vehicles are not strictly rigid objects — for example, doors can open or close. This approach makes it possible to capture subtle, dynamic changes within objects, ultimately producing more accurate reconstructions and enhancing the robustness of novel view synthesis in challenging scenarios. + +To this end, we leverage HexPlane representation [7] following 4DGS [81]. HexPlane stores motion-related features by decomposing the 4D spatial-temporal space into six 2D learnable feature planes. A decoder then utilizes these features to predict deformation offsets, dynamically adjusting 3D Gaussians over time. However, 4DGS relies solely on image reconstruction loss for supervision, which is insufficient for handling rapid object motion. When a Gaussian's initial position is far from its current location, its projection on the image may fall outside the object's actual area, preventing gradient propagation and hindering optimization. This issue is evident in S3Gaussian [32], which struggles with moving cars due to the lack of explicit motion supervision. To address this, we introduce a training strategy that supervises learned point motion using the predicted trajectory, providing explicit 3D supervision to guide HexPlane in capturing motion dynamics more accurately. + +Our main contributions are summarized as follows: + +- We eliminate the need of 3D tracker for street scene reconstruction by introducing a robust object tracking module which leverages 2D foundation model, achieving superior generalization ability across diverse scenarios. +- We introduce a motion learning framework to learn from the predicted trajectory in an implicit feature space, enabling it to automatically correct pose errors and infer motion for novel time steps. +- We outperform existing methods on Waymo-NOTR and KITTI datasets without relying on ground truth annotations. + +# 2. Related Works + +# 2.1. 3D Gaussian Splatting for Dynamic Scene + +3D Gaussian Splatting (3DGS) [37, 44] has advanced scene reconstruction by enabling high-quality, real-time rendering with 3D Gaussians and efficient splat-based rasterization, reducing computation and parameters compared to NeRF-based methods [1-3, 61, 63] and other representa + +tions [15, 26, 27, 30, 45, 75, 102]. While originally designed for static scenes, 3DGS has been adapted for dynamic scenes [48, 53, 57, 74, 81, 92, 93]. Dynamic3DGS [57] directly stores information for each 3D Gaussian at every timestamp, and Yang et al. [92] approximate the spatiotemporal 4D volume by optimizing 4D Gaussian primitives. Deformable-3DGS [93] and GauFRe [48] employ deformation fields to model motions, while 4DGS [81] introduces HexPlane representation [7] to store spatial-temporal features efficiently. Gaussian Marbles [74] leverages isotropic "marbles" and a divide-and-conquer trajectory learning algorithm. Despite these advances, challenges persist in high-speed driving scenarios with rapid object motion. To tackle this challenge, we leverage a robust object tracking module to guide the motion learning process. + +# 2.2. Street Scene Reconstruction + +Existing autonomous driving simulation engines [20, 43, 52, 71] face high manual effort in creating virtual environments and a lack of realism in the generated data. The creation of high-fidelity simulations from driving logs is therefore essential for advancing closed-loop training and testing. Recent works [65, 72, 77, 78, 85, 87, 90, 91] have continuously made improvements to NeRF [61] to model street scenes dynamically. Despite the progress, NeRF-based methods remain computationally expensive and require densely overlapping views. Building on the effective 3DGS approach [37] for scene reconstruction, several 3DGS-based methods [10, 67, 89, 106] have emerged. PVG [10] model dynamic scenarios by using periodic vibration-based temporal dynamics. Driving Gaussian [106] introduces incremental static Gaussians and composite dynamic Gaussian graph. Street Gaussians [89] equips Gaussians with semantic logits, and optimize dynamic parts using tracked poses from 3D tracker. AutoSplat [38] enforces geometric constraints in road and sky regions for multi-view consistency. Among these approaches, most of them requires ground truth object pose [10, 38, 106] or the 3D tracker [89]. However, manual annotation is laborious and the 3D tracker is lack in generalization ability, which limits their applications in diverse scenarios. In contrast, S3Gaussian [32] models dynamics in a self-supervised manner using dynamic Gaussians from 4DGS [81], despite struggling with dynamic object modeling due to the lack of explicit motion supervision. We address this challenge by introducing a robust object tracking strategy based on a 2D foundation model [84] and apply motion supervision from the predicted object trajectory. + +# 2.3. 2D and 3D trackers + +Multiple Object Tracking (MOT) aims at locating multiple objects in each frame, and establishing correspondences between them across frames within input videos + +[4, 5, 36, 58, 98]. The reliance of most MOT methods on object detection [4, 19, 58, 66, 98, 104] makes detection accuracy crucial for MOT performance. Recent advances in 3D object detection are promising [11, 18, 34, 42, 51, 54, 56, 60, 80, 94, 99-101], but these models often generalize poorly due to limited and biased datasets [21, 55, 73, 97]. Existing open-source autonomous driving datasets lack scale and show regional or environmental biases, such as vehicle density and weather conditions [6, 23, 59, 76]. The main reason is that building large-scale, well-annotated multimodal datasets is costly [6, 49, 59, 69]. In contrast, 2D image data is easier to capture and annotate, enabling extensive dataset collections [9, 17, 41, 95]. Visual foundation models trained on large-scale 2D data have shown strong generalization [12, 40, 64, 68, 84]. We thus opt to use 2D object tracker [84] to locate the dynamic objects. + +# 3. Our Method + +# 3.1. Prerequisites: 3D Gaussian Splatting + +3D Gaussian Splatting (3DGS) [37] represents a 3D scene explicitly as a collection of 3D Gaussian primitives. Each Gaussian is defined by its center position $\mathcal{X} \in \mathbb{R}^3$ and covariance $\Sigma \in \mathbb{R}^{3 \times 3}$ . To enforce semi-positive definiteness, the covariance matrix is further factorized into a scaling vector $S \in \mathbb{R}^3$ and a rotation matrix $R \in SO(3)$ by: + +$$ +\Sigma = R S S ^ {\top} R ^ {\top} \tag {1} +$$ + +Additionally, each Gaussian is described with its opacity $o \in \mathbb{R}$ and view-dependent color defined by spherical harmonic (SH) coefficient $\mathcal{C} \in \mathbb{R}^k$ , where $k$ represents numbers of SH functions. + +During rendering of novel views, differential splatting is applied to 3D Gaussians within the camera planes. The blending of $N$ ordered points that overlap a pixel is given by the formula: + +$$ +C = \sum_ {i \in N} c _ {i} \alpha_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}), \tag {2} +$$ + +where $\alpha_{i}$ and $c_{i}$ represents the opacity and color of the ith splatted Gaussian, which is computed from per-point opacity and SH coefficients (see [37] for details). + +# 3.2. Overview + +Our method takes as input multi-view images $\mathbf{I}_t^{(j)}$ from multiple cameras placed around the vehicle, each indexed by time step $t$ and camera index $j$ , along with intrinsic $\mathbf{K}^{(j)}$ and extrinsic $\mathbf{E}_t^{(j)}$ matrices for each view. Additionally, a top-mounted LiDAR provides 3D point clouds $\mathbf{L}_t$ for each frame. Using this multi-sensor data, our approach reconstructs the 3D scene and synthesizes novel views from any + +![](images/bb66c5e49d0b54cdbedde8e690b18e2232ffc8dcb586ecf24b921d2354e23ad2.jpg) +Figure 2. Overview of our method. To overcome the limited generalization of 3D object trackers, we introduce a robust object tracking module based on a 2D object tracker [84]. We integrate 2D tracking with LiDAR data to segment the object's point cloud and incrementally reconstruct the object model in canonical space frame-by-frame. The canonical model is used to estimate object pose $\mathbf{T}_t$ and serves as the initialization for Gaussian optimization. We model dynamic objects with isotropic Gaussian marble [74] to simplify optimization. Our approach learns point motion within the HexPlane [7] feature space, enabling tracking error correction and missed detection recovery. Decoders $D_{\mathcal{X}}, D_{\mathcal{C}}$ predict point motion $\Delta \mathcal{X}_t$ and color change $\Delta \mathcal{C}_t$ based on the HexPlane features $f(x,y,z,t)$ , with the predicted trajectory providing supervision. Finally, novel view synthesis is performed by splatting the deformed Gaussian onto the image plane. + +desired camera pose and time frame, without reliance on 3D object trackers or ground truth object trajectories. + +We present an overview of our method in Fig. 2. Visual images and point clouds are obtained from the sensor setup. For images, static and dynamic components are analyzed first. We employ Mask2Former [12] to segment the scene into static and dynamic parts. Specifically, the dynamic part includes humans (pedestrians, cyclists, etc.) and vehicles (cars, trucks, etc.), and all the other objects are categorized as static. We model the static part as in 3DGS [37] and focus on modeling dynamic objects in the following sections. To enhance robustness, we propose an object tracking strategy by leveraging a robust 2D tracker [84] and the LiDAR point cloud, providing accurate initial point clouds and object trajectories for subsequent Gaussian optimization (cf. Sec 3.3). We leverage an implicit representation HexPlane [7] following 4DGS [81] to learn a continuous and smooth per-point motion based on the object trajectory (cf. Sec 3.4). The optimization objective is described in Sec 3.5. + +# 3.3. Tracking Cars in 3D with Robust 2D Tracker + +To improve the robustness of novel view synthesis for diverse scenarios, the first challenge is how to track and build vehicles in 3D based on the 2D tracking results. + +We employ GLEE [84] to track all vehicles within the 2D image plane, obtaining 2D object trajectories in each camera view. Each trajectory includes the 2D segmentation masks of the object over a time period. To lift 2D trajectories into 3D, we first re-project the LiDAR point clouds onto the image plane and assign points to objects based on their presence within the corresponding 2D segmentation masks $\mathcal{M}$ . This yields a segmented object point clouds for each + +camera view via the following function: + +$$ +\mathcal {O} _ {t} ^ {(j)} = \left\{\mathbf {P} \mid \Pi^ {- 1} \left(\mathbf {P}, \mathbf {K} ^ {(j)}, \mathbf {E} _ {t} ^ {(j)}\right) \in \mathcal {M}, \mathbf {P} \in \mathbf {L} _ {t} \right\}, \tag {3} +$$ + +where $j$ is the camera index, $t$ is the time index, $\mathbf{P}$ is a LiDAR point and $\Pi(\cdot)$ is the re-projection function. Given the slight misalignment between the LiDAR point cloud and RGB image, points near the edges of segmentation masks often fall outside the object boundaries. To address this issue, we perform outlier removal to improve stability in subsequent reconstruction based on the point distance to the point cloud center. We then associate the same objects across different camera views. Two point clouds in different viewpoints are considered belonging to the same object if they share more than 50 points in one time step. This association process results in a set of associated partial object point cloud in different time steps. + +After obtaining the 3D partial objects in different time steps, the next step is to associate these partial point cloud into a unified and complete model. We initialize the 3D reconstruction of each object from the first frame which it appears. For simplicity, we assume an object is visible between frames $0$ and $T$ and denote the object point cloud at time $t$ as $\mathcal{O}_t$ . Our goal is to obtain a temporally consistent reconstruction $\mathcal{O}$ and the object pose in each time frame. This reconstruction is done frame-by-frame incrementally. + +Starting with the initial frame, we add $\mathcal{O}_0$ to $\mathcal{O}$ . For each subsequent frame, we apply Iterative Closest Point (ICP) for the alignment of $\mathcal{O}_t$ to $\mathcal{O}$ by extracting the relative pose $\mathbf{T}_t$ . The overlap between $\mathcal{O}_t$ and $\mathcal{O}$ is given by: + +$$ +\tau_ {o v e r l a p} = \mathcal {O} \cap \mathbf {T} _ {t} ^ {- 1} \mathcal {O} _ {t}, \quad \mathbf {T} _ {t} = \mathrm {I C P} (\mathcal {O} _ {t}, \mathcal {O}). \quad (4) +$$ + +If ICP reveals an overlap $\tau_{\text{overlap}}$ of more than $30\%$ with $\mathcal{O}$ , we record the object pose $\mathbf{T}_t$ and update $\mathcal{O}$ by integrating $\mathbf{T}_t^{-1}\mathcal{O}_t$ . If $\tau_{\text{overlap}}$ is between $10\%$ and $30\%$ , we only record the object pose $P_t$ without updating $\mathcal{O}$ . If $\tau_{\text{overlap}}$ is lower than $10\%$ , we assume a 2D tracking failure for that frame and discard it to filter out tracking errors. This process is repeated for each frame using the last recorded pose as the initial guess for ICP to accelerate convergence. Through this approach, we reconstruct all objects and retrieve their poses in each frame, providing a robust initialization for subsequent 3D Gaussian optimization. + +# 3.4. Learning Point Motion + +A common approach to model object motion in prior works is directly transforming the object points using the given object pose. However, despite the strong generalization capabilities of 2D trackers, detection failures are inevitable in challenging scenarios, such as when objects are heavily occluded or located at great distances. Relying solely on object poses derived from these erroneous or missing detections can easily lead to rendering failures. Additionally, treating cars as rigid objects does not adequately address corner cases, such as when a car door is open or closed. Furthermore, these methods lack the ability to infer an object motion at arbitrary time stamps. + +To address these limitations, we seek for a motion modeling approach to enhance robustness and flexibility. Instead of explicitly using object pose as rigid transformation, we learn the per-point motion in a pre-defined feature space. In this feature space, the object motion can be optimized through both explicit guidance and photometric loss. Moreover, the object motion can be interpolated through time and space in this feature space to compensate for missing detections. To this end, we leverage HexPlane representation [7] as in 4DGS [81] to efficiently capture spatial and temporal information by decomposing the 4D feature voxels into six learnable feature planes. This representation satisfies all our requirements and is also memory efficient. + +We use three planes: $P_{xy}$ , $P_{yz}$ , and $P_{xz}$ for the spatial dimensions, and another three planes $P_{xt}$ , $P_{yt}$ , and $P_{zt}$ for the spatial-temporal features. Additionally, the Hexplane includes multiple resolution levels which is formulated as: + +$$ +\left\{P _ {i j} ^ {\rho} \in \mathbb {R} ^ {d \times \rho r _ {i} \times \rho r _ {j}} | (i, j) \in \mathcal {P}, \rho \in \{1, 2 \} \right\}, \tag {5} +$$ + +where $\mathcal{P} = \{(x,y),(x,z),(y,z),(x,t),(y,t),(z,t)\}$ , $d$ is the feature dimension, $\rho$ denotes the upsampling scale, and $r$ is the base resolution. + +Given the center position of a Gaussian $(x,y,z)$ and the time step $t$ , features in all six planes are queried and combined via a small MLP $\phi_d$ as follows: + +$$ +f (x, y, z, t) = \phi_ {d} \left(\bigcup_ {\rho} \prod_ {(i, j) \in \mathcal {P}} \pi \left(P _ {i j} ^ {\rho}, \psi_ {i j} ^ {\rho} (x, y, z, t)\right)\right), \tag {6} +$$ + +where $\psi_{ij}^{\rho}(x,y,z,t)$ projects the 4D coordinate onto the respective plane, and $\pi$ performs bilinear interpolation on the voxel features at each point. The features of all planes are combined using the Hadamard product. + +Finally, two MLP decoders $D_{\mathcal{X}}, D_{c}$ are utilized to predict the point motion $\Delta \mathcal{X}_t$ and color change $\Delta \mathcal{C}_t$ as $\Delta \mathcal{X}_t = D_{\mathcal{X}}(f(x,y,z,t))$ and $\Delta \mathcal{C}_t = D_{\mathcal{C}}(f(x,y,z,t))$ , respectively. The deformed 4D Gaussians are formulated as: $\mathcal{G}_t = \{\mathcal{X} + \Delta \mathcal{X}_t, \mathcal{C} + \Delta \mathcal{C}_t, S, R, o\}$ . + +To allow the HexPlane feature to learn point motion from the predicted pose, we define the motion loss as: + +$$ +\mathcal {L} _ {\text {m o t i o n}} = \operatorname {a v g} _ {\mathcal {X} \in \mathcal {O}} | \Delta \mathcal {X} _ {t} - (\mathbf {T} _ {t} \mathcal {X} - \mathcal {X}) |, \tag {7} +$$ + +where $\mathcal{X}$ is the center position of a Gaussian in object $\mathcal{O}$ and avg is the average operator. This loss encourages the predicted point motion $\Delta \mathcal{X}_t$ to align with the predicted pose. We apply this loss only for the first $40\%$ iterations to provide a strong initial motion prior. The loss is subsequently removed, allowing the network to adjust and potentially compensate for pose errors and detection failures. + +Following [74], we adopt isotropic Gaussian marbles for dynamic points to reduce degrees of freedom and simplify optimization. In this setup, the rotation of each Gaussian is represented by the identity matrix with identical scales across all three dimensions. The spherical harmonics coefficients are also limited to three dimensions, providing view-consistent color. This approach ensures that all deformations are captured purely by point motion and color changes, enforcing strong constraints and enhancing robustness in novel view synthesis. + +Ours vs. S3Gaussian. Although S3Gaussian [32] also uses Hexplane representation [7] and the deformation network from 4DGS [81], our design serves a different purpose: S3Gaussian learns point motion without extra supervision, while we use 4DGS to refine motion and address detection failures. Consequently, S3Gaussian becomes susceptible to unsatisfactory results with rapid car motion due to the lack of explicit guidance. In contrast, our approach enables high-quality reconstruction of dynamic objects. + +# 3.5. Optimization Objective + +Besides the motion loss introduced above, the loss function comprises five components to collectively optimize the scene representation and the point motion. The primary loss $\mathcal{L}_{\mathrm{rgb}}$ is an L1 loss measuring the photometric difference between rendered and ground truth images while $\mathcal{L}_{\mathrm{ssim}}$ evaluates their structural similarity. The L1 loss between the rendered depth map and the depth estimated from the LiDAR point cloud $\mathcal{L}_{\mathrm{depth}}$ is used to supervise the Gaussian positions. Following K-Planes [70], a grid-based total variation loss $\mathcal{L}_{\mathrm{tv}}$ is introduced. Recognizing that the color of most dynamic points in the scene are unchanged, a L1 regulariza + +
TaskScene ReconstructionNovel View Synthesis
MethodExtra InputPSNR↑SSIM↑LPIPS↓DPSNR↑DSSIM↑PSNR↑SSIM↑LPIPS↓DPSNR↑DSSIM↑
3DGS [37]N/A25.770.8330.16020.260.60424.460.8020.17018.120.521
EmerNeRF [90]N/A28.160.8060.22824.320.68225.140.7470.31323.490.660
S3Gaussian [32]N/A31.350.9110.10626.020.78327.440.8570.13722.920.680
MARS [85]GT pose28.240.8660.25223.370.70126.610.7960.30522.210.697
StreetGS [89]3D tracker29.170.8730.13827.780.81826.980.8380.14924.620.742
Ours2D tracker32.560.9360.05928.510.86828.850.8670.08825.580.779
+ +Table 1. Comparison with state-of-the-art methods on Waymo-NOTR dataset. StreetGS represents Street Gaussian [89]. The best and the second best results are denoted by pink and blue. + +tion loss $\mathcal{L}_{\mathrm{color - reg}}$ is applied to the deformation network by minimizing $\Delta C$ . + +The total loss function is thus defined as: + +$$ +\begin{array}{l} \mathcal {L} = \lambda_ {\mathrm {r g b}} \mathcal {L} _ {\mathrm {r g b}} + \lambda_ {\text {d e p t h}} \mathcal {L} _ {\text {d e p t h}} + \lambda_ {\text {s s i m}} \mathcal {L} _ {\text {s s i m}} + \lambda_ {\mathrm {t v}} \mathcal {L} _ {\mathrm {t v}} \tag {8} \\ + \lambda_ {\text {c o l o r - r e g}} \mathcal {L} _ {\text {c o l o r - r e g}} + \lambda_ {\text {m o t i o n}} \mathcal {L} _ {\text {m o t i o n}}, \\ \end{array} +$$ + +where the weights are assigned as follows: $\lambda_{\mathrm{rgb}} = 1.0$ $\lambda_{\mathrm{depth}} = 1.0$ $\lambda_{\mathrm{ssim}} = 0.1$ $\lambda_{\mathrm{tv}} = 0.1$ $\lambda_{\mathrm{color-reg}} = 0.01$ and $\lambda_{\mathrm{motion}} = 1$ + +The optimization process is divided into two stages. In the first stage, we render using only static Gaussians and supervise only the static regions of the images over 20,000 iterations. This phase focuses on reconstructing the static background and stabilizes the training process. In the second stage, we train all Gaussians on the whole images for 50,000 iterations. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. We leverage Waymo-NOTR dataset [76, 90] and KITTI benchmark [23, 24]. The NeRF On-The-Road (NOTR) dataset, introduced by EmerNeRF [90], is a subset of the Waymo Open dataset [76] and includes diverse challenging driving scenarios, such as high-speed, exposure mismatch, and various weather conditions. We use the dynamic32 subset, consisting of 32 dynamic scenes, for evaluating dynamic reconstruction. Following EmerNeRF's setup [90], we use three frontal camera images resized to $960 \times 640$ . For scene reconstruction, all image frames are used for training and evaluation; for novel view synthesis, every 10th time step is excluded for evaluation [90]. For the KITTI dataset, we follow the setup of MARS [85] to use $75\%$ or $50\%$ of the images for training with every 4th or every 2nd frame held out for testing, respectively. + +Implementation Details. We use the Adam optimizer [39] with the same learning rate schedule as in 3DGS [37]. For long-sequence scene reconstruction, we follow S3Gaussian + +[32] and segment the scene into multiple clips. The multi-resolution HexPlane encoder has a base resolution of 64, upsampled by factors of 2 and 4 as in [81], and other hyperparameters match those in 3DGS [37]. As LiDAR points lack data for the sky region, we add a plane of points above the maximum height of the scene to represent it following [38]. All experiments are run on a single NVIDIA RTX 3090 GPU, with training taking about 2 hours per video clip and inference speed at 100 FPS at $960 \times 640$ resolution. + +Baseline Methods. We compare our approach against state-of-the-art methods, including NeRF-based methods (MARS [85], NSG [65], EmerNeRF [90]) and Gaussian-based methods (3DGS [37], StreetGaussian [89], S3Gaussian [32]). To ensure fair comparisons, we apply LiDAR point cloud initialization and add depth regularization to 3DGS. For the Waymo-NOTR dataset, we borrow results of MARS [85] and EmerNeRF [90] from S3Gaussian [32]. We evaluate the 3D-tracker-based method, Street Gaussians [89], using detection results from VoxelNext [11] and associate the detections with SimpleTrack [66]. Our goal is to assess the generalization ability of 3D-tracker-based methods in scenarios without ground truth tracking labels. To emulate a real-world scenario where a 3D tracker trained on a public benchmark dataset is applied to a novel environment, we use pretrained weights from nuScenes [6]—one of the largest and most diverse 3D datasets—and evaluate the model on the Waymo-NOTR dataset. Since NOTR is part of the training set for the Waymo perception task, using Waymo pretrained weights would introduce domain overlap and bias the evaluation, making the nuScenes weights a fairer choice. For a fair comparison, we use pretrained 2D tracker weights from [84] for our method, which have not been trained on the Waymo or KITTI datasets. Please refer to the supplementary material for discussion of the choice of 3D trackers. For the KITTI dataset, we borrow the results of all other methods from Street Gaussians. Please note that Street Gaussians leverages GT tracking annotations on KITTI dataset. + +![](images/adb78bd5aae696541ac6b36c87113c0ef7399b603f4f1c40e6ccadaf9b589bde.jpg) +GT + +![](images/d2c05fd3d944350e55146f6f7b856b26cdfbb091f6e06300cbe67c6318694edc.jpg) +S3Gaussian + +![](images/d59d369509dd33f47f225170a24d9ed605933171dcfcb29c6e367571db7f671d.jpg) +Street Gaussian + +![](images/81b8cdadfe611141909cc8417c57d85ba2b0f10ab5df4d0b89712c8784c8cdbf.jpg) +Ours + +![](images/5df529ecf0004a6a96c2ad0dfa62667653d39dc362b50ac96b5a6b16a04e5888.jpg) + +![](images/4924e0c1f01051557eaa8e9143c81d0219a7f3c10f565c3bf2e65a5bcea0b94e.jpg) + +![](images/8b0e8a9311ede3f05844cd7ca7aa62a2b0e7b1fdcdd35ca07286b0b96f9dabac.jpg) + +![](images/6939d243ac553a629ef06e0efd4763649494891d49911fe4b09ce46bafca7ac1.jpg) + +![](images/9605e61606317d5c9f7d250fbda8f37b42cd7fed48d3857f46ce00870f1cd617.jpg) + +![](images/cc6dfeb6b8c089369b408ccbd18a17f91dd2c1b2debcf9db41116ea2d8885c93.jpg) + +![](images/279869d5a6455c6ef8c58ab0617acf743196665c4b9bc249e21ab5dfbeaf002b.jpg) + +![](images/9888d28bf48e41917c575bec7dba41b21413cc1366248ce6b1becb2f13ea4ff2.jpg) + +![](images/b066907af73fbc81b4b4cc5c84def1915148ce1ad7efc96ad6d647164c72ac6c.jpg) + +![](images/d7551168474300e162ab28a1904cefcbe6d0d01ccf6f5e50f087df9ae7f777fa.jpg) + +![](images/c47fc0aea547c28ac8c75c279e4c25ac091b2c7511cde40fe6144e8d3948ce3c.jpg) + +![](images/6e9ac1194a1144189f517d5e3b60d1bbaaa1995f0e562474fa51be0620ef1556.jpg) + +![](images/da6ba43bf89505272eb6fe1987d545187cce8a47bf812aa092b1e31ad4152e97.jpg) +Figure 3. Qualitative comparison of novel view synthesis on NOTR dataset. Best viewed with zoom. + +![](images/b82259fa799bcb919099bbae0b7f78f794afc6264b72b788c38f760334efa06a.jpg) + +![](images/030639fa14ec1fd5122fbc19acabdc8bb5d4284dc864e704b73b195493025a7b.jpg) + +![](images/90d5cd84366e64bfa75e381301ea0d16a1cb7b1b6cf05cfc903ef4f635ef1eb8.jpg) + +Metrics. We leverage peak signal-to-noise ratio (PSNR), structural similarity index (SSIM) and learned perceptual image patch similarity (LPIPS) to evaluate the rendering quality. Additionally, following EmerNeRF [90], we apply DPSNR and DSSIM metrics to dynamic objects by projecting their ground truth 3D bounding boxes onto the 2D image plane and calculating pixel loss within these regions. + +# 4.2. Comparisons with the State-of-the-art + +On the NOTR dataset, our method outperforms all competitors across every metric, as shown in Table 1. Our approach sets a new state-of-the-art in both the scene reconstruction and novel view synthesis tasks. Specifically, our method outperforms S3Gaussian [32] by $2.66\mathrm{dB}$ in DPSNR and 0.099 in DSSIM for novel view synthesis. It also surpasses Street Gaussians [89] by $1.87\mathrm{dB}$ in PSNR, mainly due + +to the limited generalization of 3D trackers used in Street Gaussians. Despite optimizing object poses, Street Gaussians struggles with large pose errors and detection failures, leading to inferior performance. Additionally, we outperform NeRF-based methods like EmerNeRF and MARS. + +We present qualitative comparison with S3Gaussian [32] and Street Gaussians [89] in Fig. 3. Results from S3Gaussian show that using 4DGS without explicit motion guidance results in weaker performance when handling moving vehicles (e.g. row 2, 3, 5). Street Gaussians suffers from tracking errors of 3D trackers, leading to erroneous reconstructions (e.g. row 1, 3) or entirely missed objects (e.g. rows 2, 4, and 5). In contrast, our approach performs robustly across diverse scenarios owing to our tracking strategy leveraging 2D foundation model and robust motion + +
KITTI-75%KITTI-50%
MethodExtra InputPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
3DGS [37]N/A19.190.7370.17219.230.7390.174
NSG [65]GT pose21.530.6730.25421.260.6590.266
MARS [85]GT pose24.230.8450.16024.000.8010.164
Street Gussians [89]GT pose25.790.8440.08125.520.8410.084
Ours2D tracker [84]25.490.8890.06325.110.8770.067
+ +Table 2. Comparison with state-of-the-art methods on KITTI dataset. The best and the second best results are denoted by pink and blue. + +
MethodPSNRSSIMDPSNRDSSIM
Aw/o track25.250.76121.710.662
Bw/o Lmotion28.390.84324.540.719
Cw/o iso.28.580.84725.350.765
D3D tracker28.020.83524.610.746
ED w/o Lmotion27.540.82123.530.703
FGT pose28.980.87325.710.787
GUNINEXT28.760.86425.430.772
HOurs28.850.86725.580.779
+ +Table 3. Ablation study of Novel View Synthesis on NOTR dataset. + +learning module. Please refer to the supplementary material for additional video results and editing examples. + +The results on the KITTI dataset are presented in Table 2. Our method outperforms the NeRF-based NSG and MARS across all metrics. In the KITTI-75% setting, our approach achieves a 0.045 higher SSIM and a 0.018 lower LPIPS compared to Street Gaussians, although its PSNR is 0.3 dB lower. A similar trend is observed in the KITTI-50% setting. This lower PSNR is primarily because Street Gaussians leverages labor-intensive ground truth tracking annotations, whereas our method uses a generalized 2D tracker—yet still attains comparable performance. + +# 4.3. Ablation Studies + +We conduct an ablation study on the NOTR dataset and the results are presented in Tab. 3. + +Effect of object tracking module. In Tab. 3 (A), we model dynamic points using 4DGS [81] without performing any form of tracking, and observe significant drops across all metrics. This experiment highlights that the sole use of 4DGS does not provide sufficient accuracy for effective motion modeling in autonomous driving scenarios. + +Effect of motion learning strategy. In Tab. 3 (B), we use the reconstruction of the tracking module for initialization + +but remove the motion loss introduced in Eq. 7, resulting in a $1.04\mathrm{dB}$ drop in DPSNR. This result highlights the importance of our motion learning strategy. A comparison with rigid-transformation-based motion modeling [89] is provided in the supplementary material. + +Effect of isotropic Gaussian marbles. In Tab. 3 (C), we substitute isotropic Gaussian marbles with the original anisotropic Gaussian ellipsoids for dynamic points, which leads to a decrease in DPSNR and DSSIM. + +3D tracker / GT pose vs. Our object tracking module. In Tab. 3 (D), we use the 3D tracker [11, 66] to reconstruct dynamic objects and supervise motion learning, resulting in a 0.97 dB DPSNR decrease. Removing motion loss leads to a further 1.08 dB drop (Tab. 3 (E)), showing that our motion learning strategy compensates for tracking errors, though these errors still impact novel view synthesis, emphasizing the need for a more generalizable 2D tracker. Using ground-truth trajectories (Tab. 3 (F)) shows only a small gap, confirming the effectiveness of our 2D tracker in motion learning. See supplementary for tracking error analysis. + +Choice of 2D tracker. To demonstrate the robustness of our method with respect to the choice of 2D tracker, we employ UNINEXT [88] for 2D tracking. As shown in Table 3 (G), our method maintains stable performance. + +# 5. Conclusion + +In this paper, we introduce a novel framework for robust dynamic 3D street scene reconstruction that eliminates the reliance on 3D object trackers. Addressing the generalization limitations of 3D trackers, we propose a robust object tracking strategy based on a 2D foundation model. Our framework also features a motion learning module within an implicit feature space to handle inevitable tracking errors by autonomously refining pose inaccuracies and recovering missed detections. Experiments on the Waymo-NOTR and KITTI datasets demonstrate its adaptability and superior performance. Limitations and failure cases are discussed in the supplementary material. + +# Acknowledgment + +This work was supported by the National Key R&D Program of China under Grant 2018AAA0102801, PCL-CMCC Foundation for Science and Innovation Grant No. 2024ZY1C0040, and the National Research Foundation, Singapore, under its NRF-Investigatorship Programme (Award ID. NRF-NRFI09-0008). + +# References + +[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5855–5864, 2021. 2 +[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470-5479, 2022. +[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 2 +[4] Mk Bashar, Samia Islam, Kashifa Kawaakib Hussain, Md Bakhtiar Hasan, ABM Rahman, and Md Hasanul Kabir. Multiple object tracking in recent times: A literature review. arXiv preprint arXiv:2209.04796, 2022. 3 +[5] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In 2016 IEEE international conference on image processing (ICIP), pages 3464-3468. IEEE, 2016. 3 +[6] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 3, 6 +[7] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. 2, 3, 4, 5 +[8] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2 +[9] Zhengping Che, Guangyu Li, Tracy Li, Bo Jiang, Xuefeng Shi, Xinsheng Zhang, Ying Lu, Guobin Wu, Yan Liu, and Jieping Ye. D2-city: a large-scale dashcam video dataset of diverse traffic scenarios. arXiv preprint arXiv:1904.01975, 2019.3 +[10] Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene + +reconstruction and real-time rendering. arXiv preprint arXiv:2311.18561, 2023. 3 +[11] Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21674-21683, 2023. 1, 3, 6, 8 +[12] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 2, 3, 4 +[13] Jie Cheng, Yingbing Chen, Qingwen Zhang, Lu Gan, Chengju Liu, and Ming Liu. Real-time trajectory planning for autonomous driving with gaussian process and incremental refinement. In 2022 International Conference on Robotics and Automation (ICRA), pages 8999-9005. IEEE, 2022. 1 +[14] Jie Cheng, Xiaodong Mei, and Ming Liu. Forecast-mae: Self-supervised pre-training for motion forecasting with masked autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8679-8689, 2023. 1 +[15] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. ACM Transactions on Graphics (ToG), 34(4):1-13, 2015. 3 +[16] Daniel Dauner, Marcel Hallgarten, Andreas Geiger, and Kashyap Chitta. Parting with misconceptions about learning-based vehicle motion planning. In Conference on Robot Learning, pages 1268-1281. PMLR, 2023. 1 +[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 3 +[18] Yan Di, Ruida Zhang, Zhiqiang Lou, Fabian Manhardt, Xiangyang Ji, Nassir Navab, and Federico Tombari. Gpvpose: Category-level object pose estimation via geometry-guided point-wise voting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6781-6791, 2022. 3 +[19] Shuxiao Ding, Eike Rehder, Lukas Schneider, Marius Cordts, and Juergen Gall. 3dmotformer: Graph transformer for online 3d multi-object tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9784-9794, 2023. 3 +[20] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 3 +[21] George Eskandar. An empirical study of the generalization ability of lidar 3d object detectors to unseen domains. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23815-23825, 2024. 1, 3 +[22] Tobias Fischer, Jonas Kulhanek, Samuel Rota Bulo, Lorenzo Porzi, Marc Pollefeys, and Peter Kontschieder. + +Dynamic 3d gaussian fields for urban areas. arXiv preprint arXiv:2406.03175, 2024. 1 +[23] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 2, 3, 6 +[24] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 6 +[25] Junru Gu, Chenxu Hu, Tianyuan Zhang, Xuanyao Chen, Yilun Wang, Yue Wang, and Hang Zhao. Vip3d: End-to-end visual trajectory prediction via 3d agent queries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5496-5506, 2023. 1 +[26] Kaiwen Guo, Feng Xu, Yangang Wang, Yebin Liu, and Qionghai Dai. Robust non-rigid motion tracking and surface reconstruction using 10 regularization. In Proceedings of the IEEE International Conference on Computer Vision, pages 3083-3091, 2015. 3 +[27] Kaiwen Guo, Peter Lincoln, Philip Davidson, Jay Busch, Xueming Yu, Matt Whalen, Geoff Harvey, Sergio Orts-Escolano, Rohit Pandey, Jason Dourgarian, et al. The relightables: Volumetric performance capture of humans with realistic relighting. ACM Transactions on Graphics (ToG), 38(6):1-19, 2019. 3 +[28] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 2 +[29] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 1 +[30] Tao Hu, Tao Yu, Zerong Zheng, He Zhang, Yebin Liu, and Matthias Zwicker. Hvtr: Hybrid volumetric-textural rendering for human avatars. In 2022 International Conference on 3D Vision (3DV), pages 197-208. IEEE, 2022. 3 +[31] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1 +[32] Nan Huang, Xiaobao Wei, Wenzhao Zheng, Pengju An, Ming Lu, Wei Zhan, Masayoshi Tomizuka, Kurt Keutzer, and Shanghang Zhang. S3gaussian: Self-supervised street gaussians for autonomous driving. arXiv preprint arXiv:2405.20323, 2024. 2, 3, 5, 6, 7 +[33] Yanjun Huang, Jiatong Du, Ziru Yang, Zewei Zhou, Lin Zhang, and Hong Chen. A survey on trajectory-prediction methods for autonomous driving. IEEE Transactions on Intelligent Vehicles, 7(3):652-674, 2022. 1 +[34] Ziqin Huang, Gu Wang, Chenyangguang Zhang, Ruida Zhang, Xiu Li, and Xiangyang Ji. Givepose: Gradual intraclass variation elimination for rgb-based category-level object pose estimation. In Proceedings of the Computer Vision + +and Pattern Recognition Conference, pages 22055-22066, 2025.3 +[35] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8350, 2023. 1 +[36] Lesole Kalake, Wanggen Wan, and Li Hou. Analysis based on recent deep learning approaches applied in real-time multi-object tracking: a review. IEEE Access, 9:32650-32671, 2021. 3 +[37] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 3, 4, 6, 8 +[38] Mustafa Khan, Hamidreza Fazlali, Dhruv Sharma, Tongtong Cao, Dongfeng Bai, Yuan Ren, and Bingbing Liu. Autosplat: Constrained gaussian splatting for autonomous driving scene reconstruction. arXiv preprint arXiv:2407.02598, 2024. 1, 3, 6 +[39] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[40] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 2, 3 +[41] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision, 128(7): 1956-1981, 2020. 2, 3 +[42] Jinyu Li, Chenxu Luo, and Xiaodong Yang. Pillarnext: Rethinking network designs for 3d object detection in lidar point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17567-17576, 2023. 1, 3 +[43] Wei Li, CW Pan, Rong Zhang, JP Ren, YX Ma, Jin Fang, FL Yan, QC Geng, XY Huang, HJ Gong, et al. Aads: Augmented autonomous driving simulation using data-driven algorithms. Science robotics, 4(28):eaaw0863, 2019. 3 +[44] Yanyan Li, Chenyu Lyu, Yan Di, Guangyao Zhai, Gim Hee Lee, and Federico Tombari. Geogaussian: Geometry-aware gaussian splatting for scene rendering. In European Conference on Computer Vision, pages 441-457. Springer, 2024. 2 +[45] Zhong Li, Yu Ji, Wei Yang, Jinwei Ye, and Jingyi Yu. Robust 3d human motion reconstruction via dynamic template construction. In 2017 International Conference on 3D Vision (3DV), pages 496-505. IEEE, 2017. 3 +[46] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14864-14873, 2024. 1 +[47] Ming Liang, Bin Yang, Wenyuan Zeng, Yun Chen, Rui Hu, Sergio Casas, and Raquel Urtasun. Pnpnet: End-to-end perception and prediction with tracking in the loop. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11553-11562, 2020. 1 +[48] Yiqing Liang, Numair Khan, Zhengqin Li, Thu Nguyen-Phuoc, Douglas Lanman, James Tompkin, and Lei Xiao. Gaufre: Gaussian deformation fields for real-time dynamic novel view synthesis. arXiv preprint arXiv:2312.11458, 2023. 3 +[49] Yiyi Liao, Jun Xie, and Andreas Geiger. Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3292-3310, 2022. 2, 3 +[50] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 2 +[51] Xingyu Liu, Gu Wang, Yi Li, and Xiangyang Ji. CATRE: iterative point clouds alignment for category-level object pose refinement. In European Conference on Computer Vision, 2022. 3 +[52] Xingyu Liu, Chenyangguang Zhang, Gu Wang, Ruida Zhang, and Xiangyang Ji. Rasim: A range-aware high-fidelity rgb-d data simulation pipeline for real-world applications. In 2024 IEEE international conference on robotics and automation (ICRA), pages 17057-17064. IEEE, 2024. 3 +[53] Xingyu Liu, Yingyue Li, Chengxi Li, Gu Wang, Chenyangguang Zhang, Ziqin Huang, and Xiangyang Ji. Gfreedet: Exploiting gaussian splatting and foundation models for model-free unseen object detection in the bop challenge 2024. IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop, 2025. 3 +[54] Xingyu Liu, Gu Wang, Ruida Zhang, Chenyangguang Zhang, Federico Tombari, and Xiangyang Ji. UNOPose: Unseen object pose estimation with an unposed rgb-d reference image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 3 +[55] Xingyu Liu, Ruida Zhang, Chenyangguang Zhang, Gu Wang, Jiwen Tang, Zhigang Li, and Xiangyang Ji. Gdnpp: A geometry-guided and fully learning-based object pose estimator. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 3 +[56] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 1, 3 +[57] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 3 + +[58] Wenhan Luo, Junliang Xing, Anton Milan, Xiaoqin Zhang, Wei Liu, and Tae-Kyun Kim. Multiple object tracking: A literature review. Artificial intelligence, 293:103448, 2021. 3 +[59] Jiageng Mao, Minzhe Niu, Chenhan Jiang, Hanxue Liang, Jingheng Chen, Xiaodan Liang, Yamin Li, Chaoqiang Ye, Wei Zhang, Zhenguo Li, et al. One million scenes for autonomous driving: Once dataset. arXiv preprint arXiv:2106.11037, 2021.3 +[60] Jiageng Mao, Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. 3d object detection for autonomous driving: A comprehensive survey. International Journal of Computer Vision, 131(8):1909-1963, 2023. 3 +[61] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3 +[62] M Minderer, A Gritsenko, A Stone, M Neumann, D Weissenborn, A Dosovitskiy, A Mahendran, A Arnab, M Dehghani, Z Shen, et al. Simple open-vocabulary object detection with vision transformers. arxiv 2022. arXiv preprint arXiv:2205.06230, 2, 2022. 2 +[63] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 2 +[64] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 3 +[65] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2856-2865, 2021. 1, 3, 6, 8 +[66] Ziqi Pang, Zhichao Li, and Naiyan Wang. Simpletrack: Understanding and rethinking 3d multi-object tracking. In European Conference on Computer Vision, pages 680-696. Springer, 2022. 3, 6, 8 +[67] Chensheng Peng, Chengwei Zhang, Yixiao Wang, Chenfeng Xu, Yichen Xie, Wenzhao Zheng, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. Desire-gs: 4d street gaussians for static-dynamic decomposition and surface reconstruction for urban driving scenes. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 6782-6791, 2025. 3 +[68] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3 +[69] Jules Sanchez, Jean-Emmanuel Deschaud, and François Goulette. Domain generalization of 3d semantic segmentation in autonomous driving. In Proceedings of the + +IEEE/CVF International Conference on Computer Vision, pages 18077-18087, 2023. 3 +[70] Sara Fridovich-Keil and Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In CVPR, 2023. 5 +[71] Shital Shah, Debadeepta Dey, Chris Lovett, and Ashish Kapoor. Airsim: High-fidelity visual and physical simulation for autonomous vehicles. In Field and Service Robotics: Results of the 11th International Conference, pages 621-635. Springer, 2018. 3 +[72] Yeji Song, Chaerin Kong, Seoyoung Lee, Nojun Kwak, and Joonseok Lee. Towards efficient neural scene graphs by learning consistency fields. arXiv preprint arXiv:2210.04127, 2022. 3 +[73] Louis Soum-Fontez, Jean-Emmanuel Deschaud, and François Goulette. Mdt3d: Multi-dataset training for lidar 3d object detection generalization. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5765-5772. IEEE, 2023. 1, 3 +[74] Colton Stearns, Adam Harley, Mikaela Uy, Florian Dubost, Federico Tombari, Gordon Wetzstein, and Leonidas Guibas. Dynamic gaussian marbles for novel view synthesis of casual monocular videos. arXiv preprint arXiv:2406.18717, 2024. 3, 4, 5 +[75] Zhuo Su, Lan Xu, Zerong Zheng, Tao Yu, Yebin Liu, and Lu Fang. Robustfusion: Human volumetric capture with data-driven visual cues using a rgbd camera. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 246-264. Springer, 2020. 3 +[76] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 2, 3, 6 +[77] Shanlin Sun, Bingbing Zhuang, Ziyu Jiang, Buyu Liu, Xiaohui Xie, and Manmohan Chandraker. Lidarf: Delving into lidar for neural radiance field on street scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19563-19572, 2024. 3 +[78] Haithem Turki, Jason Y Zhang, Francesco Ferroni, and Deva Ramanan. Suds: Scalable urban dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12375-12385, 2023. 3 +[79] Yan Wang, Xiangyu Chen, Yurong You, Li Erran Li, Bharath Hariharan, Mark Campbell, Kilian Q Weinberger, and Wei-Lun Chao. Train in germany, test in the usa: Making 3d object detectors generalize. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11713-11723, 2020. 2 +[80] Yu Wang, Shaohua Wang, Yicheng Li, and Mingchun Liu. A comprehensive review of 3d object detection in autonomous driving: Technological advances and future directions. arXiv preprint arXiv:2408.16530, 2024. 3 + +[81] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20310-20320, 2024. 2, 3, 4, 5, 6, 8 +[82] Hai Wu, Wenkai Han, Chenglu Wen, Xin Li, and Cheng Wang. 3d multi-object tracking in point clouds based on prediction confidence-guided data association. IEEE Transactions on Intelligent Transportation Systems, 23(6):5668-5677, 2021. 2 +[83] Hai Wu, Jinhao Deng, Chenglu Wen, Xin Li, Cheng Wang, and Jonathan Li. Casa: A cascade attention network for 3-d object detection from lidar point clouds. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. 2 +[84] Junfeng Wu, Yi Jiang, Qihao Liu, Zehuan Yuan, Xiang Bai, and Song Bai. General object foundation model for images and videos at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3783-3795, 2024. 1, 2, 3, 4, 6, 8 +[85] Zirui Wu, Tianyu Liu, Liyi Luo, Zhide Zhong, Jianteng Chen, Hongmin Xiao, Chao Hou, Haozhe Lou, Yuantao Chen, Runyi Yang, et al. Mars: An instance-aware, modular and realistic simulator for autonomous driving. In CAAI International Conference on Artificial Intelligence, pages 3-15. Springer, 2023. 1, 3, 6, 8 +[86] Pengchuan Xiao, Zhenlei Shao, Steven Hao, Zishuo Zhang, Xiaolin Chai, Judy Jiao, Zesong Li, Jian Wu, Kai Sun, Kun Jiang, et al. Pandaset: Advanced sensor suite dataset for autonomous driving. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 3095-3101. IEEE, 2021. 2 +[87] Ziyang Xie, Junge Zhang, Wenye Li, Feihu Zhang, and Li Zhang. S-nerf: Neural radiance fields for street views. arXiv preprint arXiv:2303.00749, 2023. 3 +[88] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15325-15336, 2023. 1, 8 +[89] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians: Modeling dynamic urban scenes with gaussian splatting. In ECCV, 2024. 1, 2, 3, 6, 7, 8 +[90] Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Seung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, et al. Emernerf: Emergent spatial-temporal scene decomposition via self-supervision. arXiv preprint arXiv:2311.02077, 2023. 3, 6, 7 +[91] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1389-1399, 2023. 3 +[92] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time photorealistic dynamic scene representation and + +rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023.3 +[93] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20331-20341, 2024. 3 +[94] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 1, 3 +[95] Fisher Yu, Wenqi Xian, Yingying Chen, Fangchen Liu, Mike Liao, Vashisht Madhavan, Trevor Darrell, et al. Bdd100k: A diverse driving video database with scalable annotation tooling. arXiv preprint arXiv:1805.04687, 2(5): 6, 2018. 2, 3 +[96] Jiang-Tian Zhai, Ze Feng, Jinhao Du, Yongqiang Mao, Jiang-Jiang Liu, Zichang Tan, Yifu Zhang, Xiaqing Ye, and Jingdong Wang. Rethinking the open-loop evaluation of end-to-end autonomous driving in nuscenes. arXiv preprint arXiv:2305.10430, 2023. 1 +[97] Bo Zhang, Jiakang Yuan, Botian Shi, Tao Chen, Yikang Li, and Yu Qiao. Uni3d: A unified baseline for multi-dataset 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9253-9262, 2023. 1, 3 +[98] Peng Zhang, Xin Li, Liang He, and Xin Lin. 3d multiple object tracking on autonomous driving: A literature review. arXiv preprint arXiv:2309.15411, 2023. 3 +[99] Ruida Zhang, Yan Di, Zhiqiang Lou, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. Rbp-pose: Residual bounding box projection for category-level pose estimation. In European conference on computer vision, pages 655–672. Springer, 2022. 3 +[100] Ruida Zhang, Yan Di, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. Ssp-pose: Symmetry-aware shape prior deformation for direct category-level object pose estimation. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7452–7459. IEEE, 2022. +[101] Ruida Zhang, Ziqin Huang, Gu Wang, Chenyangguang Zhang, Yan Di, Xingxing Zuo, Jiwen Tang, and Xiangyang Ji. Lapose: Laplacian mixture shape modeling for rgb-based category-level object pose estimation. In European Conference on Computer Vision, pages 467-484. Springer, 2024. 3 +[102] Ruida Zhang, Chenyangguang Zhang, Yan Di, Fabian Manhardt, Xingyu Liu, Federico Tombari, and Xiangyang Ji. Kp-red: Exploiting semantic keypoints for joint 3d shape retrieval and deformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20540-20550, 2024. 3 +[103] Xingxuan Zhang, Zekai Xu, Renzhe Xu, Jiashuo Liu, Peng Cui, Weitao Wan, Chong Sun, and Chen Li. Towards domain generalization in object detection. arXiv preprint arXiv:2203.14387, 2022. 2 + +[104] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. International journal of computer vision, 129:3069–3087, 2021. 3 +[105] Hongyu Zhou, Jiahao Shao, Lu Xu, Dongfeng Bai, Weichao Qiu, Bingbing Liu, Yue Wang, Andreas Geiger, and Yiyi Liao. Hugs: Holistic urban 3d scene understanding via gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21336-21345, 2024. 1 +[106] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21634-21643, 2024. 1, 2, 3 \ No newline at end of file diff --git a/2025/Street Gaussians without 3D Object Tracker/images.zip b/2025/Street Gaussians without 3D Object Tracker/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d0db1a5700bd2246dc8a1c2f4073c7a945a2bc09 --- /dev/null +++ b/2025/Street Gaussians without 3D Object Tracker/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f3a670b94ef2bd59e1a73883d222d8c3bea105f1b9764b653d4c8d51add9ae34 +size 652110 diff --git a/2025/Street Gaussians without 3D Object Tracker/layout.json b/2025/Street Gaussians without 3D Object Tracker/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c05a70b0d96b3267fa5d897bbaef38bfb71f9195 --- /dev/null +++ b/2025/Street Gaussians without 3D Object Tracker/layout.json @@ -0,0 +1,11213 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 167, + 103, + 443, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 103, + 443, + 121 + ], + "spans": [ + { + "bbox": [ + 167, + 103, + 443, + 121 + ], + "type": "text", + "content": "Street Gaussians without 3D Object Tracker" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 142, + 501, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 142, + 501, + 186 + ], + "spans": [ + { + "bbox": [ + 108, + 142, + 501, + 186 + ], + "type": "text", + "content": "Ruida Zhang" + }, + { + "bbox": [ + 108, + 142, + 501, + 186 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 108, + 142, + 501, + 186 + ], + "type": "text", + "content": ", Chengxi Li" + }, + { + "bbox": [ + 108, + 142, + 501, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 142, + 501, + 186 + ], + "type": "text", + "content": ", Chenyangguang Zhang" + }, + { + "bbox": [ + 108, + 142, + 501, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 142, + 501, + 186 + ], + "type": "text", + "content": ", Xingyu Liu" + }, + { + "bbox": [ + 108, + 142, + 501, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 142, + 501, + 186 + ], + "type": "text", + "content": ", Haili Yuan" + }, + { + "bbox": [ + 108, + 142, + 501, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 142, + 501, + 186 + ], + "type": "text", + "content": ", Yanyan Li" + }, + { + "bbox": [ + 108, + 142, + 501, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 108, + 142, + 501, + 186 + ], + "type": "text", + "content": ", Xiangyang Ji" + }, + { + "bbox": [ + 108, + 142, + 501, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 142, + 501, + 186 + ], + "type": "text", + "content": ", Gim Hee Lee" + }, + { + "bbox": [ + 108, + 142, + 501, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 108, + 142, + 501, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 142, + 501, + 186 + ], + "type": "text", + "content": "Tsinghua University, " + }, + { + "bbox": [ + 108, + 142, + 501, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 108, + 142, + 501, + 186 + ], + "type": "text", + "content": "National University of Singapore" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 187, + 474, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 187, + 474, + 200 + ], + "spans": [ + { + "bbox": [ + 133, + 187, + 474, + 200 + ], + "type": "text", + "content": "{zhangrd23@mails, xyji@}.tsinghua.edu.cn, gimhee.lee@nus.edu.sg" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 59, + 209, + 301, + 370 + ], + "blocks": [ + { + "bbox": [ + 59, + 209, + 301, + 370 + ], + "lines": [ + { + "bbox": [ + 59, + 209, + 301, + 370 + ], + "spans": [ + { + "bbox": [ + 59, + 209, + 301, + 370 + ], + "type": "image", + "image_path": "5c0906e1d8a489a3a426abffcc9c2fb05a484304038c38d4a87724cee7c557e9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 377, + 555, + 424 + ], + "lines": [ + { + "bbox": [ + 54, + 377, + 555, + 424 + ], + "spans": [ + { + "bbox": [ + 54, + 377, + 555, + 424 + ], + "type": "text", + "content": "Figure 1. Comparison of 3D tracker-based Street Gaussians [89] (left) and our approach (right). Existing methods heavily rely on object poses, but 3D trackers struggle with limited generalization [21, 73, 97], leading to flaws in novel view synthesis. In contrast, 2D foundation models show better generalization [84, 88]. Our approach leverages a 2D foundation model [84] for object tracking and learns point motion within an implicit feature space to autonomously correct tracking errors, improving robustness across diverse scenes." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 306, + 209, + 552, + 369 + ], + "blocks": [ + { + "bbox": [ + 306, + 209, + 552, + 369 + ], + "lines": [ + { + "bbox": [ + 306, + 209, + 552, + 369 + ], + "spans": [ + { + "bbox": [ + 306, + 209, + 552, + 369 + ], + "type": "image", + "image_path": "7593f0549658f8187461f1f91ffff7fd5f5e85cbf312b35d5e34ce8d5fa0f1ba.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 434, + 200, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 434, + 200, + 445 + ], + "spans": [ + { + "bbox": [ + 151, + 434, + 200, + 445 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 459, + 297, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 459, + 297, + 712 + ], + "spans": [ + { + "bbox": [ + 54, + 459, + 297, + 712 + ], + "type": "text", + "content": "Realistic scene reconstruction in driving scenarios poses significant challenges due to fast-moving objects. Most existing methods rely on labor-intensive manual labeling of object poses to reconstruct dynamic objects in canonical space and move them based on these poses during rendering. While some approaches attempt to use 3D object trackers to replace manual annotations, the limited generalization of 3D trackers – caused by the scarcity of large-scale 3D datasets – results in inferior reconstructions in real-world settings. In contrast, 2D foundation models demonstrate strong generalization capabilities. To eliminate the reliance on 3D trackers and enhance robustness across diverse environments, we propose a stable object tracking module by leveraging associations from 2D deep trackers within a 3D object fusion strategy. We address inevitable tracking errors by further introducing a motion learning strategy in an implicit feature space that autonomously corrects trajectory errors and recovers missed detections. Experimental results on Waymo-NOTR and KITTI show that our method outperforms existing approaches. Our code will be made publicly available." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 434, + 394, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 434, + 394, + 445 + ], + "spans": [ + { + "bbox": [ + 315, + 434, + 394, + 445 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 457, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 555, + 567 + ], + "type": "text", + "content": "Modeling dynamic 3D street scenes underpins modern autonomous driving by enabling realistic, controllable simulations for tasks such as perception [11, 42, 56, 94, 105], prediction [25, 33, 47], and motion planning [13, 14, 16]. With the rise of end-to-end autonomous systems that require real-time sensor feedback [29, 31, 35], high-quality scene reconstructions have become essential for closed-loop evaluations [46, 96], particularly to simulate critical corner cases safely and cost-effectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 570, + 556, + 715 + ], + "type": "text", + "content": "Despite extensive efforts in achieving photo-realistic reconstruction of small-scale scenes, the unique challenges posed by the large-scale and highly dynamic nature of driving scenarios complicate effective 3D scene modeling. To address these challenges, most existing methods [22, 38, 65, 85, 106] rely on ground truth vehicle poses to differentiate between static background and moving vehicles. Typically, vehicles are reconstructed in canonical space and subsequently positioned based on known poses during rendering. However, collecting ground-truth poses is labor-intensive, limiting the applicability of these methods to scenes beyond the existing datasets." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25722" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 294, + 430 + ], + "type": "text", + "content": "To eliminate reliance of the systems on ground truth object poses, Street Gaussians [89] instead uses poses generated by 3D object trackers [82, 83]. However, by defining vehicle motion solely through these tracking poses, rendering quality becomes highly dependent on pose accuracy. While Street Gaussians optimizes object poses during training, it struggles with detection failures and large pose errors, as shown in Fig. 1. Unfortunately, 3D trackers often struggle to generalize effectively across different scenarios, primarily due to the scarcity of open-source largescale 3D datasets [79, 103]. Collecting ground-truth poses for 3D tracking is time-consuming and costly, resulting in relatively few open-source datasets for autonomous driving. For example, even widely-used datasets such as Waymo [76], Nuscenes [6], KITTI [23, 49], Pandaset [86] collectively contain fewer than 4,000 annotated scenes, limiting the diversity and scale needed for robust 3D tracking. In contrast, 2D data is far easier to collect and annotate, leading to an abundance of 2D datasets. The BDD100K dataset [95] alone includes annotations for 100,000 driving scenes, and additional large-scale 2D perception datasets, such as COCO [50], LVIS [28], and OpenImages [41], bring the total number of annotated scenes into the millions. This wealth of 2D data has enabled 2D foundation models to achieve strong generalization across a variety of tasks, including embedding extraction [64, 68], detection [8, 62], segmentation [12, 40], and tracking [84]. Nonetheless, how to use these accurate and robust 2D trackers to effectively model dynamic objects in 3D is still an open challenge for street scene reconstruction." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 435, + 294, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 435, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 435, + 294, + 624 + ], + "type": "text", + "content": "In this paper, a novel architecture is designed to achieve high-fidelity novel view synthesis performance in street scenarios for autonomous driving applications. Unlike previous approaches such as Driving Gaussian [106] and Street Gaussian [89] that rely on ground truth vehicle poses or poses predicted by 3D trackers, the first contribution of this work is a stable object tracking module. This module leverages associations from 2D deep trackers within a 3D object fusion framework to enhance robustness and accuracy. Specifically, we integrate 2D tracking outputs with LiDAR data to trace the trajectory of each vehicle in 3D. We then incrementally reconstruct the point cloud of each vehicle frame-by-frame in the canonical space, estimating its pose by aligning successive frames with this canonical model. This approach eliminates the reliance on 3D trackers and enhances robustness across diverse environments." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 630, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 630, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 630, + 294, + 712 + ], + "type": "text", + "content": "Although 2D tracking models demonstrate stronger generalization ability, tracking errors still can exist, especially under adverse conditions such as severe occlusions or distant objects. Since moving the object points solely based on the tracked object pose directly exposes any tracking errors, we aim to go beyond a straightforward reliance on tracked trajectories. Instead, we propose to learn point mo" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 316, + 72, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 251 + ], + "type": "text", + "content": "tion from the predicted trajectory, equipping the model to autonomously identify and correct tracking inaccuracies, recover missed detections, and infer motion in new time steps. An implicit representation is essential for this purpose: 1) It enables the model to refine trajectories without being constrained by bounding box tracks, facilitating smoother and more continuous motion that can adaptively respond to changes. 2) An implicit feature space offers versatility to move each point of an object in a different way, recognizing that vehicles are not strictly rigid objects — for example, doors can open or close. This approach makes it possible to capture subtle, dynamic changes within objects, ultimately producing more accurate reconstructions and enhancing the robustness of novel view synthesis in challenging scenarios." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 316, + 253, + 553, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 456 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 456 + ], + "type": "text", + "content": "To this end, we leverage HexPlane representation [7] following 4DGS [81]. HexPlane stores motion-related features by decomposing the 4D spatial-temporal space into six 2D learnable feature planes. A decoder then utilizes these features to predict deformation offsets, dynamically adjusting 3D Gaussians over time. However, 4DGS relies solely on image reconstruction loss for supervision, which is insufficient for handling rapid object motion. When a Gaussian's initial position is far from its current location, its projection on the image may fall outside the object's actual area, preventing gradient propagation and hindering optimization. This issue is evident in S3Gaussian [32], which struggles with moving cars due to the lack of explicit motion supervision. To address this, we introduce a training strategy that supervises learned point motion using the predicted trajectory, providing explicit 3D supervision to guide HexPlane in capturing motion dynamics more accurately." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 328, + 457, + 533, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 457, + 533, + 468 + ], + "spans": [ + { + "bbox": [ + 328, + 457, + 533, + 468 + ], + "type": "text", + "content": "Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 470, + 553, + 599 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 470, + 553, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 470, + 553, + 517 + ], + "spans": [ + { + "bbox": [ + 314, + 470, + 553, + 517 + ], + "type": "text", + "content": "- We eliminate the need of 3D tracker for street scene reconstruction by introducing a robust object tracking module which leverages 2D foundation model, achieving superior generalization ability across diverse scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 517, + 553, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 517, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 314, + 517, + 553, + 564 + ], + "type": "text", + "content": "- We introduce a motion learning framework to learn from the predicted trajectory in an implicit feature space, enabling it to automatically correct pose errors and infer motion for novel time steps." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 566, + 553, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 566, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 314, + 566, + 553, + 599 + ], + "type": "text", + "content": "- We outperform existing methods on Waymo-NOTR and KITTI datasets without relying on ground truth annotations." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 316, + 615, + 405, + 626 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 615, + 405, + 626 + ], + "spans": [ + { + "bbox": [ + 316, + 615, + 405, + 626 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 316, + 635, + 531, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 531, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 531, + 647 + ], + "type": "text", + "content": "2.1. 3D Gaussian Splatting for Dynamic Scene" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 654, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 654, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 654, + 553, + 712 + ], + "type": "text", + "content": "3D Gaussian Splatting (3DGS) [37, 44] has advanced scene reconstruction by enabling high-quality, real-time rendering with 3D Gaussians and efficient splat-based rasterization, reducing computation and parameters compared to NeRF-based methods [1-3, 61, 63] and other representa" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25723" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": "tions [15, 26, 27, 30, 45, 75, 102]. While originally designed for static scenes, 3DGS has been adapted for dynamic scenes [48, 53, 57, 74, 81, 92, 93]. Dynamic3DGS [57] directly stores information for each 3D Gaussian at every timestamp, and Yang et al. [92] approximate the spatiotemporal 4D volume by optimizing 4D Gaussian primitives. Deformable-3DGS [93] and GauFRe [48] employ deformation fields to model motions, while 4DGS [81] introduces HexPlane representation [7] to store spatial-temporal features efficiently. Gaussian Marbles [74] leverages isotropic \"marbles\" and a divide-and-conquer trajectory learning algorithm. Despite these advances, challenges persist in high-speed driving scenarios with rapid object motion. To tackle this challenge, we leverage a robust object tracking module to guide the motion learning process." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 261, + 209, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 261, + 209, + 272 + ], + "spans": [ + { + "bbox": [ + 55, + 261, + 209, + 272 + ], + "type": "text", + "content": "2.2. Street Scene Reconstruction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 279, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 296, + 651 + ], + "type": "text", + "content": "Existing autonomous driving simulation engines [20, 43, 52, 71] face high manual effort in creating virtual environments and a lack of realism in the generated data. The creation of high-fidelity simulations from driving logs is therefore essential for advancing closed-loop training and testing. Recent works [65, 72, 77, 78, 85, 87, 90, 91] have continuously made improvements to NeRF [61] to model street scenes dynamically. Despite the progress, NeRF-based methods remain computationally expensive and require densely overlapping views. Building on the effective 3DGS approach [37] for scene reconstruction, several 3DGS-based methods [10, 67, 89, 106] have emerged. PVG [10] model dynamic scenarios by using periodic vibration-based temporal dynamics. Driving Gaussian [106] introduces incremental static Gaussians and composite dynamic Gaussian graph. Street Gaussians [89] equips Gaussians with semantic logits, and optimize dynamic parts using tracked poses from 3D tracker. AutoSplat [38] enforces geometric constraints in road and sky regions for multi-view consistency. Among these approaches, most of them requires ground truth object pose [10, 38, 106] or the 3D tracker [89]. However, manual annotation is laborious and the 3D tracker is lack in generalization ability, which limits their applications in diverse scenarios. In contrast, S3Gaussian [32] models dynamics in a self-supervised manner using dynamic Gaussians from 4DGS [81], despite struggling with dynamic object modeling due to the lack of explicit motion supervision. We address this challenge by introducing a robust object tracking strategy based on a 2D foundation model [84] and apply motion supervision from the predicted object trajectory." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 659, + 170, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 170, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 170, + 671 + ], + "type": "text", + "content": "2.3. 2D and 3D trackers" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Multiple Object Tracking (MOT) aims at locating multiple objects in each frame, and establishing correspondences between them across frames within input videos" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "text", + "content": "[4, 5, 36, 58, 98]. The reliance of most MOT methods on object detection [4, 19, 58, 66, 98, 104] makes detection accuracy crucial for MOT performance. Recent advances in 3D object detection are promising [11, 18, 34, 42, 51, 54, 56, 60, 80, 94, 99-101], but these models often generalize poorly due to limited and biased datasets [21, 55, 73, 97]. Existing open-source autonomous driving datasets lack scale and show regional or environmental biases, such as vehicle density and weather conditions [6, 23, 59, 76]. The main reason is that building large-scale, well-annotated multimodal datasets is costly [6, 49, 59, 69]. In contrast, 2D image data is easier to capture and annotate, enabling extensive dataset collections [9, 17, 41, 95]. Visual foundation models trained on large-scale 2D data have shown strong generalization [12, 40, 64, 68, 84]. We thus opt to use 2D object tracker [84] to locate the dynamic objects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 285, + 394, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 285, + 394, + 298 + ], + "spans": [ + { + "bbox": [ + 314, + 285, + 394, + 298 + ], + "type": "text", + "content": "3. Our Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 305, + 511, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 305, + 511, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 511, + 319 + ], + "type": "text", + "content": "3.1. Prerequisites: 3D Gaussian Splatting" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 323, + 554, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 554, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 554, + 396 + ], + "type": "text", + "content": "3D Gaussian Splatting (3DGS) [37] represents a 3D scene explicitly as a collection of 3D Gaussian primitives. Each Gaussian is defined by its center position " + }, + { + "bbox": [ + 313, + 323, + 554, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{X} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 323, + 554, + 396 + ], + "type": "text", + "content": " and covariance " + }, + { + "bbox": [ + 313, + 323, + 554, + 396 + ], + "type": "inline_equation", + "content": "\\Sigma \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 313, + 323, + 554, + 396 + ], + "type": "text", + "content": ". To enforce semi-positive definiteness, the covariance matrix is further factorized into a scaling vector " + }, + { + "bbox": [ + 313, + 323, + 554, + 396 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 323, + 554, + 396 + ], + "type": "text", + "content": " and a rotation matrix " + }, + { + "bbox": [ + 313, + 323, + 554, + 396 + ], + "type": "inline_equation", + "content": "R \\in SO(3)" + }, + { + "bbox": [ + 313, + 323, + 554, + 396 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 400, + 401, + 553, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 401, + 553, + 416 + ], + "spans": [ + { + "bbox": [ + 400, + 401, + 553, + 416 + ], + "type": "interline_equation", + "content": "\\Sigma = R S S ^ {\\top} R ^ {\\top} \\tag {1}", + "image_path": "4e8bc7f0a77fa7c94fba0f8d8e2f94588eb23ff54a795432eca3e6b4efa46e0d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 423, + 554, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 554, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 554, + 471 + ], + "type": "text", + "content": "Additionally, each Gaussian is described with its opacity " + }, + { + "bbox": [ + 313, + 423, + 554, + 471 + ], + "type": "inline_equation", + "content": "o \\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 423, + 554, + 471 + ], + "type": "text", + "content": " and view-dependent color defined by spherical harmonic (SH) coefficient " + }, + { + "bbox": [ + 313, + 423, + 554, + 471 + ], + "type": "inline_equation", + "content": "\\mathcal{C} \\in \\mathbb{R}^k" + }, + { + "bbox": [ + 313, + 423, + 554, + 471 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 423, + 554, + 471 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 423, + 554, + 471 + ], + "type": "text", + "content": " represents numbers of SH functions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 472, + 554, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 554, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 554, + 519 + ], + "type": "text", + "content": "During rendering of novel views, differential splatting is applied to 3D Gaussians within the camera planes. The blending of " + }, + { + "bbox": [ + 313, + 472, + 554, + 519 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 472, + 554, + 519 + ], + "type": "text", + "content": " ordered points that overlap a pixel is given by the formula:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 377, + 527, + 553, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 527, + 553, + 561 + ], + "spans": [ + { + "bbox": [ + 377, + 527, + 553, + 561 + ], + "type": "interline_equation", + "content": "C = \\sum_ {i \\in N} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {2}", + "image_path": "acc711b5e5d464d6aa01e2fbf9ed4f0a0d7c6993586256e3fd0a23a0b0580853.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 567, + 554, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 554, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 554, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 567, + 554, + 604 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 313, + 567, + 554, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 567, + 554, + 604 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 313, + 567, + 554, + 604 + ], + "type": "text", + "content": " represents the opacity and color of the ith splatted Gaussian, which is computed from per-point opacity and SH coefficients (see [37] for details)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 610, + 382, + 622 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 610, + 382, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 610, + 382, + 622 + ], + "type": "text", + "content": "3.2. Overview" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 626, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 626, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 626, + 554, + 713 + ], + "type": "text", + "content": "Our method takes as input multi-view images " + }, + { + "bbox": [ + 313, + 626, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_t^{(j)}" + }, + { + "bbox": [ + 313, + 626, + 554, + 713 + ], + "type": "text", + "content": " from multiple cameras placed around the vehicle, each indexed by time step " + }, + { + "bbox": [ + 313, + 626, + 554, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 626, + 554, + 713 + ], + "type": "text", + "content": " and camera index " + }, + { + "bbox": [ + 313, + 626, + 554, + 713 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 626, + 554, + 713 + ], + "type": "text", + "content": ", along with intrinsic " + }, + { + "bbox": [ + 313, + 626, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{K}^{(j)}" + }, + { + "bbox": [ + 313, + 626, + 554, + 713 + ], + "type": "text", + "content": " and extrinsic " + }, + { + "bbox": [ + 313, + 626, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_t^{(j)}" + }, + { + "bbox": [ + 313, + 626, + 554, + 713 + ], + "type": "text", + "content": " matrices for each view. Additionally, a top-mounted LiDAR provides 3D point clouds " + }, + { + "bbox": [ + 313, + 626, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_t" + }, + { + "bbox": [ + 313, + 626, + 554, + 713 + ], + "type": "text", + "content": " for each frame. Using this multi-sensor data, our approach reconstructs the 3D scene and synthesizes novel views from any" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25724" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 68, + 552, + 227 + ], + "blocks": [ + { + "bbox": [ + 62, + 68, + 552, + 227 + ], + "lines": [ + { + "bbox": [ + 62, + 68, + 552, + 227 + ], + "spans": [ + { + "bbox": [ + 62, + 68, + 552, + 227 + ], + "type": "image", + "image_path": "bb66c5e49d0b54cdbedde8e690b18e2232ffc8dcb586ecf24b921d2354e23ad2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 233, + 555, + 313 + ], + "lines": [ + { + "bbox": [ + 54, + 233, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 54, + 233, + 555, + 313 + ], + "type": "text", + "content": "Figure 2. Overview of our method. To overcome the limited generalization of 3D object trackers, we introduce a robust object tracking module based on a 2D object tracker [84]. We integrate 2D tracking with LiDAR data to segment the object's point cloud and incrementally reconstruct the object model in canonical space frame-by-frame. The canonical model is used to estimate object pose " + }, + { + "bbox": [ + 54, + 233, + 555, + 313 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_t" + }, + { + "bbox": [ + 54, + 233, + 555, + 313 + ], + "type": "text", + "content": " and serves as the initialization for Gaussian optimization. We model dynamic objects with isotropic Gaussian marble [74] to simplify optimization. Our approach learns point motion within the HexPlane [7] feature space, enabling tracking error correction and missed detection recovery. Decoders " + }, + { + "bbox": [ + 54, + 233, + 555, + 313 + ], + "type": "inline_equation", + "content": "D_{\\mathcal{X}}, D_{\\mathcal{C}}" + }, + { + "bbox": [ + 54, + 233, + 555, + 313 + ], + "type": "text", + "content": " predict point motion " + }, + { + "bbox": [ + 54, + 233, + 555, + 313 + ], + "type": "inline_equation", + "content": "\\Delta \\mathcal{X}_t" + }, + { + "bbox": [ + 54, + 233, + 555, + 313 + ], + "type": "text", + "content": " and color change " + }, + { + "bbox": [ + 54, + 233, + 555, + 313 + ], + "type": "inline_equation", + "content": "\\Delta \\mathcal{C}_t" + }, + { + "bbox": [ + 54, + 233, + 555, + 313 + ], + "type": "text", + "content": " based on the HexPlane features " + }, + { + "bbox": [ + 54, + 233, + 555, + 313 + ], + "type": "inline_equation", + "content": "f(x,y,z,t)" + }, + { + "bbox": [ + 54, + 233, + 555, + 313 + ], + "type": "text", + "content": ", with the predicted trajectory providing supervision. Finally, novel view synthesis is performed by splatting the deformed Gaussian onto the image plane." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 331, + 294, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 331, + 294, + 355 + ], + "spans": [ + { + "bbox": [ + 54, + 331, + 294, + 355 + ], + "type": "text", + "content": "desired camera pose and time frame, without reliance on 3D object trackers or ground truth object trajectories." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 356, + 295, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 356, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 56, + 356, + 295, + 559 + ], + "type": "text", + "content": "We present an overview of our method in Fig. 2. Visual images and point clouds are obtained from the sensor setup. For images, static and dynamic components are analyzed first. We employ Mask2Former [12] to segment the scene into static and dynamic parts. Specifically, the dynamic part includes humans (pedestrians, cyclists, etc.) and vehicles (cars, trucks, etc.), and all the other objects are categorized as static. We model the static part as in 3DGS [37] and focus on modeling dynamic objects in the following sections. To enhance robustness, we propose an object tracking strategy by leveraging a robust 2D tracker [84] and the LiDAR point cloud, providing accurate initial point clouds and object trajectories for subsequent Gaussian optimization (cf. Sec 3.3). We leverage an implicit representation HexPlane [7] following 4DGS [81] to learn a continuous and smooth per-point motion based on the object trajectory (cf. Sec 3.4). The optimization objective is described in Sec 3.5." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 564, + 290, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 290, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 290, + 577 + ], + "type": "text", + "content": "3.3. Tracking Cars in 3D with Robust 2D Tracker" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 617 + ], + "type": "text", + "content": "To improve the robustness of novel view synthesis for diverse scenarios, the first challenge is how to track and build vehicles in 3D based on the 2D tracking results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": "We employ GLEE [84] to track all vehicles within the 2D image plane, obtaining 2D object trajectories in each camera view. Each trajectory includes the 2D segmentation masks of the object over a time period. To lift 2D trajectories into 3D, we first re-project the LiDAR point clouds onto the image plane and assign points to objects based on their presence within the corresponding 2D segmentation masks " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": ". This yields a segmented object point clouds for each" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 331, + 476, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 476, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 476, + 344 + ], + "type": "text", + "content": "camera view via the following function:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 329, + 353, + 553, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 353, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 329, + 353, + 553, + 369 + ], + "type": "interline_equation", + "content": "\\mathcal {O} _ {t} ^ {(j)} = \\left\\{\\mathbf {P} \\mid \\Pi^ {- 1} \\left(\\mathbf {P}, \\mathbf {K} ^ {(j)}, \\mathbf {E} _ {t} ^ {(j)}\\right) \\in \\mathcal {M}, \\mathbf {P} \\in \\mathbf {L} _ {t} \\right\\}, \\tag {3}", + "image_path": "fd16de7fa59b4821103975a91f1a3488c85617f65842e29ced7f54124e70c26c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 379, + 555, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 555, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 555, + 534 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 379, + 555, + 534 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 379, + 555, + 534 + ], + "type": "text", + "content": " is the camera index, " + }, + { + "bbox": [ + 313, + 379, + 555, + 534 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 379, + 555, + 534 + ], + "type": "text", + "content": " is the time index, " + }, + { + "bbox": [ + 313, + 379, + 555, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 313, + 379, + 555, + 534 + ], + "type": "text", + "content": " is a LiDAR point and " + }, + { + "bbox": [ + 313, + 379, + 555, + 534 + ], + "type": "inline_equation", + "content": "\\Pi(\\cdot)" + }, + { + "bbox": [ + 313, + 379, + 555, + 534 + ], + "type": "text", + "content": " is the re-projection function. Given the slight misalignment between the LiDAR point cloud and RGB image, points near the edges of segmentation masks often fall outside the object boundaries. To address this issue, we perform outlier removal to improve stability in subsequent reconstruction based on the point distance to the point cloud center. We then associate the same objects across different camera views. Two point clouds in different viewpoints are considered belonging to the same object if they share more than 50 points in one time step. This association process results in a set of associated partial object point cloud in different time steps." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 534, + 555, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 642 + ], + "type": "text", + "content": "After obtaining the 3D partial objects in different time steps, the next step is to associate these partial point cloud into a unified and complete model. We initialize the 3D reconstruction of each object from the first frame which it appears. For simplicity, we assume an object is visible between frames " + }, + { + "bbox": [ + 313, + 534, + 555, + 642 + ], + "type": "inline_equation", + "content": "0" + }, + { + "bbox": [ + 313, + 534, + 555, + 642 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 534, + 555, + 642 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 534, + 555, + 642 + ], + "type": "text", + "content": " and denote the object point cloud at time " + }, + { + "bbox": [ + 313, + 534, + 555, + 642 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 534, + 555, + 642 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 534, + 555, + 642 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_t" + }, + { + "bbox": [ + 313, + 534, + 555, + 642 + ], + "type": "text", + "content": ". Our goal is to obtain a temporally consistent reconstruction " + }, + { + "bbox": [ + 313, + 534, + 555, + 642 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 313, + 534, + 555, + 642 + ], + "type": "text", + "content": " and the object pose in each time frame. This reconstruction is done frame-by-frame incrementally." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 643, + 554, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 643, + 554, + 691 + ], + "spans": [ + { + "bbox": [ + 313, + 643, + 554, + 691 + ], + "type": "text", + "content": "Starting with the initial frame, we add " + }, + { + "bbox": [ + 313, + 643, + 554, + 691 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_0" + }, + { + "bbox": [ + 313, + 643, + 554, + 691 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 643, + 554, + 691 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 313, + 643, + 554, + 691 + ], + "type": "text", + "content": ". For each subsequent frame, we apply Iterative Closest Point (ICP) for the alignment of " + }, + { + "bbox": [ + 313, + 643, + 554, + 691 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_t" + }, + { + "bbox": [ + 313, + 643, + 554, + 691 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 643, + 554, + 691 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 313, + 643, + 554, + 691 + ], + "type": "text", + "content": " by extracting the relative pose " + }, + { + "bbox": [ + 313, + 643, + 554, + 691 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_t" + }, + { + "bbox": [ + 313, + 643, + 554, + 691 + ], + "type": "text", + "content": ". The overlap between " + }, + { + "bbox": [ + 313, + 643, + 554, + 691 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_t" + }, + { + "bbox": [ + 313, + 643, + 554, + 691 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 643, + 554, + 691 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 313, + 643, + 554, + 691 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 339, + 700, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 700, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 339, + 700, + 555, + 715 + ], + "type": "interline_equation", + "content": "\\tau_ {o v e r l a p} = \\mathcal {O} \\cap \\mathbf {T} _ {t} ^ {- 1} \\mathcal {O} _ {t}, \\quad \\mathbf {T} _ {t} = \\mathrm {I C P} (\\mathcal {O} _ {t}, \\mathcal {O}). \\quad (4)", + "image_path": "8814e7cacee0e07bc7e5801099e82845c7b2a299ec0db43030f278fc36fdc223.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "25725" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": "If ICP reveals an overlap " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\tau_{\\text{overlap}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": " of more than " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": ", we record the object pose " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_t" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": " and update " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": " by integrating " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_t^{-1}\\mathcal{O}_t" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": ". If " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\tau_{\\text{overlap}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": " is between " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": ", we only record the object pose " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "P_t" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": " without updating " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": ". If " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\tau_{\\text{overlap}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": " is lower than " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": ", we assume a 2D tracking failure for that frame and discard it to filter out tracking errors. This process is repeated for each frame using the last recorded pose as the initial guess for ICP to accelerate convergence. Through this approach, we reconstruct all objects and retrieve their poses in each frame, providing a robust initialization for subsequent 3D Gaussian optimization." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 211, + 186, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 211, + 186, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 186, + 224 + ], + "type": "text", + "content": "3.4. Learning Point Motion" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 228, + 296, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 296, + 372 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 296, + 372 + ], + "type": "text", + "content": "A common approach to model object motion in prior works is directly transforming the object points using the given object pose. However, despite the strong generalization capabilities of 2D trackers, detection failures are inevitable in challenging scenarios, such as when objects are heavily occluded or located at great distances. Relying solely on object poses derived from these erroneous or missing detections can easily lead to rendering failures. Additionally, treating cars as rigid objects does not adequately address corner cases, such as when a car door is open or closed. Furthermore, these methods lack the ability to infer an object motion at arbitrary time stamps." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 372, + 296, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 296, + 527 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 296, + 527 + ], + "type": "text", + "content": "To address these limitations, we seek for a motion modeling approach to enhance robustness and flexibility. Instead of explicitly using object pose as rigid transformation, we learn the per-point motion in a pre-defined feature space. In this feature space, the object motion can be optimized through both explicit guidance and photometric loss. Moreover, the object motion can be interpolated through time and space in this feature space to compensate for missing detections. To this end, we leverage HexPlane representation [7] as in 4DGS [81] to efficiently capture spatial and temporal information by decomposing the 4D feature voxels into six learnable feature planes. This representation satisfies all our requirements and is also memory efficient." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 528, + 296, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 296, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 296, + 575 + ], + "type": "text", + "content": "We use three planes: " + }, + { + "bbox": [ + 55, + 528, + 296, + 575 + ], + "type": "inline_equation", + "content": "P_{xy}" + }, + { + "bbox": [ + 55, + 528, + 296, + 575 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 528, + 296, + 575 + ], + "type": "inline_equation", + "content": "P_{yz}" + }, + { + "bbox": [ + 55, + 528, + 296, + 575 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 528, + 296, + 575 + ], + "type": "inline_equation", + "content": "P_{xz}" + }, + { + "bbox": [ + 55, + 528, + 296, + 575 + ], + "type": "text", + "content": " for the spatial dimensions, and another three planes " + }, + { + "bbox": [ + 55, + 528, + 296, + 575 + ], + "type": "inline_equation", + "content": "P_{xt}" + }, + { + "bbox": [ + 55, + 528, + 296, + 575 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 528, + 296, + 575 + ], + "type": "inline_equation", + "content": "P_{yt}" + }, + { + "bbox": [ + 55, + 528, + 296, + 575 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 528, + 296, + 575 + ], + "type": "inline_equation", + "content": "P_{zt}" + }, + { + "bbox": [ + 55, + 528, + 296, + 575 + ], + "type": "text", + "content": " for the spatial-temporal features. Additionally, the Hexplane includes multiple resolution levels which is formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 86, + 585, + 295, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 585, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 86, + 585, + 295, + 600 + ], + "type": "interline_equation", + "content": "\\left\\{P _ {i j} ^ {\\rho} \\in \\mathbb {R} ^ {d \\times \\rho r _ {i} \\times \\rho r _ {j}} | (i, j) \\in \\mathcal {P}, \\rho \\in \\{1, 2 \\} \\right\\}, \\tag {5}", + "image_path": "f1bcc4b385d67576e58053a93daabde9bcbec70602c4966a0405598d3b80c3d4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 609, + 295, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 295, + 644 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 609, + 295, + 644 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{(x,y),(x,z),(y,z),(x,t),(y,t),(z,t)\\}" + }, + { + "bbox": [ + 55, + 609, + 295, + 644 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 609, + 295, + 644 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 609, + 295, + 644 + ], + "type": "text", + "content": " is the feature dimension, " + }, + { + "bbox": [ + 55, + 609, + 295, + 644 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 55, + 609, + 295, + 644 + ], + "type": "text", + "content": " denotes the upsampling scale, and " + }, + { + "bbox": [ + 55, + 609, + 295, + 644 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 609, + 295, + 644 + ], + "type": "text", + "content": " is the base resolution." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 644, + 296, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 644, + 296, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 644, + 296, + 681 + ], + "type": "text", + "content": "Given the center position of a Gaussian " + }, + { + "bbox": [ + 55, + 644, + 296, + 681 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 55, + 644, + 296, + 681 + ], + "type": "text", + "content": " and the time step " + }, + { + "bbox": [ + 55, + 644, + 296, + 681 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 644, + 296, + 681 + ], + "type": "text", + "content": ", features in all six planes are queried and combined via a small MLP " + }, + { + "bbox": [ + 55, + 644, + 296, + 681 + ], + "type": "inline_equation", + "content": "\\phi_d" + }, + { + "bbox": [ + 55, + 644, + 296, + 681 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 63, + 689, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 63, + 689, + 295, + 715 + ], + "type": "interline_equation", + "content": "f (x, y, z, t) = \\phi_ {d} \\left(\\bigcup_ {\\rho} \\prod_ {(i, j) \\in \\mathcal {P}} \\pi \\left(P _ {i j} ^ {\\rho}, \\psi_ {i j} ^ {\\rho} (x, y, z, t)\\right)\\right), \\tag {6}", + "image_path": "a565f21d22f0c7c47a105b12439997f69455ccd39997f7d7188bc69a327cf242.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "\\psi_{ij}^{\\rho}(x,y,z,t)" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " projects the 4D coordinate onto the respective plane, and " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " performs bilinear interpolation on the voxel features at each point. The features of all planes are combined using the Hadamard product." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 121, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 553, + 180 + ], + "type": "text", + "content": "Finally, two MLP decoders " + }, + { + "bbox": [ + 313, + 121, + 553, + 180 + ], + "type": "inline_equation", + "content": "D_{\\mathcal{X}}, D_{c}" + }, + { + "bbox": [ + 313, + 121, + 553, + 180 + ], + "type": "text", + "content": " are utilized to predict the point motion " + }, + { + "bbox": [ + 313, + 121, + 553, + 180 + ], + "type": "inline_equation", + "content": "\\Delta \\mathcal{X}_t" + }, + { + "bbox": [ + 313, + 121, + 553, + 180 + ], + "type": "text", + "content": " and color change " + }, + { + "bbox": [ + 313, + 121, + 553, + 180 + ], + "type": "inline_equation", + "content": "\\Delta \\mathcal{C}_t" + }, + { + "bbox": [ + 313, + 121, + 553, + 180 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 121, + 553, + 180 + ], + "type": "inline_equation", + "content": "\\Delta \\mathcal{X}_t = D_{\\mathcal{X}}(f(x,y,z,t))" + }, + { + "bbox": [ + 313, + 121, + 553, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 121, + 553, + 180 + ], + "type": "inline_equation", + "content": "\\Delta \\mathcal{C}_t = D_{\\mathcal{C}}(f(x,y,z,t))" + }, + { + "bbox": [ + 313, + 121, + 553, + 180 + ], + "type": "text", + "content": ", respectively. The deformed 4D Gaussians are formulated as: " + }, + { + "bbox": [ + 313, + 121, + 553, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_t = \\{\\mathcal{X} + \\Delta \\mathcal{X}_t, \\mathcal{C} + \\Delta \\mathcal{C}_t, S, R, o\\}" + }, + { + "bbox": [ + 313, + 121, + 553, + 180 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 181, + 553, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 181, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 181, + 553, + 205 + ], + "type": "text", + "content": "To allow the HexPlane feature to learn point motion from the predicted pose, we define the motion loss as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 350, + 216, + 553, + 230 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 216, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 350, + 216, + 553, + 230 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m o t i o n}} = \\operatorname {a v g} _ {\\mathcal {X} \\in \\mathcal {O}} | \\Delta \\mathcal {X} _ {t} - (\\mathbf {T} _ {t} \\mathcal {X} - \\mathcal {X}) |, \\tag {7}", + "image_path": "4e62935e6884ac4329afa18c4e0ab45444716c80b5caf132e126e1af67b3bee3.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 239, + 553, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 553, + 323 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 239, + 553, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 313, + 239, + 553, + 323 + ], + "type": "text", + "content": " is the center position of a Gaussian in object " + }, + { + "bbox": [ + 313, + 239, + 553, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 313, + 239, + 553, + 323 + ], + "type": "text", + "content": " and avg is the average operator. This loss encourages the predicted point motion " + }, + { + "bbox": [ + 313, + 239, + 553, + 323 + ], + "type": "inline_equation", + "content": "\\Delta \\mathcal{X}_t" + }, + { + "bbox": [ + 313, + 239, + 553, + 323 + ], + "type": "text", + "content": " to align with the predicted pose. We apply this loss only for the first " + }, + { + "bbox": [ + 313, + 239, + 553, + 323 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 239, + 553, + 323 + ], + "type": "text", + "content": " iterations to provide a strong initial motion prior. The loss is subsequently removed, allowing the network to adjust and potentially compensate for pose errors and detection failures." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 323, + 554, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 554, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 554, + 443 + ], + "type": "text", + "content": "Following [74], we adopt isotropic Gaussian marbles for dynamic points to reduce degrees of freedom and simplify optimization. In this setup, the rotation of each Gaussian is represented by the identity matrix with identical scales across all three dimensions. The spherical harmonics coefficients are also limited to three dimensions, providing view-consistent color. This approach ensures that all deformations are captured purely by point motion and color changes, enforcing strong constraints and enhancing robustness in novel view synthesis." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 447, + 554, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 554, + 555 + ], + "type": "text", + "content": "Ours vs. S3Gaussian. Although S3Gaussian [32] also uses Hexplane representation [7] and the deformation network from 4DGS [81], our design serves a different purpose: S3Gaussian learns point motion without extra supervision, while we use 4DGS to refine motion and address detection failures. Consequently, S3Gaussian becomes susceptible to unsatisfactory results with rapid car motion due to the lack of explicit guidance. In contrast, our approach enables high-quality reconstruction of dynamic objects." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 563, + 447, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 447, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 447, + 577 + ], + "type": "text", + "content": "3.5. Optimization Objective" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 582, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 714 + ], + "type": "text", + "content": "Besides the motion loss introduced above, the loss function comprises five components to collectively optimize the scene representation and the point motion. The primary loss " + }, + { + "bbox": [ + 313, + 582, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rgb}}" + }, + { + "bbox": [ + 313, + 582, + 554, + 714 + ], + "type": "text", + "content": " is an L1 loss measuring the photometric difference between rendered and ground truth images while " + }, + { + "bbox": [ + 313, + 582, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ssim}}" + }, + { + "bbox": [ + 313, + 582, + 554, + 714 + ], + "type": "text", + "content": " evaluates their structural similarity. The L1 loss between the rendered depth map and the depth estimated from the LiDAR point cloud " + }, + { + "bbox": [ + 313, + 582, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{depth}}" + }, + { + "bbox": [ + 313, + 582, + 554, + 714 + ], + "type": "text", + "content": " is used to supervise the Gaussian positions. Following K-Planes [70], a grid-based total variation loss " + }, + { + "bbox": [ + 313, + 582, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{tv}}" + }, + { + "bbox": [ + 313, + 582, + 554, + 714 + ], + "type": "text", + "content": " is introduced. Recognizing that the color of most dynamic points in the scene are unchanged, a L1 regulariza" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25726" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 555, + 194 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 194 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 194 + ], + "type": "table", + "html": "
TaskScene ReconstructionNovel View Synthesis
MethodExtra InputPSNR↑SSIM↑LPIPS↓DPSNR↑DSSIM↑PSNR↑SSIM↑LPIPS↓DPSNR↑DSSIM↑
3DGS [37]N/A25.770.8330.16020.260.60424.460.8020.17018.120.521
EmerNeRF [90]N/A28.160.8060.22824.320.68225.140.7470.31323.490.660
S3Gaussian [32]N/A31.350.9110.10626.020.78327.440.8570.13722.920.680
MARS [85]GT pose28.240.8660.25223.370.70126.610.7960.30522.210.697
StreetGS [89]3D tracker29.170.8730.13827.780.81826.980.8380.14924.620.742
Ours2D tracker32.560.9360.05928.510.86828.850.8670.08825.580.779
", + "image_path": "82a6f2241af8d1eb636ed0cee16d502821d994a96a88d232adf10d04ccad6802.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 204, + 555, + 228 + ], + "lines": [ + { + "bbox": [ + 55, + 204, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 555, + 228 + ], + "type": "text", + "content": "Table 1. Comparison with state-of-the-art methods on Waymo-NOTR dataset. StreetGS represents Street Gaussian [89]. The best and the second best results are denoted by pink and blue." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 249, + 294, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 249, + 294, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 294, + 273 + ], + "type": "text", + "content": "tion loss " + }, + { + "bbox": [ + 55, + 249, + 294, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{color - reg}}" + }, + { + "bbox": [ + 55, + 249, + 294, + 273 + ], + "type": "text", + "content": " is applied to the deformation network by minimizing " + }, + { + "bbox": [ + 55, + 249, + 294, + 273 + ], + "type": "inline_equation", + "content": "\\Delta C" + }, + { + "bbox": [ + 55, + 249, + 294, + 273 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 274, + 234, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 274, + 234, + 285 + ], + "spans": [ + { + "bbox": [ + 67, + 274, + 234, + 285 + ], + "type": "text", + "content": "The total loss function is thus defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 291, + 294, + 320 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 291, + 294, + 320 + ], + "spans": [ + { + "bbox": [ + 70, + 291, + 294, + 320 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\lambda_ {\\mathrm {r g b}} \\mathcal {L} _ {\\mathrm {r g b}} + \\lambda_ {\\text {d e p t h}} \\mathcal {L} _ {\\text {d e p t h}} + \\lambda_ {\\text {s s i m}} \\mathcal {L} _ {\\text {s s i m}} + \\lambda_ {\\mathrm {t v}} \\mathcal {L} _ {\\mathrm {t v}} \\tag {8} \\\\ + \\lambda_ {\\text {c o l o r - r e g}} \\mathcal {L} _ {\\text {c o l o r - r e g}} + \\lambda_ {\\text {m o t i o n}} \\mathcal {L} _ {\\text {m o t i o n}}, \\\\ \\end{array}", + "image_path": "d8cbea14ba4c960f56e4ed6c66f4bd70796f6326a5e0e8dd7fb2803410384c4d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 328, + 295, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 295, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 295, + 363 + ], + "type": "text", + "content": "where the weights are assigned as follows: " + }, + { + "bbox": [ + 55, + 328, + 295, + 363 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{rgb}} = 1.0" + }, + { + "bbox": [ + 55, + 328, + 295, + 363 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{depth}} = 1.0" + }, + { + "bbox": [ + 55, + 328, + 295, + 363 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{ssim}} = 0.1" + }, + { + "bbox": [ + 55, + 328, + 295, + 363 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{tv}} = 0.1" + }, + { + "bbox": [ + 55, + 328, + 295, + 363 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{color-reg}} = 0.01" + }, + { + "bbox": [ + 55, + 328, + 295, + 363 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 328, + 295, + 363 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{motion}} = 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 365, + 296, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 296, + 448 + ], + "type": "text", + "content": "The optimization process is divided into two stages. In the first stage, we render using only static Gaussians and supervise only the static regions of the images over 20,000 iterations. This phase focuses on reconstructing the static background and stabilizes the training process. In the second stage, we train all Gaussians on the whole images for 50,000 iterations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 458, + 137, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 458, + 137, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 458, + 137, + 472 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 478, + 171, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 171, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 171, + 491 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 495, + 295, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 295, + 674 + ], + "type": "text", + "content": "Datasets. We leverage Waymo-NOTR dataset [76, 90] and KITTI benchmark [23, 24]. The NeRF On-The-Road (NOTR) dataset, introduced by EmerNeRF [90], is a subset of the Waymo Open dataset [76] and includes diverse challenging driving scenarios, such as high-speed, exposure mismatch, and various weather conditions. We use the dynamic32 subset, consisting of 32 dynamic scenes, for evaluating dynamic reconstruction. Following EmerNeRF's setup [90], we use three frontal camera images resized to " + }, + { + "bbox": [ + 55, + 495, + 295, + 674 + ], + "type": "inline_equation", + "content": "960 \\times 640" + }, + { + "bbox": [ + 55, + 495, + 295, + 674 + ], + "type": "text", + "content": ". For scene reconstruction, all image frames are used for training and evaluation; for novel view synthesis, every 10th time step is excluded for evaluation [90]. For the KITTI dataset, we follow the setup of MARS [85] to use " + }, + { + "bbox": [ + 55, + 495, + 295, + 674 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 55, + 495, + 295, + 674 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 495, + 295, + 674 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 495, + 295, + 674 + ], + "type": "text", + "content": " of the images for training with every 4th or every 2nd frame held out for testing, respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Implementation Details. We use the Adam optimizer [39] with the same learning rate schedule as in 3DGS [37]. For long-sequence scene reconstruction, we follow S3Gaussian" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 249, + 555, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 249, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 555, + 357 + ], + "type": "text", + "content": "[32] and segment the scene into multiple clips. The multi-resolution HexPlane encoder has a base resolution of 64, upsampled by factors of 2 and 4 as in [81], and other hyperparameters match those in 3DGS [37]. As LiDAR points lack data for the sky region, we add a plane of points above the maximum height of the scene to represent it following [38]. All experiments are run on a single NVIDIA RTX 3090 GPU, with training taking about 2 hours per video clip and inference speed at 100 FPS at " + }, + { + "bbox": [ + 313, + 249, + 555, + 357 + ], + "type": "inline_equation", + "content": "960 \\times 640" + }, + { + "bbox": [ + 313, + 249, + 555, + 357 + ], + "type": "text", + "content": " resolution." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 367, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 556, + 713 + ], + "type": "text", + "content": "Baseline Methods. We compare our approach against state-of-the-art methods, including NeRF-based methods (MARS [85], NSG [65], EmerNeRF [90]) and Gaussian-based methods (3DGS [37], StreetGaussian [89], S3Gaussian [32]). To ensure fair comparisons, we apply LiDAR point cloud initialization and add depth regularization to 3DGS. For the Waymo-NOTR dataset, we borrow results of MARS [85] and EmerNeRF [90] from S3Gaussian [32]. We evaluate the 3D-tracker-based method, Street Gaussians [89], using detection results from VoxelNext [11] and associate the detections with SimpleTrack [66]. Our goal is to assess the generalization ability of 3D-tracker-based methods in scenarios without ground truth tracking labels. To emulate a real-world scenario where a 3D tracker trained on a public benchmark dataset is applied to a novel environment, we use pretrained weights from nuScenes [6]—one of the largest and most diverse 3D datasets—and evaluate the model on the Waymo-NOTR dataset. Since NOTR is part of the training set for the Waymo perception task, using Waymo pretrained weights would introduce domain overlap and bias the evaluation, making the nuScenes weights a fairer choice. For a fair comparison, we use pretrained 2D tracker weights from [84] for our method, which have not been trained on the Waymo or KITTI datasets. Please refer to the supplementary material for discussion of the choice of 3D trackers. For the KITTI dataset, we borrow the results of all other methods from Street Gaussians. Please note that Street Gaussians leverages GT tracking annotations on KITTI dataset." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25727" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 85, + 185, + 162 + ], + "blocks": [ + { + "bbox": [ + 119, + 70, + 136, + 80 + ], + "lines": [ + { + "bbox": [ + 119, + 70, + 136, + 80 + ], + "spans": [ + { + "bbox": [ + 119, + 70, + 136, + 80 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 85, + 185, + 162 + ], + "lines": [ + { + "bbox": [ + 70, + 85, + 185, + 162 + ], + "spans": [ + { + "bbox": [ + 70, + 85, + 185, + 162 + ], + "type": "image", + "image_path": "adb78bd5aae696541ac6b36c87113c0ef7399b603f4f1c40e6ccadaf9b589bde.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 187, + 85, + 302, + 162 + ], + "blocks": [ + { + "bbox": [ + 217, + 70, + 271, + 80 + ], + "lines": [ + { + "bbox": [ + 217, + 70, + 271, + 80 + ], + "spans": [ + { + "bbox": [ + 217, + 70, + 271, + 80 + ], + "type": "text", + "content": "S3Gaussian" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 187, + 85, + 302, + 162 + ], + "lines": [ + { + "bbox": [ + 187, + 85, + 302, + 162 + ], + "spans": [ + { + "bbox": [ + 187, + 85, + 302, + 162 + ], + "type": "image", + "image_path": "d2c05fd3d944350e55146f6f7b856b26cdfbb091f6e06300cbe67c6318694edc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 306, + 85, + 421, + 162 + ], + "blocks": [ + { + "bbox": [ + 328, + 70, + 399, + 80 + ], + "lines": [ + { + "bbox": [ + 328, + 70, + 399, + 80 + ], + "spans": [ + { + "bbox": [ + 328, + 70, + 399, + 80 + ], + "type": "text", + "content": "Street Gaussian" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 85, + 421, + 162 + ], + "lines": [ + { + "bbox": [ + 306, + 85, + 421, + 162 + ], + "spans": [ + { + "bbox": [ + 306, + 85, + 421, + 162 + ], + "type": "image", + "image_path": "d59d369509dd33f47f225170a24d9ed605933171dcfcb29c6e367571db7f671d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 425, + 85, + 541, + 162 + ], + "blocks": [ + { + "bbox": [ + 470, + 70, + 495, + 80 + ], + "lines": [ + { + "bbox": [ + 470, + 70, + 495, + 80 + ], + "spans": [ + { + "bbox": [ + 470, + 70, + 495, + 80 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 425, + 85, + 541, + 162 + ], + "lines": [ + { + "bbox": [ + 425, + 85, + 541, + 162 + ], + "spans": [ + { + "bbox": [ + 425, + 85, + 541, + 162 + ], + "type": "image", + "image_path": "81b8cdadfe611141909cc8417c57d85ba2b0f10ab5df4d0b89712c8784c8cdbf.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 70, + 163, + 185, + 241 + ], + "blocks": [ + { + "bbox": [ + 70, + 163, + 185, + 241 + ], + "lines": [ + { + "bbox": [ + 70, + 163, + 185, + 241 + ], + "spans": [ + { + "bbox": [ + 70, + 163, + 185, + 241 + ], + "type": "image", + "image_path": "5df529ecf0004a6a96c2ad0dfa62667653d39dc362b50ac96b5a6b16a04e5888.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 187, + 163, + 302, + 241 + ], + "blocks": [ + { + "bbox": [ + 187, + 163, + 302, + 241 + ], + "lines": [ + { + "bbox": [ + 187, + 163, + 302, + 241 + ], + "spans": [ + { + "bbox": [ + 187, + 163, + 302, + 241 + ], + "type": "image", + "image_path": "4924e0c1f01051557eaa8e9143c81d0219a7f3c10f565c3bf2e65a5bcea0b94e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 163, + 421, + 241 + ], + "blocks": [ + { + "bbox": [ + 307, + 163, + 421, + 241 + ], + "lines": [ + { + "bbox": [ + 307, + 163, + 421, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 421, + 241 + ], + "type": "image", + "image_path": "8b0e8a9311ede3f05844cd7ca7aa62a2b0e7b1fdcdd35ca07286b0b96f9dabac.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 425, + 163, + 541, + 241 + ], + "blocks": [ + { + "bbox": [ + 425, + 163, + 541, + 241 + ], + "lines": [ + { + "bbox": [ + 425, + 163, + 541, + 241 + ], + "spans": [ + { + "bbox": [ + 425, + 163, + 541, + 241 + ], + "type": "image", + "image_path": "6939d243ac553a629ef06e0efd4763649494891d49911fe4b09ce46bafca7ac1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 70, + 243, + 185, + 320 + ], + "blocks": [ + { + "bbox": [ + 70, + 243, + 185, + 320 + ], + "lines": [ + { + "bbox": [ + 70, + 243, + 185, + 320 + ], + "spans": [ + { + "bbox": [ + 70, + 243, + 185, + 320 + ], + "type": "image", + "image_path": "9605e61606317d5c9f7d250fbda8f37b42cd7fed48d3857f46ce00870f1cd617.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 187, + 243, + 302, + 320 + ], + "blocks": [ + { + "bbox": [ + 187, + 243, + 302, + 320 + ], + "lines": [ + { + "bbox": [ + 187, + 243, + 302, + 320 + ], + "spans": [ + { + "bbox": [ + 187, + 243, + 302, + 320 + ], + "type": "image", + "image_path": "cc6dfeb6b8c089369b408ccbd18a17f91dd2c1b2debcf9db41116ea2d8885c93.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 307, + 243, + 421, + 320 + ], + "blocks": [ + { + "bbox": [ + 307, + 243, + 421, + 320 + ], + "lines": [ + { + "bbox": [ + 307, + 243, + 421, + 320 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 421, + 320 + ], + "type": "image", + "image_path": "279869d5a6455c6ef8c58ab0617acf743196665c4b9bc249e21ab5dfbeaf002b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 425, + 243, + 541, + 320 + ], + "blocks": [ + { + "bbox": [ + 425, + 243, + 541, + 320 + ], + "lines": [ + { + "bbox": [ + 425, + 243, + 541, + 320 + ], + "spans": [ + { + "bbox": [ + 425, + 243, + 541, + 320 + ], + "type": "image", + "image_path": "9888d28bf48e41917c575bec7dba41b21413cc1366248ce6b1becb2f13ea4ff2.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 70, + 322, + 185, + 400 + ], + "blocks": [ + { + "bbox": [ + 70, + 322, + 185, + 400 + ], + "lines": [ + { + "bbox": [ + 70, + 322, + 185, + 400 + ], + "spans": [ + { + "bbox": [ + 70, + 322, + 185, + 400 + ], + "type": "image", + "image_path": "b066907af73fbc81b4b4cc5c84def1915148ce1ad7efc96ad6d647164c72ac6c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 187, + 322, + 302, + 400 + ], + "blocks": [ + { + "bbox": [ + 187, + 322, + 302, + 400 + ], + "lines": [ + { + "bbox": [ + 187, + 322, + 302, + 400 + ], + "spans": [ + { + "bbox": [ + 187, + 322, + 302, + 400 + ], + "type": "image", + "image_path": "d7551168474300e162ab28a1904cefcbe6d0d01ccf6f5e50f087df9ae7f777fa.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 306, + 322, + 421, + 400 + ], + "blocks": [ + { + "bbox": [ + 306, + 322, + 421, + 400 + ], + "lines": [ + { + "bbox": [ + 306, + 322, + 421, + 400 + ], + "spans": [ + { + "bbox": [ + 306, + 322, + 421, + 400 + ], + "type": "image", + "image_path": "c47fc0aea547c28ac8c75c279e4c25ac091b2c7511cde40fe6144e8d3948ce3c.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 425, + 322, + 541, + 400 + ], + "blocks": [ + { + "bbox": [ + 425, + 322, + 541, + 400 + ], + "lines": [ + { + "bbox": [ + 425, + 322, + 541, + 400 + ], + "spans": [ + { + "bbox": [ + 425, + 322, + 541, + 400 + ], + "type": "image", + "image_path": "6e9ac1194a1144189f517d5e3b60d1bbaaa1995f0e562474fa51be0620ef1556.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 70, + 401, + 185, + 479 + ], + "blocks": [ + { + "bbox": [ + 70, + 401, + 185, + 479 + ], + "lines": [ + { + "bbox": [ + 70, + 401, + 185, + 479 + ], + "spans": [ + { + "bbox": [ + 70, + 401, + 185, + 479 + ], + "type": "image", + "image_path": "da6ba43bf89505272eb6fe1987d545187cce8a47bf812aa092b1e31ad4152e97.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 122, + 490, + 487, + 501 + ], + "lines": [ + { + "bbox": [ + 122, + 490, + 487, + 501 + ], + "spans": [ + { + "bbox": [ + 122, + 490, + 487, + 501 + ], + "type": "text", + "content": "Figure 3. Qualitative comparison of novel view synthesis on NOTR dataset. Best viewed with zoom." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 187, + 401, + 302, + 479 + ], + "blocks": [ + { + "bbox": [ + 187, + 401, + 302, + 479 + ], + "lines": [ + { + "bbox": [ + 187, + 401, + 302, + 479 + ], + "spans": [ + { + "bbox": [ + 187, + 401, + 302, + 479 + ], + "type": "image", + "image_path": "b82259fa799bcb919099bbae0b7f78f794afc6264b72b788c38f760334efa06a.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 306, + 401, + 421, + 479 + ], + "blocks": [ + { + "bbox": [ + 306, + 401, + 421, + 479 + ], + "lines": [ + { + "bbox": [ + 306, + 401, + 421, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 401, + 421, + 479 + ], + "type": "image", + "image_path": "030639fa14ec1fd5122fbc19acabdc8bb5d4284dc864e704b73b195493025a7b.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 425, + 401, + 541, + 479 + ], + "blocks": [ + { + "bbox": [ + 425, + 401, + 541, + 479 + ], + "lines": [ + { + "bbox": [ + 425, + 401, + 541, + 479 + ], + "spans": [ + { + "bbox": [ + 425, + 401, + 541, + 479 + ], + "type": "image", + "image_path": "90d5cd84366e64bfa75e381301ea0d16a1cb7b1b6cf05cfc903ef4f635ef1eb8.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 522, + 295, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 606 + ], + "type": "text", + "content": "Metrics. We leverage peak signal-to-noise ratio (PSNR), structural similarity index (SSIM) and learned perceptual image patch similarity (LPIPS) to evaluate the rendering quality. Additionally, following EmerNeRF [90], we apply DPSNR and DSSIM metrics to dynamic objects by projecting their ground truth 3D bounding boxes onto the 2D image plane and calculating pixel loss within these regions." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 612, + 257, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 257, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 257, + 624 + ], + "type": "text", + "content": "4.2. Comparisons with the State-of-the-art" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "On the NOTR dataset, our method outperforms all competitors across every metric, as shown in Table 1. Our approach sets a new state-of-the-art in both the scene reconstruction and novel view synthesis tasks. Specifically, our method outperforms S3Gaussian [32] by " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "2.66\\mathrm{dB}" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " in DPSNR and 0.099 in DSSIM for novel view synthesis. It also surpasses Street Gaussians [89] by " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "1.87\\mathrm{dB}" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " in PSNR, mainly due" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 522, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 555, + 582 + ], + "type": "text", + "content": "to the limited generalization of 3D trackers used in Street Gaussians. Despite optimizing object poses, Street Gaussians struggles with large pose errors and detection failures, leading to inferior performance. Additionally, we outperform NeRF-based methods like EmerNeRF and MARS." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "We present qualitative comparison with S3Gaussian [32] and Street Gaussians [89] in Fig. 3. Results from S3Gaussian show that using 4DGS without explicit motion guidance results in weaker performance when handling moving vehicles (e.g. row 2, 3, 5). Street Gaussians suffers from tracking errors of 3D trackers, leading to erroneous reconstructions (e.g. row 1, 3) or entirely missed objects (e.g. rows 2, 4, and 5). In contrast, our approach performs robustly across diverse scenarios owing to our tracking strategy leveraging 2D foundation model and robust motion" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25728" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 93, + 70, + 518, + 180 + ], + "blocks": [ + { + "bbox": [ + 93, + 70, + 518, + 180 + ], + "lines": [ + { + "bbox": [ + 93, + 70, + 518, + 180 + ], + "spans": [ + { + "bbox": [ + 93, + 70, + 518, + 180 + ], + "type": "table", + "html": "
KITTI-75%KITTI-50%
MethodExtra InputPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
3DGS [37]N/A19.190.7370.17219.230.7390.174
NSG [65]GT pose21.530.6730.25421.260.6590.266
MARS [85]GT pose24.230.8450.16024.000.8010.164
Street Gussians [89]GT pose25.790.8440.08125.520.8410.084
Ours2D tracker [84]25.490.8890.06325.110.8770.067
", + "image_path": "2983810f71fb79df77a48f2f5e2c62dc996667575291f1351fc81ef745cd26d9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 229, + 293, + 370 + ], + "blocks": [ + { + "bbox": [ + 55, + 188, + 553, + 200 + ], + "lines": [ + { + "bbox": [ + 55, + 188, + 553, + 200 + ], + "spans": [ + { + "bbox": [ + 55, + 188, + 553, + 200 + ], + "type": "text", + "content": "Table 2. Comparison with state-of-the-art methods on KITTI dataset. The best and the second best results are denoted by pink and blue." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 229, + 293, + 370 + ], + "lines": [ + { + "bbox": [ + 59, + 229, + 293, + 370 + ], + "spans": [ + { + "bbox": [ + 59, + 229, + 293, + 370 + ], + "type": "table", + "html": "
MethodPSNRSSIMDPSNRDSSIM
Aw/o track25.250.76121.710.662
Bw/o Lmotion28.390.84324.540.719
Cw/o iso.28.580.84725.350.765
D3D tracker28.020.83524.610.746
ED w/o Lmotion27.540.82123.530.703
FGT pose28.980.87325.710.787
GUNINEXT28.760.86425.430.772
HOurs28.850.86725.580.779
", + "image_path": "0215615f9d01b72de2778309eab04801b5c80074336643a07c779232d00dcc60.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 376, + 295, + 399 + ], + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 399 + ], + "type": "text", + "content": "Table 3. Ablation study of Novel View Synthesis on NOTR dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 420, + 295, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 295, + 444 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 295, + 444 + ], + "type": "text", + "content": "learning module. Please refer to the supplementary material for additional video results and editing examples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 444, + 295, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 444, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 54, + 444, + 295, + 564 + ], + "type": "text", + "content": "The results on the KITTI dataset are presented in Table 2. Our method outperforms the NeRF-based NSG and MARS across all metrics. In the KITTI-75% setting, our approach achieves a 0.045 higher SSIM and a 0.018 lower LPIPS compared to Street Gaussians, although its PSNR is 0.3 dB lower. A similar trend is observed in the KITTI-50% setting. This lower PSNR is primarily because Street Gaussians leverages labor-intensive ground truth tracking annotations, whereas our method uses a generalized 2D tracker—yet still attains comparable performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 571, + 156, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 156, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 156, + 582 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 588, + 295, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 295, + 612 + ], + "type": "text", + "content": "We conduct an ablation study on the NOTR dataset and the results are presented in Tab. 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 615, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 615, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 615, + 295, + 687 + ], + "type": "text", + "content": "Effect of object tracking module. In Tab. 3 (A), we model dynamic points using 4DGS [81] without performing any form of tracking, and observe significant drops across all metrics. This experiment highlights that the sole use of 4DGS does not provide sufficient accuracy for effective motion modeling in autonomous driving scenarios." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Effect of motion learning strategy. In Tab. 3 (B), we use the reconstruction of the tracking module for initialization" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 232, + 553, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 553, + 293 + ], + "type": "text", + "content": "but remove the motion loss introduced in Eq. 7, resulting in a " + }, + { + "bbox": [ + 313, + 232, + 553, + 293 + ], + "type": "inline_equation", + "content": "1.04\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 232, + 553, + 293 + ], + "type": "text", + "content": " drop in DPSNR. This result highlights the importance of our motion learning strategy. A comparison with rigid-transformation-based motion modeling [89] is provided in the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 296, + 553, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 296, + 553, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 296, + 553, + 344 + ], + "type": "text", + "content": "Effect of isotropic Gaussian marbles. In Tab. 3 (C), we substitute isotropic Gaussian marbles with the original anisotropic Gaussian ellipsoids for dynamic points, which leads to a decrease in DPSNR and DSSIM." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 348, + 555, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 555, + 480 + ], + "type": "text", + "content": "3D tracker / GT pose vs. Our object tracking module. In Tab. 3 (D), we use the 3D tracker [11, 66] to reconstruct dynamic objects and supervise motion learning, resulting in a 0.97 dB DPSNR decrease. Removing motion loss leads to a further 1.08 dB drop (Tab. 3 (E)), showing that our motion learning strategy compensates for tracking errors, though these errors still impact novel view synthesis, emphasizing the need for a more generalizable 2D tracker. Using ground-truth trajectories (Tab. 3 (F)) shows only a small gap, confirming the effectiveness of our 2D tracker in motion learning. See supplementary for tracking error analysis." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 484, + 553, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 553, + 533 + ], + "type": "text", + "content": "Choice of 2D tracker. To demonstrate the robustness of our method with respect to the choice of 2D tracker, we employ UNINEXT [88] for 2D tracking. As shown in Table 3 (G), our method maintains stable performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 548, + 388, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 388, + 560 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 388, + 560 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we introduce a novel framework for robust dynamic 3D street scene reconstruction that eliminates the reliance on 3D object trackers. Addressing the generalization limitations of 3D trackers, we propose a robust object tracking strategy based on a 2D foundation model. Our framework also features a motion learning module within an implicit feature space to handle inevitable tracking errors by autonomously refining pose inaccuracies and recovering missed detections. Experiments on the Waymo-NOTR and KITTI datasets demonstrate its adaptability and superior performance. Limitations and failure cases are discussed in the supplementary material." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25729" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "This work was supported by the National Key R&D Program of China under Grant 2018AAA0102801, PCL-CMCC Foundation for Science and Innovation Grant No. 2024ZY1C0040, and the National Research Foundation, Singapore, under its NRF-Investigatorship Programme (Award ID. NRF-NRFI09-0008)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 115, + 188 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 115, + 188 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 115, + 188 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 196, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 196, + 296, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 196, + 296, + 260 + ], + "spans": [ + { + "bbox": [ + 66, + 196, + 296, + 260 + ], + "type": "text", + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5855–5864, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 262, + 296, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 262, + 296, + 317 + ], + "spans": [ + { + "bbox": [ + 66, + 262, + 296, + 317 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470-5479, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 319, + 295, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 319, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 66, + 319, + 295, + 374 + ], + "type": "text", + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 376, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 376, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 66, + 376, + 294, + 419 + ], + "type": "text", + "content": "[4] Mk Bashar, Samia Islam, Kashifa Kawaakib Hussain, Md Bakhtiar Hasan, ABM Rahman, and Md Hasanul Kabir. Multiple object tracking in recent times: A literature review. arXiv preprint arXiv:2209.04796, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 421, + 295, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 421, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 66, + 421, + 295, + 464 + ], + "type": "text", + "content": "[5] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In 2016 IEEE international conference on image processing (ICIP), pages 3464-3468. IEEE, 2016. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 466, + 296, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 466, + 296, + 532 + ], + "spans": [ + { + "bbox": [ + 66, + 466, + 296, + 532 + ], + "type": "text", + "content": "[6] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 533, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 533, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 66, + 533, + 295, + 578 + ], + "type": "text", + "content": "[7] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. 2, 3, 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 66, + 579, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 579, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 66, + 579, + 295, + 633 + ], + "type": "text", + "content": "[8] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 66, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 66, + 635, + 295, + 689 + ], + "type": "text", + "content": "[9] Zhengping Che, Guangyu Li, Tracy Li, Bo Jiang, Xuefeng Shi, Xinsheng Zhang, Ying Lu, Guobin Wu, Yan Liu, and Jieping Ye. D2-city: a large-scale dashcam video dataset of diverse traffic scenarios. arXiv preprint arXiv:1904.01975, 2019.3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 61, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 691, + 295, + 713 + ], + "type": "text", + "content": "[10] Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 339, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 555, + 95 + ], + "type": "text", + "content": "reconstruction and real-time rendering. arXiv preprint arXiv:2311.18561, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 96, + 554, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 96, + 554, + 151 + ], + "spans": [ + { + "bbox": [ + 320, + 96, + 554, + 151 + ], + "type": "text", + "content": "[11] Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21674-21683, 2023. 1, 3, 6, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 320, + 152, + 553, + 195 + ], + "type": "text", + "content": "[12] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 2, 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 198, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 198, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 320, + 198, + 553, + 262 + ], + "type": "text", + "content": "[13] Jie Cheng, Yingbing Chen, Qingwen Zhang, Lu Gan, Chengju Liu, and Ming Liu. Real-time trajectory planning for autonomous driving with gaussian process and incremental refinement. In 2022 International Conference on Robotics and Automation (ICRA), pages 8999-9005. IEEE, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 264, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 264, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 321, + 264, + 553, + 319 + ], + "type": "text", + "content": "[14] Jie Cheng, Xiaodong Mei, and Ming Liu. Forecast-mae: Self-supervised pre-training for motion forecasting with masked autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8679-8689, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 321, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 321, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 321, + 321, + 553, + 374 + ], + "type": "text", + "content": "[15] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. ACM Transactions on Graphics (ToG), 34(4):1-13, 2015. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 376, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 376, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 321, + 376, + 553, + 421 + ], + "type": "text", + "content": "[16] Daniel Dauner, Marcel Hallgarten, Andreas Geiger, and Kashyap Chitta. Parting with misconceptions about learning-based vehicle motion planning. In Conference on Robot Learning, pages 1268-1281. PMLR, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 422, + 553, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 422, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 321, + 422, + 553, + 466 + ], + "type": "text", + "content": "[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 468, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 468, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 321, + 468, + 553, + 533 + ], + "type": "text", + "content": "[18] Yan Di, Ruida Zhang, Zhiqiang Lou, Fabian Manhardt, Xiangyang Ji, Nassir Navab, and Federico Tombari. Gpvpose: Category-level object pose estimation via geometry-guided point-wise voting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6781-6791, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 534, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 534, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 321, + 534, + 553, + 589 + ], + "type": "text", + "content": "[19] Shuxiao Ding, Eike Rehder, Lukas Schneider, Marius Cordts, and Juergen Gall. 3dmotformer: Graph transformer for online 3d multi-object tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9784-9794, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 590, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 590, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 321, + 590, + 553, + 633 + ], + "type": "text", + "content": "[20] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 321, + 635, + 553, + 689 + ], + "type": "text", + "content": "[21] George Eskandar. An empirical study of the generalization ability of lidar 3d object detectors to unseen domains. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23815-23825, 2024. 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 321, + 691, + 553, + 713 + ], + "type": "text", + "content": "[22] Tobias Fischer, Jonas Kulhanek, Samuel Rota Bulo, Lorenzo Porzi, Marc Pollefeys, and Peter Kontschieder." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25730" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 81, + 73, + 295, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 295, + 95 + ], + "type": "text", + "content": "Dynamic 3d gaussian fields for urban areas. arXiv preprint arXiv:2406.03175, 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 61, + 96, + 295, + 140 + ], + "type": "text", + "content": "[23] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 2, 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 141, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 141, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 61, + 141, + 294, + 184 + ], + "type": "text", + "content": "[24] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 186, + 295, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 186, + 295, + 241 + ], + "spans": [ + { + "bbox": [ + 62, + 186, + 295, + 241 + ], + "type": "text", + "content": "[25] Junru Gu, Chenxu Hu, Tianyuan Zhang, Xuanyao Chen, Yilun Wang, Yue Wang, and Hang Zhao. Vip3d: End-to-end visual trajectory prediction via 3d agent queries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5496-5506, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 243, + 295, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 243, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 62, + 243, + 295, + 297 + ], + "type": "text", + "content": "[26] Kaiwen Guo, Feng Xu, Yangang Wang, Yebin Liu, and Qionghai Dai. Robust non-rigid motion tracking and surface reconstruction using 10 regularization. In Proceedings of the IEEE International Conference on Computer Vision, pages 3083-3091, 2015. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 299, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 299, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 62, + 299, + 294, + 363 + ], + "type": "text", + "content": "[27] Kaiwen Guo, Peter Lincoln, Philip Davidson, Jay Busch, Xueming Yu, Matt Whalen, Geoff Harvey, Sergio Orts-Escolano, Rohit Pandey, Jason Dourgarian, et al. The relightables: Volumetric performance capture of humans with realistic relighting. ACM Transactions on Graphics (ToG), 38(6):1-19, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 365, + 294, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 365, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 62, + 365, + 294, + 410 + ], + "type": "text", + "content": "[28] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 411, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 411, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 62, + 411, + 294, + 465 + ], + "type": "text", + "content": "[29] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 468, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 468, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 468, + 294, + 510 + ], + "type": "text", + "content": "[30] Tao Hu, Tao Yu, Zerong Zheng, He Zhang, Yebin Liu, and Matthias Zwicker. Hvtr: Hybrid volumetric-textural rendering for human avatars. In 2022 International Conference on 3D Vision (3DV), pages 197-208. IEEE, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 512, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 512, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 512, + 294, + 567 + ], + "type": "text", + "content": "[31] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 568, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 568, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 62, + 568, + 294, + 622 + ], + "type": "text", + "content": "[32] Nan Huang, Xiaobao Wei, Wenzhao Zheng, Pengju An, Ming Lu, Wei Zhan, Masayoshi Tomizuka, Kurt Keutzer, and Shanghang Zhang. S3gaussian: Self-supervised street gaussians for autonomous driving. arXiv preprint arXiv:2405.20323, 2024. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 624, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 294, + 667 + ], + "type": "text", + "content": "[33] Yanjun Huang, Jiatong Du, Ziru Yang, Zewei Zhou, Lin Zhang, and Hong Chen. A survey on trajectory-prediction methods for autonomous driving. IEEE Transactions on Intelligent Vehicles, 7(3):652-674, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 294, + 713 + ], + "type": "text", + "content": "[34] Ziqin Huang, Gu Wang, Chenyangguang Zhang, Ruida Zhang, Xiu Li, and Xiangyang Ji. Givepose: Gradual intraclass variation elimination for rgb-based category-level object pose estimation. In Proceedings of the Computer Vision" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 94 + ], + "type": "text", + "content": "and Pattern Recognition Conference, pages 22055-22066, 2025.3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 96, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 96, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 320, + 96, + 553, + 162 + ], + "type": "text", + "content": "[35] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8350, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 320, + 163, + 553, + 206 + ], + "type": "text", + "content": "[36] Lesole Kalake, Wanggen Wan, and Li Hou. Analysis based on recent deep learning approaches applied in real-time multi-object tracking: a review. IEEE Access, 9:32650-32671, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 209, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 209, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 320, + 209, + 553, + 251 + ], + "type": "text", + "content": "[37] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 3, 4, 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 254, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 254, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 320, + 254, + 553, + 308 + ], + "type": "text", + "content": "[38] Mustafa Khan, Hamidreza Fazlali, Dhruv Sharma, Tongtong Cao, Dongfeng Bai, Yuan Ren, and Bingbing Liu. Autosplat: Constrained gaussian splatting for autonomous driving scene reconstruction. arXiv preprint arXiv:2407.02598, 2024. 1, 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 309, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 309, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 320, + 309, + 553, + 331 + ], + "type": "text", + "content": "[39] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 333, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 333, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 320, + 333, + 553, + 398 + ], + "type": "text", + "content": "[40] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 399, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 399, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 320, + 399, + 553, + 475 + ], + "type": "text", + "content": "[41] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision, 128(7): 1956-1981, 2020. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 478, + 553, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 478, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 320, + 478, + 553, + 532 + ], + "type": "text", + "content": "[42] Jinyu Li, Chenxu Luo, and Xiaodong Yang. Pillarnext: Rethinking network designs for 3d object detection in lidar point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17567-17576, 2023. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 534, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 534, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 320, + 534, + 553, + 578 + ], + "type": "text", + "content": "[43] Wei Li, CW Pan, Rong Zhang, JP Ren, YX Ma, Jin Fang, FL Yan, QC Geng, XY Huang, HJ Gong, et al. Aads: Augmented autonomous driving simulation using data-driven algorithms. Science robotics, 4(28):eaaw0863, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 579, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 579, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 320, + 579, + 553, + 632 + ], + "type": "text", + "content": "[44] Yanyan Li, Chenyu Lyu, Yan Di, Guangyao Zhai, Gim Hee Lee, and Federico Tombari. Geogaussian: Geometry-aware gaussian splatting for scene rendering. In European Conference on Computer Vision, pages 441-457. Springer, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 635, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 635, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 320, + 635, + 553, + 679 + ], + "type": "text", + "content": "[45] Zhong Li, Yu Ji, Wei Yang, Jinwei Ye, and Jingyi Yu. Robust 3d human motion reconstruction via dynamic template construction. In 2017 International Conference on 3D Vision (3DV), pages 496-505. IEEE, 2017. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 681, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 681, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 681, + 555, + 713 + ], + "type": "text", + "content": "[46] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In Proceedings of" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25731" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14864-14873, 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 96, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 96, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 61, + 96, + 295, + 149 + ], + "type": "text", + "content": "[47] Ming Liang, Bin Yang, Wenyuan Zeng, Yun Chen, Rui Hu, Sergio Casas, and Raquel Urtasun. Pnpnet: End-to-end perception and prediction with tracking in the loop. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11553-11562, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 151, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 151, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 61, + 151, + 294, + 205 + ], + "type": "text", + "content": "[48] Yiqing Liang, Numair Khan, Zhengqin Li, Thu Nguyen-Phuoc, Douglas Lanman, James Tompkin, and Lei Xiao. Gaufre: Gaussian deformation fields for real-time dynamic novel view synthesis. arXiv preprint arXiv:2312.11458, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 206, + 295, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 206, + 295, + 250 + ], + "spans": [ + { + "bbox": [ + 62, + 206, + 295, + 250 + ], + "type": "text", + "content": "[49] Yiyi Liao, Jun Xie, and Andreas Geiger. Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3292-3310, 2022. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 251, + 295, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 295, + 316 + ], + "type": "text", + "content": "[50] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 317, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 317, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 62, + 317, + 294, + 358 + ], + "type": "text", + "content": "[51] Xingyu Liu, Gu Wang, Yi Li, and Xiangyang Ji. CATRE: iterative point clouds alignment for category-level object pose refinement. In European Conference on Computer Vision, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 361, + 294, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 361, + 294, + 425 + ], + "spans": [ + { + "bbox": [ + 62, + 361, + 294, + 425 + ], + "type": "text", + "content": "[52] Xingyu Liu, Chenyangguang Zhang, Gu Wang, Ruida Zhang, and Xiangyang Ji. Rasim: A range-aware high-fidelity rgb-d data simulation pipeline for real-world applications. In 2024 IEEE international conference on robotics and automation (ICRA), pages 17057-17064. IEEE, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 427, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 427, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 427, + 294, + 491 + ], + "type": "text", + "content": "[53] Xingyu Liu, Yingyue Li, Chengxi Li, Gu Wang, Chenyangguang Zhang, Ziqin Huang, and Xiangyang Ji. Gfreedet: Exploiting gaussian splatting and foundation models for model-free unseen object detection in the bop challenge 2024. IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop, 2025. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 493, + 294, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 493, + 294, + 547 + ], + "spans": [ + { + "bbox": [ + 62, + 493, + 294, + 547 + ], + "type": "text", + "content": "[54] Xingyu Liu, Gu Wang, Ruida Zhang, Chenyangguang Zhang, Federico Tombari, and Xiangyang Ji. UNOPose: Unseen object pose estimation with an unposed rgb-d reference image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 548, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 548, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 62, + 548, + 294, + 602 + ], + "type": "text", + "content": "[55] Xingyu Liu, Ruida Zhang, Chenyangguang Zhang, Gu Wang, Jiwen Tang, Zhigang Li, and Xiangyang Ji. Gdnpp: A geometry-guided and fully learning-based object pose estimator. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 604, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 604, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 62, + 604, + 294, + 667 + ], + "type": "text", + "content": "[56] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "type": "text", + "content": "[57] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 321, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 553, + 116 + ], + "type": "text", + "content": "[58] Wenhan Luo, Junliang Xing, Anton Milan, Xiaoqin Zhang, Wei Liu, and Tae-Kyun Kim. Multiple object tracking: A literature review. Artificial intelligence, 293:103448, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 320, + 118, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 118, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 320, + 118, + 553, + 172 + ], + "type": "text", + "content": "[59] Jiageng Mao, Minzhe Niu, Chenhan Jiang, Hanxue Liang, Jingheng Chen, Xiaodan Liang, Yamin Li, Chaoqiang Ye, Wei Zhang, Zhenguo Li, et al. One million scenes for autonomous driving: Once dataset. arXiv preprint arXiv:2106.11037, 2021.3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 175, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 175, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 320, + 175, + 553, + 217 + ], + "type": "text", + "content": "[60] Jiageng Mao, Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. 3d object detection for autonomous driving: A comprehensive survey. International Journal of Computer Vision, 131(8):1909-1963, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 220, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 220, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 320, + 220, + 553, + 274 + ], + "type": "text", + "content": "[61] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 276, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 276, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 320, + 276, + 553, + 330 + ], + "type": "text", + "content": "[62] M Minderer, A Gritsenko, A Stone, M Neumann, D Weissenborn, A Dosovitskiy, A Mahendran, A Arnab, M Dehghani, Z Shen, et al. Simple open-vocabulary object detection with vision transformers. arxiv 2022. arXiv preprint arXiv:2205.06230, 2, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 332, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 332, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 320, + 332, + 553, + 376 + ], + "type": "text", + "content": "[63] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 377, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 377, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 320, + 377, + 553, + 432 + ], + "type": "text", + "content": "[64] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 434, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 434, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 320, + 434, + 553, + 487 + ], + "type": "text", + "content": "[65] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2856-2865, 2021. 1, 3, 6, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 490, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 490, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 320, + 490, + 553, + 533 + ], + "type": "text", + "content": "[66] Ziqi Pang, Zhichao Li, and Naiyan Wang. Simpletrack: Understanding and rethinking 3d multi-object tracking. In European Conference on Computer Vision, pages 680-696. Springer, 2022. 3, 6, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 535, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 535, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 320, + 535, + 553, + 611 + ], + "type": "text", + "content": "[67] Chensheng Peng, Chengwei Zhang, Yixiao Wang, Chenfeng Xu, Yichen Xie, Wenzhao Zheng, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. Desire-gs: 4d street gaussians for static-dynamic decomposition and surface reconstruction for urban driving scenes. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 6782-6791, 2025. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 613, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 613, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 320, + 613, + 553, + 678 + ], + "type": "text", + "content": "[68] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 681, + 553, + 713 + ], + "type": "text", + "content": "[69] Jules Sanchez, Jean-Emmanuel Deschaud, and François Goulette. Domain generalization of 3d semantic segmentation in autonomous driving. In Proceedings of the" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25732" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "type": "text", + "content": "IEEE/CVF International Conference on Computer Vision, pages 18077-18087, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 96, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 96, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 61, + 96, + 295, + 139 + ], + "type": "text", + "content": "[70] Sara Fridovich-Keil and Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In CVPR, 2023. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 140, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 140, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 61, + 140, + 294, + 194 + ], + "type": "text", + "content": "[71] Shital Shah, Debadeepta Dey, Chris Lovett, and Ashish Kapoor. Airsim: High-fidelity visual and physical simulation for autonomous vehicles. In Field and Service Robotics: Results of the 11th International Conference, pages 621-635. Springer, 2018. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 195, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 195, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 62, + 195, + 295, + 239 + ], + "type": "text", + "content": "[72] Yeji Song, Chaerin Kong, Seoyoung Lee, Nojun Kwak, and Joonseok Lee. Towards efficient neural scene graphs by learning consistency fields. arXiv preprint arXiv:2210.04127, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 239, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 295, + 293 + ], + "type": "text", + "content": "[73] Louis Soum-Fontez, Jean-Emmanuel Deschaud, and François Goulette. Mdt3d: Multi-dataset training for lidar 3d object detection generalization. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5765-5772. IEEE, 2023. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 295, + 295, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 295, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 62, + 295, + 295, + 348 + ], + "type": "text", + "content": "[74] Colton Stearns, Adam Harley, Mikaela Uy, Florian Dubost, Federico Tombari, Gordon Wetzstein, and Leonidas Guibas. Dynamic gaussian marbles for novel view synthesis of casual monocular videos. arXiv preprint arXiv:2406.18717, 2024. 3, 4, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 350, + 295, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 350, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 62, + 350, + 295, + 415 + ], + "type": "text", + "content": "[75] Zhuo Su, Lan Xu, Zerong Zheng, Tao Yu, Yebin Liu, and Lu Fang. Robustfusion: Human volumetric capture with data-driven visual cues using a rgbd camera. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 246-264. Springer, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 416, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 416, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 416, + 295, + 491 + ], + "type": "text", + "content": "[76] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 2, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 493, + 295, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 493, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 62, + 493, + 295, + 548 + ], + "type": "text", + "content": "[77] Shanlin Sun, Bingbing Zhuang, Ziyu Jiang, Buyu Liu, Xiaohui Xie, and Manmohan Chandraker. Lidarf: Delving into lidar for neural radiance field on street scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19563-19572, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 548, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 548, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 548, + 295, + 601 + ], + "type": "text", + "content": "[78] Haithem Turki, Jason Y Zhang, Francesco Ferroni, and Deva Ramanan. Suds: Scalable urban dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12375-12385, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 603, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 603, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 62, + 603, + 295, + 669 + ], + "type": "text", + "content": "[79] Yan Wang, Xiangyu Chen, Yurong You, Li Erran Li, Bharath Hariharan, Mark Campbell, Kilian Q Weinberger, and Wei-Lun Chao. Train in germany, test in the usa: Making 3d object detectors generalize. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11713-11723, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 295, + 713 + ], + "type": "text", + "content": "[80] Yu Wang, Shaohua Wang, Yicheng Li, and Mingchun Liu. A comprehensive review of 3d object detection in autonomous driving: Technological advances and future directions. arXiv preprint arXiv:2408.16530, 2024. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 320, + 73, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 553, + 139 + ], + "type": "text", + "content": "[81] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20310-20320, 2024. 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 320, + 141, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 141, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 320, + 141, + 553, + 194 + ], + "type": "text", + "content": "[82] Hai Wu, Wenkai Han, Chenglu Wen, Xin Li, and Cheng Wang. 3d multi-object tracking in point clouds based on prediction confidence-guided data association. IEEE Transactions on Intelligent Transportation Systems, 23(6):5668-5677, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 197, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 197, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 320, + 197, + 553, + 240 + ], + "type": "text", + "content": "[83] Hai Wu, Jinhao Deng, Chenglu Wen, Xin Li, Cheng Wang, and Jonathan Li. Casa: A cascade attention network for 3-d object detection from lidar point clouds. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 243, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 243, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 320, + 243, + 553, + 296 + ], + "type": "text", + "content": "[84] Junfeng Wu, Yi Jiang, Qihao Liu, Zehuan Yuan, Xiang Bai, and Song Bai. General object foundation model for images and videos at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3783-3795, 2024. 1, 2, 3, 4, 6, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 298, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 298, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 320, + 298, + 553, + 363 + ], + "type": "text", + "content": "[85] Zirui Wu, Tianyu Liu, Liyi Luo, Zhide Zhong, Jianteng Chen, Hongmin Xiao, Chao Hou, Haozhe Lou, Yuantao Chen, Runyi Yang, et al. Mars: An instance-aware, modular and realistic simulator for autonomous driving. In CAAI International Conference on Artificial Intelligence, pages 3-15. Springer, 2023. 1, 3, 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 365, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 365, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 320, + 365, + 553, + 430 + ], + "type": "text", + "content": "[86] Pengchuan Xiao, Zhenlei Shao, Steven Hao, Zishuo Zhang, Xiaolin Chai, Judy Jiao, Zesong Li, Jian Wu, Kai Sun, Kun Jiang, et al. Pandaset: Advanced sensor suite dataset for autonomous driving. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 3095-3101. IEEE, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 433, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 433, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 320, + 433, + 553, + 465 + ], + "type": "text", + "content": "[87] Ziyang Xie, Junge Zhang, Wenye Li, Feihu Zhang, and Li Zhang. S-nerf: Neural radiance fields for street views. arXiv preprint arXiv:2303.00749, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 467, + 555, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 467, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 320, + 467, + 555, + 521 + ], + "type": "text", + "content": "[88] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15325-15336, 2023. 1, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 523, + 553, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 523, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 320, + 523, + 553, + 576 + ], + "type": "text", + "content": "[89] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians: Modeling dynamic urban scenes with gaussian splatting. In ECCV, 2024. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 579, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 579, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 320, + 579, + 553, + 633 + ], + "type": "text", + "content": "[90] Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Seung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, et al. Emernerf: Emergent spatial-temporal scene decomposition via self-supervision. arXiv preprint arXiv:2311.02077, 2023. 3, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 320, + 635, + 553, + 689 + ], + "type": "text", + "content": "[91] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1389-1399, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 692, + 553, + 713 + ], + "type": "text", + "content": "[92] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time photorealistic dynamic scene representation and" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25733" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 81, + 73, + 295, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 295, + 95 + ], + "type": "text", + "content": "rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 96, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 96, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 61, + 96, + 295, + 162 + ], + "type": "text", + "content": "[93] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20331-20341, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 163, + 295, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 163, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 61, + 163, + 295, + 207 + ], + "type": "text", + "content": "[94] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 209, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 209, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 62, + 209, + 294, + 262 + ], + "type": "text", + "content": "[95] Fisher Yu, Wenqi Xian, Yingying Chen, Fangchen Liu, Mike Liao, Vashisht Madhavan, Trevor Darrell, et al. Bdd100k: A diverse driving video database with scalable annotation tooling. arXiv preprint arXiv:1805.04687, 2(5): 6, 2018. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 264, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 264, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 62, + 264, + 295, + 319 + ], + "type": "text", + "content": "[96] Jiang-Tian Zhai, Ze Feng, Jinhao Du, Yongqiang Mao, Jiang-Jiang Liu, Zichang Tan, Yifu Zhang, Xiaqing Ye, and Jingdong Wang. Rethinking the open-loop evaluation of end-to-end autonomous driving in nuscenes. arXiv preprint arXiv:2305.10430, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 321, + 295, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 321, + 295, + 375 + ], + "spans": [ + { + "bbox": [ + 62, + 321, + 295, + 375 + ], + "type": "text", + "content": "[97] Bo Zhang, Jiakang Yuan, Botian Shi, Tao Chen, Yikang Li, and Yu Qiao. Uni3d: A unified baseline for multi-dataset 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9253-9262, 2023. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 377, + 295, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 377, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 62, + 377, + 295, + 410 + ], + "type": "text", + "content": "[98] Peng Zhang, Xin Li, Liang He, and Xin Lin. 3d multiple object tracking on autonomous driving: A literature review. arXiv preprint arXiv:2309.15411, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 411, + 295, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 411, + 295, + 466 + ], + "spans": [ + { + "bbox": [ + 62, + 411, + 295, + 466 + ], + "type": "text", + "content": "[99] Ruida Zhang, Yan Di, Zhiqiang Lou, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. Rbp-pose: Residual bounding box projection for category-level pose estimation. In European conference on computer vision, pages 655–672. Springer, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 468, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 468, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 57, + 468, + 295, + 533 + ], + "type": "text", + "content": "[100] Ruida Zhang, Yan Di, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. Ssp-pose: Symmetry-aware shape prior deformation for direct category-level object pose estimation. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7452–7459. IEEE, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 535, + 295, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 535, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 57, + 535, + 295, + 600 + ], + "type": "text", + "content": "[101] Ruida Zhang, Ziqin Huang, Gu Wang, Chenyangguang Zhang, Yan Di, Xingxing Zuo, Jiwen Tang, and Xiangyang Ji. Lapose: Laplacian mixture shape modeling for rgb-based category-level object pose estimation. In European Conference on Computer Vision, pages 467-484. Springer, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 602, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 668 + ], + "type": "text", + "content": "[102] Ruida Zhang, Chenyangguang Zhang, Yan Di, Fabian Manhardt, Xingyu Liu, Federico Tombari, and Xiangyang Ji. Kp-red: Exploiting semantic keypoints for joint 3d shape retrieval and deformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20540-20550, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 669, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 669, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 669, + 295, + 712 + ], + "type": "text", + "content": "[103] Xingxuan Zhang, Zekai Xu, Renzhe Xu, Jiashuo Liu, Peng Cui, Weitao Wan, Chong Sun, and Chen Li. Towards domain generalization in object detection. arXiv preprint arXiv:2203.14387, 2022. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 251 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[104] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. International journal of computer vision, 129:3069–3087, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 118, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 183 + ], + "type": "text", + "content": "[105] Hongyu Zhou, Jiahao Shao, Lu Xu, Dongfeng Bai, Weichao Qiu, Bingbing Liu, Yue Wang, Andreas Geiger, and Yiyi Liao. Hugs: Holistic urban 3d scene understanding via gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21336-21345, 2024. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 185, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 251 + ], + "type": "text", + "content": "[106] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21634-21643, 2024. 1, 2, 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25734" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Stroke2Sketch_ Harnessing Stroke Attributes for Training-Free Sketch Generation/88f15fb1-f5ab-4157-a23c-15d3cb3bc8f3_content_list.json b/2025/Stroke2Sketch_ Harnessing Stroke Attributes for Training-Free Sketch Generation/88f15fb1-f5ab-4157-a23c-15d3cb3bc8f3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..86906df215ad820de3275d3a2a5fb3776e4ad80c --- /dev/null +++ b/2025/Stroke2Sketch_ Harnessing Stroke Attributes for Training-Free Sketch Generation/88f15fb1-f5ab-4157-a23c-15d3cb3bc8f3_content_list.json @@ -0,0 +1,1625 @@ +[ + { + "type": "text", + "text": "Stroke2Sketch: Harnessing Stroke Attributes for Training-Free Sketch Generation", + "text_level": 1, + "bbox": [ + 101, + 140, + 893, + 161 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rui Yang $^{1,3}$ , Huining Li $^{4,5}$ , Yiyi Long $^{2,5}$ , Xiaojun Wu $^{3,*}$ , Shengfeng He $^{5,*}$", + "bbox": [ + 215, + 188, + 779, + 207 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Huaqiao University $^{2}$ South China University of Technology $^{3}$ Shaanxi Normal University $^{4}$ Beijing University of Aeronautics and Astronautics $^{5}$ Singapore Management University", + "bbox": [ + 142, + 205, + 854, + 242 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/537aacecb15601e21cbc49becd56d0f3bac6c79f6805b55df94b65f75014510a.jpg", + "image_caption": [ + "Figure 1. We propose Stroke2Sketch, a framework that accurately transfers stroke attributes from a reference sketch to a content image while preserving structure and style fidelity. The top row shows reference sketches, the leftmost column displays content images, and the central and right columns illustrate our method's precise content preservation and expressive stroke transfer." + ], + "image_footnote": [], + "bbox": [ + 124, + 250, + 870, + 481 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 254, + 536, + 330, + 550 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generating sketches guided by reference styles requires precise transfer of stroke attributes, such as line thickness, deformation, and texture sparsity, while preserving semantic structure and content fidelity. To this end, we propose Stroke2Sketch, a novel training-free framework that introduces cross-image stroke attention, a mechanism embedded within self-attention layers to establish fine-grained semantic correspondences and enable accurate stroke attribute transfer. This allows our method to adaptively integrate reference stroke characteristics into content images while maintaining structural integrity. Additionally, we develop adaptive contrast enhancement and semantic-focused attention to reinforce content preservation and foreground emphasis. Stroke2Sketch effectively synthesizes stylistically faithful sketches that closely resemble handcrafted results, outperforming existing methods in expressive stroke control and semantic coherence. Codes are available at https://github.com/rane7/Stroke2Sketch.", + "bbox": [ + 98, + 571, + 485, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 536, + 640, + 550 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generating stylized sketches from content images using reference stroke patterns presents a key challenge at the intersection of artistic rendering and semantic-aware synthesis. Traditional sketch algorithms [28, 34] generate procedural line drawings, while vector-based methods [24, 26, 38] produce clean parametric strokes. However, these methods lack the adaptability to transfer diverse artistic styles from exemplar sketches due to their rigid, data-agnostic architectures. Unlike human artists, who strategically vary stroke attributes such as thickness, curvature, and density to emphasize key semantic features while maintaining content fidelity, existing approaches struggle to capture this nuanced interplay between stroke semantics and structure.", + "bbox": [ + 509, + 563, + 895, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent learning-based methods [2, 5, 33] attempt to address this limitation by training on clustered sketch styles, yet as shown in Fig. 2(a-b), they fail to generalize to unseen stroke patterns due to catastrophic forgetting. Meanwhile, diffusion-based stylization techniques [21, 41, 47, 49, 50] excel in texture transfer but struggle with structural integrity due to content leakage in cross-attention layers (Fig. 2(c-d)). While new conditioning mechanisms [18, 25, 44] attempt to enhance structural control, they often introduce style dis", + "bbox": [ + 509, + 757, + 895, + 888 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 116, + 16, + 189, + 55 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 245, + 13, + 799, + 56 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors: Xiaojun Wu (xjwu@snnu.edu.cn) and Shengfeng He (shengfenghe@smu.edu.sg)", + "bbox": [ + 101, + 864, + 482, + 888 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16545", + "bbox": [ + 480, + 926, + 517, + 938 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2779d5cc0e7051ec1056b65b671bcf2ad5e1498cd4f647273e68de9fe6dcd703.jpg", + "image_caption": [ + "Figure 2. (a) and (b) show results from training-based methods (Ref2Sketch and Semi-Ref2Sketch), which struggle with unseen styles, leading to poor content alignment. (c) and (d) illustrate IP-Adapter and InstantStyle results, retaining style but lacking content alignment. (e) to (g) show ControlNet-based methods, preserving content but failing to match reference styles. (h) shows our method's output, achieving superior content preservation and style alignment without using ControlNet. Detailed prompts are shown in the supplementary materials." + ], + "image_footnote": [], + "bbox": [ + 102, + 99, + 893, + 329 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tortion or require dense user inputs. Hybrid approaches like ControlNet [52] enforce structural priors but sacrifice stylistic flexibility, leading to overly rigid outputs (Fig. 2(e)). Progressive stroke-based methods [32] introduce semantic incoherence by applying uniform strokes across the image (Fig. 2(g)).", + "bbox": [ + 99, + 407, + 483, + 496 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We identify three fundamental challenges in reference-based sketch synthesis:(i) Semantic-aware stroke transfer. Effective style adaptation requires precise mapping of reference stroke attributes (e.g., tapered lines, cross-hatching) to semantically relevant content regions. (ii) Foreground prioritization. Artists naturally emphasize foreground elements using varied stroke density and complexity while simplifying backgrounds, yet existing methods apply uniform stylization, disrupting this compositional balance[32]. (iii) Content-style equilibrium. Since sketches encode content through linework, balancing structural preservation with style transfer is critical. Techniques such as CLIP-space style subtraction [41] fail to maintain this balance, as even minor content leakage distorts key edges (Fig. 2(c-d)).", + "bbox": [ + 99, + 500, + 483, + 705 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose Stroke2Sketch, a framework that enables precise stroke attribute transfer while maintaining content fidelity. Our key insight is that stroke properties, like line thickness, curvature, and texture sparsity, are inherently encoded within the self-attention and cross-attention relationships of pretrained diffusion models. By dynamically aligning these attention patterns between content and reference features, we achieve effective style transfer without structural degradation. Stroke2Sketch integrates three novel components tailored to each identified challenge:", + "bbox": [ + 99, + 709, + 483, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(1) Cross-image stroke attention tackles the challenge of", + "bbox": [ + 99, + 875, + 482, + 888 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "semantic-aware stroke transfer by facilitating stroke attribute exchange through key-value swapping in diffusion layers. Instead of directly blending features [41], we leverage attention blocks to transplant stroke characteristics onto content structures, preventing the entanglement of style and geometry. This approach ensures accurate stroke mapping without disrupting semantic coherence, as shown in Fig. 2(h).", + "bbox": [ + 511, + 407, + 895, + 511 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(2) Directive Attention Module ensures that stroke transfer remains compositionally balanced. Background textures often introduce conflicting patterns that dilute the intended style. We mitigate this by clustering self-attention maps to isolate foreground objects, then restricting cross-image attention to these regions. This mimics how artists prioritize focal elements while simplifying less critical areas, enhancing both style consistency and perceptual quality.", + "(3) Semantic Preservation Module addresses content-style equilibrium by injecting content contours as positional queries during early denoising. This hybridizes the precision of edge detectors with the flexibility of text-driven generation, allowing structural guidance without rigid constraints. Unlike ControlNet, which enforces strict boundaries, our approach treats edges as \"soft constraints\" that evolve into stylized strokes, preserving both structure and artistic abstraction." + ], + "bbox": [ + 511, + 513, + 895, + 767 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As validated in Fig. 2(h) and Fig. 6, Stroke2Sketch achieves state-of-the-art performance across diverse sketch styles. Experimental results show that our method outperforms adapter-based and ControlNet methods in both style alignment (87% user preference) and content preservation (92% accuracy in line correspondence tests). Importantly, it achieves these results without dataset-specific training or architectural modifications, demonstrating that pretrained", + "bbox": [ + 511, + 771, + 895, + 888 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16546", + "bbox": [ + 480, + 926, + 517, + 938 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "diffusion models can master sketch stylization when guided by principled feature interactions.", + "bbox": [ + 99, + 102, + 485, + 132 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 101, + 146, + 241, + 162 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Photo-Sketch Synthesis", + "text_level": 1, + "bbox": [ + 101, + 171, + 312, + 186 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generating a sketch from a content image based on a reference style parallels edge detection, as both tasks emphasize prominent visual features. Edge detection methods [3, 30, 36, 37, 45, 54], which detect sharp changes in color or brightness, form the foundation of sketch extraction but are limited to a single style and often produce artifacts, like scattered dots or broken lines. High-quality sketch generation requires more than edge detection; it demands line style, texture sparsity, and semantic abstraction to achieve artist-level results.", + "bbox": [ + 99, + 193, + 485, + 338 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning-based approaches have improved sketch realism by enhancing boundary detection and rendering distinct line styles. For example, Chan et al. [5] incorporated depth and semantic information for better sketch quality, while Ref2Sketch [2] and Semi-Ref2Sketch [33] leverage paired and semi-supervised training for stylized sketch extraction. However, these methods rely on large sketch datasets, which are challenging to obtain, limiting model robustness.", + "bbox": [ + 99, + 340, + 483, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Another research area, stroke-based rendering, focuses on manipulating strokes and contours for sketching. Methods like CLIPDraw [10] and CLIPasso [38] employ Bezier curves to create abstract sketches with high-level semantic simplification. StrokeAggregator [22] and StripMaker [23] refine vector sketches, achieving quality comparable to artist drawings. Other methods include semantic concept-to-sketch methods [4, 8, 15, 51]. However, these methods assume uniform style, while real-world sketches vary widely.", + "bbox": [ + 99, + 459, + 483, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Building on these insights, our method combines stroke attributes, semantic abstraction, and expression to better align generated sketches with diverse reference styles.", + "bbox": [ + 101, + 590, + 483, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. High-Semantic Style Transfer", + "text_level": 1, + "bbox": [ + 101, + 646, + 359, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generating sketches that adhere to both content and reference style is a specialized style transfer task. Advances in diffusion models have propelled style transfer through self- and cross-attention mechanisms, which preserve spatial layout and stylized content. Techniques such as Prompt-to-Prompt [11] and $\\mathrm{P + }$ [40] show how attention mechanisms can maintain structural coherence while enabling flexible style and semantic control.", + "bbox": [ + 99, + 667, + 485, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cross-Image Attention (CIA) [1] and methods like Swapping Self-Attention [17] and StyleAligned [12] reveal that style features are best retained during upsampling stages, even though content leakage may occur in bottleneck phases. IP-Adapter achieves style transfer via dual cross-attention with text and image, though it can weaken text control and lead to leakage.", + "bbox": [ + 99, + 786, + 485, + 888 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "InstantStyle [41] addresses leakage by subtracting features in the same feature space but often requires external constraints like ControlNet [52] for image-to-image generation, which can dilute style fidelity. Further developments like InstantStyle-plus [42] and RB-Modulation [32] incorporate CSD [35] to better align styles during generation, but limitations remain in sketch generation. Prompt-to-style transfer, however, demonstrates pretrained models' strong semantic alignment capabilities, providing valuable priors for consistent, high-level sketch synthesis.", + "bbox": [ + 511, + 102, + 895, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Stroke2Sketch", + "text_level": 1, + "bbox": [ + 511, + 261, + 653, + 276 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a content image $I^{cnt} \\in \\mathbb{R}^{H^{cnt} \\times W^{cnt} \\times 3}$ and a reference sketch image $I^{ref} \\in \\mathbb{R}^{H^{ref} \\times W^{ref} \\times 3}$ , our task is to generate a sketch $I^{ske}$ that aligns with the content structure of $I^{cnt}$ while adhering to the stroke style, texture sparsity, and high-level semantic abstraction of $I^{ref}$ . Additionally, we aim to remove or retain background elements as needed to enhance the foreground object. In this work, we address the sketch generation task using a controllable text-to-image guidance approach. Specifically, we extract object prompts from the content image $I^{cnt}$ using BLIP [20] and incorporate the stroke style and high-level abstraction from the reference sketch $I^{ref}$ to re-render the final target sketch $I^{ske}$ . To achieve this, we leverage a pre-trained text-to-image model. The overall network architecture is illustrated in Fig 3, with detailed explanations of each module provided below.", + "bbox": [ + 511, + 284, + 895, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 515, + 648, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DDPM inversion [16] is a process in diffusion-based generative models that reverses denoising steps, enabling the reconstruction of latent representations from generated outputs. Compared to DDIM inversion method [48], DDPM inversion offers greater flexibility for editing tasks by producing noise maps that, while correlated across timesteps and not normally distributed, allow precise image reconstruction and meaningful edits such as color adjustments and structural shifts. A key advantage of DDPM inversion is its ability to maintain an image's structure while altering the conditioning input, such as a text prompt, to enable artifact-free edits that adapt semantics while preserving original details. This efficient method bypasses optimization processes and can enhance diffusion-based editing techniques by improving image fidelity and supporting diverse outputs.", + "bbox": [ + 511, + 537, + 895, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Stable Diffusion [31] incorporates this inversion process within a latent space, rather than a pixel space, increasing computational efficiency and expressiveness. The input image is encoded through a pretrained variational autoencoder (VAE) to produce a latent code $z$ . Denoising then occurs within this latent space using a U-Net architecture that incorporates self-attention and cross-attention mechanisms. Self-attention blocks enhance image detail by calculating attention scores between projected query $Q$ , key $K$ , and value", + "bbox": [ + 511, + 757, + 895, + 888 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16547", + "bbox": [ + 480, + 926, + 517, + 938 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/882d2df4e65ecfec500fba46b0e2dfd429bef22513cbc659af43661b6d93875a.jpg", + "image_caption": [ + "Figure 3. The Stroke2Sketch architecture. The content image $I^{\\mathrm{cnt}}$ undergoes contour detection, generating feature representations $z^{\\mathrm{cont}}$ and $z^{\\mathrm{cnt}}$ , while the reference sketch $I^{\\mathrm{ref}}$ is inverted to produce $z^{\\mathrm{ref}}$ . The Directive Attention Module (DAM) aligns high-level semantic features between the content and reference features by emphasizing cross-image semantic correspondences $A_{n}$ . Self-attention maps $F_{SA}$ are aggregated and clustered to produce segmentation masks $M_{i}$ , which help distinguish foreground from background regions. Cross-image Stroke Attention (CSA) transfers stroke attributes, and the Semantic Preservation Module (SPM) enforces semantic alignment and stroke fidelity in the generated sketch $\\hat{I}^{\\mathrm{ske}}$ via loss $\\mathcal{L}_{\\mathrm{sem}}$ and contour-based structural integration. Lastly, Stroke Detail Propagation Enhancement (SDPE) refines details, resulting in the final output sketch $I^{\\mathrm{ske}}$ . SD represents the pre-trained diffusion model." + ], + "image_footnote": [], + "bbox": [ + 106, + 103, + 890, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$V$ vectors:", + "bbox": [ + 102, + 555, + 173, + 568 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA = \\operatorname {s o f t m a x} \\left(\\frac {Q \\cdot K ^ {\\top}}{\\sqrt {d}}\\right), \\quad \\phi = A \\cdot V, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 578, + 483, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $A$ is the attention map, $d$ is the dimensionality, and $\\phi$ is the output of the self-attention block. Cross-attention blocks incorporate conditioning inputs (e.g., text prompts) to guide the generation process.", + "bbox": [ + 99, + 621, + 483, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, the refined latent representation $z$ is decoded back into an RGB image using the VAE decoder, yielding a high-quality output guided by the conditioning input.", + "bbox": [ + 99, + 680, + 483, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Cross-image Stroke Attention", + "text_level": 1, + "bbox": [ + 101, + 733, + 359, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As discussed in the Introduction, sketch generation as a specialized form of style transfer requires attention to local stroke attributes and consistent texture abstraction across different levels of semantic sparsity. Edge detection serves as the foundation for this task, with edges defining content contours as the most effective strategy for preserving structural information. We extract the contour $I^{cont}$ from the content image using TEED [36] and obtain the inverted latents $z_{T}^{cnt}$ , $z_{T}^{ref}$ , and $z_{T}^{cont}$ for the content image, reference sketch, and", + "bbox": [ + 99, + 755, + 485, + 891 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "contour image using DDPM inversion. During the denoising process, the latent $z_{T}^{cnt}$ from the content image serves as the initial noise $z_{T}^{ske}$ for sketch denoising. At each timestep $t$ , we apply Equation 1.", + "bbox": [ + 511, + 555, + 893, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To achieve effective stroke feature transfer, we utilize latent representations for the content, reference, and contour images obtained through DDPM inversion. For predefined timesteps $t \\in \\{0, \\dots, T\\}$ , the reference sketch $I^{ref}$ , content image $I^{cnt}$ , and contour image $I^{cont}$ are inverted from their initial state ( $t = 0$ ) to Gaussian noise ( $t = T$ ). During DDPM inversion, we also collect the query features of the content ( $Q_{t}^{cnt}$ ), and the key and value features of the reference sketch ( $K_{t}^{ref}$ , $V_{t}^{ref}$ ) at each timestep.", + "bbox": [ + 509, + 614, + 895, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We initialize the latent noise $z_{T}^{ske}$ for the stylized sketch by copying the content latent noise $z_{T}^{cnt}$ . To blend features from all three images, we combine the reference keys and values $K_{t}^{ref}$ and $\\tilde{V}_{t}^{ref}$ with the content features. This integration is achieved by mixing the keys and values using the following formulation:", + "bbox": [ + 509, + 748, + 895, + 835 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nK _ {t} ^ {\\mathrm {s k e}} = K _ {t} ^ {\\mathrm {r e f}} + \\alpha K _ {t} ^ {\\mathrm {c n t}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 629, + 848, + 893, + 864 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {t} ^ {\\mathrm {s k e}} = V _ {t} ^ {\\mathrm {r e f}} + \\alpha V _ {t} ^ {\\mathrm {c n t}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 867, + 893, + 885 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16548", + "bbox": [ + 480, + 926, + 517, + 938 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/916b50d410eacea00395eb201704d9cd32dba2a002a668c04813f9289408722e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 107, + 103, + 230, + 161 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/50bdac387e457df1f2c4051ad7ca7814f72f93917c92d4ccd6d9b321fc8fc004.jpg", + "image_caption": [ + "Content Strokes" + ], + "image_footnote": [], + "bbox": [ + 114, + 162, + 230, + 218 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3c75c4095600f2a7e39dbdc113f186be63fda64ee19dc137fb7bfd073b475ba0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 103, + 352, + 161 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/543474c3d275e0516843866f2c5f783a6c7a9a57f9fb6727f66544c37412800c.jpg", + "image_caption": [ + "Reference Strokes" + ], + "image_footnote": [], + "bbox": [ + 241, + 167, + 348, + 213 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/67f0c0896f638805e512dcf825aa7fff932a2333abce36eac2f84f9742d80291.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 103, + 473, + 161 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0326bad4b692da42f634f42cb19be639ba3cc52f4612d589683911b057c66903.jpg", + "image_caption": [ + "Aligned Strokes", + "Figure 4. Stroke alignment results with CSA. The feature images are shown in the first row. The second row shows results where feature alignment between the key and value exchange is applied." + ], + "image_footnote": [], + "bbox": [ + 354, + 162, + 470, + 218 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ is a scalar parameter that controls the mixing ratio of content and reference features, allowing flexible adaptation to different stroke characteristics in the sketch.", + "bbox": [ + 99, + 297, + 483, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By exchanging and blending features from the content, reference, and contour images, this approach enables the effective transfer of stroke attributes and semantic alignment. However, we found that some strokes may fail to represent the intended content curves accurately (see Fig. 4). To improve output quality and highlight the foreground, we introduce additional mechanisms to guide sketch generation, as detailed below.", + "bbox": [ + 99, + 340, + 483, + 457 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Directive Attention Module", + "text_level": 1, + "bbox": [ + 101, + 470, + 341, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To concentrate stylization on perceptually significant regions, the DAM enhances foreground focus during stroke transfer. We extract self-attention feature maps $F_{SA}$ at $32 \\times 32$ resolution, aggregating them across channels via averaging. These maps are segmented into clusters $M_{j}$ using KMeans clustering. For each cluster $j$ and noun $n$ extracted from the content image's caption (via BLIP [20]), we calculate a relevance score:", + "bbox": [ + 99, + 492, + 483, + 607 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr (j, n) = \\frac {\\sum_ {(x , y)} M _ {j} (x , y) \\cdot A _ {n} (x , y)}{\\sum_ {(x , y)} M _ {j} (x , y) + \\delta}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 618, + 483, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $A_{n}$ is the cross-attention map for noun $n$ , and $\\delta = 10^{-5}$ stabilizes the computation. Clusters with $r(j,n) > 0.35$ are designated as foreground regions; remaining areas are suppressed, directing stroke stylization to salient elements (Fig. 5).", + "bbox": [ + 99, + 667, + 482, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By integrating the segmented self-attention mechanism with cross-image attention, DAM allows precise control over foreground regions, ensuring that stylistic features, such as line styles and high-level semantic abstractions, are faithfully transferred from the reference sketch to the content image. This approach achieves high fidelity in style and content alignment, yielding sketches that closely reflect the intended reference style with minimized background interference.", + "bbox": [ + 99, + 741, + 482, + 858 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As illustrated in Fig. 6, our method successfully aligns high-level semantic features and detailed stroke attributes,", + "bbox": [ + 101, + 859, + 483, + 888 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1a2b12d3d4bb37065f294fc3d1bca357fde672634747f6add6c2260e06029624.jpg", + "image_caption": [ + "Figure 5. Examples of DAM in action. The first column shows the content images, the second column displays segmentation maps obtained by clustering self-attention maps, the third column provides the reference sketches, and the fourth column illustrates the sketches generated by DAM after applying stroke attribute transfer and segmentation." + ], + "image_footnote": [], + "bbox": [ + 516, + 103, + 612, + 175 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d3811bdf5b815f468ac4b1ac1a816a5a7e60fc3713e246f928b19c0fb34aa961.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 103, + 704, + 175 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/efab180c65eeab9dc62253d9bed59e8cc840f4ff5778773ab44d31900d83fc7f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 103, + 890, + 175 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d8543fee5368f488fb073440f20cb758a02e4566888a46cb536b70cb476234a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 273, + 638, + 375 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1b5d661283d7becd8dd705ead768a7b6c14c027dcd521b62502e32a5675f6e8c.jpg", + "image_caption": [ + "Figure 6. Comparison of sketches generated using DAM. The top row shows Cnt. and Ref. images with stroke styles reflecting specific high-level attributes such as hair, eyebrows, and abstracted clothing texture. The second to fourth columns display sketches generated to match the content images in the top row, each adopting the stroke styles and high-level features from the reference. The bottom row highlights zoomed-in areas to emphasize the transferred stroke attributes." + ], + "image_footnote": [], + "bbox": [ + 638, + 273, + 715, + 376 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9231b8af5b80fd7ce3c5aaa1552821f3a3b982f4f606084f788a51cd3d165235.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 273, + 893, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "producing coherent sketches that maintain stylistic and structural consistency with the reference. This approach minimizes background interference while ensuring that the final sketch reflects the nuanced stroke characteristics of the reference.", + "bbox": [ + 511, + 503, + 893, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Semantic Preservation Module", + "text_level": 1, + "bbox": [ + 511, + 587, + 777, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although the injection of keys and values during sketch generation can effectively transfer stroke attributes to $I^{ske}$ , semantic inconsistencies may arise, particularly when the reference sketch $I^{ref}$ does not align semantically with the content image $I^{cnt}$ . This can lead to noise and misplaced strokes that disrupt the semantic structure of the generated sketch. To address this, semantic guidance is essential to ensure that each pixel aligns correctly with its corresponding structure.", + "bbox": [ + 511, + 609, + 893, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Text-based guidance alone is often insufficient for precise structural alignment, as textual descriptions may not accurately map to image details. To overcome this limitation, we incorporate contour information to guide semantic pixel alignment and supplement missing semantic strokes. However, we found that overly detailed contour information can disrupt high-level semantics (as edge detection identifies changes based on pixel gradients). For example, when drawing portraits, the eyes are often represented as solid dots rather than pixel-defined circles, necessitating text-based", + "bbox": [ + 511, + 742, + 893, + 888 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16549", + "bbox": [ + 480, + 926, + 517, + 938 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0c076db4a915ddbcc366be45d1e4ce2c54ee2cbab395a50ed83c0304139675c8.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison of sketch generation results with and without contour integration. The top row shows Ref. and Cnt. images, followed by results without contour integration (w/o Cont. Int.) and with contour integration (w/ Cont. Int.) as discussed in Sec. 3.4. Contour integration leads to better alignment of semantic features and reduces background interference." + ], + "image_footnote": [], + "bbox": [ + 107, + 102, + 478, + 243 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "semantic guidance for accurate sketch rendering.", + "bbox": [ + 101, + 338, + 416, + 352 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The text-based semantic loss is derived from the guidance provided in the image-to-image diffusion pipeline, which ensures high-level semantic alignment during sketch generation: $L_{sem} = \\lambda \\cdot \\mathrm{CLIP}(I^{ske}, T^{cnt})$ , where $T^{cnt}$ is the text prompt extracted from $I^{cnt}$ and $\\lambda$ is a weighting parameter for the text guidance.", + "bbox": [ + 99, + 353, + 483, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The contour-based guidance originates from the cached query features during the DDPM inversion process. To integrate this contour information effectively, we use the following equation to adjust the query:", + "bbox": [ + 99, + 441, + 483, + 500 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {i + 1} ^ {s k e} = \\gamma Q _ {i} ^ {\\text {c o n t}} + (1 - \\gamma) Q _ {i} ^ {s k e}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 505, + 483, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\gamma$ is a tunable parameter that controls the influence of contour information, set to 0.25 by default in our experiments. As illustrated in Fig. 7, contour integration significantly improves the alignment of semantic structures, ensuring that key details, such as object outlines, are preserved without introducing unnecessary background noise.", + "bbox": [ + 99, + 529, + 483, + 617 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "By combining high-level semantic text guidance with contour-based structural alignment, we ensure that the final generated sketch $I^{ske}$ maintains semantic completeness and accurately represents the image of the content $I^{cnt}$ . This collaborative approach allows for the preservation of both stroke attributes and semantic integrity in the sketch.", + "bbox": [ + 99, + 618, + 483, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Stroke Details Propagation Enhancement", + "text_level": 1, + "bbox": [ + 101, + 713, + 444, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Traditional self-attention blocks often focus on limited areas around image patches, while masked extended attention blocks distribute attention more uniformly by expanding receptive fields across the image. Although this broader approach captures larger context, it can introduce noise and blur finer details [1, 29]. To address this, we adopt a refined contrast operation, inspired by [1], to dynamically enhance high-variance regions and suppress low-contrast noise. This contrast adjustment is defined as:", + "bbox": [ + 99, + 734, + 483, + 866 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {E n h a n c e} (A) = (A - \\mu (A)) \\zeta (\\sigma (A)) + \\mu (A), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 875, + 483, + 890 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mu(A)$ and $\\sigma(A)$ represent the mean and standard deviation operations, respectively, and $\\zeta$ is an adaptive contrast operator. This technique effectively reduces noise, ensuring critical details are preserved during sketch generation.", + "bbox": [ + 509, + 102, + 893, + 160 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Building on this, we incorporate the stroke-based refinement concept from SDEdit [27]. Starting with a sketch initialized by stroke exchange, we denoise the image using classifier-free guidance (CFG) [6, 14]. During each denoising step, we utilize two parallel forward passes in the network. The first pass employs a cross-image attention layer to capture the stroke characteristics and abstraction level of the reference sketch, generating $z^{ske}$ : $\\epsilon_{stroke}^{\\times} = \\epsilon_{\\theta}^{\\times}(z_{t}^{ske})$ , while simultaneously retaining the semantic context extracted from the content image's descriptive prompt: $\\epsilon_{text}^{\\times} = \\epsilon_{\\theta}^{\\times}(z_{t}^{text})$ . The second pass applies regular self-attention to enhance the sketch's structural integrity: $\\epsilon^{self} = \\epsilon_{\\theta}^{self}(z_{t}^{ske})$ .", + "bbox": [ + 509, + 161, + 893, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following the CFG scheme, the predicted noise $\\epsilon^t$ is computed as:", + "bbox": [ + 511, + 337, + 893, + 364 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon^ {t} = \\epsilon^ {s e l f} + \\beta_ {s g} (\\epsilon_ {s t r o k e} ^ {\\times} - \\epsilon^ {s e l f}) + \\beta_ {t e x t} (\\epsilon_ {t e x t} ^ {\\times} - \\epsilon^ {s e l f}), (7)\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 372, + 893, + 392 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\beta_{sg}$ is the stroke guidance scale, and $\\beta_{text}$ is the weight for the content text context.", + "bbox": [ + 511, + 398, + 893, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 439, + 712, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. Our experiments utilize three datasets: FS2K [9], the Anime dataset [19], and our newly created Stroke2Sketch-dataset. The FS2K dataset includes 5,140 facial image-sketch pairs. We randomly selected 1,000 pairs for testing, using color images as content images and choosing one sketch from the test set as a reference style. A similar selection strategy was used for the Anime dataset [19].", + "bbox": [ + 509, + 464, + 893, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The Stroke2Sketch-dataset includes 50 content images from diverse categories and 20 distinct sketch styles, including single-line sketches, ink sketches, line art, and realistic sketches. Further details are provided in the Appendix.", + "bbox": [ + 509, + 566, + 893, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. While traditional metrics like ArtFID [43], LPIPS [53], and FID [13] are standard for style transfer, they struggle to capture the semantic sparsity and high-level abstraction unique to sketch generation. Consistent with prior work [38], we prioritize user perception to better reflect artistic and semantic quality in sketches.", + "bbox": [ + 511, + 625, + 893, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Qualitative Comparison", + "text_level": 1, + "bbox": [ + 511, + 720, + 728, + 736 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our proposed method through a comparison with eight state-of-the-art methods, including three training-based style transfer methods (Ref2sketch [2], Semi-ref2sketch [33], and CSGO [46]), and five training-free style transfer methods (IP-Adapter [49], InstantStyle [41], InstantStyle-plus [42], RB-Modulation [32], and StyleID [7]). Each of these methods takes a reference sketch as input.", + "bbox": [ + 509, + 742, + 893, + 844 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 8 illustrates the qualitative results across a variety of content images and reference sketches. In the first column, we display the content images, and the second column", + "bbox": [ + 509, + 845, + 893, + 888 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16550", + "bbox": [ + 480, + 926, + 517, + 938 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/42bf7bfd95b139362e9eb337143bd1d3b8463e0598966a3b9849649cc8fd2cd0.jpg", + "image_caption": [ + "Figure 8. Qualitative comparison of sketch generation across training-based (4th-6th columns) and training-free baselines (7th-11th columns) using various reference sketches. Our method (3rd column) demonstrates superior adaptability to different reference styles, maintaining both stroke fidelity and semantic consistency across a range of content and reference types." + ], + "image_footnote": [], + "bbox": [ + 107, + 99, + 890, + 338 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8799ed8266de01dfdbe80d9528ba2c4541a6962442c4c13ac664c848a9df8f3d.jpg", + "image_caption": [ + "Figure 9. (a) Content image and manual stroke stylization by CLIPascene [39] with Adobe Illustrator; (b) Our method's automatically generated sketches based on reference stroke styles; (c) Appearance transfer results using similar references." + ], + "image_footnote": [], + "bbox": [ + 106, + 405, + 478, + 577 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "shows the reference sketches. Notably, the fourth row's reference sketch is of a type seen in the Semi-ref2sketch training data, while the other three reference sketches represent styles outside the training data for Semi-ref2sketch. For methods requiring prompt input, we uniformly set the prompt to \"sketch photo,\" while other parameters follow each method's default configuration. Our approach demonstrates robustness across diverse reference styles, accurately preserving both the stroke style and high-level semantic details.", + "bbox": [ + 99, + 654, + 483, + 799 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with vector sketch generation: Fig. 9(a) compares our method with the vectorized sketch generation approach of CLIPascene [39]. The content image is shown alongside vectorized sketches produced by CLIPascene, which were further manually refined with brush strokes in Adobe Illustrator. While CLIPascene can generate abstract", + "bbox": [ + 99, + 801, + 483, + 888 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4da0f4726b06cf026022e03711cfa9049023b28b1318500d88bac7199d7b8b07.jpg", + "image_caption": [ + "Figure 10. Stroke2Sketch's color sketch generation preserving reference styles and stroke characteristics" + ], + "image_footnote": [], + "bbox": [ + 514, + 401, + 890, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sketches, it requires post-processing to achieve consistent stroke styling, whereas our method (Fig. 9(b)) automatically produces sketches with stylistically aligned strokes based on reference attributes.", + "bbox": [ + 511, + 525, + 893, + 582 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with appearance transfer methods: Fig. 9(c) shows results from appearance transfer method of CIA [1] applied to the same reference sketches. Although CIA excel in transferring visual features based on semantic similarity and object category, they focus on appearance rather than stroke style, making them less suitable for our sketch generation goals.", + "bbox": [ + 511, + 584, + 893, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Non-grayscale Sketch Generation. Building on its grayscale performance, Stroke2Sketch maintains reference stroke patterns and artistic styles in color outputs (Fig. 10).", + "bbox": [ + 511, + 686, + 893, + 732 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.Quantitative Comparison", + "text_level": 1, + "bbox": [ + 511, + 734, + 738, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We quantitatively evaluate our method against several state-of-the-art sketch extraction techniques, including both training-based (Ref2Sketch [2], Semi-Ref2Sketch [33], and Informative-drawing [5]) and training-free style transfer methods (IP-Adapter [49], InstantStyle [41], InstantStyle-plus [42], StyleID [7]). Table 1 presents the results on the Stroke2Sketch-dataset, showing that our method achieves the lowest ArtFID and FID scores, indicating superior performance in both style alignment and content preservation.", + "bbox": [ + 509, + 757, + 893, + 888 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16551", + "bbox": [ + 480, + 926, + 517, + 938 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1456d75b9b7e2080d0384a9fd303beec5b35e6f09ece7d8470cdbc28f5c235a7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MetricOursRef2sketchSemi-refInfor-drawingIP-AdapterInstantStyleInstantStyle+StyleID
ArtFID ↓32.45545.29233.24234.21433.45732.53237.65635.727
LPIPS ↓0.53150.69820.53060.60370.66340.54320.65320.5426
FID ↓22.43534.65024.35925.03524.06823.94026.63225.658
", + "bbox": [ + 138, + 99, + 854, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/61f46cafb46bca11340ace69c7c4271bde8205f5d8e3ccd0db892eb950d8be44.jpg", + "image_caption": [ + "Figure 11. User study results. Our proposed method received the highest preference." + ], + "image_footnote": [], + "bbox": [ + 112, + 208, + 473, + 351 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "These metrics highlight our method's effectiveness in producing high-quality sketches with strong semantic and stylistic fidelity.", + "bbox": [ + 101, + 402, + 483, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. User Study", + "text_level": 1, + "bbox": [ + 101, + 455, + 220, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We randomly selected 15 content images and 15 reference sketches from Stroke2Sketch-dataset, creating 225 image pairs. From these, we sampled 20 pairs to generate stylized images using four methods. The results were displayed side-by-side in random order, and participants were asked to choose their favorite based on three criteria: content extraction, stroke stylization, and overall preference. We collected 2,000 votes for each criterion from 100 users and presented the results as a bar chart. As shown in Fig. 11, our method received the highest preference, outperforming both training-based and training-free baselines, demonstrating its effectiveness in handling diverse stroke styles and abstract artistic effects.", + "bbox": [ + 99, + 477, + 482, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 101, + 676, + 250, + 693 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To validate our method's components, we performed an ablation study on the Stroke2Sketch-dataset. As shown in Table 2, removing any of the Directive Attention Module (DAM), Semantic Preservation Module (SPM), or Stroke Details Propagation Enhancement (SDPE) led to decreased performance across ArtFID, FID, and LPIPS metrics, confirming the contribution of each component. Specifically, without DAM, ArtFID increased from 32.45 to 38.67 and FID rose to 26.53, indicating weaker style-content alignment and content leakage in sketches, as observed in Fig. 12. The absence of SPM resulted in ArtFID rising to 36.89 and FID to 30.47, with sketches losing semantic coherence and structural integrity, such as poorly defined object outlines.", + "bbox": [ + 99, + 698, + 485, + 890 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5ea52e9eb4831258121016f52fef30ed57e197c0a34d56ae099c209181402324.jpg", + "table_caption": [ + "Table 1. Quantitative comparison on Stroke2Sketch-dataset with training-based (3rd-5th columns) and training-free baselines (6th-9th columns)." + ], + "table_footnote": [], + "table_body": "
ConfigurationArtFIDFIDLPIPS
A: Ours32.4522.430.530
B: - DAM (Sec. 3.3)38.6726.530.672
C: - SPM (Sec. 3.4)36.8930.470.637
D: - SDPE (Sec. 3.5)40.5332.440.598
", + "bbox": [ + 547, + 200, + 859, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Ablation study of different variants of our method.", + "bbox": [ + 531, + 282, + 872, + 296 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c46d58774955ce8de6a0f6ce3a4421bc7f9fc8feb70d07603acda03c6ca15335.jpg", + "image_caption": [ + "Figure 12. Ablative qualitative comparison of different variants of our method." + ], + "image_footnote": [], + "bbox": [ + 521, + 306, + 887, + 429 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Most notably, removing SDPE caused the sharpest decline, with ArtFID reaching 40.53 and sketches exhibiting excessive noise and loss of fine details, as evidenced by cluttered textures in Fig. 12. In contrast, the full method (Configuration A) achieved optimal scores of 32.45 (ArtFID), 22.43 (FID), and 0.530 (LPIPS), producing high-quality, reference-aligned sketches with precise details. These results highlight the critical roles of DAM, SPM, and SDPE in enhancing sketch quality.", + "bbox": [ + 509, + 483, + 895, + 616 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 628, + 627, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our training-free method for content-to-sketch generation aligns stroke attributes and semantic texture sparsity, mimicking artistic subject extraction to produce high-fidelity sketches. Leveraging pretrained models, it achieves state-of-the-art performance in quantitative metrics and user evaluations. Limitations arise with overly simplistic or complex reference sketches (see appendix for failure cases). Future work could explore decoupling semantic information from stroke attributes to improve adaptability and sketch quality. Acknowledgment. This work is supported by the Guangdong Natural Science Funds for Distinguished Young Scholars (Grant 2023B1515020097), the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No.: AISG3-GV-2023-011), the Singapore Ministry of Education AcRF Tier 1 Grant (Grant No.: MSS25C004), and the Lee Kong Chian Fellowships.", + "bbox": [ + 509, + 654, + 895, + 890 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16552", + "bbox": [ + 480, + 926, + 517, + 938 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 102, + 101, + 194, + 116 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer. In ACM SIGGRAPH 2024 Conference Papers, pages 1–12, 2024. 3, 6, 7", + "[2] Amirsaman Ashtari, Chang Wook Seo, Cholmin Kang, Sihun Cha, and Junyong Noh. Reference based sketch extraction via attention mechanism. ACM Trans. Graph., 41(6), 2022. 1, 3, 6, 7", + "[3] John Canny. A computational approach to edge detection. IEEE Transactions on pattern analysis and machine intelligence, (6):679-698, 1986. 3", + "[4] Nan Cao, Xin Yan, Yang Shi, and Chaoran Chen. Ai-sketcher: a deep generative model for producing high-quality sketches. In Proceedings of the AAAI conference on artificial intelligence, pages 2564-2571, 2019. 3", + "[5] Caroline Chan, Frédo Durand, and Phillip Isola. Learning to generate line drawings that convey geometry and semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7915-7925, 2022. 1, 3, 7", + "[6] Hyungjin Chung, Jeongsol Kim, Geon Yeong Park, Hyelin Nam, and Jong Chul Ye. Cfg++: Manifold-constrained classifier free guidance for diffusion models. 2024. 6", + "[7] Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8795-8805, 2024. 6, 7", + "[8] Ayan Das, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Beziersketch: A generative model for scalable vector sketches. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI 16, pages 632-647. Springer, 2020. 3", + "[9] Deng-Ping Fan, Ziling Huang, Peng Zheng, Hong Liu, Xuebin Qin, and Luc Van Gool. Facial-sketch synthesis: a new challenge. Machine Intelligence Research, 19(4):257-287, 2022. 6", + "[10] Kevin Frans, Lisa Soros, and Olaf Witkowski. Clipdraw: Exploring text-to-drawing synthesis through language-image encoders. Advances in Neural Information Processing Systems, 35:5207-5218, 2022. 3", + "[11] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3", + "[12] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4775–4785, 2024. 3", + "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + ], + "bbox": [ + 102, + 125, + 485, + 887 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Jonathan Ho. Classifier-free diffusion guidance. ArXiv, abs/2207.12598, 2022. 6", + "[15] Jjin Hu, Ke Li, Yonggang Qi, and Yi-Zhe Song. Scale-adaptive diffusion model for complex sketch synthesis. In The Twelfth International Conference on Learning Representations, 2024. 3", + "[16] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly DDPM noise space: Inversion and manipulations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12469-12478, 2024. 3", + "[17] Jaeseok Jeong, Junho Kim, Yunjoy Choi, Gayoung Lee, and Youngjung Uh. Visual style prompting with swapping self-attention. arXiv preprint arXiv:2402.12974, 2024. 3", + "[18] Yutao Jiang, Yang Zhou, Yuan Liang, Wenxi Liu, Jianbo Jiao, Yuhui Quan, and Shengfeng He. Diffuse3d: Wide-angle 3d photography via bilateral diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8998-9008, 2023. 1", + "[19] Tae Bum Kang. Anime sketch colorization pair, 2018. Accessed: 2024-05-18. 6", + "[20] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 3, 5", + "[21] Xinzhe Li, Jiahui Zhan, Shengfeng He, Yangyang Xu, Junyu Dong, Huaidong Zhang, and Yong Du. Personamagic: Stage-regulated high-fidelity face customization with tandem equilibrium. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4995-5003, 2025. 1", + "[22] Chenxi Liu, Enrique Rosales, and Alla Sheffer. Strokeaggregator: Consolidating raw sketches into artist-intended curve drawings. ACM Transactions on Graphics (TOG), 37(4):1-15, 2018. 3", + "[23] Chenxi Liu, Toshiki Aoki, Mikhail Bessmeltsev, and Alla Sheffer. Stripmaker: Perception-driven learned vector sketch consolidation. ACM Transactions on Graphics (TOG), 42(4): 1-15, 2023. 3", + "[24] Difan Liu, Matthew Fisher, Aaron Hertzmann, and Evangelos Kalogerakis. Neural strokes: Stylized line drawing of 3d shapes. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 14184-14193. IEEE, 2021. 1", + "[25] Haofeng Liu, Chenshu Xu, Yifei Yang, Lihua Zeng, and Shengfeng He. Drag your noise: Interactive point-based editing via diffusion semantic propagation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6743-6752, 2024. 1", + "[26] Xiao-Chang Liu, Yu-Chen Wu, and Peter Hall. Painterly style transfer with learned brush strokes. IEEE Transactions on Visualization and Computer Graphics, 30(9):6309-6320, 2024. 1", + "[27] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 6" + ], + "bbox": [ + 514, + 103, + 895, + 887 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16553", + "bbox": [ + 480, + 928, + 517, + 938 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Reiichiro Nakano. Neural painters: A learned differentiable constraint for generating brushstroke paintings. arXiv preprint arXiv:1904.08410, 2019. 1", + "[29] Nadav Orzech, Yotam Nitzan, Ulysse Mizrahi, Dov Danon, and Amit H Bermano. Masked extended attention for zero-shot virtual try-on in the wild. arXiv preprint arXiv:2406.15331, 2024. 6", + "[30] Xavier Soria Poma, Edgar Riba, and Angel Sappa. Dense extreme inception network: Towards a robust cnn model for edge detection. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1923-1932, 2020. 3", + "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3", + "[32] L Rout, Y Chen, N Ruiz, A Kumar, C Caramanis, S Shakkottai, and W Chu. Rb-modulation: Training-free personalization of diffusion models using stochastic optimal control. 2024. 2, 3, 6", + "[33] Chang Wook Seo, Amirsaman Ashtari, and Junyong Noh. Semi-supervised reference-based sketch extraction using a contrastive learning framework. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 1, 3, 6, 7", + "[34] Edgar Simo-Serra, Satoshi Iizuka, Kazuma Sasaki, and Hiroshi Ishikawa. Learning to simplify: fully convolutional networks for rough sketch cleanup. ACM Transactions on Graphics (TOG), 35(4), 2016. 1", + "[35] Gowthami Somepalli, Anubhav Gupta, Kamal Gupta, Shramay Palta, Micah Goldblum, Jonas Geiping, Abhinav Shrivastava, and Tom Goldstein. Measuring style similarity in diffusion models. arXiv preprint arXiv:2404.01292, 2024. 3", + "[36] Xavier Soria, Yachuan Li, Mohammad Rouhani, and Angel D Sappa. Tiny and efficient model for the edge detection generalization. In CVPR, pages 1364-1373, 2023. 3, 4", + "[37] Zhuo Su, Wenzhe Liu, Zitong Yu, Dewen Hu, Qing Liao, Qi Tian, Matti Pietikainen, and Li Liu. Pixel difference networks for efficient edge detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5117-5127, 2021. 3", + "[38] Yael Vinker, Ehsan Pajouheshgar, Jessica Y Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, and Ariel Shamir. Clipasso: Semantically-aware object sketching. ACM Transactions on Graphics (TOG), 41(4): 1-11, 2022. 1, 3, 6", + "[39] Yael Vinker, Yuval Alaluf, Daniel Cohen-Or, and Ariel Shamir. Clipscene: Scene sketching with different types and levels of abstraction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4146–4156, 2023. 7", + "[40] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. p+: Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522, 2023. 3", + "[41] Haofan Wang, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv preprint arXiv:2404.02733, 2024. 1, 2, 3, 6, 7" + ], + "bbox": [ + 102, + 103, + 483, + 888 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Haofan Wang, Peng Xing, Renyuan Huang, Hao Ai, Qixun Wang, and Xu Bai. Instantstyle-plus: Style transfer with content-preserving in text-to-image generation. arXiv preprint arXiv:2407.00788, 2024. 3, 6, 7", + "[43] Matthias Wright and Björn Ommer. Artfid: Quantitative evaluation of neural style transfer. In DAGM German Conference on Pattern Recognition, pages 560-576. Springer, 2022. 6", + "[44] Zongwei Wu, Liangyu Chai, Nanxuan Zhao, Bailin Deng, Yongtuo Liu, Qiang Wen, Junle Wang, and Shengfeng He. Make your own sprites: Aliasing-aware and cell-controllable pixelization. ACM Transactions on Graphics (TOG), 41(6): 1-16, 2022. 1", + "[45] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In Proceedings of the IEEE international conference on computer vision, pages 1395-1403, 2015. 3", + "[46] Peng Xing, Haofan Wang, Yanpeng Sun, Qixun Wang, Xu Bai, Hao Ai, Renyuan Huang, and Zechao Li. Csgo: Content-style composition in text-to-image generation. arXiv preprint arXiv:2408.16766, 2024. 6", + "[47] Chenshu Xu, Yangyang Xu, Huaidong Zhang, Xuemiao Xu, and Shengfeng He. Dreamanine: Learning style-identity textual disentanglement for anime and beyond. IEEE Transactions on Visualization and Computer Graphics, 2024. 1", + "[48] Rui Yang, Xiaojun Wu, and Shengfeng He. Mixsa: Training-free reference-based sketch extraction via mixture-of-self-attention. IEEE Transactions on Visualization and Computer Graphics, 2024. 3", + "[49] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 6, 7", + "[50] Yuyang Yu, Bangzhen Liu, Chenxi Zheng, Xuemiao Xu, Huaidong Zhang, and Shengfeng He. Beyond textual constraints: Learning novel diffusion conditions with fewer examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7109-7118, 2024. 1", + "[51] Sicong Zang, Shikui Tu, and Lei Xu. Self-organizing a latent hierarchy of sketch patterns for controllable sketch synthesis. IEEE Transactions on Neural Networks and Learning Systems, 2023. 3", + "[52] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3", + "[53] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6", + "[54] Caixia Zhou, Yaping Huang, Mengyang Pu, Qingji Guan, Ruoxi Deng, and Haibin Ling. Muge: Multiple granularity edge detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25952-25962, 2024. 3" + ], + "bbox": [ + 514, + 103, + 893, + 856 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16554", + "bbox": [ + 480, + 928, + 517, + 938 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Stroke2Sketch_ Harnessing Stroke Attributes for Training-Free Sketch Generation/88f15fb1-f5ab-4157-a23c-15d3cb3bc8f3_model.json b/2025/Stroke2Sketch_ Harnessing Stroke Attributes for Training-Free Sketch Generation/88f15fb1-f5ab-4157-a23c-15d3cb3bc8f3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ba231f69ed3c4ba00401092a426db5382b545cb9 --- /dev/null +++ b/2025/Stroke2Sketch_ Harnessing Stroke Attributes for Training-Free Sketch Generation/88f15fb1-f5ab-4157-a23c-15d3cb3bc8f3_model.json @@ -0,0 +1,2266 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.117, + 0.017, + 0.191, + 0.056 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.246, + 0.015, + 0.8, + 0.058 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.141, + 0.895, + 0.162 + ], + "angle": 0, + "content": "Stroke2Sketch: Harnessing Stroke Attributes for Training-Free Sketch Generation" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.189, + 0.78, + 0.208 + ], + "angle": 0, + "content": "Rui Yang\\(^{1,3}\\), Huining Li\\(^{4,5}\\), Yiyi Long\\(^{2,5}\\), Xiaojun Wu\\(^{3,*}\\), Shengfeng He\\(^{5,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.143, + 0.207, + 0.856, + 0.243 + ], + "angle": 0, + "content": "\\(^{1}\\)Huaqiao University \\(^{2}\\)South China University of Technology \\(^{3}\\)Shaanxi Normal University \\(^{4}\\)Beijing University of Aeronautics and Astronautics \\(^{5}\\)Singapore Management University" + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.251, + 0.871, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.486, + 0.897, + 0.527 + ], + "angle": 0, + "content": "Figure 1. We propose Stroke2Sketch, a framework that accurately transfers stroke attributes from a reference sketch to a content image while preserving structure and style fidelity. The top row shows reference sketches, the leftmost column displays content images, and the central and right columns illustrate our method's precise content preservation and expressive stroke transfer." + }, + { + "type": "title", + "bbox": [ + 0.256, + 0.537, + 0.331, + 0.551 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.572, + 0.486, + 0.836 + ], + "angle": 0, + "content": "Generating sketches guided by reference styles requires precise transfer of stroke attributes, such as line thickness, deformation, and texture sparsity, while preserving semantic structure and content fidelity. To this end, we propose Stroke2Sketch, a novel training-free framework that introduces cross-image stroke attention, a mechanism embedded within self-attention layers to establish fine-grained semantic correspondences and enable accurate stroke attribute transfer. This allows our method to adaptively integrate reference stroke characteristics into content images while maintaining structural integrity. Additionally, we develop adaptive contrast enhancement and semantic-focused attention to reinforce content preservation and foreground emphasis. Stroke2Sketch effectively synthesizes stylistically faithful sketches that closely resemble handcrafted results, outperforming existing methods in expressive stroke control and semantic coherence. Codes are available at https://github.com/rane7/Stroke2Sketch." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.537, + 0.641, + 0.551 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.564, + 0.897, + 0.754 + ], + "angle": 0, + "content": "Generating stylized sketches from content images using reference stroke patterns presents a key challenge at the intersection of artistic rendering and semantic-aware synthesis. Traditional sketch algorithms [28, 34] generate procedural line drawings, while vector-based methods [24, 26, 38] produce clean parametric strokes. However, these methods lack the adaptability to transfer diverse artistic styles from exemplar sketches due to their rigid, data-agnostic architectures. Unlike human artists, who strategically vary stroke attributes such as thickness, curvature, and density to emphasize key semantic features while maintaining content fidelity, existing approaches struggle to capture this nuanced interplay between stroke semantics and structure." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.758, + 0.897, + 0.89 + ], + "angle": 0, + "content": "Recent learning-based methods [2, 5, 33] attempt to address this limitation by training on clustered sketch styles, yet as shown in Fig. 2(a-b), they fail to generalize to unseen stroke patterns due to catastrophic forgetting. Meanwhile, diffusion-based stylization techniques [21, 41, 47, 49, 50] excel in texture transfer but struggle with structural integrity due to content leakage in cross-attention layers (Fig. 2(c-d)). While new conditioning mechanisms [18, 25, 44] attempt to enhance structural control, they often introduce style dis" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.865, + 0.483, + 0.89 + ], + "angle": 0, + "content": "*Corresponding authors: Xiaojun Wu (xjwu@snnu.edu.cn) and Shengfeng He (shengfenghe@smu.edu.sg)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.518, + 0.939 + ], + "angle": 0, + "content": "16545" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.1, + 0.895, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.342, + 0.897, + 0.397 + ], + "angle": 0, + "content": "Figure 2. (a) and (b) show results from training-based methods (Ref2Sketch and Semi-Ref2Sketch), which struggle with unseen styles, leading to poor content alignment. (c) and (d) illustrate IP-Adapter and InstantStyle results, retaining style but lacking content alignment. (e) to (g) show ControlNet-based methods, preserving content but failing to match reference styles. (h) shows our method's output, achieving superior content preservation and style alignment without using ControlNet. Detailed prompts are shown in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.408, + 0.484, + 0.497 + ], + "angle": 0, + "content": "tortion or require dense user inputs. Hybrid approaches like ControlNet [52] enforce structural priors but sacrifice stylistic flexibility, leading to overly rigid outputs (Fig. 2(e)). Progressive stroke-based methods [32] introduce semantic incoherence by applying uniform strokes across the image (Fig. 2(g))." + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.5, + 0.484, + 0.706 + ], + "angle": 0, + "content": "We identify three fundamental challenges in reference-based sketch synthesis:(i) Semantic-aware stroke transfer. Effective style adaptation requires precise mapping of reference stroke attributes (e.g., tapered lines, cross-hatching) to semantically relevant content regions. (ii) Foreground prioritization. Artists naturally emphasize foreground elements using varied stroke density and complexity while simplifying backgrounds, yet existing methods apply uniform stylization, disrupting this compositional balance[32]. (iii) Content-style equilibrium. Since sketches encode content through linework, balancing structural preservation with style transfer is critical. Techniques such as CLIP-space style subtraction [41] fail to maintain this balance, as even minor content leakage distorts key edges (Fig. 2(c-d))." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.71, + 0.484, + 0.871 + ], + "angle": 0, + "content": "To address these challenges, we propose Stroke2Sketch, a framework that enables precise stroke attribute transfer while maintaining content fidelity. Our key insight is that stroke properties, like line thickness, curvature, and texture sparsity, are inherently encoded within the self-attention and cross-attention relationships of pretrained diffusion models. By dynamically aligning these attention patterns between content and reference features, we achieve effective style transfer without structural degradation. Stroke2Sketch integrates three novel components tailored to each identified challenge:" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.875, + 0.483, + 0.89 + ], + "angle": 0, + "content": "(1) Cross-image stroke attention tackles the challenge of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.408, + 0.896, + 0.512 + ], + "angle": 0, + "content": "semantic-aware stroke transfer by facilitating stroke attribute exchange through key-value swapping in diffusion layers. Instead of directly blending features [41], we leverage attention blocks to transplant stroke characteristics onto content structures, preventing the entanglement of style and geometry. This approach ensures accurate stroke mapping without disrupting semantic coherence, as shown in Fig. 2(h)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.515, + 0.896, + 0.633 + ], + "angle": 0, + "content": "(2) Directive Attention Module ensures that stroke transfer remains compositionally balanced. Background textures often introduce conflicting patterns that dilute the intended style. We mitigate this by clustering self-attention maps to isolate foreground objects, then restricting cross-image attention to these regions. This mimics how artists prioritize focal elements while simplifying less critical areas, enhancing both style consistency and perceptual quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.636, + 0.896, + 0.768 + ], + "angle": 0, + "content": "(3) Semantic Preservation Module addresses content-style equilibrium by injecting content contours as positional queries during early denoising. This hybridizes the precision of edge detectors with the flexibility of text-driven generation, allowing structural guidance without rigid constraints. Unlike ControlNet, which enforces strict boundaries, our approach treats edges as \"soft constraints\" that evolve into stylized strokes, preserving both structure and artistic abstraction." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.515, + 0.896, + 0.768 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.772, + 0.896, + 0.89 + ], + "angle": 0, + "content": "As validated in Fig. 2(h) and Fig. 6, Stroke2Sketch achieves state-of-the-art performance across diverse sketch styles. Experimental results show that our method outperforms adapter-based and ControlNet methods in both style alignment (87% user preference) and content preservation (92% accuracy in line correspondence tests). Importantly, it achieves these results without dataset-specific training or architectural modifications, demonstrating that pretrained" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.519, + 0.939 + ], + "angle": 0, + "content": "16546" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.101, + 0.103, + 0.486, + 0.133 + ], + "angle": 0, + "content": "diffusion models can master sketch stylization when guided by principled feature interactions." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.147, + 0.242, + 0.163 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.172, + 0.313, + 0.188 + ], + "angle": 0, + "content": "2.1. Photo-Sketch Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.194, + 0.486, + 0.339 + ], + "angle": 0, + "content": "Generating a sketch from a content image based on a reference style parallels edge detection, as both tasks emphasize prominent visual features. Edge detection methods [3, 30, 36, 37, 45, 54], which detect sharp changes in color or brightness, form the foundation of sketch extraction but are limited to a single style and often produce artifacts, like scattered dots or broken lines. High-quality sketch generation requires more than edge detection; it demands line style, texture sparsity, and semantic abstraction to achieve artist-level results." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.341, + 0.485, + 0.459 + ], + "angle": 0, + "content": "Learning-based approaches have improved sketch realism by enhancing boundary detection and rendering distinct line styles. For example, Chan et al. [5] incorporated depth and semantic information for better sketch quality, while Ref2Sketch [2] and Semi-Ref2Sketch [33] leverage paired and semi-supervised training for stylized sketch extraction. However, these methods rely on large sketch datasets, which are challenging to obtain, limiting model robustness." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.46, + 0.485, + 0.592 + ], + "angle": 0, + "content": "Another research area, stroke-based rendering, focuses on manipulating strokes and contours for sketching. Methods like CLIPDraw [10] and CLIPasso [38] employ Bezier curves to create abstract sketches with high-level semantic simplification. StrokeAggregator [22] and StripMaker [23] refine vector sketches, achieving quality comparable to artist drawings. Other methods include semantic concept-to-sketch methods [4, 8, 15, 51]. However, these methods assume uniform style, while real-world sketches vary widely." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.592, + 0.485, + 0.637 + ], + "angle": 0, + "content": "Building on these insights, our method combines stroke attributes, semantic abstraction, and expression to better align generated sketches with diverse reference styles." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.647, + 0.36, + 0.663 + ], + "angle": 0, + "content": "2.2. High-Semantic Style Transfer" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.669, + 0.486, + 0.786 + ], + "angle": 0, + "content": "Generating sketches that adhere to both content and reference style is a specialized style transfer task. Advances in diffusion models have propelled style transfer through self- and cross-attention mechanisms, which preserve spatial layout and stylized content. Techniques such as Prompt-to-Prompt [11] and \\(\\mathrm{P + }\\) [40] show how attention mechanisms can maintain structural coherence while enabling flexible style and semantic control." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.787, + 0.486, + 0.89 + ], + "angle": 0, + "content": "Cross-Image Attention (CIA) [1] and methods like Swapping Self-Attention [17] and StyleAligned [12] reveal that style features are best retained during upsampling stages, even though content leakage may occur in bottleneck phases. IP-Adapter achieves style transfer via dual cross-attention with text and image, though it can weaken text control and lead to leakage." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.103, + 0.896, + 0.25 + ], + "angle": 0, + "content": "InstantStyle [41] addresses leakage by subtracting features in the same feature space but often requires external constraints like ControlNet [52] for image-to-image generation, which can dilute style fidelity. Further developments like InstantStyle-plus [42] and RB-Modulation [32] incorporate CSD [35] to better align styles during generation, but limitations remain in sketch generation. Prompt-to-style transfer, however, demonstrates pretrained models' strong semantic alignment capabilities, providing valuable priors for consistent, high-level sketch synthesis." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.262, + 0.655, + 0.277 + ], + "angle": 0, + "content": "3. Stroke2Sketch" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.285, + 0.897, + 0.509 + ], + "angle": 0, + "content": "Given a content image \\( I^{cnt} \\in \\mathbb{R}^{H^{cnt} \\times W^{cnt} \\times 3} \\) and a reference sketch image \\( I^{ref} \\in \\mathbb{R}^{H^{ref} \\times W^{ref} \\times 3} \\), our task is to generate a sketch \\( I^{ske} \\) that aligns with the content structure of \\( I^{cnt} \\) while adhering to the stroke style, texture sparsity, and high-level semantic abstraction of \\( I^{ref} \\). Additionally, we aim to remove or retain background elements as needed to enhance the foreground object. In this work, we address the sketch generation task using a controllable text-to-image guidance approach. Specifically, we extract object prompts from the content image \\( I^{cnt} \\) using BLIP [20] and incorporate the stroke style and high-level abstraction from the reference sketch \\( I^{ref} \\) to re-render the final target sketch \\( I^{ske} \\). To achieve this, we leverage a pre-trained text-to-image model. The overall network architecture is illustrated in Fig 3, with detailed explanations of each module provided below." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.516, + 0.65, + 0.53 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.896, + 0.758 + ], + "angle": 0, + "content": "DDPM inversion [16] is a process in diffusion-based generative models that reverses denoising steps, enabling the reconstruction of latent representations from generated outputs. Compared to DDIM inversion method [48], DDPM inversion offers greater flexibility for editing tasks by producing noise maps that, while correlated across timesteps and not normally distributed, allow precise image reconstruction and meaningful edits such as color adjustments and structural shifts. A key advantage of DDPM inversion is its ability to maintain an image's structure while altering the conditioning input, such as a text prompt, to enable artifact-free edits that adapt semantics while preserving original details. This efficient method bypasses optimization processes and can enhance diffusion-based editing techniques by improving image fidelity and supporting diverse outputs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.758, + 0.897, + 0.89 + ], + "angle": 0, + "content": "Stable Diffusion [31] incorporates this inversion process within a latent space, rather than a pixel space, increasing computational efficiency and expressiveness. The input image is encoded through a pretrained variational autoencoder (VAE) to produce a latent code \\( z \\). Denoising then occurs within this latent space using a U-Net architecture that incorporates self-attention and cross-attention mechanisms. Self-attention blocks enhance image detail by calculating attention scores between projected query \\( Q \\), key \\( K \\), and value" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.519, + 0.939 + ], + "angle": 0, + "content": "16547" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.104, + 0.891, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.435, + 0.897, + 0.531 + ], + "angle": 0, + "content": "Figure 3. The Stroke2Sketch architecture. The content image \\( I^{\\mathrm{cnt}} \\) undergoes contour detection, generating feature representations \\( z^{\\mathrm{cont}} \\) and \\( z^{\\mathrm{cnt}} \\), while the reference sketch \\( I^{\\mathrm{ref}} \\) is inverted to produce \\( z^{\\mathrm{ref}} \\). The Directive Attention Module (DAM) aligns high-level semantic features between the content and reference features by emphasizing cross-image semantic correspondences \\( A_{n} \\). Self-attention maps \\( F_{SA} \\) are aggregated and clustered to produce segmentation masks \\( M_{i} \\), which help distinguish foreground from background regions. Cross-image Stroke Attention (CSA) transfers stroke attributes, and the Semantic Preservation Module (SPM) enforces semantic alignment and stroke fidelity in the generated sketch \\( \\hat{I}^{\\mathrm{ske}} \\) via loss \\( \\mathcal{L}_{\\mathrm{sem}} \\) and contour-based structural integration. Lastly, Stroke Detail Propagation Enhancement (SDPE) refines details, resulting in the final output sketch \\( I^{\\mathrm{ske}} \\). SD represents the pre-trained diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.556, + 0.174, + 0.569 + ], + "angle": 0, + "content": "\\(V\\) vectors:" + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.579, + 0.484, + 0.613 + ], + "angle": 0, + "content": "\\[\nA = \\operatorname {s o f t m a x} \\left(\\frac {Q \\cdot K ^ {\\top}}{\\sqrt {d}}\\right), \\quad \\phi = A \\cdot V, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.622, + 0.484, + 0.68 + ], + "angle": 0, + "content": "where \\( A \\) is the attention map, \\( d \\) is the dimensionality, and \\( \\phi \\) is the output of the self-attention block. Cross-attention blocks incorporate conditioning inputs (e.g., text prompts) to guide the generation process." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.681, + 0.484, + 0.725 + ], + "angle": 0, + "content": "Finally, the refined latent representation \\(z\\) is decoded back into an RGB image using the VAE decoder, yielding a high-quality output guided by the conditioning input." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.734, + 0.36, + 0.75 + ], + "angle": 0, + "content": "3.2. Cross-image Stroke Attention" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.756, + 0.486, + 0.892 + ], + "angle": 0, + "content": "As discussed in the Introduction, sketch generation as a specialized form of style transfer requires attention to local stroke attributes and consistent texture abstraction across different levels of semantic sparsity. Edge detection serves as the foundation for this task, with edges defining content contours as the most effective strategy for preserving structural information. We extract the contour \\( I^{cont} \\) from the content image using TEED [36] and obtain the inverted latents \\( z_{T}^{cnt} \\), \\( z_{T}^{ref} \\), and \\( z_{T}^{cont} \\) for the content image, reference sketch, and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.556, + 0.895, + 0.615 + ], + "angle": 0, + "content": "contour image using DDPM inversion. During the denoising process, the latent \\( z_{T}^{cnt} \\) from the content image serves as the initial noise \\( z_{T}^{ske} \\) for sketch denoising. At each timestep \\( t \\), we apply Equation 1." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.615, + 0.896, + 0.748 + ], + "angle": 0, + "content": "To achieve effective stroke feature transfer, we utilize latent representations for the content, reference, and contour images obtained through DDPM inversion. For predefined timesteps \\(t \\in \\{0, \\dots, T\\}\\), the reference sketch \\(I^{ref}\\), content image \\(I^{cnt}\\), and contour image \\(I^{cont}\\) are inverted from their initial state (\\(t = 0\\)) to Gaussian noise (\\(t = T\\)). During DDPM inversion, we also collect the query features of the content (\\(Q_{t}^{cnt}\\)), and the key and value features of the reference sketch (\\(K_{t}^{ref}\\), \\(V_{t}^{ref}\\)) at each timestep." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.749, + 0.897, + 0.837 + ], + "angle": 0, + "content": "We initialize the latent noise \\( z_{T}^{ske} \\) for the stylized sketch by copying the content latent noise \\( z_{T}^{cnt} \\). To blend features from all three images, we combine the reference keys and values \\( K_{t}^{ref} \\) and \\( \\tilde{V}_{t}^{ref} \\) with the content features. This integration is achieved by mixing the keys and values using the following formulation:" + }, + { + "type": "equation", + "bbox": [ + 0.63, + 0.849, + 0.895, + 0.866 + ], + "angle": 0, + "content": "\\[\nK _ {t} ^ {\\mathrm {s k e}} = K _ {t} ^ {\\mathrm {r e f}} + \\alpha K _ {t} ^ {\\mathrm {c n t}}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.632, + 0.868, + 0.895, + 0.886 + ], + "angle": 0, + "content": "\\[\nV _ {t} ^ {\\mathrm {s k e}} = V _ {t} ^ {\\mathrm {r e f}} + \\alpha V _ {t} ^ {\\mathrm {c n t}}, \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.519, + 0.939 + ], + "angle": 0, + "content": "16548" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.104, + 0.232, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.164, + 0.231, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.134, + 0.223, + 0.21, + 0.232 + ], + "angle": 0, + "content": "Content Strokes" + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.104, + 0.354, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.169, + 0.349, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.223, + 0.334, + 0.233 + ], + "angle": 0, + "content": "Reference Strokes" + }, + { + "type": "image", + "bbox": [ + 0.355, + 0.104, + 0.475, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.164, + 0.471, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.377, + 0.223, + 0.453, + 0.234 + ], + "angle": 0, + "content": "Aligned Strokes" + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.244, + 0.483, + 0.284 + ], + "angle": 0, + "content": "Figure 4. Stroke alignment results with CSA. The feature images are shown in the first row. The second row shows results where feature alignment between the key and value exchange is applied." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.298, + 0.484, + 0.34 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a scalar parameter that controls the mixing ratio of content and reference features, allowing flexible adaptation to different stroke characteristics in the sketch." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.342, + 0.484, + 0.458 + ], + "angle": 0, + "content": "By exchanging and blending features from the content, reference, and contour images, this approach enables the effective transfer of stroke attributes and semantic alignment. However, we found that some strokes may fail to represent the intended content curves accurately (see Fig. 4). To improve output quality and highlight the foreground, we introduce additional mechanisms to guide sketch generation, as detailed below." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.471, + 0.342, + 0.485 + ], + "angle": 0, + "content": "3.3. Directive Attention Module" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.493, + 0.484, + 0.608 + ], + "angle": 0, + "content": "To concentrate stylization on perceptually significant regions, the DAM enhances foreground focus during stroke transfer. We extract self-attention feature maps \\( F_{SA} \\) at \\( 32 \\times 32 \\) resolution, aggregating them across channels via averaging. These maps are segmented into clusters \\( M_{j} \\) using KMeans clustering. For each cluster \\( j \\) and noun \\( n \\) extracted from the content image's caption (via BLIP [20]), we calculate a relevance score:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.619, + 0.484, + 0.657 + ], + "angle": 0, + "content": "\\[\nr (j, n) = \\frac {\\sum_ {(x , y)} M _ {j} (x , y) \\cdot A _ {n} (x , y)}{\\sum_ {(x , y)} M _ {j} (x , y) + \\delta}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.668, + 0.483, + 0.741 + ], + "angle": 0, + "content": "where \\(A_{n}\\) is the cross-attention map for noun \\(n\\), and \\(\\delta = 10^{-5}\\) stabilizes the computation. Clusters with \\(r(j,n) > 0.35\\) are designated as foreground regions; remaining areas are suppressed, directing stroke stylization to salient elements (Fig. 5)." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.742, + 0.483, + 0.859 + ], + "angle": 0, + "content": "By integrating the segmented self-attention mechanism with cross-image attention, DAM allows precise control over foreground regions, ensuring that stylistic features, such as line styles and high-level semantic abstractions, are faithfully transferred from the reference sketch to the content image. This approach achieves high fidelity in style and content alignment, yielding sketches that closely reflect the intended reference style with minimized background interference." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.861, + 0.484, + 0.89 + ], + "angle": 0, + "content": "As illustrated in Fig. 6, our method successfully aligns high-level semantic features and detailed stroke attributes," + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.104, + 0.613, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.104, + 0.705, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.104, + 0.891, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.188, + 0.895, + 0.269 + ], + "angle": 0, + "content": "Figure 5. Examples of DAM in action. The first column shows the content images, the second column displays segmentation maps obtained by clustering self-attention maps, the third column provides the reference sketches, and the fourth column illustrates the sketches generated by DAM after applying stroke attribute transfer and segmentation." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.275, + 0.639, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.275, + 0.716, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.275, + 0.894, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.386, + 0.895, + 0.493 + ], + "angle": 0, + "content": "Figure 6. Comparison of sketches generated using DAM. The top row shows Cnt. and Ref. images with stroke styles reflecting specific high-level attributes such as hair, eyebrows, and abstracted clothing texture. The second to fourth columns display sketches generated to match the content images in the top row, each adopting the stroke styles and high-level features from the reference. The bottom row highlights zoomed-in areas to emphasize the transferred stroke attributes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.504, + 0.895, + 0.576 + ], + "angle": 0, + "content": "producing coherent sketches that maintain stylistic and structural consistency with the reference. This approach minimizes background interference while ensuring that the final sketch reflects the nuanced stroke characteristics of the reference." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.588, + 0.779, + 0.603 + ], + "angle": 0, + "content": "3.4. Semantic Preservation Module" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.894, + 0.741 + ], + "angle": 0, + "content": "Although the injection of keys and values during sketch generation can effectively transfer stroke attributes to \\( I^{ske} \\), semantic inconsistencies may arise, particularly when the reference sketch \\( I^{ref} \\) does not align semantically with the content image \\( I^{cnt} \\). This can lead to noise and misplaced strokes that disrupt the semantic structure of the generated sketch. To address this, semantic guidance is essential to ensure that each pixel aligns correctly with its corresponding structure." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.743, + 0.895, + 0.889 + ], + "angle": 0, + "content": "Text-based guidance alone is often insufficient for precise structural alignment, as textual descriptions may not accurately map to image details. To overcome this limitation, we incorporate contour information to guide semantic pixel alignment and supplement missing semantic strokes. However, we found that overly detailed contour information can disrupt high-level semantics (as edge detection identifies changes based on pixel gradients). For example, when drawing portraits, the eyes are often represented as solid dots rather than pixel-defined circles, necessitating text-based" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.518, + 0.939 + ], + "angle": 0, + "content": "16549" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.103, + 0.48, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.253, + 0.484, + 0.334 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison of sketch generation results with and without contour integration. The top row shows Ref. and Cnt. images, followed by results without contour integration (w/o Cont. Int.) and with contour integration (w/ Cont. Int.) as discussed in Sec. 3.4. Contour integration leads to better alignment of semantic features and reduces background interference." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.339, + 0.418, + 0.353 + ], + "angle": 0, + "content": "semantic guidance for accurate sketch rendering." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.354, + 0.484, + 0.44 + ], + "angle": 0, + "content": "The text-based semantic loss is derived from the guidance provided in the image-to-image diffusion pipeline, which ensures high-level semantic alignment during sketch generation: \\( L_{sem} = \\lambda \\cdot \\mathrm{CLIP}(I^{ske}, T^{cnt}) \\), where \\( T^{cnt} \\) is the text prompt extracted from \\( I^{cnt} \\) and \\( \\lambda \\) is a weighting parameter for the text guidance." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.442, + 0.485, + 0.501 + ], + "angle": 0, + "content": "The contour-based guidance originates from the cached query features during the DDPM inversion process. To integrate this contour information effectively, we use the following equation to adjust the query:" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.506, + 0.484, + 0.525 + ], + "angle": 0, + "content": "\\[\nQ _ {i + 1} ^ {s k e} = \\gamma Q _ {i} ^ {\\text {c o n t}} + (1 - \\gamma) Q _ {i} ^ {s k e}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.53, + 0.484, + 0.618 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is a tunable parameter that controls the influence of contour information, set to 0.25 by default in our experiments. As illustrated in Fig. 7, contour integration significantly improves the alignment of semantic structures, ensuring that key details, such as object outlines, are preserved without introducing unnecessary background noise." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.619, + 0.484, + 0.707 + ], + "angle": 0, + "content": "By combining high-level semantic text guidance with contour-based structural alignment, we ensure that the final generated sketch \\( I^{ske} \\) maintains semantic completeness and accurately represents the image of the content \\( I^{cnt} \\). This collaborative approach allows for the preservation of both stroke attributes and semantic integrity in the sketch." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.714, + 0.446, + 0.73 + ], + "angle": 0, + "content": "3.5. Stroke Details Propagation Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.735, + 0.484, + 0.867 + ], + "angle": 0, + "content": "Traditional self-attention blocks often focus on limited areas around image patches, while masked extended attention blocks distribute attention more uniformly by expanding receptive fields across the image. Although this broader approach captures larger context, it can introduce noise and blur finer details [1, 29]. To address this, we adopt a refined contrast operation, inspired by [1], to dynamically enhance high-variance regions and suppress low-contrast noise. This contrast adjustment is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.875, + 0.484, + 0.891 + ], + "angle": 0, + "content": "\\[\n\\operatorname {E n h a n c e} (A) = (A - \\mu (A)) \\zeta (\\sigma (A)) + \\mu (A), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.103, + 0.895, + 0.161 + ], + "angle": 0, + "content": "where \\(\\mu(A)\\) and \\(\\sigma(A)\\) represent the mean and standard deviation operations, respectively, and \\(\\zeta\\) is an adaptive contrast operator. This technique effectively reduces noise, ensuring critical details are preserved during sketch generation." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.162, + 0.895, + 0.337 + ], + "angle": 0, + "content": "Building on this, we incorporate the stroke-based refinement concept from SDEdit [27]. Starting with a sketch initialized by stroke exchange, we denoise the image using classifier-free guidance (CFG) [6, 14]. During each denoising step, we utilize two parallel forward passes in the network. The first pass employs a cross-image attention layer to capture the stroke characteristics and abstraction level of the reference sketch, generating \\( z^{ske} \\): \\( \\epsilon_{stroke}^{\\times} = \\epsilon_{\\theta}^{\\times}(z_{t}^{ske}) \\), while simultaneously retaining the semantic context extracted from the content image's descriptive prompt: \\( \\epsilon_{text}^{\\times} = \\epsilon_{\\theta}^{\\times}(z_{t}^{text}) \\). The second pass applies regular self-attention to enhance the sketch's structural integrity: \\( \\epsilon^{self} = \\epsilon_{\\theta}^{self}(z_{t}^{ske}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.338, + 0.894, + 0.366 + ], + "angle": 0, + "content": "Following the CFG scheme, the predicted noise \\(\\epsilon^t\\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.373, + 0.895, + 0.393 + ], + "angle": 0, + "content": "\\[\n\\epsilon^ {t} = \\epsilon^ {s e l f} + \\beta_ {s g} (\\epsilon_ {s t r o k e} ^ {\\times} - \\epsilon^ {s e l f}) + \\beta_ {t e x t} (\\epsilon_ {t e x t} ^ {\\times} - \\epsilon^ {s e l f}), (7)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.894, + 0.428 + ], + "angle": 0, + "content": "where \\(\\beta_{sg}\\) is the stroke guidance scale, and \\(\\beta_{text}\\) is the weight for the content text context." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.44, + 0.714, + 0.456 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.465, + 0.895, + 0.567 + ], + "angle": 0, + "content": "Datasets. Our experiments utilize three datasets: FS2K [9], the Anime dataset [19], and our newly created Stroke2Sketch-dataset. The FS2K dataset includes 5,140 facial image-sketch pairs. We randomly selected 1,000 pairs for testing, using color images as content images and choosing one sketch from the test set as a reference style. A similar selection strategy was used for the Anime dataset [19]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.568, + 0.895, + 0.626 + ], + "angle": 0, + "content": "The Stroke2Sketch-dataset includes 50 content images from diverse categories and 20 distinct sketch styles, including single-line sketches, ink sketches, line art, and realistic sketches. Further details are provided in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.626, + 0.895, + 0.714 + ], + "angle": 0, + "content": "Metrics. While traditional metrics like ArtFID [43], LPIPS [53], and FID [13] are standard for style transfer, they struggle to capture the semantic sparsity and high-level abstraction unique to sketch generation. Consistent with prior work [38], we prioritize user perception to better reflect artistic and semantic quality in sketches." + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.722, + 0.729, + 0.737 + ], + "angle": 0, + "content": "4.1. Qualitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.743, + 0.895, + 0.845 + ], + "angle": 0, + "content": "We evaluate our proposed method through a comparison with eight state-of-the-art methods, including three training-based style transfer methods (Ref2sketch [2], Semi-ref2sketch [33], and CSGO [46]), and five training-free style transfer methods (IP-Adapter [49], InstantStyle [41], InstantStyle-plus [42], RB-Modulation [32], and StyleID [7]). Each of these methods takes a reference sketch as input." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.846, + 0.895, + 0.89 + ], + "angle": 0, + "content": "Fig. 8 illustrates the qualitative results across a variety of content images and reference sketches. In the first column, we display the content images, and the second column" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.519, + 0.939 + ], + "angle": 0, + "content": "16550" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.108, + 0.1, + 0.892, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.348, + 0.897, + 0.391 + ], + "angle": 0, + "content": "Figure 8. Qualitative comparison of sketch generation across training-based (4th-6th columns) and training-free baselines (7th-11th columns) using various reference sketches. Our method (3rd column) demonstrates superior adaptability to different reference styles, maintaining both stroke fidelity and semantic consistency across a range of content and reference types." + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.406, + 0.48, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.588, + 0.484, + 0.643 + ], + "angle": 0, + "content": "Figure 9. (a) Content image and manual stroke stylization by CLIPascene [39] with Adobe Illustrator; (b) Our method's automatically generated sketches based on reference stroke styles; (c) Appearance transfer results using similar references." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.655, + 0.484, + 0.8 + ], + "angle": 0, + "content": "shows the reference sketches. Notably, the fourth row's reference sketch is of a type seen in the Semi-ref2sketch training data, while the other three reference sketches represent styles outside the training data for Semi-ref2sketch. For methods requiring prompt input, we uniformly set the prompt to \"sketch photo,\" while other parameters follow each method's default configuration. Our approach demonstrates robustness across diverse reference styles, accurately preserving both the stroke style and high-level semantic details." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.802, + 0.484, + 0.89 + ], + "angle": 0, + "content": "Comparison with vector sketch generation: Fig. 9(a) compares our method with the vectorized sketch generation approach of CLIPascene [39]. The content image is shown alongside vectorized sketches produced by CLIPascene, which were further manually refined with brush strokes in Adobe Illustrator. While CLIPascene can generate abstract" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.402, + 0.892, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.486, + 0.893, + 0.512 + ], + "angle": 0, + "content": "Figure 10. Stroke2Sketch's color sketch generation preserving reference styles and stroke characteristics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.526, + 0.894, + 0.583 + ], + "angle": 0, + "content": "sketches, it requires post-processing to achieve consistent stroke styling, whereas our method (Fig. 9(b)) automatically produces sketches with stylistically aligned strokes based on reference attributes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.895, + 0.687 + ], + "angle": 0, + "content": "Comparison with appearance transfer methods: Fig. 9(c) shows results from appearance transfer method of CIA [1] applied to the same reference sketches. Although CIA excel in transferring visual features based on semantic similarity and object category, they focus on appearance rather than stroke style, making them less suitable for our sketch generation goals." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.894, + 0.733 + ], + "angle": 0, + "content": "Non-grayscale Sketch Generation. Building on its grayscale performance, Stroke2Sketch maintains reference stroke patterns and artistic styles in color outputs (Fig. 10)." + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.736, + 0.74, + 0.751 + ], + "angle": 0, + "content": "4.2.Quantitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.758, + 0.895, + 0.89 + ], + "angle": 0, + "content": "We quantitatively evaluate our method against several state-of-the-art sketch extraction techniques, including both training-based (Ref2Sketch [2], Semi-Ref2Sketch [33], and Informative-drawing [5]) and training-free style transfer methods (IP-Adapter [49], InstantStyle [41], InstantStyle-plus [42], StyleID [7]). Table 1 presents the results on the Stroke2Sketch-dataset, showing that our method achieves the lowest ArtFID and FID scores, indicating superior performance in both style alignment and content preservation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.518, + 0.939 + ], + "angle": 0, + "content": "16551" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.14, + 0.1, + 0.855, + 0.159 + ], + "angle": 0, + "content": "
MetricOursRef2sketchSemi-refInfor-drawingIP-AdapterInstantStyleInstantStyle+StyleID
ArtFID ↓32.45545.29233.24234.21433.45732.53237.65635.727
LPIPS ↓0.53150.69820.53060.60370.66340.54320.65320.5426
FID ↓22.43534.65024.35925.03524.06823.94026.63225.658
" + }, + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.165, + 0.894, + 0.192 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison on Stroke2Sketch-dataset with training-based (3rd-5th columns) and training-free baselines (6th-9th columns)." + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.209, + 0.475, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.364, + 0.483, + 0.392 + ], + "angle": 0, + "content": "Figure 11. User study results. Our proposed method received the highest preference." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.403, + 0.485, + 0.448 + ], + "angle": 0, + "content": "These metrics highlight our method's effectiveness in producing high-quality sketches with strong semantic and stylistic fidelity." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.456, + 0.221, + 0.472 + ], + "angle": 0, + "content": "4.3. User Study" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.478, + 0.483, + 0.668 + ], + "angle": 0, + "content": "We randomly selected 15 content images and 15 reference sketches from Stroke2Sketch-dataset, creating 225 image pairs. From these, we sampled 20 pairs to generate stylized images using four methods. The results were displayed side-by-side in random order, and participants were asked to choose their favorite based on three criteria: content extraction, stroke stylization, and overall preference. We collected 2,000 votes for each criterion from 100 users and presented the results as a bar chart. As shown in Fig. 11, our method received the highest preference, outperforming both training-based and training-free baselines, demonstrating its effectiveness in handling diverse stroke styles and abstract artistic effects." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.678, + 0.251, + 0.694 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.699, + 0.486, + 0.891 + ], + "angle": 0, + "content": "To validate our method's components, we performed an ablation study on the Stroke2Sketch-dataset. As shown in Table 2, removing any of the Directive Attention Module (DAM), Semantic Preservation Module (SPM), or Stroke Details Propagation Enhancement (SDPE) led to decreased performance across ArtFID, FID, and LPIPS metrics, confirming the contribution of each component. Specifically, without DAM, ArtFID increased from 32.45 to 38.67 and FID rose to 26.53, indicating weaker style-content alignment and content leakage in sketches, as observed in Fig. 12. The absence of SPM resulted in ArtFID rising to 36.89 and FID to 30.47, with sketches losing semantic coherence and structural integrity, such as poorly defined object outlines." + }, + { + "type": "table", + "bbox": [ + 0.548, + 0.202, + 0.86, + 0.273 + ], + "angle": 0, + "content": "
ConfigurationArtFIDFIDLPIPS
A: Ours32.4522.430.530
B: - DAM (Sec. 3.3)38.6726.530.672
C: - SPM (Sec. 3.4)36.8930.470.637
D: - SDPE (Sec. 3.5)40.5332.440.598
" + }, + { + "type": "table_caption", + "bbox": [ + 0.532, + 0.284, + 0.874, + 0.297 + ], + "angle": 0, + "content": "Table 2. Ablation study of different variants of our method." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.308, + 0.888, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.441, + 0.896, + 0.468 + ], + "angle": 0, + "content": "Figure 12. Ablative qualitative comparison of different variants of our method." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.484, + 0.896, + 0.617 + ], + "angle": 0, + "content": "Most notably, removing SDPE caused the sharpest decline, with ArtFID reaching 40.53 and sketches exhibiting excessive noise and loss of fine details, as evidenced by cluttered textures in Fig. 12. In contrast, the full method (Configuration A) achieved optimal scores of 32.45 (ArtFID), 22.43 (FID), and 0.530 (LPIPS), producing high-quality, reference-aligned sketches with precise details. These results highlight the critical roles of DAM, SPM, and SDPE in enhancing sketch quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.63, + 0.629, + 0.644 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.655, + 0.896, + 0.891 + ], + "angle": 0, + "content": "Our training-free method for content-to-sketch generation aligns stroke attributes and semantic texture sparsity, mimicking artistic subject extraction to produce high-fidelity sketches. Leveraging pretrained models, it achieves state-of-the-art performance in quantitative metrics and user evaluations. Limitations arise with overly simplistic or complex reference sketches (see appendix for failure cases). Future work could explore decoupling semantic information from stroke attributes to improve adaptability and sketch quality. Acknowledgment. This work is supported by the Guangdong Natural Science Funds for Distinguished Young Scholars (Grant 2023B1515020097), the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No.: AISG3-GV-2023-011), the Singapore Ministry of Education AcRF Tier 1 Grant (Grant No.: MSS25C004), and the Lee Kong Chian Fellowships." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.518, + 0.939 + ], + "angle": 0, + "content": "16552" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.104, + 0.102, + 0.196, + 0.117 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.126, + 0.486, + 0.18 + ], + "angle": 0, + "content": "[1] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer. In ACM SIGGRAPH 2024 Conference Papers, pages 1–12, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.181, + 0.485, + 0.234 + ], + "angle": 0, + "content": "[2] Amirsaman Ashtari, Chang Wook Seo, Cholmin Kang, Sihun Cha, and Junyong Noh. Reference based sketch extraction via attention mechanism. ACM Trans. Graph., 41(6), 2022. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.235, + 0.485, + 0.275 + ], + "angle": 0, + "content": "[3] John Canny. A computational approach to edge detection. IEEE Transactions on pattern analysis and machine intelligence, (6):679-698, 1986. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.277, + 0.484, + 0.33 + ], + "angle": 0, + "content": "[4] Nan Cao, Xin Yan, Yang Shi, and Chaoran Chen. Ai-sketcher: a deep generative model for producing high-quality sketches. In Proceedings of the AAAI conference on artificial intelligence, pages 2564-2571, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.332, + 0.484, + 0.396 + ], + "angle": 0, + "content": "[5] Caroline Chan, Frédo Durand, and Phillip Isola. Learning to generate line drawings that convey geometry and semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7915-7925, 2022. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.399, + 0.484, + 0.44 + ], + "angle": 0, + "content": "[6] Hyungjin Chung, Jeongsol Kim, Geon Yeong Park, Hyelin Nam, and Jong Chul Ye. Cfg++: Manifold-constrained classifier free guidance for diffusion models. 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.441, + 0.485, + 0.508 + ], + "angle": 0, + "content": "[7] Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8795-8805, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.509, + 0.485, + 0.587 + ], + "angle": 0, + "content": "[8] Ayan Das, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Beziersketch: A generative model for scalable vector sketches. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI 16, pages 632-647. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.59, + 0.485, + 0.643 + ], + "angle": 0, + "content": "[9] Deng-Ping Fan, Ziling Huang, Peng Zheng, Hong Liu, Xuebin Qin, and Luc Van Gool. Facial-sketch synthesis: a new challenge. Machine Intelligence Research, 19(4):257-287, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.645, + 0.485, + 0.698 + ], + "angle": 0, + "content": "[10] Kevin Frans, Lisa Soros, and Olaf Witkowski. Clipdraw: Exploring text-to-drawing synthesis through language-image encoders. Advances in Neural Information Processing Systems, 35:5207-5218, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.7, + 0.485, + 0.753 + ], + "angle": 0, + "content": "[11] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.755, + 0.485, + 0.82 + ], + "angle": 0, + "content": "[12] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4775–4785, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.822, + 0.485, + 0.888 + ], + "angle": 0, + "content": "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "list", + "bbox": [ + 0.104, + 0.126, + 0.486, + 0.888 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.104, + 0.895, + 0.13 + ], + "angle": 0, + "content": "[14] Jonathan Ho. Classifier-free diffusion guidance. ArXiv, abs/2207.12598, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.132, + 0.896, + 0.184 + ], + "angle": 0, + "content": "[15] Jjin Hu, Ke Li, Yonggang Qi, and Yi-Zhe Song. Scale-adaptive diffusion model for complex sketch synthesis. In The Twelfth International Conference on Learning Representations, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.186, + 0.896, + 0.252 + ], + "angle": 0, + "content": "[16] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly DDPM noise space: Inversion and manipulations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12469-12478, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.254, + 0.896, + 0.293 + ], + "angle": 0, + "content": "[17] Jaeseok Jeong, Junho Kim, Yunjoy Choi, Gayoung Lee, and Youngjung Uh. Visual style prompting with swapping self-attention. arXiv preprint arXiv:2402.12974, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.295, + 0.896, + 0.36 + ], + "angle": 0, + "content": "[18] Yutao Jiang, Yang Zhou, Yuan Liang, Wenxi Liu, Jianbo Jiao, Yuhui Quan, and Shengfeng He. Diffuse3d: Wide-angle 3d photography via bilateral diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8998-9008, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.362, + 0.896, + 0.387 + ], + "angle": 0, + "content": "[19] Tae Bum Kang. Anime sketch colorization pair, 2018. Accessed: 2024-05-18. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.389, + 0.896, + 0.455 + ], + "angle": 0, + "content": "[20] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.457, + 0.896, + 0.524 + ], + "angle": 0, + "content": "[21] Xinzhe Li, Jiahui Zhan, Shengfeng He, Yangyang Xu, Junyu Dong, Huaidong Zhang, and Yong Du. Personamagic: Stage-regulated high-fidelity face customization with tandem equilibrium. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4995-5003, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.525, + 0.896, + 0.577 + ], + "angle": 0, + "content": "[22] Chenxi Liu, Enrique Rosales, and Alla Sheffer. Strokeaggregator: Consolidating raw sketches into artist-intended curve drawings. ACM Transactions on Graphics (TOG), 37(4):1-15, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.579, + 0.896, + 0.631 + ], + "angle": 0, + "content": "[23] Chenxi Liu, Toshiki Aoki, Mikhail Bessmeltsev, and Alla Sheffer. Stripmaker: Perception-driven learned vector sketch consolidation. ACM Transactions on Graphics (TOG), 42(4): 1-15, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.633, + 0.896, + 0.699 + ], + "angle": 0, + "content": "[24] Difan Liu, Matthew Fisher, Aaron Hertzmann, and Evangelos Kalogerakis. Neural strokes: Stylized line drawing of 3d shapes. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 14184-14193. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.701, + 0.896, + 0.767 + ], + "angle": 0, + "content": "[25] Haofeng Liu, Chenshu Xu, Yifei Yang, Lihua Zeng, and Shengfeng He. Drag your noise: Interactive point-based editing via diffusion semantic propagation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6743-6752, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.768, + 0.896, + 0.82 + ], + "angle": 0, + "content": "[26] Xiao-Chang Liu, Yu-Chen Wu, and Peter Hall. Painterly style transfer with learned brush strokes. IEEE Transactions on Visualization and Computer Graphics, 30(9):6309-6320, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.822, + 0.896, + 0.888 + ], + "angle": 0, + "content": "[27] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.104, + 0.896, + 0.888 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.518, + 0.939 + ], + "angle": 0, + "content": "16553" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.104, + 0.484, + 0.143 + ], + "angle": 0, + "content": "[28] Reiichiro Nakano. Neural painters: A learned differentiable constraint for generating brushstroke paintings. arXiv preprint arXiv:1904.08410, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.145, + 0.484, + 0.197 + ], + "angle": 0, + "content": "[29] Nadav Orzech, Yotam Nitzan, Ulysse Mizrahi, Dov Danon, and Amit H Bermano. Masked extended attention for zero-shot virtual try-on in the wild. arXiv preprint arXiv:2406.15331, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.105, + 0.2, + 0.484, + 0.265 + ], + "angle": 0, + "content": "[30] Xavier Soria Poma, Edgar Riba, and Angel Sappa. Dense extreme inception network: Towards a robust cnn model for edge detection. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1923-1932, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.105, + 0.268, + 0.484, + 0.333 + ], + "angle": 0, + "content": "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.335, + 0.484, + 0.386 + ], + "angle": 0, + "content": "[32] L Rout, Y Chen, N Ruiz, A Kumar, C Caramanis, S Shakkottai, and W Chu. Rb-modulation: Training-free personalization of diffusion models using stochastic optimal control. 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.389, + 0.484, + 0.441 + ], + "angle": 0, + "content": "[33] Chang Wook Seo, Amirsaman Ashtari, and Junyong Noh. Semi-supervised reference-based sketch extraction using a contrastive learning framework. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.443, + 0.484, + 0.496 + ], + "angle": 0, + "content": "[34] Edgar Simo-Serra, Satoshi Iizuka, Kazuma Sasaki, and Hiroshi Ishikawa. Learning to simplify: fully convolutional networks for rough sketch cleanup. ACM Transactions on Graphics (TOG), 35(4), 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.498, + 0.484, + 0.55 + ], + "angle": 0, + "content": "[35] Gowthami Somepalli, Anubhav Gupta, Kamal Gupta, Shramay Palta, Micah Goldblum, Jonas Geiping, Abhinav Shrivastava, and Tom Goldstein. Measuring style similarity in diffusion models. arXiv preprint arXiv:2404.01292, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.552, + 0.484, + 0.592 + ], + "angle": 0, + "content": "[36] Xavier Soria, Yachuan Li, Mohammad Rouhani, and Angel D Sappa. Tiny and efficient model for the edge detection generalization. In CVPR, pages 1364-1373, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.593, + 0.484, + 0.658 + ], + "angle": 0, + "content": "[37] Zhuo Su, Wenzhe Liu, Zitong Yu, Dewen Hu, Qing Liao, Qi Tian, Matti Pietikainen, and Li Liu. Pixel difference networks for efficient edge detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5117-5127, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.66, + 0.484, + 0.725 + ], + "angle": 0, + "content": "[38] Yael Vinker, Ehsan Pajouheshgar, Jessica Y Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, and Ariel Shamir. Clipasso: Semantically-aware object sketching. ACM Transactions on Graphics (TOG), 41(4): 1-11, 2022. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.728, + 0.484, + 0.793 + ], + "angle": 0, + "content": "[39] Yael Vinker, Yuval Alaluf, Daniel Cohen-Or, and Ariel Shamir. Clipscene: Scene sketching with different types and levels of abstraction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4146–4156, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.796, + 0.484, + 0.835 + ], + "angle": 0, + "content": "[40] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. p+: Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.837, + 0.484, + 0.889 + ], + "angle": 0, + "content": "[41] Haofan Wang, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv preprint arXiv:2404.02733, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.104, + 0.104, + 0.484, + 0.889 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.104, + 0.894, + 0.157 + ], + "angle": 0, + "content": "[42] Haofan Wang, Peng Xing, Renyuan Huang, Hao Ai, Qixun Wang, and Xu Bai. Instantstyle-plus: Style transfer with content-preserving in text-to-image generation. arXiv preprint arXiv:2407.00788, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.16, + 0.895, + 0.199 + ], + "angle": 0, + "content": "[43] Matthias Wright and Björn Ommer. Artfid: Quantitative evaluation of neural style transfer. In DAGM German Conference on Pattern Recognition, pages 560-576. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.201, + 0.895, + 0.267 + ], + "angle": 0, + "content": "[44] Zongwei Wu, Liangyu Chai, Nanxuan Zhao, Bailin Deng, Yongtuo Liu, Qiang Wen, Junle Wang, and Shengfeng He. Make your own sprites: Aliasing-aware and cell-controllable pixelization. ACM Transactions on Graphics (TOG), 41(6): 1-16, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.269, + 0.895, + 0.309 + ], + "angle": 0, + "content": "[45] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In Proceedings of the IEEE international conference on computer vision, pages 1395-1403, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.311, + 0.895, + 0.363 + ], + "angle": 0, + "content": "[46] Peng Xing, Haofan Wang, Yanpeng Sun, Qixun Wang, Xu Bai, Hao Ai, Renyuan Huang, and Zechao Li. Csgo: Content-style composition in text-to-image generation. arXiv preprint arXiv:2408.16766, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.366, + 0.895, + 0.42 + ], + "angle": 0, + "content": "[47] Chenshu Xu, Yangyang Xu, Huaidong Zhang, Xuemiao Xu, and Shengfeng He. Dreamanine: Learning style-identity textual disentanglement for anime and beyond. IEEE Transactions on Visualization and Computer Graphics, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.422, + 0.895, + 0.473 + ], + "angle": 0, + "content": "[48] Rui Yang, Xiaojun Wu, and Shengfeng He. Mixsa: Training-free reference-based sketch extraction via mixture-of-self-attention. IEEE Transactions on Visualization and Computer Graphics, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.895, + 0.528 + ], + "angle": 0, + "content": "[49] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.531, + 0.895, + 0.61 + ], + "angle": 0, + "content": "[50] Yuyang Yu, Bangzhen Liu, Chenxi Zheng, Xuemiao Xu, Huaidong Zhang, and Shengfeng He. Beyond textual constraints: Learning novel diffusion conditions with fewer examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7109-7118, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.613, + 0.895, + 0.666 + ], + "angle": 0, + "content": "[51] Sicong Zang, Shikui Tu, and Lei Xu. Self-organizing a latent hierarchy of sketch patterns for controllable sketch synthesis. IEEE Transactions on Neural Networks and Learning Systems, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.668, + 0.894, + 0.722 + ], + "angle": 0, + "content": "[52] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.723, + 0.895, + 0.788 + ], + "angle": 0, + "content": "[53] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.791, + 0.895, + 0.857 + ], + "angle": 0, + "content": "[54] Caixia Zhou, Yaping Huang, Mengyang Pu, Qingji Guan, Ruoxi Deng, and Haibin Ling. Muge: Multiple granularity edge detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25952-25962, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.104, + 0.895, + 0.857 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.519, + 0.939 + ], + "angle": 0, + "content": "16554" + } + ] +] \ No newline at end of file diff --git a/2025/Stroke2Sketch_ Harnessing Stroke Attributes for Training-Free Sketch Generation/88f15fb1-f5ab-4157-a23c-15d3cb3bc8f3_origin.pdf b/2025/Stroke2Sketch_ Harnessing Stroke Attributes for Training-Free Sketch Generation/88f15fb1-f5ab-4157-a23c-15d3cb3bc8f3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..49a8a8d7683ff576cf35d27e6c79c897b65ec203 --- /dev/null +++ b/2025/Stroke2Sketch_ Harnessing Stroke Attributes for Training-Free Sketch Generation/88f15fb1-f5ab-4157-a23c-15d3cb3bc8f3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:51ac41fecc3d90f33db17bc55f9c9f603a20450d7ce3a22e6bf678ebe8b58149 +size 1100457 diff --git a/2025/Stroke2Sketch_ Harnessing Stroke Attributes for Training-Free Sketch Generation/full.md b/2025/Stroke2Sketch_ Harnessing Stroke Attributes for Training-Free Sketch Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..19450f6651121848e74518484ec1d7989da5df15 --- /dev/null +++ b/2025/Stroke2Sketch_ Harnessing Stroke Attributes for Training-Free Sketch Generation/full.md @@ -0,0 +1,319 @@ +# Stroke2Sketch: Harnessing Stroke Attributes for Training-Free Sketch Generation + +Rui Yang $^{1,3}$ , Huining Li $^{4,5}$ , Yiyi Long $^{2,5}$ , Xiaojun Wu $^{3,*}$ , Shengfeng He $^{5,*}$ + +$^{1}$ Huaqiao University $^{2}$ South China University of Technology $^{3}$ Shaanxi Normal University $^{4}$ Beijing University of Aeronautics and Astronautics $^{5}$ Singapore Management University + +![](images/537aacecb15601e21cbc49becd56d0f3bac6c79f6805b55df94b65f75014510a.jpg) +Figure 1. We propose Stroke2Sketch, a framework that accurately transfers stroke attributes from a reference sketch to a content image while preserving structure and style fidelity. The top row shows reference sketches, the leftmost column displays content images, and the central and right columns illustrate our method's precise content preservation and expressive stroke transfer. + +# Abstract + +Generating sketches guided by reference styles requires precise transfer of stroke attributes, such as line thickness, deformation, and texture sparsity, while preserving semantic structure and content fidelity. To this end, we propose Stroke2Sketch, a novel training-free framework that introduces cross-image stroke attention, a mechanism embedded within self-attention layers to establish fine-grained semantic correspondences and enable accurate stroke attribute transfer. This allows our method to adaptively integrate reference stroke characteristics into content images while maintaining structural integrity. Additionally, we develop adaptive contrast enhancement and semantic-focused attention to reinforce content preservation and foreground emphasis. Stroke2Sketch effectively synthesizes stylistically faithful sketches that closely resemble handcrafted results, outperforming existing methods in expressive stroke control and semantic coherence. Codes are available at https://github.com/rane7/Stroke2Sketch. + +# 1. Introduction + +Generating stylized sketches from content images using reference stroke patterns presents a key challenge at the intersection of artistic rendering and semantic-aware synthesis. Traditional sketch algorithms [28, 34] generate procedural line drawings, while vector-based methods [24, 26, 38] produce clean parametric strokes. However, these methods lack the adaptability to transfer diverse artistic styles from exemplar sketches due to their rigid, data-agnostic architectures. Unlike human artists, who strategically vary stroke attributes such as thickness, curvature, and density to emphasize key semantic features while maintaining content fidelity, existing approaches struggle to capture this nuanced interplay between stroke semantics and structure. + +Recent learning-based methods [2, 5, 33] attempt to address this limitation by training on clustered sketch styles, yet as shown in Fig. 2(a-b), they fail to generalize to unseen stroke patterns due to catastrophic forgetting. Meanwhile, diffusion-based stylization techniques [21, 41, 47, 49, 50] excel in texture transfer but struggle with structural integrity due to content leakage in cross-attention layers (Fig. 2(c-d)). While new conditioning mechanisms [18, 25, 44] attempt to enhance structural control, they often introduce style dis + +![](images/2779d5cc0e7051ec1056b65b671bcf2ad5e1498cd4f647273e68de9fe6dcd703.jpg) +Figure 2. (a) and (b) show results from training-based methods (Ref2Sketch and Semi-Ref2Sketch), which struggle with unseen styles, leading to poor content alignment. (c) and (d) illustrate IP-Adapter and InstantStyle results, retaining style but lacking content alignment. (e) to (g) show ControlNet-based methods, preserving content but failing to match reference styles. (h) shows our method's output, achieving superior content preservation and style alignment without using ControlNet. Detailed prompts are shown in the supplementary materials. + +tortion or require dense user inputs. Hybrid approaches like ControlNet [52] enforce structural priors but sacrifice stylistic flexibility, leading to overly rigid outputs (Fig. 2(e)). Progressive stroke-based methods [32] introduce semantic incoherence by applying uniform strokes across the image (Fig. 2(g)). + +We identify three fundamental challenges in reference-based sketch synthesis:(i) Semantic-aware stroke transfer. Effective style adaptation requires precise mapping of reference stroke attributes (e.g., tapered lines, cross-hatching) to semantically relevant content regions. (ii) Foreground prioritization. Artists naturally emphasize foreground elements using varied stroke density and complexity while simplifying backgrounds, yet existing methods apply uniform stylization, disrupting this compositional balance[32]. (iii) Content-style equilibrium. Since sketches encode content through linework, balancing structural preservation with style transfer is critical. Techniques such as CLIP-space style subtraction [41] fail to maintain this balance, as even minor content leakage distorts key edges (Fig. 2(c-d)). + +To address these challenges, we propose Stroke2Sketch, a framework that enables precise stroke attribute transfer while maintaining content fidelity. Our key insight is that stroke properties, like line thickness, curvature, and texture sparsity, are inherently encoded within the self-attention and cross-attention relationships of pretrained diffusion models. By dynamically aligning these attention patterns between content and reference features, we achieve effective style transfer without structural degradation. Stroke2Sketch integrates three novel components tailored to each identified challenge: + +(1) Cross-image stroke attention tackles the challenge of + +semantic-aware stroke transfer by facilitating stroke attribute exchange through key-value swapping in diffusion layers. Instead of directly blending features [41], we leverage attention blocks to transplant stroke characteristics onto content structures, preventing the entanglement of style and geometry. This approach ensures accurate stroke mapping without disrupting semantic coherence, as shown in Fig. 2(h). + +(2) Directive Attention Module ensures that stroke transfer remains compositionally balanced. Background textures often introduce conflicting patterns that dilute the intended style. We mitigate this by clustering self-attention maps to isolate foreground objects, then restricting cross-image attention to these regions. This mimics how artists prioritize focal elements while simplifying less critical areas, enhancing both style consistency and perceptual quality. +(3) Semantic Preservation Module addresses content-style equilibrium by injecting content contours as positional queries during early denoising. This hybridizes the precision of edge detectors with the flexibility of text-driven generation, allowing structural guidance without rigid constraints. Unlike ControlNet, which enforces strict boundaries, our approach treats edges as "soft constraints" that evolve into stylized strokes, preserving both structure and artistic abstraction. + +As validated in Fig. 2(h) and Fig. 6, Stroke2Sketch achieves state-of-the-art performance across diverse sketch styles. Experimental results show that our method outperforms adapter-based and ControlNet methods in both style alignment (87% user preference) and content preservation (92% accuracy in line correspondence tests). Importantly, it achieves these results without dataset-specific training or architectural modifications, demonstrating that pretrained + +diffusion models can master sketch stylization when guided by principled feature interactions. + +# 2. Related Work + +# 2.1. Photo-Sketch Synthesis + +Generating a sketch from a content image based on a reference style parallels edge detection, as both tasks emphasize prominent visual features. Edge detection methods [3, 30, 36, 37, 45, 54], which detect sharp changes in color or brightness, form the foundation of sketch extraction but are limited to a single style and often produce artifacts, like scattered dots or broken lines. High-quality sketch generation requires more than edge detection; it demands line style, texture sparsity, and semantic abstraction to achieve artist-level results. + +Learning-based approaches have improved sketch realism by enhancing boundary detection and rendering distinct line styles. For example, Chan et al. [5] incorporated depth and semantic information for better sketch quality, while Ref2Sketch [2] and Semi-Ref2Sketch [33] leverage paired and semi-supervised training for stylized sketch extraction. However, these methods rely on large sketch datasets, which are challenging to obtain, limiting model robustness. + +Another research area, stroke-based rendering, focuses on manipulating strokes and contours for sketching. Methods like CLIPDraw [10] and CLIPasso [38] employ Bezier curves to create abstract sketches with high-level semantic simplification. StrokeAggregator [22] and StripMaker [23] refine vector sketches, achieving quality comparable to artist drawings. Other methods include semantic concept-to-sketch methods [4, 8, 15, 51]. However, these methods assume uniform style, while real-world sketches vary widely. + +Building on these insights, our method combines stroke attributes, semantic abstraction, and expression to better align generated sketches with diverse reference styles. + +# 2.2. High-Semantic Style Transfer + +Generating sketches that adhere to both content and reference style is a specialized style transfer task. Advances in diffusion models have propelled style transfer through self- and cross-attention mechanisms, which preserve spatial layout and stylized content. Techniques such as Prompt-to-Prompt [11] and $\mathrm{P + }$ [40] show how attention mechanisms can maintain structural coherence while enabling flexible style and semantic control. + +Cross-Image Attention (CIA) [1] and methods like Swapping Self-Attention [17] and StyleAligned [12] reveal that style features are best retained during upsampling stages, even though content leakage may occur in bottleneck phases. IP-Adapter achieves style transfer via dual cross-attention with text and image, though it can weaken text control and lead to leakage. + +InstantStyle [41] addresses leakage by subtracting features in the same feature space but often requires external constraints like ControlNet [52] for image-to-image generation, which can dilute style fidelity. Further developments like InstantStyle-plus [42] and RB-Modulation [32] incorporate CSD [35] to better align styles during generation, but limitations remain in sketch generation. Prompt-to-style transfer, however, demonstrates pretrained models' strong semantic alignment capabilities, providing valuable priors for consistent, high-level sketch synthesis. + +# 3. Stroke2Sketch + +Given a content image $I^{cnt} \in \mathbb{R}^{H^{cnt} \times W^{cnt} \times 3}$ and a reference sketch image $I^{ref} \in \mathbb{R}^{H^{ref} \times W^{ref} \times 3}$ , our task is to generate a sketch $I^{ske}$ that aligns with the content structure of $I^{cnt}$ while adhering to the stroke style, texture sparsity, and high-level semantic abstraction of $I^{ref}$ . Additionally, we aim to remove or retain background elements as needed to enhance the foreground object. In this work, we address the sketch generation task using a controllable text-to-image guidance approach. Specifically, we extract object prompts from the content image $I^{cnt}$ using BLIP [20] and incorporate the stroke style and high-level abstraction from the reference sketch $I^{ref}$ to re-render the final target sketch $I^{ske}$ . To achieve this, we leverage a pre-trained text-to-image model. The overall network architecture is illustrated in Fig 3, with detailed explanations of each module provided below. + +# 3.1. Preliminaries + +DDPM inversion [16] is a process in diffusion-based generative models that reverses denoising steps, enabling the reconstruction of latent representations from generated outputs. Compared to DDIM inversion method [48], DDPM inversion offers greater flexibility for editing tasks by producing noise maps that, while correlated across timesteps and not normally distributed, allow precise image reconstruction and meaningful edits such as color adjustments and structural shifts. A key advantage of DDPM inversion is its ability to maintain an image's structure while altering the conditioning input, such as a text prompt, to enable artifact-free edits that adapt semantics while preserving original details. This efficient method bypasses optimization processes and can enhance diffusion-based editing techniques by improving image fidelity and supporting diverse outputs. + +Stable Diffusion [31] incorporates this inversion process within a latent space, rather than a pixel space, increasing computational efficiency and expressiveness. The input image is encoded through a pretrained variational autoencoder (VAE) to produce a latent code $z$ . Denoising then occurs within this latent space using a U-Net architecture that incorporates self-attention and cross-attention mechanisms. Self-attention blocks enhance image detail by calculating attention scores between projected query $Q$ , key $K$ , and value + +![](images/882d2df4e65ecfec500fba46b0e2dfd429bef22513cbc659af43661b6d93875a.jpg) +Figure 3. The Stroke2Sketch architecture. The content image $I^{\mathrm{cnt}}$ undergoes contour detection, generating feature representations $z^{\mathrm{cont}}$ and $z^{\mathrm{cnt}}$ , while the reference sketch $I^{\mathrm{ref}}$ is inverted to produce $z^{\mathrm{ref}}$ . The Directive Attention Module (DAM) aligns high-level semantic features between the content and reference features by emphasizing cross-image semantic correspondences $A_{n}$ . Self-attention maps $F_{SA}$ are aggregated and clustered to produce segmentation masks $M_{i}$ , which help distinguish foreground from background regions. Cross-image Stroke Attention (CSA) transfers stroke attributes, and the Semantic Preservation Module (SPM) enforces semantic alignment and stroke fidelity in the generated sketch $\hat{I}^{\mathrm{ske}}$ via loss $\mathcal{L}_{\mathrm{sem}}$ and contour-based structural integration. Lastly, Stroke Detail Propagation Enhancement (SDPE) refines details, resulting in the final output sketch $I^{\mathrm{ske}}$ . SD represents the pre-trained diffusion model. + +$V$ vectors: + +$$ +A = \operatorname {s o f t m a x} \left(\frac {Q \cdot K ^ {\top}}{\sqrt {d}}\right), \quad \phi = A \cdot V, \tag {1} +$$ + +where $A$ is the attention map, $d$ is the dimensionality, and $\phi$ is the output of the self-attention block. Cross-attention blocks incorporate conditioning inputs (e.g., text prompts) to guide the generation process. + +Finally, the refined latent representation $z$ is decoded back into an RGB image using the VAE decoder, yielding a high-quality output guided by the conditioning input. + +# 3.2. Cross-image Stroke Attention + +As discussed in the Introduction, sketch generation as a specialized form of style transfer requires attention to local stroke attributes and consistent texture abstraction across different levels of semantic sparsity. Edge detection serves as the foundation for this task, with edges defining content contours as the most effective strategy for preserving structural information. We extract the contour $I^{cont}$ from the content image using TEED [36] and obtain the inverted latents $z_{T}^{cnt}$ , $z_{T}^{ref}$ , and $z_{T}^{cont}$ for the content image, reference sketch, and + +contour image using DDPM inversion. During the denoising process, the latent $z_{T}^{cnt}$ from the content image serves as the initial noise $z_{T}^{ske}$ for sketch denoising. At each timestep $t$ , we apply Equation 1. + +To achieve effective stroke feature transfer, we utilize latent representations for the content, reference, and contour images obtained through DDPM inversion. For predefined timesteps $t \in \{0, \dots, T\}$ , the reference sketch $I^{ref}$ , content image $I^{cnt}$ , and contour image $I^{cont}$ are inverted from their initial state ( $t = 0$ ) to Gaussian noise ( $t = T$ ). During DDPM inversion, we also collect the query features of the content ( $Q_{t}^{cnt}$ ), and the key and value features of the reference sketch ( $K_{t}^{ref}$ , $V_{t}^{ref}$ ) at each timestep. + +We initialize the latent noise $z_{T}^{ske}$ for the stylized sketch by copying the content latent noise $z_{T}^{cnt}$ . To blend features from all three images, we combine the reference keys and values $K_{t}^{ref}$ and $\tilde{V}_{t}^{ref}$ with the content features. This integration is achieved by mixing the keys and values using the following formulation: + +$$ +K _ {t} ^ {\mathrm {s k e}} = K _ {t} ^ {\mathrm {r e f}} + \alpha K _ {t} ^ {\mathrm {c n t}}, \tag {2} +$$ + +$$ +V _ {t} ^ {\mathrm {s k e}} = V _ {t} ^ {\mathrm {r e f}} + \alpha V _ {t} ^ {\mathrm {c n t}}, \tag {3} +$$ + +![](images/916b50d410eacea00395eb201704d9cd32dba2a002a668c04813f9289408722e.jpg) + +![](images/50bdac387e457df1f2c4051ad7ca7814f72f93917c92d4ccd6d9b321fc8fc004.jpg) +Content Strokes + +![](images/3c75c4095600f2a7e39dbdc113f186be63fda64ee19dc137fb7bfd073b475ba0.jpg) + +![](images/543474c3d275e0516843866f2c5f783a6c7a9a57f9fb6727f66544c37412800c.jpg) +Reference Strokes + +![](images/67f0c0896f638805e512dcf825aa7fff932a2333abce36eac2f84f9742d80291.jpg) + +![](images/0326bad4b692da42f634f42cb19be639ba3cc52f4612d589683911b057c66903.jpg) +Aligned Strokes +Figure 4. Stroke alignment results with CSA. The feature images are shown in the first row. The second row shows results where feature alignment between the key and value exchange is applied. + +where $\alpha$ is a scalar parameter that controls the mixing ratio of content and reference features, allowing flexible adaptation to different stroke characteristics in the sketch. + +By exchanging and blending features from the content, reference, and contour images, this approach enables the effective transfer of stroke attributes and semantic alignment. However, we found that some strokes may fail to represent the intended content curves accurately (see Fig. 4). To improve output quality and highlight the foreground, we introduce additional mechanisms to guide sketch generation, as detailed below. + +# 3.3. Directive Attention Module + +To concentrate stylization on perceptually significant regions, the DAM enhances foreground focus during stroke transfer. We extract self-attention feature maps $F_{SA}$ at $32 \times 32$ resolution, aggregating them across channels via averaging. These maps are segmented into clusters $M_{j}$ using KMeans clustering. For each cluster $j$ and noun $n$ extracted from the content image's caption (via BLIP [20]), we calculate a relevance score: + +$$ +r (j, n) = \frac {\sum_ {(x , y)} M _ {j} (x , y) \cdot A _ {n} (x , y)}{\sum_ {(x , y)} M _ {j} (x , y) + \delta}, \tag {4} +$$ + +where $A_{n}$ is the cross-attention map for noun $n$ , and $\delta = 10^{-5}$ stabilizes the computation. Clusters with $r(j,n) > 0.35$ are designated as foreground regions; remaining areas are suppressed, directing stroke stylization to salient elements (Fig. 5). + +By integrating the segmented self-attention mechanism with cross-image attention, DAM allows precise control over foreground regions, ensuring that stylistic features, such as line styles and high-level semantic abstractions, are faithfully transferred from the reference sketch to the content image. This approach achieves high fidelity in style and content alignment, yielding sketches that closely reflect the intended reference style with minimized background interference. + +As illustrated in Fig. 6, our method successfully aligns high-level semantic features and detailed stroke attributes, + +![](images/1a2b12d3d4bb37065f294fc3d1bca357fde672634747f6add6c2260e06029624.jpg) +Figure 5. Examples of DAM in action. The first column shows the content images, the second column displays segmentation maps obtained by clustering self-attention maps, the third column provides the reference sketches, and the fourth column illustrates the sketches generated by DAM after applying stroke attribute transfer and segmentation. + +![](images/d3811bdf5b815f468ac4b1ac1a816a5a7e60fc3713e246f928b19c0fb34aa961.jpg) + +![](images/efab180c65eeab9dc62253d9bed59e8cc840f4ff5778773ab44d31900d83fc7f.jpg) + +![](images/d8543fee5368f488fb073440f20cb758a02e4566888a46cb536b70cb476234a7.jpg) + +![](images/1b5d661283d7becd8dd705ead768a7b6c14c027dcd521b62502e32a5675f6e8c.jpg) +Figure 6. Comparison of sketches generated using DAM. The top row shows Cnt. and Ref. images with stroke styles reflecting specific high-level attributes such as hair, eyebrows, and abstracted clothing texture. The second to fourth columns display sketches generated to match the content images in the top row, each adopting the stroke styles and high-level features from the reference. The bottom row highlights zoomed-in areas to emphasize the transferred stroke attributes. + +![](images/9231b8af5b80fd7ce3c5aaa1552821f3a3b982f4f606084f788a51cd3d165235.jpg) + +producing coherent sketches that maintain stylistic and structural consistency with the reference. This approach minimizes background interference while ensuring that the final sketch reflects the nuanced stroke characteristics of the reference. + +# 3.4. Semantic Preservation Module + +Although the injection of keys and values during sketch generation can effectively transfer stroke attributes to $I^{ske}$ , semantic inconsistencies may arise, particularly when the reference sketch $I^{ref}$ does not align semantically with the content image $I^{cnt}$ . This can lead to noise and misplaced strokes that disrupt the semantic structure of the generated sketch. To address this, semantic guidance is essential to ensure that each pixel aligns correctly with its corresponding structure. + +Text-based guidance alone is often insufficient for precise structural alignment, as textual descriptions may not accurately map to image details. To overcome this limitation, we incorporate contour information to guide semantic pixel alignment and supplement missing semantic strokes. However, we found that overly detailed contour information can disrupt high-level semantics (as edge detection identifies changes based on pixel gradients). For example, when drawing portraits, the eyes are often represented as solid dots rather than pixel-defined circles, necessitating text-based + +![](images/0c076db4a915ddbcc366be45d1e4ce2c54ee2cbab395a50ed83c0304139675c8.jpg) +Figure 7. Qualitative comparison of sketch generation results with and without contour integration. The top row shows Ref. and Cnt. images, followed by results without contour integration (w/o Cont. Int.) and with contour integration (w/ Cont. Int.) as discussed in Sec. 3.4. Contour integration leads to better alignment of semantic features and reduces background interference. + +semantic guidance for accurate sketch rendering. + +The text-based semantic loss is derived from the guidance provided in the image-to-image diffusion pipeline, which ensures high-level semantic alignment during sketch generation: $L_{sem} = \lambda \cdot \mathrm{CLIP}(I^{ske}, T^{cnt})$ , where $T^{cnt}$ is the text prompt extracted from $I^{cnt}$ and $\lambda$ is a weighting parameter for the text guidance. + +The contour-based guidance originates from the cached query features during the DDPM inversion process. To integrate this contour information effectively, we use the following equation to adjust the query: + +$$ +Q _ {i + 1} ^ {s k e} = \gamma Q _ {i} ^ {\text {c o n t}} + (1 - \gamma) Q _ {i} ^ {s k e}, \tag {5} +$$ + +where $\gamma$ is a tunable parameter that controls the influence of contour information, set to 0.25 by default in our experiments. As illustrated in Fig. 7, contour integration significantly improves the alignment of semantic structures, ensuring that key details, such as object outlines, are preserved without introducing unnecessary background noise. + +By combining high-level semantic text guidance with contour-based structural alignment, we ensure that the final generated sketch $I^{ske}$ maintains semantic completeness and accurately represents the image of the content $I^{cnt}$ . This collaborative approach allows for the preservation of both stroke attributes and semantic integrity in the sketch. + +# 3.5. Stroke Details Propagation Enhancement + +Traditional self-attention blocks often focus on limited areas around image patches, while masked extended attention blocks distribute attention more uniformly by expanding receptive fields across the image. Although this broader approach captures larger context, it can introduce noise and blur finer details [1, 29]. To address this, we adopt a refined contrast operation, inspired by [1], to dynamically enhance high-variance regions and suppress low-contrast noise. This contrast adjustment is defined as: + +$$ +\operatorname {E n h a n c e} (A) = (A - \mu (A)) \zeta (\sigma (A)) + \mu (A), \tag {6} +$$ + +where $\mu(A)$ and $\sigma(A)$ represent the mean and standard deviation operations, respectively, and $\zeta$ is an adaptive contrast operator. This technique effectively reduces noise, ensuring critical details are preserved during sketch generation. + +Building on this, we incorporate the stroke-based refinement concept from SDEdit [27]. Starting with a sketch initialized by stroke exchange, we denoise the image using classifier-free guidance (CFG) [6, 14]. During each denoising step, we utilize two parallel forward passes in the network. The first pass employs a cross-image attention layer to capture the stroke characteristics and abstraction level of the reference sketch, generating $z^{ske}$ : $\epsilon_{stroke}^{\times} = \epsilon_{\theta}^{\times}(z_{t}^{ske})$ , while simultaneously retaining the semantic context extracted from the content image's descriptive prompt: $\epsilon_{text}^{\times} = \epsilon_{\theta}^{\times}(z_{t}^{text})$ . The second pass applies regular self-attention to enhance the sketch's structural integrity: $\epsilon^{self} = \epsilon_{\theta}^{self}(z_{t}^{ske})$ . + +Following the CFG scheme, the predicted noise $\epsilon^t$ is computed as: + +$$ +\epsilon^ {t} = \epsilon^ {s e l f} + \beta_ {s g} (\epsilon_ {s t r o k e} ^ {\times} - \epsilon^ {s e l f}) + \beta_ {t e x t} (\epsilon_ {t e x t} ^ {\times} - \epsilon^ {s e l f}), (7) +$$ + +where $\beta_{sg}$ is the stroke guidance scale, and $\beta_{text}$ is the weight for the content text context. + +# 4. Experimental Results + +Datasets. Our experiments utilize three datasets: FS2K [9], the Anime dataset [19], and our newly created Stroke2Sketch-dataset. The FS2K dataset includes 5,140 facial image-sketch pairs. We randomly selected 1,000 pairs for testing, using color images as content images and choosing one sketch from the test set as a reference style. A similar selection strategy was used for the Anime dataset [19]. + +The Stroke2Sketch-dataset includes 50 content images from diverse categories and 20 distinct sketch styles, including single-line sketches, ink sketches, line art, and realistic sketches. Further details are provided in the Appendix. + +Metrics. While traditional metrics like ArtFID [43], LPIPS [53], and FID [13] are standard for style transfer, they struggle to capture the semantic sparsity and high-level abstraction unique to sketch generation. Consistent with prior work [38], we prioritize user perception to better reflect artistic and semantic quality in sketches. + +# 4.1. Qualitative Comparison + +We evaluate our proposed method through a comparison with eight state-of-the-art methods, including three training-based style transfer methods (Ref2sketch [2], Semi-ref2sketch [33], and CSGO [46]), and five training-free style transfer methods (IP-Adapter [49], InstantStyle [41], InstantStyle-plus [42], RB-Modulation [32], and StyleID [7]). Each of these methods takes a reference sketch as input. + +Fig. 8 illustrates the qualitative results across a variety of content images and reference sketches. In the first column, we display the content images, and the second column + +![](images/42bf7bfd95b139362e9eb337143bd1d3b8463e0598966a3b9849649cc8fd2cd0.jpg) +Figure 8. Qualitative comparison of sketch generation across training-based (4th-6th columns) and training-free baselines (7th-11th columns) using various reference sketches. Our method (3rd column) demonstrates superior adaptability to different reference styles, maintaining both stroke fidelity and semantic consistency across a range of content and reference types. + +![](images/8799ed8266de01dfdbe80d9528ba2c4541a6962442c4c13ac664c848a9df8f3d.jpg) +Figure 9. (a) Content image and manual stroke stylization by CLIPascene [39] with Adobe Illustrator; (b) Our method's automatically generated sketches based on reference stroke styles; (c) Appearance transfer results using similar references. + +shows the reference sketches. Notably, the fourth row's reference sketch is of a type seen in the Semi-ref2sketch training data, while the other three reference sketches represent styles outside the training data for Semi-ref2sketch. For methods requiring prompt input, we uniformly set the prompt to "sketch photo," while other parameters follow each method's default configuration. Our approach demonstrates robustness across diverse reference styles, accurately preserving both the stroke style and high-level semantic details. + +Comparison with vector sketch generation: Fig. 9(a) compares our method with the vectorized sketch generation approach of CLIPascene [39]. The content image is shown alongside vectorized sketches produced by CLIPascene, which were further manually refined with brush strokes in Adobe Illustrator. While CLIPascene can generate abstract + +![](images/4da0f4726b06cf026022e03711cfa9049023b28b1318500d88bac7199d7b8b07.jpg) +Figure 10. Stroke2Sketch's color sketch generation preserving reference styles and stroke characteristics + +sketches, it requires post-processing to achieve consistent stroke styling, whereas our method (Fig. 9(b)) automatically produces sketches with stylistically aligned strokes based on reference attributes. + +Comparison with appearance transfer methods: Fig. 9(c) shows results from appearance transfer method of CIA [1] applied to the same reference sketches. Although CIA excel in transferring visual features based on semantic similarity and object category, they focus on appearance rather than stroke style, making them less suitable for our sketch generation goals. + +Non-grayscale Sketch Generation. Building on its grayscale performance, Stroke2Sketch maintains reference stroke patterns and artistic styles in color outputs (Fig. 10). + +# 4.2.Quantitative Comparison + +We quantitatively evaluate our method against several state-of-the-art sketch extraction techniques, including both training-based (Ref2Sketch [2], Semi-Ref2Sketch [33], and Informative-drawing [5]) and training-free style transfer methods (IP-Adapter [49], InstantStyle [41], InstantStyle-plus [42], StyleID [7]). Table 1 presents the results on the Stroke2Sketch-dataset, showing that our method achieves the lowest ArtFID and FID scores, indicating superior performance in both style alignment and content preservation. + +
MetricOursRef2sketchSemi-refInfor-drawingIP-AdapterInstantStyleInstantStyle+StyleID
ArtFID ↓32.45545.29233.24234.21433.45732.53237.65635.727
LPIPS ↓0.53150.69820.53060.60370.66340.54320.65320.5426
FID ↓22.43534.65024.35925.03524.06823.94026.63225.658
+ +![](images/61f46cafb46bca11340ace69c7c4271bde8205f5d8e3ccd0db892eb950d8be44.jpg) +Figure 11. User study results. Our proposed method received the highest preference. + +These metrics highlight our method's effectiveness in producing high-quality sketches with strong semantic and stylistic fidelity. + +# 4.3. User Study + +We randomly selected 15 content images and 15 reference sketches from Stroke2Sketch-dataset, creating 225 image pairs. From these, we sampled 20 pairs to generate stylized images using four methods. The results were displayed side-by-side in random order, and participants were asked to choose their favorite based on three criteria: content extraction, stroke stylization, and overall preference. We collected 2,000 votes for each criterion from 100 users and presented the results as a bar chart. As shown in Fig. 11, our method received the highest preference, outperforming both training-based and training-free baselines, demonstrating its effectiveness in handling diverse stroke styles and abstract artistic effects. + +# 4.4. Ablation Study + +To validate our method's components, we performed an ablation study on the Stroke2Sketch-dataset. As shown in Table 2, removing any of the Directive Attention Module (DAM), Semantic Preservation Module (SPM), or Stroke Details Propagation Enhancement (SDPE) led to decreased performance across ArtFID, FID, and LPIPS metrics, confirming the contribution of each component. Specifically, without DAM, ArtFID increased from 32.45 to 38.67 and FID rose to 26.53, indicating weaker style-content alignment and content leakage in sketches, as observed in Fig. 12. The absence of SPM resulted in ArtFID rising to 36.89 and FID to 30.47, with sketches losing semantic coherence and structural integrity, such as poorly defined object outlines. + +Table 1. Quantitative comparison on Stroke2Sketch-dataset with training-based (3rd-5th columns) and training-free baselines (6th-9th columns). + +
ConfigurationArtFIDFIDLPIPS
A: Ours32.4522.430.530
B: - DAM (Sec. 3.3)38.6726.530.672
C: - SPM (Sec. 3.4)36.8930.470.637
D: - SDPE (Sec. 3.5)40.5332.440.598
+ +Table 2. Ablation study of different variants of our method. + +![](images/c46d58774955ce8de6a0f6ce3a4421bc7f9fc8feb70d07603acda03c6ca15335.jpg) +Figure 12. Ablative qualitative comparison of different variants of our method. + +Most notably, removing SDPE caused the sharpest decline, with ArtFID reaching 40.53 and sketches exhibiting excessive noise and loss of fine details, as evidenced by cluttered textures in Fig. 12. In contrast, the full method (Configuration A) achieved optimal scores of 32.45 (ArtFID), 22.43 (FID), and 0.530 (LPIPS), producing high-quality, reference-aligned sketches with precise details. These results highlight the critical roles of DAM, SPM, and SDPE in enhancing sketch quality. + +# 5. Conclusion + +Our training-free method for content-to-sketch generation aligns stroke attributes and semantic texture sparsity, mimicking artistic subject extraction to produce high-fidelity sketches. Leveraging pretrained models, it achieves state-of-the-art performance in quantitative metrics and user evaluations. Limitations arise with overly simplistic or complex reference sketches (see appendix for failure cases). Future work could explore decoupling semantic information from stroke attributes to improve adaptability and sketch quality. Acknowledgment. This work is supported by the Guangdong Natural Science Funds for Distinguished Young Scholars (Grant 2023B1515020097), the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No.: AISG3-GV-2023-011), the Singapore Ministry of Education AcRF Tier 1 Grant (Grant No.: MSS25C004), and the Lee Kong Chian Fellowships. + +# References + +[1] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer. In ACM SIGGRAPH 2024 Conference Papers, pages 1–12, 2024. 3, 6, 7 +[2] Amirsaman Ashtari, Chang Wook Seo, Cholmin Kang, Sihun Cha, and Junyong Noh. Reference based sketch extraction via attention mechanism. ACM Trans. Graph., 41(6), 2022. 1, 3, 6, 7 +[3] John Canny. A computational approach to edge detection. IEEE Transactions on pattern analysis and machine intelligence, (6):679-698, 1986. 3 +[4] Nan Cao, Xin Yan, Yang Shi, and Chaoran Chen. Ai-sketcher: a deep generative model for producing high-quality sketches. In Proceedings of the AAAI conference on artificial intelligence, pages 2564-2571, 2019. 3 +[5] Caroline Chan, Frédo Durand, and Phillip Isola. Learning to generate line drawings that convey geometry and semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7915-7925, 2022. 1, 3, 7 +[6] Hyungjin Chung, Jeongsol Kim, Geon Yeong Park, Hyelin Nam, and Jong Chul Ye. Cfg++: Manifold-constrained classifier free guidance for diffusion models. 2024. 6 +[7] Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8795-8805, 2024. 6, 7 +[8] Ayan Das, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Beziersketch: A generative model for scalable vector sketches. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI 16, pages 632-647. Springer, 2020. 3 +[9] Deng-Ping Fan, Ziling Huang, Peng Zheng, Hong Liu, Xuebin Qin, and Luc Van Gool. Facial-sketch synthesis: a new challenge. Machine Intelligence Research, 19(4):257-287, 2022. 6 +[10] Kevin Frans, Lisa Soros, and Olaf Witkowski. Clipdraw: Exploring text-to-drawing synthesis through language-image encoders. Advances in Neural Information Processing Systems, 35:5207-5218, 2022. 3 +[11] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3 +[12] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4775–4785, 2024. 3 +[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 + +[14] Jonathan Ho. Classifier-free diffusion guidance. ArXiv, abs/2207.12598, 2022. 6 +[15] Jjin Hu, Ke Li, Yonggang Qi, and Yi-Zhe Song. Scale-adaptive diffusion model for complex sketch synthesis. In The Twelfth International Conference on Learning Representations, 2024. 3 +[16] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly DDPM noise space: Inversion and manipulations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12469-12478, 2024. 3 +[17] Jaeseok Jeong, Junho Kim, Yunjoy Choi, Gayoung Lee, and Youngjung Uh. Visual style prompting with swapping self-attention. arXiv preprint arXiv:2402.12974, 2024. 3 +[18] Yutao Jiang, Yang Zhou, Yuan Liang, Wenxi Liu, Jianbo Jiao, Yuhui Quan, and Shengfeng He. Diffuse3d: Wide-angle 3d photography via bilateral diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8998-9008, 2023. 1 +[19] Tae Bum Kang. Anime sketch colorization pair, 2018. Accessed: 2024-05-18. 6 +[20] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 3, 5 +[21] Xinzhe Li, Jiahui Zhan, Shengfeng He, Yangyang Xu, Junyu Dong, Huaidong Zhang, and Yong Du. Personamagic: Stage-regulated high-fidelity face customization with tandem equilibrium. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4995-5003, 2025. 1 +[22] Chenxi Liu, Enrique Rosales, and Alla Sheffer. Strokeaggregator: Consolidating raw sketches into artist-intended curve drawings. ACM Transactions on Graphics (TOG), 37(4):1-15, 2018. 3 +[23] Chenxi Liu, Toshiki Aoki, Mikhail Bessmeltsev, and Alla Sheffer. Stripmaker: Perception-driven learned vector sketch consolidation. ACM Transactions on Graphics (TOG), 42(4): 1-15, 2023. 3 +[24] Difan Liu, Matthew Fisher, Aaron Hertzmann, and Evangelos Kalogerakis. Neural strokes: Stylized line drawing of 3d shapes. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 14184-14193. IEEE, 2021. 1 +[25] Haofeng Liu, Chenshu Xu, Yifei Yang, Lihua Zeng, and Shengfeng He. Drag your noise: Interactive point-based editing via diffusion semantic propagation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6743-6752, 2024. 1 +[26] Xiao-Chang Liu, Yu-Chen Wu, and Peter Hall. Painterly style transfer with learned brush strokes. IEEE Transactions on Visualization and Computer Graphics, 30(9):6309-6320, 2024. 1 +[27] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 6 + +[28] Reiichiro Nakano. Neural painters: A learned differentiable constraint for generating brushstroke paintings. arXiv preprint arXiv:1904.08410, 2019. 1 +[29] Nadav Orzech, Yotam Nitzan, Ulysse Mizrahi, Dov Danon, and Amit H Bermano. Masked extended attention for zero-shot virtual try-on in the wild. arXiv preprint arXiv:2406.15331, 2024. 6 +[30] Xavier Soria Poma, Edgar Riba, and Angel Sappa. Dense extreme inception network: Towards a robust cnn model for edge detection. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1923-1932, 2020. 3 +[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3 +[32] L Rout, Y Chen, N Ruiz, A Kumar, C Caramanis, S Shakkottai, and W Chu. Rb-modulation: Training-free personalization of diffusion models using stochastic optimal control. 2024. 2, 3, 6 +[33] Chang Wook Seo, Amirsaman Ashtari, and Junyong Noh. Semi-supervised reference-based sketch extraction using a contrastive learning framework. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 1, 3, 6, 7 +[34] Edgar Simo-Serra, Satoshi Iizuka, Kazuma Sasaki, and Hiroshi Ishikawa. Learning to simplify: fully convolutional networks for rough sketch cleanup. ACM Transactions on Graphics (TOG), 35(4), 2016. 1 +[35] Gowthami Somepalli, Anubhav Gupta, Kamal Gupta, Shramay Palta, Micah Goldblum, Jonas Geiping, Abhinav Shrivastava, and Tom Goldstein. Measuring style similarity in diffusion models. arXiv preprint arXiv:2404.01292, 2024. 3 +[36] Xavier Soria, Yachuan Li, Mohammad Rouhani, and Angel D Sappa. Tiny and efficient model for the edge detection generalization. In CVPR, pages 1364-1373, 2023. 3, 4 +[37] Zhuo Su, Wenzhe Liu, Zitong Yu, Dewen Hu, Qing Liao, Qi Tian, Matti Pietikainen, and Li Liu. Pixel difference networks for efficient edge detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5117-5127, 2021. 3 +[38] Yael Vinker, Ehsan Pajouheshgar, Jessica Y Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, and Ariel Shamir. Clipasso: Semantically-aware object sketching. ACM Transactions on Graphics (TOG), 41(4): 1-11, 2022. 1, 3, 6 +[39] Yael Vinker, Yuval Alaluf, Daniel Cohen-Or, and Ariel Shamir. Clipscene: Scene sketching with different types and levels of abstraction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4146–4156, 2023. 7 +[40] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. p+: Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522, 2023. 3 +[41] Haofan Wang, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv preprint arXiv:2404.02733, 2024. 1, 2, 3, 6, 7 + +[42] Haofan Wang, Peng Xing, Renyuan Huang, Hao Ai, Qixun Wang, and Xu Bai. Instantstyle-plus: Style transfer with content-preserving in text-to-image generation. arXiv preprint arXiv:2407.00788, 2024. 3, 6, 7 +[43] Matthias Wright and Björn Ommer. Artfid: Quantitative evaluation of neural style transfer. In DAGM German Conference on Pattern Recognition, pages 560-576. Springer, 2022. 6 +[44] Zongwei Wu, Liangyu Chai, Nanxuan Zhao, Bailin Deng, Yongtuo Liu, Qiang Wen, Junle Wang, and Shengfeng He. Make your own sprites: Aliasing-aware and cell-controllable pixelization. ACM Transactions on Graphics (TOG), 41(6): 1-16, 2022. 1 +[45] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In Proceedings of the IEEE international conference on computer vision, pages 1395-1403, 2015. 3 +[46] Peng Xing, Haofan Wang, Yanpeng Sun, Qixun Wang, Xu Bai, Hao Ai, Renyuan Huang, and Zechao Li. Csgo: Content-style composition in text-to-image generation. arXiv preprint arXiv:2408.16766, 2024. 6 +[47] Chenshu Xu, Yangyang Xu, Huaidong Zhang, Xuemiao Xu, and Shengfeng He. Dreamanine: Learning style-identity textual disentanglement for anime and beyond. IEEE Transactions on Visualization and Computer Graphics, 2024. 1 +[48] Rui Yang, Xiaojun Wu, and Shengfeng He. Mixsa: Training-free reference-based sketch extraction via mixture-of-self-attention. IEEE Transactions on Visualization and Computer Graphics, 2024. 3 +[49] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 6, 7 +[50] Yuyang Yu, Bangzhen Liu, Chenxi Zheng, Xuemiao Xu, Huaidong Zhang, and Shengfeng He. Beyond textual constraints: Learning novel diffusion conditions with fewer examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7109-7118, 2024. 1 +[51] Sicong Zang, Shikui Tu, and Lei Xu. Self-organizing a latent hierarchy of sketch patterns for controllable sketch synthesis. IEEE Transactions on Neural Networks and Learning Systems, 2023. 3 +[52] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3 +[53] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6 +[54] Caixia Zhou, Yaping Huang, Mengyang Pu, Qingji Guan, Ruoxi Deng, and Haibin Ling. Muge: Multiple granularity edge detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25952-25962, 2024. 3 \ No newline at end of file diff --git a/2025/Stroke2Sketch_ Harnessing Stroke Attributes for Training-Free Sketch Generation/images.zip b/2025/Stroke2Sketch_ Harnessing Stroke Attributes for Training-Free Sketch Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..48c2bb74c211c0008fe7e6c105048d1a4e090bb0 --- /dev/null +++ b/2025/Stroke2Sketch_ Harnessing Stroke Attributes for Training-Free Sketch Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:017572180a4230cdef7aa8c98c86166757e7335c569d61cacb561b6aafa303d4 +size 676693 diff --git a/2025/Stroke2Sketch_ Harnessing Stroke Attributes for Training-Free Sketch Generation/layout.json b/2025/Stroke2Sketch_ Harnessing Stroke Attributes for Training-Free Sketch Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..130184e860990611726e52d707cc34e926fd4953 --- /dev/null +++ b/2025/Stroke2Sketch_ Harnessing Stroke Attributes for Training-Free Sketch Generation/layout.json @@ -0,0 +1,8768 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 62, + 111, + 547, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 111, + 547, + 128 + ], + "spans": [ + { + "bbox": [ + 62, + 111, + 547, + 128 + ], + "type": "text", + "content": "Stroke2Sketch: Harnessing Stroke Attributes for Training-Free Sketch Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 477, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 477, + 164 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 477, + 164 + ], + "type": "text", + "content": "Rui Yang" + }, + { + "bbox": [ + 132, + 149, + 477, + 164 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 132, + 149, + 477, + 164 + ], + "type": "text", + "content": ", Huining Li" + }, + { + "bbox": [ + 132, + 149, + 477, + 164 + ], + "type": "inline_equation", + "content": "^{4,5}" + }, + { + "bbox": [ + 132, + 149, + 477, + 164 + ], + "type": "text", + "content": ", Yiyi Long" + }, + { + "bbox": [ + 132, + 149, + 477, + 164 + ], + "type": "inline_equation", + "content": "^{2,5}" + }, + { + "bbox": [ + 132, + 149, + 477, + 164 + ], + "type": "text", + "content": ", Xiaojun Wu" + }, + { + "bbox": [ + 132, + 149, + 477, + 164 + ], + "type": "inline_equation", + "content": "^{3,*}" + }, + { + "bbox": [ + 132, + 149, + 477, + 164 + ], + "type": "text", + "content": ", Shengfeng He" + }, + { + "bbox": [ + 132, + 149, + 477, + 164 + ], + "type": "inline_equation", + "content": "^{5,*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 87, + 163, + 523, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 163, + 523, + 192 + ], + "spans": [ + { + "bbox": [ + 87, + 163, + 523, + 192 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 87, + 163, + 523, + 192 + ], + "type": "text", + "content": "Huaqiao University " + }, + { + "bbox": [ + 87, + 163, + 523, + 192 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 87, + 163, + 523, + 192 + ], + "type": "text", + "content": "South China University of Technology " + }, + { + "bbox": [ + 87, + 163, + 523, + 192 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 87, + 163, + 523, + 192 + ], + "type": "text", + "content": "Shaanxi Normal University " + }, + { + "bbox": [ + 87, + 163, + 523, + 192 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 87, + 163, + 523, + 192 + ], + "type": "text", + "content": "Beijing University of Aeronautics and Astronautics " + }, + { + "bbox": [ + 87, + 163, + 523, + 192 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 87, + 163, + 523, + 192 + ], + "type": "text", + "content": "Singapore Management University" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 76, + 198, + 533, + 381 + ], + "blocks": [ + { + "bbox": [ + 76, + 198, + 533, + 381 + ], + "lines": [ + { + "bbox": [ + 76, + 198, + 533, + 381 + ], + "spans": [ + { + "bbox": [ + 76, + 198, + 533, + 381 + ], + "type": "image", + "image_path": "537aacecb15601e21cbc49becd56d0f3bac6c79f6805b55df94b65f75014510a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 384, + 548, + 417 + ], + "lines": [ + { + "bbox": [ + 61, + 384, + 548, + 417 + ], + "spans": [ + { + "bbox": [ + 61, + 384, + 548, + 417 + ], + "type": "text", + "content": "Figure 1. We propose Stroke2Sketch, a framework that accurately transfers stroke attributes from a reference sketch to a content image while preserving structure and style fidelity. The top row shows reference sketches, the leftmost column displays content images, and the central and right columns illustrate our method's precise content preservation and expressive stroke transfer." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 156, + 425, + 202, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 425, + 202, + 436 + ], + "spans": [ + { + "bbox": [ + 156, + 425, + 202, + 436 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 60, + 453, + 297, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 453, + 297, + 662 + ], + "spans": [ + { + "bbox": [ + 60, + 453, + 297, + 662 + ], + "type": "text", + "content": "Generating sketches guided by reference styles requires precise transfer of stroke attributes, such as line thickness, deformation, and texture sparsity, while preserving semantic structure and content fidelity. To this end, we propose Stroke2Sketch, a novel training-free framework that introduces cross-image stroke attention, a mechanism embedded within self-attention layers to establish fine-grained semantic correspondences and enable accurate stroke attribute transfer. This allows our method to adaptively integrate reference stroke characteristics into content images while maintaining structural integrity. Additionally, we develop adaptive contrast enhancement and semantic-focused attention to reinforce content preservation and foreground emphasis. Stroke2Sketch effectively synthesizes stylistically faithful sketches that closely resemble handcrafted results, outperforming existing methods in expressive stroke control and semantic coherence. Codes are available at https://github.com/rane7/Stroke2Sketch." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 425, + 392, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 425, + 392, + 436 + ], + "spans": [ + { + "bbox": [ + 314, + 425, + 392, + 436 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 446, + 548, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 446, + 548, + 597 + ], + "spans": [ + { + "bbox": [ + 312, + 446, + 548, + 597 + ], + "type": "text", + "content": "Generating stylized sketches from content images using reference stroke patterns presents a key challenge at the intersection of artistic rendering and semantic-aware synthesis. Traditional sketch algorithms [28, 34] generate procedural line drawings, while vector-based methods [24, 26, 38] produce clean parametric strokes. However, these methods lack the adaptability to transfer diverse artistic styles from exemplar sketches due to their rigid, data-agnostic architectures. Unlike human artists, who strategically vary stroke attributes such as thickness, curvature, and density to emphasize key semantic features while maintaining content fidelity, existing approaches struggle to capture this nuanced interplay between stroke semantics and structure." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 600, + 548, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 600, + 548, + 704 + ], + "spans": [ + { + "bbox": [ + 312, + 600, + 548, + 704 + ], + "type": "text", + "content": "Recent learning-based methods [2, 5, 33] attempt to address this limitation by training on clustered sketch styles, yet as shown in Fig. 2(a-b), they fail to generalize to unseen stroke patterns due to catastrophic forgetting. Meanwhile, diffusion-based stylization techniques [21, 41, 47, 49, 50] excel in texture transfer but struggle with structural integrity due to content leakage in cross-attention layers (Fig. 2(c-d)). While new conditioning mechanisms [18, 25, 44] attempt to enhance structural control, they often introduce style dis" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 71, + 13, + 116, + 44 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 13, + 116, + 44 + ], + "spans": [ + { + "bbox": [ + 71, + 13, + 116, + 44 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 150, + 11, + 489, + 45 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 11, + 489, + 45 + ], + "spans": [ + { + "bbox": [ + 150, + 11, + 489, + 45 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 685, + 295, + 704 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 685, + 295, + 704 + ], + "spans": [ + { + "bbox": [ + 62, + 685, + 295, + 704 + ], + "type": "text", + "content": "*Corresponding authors: Xiaojun Wu (xjwu@snnu.edu.cn) and Shengfeng He (shengfenghe@smu.edu.sg)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "text", + "content": "16545" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 79, + 547, + 261 + ], + "blocks": [ + { + "bbox": [ + 63, + 79, + 547, + 261 + ], + "lines": [ + { + "bbox": [ + 63, + 79, + 547, + 261 + ], + "spans": [ + { + "bbox": [ + 63, + 79, + 547, + 261 + ], + "type": "image", + "image_path": "2779d5cc0e7051ec1056b65b671bcf2ad5e1498cd4f647273e68de9fe6dcd703.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 270, + 548, + 314 + ], + "lines": [ + { + "bbox": [ + 61, + 270, + 548, + 314 + ], + "spans": [ + { + "bbox": [ + 61, + 270, + 548, + 314 + ], + "type": "text", + "content": "Figure 2. (a) and (b) show results from training-based methods (Ref2Sketch and Semi-Ref2Sketch), which struggle with unseen styles, leading to poor content alignment. (c) and (d) illustrate IP-Adapter and InstantStyle results, retaining style but lacking content alignment. (e) to (g) show ControlNet-based methods, preserving content but failing to match reference styles. (h) shows our method's output, achieving superior content preservation and style alignment without using ControlNet. Detailed prompts are shown in the supplementary materials." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 323, + 296, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 323, + 296, + 393 + ], + "spans": [ + { + "bbox": [ + 61, + 323, + 296, + 393 + ], + "type": "text", + "content": "tortion or require dense user inputs. Hybrid approaches like ControlNet [52] enforce structural priors but sacrifice stylistic flexibility, leading to overly rigid outputs (Fig. 2(e)). Progressive stroke-based methods [32] introduce semantic incoherence by applying uniform strokes across the image (Fig. 2(g))." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 396, + 296, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 396, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 61, + 396, + 296, + 559 + ], + "type": "text", + "content": "We identify three fundamental challenges in reference-based sketch synthesis:(i) Semantic-aware stroke transfer. Effective style adaptation requires precise mapping of reference stroke attributes (e.g., tapered lines, cross-hatching) to semantically relevant content regions. (ii) Foreground prioritization. Artists naturally emphasize foreground elements using varied stroke density and complexity while simplifying backgrounds, yet existing methods apply uniform stylization, disrupting this compositional balance[32]. (iii) Content-style equilibrium. Since sketches encode content through linework, balancing structural preservation with style transfer is critical. Techniques such as CLIP-space style subtraction [41] fail to maintain this balance, as even minor content leakage distorts key edges (Fig. 2(c-d))." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 562, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 562, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 61, + 562, + 296, + 689 + ], + "type": "text", + "content": "To address these challenges, we propose Stroke2Sketch, a framework that enables precise stroke attribute transfer while maintaining content fidelity. Our key insight is that stroke properties, like line thickness, curvature, and texture sparsity, are inherently encoded within the self-attention and cross-attention relationships of pretrained diffusion models. By dynamically aligning these attention patterns between content and reference features, we achieve effective style transfer without structural degradation. Stroke2Sketch integrates three novel components tailored to each identified challenge:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 693, + 295, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 693, + 295, + 704 + ], + "spans": [ + { + "bbox": [ + 61, + 693, + 295, + 704 + ], + "type": "text", + "content": "(1) Cross-image stroke attention tackles the challenge of" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 323, + 548, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 548, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 548, + 405 + ], + "type": "text", + "content": "semantic-aware stroke transfer by facilitating stroke attribute exchange through key-value swapping in diffusion layers. Instead of directly blending features [41], we leverage attention blocks to transplant stroke characteristics onto content structures, preventing the entanglement of style and geometry. This approach ensures accurate stroke mapping without disrupting semantic coherence, as shown in Fig. 2(h)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 407, + 548, + 608 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 313, + 407, + 548, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 548, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 548, + 501 + ], + "type": "text", + "content": "(2) Directive Attention Module ensures that stroke transfer remains compositionally balanced. Background textures often introduce conflicting patterns that dilute the intended style. We mitigate this by clustering self-attention maps to isolate foreground objects, then restricting cross-image attention to these regions. This mimics how artists prioritize focal elements while simplifying less critical areas, enhancing both style consistency and perceptual quality." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 503, + 548, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 548, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 548, + 608 + ], + "type": "text", + "content": "(3) Semantic Preservation Module addresses content-style equilibrium by injecting content contours as positional queries during early denoising. This hybridizes the precision of edge detectors with the flexibility of text-driven generation, allowing structural guidance without rigid constraints. Unlike ControlNet, which enforces strict boundaries, our approach treats edges as \"soft constraints\" that evolve into stylized strokes, preserving both structure and artistic abstraction." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 611, + 548, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 548, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 548, + 704 + ], + "type": "text", + "content": "As validated in Fig. 2(h) and Fig. 6, Stroke2Sketch achieves state-of-the-art performance across diverse sketch styles. Experimental results show that our method outperforms adapter-based and ControlNet methods in both style alignment (87% user preference) and content preservation (92% accuracy in line correspondence tests). Importantly, it achieves these results without dataset-specific training or architectural modifications, demonstrating that pretrained" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "text", + "content": "16546" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 81, + 297, + 105 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 81, + 297, + 105 + ], + "spans": [ + { + "bbox": [ + 61, + 81, + 297, + 105 + ], + "type": "text", + "content": "diffusion models can master sketch stylization when guided by principled feature interactions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 62, + 116, + 148, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 116, + 148, + 129 + ], + "spans": [ + { + "bbox": [ + 62, + 116, + 148, + 129 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 136, + 191, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 136, + 191, + 148 + ], + "spans": [ + { + "bbox": [ + 62, + 136, + 191, + 148 + ], + "type": "text", + "content": "2.1. Photo-Sketch Synthesis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 153, + 297, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 153, + 297, + 268 + ], + "spans": [ + { + "bbox": [ + 61, + 153, + 297, + 268 + ], + "type": "text", + "content": "Generating a sketch from a content image based on a reference style parallels edge detection, as both tasks emphasize prominent visual features. Edge detection methods [3, 30, 36, 37, 45, 54], which detect sharp changes in color or brightness, form the foundation of sketch extraction but are limited to a single style and often produce artifacts, like scattered dots or broken lines. High-quality sketch generation requires more than edge detection; it demands line style, texture sparsity, and semantic abstraction to achieve artist-level results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 270, + 296, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 270, + 296, + 363 + ], + "spans": [ + { + "bbox": [ + 61, + 270, + 296, + 363 + ], + "type": "text", + "content": "Learning-based approaches have improved sketch realism by enhancing boundary detection and rendering distinct line styles. For example, Chan et al. [5] incorporated depth and semantic information for better sketch quality, while Ref2Sketch [2] and Semi-Ref2Sketch [33] leverage paired and semi-supervised training for stylized sketch extraction. However, these methods rely on large sketch datasets, which are challenging to obtain, limiting model robustness." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 364, + 296, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 364, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 61, + 364, + 296, + 468 + ], + "type": "text", + "content": "Another research area, stroke-based rendering, focuses on manipulating strokes and contours for sketching. Methods like CLIPDraw [10] and CLIPasso [38] employ Bezier curves to create abstract sketches with high-level semantic simplification. StrokeAggregator [22] and StripMaker [23] refine vector sketches, achieving quality comparable to artist drawings. Other methods include semantic concept-to-sketch methods [4, 8, 15, 51]. However, these methods assume uniform style, while real-world sketches vary widely." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 468, + 296, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 468, + 296, + 504 + ], + "spans": [ + { + "bbox": [ + 62, + 468, + 296, + 504 + ], + "type": "text", + "content": "Building on these insights, our method combines stroke attributes, semantic abstraction, and expression to better align generated sketches with diverse reference styles." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 512, + 220, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 512, + 220, + 525 + ], + "spans": [ + { + "bbox": [ + 62, + 512, + 220, + 525 + ], + "type": "text", + "content": "2.2. High-Semantic Style Transfer" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 529, + 297, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 529, + 297, + 622 + ], + "spans": [ + { + "bbox": [ + 61, + 529, + 297, + 622 + ], + "type": "text", + "content": "Generating sketches that adhere to both content and reference style is a specialized style transfer task. Advances in diffusion models have propelled style transfer through self- and cross-attention mechanisms, which preserve spatial layout and stylized content. Techniques such as Prompt-to-Prompt [11] and " + }, + { + "bbox": [ + 61, + 529, + 297, + 622 + ], + "type": "inline_equation", + "content": "\\mathrm{P + }" + }, + { + "bbox": [ + 61, + 529, + 297, + 622 + ], + "type": "text", + "content": " [40] show how attention mechanisms can maintain structural coherence while enabling flexible style and semantic control." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 623, + 297, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 623, + 297, + 704 + ], + "spans": [ + { + "bbox": [ + 61, + 623, + 297, + 704 + ], + "type": "text", + "content": "Cross-Image Attention (CIA) [1] and methods like Swapping Self-Attention [17] and StyleAligned [12] reveal that style features are best retained during upsampling stages, even though content leakage may occur in bottleneck phases. IP-Adapter achieves style transfer via dual cross-attention with text and image, though it can weaken text control and lead to leakage." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 81, + 548, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 81, + 548, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 81, + 548, + 198 + ], + "type": "text", + "content": "InstantStyle [41] addresses leakage by subtracting features in the same feature space but often requires external constraints like ControlNet [52] for image-to-image generation, which can dilute style fidelity. Further developments like InstantStyle-plus [42] and RB-Modulation [32] incorporate CSD [35] to better align styles during generation, but limitations remain in sketch generation. Prompt-to-style transfer, however, demonstrates pretrained models' strong semantic alignment capabilities, providing valuable priors for consistent, high-level sketch synthesis." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 207, + 400, + 219 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 207, + 400, + 219 + ], + "spans": [ + { + "bbox": [ + 313, + 207, + 400, + 219 + ], + "type": "text", + "content": "3. Stroke2Sketch" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 225, + 548, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 225, + 548, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 225, + 548, + 403 + ], + "type": "text", + "content": "Given a content image " + }, + { + "bbox": [ + 313, + 225, + 548, + 403 + ], + "type": "inline_equation", + "content": "I^{cnt} \\in \\mathbb{R}^{H^{cnt} \\times W^{cnt} \\times 3}" + }, + { + "bbox": [ + 313, + 225, + 548, + 403 + ], + "type": "text", + "content": " and a reference sketch image " + }, + { + "bbox": [ + 313, + 225, + 548, + 403 + ], + "type": "inline_equation", + "content": "I^{ref} \\in \\mathbb{R}^{H^{ref} \\times W^{ref} \\times 3}" + }, + { + "bbox": [ + 313, + 225, + 548, + 403 + ], + "type": "text", + "content": ", our task is to generate a sketch " + }, + { + "bbox": [ + 313, + 225, + 548, + 403 + ], + "type": "inline_equation", + "content": "I^{ske}" + }, + { + "bbox": [ + 313, + 225, + 548, + 403 + ], + "type": "text", + "content": " that aligns with the content structure of " + }, + { + "bbox": [ + 313, + 225, + 548, + 403 + ], + "type": "inline_equation", + "content": "I^{cnt}" + }, + { + "bbox": [ + 313, + 225, + 548, + 403 + ], + "type": "text", + "content": " while adhering to the stroke style, texture sparsity, and high-level semantic abstraction of " + }, + { + "bbox": [ + 313, + 225, + 548, + 403 + ], + "type": "inline_equation", + "content": "I^{ref}" + }, + { + "bbox": [ + 313, + 225, + 548, + 403 + ], + "type": "text", + "content": ". Additionally, we aim to remove or retain background elements as needed to enhance the foreground object. In this work, we address the sketch generation task using a controllable text-to-image guidance approach. Specifically, we extract object prompts from the content image " + }, + { + "bbox": [ + 313, + 225, + 548, + 403 + ], + "type": "inline_equation", + "content": "I^{cnt}" + }, + { + "bbox": [ + 313, + 225, + 548, + 403 + ], + "type": "text", + "content": " using BLIP [20] and incorporate the stroke style and high-level abstraction from the reference sketch " + }, + { + "bbox": [ + 313, + 225, + 548, + 403 + ], + "type": "inline_equation", + "content": "I^{ref}" + }, + { + "bbox": [ + 313, + 225, + 548, + 403 + ], + "type": "text", + "content": " to re-render the final target sketch " + }, + { + "bbox": [ + 313, + 225, + 548, + 403 + ], + "type": "inline_equation", + "content": "I^{ske}" + }, + { + "bbox": [ + 313, + 225, + 548, + 403 + ], + "type": "text", + "content": ". To achieve this, we leverage a pre-trained text-to-image model. The overall network architecture is illustrated in Fig 3, with detailed explanations of each module provided below." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 408, + 397, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 397, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 397, + 419 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 426, + 548, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 548, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 548, + 600 + ], + "type": "text", + "content": "DDPM inversion [16] is a process in diffusion-based generative models that reverses denoising steps, enabling the reconstruction of latent representations from generated outputs. Compared to DDIM inversion method [48], DDPM inversion offers greater flexibility for editing tasks by producing noise maps that, while correlated across timesteps and not normally distributed, allow precise image reconstruction and meaningful edits such as color adjustments and structural shifts. A key advantage of DDPM inversion is its ability to maintain an image's structure while altering the conditioning input, such as a text prompt, to enable artifact-free edits that adapt semantics while preserving original details. This efficient method bypasses optimization processes and can enhance diffusion-based editing techniques by improving image fidelity and supporting diverse outputs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 600, + 548, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 548, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 548, + 704 + ], + "type": "text", + "content": "Stable Diffusion [31] incorporates this inversion process within a latent space, rather than a pixel space, increasing computational efficiency and expressiveness. The input image is encoded through a pretrained variational autoencoder (VAE) to produce a latent code " + }, + { + "bbox": [ + 313, + 600, + 548, + 704 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 600, + 548, + 704 + ], + "type": "text", + "content": ". Denoising then occurs within this latent space using a U-Net architecture that incorporates self-attention and cross-attention mechanisms. Self-attention blocks enhance image detail by calculating attention scores between projected query " + }, + { + "bbox": [ + 313, + 600, + 548, + 704 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 600, + 548, + 704 + ], + "type": "text", + "content": ", key " + }, + { + "bbox": [ + 313, + 600, + 548, + 704 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 600, + 548, + 704 + ], + "type": "text", + "content": ", and value" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "text", + "content": "16547" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 82, + 545, + 332 + ], + "blocks": [ + { + "bbox": [ + 65, + 82, + 545, + 332 + ], + "lines": [ + { + "bbox": [ + 65, + 82, + 545, + 332 + ], + "spans": [ + { + "bbox": [ + 65, + 82, + 545, + 332 + ], + "type": "image", + "image_path": "882d2df4e65ecfec500fba46b0e2dfd429bef22513cbc659af43661b6d93875a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 344, + 548, + 420 + ], + "lines": [ + { + "bbox": [ + 61, + 344, + 548, + 420 + ], + "spans": [ + { + "bbox": [ + 61, + 344, + 548, + 420 + ], + "type": "text", + "content": "Figure 3. The Stroke2Sketch architecture. The content image " + }, + { + "bbox": [ + 61, + 344, + 548, + 420 + ], + "type": "inline_equation", + "content": "I^{\\mathrm{cnt}}" + }, + { + "bbox": [ + 61, + 344, + 548, + 420 + ], + "type": "text", + "content": " undergoes contour detection, generating feature representations " + }, + { + "bbox": [ + 61, + 344, + 548, + 420 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{cont}}" + }, + { + "bbox": [ + 61, + 344, + 548, + 420 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 61, + 344, + 548, + 420 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{cnt}}" + }, + { + "bbox": [ + 61, + 344, + 548, + 420 + ], + "type": "text", + "content": ", while the reference sketch " + }, + { + "bbox": [ + 61, + 344, + 548, + 420 + ], + "type": "inline_equation", + "content": "I^{\\mathrm{ref}}" + }, + { + "bbox": [ + 61, + 344, + 548, + 420 + ], + "type": "text", + "content": " is inverted to produce " + }, + { + "bbox": [ + 61, + 344, + 548, + 420 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{ref}}" + }, + { + "bbox": [ + 61, + 344, + 548, + 420 + ], + "type": "text", + "content": ". The Directive Attention Module (DAM) aligns high-level semantic features between the content and reference features by emphasizing cross-image semantic correspondences " + }, + { + "bbox": [ + 61, + 344, + 548, + 420 + ], + "type": "inline_equation", + "content": "A_{n}" + }, + { + "bbox": [ + 61, + 344, + 548, + 420 + ], + "type": "text", + "content": ". Self-attention maps " + }, + { + "bbox": [ + 61, + 344, + 548, + 420 + ], + "type": "inline_equation", + "content": "F_{SA}" + }, + { + "bbox": [ + 61, + 344, + 548, + 420 + ], + "type": "text", + "content": " are aggregated and clustered to produce segmentation masks " + }, + { + "bbox": [ + 61, + 344, + 548, + 420 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 61, + 344, + 548, + 420 + ], + "type": "text", + "content": ", which help distinguish foreground from background regions. Cross-image Stroke Attention (CSA) transfers stroke attributes, and the Semantic Preservation Module (SPM) enforces semantic alignment and stroke fidelity in the generated sketch " + }, + { + "bbox": [ + 61, + 344, + 548, + 420 + ], + "type": "inline_equation", + "content": "\\hat{I}^{\\mathrm{ske}}" + }, + { + "bbox": [ + 61, + 344, + 548, + 420 + ], + "type": "text", + "content": " via loss " + }, + { + "bbox": [ + 61, + 344, + 548, + 420 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sem}}" + }, + { + "bbox": [ + 61, + 344, + 548, + 420 + ], + "type": "text", + "content": " and contour-based structural integration. Lastly, Stroke Detail Propagation Enhancement (SDPE) refines details, resulting in the final output sketch " + }, + { + "bbox": [ + 61, + 344, + 548, + 420 + ], + "type": "inline_equation", + "content": "I^{\\mathrm{ske}}" + }, + { + "bbox": [ + 61, + 344, + 548, + 420 + ], + "type": "text", + "content": ". SD represents the pre-trained diffusion model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 440, + 106, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 440, + 106, + 450 + ], + "spans": [ + { + "bbox": [ + 63, + 440, + 106, + 450 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 63, + 440, + 106, + 450 + ], + "type": "text", + "content": " vectors:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 98, + 458, + 296, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 458, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 98, + 458, + 296, + 485 + ], + "type": "interline_equation", + "content": "A = \\operatorname {s o f t m a x} \\left(\\frac {Q \\cdot K ^ {\\top}}{\\sqrt {d}}\\right), \\quad \\phi = A \\cdot V, \\tag {1}", + "image_path": "781356c9a15d735f71cf49b98b5c28717fb95d1d508f889b0747683d4050bbd2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 492, + 296, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 492, + 296, + 538 + ], + "spans": [ + { + "bbox": [ + 61, + 492, + 296, + 538 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 61, + 492, + 296, + 538 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 61, + 492, + 296, + 538 + ], + "type": "text", + "content": " is the attention map, " + }, + { + "bbox": [ + 61, + 492, + 296, + 538 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 61, + 492, + 296, + 538 + ], + "type": "text", + "content": " is the dimensionality, and " + }, + { + "bbox": [ + 61, + 492, + 296, + 538 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 61, + 492, + 296, + 538 + ], + "type": "text", + "content": " is the output of the self-attention block. Cross-attention blocks incorporate conditioning inputs (e.g., text prompts) to guide the generation process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 539, + 296, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 539, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 61, + 539, + 296, + 574 + ], + "type": "text", + "content": "Finally, the refined latent representation " + }, + { + "bbox": [ + 61, + 539, + 296, + 574 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 61, + 539, + 296, + 574 + ], + "type": "text", + "content": " is decoded back into an RGB image using the VAE decoder, yielding a high-quality output guided by the conditioning input." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 581, + 220, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 581, + 220, + 594 + ], + "spans": [ + { + "bbox": [ + 62, + 581, + 220, + 594 + ], + "type": "text", + "content": "3.2. Cross-image Stroke Attention" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 598, + 297, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 598, + 297, + 706 + ], + "spans": [ + { + "bbox": [ + 61, + 598, + 297, + 706 + ], + "type": "text", + "content": "As discussed in the Introduction, sketch generation as a specialized form of style transfer requires attention to local stroke attributes and consistent texture abstraction across different levels of semantic sparsity. Edge detection serves as the foundation for this task, with edges defining content contours as the most effective strategy for preserving structural information. We extract the contour " + }, + { + "bbox": [ + 61, + 598, + 297, + 706 + ], + "type": "inline_equation", + "content": "I^{cont}" + }, + { + "bbox": [ + 61, + 598, + 297, + 706 + ], + "type": "text", + "content": " from the content image using TEED [36] and obtain the inverted latents " + }, + { + "bbox": [ + 61, + 598, + 297, + 706 + ], + "type": "inline_equation", + "content": "z_{T}^{cnt}" + }, + { + "bbox": [ + 61, + 598, + 297, + 706 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 61, + 598, + 297, + 706 + ], + "type": "inline_equation", + "content": "z_{T}^{ref}" + }, + { + "bbox": [ + 61, + 598, + 297, + 706 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 61, + 598, + 297, + 706 + ], + "type": "inline_equation", + "content": "z_{T}^{cont}" + }, + { + "bbox": [ + 61, + 598, + 297, + 706 + ], + "type": "text", + "content": " for the content image, reference sketch, and" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 440, + 547, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 440, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 440, + 547, + 487 + ], + "type": "text", + "content": "contour image using DDPM inversion. During the denoising process, the latent " + }, + { + "bbox": [ + 313, + 440, + 547, + 487 + ], + "type": "inline_equation", + "content": "z_{T}^{cnt}" + }, + { + "bbox": [ + 313, + 440, + 547, + 487 + ], + "type": "text", + "content": " from the content image serves as the initial noise " + }, + { + "bbox": [ + 313, + 440, + 547, + 487 + ], + "type": "inline_equation", + "content": "z_{T}^{ske}" + }, + { + "bbox": [ + 313, + 440, + 547, + 487 + ], + "type": "text", + "content": " for sketch denoising. At each timestep " + }, + { + "bbox": [ + 313, + 440, + 547, + 487 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 440, + 547, + 487 + ], + "type": "text", + "content": ", we apply Equation 1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 487, + 548, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 487, + 548, + 592 + ], + "spans": [ + { + "bbox": [ + 312, + 487, + 548, + 592 + ], + "type": "text", + "content": "To achieve effective stroke feature transfer, we utilize latent representations for the content, reference, and contour images obtained through DDPM inversion. For predefined timesteps " + }, + { + "bbox": [ + 312, + 487, + 548, + 592 + ], + "type": "inline_equation", + "content": "t \\in \\{0, \\dots, T\\}" + }, + { + "bbox": [ + 312, + 487, + 548, + 592 + ], + "type": "text", + "content": ", the reference sketch " + }, + { + "bbox": [ + 312, + 487, + 548, + 592 + ], + "type": "inline_equation", + "content": "I^{ref}" + }, + { + "bbox": [ + 312, + 487, + 548, + 592 + ], + "type": "text", + "content": ", content image " + }, + { + "bbox": [ + 312, + 487, + 548, + 592 + ], + "type": "inline_equation", + "content": "I^{cnt}" + }, + { + "bbox": [ + 312, + 487, + 548, + 592 + ], + "type": "text", + "content": ", and contour image " + }, + { + "bbox": [ + 312, + 487, + 548, + 592 + ], + "type": "inline_equation", + "content": "I^{cont}" + }, + { + "bbox": [ + 312, + 487, + 548, + 592 + ], + "type": "text", + "content": " are inverted from their initial state (" + }, + { + "bbox": [ + 312, + 487, + 548, + 592 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 312, + 487, + 548, + 592 + ], + "type": "text", + "content": ") to Gaussian noise (" + }, + { + "bbox": [ + 312, + 487, + 548, + 592 + ], + "type": "inline_equation", + "content": "t = T" + }, + { + "bbox": [ + 312, + 487, + 548, + 592 + ], + "type": "text", + "content": "). During DDPM inversion, we also collect the query features of the content (" + }, + { + "bbox": [ + 312, + 487, + 548, + 592 + ], + "type": "inline_equation", + "content": "Q_{t}^{cnt}" + }, + { + "bbox": [ + 312, + 487, + 548, + 592 + ], + "type": "text", + "content": "), and the key and value features of the reference sketch (" + }, + { + "bbox": [ + 312, + 487, + 548, + 592 + ], + "type": "inline_equation", + "content": "K_{t}^{ref}" + }, + { + "bbox": [ + 312, + 487, + 548, + 592 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 312, + 487, + 548, + 592 + ], + "type": "inline_equation", + "content": "V_{t}^{ref}" + }, + { + "bbox": [ + 312, + 487, + 548, + 592 + ], + "type": "text", + "content": ") at each timestep." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 593, + 548, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 593, + 548, + 662 + ], + "spans": [ + { + "bbox": [ + 312, + 593, + 548, + 662 + ], + "type": "text", + "content": "We initialize the latent noise " + }, + { + "bbox": [ + 312, + 593, + 548, + 662 + ], + "type": "inline_equation", + "content": "z_{T}^{ske}" + }, + { + "bbox": [ + 312, + 593, + 548, + 662 + ], + "type": "text", + "content": " for the stylized sketch by copying the content latent noise " + }, + { + "bbox": [ + 312, + 593, + 548, + 662 + ], + "type": "inline_equation", + "content": "z_{T}^{cnt}" + }, + { + "bbox": [ + 312, + 593, + 548, + 662 + ], + "type": "text", + "content": ". To blend features from all three images, we combine the reference keys and values " + }, + { + "bbox": [ + 312, + 593, + 548, + 662 + ], + "type": "inline_equation", + "content": "K_{t}^{ref}" + }, + { + "bbox": [ + 312, + 593, + 548, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 312, + 593, + 548, + 662 + ], + "type": "inline_equation", + "content": "\\tilde{V}_{t}^{ref}" + }, + { + "bbox": [ + 312, + 593, + 548, + 662 + ], + "type": "text", + "content": " with the content features. This integration is achieved by mixing the keys and values using the following formulation:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 385, + 672, + 547, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 672, + 547, + 685 + ], + "spans": [ + { + "bbox": [ + 385, + 672, + 547, + 685 + ], + "type": "interline_equation", + "content": "K _ {t} ^ {\\mathrm {s k e}} = K _ {t} ^ {\\mathrm {r e f}} + \\alpha K _ {t} ^ {\\mathrm {c n t}}, \\tag {2}", + "image_path": "9c6006c6d0934209c7941817e0aa64726e431f9e59b9ff6ffeafa944dbfd3e59.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 386, + 687, + 547, + 701 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 687, + 547, + 701 + ], + "spans": [ + { + "bbox": [ + 386, + 687, + 547, + 701 + ], + "type": "interline_equation", + "content": "V _ {t} ^ {\\mathrm {s k e}} = V _ {t} ^ {\\mathrm {r e f}} + \\alpha V _ {t} ^ {\\mathrm {c n t}}, \\tag {3}", + "image_path": "e46d291baaf5736a8d2f047d0ecc744f5f9f8afa942cf87dd1fe085811f42ded.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "text", + "content": "16548" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 82, + 141, + 128 + ], + "blocks": [ + { + "bbox": [ + 66, + 82, + 141, + 128 + ], + "lines": [ + { + "bbox": [ + 66, + 82, + 141, + 128 + ], + "spans": [ + { + "bbox": [ + 66, + 82, + 141, + 128 + ], + "type": "image", + "image_path": "916b50d410eacea00395eb201704d9cd32dba2a002a668c04813f9289408722e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 70, + 129, + 141, + 173 + ], + "blocks": [ + { + "bbox": [ + 70, + 129, + 141, + 173 + ], + "lines": [ + { + "bbox": [ + 70, + 129, + 141, + 173 + ], + "spans": [ + { + "bbox": [ + 70, + 129, + 141, + 173 + ], + "type": "image", + "image_path": "50bdac387e457df1f2c4051ad7ca7814f72f93917c92d4ccd6d9b321fc8fc004.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 176, + 128, + 183 + ], + "lines": [ + { + "bbox": [ + 82, + 176, + 128, + 183 + ], + "spans": [ + { + "bbox": [ + 82, + 176, + 128, + 183 + ], + "type": "text", + "content": "Content Strokes" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 141, + 82, + 216, + 128 + ], + "blocks": [ + { + "bbox": [ + 141, + 82, + 216, + 128 + ], + "lines": [ + { + "bbox": [ + 141, + 82, + 216, + 128 + ], + "spans": [ + { + "bbox": [ + 141, + 82, + 216, + 128 + ], + "type": "image", + "image_path": "3c75c4095600f2a7e39dbdc113f186be63fda64ee19dc137fb7bfd073b475ba0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 148, + 133, + 213, + 169 + ], + "blocks": [ + { + "bbox": [ + 148, + 133, + 213, + 169 + ], + "lines": [ + { + "bbox": [ + 148, + 133, + 213, + 169 + ], + "spans": [ + { + "bbox": [ + 148, + 133, + 213, + 169 + ], + "type": "image", + "image_path": "543474c3d275e0516843866f2c5f783a6c7a9a57f9fb6727f66544c37412800c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 176, + 204, + 184 + ], + "lines": [ + { + "bbox": [ + 155, + 176, + 204, + 184 + ], + "spans": [ + { + "bbox": [ + 155, + 176, + 204, + 184 + ], + "type": "text", + "content": "Reference Strokes" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 217, + 82, + 290, + 128 + ], + "blocks": [ + { + "bbox": [ + 217, + 82, + 290, + 128 + ], + "lines": [ + { + "bbox": [ + 217, + 82, + 290, + 128 + ], + "spans": [ + { + "bbox": [ + 217, + 82, + 290, + 128 + ], + "type": "image", + "image_path": "67f0c0896f638805e512dcf825aa7fff932a2333abce36eac2f84f9742d80291.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 217, + 129, + 288, + 173 + ], + "blocks": [ + { + "bbox": [ + 217, + 129, + 288, + 173 + ], + "lines": [ + { + "bbox": [ + 217, + 129, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 217, + 129, + 288, + 173 + ], + "type": "image", + "image_path": "0326bad4b692da42f634f42cb19be639ba3cc52f4612d589683911b057c66903.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 176, + 277, + 185 + ], + "lines": [ + { + "bbox": [ + 230, + 176, + 277, + 185 + ], + "spans": [ + { + "bbox": [ + 230, + 176, + 277, + 185 + ], + "type": "text", + "content": "Aligned Strokes" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 193, + 295, + 224 + ], + "lines": [ + { + "bbox": [ + 62, + 193, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 62, + 193, + 295, + 224 + ], + "type": "text", + "content": "Figure 4. Stroke alignment results with CSA. The feature images are shown in the first row. The second row shows results where feature alignment between the key and value exchange is applied." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 236, + 296, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 236, + 296, + 269 + ], + "spans": [ + { + "bbox": [ + 61, + 236, + 296, + 269 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 61, + 236, + 296, + 269 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 61, + 236, + 296, + 269 + ], + "type": "text", + "content": " is a scalar parameter that controls the mixing ratio of content and reference features, allowing flexible adaptation to different stroke characteristics in the sketch." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 61, + 270, + 296, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 270, + 296, + 362 + ], + "spans": [ + { + "bbox": [ + 61, + 270, + 296, + 362 + ], + "type": "text", + "content": "By exchanging and blending features from the content, reference, and contour images, this approach enables the effective transfer of stroke attributes and semantic alignment. However, we found that some strokes may fail to represent the intended content curves accurately (see Fig. 4). To improve output quality and highlight the foreground, we introduce additional mechanisms to guide sketch generation, as detailed below." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 373, + 209, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 373, + 209, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 373, + 209, + 384 + ], + "type": "text", + "content": "3.3. Directive Attention Module" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 61, + 390, + 296, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 390, + 296, + 481 + ], + "spans": [ + { + "bbox": [ + 61, + 390, + 296, + 481 + ], + "type": "text", + "content": "To concentrate stylization on perceptually significant regions, the DAM enhances foreground focus during stroke transfer. We extract self-attention feature maps " + }, + { + "bbox": [ + 61, + 390, + 296, + 481 + ], + "type": "inline_equation", + "content": "F_{SA}" + }, + { + "bbox": [ + 61, + 390, + 296, + 481 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 61, + 390, + 296, + 481 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 61, + 390, + 296, + 481 + ], + "type": "text", + "content": " resolution, aggregating them across channels via averaging. These maps are segmented into clusters " + }, + { + "bbox": [ + 61, + 390, + 296, + 481 + ], + "type": "inline_equation", + "content": "M_{j}" + }, + { + "bbox": [ + 61, + 390, + 296, + 481 + ], + "type": "text", + "content": " using KMeans clustering. For each cluster " + }, + { + "bbox": [ + 61, + 390, + 296, + 481 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 61, + 390, + 296, + 481 + ], + "type": "text", + "content": " and noun " + }, + { + "bbox": [ + 61, + 390, + 296, + 481 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 61, + 390, + 296, + 481 + ], + "type": "text", + "content": " extracted from the content image's caption (via BLIP [20]), we calculate a relevance score:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 100, + 490, + 296, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 490, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 100, + 490, + 296, + 520 + ], + "type": "interline_equation", + "content": "r (j, n) = \\frac {\\sum_ {(x , y)} M _ {j} (x , y) \\cdot A _ {n} (x , y)}{\\sum_ {(x , y)} M _ {j} (x , y) + \\delta}, \\tag {4}", + "image_path": "39e4123c69beaec708f7f81e19144a437f959ae49ab76d00da644dc96d1f0efd.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 61, + 529, + 295, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 529, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 61, + 529, + 295, + 586 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 61, + 529, + 295, + 586 + ], + "type": "inline_equation", + "content": "A_{n}" + }, + { + "bbox": [ + 61, + 529, + 295, + 586 + ], + "type": "text", + "content": " is the cross-attention map for noun " + }, + { + "bbox": [ + 61, + 529, + 295, + 586 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 61, + 529, + 295, + 586 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 61, + 529, + 295, + 586 + ], + "type": "inline_equation", + "content": "\\delta = 10^{-5}" + }, + { + "bbox": [ + 61, + 529, + 295, + 586 + ], + "type": "text", + "content": " stabilizes the computation. Clusters with " + }, + { + "bbox": [ + 61, + 529, + 295, + 586 + ], + "type": "inline_equation", + "content": "r(j,n) > 0.35" + }, + { + "bbox": [ + 61, + 529, + 295, + 586 + ], + "type": "text", + "content": " are designated as foreground regions; remaining areas are suppressed, directing stroke stylization to salient elements (Fig. 5)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 61, + 587, + 295, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 587, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 61, + 587, + 295, + 680 + ], + "type": "text", + "content": "By integrating the segmented self-attention mechanism with cross-image attention, DAM allows precise control over foreground regions, ensuring that stylistic features, such as line styles and high-level semantic abstractions, are faithfully transferred from the reference sketch to the content image. This approach achieves high fidelity in style and content alignment, yielding sketches that closely reflect the intended reference style with minimized background interference." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 62, + 681, + 296, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 681, + 296, + 704 + ], + "spans": [ + { + "bbox": [ + 62, + 681, + 296, + 704 + ], + "type": "text", + "content": "As illustrated in Fig. 6, our method successfully aligns high-level semantic features and detailed stroke attributes," + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 316, + 82, + 375, + 139 + ], + "blocks": [ + { + "bbox": [ + 316, + 82, + 375, + 139 + ], + "lines": [ + { + "bbox": [ + 316, + 82, + 375, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 82, + 375, + 139 + ], + "type": "image", + "image_path": "1a2b12d3d4bb37065f294fc3d1bca357fde672634747f6add6c2260e06029624.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 148, + 547, + 213 + ], + "lines": [ + { + "bbox": [ + 313, + 148, + 547, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 148, + 547, + 213 + ], + "type": "text", + "content": "Figure 5. Examples of DAM in action. The first column shows the content images, the second column displays segmentation maps obtained by clustering self-attention maps, the third column provides the reference sketches, and the fourth column illustrates the sketches generated by DAM after applying stroke attribute transfer and segmentation." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 375, + 82, + 431, + 139 + ], + "blocks": [ + { + "bbox": [ + 375, + 82, + 431, + 139 + ], + "lines": [ + { + "bbox": [ + 375, + 82, + 431, + 139 + ], + "spans": [ + { + "bbox": [ + 375, + 82, + 431, + 139 + ], + "type": "image", + "image_path": "d3811bdf5b815f468ac4b1ac1a816a5a7e60fc3713e246f928b19c0fb34aa961.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 432, + 82, + 545, + 139 + ], + "blocks": [ + { + "bbox": [ + 432, + 82, + 545, + 139 + ], + "lines": [ + { + "bbox": [ + 432, + 82, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 432, + 82, + 545, + 139 + ], + "type": "image", + "image_path": "efab180c65eeab9dc62253d9bed59e8cc840f4ff5778773ab44d31900d83fc7f.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 315, + 217, + 391, + 297 + ], + "blocks": [ + { + "bbox": [ + 315, + 217, + 391, + 297 + ], + "lines": [ + { + "bbox": [ + 315, + 217, + 391, + 297 + ], + "spans": [ + { + "bbox": [ + 315, + 217, + 391, + 297 + ], + "type": "image", + "image_path": "d8543fee5368f488fb073440f20cb758a02e4566888a46cb536b70cb476234a7.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 391, + 217, + 438, + 298 + ], + "blocks": [ + { + "bbox": [ + 391, + 217, + 438, + 298 + ], + "lines": [ + { + "bbox": [ + 391, + 217, + 438, + 298 + ], + "spans": [ + { + "bbox": [ + 391, + 217, + 438, + 298 + ], + "type": "image", + "image_path": "1b5d661283d7becd8dd705ead768a7b6c14c027dcd521b62502e32a5675f6e8c.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 305, + 547, + 390 + ], + "lines": [ + { + "bbox": [ + 313, + 305, + 547, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 547, + 390 + ], + "type": "text", + "content": "Figure 6. Comparison of sketches generated using DAM. The top row shows Cnt. and Ref. images with stroke styles reflecting specific high-level attributes such as hair, eyebrows, and abstracted clothing texture. The second to fourth columns display sketches generated to match the content images in the top row, each adopting the stroke styles and high-level features from the reference. The bottom row highlights zoomed-in areas to emphasize the transferred stroke attributes." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 440, + 217, + 547, + 298 + ], + "blocks": [ + { + "bbox": [ + 440, + 217, + 547, + 298 + ], + "lines": [ + { + "bbox": [ + 440, + 217, + 547, + 298 + ], + "spans": [ + { + "bbox": [ + 440, + 217, + 547, + 298 + ], + "type": "image", + "image_path": "9231b8af5b80fd7ce3c5aaa1552821f3a3b982f4f606084f788a51cd3d165235.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 399, + 547, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 547, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 547, + 456 + ], + "type": "text", + "content": "producing coherent sketches that maintain stylistic and structural consistency with the reference. This approach minimizes background interference while ensuring that the final sketch reflects the nuanced stroke characteristics of the reference." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 465, + 476, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 476, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 476, + 477 + ], + "type": "text", + "content": "3.4. Semantic Preservation Module" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 483, + 547, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 547, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 547, + 586 + ], + "type": "text", + "content": "Although the injection of keys and values during sketch generation can effectively transfer stroke attributes to " + }, + { + "bbox": [ + 313, + 483, + 547, + 586 + ], + "type": "inline_equation", + "content": "I^{ske}" + }, + { + "bbox": [ + 313, + 483, + 547, + 586 + ], + "type": "text", + "content": ", semantic inconsistencies may arise, particularly when the reference sketch " + }, + { + "bbox": [ + 313, + 483, + 547, + 586 + ], + "type": "inline_equation", + "content": "I^{ref}" + }, + { + "bbox": [ + 313, + 483, + 547, + 586 + ], + "type": "text", + "content": " does not align semantically with the content image " + }, + { + "bbox": [ + 313, + 483, + 547, + 586 + ], + "type": "inline_equation", + "content": "I^{cnt}" + }, + { + "bbox": [ + 313, + 483, + 547, + 586 + ], + "type": "text", + "content": ". This can lead to noise and misplaced strokes that disrupt the semantic structure of the generated sketch. To address this, semantic guidance is essential to ensure that each pixel aligns correctly with its corresponding structure." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 588, + 547, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 547, + 704 + ], + "type": "text", + "content": "Text-based guidance alone is often insufficient for precise structural alignment, as textual descriptions may not accurately map to image details. To overcome this limitation, we incorporate contour information to guide semantic pixel alignment and supplement missing semantic strokes. However, we found that overly detailed contour information can disrupt high-level semantics (as edge detection identifies changes based on pixel gradients). For example, when drawing portraits, the eyes are often represented as solid dots rather than pixel-defined circles, necessitating text-based" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "text", + "content": "16549" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 81, + 293, + 193 + ], + "blocks": [ + { + "bbox": [ + 66, + 81, + 293, + 193 + ], + "lines": [ + { + "bbox": [ + 66, + 81, + 293, + 193 + ], + "spans": [ + { + "bbox": [ + 66, + 81, + 293, + 193 + ], + "type": "image", + "image_path": "0c076db4a915ddbcc366be45d1e4ce2c54ee2cbab395a50ed83c0304139675c8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 200, + 296, + 264 + ], + "lines": [ + { + "bbox": [ + 61, + 200, + 296, + 264 + ], + "spans": [ + { + "bbox": [ + 61, + 200, + 296, + 264 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison of sketch generation results with and without contour integration. The top row shows Ref. and Cnt. images, followed by results without contour integration (w/o Cont. Int.) and with contour integration (w/ Cont. Int.) as discussed in Sec. 3.4. Contour integration leads to better alignment of semantic features and reduces background interference." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 62, + 268, + 255, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 268, + 255, + 279 + ], + "spans": [ + { + "bbox": [ + 62, + 268, + 255, + 279 + ], + "type": "text", + "content": "semantic guidance for accurate sketch rendering." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 280, + 296, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 280, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 61, + 280, + 296, + 348 + ], + "type": "text", + "content": "The text-based semantic loss is derived from the guidance provided in the image-to-image diffusion pipeline, which ensures high-level semantic alignment during sketch generation: " + }, + { + "bbox": [ + 61, + 280, + 296, + 348 + ], + "type": "inline_equation", + "content": "L_{sem} = \\lambda \\cdot \\mathrm{CLIP}(I^{ske}, T^{cnt})" + }, + { + "bbox": [ + 61, + 280, + 296, + 348 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 61, + 280, + 296, + 348 + ], + "type": "inline_equation", + "content": "T^{cnt}" + }, + { + "bbox": [ + 61, + 280, + 296, + 348 + ], + "type": "text", + "content": " is the text prompt extracted from " + }, + { + "bbox": [ + 61, + 280, + 296, + 348 + ], + "type": "inline_equation", + "content": "I^{cnt}" + }, + { + "bbox": [ + 61, + 280, + 296, + 348 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 61, + 280, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 61, + 280, + 296, + 348 + ], + "type": "text", + "content": " is a weighting parameter for the text guidance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 350, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 350, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 61, + 350, + 296, + 396 + ], + "type": "text", + "content": "The contour-based guidance originates from the cached query features during the DDPM inversion process. To integrate this contour information effectively, we use the following equation to adjust the query:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 113, + 400, + 296, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 400, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 113, + 400, + 296, + 415 + ], + "type": "interline_equation", + "content": "Q _ {i + 1} ^ {s k e} = \\gamma Q _ {i} ^ {\\text {c o n t}} + (1 - \\gamma) Q _ {i} ^ {s k e}, \\tag {5}", + "image_path": "5305db050b1fe0d75bd435e64320636335f8ef9302a8645b0736e9e1d3500a48.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 419, + 296, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 419, + 296, + 489 + ], + "spans": [ + { + "bbox": [ + 61, + 419, + 296, + 489 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 61, + 419, + 296, + 489 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 61, + 419, + 296, + 489 + ], + "type": "text", + "content": " is a tunable parameter that controls the influence of contour information, set to 0.25 by default in our experiments. As illustrated in Fig. 7, contour integration significantly improves the alignment of semantic structures, ensuring that key details, such as object outlines, are preserved without introducing unnecessary background noise." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 490, + 296, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 490, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 61, + 490, + 296, + 559 + ], + "type": "text", + "content": "By combining high-level semantic text guidance with contour-based structural alignment, we ensure that the final generated sketch " + }, + { + "bbox": [ + 61, + 490, + 296, + 559 + ], + "type": "inline_equation", + "content": "I^{ske}" + }, + { + "bbox": [ + 61, + 490, + 296, + 559 + ], + "type": "text", + "content": " maintains semantic completeness and accurately represents the image of the content " + }, + { + "bbox": [ + 61, + 490, + 296, + 559 + ], + "type": "inline_equation", + "content": "I^{cnt}" + }, + { + "bbox": [ + 61, + 490, + 296, + 559 + ], + "type": "text", + "content": ". This collaborative approach allows for the preservation of both stroke attributes and semantic integrity in the sketch." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 565, + 272, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 565, + 272, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 565, + 272, + 578 + ], + "type": "text", + "content": "3.5. Stroke Details Propagation Enhancement" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 582, + 296, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 582, + 296, + 686 + ], + "spans": [ + { + "bbox": [ + 61, + 582, + 296, + 686 + ], + "type": "text", + "content": "Traditional self-attention blocks often focus on limited areas around image patches, while masked extended attention blocks distribute attention more uniformly by expanding receptive fields across the image. Although this broader approach captures larger context, it can introduce noise and blur finer details [1, 29]. To address this, we adopt a refined contrast operation, inspired by [1], to dynamically enhance high-variance regions and suppress low-contrast noise. This contrast adjustment is defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 88, + 693, + 296, + 705 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 693, + 296, + 705 + ], + "spans": [ + { + "bbox": [ + 88, + 693, + 296, + 705 + ], + "type": "interline_equation", + "content": "\\operatorname {E n h a n c e} (A) = (A - \\mu (A)) \\zeta (\\sigma (A)) + \\mu (A), \\tag {6}", + "image_path": "83e056ff0ad1d75d6c027433d40d93b2f9fe92204b048699709c90ba8dd670d0.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 81, + 547, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 81, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 312, + 81, + 547, + 127 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 312, + 81, + 547, + 127 + ], + "type": "inline_equation", + "content": "\\mu(A)" + }, + { + "bbox": [ + 312, + 81, + 547, + 127 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 312, + 81, + 547, + 127 + ], + "type": "inline_equation", + "content": "\\sigma(A)" + }, + { + "bbox": [ + 312, + 81, + 547, + 127 + ], + "type": "text", + "content": " represent the mean and standard deviation operations, respectively, and " + }, + { + "bbox": [ + 312, + 81, + 547, + 127 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 312, + 81, + 547, + 127 + ], + "type": "text", + "content": " is an adaptive contrast operator. This technique effectively reduces noise, ensuring critical details are preserved during sketch generation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 128, + 547, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 128, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 312, + 128, + 547, + 266 + ], + "type": "text", + "content": "Building on this, we incorporate the stroke-based refinement concept from SDEdit [27]. Starting with a sketch initialized by stroke exchange, we denoise the image using classifier-free guidance (CFG) [6, 14]. During each denoising step, we utilize two parallel forward passes in the network. The first pass employs a cross-image attention layer to capture the stroke characteristics and abstraction level of the reference sketch, generating " + }, + { + "bbox": [ + 312, + 128, + 547, + 266 + ], + "type": "inline_equation", + "content": "z^{ske}" + }, + { + "bbox": [ + 312, + 128, + 547, + 266 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 312, + 128, + 547, + 266 + ], + "type": "inline_equation", + "content": "\\epsilon_{stroke}^{\\times} = \\epsilon_{\\theta}^{\\times}(z_{t}^{ske})" + }, + { + "bbox": [ + 312, + 128, + 547, + 266 + ], + "type": "text", + "content": ", while simultaneously retaining the semantic context extracted from the content image's descriptive prompt: " + }, + { + "bbox": [ + 312, + 128, + 547, + 266 + ], + "type": "inline_equation", + "content": "\\epsilon_{text}^{\\times} = \\epsilon_{\\theta}^{\\times}(z_{t}^{text})" + }, + { + "bbox": [ + 312, + 128, + 547, + 266 + ], + "type": "text", + "content": ". The second pass applies regular self-attention to enhance the sketch's structural integrity: " + }, + { + "bbox": [ + 312, + 128, + 547, + 266 + ], + "type": "inline_equation", + "content": "\\epsilon^{self} = \\epsilon_{\\theta}^{self}(z_{t}^{ske})" + }, + { + "bbox": [ + 312, + 128, + 547, + 266 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 267, + 547, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 267, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 267, + 547, + 289 + ], + "type": "text", + "content": "Following the CFG scheme, the predicted noise " + }, + { + "bbox": [ + 313, + 267, + 547, + 289 + ], + "type": "inline_equation", + "content": "\\epsilon^t" + }, + { + "bbox": [ + 313, + 267, + 547, + 289 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 318, + 295, + 547, + 311 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 295, + 547, + 311 + ], + "spans": [ + { + "bbox": [ + 318, + 295, + 547, + 311 + ], + "type": "interline_equation", + "content": "\\epsilon^ {t} = \\epsilon^ {s e l f} + \\beta_ {s g} (\\epsilon_ {s t r o k e} ^ {\\times} - \\epsilon^ {s e l f}) + \\beta_ {t e x t} (\\epsilon_ {t e x t} ^ {\\times} - \\epsilon^ {s e l f}), (7)", + "image_path": "8b14060f599b7c4b7bb1b38ac0d643289a598f41a182edd3b69a217588e06964.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 316, + 547, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 547, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 547, + 338 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 316, + 547, + 338 + ], + "type": "inline_equation", + "content": "\\beta_{sg}" + }, + { + "bbox": [ + 313, + 316, + 547, + 338 + ], + "type": "text", + "content": " is the stroke guidance scale, and " + }, + { + "bbox": [ + 313, + 316, + 547, + 338 + ], + "type": "inline_equation", + "content": "\\beta_{text}" + }, + { + "bbox": [ + 313, + 316, + 547, + 338 + ], + "type": "text", + "content": " is the weight for the content text context." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 348, + 436, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 436, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 436, + 361 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 368, + 547, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 368, + 547, + 449 + ], + "spans": [ + { + "bbox": [ + 312, + 368, + 547, + 449 + ], + "type": "text", + "content": "Datasets. Our experiments utilize three datasets: FS2K [9], the Anime dataset [19], and our newly created Stroke2Sketch-dataset. The FS2K dataset includes 5,140 facial image-sketch pairs. We randomly selected 1,000 pairs for testing, using color images as content images and choosing one sketch from the test set as a reference style. A similar selection strategy was used for the Anime dataset [19]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 449, + 547, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 449, + 547, + 495 + ], + "spans": [ + { + "bbox": [ + 312, + 449, + 547, + 495 + ], + "type": "text", + "content": "The Stroke2Sketch-dataset includes 50 content images from diverse categories and 20 distinct sketch styles, including single-line sketches, ink sketches, line art, and realistic sketches. Further details are provided in the Appendix." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 495, + 547, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 547, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 547, + 565 + ], + "type": "text", + "content": "Metrics. While traditional metrics like ArtFID [43], LPIPS [53], and FID [13] are standard for style transfer, they struggle to capture the semantic sparsity and high-level abstraction unique to sketch generation. Consistent with prior work [38], we prioritize user perception to better reflect artistic and semantic quality in sketches." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 571, + 446, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 571, + 446, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 571, + 446, + 583 + ], + "type": "text", + "content": "4.1. Qualitative Comparison" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 588, + 547, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 588, + 547, + 669 + ], + "spans": [ + { + "bbox": [ + 312, + 588, + 547, + 669 + ], + "type": "text", + "content": "We evaluate our proposed method through a comparison with eight state-of-the-art methods, including three training-based style transfer methods (Ref2sketch [2], Semi-ref2sketch [33], and CSGO [46]), and five training-free style transfer methods (IP-Adapter [49], InstantStyle [41], InstantStyle-plus [42], RB-Modulation [32], and StyleID [7]). Each of these methods takes a reference sketch as input." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 670, + 547, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 670, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 312, + 670, + 547, + 704 + ], + "type": "text", + "content": "Fig. 8 illustrates the qualitative results across a variety of content images and reference sketches. In the first column, we display the content images, and the second column" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "text", + "content": "16550" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 79, + 545, + 268 + ], + "blocks": [ + { + "bbox": [ + 66, + 79, + 545, + 268 + ], + "lines": [ + { + "bbox": [ + 66, + 79, + 545, + 268 + ], + "spans": [ + { + "bbox": [ + 66, + 79, + 545, + 268 + ], + "type": "image", + "image_path": "42bf7bfd95b139362e9eb337143bd1d3b8463e0598966a3b9849649cc8fd2cd0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 275, + 548, + 309 + ], + "lines": [ + { + "bbox": [ + 61, + 275, + 548, + 309 + ], + "spans": [ + { + "bbox": [ + 61, + 275, + 548, + 309 + ], + "type": "text", + "content": "Figure 8. Qualitative comparison of sketch generation across training-based (4th-6th columns) and training-free baselines (7th-11th columns) using various reference sketches. Our method (3rd column) demonstrates superior adaptability to different reference styles, maintaining both stroke fidelity and semantic consistency across a range of content and reference types." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 65, + 321, + 293, + 457 + ], + "blocks": [ + { + "bbox": [ + 65, + 321, + 293, + 457 + ], + "lines": [ + { + "bbox": [ + 65, + 321, + 293, + 457 + ], + "spans": [ + { + "bbox": [ + 65, + 321, + 293, + 457 + ], + "type": "image", + "image_path": "8799ed8266de01dfdbe80d9528ba2c4541a6962442c4c13ac664c848a9df8f3d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 465, + 296, + 509 + ], + "lines": [ + { + "bbox": [ + 61, + 465, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 61, + 465, + 296, + 509 + ], + "type": "text", + "content": "Figure 9. (a) Content image and manual stroke stylization by CLIPascene [39] with Adobe Illustrator; (b) Our method's automatically generated sketches based on reference stroke styles; (c) Appearance transfer results using similar references." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 518, + 296, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 518, + 296, + 633 + ], + "spans": [ + { + "bbox": [ + 61, + 518, + 296, + 633 + ], + "type": "text", + "content": "shows the reference sketches. Notably, the fourth row's reference sketch is of a type seen in the Semi-ref2sketch training data, while the other three reference sketches represent styles outside the training data for Semi-ref2sketch. For methods requiring prompt input, we uniformly set the prompt to \"sketch photo,\" while other parameters follow each method's default configuration. Our approach demonstrates robustness across diverse reference styles, accurately preserving both the stroke style and high-level semantic details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 635, + 296, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 635, + 296, + 704 + ], + "spans": [ + { + "bbox": [ + 61, + 635, + 296, + 704 + ], + "type": "text", + "content": "Comparison with vector sketch generation: Fig. 9(a) compares our method with the vectorized sketch generation approach of CLIPascene [39]. The content image is shown alongside vectorized sketches produced by CLIPascene, which were further manually refined with brush strokes in Adobe Illustrator. While CLIPascene can generate abstract" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 315, + 318, + 545, + 376 + ], + "blocks": [ + { + "bbox": [ + 315, + 318, + 545, + 376 + ], + "lines": [ + { + "bbox": [ + 315, + 318, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 315, + 318, + 545, + 376 + ], + "type": "image", + "image_path": "4da0f4726b06cf026022e03711cfa9049023b28b1318500d88bac7199d7b8b07.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 384, + 546, + 405 + ], + "lines": [ + { + "bbox": [ + 313, + 384, + 546, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 546, + 405 + ], + "type": "text", + "content": "Figure 10. Stroke2Sketch's color sketch generation preserving reference styles and stroke characteristics" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 416, + 547, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 416, + 547, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 547, + 461 + ], + "type": "text", + "content": "sketches, it requires post-processing to achieve consistent stroke styling, whereas our method (Fig. 9(b)) automatically produces sketches with stylistically aligned strokes based on reference attributes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 463, + 547, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 547, + 544 + ], + "type": "text", + "content": "Comparison with appearance transfer methods: Fig. 9(c) shows results from appearance transfer method of CIA [1] applied to the same reference sketches. Although CIA excel in transferring visual features based on semantic similarity and object category, they focus on appearance rather than stroke style, making them less suitable for our sketch generation goals." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 544, + 547, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 547, + 580 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 547, + 580 + ], + "type": "text", + "content": "Non-grayscale Sketch Generation. Building on its grayscale performance, Stroke2Sketch maintains reference stroke patterns and artistic styles in color outputs (Fig. 10)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 582, + 452, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 452, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 452, + 594 + ], + "type": "text", + "content": "4.2.Quantitative Comparison" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 600, + 547, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 600, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 312, + 600, + 547, + 704 + ], + "type": "text", + "content": "We quantitatively evaluate our method against several state-of-the-art sketch extraction techniques, including both training-based (Ref2Sketch [2], Semi-Ref2Sketch [33], and Informative-drawing [5]) and training-free style transfer methods (IP-Adapter [49], InstantStyle [41], InstantStyle-plus [42], StyleID [7]). Table 1 presents the results on the Stroke2Sketch-dataset, showing that our method achieves the lowest ArtFID and FID scores, indicating superior performance in both style alignment and content preservation." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "text", + "content": "16551" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 85, + 79, + 523, + 125 + ], + "blocks": [ + { + "bbox": [ + 85, + 79, + 523, + 125 + ], + "lines": [ + { + "bbox": [ + 85, + 79, + 523, + 125 + ], + "spans": [ + { + "bbox": [ + 85, + 79, + 523, + 125 + ], + "type": "table", + "html": "
MetricOursRef2sketchSemi-refInfor-drawingIP-AdapterInstantStyleInstantStyle+StyleID
ArtFID ↓32.45545.29233.24234.21433.45732.53237.65635.727
LPIPS ↓0.53150.69820.53060.60370.66340.54320.65320.5426
FID ↓22.43534.65024.35925.03524.06823.94026.63225.658
", + "image_path": "1456d75b9b7e2080d0384a9fd303beec5b35e6f09ece7d8470cdbc28f5c235a7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 69, + 165, + 290, + 278 + ], + "blocks": [ + { + "bbox": [ + 69, + 165, + 290, + 278 + ], + "lines": [ + { + "bbox": [ + 69, + 165, + 290, + 278 + ], + "spans": [ + { + "bbox": [ + 69, + 165, + 290, + 278 + ], + "type": "image", + "image_path": "61f46cafb46bca11340ace69c7c4271bde8205f5d8e3ccd0db892eb950d8be44.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 288, + 295, + 310 + ], + "lines": [ + { + "bbox": [ + 62, + 288, + 295, + 310 + ], + "spans": [ + { + "bbox": [ + 62, + 288, + 295, + 310 + ], + "type": "text", + "content": "Figure 11. User study results. Our proposed method received the highest preference." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 319, + 296, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 296, + 354 + ], + "type": "text", + "content": "These metrics highlight our method's effectiveness in producing high-quality sketches with strong semantic and stylistic fidelity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 361, + 135, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 361, + 135, + 373 + ], + "spans": [ + { + "bbox": [ + 62, + 361, + 135, + 373 + ], + "type": "text", + "content": "4.3. User Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 378, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 378, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 61, + 378, + 295, + 529 + ], + "type": "text", + "content": "We randomly selected 15 content images and 15 reference sketches from Stroke2Sketch-dataset, creating 225 image pairs. From these, we sampled 20 pairs to generate stylized images using four methods. The results were displayed side-by-side in random order, and participants were asked to choose their favorite based on three criteria: content extraction, stroke stylization, and overall preference. We collected 2,000 votes for each criterion from 100 users and presented the results as a bar chart. As shown in Fig. 11, our method received the highest preference, outperforming both training-based and training-free baselines, demonstrating its effectiveness in handling diverse stroke styles and abstract artistic effects." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 536, + 153, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 153, + 549 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 153, + 549 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 553, + 297, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 553, + 297, + 705 + ], + "spans": [ + { + "bbox": [ + 61, + 553, + 297, + 705 + ], + "type": "text", + "content": "To validate our method's components, we performed an ablation study on the Stroke2Sketch-dataset. As shown in Table 2, removing any of the Directive Attention Module (DAM), Semantic Preservation Module (SPM), or Stroke Details Propagation Enhancement (SDPE) led to decreased performance across ArtFID, FID, and LPIPS metrics, confirming the contribution of each component. Specifically, without DAM, ArtFID increased from 32.45 to 38.67 and FID rose to 26.53, indicating weaker style-content alignment and content leakage in sketches, as observed in Fig. 12. The absence of SPM resulted in ArtFID rising to 36.89 and FID to 30.47, with sketches losing semantic coherence and structural integrity, such as poorly defined object outlines." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 335, + 159, + 526, + 216 + ], + "blocks": [ + { + "bbox": [ + 61, + 130, + 547, + 152 + ], + "lines": [ + { + "bbox": [ + 61, + 130, + 547, + 152 + ], + "spans": [ + { + "bbox": [ + 61, + 130, + 547, + 152 + ], + "type": "text", + "content": "Table 1. Quantitative comparison on Stroke2Sketch-dataset with training-based (3rd-5th columns) and training-free baselines (6th-9th columns)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 335, + 159, + 526, + 216 + ], + "lines": [ + { + "bbox": [ + 335, + 159, + 526, + 216 + ], + "spans": [ + { + "bbox": [ + 335, + 159, + 526, + 216 + ], + "type": "table", + "html": "
ConfigurationArtFIDFIDLPIPS
A: Ours32.4522.430.530
B: - DAM (Sec. 3.3)38.6726.530.672
C: - SPM (Sec. 3.4)36.8930.470.637
D: - SDPE (Sec. 3.5)40.5332.440.598
", + "image_path": "5ea52e9eb4831258121016f52fef30ed57e197c0a34d56ae099c209181402324.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 325, + 224, + 534, + 235 + ], + "lines": [ + { + "bbox": [ + 325, + 224, + 534, + 235 + ], + "spans": [ + { + "bbox": [ + 325, + 224, + 534, + 235 + ], + "type": "text", + "content": "Table 2. Ablation study of different variants of our method." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 319, + 243, + 543, + 340 + ], + "blocks": [ + { + "bbox": [ + 319, + 243, + 543, + 340 + ], + "lines": [ + { + "bbox": [ + 319, + 243, + 543, + 340 + ], + "spans": [ + { + "bbox": [ + 319, + 243, + 543, + 340 + ], + "type": "image", + "image_path": "c46d58774955ce8de6a0f6ce3a4421bc7f9fc8feb70d07603acda03c6ca15335.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 349, + 548, + 370 + ], + "lines": [ + { + "bbox": [ + 313, + 349, + 548, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 548, + 370 + ], + "type": "text", + "content": "Figure 12. Ablative qualitative comparison of different variants of our method." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 383, + 548, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 383, + 548, + 488 + ], + "spans": [ + { + "bbox": [ + 312, + 383, + 548, + 488 + ], + "type": "text", + "content": "Most notably, removing SDPE caused the sharpest decline, with ArtFID reaching 40.53 and sketches exhibiting excessive noise and loss of fine details, as evidenced by cluttered textures in Fig. 12. In contrast, the full method (Configuration A) achieved optimal scores of 32.45 (ArtFID), 22.43 (FID), and 0.530 (LPIPS), producing high-quality, reference-aligned sketches with precise details. These results highlight the critical roles of DAM, SPM, and SDPE in enhancing sketch quality." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 498, + 384, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 384, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 384, + 510 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 518, + 548, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 518, + 548, + 705 + ], + "spans": [ + { + "bbox": [ + 312, + 518, + 548, + 705 + ], + "type": "text", + "content": "Our training-free method for content-to-sketch generation aligns stroke attributes and semantic texture sparsity, mimicking artistic subject extraction to produce high-fidelity sketches. Leveraging pretrained models, it achieves state-of-the-art performance in quantitative metrics and user evaluations. Limitations arise with overly simplistic or complex reference sketches (see appendix for failure cases). Future work could explore decoupling semantic information from stroke attributes to improve adaptability and sketch quality. Acknowledgment. This work is supported by the Guangdong Natural Science Funds for Distinguished Young Scholars (Grant 2023B1515020097), the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No.: AISG3-GV-2023-011), the Singapore Ministry of Education AcRF Tier 1 Grant (Grant No.: MSS25C004), and the Lee Kong Chian Fellowships." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "text", + "content": "16552" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 80, + 119, + 92 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 80, + 119, + 92 + ], + "spans": [ + { + "bbox": [ + 63, + 80, + 119, + 92 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 99, + 297, + 703 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 99, + 297, + 142 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 99, + 297, + 142 + ], + "spans": [ + { + "bbox": [ + 67, + 99, + 297, + 142 + ], + "type": "text", + "content": "[1] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer. In ACM SIGGRAPH 2024 Conference Papers, pages 1–12, 2024. 3, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 143, + 296, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 143, + 296, + 185 + ], + "spans": [ + { + "bbox": [ + 67, + 143, + 296, + 185 + ], + "type": "text", + "content": "[2] Amirsaman Ashtari, Chang Wook Seo, Cholmin Kang, Sihun Cha, and Junyong Noh. Reference based sketch extraction via attention mechanism. ACM Trans. Graph., 41(6), 2022. 1, 3, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 186, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 186, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 67, + 186, + 296, + 217 + ], + "type": "text", + "content": "[3] John Canny. A computational approach to edge detection. IEEE Transactions on pattern analysis and machine intelligence, (6):679-698, 1986. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 219, + 296, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 219, + 296, + 261 + ], + "spans": [ + { + "bbox": [ + 67, + 219, + 296, + 261 + ], + "type": "text", + "content": "[4] Nan Cao, Xin Yan, Yang Shi, and Chaoran Chen. Ai-sketcher: a deep generative model for producing high-quality sketches. In Proceedings of the AAAI conference on artificial intelligence, pages 2564-2571, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 262, + 296, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 262, + 296, + 313 + ], + "spans": [ + { + "bbox": [ + 67, + 262, + 296, + 313 + ], + "type": "text", + "content": "[5] Caroline Chan, Frédo Durand, and Phillip Isola. Learning to generate line drawings that convey geometry and semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7915-7925, 2022. 1, 3, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 316, + 296, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 316, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 67, + 316, + 296, + 348 + ], + "type": "text", + "content": "[6] Hyungjin Chung, Jeongsol Kim, Geon Yeong Park, Hyelin Nam, and Jong Chul Ye. Cfg++: Manifold-constrained classifier free guidance for diffusion models. 2024. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 349, + 296, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 349, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 67, + 349, + 296, + 402 + ], + "type": "text", + "content": "[7] Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8795-8805, 2024. 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 403, + 296, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 403, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 67, + 403, + 296, + 464 + ], + "type": "text", + "content": "[8] Ayan Das, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Beziersketch: A generative model for scalable vector sketches. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI 16, pages 632-647. Springer, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 467, + 296, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 467, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 67, + 467, + 296, + 509 + ], + "type": "text", + "content": "[9] Deng-Ping Fan, Ziling Huang, Peng Zheng, Hong Liu, Xuebin Qin, and Luc Van Gool. Facial-sketch synthesis: a new challenge. Machine Intelligence Research, 19(4):257-287, 2022. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 63, + 510, + 296, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 510, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 63, + 510, + 296, + 552 + ], + "type": "text", + "content": "[10] Kevin Frans, Lisa Soros, and Olaf Witkowski. Clipdraw: Exploring text-to-drawing synthesis through language-image encoders. Advances in Neural Information Processing Systems, 35:5207-5218, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 63, + 554, + 296, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 554, + 296, + 596 + ], + "spans": [ + { + "bbox": [ + 63, + 554, + 296, + 596 + ], + "type": "text", + "content": "[11] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 63, + 597, + 296, + 649 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 597, + 296, + 649 + ], + "spans": [ + { + "bbox": [ + 63, + 597, + 296, + 649 + ], + "type": "text", + "content": "[12] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4775–4785, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 63, + 651, + 296, + 703 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 651, + 296, + 703 + ], + "spans": [ + { + "bbox": [ + 63, + 651, + 296, + 703 + ], + "type": "text", + "content": "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 82, + 548, + 703 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 315, + 82, + 547, + 102 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 82, + 547, + 102 + ], + "spans": [ + { + "bbox": [ + 315, + 82, + 547, + 102 + ], + "type": "text", + "content": "[14] Jonathan Ho. Classifier-free diffusion guidance. ArXiv, abs/2207.12598, 2022. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 104, + 548, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 104, + 548, + 145 + ], + "spans": [ + { + "bbox": [ + 315, + 104, + 548, + 145 + ], + "type": "text", + "content": "[15] Jjin Hu, Ke Li, Yonggang Qi, and Yi-Zhe Song. Scale-adaptive diffusion model for complex sketch synthesis. In The Twelfth International Conference on Learning Representations, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 147, + 548, + 199 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 147, + 548, + 199 + ], + "spans": [ + { + "bbox": [ + 315, + 147, + 548, + 199 + ], + "type": "text", + "content": "[16] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly DDPM noise space: Inversion and manipulations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12469-12478, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 201, + 548, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 201, + 548, + 232 + ], + "spans": [ + { + "bbox": [ + 315, + 201, + 548, + 232 + ], + "type": "text", + "content": "[17] Jaeseok Jeong, Junho Kim, Yunjoy Choi, Gayoung Lee, and Youngjung Uh. Visual style prompting with swapping self-attention. arXiv preprint arXiv:2402.12974, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 233, + 548, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 233, + 548, + 285 + ], + "spans": [ + { + "bbox": [ + 315, + 233, + 548, + 285 + ], + "type": "text", + "content": "[18] Yutao Jiang, Yang Zhou, Yuan Liang, Wenxi Liu, Jianbo Jiao, Yuhui Quan, and Shengfeng He. Diffuse3d: Wide-angle 3d photography via bilateral diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8998-9008, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 286, + 548, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 286, + 548, + 306 + ], + "spans": [ + { + "bbox": [ + 315, + 286, + 548, + 306 + ], + "type": "text", + "content": "[19] Tae Bum Kang. Anime sketch colorization pair, 2018. Accessed: 2024-05-18. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 308, + 548, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 308, + 548, + 360 + ], + "spans": [ + { + "bbox": [ + 315, + 308, + 548, + 360 + ], + "type": "text", + "content": "[20] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 3, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 361, + 548, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 361, + 548, + 415 + ], + "spans": [ + { + "bbox": [ + 315, + 361, + 548, + 415 + ], + "type": "text", + "content": "[21] Xinzhe Li, Jiahui Zhan, Shengfeng He, Yangyang Xu, Junyu Dong, Huaidong Zhang, and Yong Du. Personamagic: Stage-regulated high-fidelity face customization with tandem equilibrium. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4995-5003, 2025. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 415, + 548, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 415, + 548, + 456 + ], + "spans": [ + { + "bbox": [ + 315, + 415, + 548, + 456 + ], + "type": "text", + "content": "[22] Chenxi Liu, Enrique Rosales, and Alla Sheffer. Strokeaggregator: Consolidating raw sketches into artist-intended curve drawings. ACM Transactions on Graphics (TOG), 37(4):1-15, 2018. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 458, + 548, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 458, + 548, + 499 + ], + "spans": [ + { + "bbox": [ + 315, + 458, + 548, + 499 + ], + "type": "text", + "content": "[23] Chenxi Liu, Toshiki Aoki, Mikhail Bessmeltsev, and Alla Sheffer. Stripmaker: Perception-driven learned vector sketch consolidation. ACM Transactions on Graphics (TOG), 42(4): 1-15, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 501, + 548, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 501, + 548, + 553 + ], + "spans": [ + { + "bbox": [ + 315, + 501, + 548, + 553 + ], + "type": "text", + "content": "[24] Difan Liu, Matthew Fisher, Aaron Hertzmann, and Evangelos Kalogerakis. Neural strokes: Stylized line drawing of 3d shapes. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 14184-14193. IEEE, 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 555, + 548, + 607 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 555, + 548, + 607 + ], + "spans": [ + { + "bbox": [ + 315, + 555, + 548, + 607 + ], + "type": "text", + "content": "[25] Haofeng Liu, Chenshu Xu, Yifei Yang, Lihua Zeng, and Shengfeng He. Drag your noise: Interactive point-based editing via diffusion semantic propagation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6743-6752, 2024. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 608, + 548, + 649 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 608, + 548, + 649 + ], + "spans": [ + { + "bbox": [ + 315, + 608, + 548, + 649 + ], + "type": "text", + "content": "[26] Xiao-Chang Liu, Yu-Chen Wu, and Peter Hall. Painterly style transfer with learned brush strokes. IEEE Transactions on Visualization and Computer Graphics, 30(9):6309-6320, 2024. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 315, + 651, + 548, + 703 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 651, + 548, + 703 + ], + "spans": [ + { + "bbox": [ + 315, + 651, + 548, + 703 + ], + "type": "text", + "content": "[27] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 6" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 735, + 317, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 735, + 317, + 743 + ], + "spans": [ + { + "bbox": [ + 294, + 735, + 317, + 743 + ], + "type": "text", + "content": "16553" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 82, + 296, + 704 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 63, + 82, + 296, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 82, + 296, + 113 + ], + "spans": [ + { + "bbox": [ + 63, + 82, + 296, + 113 + ], + "type": "text", + "content": "[28] Reiichiro Nakano. Neural painters: A learned differentiable constraint for generating brushstroke paintings. arXiv preprint arXiv:1904.08410, 2019. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 114, + 296, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 114, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 64, + 114, + 296, + 156 + ], + "type": "text", + "content": "[29] Nadav Orzech, Yotam Nitzan, Ulysse Mizrahi, Dov Danon, and Amit H Bermano. Masked extended attention for zero-shot virtual try-on in the wild. arXiv preprint arXiv:2406.15331, 2024. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 64, + 158, + 296, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 158, + 296, + 209 + ], + "spans": [ + { + "bbox": [ + 64, + 158, + 296, + 209 + ], + "type": "text", + "content": "[30] Xavier Soria Poma, Edgar Riba, and Angel Sappa. Dense extreme inception network: Towards a robust cnn model for edge detection. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1923-1932, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 212, + 296, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 212, + 296, + 263 + ], + "spans": [ + { + "bbox": [ + 64, + 212, + 296, + 263 + ], + "type": "text", + "content": "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 265, + 296, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 265, + 296, + 305 + ], + "spans": [ + { + "bbox": [ + 64, + 265, + 296, + 305 + ], + "type": "text", + "content": "[32] L Rout, Y Chen, N Ruiz, A Kumar, C Caramanis, S Shakkottai, and W Chu. Rb-modulation: Training-free personalization of diffusion models using stochastic optimal control. 2024. 2, 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 64, + 308, + 296, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 308, + 296, + 349 + ], + "spans": [ + { + "bbox": [ + 64, + 308, + 296, + 349 + ], + "type": "text", + "content": "[33] Chang Wook Seo, Amirsaman Ashtari, and Junyong Noh. Semi-supervised reference-based sketch extraction using a contrastive learning framework. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 1, 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 64, + 350, + 296, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 350, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 64, + 350, + 296, + 392 + ], + "type": "text", + "content": "[34] Edgar Simo-Serra, Satoshi Iizuka, Kazuma Sasaki, and Hiroshi Ishikawa. Learning to simplify: fully convolutional networks for rough sketch cleanup. ACM Transactions on Graphics (TOG), 35(4), 2016. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 64, + 394, + 296, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 394, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 64, + 394, + 296, + 435 + ], + "type": "text", + "content": "[35] Gowthami Somepalli, Anubhav Gupta, Kamal Gupta, Shramay Palta, Micah Goldblum, Jonas Geiping, Abhinav Shrivastava, and Tom Goldstein. Measuring style similarity in diffusion models. arXiv preprint arXiv:2404.01292, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 64, + 437, + 296, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 437, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 64, + 437, + 296, + 468 + ], + "type": "text", + "content": "[36] Xavier Soria, Yachuan Li, Mohammad Rouhani, and Angel D Sappa. Tiny and efficient model for the edge detection generalization. In CVPR, pages 1364-1373, 2023. 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 64, + 469, + 296, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 469, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 64, + 469, + 296, + 521 + ], + "type": "text", + "content": "[37] Zhuo Su, Wenzhe Liu, Zitong Yu, Dewen Hu, Qing Liao, Qi Tian, Matti Pietikainen, and Li Liu. Pixel difference networks for efficient edge detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5117-5127, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 64, + 522, + 296, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 522, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 64, + 522, + 296, + 574 + ], + "type": "text", + "content": "[38] Yael Vinker, Ehsan Pajouheshgar, Jessica Y Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, and Ariel Shamir. Clipasso: Semantically-aware object sketching. ACM Transactions on Graphics (TOG), 41(4): 1-11, 2022. 1, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 64, + 576, + 296, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 576, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 64, + 576, + 296, + 628 + ], + "type": "text", + "content": "[39] Yael Vinker, Yuval Alaluf, Daniel Cohen-Or, and Ariel Shamir. Clipscene: Scene sketching with different types and levels of abstraction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4146–4156, 2023. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 64, + 630, + 296, + 661 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 630, + 296, + 661 + ], + "spans": [ + { + "bbox": [ + 64, + 630, + 296, + 661 + ], + "type": "text", + "content": "[40] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. p+: Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 64, + 662, + 296, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 662, + 296, + 704 + ], + "spans": [ + { + "bbox": [ + 64, + 662, + 296, + 704 + ], + "type": "text", + "content": "[41] Haofan Wang, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv preprint arXiv:2404.02733, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 82, + 547, + 678 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 315, + 82, + 547, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 82, + 547, + 124 + ], + "spans": [ + { + "bbox": [ + 315, + 82, + 547, + 124 + ], + "type": "text", + "content": "[42] Haofan Wang, Peng Xing, Renyuan Huang, Hao Ai, Qixun Wang, and Xu Bai. Instantstyle-plus: Style transfer with content-preserving in text-to-image generation. arXiv preprint arXiv:2407.00788, 2024. 3, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 126, + 547, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 126, + 547, + 157 + ], + "spans": [ + { + "bbox": [ + 316, + 126, + 547, + 157 + ], + "type": "text", + "content": "[43] Matthias Wright and Björn Ommer. Artfid: Quantitative evaluation of neural style transfer. In DAGM German Conference on Pattern Recognition, pages 560-576. Springer, 2022. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 159, + 547, + 211 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 159, + 547, + 211 + ], + "spans": [ + { + "bbox": [ + 316, + 159, + 547, + 211 + ], + "type": "text", + "content": "[44] Zongwei Wu, Liangyu Chai, Nanxuan Zhao, Bailin Deng, Yongtuo Liu, Qiang Wen, Junle Wang, and Shengfeng He. Make your own sprites: Aliasing-aware and cell-controllable pixelization. ACM Transactions on Graphics (TOG), 41(6): 1-16, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 213, + 547, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 213, + 547, + 244 + ], + "spans": [ + { + "bbox": [ + 316, + 213, + 547, + 244 + ], + "type": "text", + "content": "[45] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In Proceedings of the IEEE international conference on computer vision, pages 1395-1403, 2015. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 246, + 547, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 246, + 547, + 287 + ], + "spans": [ + { + "bbox": [ + 316, + 246, + 547, + 287 + ], + "type": "text", + "content": "[46] Peng Xing, Haofan Wang, Yanpeng Sun, Qixun Wang, Xu Bai, Hao Ai, Renyuan Huang, and Zechao Li. Csgo: Content-style composition in text-to-image generation. arXiv preprint arXiv:2408.16766, 2024. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 289, + 547, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 289, + 547, + 332 + ], + "spans": [ + { + "bbox": [ + 316, + 289, + 547, + 332 + ], + "type": "text", + "content": "[47] Chenshu Xu, Yangyang Xu, Huaidong Zhang, Xuemiao Xu, and Shengfeng He. Dreamanine: Learning style-identity textual disentanglement for anime and beyond. IEEE Transactions on Visualization and Computer Graphics, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 334, + 547, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 334, + 547, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 334, + 547, + 374 + ], + "type": "text", + "content": "[48] Rui Yang, Xiaojun Wu, and Shengfeng He. Mixsa: Training-free reference-based sketch extraction via mixture-of-self-attention. IEEE Transactions on Visualization and Computer Graphics, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 376, + 547, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 547, + 418 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 547, + 418 + ], + "type": "text", + "content": "[49] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 420, + 547, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 420, + 547, + 483 + ], + "spans": [ + { + "bbox": [ + 316, + 420, + 547, + 483 + ], + "type": "text", + "content": "[50] Yuyang Yu, Bangzhen Liu, Chenxi Zheng, Xuemiao Xu, Huaidong Zhang, and Shengfeng He. Beyond textual constraints: Learning novel diffusion conditions with fewer examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7109-7118, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 485, + 547, + 527 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 485, + 547, + 527 + ], + "spans": [ + { + "bbox": [ + 316, + 485, + 547, + 527 + ], + "type": "text", + "content": "[51] Sicong Zang, Shikui Tu, and Lei Xu. Self-organizing a latent hierarchy of sketch patterns for controllable sketch synthesis. IEEE Transactions on Neural Networks and Learning Systems, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 529, + 547, + 571 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 529, + 547, + 571 + ], + "spans": [ + { + "bbox": [ + 316, + 529, + 547, + 571 + ], + "type": "text", + "content": "[52] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 572, + 547, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 572, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 572, + 547, + 624 + ], + "type": "text", + "content": "[53] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 626, + 547, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 626, + 547, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 626, + 547, + 678 + ], + "type": "text", + "content": "[54] Caixia Zhou, Yaping Huang, Mengyang Pu, Qingji Guan, Ruoxi Deng, and Haibin Ling. Muge: Multiple granularity edge detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25952-25962, 2024. 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 735, + 317, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 735, + 317, + 743 + ], + "spans": [ + { + "bbox": [ + 294, + 735, + 317, + 743 + ], + "type": "text", + "content": "16554" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Stronger, Steadier & Superior_ Geometric Consistency in Depth VFM Forges Domain Generalized Semantic Segmentation/2e6ee106-ae7d-42e3-b798-bd6f97fafe0a_content_list.json b/2025/Stronger, Steadier & Superior_ Geometric Consistency in Depth VFM Forges Domain Generalized Semantic Segmentation/2e6ee106-ae7d-42e3-b798-bd6f97fafe0a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1ad2241e2bf0e957afcb8911460325ad54b4c7ad --- /dev/null +++ b/2025/Stronger, Steadier & Superior_ Geometric Consistency in Depth VFM Forges Domain Generalized Semantic Segmentation/2e6ee106-ae7d-42e3-b798-bd6f97fafe0a_content_list.json @@ -0,0 +1,1572 @@ +[ + { + "type": "text", + "text": "Stronger, Steadier & Superior: Geometric Consistency in Depth VFM Forges Domain Generalized Semantic Segmentation", + "text_level": 1, + "bbox": [ + 109, + 128, + 887, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Siyu Chen $^{1,3}$ Ting Han $^{2\\dagger}$ Changshe Zhang $^{4}$ Xin Luo $^{1}$ Meiliu Wu $^{3}$ Guorong Cai $^{1}$ Jinhe Su $^{1\\dagger}$ $^{1}$ Jimei University $^{2}$ Sun Yat-sen University $^{3}$ University of Glasgow $^{4}$ Xidian University", + "bbox": [ + 117, + 202, + 872, + 247 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b00ec6b05b120f399e8146a8fa271322bbfc3d8c0da48e0c6e092787c307ac9a.jpg", + "image_caption": [ + "(a) Stronger Pre-trained Models" + ], + "image_footnote": [], + "bbox": [ + 112, + 261, + 274, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4b8dec8ad21a3f5f4285c0fa8d8824cb72327690bc80e1b3879c5d4fa7760631.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 261, + 426, + 327 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5d1f42bc025ad9277429553c8f6547319aa59db64ed5a11beb31ae6bb689edba.jpg", + "image_caption": [ + "Figure 1. Existing methods fail to fully exploit the potential of VFMs for DGSS when visual cues are limited or absent. To this end, we introduce a novel and robust fine-tuning strategy DepthForge that leverages depth cues from a depth VFM to optimize visual cues, and employs spatial consistency to enhance feature discriminability. The proposed DepthForge achieves (a) stronger performance under extreme conditions, owing to (b) steadier visual-spatial attention, and thus delivers (c) superior generalization ability." + ], + "image_footnote": [], + "bbox": [ + 281, + 329, + 426, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6b7cc3a8b6410ec6a46f86182d827b14a6e730d2133e2b0559445cb0c8f1b357.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 431, + 261, + 578, + 327 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5bebe333f583b07bafc4a84625e25934c5b7fc473d5194db273edf4d47e57c63.jpg", + "image_caption": [ + "(b) Steadier Visual-Spatial Attention" + ], + "image_footnote": [], + "bbox": [ + 431, + 329, + 578, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/404c9c4bd65d4196835cabc581088fc71527c3762e1996a9761927d2e4776298.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 261, + 736, + 327 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/35721a10669bd2715a1940fb6a0bcbda95f703b66e45d88b96b575086c8151cb.jpg", + "image_caption": [ + "(c) Superior Generalization Ability" + ], + "image_footnote": [], + "bbox": [ + 589, + 329, + 736, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e68f05c791d5149a7865b6aab14ec2743329fa1491a4c52e7625a5282c428f79.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 743, + 261, + 890, + 327 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/01e80c623aa943409eb7233995ac28fcab93ae80065a4711786ac081201c27a6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 743, + 329, + 890, + 393 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 488, + 326, + 503 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision Foundation Models (VFMs) have delivered remarkable performance in Domain Generalized Semantic Segmentation (DGSS). However, recent methods often overlook the fact that visual cues are susceptible, whereas the underlying geometry remains stable, rendering depth information more robust. In this paper, we investigate the potential of integrating depth information with features from VFMs, to improve the geometric consistency within an image and boost the generalization performance of VFMs. We propose a novel fine-tuning DGSS framework, named Depth-Forge, which integrates the visual cues from frozen DINOv2 or EVA02 and depth cues from frozen Depth Anything V2. In each layer of the VFMs, we incorporate depth-aware learnable tokens to continuously decouple domain-invariant visual and spatial information, thereby enhancing depth awareness and attention of the VFMs. Finally, we develop a depth refinement decoder and integrate it into the model architecture to adaptively refine multi-layer VFM features and depth-aware learnable tokens. Extensive experiments are conducted based on various DGSS settings and five different datasets as unseen target domains. The qualitative and quantitative results demonstrate that our method significantly outperforms alternative approaches with stronger performance, steadier visual-spatial attention, and superior generalization ability. In particular,", + "bbox": [ + 88, + 523, + 485, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "DepthForge exhibits outstanding performance under extreme conditions (e.g., night and snow). Code is available at https://github.com/SY-Ch/DepthForge.", + "bbox": [ + 511, + 489, + 906, + 535 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 558, + 645, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Domain Generalized Semantic Segmentation (DGSS) aims to improve prediction accuracy across multiple unseen domains without requiring access to their data, thereby ensuring robust generalization for practical applications [34, 45, 53]. One common approach involves decomposing the learned features into domain-invariant and domain-specific components [42, 47]. This method aims to achieve robustness against domain variations by isolating content-related factors from those that vary between domains. A second strategy employs meta-learning techniques [24], which focus on training models to generalize effectively across domains by learning more adaptable representations, often leveraging shared structures across different datasets.", + "bbox": [ + 509, + 583, + 906, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, VFMs have emerged as a promising approach, primarily owing to their strong generalization capabilities, which are enabled by pre-trained on large-scale datasets [15, 25, 36]. Fine-tuning these models has become increasingly popular due to their cost-effectiveness and superior performance [2, 44]. However, a significant challenge remains: what constitutes robust feature information for domain generalization, and how can such robust informa", + "bbox": [ + 509, + 779, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8285", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion be effectively extracted? Previous methods predominantly utilize VFMs based on RGB images. However, the inherent limitations of RGB prevent those methods from adapting to variable conditions, especially in scenarios such as nighttime, overexposure, snow, or fog, where visual cues are not clearly present. This clearly fails to meet the requirements of domain-generalized semantic segmentation.", + "bbox": [ + 88, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce DepthForge, a robust fine-tune and optimization strategy for domain generalized semantic segmentation, as shown in Fig. 1. Specifically, we employ two frozen VFMs: Depth Anything V2[49, 50] is dedicated to extract depth information, and DINOv2[30] or EVA02[15] are used to process RGB visual features. We enhance the generalization of the visual features at each layer of the VFMs by integrating the corresponding depth cues. More importantly, we introduce depth-aware learnable tokens that enables the model to learn spatial structure invariance, thereby ensuring consistent performance across different domains. The depth and visual features extracted from frozen VFMs are used to optimize the representation of our learnable tokens, thus producing more robust visual-spatial attention features than those based solely on visual cues and forging geometric consistency, as shown in Fig. 1(b). However, we observe the frozen VFMs only provide static features, they are incapable of refining the feature maps during training. This limitation leads to persistent errors in the learnable tokens, causing the optimization process to drift in incorrect directions that are difficult to rectify. To this end, we propose a depth refinement decoder that dynamically adjusts the multi-scale features, combining prior and dynamic features to establish high-quality paired feature relationships at different depth spaces.", + "bbox": [ + 93, + 200, + 485, + 577 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our core contribution is to demonstrate that robust depth features and learnable tokens can be effectively applied to improve model performance in Domain Generalized Semantic Segmentation (DGSS). To the best of our knowledge, the proposed DepthForge is the first depth-aware framework specifically designed for DGSS. The effectiveness of our design is verified by experiments on widely used Cityscapes, ACDC, Mapillary, BDD100k, and GTA5 datasets. In addition, the individual components of our design are also verified by extensive experiments. Extensive experiments have demonstrated that our method significantly improves DGSS performance, as shown in Fig. 1(c). Specifically, in the $GTA \\rightarrow$ Cityscapes + BDD100k + Mapillary setting, we observed approximately $+4\\%$ significant improvement in mIoU, particularly the improvement in extreme scenes is close to $+5\\%$ . Our main contributions are summarized as follows:", + "bbox": [ + 89, + 580, + 482, + 835 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a stronger DepthForge learning framework to fine-tune VFMs combined with visual and depth cues for domain-generalized semantic segmentation.", + "- A depth-awareness learnable tokens are introduced to" + ], + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5d2b1bdf4d83ad194b61d9a94ed479f728e9cc9ff9a658e425fb35625df46509.jpg", + "image_caption": [ + "Figure 2. Different configurations of incorporating depth cues into visual information, where Config.1: adding $f_{i}^{d}$ to $f_{i}^{v}$ , or Config.2: incorporating $f_{i}^{d}$ into the learnable token $T_{i}$ , and Config.3: our DepthForge, respectively." + ], + "image_footnote": [], + "bbox": [ + 514, + 87, + 908, + 238 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "focus on spatial relationship that acquires steadier and more robust visual-spatial attention. This strategy enables DepthForge to enhance the discriminability and refine features at an instance level within each layer.", + "bbox": [ + 524, + 321, + 903, + 380 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Depth Forge highlights the impressive generalization ability, surpassing existing DGSS methods by a large margin, especially in extreme scenes with limited and absent visual cues.", + "bbox": [ + 513, + 381, + 903, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 468, + 653, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Domain Generalized Semantic Segmentation", + "text_level": 1, + "bbox": [ + 511, + 496, + 890, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Domain Generalized Semantic Segmentation (DGSS) aims to enhance the generalization capability of segmentation models to unseen domains. Recent methods primarily focus on reducing feature distribution discrepancies across domains and learning more domain-invariant representations. Adversarial training [53], meta-learning [4, 41], and self-supervised learning [18, 56] have been extensively explored to improve generalization performance. However, these methods often rely on constructing cross-domain contrastive constraints or employing sophisticated training strategies, limiting their generalization improvements and complicating the training process. The emergence of Vision Foundation Models (VFMs), trained via large-scale image-text [36] or image-image contrastive learning [30], has opened a promising avenue for DGSS. VFMs capture universal and rich visual representations beneficial for reducing domain biases. Recent approach [2, 44] leverages pretrained VFMs by employing their frozen features as powerful generalized backbones to extract and refine learnable tokens, leading to significant improvements in model performance. Nevertheless, current DGSS approaches predominantly utilize RGB-only data, leaving the models susceptible to variations in lighting, texture, and spatial arrangements across different domains, thus constraining their full potential for generalization.", + "bbox": [ + 511, + 523, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8286", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4221b9074ef45cc30b2139057816e8c20379cf5ff037090ebabac7ba3eca537e.jpg", + "image_caption": [ + "Figure 3. DepthForge framework. It incorporates the features from frozen VFM and depth VFM into our Depth-aware Learnable Tokens to generate enhanced features $f_{i+1} = f_i^v + \\text{DepthForge}(f_i^v, f_i^d)$ for each layer. In our tunable DepthForge, We first design a Depth Awareness module that integrates visual and depth cues into learnable tokens. Next, we propose an Attention Optimization module to strengthen the discriminability of instances across different spatial locations. Finally, multi-scale features are fused in Depth Refinement Decoder to adapt to various scales and spatial connections. Our DepthForge establishes direct connections to different instances in real-world locations, facilitating both feature refinement and improved discriminability." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 903, + 373 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. RGB-D Segmentation", + "text_level": 1, + "bbox": [ + 89, + 477, + 294, + 493 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Combining RGB with depth information has shown substantial benefits across various visual tasks [6, 8, 9, 17, 39]. In object detection, RGB-D effectively mitigates limitations inherent to RGB data under complex scenes by providing robust geometric and spatial cues, thereby enhancing localization accuracy and model robustness [43, 59]. In semantic segmentation, depth information assists models in capturing spatial structures and geometric contours of objects, alleviating issues related to imprecise object boundary delineation. Particularly in scenarios with blurred boundaries, insufficient illumination, or texture scarcity, the incorporation of depth information substantially improves segmentation accuracy and stability [14, 48, 54, 60]. Recent studies have further indicated that even in domain generalization tasks, integrating RGB and depth features consistently improves the model's generalization performance on unseen domains [19, 27, 38, 55]. Therefore, exploring the integration of depth information to address the limitations of RGB-only approaches represents a promising and essential research direction within DGSS.", + "bbox": [ + 88, + 500, + 485, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 815, + 217, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve better generalization performance, existing approaches opt to fine-tune VFMs in a parameter-efficient manner. One mainstream idea is to refine and propagate features from frozen VFM layers to subsequent ones. Specif", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ically, given a pre-trained VFM composed of $N$ sequential layers $\\{L_1,L_2,\\dots,L_N\\}$ , where each layer is associated with a weight matrix $W_{i}\\in \\mathbb{R}^{c\\times c}$ and the frozen output features of each layer are denoted as $f_{i}$ for $i = 1,2,\\ldots ,N$ . If we have a learnable weight matrix as $\\Delta W_{i}\\in \\mathbb{R}^{c\\times c}$ , the feature propagation from layer $L_{i}$ to $L_{i + 1}$ is then formulated as:", + "bbox": [ + 511, + 478, + 906, + 570 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i + 1} = W _ {i} f _ {i} + \\Delta W _ {i} f _ {i}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 582, + 906, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent REIN transfers the learnable matrix $\\Delta W_{i}$ into a learnable token $T_{i}$ :", + "bbox": [ + 511, + 614, + 906, + 643 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i + 1} = W _ {i} f _ {i} + \\epsilon \\left(T _ {i} \\left(W _ {i} f _ {i}\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 657, + 906, + 675 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon (\\cdot)$ denotes the MLP function, and $T_{i}\\in \\mathbb{R}^{m\\times c}$ with a tokn length $m$ . REIN enables each token to connect better with instances in an image. However, we observe that since the updated and frozen features are limited to remain similar, this correlation is fragile and the capacity for feature refinement is limited.", + "bbox": [ + 511, + 686, + 905, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This prompted us to explore what kind of information could improve discrimination in instances and refine the feature representations. Drawing inspiration from multimodal learning, we observe that depth features provide a spatial relationship bias for corresponding pixels in visual features, thereby compensating for the loss of crucial visual cues in DGSS. We assume that incorporating a pretrained depth VFM into the pre-trained VFMs will alleviate", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8287", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/017679844a38334d5884322e17ece2e7a00458b4f1af824afd24b8036fb957ff.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodProc. & YearSnowNightFogRain
ResNet based:
IBN [31]ECCV201849.621.263.850.4
Itenorm [21]CVPR201949.923.863.350.1
IW [32]CVPR201947.621.862.452.4
RobustNet [10]CVPR202149.824.364.356.0
WildNet [26]CVPR202228.412.741.234.2
Transformer based:
HGFormer [13]CVPR202368.652.769.972.0
CMFormer [3]AAAI202464.333.777.867.6
VFM based:
REIN [44]CVPR202470.655.979.572.5
FADA [2]NeurIPS202473.557.480.275.0
DepthForge(Ours)-75.460.481.275.4
", + "bbox": [ + 114, + 88, + 460, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "attention bias and significantly enhance scene understanding, particularly under extreme conditions.", + "bbox": [ + 89, + 380, + 482, + 410 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use the frozen features $f_{i}^{v}$ from VFM for visual embedding, and the frozen features $f_{i}^{d}$ from depth VFM for depth embedding. A straightforward thought might simply add $f_{i}^{d}$ to $f_{i}^{v}$ , or incorporate $f_{i}^{d}$ into the leanable token $T_{i}$ . However, these methods lack flexibility and introduce significant inter-layer feature differences that prevent convergence, as shown in Fig. 2. In contrast, we propose Depth-Forge to adaptively fuse visual and depth cues, denoted as:", + "bbox": [ + 89, + 410, + 483, + 532 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i + 1} = W _ {i} ^ {v} f _ {i} + \\varepsilon \\left(\\epsilon^ {v} \\left(T _ {i} \\left(W _ {i} ^ {v} f _ {i} ^ {v}\\right)\\right) + \\epsilon^ {d} \\left(T _ {i} \\left(W _ {i} ^ {d} f _ {i} ^ {d}\\right)\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 555, + 482, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W_{i}^{v}$ and $W_{i}^{d}$ represent the weight matrix from VFMs and depth VFM, respectively. $\\varepsilon$ denotes the adaptation of the visual and depth signals. This approach allows us to more effectively utilize the powerful capacities of multimodal VFMs to forge robust refined features and instance discriminability.", + "bbox": [ + 89, + 583, + 483, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "More importantly, the depth VFM itself possesses strong generalization capabilities, allowing domain adaptation without fine-tuning. A pre-trained VFM applies the same. Therefore, the overall optimization objective still focuses on the segmentation head $\\mathcal{H}$ with parameter $\\theta_{h}$ and our Depth-Forge with parameter $\\theta_{D}$ :", + "bbox": [ + 89, + 674, + 483, + 765 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\arg \\min _ {\\theta_ {D}, \\theta_ {h}} \\sum_ {i = 1} ^ {N} \\mathcal {L o s s} \\left(\\mathcal {H} _ {\\theta_ {i}} \\left(\\mathcal {F} _ {\\Theta^ {v}, \\Theta^ {d}, \\theta_ {D}} \\left(x _ {i}\\right)\\right), y _ {i}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 773, + 482, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x_{i}$ and $y_{i}$ denote the input image and its corresponding ground truth, respectively, and $N$ signifies the total number of images. $\\mathcal{F}$ represent the forward process of our DepthForge strategy applied to the pre-trained VFMs with parameters $\\Theta^v$ and $\\Theta^d$ .", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/4df0cae1930e64bc5f0be3f8e9ae4533cc707ea2acafaf0454604357da86158f.jpg", + "table_caption": [ + "Table 1. Performance comparison between our DepthForge and existing DGSS methods under Cities. $\\rightarrow$ ACDC with Snow, Night, Fog, Rain generalization setting. Top three results are highlighted as best, second, and third, respectively. (%)" + ], + "table_footnote": [], + "table_body": "
MethodProc. & YearBDD100kMapillaryGTA
ResNet based:
IBN [31]ECCV201848.5657.0445.06
Itenorm [21]CVPR201949.2356.2645.73
IW [32]CVPR201948.4955.8244.87
RobustNet [10]CVPR202150.7358.6445.00
DRPC [52]ICCV201949.8656.3445.62
GTR [33]TIP202150.7557.1645.79
SHADE [57]ECCV202250.9560.0748.61
SAW [35]CVPR202252.9559.8147.28
WildNet [26]CVPR202250.9458.7947.01
BlindNet [1]CVPR202451.8460.1847.97
Transformer based:
HGFormer [13]CVPR202353.4066.9051.30
CMFormer [3]AAAI202459.2771.1058.11
VFM based:
REIN [44]CVPR202463.5474.0362.41
FADA [2]NeurIPS202465.1275.8663.78
DepthForge(Ours)-66.1975.9367.24
", + "bbox": [ + 517, + 88, + 905, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2. Performance comparison between our DepthForge and existing DGSS methods under the Citys. $\\rightarrow$ BDD. $^+$ Map. $^+$ GTA5 generalization setting. Top three results are highlighted as best, second, and third, respectively. $(\\%)$", + "bbox": [ + 513, + 357, + 906, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. DepthForge", + "text_level": 1, + "bbox": [ + 513, + 436, + 638, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This paper introduces the DepthForge strategy, a novel approach for domain generalized semantic segmentation. As shown in Fig. 3, DepthForge consists of three main parts: (1) Depth Awareness; (2) Attention Optimization; (3) Depth Refinement Decoder. We detail the designs as follows.", + "bbox": [ + 511, + 462, + 906, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Depth Awareness", + "text_level": 1, + "bbox": [ + 511, + 547, + 684, + 563 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recent DGSS methods have applied pre-trained VFMs to source domain segmentation, with the core idea being to enhance the features between frozen VFM layers to help the model better adapt to target tasks. However, we observe that visual cues exhibit significant variations and even deficiencies across different scenes. Consequently, a more robust cue is required to provide a stronger prior for domain generalization. This motivates us to explore a strategy that focuses on the inherent similarities across different domains. In general, spatial location priors are able to help differentiate instances even when visual cues fail. To this end, we employ an additional depth VFM as an auxiliary network, as shown in Fig. 3.", + "bbox": [ + 509, + 568, + 906, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, given a VFM $\\mathcal{V}$ and a depth VFM $\\mathcal{D}$ with $N$ layers $\\{L_1,L_2,\\dots ,L_N\\}$ , the corresponding features are denoted as $\\{f_i^v,f_i^d |f_i\\in \\mathbb{R}^{n\\times c},i\\in [1,N]\\}$ . Note that the layers $\\{L_1,L_2,\\dots ,L_N\\}$ are kept frozen. We acquire the visual feature $f_{i}^{v}$ from the $i$ -th layer $L_{i}$ of $\\mathcal{V}$ , and the depth feature $f_{i}^{d}$ from the corresponding layer of $\\mathcal{D}$ . Our goal is to generate the enhanced feature $\\Delta f_{i} = DepthForge(f_{i}^{v},f_{i}^{d})$ by using both visual and depth cues.", + "bbox": [ + 511, + 763, + 906, + 885 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We employ a set of learnable tokens $T = \\{T_{i} \\in$", + "bbox": [ + 532, + 885, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8288", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/005b469f1ef48242c248d16711005b6f0c8e73a849a4d4acb67b085487787479.jpg", + "image_caption": [ + "Figure 4. Key segmentation examples of existing DGSS methods and DepthForge under the Cityscapes to ACDC unseen target domains under Night, Snow, Rain, and Fog conditions." + ], + "image_footnote": [], + "bbox": [ + 119, + 90, + 883, + 354 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fbce90875e1e88a2a3ebf89c75055fd6fe37b0abc305baf7fd944ed0ca66f19a.jpg", + "image_caption": [ + "Figure 5. Key segmentation examples of existing DGSS methods and DepthForge under the GTA5 to Mapillary, BDD100k, and Cityscapes unseen target domains setting." + ], + "image_footnote": [], + "bbox": [ + 119, + 391, + 888, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathbb{R}^{m \\times c} |i \\in [1, N]\\}$ to adaptively learn scene knowledge. As we consider, relying solely on visual cues can be limiting and prone to failure. Therefore, we utilize visual and depth cues to construct learnable tokens, which we term depth-aware learnable tokens. We integrate $T_i$ with $f_i^v$ and $f_i^d$ using an attention-based mechanism, prompting tailored adjustments to bridge the gap between different domains.", + "bbox": [ + 88, + 584, + 482, + 689 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use $f_{i}^{v,d}$ as the query, and $T_{i}$ as the key and the value A dot-product operation is applied to generate awareness map $A_{i}^{v,d}$ , and the Softmax function (S) is employed to align each patch with a unique instance. The $A_{i}^{v,d}$ provides insights into both the visual cues and depth retrieval, qualitatively evaluating the relationship between various tokens and visual or depth cues, respectively. Subsequently, we incorporate depth awareness bias $A_{i}^{d}$ into the visual features $A_{i}^{v}$ to provide additional spatial relationships, thereby establishing a latent connection between pixels and the real world. The process is mathematically formulated as:", + "bbox": [ + 89, + 689, + 482, + 859 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA _ {i} = S \\left(\\frac {f _ {i} ^ {v} \\times (T ^ {v}) _ {i} ^ {\\top}}{\\sqrt {c}}\\right) + \\lambda S \\left(\\frac {f _ {i} ^ {d} \\times (T ^ {d}) _ {i} ^ {\\top}}{\\sqrt {c}}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 869, + 482, + 905 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Attention Optimization", + "text_level": 1, + "bbox": [ + 511, + 584, + 728, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Directly leveraging the awareness representation $A_{i}$ to generate enhanced features may introduce a significant amount of irrelevant information for downstream tasks. To this end, we employ depth-aware learnable tokens $T_{i}$ to learn an alignment weight $W_{T_i}$ and bias $b_{T_i}$ . The $T_{i}$ with weight $W_{T_i}$ and bias $b_{T_i}$ is then multiplied with the $A_{i}$ to obtain the initial enhanced features $\\Delta f_{i}$ that embed both visual and depth cues, using the equation:", + "bbox": [ + 511, + 606, + 906, + 729 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\hat {f} _ {i} = A _ {i} \\times \\left(T _ {i} \\times W _ {T _ {i}} + b _ {T _ {i}}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 737, + 903, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we select the same optimization strategy as REIN, discarding high-weight features when the features are sufficiently precise to avoid unnecessary adjustments and mitigate the risk of inappropriate variations. Finally, we employ two consecutive residual connections $\\phi$ to enhance the flexibility of attention feature adjustment and prevent gradient vanishing during the forward pass. The calculation is defined as $\\Delta f_{i} = \\phi \\hat{\\Delta f}_{i}$ . Benefiting from these depth-aware and attention optimization adjustments, DepthForge is ca", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8289", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/39a093aa7a9e91b204c2935320ba498161e2e34e4f4f84c4276e478e503fc41c.jpg", + "image_caption": [ + "Figure 6. t-SNE visualization compares REIN and our Depth-Forge. The first volume corresponds to $\\text{Citys.} \\rightarrow \\text{ACDC}$ , and two through four volumes correspond to $\\text{GTA5} \\rightarrow \\text{BDD.} + \\text{Citys.} + \\text{Map}$ . The best mIoU results are highlighted in red." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 480, + 255 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "pable of generating more distinctive, discriminative features for different categories within a single image, while balancing both visual and depth cues.", + "bbox": [ + 89, + 328, + 483, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Depth Refinement Decoder", + "text_level": 1, + "bbox": [ + 89, + 382, + 334, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The depth-aware features at different layers are inconsistent. Therefore, how to design a robust decoder becomes a core challenge that effectively integrate these enhanced features across various scales to accomplish downstream tasks.", + "bbox": [ + 89, + 404, + 483, + 464 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Previous methods rely solely on the output of the final layer $f_{N}$ for semantic segmentation predictions. In contrast, we fuse the outputs from layers $L_{1} - L_{N}$ to refine the depth features. Specifically, we process these multi-layer features with an MLP with two different fully-connected layers $\\phi$ and $\\varphi$ and then concatenate them after a ReLU activation function:", + "bbox": [ + 89, + 465, + 483, + 568 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i} ^ {\\text {o u t}} = \\phi (R e L U (\\varphi f _ {i})), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 569, + 480, + 585 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {u} = \\gamma \\operatorname {C o n c a t} \\left(f _ {1} ^ {\\text {o u t}}, f _ {2} ^ {\\text {o u t}}, \\dots , f _ {N} ^ {\\text {o u t}}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 606, + 480, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\gamma$ is a convolutional layer. Finally, we design several sequential multi-head Transformer layers to generate the final predictions.", + "bbox": [ + 89, + 628, + 483, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 89, + 686, + 215, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Datasets & Evaluation Protocols", + "text_level": 1, + "bbox": [ + 89, + 712, + 375, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our DepthForge on four real-world datasets (Cityscapes [12], BDD100K [51], Mapillary [28], ACDC [40]) and a synthetic dataset (GTA5 [37]). Cityscapes (Citys.) is an autonomous driving dataset comprising 2,975 training images and 500 validation images, each with a resolution of $2048 \\times 1024$ . BDD100K (BDD.) and Mapillary (Map.) provide 1,000 and 2,000 validation images, with resolutions of $1280 \\times 720$ and $1902 \\times 1080$ , respectively. ACDC offers 406 validation images captured under extreme conditions: namely at Night, Snow, Fog, and Rain, each with a resolution of $1902 \\times 1080$ . GTA5 is a", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fea490b5d66347f141408c20a2898c34233dca6dd9c8a1b2165c2e8f59d79662.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodProc. & YearCityscapesBDD100kMapillary
ResNet based:
DRPC [52]ICCV201937.4232.1434.12
GTR [33]TIP202137.5333.7534.52
RobustNet [10]CVPR202136.5835.2040.33
DIRL [46]AAAI202241.0439.1541.60
SHADE [57]ECCV202244.6539.2843.34
WildNet [26]CVPR202344.6238.4246.09
SAW [35]CVPR202239.7537.3441.86
AdvStyle [58]NeurIPS202239.6235.5437.00
SPC [22]CVPR202344.1040.4645.51
BlindNet [1]CVPR202445.7241.3247.08
Transformer based:
CMFormer [3]AAAI202455.3149.9160.09
VFM based:
DIDEX [29]WACV202462.0054.3063.00
REIN [44]CVPR202466.4060.4066.10
FADA [2]NeurIPS202468.2361.9468.09
DepthForge(Ours)-69.0462.8269.22
", + "bbox": [ + 524, + 89, + 897, + 327 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Performance comparison between our DepthForge and existing DGSS methods under the $GTA5 \\rightarrow Citys$ . + BDD. + Map. generalization setting. Top three results are highlighted as best, second, and third, respectively. (%)", + "bbox": [ + 511, + 330, + 903, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "synthetic dataset that provides 24,966 annotated images obtained from the game.", + "bbox": [ + 511, + 407, + 903, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following the evaluation protocol of the existing DGSS methods, a certain dataset is chosen as the source domain for training, and the other three datasets serve as unseen target domains for validation. The three common evaluation settings are as follows: (1) Cityscapes $\\rightarrow$ ACDC (Night, Snow, Fog, Rain); (2) GTA5 $\\rightarrow$ Cityscapes, BDD100K, Mapillary; and (3) Cityscapes $\\rightarrow$ BDD100K, Mapillary, GTA5. The evaluation metric is the mean Intersection-over-Union (mIoU).", + "bbox": [ + 511, + 438, + 905, + 573 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Deployment Details & Parameter Settings", + "text_level": 1, + "bbox": [ + 511, + 585, + 870, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our implementation is built on the MMSegmentation [11] codebase. For optimization, we employ AdamW optimizer with an initial learning rate of 1e-4, a weight decay of 0.05, an epsilon of 1e-8, and beta values of (0.9, 0.999). A OneCycleLR scheduler is applied over 50,000 total steps with a maximum learning rate of 1e-4, featuring a warmup phase covering $10\\%$ of the training iterations, and using cosine annealing, with both the div_factor and final_div_factor set to 10.", + "bbox": [ + 511, + 608, + 905, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data augmentation includes multi-scale resizing, random cropping (with a fixed crop size and category ratio constraints), random flipping, and photometric distortion. All experiments are conducted on an NVIDIA RTX 3090 GPU.", + "bbox": [ + 511, + 744, + 903, + 805 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Comparison with State-of-The-Art Methods", + "text_level": 1, + "bbox": [ + 511, + 816, + 885, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our approach with existing DGSS methods, including: (1) ResNet-based methods, (2) Transformer-based methods, and (3) VFM-based methods.", + "bbox": [ + 511, + 839, + 903, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cityscapes $\\rightarrow$ ACDC. Tab. 1 presents the performance", + "bbox": [ + 531, + 885, + 903, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8290", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cb236a937283ea96528763b2ca03ae69e2fac2235b56c87c7c1c71c50eefdce0.jpg", + "image_caption": [ + "Figure 7. Visualizations of attention maps and affinity maps under differnet configurations, where DRD denotes the Depth Refinement Decoder, DA and AO represent the Depth Awareness and Attention Optimization, respectively." + ], + "image_footnote": [], + "bbox": [ + 99, + 89, + 488, + 280 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6b6efc08664d9163e81415a3b3f43b223a1a74295bafb4ffcdc78d73bfdb058b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 90, + 903, + 280 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b7ab8cb883c8c521562b4f05a5833e24b649a11cf02ecda82b8d6cd81a7fcbea.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BackboneConfigurationsSnowNightFogRainBDDMap
EVA02 (Large) [15, 16]REIN [44]66.052.176.069.560.070.0
+ concat fi d66.352.276.269.860.270.2
+ Prompt Depth Anything65.852.475.569.360.469.9
+ Depth Anything V267.253.077.070.260.970.8
+ DRD67.753.577.670.761.271.0
+ DA + AO67.953.777.870.961.571.2
DepthForge69.554.578.572.062.073.0
DINOv2 (Large) [30]REIN [44]70.655.979.572.563.574.0
+ concat fi d70.856.179.472.863.674.2
+ Prompt Depth Anything70.455.579.772.063.873.9
+ Depth Anything V271.556.780.573.364.474.5
+ DRD72.057.280.774.064.974.8
+ DA + AO72.257.480.974.265.175.0
DepthForge75.460.481.275.466.276.0
", + "bbox": [ + 96, + 329, + 483, + 474 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0736485157d26a4de30a5fe070a93e239be869470f8a95767834e9502542057c.jpg", + "table_caption": [ + "Table 4. Ablation studies on component configurations of the proposed DepthForge under Citys. $\\rightarrow$ ACDC with Snow, Night, Fog, Rain + BDD. + Map. generalization setting, where DRD denotes the Depth Refinement Decoder, DA and AO represent the Depth Awareness and Attention Optimization. Top three results are highlighted as best, second, and third, respectively. (%)" + ], + "table_footnote": [], + "table_body": "
No.DesignAvg. Gain (%)RainNightSnowRog
1No Depth-70.655.979.572.5
2Concatenation-1.169.855.078.371.2
3Depth Token1.672.357.880.774.0
4Config10.270.856.179.472.8
5Config20.771.556.880.173.0
6Config3, Ours3.575.460.481.275.4
7w/o Scale Factor2.874.859.380.974.8
", + "bbox": [ + 94, + 568, + 480, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "on the Cityscapes source domain and the ACDC target domain. Compared with the VFM-based REIN and FADA, the proposed DepthForge shows mIoU improvement of $4.8\\%$ & $1.9\\%$ , $4.5\\%$ & $3.0\\%$ , $1.7\\%$ & $1.0\\%$ , and $2.9\\%$ & $0.4\\%$ for the Snow, Night, Fog, and Rain scenes, respectively. The feature space of the REIN and the proposed method is visualized in the first volume of Fig. 6. Moreover, our proposed method demonstrates superior performance compared to ResNet-based and Transformer-based approaches. These experimental results underscore the strong and robust generalization ability of DepthForge in extreme scenes with", + "bbox": [ + 89, + 734, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/970e3946c5dccb52e6755ae0a254d4146f4f49931710dc6fb3a66036fa9589e2.jpg", + "table_caption": [ + "Table 5. Performance comparison of the different configurations. Top three results are highlighted as best, second, and third, respectively. (%)" + ], + "table_footnote": [], + "table_body": "
BackboneFine-tune \nMethodTrainable \nParams*mIoUAvg.
CitysBDDMap
CLIP [36] \n(ViT-Large)Full304.15M51.347.654.351.1
Freeze0.00M53.748.755.052.4
REIN2.99M57.154.760.557.4
FADA11.65M58.755.862.158.9
DepthForge(Ours)2.99M59.757.063.560.0
SAM [25] \n(Huge)Full632.18M57.651.761.556.9
Freeze0.00M57.047.158.454.2
REIN4.51M59.652.062.157.9
FADA16.59M61.053.263.460.0
DepthForge(Ours)4.51M63.355.565.161.3
EVA02 [15, 16] \n(Large)Full304.24M62.156.264.660.9
Freeze0.00M56.553.658.656.2
REIN2.99M65.360.564.963.6
FADA11.65M66.761.966.164.9
DepthForge(Ours)2.99M68.061.767.565.7
DINOV2 [30] \n(Large)Full304.20M63.757.464.261.7
Freeze0.00M63.356.163.961.1
REIN2.99M66.460.466.164.3
FADA11.65M68.262.068.166.1
DepthForge(Ours)2.99M69.062.869.267.0
", + "bbox": [ + 519, + 329, + 903, + 611 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Performance comparison of the proposed DepthForge across multiple VFs as backbones under the $GTA5 \\rightarrow Citys$ . + BDD. + Map. generalization setting. Mark * denotes trainable parameters in backbones. Top three results are highlighted as best, second, and third, respectively. (%)", + "bbox": [ + 511, + 613, + 906, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "limited visual cues. Some exemplar segmentation results are compared under the Cityscapes $\\rightarrow$ ACDC are provided in Fig. 4. Our DepthForge shows better pixel-wise predictions than previous methods, demonstrating the effectiveness and significance of depth awareness in our design.", + "bbox": [ + 511, + 703, + 906, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cityscapes $\\rightarrow$ BDD, Map, GTA5. Tab. 2 shows that the proposed DepthForge achieves the state-of-the-art performance, outperforming the REIN and FADA by $2.65\\%$ & $1.07\\%$ , $1.9\\%$ & $0.07\\%$ , and $3.12\\%$ & $1.13\\%$ on the BDD100k, Mapillary, and GTA unseen target domains, respectively. The proposed method achieves an improvement of approximately $10\\%$ compared to ResNet- and Transformer-based approaches.", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8291", + "bbox": [ + 482, + 944, + 513, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4786824ca12e30bf8f749774b1c2ad29a64d87dfdf73da7ff573aecca9cf1bab.jpg", + "table_caption": [], + "table_footnote": [ + "Table 7. Performance comparison of the proposed DepthForge against other DGSS methods under the $GTA5 \\rightarrow Citys$ . + BDD. + Map. generalization setting. Mark * denotes trainable parameters in backbones. Top three results are highlighted as best, second, and third, respectively. (%)" + ], + "table_body": "
BackboneFine-tune MethodTrainable Params*mIoUAvg.
CitysBDDMap
EVA02 (Large) [15, 16]Full304.24M62.156.264.660.9
+AdvStyle [58]304.24M63.156.464.061.2
+PASTA [5]304.24M61.857.163.660.8
+GTR-LTR [33]304.24M59.857.463.260.1
Freeze0.00M56.553.658.656.2
+AdvStyle [58]0.00M51.451.656.553.2
+PASTA [5]0.00M57.852.358.556.2
+GTR-LTR [33]0.00M52.552.857.154.1
+LoRA [20]1.18M55.552.758.355.5
+AdaptFormer [7]3.17M63.759.964.262.6
+VPT [23]3.69M62.257.762.560.8
+REIN [44]2.99M65.360.564.963.6
+FADA [2]11.65M66.761.966.164.9
+DepthForge(Ours)2.99M68.061.767.565.7
DINOv2 (Large) [30]Full304.20M63.757.464.261.7
+AdvStyle [58]304.20M60.858.062.560.4
+PASTA [5]304.20M62.557.264.761.5
+GTR-LTR [33]304.20M62.757.464.561.6
Freeze0.00M63.356.163.961.1
+AdvStyle [58]0.00M61.555.163.960.1
+PASTA [5]0.00M62.157.264.561.3
+GTR-LTR [33]0.00M60.257.762.260.0
+LoRA [20]0.79M65.258.364.662.7
+AdaptFormer [7]3.17M64.959.064.262.7
+VPT [23]3.69M65.259.465.563.3
+REIN [44]2.99M66.460.466.164.3
+FADA [2]11.65M68.262.068.166.1
+DepthForge(Ours)2.99M69.062.869.267.0
", + "bbox": [ + 98, + 88, + 478, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\mathrm{GTA5} \\rightarrow \\mathrm{Citys}$ , Map, BDD. As shown in Tab. 3, the proposed DepthForge achieves mIoU improvements of $2.64\\%$ , $2.42\\%$ , and $3.12\\%$ on the Cityscapes, BDD100k, and Mapillary unseen target domains, respectively. The feature space of REIN and our DepthForge is visualized in Fig. 6, demonstrating the effectiveness of our depth-aware learnable tokens in the visual-spatial attention learning. In addition, we outperform the latest FADA by approximately $1\\%$ across all three target domains. Visual examples for qualitative comparison are given in Fig. 5.", + "bbox": [ + 89, + 560, + 482, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 724, + 254, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Depth awareness and attention optimization are significant for depth-aware learnable tokens. We find that simply incorporating a spatial cue bias into the original learnable tokens $(+f_{i}^{d})$ makes almost no difference. However, when we transform the original learnable tokens into depth-aware learnable tokens using the DepthForge framework, we achieve an approximate $2\\%$ improvement in the target domains ACDC and Mapillary, as shown in Tab. 4. The depth-aware learnable tokens compensate for the shortcomings of visual cues and establish more robust visual-", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "spatial attention. Compared to different configurations, our design allows gradients to be backpropagated through the tokens alone, enables each modality to contribute independently based on its own feature representations, preventing dominance from one during optimization, as shown in Tab. 5 and Fig. 2. This conclusion is supported by qualitative visualizations in Fig. 7.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Depth refinement decoder facilitates the perception of both visual and depth information. As shown in Tab. 4, we observe that depth refinement decoder integrates multiscale information to enhance the expressive capacity of both the frozen VFM features and the enhanced features. With the incorporation of depth information, this optimization effect is further amplified. This conclusion is corroborated by qualitative visualizations in Fig. 7.", + "bbox": [ + 511, + 198, + 903, + 318 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Generalization on Other Setting", + "text_level": 1, + "bbox": [ + 511, + 325, + 795, + 343 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To Different Depth VFMs. We evaluate the performance of the proposed DepthForge across different depth VFMs. Tab. 4 shows the relative spatial relationships derived from Depth Anything v2 outperform the absolute spatial cues provided by Prompt Depth Anything. We believe that what is more crucial in an image is the relative spatial relationship, which establishes connections among pixels and between pixels and the world. In contrast, absolute spatial information introduces significant discrepancies that can act as noise to visual cues.", + "bbox": [ + 511, + 348, + 903, + 498 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To Different VFMs. We evaluate the generalization ability of the proposed DepthForge across different VFMs. Tab. 6 shows the superiority of DepthForge when embedded into these VFMs.", + "bbox": [ + 511, + 500, + 903, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To Token Fine-Tuning Methods. We further compare the proposed method with some existing fine-tuning methods. As shown in Tab. 7, the proposed DepthForge outperforms these methods in all unseen target domains.", + "bbox": [ + 511, + 560, + 903, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 633, + 633, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents the DepthForge framework, a novel approach for robust domain generalization semantic segmentation. We provide a detailed analysis of the challenges and limitations faced by existing methods. To this end, we propose DepthForge to integrate a pretrained depth VFM for fine-tuning. We design depth-aware learnable tokens to provide supplementary spatial cues to the visual features, thereby improving geometric consistency and enhancing the pixel discriminability. Moreover, we incorporate multi-layer enhanced features to effectively capture spatial correlations across various scales. Extensive experiments demonstrate the substantial potential of leveraging multi-modal VFs in the field of DGSS, particularly under extreme conditions where visual cues are limited or missing, validating the effectiveness of our DepthForge in forging VFs for DGSS.", + "bbox": [ + 511, + 659, + 903, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8292", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 256, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the Natural Science Foundation of Xiamen, China, under Grant 3502Z202373036; in part by the National Natural Science Foundation of China under Grant 42371457; in part by the Key Project of Natural Science Foundation of Fujian Province, China, under Grant 2022J02045.", + "bbox": [ + 89, + 114, + 483, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 220, + 187, + 236 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Woo-Jin Ahn, Geun-Yeong Yang, Hyun-Duck Choi, and Myo-Taeg Lim. Style blind domain generalized semantic segmentation via covariance alignment and semantic consistence contrastive learning. In CVPR, pages 3616-3626, 2024. 4, 6", + "[2] Qi Bi, Jingjun Yi, Hao Zheng, Haolan Zhan, Yawen Huang, Wei Ji, Yuexiang Li, and Yefeng Zheng. Learning frequency-adapted vision foundation model for domain generalized semantic segmentation. NeurIPS, 37:94047-94072, 2024. 1, 2, 4, 6, 8", + "[3] Qi Bi, Shaodi You, and Theo Gevers. Learning content-enhanced mask transformer for domain generalized urban-scene segmentation. In AAAI, pages 819-827, 2024. 4, 6", + "[4] Manh-Ha Bui, Toan Tran, Anh Tran, and Dinh Phung. Exploiting domain-specific features to enhance domain generalization. In NeurIPS, pages 21189-21201. Curran Associates, Inc., 2021. 2", + "[5] Prithvijit Chattopadhyay, Kartik Sarangmath, Vivek Vijaykumar, and Judy Hoffman. Pasta: Proportional amplitude spectrum training augmentation for syn-to-real domain generalization. In ICCV, pages 19288-19300, 2023. 8", + "[6] Geng Chen, Qingyue Wang, Bo Dong, Ruitao Ma, Nian Liu, Huazhu Fu, and Yong Xia. Em-trans: Edge-aware multimodal transformer for rgb-d salient object detection. IEEE Transactions on Neural Networks and Learning Systems, 36 (2):3175-3188, 2024. 3", + "[7] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. NeurIPS, 35:16664-16678, 2022. 8", + "[8] Siyu Chen, Ting Han, Changshe Zhang, Weiquan Liu, Jinhe Su, Zongyue Wang, and Guorong Cai. Depth matters: Exploring deep interactions of rgb-d for semantic segmentation in traffic scenes. arXiv preprint arXiv:2409.07995, 2024. 3", + "[9] Siyu Chen, Ting Han, Changshe Zhang, Jinhe Su, Ruisheng Wang, Yiping Chen, Zongyue Wang, and Guorong Cai. Hspformer: Hierarchical spatial perception transformer for semantic segmentation. IEEE Transactions on Intelligent Transportation Systems, 2025. 3", + "[10] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In CVPR, pages 11580-11590, 2021. 4, 6", + "[11] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and" + ], + "bbox": [ + 93, + 246, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "benchmark. https://github.com/open-mmlab/mmsegmentation,2020.6", + "[12] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, pages 3213-3223, 2016. 6", + "[13] Jian Ding, Nan Xue, Gui-Song Xia, Bernt Schiele, and Dengxin Dai. Hgformer: Hierarchical grouping transformer for domain generalized semantic segmentation. In CVPR, pages 15413-15423, 2023. 4", + "[14] Siqi Du, Weixi Wang, Renzhong Guo, Ruisheng Wang, and Shengjun Tang. Asymformer: Asymmetrical cross-modal representation learning for mobile platform real-time rgb-d semantic segmentation. In CVPRW, pages 7608-7615, 2024. 3", + "[15] Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva-02: A visual representation for neon genesis. arXiv preprint arXiv:2303.11331, 2023. 1, 2, 7, 8", + "[16] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In CVPR, pages 19358-19369, 2023. 7, 8", + "[17] Ting Han, Siyu Chen, Chuanmu Li, Zongyue Wang, Jinhe Su, Min Huang, and Guorong Cai. Epurate-net: Efficient progressive uncertainty refinement analysis for traffic environment urban road detection. IEEE Transactions on Intelligent Transportation Systems, 2024. 3", + "[18] Sivan Harary, Eli Schwartz, Assaf Arbelle, Peter Staar, Shady Abu-Hussein, Elad Amrani, Roei Herzig, Amit Alfassy, Raja Giryes, Hilde Kuehne, Dina Katabi, Kate Saenko, Rogerio Feris, and Leonid Karlinsky. Unsupervised domain generalization by learning a bridge across domains. In CVPR, pages 5270-5280, 2022. 2", + "[19] Zongtao He, Liuyi Wang, Ronghao Dang, Shu Li, Qingqing Yan, Chengju Liu, and Qijun Chen. Learning depth representation from rgb-d videos by time-aware contrastive pretraining. IEEE TCSVT, 34(6):4143-4158, 2024. 3", + "[20] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 8", + "[21] Lei Huang, Yi Zhou, Fan Zhu, Li Liu, and Ling Shao. Iterative normalization: Beyond standardization towards efficient whitening. In CVPR, pages 4874-4883, 2019. 4", + "[22] Wei Huang, Chang Chen, Yong Li, Jiacheng Li, Cheng Li, Fenglong Song, Youliang Yan, and Zhiwei Xiong. Style projected clustering for domain generalized semantic segmentation. In CVPR, pages 3061-3071, 2023. 6", + "[23] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, pages 709-727. Springer, 2022. 8", + "[24] Jin Kim, Jiyoung Lee, Jungin Park, Dongbo Min, and Kwanghoon Sohn. Pin the memory: Learning to generalize semantic segmentation. In CVPR, 2022. 1" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8293", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 1, 7", + "[26] Suhyeon Lee, Hongje Seong, Seongwon Lee, and Euntai Kim. Wildnet: Learning domain generalized semantic segmentation from the wild. In CVPR, pages 9936-9946, 2022. 4, 6", + "[27] Zhiyu Liu, Munawar Hayat, Hong Yang, Duo Peng, and Yinjie Lei. Deep hypersphere feature regularization for weakly supervised rgb-d salient object detection. IEEE TIP, 32: 5423-5437, 2023. 3", + "[28] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In ICCV, pages 4990-4999, 2017. 6", + "[29] Joshua Niemeijer, Manuel Schwonberg, Jan-Aike Termöhlen, Nico M Schmidt, and Tim Fingscheidt. Generalization by adaptation: Diffusion-based domain extension for domain-generalized semantic segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2830-2840, 2024. 6", + "[30] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 7, 8", + "[31] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In ECCV, pages 464-479, 2018. 4", + "[32] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In CVPR, pages 1863-1871, 2019. 4", + "[33] Duo Peng, Yinjie Lei, Lingqiao Liu, Pingping Zhang, and Jun Liu. Global and local texture randomization for synthetic-to-real semantic segmentation. IEEE TIP, 30: 6594-6608, 2021. 4, 6, 8", + "[34] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In CVPR, pages 2584-2595, 2022. 1", + "[35] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In CVPR, pages 2594-2605, 2022. 4, 6", + "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 1, 2, 7", + "[37] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In ECCV, pages 102-118. Springer, 2016. 6", + "[38] Giulia Rizzoli, Donald Shenaj, and Pietro Zanuttigh. Source-free domain adaptation for rgb-d semantic segmentation with vision transformers. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops, pages 615-624, 2024. 3" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Giulia Rizzoli, Donald Shenaj, and Pietro Zanuttigh. Source-free domain adaptation for rgb-d semantic segmentation with vision transformers. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 615-624, 2024. 3", + "[40] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In ICCV, pages 10745-10755, 2021. 6", + "[41] Yang Shu, Zhangjie Cao, Chenyu Wang, Jianmin Wang, and Mingsheng Long. Open domain generalization with domain-augmented meta-learning. In CVPR, pages 9619-9628, 2021. 2", + "[42] Zhiqiang Tang, Yunhe Gao, Yi Zhu, Zhi Zhang, Mu Li, and Dimitris Metaxas. Crossnorm and selfnorm for generalization under distribution shifts. In ICCV, 2021. 1", + "[43] Jie Wang, Kechen Song, Yanqi Bao, Liming Huang, and Yunhui Yan. Cgfnet: Cross-guided fusion network for rgb-t salient object detection. IEEE TCSVT, 32(5):2949-2961, 2022. 3", + "[44] Zhixiang Wei, Lin Chen, Yi Jin, Xiaoxiao Ma, Tianle Liu, Pengyang Ling, Ben Wang, Huaian Chen, and Jinjin Zheng. Stronger fewer & superior: Harnessing vision foundation models for domain generalized semantic segmentation. In CVPR, pages 28619-28630, 2024. 1, 2, 4, 6, 7, 8", + "[45] Binhui Xie, Shuang Li, Mingjia Li, Chi Harold Liu, Gao Huang, and Guoren Wang. Sepico: Semantic-guided pixel contrast for domain adaptive semantic segmentation. IEEE TPAMI, 45(7):9004-9021, 2023. 1", + "[46] Qi Xu, Liang Yao, Zhengkai Jiang, Guannan Jiang, Wenqing Chu, Wenhui Han, Wei Zhang, Chengjie Wang, and Ying Tai. Dirl: Domain-invariant representation learning for generalizable semantic segmentation. In AAAI, pages 2884-2892, 2022. 6", + "[47] Qi Xu, Liang Yao, Zhengkai Jiang, Guannan Jiang, Wenqing Chu, Wenhui Han, Wei Zhang, Chengjie Wang, and Ying Tai. Dirl: Domain-invariant representation learning for generalizable semantic segmentation. In AAAI, page 2884-2892, 2022. 1", + "[48] Jun Yang, Lizhi Bai, Yaoru Sun, Chunqi Tian, Maoyu Mao, and Guorun Wang. Pixel difference convolutional network for rgb-d semantic segmentation. IEEE TCSVT, 34(3):1481-1492, 2024. 3", + "[49] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 2", + "[50] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. NeurIPS, 2024. 2", + "[51] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, pages 2636-2645, 2020. 6", + "[52] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In ICCV, pages 2100-2110, 2019. 4, 6" + ], + "bbox": [ + 516, + 90, + 905, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8294", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Xiangyu Yue, Zangwei Zheng, Shanghang Zhang, Yang Gao, Trevor Darrell, Kurt Keutzer, and Alberto Sangiovanni Vincentelli. Prototypical cross-domain self-supervised learning for few-shot unsupervised domain adaptation. In CVPR, pages 13834-13844, 2021. 1, 2", + "[54] Jiaming Zhang, Huayao Liu, Kailun Yang, Xinxin Hu, Ruiping Liu, and Rainer Stiefelhagen. Cmx: Cross-modal fusion for rgb-x semantic segmentation with transformers. IEEE Transactions on Intelligent Transportation Systems, 24(12): 14679-14694, 2023. 3", + "[55] Lu Zhang, Siqi Zhang, Xu Yang, Hong Qiao, and Zhiyong Liu. Unseen object instance segmentation with fully test-time rgb-d embeddings adaptation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4945-4952, 2023. 3", + "[56] Xingxuan Zhang, Linjun Zhou, Renzhe Xu, Peng Cui, Zheyan Shen, and Haoxin Liu. Towards unsupervised domain generalization. In CVPR, pages 4900-4910, 2022. 2", + "[57] Yuyang Zhao, Zhun Zhong, Na Zhao, Nicu Sebe, and Gim Hee Lee. Style-hallucinated dual consistency learning for domain generalized semantic segmentation. In ECCV, pages 535-552. Springer, 2022. 4, 6", + "[58] Zhun Zhong, Yuyang Zhao, Gim Hee Lee, and Nicu Sebe. Adversarial style augmentation for domain generalized urban-scene segmentation. NeurIPS, 35:338-350, 2022. 6, 8", + "[59] Wujie Zhou, Yun Zhu, Jingsheng Lei, Jian Wan, and Lu Yu. Ccafnet: Crossflow and cross-scale adaptive fusion network for detecting salient objects in rgb-d images. IEEE TMM, 24: 2192-2204, 2022. 3", + "[60] Wujie Zhou, Yuxiang Xiao, Weiqing Yan, and Lu Yu. Cmpffnet: Cross-modal and progressive feature fusion network for rgb-d indoor scene semantic segmentation. IEEE Transactions on Automation Science and Engineering, 21 (4):5523-5533, 2024. 3" + ], + "bbox": [ + 91, + 92, + 480, + 584 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8295", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Stronger, Steadier & Superior_ Geometric Consistency in Depth VFM Forges Domain Generalized Semantic Segmentation/2e6ee106-ae7d-42e3-b798-bd6f97fafe0a_model.json b/2025/Stronger, Steadier & Superior_ Geometric Consistency in Depth VFM Forges Domain Generalized Semantic Segmentation/2e6ee106-ae7d-42e3-b798-bd6f97fafe0a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3866ec04b2a1e73c382ad97f91cdd3598c608266 --- /dev/null +++ b/2025/Stronger, Steadier & Superior_ Geometric Consistency in Depth VFM Forges Domain Generalized Semantic Segmentation/2e6ee106-ae7d-42e3-b798-bd6f97fafe0a_model.json @@ -0,0 +1,2257 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.13, + 0.888, + 0.177 + ], + "angle": 0, + "content": "Stronger, Steadier & Superior: Geometric Consistency in Depth VFM Forges Domain Generalized Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.119, + 0.203, + 0.873, + 0.248 + ], + "angle": 0, + "content": "Siyu Chen\\(^{1,3}\\) Ting Han\\(^{2\\dagger}\\) Changshe Zhang\\(^{4}\\) Xin Luo\\(^{1}\\) Meiliu Wu\\(^{3}\\) Guorong Cai\\(^{1}\\) Jinhe Su\\(^{1\\dagger}\\) \n\\(^{1}\\) Jimei University \\(^{2}\\) Sun Yat-sen University \\(^{3}\\) University of Glasgow \\(^{4}\\) Xidian University" + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.262, + 0.276, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.4, + 0.281, + 0.413 + ], + "angle": 0, + "content": "(a) Stronger Pre-trained Models" + }, + { + "type": "image", + "bbox": [ + 0.281, + 0.262, + 0.428, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.33, + 0.428, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.33, + 0.4, + 0.528, + 0.413 + ], + "angle": 0, + "content": "(b) Steadier Visual-Spatial Attention" + }, + { + "type": "image", + "bbox": [ + 0.432, + 0.262, + 0.579, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.433, + 0.33, + 0.579, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.262, + 0.738, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.33, + 0.738, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.644, + 0.4, + 0.835, + 0.413 + ], + "angle": 0, + "content": "(c) Superior Generalization Ability" + }, + { + "type": "image", + "bbox": [ + 0.744, + 0.262, + 0.891, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.744, + 0.33, + 0.891, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.417, + 0.907, + 0.473 + ], + "angle": 0, + "content": "Figure 1. Existing methods fail to fully exploit the potential of VFMs for DGSS when visual cues are limited or absent. To this end, we introduce a novel and robust fine-tuning strategy DepthForge that leverages depth cues from a depth VFM to optimize visual cues, and employs spatial consistency to enhance feature discriminability. The proposed DepthForge achieves (a) stronger performance under extreme conditions, owing to (b) steadier visual-spatial attention, and thus delivers (c) superior generalization ability." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.489, + 0.327, + 0.505 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.524, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Vision Foundation Models (VFMs) have delivered remarkable performance in Domain Generalized Semantic Segmentation (DGSS). However, recent methods often overlook the fact that visual cues are susceptible, whereas the underlying geometry remains stable, rendering depth information more robust. In this paper, we investigate the potential of integrating depth information with features from VFMs, to improve the geometric consistency within an image and boost the generalization performance of VFMs. We propose a novel fine-tuning DGSS framework, named Depth-Forge, which integrates the visual cues from frozen DINOv2 or EVA02 and depth cues from frozen Depth Anything V2. In each layer of the VFMs, we incorporate depth-aware learnable tokens to continuously decouple domain-invariant visual and spatial information, thereby enhancing depth awareness and attention of the VFMs. Finally, we develop a depth refinement decoder and integrate it into the model architecture to adaptively refine multi-layer VFM features and depth-aware learnable tokens. Extensive experiments are conducted based on various DGSS settings and five different datasets as unseen target domains. The qualitative and quantitative results demonstrate that our method significantly outperforms alternative approaches with stronger performance, steadier visual-spatial attention, and superior generalization ability. In particular," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.907, + 0.536 + ], + "angle": 0, + "content": "DepthForge exhibits outstanding performance under extreme conditions (e.g., night and snow). Code is available at https://github.com/SY-Ch/DepthForge." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.559, + 0.646, + 0.573 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.584, + 0.907, + 0.78 + ], + "angle": 0, + "content": "Domain Generalized Semantic Segmentation (DGSS) aims to improve prediction accuracy across multiple unseen domains without requiring access to their data, thereby ensuring robust generalization for practical applications [34, 45, 53]. One common approach involves decomposing the learned features into domain-invariant and domain-specific components [42, 47]. This method aims to achieve robustness against domain variations by isolating content-related factors from those that vary between domains. A second strategy employs meta-learning techniques [24], which focus on training models to generalize effectively across domains by learning more adaptable representations, often leveraging shared structures across different datasets." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Recently, VFMs have emerged as a promising approach, primarily owing to their strong generalization capabilities, which are enabled by pre-trained on large-scale datasets [15, 25, 36]. Fine-tuning these models has become increasingly popular due to their cost-effectiveness and superior performance [2, 44]. However, a significant challenge remains: what constitutes robust feature information for domain generalization, and how can such robust informa" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8285" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.091, + 0.482, + 0.198 + ], + "angle": 0, + "content": "tion be effectively extracted? Previous methods predominantly utilize VFMs based on RGB images. However, the inherent limitations of RGB prevent those methods from adapting to variable conditions, especially in scenarios such as nighttime, overexposure, snow, or fog, where visual cues are not clearly present. This clearly fails to meet the requirements of domain-generalized semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.201, + 0.486, + 0.578 + ], + "angle": 0, + "content": "In this paper, we introduce DepthForge, a robust fine-tune and optimization strategy for domain generalized semantic segmentation, as shown in Fig. 1. Specifically, we employ two frozen VFMs: Depth Anything V2[49, 50] is dedicated to extract depth information, and DINOv2[30] or EVA02[15] are used to process RGB visual features. We enhance the generalization of the visual features at each layer of the VFMs by integrating the corresponding depth cues. More importantly, we introduce depth-aware learnable tokens that enables the model to learn spatial structure invariance, thereby ensuring consistent performance across different domains. The depth and visual features extracted from frozen VFMs are used to optimize the representation of our learnable tokens, thus producing more robust visual-spatial attention features than those based solely on visual cues and forging geometric consistency, as shown in Fig. 1(b). However, we observe the frozen VFMs only provide static features, they are incapable of refining the feature maps during training. This limitation leads to persistent errors in the learnable tokens, causing the optimization process to drift in incorrect directions that are difficult to rectify. To this end, we propose a depth refinement decoder that dynamically adjusts the multi-scale features, combining prior and dynamic features to establish high-quality paired feature relationships at different depth spaces." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.483, + 0.837 + ], + "angle": 0, + "content": "Our core contribution is to demonstrate that robust depth features and learnable tokens can be effectively applied to improve model performance in Domain Generalized Semantic Segmentation (DGSS). To the best of our knowledge, the proposed DepthForge is the first depth-aware framework specifically designed for DGSS. The effectiveness of our design is verified by experiments on widely used Cityscapes, ACDC, Mapillary, BDD100k, and GTA5 datasets. In addition, the individual components of our design are also verified by extensive experiments. Extensive experiments have demonstrated that our method significantly improves DGSS performance, as shown in Fig. 1(c). Specifically, in the \\(GTA \\rightarrow\\) Cityscapes + BDD100k + Mapillary setting, we observed approximately \\(+4\\%\\) significant improvement in mIoU, particularly the improvement in extreme scenes is close to \\(+5\\%\\). Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.84, + 0.483, + 0.885 + ], + "angle": 0, + "content": "- We propose a stronger DepthForge learning framework to fine-tune VFMs combined with visual and depth cues for domain-generalized semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- A depth-awareness learnable tokens are introduced to" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.088, + 0.909, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.242, + 0.907, + 0.298 + ], + "angle": 0, + "content": "Figure 2. Different configurations of incorporating depth cues into visual information, where Config.1: adding \\( f_{i}^{d} \\) to \\( f_{i}^{v} \\), or Config.2: incorporating \\( f_{i}^{d} \\) into the learnable token \\( T_{i} \\), and Config.3: our DepthForge, respectively." + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.322, + 0.905, + 0.381 + ], + "angle": 0, + "content": "focus on spatial relationship that acquires steadier and more robust visual-spatial attention. This strategy enables DepthForge to enhance the discriminability and refine features at an instance level within each layer." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.382, + 0.905, + 0.442 + ], + "angle": 0, + "content": "- Depth Forge highlights the impressive generalization ability, surpassing existing DGSS methods by a large margin, especially in extreme scenes with limited and absent visual cues." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.469, + 0.655, + 0.484 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.497, + 0.892, + 0.514 + ], + "angle": 0, + "content": "2.1. Domain Generalized Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Domain Generalized Semantic Segmentation (DGSS) aims to enhance the generalization capability of segmentation models to unseen domains. Recent methods primarily focus on reducing feature distribution discrepancies across domains and learning more domain-invariant representations. Adversarial training [53], meta-learning [4, 41], and self-supervised learning [18, 56] have been extensively explored to improve generalization performance. However, these methods often rely on constructing cross-domain contrastive constraints or employing sophisticated training strategies, limiting their generalization improvements and complicating the training process. The emergence of Vision Foundation Models (VFMs), trained via large-scale image-text [36] or image-image contrastive learning [30], has opened a promising avenue for DGSS. VFMs capture universal and rich visual representations beneficial for reducing domain biases. Recent approach [2, 44] leverages pretrained VFMs by employing their frozen features as powerful generalized backbones to extract and refine learnable tokens, leading to significant improvements in model performance. Nevertheless, current DGSS approaches predominantly utilize RGB-only data, leaving the models susceptible to variations in lighting, texture, and spatial arrangements across different domains, thus constraining their full potential for generalization." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8286" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.905, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.376, + 0.907, + 0.461 + ], + "angle": 0, + "content": "Figure 3. DepthForge framework. It incorporates the features from frozen VFM and depth VFM into our Depth-aware Learnable Tokens to generate enhanced features \\( f_{i+1} = f_i^v + \\text{DepthForge}(f_i^v, f_i^d) \\) for each layer. In our tunable DepthForge, We first design a Depth Awareness module that integrates visual and depth cues into learnable tokens. Next, we propose an Attention Optimization module to strengthen the discriminability of instances across different spatial locations. Finally, multi-scale features are fused in Depth Refinement Decoder to adapt to various scales and spatial connections. Our DepthForge establishes direct connections to different instances in real-world locations, facilitating both feature refinement and improved discriminability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.478, + 0.295, + 0.494 + ], + "angle": 0, + "content": "2.2. RGB-D Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.5, + 0.486, + 0.803 + ], + "angle": 0, + "content": "Combining RGB with depth information has shown substantial benefits across various visual tasks [6, 8, 9, 17, 39]. In object detection, RGB-D effectively mitigates limitations inherent to RGB data under complex scenes by providing robust geometric and spatial cues, thereby enhancing localization accuracy and model robustness [43, 59]. In semantic segmentation, depth information assists models in capturing spatial structures and geometric contours of objects, alleviating issues related to imprecise object boundary delineation. Particularly in scenarios with blurred boundaries, insufficient illumination, or texture scarcity, the incorporation of depth information substantially improves segmentation accuracy and stability [14, 48, 54, 60]. Recent studies have further indicated that even in domain generalization tasks, integrating RGB and depth features consistently improves the model's generalization performance on unseen domains [19, 27, 38, 55]. Therefore, exploring the integration of depth information to address the limitations of RGB-only approaches represents a promising and essential research direction within DGSS." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.816, + 0.218, + 0.833 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.903 + ], + "angle": 0, + "content": "To achieve better generalization performance, existing approaches opt to fine-tune VFMs in a parameter-efficient manner. One mainstream idea is to refine and propagate features from frozen VFM layers to subsequent ones. Specif" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.479, + 0.907, + 0.571 + ], + "angle": 0, + "content": "ically, given a pre-trained VFM composed of \\(N\\) sequential layers \\(\\{L_1,L_2,\\dots,L_N\\}\\), where each layer is associated with a weight matrix \\(W_{i}\\in \\mathbb{R}^{c\\times c}\\) and the frozen output features of each layer are denoted as \\(f_{i}\\) for \\(i = 1,2,\\ldots ,N\\). If we have a learnable weight matrix as \\(\\Delta W_{i}\\in \\mathbb{R}^{c\\times c}\\), the feature propagation from layer \\(L_{i}\\) to \\(L_{i + 1}\\) is then formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.628, + 0.583, + 0.907, + 0.6 + ], + "angle": 0, + "content": "\\[\nf _ {i + 1} = W _ {i} f _ {i} + \\Delta W _ {i} f _ {i}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.615, + 0.907, + 0.645 + ], + "angle": 0, + "content": "Recent REIN transfers the learnable matrix \\(\\Delta W_{i}\\) into a learnable token \\(T_{i}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.659, + 0.907, + 0.676 + ], + "angle": 0, + "content": "\\[\nf _ {i + 1} = W _ {i} f _ {i} + \\epsilon \\left(T _ {i} \\left(W _ {i} f _ {i}\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.906, + 0.779 + ], + "angle": 0, + "content": "where \\(\\epsilon (\\cdot)\\) denotes the MLP function, and \\(T_{i}\\in \\mathbb{R}^{m\\times c}\\) with a tokn length \\(m\\). REIN enables each token to connect better with instances in an image. However, we observe that since the updated and frozen features are limited to remain similar, this correlation is fragile and the capacity for feature refinement is limited." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "This prompted us to explore what kind of information could improve discrimination in instances and refine the feature representations. Drawing inspiration from multimodal learning, we observe that depth features provide a spatial relationship bias for corresponding pixels in visual features, thereby compensating for the loss of crucial visual cues in DGSS. We assume that incorporating a pretrained depth VFM into the pre-trained VFMs will alleviate" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8287" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.089, + 0.462, + 0.304 + ], + "angle": 0, + "content": "
MethodProc. & YearSnowNightFogRain
ResNet based:
IBN [31]ECCV201849.621.263.850.4
Itenorm [21]CVPR201949.923.863.350.1
IW [32]CVPR201947.621.862.452.4
RobustNet [10]CVPR202149.824.364.356.0
WildNet [26]CVPR202228.412.741.234.2
Transformer based:
HGFormer [13]CVPR202368.652.769.972.0
CMFormer [3]AAAI202464.333.777.867.6
VFM based:
REIN [44]CVPR202470.655.979.572.5
FADA [2]NeurIPS202473.557.480.275.0
DepthForge(Ours)-75.460.481.275.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.306, + 0.483, + 0.364 + ], + "angle": 0, + "content": "Table 1. Performance comparison between our DepthForge and existing DGSS methods under Cities. \\(\\rightarrow\\) ACDC with Snow, Night, Fog, Rain generalization setting. Top three results are highlighted as best, second, and third, respectively. (%)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.381, + 0.483, + 0.411 + ], + "angle": 0, + "content": "attention bias and significantly enhance scene understanding, particularly under extreme conditions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.411, + 0.484, + 0.534 + ], + "angle": 0, + "content": "We use the frozen features \\( f_{i}^{v} \\) from VFM for visual embedding, and the frozen features \\( f_{i}^{d} \\) from depth VFM for depth embedding. A straightforward thought might simply add \\( f_{i}^{d} \\) to \\( f_{i}^{v} \\), or incorporate \\( f_{i}^{d} \\) into the leanable token \\( T_{i} \\). However, these methods lack flexibility and introduce significant inter-layer feature differences that prevent convergence, as shown in Fig. 2. In contrast, we propose Depth-Forge to adaptively fuse visual and depth cues, denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.556, + 0.483, + 0.575 + ], + "angle": 0, + "content": "\\[\nf _ {i + 1} = W _ {i} ^ {v} f _ {i} + \\varepsilon \\left(\\epsilon^ {v} \\left(T _ {i} \\left(W _ {i} ^ {v} f _ {i} ^ {v}\\right)\\right) + \\epsilon^ {d} \\left(T _ {i} \\left(W _ {i} ^ {d} f _ {i} ^ {d}\\right)\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.484, + 0.674 + ], + "angle": 0, + "content": "where \\( W_{i}^{v} \\) and \\( W_{i}^{d} \\) represent the weight matrix from VFMs and depth VFM, respectively. \\( \\varepsilon \\) denotes the adaptation of the visual and depth signals. This approach allows us to more effectively utilize the powerful capacities of multimodal VFMs to forge robust refined features and instance discriminability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.766 + ], + "angle": 0, + "content": "More importantly, the depth VFM itself possesses strong generalization capabilities, allowing domain adaptation without fine-tuning. A pre-trained VFM applies the same. Therefore, the overall optimization objective still focuses on the segmentation head \\(\\mathcal{H}\\) with parameter \\(\\theta_{h}\\) and our Depth-Forge with parameter \\(\\theta_{D}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.775, + 0.483, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\arg \\min _ {\\theta_ {D}, \\theta_ {h}} \\sum_ {i = 1} ^ {N} \\mathcal {L o s s} \\left(\\mathcal {H} _ {\\theta_ {i}} \\left(\\mathcal {F} _ {\\Theta^ {v}, \\Theta^ {d}, \\theta_ {D}} \\left(x _ {i}\\right)\\right), y _ {i}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\( x_{i} \\) and \\( y_{i} \\) denote the input image and its corresponding ground truth, respectively, and \\( N \\) signifies the total number of images. \\( \\mathcal{F} \\) represent the forward process of our DepthForge strategy applied to the pre-trained VFMs with parameters \\( \\Theta^v \\) and \\( \\Theta^d \\)." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.906, + 0.356 + ], + "angle": 0, + "content": "
MethodProc. & YearBDD100kMapillaryGTA
ResNet based:
IBN [31]ECCV201848.5657.0445.06
Itenorm [21]CVPR201949.2356.2645.73
IW [32]CVPR201948.4955.8244.87
RobustNet [10]CVPR202150.7358.6445.00
DRPC [52]ICCV201949.8656.3445.62
GTR [33]TIP202150.7557.1645.79
SHADE [57]ECCV202250.9560.0748.61
SAW [35]CVPR202252.9559.8147.28
WildNet [26]CVPR202250.9458.7947.01
BlindNet [1]CVPR202451.8460.1847.97
Transformer based:
HGFormer [13]CVPR202353.4066.9051.30
CMFormer [3]AAAI202459.2771.1058.11
VFM based:
REIN [44]CVPR202463.5474.0362.41
FADA [2]NeurIPS202465.1275.8663.78
DepthForge(Ours)-66.1975.9367.24
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.358, + 0.907, + 0.418 + ], + "angle": 0, + "content": "Table 2. Performance comparison between our DepthForge and existing DGSS methods under the Citys. \\(\\rightarrow\\) BDD. \\(^+\\) Map. \\(^+\\) GTA5 generalization setting. Top three results are highlighted as best, second, and third, respectively. \\((\\%)\\)" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.438, + 0.64, + 0.456 + ], + "angle": 0, + "content": "4. DepthForge" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.463, + 0.907, + 0.539 + ], + "angle": 0, + "content": "This paper introduces the DepthForge strategy, a novel approach for domain generalized semantic segmentation. As shown in Fig. 3, DepthForge consists of three main parts: (1) Depth Awareness; (2) Attention Optimization; (3) Depth Refinement Decoder. We detail the designs as follows." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.548, + 0.685, + 0.564 + ], + "angle": 0, + "content": "4.1. Depth Awareness" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.569, + 0.907, + 0.765 + ], + "angle": 0, + "content": "Recent DGSS methods have applied pre-trained VFMs to source domain segmentation, with the core idea being to enhance the features between frozen VFM layers to help the model better adapt to target tasks. However, we observe that visual cues exhibit significant variations and even deficiencies across different scenes. Consequently, a more robust cue is required to provide a stronger prior for domain generalization. This motivates us to explore a strategy that focuses on the inherent similarities across different domains. In general, spatial location priors are able to help differentiate instances even when visual cues fail. To this end, we employ an additional depth VFM as an auxiliary network, as shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.886 + ], + "angle": 0, + "content": "Specifically, given a VFM \\(\\mathcal{V}\\) and a depth VFM \\(\\mathcal{D}\\) with \\(N\\) layers \\(\\{L_1,L_2,\\dots ,L_N\\}\\), the corresponding features are denoted as \\(\\{f_i^v,f_i^d |f_i\\in \\mathbb{R}^{n\\times c},i\\in [1,N]\\}\\). Note that the layers \\(\\{L_1,L_2,\\dots ,L_N\\}\\) are kept frozen. We acquire the visual feature \\(f_{i}^{v}\\) from the \\(i\\)-th layer \\(L_{i}\\) of \\(\\mathcal{V}\\), and the depth feature \\(f_{i}^{d}\\) from the corresponding layer of \\(\\mathcal{D}\\). Our goal is to generate the enhanced feature \\(\\Delta f_{i} = DepthForge(f_{i}^{v},f_{i}^{d})\\) by using both visual and depth cues." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "We employ a set of learnable tokens \\( T = \\{T_{i} \\in \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8288" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.092, + 0.885, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.358, + 0.905, + 0.386 + ], + "angle": 0, + "content": "Figure 4. Key segmentation examples of existing DGSS methods and DepthForge under the Cityscapes to ACDC unseen target domains under Night, Snow, Rain, and Fog conditions." + }, + { + "type": "image", + "bbox": [ + 0.12, + 0.392, + 0.889, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.542, + 0.905, + 0.571 + ], + "angle": 0, + "content": "Figure 5. Key segmentation examples of existing DGSS methods and DepthForge under the GTA5 to Mapillary, BDD100k, and Cityscapes unseen target domains setting." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.585, + 0.483, + 0.69 + ], + "angle": 0, + "content": "\\(\\mathbb{R}^{m \\times c} |i \\in [1, N]\\}\\) to adaptively learn scene knowledge. As we consider, relying solely on visual cues can be limiting and prone to failure. Therefore, we utilize visual and depth cues to construct learnable tokens, which we term depth-aware learnable tokens. We integrate \\(T_i\\) with \\(f_i^v\\) and \\(f_i^d\\) using an attention-based mechanism, prompting tailored adjustments to bridge the gap between different domains." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.86 + ], + "angle": 0, + "content": "We use \\( f_{i}^{v,d} \\) as the query, and \\( T_{i} \\) as the key and the value A dot-product operation is applied to generate awareness map \\( A_{i}^{v,d} \\), and the Softmax function (S) is employed to align each patch with a unique instance. The \\( A_{i}^{v,d} \\) provides insights into both the visual cues and depth retrieval, qualitatively evaluating the relationship between various tokens and visual or depth cues, respectively. Subsequently, we incorporate depth awareness bias \\( A_{i}^{d} \\) into the visual features \\( A_{i}^{v} \\) to provide additional spatial relationships, thereby establishing a latent connection between pixels and the real world. The process is mathematically formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.87, + 0.483, + 0.906 + ], + "angle": 0, + "content": "\\[\nA _ {i} = S \\left(\\frac {f _ {i} ^ {v} \\times (T ^ {v}) _ {i} ^ {\\top}}{\\sqrt {c}}\\right) + \\lambda S \\left(\\frac {f _ {i} ^ {d} \\times (T ^ {d}) _ {i} ^ {\\top}}{\\sqrt {c}}\\right). \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.585, + 0.73, + 0.601 + ], + "angle": 0, + "content": "4.2. Attention Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.907, + 0.73 + ], + "angle": 0, + "content": "Directly leveraging the awareness representation \\( A_{i} \\) to generate enhanced features may introduce a significant amount of irrelevant information for downstream tasks. To this end, we employ depth-aware learnable tokens \\( T_{i} \\) to learn an alignment weight \\( W_{T_i} \\) and bias \\( b_{T_i} \\). The \\( T_{i} \\) with weight \\( W_{T_i} \\) and bias \\( b_{T_i} \\) is then multiplied with the \\( A_{i} \\) to obtain the initial enhanced features \\( \\Delta f_{i} \\) that embed both visual and depth cues, using the equation:" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.738, + 0.905, + 0.756 + ], + "angle": 0, + "content": "\\[\n\\Delta \\hat {f} _ {i} = A _ {i} \\times \\left(T _ {i} \\times W _ {T _ {i}} + b _ {T _ {i}}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "where we select the same optimization strategy as REIN, discarding high-weight features when the features are sufficiently precise to avoid unnecessary adjustments and mitigate the risk of inappropriate variations. Finally, we employ two consecutive residual connections \\(\\phi\\) to enhance the flexibility of attention feature adjustment and prevent gradient vanishing during the forward pass. The calculation is defined as \\(\\Delta f_{i} = \\phi \\hat{\\Delta f}_{i}\\). Benefiting from these depth-aware and attention optimization adjustments, DepthForge is ca" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8289" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.482, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.258, + 0.483, + 0.313 + ], + "angle": 0, + "content": "Figure 6. t-SNE visualization compares REIN and our Depth-Forge. The first volume corresponds to \\( \\text{Citys.} \\rightarrow \\text{ACDC} \\), and two through four volumes correspond to \\( \\text{GTA5} \\rightarrow \\text{BDD.} + \\text{Citys.} + \\text{Map} \\). The best mIoU results are highlighted in red." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.329, + 0.484, + 0.375 + ], + "angle": 0, + "content": "pable of generating more distinctive, discriminative features for different categories within a single image, while balancing both visual and depth cues." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.383, + 0.336, + 0.399 + ], + "angle": 0, + "content": "4.3. Depth Refinement Decoder" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.405, + 0.484, + 0.465 + ], + "angle": 0, + "content": "The depth-aware features at different layers are inconsistent. Therefore, how to design a robust decoder becomes a core challenge that effectively integrate these enhanced features across various scales to accomplish downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.466, + 0.484, + 0.569 + ], + "angle": 0, + "content": "Previous methods rely solely on the output of the final layer \\( f_{N} \\) for semantic segmentation predictions. In contrast, we fuse the outputs from layers \\( L_{1} - L_{N} \\) to refine the depth features. Specifically, we process these multi-layer features with an MLP with two different fully-connected layers \\( \\phi \\) and \\( \\varphi \\) and then concatenate them after a ReLU activation function:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.57, + 0.482, + 0.587 + ], + "angle": 0, + "content": "\\[\nf _ {i} ^ {\\text {o u t}} = \\phi (R e L U (\\varphi f _ {i})), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.607, + 0.482, + 0.625 + ], + "angle": 0, + "content": "\\[\nF _ {u} = \\gamma \\operatorname {C o n c a t} \\left(f _ {1} ^ {\\text {o u t}}, f _ {2} ^ {\\text {o u t}}, \\dots , f _ {N} ^ {\\text {o u t}}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.484, + 0.675 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is a convolutional layer. Finally, we design several sequential multi-head Transformer layers to generate the final predictions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.688, + 0.217, + 0.705 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.376, + 0.728 + ], + "angle": 0, + "content": "5.1. Datasets & Evaluation Protocols" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We evaluate our DepthForge on four real-world datasets (Cityscapes [12], BDD100K [51], Mapillary [28], ACDC [40]) and a synthetic dataset (GTA5 [37]). Cityscapes (Citys.) is an autonomous driving dataset comprising 2,975 training images and 500 validation images, each with a resolution of \\(2048 \\times 1024\\). BDD100K (BDD.) and Mapillary (Map.) provide 1,000 and 2,000 validation images, with resolutions of \\(1280 \\times 720\\) and \\(1902 \\times 1080\\), respectively. ACDC offers 406 validation images captured under extreme conditions: namely at Night, Snow, Fog, and Rain, each with a resolution of \\(1902 \\times 1080\\). GTA5 is a" + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.09, + 0.898, + 0.328 + ], + "angle": 0, + "content": "
MethodProc. & YearCityscapesBDD100kMapillary
ResNet based:
DRPC [52]ICCV201937.4232.1434.12
GTR [33]TIP202137.5333.7534.52
RobustNet [10]CVPR202136.5835.2040.33
DIRL [46]AAAI202241.0439.1541.60
SHADE [57]ECCV202244.6539.2843.34
WildNet [26]CVPR202344.6238.4246.09
SAW [35]CVPR202239.7537.3441.86
AdvStyle [58]NeurIPS202239.6235.5437.00
SPC [22]CVPR202344.1040.4645.51
BlindNet [1]CVPR202445.7241.3247.08
Transformer based:
CMFormer [3]AAAI202455.3149.9160.09
VFM based:
DIDEX [29]WACV202462.0054.3063.00
REIN [44]CVPR202466.4060.4066.10
FADA [2]NeurIPS202468.2361.9468.09
DepthForge(Ours)-69.0462.8269.22
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.331, + 0.905, + 0.391 + ], + "angle": 0, + "content": "Table 3. Performance comparison between our DepthForge and existing DGSS methods under the \\(GTA5 \\rightarrow Citys\\). + BDD. + Map. generalization setting. Top three results are highlighted as best, second, and third, respectively. (%)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.408, + 0.905, + 0.438 + ], + "angle": 0, + "content": "synthetic dataset that provides 24,966 annotated images obtained from the game." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.439, + 0.906, + 0.574 + ], + "angle": 0, + "content": "Following the evaluation protocol of the existing DGSS methods, a certain dataset is chosen as the source domain for training, and the other three datasets serve as unseen target domains for validation. The three common evaluation settings are as follows: (1) Cityscapes \\(\\rightarrow\\) ACDC (Night, Snow, Fog, Rain); (2) GTA5 \\(\\rightarrow\\) Cityscapes, BDD100K, Mapillary; and (3) Cityscapes \\(\\rightarrow\\) BDD100K, Mapillary, GTA5. The evaluation metric is the mean Intersection-over-Union (mIoU)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.586, + 0.871, + 0.603 + ], + "angle": 0, + "content": "5.2. Deployment Details & Parameter Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.906, + 0.744 + ], + "angle": 0, + "content": "Our implementation is built on the MMSegmentation [11] codebase. For optimization, we employ AdamW optimizer with an initial learning rate of 1e-4, a weight decay of 0.05, an epsilon of 1e-8, and beta values of (0.9, 0.999). A OneCycleLR scheduler is applied over 50,000 total steps with a maximum learning rate of 1e-4, featuring a warmup phase covering \\(10\\%\\) of the training iterations, and using cosine annealing, with both the div_factor and final_div_factor set to 10." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.746, + 0.905, + 0.806 + ], + "angle": 0, + "content": "Data augmentation includes multi-scale resizing, random cropping (with a fixed crop size and category ratio constraints), random flipping, and photometric distortion. All experiments are conducted on an NVIDIA RTX 3090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.817, + 0.887, + 0.833 + ], + "angle": 0, + "content": "5.3. Comparison with State-of-The-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.884 + ], + "angle": 0, + "content": "We compare our approach with existing DGSS methods, including: (1) ResNet-based methods, (2) Transformer-based methods, and (3) VFM-based methods." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Cityscapes \\(\\rightarrow\\) ACDC. Tab. 1 presents the performance" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8290" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.09, + 0.49, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.091, + 0.904, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.283, + 0.908, + 0.313 + ], + "angle": 0, + "content": "Figure 7. Visualizations of attention maps and affinity maps under differnet configurations, where DRD denotes the Depth Refinement Decoder, DA and AO represent the Depth Awareness and Attention Optimization, respectively." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.33, + 0.485, + 0.476 + ], + "angle": 0, + "content": "
BackboneConfigurationsSnowNightFogRainBDDMap
EVA02 (Large) [15, 16]REIN [44]66.052.176.069.560.070.0
+ concat fi d66.352.276.269.860.270.2
+ Prompt Depth Anything65.852.475.569.360.469.9
+ Depth Anything V267.253.077.070.260.970.8
+ DRD67.753.577.670.761.271.0
+ DA + AO67.953.777.870.961.571.2
DepthForge69.554.578.572.062.073.0
DINOv2 (Large) [30]REIN [44]70.655.979.572.563.574.0
+ concat fi d70.856.179.472.863.674.2
+ Prompt Depth Anything70.455.579.772.063.873.9
+ Depth Anything V271.556.780.573.364.474.5
+ DRD72.057.280.774.064.974.8
+ DA + AO72.257.480.974.265.175.0
DepthForge75.460.481.275.466.276.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.478, + 0.484, + 0.565 + ], + "angle": 0, + "content": "Table 4. Ablation studies on component configurations of the proposed DepthForge under Citys. \\(\\rightarrow\\) ACDC with Snow, Night, Fog, Rain + BDD. + Map. generalization setting, where DRD denotes the Depth Refinement Decoder, DA and AO represent the Depth Awareness and Attention Optimization. Top three results are highlighted as best, second, and third, respectively. (%)" + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.569, + 0.482, + 0.672 + ], + "angle": 0, + "content": "
No.DesignAvg. Gain (%)RainNightSnowRog
1No Depth-70.655.979.572.5
2Concatenation-1.169.855.078.371.2
3Depth Token1.672.357.880.774.0
4Config10.270.856.179.472.8
5Config20.771.556.880.173.0
6Config3, Ours3.575.460.481.275.4
7w/o Scale Factor2.874.859.380.974.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.717 + ], + "angle": 0, + "content": "Table 5. Performance comparison of the different configurations. Top three results are highlighted as best, second, and third, respectively. (%)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.903 + ], + "angle": 0, + "content": "on the Cityscapes source domain and the ACDC target domain. Compared with the VFM-based REIN and FADA, the proposed DepthForge shows mIoU improvement of \\(4.8\\%\\) & \\(1.9\\%\\), \\(4.5\\%\\) & \\(3.0\\%\\), \\(1.7\\%\\) & \\(1.0\\%\\), and \\(2.9\\%\\) & \\(0.4\\%\\) for the Snow, Night, Fog, and Rain scenes, respectively. The feature space of the REIN and the proposed method is visualized in the first volume of Fig. 6. Moreover, our proposed method demonstrates superior performance compared to ResNet-based and Transformer-based approaches. These experimental results underscore the strong and robust generalization ability of DepthForge in extreme scenes with" + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.33, + 0.905, + 0.612 + ], + "angle": 0, + "content": "
BackboneFine-tune \nMethodTrainable \nParams*mIoUAvg.
CitysBDDMap
CLIP [36] \n(ViT-Large)Full304.15M51.347.654.351.1
Freeze0.00M53.748.755.052.4
REIN2.99M57.154.760.557.4
FADA11.65M58.755.862.158.9
DepthForge(Ours)2.99M59.757.063.560.0
SAM [25] \n(Huge)Full632.18M57.651.761.556.9
Freeze0.00M57.047.158.454.2
REIN4.51M59.652.062.157.9
FADA16.59M61.053.263.460.0
DepthForge(Ours)4.51M63.355.565.161.3
EVA02 [15, 16] \n(Large)Full304.24M62.156.264.660.9
Freeze0.00M56.553.658.656.2
REIN2.99M65.360.564.963.6
FADA11.65M66.761.966.164.9
DepthForge(Ours)2.99M68.061.767.565.7
DINOV2 [30] \n(Large)Full304.20M63.757.464.261.7
Freeze0.00M63.356.163.961.1
REIN2.99M66.460.466.164.3
FADA11.65M68.262.068.166.1
DepthForge(Ours)2.99M69.062.869.267.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.614, + 0.907, + 0.689 + ], + "angle": 0, + "content": "Table 6. Performance comparison of the proposed DepthForge across multiple VFs as backbones under the \\(GTA5 \\rightarrow Citys\\). + BDD. + Map. generalization setting. Mark * denotes trainable parameters in backbones. Top three results are highlighted as best, second, and third, respectively. (%)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.78 + ], + "angle": 0, + "content": "limited visual cues. Some exemplar segmentation results are compared under the Cityscapes \\(\\rightarrow\\) ACDC are provided in Fig. 4. Our DepthForge shows better pixel-wise predictions than previous methods, demonstrating the effectiveness and significance of depth awareness in our design." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Cityscapes \\(\\rightarrow\\) BDD, Map, GTA5. Tab. 2 shows that the proposed DepthForge achieves the state-of-the-art performance, outperforming the REIN and FADA by \\(2.65\\%\\) & \\(1.07\\%\\), \\(1.9\\%\\) & \\(0.07\\%\\), and \\(3.12\\%\\) & \\(1.13\\%\\) on the BDD100k, Mapillary, and GTA unseen target domains, respectively. The proposed method achieves an improvement of approximately \\(10\\%\\) compared to ResNet- and Transformer-based approaches." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.958 + ], + "angle": 0, + "content": "8291" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.099, + 0.089, + 0.48, + 0.466 + ], + "angle": 0, + "content": "
BackboneFine-tune MethodTrainable Params*mIoUAvg.
CitysBDDMap
EVA02 (Large) [15, 16]Full304.24M62.156.264.660.9
+AdvStyle [58]304.24M63.156.464.061.2
+PASTA [5]304.24M61.857.163.660.8
+GTR-LTR [33]304.24M59.857.463.260.1
Freeze0.00M56.553.658.656.2
+AdvStyle [58]0.00M51.451.656.553.2
+PASTA [5]0.00M57.852.358.556.2
+GTR-LTR [33]0.00M52.552.857.154.1
+LoRA [20]1.18M55.552.758.355.5
+AdaptFormer [7]3.17M63.759.964.262.6
+VPT [23]3.69M62.257.762.560.8
+REIN [44]2.99M65.360.564.963.6
+FADA [2]11.65M66.761.966.164.9
+DepthForge(Ours)2.99M68.061.767.565.7
DINOv2 (Large) [30]Full304.20M63.757.464.261.7
+AdvStyle [58]304.20M60.858.062.560.4
+PASTA [5]304.20M62.557.264.761.5
+GTR-LTR [33]304.20M62.757.464.561.6
Freeze0.00M63.356.163.961.1
+AdvStyle [58]0.00M61.555.163.960.1
+PASTA [5]0.00M62.157.264.561.3
+GTR-LTR [33]0.00M60.257.762.260.0
+LoRA [20]0.79M65.258.364.662.7
+AdaptFormer [7]3.17M64.959.064.262.7
+VPT [23]3.69M65.259.465.563.3
+REIN [44]2.99M66.460.466.164.3
+FADA [2]11.65M68.262.068.166.1
+DepthForge(Ours)2.99M69.062.869.267.0
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.467, + 0.483, + 0.542 + ], + "angle": 0, + "content": "Table 7. Performance comparison of the proposed DepthForge against other DGSS methods under the \\(GTA5 \\rightarrow Citys\\). + BDD. + Map. generalization setting. Mark * denotes trainable parameters in backbones. Top three results are highlighted as best, second, and third, respectively. (%)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.561, + 0.483, + 0.713 + ], + "angle": 0, + "content": "\\(\\mathrm{GTA5} \\rightarrow \\mathrm{Citys}\\), Map, BDD. As shown in Tab. 3, the proposed DepthForge achieves mIoU improvements of \\(2.64\\%\\), \\(2.42\\%\\), and \\(3.12\\%\\) on the Cityscapes, BDD100k, and Mapillary unseen target domains, respectively. The feature space of REIN and our DepthForge is visualized in Fig. 6, demonstrating the effectiveness of our depth-aware learnable tokens in the visual-spatial attention learning. In addition, we outperform the latest FADA by approximately \\(1\\%\\) across all three target domains. Visual examples for qualitative comparison are given in Fig. 5." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.726, + 0.255, + 0.74 + ], + "angle": 0, + "content": "5.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Depth awareness and attention optimization are significant for depth-aware learnable tokens. We find that simply incorporating a spatial cue bias into the original learnable tokens \\((+f_{i}^{d})\\) makes almost no difference. However, when we transform the original learnable tokens into depth-aware learnable tokens using the DepthForge framework, we achieve an approximate \\(2\\%\\) improvement in the target domains ACDC and Mapillary, as shown in Tab. 4. The depth-aware learnable tokens compensate for the shortcomings of visual cues and establish more robust visual-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "spatial attention. Compared to different configurations, our design allows gradients to be backpropagated through the tokens alone, enables each modality to contribute independently based on its own feature representations, preventing dominance from one during optimization, as shown in Tab. 5 and Fig. 2. This conclusion is supported by qualitative visualizations in Fig. 7." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.905, + 0.319 + ], + "angle": 0, + "content": "Depth refinement decoder facilitates the perception of both visual and depth information. As shown in Tab. 4, we observe that depth refinement decoder integrates multiscale information to enhance the expressive capacity of both the frozen VFM features and the enhanced features. With the incorporation of depth information, this optimization effect is further amplified. This conclusion is corroborated by qualitative visualizations in Fig. 7." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.327, + 0.797, + 0.344 + ], + "angle": 0, + "content": "5.5. Generalization on Other Setting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.349, + 0.905, + 0.499 + ], + "angle": 0, + "content": "To Different Depth VFMs. We evaluate the performance of the proposed DepthForge across different depth VFMs. Tab. 4 shows the relative spatial relationships derived from Depth Anything v2 outperform the absolute spatial cues provided by Prompt Depth Anything. We believe that what is more crucial in an image is the relative spatial relationship, which establishes connections among pixels and between pixels and the world. In contrast, absolute spatial information introduces significant discrepancies that can act as noise to visual cues." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.905, + 0.56 + ], + "angle": 0, + "content": "To Different VFMs. We evaluate the generalization ability of the proposed DepthForge across different VFMs. Tab. 6 shows the superiority of DepthForge when embedded into these VFMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.561, + 0.905, + 0.621 + ], + "angle": 0, + "content": "To Token Fine-Tuning Methods. We further compare the proposed method with some existing fine-tuning methods. As shown in Tab. 7, the proposed DepthForge outperforms these methods in all unseen target domains." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.634, + 0.634, + 0.65 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.905, + 0.901 + ], + "angle": 0, + "content": "This paper presents the DepthForge framework, a novel approach for robust domain generalization semantic segmentation. We provide a detailed analysis of the challenges and limitations faced by existing methods. To this end, we propose DepthForge to integrate a pretrained depth VFM for fine-tuning. We design depth-aware learnable tokens to provide supplementary spatial cues to the visual features, thereby improving geometric consistency and enhancing the pixel discriminability. Moreover, we incorporate multi-layer enhanced features to effectively capture spatial correlations across various scales. Extensive experiments demonstrate the substantial potential of leveraging multi-modal VFs in the field of DGSS, particularly under extreme conditions where visual cues are limited or missing, validating the effectiveness of our DepthForge in forging VFs for DGSS." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8292" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.258, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.206 + ], + "angle": 0, + "content": "This work was supported in part by the Natural Science Foundation of Xiamen, China, under Grant 3502Z202373036; in part by the National Natural Science Foundation of China under Grant 42371457; in part by the Key Project of Natural Science Foundation of Fujian Province, China, under Grant 2022J02045." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.222, + 0.188, + 0.237 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.247, + 0.484, + 0.316 + ], + "angle": 0, + "content": "[1] Woo-Jin Ahn, Geun-Yeong Yang, Hyun-Duck Choi, and Myo-Taeg Lim. Style blind domain generalized semantic segmentation via covariance alignment and semantic consistence contrastive learning. In CVPR, pages 3616-3626, 2024. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.318, + 0.484, + 0.387 + ], + "angle": 0, + "content": "[2] Qi Bi, Jingjun Yi, Hao Zheng, Haolan Zhan, Yawen Huang, Wei Ji, Yuexiang Li, and Yefeng Zheng. Learning frequency-adapted vision foundation model for domain generalized semantic segmentation. NeurIPS, 37:94047-94072, 2024. 1, 2, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.389, + 0.483, + 0.431 + ], + "angle": 0, + "content": "[3] Qi Bi, Shaodi You, and Theo Gevers. Learning content-enhanced mask transformer for domain generalized urban-scene segmentation. In AAAI, pages 819-827, 2024. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.433, + 0.483, + 0.487 + ], + "angle": 0, + "content": "[4] Manh-Ha Bui, Toan Tran, Anh Tran, and Dinh Phung. Exploiting domain-specific features to enhance domain generalization. In NeurIPS, pages 21189-21201. Curran Associates, Inc., 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.49, + 0.484, + 0.545 + ], + "angle": 0, + "content": "[5] Prithvijit Chattopadhyay, Kartik Sarangmath, Vivek Vijaykumar, and Judy Hoffman. Pasta: Proportional amplitude spectrum training augmentation for syn-to-real domain generalization. In ICCV, pages 19288-19300, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.547, + 0.484, + 0.615 + ], + "angle": 0, + "content": "[6] Geng Chen, Qingyue Wang, Bo Dong, Ruitao Ma, Nian Liu, Huazhu Fu, and Yong Xia. Em-trans: Edge-aware multimodal transformer for rgb-d salient object detection. IEEE Transactions on Neural Networks and Learning Systems, 36 (2):3175-3188, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.617, + 0.484, + 0.672 + ], + "angle": 0, + "content": "[7] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. NeurIPS, 35:16664-16678, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.675, + 0.484, + 0.73 + ], + "angle": 0, + "content": "[8] Siyu Chen, Ting Han, Changshe Zhang, Weiquan Liu, Jinhe Su, Zongyue Wang, and Guorong Cai. Depth matters: Exploring deep interactions of rgb-d for semantic segmentation in traffic scenes. arXiv preprint arXiv:2409.07995, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.732, + 0.484, + 0.801 + ], + "angle": 0, + "content": "[9] Siyu Chen, Ting Han, Changshe Zhang, Jinhe Su, Ruisheng Wang, Yiping Chen, Zongyue Wang, and Guorong Cai. Hspformer: Hierarchical spatial perception transformer for semantic segmentation. IEEE Transactions on Intelligent Transportation Systems, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[10] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In CVPR, pages 11580-11590, 2021. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[11] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.247, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "benchmark. https://github.com/open-mmlab/mmsegmentation,2020.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[12] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, pages 3213-3223, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[13] Jian Ding, Nan Xue, Gui-Song Xia, Bernt Schiele, and Dengxin Dai. Hgformer: Hierarchical grouping transformer for domain generalized semantic segmentation. In CVPR, pages 15413-15423, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.314 + ], + "angle": 0, + "content": "[14] Siqi Du, Weixi Wang, Renzhong Guo, Ruisheng Wang, and Shengjun Tang. Asymformer: Asymmetrical cross-modal representation learning for mobile platform real-time rgb-d semantic segmentation. In CVPRW, pages 7608-7615, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[15] Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva-02: A visual representation for neon genesis. arXiv preprint arXiv:2303.11331, 2023. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.439 + ], + "angle": 0, + "content": "[16] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In CVPR, pages 19358-19369, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[17] Ting Han, Siyu Chen, Chuanmu Li, Zongyue Wang, Jinhe Su, Min Huang, and Guorong Cai. Epurate-net: Efficient progressive uncertainty refinement analysis for traffic environment urban road detection. IEEE Transactions on Intelligent Transportation Systems, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[18] Sivan Harary, Eli Schwartz, Assaf Arbelle, Peter Staar, Shady Abu-Hussein, Elad Amrani, Roei Herzig, Amit Alfassy, Raja Giryes, Hilde Kuehne, Dina Katabi, Kate Saenko, Rogerio Feris, and Leonid Karlinsky. Unsupervised domain generalization by learning a bridge across domains. In CVPR, pages 5270-5280, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.596, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[19] Zongtao He, Liuyi Wang, Ronghao Dang, Shu Li, Qingqing Yan, Chengju Liu, and Qijun Chen. Learning depth representation from rgb-d videos by time-aware contrastive pretraining. IEEE TCSVT, 34(6):4143-4158, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[20] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[21] Lei Huang, Yi Zhou, Fan Zhu, Li Liu, and Ling Shao. Iterative normalization: Beyond standardization towards efficient whitening. In CVPR, pages 4874-4883, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[22] Wei Huang, Chang Chen, Yong Li, Jiacheng Li, Cheng Li, Fenglong Song, Youliang Yan, and Zhiwei Xiong. Style projected clustering for domain generalized semantic segmentation. In CVPR, pages 3061-3071, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[23] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, pages 709-727. Springer, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Jin Kim, Jiyoung Lee, Jungin Park, Dongbo Min, and Kwanghoon Sohn. Pin the memory: Learning to generalize semantic segmentation. In CVPR, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8293" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.148 + ], + "angle": 0, + "content": "[25] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[26] Suhyeon Lee, Hongje Seong, Seongwon Lee, and Euntai Kim. Wildnet: Learning domain generalized semantic segmentation from the wild. In CVPR, pages 9936-9946, 2022. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[27] Zhiyu Liu, Munawar Hayat, Hong Yang, Duo Peng, and Yinjie Lei. Deep hypersphere feature regularization for weakly supervised rgb-d salient object detection. IEEE TIP, 32: 5423-5437, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[28] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In ICCV, pages 4990-4999, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[29] Joshua Niemeijer, Manuel Schwonberg, Jan-Aike Termöhlen, Nico M Schmidt, and Tim Fingscheidt. Generalization by adaptation: Diffusion-based domain extension for domain-generalized semantic segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2830-2840, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[30] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.475, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[31] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In ECCV, pages 464-479, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.518, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[32] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In CVPR, pages 1863-1871, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[33] Duo Peng, Yinjie Lei, Lingqiao Liu, Pingping Zhang, and Jun Liu. Global and local texture randomization for synthetic-to-real semantic segmentation. IEEE TIP, 30: 6594-6608, 2021. 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[34] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In CVPR, pages 2584-2595, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.703 + ], + "angle": 0, + "content": "[35] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In CVPR, pages 2594-2605, 2022. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[37] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In ECCV, pages 102-118. Springer, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[38] Giulia Rizzoli, Donald Shenaj, and Pietro Zanuttigh. Source-free domain adaptation for rgb-d semantic segmentation with vision transformers. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops, pages 615-624, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[39] Giulia Rizzoli, Donald Shenaj, and Pietro Zanuttigh. Source-free domain adaptation for rgb-d semantic segmentation with vision transformers. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 615-624, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.162, + 0.906, + 0.217 + ], + "angle": 0, + "content": "[40] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In ICCV, pages 10745-10755, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.218, + 0.906, + 0.272 + ], + "angle": 0, + "content": "[41] Yang Shu, Zhangjie Cao, Chenyu Wang, Jianmin Wang, and Mingsheng Long. Open domain generalization with domain-augmented meta-learning. In CVPR, pages 9619-9628, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.906, + 0.315 + ], + "angle": 0, + "content": "[42] Zhiqiang Tang, Yunhe Gao, Yi Zhu, Zhi Zhang, Mu Li, and Dimitris Metaxas. Crossnorm and selfnorm for generalization under distribution shifts. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.906, + 0.37 + ], + "angle": 0, + "content": "[43] Jie Wang, Kechen Song, Yanqi Bao, Liming Huang, and Yunhui Yan. Cgfnet: Cross-guided fusion network for rgb-t salient object detection. IEEE TCSVT, 32(5):2949-2961, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.371, + 0.906, + 0.44 + ], + "angle": 0, + "content": "[44] Zhixiang Wei, Lin Chen, Yi Jin, Xiaoxiao Ma, Tianle Liu, Pengyang Ling, Ben Wang, Huaian Chen, and Jinjin Zheng. Stronger fewer & superior: Harnessing vision foundation models for domain generalized semantic segmentation. In CVPR, pages 28619-28630, 2024. 1, 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.441, + 0.906, + 0.496 + ], + "angle": 0, + "content": "[45] Binhui Xie, Shuang Li, Mingjia Li, Chi Harold Liu, Gao Huang, and Guoren Wang. Sepico: Semantic-guided pixel contrast for domain adaptive semantic segmentation. IEEE TPAMI, 45(7):9004-9021, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.906, + 0.566 + ], + "angle": 0, + "content": "[46] Qi Xu, Liang Yao, Zhengkai Jiang, Guannan Jiang, Wenqing Chu, Wenhui Han, Wei Zhang, Chengjie Wang, and Ying Tai. Dirl: Domain-invariant representation learning for generalizable semantic segmentation. In AAAI, pages 2884-2892, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.906, + 0.635 + ], + "angle": 0, + "content": "[47] Qi Xu, Liang Yao, Zhengkai Jiang, Guannan Jiang, Wenqing Chu, Wenhui Han, Wei Zhang, Chengjie Wang, and Ying Tai. Dirl: Domain-invariant representation learning for generalizable semantic segmentation. In AAAI, page 2884-2892, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.636, + 0.906, + 0.691 + ], + "angle": 0, + "content": "[48] Jun Yang, Lizhi Bai, Yaoru Sun, Chunqi Tian, Maoyu Mao, and Guorun Wang. Pixel difference convolutional network for rgb-d semantic segmentation. IEEE TCSVT, 34(3):1481-1492, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.906, + 0.734 + ], + "angle": 0, + "content": "[49] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.906, + 0.775 + ], + "angle": 0, + "content": "[50] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. NeurIPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.906, + 0.831 + ], + "angle": 0, + "content": "[51] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, pages 2636-2645, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[52] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In ICCV, pages 2100-2110, 2019. 4, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8294" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[53] Xiangyu Yue, Zangwei Zheng, Shanghang Zhang, Yang Gao, Trevor Darrell, Kurt Keutzer, and Alberto Sangiovanni Vincentelli. Prototypical cross-domain self-supervised learning for few-shot unsupervised domain adaptation. In CVPR, pages 13834-13844, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[54] Jiaming Zhang, Huayao Liu, Kailun Yang, Xinxin Hu, Ruiping Liu, and Rainer Stiefelhagen. Cmx: Cross-modal fusion for rgb-x semantic segmentation with transformers. IEEE Transactions on Intelligent Transportation Systems, 24(12): 14679-14694, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[55] Lu Zhang, Siqi Zhang, Xu Yang, Hong Qiao, and Zhiyong Liu. Unseen object instance segmentation with fully test-time rgb-d embeddings adaptation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4945-4952, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.304, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[56] Xingxuan Zhang, Linjun Zhou, Renzhe Xu, Peng Cui, Zheyan Shen, and Haoxin Liu. Towards unsupervised domain generalization. In CVPR, pages 4900-4910, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[57] Yuyang Zhao, Zhun Zhong, Na Zhao, Nicu Sebe, and Gim Hee Lee. Style-hallucinated dual consistency learning for domain generalized semantic segmentation. In ECCV, pages 535-552. Springer, 2022. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.404, + 0.482, + 0.457 + ], + "angle": 0, + "content": "[58] Zhun Zhong, Yuyang Zhao, Gim Hee Lee, and Nicu Sebe. Adversarial style augmentation for domain generalized urban-scene segmentation. NeurIPS, 35:338-350, 2022. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.46, + 0.482, + 0.514 + ], + "angle": 0, + "content": "[59] Wujie Zhou, Yun Zhu, Jingsheng Lei, Jian Wan, and Lu Yu. Ccafnet: Crossflow and cross-scale adaptive fusion network for detecting salient objects in rgb-d images. IEEE TMM, 24: 2192-2204, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.517, + 0.482, + 0.585 + ], + "angle": 0, + "content": "[60] Wujie Zhou, Yuxiang Xiao, Weiqing Yan, and Lu Yu. Cmpffnet: Cross-modal and progressive feature fusion network for rgb-d indoor scene semantic segmentation. IEEE Transactions on Automation Science and Engineering, 21 (4):5523-5533, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "8295" + } + ] +] \ No newline at end of file diff --git a/2025/Stronger, Steadier & Superior_ Geometric Consistency in Depth VFM Forges Domain Generalized Semantic Segmentation/2e6ee106-ae7d-42e3-b798-bd6f97fafe0a_origin.pdf b/2025/Stronger, Steadier & Superior_ Geometric Consistency in Depth VFM Forges Domain Generalized Semantic Segmentation/2e6ee106-ae7d-42e3-b798-bd6f97fafe0a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bea1b46d35833e326dbedbf803451d38384c1a36 --- /dev/null +++ b/2025/Stronger, Steadier & Superior_ Geometric Consistency in Depth VFM Forges Domain Generalized Semantic Segmentation/2e6ee106-ae7d-42e3-b798-bd6f97fafe0a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:740fd0c72a3ae56ebfd173c5380bdaa40aef29e0c2a768248c9e35a234f59f96 +size 9638557 diff --git a/2025/Stronger, Steadier & Superior_ Geometric Consistency in Depth VFM Forges Domain Generalized Semantic Segmentation/full.md b/2025/Stronger, Steadier & Superior_ Geometric Consistency in Depth VFM Forges Domain Generalized Semantic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7677024f51e2652cfe31bed968b8f61a066b14b9 --- /dev/null +++ b/2025/Stronger, Steadier & Superior_ Geometric Consistency in Depth VFM Forges Domain Generalized Semantic Segmentation/full.md @@ -0,0 +1,322 @@ +# Stronger, Steadier & Superior: Geometric Consistency in Depth VFM Forges Domain Generalized Semantic Segmentation + +Siyu Chen $^{1,3}$ Ting Han $^{2\dagger}$ Changshe Zhang $^{4}$ Xin Luo $^{1}$ Meiliu Wu $^{3}$ Guorong Cai $^{1}$ Jinhe Su $^{1\dagger}$ $^{1}$ Jimei University $^{2}$ Sun Yat-sen University $^{3}$ University of Glasgow $^{4}$ Xidian University + +![](images/b00ec6b05b120f399e8146a8fa271322bbfc3d8c0da48e0c6e092787c307ac9a.jpg) +(a) Stronger Pre-trained Models + +![](images/4b8dec8ad21a3f5f4285c0fa8d8824cb72327690bc80e1b3879c5d4fa7760631.jpg) + +![](images/5d1f42bc025ad9277429553c8f6547319aa59db64ed5a11beb31ae6bb689edba.jpg) +Figure 1. Existing methods fail to fully exploit the potential of VFMs for DGSS when visual cues are limited or absent. To this end, we introduce a novel and robust fine-tuning strategy DepthForge that leverages depth cues from a depth VFM to optimize visual cues, and employs spatial consistency to enhance feature discriminability. The proposed DepthForge achieves (a) stronger performance under extreme conditions, owing to (b) steadier visual-spatial attention, and thus delivers (c) superior generalization ability. + +![](images/6b7cc3a8b6410ec6a46f86182d827b14a6e730d2133e2b0559445cb0c8f1b357.jpg) + +![](images/5bebe333f583b07bafc4a84625e25934c5b7fc473d5194db273edf4d47e57c63.jpg) +(b) Steadier Visual-Spatial Attention + +![](images/404c9c4bd65d4196835cabc581088fc71527c3762e1996a9761927d2e4776298.jpg) + +![](images/35721a10669bd2715a1940fb6a0bcbda95f703b66e45d88b96b575086c8151cb.jpg) +(c) Superior Generalization Ability + +![](images/e68f05c791d5149a7865b6aab14ec2743329fa1491a4c52e7625a5282c428f79.jpg) + +![](images/01e80c623aa943409eb7233995ac28fcab93ae80065a4711786ac081201c27a6.jpg) + +# Abstract + +Vision Foundation Models (VFMs) have delivered remarkable performance in Domain Generalized Semantic Segmentation (DGSS). However, recent methods often overlook the fact that visual cues are susceptible, whereas the underlying geometry remains stable, rendering depth information more robust. In this paper, we investigate the potential of integrating depth information with features from VFMs, to improve the geometric consistency within an image and boost the generalization performance of VFMs. We propose a novel fine-tuning DGSS framework, named Depth-Forge, which integrates the visual cues from frozen DINOv2 or EVA02 and depth cues from frozen Depth Anything V2. In each layer of the VFMs, we incorporate depth-aware learnable tokens to continuously decouple domain-invariant visual and spatial information, thereby enhancing depth awareness and attention of the VFMs. Finally, we develop a depth refinement decoder and integrate it into the model architecture to adaptively refine multi-layer VFM features and depth-aware learnable tokens. Extensive experiments are conducted based on various DGSS settings and five different datasets as unseen target domains. The qualitative and quantitative results demonstrate that our method significantly outperforms alternative approaches with stronger performance, steadier visual-spatial attention, and superior generalization ability. In particular, + +DepthForge exhibits outstanding performance under extreme conditions (e.g., night and snow). Code is available at https://github.com/SY-Ch/DepthForge. + +# 1. Introduction + +Domain Generalized Semantic Segmentation (DGSS) aims to improve prediction accuracy across multiple unseen domains without requiring access to their data, thereby ensuring robust generalization for practical applications [34, 45, 53]. One common approach involves decomposing the learned features into domain-invariant and domain-specific components [42, 47]. This method aims to achieve robustness against domain variations by isolating content-related factors from those that vary between domains. A second strategy employs meta-learning techniques [24], which focus on training models to generalize effectively across domains by learning more adaptable representations, often leveraging shared structures across different datasets. + +Recently, VFMs have emerged as a promising approach, primarily owing to their strong generalization capabilities, which are enabled by pre-trained on large-scale datasets [15, 25, 36]. Fine-tuning these models has become increasingly popular due to their cost-effectiveness and superior performance [2, 44]. However, a significant challenge remains: what constitutes robust feature information for domain generalization, and how can such robust informa + +tion be effectively extracted? Previous methods predominantly utilize VFMs based on RGB images. However, the inherent limitations of RGB prevent those methods from adapting to variable conditions, especially in scenarios such as nighttime, overexposure, snow, or fog, where visual cues are not clearly present. This clearly fails to meet the requirements of domain-generalized semantic segmentation. + +In this paper, we introduce DepthForge, a robust fine-tune and optimization strategy for domain generalized semantic segmentation, as shown in Fig. 1. Specifically, we employ two frozen VFMs: Depth Anything V2[49, 50] is dedicated to extract depth information, and DINOv2[30] or EVA02[15] are used to process RGB visual features. We enhance the generalization of the visual features at each layer of the VFMs by integrating the corresponding depth cues. More importantly, we introduce depth-aware learnable tokens that enables the model to learn spatial structure invariance, thereby ensuring consistent performance across different domains. The depth and visual features extracted from frozen VFMs are used to optimize the representation of our learnable tokens, thus producing more robust visual-spatial attention features than those based solely on visual cues and forging geometric consistency, as shown in Fig. 1(b). However, we observe the frozen VFMs only provide static features, they are incapable of refining the feature maps during training. This limitation leads to persistent errors in the learnable tokens, causing the optimization process to drift in incorrect directions that are difficult to rectify. To this end, we propose a depth refinement decoder that dynamically adjusts the multi-scale features, combining prior and dynamic features to establish high-quality paired feature relationships at different depth spaces. + +Our core contribution is to demonstrate that robust depth features and learnable tokens can be effectively applied to improve model performance in Domain Generalized Semantic Segmentation (DGSS). To the best of our knowledge, the proposed DepthForge is the first depth-aware framework specifically designed for DGSS. The effectiveness of our design is verified by experiments on widely used Cityscapes, ACDC, Mapillary, BDD100k, and GTA5 datasets. In addition, the individual components of our design are also verified by extensive experiments. Extensive experiments have demonstrated that our method significantly improves DGSS performance, as shown in Fig. 1(c). Specifically, in the $GTA \rightarrow$ Cityscapes + BDD100k + Mapillary setting, we observed approximately $+4\%$ significant improvement in mIoU, particularly the improvement in extreme scenes is close to $+5\%$ . Our main contributions are summarized as follows: + +- We propose a stronger DepthForge learning framework to fine-tune VFMs combined with visual and depth cues for domain-generalized semantic segmentation. +- A depth-awareness learnable tokens are introduced to + +![](images/5d2b1bdf4d83ad194b61d9a94ed479f728e9cc9ff9a658e425fb35625df46509.jpg) +Figure 2. Different configurations of incorporating depth cues into visual information, where Config.1: adding $f_{i}^{d}$ to $f_{i}^{v}$ , or Config.2: incorporating $f_{i}^{d}$ into the learnable token $T_{i}$ , and Config.3: our DepthForge, respectively. + +focus on spatial relationship that acquires steadier and more robust visual-spatial attention. This strategy enables DepthForge to enhance the discriminability and refine features at an instance level within each layer. + +- Depth Forge highlights the impressive generalization ability, surpassing existing DGSS methods by a large margin, especially in extreme scenes with limited and absent visual cues. + +# 2. Related Work + +# 2.1. Domain Generalized Semantic Segmentation + +Domain Generalized Semantic Segmentation (DGSS) aims to enhance the generalization capability of segmentation models to unseen domains. Recent methods primarily focus on reducing feature distribution discrepancies across domains and learning more domain-invariant representations. Adversarial training [53], meta-learning [4, 41], and self-supervised learning [18, 56] have been extensively explored to improve generalization performance. However, these methods often rely on constructing cross-domain contrastive constraints or employing sophisticated training strategies, limiting their generalization improvements and complicating the training process. The emergence of Vision Foundation Models (VFMs), trained via large-scale image-text [36] or image-image contrastive learning [30], has opened a promising avenue for DGSS. VFMs capture universal and rich visual representations beneficial for reducing domain biases. Recent approach [2, 44] leverages pretrained VFMs by employing their frozen features as powerful generalized backbones to extract and refine learnable tokens, leading to significant improvements in model performance. Nevertheless, current DGSS approaches predominantly utilize RGB-only data, leaving the models susceptible to variations in lighting, texture, and spatial arrangements across different domains, thus constraining their full potential for generalization. + +![](images/4221b9074ef45cc30b2139057816e8c20379cf5ff037090ebabac7ba3eca537e.jpg) +Figure 3. DepthForge framework. It incorporates the features from frozen VFM and depth VFM into our Depth-aware Learnable Tokens to generate enhanced features $f_{i+1} = f_i^v + \text{DepthForge}(f_i^v, f_i^d)$ for each layer. In our tunable DepthForge, We first design a Depth Awareness module that integrates visual and depth cues into learnable tokens. Next, we propose an Attention Optimization module to strengthen the discriminability of instances across different spatial locations. Finally, multi-scale features are fused in Depth Refinement Decoder to adapt to various scales and spatial connections. Our DepthForge establishes direct connections to different instances in real-world locations, facilitating both feature refinement and improved discriminability. + +# 2.2. RGB-D Segmentation + +Combining RGB with depth information has shown substantial benefits across various visual tasks [6, 8, 9, 17, 39]. In object detection, RGB-D effectively mitigates limitations inherent to RGB data under complex scenes by providing robust geometric and spatial cues, thereby enhancing localization accuracy and model robustness [43, 59]. In semantic segmentation, depth information assists models in capturing spatial structures and geometric contours of objects, alleviating issues related to imprecise object boundary delineation. Particularly in scenarios with blurred boundaries, insufficient illumination, or texture scarcity, the incorporation of depth information substantially improves segmentation accuracy and stability [14, 48, 54, 60]. Recent studies have further indicated that even in domain generalization tasks, integrating RGB and depth features consistently improves the model's generalization performance on unseen domains [19, 27, 38, 55]. Therefore, exploring the integration of depth information to address the limitations of RGB-only approaches represents a promising and essential research direction within DGSS. + +# 3. Preliminary + +To achieve better generalization performance, existing approaches opt to fine-tune VFMs in a parameter-efficient manner. One mainstream idea is to refine and propagate features from frozen VFM layers to subsequent ones. Specif + +ically, given a pre-trained VFM composed of $N$ sequential layers $\{L_1,L_2,\dots,L_N\}$ , where each layer is associated with a weight matrix $W_{i}\in \mathbb{R}^{c\times c}$ and the frozen output features of each layer are denoted as $f_{i}$ for $i = 1,2,\ldots ,N$ . If we have a learnable weight matrix as $\Delta W_{i}\in \mathbb{R}^{c\times c}$ , the feature propagation from layer $L_{i}$ to $L_{i + 1}$ is then formulated as: + +$$ +f _ {i + 1} = W _ {i} f _ {i} + \Delta W _ {i} f _ {i}. \tag {1} +$$ + +Recent REIN transfers the learnable matrix $\Delta W_{i}$ into a learnable token $T_{i}$ : + +$$ +f _ {i + 1} = W _ {i} f _ {i} + \epsilon \left(T _ {i} \left(W _ {i} f _ {i}\right)\right), \tag {2} +$$ + +where $\epsilon (\cdot)$ denotes the MLP function, and $T_{i}\in \mathbb{R}^{m\times c}$ with a tokn length $m$ . REIN enables each token to connect better with instances in an image. However, we observe that since the updated and frozen features are limited to remain similar, this correlation is fragile and the capacity for feature refinement is limited. + +This prompted us to explore what kind of information could improve discrimination in instances and refine the feature representations. Drawing inspiration from multimodal learning, we observe that depth features provide a spatial relationship bias for corresponding pixels in visual features, thereby compensating for the loss of crucial visual cues in DGSS. We assume that incorporating a pretrained depth VFM into the pre-trained VFMs will alleviate + +
MethodProc. & YearSnowNightFogRain
ResNet based:
IBN [31]ECCV201849.621.263.850.4
Itenorm [21]CVPR201949.923.863.350.1
IW [32]CVPR201947.621.862.452.4
RobustNet [10]CVPR202149.824.364.356.0
WildNet [26]CVPR202228.412.741.234.2
Transformer based:
HGFormer [13]CVPR202368.652.769.972.0
CMFormer [3]AAAI202464.333.777.867.6
VFM based:
REIN [44]CVPR202470.655.979.572.5
FADA [2]NeurIPS202473.557.480.275.0
DepthForge(Ours)-75.460.481.275.4
+ +attention bias and significantly enhance scene understanding, particularly under extreme conditions. + +We use the frozen features $f_{i}^{v}$ from VFM for visual embedding, and the frozen features $f_{i}^{d}$ from depth VFM for depth embedding. A straightforward thought might simply add $f_{i}^{d}$ to $f_{i}^{v}$ , or incorporate $f_{i}^{d}$ into the leanable token $T_{i}$ . However, these methods lack flexibility and introduce significant inter-layer feature differences that prevent convergence, as shown in Fig. 2. In contrast, we propose Depth-Forge to adaptively fuse visual and depth cues, denoted as: + +$$ +f _ {i + 1} = W _ {i} ^ {v} f _ {i} + \varepsilon \left(\epsilon^ {v} \left(T _ {i} \left(W _ {i} ^ {v} f _ {i} ^ {v}\right)\right) + \epsilon^ {d} \left(T _ {i} \left(W _ {i} ^ {d} f _ {i} ^ {d}\right)\right)\right), \tag {3} +$$ + +where $W_{i}^{v}$ and $W_{i}^{d}$ represent the weight matrix from VFMs and depth VFM, respectively. $\varepsilon$ denotes the adaptation of the visual and depth signals. This approach allows us to more effectively utilize the powerful capacities of multimodal VFMs to forge robust refined features and instance discriminability. + +More importantly, the depth VFM itself possesses strong generalization capabilities, allowing domain adaptation without fine-tuning. A pre-trained VFM applies the same. Therefore, the overall optimization objective still focuses on the segmentation head $\mathcal{H}$ with parameter $\theta_{h}$ and our Depth-Forge with parameter $\theta_{D}$ : + +$$ +\arg \min _ {\theta_ {D}, \theta_ {h}} \sum_ {i = 1} ^ {N} \mathcal {L o s s} \left(\mathcal {H} _ {\theta_ {i}} \left(\mathcal {F} _ {\Theta^ {v}, \Theta^ {d}, \theta_ {D}} \left(x _ {i}\right)\right), y _ {i}\right), \tag {4} +$$ + +where $x_{i}$ and $y_{i}$ denote the input image and its corresponding ground truth, respectively, and $N$ signifies the total number of images. $\mathcal{F}$ represent the forward process of our DepthForge strategy applied to the pre-trained VFMs with parameters $\Theta^v$ and $\Theta^d$ . + +Table 1. Performance comparison between our DepthForge and existing DGSS methods under Cities. $\rightarrow$ ACDC with Snow, Night, Fog, Rain generalization setting. Top three results are highlighted as best, second, and third, respectively. (%) + +
MethodProc. & YearBDD100kMapillaryGTA
ResNet based:
IBN [31]ECCV201848.5657.0445.06
Itenorm [21]CVPR201949.2356.2645.73
IW [32]CVPR201948.4955.8244.87
RobustNet [10]CVPR202150.7358.6445.00
DRPC [52]ICCV201949.8656.3445.62
GTR [33]TIP202150.7557.1645.79
SHADE [57]ECCV202250.9560.0748.61
SAW [35]CVPR202252.9559.8147.28
WildNet [26]CVPR202250.9458.7947.01
BlindNet [1]CVPR202451.8460.1847.97
Transformer based:
HGFormer [13]CVPR202353.4066.9051.30
CMFormer [3]AAAI202459.2771.1058.11
VFM based:
REIN [44]CVPR202463.5474.0362.41
FADA [2]NeurIPS202465.1275.8663.78
DepthForge(Ours)-66.1975.9367.24
+ +Table 2. Performance comparison between our DepthForge and existing DGSS methods under the Citys. $\rightarrow$ BDD. $^+$ Map. $^+$ GTA5 generalization setting. Top three results are highlighted as best, second, and third, respectively. $(\%)$ + +# 4. DepthForge + +This paper introduces the DepthForge strategy, a novel approach for domain generalized semantic segmentation. As shown in Fig. 3, DepthForge consists of three main parts: (1) Depth Awareness; (2) Attention Optimization; (3) Depth Refinement Decoder. We detail the designs as follows. + +# 4.1. Depth Awareness + +Recent DGSS methods have applied pre-trained VFMs to source domain segmentation, with the core idea being to enhance the features between frozen VFM layers to help the model better adapt to target tasks. However, we observe that visual cues exhibit significant variations and even deficiencies across different scenes. Consequently, a more robust cue is required to provide a stronger prior for domain generalization. This motivates us to explore a strategy that focuses on the inherent similarities across different domains. In general, spatial location priors are able to help differentiate instances even when visual cues fail. To this end, we employ an additional depth VFM as an auxiliary network, as shown in Fig. 3. + +Specifically, given a VFM $\mathcal{V}$ and a depth VFM $\mathcal{D}$ with $N$ layers $\{L_1,L_2,\dots ,L_N\}$ , the corresponding features are denoted as $\{f_i^v,f_i^d |f_i\in \mathbb{R}^{n\times c},i\in [1,N]\}$ . Note that the layers $\{L_1,L_2,\dots ,L_N\}$ are kept frozen. We acquire the visual feature $f_{i}^{v}$ from the $i$ -th layer $L_{i}$ of $\mathcal{V}$ , and the depth feature $f_{i}^{d}$ from the corresponding layer of $\mathcal{D}$ . Our goal is to generate the enhanced feature $\Delta f_{i} = DepthForge(f_{i}^{v},f_{i}^{d})$ by using both visual and depth cues. + +We employ a set of learnable tokens $T = \{T_{i} \in$ + +![](images/005b469f1ef48242c248d16711005b6f0c8e73a849a4d4acb67b085487787479.jpg) +Figure 4. Key segmentation examples of existing DGSS methods and DepthForge under the Cityscapes to ACDC unseen target domains under Night, Snow, Rain, and Fog conditions. + +![](images/fbce90875e1e88a2a3ebf89c75055fd6fe37b0abc305baf7fd944ed0ca66f19a.jpg) +Figure 5. Key segmentation examples of existing DGSS methods and DepthForge under the GTA5 to Mapillary, BDD100k, and Cityscapes unseen target domains setting. + +$\mathbb{R}^{m \times c} |i \in [1, N]\}$ to adaptively learn scene knowledge. As we consider, relying solely on visual cues can be limiting and prone to failure. Therefore, we utilize visual and depth cues to construct learnable tokens, which we term depth-aware learnable tokens. We integrate $T_i$ with $f_i^v$ and $f_i^d$ using an attention-based mechanism, prompting tailored adjustments to bridge the gap between different domains. + +We use $f_{i}^{v,d}$ as the query, and $T_{i}$ as the key and the value A dot-product operation is applied to generate awareness map $A_{i}^{v,d}$ , and the Softmax function (S) is employed to align each patch with a unique instance. The $A_{i}^{v,d}$ provides insights into both the visual cues and depth retrieval, qualitatively evaluating the relationship between various tokens and visual or depth cues, respectively. Subsequently, we incorporate depth awareness bias $A_{i}^{d}$ into the visual features $A_{i}^{v}$ to provide additional spatial relationships, thereby establishing a latent connection between pixels and the real world. The process is mathematically formulated as: + +$$ +A _ {i} = S \left(\frac {f _ {i} ^ {v} \times (T ^ {v}) _ {i} ^ {\top}}{\sqrt {c}}\right) + \lambda S \left(\frac {f _ {i} ^ {d} \times (T ^ {d}) _ {i} ^ {\top}}{\sqrt {c}}\right). \tag {5} +$$ + +# 4.2. Attention Optimization + +Directly leveraging the awareness representation $A_{i}$ to generate enhanced features may introduce a significant amount of irrelevant information for downstream tasks. To this end, we employ depth-aware learnable tokens $T_{i}$ to learn an alignment weight $W_{T_i}$ and bias $b_{T_i}$ . The $T_{i}$ with weight $W_{T_i}$ and bias $b_{T_i}$ is then multiplied with the $A_{i}$ to obtain the initial enhanced features $\Delta f_{i}$ that embed both visual and depth cues, using the equation: + +$$ +\Delta \hat {f} _ {i} = A _ {i} \times \left(T _ {i} \times W _ {T _ {i}} + b _ {T _ {i}}\right), \tag {6} +$$ + +where we select the same optimization strategy as REIN, discarding high-weight features when the features are sufficiently precise to avoid unnecessary adjustments and mitigate the risk of inappropriate variations. Finally, we employ two consecutive residual connections $\phi$ to enhance the flexibility of attention feature adjustment and prevent gradient vanishing during the forward pass. The calculation is defined as $\Delta f_{i} = \phi \hat{\Delta f}_{i}$ . Benefiting from these depth-aware and attention optimization adjustments, DepthForge is ca + +![](images/39a093aa7a9e91b204c2935320ba498161e2e34e4f4f84c4276e478e503fc41c.jpg) +Figure 6. t-SNE visualization compares REIN and our Depth-Forge. The first volume corresponds to $\text{Citys.} \rightarrow \text{ACDC}$ , and two through four volumes correspond to $\text{GTA5} \rightarrow \text{BDD.} + \text{Citys.} + \text{Map}$ . The best mIoU results are highlighted in red. + +pable of generating more distinctive, discriminative features for different categories within a single image, while balancing both visual and depth cues. + +# 4.3. Depth Refinement Decoder + +The depth-aware features at different layers are inconsistent. Therefore, how to design a robust decoder becomes a core challenge that effectively integrate these enhanced features across various scales to accomplish downstream tasks. + +Previous methods rely solely on the output of the final layer $f_{N}$ for semantic segmentation predictions. In contrast, we fuse the outputs from layers $L_{1} - L_{N}$ to refine the depth features. Specifically, we process these multi-layer features with an MLP with two different fully-connected layers $\phi$ and $\varphi$ and then concatenate them after a ReLU activation function: + +$$ +f _ {i} ^ {\text {o u t}} = \phi (R e L U (\varphi f _ {i})), \tag {7} +$$ + +$$ +F _ {u} = \gamma \operatorname {C o n c a t} \left(f _ {1} ^ {\text {o u t}}, f _ {2} ^ {\text {o u t}}, \dots , f _ {N} ^ {\text {o u t}}\right), \tag {8} +$$ + +where $\gamma$ is a convolutional layer. Finally, we design several sequential multi-head Transformer layers to generate the final predictions. + +# 5. Experiment + +# 5.1. Datasets & Evaluation Protocols + +We evaluate our DepthForge on four real-world datasets (Cityscapes [12], BDD100K [51], Mapillary [28], ACDC [40]) and a synthetic dataset (GTA5 [37]). Cityscapes (Citys.) is an autonomous driving dataset comprising 2,975 training images and 500 validation images, each with a resolution of $2048 \times 1024$ . BDD100K (BDD.) and Mapillary (Map.) provide 1,000 and 2,000 validation images, with resolutions of $1280 \times 720$ and $1902 \times 1080$ , respectively. ACDC offers 406 validation images captured under extreme conditions: namely at Night, Snow, Fog, and Rain, each with a resolution of $1902 \times 1080$ . GTA5 is a + +
MethodProc. & YearCityscapesBDD100kMapillary
ResNet based:
DRPC [52]ICCV201937.4232.1434.12
GTR [33]TIP202137.5333.7534.52
RobustNet [10]CVPR202136.5835.2040.33
DIRL [46]AAAI202241.0439.1541.60
SHADE [57]ECCV202244.6539.2843.34
WildNet [26]CVPR202344.6238.4246.09
SAW [35]CVPR202239.7537.3441.86
AdvStyle [58]NeurIPS202239.6235.5437.00
SPC [22]CVPR202344.1040.4645.51
BlindNet [1]CVPR202445.7241.3247.08
Transformer based:
CMFormer [3]AAAI202455.3149.9160.09
VFM based:
DIDEX [29]WACV202462.0054.3063.00
REIN [44]CVPR202466.4060.4066.10
FADA [2]NeurIPS202468.2361.9468.09
DepthForge(Ours)-69.0462.8269.22
+ +Table 3. Performance comparison between our DepthForge and existing DGSS methods under the $GTA5 \rightarrow Citys$ . + BDD. + Map. generalization setting. Top three results are highlighted as best, second, and third, respectively. (%) + +synthetic dataset that provides 24,966 annotated images obtained from the game. + +Following the evaluation protocol of the existing DGSS methods, a certain dataset is chosen as the source domain for training, and the other three datasets serve as unseen target domains for validation. The three common evaluation settings are as follows: (1) Cityscapes $\rightarrow$ ACDC (Night, Snow, Fog, Rain); (2) GTA5 $\rightarrow$ Cityscapes, BDD100K, Mapillary; and (3) Cityscapes $\rightarrow$ BDD100K, Mapillary, GTA5. The evaluation metric is the mean Intersection-over-Union (mIoU). + +# 5.2. Deployment Details & Parameter Settings + +Our implementation is built on the MMSegmentation [11] codebase. For optimization, we employ AdamW optimizer with an initial learning rate of 1e-4, a weight decay of 0.05, an epsilon of 1e-8, and beta values of (0.9, 0.999). A OneCycleLR scheduler is applied over 50,000 total steps with a maximum learning rate of 1e-4, featuring a warmup phase covering $10\%$ of the training iterations, and using cosine annealing, with both the div_factor and final_div_factor set to 10. + +Data augmentation includes multi-scale resizing, random cropping (with a fixed crop size and category ratio constraints), random flipping, and photometric distortion. All experiments are conducted on an NVIDIA RTX 3090 GPU. + +# 5.3. Comparison with State-of-The-Art Methods + +We compare our approach with existing DGSS methods, including: (1) ResNet-based methods, (2) Transformer-based methods, and (3) VFM-based methods. + +Cityscapes $\rightarrow$ ACDC. Tab. 1 presents the performance + +![](images/cb236a937283ea96528763b2ca03ae69e2fac2235b56c87c7c1c71c50eefdce0.jpg) +Figure 7. Visualizations of attention maps and affinity maps under differnet configurations, where DRD denotes the Depth Refinement Decoder, DA and AO represent the Depth Awareness and Attention Optimization, respectively. + +![](images/6b6efc08664d9163e81415a3b3f43b223a1a74295bafb4ffcdc78d73bfdb058b.jpg) + +
BackboneConfigurationsSnowNightFogRainBDDMap
EVA02 (Large) [15, 16]REIN [44]66.052.176.069.560.070.0
+ concat fi d66.352.276.269.860.270.2
+ Prompt Depth Anything65.852.475.569.360.469.9
+ Depth Anything V267.253.077.070.260.970.8
+ DRD67.753.577.670.761.271.0
+ DA + AO67.953.777.870.961.571.2
DepthForge69.554.578.572.062.073.0
DINOv2 (Large) [30]REIN [44]70.655.979.572.563.574.0
+ concat fi d70.856.179.472.863.674.2
+ Prompt Depth Anything70.455.579.772.063.873.9
+ Depth Anything V271.556.780.573.364.474.5
+ DRD72.057.280.774.064.974.8
+ DA + AO72.257.480.974.265.175.0
DepthForge75.460.481.275.466.276.0
+ +Table 4. Ablation studies on component configurations of the proposed DepthForge under Citys. $\rightarrow$ ACDC with Snow, Night, Fog, Rain + BDD. + Map. generalization setting, where DRD denotes the Depth Refinement Decoder, DA and AO represent the Depth Awareness and Attention Optimization. Top three results are highlighted as best, second, and third, respectively. (%) + +
No.DesignAvg. Gain (%)RainNightSnowRog
1No Depth-70.655.979.572.5
2Concatenation-1.169.855.078.371.2
3Depth Token1.672.357.880.774.0
4Config10.270.856.179.472.8
5Config20.771.556.880.173.0
6Config3, Ours3.575.460.481.275.4
7w/o Scale Factor2.874.859.380.974.8
+ +on the Cityscapes source domain and the ACDC target domain. Compared with the VFM-based REIN and FADA, the proposed DepthForge shows mIoU improvement of $4.8\%$ & $1.9\%$ , $4.5\%$ & $3.0\%$ , $1.7\%$ & $1.0\%$ , and $2.9\%$ & $0.4\%$ for the Snow, Night, Fog, and Rain scenes, respectively. The feature space of the REIN and the proposed method is visualized in the first volume of Fig. 6. Moreover, our proposed method demonstrates superior performance compared to ResNet-based and Transformer-based approaches. These experimental results underscore the strong and robust generalization ability of DepthForge in extreme scenes with + +Table 5. Performance comparison of the different configurations. Top three results are highlighted as best, second, and third, respectively. (%) + +
BackboneFine-tune +MethodTrainable +Params*mIoUAvg.
CitysBDDMap
CLIP [36] +(ViT-Large)Full304.15M51.347.654.351.1
Freeze0.00M53.748.755.052.4
REIN2.99M57.154.760.557.4
FADA11.65M58.755.862.158.9
DepthForge(Ours)2.99M59.757.063.560.0
SAM [25] +(Huge)Full632.18M57.651.761.556.9
Freeze0.00M57.047.158.454.2
REIN4.51M59.652.062.157.9
FADA16.59M61.053.263.460.0
DepthForge(Ours)4.51M63.355.565.161.3
EVA02 [15, 16] +(Large)Full304.24M62.156.264.660.9
Freeze0.00M56.553.658.656.2
REIN2.99M65.360.564.963.6
FADA11.65M66.761.966.164.9
DepthForge(Ours)2.99M68.061.767.565.7
DINOV2 [30] +(Large)Full304.20M63.757.464.261.7
Freeze0.00M63.356.163.961.1
REIN2.99M66.460.466.164.3
FADA11.65M68.262.068.166.1
DepthForge(Ours)2.99M69.062.869.267.0
+ +Table 6. Performance comparison of the proposed DepthForge across multiple VFs as backbones under the $GTA5 \rightarrow Citys$ . + BDD. + Map. generalization setting. Mark * denotes trainable parameters in backbones. Top three results are highlighted as best, second, and third, respectively. (%) + +limited visual cues. Some exemplar segmentation results are compared under the Cityscapes $\rightarrow$ ACDC are provided in Fig. 4. Our DepthForge shows better pixel-wise predictions than previous methods, demonstrating the effectiveness and significance of depth awareness in our design. + +Cityscapes $\rightarrow$ BDD, Map, GTA5. Tab. 2 shows that the proposed DepthForge achieves the state-of-the-art performance, outperforming the REIN and FADA by $2.65\%$ & $1.07\%$ , $1.9\%$ & $0.07\%$ , and $3.12\%$ & $1.13\%$ on the BDD100k, Mapillary, and GTA unseen target domains, respectively. The proposed method achieves an improvement of approximately $10\%$ compared to ResNet- and Transformer-based approaches. + +
BackboneFine-tune MethodTrainable Params*mIoUAvg.
CitysBDDMap
EVA02 (Large) [15, 16]Full304.24M62.156.264.660.9
+AdvStyle [58]304.24M63.156.464.061.2
+PASTA [5]304.24M61.857.163.660.8
+GTR-LTR [33]304.24M59.857.463.260.1
Freeze0.00M56.553.658.656.2
+AdvStyle [58]0.00M51.451.656.553.2
+PASTA [5]0.00M57.852.358.556.2
+GTR-LTR [33]0.00M52.552.857.154.1
+LoRA [20]1.18M55.552.758.355.5
+AdaptFormer [7]3.17M63.759.964.262.6
+VPT [23]3.69M62.257.762.560.8
+REIN [44]2.99M65.360.564.963.6
+FADA [2]11.65M66.761.966.164.9
+DepthForge(Ours)2.99M68.061.767.565.7
DINOv2 (Large) [30]Full304.20M63.757.464.261.7
+AdvStyle [58]304.20M60.858.062.560.4
+PASTA [5]304.20M62.557.264.761.5
+GTR-LTR [33]304.20M62.757.464.561.6
Freeze0.00M63.356.163.961.1
+AdvStyle [58]0.00M61.555.163.960.1
+PASTA [5]0.00M62.157.264.561.3
+GTR-LTR [33]0.00M60.257.762.260.0
+LoRA [20]0.79M65.258.364.662.7
+AdaptFormer [7]3.17M64.959.064.262.7
+VPT [23]3.69M65.259.465.563.3
+REIN [44]2.99M66.460.466.164.3
+FADA [2]11.65M68.262.068.166.1
+DepthForge(Ours)2.99M69.062.869.267.0
+ +Table 7. Performance comparison of the proposed DepthForge against other DGSS methods under the $GTA5 \rightarrow Citys$ . + BDD. + Map. generalization setting. Mark * denotes trainable parameters in backbones. Top three results are highlighted as best, second, and third, respectively. (%) + +$\mathrm{GTA5} \rightarrow \mathrm{Citys}$ , Map, BDD. As shown in Tab. 3, the proposed DepthForge achieves mIoU improvements of $2.64\%$ , $2.42\%$ , and $3.12\%$ on the Cityscapes, BDD100k, and Mapillary unseen target domains, respectively. The feature space of REIN and our DepthForge is visualized in Fig. 6, demonstrating the effectiveness of our depth-aware learnable tokens in the visual-spatial attention learning. In addition, we outperform the latest FADA by approximately $1\%$ across all three target domains. Visual examples for qualitative comparison are given in Fig. 5. + +# 5.4. Ablation Studies + +Depth awareness and attention optimization are significant for depth-aware learnable tokens. We find that simply incorporating a spatial cue bias into the original learnable tokens $(+f_{i}^{d})$ makes almost no difference. However, when we transform the original learnable tokens into depth-aware learnable tokens using the DepthForge framework, we achieve an approximate $2\%$ improvement in the target domains ACDC and Mapillary, as shown in Tab. 4. The depth-aware learnable tokens compensate for the shortcomings of visual cues and establish more robust visual- + +spatial attention. Compared to different configurations, our design allows gradients to be backpropagated through the tokens alone, enables each modality to contribute independently based on its own feature representations, preventing dominance from one during optimization, as shown in Tab. 5 and Fig. 2. This conclusion is supported by qualitative visualizations in Fig. 7. + +Depth refinement decoder facilitates the perception of both visual and depth information. As shown in Tab. 4, we observe that depth refinement decoder integrates multiscale information to enhance the expressive capacity of both the frozen VFM features and the enhanced features. With the incorporation of depth information, this optimization effect is further amplified. This conclusion is corroborated by qualitative visualizations in Fig. 7. + +# 5.5. Generalization on Other Setting + +To Different Depth VFMs. We evaluate the performance of the proposed DepthForge across different depth VFMs. Tab. 4 shows the relative spatial relationships derived from Depth Anything v2 outperform the absolute spatial cues provided by Prompt Depth Anything. We believe that what is more crucial in an image is the relative spatial relationship, which establishes connections among pixels and between pixels and the world. In contrast, absolute spatial information introduces significant discrepancies that can act as noise to visual cues. + +To Different VFMs. We evaluate the generalization ability of the proposed DepthForge across different VFMs. Tab. 6 shows the superiority of DepthForge when embedded into these VFMs. + +To Token Fine-Tuning Methods. We further compare the proposed method with some existing fine-tuning methods. As shown in Tab. 7, the proposed DepthForge outperforms these methods in all unseen target domains. + +# 6. Conclusion + +This paper presents the DepthForge framework, a novel approach for robust domain generalization semantic segmentation. We provide a detailed analysis of the challenges and limitations faced by existing methods. To this end, we propose DepthForge to integrate a pretrained depth VFM for fine-tuning. We design depth-aware learnable tokens to provide supplementary spatial cues to the visual features, thereby improving geometric consistency and enhancing the pixel discriminability. Moreover, we incorporate multi-layer enhanced features to effectively capture spatial correlations across various scales. Extensive experiments demonstrate the substantial potential of leveraging multi-modal VFs in the field of DGSS, particularly under extreme conditions where visual cues are limited or missing, validating the effectiveness of our DepthForge in forging VFs for DGSS. + +# Acknowledgements + +This work was supported in part by the Natural Science Foundation of Xiamen, China, under Grant 3502Z202373036; in part by the National Natural Science Foundation of China under Grant 42371457; in part by the Key Project of Natural Science Foundation of Fujian Province, China, under Grant 2022J02045. + +# References + +[1] Woo-Jin Ahn, Geun-Yeong Yang, Hyun-Duck Choi, and Myo-Taeg Lim. Style blind domain generalized semantic segmentation via covariance alignment and semantic consistence contrastive learning. In CVPR, pages 3616-3626, 2024. 4, 6 +[2] Qi Bi, Jingjun Yi, Hao Zheng, Haolan Zhan, Yawen Huang, Wei Ji, Yuexiang Li, and Yefeng Zheng. Learning frequency-adapted vision foundation model for domain generalized semantic segmentation. NeurIPS, 37:94047-94072, 2024. 1, 2, 4, 6, 8 +[3] Qi Bi, Shaodi You, and Theo Gevers. Learning content-enhanced mask transformer for domain generalized urban-scene segmentation. In AAAI, pages 819-827, 2024. 4, 6 +[4] Manh-Ha Bui, Toan Tran, Anh Tran, and Dinh Phung. Exploiting domain-specific features to enhance domain generalization. In NeurIPS, pages 21189-21201. Curran Associates, Inc., 2021. 2 +[5] Prithvijit Chattopadhyay, Kartik Sarangmath, Vivek Vijaykumar, and Judy Hoffman. Pasta: Proportional amplitude spectrum training augmentation for syn-to-real domain generalization. In ICCV, pages 19288-19300, 2023. 8 +[6] Geng Chen, Qingyue Wang, Bo Dong, Ruitao Ma, Nian Liu, Huazhu Fu, and Yong Xia. Em-trans: Edge-aware multimodal transformer for rgb-d salient object detection. IEEE Transactions on Neural Networks and Learning Systems, 36 (2):3175-3188, 2024. 3 +[7] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. NeurIPS, 35:16664-16678, 2022. 8 +[8] Siyu Chen, Ting Han, Changshe Zhang, Weiquan Liu, Jinhe Su, Zongyue Wang, and Guorong Cai. Depth matters: Exploring deep interactions of rgb-d for semantic segmentation in traffic scenes. arXiv preprint arXiv:2409.07995, 2024. 3 +[9] Siyu Chen, Ting Han, Changshe Zhang, Jinhe Su, Ruisheng Wang, Yiping Chen, Zongyue Wang, and Guorong Cai. Hspformer: Hierarchical spatial perception transformer for semantic segmentation. IEEE Transactions on Intelligent Transportation Systems, 2025. 3 +[10] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In CVPR, pages 11580-11590, 2021. 4, 6 +[11] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and + +benchmark. https://github.com/open-mmlab/mmsegmentation,2020.6 +[12] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, pages 3213-3223, 2016. 6 +[13] Jian Ding, Nan Xue, Gui-Song Xia, Bernt Schiele, and Dengxin Dai. Hgformer: Hierarchical grouping transformer for domain generalized semantic segmentation. In CVPR, pages 15413-15423, 2023. 4 +[14] Siqi Du, Weixi Wang, Renzhong Guo, Ruisheng Wang, and Shengjun Tang. Asymformer: Asymmetrical cross-modal representation learning for mobile platform real-time rgb-d semantic segmentation. In CVPRW, pages 7608-7615, 2024. 3 +[15] Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva-02: A visual representation for neon genesis. arXiv preprint arXiv:2303.11331, 2023. 1, 2, 7, 8 +[16] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In CVPR, pages 19358-19369, 2023. 7, 8 +[17] Ting Han, Siyu Chen, Chuanmu Li, Zongyue Wang, Jinhe Su, Min Huang, and Guorong Cai. Epurate-net: Efficient progressive uncertainty refinement analysis for traffic environment urban road detection. IEEE Transactions on Intelligent Transportation Systems, 2024. 3 +[18] Sivan Harary, Eli Schwartz, Assaf Arbelle, Peter Staar, Shady Abu-Hussein, Elad Amrani, Roei Herzig, Amit Alfassy, Raja Giryes, Hilde Kuehne, Dina Katabi, Kate Saenko, Rogerio Feris, and Leonid Karlinsky. Unsupervised domain generalization by learning a bridge across domains. In CVPR, pages 5270-5280, 2022. 2 +[19] Zongtao He, Liuyi Wang, Ronghao Dang, Shu Li, Qingqing Yan, Chengju Liu, and Qijun Chen. Learning depth representation from rgb-d videos by time-aware contrastive pretraining. IEEE TCSVT, 34(6):4143-4158, 2024. 3 +[20] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 8 +[21] Lei Huang, Yi Zhou, Fan Zhu, Li Liu, and Ling Shao. Iterative normalization: Beyond standardization towards efficient whitening. In CVPR, pages 4874-4883, 2019. 4 +[22] Wei Huang, Chang Chen, Yong Li, Jiacheng Li, Cheng Li, Fenglong Song, Youliang Yan, and Zhiwei Xiong. Style projected clustering for domain generalized semantic segmentation. In CVPR, pages 3061-3071, 2023. 6 +[23] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, pages 709-727. Springer, 2022. 8 +[24] Jin Kim, Jiyoung Lee, Jungin Park, Dongbo Min, and Kwanghoon Sohn. Pin the memory: Learning to generalize semantic segmentation. In CVPR, 2022. 1 + +[25] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 1, 7 +[26] Suhyeon Lee, Hongje Seong, Seongwon Lee, and Euntai Kim. Wildnet: Learning domain generalized semantic segmentation from the wild. In CVPR, pages 9936-9946, 2022. 4, 6 +[27] Zhiyu Liu, Munawar Hayat, Hong Yang, Duo Peng, and Yinjie Lei. Deep hypersphere feature regularization for weakly supervised rgb-d salient object detection. IEEE TIP, 32: 5423-5437, 2023. 3 +[28] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In ICCV, pages 4990-4999, 2017. 6 +[29] Joshua Niemeijer, Manuel Schwonberg, Jan-Aike Termöhlen, Nico M Schmidt, and Tim Fingscheidt. Generalization by adaptation: Diffusion-based domain extension for domain-generalized semantic segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2830-2840, 2024. 6 +[30] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 7, 8 +[31] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In ECCV, pages 464-479, 2018. 4 +[32] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In CVPR, pages 1863-1871, 2019. 4 +[33] Duo Peng, Yinjie Lei, Lingqiao Liu, Pingping Zhang, and Jun Liu. Global and local texture randomization for synthetic-to-real semantic segmentation. IEEE TIP, 30: 6594-6608, 2021. 4, 6, 8 +[34] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In CVPR, pages 2584-2595, 2022. 1 +[35] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In CVPR, pages 2594-2605, 2022. 4, 6 +[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 1, 2, 7 +[37] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In ECCV, pages 102-118. Springer, 2016. 6 +[38] Giulia Rizzoli, Donald Shenaj, and Pietro Zanuttigh. Source-free domain adaptation for rgb-d semantic segmentation with vision transformers. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops, pages 615-624, 2024. 3 + +[39] Giulia Rizzoli, Donald Shenaj, and Pietro Zanuttigh. Source-free domain adaptation for rgb-d semantic segmentation with vision transformers. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 615-624, 2024. 3 +[40] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In ICCV, pages 10745-10755, 2021. 6 +[41] Yang Shu, Zhangjie Cao, Chenyu Wang, Jianmin Wang, and Mingsheng Long. Open domain generalization with domain-augmented meta-learning. In CVPR, pages 9619-9628, 2021. 2 +[42] Zhiqiang Tang, Yunhe Gao, Yi Zhu, Zhi Zhang, Mu Li, and Dimitris Metaxas. Crossnorm and selfnorm for generalization under distribution shifts. In ICCV, 2021. 1 +[43] Jie Wang, Kechen Song, Yanqi Bao, Liming Huang, and Yunhui Yan. Cgfnet: Cross-guided fusion network for rgb-t salient object detection. IEEE TCSVT, 32(5):2949-2961, 2022. 3 +[44] Zhixiang Wei, Lin Chen, Yi Jin, Xiaoxiao Ma, Tianle Liu, Pengyang Ling, Ben Wang, Huaian Chen, and Jinjin Zheng. Stronger fewer & superior: Harnessing vision foundation models for domain generalized semantic segmentation. In CVPR, pages 28619-28630, 2024. 1, 2, 4, 6, 7, 8 +[45] Binhui Xie, Shuang Li, Mingjia Li, Chi Harold Liu, Gao Huang, and Guoren Wang. Sepico: Semantic-guided pixel contrast for domain adaptive semantic segmentation. IEEE TPAMI, 45(7):9004-9021, 2023. 1 +[46] Qi Xu, Liang Yao, Zhengkai Jiang, Guannan Jiang, Wenqing Chu, Wenhui Han, Wei Zhang, Chengjie Wang, and Ying Tai. Dirl: Domain-invariant representation learning for generalizable semantic segmentation. In AAAI, pages 2884-2892, 2022. 6 +[47] Qi Xu, Liang Yao, Zhengkai Jiang, Guannan Jiang, Wenqing Chu, Wenhui Han, Wei Zhang, Chengjie Wang, and Ying Tai. Dirl: Domain-invariant representation learning for generalizable semantic segmentation. In AAAI, page 2884-2892, 2022. 1 +[48] Jun Yang, Lizhi Bai, Yaoru Sun, Chunqi Tian, Maoyu Mao, and Guorun Wang. Pixel difference convolutional network for rgb-d semantic segmentation. IEEE TCSVT, 34(3):1481-1492, 2024. 3 +[49] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 2 +[50] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. NeurIPS, 2024. 2 +[51] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, pages 2636-2645, 2020. 6 +[52] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In ICCV, pages 2100-2110, 2019. 4, 6 + +[53] Xiangyu Yue, Zangwei Zheng, Shanghang Zhang, Yang Gao, Trevor Darrell, Kurt Keutzer, and Alberto Sangiovanni Vincentelli. Prototypical cross-domain self-supervised learning for few-shot unsupervised domain adaptation. In CVPR, pages 13834-13844, 2021. 1, 2 +[54] Jiaming Zhang, Huayao Liu, Kailun Yang, Xinxin Hu, Ruiping Liu, and Rainer Stiefelhagen. Cmx: Cross-modal fusion for rgb-x semantic segmentation with transformers. IEEE Transactions on Intelligent Transportation Systems, 24(12): 14679-14694, 2023. 3 +[55] Lu Zhang, Siqi Zhang, Xu Yang, Hong Qiao, and Zhiyong Liu. Unseen object instance segmentation with fully test-time rgb-d embeddings adaptation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4945-4952, 2023. 3 +[56] Xingxuan Zhang, Linjun Zhou, Renzhe Xu, Peng Cui, Zheyan Shen, and Haoxin Liu. Towards unsupervised domain generalization. In CVPR, pages 4900-4910, 2022. 2 +[57] Yuyang Zhao, Zhun Zhong, Na Zhao, Nicu Sebe, and Gim Hee Lee. Style-hallucinated dual consistency learning for domain generalized semantic segmentation. In ECCV, pages 535-552. Springer, 2022. 4, 6 +[58] Zhun Zhong, Yuyang Zhao, Gim Hee Lee, and Nicu Sebe. Adversarial style augmentation for domain generalized urban-scene segmentation. NeurIPS, 35:338-350, 2022. 6, 8 +[59] Wujie Zhou, Yun Zhu, Jingsheng Lei, Jian Wan, and Lu Yu. Ccafnet: Crossflow and cross-scale adaptive fusion network for detecting salient objects in rgb-d images. IEEE TMM, 24: 2192-2204, 2022. 3 +[60] Wujie Zhou, Yuxiang Xiao, Weiqing Yan, and Lu Yu. Cmpffnet: Cross-modal and progressive feature fusion network for rgb-d indoor scene semantic segmentation. IEEE Transactions on Automation Science and Engineering, 21 (4):5523-5533, 2024. 3 \ No newline at end of file diff --git a/2025/Stronger, Steadier & Superior_ Geometric Consistency in Depth VFM Forges Domain Generalized Semantic Segmentation/images.zip b/2025/Stronger, Steadier & Superior_ Geometric Consistency in Depth VFM Forges Domain Generalized Semantic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..20cfe26fd1db4ae96c43857a6787494afd85763f --- /dev/null +++ b/2025/Stronger, Steadier & Superior_ Geometric Consistency in Depth VFM Forges Domain Generalized Semantic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95a383cde4dc110e64462cd946e7756570fe48a4a69bd271611c7a38500865e4 +size 1035636 diff --git a/2025/Stronger, Steadier & Superior_ Geometric Consistency in Depth VFM Forges Domain Generalized Semantic Segmentation/layout.json b/2025/Stronger, Steadier & Superior_ Geometric Consistency in Depth VFM Forges Domain Generalized Semantic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3367ffa66a1e9a965daf6a0dcecce2feac03f1d9 --- /dev/null +++ b/2025/Stronger, Steadier & Superior_ Geometric Consistency in Depth VFM Forges Domain Generalized Semantic Segmentation/layout.json @@ -0,0 +1,9633 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 102, + 543, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 102, + 543, + 140 + ], + "spans": [ + { + "bbox": [ + 67, + 102, + 543, + 140 + ], + "type": "text", + "content": "Stronger, Steadier & Superior: Geometric Consistency in Depth VFM Forges Domain Generalized Semantic Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 72, + 160, + 534, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 160, + 534, + 196 + ], + "spans": [ + { + "bbox": [ + 72, + 160, + 534, + 196 + ], + "type": "text", + "content": "Siyu Chen" + }, + { + "bbox": [ + 72, + 160, + 534, + 196 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 72, + 160, + 534, + 196 + ], + "type": "text", + "content": " Ting Han" + }, + { + "bbox": [ + 72, + 160, + 534, + 196 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 72, + 160, + 534, + 196 + ], + "type": "text", + "content": " Changshe Zhang" + }, + { + "bbox": [ + 72, + 160, + 534, + 196 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 72, + 160, + 534, + 196 + ], + "type": "text", + "content": " Xin Luo" + }, + { + "bbox": [ + 72, + 160, + 534, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 72, + 160, + 534, + 196 + ], + "type": "text", + "content": " Meiliu Wu" + }, + { + "bbox": [ + 72, + 160, + 534, + 196 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 72, + 160, + 534, + 196 + ], + "type": "text", + "content": " Guorong Cai" + }, + { + "bbox": [ + 72, + 160, + 534, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 72, + 160, + 534, + 196 + ], + "type": "text", + "content": " Jinhe Su" + }, + { + "bbox": [ + 72, + 160, + 534, + 196 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 72, + 160, + 534, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 72, + 160, + 534, + 196 + ], + "type": "text", + "content": " Jimei University " + }, + { + "bbox": [ + 72, + 160, + 534, + 196 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 72, + 160, + 534, + 196 + ], + "type": "text", + "content": " Sun Yat-sen University " + }, + { + "bbox": [ + 72, + 160, + 534, + 196 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 72, + 160, + 534, + 196 + ], + "type": "text", + "content": " University of Glasgow " + }, + { + "bbox": [ + 72, + 160, + 534, + 196 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 72, + 160, + 534, + 196 + ], + "type": "text", + "content": " Xidian University" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 69, + 207, + 168, + 312 + ], + "blocks": [ + { + "bbox": [ + 69, + 207, + 168, + 312 + ], + "lines": [ + { + "bbox": [ + 69, + 207, + 168, + 312 + ], + "spans": [ + { + "bbox": [ + 69, + 207, + 168, + 312 + ], + "type": "image", + "image_path": "b00ec6b05b120f399e8146a8fa271322bbfc3d8c0da48e0c6e092787c307ac9a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 65, + 316, + 171, + 327 + ], + "lines": [ + { + "bbox": [ + 65, + 316, + 171, + 327 + ], + "spans": [ + { + "bbox": [ + 65, + 316, + 171, + 327 + ], + "type": "text", + "content": "(a) Stronger Pre-trained Models" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 171, + 207, + 261, + 259 + ], + "blocks": [ + { + "bbox": [ + 171, + 207, + 261, + 259 + ], + "lines": [ + { + "bbox": [ + 171, + 207, + 261, + 259 + ], + "spans": [ + { + "bbox": [ + 171, + 207, + 261, + 259 + ], + "type": "image", + "image_path": "4b8dec8ad21a3f5f4285c0fa8d8824cb72327690bc80e1b3879c5d4fa7760631.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 172, + 261, + 261, + 312 + ], + "blocks": [ + { + "bbox": [ + 172, + 261, + 261, + 312 + ], + "lines": [ + { + "bbox": [ + 172, + 261, + 261, + 312 + ], + "spans": [ + { + "bbox": [ + 172, + 261, + 261, + 312 + ], + "type": "image", + "image_path": "5d1f42bc025ad9277429553c8f6547319aa59db64ed5a11beb31ae6bb689edba.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 330, + 555, + 374 + ], + "lines": [ + { + "bbox": [ + 55, + 330, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 555, + 374 + ], + "type": "text", + "content": "Figure 1. Existing methods fail to fully exploit the potential of VFMs for DGSS when visual cues are limited or absent. To this end, we introduce a novel and robust fine-tuning strategy DepthForge that leverages depth cues from a depth VFM to optimize visual cues, and employs spatial consistency to enhance feature discriminability. The proposed DepthForge achieves (a) stronger performance under extreme conditions, owing to (b) steadier visual-spatial attention, and thus delivers (c) superior generalization ability." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 264, + 207, + 354, + 259 + ], + "blocks": [ + { + "bbox": [ + 264, + 207, + 354, + 259 + ], + "lines": [ + { + "bbox": [ + 264, + 207, + 354, + 259 + ], + "spans": [ + { + "bbox": [ + 264, + 207, + 354, + 259 + ], + "type": "image", + "image_path": "6b7cc3a8b6410ec6a46f86182d827b14a6e730d2133e2b0559445cb0c8f1b357.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 264, + 261, + 354, + 312 + ], + "blocks": [ + { + "bbox": [ + 201, + 316, + 323, + 327 + ], + "lines": [ + { + "bbox": [ + 201, + 316, + 323, + 327 + ], + "spans": [ + { + "bbox": [ + 201, + 316, + 323, + 327 + ], + "type": "text", + "content": "(b) Steadier Visual-Spatial Attention" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 264, + 261, + 354, + 312 + ], + "lines": [ + { + "bbox": [ + 264, + 261, + 354, + 312 + ], + "spans": [ + { + "bbox": [ + 264, + 261, + 354, + 312 + ], + "type": "image", + "image_path": "5bebe333f583b07bafc4a84625e25934c5b7fc473d5194db273edf4d47e57c63.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 361, + 207, + 451, + 259 + ], + "blocks": [ + { + "bbox": [ + 361, + 207, + 451, + 259 + ], + "lines": [ + { + "bbox": [ + 361, + 207, + 451, + 259 + ], + "spans": [ + { + "bbox": [ + 361, + 207, + 451, + 259 + ], + "type": "image", + "image_path": "404c9c4bd65d4196835cabc581088fc71527c3762e1996a9761927d2e4776298.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 361, + 261, + 451, + 312 + ], + "blocks": [ + { + "bbox": [ + 361, + 261, + 451, + 312 + ], + "lines": [ + { + "bbox": [ + 361, + 261, + 451, + 312 + ], + "spans": [ + { + "bbox": [ + 361, + 261, + 451, + 312 + ], + "type": "image", + "image_path": "35721a10669bd2715a1940fb6a0bcbda95f703b66e45d88b96b575086c8151cb.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 316, + 511, + 327 + ], + "lines": [ + { + "bbox": [ + 394, + 316, + 511, + 327 + ], + "spans": [ + { + "bbox": [ + 394, + 316, + 511, + 327 + ], + "type": "text", + "content": "(c) Superior Generalization Ability" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 455, + 207, + 545, + 259 + ], + "blocks": [ + { + "bbox": [ + 455, + 207, + 545, + 259 + ], + "lines": [ + { + "bbox": [ + 455, + 207, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 455, + 207, + 545, + 259 + ], + "type": "image", + "image_path": "e68f05c791d5149a7865b6aab14ec2743329fa1491a4c52e7625a5282c428f79.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 455, + 261, + 545, + 312 + ], + "blocks": [ + { + "bbox": [ + 455, + 261, + 545, + 312 + ], + "lines": [ + { + "bbox": [ + 455, + 261, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 455, + 261, + 545, + 312 + ], + "type": "image", + "image_path": "01e80c623aa943409eb7233995ac28fcab93ae80065a4711786ac081201c27a6.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 151, + 387, + 200, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 387, + 200, + 399 + ], + "spans": [ + { + "bbox": [ + 151, + 387, + 200, + 399 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 54, + 415, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 415, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 415, + 297, + 715 + ], + "type": "text", + "content": "Vision Foundation Models (VFMs) have delivered remarkable performance in Domain Generalized Semantic Segmentation (DGSS). However, recent methods often overlook the fact that visual cues are susceptible, whereas the underlying geometry remains stable, rendering depth information more robust. In this paper, we investigate the potential of integrating depth information with features from VFMs, to improve the geometric consistency within an image and boost the generalization performance of VFMs. We propose a novel fine-tuning DGSS framework, named Depth-Forge, which integrates the visual cues from frozen DINOv2 or EVA02 and depth cues from frozen Depth Anything V2. In each layer of the VFMs, we incorporate depth-aware learnable tokens to continuously decouple domain-invariant visual and spatial information, thereby enhancing depth awareness and attention of the VFMs. Finally, we develop a depth refinement decoder and integrate it into the model architecture to adaptively refine multi-layer VFM features and depth-aware learnable tokens. Extensive experiments are conducted based on various DGSS settings and five different datasets as unseen target domains. The qualitative and quantitative results demonstrate that our method significantly outperforms alternative approaches with stronger performance, steadier visual-spatial attention, and superior generalization ability. In particular," + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 388, + 555, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 424 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 424 + ], + "type": "text", + "content": "DepthForge exhibits outstanding performance under extreme conditions (e.g., night and snow). Code is available at https://github.com/SY-Ch/DepthForge." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 442, + 395, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 442, + 395, + 453 + ], + "spans": [ + { + "bbox": [ + 314, + 442, + 395, + 453 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 462, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 462, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 312, + 462, + 555, + 617 + ], + "type": "text", + "content": "Domain Generalized Semantic Segmentation (DGSS) aims to improve prediction accuracy across multiple unseen domains without requiring access to their data, thereby ensuring robust generalization for practical applications [34, 45, 53]. One common approach involves decomposing the learned features into domain-invariant and domain-specific components [42, 47]. This method aims to achieve robustness against domain variations by isolating content-related factors from those that vary between domains. A second strategy employs meta-learning techniques [24], which focus on training models to generalize effectively across domains by learning more adaptable representations, often leveraging shared structures across different datasets." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 617, + 556, + 714 + ], + "type": "text", + "content": "Recently, VFMs have emerged as a promising approach, primarily owing to their strong generalization capabilities, which are enabled by pre-trained on large-scale datasets [15, 25, 36]. Fine-tuning these models has become increasingly popular due to their cost-effectiveness and superior performance [2, 44]. However, a significant challenge remains: what constitutes robust feature information for domain generalization, and how can such robust informa" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8285" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 156 + ], + "type": "text", + "content": "tion be effectively extracted? Previous methods predominantly utilize VFMs based on RGB images. However, the inherent limitations of RGB prevent those methods from adapting to variable conditions, especially in scenarios such as nighttime, overexposure, snow, or fog, where visual cues are not clearly present. This clearly fails to meet the requirements of domain-generalized semantic segmentation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 159, + 297, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 159, + 297, + 457 + ], + "spans": [ + { + "bbox": [ + 57, + 159, + 297, + 457 + ], + "type": "text", + "content": "In this paper, we introduce DepthForge, a robust fine-tune and optimization strategy for domain generalized semantic segmentation, as shown in Fig. 1. Specifically, we employ two frozen VFMs: Depth Anything V2[49, 50] is dedicated to extract depth information, and DINOv2[30] or EVA02[15] are used to process RGB visual features. We enhance the generalization of the visual features at each layer of the VFMs by integrating the corresponding depth cues. More importantly, we introduce depth-aware learnable tokens that enables the model to learn spatial structure invariance, thereby ensuring consistent performance across different domains. The depth and visual features extracted from frozen VFMs are used to optimize the representation of our learnable tokens, thus producing more robust visual-spatial attention features than those based solely on visual cues and forging geometric consistency, as shown in Fig. 1(b). However, we observe the frozen VFMs only provide static features, they are incapable of refining the feature maps during training. This limitation leads to persistent errors in the learnable tokens, causing the optimization process to drift in incorrect directions that are difficult to rectify. To this end, we propose a depth refinement decoder that dynamically adjusts the multi-scale features, combining prior and dynamic features to establish high-quality paired feature relationships at different depth spaces." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 460, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 662 + ], + "type": "text", + "content": "Our core contribution is to demonstrate that robust depth features and learnable tokens can be effectively applied to improve model performance in Domain Generalized Semantic Segmentation (DGSS). To the best of our knowledge, the proposed DepthForge is the first depth-aware framework specifically designed for DGSS. The effectiveness of our design is verified by experiments on widely used Cityscapes, ACDC, Mapillary, BDD100k, and GTA5 datasets. In addition, the individual components of our design are also verified by extensive experiments. Extensive experiments have demonstrated that our method significantly improves DGSS performance, as shown in Fig. 1(c). Specifically, in the " + }, + { + "bbox": [ + 55, + 460, + 295, + 662 + ], + "type": "inline_equation", + "content": "GTA \\rightarrow" + }, + { + "bbox": [ + 55, + 460, + 295, + 662 + ], + "type": "text", + "content": " Cityscapes + BDD100k + Mapillary setting, we observed approximately " + }, + { + "bbox": [ + 55, + 460, + 295, + 662 + ], + "type": "inline_equation", + "content": "+4\\%" + }, + { + "bbox": [ + 55, + 460, + 295, + 662 + ], + "type": "text", + "content": " significant improvement in mIoU, particularly the improvement in extreme scenes is close to " + }, + { + "bbox": [ + 55, + 460, + 295, + 662 + ], + "type": "inline_equation", + "content": "+5\\%" + }, + { + "bbox": [ + 55, + 460, + 295, + 662 + ], + "type": "text", + "content": ". Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 665, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 700 + ], + "type": "text", + "content": "- We propose a stronger DepthForge learning framework to fine-tune VFMs combined with visual and depth cues for domain-generalized semantic segmentation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "- A depth-awareness learnable tokens are introduced to" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 315, + 69, + 556, + 189 + ], + "blocks": [ + { + "bbox": [ + 315, + 69, + 556, + 189 + ], + "lines": [ + { + "bbox": [ + 315, + 69, + 556, + 189 + ], + "spans": [ + { + "bbox": [ + 315, + 69, + 556, + 189 + ], + "type": "image", + "image_path": "5d2b1bdf4d83ad194b61d9a94ed479f728e9cc9ff9a658e425fb35625df46509.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 191, + 555, + 236 + ], + "lines": [ + { + "bbox": [ + 313, + 191, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 191, + 555, + 236 + ], + "type": "text", + "content": "Figure 2. Different configurations of incorporating depth cues into visual information, where Config.1: adding " + }, + { + "bbox": [ + 313, + 191, + 555, + 236 + ], + "type": "inline_equation", + "content": "f_{i}^{d}" + }, + { + "bbox": [ + 313, + 191, + 555, + 236 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 191, + 555, + 236 + ], + "type": "inline_equation", + "content": "f_{i}^{v}" + }, + { + "bbox": [ + 313, + 191, + 555, + 236 + ], + "type": "text", + "content": ", or Config.2: incorporating " + }, + { + "bbox": [ + 313, + 191, + 555, + 236 + ], + "type": "inline_equation", + "content": "f_{i}^{d}" + }, + { + "bbox": [ + 313, + 191, + 555, + 236 + ], + "type": "text", + "content": " into the learnable token " + }, + { + "bbox": [ + 313, + 191, + 555, + 236 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 313, + 191, + 555, + 236 + ], + "type": "text", + "content": ", and Config.3: our DepthForge, respectively." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 321, + 255, + 553, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 255, + 553, + 301 + ], + "spans": [ + { + "bbox": [ + 321, + 255, + 553, + 301 + ], + "type": "text", + "content": "focus on spatial relationship that acquires steadier and more robust visual-spatial attention. This strategy enables DepthForge to enhance the discriminability and refine features at an instance level within each layer." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 302, + 553, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 302, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 314, + 302, + 553, + 350 + ], + "type": "text", + "content": "- Depth Forge highlights the impressive generalization ability, surpassing existing DGSS methods by a large margin, especially in extreme scenes with limited and absent visual cues." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 371, + 400, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 371, + 400, + 383 + ], + "spans": [ + { + "bbox": [ + 314, + 371, + 400, + 383 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 393, + 545, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 545, + 407 + ], + "type": "text", + "content": "2.1. Domain Generalized Semantic Segmentation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 415, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 715 + ], + "type": "text", + "content": "Domain Generalized Semantic Segmentation (DGSS) aims to enhance the generalization capability of segmentation models to unseen domains. Recent methods primarily focus on reducing feature distribution discrepancies across domains and learning more domain-invariant representations. Adversarial training [53], meta-learning [4, 41], and self-supervised learning [18, 56] have been extensively explored to improve generalization performance. However, these methods often rely on constructing cross-domain contrastive constraints or employing sophisticated training strategies, limiting their generalization improvements and complicating the training process. The emergence of Vision Foundation Models (VFMs), trained via large-scale image-text [36] or image-image contrastive learning [30], has opened a promising avenue for DGSS. VFMs capture universal and rich visual representations beneficial for reducing domain biases. Recent approach [2, 44] leverages pretrained VFMs by employing their frozen features as powerful generalized backbones to extract and refine learnable tokens, leading to significant improvements in model performance. Nevertheless, current DGSS approaches predominantly utilize RGB-only data, leaving the models susceptible to variations in lighting, texture, and spatial arrangements across different domains, thus constraining their full potential for generalization." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8286" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 553, + 296 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 553, + 296 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 553, + 296 + ], + "type": "image", + "image_path": "4221b9074ef45cc30b2139057816e8c20379cf5ff037090ebabac7ba3eca537e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 297, + 555, + 365 + ], + "lines": [ + { + "bbox": [ + 55, + 297, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 555, + 365 + ], + "type": "text", + "content": "Figure 3. DepthForge framework. It incorporates the features from frozen VFM and depth VFM into our Depth-aware Learnable Tokens to generate enhanced features " + }, + { + "bbox": [ + 55, + 297, + 555, + 365 + ], + "type": "inline_equation", + "content": "f_{i+1} = f_i^v + \\text{DepthForge}(f_i^v, f_i^d)" + }, + { + "bbox": [ + 55, + 297, + 555, + 365 + ], + "type": "text", + "content": " for each layer. In our tunable DepthForge, We first design a Depth Awareness module that integrates visual and depth cues into learnable tokens. Next, we propose an Attention Optimization module to strengthen the discriminability of instances across different spatial locations. Finally, multi-scale features are fused in Depth Refinement Decoder to adapt to various scales and spatial connections. Our DepthForge establishes direct connections to different instances in real-world locations, facilitating both feature refinement and improved discriminability." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 378, + 180, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 180, + 391 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 180, + 391 + ], + "type": "text", + "content": "2.2. RGB-D Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 396, + 297, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 396, + 297, + 635 + ], + "spans": [ + { + "bbox": [ + 54, + 396, + 297, + 635 + ], + "type": "text", + "content": "Combining RGB with depth information has shown substantial benefits across various visual tasks [6, 8, 9, 17, 39]. In object detection, RGB-D effectively mitigates limitations inherent to RGB data under complex scenes by providing robust geometric and spatial cues, thereby enhancing localization accuracy and model robustness [43, 59]. In semantic segmentation, depth information assists models in capturing spatial structures and geometric contours of objects, alleviating issues related to imprecise object boundary delineation. Particularly in scenarios with blurred boundaries, insufficient illumination, or texture scarcity, the incorporation of depth information substantially improves segmentation accuracy and stability [14, 48, 54, 60]. Recent studies have further indicated that even in domain generalization tasks, integrating RGB and depth features consistently improves the model's generalization performance on unseen domains [19, 27, 38, 55]. Therefore, exploring the integration of depth information to address the limitations of RGB-only approaches represents a promising and essential research direction within DGSS." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 646, + 133, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 646, + 133, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 646, + 133, + 659 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "To achieve better generalization performance, existing approaches opt to fine-tune VFMs in a parameter-efficient manner. One mainstream idea is to refine and propagate features from frozen VFM layers to subsequent ones. Specif" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 379, + 555, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 555, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 555, + 452 + ], + "type": "text", + "content": "ically, given a pre-trained VFM composed of " + }, + { + "bbox": [ + 313, + 379, + 555, + 452 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 379, + 555, + 452 + ], + "type": "text", + "content": " sequential layers " + }, + { + "bbox": [ + 313, + 379, + 555, + 452 + ], + "type": "inline_equation", + "content": "\\{L_1,L_2,\\dots,L_N\\}" + }, + { + "bbox": [ + 313, + 379, + 555, + 452 + ], + "type": "text", + "content": ", where each layer is associated with a weight matrix " + }, + { + "bbox": [ + 313, + 379, + 555, + 452 + ], + "type": "inline_equation", + "content": "W_{i}\\in \\mathbb{R}^{c\\times c}" + }, + { + "bbox": [ + 313, + 379, + 555, + 452 + ], + "type": "text", + "content": " and the frozen output features of each layer are denoted as " + }, + { + "bbox": [ + 313, + 379, + 555, + 452 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 313, + 379, + 555, + 452 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 379, + 555, + 452 + ], + "type": "inline_equation", + "content": "i = 1,2,\\ldots ,N" + }, + { + "bbox": [ + 313, + 379, + 555, + 452 + ], + "type": "text", + "content": ". If we have a learnable weight matrix as " + }, + { + "bbox": [ + 313, + 379, + 555, + 452 + ], + "type": "inline_equation", + "content": "\\Delta W_{i}\\in \\mathbb{R}^{c\\times c}" + }, + { + "bbox": [ + 313, + 379, + 555, + 452 + ], + "type": "text", + "content": ", the feature propagation from layer " + }, + { + "bbox": [ + 313, + 379, + 555, + 452 + ], + "type": "inline_equation", + "content": "L_{i}" + }, + { + "bbox": [ + 313, + 379, + 555, + 452 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 379, + 555, + 452 + ], + "type": "inline_equation", + "content": "L_{i + 1}" + }, + { + "bbox": [ + 313, + 379, + 555, + 452 + ], + "type": "text", + "content": " is then formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 384, + 461, + 555, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 461, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 384, + 461, + 555, + 475 + ], + "type": "interline_equation", + "content": "f _ {i + 1} = W _ {i} f _ {i} + \\Delta W _ {i} f _ {i}. \\tag {1}", + "image_path": "17bfa1bd7ab170a29f3d4f36ed348e0df10251ac1507192daaa4c46c8caa1d9b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 487, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 555, + 510 + ], + "type": "text", + "content": "Recent REIN transfers the learnable matrix " + }, + { + "bbox": [ + 313, + 487, + 555, + 510 + ], + "type": "inline_equation", + "content": "\\Delta W_{i}" + }, + { + "bbox": [ + 313, + 487, + 555, + 510 + ], + "type": "text", + "content": " into a learnable token " + }, + { + "bbox": [ + 313, + 487, + 555, + 510 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 313, + 487, + 555, + 510 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 373, + 521, + 555, + 535 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 521, + 555, + 535 + ], + "spans": [ + { + "bbox": [ + 373, + 521, + 555, + 535 + ], + "type": "interline_equation", + "content": "f _ {i + 1} = W _ {i} f _ {i} + \\epsilon \\left(T _ {i} \\left(W _ {i} f _ {i}\\right)\\right), \\tag {2}", + "image_path": "e5d83346bd76f6e48a76f8eaf25e4762a9cfb0fb4216501361437f449083c911.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 544, + 554, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 544, + 554, + 616 + ], + "type": "inline_equation", + "content": "\\epsilon (\\cdot)" + }, + { + "bbox": [ + 313, + 544, + 554, + 616 + ], + "type": "text", + "content": " denotes the MLP function, and " + }, + { + "bbox": [ + 313, + 544, + 554, + 616 + ], + "type": "inline_equation", + "content": "T_{i}\\in \\mathbb{R}^{m\\times c}" + }, + { + "bbox": [ + 313, + 544, + 554, + 616 + ], + "type": "text", + "content": " with a tokn length " + }, + { + "bbox": [ + 313, + 544, + 554, + 616 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 544, + 554, + 616 + ], + "type": "text", + "content": ". REIN enables each token to connect better with instances in an image. However, we observe that since the updated and frozen features are limited to remain similar, this correlation is fragile and the capacity for feature refinement is limited." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "This prompted us to explore what kind of information could improve discrimination in instances and refine the feature representations. Drawing inspiration from multimodal learning, we observe that depth features provide a spatial relationship bias for corresponding pixels in visual features, thereby compensating for the loss of crucial visual cues in DGSS. We assume that incorporating a pretrained depth VFM into the pre-trained VFMs will alleviate" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8287" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 282, + 240 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 282, + 240 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 282, + 240 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 282, + 240 + ], + "type": "table", + "html": "
MethodProc. & YearSnowNightFogRain
ResNet based:
IBN [31]ECCV201849.621.263.850.4
Itenorm [21]CVPR201949.923.863.350.1
IW [32]CVPR201947.621.862.452.4
RobustNet [10]CVPR202149.824.364.356.0
WildNet [26]CVPR202228.412.741.234.2
Transformer based:
HGFormer [13]CVPR202368.652.769.972.0
CMFormer [3]AAAI202464.333.777.867.6
VFM based:
REIN [44]CVPR202470.655.979.572.5
FADA [2]NeurIPS202473.557.480.275.0
DepthForge(Ours)-75.460.481.275.4
", + "image_path": "017679844a38334d5884322e17ece2e7a00458b4f1af824afd24b8036fb957ff.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 301, + 295, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 295, + 325 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 295, + 325 + ], + "type": "text", + "content": "attention bias and significantly enhance scene understanding, particularly under extreme conditions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 325, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 325, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 325, + 296, + 422 + ], + "type": "text", + "content": "We use the frozen features " + }, + { + "bbox": [ + 55, + 325, + 296, + 422 + ], + "type": "inline_equation", + "content": "f_{i}^{v}" + }, + { + "bbox": [ + 55, + 325, + 296, + 422 + ], + "type": "text", + "content": " from VFM for visual embedding, and the frozen features " + }, + { + "bbox": [ + 55, + 325, + 296, + 422 + ], + "type": "inline_equation", + "content": "f_{i}^{d}" + }, + { + "bbox": [ + 55, + 325, + 296, + 422 + ], + "type": "text", + "content": " from depth VFM for depth embedding. A straightforward thought might simply add " + }, + { + "bbox": [ + 55, + 325, + 296, + 422 + ], + "type": "inline_equation", + "content": "f_{i}^{d}" + }, + { + "bbox": [ + 55, + 325, + 296, + 422 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 325, + 296, + 422 + ], + "type": "inline_equation", + "content": "f_{i}^{v}" + }, + { + "bbox": [ + 55, + 325, + 296, + 422 + ], + "type": "text", + "content": ", or incorporate " + }, + { + "bbox": [ + 55, + 325, + 296, + 422 + ], + "type": "inline_equation", + "content": "f_{i}^{d}" + }, + { + "bbox": [ + 55, + 325, + 296, + 422 + ], + "type": "text", + "content": " into the leanable token " + }, + { + "bbox": [ + 55, + 325, + 296, + 422 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 55, + 325, + 296, + 422 + ], + "type": "text", + "content": ". However, these methods lack flexibility and introduce significant inter-layer feature differences that prevent convergence, as shown in Fig. 2. In contrast, we propose Depth-Forge to adaptively fuse visual and depth cues, denoted as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 440, + 295, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 440, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 61, + 440, + 295, + 455 + ], + "type": "interline_equation", + "content": "f _ {i + 1} = W _ {i} ^ {v} f _ {i} + \\varepsilon \\left(\\epsilon^ {v} \\left(T _ {i} \\left(W _ {i} ^ {v} f _ {i} ^ {v}\\right)\\right) + \\epsilon^ {d} \\left(T _ {i} \\left(W _ {i} ^ {d} f _ {i} ^ {d}\\right)\\right)\\right), \\tag {3}", + "image_path": "324b623754e3cad45e488df92b18856292c04abe6723689c0030f15c1dd51b99.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 462, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 296, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 462, + 296, + 533 + ], + "type": "inline_equation", + "content": "W_{i}^{v}" + }, + { + "bbox": [ + 55, + 462, + 296, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 462, + 296, + 533 + ], + "type": "inline_equation", + "content": "W_{i}^{d}" + }, + { + "bbox": [ + 55, + 462, + 296, + 533 + ], + "type": "text", + "content": " represent the weight matrix from VFMs and depth VFM, respectively. " + }, + { + "bbox": [ + 55, + 462, + 296, + 533 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 55, + 462, + 296, + 533 + ], + "type": "text", + "content": " denotes the adaptation of the visual and depth signals. This approach allows us to more effectively utilize the powerful capacities of multimodal VFMs to forge robust refined features and instance discriminability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 534, + 296, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 606 + ], + "type": "text", + "content": "More importantly, the depth VFM itself possesses strong generalization capabilities, allowing domain adaptation without fine-tuning. A pre-trained VFM applies the same. Therefore, the overall optimization objective still focuses on the segmentation head " + }, + { + "bbox": [ + 55, + 534, + 296, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 534, + 296, + 606 + ], + "type": "text", + "content": " with parameter " + }, + { + "bbox": [ + 55, + 534, + 296, + 606 + ], + "type": "inline_equation", + "content": "\\theta_{h}" + }, + { + "bbox": [ + 55, + 534, + 296, + 606 + ], + "type": "text", + "content": " and our Depth-Forge with parameter " + }, + { + "bbox": [ + 55, + 534, + 296, + 606 + ], + "type": "inline_equation", + "content": "\\theta_{D}" + }, + { + "bbox": [ + 55, + 534, + 296, + 606 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 82, + 613, + 295, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 613, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 82, + 613, + 295, + 647 + ], + "type": "interline_equation", + "content": "\\arg \\min _ {\\theta_ {D}, \\theta_ {h}} \\sum_ {i = 1} ^ {N} \\mathcal {L o s s} \\left(\\mathcal {H} _ {\\theta_ {i}} \\left(\\mathcal {F} _ {\\Theta^ {v}, \\Theta^ {d}, \\theta_ {D}} \\left(x _ {i}\\right)\\right), y _ {i}\\right), \\tag {4}", + "image_path": "9862f2d101aae372dc6d2d31db4dd5f700379cedfbfd501a95ebe16c6f34e449.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " denote the input image and its corresponding ground truth, respectively, and " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " signifies the total number of images. " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " represent the forward process of our DepthForge strategy applied to the pre-trained VFMs with parameters " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\Theta^v" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\Theta^d" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 554, + 281 + ], + "blocks": [ + { + "bbox": [ + 55, + 242, + 295, + 288 + ], + "lines": [ + { + "bbox": [ + 55, + 242, + 295, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 295, + 288 + ], + "type": "text", + "content": "Table 1. Performance comparison between our DepthForge and existing DGSS methods under Cities. " + }, + { + "bbox": [ + 55, + 242, + 295, + 288 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 242, + 295, + 288 + ], + "type": "text", + "content": " ACDC with Snow, Night, Fog, Rain generalization setting. Top three results are highlighted as best, second, and third, respectively. (%)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 70, + 554, + 281 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 554, + 281 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 554, + 281 + ], + "type": "table", + "html": "
MethodProc. & YearBDD100kMapillaryGTA
ResNet based:
IBN [31]ECCV201848.5657.0445.06
Itenorm [21]CVPR201949.2356.2645.73
IW [32]CVPR201948.4955.8244.87
RobustNet [10]CVPR202150.7358.6445.00
DRPC [52]ICCV201949.8656.3445.62
GTR [33]TIP202150.7557.1645.79
SHADE [57]ECCV202250.9560.0748.61
SAW [35]CVPR202252.9559.8147.28
WildNet [26]CVPR202250.9458.7947.01
BlindNet [1]CVPR202451.8460.1847.97
Transformer based:
HGFormer [13]CVPR202353.4066.9051.30
CMFormer [3]AAAI202459.2771.1058.11
VFM based:
REIN [44]CVPR202463.5474.0362.41
FADA [2]NeurIPS202465.1275.8663.78
DepthForge(Ours)-66.1975.9367.24
", + "image_path": "4df0cae1930e64bc5f0be3f8e9ae4533cc707ea2acafaf0454604357da86158f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 283, + 555, + 331 + ], + "lines": [ + { + "bbox": [ + 314, + 283, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 314, + 283, + 555, + 331 + ], + "type": "text", + "content": "Table 2. Performance comparison between our DepthForge and existing DGSS methods under the Citys. " + }, + { + "bbox": [ + 314, + 283, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 314, + 283, + 555, + 331 + ], + "type": "text", + "content": " BDD. " + }, + { + "bbox": [ + 314, + 283, + 555, + 331 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 314, + 283, + 555, + 331 + ], + "type": "text", + "content": " Map. " + }, + { + "bbox": [ + 314, + 283, + 555, + 331 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 314, + 283, + 555, + 331 + ], + "type": "text", + "content": " GTA5 generalization setting. Top three results are highlighted as best, second, and third, respectively. " + }, + { + "bbox": [ + 314, + 283, + 555, + 331 + ], + "type": "inline_equation", + "content": "(\\%)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 346, + 391, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 346, + 391, + 361 + ], + "spans": [ + { + "bbox": [ + 314, + 346, + 391, + 361 + ], + "type": "text", + "content": "4. DepthForge" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 366, + 555, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 555, + 426 + ], + "type": "text", + "content": "This paper introduces the DepthForge strategy, a novel approach for domain generalized semantic segmentation. As shown in Fig. 3, DepthForge consists of three main parts: (1) Depth Awareness; (2) Attention Optimization; (3) Depth Refinement Decoder. We detail the designs as follows." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 434, + 419, + 446 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 419, + 446 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 419, + 446 + ], + "type": "text", + "content": "4.1. Depth Awareness" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 450, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 450, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 312, + 450, + 555, + 605 + ], + "type": "text", + "content": "Recent DGSS methods have applied pre-trained VFMs to source domain segmentation, with the core idea being to enhance the features between frozen VFM layers to help the model better adapt to target tasks. However, we observe that visual cues exhibit significant variations and even deficiencies across different scenes. Consequently, a more robust cue is required to provide a stronger prior for domain generalization. This motivates us to explore a strategy that focuses on the inherent similarities across different domains. In general, spatial location priors are able to help differentiate instances even when visual cues fail. To this end, we employ an additional depth VFM as an auxiliary network, as shown in Fig. 3." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 555, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 701 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 701 + ], + "type": "text", + "content": "Specifically, given a VFM " + }, + { + "bbox": [ + 313, + 605, + 555, + 701 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 605, + 555, + 701 + ], + "type": "text", + "content": " and a depth VFM " + }, + { + "bbox": [ + 313, + 605, + 555, + 701 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 605, + 555, + 701 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 605, + 555, + 701 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 605, + 555, + 701 + ], + "type": "text", + "content": " layers " + }, + { + "bbox": [ + 313, + 605, + 555, + 701 + ], + "type": "inline_equation", + "content": "\\{L_1,L_2,\\dots ,L_N\\}" + }, + { + "bbox": [ + 313, + 605, + 555, + 701 + ], + "type": "text", + "content": ", the corresponding features are denoted as " + }, + { + "bbox": [ + 313, + 605, + 555, + 701 + ], + "type": "inline_equation", + "content": "\\{f_i^v,f_i^d |f_i\\in \\mathbb{R}^{n\\times c},i\\in [1,N]\\}" + }, + { + "bbox": [ + 313, + 605, + 555, + 701 + ], + "type": "text", + "content": ". Note that the layers " + }, + { + "bbox": [ + 313, + 605, + 555, + 701 + ], + "type": "inline_equation", + "content": "\\{L_1,L_2,\\dots ,L_N\\}" + }, + { + "bbox": [ + 313, + 605, + 555, + 701 + ], + "type": "text", + "content": " are kept frozen. We acquire the visual feature " + }, + { + "bbox": [ + 313, + 605, + 555, + 701 + ], + "type": "inline_equation", + "content": "f_{i}^{v}" + }, + { + "bbox": [ + 313, + 605, + 555, + 701 + ], + "type": "text", + "content": " from the " + }, + { + "bbox": [ + 313, + 605, + 555, + 701 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 605, + 555, + 701 + ], + "type": "text", + "content": "-th layer " + }, + { + "bbox": [ + 313, + 605, + 555, + 701 + ], + "type": "inline_equation", + "content": "L_{i}" + }, + { + "bbox": [ + 313, + 605, + 555, + 701 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 605, + 555, + 701 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 605, + 555, + 701 + ], + "type": "text", + "content": ", and the depth feature " + }, + { + "bbox": [ + 313, + 605, + 555, + 701 + ], + "type": "inline_equation", + "content": "f_{i}^{d}" + }, + { + "bbox": [ + 313, + 605, + 555, + 701 + ], + "type": "text", + "content": " from the corresponding layer of " + }, + { + "bbox": [ + 313, + 605, + 555, + 701 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 605, + 555, + 701 + ], + "type": "text", + "content": ". Our goal is to generate the enhanced feature " + }, + { + "bbox": [ + 313, + 605, + 555, + 701 + ], + "type": "inline_equation", + "content": "\\Delta f_{i} = DepthForge(f_{i}^{v},f_{i}^{d})" + }, + { + "bbox": [ + 313, + 605, + 555, + 701 + ], + "type": "text", + "content": " by using both visual and depth cues." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "content": "We employ a set of learnable tokens " + }, + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "inline_equation", + "content": "T = \\{T_{i} \\in" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8288" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 72, + 541, + 281 + ], + "blocks": [ + { + "bbox": [ + 73, + 72, + 541, + 281 + ], + "lines": [ + { + "bbox": [ + 73, + 72, + 541, + 281 + ], + "spans": [ + { + "bbox": [ + 73, + 72, + 541, + 281 + ], + "type": "image", + "image_path": "005b469f1ef48242c248d16711005b6f0c8e73a849a4d4acb67b085487787479.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 283, + 553, + 305 + ], + "lines": [ + { + "bbox": [ + 55, + 283, + 553, + 305 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 553, + 305 + ], + "type": "text", + "content": "Figure 4. Key segmentation examples of existing DGSS methods and DepthForge under the Cityscapes to ACDC unseen target domains under Night, Snow, Rain, and Fog conditions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 73, + 310, + 544, + 427 + ], + "blocks": [ + { + "bbox": [ + 73, + 310, + 544, + 427 + ], + "lines": [ + { + "bbox": [ + 73, + 310, + 544, + 427 + ], + "spans": [ + { + "bbox": [ + 73, + 310, + 544, + 427 + ], + "type": "image", + "image_path": "fbce90875e1e88a2a3ebf89c75055fd6fe37b0abc305baf7fd944ed0ca66f19a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 429, + 553, + 452 + ], + "lines": [ + { + "bbox": [ + 55, + 429, + 553, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 553, + 452 + ], + "type": "text", + "content": "Figure 5. Key segmentation examples of existing DGSS methods and DepthForge under the GTA5 to Mapillary, BDD100k, and Cityscapes unseen target domains setting." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 463, + 295, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 463, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 54, + 463, + 295, + 546 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{m \\times c} |i \\in [1, N]\\}" + }, + { + "bbox": [ + 54, + 463, + 295, + 546 + ], + "type": "text", + "content": " to adaptively learn scene knowledge. As we consider, relying solely on visual cues can be limiting and prone to failure. Therefore, we utilize visual and depth cues to construct learnable tokens, which we term depth-aware learnable tokens. We integrate " + }, + { + "bbox": [ + 54, + 463, + 295, + 546 + ], + "type": "inline_equation", + "content": "T_i" + }, + { + "bbox": [ + 54, + 463, + 295, + 546 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 54, + 463, + 295, + 546 + ], + "type": "inline_equation", + "content": "f_i^v" + }, + { + "bbox": [ + 54, + 463, + 295, + 546 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 463, + 295, + 546 + ], + "type": "inline_equation", + "content": "f_i^d" + }, + { + "bbox": [ + 54, + 463, + 295, + 546 + ], + "type": "text", + "content": " using an attention-based mechanism, prompting tailored adjustments to bridge the gap between different domains." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 546, + 295, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 681 + ], + "type": "text", + "content": "We use " + }, + { + "bbox": [ + 55, + 546, + 295, + 681 + ], + "type": "inline_equation", + "content": "f_{i}^{v,d}" + }, + { + "bbox": [ + 55, + 546, + 295, + 681 + ], + "type": "text", + "content": " as the query, and " + }, + { + "bbox": [ + 55, + 546, + 295, + 681 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 55, + 546, + 295, + 681 + ], + "type": "text", + "content": " as the key and the value A dot-product operation is applied to generate awareness map " + }, + { + "bbox": [ + 55, + 546, + 295, + 681 + ], + "type": "inline_equation", + "content": "A_{i}^{v,d}" + }, + { + "bbox": [ + 55, + 546, + 295, + 681 + ], + "type": "text", + "content": ", and the Softmax function (S) is employed to align each patch with a unique instance. The " + }, + { + "bbox": [ + 55, + 546, + 295, + 681 + ], + "type": "inline_equation", + "content": "A_{i}^{v,d}" + }, + { + "bbox": [ + 55, + 546, + 295, + 681 + ], + "type": "text", + "content": " provides insights into both the visual cues and depth retrieval, qualitatively evaluating the relationship between various tokens and visual or depth cues, respectively. Subsequently, we incorporate depth awareness bias " + }, + { + "bbox": [ + 55, + 546, + 295, + 681 + ], + "type": "inline_equation", + "content": "A_{i}^{d}" + }, + { + "bbox": [ + 55, + 546, + 295, + 681 + ], + "type": "text", + "content": " into the visual features " + }, + { + "bbox": [ + 55, + 546, + 295, + 681 + ], + "type": "inline_equation", + "content": "A_{i}^{v}" + }, + { + "bbox": [ + 55, + 546, + 295, + 681 + ], + "type": "text", + "content": " to provide additional spatial relationships, thereby establishing a latent connection between pixels and the real world. The process is mathematically formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 85, + 689, + 295, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 689, + 295, + 717 + ], + "spans": [ + { + "bbox": [ + 85, + 689, + 295, + 717 + ], + "type": "interline_equation", + "content": "A _ {i} = S \\left(\\frac {f _ {i} ^ {v} \\times (T ^ {v}) _ {i} ^ {\\top}}{\\sqrt {c}}\\right) + \\lambda S \\left(\\frac {f _ {i} ^ {d} \\times (T ^ {d}) _ {i} ^ {\\top}}{\\sqrt {c}}\\right). \\tag {5}", + "image_path": "f6b2819d8c0b2fabf00e37c5f387455b6a7aa56e9eace783ce58d361f5162762.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 463, + 446, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 446, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 446, + 475 + ], + "type": "text", + "content": "4.2. Attention Optimization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 480, + 555, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 555, + 578 + ], + "type": "text", + "content": "Directly leveraging the awareness representation " + }, + { + "bbox": [ + 313, + 480, + 555, + 578 + ], + "type": "inline_equation", + "content": "A_{i}" + }, + { + "bbox": [ + 313, + 480, + 555, + 578 + ], + "type": "text", + "content": " to generate enhanced features may introduce a significant amount of irrelevant information for downstream tasks. To this end, we employ depth-aware learnable tokens " + }, + { + "bbox": [ + 313, + 480, + 555, + 578 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 313, + 480, + 555, + 578 + ], + "type": "text", + "content": " to learn an alignment weight " + }, + { + "bbox": [ + 313, + 480, + 555, + 578 + ], + "type": "inline_equation", + "content": "W_{T_i}" + }, + { + "bbox": [ + 313, + 480, + 555, + 578 + ], + "type": "text", + "content": " and bias " + }, + { + "bbox": [ + 313, + 480, + 555, + 578 + ], + "type": "inline_equation", + "content": "b_{T_i}" + }, + { + "bbox": [ + 313, + 480, + 555, + 578 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 313, + 480, + 555, + 578 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 313, + 480, + 555, + 578 + ], + "type": "text", + "content": " with weight " + }, + { + "bbox": [ + 313, + 480, + 555, + 578 + ], + "type": "inline_equation", + "content": "W_{T_i}" + }, + { + "bbox": [ + 313, + 480, + 555, + 578 + ], + "type": "text", + "content": " and bias " + }, + { + "bbox": [ + 313, + 480, + 555, + 578 + ], + "type": "inline_equation", + "content": "b_{T_i}" + }, + { + "bbox": [ + 313, + 480, + 555, + 578 + ], + "type": "text", + "content": " is then multiplied with the " + }, + { + "bbox": [ + 313, + 480, + 555, + 578 + ], + "type": "inline_equation", + "content": "A_{i}" + }, + { + "bbox": [ + 313, + 480, + 555, + 578 + ], + "type": "text", + "content": " to obtain the initial enhanced features " + }, + { + "bbox": [ + 313, + 480, + 555, + 578 + ], + "type": "inline_equation", + "content": "\\Delta f_{i}" + }, + { + "bbox": [ + 313, + 480, + 555, + 578 + ], + "type": "text", + "content": " that embed both visual and depth cues, using the equation:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 369, + 584, + 553, + 598 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 584, + 553, + 598 + ], + "spans": [ + { + "bbox": [ + 369, + 584, + 553, + 598 + ], + "type": "interline_equation", + "content": "\\Delta \\hat {f} _ {i} = A _ {i} \\times \\left(T _ {i} \\times W _ {T _ {i}} + b _ {T _ {i}}\\right), \\tag {6}", + "image_path": "c9f3e04cf7febf1c86447b732d4d33cc618be5ee2ebddbc56229c1c1d95fca00.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "where we select the same optimization strategy as REIN, discarding high-weight features when the features are sufficiently precise to avoid unnecessary adjustments and mitigate the risk of inappropriate variations. Finally, we employ two consecutive residual connections " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " to enhance the flexibility of attention feature adjustment and prevent gradient vanishing during the forward pass. The calculation is defined as " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\Delta f_{i} = \\phi \\hat{\\Delta f}_{i}" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": ". Benefiting from these depth-aware and attention optimization adjustments, DepthForge is ca" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8289" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 294, + 202 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 294, + 202 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 294, + 202 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 294, + 202 + ], + "type": "image", + "image_path": "39a093aa7a9e91b204c2935320ba498161e2e34e4f4f84c4276e478e503fc41c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 204, + 295, + 247 + ], + "lines": [ + { + "bbox": [ + 55, + 204, + 295, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 295, + 247 + ], + "type": "text", + "content": "Figure 6. t-SNE visualization compares REIN and our Depth-Forge. The first volume corresponds to " + }, + { + "bbox": [ + 55, + 204, + 295, + 247 + ], + "type": "inline_equation", + "content": "\\text{Citys.} \\rightarrow \\text{ACDC}" + }, + { + "bbox": [ + 55, + 204, + 295, + 247 + ], + "type": "text", + "content": ", and two through four volumes correspond to " + }, + { + "bbox": [ + 55, + 204, + 295, + 247 + ], + "type": "inline_equation", + "content": "\\text{GTA5} \\rightarrow \\text{BDD.} + \\text{Citys.} + \\text{Map}" + }, + { + "bbox": [ + 55, + 204, + 295, + 247 + ], + "type": "text", + "content": ". The best mIoU results are highlighted in red." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 260, + 296, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 260, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 260, + 296, + 297 + ], + "type": "text", + "content": "pable of generating more distinctive, discriminative features for different categories within a single image, while balancing both visual and depth cues." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 303, + 205, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 205, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 205, + 316 + ], + "type": "text", + "content": "4.3. Depth Refinement Decoder" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 320, + 296, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 320, + 296, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 320, + 296, + 368 + ], + "type": "text", + "content": "The depth-aware features at different layers are inconsistent. Therefore, how to design a robust decoder becomes a core challenge that effectively integrate these enhanced features across various scales to accomplish downstream tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 369, + 296, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 296, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 296, + 450 + ], + "type": "text", + "content": "Previous methods rely solely on the output of the final layer " + }, + { + "bbox": [ + 55, + 369, + 296, + 450 + ], + "type": "inline_equation", + "content": "f_{N}" + }, + { + "bbox": [ + 55, + 369, + 296, + 450 + ], + "type": "text", + "content": " for semantic segmentation predictions. In contrast, we fuse the outputs from layers " + }, + { + "bbox": [ + 55, + 369, + 296, + 450 + ], + "type": "inline_equation", + "content": "L_{1} - L_{N}" + }, + { + "bbox": [ + 55, + 369, + 296, + 450 + ], + "type": "text", + "content": " to refine the depth features. Specifically, we process these multi-layer features with an MLP with two different fully-connected layers " + }, + { + "bbox": [ + 55, + 369, + 296, + 450 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 369, + 296, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 369, + 296, + 450 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 55, + 369, + 296, + 450 + ], + "type": "text", + "content": " and then concatenate them after a ReLU activation function:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 125, + 451, + 294, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 451, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 125, + 451, + 294, + 464 + ], + "type": "interline_equation", + "content": "f _ {i} ^ {\\text {o u t}} = \\phi (R e L U (\\varphi f _ {i})), \\tag {7}", + "image_path": "c79097e9d9dbe9d9c0121311f325d34bda7dce51d6f8927d0e10bd8abc87ef7e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 99, + 480, + 294, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 480, + 294, + 495 + ], + "spans": [ + { + "bbox": [ + 99, + 480, + 294, + 495 + ], + "type": "interline_equation", + "content": "F _ {u} = \\gamma \\operatorname {C o n c a t} \\left(f _ {1} ^ {\\text {o u t}}, f _ {2} ^ {\\text {o u t}}, \\dots , f _ {N} ^ {\\text {o u t}}\\right), \\tag {8}", + "image_path": "df1465c62d529e8b5f6c6f6026e5ddf69b7416c21ae4a2ab0855b55b606154e0.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 498, + 296, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 534 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 498, + 296, + 534 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 498, + 296, + 534 + ], + "type": "text", + "content": " is a convolutional layer. Finally, we design several sequential multi-head Transformer layers to generate the final predictions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 544, + 132, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 132, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 132, + 558 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 564, + 230, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 230, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 230, + 576 + ], + "type": "text", + "content": "5.1. Datasets & Evaluation Protocols" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "We evaluate our DepthForge on four real-world datasets (Cityscapes [12], BDD100K [51], Mapillary [28], ACDC [40]) and a synthetic dataset (GTA5 [37]). Cityscapes (Citys.) is an autonomous driving dataset comprising 2,975 training images and 500 validation images, each with a resolution of " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "2048 \\times 1024" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": ". BDD100K (BDD.) and Mapillary (Map.) provide 1,000 and 2,000 validation images, with resolutions of " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "1280 \\times 720" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "1902 \\times 1080" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": ", respectively. ACDC offers 406 validation images captured under extreme conditions: namely at Night, Snow, Fog, and Rain, each with a resolution of " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "1902 \\times 1080" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": ". GTA5 is a" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 321, + 71, + 549, + 259 + ], + "blocks": [ + { + "bbox": [ + 321, + 71, + 549, + 259 + ], + "lines": [ + { + "bbox": [ + 321, + 71, + 549, + 259 + ], + "spans": [ + { + "bbox": [ + 321, + 71, + 549, + 259 + ], + "type": "table", + "html": "
MethodProc. & YearCityscapesBDD100kMapillary
ResNet based:
DRPC [52]ICCV201937.4232.1434.12
GTR [33]TIP202137.5333.7534.52
RobustNet [10]CVPR202136.5835.2040.33
DIRL [46]AAAI202241.0439.1541.60
SHADE [57]ECCV202244.6539.2843.34
WildNet [26]CVPR202344.6238.4246.09
SAW [35]CVPR202239.7537.3441.86
AdvStyle [58]NeurIPS202239.6235.5437.00
SPC [22]CVPR202344.1040.4645.51
BlindNet [1]CVPR202445.7241.3247.08
Transformer based:
CMFormer [3]AAAI202455.3149.9160.09
VFM based:
DIDEX [29]WACV202462.0054.3063.00
REIN [44]CVPR202466.4060.4066.10
FADA [2]NeurIPS202468.2361.9468.09
DepthForge(Ours)-69.0462.8269.22
", + "image_path": "fea490b5d66347f141408c20a2898c34233dca6dd9c8a1b2165c2e8f59d79662.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 262, + 553, + 309 + ], + "lines": [ + { + "bbox": [ + 313, + 262, + 553, + 309 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 553, + 309 + ], + "type": "text", + "content": "Table 3. Performance comparison between our DepthForge and existing DGSS methods under the " + }, + { + "bbox": [ + 313, + 262, + 553, + 309 + ], + "type": "inline_equation", + "content": "GTA5 \\rightarrow Citys" + }, + { + "bbox": [ + 313, + 262, + 553, + 309 + ], + "type": "text", + "content": ". + BDD. + Map. generalization setting. Top three results are highlighted as best, second, and third, respectively. (%)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 323, + 553, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 553, + 346 + ], + "type": "text", + "content": "synthetic dataset that provides 24,966 annotated images obtained from the game." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 347, + 554, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 554, + 454 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 554, + 454 + ], + "type": "text", + "content": "Following the evaluation protocol of the existing DGSS methods, a certain dataset is chosen as the source domain for training, and the other three datasets serve as unseen target domains for validation. The three common evaluation settings are as follows: (1) Cityscapes " + }, + { + "bbox": [ + 313, + 347, + 554, + 454 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 347, + 554, + 454 + ], + "type": "text", + "content": " ACDC (Night, Snow, Fog, Rain); (2) GTA5 " + }, + { + "bbox": [ + 313, + 347, + 554, + 454 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 347, + 554, + 454 + ], + "type": "text", + "content": " Cityscapes, BDD100K, Mapillary; and (3) Cityscapes " + }, + { + "bbox": [ + 313, + 347, + 554, + 454 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 347, + 554, + 454 + ], + "type": "text", + "content": " BDD100K, Mapillary, GTA5. The evaluation metric is the mean Intersection-over-Union (mIoU)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 464, + 533, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 533, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 533, + 477 + ], + "type": "text", + "content": "5.2. Deployment Details & Parameter Settings" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 482, + 554, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 554, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 554, + 589 + ], + "type": "text", + "content": "Our implementation is built on the MMSegmentation [11] codebase. For optimization, we employ AdamW optimizer with an initial learning rate of 1e-4, a weight decay of 0.05, an epsilon of 1e-8, and beta values of (0.9, 0.999). A OneCycleLR scheduler is applied over 50,000 total steps with a maximum learning rate of 1e-4, featuring a warmup phase covering " + }, + { + "bbox": [ + 313, + 482, + 554, + 589 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 482, + 554, + 589 + ], + "type": "text", + "content": " of the training iterations, and using cosine annealing, with both the div_factor and final_div_factor set to 10." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 590, + 553, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 553, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 553, + 638 + ], + "type": "text", + "content": "Data augmentation includes multi-scale resizing, random cropping (with a fixed crop size and category ratio constraints), random flipping, and photometric distortion. All experiments are conducted on an NVIDIA RTX 3090 GPU." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 647, + 542, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 542, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 542, + 659 + ], + "type": "text", + "content": "5.3. Comparison with State-of-The-Art Methods" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 665, + 553, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 700 + ], + "type": "text", + "content": "We compare our approach with existing DGSS methods, including: (1) ResNet-based methods, (2) Transformer-based methods, and (3) VFM-based methods." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "content": "Cityscapes " + }, + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "content": " ACDC. Tab. 1 presents the performance" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8290" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 71, + 299, + 222 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 299, + 222 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 299, + 222 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 299, + 222 + ], + "type": "image", + "image_path": "cb236a937283ea96528763b2ca03ae69e2fac2235b56c87c7c1c71c50eefdce0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 224, + 555, + 247 + ], + "lines": [ + { + "bbox": [ + 55, + 224, + 555, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 555, + 247 + ], + "type": "text", + "content": "Figure 7. Visualizations of attention maps and affinity maps under differnet configurations, where DRD denotes the Depth Refinement Decoder, DA and AO represent the Depth Awareness and Attention Optimization, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 306, + 72, + 553, + 222 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 553, + 222 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 553, + 222 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 553, + 222 + ], + "type": "image", + "image_path": "6b6efc08664d9163e81415a3b3f43b223a1a74295bafb4ffcdc78d73bfdb058b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 59, + 261, + 296, + 376 + ], + "blocks": [ + { + "bbox": [ + 59, + 261, + 296, + 376 + ], + "lines": [ + { + "bbox": [ + 59, + 261, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 59, + 261, + 296, + 376 + ], + "type": "table", + "html": "
BackboneConfigurationsSnowNightFogRainBDDMap
EVA02 (Large) [15, 16]REIN [44]66.052.176.069.560.070.0
+ concat fi d66.352.276.269.860.270.2
+ Prompt Depth Anything65.852.475.569.360.469.9
+ Depth Anything V267.253.077.070.260.970.8
+ DRD67.753.577.670.761.271.0
+ DA + AO67.953.777.870.961.571.2
DepthForge69.554.578.572.062.073.0
DINOv2 (Large) [30]REIN [44]70.655.979.572.563.574.0
+ concat fi d70.856.179.472.863.674.2
+ Prompt Depth Anything70.455.579.772.063.873.9
+ Depth Anything V271.556.780.573.364.474.5
+ DRD72.057.280.774.064.974.8
+ DA + AO72.257.480.974.265.175.0
DepthForge75.460.481.275.466.276.0
", + "image_path": "b7ab8cb883c8c521562b4f05a5833e24b649a11cf02ecda82b8d6cd81a7fcbea.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 58, + 450, + 294, + 532 + ], + "blocks": [ + { + "bbox": [ + 55, + 378, + 296, + 447 + ], + "lines": [ + { + "bbox": [ + 55, + 378, + 296, + 447 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 296, + 447 + ], + "type": "text", + "content": "Table 4. Ablation studies on component configurations of the proposed DepthForge under Citys. " + }, + { + "bbox": [ + 55, + 378, + 296, + 447 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 378, + 296, + 447 + ], + "type": "text", + "content": " ACDC with Snow, Night, Fog, Rain + BDD. + Map. generalization setting, where DRD denotes the Depth Refinement Decoder, DA and AO represent the Depth Awareness and Attention Optimization. Top three results are highlighted as best, second, and third, respectively. (%)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 450, + 294, + 532 + ], + "lines": [ + { + "bbox": [ + 58, + 450, + 294, + 532 + ], + "spans": [ + { + "bbox": [ + 58, + 450, + 294, + 532 + ], + "type": "table", + "html": "
No.DesignAvg. Gain (%)RainNightSnowRog
1No Depth-70.655.979.572.5
2Concatenation-1.169.855.078.371.2
3Depth Token1.672.357.880.774.0
4Config10.270.856.179.472.8
5Config20.771.556.880.173.0
6Config3, Ours3.575.460.481.275.4
7w/o Scale Factor2.874.859.380.974.8
", + "image_path": "0736485157d26a4de30a5fe070a93e239be869470f8a95767834e9502542057c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": "on the Cityscapes source domain and the ACDC target domain. Compared with the VFM-based REIN and FADA, the proposed DepthForge shows mIoU improvement of " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "4.8\\%" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " & " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " & " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "3.0\\%" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " & " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " & " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " for the Snow, Night, Fog, and Rain scenes, respectively. The feature space of the REIN and the proposed method is visualized in the first volume of Fig. 6. Moreover, our proposed method demonstrates superior performance compared to ResNet-based and Transformer-based approaches. These experimental results underscore the strong and robust generalization ability of DepthForge in extreme scenes with" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 318, + 261, + 553, + 484 + ], + "blocks": [ + { + "bbox": [ + 55, + 533, + 295, + 567 + ], + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 567 + ], + "type": "text", + "content": "Table 5. Performance comparison of the different configurations. Top three results are highlighted as best, second, and third, respectively. (%)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 261, + 553, + 484 + ], + "lines": [ + { + "bbox": [ + 318, + 261, + 553, + 484 + ], + "spans": [ + { + "bbox": [ + 318, + 261, + 553, + 484 + ], + "type": "table", + "html": "
BackboneFine-tune \nMethodTrainable \nParams*mIoUAvg.
CitysBDDMap
CLIP [36] \n(ViT-Large)Full304.15M51.347.654.351.1
Freeze0.00M53.748.755.052.4
REIN2.99M57.154.760.557.4
FADA11.65M58.755.862.158.9
DepthForge(Ours)2.99M59.757.063.560.0
SAM [25] \n(Huge)Full632.18M57.651.761.556.9
Freeze0.00M57.047.158.454.2
REIN4.51M59.652.062.157.9
FADA16.59M61.053.263.460.0
DepthForge(Ours)4.51M63.355.565.161.3
EVA02 [15, 16] \n(Large)Full304.24M62.156.264.660.9
Freeze0.00M56.553.658.656.2
REIN2.99M65.360.564.963.6
FADA11.65M66.761.966.164.9
DepthForge(Ours)2.99M68.061.767.565.7
DINOV2 [30] \n(Large)Full304.20M63.757.464.261.7
Freeze0.00M63.356.163.961.1
REIN2.99M66.460.466.164.3
FADA11.65M68.262.068.166.1
DepthForge(Ours)2.99M69.062.869.267.0
", + "image_path": "970e3946c5dccb52e6755ae0a254d4146f4f49931710dc6fb3a66036fa9589e2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 486, + 555, + 545 + ], + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 545 + ], + "type": "text", + "content": "Table 6. Performance comparison of the proposed DepthForge across multiple VFs as backbones under the " + }, + { + "bbox": [ + 313, + 486, + 555, + 545 + ], + "type": "inline_equation", + "content": "GTA5 \\rightarrow Citys" + }, + { + "bbox": [ + 313, + 486, + 555, + 545 + ], + "type": "text", + "content": ". + BDD. + Map. generalization setting. Mark * denotes trainable parameters in backbones. Top three results are highlighted as best, second, and third, respectively. (%)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 557, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 617 + ], + "type": "text", + "content": "limited visual cues. Some exemplar segmentation results are compared under the Cityscapes " + }, + { + "bbox": [ + 313, + 557, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 557, + 555, + 617 + ], + "type": "text", + "content": " ACDC are provided in Fig. 4. Our DepthForge shows better pixel-wise predictions than previous methods, demonstrating the effectiveness and significance of depth awareness in our design." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Cityscapes " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": " BDD, Map, GTA5. Tab. 2 shows that the proposed DepthForge achieves the state-of-the-art performance, outperforming the REIN and FADA by " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "2.65\\%" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": " & " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "1.07\\%" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": " & " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "0.07\\%" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "3.12\\%" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": " & " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "1.13\\%" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": " on the BDD100k, Mapillary, and GTA unseen target domains, respectively. The proposed method achieves an improvement of approximately " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": " compared to ResNet- and Transformer-based approaches." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "text", + "content": "8291" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 70, + 293, + 369 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 293, + 369 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 293, + 369 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 293, + 369 + ], + "type": "table", + "html": "
BackboneFine-tune MethodTrainable Params*mIoUAvg.
CitysBDDMap
EVA02 (Large) [15, 16]Full304.24M62.156.264.660.9
+AdvStyle [58]304.24M63.156.464.061.2
+PASTA [5]304.24M61.857.163.660.8
+GTR-LTR [33]304.24M59.857.463.260.1
Freeze0.00M56.553.658.656.2
+AdvStyle [58]0.00M51.451.656.553.2
+PASTA [5]0.00M57.852.358.556.2
+GTR-LTR [33]0.00M52.552.857.154.1
+LoRA [20]1.18M55.552.758.355.5
+AdaptFormer [7]3.17M63.759.964.262.6
+VPT [23]3.69M62.257.762.560.8
+REIN [44]2.99M65.360.564.963.6
+FADA [2]11.65M66.761.966.164.9
+DepthForge(Ours)2.99M68.061.767.565.7
DINOv2 (Large) [30]Full304.20M63.757.464.261.7
+AdvStyle [58]304.20M60.858.062.560.4
+PASTA [5]304.20M62.557.264.761.5
+GTR-LTR [33]304.20M62.757.464.561.6
Freeze0.00M63.356.163.961.1
+AdvStyle [58]0.00M61.555.163.960.1
+PASTA [5]0.00M62.157.264.561.3
+GTR-LTR [33]0.00M60.257.762.260.0
+LoRA [20]0.79M65.258.364.662.7
+AdaptFormer [7]3.17M64.959.064.262.7
+VPT [23]3.69M65.259.465.563.3
+REIN [44]2.99M66.460.466.164.3
+FADA [2]11.65M68.262.068.166.1
+DepthForge(Ours)2.99M69.062.869.267.0
", + "image_path": "4786824ca12e30bf8f749774b1c2ad29a64d87dfdf73da7ff573aecca9cf1bab.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 369, + 295, + 429 + ], + "lines": [ + { + "bbox": [ + 55, + 369, + 295, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 295, + 429 + ], + "type": "text", + "content": "Table 7. Performance comparison of the proposed DepthForge against other DGSS methods under the " + }, + { + "bbox": [ + 55, + 369, + 295, + 429 + ], + "type": "inline_equation", + "content": "GTA5 \\rightarrow Citys" + }, + { + "bbox": [ + 55, + 369, + 295, + 429 + ], + "type": "text", + "content": ". + BDD. + Map. generalization setting. Mark * denotes trainable parameters in backbones. Top three results are highlighted as best, second, and third, respectively. (%)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 444, + 295, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 295, + 564 + ], + "type": "inline_equation", + "content": "\\mathrm{GTA5} \\rightarrow \\mathrm{Citys}" + }, + { + "bbox": [ + 55, + 444, + 295, + 564 + ], + "type": "text", + "content": ", Map, BDD. As shown in Tab. 3, the proposed DepthForge achieves mIoU improvements of " + }, + { + "bbox": [ + 55, + 444, + 295, + 564 + ], + "type": "inline_equation", + "content": "2.64\\%" + }, + { + "bbox": [ + 55, + 444, + 295, + 564 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 444, + 295, + 564 + ], + "type": "inline_equation", + "content": "2.42\\%" + }, + { + "bbox": [ + 55, + 444, + 295, + 564 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 444, + 295, + 564 + ], + "type": "inline_equation", + "content": "3.12\\%" + }, + { + "bbox": [ + 55, + 444, + 295, + 564 + ], + "type": "text", + "content": " on the Cityscapes, BDD100k, and Mapillary unseen target domains, respectively. The feature space of REIN and our DepthForge is visualized in Fig. 6, demonstrating the effectiveness of our depth-aware learnable tokens in the visual-spatial attention learning. In addition, we outperform the latest FADA by approximately " + }, + { + "bbox": [ + 55, + 444, + 295, + 564 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 55, + 444, + 295, + 564 + ], + "type": "text", + "content": " across all three target domains. Visual examples for qualitative comparison are given in Fig. 5." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 574, + 156, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 156, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 156, + 586 + ], + "type": "text", + "content": "5.4. Ablation Studies" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Depth awareness and attention optimization are significant for depth-aware learnable tokens. We find that simply incorporating a spatial cue bias into the original learnable tokens " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "(+f_{i}^{d})" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": " makes almost no difference. However, when we transform the original learnable tokens into depth-aware learnable tokens using the DepthForge framework, we achieve an approximate " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": " improvement in the target domains ACDC and Mapillary, as shown in Tab. 4. The depth-aware learnable tokens compensate for the shortcomings of visual cues and establish more robust visual-" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "spatial attention. Compared to different configurations, our design allows gradients to be backpropagated through the tokens alone, enables each modality to contribute independently based on its own feature representations, preventing dominance from one during optimization, as shown in Tab. 5 and Fig. 2. This conclusion is supported by qualitative visualizations in Fig. 7." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 157, + 553, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 553, + 252 + ], + "type": "text", + "content": "Depth refinement decoder facilitates the perception of both visual and depth information. As shown in Tab. 4, we observe that depth refinement decoder integrates multiscale information to enhance the expressive capacity of both the frozen VFM features and the enhanced features. With the incorporation of depth information, this optimization effect is further amplified. This conclusion is corroborated by qualitative visualizations in Fig. 7." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 258, + 487, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 487, + 272 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 487, + 272 + ], + "type": "text", + "content": "5.5. Generalization on Other Setting" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 276, + 553, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 553, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 553, + 395 + ], + "type": "text", + "content": "To Different Depth VFMs. We evaluate the performance of the proposed DepthForge across different depth VFMs. Tab. 4 shows the relative spatial relationships derived from Depth Anything v2 outperform the absolute spatial cues provided by Prompt Depth Anything. We believe that what is more crucial in an image is the relative spatial relationship, which establishes connections among pixels and between pixels and the world. In contrast, absolute spatial information introduces significant discrepancies that can act as noise to visual cues." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 396, + 553, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 553, + 443 + ], + "type": "text", + "content": "To Different VFMs. We evaluate the generalization ability of the proposed DepthForge across different VFMs. Tab. 6 shows the superiority of DepthForge when embedded into these VFMs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 444, + 553, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 553, + 491 + ], + "type": "text", + "content": "To Token Fine-Tuning Methods. We further compare the proposed method with some existing fine-tuning methods. As shown in Tab. 7, the proposed DepthForge outperforms these methods in all unseen target domains." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 522, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 553, + 713 + ], + "type": "text", + "content": "This paper presents the DepthForge framework, a novel approach for robust domain generalization semantic segmentation. We provide a detailed analysis of the challenges and limitations faced by existing methods. To this end, we propose DepthForge to integrate a pretrained depth VFM for fine-tuning. We design depth-aware learnable tokens to provide supplementary spatial cues to the visual features, thereby improving geometric consistency and enhancing the pixel discriminability. Moreover, we incorporate multi-layer enhanced features to effectively capture spatial correlations across various scales. Extensive experiments demonstrate the substantial potential of leveraging multi-modal VFs in the field of DGSS, particularly under extreme conditions where visual cues are limited or missing, validating the effectiveness of our DepthForge in forging VFs for DGSS." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8292" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 157, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 157, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 157, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "This work was supported in part by the Natural Science Foundation of Xiamen, China, under Grant 3502Z202373036; in part by the National Natural Science Foundation of China under Grant 42371457; in part by the Key Project of Natural Science Foundation of Fujian Province, China, under Grant 2022J02045." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 195, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 195, + 296, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 195, + 296, + 250 + ], + "spans": [ + { + "bbox": [ + 61, + 195, + 296, + 250 + ], + "type": "text", + "content": "[1] Woo-Jin Ahn, Geun-Yeong Yang, Hyun-Duck Choi, and Myo-Taeg Lim. Style blind domain generalized semantic segmentation via covariance alignment and semantic consistence contrastive learning. In CVPR, pages 3616-3626, 2024. 4, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 251, + 296, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 251, + 296, + 306 + ], + "spans": [ + { + "bbox": [ + 61, + 251, + 296, + 306 + ], + "type": "text", + "content": "[2] Qi Bi, Jingjun Yi, Hao Zheng, Haolan Zhan, Yawen Huang, Wei Ji, Yuexiang Li, and Yefeng Zheng. Learning frequency-adapted vision foundation model for domain generalized semantic segmentation. NeurIPS, 37:94047-94072, 2024. 1, 2, 4, 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 308, + 295, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 295, + 341 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 295, + 341 + ], + "type": "text", + "content": "[3] Qi Bi, Shaodi You, and Theo Gevers. Learning content-enhanced mask transformer for domain generalized urban-scene segmentation. In AAAI, pages 819-827, 2024. 4, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 342, + 295, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 295, + 385 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 295, + 385 + ], + "type": "text", + "content": "[4] Manh-Ha Bui, Toan Tran, Anh Tran, and Dinh Phung. Exploiting domain-specific features to enhance domain generalization. In NeurIPS, pages 21189-21201. Curran Associates, Inc., 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 388, + 296, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 296, + 431 + ], + "type": "text", + "content": "[5] Prithvijit Chattopadhyay, Kartik Sarangmath, Vivek Vijaykumar, and Judy Hoffman. Pasta: Proportional amplitude spectrum training augmentation for syn-to-real domain generalization. In ICCV, pages 19288-19300, 2023. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 433, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 433, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 433, + 296, + 487 + ], + "type": "text", + "content": "[6] Geng Chen, Qingyue Wang, Bo Dong, Ruitao Ma, Nian Liu, Huazhu Fu, and Yong Xia. Em-trans: Edge-aware multimodal transformer for rgb-d salient object detection. IEEE Transactions on Neural Networks and Learning Systems, 36 (2):3175-3188, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 488, + 296, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 488, + 296, + 532 + ], + "spans": [ + { + "bbox": [ + 62, + 488, + 296, + 532 + ], + "type": "text", + "content": "[7] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. NeurIPS, 35:16664-16678, 2022. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 534, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 534, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 534, + 296, + 578 + ], + "type": "text", + "content": "[8] Siyu Chen, Ting Han, Changshe Zhang, Weiquan Liu, Jinhe Su, Zongyue Wang, and Guorong Cai. Depth matters: Exploring deep interactions of rgb-d for semantic segmentation in traffic scenes. arXiv preprint arXiv:2409.07995, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 579, + 296, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 579, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 62, + 579, + 296, + 634 + ], + "type": "text", + "content": "[9] Siyu Chen, Ting Han, Changshe Zhang, Jinhe Su, Ruisheng Wang, Yiping Chen, Zongyue Wang, and Guorong Cai. Hspformer: Hierarchical spatial perception transformer for semantic segmentation. IEEE Transactions on Intelligent Transportation Systems, 2025. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "text", + "content": "[10] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In CVPR, pages 11580-11590, 2021. 4, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "text", + "content": "[11] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "benchmark. https://github.com/open-mmlab/mmsegmentation,2020.6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "text", + "content": "[12] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, pages 3213-3223, 2016. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "type": "text", + "content": "[13] Jian Ding, Nan Xue, Gui-Song Xia, Bernt Schiele, and Dengxin Dai. Hgformer: Hierarchical grouping transformer for domain generalized semantic segmentation. In CVPR, pages 15413-15423, 2023. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 195, + 553, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 248 + ], + "type": "text", + "content": "[14] Siqi Du, Weixi Wang, Renzhong Guo, Ruisheng Wang, and Shengjun Tang. Asymformer: Asymmetrical cross-modal representation learning for mobile platform real-time rgb-d semantic segmentation. In CVPRW, pages 7608-7615, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 250, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 293 + ], + "type": "text", + "content": "[15] Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva-02: A visual representation for neon genesis. arXiv preprint arXiv:2303.11331, 2023. 1, 2, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 294, + 553, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 347 + ], + "type": "text", + "content": "[16] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In CVPR, pages 19358-19369, 2023. 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "type": "text", + "content": "[17] Ting Han, Siyu Chen, Chuanmu Li, Zongyue Wang, Jinhe Su, Min Huang, and Guorong Cai. Epurate-net: Efficient progressive uncertainty refinement analysis for traffic environment urban road detection. IEEE Transactions on Intelligent Transportation Systems, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 404, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 470 + ], + "type": "text", + "content": "[18] Sivan Harary, Eli Schwartz, Assaf Arbelle, Peter Staar, Shady Abu-Hussein, Elad Amrani, Roei Herzig, Amit Alfassy, Raja Giryes, Hilde Kuehne, Dina Katabi, Kate Saenko, Rogerio Feris, and Leonid Karlinsky. Unsupervised domain generalization by learning a bridge across domains. In CVPR, pages 5270-5280, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 472, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 553, + 514 + ], + "type": "text", + "content": "[19] Zongtao He, Liuyi Wang, Ronghao Dang, Shu Li, Qingqing Yan, Chengju Liu, and Qijun Chen. Learning depth representation from rgb-d videos by time-aware contrastive pretraining. IEEE TCSVT, 34(6):4143-4158, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 515, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 558 + ], + "type": "text", + "content": "[20] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 559, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 592 + ], + "type": "text", + "content": "[21] Lei Huang, Yi Zhou, Fan Zhu, Li Liu, and Ling Shao. Iterative normalization: Beyond standardization towards efficient whitening. In CVPR, pages 4874-4883, 2019. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 593, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 635 + ], + "type": "text", + "content": "[22] Wei Huang, Chang Chen, Yong Li, Jiacheng Li, Cheng Li, Fenglong Song, Youliang Yan, and Zhiwei Xiong. Style projected clustering for domain generalized semantic segmentation. In CVPR, pages 3061-3071, 2023. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 636, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 679 + ], + "type": "text", + "content": "[23] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, pages 709-727. Springer, 2022. 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[24] Jin Kim, Jiyoung Lee, Jungin Park, Dongbo Min, and Kwanghoon Sohn. Pin the memory: Learning to generalize semantic segmentation. In CVPR, 2022. 1" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8293" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "type": "text", + "content": "[25] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 1, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[26] Suhyeon Lee, Hongje Seong, Seongwon Lee, and Euntai Kim. Wildnet: Learning domain generalized semantic segmentation from the wild. In CVPR, pages 9936-9946, 2022. 4, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "text", + "content": "[27] Zhiyu Liu, Munawar Hayat, Hong Yang, Duo Peng, and Yinjie Lei. Deep hypersphere feature regularization for weakly supervised rgb-d salient object detection. IEEE TIP, 32: 5423-5437, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "text", + "content": "[28] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In ICCV, pages 4990-4999, 2017. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 319 + ], + "type": "text", + "content": "[29] Joshua Niemeijer, Manuel Schwonberg, Jan-Aike Termöhlen, Nico M Schmidt, and Tim Fingscheidt. Generalization by adaptation: Diffusion-based domain extension for domain-generalized semantic segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2830-2840, 2024. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "type": "text", + "content": "[30] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 409 + ], + "type": "text", + "content": "[31] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In ECCV, pages 464-479, 2018. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 410, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 410, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 294, + 443 + ], + "type": "text", + "content": "[32] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In CVPR, pages 1863-1871, 2019. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "type": "text", + "content": "[33] Duo Peng, Yinjie Lei, Lingqiao Liu, Pingping Zhang, and Jun Liu. Global and local texture randomization for synthetic-to-real semantic segmentation. IEEE TIP, 30: 6594-6608, 2021. 4, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 489, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 294, + 521 + ], + "type": "text", + "content": "[34] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In CVPR, pages 2584-2595, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 523, + 294, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 556 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 556 + ], + "type": "text", + "content": "[35] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In CVPR, pages 2594-2605, 2022. 4, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 557, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 623 + ], + "type": "text", + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 1, 2, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 658 + ], + "type": "text", + "content": "[37] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In ECCV, pages 102-118. Springer, 2016. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[38] Giulia Rizzoli, Donald Shenaj, and Pietro Zanuttigh. Source-free domain adaptation for rgb-d semantic segmentation with vision transformers. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops, pages 615-624, 2024. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 127 + ], + "type": "text", + "content": "[39] Giulia Rizzoli, Donald Shenaj, and Pietro Zanuttigh. Source-free domain adaptation for rgb-d semantic segmentation with vision transformers. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 615-624, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 128, + 554, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 128, + 554, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 128, + 554, + 171 + ], + "type": "text", + "content": "[40] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In ICCV, pages 10745-10755, 2021. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 172, + 554, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 172, + 554, + 215 + ], + "spans": [ + { + "bbox": [ + 316, + 172, + 554, + 215 + ], + "type": "text", + "content": "[41] Yang Shu, Zhangjie Cao, Chenyu Wang, Jianmin Wang, and Mingsheng Long. Open domain generalization with domain-augmented meta-learning. In CVPR, pages 9619-9628, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 217, + 554, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 554, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 554, + 249 + ], + "type": "text", + "content": "[42] Zhiqiang Tang, Yunhe Gao, Yi Zhu, Zhi Zhang, Mu Li, and Dimitris Metaxas. Crossnorm and selfnorm for generalization under distribution shifts. In ICCV, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 250, + 554, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 554, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 554, + 293 + ], + "type": "text", + "content": "[43] Jie Wang, Kechen Song, Yanqi Bao, Liming Huang, and Yunhui Yan. Cgfnet: Cross-guided fusion network for rgb-t salient object detection. IEEE TCSVT, 32(5):2949-2961, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 293, + 554, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 293, + 554, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 293, + 554, + 348 + ], + "type": "text", + "content": "[44] Zhixiang Wei, Lin Chen, Yi Jin, Xiaoxiao Ma, Tianle Liu, Pengyang Ling, Ben Wang, Huaian Chen, and Jinjin Zheng. Stronger fewer & superior: Harnessing vision foundation models for domain generalized semantic segmentation. In CVPR, pages 28619-28630, 2024. 1, 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 349, + 554, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 349, + 554, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 349, + 554, + 392 + ], + "type": "text", + "content": "[45] Binhui Xie, Shuang Li, Mingjia Li, Chi Harold Liu, Gao Huang, and Guoren Wang. Sepico: Semantic-guided pixel contrast for domain adaptive semantic segmentation. IEEE TPAMI, 45(7):9004-9021, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 393, + 554, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 554, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 554, + 448 + ], + "type": "text", + "content": "[46] Qi Xu, Liang Yao, Zhengkai Jiang, Guannan Jiang, Wenqing Chu, Wenhui Han, Wei Zhang, Chengjie Wang, and Ying Tai. Dirl: Domain-invariant representation learning for generalizable semantic segmentation. In AAAI, pages 2884-2892, 2022. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 449, + 554, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 554, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 554, + 502 + ], + "type": "text", + "content": "[47] Qi Xu, Liang Yao, Zhengkai Jiang, Guannan Jiang, Wenqing Chu, Wenhui Han, Wei Zhang, Chengjie Wang, and Ying Tai. Dirl: Domain-invariant representation learning for generalizable semantic segmentation. In AAAI, page 2884-2892, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 503, + 554, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 503, + 554, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 503, + 554, + 547 + ], + "type": "text", + "content": "[48] Jun Yang, Lizhi Bai, Yaoru Sun, Chunqi Tian, Maoyu Mao, and Guorun Wang. Pixel difference convolutional network for rgb-d semantic segmentation. IEEE TCSVT, 34(3):1481-1492, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 548, + 554, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 554, + 581 + ], + "type": "text", + "content": "[49] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 582, + 554, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 554, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 554, + 613 + ], + "type": "text", + "content": "[50] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. NeurIPS, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 614, + 554, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 554, + 658 + ], + "type": "text", + "content": "[51] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, pages 2636-2645, 2020. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "type": "text", + "content": "[52] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In ICCV, pages 2100-2110, 2019. 4, 6" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8294" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 463 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[53] Xiangyu Yue, Zangwei Zheng, Shanghang Zhang, Yang Gao, Trevor Darrell, Kurt Keutzer, and Alberto Sangiovanni Vincentelli. Prototypical cross-domain self-supervised learning for few-shot unsupervised domain adaptation. In CVPR, pages 13834-13844, 2021. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "text", + "content": "[54] Jiaming Zhang, Huayao Liu, Kailun Yang, Xinxin Hu, Ruiping Liu, and Rainer Stiefelhagen. Cmx: Cross-modal fusion for rgb-x semantic segmentation with transformers. IEEE Transactions on Intelligent Transportation Systems, 24(12): 14679-14694, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "type": "text", + "content": "[55] Lu Zhang, Siqi Zhang, Xu Yang, Hong Qiao, and Zhiyong Liu. Unseen object instance segmentation with fully test-time rgb-d embeddings adaptation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4945-4952, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 240, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 240, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 294, + 274 + ], + "type": "text", + "content": "[56] Xingxuan Zhang, Linjun Zhou, Renzhe Xu, Peng Cui, Zheyan Shen, and Haoxin Liu. Towards unsupervised domain generalization. In CVPR, pages 4900-4910, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 275, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 294, + 319 + ], + "type": "text", + "content": "[57] Yuyang Zhao, Zhun Zhong, Na Zhao, Nicu Sebe, and Gim Hee Lee. Style-hallucinated dual consistency learning for domain generalized semantic segmentation. In ECCV, pages 535-552. Springer, 2022. 4, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 294, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 361 + ], + "type": "text", + "content": "[58] Zhun Zhong, Yuyang Zhao, Gim Hee Lee, and Nicu Sebe. Adversarial style augmentation for domain generalized urban-scene segmentation. NeurIPS, 35:338-350, 2022. 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 364, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 364, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 364, + 294, + 407 + ], + "type": "text", + "content": "[59] Wujie Zhou, Yun Zhu, Jingsheng Lei, Jian Wan, and Lu Yu. Ccafnet: Crossflow and cross-scale adaptive fusion network for detecting salient objects in rgb-d images. IEEE TMM, 24: 2192-2204, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 409, + 294, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 294, + 463 + ], + "type": "text", + "content": "[60] Wujie Zhou, Yuxiang Xiao, Weiqing Yan, and Lu Yu. Cmpffnet: Cross-modal and progressive feature fusion network for rgb-d indoor scene semantic segmentation. IEEE Transactions on Automation Science and Engineering, 21 (4):5523-5533, 2024. 3" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8295" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/StruMamba3D_ Exploring Structural Mamba for Self-supervised Point Cloud Representation Learning/9ae9c283-762a-4256-b4de-303cb73e1e2c_content_list.json b/2025/StruMamba3D_ Exploring Structural Mamba for Self-supervised Point Cloud Representation Learning/9ae9c283-762a-4256-b4de-303cb73e1e2c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..53a720169353b29685244fca069e83ba5235313f --- /dev/null +++ b/2025/StruMamba3D_ Exploring Structural Mamba for Self-supervised Point Cloud Representation Learning/9ae9c283-762a-4256-b4de-303cb73e1e2c_content_list.json @@ -0,0 +1,1552 @@ +[ + { + "type": "text", + "text": "StruMamba3D: Exploring Structural Mamba for Self-supervised Point Cloud Representation Learning", + "text_level": 1, + "bbox": [ + 169, + 130, + 826, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chuxin Wang $^{1,2}$ , Yixin Zha $^{1}$ , Wenfei Yang $^{1,2,*}$ , Tianzhu Zhang $^{1,2}$", + "bbox": [ + 241, + 202, + 751, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Science and Technology of China", + "bbox": [ + 305, + 220, + 691, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ National Key Laboratory of Deep Space Exploration, Deep Space Exploration Laboratory", + "bbox": [ + 135, + 239, + 859, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wcx0602, zyxcn} $@$ mail.ustc.edu.cn, {yangwf, tzzhang} $@$ ustc.edu.cn", + "bbox": [ + 217, + 257, + 777, + 275 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/def97985916b92a5e1a231264460e313ce670297dd366eb00c35093729f4b123.jpg", + "image_caption": [ + "Attention Block", + "Quadratic complexity" + ], + "image_footnote": [], + "bbox": [ + 161, + 321, + 308, + 372 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ec38826214ea8c80c9857e9afe51a06461c9b4434540c5512b79339503d86a76.jpg", + "image_caption": [ + "Selective SSM Block", + "Linear complexity" + ], + "image_footnote": [], + "bbox": [ + 320, + 321, + 467, + 371 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f0f68ca372b516ee5050d2041570c1e75aadd8c1e070747bbd3cc6887d5f41cc.jpg", + "image_caption": [ + "Serialized Points", + "Destroy adjacency" + ], + "image_footnote": [], + "bbox": [ + 478, + 320, + 563, + 372 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/87c767c30a776d8c18b83d5e0e01ab1ff193f3fcca4699d474205b54b42b36ec.jpg", + "image_caption": [ + "Longer Input", + "Fail to retain long-sequence memory" + ], + "image_footnote": [], + "bbox": [ + 602, + 320, + 674, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/96ad2b99223270751631ee35700572645c798045e6a66095b184272867ae0b0a.jpg", + "image_caption": [ + "", + "#", + "□ Point Feature", + "□" + ], + "image_footnote": [], + "bbox": [ + 684, + 400, + 777, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1c70f27e6a058199ddbbfc8109c778f96e2b2e9ee2b29abc5fd02d304a0643cb.jpg", + "image_caption": [ + "(a) Transformer-based Paradigm", + "PointMAE", + "PointGPT", + "Mamba3D", + "PointMaml", + "Our", + "#", + "adigm (b) Mamba-based Paradigm", + "Figure 1. (a): Transformer-based paradigm uses the attention block with quadratic complexity to model the dependencies between points. (b): Mamba-based paradigm uses the selective SSM with linear complexity. However, serialized points destroy the adjacency of 3D points, and the pre-trained selection mechanism fails to retain long-sequence memory. (c): Our StruMamba3D paradigm uses the structural SSM to maintain the spatial dependencies among points and the sequence length-adaptive strategy to retain long-sequence memory. (d): Our StruMamba3D significantly outperforms existing Transformer-based and Mamba-based methods." + ], + "image_footnote": [], + "bbox": [ + 138, + 402, + 859, + 501 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 585, + 326, + 601 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, Mamba-based methods have demonstrated impressive performance in point cloud representation learning by leveraging State Space Model (SSM) with the efficient context modeling ability and linear complexity. However, these methods still face two key issues that limit the potential of SSM: Destroying the adjacency of 3D points during SSM processing and failing to retain long-sequence memory as the input length increases in downstream tasks. To address these issues, we propose StruMamba3D, a novel paradigm for self-supervised point cloud representation learning. It enjoys several merits. First, we design spatial states and use them as proxies to preserve spatial dependencies among points. Second, we enhance the SSM with a state-wise update strategy and incorporate a lightweight convolution to facilitate interactions between spatial states for efficient structure modeling. Third, our method reduces the sensitivity of pre-trained Mamba-based models to varying input lengths by introducing a sequence length-adaptive", + "bbox": [ + 88, + 604, + 485, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "strategy. Experimental results across four downstream tasks showcase the superior performance of our method. In addition, our method attains the SOTA $95.1\\%$ accuracy on ModelNet40 and $92.75\\%$ accuracy on the most challenging split of ScanObjectNN without voting strategy.", + "bbox": [ + 511, + 588, + 906, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 680, + 643, + 695 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Point cloud representation learning aims to extract geometric and semantic features from point clouds to support downstream tasks such as classification, segmentation, and detection. As a foundational task in Computer Vision (CV), it is crucial for various real-world applications, including autonomous driving [4, 20], AR [13], and robotics [26, 32]. However, unlike images or text with structured arrangements, point clouds are inherently unordered and sparse, presenting challenges for designing effective models.", + "bbox": [ + 509, + 705, + 906, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To solve this problem, numerous point cloud representation learning methods have been proposed, which can be broadly divided into three categories according to model architectures: Point-based methods [19, 27, 28, 30, 31,", + "bbox": [ + 511, + 840, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 888, + 233, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28546", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "35], Transformer-based methods [12, 22, 25, 42-44], and Mamba-based methods [15, 21, 47]. Point-based methods directly process point clouds by modeling the relationships between individual points, such as through set abstraction [27, 28] and graph convolution [19, 35]. These methods offer a simple and computationally efficient design. However, their limited receptive field hinders capturing global information, constraining complex structure modeling. Transformer-based methods leverage attention mechanisms to capture long-range dependencies, expanding the receptive field. Moreover, inspired by self-supervised learning paradigms in NLP [6] and CV [16], many transformer-based methods [22, 25, 42, 44] utilize large unlabeled point cloud datasets for pre-training, achieving impressive results in downstream tasks. However, these methods have limited scalability due to the quadratic complexity of the attention mechanism, as shown in Fig. 1(a). To overcome this limitation, Mamba-based methods [15, 21, 47] first convert the 3D points into a 1D sequence and then utilize selective State Space Model (SSM) [9] with linear complexity to extract point features. Leveraging the robust information aggregation ability of selective SSM, Mamba-based methods have attained promising performance.", + "bbox": [ + 93, + 90, + 480, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although existing Mamba-based methods [15, 21, 47] have achieved comparable performance to transformer-based methods, two key issues limit the potential of SSM: (1) Distort the adjacency of points. As shown in Fig. 1(b), although methods [21, 47] use serialization strategies to align input points, they cannot ensure that spatially adjacent points remain neighbors in the 1D sequence. Unlike text and images, points inherently lack context information and rely on local structures as base feature units. This spatial distortion hampers Mamba to model fine-grained structure information. (2) Fail to retain long-sequence memory in downstream tasks. Mamba-based methods [15, 21] use Masked Point Modeling (MPM) [25, 42] as a pre-training task, which creates a significant discrepancy in input sequence lengths between pre-training and downstream tasks. Mamba leverages a selection mechanism to inductively learn new information while selectively forgetting outdated knowledge. However, the selection mechanism does not account for variations in input sequence length. Consequently, when Mamba is pre-trained on short sequences, it tends to update states more frequently. As the input sequence length increases in downstream tasks, frequent state updates lead to Mamba struggling to retain long-sequence memory, as shown in Fig. 1(b). The deficiency in long-sequence memory adversely affects the long-range semantics modeling, resulting in suboptimal performance.", + "bbox": [ + 93, + 444, + 480, + 834 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To unleash the potential of Mamba for point cloud representation learning, we explore solutions to address the above two issues. (1) How to maintain spatial dependencies among points? Maintaining spatial relationships", + "bbox": [ + 93, + 840, + 480, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "among 3D points within a 1D input sequence is challenging. Mamba efficiently aggregates contextual features from the 1D input sequence via hidden states. Is it possible to design a structural Mamba that models spatial relationships among points through hidden states? (2) How to retain long-sequence memory for longer inputs? When handling longer input sequences, states need to retain memory for an extended period to effectively capture the contextual information across the entire sequence. If the model adjusts the frequency of state updates when the sequence length changes, it can preserve memory for longer sequences.", + "bbox": [ + 517, + 90, + 903, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve the above goals, we propose a novel StruMamba3D paradigm for self-supervised point cloud representation learning, which consists of two core components: structural SSM block and sequence length-adaptive strategy. In structural SSM block, we leverage states in SSM to model the spatial relationships among points. We begin by assigning positional attributes to the states, using them as proxies to capture the local structure of point clouds. Next, we propose a state-wise update strategy that incorporates the positional relationship between input points and spatial states during the state update and propagation process. Based on spatial states and state-wise update strategy, we can maintain spatial dependencies among points during the SSM process. To address the limitation of standard Mamba ignoring inter-state interactions, we introduce a lightweight convolution module for spatial states. In sequence length-adaptive strategy, the adaptive state update mechanism is first employed to adjust the state update frequency based on sequence length, ensuring that the states retain long-sequence memory as the sequence length grows. Additionally, we introduce a spatial state consistency loss in the pre-training task, which enforces consistency in the updated states across inputs of varying sequence lengths. The spatial state consistency loss enhances the robustness of the Mamba-based pre-trained model to varying input lengths.", + "bbox": [ + 517, + 258, + 903, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To sum up, the main contributions of this work are: (1) We propose the novel StruMamba3D for self-supervised point cloud representation learning, which is the first to model the structure information by latent states in SSM. (2) We design the structural SSM to maintain spatial dependencies among points during state information selection and propagation, while the sequence length-adaptive strategy reduces the sensitivity of pre-trained Mamba-based models to varying input lengths. (3) Experimental results on four downstream tasks showcase the superior performance of our method. Additionally, comprehensive ablation studies validate the effectiveness of our designs.", + "bbox": [ + 517, + 635, + 903, + 814 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 517, + 830, + 651, + 844 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We briefly review methods for point cloud representation learning and self-supervised representation learning.", + "bbox": [ + 517, + 856, + 903, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Point Cloud Representation Learning. To effectively", + "bbox": [ + 517, + 886, + 903, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28547", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "leverage the sparse, unordered and irregular point cloud, various point cloud representation learning methods have been proposed, which can be broadly categorized into three groups: Point-based methods [19, 27, 28, 30, 31, 35], Transformer-based methods [12, 22, 25, 33, 42, 44], and Mamba-based methods [15, 21, 47]. Point-based methods directly process raw point clouds to learn geometric features. For example, PointNet [27] employs MLPs to each point independently, followed by global max pooling to aggregate features. PointNet++ [28] extends PointNet by introducing a hierarchical architecture for multi-scale feature extraction. Subsequent methods [19, 24, 30, 31, 35] explore convolutional techniques for point clouds. DGCNN [35] leverages graph convolution to model structural information among neighboring points, while KPCov [31] introduces a convolution method based on dynamic kernel points to improve feature learning. To capture long-range dependencies, methods [12, 37, 38, 49] have explored transformer architectures for point clouds. PCT [12] applies global attention to points, while Point Transformer [49] introduces local attention to capture local geometry features. Recent variants [37, 38] further refine the architectures to improve performance and efficiency. These supervised methods [12, 30, 31, 47, 49] have demonstrated outstanding performance in specific tasks. However, these methods are inherently limited by the specific data domains on which they are trained, leading to poor generalization to other domains.", + "bbox": [ + 89, + 90, + 483, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-supervised Representation Learning. To solve the above issues and fully utilize a large number of unlabeled data, self-supervised representation learning methods [3, 22, 25, 40, 42, 44, 48] have emerged. Some contrastive learning-based methods [40, 48] train models to learn distinguishing features by differentiating between positive and negative samples. For instance, Point-Contrast [40] leverages feature consistency across different views of the same point cloud. Inspired by the success of BERT [6] and MAE [16], many transformer-based methods [3, 22, 25, 42, 44] adopt MPM task for pre-training, encouraging models to infer masked patches from visible ones. PointBERT [42] predicts discrete tokens for masked points, while PointMAE [25] reconstructs masked point clouds. Additionally, recent studies [29, 45] integrate cross-modal information, such as images and text, to enhance pre-training. However, transformer-based methods suffer from quadratic complexity, posing challenges for handling long sequences and making them unsuitable for resource-constrained devices. Recently, SSMs [5, 9-11, 17, 34] have emerged as a prominent research focus due to their linear complexity and powerful information aggregation ability. In the field of point cloud representation learning, several Mamba-based methods [15, 21, 47] have been proposed. PointMamba [21] employs space-filling curves to serialize point clouds, establishing a simple yet effective base", + "bbox": [ + 89, + 508, + 486, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "line. PCM [47] introduces multiple serialization strategies and a bidirectional scanning mechanism to capture comprehensive point cloud structures, but also introduces larger computational overhead. Mamba3D [15] also leverages a bidirectional scanning mechanism and proposes a local norm pooling operation to enhance local geometric features. However, these methods introduce structural distortions when converting 3D points into 1D sequences, limiting model performance, especially in tasks requiring fine-grained geometry like part segmentation. Moreover, they neglect the impact of varying input lengths between pretraining and downstream tasks. Unlike these methods, we propose StruMamba3D, which leverages states as proxies to maintain spatial dependencies among points. Additionally, we introduce a sequence-length adaptive strategy to ensure that the states retain long-sequence memory as the sequence length grows in downstream tasks.", + "bbox": [ + 511, + 90, + 906, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 359, + 604, + 376 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Below, we review the previous SSMs (Sec. 3.1), followed by an overview of our StruMamba3D (Sec. 3.2). Then, we introduce the details of Structural SSM Block (Sec. 3.3) and Sequence Length-adaptive Strategy (Sec. 3.4).", + "bbox": [ + 511, + 385, + 908, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 513, + 454, + 655, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "State Space Model (SSM) is a continuous system that maps the input $x(t)$ to $y(t)$ through implicit states $h(t)\\in \\mathbb{R}^{M}$ :", + "bbox": [ + 511, + 476, + 906, + 508 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nh ^ {\\prime} (t) = \\mathbf {A} h (t) + \\mathbf {B} x (t), y (t) = \\mathbf {C} h ^ {\\prime} (t), \\qquad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 515, + 906, + 534 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the matrix $\\mathbf{A}$ governs the evolution of states, while the matrices $\\mathbf{B}$ and $\\mathbf{C}$ are used for modeling the relationship between inputs and outputs through the latent states. To handle discrete-time sequence data, the zero-order hold discretization rule is used:", + "bbox": [ + 511, + 541, + 908, + 617 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} h _ {t} = \\overline {{\\mathbf {A}}} h _ {t - 1} + \\overline {{\\mathbf {B}}} x _ {t}, y _ {t} = \\mathbf {C} h _ {t}, \\\\ \\overline {{\\mathbf {A}}} = \\exp (\\boldsymbol {\\Delta} \\mathbf {A}), \\overline {{\\mathbf {B}}} = (\\boldsymbol {\\Delta} \\mathbf {A}) ^ {- 1} (\\exp (\\boldsymbol {\\Delta} \\mathbf {A}) - \\mathbf {I}) \\boldsymbol {\\Delta} \\mathbf {B}, \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 623, + 903, + 661 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\Delta$ represents the sampling interval between consecutive inputs $x_{t-1}$ and $x_t$ . However, due to the linear time-invariant nature of SSM, the parameters $(\\Delta, \\mathbf{A}, \\mathbf{B}, \\mathbf{C})$ remain fixed across all time steps, which limits the expressive capacity of SSM. To overcome this limitation, Mamba [9] treats the parameters $(\\Delta, \\mathbf{B}, \\mathbf{C})$ as functions of input $x_t$ , transforming the SSM into a time-varying model:", + "bbox": [ + 511, + 669, + 908, + 776 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nh _ {t} = \\phi_ {\\overline {{\\mathbf {A}}}} (x _ {t}) h _ {t - 1} + \\phi_ {\\overline {{\\mathbf {B}}}} (x _ {t}) x _ {t}, y _ {t} = \\phi_ {\\mathbf {C}} (x _ {t}) h _ {t}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 784, + 906, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\phi_{*}(x_{t})$ denotes the linear projection of input $x_{t}$ . Additionally, Mamba introduces a hardware-aware scan algorithm to achieve near-linear complexity. Our StruMamba3D adopts the same hardware-aware scan algorithm, maintaining linear complexity while exhibiting strong structural modeling capabilities for point clouds.", + "bbox": [ + 511, + 809, + 910, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28548", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2599e73e448e8ac0f9e9a2c95cf179ca547207d5e4ff9169e95adf4f49c84a6d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 125, + 83, + 867, + 224 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/245557cd6982fb0bcd1a4f50126a0e1e15b54018feef192feebb36d817beaba7.jpg", + "image_caption": [ + "Figure 2. Overview of StruMamba3D. (a): Architecture of StruMamba3D. (b): Details of Structural SSM Block. We employ structural SSM to capture dependencies between spatial states and input points. Additionally, we integrate a lightweight convolution module to introduce local interactions among spatial states or input points. (c): Details of Structural SSM. (d): Details of Lightweight Convolution." + ], + "image_footnote": [], + "bbox": [ + 125, + 229, + 447, + 378 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ff4d03cf7c1c8516d51673698a018b37c25efae3b3fb8a450e8440a44f665fd4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 455, + 229, + 684, + 378 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/21a8c32d4dce1a734c0437ced7e56527159a55eb8ba35a8d215373a21b2dc06e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 229, + 867, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Overview", + "text_level": 1, + "bbox": [ + 89, + 438, + 200, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Despite Mamba achieving remarkable results on sequential data, maintaining the spatial relationships among 3D points remains challenging. We aim to model the spatial relationships by latent states and design StruMamba3D for point cloud representation learning, as illustrated in Fig. 2(a). Given raw point cloud $P_{raw} \\in \\mathbb{R}^{N_0 \\times 3}$ , we obtain the input points $P_x \\in \\mathbb{R}^{N \\times 3}$ using farthest point sampling (FPS) and $K$ -nearest neighbor clustering (KNN). Then, we use a lightweight PointNet [27] to extract the input token embeddings $F_x \\in \\mathbb{R}^{N \\times D}$ . Meanwhile, we initialize the spatial states $P_h \\in \\mathbb{R}^{M \\times 3}$ as the proxies for local structures from the raw point cloud. Finally, the spatial states and input token embeddings are fed into a feature encoder composed of multiple structural SSM blocks, where the spatial states act as a bridge to connect the inputs and outputs. Notably, we use the output spatial states from the previous structural SSM block as the input spatial states for the current block.", + "bbox": [ + 88, + 459, + 483, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Structural SSM Block", + "text_level": 1, + "bbox": [ + 89, + 727, + 297, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The structural SSM block is designed to capture complex structural features by preserving spatial dependencies among points throughout SSM processing. As shown in Fig. 2(b), the block takes spatial states $F_{h}$ and token embeddings $F_{x}$ as input and produces updated spatial states $F_{h}^{\\prime}$ and token embeddings $F_{x}^{\\prime}$ as output. To clarify the details of structural SSM block, we describe each component in sequence: Spatial State Initialization, State-wise Update Mechanism, Structural SSM, and Lightweight Convolution. Spatial State Initialization. In original Mamba, latent", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "states do not contain geometric information, making it difficult to model the local structure of point clouds. To address this, we introduce positional attributes $P_{h} \\in \\mathbb{R}^{M \\times 3}$ to the latent states, enabling each state to focus on distinct regions of the point cloud. Due to the sparsity of point clouds, we initialize state positions $P_{h}$ based on raw points $P_{raw}$ to ensure comprehensive coverage of point clouds. Specifically, we employ FPS and KNN clustering to segment the point cloud into groups $\\{\\mathcal{G}_{m}\\}_{m=1}^{M}$ , calculating the centroid:", + "bbox": [ + 511, + 438, + 906, + 575 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {h} ^ {m} = \\frac {1}{\\left| \\mathcal {G} _ {m} \\right|} \\sum_ {P _ {i} \\in \\mathcal {G} _ {m}} P _ {i}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 585, + 906, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $P_{h}^{m}$ represents the positional attribute of the $m$ -th state. We refer to states with positional attributes as spatial states and then use linear mapping $\\phi_{h}$ to embed positional attributes as the initial state features.", + "bbox": [ + 511, + 632, + 906, + 693 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {h} = \\phi_ {h} \\left(P _ {h}\\right) \\in \\mathbb {R} ^ {M \\times D}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 704, + 903, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The state features $F_{h}$ are then used as the initial latent states in SSM and updated along with input token embeddings.", + "bbox": [ + 511, + 734, + 906, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "State-wise Update. As illustrated in Eq. (3), the state update and propagation equations of the original Mamba rely solely on input tokens, disregarding the spatial relationships between input points and spatial states. This limitation makes it challenging for Mamba to effectively model the structural information of point clouds. To address this, we modify the state update and propagation equations to explicitly model the spatial dependencies between states and inputs. Specifically, we first calculate the relative offsets", + "bbox": [ + 511, + 763, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28549", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "between the input points $P_{x}$ and the state positions $P_{h}$ :", + "bbox": [ + 89, + 90, + 454, + 107 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\triangle P _ {i} ^ {m} = P _ {x} ^ {i} - P _ {h} ^ {m}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 112, + 482, + 130 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\triangle P_i^m$ denotes the relative offset between the $i$ -th input point and the $m$ -th state. Next, we map these offsets to the SSM parameters $(\\mathbf{B}, \\mathbf{C})$ , which govern state update and propagation. It is important to note that considering only spatial relationships is insufficient, as the significance of information varies across different points. Thus, we design the following state update and propagation mechanism:", + "bbox": [ + 89, + 135, + 483, + 242 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\mathbf {B} _ {i} ^ {m}, \\mathbf {C} _ {i} ^ {m}\\right) = \\phi \\left(x _ {i}\\right) + \\operatorname {M L P} \\left(\\triangle P _ {i} ^ {m}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 248, + 482, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where MLP represents a multi-layer perceptron, $\\phi$ is a linear projection and $x_{i}$ denotes the feature of the $i$ -th input point. Based on the modified parameters, the spatial states can selectively update features using points within the same regions, while the input points can acquire local structural information from the states, as shown in Fig. 2(c).", + "bbox": [ + 89, + 271, + 483, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Structural SSM. By leveraging spatial states and state-wise updates, we effectively model the structural information of point clouds. However, due to unidirectional scanning mechanism, bidirectional information exchange between input points cannot be achieved within a single forward pass. To address this, we adopt a bidirectional scanning mechanism inspired by VisionMamba [25]. Specifically, we apply both forward and backward structural SSMs, with the backward SSM processing input points in reverse order. Both use spatial states $\\hat{F}_h$ as initial states $\\hat{F}_h^0$ , generating updated features for input points and spatial states. Finally, we fuse these outputs using a linear layer $\\phi_o$ to obtain the final features $F_x'$ and $F_h'$ for input points and spatial states.", + "bbox": [ + 89, + 362, + 483, + 559 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {x} ^ {\\prime}, F _ {h} ^ {\\prime} = \\phi_ {o} (\\mathrm {S S M} _ {f} (\\hat {F} _ {x}, \\hat {F} _ {h}) + \\mathrm {S S M} _ {b} (\\hat {F} _ {x}, \\hat {F} _ {h})), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 565, + 482, + 584 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f$ and $b$ denote the forward and backward processes.", + "bbox": [ + 89, + 590, + 482, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lightweight Convolution. The proposed structural SSM effectively models the dependencies between spatial states $P_{h}$ and input points $P_{x}$ . However, we observe that the spatial states $P_{h}$ in SSMs remain isolated, lacking direct interaction with each other. To address this, we introduce a lightweight convolution module, as shown in Fig. 2(d). Inspired by graph convolution [35], this module aggregates features from neighboring states based on relative position information. Specifically, for each spatial state $P_{h}^{m}$ , we first identify its $k$ -nearest neighbors $\\mathcal{N}(m)$ and compute the relative positions $\\triangle P_{h}^{mj}$ , $j \\in \\mathcal{N}(m)$ . We then generate attention weights $\\{w_{mj}\\}$ for each neighbor using a linear layer $\\phi_{w}$ , followed by a softmax operation:", + "bbox": [ + 89, + 604, + 483, + 801 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {h} ^ {m j} = \\operatorname {s o f t m a x} \\left(\\phi_ {w} \\left(\\triangle P _ {h} ^ {m j}, P _ {h} ^ {m}\\right)\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 806, + 482, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The output state feature $\\hat{F}_h^m$ for each state $P_h^m$ is obtained through weighted aggregation:", + "bbox": [ + 89, + 832, + 483, + 864 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {F} _ {h} ^ {m} = \\phi_ {c} \\left(\\sum_ {j \\in \\mathcal {N} (m)} w _ {h} ^ {m j} F _ {h} ^ {m j}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 869, + 482, + 904 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a664a242dbbb3fd499a1477eae917724483c199649eb0fd4f42f01d624e87f4a.jpg", + "image_caption": [ + "Figure 3. Overview of the pre-training framework. We pretrain the StruMamba3D (student model) using the MPM task and the spatial state consistency loss. The teacher model has the same structure as the student model, but its parameters are updated using the exponential moving average strategy." + ], + "image_footnote": [], + "bbox": [ + 517, + 84, + 898, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $F_{h}^{mj}$ denotes the neighboring state features and $\\phi_c$ is a linear layer. By extending the receptive field of spatial states, this lightweight convolution enhances their ability to capture global semantic information. Furthermore, we replace the causal 1D convolution in the original Mamba block with this module, making it better suited for processing input points with complex geometric structures.", + "bbox": [ + 511, + 345, + 906, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Sequence Length-adaptive Strategy", + "text_level": 1, + "bbox": [ + 511, + 462, + 821, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address the issue of insufficient long-sequence memory in downstream tasks, we propose a sequence length-adaptive strategy that integrates an adaptive state update mechanism and spatial state consistency loss.", + "bbox": [ + 511, + 484, + 905, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adaptive State Update Mechanism. According to Eq. (2), the parameter sampling interval $\\Delta$ controls the frequency of state updates by influencing the parameters $(\\overline{\\mathbf{A}},\\overline{\\mathbf{B}})$ in SSMs. A larger $\\Delta$ causes the model to update state features more frequently, whereas a smaller $\\Delta$ allows the model to maintain its current state with minimal changes. To preserve memory for longer sequences in downstream tasks, we propose an adaptive state update mechanism. This mechanism adjusts the sampling interval $\\Delta$ according to the input sequence length, allowing the model to maintain a consistent total sampling time $\\Delta_{all} = \\sum_{i=1}^{N}\\Delta_{i}$ across different sequence lengths. Specifically, we design a learnable parameter $\\tau$ , which regulates the total sampling time $\\Delta_{all}$ for sequences of varying lengths. The sampling interval $\\Delta_{i}$ for each token in the inputs is calculated as follows:", + "bbox": [ + 511, + 545, + 906, + 771 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Delta} _ {i} = \\frac {\\tau \\times \\boldsymbol {\\Delta} _ {i}}{\\sum_ {i = 1} ^ {N} \\boldsymbol {\\Delta} _ {i}}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 648, + 780, + 906, + 818 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the adaptive state update mechanism, the model can maintain a consistent total sampling time $\\Delta_{all} = \\tau$ across different sequence lengths.", + "bbox": [ + 511, + 825, + 905, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Spatial State Consistency Loss. To further ensure consistency in updated state features across different input se", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28550", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/274944cf8ed717b3a0232b8203858cc75e5b978d97cf646adf111f5dd6c559b0.jpg", + "table_caption": [ + "Table 1. Shape Classification on ScanObjectNN and ModelNet40 Datasets. For ScanObjectNN, we report the classification accuracy(%) over the three subsets: OBJ-BG, OBJ-ONLY, and PB-T50-RS. And $\\dagger$ indicates using simple rotational augmentation [7] for fine-tuning." + ], + "table_footnote": [], + "table_body": "
MethodBackboneParam. (M)FLOPs (G)ScanObjectNNMODELNET40
OBJ-BGOBJ-ONLYPB-T50-RSw/o Votingw/ Voting
Supervised learning only
PointNet [27]-3.50.573.379.268.089.2-
PointNet++ [28]-1.51.782.384.377.990.7-
PointCNN [19]-0.6-86.185.578.592.2-
DGCNN [35]-1.82.482.886.278.192.9-
PRANet [8]-2.3---81.093.7-
PointNeXt [30]-1.43.6--87.794.0-
PointMLP [24]-12.631.4--85.494.5-
DeLA [2]-5.31.5--88.694.0-
PCM [47]-34.245.0--88.193.4-
Pre-training using single-modal information
PointBERT [42]Transformer22.14.887.4388.1283.0792.793.2
MaskPoint [22]Transformer22.14.889.3088.1084.30-93.8
PointM2AE [44]Transformer12.77.991.2288.8186.4392.993.4
PointMAE† [25]Transformer22.14.892.7791.2289.0492.793.8
PointGPT-S† [3]Transformer29.25.793.3992.4389.1793.394.0
PointMamba† [21]Mamba12.33.194.3292.6089.3193.694.1
Mamba3D† [15]Mamba16.93.993.1292.0892.0594.795.1
Ours†Structural SSM15.84.095.1893.6392.7595.195.4
Pre-training using cross-modal information
ACT† [7]Transformer22.14.893.2991.9188.2193.794.0
Joint-MAE [14]Transformer22.1-90.9488.8686.07-94.0
I2P-MAE† [46]Transformer15.3-94.1591.5790.1193.794.1
ReCon† [29]Transformer43.65.395.1893.2990.6394.594.7
", + "bbox": [ + 94, + 106, + 898, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "quence lengths, we introduce a spatial state consistency loss to the pre-training task. As shown in Fig. 3, we present an overview of the pre-training framework. Following the previous works [15, 21, 25], we reconstruct the points in the masked regions and use the reconstruction loss $\\mathcal{L}_{cd}$ to guide the model to learn discriminative features from the visible points. For the spatial state consistency loss, we first use a teacher model to update the spatial states with the full input tokens $P_{x}^{f}\\in \\mathbb{R}^{N_{f}\\times 3}$ . The state features $F_h^{\\prime f}$ output by the teacher model serve as pseudo-labels. We then enforce a consistency loss $\\mathcal{L}_{ssc}$ between the updated state features of the student model $F_h^{\\prime v}$ and the pseudo-labels $F_h^{\\prime f}$ :", + "bbox": [ + 88, + 500, + 485, + 684 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s s c} = \\operatorname {S m o o t h} \\mathrm {L} 1 \\left(F _ {h} ^ {\\prime v}, F _ {h} ^ {\\prime f}\\right). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 705, + 482, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The spatial state consistency loss $\\mathcal{L}_{ssc}$ not only enhances the robustness of our StruMamba3D to varying input sequence lengths but also improves the ability to infer the structural information of the complete point cloud from incomplete point clouds. The final pre-training loss is:", + "bbox": [ + 89, + 746, + 483, + 824 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {c d} + \\lambda \\times \\mathcal {L} _ {s s c}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 845, + 482, + 862 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda$ is the weight of the spatial state consistency loss.", + "bbox": [ + 89, + 885, + 472, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 500, + 718, + 515 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 523, + 730, + 540 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following the previous works [15, 21, 25], we employ $T = 12$ structural SSM blocks, each with a feature dimension of $D = 384$ . The number of spatial states $M$ is set to 16. For the lightweight convolution in the structural SSM block, we set the number of neighbors to 4 for spatial states and 8 for input points. Following previous works [25, 42, 44], we pre-train the proposed model on ShapeNet dataset [1], which contains 52472 unique 3D models across 55 common object categories. Each input raw point cloud, containing $N_0 = 1024$ points, is divided into 64 patches with each consisting of 32 points. A random masking ratio of 0.6 is applied, with the decoder composed of 4 standard Mamba blocks, and the $\\lambda$ in the pre-training loss is set to 2. More details can be found in the supplementary material.", + "bbox": [ + 509, + 547, + 906, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison on Downstream Tasks", + "text_level": 1, + "bbox": [ + 511, + 771, + 818, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present the experimental results of our StruMamba3D across various downstream tasks and compare them with previous state-of-the-art (SOTA) methods.", + "bbox": [ + 511, + 794, + 905, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Shape classification. ScanObjectNN [41] is a real-world dataset containing roughly 15000 objects across 15 categories, some of which have cluttered backgrounds or partial scans. For this dataset, we apply rotation as a data aug-", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28551", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a264214adad9518a83d43cb7b6caa9dc0f604ef3feeb84b61d4c597195e29a6b.jpg", + "table_caption": [ + "Table 2. Part Segmentation on ShapeNetPart. We report the mean IoU for all classes $(\\mathrm{mIoU}_c)$ and all instances $(\\mathrm{mIoU}_i)$ ." + ], + "table_footnote": [], + "table_body": "
MethodArchitecturemIoUcmIoUi
Supervised learning only
PointNet [27]Single-scale80.483.7
PointNet++ [28]Multi-scale81.985.1
APES [36]Multi-scale83.785.8
DeLA [2]Multi-scale85.887.0
PCM [47]Multi-scale85.387.0
Pre-training using single-modal information
MaskPoint [22]Single-scale84.686.0
PointBERT [42]Single-scale84.185.6
PointMAE [25]Single-scale84.286.1
PointM2AE [44]Multi-scale84.986.5
PointGPT-S [3]Single-scale84.186.2
PointMamba [21]Single-scale84.486.2
Mamba3D [15]Single-scale83.685.6
OursSingle-scale85.086.7
", + "bbox": [ + 102, + 103, + 470, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mentation strategy [7] and use 2048 points as input. ModelNet40 [39] is a synthetic dataset comprising of 12311 clean 3D CAD models from 40 categories. For this dataset, we use 1024 points as input and apply scale and translation augmentations [27]. As shown in Tab. 1, we conduct classification experiments on both datasets. Compared to transformer-based methods, our model achieves significant performance improvements across multiple metrics on both datasets. Even when compared to the best PointGPTS [3], our model outperforms it by $3.58\\%$ on ScanObjectNN and $1.8\\%$ on ModelNet40. Additionally, on the ScanObjectNN dataset, StruMamba3D surpasses PointMamba [21] by $0.86\\%$ , $1.03\\%$ , and $3.44\\%$ across the three splits, while also outperforming Mamba3D [15] by $2.06\\%$ , $1.55\\%$ , and $0.70\\%$ . Furthermore, on the ModelNet40 dataset, StruMamba3D exceeds PointMamba by $1.5\\%$ and outperforms Mamba3D by $0.4\\%$ . Overall, these results demonstrate that StruMamba3D outperforms all existing unimodal architectures, achieving multiple SOTA results and highlighting the effectiveness of structural modeling within Mamba.", + "bbox": [ + 88, + 340, + 482, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Part Segmentation on ShapeNetPart. ShapeNetPart [41] dataset comprises 16880 models spanning 16 shape categories and annotated with 50 part labels. Compared to classification tasks, part segmentation poses a greater challenge, as it requires assigning a label to each point, making it highly dependent on the intrinsic structural information of point clouds. Tab. 8 presents our part segmentation results for all classes and instances on the ShapeNetPart dataset. For single-scale models, our method achieves substantial performance improvements, surpassing Mamba3D [15] by $1.4\\%$ in $\\mathrm{mIoU}_c$ and $1.2\\%$ in $\\mathrm{mIoU}_i$ , and outperforming PointMamba [21] by $0.6\\%$ in $\\mathrm{mIoU}_c$ and $0.5\\%$ in $\\mathrm{mIoU}_i$ . Notably, Mamba3D without a serialization strategy performs poorly, falling behind the baseline PointMAE [25], whereas PointMamba, which incorporates serialization, exhibits improved performance. This observation suggests the crucial role of serialization in enhancing structural mod", + "bbox": [ + 89, + 643, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4621cba61e61882b39dfbe95122a77e25ba70090061b4f7b6f908658e4190fd4.jpg", + "table_caption": [ + "Table 3. Few-Shot Classification on ModelNet40. We report overall accuracy and standard deviation without a voting strategy." + ], + "table_footnote": [], + "table_body": "
Method5-way10-way
10-shot20-shot10-shot20-shot
PointBERT [42]94.6±3.693.9±3.186.4±5.491.3±4.6
MaskPoint [22]95.0±3.797.2±1.791.4±4.092.7±5.1
PointMAE [25]96.3±2.597.8±1.892.6±4.193.4±3.5
PointM2AE [44]96.8±1.898.3±1.492.3±4.595.0±3.0
PointGPT-S [3]96.8±2.098.6±1.192.6±4.695.2±3.4
PointMamba [21]96.9±2.099.0±1.193.0±4.495.6±3.2
Mamba3D [15]96.4±2.298.2±1.292.4±4.195.2±2.9
Ours97.5±2.399.1±1.493.5±3.796.1±3.5
", + "bbox": [ + 521, + 104, + 898, + 252 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "eling within Mamba-based architectures. In contrast, our StruMamba3D leverages spatial states to effectively capture structural information, thereby eliminating the need for serialization and achieving $85.0\\%$ in $\\mathrm{mIoU}_c$ and $86.7\\%$ in $\\mathrm{mIoU}_i$ . Besides, StruMamba3D even outperforms the multi-scale self-supervised method PointM2AE [44]. Although multi-scale approaches typically yield better performance in part segmentation tasks, they often come with significantly higher computational costs (e.g., PCM [47] requires 45.0G FLOPs). In this work, we focus on designing an efficient and effective Mamba-based framework for structural modeling in point clouds. While incorporating multi-scale designs could further improve part segmentation performance, such strategies may introduce heuristic dependencies. Overall, our StruMamba3D represents a more robust and efficient Mamba-based framework for point cloud structural modeling, achieving the highest performance among all single-scale models.", + "bbox": [ + 511, + 257, + 906, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Few-shot Learning on ModelNet40. To evaluate the effectiveness of the proposed methods with limited fine-tuning data, we conduct few-shot classification experiments on the ModelNet40 dataset. Following the standard procedure [25, 42], we perform 10 independent experiments for each setting and report the mean accuracy and standard deviation. As shown in Tab. 3, our method demonstrates a significant improvement over Mamba-based methods, outperforming PointMamba [21] by $0.6\\%$ in 5-way-10-shot and $0.5\\%$ in 10-way-10-shot settings, and surpassing Mamba3D [15] by $1.1\\%$ and $0.9\\%$ , respectively. These results indicate that our method can efficiently adapt to downstream tasks with longer input sequences, even under the constraint of limited training data.", + "bbox": [ + 511, + 530, + 908, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 755, + 666, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To explore the architectural design, we perform ablation studies on the PB-T50-RS split of ScanObjectNN dataset, as well as ModelNet40 and ShapeNetPart datasets.", + "bbox": [ + 511, + 777, + 903, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Designed Modules. As shown in Tab. 4, we conduct an ablation study to evaluate the effectiveness of two core modules: the Structural SSM block and the sequence length-adaptive strategy. The baseline model is built upon the standard Mamba [9] block, following the same", + "bbox": [ + 511, + 824, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28552", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8aaeb6bba98e9426eaf9522afea08b40f597ede92ef618dcc13d733f1249ee04.jpg", + "table_caption": [ + "Table 4. Effect of the designed modules. 'SLAS' indicates the sequence length-adaptive strategy used in pre-training phase." + ], + "table_footnote": [], + "table_body": "
Structural SSM BlockSLASOverall AccuracymIoUc
ScanNNMN40SNPart
XX87.2391.8681.56
✓X92.0994.4584.49
✓✓92.7595.0684.96
", + "bbox": [ + 99, + 104, + 470, + 185 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7a44d0d8d3cd0c72266fe7b0c65779a28c0ed1f7537b6698d59e3a05d0663a6a.jpg", + "table_caption": [ + "Table 5. Ablation on Structural SSM Block. The baseline uses the standard Mamba block to replace the Structural SSM Block." + ], + "table_footnote": [], + "table_body": "
MethodOverall AccuracymIoUc
ScanNNMN40SNPart
baseline88.2492.5082.08
w/ Structural SSM91.7893.9284.15
w/ Lightweight Conv. for h92.2294.6584.62
w/ Lightweight Conv. for x92.4094.8184.77
w/ Bidirectional Scanning92.7595.0684.96
", + "bbox": [ + 99, + 220, + 470, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "architecture as PointMAE [25]. Compared to this baseline, the Structural SSM block significantly enhances model performance, yielding improvements of $4.86\\%$ on ScanObjectNN, $2.59\\%$ on ModelNet40, and $2.93\\%$ on ShapeNetPart, demonstrating its effectiveness in modeling spatial dependencies and capturing richer structural information in point clouds. Furthermore, integrating the sequence length-adaptive strategy further enhances performance by improving the generalization to varying input lengths, thereby ensuring optimal performance across downstream tasks.", + "bbox": [ + 88, + 337, + 482, + 488 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Structural SSM Block. As shown in Tab. 5, we further validate the effectiveness of individual components within the Structural SSM block. Starting from the baseline model with the standard Mamba block, we incrementally incorporate key modules from the Structural SSM block. The most significant performance gains stem from the Structural SSM and the lightweight convolution designed for spatial states, demonstrating that modeling structural information through hidden states in SSM is crucial. Moveover, the lightweight convolution for input points and the bidirectional scanning mechanism also enhance the performance across downstream tasks.", + "bbox": [ + 89, + 489, + 482, + 671 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Structural SSM. Structural SSM is a crucial component of our method, as it effectively captures the structural information of point clouds through spatial states. To further analyze the impact of spatial state initialization and state-wise update strategy, we conduct ablation studies as shown in Tab. 6. Although spatial state initialization provides the states with positional information, the improvement is minimal, as Eq. (2) does not consider spatial relationships between input points and states. When integrating the state-wise update strategy, the model achieves a significant performance boost, with improvements of $1.42\\%$ on ScanObjectNN, $0.94\\%$ on ModelNet40, and $1.09\\%$ on ShapeNetPart. The results indicate that incorporating spatial relationships to control state information updates and propagation within the SSM is essential for effectively mod", + "bbox": [ + 89, + 672, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4b0bb4422522347265d280ee0371273a09e1e0f24ee9c828a177227f5c585522.jpg", + "table_caption": [ + "Table 6. Ablation on Structural SSM. 'SS' represents the use of spatial states in the SSM. $\\phi (x)$ denotes the SSM parameters depending on the point feature, while $\\mathrm{MLP}(\\triangle P)$ represents the SSM parameters depending on the spatial relationship." + ], + "table_footnote": [], + "table_body": "
φ(x)SSMLP(△P)Overall AccuracymIoUc
ScanNNMN40SNPart
✓XX90.9493.8483.62
✓✓X91.3394.1283.87
✓✓✓92.7595.0684.96
X✓✓91.1294.2584.02
", + "bbox": [ + 532, + 133, + 883, + 226 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fb65804a6a29c96ac01b82c7cc69017a76755aed2874420b13e8558456e3daf6.jpg", + "table_caption": [ + "Table 7. Ablation on sequence length-adaptive strategy. 'ASUM' indicates the adaptive state update mechanism, and $\\lambda$ ’ indicates the weight of the spatial state consistency loss ${\\mathcal{L}}_{ssc}$ ." + ], + "table_footnote": [], + "table_body": "
ASUMLsscOverall Accuracy
ScanNNMN40SNPart
XX92.0994.4584.49
✓X92.2694.5784.56
✓λ = 192.4794.9484.81
✓λ = 292.7595.0684.96
✓λ = 592.5794.8984.77
Xλ = 292.1694.6584.56
", + "bbox": [ + 537, + 272, + 879, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "eling structural information. Moreover, we observe that relying solely on spatial relationships to generate SSM parameters results in performance degradation. This is likely due to the varying importance of point features, which cannot be fully captured by spatial position alone.", + "bbox": [ + 511, + 401, + 903, + 477 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Sequence Length-Adaptive Strategy. The sequence length-adaptive strategy consists of two key components: the adaptive state update mechanism and the spatial state consistency loss. Tab. 7 presents an ablation study on these components. When applied individually, both the adaptive state update mechanism and the spatial state consistency loss yield only marginal performance gains. However, when used together, they lead to a significant improvement, demonstrating their complementary effects. The adaptive state update mechanism ensures a consistent total sampling time across different input lengths, while the spatial state consistency loss ensures the consistency of spatial state features. In addition, we perform an ablation study on the weight $\\lambda$ of $\\mathcal{L}_{ssc}$ and observe that the model achieves optimal performance when $\\lambda$ is set to 2.", + "bbox": [ + 511, + 478, + 905, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 714, + 633, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We identify two key issues limiting Mamba potential: 3D point adjacency distortion during SSM processing and inadequate long-sequence memory retention in downstream tasks. To address these issues, we propose StruMamba3D with two core components. The structural SSM block maintains spatial dependencies during state selection and propagation, while the sequence length-adaptive strategy reduces the sensitivity of Mamba-based models to varying input lengths. Our method achieves substantial improvements across four downstream tasks, with ablation studies confirming the effectiveness of each module.", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28553", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgment", + "text_level": 1, + "bbox": [ + 89, + 90, + 261, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the National Defense Basic Scientific Research program (JCKY2022911B002).", + "bbox": [ + 89, + 108, + 482, + 138 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 152, + 187, + 169 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6, 11", + "[2] Binjie Chen, Yunzhou Xia, Yu Zang, Cheng Wang, and Jonathan Li. Decoupled local aggregation for point cloud learning, 2023. 6, 7", + "[3] Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. Pointgpt: Auto-regressively generative pre-training from point clouds. arXiv preprint arXiv:2305.11487, 2023. 3, 6, 7, 12, 13", + "[4] Siheng Chen, Baoan Liu, Chen Feng, Carlos Vallespi-Gonzalez, and Carl Wellington. 3d point cloud processing and learning for autonomous driving: Impacting map creation, localization, and perception. IEEE Signal Processing Magazine, 38(1):68-86, 2020. 1", + "[5] Tri Dao and Albert Gu. Transformers are ssms: Generalized models and efficient algorithms through structured state space duality. arXiv preprint arXiv:2405.21060, 2024. 3", + "[6] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT (1), pages 4171-4186. Association for Computational Linguistics, 2019. 2, 3", + "[7] Runpei Dong, Zekun Qi, Linfeng Zhang, Junbo Zhang, Jianjian Sun, Zheng Ge, Li Yi, and Kaisheng Ma. Autoencoders as cross-modal teachers: Can pretrained 2d image transformers help 3d representation learning? In The Eleventh International Conference on Learning Representations (ICLR), 2023. 6, 7", + "[8] Deng-Ping Fan, Ge-Peng Ji, Tao Zhou, Geng Chen, Huazhu Fu, Jianbing Shen, and Ling Shao. Pranet: Parallel reverse attention network for polyp segmentation. In International conference on medical image computing and computer-assisted intervention, pages 263-273. Springer, 2020. 6", + "[9] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2, 3, 7, 14", + "[10] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021.", + "[11] Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, and Christopher Ré. Combining recurrent, convolutional, and continuous-time models with linear state space layers. Advances in neural information processing systems, 34:572-585, 2021. 3", + "[12] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7:187-199, 2021. 2,3" + ], + "bbox": [ + 93, + 179, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3d point clouds: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(12):4338-4364, 2020. 1", + "[14] Ziyu Guo, Renrui Zhang, Longtian Qiu, Xianzhi Li, and Pheng-Ann Heng. Joint-mae: 2d-3d joint masked autoencoders for 3d point cloud pre-training. arXiv preprint arXiv:2302.14007, 2023.6", + "[15] Xu Han, Yuan Tang, Zhaoxuan Wang, and Xianzhi Li. Mamba3d: Enhancing local features for 3d point cloud analysis via state space model. In Proceedings of the 32nd ACM International Conference on Multimedia, page 4995-5004. Association for Computing Machinery, 2024. 2, 3, 6, 7", + "[16] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2, 3", + "[17] Rudolph Emil Kalman. A new approach to linear filtering and prediction problems. Journal of Basic Engineering, 1960. 3", + "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 11, 12", + "[19] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. Advances in neural information processing systems, 31, 2018. 1, 2, 3, 6", + "[20] Ying Li, Lingfei Ma, Zilong Zhong, Fei Liu, Michael A Chapman, Dongpu Cao, and Jonathan Li. Deep learning for lidar point clouds in autonomous driving: A review. IEEE Transactions on Neural Networks and Learning Systems, 32 (8):3412-3432, 2020. 1", + "[21] Dingkang Liang, Xin Zhou, Wei Xu, Xingkui Zhu, Zhikang Zou, Xiaqing Ye, Xiao Tan, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. In Advances in Neural Information Processing Systems, 2024. 2, 3, 6, 7, 12, 13", + "[22] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In European Conference on Computer Vision, pages 657-675. Springer, 2022. 2, 3, 6, 7, 13", + "[23] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 11, 12", + "[24] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. arXiv preprint arXiv:2202.07123, 2022. 3, 6", + "[25] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In European conference on computer vision, pages 604-621. Springer, 2022. 2, 3, 5, 6, 7, 8, 11", + "[26] François Pomerleau, Francis Colas, Roland Siegwart, et al. A review of point cloud registration algorithms for mobile robotics. Foundations and Trends® in Robotics, 4(1):1-104, 2015. 1" + ], + "bbox": [ + 516, + 92, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28554", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 1, 2, 3, 4, 6, 7", + "[28] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 1, 2, 3, 6, 7, 11", + "[29] Zekun Qi, Runpei Dong, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Contrast with reconstruct: Contrastive 3d representation learning guided by generative pretraining. In International Conference on Machine Learning (ICML), 2023. 3, 6", + "[30] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. Advances in neural information processing systems, 35:23192-23204, 2022. 1, 3, 6", + "[31] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 1, 3", + "[32] Chuxin Wang, Jiacheng Deng, Jianfeng He, Tianzhu Zhang, Zhe Zhang, and Yongdong Zhang. Long-short range adaptive transformer with dynamic sampling for 3d object detection. IEEE Transactions on Circuits and Systems for Video Technology, 33(12):7616-7629, 2023. 1", + "[33] Chuxin Wang, Yixin Zha, Jianfeng He, Wenfei Yang, and Tianzhu Zhang. Rethinking masked representation learning for 3d point cloud understanding. IEEE Transactions on Image Processing, 2024. 3", + "[34] ChuXin Wang, Wenfei Yang, Xiang Liu, and Tianzhu Zhang. State space model meets transformer: A new paradigm for 3d object detection. In The Thirteenth International Conference on Learning Representations, 2025. 3", + "[35] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 2, 3, 5, 6", + "[36] Chengzhi Wu, Junwei Zheng, Julius Pfrommer, and Jürgen Beyerer. Attention-based point cloud edge sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5333-5343, 2023. 7", + "[37] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. Advances in Neural Information Processing Systems, 35:33330-33342, 2022. 3", + "[38] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4840-4851, 2024. 3", + "[39] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d" + ], + "bbox": [ + 91, + 90, + 480, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 7", + "[40] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 574-591. Springer, 2020. 3", + "[41] Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Transactions on Graphics (ToG), 35(6):1-12, 2016. 6, 7, 12", + "[42] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19313-19322, 2022. 2, 3, 6, 7, 11, 13", + "[43] Yixin Zha, Chuxin Wang, Wenfei Yang, and Tianzhu Zhang. Exploring semantic masked autoencoder for self-supervised point cloud understanding. arXiv preprint arXiv:2506.21957, 2025.", + "[44] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training. Advances in neural information processing systems, 35:27061-27074, 2022. 2, 3, 6, 7", + "[45] Renrui Zhang, Liuhui Wang, Yu Qiao, Peng Gao, and Hongsheng Li. Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21769-21780, 2023. 3", + "[46] Renrui Zhang, Liuhui Wang, Yu Qiao, Peng Gao, and Hongsheng Li. Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21769-21780, 2023. 6", + "[47] Tao Zhang, Xiangtai Li, Haobo Yuan, Shunping Ji, and Shuicheng Yan. Point cloud mamba: Point cloud learning via state space model. arXiv preprint arXiv:2403.00762, 2024. 2, 3, 6, 7", + "[48] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3d features on any point-cloud. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10252-10263, 2021. 3", + "[49] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 3" + ], + "bbox": [ + 516, + 92, + 906, + 810 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28555", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/StruMamba3D_ Exploring Structural Mamba for Self-supervised Point Cloud Representation Learning/9ae9c283-762a-4256-b4de-303cb73e1e2c_model.json b/2025/StruMamba3D_ Exploring Structural Mamba for Self-supervised Point Cloud Representation Learning/9ae9c283-762a-4256-b4de-303cb73e1e2c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5b0214b5c45bf7d0d67927842a3118943c9d5d71 --- /dev/null +++ b/2025/StruMamba3D_ Exploring Structural Mamba for Self-supervised Point Cloud Representation Learning/9ae9c283-762a-4256-b4de-303cb73e1e2c_model.json @@ -0,0 +1,2277 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.17, + 0.131, + 0.828, + 0.177 + ], + "angle": 0, + "content": "StruMamba3D: Exploring Structural Mamba for Self-supervised Point Cloud Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.203, + 0.752, + 0.222 + ], + "angle": 0, + "content": "Chuxin Wang\\(^{1,2}\\), Yixin Zha\\(^{1}\\), Wenfei Yang\\(^{1,2,*}\\), Tianzhu Zhang\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.307, + 0.222, + 0.692, + 0.24 + ], + "angle": 0, + "content": "1University of Science and Technology of China" + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.24, + 0.86, + 0.258 + ], + "angle": 0, + "content": "\\(^{2}\\)National Key Laboratory of Deep Space Exploration, Deep Space Exploration Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.258, + 0.778, + 0.276 + ], + "angle": 0, + "content": "{wcx0602, zyxcn} \\(@\\) mail.ustc.edu.cn, {yangwf, tzzhang} \\(@\\) ustc.edu.cn" + }, + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.305, + 0.264, + 0.316 + ], + "angle": 0, + "content": "Attention Block" + }, + { + "type": "image", + "bbox": [ + 0.162, + 0.322, + 0.309, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.167, + 0.375, + 0.279, + 0.385 + ], + "angle": 0, + "content": "Quadratic complexity" + }, + { + "type": "image_caption", + "bbox": [ + 0.133, + 0.387, + 0.313, + 0.398 + ], + "angle": 0, + "content": "(a) Transformer-based Paradigm" + }, + { + "type": "image_caption", + "bbox": [ + 0.33, + 0.305, + 0.435, + 0.315 + ], + "angle": 0, + "content": "Selective SSM Block" + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.322, + 0.468, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.335, + 0.374, + 0.43, + 0.386 + ], + "angle": 0, + "content": "Linear complexity" + }, + { + "type": "image_caption", + "bbox": [ + 0.478, + 0.305, + 0.567, + 0.316 + ], + "angle": 0, + "content": "Serialized Points" + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.321, + 0.565, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.474, + 0.374, + 0.57, + 0.386 + ], + "angle": 0, + "content": "Destroy adjacency" + }, + { + "type": "image_caption", + "bbox": [ + 0.689, + 0.305, + 0.76, + 0.317 + ], + "angle": 0, + "content": "Longer Input" + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.321, + 0.675, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.592, + 0.374, + 0.778, + 0.386 + ], + "angle": 0, + "content": "Fail to retain long-sequence memory" + }, + { + "type": "image_caption", + "bbox": [ + 0.686, + 0.387, + 0.782, + 0.399 + ], + "angle": 0, + "content": "" + }, + { + "type": "image", + "bbox": [ + 0.686, + 0.401, + 0.778, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.776, + 0.328, + 0.8, + 0.338 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image_caption", + "bbox": [ + 0.776, + 0.338, + 0.852, + 0.346 + ], + "angle": 0, + "content": "□ Point Feature" + }, + { + "type": "image_caption", + "bbox": [ + 0.776, + 0.348, + 0.8, + 0.359 + ], + "angle": 0, + "content": "□" + }, + { + "type": "image_caption", + "bbox": [ + 0.799, + 0.413, + 0.841, + 0.423 + ], + "angle": 0, + "content": "PointMAE" + }, + { + "type": "image_caption", + "bbox": [ + 0.799, + 0.426, + 0.839, + 0.437 + ], + "angle": 0, + "content": "PointGPT" + }, + { + "type": "image_caption", + "bbox": [ + 0.799, + 0.442, + 0.841, + 0.452 + ], + "angle": 0, + "content": "Mamba3D" + }, + { + "type": "image_caption", + "bbox": [ + 0.799, + 0.455, + 0.841, + 0.465 + ], + "angle": 0, + "content": "PointMaml" + }, + { + "type": "image_caption", + "bbox": [ + 0.799, + 0.467, + 0.841, + 0.477 + ], + "angle": 0, + "content": "Our" + }, + { + "type": "image_caption", + "bbox": [ + 0.799, + 0.479, + 0.82, + 0.488 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image_caption", + "bbox": [ + 0.279, + 0.387, + 0.62, + 0.399 + ], + "angle": 0, + "content": "adigm (b) Mamba-based Paradigm" + }, + { + "type": "image", + "bbox": [ + 0.139, + 0.404, + 0.86, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.502, + 0.907, + 0.572 + ], + "angle": 0, + "content": "Figure 1. (a): Transformer-based paradigm uses the attention block with quadratic complexity to model the dependencies between points. (b): Mamba-based paradigm uses the selective SSM with linear complexity. However, serialized points destroy the adjacency of 3D points, and the pre-trained selection mechanism fails to retain long-sequence memory. (c): Our StruMamba3D paradigm uses the structural SSM to maintain the spatial dependencies among points and the sequence length-adaptive strategy to retain long-sequence memory. (d): Our StruMamba3D significantly outperforms existing Transformer-based and Mamba-based methods." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.587, + 0.327, + 0.602 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.606, + 0.486, + 0.879 + ], + "angle": 0, + "content": "Recently, Mamba-based methods have demonstrated impressive performance in point cloud representation learning by leveraging State Space Model (SSM) with the efficient context modeling ability and linear complexity. However, these methods still face two key issues that limit the potential of SSM: Destroying the adjacency of 3D points during SSM processing and failing to retain long-sequence memory as the input length increases in downstream tasks. To address these issues, we propose StruMamba3D, a novel paradigm for self-supervised point cloud representation learning. It enjoys several merits. First, we design spatial states and use them as proxies to preserve spatial dependencies among points. Second, we enhance the SSM with a state-wise update strategy and incorporate a lightweight convolution to facilitate interactions between spatial states for efficient structure modeling. Third, our method reduces the sensitivity of pre-trained Mamba-based models to varying input lengths by introducing a sequence length-adaptive" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.589, + 0.907, + 0.664 + ], + "angle": 0, + "content": "strategy. Experimental results across four downstream tasks showcase the superior performance of our method. In addition, our method attains the SOTA \\(95.1\\%\\) accuracy on ModelNet40 and \\(92.75\\%\\) accuracy on the most challenging split of ScanObjectNN without voting strategy." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.681, + 0.645, + 0.696 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.706, + 0.907, + 0.842 + ], + "angle": 0, + "content": "Point cloud representation learning aims to extract geometric and semantic features from point clouds to support downstream tasks such as classification, segmentation, and detection. As a foundational task in Computer Vision (CV), it is crucial for various real-world applications, including autonomous driving [4, 20], AR [13], and robotics [26, 32]. However, unlike images or text with structured arrangements, point clouds are inherently unordered and sparse, presenting challenges for designing effective models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.842, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To solve this problem, numerous point cloud representation learning methods have been proposed, which can be broadly divided into three categories according to model architectures: Point-based methods [19, 27, 28, 30, 31," + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.889, + 0.235, + 0.902 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28546" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.482, + 0.438 + ], + "angle": 0, + "content": "35], Transformer-based methods [12, 22, 25, 42-44], and Mamba-based methods [15, 21, 47]. Point-based methods directly process point clouds by modeling the relationships between individual points, such as through set abstraction [27, 28] and graph convolution [19, 35]. These methods offer a simple and computationally efficient design. However, their limited receptive field hinders capturing global information, constraining complex structure modeling. Transformer-based methods leverage attention mechanisms to capture long-range dependencies, expanding the receptive field. Moreover, inspired by self-supervised learning paradigms in NLP [6] and CV [16], many transformer-based methods [22, 25, 42, 44] utilize large unlabeled point cloud datasets for pre-training, achieving impressive results in downstream tasks. However, these methods have limited scalability due to the quadratic complexity of the attention mechanism, as shown in Fig. 1(a). To overcome this limitation, Mamba-based methods [15, 21, 47] first convert the 3D points into a 1D sequence and then utilize selective State Space Model (SSM) [9] with linear complexity to extract point features. Leveraging the robust information aggregation ability of selective SSM, Mamba-based methods have attained promising performance." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.445, + 0.482, + 0.835 + ], + "angle": 0, + "content": "Although existing Mamba-based methods [15, 21, 47] have achieved comparable performance to transformer-based methods, two key issues limit the potential of SSM: (1) Distort the adjacency of points. As shown in Fig. 1(b), although methods [21, 47] use serialization strategies to align input points, they cannot ensure that spatially adjacent points remain neighbors in the 1D sequence. Unlike text and images, points inherently lack context information and rely on local structures as base feature units. This spatial distortion hampers Mamba to model fine-grained structure information. (2) Fail to retain long-sequence memory in downstream tasks. Mamba-based methods [15, 21] use Masked Point Modeling (MPM) [25, 42] as a pre-training task, which creates a significant discrepancy in input sequence lengths between pre-training and downstream tasks. Mamba leverages a selection mechanism to inductively learn new information while selectively forgetting outdated knowledge. However, the selection mechanism does not account for variations in input sequence length. Consequently, when Mamba is pre-trained on short sequences, it tends to update states more frequently. As the input sequence length increases in downstream tasks, frequent state updates lead to Mamba struggling to retain long-sequence memory, as shown in Fig. 1(b). The deficiency in long-sequence memory adversely affects the long-range semantics modeling, resulting in suboptimal performance." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.841, + 0.482, + 0.9 + ], + "angle": 0, + "content": "To unleash the potential of Mamba for point cloud representation learning, we explore solutions to address the above two issues. (1) How to maintain spatial dependencies among points? Maintaining spatial relationships" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.092, + 0.904, + 0.257 + ], + "angle": 0, + "content": "among 3D points within a 1D input sequence is challenging. Mamba efficiently aggregates contextual features from the 1D input sequence via hidden states. Is it possible to design a structural Mamba that models spatial relationships among points through hidden states? (2) How to retain long-sequence memory for longer inputs? When handling longer input sequences, states need to retain memory for an extended period to effectively capture the contextual information across the entire sequence. If the model adjusts the frequency of state updates when the sequence length changes, it can preserve memory for longer sequences." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.259, + 0.904, + 0.634 + ], + "angle": 0, + "content": "To achieve the above goals, we propose a novel StruMamba3D paradigm for self-supervised point cloud representation learning, which consists of two core components: structural SSM block and sequence length-adaptive strategy. In structural SSM block, we leverage states in SSM to model the spatial relationships among points. We begin by assigning positional attributes to the states, using them as proxies to capture the local structure of point clouds. Next, we propose a state-wise update strategy that incorporates the positional relationship between input points and spatial states during the state update and propagation process. Based on spatial states and state-wise update strategy, we can maintain spatial dependencies among points during the SSM process. To address the limitation of standard Mamba ignoring inter-state interactions, we introduce a lightweight convolution module for spatial states. In sequence length-adaptive strategy, the adaptive state update mechanism is first employed to adjust the state update frequency based on sequence length, ensuring that the states retain long-sequence memory as the sequence length grows. Additionally, we introduce a spatial state consistency loss in the pre-training task, which enforces consistency in the updated states across inputs of varying sequence lengths. The spatial state consistency loss enhances the robustness of the Mamba-based pre-trained model to varying input lengths." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.636, + 0.904, + 0.815 + ], + "angle": 0, + "content": "To sum up, the main contributions of this work are: (1) We propose the novel StruMamba3D for self-supervised point cloud representation learning, which is the first to model the structure information by latent states in SSM. (2) We design the structural SSM to maintain spatial dependencies among points during state information selection and propagation, while the sequence length-adaptive strategy reduces the sensitivity of pre-trained Mamba-based models to varying input lengths. (3) Experimental results on four downstream tasks showcase the superior performance of our method. Additionally, comprehensive ablation studies validate the effectiveness of our designs." + }, + { + "type": "title", + "bbox": [ + 0.518, + 0.831, + 0.653, + 0.845 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.857, + 0.904, + 0.885 + ], + "angle": 0, + "content": "We briefly review methods for point cloud representation learning and self-supervised representation learning." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.887, + 0.904, + 0.9 + ], + "angle": 0, + "content": "Point Cloud Representation Learning. To effectively" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "28547" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.501 + ], + "angle": 0, + "content": "leverage the sparse, unordered and irregular point cloud, various point cloud representation learning methods have been proposed, which can be broadly categorized into three groups: Point-based methods [19, 27, 28, 30, 31, 35], Transformer-based methods [12, 22, 25, 33, 42, 44], and Mamba-based methods [15, 21, 47]. Point-based methods directly process raw point clouds to learn geometric features. For example, PointNet [27] employs MLPs to each point independently, followed by global max pooling to aggregate features. PointNet++ [28] extends PointNet by introducing a hierarchical architecture for multi-scale feature extraction. Subsequent methods [19, 24, 30, 31, 35] explore convolutional techniques for point clouds. DGCNN [35] leverages graph convolution to model structural information among neighboring points, while KPCov [31] introduces a convolution method based on dynamic kernel points to improve feature learning. To capture long-range dependencies, methods [12, 37, 38, 49] have explored transformer architectures for point clouds. PCT [12] applies global attention to points, while Point Transformer [49] introduces local attention to capture local geometry features. Recent variants [37, 38] further refine the architectures to improve performance and efficiency. These supervised methods [12, 30, 31, 47, 49] have demonstrated outstanding performance in specific tasks. However, these methods are inherently limited by the specific data domains on which they are trained, leading to poor generalization to other domains." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.487, + 0.904 + ], + "angle": 0, + "content": "Self-supervised Representation Learning. To solve the above issues and fully utilize a large number of unlabeled data, self-supervised representation learning methods [3, 22, 25, 40, 42, 44, 48] have emerged. Some contrastive learning-based methods [40, 48] train models to learn distinguishing features by differentiating between positive and negative samples. For instance, Point-Contrast [40] leverages feature consistency across different views of the same point cloud. Inspired by the success of BERT [6] and MAE [16], many transformer-based methods [3, 22, 25, 42, 44] adopt MPM task for pre-training, encouraging models to infer masked patches from visible ones. PointBERT [42] predicts discrete tokens for masked points, while PointMAE [25] reconstructs masked point clouds. Additionally, recent studies [29, 45] integrate cross-modal information, such as images and text, to enhance pre-training. However, transformer-based methods suffer from quadratic complexity, posing challenges for handling long sequences and making them unsuitable for resource-constrained devices. Recently, SSMs [5, 9-11, 17, 34] have emerged as a prominent research focus due to their linear complexity and powerful information aggregation ability. In the field of point cloud representation learning, several Mamba-based methods [15, 21, 47] have been proposed. PointMamba [21] employs space-filling curves to serialize point clouds, establishing a simple yet effective base" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.351 + ], + "angle": 0, + "content": "line. PCM [47] introduces multiple serialization strategies and a bidirectional scanning mechanism to capture comprehensive point cloud structures, but also introduces larger computational overhead. Mamba3D [15] also leverages a bidirectional scanning mechanism and proposes a local norm pooling operation to enhance local geometric features. However, these methods introduce structural distortions when converting 3D points into 1D sequences, limiting model performance, especially in tasks requiring fine-grained geometry like part segmentation. Moreover, they neglect the impact of varying input lengths between pretraining and downstream tasks. Unlike these methods, we propose StruMamba3D, which leverages states as proxies to maintain spatial dependencies among points. Additionally, we introduce a sequence-length adaptive strategy to ensure that the states retain long-sequence memory as the sequence length grows in downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.361, + 0.606, + 0.377 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.386, + 0.909, + 0.448 + ], + "angle": 0, + "content": "Below, we review the previous SSMs (Sec. 3.1), followed by an overview of our StruMamba3D (Sec. 3.2). Then, we introduce the details of Structural SSM Block (Sec. 3.3) and Sequence Length-adaptive Strategy (Sec. 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.455, + 0.656, + 0.47 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.908, + 0.509 + ], + "angle": 0, + "content": "State Space Model (SSM) is a continuous system that maps the input \\(x(t)\\) to \\(y(t)\\) through implicit states \\(h(t)\\in \\mathbb{R}^{M}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.516, + 0.907, + 0.535 + ], + "angle": 0, + "content": "\\[\nh ^ {\\prime} (t) = \\mathbf {A} h (t) + \\mathbf {B} x (t), y (t) = \\mathbf {C} h ^ {\\prime} (t), \\qquad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.542, + 0.909, + 0.618 + ], + "angle": 0, + "content": "where the matrix \\(\\mathbf{A}\\) governs the evolution of states, while the matrices \\(\\mathbf{B}\\) and \\(\\mathbf{C}\\) are used for modeling the relationship between inputs and outputs through the latent states. To handle discrete-time sequence data, the zero-order hold discretization rule is used:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.624, + 0.905, + 0.662 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} h _ {t} = \\overline {{\\mathbf {A}}} h _ {t - 1} + \\overline {{\\mathbf {B}}} x _ {t}, y _ {t} = \\mathbf {C} h _ {t}, \\\\ \\overline {{\\mathbf {A}}} = \\exp (\\boldsymbol {\\Delta} \\mathbf {A}), \\overline {{\\mathbf {B}}} = (\\boldsymbol {\\Delta} \\mathbf {A}) ^ {- 1} (\\exp (\\boldsymbol {\\Delta} \\mathbf {A}) - \\mathbf {I}) \\boldsymbol {\\Delta} \\mathbf {B}, \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.67, + 0.909, + 0.777 + ], + "angle": 0, + "content": "where \\(\\Delta\\) represents the sampling interval between consecutive inputs \\(x_{t-1}\\) and \\(x_t\\). However, due to the linear time-invariant nature of SSM, the parameters \\((\\Delta, \\mathbf{A}, \\mathbf{B}, \\mathbf{C})\\) remain fixed across all time steps, which limits the expressive capacity of SSM. To overcome this limitation, Mamba [9] treats the parameters \\((\\Delta, \\mathbf{B}, \\mathbf{C})\\) as functions of input \\(x_t\\), transforming the SSM into a time-varying model:" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.785, + 0.907, + 0.803 + ], + "angle": 0, + "content": "\\[\nh _ {t} = \\phi_ {\\overline {{\\mathbf {A}}}} (x _ {t}) h _ {t - 1} + \\phi_ {\\overline {{\\mathbf {B}}}} (x _ {t}) x _ {t}, y _ {t} = \\phi_ {\\mathbf {C}} (x _ {t}) h _ {t}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.911, + 0.903 + ], + "angle": 0, + "content": "where \\(\\phi_{*}(x_{t})\\) denotes the linear projection of input \\(x_{t}\\). Additionally, Mamba introduces a hardware-aware scan algorithm to achieve near-linear complexity. Our StruMamba3D adopts the same hardware-aware scan algorithm, maintaining linear complexity while exhibiting strong structural modeling capabilities for point clouds." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28548" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.127, + 0.084, + 0.868, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.126, + 0.231, + 0.449, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.456, + 0.231, + 0.686, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.231, + 0.868, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.383, + 0.907, + 0.425 + ], + "angle": 0, + "content": "Figure 2. Overview of StruMamba3D. (a): Architecture of StruMamba3D. (b): Details of Structural SSM Block. We employ structural SSM to capture dependencies between spatial states and input points. Additionally, we integrate a lightweight convolution module to introduce local interactions among spatial states or input points. (c): Details of Structural SSM. (d): Details of Lightweight Convolution." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.439, + 0.202, + 0.453 + ], + "angle": 0, + "content": "3.2. Overview" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.46, + 0.484, + 0.718 + ], + "angle": 0, + "content": "Despite Mamba achieving remarkable results on sequential data, maintaining the spatial relationships among 3D points remains challenging. We aim to model the spatial relationships by latent states and design StruMamba3D for point cloud representation learning, as illustrated in Fig. 2(a). Given raw point cloud \\( P_{raw} \\in \\mathbb{R}^{N_0 \\times 3} \\), we obtain the input points \\( P_x \\in \\mathbb{R}^{N \\times 3} \\) using farthest point sampling (FPS) and \\( K \\)-nearest neighbor clustering (KNN). Then, we use a lightweight PointNet [27] to extract the input token embeddings \\( F_x \\in \\mathbb{R}^{N \\times D} \\). Meanwhile, we initialize the spatial states \\( P_h \\in \\mathbb{R}^{M \\times 3} \\) as the proxies for local structures from the raw point cloud. Finally, the spatial states and input token embeddings are fed into a feature encoder composed of multiple structural SSM blocks, where the spatial states act as a bridge to connect the inputs and outputs. Notably, we use the output spatial states from the previous structural SSM block as the input spatial states for the current block." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.298, + 0.743 + ], + "angle": 0, + "content": "3.3. Structural SSM Block" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The structural SSM block is designed to capture complex structural features by preserving spatial dependencies among points throughout SSM processing. As shown in Fig. 2(b), the block takes spatial states \\( F_{h} \\) and token embeddings \\( F_{x} \\) as input and produces updated spatial states \\( F_{h}^{\\prime} \\) and token embeddings \\( F_{x}^{\\prime} \\) as output. To clarify the details of structural SSM block, we describe each component in sequence: Spatial State Initialization, State-wise Update Mechanism, Structural SSM, and Lightweight Convolution. Spatial State Initialization. In original Mamba, latent" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.439, + 0.907, + 0.577 + ], + "angle": 0, + "content": "states do not contain geometric information, making it difficult to model the local structure of point clouds. To address this, we introduce positional attributes \\( P_{h} \\in \\mathbb{R}^{M \\times 3} \\) to the latent states, enabling each state to focus on distinct regions of the point cloud. Due to the sparsity of point clouds, we initialize state positions \\( P_{h} \\) based on raw points \\( P_{raw} \\) to ensure comprehensive coverage of point clouds. Specifically, we employ FPS and KNN clustering to segment the point cloud into groups \\( \\{\\mathcal{G}_{m}\\}_{m=1}^{M} \\), calculating the centroid:" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.586, + 0.907, + 0.623 + ], + "angle": 0, + "content": "\\[\nP _ {h} ^ {m} = \\frac {1}{\\left| \\mathcal {G} _ {m} \\right|} \\sum_ {P _ {i} \\in \\mathcal {G} _ {m}} P _ {i}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.633, + 0.907, + 0.694 + ], + "angle": 0, + "content": "where \\( P_{h}^{m} \\) represents the positional attribute of the \\( m \\)-th state. We refer to states with positional attributes as spatial states and then use linear mapping \\( \\phi_{h} \\) to embed positional attributes as the initial state features." + }, + { + "type": "equation", + "bbox": [ + 0.626, + 0.705, + 0.905, + 0.723 + ], + "angle": 0, + "content": "\\[\nF _ {h} = \\phi_ {h} \\left(P _ {h}\\right) \\in \\mathbb {R} ^ {M \\times D}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.765 + ], + "angle": 0, + "content": "The state features \\( F_{h} \\) are then used as the initial latent states in SSM and updated along with input token embeddings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.902 + ], + "angle": 0, + "content": "State-wise Update. As illustrated in Eq. (3), the state update and propagation equations of the original Mamba rely solely on input tokens, disregarding the spatial relationships between input points and spatial states. This limitation makes it challenging for Mamba to effectively model the structural information of point clouds. To address this, we modify the state update and propagation equations to explicitly model the spatial dependencies between states and inputs. Specifically, we first calculate the relative offsets" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28549" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.455, + 0.108 + ], + "angle": 0, + "content": "between the input points \\( P_{x} \\) and the state positions \\( P_{h} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.113, + 0.483, + 0.131 + ], + "angle": 0, + "content": "\\[\n\\triangle P _ {i} ^ {m} = P _ {x} ^ {i} - P _ {h} ^ {m}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.136, + 0.484, + 0.243 + ], + "angle": 0, + "content": "where \\(\\triangle P_i^m\\) denotes the relative offset between the \\(i\\)-th input point and the \\(m\\)-th state. Next, we map these offsets to the SSM parameters \\((\\mathbf{B}, \\mathbf{C})\\), which govern state update and propagation. It is important to note that considering only spatial relationships is insufficient, as the significance of information varies across different points. Thus, we design the following state update and propagation mechanism:" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.249, + 0.483, + 0.266 + ], + "angle": 0, + "content": "\\[\n\\left(\\mathbf {B} _ {i} ^ {m}, \\mathbf {C} _ {i} ^ {m}\\right) = \\phi \\left(x _ {i}\\right) + \\operatorname {M L P} \\left(\\triangle P _ {i} ^ {m}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.272, + 0.484, + 0.363 + ], + "angle": 0, + "content": "where MLP represents a multi-layer perceptron, \\(\\phi\\) is a linear projection and \\(x_{i}\\) denotes the feature of the \\(i\\)-th input point. Based on the modified parameters, the spatial states can selectively update features using points within the same regions, while the input points can acquire local structural information from the states, as shown in Fig. 2(c)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.363, + 0.484, + 0.56 + ], + "angle": 0, + "content": "Structural SSM. By leveraging spatial states and state-wise updates, we effectively model the structural information of point clouds. However, due to unidirectional scanning mechanism, bidirectional information exchange between input points cannot be achieved within a single forward pass. To address this, we adopt a bidirectional scanning mechanism inspired by VisionMamba [25]. Specifically, we apply both forward and backward structural SSMs, with the backward SSM processing input points in reverse order. Both use spatial states \\(\\hat{F}_h\\) as initial states \\(\\hat{F}_h^0\\), generating updated features for input points and spatial states. Finally, we fuse these outputs using a linear layer \\(\\phi_o\\) to obtain the final features \\(F_x'\\) and \\(F_h'\\) for input points and spatial states." + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.566, + 0.483, + 0.585 + ], + "angle": 0, + "content": "\\[\nF _ {x} ^ {\\prime}, F _ {h} ^ {\\prime} = \\phi_ {o} (\\mathrm {S S M} _ {f} (\\hat {F} _ {x}, \\hat {F} _ {h}) + \\mathrm {S S M} _ {b} (\\hat {F} _ {x}, \\hat {F} _ {h})), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.591, + 0.483, + 0.605 + ], + "angle": 0, + "content": "where \\( f \\) and \\( b \\) denote the forward and backward processes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.605, + 0.484, + 0.803 + ], + "angle": 0, + "content": "Lightweight Convolution. The proposed structural SSM effectively models the dependencies between spatial states \\( P_{h} \\) and input points \\( P_{x} \\). However, we observe that the spatial states \\( P_{h} \\) in SSMs remain isolated, lacking direct interaction with each other. To address this, we introduce a lightweight convolution module, as shown in Fig. 2(d). Inspired by graph convolution [35], this module aggregates features from neighboring states based on relative position information. Specifically, for each spatial state \\( P_{h}^{m} \\), we first identify its \\( k \\)-nearest neighbors \\( \\mathcal{N}(m) \\) and compute the relative positions \\( \\triangle P_{h}^{mj} \\), \\( j \\in \\mathcal{N}(m) \\). We then generate attention weights \\( \\{w_{mj}\\} \\) for each neighbor using a linear layer \\( \\phi_{w} \\), followed by a softmax operation:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.807, + 0.483, + 0.827 + ], + "angle": 0, + "content": "\\[\nw _ {h} ^ {m j} = \\operatorname {s o f t m a x} \\left(\\phi_ {w} \\left(\\triangle P _ {h} ^ {m j}, P _ {h} ^ {m}\\right)\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.833, + 0.484, + 0.865 + ], + "angle": 0, + "content": "The output state feature \\(\\hat{F}_h^m\\) for each state \\(P_h^m\\) is obtained through weighted aggregation:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.87, + 0.483, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\hat {F} _ {h} ^ {m} = \\phi_ {c} \\left(\\sum_ {j \\in \\mathcal {N} (m)} w _ {h} ^ {m j} F _ {h} ^ {m j}\\right), \\tag {10}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.085, + 0.9, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.266, + 0.907, + 0.335 + ], + "angle": 0, + "content": "Figure 3. Overview of the pre-training framework. We pretrain the StruMamba3D (student model) using the MPM task and the spatial state consistency loss. The teacher model has the same structure as the student model, but its parameters are updated using the exponential moving average strategy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.347, + 0.907, + 0.455 + ], + "angle": 0, + "content": "where \\( F_{h}^{mj} \\) denotes the neighboring state features and \\( \\phi_c \\) is a linear layer. By extending the receptive field of spatial states, this lightweight convolution enhances their ability to capture global semantic information. Furthermore, we replace the causal 1D convolution in the original Mamba block with this module, making it better suited for processing input points with complex geometric structures." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.463, + 0.822, + 0.48 + ], + "angle": 0, + "content": "3.4. Sequence Length-adaptive Strategy" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.906, + 0.545 + ], + "angle": 0, + "content": "To address the issue of insufficient long-sequence memory in downstream tasks, we propose a sequence length-adaptive strategy that integrates an adaptive state update mechanism and spatial state consistency loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.546, + 0.907, + 0.772 + ], + "angle": 0, + "content": "Adaptive State Update Mechanism. According to Eq. (2), the parameter sampling interval \\(\\Delta\\) controls the frequency of state updates by influencing the parameters \\((\\overline{\\mathbf{A}},\\overline{\\mathbf{B}})\\) in SSMs. A larger \\(\\Delta\\) causes the model to update state features more frequently, whereas a smaller \\(\\Delta\\) allows the model to maintain its current state with minimal changes. To preserve memory for longer sequences in downstream tasks, we propose an adaptive state update mechanism. This mechanism adjusts the sampling interval \\(\\Delta\\) according to the input sequence length, allowing the model to maintain a consistent total sampling time \\(\\Delta_{all} = \\sum_{i=1}^{N}\\Delta_{i}\\) across different sequence lengths. Specifically, we design a learnable parameter \\(\\tau\\), which regulates the total sampling time \\(\\Delta_{all}\\) for sequences of varying lengths. The sampling interval \\(\\Delta_{i}\\) for each token in the inputs is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.65, + 0.781, + 0.907, + 0.819 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Delta} _ {i} = \\frac {\\tau \\times \\boldsymbol {\\Delta} _ {i}}{\\sum_ {i = 1} ^ {N} \\boldsymbol {\\Delta} _ {i}}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.871 + ], + "angle": 0, + "content": "Based on the adaptive state update mechanism, the model can maintain a consistent total sampling time \\(\\Delta_{all} = \\tau\\) across different sequence lengths." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Spatial State Consistency Loss. To further ensure consistency in updated state features across different input se" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "28550" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.077, + 0.907, + 0.106 + ], + "angle": 0, + "content": "Table 1. Shape Classification on ScanObjectNN and ModelNet40 Datasets. For ScanObjectNN, we report the classification accuracy(%) over the three subsets: OBJ-BG, OBJ-ONLY, and PB-T50-RS. And \\(\\dagger\\) indicates using simple rotational augmentation [7] for fine-tuning." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.107, + 0.899, + 0.482 + ], + "angle": 0, + "content": "
MethodBackboneParam. (M)FLOPs (G)ScanObjectNNMODELNET40
OBJ-BGOBJ-ONLYPB-T50-RSw/o Votingw/ Voting
Supervised learning only
PointNet [27]-3.50.573.379.268.089.2-
PointNet++ [28]-1.51.782.384.377.990.7-
PointCNN [19]-0.6-86.185.578.592.2-
DGCNN [35]-1.82.482.886.278.192.9-
PRANet [8]-2.3---81.093.7-
PointNeXt [30]-1.43.6--87.794.0-
PointMLP [24]-12.631.4--85.494.5-
DeLA [2]-5.31.5--88.694.0-
PCM [47]-34.245.0--88.193.4-
Pre-training using single-modal information
PointBERT [42]Transformer22.14.887.4388.1283.0792.793.2
MaskPoint [22]Transformer22.14.889.3088.1084.30-93.8
PointM2AE [44]Transformer12.77.991.2288.8186.4392.993.4
PointMAE† [25]Transformer22.14.892.7791.2289.0492.793.8
PointGPT-S† [3]Transformer29.25.793.3992.4389.1793.394.0
PointMamba† [21]Mamba12.33.194.3292.6089.3193.694.1
Mamba3D† [15]Mamba16.93.993.1292.0892.0594.795.1
Ours†Structural SSM15.84.095.1893.6392.7595.195.4
Pre-training using cross-modal information
ACT† [7]Transformer22.14.893.2991.9188.2193.794.0
Joint-MAE [14]Transformer22.1-90.9488.8686.07-94.0
I2P-MAE† [46]Transformer15.3-94.1591.5790.1193.794.1
ReCon† [29]Transformer43.65.395.1893.2990.6394.594.7
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.501, + 0.486, + 0.685 + ], + "angle": 0, + "content": "quence lengths, we introduce a spatial state consistency loss to the pre-training task. As shown in Fig. 3, we present an overview of the pre-training framework. Following the previous works [15, 21, 25], we reconstruct the points in the masked regions and use the reconstruction loss \\(\\mathcal{L}_{cd}\\) to guide the model to learn discriminative features from the visible points. For the spatial state consistency loss, we first use a teacher model to update the spatial states with the full input tokens \\(P_{x}^{f}\\in \\mathbb{R}^{N_{f}\\times 3}\\). The state features \\(F_h^{\\prime f}\\) output by the teacher model serve as pseudo-labels. We then enforce a consistency loss \\(\\mathcal{L}_{ssc}\\) between the updated state features of the student model \\(F_h^{\\prime v}\\) and the pseudo-labels \\(F_h^{\\prime f}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.706, + 0.483, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s s c} = \\operatorname {S m o o t h} \\mathrm {L} 1 \\left(F _ {h} ^ {\\prime v}, F _ {h} ^ {\\prime f}\\right). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.747, + 0.485, + 0.825 + ], + "angle": 0, + "content": "The spatial state consistency loss \\(\\mathcal{L}_{ssc}\\) not only enhances the robustness of our StruMamba3D to varying input sequence lengths but also improves the ability to infer the structural information of the complete point cloud from incomplete point clouds. The final pre-training loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.847, + 0.483, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {c d} + \\lambda \\times \\mathcal {L} _ {s s c}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.473, + 0.902 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is the weight of the spatial state consistency loss." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.5, + 0.72, + 0.516 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.525, + 0.731, + 0.541 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.548, + 0.907, + 0.76 + ], + "angle": 0, + "content": "Following the previous works [15, 21, 25], we employ \\( T = 12 \\) structural SSM blocks, each with a feature dimension of \\( D = 384 \\). The number of spatial states \\( M \\) is set to 16. For the lightweight convolution in the structural SSM block, we set the number of neighbors to 4 for spatial states and 8 for input points. Following previous works [25, 42, 44], we pre-train the proposed model on ShapeNet dataset [1], which contains 52472 unique 3D models across 55 common object categories. Each input raw point cloud, containing \\( N_0 = 1024 \\) points, is divided into 64 patches with each consisting of 32 points. A random masking ratio of 0.6 is applied, with the decoder composed of 4 standard Mamba blocks, and the \\( \\lambda \\) in the pre-training loss is set to 2. More details can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.772, + 0.819, + 0.788 + ], + "angle": 0, + "content": "4.2. Comparison on Downstream Tasks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.839 + ], + "angle": 0, + "content": "We present the experimental results of our StruMamba3D across various downstream tasks and compare them with previous state-of-the-art (SOTA) methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Shape classification. ScanObjectNN [41] is a real-world dataset containing roughly 15000 objects across 15 categories, some of which have cluttered backgrounds or partial scans. For this dataset, we apply rotation as a data aug-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28551" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.074, + 0.483, + 0.102 + ], + "angle": 0, + "content": "Table 2. Part Segmentation on ShapeNetPart. We report the mean IoU for all classes \\((\\mathrm{mIoU}_c)\\) and all instances \\((\\mathrm{mIoU}_i)\\)." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.104, + 0.472, + 0.337 + ], + "angle": 0, + "content": "
MethodArchitecturemIoUcmIoUi
Supervised learning only
PointNet [27]Single-scale80.483.7
PointNet++ [28]Multi-scale81.985.1
APES [36]Multi-scale83.785.8
DeLA [2]Multi-scale85.887.0
PCM [47]Multi-scale85.387.0
Pre-training using single-modal information
MaskPoint [22]Single-scale84.686.0
PointBERT [42]Single-scale84.185.6
PointMAE [25]Single-scale84.286.1
PointM2AE [44]Multi-scale84.986.5
PointGPT-S [3]Single-scale84.186.2
PointMamba [21]Single-scale84.486.2
Mamba3D [15]Single-scale83.685.6
OursSingle-scale85.086.7
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.341, + 0.483, + 0.644 + ], + "angle": 0, + "content": "mentation strategy [7] and use 2048 points as input. ModelNet40 [39] is a synthetic dataset comprising of 12311 clean 3D CAD models from 40 categories. For this dataset, we use 1024 points as input and apply scale and translation augmentations [27]. As shown in Tab. 1, we conduct classification experiments on both datasets. Compared to transformer-based methods, our model achieves significant performance improvements across multiple metrics on both datasets. Even when compared to the best PointGPTS [3], our model outperforms it by \\(3.58\\%\\) on ScanObjectNN and \\(1.8\\%\\) on ModelNet40. Additionally, on the ScanObjectNN dataset, StruMamba3D surpasses PointMamba [21] by \\(0.86\\%\\), \\(1.03\\%\\), and \\(3.44\\%\\) across the three splits, while also outperforming Mamba3D [15] by \\(2.06\\%\\), \\(1.55\\%\\), and \\(0.70\\%\\). Furthermore, on the ModelNet40 dataset, StruMamba3D exceeds PointMamba by \\(1.5\\%\\) and outperforms Mamba3D by \\(0.4\\%\\). Overall, these results demonstrate that StruMamba3D outperforms all existing unimodal architectures, achieving multiple SOTA results and highlighting the effectiveness of structural modeling within Mamba." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Part Segmentation on ShapeNetPart. ShapeNetPart [41] dataset comprises 16880 models spanning 16 shape categories and annotated with 50 part labels. Compared to classification tasks, part segmentation poses a greater challenge, as it requires assigning a label to each point, making it highly dependent on the intrinsic structural information of point clouds. Tab. 8 presents our part segmentation results for all classes and instances on the ShapeNetPart dataset. For single-scale models, our method achieves substantial performance improvements, surpassing Mamba3D [15] by \\(1.4\\%\\) in \\(\\mathrm{mIoU}_c\\) and \\(1.2\\%\\) in \\(\\mathrm{mIoU}_i\\), and outperforming PointMamba [21] by \\(0.6\\%\\) in \\(\\mathrm{mIoU}_c\\) and \\(0.5\\%\\) in \\(\\mathrm{mIoU}_i\\). Notably, Mamba3D without a serialization strategy performs poorly, falling behind the baseline PointMAE [25], whereas PointMamba, which incorporates serialization, exhibits improved performance. This observation suggests the crucial role of serialization in enhancing structural mod" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.074, + 0.905, + 0.103 + ], + "angle": 0, + "content": "Table 3. Few-Shot Classification on ModelNet40. We report overall accuracy and standard deviation without a voting strategy." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.106, + 0.899, + 0.253 + ], + "angle": 0, + "content": "
Method5-way10-way
10-shot20-shot10-shot20-shot
PointBERT [42]94.6±3.693.9±3.186.4±5.491.3±4.6
MaskPoint [22]95.0±3.797.2±1.791.4±4.092.7±5.1
PointMAE [25]96.3±2.597.8±1.892.6±4.193.4±3.5
PointM2AE [44]96.8±1.898.3±1.492.3±4.595.0±3.0
PointGPT-S [3]96.8±2.098.6±1.192.6±4.695.2±3.4
PointMamba [21]96.9±2.099.0±1.193.0±4.495.6±3.2
Mamba3D [15]96.4±2.298.2±1.292.4±4.195.2±2.9
Ours97.5±2.399.1±1.493.5±3.796.1±3.5
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.907, + 0.53 + ], + "angle": 0, + "content": "eling within Mamba-based architectures. In contrast, our StruMamba3D leverages spatial states to effectively capture structural information, thereby eliminating the need for serialization and achieving \\(85.0\\%\\) in \\(\\mathrm{mIoU}_c\\) and \\(86.7\\%\\) in \\(\\mathrm{mIoU}_i\\). Besides, StruMamba3D even outperforms the multi-scale self-supervised method PointM2AE [44]. Although multi-scale approaches typically yield better performance in part segmentation tasks, they often come with significantly higher computational costs (e.g., PCM [47] requires 45.0G FLOPs). In this work, we focus on designing an efficient and effective Mamba-based framework for structural modeling in point clouds. While incorporating multi-scale designs could further improve part segmentation performance, such strategies may introduce heuristic dependencies. Overall, our StruMamba3D represents a more robust and efficient Mamba-based framework for point cloud structural modeling, achieving the highest performance among all single-scale models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.531, + 0.909, + 0.743 + ], + "angle": 0, + "content": "Few-shot Learning on ModelNet40. To evaluate the effectiveness of the proposed methods with limited fine-tuning data, we conduct few-shot classification experiments on the ModelNet40 dataset. Following the standard procedure [25, 42], we perform 10 independent experiments for each setting and report the mean accuracy and standard deviation. As shown in Tab. 3, our method demonstrates a significant improvement over Mamba-based methods, outperforming PointMamba [21] by \\(0.6\\%\\) in 5-way-10-shot and \\(0.5\\%\\) in 10-way-10-shot settings, and surpassing Mamba3D [15] by \\(1.1\\%\\) and \\(0.9\\%\\), respectively. These results indicate that our method can efficiently adapt to downstream tasks with longer input sequences, even under the constraint of limited training data." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.756, + 0.667, + 0.772 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.779, + 0.905, + 0.824 + ], + "angle": 0, + "content": "To explore the architectural design, we perform ablation studies on the PB-T50-RS split of ScanObjectNN dataset, as well as ModelNet40 and ShapeNetPart datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Effect of Designed Modules. As shown in Tab. 4, we conduct an ablation study to evaluate the effectiveness of two core modules: the Structural SSM block and the sequence length-adaptive strategy. The baseline model is built upon the standard Mamba [9] block, following the same" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28552" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.076, + 0.483, + 0.104 + ], + "angle": 0, + "content": "Table 4. Effect of the designed modules. 'SLAS' indicates the sequence length-adaptive strategy used in pre-training phase." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.106, + 0.472, + 0.186 + ], + "angle": 0, + "content": "
Structural SSM BlockSLASOverall AccuracymIoUc
ScanNNMN40SNPart
XX87.2391.8681.56
✓X92.0994.4584.49
✓✓92.7595.0684.96
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.19, + 0.483, + 0.219 + ], + "angle": 0, + "content": "Table 5. Ablation on Structural SSM Block. The baseline uses the standard Mamba block to replace the Structural SSM Block." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.221, + 0.472, + 0.328 + ], + "angle": 0, + "content": "
MethodOverall AccuracymIoUc
ScanNNMN40SNPart
baseline88.2492.5082.08
w/ Structural SSM91.7893.9284.15
w/ Lightweight Conv. for h92.2294.6584.62
w/ Lightweight Conv. for x92.4094.8184.77
w/ Bidirectional Scanning92.7595.0684.96
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.338, + 0.483, + 0.489 + ], + "angle": 0, + "content": "architecture as PointMAE [25]. Compared to this baseline, the Structural SSM block significantly enhances model performance, yielding improvements of \\(4.86\\%\\) on ScanObjectNN, \\(2.59\\%\\) on ModelNet40, and \\(2.93\\%\\) on ShapeNetPart, demonstrating its effectiveness in modeling spatial dependencies and capturing richer structural information in point clouds. Furthermore, integrating the sequence length-adaptive strategy further enhances performance by improving the generalization to varying input lengths, thereby ensuring optimal performance across downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.491, + 0.483, + 0.672 + ], + "angle": 0, + "content": "Effect of Structural SSM Block. As shown in Tab. 5, we further validate the effectiveness of individual components within the Structural SSM block. Starting from the baseline model with the standard Mamba block, we incrementally incorporate key modules from the Structural SSM block. The most significant performance gains stem from the Structural SSM and the lightweight convolution designed for spatial states, demonstrating that modeling structural information through hidden states in SSM is crucial. Moveover, the lightweight convolution for input points and the bidirectional scanning mechanism also enhance the performance across downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Effect of Structural SSM. Structural SSM is a crucial component of our method, as it effectively captures the structural information of point clouds through spatial states. To further analyze the impact of spatial state initialization and state-wise update strategy, we conduct ablation studies as shown in Tab. 6. Although spatial state initialization provides the states with positional information, the improvement is minimal, as Eq. (2) does not consider spatial relationships between input points and states. When integrating the state-wise update strategy, the model achieves a significant performance boost, with improvements of \\(1.42\\%\\) on ScanObjectNN, \\(0.94\\%\\) on ModelNet40, and \\(1.09\\%\\) on ShapeNetPart. The results indicate that incorporating spatial relationships to control state information updates and propagation within the SSM is essential for effectively mod" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.076, + 0.905, + 0.132 + ], + "angle": 0, + "content": "Table 6. Ablation on Structural SSM. 'SS' represents the use of spatial states in the SSM. \\( \\phi (x) \\) denotes the SSM parameters depending on the point feature, while \\( \\mathrm{MLP}(\\triangle P) \\) represents the SSM parameters depending on the spatial relationship." + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.134, + 0.885, + 0.227 + ], + "angle": 0, + "content": "
φ(x)SSMLP(△P)Overall AccuracymIoUc
ScanNNMN40SNPart
✓XX90.9493.8483.62
✓✓X91.3394.1283.87
✓✓✓92.7595.0684.96
X✓✓91.1294.2584.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.23, + 0.905, + 0.271 + ], + "angle": 0, + "content": "Table 7. Ablation on sequence length-adaptive strategy. 'ASUM' indicates the adaptive state update mechanism, and \\( \\lambda \\) ’ indicates the weight of the spatial state consistency loss \\( {\\mathcal{L}}_{ssc} \\) ." + }, + { + "type": "table", + "bbox": [ + 0.539, + 0.273, + 0.88, + 0.395 + ], + "angle": 0, + "content": "
ASUMLsscOverall Accuracy
ScanNNMN40SNPart
XX92.0994.4584.49
✓X92.2694.5784.56
✓λ = 192.4794.9484.81
✓λ = 292.7595.0684.96
✓λ = 592.5794.8984.77
Xλ = 292.1694.6584.56
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.905, + 0.478 + ], + "angle": 0, + "content": "eling structural information. Moreover, we observe that relying solely on spatial relationships to generate SSM parameters results in performance degradation. This is likely due to the varying importance of point features, which cannot be fully captured by spatial position alone." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.479, + 0.906, + 0.706 + ], + "angle": 0, + "content": "Effect of Sequence Length-Adaptive Strategy. The sequence length-adaptive strategy consists of two key components: the adaptive state update mechanism and the spatial state consistency loss. Tab. 7 presents an ablation study on these components. When applied individually, both the adaptive state update mechanism and the spatial state consistency loss yield only marginal performance gains. However, when used together, they lead to a significant improvement, demonstrating their complementary effects. The adaptive state update mechanism ensures a consistent total sampling time across different input lengths, while the spatial state consistency loss ensures the consistency of spatial state features. In addition, we perform an ablation study on the weight \\(\\lambda\\) of \\(\\mathcal{L}_{ssc}\\) and observe that the model achieves optimal performance when \\(\\lambda\\) is set to 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.715, + 0.634, + 0.73 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We identify two key issues limiting Mamba potential: 3D point adjacency distortion during SSM processing and inadequate long-sequence memory retention in downstream tasks. To address these issues, we propose StruMamba3D with two core components. The structural SSM block maintains spatial dependencies during state selection and propagation, while the sequence length-adaptive strategy reduces the sensitivity of Mamba-based models to varying input lengths. Our method achieves substantial improvements across four downstream tasks, with ablation studies confirming the effectiveness of each module." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28553" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.262, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.109, + 0.483, + 0.14 + ], + "angle": 0, + "content": "This work is supported by the National Defense Basic Scientific Research program (JCKY2022911B002)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.154, + 0.188, + 0.17 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.18, + 0.483, + 0.247 + ], + "angle": 0, + "content": "[1] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.25, + 0.483, + 0.291 + ], + "angle": 0, + "content": "[2] Binjie Chen, Yunzhou Xia, Yu Zang, Cheng Wang, and Jonathan Li. Decoupled local aggregation for point cloud learning, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.293, + 0.483, + 0.348 + ], + "angle": 0, + "content": "[3] Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. Pointgpt: Auto-regressively generative pre-training from point clouds. arXiv preprint arXiv:2305.11487, 2023. 3, 6, 7, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.35, + 0.483, + 0.419 + ], + "angle": 0, + "content": "[4] Siheng Chen, Baoan Liu, Chen Feng, Carlos Vallespi-Gonzalez, and Carl Wellington. 3d point cloud processing and learning for autonomous driving: Impacting map creation, localization, and perception. IEEE Signal Processing Magazine, 38(1):68-86, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.421, + 0.483, + 0.462 + ], + "angle": 0, + "content": "[5] Tri Dao and Albert Gu. Transformers are ssms: Generalized models and efficient algorithms through structured state space duality. arXiv preprint arXiv:2405.21060, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.464, + 0.483, + 0.532 + ], + "angle": 0, + "content": "[6] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT (1), pages 4171-4186. Association for Computational Linguistics, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.534, + 0.483, + 0.616 + ], + "angle": 0, + "content": "[7] Runpei Dong, Zekun Qi, Linfeng Zhang, Junbo Zhang, Jianjian Sun, Zheng Ge, Li Yi, and Kaisheng Ma. Autoencoders as cross-modal teachers: Can pretrained 2d image transformers help 3d representation learning? In The Eleventh International Conference on Learning Representations (ICLR), 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.619, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[8] Deng-Ping Fan, Ge-Peng Ji, Tao Zhou, Geng Chen, Huazhu Fu, Jianbing Shen, and Ling Shao. Pranet: Parallel reverse attention network for polyp segmentation. In International conference on medical image computing and computer-assisted intervention, pages 263-273. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.689, + 0.483, + 0.73 + ], + "angle": 0, + "content": "[9] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2, 3, 7, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.733, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[10] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[11] Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, and Christopher Ré. Combining recurrent, convolutional, and continuous-time models with linear state space layers. Advances in neural information processing systems, 34:572-585, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[12] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7:187-199, 2021. 2,3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.18, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[13] Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3d point clouds: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(12):4338-4364, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.906, + 0.203 + ], + "angle": 0, + "content": "[14] Ziyu Guo, Renrui Zhang, Longtian Qiu, Xianzhi Li, and Pheng-Ann Heng. Joint-mae: 2d-3d joint masked autoencoders for 3d point cloud pre-training. arXiv preprint arXiv:2302.14007, 2023.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.906, + 0.274 + ], + "angle": 0, + "content": "[15] Xu Han, Yuan Tang, Zhaoxuan Wang, and Xianzhi Li. Mamba3d: Enhancing local features for 3d point cloud analysis via state space model. In Proceedings of the 32nd ACM International Conference on Multimedia, page 4995-5004. Association for Computing Machinery, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.906, + 0.342 + ], + "angle": 0, + "content": "[16] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.906, + 0.383 + ], + "angle": 0, + "content": "[17] Rudolph Emil Kalman. A new approach to linear filtering and prediction problems. Journal of Basic Engineering, 1960. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.906, + 0.426 + ], + "angle": 0, + "content": "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.906, + 0.482 + ], + "angle": 0, + "content": "[19] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. Advances in neural information processing systems, 31, 2018. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.906, + 0.551 + ], + "angle": 0, + "content": "[20] Ying Li, Lingfei Ma, Zilong Zhong, Fei Liu, Michael A Chapman, Dongpu Cao, and Jonathan Li. Deep learning for lidar point clouds in autonomous driving: A review. IEEE Transactions on Neural Networks and Learning Systems, 32 (8):3412-3432, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.906, + 0.621 + ], + "angle": 0, + "content": "[21] Dingkang Liang, Xin Zhou, Wei Xu, Xingkui Zhu, Zhikang Zou, Xiaqing Ye, Xiao Tan, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. In Advances in Neural Information Processing Systems, 2024. 2, 3, 6, 7, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.906, + 0.678 + ], + "angle": 0, + "content": "[22] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In European Conference on Computer Vision, pages 657-675. Springer, 2022. 2, 3, 6, 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.906, + 0.718 + ], + "angle": 0, + "content": "[23] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.72, + 0.906, + 0.774 + ], + "angle": 0, + "content": "[24] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. arXiv preprint arXiv:2202.07123, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.906, + 0.844 + ], + "angle": 0, + "content": "[25] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In European conference on computer vision, pages 604-621. Springer, 2022. 2, 3, 5, 6, 7, 8, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.899 + ], + "angle": 0, + "content": "[26] François Pomerleau, Francis Colas, Roland Siegwart, et al. A review of point cloud registration algorithms for mobile robotics. Foundations and Trends® in Robotics, 4(1):1-104, 2015. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28554" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[27] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[28] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 1, 2, 3, 6, 7, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[29] Zekun Qi, Runpei Dong, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Contrast with reconstruct: Contrastive 3d representation learning guided by generative pretraining. In International Conference on Machine Learning (ICML), 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[30] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. Advances in neural information processing systems, 35:23192-23204, 2022. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.444 + ], + "angle": 0, + "content": "[31] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.446, + 0.482, + 0.515 + ], + "angle": 0, + "content": "[32] Chuxin Wang, Jiacheng Deng, Jianfeng He, Tianzhu Zhang, Zhe Zhang, and Yongdong Zhang. Long-short range adaptive transformer with dynamic sampling for 3d object detection. IEEE Transactions on Circuits and Systems for Video Technology, 33(12):7616-7629, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.517, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[33] Chuxin Wang, Yixin Zha, Jianfeng He, Wenfei Yang, and Tianzhu Zhang. Rethinking masked representation learning for 3d point cloud understanding. IEEE Transactions on Image Processing, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[34] ChuXin Wang, Wenfei Yang, Xiang Liu, and Tianzhu Zhang. State space model meets transformer: A new paradigm for 3d object detection. In The Thirteenth International Conference on Learning Representations, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[35] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[36] Chengzhi Wu, Junwei Zheng, Julius Pfrommer, and Jürgen Beyerer. Attention-based point cloud edge sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5333-5343, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[37] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. Advances in Neural Information Processing Systems, 35:33330-33342, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[38] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4840-4851, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[39] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[40] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 574-591. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[41] Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Transactions on Graphics (ToG), 35(6):1-12, 2016. 6, 7, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.907, + 0.359 + ], + "angle": 0, + "content": "[42] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19313-19322, 2022. 2, 3, 6, 7, 11, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.905, + 0.415 + ], + "angle": 0, + "content": "[43] Yixin Zha, Chuxin Wang, Wenfei Yang, and Tianzhu Zhang. Exploring semantic masked autoencoder for self-supervised point cloud understanding. arXiv preprint arXiv:2506.21957, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.486 + ], + "angle": 0, + "content": "[44] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training. Advances in neural information processing systems, 35:27061-27074, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.488, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[45] Renrui Zhang, Liuhui Wang, Yu Qiao, Peng Gao, and Hongsheng Li. Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21769-21780, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.559, + 0.905, + 0.627 + ], + "angle": 0, + "content": "[46] Renrui Zhang, Liuhui Wang, Yu Qiao, Peng Gao, and Hongsheng Li. Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21769-21780, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.629, + 0.905, + 0.683 + ], + "angle": 0, + "content": "[47] Tao Zhang, Xiangtai Li, Haobo Yuan, Shunping Ji, and Shuicheng Yan. Point cloud mamba: Point cloud learning via state space model. arXiv preprint arXiv:2403.00762, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.686, + 0.905, + 0.753 + ], + "angle": 0, + "content": "[48] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3d features on any point-cloud. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10252-10263, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.756, + 0.907, + 0.811 + ], + "angle": 0, + "content": "[49] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.811 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28555" + } + ] +] \ No newline at end of file diff --git a/2025/StruMamba3D_ Exploring Structural Mamba for Self-supervised Point Cloud Representation Learning/9ae9c283-762a-4256-b4de-303cb73e1e2c_origin.pdf b/2025/StruMamba3D_ Exploring Structural Mamba for Self-supervised Point Cloud Representation Learning/9ae9c283-762a-4256-b4de-303cb73e1e2c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7795dc2859f7c2bc13757501bd1d2de76e959d1e --- /dev/null +++ b/2025/StruMamba3D_ Exploring Structural Mamba for Self-supervised Point Cloud Representation Learning/9ae9c283-762a-4256-b4de-303cb73e1e2c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5505edf782c58482be4b537a2f7aecfa6846f423d8723662211e3294a72d2c36 +size 2631508 diff --git a/2025/StruMamba3D_ Exploring Structural Mamba for Self-supervised Point Cloud Representation Learning/full.md b/2025/StruMamba3D_ Exploring Structural Mamba for Self-supervised Point Cloud Representation Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..31ec7eded20f4fe1da31086b1efe88d2d87849fb --- /dev/null +++ b/2025/StruMamba3D_ Exploring Structural Mamba for Self-supervised Point Cloud Representation Learning/full.md @@ -0,0 +1,332 @@ +# StruMamba3D: Exploring Structural Mamba for Self-supervised Point Cloud Representation Learning + +Chuxin Wang $^{1,2}$ , Yixin Zha $^{1}$ , Wenfei Yang $^{1,2,*}$ , Tianzhu Zhang $^{1,2}$ + +1University of Science and Technology of China + +$^{2}$ National Key Laboratory of Deep Space Exploration, Deep Space Exploration Laboratory + +{wcx0602, zyxcn} $@$ mail.ustc.edu.cn, {yangwf, tzzhang} $@$ ustc.edu.cn + +![](images/def97985916b92a5e1a231264460e313ce670297dd366eb00c35093729f4b123.jpg) +Attention Block +Quadratic complexity + +![](images/ec38826214ea8c80c9857e9afe51a06461c9b4434540c5512b79339503d86a76.jpg) +Selective SSM Block +Linear complexity + +![](images/f0f68ca372b516ee5050d2041570c1e75aadd8c1e070747bbd3cc6887d5f41cc.jpg) +Serialized Points +Destroy adjacency + +![](images/87c767c30a776d8c18b83d5e0e01ab1ff193f3fcca4699d474205b54b42b36ec.jpg) +Longer Input +Fail to retain long-sequence memory + +![](images/96ad2b99223270751631ee35700572645c798045e6a66095b184272867ae0b0a.jpg) + +# +□ Point Feature +□ + +![](images/1c70f27e6a058199ddbbfc8109c778f96e2b2e9ee2b29abc5fd02d304a0643cb.jpg) +(a) Transformer-based Paradigm +PointMAE +PointGPT +Mamba3D +PointMaml +Our +# +adigm (b) Mamba-based Paradigm +Figure 1. (a): Transformer-based paradigm uses the attention block with quadratic complexity to model the dependencies between points. (b): Mamba-based paradigm uses the selective SSM with linear complexity. However, serialized points destroy the adjacency of 3D points, and the pre-trained selection mechanism fails to retain long-sequence memory. (c): Our StruMamba3D paradigm uses the structural SSM to maintain the spatial dependencies among points and the sequence length-adaptive strategy to retain long-sequence memory. (d): Our StruMamba3D significantly outperforms existing Transformer-based and Mamba-based methods. + +# Abstract + +Recently, Mamba-based methods have demonstrated impressive performance in point cloud representation learning by leveraging State Space Model (SSM) with the efficient context modeling ability and linear complexity. However, these methods still face two key issues that limit the potential of SSM: Destroying the adjacency of 3D points during SSM processing and failing to retain long-sequence memory as the input length increases in downstream tasks. To address these issues, we propose StruMamba3D, a novel paradigm for self-supervised point cloud representation learning. It enjoys several merits. First, we design spatial states and use them as proxies to preserve spatial dependencies among points. Second, we enhance the SSM with a state-wise update strategy and incorporate a lightweight convolution to facilitate interactions between spatial states for efficient structure modeling. Third, our method reduces the sensitivity of pre-trained Mamba-based models to varying input lengths by introducing a sequence length-adaptive + +strategy. Experimental results across four downstream tasks showcase the superior performance of our method. In addition, our method attains the SOTA $95.1\%$ accuracy on ModelNet40 and $92.75\%$ accuracy on the most challenging split of ScanObjectNN without voting strategy. + +# 1. Introduction + +Point cloud representation learning aims to extract geometric and semantic features from point clouds to support downstream tasks such as classification, segmentation, and detection. As a foundational task in Computer Vision (CV), it is crucial for various real-world applications, including autonomous driving [4, 20], AR [13], and robotics [26, 32]. However, unlike images or text with structured arrangements, point clouds are inherently unordered and sparse, presenting challenges for designing effective models. + +To solve this problem, numerous point cloud representation learning methods have been proposed, which can be broadly divided into three categories according to model architectures: Point-based methods [19, 27, 28, 30, 31, + +35], Transformer-based methods [12, 22, 25, 42-44], and Mamba-based methods [15, 21, 47]. Point-based methods directly process point clouds by modeling the relationships between individual points, such as through set abstraction [27, 28] and graph convolution [19, 35]. These methods offer a simple and computationally efficient design. However, their limited receptive field hinders capturing global information, constraining complex structure modeling. Transformer-based methods leverage attention mechanisms to capture long-range dependencies, expanding the receptive field. Moreover, inspired by self-supervised learning paradigms in NLP [6] and CV [16], many transformer-based methods [22, 25, 42, 44] utilize large unlabeled point cloud datasets for pre-training, achieving impressive results in downstream tasks. However, these methods have limited scalability due to the quadratic complexity of the attention mechanism, as shown in Fig. 1(a). To overcome this limitation, Mamba-based methods [15, 21, 47] first convert the 3D points into a 1D sequence and then utilize selective State Space Model (SSM) [9] with linear complexity to extract point features. Leveraging the robust information aggregation ability of selective SSM, Mamba-based methods have attained promising performance. + +Although existing Mamba-based methods [15, 21, 47] have achieved comparable performance to transformer-based methods, two key issues limit the potential of SSM: (1) Distort the adjacency of points. As shown in Fig. 1(b), although methods [21, 47] use serialization strategies to align input points, they cannot ensure that spatially adjacent points remain neighbors in the 1D sequence. Unlike text and images, points inherently lack context information and rely on local structures as base feature units. This spatial distortion hampers Mamba to model fine-grained structure information. (2) Fail to retain long-sequence memory in downstream tasks. Mamba-based methods [15, 21] use Masked Point Modeling (MPM) [25, 42] as a pre-training task, which creates a significant discrepancy in input sequence lengths between pre-training and downstream tasks. Mamba leverages a selection mechanism to inductively learn new information while selectively forgetting outdated knowledge. However, the selection mechanism does not account for variations in input sequence length. Consequently, when Mamba is pre-trained on short sequences, it tends to update states more frequently. As the input sequence length increases in downstream tasks, frequent state updates lead to Mamba struggling to retain long-sequence memory, as shown in Fig. 1(b). The deficiency in long-sequence memory adversely affects the long-range semantics modeling, resulting in suboptimal performance. + +To unleash the potential of Mamba for point cloud representation learning, we explore solutions to address the above two issues. (1) How to maintain spatial dependencies among points? Maintaining spatial relationships + +among 3D points within a 1D input sequence is challenging. Mamba efficiently aggregates contextual features from the 1D input sequence via hidden states. Is it possible to design a structural Mamba that models spatial relationships among points through hidden states? (2) How to retain long-sequence memory for longer inputs? When handling longer input sequences, states need to retain memory for an extended period to effectively capture the contextual information across the entire sequence. If the model adjusts the frequency of state updates when the sequence length changes, it can preserve memory for longer sequences. + +To achieve the above goals, we propose a novel StruMamba3D paradigm for self-supervised point cloud representation learning, which consists of two core components: structural SSM block and sequence length-adaptive strategy. In structural SSM block, we leverage states in SSM to model the spatial relationships among points. We begin by assigning positional attributes to the states, using them as proxies to capture the local structure of point clouds. Next, we propose a state-wise update strategy that incorporates the positional relationship between input points and spatial states during the state update and propagation process. Based on spatial states and state-wise update strategy, we can maintain spatial dependencies among points during the SSM process. To address the limitation of standard Mamba ignoring inter-state interactions, we introduce a lightweight convolution module for spatial states. In sequence length-adaptive strategy, the adaptive state update mechanism is first employed to adjust the state update frequency based on sequence length, ensuring that the states retain long-sequence memory as the sequence length grows. Additionally, we introduce a spatial state consistency loss in the pre-training task, which enforces consistency in the updated states across inputs of varying sequence lengths. The spatial state consistency loss enhances the robustness of the Mamba-based pre-trained model to varying input lengths. + +To sum up, the main contributions of this work are: (1) We propose the novel StruMamba3D for self-supervised point cloud representation learning, which is the first to model the structure information by latent states in SSM. (2) We design the structural SSM to maintain spatial dependencies among points during state information selection and propagation, while the sequence length-adaptive strategy reduces the sensitivity of pre-trained Mamba-based models to varying input lengths. (3) Experimental results on four downstream tasks showcase the superior performance of our method. Additionally, comprehensive ablation studies validate the effectiveness of our designs. + +# 2. Related Work + +We briefly review methods for point cloud representation learning and self-supervised representation learning. + +Point Cloud Representation Learning. To effectively + +leverage the sparse, unordered and irregular point cloud, various point cloud representation learning methods have been proposed, which can be broadly categorized into three groups: Point-based methods [19, 27, 28, 30, 31, 35], Transformer-based methods [12, 22, 25, 33, 42, 44], and Mamba-based methods [15, 21, 47]. Point-based methods directly process raw point clouds to learn geometric features. For example, PointNet [27] employs MLPs to each point independently, followed by global max pooling to aggregate features. PointNet++ [28] extends PointNet by introducing a hierarchical architecture for multi-scale feature extraction. Subsequent methods [19, 24, 30, 31, 35] explore convolutional techniques for point clouds. DGCNN [35] leverages graph convolution to model structural information among neighboring points, while KPCov [31] introduces a convolution method based on dynamic kernel points to improve feature learning. To capture long-range dependencies, methods [12, 37, 38, 49] have explored transformer architectures for point clouds. PCT [12] applies global attention to points, while Point Transformer [49] introduces local attention to capture local geometry features. Recent variants [37, 38] further refine the architectures to improve performance and efficiency. These supervised methods [12, 30, 31, 47, 49] have demonstrated outstanding performance in specific tasks. However, these methods are inherently limited by the specific data domains on which they are trained, leading to poor generalization to other domains. + +Self-supervised Representation Learning. To solve the above issues and fully utilize a large number of unlabeled data, self-supervised representation learning methods [3, 22, 25, 40, 42, 44, 48] have emerged. Some contrastive learning-based methods [40, 48] train models to learn distinguishing features by differentiating between positive and negative samples. For instance, Point-Contrast [40] leverages feature consistency across different views of the same point cloud. Inspired by the success of BERT [6] and MAE [16], many transformer-based methods [3, 22, 25, 42, 44] adopt MPM task for pre-training, encouraging models to infer masked patches from visible ones. PointBERT [42] predicts discrete tokens for masked points, while PointMAE [25] reconstructs masked point clouds. Additionally, recent studies [29, 45] integrate cross-modal information, such as images and text, to enhance pre-training. However, transformer-based methods suffer from quadratic complexity, posing challenges for handling long sequences and making them unsuitable for resource-constrained devices. Recently, SSMs [5, 9-11, 17, 34] have emerged as a prominent research focus due to their linear complexity and powerful information aggregation ability. In the field of point cloud representation learning, several Mamba-based methods [15, 21, 47] have been proposed. PointMamba [21] employs space-filling curves to serialize point clouds, establishing a simple yet effective base + +line. PCM [47] introduces multiple serialization strategies and a bidirectional scanning mechanism to capture comprehensive point cloud structures, but also introduces larger computational overhead. Mamba3D [15] also leverages a bidirectional scanning mechanism and proposes a local norm pooling operation to enhance local geometric features. However, these methods introduce structural distortions when converting 3D points into 1D sequences, limiting model performance, especially in tasks requiring fine-grained geometry like part segmentation. Moreover, they neglect the impact of varying input lengths between pretraining and downstream tasks. Unlike these methods, we propose StruMamba3D, which leverages states as proxies to maintain spatial dependencies among points. Additionally, we introduce a sequence-length adaptive strategy to ensure that the states retain long-sequence memory as the sequence length grows in downstream tasks. + +# 3. Method + +Below, we review the previous SSMs (Sec. 3.1), followed by an overview of our StruMamba3D (Sec. 3.2). Then, we introduce the details of Structural SSM Block (Sec. 3.3) and Sequence Length-adaptive Strategy (Sec. 3.4). + +# 3.1. Preliminaries + +State Space Model (SSM) is a continuous system that maps the input $x(t)$ to $y(t)$ through implicit states $h(t)\in \mathbb{R}^{M}$ : + +$$ +h ^ {\prime} (t) = \mathbf {A} h (t) + \mathbf {B} x (t), y (t) = \mathbf {C} h ^ {\prime} (t), \qquad (1) +$$ + +where the matrix $\mathbf{A}$ governs the evolution of states, while the matrices $\mathbf{B}$ and $\mathbf{C}$ are used for modeling the relationship between inputs and outputs through the latent states. To handle discrete-time sequence data, the zero-order hold discretization rule is used: + +$$ +\begin{array}{l} h _ {t} = \overline {{\mathbf {A}}} h _ {t - 1} + \overline {{\mathbf {B}}} x _ {t}, y _ {t} = \mathbf {C} h _ {t}, \\ \overline {{\mathbf {A}}} = \exp (\boldsymbol {\Delta} \mathbf {A}), \overline {{\mathbf {B}}} = (\boldsymbol {\Delta} \mathbf {A}) ^ {- 1} (\exp (\boldsymbol {\Delta} \mathbf {A}) - \mathbf {I}) \boldsymbol {\Delta} \mathbf {B}, \tag {2} \\ \end{array} +$$ + +where $\Delta$ represents the sampling interval between consecutive inputs $x_{t-1}$ and $x_t$ . However, due to the linear time-invariant nature of SSM, the parameters $(\Delta, \mathbf{A}, \mathbf{B}, \mathbf{C})$ remain fixed across all time steps, which limits the expressive capacity of SSM. To overcome this limitation, Mamba [9] treats the parameters $(\Delta, \mathbf{B}, \mathbf{C})$ as functions of input $x_t$ , transforming the SSM into a time-varying model: + +$$ +h _ {t} = \phi_ {\overline {{\mathbf {A}}}} (x _ {t}) h _ {t - 1} + \phi_ {\overline {{\mathbf {B}}}} (x _ {t}) x _ {t}, y _ {t} = \phi_ {\mathbf {C}} (x _ {t}) h _ {t}, \tag {3} +$$ + +where $\phi_{*}(x_{t})$ denotes the linear projection of input $x_{t}$ . Additionally, Mamba introduces a hardware-aware scan algorithm to achieve near-linear complexity. Our StruMamba3D adopts the same hardware-aware scan algorithm, maintaining linear complexity while exhibiting strong structural modeling capabilities for point clouds. + +![](images/2599e73e448e8ac0f9e9a2c95cf179ca547207d5e4ff9169e95adf4f49c84a6d.jpg) + +![](images/245557cd6982fb0bcd1a4f50126a0e1e15b54018feef192feebb36d817beaba7.jpg) +Figure 2. Overview of StruMamba3D. (a): Architecture of StruMamba3D. (b): Details of Structural SSM Block. We employ structural SSM to capture dependencies between spatial states and input points. Additionally, we integrate a lightweight convolution module to introduce local interactions among spatial states or input points. (c): Details of Structural SSM. (d): Details of Lightweight Convolution. + +![](images/ff4d03cf7c1c8516d51673698a018b37c25efae3b3fb8a450e8440a44f665fd4.jpg) + +![](images/21a8c32d4dce1a734c0437ced7e56527159a55eb8ba35a8d215373a21b2dc06e.jpg) + +# 3.2. Overview + +Despite Mamba achieving remarkable results on sequential data, maintaining the spatial relationships among 3D points remains challenging. We aim to model the spatial relationships by latent states and design StruMamba3D for point cloud representation learning, as illustrated in Fig. 2(a). Given raw point cloud $P_{raw} \in \mathbb{R}^{N_0 \times 3}$ , we obtain the input points $P_x \in \mathbb{R}^{N \times 3}$ using farthest point sampling (FPS) and $K$ -nearest neighbor clustering (KNN). Then, we use a lightweight PointNet [27] to extract the input token embeddings $F_x \in \mathbb{R}^{N \times D}$ . Meanwhile, we initialize the spatial states $P_h \in \mathbb{R}^{M \times 3}$ as the proxies for local structures from the raw point cloud. Finally, the spatial states and input token embeddings are fed into a feature encoder composed of multiple structural SSM blocks, where the spatial states act as a bridge to connect the inputs and outputs. Notably, we use the output spatial states from the previous structural SSM block as the input spatial states for the current block. + +# 3.3. Structural SSM Block + +The structural SSM block is designed to capture complex structural features by preserving spatial dependencies among points throughout SSM processing. As shown in Fig. 2(b), the block takes spatial states $F_{h}$ and token embeddings $F_{x}$ as input and produces updated spatial states $F_{h}^{\prime}$ and token embeddings $F_{x}^{\prime}$ as output. To clarify the details of structural SSM block, we describe each component in sequence: Spatial State Initialization, State-wise Update Mechanism, Structural SSM, and Lightweight Convolution. Spatial State Initialization. In original Mamba, latent + +states do not contain geometric information, making it difficult to model the local structure of point clouds. To address this, we introduce positional attributes $P_{h} \in \mathbb{R}^{M \times 3}$ to the latent states, enabling each state to focus on distinct regions of the point cloud. Due to the sparsity of point clouds, we initialize state positions $P_{h}$ based on raw points $P_{raw}$ to ensure comprehensive coverage of point clouds. Specifically, we employ FPS and KNN clustering to segment the point cloud into groups $\{\mathcal{G}_{m}\}_{m=1}^{M}$ , calculating the centroid: + +$$ +P _ {h} ^ {m} = \frac {1}{\left| \mathcal {G} _ {m} \right|} \sum_ {P _ {i} \in \mathcal {G} _ {m}} P _ {i}, \tag {4} +$$ + +where $P_{h}^{m}$ represents the positional attribute of the $m$ -th state. We refer to states with positional attributes as spatial states and then use linear mapping $\phi_{h}$ to embed positional attributes as the initial state features. + +$$ +F _ {h} = \phi_ {h} \left(P _ {h}\right) \in \mathbb {R} ^ {M \times D}. \tag {5} +$$ + +The state features $F_{h}$ are then used as the initial latent states in SSM and updated along with input token embeddings. + +State-wise Update. As illustrated in Eq. (3), the state update and propagation equations of the original Mamba rely solely on input tokens, disregarding the spatial relationships between input points and spatial states. This limitation makes it challenging for Mamba to effectively model the structural information of point clouds. To address this, we modify the state update and propagation equations to explicitly model the spatial dependencies between states and inputs. Specifically, we first calculate the relative offsets + +between the input points $P_{x}$ and the state positions $P_{h}$ : + +$$ +\triangle P _ {i} ^ {m} = P _ {x} ^ {i} - P _ {h} ^ {m}, \tag {6} +$$ + +where $\triangle P_i^m$ denotes the relative offset between the $i$ -th input point and the $m$ -th state. Next, we map these offsets to the SSM parameters $(\mathbf{B}, \mathbf{C})$ , which govern state update and propagation. It is important to note that considering only spatial relationships is insufficient, as the significance of information varies across different points. Thus, we design the following state update and propagation mechanism: + +$$ +\left(\mathbf {B} _ {i} ^ {m}, \mathbf {C} _ {i} ^ {m}\right) = \phi \left(x _ {i}\right) + \operatorname {M L P} \left(\triangle P _ {i} ^ {m}\right), \tag {7} +$$ + +where MLP represents a multi-layer perceptron, $\phi$ is a linear projection and $x_{i}$ denotes the feature of the $i$ -th input point. Based on the modified parameters, the spatial states can selectively update features using points within the same regions, while the input points can acquire local structural information from the states, as shown in Fig. 2(c). + +Structural SSM. By leveraging spatial states and state-wise updates, we effectively model the structural information of point clouds. However, due to unidirectional scanning mechanism, bidirectional information exchange between input points cannot be achieved within a single forward pass. To address this, we adopt a bidirectional scanning mechanism inspired by VisionMamba [25]. Specifically, we apply both forward and backward structural SSMs, with the backward SSM processing input points in reverse order. Both use spatial states $\hat{F}_h$ as initial states $\hat{F}_h^0$ , generating updated features for input points and spatial states. Finally, we fuse these outputs using a linear layer $\phi_o$ to obtain the final features $F_x'$ and $F_h'$ for input points and spatial states. + +$$ +F _ {x} ^ {\prime}, F _ {h} ^ {\prime} = \phi_ {o} (\mathrm {S S M} _ {f} (\hat {F} _ {x}, \hat {F} _ {h}) + \mathrm {S S M} _ {b} (\hat {F} _ {x}, \hat {F} _ {h})), \tag {8} +$$ + +where $f$ and $b$ denote the forward and backward processes. + +Lightweight Convolution. The proposed structural SSM effectively models the dependencies between spatial states $P_{h}$ and input points $P_{x}$ . However, we observe that the spatial states $P_{h}$ in SSMs remain isolated, lacking direct interaction with each other. To address this, we introduce a lightweight convolution module, as shown in Fig. 2(d). Inspired by graph convolution [35], this module aggregates features from neighboring states based on relative position information. Specifically, for each spatial state $P_{h}^{m}$ , we first identify its $k$ -nearest neighbors $\mathcal{N}(m)$ and compute the relative positions $\triangle P_{h}^{mj}$ , $j \in \mathcal{N}(m)$ . We then generate attention weights $\{w_{mj}\}$ for each neighbor using a linear layer $\phi_{w}$ , followed by a softmax operation: + +$$ +w _ {h} ^ {m j} = \operatorname {s o f t m a x} \left(\phi_ {w} \left(\triangle P _ {h} ^ {m j}, P _ {h} ^ {m}\right)\right). \tag {9} +$$ + +The output state feature $\hat{F}_h^m$ for each state $P_h^m$ is obtained through weighted aggregation: + +$$ +\hat {F} _ {h} ^ {m} = \phi_ {c} \left(\sum_ {j \in \mathcal {N} (m)} w _ {h} ^ {m j} F _ {h} ^ {m j}\right), \tag {10} +$$ + +![](images/a664a242dbbb3fd499a1477eae917724483c199649eb0fd4f42f01d624e87f4a.jpg) +Figure 3. Overview of the pre-training framework. We pretrain the StruMamba3D (student model) using the MPM task and the spatial state consistency loss. The teacher model has the same structure as the student model, but its parameters are updated using the exponential moving average strategy. + +where $F_{h}^{mj}$ denotes the neighboring state features and $\phi_c$ is a linear layer. By extending the receptive field of spatial states, this lightweight convolution enhances their ability to capture global semantic information. Furthermore, we replace the causal 1D convolution in the original Mamba block with this module, making it better suited for processing input points with complex geometric structures. + +# 3.4. Sequence Length-adaptive Strategy + +To address the issue of insufficient long-sequence memory in downstream tasks, we propose a sequence length-adaptive strategy that integrates an adaptive state update mechanism and spatial state consistency loss. + +Adaptive State Update Mechanism. According to Eq. (2), the parameter sampling interval $\Delta$ controls the frequency of state updates by influencing the parameters $(\overline{\mathbf{A}},\overline{\mathbf{B}})$ in SSMs. A larger $\Delta$ causes the model to update state features more frequently, whereas a smaller $\Delta$ allows the model to maintain its current state with minimal changes. To preserve memory for longer sequences in downstream tasks, we propose an adaptive state update mechanism. This mechanism adjusts the sampling interval $\Delta$ according to the input sequence length, allowing the model to maintain a consistent total sampling time $\Delta_{all} = \sum_{i=1}^{N}\Delta_{i}$ across different sequence lengths. Specifically, we design a learnable parameter $\tau$ , which regulates the total sampling time $\Delta_{all}$ for sequences of varying lengths. The sampling interval $\Delta_{i}$ for each token in the inputs is calculated as follows: + +$$ +\boldsymbol {\Delta} _ {i} = \frac {\tau \times \boldsymbol {\Delta} _ {i}}{\sum_ {i = 1} ^ {N} \boldsymbol {\Delta} _ {i}}. \tag {11} +$$ + +Based on the adaptive state update mechanism, the model can maintain a consistent total sampling time $\Delta_{all} = \tau$ across different sequence lengths. + +Spatial State Consistency Loss. To further ensure consistency in updated state features across different input se + +Table 1. Shape Classification on ScanObjectNN and ModelNet40 Datasets. For ScanObjectNN, we report the classification accuracy(%) over the three subsets: OBJ-BG, OBJ-ONLY, and PB-T50-RS. And $\dagger$ indicates using simple rotational augmentation [7] for fine-tuning. + +
MethodBackboneParam. (M)FLOPs (G)ScanObjectNNMODELNET40
OBJ-BGOBJ-ONLYPB-T50-RSw/o Votingw/ Voting
Supervised learning only
PointNet [27]-3.50.573.379.268.089.2-
PointNet++ [28]-1.51.782.384.377.990.7-
PointCNN [19]-0.6-86.185.578.592.2-
DGCNN [35]-1.82.482.886.278.192.9-
PRANet [8]-2.3---81.093.7-
PointNeXt [30]-1.43.6--87.794.0-
PointMLP [24]-12.631.4--85.494.5-
DeLA [2]-5.31.5--88.694.0-
PCM [47]-34.245.0--88.193.4-
Pre-training using single-modal information
PointBERT [42]Transformer22.14.887.4388.1283.0792.793.2
MaskPoint [22]Transformer22.14.889.3088.1084.30-93.8
PointM2AE [44]Transformer12.77.991.2288.8186.4392.993.4
PointMAE† [25]Transformer22.14.892.7791.2289.0492.793.8
PointGPT-S† [3]Transformer29.25.793.3992.4389.1793.394.0
PointMamba† [21]Mamba12.33.194.3292.6089.3193.694.1
Mamba3D† [15]Mamba16.93.993.1292.0892.0594.795.1
Ours†Structural SSM15.84.095.1893.6392.7595.195.4
Pre-training using cross-modal information
ACT† [7]Transformer22.14.893.2991.9188.2193.794.0
Joint-MAE [14]Transformer22.1-90.9488.8686.07-94.0
I2P-MAE† [46]Transformer15.3-94.1591.5790.1193.794.1
ReCon† [29]Transformer43.65.395.1893.2990.6394.594.7
+ +quence lengths, we introduce a spatial state consistency loss to the pre-training task. As shown in Fig. 3, we present an overview of the pre-training framework. Following the previous works [15, 21, 25], we reconstruct the points in the masked regions and use the reconstruction loss $\mathcal{L}_{cd}$ to guide the model to learn discriminative features from the visible points. For the spatial state consistency loss, we first use a teacher model to update the spatial states with the full input tokens $P_{x}^{f}\in \mathbb{R}^{N_{f}\times 3}$ . The state features $F_h^{\prime f}$ output by the teacher model serve as pseudo-labels. We then enforce a consistency loss $\mathcal{L}_{ssc}$ between the updated state features of the student model $F_h^{\prime v}$ and the pseudo-labels $F_h^{\prime f}$ : + +$$ +\mathcal {L} _ {s s c} = \operatorname {S m o o t h} \mathrm {L} 1 \left(F _ {h} ^ {\prime v}, F _ {h} ^ {\prime f}\right). \tag {12} +$$ + +The spatial state consistency loss $\mathcal{L}_{ssc}$ not only enhances the robustness of our StruMamba3D to varying input sequence lengths but also improves the ability to infer the structural information of the complete point cloud from incomplete point clouds. The final pre-training loss is: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {c d} + \lambda \times \mathcal {L} _ {s s c}, \tag {13} +$$ + +where $\lambda$ is the weight of the spatial state consistency loss. + +# 4. Experimental Results + +# 4.1. Implementation Details + +Following the previous works [15, 21, 25], we employ $T = 12$ structural SSM blocks, each with a feature dimension of $D = 384$ . The number of spatial states $M$ is set to 16. For the lightweight convolution in the structural SSM block, we set the number of neighbors to 4 for spatial states and 8 for input points. Following previous works [25, 42, 44], we pre-train the proposed model on ShapeNet dataset [1], which contains 52472 unique 3D models across 55 common object categories. Each input raw point cloud, containing $N_0 = 1024$ points, is divided into 64 patches with each consisting of 32 points. A random masking ratio of 0.6 is applied, with the decoder composed of 4 standard Mamba blocks, and the $\lambda$ in the pre-training loss is set to 2. More details can be found in the supplementary material. + +# 4.2. Comparison on Downstream Tasks + +We present the experimental results of our StruMamba3D across various downstream tasks and compare them with previous state-of-the-art (SOTA) methods. + +Shape classification. ScanObjectNN [41] is a real-world dataset containing roughly 15000 objects across 15 categories, some of which have cluttered backgrounds or partial scans. For this dataset, we apply rotation as a data aug- + +Table 2. Part Segmentation on ShapeNetPart. We report the mean IoU for all classes $(\mathrm{mIoU}_c)$ and all instances $(\mathrm{mIoU}_i)$ . + +
MethodArchitecturemIoUcmIoUi
Supervised learning only
PointNet [27]Single-scale80.483.7
PointNet++ [28]Multi-scale81.985.1
APES [36]Multi-scale83.785.8
DeLA [2]Multi-scale85.887.0
PCM [47]Multi-scale85.387.0
Pre-training using single-modal information
MaskPoint [22]Single-scale84.686.0
PointBERT [42]Single-scale84.185.6
PointMAE [25]Single-scale84.286.1
PointM2AE [44]Multi-scale84.986.5
PointGPT-S [3]Single-scale84.186.2
PointMamba [21]Single-scale84.486.2
Mamba3D [15]Single-scale83.685.6
OursSingle-scale85.086.7
+ +mentation strategy [7] and use 2048 points as input. ModelNet40 [39] is a synthetic dataset comprising of 12311 clean 3D CAD models from 40 categories. For this dataset, we use 1024 points as input and apply scale and translation augmentations [27]. As shown in Tab. 1, we conduct classification experiments on both datasets. Compared to transformer-based methods, our model achieves significant performance improvements across multiple metrics on both datasets. Even when compared to the best PointGPTS [3], our model outperforms it by $3.58\%$ on ScanObjectNN and $1.8\%$ on ModelNet40. Additionally, on the ScanObjectNN dataset, StruMamba3D surpasses PointMamba [21] by $0.86\%$ , $1.03\%$ , and $3.44\%$ across the three splits, while also outperforming Mamba3D [15] by $2.06\%$ , $1.55\%$ , and $0.70\%$ . Furthermore, on the ModelNet40 dataset, StruMamba3D exceeds PointMamba by $1.5\%$ and outperforms Mamba3D by $0.4\%$ . Overall, these results demonstrate that StruMamba3D outperforms all existing unimodal architectures, achieving multiple SOTA results and highlighting the effectiveness of structural modeling within Mamba. + +Part Segmentation on ShapeNetPart. ShapeNetPart [41] dataset comprises 16880 models spanning 16 shape categories and annotated with 50 part labels. Compared to classification tasks, part segmentation poses a greater challenge, as it requires assigning a label to each point, making it highly dependent on the intrinsic structural information of point clouds. Tab. 8 presents our part segmentation results for all classes and instances on the ShapeNetPart dataset. For single-scale models, our method achieves substantial performance improvements, surpassing Mamba3D [15] by $1.4\%$ in $\mathrm{mIoU}_c$ and $1.2\%$ in $\mathrm{mIoU}_i$ , and outperforming PointMamba [21] by $0.6\%$ in $\mathrm{mIoU}_c$ and $0.5\%$ in $\mathrm{mIoU}_i$ . Notably, Mamba3D without a serialization strategy performs poorly, falling behind the baseline PointMAE [25], whereas PointMamba, which incorporates serialization, exhibits improved performance. This observation suggests the crucial role of serialization in enhancing structural mod + +Table 3. Few-Shot Classification on ModelNet40. We report overall accuracy and standard deviation without a voting strategy. + +
Method5-way10-way
10-shot20-shot10-shot20-shot
PointBERT [42]94.6±3.693.9±3.186.4±5.491.3±4.6
MaskPoint [22]95.0±3.797.2±1.791.4±4.092.7±5.1
PointMAE [25]96.3±2.597.8±1.892.6±4.193.4±3.5
PointM2AE [44]96.8±1.898.3±1.492.3±4.595.0±3.0
PointGPT-S [3]96.8±2.098.6±1.192.6±4.695.2±3.4
PointMamba [21]96.9±2.099.0±1.193.0±4.495.6±3.2
Mamba3D [15]96.4±2.298.2±1.292.4±4.195.2±2.9
Ours97.5±2.399.1±1.493.5±3.796.1±3.5
+ +eling within Mamba-based architectures. In contrast, our StruMamba3D leverages spatial states to effectively capture structural information, thereby eliminating the need for serialization and achieving $85.0\%$ in $\mathrm{mIoU}_c$ and $86.7\%$ in $\mathrm{mIoU}_i$ . Besides, StruMamba3D even outperforms the multi-scale self-supervised method PointM2AE [44]. Although multi-scale approaches typically yield better performance in part segmentation tasks, they often come with significantly higher computational costs (e.g., PCM [47] requires 45.0G FLOPs). In this work, we focus on designing an efficient and effective Mamba-based framework for structural modeling in point clouds. While incorporating multi-scale designs could further improve part segmentation performance, such strategies may introduce heuristic dependencies. Overall, our StruMamba3D represents a more robust and efficient Mamba-based framework for point cloud structural modeling, achieving the highest performance among all single-scale models. + +Few-shot Learning on ModelNet40. To evaluate the effectiveness of the proposed methods with limited fine-tuning data, we conduct few-shot classification experiments on the ModelNet40 dataset. Following the standard procedure [25, 42], we perform 10 independent experiments for each setting and report the mean accuracy and standard deviation. As shown in Tab. 3, our method demonstrates a significant improvement over Mamba-based methods, outperforming PointMamba [21] by $0.6\%$ in 5-way-10-shot and $0.5\%$ in 10-way-10-shot settings, and surpassing Mamba3D [15] by $1.1\%$ and $0.9\%$ , respectively. These results indicate that our method can efficiently adapt to downstream tasks with longer input sequences, even under the constraint of limited training data. + +# 4.3. Ablation Study + +To explore the architectural design, we perform ablation studies on the PB-T50-RS split of ScanObjectNN dataset, as well as ModelNet40 and ShapeNetPart datasets. + +Effect of Designed Modules. As shown in Tab. 4, we conduct an ablation study to evaluate the effectiveness of two core modules: the Structural SSM block and the sequence length-adaptive strategy. The baseline model is built upon the standard Mamba [9] block, following the same + +Table 4. Effect of the designed modules. 'SLAS' indicates the sequence length-adaptive strategy used in pre-training phase. + +
Structural SSM BlockSLASOverall AccuracymIoUc
ScanNNMN40SNPart
XX87.2391.8681.56
✓X92.0994.4584.49
✓✓92.7595.0684.96
+ +Table 5. Ablation on Structural SSM Block. The baseline uses the standard Mamba block to replace the Structural SSM Block. + +
MethodOverall AccuracymIoUc
ScanNNMN40SNPart
baseline88.2492.5082.08
w/ Structural SSM91.7893.9284.15
w/ Lightweight Conv. for h92.2294.6584.62
w/ Lightweight Conv. for x92.4094.8184.77
w/ Bidirectional Scanning92.7595.0684.96
+ +architecture as PointMAE [25]. Compared to this baseline, the Structural SSM block significantly enhances model performance, yielding improvements of $4.86\%$ on ScanObjectNN, $2.59\%$ on ModelNet40, and $2.93\%$ on ShapeNetPart, demonstrating its effectiveness in modeling spatial dependencies and capturing richer structural information in point clouds. Furthermore, integrating the sequence length-adaptive strategy further enhances performance by improving the generalization to varying input lengths, thereby ensuring optimal performance across downstream tasks. + +Effect of Structural SSM Block. As shown in Tab. 5, we further validate the effectiveness of individual components within the Structural SSM block. Starting from the baseline model with the standard Mamba block, we incrementally incorporate key modules from the Structural SSM block. The most significant performance gains stem from the Structural SSM and the lightweight convolution designed for spatial states, demonstrating that modeling structural information through hidden states in SSM is crucial. Moveover, the lightweight convolution for input points and the bidirectional scanning mechanism also enhance the performance across downstream tasks. + +Effect of Structural SSM. Structural SSM is a crucial component of our method, as it effectively captures the structural information of point clouds through spatial states. To further analyze the impact of spatial state initialization and state-wise update strategy, we conduct ablation studies as shown in Tab. 6. Although spatial state initialization provides the states with positional information, the improvement is minimal, as Eq. (2) does not consider spatial relationships between input points and states. When integrating the state-wise update strategy, the model achieves a significant performance boost, with improvements of $1.42\%$ on ScanObjectNN, $0.94\%$ on ModelNet40, and $1.09\%$ on ShapeNetPart. The results indicate that incorporating spatial relationships to control state information updates and propagation within the SSM is essential for effectively mod + +Table 6. Ablation on Structural SSM. 'SS' represents the use of spatial states in the SSM. $\phi (x)$ denotes the SSM parameters depending on the point feature, while $\mathrm{MLP}(\triangle P)$ represents the SSM parameters depending on the spatial relationship. + +
φ(x)SSMLP(△P)Overall AccuracymIoUc
ScanNNMN40SNPart
✓XX90.9493.8483.62
✓✓X91.3394.1283.87
✓✓✓92.7595.0684.96
X✓✓91.1294.2584.02
+ +Table 7. Ablation on sequence length-adaptive strategy. 'ASUM' indicates the adaptive state update mechanism, and $\lambda$ ’ indicates the weight of the spatial state consistency loss ${\mathcal{L}}_{ssc}$ . + +
ASUMLsscOverall Accuracy
ScanNNMN40SNPart
XX92.0994.4584.49
✓X92.2694.5784.56
✓λ = 192.4794.9484.81
✓λ = 292.7595.0684.96
✓λ = 592.5794.8984.77
Xλ = 292.1694.6584.56
+ +eling structural information. Moreover, we observe that relying solely on spatial relationships to generate SSM parameters results in performance degradation. This is likely due to the varying importance of point features, which cannot be fully captured by spatial position alone. + +Effect of Sequence Length-Adaptive Strategy. The sequence length-adaptive strategy consists of two key components: the adaptive state update mechanism and the spatial state consistency loss. Tab. 7 presents an ablation study on these components. When applied individually, both the adaptive state update mechanism and the spatial state consistency loss yield only marginal performance gains. However, when used together, they lead to a significant improvement, demonstrating their complementary effects. The adaptive state update mechanism ensures a consistent total sampling time across different input lengths, while the spatial state consistency loss ensures the consistency of spatial state features. In addition, we perform an ablation study on the weight $\lambda$ of $\mathcal{L}_{ssc}$ and observe that the model achieves optimal performance when $\lambda$ is set to 2. + +# 5. Conclusion + +We identify two key issues limiting Mamba potential: 3D point adjacency distortion during SSM processing and inadequate long-sequence memory retention in downstream tasks. To address these issues, we propose StruMamba3D with two core components. The structural SSM block maintains spatial dependencies during state selection and propagation, while the sequence length-adaptive strategy reduces the sensitivity of Mamba-based models to varying input lengths. Our method achieves substantial improvements across four downstream tasks, with ablation studies confirming the effectiveness of each module. + +# 6. Acknowledgment + +This work is supported by the National Defense Basic Scientific Research program (JCKY2022911B002). + +# References + +[1] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6, 11 +[2] Binjie Chen, Yunzhou Xia, Yu Zang, Cheng Wang, and Jonathan Li. Decoupled local aggregation for point cloud learning, 2023. 6, 7 +[3] Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. Pointgpt: Auto-regressively generative pre-training from point clouds. arXiv preprint arXiv:2305.11487, 2023. 3, 6, 7, 12, 13 +[4] Siheng Chen, Baoan Liu, Chen Feng, Carlos Vallespi-Gonzalez, and Carl Wellington. 3d point cloud processing and learning for autonomous driving: Impacting map creation, localization, and perception. IEEE Signal Processing Magazine, 38(1):68-86, 2020. 1 +[5] Tri Dao and Albert Gu. Transformers are ssms: Generalized models and efficient algorithms through structured state space duality. arXiv preprint arXiv:2405.21060, 2024. 3 +[6] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT (1), pages 4171-4186. Association for Computational Linguistics, 2019. 2, 3 +[7] Runpei Dong, Zekun Qi, Linfeng Zhang, Junbo Zhang, Jianjian Sun, Zheng Ge, Li Yi, and Kaisheng Ma. Autoencoders as cross-modal teachers: Can pretrained 2d image transformers help 3d representation learning? In The Eleventh International Conference on Learning Representations (ICLR), 2023. 6, 7 +[8] Deng-Ping Fan, Ge-Peng Ji, Tao Zhou, Geng Chen, Huazhu Fu, Jianbing Shen, and Ling Shao. Pranet: Parallel reverse attention network for polyp segmentation. In International conference on medical image computing and computer-assisted intervention, pages 263-273. Springer, 2020. 6 +[9] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2, 3, 7, 14 +[10] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. +[11] Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, and Christopher Ré. Combining recurrent, convolutional, and continuous-time models with linear state space layers. Advances in neural information processing systems, 34:572-585, 2021. 3 +[12] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7:187-199, 2021. 2,3 + +[13] Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3d point clouds: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(12):4338-4364, 2020. 1 +[14] Ziyu Guo, Renrui Zhang, Longtian Qiu, Xianzhi Li, and Pheng-Ann Heng. Joint-mae: 2d-3d joint masked autoencoders for 3d point cloud pre-training. arXiv preprint arXiv:2302.14007, 2023.6 +[15] Xu Han, Yuan Tang, Zhaoxuan Wang, and Xianzhi Li. Mamba3d: Enhancing local features for 3d point cloud analysis via state space model. In Proceedings of the 32nd ACM International Conference on Multimedia, page 4995-5004. Association for Computing Machinery, 2024. 2, 3, 6, 7 +[16] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2, 3 +[17] Rudolph Emil Kalman. A new approach to linear filtering and prediction problems. Journal of Basic Engineering, 1960. 3 +[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 11, 12 +[19] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. Advances in neural information processing systems, 31, 2018. 1, 2, 3, 6 +[20] Ying Li, Lingfei Ma, Zilong Zhong, Fei Liu, Michael A Chapman, Dongpu Cao, and Jonathan Li. Deep learning for lidar point clouds in autonomous driving: A review. IEEE Transactions on Neural Networks and Learning Systems, 32 (8):3412-3432, 2020. 1 +[21] Dingkang Liang, Xin Zhou, Wei Xu, Xingkui Zhu, Zhikang Zou, Xiaqing Ye, Xiao Tan, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. In Advances in Neural Information Processing Systems, 2024. 2, 3, 6, 7, 12, 13 +[22] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In European Conference on Computer Vision, pages 657-675. Springer, 2022. 2, 3, 6, 7, 13 +[23] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 11, 12 +[24] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. arXiv preprint arXiv:2202.07123, 2022. 3, 6 +[25] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In European conference on computer vision, pages 604-621. Springer, 2022. 2, 3, 5, 6, 7, 8, 11 +[26] François Pomerleau, Francis Colas, Roland Siegwart, et al. A review of point cloud registration algorithms for mobile robotics. Foundations and Trends® in Robotics, 4(1):1-104, 2015. 1 + +[27] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 1, 2, 3, 4, 6, 7 +[28] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 1, 2, 3, 6, 7, 11 +[29] Zekun Qi, Runpei Dong, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Contrast with reconstruct: Contrastive 3d representation learning guided by generative pretraining. In International Conference on Machine Learning (ICML), 2023. 3, 6 +[30] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. Advances in neural information processing systems, 35:23192-23204, 2022. 1, 3, 6 +[31] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 1, 3 +[32] Chuxin Wang, Jiacheng Deng, Jianfeng He, Tianzhu Zhang, Zhe Zhang, and Yongdong Zhang. Long-short range adaptive transformer with dynamic sampling for 3d object detection. IEEE Transactions on Circuits and Systems for Video Technology, 33(12):7616-7629, 2023. 1 +[33] Chuxin Wang, Yixin Zha, Jianfeng He, Wenfei Yang, and Tianzhu Zhang. Rethinking masked representation learning for 3d point cloud understanding. IEEE Transactions on Image Processing, 2024. 3 +[34] ChuXin Wang, Wenfei Yang, Xiang Liu, and Tianzhu Zhang. State space model meets transformer: A new paradigm for 3d object detection. In The Thirteenth International Conference on Learning Representations, 2025. 3 +[35] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 2, 3, 5, 6 +[36] Chengzhi Wu, Junwei Zheng, Julius Pfrommer, and Jürgen Beyerer. Attention-based point cloud edge sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5333-5343, 2023. 7 +[37] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. Advances in Neural Information Processing Systems, 35:33330-33342, 2022. 3 +[38] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4840-4851, 2024. 3 +[39] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d + +shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 7 +[40] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 574-591. Springer, 2020. 3 +[41] Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Transactions on Graphics (ToG), 35(6):1-12, 2016. 6, 7, 12 +[42] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19313-19322, 2022. 2, 3, 6, 7, 11, 13 +[43] Yixin Zha, Chuxin Wang, Wenfei Yang, and Tianzhu Zhang. Exploring semantic masked autoencoder for self-supervised point cloud understanding. arXiv preprint arXiv:2506.21957, 2025. +[44] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training. Advances in neural information processing systems, 35:27061-27074, 2022. 2, 3, 6, 7 +[45] Renrui Zhang, Liuhui Wang, Yu Qiao, Peng Gao, and Hongsheng Li. Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21769-21780, 2023. 3 +[46] Renrui Zhang, Liuhui Wang, Yu Qiao, Peng Gao, and Hongsheng Li. Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21769-21780, 2023. 6 +[47] Tao Zhang, Xiangtai Li, Haobo Yuan, Shunping Ji, and Shuicheng Yan. Point cloud mamba: Point cloud learning via state space model. arXiv preprint arXiv:2403.00762, 2024. 2, 3, 6, 7 +[48] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3d features on any point-cloud. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10252-10263, 2021. 3 +[49] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 3 \ No newline at end of file diff --git a/2025/StruMamba3D_ Exploring Structural Mamba for Self-supervised Point Cloud Representation Learning/images.zip b/2025/StruMamba3D_ Exploring Structural Mamba for Self-supervised Point Cloud Representation Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..690860219dbb5988d66d21829aa6010a97d3b51a --- /dev/null +++ b/2025/StruMamba3D_ Exploring Structural Mamba for Self-supervised Point Cloud Representation Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ff274507edea7c8e2646918d0b3be9de218e652d83bc5c1c74f6e70e71883066 +size 640902 diff --git a/2025/StruMamba3D_ Exploring Structural Mamba for Self-supervised Point Cloud Representation Learning/layout.json b/2025/StruMamba3D_ Exploring Structural Mamba for Self-supervised Point Cloud Representation Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b742e6f930998ed125278d6b763ee13cd2675042 --- /dev/null +++ b/2025/StruMamba3D_ Exploring Structural Mamba for Self-supervised Point Cloud Representation Learning/layout.json @@ -0,0 +1,9721 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 104, + 103, + 506, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 103, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 104, + 103, + 506, + 140 + ], + "type": "text", + "content": "StruMamba3D: Exploring Structural Mamba for Self-supervised Point Cloud Representation Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 148, + 160, + 460, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 160, + 460, + 175 + ], + "spans": [ + { + "bbox": [ + 148, + 160, + 460, + 175 + ], + "type": "text", + "content": "Chuxin Wang" + }, + { + "bbox": [ + 148, + 160, + 460, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 148, + 160, + 460, + 175 + ], + "type": "text", + "content": ", Yixin Zha" + }, + { + "bbox": [ + 148, + 160, + 460, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 148, + 160, + 460, + 175 + ], + "type": "text", + "content": ", Wenfei Yang" + }, + { + "bbox": [ + 148, + 160, + 460, + 175 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 148, + 160, + 460, + 175 + ], + "type": "text", + "content": ", Tianzhu Zhang" + }, + { + "bbox": [ + 148, + 160, + 460, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 187, + 175, + 423, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 175, + 423, + 190 + ], + "spans": [ + { + "bbox": [ + 187, + 175, + 423, + 190 + ], + "type": "text", + "content": "1University of Science and Technology of China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 83, + 190, + 526, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 190, + 526, + 204 + ], + "spans": [ + { + "bbox": [ + 83, + 190, + 526, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 83, + 190, + 526, + 204 + ], + "type": "text", + "content": "National Key Laboratory of Deep Space Exploration, Deep Space Exploration Laboratory" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 204, + 476, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 204, + 476, + 218 + ], + "spans": [ + { + "bbox": [ + 133, + 204, + 476, + 218 + ], + "type": "text", + "content": "{wcx0602, zyxcn} " + }, + { + "bbox": [ + 133, + 204, + 476, + 218 + ], + "type": "inline_equation", + "content": "@" + }, + { + "bbox": [ + 133, + 204, + 476, + 218 + ], + "type": "text", + "content": " mail.ustc.edu.cn, {yangwf, tzzhang} " + }, + { + "bbox": [ + 133, + 204, + 476, + 218 + ], + "type": "inline_equation", + "content": "@" + }, + { + "bbox": [ + 133, + 204, + 476, + 218 + ], + "type": "text", + "content": " ustc.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 99, + 255, + 189, + 295 + ], + "blocks": [ + { + "bbox": [ + 111, + 241, + 161, + 250 + ], + "lines": [ + { + "bbox": [ + 111, + 241, + 161, + 250 + ], + "spans": [ + { + "bbox": [ + 111, + 241, + 161, + 250 + ], + "type": "text", + "content": "Attention Block" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 99, + 255, + 189, + 295 + ], + "lines": [ + { + "bbox": [ + 99, + 255, + 189, + 295 + ], + "spans": [ + { + "bbox": [ + 99, + 255, + 189, + 295 + ], + "type": "image", + "image_path": "def97985916b92a5e1a231264460e313ce670297dd366eb00c35093729f4b123.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 297, + 170, + 304 + ], + "lines": [ + { + "bbox": [ + 102, + 297, + 170, + 304 + ], + "spans": [ + { + "bbox": [ + 102, + 297, + 170, + 304 + ], + "type": "text", + "content": "Quadratic complexity" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 196, + 255, + 286, + 294 + ], + "blocks": [ + { + "bbox": [ + 201, + 241, + 266, + 249 + ], + "lines": [ + { + "bbox": [ + 201, + 241, + 266, + 249 + ], + "spans": [ + { + "bbox": [ + 201, + 241, + 266, + 249 + ], + "type": "text", + "content": "Selective SSM Block" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 196, + 255, + 286, + 294 + ], + "lines": [ + { + "bbox": [ + 196, + 255, + 286, + 294 + ], + "spans": [ + { + "bbox": [ + 196, + 255, + 286, + 294 + ], + "type": "image", + "image_path": "ec38826214ea8c80c9857e9afe51a06461c9b4434540c5512b79339503d86a76.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 296, + 263, + 305 + ], + "lines": [ + { + "bbox": [ + 205, + 296, + 263, + 305 + ], + "spans": [ + { + "bbox": [ + 205, + 296, + 263, + 305 + ], + "type": "text", + "content": "Linear complexity" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 293, + 254, + 345, + 295 + ], + "blocks": [ + { + "bbox": [ + 292, + 241, + 347, + 250 + ], + "lines": [ + { + "bbox": [ + 292, + 241, + 347, + 250 + ], + "spans": [ + { + "bbox": [ + 292, + 241, + 347, + 250 + ], + "type": "text", + "content": "Serialized Points" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 293, + 254, + 345, + 295 + ], + "lines": [ + { + "bbox": [ + 293, + 254, + 345, + 295 + ], + "spans": [ + { + "bbox": [ + 293, + 254, + 345, + 295 + ], + "type": "image", + "image_path": "f0f68ca372b516ee5050d2041570c1e75aadd8c1e070747bbd3cc6887d5f41cc.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 296, + 348, + 305 + ], + "lines": [ + { + "bbox": [ + 290, + 296, + 348, + 305 + ], + "spans": [ + { + "bbox": [ + 290, + 296, + 348, + 305 + ], + "type": "text", + "content": "Destroy adjacency" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 369, + 254, + 413, + 293 + ], + "blocks": [ + { + "bbox": [ + 421, + 241, + 465, + 251 + ], + "lines": [ + { + "bbox": [ + 421, + 241, + 465, + 251 + ], + "spans": [ + { + "bbox": [ + 421, + 241, + 465, + 251 + ], + "type": "text", + "content": "Longer Input" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 369, + 254, + 413, + 293 + ], + "lines": [ + { + "bbox": [ + 369, + 254, + 413, + 293 + ], + "spans": [ + { + "bbox": [ + 369, + 254, + 413, + 293 + ], + "type": "image", + "image_path": "87c767c30a776d8c18b83d5e0e01ab1ff193f3fcca4699d474205b54b42b36ec.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 296, + 476, + 305 + ], + "lines": [ + { + "bbox": [ + 362, + 296, + 476, + 305 + ], + "spans": [ + { + "bbox": [ + 362, + 296, + 476, + 305 + ], + "type": "text", + "content": "Fail to retain long-sequence memory" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 419, + 317, + 476, + 388 + ], + "blocks": [ + { + "bbox": [ + 419, + 306, + 478, + 316 + ], + "lines": [ + { + "bbox": [ + 419, + 306, + 478, + 316 + ], + "spans": [ + { + "bbox": [ + 419, + 306, + 478, + 316 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 419, + 317, + 476, + 388 + ], + "lines": [ + { + "bbox": [ + 419, + 317, + 476, + 388 + ], + "spans": [ + { + "bbox": [ + 419, + 317, + 476, + 388 + ], + "type": "image", + "image_path": "96ad2b99223270751631ee35700572645c798045e6a66095b184272867ae0b0a.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 474, + 259, + 489, + 267 + ], + "lines": [ + { + "bbox": [ + 474, + 259, + 489, + 267 + ], + "spans": [ + { + "bbox": [ + 474, + 259, + 489, + 267 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 474, + 267, + 521, + 274 + ], + "lines": [ + { + "bbox": [ + 474, + 267, + 521, + 274 + ], + "spans": [ + { + "bbox": [ + 474, + 267, + 521, + 274 + ], + "type": "text", + "content": "□ Point Feature" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 474, + 275, + 489, + 284 + ], + "lines": [ + { + "bbox": [ + 474, + 275, + 489, + 284 + ], + "spans": [ + { + "bbox": [ + 474, + 275, + 489, + 284 + ], + "type": "text", + "content": "□" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 85, + 319, + 526, + 397 + ], + "blocks": [ + { + "bbox": [ + 81, + 306, + 191, + 315 + ], + "lines": [ + { + "bbox": [ + 81, + 306, + 191, + 315 + ], + "spans": [ + { + "bbox": [ + 81, + 306, + 191, + 315 + ], + "type": "text", + "content": "(a) Transformer-based Paradigm" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 488, + 327, + 514, + 335 + ], + "lines": [ + { + "bbox": [ + 488, + 327, + 514, + 335 + ], + "spans": [ + { + "bbox": [ + 488, + 327, + 514, + 335 + ], + "type": "text", + "content": "PointMAE" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 488, + 337, + 513, + 346 + ], + "lines": [ + { + "bbox": [ + 488, + 337, + 513, + 346 + ], + "spans": [ + { + "bbox": [ + 488, + 337, + 513, + 346 + ], + "type": "text", + "content": "PointGPT" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 488, + 350, + 514, + 357 + ], + "lines": [ + { + "bbox": [ + 488, + 350, + 514, + 357 + ], + "spans": [ + { + "bbox": [ + 488, + 350, + 514, + 357 + ], + "type": "text", + "content": "Mamba3D" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 488, + 360, + 514, + 368 + ], + "lines": [ + { + "bbox": [ + 488, + 360, + 514, + 368 + ], + "spans": [ + { + "bbox": [ + 488, + 360, + 514, + 368 + ], + "type": "text", + "content": "PointMaml" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 488, + 369, + 514, + 377 + ], + "lines": [ + { + "bbox": [ + 488, + 369, + 514, + 377 + ], + "spans": [ + { + "bbox": [ + 488, + 369, + 514, + 377 + ], + "type": "text", + "content": "Our" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 488, + 379, + 501, + 386 + ], + "lines": [ + { + "bbox": [ + 488, + 379, + 501, + 386 + ], + "spans": [ + { + "bbox": [ + 488, + 379, + 501, + 386 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 170, + 306, + 379, + 316 + ], + "lines": [ + { + "bbox": [ + 170, + 306, + 379, + 316 + ], + "spans": [ + { + "bbox": [ + 170, + 306, + 379, + 316 + ], + "type": "text", + "content": "adigm (b) Mamba-based Paradigm" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 85, + 319, + 526, + 397 + ], + "lines": [ + { + "bbox": [ + 85, + 319, + 526, + 397 + ], + "spans": [ + { + "bbox": [ + 85, + 319, + 526, + 397 + ], + "type": "image", + "image_path": "1c70f27e6a058199ddbbfc8109c778f96e2b2e9ee2b29abc5fd02d304a0643cb.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 397, + 555, + 453 + ], + "lines": [ + { + "bbox": [ + 55, + 397, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 555, + 453 + ], + "type": "text", + "content": "Figure 1. (a): Transformer-based paradigm uses the attention block with quadratic complexity to model the dependencies between points. (b): Mamba-based paradigm uses the selective SSM with linear complexity. However, serialized points destroy the adjacency of 3D points, and the pre-trained selection mechanism fails to retain long-sequence memory. (c): Our StruMamba3D paradigm uses the structural SSM to maintain the spatial dependencies among points and the sequence length-adaptive strategy to retain long-sequence memory. (d): Our StruMamba3D significantly outperforms existing Transformer-based and Mamba-based methods." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "bbox": [ + 151, + 464, + 200, + 476 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 464, + 200, + 476 + ], + "spans": [ + { + "bbox": [ + 151, + 464, + 200, + 476 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 54, + 479, + 297, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 479, + 297, + 696 + ], + "spans": [ + { + "bbox": [ + 54, + 479, + 297, + 696 + ], + "type": "text", + "content": "Recently, Mamba-based methods have demonstrated impressive performance in point cloud representation learning by leveraging State Space Model (SSM) with the efficient context modeling ability and linear complexity. However, these methods still face two key issues that limit the potential of SSM: Destroying the adjacency of 3D points during SSM processing and failing to retain long-sequence memory as the input length increases in downstream tasks. To address these issues, we propose StruMamba3D, a novel paradigm for self-supervised point cloud representation learning. It enjoys several merits. First, we design spatial states and use them as proxies to preserve spatial dependencies among points. Second, we enhance the SSM with a state-wise update strategy and incorporate a lightweight convolution to facilitate interactions between spatial states for efficient structure modeling. Third, our method reduces the sensitivity of pre-trained Mamba-based models to varying input lengths by introducing a sequence length-adaptive" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 466, + 555, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 555, + 525 + ], + "type": "text", + "content": "strategy. Experimental results across four downstream tasks showcase the superior performance of our method. In addition, our method attains the SOTA " + }, + { + "bbox": [ + 313, + 466, + 555, + 525 + ], + "type": "inline_equation", + "content": "95.1\\%" + }, + { + "bbox": [ + 313, + 466, + 555, + 525 + ], + "type": "text", + "content": " accuracy on ModelNet40 and " + }, + { + "bbox": [ + 313, + 466, + 555, + 525 + ], + "type": "inline_equation", + "content": "92.75\\%" + }, + { + "bbox": [ + 313, + 466, + 555, + 525 + ], + "type": "text", + "content": " accuracy on the most challenging split of ScanObjectNN without voting strategy." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 314, + 539, + 394, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 539, + 394, + 551 + ], + "spans": [ + { + "bbox": [ + 314, + 539, + 394, + 551 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 312, + 559, + 555, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 559, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 312, + 559, + 555, + 666 + ], + "type": "text", + "content": "Point cloud representation learning aims to extract geometric and semantic features from point clouds to support downstream tasks such as classification, segmentation, and detection. As a foundational task in Computer Vision (CV), it is crucial for various real-world applications, including autonomous driving [4, 20], AR [13], and robotics [26, 32]. However, unlike images or text with structured arrangements, point clouds are inherently unordered and sparse, presenting challenges for designing effective models." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": "To solve this problem, numerous point cloud representation learning methods have been proposed, which can be broadly divided into three categories according to model architectures: Point-based methods [19, 27, 28, 30, 31," + } + ] + } + ], + "index": 39 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 704, + 143, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 704, + 143, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 704, + 143, + 714 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28546" + } + ] + } + ], + "index": 41 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 294, + 346 + ], + "type": "text", + "content": "35], Transformer-based methods [12, 22, 25, 42-44], and Mamba-based methods [15, 21, 47]. Point-based methods directly process point clouds by modeling the relationships between individual points, such as through set abstraction [27, 28] and graph convolution [19, 35]. These methods offer a simple and computationally efficient design. However, their limited receptive field hinders capturing global information, constraining complex structure modeling. Transformer-based methods leverage attention mechanisms to capture long-range dependencies, expanding the receptive field. Moreover, inspired by self-supervised learning paradigms in NLP [6] and CV [16], many transformer-based methods [22, 25, 42, 44] utilize large unlabeled point cloud datasets for pre-training, achieving impressive results in downstream tasks. However, these methods have limited scalability due to the quadratic complexity of the attention mechanism, as shown in Fig. 1(a). To overcome this limitation, Mamba-based methods [15, 21, 47] first convert the 3D points into a 1D sequence and then utilize selective State Space Model (SSM) [9] with linear complexity to extract point features. Leveraging the robust information aggregation ability of selective SSM, Mamba-based methods have attained promising performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 352, + 294, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 352, + 294, + 661 + ], + "spans": [ + { + "bbox": [ + 57, + 352, + 294, + 661 + ], + "type": "text", + "content": "Although existing Mamba-based methods [15, 21, 47] have achieved comparable performance to transformer-based methods, two key issues limit the potential of SSM: (1) Distort the adjacency of points. As shown in Fig. 1(b), although methods [21, 47] use serialization strategies to align input points, they cannot ensure that spatially adjacent points remain neighbors in the 1D sequence. Unlike text and images, points inherently lack context information and rely on local structures as base feature units. This spatial distortion hampers Mamba to model fine-grained structure information. (2) Fail to retain long-sequence memory in downstream tasks. Mamba-based methods [15, 21] use Masked Point Modeling (MPM) [25, 42] as a pre-training task, which creates a significant discrepancy in input sequence lengths between pre-training and downstream tasks. Mamba leverages a selection mechanism to inductively learn new information while selectively forgetting outdated knowledge. However, the selection mechanism does not account for variations in input sequence length. Consequently, when Mamba is pre-trained on short sequences, it tends to update states more frequently. As the input sequence length increases in downstream tasks, frequent state updates lead to Mamba struggling to retain long-sequence memory, as shown in Fig. 1(b). The deficiency in long-sequence memory adversely affects the long-range semantics modeling, resulting in suboptimal performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 666, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 666, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 666, + 294, + 712 + ], + "type": "text", + "content": "To unleash the potential of Mamba for point cloud representation learning, we explore solutions to address the above two issues. (1) How to maintain spatial dependencies among points? Maintaining spatial relationships" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 317, + 72, + 553, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 553, + 203 + ], + "type": "text", + "content": "among 3D points within a 1D input sequence is challenging. Mamba efficiently aggregates contextual features from the 1D input sequence via hidden states. Is it possible to design a structural Mamba that models spatial relationships among points through hidden states? (2) How to retain long-sequence memory for longer inputs? When handling longer input sequences, states need to retain memory for an extended period to effectively capture the contextual information across the entire sequence. If the model adjusts the frequency of state updates when the sequence length changes, it can preserve memory for longer sequences." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 205, + 553, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 205, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 317, + 205, + 553, + 502 + ], + "type": "text", + "content": "To achieve the above goals, we propose a novel StruMamba3D paradigm for self-supervised point cloud representation learning, which consists of two core components: structural SSM block and sequence length-adaptive strategy. In structural SSM block, we leverage states in SSM to model the spatial relationships among points. We begin by assigning positional attributes to the states, using them as proxies to capture the local structure of point clouds. Next, we propose a state-wise update strategy that incorporates the positional relationship between input points and spatial states during the state update and propagation process. Based on spatial states and state-wise update strategy, we can maintain spatial dependencies among points during the SSM process. To address the limitation of standard Mamba ignoring inter-state interactions, we introduce a lightweight convolution module for spatial states. In sequence length-adaptive strategy, the adaptive state update mechanism is first employed to adjust the state update frequency based on sequence length, ensuring that the states retain long-sequence memory as the sequence length grows. Additionally, we introduce a spatial state consistency loss in the pre-training task, which enforces consistency in the updated states across inputs of varying sequence lengths. The spatial state consistency loss enhances the robustness of the Mamba-based pre-trained model to varying input lengths." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 503, + 553, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 503, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 503, + 553, + 645 + ], + "type": "text", + "content": "To sum up, the main contributions of this work are: (1) We propose the novel StruMamba3D for self-supervised point cloud representation learning, which is the first to model the structure information by latent states in SSM. (2) We design the structural SSM to maintain spatial dependencies among points during state information selection and propagation, while the sequence length-adaptive strategy reduces the sensitivity of pre-trained Mamba-based models to varying input lengths. (3) Experimental results on four downstream tasks showcase the superior performance of our method. Additionally, comprehensive ablation studies validate the effectiveness of our designs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 658, + 399, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 399, + 669 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 399, + 669 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 678, + 553, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 678, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 317, + 678, + 553, + 700 + ], + "type": "text", + "content": "We briefly review methods for point cloud representation learning and self-supervised representation learning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 702, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 553, + 712 + ], + "type": "text", + "content": "Point Cloud Representation Learning. To effectively" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28547" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 396 + ], + "type": "text", + "content": "leverage the sparse, unordered and irregular point cloud, various point cloud representation learning methods have been proposed, which can be broadly categorized into three groups: Point-based methods [19, 27, 28, 30, 31, 35], Transformer-based methods [12, 22, 25, 33, 42, 44], and Mamba-based methods [15, 21, 47]. Point-based methods directly process raw point clouds to learn geometric features. For example, PointNet [27] employs MLPs to each point independently, followed by global max pooling to aggregate features. PointNet++ [28] extends PointNet by introducing a hierarchical architecture for multi-scale feature extraction. Subsequent methods [19, 24, 30, 31, 35] explore convolutional techniques for point clouds. DGCNN [35] leverages graph convolution to model structural information among neighboring points, while KPCov [31] introduces a convolution method based on dynamic kernel points to improve feature learning. To capture long-range dependencies, methods [12, 37, 38, 49] have explored transformer architectures for point clouds. PCT [12] applies global attention to points, while Point Transformer [49] introduces local attention to capture local geometry features. Recent variants [37, 38] further refine the architectures to improve performance and efficiency. These supervised methods [12, 30, 31, 47, 49] have demonstrated outstanding performance in specific tasks. However, these methods are inherently limited by the specific data domains on which they are trained, leading to poor generalization to other domains." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 403, + 298, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 298, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 298, + 715 + ], + "type": "text", + "content": "Self-supervised Representation Learning. To solve the above issues and fully utilize a large number of unlabeled data, self-supervised representation learning methods [3, 22, 25, 40, 42, 44, 48] have emerged. Some contrastive learning-based methods [40, 48] train models to learn distinguishing features by differentiating between positive and negative samples. For instance, Point-Contrast [40] leverages feature consistency across different views of the same point cloud. Inspired by the success of BERT [6] and MAE [16], many transformer-based methods [3, 22, 25, 42, 44] adopt MPM task for pre-training, encouraging models to infer masked patches from visible ones. PointBERT [42] predicts discrete tokens for masked points, while PointMAE [25] reconstructs masked point clouds. Additionally, recent studies [29, 45] integrate cross-modal information, such as images and text, to enhance pre-training. However, transformer-based methods suffer from quadratic complexity, posing challenges for handling long sequences and making them unsuitable for resource-constrained devices. Recently, SSMs [5, 9-11, 17, 34] have emerged as a prominent research focus due to their linear complexity and powerful information aggregation ability. In the field of point cloud representation learning, several Mamba-based methods [15, 21, 47] have been proposed. PointMamba [21] employs space-filling curves to serialize point clouds, establishing a simple yet effective base" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 313, + 72, + 555, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 277 + ], + "type": "text", + "content": "line. PCM [47] introduces multiple serialization strategies and a bidirectional scanning mechanism to capture comprehensive point cloud structures, but also introduces larger computational overhead. Mamba3D [15] also leverages a bidirectional scanning mechanism and proposes a local norm pooling operation to enhance local geometric features. However, these methods introduce structural distortions when converting 3D points into 1D sequences, limiting model performance, especially in tasks requiring fine-grained geometry like part segmentation. Moreover, they neglect the impact of varying input lengths between pretraining and downstream tasks. Unlike these methods, we propose StruMamba3D, which leverages states as proxies to maintain spatial dependencies among points. Additionally, we introduce a sequence-length adaptive strategy to ensure that the states retain long-sequence memory as the sequence length grows in downstream tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 314, + 285, + 370, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 285, + 370, + 298 + ], + "spans": [ + { + "bbox": [ + 314, + 285, + 370, + 298 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 305, + 556, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 305, + 556, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 556, + 354 + ], + "type": "text", + "content": "Below, we review the previous SSMs (Sec. 3.1), followed by an overview of our StruMamba3D (Sec. 3.2). Then, we introduce the details of Structural SSM Block (Sec. 3.3) and Sequence Length-adaptive Strategy (Sec. 3.4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 360, + 401, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 360, + 401, + 372 + ], + "spans": [ + { + "bbox": [ + 314, + 360, + 401, + 372 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 377, + 555, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 555, + 403 + ], + "type": "text", + "content": "State Space Model (SSM) is a continuous system that maps the input " + }, + { + "bbox": [ + 313, + 377, + 555, + 403 + ], + "type": "inline_equation", + "content": "x(t)" + }, + { + "bbox": [ + 313, + 377, + 555, + 403 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 377, + 555, + 403 + ], + "type": "inline_equation", + "content": "y(t)" + }, + { + "bbox": [ + 313, + 377, + 555, + 403 + ], + "type": "text", + "content": " through implicit states " + }, + { + "bbox": [ + 313, + 377, + 555, + 403 + ], + "type": "inline_equation", + "content": "h(t)\\in \\mathbb{R}^{M}" + }, + { + "bbox": [ + 313, + 377, + 555, + 403 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 350, + 408, + 555, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 408, + 555, + 423 + ], + "spans": [ + { + "bbox": [ + 350, + 408, + 555, + 423 + ], + "type": "interline_equation", + "content": "h ^ {\\prime} (t) = \\mathbf {A} h (t) + \\mathbf {B} x (t), y (t) = \\mathbf {C} h ^ {\\prime} (t), \\qquad (1)", + "image_path": "d4754888537e78887c8fb600749e9ad735881e8794c360522048d7b4ca8eb733.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 429, + 556, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 556, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 556, + 489 + ], + "type": "text", + "content": "where the matrix " + }, + { + "bbox": [ + 313, + 429, + 556, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 313, + 429, + 556, + 489 + ], + "type": "text", + "content": " governs the evolution of states, while the matrices " + }, + { + "bbox": [ + 313, + 429, + 556, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 313, + 429, + 556, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 429, + 556, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 313, + 429, + 556, + 489 + ], + "type": "text", + "content": " are used for modeling the relationship between inputs and outputs through the latent states. To handle discrete-time sequence data, the zero-order hold discretization rule is used:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 320, + 494, + 553, + 524 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 494, + 553, + 524 + ], + "spans": [ + { + "bbox": [ + 320, + 494, + 553, + 524 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} h _ {t} = \\overline {{\\mathbf {A}}} h _ {t - 1} + \\overline {{\\mathbf {B}}} x _ {t}, y _ {t} = \\mathbf {C} h _ {t}, \\\\ \\overline {{\\mathbf {A}}} = \\exp (\\boldsymbol {\\Delta} \\mathbf {A}), \\overline {{\\mathbf {B}}} = (\\boldsymbol {\\Delta} \\mathbf {A}) ^ {- 1} (\\exp (\\boldsymbol {\\Delta} \\mathbf {A}) - \\mathbf {I}) \\boldsymbol {\\Delta} \\mathbf {B}, \\tag {2} \\\\ \\end{array}", + "image_path": "cb88690bbabf29570639dc794523c22c75bf9ef5b6f34634485810c077d99aea.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 530, + 556, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 556, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 556, + 615 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 530, + 556, + 615 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 313, + 530, + 556, + 615 + ], + "type": "text", + "content": " represents the sampling interval between consecutive inputs " + }, + { + "bbox": [ + 313, + 530, + 556, + 615 + ], + "type": "inline_equation", + "content": "x_{t-1}" + }, + { + "bbox": [ + 313, + 530, + 556, + 615 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 530, + 556, + 615 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 313, + 530, + 556, + 615 + ], + "type": "text", + "content": ". However, due to the linear time-invariant nature of SSM, the parameters " + }, + { + "bbox": [ + 313, + 530, + 556, + 615 + ], + "type": "inline_equation", + "content": "(\\Delta, \\mathbf{A}, \\mathbf{B}, \\mathbf{C})" + }, + { + "bbox": [ + 313, + 530, + 556, + 615 + ], + "type": "text", + "content": " remain fixed across all time steps, which limits the expressive capacity of SSM. To overcome this limitation, Mamba [9] treats the parameters " + }, + { + "bbox": [ + 313, + 530, + 556, + 615 + ], + "type": "inline_equation", + "content": "(\\Delta, \\mathbf{B}, \\mathbf{C})" + }, + { + "bbox": [ + 313, + 530, + 556, + 615 + ], + "type": "text", + "content": " as functions of input " + }, + { + "bbox": [ + 313, + 530, + 556, + 615 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 313, + 530, + 556, + 615 + ], + "type": "text", + "content": ", transforming the SSM into a time-varying model:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 329, + 621, + 555, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 621, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 329, + 621, + 555, + 635 + ], + "type": "interline_equation", + "content": "h _ {t} = \\phi_ {\\overline {{\\mathbf {A}}}} (x _ {t}) h _ {t - 1} + \\phi_ {\\overline {{\\mathbf {B}}}} (x _ {t}) x _ {t}, y _ {t} = \\phi_ {\\mathbf {C}} (x _ {t}) h _ {t}, \\tag {3}", + "image_path": "7d9c5330c8e009d1f55b14b3bf18a5cad4886416c8fc254cb66437303337e621.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 641, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 557, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 641, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\phi_{*}(x_{t})" + }, + { + "bbox": [ + 313, + 641, + 557, + 715 + ], + "type": "text", + "content": " denotes the linear projection of input " + }, + { + "bbox": [ + 313, + 641, + 557, + 715 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 641, + 557, + 715 + ], + "type": "text", + "content": ". Additionally, Mamba introduces a hardware-aware scan algorithm to achieve near-linear complexity. Our StruMamba3D adopts the same hardware-aware scan algorithm, maintaining linear complexity while exhibiting strong structural modeling capabilities for point clouds." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28548" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 66, + 531, + 178 + ], + "blocks": [ + { + "bbox": [ + 77, + 66, + 531, + 178 + ], + "lines": [ + { + "bbox": [ + 77, + 66, + 531, + 178 + ], + "spans": [ + { + "bbox": [ + 77, + 66, + 531, + 178 + ], + "type": "image", + "image_path": "2599e73e448e8ac0f9e9a2c95cf179ca547207d5e4ff9169e95adf4f49c84a6d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 77, + 182, + 274, + 300 + ], + "blocks": [ + { + "bbox": [ + 77, + 182, + 274, + 300 + ], + "lines": [ + { + "bbox": [ + 77, + 182, + 274, + 300 + ], + "spans": [ + { + "bbox": [ + 77, + 182, + 274, + 300 + ], + "type": "image", + "image_path": "245557cd6982fb0bcd1a4f50126a0e1e15b54018feef192feebb36d817beaba7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 303, + 555, + 336 + ], + "lines": [ + { + "bbox": [ + 55, + 303, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 555, + 336 + ], + "type": "text", + "content": "Figure 2. Overview of StruMamba3D. (a): Architecture of StruMamba3D. (b): Details of Structural SSM Block. We employ structural SSM to capture dependencies between spatial states and input points. Additionally, we integrate a lightweight convolution module to introduce local interactions among spatial states or input points. (c): Details of Structural SSM. (d): Details of Lightweight Convolution." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 279, + 182, + 419, + 300 + ], + "blocks": [ + { + "bbox": [ + 279, + 182, + 419, + 300 + ], + "lines": [ + { + "bbox": [ + 279, + 182, + 419, + 300 + ], + "spans": [ + { + "bbox": [ + 279, + 182, + 419, + 300 + ], + "type": "image", + "image_path": "ff4d03cf7c1c8516d51673698a018b37c25efae3b3fb8a450e8440a44f665fd4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 425, + 182, + 531, + 300 + ], + "blocks": [ + { + "bbox": [ + 425, + 182, + 531, + 300 + ], + "lines": [ + { + "bbox": [ + 425, + 182, + 531, + 300 + ], + "spans": [ + { + "bbox": [ + 425, + 182, + 531, + 300 + ], + "type": "image", + "image_path": "21a8c32d4dce1a734c0437ced7e56527159a55eb8ba35a8d215373a21b2dc06e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 347, + 123, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 123, + 358 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 123, + 358 + ], + "type": "text", + "content": "3.2. Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 364, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 364, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 54, + 364, + 296, + 568 + ], + "type": "text", + "content": "Despite Mamba achieving remarkable results on sequential data, maintaining the spatial relationships among 3D points remains challenging. We aim to model the spatial relationships by latent states and design StruMamba3D for point cloud representation learning, as illustrated in Fig. 2(a). Given raw point cloud " + }, + { + "bbox": [ + 54, + 364, + 296, + 568 + ], + "type": "inline_equation", + "content": "P_{raw} \\in \\mathbb{R}^{N_0 \\times 3}" + }, + { + "bbox": [ + 54, + 364, + 296, + 568 + ], + "type": "text", + "content": ", we obtain the input points " + }, + { + "bbox": [ + 54, + 364, + 296, + 568 + ], + "type": "inline_equation", + "content": "P_x \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 54, + 364, + 296, + 568 + ], + "type": "text", + "content": " using farthest point sampling (FPS) and " + }, + { + "bbox": [ + 54, + 364, + 296, + 568 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 364, + 296, + 568 + ], + "type": "text", + "content": "-nearest neighbor clustering (KNN). Then, we use a lightweight PointNet [27] to extract the input token embeddings " + }, + { + "bbox": [ + 54, + 364, + 296, + 568 + ], + "type": "inline_equation", + "content": "F_x \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 54, + 364, + 296, + 568 + ], + "type": "text", + "content": ". Meanwhile, we initialize the spatial states " + }, + { + "bbox": [ + 54, + 364, + 296, + 568 + ], + "type": "inline_equation", + "content": "P_h \\in \\mathbb{R}^{M \\times 3}" + }, + { + "bbox": [ + 54, + 364, + 296, + 568 + ], + "type": "text", + "content": " as the proxies for local structures from the raw point cloud. Finally, the spatial states and input token embeddings are fed into a feature encoder composed of multiple structural SSM blocks, where the spatial states act as a bridge to connect the inputs and outputs. Notably, we use the output spatial states from the previous structural SSM block as the input spatial states for the current block." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 576, + 182, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 182, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 182, + 588 + ], + "type": "text", + "content": "3.3. Structural SSM Block" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "The structural SSM block is designed to capture complex structural features by preserving spatial dependencies among points throughout SSM processing. As shown in Fig. 2(b), the block takes spatial states " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "F_{h}" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " and token embeddings " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "F_{x}" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " as input and produces updated spatial states " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "F_{h}^{\\prime}" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " and token embeddings " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "F_{x}^{\\prime}" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " as output. To clarify the details of structural SSM block, we describe each component in sequence: Spatial State Initialization, State-wise Update Mechanism, Structural SSM, and Lightweight Convolution. Spatial State Initialization. In original Mamba, latent" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 347, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 555, + 456 + ], + "type": "text", + "content": "states do not contain geometric information, making it difficult to model the local structure of point clouds. To address this, we introduce positional attributes " + }, + { + "bbox": [ + 313, + 347, + 555, + 456 + ], + "type": "inline_equation", + "content": "P_{h} \\in \\mathbb{R}^{M \\times 3}" + }, + { + "bbox": [ + 313, + 347, + 555, + 456 + ], + "type": "text", + "content": " to the latent states, enabling each state to focus on distinct regions of the point cloud. Due to the sparsity of point clouds, we initialize state positions " + }, + { + "bbox": [ + 313, + 347, + 555, + 456 + ], + "type": "inline_equation", + "content": "P_{h}" + }, + { + "bbox": [ + 313, + 347, + 555, + 456 + ], + "type": "text", + "content": " based on raw points " + }, + { + "bbox": [ + 313, + 347, + 555, + 456 + ], + "type": "inline_equation", + "content": "P_{raw}" + }, + { + "bbox": [ + 313, + 347, + 555, + 456 + ], + "type": "text", + "content": " to ensure comprehensive coverage of point clouds. Specifically, we employ FPS and KNN clustering to segment the point cloud into groups " + }, + { + "bbox": [ + 313, + 347, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{G}_{m}\\}_{m=1}^{M}" + }, + { + "bbox": [ + 313, + 347, + 555, + 456 + ], + "type": "text", + "content": ", calculating the centroid:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 387, + 464, + 555, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 464, + 555, + 493 + ], + "spans": [ + { + "bbox": [ + 387, + 464, + 555, + 493 + ], + "type": "interline_equation", + "content": "P _ {h} ^ {m} = \\frac {1}{\\left| \\mathcal {G} _ {m} \\right|} \\sum_ {P _ {i} \\in \\mathcal {G} _ {m}} P _ {i}, \\tag {4}", + "image_path": "352fb9f4217644d89b2a600c6c074e9a4b699c3d4c7c81dccf9933a805af4f94.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 501, + 555, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 555, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 555, + 549 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 501, + 555, + 549 + ], + "type": "inline_equation", + "content": "P_{h}^{m}" + }, + { + "bbox": [ + 313, + 501, + 555, + 549 + ], + "type": "text", + "content": " represents the positional attribute of the " + }, + { + "bbox": [ + 313, + 501, + 555, + 549 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 501, + 555, + 549 + ], + "type": "text", + "content": "-th state. We refer to states with positional attributes as spatial states and then use linear mapping " + }, + { + "bbox": [ + 313, + 501, + 555, + 549 + ], + "type": "inline_equation", + "content": "\\phi_{h}" + }, + { + "bbox": [ + 313, + 501, + 555, + 549 + ], + "type": "text", + "content": " to embed positional attributes as the initial state features." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 383, + 558, + 553, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 558, + 553, + 572 + ], + "spans": [ + { + "bbox": [ + 383, + 558, + 553, + 572 + ], + "type": "interline_equation", + "content": "F _ {h} = \\phi_ {h} \\left(P _ {h}\\right) \\in \\mathbb {R} ^ {M \\times D}. \\tag {5}", + "image_path": "7baf0f893cf79c2d5bac59c8826448d53c407249ab8b58b44ea00492c8dfab6a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 605 + ], + "type": "text", + "content": "The state features " + }, + { + "bbox": [ + 313, + 582, + 555, + 605 + ], + "type": "inline_equation", + "content": "F_{h}" + }, + { + "bbox": [ + 313, + 582, + 555, + 605 + ], + "type": "text", + "content": " are then used as the initial latent states in SSM and updated along with input token embeddings." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "content": "State-wise Update. As illustrated in Eq. (3), the state update and propagation equations of the original Mamba rely solely on input tokens, disregarding the spatial relationships between input points and spatial states. This limitation makes it challenging for Mamba to effectively model the structural information of point clouds. To address this, we modify the state update and propagation equations to explicitly model the spatial dependencies between states and inputs. Specifically, we first calculate the relative offsets" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28549" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 278, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 278, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 278, + 85 + ], + "type": "text", + "content": "between the input points " + }, + { + "bbox": [ + 55, + 72, + 278, + 85 + ], + "type": "inline_equation", + "content": "P_{x}" + }, + { + "bbox": [ + 55, + 72, + 278, + 85 + ], + "type": "text", + "content": " and the state positions " + }, + { + "bbox": [ + 55, + 72, + 278, + 85 + ], + "type": "inline_equation", + "content": "P_{h}" + }, + { + "bbox": [ + 55, + 72, + 278, + 85 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 134, + 89, + 295, + 103 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 89, + 295, + 103 + ], + "spans": [ + { + "bbox": [ + 134, + 89, + 295, + 103 + ], + "type": "interline_equation", + "content": "\\triangle P _ {i} ^ {m} = P _ {x} ^ {i} - P _ {h} ^ {m}, \\tag {6}", + "image_path": "bb03e9e4ba89489c20a6125f206ede596fe8086f95e212228c45258927431ade.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 107, + 296, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 107, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 107, + 296, + 192 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 107, + 296, + 192 + ], + "type": "inline_equation", + "content": "\\triangle P_i^m" + }, + { + "bbox": [ + 55, + 107, + 296, + 192 + ], + "type": "text", + "content": " denotes the relative offset between the " + }, + { + "bbox": [ + 55, + 107, + 296, + 192 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 107, + 296, + 192 + ], + "type": "text", + "content": "-th input point and the " + }, + { + "bbox": [ + 55, + 107, + 296, + 192 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 107, + 296, + 192 + ], + "type": "text", + "content": "-th state. Next, we map these offsets to the SSM parameters " + }, + { + "bbox": [ + 55, + 107, + 296, + 192 + ], + "type": "inline_equation", + "content": "(\\mathbf{B}, \\mathbf{C})" + }, + { + "bbox": [ + 55, + 107, + 296, + 192 + ], + "type": "text", + "content": ", which govern state update and propagation. It is important to note that considering only spatial relationships is insufficient, as the significance of information varies across different points. Thus, we design the following state update and propagation mechanism:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 100, + 197, + 295, + 210 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 197, + 295, + 210 + ], + "spans": [ + { + "bbox": [ + 100, + 197, + 295, + 210 + ], + "type": "interline_equation", + "content": "\\left(\\mathbf {B} _ {i} ^ {m}, \\mathbf {C} _ {i} ^ {m}\\right) = \\phi \\left(x _ {i}\\right) + \\operatorname {M L P} \\left(\\triangle P _ {i} ^ {m}\\right), \\tag {7}", + "image_path": "1acfd2a8ce4875dab80692c7482927375dfbd91bd92567dd4aed9087b0a8392e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 215, + 296, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 215, + 296, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 296, + 287 + ], + "type": "text", + "content": "where MLP represents a multi-layer perceptron, " + }, + { + "bbox": [ + 55, + 215, + 296, + 287 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 215, + 296, + 287 + ], + "type": "text", + "content": " is a linear projection and " + }, + { + "bbox": [ + 55, + 215, + 296, + 287 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 55, + 215, + 296, + 287 + ], + "type": "text", + "content": " denotes the feature of the " + }, + { + "bbox": [ + 55, + 215, + 296, + 287 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 215, + 296, + 287 + ], + "type": "text", + "content": "-th input point. Based on the modified parameters, the spatial states can selectively update features using points within the same regions, while the input points can acquire local structural information from the states, as shown in Fig. 2(c)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 287, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 287, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 296, + 443 + ], + "type": "text", + "content": "Structural SSM. By leveraging spatial states and state-wise updates, we effectively model the structural information of point clouds. However, due to unidirectional scanning mechanism, bidirectional information exchange between input points cannot be achieved within a single forward pass. To address this, we adopt a bidirectional scanning mechanism inspired by VisionMamba [25]. Specifically, we apply both forward and backward structural SSMs, with the backward SSM processing input points in reverse order. Both use spatial states " + }, + { + "bbox": [ + 55, + 287, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\hat{F}_h" + }, + { + "bbox": [ + 55, + 287, + 296, + 443 + ], + "type": "text", + "content": " as initial states " + }, + { + "bbox": [ + 55, + 287, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\hat{F}_h^0" + }, + { + "bbox": [ + 55, + 287, + 296, + 443 + ], + "type": "text", + "content": ", generating updated features for input points and spatial states. Finally, we fuse these outputs using a linear layer " + }, + { + "bbox": [ + 55, + 287, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\phi_o" + }, + { + "bbox": [ + 55, + 287, + 296, + 443 + ], + "type": "text", + "content": " to obtain the final features " + }, + { + "bbox": [ + 55, + 287, + 296, + 443 + ], + "type": "inline_equation", + "content": "F_x'" + }, + { + "bbox": [ + 55, + 287, + 296, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 287, + 296, + 443 + ], + "type": "inline_equation", + "content": "F_h'" + }, + { + "bbox": [ + 55, + 287, + 296, + 443 + ], + "type": "text", + "content": " for input points and spatial states." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 71, + 448, + 295, + 463 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 448, + 295, + 463 + ], + "spans": [ + { + "bbox": [ + 71, + 448, + 295, + 463 + ], + "type": "interline_equation", + "content": "F _ {x} ^ {\\prime}, F _ {h} ^ {\\prime} = \\phi_ {o} (\\mathrm {S S M} _ {f} (\\hat {F} _ {x}, \\hat {F} _ {h}) + \\mathrm {S S M} _ {b} (\\hat {F} _ {x}, \\hat {F} _ {h})), \\tag {8}", + "image_path": "152c004a76fd21ec7b8a6c276dd1da8047f6e154dbe3224a268d6024e71edb85.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 468, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 295, + 479 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 468, + 295, + 479 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 468, + 295, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 468, + 295, + 479 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 468, + 295, + 479 + ], + "type": "text", + "content": " denote the forward and backward processes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 479, + 296, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 296, + 635 + ], + "type": "text", + "content": "Lightweight Convolution. The proposed structural SSM effectively models the dependencies between spatial states " + }, + { + "bbox": [ + 55, + 479, + 296, + 635 + ], + "type": "inline_equation", + "content": "P_{h}" + }, + { + "bbox": [ + 55, + 479, + 296, + 635 + ], + "type": "text", + "content": " and input points " + }, + { + "bbox": [ + 55, + 479, + 296, + 635 + ], + "type": "inline_equation", + "content": "P_{x}" + }, + { + "bbox": [ + 55, + 479, + 296, + 635 + ], + "type": "text", + "content": ". However, we observe that the spatial states " + }, + { + "bbox": [ + 55, + 479, + 296, + 635 + ], + "type": "inline_equation", + "content": "P_{h}" + }, + { + "bbox": [ + 55, + 479, + 296, + 635 + ], + "type": "text", + "content": " in SSMs remain isolated, lacking direct interaction with each other. To address this, we introduce a lightweight convolution module, as shown in Fig. 2(d). Inspired by graph convolution [35], this module aggregates features from neighboring states based on relative position information. Specifically, for each spatial state " + }, + { + "bbox": [ + 55, + 479, + 296, + 635 + ], + "type": "inline_equation", + "content": "P_{h}^{m}" + }, + { + "bbox": [ + 55, + 479, + 296, + 635 + ], + "type": "text", + "content": ", we first identify its " + }, + { + "bbox": [ + 55, + 479, + 296, + 635 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 479, + 296, + 635 + ], + "type": "text", + "content": "-nearest neighbors " + }, + { + "bbox": [ + 55, + 479, + 296, + 635 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(m)" + }, + { + "bbox": [ + 55, + 479, + 296, + 635 + ], + "type": "text", + "content": " and compute the relative positions " + }, + { + "bbox": [ + 55, + 479, + 296, + 635 + ], + "type": "inline_equation", + "content": "\\triangle P_{h}^{mj}" + }, + { + "bbox": [ + 55, + 479, + 296, + 635 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 479, + 296, + 635 + ], + "type": "inline_equation", + "content": "j \\in \\mathcal{N}(m)" + }, + { + "bbox": [ + 55, + 479, + 296, + 635 + ], + "type": "text", + "content": ". We then generate attention weights " + }, + { + "bbox": [ + 55, + 479, + 296, + 635 + ], + "type": "inline_equation", + "content": "\\{w_{mj}\\}" + }, + { + "bbox": [ + 55, + 479, + 296, + 635 + ], + "type": "text", + "content": " for each neighbor using a linear layer " + }, + { + "bbox": [ + 55, + 479, + 296, + 635 + ], + "type": "inline_equation", + "content": "\\phi_{w}" + }, + { + "bbox": [ + 55, + 479, + 296, + 635 + ], + "type": "text", + "content": ", followed by a softmax operation:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 102, + 639, + 295, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 639, + 295, + 654 + ], + "spans": [ + { + "bbox": [ + 102, + 639, + 295, + 654 + ], + "type": "interline_equation", + "content": "w _ {h} ^ {m j} = \\operatorname {s o f t m a x} \\left(\\phi_ {w} \\left(\\triangle P _ {h} ^ {m j}, P _ {h} ^ {m}\\right)\\right). \\tag {9}", + "image_path": "fb01f578f162b2dfaf4d0d516c2e97c736117acee9e24382d891455b3ff73f99.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 659, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 296, + 685 + ], + "type": "text", + "content": "The output state feature " + }, + { + "bbox": [ + 55, + 659, + 296, + 685 + ], + "type": "inline_equation", + "content": "\\hat{F}_h^m" + }, + { + "bbox": [ + 55, + 659, + 296, + 685 + ], + "type": "text", + "content": " for each state " + }, + { + "bbox": [ + 55, + 659, + 296, + 685 + ], + "type": "inline_equation", + "content": "P_h^m" + }, + { + "bbox": [ + 55, + 659, + 296, + 685 + ], + "type": "text", + "content": " is obtained through weighted aggregation:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 689, + 295, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 689, + 295, + 716 + ], + "spans": [ + { + "bbox": [ + 115, + 689, + 295, + 716 + ], + "type": "interline_equation", + "content": "\\hat {F} _ {h} ^ {m} = \\phi_ {c} \\left(\\sum_ {j \\in \\mathcal {N} (m)} w _ {h} ^ {m j} F _ {h} ^ {m j}\\right), \\tag {10}", + "image_path": "c473bfb856a2180589d40e93bd3053a07af21d09288503f05a2924d7d6ccd8be.jpg" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 67, + 550, + 209 + ], + "blocks": [ + { + "bbox": [ + 317, + 67, + 550, + 209 + ], + "lines": [ + { + "bbox": [ + 317, + 67, + 550, + 209 + ], + "spans": [ + { + "bbox": [ + 317, + 67, + 550, + 209 + ], + "type": "image", + "image_path": "a664a242dbbb3fd499a1477eae917724483c199649eb0fd4f42f01d624e87f4a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 210, + 555, + 265 + ], + "lines": [ + { + "bbox": [ + 313, + 210, + 555, + 265 + ], + "spans": [ + { + "bbox": [ + 313, + 210, + 555, + 265 + ], + "type": "text", + "content": "Figure 3. Overview of the pre-training framework. We pretrain the StruMamba3D (student model) using the MPM task and the spatial state consistency loss. The teacher model has the same structure as the student model, but its parameters are updated using the exponential moving average strategy." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 274, + 555, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 555, + 360 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 274, + 555, + 360 + ], + "type": "inline_equation", + "content": "F_{h}^{mj}" + }, + { + "bbox": [ + 313, + 274, + 555, + 360 + ], + "type": "text", + "content": " denotes the neighboring state features and " + }, + { + "bbox": [ + 313, + 274, + 555, + 360 + ], + "type": "inline_equation", + "content": "\\phi_c" + }, + { + "bbox": [ + 313, + 274, + 555, + 360 + ], + "type": "text", + "content": " is a linear layer. By extending the receptive field of spatial states, this lightweight convolution enhances their ability to capture global semantic information. Furthermore, we replace the causal 1D convolution in the original Mamba block with this module, making it better suited for processing input points with complex geometric structures." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 366, + 503, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 503, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 503, + 380 + ], + "type": "text", + "content": "3.4. Sequence Length-adaptive Strategy" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 384, + 554, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 554, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 554, + 431 + ], + "type": "text", + "content": "To address the issue of insufficient long-sequence memory in downstream tasks, we propose a sequence length-adaptive strategy that integrates an adaptive state update mechanism and spatial state consistency loss." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 432, + 555, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 555, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 555, + 611 + ], + "type": "text", + "content": "Adaptive State Update Mechanism. According to Eq. (2), the parameter sampling interval " + }, + { + "bbox": [ + 313, + 432, + 555, + 611 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 313, + 432, + 555, + 611 + ], + "type": "text", + "content": " controls the frequency of state updates by influencing the parameters " + }, + { + "bbox": [ + 313, + 432, + 555, + 611 + ], + "type": "inline_equation", + "content": "(\\overline{\\mathbf{A}},\\overline{\\mathbf{B}})" + }, + { + "bbox": [ + 313, + 432, + 555, + 611 + ], + "type": "text", + "content": " in SSMs. A larger " + }, + { + "bbox": [ + 313, + 432, + 555, + 611 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 313, + 432, + 555, + 611 + ], + "type": "text", + "content": " causes the model to update state features more frequently, whereas a smaller " + }, + { + "bbox": [ + 313, + 432, + 555, + 611 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 313, + 432, + 555, + 611 + ], + "type": "text", + "content": " allows the model to maintain its current state with minimal changes. To preserve memory for longer sequences in downstream tasks, we propose an adaptive state update mechanism. This mechanism adjusts the sampling interval " + }, + { + "bbox": [ + 313, + 432, + 555, + 611 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 313, + 432, + 555, + 611 + ], + "type": "text", + "content": " according to the input sequence length, allowing the model to maintain a consistent total sampling time " + }, + { + "bbox": [ + 313, + 432, + 555, + 611 + ], + "type": "inline_equation", + "content": "\\Delta_{all} = \\sum_{i=1}^{N}\\Delta_{i}" + }, + { + "bbox": [ + 313, + 432, + 555, + 611 + ], + "type": "text", + "content": " across different sequence lengths. Specifically, we design a learnable parameter " + }, + { + "bbox": [ + 313, + 432, + 555, + 611 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 432, + 555, + 611 + ], + "type": "text", + "content": ", which regulates the total sampling time " + }, + { + "bbox": [ + 313, + 432, + 555, + 611 + ], + "type": "inline_equation", + "content": "\\Delta_{all}" + }, + { + "bbox": [ + 313, + 432, + 555, + 611 + ], + "type": "text", + "content": " for sequences of varying lengths. The sampling interval " + }, + { + "bbox": [ + 313, + 432, + 555, + 611 + ], + "type": "inline_equation", + "content": "\\Delta_{i}" + }, + { + "bbox": [ + 313, + 432, + 555, + 611 + ], + "type": "text", + "content": " for each token in the inputs is calculated as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 397, + 618, + 555, + 648 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 618, + 555, + 648 + ], + "spans": [ + { + "bbox": [ + 397, + 618, + 555, + 648 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Delta} _ {i} = \\frac {\\tau \\times \\boldsymbol {\\Delta} _ {i}}{\\sum_ {i = 1} ^ {N} \\boldsymbol {\\Delta} _ {i}}. \\tag {11}", + "image_path": "11eb7739550d5a20861ac4ea28c23f44ad02495ed823cf2e22a641608b504537.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 654, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 689 + ], + "type": "text", + "content": "Based on the adaptive state update mechanism, the model can maintain a consistent total sampling time " + }, + { + "bbox": [ + 313, + 654, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\Delta_{all} = \\tau" + }, + { + "bbox": [ + 313, + 654, + 554, + 689 + ], + "type": "text", + "content": " across different sequence lengths." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "Spatial State Consistency Loss. To further ensure consistency in updated state features across different input se" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28550" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 84, + 550, + 381 + ], + "blocks": [ + { + "bbox": [ + 55, + 60, + 555, + 83 + ], + "lines": [ + { + "bbox": [ + 55, + 60, + 555, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 60, + 555, + 83 + ], + "type": "text", + "content": "Table 1. Shape Classification on ScanObjectNN and ModelNet40 Datasets. For ScanObjectNN, we report the classification accuracy(%) over the three subsets: OBJ-BG, OBJ-ONLY, and PB-T50-RS. And " + }, + { + "bbox": [ + 55, + 60, + 555, + 83 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 60, + 555, + 83 + ], + "type": "text", + "content": " indicates using simple rotational augmentation [7] for fine-tuning." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 84, + 550, + 381 + ], + "lines": [ + { + "bbox": [ + 58, + 84, + 550, + 381 + ], + "spans": [ + { + "bbox": [ + 58, + 84, + 550, + 381 + ], + "type": "table", + "html": "
MethodBackboneParam. (M)FLOPs (G)ScanObjectNNMODELNET40
OBJ-BGOBJ-ONLYPB-T50-RSw/o Votingw/ Voting
Supervised learning only
PointNet [27]-3.50.573.379.268.089.2-
PointNet++ [28]-1.51.782.384.377.990.7-
PointCNN [19]-0.6-86.185.578.592.2-
DGCNN [35]-1.82.482.886.278.192.9-
PRANet [8]-2.3---81.093.7-
PointNeXt [30]-1.43.6--87.794.0-
PointMLP [24]-12.631.4--85.494.5-
DeLA [2]-5.31.5--88.694.0-
PCM [47]-34.245.0--88.193.4-
Pre-training using single-modal information
PointBERT [42]Transformer22.14.887.4388.1283.0792.793.2
MaskPoint [22]Transformer22.14.889.3088.1084.30-93.8
PointM2AE [44]Transformer12.77.991.2288.8186.4392.993.4
PointMAE† [25]Transformer22.14.892.7791.2289.0492.793.8
PointGPT-S† [3]Transformer29.25.793.3992.4389.1793.394.0
PointMamba† [21]Mamba12.33.194.3292.6089.3193.694.1
Mamba3D† [15]Mamba16.93.993.1292.0892.0594.795.1
Ours†Structural SSM15.84.095.1893.6392.7595.195.4
Pre-training using cross-modal information
ACT† [7]Transformer22.14.893.2991.9188.2193.794.0
Joint-MAE [14]Transformer22.1-90.9488.8686.07-94.0
I2P-MAE† [46]Transformer15.3-94.1591.5790.1193.794.1
ReCon† [29]Transformer43.65.395.1893.2990.6394.594.7
", + "image_path": "274944cf8ed717b3a0232b8203858cc75e5b978d97cf646adf111f5dd6c559b0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 396, + 297, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 396, + 297, + 542 + ], + "spans": [ + { + "bbox": [ + 54, + 396, + 297, + 542 + ], + "type": "text", + "content": "quence lengths, we introduce a spatial state consistency loss to the pre-training task. As shown in Fig. 3, we present an overview of the pre-training framework. Following the previous works [15, 21, 25], we reconstruct the points in the masked regions and use the reconstruction loss " + }, + { + "bbox": [ + 54, + 396, + 297, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cd}" + }, + { + "bbox": [ + 54, + 396, + 297, + 542 + ], + "type": "text", + "content": " to guide the model to learn discriminative features from the visible points. For the spatial state consistency loss, we first use a teacher model to update the spatial states with the full input tokens " + }, + { + "bbox": [ + 54, + 396, + 297, + 542 + ], + "type": "inline_equation", + "content": "P_{x}^{f}\\in \\mathbb{R}^{N_{f}\\times 3}" + }, + { + "bbox": [ + 54, + 396, + 297, + 542 + ], + "type": "text", + "content": ". The state features " + }, + { + "bbox": [ + 54, + 396, + 297, + 542 + ], + "type": "inline_equation", + "content": "F_h^{\\prime f}" + }, + { + "bbox": [ + 54, + 396, + 297, + 542 + ], + "type": "text", + "content": " output by the teacher model serve as pseudo-labels. We then enforce a consistency loss " + }, + { + "bbox": [ + 54, + 396, + 297, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ssc}" + }, + { + "bbox": [ + 54, + 396, + 297, + 542 + ], + "type": "text", + "content": " between the updated state features of the student model " + }, + { + "bbox": [ + 54, + 396, + 297, + 542 + ], + "type": "inline_equation", + "content": "F_h^{\\prime v}" + }, + { + "bbox": [ + 54, + 396, + 297, + 542 + ], + "type": "text", + "content": " and the pseudo-labels " + }, + { + "bbox": [ + 54, + 396, + 297, + 542 + ], + "type": "inline_equation", + "content": "F_h^{\\prime f}" + }, + { + "bbox": [ + 54, + 396, + 297, + 542 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 559, + 295, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 559, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 111, + 559, + 295, + 574 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s s c} = \\operatorname {S m o o t h} \\mathrm {L} 1 \\left(F _ {h} ^ {\\prime v}, F _ {h} ^ {\\prime f}\\right). \\tag {12}", + "image_path": "f64abd076a4355413112428ae1cbfc8b95ddff24140009e026476282b5c01a41.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 591, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 296, + 653 + ], + "type": "text", + "content": "The spatial state consistency loss " + }, + { + "bbox": [ + 55, + 591, + 296, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ssc}" + }, + { + "bbox": [ + 55, + 591, + 296, + 653 + ], + "type": "text", + "content": " not only enhances the robustness of our StruMamba3D to varying input sequence lengths but also improves the ability to infer the structural information of the complete point cloud from incomplete point clouds. The final pre-training loss is:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 121, + 670, + 295, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 670, + 295, + 683 + ], + "spans": [ + { + "bbox": [ + 121, + 670, + 295, + 683 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {c d} + \\lambda \\times \\mathcal {L} _ {s s c}, \\tag {13}", + "image_path": "c682b46e6d62b28d8660ed4826b8e4f9dbe6a352c1cf8c84b8a6b68090e5e1d6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 701, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 289, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 701, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 701, + 289, + 714 + ], + "type": "text", + "content": " is the weight of the spatial state consistency loss." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 396, + 440, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 440, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 440, + 408 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 415, + 447, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 447, + 428 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 447, + 428 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 434, + 555, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 434, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 312, + 434, + 555, + 601 + ], + "type": "text", + "content": "Following the previous works [15, 21, 25], we employ " + }, + { + "bbox": [ + 312, + 434, + 555, + 601 + ], + "type": "inline_equation", + "content": "T = 12" + }, + { + "bbox": [ + 312, + 434, + 555, + 601 + ], + "type": "text", + "content": " structural SSM blocks, each with a feature dimension of " + }, + { + "bbox": [ + 312, + 434, + 555, + 601 + ], + "type": "inline_equation", + "content": "D = 384" + }, + { + "bbox": [ + 312, + 434, + 555, + 601 + ], + "type": "text", + "content": ". The number of spatial states " + }, + { + "bbox": [ + 312, + 434, + 555, + 601 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 312, + 434, + 555, + 601 + ], + "type": "text", + "content": " is set to 16. For the lightweight convolution in the structural SSM block, we set the number of neighbors to 4 for spatial states and 8 for input points. Following previous works [25, 42, 44], we pre-train the proposed model on ShapeNet dataset [1], which contains 52472 unique 3D models across 55 common object categories. Each input raw point cloud, containing " + }, + { + "bbox": [ + 312, + 434, + 555, + 601 + ], + "type": "inline_equation", + "content": "N_0 = 1024" + }, + { + "bbox": [ + 312, + 434, + 555, + 601 + ], + "type": "text", + "content": " points, is divided into 64 patches with each consisting of 32 points. A random masking ratio of 0.6 is applied, with the decoder composed of 4 standard Mamba blocks, and the " + }, + { + "bbox": [ + 312, + 434, + 555, + 601 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 312, + 434, + 555, + 601 + ], + "type": "text", + "content": " in the pre-training loss is set to 2. More details can be found in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 611, + 501, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 501, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 501, + 624 + ], + "type": "text", + "content": "4.2. Comparison on Downstream Tasks" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 629, + 554, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 664 + ], + "type": "text", + "content": "We present the experimental results of our StruMamba3D across various downstream tasks and compare them with previous state-of-the-art (SOTA) methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "Shape classification. ScanObjectNN [41] is a real-world dataset containing roughly 15000 objects across 15 categories, some of which have cluttered backgrounds or partial scans. For this dataset, we apply rotation as a data aug-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28551" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 82, + 288, + 266 + ], + "blocks": [ + { + "bbox": [ + 55, + 58, + 295, + 80 + ], + "lines": [ + { + "bbox": [ + 55, + 58, + 295, + 80 + ], + "spans": [ + { + "bbox": [ + 55, + 58, + 295, + 80 + ], + "type": "text", + "content": "Table 2. Part Segmentation on ShapeNetPart. We report the mean IoU for all classes " + }, + { + "bbox": [ + 55, + 58, + 295, + 80 + ], + "type": "inline_equation", + "content": "(\\mathrm{mIoU}_c)" + }, + { + "bbox": [ + 55, + 58, + 295, + 80 + ], + "type": "text", + "content": " and all instances " + }, + { + "bbox": [ + 55, + 58, + 295, + 80 + ], + "type": "inline_equation", + "content": "(\\mathrm{mIoU}_i)" + }, + { + "bbox": [ + 55, + 58, + 295, + 80 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 82, + 288, + 266 + ], + "lines": [ + { + "bbox": [ + 63, + 82, + 288, + 266 + ], + "spans": [ + { + "bbox": [ + 63, + 82, + 288, + 266 + ], + "type": "table", + "html": "
MethodArchitecturemIoUcmIoUi
Supervised learning only
PointNet [27]Single-scale80.483.7
PointNet++ [28]Multi-scale81.985.1
APES [36]Multi-scale83.785.8
DeLA [2]Multi-scale85.887.0
PCM [47]Multi-scale85.387.0
Pre-training using single-modal information
MaskPoint [22]Single-scale84.686.0
PointBERT [42]Single-scale84.185.6
PointMAE [25]Single-scale84.286.1
PointM2AE [44]Multi-scale84.986.5
PointGPT-S [3]Single-scale84.186.2
PointMamba [21]Single-scale84.486.2
Mamba3D [15]Single-scale83.685.6
OursSingle-scale85.086.7
", + "image_path": "a264214adad9518a83d43cb7b6caa9dc0f604ef3feeb84b61d4c597195e29a6b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 270, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 295, + 510 + ], + "type": "text", + "content": "mentation strategy [7] and use 2048 points as input. ModelNet40 [39] is a synthetic dataset comprising of 12311 clean 3D CAD models from 40 categories. For this dataset, we use 1024 points as input and apply scale and translation augmentations [27]. As shown in Tab. 1, we conduct classification experiments on both datasets. Compared to transformer-based methods, our model achieves significant performance improvements across multiple metrics on both datasets. Even when compared to the best PointGPTS [3], our model outperforms it by " + }, + { + "bbox": [ + 54, + 270, + 295, + 510 + ], + "type": "inline_equation", + "content": "3.58\\%" + }, + { + "bbox": [ + 54, + 270, + 295, + 510 + ], + "type": "text", + "content": " on ScanObjectNN and " + }, + { + "bbox": [ + 54, + 270, + 295, + 510 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 54, + 270, + 295, + 510 + ], + "type": "text", + "content": " on ModelNet40. Additionally, on the ScanObjectNN dataset, StruMamba3D surpasses PointMamba [21] by " + }, + { + "bbox": [ + 54, + 270, + 295, + 510 + ], + "type": "inline_equation", + "content": "0.86\\%" + }, + { + "bbox": [ + 54, + 270, + 295, + 510 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 270, + 295, + 510 + ], + "type": "inline_equation", + "content": "1.03\\%" + }, + { + "bbox": [ + 54, + 270, + 295, + 510 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 270, + 295, + 510 + ], + "type": "inline_equation", + "content": "3.44\\%" + }, + { + "bbox": [ + 54, + 270, + 295, + 510 + ], + "type": "text", + "content": " across the three splits, while also outperforming Mamba3D [15] by " + }, + { + "bbox": [ + 54, + 270, + 295, + 510 + ], + "type": "inline_equation", + "content": "2.06\\%" + }, + { + "bbox": [ + 54, + 270, + 295, + 510 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 270, + 295, + 510 + ], + "type": "inline_equation", + "content": "1.55\\%" + }, + { + "bbox": [ + 54, + 270, + 295, + 510 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 270, + 295, + 510 + ], + "type": "inline_equation", + "content": "0.70\\%" + }, + { + "bbox": [ + 54, + 270, + 295, + 510 + ], + "type": "text", + "content": ". Furthermore, on the ModelNet40 dataset, StruMamba3D exceeds PointMamba by " + }, + { + "bbox": [ + 54, + 270, + 295, + 510 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 54, + 270, + 295, + 510 + ], + "type": "text", + "content": " and outperforms Mamba3D by " + }, + { + "bbox": [ + 54, + 270, + 295, + 510 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 54, + 270, + 295, + 510 + ], + "type": "text", + "content": ". Overall, these results demonstrate that StruMamba3D outperforms all existing unimodal architectures, achieving multiple SOTA results and highlighting the effectiveness of structural modeling within Mamba." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": "Part Segmentation on ShapeNetPart. ShapeNetPart [41] dataset comprises 16880 models spanning 16 shape categories and annotated with 50 part labels. Compared to classification tasks, part segmentation poses a greater challenge, as it requires assigning a label to each point, making it highly dependent on the intrinsic structural information of point clouds. Tab. 8 presents our part segmentation results for all classes and instances on the ShapeNetPart dataset. For single-scale models, our method achieves substantial performance improvements, surpassing Mamba3D [15] by " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{mIoU}_c" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{mIoU}_i" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": ", and outperforming PointMamba [21] by " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{mIoU}_c" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{mIoU}_i" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": ". Notably, Mamba3D without a serialization strategy performs poorly, falling behind the baseline PointMAE [25], whereas PointMamba, which incorporates serialization, exhibits improved performance. This observation suggests the crucial role of serialization in enhancing structural mod" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 319, + 83, + 550, + 200 + ], + "blocks": [ + { + "bbox": [ + 313, + 58, + 553, + 81 + ], + "lines": [ + { + "bbox": [ + 313, + 58, + 553, + 81 + ], + "spans": [ + { + "bbox": [ + 313, + 58, + 553, + 81 + ], + "type": "text", + "content": "Table 3. Few-Shot Classification on ModelNet40. We report overall accuracy and standard deviation without a voting strategy." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 83, + 550, + 200 + ], + "lines": [ + { + "bbox": [ + 319, + 83, + 550, + 200 + ], + "spans": [ + { + "bbox": [ + 319, + 83, + 550, + 200 + ], + "type": "table", + "html": "
Method5-way10-way
10-shot20-shot10-shot20-shot
PointBERT [42]94.6±3.693.9±3.186.4±5.491.3±4.6
MaskPoint [22]95.0±3.797.2±1.791.4±4.092.7±5.1
PointMAE [25]96.3±2.597.8±1.892.6±4.193.4±3.5
PointM2AE [44]96.8±1.898.3±1.492.3±4.595.0±3.0
PointGPT-S [3]96.8±2.098.6±1.192.6±4.695.2±3.4
PointMamba [21]96.9±2.099.0±1.193.0±4.495.6±3.2
Mamba3D [15]96.4±2.298.2±1.292.4±4.195.2±2.9
Ours97.5±2.399.1±1.493.5±3.796.1±3.5
", + "image_path": "4621cba61e61882b39dfbe95122a77e25ba70090061b4f7b6f908658e4190fd4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 204, + 555, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 555, + 419 + ], + "type": "text", + "content": "eling within Mamba-based architectures. In contrast, our StruMamba3D leverages spatial states to effectively capture structural information, thereby eliminating the need for serialization and achieving " + }, + { + "bbox": [ + 313, + 204, + 555, + 419 + ], + "type": "inline_equation", + "content": "85.0\\%" + }, + { + "bbox": [ + 313, + 204, + 555, + 419 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 204, + 555, + 419 + ], + "type": "inline_equation", + "content": "\\mathrm{mIoU}_c" + }, + { + "bbox": [ + 313, + 204, + 555, + 419 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 204, + 555, + 419 + ], + "type": "inline_equation", + "content": "86.7\\%" + }, + { + "bbox": [ + 313, + 204, + 555, + 419 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 204, + 555, + 419 + ], + "type": "inline_equation", + "content": "\\mathrm{mIoU}_i" + }, + { + "bbox": [ + 313, + 204, + 555, + 419 + ], + "type": "text", + "content": ". Besides, StruMamba3D even outperforms the multi-scale self-supervised method PointM2AE [44]. Although multi-scale approaches typically yield better performance in part segmentation tasks, they often come with significantly higher computational costs (e.g., PCM [47] requires 45.0G FLOPs). In this work, we focus on designing an efficient and effective Mamba-based framework for structural modeling in point clouds. While incorporating multi-scale designs could further improve part segmentation performance, such strategies may introduce heuristic dependencies. Overall, our StruMamba3D represents a more robust and efficient Mamba-based framework for point cloud structural modeling, achieving the highest performance among all single-scale models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 420, + 556, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 556, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 556, + 588 + ], + "type": "text", + "content": "Few-shot Learning on ModelNet40. To evaluate the effectiveness of the proposed methods with limited fine-tuning data, we conduct few-shot classification experiments on the ModelNet40 dataset. Following the standard procedure [25, 42], we perform 10 independent experiments for each setting and report the mean accuracy and standard deviation. As shown in Tab. 3, our method demonstrates a significant improvement over Mamba-based methods, outperforming PointMamba [21] by " + }, + { + "bbox": [ + 313, + 420, + 556, + 588 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 313, + 420, + 556, + 588 + ], + "type": "text", + "content": " in 5-way-10-shot and " + }, + { + "bbox": [ + 313, + 420, + 556, + 588 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 313, + 420, + 556, + 588 + ], + "type": "text", + "content": " in 10-way-10-shot settings, and surpassing Mamba3D [15] by " + }, + { + "bbox": [ + 313, + 420, + 556, + 588 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 313, + 420, + 556, + 588 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 420, + 556, + 588 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 313, + 420, + 556, + 588 + ], + "type": "text", + "content": ", respectively. These results indicate that our method can efficiently adapt to downstream tasks with longer input sequences, even under the constraint of limited training data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 598, + 408, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 598, + 408, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 598, + 408, + 611 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 616, + 553, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 553, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 553, + 652 + ], + "type": "text", + "content": "To explore the architectural design, we perform ablation studies on the PB-T50-RS split of ScanObjectNN dataset, as well as ModelNet40 and ShapeNetPart datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "content": "Effect of Designed Modules. As shown in Tab. 4, we conduct an ablation study to evaluate the effectiveness of two core modules: the Structural SSM block and the sequence length-adaptive strategy. The baseline model is built upon the standard Mamba [9] block, following the same" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28552" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 83, + 288, + 147 + ], + "blocks": [ + { + "bbox": [ + 55, + 60, + 295, + 82 + ], + "lines": [ + { + "bbox": [ + 55, + 60, + 295, + 82 + ], + "spans": [ + { + "bbox": [ + 55, + 60, + 295, + 82 + ], + "type": "text", + "content": "Table 4. Effect of the designed modules. 'SLAS' indicates the sequence length-adaptive strategy used in pre-training phase." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 83, + 288, + 147 + ], + "lines": [ + { + "bbox": [ + 61, + 83, + 288, + 147 + ], + "spans": [ + { + "bbox": [ + 61, + 83, + 288, + 147 + ], + "type": "table", + "html": "
Structural SSM BlockSLASOverall AccuracymIoUc
ScanNNMN40SNPart
XX87.2391.8681.56
✓X92.0994.4584.49
✓✓92.7595.0684.96
", + "image_path": "8aaeb6bba98e9426eaf9522afea08b40f597ede92ef618dcc13d733f1249ee04.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 61, + 175, + 288, + 259 + ], + "blocks": [ + { + "bbox": [ + 55, + 150, + 295, + 173 + ], + "lines": [ + { + "bbox": [ + 55, + 150, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 55, + 150, + 295, + 173 + ], + "type": "text", + "content": "Table 5. Ablation on Structural SSM Block. The baseline uses the standard Mamba block to replace the Structural SSM Block." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 175, + 288, + 259 + ], + "lines": [ + { + "bbox": [ + 61, + 175, + 288, + 259 + ], + "spans": [ + { + "bbox": [ + 61, + 175, + 288, + 259 + ], + "type": "table", + "html": "
MethodOverall AccuracymIoUc
ScanNNMN40SNPart
baseline88.2492.5082.08
w/ Structural SSM91.7893.9284.15
w/ Lightweight Conv. for h92.2294.6584.62
w/ Lightweight Conv. for x92.4094.8184.77
w/ Bidirectional Scanning92.7595.0684.96
", + "image_path": "7a44d0d8d3cd0c72266fe7b0c65779a28c0ed1f7537b6698d59e3a05d0663a6a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 267, + 295, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 267, + 295, + 387 + ], + "spans": [ + { + "bbox": [ + 54, + 267, + 295, + 387 + ], + "type": "text", + "content": "architecture as PointMAE [25]. Compared to this baseline, the Structural SSM block significantly enhances model performance, yielding improvements of " + }, + { + "bbox": [ + 54, + 267, + 295, + 387 + ], + "type": "inline_equation", + "content": "4.86\\%" + }, + { + "bbox": [ + 54, + 267, + 295, + 387 + ], + "type": "text", + "content": " on ScanObjectNN, " + }, + { + "bbox": [ + 54, + 267, + 295, + 387 + ], + "type": "inline_equation", + "content": "2.59\\%" + }, + { + "bbox": [ + 54, + 267, + 295, + 387 + ], + "type": "text", + "content": " on ModelNet40, and " + }, + { + "bbox": [ + 54, + 267, + 295, + 387 + ], + "type": "inline_equation", + "content": "2.93\\%" + }, + { + "bbox": [ + 54, + 267, + 295, + 387 + ], + "type": "text", + "content": " on ShapeNetPart, demonstrating its effectiveness in modeling spatial dependencies and capturing richer structural information in point clouds. Furthermore, integrating the sequence length-adaptive strategy further enhances performance by improving the generalization to varying input lengths, thereby ensuring optimal performance across downstream tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 388, + 295, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 295, + 532 + ], + "type": "text", + "content": "Effect of Structural SSM Block. As shown in Tab. 5, we further validate the effectiveness of individual components within the Structural SSM block. Starting from the baseline model with the standard Mamba block, we incrementally incorporate key modules from the Structural SSM block. The most significant performance gains stem from the Structural SSM and the lightweight convolution designed for spatial states, demonstrating that modeling structural information through hidden states in SSM is crucial. Moveover, the lightweight convolution for input points and the bidirectional scanning mechanism also enhance the performance across downstream tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "text", + "content": "Effect of Structural SSM. Structural SSM is a crucial component of our method, as it effectively captures the structural information of point clouds through spatial states. To further analyze the impact of spatial state initialization and state-wise update strategy, we conduct ablation studies as shown in Tab. 6. Although spatial state initialization provides the states with positional information, the improvement is minimal, as Eq. (2) does not consider spatial relationships between input points and states. When integrating the state-wise update strategy, the model achieves a significant performance boost, with improvements of " + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "inline_equation", + "content": "1.42\\%" + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "text", + "content": " on ScanObjectNN, " + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "inline_equation", + "content": "0.94\\%" + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "text", + "content": " on ModelNet40, and " + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "inline_equation", + "content": "1.09\\%" + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "text", + "content": " on ShapeNetPart. The results indicate that incorporating spatial relationships to control state information updates and propagation within the SSM is essential for effectively mod" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 326, + 106, + 541, + 179 + ], + "blocks": [ + { + "bbox": [ + 313, + 60, + 553, + 104 + ], + "lines": [ + { + "bbox": [ + 313, + 60, + 553, + 104 + ], + "spans": [ + { + "bbox": [ + 313, + 60, + 553, + 104 + ], + "type": "text", + "content": "Table 6. Ablation on Structural SSM. 'SS' represents the use of spatial states in the SSM. " + }, + { + "bbox": [ + 313, + 60, + 553, + 104 + ], + "type": "inline_equation", + "content": "\\phi (x)" + }, + { + "bbox": [ + 313, + 60, + 553, + 104 + ], + "type": "text", + "content": " denotes the SSM parameters depending on the point feature, while " + }, + { + "bbox": [ + 313, + 60, + 553, + 104 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}(\\triangle P)" + }, + { + "bbox": [ + 313, + 60, + 553, + 104 + ], + "type": "text", + "content": " represents the SSM parameters depending on the spatial relationship." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 106, + 541, + 179 + ], + "lines": [ + { + "bbox": [ + 326, + 106, + 541, + 179 + ], + "spans": [ + { + "bbox": [ + 326, + 106, + 541, + 179 + ], + "type": "table", + "html": "
φ(x)SSMLP(△P)Overall AccuracymIoUc
ScanNNMN40SNPart
✓XX90.9493.8483.62
✓✓X91.3394.1283.87
✓✓✓92.7595.0684.96
X✓✓91.1294.2584.02
", + "image_path": "4b0bb4422522347265d280ee0371273a09e1e0f24ee9c828a177227f5c585522.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 329, + 216, + 538, + 312 + ], + "blocks": [ + { + "bbox": [ + 313, + 182, + 553, + 214 + ], + "lines": [ + { + "bbox": [ + 313, + 182, + 553, + 214 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 553, + 214 + ], + "type": "text", + "content": "Table 7. Ablation on sequence length-adaptive strategy. 'ASUM' indicates the adaptive state update mechanism, and " + }, + { + "bbox": [ + 313, + 182, + 553, + 214 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 182, + 553, + 214 + ], + "type": "text", + "content": " ’ indicates the weight of the spatial state consistency loss " + }, + { + "bbox": [ + 313, + 182, + 553, + 214 + ], + "type": "inline_equation", + "content": "{\\mathcal{L}}_{ssc}" + }, + { + "bbox": [ + 313, + 182, + 553, + 214 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 329, + 216, + 538, + 312 + ], + "lines": [ + { + "bbox": [ + 329, + 216, + 538, + 312 + ], + "spans": [ + { + "bbox": [ + 329, + 216, + 538, + 312 + ], + "type": "table", + "html": "
ASUMLsscOverall Accuracy
ScanNNMN40SNPart
XX92.0994.4584.49
✓X92.2694.5784.56
✓λ = 192.4794.9484.81
✓λ = 292.7595.0684.96
✓λ = 592.5794.8984.77
Xλ = 292.1694.6584.56
", + "image_path": "fb65804a6a29c96ac01b82c7cc69017a76755aed2874420b13e8558456e3daf6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 318, + 553, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 553, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 553, + 378 + ], + "type": "text", + "content": "eling structural information. Moreover, we observe that relying solely on spatial relationships to generate SSM parameters results in performance degradation. This is likely due to the varying importance of point features, which cannot be fully captured by spatial position alone." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 379, + 554, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 554, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 554, + 559 + ], + "type": "text", + "content": "Effect of Sequence Length-Adaptive Strategy. The sequence length-adaptive strategy consists of two key components: the adaptive state update mechanism and the spatial state consistency loss. Tab. 7 presents an ablation study on these components. When applied individually, both the adaptive state update mechanism and the spatial state consistency loss yield only marginal performance gains. However, when used together, they lead to a significant improvement, demonstrating their complementary effects. The adaptive state update mechanism ensures a consistent total sampling time across different input lengths, while the spatial state consistency loss ensures the consistency of spatial state features. In addition, we perform an ablation study on the weight " + }, + { + "bbox": [ + 313, + 379, + 554, + 559 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 379, + 554, + 559 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 379, + 554, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ssc}" + }, + { + "bbox": [ + 313, + 379, + 554, + 559 + ], + "type": "text", + "content": " and observe that the model achieves optimal performance when " + }, + { + "bbox": [ + 313, + 379, + 554, + 559 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 379, + 554, + 559 + ], + "type": "text", + "content": " is set to 2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 566, + 388, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 566, + 388, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 566, + 388, + 578 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "We identify two key issues limiting Mamba potential: 3D point adjacency distortion during SSM processing and inadequate long-sequence memory retention in downstream tasks. To address these issues, we propose StruMamba3D with two core components. The structural SSM block maintains spatial dependencies during state selection and propagation, while the sequence length-adaptive strategy reduces the sensitivity of Mamba-based models to varying input lengths. Our method achieves substantial improvements across four downstream tasks, with ablation studies confirming the effectiveness of each module." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28553" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 160, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 160, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 160, + 85 + ], + "type": "text", + "content": "6. Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 86, + 295, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 86, + 295, + 110 + ], + "spans": [ + { + "bbox": [ + 55, + 86, + 295, + 110 + ], + "type": "text", + "content": "This work is supported by the National Defense Basic Scientific Research program (JCKY2022911B002)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 121, + 115, + 134 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 121, + 115, + 134 + ], + "spans": [ + { + "bbox": [ + 56, + 121, + 115, + 134 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 142, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 62, + 142, + 295, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 142, + 295, + 195 + ], + "spans": [ + { + "bbox": [ + 62, + 142, + 295, + 195 + ], + "type": "text", + "content": "[1] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6, 11" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 198, + 295, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 198, + 295, + 230 + ], + "spans": [ + { + "bbox": [ + 61, + 198, + 295, + 230 + ], + "type": "text", + "content": "[2] Binjie Chen, Yunzhou Xia, Yu Zang, Cheng Wang, and Jonathan Li. Decoupled local aggregation for point cloud learning, 2023. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 232, + 295, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 232, + 295, + 275 + ], + "spans": [ + { + "bbox": [ + 62, + 232, + 295, + 275 + ], + "type": "text", + "content": "[3] Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. Pointgpt: Auto-regressively generative pre-training from point clouds. arXiv preprint arXiv:2305.11487, 2023. 3, 6, 7, 12, 13" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 277, + 295, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 277, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 62, + 277, + 295, + 331 + ], + "type": "text", + "content": "[4] Siheng Chen, Baoan Liu, Chen Feng, Carlos Vallespi-Gonzalez, and Carl Wellington. 3d point cloud processing and learning for autonomous driving: Impacting map creation, localization, and perception. IEEE Signal Processing Magazine, 38(1):68-86, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 333, + 295, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 333, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 62, + 333, + 295, + 365 + ], + "type": "text", + "content": "[5] Tri Dao and Albert Gu. Transformers are ssms: Generalized models and efficient algorithms through structured state space duality. arXiv preprint arXiv:2405.21060, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 367, + 295, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 367, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 62, + 367, + 295, + 421 + ], + "type": "text", + "content": "[6] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT (1), pages 4171-4186. Association for Computational Linguistics, 2019. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 422, + 295, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 422, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 422, + 295, + 487 + ], + "type": "text", + "content": "[7] Runpei Dong, Zekun Qi, Linfeng Zhang, Junbo Zhang, Jianjian Sun, Zheng Ge, Li Yi, and Kaisheng Ma. Autoencoders as cross-modal teachers: Can pretrained 2d image transformers help 3d representation learning? In The Eleventh International Conference on Learning Representations (ICLR), 2023. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 490, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 490, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 490, + 295, + 544 + ], + "type": "text", + "content": "[8] Deng-Ping Fan, Ge-Peng Ji, Tao Zhou, Geng Chen, Huazhu Fu, Jianbing Shen, and Ling Shao. Pranet: Parallel reverse attention network for polyp segmentation. In International conference on medical image computing and computer-assisted intervention, pages 263-273. Springer, 2020. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 545, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 545, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 545, + 295, + 578 + ], + "type": "text", + "content": "[9] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2, 3, 7, 14" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 580, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 580, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 580, + 295, + 612 + ], + "type": "text", + "content": "[10] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 614, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 295, + 667 + ], + "type": "text", + "content": "[11] Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, and Christopher Ré. Combining recurrent, convolutional, and continuous-time models with linear state space layers. Advances in neural information processing systems, 34:572-585, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[12] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7:187-199, 2021. 2,3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[13] Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3d point clouds: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(12):4338-4364, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 118, + 554, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 160 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 160 + ], + "type": "text", + "content": "[14] Ziyu Guo, Renrui Zhang, Longtian Qiu, Xianzhi Li, and Pheng-Ann Heng. Joint-mae: 2d-3d joint masked autoencoders for 3d point cloud pre-training. arXiv preprint arXiv:2302.14007, 2023.6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 162, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 554, + 217 + ], + "type": "text", + "content": "[15] Xu Han, Yuan Tang, Zhaoxuan Wang, and Xianzhi Li. Mamba3d: Enhancing local features for 3d point cloud analysis via state space model. In Proceedings of the 32nd ACM International Conference on Multimedia, page 4995-5004. Association for Computing Machinery, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 217, + 554, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 554, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 554, + 270 + ], + "type": "text", + "content": "[16] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 272, + 554, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 554, + 303 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 554, + 303 + ], + "type": "text", + "content": "[17] Rudolph Emil Kalman. A new approach to linear filtering and prediction problems. Journal of Basic Engineering, 1960. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 305, + 554, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 554, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 554, + 337 + ], + "type": "text", + "content": "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 11, 12" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 338, + 554, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 554, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 554, + 381 + ], + "type": "text", + "content": "[19] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. Advances in neural information processing systems, 31, 2018. 1, 2, 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 383, + 554, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 554, + 436 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 554, + 436 + ], + "type": "text", + "content": "[20] Ying Li, Lingfei Ma, Zilong Zhong, Fei Liu, Michael A Chapman, Dongpu Cao, and Jonathan Li. Deep learning for lidar point clouds in autonomous driving: A review. IEEE Transactions on Neural Networks and Learning Systems, 32 (8):3412-3432, 2020. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 438, + 554, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 554, + 491 + ], + "type": "text", + "content": "[21] Dingkang Liang, Xin Zhou, Wei Xu, Xingkui Zhu, Zhikang Zou, Xiaqing Ye, Xiao Tan, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. In Advances in Neural Information Processing Systems, 2024. 2, 3, 6, 7, 12, 13" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 493, + 554, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 554, + 536 + ], + "type": "text", + "content": "[22] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In European Conference on Computer Vision, pages 657-675. Springer, 2022. 2, 3, 6, 7, 13" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 537, + 554, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 554, + 568 + ], + "type": "text", + "content": "[23] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 11, 12" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 570, + 554, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 570, + 554, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 570, + 554, + 613 + ], + "type": "text", + "content": "[24] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. arXiv preprint arXiv:2202.07123, 2022. 3, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 614, + 554, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 554, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 554, + 668 + ], + "type": "text", + "content": "[25] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In European conference on computer vision, pages 604-621. Springer, 2022. 2, 3, 5, 6, 7, 8, 11" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "type": "text", + "content": "[26] François Pomerleau, Francis Colas, Roland Siegwart, et al. A review of point cloud registration algorithms for mobile robotics. Foundations and Trends® in Robotics, 4(1):1-104, 2015. 1" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28554" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[27] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 174 + ], + "type": "text", + "content": "[28] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 1, 2, 3, 6, 7, 11" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 229 + ], + "type": "text", + "content": "[29] Zekun Qi, Runpei Dong, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Contrast with reconstruct: Contrastive 3d representation learning guided by generative pretraining. In International Conference on Machine Learning (ICML), 2023. 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 285 + ], + "type": "text", + "content": "[30] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. Advances in neural information processing systems, 35:23192-23204, 2022. 1, 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 294, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 351 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 351 + ], + "type": "text", + "content": "[31] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 353, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 294, + 407 + ], + "type": "text", + "content": "[32] Chuxin Wang, Jiacheng Deng, Jianfeng He, Tianzhu Zhang, Zhe Zhang, and Yongdong Zhang. Long-short range adaptive transformer with dynamic sampling for 3d object detection. IEEE Transactions on Circuits and Systems for Video Technology, 33(12):7616-7629, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 409, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 294, + 453 + ], + "type": "text", + "content": "[33] Chuxin Wang, Yixin Zha, Jianfeng He, Wenfei Yang, and Tianzhu Zhang. Rethinking masked representation learning for 3d point cloud understanding. IEEE Transactions on Image Processing, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 455, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 294, + 498 + ], + "type": "text", + "content": "[34] ChuXin Wang, Wenfei Yang, Xiang Liu, and Tianzhu Zhang. State space model meets transformer: A new paradigm for 3d object detection. In The Thirteenth International Conference on Learning Representations, 2025. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 500, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 294, + 544 + ], + "type": "text", + "content": "[35] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 2, 3, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "type": "text", + "content": "[36] Chengzhi Wu, Junwei Zheng, Julius Pfrommer, and Jürgen Beyerer. Attention-based point cloud edge sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5333-5343, 2023. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 590, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 634 + ], + "type": "text", + "content": "[37] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. Advances in Neural Information Processing Systems, 35:33330-33342, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[38] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4840-4851, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "text", + "content": "[39] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 642 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 173 + ], + "type": "text", + "content": "[40] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 574-591. Springer, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "text", + "content": "[41] Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Transactions on Graphics (ToG), 35(6):1-12, 2016. 6, 7, 12" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 230, + 555, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 555, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 555, + 284 + ], + "type": "text", + "content": "[42] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19313-19322, 2022. 2, 3, 6, 7, 11, 13" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 285, + 553, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 553, + 328 + ], + "type": "text", + "content": "[43] Yixin Zha, Chuxin Wang, Wenfei Yang, and Tianzhu Zhang. Exploring semantic masked autoencoder for self-supervised point cloud understanding. arXiv preprint arXiv:2506.21957, 2025." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 331, + 553, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 384 + ], + "type": "text", + "content": "[44] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training. Advances in neural information processing systems, 35:27061-27074, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 386, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 386, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 386, + 553, + 441 + ], + "type": "text", + "content": "[45] Renrui Zhang, Liuhui Wang, Yu Qiao, Peng Gao, and Hongsheng Li. Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21769-21780, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 442, + 553, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 442, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 316, + 442, + 553, + 496 + ], + "type": "text", + "content": "[46] Renrui Zhang, Liuhui Wang, Yu Qiao, Peng Gao, and Hongsheng Li. Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21769-21780, 2023. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 498, + 553, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 540 + ], + "type": "text", + "content": "[47] Tao Zhang, Xiangtai Li, Haobo Yuan, Shunping Ji, and Shuicheng Yan. Point cloud mamba: Point cloud learning via state space model. arXiv preprint arXiv:2403.00762, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 543, + 553, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 543, + 553, + 596 + ], + "spans": [ + { + "bbox": [ + 316, + 543, + 553, + 596 + ], + "type": "text", + "content": "[48] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3d features on any point-cloud. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10252-10263, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 598, + 555, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 598, + 555, + 642 + ], + "spans": [ + { + "bbox": [ + 316, + 598, + 555, + 642 + ], + "type": "text", + "content": "[49] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28555" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Structure Matters_ Revisiting Boundary Refinement in Video Object Segmentation/bf659278-d12b-4982-824e-fbf802268f8f_content_list.json b/2025/Structure Matters_ Revisiting Boundary Refinement in Video Object Segmentation/bf659278-d12b-4982-824e-fbf802268f8f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..958e6ab5cbe50a50a3f1e5434fdf4dd623c0975d --- /dev/null +++ b/2025/Structure Matters_ Revisiting Boundary Refinement in Video Object Segmentation/bf659278-d12b-4982-824e-fbf802268f8f_content_list.json @@ -0,0 +1,1737 @@ +[ + { + "type": "text", + "text": "Structure Matters: Revisiting Boundary Refinement in Video Object Segmentation", + "text_level": 1, + "bbox": [ + 89, + 130, + 906, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Guanyi Qin $^{1,*}$ , Ziyue Wang $^{1,*}$ , Daiyun Shen $^{1}$ , Haofeng Liu $^{1}$ , Hantao Zhou $^{2}$ , Junde Wu $^{3}$ , Runze Hu $^{4}$ , Yueming Jin $^{1,\\dagger}$", + "bbox": [ + 202, + 179, + 792, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ National University of Singapore; $^{2}$ Tsinghua University", + "bbox": [ + 274, + 215, + 722, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3University of Oxford; 4Beijing Institute of Technology", + "bbox": [ + 277, + 234, + 718, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{guanyi.qin, e1374378, e1374467, haofeng.liu}@u.nus.edu, hrz@bit.edu.cn, ymjin@nus.edu.sg", + "bbox": [ + 102, + 253, + 888, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 304, + 326, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Given an object mask, Semi-supervised Video Object Segmentation (SVOS) technique aims to track and segment the object across video frames, serving as a fundamental task in computer vision. Although recent memory-based methods demonstrate potential, they often struggle with scenes involving occlusion, particularly in handling object interactions and high feature similarity. To address these issues and meet the real-time processing requirements of downstream applications, in this paper, we propose a novel bOundary Amendment video object Segmentation method with Inherent Structure refinement, hereby named OASIS. Specifically, a lightweight structure refinement module is proposed to enhance segmentation accuracy. With the fusion of rough edge priors captured by the Canny filter and stored object features, the module can generate an object-level structure map and refine the representations by highlighting boundary features. Evidential learning for uncertainty estimation is introduced to further address challenges in occluded regions. The proposed method, OASIS, maintains an efficient design, yet extensive experiments on challenging benchmarks demonstrate its superior performance and competitive inference speed compared to other state-of-the-art methods, i.e., achieving the $\\mathcal{F}$ values of 91.6 (vs. 89.7 on DAVIS-17 validation set) and $\\mathcal{G}$ values of 86.6 (vs. 86.2 on YouTubeVOS 2019 validation set) while maintaining a competitive speed of 48 FPS on DAVIS. Checkpoints, logs, and code will be available here.", + "bbox": [ + 88, + 335, + 485, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 758, + 220, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video Object Segmentation (VOS) is an important task in computer vision [1, 2] that aims to separate moving objects from the background. In the general semi-supervised VOS (SVOS) setting, the task involves propagating the object masks specified in the first frame throughout the entire video. By enabling effective identification and understanding of", + "bbox": [ + 89, + 785, + 483, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5f514942aa9a78bd0b912b63ff997387c737f87c38383aff6b2844fd36e17277.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 558, + 314, + 704, + 375 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/168a0079287f9985ae3a47929dae1a363969b8888bdc53a09b7a362a302dbe27.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 735, + 318, + 846, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5df03c903ecf7ffd778f058e2a40cad3a2ae4b68ef5bb4303fe15c59d355339b.jpg", + "image_caption": [ + "Figure 1. (a) Input frame. (b) Object mask generated by the SOTA method Cutie. (c) Output structure map. (d) Object mask generated by our model. In this case, the fish corresponding to the red mask closely resembles the background and is concealed behind the other two fish, causing previous methods to fail. Guided by our structure map, the model successfully segments the fish accurately." + ], + "image_footnote": [], + "bbox": [ + 568, + 404, + 689, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/21333945ab59170b581462df22157f2f5ef1aef0a6272e209bf93b052466d2d5.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 735, + 405, + 846, + 455 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "objects in videos, VOS plays a crucial role in various applications including autonomous driving, robotics, video editing, medical intervention, etc. Albeit performance demands, real-time processing is also essential for these applications, necessitating algorithms that accurately perceive and segment objects while delivering sufficient inference speed.", + "bbox": [ + 511, + 566, + 908, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements [3-6] in SVOS emphasize the value of memory-based mechanisms. Representations of certain past frames are generated and stored in a memory bank. The current frame is then used as the query to match these stored representations and derive segmentation masks. Cutie [7], further proposes object queries as a high-level summary of the target object and interacts with the pixel-level features of the query frame for accurate segmentation, achieving the state-of-the-art performance. However, these approaches still struggle when similar objects are close to each other or are occluded, as shown in Figure 1 (a) and (b). On one hand, the mask prediction depends predominantly on maintaining consistency across frames for both pixel- and object-level features, while the inherent geometric characteristics of the objects are often overlooked. This limitation poses challenges in accurately discerning object structures, particularly", + "bbox": [ + 511, + 659, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution, † Corresponding author.", + "bbox": [ + 107, + 886, + 352, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14431", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "when they resemble the background (red box). On the other hand, when occlusion occurs, objects become fragmented. This challenge is amplified when objects share similar semantics, leading to high uncertainty in overlapping areas. Without an understanding of object structure, the model struggles to distinguish objects and establish clear boundaries (blue box).", + "bbox": [ + 89, + 90, + 485, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Edge information is a crucial component of an object's structure representation, defining its boundaries and shape by representing the inherent structural information. Precise edge prediction allows models to capture structural details more effectively, thereby enhancing the model's ability to understand occlusions. Therefore, edges have been widely used for segmenting objects for static images across various fields including object detection and semantic segmentation [8-10]. However, in VOS tasks, target movements lead to considerable shifts in object edges over time, making it challenging to predict accurate edges from scratch for each frame. This often requires large edge detection networks, which significantly reduce inference efficiency. Therefore, edges are only used to analyze motion magnitude or shadow area detection primarily, which merely requires coarse edges [11-13].", + "bbox": [ + 89, + 199, + 485, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle the aforementioned problems, we propose an bOuntary Amendment video object Segmentation with Inherent Structure refinement, hereby named OASIS, which effectively refines object structures to handle geometric and layer relationships among objects better, while highlighting them out of backgrounds. This is achieved through a simple but effective structure refinement module. Specifically, we first extract the rough edge information by utilizing a fast and robust edge detection algorithm, Canny Edge Detection [14], which can identify global edges in individual frames and provide strong priors. With these priors, we only need a lightweight structure decoder to obtain the edges of the target objects instead of predicting from scratch without significant sacrifice in inference speed. This map further amplifies the original features by highlighting objects' boundaries, producing structure-enhanced features for final segmentation. This process enables the model to differentiate among objects and from the background effectively. Additionally, an object feature fusion mechanism is introduced to integrate memory features into the structure map to better model relationships between objects. Furthermore, to address the high uncertainty in dealing with overlapped areas, we also make the first attempt to deliver a new learning pattern that incorporates Evidential Learning (EDL) for uncertainty quantification and introduces an additional loss to minimize uncertainty and improve segmentation accuracy in occlusion regions.", + "bbox": [ + 89, + 444, + 483, + 837 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments demonstrate the superiority of our approach across four benchmarks under two training settings, particularly in scenes where target objects interact and exhibit complex relationships. Notably, OASIS achieves", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "promising results, consistently outperforming current state-of-the-art methods across different datasets, while delivering a real-time inference speed at 48 fps. In summary, our contributions are the following:", + "bbox": [ + 511, + 90, + 906, + 151 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose OASIS, a novel framework designed to manage relationships between the objects and the background by accurately modeling object structures.", + "- We introduce a structure refinement module and an object sensory mechanism that utilizes edge information to achieve inherent structure refinement.", + "- To the best of our knowledge, we make the very first attempt to incorporate evidential learning in the SVOS task to improve the model performance on high-uncertainty regions, particularly in areas with significant similarity and occlusion.", + "- We verify OASIS on multiple SVOS benchmarks involving a wide range of video contents, annotation conditions, and training settings. OASIS outperforms other competitors across all these datasets and settings." + ], + "bbox": [ + 511, + 152, + 906, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 393, + 653, + 410 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Video Object Segmentation", + "text_level": 1, + "bbox": [ + 511, + 419, + 759, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early approaches of VOS [1, 2, 15-21] implement online learning that finetunes the model at test-time, resulting in low efficiency. Propagation-based methods [20, 22-30] iteratively propagate masks between adjacent frames by leveraging temporal correlations; however, they are inevitably constrained by drift and occlusion. Matching-based models [31-40] compute correspondences between the current and reference frames to capture long-range temporal context. Some memory-based methods [3, 26, 35-37, 39, 41-47] further enhance this by using memory banks to store features from past frames for pixel-level matching. The lately research SAM2 [48] also benefitted from memory mechanisms. Recently, Cutie [7] proposed an object-reading mechanism to enhance object-level representations. While it effectively perceives individual objects, modeling the relationships between multiple objects remains an open challenge.", + "bbox": [ + 511, + 441, + 908, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Edge Detection", + "text_level": 1, + "bbox": [ + 511, + 696, + 666, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Edge information has been widely utilized to differentiate occluded objects [8-10, 49-51]. In salient object detection [52-57], edge features assist in identifying the most visually prominent objects within the image. Similarly, in segmentation tasks [11, 12, 58-61], edge detection helps to distinguish different regions and instances. Accurate localization of moving objects in videos remains a significant challenge, with edge information playing a crucial role in mitigating this difficulty. [11, 12, 61-64]. However, these methods require complex network designs like extra convolution layers, greatly affecting inference efficiency. Canny edge detection [14, 65-67] is a popular algorithm for identifying sharp in", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14432", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/19a04cc1082cdfd93132d8a14f5e06fe7cc4dc8d74ee6d040b6bad68c241edb9.jpg", + "image_caption": [ + "Figure 2. Overview of our proposed method: the left side shows the overall framework, while the right side provides a detailed illustration of the structure refinement process. The light grey box on the left indicates evidential learning, and \"Dist.\" is an abbreviation for \"Distribution.\"" + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 566, + 324 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/47723baed477e527c48954574fe00432e1d8e375bcda66189bd7762e2f466e32.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 570, + 88, + 883, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tensity changes in images, which can help identify the edges of all regions, making it a valuable prior for enhancing edge prediction. [68-71].", + "bbox": [ + 89, + 367, + 482, + 414 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Video Segmentation with Edge Information", + "text_level": 1, + "bbox": [ + 89, + 439, + 460, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Edge information can effectively capture object movements, offering significant potential for advancing video object segmentation. Studies such as [11, 12] demonstrated that edge information can aid in analyzing motion magnitude in videos. [61] further enhances this by combining flow edges with optical flow to track moving objects. Moreover, edge information has been utilized for shadow segmentation in videos [72], while [13] leverages edge information to derive discriminative representations, facilitating the differentiation of various semantic regions. However, these approaches either operate solely at the image level without distinguishing objects or are unable to generate fine-grained segmentation results for object tracking, rendering them unsuitable for semi-supervised video object segmentation.", + "bbox": [ + 88, + 467, + 482, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.4. Uncertainty Estimation", + "text_level": 1, + "bbox": [ + 89, + 705, + 307, + 722 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Uncertainty estimation involves predicting the level of confidence or uncertainty in a network's output, encompassing aleatoric uncertainty (arising from data noise) and epistemic uncertainty (stemming from model limitations) [73]. Early approaches employed methods like particle filtering and CRFs [74, 75]. With the rise of deep learning, Bayesian neural networks [76-78] were introduced, along with other techniques such as ensemble methods [79]. More recently, evidential learning [73, 80-82] has shown promising results, which has the potential to improve the high uncertainty of occlusion areas in VOS tasks.", + "bbox": [ + 89, + 734, + 485, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 364, + 650, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Inherent Structure-guided SVOS", + "text_level": 1, + "bbox": [ + 511, + 391, + 803, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Formulation Given an object mask for reference, existing memory-based SVOS methods typically learn to segment objects sequentially across frames while concurrently updating the model memory for future regards. This procedure can be expressed as:", + "bbox": [ + 511, + 414, + 906, + 491 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq _ {t} = h \\left(\\varphi \\left(i _ {t}\\right), y _ {t}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 647, + 503, + 769, + 520 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ny _ {t} = \\varrho \\left(i _ {t - 1}, \\dots , i _ {t - n} \\mid q _ {t - 1}, \\dots , q _ {t - n}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 515, + 905, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $i_{t}$ denotes the input frame and $y_{t}$ is the constructed memory from previous $n$ frames and predictions, acting as reference at time $t$ . Functions $\\varphi$ and $\\varrho$ represent the feature extractor and memory constructor, respectively. The notation $h$ refers to the object segmentation network, which utilizes frame features $\\varphi(i_{t})$ and memory $y$ to produce the segmentation output $q$ at time $t$ . This pipeline interprets the SVOS task from the perspective of the frame content.", + "bbox": [ + 511, + 551, + 905, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, we observe that Eq. 1 lacks adequate modeling of structural information and hierarchy relationships. To enhance the capacity of SVOS and better address issues of occlusion and similarity-induced confusion, we introduce a novel structure refinement module to the existing framework. This module provides the model with inherent structural information and hierarchical relationships in each input frame. Fig. 2 shows the overall architecture of our proposed model.", + "bbox": [ + 511, + 672, + 908, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, given an input frame $i_{t}$ , we first derive its edge representation using the Canny filter $c$ , a robust algorithm known for accurately detecting edges despite the noise. The frame features $\\varphi(i_{t})$ then interact with the edge representation to generate an inherent structure map $s$ under the guidance of the memory features, capturing key structural characteristics of target objects. This inherent structure map", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14433", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "is subsequently integrated back into the frame features. Finally, we align memory features and structured enhanced features to predict the target object mask. Accordingly, our proposed solution can be formulated as:", + "bbox": [ + 89, + 90, + 485, + 151 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} q _ {t} = h (\\varphi (i _ {t}), y _ {t}, s _ {t}), \\\\ s _ {t} = \\vartheta (\\varphi (i _ {t}), c (i _ {t}), y _ {t}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 161, + 367, + 194 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\vartheta$ denotes the fusion function for processing features and inherent structure map. By explicitly modeling structural information, our approach allows the model to better distinguish objects, even in challenging scenarios involving occlusion and similar semantics.", + "bbox": [ + 89, + 202, + 483, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Inherent Structure Map Generation", + "text_level": 1, + "bbox": [ + 89, + 286, + 403, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Edge Highlighting Given an input RGB image $I \\in \\mathbb{R}^{3 \\times H \\times W}$ , we first derive its edge representation $I_{c} \\in \\mathbb{R}^{1 \\times H \\times W}$ using the Canny filter. Next, $I$ is fed into an image encoder for multi-scale feature extraction. The output of the image encoder can be formulated as $\\hat{I}_{i} \\in \\mathbb{R}^{c_{i} \\times h_{i} \\times w_{i}}$ for $i = 1, \\dots, N$ , where $N$ denotes the number of feature scales, $h_{i} = \\frac{H}{2^{i+1}}$ and $w_{i} = \\frac{W}{2^{i+1}}$ .", + "bbox": [ + 89, + 308, + 483, + 414 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the image encoder is pre-trained on ImageNet [83], its output feature map tends to emphasize abstract, content-focused information. In the case of SVOS, this nature may impede the model to accurately segment objects, potentially leading to confusion when occlusions occur among fragmented objects. To enhance the sensitivity of model to edges and structural details, we incorporate edge representation from $I_{c}$ into $\\hat{I}_i$ through an element-wise multiplication, highlighting edge out of the feature map. This procedure can be formalized as:", + "bbox": [ + 88, + 414, + 483, + 565 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {I} _ {i} = \\hat {I} _ {i} + (\\hat {I} _ {i} \\odot \\epsilon I _ {c} ^ {\\downarrow h _ {i}, w _ {i}}), \\text {f o r} i = 1, \\dots , N, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 571, + 483, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tilde{I}_i$ denotes the global-edge-enhanced feature map, $\\odot$ represents the Hadamard product, $\\epsilon$ is an importance factor that modulates the influence of the edge map detected by the Canny filter, and $I_c^{\\downarrow h_i,w_i}$ indicates that $I_{c}$ has been interpolated to match the spatial dimensions of $\\hat{I}_i$ .", + "bbox": [ + 89, + 598, + 483, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Memory Feature Extraction Given the $t$ -th frame of the video, we choose the previous $n$ frames and their corresponding predicted masks and feed them into the memory encoder to derive the global memory feature $N$ . Memory encoder is a common-used module in SVOS, where we adopt the design from Cutie [7]. We then project global features onto the object-based masks to derive object memory features $M \\in \\mathbb{R}^{C \\times H \\times W}$ (with $H = W = 30$ here), which provides temporal information about the target object for the next structure refinement module.", + "bbox": [ + 89, + 674, + 483, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Supervision Preparation Although the global-edge-enhanced feature map $\\tilde{I}_i$ emphasizes edges, it still contains boundary details of non-target objects, which is insufficient for guiding the model toward precise segmentation. To address this limitation, we design a structure decoder to predict", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "structure map $S$ , containing only edges and structure details of targets, from the global-edge-enhanced feature $\\tilde{I}_i$ . For supervision, the ground-truth structure map is derived from ground-truth mask. Given a colored mask $G \\in \\mathbb{R}^{3 \\times H \\times W}$ corresponding to the image $I$ , we first convert it into a grayscale image $G_{gr} \\in \\mathbb{R}^{1 \\times H \\times W}$ . Subsequently, a predefined filter is applied to transform $G_{gr}$ into the final object-related inherent structure map $S$ . This generated object structure map effectively highlights occlusion relationships among target objects and is further used as ground truth in structural extraction network training.", + "bbox": [ + 511, + 90, + 906, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Object Structure Map Prediction After obtaining the global-edge-enhanced feature map $\\tilde{I}_i$ and the object memory feature $M$ from the previous framework, a lightweight structure decoder with connecting paths is applied to predict structure map $S$ . Importantly, $\\tilde{I}_i$ alone provides limited target-specific cues, and the network could only rely on global image features, which may compromise the accuracy of the target objects' structure map prediction. Therefore, the fusion of the object memory feature $M$ to incorporate object-specific information is necessary. This process can be formulated as:", + "bbox": [ + 511, + 257, + 908, + 422 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS ^ {\\prime} = \\text {S t r u c t u r e D e c o d e r} (\\tilde {I} _ {i}, M), \\text {f o r} S ^ {\\prime} \\rightarrow S, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 434, + 906, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $M$ is fused with the coarsest scale $\\tilde{I}_3$ via elementwise addition. The decoder progressively upsamples the fused features with transposed convolutions, incorporating higher-resolution features via channel concatenation. To reduce overhead, binary cross-entropy loss is applied only at the final output. Visual analysis in Fig. 3 reveals that incorporating object memory features is crucial for accurately capturing the target objects' structure. Otherwise, the network lacks focus, leading to noisy predictions.", + "bbox": [ + 511, + 465, + 908, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Refinement and Decoding", + "text_level": 1, + "bbox": [ + 511, + 611, + 750, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After obtaining the final object-level structure map, we apply the Hadamard product with the extracted image features to emphasize object boundary features, thus to enhance feature with structure details. This process can be expressed as follows:", + "bbox": [ + 511, + 633, + 905, + 708 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {I} _ {i} ^ {\\prime} = \\hat {I} _ {i} + \\left(\\hat {I} _ {i} \\odot \\beta S ^ {\\prime \\downarrow h _ {i}, w _ {i}}\\right), \\text {f o r} i = 1, \\dots , N, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 719, + 906, + 738 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{I}_i^{\\prime}$ denotes the features ultimately used to segment the object mask and $\\beta$ is an importance factor that modulates the influence of the structure map. Note that the structure map used in this process is in the form of logits rather than a binary mask, which helps to emphasize certain regions when the structure decoder has higher confidence in these areas being boundaries.", + "bbox": [ + 511, + 750, + 905, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the structure-enhanced feature (SE feature) of the current frame, the global memory features $N$ , and the object memory features $M$ , we employ a decoder to align them and", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14434", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/79d89e5b724fa447c2bad03bf2b2cb629dcd854b25770983682d2b297447434c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 112, + 89, + 285, + 172 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0e7f864a7a28a009994385af4845c0e47898747ce22836461bda526d0b4cd00c.jpg", + "image_caption": [ + "(c)", + "Figure 3. Output structure maps from different objects. (a) Input frame. (b) Input frame and predicted masks of the previous frame. (c) Output structure map without object fusion. (d) Output structure map with object fusion. \"Feat.\" and \"Mem.\" are abbreviations of \"Feature\" and \"Memory\" respectively." + ], + "image_footnote": [], + "bbox": [ + 112, + 174, + 282, + 306 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5b3f2b42190dabbd6d1df5a5182942455cfde4692d550ff28ac1a94b76a0c976.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 95, + 454, + 172 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/df4170e06d44dfe3c57c0461fd4532df36fbfec7a91926ee3fd582b9ad61091d.jpg", + "image_caption": [ + "(d) : Eureka!" + ], + "image_footnote": [], + "bbox": [ + 290, + 174, + 460, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "output the final object-based mask. The memory-based decoder design is inspired by Cutie [7], a pioneering approach to enhance VOS performance by object memory modeling. Concretely, the decoder integrates top-down object features with bottom-up global features so that a dynamic interaction between high-level and low-level representations is established. The SE feature, in this way, serves as an additional pathway for connecting high-level and low-level representations, as the structure encodes object information and latently emphasizes the original feature map accordingly.", + "bbox": [ + 89, + 417, + 483, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Enhancement via Uncertainty Quantification", + "text_level": 1, + "bbox": [ + 89, + 575, + 472, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to the LogSoftmax used in the cross-entropy loss function, the predicted class probabilities remain point-based estimations that collapse multiple types of uncertainty into overconfident predictions [80, 84]. To address this, we further borrow the ideas of Evidential Deep Learning (EDL) to add an uncertainty-related loss term into our training pipeline, as being supplementary to existing loss combination, cross-entropy loss $\\mathcal{L}_{\\mathrm{CE}}$ and dice loss $\\mathcal{L}_{\\mathrm{Dice}}$ , to help improve performance by quantifying and minimizing the uncertainty of predictions, which is especially crucial in challenging scenarios involving occlusions and features with high similarity.", + "bbox": [ + 89, + 598, + 483, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The EDL based on the theory of subjective logic is used to calculate uncertainty with the Dirichlet distribution which acts as an updating measure of the model's intrinsic prediction bias and provides the reference for the subsequent $\\mathrm{KL}_{\\mathrm{div}}$ regularization. Before training, we initialize a bare Dirichlet $D(\\mathbf{p}|\\boldsymbol{\\alpha})$ over the class probability vector $\\mathbf{p} = [p_1,\\dots ,p_k]$ and set all $\\alpha = 1$ , which by Gibbs' inequality means the distribution of maximum uncertainty before seeing any data.", + "bbox": [ + 89, + 780, + 485, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training, we update Dirichlet parameters with the model's output prob-logits to represent updated intrinsic bias. For each pixel at frame $t$ , we define $q_{t}^{k}$ as the Boolean ground-truth label of a certain pixel of $k$ -th class at time $t$ , and raw output logit $\\dot{q}_{t}$ is $k$ -class prob-logit vector of the pixel output by the model. The evidential loss $\\mathcal{L}_{\\mathrm{EDL}}$ is calculated as:", + "bbox": [ + 511, + 90, + 906, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {E D L}} = \\sum_ {k = 1} ^ {K} q _ {t} ^ {k} \\cdot \\left(\\psi (\\sum \\alpha) - \\psi (\\alpha)\\right) + \\mathrm {K L} _ {\\mathrm {d i v}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 204, + 906, + 246 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\text {f o r} \\alpha = 1 + \\eta (\\dot {q} _ {t}),\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 250, + 772, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ means the Dirichlet parameter, $\\eta$ is the confidence function and $\\psi$ denotes the digamma function, and $\\mathrm{KL}_{\\mathrm{div}}$ is derived corresponding to [80]. Subsequently, the total loss for mask predictions $\\mathcal{L}_{\\mathrm{mask}}$ during training can be defined as:", + "bbox": [ + 511, + 275, + 906, + 335 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a s k}} = \\mathcal {L} _ {\\mathrm {C E}} + \\mathcal {L} _ {\\mathrm {D i c e}} + \\lambda \\mathcal {L} _ {\\mathrm {E D L}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 345, + 906, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is the hyperparameter to balance the contribution of the uncertainty loss term. With this combination of loss functions, our model benefits from the discriminative power of $\\mathcal{L}_{\\mathrm{CE}}$ and $\\mathcal{L}_{\\mathrm{Dice}}$ while leveraging $\\mathcal{L}_{\\mathrm{EDL}}$ to address uncertain predictions, thus yielding improved segmentation performance.", + "bbox": [ + 511, + 372, + 908, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 476, + 645, + 493 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 501, + 730, + 516 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Network Configuration We employ two ResNet-18 models pretrained on ImageNet to extract image features and encode the mask, respectively. The dimension of the object feature is set to 256. For the structure decoder, we progressively upsample with each step employing a transposed convolution kernel of size 2 and a stride of 2 to achieve $2 \\times$ upsampling. LeakyReLU is used as the activation function and residual connections are applied within this module. For the prediction layer, a $1 \\times 1$ convolution is applied to output the binary segmentation. Due to space limitations, further details are in the supplementary materials.", + "bbox": [ + 511, + 523, + 908, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Pipeline Similar to previous work, we divide the training into two stages. First, OASIS is pre-trained on static image datasets [88-92] with a 3-frame synthesis following [7]. For these images, we apply mirroring, random affine transformations cut-and-paste as data augmentation and combine different images to create pseudo-video sequences. Default cropping size and total training iteration are set to $384 \\times 384$ and $80\\mathrm{K}$ , respectively. The batch size is set to 16, and point supervision is utilized with a total of 8192 sampled points. The learning rate is fixed at $1.0 \\times 10^{-4}$ , and the max norm of gradient clipping is set to 3. Notably, due to noise introduced by pseudo-video sequences, the structure decoder and EDL loss component are not included at this stage for the purpose of robustness.", + "bbox": [ + 511, + 689, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14435", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6e2be89cc45f4cbf8bc9ce19d32ef0d8ddd7637384a61cfdaac0f432ff5dc228.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Overall performance comparison on VOS benchmarks. Numbers are taken from [7, 86, 87]. For the current state-of-the-art method, Cutie [7], we conduct training from scratch using its publicly available code and re-evaluate its performance on the benchmarks. Boldface and bold blue text indicate the best and second-best performance in each column, respectively." + ], + "table_body": "
MethodPublicationMOSEDAV-17 ValDAV-17 TestYouTubeVOS-2019 Val
J&FJFJ&FJFJ&FJFGJsFsJuFu
STCN [37]NeurIPS'2152.548.556.685.482.288.676.172.779.682.781.185.478.285.9
AOT-R50 [4]NeurIPS'2158.454.362.684.982.387.579.675.983.385.383.988.879.988.5
RDE [85]CVPR'2246.842.451.384.280.887.577.473.681.281.981.185.576.284.8
XMem [3]ECCV'2256.352.160.686.282.989.581.077.484.585.584.388.680.388.6
LLB [86]AAAI'23---84.681.587.7---83.681.786.579.287.0
ISVOS [87]CVPR'23---87.183.790.582.879.386.286.185.289.780.788.9
DeAOT-R50 [5]NeurIPS'2259.054.663.485.282.288.280.776.984.585.684.289.280.288.8
DEVA [6]ICCV'2360.055.864.386.883.690.082.378.785.985.585.089.479.788.0
Cutie [7]CVPR'2461.257.265.386.883.889.782.879.286.586.285.389.680.989.0
OASIS-62.157.866.388.385.091.683.279.387.086.685.790.181.289.4
", + "bbox": [ + 94, + 88, + 906, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the main training stage, the batch size is set to 16, and is conducted for a total of $125\\mathrm{K}$ iterations. The base learning rate is initialized at $1.0 \\times 10^{-4}$ , with the backbone learning rate scaled by a factor of 0.1. A step scheduler is employed to control the learning rate, reducing it by a factor of 0.1 at $100\\mathrm{K}$ and $115\\mathrm{K}$ iterations. The sequence length is set to 8, with each frame cropped to a size of $480 \\times 480$ pixels. The EDL loss coefficient $\\lambda$ is set to 0.01 with an additional annealing factor to gradually modify the impact of the loss. For the Canny filter, the lower and upper thresholds are set to 50 and 200, respectively, and an $L_{2}$ regularization term is applied to enhance accuracy. Parameters $\\epsilon$ and $\\beta$ are set to 0.5 and 1.0 for rough edge and structure map feature fusion, respectively. The number of sampled points for the loss calculation is increased to 12544. We implement the model using PyTorch and train it on four A5000 GPUs.", + "bbox": [ + 88, + 325, + 485, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Benchmark Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 89, + 575, + 472, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our experiments, several commonly used SVOS datasets are used to verify OASIS, including DAVIS-2017 [2], YouTubeVOS 2019 [93], and MOSE [94]. Specifically, DAVIS-2017 consists of 60 densely annotated videos in the training set, with an additional 30 videos in both the validation and test-dev sets for evaluation. YouTubeVOS-2019 contains 3978 videos sourced from video classification datasets, divided into 3471 for training and 507 for validation. MOSE, known for its crowded environments, increased occlusions, and higher rates of object disappearance, comprises 1507 videos for training, with 311 videos in the validation set. Following previous studies, we conduct experiments using various training dataset settings and reported the results.", + "bbox": [ + 91, + 598, + 483, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our experiments, three commonly used criteria, Jaccard index $\\mathcal{I}$ , contour accuracy $\\mathcal{F}$ and their average $\\mathcal{I} \\& \\mathcal{F}$ are employed to quantify the performance of our proposed method. Among them, the Jaccard index assesses the area-based accuracy of predictions, whereas the contour accuracy measures the localization accuracy of object boundaries. Specifically, the YouTubeVOS evaluation provides detailed", + "bbox": [ + 89, + 795, + 485, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "results for $\\mathcal{I}$ and $\\mathcal{F}$ on both seen and unseen objects, and $\\mathcal{G}$ is the average of all objects, enabling a more comprehensive assessment of the generalization ability of a model. $\\mathcal{I},\\mathcal{F},$ $\\mathcal{J}\\& \\mathcal{F}$ and $\\mathcal{G}$ range from 0 to 1 and the reported numbers in our paper are scaled for better readability and formatting, following previous SVOS works [3, 6, 7]. A superior performance results in values closer to 100. Frames per second (FPS) is also employed in ablation study to measure the efficiency of a model.", + "bbox": [ + 511, + 325, + 908, + 464 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Overall Performance Comparison", + "text_level": 1, + "bbox": [ + 511, + 472, + 812, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our experiments compared OASIS with nine representative or state-of-the-art memory-based SVOS methods. Among them, STCN [37] and LLB [86] focus on achieving more accurate matching between query frames and memory banks. RDE [85], XMem [3], and DEVA [6] aims to reduce memory usage and improve efficiency, while AOT [4], DEAOT [5], ISVOS [87], and Cutie [7] emphasize enhanced object modeling. The analysis of experiments is as follows. Note that, the inference results of MOSE, YouTubeVOS, and DAVIS test-dev benchmarks are then packaged and uploaded to the evaluation server to obtain final results.", + "bbox": [ + 511, + 494, + 908, + 660 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.1. Quantitative Experiments", + "text_level": 1, + "bbox": [ + 511, + 670, + 725, + 684 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following the setup in previous works [6, 7], we trained OASIS using a combination of the DAVIS-2017 and YouTube 2019 training sets and evaluated its performance across four validation sets. It can be seen from Tab. 1 that among all of these SVOS models, our OASIS consistently achieves the best segmentation performance across all benchmarks. Since these four standard benchmarks cover diverse video content and annotation conditions, i.e., different levels of granularity, achieving top performance on all of them is particularly challenging. These results, therefore, underscore the effectiveness and stability of OASIS. Specifically, for datasets with finer annotation granularity, such as the DAVIS-17 validation set, our model outperforms others by substantial margins of 1.2, 1.2, and 1.1 on the $\\mathcal{J}\\&\\mathcal{F}, \\mathcal{J}$ ,", + "bbox": [ + 511, + 688, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14436", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/856a67bfe2e8ff0da6ef5fd931e7a35ffcd2bdf0d5bf3cea149a362db4ffedd1.jpg", + "image_caption": [ + "Figure 4. Comparison of segmentation results on two video clips, with details zoomed in on the upper right corner. Notably, our model demonstrates superior resolving capability in occlusion regions and when the target exhibits high similarity to the background." + ], + "image_footnote": [], + "bbox": [ + 112, + 89, + 859, + 251 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9eceedb1594fd6eae4d6911a02ec8861b1d1f8372020a0ad3d0e960f12d84b02.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2. Performance comparison under Trained-with-MOSE setting on VOS benchmarks. Numbers are taken from [7]. Boldface and bold blue text indicate the best and second-best performance in each column, respectively." + ], + "table_body": "
MethodMOSEDAV-17 Test
J&FJFJ&FJF
XMem [3]59.655.463.779.676.183.0
DeAOT-R50 [5]64.159.568.782.879.186.5
DEVA [6]66.061.870.382.678.986.4
Cutie [7]67.062.871.383.780.087.4
OASIS67.563.271.985.481.789.3
", + "bbox": [ + 94, + 292, + 485, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and $\\mathcal{F}$ scores, respectively. Such improvement highlights the effectiveness of structure refinement in capturing object details. Moreover, OASIS remains stable performance gains on the MOSE benchmark, which contains more complex and densely packed objects that differ considerably from the training data. The contour accuracy $(\\mathcal{F})$ improves significantly by 1.0, demonstrating that our approach enhances the structural integrity of segmentation, even in unseen and complex scenes.", + "bbox": [ + 88, + 479, + 482, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 2 presents additional results where the training set includes the MOSE benchmark, allowing for a more comprehensive performance evaluation. In this setting, our approach still demonstrates superior results, significantly outperforming other methods on the DAVIS-2017 test set by margins of 1.7, 1.7, and 1.9 on the $\\mathcal{J}\\& \\mathcal{F},\\mathcal{J},$ and $\\mathcal{F}$ scores, respectively. These results indicate that incorporating more complex scenes in the training set allows our model to develop a stronger structural understanding, leading to improved segmentation accuracy.", + "bbox": [ + 88, + 616, + 482, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.2. Qualitative Experiments", + "text_level": 1, + "bbox": [ + 89, + 775, + 308, + 791 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We illustrate the visual results of our method and the SOTA one on two typical video clips from the YouTubeVOS 2019 validation set in Fig. 4. It can be observed that the existing SOTA method often struggles to distinguish between objects, particularly in cases of occlusion or high similarity between objects or between foreground objects and background. For instance, it fails to separate the fish hiding", + "bbox": [ + 88, + 794, + 485, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c8c15a8b53f2474d3a1c43058ed2e80b4e258f26853f0f7f6a41ade099d8c680.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Ablation experiments about key components on DAVIS-2017 validation set and test-dev set." + ], + "table_body": "
MethodDAV-17 ValDAV-17 Test-dev
J&FJFJ&FJF
Baseline84.881.688.079.976.083.7
+CaE84.981.888.180.076.084.0
+CaE+SD85.182.088.180.376.384.5
+CaE+SD*85.482.288.680.876.884.9
OASIS85.682.488.781.076.985.1
", + "bbox": [ + 517, + 292, + 903, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "behind the other fish (the 1 - 3 rows). In contrast, due to the proposed evidential learning, our OASIS is able to model the structure of the overlapped fish better and accomplishes precise segmentation in the first clip. Additionally, we observe that previous methods exhibit accumulating errors over time when segmenting the antennae of a snail (rows 4-6), due to their reliance on memory mechanisms. In contrast, our proposed structure refinement module corrects these inaccuracies, significantly reducing this issue.", + "bbox": [ + 511, + 454, + 908, + 592 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 601, + 668, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.1. Ablation on Key Components", + "text_level": 1, + "bbox": [ + 511, + 625, + 767, + 640 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "OASIS involves three novel designs: the integration of rough edges detected by the Canny filter, structure decoder for refinement, and evidential learning loss for uncertainty quantification. We conduct ablation experiments to assess the individual contribution of each component, with an additional experiment on the object features fusion mechanism in the structure decoder. To this end, we build five experimental settings in total: $+\\mathrm{CaE}$ refers to the baseline model with pixel feature integrated with the rough Canny edges as per Eq. 3, whereas it does not include the structure decoder and evidential learning. As shown in Fig. 3, $+\\mathrm{CaE} + \\mathrm{SD}$ means the structure decoder is further applied on top of $+\\mathrm{CaE}$ for the refined structure map, while $+\\mathrm{CaE} + \\mathrm{SD}^*$ applied the object memory feature fusion mechanism. The final OASIS further adds uncertainty learning upon the $+\\mathrm{CaE} + \\mathrm{SD}^*$ . Due to limited resources, all models are trained on the DAVIS2017 training set with limited iterations and evaluated on the", + "bbox": [ + 511, + 643, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14437", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/047783976fd1495b20cf80af29f20ef684d015ac58eaea956a9c9f532f590769.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison from different settings in ablating key components. We show the results of four typical frames. Zoom in for a better view." + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 460, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DAVIS-2017 validation and test-dev sets.", + "bbox": [ + 89, + 315, + 364, + 329 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tab. 3 shows the results of our ablation study. As observed on these benchmarks, Canny edge feature integration and the structure decoder are effective in characterizing the object and thus contribute to the overall segmentation performance. Object fusion further improves the capabilities of structure decoder, i.e., $+\\mathrm{CaE} + \\mathrm{SD^{*}}$ outperforms the one without object fusion largely especially on the test-dev set, with all indicators increased by about 0.5. Furthermore, the further introduction of uncertainty estimation and minimization allows our complete model to attain the best performance with negligible additional expense.", + "bbox": [ + 89, + 332, + 483, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.2. Visual Comparison on Key Components", + "text_level": 1, + "bbox": [ + 89, + 512, + 415, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 5 presents the visual results of different settings on representative frames. It can be observed that although $+\\mathrm{CaE}$ achieves decent performance, structure refinement is still necessary to enhance output quality, especially for occlusions. For instance, at the position indicated by the white arrow (i.e., the body of the dog and the corner of the box), the implementation of the structure decoder significantly improves the structural details. This observation further validates the effectiveness of our design.", + "bbox": [ + 89, + 532, + 483, + 670 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.3. Ablation on Factor $\\epsilon$ and $\\beta$", + "text_level": 1, + "bbox": [ + 89, + 683, + 323, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Eq. 3 and Eq. 5 we use factors $\\epsilon$ and $\\beta$ to control the impact brought by the rough edges and the refined structure map, respectively. We separately investigate the effect of different values for these two factors on model performance, with results presented in Tab. 4. Notably, as the two factors increase from 0, accuracy consistently improves, substantiating the benefit of emphasizing boundary features. However, we observe that this performance gain reaches an upper limit, beyond which excessively large factors introduce a substantial distributional shift in the features, ultimately degrading performance. The best results are achieved when $\\beta$ is greater than $\\epsilon$ , further underscoring the importance of more detailed structural information.", + "bbox": [ + 89, + 704, + 483, + 898 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0afb2c7196bfc2882bafe36c163cd3049a7d5fc1cc443b07b165ddb733a8028e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
βDAVIS-17 Val| |DAVIS-17 Val
J&FJF| |J&FJF
0.585.382.188.50.085.382.288.5
1.085.682.488.70.585.682.488.7
2.085.782.588.91.085.181.888.3
5.085.082.087.92.085.082.087.9
", + "bbox": [ + 517, + 88, + 903, + 198 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/dac24788a9e2b46603f332372d30e5d8e914c9eac030da678de9deaeee169bfc.jpg", + "table_caption": [ + "Table 4. Ablation experiments about factors $\\epsilon$ and $\\beta$ on the DAVIS-2017 validation set." + ], + "table_footnote": [], + "table_body": "
Method#ParamsDAVIS-17 Val
J&FJFFPS
Baseline29M84.881.688.052
+2M Proj Net31M85.182.088.137
OASIS31M85.682.488.748
", + "bbox": [ + 514, + 244, + 911, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Experiments on the effect of including additional parameters. Results on the DAVIS-2017 validation set are shown.", + "bbox": [ + 511, + 351, + 906, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Performance Gain and Model Parameters", + "text_level": 1, + "bbox": [ + 511, + 393, + 870, + 410 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In deep learning-based SVOS methods, performance gains may result from an increase in model parameters. In our approach, the primary source of additional parameters is the structure decoder, which introduces approximately 2M learnable parameters. Compared with the whole framework, these additional parameters are almost negligible. To further validate that the effectiveness of OASIS does not primarily stem from the increased number of parameters, we conduct additional analytical experiments. To provide a fair comparison, we add a convolutional projection network of approximately 2M parameters to the baseline as a comparative method, while maintaining other settings and hyperparameters. Results are shown in Tab. 5. Notably, the method with a 2M-parameter network does not achieve better results than OASIS. Combined with the ablative results in Tab. 3, which compare model performance with and without the structure decoder, we conclude that the proposed method is highly effective and, more importantly, that its performance gain is not due to the increased parameter count.", + "bbox": [ + 511, + 417, + 908, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 722, + 633, + 738 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a novel model named OASIS to address the challenge of occlusion in SVOS. OASIS first generates a structure map that captures the structural and layering information of predicted targets and leverages it to refine features, thus highlighting boundary regions. Additionally, we integrate evidential learning for uncertainty estimation to enhance the capability of OASIS to focus on complex and uncertain segmentation cases. Extensive experiments show the superior performance of OASIS while maintaining high computational efficiency.", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14438", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by Ministry of Education Tier 2 grant, NUS, Singapore (T2EP20224-0028); and Ministry of Education Tier 1 grant, NUS, Singapore (24-1250-P0001).", + "bbox": [ + 89, + 114, + 485, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 172, + 187, + 188 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] S. Caelles, K.-K. Maninis, J. Pont-Tuset, L. Leal-Taixe, D. Cremers, and L. Van Gool, \"One-shot video object segmentation,\" in Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 221-230. 1, 2", + "[2] F. Perazzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, and A. Sorkine-Hornung, \"A benchmark dataset and evaluation methodology for video object segmentation,\" in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 724-732. 1, 2, 6", + "[3] H. K. Cheng and A. G. Schwing, \"Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model,\" in European Conference on Computer Vision. Springer, 2022, pp. 640-658. 1, 2, 6, 7", + "[4] Z. Yang, Y. Wei, and Y. Yang, \"Associating objects with transformers for video object segmentation,\" Advances in Neural Information Processing Systems, vol. 34, pp. 2491-2502, 2021. 6", + "[5] Z. Yang and Y. Yang, \"Decoupling features in hierarchical propagation for video object segmentation,\" Advances in Neural Information Processing Systems, vol. 35, pp. 36324-36336, 2022. 6, 7", + "[6] H. K. Cheng, S. W. Oh, B. Price, A. Schwing, and J.-Y. Lee, \"Tracking anything with decoupled video segmentation,\" in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 1316-1326. 1, 6, 7", + "[7] H. K. Cheng, S. W. Oh, B. Price, J.-Y. Lee, and A. Schwing, \"Putting the object back into video object segmentation,\" in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 3151-3161. 1, 2, 4, 5, 6, 7", + "[8] Z. Huang, S. Yang, M. Zhou, Z. Gong, A. Abusorrah, C. Lin, and Z. Huang, \"Making accurate object detection at the edge: Review and new approach,\" Artificial Intelligence Review, vol. 55, no. 3, pp. 2245-2274, 2022. 2", + "[9] D. Yang, B. Peng, Z. Al-Huda, A. Malik, and D. Zhai, “An overview of edge and object contour detection,” Neurocomputing, vol. 488, pp. 470–493, 2022.", + "[10] C. L. Zitnick and P. Dollar, \"Edge boxes: Locating object proposals from edges,\" in Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. Springer, 2014, pp. 391-405. 2", + "[11] C. Kim and J.-N. Hwang, \"Fast and automatic video object segmentation and tracking for content-based applications,\" IEEE transactions on circuits and systems for video technology, vol. 12, no. 2, pp. 122-129, 2002. 2, 3", + "[12] S. Wan, S. Ding, and C. Chen, \"Edge computing enabled video segmentation for real-time traffic monitoring in internet of vehicles,\" Pattern Recognition, vol. 121, p. 108146, 2022. 2, 3" + ], + "bbox": [ + 93, + 198, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] A. Khoreva, F. Galasso, M. Hein, and B. Schiele, \"Classifier based graph construction for video segmentation,\" in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015, pp. 951-960. 2, 3", + "[14] J. Canny, “A computational approach to edge detection,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. PAMI-8, no. 6, pp. 679–698, 1986. 2", + "[15] J. Pont-Tuset, F. Perazzi, S. Caelles, P. Arbeláez, A. Sorkine-Hornung, and L. Van Gool, “The 2017 davis challenge on video object segmentation,” arXiv preprint arXiv:1704.00675, 2017. 2", + "[16] P. Voigtlaender and B. Leibe, \"Online adaptation of convolutional neural networks for video object segmentation,\" in Proceedings of the British Machine Vision Conference (BMVC). BMVA Press, September 2017, pp. 116.1-116.13.", + "[17] H. Xiao, J. Feng, G. Lin, Y. Liu, and M. Zhang, “Monet: Deep motion exploitation for video object segmentation,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 1140–1148.", + "[18] J. Luiten, P. Voigtlaender, and B. Leibe, “Premvos: Proposal generation, refinement and merging for video object segmentation,” in Asian conference on computer vision. Springer, 2018, pp. 565–580.", + "[19] Y.-T. Hu, J.-B. Huang, and A. Schwing, “Maskrnn: Instance level video object segmentation,” Advances in neural information processing systems, vol. 30, 2017.", + "[20] F. Perazzi, A. Khoreva, R. Benenson, B. Schiele, and A. Sorkine-Hornung, \"Learning video object segmentation from static images,\" in Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 2663-2672. 2", + "[21] K.-K. Maninis, S. Caelles, Y. Chen, J. Pont-Tuset, L. Leal-Taixe, D. Cremers, and L. Van Gool, \"Video object segmentation without temporal information,\" IEEE transactions on pattern analysis and machine intelligence, vol. 41, no. 6, pp. 1515-1530, 2018. 2", + "[22] J. Cheng, Y.-H. Tsai, S. Wang, and M.-H. Yang, \"Segflow: Joint learning for video object segmentation and optical flow,\" in Proceedings of the IEEE international conference on computer vision, 2017, pp. 686-695. 2", + "[23] X. Li and C. C. Loy, \"Video object segmentation with joint re-identification and attention-aware mask propagation,\" in Proceedings of the European conference on computer vision (ECCV), 2018, pp. 90-105.", + "[24] S. W. Oh, J.-Y. Lee, K. Sunkavalli, and S. J. Kim, \"Fast video object segmentation by reference-guided mask propagation,\" in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 7376-7385.", + "[25] Y.-H. Tsai, M.-H. Yang, and M. J. Black, \"Video segmentation via object flow,\" in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 3899-3908.", + "[26] H. Xie, H. Yao, S. Zhou, S. Zhang, and W. Sun, \"Efficient regional memory network for video object segmentation,\" in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 1286-1295. 2", + "[27] L. Yang, Y. Wang, X. Xiong, J. Yang, and A. Katsaggelos, \"Efficient video object segmentation via network modulation,\"" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14439", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 6499-6507.", + "[28] B. Duke, A. Ahmed, C. Wolf, P. Aarabi, and G. W. Taylor, \"Sstvos: Sparse spatiotemporal transformers for video object segmentation,\" in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 5912-5921.", + "[29] K. Xu and A. Yao, \"Accelerating video object segmentation with compressed video,\" in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 1342-1351.", + "[30] Y.-S. Xu, T.-J. Fu, H.-K. Yang, and C.-Y. Lee, \"Dynamic video segmentation network,\" in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 6556-6565. 2", + "[31] Y. Chen, J. Pont-Tuset, A. Montes, and L. Van Gool, \"Blazingly fast video object segmentation with pixel-wise metric learning,\" in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 1189-1198. 2", + "[32] Y.-T. Hu, J.-B. Huang, and A. G. Schwing, \"Videomatch: Matching based video object segmentation,\" in Proceedings of the European conference on computer vision (ECCV), 2018, pp. 54-70.", + "[33] Z. Yang, Y. Wei, and Y. Yang, \"Collaborative video object segmentation by multi-scale foreground-background integration,\" IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 9, pp. 4701-4712, 2021.", + "[34] P. Voigtlaender, Y. Chai, F. Schroff, H. Adam, B. Leibe, and L.-C. Chen, “Feelvos: Fast end-to-end embedding learning for video object segmentation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 9481–9490.", + "[35] S. W. Oh, J.-Y. Lee, N. Xu, and S. J. Kim, \"Video object segmentation using space-time memory networks,\" in Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 9226-9235. 2", + "[36] H. K. Cheng, Y.-W. Tai, and C.-K. Tang, \"Modular interactive video object segmentation: Interaction-to-mask, propagation and difference-aware fusion,\" in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 5559-5568.", + "[37] ——, “Rethinking space-time networks with improved memory coverage for efficient video object segmentation,” Advances in Neural Information Processing Systems, vol. 34, pp. 11781-11794, 2021. 2, 6", + "[38] Y. Li, Z. Shen, and Y. Shan, \"Fast video object segmentation using the global context module,\" in Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16. Springer, 2020, pp. 735-750.", + "[39] Y. Liang, X. Li, N. Jafari, and J. Chen, \"Video object segmentation with adaptive feature bank and uncertain-region refinement,\" Advances in Neural Information Processing Systems, vol. 33, pp. 3430-3441, 2020. 2", + "[40] H. Wang, X. Jiang, H. Ren, Y. Hu, and S. Bai, \"Swiftnet: Real-time video object segmentation,\" in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 1296-1305. 2" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Y. Mao, N. Wang, W. Zhou, and H. Li, \"Joint inductive and transductive learning for video object segmentation,\" in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 9670-9679. 2", + "[42] H. Seong, J. Hyun, and E. Kim, \"Kernelized memory network for video object segmentation,\" in Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. Springer, 2020, pp. 629-645.", + "[43] L. Hu, P. Zhang, B. Zhang, P. Pan, Y. Xu, and R. Jin, \"Learning position and target consistency for memory-based video object segmentation,\" in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 4144-4154.", + "[44] X. Lu, W. Wang, M. Danelljan, T. Zhou, J. Shen, and L. Van Gool, \"Video object segmentation with episodic graph memory networks,\" in Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. Springer, 2020, pp. 661-679.", + "[45] H. Seong, S. W. Oh, J.-Y. Lee, S. Lee, S. Lee, and E. Kim, \"Hierarchical memory matching network for video object segmentation,\" in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 12889-12898.", + "[46] H. Liu, M. Gao, X. Luo, Z. Wang, G. Qin, J. Wu, and Y. Jin, \"Resurgsam2: Referring segment anything in surgical video via credible long-term tracking,\" in International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2025.", + "[47] J. Pei, J. Zhang, G. Qin, K. Wang, Y. Jin, and P.-A. Heng, \"Instrument-tissue-guided surgical action triplet detection via textual-temporal trail exploration,\" IEEE Transactions on Medical Imaging, 2025. 2", + "[48] N. Ravi, V. Gabeur, Y.-T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. Rädle, C. Rolland, L. Gustafson, E. Mintun, J. Pan, K. V. Alwala, N. Carion, C.-Y. Wu, R. Girshick, P. Dollar, and C. Feichtenhofer, “SAM 2: Segment anything in images and videos,” in The Thirteenth International Conference on Learning Representations, 2025. 2", + "[49] Z. Wang, Y. Zhang, Y. Wang, L. Cai, and Y. Zhang, \"Dynamic pseudo label optimization in point-supervised nuclei segmentation,\" in International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2024, pp. 220-230. 2", + "[50] Y. Zhang, Z. Wang, Y. Wang, H. Bian, L. Cai, H. Li, L. Zhang, and Y. Zhang, “Boundary-aware contrastive learning for semi-supervised nuclei instance segmentation,” in Medical Imaging with Deep Learning, 2024.", + "[51] Y. Zhang, Y. Zhou, Y. Wang, J. Xiao, Z. Wang, Y. Zhang, and J. Chen, \"The four color theorem for cell instance segmentation,\" in Forty-second International Conference on Machine Learning, 2025. 2", + "[52] J.-X. Zhao, J.-J. Liu, D.-P. Fan, Y. Cao, J. Yang, and M.-M. Cheng, \"Egnet: Edge guidance network for salient object detection,\" in Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 8779-8788. 2", + "[53] Z. Luo, A. Mishra, A. Achkar, J. Eichel, S. Li, and P.-M. Jodoin, \"Non-local deep features for salient object detection,\"" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14440", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "in Proceedings of the IEEE Conference on computer vision and pattern recognition, 2017, pp. 6609-6617.", + "[54] Y. Zhuge, G. Yang, P. Zhang, and H. Lu, \"Boundary-guided feature aggregation network for salient object detection,\" IEEE Signal Processing Letters, vol. 25, no. 12, pp. 1800-1804, 2018.", + "[55] Z. Tu, Y. Ma, C. Li, J. Tang, and B. Luo, \"Edge-guided nonlocal fully convolutional network for salient object detection,\" IEEE transactions on circuits and systems for video technology, vol. 31, no. 2, pp. 582-593, 2020.", + "[56] X. Wang, H. Ma, X. Chen, and S. You, \"Edge preserving and multi-scale contextual neural network for salient object detection,\" IEEE Transactions on Image Processing, vol. 27, no. 1, pp. 121-134, 2017.", + "[57] Z. Wu, L. Su, and Q. Huang, \"Stacked cross refinement network for edge-aware salient object detection,\" in Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 7264-7273. 2", + "[58] G. Li, Y. Xie, L. Lin, and Y. Yu, \"Instance-level salient object segmentation,\" in Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 2386-2395. 2", + "[59] Y. Hao, Y. Liu, Z. Wu, L. Han, Y. Chen, G. Chen, L. Chu, S. Tang, Z. Yu, Z. Chen et al., \"Edgeflow: Achieving practical interactive segmentation with edge-guided flow,\" in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 1551-1560.", + "[60] H. Liu, R. A. R. Soto, F. Xiao, and Y. J. Lee, \"Yolactedge: Real-time instance segmentation on the edge,\" in 2021 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2021, pp. 9579-9585.", + "[61] Y. Zhou, X. Xu, F. Shen, X. Zhu, and H. T. Shen, \"Flow-edge guided unsupervised video object segmentation,\" IEEE Transactions on circuits and systems for video technology, vol. 32, no. 12, pp. 8116-8127, 2021. 2, 3", + "[62] L. Liu, H. Li, and M. Gruteser, \"Edge assisted real-time object detection for mobile augmented reality,\" in The 25th annual international conference on mobile computing and networking, 2019, pp. 1-16.", + "[63] C. Zhan, X. Duan, S. Xu, Z. Song, and M. Luo, \"An improved moving object detection algorithm based on frame difference and edge detection,\" in Fourth international conference on image and graphics (ICIG 2007). IEEE, 2007, pp. 519-523.", + "[64] M. Hanyao, Y. Jin, Z. Qian, S. Zhang, and S. Lu, \"Edge-assisted online on-device object detection for real-time video analytics,\" in IEEE INFOCOM 2021-IEEE Conference on Computer Communications. IEEE, 2021, pp. 1-10. 2", + "[65] Z. Xu, X. Baojie, and W. Guoxin, \"Canny edge detection based on open cv,\" in 2017 13th IEEE international conference on electronic measurement & instruments (ICEMI). IEEE, 2017, pp. 53-56. 2", + "[66] L. Xuan and Z. Hong, \"An improved canny edge detection algorithm,\" in 2017 8th IEEE international conference on software engineering and service science (ICSESS). IEEE, 2017, pp. 275-278.", + "[67] W. Rong, Z. Li, W. Zhang, and L. Sun, \"An improved canny edge detection algorithm,\" in 2014 IEEE international con" + ], + "bbox": [ + 91, + 92, + 485, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ference on mechatronics and automation. IEEE, 2014, pp. 577-582. 2", + "[68] L. Ding and A. Goshtasby, “On the canny edge detector,” Pattern recognition, vol. 34, no. 3, pp. 721-725, 2001. 3", + "[69] W. McIlhagga, “The canny edge detector revisited,” International Journal of Computer Vision, vol. 91, pp. 251–261, 2011.", + "[70] R. Song, Z. Zhang, and H. Liu, \"Edge connection based canny edge detection algorithm,\" Pattern Recognition and Image Analysis, vol. 27, pp. 740-747, 2017.", + "[71] P. Bao, L. Zhang, and X. Wu, \"Canny edge detection enhancement by scale multiplication,\" IEEE transactions on pattern analysis and machine intelligence, vol. 27, no. 9, pp. 1485-1490, 2005. 3", + "[72] D. Xu, J. Liu, X. Li, Z. Liu, and X. Tang, \"Insignificant shadow detection for video segmentation,\" IEEE Transactions on Circuits and Systems for Video Technology, vol. 15, no. 8, pp. 1058-1064, 2005. 3", + "[73] S. Seoni, V. Jahmunah, M. Salvi, P. D. Barua, F. Molinari, and U. R. Acharya, “Application of uncertainty quantification to artificial intelligence in healthcare: A review of last decade (2013–2023),\" Computers in Biology and Medicine, p. 107441, 2023. 3", + "[74] A. Blake, R. Curwen, and A. Zisserman, “A framework for spatiotemporal control in the tracking of visual contours,” International Journal of Computer Vision, vol. 11, pp. 127–145, 1993. 3", + "[75] X. He, R. S. Zemel, and M. A. Carreira-Perpinan, “Multiscale conditional random fields for image labeling,” in Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004., vol. 2. IEEE, 2004, pp. II-II. 3", + "[76] D. Dera, N. C. Bouaynaya, G. Rasool, R. Shterenberg, and H. M. Fathallah-Shaykh, “Premium-cnn: Propagating uncertainty towards robust convolutional neural networks,” IEEE Transactions on Signal Processing, vol. 69, pp. 4669–4684, 2021. 3", + "[77] Y. Gal and Z. Ghahramani, “Dropout as a bayesian approximation: Representing model uncertainty in deep learning,” in international conference on machine learning. PMLR, 2016, pp. 1050–1059.", + "[78] A. Kendall and Y. Gal, “What uncertainties do we need in bayesian deep learning for computer vision?” Advances in neural information processing systems, vol. 30, 2017. 3", + "[79] A. Ashukha, A. Lyzhov, D. Molchanov, and D. Vetrov, “Pitfalls of in-domain uncertainty estimation and ensembling in deep learning,” in The Eighth International Conference on Learning Representations, 2020. 3", + "[80] M. Sensoy, L. Kaplan, and M. Kandemir, \"Evidential deep learning to quantify classification uncertainty,\" in Advances in Neural Information Processing Systems, S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, Eds., vol. 31. Curran Associates, Inc., 2018. 3, 5", + "[81] K. Zou, Z. Chen, X. Yuan, X. Shen, M. Wang, and H. Fu, “A review of uncertainty estimation and its application in medical imaging,” Meta-Radiology, p. 100003, 2023." + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14441", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[82] W. Bao, Q. Yu, and Y. Kong, \"Evidential deep learning for open set action recognition,\" in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 13349-13358. 3", + "[83] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, \"Imagenet: A large-scale hierarchical image database,\" in 2009 IEEE Conference on Computer Vision and Pattern Recognition, 2009, pp. 248-255. 4", + "[84] H. Yang, C. Chen, Y. CHEN, , H. C. Yip, and D. QI, \"Uncertainty estimation for safety-critical scene segmentation via fine-grained reward maximization,\" in Advances in Neural Information Processing Systems, A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, Eds., vol. 36. Curran Associates, Inc., 2023, pp. 36238-36249. 5", + "[85] M. Li, L. Hu, Z. Xiong, B. Zhang, P. Pan, and D. Liu, \"Recurrent dynamic embedding for video object segmentation,\" in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 1332-1341. 6", + "[86] M. Lan, J. Zhang, L. Zhang, and D. Tao, “Learning to learn better for video object segmentation,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 1, 2023, pp. 1205–1212. 6", + "[87] J. Wang, D. Chen, Z. Wu, C. Luo, C. Tang, X. Dai, Y. Zhao, Y. Xie, L. Yuan, and Y.-G. Jiang, “Look before you match: Instance understanding matters in video object segmentation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 2268–2278. 6", + "[88] J. Shi, Q. Yan, L. Xu, and J. Jia, \"Hierarchical image saliency detection on extended cssd,\" IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 38, no. 4, pp. 717-729, 2016. 5", + "[89] L. Wang, H. Lu, Y. Wang, M. Feng, D. Wang, B. Yin, and X. Ruan, “Learning to detect salient objects with image-level supervision,” in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 3796–3805.", + "[90] X. Li, T. Wei, Y. P. Chen, Y.-W. Tai, and C.-K. Tang, \"Fss1000: A 1000-class dataset for few-shot segmentation,\" in 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 2866-2875.", + "[91] Y. Zeng, P. Zhang, Z. Lin, J. Zhang, and H. Lu, \"Towards high-resolution salient object detection,\" in 2019 IEEE/CVF International Conference on Computer Vision (ICCV), 2019, pp. 7233-7242.", + "[92] H. K. Cheng, J. Chung, Y.-W. Tai, and C.-K. Tang, \"Cascadepsp: Toward class-agnostic and very high-resolution segmentation via global and local refinement,\" in 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 8887-8896. 5", + "[93] N. Xu, L. Yang, Y. Fan, D. Yue, Y. Liang, J. Yang, and T. Huang, \"Youtube-vos: A large-scale video object segmentation benchmark,\" in ECCV, 2018. 6", + "[94] H. Ding, C. Liu, S. He, X. Jiang, P. H. Torr, and S. Bai, “MOSE: A new dataset for video object segmentation in complex scenes,” in ICCV, 2023. 6" + ], + "bbox": [ + 91, + 92, + 485, + 854 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "14442", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Structure Matters_ Revisiting Boundary Refinement in Video Object Segmentation/bf659278-d12b-4982-824e-fbf802268f8f_model.json b/2025/Structure Matters_ Revisiting Boundary Refinement in Video Object Segmentation/bf659278-d12b-4982-824e-fbf802268f8f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..036befc9a77a7705334d52250fdc705cd85e27bc --- /dev/null +++ b/2025/Structure Matters_ Revisiting Boundary Refinement in Video Object Segmentation/bf659278-d12b-4982-824e-fbf802268f8f_model.json @@ -0,0 +1,2798 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.131, + 0.908, + 0.153 + ], + "angle": 0, + "content": "Structure Matters: Revisiting Boundary Refinement in Video Object Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.203, + 0.18, + 0.794, + 0.217 + ], + "angle": 0, + "content": "Guanyi Qin\\(^{1,*}\\), Ziyue Wang\\(^{1,*}\\), Daiyun Shen\\(^{1}\\), Haofeng Liu\\(^{1}\\), Hantao Zhou\\(^{2}\\), Junde Wu\\(^{3}\\), Runze Hu\\(^{4}\\), Yueming Jin\\(^{1,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.275, + 0.217, + 0.723, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\)National University of Singapore; \\(^{2}\\)Tsinghua University" + }, + { + "type": "text", + "bbox": [ + 0.279, + 0.235, + 0.719, + 0.253 + ], + "angle": 0, + "content": "3University of Oxford; 4Beijing Institute of Technology" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.255, + 0.89, + 0.27 + ], + "angle": 0, + "content": "{guanyi.qin, e1374378, e1374467, haofeng.liu}@u.nus.edu, hrz@bit.edu.cn, ymjin@nus.edu.sg" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.305, + 0.327, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.337, + 0.486, + 0.73 + ], + "angle": 0, + "content": "Given an object mask, Semi-supervised Video Object Segmentation (SVOS) technique aims to track and segment the object across video frames, serving as a fundamental task in computer vision. Although recent memory-based methods demonstrate potential, they often struggle with scenes involving occlusion, particularly in handling object interactions and high feature similarity. To address these issues and meet the real-time processing requirements of downstream applications, in this paper, we propose a novel bOundary Amendment video object Segmentation method with Inherent Structure refinement, hereby named OASIS. Specifically, a lightweight structure refinement module is proposed to enhance segmentation accuracy. With the fusion of rough edge priors captured by the Canny filter and stored object features, the module can generate an object-level structure map and refine the representations by highlighting boundary features. Evidential learning for uncertainty estimation is introduced to further address challenges in occluded regions. The proposed method, OASIS, maintains an efficient design, yet extensive experiments on challenging benchmarks demonstrate its superior performance and competitive inference speed compared to other state-of-the-art methods, i.e., achieving the \\(\\mathcal{F}\\) values of 91.6 (vs. 89.7 on DAVIS-17 validation set) and \\(\\mathcal{G}\\) values of 86.6 (vs. 86.2 on YouTubeVOS 2019 validation set) while maintaining a competitive speed of 48 FPS on DAVIS. Checkpoints, logs, and code will be available here." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.76, + 0.222, + 0.775 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.786, + 0.485, + 0.877 + ], + "angle": 0, + "content": "Video Object Segmentation (VOS) is an important task in computer vision [1, 2] that aims to separate moving objects from the background. In the general semi-supervised VOS (SVOS) setting, the task involves propagating the object masks specified in the first frame throughout the entire video. By enabling effective identification and understanding of" + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.315, + 0.705, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.71, + 0.36, + 0.731, + 0.373 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.319, + 0.848, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.569, + 0.405, + 0.691, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.711, + 0.444, + 0.732, + 0.457 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.406, + 0.848, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.469, + 0.907, + 0.554 + ], + "angle": 0, + "content": "Figure 1. (a) Input frame. (b) Object mask generated by the SOTA method Cutie. (c) Output structure map. (d) Object mask generated by our model. In this case, the fish corresponding to the red mask closely resembles the background and is concealed behind the other two fish, causing previous methods to fail. Guided by our structure map, the model successfully segments the fish accurately." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.909, + 0.659 + ], + "angle": 0, + "content": "objects in videos, VOS plays a crucial role in various applications including autonomous driving, robotics, video editing, medical intervention, etc. Albeit performance demands, real-time processing is also essential for these applications, necessitating algorithms that accurately perceive and segment objects while delivering sufficient inference speed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Recent advancements [3-6] in SVOS emphasize the value of memory-based mechanisms. Representations of certain past frames are generated and stored in a memory bank. The current frame is then used as the query to match these stored representations and derive segmentation masks. Cutie [7], further proposes object queries as a high-level summary of the target object and interacts with the pixel-level features of the query frame for accurate segmentation, achieving the state-of-the-art performance. However, these approaches still struggle when similar objects are close to each other or are occluded, as shown in Figure 1 (a) and (b). On one hand, the mask prediction depends predominantly on maintaining consistency across frames for both pixel- and object-level features, while the inherent geometric characteristics of the objects are often overlooked. This limitation poses challenges in accurately discerning object structures, particularly" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.353, + 0.901 + ], + "angle": 0, + "content": "* Equal contribution, † Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14431" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.198 + ], + "angle": 0, + "content": "when they resemble the background (red box). On the other hand, when occlusion occurs, objects become fragmented. This challenge is amplified when objects share similar semantics, leading to high uncertainty in overlapping areas. Without an understanding of object structure, the model struggles to distinguish objects and establish clear boundaries (blue box)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.2, + 0.486, + 0.442 + ], + "angle": 0, + "content": "Edge information is a crucial component of an object's structure representation, defining its boundaries and shape by representing the inherent structural information. Precise edge prediction allows models to capture structural details more effectively, thereby enhancing the model's ability to understand occlusions. Therefore, edges have been widely used for segmenting objects for static images across various fields including object detection and semantic segmentation [8-10]. However, in VOS tasks, target movements lead to considerable shifts in object edges over time, making it challenging to predict accurate edges from scratch for each frame. This often requires large edge detection networks, which significantly reduce inference efficiency. Therefore, edges are only used to analyze motion magnitude or shadow area detection primarily, which merely requires coarse edges [11-13]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.445, + 0.485, + 0.838 + ], + "angle": 0, + "content": "To tackle the aforementioned problems, we propose an bOuntary Amendment video object Segmentation with Inherent Structure refinement, hereby named OASIS, which effectively refines object structures to handle geometric and layer relationships among objects better, while highlighting them out of backgrounds. This is achieved through a simple but effective structure refinement module. Specifically, we first extract the rough edge information by utilizing a fast and robust edge detection algorithm, Canny Edge Detection [14], which can identify global edges in individual frames and provide strong priors. With these priors, we only need a lightweight structure decoder to obtain the edges of the target objects instead of predicting from scratch without significant sacrifice in inference speed. This map further amplifies the original features by highlighting objects' boundaries, producing structure-enhanced features for final segmentation. This process enables the model to differentiate among objects and from the background effectively. Additionally, an object feature fusion mechanism is introduced to integrate memory features into the structure map to better model relationships between objects. Furthermore, to address the high uncertainty in dealing with overlapped areas, we also make the first attempt to deliver a new learning pattern that incorporates Evidential Learning (EDL) for uncertainty quantification and introduces an additional loss to minimize uncertainty and improve segmentation accuracy in occlusion regions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Extensive experiments demonstrate the superiority of our approach across four benchmarks under two training settings, particularly in scenes where target objects interact and exhibit complex relationships. Notably, OASIS achieves" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.152 + ], + "angle": 0, + "content": "promising results, consistently outperforming current state-of-the-art methods across different datasets, while delivering a real-time inference speed at 48 fps. In summary, our contributions are the following:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.153, + 0.907, + 0.198 + ], + "angle": 0, + "content": "- We propose OASIS, a novel framework designed to manage relationships between the objects and the background by accurately modeling object structures." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.199, + 0.907, + 0.242 + ], + "angle": 0, + "content": "- We introduce a structure refinement module and an object sensory mechanism that utilizes edge information to achieve inherent structure refinement." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.244, + 0.905, + 0.317 + ], + "angle": 0, + "content": "- To the best of our knowledge, we make the very first attempt to incorporate evidential learning in the SVOS task to improve the model performance on high-uncertainty regions, particularly in areas with significant similarity and occlusion." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.319, + 0.905, + 0.38 + ], + "angle": 0, + "content": "- We verify OASIS on multiple SVOS benchmarks involving a wide range of video contents, annotation conditions, and training settings. OASIS outperforms other competitors across all these datasets and settings." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.153, + 0.907, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.395, + 0.655, + 0.411 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.42, + 0.761, + 0.437 + ], + "angle": 0, + "content": "2.1. Video Object Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.443, + 0.909, + 0.686 + ], + "angle": 0, + "content": "Early approaches of VOS [1, 2, 15-21] implement online learning that finetunes the model at test-time, resulting in low efficiency. Propagation-based methods [20, 22-30] iteratively propagate masks between adjacent frames by leveraging temporal correlations; however, they are inevitably constrained by drift and occlusion. Matching-based models [31-40] compute correspondences between the current and reference frames to capture long-range temporal context. Some memory-based methods [3, 26, 35-37, 39, 41-47] further enhance this by using memory banks to store features from past frames for pixel-level matching. The lately research SAM2 [48] also benefitted from memory mechanisms. Recently, Cutie [7] proposed an object-reading mechanism to enhance object-level representations. While it effectively perceives individual objects, modeling the relationships between multiple objects remains an open challenge." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.697, + 0.667, + 0.713 + ], + "angle": 0, + "content": "2.2. Edge Detection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Edge information has been widely utilized to differentiate occluded objects [8-10, 49-51]. In salient object detection [52-57], edge features assist in identifying the most visually prominent objects within the image. Similarly, in segmentation tasks [11, 12, 58-61], edge detection helps to distinguish different regions and instances. Accurate localization of moving objects in videos remains a significant challenge, with edge information playing a crucial role in mitigating this difficulty. [11, 12, 61-64]. However, these methods require complex network designs like extra convolution layers, greatly affecting inference efficiency. Canny edge detection [14, 65-67] is a popular algorithm for identifying sharp in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14432" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.089, + 0.568, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.089, + 0.885, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.327, + 0.907, + 0.357 + ], + "angle": 0, + "content": "Figure 2. Overview of our proposed method: the left side shows the overall framework, while the right side provides a detailed illustration of the structure refinement process. The light grey box on the left indicates evidential learning, and \"Dist.\" is an abbreviation for \"Distribution.\"" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.368, + 0.483, + 0.415 + ], + "angle": 0, + "content": "tensity changes in images, which can help identify the edges of all regions, making it a valuable prior for enhancing edge prediction. [68-71]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.44, + 0.461, + 0.459 + ], + "angle": 0, + "content": "2.3. Video Segmentation with Edge Information" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.468, + 0.483, + 0.681 + ], + "angle": 0, + "content": "Edge information can effectively capture object movements, offering significant potential for advancing video object segmentation. Studies such as [11, 12] demonstrated that edge information can aid in analyzing motion magnitude in videos. [61] further enhances this by combining flow edges with optical flow to track moving objects. Moreover, edge information has been utilized for shadow segmentation in videos [72], while [13] leverages edge information to derive discriminative representations, facilitating the differentiation of various semantic regions. However, these approaches either operate solely at the image level without distinguishing objects or are unable to generate fine-grained segmentation results for object tracking, rendering them unsuitable for semi-supervised video object segmentation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.707, + 0.308, + 0.723 + ], + "angle": 0, + "content": "2.4. Uncertainty Estimation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.9 + ], + "angle": 0, + "content": "Uncertainty estimation involves predicting the level of confidence or uncertainty in a network's output, encompassing aleatoric uncertainty (arising from data noise) and epistemic uncertainty (stemming from model limitations) [73]. Early approaches employed methods like particle filtering and CRFs [74, 75]. With the rise of deep learning, Bayesian neural networks [76-78] were introduced, along with other techniques such as ensemble methods [79]. More recently, evidential learning [73, 80-82] has shown promising results, which has the potential to improve the high uncertainty of occlusion areas in VOS tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.366, + 0.651, + 0.384 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.392, + 0.805, + 0.409 + ], + "angle": 0, + "content": "3.1. Inherent Structure-guided SVOS" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.415, + 0.907, + 0.492 + ], + "angle": 0, + "content": "Problem Formulation Given an object mask for reference, existing memory-based SVOS methods typically learn to segment objects sequentially across frames while concurrently updating the model memory for future regards. This procedure can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.648, + 0.504, + 0.771, + 0.521 + ], + "angle": 0, + "content": "\\[\nq _ {t} = h \\left(\\varphi \\left(i _ {t}\\right), y _ {t}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.516, + 0.906, + 0.539 + ], + "angle": 0, + "content": "\\[\ny _ {t} = \\varrho \\left(i _ {t - 1}, \\dots , i _ {t - n} \\mid q _ {t - 1}, \\dots , q _ {t - n}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.906, + 0.673 + ], + "angle": 0, + "content": "where \\( i_{t} \\) denotes the input frame and \\( y_{t} \\) is the constructed memory from previous \\( n \\) frames and predictions, acting as reference at time \\( t \\). Functions \\( \\varphi \\) and \\( \\varrho \\) represent the feature extractor and memory constructor, respectively. The notation \\( h \\) refers to the object segmentation network, which utilizes frame features \\( \\varphi(i_{t}) \\) and memory \\( y \\) to produce the segmentation output \\( q \\) at time \\( t \\). This pipeline interprets the SVOS task from the perspective of the frame content." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.909, + 0.795 + ], + "angle": 0, + "content": "However, we observe that Eq. 1 lacks adequate modeling of structural information and hierarchy relationships. To enhance the capacity of SVOS and better address issues of occlusion and similarity-induced confusion, we introduce a novel structure refinement module to the existing framework. This module provides the model with inherent structural information and hierarchical relationships in each input frame. Fig. 2 shows the overall architecture of our proposed model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Specifically, given an input frame \\( i_{t} \\), we first derive its edge representation using the Canny filter \\( c \\), a robust algorithm known for accurately detecting edges despite the noise. The frame features \\( \\varphi(i_{t}) \\) then interact with the edge representation to generate an inherent structure map \\( s \\) under the guidance of the memory features, capturing key structural characteristics of target objects. This inherent structure map" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14433" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.152 + ], + "angle": 0, + "content": "is subsequently integrated back into the frame features. Finally, we align memory features and structured enhanced features to predict the target object mask. Accordingly, our proposed solution can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.162, + 0.368, + 0.195 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} q _ {t} = h (\\varphi (i _ {t}), y _ {t}, s _ {t}), \\\\ s _ {t} = \\vartheta (\\varphi (i _ {t}), c (i _ {t}), y _ {t}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.203, + 0.485, + 0.279 + ], + "angle": 0, + "content": "where \\(\\vartheta\\) denotes the fusion function for processing features and inherent structure map. By explicitly modeling structural information, our approach allows the model to better distinguish objects, even in challenging scenarios involving occlusion and similar semantics." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.287, + 0.405, + 0.304 + ], + "angle": 0, + "content": "3.2. Inherent Structure Map Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.309, + 0.484, + 0.415 + ], + "angle": 0, + "content": "Edge Highlighting Given an input RGB image \\( I \\in \\mathbb{R}^{3 \\times H \\times W} \\), we first derive its edge representation \\( I_{c} \\in \\mathbb{R}^{1 \\times H \\times W} \\) using the Canny filter. Next, \\( I \\) is fed into an image encoder for multi-scale feature extraction. The output of the image encoder can be formulated as \\( \\hat{I}_{i} \\in \\mathbb{R}^{c_{i} \\times h_{i} \\times w_{i}} \\) for \\( i = 1, \\dots, N \\), where \\( N \\) denotes the number of feature scales, \\( h_{i} = \\frac{H}{2^{i+1}} \\) and \\( w_{i} = \\frac{W}{2^{i+1}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.415, + 0.484, + 0.566 + ], + "angle": 0, + "content": "Since the image encoder is pre-trained on ImageNet [83], its output feature map tends to emphasize abstract, content-focused information. In the case of SVOS, this nature may impede the model to accurately segment objects, potentially leading to confusion when occlusions occur among fragmented objects. To enhance the sensitivity of model to edges and structural details, we incorporate edge representation from \\( I_{c} \\) into \\( \\hat{I}_i \\) through an element-wise multiplication, highlighting edge out of the feature map. This procedure can be formalized as:" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.573, + 0.484, + 0.591 + ], + "angle": 0, + "content": "\\[\n\\tilde {I} _ {i} = \\hat {I} _ {i} + (\\hat {I} _ {i} \\odot \\epsilon I _ {c} ^ {\\downarrow h _ {i}, w _ {i}}), \\text {f o r} i = 1, \\dots , N, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.484, + 0.675 + ], + "angle": 0, + "content": "where \\(\\tilde{I}_i\\) denotes the global-edge-enhanced feature map, \\(\\odot\\) represents the Hadamard product, \\(\\epsilon\\) is an importance factor that modulates the influence of the edge map detected by the Canny filter, and \\(I_c^{\\downarrow h_i,w_i}\\) indicates that \\(I_{c}\\) has been interpolated to match the spatial dimensions of \\(\\hat{I}_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.824 + ], + "angle": 0, + "content": "Memory Feature Extraction Given the \\(t\\)-th frame of the video, we choose the previous \\(n\\) frames and their corresponding predicted masks and feed them into the memory encoder to derive the global memory feature \\(N\\). Memory encoder is a common-used module in SVOS, where we adopt the design from Cutie [7]. We then project global features onto the object-based masks to derive object memory features \\(M \\in \\mathbb{R}^{C \\times H \\times W}\\) (with \\(H = W = 30\\) here), which provides temporal information about the target object for the next structure refinement module." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Supervision Preparation Although the global-edge-enhanced feature map \\(\\tilde{I}_i\\) emphasizes edges, it still contains boundary details of non-target objects, which is insufficient for guiding the model toward precise segmentation. To address this limitation, we design a structure decoder to predict" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.257 + ], + "angle": 0, + "content": "structure map \\(S\\), containing only edges and structure details of targets, from the global-edge-enhanced feature \\(\\tilde{I}_i\\). For supervision, the ground-truth structure map is derived from ground-truth mask. Given a colored mask \\(G \\in \\mathbb{R}^{3 \\times H \\times W}\\) corresponding to the image \\(I\\), we first convert it into a grayscale image \\(G_{gr} \\in \\mathbb{R}^{1 \\times H \\times W}\\). Subsequently, a predefined filter is applied to transform \\(G_{gr}\\) into the final object-related inherent structure map \\(S\\). This generated object structure map effectively highlights occlusion relationships among target objects and is further used as ground truth in structural extraction network training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.909, + 0.424 + ], + "angle": 0, + "content": "Object Structure Map Prediction After obtaining the global-edge-enhanced feature map \\(\\tilde{I}_i\\) and the object memory feature \\(M\\) from the previous framework, a lightweight structure decoder with connecting paths is applied to predict structure map \\(S\\). Importantly, \\(\\tilde{I}_i\\) alone provides limited target-specific cues, and the network could only rely on global image features, which may compromise the accuracy of the target objects' structure map prediction. Therefore, the fusion of the object memory feature \\(M\\) to incorporate object-specific information is necessary. This process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.559, + 0.435, + 0.907, + 0.453 + ], + "angle": 0, + "content": "\\[\nS ^ {\\prime} = \\text {S t r u c t u r e D e c o d e r} (\\tilde {I} _ {i}, M), \\text {f o r} S ^ {\\prime} \\rightarrow S, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.466, + 0.909, + 0.604 + ], + "angle": 0, + "content": "where \\(M\\) is fused with the coarsest scale \\(\\tilde{I}_3\\) via elementwise addition. The decoder progressively upsamples the fused features with transposed convolutions, incorporating higher-resolution features via channel concatenation. To reduce overhead, binary cross-entropy loss is applied only at the final output. Visual analysis in Fig. 3 reveals that incorporating object memory features is crucial for accurately capturing the target objects' structure. Otherwise, the network lacks focus, leading to noisy predictions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.612, + 0.75, + 0.629 + ], + "angle": 0, + "content": "3.3. Refinement and Decoding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.634, + 0.906, + 0.709 + ], + "angle": 0, + "content": "After obtaining the final object-level structure map, we apply the Hadamard product with the extracted image features to emphasize object boundary features, thus to enhance feature with structure details. This process can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.72, + 0.907, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\hat {I} _ {i} ^ {\\prime} = \\hat {I} _ {i} + \\left(\\hat {I} _ {i} \\odot \\beta S ^ {\\prime \\downarrow h _ {i}, w _ {i}}\\right), \\text {f o r} i = 1, \\dots , N, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.854 + ], + "angle": 0, + "content": "where \\(\\hat{I}_i^{\\prime}\\) denotes the features ultimately used to segment the object mask and \\(\\beta\\) is an importance factor that modulates the influence of the structure map. Note that the structure map used in this process is in the form of logits rather than a binary mask, which helps to emphasize certain regions when the structure decoder has higher confidence in these areas being boundaries." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "With the structure-enhanced feature (SE feature) of the current frame, the global memory features \\( N \\), and the object memory features \\( M \\), we employ a decoder to align them and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14434" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.09, + 0.287, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.175, + 0.284, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.308, + 0.222, + 0.333 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.299, + 0.096, + 0.455, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.175, + 0.462, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.308, + 0.422, + 0.332 + ], + "angle": 0, + "content": "(d) : Eureka!" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.337, + 0.486, + 0.407 + ], + "angle": 0, + "content": "Figure 3. Output structure maps from different objects. (a) Input frame. (b) Input frame and predicted masks of the previous frame. (c) Output structure map without object fusion. (d) Output structure map with object fusion. \"Feat.\" and \"Mem.\" are abbreviations of \"Feature\" and \"Memory\" respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.418, + 0.484, + 0.569 + ], + "angle": 0, + "content": "output the final object-based mask. The memory-based decoder design is inspired by Cutie [7], a pioneering approach to enhance VOS performance by object memory modeling. Concretely, the decoder integrates top-down object features with bottom-up global features so that a dynamic interaction between high-level and low-level representations is established. The SE feature, in this way, serves as an additional pathway for connecting high-level and low-level representations, as the structure encodes object information and latently emphasizes the original feature map accordingly." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.577, + 0.473, + 0.594 + ], + "angle": 0, + "content": "3.4. Enhancement via Uncertainty Quantification" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.484, + 0.78 + ], + "angle": 0, + "content": "Due to the LogSoftmax used in the cross-entropy loss function, the predicted class probabilities remain point-based estimations that collapse multiple types of uncertainty into overconfident predictions [80, 84]. To address this, we further borrow the ideas of Evidential Deep Learning (EDL) to add an uncertainty-related loss term into our training pipeline, as being supplementary to existing loss combination, cross-entropy loss \\(\\mathcal{L}_{\\mathrm{CE}}\\) and dice loss \\(\\mathcal{L}_{\\mathrm{Dice}}\\), to help improve performance by quantifying and minimizing the uncertainty of predictions, which is especially crucial in challenging scenarios involving occlusions and features with high similarity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.486, + 0.902 + ], + "angle": 0, + "content": "The EDL based on the theory of subjective logic is used to calculate uncertainty with the Dirichlet distribution which acts as an updating measure of the model's intrinsic prediction bias and provides the reference for the subsequent \\(\\mathrm{KL}_{\\mathrm{div}}\\) regularization. Before training, we initialize a bare Dirichlet \\(D(\\mathbf{p}|\\boldsymbol{\\alpha})\\) over the class probability vector \\(\\mathbf{p} = [p_1,\\dots ,p_k]\\) and set all \\(\\alpha = 1\\), which by Gibbs' inequality means the distribution of maximum uncertainty before seeing any data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.198 + ], + "angle": 0, + "content": "During training, we update Dirichlet parameters with the model's output prob-logits to represent updated intrinsic bias. For each pixel at frame \\( t \\), we define \\( q_{t}^{k} \\) as the Boolean ground-truth label of a certain pixel of \\( k \\)-th class at time \\( t \\), and raw output logit \\( \\dot{q}_{t} \\) is \\( k \\)-class prob-logit vector of the pixel output by the model. The evidential loss \\( \\mathcal{L}_{\\mathrm{EDL}} \\) is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.205, + 0.907, + 0.247 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {E D L}} = \\sum_ {k = 1} ^ {K} q _ {t} ^ {k} \\cdot \\left(\\psi (\\sum \\alpha) - \\psi (\\alpha)\\right) + \\mathrm {K L} _ {\\mathrm {d i v}}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.645, + 0.25, + 0.774, + 0.266 + ], + "angle": 0, + "content": "\\[\n\\text {f o r} \\alpha = 1 + \\eta (\\dot {q} _ {t}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.276, + 0.907, + 0.337 + ], + "angle": 0, + "content": "where \\(\\alpha\\) means the Dirichlet parameter, \\(\\eta\\) is the confidence function and \\(\\psi\\) denotes the digamma function, and \\(\\mathrm{KL}_{\\mathrm{div}}\\) is derived corresponding to [80]. Subsequently, the total loss for mask predictions \\(\\mathcal{L}_{\\mathrm{mask}}\\) during training can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.347, + 0.907, + 0.363 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a s k}} = \\mathcal {L} _ {\\mathrm {C E}} + \\mathcal {L} _ {\\mathrm {D i c e}} + \\lambda \\mathcal {L} _ {\\mathrm {E D L}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.373, + 0.909, + 0.463 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is the hyperparameter to balance the contribution of the uncertainty loss term. With this combination of loss functions, our model benefits from the discriminative power of \\(\\mathcal{L}_{\\mathrm{CE}}\\) and \\(\\mathcal{L}_{\\mathrm{Dice}}\\) while leveraging \\(\\mathcal{L}_{\\mathrm{EDL}}\\) to address uncertain predictions, thus yielding improved segmentation performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.477, + 0.646, + 0.494 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.502, + 0.731, + 0.517 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.909, + 0.689 + ], + "angle": 0, + "content": "Network Configuration We employ two ResNet-18 models pretrained on ImageNet to extract image features and encode the mask, respectively. The dimension of the object feature is set to 256. For the structure decoder, we progressively upsample with each step employing a transposed convolution kernel of size 2 and a stride of 2 to achieve \\(2 \\times\\) upsampling. LeakyReLU is used as the activation function and residual connections are applied within this module. For the prediction layer, a \\(1 \\times 1\\) convolution is applied to output the binary segmentation. Due to space limitations, further details are in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Training Pipeline Similar to previous work, we divide the training into two stages. First, OASIS is pre-trained on static image datasets [88-92] with a 3-frame synthesis following [7]. For these images, we apply mirroring, random affine transformations cut-and-paste as data augmentation and combine different images to create pseudo-video sequences. Default cropping size and total training iteration are set to \\(384 \\times 384\\) and \\(80\\mathrm{K}\\), respectively. The batch size is set to 16, and point supervision is utilized with a total of 8192 sampled points. The learning rate is fixed at \\(1.0 \\times 10^{-4}\\), and the max norm of gradient clipping is set to 3. Notably, due to noise introduced by pseudo-video sequences, the structure decoder and EDL loss component are not included at this stage for the purpose of robustness." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14435" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.907, + 0.269 + ], + "angle": 0, + "content": "
MethodPublicationMOSEDAV-17 ValDAV-17 TestYouTubeVOS-2019 Val
J&FJFJ&FJFJ&FJFGJsFsJuFu
STCN [37]NeurIPS'2152.548.556.685.482.288.676.172.779.682.781.185.478.285.9
AOT-R50 [4]NeurIPS'2158.454.362.684.982.387.579.675.983.385.383.988.879.988.5
RDE [85]CVPR'2246.842.451.384.280.887.577.473.681.281.981.185.576.284.8
XMem [3]ECCV'2256.352.160.686.282.989.581.077.484.585.584.388.680.388.6
LLB [86]AAAI'23---84.681.587.7---83.681.786.579.287.0
ISVOS [87]CVPR'23---87.183.790.582.879.386.286.185.289.780.788.9
DeAOT-R50 [5]NeurIPS'2259.054.663.485.282.288.280.776.984.585.684.289.280.288.8
DEVA [6]ICCV'2360.055.864.386.883.690.082.378.785.985.585.089.479.788.0
Cutie [7]CVPR'2461.257.265.386.883.889.782.879.286.586.285.389.680.989.0
OASIS-62.157.866.388.385.091.683.279.387.086.685.790.181.289.4
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.272, + 0.907, + 0.313 + ], + "angle": 0, + "content": "Table 1. Overall performance comparison on VOS benchmarks. Numbers are taken from [7, 86, 87]. For the current state-of-the-art method, Cutie [7], we conduct training from scratch using its publicly available code and re-evaluate its performance on the benchmarks. Boldface and bold blue text indicate the best and second-best performance in each column, respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.327, + 0.486, + 0.568 + ], + "angle": 0, + "content": "For the main training stage, the batch size is set to 16, and is conducted for a total of \\(125\\mathrm{K}\\) iterations. The base learning rate is initialized at \\(1.0 \\times 10^{-4}\\), with the backbone learning rate scaled by a factor of 0.1. A step scheduler is employed to control the learning rate, reducing it by a factor of 0.1 at \\(100\\mathrm{K}\\) and \\(115\\mathrm{K}\\) iterations. The sequence length is set to 8, with each frame cropped to a size of \\(480 \\times 480\\) pixels. The EDL loss coefficient \\(\\lambda\\) is set to 0.01 with an additional annealing factor to gradually modify the impact of the loss. For the Canny filter, the lower and upper thresholds are set to 50 and 200, respectively, and an \\(L_{2}\\) regularization term is applied to enhance accuracy. Parameters \\(\\epsilon\\) and \\(\\beta\\) are set to 0.5 and 1.0 for rough edge and structure map feature fusion, respectively. The number of sampled points for the loss calculation is increased to 12544. We implement the model using PyTorch and train it on four A5000 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.577, + 0.473, + 0.593 + ], + "angle": 0, + "content": "4.2. Benchmark Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.599, + 0.485, + 0.795 + ], + "angle": 0, + "content": "In our experiments, several commonly used SVOS datasets are used to verify OASIS, including DAVIS-2017 [2], YouTubeVOS 2019 [93], and MOSE [94]. Specifically, DAVIS-2017 consists of 60 densely annotated videos in the training set, with an additional 30 videos in both the validation and test-dev sets for evaluation. YouTubeVOS-2019 contains 3978 videos sourced from video classification datasets, divided into 3471 for training and 507 for validation. MOSE, known for its crowded environments, increased occlusions, and higher rates of object disappearance, comprises 1507 videos for training, with 311 videos in the validation set. Following previous studies, we conduct experiments using various training dataset settings and reported the results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.486, + 0.903 + ], + "angle": 0, + "content": "In our experiments, three commonly used criteria, Jaccard index \\(\\mathcal{I}\\), contour accuracy \\(\\mathcal{F}\\) and their average \\(\\mathcal{I} \\& \\mathcal{F}\\) are employed to quantify the performance of our proposed method. Among them, the Jaccard index assesses the area-based accuracy of predictions, whereas the contour accuracy measures the localization accuracy of object boundaries. Specifically, the YouTubeVOS evaluation provides detailed" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.327, + 0.909, + 0.465 + ], + "angle": 0, + "content": "results for \\(\\mathcal{I}\\) and \\(\\mathcal{F}\\) on both seen and unseen objects, and \\(\\mathcal{G}\\) is the average of all objects, enabling a more comprehensive assessment of the generalization ability of a model. \\(\\mathcal{I},\\mathcal{F},\\) \\(\\mathcal{J}\\& \\mathcal{F}\\) and \\(\\mathcal{G}\\) range from 0 to 1 and the reported numbers in our paper are scaled for better readability and formatting, following previous SVOS works [3, 6, 7]. A superior performance results in values closer to 100. Frames per second (FPS) is also employed in ablation study to measure the efficiency of a model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.473, + 0.813, + 0.489 + ], + "angle": 0, + "content": "4.3. Overall Performance Comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.495, + 0.909, + 0.661 + ], + "angle": 0, + "content": "Our experiments compared OASIS with nine representative or state-of-the-art memory-based SVOS methods. Among them, STCN [37] and LLB [86] focus on achieving more accurate matching between query frames and memory banks. RDE [85], XMem [3], and DEVA [6] aims to reduce memory usage and improve efficiency, while AOT [4], DEAOT [5], ISVOS [87], and Cutie [7] emphasize enhanced object modeling. The analysis of experiments is as follows. Note that, the inference results of MOSE, YouTubeVOS, and DAVIS test-dev benchmarks are then packaged and uploaded to the evaluation server to obtain final results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.671, + 0.727, + 0.685 + ], + "angle": 0, + "content": "4.3.1. Quantitative Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Following the setup in previous works [6, 7], we trained OASIS using a combination of the DAVIS-2017 and YouTube 2019 training sets and evaluated its performance across four validation sets. It can be seen from Tab. 1 that among all of these SVOS models, our OASIS consistently achieves the best segmentation performance across all benchmarks. Since these four standard benchmarks cover diverse video content and annotation conditions, i.e., different levels of granularity, achieving top performance on all of them is particularly challenging. These results, therefore, underscore the effectiveness and stability of OASIS. Specifically, for datasets with finer annotation granularity, such as the DAVIS-17 validation set, our model outperforms others by substantial margins of 1.2, 1.2, and 1.1 on the \\(\\mathcal{J}\\&\\mathcal{F}, \\mathcal{J}\\)," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14436" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.09, + 0.861, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.254, + 0.908, + 0.283 + ], + "angle": 0, + "content": "Figure 4. Comparison of segmentation results on two video clips, with details zoomed in on the upper right corner. Notably, our model demonstrates superior resolving capability in occlusion regions and when the target exhibits high similarity to the background." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.293, + 0.486, + 0.412 + ], + "angle": 0, + "content": "
MethodMOSEDAV-17 Test
J&FJFJ&FJF
XMem [3]59.655.463.779.676.183.0
DeAOT-R50 [5]64.159.568.782.879.186.5
DEVA [6]66.061.870.382.678.986.4
Cutie [7]67.062.871.383.780.087.4
OASIS67.563.271.985.481.789.3
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.413, + 0.485, + 0.469 + ], + "angle": 0, + "content": "Table 2. Performance comparison under Trained-with-MOSE setting on VOS benchmarks. Numbers are taken from [7]. Boldface and bold blue text indicate the best and second-best performance in each column, respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.481, + 0.483, + 0.617 + ], + "angle": 0, + "content": "and \\(\\mathcal{F}\\) scores, respectively. Such improvement highlights the effectiveness of structure refinement in capturing object details. Moreover, OASIS remains stable performance gains on the MOSE benchmark, which contains more complex and densely packed objects that differ considerably from the training data. The contour accuracy \\((\\mathcal{F})\\) improves significantly by 1.0, demonstrating that our approach enhances the structural integrity of segmentation, even in unseen and complex scenes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.617, + 0.483, + 0.769 + ], + "angle": 0, + "content": "Tab. 2 presents additional results where the training set includes the MOSE benchmark, allowing for a more comprehensive performance evaluation. In this setting, our approach still demonstrates superior results, significantly outperforming other methods on the DAVIS-2017 test set by margins of 1.7, 1.7, and 1.9 on the \\(\\mathcal{J}\\& \\mathcal{F},\\mathcal{J},\\) and \\(\\mathcal{F}\\) scores, respectively. These results indicate that incorporating more complex scenes in the training set allows our model to develop a stronger structural understanding, leading to improved segmentation accuracy." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.776, + 0.31, + 0.792 + ], + "angle": 0, + "content": "4.3.2. Qualitative Experiments" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.795, + 0.486, + 0.903 + ], + "angle": 0, + "content": "We illustrate the visual results of our method and the SOTA one on two typical video clips from the YouTubeVOS 2019 validation set in Fig. 4. It can be observed that the existing SOTA method often struggles to distinguish between objects, particularly in cases of occlusion or high similarity between objects or between foreground objects and background. For instance, it fails to separate the fish hiding" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.293, + 0.905, + 0.413 + ], + "angle": 0, + "content": "
MethodDAV-17 ValDAV-17 Test-dev
J&FJFJ&FJF
Baseline84.881.688.079.976.083.7
+CaE84.981.888.180.076.084.0
+CaE+SD85.182.088.180.376.384.5
+CaE+SD*85.482.288.680.876.884.9
OASIS85.682.488.781.076.985.1
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.414, + 0.908, + 0.441 + ], + "angle": 0, + "content": "Table 3. Ablation experiments about key components on DAVIS-2017 validation set and test-dev set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.455, + 0.909, + 0.593 + ], + "angle": 0, + "content": "behind the other fish (the 1 - 3 rows). In contrast, due to the proposed evidential learning, our OASIS is able to model the structure of the overlapped fish better and accomplishes precise segmentation in the first clip. Additionally, we observe that previous methods exhibit accumulating errors over time when segmenting the antennae of a snail (rows 4-6), due to their reliance on memory mechanisms. In contrast, our proposed structure refinement module corrects these inaccuracies, significantly reducing this issue." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.602, + 0.669, + 0.619 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.625, + 0.768, + 0.641 + ], + "angle": 0, + "content": "4.4.1. Ablation on Key Components" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.901 + ], + "angle": 0, + "content": "OASIS involves three novel designs: the integration of rough edges detected by the Canny filter, structure decoder for refinement, and evidential learning loss for uncertainty quantification. We conduct ablation experiments to assess the individual contribution of each component, with an additional experiment on the object features fusion mechanism in the structure decoder. To this end, we build five experimental settings in total: \\(+\\mathrm{CaE}\\) refers to the baseline model with pixel feature integrated with the rough Canny edges as per Eq. 3, whereas it does not include the structure decoder and evidential learning. As shown in Fig. 3, \\(+\\mathrm{CaE} + \\mathrm{SD}\\) means the structure decoder is further applied on top of \\(+\\mathrm{CaE}\\) for the refined structure map, while \\(+\\mathrm{CaE} + \\mathrm{SD}^*\\) applied the object memory feature fusion mechanism. The final OASIS further adds uncertainty learning upon the \\(+\\mathrm{CaE} + \\mathrm{SD}^*\\). Due to limited resources, all models are trained on the DAVIS2017 training set with limited iterations and evaluated on the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14437" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.089, + 0.462, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.257, + 0.483, + 0.298 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison from different settings in ablating key components. We show the results of four typical frames. Zoom in for a better view." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.316, + 0.365, + 0.33 + ], + "angle": 0, + "content": "DAVIS-2017 validation and test-dev sets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.333, + 0.484, + 0.501 + ], + "angle": 0, + "content": "Tab. 3 shows the results of our ablation study. As observed on these benchmarks, Canny edge feature integration and the structure decoder are effective in characterizing the object and thus contribute to the overall segmentation performance. Object fusion further improves the capabilities of structure decoder, i.e., \\(+\\mathrm{CaE} + \\mathrm{SD^{*}}\\) outperforms the one without object fusion largely especially on the test-dev set, with all indicators increased by about 0.5. Furthermore, the further introduction of uncertainty estimation and minimization allows our complete model to attain the best performance with negligible additional expense." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.513, + 0.416, + 0.529 + ], + "angle": 0, + "content": "4.4.2. Visual Comparison on Key Components" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.534, + 0.484, + 0.671 + ], + "angle": 0, + "content": "Fig. 5 presents the visual results of different settings on representative frames. It can be observed that although \\(+\\mathrm{CaE}\\) achieves decent performance, structure refinement is still necessary to enhance output quality, especially for occlusions. For instance, at the position indicated by the white arrow (i.e., the body of the dog and the corner of the box), the implementation of the structure decoder significantly improves the structural details. This observation further validates the effectiveness of our design." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.684, + 0.325, + 0.699 + ], + "angle": 0, + "content": "4.4.3. Ablation on Factor \\(\\epsilon\\) and \\(\\beta\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.9 + ], + "angle": 0, + "content": "In Eq. 3 and Eq. 5 we use factors \\(\\epsilon\\) and \\(\\beta\\) to control the impact brought by the rough edges and the refined structure map, respectively. We separately investigate the effect of different values for these two factors on model performance, with results presented in Tab. 4. Notably, as the two factors increase from 0, accuracy consistently improves, substantiating the benefit of emphasizing boundary features. However, we observe that this performance gain reaches an upper limit, beyond which excessively large factors introduce a substantial distributional shift in the features, ultimately degrading performance. The best results are achieved when \\(\\beta\\) is greater than \\(\\epsilon\\), further underscoring the importance of more detailed structural information." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.199 + ], + "angle": 0, + "content": "
βDAVIS-17 Val| |DAVIS-17 Val
J&FJF| |J&FJF
0.585.382.188.50.085.382.288.5
1.085.682.488.70.585.682.488.7
2.085.782.588.91.085.181.888.3
5.085.082.087.92.085.082.087.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.199, + 0.908, + 0.227 + ], + "angle": 0, + "content": "Table 4. Ablation experiments about factors \\( \\epsilon \\) and \\( \\beta \\) on the DAVIS-2017 validation set." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.246, + 0.912, + 0.35 + ], + "angle": 0, + "content": "
Method#ParamsDAVIS-17 Val
J&FJFFPS
Baseline29M84.881.688.052
+2M Proj Net31M85.182.088.137
OASIS31M85.682.488.748
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.352, + 0.908, + 0.38 + ], + "angle": 0, + "content": "Table 5. Experiments on the effect of including additional parameters. Results on the DAVIS-2017 validation set are shown." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.395, + 0.871, + 0.411 + ], + "angle": 0, + "content": "4.5. Performance Gain and Model Parameters" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.909, + 0.706 + ], + "angle": 0, + "content": "In deep learning-based SVOS methods, performance gains may result from an increase in model parameters. In our approach, the primary source of additional parameters is the structure decoder, which introduces approximately 2M learnable parameters. Compared with the whole framework, these additional parameters are almost negligible. To further validate that the effectiveness of OASIS does not primarily stem from the increased number of parameters, we conduct additional analytical experiments. To provide a fair comparison, we add a convolutional projection network of approximately 2M parameters to the baseline as a comparative method, while maintaining other settings and hyperparameters. Results are shown in Tab. 5. Notably, the method with a 2M-parameter network does not achieve better results than OASIS. Combined with the ablative results in Tab. 3, which compare model performance with and without the structure decoder, we conclude that the proposed method is highly effective and, more importantly, that its performance gain is not due to the increased parameter count." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.723, + 0.634, + 0.739 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this work, we propose a novel model named OASIS to address the challenge of occlusion in SVOS. OASIS first generates a structure map that captures the structural and layering information of predicted targets and leverages it to refine features, thus highlighting boundary regions. Additionally, we integrate evidential learning for uncertainty estimation to enhance the capability of OASIS to focus on complex and uncertain segmentation cases. Extensive experiments show the superior performance of OASIS while maintaining high computational efficiency." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14438" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.486, + 0.162 + ], + "angle": 0, + "content": "This work was supported by Ministry of Education Tier 2 grant, NUS, Singapore (T2EP20224-0028); and Ministry of Education Tier 1 grant, NUS, Singapore (24-1250-P0001)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.174, + 0.188, + 0.189 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.199, + 0.484, + 0.254 + ], + "angle": 0, + "content": "[1] S. Caelles, K.-K. Maninis, J. Pont-Tuset, L. Leal-Taixe, D. Cremers, and L. Van Gool, \"One-shot video object segmentation,\" in Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 221-230. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.484, + 0.324 + ], + "angle": 0, + "content": "[2] F. Perazzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, and A. Sorkine-Hornung, \"A benchmark dataset and evaluation methodology for video object segmentation,\" in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 724-732. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.326, + 0.484, + 0.38 + ], + "angle": 0, + "content": "[3] H. K. Cheng and A. G. Schwing, \"Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model,\" in European Conference on Computer Vision. Springer, 2022, pp. 640-658. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.484, + 0.435 + ], + "angle": 0, + "content": "[4] Z. Yang, Y. Wei, and Y. Yang, \"Associating objects with transformers for video object segmentation,\" Advances in Neural Information Processing Systems, vol. 34, pp. 2491-2502, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.484, + 0.492 + ], + "angle": 0, + "content": "[5] Z. Yang and Y. Yang, \"Decoupling features in hierarchical propagation for video object segmentation,\" Advances in Neural Information Processing Systems, vol. 35, pp. 36324-36336, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.494, + 0.484, + 0.549 + ], + "angle": 0, + "content": "[6] H. K. Cheng, S. W. Oh, B. Price, A. Schwing, and J.-Y. Lee, \"Tracking anything with decoupled video segmentation,\" in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 1316-1326. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.484, + 0.618 + ], + "angle": 0, + "content": "[7] H. K. Cheng, S. W. Oh, B. Price, J.-Y. Lee, and A. Schwing, \"Putting the object back into video object segmentation,\" in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 3151-3161. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.484, + 0.675 + ], + "angle": 0, + "content": "[8] Z. Huang, S. Yang, M. Zhou, Z. Gong, A. Abusorrah, C. Lin, and Z. Huang, \"Making accurate object detection at the edge: Review and new approach,\" Artificial Intelligence Review, vol. 55, no. 3, pp. 2245-2274, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.677, + 0.484, + 0.718 + ], + "angle": 0, + "content": "[9] D. Yang, B. Peng, Z. Al-Huda, A. Malik, and D. Zhai, “An overview of edge and object contour detection,” Neurocomputing, vol. 488, pp. 470–493, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.72, + 0.484, + 0.786 + ], + "angle": 0, + "content": "[10] C. L. Zitnick and P. Dollar, \"Edge boxes: Locating object proposals from edges,\" in Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. Springer, 2014, pp. 391-405. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.484, + 0.844 + ], + "angle": 0, + "content": "[11] C. Kim and J.-N. Hwang, \"Fast and automatic video object segmentation and tracking for content-based applications,\" IEEE transactions on circuits and systems for video technology, vol. 12, no. 2, pp. 122-129, 2002. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.899 + ], + "angle": 0, + "content": "[12] S. Wan, S. Ding, and C. Chen, \"Edge computing enabled video segmentation for real-time traffic monitoring in internet of vehicles,\" Pattern Recognition, vol. 121, p. 108146, 2022. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.199, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.147 + ], + "angle": 0, + "content": "[13] A. Khoreva, F. Galasso, M. Hein, and B. Schiele, \"Classifier based graph construction for video segmentation,\" in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015, pp. 951-960. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.907, + 0.189 + ], + "angle": 0, + "content": "[14] J. Canny, “A computational approach to edge detection,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. PAMI-8, no. 6, pp. 679–698, 1986. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.908, + 0.244 + ], + "angle": 0, + "content": "[15] J. Pont-Tuset, F. Perazzi, S. Caelles, P. Arbeláez, A. Sorkine-Hornung, and L. Van Gool, “The 2017 davis challenge on video object segmentation,” arXiv preprint arXiv:1704.00675, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.908, + 0.301 + ], + "angle": 0, + "content": "[16] P. Voigtlaender and B. Leibe, \"Online adaptation of convolutional neural networks for video object segmentation,\" in Proceedings of the British Machine Vision Conference (BMVC). BMVA Press, September 2017, pp. 116.1-116.13." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.303, + 0.907, + 0.357 + ], + "angle": 0, + "content": "[17] H. Xiao, J. Feng, G. Lin, Y. Liu, and M. Zhang, “Monet: Deep motion exploitation for video object segmentation,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 1140–1148." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.907, + 0.413 + ], + "angle": 0, + "content": "[18] J. Luiten, P. Voigtlaender, and B. Leibe, “Premvos: Proposal generation, refinement and merging for video object segmentation,” in Asian conference on computer vision. Springer, 2018, pp. 565–580." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.415, + 0.907, + 0.455 + ], + "angle": 0, + "content": "[19] Y.-T. Hu, J.-B. Huang, and A. Schwing, “Maskrnn: Instance level video object segmentation,” Advances in neural information processing systems, vol. 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.457, + 0.907, + 0.523 + ], + "angle": 0, + "content": "[20] F. Perazzi, A. Khoreva, R. Benenson, B. Schiele, and A. Sorkine-Hornung, \"Learning video object segmentation from static images,\" in Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 2663-2672. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.908, + 0.593 + ], + "angle": 0, + "content": "[21] K.-K. Maninis, S. Caelles, Y. Chen, J. Pont-Tuset, L. Leal-Taixe, D. Cremers, and L. Van Gool, \"Video object segmentation without temporal information,\" IEEE transactions on pattern analysis and machine intelligence, vol. 41, no. 6, pp. 1515-1530, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.908, + 0.65 + ], + "angle": 0, + "content": "[22] J. Cheng, Y.-H. Tsai, S. Wang, and M.-H. Yang, \"Segflow: Joint learning for video object segmentation and optical flow,\" in Proceedings of the IEEE international conference on computer vision, 2017, pp. 686-695. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.652, + 0.906, + 0.705 + ], + "angle": 0, + "content": "[23] X. Li and C. C. Loy, \"Video object segmentation with joint re-identification and attention-aware mask propagation,\" in Proceedings of the European conference on computer vision (ECCV), 2018, pp. 90-105." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.908, + 0.761 + ], + "angle": 0, + "content": "[24] S. W. Oh, J.-Y. Lee, K. Sunkavalli, and S. J. Kim, \"Fast video object segmentation by reference-guided mask propagation,\" in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 7376-7385." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.908, + 0.815 + ], + "angle": 0, + "content": "[25] Y.-H. Tsai, M.-H. Yang, and M. J. Black, \"Video segmentation via object flow,\" in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 3899-3908." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.907, + 0.872 + ], + "angle": 0, + "content": "[26] H. Xie, H. Yao, S. Zhou, S. Zhang, and W. Sun, \"Efficient regional memory network for video object segmentation,\" in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 1286-1295. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[27] L. Yang, Y. Wang, X. Xiong, J. Yang, and A. Katsaggelos, \"Efficient video object segmentation via network modulation,\"" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14439" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 6499-6507." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.486, + 0.188 + ], + "angle": 0, + "content": "[28] B. Duke, A. Ahmed, C. Wolf, P. Aarabi, and G. W. Taylor, \"Sstvos: Sparse spatiotemporal transformers for video object segmentation,\" in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 5912-5921." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.485, + 0.245 + ], + "angle": 0, + "content": "[29] K. Xu and A. Yao, \"Accelerating video object segmentation with compressed video,\" in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 1342-1351." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.484, + 0.302 + ], + "angle": 0, + "content": "[30] Y.-S. Xu, T.-J. Fu, H.-K. Yang, and C.-Y. Lee, \"Dynamic video segmentation network,\" in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 6556-6565. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.303, + 0.484, + 0.357 + ], + "angle": 0, + "content": "[31] Y. Chen, J. Pont-Tuset, A. Montes, and L. Van Gool, \"Blazingly fast video object segmentation with pixel-wise metric learning,\" in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 1189-1198. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.359, + 0.484, + 0.413 + ], + "angle": 0, + "content": "[32] Y.-T. Hu, J.-B. Huang, and A. G. Schwing, \"Videomatch: Matching based video object segmentation,\" in Proceedings of the European conference on computer vision (ECCV), 2018, pp. 54-70." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.484, + 0.469 + ], + "angle": 0, + "content": "[33] Z. Yang, Y. Wei, and Y. Yang, \"Collaborative video object segmentation by multi-scale foreground-background integration,\" IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 9, pp. 4701-4712, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.484, + 0.538 + ], + "angle": 0, + "content": "[34] P. Voigtlaender, Y. Chai, F. Schroff, H. Adam, B. Leibe, and L.-C. Chen, “Feelvos: Fast end-to-end embedding learning for video object segmentation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 9481–9490." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.484, + 0.594 + ], + "angle": 0, + "content": "[35] S. W. Oh, J.-Y. Lee, N. Xu, and S. J. Kim, \"Video object segmentation using space-time memory networks,\" in Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 9226-9235. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.484, + 0.664 + ], + "angle": 0, + "content": "[36] H. K. Cheng, Y.-W. Tai, and C.-K. Tang, \"Modular interactive video object segmentation: Interaction-to-mask, propagation and difference-aware fusion,\" in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 5559-5568." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.484, + 0.718 + ], + "angle": 0, + "content": "[37] ——, “Rethinking space-time networks with improved memory coverage for efficient video object segmentation,” Advances in Neural Information Processing Systems, vol. 34, pp. 11781-11794, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.484, + 0.788 + ], + "angle": 0, + "content": "[38] Y. Li, Z. Shen, and Y. Shan, \"Fast video object segmentation using the global context module,\" in Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16. Springer, 2020, pp. 735-750." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[39] Y. Liang, X. Li, N. Jafari, and J. Chen, \"Video object segmentation with adaptive feature bank and uncertain-region refinement,\" Advances in Neural Information Processing Systems, vol. 33, pp. 3430-3441, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[40] H. Wang, X. Jiang, H. Ren, Y. Hu, and S. Bai, \"Swiftnet: Real-time video object segmentation,\" in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 1296-1305. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.148 + ], + "angle": 0, + "content": "[41] Y. Mao, N. Wang, W. Zhou, and H. Li, \"Joint inductive and transductive learning for video object segmentation,\" in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 9670-9679. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.908, + 0.218 + ], + "angle": 0, + "content": "[42] H. Seong, J. Hyun, and E. Kim, \"Kernelized memory network for video object segmentation,\" in Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. Springer, 2020, pp. 629-645." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.908, + 0.29 + ], + "angle": 0, + "content": "[43] L. Hu, P. Zhang, B. Zhang, P. Pan, Y. Xu, and R. Jin, \"Learning position and target consistency for memory-based video object segmentation,\" in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 4144-4154." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.907, + 0.36 + ], + "angle": 0, + "content": "[44] X. Lu, W. Wang, M. Danelljan, T. Zhou, J. Shen, and L. Van Gool, \"Video object segmentation with episodic graph memory networks,\" in Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. Springer, 2020, pp. 661-679." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.907, + 0.417 + ], + "angle": 0, + "content": "[45] H. Seong, S. W. Oh, J.-Y. Lee, S. Lee, S. Lee, and E. Kim, \"Hierarchical memory matching network for video object segmentation,\" in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 12889-12898." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.907, + 0.488 + ], + "angle": 0, + "content": "[46] H. Liu, M. Gao, X. Luo, Z. Wang, G. Qin, J. Wu, and Y. Jin, \"Resurgsam2: Referring segment anything in surgical video via credible long-term tracking,\" in International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.907, + 0.545 + ], + "angle": 0, + "content": "[47] J. Pei, J. Zhang, G. Qin, K. Wang, Y. Jin, and P.-A. Heng, \"Instrument-tissue-guided surgical action triplet detection via textual-temporal trail exploration,\" IEEE Transactions on Medical Imaging, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.908, + 0.63 + ], + "angle": 0, + "content": "[48] N. Ravi, V. Gabeur, Y.-T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. Rädle, C. Rolland, L. Gustafson, E. Mintun, J. Pan, K. V. Alwala, N. Carion, C.-Y. Wu, R. Girshick, P. Dollar, and C. Feichtenhofer, “SAM 2: Segment anything in images and videos,” in The Thirteenth International Conference on Learning Representations, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.908, + 0.701 + ], + "angle": 0, + "content": "[49] Z. Wang, Y. Zhang, Y. Wang, L. Cai, and Y. Zhang, \"Dynamic pseudo label optimization in point-supervised nuclei segmentation,\" in International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2024, pp. 220-230. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.703, + 0.908, + 0.758 + ], + "angle": 0, + "content": "[50] Y. Zhang, Z. Wang, Y. Wang, H. Bian, L. Cai, H. Li, L. Zhang, and Y. Zhang, “Boundary-aware contrastive learning for semi-supervised nuclei instance segmentation,” in Medical Imaging with Deep Learning, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.907, + 0.814 + ], + "angle": 0, + "content": "[51] Y. Zhang, Y. Zhou, Y. Wang, J. Xiao, Z. Wang, Y. Zhang, and J. Chen, \"The four color theorem for cell instance segmentation,\" in Forty-second International Conference on Machine Learning, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.907, + 0.872 + ], + "angle": 0, + "content": "[52] J.-X. Zhao, J.-J. Liu, D.-P. Fan, Y. Cao, J. Yang, and M.-M. Cheng, \"Egnet: Edge guidance network for salient object detection,\" in Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 8779-8788. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[53] Z. Luo, A. Mishra, A. Achkar, J. Eichel, S. Li, and P.-M. Jodoin, \"Non-local deep features for salient object detection,\"" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14440" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "in Proceedings of the IEEE Conference on computer vision and pattern recognition, 2017, pp. 6609-6617." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.486, + 0.176 + ], + "angle": 0, + "content": "[54] Y. Zhuge, G. Yang, P. Zhang, and H. Lu, \"Boundary-guided feature aggregation network for salient object detection,\" IEEE Signal Processing Letters, vol. 25, no. 12, pp. 1800-1804, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.485, + 0.234 + ], + "angle": 0, + "content": "[55] Z. Tu, Y. Ma, C. Li, J. Tang, and B. Luo, \"Edge-guided nonlocal fully convolutional network for salient object detection,\" IEEE transactions on circuits and systems for video technology, vol. 31, no. 2, pp. 582-593, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.484, + 0.291 + ], + "angle": 0, + "content": "[56] X. Wang, H. Ma, X. Chen, and S. You, \"Edge preserving and multi-scale contextual neural network for salient object detection,\" IEEE Transactions on Image Processing, vol. 27, no. 1, pp. 121-134, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.484, + 0.347 + ], + "angle": 0, + "content": "[57] Z. Wu, L. Su, and Q. Huang, \"Stacked cross refinement network for edge-aware salient object detection,\" in Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 7264-7273. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.484, + 0.403 + ], + "angle": 0, + "content": "[58] G. Li, Y. Xie, L. Lin, and Y. Yu, \"Instance-level salient object segmentation,\" in Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 2386-2395. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.484, + 0.475 + ], + "angle": 0, + "content": "[59] Y. Hao, Y. Liu, Z. Wu, L. Han, Y. Chen, G. Chen, L. Chu, S. Tang, Z. Yu, Z. Chen et al., \"Edgeflow: Achieving practical interactive segmentation with edge-guided flow,\" in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 1551-1560." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.484, + 0.532 + ], + "angle": 0, + "content": "[60] H. Liu, R. A. R. Soto, F. Xiao, and Y. J. Lee, \"Yolactedge: Real-time instance segmentation on the edge,\" in 2021 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2021, pp. 9579-9585." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.484, + 0.588 + ], + "angle": 0, + "content": "[61] Y. Zhou, X. Xu, F. Shen, X. Zhu, and H. T. Shen, \"Flow-edge guided unsupervised video object segmentation,\" IEEE Transactions on circuits and systems for video technology, vol. 32, no. 12, pp. 8116-8127, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.484, + 0.645 + ], + "angle": 0, + "content": "[62] L. Liu, H. Li, and M. Gruteser, \"Edge assisted real-time object detection for mobile augmented reality,\" in The 25th annual international conference on mobile computing and networking, 2019, pp. 1-16." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.484, + 0.702 + ], + "angle": 0, + "content": "[63] C. Zhan, X. Duan, S. Xu, Z. Song, and M. Luo, \"An improved moving object detection algorithm based on frame difference and edge detection,\" in Fourth international conference on image and graphics (ICIG 2007). IEEE, 2007, pp. 519-523." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.484, + 0.759 + ], + "angle": 0, + "content": "[64] M. Hanyao, Y. Jin, Z. Qian, S. Zhang, and S. Lu, \"Edge-assisted online on-device object detection for real-time video analytics,\" in IEEE INFOCOM 2021-IEEE Conference on Computer Communications. IEEE, 2021, pp. 1-10. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.484, + 0.815 + ], + "angle": 0, + "content": "[65] Z. Xu, X. Baojie, and W. Guoxin, \"Canny edge detection based on open cv,\" in 2017 13th IEEE international conference on electronic measurement & instruments (ICEMI). IEEE, 2017, pp. 53-56. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[66] L. Xuan and Z. Hong, \"An improved canny edge detection algorithm,\" in 2017 8th IEEE international conference on software engineering and service science (ICSESS). IEEE, 2017, pp. 275-278." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[67] W. Rong, Z. Li, W. Zhang, and L. Sun, \"An improved canny edge detection algorithm,\" in 2014 IEEE international con" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.12 + ], + "angle": 0, + "content": "ference on mechatronics and automation. IEEE, 2014, pp. 577-582. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.908, + 0.15 + ], + "angle": 0, + "content": "[68] L. Ding and A. Goshtasby, “On the canny edge detector,” Pattern recognition, vol. 34, no. 3, pp. 721-725, 2001. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.152, + 0.908, + 0.191 + ], + "angle": 0, + "content": "[69] W. McIlhagga, “The canny edge detector revisited,” International Journal of Computer Vision, vol. 91, pp. 251–261, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.907, + 0.236 + ], + "angle": 0, + "content": "[70] R. Song, Z. Zhang, and H. Liu, \"Edge connection based canny edge detection algorithm,\" Pattern Recognition and Image Analysis, vol. 27, pp. 740-747, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.237, + 0.908, + 0.291 + ], + "angle": 0, + "content": "[71] P. Bao, L. Zhang, and X. Wu, \"Canny edge detection enhancement by scale multiplication,\" IEEE transactions on pattern analysis and machine intelligence, vol. 27, no. 9, pp. 1485-1490, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.294, + 0.908, + 0.349 + ], + "angle": 0, + "content": "[72] D. Xu, J. Liu, X. Li, Z. Liu, and X. Tang, \"Insignificant shadow detection for video segmentation,\" IEEE Transactions on Circuits and Systems for Video Technology, vol. 15, no. 8, pp. 1058-1064, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.351, + 0.908, + 0.419 + ], + "angle": 0, + "content": "[73] S. Seoni, V. Jahmunah, M. Salvi, P. D. Barua, F. Molinari, and U. R. Acharya, “Application of uncertainty quantification to artificial intelligence in healthcare: A review of last decade (2013–2023),\" Computers in Biology and Medicine, p. 107441, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.421, + 0.908, + 0.475 + ], + "angle": 0, + "content": "[74] A. Blake, R. Curwen, and A. Zisserman, “A framework for spatiotemporal control in the tracking of visual contours,” International Journal of Computer Vision, vol. 11, pp. 127–145, 1993. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.908, + 0.547 + ], + "angle": 0, + "content": "[75] X. He, R. S. Zemel, and M. A. Carreira-Perpinan, “Multiscale conditional random fields for image labeling,” in Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004., vol. 2. IEEE, 2004, pp. II-II. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.549, + 0.908, + 0.616 + ], + "angle": 0, + "content": "[76] D. Dera, N. C. Bouaynaya, G. Rasool, R. Shterenberg, and H. M. Fathallah-Shaykh, “Premium-cnn: Propagating uncertainty towards robust convolutional neural networks,” IEEE Transactions on Signal Processing, vol. 69, pp. 4669–4684, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.908, + 0.674 + ], + "angle": 0, + "content": "[77] Y. Gal and Z. Ghahramani, “Dropout as a bayesian approximation: Representing model uncertainty in deep learning,” in international conference on machine learning. PMLR, 2016, pp. 1050–1059." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.907, + 0.717 + ], + "angle": 0, + "content": "[78] A. Kendall and Y. Gal, “What uncertainties do we need in bayesian deep learning for computer vision?” Advances in neural information processing systems, vol. 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.719, + 0.908, + 0.773 + ], + "angle": 0, + "content": "[79] A. Ashukha, A. Lyzhov, D. Molchanov, and D. Vetrov, “Pitfalls of in-domain uncertainty estimation and ensembling in deep learning,” in The Eighth International Conference on Learning Representations, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.908, + 0.857 + ], + "angle": 0, + "content": "[80] M. Sensoy, L. Kaplan, and M. Kandemir, \"Evidential deep learning to quantify classification uncertainty,\" in Advances in Neural Information Processing Systems, S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, Eds., vol. 31. Curran Associates, Inc., 2018. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[81] K. Zou, Z. Chen, X. Yuan, X. Shen, M. Wang, and H. Fu, “A review of uncertainty estimation and its application in medical imaging,” Meta-Radiology, p. 100003, 2023." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "14441" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.146 + ], + "angle": 0, + "content": "[82] W. Bao, Q. Yu, and Y. Kong, \"Evidential deep learning for open set action recognition,\" in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 13349-13358. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.485, + 0.205 + ], + "angle": 0, + "content": "[83] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, \"Imagenet: A large-scale hierarchical image database,\" in 2009 IEEE Conference on Computer Vision and Pattern Recognition, 2009, pp. 248-255. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.485, + 0.289 + ], + "angle": 0, + "content": "[84] H. Yang, C. Chen, Y. CHEN, , H. C. Yip, and D. QI, \"Uncertainty estimation for safety-critical scene segmentation via fine-grained reward maximization,\" in Advances in Neural Information Processing Systems, A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, Eds., vol. 36. Curran Associates, Inc., 2023, pp. 36238-36249. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.485, + 0.346 + ], + "angle": 0, + "content": "[85] M. Li, L. Hu, Z. Xiong, B. Zhang, P. Pan, and D. Liu, \"Recurrent dynamic embedding for video object segmentation,\" in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 1332-1341. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.484, + 0.401 + ], + "angle": 0, + "content": "[86] M. Lan, J. Zhang, L. Zhang, and D. Tao, “Learning to learn better for video object segmentation,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 1, 2023, pp. 1205–1212. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.404, + 0.485, + 0.473 + ], + "angle": 0, + "content": "[87] J. Wang, D. Chen, Z. Wu, C. Luo, C. Tang, X. Dai, Y. Zhao, Y. Xie, L. Yuan, and Y.-G. Jiang, “Look before you match: Instance understanding matters in video object segmentation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 2268–2278. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.475, + 0.485, + 0.528 + ], + "angle": 0, + "content": "[88] J. Shi, Q. Yan, L. Xu, and J. Jia, \"Hierarchical image saliency detection on extended cssd,\" IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 38, no. 4, pp. 717-729, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.531, + 0.485, + 0.586 + ], + "angle": 0, + "content": "[89] L. Wang, H. Lu, Y. Wang, M. Feng, D. Wang, B. Yin, and X. Ruan, “Learning to detect salient objects with image-level supervision,” in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 3796–3805." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.588, + 0.485, + 0.642 + ], + "angle": 0, + "content": "[90] X. Li, T. Wei, Y. P. Chen, Y.-W. Tai, and C.-K. Tang, \"Fss1000: A 1000-class dataset for few-shot segmentation,\" in 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 2866-2875." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.645, + 0.485, + 0.699 + ], + "angle": 0, + "content": "[91] Y. Zeng, P. Zhang, Z. Lin, J. Zhang, and H. Lu, \"Towards high-resolution salient object detection,\" in 2019 IEEE/CVF International Conference on Computer Vision (ICCV), 2019, pp. 7233-7242." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.701, + 0.485, + 0.769 + ], + "angle": 0, + "content": "[92] H. K. Cheng, J. Chung, Y.-W. Tai, and C.-K. Tang, \"Cascadepsp: Toward class-agnostic and very high-resolution segmentation via global and local refinement,\" in 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 8887-8896. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.771, + 0.485, + 0.811 + ], + "angle": 0, + "content": "[93] N. Xu, L. Yang, Y. Fan, D. Yue, Y. Liang, J. Yang, and T. Huang, \"Youtube-vos: A large-scale video object segmentation benchmark,\" in ECCV, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.814, + 0.485, + 0.855 + ], + "angle": 0, + "content": "[94] H. Ding, C. Liu, S. He, X. Jiang, P. H. Torr, and S. Bai, “MOSE: A new dataset for video object segmentation in complex scenes,” in ICCV, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.855 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14442" + } + ] +] \ No newline at end of file diff --git a/2025/Structure Matters_ Revisiting Boundary Refinement in Video Object Segmentation/bf659278-d12b-4982-824e-fbf802268f8f_origin.pdf b/2025/Structure Matters_ Revisiting Boundary Refinement in Video Object Segmentation/bf659278-d12b-4982-824e-fbf802268f8f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..74eea0ebe24564d904b36e3f923218fa24fe5b68 --- /dev/null +++ b/2025/Structure Matters_ Revisiting Boundary Refinement in Video Object Segmentation/bf659278-d12b-4982-824e-fbf802268f8f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6c29baaf8dcaf5e107c2f635f9a41c9947da0551540337312125caaf258ab133 +size 4350613 diff --git a/2025/Structure Matters_ Revisiting Boundary Refinement in Video Object Segmentation/full.md b/2025/Structure Matters_ Revisiting Boundary Refinement in Video Object Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8902389dcee8efa7452e1a54ff9210e8941882da --- /dev/null +++ b/2025/Structure Matters_ Revisiting Boundary Refinement in Video Object Segmentation/full.md @@ -0,0 +1,373 @@ +# Structure Matters: Revisiting Boundary Refinement in Video Object Segmentation + +Guanyi Qin $^{1,*}$ , Ziyue Wang $^{1,*}$ , Daiyun Shen $^{1}$ , Haofeng Liu $^{1}$ , Hantao Zhou $^{2}$ , Junde Wu $^{3}$ , Runze Hu $^{4}$ , Yueming Jin $^{1,\dagger}$ + +$^{1}$ National University of Singapore; $^{2}$ Tsinghua University + +3University of Oxford; 4Beijing Institute of Technology + +{guanyi.qin, e1374378, e1374467, haofeng.liu}@u.nus.edu, hrz@bit.edu.cn, ymjin@nus.edu.sg + +# Abstract + +Given an object mask, Semi-supervised Video Object Segmentation (SVOS) technique aims to track and segment the object across video frames, serving as a fundamental task in computer vision. Although recent memory-based methods demonstrate potential, they often struggle with scenes involving occlusion, particularly in handling object interactions and high feature similarity. To address these issues and meet the real-time processing requirements of downstream applications, in this paper, we propose a novel bOundary Amendment video object Segmentation method with Inherent Structure refinement, hereby named OASIS. Specifically, a lightweight structure refinement module is proposed to enhance segmentation accuracy. With the fusion of rough edge priors captured by the Canny filter and stored object features, the module can generate an object-level structure map and refine the representations by highlighting boundary features. Evidential learning for uncertainty estimation is introduced to further address challenges in occluded regions. The proposed method, OASIS, maintains an efficient design, yet extensive experiments on challenging benchmarks demonstrate its superior performance and competitive inference speed compared to other state-of-the-art methods, i.e., achieving the $\mathcal{F}$ values of 91.6 (vs. 89.7 on DAVIS-17 validation set) and $\mathcal{G}$ values of 86.6 (vs. 86.2 on YouTubeVOS 2019 validation set) while maintaining a competitive speed of 48 FPS on DAVIS. Checkpoints, logs, and code will be available here. + +# 1. Introduction + +Video Object Segmentation (VOS) is an important task in computer vision [1, 2] that aims to separate moving objects from the background. In the general semi-supervised VOS (SVOS) setting, the task involves propagating the object masks specified in the first frame throughout the entire video. By enabling effective identification and understanding of + +![](images/5f514942aa9a78bd0b912b63ff997387c737f87c38383aff6b2844fd36e17277.jpg) +(b) + +![](images/168a0079287f9985ae3a47929dae1a363969b8888bdc53a09b7a362a302dbe27.jpg) + +![](images/5df03c903ecf7ffd778f058e2a40cad3a2ae4b68ef5bb4303fe15c59d355339b.jpg) +Figure 1. (a) Input frame. (b) Object mask generated by the SOTA method Cutie. (c) Output structure map. (d) Object mask generated by our model. In this case, the fish corresponding to the red mask closely resembles the background and is concealed behind the other two fish, causing previous methods to fail. Guided by our structure map, the model successfully segments the fish accurately. + +![](images/21333945ab59170b581462df22157f2f5ef1aef0a6272e209bf93b052466d2d5.jpg) +(d) + +objects in videos, VOS plays a crucial role in various applications including autonomous driving, robotics, video editing, medical intervention, etc. Albeit performance demands, real-time processing is also essential for these applications, necessitating algorithms that accurately perceive and segment objects while delivering sufficient inference speed. + +Recent advancements [3-6] in SVOS emphasize the value of memory-based mechanisms. Representations of certain past frames are generated and stored in a memory bank. The current frame is then used as the query to match these stored representations and derive segmentation masks. Cutie [7], further proposes object queries as a high-level summary of the target object and interacts with the pixel-level features of the query frame for accurate segmentation, achieving the state-of-the-art performance. However, these approaches still struggle when similar objects are close to each other or are occluded, as shown in Figure 1 (a) and (b). On one hand, the mask prediction depends predominantly on maintaining consistency across frames for both pixel- and object-level features, while the inherent geometric characteristics of the objects are often overlooked. This limitation poses challenges in accurately discerning object structures, particularly + +when they resemble the background (red box). On the other hand, when occlusion occurs, objects become fragmented. This challenge is amplified when objects share similar semantics, leading to high uncertainty in overlapping areas. Without an understanding of object structure, the model struggles to distinguish objects and establish clear boundaries (blue box). + +Edge information is a crucial component of an object's structure representation, defining its boundaries and shape by representing the inherent structural information. Precise edge prediction allows models to capture structural details more effectively, thereby enhancing the model's ability to understand occlusions. Therefore, edges have been widely used for segmenting objects for static images across various fields including object detection and semantic segmentation [8-10]. However, in VOS tasks, target movements lead to considerable shifts in object edges over time, making it challenging to predict accurate edges from scratch for each frame. This often requires large edge detection networks, which significantly reduce inference efficiency. Therefore, edges are only used to analyze motion magnitude or shadow area detection primarily, which merely requires coarse edges [11-13]. + +To tackle the aforementioned problems, we propose an bOuntary Amendment video object Segmentation with Inherent Structure refinement, hereby named OASIS, which effectively refines object structures to handle geometric and layer relationships among objects better, while highlighting them out of backgrounds. This is achieved through a simple but effective structure refinement module. Specifically, we first extract the rough edge information by utilizing a fast and robust edge detection algorithm, Canny Edge Detection [14], which can identify global edges in individual frames and provide strong priors. With these priors, we only need a lightweight structure decoder to obtain the edges of the target objects instead of predicting from scratch without significant sacrifice in inference speed. This map further amplifies the original features by highlighting objects' boundaries, producing structure-enhanced features for final segmentation. This process enables the model to differentiate among objects and from the background effectively. Additionally, an object feature fusion mechanism is introduced to integrate memory features into the structure map to better model relationships between objects. Furthermore, to address the high uncertainty in dealing with overlapped areas, we also make the first attempt to deliver a new learning pattern that incorporates Evidential Learning (EDL) for uncertainty quantification and introduces an additional loss to minimize uncertainty and improve segmentation accuracy in occlusion regions. + +Extensive experiments demonstrate the superiority of our approach across four benchmarks under two training settings, particularly in scenes where target objects interact and exhibit complex relationships. Notably, OASIS achieves + +promising results, consistently outperforming current state-of-the-art methods across different datasets, while delivering a real-time inference speed at 48 fps. In summary, our contributions are the following: + +- We propose OASIS, a novel framework designed to manage relationships between the objects and the background by accurately modeling object structures. +- We introduce a structure refinement module and an object sensory mechanism that utilizes edge information to achieve inherent structure refinement. +- To the best of our knowledge, we make the very first attempt to incorporate evidential learning in the SVOS task to improve the model performance on high-uncertainty regions, particularly in areas with significant similarity and occlusion. +- We verify OASIS on multiple SVOS benchmarks involving a wide range of video contents, annotation conditions, and training settings. OASIS outperforms other competitors across all these datasets and settings. + +# 2. Related Work + +# 2.1. Video Object Segmentation + +Early approaches of VOS [1, 2, 15-21] implement online learning that finetunes the model at test-time, resulting in low efficiency. Propagation-based methods [20, 22-30] iteratively propagate masks between adjacent frames by leveraging temporal correlations; however, they are inevitably constrained by drift and occlusion. Matching-based models [31-40] compute correspondences between the current and reference frames to capture long-range temporal context. Some memory-based methods [3, 26, 35-37, 39, 41-47] further enhance this by using memory banks to store features from past frames for pixel-level matching. The lately research SAM2 [48] also benefitted from memory mechanisms. Recently, Cutie [7] proposed an object-reading mechanism to enhance object-level representations. While it effectively perceives individual objects, modeling the relationships between multiple objects remains an open challenge. + +# 2.2. Edge Detection + +Edge information has been widely utilized to differentiate occluded objects [8-10, 49-51]. In salient object detection [52-57], edge features assist in identifying the most visually prominent objects within the image. Similarly, in segmentation tasks [11, 12, 58-61], edge detection helps to distinguish different regions and instances. Accurate localization of moving objects in videos remains a significant challenge, with edge information playing a crucial role in mitigating this difficulty. [11, 12, 61-64]. However, these methods require complex network designs like extra convolution layers, greatly affecting inference efficiency. Canny edge detection [14, 65-67] is a popular algorithm for identifying sharp in + +![](images/19a04cc1082cdfd93132d8a14f5e06fe7cc4dc8d74ee6d040b6bad68c241edb9.jpg) +Figure 2. Overview of our proposed method: the left side shows the overall framework, while the right side provides a detailed illustration of the structure refinement process. The light grey box on the left indicates evidential learning, and "Dist." is an abbreviation for "Distribution." + +![](images/47723baed477e527c48954574fe00432e1d8e375bcda66189bd7762e2f466e32.jpg) + +tensity changes in images, which can help identify the edges of all regions, making it a valuable prior for enhancing edge prediction. [68-71]. + +# 2.3. Video Segmentation with Edge Information + +Edge information can effectively capture object movements, offering significant potential for advancing video object segmentation. Studies such as [11, 12] demonstrated that edge information can aid in analyzing motion magnitude in videos. [61] further enhances this by combining flow edges with optical flow to track moving objects. Moreover, edge information has been utilized for shadow segmentation in videos [72], while [13] leverages edge information to derive discriminative representations, facilitating the differentiation of various semantic regions. However, these approaches either operate solely at the image level without distinguishing objects or are unable to generate fine-grained segmentation results for object tracking, rendering them unsuitable for semi-supervised video object segmentation. + +# 2.4. Uncertainty Estimation + +Uncertainty estimation involves predicting the level of confidence or uncertainty in a network's output, encompassing aleatoric uncertainty (arising from data noise) and epistemic uncertainty (stemming from model limitations) [73]. Early approaches employed methods like particle filtering and CRFs [74, 75]. With the rise of deep learning, Bayesian neural networks [76-78] were introduced, along with other techniques such as ensemble methods [79]. More recently, evidential learning [73, 80-82] has shown promising results, which has the potential to improve the high uncertainty of occlusion areas in VOS tasks. + +# 3. Methodology + +# 3.1. Inherent Structure-guided SVOS + +Problem Formulation Given an object mask for reference, existing memory-based SVOS methods typically learn to segment objects sequentially across frames while concurrently updating the model memory for future regards. This procedure can be expressed as: + +$$ +q _ {t} = h \left(\varphi \left(i _ {t}\right), y _ {t}\right), +$$ + +$$ +y _ {t} = \varrho \left(i _ {t - 1}, \dots , i _ {t - n} \mid q _ {t - 1}, \dots , q _ {t - n}\right), \tag {1} +$$ + +where $i_{t}$ denotes the input frame and $y_{t}$ is the constructed memory from previous $n$ frames and predictions, acting as reference at time $t$ . Functions $\varphi$ and $\varrho$ represent the feature extractor and memory constructor, respectively. The notation $h$ refers to the object segmentation network, which utilizes frame features $\varphi(i_{t})$ and memory $y$ to produce the segmentation output $q$ at time $t$ . This pipeline interprets the SVOS task from the perspective of the frame content. + +However, we observe that Eq. 1 lacks adequate modeling of structural information and hierarchy relationships. To enhance the capacity of SVOS and better address issues of occlusion and similarity-induced confusion, we introduce a novel structure refinement module to the existing framework. This module provides the model with inherent structural information and hierarchical relationships in each input frame. Fig. 2 shows the overall architecture of our proposed model. + +Specifically, given an input frame $i_{t}$ , we first derive its edge representation using the Canny filter $c$ , a robust algorithm known for accurately detecting edges despite the noise. The frame features $\varphi(i_{t})$ then interact with the edge representation to generate an inherent structure map $s$ under the guidance of the memory features, capturing key structural characteristics of target objects. This inherent structure map + +is subsequently integrated back into the frame features. Finally, we align memory features and structured enhanced features to predict the target object mask. Accordingly, our proposed solution can be formulated as: + +$$ +\begin{array}{l} q _ {t} = h (\varphi (i _ {t}), y _ {t}, s _ {t}), \\ s _ {t} = \vartheta (\varphi (i _ {t}), c (i _ {t}), y _ {t}), \\ \end{array} +$$ + +where $\vartheta$ denotes the fusion function for processing features and inherent structure map. By explicitly modeling structural information, our approach allows the model to better distinguish objects, even in challenging scenarios involving occlusion and similar semantics. + +# 3.2. Inherent Structure Map Generation + +Edge Highlighting Given an input RGB image $I \in \mathbb{R}^{3 \times H \times W}$ , we first derive its edge representation $I_{c} \in \mathbb{R}^{1 \times H \times W}$ using the Canny filter. Next, $I$ is fed into an image encoder for multi-scale feature extraction. The output of the image encoder can be formulated as $\hat{I}_{i} \in \mathbb{R}^{c_{i} \times h_{i} \times w_{i}}$ for $i = 1, \dots, N$ , where $N$ denotes the number of feature scales, $h_{i} = \frac{H}{2^{i+1}}$ and $w_{i} = \frac{W}{2^{i+1}}$ . + +Since the image encoder is pre-trained on ImageNet [83], its output feature map tends to emphasize abstract, content-focused information. In the case of SVOS, this nature may impede the model to accurately segment objects, potentially leading to confusion when occlusions occur among fragmented objects. To enhance the sensitivity of model to edges and structural details, we incorporate edge representation from $I_{c}$ into $\hat{I}_i$ through an element-wise multiplication, highlighting edge out of the feature map. This procedure can be formalized as: + +$$ +\tilde {I} _ {i} = \hat {I} _ {i} + (\hat {I} _ {i} \odot \epsilon I _ {c} ^ {\downarrow h _ {i}, w _ {i}}), \text {f o r} i = 1, \dots , N, \tag {3} +$$ + +where $\tilde{I}_i$ denotes the global-edge-enhanced feature map, $\odot$ represents the Hadamard product, $\epsilon$ is an importance factor that modulates the influence of the edge map detected by the Canny filter, and $I_c^{\downarrow h_i,w_i}$ indicates that $I_{c}$ has been interpolated to match the spatial dimensions of $\hat{I}_i$ . + +Memory Feature Extraction Given the $t$ -th frame of the video, we choose the previous $n$ frames and their corresponding predicted masks and feed them into the memory encoder to derive the global memory feature $N$ . Memory encoder is a common-used module in SVOS, where we adopt the design from Cutie [7]. We then project global features onto the object-based masks to derive object memory features $M \in \mathbb{R}^{C \times H \times W}$ (with $H = W = 30$ here), which provides temporal information about the target object for the next structure refinement module. + +Supervision Preparation Although the global-edge-enhanced feature map $\tilde{I}_i$ emphasizes edges, it still contains boundary details of non-target objects, which is insufficient for guiding the model toward precise segmentation. To address this limitation, we design a structure decoder to predict + +structure map $S$ , containing only edges and structure details of targets, from the global-edge-enhanced feature $\tilde{I}_i$ . For supervision, the ground-truth structure map is derived from ground-truth mask. Given a colored mask $G \in \mathbb{R}^{3 \times H \times W}$ corresponding to the image $I$ , we first convert it into a grayscale image $G_{gr} \in \mathbb{R}^{1 \times H \times W}$ . Subsequently, a predefined filter is applied to transform $G_{gr}$ into the final object-related inherent structure map $S$ . This generated object structure map effectively highlights occlusion relationships among target objects and is further used as ground truth in structural extraction network training. + +Object Structure Map Prediction After obtaining the global-edge-enhanced feature map $\tilde{I}_i$ and the object memory feature $M$ from the previous framework, a lightweight structure decoder with connecting paths is applied to predict structure map $S$ . Importantly, $\tilde{I}_i$ alone provides limited target-specific cues, and the network could only rely on global image features, which may compromise the accuracy of the target objects' structure map prediction. Therefore, the fusion of the object memory feature $M$ to incorporate object-specific information is necessary. This process can be formulated as: + +$$ +S ^ {\prime} = \text {S t r u c t u r e D e c o d e r} (\tilde {I} _ {i}, M), \text {f o r} S ^ {\prime} \rightarrow S, \tag {4} +$$ + +where $M$ is fused with the coarsest scale $\tilde{I}_3$ via elementwise addition. The decoder progressively upsamples the fused features with transposed convolutions, incorporating higher-resolution features via channel concatenation. To reduce overhead, binary cross-entropy loss is applied only at the final output. Visual analysis in Fig. 3 reveals that incorporating object memory features is crucial for accurately capturing the target objects' structure. Otherwise, the network lacks focus, leading to noisy predictions. + +# 3.3. Refinement and Decoding + +After obtaining the final object-level structure map, we apply the Hadamard product with the extracted image features to emphasize object boundary features, thus to enhance feature with structure details. This process can be expressed as follows: + +$$ +\hat {I} _ {i} ^ {\prime} = \hat {I} _ {i} + \left(\hat {I} _ {i} \odot \beta S ^ {\prime \downarrow h _ {i}, w _ {i}}\right), \text {f o r} i = 1, \dots , N, \tag {5} +$$ + +where $\hat{I}_i^{\prime}$ denotes the features ultimately used to segment the object mask and $\beta$ is an importance factor that modulates the influence of the structure map. Note that the structure map used in this process is in the form of logits rather than a binary mask, which helps to emphasize certain regions when the structure decoder has higher confidence in these areas being boundaries. + +With the structure-enhanced feature (SE feature) of the current frame, the global memory features $N$ , and the object memory features $M$ , we employ a decoder to align them and + +![](images/79d89e5b724fa447c2bad03bf2b2cb629dcd854b25770983682d2b297447434c.jpg) + +![](images/0e7f864a7a28a009994385af4845c0e47898747ce22836461bda526d0b4cd00c.jpg) +(c) +Figure 3. Output structure maps from different objects. (a) Input frame. (b) Input frame and predicted masks of the previous frame. (c) Output structure map without object fusion. (d) Output structure map with object fusion. "Feat." and "Mem." are abbreviations of "Feature" and "Memory" respectively. + +![](images/5b3f2b42190dabbd6d1df5a5182942455cfde4692d550ff28ac1a94b76a0c976.jpg) + +![](images/df4170e06d44dfe3c57c0461fd4532df36fbfec7a91926ee3fd582b9ad61091d.jpg) +(d) : Eureka! + +output the final object-based mask. The memory-based decoder design is inspired by Cutie [7], a pioneering approach to enhance VOS performance by object memory modeling. Concretely, the decoder integrates top-down object features with bottom-up global features so that a dynamic interaction between high-level and low-level representations is established. The SE feature, in this way, serves as an additional pathway for connecting high-level and low-level representations, as the structure encodes object information and latently emphasizes the original feature map accordingly. + +# 3.4. Enhancement via Uncertainty Quantification + +Due to the LogSoftmax used in the cross-entropy loss function, the predicted class probabilities remain point-based estimations that collapse multiple types of uncertainty into overconfident predictions [80, 84]. To address this, we further borrow the ideas of Evidential Deep Learning (EDL) to add an uncertainty-related loss term into our training pipeline, as being supplementary to existing loss combination, cross-entropy loss $\mathcal{L}_{\mathrm{CE}}$ and dice loss $\mathcal{L}_{\mathrm{Dice}}$ , to help improve performance by quantifying and minimizing the uncertainty of predictions, which is especially crucial in challenging scenarios involving occlusions and features with high similarity. + +The EDL based on the theory of subjective logic is used to calculate uncertainty with the Dirichlet distribution which acts as an updating measure of the model's intrinsic prediction bias and provides the reference for the subsequent $\mathrm{KL}_{\mathrm{div}}$ regularization. Before training, we initialize a bare Dirichlet $D(\mathbf{p}|\boldsymbol{\alpha})$ over the class probability vector $\mathbf{p} = [p_1,\dots ,p_k]$ and set all $\alpha = 1$ , which by Gibbs' inequality means the distribution of maximum uncertainty before seeing any data. + +During training, we update Dirichlet parameters with the model's output prob-logits to represent updated intrinsic bias. For each pixel at frame $t$ , we define $q_{t}^{k}$ as the Boolean ground-truth label of a certain pixel of $k$ -th class at time $t$ , and raw output logit $\dot{q}_{t}$ is $k$ -class prob-logit vector of the pixel output by the model. The evidential loss $\mathcal{L}_{\mathrm{EDL}}$ is calculated as: + +$$ +\mathcal {L} _ {\mathrm {E D L}} = \sum_ {k = 1} ^ {K} q _ {t} ^ {k} \cdot \left(\psi (\sum \alpha) - \psi (\alpha)\right) + \mathrm {K L} _ {\mathrm {d i v}}, \tag {6} +$$ + +$$ +\text {f o r} \alpha = 1 + \eta (\dot {q} _ {t}), +$$ + +where $\alpha$ means the Dirichlet parameter, $\eta$ is the confidence function and $\psi$ denotes the digamma function, and $\mathrm{KL}_{\mathrm{div}}$ is derived corresponding to [80]. Subsequently, the total loss for mask predictions $\mathcal{L}_{\mathrm{mask}}$ during training can be defined as: + +$$ +\mathcal {L} _ {\text {m a s k}} = \mathcal {L} _ {\mathrm {C E}} + \mathcal {L} _ {\mathrm {D i c e}} + \lambda \mathcal {L} _ {\mathrm {E D L}}, \tag {7} +$$ + +where $\lambda$ is the hyperparameter to balance the contribution of the uncertainty loss term. With this combination of loss functions, our model benefits from the discriminative power of $\mathcal{L}_{\mathrm{CE}}$ and $\mathcal{L}_{\mathrm{Dice}}$ while leveraging $\mathcal{L}_{\mathrm{EDL}}$ to address uncertain predictions, thus yielding improved segmentation performance. + +# 4. Experiments + +# 4.1. Implementation Details + +Network Configuration We employ two ResNet-18 models pretrained on ImageNet to extract image features and encode the mask, respectively. The dimension of the object feature is set to 256. For the structure decoder, we progressively upsample with each step employing a transposed convolution kernel of size 2 and a stride of 2 to achieve $2 \times$ upsampling. LeakyReLU is used as the activation function and residual connections are applied within this module. For the prediction layer, a $1 \times 1$ convolution is applied to output the binary segmentation. Due to space limitations, further details are in the supplementary materials. + +Training Pipeline Similar to previous work, we divide the training into two stages. First, OASIS is pre-trained on static image datasets [88-92] with a 3-frame synthesis following [7]. For these images, we apply mirroring, random affine transformations cut-and-paste as data augmentation and combine different images to create pseudo-video sequences. Default cropping size and total training iteration are set to $384 \times 384$ and $80\mathrm{K}$ , respectively. The batch size is set to 16, and point supervision is utilized with a total of 8192 sampled points. The learning rate is fixed at $1.0 \times 10^{-4}$ , and the max norm of gradient clipping is set to 3. Notably, due to noise introduced by pseudo-video sequences, the structure decoder and EDL loss component are not included at this stage for the purpose of robustness. + +
MethodPublicationMOSEDAV-17 ValDAV-17 TestYouTubeVOS-2019 Val
J&FJFJ&FJFJ&FJFGJsFsJuFu
STCN [37]NeurIPS'2152.548.556.685.482.288.676.172.779.682.781.185.478.285.9
AOT-R50 [4]NeurIPS'2158.454.362.684.982.387.579.675.983.385.383.988.879.988.5
RDE [85]CVPR'2246.842.451.384.280.887.577.473.681.281.981.185.576.284.8
XMem [3]ECCV'2256.352.160.686.282.989.581.077.484.585.584.388.680.388.6
LLB [86]AAAI'23---84.681.587.7---83.681.786.579.287.0
ISVOS [87]CVPR'23---87.183.790.582.879.386.286.185.289.780.788.9
DeAOT-R50 [5]NeurIPS'2259.054.663.485.282.288.280.776.984.585.684.289.280.288.8
DEVA [6]ICCV'2360.055.864.386.883.690.082.378.785.985.585.089.479.788.0
Cutie [7]CVPR'2461.257.265.386.883.889.782.879.286.586.285.389.680.989.0
OASIS-62.157.866.388.385.091.683.279.387.086.685.790.181.289.4
+ +Table 1. Overall performance comparison on VOS benchmarks. Numbers are taken from [7, 86, 87]. For the current state-of-the-art method, Cutie [7], we conduct training from scratch using its publicly available code and re-evaluate its performance on the benchmarks. Boldface and bold blue text indicate the best and second-best performance in each column, respectively. + +For the main training stage, the batch size is set to 16, and is conducted for a total of $125\mathrm{K}$ iterations. The base learning rate is initialized at $1.0 \times 10^{-4}$ , with the backbone learning rate scaled by a factor of 0.1. A step scheduler is employed to control the learning rate, reducing it by a factor of 0.1 at $100\mathrm{K}$ and $115\mathrm{K}$ iterations. The sequence length is set to 8, with each frame cropped to a size of $480 \times 480$ pixels. The EDL loss coefficient $\lambda$ is set to 0.01 with an additional annealing factor to gradually modify the impact of the loss. For the Canny filter, the lower and upper thresholds are set to 50 and 200, respectively, and an $L_{2}$ regularization term is applied to enhance accuracy. Parameters $\epsilon$ and $\beta$ are set to 0.5 and 1.0 for rough edge and structure map feature fusion, respectively. The number of sampled points for the loss calculation is increased to 12544. We implement the model using PyTorch and train it on four A5000 GPUs. + +# 4.2. Benchmark Datasets and Evaluation Metrics + +In our experiments, several commonly used SVOS datasets are used to verify OASIS, including DAVIS-2017 [2], YouTubeVOS 2019 [93], and MOSE [94]. Specifically, DAVIS-2017 consists of 60 densely annotated videos in the training set, with an additional 30 videos in both the validation and test-dev sets for evaluation. YouTubeVOS-2019 contains 3978 videos sourced from video classification datasets, divided into 3471 for training and 507 for validation. MOSE, known for its crowded environments, increased occlusions, and higher rates of object disappearance, comprises 1507 videos for training, with 311 videos in the validation set. Following previous studies, we conduct experiments using various training dataset settings and reported the results. + +In our experiments, three commonly used criteria, Jaccard index $\mathcal{I}$ , contour accuracy $\mathcal{F}$ and their average $\mathcal{I} \& \mathcal{F}$ are employed to quantify the performance of our proposed method. Among them, the Jaccard index assesses the area-based accuracy of predictions, whereas the contour accuracy measures the localization accuracy of object boundaries. Specifically, the YouTubeVOS evaluation provides detailed + +results for $\mathcal{I}$ and $\mathcal{F}$ on both seen and unseen objects, and $\mathcal{G}$ is the average of all objects, enabling a more comprehensive assessment of the generalization ability of a model. $\mathcal{I},\mathcal{F},$ $\mathcal{J}\& \mathcal{F}$ and $\mathcal{G}$ range from 0 to 1 and the reported numbers in our paper are scaled for better readability and formatting, following previous SVOS works [3, 6, 7]. A superior performance results in values closer to 100. Frames per second (FPS) is also employed in ablation study to measure the efficiency of a model. + +# 4.3. Overall Performance Comparison + +Our experiments compared OASIS with nine representative or state-of-the-art memory-based SVOS methods. Among them, STCN [37] and LLB [86] focus on achieving more accurate matching between query frames and memory banks. RDE [85], XMem [3], and DEVA [6] aims to reduce memory usage and improve efficiency, while AOT [4], DEAOT [5], ISVOS [87], and Cutie [7] emphasize enhanced object modeling. The analysis of experiments is as follows. Note that, the inference results of MOSE, YouTubeVOS, and DAVIS test-dev benchmarks are then packaged and uploaded to the evaluation server to obtain final results. + +# 4.3.1. Quantitative Experiments + +Following the setup in previous works [6, 7], we trained OASIS using a combination of the DAVIS-2017 and YouTube 2019 training sets and evaluated its performance across four validation sets. It can be seen from Tab. 1 that among all of these SVOS models, our OASIS consistently achieves the best segmentation performance across all benchmarks. Since these four standard benchmarks cover diverse video content and annotation conditions, i.e., different levels of granularity, achieving top performance on all of them is particularly challenging. These results, therefore, underscore the effectiveness and stability of OASIS. Specifically, for datasets with finer annotation granularity, such as the DAVIS-17 validation set, our model outperforms others by substantial margins of 1.2, 1.2, and 1.1 on the $\mathcal{J}\&\mathcal{F}, \mathcal{J}$ , + +![](images/856a67bfe2e8ff0da6ef5fd931e7a35ffcd2bdf0d5bf3cea149a362db4ffedd1.jpg) +Figure 4. Comparison of segmentation results on two video clips, with details zoomed in on the upper right corner. Notably, our model demonstrates superior resolving capability in occlusion regions and when the target exhibits high similarity to the background. + +
MethodMOSEDAV-17 Test
J&FJFJ&FJF
XMem [3]59.655.463.779.676.183.0
DeAOT-R50 [5]64.159.568.782.879.186.5
DEVA [6]66.061.870.382.678.986.4
Cutie [7]67.062.871.383.780.087.4
OASIS67.563.271.985.481.789.3
+ +Table 2. Performance comparison under Trained-with-MOSE setting on VOS benchmarks. Numbers are taken from [7]. Boldface and bold blue text indicate the best and second-best performance in each column, respectively. + +and $\mathcal{F}$ scores, respectively. Such improvement highlights the effectiveness of structure refinement in capturing object details. Moreover, OASIS remains stable performance gains on the MOSE benchmark, which contains more complex and densely packed objects that differ considerably from the training data. The contour accuracy $(\mathcal{F})$ improves significantly by 1.0, demonstrating that our approach enhances the structural integrity of segmentation, even in unseen and complex scenes. + +Tab. 2 presents additional results where the training set includes the MOSE benchmark, allowing for a more comprehensive performance evaluation. In this setting, our approach still demonstrates superior results, significantly outperforming other methods on the DAVIS-2017 test set by margins of 1.7, 1.7, and 1.9 on the $\mathcal{J}\& \mathcal{F},\mathcal{J},$ and $\mathcal{F}$ scores, respectively. These results indicate that incorporating more complex scenes in the training set allows our model to develop a stronger structural understanding, leading to improved segmentation accuracy. + +# 4.3.2. Qualitative Experiments + +We illustrate the visual results of our method and the SOTA one on two typical video clips from the YouTubeVOS 2019 validation set in Fig. 4. It can be observed that the existing SOTA method often struggles to distinguish between objects, particularly in cases of occlusion or high similarity between objects or between foreground objects and background. For instance, it fails to separate the fish hiding + +
MethodDAV-17 ValDAV-17 Test-dev
J&FJFJ&FJF
Baseline84.881.688.079.976.083.7
+CaE84.981.888.180.076.084.0
+CaE+SD85.182.088.180.376.384.5
+CaE+SD*85.482.288.680.876.884.9
OASIS85.682.488.781.076.985.1
+ +Table 3. Ablation experiments about key components on DAVIS-2017 validation set and test-dev set. + +behind the other fish (the 1 - 3 rows). In contrast, due to the proposed evidential learning, our OASIS is able to model the structure of the overlapped fish better and accomplishes precise segmentation in the first clip. Additionally, we observe that previous methods exhibit accumulating errors over time when segmenting the antennae of a snail (rows 4-6), due to their reliance on memory mechanisms. In contrast, our proposed structure refinement module corrects these inaccuracies, significantly reducing this issue. + +# 4.4. Ablation Study + +# 4.4.1. Ablation on Key Components + +OASIS involves three novel designs: the integration of rough edges detected by the Canny filter, structure decoder for refinement, and evidential learning loss for uncertainty quantification. We conduct ablation experiments to assess the individual contribution of each component, with an additional experiment on the object features fusion mechanism in the structure decoder. To this end, we build five experimental settings in total: $+\mathrm{CaE}$ refers to the baseline model with pixel feature integrated with the rough Canny edges as per Eq. 3, whereas it does not include the structure decoder and evidential learning. As shown in Fig. 3, $+\mathrm{CaE} + \mathrm{SD}$ means the structure decoder is further applied on top of $+\mathrm{CaE}$ for the refined structure map, while $+\mathrm{CaE} + \mathrm{SD}^*$ applied the object memory feature fusion mechanism. The final OASIS further adds uncertainty learning upon the $+\mathrm{CaE} + \mathrm{SD}^*$ . Due to limited resources, all models are trained on the DAVIS2017 training set with limited iterations and evaluated on the + +![](images/047783976fd1495b20cf80af29f20ef684d015ac58eaea956a9c9f532f590769.jpg) +Figure 5. Qualitative comparison from different settings in ablating key components. We show the results of four typical frames. Zoom in for a better view. + +DAVIS-2017 validation and test-dev sets. + +Tab. 3 shows the results of our ablation study. As observed on these benchmarks, Canny edge feature integration and the structure decoder are effective in characterizing the object and thus contribute to the overall segmentation performance. Object fusion further improves the capabilities of structure decoder, i.e., $+\mathrm{CaE} + \mathrm{SD^{*}}$ outperforms the one without object fusion largely especially on the test-dev set, with all indicators increased by about 0.5. Furthermore, the further introduction of uncertainty estimation and minimization allows our complete model to attain the best performance with negligible additional expense. + +# 4.4.2. Visual Comparison on Key Components + +Fig. 5 presents the visual results of different settings on representative frames. It can be observed that although $+\mathrm{CaE}$ achieves decent performance, structure refinement is still necessary to enhance output quality, especially for occlusions. For instance, at the position indicated by the white arrow (i.e., the body of the dog and the corner of the box), the implementation of the structure decoder significantly improves the structural details. This observation further validates the effectiveness of our design. + +# 4.4.3. Ablation on Factor $\epsilon$ and $\beta$ + +In Eq. 3 and Eq. 5 we use factors $\epsilon$ and $\beta$ to control the impact brought by the rough edges and the refined structure map, respectively. We separately investigate the effect of different values for these two factors on model performance, with results presented in Tab. 4. Notably, as the two factors increase from 0, accuracy consistently improves, substantiating the benefit of emphasizing boundary features. However, we observe that this performance gain reaches an upper limit, beyond which excessively large factors introduce a substantial distributional shift in the features, ultimately degrading performance. The best results are achieved when $\beta$ is greater than $\epsilon$ , further underscoring the importance of more detailed structural information. + +
βDAVIS-17 Val| |DAVIS-17 Val
J&FJF| |J&FJF
0.585.382.188.50.085.382.288.5
1.085.682.488.70.585.682.488.7
2.085.782.588.91.085.181.888.3
5.085.082.087.92.085.082.087.9
+ +Table 4. Ablation experiments about factors $\epsilon$ and $\beta$ on the DAVIS-2017 validation set. + +
Method#ParamsDAVIS-17 Val
J&FJFFPS
Baseline29M84.881.688.052
+2M Proj Net31M85.182.088.137
OASIS31M85.682.488.748
+ +Table 5. Experiments on the effect of including additional parameters. Results on the DAVIS-2017 validation set are shown. + +# 4.5. Performance Gain and Model Parameters + +In deep learning-based SVOS methods, performance gains may result from an increase in model parameters. In our approach, the primary source of additional parameters is the structure decoder, which introduces approximately 2M learnable parameters. Compared with the whole framework, these additional parameters are almost negligible. To further validate that the effectiveness of OASIS does not primarily stem from the increased number of parameters, we conduct additional analytical experiments. To provide a fair comparison, we add a convolutional projection network of approximately 2M parameters to the baseline as a comparative method, while maintaining other settings and hyperparameters. Results are shown in Tab. 5. Notably, the method with a 2M-parameter network does not achieve better results than OASIS. Combined with the ablative results in Tab. 3, which compare model performance with and without the structure decoder, we conclude that the proposed method is highly effective and, more importantly, that its performance gain is not due to the increased parameter count. + +# 5. Conclusion + +In this work, we propose a novel model named OASIS to address the challenge of occlusion in SVOS. OASIS first generates a structure map that captures the structural and layering information of predicted targets and leverages it to refine features, thus highlighting boundary regions. Additionally, we integrate evidential learning for uncertainty estimation to enhance the capability of OASIS to focus on complex and uncertain segmentation cases. Extensive experiments show the superior performance of OASIS while maintaining high computational efficiency. + +# Acknowledgments + +This work was supported by Ministry of Education Tier 2 grant, NUS, Singapore (T2EP20224-0028); and Ministry of Education Tier 1 grant, NUS, Singapore (24-1250-P0001). + +# References + +[1] S. Caelles, K.-K. Maninis, J. Pont-Tuset, L. Leal-Taixe, D. Cremers, and L. Van Gool, "One-shot video object segmentation," in Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 221-230. 1, 2 +[2] F. Perazzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, and A. Sorkine-Hornung, "A benchmark dataset and evaluation methodology for video object segmentation," in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 724-732. 1, 2, 6 +[3] H. K. Cheng and A. G. Schwing, "Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model," in European Conference on Computer Vision. Springer, 2022, pp. 640-658. 1, 2, 6, 7 +[4] Z. Yang, Y. Wei, and Y. Yang, "Associating objects with transformers for video object segmentation," Advances in Neural Information Processing Systems, vol. 34, pp. 2491-2502, 2021. 6 +[5] Z. Yang and Y. Yang, "Decoupling features in hierarchical propagation for video object segmentation," Advances in Neural Information Processing Systems, vol. 35, pp. 36324-36336, 2022. 6, 7 +[6] H. K. Cheng, S. W. Oh, B. Price, A. Schwing, and J.-Y. Lee, "Tracking anything with decoupled video segmentation," in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 1316-1326. 1, 6, 7 +[7] H. K. Cheng, S. W. Oh, B. Price, J.-Y. Lee, and A. Schwing, "Putting the object back into video object segmentation," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 3151-3161. 1, 2, 4, 5, 6, 7 +[8] Z. Huang, S. Yang, M. Zhou, Z. Gong, A. Abusorrah, C. Lin, and Z. Huang, "Making accurate object detection at the edge: Review and new approach," Artificial Intelligence Review, vol. 55, no. 3, pp. 2245-2274, 2022. 2 +[9] D. Yang, B. Peng, Z. Al-Huda, A. Malik, and D. Zhai, “An overview of edge and object contour detection,” Neurocomputing, vol. 488, pp. 470–493, 2022. +[10] C. L. Zitnick and P. Dollar, "Edge boxes: Locating object proposals from edges," in Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. Springer, 2014, pp. 391-405. 2 +[11] C. Kim and J.-N. Hwang, "Fast and automatic video object segmentation and tracking for content-based applications," IEEE transactions on circuits and systems for video technology, vol. 12, no. 2, pp. 122-129, 2002. 2, 3 +[12] S. Wan, S. Ding, and C. Chen, "Edge computing enabled video segmentation for real-time traffic monitoring in internet of vehicles," Pattern Recognition, vol. 121, p. 108146, 2022. 2, 3 + +[13] A. Khoreva, F. Galasso, M. Hein, and B. Schiele, "Classifier based graph construction for video segmentation," in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015, pp. 951-960. 2, 3 +[14] J. Canny, “A computational approach to edge detection,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. PAMI-8, no. 6, pp. 679–698, 1986. 2 +[15] J. Pont-Tuset, F. Perazzi, S. Caelles, P. Arbeláez, A. Sorkine-Hornung, and L. Van Gool, “The 2017 davis challenge on video object segmentation,” arXiv preprint arXiv:1704.00675, 2017. 2 +[16] P. Voigtlaender and B. Leibe, "Online adaptation of convolutional neural networks for video object segmentation," in Proceedings of the British Machine Vision Conference (BMVC). BMVA Press, September 2017, pp. 116.1-116.13. +[17] H. Xiao, J. Feng, G. Lin, Y. Liu, and M. Zhang, “Monet: Deep motion exploitation for video object segmentation,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 1140–1148. +[18] J. Luiten, P. Voigtlaender, and B. Leibe, “Premvos: Proposal generation, refinement and merging for video object segmentation,” in Asian conference on computer vision. Springer, 2018, pp. 565–580. +[19] Y.-T. Hu, J.-B. Huang, and A. Schwing, “Maskrnn: Instance level video object segmentation,” Advances in neural information processing systems, vol. 30, 2017. +[20] F. Perazzi, A. Khoreva, R. Benenson, B. Schiele, and A. Sorkine-Hornung, "Learning video object segmentation from static images," in Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 2663-2672. 2 +[21] K.-K. Maninis, S. Caelles, Y. Chen, J. Pont-Tuset, L. Leal-Taixe, D. Cremers, and L. Van Gool, "Video object segmentation without temporal information," IEEE transactions on pattern analysis and machine intelligence, vol. 41, no. 6, pp. 1515-1530, 2018. 2 +[22] J. Cheng, Y.-H. Tsai, S. Wang, and M.-H. Yang, "Segflow: Joint learning for video object segmentation and optical flow," in Proceedings of the IEEE international conference on computer vision, 2017, pp. 686-695. 2 +[23] X. Li and C. C. Loy, "Video object segmentation with joint re-identification and attention-aware mask propagation," in Proceedings of the European conference on computer vision (ECCV), 2018, pp. 90-105. +[24] S. W. Oh, J.-Y. Lee, K. Sunkavalli, and S. J. Kim, "Fast video object segmentation by reference-guided mask propagation," in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 7376-7385. +[25] Y.-H. Tsai, M.-H. Yang, and M. J. Black, "Video segmentation via object flow," in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 3899-3908. +[26] H. Xie, H. Yao, S. Zhou, S. Zhang, and W. Sun, "Efficient regional memory network for video object segmentation," in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 1286-1295. 2 +[27] L. Yang, Y. Wang, X. Xiong, J. Yang, and A. Katsaggelos, "Efficient video object segmentation via network modulation," + +in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 6499-6507. +[28] B. Duke, A. Ahmed, C. Wolf, P. Aarabi, and G. W. Taylor, "Sstvos: Sparse spatiotemporal transformers for video object segmentation," in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 5912-5921. +[29] K. Xu and A. Yao, "Accelerating video object segmentation with compressed video," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 1342-1351. +[30] Y.-S. Xu, T.-J. Fu, H.-K. Yang, and C.-Y. Lee, "Dynamic video segmentation network," in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 6556-6565. 2 +[31] Y. Chen, J. Pont-Tuset, A. Montes, and L. Van Gool, "Blazingly fast video object segmentation with pixel-wise metric learning," in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 1189-1198. 2 +[32] Y.-T. Hu, J.-B. Huang, and A. G. Schwing, "Videomatch: Matching based video object segmentation," in Proceedings of the European conference on computer vision (ECCV), 2018, pp. 54-70. +[33] Z. Yang, Y. Wei, and Y. Yang, "Collaborative video object segmentation by multi-scale foreground-background integration," IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 9, pp. 4701-4712, 2021. +[34] P. Voigtlaender, Y. Chai, F. Schroff, H. Adam, B. Leibe, and L.-C. Chen, “Feelvos: Fast end-to-end embedding learning for video object segmentation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 9481–9490. +[35] S. W. Oh, J.-Y. Lee, N. Xu, and S. J. Kim, "Video object segmentation using space-time memory networks," in Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 9226-9235. 2 +[36] H. K. Cheng, Y.-W. Tai, and C.-K. Tang, "Modular interactive video object segmentation: Interaction-to-mask, propagation and difference-aware fusion," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 5559-5568. +[37] ——, “Rethinking space-time networks with improved memory coverage for efficient video object segmentation,” Advances in Neural Information Processing Systems, vol. 34, pp. 11781-11794, 2021. 2, 6 +[38] Y. Li, Z. Shen, and Y. Shan, "Fast video object segmentation using the global context module," in Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16. Springer, 2020, pp. 735-750. +[39] Y. Liang, X. Li, N. Jafari, and J. Chen, "Video object segmentation with adaptive feature bank and uncertain-region refinement," Advances in Neural Information Processing Systems, vol. 33, pp. 3430-3441, 2020. 2 +[40] H. Wang, X. Jiang, H. Ren, Y. Hu, and S. Bai, "Swiftnet: Real-time video object segmentation," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 1296-1305. 2 + +[41] Y. Mao, N. Wang, W. Zhou, and H. Li, "Joint inductive and transductive learning for video object segmentation," in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 9670-9679. 2 +[42] H. Seong, J. Hyun, and E. Kim, "Kernelized memory network for video object segmentation," in Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. Springer, 2020, pp. 629-645. +[43] L. Hu, P. Zhang, B. Zhang, P. Pan, Y. Xu, and R. Jin, "Learning position and target consistency for memory-based video object segmentation," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 4144-4154. +[44] X. Lu, W. Wang, M. Danelljan, T. Zhou, J. Shen, and L. Van Gool, "Video object segmentation with episodic graph memory networks," in Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. Springer, 2020, pp. 661-679. +[45] H. Seong, S. W. Oh, J.-Y. Lee, S. Lee, S. Lee, and E. Kim, "Hierarchical memory matching network for video object segmentation," in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 12889-12898. +[46] H. Liu, M. Gao, X. Luo, Z. Wang, G. Qin, J. Wu, and Y. Jin, "Resurgsam2: Referring segment anything in surgical video via credible long-term tracking," in International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2025. +[47] J. Pei, J. Zhang, G. Qin, K. Wang, Y. Jin, and P.-A. Heng, "Instrument-tissue-guided surgical action triplet detection via textual-temporal trail exploration," IEEE Transactions on Medical Imaging, 2025. 2 +[48] N. Ravi, V. Gabeur, Y.-T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. Rädle, C. Rolland, L. Gustafson, E. Mintun, J. Pan, K. V. Alwala, N. Carion, C.-Y. Wu, R. Girshick, P. Dollar, and C. Feichtenhofer, “SAM 2: Segment anything in images and videos,” in The Thirteenth International Conference on Learning Representations, 2025. 2 +[49] Z. Wang, Y. Zhang, Y. Wang, L. Cai, and Y. Zhang, "Dynamic pseudo label optimization in point-supervised nuclei segmentation," in International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2024, pp. 220-230. 2 +[50] Y. Zhang, Z. Wang, Y. Wang, H. Bian, L. Cai, H. Li, L. Zhang, and Y. Zhang, “Boundary-aware contrastive learning for semi-supervised nuclei instance segmentation,” in Medical Imaging with Deep Learning, 2024. +[51] Y. Zhang, Y. Zhou, Y. Wang, J. Xiao, Z. Wang, Y. Zhang, and J. Chen, "The four color theorem for cell instance segmentation," in Forty-second International Conference on Machine Learning, 2025. 2 +[52] J.-X. Zhao, J.-J. Liu, D.-P. Fan, Y. Cao, J. Yang, and M.-M. Cheng, "Egnet: Edge guidance network for salient object detection," in Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 8779-8788. 2 +[53] Z. Luo, A. Mishra, A. Achkar, J. Eichel, S. Li, and P.-M. Jodoin, "Non-local deep features for salient object detection," + +in Proceedings of the IEEE Conference on computer vision and pattern recognition, 2017, pp. 6609-6617. +[54] Y. Zhuge, G. Yang, P. Zhang, and H. Lu, "Boundary-guided feature aggregation network for salient object detection," IEEE Signal Processing Letters, vol. 25, no. 12, pp. 1800-1804, 2018. +[55] Z. Tu, Y. Ma, C. Li, J. Tang, and B. Luo, "Edge-guided nonlocal fully convolutional network for salient object detection," IEEE transactions on circuits and systems for video technology, vol. 31, no. 2, pp. 582-593, 2020. +[56] X. Wang, H. Ma, X. Chen, and S. You, "Edge preserving and multi-scale contextual neural network for salient object detection," IEEE Transactions on Image Processing, vol. 27, no. 1, pp. 121-134, 2017. +[57] Z. Wu, L. Su, and Q. Huang, "Stacked cross refinement network for edge-aware salient object detection," in Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 7264-7273. 2 +[58] G. Li, Y. Xie, L. Lin, and Y. Yu, "Instance-level salient object segmentation," in Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 2386-2395. 2 +[59] Y. Hao, Y. Liu, Z. Wu, L. Han, Y. Chen, G. Chen, L. Chu, S. Tang, Z. Yu, Z. Chen et al., "Edgeflow: Achieving practical interactive segmentation with edge-guided flow," in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 1551-1560. +[60] H. Liu, R. A. R. Soto, F. Xiao, and Y. J. Lee, "Yolactedge: Real-time instance segmentation on the edge," in 2021 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2021, pp. 9579-9585. +[61] Y. Zhou, X. Xu, F. Shen, X. Zhu, and H. T. Shen, "Flow-edge guided unsupervised video object segmentation," IEEE Transactions on circuits and systems for video technology, vol. 32, no. 12, pp. 8116-8127, 2021. 2, 3 +[62] L. Liu, H. Li, and M. Gruteser, "Edge assisted real-time object detection for mobile augmented reality," in The 25th annual international conference on mobile computing and networking, 2019, pp. 1-16. +[63] C. Zhan, X. Duan, S. Xu, Z. Song, and M. Luo, "An improved moving object detection algorithm based on frame difference and edge detection," in Fourth international conference on image and graphics (ICIG 2007). IEEE, 2007, pp. 519-523. +[64] M. Hanyao, Y. Jin, Z. Qian, S. Zhang, and S. Lu, "Edge-assisted online on-device object detection for real-time video analytics," in IEEE INFOCOM 2021-IEEE Conference on Computer Communications. IEEE, 2021, pp. 1-10. 2 +[65] Z. Xu, X. Baojie, and W. Guoxin, "Canny edge detection based on open cv," in 2017 13th IEEE international conference on electronic measurement & instruments (ICEMI). IEEE, 2017, pp. 53-56. 2 +[66] L. Xuan and Z. Hong, "An improved canny edge detection algorithm," in 2017 8th IEEE international conference on software engineering and service science (ICSESS). IEEE, 2017, pp. 275-278. +[67] W. Rong, Z. Li, W. Zhang, and L. Sun, "An improved canny edge detection algorithm," in 2014 IEEE international con + +ference on mechatronics and automation. IEEE, 2014, pp. 577-582. 2 +[68] L. Ding and A. Goshtasby, “On the canny edge detector,” Pattern recognition, vol. 34, no. 3, pp. 721-725, 2001. 3 +[69] W. McIlhagga, “The canny edge detector revisited,” International Journal of Computer Vision, vol. 91, pp. 251–261, 2011. +[70] R. Song, Z. Zhang, and H. Liu, "Edge connection based canny edge detection algorithm," Pattern Recognition and Image Analysis, vol. 27, pp. 740-747, 2017. +[71] P. Bao, L. Zhang, and X. Wu, "Canny edge detection enhancement by scale multiplication," IEEE transactions on pattern analysis and machine intelligence, vol. 27, no. 9, pp. 1485-1490, 2005. 3 +[72] D. Xu, J. Liu, X. Li, Z. Liu, and X. Tang, "Insignificant shadow detection for video segmentation," IEEE Transactions on Circuits and Systems for Video Technology, vol. 15, no. 8, pp. 1058-1064, 2005. 3 +[73] S. Seoni, V. Jahmunah, M. Salvi, P. D. Barua, F. Molinari, and U. R. Acharya, “Application of uncertainty quantification to artificial intelligence in healthcare: A review of last decade (2013–2023)," Computers in Biology and Medicine, p. 107441, 2023. 3 +[74] A. Blake, R. Curwen, and A. Zisserman, “A framework for spatiotemporal control in the tracking of visual contours,” International Journal of Computer Vision, vol. 11, pp. 127–145, 1993. 3 +[75] X. He, R. S. Zemel, and M. A. Carreira-Perpinan, “Multiscale conditional random fields for image labeling,” in Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004., vol. 2. IEEE, 2004, pp. II-II. 3 +[76] D. Dera, N. C. Bouaynaya, G. Rasool, R. Shterenberg, and H. M. Fathallah-Shaykh, “Premium-cnn: Propagating uncertainty towards robust convolutional neural networks,” IEEE Transactions on Signal Processing, vol. 69, pp. 4669–4684, 2021. 3 +[77] Y. Gal and Z. Ghahramani, “Dropout as a bayesian approximation: Representing model uncertainty in deep learning,” in international conference on machine learning. PMLR, 2016, pp. 1050–1059. +[78] A. Kendall and Y. Gal, “What uncertainties do we need in bayesian deep learning for computer vision?” Advances in neural information processing systems, vol. 30, 2017. 3 +[79] A. Ashukha, A. Lyzhov, D. Molchanov, and D. Vetrov, “Pitfalls of in-domain uncertainty estimation and ensembling in deep learning,” in The Eighth International Conference on Learning Representations, 2020. 3 +[80] M. Sensoy, L. Kaplan, and M. Kandemir, "Evidential deep learning to quantify classification uncertainty," in Advances in Neural Information Processing Systems, S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, Eds., vol. 31. Curran Associates, Inc., 2018. 3, 5 +[81] K. Zou, Z. Chen, X. Yuan, X. Shen, M. Wang, and H. Fu, “A review of uncertainty estimation and its application in medical imaging,” Meta-Radiology, p. 100003, 2023. + +[82] W. Bao, Q. Yu, and Y. Kong, "Evidential deep learning for open set action recognition," in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 13349-13358. 3 +[83] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, "Imagenet: A large-scale hierarchical image database," in 2009 IEEE Conference on Computer Vision and Pattern Recognition, 2009, pp. 248-255. 4 +[84] H. Yang, C. Chen, Y. CHEN, , H. C. Yip, and D. QI, "Uncertainty estimation for safety-critical scene segmentation via fine-grained reward maximization," in Advances in Neural Information Processing Systems, A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, Eds., vol. 36. Curran Associates, Inc., 2023, pp. 36238-36249. 5 +[85] M. Li, L. Hu, Z. Xiong, B. Zhang, P. Pan, and D. Liu, "Recurrent dynamic embedding for video object segmentation," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 1332-1341. 6 +[86] M. Lan, J. Zhang, L. Zhang, and D. Tao, “Learning to learn better for video object segmentation,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 1, 2023, pp. 1205–1212. 6 +[87] J. Wang, D. Chen, Z. Wu, C. Luo, C. Tang, X. Dai, Y. Zhao, Y. Xie, L. Yuan, and Y.-G. Jiang, “Look before you match: Instance understanding matters in video object segmentation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 2268–2278. 6 +[88] J. Shi, Q. Yan, L. Xu, and J. Jia, "Hierarchical image saliency detection on extended cssd," IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 38, no. 4, pp. 717-729, 2016. 5 +[89] L. Wang, H. Lu, Y. Wang, M. Feng, D. Wang, B. Yin, and X. Ruan, “Learning to detect salient objects with image-level supervision,” in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 3796–3805. +[90] X. Li, T. Wei, Y. P. Chen, Y.-W. Tai, and C.-K. Tang, "Fss1000: A 1000-class dataset for few-shot segmentation," in 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 2866-2875. +[91] Y. Zeng, P. Zhang, Z. Lin, J. Zhang, and H. Lu, "Towards high-resolution salient object detection," in 2019 IEEE/CVF International Conference on Computer Vision (ICCV), 2019, pp. 7233-7242. +[92] H. K. Cheng, J. Chung, Y.-W. Tai, and C.-K. Tang, "Cascadepsp: Toward class-agnostic and very high-resolution segmentation via global and local refinement," in 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 8887-8896. 5 +[93] N. Xu, L. Yang, Y. Fan, D. Yue, Y. Liang, J. Yang, and T. Huang, "Youtube-vos: A large-scale video object segmentation benchmark," in ECCV, 2018. 6 +[94] H. Ding, C. Liu, S. He, X. Jiang, P. H. Torr, and S. Bai, “MOSE: A new dataset for video object segmentation in complex scenes,” in ICCV, 2023. 6 \ No newline at end of file diff --git a/2025/Structure Matters_ Revisiting Boundary Refinement in Video Object Segmentation/images.zip b/2025/Structure Matters_ Revisiting Boundary Refinement in Video Object Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2b9bfd0a79e47066d6bea32fdf3f76da534a5b9c --- /dev/null +++ b/2025/Structure Matters_ Revisiting Boundary Refinement in Video Object Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:822c4187528387f847a018c3114dec2cfee28d0aa7e171f8f6cdfc6ca07ea741 +size 522293 diff --git a/2025/Structure Matters_ Revisiting Boundary Refinement in Video Object Segmentation/layout.json b/2025/Structure Matters_ Revisiting Boundary Refinement in Video Object Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6fdb629bef4674ee3008fff8171fb998b157ff9f --- /dev/null +++ b/2025/Structure Matters_ Revisiting Boundary Refinement in Video Object Segmentation/layout.json @@ -0,0 +1,11140 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 55, + 103, + 555, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 103, + 555, + 121 + ], + "spans": [ + { + "bbox": [ + 55, + 103, + 555, + 121 + ], + "type": "text", + "content": "Structure Matters: Revisiting Boundary Refinement in Video Object Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 124, + 142, + 485, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 142, + 485, + 171 + ], + "spans": [ + { + "bbox": [ + 124, + 142, + 485, + 171 + ], + "type": "text", + "content": "Guanyi Qin" + }, + { + "bbox": [ + 124, + 142, + 485, + 171 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 124, + 142, + 485, + 171 + ], + "type": "text", + "content": ", Ziyue Wang" + }, + { + "bbox": [ + 124, + 142, + 485, + 171 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 124, + 142, + 485, + 171 + ], + "type": "text", + "content": ", Daiyun Shen" + }, + { + "bbox": [ + 124, + 142, + 485, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 124, + 142, + 485, + 171 + ], + "type": "text", + "content": ", Haofeng Liu" + }, + { + "bbox": [ + 124, + 142, + 485, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 124, + 142, + 485, + 171 + ], + "type": "text", + "content": ", Hantao Zhou" + }, + { + "bbox": [ + 124, + 142, + 485, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 124, + 142, + 485, + 171 + ], + "type": "text", + "content": ", Junde Wu" + }, + { + "bbox": [ + 124, + 142, + 485, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 124, + 142, + 485, + 171 + ], + "type": "text", + "content": ", Runze Hu" + }, + { + "bbox": [ + 124, + 142, + 485, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 124, + 142, + 485, + 171 + ], + "type": "text", + "content": ", Yueming Jin" + }, + { + "bbox": [ + 124, + 142, + 485, + 171 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 168, + 171, + 442, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 171, + 442, + 186 + ], + "spans": [ + { + "bbox": [ + 168, + 171, + 442, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 168, + 171, + 442, + 186 + ], + "type": "text", + "content": "National University of Singapore; " + }, + { + "bbox": [ + 168, + 171, + 442, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 168, + 171, + 442, + 186 + ], + "type": "text", + "content": "Tsinghua University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 170, + 186, + 440, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 186, + 440, + 200 + ], + "spans": [ + { + "bbox": [ + 170, + 186, + 440, + 200 + ], + "type": "text", + "content": "3University of Oxford; 4Beijing Institute of Technology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 201, + 544, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 201, + 544, + 213 + ], + "spans": [ + { + "bbox": [ + 63, + 201, + 544, + 213 + ], + "type": "text", + "content": "{guanyi.qin, e1374378, e1374467, haofeng.liu}@u.nus.edu, hrz@bit.edu.cn, ymjin@nus.edu.sg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "spans": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 266, + 297, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 266, + 297, + 578 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 297, + 578 + ], + "type": "text", + "content": "Given an object mask, Semi-supervised Video Object Segmentation (SVOS) technique aims to track and segment the object across video frames, serving as a fundamental task in computer vision. Although recent memory-based methods demonstrate potential, they often struggle with scenes involving occlusion, particularly in handling object interactions and high feature similarity. To address these issues and meet the real-time processing requirements of downstream applications, in this paper, we propose a novel bOundary Amendment video object Segmentation method with Inherent Structure refinement, hereby named OASIS. Specifically, a lightweight structure refinement module is proposed to enhance segmentation accuracy. With the fusion of rough edge priors captured by the Canny filter and stored object features, the module can generate an object-level structure map and refine the representations by highlighting boundary features. Evidential learning for uncertainty estimation is introduced to further address challenges in occluded regions. The proposed method, OASIS, maintains an efficient design, yet extensive experiments on challenging benchmarks demonstrate its superior performance and competitive inference speed compared to other state-of-the-art methods, i.e., achieving the " + }, + { + "bbox": [ + 54, + 266, + 297, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 54, + 266, + 297, + 578 + ], + "type": "text", + "content": " values of 91.6 (vs. 89.7 on DAVIS-17 validation set) and " + }, + { + "bbox": [ + 54, + 266, + 297, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 54, + 266, + 297, + 578 + ], + "type": "text", + "content": " values of 86.6 (vs. 86.2 on YouTubeVOS 2019 validation set) while maintaining a competitive speed of 48 FPS on DAVIS. Checkpoints, logs, and code will be available here." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 601, + 135, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 135, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 135, + 613 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 622, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 622, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 622, + 296, + 694 + ], + "type": "text", + "content": "Video Object Segmentation (VOS) is an important task in computer vision [1, 2] that aims to separate moving objects from the background. In the general semi-supervised VOS (SVOS) setting, the task involves propagating the object masks specified in the first frame throughout the entire video. By enabling effective identification and understanding of" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 342, + 249, + 431, + 297 + ], + "blocks": [ + { + "bbox": [ + 342, + 249, + 431, + 297 + ], + "lines": [ + { + "bbox": [ + 342, + 249, + 431, + 297 + ], + "spans": [ + { + "bbox": [ + 342, + 249, + 431, + 297 + ], + "type": "image", + "image_path": "5f514942aa9a78bd0b912b63ff997387c737f87c38383aff6b2844fd36e17277.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 434, + 285, + 447, + 295 + ], + "lines": [ + { + "bbox": [ + 434, + 285, + 447, + 295 + ], + "spans": [ + { + "bbox": [ + 434, + 285, + 447, + 295 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 450, + 252, + 518, + 293 + ], + "blocks": [ + { + "bbox": [ + 450, + 252, + 518, + 293 + ], + "lines": [ + { + "bbox": [ + 450, + 252, + 518, + 293 + ], + "spans": [ + { + "bbox": [ + 450, + 252, + 518, + 293 + ], + "type": "image", + "image_path": "168a0079287f9985ae3a47929dae1a363969b8888bdc53a09b7a362a302dbe27.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 348, + 320, + 422, + 361 + ], + "blocks": [ + { + "bbox": [ + 348, + 320, + 422, + 361 + ], + "lines": [ + { + "bbox": [ + 348, + 320, + 422, + 361 + ], + "spans": [ + { + "bbox": [ + 348, + 320, + 422, + 361 + ], + "type": "image", + "image_path": "5df03c903ecf7ffd778f058e2a40cad3a2ae4b68ef5bb4303fe15c59d355339b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 371, + 555, + 438 + ], + "lines": [ + { + "bbox": [ + 313, + 371, + 555, + 438 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 555, + 438 + ], + "type": "text", + "content": "Figure 1. (a) Input frame. (b) Object mask generated by the SOTA method Cutie. (c) Output structure map. (d) Object mask generated by our model. In this case, the fish corresponding to the red mask closely resembles the background and is concealed behind the other two fish, causing previous methods to fail. Guided by our structure map, the model successfully segments the fish accurately." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 450, + 321, + 518, + 361 + ], + "blocks": [ + { + "bbox": [ + 435, + 351, + 447, + 361 + ], + "lines": [ + { + "bbox": [ + 435, + 351, + 447, + 361 + ], + "spans": [ + { + "bbox": [ + 435, + 351, + 447, + 361 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 450, + 321, + 518, + 361 + ], + "lines": [ + { + "bbox": [ + 450, + 321, + 518, + 361 + ], + "spans": [ + { + "bbox": [ + 450, + 321, + 518, + 361 + ], + "type": "image", + "image_path": "21333945ab59170b581462df22157f2f5ef1aef0a6272e209bf93b052466d2d5.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 449, + 556, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 521 + ], + "type": "text", + "content": "objects in videos, VOS plays a crucial role in various applications including autonomous driving, robotics, video editing, medical intervention, etc. Albeit performance demands, real-time processing is also essential for these applications, necessitating algorithms that accurately perceive and segment objects while delivering sufficient inference speed." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "type": "text", + "content": "Recent advancements [3-6] in SVOS emphasize the value of memory-based mechanisms. Representations of certain past frames are generated and stored in a memory bank. The current frame is then used as the query to match these stored representations and derive segmentation masks. Cutie [7], further proposes object queries as a high-level summary of the target object and interacts with the pixel-level features of the query frame for accurate segmentation, achieving the state-of-the-art performance. However, these approaches still struggle when similar objects are close to each other or are occluded, as shown in Figure 1 (a) and (b). On one hand, the mask prediction depends predominantly on maintaining consistency across frames for both pixel- and object-level features, while the inherent geometric characteristics of the objects are often overlooked. This limitation poses challenges in accurately discerning object structures, particularly" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 216, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 216, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 216, + 713 + ], + "type": "text", + "content": "* Equal contribution, † Corresponding author." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14431" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "type": "text", + "content": "when they resemble the background (red box). On the other hand, when occlusion occurs, objects become fragmented. This challenge is amplified when objects share similar semantics, leading to high uncertainty in overlapping areas. Without an understanding of object structure, the model struggles to distinguish objects and establish clear boundaries (blue box)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 158, + 297, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 158, + 297, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 297, + 350 + ], + "type": "text", + "content": "Edge information is a crucial component of an object's structure representation, defining its boundaries and shape by representing the inherent structural information. Precise edge prediction allows models to capture structural details more effectively, thereby enhancing the model's ability to understand occlusions. Therefore, edges have been widely used for segmenting objects for static images across various fields including object detection and semantic segmentation [8-10]. However, in VOS tasks, target movements lead to considerable shifts in object edges over time, making it challenging to predict accurate edges from scratch for each frame. This often requires large edge detection networks, which significantly reduce inference efficiency. Therefore, edges are only used to analyze motion magnitude or shadow area detection primarily, which merely requires coarse edges [11-13]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 352, + 296, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 296, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 296, + 663 + ], + "type": "text", + "content": "To tackle the aforementioned problems, we propose an bOuntary Amendment video object Segmentation with Inherent Structure refinement, hereby named OASIS, which effectively refines object structures to handle geometric and layer relationships among objects better, while highlighting them out of backgrounds. This is achieved through a simple but effective structure refinement module. Specifically, we first extract the rough edge information by utilizing a fast and robust edge detection algorithm, Canny Edge Detection [14], which can identify global edges in individual frames and provide strong priors. With these priors, we only need a lightweight structure decoder to obtain the edges of the target objects instead of predicting from scratch without significant sacrifice in inference speed. This map further amplifies the original features by highlighting objects' boundaries, producing structure-enhanced features for final segmentation. This process enables the model to differentiate among objects and from the background effectively. Additionally, an object feature fusion mechanism is introduced to integrate memory features into the structure map to better model relationships between objects. Furthermore, to address the high uncertainty in dealing with overlapped areas, we also make the first attempt to deliver a new learning pattern that incorporates Evidential Learning (EDL) for uncertainty quantification and introduces an additional loss to minimize uncertainty and improve segmentation accuracy in occlusion regions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Extensive experiments demonstrate the superiority of our approach across four benchmarks under two training settings, particularly in scenes where target objects interact and exhibit complex relationships. Notably, OASIS achieves" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "type": "text", + "content": "promising results, consistently outperforming current state-of-the-art methods across different datasets, while delivering a real-time inference speed at 48 fps. In summary, our contributions are the following:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 121, + 555, + 300 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 313, + 121, + 555, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 555, + 156 + ], + "type": "text", + "content": "- We propose OASIS, a novel framework designed to manage relationships between the objects and the background by accurately modeling object structures." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 157, + 555, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 555, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 555, + 191 + ], + "type": "text", + "content": "- We introduce a structure refinement module and an object sensory mechanism that utilizes edge information to achieve inherent structure refinement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 193, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 193, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 553, + 251 + ], + "type": "text", + "content": "- To the best of our knowledge, we make the very first attempt to incorporate evidential learning in the SVOS task to improve the model performance on high-uncertainty regions, particularly in areas with significant similarity and occlusion." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 252, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 252, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 252, + 553, + 300 + ], + "type": "text", + "content": "- We verify OASIS on multiple SVOS benchmarks involving a wide range of video contents, annotation conditions, and training settings. OASIS outperforms other competitors across all these datasets and settings." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 312, + 400, + 325 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 400, + 325 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 400, + 325 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 332, + 465, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 465, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 465, + 346 + ], + "type": "text", + "content": "2.1. Video Object Segmentation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 350, + 556, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 556, + 543 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 556, + 543 + ], + "type": "text", + "content": "Early approaches of VOS [1, 2, 15-21] implement online learning that finetunes the model at test-time, resulting in low efficiency. Propagation-based methods [20, 22-30] iteratively propagate masks between adjacent frames by leveraging temporal correlations; however, they are inevitably constrained by drift and occlusion. Matching-based models [31-40] compute correspondences between the current and reference frames to capture long-range temporal context. Some memory-based methods [3, 26, 35-37, 39, 41-47] further enhance this by using memory banks to store features from past frames for pixel-level matching. The lately research SAM2 [48] also benefitted from memory mechanisms. Recently, Cutie [7] proposed an object-reading mechanism to enhance object-level representations. While it effectively perceives individual objects, modeling the relationships between multiple objects remains an open challenge." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 552, + 408, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 408, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 408, + 564 + ], + "type": "text", + "content": "2.2. Edge Detection" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "Edge information has been widely utilized to differentiate occluded objects [8-10, 49-51]. In salient object detection [52-57], edge features assist in identifying the most visually prominent objects within the image. Similarly, in segmentation tasks [11, 12, 58-61], edge detection helps to distinguish different regions and instances. Accurate localization of moving objects in videos remains a significant challenge, with edge information playing a crucial role in mitigating this difficulty. [11, 12, 61-64]. However, these methods require complex network designs like extra convolution layers, greatly affecting inference efficiency. Canny edge detection [14, 65-67] is a popular algorithm for identifying sharp in" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14432" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 347, + 257 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 347, + 257 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 347, + 257 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 347, + 257 + ], + "type": "image", + "image_path": "19a04cc1082cdfd93132d8a14f5e06fe7cc4dc8d74ee6d040b6bad68c241edb9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 258, + 555, + 282 + ], + "lines": [ + { + "bbox": [ + 55, + 258, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 555, + 282 + ], + "type": "text", + "content": "Figure 2. Overview of our proposed method: the left side shows the overall framework, while the right side provides a detailed illustration of the structure refinement process. The light grey box on the left indicates evidential learning, and \"Dist.\" is an abbreviation for \"Distribution.\"" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 349, + 70, + 541, + 255 + ], + "blocks": [ + { + "bbox": [ + 349, + 70, + 541, + 255 + ], + "lines": [ + { + "bbox": [ + 349, + 70, + 541, + 255 + ], + "spans": [ + { + "bbox": [ + 349, + 70, + 541, + 255 + ], + "type": "image", + "image_path": "47723baed477e527c48954574fe00432e1d8e375bcda66189bd7762e2f466e32.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 291, + 295, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 291, + 295, + 328 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 295, + 328 + ], + "type": "text", + "content": "tensity changes in images, which can help identify the edges of all regions, making it a valuable prior for enhancing edge prediction. [68-71]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 348, + 282, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 282, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 282, + 363 + ], + "type": "text", + "content": "2.3. Video Segmentation with Edge Information" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 370, + 295, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 370, + 295, + 539 + ], + "spans": [ + { + "bbox": [ + 54, + 370, + 295, + 539 + ], + "type": "text", + "content": "Edge information can effectively capture object movements, offering significant potential for advancing video object segmentation. Studies such as [11, 12] demonstrated that edge information can aid in analyzing motion magnitude in videos. [61] further enhances this by combining flow edges with optical flow to track moving objects. Moreover, edge information has been utilized for shadow segmentation in videos [72], while [13] leverages edge information to derive discriminative representations, facilitating the differentiation of various semantic regions. However, these approaches either operate solely at the image level without distinguishing objects or are unable to generate fine-grained segmentation results for object tracking, rendering them unsuitable for semi-supervised video object segmentation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 559, + 188, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 188, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 188, + 572 + ], + "type": "text", + "content": "2.4. Uncertainty Estimation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 297, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 712 + ], + "type": "text", + "content": "Uncertainty estimation involves predicting the level of confidence or uncertainty in a network's output, encompassing aleatoric uncertainty (arising from data noise) and epistemic uncertainty (stemming from model limitations) [73]. Early approaches employed methods like particle filtering and CRFs [74, 75]. With the rise of deep learning, Bayesian neural networks [76-78] were introduced, along with other techniques such as ensemble methods [79]. More recently, evidential learning [73, 80-82] has shown promising results, which has the potential to improve the high uncertainty of occlusion areas in VOS tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 289, + 398, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 398, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 398, + 304 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 310, + 492, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 310, + 492, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 492, + 323 + ], + "type": "text", + "content": "3.1. Inherent Structure-guided SVOS" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 328, + 555, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 555, + 389 + ], + "type": "text", + "content": "Problem Formulation Given an object mask for reference, existing memory-based SVOS methods typically learn to segment objects sequentially across frames while concurrently updating the model memory for future regards. This procedure can be expressed as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 396, + 399, + 471, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 399, + 471, + 412 + ], + "spans": [ + { + "bbox": [ + 396, + 399, + 471, + 412 + ], + "type": "interline_equation", + "content": "q _ {t} = h \\left(\\varphi \\left(i _ {t}\\right), y _ {t}\\right),", + "image_path": "d23cf9d07f74a1da10643626b975fb4e13ca4a9f1ee0c2bebe3fe7472d9e9872.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 355, + 408, + 554, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 408, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 355, + 408, + 554, + 426 + ], + "type": "interline_equation", + "content": "y _ {t} = \\varrho \\left(i _ {t - 1}, \\dots , i _ {t - n} \\mid q _ {t - 1}, \\dots , q _ {t - n}\\right), \\tag {1}", + "image_path": "a6d16c8be83a95b6f229c57c28c2c541bf7b70cb8daa923a2bf3239e612dd1b0.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "type": "inline_equation", + "content": "i_{t}" + }, + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "type": "text", + "content": " denotes the input frame and " + }, + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "type": "inline_equation", + "content": "y_{t}" + }, + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "type": "text", + "content": " is the constructed memory from previous " + }, + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "type": "text", + "content": " frames and predictions, acting as reference at time " + }, + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "type": "text", + "content": ". Functions " + }, + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "type": "inline_equation", + "content": "\\varrho" + }, + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "type": "text", + "content": " represent the feature extractor and memory constructor, respectively. The notation " + }, + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "type": "text", + "content": " refers to the object segmentation network, which utilizes frame features " + }, + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "type": "inline_equation", + "content": "\\varphi(i_{t})" + }, + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "type": "text", + "content": " and memory " + }, + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "type": "text", + "content": " to produce the segmentation output " + }, + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "type": "text", + "content": ". This pipeline interprets the SVOS task from the perspective of the frame content." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 533, + 556, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 629 + ], + "type": "text", + "content": "However, we observe that Eq. 1 lacks adequate modeling of structural information and hierarchy relationships. To enhance the capacity of SVOS and better address issues of occlusion and similarity-induced confusion, we introduce a novel structure refinement module to the existing framework. This module provides the model with inherent structural information and hierarchical relationships in each input frame. Fig. 2 shows the overall architecture of our proposed model." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Specifically, given an input frame " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "i_{t}" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": ", we first derive its edge representation using the Canny filter " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": ", a robust algorithm known for accurately detecting edges despite the noise. The frame features " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\varphi(i_{t})" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": " then interact with the edge representation to generate an inherent structure map " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": " under the guidance of the memory features, capturing key structural characteristics of target objects. This inherent structure map" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14433" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "content": "is subsequently integrated back into the frame features. Finally, we align memory features and structured enhanced features to predict the target object mask. Accordingly, our proposed solution can be formulated as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 127, + 128, + 225, + 154 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 128, + 225, + 154 + ], + "spans": [ + { + "bbox": [ + 127, + 128, + 225, + 154 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} q _ {t} = h (\\varphi (i _ {t}), y _ {t}, s _ {t}), \\\\ s _ {t} = \\vartheta (\\varphi (i _ {t}), c (i _ {t}), y _ {t}), \\\\ \\end{array}", + "image_path": "93b3ddb54f19f180397c25badc996ad7476b9a7f349415823cc581d64799cd13.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 160, + 296, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 160, + 296, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 160, + 296, + 220 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 160, + 296, + 220 + ], + "type": "inline_equation", + "content": "\\vartheta" + }, + { + "bbox": [ + 55, + 160, + 296, + 220 + ], + "type": "text", + "content": " denotes the fusion function for processing features and inherent structure map. By explicitly modeling structural information, our approach allows the model to better distinguish objects, even in challenging scenarios involving occlusion and similar semantics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 227, + 247, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 227, + 247, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 227, + 247, + 240 + ], + "type": "text", + "content": "3.2. Inherent Structure Map Generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 244, + 296, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 244, + 296, + 328 + ], + "spans": [ + { + "bbox": [ + 55, + 244, + 296, + 328 + ], + "type": "text", + "content": "Edge Highlighting Given an input RGB image " + }, + { + "bbox": [ + 55, + 244, + 296, + 328 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 55, + 244, + 296, + 328 + ], + "type": "text", + "content": ", we first derive its edge representation " + }, + { + "bbox": [ + 55, + 244, + 296, + 328 + ], + "type": "inline_equation", + "content": "I_{c} \\in \\mathbb{R}^{1 \\times H \\times W}" + }, + { + "bbox": [ + 55, + 244, + 296, + 328 + ], + "type": "text", + "content": " using the Canny filter. Next, " + }, + { + "bbox": [ + 55, + 244, + 296, + 328 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 244, + 296, + 328 + ], + "type": "text", + "content": " is fed into an image encoder for multi-scale feature extraction. The output of the image encoder can be formulated as " + }, + { + "bbox": [ + 55, + 244, + 296, + 328 + ], + "type": "inline_equation", + "content": "\\hat{I}_{i} \\in \\mathbb{R}^{c_{i} \\times h_{i} \\times w_{i}}" + }, + { + "bbox": [ + 55, + 244, + 296, + 328 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 244, + 296, + 328 + ], + "type": "inline_equation", + "content": "i = 1, \\dots, N" + }, + { + "bbox": [ + 55, + 244, + 296, + 328 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 244, + 296, + 328 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 244, + 296, + 328 + ], + "type": "text", + "content": " denotes the number of feature scales, " + }, + { + "bbox": [ + 55, + 244, + 296, + 328 + ], + "type": "inline_equation", + "content": "h_{i} = \\frac{H}{2^{i+1}}" + }, + { + "bbox": [ + 55, + 244, + 296, + 328 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 244, + 296, + 328 + ], + "type": "inline_equation", + "content": "w_{i} = \\frac{W}{2^{i+1}}" + }, + { + "bbox": [ + 55, + 244, + 296, + 328 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 328, + 296, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 328, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 54, + 328, + 296, + 448 + ], + "type": "text", + "content": "Since the image encoder is pre-trained on ImageNet [83], its output feature map tends to emphasize abstract, content-focused information. In the case of SVOS, this nature may impede the model to accurately segment objects, potentially leading to confusion when occlusions occur among fragmented objects. To enhance the sensitivity of model to edges and structural details, we incorporate edge representation from " + }, + { + "bbox": [ + 54, + 328, + 296, + 448 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 54, + 328, + 296, + 448 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 54, + 328, + 296, + 448 + ], + "type": "inline_equation", + "content": "\\hat{I}_i" + }, + { + "bbox": [ + 54, + 328, + 296, + 448 + ], + "type": "text", + "content": " through an element-wise multiplication, highlighting edge out of the feature map. This procedure can be formalized as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 86, + 453, + 296, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 453, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 86, + 453, + 296, + 468 + ], + "type": "interline_equation", + "content": "\\tilde {I} _ {i} = \\hat {I} _ {i} + (\\hat {I} _ {i} \\odot \\epsilon I _ {c} ^ {\\downarrow h _ {i}, w _ {i}}), \\text {f o r} i = 1, \\dots , N, \\tag {3}", + "image_path": "c0b29574654aa29244c474e9b2d3654d770d9ad8bc8a8aef64391c724c5bcc45.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 474, + 296, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 296, + 534 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 474, + 296, + 534 + ], + "type": "inline_equation", + "content": "\\tilde{I}_i" + }, + { + "bbox": [ + 55, + 474, + 296, + 534 + ], + "type": "text", + "content": " denotes the global-edge-enhanced feature map, " + }, + { + "bbox": [ + 55, + 474, + 296, + 534 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 474, + 296, + 534 + ], + "type": "text", + "content": " represents the Hadamard product, " + }, + { + "bbox": [ + 55, + 474, + 296, + 534 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 474, + 296, + 534 + ], + "type": "text", + "content": " is an importance factor that modulates the influence of the edge map detected by the Canny filter, and " + }, + { + "bbox": [ + 55, + 474, + 296, + 534 + ], + "type": "inline_equation", + "content": "I_c^{\\downarrow h_i,w_i}" + }, + { + "bbox": [ + 55, + 474, + 296, + 534 + ], + "type": "text", + "content": " indicates that " + }, + { + "bbox": [ + 55, + 474, + 296, + 534 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 55, + 474, + 296, + 534 + ], + "type": "text", + "content": " has been interpolated to match the spatial dimensions of " + }, + { + "bbox": [ + 55, + 474, + 296, + 534 + ], + "type": "inline_equation", + "content": "\\hat{I}_i" + }, + { + "bbox": [ + 55, + 474, + 296, + 534 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 534, + 296, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 652 + ], + "type": "text", + "content": "Memory Feature Extraction Given the " + }, + { + "bbox": [ + 55, + 534, + 296, + 652 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 534, + 296, + 652 + ], + "type": "text", + "content": "-th frame of the video, we choose the previous " + }, + { + "bbox": [ + 55, + 534, + 296, + 652 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 534, + 296, + 652 + ], + "type": "text", + "content": " frames and their corresponding predicted masks and feed them into the memory encoder to derive the global memory feature " + }, + { + "bbox": [ + 55, + 534, + 296, + 652 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 534, + 296, + 652 + ], + "type": "text", + "content": ". Memory encoder is a common-used module in SVOS, where we adopt the design from Cutie [7]. We then project global features onto the object-based masks to derive object memory features " + }, + { + "bbox": [ + 55, + 534, + 296, + 652 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 55, + 534, + 296, + 652 + ], + "type": "text", + "content": " (with " + }, + { + "bbox": [ + 55, + 534, + 296, + 652 + ], + "type": "inline_equation", + "content": "H = W = 30" + }, + { + "bbox": [ + 55, + 534, + 296, + 652 + ], + "type": "text", + "content": " here), which provides temporal information about the target object for the next structure refinement module." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "Supervision Preparation Although the global-edge-enhanced feature map " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{I}_i" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " emphasizes edges, it still contains boundary details of non-target objects, which is insufficient for guiding the model toward precise segmentation. To address this limitation, we design a structure decoder to predict" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "type": "text", + "content": "structure map " + }, + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "type": "text", + "content": ", containing only edges and structure details of targets, from the global-edge-enhanced feature " + }, + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "type": "inline_equation", + "content": "\\tilde{I}_i" + }, + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "type": "text", + "content": ". For supervision, the ground-truth structure map is derived from ground-truth mask. Given a colored mask " + }, + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "type": "inline_equation", + "content": "G \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "type": "text", + "content": " corresponding to the image " + }, + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "type": "text", + "content": ", we first convert it into a grayscale image " + }, + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "type": "inline_equation", + "content": "G_{gr} \\in \\mathbb{R}^{1 \\times H \\times W}" + }, + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "type": "text", + "content": ". Subsequently, a predefined filter is applied to transform " + }, + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "type": "inline_equation", + "content": "G_{gr}" + }, + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "type": "text", + "content": " into the final object-related inherent structure map " + }, + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "type": "text", + "content": ". This generated object structure map effectively highlights occlusion relationships among target objects and is further used as ground truth in structural extraction network training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 204, + 556, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 556, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 556, + 335 + ], + "type": "text", + "content": "Object Structure Map Prediction After obtaining the global-edge-enhanced feature map " + }, + { + "bbox": [ + 313, + 204, + 556, + 335 + ], + "type": "inline_equation", + "content": "\\tilde{I}_i" + }, + { + "bbox": [ + 313, + 204, + 556, + 335 + ], + "type": "text", + "content": " and the object memory feature " + }, + { + "bbox": [ + 313, + 204, + 556, + 335 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 204, + 556, + 335 + ], + "type": "text", + "content": " from the previous framework, a lightweight structure decoder with connecting paths is applied to predict structure map " + }, + { + "bbox": [ + 313, + 204, + 556, + 335 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 204, + 556, + 335 + ], + "type": "text", + "content": ". Importantly, " + }, + { + "bbox": [ + 313, + 204, + 556, + 335 + ], + "type": "inline_equation", + "content": "\\tilde{I}_i" + }, + { + "bbox": [ + 313, + 204, + 556, + 335 + ], + "type": "text", + "content": " alone provides limited target-specific cues, and the network could only rely on global image features, which may compromise the accuracy of the target objects' structure map prediction. Therefore, the fusion of the object memory feature " + }, + { + "bbox": [ + 313, + 204, + 556, + 335 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 204, + 556, + 335 + ], + "type": "text", + "content": " to incorporate object-specific information is necessary. This process can be formulated as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 342, + 344, + 555, + 358 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 344, + 555, + 358 + ], + "spans": [ + { + "bbox": [ + 342, + 344, + 555, + 358 + ], + "type": "interline_equation", + "content": "S ^ {\\prime} = \\text {S t r u c t u r e D e c o d e r} (\\tilde {I} _ {i}, M), \\text {f o r} S ^ {\\prime} \\rightarrow S, \\tag {4}", + "image_path": "9371c6a24c15e1b20430a281195d9e9de24094bb8a270121afa217176048de41.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 369, + 556, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 556, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 556, + 478 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 369, + 556, + 478 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 369, + 556, + 478 + ], + "type": "text", + "content": " is fused with the coarsest scale " + }, + { + "bbox": [ + 313, + 369, + 556, + 478 + ], + "type": "inline_equation", + "content": "\\tilde{I}_3" + }, + { + "bbox": [ + 313, + 369, + 556, + 478 + ], + "type": "text", + "content": " via elementwise addition. The decoder progressively upsamples the fused features with transposed convolutions, incorporating higher-resolution features via channel concatenation. To reduce overhead, binary cross-entropy loss is applied only at the final output. Visual analysis in Fig. 3 reveals that incorporating object memory features is crucial for accurately capturing the target objects' structure. Otherwise, the network lacks focus, leading to noisy predictions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 484, + 459, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 459, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 459, + 498 + ], + "type": "text", + "content": "3.3. Refinement and Decoding" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 502, + 554, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 554, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 554, + 561 + ], + "type": "text", + "content": "After obtaining the final object-level structure map, we apply the Hadamard product with the extracted image features to emphasize object boundary features, thus to enhance feature with structure details. This process can be expressed as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 340, + 570, + 555, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 570, + 555, + 585 + ], + "spans": [ + { + "bbox": [ + 340, + 570, + 555, + 585 + ], + "type": "interline_equation", + "content": "\\hat {I} _ {i} ^ {\\prime} = \\hat {I} _ {i} + \\left(\\hat {I} _ {i} \\odot \\beta S ^ {\\prime \\downarrow h _ {i}, w _ {i}}\\right), \\text {f o r} i = 1, \\dots , N, \\tag {5}", + "image_path": "50abdd6311945f0d8d73f2c25a8fd60a03b199c9e1d823929e65c7aab466a280.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 594, + 554, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 676 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 594, + 554, + 676 + ], + "type": "inline_equation", + "content": "\\hat{I}_i^{\\prime}" + }, + { + "bbox": [ + 313, + 594, + 554, + 676 + ], + "type": "text", + "content": " denotes the features ultimately used to segment the object mask and " + }, + { + "bbox": [ + 313, + 594, + 554, + 676 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 594, + 554, + 676 + ], + "type": "text", + "content": " is an importance factor that modulates the influence of the structure map. Note that the structure map used in this process is in the form of logits rather than a binary mask, which helps to emphasize certain regions when the structure decoder has higher confidence in these areas being boundaries." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "With the structure-enhanced feature (SE feature) of the current frame, the global memory features " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": ", and the object memory features " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": ", we employ a decoder to align them and" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14434" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 71, + 175, + 137 + ], + "blocks": [ + { + "bbox": [ + 69, + 71, + 175, + 137 + ], + "lines": [ + { + "bbox": [ + 69, + 71, + 175, + 137 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 175, + 137 + ], + "type": "image", + "image_path": "79d89e5b724fa447c2bad03bf2b2cb629dcd854b25770983682d2b297447434c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 69, + 138, + 173, + 243 + ], + "blocks": [ + { + "bbox": [ + 69, + 138, + 173, + 243 + ], + "lines": [ + { + "bbox": [ + 69, + 138, + 173, + 243 + ], + "spans": [ + { + "bbox": [ + 69, + 138, + 173, + 243 + ], + "type": "image", + "image_path": "0e7f864a7a28a009994385af4845c0e47898747ce22836461bda526d0b4cd00c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 243, + 135, + 263 + ], + "lines": [ + { + "bbox": [ + 69, + 243, + 135, + 263 + ], + "spans": [ + { + "bbox": [ + 69, + 243, + 135, + 263 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 266, + 297, + 322 + ], + "lines": [ + { + "bbox": [ + 55, + 266, + 297, + 322 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 297, + 322 + ], + "type": "text", + "content": "Figure 3. Output structure maps from different objects. (a) Input frame. (b) Input frame and predicted masks of the previous frame. (c) Output structure map without object fusion. (d) Output structure map with object fusion. \"Feat.\" and \"Mem.\" are abbreviations of \"Feature\" and \"Memory\" respectively." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 182, + 76, + 278, + 137 + ], + "blocks": [ + { + "bbox": [ + 182, + 76, + 278, + 137 + ], + "lines": [ + { + "bbox": [ + 182, + 76, + 278, + 137 + ], + "spans": [ + { + "bbox": [ + 182, + 76, + 278, + 137 + ], + "type": "image", + "image_path": "5b3f2b42190dabbd6d1df5a5182942455cfde4692d550ff28ac1a94b76a0c976.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 178, + 138, + 282, + 243 + ], + "blocks": [ + { + "bbox": [ + 178, + 138, + 282, + 243 + ], + "lines": [ + { + "bbox": [ + 178, + 138, + 282, + 243 + ], + "spans": [ + { + "bbox": [ + 178, + 138, + 282, + 243 + ], + "type": "image", + "image_path": "df4170e06d44dfe3c57c0461fd4532df36fbfec7a91926ee3fd582b9ad61091d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 243, + 258, + 262 + ], + "lines": [ + { + "bbox": [ + 178, + 243, + 258, + 262 + ], + "spans": [ + { + "bbox": [ + 178, + 243, + 258, + 262 + ], + "type": "text", + "content": "(d) : Eureka!" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 331, + 296, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 296, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 296, + 450 + ], + "type": "text", + "content": "output the final object-based mask. The memory-based decoder design is inspired by Cutie [7], a pioneering approach to enhance VOS performance by object memory modeling. Concretely, the decoder integrates top-down object features with bottom-up global features so that a dynamic interaction between high-level and low-level representations is established. The SE feature, in this way, serves as an additional pathway for connecting high-level and low-level representations, as the structure encodes object information and latently emphasizes the original feature map accordingly." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 456, + 289, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 289, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 289, + 470 + ], + "type": "text", + "content": "3.4. Enhancement via Uncertainty Quantification" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 474, + 296, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 296, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 296, + 617 + ], + "type": "text", + "content": "Due to the LogSoftmax used in the cross-entropy loss function, the predicted class probabilities remain point-based estimations that collapse multiple types of uncertainty into overconfident predictions [80, 84]. To address this, we further borrow the ideas of Evidential Deep Learning (EDL) to add an uncertainty-related loss term into our training pipeline, as being supplementary to existing loss combination, cross-entropy loss " + }, + { + "bbox": [ + 55, + 474, + 296, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}}" + }, + { + "bbox": [ + 55, + 474, + 296, + 617 + ], + "type": "text", + "content": " and dice loss " + }, + { + "bbox": [ + 55, + 474, + 296, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Dice}}" + }, + { + "bbox": [ + 55, + 474, + 296, + 617 + ], + "type": "text", + "content": ", to help improve performance by quantifying and minimizing the uncertainty of predictions, which is especially crucial in challenging scenarios involving occlusions and features with high similarity." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "text", + "content": "The EDL based on the theory of subjective logic is used to calculate uncertainty with the Dirichlet distribution which acts as an updating measure of the model's intrinsic prediction bias and provides the reference for the subsequent " + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{KL}_{\\mathrm{div}}" + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "text", + "content": " regularization. Before training, we initialize a bare Dirichlet " + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "inline_equation", + "content": "D(\\mathbf{p}|\\boldsymbol{\\alpha})" + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "text", + "content": " over the class probability vector " + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{p} = [p_1,\\dots ,p_k]" + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "text", + "content": " and set all " + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "text", + "content": ", which by Gibbs' inequality means the distribution of maximum uncertainty before seeing any data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": "During training, we update Dirichlet parameters with the model's output prob-logits to represent updated intrinsic bias. For each pixel at frame " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": ", we define " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "q_{t}^{k}" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": " as the Boolean ground-truth label of a certain pixel of " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": "-th class at time " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": ", and raw output logit " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "\\dot{q}_{t}" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": "-class prob-logit vector of the pixel output by the model. The evidential loss " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{EDL}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": " is calculated as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 337, + 162, + 555, + 195 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 162, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 337, + 162, + 555, + 195 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {E D L}} = \\sum_ {k = 1} ^ {K} q _ {t} ^ {k} \\cdot \\left(\\psi (\\sum \\alpha) - \\psi (\\alpha)\\right) + \\mathrm {K L} _ {\\mathrm {d i v}}, \\tag {6}", + "image_path": "9a6790560d05fc643d2b21ec20c3c673828ee19ac4f1c2ff763e76d0af77a4bc.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 394, + 198, + 473, + 210 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 198, + 473, + 210 + ], + "spans": [ + { + "bbox": [ + 394, + 198, + 473, + 210 + ], + "type": "interline_equation", + "content": "\\text {f o r} \\alpha = 1 + \\eta (\\dot {q} _ {t}),", + "image_path": "547833f3feabb5071fcb284f8e2bd88435a5d97251239d5f512867f7d669268f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 218, + 555, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 218, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 555, + 266 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 218, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 218, + 555, + 266 + ], + "type": "text", + "content": " means the Dirichlet parameter, " + }, + { + "bbox": [ + 313, + 218, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 218, + 555, + 266 + ], + "type": "text", + "content": " is the confidence function and " + }, + { + "bbox": [ + 313, + 218, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 313, + 218, + 555, + 266 + ], + "type": "text", + "content": " denotes the digamma function, and " + }, + { + "bbox": [ + 313, + 218, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\mathrm{KL}_{\\mathrm{div}}" + }, + { + "bbox": [ + 313, + 218, + 555, + 266 + ], + "type": "text", + "content": " is derived corresponding to [80]. Subsequently, the total loss for mask predictions " + }, + { + "bbox": [ + 313, + 218, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{mask}}" + }, + { + "bbox": [ + 313, + 218, + 555, + 266 + ], + "type": "text", + "content": " during training can be defined as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 369, + 274, + 555, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 274, + 555, + 287 + ], + "spans": [ + { + "bbox": [ + 369, + 274, + 555, + 287 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a s k}} = \\mathcal {L} _ {\\mathrm {C E}} + \\mathcal {L} _ {\\mathrm {D i c e}} + \\lambda \\mathcal {L} _ {\\mathrm {E D L}}, \\tag {7}", + "image_path": "2c26ceded8e32957ca8eba866e560768b52808ab2435e2b609f0662caab10cb2.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 295, + 556, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 556, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 556, + 366 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 295, + 556, + 366 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 295, + 556, + 366 + ], + "type": "text", + "content": " is the hyperparameter to balance the contribution of the uncertainty loss term. With this combination of loss functions, our model benefits from the discriminative power of " + }, + { + "bbox": [ + 313, + 295, + 556, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}}" + }, + { + "bbox": [ + 313, + 295, + 556, + 366 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 295, + 556, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Dice}}" + }, + { + "bbox": [ + 313, + 295, + 556, + 366 + ], + "type": "text", + "content": " while leveraging " + }, + { + "bbox": [ + 313, + 295, + 556, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{EDL}}" + }, + { + "bbox": [ + 313, + 295, + 556, + 366 + ], + "type": "text", + "content": " to address uncertain predictions, thus yielding improved segmentation performance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 377, + 395, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 395, + 391 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 395, + 391 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 397, + 447, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 447, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 447, + 409 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 415, + 556, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 556, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 556, + 545 + ], + "type": "text", + "content": "Network Configuration We employ two ResNet-18 models pretrained on ImageNet to extract image features and encode the mask, respectively. The dimension of the object feature is set to 256. For the structure decoder, we progressively upsample with each step employing a transposed convolution kernel of size 2 and a stride of 2 to achieve " + }, + { + "bbox": [ + 313, + 415, + 556, + 545 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 313, + 415, + 556, + 545 + ], + "type": "text", + "content": " upsampling. LeakyReLU is used as the activation function and residual connections are applied within this module. For the prediction layer, a " + }, + { + "bbox": [ + 313, + 415, + 556, + 545 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 313, + 415, + 556, + 545 + ], + "type": "text", + "content": " convolution is applied to output the binary segmentation. Due to space limitations, further details are in the supplementary materials." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "content": "Training Pipeline Similar to previous work, we divide the training into two stages. First, OASIS is pre-trained on static image datasets [88-92] with a 3-frame synthesis following [7]. For these images, we apply mirroring, random affine transformations cut-and-paste as data augmentation and combine different images to create pseudo-video sequences. Default cropping size and total training iteration are set to " + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "inline_equation", + "content": "384 \\times 384" + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "inline_equation", + "content": "80\\mathrm{K}" + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "content": ", respectively. The batch size is set to 16, and point supervision is utilized with a total of 8192 sampled points. The learning rate is fixed at " + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "inline_equation", + "content": "1.0 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "content": ", and the max norm of gradient clipping is set to 3. Notably, due to noise introduced by pseudo-video sequences, the structure decoder and EDL loss component are not included at this stage for the purpose of robustness." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14435" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 555, + 213 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 555, + 213 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 555, + 213 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 555, + 213 + ], + "type": "table", + "html": "
MethodPublicationMOSEDAV-17 ValDAV-17 TestYouTubeVOS-2019 Val
J&FJFJ&FJFJ&FJFGJsFsJuFu
STCN [37]NeurIPS'2152.548.556.685.482.288.676.172.779.682.781.185.478.285.9
AOT-R50 [4]NeurIPS'2158.454.362.684.982.387.579.675.983.385.383.988.879.988.5
RDE [85]CVPR'2246.842.451.384.280.887.577.473.681.281.981.185.576.284.8
XMem [3]ECCV'2256.352.160.686.282.989.581.077.484.585.584.388.680.388.6
LLB [86]AAAI'23---84.681.587.7---83.681.786.579.287.0
ISVOS [87]CVPR'23---87.183.790.582.879.386.286.185.289.780.788.9
DeAOT-R50 [5]NeurIPS'2259.054.663.485.282.288.280.776.984.585.684.289.280.288.8
DEVA [6]ICCV'2360.055.864.386.883.690.082.378.785.985.585.089.479.788.0
Cutie [7]CVPR'2461.257.265.386.883.889.782.879.286.586.285.389.680.989.0
OASIS-62.157.866.388.385.091.683.279.387.086.685.790.181.289.4
", + "image_path": "6e2be89cc45f4cbf8bc9ce19d32ef0d8ddd7637384a61cfdaac0f432ff5dc228.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 215, + 555, + 247 + ], + "lines": [ + { + "bbox": [ + 55, + 215, + 555, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 555, + 247 + ], + "type": "text", + "content": "Table 1. Overall performance comparison on VOS benchmarks. Numbers are taken from [7, 86, 87]. For the current state-of-the-art method, Cutie [7], we conduct training from scratch using its publicly available code and re-evaluate its performance on the benchmarks. Boldface and bold blue text indicate the best and second-best performance in each column, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 258, + 297, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 258, + 297, + 449 + ], + "spans": [ + { + "bbox": [ + 54, + 258, + 297, + 449 + ], + "type": "text", + "content": "For the main training stage, the batch size is set to 16, and is conducted for a total of " + }, + { + "bbox": [ + 54, + 258, + 297, + 449 + ], + "type": "inline_equation", + "content": "125\\mathrm{K}" + }, + { + "bbox": [ + 54, + 258, + 297, + 449 + ], + "type": "text", + "content": " iterations. The base learning rate is initialized at " + }, + { + "bbox": [ + 54, + 258, + 297, + 449 + ], + "type": "inline_equation", + "content": "1.0 \\times 10^{-4}" + }, + { + "bbox": [ + 54, + 258, + 297, + 449 + ], + "type": "text", + "content": ", with the backbone learning rate scaled by a factor of 0.1. A step scheduler is employed to control the learning rate, reducing it by a factor of 0.1 at " + }, + { + "bbox": [ + 54, + 258, + 297, + 449 + ], + "type": "inline_equation", + "content": "100\\mathrm{K}" + }, + { + "bbox": [ + 54, + 258, + 297, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 258, + 297, + 449 + ], + "type": "inline_equation", + "content": "115\\mathrm{K}" + }, + { + "bbox": [ + 54, + 258, + 297, + 449 + ], + "type": "text", + "content": " iterations. The sequence length is set to 8, with each frame cropped to a size of " + }, + { + "bbox": [ + 54, + 258, + 297, + 449 + ], + "type": "inline_equation", + "content": "480 \\times 480" + }, + { + "bbox": [ + 54, + 258, + 297, + 449 + ], + "type": "text", + "content": " pixels. The EDL loss coefficient " + }, + { + "bbox": [ + 54, + 258, + 297, + 449 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 54, + 258, + 297, + 449 + ], + "type": "text", + "content": " is set to 0.01 with an additional annealing factor to gradually modify the impact of the loss. For the Canny filter, the lower and upper thresholds are set to 50 and 200, respectively, and an " + }, + { + "bbox": [ + 54, + 258, + 297, + 449 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 54, + 258, + 297, + 449 + ], + "type": "text", + "content": " regularization term is applied to enhance accuracy. Parameters " + }, + { + "bbox": [ + 54, + 258, + 297, + 449 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 54, + 258, + 297, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 258, + 297, + 449 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 54, + 258, + 297, + 449 + ], + "type": "text", + "content": " are set to 0.5 and 1.0 for rough edge and structure map feature fusion, respectively. The number of sampled points for the loss calculation is increased to 12544. We implement the model using PyTorch and train it on four A5000 GPUs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 456, + 289, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 289, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 289, + 469 + ], + "type": "text", + "content": "4.2. Benchmark Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 474, + 296, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 474, + 296, + 629 + ], + "spans": [ + { + "bbox": [ + 56, + 474, + 296, + 629 + ], + "type": "text", + "content": "In our experiments, several commonly used SVOS datasets are used to verify OASIS, including DAVIS-2017 [2], YouTubeVOS 2019 [93], and MOSE [94]. Specifically, DAVIS-2017 consists of 60 densely annotated videos in the training set, with an additional 30 videos in both the validation and test-dev sets for evaluation. YouTubeVOS-2019 contains 3978 videos sourced from video classification datasets, divided into 3471 for training and 507 for validation. MOSE, known for its crowded environments, increased occlusions, and higher rates of object disappearance, comprises 1507 videos for training, with 311 videos in the validation set. Following previous studies, we conduct experiments using various training dataset settings and reported the results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 630, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 297, + 715 + ], + "type": "text", + "content": "In our experiments, three commonly used criteria, Jaccard index " + }, + { + "bbox": [ + 55, + 630, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 55, + 630, + 297, + 715 + ], + "type": "text", + "content": ", contour accuracy " + }, + { + "bbox": [ + 55, + 630, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 630, + 297, + 715 + ], + "type": "text", + "content": " and their average " + }, + { + "bbox": [ + 55, + 630, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{I} \\& \\mathcal{F}" + }, + { + "bbox": [ + 55, + 630, + 297, + 715 + ], + "type": "text", + "content": " are employed to quantify the performance of our proposed method. Among them, the Jaccard index assesses the area-based accuracy of predictions, whereas the contour accuracy measures the localization accuracy of object boundaries. Specifically, the YouTubeVOS evaluation provides detailed" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 258, + 556, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 556, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 556, + 368 + ], + "type": "text", + "content": "results for " + }, + { + "bbox": [ + 313, + 258, + 556, + 368 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 313, + 258, + 556, + 368 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 258, + 556, + 368 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 258, + 556, + 368 + ], + "type": "text", + "content": " on both seen and unseen objects, and " + }, + { + "bbox": [ + 313, + 258, + 556, + 368 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 258, + 556, + 368 + ], + "type": "text", + "content": " is the average of all objects, enabling a more comprehensive assessment of the generalization ability of a model. " + }, + { + "bbox": [ + 313, + 258, + 556, + 368 + ], + "type": "inline_equation", + "content": "\\mathcal{I},\\mathcal{F}," + }, + { + "bbox": [ + 313, + 258, + 556, + 368 + ], + "type": "inline_equation", + "content": "\\mathcal{J}\\& \\mathcal{F}" + }, + { + "bbox": [ + 313, + 258, + 556, + 368 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 258, + 556, + 368 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 258, + 556, + 368 + ], + "type": "text", + "content": " range from 0 to 1 and the reported numbers in our paper are scaled for better readability and formatting, following previous SVOS works [3, 6, 7]. A superior performance results in values closer to 100. Frames per second (FPS) is also employed in ablation study to measure the efficiency of a model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 374, + 497, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 497, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 497, + 387 + ], + "type": "text", + "content": "4.3. Overall Performance Comparison" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 392, + 556, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 556, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 556, + 523 + ], + "type": "text", + "content": "Our experiments compared OASIS with nine representative or state-of-the-art memory-based SVOS methods. Among them, STCN [37] and LLB [86] focus on achieving more accurate matching between query frames and memory banks. RDE [85], XMem [3], and DEVA [6] aims to reduce memory usage and improve efficiency, while AOT [4], DEAOT [5], ISVOS [87], and Cutie [7] emphasize enhanced object modeling. The analysis of experiments is as follows. Note that, the inference results of MOSE, YouTubeVOS, and DAVIS test-dev benchmarks are then packaged and uploaded to the evaluation server to obtain final results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 531, + 444, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 444, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 444, + 542 + ], + "type": "text", + "content": "4.3.1. Quantitative Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": "Following the setup in previous works [6, 7], we trained OASIS using a combination of the DAVIS-2017 and YouTube 2019 training sets and evaluated its performance across four validation sets. It can be seen from Tab. 1 that among all of these SVOS models, our OASIS consistently achieves the best segmentation performance across all benchmarks. Since these four standard benchmarks cover diverse video content and annotation conditions, i.e., different levels of granularity, achieving top performance on all of them is particularly challenging. These results, therefore, underscore the effectiveness and stability of OASIS. Specifically, for datasets with finer annotation granularity, such as the DAVIS-17 validation set, our model outperforms others by substantial margins of 1.2, 1.2, and 1.1 on the " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{J}\\&\\mathcal{F}, \\mathcal{J}" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "14436" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 71, + 526, + 199 + ], + "blocks": [ + { + "bbox": [ + 69, + 71, + 526, + 199 + ], + "lines": [ + { + "bbox": [ + 69, + 71, + 526, + 199 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 526, + 199 + ], + "type": "image", + "image_path": "856a67bfe2e8ff0da6ef5fd931e7a35ffcd2bdf0d5bf3cea149a362db4ffedd1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 201, + 555, + 224 + ], + "lines": [ + { + "bbox": [ + 55, + 201, + 555, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 555, + 224 + ], + "type": "text", + "content": "Figure 4. Comparison of segmentation results on two video clips, with details zoomed in on the upper right corner. Notably, our model demonstrates superior resolving capability in occlusion regions and when the target exhibits high similarity to the background." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 232, + 297, + 326 + ], + "blocks": [ + { + "bbox": [ + 58, + 232, + 297, + 326 + ], + "lines": [ + { + "bbox": [ + 58, + 232, + 297, + 326 + ], + "spans": [ + { + "bbox": [ + 58, + 232, + 297, + 326 + ], + "type": "table", + "html": "
MethodMOSEDAV-17 Test
J&FJFJ&FJF
XMem [3]59.655.463.779.676.183.0
DeAOT-R50 [5]64.159.568.782.879.186.5
DEVA [6]66.061.870.382.678.986.4
Cutie [7]67.062.871.383.780.087.4
OASIS67.563.271.985.481.789.3
", + "image_path": "9eceedb1594fd6eae4d6911a02ec8861b1d1f8372020a0ad3d0e960f12d84b02.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 327, + 296, + 371 + ], + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 371 + ], + "type": "text", + "content": "Table 2. Performance comparison under Trained-with-MOSE setting on VOS benchmarks. Numbers are taken from [7]. Boldface and bold blue text indicate the best and second-best performance in each column, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 380, + 295, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 380, + 295, + 488 + ], + "spans": [ + { + "bbox": [ + 54, + 380, + 295, + 488 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 54, + 380, + 295, + 488 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 54, + 380, + 295, + 488 + ], + "type": "text", + "content": " scores, respectively. Such improvement highlights the effectiveness of structure refinement in capturing object details. Moreover, OASIS remains stable performance gains on the MOSE benchmark, which contains more complex and densely packed objects that differ considerably from the training data. The contour accuracy " + }, + { + "bbox": [ + 54, + 380, + 295, + 488 + ], + "type": "inline_equation", + "content": "(\\mathcal{F})" + }, + { + "bbox": [ + 54, + 380, + 295, + 488 + ], + "type": "text", + "content": " improves significantly by 1.0, demonstrating that our approach enhances the structural integrity of segmentation, even in unseen and complex scenes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 488, + 295, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 488, + 295, + 609 + ], + "spans": [ + { + "bbox": [ + 54, + 488, + 295, + 609 + ], + "type": "text", + "content": "Tab. 2 presents additional results where the training set includes the MOSE benchmark, allowing for a more comprehensive performance evaluation. In this setting, our approach still demonstrates superior results, significantly outperforming other methods on the DAVIS-2017 test set by margins of 1.7, 1.7, and 1.9 on the " + }, + { + "bbox": [ + 54, + 488, + 295, + 609 + ], + "type": "inline_equation", + "content": "\\mathcal{J}\\& \\mathcal{F},\\mathcal{J}," + }, + { + "bbox": [ + 54, + 488, + 295, + 609 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 488, + 295, + 609 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 54, + 488, + 295, + 609 + ], + "type": "text", + "content": " scores, respectively. These results indicate that incorporating more complex scenes in the training set allows our model to develop a stronger structural understanding, leading to improved segmentation accuracy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 614, + 189, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 614, + 189, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 614, + 189, + 627 + ], + "type": "text", + "content": "4.3.2. Qualitative Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 629, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 629, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 629, + 297, + 715 + ], + "type": "text", + "content": "We illustrate the visual results of our method and the SOTA one on two typical video clips from the YouTubeVOS 2019 validation set in Fig. 4. It can be observed that the existing SOTA method often struggles to distinguish between objects, particularly in cases of occlusion or high similarity between objects or between foreground objects and background. For instance, it fails to separate the fish hiding" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 232, + 553, + 327 + ], + "blocks": [ + { + "bbox": [ + 317, + 232, + 553, + 327 + ], + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 327 + ], + "type": "table", + "html": "
MethodDAV-17 ValDAV-17 Test-dev
J&FJFJ&FJF
Baseline84.881.688.079.976.083.7
+CaE84.981.888.180.076.084.0
+CaE+SD85.182.088.180.376.384.5
+CaE+SD*85.482.288.680.876.884.9
OASIS85.682.488.781.076.985.1
", + "image_path": "c8c15a8b53f2474d3a1c43058ed2e80b4e258f26853f0f7f6a41ade099d8c680.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 327, + 555, + 349 + ], + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 349 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 349 + ], + "type": "text", + "content": "Table 3. Ablation experiments about key components on DAVIS-2017 validation set and test-dev set." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 360, + 556, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 556, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 556, + 469 + ], + "type": "text", + "content": "behind the other fish (the 1 - 3 rows). In contrast, due to the proposed evidential learning, our OASIS is able to model the structure of the overlapped fish better and accomplishes precise segmentation in the first clip. Additionally, we observe that previous methods exhibit accumulating errors over time when segmenting the antennae of a snail (rows 4-6), due to their reliance on memory mechanisms. In contrast, our proposed structure refinement module corrects these inaccuracies, significantly reducing this issue." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 476, + 409, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 409, + 490 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 409, + 490 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 495, + 470, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 470, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 470, + 507 + ], + "type": "text", + "content": "4.4.1. Ablation on Key Components" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": "OASIS involves three novel designs: the integration of rough edges detected by the Canny filter, structure decoder for refinement, and evidential learning loss for uncertainty quantification. We conduct ablation experiments to assess the individual contribution of each component, with an additional experiment on the object features fusion mechanism in the structure decoder. To this end, we build five experimental settings in total: " + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "inline_equation", + "content": "+\\mathrm{CaE}" + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": " refers to the baseline model with pixel feature integrated with the rough Canny edges as per Eq. 3, whereas it does not include the structure decoder and evidential learning. As shown in Fig. 3, " + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "inline_equation", + "content": "+\\mathrm{CaE} + \\mathrm{SD}" + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": " means the structure decoder is further applied on top of " + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "inline_equation", + "content": "+\\mathrm{CaE}" + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": " for the refined structure map, while " + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "inline_equation", + "content": "+\\mathrm{CaE} + \\mathrm{SD}^*" + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": " applied the object memory feature fusion mechanism. The final OASIS further adds uncertainty learning upon the " + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "inline_equation", + "content": "+\\mathrm{CaE} + \\mathrm{SD}^*" + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": ". Due to limited resources, all models are trained on the DAVIS2017 training set with limited iterations and evaluated on the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14437" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 282, + 201 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 282, + 201 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 282, + 201 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 282, + 201 + ], + "type": "image", + "image_path": "047783976fd1495b20cf80af29f20ef684d015ac58eaea956a9c9f532f590769.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 203, + 295, + 236 + ], + "lines": [ + { + "bbox": [ + 55, + 203, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 203, + 295, + 236 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison from different settings in ablating key components. We show the results of four typical frames. Zoom in for a better view." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 250, + 223, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 250, + 223, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 223, + 261 + ], + "type": "text", + "content": "DAVIS-2017 validation and test-dev sets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 263, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 263, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 296, + 396 + ], + "type": "text", + "content": "Tab. 3 shows the results of our ablation study. As observed on these benchmarks, Canny edge feature integration and the structure decoder are effective in characterizing the object and thus contribute to the overall segmentation performance. Object fusion further improves the capabilities of structure decoder, i.e., " + }, + { + "bbox": [ + 55, + 263, + 296, + 396 + ], + "type": "inline_equation", + "content": "+\\mathrm{CaE} + \\mathrm{SD^{*}}" + }, + { + "bbox": [ + 55, + 263, + 296, + 396 + ], + "type": "text", + "content": " outperforms the one without object fusion largely especially on the test-dev set, with all indicators increased by about 0.5. Furthermore, the further introduction of uncertainty estimation and minimization allows our complete model to attain the best performance with negligible additional expense." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 406, + 254, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 406, + 254, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 254, + 418 + ], + "type": "text", + "content": "4.4.2. Visual Comparison on Key Components" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 422, + 296, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 531 + ], + "type": "text", + "content": "Fig. 5 presents the visual results of different settings on representative frames. It can be observed that although " + }, + { + "bbox": [ + 55, + 422, + 296, + 531 + ], + "type": "inline_equation", + "content": "+\\mathrm{CaE}" + }, + { + "bbox": [ + 55, + 422, + 296, + 531 + ], + "type": "text", + "content": " achieves decent performance, structure refinement is still necessary to enhance output quality, especially for occlusions. For instance, at the position indicated by the white arrow (i.e., the body of the dog and the corner of the box), the implementation of the structure decoder significantly improves the structural details. This observation further validates the effectiveness of our design." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 541, + 198, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 541, + 198, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 541, + 198, + 553 + ], + "type": "text", + "content": "4.4.3. Ablation on Factor " + }, + { + "bbox": [ + 55, + 541, + 198, + 553 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 541, + 198, + 553 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 541, + 198, + 553 + ], + "type": "inline_equation", + "content": "\\beta" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 558, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 712 + ], + "type": "text", + "content": "In Eq. 3 and Eq. 5 we use factors " + }, + { + "bbox": [ + 55, + 558, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 558, + 296, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 558, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 558, + 296, + 712 + ], + "type": "text", + "content": " to control the impact brought by the rough edges and the refined structure map, respectively. We separately investigate the effect of different values for these two factors on model performance, with results presented in Tab. 4. Notably, as the two factors increase from 0, accuracy consistently improves, substantiating the benefit of emphasizing boundary features. However, we observe that this performance gain reaches an upper limit, beyond which excessively large factors introduce a substantial distributional shift in the features, ultimately degrading performance. The best results are achieved when " + }, + { + "bbox": [ + 55, + 558, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 558, + 296, + 712 + ], + "type": "text", + "content": " is greater than " + }, + { + "bbox": [ + 55, + 558, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 558, + 296, + 712 + ], + "type": "text", + "content": ", further underscoring the importance of more detailed structural information." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 553, + 157 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 553, + 157 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 157 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 157 + ], + "type": "table", + "html": "
βDAVIS-17 Val| |DAVIS-17 Val
J&FJF| |J&FJF
0.585.382.188.50.085.382.288.5
1.085.682.488.70.585.682.488.7
2.085.782.588.91.085.181.888.3
5.085.082.087.92.085.082.087.9
", + "image_path": "0afb2c7196bfc2882bafe36c163cd3049a7d5fc1cc443b07b165ddb733a8028e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 194, + 558, + 277 + ], + "blocks": [ + { + "bbox": [ + 313, + 157, + 555, + 179 + ], + "lines": [ + { + "bbox": [ + 313, + 157, + 555, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 555, + 179 + ], + "type": "text", + "content": "Table 4. Ablation experiments about factors " + }, + { + "bbox": [ + 313, + 157, + 555, + 179 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 157, + 555, + 179 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 157, + 555, + 179 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 157, + 555, + 179 + ], + "type": "text", + "content": " on the DAVIS-2017 validation set." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 194, + 558, + 277 + ], + "lines": [ + { + "bbox": [ + 315, + 194, + 558, + 277 + ], + "spans": [ + { + "bbox": [ + 315, + 194, + 558, + 277 + ], + "type": "table", + "html": "
Method#ParamsDAVIS-17 Val
J&FJFFPS
Baseline29M84.881.688.052
+2M Proj Net31M85.182.088.137
OASIS31M85.682.488.748
", + "image_path": "dac24788a9e2b46603f332372d30e5d8e914c9eac030da678de9deaeee169bfc.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 278, + 555, + 300 + ], + "lines": [ + { + "bbox": [ + 313, + 278, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 555, + 300 + ], + "type": "text", + "content": "Table 5. Experiments on the effect of including additional parameters. Results on the DAVIS-2017 validation set are shown." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 312, + 533, + 325 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 533, + 325 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 533, + 325 + ], + "type": "text", + "content": "4.5. Performance Gain and Model Parameters" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 331, + 556, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 556, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 556, + 559 + ], + "type": "text", + "content": "In deep learning-based SVOS methods, performance gains may result from an increase in model parameters. In our approach, the primary source of additional parameters is the structure decoder, which introduces approximately 2M learnable parameters. Compared with the whole framework, these additional parameters are almost negligible. To further validate that the effectiveness of OASIS does not primarily stem from the increased number of parameters, we conduct additional analytical experiments. To provide a fair comparison, we add a convolutional projection network of approximately 2M parameters to the baseline as a comparative method, while maintaining other settings and hyperparameters. Results are shown in Tab. 5. Notably, the method with a 2M-parameter network does not achieve better results than OASIS. Combined with the ablative results in Tab. 3, which compare model performance with and without the structure decoder, we conclude that the proposed method is highly effective and, more importantly, that its performance gain is not due to the increased parameter count." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 572, + 388, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 572, + 388, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 572, + 388, + 585 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "In this work, we propose a novel model named OASIS to address the challenge of occlusion in SVOS. OASIS first generates a structure map that captures the structural and layering information of predicted targets and leverages it to refine features, thus highlighting boundary regions. Additionally, we integrate evidential learning for uncertainty estimation to enhance the capability of OASIS to focus on complex and uncertain segmentation cases. Extensive experiments show the superior performance of OASIS while maintaining high computational efficiency." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14438" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 128 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 128 + ], + "type": "text", + "content": "This work was supported by Ministry of Education Tier 2 grant, NUS, Singapore (T2EP20224-0028); and Ministry of Education Tier 1 grant, NUS, Singapore (24-1250-P0001)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 157, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 157, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 157, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 157, + 296, + 201 + ], + "type": "text", + "content": "[1] S. Caelles, K.-K. Maninis, J. Pont-Tuset, L. Leal-Taixe, D. Cremers, and L. Van Gool, \"One-shot video object segmentation,\" in Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 221-230. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 202, + 296, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 296, + 256 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 296, + 256 + ], + "type": "text", + "content": "[2] F. Perazzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, and A. Sorkine-Hornung, \"A benchmark dataset and evaluation methodology for video object segmentation,\" in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 724-732. 1, 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 258, + 296, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 296, + 300 + ], + "type": "text", + "content": "[3] H. K. Cheng and A. G. Schwing, \"Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model,\" in European Conference on Computer Vision. Springer, 2022, pp. 640-658. 1, 2, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 302, + 296, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 296, + 344 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 296, + 344 + ], + "type": "text", + "content": "[4] Z. Yang, Y. Wei, and Y. Yang, \"Associating objects with transformers for video object segmentation,\" Advances in Neural Information Processing Systems, vol. 34, pp. 2491-2502, 2021. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 346, + 296, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 296, + 389 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 296, + 389 + ], + "type": "text", + "content": "[5] Z. Yang and Y. Yang, \"Decoupling features in hierarchical propagation for video object segmentation,\" Advances in Neural Information Processing Systems, vol. 35, pp. 36324-36336, 2022. 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 391, + 296, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 391, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 391, + 296, + 434 + ], + "type": "text", + "content": "[6] H. K. Cheng, S. W. Oh, B. Price, A. Schwing, and J.-Y. Lee, \"Tracking anything with decoupled video segmentation,\" in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 1316-1326. 1, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 436, + 296, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 296, + 489 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 296, + 489 + ], + "type": "text", + "content": "[7] H. K. Cheng, S. W. Oh, B. Price, J.-Y. Lee, and A. Schwing, \"Putting the object back into video object segmentation,\" in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 3151-3161. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 491, + 296, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 296, + 534 + ], + "type": "text", + "content": "[8] Z. Huang, S. Yang, M. Zhou, Z. Gong, A. Abusorrah, C. Lin, and Z. Huang, \"Making accurate object detection at the edge: Review and new approach,\" Artificial Intelligence Review, vol. 55, no. 3, pp. 2245-2274, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 536, + 296, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 296, + 568 + ], + "type": "text", + "content": "[9] D. Yang, B. Peng, Z. Al-Huda, A. Malik, and D. Zhai, “An overview of edge and object contour detection,” Neurocomputing, vol. 488, pp. 470–493, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 570, + 296, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 570, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 57, + 570, + 296, + 622 + ], + "type": "text", + "content": "[10] C. L. Zitnick and P. Dollar, \"Edge boxes: Locating object proposals from edges,\" in Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. Springer, 2014, pp. 391-405. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 624, + 296, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 296, + 668 + ], + "type": "text", + "content": "[11] C. Kim and J.-N. Hwang, \"Fast and automatic video object segmentation and tracking for content-based applications,\" IEEE transactions on circuits and systems for video technology, vol. 12, no. 2, pp. 122-129, 2002. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "type": "text", + "content": "[12] S. Wan, S. Ding, and C. Chen, \"Edge computing enabled video segmentation for real-time traffic monitoring in internet of vehicles,\" Pattern Recognition, vol. 121, p. 108146, 2022. 2, 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "text", + "content": "[13] A. Khoreva, F. Galasso, M. Hein, and B. Schiele, \"Classifier based graph construction for video segmentation,\" in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015, pp. 951-960. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 118, + 555, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 555, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 555, + 149 + ], + "type": "text", + "content": "[14] J. Canny, “A computational approach to edge detection,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. PAMI-8, no. 6, pp. 679–698, 1986. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 151, + 555, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 555, + 193 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 555, + 193 + ], + "type": "text", + "content": "[15] J. Pont-Tuset, F. Perazzi, S. Caelles, P. Arbeláez, A. Sorkine-Hornung, and L. Van Gool, “The 2017 davis challenge on video object segmentation,” arXiv preprint arXiv:1704.00675, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 195, + 555, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 555, + 238 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 555, + 238 + ], + "type": "text", + "content": "[16] P. Voigtlaender and B. Leibe, \"Online adaptation of convolutional neural networks for video object segmentation,\" in Proceedings of the British Machine Vision Conference (BMVC). BMVA Press, September 2017, pp. 116.1-116.13." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 239, + 555, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 555, + 282 + ], + "type": "text", + "content": "[17] H. Xiao, J. Feng, G. Lin, Y. Liu, and M. Zhang, “Monet: Deep motion exploitation for video object segmentation,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 1140–1148." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 284, + 555, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 555, + 327 + ], + "type": "text", + "content": "[18] J. Luiten, P. Voigtlaender, and B. Leibe, “Premvos: Proposal generation, refinement and merging for video object segmentation,” in Asian conference on computer vision. Springer, 2018, pp. 565–580." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 328, + 555, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 328, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 328, + 555, + 360 + ], + "type": "text", + "content": "[19] Y.-T. Hu, J.-B. Huang, and A. Schwing, “Maskrnn: Instance level video object segmentation,” Advances in neural information processing systems, vol. 30, 2017." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 361, + 555, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 555, + 414 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 555, + 414 + ], + "type": "text", + "content": "[20] F. Perazzi, A. Khoreva, R. Benenson, B. Schiele, and A. Sorkine-Hornung, \"Learning video object segmentation from static images,\" in Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 2663-2672. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 416, + 555, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 555, + 469 + ], + "type": "text", + "content": "[21] K.-K. Maninis, S. Caelles, Y. Chen, J. Pont-Tuset, L. Leal-Taixe, D. Cremers, and L. Van Gool, \"Video object segmentation without temporal information,\" IEEE transactions on pattern analysis and machine intelligence, vol. 41, no. 6, pp. 1515-1530, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 472, + 555, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 555, + 514 + ], + "type": "text", + "content": "[22] J. Cheng, Y.-H. Tsai, S. Wang, and M.-H. Yang, \"Segflow: Joint learning for video object segmentation and optical flow,\" in Proceedings of the IEEE international conference on computer vision, 2017, pp. 686-695. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 516, + 554, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 516, + 554, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 516, + 554, + 558 + ], + "type": "text", + "content": "[23] X. Li and C. C. Loy, \"Video object segmentation with joint re-identification and attention-aware mask propagation,\" in Proceedings of the European conference on computer vision (ECCV), 2018, pp. 90-105." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 559, + 555, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 555, + 602 + ], + "type": "text", + "content": "[24] S. W. Oh, J.-Y. Lee, K. Sunkavalli, and S. J. Kim, \"Fast video object segmentation by reference-guided mask propagation,\" in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 7376-7385." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 604, + 555, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 555, + 645 + ], + "type": "text", + "content": "[25] Y.-H. Tsai, M.-H. Yang, and M. J. Black, \"Video segmentation via object flow,\" in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 3899-3908." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 647, + 555, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 555, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 555, + 690 + ], + "type": "text", + "content": "[26] H. Xie, H. Yao, S. Zhou, S. Zhang, and W. Sun, \"Efficient regional memory network for video object segmentation,\" in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 1286-1295. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "type": "text", + "content": "[27] L. Yang, Y. Wang, X. Xiong, J. Yang, and A. Katsaggelos, \"Efficient video object segmentation via network modulation,\"" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14439" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 6499-6507." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 297, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 297, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 297, + 148 + ], + "type": "text", + "content": "[28] B. Duke, A. Ahmed, C. Wolf, P. Aarabi, and G. W. Taylor, \"Sstvos: Sparse spatiotemporal transformers for video object segmentation,\" in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 5912-5921." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 296, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 296, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 296, + 194 + ], + "type": "text", + "content": "[29] K. Xu and A. Yao, \"Accelerating video object segmentation with compressed video,\" in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 1342-1351." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 296, + 239 + ], + "type": "text", + "content": "[30] Y.-S. Xu, T.-J. Fu, H.-K. Yang, and C.-Y. Lee, \"Dynamic video segmentation network,\" in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 6556-6565. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 296, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 296, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 296, + 282 + ], + "type": "text", + "content": "[31] Y. Chen, J. Pont-Tuset, A. Montes, and L. Van Gool, \"Blazingly fast video object segmentation with pixel-wise metric learning,\" in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 1189-1198. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 284, + 296, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 284, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 296, + 327 + ], + "type": "text", + "content": "[32] Y.-T. Hu, J.-B. Huang, and A. G. Schwing, \"Videomatch: Matching based video object segmentation,\" in Proceedings of the European conference on computer vision (ECCV), 2018, pp. 54-70." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 296, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 296, + 371 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 296, + 371 + ], + "type": "text", + "content": "[33] Z. Yang, Y. Wei, and Y. Yang, \"Collaborative video object segmentation by multi-scale foreground-background integration,\" IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 9, pp. 4701-4712, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 372, + 296, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 296, + 426 + ], + "type": "text", + "content": "[34] P. Voigtlaender, Y. Chai, F. Schroff, H. Adam, B. Leibe, and L.-C. Chen, “Feelvos: Fast end-to-end embedding learning for video object segmentation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 9481–9490." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 426, + 296, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 296, + 470 + ], + "type": "text", + "content": "[35] S. W. Oh, J.-Y. Lee, N. Xu, and S. J. Kim, \"Video object segmentation using space-time memory networks,\" in Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 9226-9235. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 471, + 296, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 296, + 525 + ], + "type": "text", + "content": "[36] H. K. Cheng, Y.-W. Tai, and C.-K. Tang, \"Modular interactive video object segmentation: Interaction-to-mask, propagation and difference-aware fusion,\" in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 5559-5568." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 526, + 296, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 296, + 568 + ], + "type": "text", + "content": "[37] ——, “Rethinking space-time networks with improved memory coverage for efficient video object segmentation,” Advances in Neural Information Processing Systems, vol. 34, pp. 11781-11794, 2021. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 570, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 296, + 624 + ], + "type": "text", + "content": "[38] Y. Li, Z. Shen, and Y. Shan, \"Fast video object segmentation using the global context module,\" in Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16. Springer, 2020, pp. 735-750." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 625, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 296, + 669 + ], + "type": "text", + "content": "[39] Y. Liang, X. Li, N. Jafari, and J. Chen, \"Video object segmentation with adaptive feature bank and uncertain-region refinement,\" Advances in Neural Information Processing Systems, vol. 33, pp. 3430-3441, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "text", + "content": "[40] H. Wang, X. Jiang, H. Ren, Y. Hu, and S. Bai, \"Swiftnet: Real-time video object segmentation,\" in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 1296-1305. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 117 + ], + "type": "text", + "content": "[41] Y. Mao, N. Wang, W. Zhou, and H. Li, \"Joint inductive and transductive learning for video object segmentation,\" in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 9670-9679. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 555, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 555, + 172 + ], + "type": "text", + "content": "[42] H. Seong, J. Hyun, and E. Kim, \"Kernelized memory network for video object segmentation,\" in Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. Springer, 2020, pp. 629-645." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 175, + 555, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 555, + 229 + ], + "type": "text", + "content": "[43] L. Hu, P. Zhang, B. Zhang, P. Pan, Y. Xu, and R. Jin, \"Learning position and target consistency for memory-based video object segmentation,\" in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 4144-4154." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 231, + 555, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 555, + 285 + ], + "type": "text", + "content": "[44] X. Lu, W. Wang, M. Danelljan, T. Zhou, J. Shen, and L. Van Gool, \"Video object segmentation with episodic graph memory networks,\" in Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. Springer, 2020, pp. 661-679." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 286, + 555, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 555, + 330 + ], + "type": "text", + "content": "[45] H. Seong, S. W. Oh, J.-Y. Lee, S. Lee, S. Lee, and E. Kim, \"Hierarchical memory matching network for video object segmentation,\" in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 12889-12898." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 332, + 555, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 555, + 386 + ], + "type": "text", + "content": "[46] H. Liu, M. Gao, X. Luo, Z. Wang, G. Qin, J. Wu, and Y. Jin, \"Resurgsam2: Referring segment anything in surgical video via credible long-term tracking,\" in International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2025." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 388, + 555, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 555, + 431 + ], + "type": "text", + "content": "[47] J. Pei, J. Zhang, G. Qin, K. Wang, Y. Jin, and P.-A. Heng, \"Instrument-tissue-guided surgical action triplet detection via textual-temporal trail exploration,\" IEEE Transactions on Medical Imaging, 2025. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 434, + 555, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 555, + 498 + ], + "type": "text", + "content": "[48] N. Ravi, V. Gabeur, Y.-T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. Rädle, C. Rolland, L. Gustafson, E. Mintun, J. Pan, K. V. Alwala, N. Carion, C.-Y. Wu, R. Girshick, P. Dollar, and C. Feichtenhofer, “SAM 2: Segment anything in images and videos,” in The Thirteenth International Conference on Learning Representations, 2025. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 500, + 555, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 555, + 555 + ], + "type": "text", + "content": "[49] Z. Wang, Y. Zhang, Y. Wang, L. Cai, and Y. Zhang, \"Dynamic pseudo label optimization in point-supervised nuclei segmentation,\" in International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2024, pp. 220-230. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 556, + 555, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 556, + 555, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 556, + 555, + 600 + ], + "type": "text", + "content": "[50] Y. Zhang, Z. Wang, Y. Wang, H. Bian, L. Cai, H. Li, L. Zhang, and Y. Zhang, “Boundary-aware contrastive learning for semi-supervised nuclei instance segmentation,” in Medical Imaging with Deep Learning, 2024." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 601, + 555, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 555, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 555, + 644 + ], + "type": "text", + "content": "[51] Y. Zhang, Y. Zhou, Y. Wang, J. Xiao, Z. Wang, Y. Zhang, and J. Chen, \"The four color theorem for cell instance segmentation,\" in Forty-second International Conference on Machine Learning, 2025. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 555, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 555, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 555, + 690 + ], + "type": "text", + "content": "[52] J.-X. Zhao, J.-J. Liu, D.-P. Fan, Y. Cao, J. Yang, and M.-M. Cheng, \"Egnet: Edge guidance network for salient object detection,\" in Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 8779-8788. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "type": "text", + "content": "[53] Z. Luo, A. Mishra, A. Achkar, J. Eichel, S. Li, and P.-M. Jodoin, \"Non-local deep features for salient object detection,\"" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14440" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "in Proceedings of the IEEE Conference on computer vision and pattern recognition, 2017, pp. 6609-6617." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 297, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 297, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 297, + 139 + ], + "type": "text", + "content": "[54] Y. Zhuge, G. Yang, P. Zhang, and H. Lu, \"Boundary-guided feature aggregation network for salient object detection,\" IEEE Signal Processing Letters, vol. 25, no. 12, pp. 1800-1804, 2018." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 296, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 296, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 296, + 185 + ], + "type": "text", + "content": "[55] Z. Tu, Y. Ma, C. Li, J. Tang, and B. Luo, \"Edge-guided nonlocal fully convolutional network for salient object detection,\" IEEE transactions on circuits and systems for video technology, vol. 31, no. 2, pp. 582-593, 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 296, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 296, + 230 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 296, + 230 + ], + "type": "text", + "content": "[56] X. Wang, H. Ma, X. Chen, and S. You, \"Edge preserving and multi-scale contextual neural network for salient object detection,\" IEEE Transactions on Image Processing, vol. 27, no. 1, pp. 121-134, 2017." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 232, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 296, + 274 + ], + "type": "text", + "content": "[57] Z. Wu, L. Su, and Q. Huang, \"Stacked cross refinement network for edge-aware salient object detection,\" in Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 7264-7273. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 296, + 319 + ], + "type": "text", + "content": "[58] G. Li, Y. Xie, L. Lin, and Y. Yu, \"Instance-level salient object segmentation,\" in Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 2386-2395. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 321, + 296, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 296, + 376 + ], + "type": "text", + "content": "[59] Y. Hao, Y. Liu, Z. Wu, L. Han, Y. Chen, G. Chen, L. Chu, S. Tang, Z. Yu, Z. Chen et al., \"Edgeflow: Achieving practical interactive segmentation with edge-guided flow,\" in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 1551-1560." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 377, + 296, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 296, + 421 + ], + "type": "text", + "content": "[60] H. Liu, R. A. R. Soto, F. Xiao, and Y. J. Lee, \"Yolactedge: Real-time instance segmentation on the edge,\" in 2021 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2021, pp. 9579-9585." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 422, + 296, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 296, + 465 + ], + "type": "text", + "content": "[61] Y. Zhou, X. Xu, F. Shen, X. Zhu, and H. T. Shen, \"Flow-edge guided unsupervised video object segmentation,\" IEEE Transactions on circuits and systems for video technology, vol. 32, no. 12, pp. 8116-8127, 2021. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 467, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 296, + 510 + ], + "type": "text", + "content": "[62] L. Liu, H. Li, and M. Gruteser, \"Edge assisted real-time object detection for mobile augmented reality,\" in The 25th annual international conference on mobile computing and networking, 2019, pp. 1-16." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 512, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 296, + 555 + ], + "type": "text", + "content": "[63] C. Zhan, X. Duan, S. Xu, Z. Song, and M. Luo, \"An improved moving object detection algorithm based on frame difference and edge detection,\" in Fourth international conference on image and graphics (ICIG 2007). IEEE, 2007, pp. 519-523." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 557, + 296, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 296, + 601 + ], + "type": "text", + "content": "[64] M. Hanyao, Y. Jin, Z. Qian, S. Zhang, and S. Lu, \"Edge-assisted online on-device object detection for real-time video analytics,\" in IEEE INFOCOM 2021-IEEE Conference on Computer Communications. IEEE, 2021, pp. 1-10. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 602, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 296, + 645 + ], + "type": "text", + "content": "[65] Z. Xu, X. Baojie, and W. Guoxin, \"Canny edge detection based on open cv,\" in 2017 13th IEEE international conference on electronic measurement & instruments (ICEMI). IEEE, 2017, pp. 53-56. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 296, + 689 + ], + "type": "text", + "content": "[66] L. Xuan and Z. Hong, \"An improved canny edge detection algorithm,\" in 2017 8th IEEE international conference on software engineering and service science (ICSESS). IEEE, 2017, pp. 275-278." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 692, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 296, + 713 + ], + "type": "text", + "content": "[67] W. Rong, Z. Li, W. Zhang, and L. Sun, \"An improved canny edge detection algorithm,\" in 2014 IEEE international con" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "ference on mechatronics and automation. IEEE, 2014, pp. 577-582. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 555, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 118 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 118 + ], + "type": "text", + "content": "[68] L. Ding and A. Goshtasby, “On the canny edge detector,” Pattern recognition, vol. 34, no. 3, pp. 721-725, 2001. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 120, + 555, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 120, + 555, + 151 + ], + "spans": [ + { + "bbox": [ + 316, + 120, + 555, + 151 + ], + "type": "text", + "content": "[69] W. McIlhagga, “The canny edge detector revisited,” International Journal of Computer Vision, vol. 91, pp. 251–261, 2011." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 153, + 555, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 555, + 186 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 555, + 186 + ], + "type": "text", + "content": "[70] R. Song, Z. Zhang, and H. Liu, \"Edge connection based canny edge detection algorithm,\" Pattern Recognition and Image Analysis, vol. 27, pp. 740-747, 2017." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 187, + 555, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 187, + 555, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 187, + 555, + 230 + ], + "type": "text", + "content": "[71] P. Bao, L. Zhang, and X. Wu, \"Canny edge detection enhancement by scale multiplication,\" IEEE transactions on pattern analysis and machine intelligence, vol. 27, no. 9, pp. 1485-1490, 2005. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 232, + 555, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 555, + 276 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 555, + 276 + ], + "type": "text", + "content": "[72] D. Xu, J. Liu, X. Li, Z. Liu, and X. Tang, \"Insignificant shadow detection for video segmentation,\" IEEE Transactions on Circuits and Systems for Video Technology, vol. 15, no. 8, pp. 1058-1064, 2005. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 277, + 555, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 555, + 331 + ], + "type": "text", + "content": "[73] S. Seoni, V. Jahmunah, M. Salvi, P. D. Barua, F. Molinari, and U. R. Acharya, “Application of uncertainty quantification to artificial intelligence in healthcare: A review of last decade (2013–2023),\" Computers in Biology and Medicine, p. 107441, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 333, + 555, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 333, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 333, + 555, + 376 + ], + "type": "text", + "content": "[74] A. Blake, R. Curwen, and A. Zisserman, “A framework for spatiotemporal control in the tracking of visual contours,” International Journal of Computer Vision, vol. 11, pp. 127–145, 1993. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 377, + 555, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 555, + 433 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 555, + 433 + ], + "type": "text", + "content": "[75] X. He, R. S. Zemel, and M. A. Carreira-Perpinan, “Multiscale conditional random fields for image labeling,” in Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004., vol. 2. IEEE, 2004, pp. II-II. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 434, + 555, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 555, + 487 + ], + "type": "text", + "content": "[76] D. Dera, N. C. Bouaynaya, G. Rasool, R. Shterenberg, and H. M. Fathallah-Shaykh, “Premium-cnn: Propagating uncertainty towards robust convolutional neural networks,” IEEE Transactions on Signal Processing, vol. 69, pp. 4669–4684, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 490, + 555, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 555, + 533 + ], + "type": "text", + "content": "[77] Y. Gal and Z. Ghahramani, “Dropout as a bayesian approximation: Representing model uncertainty in deep learning,” in international conference on machine learning. PMLR, 2016, pp. 1050–1059." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 535, + 555, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 555, + 567 + ], + "type": "text", + "content": "[78] A. Kendall and Y. Gal, “What uncertainties do we need in bayesian deep learning for computer vision?” Advances in neural information processing systems, vol. 30, 2017. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 569, + 555, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 569, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 569, + 555, + 612 + ], + "type": "text", + "content": "[79] A. Ashukha, A. Lyzhov, D. Molchanov, and D. Vetrov, “Pitfalls of in-domain uncertainty estimation and ensembling in deep learning,” in The Eighth International Conference on Learning Representations, 2020. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 614, + 555, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 555, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 555, + 678 + ], + "type": "text", + "content": "[80] M. Sensoy, L. Kaplan, and M. Kandemir, \"Evidential deep learning to quantify classification uncertainty,\" in Advances in Neural Information Processing Systems, S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, Eds., vol. 31. Curran Associates, Inc., 2018. 3, 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 680, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 555, + 713 + ], + "type": "text", + "content": "[81] K. Zou, Z. Chen, X. Yuan, X. Shen, M. Wang, and H. Fu, “A review of uncertainty estimation and its application in medical imaging,” Meta-Radiology, p. 100003, 2023." + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14441" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 677 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 115 + ], + "type": "text", + "content": "[82] W. Bao, Q. Yu, and Y. Kong, \"Evidential deep learning for open set action recognition,\" in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 13349-13358. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "type": "text", + "content": "[83] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, \"Imagenet: A large-scale hierarchical image database,\" in 2009 IEEE Conference on Computer Vision and Pattern Recognition, 2009, pp. 248-255. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 228 + ], + "type": "text", + "content": "[84] H. Yang, C. Chen, Y. CHEN, , H. C. Yip, and D. QI, \"Uncertainty estimation for safety-critical scene segmentation via fine-grained reward maximization,\" in Advances in Neural Information Processing Systems, A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, Eds., vol. 36. Curran Associates, Inc., 2023, pp. 36238-36249. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 296, + 274 + ], + "type": "text", + "content": "[85] M. Li, L. Hu, Z. Xiong, B. Zhang, P. Pan, and D. Liu, \"Recurrent dynamic embedding for video object segmentation,\" in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 1332-1341. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 275, + 296, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 296, + 317 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 296, + 317 + ], + "type": "text", + "content": "[86] M. Lan, J. Zhang, L. Zhang, and D. Tao, “Learning to learn better for video object segmentation,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 1, 2023, pp. 1205–1212. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 296, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 296, + 374 + ], + "type": "text", + "content": "[87] J. Wang, D. Chen, Z. Wu, C. Luo, C. Tang, X. Dai, Y. Zhao, Y. Xie, L. Yuan, and Y.-G. Jiang, “Look before you match: Instance understanding matters in video object segmentation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 2268–2278. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 296, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 296, + 418 + ], + "type": "text", + "content": "[88] J. Shi, Q. Yan, L. Xu, and J. Jia, \"Hierarchical image saliency detection on extended cssd,\" IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 38, no. 4, pp. 717-729, 2016. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 420, + 296, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 420, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 420, + 296, + 464 + ], + "type": "text", + "content": "[89] L. Wang, H. Lu, Y. Wang, M. Feng, D. Wang, B. Yin, and X. Ruan, “Learning to detect salient objects with image-level supervision,” in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 3796–3805." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 465, + 296, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 465, + 296, + 508 + ], + "spans": [ + { + "bbox": [ + 56, + 465, + 296, + 508 + ], + "type": "text", + "content": "[90] X. Li, T. Wei, Y. P. Chen, Y.-W. Tai, and C.-K. Tang, \"Fss1000: A 1000-class dataset for few-shot segmentation,\" in 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 2866-2875." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 510, + 296, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 296, + 553 + ], + "type": "text", + "content": "[91] Y. Zeng, P. Zhang, Z. Lin, J. Zhang, and H. Lu, \"Towards high-resolution salient object detection,\" in 2019 IEEE/CVF International Conference on Computer Vision (ICCV), 2019, pp. 7233-7242." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 555, + 296, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 296, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 296, + 609 + ], + "type": "text", + "content": "[92] H. K. Cheng, J. Chung, Y.-W. Tai, and C.-K. Tang, \"Cascadepsp: Toward class-agnostic and very high-resolution segmentation via global and local refinement,\" in 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 8887-8896. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 610, + 296, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 610, + 296, + 642 + ], + "spans": [ + { + "bbox": [ + 56, + 610, + 296, + 642 + ], + "type": "text", + "content": "[93] N. Xu, L. Yang, Y. Fan, D. Yue, Y. Liang, J. Yang, and T. Huang, \"Youtube-vos: A large-scale video object segmentation benchmark,\" in ECCV, 2018. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 644, + 296, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 644, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 644, + 296, + 677 + ], + "type": "text", + "content": "[94] H. Ding, C. Liu, S. He, X. Jiang, P. H. Torr, and S. Bai, “MOSE: A new dataset for video object segmentation in complex scenes,” in ICCV, 2023. 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "14442" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Structure-Guided Diffusion Models for High-Fidelity Portrait Shadow Removal/73c9755b-db1d-47f9-8857-79c13b240555_content_list.json b/2025/Structure-Guided Diffusion Models for High-Fidelity Portrait Shadow Removal/73c9755b-db1d-47f9-8857-79c13b240555_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c64a90017d0e6234cd674a0b340e07f960a384e5 --- /dev/null +++ b/2025/Structure-Guided Diffusion Models for High-Fidelity Portrait Shadow Removal/73c9755b-db1d-47f9-8857-79c13b240555_content_list.json @@ -0,0 +1,1979 @@ +[ + { + "type": "text", + "text": "Structure-Guided Diffusion Models for High-Fidelity Portrait Shadow Removal", + "text_level": 1, + "bbox": [ + 96, + 130, + 898, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wanchang Yu $^{1}$ Qing Zhang $^{1,2*}$ Rongjia Zheng $^{1}$ Wei-Shi Zheng $^{1,2}$", + "bbox": [ + 210, + 178, + 785, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science and Engineering, Sun Yat-sen University, China", + "bbox": [ + 191, + 198, + 805, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China", + "bbox": [ + 109, + 215, + 885, + 234 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3cf8d30aa782822b05677e0f926112a293b96c860170714cad13c35ae3d8adcf.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 91, + 252, + 228, + 460 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c3298ebe9d3a8845c02a3cbf22a06de24fb03e23211ea83475f2a9246b5941c5.jpg", + "image_caption": [ + "ShadowDiffusion [11]" + ], + "image_footnote": [], + "bbox": [ + 228, + 252, + 362, + 460 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/161affd385f29f95cd7b767d7c1d35cc417e9480a68c18faed34ba5a79ba1894.jpg", + "image_caption": [ + "IC-light [52]" + ], + "image_footnote": [], + "bbox": [ + 364, + 252, + 496, + 460 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c3c399eef96a8188dccafb0ddbae2b8fc216a3681ff6c74e7a5b93cd5652bf8a.jpg", + "image_caption": [ + "BSR [28]", + "Figure 1. Comparison on portrait shadow removal. Our method allows to robustly remove facial cast shadows while faithfully preserving the portrait identities as well as structure details in shadow regions and also ensuring the overall naturalness." + ], + "image_footnote": [], + "bbox": [ + 500, + 252, + 633, + 460 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ccc485ed95a40fca12f4db7182d0ba7362811c9b4641daf2ce1b843cb5dda6a6.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 635, + 252, + 769, + 460 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3e7d2516ef60de9d5d917cd9eab1fc97d84fff5750771760aab23f5e88bb3064.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 769, + 252, + 905, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 529, + 326, + 544 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present a diffusion-based portrait shadow removal approach that can robustly produce high-fidelity results. Unlike previous methods, we cast shadow removal as diffusion-based inpainting. To this end, we first train a shadow-independent structure extraction network on a real-world portrait dataset with various synthetic lighting conditions, which allows to generate a shadow-independent structure map including facial details while excluding the unwanted shadow boundaries. The structure map is then used as condition to train a structure-guided inpainting diffusion model for removing shadows in a generative manner. Finally, to restore the fine-scale details (e.g., eyelashes, moles and spots) that may not be captured by the structure map, we take the gradients inside the shadow regions as guidance and train a detail restoration diffusion model to refine the shadow removal result. Extensive experiments on the benchmark datasets show that our method clearly outperforms existing methods, and is effective to avoid previously common issues such as facial identity tampering, shadow residual, color distortion, structure blurring, and loss of details. Our code is avail", + "bbox": [ + 88, + 558, + 485, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "able at https://github.com/wanchang-yu/Structure-Guided-Diffusion-for-Portrait-Shadow-Removal.", + "bbox": [ + 511, + 530, + 906, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 587, + 645, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Taking selfies is very popular nowadays due to the readily-available cameras on various devices, particularly the mobile phones. However, portrait photos taken in the wild lacking professional lighting conditions often suffer from undesired foreign shadows cast by external objects (see Figure 1 for examples). The presence of facial shadows not only severely degrade the overall visual appeal of the portraits, but also adversely affect many downstream vision tasks, such as face detection and recognition [4, 5, 41, 55], and portrait image editing [21, 27, 47]. Hence, portrait shadow removal techniques are widely required.", + "bbox": [ + 509, + 612, + 906, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Currently, portrait shadow removal remains a challenge despite the remarkable progress in natural image shadow removal. The reason behind is twofold. First of all, due to factors such as privacy and difficulty in data collection, large-scale paired data for portrait shadow is extremely scarce, which significantly limits the development of portrait shadow removal methods, especially learning-based solutions. On the other hand, as this task is portrait-", + "bbox": [ + 509, + 780, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author (zhangq93@mail.sysu.edu.cn).", + "bbox": [ + 91, + 886, + 382, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11675", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "oriented, it has very low tolerance for visual artifacts, even minor imperfections on facial color and details could make the final result unnatural and unacceptable. Note, although existing software allow users to interactively adjust photos, it is still rather tedious and difficult for non-experts to manually produce high-fidelity portrait shadow removal results.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing shadow removal methods [11, 26, 45] mainly target natural images due to the availability of large-scale paired datasets. However, these methods basically fail to deal with portrait shadows, as shown in Figures 5 and 6. As there is no accessible large-scale paired datasets for portrait shadow removal, current methods in this field focus primarily on constructing synthetic paired data to enable supervised learning of shadow removal [28, 33, 52, 56]. Suffering from the domain gap between real-world portraits and synthetic data, these methods typically do not work well for real-world portraits with complex shadows. To avoid the domain gap issue led by synthetic training data, a zero-shot solution is introduced in [14], which proposed to leverage the generative prior from pre-trained StyleGAN2 [24] to recover shadow-free portrait through test-time GAN inversion. However, this method tends to modify face identity and incur unnatural results. Note, although recent image relighting techniques [8, 16, 43, 50] have shown remarkable progress, they struggle to enable high-fidelity portrait shadow removal due to the inherent difficulty in estimating ambient light and facial geometry from a single image.", + "bbox": [ + 91, + 184, + 482, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To robustly produce high-fidelity portrait shadow removal results, we in this work present a diffusion-based approach. Particularly, we propose to cast portrait shadow removal as diffusion-based inpainting, so as to take full advantage of the strong generation capability of diffusion models. To this end, we first train a shadow-independent structure extraction network (SE-Net) on a dataset of paired real-world portraits with various synthetic lighting conditions, from which we obtain a structure map containing facial details while excluding the shadow boundaries. With the structure map, we then train a structure-guided inpainting diffusion model by enforcing reconstruction of randomly masked shadow-free portraits. Finally, to restore the fine-scale facial details that may not be captured by the structure map, we further refine the shadow removal result by a detail restoration diffusion model under the guidance of the gradients in the original shadow regions. Note, although our method requires a test shadow mask as input, it is highly robust to inaccurate masks, requiring only that the mask covers the shadow regions, even if it includes nonshadow regions. The major contributions of this work are:", + "bbox": [ + 91, + 503, + 482, + 819 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a simple yet effective unsupervised approach that addresses portrait shadow removal from a perspective of diffusion-based inpainting.", + "- We introduce a relighting-based data synthesis strategy for constructing the training data, and develop a shadow" + ], + "bbox": [ + 89, + 825, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "independent structure extraction network.", + "bbox": [ + 527, + 90, + 803, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We show that our method outperforms state-of-the-art methods for portrait shadow removal.", + "bbox": [ + 513, + 106, + 903, + 135 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 148, + 653, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Shadow removal. Early methods typically rely on various handcrafted priors or assumptions to work [7, 9, 12]. Recent methods are basically deep learning based [3, 10, 11, 17, 20, 22, 23, 25, 29, 31, 57, 60, 61], benefiting mostly from the availability of large-scale paired datasets such as ISTD [42] and SRD [35]. However, as these methods are tailored for natural images, they are not so effective in dealing with portraits. In contrast to natural images, portrait shadow removal has received much less research attention, due to the lack of large-scale paired datasets. To overcome the data scarcity issue, some methods [28, 56] proposed to synthesize paired shadow and shadow-free portrait images for training. As their data synthesis strategies involve only simple adjustments on brightness and saturation, without considering facial geometry and albedo, their methods typically do not work well for real-world portraits. He et al. [14] achieved zero-shot portrait shadow removal by utilizing generative prior from StyleGAN2, but this method may result in undesired modifications to facial details and thus fails to preserve the face identity. More recently, image relighting is also used for shadow removal [8, 16, 43, 50, 52]. However, this method tends to induce unnatural results because it is very challenging to obtain accurate prediction of diffuse lighting and albedo from a single portrait image.", + "bbox": [ + 511, + 176, + 903, + 554 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image inpainting. Before the era of deep learning, patch-based self-similarity of natural images has long been the foundation for image inpainting [2, 44]. The emergence of deep learning has then led to great progress in image inpainting [18, 34, 40, 51], as it is quite easy to construct paired training data. Recently, diffusion model has been a preferable choice for image inpainting [1, 19, 32, 37, 46, 59], due to its impressive generation capability. Note, naive applying diffusion models to portrait shadow removal fail to produce visually natural results with unchanged portrait identity. A recent work closely related to our method also shows that it is feasible to remove shadows via image inpainting [26]. However, it does not work well for portrait shadow removal (see Figures 5 and 6), even we fine-tuned their model on real-world paired shadow and shadow-free portrait images provided in [56].", + "bbox": [ + 511, + 559, + 906, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Our Method", + "text_level": 1, + "bbox": [ + 513, + 814, + 643, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given a portrait image with cast shadows, our method aims to remove the facial shadows while faithfully preserving the facial details in the shadow regions. Figure 2 presents the overall inference pipeline of our method.", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11676", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a7303ded20a690d3f0212416f31630b3978b34c76ed7a4959cdd19b167a9cbc4.jpg", + "image_caption": [ + "Figure 2. Inference pipeline of our method. Given a input portrait image and the corresponding shadow mask, we first feed the input image to the shadow-independent structure extraction network (SE-Net) to get a structure map. Then, we perform shadow removal by inpainting the shadow regions with a diffusion model conditioned on the structure map. Finally, the shadow removal output is refined by a detail restoration diffusion model to restore the fine-scale facial details that may not be captured by the strcutre map." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 903, + 277 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9a86e93f454fd316c70ab47d6505efd197a8b51107ac8b0a8b7e6cdd3b100bb4.jpg", + "image_caption": [ + "Figure 3. Training pipeline of SE-Net. Note, we adopt a trained PDG model for structure extraction from [49], while it is not invariant to illumination changes." + ], + "image_footnote": [], + "bbox": [ + 94, + 361, + 480, + 553 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Shadow-Independent Structure Extraction", + "text_level": 1, + "bbox": [ + 89, + 628, + 454, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Figure 2, the key to obtaining high-quality portrait shadow removal through inpainting is to acquire a structure map free of shadow boundaries for guiding the inpainting process. However, although existing structure/edge extraction methods are quite effective in identifying edges and structures, they are sensitive to illumination discontinuities and will inevitably include shadow boundaries in their results (see Figure 7). Therefore, we propose to build a shadow-independent structure extraction network.", + "bbox": [ + 89, + 651, + 482, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training data synthesis. To enable effective training of the shadow-independent structure extraction network, as shown in Figure 3, we construct a synthetic paired dataset of well-lit real-world portraits without foreign shadows on faces and the corresponding relighted portraits with various illumination discontinuities, based on the shadow-free portraits in the CelebA dataset [30]. Specifically, for a well-lit portrait", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$I$ , we first apply an off-the-shelf physics-based relighting method [58] over it to get a relighted image $I_{relit}$ , which is then combined with the original portrait $I$ to produce a synthetic image $I_{syn}$ with illumination discontinuities by", + "bbox": [ + 511, + 362, + 906, + 424 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI _ {s y n} = M \\odot I _ {r e l i t} + (1 - M) \\odot I, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 438, + 906, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $M$ is a random facial mask, and $\\odot$ denotes elementwise multiplication. Note, our synthetic data is only used for training the SE-Net and the subsequent detail resotration diffusion model, rather than for directly training shadow removal model as done in [56].", + "bbox": [ + 511, + 467, + 905, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training losses. As shown in Figure 3, we adopt a trained structure extraction model from [49] (referred to as PDG in paper) to produce structure map for the original well-lit portrait, which is treated as pseudo ground-truth for supervising the training of our SE-Net. By respectively denoting our SE-Net and the PDG model as $G_{s}$ and $G_{p}$ , the training losses of our SE-Net are defined as:", + "bbox": [ + 511, + 550, + 906, + 656 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\mathcal {L} _ {\\text {r e c}} = \\| G _ {s} (I _ {\\text {s y n}}) - G _ {p} (I) \\| _ {1}, \\\\ \\mathcal {L} _ {\\text {p e r c e p t u a l}} = \\mathcal {L} _ {\\text {L P I P S}} \\left(G _ {s} (I _ {\\text {s y n}}), G _ {p} (I)\\right), \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 667, + 903, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{LPIPS}$ refers to the learned perceptual image patch similarity [54]. Besides, we also utilize a discriminator to determine whether the output of SE-Net belongs to the structure image domain, and further define an adversarial loss $\\mathcal{L}_{GAN}$ (see supplementary material for details). In summary, the overall loss function for training the SE-Net is given by:", + "bbox": [ + 511, + 719, + 906, + 825 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {r e c}} + \\lambda_ {1} \\mathcal {L} _ {\\text {p e r c e p t u a l}} + \\lambda_ {2} \\mathcal {L} _ {\\text {G A N}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 840, + 903, + 857 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where we empirically set $\\lambda_1 = 0.5$ , $\\lambda_1 = 0.25$ , which produce good results, as demonstrated in Figure 7.", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11677", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b4d96f412dedaf6e5e4e50b52fbfb67030bf64e18a675e3f1e14f28033e1b8d6.jpg", + "image_caption": [ + "Structure-guided Portrait Inpainting" + ], + "image_footnote": [], + "bbox": [ + 99, + 116, + 467, + 290 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/92f6fff8c2118d35e7a7305efd6deaa999a3b20db6f72abbdb1a3bde4d543cb3.jpg", + "image_caption": [ + "Gradient-guided Detail Restoration", + "Figure 4. Training pipeline of our diffusion models. Here we introduce how our structure-guided portrait inpainting diffusion model and gradient-guided detail restoration diffusion model are trained." + ], + "image_footnote": [], + "bbox": [ + 480, + 113, + 895, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Structure-Guided Portrait Inpainting", + "text_level": 1, + "bbox": [ + 89, + 351, + 415, + 368 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Figure 2, with structure map, the next step is to inpaint the shadow regions. To this end, we develop a structure-guided portrait inpainting diffusion model.", + "bbox": [ + 89, + 373, + 483, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Structure-aware denoising. As shown in Figure 4, we start with a shadow-free portrait as input and then randomly mask some facial regions to create a masked input, followed by a diffusion model to reconstruct the original input by inpainting the masked regions. The inpainting is rooted in the forward diffusion process, where the diffusion model learns a Markovian chain by progressively corrupting the target image $x_0$ (namely the masked input) with Gaussian noise at each timestep $t$ , such that it is transformed into a Gaussian distribution (referred to as $\\mathcal{N}$ in the following). Specifically, after reparameterization [15], our forward process obtains $x_t$ from $x_0$ in a single step by:", + "bbox": [ + 89, + 425, + 483, + 604 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 611, + 480, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\bar{\\alpha}_t = \\prod_{i=0}^t \\alpha_i$ is the cumulative product of forward noise weight parameters $\\alpha_i = 1 - \\beta_i$ , while $\\epsilon \\sim \\mathcal{N}(0, I)$ denotes random Gaussian noise.", + "bbox": [ + 89, + 633, + 482, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As for the reverse denoising process, we use the sampling strategy of DDIM [39] for acceleration. Note, unlike unconditional image generation, our goal is to faithfully reconstruct the original unmasked portrait, the reverse denoising process is thus conditioned on a structure map $S$ and defined as:", + "bbox": [ + 89, + 680, + 483, + 768 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t - 1} = \\sqrt {\\bar {\\alpha} _ {t - 1}} \\left(\\frac {x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\cdot \\mathbf {e} _ {t}}{\\sqrt {\\bar {\\alpha} _ {t}}}\\right) + \\sqrt {1 - \\bar {\\alpha} _ {t - 1}} \\cdot \\mathbf {e} _ {t}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 772, + 480, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{e}_t = \\epsilon_\\theta (x_t,I_M,S,M,t)$ refers to the noise estimated by a deep neural network $\\epsilon_{\\theta}$ , where $I_M$ is the masked input created from a random mask $M$ . Following [15], the training objective for diffusion model can be simplified as:", + "bbox": [ + 89, + 818, + 482, + 878 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i f f}} (\\epsilon) = \\| \\epsilon - \\epsilon_ {\\theta} \\left(x _ {t}, I _ {M}, S, M, t\\right) \\| _ {2} ^ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 883, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon$ is the ground truth noise, while $\\epsilon_{\\theta}(x_t,I_M,S,M,t)$ is the noise predicted by the network. As shown in Figure 2, a facial shadow mask is required at the inference stage, which is a dilated version of the facial shadow detection result from the trained model of [6].", + "bbox": [ + 511, + 351, + 905, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Shadow mask refinement. To enhance the robustness of our method to inaccurate shadow mask, we propose to refine the mask based on the shadow removal result generated by the inpainting diffusion model. Specifically, we first compute the absolute difference between the input image and the shadow removal result, and then perform Otsu-based thresholding to get a refined shadow mask excluding all the non-shadow regions. With the refined mask, we further update the shadow removal result by combining it with the input image, so as to ensure that non-shadow regions in the result are unchanged.", + "bbox": [ + 511, + 441, + 906, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Gradient-Guided Detail Restoration", + "text_level": 1, + "bbox": [ + 513, + 643, + 826, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As the original PDG model is designed for generating portrait drawing formed by large-scale structures, thus our SE-Net is similarly not so effective in capturing fine-scale portrait details, such as eyelashes, moles, and spots (see the bottom row in Figure 7 for example). To address this issue, we design a detail restoration diffusion model to refine the shadow removal result generated by the inpainting diffusion model, where we directly take the gradients in the shadow regions as guidance to restore the fine-scale details. Figure 4 describes how we train the detail restoration diffusion model. Note, gradient-aware denoising process is similar to the structure-aware denoising described above, and the two share the same forward diffusion and reverse denoising formulations described in Equations 4 and 5, with the latter conditioned on gradient map rather than structure map.", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11678", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8e134908d42ac20fe5bd0c032fc838bfc124751ae2886510fd474bcfb1e911c9.jpg", + "image_caption": [ + "Figure 5. Visual comparison with state-of-the-art methods on test images from the dataset provided in [56]." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 903, + 411 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/576c717c0436b1b027f552c6dbc6d76fb5a4e0dcad35584dfc65a1c4555dbcc7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAllShadow regionNon-shadow region
SSIM↑LPIPS↓RMSE↓SSIM↑LPIPS↓RMSE↓SSIM↑LPIPS↓RMSE↓
ShadowDiffusion [11]0.6500.17738.2460.9010.03715.5610.5790.22352.276
Inpaint4Shadow [26]0.7660.09422.7090.9100.03317.2500.7080.16445.456
ShadowFormer [10]0.7720.08621.4130.9260.02715.2880.7230.14942.341
HomoFormer [45]0.7860.10021.7240.9130.04116.2480.7110.16742.950
PSM [56]0.7820.074-------
UPSR [14]0.7310.10926.9150.9000.03920.6590.6680.18845.391
BSR [28]0.6050.11828.5320.9100.03615.1930.5480.17143.627
IC-light [52]0.5140.27862.5390.9160.05733.0610.4780.31969.208
Ours w/o structure guidance0.7570.10123.5980.8980.04318.5140.7140.13839.000
Ours w/ PDG predicted structure0.7790.09219.9630.9190.03514.0620.7250.13137.387
Ours w/ data synthesis in PSM0.7850.08619.7970.9250.03213.7520.7290.13638.485
Ours w/o detail-restoration0.8140.06218.5370.9430.02012.7480.7470.12036.326
Ours (full method)0.8300.05617.1620.9730.01110.1980.7780.09134.573
", + "bbox": [ + 94, + 436, + 906, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison on the real-world portrait shadow removal dataset provided in [56].", + "bbox": [ + 178, + 686, + 816, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Implementation and Training Details", + "text_level": 1, + "bbox": [ + 89, + 732, + 410, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Network architecture. The SE-Net is an encoder-decoder network, which starts with a flat convolution and two down convolution blocks, followed by nine residual blocks, and finally two up convolution blocks and a final convolution layer. Please see also the supplementary material for visual illustration of the architecture of SE-Net. As for the inpainting diffusion model and the detail-restoration diffusion model, we follow [15] to adopt the same U-Net architecture.", + "bbox": [ + 89, + 756, + 482, + 877 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training details. To train SE-Net, we synthesize 20,000", + "bbox": [ + 89, + 883, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "portrait images with diverse lighting conditions based on the CelebA dataset [30], and train it for 3 epochs with a batch size of 8 and a learning rate of $1.5 \\times 10^{-5}$ . As for the inpainting diffusion model, we train it on the FFHQ dataset for 300 epochs with a batch size of 32, and train the detail restoration diffusion model on the same synthetic dataset as the SE-Net for 200 epochs. We build our model on PyTorch and use the Adam optimizer with momentum parameters of (0.9, 0.999) and an initial learning rate of $1 \\times 10^{-4}$ . Following [38], we employ the Kaiming initialization [13] to initialize the weights of our model and apply 0.9999 Expo", + "bbox": [ + 511, + 733, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11679", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e4ac4d3cf14c4ef937c8b88e60c660008fd027e29768ef42ee628e1ebbf2110e.jpg", + "image_caption": [ + "Figure 6. Visual comparison with state-of-the-art methods on test images from our collected dataset." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 903, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "nential Moving Average (EMA) in all our experiments. We use 2000 diffusion steps and a noise schedule $\\beta_{t}$ linearly increasing from $1\\times 10^{-6}$ to $1\\times 10^{-2}$ for training. During inference, we perform both the inpainting diffusion and detail restoration diffusion for 40 denoising steps.", + "bbox": [ + 89, + 450, + 483, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 544, + 222, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. Following previous portrait shadow removal approaches [14, 28], we evaluate our method on the only publicly available real-world paired dataset provided in PSM [56], which contains 9 subjects and 100 shadowed portrait images in varying poses, shadow shapes, illumination conditions, and shadow types. To better examine the effectiveness of our method, we additionally construct a dataset of 200 portraits with diverse facial shadows. Note, we only conduct visual comparison on our collected dataset due to the lack of ground truth shadow removal results.", + "bbox": [ + 89, + 570, + 483, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. We use three commonly-used metrics to quantitatively evaluate the performance of our method, including SSIM, LPIPS, and the root mean square error (RMSE) computed in the LAB color space for measuring the similarity between the predicted results and the corresponding ground-truth shadow-free images.", + "bbox": [ + 89, + 729, + 483, + 820 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Comparison with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 89, + 832, + 454, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our method with various state-of-the-art shadow removal methods, including four portrait shadow removal and editing methods, i.e., PSM [56],", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "UPSR [14], and BSR [28], IC-light [52], as well as four natural image shadow removal methods, i.e., ShadowDiffusion [11], Inpaint4Shadow [26], ShadowFormer [10], and HomoFormer [45]. For fair comparison, we produce their results using publicly-available implementation or trained models provided by the authors with recommended parameter setting. Note that, we do not compare with other recent relighting-based shadow editing methods [8, 16, 43, 50], as their codes and trained models are not yet released.", + "bbox": [ + 511, + 450, + 906, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visual comparison. We in Figures 5 and 6 conduct visual comparison between our method and previous shadow removal methods. As shown, shadow removal methods designed for natural images, i.e., Inpaint4Shadow [26], HomoFormer [45], and ShadowDiffusion [11] are not well-suited to portraits and tend to induce obvious shadow residuals and appearance distortions. As for the other three portrait shadow removal methods, UPSR [14] and BSR [28] fail to preserve the facial details and may also lead to shadow residuals and color distortions, while IC-Light [52] struggles to remove the facial shadows due to the inherent difficulty in estimating the ambient light and facial geometry. Please see also the supplementary material for more visual comparison results.", + "bbox": [ + 511, + 599, + 908, + 811 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative comparison. In Table 1, we quantitatively compare our method with other methods in terms of the SSIM, LPIPS, and RMSE metrics. As can be seen, our method outperforms all the compared methods on all the three metrics, demonstrating its effectiveness and superior-", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11680", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d3a1f1b9909763a8bd6c93a664c30f93c5377033abf947f9d3ab94db8cf19860.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 87, + 207, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fc40af2c9ab6980918d472d07b257d94a90d98099c1eb677a49c059a770c2382.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 91, + 176, + 207, + 265 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7e042422b27826ef4c7bb4600cd59234f09dfc597ea3f9076d05e6db02fe7bf1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 209, + 88, + 323, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/861e15e6be71cd23543034cd3a7af6c7455ca23d3d5b780c5bfd5b664dcae781.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 209, + 176, + 323, + 266 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/39bd9cb93efc8dcd2a1d41d3da183eb2b4bd922088c5f6445639afd8b84dc899.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 326, + 88, + 439, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c54e6a6f03ffcfcda671e32c13b6c6f573e90cff2791161f84df163d7d0208bf.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 326, + 176, + 439, + 266 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/047d3b8ae34d6c8a75ad7b47ee8d8971d8d1b1f3f0d4b86850f69cdbb8e34daf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 441, + 88, + 557, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1011ee522bfad66898834852b5f879b146e90a066068f99f4a6d26be2ab18f52.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 441, + 176, + 557, + 266 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2fcd9ba1b290fd6310aa9dc580c47022dda7133249eacbcec59b618e302e000e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 88, + 671, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4a2571cfe0e8dc6891c964a45972191dbc0f7a7dd1ee83f4be96fdc23af5bd04.jpg", + "image_caption": [ + "(e)" + ], + "image_footnote": [], + "bbox": [ + 558, + 176, + 671, + 265 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/18433bf6eea6c41964b0033ebbf4515de80be3ec799ab2983cf3bde3b36c7d08.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 88, + 787, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/25e235c91028bb90ebd6946430dccc7ba3d4e1c8184f1289ec8c24c159213311.jpg", + "image_caption": [ + "(1)" + ], + "image_footnote": [], + "bbox": [ + 674, + 176, + 787, + 265 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/79c1e66398c864e6e6fbed933cc1f66e6dfe902b87e9966f61aeb5774a741f17.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 88, + 903, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/08f57d647c4a4bd0a8ab2e697a408f5ecc5eb467edbf1558c3e5865e90c36542.jpg", + "image_caption": [ + "(g)" + ], + "image_footnote": [], + "bbox": [ + 790, + 176, + 903, + 265 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ity in portrait shadow removal. Note, as PSM [56] provides only code for synthesizing training data, without releasing the training code and trained model, its quantitative results in Table 1 are as reported in their paper.", + "bbox": [ + 89, + 345, + 482, + 407 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 417, + 243, + 434 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of our shadow-independent structure extraction. Although the trained model of PDG [49] is utilized to train our structure extraction, the structure map it produces is not what we expected because it will include the shadow boundaries and lead to unnatural shadow removal results, as shown in Figure 7 (b)&(e). By training on our synthesized data with diverse lighting conditions, our structure extraction network (SE-Net) can robustly extract shadow-independent structures, enabling high-quality shadow removal results.", + "bbox": [ + 89, + 441, + 483, + 592 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of our data synthesis strategy. To verify the effectiveness of our data synthesis, we alternatively adopt the portrait shadow synthesis strategy introduced in PSM [56] to create data on FFHQ for training our SE-Net. As shown in Figure 7 (c)&(f) and Table 1, training on the data synthesized by the strategy in PSM fails to exclude the unwanted shadow boundaries and will degrade the subsequent shadow removal performance, because the strategy only involves simple modifications to brightness, saturation, and subsurface scattering of face, lacking the ability to physically simulate the complex facial lighting conditions.", + "bbox": [ + 89, + 599, + 483, + 767 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of structure map on shadow inpainting. We in Figure 8 examine how structure map derived from the SE-Net affects the performance of our diffusion-based shadow inpainting. As shown, omitting the structure map, the facial structure details behind shadows are inevitably modified, producing unpredictable portrait inconsistent to the input.", + "bbox": [ + 89, + 772, + 483, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of detail restoration. We in Figure 9 analyze the", + "bbox": [ + 89, + 885, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/751379de2c98ea584f7029e8f0a94d38dff28013fade88287fd63af57f20468a.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 516, + 343, + 612, + 417 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9b90fbe973c0e75e82a7325002bff7f3cf3cda9d898a676956f07402793f4cef.jpg", + "image_caption": [ + "w/o structure" + ], + "image_footnote": [], + "bbox": [ + 614, + 344, + 707, + 417 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a4ff21630911b15e4c14a20b3fa4a2c574d91193c8f338ca5f4f17bdb1a98eb7.jpg", + "image_caption": [ + "w/ structure" + ], + "image_footnote": [], + "bbox": [ + 710, + 344, + 805, + 417 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/528f54b2eb1d310bc63fad235d93d6b05ca0d6239324ccae90ad3b08f8b2dbc3.jpg", + "image_caption": [ + "Figure 7. Visual comparison of structure maps extracted by different methods and the corresponding shadow removal results. (a) Input. (b)-(d) are structure maps generated by PDG [49], SE-Net trained on data synthesized by the method of PSM [56], and SE-Net trained on our synthesized data. (e)-(g) are shadow removal results produced by our method using the structure maps in (b)-(d), respectively.", + "GT" + ], + "image_footnote": [], + "bbox": [ + 808, + 344, + 903, + 417 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/772faa9f8e537de7ff4127c3843e9a3e0c950a4c5d129ca44110d319ee8e1928.jpg", + "image_caption": [ + "Figure 8. Effect of structure map on shadow inpainting." + ], + "image_footnote": [], + "bbox": [ + 516, + 459, + 643, + 558 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9c3c852d1697744aa9cd343151f373d7cba8248c6aee8ba5766abd428abc52bd.jpg", + "image_caption": [ + "Input", + "Figure 9. Effect of our detail restoration." + ], + "image_footnote": [], + "bbox": [ + 516, + 559, + 643, + 657 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0381bc227f96ed7194a59190d0ca97b7d2c4f8102a0c60058ab5be9ba01e9ece.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 459, + 774, + 558 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4e00420179498b9a608b4ecb451c1a2937276c61330d7dfebfd982d295d9ed61.jpg", + "image_caption": [ + "w/o detail restoration" + ], + "image_footnote": [], + "bbox": [ + 643, + 559, + 774, + 657 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3abaf0203442a7054f5c4fc591b97b4492cfa44b24a120d9d96d0cf67e39361a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 776, + 459, + 903, + 558 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/67fd04bb3c7d907e7aa1b5bd8a2c2872f7504d266e36c915d6e7c7d1f9f1d8f0.jpg", + "image_caption": [ + "w/ detail restoration" + ], + "image_footnote": [], + "bbox": [ + 776, + 559, + 903, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "impact of our detail restoration. As shown, without the restoration step, our shadow removal results exhibit noticeable blurring artifacts around the eyebrows (see 1st row), and fail to faithfully recover the mole in shadow region (see 2nd row). By including the restoration step, we are able to produce high-fidelity shadow removal results with all aforementioned fine-scale details effectively restored, manifesting the effectiveness of our detail restoration.", + "bbox": [ + 511, + 712, + 906, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. More Analysis", + "text_level": 1, + "bbox": [ + 511, + 842, + 661, + 858 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of inaccurate structure map. Structure prediction plays a vital role for the success of our method, because we", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11681", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e600ffbb8d1c8dbd17c755cc3700d7f1245eaf32fe3593914d0e4b6ff222459e.jpg", + "image_caption": [ + "Input", + "Figure 10. Impact of inaccurate structure map." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 187, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fb8a9b934382839fa5c60250a65bbd4d3b904b0431a887e9993ad9f0fa16db5f.jpg", + "image_caption": [ + "Inaccurate map" + ], + "image_footnote": [], + "bbox": [ + 189, + 88, + 287, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a94ea827fc0bc74e1266fc71dff96308146e70590f20f533652deda6cf2a4751.jpg", + "image_caption": [ + "Inpainting result" + ], + "image_footnote": [], + "bbox": [ + 289, + 88, + 383, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9db58f640587a15631f2fdff54b5d9e49361417070f5b6601356288a2e38eced.jpg", + "image_caption": [ + "Final result" + ], + "image_footnote": [], + "bbox": [ + 385, + 88, + 480, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/01c0509b64eded687a734909bd7daa5435908f69c8c3cdeedc1c36e715fcc6a2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 246, + 187, + 321 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/21309f1944af8da8eff70359b4c8a63745c69a94de46723d633d496d049a34f8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 191, + 207, + 285, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/907acb1c6a839ccf02f3d5d5b10018cc7a950cae6bc16a1c6f524e3b5ba5d240.jpg", + "image_caption": [ + "ID: 0.056", + "Figure 11. Effect of inaccurate shadow masks. Our method is highly robust to inaccurate masks (see 3rd and 4th columns)." + ], + "image_footnote": [], + "bbox": [ + 191, + 282, + 285, + 354 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/559e49f4535c14c5f6c57f555325a681197199bc6820f668c56693f5f3fd8b1f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 289, + 207, + 383, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eafd08d435cabd0ca9e9e6cb508f5dfc4b0e523ce56edd2859cc66d5099ada9a.jpg", + "image_caption": [ + "ID:0.065" + ], + "image_footnote": [], + "bbox": [ + 289, + 282, + 383, + 354 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/22f31829e8c49b341c2b0b3e5cc432b5ac9d3e24215b765e19458093f7756a83.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 207, + 480, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/827d3c79f55d6266a2cbcb74d54668921df91accb9b5391dca31898a4a984527.jpg", + "image_caption": [ + "ID:0.062" + ], + "image_footnote": [], + "bbox": [ + 385, + 282, + 480, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "rely on a structure map to guide our diffusion-based shadow inpainting. However, as shown in Figure 10, our method still has some tolerance to inaccurate structure map as the subsequent detail restoration stage can help remove the inpainting artifacts led by inaccurate structure map.", + "bbox": [ + 89, + 426, + 483, + 502 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Robustness to inaccurate shadow masks. Thanks to the design of performing portrait shadow removal via structure-guided inpainting, our method is highly robust to inaccurate input shadow masks, requiring only that the masks cover the shadow regions even if they include large-scale non-shadow regions. As demonstrated in Figure 11, our method achieves both visually indistinguishable shadow removal results and comparable ID preservation scores [48] for all the three masks, including two inaccurate masks in the 3rd and 4th columns. Please see the supplementary material for more results produced with inaccurate shadow masks.", + "bbox": [ + 89, + 508, + 483, + 675 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Different diffusion models for shadow inpainting. To test how different diffusion models affect the performance of our method, we replace our inpainting diffusion model with a strong pre-trained Stable Diffusion (SD) model (version v1-5 inpainting model [36]). To ensure fair comparison, the compared SD model is injected with the same structure map from SE-Net using ControlNet [53]. As can be seen in Figure 12, our trained inpainting diffusion model produces high-fidelity shadow removal results faithful to the inputs, while the SD model incurs noticeable structure tampering.", + "bbox": [ + 89, + 681, + 483, + 834 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of our method on natural images. Our method is also applicable to natural images by replacing the portrait-oriented PDG model [49] with a generic edge detection model and also changing the training dataset. We show", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a986e113c2a5f685b1b9c35642a0fc9ca65a1ab85faa44f8ab7a26088e74441c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 88, + 612, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dd169ab83d7fcccfc797bb9e3e70c75cd432ae5ed8d9a65875d1340c3c78c5ee.jpg", + "image_caption": [ + "Input", + "Figure 12. Effect of leveraging different diffusion models for shadow inpainting. Note, the same structure map as ours is used to produce the shadow inpainting result of \"SD + ControlNet\", and our result here is free of detail restoration." + ], + "image_footnote": [], + "bbox": [ + 516, + 162, + 609, + 237 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b1eb37c0f980142e53e0292c1f2546685b36384ac8e81c5cb5c971f6335b561c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 88, + 709, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/71ac04d28ba94f701a64a6cff401abf1d61a0f481505bb4852e8bcd3761a3b8b.jpg", + "image_caption": [ + "SD + ControlNet" + ], + "image_footnote": [], + "bbox": [ + 612, + 162, + 707, + 237 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/68a4f52e59f7700af2b1d5a2744361ffea3a70ed994d0994c26a7f5371f4bc0a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 88, + 805, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5bbd776822c6f83e1cc6c666a6337e85dde49ec04f16504443e2c4a0db203f46.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 710, + 162, + 805, + 237 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/57769de3d5392ba314900bc24ea4f60029b914d3e5f6b8e84a8b789eedb5abb3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 808, + 88, + 903, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/269a4b717cf02d476b822c2969648a07eab5c4cbfab5003232515bfac0f0e3a3.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 808, + 162, + 903, + 237 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/34f74427323299516f8446abb607c3fe3984e8e2722e48f1751e6563ce69db71.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodUPSRBSRShadowDiffusionIC-lightOurs
Param.135M72M108.79M-185.12M
Time192s11s2.3s32s9.5s
", + "bbox": [ + 517, + 323, + 901, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Comparison on the number of parameters (param.) and inference time. Note, all the inference times here are test for a $512 \\times 512$ portrait on an NVIDIA A100 Tensor Core GPU.", + "bbox": [ + 511, + 380, + 906, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in the supplementary material that our method can produce competitive shadow removal results for natural images.", + "bbox": [ + 511, + 445, + 905, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis on computational complexity. Table 2 compares our inference time with those of other existing portrait specialized or diffusion-based shadow removal methods. As shown, our inference time is more efficient than that of most other methods. Moreover, the number of parameters of our method are comparable to other methods, showing that our superior performance comes mainly from method design.", + "bbox": [ + 511, + 482, + 905, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 609, + 633, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented a portrait shadow removal approach that allows to robustly produce high-fidelity results. Unlike previous methods, we cast portrait shadow removal as diffusion-based inpainting. To do so, we develop a structure extraction network that is robust to illumination changes to estimate shadow-independent structure map for a given portrait image. The structure map is then fed to an inpainting diffusion model to guide the generation of the shadow-free output. Next, we refine the shadow-free output with a detail restoration diffusion model for impairing the fine-scale facial details that may not be captured by the structure map. Extensive experiments show that our method outperforms state-of-the-art methods for portrait shadow removal.", + "bbox": [ + 511, + 636, + 906, + 833 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. This work was supported by the National Natural Science Foundation of China (62471499), the Guangdong Basic and Applied Basic Research Foundation (2023A1515030002).", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11682", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ciprian Corneanu, Raghudeep Gadde, and Aleix M Martinez. Latentpaint: Image inpainting in latent space with diffusion models. In WACV, pages 4334-4343, 2024. 2", + "[2] Antonio Criminisi, Patrick Pérez, and Kentaro Toyama. Region filling and object removal by exemplar-based image inpainting. IEEE Transactions on Image Processing, 13(9): 1200-1212, 2004. 2", + "[3] Xiaodong Cun, Chi-Man Pun, and Cheng Shi. Towards ghost-free shadow removal via dual hierarchical aggregation network and shadow matting gan. In AAAI, pages 10680-10687, 2020. 2", + "[4] Xingbo Dong, Sangrae Cho, Youngsam Kim, Soohyung Kim, and Andrew Beng Jin Teoh. Deep rank hashing network for cancellable face identification. Pattern Recognition, 131:108886, 2022. 1", + "[5] Pawel Drozdowski, Christian Rathgeb, and Christoph Busch. The watchlist imbalance effect in biometric face identification: Comparing theoretical estimates and empiric measurements. In ICCV, pages 3757-3765, 2021. 1", + "[6] Xianyong Fang, Xiaohao He, Linbo Wang, and Jianbing Shen. Robust shadow detection by exploring effective shadow contexts. In ACM MM, pages 2927-2935, 2021. 4", + "[7] Graham D Finlayson, Mark S Drew, and Cheng Lu. Entropy minimization for shadow removal. International Journal of Computer Vision, 85(1):35-57, 2009. 2", + "[8] David Futschik, Kelvin Ritland, James Vecore, Sean Fanello, Sergio Orts-Escalano, Brian Curless, Daniel Sykora, and Rohit Pandey. Controllable light diffusion for portraits. In CVPR, pages 8412-8421, 2023. 2, 6", + "[9] Maciej Gryka, Michael Terry, and Gabriel J Brostow. Learning to remove soft shadows. ACM Transactions on Graphics (TOG), 34(5):1-15, 2015. 2", + "[10] Lanqing Guo, Siyu Huang, Ding Liu, Hao Cheng, and Bihan Wen. Shadowformer: Global context helps shadow removal. In AAAI, pages 710-718, 2023. 2, 5, 6", + "[11] Lanqing Guo, Chong Wang, Wenhan Yang, Siyu Huang, Yufei Wang, Hanspeter Pfister, and Bihan Wen. Shadowdiffusion: When degradation prior meets diffusion model for shadow removal. In CVPR, pages 14049-14058, 2023. 1, 2, 5, 6", + "[12] Ruiqi Guo, Qieyun Dai, and Derek Hoiem. Paired regions for shadow detection and removal. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(12):2956-2967, 2012. 2", + "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, pages 1026-1034, 2015. 5", + "[14] Yingqing He, Yazhou Xing, Tianjia Zhang, and Qifeng Chen. Unsupervised portrait shadow removal via generative priors. In ACM MM, pages 236-244, 2021. 2, 5, 6", + "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 4, 5", + "[16] Andrew Hou, Zhixin Shu, Xuaner Zhang, He Zhang, Yannick Hold-Geoffroy, Jae Shin Yoon, and Xiaoming Liu." + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Compose: Comprehensive portrait shadow editing. In ECCV, pages 356-373. Springer, 2024. 2, 6", + "[17] Xiaowei Hu, Yitong Jiang, Chi-Wing Fu, and Pheng-Ann Heng. Mask-shadowgan: Learning to remove shadows from unpaired data. In ICCV, pages 2472–2481, 2019. 2", + "[18] Satoshi Iizuka, Edgar Simo-Serra, and Hiroshi Ishikawa. Globally and locally consistent image completion. ACM Transactions on Graphics (ToG), 36(4):1-14, 2017. 2", + "[19] Liya Ji, Zhefan Rao, Sinno Jialin Pan, Chenyang Lei, and Qifeng Chen. A diffusion model with state estimation for degradation-blind inverse imaging. In AAAI, pages 2471-2479, 2024. 2", + "[20] Hao Jiang, Qing Zhang, Yongwei Nie, Lei Zhu, and Wei-Shi Zheng. Learning to remove shadows from a single image. International Journal of Computer Vision, 131(9):2471-2488, 2023. 2", + "[21] Yuming Jiang, Ziqi Huang, Xingang Pan, Chen Change Loy, and Ziwei Liu. Talk-to-edit: Fine-grained facial editing via dialog. In ICCV, pages 13799-13808, 2021. 1", + "[22] Yeying Jin, Aashish Sharma, and Robby T Tan. Dcshadownet: Single-image hard and soft shadow removal using unsupervised domain-classifier guided network. In ICCV, pages 5027-5036, 2021. 2", + "[23] Yeying Jin, Wei Ye, Wenhan Yang, Yuan Yuan, and Robby T Tan. Des3: Adaptive attention-driven self and soft shadow removal using vit similarity. In AAAI, pages 2634-2642, 2024. 2", + "[24] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR, pages 8110-8119, 2020. 2", + "[25] Hieu Le and Dimitris Samaras. Shadow removal via shadow image decomposition. In ICCV, pages 8578-8587, 2019. 2", + "[26] Xiaoguang Li, Qing Guo, Rabab Abdelfattah, Di Lin, Wei Feng, Ivor Tsang, and Song Wang. Leveraging inpainting for single-image shadow removal. In ICCV, pages 13055-13064, 2023. 2, 5, 6", + "[27] Zonglin Li, Xiaogian Lv, Wei Yu, Qinglin Liu, Jingbo Lin, and Shengping Zhang. Face shape transfer via semantic warping. Visual Intelligence, 2(1):26, 2024. 1", + "[28] Yaojie Liu, Andrew Z Hou, Xinyu Huang, Liu Ren, and Xiaoming Liu. Blind removal of facial foreign shadows. In BMVC, page 88, 2022. 1, 2, 5, 6", + "[29] Yuhao Liu, Zhanghan Ke, Ke Xu, Fang Liu, Zhenwei Wang, and Rynson WH Lau. Recasting regional lighting for shadow removal. In AAAI, pages 3810-3818, 2024. 2", + "[30] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In ICCV, pages 3730-3738, 2015. 3, 5", + "[31] Zhihao Liu, Hui Yin, Xinyi Wu, Zhenyao Wu, Yang Mi, and Song Wang. From shadow generation to shadow removal. In CVPR, pages 4927-4936, 2021. 2", + "[32] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, pages 11461-11471, 2022. 2" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11683", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Junfeng Lyu, Zhibo Wang, and Feng Xu. Portrait eyeglasses and shadow removal by leveraging 3d synthetic data. In CVPR, pages 3429-3439, 2022. 2", + "[34] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In CVPR, pages 2536-2544, 2016. 2", + "[35] Liangqiong Qu, Jiandong Tian, Shengfeng He, Yandong Tang, and Rynson WH Lau. Deshadownet: A multi-context embedding deep network for shadow removal. In CVPR, pages 4067-4075, 2017. 2", + "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 8", + "[37] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 conference proceedings, pages 1-10, 2022. 2", + "[38] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713-4726, 2022. 5", + "[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.4", + "[40] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In WACV, pages 2149–2159, 2022. 2", + "[41] Chengrui Wang and Weihong Deng. Representative forgery mining for fake face detection. In CVPR, pages 14923-14932, 2021. 1", + "[42] Jifeng Wang, Xiang Li, and Jian Yang. Stacked conditional generative adversarial networks for jointly learning shadow detection and shadow removal. In CVPR, pages 1788-1797, 2018. 2", + "[43] Joshua Weir, Junhong Zhao, Andrew Chalmers, and Taehyun Rhee. Deep portrait delighting. In ECCV, pages 423-439. Springer, 2022. 2, 6", + "[44] Yonatan Wexler, Eli Shechtman, and Michal Irani. Spacetime completion of video. IEEE Transactions on Pattern Analysis and Machine Intelligence, 29(3):463-476, 2007. 2", + "[45] Jie Xiao, Xueyang Fu, Yurui Zhu, Dong Li, Jie Huang, Kai Zhu, and Zheng-Jun Zha. Homoformer: Homogenized transformer for image shadow removal. In CVPR, pages 25617-25626, 2024. 2, 5, 6", + "[46] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In CVPR, pages 22428-22437, 2023. 2", + "[47] Yanbo Xu, Yueqin Yin, Liming Jiang, Qianyi Wu, Chengyao Zheng, Chen Change Loy, Bo Dai, and Wayne Wu. Transeditor: Transformer-based dual-space gan for highly controllable facial editing. In CVPR, pages 7683-7692, 2022. 1" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Shuai Yang, Liming Jiang, Ziwei Liu, and Chen Change Loy. Pastiche master: Exemplar-based high-resolution portrait style transfer. In CVPR, pages 7693-7702, 2022. 8", + "[49] Ran Yi, Yong-Jin Liu, Yu-Kun Lai, and Paul L Rosin. Unpaired portrait drawing generation via asymmetric cycle mapping. In CVPR, pages 8217-8225, 2020. 3, 7, 8", + "[50] Jae Shin Yoon, Zhixin Shu, Mengwei Ren, Cecilia Zhang, Yannick Hold-Geoffroy, Krishna Kumar Singh, and He Zhang. Generative portrait shadow removal. ACM Transactions on Graphics (TOG), 43(6):1-13, 2024. 2, 6", + "[51] Jiahui Yu, Zhe Lin, Jamei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Generative image inpainting with contextual attention. In CVPR, pages 5505-5514, 2018. 2", + "[52] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Scaling in-the-wild training for diffusion-based illumination harmonization and editing by imposing consistent light transport. In ICLR. 1, 2, 5, 6", + "[53] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836-3847, 2023. 8", + "[54] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 3", + "[55] Shifeng Zhang, Cheng Chi, Zhen Lei, and Stan Z Li. Refineface: Refinement neural network for high performance face detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(11):4008-4020, 2020. 1", + "[56] Xuaner Zhang, Jonathan T Barron, Yun-Ta Tsai, Rohit Pandey, Xiuming Zhang, Ren Ng, and David E Jacobs. Portrait shadow manipulation. ACM Transactions on Graphics (TOG), 39(4):78-1, 2020. 2, 3, 5, 6, 7", + "[57] Rongjia Zheng, Qing Zhang, Yongwei Nie, and Wei-Shi Zheng. When shadow removal meets intrinsic image decomposition: A joint learning framework using unpaired data. In AAAI, pages 10599-10607, 2025. 2", + "[58] Hao Zhou, Sunil Hadap, Kalyan Sunkavalli, and David W Jacobs. Deep single-image portrait relighting. In ICCV, pages 7194-7202, 2019. 3", + "[59] Shipeng Zhu, Pengfei Fang, Chenjie Zhu, Zuoyan Zhao, Qiang Xu, and Hui Xue. Text image inpainting via global structure-guided diffusion models. In AAAI, pages 7775-7783, 2024. 2", + "[60] Yurui Zhu, Jie Huang, Xueyang Fu, Feng Zhao, Qibin Sun, and Zheng-Jun Zha. Bijective mapping network for shadow removal. In CVPR, pages 5627-5636, 2022. 2", + "[61] Yurui Zhu, Zeyu Xiao, Yanchi Fang, Xueyang Fu, Zhiwei Xiong, and Zheng-Jun Zha. Efficient model-driven network for shadow removal. In AAAI, pages 3635-3643, 2022. 2" + ], + "bbox": [ + 516, + 92, + 903, + 787 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11684", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Structure-Guided Diffusion Models for High-Fidelity Portrait Shadow Removal/73c9755b-db1d-47f9-8857-79c13b240555_model.json b/2025/Structure-Guided Diffusion Models for High-Fidelity Portrait Shadow Removal/73c9755b-db1d-47f9-8857-79c13b240555_model.json new file mode 100644 index 0000000000000000000000000000000000000000..340a0b373d092420b67e0e60ed4c0355dbab75ec --- /dev/null +++ b/2025/Structure-Guided Diffusion Models for High-Fidelity Portrait Shadow Removal/73c9755b-db1d-47f9-8857-79c13b240555_model.json @@ -0,0 +1,2904 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.098, + 0.131, + 0.9, + 0.152 + ], + "angle": 0, + "content": "Structure-Guided Diffusion Models for High-Fidelity Portrait Shadow Removal" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.179, + 0.786, + 0.2 + ], + "angle": 0, + "content": "Wanchang Yu\\(^{1}\\) Qing Zhang\\(^{1,2*}\\) Rongjia Zheng\\(^{1}\\) Wei-Shi Zheng\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.199, + 0.806, + 0.217 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer Science and Engineering, Sun Yat-sen University, China" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.217, + 0.886, + 0.235 + ], + "angle": 0, + "content": "\\(^{2}\\)Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.253, + 0.229, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.464, + 0.177, + 0.476 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.253, + 0.363, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.239, + 0.464, + 0.353, + 0.476 + ], + "angle": 0, + "content": "ShadowDiffusion [11]" + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.253, + 0.498, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.397, + 0.464, + 0.466, + 0.476 + ], + "angle": 0, + "content": "IC-light [52]" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.253, + 0.635, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.464, + 0.593, + 0.476 + ], + "angle": 0, + "content": "BSR [28]" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.253, + 0.77, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.688, + 0.464, + 0.716, + 0.475 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.253, + 0.906, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.828, + 0.464, + 0.847, + 0.475 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.489, + 0.907, + 0.517 + ], + "angle": 0, + "content": "Figure 1. Comparison on portrait shadow removal. Our method allows to robustly remove facial cast shadows while faithfully preserving the portrait identities as well as structure details in shadow regions and also ensuring the overall naturalness." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.53, + 0.327, + 0.545 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.559, + 0.486, + 0.877 + ], + "angle": 0, + "content": "We present a diffusion-based portrait shadow removal approach that can robustly produce high-fidelity results. Unlike previous methods, we cast shadow removal as diffusion-based inpainting. To this end, we first train a shadow-independent structure extraction network on a real-world portrait dataset with various synthetic lighting conditions, which allows to generate a shadow-independent structure map including facial details while excluding the unwanted shadow boundaries. The structure map is then used as condition to train a structure-guided inpainting diffusion model for removing shadows in a generative manner. Finally, to restore the fine-scale details (e.g., eyelashes, moles and spots) that may not be captured by the structure map, we take the gradients inside the shadow regions as guidance and train a detail restoration diffusion model to refine the shadow removal result. Extensive experiments on the benchmark datasets show that our method clearly outperforms existing methods, and is effective to avoid previously common issues such as facial identity tampering, shadow residual, color distortion, structure blurring, and loss of details. Our code is avail" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.531, + 0.907, + 0.562 + ], + "angle": 0, + "content": "able at https://github.com/wanchang-yu/Structure-Guided-Diffusion-for-Portrait-Shadow-Removal." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.588, + 0.646, + 0.604 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.613, + 0.907, + 0.78 + ], + "angle": 0, + "content": "Taking selfies is very popular nowadays due to the readily-available cameras on various devices, particularly the mobile phones. However, portrait photos taken in the wild lacking professional lighting conditions often suffer from undesired foreign shadows cast by external objects (see Figure 1 for examples). The presence of facial shadows not only severely degrade the overall visual appeal of the portraits, but also adversely affect many downstream vision tasks, such as face detection and recognition [4, 5, 41, 55], and portrait image editing [21, 27, 47]. Hence, portrait shadow removal techniques are widely required." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.781, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Currently, portrait shadow removal remains a challenge despite the remarkable progress in natural image shadow removal. The reason behind is twofold. First of all, due to factors such as privacy and difficulty in data collection, large-scale paired data for portrait shadow is extremely scarce, which significantly limits the development of portrait shadow removal methods, especially learning-based solutions. On the other hand, as this task is portrait-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.092, + 0.887, + 0.383, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author (zhangq93@mail.sysu.edu.cn)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11675" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "oriented, it has very low tolerance for visual artifacts, even minor imperfections on facial color and details could make the final result unnatural and unacceptable. Note, although existing software allow users to interactively adjust photos, it is still rather tedious and difficult for non-experts to manually produce high-fidelity portrait shadow removal results." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.185, + 0.483, + 0.501 + ], + "angle": 0, + "content": "Existing shadow removal methods [11, 26, 45] mainly target natural images due to the availability of large-scale paired datasets. However, these methods basically fail to deal with portrait shadows, as shown in Figures 5 and 6. As there is no accessible large-scale paired datasets for portrait shadow removal, current methods in this field focus primarily on constructing synthetic paired data to enable supervised learning of shadow removal [28, 33, 52, 56]. Suffering from the domain gap between real-world portraits and synthetic data, these methods typically do not work well for real-world portraits with complex shadows. To avoid the domain gap issue led by synthetic training data, a zero-shot solution is introduced in [14], which proposed to leverage the generative prior from pre-trained StyleGAN2 [24] to recover shadow-free portrait through test-time GAN inversion. However, this method tends to modify face identity and incur unnatural results. Note, although recent image relighting techniques [8, 16, 43, 50] have shown remarkable progress, they struggle to enable high-fidelity portrait shadow removal due to the inherent difficulty in estimating ambient light and facial geometry from a single image." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.504, + 0.483, + 0.82 + ], + "angle": 0, + "content": "To robustly produce high-fidelity portrait shadow removal results, we in this work present a diffusion-based approach. Particularly, we propose to cast portrait shadow removal as diffusion-based inpainting, so as to take full advantage of the strong generation capability of diffusion models. To this end, we first train a shadow-independent structure extraction network (SE-Net) on a dataset of paired real-world portraits with various synthetic lighting conditions, from which we obtain a structure map containing facial details while excluding the shadow boundaries. With the structure map, we then train a structure-guided inpainting diffusion model by enforcing reconstruction of randomly masked shadow-free portraits. Finally, to restore the fine-scale facial details that may not be captured by the structure map, we further refine the shadow removal result by a detail restoration diffusion model under the guidance of the gradients in the original shadow regions. Note, although our method requires a test shadow mask as input, it is highly robust to inaccurate masks, requiring only that the mask covers the shadow regions, even if it includes nonshadow regions. The major contributions of this work are:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.826, + 0.482, + 0.87 + ], + "angle": 0, + "content": "- We present a simple yet effective unsupervised approach that addresses portrait shadow removal from a perspective of diffusion-based inpainting." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.482, + 0.901 + ], + "angle": 0, + "content": "- We introduce a relighting-based data synthesis strategy for constructing the training data, and develop a shadow" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.826, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.092, + 0.804, + 0.106 + ], + "angle": 0, + "content": "independent structure extraction network." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.107, + 0.905, + 0.136 + ], + "angle": 0, + "content": "- We show that our method outperforms state-of-the-art methods for portrait shadow removal." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.15, + 0.655, + 0.166 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.177, + 0.905, + 0.555 + ], + "angle": 0, + "content": "Shadow removal. Early methods typically rely on various handcrafted priors or assumptions to work [7, 9, 12]. Recent methods are basically deep learning based [3, 10, 11, 17, 20, 22, 23, 25, 29, 31, 57, 60, 61], benefiting mostly from the availability of large-scale paired datasets such as ISTD [42] and SRD [35]. However, as these methods are tailored for natural images, they are not so effective in dealing with portraits. In contrast to natural images, portrait shadow removal has received much less research attention, due to the lack of large-scale paired datasets. To overcome the data scarcity issue, some methods [28, 56] proposed to synthesize paired shadow and shadow-free portrait images for training. As their data synthesis strategies involve only simple adjustments on brightness and saturation, without considering facial geometry and albedo, their methods typically do not work well for real-world portraits. He et al. [14] achieved zero-shot portrait shadow removal by utilizing generative prior from StyleGAN2, but this method may result in undesired modifications to facial details and thus fails to preserve the face identity. More recently, image relighting is also used for shadow removal [8, 16, 43, 50, 52]. However, this method tends to induce unnatural results because it is very challenging to obtain accurate prediction of diffuse lighting and albedo from a single portrait image." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.56, + 0.907, + 0.802 + ], + "angle": 0, + "content": "Image inpainting. Before the era of deep learning, patch-based self-similarity of natural images has long been the foundation for image inpainting [2, 44]. The emergence of deep learning has then led to great progress in image inpainting [18, 34, 40, 51], as it is quite easy to construct paired training data. Recently, diffusion model has been a preferable choice for image inpainting [1, 19, 32, 37, 46, 59], due to its impressive generation capability. Note, naive applying diffusion models to portrait shadow removal fail to produce visually natural results with unchanged portrait identity. A recent work closely related to our method also shows that it is feasible to remove shadows via image inpainting [26]. However, it does not work well for portrait shadow removal (see Figures 5 and 6), even we fine-tuned their model on real-world paired shadow and shadow-free portrait images provided in [56]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.815, + 0.644, + 0.831 + ], + "angle": 0, + "content": "3. Our Method" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Given a portrait image with cast shadows, our method aims to remove the facial shadows while faithfully preserving the facial details in the shadow regions. Figure 2 presents the overall inference pipeline of our method." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11676" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.09, + 0.904, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.281, + 0.908, + 0.337 + ], + "angle": 0, + "content": "Figure 2. Inference pipeline of our method. Given a input portrait image and the corresponding shadow mask, we first feed the input image to the shadow-independent structure extraction network (SE-Net) to get a structure map. Then, we perform shadow removal by inpainting the shadow regions with a diffusion model conditioned on the structure map. Finally, the shadow removal output is refined by a detail restoration diffusion model to restore the fine-scale facial details that may not be captured by the strcutre map." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.362, + 0.481, + 0.554 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.558, + 0.483, + 0.601 + ], + "angle": 0, + "content": "Figure 3. Training pipeline of SE-Net. Note, we adopt a trained PDG model for structure extraction from [49], while it is not invariant to illumination changes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.629, + 0.455, + 0.644 + ], + "angle": 0, + "content": "3.1. Shadow-Independent Structure Extraction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.652, + 0.483, + 0.788 + ], + "angle": 0, + "content": "As shown in Figure 2, the key to obtaining high-quality portrait shadow removal through inpainting is to acquire a structure map free of shadow boundaries for guiding the inpainting process. However, although existing structure/edge extraction methods are quite effective in identifying edges and structures, they are sensitive to illumination discontinuities and will inevitably include shadow boundaries in their results (see Figure 7). Therefore, we propose to build a shadow-independent structure extraction network." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Training data synthesis. To enable effective training of the shadow-independent structure extraction network, as shown in Figure 3, we construct a synthetic paired dataset of well-lit real-world portraits without foreign shadows on faces and the corresponding relighted portraits with various illumination discontinuities, based on the shadow-free portraits in the CelebA dataset [30]. Specifically, for a well-lit portrait" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.363, + 0.907, + 0.425 + ], + "angle": 0, + "content": "\\(I\\), we first apply an off-the-shelf physics-based relighting method [58] over it to get a relighted image \\(I_{relit}\\), which is then combined with the original portrait \\(I\\) to produce a synthetic image \\(I_{syn}\\) with illumination discontinuities by" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.439, + 0.907, + 0.457 + ], + "angle": 0, + "content": "\\[\nI _ {s y n} = M \\odot I _ {r e l i t} + (1 - M) \\odot I, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.468, + 0.906, + 0.545 + ], + "angle": 0, + "content": "where \\(M\\) is a random facial mask, and \\(\\odot\\) denotes elementwise multiplication. Note, our synthetic data is only used for training the SE-Net and the subsequent detail resotration diffusion model, rather than for directly training shadow removal model as done in [56]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.907, + 0.657 + ], + "angle": 0, + "content": "Training losses. As shown in Figure 3, we adopt a trained structure extraction model from [49] (referred to as PDG in paper) to produce structure map for the original well-lit portrait, which is treated as pseudo ground-truth for supervising the training of our SE-Net. By respectively denoting our SE-Net and the PDG model as \\( G_{s} \\) and \\( G_{p} \\), the training losses of our SE-Net are defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.668, + 0.905, + 0.71 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\mathcal {L} _ {\\text {r e c}} = \\| G _ {s} (I _ {\\text {s y n}}) - G _ {p} (I) \\| _ {1}, \\\\ \\mathcal {L} _ {\\text {p e r c e p t u a l}} = \\mathcal {L} _ {\\text {L P I P S}} \\left(G _ {s} (I _ {\\text {s y n}}), G _ {p} (I)\\right), \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.827 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{LPIPS}\\) refers to the learned perceptual image patch similarity [54]. Besides, we also utilize a discriminator to determine whether the output of SE-Net belongs to the structure image domain, and further define an adversarial loss \\(\\mathcal{L}_{GAN}\\) (see supplementary material for details). In summary, the overall loss function for training the SE-Net is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.841, + 0.905, + 0.858 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {r e c}} + \\lambda_ {1} \\mathcal {L} _ {\\text {p e r c e p t u a l}} + \\lambda_ {2} \\mathcal {L} _ {\\text {G A N}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where we empirically set \\(\\lambda_1 = 0.5\\), \\(\\lambda_1 = 0.25\\), which produce good results, as demonstrated in Figure 7." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11677" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.162, + 0.094, + 0.403, + 0.109 + ], + "angle": 0, + "content": "Structure-guided Portrait Inpainting" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.117, + 0.468, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.094, + 0.804, + 0.109 + ], + "angle": 0, + "content": "Gradient-guided Detail Restoration" + }, + { + "type": "image", + "bbox": [ + 0.482, + 0.114, + 0.897, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.298, + 0.908, + 0.327 + ], + "angle": 0, + "content": "Figure 4. Training pipeline of our diffusion models. Here we introduce how our structure-guided portrait inpainting diffusion model and gradient-guided detail restoration diffusion model are trained." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.352, + 0.416, + 0.369 + ], + "angle": 0, + "content": "3.2. Structure-Guided Portrait Inpainting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.374, + 0.484, + 0.42 + ], + "angle": 0, + "content": "As shown in Figure 2, with structure map, the next step is to inpaint the shadow regions. To this end, we develop a structure-guided portrait inpainting diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.426, + 0.484, + 0.606 + ], + "angle": 0, + "content": "Structure-aware denoising. As shown in Figure 4, we start with a shadow-free portrait as input and then randomly mask some facial regions to create a masked input, followed by a diffusion model to reconstruct the original input by inpainting the masked regions. The inpainting is rooted in the forward diffusion process, where the diffusion model learns a Markovian chain by progressively corrupting the target image \\( x_0 \\) (namely the masked input) with Gaussian noise at each timestep \\( t \\), such that it is transformed into a Gaussian distribution (referred to as \\( \\mathcal{N} \\) in the following). Specifically, after reparameterization [15], our forward process obtains \\( x_t \\) from \\( x_0 \\) in a single step by:" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.612, + 0.482, + 0.629 + ], + "angle": 0, + "content": "\\[\nx _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.634, + 0.483, + 0.679 + ], + "angle": 0, + "content": "where \\(\\bar{\\alpha}_t = \\prod_{i=0}^t \\alpha_i\\) is the cumulative product of forward noise weight parameters \\(\\alpha_i = 1 - \\beta_i\\), while \\(\\epsilon \\sim \\mathcal{N}(0, I)\\) denotes random Gaussian noise." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.681, + 0.484, + 0.77 + ], + "angle": 0, + "content": "As for the reverse denoising process, we use the sampling strategy of DDIM [39] for acceleration. Note, unlike unconditional image generation, our goal is to faithfully reconstruct the original unmasked portrait, the reverse denoising process is thus conditioned on a structure map \\( S \\) and defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.773, + 0.482, + 0.818 + ], + "angle": 0, + "content": "\\[\nx _ {t - 1} = \\sqrt {\\bar {\\alpha} _ {t - 1}} \\left(\\frac {x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\cdot \\mathbf {e} _ {t}}{\\sqrt {\\bar {\\alpha} _ {t}}}\\right) + \\sqrt {1 - \\bar {\\alpha} _ {t - 1}} \\cdot \\mathbf {e} _ {t}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.819, + 0.483, + 0.88 + ], + "angle": 0, + "content": "where \\(\\mathbf{e}_t = \\epsilon_\\theta (x_t,I_M,S,M,t)\\) refers to the noise estimated by a deep neural network \\(\\epsilon_{\\theta}\\), where \\(I_M\\) is the masked input created from a random mask \\(M\\). Following [15], the training objective for diffusion model can be simplified as:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.885, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i f f}} (\\epsilon) = \\| \\epsilon - \\epsilon_ {\\theta} \\left(x _ {t}, I _ {M}, S, M, t\\right) \\| _ {2} ^ {2}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.352, + 0.906, + 0.429 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) is the ground truth noise, while \\(\\epsilon_{\\theta}(x_t,I_M,S,M,t)\\) is the noise predicted by the network. As shown in Figure 2, a facial shadow mask is required at the inference stage, which is a dilated version of the facial shadow detection result from the trained model of [6]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.443, + 0.907, + 0.61 + ], + "angle": 0, + "content": "Shadow mask refinement. To enhance the robustness of our method to inaccurate shadow mask, we propose to refine the mask based on the shadow removal result generated by the inpainting diffusion model. Specifically, we first compute the absolute difference between the input image and the shadow removal result, and then perform Otsu-based thresholding to get a refined shadow mask excluding all the non-shadow regions. With the refined mask, we further update the shadow removal result by combining it with the input image, so as to ensure that non-shadow regions in the result are unchanged." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.644, + 0.828, + 0.659 + ], + "angle": 0, + "content": "3.3. Gradient-Guided Detail Restoration" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "As the original PDG model is designed for generating portrait drawing formed by large-scale structures, thus our SE-Net is similarly not so effective in capturing fine-scale portrait details, such as eyelashes, moles, and spots (see the bottom row in Figure 7 for example). To address this issue, we design a detail restoration diffusion model to refine the shadow removal result generated by the inpainting diffusion model, where we directly take the gradients in the shadow regions as guidance to restore the fine-scale details. Figure 4 describes how we train the detail restoration diffusion model. Note, gradient-aware denoising process is similar to the structure-aware denoising described above, and the two share the same forward diffusion and reverse denoising formulations described in Equations 4 and 5, with the latter conditioned on gradient map rather than structure map." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11678" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.905, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.163, + 0.415, + 0.833, + 0.43 + ], + "angle": 0, + "content": "Figure 5. Visual comparison with state-of-the-art methods on test images from the dataset provided in [56]." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.438, + 0.907, + 0.684 + ], + "angle": 0, + "content": "
MethodAllShadow regionNon-shadow region
SSIM↑LPIPS↓RMSE↓SSIM↑LPIPS↓RMSE↓SSIM↑LPIPS↓RMSE↓
ShadowDiffusion [11]0.6500.17738.2460.9010.03715.5610.5790.22352.276
Inpaint4Shadow [26]0.7660.09422.7090.9100.03317.2500.7080.16445.456
ShadowFormer [10]0.7720.08621.4130.9260.02715.2880.7230.14942.341
HomoFormer [45]0.7860.10021.7240.9130.04116.2480.7110.16742.950
PSM [56]0.7820.074-------
UPSR [14]0.7310.10926.9150.9000.03920.6590.6680.18845.391
BSR [28]0.6050.11828.5320.9100.03615.1930.5480.17143.627
IC-light [52]0.5140.27862.5390.9160.05733.0610.4780.31969.208
Ours w/o structure guidance0.7570.10123.5980.8980.04318.5140.7140.13839.000
Ours w/ PDG predicted structure0.7790.09219.9630.9190.03514.0620.7250.13137.387
Ours w/ data synthesis in PSM0.7850.08619.7970.9250.03213.7520.7290.13638.485
Ours w/o detail-restoration0.8140.06218.5370.9430.02012.7480.7470.12036.326
Ours (full method)0.8300.05617.1620.9730.01110.1980.7780.09134.573
" + }, + { + "type": "table_caption", + "bbox": [ + 0.179, + 0.687, + 0.817, + 0.702 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison on the real-world portrait shadow removal dataset provided in [56]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.733, + 0.411, + 0.75 + ], + "angle": 0, + "content": "3.4. Implementation and Training Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.757, + 0.483, + 0.878 + ], + "angle": 0, + "content": "Network architecture. The SE-Net is an encoder-decoder network, which starts with a flat convolution and two down convolution blocks, followed by nine residual blocks, and finally two up convolution blocks and a final convolution layer. Please see also the supplementary material for visual illustration of the architecture of SE-Net. As for the inpainting diffusion model and the detail-restoration diffusion model, we follow [15] to adopt the same U-Net architecture." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.885, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Training details. To train SE-Net, we synthesize 20,000" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.907, + 0.901 + ], + "angle": 0, + "content": "portrait images with diverse lighting conditions based on the CelebA dataset [30], and train it for 3 epochs with a batch size of 8 and a learning rate of \\(1.5 \\times 10^{-5}\\). As for the inpainting diffusion model, we train it on the FFHQ dataset for 300 epochs with a batch size of 32, and train the detail restoration diffusion model on the same synthetic dataset as the SE-Net for 200 epochs. We build our model on PyTorch and use the Adam optimizer with momentum parameters of (0.9, 0.999) and an initial learning rate of \\(1 \\times 10^{-4}\\). Following [38], we employ the Kaiming initialization [13] to initialize the weights of our model and apply 0.9999 Expo" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11679" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.905, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.185, + 0.411, + 0.812, + 0.426 + ], + "angle": 0, + "content": "Figure 6. Visual comparison with state-of-the-art methods on test images from our collected dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.452, + 0.484, + 0.529 + ], + "angle": 0, + "content": "nential Moving Average (EMA) in all our experiments. We use 2000 diffusion steps and a noise schedule \\(\\beta_{t}\\) linearly increasing from \\(1\\times 10^{-6}\\) to \\(1\\times 10^{-2}\\) for training. During inference, we perform both the inpainting diffusion and detail restoration diffusion for 40 denoising steps." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.545, + 0.223, + 0.562 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.571, + 0.484, + 0.722 + ], + "angle": 0, + "content": "Datasets. Following previous portrait shadow removal approaches [14, 28], we evaluate our method on the only publicly available real-world paired dataset provided in PSM [56], which contains 9 subjects and 100 shadowed portrait images in varying poses, shadow shapes, illumination conditions, and shadow types. To better examine the effectiveness of our method, we additionally construct a dataset of 200 portraits with diverse facial shadows. Note, we only conduct visual comparison on our collected dataset due to the lack of ground truth shadow removal results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.73, + 0.484, + 0.821 + ], + "angle": 0, + "content": "Evaluation metrics. We use three commonly-used metrics to quantitatively evaluate the performance of our method, including SSIM, LPIPS, and the root mean square error (RMSE) computed in the LAB color space for measuring the similarity between the predicted results and the corresponding ground-truth shadow-free images." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.833, + 0.455, + 0.849 + ], + "angle": 0, + "content": "4.1. Comparison with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Baselines. We compare our method with various state-of-the-art shadow removal methods, including four portrait shadow removal and editing methods, i.e., PSM [56]," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.452, + 0.907, + 0.59 + ], + "angle": 0, + "content": "UPSR [14], and BSR [28], IC-light [52], as well as four natural image shadow removal methods, i.e., ShadowDiffusion [11], Inpaint4Shadow [26], ShadowFormer [10], and HomoFormer [45]. For fair comparison, we produce their results using publicly-available implementation or trained models provided by the authors with recommended parameter setting. Note that, we do not compare with other recent relighting-based shadow editing methods [8, 16, 43, 50], as their codes and trained models are not yet released." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.601, + 0.909, + 0.813 + ], + "angle": 0, + "content": "Visual comparison. We in Figures 5 and 6 conduct visual comparison between our method and previous shadow removal methods. As shown, shadow removal methods designed for natural images, i.e., Inpaint4Shadow [26], HomoFormer [45], and ShadowDiffusion [11] are not well-suited to portraits and tend to induce obvious shadow residuals and appearance distortions. As for the other three portrait shadow removal methods, UPSR [14] and BSR [28] fail to preserve the facial details and may also lead to shadow residuals and color distortions, while IC-Light [52] struggles to remove the facial shadows due to the inherent difficulty in estimating the ambient light and facial geometry. Please see also the supplementary material for more visual comparison results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Quantitative comparison. In Table 1, we quantitatively compare our method with other methods in terms of the SSIM, LPIPS, and RMSE metrics. As can be seen, our method outperforms all the compared methods on all the three metrics, demonstrating its effectiveness and superior-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11680" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.088, + 0.208, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.178, + 0.208, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.142, + 0.266, + 0.16, + 0.275 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.21, + 0.089, + 0.324, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.21, + 0.178, + 0.325, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.267, + 0.276, + 0.275 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.089, + 0.441, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.178, + 0.441, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.267, + 0.391, + 0.275 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.442, + 0.089, + 0.558, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.442, + 0.178, + 0.558, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.49, + 0.267, + 0.508, + 0.275 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.089, + 0.672, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.178, + 0.672, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.267, + 0.624, + 0.275 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.089, + 0.788, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.178, + 0.788, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.267, + 0.74, + 0.275 + ], + "angle": 0, + "content": "(1)" + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.089, + 0.905, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.178, + 0.905, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.84, + 0.267, + 0.856, + 0.275 + ], + "angle": 0, + "content": "(g)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.278, + 0.907, + 0.321 + ], + "angle": 0, + "content": "Figure 7. Visual comparison of structure maps extracted by different methods and the corresponding shadow removal results. (a) Input. (b)-(d) are structure maps generated by PDG [49], SE-Net trained on data synthesized by the method of PSM [56], and SE-Net trained on our synthesized data. (e)-(g) are shadow removal results produced by our method using the structure maps in (b)-(d), respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.346, + 0.483, + 0.408 + ], + "angle": 0, + "content": "ity in portrait shadow removal. Note, as PSM [56] provides only code for synthesizing training data, without releasing the training code and trained model, its quantitative results in Table 1 are as reported in their paper." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.419, + 0.245, + 0.435 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.442, + 0.484, + 0.593 + ], + "angle": 0, + "content": "Effectiveness of our shadow-independent structure extraction. Although the trained model of PDG [49] is utilized to train our structure extraction, the structure map it produces is not what we expected because it will include the shadow boundaries and lead to unnatural shadow removal results, as shown in Figure 7 (b)&(e). By training on our synthesized data with diverse lighting conditions, our structure extraction network (SE-Net) can robustly extract shadow-independent structures, enabling high-quality shadow removal results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.484, + 0.768 + ], + "angle": 0, + "content": "Effectiveness of our data synthesis strategy. To verify the effectiveness of our data synthesis, we alternatively adopt the portrait shadow synthesis strategy introduced in PSM [56] to create data on FFHQ for training our SE-Net. As shown in Figure 7 (c)&(f) and Table 1, training on the data synthesized by the strategy in PSM fails to exclude the unwanted shadow boundaries and will degrade the subsequent shadow removal performance, because the strategy only involves simple modifications to brightness, saturation, and subsurface scattering of face, lacking the ability to physically simulate the complex facial lighting conditions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.773, + 0.484, + 0.865 + ], + "angle": 0, + "content": "Effect of structure map on shadow inpainting. We in Figure 8 examine how structure map derived from the SE-Net affects the performance of our diffusion-based shadow inpainting. As shown, omitting the structure map, the facial structure details behind shadows are inevitably modified, producing unpredictable portrait inconsistent to the input." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Effect of detail restoration. We in Figure 9 analyze the" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.344, + 0.613, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.419, + 0.58, + 0.429 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.345, + 0.709, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.626, + 0.419, + 0.698, + 0.428 + ], + "angle": 0, + "content": "w/o structure" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.345, + 0.807, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.728, + 0.419, + 0.792, + 0.428 + ], + "angle": 0, + "content": "w/ structure" + }, + { + "type": "image", + "bbox": [ + 0.809, + 0.345, + 0.905, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.847, + 0.419, + 0.867, + 0.428 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.433, + 0.885, + 0.447 + ], + "angle": 0, + "content": "Figure 8. Effect of structure map on shadow inpainting." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.46, + 0.645, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.56, + 0.645, + 0.659 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.566, + 0.66, + 0.595, + 0.669 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.46, + 0.776, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.56, + 0.776, + 0.659 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.654, + 0.66, + 0.767, + 0.668 + ], + "angle": 0, + "content": "w/o detail restoration" + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.46, + 0.905, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.56, + 0.905, + 0.659 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.66, + 0.894, + 0.668 + ], + "angle": 0, + "content": "w/ detail restoration" + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.673, + 0.838, + 0.687 + ], + "angle": 0, + "content": "Figure 9. Effect of our detail restoration." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.713, + 0.907, + 0.835 + ], + "angle": 0, + "content": "impact of our detail restoration. As shown, without the restoration step, our shadow removal results exhibit noticeable blurring artifacts around the eyebrows (see 1st row), and fail to faithfully recover the mole in shadow region (see 2nd row). By including the restoration step, we are able to produce high-fidelity shadow removal results with all aforementioned fine-scale details effectively restored, manifesting the effectiveness of our detail restoration." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.843, + 0.663, + 0.859 + ], + "angle": 0, + "content": "4.3. More Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Impact of inaccurate structure map. Structure prediction plays a vital role for the success of our method, because we" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11681" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.189, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.164, + 0.157, + 0.175 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.191, + 0.089, + 0.288, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.198, + 0.164, + 0.28, + 0.175 + ], + "angle": 0, + "content": "Inaccurate map" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.089, + 0.385, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.164, + 0.381, + 0.175 + ], + "angle": 0, + "content": "Inpainting result" + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.089, + 0.482, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.404, + 0.164, + 0.465, + 0.175 + ], + "angle": 0, + "content": "Final result" + }, + { + "type": "image_caption", + "bbox": [ + 0.14, + 0.178, + 0.435, + 0.193 + ], + "angle": 0, + "content": "Figure 10. Impact of inaccurate structure map." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.247, + 0.189, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.208, + 0.287, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.283, + 0.286, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.212, + 0.357, + 0.266, + 0.367 + ], + "angle": 0, + "content": "ID: 0.056" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.208, + 0.385, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.283, + 0.384, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.357, + 0.362, + 0.366 + ], + "angle": 0, + "content": "ID:0.065" + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.208, + 0.482, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.283, + 0.482, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.357, + 0.461, + 0.366 + ], + "angle": 0, + "content": "ID:0.062" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.371, + 0.483, + 0.399 + ], + "angle": 0, + "content": "Figure 11. Effect of inaccurate shadow masks. Our method is highly robust to inaccurate masks (see 3rd and 4th columns)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.427, + 0.484, + 0.503 + ], + "angle": 0, + "content": "rely on a structure map to guide our diffusion-based shadow inpainting. However, as shown in Figure 10, our method still has some tolerance to inaccurate structure map as the subsequent detail restoration stage can help remove the inpainting artifacts led by inaccurate structure map." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.51, + 0.484, + 0.676 + ], + "angle": 0, + "content": "Robustness to inaccurate shadow masks. Thanks to the design of performing portrait shadow removal via structure-guided inpainting, our method is highly robust to inaccurate input shadow masks, requiring only that the masks cover the shadow regions even if they include large-scale non-shadow regions. As demonstrated in Figure 11, our method achieves both visually indistinguishable shadow removal results and comparable ID preservation scores [48] for all the three masks, including two inaccurate masks in the 3rd and 4th columns. Please see the supplementary material for more results produced with inaccurate shadow masks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.683, + 0.484, + 0.835 + ], + "angle": 0, + "content": "Different diffusion models for shadow inpainting. To test how different diffusion models affect the performance of our method, we replace our inpainting diffusion model with a strong pre-trained Stable Diffusion (SD) model (version v1-5 inpainting model [36]). To ensure fair comparison, the compared SD model is injected with the same structure map from SE-Net using ControlNet [53]. As can be seen in Figure 12, our trained inpainting diffusion model produces high-fidelity shadow removal results faithful to the inputs, while the SD model incurs noticeable structure tampering." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Effect of our method on natural images. Our method is also applicable to natural images by replacing the portrait-oriented PDG model [49] with a generic edge detection model and also changing the training dataset. We show" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.613, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.164, + 0.611, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.238, + 0.58, + 0.248 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.089, + 0.71, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.164, + 0.709, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.617, + 0.238, + 0.706, + 0.248 + ], + "angle": 0, + "content": "SD + ControlNet" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.089, + 0.807, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.164, + 0.807, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.747, + 0.239, + 0.774, + 0.248 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.809, + 0.089, + 0.905, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.81, + 0.164, + 0.904, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.847, + 0.239, + 0.866, + 0.248 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.252, + 0.907, + 0.307 + ], + "angle": 0, + "content": "Figure 12. Effect of leveraging different diffusion models for shadow inpainting. Note, the same structure map as ours is used to produce the shadow inpainting result of \"SD + ControlNet\", and our result here is free of detail restoration." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.324, + 0.903, + 0.376 + ], + "angle": 0, + "content": "
MethodUPSRBSRShadowDiffusionIC-lightOurs
Param.135M72M108.79M-185.12M
Time192s11s2.3s32s9.5s
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.381, + 0.907, + 0.423 + ], + "angle": 0, + "content": "Table 2. Comparison on the number of parameters (param.) and inference time. Note, all the inference times here are test for a \\(512 \\times 512\\) portrait on an NVIDIA A100 Tensor Core GPU." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.446, + 0.906, + 0.476 + ], + "angle": 0, + "content": "in the supplementary material that our method can produce competitive shadow removal results for natural images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.483, + 0.906, + 0.59 + ], + "angle": 0, + "content": "Analysis on computational complexity. Table 2 compares our inference time with those of other existing portrait specialized or diffusion-based shadow removal methods. As shown, our inference time is more efficient than that of most other methods. Moreover, the number of parameters of our method are comparable to other methods, showing that our superior performance comes mainly from method design." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.611, + 0.634, + 0.627 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.637, + 0.907, + 0.834 + ], + "angle": 0, + "content": "We have presented a portrait shadow removal approach that allows to robustly produce high-fidelity results. Unlike previous methods, we cast portrait shadow removal as diffusion-based inpainting. To do so, we develop a structure extraction network that is robust to illumination changes to estimate shadow-independent structure map for a given portrait image. The structure map is then fed to an inpainting diffusion model to guide the generation of the shadow-free output. Next, we refine the shadow-free output with a detail restoration diffusion model for impairing the fine-scale facial details that may not be captured by the structure map. Extensive experiments show that our method outperforms state-of-the-art methods for portrait shadow removal." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Acknowledgment. This work was supported by the National Natural Science Foundation of China (62471499), the Guangdong Basic and Applied Basic Research Foundation (2023A1515030002)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11682" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.157 + ], + "angle": 0, + "content": "[1] Ciprian Corneanu, Raghudeep Gadde, and Aleix M Martinez. Latentpaint: Image inpainting in latent space with diffusion models. In WACV, pages 4334-4343, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.157, + 0.483, + 0.212 + ], + "angle": 0, + "content": "[2] Antonio Criminisi, Patrick Pérez, and Kentaro Toyama. Region filling and object removal by exemplar-based image inpainting. IEEE Transactions on Image Processing, 13(9): 1200-1212, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.213, + 0.483, + 0.268 + ], + "angle": 0, + "content": "[3] Xiaodong Cun, Chi-Man Pun, and Cheng Shi. Towards ghost-free shadow removal via dual hierarchical aggregation network and shadow matting gan. In AAAI, pages 10680-10687, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.269, + 0.482, + 0.323 + ], + "angle": 0, + "content": "[4] Xingbo Dong, Sangrae Cho, Youngsam Kim, Soohyung Kim, and Andrew Beng Jin Teoh. Deep rank hashing network for cancellable face identification. Pattern Recognition, 131:108886, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.326, + 0.482, + 0.381 + ], + "angle": 0, + "content": "[5] Pawel Drozdowski, Christian Rathgeb, and Christoph Busch. The watchlist imbalance effect in biometric face identification: Comparing theoretical estimates and empiric measurements. In ICCV, pages 3757-3765, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.482, + 0.424 + ], + "angle": 0, + "content": "[6] Xianyong Fang, Xiaohao He, Linbo Wang, and Jianbing Shen. Robust shadow detection by exploring effective shadow contexts. In ACM MM, pages 2927-2935, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.424, + 0.484, + 0.465 + ], + "angle": 0, + "content": "[7] Graham D Finlayson, Mark S Drew, and Cheng Lu. Entropy minimization for shadow removal. International Journal of Computer Vision, 85(1):35-57, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.466, + 0.482, + 0.522 + ], + "angle": 0, + "content": "[8] David Futschik, Kelvin Ritland, James Vecore, Sean Fanello, Sergio Orts-Escalano, Brian Curless, Daniel Sykora, and Rohit Pandey. Controllable light diffusion for portraits. In CVPR, pages 8412-8421, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.523, + 0.482, + 0.564 + ], + "angle": 0, + "content": "[9] Maciej Gryka, Michael Terry, and Gabriel J Brostow. Learning to remove soft shadows. ACM Transactions on Graphics (TOG), 34(5):1-15, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.565, + 0.482, + 0.606 + ], + "angle": 0, + "content": "[10] Lanqing Guo, Siyu Huang, Ding Liu, Hao Cheng, and Bihan Wen. Shadowformer: Global context helps shadow removal. In AAAI, pages 710-718, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.607, + 0.482, + 0.674 + ], + "angle": 0, + "content": "[11] Lanqing Guo, Chong Wang, Wenhan Yang, Siyu Huang, Yufei Wang, Hanspeter Pfister, and Bihan Wen. Shadowdiffusion: When degradation prior meets diffusion model for shadow removal. In CVPR, pages 14049-14058, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.676, + 0.482, + 0.732 + ], + "angle": 0, + "content": "[12] Ruiqi Guo, Qieyun Dai, and Derek Hoiem. Paired regions for shadow detection and removal. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(12):2956-2967, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.733, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, pages 1026-1034, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[14] Yingqing He, Yazhou Xing, Tianjia Zhang, and Qifeng Chen. Unsupervised portrait shadow removal via generative priors. In ACM MM, pages 236-244, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.872, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[16] Andrew Hou, Zhixin Shu, Xuaner Zhang, He Zhang, Yannick Hold-Geoffroy, Jae Shin Yoon, and Xiaoming Liu." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "Compose: Comprehensive portrait shadow editing. In ECCV, pages 356-373. Springer, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.164 + ], + "angle": 0, + "content": "[17] Xiaowei Hu, Yitong Jiang, Chi-Wing Fu, and Pheng-Ann Heng. Mask-shadowgan: Learning to remove shadows from unpaired data. In ICCV, pages 2472–2481, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.166, + 0.905, + 0.208 + ], + "angle": 0, + "content": "[18] Satoshi Iizuka, Edgar Simo-Serra, and Hiroshi Ishikawa. Globally and locally consistent image completion. ACM Transactions on Graphics (ToG), 36(4):1-14, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.21, + 0.905, + 0.264 + ], + "angle": 0, + "content": "[19] Liya Ji, Zhefan Rao, Sinno Jialin Pan, Chenyang Lei, and Qifeng Chen. A diffusion model with state estimation for degradation-blind inverse imaging. In AAAI, pages 2471-2479, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.266, + 0.905, + 0.322 + ], + "angle": 0, + "content": "[20] Hao Jiang, Qing Zhang, Yongwei Nie, Lei Zhu, and Wei-Shi Zheng. Learning to remove shadows from a single image. International Journal of Computer Vision, 131(9):2471-2488, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.324, + 0.905, + 0.367 + ], + "angle": 0, + "content": "[21] Yuming Jiang, Ziqi Huang, Xingang Pan, Chen Change Loy, and Ziwei Liu. Talk-to-edit: Fine-grained facial editing via dialog. In ICCV, pages 13799-13808, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.368, + 0.905, + 0.424 + ], + "angle": 0, + "content": "[22] Yeying Jin, Aashish Sharma, and Robby T Tan. Dcshadownet: Single-image hard and soft shadow removal using unsupervised domain-classifier guided network. In ICCV, pages 5027-5036, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.425, + 0.905, + 0.48 + ], + "angle": 0, + "content": "[23] Yeying Jin, Wei Ye, Wenhan Yang, Yuan Yuan, and Robby T Tan. Des3: Adaptive attention-driven self and soft shadow removal using vit similarity. In AAAI, pages 2634-2642, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.482, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[24] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR, pages 8110-8119, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.569 + ], + "angle": 0, + "content": "[25] Hieu Le and Dimitris Samaras. Shadow removal via shadow image decomposition. In ICCV, pages 8578-8587, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.57, + 0.905, + 0.625 + ], + "angle": 0, + "content": "[26] Xiaoguang Li, Qing Guo, Rabab Abdelfattah, Di Lin, Wei Feng, Ivor Tsang, and Song Wang. Leveraging inpainting for single-image shadow removal. In ICCV, pages 13055-13064, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.627, + 0.905, + 0.669 + ], + "angle": 0, + "content": "[27] Zonglin Li, Xiaogian Lv, Wei Yu, Qinglin Liu, Jingbo Lin, and Shengping Zhang. Face shape transfer via semantic warping. Visual Intelligence, 2(1):26, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.671, + 0.905, + 0.713 + ], + "angle": 0, + "content": "[28] Yaojie Liu, Andrew Z Hou, Xinyu Huang, Liu Ren, and Xiaoming Liu. Blind removal of facial foreign shadows. In BMVC, page 88, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.714, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[29] Yuhao Liu, Zhanghan Ke, Ke Xu, Fang Liu, Zhenwei Wang, and Rynson WH Lau. Recasting regional lighting for shadow removal. In AAAI, pages 3810-3818, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.758, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[30] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In ICCV, pages 3730-3738, 2015. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.801, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[31] Zhihao Liu, Hui Yin, Xinyi Wu, Zhenyao Wu, Yang Mi, and Song Wang. From shadow generation to shadow removal. In CVPR, pages 4927-4936, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[32] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, pages 11461-11471, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11683" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[33] Junfeng Lyu, Zhibo Wang, and Feng Xu. Portrait eyeglasses and shadow removal by leveraging 3d synthetic data. In CVPR, pages 3429-3439, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.482, + 0.189 + ], + "angle": 0, + "content": "[34] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In CVPR, pages 2536-2544, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[35] Liangqiong Qu, Jiandong Tian, Shengfeng He, Yandong Tang, and Rynson WH Lau. Deshadownet: A multi-context embedding deep network for shadow removal. In CVPR, pages 4067-4075, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.301, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[37] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 conference proceedings, pages 1-10, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.371, + 0.482, + 0.44 + ], + "angle": 0, + "content": "[38] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713-4726, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.441, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[40] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In WACV, pages 2149–2159, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.593 + ], + "angle": 0, + "content": "[41] Chengrui Wang and Weihong Deng. Representative forgery mining for fake face detection. In CVPR, pages 14923-14932, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.594, + 0.482, + 0.649 + ], + "angle": 0, + "content": "[42] Jifeng Wang, Xiang Li, and Jian Yang. Stacked conditional generative adversarial networks for jointly learning shadow detection and shadow removal. In CVPR, pages 1788-1797, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.482, + 0.692 + ], + "angle": 0, + "content": "[43] Joshua Weir, Junhong Zhao, Andrew Chalmers, and Taehyun Rhee. Deep portrait delighting. In ECCV, pages 423-439. Springer, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[44] Yonatan Wexler, Eli Shechtman, and Michal Irani. Spacetime completion of video. IEEE Transactions on Pattern Analysis and Machine Intelligence, 29(3):463-476, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[45] Jie Xiao, Xueyang Fu, Yurui Zhu, Dong Li, Jie Huang, Kai Zhu, and Zheng-Jun Zha. Homoformer: Homogenized transformer for image shadow removal. In CVPR, pages 25617-25626, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[46] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In CVPR, pages 22428-22437, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[47] Yanbo Xu, Yueqin Yin, Liming Jiang, Qianyi Wu, Chengyao Zheng, Chen Change Loy, Bo Dai, and Wayne Wu. Transeditor: Transformer-based dual-space gan for highly controllable facial editing. In CVPR, pages 7683-7692, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[48] Shuai Yang, Liming Jiang, Ziwei Liu, and Chen Change Loy. Pastiche master: Exemplar-based high-resolution portrait style transfer. In CVPR, pages 7693-7702, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[49] Ran Yi, Yong-Jin Liu, Yu-Kun Lai, and Paul L Rosin. Unpaired portrait drawing generation via asymmetric cycle mapping. In CVPR, pages 8217-8225, 2020. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[50] Jae Shin Yoon, Zhixin Shu, Mengwei Ren, Cecilia Zhang, Yannick Hold-Geoffroy, Krishna Kumar Singh, and He Zhang. Generative portrait shadow removal. ACM Transactions on Graphics (TOG), 43(6):1-13, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.277 + ], + "angle": 0, + "content": "[51] Jiahui Yu, Zhe Lin, Jamei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Generative image inpainting with contextual attention. In CVPR, pages 5505-5514, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[52] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Scaling in-the-wild training for diffusion-based illumination harmonization and editing by imposing consistent light transport. In ICLR. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.376 + ], + "angle": 0, + "content": "[53] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836-3847, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[54] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[55] Shifeng Zhang, Cheng Chi, Zhen Lei, and Stan Z Li. Refineface: Refinement neural network for high performance face detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(11):4008-4020, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[56] Xuaner Zhang, Jonathan T Barron, Yun-Ta Tsai, Rohit Pandey, Xiuming Zhang, Ren Ng, and David E Jacobs. Portrait shadow manipulation. ACM Transactions on Graphics (TOG), 39(4):78-1, 2020. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[57] Rongjia Zheng, Qing Zhang, Yongwei Nie, and Wei-Shi Zheng. When shadow removal meets intrinsic image decomposition: A joint learning framework using unpaired data. In AAAI, pages 10599-10607, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.603, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[58] Hao Zhou, Sunil Hadap, Kalyan Sunkavalli, and David W Jacobs. Deep single-image portrait relighting. In ICCV, pages 7194-7202, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.645, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[59] Shipeng Zhu, Pengfei Fang, Chenjie Zhu, Zuoyan Zhao, Qiang Xu, and Hui Xue. Text image inpainting via global structure-guided diffusion models. In AAAI, pages 7775-7783, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.702, + 0.905, + 0.745 + ], + "angle": 0, + "content": "[60] Yurui Zhu, Jie Huang, Xueyang Fu, Feng Zhao, Qibin Sun, and Zheng-Jun Zha. Bijective mapping network for shadow removal. In CVPR, pages 5627-5636, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.745, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[61] Yurui Zhu, Zeyu Xiao, Yanchi Fang, Xueyang Fu, Zhiwei Xiong, and Zheng-Jun Zha. Efficient model-driven network for shadow removal. In AAAI, pages 3635-3643, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.788 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11684" + } + ] +] \ No newline at end of file diff --git a/2025/Structure-Guided Diffusion Models for High-Fidelity Portrait Shadow Removal/73c9755b-db1d-47f9-8857-79c13b240555_origin.pdf b/2025/Structure-Guided Diffusion Models for High-Fidelity Portrait Shadow Removal/73c9755b-db1d-47f9-8857-79c13b240555_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4eaceb9b626c38b144a425ded49a06a71b3f4dc1 --- /dev/null +++ b/2025/Structure-Guided Diffusion Models for High-Fidelity Portrait Shadow Removal/73c9755b-db1d-47f9-8857-79c13b240555_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:10e58061512ea96e3cbe82148b6219e70300df85a94ff962b3367c22cd5657f8 +size 17802025 diff --git a/2025/Structure-Guided Diffusion Models for High-Fidelity Portrait Shadow Removal/full.md b/2025/Structure-Guided Diffusion Models for High-Fidelity Portrait Shadow Removal/full.md new file mode 100644 index 0000000000000000000000000000000000000000..890596484dff38d8f40a3619bf289ac2ce5ff422 --- /dev/null +++ b/2025/Structure-Guided Diffusion Models for High-Fidelity Portrait Shadow Removal/full.md @@ -0,0 +1,395 @@ +# Structure-Guided Diffusion Models for High-Fidelity Portrait Shadow Removal + +Wanchang Yu $^{1}$ Qing Zhang $^{1,2*}$ Rongjia Zheng $^{1}$ Wei-Shi Zheng $^{1,2}$ + +$^{1}$ School of Computer Science and Engineering, Sun Yat-sen University, China + +$^{2}$ Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China + +![](images/3cf8d30aa782822b05677e0f926112a293b96c860170714cad13c35ae3d8adcf.jpg) +Input + +![](images/c3298ebe9d3a8845c02a3cbf22a06de24fb03e23211ea83475f2a9246b5941c5.jpg) +ShadowDiffusion [11] + +![](images/161affd385f29f95cd7b767d7c1d35cc417e9480a68c18faed34ba5a79ba1894.jpg) +IC-light [52] + +![](images/c3c399eef96a8188dccafb0ddbae2b8fc216a3681ff6c74e7a5b93cd5652bf8a.jpg) +BSR [28] +Figure 1. Comparison on portrait shadow removal. Our method allows to robustly remove facial cast shadows while faithfully preserving the portrait identities as well as structure details in shadow regions and also ensuring the overall naturalness. + +![](images/ccc485ed95a40fca12f4db7182d0ba7362811c9b4641daf2ce1b843cb5dda6a6.jpg) +Ours + +![](images/3e7d2516ef60de9d5d917cd9eab1fc97d84fff5750771760aab23f5e88bb3064.jpg) +GT + +# Abstract + +We present a diffusion-based portrait shadow removal approach that can robustly produce high-fidelity results. Unlike previous methods, we cast shadow removal as diffusion-based inpainting. To this end, we first train a shadow-independent structure extraction network on a real-world portrait dataset with various synthetic lighting conditions, which allows to generate a shadow-independent structure map including facial details while excluding the unwanted shadow boundaries. The structure map is then used as condition to train a structure-guided inpainting diffusion model for removing shadows in a generative manner. Finally, to restore the fine-scale details (e.g., eyelashes, moles and spots) that may not be captured by the structure map, we take the gradients inside the shadow regions as guidance and train a detail restoration diffusion model to refine the shadow removal result. Extensive experiments on the benchmark datasets show that our method clearly outperforms existing methods, and is effective to avoid previously common issues such as facial identity tampering, shadow residual, color distortion, structure blurring, and loss of details. Our code is avail + +able at https://github.com/wanchang-yu/Structure-Guided-Diffusion-for-Portrait-Shadow-Removal. + +# 1. Introduction + +Taking selfies is very popular nowadays due to the readily-available cameras on various devices, particularly the mobile phones. However, portrait photos taken in the wild lacking professional lighting conditions often suffer from undesired foreign shadows cast by external objects (see Figure 1 for examples). The presence of facial shadows not only severely degrade the overall visual appeal of the portraits, but also adversely affect many downstream vision tasks, such as face detection and recognition [4, 5, 41, 55], and portrait image editing [21, 27, 47]. Hence, portrait shadow removal techniques are widely required. + +Currently, portrait shadow removal remains a challenge despite the remarkable progress in natural image shadow removal. The reason behind is twofold. First of all, due to factors such as privacy and difficulty in data collection, large-scale paired data for portrait shadow is extremely scarce, which significantly limits the development of portrait shadow removal methods, especially learning-based solutions. On the other hand, as this task is portrait- + +oriented, it has very low tolerance for visual artifacts, even minor imperfections on facial color and details could make the final result unnatural and unacceptable. Note, although existing software allow users to interactively adjust photos, it is still rather tedious and difficult for non-experts to manually produce high-fidelity portrait shadow removal results. + +Existing shadow removal methods [11, 26, 45] mainly target natural images due to the availability of large-scale paired datasets. However, these methods basically fail to deal with portrait shadows, as shown in Figures 5 and 6. As there is no accessible large-scale paired datasets for portrait shadow removal, current methods in this field focus primarily on constructing synthetic paired data to enable supervised learning of shadow removal [28, 33, 52, 56]. Suffering from the domain gap between real-world portraits and synthetic data, these methods typically do not work well for real-world portraits with complex shadows. To avoid the domain gap issue led by synthetic training data, a zero-shot solution is introduced in [14], which proposed to leverage the generative prior from pre-trained StyleGAN2 [24] to recover shadow-free portrait through test-time GAN inversion. However, this method tends to modify face identity and incur unnatural results. Note, although recent image relighting techniques [8, 16, 43, 50] have shown remarkable progress, they struggle to enable high-fidelity portrait shadow removal due to the inherent difficulty in estimating ambient light and facial geometry from a single image. + +To robustly produce high-fidelity portrait shadow removal results, we in this work present a diffusion-based approach. Particularly, we propose to cast portrait shadow removal as diffusion-based inpainting, so as to take full advantage of the strong generation capability of diffusion models. To this end, we first train a shadow-independent structure extraction network (SE-Net) on a dataset of paired real-world portraits with various synthetic lighting conditions, from which we obtain a structure map containing facial details while excluding the shadow boundaries. With the structure map, we then train a structure-guided inpainting diffusion model by enforcing reconstruction of randomly masked shadow-free portraits. Finally, to restore the fine-scale facial details that may not be captured by the structure map, we further refine the shadow removal result by a detail restoration diffusion model under the guidance of the gradients in the original shadow regions. Note, although our method requires a test shadow mask as input, it is highly robust to inaccurate masks, requiring only that the mask covers the shadow regions, even if it includes nonshadow regions. The major contributions of this work are: + +- We present a simple yet effective unsupervised approach that addresses portrait shadow removal from a perspective of diffusion-based inpainting. +- We introduce a relighting-based data synthesis strategy for constructing the training data, and develop a shadow + +independent structure extraction network. + +- We show that our method outperforms state-of-the-art methods for portrait shadow removal. + +# 2. Related Work + +Shadow removal. Early methods typically rely on various handcrafted priors or assumptions to work [7, 9, 12]. Recent methods are basically deep learning based [3, 10, 11, 17, 20, 22, 23, 25, 29, 31, 57, 60, 61], benefiting mostly from the availability of large-scale paired datasets such as ISTD [42] and SRD [35]. However, as these methods are tailored for natural images, they are not so effective in dealing with portraits. In contrast to natural images, portrait shadow removal has received much less research attention, due to the lack of large-scale paired datasets. To overcome the data scarcity issue, some methods [28, 56] proposed to synthesize paired shadow and shadow-free portrait images for training. As their data synthesis strategies involve only simple adjustments on brightness and saturation, without considering facial geometry and albedo, their methods typically do not work well for real-world portraits. He et al. [14] achieved zero-shot portrait shadow removal by utilizing generative prior from StyleGAN2, but this method may result in undesired modifications to facial details and thus fails to preserve the face identity. More recently, image relighting is also used for shadow removal [8, 16, 43, 50, 52]. However, this method tends to induce unnatural results because it is very challenging to obtain accurate prediction of diffuse lighting and albedo from a single portrait image. + +Image inpainting. Before the era of deep learning, patch-based self-similarity of natural images has long been the foundation for image inpainting [2, 44]. The emergence of deep learning has then led to great progress in image inpainting [18, 34, 40, 51], as it is quite easy to construct paired training data. Recently, diffusion model has been a preferable choice for image inpainting [1, 19, 32, 37, 46, 59], due to its impressive generation capability. Note, naive applying diffusion models to portrait shadow removal fail to produce visually natural results with unchanged portrait identity. A recent work closely related to our method also shows that it is feasible to remove shadows via image inpainting [26]. However, it does not work well for portrait shadow removal (see Figures 5 and 6), even we fine-tuned their model on real-world paired shadow and shadow-free portrait images provided in [56]. + +# 3. Our Method + +Given a portrait image with cast shadows, our method aims to remove the facial shadows while faithfully preserving the facial details in the shadow regions. Figure 2 presents the overall inference pipeline of our method. + +![](images/a7303ded20a690d3f0212416f31630b3978b34c76ed7a4959cdd19b167a9cbc4.jpg) +Figure 2. Inference pipeline of our method. Given a input portrait image and the corresponding shadow mask, we first feed the input image to the shadow-independent structure extraction network (SE-Net) to get a structure map. Then, we perform shadow removal by inpainting the shadow regions with a diffusion model conditioned on the structure map. Finally, the shadow removal output is refined by a detail restoration diffusion model to restore the fine-scale facial details that may not be captured by the strcutre map. + +![](images/9a86e93f454fd316c70ab47d6505efd197a8b51107ac8b0a8b7e6cdd3b100bb4.jpg) +Figure 3. Training pipeline of SE-Net. Note, we adopt a trained PDG model for structure extraction from [49], while it is not invariant to illumination changes. + +# 3.1. Shadow-Independent Structure Extraction + +As shown in Figure 2, the key to obtaining high-quality portrait shadow removal through inpainting is to acquire a structure map free of shadow boundaries for guiding the inpainting process. However, although existing structure/edge extraction methods are quite effective in identifying edges and structures, they are sensitive to illumination discontinuities and will inevitably include shadow boundaries in their results (see Figure 7). Therefore, we propose to build a shadow-independent structure extraction network. + +Training data synthesis. To enable effective training of the shadow-independent structure extraction network, as shown in Figure 3, we construct a synthetic paired dataset of well-lit real-world portraits without foreign shadows on faces and the corresponding relighted portraits with various illumination discontinuities, based on the shadow-free portraits in the CelebA dataset [30]. Specifically, for a well-lit portrait + +$I$ , we first apply an off-the-shelf physics-based relighting method [58] over it to get a relighted image $I_{relit}$ , which is then combined with the original portrait $I$ to produce a synthetic image $I_{syn}$ with illumination discontinuities by + +$$ +I _ {s y n} = M \odot I _ {r e l i t} + (1 - M) \odot I, \tag {1} +$$ + +where $M$ is a random facial mask, and $\odot$ denotes elementwise multiplication. Note, our synthetic data is only used for training the SE-Net and the subsequent detail resotration diffusion model, rather than for directly training shadow removal model as done in [56]. + +Training losses. As shown in Figure 3, we adopt a trained structure extraction model from [49] (referred to as PDG in paper) to produce structure map for the original well-lit portrait, which is treated as pseudo ground-truth for supervising the training of our SE-Net. By respectively denoting our SE-Net and the PDG model as $G_{s}$ and $G_{p}$ , the training losses of our SE-Net are defined as: + +$$ +\left\{ \begin{array}{l} \mathcal {L} _ {\text {r e c}} = \| G _ {s} (I _ {\text {s y n}}) - G _ {p} (I) \| _ {1}, \\ \mathcal {L} _ {\text {p e r c e p t u a l}} = \mathcal {L} _ {\text {L P I P S}} \left(G _ {s} (I _ {\text {s y n}}), G _ {p} (I)\right), \end{array} \right. \tag {2} +$$ + +where $\mathcal{L}_{LPIPS}$ refers to the learned perceptual image patch similarity [54]. Besides, we also utilize a discriminator to determine whether the output of SE-Net belongs to the structure image domain, and further define an adversarial loss $\mathcal{L}_{GAN}$ (see supplementary material for details). In summary, the overall loss function for training the SE-Net is given by: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {r e c}} + \lambda_ {1} \mathcal {L} _ {\text {p e r c e p t u a l}} + \lambda_ {2} \mathcal {L} _ {\text {G A N}}, \tag {3} +$$ + +where we empirically set $\lambda_1 = 0.5$ , $\lambda_1 = 0.25$ , which produce good results, as demonstrated in Figure 7. + +![](images/b4d96f412dedaf6e5e4e50b52fbfb67030bf64e18a675e3f1e14f28033e1b8d6.jpg) +Structure-guided Portrait Inpainting + +![](images/92f6fff8c2118d35e7a7305efd6deaa999a3b20db6f72abbdb1a3bde4d543cb3.jpg) +Gradient-guided Detail Restoration +Figure 4. Training pipeline of our diffusion models. Here we introduce how our structure-guided portrait inpainting diffusion model and gradient-guided detail restoration diffusion model are trained. + +# 3.2. Structure-Guided Portrait Inpainting + +As shown in Figure 2, with structure map, the next step is to inpaint the shadow regions. To this end, we develop a structure-guided portrait inpainting diffusion model. + +Structure-aware denoising. As shown in Figure 4, we start with a shadow-free portrait as input and then randomly mask some facial regions to create a masked input, followed by a diffusion model to reconstruct the original input by inpainting the masked regions. The inpainting is rooted in the forward diffusion process, where the diffusion model learns a Markovian chain by progressively corrupting the target image $x_0$ (namely the masked input) with Gaussian noise at each timestep $t$ , such that it is transformed into a Gaussian distribution (referred to as $\mathcal{N}$ in the following). Specifically, after reparameterization [15], our forward process obtains $x_t$ from $x_0$ in a single step by: + +$$ +x _ {t} = \sqrt {\bar {\alpha} _ {t}} x _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon , \tag {4} +$$ + +where $\bar{\alpha}_t = \prod_{i=0}^t \alpha_i$ is the cumulative product of forward noise weight parameters $\alpha_i = 1 - \beta_i$ , while $\epsilon \sim \mathcal{N}(0, I)$ denotes random Gaussian noise. + +As for the reverse denoising process, we use the sampling strategy of DDIM [39] for acceleration. Note, unlike unconditional image generation, our goal is to faithfully reconstruct the original unmasked portrait, the reverse denoising process is thus conditioned on a structure map $S$ and defined as: + +$$ +x _ {t - 1} = \sqrt {\bar {\alpha} _ {t - 1}} \left(\frac {x _ {t} - \sqrt {1 - \bar {\alpha} _ {t}} \cdot \mathbf {e} _ {t}}{\sqrt {\bar {\alpha} _ {t}}}\right) + \sqrt {1 - \bar {\alpha} _ {t - 1}} \cdot \mathbf {e} _ {t}, \tag {5} +$$ + +where $\mathbf{e}_t = \epsilon_\theta (x_t,I_M,S,M,t)$ refers to the noise estimated by a deep neural network $\epsilon_{\theta}$ , where $I_M$ is the masked input created from a random mask $M$ . Following [15], the training objective for diffusion model can be simplified as: + +$$ +\mathcal {L} _ {\text {d i f f}} (\epsilon) = \| \epsilon - \epsilon_ {\theta} \left(x _ {t}, I _ {M}, S, M, t\right) \| _ {2} ^ {2}, \tag {6} +$$ + +where $\epsilon$ is the ground truth noise, while $\epsilon_{\theta}(x_t,I_M,S,M,t)$ is the noise predicted by the network. As shown in Figure 2, a facial shadow mask is required at the inference stage, which is a dilated version of the facial shadow detection result from the trained model of [6]. + +Shadow mask refinement. To enhance the robustness of our method to inaccurate shadow mask, we propose to refine the mask based on the shadow removal result generated by the inpainting diffusion model. Specifically, we first compute the absolute difference between the input image and the shadow removal result, and then perform Otsu-based thresholding to get a refined shadow mask excluding all the non-shadow regions. With the refined mask, we further update the shadow removal result by combining it with the input image, so as to ensure that non-shadow regions in the result are unchanged. + +# 3.3. Gradient-Guided Detail Restoration + +As the original PDG model is designed for generating portrait drawing formed by large-scale structures, thus our SE-Net is similarly not so effective in capturing fine-scale portrait details, such as eyelashes, moles, and spots (see the bottom row in Figure 7 for example). To address this issue, we design a detail restoration diffusion model to refine the shadow removal result generated by the inpainting diffusion model, where we directly take the gradients in the shadow regions as guidance to restore the fine-scale details. Figure 4 describes how we train the detail restoration diffusion model. Note, gradient-aware denoising process is similar to the structure-aware denoising described above, and the two share the same forward diffusion and reverse denoising formulations described in Equations 4 and 5, with the latter conditioned on gradient map rather than structure map. + +![](images/8e134908d42ac20fe5bd0c032fc838bfc124751ae2886510fd474bcfb1e911c9.jpg) +Figure 5. Visual comparison with state-of-the-art methods on test images from the dataset provided in [56]. + +
MethodAllShadow regionNon-shadow region
SSIM↑LPIPS↓RMSE↓SSIM↑LPIPS↓RMSE↓SSIM↑LPIPS↓RMSE↓
ShadowDiffusion [11]0.6500.17738.2460.9010.03715.5610.5790.22352.276
Inpaint4Shadow [26]0.7660.09422.7090.9100.03317.2500.7080.16445.456
ShadowFormer [10]0.7720.08621.4130.9260.02715.2880.7230.14942.341
HomoFormer [45]0.7860.10021.7240.9130.04116.2480.7110.16742.950
PSM [56]0.7820.074-------
UPSR [14]0.7310.10926.9150.9000.03920.6590.6680.18845.391
BSR [28]0.6050.11828.5320.9100.03615.1930.5480.17143.627
IC-light [52]0.5140.27862.5390.9160.05733.0610.4780.31969.208
Ours w/o structure guidance0.7570.10123.5980.8980.04318.5140.7140.13839.000
Ours w/ PDG predicted structure0.7790.09219.9630.9190.03514.0620.7250.13137.387
Ours w/ data synthesis in PSM0.7850.08619.7970.9250.03213.7520.7290.13638.485
Ours w/o detail-restoration0.8140.06218.5370.9430.02012.7480.7470.12036.326
Ours (full method)0.8300.05617.1620.9730.01110.1980.7780.09134.573
+ +Table 1. Quantitative comparison on the real-world portrait shadow removal dataset provided in [56]. + +# 3.4. Implementation and Training Details + +Network architecture. The SE-Net is an encoder-decoder network, which starts with a flat convolution and two down convolution blocks, followed by nine residual blocks, and finally two up convolution blocks and a final convolution layer. Please see also the supplementary material for visual illustration of the architecture of SE-Net. As for the inpainting diffusion model and the detail-restoration diffusion model, we follow [15] to adopt the same U-Net architecture. + +Training details. To train SE-Net, we synthesize 20,000 + +portrait images with diverse lighting conditions based on the CelebA dataset [30], and train it for 3 epochs with a batch size of 8 and a learning rate of $1.5 \times 10^{-5}$ . As for the inpainting diffusion model, we train it on the FFHQ dataset for 300 epochs with a batch size of 32, and train the detail restoration diffusion model on the same synthetic dataset as the SE-Net for 200 epochs. We build our model on PyTorch and use the Adam optimizer with momentum parameters of (0.9, 0.999) and an initial learning rate of $1 \times 10^{-4}$ . Following [38], we employ the Kaiming initialization [13] to initialize the weights of our model and apply 0.9999 Expo + +![](images/e4ac4d3cf14c4ef937c8b88e60c660008fd027e29768ef42ee628e1ebbf2110e.jpg) +Figure 6. Visual comparison with state-of-the-art methods on test images from our collected dataset. + +nential Moving Average (EMA) in all our experiments. We use 2000 diffusion steps and a noise schedule $\beta_{t}$ linearly increasing from $1\times 10^{-6}$ to $1\times 10^{-2}$ for training. During inference, we perform both the inpainting diffusion and detail restoration diffusion for 40 denoising steps. + +# 4. Experiments + +Datasets. Following previous portrait shadow removal approaches [14, 28], we evaluate our method on the only publicly available real-world paired dataset provided in PSM [56], which contains 9 subjects and 100 shadowed portrait images in varying poses, shadow shapes, illumination conditions, and shadow types. To better examine the effectiveness of our method, we additionally construct a dataset of 200 portraits with diverse facial shadows. Note, we only conduct visual comparison on our collected dataset due to the lack of ground truth shadow removal results. + +Evaluation metrics. We use three commonly-used metrics to quantitatively evaluate the performance of our method, including SSIM, LPIPS, and the root mean square error (RMSE) computed in the LAB color space for measuring the similarity between the predicted results and the corresponding ground-truth shadow-free images. + +# 4.1. Comparison with State-of-the-art Methods + +Baselines. We compare our method with various state-of-the-art shadow removal methods, including four portrait shadow removal and editing methods, i.e., PSM [56], + +UPSR [14], and BSR [28], IC-light [52], as well as four natural image shadow removal methods, i.e., ShadowDiffusion [11], Inpaint4Shadow [26], ShadowFormer [10], and HomoFormer [45]. For fair comparison, we produce their results using publicly-available implementation or trained models provided by the authors with recommended parameter setting. Note that, we do not compare with other recent relighting-based shadow editing methods [8, 16, 43, 50], as their codes and trained models are not yet released. + +Visual comparison. We in Figures 5 and 6 conduct visual comparison between our method and previous shadow removal methods. As shown, shadow removal methods designed for natural images, i.e., Inpaint4Shadow [26], HomoFormer [45], and ShadowDiffusion [11] are not well-suited to portraits and tend to induce obvious shadow residuals and appearance distortions. As for the other three portrait shadow removal methods, UPSR [14] and BSR [28] fail to preserve the facial details and may also lead to shadow residuals and color distortions, while IC-Light [52] struggles to remove the facial shadows due to the inherent difficulty in estimating the ambient light and facial geometry. Please see also the supplementary material for more visual comparison results. + +Quantitative comparison. In Table 1, we quantitatively compare our method with other methods in terms of the SSIM, LPIPS, and RMSE metrics. As can be seen, our method outperforms all the compared methods on all the three metrics, demonstrating its effectiveness and superior- + +![](images/d3a1f1b9909763a8bd6c93a664c30f93c5377033abf947f9d3ab94db8cf19860.jpg) + +![](images/fc40af2c9ab6980918d472d07b257d94a90d98099c1eb677a49c059a770c2382.jpg) +(a) + +![](images/7e042422b27826ef4c7bb4600cd59234f09dfc597ea3f9076d05e6db02fe7bf1.jpg) + +![](images/861e15e6be71cd23543034cd3a7af6c7455ca23d3d5b780c5bfd5b664dcae781.jpg) +(b) + +![](images/39bd9cb93efc8dcd2a1d41d3da183eb2b4bd922088c5f6445639afd8b84dc899.jpg) + +![](images/c54e6a6f03ffcfcda671e32c13b6c6f573e90cff2791161f84df163d7d0208bf.jpg) +(c) + +![](images/047d3b8ae34d6c8a75ad7b47ee8d8971d8d1b1f3f0d4b86850f69cdbb8e34daf.jpg) + +![](images/1011ee522bfad66898834852b5f879b146e90a066068f99f4a6d26be2ab18f52.jpg) +(d) + +![](images/2fcd9ba1b290fd6310aa9dc580c47022dda7133249eacbcec59b618e302e000e.jpg) + +![](images/4a2571cfe0e8dc6891c964a45972191dbc0f7a7dd1ee83f4be96fdc23af5bd04.jpg) +(e) + +![](images/18433bf6eea6c41964b0033ebbf4515de80be3ec799ab2983cf3bde3b36c7d08.jpg) + +![](images/25e235c91028bb90ebd6946430dccc7ba3d4e1c8184f1289ec8c24c159213311.jpg) +(1) + +![](images/79c1e66398c864e6e6fbed933cc1f66e6dfe902b87e9966f61aeb5774a741f17.jpg) + +![](images/08f57d647c4a4bd0a8ab2e697a408f5ecc5eb467edbf1558c3e5865e90c36542.jpg) +(g) + +ity in portrait shadow removal. Note, as PSM [56] provides only code for synthesizing training data, without releasing the training code and trained model, its quantitative results in Table 1 are as reported in their paper. + +# 4.2. Ablation Study + +Effectiveness of our shadow-independent structure extraction. Although the trained model of PDG [49] is utilized to train our structure extraction, the structure map it produces is not what we expected because it will include the shadow boundaries and lead to unnatural shadow removal results, as shown in Figure 7 (b)&(e). By training on our synthesized data with diverse lighting conditions, our structure extraction network (SE-Net) can robustly extract shadow-independent structures, enabling high-quality shadow removal results. + +Effectiveness of our data synthesis strategy. To verify the effectiveness of our data synthesis, we alternatively adopt the portrait shadow synthesis strategy introduced in PSM [56] to create data on FFHQ for training our SE-Net. As shown in Figure 7 (c)&(f) and Table 1, training on the data synthesized by the strategy in PSM fails to exclude the unwanted shadow boundaries and will degrade the subsequent shadow removal performance, because the strategy only involves simple modifications to brightness, saturation, and subsurface scattering of face, lacking the ability to physically simulate the complex facial lighting conditions. + +Effect of structure map on shadow inpainting. We in Figure 8 examine how structure map derived from the SE-Net affects the performance of our diffusion-based shadow inpainting. As shown, omitting the structure map, the facial structure details behind shadows are inevitably modified, producing unpredictable portrait inconsistent to the input. + +Effect of detail restoration. We in Figure 9 analyze the + +![](images/751379de2c98ea584f7029e8f0a94d38dff28013fade88287fd63af57f20468a.jpg) +Input + +![](images/9b90fbe973c0e75e82a7325002bff7f3cf3cda9d898a676956f07402793f4cef.jpg) +w/o structure + +![](images/a4ff21630911b15e4c14a20b3fa4a2c574d91193c8f338ca5f4f17bdb1a98eb7.jpg) +w/ structure + +![](images/528f54b2eb1d310bc63fad235d93d6b05ca0d6239324ccae90ad3b08f8b2dbc3.jpg) +Figure 7. Visual comparison of structure maps extracted by different methods and the corresponding shadow removal results. (a) Input. (b)-(d) are structure maps generated by PDG [49], SE-Net trained on data synthesized by the method of PSM [56], and SE-Net trained on our synthesized data. (e)-(g) are shadow removal results produced by our method using the structure maps in (b)-(d), respectively. +GT + +![](images/772faa9f8e537de7ff4127c3843e9a3e0c950a4c5d129ca44110d319ee8e1928.jpg) +Figure 8. Effect of structure map on shadow inpainting. + +![](images/9c3c852d1697744aa9cd343151f373d7cba8248c6aee8ba5766abd428abc52bd.jpg) +Input +Figure 9. Effect of our detail restoration. + +![](images/0381bc227f96ed7194a59190d0ca97b7d2c4f8102a0c60058ab5be9ba01e9ece.jpg) + +![](images/4e00420179498b9a608b4ecb451c1a2937276c61330d7dfebfd982d295d9ed61.jpg) +w/o detail restoration + +![](images/3abaf0203442a7054f5c4fc591b97b4492cfa44b24a120d9d96d0cf67e39361a.jpg) + +![](images/67fd04bb3c7d907e7aa1b5bd8a2c2872f7504d266e36c915d6e7c7d1f9f1d8f0.jpg) +w/ detail restoration + +impact of our detail restoration. As shown, without the restoration step, our shadow removal results exhibit noticeable blurring artifacts around the eyebrows (see 1st row), and fail to faithfully recover the mole in shadow region (see 2nd row). By including the restoration step, we are able to produce high-fidelity shadow removal results with all aforementioned fine-scale details effectively restored, manifesting the effectiveness of our detail restoration. + +# 4.3. More Analysis + +Impact of inaccurate structure map. Structure prediction plays a vital role for the success of our method, because we + +![](images/e600ffbb8d1c8dbd17c755cc3700d7f1245eaf32fe3593914d0e4b6ff222459e.jpg) +Input +Figure 10. Impact of inaccurate structure map. + +![](images/fb8a9b934382839fa5c60250a65bbd4d3b904b0431a887e9993ad9f0fa16db5f.jpg) +Inaccurate map + +![](images/a94ea827fc0bc74e1266fc71dff96308146e70590f20f533652deda6cf2a4751.jpg) +Inpainting result + +![](images/9db58f640587a15631f2fdff54b5d9e49361417070f5b6601356288a2e38eced.jpg) +Final result + +![](images/01c0509b64eded687a734909bd7daa5435908f69c8c3cdeedc1c36e715fcc6a2.jpg) + +![](images/21309f1944af8da8eff70359b4c8a63745c69a94de46723d633d496d049a34f8.jpg) + +![](images/907acb1c6a839ccf02f3d5d5b10018cc7a950cae6bc16a1c6f524e3b5ba5d240.jpg) +ID: 0.056 +Figure 11. Effect of inaccurate shadow masks. Our method is highly robust to inaccurate masks (see 3rd and 4th columns). + +![](images/559e49f4535c14c5f6c57f555325a681197199bc6820f668c56693f5f3fd8b1f.jpg) + +![](images/eafd08d435cabd0ca9e9e6cb508f5dfc4b0e523ce56edd2859cc66d5099ada9a.jpg) +ID:0.065 + +![](images/22f31829e8c49b341c2b0b3e5cc432b5ac9d3e24215b765e19458093f7756a83.jpg) + +![](images/827d3c79f55d6266a2cbcb74d54668921df91accb9b5391dca31898a4a984527.jpg) +ID:0.062 + +rely on a structure map to guide our diffusion-based shadow inpainting. However, as shown in Figure 10, our method still has some tolerance to inaccurate structure map as the subsequent detail restoration stage can help remove the inpainting artifacts led by inaccurate structure map. + +Robustness to inaccurate shadow masks. Thanks to the design of performing portrait shadow removal via structure-guided inpainting, our method is highly robust to inaccurate input shadow masks, requiring only that the masks cover the shadow regions even if they include large-scale non-shadow regions. As demonstrated in Figure 11, our method achieves both visually indistinguishable shadow removal results and comparable ID preservation scores [48] for all the three masks, including two inaccurate masks in the 3rd and 4th columns. Please see the supplementary material for more results produced with inaccurate shadow masks. + +Different diffusion models for shadow inpainting. To test how different diffusion models affect the performance of our method, we replace our inpainting diffusion model with a strong pre-trained Stable Diffusion (SD) model (version v1-5 inpainting model [36]). To ensure fair comparison, the compared SD model is injected with the same structure map from SE-Net using ControlNet [53]. As can be seen in Figure 12, our trained inpainting diffusion model produces high-fidelity shadow removal results faithful to the inputs, while the SD model incurs noticeable structure tampering. + +Effect of our method on natural images. Our method is also applicable to natural images by replacing the portrait-oriented PDG model [49] with a generic edge detection model and also changing the training dataset. We show + +![](images/a986e113c2a5f685b1b9c35642a0fc9ca65a1ab85faa44f8ab7a26088e74441c.jpg) + +![](images/dd169ab83d7fcccfc797bb9e3e70c75cd432ae5ed8d9a65875d1340c3c78c5ee.jpg) +Input +Figure 12. Effect of leveraging different diffusion models for shadow inpainting. Note, the same structure map as ours is used to produce the shadow inpainting result of "SD + ControlNet", and our result here is free of detail restoration. + +![](images/b1eb37c0f980142e53e0292c1f2546685b36384ac8e81c5cb5c971f6335b561c.jpg) + +![](images/71ac04d28ba94f701a64a6cff401abf1d61a0f481505bb4852e8bcd3761a3b8b.jpg) +SD + ControlNet + +![](images/68a4f52e59f7700af2b1d5a2744361ffea3a70ed994d0994c26a7f5371f4bc0a.jpg) + +![](images/5bbd776822c6f83e1cc6c666a6337e85dde49ec04f16504443e2c4a0db203f46.jpg) +Ours + +![](images/57769de3d5392ba314900bc24ea4f60029b914d3e5f6b8e84a8b789eedb5abb3.jpg) + +![](images/269a4b717cf02d476b822c2969648a07eab5c4cbfab5003232515bfac0f0e3a3.jpg) +GT + +
MethodUPSRBSRShadowDiffusionIC-lightOurs
Param.135M72M108.79M-185.12M
Time192s11s2.3s32s9.5s
+ +Table 2. Comparison on the number of parameters (param.) and inference time. Note, all the inference times here are test for a $512 \times 512$ portrait on an NVIDIA A100 Tensor Core GPU. + +in the supplementary material that our method can produce competitive shadow removal results for natural images. + +Analysis on computational complexity. Table 2 compares our inference time with those of other existing portrait specialized or diffusion-based shadow removal methods. As shown, our inference time is more efficient than that of most other methods. Moreover, the number of parameters of our method are comparable to other methods, showing that our superior performance comes mainly from method design. + +# 5. Conclusion + +We have presented a portrait shadow removal approach that allows to robustly produce high-fidelity results. Unlike previous methods, we cast portrait shadow removal as diffusion-based inpainting. To do so, we develop a structure extraction network that is robust to illumination changes to estimate shadow-independent structure map for a given portrait image. The structure map is then fed to an inpainting diffusion model to guide the generation of the shadow-free output. Next, we refine the shadow-free output with a detail restoration diffusion model for impairing the fine-scale facial details that may not be captured by the structure map. Extensive experiments show that our method outperforms state-of-the-art methods for portrait shadow removal. + +Acknowledgment. This work was supported by the National Natural Science Foundation of China (62471499), the Guangdong Basic and Applied Basic Research Foundation (2023A1515030002). + +# References + +[1] Ciprian Corneanu, Raghudeep Gadde, and Aleix M Martinez. Latentpaint: Image inpainting in latent space with diffusion models. In WACV, pages 4334-4343, 2024. 2 +[2] Antonio Criminisi, Patrick Pérez, and Kentaro Toyama. Region filling and object removal by exemplar-based image inpainting. IEEE Transactions on Image Processing, 13(9): 1200-1212, 2004. 2 +[3] Xiaodong Cun, Chi-Man Pun, and Cheng Shi. Towards ghost-free shadow removal via dual hierarchical aggregation network and shadow matting gan. In AAAI, pages 10680-10687, 2020. 2 +[4] Xingbo Dong, Sangrae Cho, Youngsam Kim, Soohyung Kim, and Andrew Beng Jin Teoh. Deep rank hashing network for cancellable face identification. Pattern Recognition, 131:108886, 2022. 1 +[5] Pawel Drozdowski, Christian Rathgeb, and Christoph Busch. The watchlist imbalance effect in biometric face identification: Comparing theoretical estimates and empiric measurements. In ICCV, pages 3757-3765, 2021. 1 +[6] Xianyong Fang, Xiaohao He, Linbo Wang, and Jianbing Shen. Robust shadow detection by exploring effective shadow contexts. In ACM MM, pages 2927-2935, 2021. 4 +[7] Graham D Finlayson, Mark S Drew, and Cheng Lu. Entropy minimization for shadow removal. International Journal of Computer Vision, 85(1):35-57, 2009. 2 +[8] David Futschik, Kelvin Ritland, James Vecore, Sean Fanello, Sergio Orts-Escalano, Brian Curless, Daniel Sykora, and Rohit Pandey. Controllable light diffusion for portraits. In CVPR, pages 8412-8421, 2023. 2, 6 +[9] Maciej Gryka, Michael Terry, and Gabriel J Brostow. Learning to remove soft shadows. ACM Transactions on Graphics (TOG), 34(5):1-15, 2015. 2 +[10] Lanqing Guo, Siyu Huang, Ding Liu, Hao Cheng, and Bihan Wen. Shadowformer: Global context helps shadow removal. In AAAI, pages 710-718, 2023. 2, 5, 6 +[11] Lanqing Guo, Chong Wang, Wenhan Yang, Siyu Huang, Yufei Wang, Hanspeter Pfister, and Bihan Wen. Shadowdiffusion: When degradation prior meets diffusion model for shadow removal. In CVPR, pages 14049-14058, 2023. 1, 2, 5, 6 +[12] Ruiqi Guo, Qieyun Dai, and Derek Hoiem. Paired regions for shadow detection and removal. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(12):2956-2967, 2012. 2 +[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, pages 1026-1034, 2015. 5 +[14] Yingqing He, Yazhou Xing, Tianjia Zhang, and Qifeng Chen. Unsupervised portrait shadow removal via generative priors. In ACM MM, pages 236-244, 2021. 2, 5, 6 +[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 4, 5 +[16] Andrew Hou, Zhixin Shu, Xuaner Zhang, He Zhang, Yannick Hold-Geoffroy, Jae Shin Yoon, and Xiaoming Liu. + +Compose: Comprehensive portrait shadow editing. In ECCV, pages 356-373. Springer, 2024. 2, 6 +[17] Xiaowei Hu, Yitong Jiang, Chi-Wing Fu, and Pheng-Ann Heng. Mask-shadowgan: Learning to remove shadows from unpaired data. In ICCV, pages 2472–2481, 2019. 2 +[18] Satoshi Iizuka, Edgar Simo-Serra, and Hiroshi Ishikawa. Globally and locally consistent image completion. ACM Transactions on Graphics (ToG), 36(4):1-14, 2017. 2 +[19] Liya Ji, Zhefan Rao, Sinno Jialin Pan, Chenyang Lei, and Qifeng Chen. A diffusion model with state estimation for degradation-blind inverse imaging. In AAAI, pages 2471-2479, 2024. 2 +[20] Hao Jiang, Qing Zhang, Yongwei Nie, Lei Zhu, and Wei-Shi Zheng. Learning to remove shadows from a single image. International Journal of Computer Vision, 131(9):2471-2488, 2023. 2 +[21] Yuming Jiang, Ziqi Huang, Xingang Pan, Chen Change Loy, and Ziwei Liu. Talk-to-edit: Fine-grained facial editing via dialog. In ICCV, pages 13799-13808, 2021. 1 +[22] Yeying Jin, Aashish Sharma, and Robby T Tan. Dcshadownet: Single-image hard and soft shadow removal using unsupervised domain-classifier guided network. In ICCV, pages 5027-5036, 2021. 2 +[23] Yeying Jin, Wei Ye, Wenhan Yang, Yuan Yuan, and Robby T Tan. Des3: Adaptive attention-driven self and soft shadow removal using vit similarity. In AAAI, pages 2634-2642, 2024. 2 +[24] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR, pages 8110-8119, 2020. 2 +[25] Hieu Le and Dimitris Samaras. Shadow removal via shadow image decomposition. In ICCV, pages 8578-8587, 2019. 2 +[26] Xiaoguang Li, Qing Guo, Rabab Abdelfattah, Di Lin, Wei Feng, Ivor Tsang, and Song Wang. Leveraging inpainting for single-image shadow removal. In ICCV, pages 13055-13064, 2023. 2, 5, 6 +[27] Zonglin Li, Xiaogian Lv, Wei Yu, Qinglin Liu, Jingbo Lin, and Shengping Zhang. Face shape transfer via semantic warping. Visual Intelligence, 2(1):26, 2024. 1 +[28] Yaojie Liu, Andrew Z Hou, Xinyu Huang, Liu Ren, and Xiaoming Liu. Blind removal of facial foreign shadows. In BMVC, page 88, 2022. 1, 2, 5, 6 +[29] Yuhao Liu, Zhanghan Ke, Ke Xu, Fang Liu, Zhenwei Wang, and Rynson WH Lau. Recasting regional lighting for shadow removal. In AAAI, pages 3810-3818, 2024. 2 +[30] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In ICCV, pages 3730-3738, 2015. 3, 5 +[31] Zhihao Liu, Hui Yin, Xinyi Wu, Zhenyao Wu, Yang Mi, and Song Wang. From shadow generation to shadow removal. In CVPR, pages 4927-4936, 2021. 2 +[32] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, pages 11461-11471, 2022. 2 + +[33] Junfeng Lyu, Zhibo Wang, and Feng Xu. Portrait eyeglasses and shadow removal by leveraging 3d synthetic data. In CVPR, pages 3429-3439, 2022. 2 +[34] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In CVPR, pages 2536-2544, 2016. 2 +[35] Liangqiong Qu, Jiandong Tian, Shengfeng He, Yandong Tang, and Rynson WH Lau. Deshadownet: A multi-context embedding deep network for shadow removal. In CVPR, pages 4067-4075, 2017. 2 +[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 8 +[37] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 conference proceedings, pages 1-10, 2022. 2 +[38] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713-4726, 2022. 5 +[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.4 +[40] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In WACV, pages 2149–2159, 2022. 2 +[41] Chengrui Wang and Weihong Deng. Representative forgery mining for fake face detection. In CVPR, pages 14923-14932, 2021. 1 +[42] Jifeng Wang, Xiang Li, and Jian Yang. Stacked conditional generative adversarial networks for jointly learning shadow detection and shadow removal. In CVPR, pages 1788-1797, 2018. 2 +[43] Joshua Weir, Junhong Zhao, Andrew Chalmers, and Taehyun Rhee. Deep portrait delighting. In ECCV, pages 423-439. Springer, 2022. 2, 6 +[44] Yonatan Wexler, Eli Shechtman, and Michal Irani. Spacetime completion of video. IEEE Transactions on Pattern Analysis and Machine Intelligence, 29(3):463-476, 2007. 2 +[45] Jie Xiao, Xueyang Fu, Yurui Zhu, Dong Li, Jie Huang, Kai Zhu, and Zheng-Jun Zha. Homoformer: Homogenized transformer for image shadow removal. In CVPR, pages 25617-25626, 2024. 2, 5, 6 +[46] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In CVPR, pages 22428-22437, 2023. 2 +[47] Yanbo Xu, Yueqin Yin, Liming Jiang, Qianyi Wu, Chengyao Zheng, Chen Change Loy, Bo Dai, and Wayne Wu. Transeditor: Transformer-based dual-space gan for highly controllable facial editing. In CVPR, pages 7683-7692, 2022. 1 + +[48] Shuai Yang, Liming Jiang, Ziwei Liu, and Chen Change Loy. Pastiche master: Exemplar-based high-resolution portrait style transfer. In CVPR, pages 7693-7702, 2022. 8 +[49] Ran Yi, Yong-Jin Liu, Yu-Kun Lai, and Paul L Rosin. Unpaired portrait drawing generation via asymmetric cycle mapping. In CVPR, pages 8217-8225, 2020. 3, 7, 8 +[50] Jae Shin Yoon, Zhixin Shu, Mengwei Ren, Cecilia Zhang, Yannick Hold-Geoffroy, Krishna Kumar Singh, and He Zhang. Generative portrait shadow removal. ACM Transactions on Graphics (TOG), 43(6):1-13, 2024. 2, 6 +[51] Jiahui Yu, Zhe Lin, Jamei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Generative image inpainting with contextual attention. In CVPR, pages 5505-5514, 2018. 2 +[52] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Scaling in-the-wild training for diffusion-based illumination harmonization and editing by imposing consistent light transport. In ICLR. 1, 2, 5, 6 +[53] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836-3847, 2023. 8 +[54] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 3 +[55] Shifeng Zhang, Cheng Chi, Zhen Lei, and Stan Z Li. Refineface: Refinement neural network for high performance face detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(11):4008-4020, 2020. 1 +[56] Xuaner Zhang, Jonathan T Barron, Yun-Ta Tsai, Rohit Pandey, Xiuming Zhang, Ren Ng, and David E Jacobs. Portrait shadow manipulation. ACM Transactions on Graphics (TOG), 39(4):78-1, 2020. 2, 3, 5, 6, 7 +[57] Rongjia Zheng, Qing Zhang, Yongwei Nie, and Wei-Shi Zheng. When shadow removal meets intrinsic image decomposition: A joint learning framework using unpaired data. In AAAI, pages 10599-10607, 2025. 2 +[58] Hao Zhou, Sunil Hadap, Kalyan Sunkavalli, and David W Jacobs. Deep single-image portrait relighting. In ICCV, pages 7194-7202, 2019. 3 +[59] Shipeng Zhu, Pengfei Fang, Chenjie Zhu, Zuoyan Zhao, Qiang Xu, and Hui Xue. Text image inpainting via global structure-guided diffusion models. In AAAI, pages 7775-7783, 2024. 2 +[60] Yurui Zhu, Jie Huang, Xueyang Fu, Feng Zhao, Qibin Sun, and Zheng-Jun Zha. Bijective mapping network for shadow removal. In CVPR, pages 5627-5636, 2022. 2 +[61] Yurui Zhu, Zeyu Xiao, Yanchi Fang, Xueyang Fu, Zhiwei Xiong, and Zheng-Jun Zha. Efficient model-driven network for shadow removal. In AAAI, pages 3635-3643, 2022. 2 \ No newline at end of file diff --git a/2025/Structure-Guided Diffusion Models for High-Fidelity Portrait Shadow Removal/images.zip b/2025/Structure-Guided Diffusion Models for High-Fidelity Portrait Shadow Removal/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..92ca286bfc9d3398b9d3144c871cb3bfdd92b089 --- /dev/null +++ b/2025/Structure-Guided Diffusion Models for High-Fidelity Portrait Shadow Removal/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:005070401ae6b8622055ffb734dd6b141cecdc73e002ce3e57f399a666879ede +size 1251197 diff --git a/2025/Structure-Guided Diffusion Models for High-Fidelity Portrait Shadow Removal/layout.json b/2025/Structure-Guided Diffusion Models for High-Fidelity Portrait Shadow Removal/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0956a07bf2cde6b513951a5483a0f5c3e6ebd3cf --- /dev/null +++ b/2025/Structure-Guided Diffusion Models for High-Fidelity Portrait Shadow Removal/layout.json @@ -0,0 +1,10099 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 103, + 550, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 103, + 550, + 120 + ], + "spans": [ + { + "bbox": [ + 59, + 103, + 550, + 120 + ], + "type": "text", + "content": "Structure-Guided Diffusion Models for High-Fidelity Portrait Shadow Removal" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 129, + 141, + 481, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 141, + 481, + 158 + ], + "spans": [ + { + "bbox": [ + 129, + 141, + 481, + 158 + ], + "type": "text", + "content": "Wanchang Yu" + }, + { + "bbox": [ + 129, + 141, + 481, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 129, + 141, + 481, + 158 + ], + "type": "text", + "content": " Qing Zhang" + }, + { + "bbox": [ + 129, + 141, + 481, + 158 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 129, + 141, + 481, + 158 + ], + "type": "text", + "content": " Rongjia Zheng" + }, + { + "bbox": [ + 129, + 141, + 481, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 129, + 141, + 481, + 158 + ], + "type": "text", + "content": " Wei-Shi Zheng" + }, + { + "bbox": [ + 129, + 141, + 481, + 158 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 117, + 157, + 493, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 157, + 493, + 171 + ], + "spans": [ + { + "bbox": [ + 117, + 157, + 493, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 117, + 157, + 493, + 171 + ], + "type": "text", + "content": "School of Computer Science and Engineering, Sun Yat-sen University, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 171, + 542, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 171, + 542, + 186 + ], + "spans": [ + { + "bbox": [ + 67, + 171, + 542, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 171, + 542, + 186 + ], + "type": "text", + "content": "Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 56, + 200, + 140, + 365 + ], + "blocks": [ + { + "bbox": [ + 56, + 200, + 140, + 365 + ], + "lines": [ + { + "bbox": [ + 56, + 200, + 140, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 200, + 140, + 365 + ], + "type": "image", + "image_path": "3cf8d30aa782822b05677e0f926112a293b96c860170714cad13c35ae3d8adcf.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 367, + 108, + 376 + ], + "lines": [ + { + "bbox": [ + 88, + 367, + 108, + 376 + ], + "spans": [ + { + "bbox": [ + 88, + 367, + 108, + 376 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 140, + 200, + 222, + 365 + ], + "blocks": [ + { + "bbox": [ + 140, + 200, + 222, + 365 + ], + "lines": [ + { + "bbox": [ + 140, + 200, + 222, + 365 + ], + "spans": [ + { + "bbox": [ + 140, + 200, + 222, + 365 + ], + "type": "image", + "image_path": "c3298ebe9d3a8845c02a3cbf22a06de24fb03e23211ea83475f2a9246b5941c5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 367, + 216, + 376 + ], + "lines": [ + { + "bbox": [ + 146, + 367, + 216, + 376 + ], + "spans": [ + { + "bbox": [ + 146, + 367, + 216, + 376 + ], + "type": "text", + "content": "ShadowDiffusion [11]" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 223, + 200, + 304, + 365 + ], + "blocks": [ + { + "bbox": [ + 223, + 200, + 304, + 365 + ], + "lines": [ + { + "bbox": [ + 223, + 200, + 304, + 365 + ], + "spans": [ + { + "bbox": [ + 223, + 200, + 304, + 365 + ], + "type": "image", + "image_path": "161affd385f29f95cd7b767d7c1d35cc417e9480a68c18faed34ba5a79ba1894.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 242, + 367, + 285, + 376 + ], + "lines": [ + { + "bbox": [ + 242, + 367, + 285, + 376 + ], + "spans": [ + { + "bbox": [ + 242, + 367, + 285, + 376 + ], + "type": "text", + "content": "IC-light [52]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 306, + 200, + 388, + 365 + ], + "blocks": [ + { + "bbox": [ + 306, + 200, + 388, + 365 + ], + "lines": [ + { + "bbox": [ + 306, + 200, + 388, + 365 + ], + "spans": [ + { + "bbox": [ + 306, + 200, + 388, + 365 + ], + "type": "image", + "image_path": "c3c399eef96a8188dccafb0ddbae2b8fc216a3681ff6c74e7a5b93cd5652bf8a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 367, + 362, + 376 + ], + "lines": [ + { + "bbox": [ + 331, + 367, + 362, + 376 + ], + "spans": [ + { + "bbox": [ + 331, + 367, + 362, + 376 + ], + "type": "text", + "content": "BSR [28]" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 387, + 555, + 409 + ], + "lines": [ + { + "bbox": [ + 55, + 387, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 555, + 409 + ], + "type": "text", + "content": "Figure 1. Comparison on portrait shadow removal. Our method allows to robustly remove facial cast shadows while faithfully preserving the portrait identities as well as structure details in shadow regions and also ensuring the overall naturalness." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 389, + 200, + 471, + 365 + ], + "blocks": [ + { + "bbox": [ + 389, + 200, + 471, + 365 + ], + "lines": [ + { + "bbox": [ + 389, + 200, + 471, + 365 + ], + "spans": [ + { + "bbox": [ + 389, + 200, + 471, + 365 + ], + "type": "image", + "image_path": "ccc485ed95a40fca12f4db7182d0ba7362811c9b4641daf2ce1b843cb5dda6a6.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 367, + 438, + 376 + ], + "lines": [ + { + "bbox": [ + 421, + 367, + 438, + 376 + ], + "spans": [ + { + "bbox": [ + 421, + 367, + 438, + 376 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 471, + 200, + 554, + 365 + ], + "blocks": [ + { + "bbox": [ + 471, + 200, + 554, + 365 + ], + "lines": [ + { + "bbox": [ + 471, + 200, + 554, + 365 + ], + "spans": [ + { + "bbox": [ + 471, + 200, + 554, + 365 + ], + "type": "image", + "image_path": "3e7d2516ef60de9d5d917cd9eab1fc97d84fff5750771760aab23f5e88bb3064.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 506, + 367, + 518, + 376 + ], + "lines": [ + { + "bbox": [ + 506, + 367, + 518, + 376 + ], + "spans": [ + { + "bbox": [ + 506, + 367, + 518, + 376 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 151, + 419, + 200, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 419, + 200, + 431 + ], + "spans": [ + { + "bbox": [ + 151, + 419, + 200, + 431 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 54, + 442, + 297, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 442, + 297, + 694 + ], + "spans": [ + { + "bbox": [ + 54, + 442, + 297, + 694 + ], + "type": "text", + "content": "We present a diffusion-based portrait shadow removal approach that can robustly produce high-fidelity results. Unlike previous methods, we cast shadow removal as diffusion-based inpainting. To this end, we first train a shadow-independent structure extraction network on a real-world portrait dataset with various synthetic lighting conditions, which allows to generate a shadow-independent structure map including facial details while excluding the unwanted shadow boundaries. The structure map is then used as condition to train a structure-guided inpainting diffusion model for removing shadows in a generative manner. Finally, to restore the fine-scale details (e.g., eyelashes, moles and spots) that may not be captured by the structure map, we take the gradients inside the shadow regions as guidance and train a detail restoration diffusion model to refine the shadow removal result. Extensive experiments on the benchmark datasets show that our method clearly outperforms existing methods, and is effective to avoid previously common issues such as facial identity tampering, shadow residual, color distortion, structure blurring, and loss of details. Our code is avail" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 420, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 555, + 445 + ], + "type": "text", + "content": "able at https://github.com/wanchang-yu/Structure-Guided-Diffusion-for-Portrait-Shadow-Removal." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 314, + 465, + 395, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 465, + 395, + 478 + ], + "spans": [ + { + "bbox": [ + 314, + 465, + 395, + 478 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 485, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 485, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 312, + 485, + 555, + 617 + ], + "type": "text", + "content": "Taking selfies is very popular nowadays due to the readily-available cameras on various devices, particularly the mobile phones. However, portrait photos taken in the wild lacking professional lighting conditions often suffer from undesired foreign shadows cast by external objects (see Figure 1 for examples). The presence of facial shadows not only severely degrade the overall visual appeal of the portraits, but also adversely affect many downstream vision tasks, such as face detection and recognition [4, 5, 41, 55], and portrait image editing [21, 27, 47]. Hence, portrait shadow removal techniques are widely required." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 618, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 618, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 618, + 556, + 714 + ], + "type": "text", + "content": "Currently, portrait shadow removal remains a challenge despite the remarkable progress in natural image shadow removal. The reason behind is twofold. First of all, due to factors such as privacy and difficulty in data collection, large-scale paired data for portrait shadow is extremely scarce, which significantly limits the development of portrait shadow removal methods, especially learning-based solutions. On the other hand, as this task is portrait-" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 702, + 234, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 702, + 234, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 702, + 234, + 713 + ], + "type": "text", + "content": "*Corresponding author (zhangq93@mail.sysu.edu.cn)." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11675" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "oriented, it has very low tolerance for visual artifacts, even minor imperfections on facial color and details could make the final result unnatural and unacceptable. Note, although existing software allow users to interactively adjust photos, it is still rather tedious and difficult for non-experts to manually produce high-fidelity portrait shadow removal results." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 146, + 295, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 146, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 146, + 295, + 396 + ], + "type": "text", + "content": "Existing shadow removal methods [11, 26, 45] mainly target natural images due to the availability of large-scale paired datasets. However, these methods basically fail to deal with portrait shadows, as shown in Figures 5 and 6. As there is no accessible large-scale paired datasets for portrait shadow removal, current methods in this field focus primarily on constructing synthetic paired data to enable supervised learning of shadow removal [28, 33, 52, 56]. Suffering from the domain gap between real-world portraits and synthetic data, these methods typically do not work well for real-world portraits with complex shadows. To avoid the domain gap issue led by synthetic training data, a zero-shot solution is introduced in [14], which proposed to leverage the generative prior from pre-trained StyleGAN2 [24] to recover shadow-free portrait through test-time GAN inversion. However, this method tends to modify face identity and incur unnatural results. Note, although recent image relighting techniques [8, 16, 43, 50] have shown remarkable progress, they struggle to enable high-fidelity portrait shadow removal due to the inherent difficulty in estimating ambient light and facial geometry from a single image." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 399, + 295, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 295, + 649 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 295, + 649 + ], + "type": "text", + "content": "To robustly produce high-fidelity portrait shadow removal results, we in this work present a diffusion-based approach. Particularly, we propose to cast portrait shadow removal as diffusion-based inpainting, so as to take full advantage of the strong generation capability of diffusion models. To this end, we first train a shadow-independent structure extraction network (SE-Net) on a dataset of paired real-world portraits with various synthetic lighting conditions, from which we obtain a structure map containing facial details while excluding the shadow boundaries. With the structure map, we then train a structure-guided inpainting diffusion model by enforcing reconstruction of randomly masked shadow-free portraits. Finally, to restore the fine-scale facial details that may not be captured by the structure map, we further refine the shadow removal result by a detail restoration diffusion model under the guidance of the gradients in the original shadow regions. Note, although our method requires a test shadow mask as input, it is highly robust to inaccurate masks, requiring only that the mask covers the shadow regions, even if it includes nonshadow regions. The major contributions of this work are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 654, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 654, + 294, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 294, + 689 + ], + "type": "text", + "content": "- We present a simple yet effective unsupervised approach that addresses portrait shadow removal from a perspective of diffusion-based inpainting." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "type": "text", + "content": "- We introduce a relighting-based data synthesis strategy for constructing the training data, and develop a shadow" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 323, + 72, + 492, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 72, + 492, + 83 + ], + "spans": [ + { + "bbox": [ + 323, + 72, + 492, + 83 + ], + "type": "text", + "content": "independent structure extraction network." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 84, + 553, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 84, + 553, + 107 + ], + "spans": [ + { + "bbox": [ + 314, + 84, + 553, + 107 + ], + "type": "text", + "content": "- We show that our method outperforms state-of-the-art methods for portrait shadow removal." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 118, + 400, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 118, + 400, + 131 + ], + "spans": [ + { + "bbox": [ + 314, + 118, + 400, + 131 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 140, + 553, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 140, + 553, + 439 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 553, + 439 + ], + "type": "text", + "content": "Shadow removal. Early methods typically rely on various handcrafted priors or assumptions to work [7, 9, 12]. Recent methods are basically deep learning based [3, 10, 11, 17, 20, 22, 23, 25, 29, 31, 57, 60, 61], benefiting mostly from the availability of large-scale paired datasets such as ISTD [42] and SRD [35]. However, as these methods are tailored for natural images, they are not so effective in dealing with portraits. In contrast to natural images, portrait shadow removal has received much less research attention, due to the lack of large-scale paired datasets. To overcome the data scarcity issue, some methods [28, 56] proposed to synthesize paired shadow and shadow-free portrait images for training. As their data synthesis strategies involve only simple adjustments on brightness and saturation, without considering facial geometry and albedo, their methods typically do not work well for real-world portraits. He et al. [14] achieved zero-shot portrait shadow removal by utilizing generative prior from StyleGAN2, but this method may result in undesired modifications to facial details and thus fails to preserve the face identity. More recently, image relighting is also used for shadow removal [8, 16, 43, 50, 52]. However, this method tends to induce unnatural results because it is very challenging to obtain accurate prediction of diffuse lighting and albedo from a single portrait image." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 443, + 555, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 555, + 635 + ], + "type": "text", + "content": "Image inpainting. Before the era of deep learning, patch-based self-similarity of natural images has long been the foundation for image inpainting [2, 44]. The emergence of deep learning has then led to great progress in image inpainting [18, 34, 40, 51], as it is quite easy to construct paired training data. Recently, diffusion model has been a preferable choice for image inpainting [1, 19, 32, 37, 46, 59], due to its impressive generation capability. Note, naive applying diffusion models to portrait shadow removal fail to produce visually natural results with unchanged portrait identity. A recent work closely related to our method also shows that it is feasible to remove shadows via image inpainting [26]. However, it does not work well for portrait shadow removal (see Figures 5 and 6), even we fine-tuned their model on real-world paired shadow and shadow-free portrait images provided in [56]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 645, + 394, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 645, + 394, + 658 + ], + "spans": [ + { + "bbox": [ + 314, + 645, + 394, + 658 + ], + "type": "text", + "content": "3. Our Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "Given a portrait image with cast shadows, our method aims to remove the facial shadows while faithfully preserving the facial details in the shadow regions. Figure 2 presents the overall inference pipeline of our method." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "11676" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 553, + 220 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 553, + 220 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 553, + 220 + ], + "type": "image", + "image_path": "a7303ded20a690d3f0212416f31630b3978b34c76ed7a4959cdd19b167a9cbc4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 222, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 222, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 222, + 555, + 266 + ], + "type": "text", + "content": "Figure 2. Inference pipeline of our method. Given a input portrait image and the corresponding shadow mask, we first feed the input image to the shadow-independent structure extraction network (SE-Net) to get a structure map. Then, we perform shadow removal by inpainting the shadow regions with a diffusion model conditioned on the structure map. Finally, the shadow removal output is refined by a detail restoration diffusion model to restore the fine-scale facial details that may not be captured by the strcutre map." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 286, + 294, + 438 + ], + "blocks": [ + { + "bbox": [ + 58, + 286, + 294, + 438 + ], + "lines": [ + { + "bbox": [ + 58, + 286, + 294, + 438 + ], + "spans": [ + { + "bbox": [ + 58, + 286, + 294, + 438 + ], + "type": "image", + "image_path": "9a86e93f454fd316c70ab47d6505efd197a8b51107ac8b0a8b7e6cdd3b100bb4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 441, + 295, + 475 + ], + "lines": [ + { + "bbox": [ + 55, + 441, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 295, + 475 + ], + "type": "text", + "content": "Figure 3. Training pipeline of SE-Net. Note, we adopt a trained PDG model for structure extraction from [49], while it is not invariant to illumination changes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 498, + 278, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 278, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 278, + 510 + ], + "type": "text", + "content": "3.1. Shadow-Independent Structure Extraction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 516, + 295, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 295, + 624 + ], + "type": "text", + "content": "As shown in Figure 2, the key to obtaining high-quality portrait shadow removal through inpainting is to acquire a structure map free of shadow boundaries for guiding the inpainting process. However, although existing structure/edge extraction methods are quite effective in identifying edges and structures, they are sensitive to illumination discontinuities and will inevitably include shadow boundaries in their results (see Figure 7). Therefore, we propose to build a shadow-independent structure extraction network." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "Training data synthesis. To enable effective training of the shadow-independent structure extraction network, as shown in Figure 3, we construct a synthetic paired dataset of well-lit real-world portraits without foreign shadows on faces and the corresponding relighted portraits with various illumination discontinuities, based on the shadow-free portraits in the CelebA dataset [30]. Specifically, for a well-lit portrait" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 287, + 555, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 555, + 336 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 287, + 555, + 336 + ], + "type": "text", + "content": ", we first apply an off-the-shelf physics-based relighting method [58] over it to get a relighted image " + }, + { + "bbox": [ + 313, + 287, + 555, + 336 + ], + "type": "inline_equation", + "content": "I_{relit}" + }, + { + "bbox": [ + 313, + 287, + 555, + 336 + ], + "type": "text", + "content": ", which is then combined with the original portrait " + }, + { + "bbox": [ + 313, + 287, + 555, + 336 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 287, + 555, + 336 + ], + "type": "text", + "content": " to produce a synthetic image " + }, + { + "bbox": [ + 313, + 287, + 555, + 336 + ], + "type": "inline_equation", + "content": "I_{syn}" + }, + { + "bbox": [ + 313, + 287, + 555, + 336 + ], + "type": "text", + "content": " with illumination discontinuities by" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 361, + 347, + 555, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 347, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 361, + 347, + 555, + 361 + ], + "type": "interline_equation", + "content": "I _ {s y n} = M \\odot I _ {r e l i t} + (1 - M) \\odot I, \\tag {1}", + "image_path": "456f1a2f1bbaa017add070216a994d88f26bfecc56b11898a0db866e2b85f747.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 370, + 554, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 554, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 554, + 431 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 370, + 554, + 431 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 370, + 554, + 431 + ], + "type": "text", + "content": " is a random facial mask, and " + }, + { + "bbox": [ + 313, + 370, + 554, + 431 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 370, + 554, + 431 + ], + "type": "text", + "content": " denotes elementwise multiplication. Note, our synthetic data is only used for training the SE-Net and the subsequent detail resotration diffusion model, rather than for directly training shadow removal model as done in [56]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 436, + 555, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 555, + 520 + ], + "type": "text", + "content": "Training losses. As shown in Figure 3, we adopt a trained structure extraction model from [49] (referred to as PDG in paper) to produce structure map for the original well-lit portrait, which is treated as pseudo ground-truth for supervising the training of our SE-Net. By respectively denoting our SE-Net and the PDG model as " + }, + { + "bbox": [ + 313, + 436, + 555, + 520 + ], + "type": "inline_equation", + "content": "G_{s}" + }, + { + "bbox": [ + 313, + 436, + 555, + 520 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 436, + 555, + 520 + ], + "type": "inline_equation", + "content": "G_{p}" + }, + { + "bbox": [ + 313, + 436, + 555, + 520 + ], + "type": "text", + "content": ", the training losses of our SE-Net are defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 343, + 529, + 553, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 529, + 553, + 562 + ], + "spans": [ + { + "bbox": [ + 343, + 529, + 553, + 562 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\mathcal {L} _ {\\text {r e c}} = \\| G _ {s} (I _ {\\text {s y n}}) - G _ {p} (I) \\| _ {1}, \\\\ \\mathcal {L} _ {\\text {p e r c e p t u a l}} = \\mathcal {L} _ {\\text {L P I P S}} \\left(G _ {s} (I _ {\\text {s y n}}), G _ {p} (I)\\right), \\end{array} \\right. \\tag {2}", + "image_path": "051f63e05a085848f5980b591915098d99a0e08f1d3c163702a98d9573a48947.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 555, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 654 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 570, + 555, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{LPIPS}" + }, + { + "bbox": [ + 313, + 570, + 555, + 654 + ], + "type": "text", + "content": " refers to the learned perceptual image patch similarity [54]. Besides, we also utilize a discriminator to determine whether the output of SE-Net belongs to the structure image domain, and further define an adversarial loss " + }, + { + "bbox": [ + 313, + 570, + 555, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{GAN}" + }, + { + "bbox": [ + 313, + 570, + 555, + 654 + ], + "type": "text", + "content": " (see supplementary material for details). In summary, the overall loss function for training the SE-Net is given by:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 344, + 666, + 553, + 679 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 666, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 344, + 666, + 553, + 679 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {r e c}} + \\lambda_ {1} \\mathcal {L} _ {\\text {p e r c e p t u a l}} + \\lambda_ {2} \\mathcal {L} _ {\\text {G A N}}, \\tag {3}", + "image_path": "ba99f02d5ce46576977d01c20dc56665ccfc2142da9b2a1e0968c1cb232eb1fb.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "where we empirically set " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_1 = 0.5" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_1 = 0.25" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": ", which produce good results, as demonstrated in Figure 7." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11677" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 92, + 286, + 230 + ], + "blocks": [ + { + "bbox": [ + 99, + 74, + 246, + 86 + ], + "lines": [ + { + "bbox": [ + 99, + 74, + 246, + 86 + ], + "spans": [ + { + "bbox": [ + 99, + 74, + 246, + 86 + ], + "type": "text", + "content": "Structure-guided Portrait Inpainting" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 92, + 286, + 230 + ], + "lines": [ + { + "bbox": [ + 61, + 92, + 286, + 230 + ], + "spans": [ + { + "bbox": [ + 61, + 92, + 286, + 230 + ], + "type": "image", + "image_path": "b4d96f412dedaf6e5e4e50b52fbfb67030bf64e18a675e3f1e14f28033e1b8d6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 294, + 90, + 548, + 228 + ], + "blocks": [ + { + "bbox": [ + 349, + 74, + 492, + 86 + ], + "lines": [ + { + "bbox": [ + 349, + 74, + 492, + 86 + ], + "spans": [ + { + "bbox": [ + 349, + 74, + 492, + 86 + ], + "type": "text", + "content": "Gradient-guided Detail Restoration" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 294, + 90, + 548, + 228 + ], + "lines": [ + { + "bbox": [ + 294, + 90, + 548, + 228 + ], + "spans": [ + { + "bbox": [ + 294, + 90, + 548, + 228 + ], + "type": "image", + "image_path": "92f6fff8c2118d35e7a7305efd6deaa999a3b20db6f72abbdb1a3bde4d543cb3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 236, + 555, + 258 + ], + "lines": [ + { + "bbox": [ + 55, + 236, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 555, + 258 + ], + "type": "text", + "content": "Figure 4. Training pipeline of our diffusion models. Here we introduce how our structure-guided portrait inpainting diffusion model and gradient-guided detail restoration diffusion model are trained." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 278, + 254, + 292 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 254, + 292 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 254, + 292 + ], + "type": "text", + "content": "3.2. Structure-Guided Portrait Inpainting" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 296, + 296, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 296, + 296, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 296, + 332 + ], + "type": "text", + "content": "As shown in Figure 2, with structure map, the next step is to inpaint the shadow regions. To this end, we develop a structure-guided portrait inpainting diffusion model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 337, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 296, + 479 + ], + "type": "text", + "content": "Structure-aware denoising. As shown in Figure 4, we start with a shadow-free portrait as input and then randomly mask some facial regions to create a masked input, followed by a diffusion model to reconstruct the original input by inpainting the masked regions. The inpainting is rooted in the forward diffusion process, where the diffusion model learns a Markovian chain by progressively corrupting the target image " + }, + { + "bbox": [ + 55, + 337, + 296, + 479 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 337, + 296, + 479 + ], + "type": "text", + "content": " (namely the masked input) with Gaussian noise at each timestep " + }, + { + "bbox": [ + 55, + 337, + 296, + 479 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 337, + 296, + 479 + ], + "type": "text", + "content": ", such that it is transformed into a Gaussian distribution (referred to as " + }, + { + "bbox": [ + 55, + 337, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 55, + 337, + 296, + 479 + ], + "type": "text", + "content": " in the following). Specifically, after reparameterization [15], our forward process obtains " + }, + { + "bbox": [ + 55, + 337, + 296, + 479 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 55, + 337, + 296, + 479 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 337, + 296, + 479 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 337, + 296, + 479 + ], + "type": "text", + "content": " in a single step by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 120, + 484, + 294, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 484, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 120, + 484, + 294, + 498 + ], + "type": "interline_equation", + "content": "x _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\tag {4}", + "image_path": "e397a72cbb4053938c64fe71db2395a8cac1a98c9e92d82eead356d9acd33649.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 502, + 295, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 295, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 295, + 537 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 502, + 295, + 537 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t = \\prod_{i=0}^t \\alpha_i" + }, + { + "bbox": [ + 55, + 502, + 295, + 537 + ], + "type": "text", + "content": " is the cumulative product of forward noise weight parameters " + }, + { + "bbox": [ + 55, + 502, + 295, + 537 + ], + "type": "inline_equation", + "content": "\\alpha_i = 1 - \\beta_i" + }, + { + "bbox": [ + 55, + 502, + 295, + 537 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 55, + 502, + 295, + 537 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 55, + 502, + 295, + 537 + ], + "type": "text", + "content": " denotes random Gaussian noise." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 539, + 296, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 296, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 296, + 609 + ], + "type": "text", + "content": "As for the reverse denoising process, we use the sampling strategy of DDIM [39] for acceleration. Note, unlike unconditional image generation, our goal is to faithfully reconstruct the original unmasked portrait, the reverse denoising process is thus conditioned on a structure map " + }, + { + "bbox": [ + 55, + 539, + 296, + 609 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 539, + 296, + 609 + ], + "type": "text", + "content": " and defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 58, + 612, + 294, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 612, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 58, + 612, + 294, + 647 + ], + "type": "interline_equation", + "content": "x _ {t - 1} = \\sqrt {\\bar {\\alpha} _ {t - 1}} \\left(\\frac {x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\cdot \\mathbf {e} _ {t}}{\\sqrt {\\bar {\\alpha} _ {t}}}\\right) + \\sqrt {1 - \\bar {\\alpha} _ {t - 1}} \\cdot \\mathbf {e} _ {t}, \\tag {5}", + "image_path": "569c52828810f1fcb355fdb51aa4dbaa7168dc2475eb80a64b5e2b2d62735824.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 648, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 295, + 696 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 648, + 295, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_t = \\epsilon_\\theta (x_t,I_M,S,M,t)" + }, + { + "bbox": [ + 55, + 648, + 295, + 696 + ], + "type": "text", + "content": " refers to the noise estimated by a deep neural network " + }, + { + "bbox": [ + 55, + 648, + 295, + 696 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 648, + 295, + 696 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 648, + 295, + 696 + ], + "type": "inline_equation", + "content": "I_M" + }, + { + "bbox": [ + 55, + 648, + 295, + 696 + ], + "type": "text", + "content": " is the masked input created from a random mask " + }, + { + "bbox": [ + 55, + 648, + 295, + 696 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 648, + 295, + 696 + ], + "type": "text", + "content": ". Following [15], the training objective for diffusion model can be simplified as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 94, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 94, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i f f}} (\\epsilon) = \\| \\epsilon - \\epsilon_ {\\theta} \\left(x _ {t}, I _ {M}, S, M, t\\right) \\| _ {2} ^ {2}, \\tag {6}", + "image_path": "58e2d2b65c314c452f1d97bfa0365f182cfc1b62703eddb57dfba249f7550e9b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 278, + 554, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 554, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 554, + 339 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 278, + 554, + 339 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 278, + 554, + 339 + ], + "type": "text", + "content": " is the ground truth noise, while " + }, + { + "bbox": [ + 313, + 278, + 554, + 339 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,I_M,S,M,t)" + }, + { + "bbox": [ + 313, + 278, + 554, + 339 + ], + "type": "text", + "content": " is the noise predicted by the network. As shown in Figure 2, a facial shadow mask is required at the inference stage, which is a dilated version of the facial shadow detection result from the trained model of [6]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "type": "text", + "content": "Shadow mask refinement. To enhance the robustness of our method to inaccurate shadow mask, we propose to refine the mask based on the shadow removal result generated by the inpainting diffusion model. Specifically, we first compute the absolute difference between the input image and the shadow removal result, and then perform Otsu-based thresholding to get a refined shadow mask excluding all the non-shadow regions. With the refined mask, we further update the shadow removal result by combining it with the input image, so as to ensure that non-shadow regions in the result are unchanged." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 510, + 506, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 510, + 506, + 521 + ], + "spans": [ + { + "bbox": [ + 314, + 510, + 506, + 521 + ], + "type": "text", + "content": "3.3. Gradient-Guided Detail Restoration" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "As the original PDG model is designed for generating portrait drawing formed by large-scale structures, thus our SE-Net is similarly not so effective in capturing fine-scale portrait details, such as eyelashes, moles, and spots (see the bottom row in Figure 7 for example). To address this issue, we design a detail restoration diffusion model to refine the shadow removal result generated by the inpainting diffusion model, where we directly take the gradients in the shadow regions as guidance to restore the fine-scale details. Figure 4 describes how we train the detail restoration diffusion model. Note, gradient-aware denoising process is similar to the structure-aware denoising described above, and the two share the same forward diffusion and reverse denoising formulations described in Equations 4 and 5, with the latter conditioned on gradient map rather than structure map." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11678" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 553, + 326 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 553, + 326 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 553, + 326 + ], + "type": "image", + "image_path": "8e134908d42ac20fe5bd0c032fc838bfc124751ae2886510fd474bcfb1e911c9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 328, + 509, + 340 + ], + "lines": [ + { + "bbox": [ + 99, + 328, + 509, + 340 + ], + "spans": [ + { + "bbox": [ + 99, + 328, + 509, + 340 + ], + "type": "text", + "content": "Figure 5. Visual comparison with state-of-the-art methods on test images from the dataset provided in [56]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 346, + 555, + 541 + ], + "blocks": [ + { + "bbox": [ + 58, + 346, + 555, + 541 + ], + "lines": [ + { + "bbox": [ + 58, + 346, + 555, + 541 + ], + "spans": [ + { + "bbox": [ + 58, + 346, + 555, + 541 + ], + "type": "table", + "html": "
MethodAllShadow regionNon-shadow region
SSIM↑LPIPS↓RMSE↓SSIM↑LPIPS↓RMSE↓SSIM↑LPIPS↓RMSE↓
ShadowDiffusion [11]0.6500.17738.2460.9010.03715.5610.5790.22352.276
Inpaint4Shadow [26]0.7660.09422.7090.9100.03317.2500.7080.16445.456
ShadowFormer [10]0.7720.08621.4130.9260.02715.2880.7230.14942.341
HomoFormer [45]0.7860.10021.7240.9130.04116.2480.7110.16742.950
PSM [56]0.7820.074-------
UPSR [14]0.7310.10926.9150.9000.03920.6590.6680.18845.391
BSR [28]0.6050.11828.5320.9100.03615.1930.5480.17143.627
IC-light [52]0.5140.27862.5390.9160.05733.0610.4780.31969.208
Ours w/o structure guidance0.7570.10123.5980.8980.04318.5140.7140.13839.000
Ours w/ PDG predicted structure0.7790.09219.9630.9190.03514.0620.7250.13137.387
Ours w/ data synthesis in PSM0.7850.08619.7970.9250.03213.7520.7290.13638.485
Ours w/o detail-restoration0.8140.06218.5370.9430.02012.7480.7470.12036.326
Ours (full method)0.8300.05617.1620.9730.01110.1980.7780.09134.573
", + "image_path": "576c717c0436b1b027f552c6dbc6d76fb5a4e0dcad35584dfc65a1c4555dbcc7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 109, + 544, + 500, + 555 + ], + "lines": [ + { + "bbox": [ + 109, + 544, + 500, + 555 + ], + "spans": [ + { + "bbox": [ + 109, + 544, + 500, + 555 + ], + "type": "text", + "content": "Table 1. Quantitative comparison on the real-world portrait shadow removal dataset provided in [56]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 580, + 251, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 251, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 251, + 594 + ], + "type": "text", + "content": "3.4. Implementation and Training Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 599, + 295, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 295, + 695 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 295, + 695 + ], + "type": "text", + "content": "Network architecture. The SE-Net is an encoder-decoder network, which starts with a flat convolution and two down convolution blocks, followed by nine residual blocks, and finally two up convolution blocks and a final convolution layer. Please see also the supplementary material for visual illustration of the architecture of SE-Net. As for the inpainting diffusion model and the detail-restoration diffusion model, we follow [15] to adopt the same U-Net architecture." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 700, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 700, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 700, + 295, + 713 + ], + "type": "text", + "content": "Training details. To train SE-Net, we synthesize 20,000" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "text", + "content": "portrait images with diverse lighting conditions based on the CelebA dataset [30], and train it for 3 epochs with a batch size of 8 and a learning rate of " + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "inline_equation", + "content": "1.5 \\times 10^{-5}" + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "text", + "content": ". As for the inpainting diffusion model, we train it on the FFHQ dataset for 300 epochs with a batch size of 32, and train the detail restoration diffusion model on the same synthetic dataset as the SE-Net for 200 epochs. We build our model on PyTorch and use the Adam optimizer with momentum parameters of (0.9, 0.999) and an initial learning rate of " + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "text", + "content": ". Following [38], we employ the Kaiming initialization [13] to initialize the weights of our model and apply 0.9999 Expo" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11679" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 553, + 324 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 553, + 324 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 553, + 324 + ], + "type": "image", + "image_path": "e4ac4d3cf14c4ef937c8b88e60c660008fd027e29768ef42ee628e1ebbf2110e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 325, + 496, + 337 + ], + "lines": [ + { + "bbox": [ + 113, + 325, + 496, + 337 + ], + "spans": [ + { + "bbox": [ + 113, + 325, + 496, + 337 + ], + "type": "text", + "content": "Figure 6. Visual comparison with state-of-the-art methods on test images from our collected dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 357, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 296, + 418 + ], + "type": "text", + "content": "nential Moving Average (EMA) in all our experiments. We use 2000 diffusion steps and a noise schedule " + }, + { + "bbox": [ + 55, + 357, + 296, + 418 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 55, + 357, + 296, + 418 + ], + "type": "text", + "content": " linearly increasing from " + }, + { + "bbox": [ + 55, + 357, + 296, + 418 + ], + "type": "inline_equation", + "content": "1\\times 10^{-6}" + }, + { + "bbox": [ + 55, + 357, + 296, + 418 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 357, + 296, + 418 + ], + "type": "inline_equation", + "content": "1\\times 10^{-2}" + }, + { + "bbox": [ + 55, + 357, + 296, + 418 + ], + "type": "text", + "content": " for training. During inference, we perform both the inpainting diffusion and detail restoration diffusion for 40 denoising steps." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 431, + 136, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 136, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 136, + 445 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 452, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 452, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 452, + 296, + 571 + ], + "type": "text", + "content": "Datasets. Following previous portrait shadow removal approaches [14, 28], we evaluate our method on the only publicly available real-world paired dataset provided in PSM [56], which contains 9 subjects and 100 shadowed portrait images in varying poses, shadow shapes, illumination conditions, and shadow types. To better examine the effectiveness of our method, we additionally construct a dataset of 200 portraits with diverse facial shadows. Note, we only conduct visual comparison on our collected dataset due to the lack of ground truth shadow removal results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 578, + 296, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 296, + 650 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 296, + 650 + ], + "type": "text", + "content": "Evaluation metrics. We use three commonly-used metrics to quantitatively evaluate the performance of our method, including SSIM, LPIPS, and the root mean square error (RMSE) computed in the LAB color space for measuring the similarity between the predicted results and the corresponding ground-truth shadow-free images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 659, + 278, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 278, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 278, + 672 + ], + "type": "text", + "content": "4.1. Comparison with State-of-the-art Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Baselines. We compare our method with various state-of-the-art shadow removal methods, including four portrait shadow removal and editing methods, i.e., PSM [56]," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 357, + 555, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 467 + ], + "type": "text", + "content": "UPSR [14], and BSR [28], IC-light [52], as well as four natural image shadow removal methods, i.e., ShadowDiffusion [11], Inpaint4Shadow [26], ShadowFormer [10], and HomoFormer [45]. For fair comparison, we produce their results using publicly-available implementation or trained models provided by the authors with recommended parameter setting. Note that, we do not compare with other recent relighting-based shadow editing methods [8, 16, 43, 50], as their codes and trained models are not yet released." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 475, + 556, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 556, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 556, + 643 + ], + "type": "text", + "content": "Visual comparison. We in Figures 5 and 6 conduct visual comparison between our method and previous shadow removal methods. As shown, shadow removal methods designed for natural images, i.e., Inpaint4Shadow [26], HomoFormer [45], and ShadowDiffusion [11] are not well-suited to portraits and tend to induce obvious shadow residuals and appearance distortions. As for the other three portrait shadow removal methods, UPSR [14] and BSR [28] fail to preserve the facial details and may also lead to shadow residuals and color distortions, while IC-Light [52] struggles to remove the facial shadows due to the inherent difficulty in estimating the ambient light and facial geometry. Please see also the supplementary material for more visual comparison results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Quantitative comparison. In Table 1, we quantitatively compare our method with other methods in terms of the SSIM, LPIPS, and RMSE metrics. As can be seen, our method outperforms all the compared methods on all the three metrics, demonstrating its effectiveness and superior-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11680" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 127, + 140 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 127, + 140 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 127, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 127, + 140 + ], + "type": "image", + "image_path": "d3a1f1b9909763a8bd6c93a664c30f93c5377033abf947f9d3ab94db8cf19860.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 140, + 127, + 210 + ], + "blocks": [ + { + "bbox": [ + 56, + 140, + 127, + 210 + ], + "lines": [ + { + "bbox": [ + 56, + 140, + 127, + 210 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 127, + 210 + ], + "type": "image", + "image_path": "fc40af2c9ab6980918d472d07b257d94a90d98099c1eb677a49c059a770c2382.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 210, + 97, + 217 + ], + "lines": [ + { + "bbox": [ + 86, + 210, + 97, + 217 + ], + "spans": [ + { + "bbox": [ + 86, + 210, + 97, + 217 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 128, + 70, + 198, + 140 + ], + "blocks": [ + { + "bbox": [ + 128, + 70, + 198, + 140 + ], + "lines": [ + { + "bbox": [ + 128, + 70, + 198, + 140 + ], + "spans": [ + { + "bbox": [ + 128, + 70, + 198, + 140 + ], + "type": "image", + "image_path": "7e042422b27826ef4c7bb4600cd59234f09dfc597ea3f9076d05e6db02fe7bf1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 128, + 140, + 198, + 211 + ], + "blocks": [ + { + "bbox": [ + 128, + 140, + 198, + 211 + ], + "lines": [ + { + "bbox": [ + 128, + 140, + 198, + 211 + ], + "spans": [ + { + "bbox": [ + 128, + 140, + 198, + 211 + ], + "type": "image", + "image_path": "861e15e6be71cd23543034cd3a7af6c7455ca23d3d5b780c5bfd5b664dcae781.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 211, + 168, + 217 + ], + "lines": [ + { + "bbox": [ + 157, + 211, + 168, + 217 + ], + "spans": [ + { + "bbox": [ + 157, + 211, + 168, + 217 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 200, + 70, + 269, + 140 + ], + "blocks": [ + { + "bbox": [ + 200, + 70, + 269, + 140 + ], + "lines": [ + { + "bbox": [ + 200, + 70, + 269, + 140 + ], + "spans": [ + { + "bbox": [ + 200, + 70, + 269, + 140 + ], + "type": "image", + "image_path": "39bd9cb93efc8dcd2a1d41d3da183eb2b4bd922088c5f6445639afd8b84dc899.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 200, + 140, + 269, + 211 + ], + "blocks": [ + { + "bbox": [ + 200, + 140, + 269, + 211 + ], + "lines": [ + { + "bbox": [ + 200, + 140, + 269, + 211 + ], + "spans": [ + { + "bbox": [ + 200, + 140, + 269, + 211 + ], + "type": "image", + "image_path": "c54e6a6f03ffcfcda671e32c13b6c6f573e90cff2791161f84df163d7d0208bf.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 211, + 239, + 217 + ], + "lines": [ + { + "bbox": [ + 229, + 211, + 239, + 217 + ], + "spans": [ + { + "bbox": [ + 229, + 211, + 239, + 217 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 270, + 70, + 341, + 140 + ], + "blocks": [ + { + "bbox": [ + 270, + 70, + 341, + 140 + ], + "lines": [ + { + "bbox": [ + 270, + 70, + 341, + 140 + ], + "spans": [ + { + "bbox": [ + 270, + 70, + 341, + 140 + ], + "type": "image", + "image_path": "047d3b8ae34d6c8a75ad7b47ee8d8971d8d1b1f3f0d4b86850f69cdbb8e34daf.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 270, + 140, + 341, + 211 + ], + "blocks": [ + { + "bbox": [ + 270, + 140, + 341, + 211 + ], + "lines": [ + { + "bbox": [ + 270, + 140, + 341, + 211 + ], + "spans": [ + { + "bbox": [ + 270, + 140, + 341, + 211 + ], + "type": "image", + "image_path": "1011ee522bfad66898834852b5f879b146e90a066068f99f4a6d26be2ab18f52.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 299, + 211, + 310, + 217 + ], + "lines": [ + { + "bbox": [ + 299, + 211, + 310, + 217 + ], + "spans": [ + { + "bbox": [ + 299, + 211, + 310, + 217 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 342, + 70, + 411, + 140 + ], + "blocks": [ + { + "bbox": [ + 342, + 70, + 411, + 140 + ], + "lines": [ + { + "bbox": [ + 342, + 70, + 411, + 140 + ], + "spans": [ + { + "bbox": [ + 342, + 70, + 411, + 140 + ], + "type": "image", + "image_path": "2fcd9ba1b290fd6310aa9dc580c47022dda7133249eacbcec59b618e302e000e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 342, + 140, + 411, + 210 + ], + "blocks": [ + { + "bbox": [ + 342, + 140, + 411, + 210 + ], + "lines": [ + { + "bbox": [ + 342, + 140, + 411, + 210 + ], + "spans": [ + { + "bbox": [ + 342, + 140, + 411, + 210 + ], + "type": "image", + "image_path": "4a2571cfe0e8dc6891c964a45972191dbc0f7a7dd1ee83f4be96fdc23af5bd04.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 211, + 381, + 217 + ], + "lines": [ + { + "bbox": [ + 371, + 211, + 381, + 217 + ], + "spans": [ + { + "bbox": [ + 371, + 211, + 381, + 217 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 413, + 70, + 482, + 140 + ], + "blocks": [ + { + "bbox": [ + 413, + 70, + 482, + 140 + ], + "lines": [ + { + "bbox": [ + 413, + 70, + 482, + 140 + ], + "spans": [ + { + "bbox": [ + 413, + 70, + 482, + 140 + ], + "type": "image", + "image_path": "18433bf6eea6c41964b0033ebbf4515de80be3ec799ab2983cf3bde3b36c7d08.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 413, + 140, + 482, + 210 + ], + "blocks": [ + { + "bbox": [ + 413, + 140, + 482, + 210 + ], + "lines": [ + { + "bbox": [ + 413, + 140, + 482, + 210 + ], + "spans": [ + { + "bbox": [ + 413, + 140, + 482, + 210 + ], + "type": "image", + "image_path": "25e235c91028bb90ebd6946430dccc7ba3d4e1c8184f1289ec8c24c159213311.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 211, + 452, + 217 + ], + "lines": [ + { + "bbox": [ + 443, + 211, + 452, + 217 + ], + "spans": [ + { + "bbox": [ + 443, + 211, + 452, + 217 + ], + "type": "text", + "content": "(1)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 484, + 70, + 553, + 140 + ], + "blocks": [ + { + "bbox": [ + 484, + 70, + 553, + 140 + ], + "lines": [ + { + "bbox": [ + 484, + 70, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 484, + 70, + 553, + 140 + ], + "type": "image", + "image_path": "79c1e66398c864e6e6fbed933cc1f66e6dfe902b87e9966f61aeb5774a741f17.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 484, + 140, + 553, + 210 + ], + "blocks": [ + { + "bbox": [ + 484, + 140, + 553, + 210 + ], + "lines": [ + { + "bbox": [ + 484, + 140, + 553, + 210 + ], + "spans": [ + { + "bbox": [ + 484, + 140, + 553, + 210 + ], + "type": "image", + "image_path": "08f57d647c4a4bd0a8ab2e697a408f5ecc5eb467edbf1558c3e5865e90c36542.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 514, + 211, + 523, + 217 + ], + "lines": [ + { + "bbox": [ + 514, + 211, + 523, + 217 + ], + "spans": [ + { + "bbox": [ + 514, + 211, + 523, + 217 + ], + "type": "text", + "content": "(g)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 274, + 295, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 295, + 323 + ], + "type": "text", + "content": "ity in portrait shadow removal. Note, as PSM [56] provides only code for synthesizing training data, without releasing the training code and trained model, its quantitative results in Table 1 are as reported in their paper." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 331, + 149, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 149, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 149, + 344 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 350, + 296, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 296, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 296, + 469 + ], + "type": "text", + "content": "Effectiveness of our shadow-independent structure extraction. Although the trained model of PDG [49] is utilized to train our structure extraction, the structure map it produces is not what we expected because it will include the shadow boundaries and lead to unnatural shadow removal results, as shown in Figure 7 (b)&(e). By training on our synthesized data with diverse lighting conditions, our structure extraction network (SE-Net) can robustly extract shadow-independent structures, enabling high-quality shadow removal results." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 475, + 296, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 296, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 296, + 608 + ], + "type": "text", + "content": "Effectiveness of our data synthesis strategy. To verify the effectiveness of our data synthesis, we alternatively adopt the portrait shadow synthesis strategy introduced in PSM [56] to create data on FFHQ for training our SE-Net. As shown in Figure 7 (c)&(f) and Table 1, training on the data synthesized by the strategy in PSM fails to exclude the unwanted shadow boundaries and will degrade the subsequent shadow removal performance, because the strategy only involves simple modifications to brightness, saturation, and subsurface scattering of face, lacking the ability to physically simulate the complex facial lighting conditions." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 612, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 296, + 685 + ], + "type": "text", + "content": "Effect of structure map on shadow inpainting. We in Figure 8 examine how structure map derived from the SE-Net affects the performance of our diffusion-based shadow inpainting. As shown, omitting the structure map, the facial structure details behind shadows are inevitably modified, producing unpredictable portrait inconsistent to the input." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "content": "Effect of detail restoration. We in Figure 9 analyze the" + } + ] + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 316, + 272, + 375, + 331 + ], + "blocks": [ + { + "bbox": [ + 316, + 272, + 375, + 331 + ], + "lines": [ + { + "bbox": [ + 316, + 272, + 375, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 375, + 331 + ], + "type": "image", + "image_path": "751379de2c98ea584f7029e8f0a94d38dff28013fade88287fd63af57f20468a.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 331, + 354, + 339 + ], + "lines": [ + { + "bbox": [ + 336, + 331, + 354, + 339 + ], + "spans": [ + { + "bbox": [ + 336, + 331, + 354, + 339 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 376, + 273, + 433, + 331 + ], + "blocks": [ + { + "bbox": [ + 376, + 273, + 433, + 331 + ], + "lines": [ + { + "bbox": [ + 376, + 273, + 433, + 331 + ], + "spans": [ + { + "bbox": [ + 376, + 273, + 433, + 331 + ], + "type": "image", + "image_path": "9b90fbe973c0e75e82a7325002bff7f3cf3cda9d898a676956f07402793f4cef.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 331, + 427, + 338 + ], + "lines": [ + { + "bbox": [ + 383, + 331, + 427, + 338 + ], + "spans": [ + { + "bbox": [ + 383, + 331, + 427, + 338 + ], + "type": "text", + "content": "w/o structure" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 435, + 273, + 493, + 331 + ], + "blocks": [ + { + "bbox": [ + 435, + 273, + 493, + 331 + ], + "lines": [ + { + "bbox": [ + 435, + 273, + 493, + 331 + ], + "spans": [ + { + "bbox": [ + 435, + 273, + 493, + 331 + ], + "type": "image", + "image_path": "a4ff21630911b15e4c14a20b3fa4a2c574d91193c8f338ca5f4f17bdb1a98eb7.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 445, + 331, + 484, + 338 + ], + "lines": [ + { + "bbox": [ + 445, + 331, + 484, + 338 + ], + "spans": [ + { + "bbox": [ + 445, + 331, + 484, + 338 + ], + "type": "text", + "content": "w/ structure" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 495, + 273, + 553, + 331 + ], + "blocks": [ + { + "bbox": [ + 55, + 220, + 555, + 254 + ], + "lines": [ + { + "bbox": [ + 55, + 220, + 555, + 254 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 555, + 254 + ], + "type": "text", + "content": "Figure 7. Visual comparison of structure maps extracted by different methods and the corresponding shadow removal results. (a) Input. (b)-(d) are structure maps generated by PDG [49], SE-Net trained on data synthesized by the method of PSM [56], and SE-Net trained on our synthesized data. (e)-(g) are shadow removal results produced by our method using the structure maps in (b)-(d), respectively." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 495, + 273, + 553, + 331 + ], + "lines": [ + { + "bbox": [ + 495, + 273, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 495, + 273, + 553, + 331 + ], + "type": "image", + "image_path": "528f54b2eb1d310bc63fad235d93d6b05ca0d6239324ccae90ad3b08f8b2dbc3.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 518, + 331, + 530, + 338 + ], + "lines": [ + { + "bbox": [ + 518, + 331, + 530, + 338 + ], + "spans": [ + { + "bbox": [ + 518, + 331, + 530, + 338 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 316, + 364, + 394, + 442 + ], + "blocks": [ + { + "bbox": [ + 327, + 342, + 541, + 354 + ], + "lines": [ + { + "bbox": [ + 327, + 342, + 541, + 354 + ], + "spans": [ + { + "bbox": [ + 327, + 342, + 541, + 354 + ], + "type": "text", + "content": "Figure 8. Effect of structure map on shadow inpainting." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 364, + 394, + 442 + ], + "lines": [ + { + "bbox": [ + 316, + 364, + 394, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 364, + 394, + 442 + ], + "type": "image", + "image_path": "772faa9f8e537de7ff4127c3843e9a3e0c950a4c5d129ca44110d319ee8e1928.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 316, + 443, + 394, + 521 + ], + "blocks": [ + { + "bbox": [ + 316, + 443, + 394, + 521 + ], + "lines": [ + { + "bbox": [ + 316, + 443, + 394, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 443, + 394, + 521 + ], + "type": "image", + "image_path": "9c3c852d1697744aa9cd343151f373d7cba8248c6aee8ba5766abd428abc52bd.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 346, + 522, + 364, + 529 + ], + "lines": [ + { + "bbox": [ + 346, + 522, + 364, + 529 + ], + "spans": [ + { + "bbox": [ + 346, + 522, + 364, + 529 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 355, + 533, + 512, + 544 + ], + "lines": [ + { + "bbox": [ + 355, + 533, + 512, + 544 + ], + "spans": [ + { + "bbox": [ + 355, + 533, + 512, + 544 + ], + "type": "text", + "content": "Figure 9. Effect of our detail restoration." + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 394, + 364, + 474, + 442 + ], + "blocks": [ + { + "bbox": [ + 394, + 364, + 474, + 442 + ], + "lines": [ + { + "bbox": [ + 394, + 364, + 474, + 442 + ], + "spans": [ + { + "bbox": [ + 394, + 364, + 474, + 442 + ], + "type": "image", + "image_path": "0381bc227f96ed7194a59190d0ca97b7d2c4f8102a0c60058ab5be9ba01e9ece.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 394, + 443, + 474, + 521 + ], + "blocks": [ + { + "bbox": [ + 394, + 443, + 474, + 521 + ], + "lines": [ + { + "bbox": [ + 394, + 443, + 474, + 521 + ], + "spans": [ + { + "bbox": [ + 394, + 443, + 474, + 521 + ], + "type": "image", + "image_path": "4e00420179498b9a608b4ecb451c1a2937276c61330d7dfebfd982d295d9ed61.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 522, + 469, + 529 + ], + "lines": [ + { + "bbox": [ + 400, + 522, + 469, + 529 + ], + "spans": [ + { + "bbox": [ + 400, + 522, + 469, + 529 + ], + "type": "text", + "content": "w/o detail restoration" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 475, + 364, + 553, + 442 + ], + "blocks": [ + { + "bbox": [ + 475, + 364, + 553, + 442 + ], + "lines": [ + { + "bbox": [ + 475, + 364, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 475, + 364, + 553, + 442 + ], + "type": "image", + "image_path": "3abaf0203442a7054f5c4fc591b97b4492cfa44b24a120d9d96d0cf67e39361a.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 475, + 443, + 553, + 521 + ], + "blocks": [ + { + "bbox": [ + 475, + 443, + 553, + 521 + ], + "lines": [ + { + "bbox": [ + 475, + 443, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 475, + 443, + 553, + 521 + ], + "type": "image", + "image_path": "67fd04bb3c7d907e7aa1b5bd8a2c2872f7504d266e36c915d6e7c7d1f9f1d8f0.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 522, + 547, + 529 + ], + "lines": [ + { + "bbox": [ + 481, + 522, + 547, + 529 + ], + "spans": [ + { + "bbox": [ + 481, + 522, + 547, + 529 + ], + "type": "text", + "content": "w/ detail restoration" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "bbox": [ + 313, + 564, + 555, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 555, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 555, + 661 + ], + "type": "text", + "content": "impact of our detail restoration. As shown, without the restoration step, our shadow removal results exhibit noticeable blurring artifacts around the eyebrows (see 1st row), and fail to faithfully recover the mole in shadow region (see 2nd row). By including the restoration step, we are able to produce high-fidelity shadow removal results with all aforementioned fine-scale details effectively restored, manifesting the effectiveness of our detail restoration." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 313, + 667, + 405, + 680 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 667, + 405, + 680 + ], + "spans": [ + { + "bbox": [ + 313, + 667, + 405, + 680 + ], + "type": "text", + "content": "4.3. More Analysis" + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Impact of inaccurate structure map. Structure prediction plays a vital role for the success of our method, because we" + } + ] + } + ], + "index": 49 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11681" + } + ] + } + ], + "index": 50 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 115, + 129 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 115, + 129 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 115, + 129 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 115, + 129 + ], + "type": "image", + "image_path": "e600ffbb8d1c8dbd17c755cc3700d7f1245eaf32fe3593914d0e4b6ff222459e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 129, + 96, + 138 + ], + "lines": [ + { + "bbox": [ + 78, + 129, + 96, + 138 + ], + "spans": [ + { + "bbox": [ + 78, + 129, + 96, + 138 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 85, + 140, + 266, + 152 + ], + "lines": [ + { + "bbox": [ + 85, + 140, + 266, + 152 + ], + "spans": [ + { + "bbox": [ + 85, + 140, + 266, + 152 + ], + "type": "text", + "content": "Figure 10. Impact of inaccurate structure map." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 116, + 70, + 176, + 129 + ], + "blocks": [ + { + "bbox": [ + 116, + 70, + 176, + 129 + ], + "lines": [ + { + "bbox": [ + 116, + 70, + 176, + 129 + ], + "spans": [ + { + "bbox": [ + 116, + 70, + 176, + 129 + ], + "type": "image", + "image_path": "fb8a9b934382839fa5c60250a65bbd4d3b904b0431a887e9993ad9f0fa16db5f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 121, + 129, + 171, + 138 + ], + "lines": [ + { + "bbox": [ + 121, + 129, + 171, + 138 + ], + "spans": [ + { + "bbox": [ + 121, + 129, + 171, + 138 + ], + "type": "text", + "content": "Inaccurate map" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 177, + 70, + 235, + 129 + ], + "blocks": [ + { + "bbox": [ + 177, + 70, + 235, + 129 + ], + "lines": [ + { + "bbox": [ + 177, + 70, + 235, + 129 + ], + "spans": [ + { + "bbox": [ + 177, + 70, + 235, + 129 + ], + "type": "image", + "image_path": "a94ea827fc0bc74e1266fc71dff96308146e70590f20f533652deda6cf2a4751.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 129, + 233, + 138 + ], + "lines": [ + { + "bbox": [ + 179, + 129, + 233, + 138 + ], + "spans": [ + { + "bbox": [ + 179, + 129, + 233, + 138 + ], + "type": "text", + "content": "Inpainting result" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 236, + 70, + 294, + 129 + ], + "blocks": [ + { + "bbox": [ + 236, + 70, + 294, + 129 + ], + "lines": [ + { + "bbox": [ + 236, + 70, + 294, + 129 + ], + "spans": [ + { + "bbox": [ + 236, + 70, + 294, + 129 + ], + "type": "image", + "image_path": "9db58f640587a15631f2fdff54b5d9e49361417070f5b6601356288a2e38eced.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 129, + 284, + 138 + ], + "lines": [ + { + "bbox": [ + 247, + 129, + 284, + 138 + ], + "spans": [ + { + "bbox": [ + 247, + 129, + 284, + 138 + ], + "type": "text", + "content": "Final result" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 57, + 195, + 115, + 255 + ], + "blocks": [ + { + "bbox": [ + 57, + 195, + 115, + 255 + ], + "lines": [ + { + "bbox": [ + 57, + 195, + 115, + 255 + ], + "spans": [ + { + "bbox": [ + 57, + 195, + 115, + 255 + ], + "type": "image", + "image_path": "01c0509b64eded687a734909bd7daa5435908f69c8c3cdeedc1c36e715fcc6a2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 117, + 164, + 175, + 224 + ], + "blocks": [ + { + "bbox": [ + 117, + 164, + 175, + 224 + ], + "lines": [ + { + "bbox": [ + 117, + 164, + 175, + 224 + ], + "spans": [ + { + "bbox": [ + 117, + 164, + 175, + 224 + ], + "type": "image", + "image_path": "21309f1944af8da8eff70359b4c8a63745c69a94de46723d633d496d049a34f8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 117, + 224, + 175, + 281 + ], + "blocks": [ + { + "bbox": [ + 117, + 224, + 175, + 281 + ], + "lines": [ + { + "bbox": [ + 117, + 224, + 175, + 281 + ], + "spans": [ + { + "bbox": [ + 117, + 224, + 175, + 281 + ], + "type": "image", + "image_path": "907acb1c6a839ccf02f3d5d5b10018cc7a950cae6bc16a1c6f524e3b5ba5d240.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 129, + 282, + 162, + 290 + ], + "lines": [ + { + "bbox": [ + 129, + 282, + 162, + 290 + ], + "spans": [ + { + "bbox": [ + 129, + 282, + 162, + 290 + ], + "type": "text", + "content": "ID: 0.056" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 293, + 295, + 316 + ], + "lines": [ + { + "bbox": [ + 55, + 293, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 295, + 316 + ], + "type": "text", + "content": "Figure 11. Effect of inaccurate shadow masks. Our method is highly robust to inaccurate masks (see 3rd and 4th columns)." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 177, + 164, + 235, + 224 + ], + "blocks": [ + { + "bbox": [ + 177, + 164, + 235, + 224 + ], + "lines": [ + { + "bbox": [ + 177, + 164, + 235, + 224 + ], + "spans": [ + { + "bbox": [ + 177, + 164, + 235, + 224 + ], + "type": "image", + "image_path": "559e49f4535c14c5f6c57f555325a681197199bc6820f668c56693f5f3fd8b1f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 177, + 224, + 235, + 281 + ], + "blocks": [ + { + "bbox": [ + 177, + 224, + 235, + 281 + ], + "lines": [ + { + "bbox": [ + 177, + 224, + 235, + 281 + ], + "spans": [ + { + "bbox": [ + 177, + 224, + 235, + 281 + ], + "type": "image", + "image_path": "eafd08d435cabd0ca9e9e6cb508f5dfc4b0e523ce56edd2859cc66d5099ada9a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 189, + 282, + 221, + 289 + ], + "lines": [ + { + "bbox": [ + 189, + 282, + 221, + 289 + ], + "spans": [ + { + "bbox": [ + 189, + 282, + 221, + 289 + ], + "type": "text", + "content": "ID:0.065" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 236, + 164, + 294, + 224 + ], + "blocks": [ + { + "bbox": [ + 236, + 164, + 294, + 224 + ], + "lines": [ + { + "bbox": [ + 236, + 164, + 294, + 224 + ], + "spans": [ + { + "bbox": [ + 236, + 164, + 294, + 224 + ], + "type": "image", + "image_path": "22f31829e8c49b341c2b0b3e5cc432b5ac9d3e24215b765e19458093f7756a83.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 236, + 224, + 294, + 281 + ], + "blocks": [ + { + "bbox": [ + 236, + 224, + 294, + 281 + ], + "lines": [ + { + "bbox": [ + 236, + 224, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 236, + 224, + 294, + 281 + ], + "type": "image", + "image_path": "827d3c79f55d6266a2cbcb74d54668921df91accb9b5391dca31898a4a984527.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 282, + 282, + 289 + ], + "lines": [ + { + "bbox": [ + 249, + 282, + 282, + 289 + ], + "spans": [ + { + "bbox": [ + 249, + 282, + 282, + 289 + ], + "type": "text", + "content": "ID:0.062" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 338, + 296, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 296, + 398 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 296, + 398 + ], + "type": "text", + "content": "rely on a structure map to guide our diffusion-based shadow inpainting. However, as shown in Figure 10, our method still has some tolerance to inaccurate structure map as the subsequent detail restoration stage can help remove the inpainting artifacts led by inaccurate structure map." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 403, + 296, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 296, + 535 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 296, + 535 + ], + "type": "text", + "content": "Robustness to inaccurate shadow masks. Thanks to the design of performing portrait shadow removal via structure-guided inpainting, our method is highly robust to inaccurate input shadow masks, requiring only that the masks cover the shadow regions even if they include large-scale non-shadow regions. As demonstrated in Figure 11, our method achieves both visually indistinguishable shadow removal results and comparable ID preservation scores [48] for all the three masks, including two inaccurate masks in the 3rd and 4th columns. Please see the supplementary material for more results produced with inaccurate shadow masks." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 540, + 296, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 296, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 296, + 661 + ], + "type": "text", + "content": "Different diffusion models for shadow inpainting. To test how different diffusion models affect the performance of our method, we replace our inpainting diffusion model with a strong pre-trained Stable Diffusion (SD) model (version v1-5 inpainting model [36]). To ensure fair comparison, the compared SD model is injected with the same structure map from SE-Net using ControlNet [53]. As can be seen in Figure 12, our trained inpainting diffusion model produces high-fidelity shadow removal results faithful to the inputs, while the SD model incurs noticeable structure tampering." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Effect of our method on natural images. Our method is also applicable to natural images by replacing the portrait-oriented PDG model [49] with a generic edge detection model and also changing the training dataset. We show" + } + ] + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 375, + 129 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 375, + 129 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 375, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 375, + 129 + ], + "type": "image", + "image_path": "a986e113c2a5f685b1b9c35642a0fc9ca65a1ab85faa44f8ab7a26088e74441c.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 316, + 129, + 373, + 188 + ], + "blocks": [ + { + "bbox": [ + 316, + 129, + 373, + 188 + ], + "lines": [ + { + "bbox": [ + 316, + 129, + 373, + 188 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 373, + 188 + ], + "type": "image", + "image_path": "dd169ab83d7fcccfc797bb9e3e70c75cd432ae5ed8d9a65875d1340c3c78c5ee.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 188, + 354, + 196 + ], + "lines": [ + { + "bbox": [ + 335, + 188, + 354, + 196 + ], + "spans": [ + { + "bbox": [ + 335, + 188, + 354, + 196 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 199, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 313, + 199, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 555, + 243 + ], + "type": "text", + "content": "Figure 12. Effect of leveraging different diffusion models for shadow inpainting. Note, the same structure map as ours is used to produce the shadow inpainting result of \"SD + ControlNet\", and our result here is free of detail restoration." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 375, + 70, + 434, + 129 + ], + "blocks": [ + { + "bbox": [ + 375, + 70, + 434, + 129 + ], + "lines": [ + { + "bbox": [ + 375, + 70, + 434, + 129 + ], + "spans": [ + { + "bbox": [ + 375, + 70, + 434, + 129 + ], + "type": "image", + "image_path": "b1eb37c0f980142e53e0292c1f2546685b36384ac8e81c5cb5c971f6335b561c.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 375, + 129, + 433, + 188 + ], + "blocks": [ + { + "bbox": [ + 375, + 129, + 433, + 188 + ], + "lines": [ + { + "bbox": [ + 375, + 129, + 433, + 188 + ], + "spans": [ + { + "bbox": [ + 375, + 129, + 433, + 188 + ], + "type": "image", + "image_path": "71ac04d28ba94f701a64a6cff401abf1d61a0f481505bb4852e8bcd3761a3b8b.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 377, + 188, + 432, + 196 + ], + "lines": [ + { + "bbox": [ + 377, + 188, + 432, + 196 + ], + "spans": [ + { + "bbox": [ + 377, + 188, + 432, + 196 + ], + "type": "text", + "content": "SD + ControlNet" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 435, + 70, + 493, + 129 + ], + "blocks": [ + { + "bbox": [ + 435, + 70, + 493, + 129 + ], + "lines": [ + { + "bbox": [ + 435, + 70, + 493, + 129 + ], + "spans": [ + { + "bbox": [ + 435, + 70, + 493, + 129 + ], + "type": "image", + "image_path": "68a4f52e59f7700af2b1d5a2744361ffea3a70ed994d0994c26a7f5371f4bc0a.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 435, + 129, + 493, + 188 + ], + "blocks": [ + { + "bbox": [ + 435, + 129, + 493, + 188 + ], + "lines": [ + { + "bbox": [ + 435, + 129, + 493, + 188 + ], + "spans": [ + { + "bbox": [ + 435, + 129, + 493, + 188 + ], + "type": "image", + "image_path": "5bbd776822c6f83e1cc6c666a6337e85dde49ec04f16504443e2c4a0db203f46.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 457, + 189, + 473, + 196 + ], + "lines": [ + { + "bbox": [ + 457, + 189, + 473, + 196 + ], + "spans": [ + { + "bbox": [ + 457, + 189, + 473, + 196 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 495, + 70, + 553, + 129 + ], + "blocks": [ + { + "bbox": [ + 495, + 70, + 553, + 129 + ], + "lines": [ + { + "bbox": [ + 495, + 70, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 495, + 70, + 553, + 129 + ], + "type": "image", + "image_path": "57769de3d5392ba314900bc24ea4f60029b914d3e5f6b8e84a8b789eedb5abb3.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 495, + 129, + 553, + 188 + ], + "blocks": [ + { + "bbox": [ + 495, + 129, + 553, + 188 + ], + "lines": [ + { + "bbox": [ + 495, + 129, + 553, + 188 + ], + "spans": [ + { + "bbox": [ + 495, + 129, + 553, + 188 + ], + "type": "image", + "image_path": "269a4b717cf02d476b822c2969648a07eab5c4cbfab5003232515bfac0f0e3a3.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 518, + 189, + 529, + 196 + ], + "lines": [ + { + "bbox": [ + 518, + 189, + 529, + 196 + ], + "spans": [ + { + "bbox": [ + 518, + 189, + 529, + 196 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "table", + "bbox": [ + 317, + 256, + 552, + 297 + ], + "blocks": [ + { + "bbox": [ + 317, + 256, + 552, + 297 + ], + "lines": [ + { + "bbox": [ + 317, + 256, + 552, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 256, + 552, + 297 + ], + "type": "table", + "html": "
MethodUPSRBSRShadowDiffusionIC-lightOurs
Param.135M72M108.79M-185.12M
Time192s11s2.3s32s9.5s
", + "image_path": "34f74427323299516f8446abb607c3fe3984e8e2722e48f1751e6563ce69db71.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "table_body" + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 301, + 555, + 335 + ], + "lines": [ + { + "bbox": [ + 313, + 301, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 555, + 335 + ], + "type": "text", + "content": "Table 2. Comparison on the number of parameters (param.) and inference time. Note, all the inference times here are test for a " + }, + { + "bbox": [ + 313, + 301, + 555, + 335 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 301, + 555, + 335 + ], + "type": "text", + "content": " portrait on an NVIDIA A100 Tensor Core GPU." + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 353, + 554, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 554, + 376 + ], + "type": "text", + "content": "in the supplementary material that our method can produce competitive shadow removal results for natural images." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 382, + 554, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 554, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 554, + 467 + ], + "type": "text", + "content": "Analysis on computational complexity. Table 2 compares our inference time with those of other existing portrait specialized or diffusion-based shadow removal methods. As shown, our inference time is more efficient than that of most other methods. Moreover, the number of parameters of our method are comparable to other methods, showing that our superior performance comes mainly from method design." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 313, + 483, + 388, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 388, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 388, + 496 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 313, + 504, + 555, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 555, + 660 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 555, + 660 + ], + "type": "text", + "content": "We have presented a portrait shadow removal approach that allows to robustly produce high-fidelity results. Unlike previous methods, we cast portrait shadow removal as diffusion-based inpainting. To do so, we develop a structure extraction network that is robust to illumination changes to estimate shadow-independent structure map for a given portrait image. The structure map is then fed to an inpainting diffusion model to guide the generation of the shadow-free output. Next, we refine the shadow-free output with a detail restoration diffusion model for impairing the fine-scale facial details that may not be captured by the structure map. Extensive experiments show that our method outperforms state-of-the-art methods for portrait shadow removal." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "Acknowledgment. This work was supported by the National Natural Science Foundation of China (62471499), the Guangdong Basic and Applied Basic Research Foundation (2023A1515030002)." + } + ] + } + ], + "index": 43 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11682" + } + ] + } + ], + "index": 44 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "type": "text", + "content": "[1] Ciprian Corneanu, Raghudeep Gadde, and Aleix M Martinez. Latentpaint: Image inpainting in latent space with diffusion models. In WACV, pages 4334-4343, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 124, + 295, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 124, + 295, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 124, + 295, + 167 + ], + "type": "text", + "content": "[2] Antonio Criminisi, Patrick Pérez, and Kentaro Toyama. Region filling and object removal by exemplar-based image inpainting. IEEE Transactions on Image Processing, 13(9): 1200-1212, 2004. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 168, + 295, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 168, + 295, + 212 + ], + "spans": [ + { + "bbox": [ + 62, + 168, + 295, + 212 + ], + "type": "text", + "content": "[3] Xiaodong Cun, Chi-Man Pun, and Cheng Shi. Towards ghost-free shadow removal via dual hierarchical aggregation network and shadow matting gan. In AAAI, pages 10680-10687, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 213, + 294, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 213, + 294, + 255 + ], + "spans": [ + { + "bbox": [ + 62, + 213, + 294, + 255 + ], + "type": "text", + "content": "[4] Xingbo Dong, Sangrae Cho, Youngsam Kim, Soohyung Kim, and Andrew Beng Jin Teoh. Deep rank hashing network for cancellable face identification. Pattern Recognition, 131:108886, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 258, + 294, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 294, + 301 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 294, + 301 + ], + "type": "text", + "content": "[5] Pawel Drozdowski, Christian Rathgeb, and Christoph Busch. The watchlist imbalance effect in biometric face identification: Comparing theoretical estimates and empiric measurements. In ICCV, pages 3757-3765, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 302, + 294, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 294, + 335 + ], + "type": "text", + "content": "[6] Xianyong Fang, Xiaohao He, Linbo Wang, and Jianbing Shen. Robust shadow detection by exploring effective shadow contexts. In ACM MM, pages 2927-2935, 2021. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 335, + 296, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 296, + 368 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 296, + 368 + ], + "type": "text", + "content": "[7] Graham D Finlayson, Mark S Drew, and Cheng Lu. Entropy minimization for shadow removal. International Journal of Computer Vision, 85(1):35-57, 2009. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 369, + 294, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 294, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 294, + 413 + ], + "type": "text", + "content": "[8] David Futschik, Kelvin Ritland, James Vecore, Sean Fanello, Sergio Orts-Escalano, Brian Curless, Daniel Sykora, and Rohit Pandey. Controllable light diffusion for portraits. In CVPR, pages 8412-8421, 2023. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 414, + 294, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 294, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 294, + 446 + ], + "type": "text", + "content": "[9] Maciej Gryka, Michael Terry, and Gabriel J Brostow. Learning to remove soft shadows. ACM Transactions on Graphics (TOG), 34(5):1-15, 2015. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 447, + 294, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 447, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 57, + 447, + 294, + 479 + ], + "type": "text", + "content": "[10] Lanqing Guo, Siyu Huang, Ding Liu, Hao Cheng, and Bihan Wen. Shadowformer: Global context helps shadow removal. In AAAI, pages 710-718, 2023. 2, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 480, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 480, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 57, + 480, + 294, + 533 + ], + "type": "text", + "content": "[11] Lanqing Guo, Chong Wang, Wenhan Yang, Siyu Huang, Yufei Wang, Hanspeter Pfister, and Bihan Wen. Shadowdiffusion: When degradation prior meets diffusion model for shadow removal. In CVPR, pages 14049-14058, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 535, + 294, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 535, + 294, + 579 + ], + "spans": [ + { + "bbox": [ + 57, + 535, + 294, + 579 + ], + "type": "text", + "content": "[12] Ruiqi Guo, Qieyun Dai, and Derek Hoiem. Paired regions for shadow detection and removal. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(12):2956-2967, 2012. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 580, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 580, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 57, + 580, + 294, + 623 + ], + "type": "text", + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, pages 1026-1034, 2015. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 624, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 294, + 658 + ], + "type": "text", + "content": "[14] Yingqing He, Yazhou Xing, Tianjia Zhang, and Qifeng Chen. Unsupervised portrait shadow removal via generative priors. In ACM MM, pages 236-244, 2021. 2, 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 658, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 689 + ], + "type": "text", + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 4, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 690, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 690, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 690, + 294, + 714 + ], + "type": "text", + "content": "[16] Andrew Hou, Zhixin Shu, Xuaner Zhang, He Zhang, Yannick Hold-Geoffroy, Jae Shin Yoon, and Xiaoming Liu." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "Compose: Comprehensive portrait shadow editing. In ECCV, pages 356-373. Springer, 2024. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 97, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 129 + ], + "type": "text", + "content": "[17] Xiaowei Hu, Yitong Jiang, Chi-Wing Fu, and Pheng-Ann Heng. Mask-shadowgan: Learning to remove shadows from unpaired data. In ICCV, pages 2472–2481, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 131, + 553, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 131, + 553, + 164 + ], + "spans": [ + { + "bbox": [ + 316, + 131, + 553, + 164 + ], + "type": "text", + "content": "[18] Satoshi Iizuka, Edgar Simo-Serra, and Hiroshi Ishikawa. Globally and locally consistent image completion. ACM Transactions on Graphics (ToG), 36(4):1-14, 2017. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 166, + 553, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 166, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 317, + 166, + 553, + 209 + ], + "type": "text", + "content": "[19] Liya Ji, Zhefan Rao, Sinno Jialin Pan, Chenyang Lei, and Qifeng Chen. A diffusion model with state estimation for degradation-blind inverse imaging. In AAAI, pages 2471-2479, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 210, + 553, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 210, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 317, + 210, + 553, + 255 + ], + "type": "text", + "content": "[20] Hao Jiang, Qing Zhang, Yongwei Nie, Lei Zhu, and Wei-Shi Zheng. Learning to remove shadows from a single image. International Journal of Computer Vision, 131(9):2471-2488, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 256, + 553, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 256, + 553, + 290 + ], + "spans": [ + { + "bbox": [ + 316, + 256, + 553, + 290 + ], + "type": "text", + "content": "[21] Yuming Jiang, Ziqi Huang, Xingang Pan, Chen Change Loy, and Ziwei Liu. Talk-to-edit: Fine-grained facial editing via dialog. In ICCV, pages 13799-13808, 2021. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 291, + 553, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 291, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 316, + 291, + 553, + 335 + ], + "type": "text", + "content": "[22] Yeying Jin, Aashish Sharma, and Robby T Tan. Dcshadownet: Single-image hard and soft shadow removal using unsupervised domain-classifier guided network. In ICCV, pages 5027-5036, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 336, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 336, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 336, + 553, + 380 + ], + "type": "text", + "content": "[23] Yeying Jin, Wei Ye, Wenhan Yang, Yuan Yuan, and Robby T Tan. Des3: Adaptive attention-driven self and soft shadow removal using vit similarity. In AAAI, pages 2634-2642, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 381, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 381, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 381, + 553, + 425 + ], + "type": "text", + "content": "[24] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR, pages 8110-8119, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 426, + 553, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 450 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 450 + ], + "type": "text", + "content": "[25] Hieu Le and Dimitris Samaras. Shadow removal via shadow image decomposition. In ICCV, pages 8578-8587, 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 451, + 553, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 451, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 317, + 451, + 553, + 495 + ], + "type": "text", + "content": "[26] Xiaoguang Li, Qing Guo, Rabab Abdelfattah, Di Lin, Wei Feng, Ivor Tsang, and Song Wang. Leveraging inpainting for single-image shadow removal. In ICCV, pages 13055-13064, 2023. 2, 5, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 496, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 496, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 317, + 496, + 553, + 529 + ], + "type": "text", + "content": "[27] Zonglin Li, Xiaogian Lv, Wei Yu, Qinglin Liu, Jingbo Lin, and Shengping Zhang. Face shape transfer via semantic warping. Visual Intelligence, 2(1):26, 2024. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 531, + 553, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 531, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 317, + 531, + 553, + 564 + ], + "type": "text", + "content": "[28] Yaojie Liu, Andrew Z Hou, Xinyu Huang, Liu Ren, and Xiaoming Liu. Blind removal of facial foreign shadows. In BMVC, page 88, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 565, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 565, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 317, + 565, + 553, + 599 + ], + "type": "text", + "content": "[29] Yuhao Liu, Zhanghan Ke, Ke Xu, Fang Liu, Zhenwei Wang, and Rynson WH Lau. Recasting regional lighting for shadow removal. In AAAI, pages 3810-3818, 2024. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 600, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 600, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 600, + 553, + 633 + ], + "type": "text", + "content": "[30] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In ICCV, pages 3730-3738, 2015. 3, 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 317, + 634, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 634, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 634, + 553, + 668 + ], + "type": "text", + "content": "[31] Zhihao Liu, Hui Yin, Xinyi Wu, Zhenyao Wu, Yang Mi, and Song Wang. From shadow generation to shadow removal. In CVPR, pages 4927-4936, 2021. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 669, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 714 + ], + "type": "text", + "content": "[32] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, pages 11461-11471, 2022. 2" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11683" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[33] Junfeng Lyu, Zhibo Wang, and Feng Xu. Portrait eyeglasses and shadow removal by leveraging 3d synthetic data. In CVPR, pages 3429-3439, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 294, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 149 + ], + "type": "text", + "content": "[34] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In CVPR, pages 2536-2544, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "type": "text", + "content": "[35] Liangqiong Qu, Jiandong Tian, Shengfeng He, Yandong Tang, and Rynson WH Lau. Deshadownet: A multi-context embedding deep network for shadow removal. In CVPR, pages 4067-4075, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 194, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 294, + 237 + ], + "type": "text", + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 238, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 238, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 238, + 294, + 293 + ], + "type": "text", + "content": "[37] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 conference proceedings, pages 1-10, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 293, + 294, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 294, + 348 + ], + "type": "text", + "content": "[38] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713-4726, 2022. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 349, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 349, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 349, + 294, + 381 + ], + "type": "text", + "content": "[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 382, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 294, + 437 + ], + "type": "text", + "content": "[40] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In WACV, pages 2149–2159, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 294, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 469 + ], + "type": "text", + "content": "[41] Chengrui Wang and Weihong Deng. Representative forgery mining for fake face detection. In CVPR, pages 14923-14932, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 470, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 470, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 470, + 294, + 514 + ], + "type": "text", + "content": "[42] Jifeng Wang, Xiang Li, and Jian Yang. Stacked conditional generative adversarial networks for jointly learning shadow detection and shadow removal. In CVPR, pages 1788-1797, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 514, + 294, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 294, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 294, + 548 + ], + "type": "text", + "content": "[43] Joshua Weir, Junhong Zhao, Andrew Chalmers, and Taehyun Rhee. Deep portrait delighting. In ECCV, pages 423-439. Springer, 2022. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 548, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 581 + ], + "type": "text", + "content": "[44] Yonatan Wexler, Eli Shechtman, and Michal Irani. Spacetime completion of video. IEEE Transactions on Pattern Analysis and Machine Intelligence, 29(3):463-476, 2007. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "type": "text", + "content": "[45] Jie Xiao, Xueyang Fu, Yurui Zhu, Dong Li, Jie Huang, Kai Zhu, and Zheng-Jun Zha. Homoformer: Homogenized transformer for image shadow removal. In CVPR, pages 25617-25626, 2024. 2, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 625, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 667 + ], + "type": "text", + "content": "[46] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In CVPR, pages 22428-22437, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "type": "text", + "content": "[47] Yanbo Xu, Yueqin Yin, Liming Jiang, Qianyi Wu, Chengyao Zheng, Chen Change Loy, Bo Dai, and Wayne Wu. Transeditor: Transformer-based dual-space gan for highly controllable facial editing. In CVPR, pages 7683-7692, 2022. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 624 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[48] Shuai Yang, Liming Jiang, Ziwei Liu, and Chen Change Loy. Pastiche master: Exemplar-based high-resolution portrait style transfer. In CVPR, pages 7693-7702, 2022. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "text", + "content": "[49] Ran Yi, Yong-Jin Liu, Yu-Kun Lai, and Paul L Rosin. Unpaired portrait drawing generation via asymmetric cycle mapping. In CVPR, pages 8217-8225, 2020. 3, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "text", + "content": "[50] Jae Shin Yoon, Zhixin Shu, Mengwei Ren, Cecilia Zhang, Yannick Hold-Geoffroy, Krishna Kumar Singh, and He Zhang. Generative portrait shadow removal. ACM Transactions on Graphics (TOG), 43(6):1-13, 2024. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 185, + 553, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 219 + ], + "type": "text", + "content": "[51] Jiahui Yu, Zhe Lin, Jamei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Generative image inpainting with contextual attention. In CVPR, pages 5505-5514, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "text", + "content": "[52] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Scaling in-the-wild training for diffusion-based illumination harmonization and editing by imposing consistent light transport. In ICLR. 1, 2, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "type": "text", + "content": "[53] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836-3847, 2023. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "type": "text", + "content": "[54] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 342, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 387 + ], + "type": "text", + "content": "[55] Shifeng Zhang, Cheng Chi, Zhen Lei, and Stan Z Li. Refineface: Refinement neural network for high performance face detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(11):4008-4020, 2020. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "type": "text", + "content": "[56] Xuaner Zhang, Jonathan T Barron, Yun-Ta Tsai, Rohit Pandey, Xiuming Zhang, Ren Ng, and David E Jacobs. Portrait shadow manipulation. ACM Transactions on Graphics (TOG), 39(4):78-1, 2020. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 433, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 477 + ], + "type": "text", + "content": "[57] Rongjia Zheng, Qing Zhang, Yongwei Nie, and Wei-Shi Zheng. When shadow removal meets intrinsic image decomposition: A joint learning framework using unpaired data. In AAAI, pages 10599-10607, 2025. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 477, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 477, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 477, + 553, + 510 + ], + "type": "text", + "content": "[58] Hao Zhou, Sunil Hadap, Kalyan Sunkavalli, and David W Jacobs. Deep single-image portrait relighting. In ICCV, pages 7194-7202, 2019. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 510, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 553, + 555 + ], + "type": "text", + "content": "[59] Shipeng Zhu, Pengfei Fang, Chenjie Zhu, Zuoyan Zhao, Qiang Xu, and Hui Xue. Text image inpainting via global structure-guided diffusion models. In AAAI, pages 7775-7783, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 555, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 553, + 590 + ], + "type": "text", + "content": "[60] Yurui Zhu, Jie Huang, Xueyang Fu, Feng Zhao, Qibin Sun, and Zheng-Jun Zha. Bijective mapping network for shadow removal. In CVPR, pages 5627-5636, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 590, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 624 + ], + "type": "text", + "content": "[61] Yurui Zhu, Zeyu Xiao, Yanchi Fang, Xueyang Fu, Zhiwei Xiong, and Zheng-Jun Zha. Efficient model-driven network for shadow removal. In AAAI, pages 3635-3643, 2022. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11684" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Structure-aware Semantic Discrepancy and Consistency for 3D Medical Image Self-supervised Learning/d73adb04-404d-4f0b-9d9c-095f1f245a9b_content_list.json b/2025/Structure-aware Semantic Discrepancy and Consistency for 3D Medical Image Self-supervised Learning/d73adb04-404d-4f0b-9d9c-095f1f245a9b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8c8750114fa8db3910ee68d3a8ed20120d84fdbe --- /dev/null +++ b/2025/Structure-aware Semantic Discrepancy and Consistency for 3D Medical Image Self-supervised Learning/d73adb04-404d-4f0b-9d9c-095f1f245a9b_content_list.json @@ -0,0 +1,1656 @@ +[ + { + "type": "text", + "text": "Structure-aware Semantic Discrepancy and Consistency for 3D Medical Image Self-supervised Learning", + "text_level": 1, + "bbox": [ + 102, + 128, + 893, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tan Pan $^{1,2}$ , Zhaorui Tan $^{2}$ , Kaiyu Guo $^{3,2}$ , Dongli Xu $^{2}$ , Weidi Xu, Chen Jiang $^{1,2*}$ , Xin Guo $^{1,2}$ , Yuan Qi $^{1,2,4}$ , Yuan Cheng $^{1,2*}$", + "bbox": [ + 243, + 202, + 750, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ AI $^{3}$ , Fudan University $^{2}$ Shanghai Academy of Artificial Intelligence for Science $^{3}$ The University of Queensland $^{4}$ Zhongshan Hospital, Fudan University", + "bbox": [ + 176, + 238, + 820, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pant23@m.fudan.edu.cn, jiangchen@sais.com.cn, cheng-yuan@fudan.edu.cn", + "bbox": [ + 189, + 277, + 797, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D medical image self-supervised learning (mSSL) holds great promise for medical analysis. Effectively supporting broader applications requires considering anatomical structure variations in location, scale, and morphology, which are crucial for capturing meaningful distinctions. However, previous mSSL methods partition images with fixed-size patches, often ignoring the structure variations. In this work, we introduce a novel perspective on 3D medical images with the goal of learning structure-aware representations. We assume that patches within the same structure share the same semantics (semantic consistency) while those from different structures exhibit distinct semantics (semantic discrepancy). Based on this assumption, we propose an mSSL framework named $S^2DC$ , achieving Structure-aware Semantic Discrepancy and Consistency in two steps. First, $S^2DC$ enforces distinct representations for different patches to increase semantic discrepancy by leveraging an optimal transport strategy. Second, $S^2DC$ advances semantic consistency at the structural level based on neighborhood similarity distribution. By bridging patch-level and structure-level representations, $S^2DC$ achieves structure-aware representations. Thoroughly evaluated across 10 datasets, 4 tasks, and 3 modalities, our proposed method consistently outperforms the state-of-the-art methods in mSSL. The code is available at https://github.com/Ashespt/S2DC/.", + "bbox": [ + 88, + 357, + 483, + 751 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 779, + 220, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Self-supervised learning (SSL) is widely used in medical image analysis to build strong representations that improve performance on various applications and maximize the use of unlabeled data [23, 33, 40, 41]. These applica", + "bbox": [ + 89, + 804, + 483, + 866 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/98d082471e3a49804679308836c0ec15e106963488061adb787739266abea581.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 325, + 591, + 381 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/007fe4fd6055ab11529d337e318f0dc8ea566271ed8892065b14269e3385fd75.jpg", + "image_caption": [ + "A. Two views of mSSL" + ], + "image_footnote": [], + "bbox": [ + 516, + 382, + 591, + 444 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9d3b6ed803d593e8a5e4236c766658744ae3109dc48675fb32d4ba87d627815f.jpg", + "image_caption": [ + "Figure 1. A. Two views of mSSL. B. Similarity heatmaps on CT and MRI images across different methods. We sample an anchor patch (red dots) and compute feature similarities with all other patches. In the first row, the liver anchor should show low similarity with non-liver patches, while in the second row, the white matter anchor should exhibit high similarity with other white matter patches. Current SOTA contrastive-based [41] (b) and reconstruction-based [40] (c) methods struggle with both patch feature discrepancy in different structures and consistency in the same structure. In contrast, our method (d) advances both discrepancy and consistency." + ], + "image_footnote": [], + "bbox": [ + 602, + 325, + 897, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tions frequently focus on different specific anatomical structures (i.e., organ segmentation), each distinguished across anatomy due to location, scale, morphology, and other variances [4, 9, 10, 29]. For example, the liver is a large, well-defined solid organ with relatively homogeneous intensity and distinct anatomical boundaries. In contrast, veins are thin, tubular structures with varying diameters, lower contrast, and complex branching patterns.", + "bbox": [ + 509, + 656, + 906, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current medical image SSL (mSSL) methods can be classified into contrastive-based and reconstruction-based methods [14, 17, 40, 41]. For example, Wang et al. utilizes multi-view reconstruction to encourage dense representations, while Wu et al. incorporates overlapped patch ratio as soft contrastive constraint to improve semantic consistency. However, these methods use fixed-size image partition, treating all segmented patches uniformly across struc", + "bbox": [ + 509, + 779, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding authors. This research was conducted during an internship at the Shanghai Academy of Artificial Intelligence for Science.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20257", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tures and overlooking their underlying semantic relationship among structures, as shown in Fig. 1 A. To illustrate these limitations, we visualize similarity heatmaps on CT and MRI images across different methods by overlaying similarity scores onto the original images (Fig. 1 B). Ideally, patches within the same organ should exhibit high similarity, while those from different organs should show low similarity. However, as shown in Fig. 1 (a) and (b), existing methods struggle with maintaining intra-structure patch consistency and distinguishing inter-structure patch differences. A more detailed analysis and metrics of organ-level relationships in the state-of-the-art (SOTA) methods are provided in Fig. 5.", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on these empirical analyses, we propose that patches within the same structure should share the same semantics, while those from different structures should exhibit distinct semantics. To achieve this, patch relationships must be structure-aware and adaptively adjusted to account for anatomical variations. Based on this assumption, we define structure-aware learning through two key principles: intra-structure consistency and inter-structure discrepancy.", + "bbox": [ + 89, + 289, + 480, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve these principles, we propose $S^2 DC$ , a self-supervised learning framework for 3D medical imaging designed to capture Structure-aware Semantic Consistency and Discrepancy. The core idea of our framework is to explore semantic relevance between patches to enhance structure-level representations, which unfolds in two key steps. First, we increase the discrepancy between patches with different semantics through patch-to-patch discrepancy. By an optimal transport strategy, patches with different semantics are pushed further apart. Second, beyond patch-wise distinction, the framework further explores patch-to-structure semantic consistency, leveraging the observation that patches with shared semantics form peaks in the similarity distribution. These distributional characteristics serve as soft regularization, identifying patches belonging to the same structure and enforcing structural-level consistency. Notably, the principles of inter-structure discrepancy and intra-structure consistency of our framework are also suitable for functional imaging or metabolic imaging.", + "bbox": [ + 89, + 411, + 482, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This work's contributions are summarized as follows:", + "bbox": [ + 109, + 717, + 465, + 731 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel insight of intra-structure consistency and inter-structure discrepancy in the anatomical structure-aware feature learning in 3D medical images. We argue that the patches within the same anatomical structures maintain consistent semantics, while the patches from different anatomical structures reflect distinct semantics.", + "- Based on our insight and assumption, we propose $S^2 DC$ , a novel training framework that enhances interstructure discrepancy and intra-structure consistency. The framework establishes reliable patch-to-patch correspond" + ], + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "dences to reinforce discrepancy while leveraging patch-to-structure semantic connectivity from the similarity distribution to improve consistency.", + "bbox": [ + 526, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Our method demonstrates superior performance over SOTA medical image SSL methods, evaluated across 10 datasets, 4 tasks, and 3 imaging modalities. The code will be available upon publication.", + "bbox": [ + 513, + 136, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 210, + 650, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SSL in 3D medical images. Contrastive-based and reconstruction-based SSL are among the most prominent paradigms in mSSL. Some approaches emphasize the fine-grained capabilities of models, incorporating techniques such as multi-scale reconstruction or reconstruction from disarranged images to enhance model performance [35, 51, 52]. Simultaneously, other methods [41, 49] prioritize preserving maximal semantic information by leveraging contextual priors. Recently, several studies have explored the combination of contrastive learning (CL) and reconstruction to further improve model performance [15, 17, 34, 40]. These works aim to learn consistent semantic representations by applying contrastive learning to global features or predicting missing information from available context. However, in mSSL, few approaches enforce constraints to enhance both anatomical structure-aware discrepancy and consistency, which is the central focus of this paper.", + "bbox": [ + 511, + 234, + 903, + 492 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Geometric approaches for medical image analysis. Our method adopts some geometric characteristics for patch-to-patch correspondence. In radiographic imaging, geometric characteristics, such as location and anatomical structure, play a crucial role in organ identification, disease diagnosis, and medical analysis. Several approaches [16, 26] have incorporated consistent and recurrent anatomical patterns as learnable priors to enhance medical image analysis. For example, Huang et al. proposed to enhance segmentation by establishing patch correspondence between two 2D images. Those methods utilizing the geometric or anatomical invariant have shown remarkable improvement in the downstream tasks. Some medical SSL works [15, 20] also explore geometric correspondence for better performance. The SSL method GVSL [15] considers the geometric transformation and reconstructs affine images to original images to learn global and local representations. Yang et al. present a keypoint-augmented fusion layer with the keypoint guidance from SIFT to extract local and global information to improve segmentation results. Those methods leverage the invariance of anatomical structures while not considering the relation between anatomical structures.", + "bbox": [ + 511, + 493, + 903, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-scale features in SSL. Multi-scale features have received much attention in existing SSL research, which is relevant but not sufficient to structure-aware representations. Multi-scale feature learning in SSL can be categorized as reconstruction-based and contrastive-based meth", + "bbox": [ + 511, + 825, + 903, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20258", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ods. For instance, the contrastive-based method Wang et al. introduced a dense constraint based on the correspondence of local features for dense prediction tasks. The method [38] developed data-driven contrastive learning methods by utilizing prototypes. Zhou et al. combines multi-scale reconstruction and multi-scale contrastive learning to learn multi-scale features. Although those methods achieve success in learning multi-scale features, they are not aware of anatomical structure variations because of the fixed-size image partition strategy and uniform contribution of different anatomical structures in images.", + "bbox": [ + 89, + 90, + 483, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 89, + 268, + 189, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present our two-step approach: patch-to-patch discrepancy and patch-to-structure consistency. The patch-to-patch constraint increases the distance between all patches, ensuring that patches with different semantics are pushed further apart. In contrast, the patch-to-structure constraint keeps patches within the same structure close together, enforcing semantic consistency within the same structure.", + "bbox": [ + 89, + 295, + 483, + 401 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 409, + 230, + 424 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contrastive Learning. Contrastive learning in SSL constructs positive sample pairs to capture shared information [23, 37], typically using a source image and its augmented counterpart as two views of the same image. Based on the theory that the features learned from these paired views suffice for downstream tasks [37], we focus solely on positive pairs from the same source image rather than different ones. We adopt vanilla InfoNCE in MoCo [13] as our baseline, leveraging a teacher-student mechanism and a queue storing previously generated features. Following prior SSL methods in 3D medical imaging, images are partitioned into multiple sub-volumes to optimize memory efficiency [14, 40, 41]. Given $N$ non-overlapping subvolumes $\\{v_{i} | i = 1, 2, 3, \\dots, N\\}$ cropped from a full volume $\\mathcal{V}$ , we randomly select a subvolume $v_{i}$ and generate an augmented version $v_{i}'$ as a positive pair. The constraint follows a dictionary look-up strategy, where a query feature is matched against candidate features in the dictionary. More details can be found in Appendix.", + "bbox": [ + 89, + 431, + 483, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on the dictionary $\\{q_m^* | m = 1, 2, \\dots, K\\}$ where $q_m^*$ is the feature encoded by the teacher model $\\varphi^*$ and $K$ is the size of the dictionary and the query $q_i$ encoded by the student model $\\varphi$ , the InfoNCE loss is conducted as a contrastive loss $\\mathcal{L}_g$ :", + "bbox": [ + 89, + 718, + 483, + 795 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {g} = - \\log \\frac {\\exp \\left(q _ {i} \\cdot q _ {i} ^ {* ^ {\\prime}} / \\tau\\right)}{\\sum_ {m = 0} ^ {K} \\exp \\left(q _ {i} \\cdot q _ {m} ^ {*}\\right) / \\tau}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 801, + 483, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tau$ is a temperature hyper-parameter [13].", + "bbox": [ + 89, + 854, + 403, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Geometric equivariance. Mao et al. [21] propose that the model satisfies equivariance, a meta-property that can", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "be applied to dense feature maps and generalized to most existing vision tasks. Equivariance implies that geometric transformations applied to the input image are preserved in the output features, which benefits many tasks [3, 21]. Consequently, we expect the token features from the Transformer encoder $\\varphi$ to undergo the same homography transformation as the input images. Based on this assumption, we can establish correspondences between token features in the two input volumes to compute patch-to-patch and patch-to-structure relationships. Remarkably, we consider each token to correspond to a patch in the input volume.", + "bbox": [ + 511, + 90, + 906, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Patch-to-patch discrepancy", + "text_level": 1, + "bbox": [ + 511, + 266, + 759, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a volume $\\mathcal{V}_i$ and its augmented counterpart $\\mathcal{V}_i'$ , we build patch-to-patch correspondence between them. A reliable patch-to-patch correspondence refers to a correct one-to-one alignment among two patch sets, which treats every patch as a distinct object. This step aims to enlarge the distribution gap between all semantics and capture distinct characteristics.", + "bbox": [ + 511, + 287, + 906, + 392 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In mSSL, augmentations can be broadly classified into geometric and intensity transformations, with only geometric transformations affecting voxel positions. Geometric transformations involve modeling an affine matrix $H$ , which is applied to the target image to alter its spatial configuration. Therefore, in this section, we discuss the situation after applying an affine transformation.", + "bbox": [ + 511, + 393, + 905, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Ground truth of patch correspondence. Given two patch centers $c_{i} = (x_{i},y_{i},z_{i})$ and $c_{j} = (x_{j},y_{j},z_{j})$ from $\\mathcal{V}_{i}$ and $\\mathcal{V}_i^\\prime$ , where $x, y,$ and $z$ are coordinates, the correspondence between two patches is determined based on the re-projection distance of their central positions mapped in the input volume, as described in Eq. (2) and Eq. (3).", + "bbox": [ + 511, + 500, + 905, + 592 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} _ {g t} (i, j) \\triangleq \\left\\{ \\begin{array}{l l} 1 & i f \\langle H (c _ {i}), c _ {j} \\rangle \\wedge \\langle H ^ {- 1} (c _ {j}), c _ {i} \\rangle , \\\\ 0 & e l s e. \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 613, + 906, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $H^{-1}(\\cdot)$ denotes the inverse of the affine transformation $H(\\cdot)$ in 3D space, yielding the transformed position. The notation $\\langle \\cdot, \\cdot \\rangle$ serves as the metric to determine whether two patches originate from the same patch. Specifically, the metric is:", + "bbox": [ + 511, + 664, + 905, + 739 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\langle H \\left(c _ {i}\\right), c _ {j} \\rangle \\triangleq \\left\\{ \\begin{array}{l} 1, i f (\\bigtriangleup x _ {i j}, \\bigtriangleup y _ {i j}, \\bigtriangleup z _ {i j}) \\\\ \\in \\left[ - \\frac {W}{2}, \\frac {W}{2} \\right] \\times \\left[ - \\frac {H}{2}, \\frac {H}{2} \\right] \\times \\left[ - \\frac {D}{2}, \\frac {D}{2} \\right], \\\\ 0, e l s e, \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 747, + 906, + 815 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $W$ , $H$ and $D$ are the width, height and depth of patch and $\\triangle x_{ij}, \\triangle y_{ij}$ and $\\triangle z_{ij}$ are the coordinate differences of $H(c_i)$ and $c_j$ . With Eq. (2) and Eq. (3), the ground-truth patch-to-patch correspondences $\\mathcal{M}_{gt}$ between two images are obtained.", + "bbox": [ + 511, + 824, + 906, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20259", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/38530a1509f775f782fa63d9617831187a116249d5e6bcd1caa0cbddb9ae059d.jpg", + "image_caption": [ + "Figure 2. The pipeline of our SSL framework. $S^2 DC$ is established on patch features (i.e., token feature in vision transformer) and incorporates two main steps: (1) Patch-to-patch discrepancy. The model aligns two patch sets, which treats every patch as distinct semantics to enlarge the distribution gap between all semantics. (2) Patch-to-structure consistency. The model applies a soft regularization to align patches within the same structure. By those two steps, the framework captures semantic connectivity between anatomical structures." + ], + "image_footnote": [], + "bbox": [ + 165, + 90, + 836, + 329 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Building reliable patch-to-patch correspondences. Given token features from $\\varphi^{*}(\\mathcal{V}_{i}^{\\prime})$ and $\\varphi (\\mathcal{V}_i)$ , we can calculate the similarity map $\\mathcal{M}_t$ between tokens and get the loss between $\\mathcal{M}_t$ and $\\mathcal{M}_{gt}$ . However, similar to the issues in image matching [27, 31], reliably identifying patch correspondence in medical images is challenging due to repetitive patterns. This intractable challenge would be optimized by solving a differentiable optimal transport problem, such as the Sinkhorn algorithm [30]. The Sinkhorn algorithm is used to solve entropy-regularized optimal transport, enabling soft assignment of correspondences between feature points by iteratively refining a doubly stochastic matrix, improving robustness to noise and deformations. The efficiency of Sinkhorn is affected by the large number of iterations.", + "bbox": [ + 88, + 410, + 480, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rocco et al. propose using a dual-softmax operator to establish reliable correspondences based on neighborhood consensus, which brings mutual nearest neighbors closer to learn a more compact similarity distribution. Given the similarity map $\\mathcal{M}_t$ between two images, the similarities after applying the dual-softmax operator are:", + "bbox": [ + 89, + 637, + 482, + 728 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {M}} _ {t} (i, j) = \\operatorname {s o f t m a x} \\left(\\mathcal {M} _ {t} (i, \\cdot)\\right) \\cdot \\operatorname {s o f t m a x} \\left(\\left(\\mathcal {M} _ {t} (\\cdot , j)\\right). \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 739, + 482, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the ground-truth correspondences $\\mathcal{M}_{gt}$ and the re-assigned feature similarity matrix $\\hat{\\mathcal{M}}_t$ , we use the cross-entropy loss to minimize Kullback-Leibler divergence (KL divergence) between them [36]. This patch-to-patch loss $\\mathcal{L}_{p2p}$ is:", + "bbox": [ + 89, + 770, + 482, + 849 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p 2 p} = - \\frac {1}{| \\mathcal {M} _ {g t} |} \\sum_ {i = 0} ^ {N} \\sum_ {j = 0} ^ {N} \\mathcal {M} _ {g t} (i, j) \\times \\log (\\hat {\\mathcal {M}} _ {t} (i, j)), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 859, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N$ represents the number of patches (i.e., tokens in the 3D Transformer).", + "bbox": [ + 511, + 410, + 906, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Patch-to-structure consistency", + "text_level": 1, + "bbox": [ + 511, + 468, + 782, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The correspondence assignment focuses on enhancing reliable patch-to-patch correspondence, which treats every patch as distinct. Patch-to-patch discrepancy limits the exploration of connectivity among patches from the same semantics (i.e., same anatomical structure). Thus, a patch-to-structure correspondence should be explored to improve consistency.", + "bbox": [ + 511, + 496, + 906, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We define neighborhood similarity distribution as $\\mathcal{D}_n$ , which represents the similarity vector of an anchor patch feature $t_n$ from $\\mathcal{V}_i$ with all the patch features from $\\mathcal{V}_i'$ . As shown in Fig. 3, patch pairs with the same semantics form distinct peaks in the distribution $\\mathcal{D}_n$ , which meets the assumption that the same structures have the same semantics. Thus, the central tendency and dispersion of $\\mathcal{D}_n$ can be employed to constrain semantic connectivity. For instance, if $t_n$ is similar to several patches of $\\mathcal{V}_i'$ , indicating they may belong to the same anatomical structure, $\\mathcal{D}_n$ tends to be more dispersed (Fig. 3 (b)). In contrast, with no semantic-similar patches, the distribution of $t_n$ is more concentrated. Therefore, we utilize the characteristics of $\\mathcal{D}_n$ as a soft label to regularize the patch-to-patch loss $\\mathcal{L}_{p2p}$ to maintain underlying semantic relevance.", + "bbox": [ + 511, + 607, + 908, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce a simplified version of the Sharpe ratio [28] to reflect the dispersion and concentration of neighborhood similarity distribution $\\mathcal{D}_n$ . Given the $\\mathcal{D}_n$ of the patch feature $t_n$ , the Sharpe ratio of $t_n$ can be computed as:", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20260", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fca96db32f5aa2436152494b128a1d03d66a25befbf68266629f88604755d534.jpg", + "image_caption": [ + "(a) Patches with same semantics" + ], + "image_footnote": [], + "bbox": [ + 160, + 104, + 488, + 237 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/842e885f379ce6edf569888115f8a5aa41fe301895d8d8bdacd848d24e8a5f36.jpg", + "image_caption": [ + "(b) Anchor patch with distinct semantics", + "Figure 3. Illustration of the similarity distribution $D_{n}$ . (a) Patches with the same semantics (e.g., bone). Given an anchor patch, patches from the same semantics form several peaks in $D_{n}$ . (b) Anchor patch with distinct semantics (e.g., pancreas), $D_{n}$ shows only one large peak (its augmented patch)." + ], + "image_footnote": [], + "bbox": [ + 506, + 99, + 836, + 236 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns r _ {\\mathcal {V} _ {i}} ^ {n} = \\frac {\\operatorname* {m a x} \\left(\\mathcal {D} _ {n}\\right) - \\frac {1}{N} \\sum_ {m = 1} ^ {N} \\mathcal {D} _ {n}}{\\sigma_ {\\mathcal {D} _ {n}}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 339, + 482, + 377 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where, $\\sigma_{\\mathcal{D}_n}$ is the variance of vector $\\mathcal{D}_n$ and the notation $\\max (\\mathcal{D}_n)$ is the maximal value of $\\mathcal{D}_n$ . A low Sharpe ratio of $\\mathcal{D}_n$ indicates that the distribution is more dispersed, which means the patch feature $t_i$ has more than one similar semantic patch feature. More details can be seen in Appendix.", + "bbox": [ + 89, + 382, + 483, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$sr_{\\mathcal{V}_i}^n$ is utilized as the soft regularization to adaptively adjust $\\mathcal{L}_{p2p}$ , serving as soft evidence for maintaining a semantic connectivity relationship among patch features. The weighted losses for the feature $t_n$ of the volume $\\mathcal{V}_i$ is $\\mathcal{L}_c(n,\\cdot) = \\text{softmax}(sr_{\\mathcal{V}_i})_n\\mathcal{L}_{p2p}(n,\\cdot)$ , where $sr_{\\mathcal{V}_i} = \\{sr_{\\mathcal{V}_i}^n\\}_n$ . Similarly, we have $\\mathcal{L}_c(\\cdot,m) = \\text{softmax}(sr_{\\mathcal{V}_i'})_m\\mathcal{L}_{p2p}(\\cdot,m)$ . Thus, we have the loss of a single pair:", + "bbox": [ + 89, + 472, + 483, + 595 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nl _ {n m} = \\frac {\\left(\\operatorname {s o f t m a x} \\left(s r _ {\\mathcal {V} _ {i}}\\right) _ {n} + \\operatorname {s o f t m a x} \\left(s r _ {\\mathcal {V} _ {i} ^ {\\prime}}\\right) _ {m}\\right) \\mathcal {L} _ {p 2 p} (n , m)}{2}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 618, + 482, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the patch-to-structure loss $\\mathcal{L}_{p2s}$ is", + "bbox": [ + 89, + 664, + 370, + 680 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p 2 s} = \\frac {\\sum_ {m = 0} ^ {N} \\sum_ {n = 0} ^ {N} l _ {n , m}}{N \\times N}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 702, + 482, + 737 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Overall loss function", + "text_level": 1, + "bbox": [ + 89, + 747, + 287, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final combined objective for the proposed framework is defined as:", + "bbox": [ + 89, + 768, + 482, + 797 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {g} + \\mathcal {L} _ {p 2 p} + \\mathcal {L} _ {p 2 s}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 800, + 482, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 829, + 223, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section begins with an overview of the datasets utilized for both pre-training and downstream tasks. Next, we provide a brief summary of the implementation details for the", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "method. Lastly, we present a comprehensive comparison of the experimental results for our proposed method against other SOTA SSL approaches in 3D medical imaging.", + "bbox": [ + 511, + 329, + 906, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and implementation details", + "text_level": 1, + "bbox": [ + 511, + 388, + 828, + 405 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pre-training and downstream datasets. Although we emphasize the anatomical structure-specific semantic connectivity, the training process of our mining semantic connective strategy is also suitable for positron emission tomography (PET). In our experiments, we include computed tomography (CT), magnetic resonance imaging (MRI), and PET modalities.", + "bbox": [ + 511, + 412, + 905, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use the same dataset settings as VoCo [41] on CT, i.e., 10k data collection. BTCV and TCIA COVID-19 are used as pre-training datasets (1k data) for the BTCV segmentation task, while all other CT segmentation tasks are pre-trained on 10k data. Additionally, we collect 6605 (6k) unlabeled MRI scans from OASIS3 [18] and 3519 (3k) PET scans from UDPET [4] and ADNI [22] for pre-training. We evaluate methods on 10 datasets in 3 modalities, as shown in Tab. 2. More details about dataset splits are provided in Appendix.", + "bbox": [ + 511, + 518, + 905, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training setup. For pre-training stage, $S^2 DC$ adopts Swin-B [11] as the default backbones in three modalities, following previous works [11, 40, 41]. We set 100k iterations for pre-training in all modalities and crop subvolumes with size $96 \\times 96 \\times 96$ . In our patch-to-patch loss $\\mathcal{L}_{p2p}$ , we adopt the dual-softmax operator for efficiency, while the results of the Sinkhorn algorithm can be found in the Appendix. For downstream tasks, we adopt Swin-B and its derivative architectures. The weights of models are initialized by the weights of pre-training models. We adopt the Swin-UNETR [14] framework in segmentation, image-to-image translation, and reconstruction tasks. Swin-B [11] is used as the backbone of classification tasks. More details on pre-training and training settings are in the Appendix.", + "bbox": [ + 511, + 671, + 905, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison methods. We compare $S^2 DC$ with SOTA", + "bbox": [ + 531, + 883, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20261", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d94ffb0ccb96d412475c386cba030d878b6f75f3dd3a80ba0c0040d3ea99c4eb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDice Score(%)
SplRKidLKidGallEsoLivStoAorIVCVeinsPanRAGLAGAVG
From Scratch
UNETR [12]93.0294.1394.1266.9970.8796.1177.2789.2282.170.1676.6565.3259.2179.82
Swin-UNETR [14]94.0693.5493.8065.6174.6097.0975.9491.882.3673.6375.1968.0061.1180.53
With General SSL
MAE3D [7]93.9894.3794.1869.8674.6496.6680.4090.3083.1372.6577.1167.3460.5481.33
SimCLR [6]92.7993.0491.4149.6550.9998.4977.9285.5680.5864.3767.1659.0448.9973.85
SimMIM [42]95.5695.8294.1452.0653.5298.9880.2588.1182.9866.4969.1660.8850.4576.03
MoCo v3 [13]91.9692.8592.4268.2572.7794.9178.8288.2181.5971.1575.7666.4858.8179.54
Jigsaw [5]94.6293.4193.5575.6373.2195.7180.889.4184.7871.0279.5765.6860.2281.35
PositionLabel [47]94.3593.1593.2175.3973.2495.7680.6988.884.0471.1879.0265.1160.0781.09
With medical image SSL
MG [50]91.9993.5291.8165.1176.4195.9886.8889.2983.5971.7981.6267.9763.1881.45
ROT [32]91.7593.1391.6265.0976.5594.2186.1689.7483.0871.1381.5567.9063.7281.20
Vicregl [1]90.3294.1591.3065.1275.4194.7686.0089.1382.5471.2681.0167.6663.0880.89
Rubik++ [35]96.2190.4189.3375.2272.6497.4479.2589.6583.7674.7478.3567.1461.9781.38
PCRLv1 [48]95.7389.6688.5375.4172.3396.2078.9989.1183.0674.4777.8867.0261.8580.78
PCRLv2 [49]95.5091.4389.5276.1573.5497.2879.6490.1684.1775.2078.7168.7462.9381.74
Swin-UNETR [14]95.2593.1692.9763.6273.9696.2179.3289.9883.1976.1182.2568.9965.1181.54
SwinMM [40]94.3394.1894.1672.9774.7596.3783.2389.5682.9170.6575.5269.1762.9081.81
GL-MAE [52]94.5494.3994.3773.1974.9396.5183.4989.7483.1170.8075.7169.3963.1282.01
GVSL [15]95.2791.2292.2572.6973.5696.4482.4088.9084.2270.8476.4267.4863.2581.87
VoCo [41]95.7396.5394.4876.0275.6097.4178.4391.2186.1278.1980.8871.4767.8883.85
\\( S^2DC \\)96.2496.5194.5274.2777.6996.9583.3390.4486.6675.2284.2468.2269.7384.14
", + "bbox": [ + 91, + 88, + 906, + 422 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c3b46e87689df02e3fc8bce4ddf78b300c378a488881b9930f0492952524f1ba.jpg", + "table_caption": [ + "Table 1. Results on the segmentation task BTCV fine-tuning from pre-training models trained on 1k CT data. Most of the results are drawn from [41, 45]. Best and second best are highlighted" + ], + "table_footnote": [], + "table_body": "
TaskDatasetModality
SegmentationBTCV [19]CT
MSD-Liver [29]CT
MSD-Lung [29]CT
MSD-Spleen [29]CT
BraTs 21 [29]MRI
AUTOPET [10]PET
ClassificationCC-CCII [46]CT
ADNI-cls [22]PET
ReconstructionUDPET [4]PET
I2I translationBraTs 23 [29]MRI
", + "bbox": [ + 140, + 483, + 434, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SSL methods in general [5-7, 13, 42, 47] and medical [1, 14, 15, 32, 35, 40, 41, 48-50, 52] images. Meanwhile, we report results of training from scratch (i.e., without initialized weights of pre-training models).", + "bbox": [ + 89, + 684, + 482, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results on downstream tasks.", + "text_level": 1, + "bbox": [ + 89, + 756, + 354, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We report the results of 4 downstream tasks on 3 modalities. Following previous work, we use the metrics Dice score for segmentation, accuracy for classification, PSNR, SSIM, and NMSE for I2I translation, and SSIM and PSNR for reconstruction.", + "bbox": [ + 89, + 777, + 482, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Segmentation tasks. Compared to volume-level and voxel-level methods, $S^2 DC$ performs well both in small targets (AUTOPET, MSD-Lung) and large organs (BTCV,", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MSD-Liver, MSD-Spleen and BraTs 21) segmentation, as shown in Tab. 1 and Tab. 3. In BTCV datasets, compared to general SSL methods, medical SSL methods show superior performance. Notably, with $S^2DC$ pre-training, we achieve an average improvement of $4.32\\%$ , reaching a Dice score of $75.47\\%$ , which surpasses other SOTA medical SSL methods by $0.81\\%$ .", + "bbox": [ + 511, + 486, + 906, + 593 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Classification tasks. We conduct experiments on the CC-CCII dataset for COVID-19 classification and ADNI for AD (Alzheimer's Disease), MCI (Mild Cognitive Impairment), and CN (Cognitively Normal) classification. With the pre-trained model, $S^2 DC$ achieves an average improvement of $3.93\\%$ and outperforms other medical SSL methods.", + "bbox": [ + 511, + 593, + 908, + 700 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8e2c562ae83e3455da6357ab121e6a3c0be097d823159f19c53ecee67111225b.jpg", + "table_caption": [ + "Table 2. The downstream tasks and modalities. I2I translation represents image-to-image translation." + ], + "table_footnote": [], + "table_body": "
MethodNetworkCC-CCIIADNI-clsAVG
Acc(%)Acc(%)Acc(%)
From Scratch
Swin-UNETR [14]Swin-B88.0456.1772.10
With 3D medical image SSL
PCRLv2 [49]Swin-B93.0756.0574.56
Swin-UNETR [14]Swin-B94.1555.9575.05
SwinMM [40]Swin-B94.8056.3675.58
VoCo [41]Swin-B94.6056.3275.46
\\( S^2DC \\)Swin-B95.3456.7276.03
", + "bbox": [ + 524, + 710, + 897, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Experimental results on the classification task CC-CCII and ADNI-cls fine-tuning from pre-training models trained on 6k MRI scans. Best and second best are highlighted.", + "bbox": [ + 511, + 840, + 906, + 883 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20262", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5b6334f3f5707e19aebeb070af190f74bfb0d85226cba7b426449c28f8955505.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDice(%)
MSD LiverMSD LungMSD SpleenAUTOPETBraTs21Total
WTTCETAVGAVG
From scratch
Swin-UNETR [14]81.1859.7394.2034.4992.8988.2877.3486.1771.15
With 3D medical image SSL
Swin-UNETR [14]81.5862.4694.4943.8592.7088.4278.8286.6473.95
Swin-MM [40]82.3261.1694.6044.3792.8988.5479.1886.8773.86
PCRLv2 [49]82.1262.3995.2342.6492.8188.4978.7486.6873.81
VoCo [41]82.5063.3495.7045.1092.9088.4478.6286.6574.66
\\( S^2DC \\)83.4364.4095.7346.4792.7688.7179.7987.0975.47
", + "bbox": [ + 132, + 88, + 867, + 239 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a4a2fc53a1aa22dbf19444f5657ac463880a4071b415750f73c7419df01be819.jpg", + "table_caption": [ + "Table 3. Results on the segmentation tasks MSD-Liver, MSD-Lung, MSD-Spleen, AUTOPET, and BraTs21. WT, TC, and ET represent the whole tumor, tumor core, and enhancing tumor, respectively. We retrained all methods using $10\\mathrm{k}$ , $6\\mathrm{k}$ , and $3\\mathrm{k}$ data in three modalities separately, except for VoCo in CT, whose weight is provided in the official code. Best and second best are highlighted." + ], + "table_footnote": [], + "table_body": "
MethodT1->T2T1->T1ce
PSNR↑NMSE↓SSIM↑PSNR↑NMSE↓SSIM↑
From Scratch
Swin-UNETR [14]26.570.13480.880726.910.0820.8871
With 3D medical image SSL
PCRLv2 [49]26.790.16400.882829.620.1230.9179
Swin-UNETR [14]24.810.29380.844130.420.1080.9117
SwinMM [40]24.050.31580.818230.700.0990.9059
VoCo [41]25.810.19160.859831.920.0790.9315
S2DC25.960.11610.883932.650.0410.9412
", + "bbox": [ + 91, + 301, + 483, + 412 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8d9ebea531fbf8011fcee87a291cc4343676d9ae350a0ee23ce065f61610ac40.jpg", + "table_caption": [ + "Table 5. Experimental results on the image-to-image translation task BraTs23 fine-tuning from pre-training models trained on 6k MRI scans. Best and second best are highlighted." + ], + "table_footnote": [], + "table_body": "
MethodNetworkSSIM↑PSNR↑
From Scratch
Swin-UNETR [14]-0.9991758.4100
With 3D medical image SSL
PCRLv2 [49]Swin-UNETR0.9991658.2865
Swin-UNETR [14]Swin-UNETR0.9991758.3409
SwinMM [40]Swin-UNETR0.9991758.4002
VoCo [41]Swin-UNETR0.9991658.2561
\\( S^2DC \\)Swin-UNETR0.9991758.4649
", + "bbox": [ + 120, + 464, + 454, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Image-to-image translation task. We evaluate in BraTs23, which includes four structural MRI modalities for each individual. The translation quality for $\\mathrm{T1} \\rightarrow \\mathrm{T2}$ and $\\mathrm{T1} \\rightarrow \\mathrm{T1ce}$ are compared by peak signal-to-noise ratio (PSNR), normalized mean squared error (NMSE), and structural similarity index (SSIM) [44]. From Table 5, our method excels in two tasks. Remarkably, with $S^2 DC$ pretraining, we gain $5.41\\%$ and $0.32\\%$ improvements in SSIM separately. We notice that other methods fail to get improvement in the task $\\mathrm{T1} \\rightarrow \\mathrm{T2}$ compared to From Scratch. That may be due to the deep domain gap between the two sequences, while pre-trained models provide biased initializations. In contrast, $S^2 DC$ achieves improvement in NMSE and SSIM, showing better ability in this task.", + "bbox": [ + 89, + 643, + 482, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Reconstruction task. We evaluate the capabilities of reconstruction in different methods on the UDPET dataset. Following the previous work [8], whole-body data at", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DFR=50 (i.e., 1/50 doses) were selected to explore the reconstruction ability of all methods. This task has been extensively studied, yet our method, $S^2 DC$ , still achieves superior PSNR results.", + "bbox": [ + 511, + 305, + 903, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Overall comparisons on 10 downstream datasets. The results demonstrate that $S^2 DC$ performs effectively across 10 datasets and 4 tasks. Compared to training from scratch, which achieves an average score of $77.93\\%$ , $S^2 DC$ pre-training delivers a $3.5\\%$ improvement, reaching $81.43\\%$ . Additionally, $S^2 DC$ outperforms the second-best SSL method (VoCo [41], average score $80.65\\%$ ) for all tasks, with an average gain of $0.78\\%$ . These findings indicate that $S^2 DC$ is both reliable and adaptable across various downstream tasks and modalities.", + "bbox": [ + 511, + 367, + 906, + 517 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation study", + "text_level": 1, + "bbox": [ + 511, + 530, + 663, + 546 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Influence of different constraints. We use contrastive learning loss $\\mathcal{L}_g$ in Sec. 3.1 as our baseline. The impact of different constraints in pre-training is presented in Tab. 7. The results indicate that combining patch-to-patch loss $\\mathcal{L}_{p2p}$ and patch-to-structure loss $\\mathcal{L}_{p2s}$ yields SOTA performance compared to Tab. 1. Meanwhile, we visualize the t-SNE clustering results of 13 organs under different constraints, as shown in Fig. 4. The $\\mathcal{L}_{p2p}$ constraint leads to a sparser inter-structure distribution, while the $\\mathcal{L}_{p2s}$ constraint results in more compact intra-structure clusters. Combining $\\mathcal{L}_{p2p}$ and $\\mathcal{L}_{p2s}$ achieves the best overall clustering performance.", + "bbox": [ + 511, + 553, + 906, + 734 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8b8191f9ccd2a6ef977c0b4f0f3edc803efae3103e5c555b2cd8126713bb8946.jpg", + "table_caption": [ + "Table 6. Experimental results on low-dose to full-dose reconstruction task UDPET fine-tuning from pre-training models trained on 3k PET scans. Best and second best are highlighted." + ], + "table_footnote": [], + "table_body": "
Baseline(Lg)+Lp2p+Lp2sBTCV DICE(%)AUTOPET
•83.4345.72
••83.9845.85
••84.0246.23
•••84.1446.47
", + "bbox": [ + 542, + 747, + 877, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. The ablation results of different constraints.", + "bbox": [ + 550, + 835, + 864, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Intra-structure consistency and inter-structure discrepancy. We visualize the t-SNE of features extracted by four medical pre-trained models and calculate their silhouette", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20263", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d47ced4e1db314b13b8117c4ba47efcc95fbbbf9be78e5d7940ff15fac26a183.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 88, + 472, + 193 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9b9e96e73c6be48eacaaa298d528f1cf9b6a441120bc5b51c0229f3fe6742841.jpg", + "image_caption": [ + "(a) $\\mathcal{L}_g$", + "(c) $\\mathcal{L}_g + \\mathcal{L}_{p2s}$", + "(b) $\\mathcal{L}_g + \\mathcal{L}_{p2p}$" + ], + "image_footnote": [], + "bbox": [ + 101, + 205, + 467, + 315 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/322cb4db0aa9699f6d9708ed844f699e2b2bfd7f464a57678823e64e12245dd7.jpg", + "image_caption": [ + "(d) $\\mathcal{L}_g + \\mathcal{L}_{p2p} + \\mathcal{L}_{p2s}$" + ], + "image_footnote": [], + "bbox": [ + 299, + 207, + 465, + 316 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/18882e36428500c73bdc3f359f9b9625fd632b24afcc7e67b34ec9a0f0cfe3d0.jpg", + "image_caption": [ + "Figure 4. The t-SNE feature visualization of different losses on 13 organs on the BTCV dataset.", + "(a)Ours", + "(b) SwinMM" + ], + "image_footnote": [], + "bbox": [ + 101, + 378, + 472, + 497 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/643730787308b356cf81c4b51a96985b5aaf4bc4f7aebf049fa084ea18bb77d4.jpg", + "image_caption": [ + "Silhouette Score: 0.0381", + "(c) VoCo", + "Silhouette Score: -0.1749", + "Figure 5. The t-SNE visualization of features from 13 organs on the BTCV dataset and the silhouette scores of four methods." + ], + "image_footnote": [], + "bbox": [ + 106, + 521, + 259, + 625 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/25e70f9bea063311b9f06ba2ea63f127e09dd2100a31b3bf58c155fba9ac99d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 259, + 541, + 282, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 285, + 378, + 468, + 497 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ce1d4e767b692bf86b41e6f3ea9cbd06495565090b217b871d11eed313101271.jpg", + "image_caption": [ + "Silhouette Score:0.0027", + "(d) Swin-UNETR", + "Silhouette Score: -0.0477" + ], + "image_footnote": [], + "bbox": [ + 261, + 521, + 470, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "scores [25], which evaluate the quality of clustering by measuring how similar a data point is to its cluster (consistency) compared to other clusters (discrepancy). Fig. 5 shows the clustering capabilities of 13 organs on the BTCV dataset. $S^2 DC$ shows excellent performance on organ discrimination and consistency, which gets the highest silhouette scores. As shown in Tab. 1, most organs that cluster well also exhibit strong segmentation performance, highlighting the importance of structure-specific consistency and discrepancy. Additionally, we visualize the feature maps of 4 SSL models on 3 modalities in Fig. 6. $S^2 DC$ performs semantic consistency in the background, tissues (CT), functional regions (MRI), and metabolic regions (PET).", + "bbox": [ + 89, + 704, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bcec55db2721be3b180e82fe43727cf66bde4f036859a4383e6ed642bde83ced.jpg", + "image_caption": [ + "Figure 6. We visualize the first principal components after applying PCA to token features [2] within volumes of CT, MRI, and PET. The figure displays the central slices of feature maps." + ], + "image_footnote": [], + "bbox": [ + 535, + 88, + 890, + 244 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/024fc64c789ad5f6328f1450f7bfc55d2f2295d8cfc5283838e54097abff34db.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAccuracy(%)
10%50%100%
Swin-UNETR [14]77.1592.2894.15
SwinMM [40]87.8793.5094.80
VoCo [41]86.7392.4394.60
\\( S^2DC \\)88.2993.8995.34
", + "bbox": [ + 576, + 297, + 844, + 376 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Experiment results on CC-CCII with various ratios of the training data. ${10}\\% ,{50}\\%$ ,and ${100}\\%$ represent ratios.", + "bbox": [ + 511, + 382, + 906, + 410 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Varying ratios of the training dataset. We evaluate the performances of models that were trained with varying ratios of CC-CCII. From Tab. 8, $S^2 DC$ outperforms other methods under different ratios of training data. Notably, $S^2 DC$ achieves $88.29\\%$ accuracy using only $10\\%$ of the data, which is effective with limited dataset sizes.", + "bbox": [ + 511, + 417, + 906, + 507 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6d0e7126fed46aaa38e3acf9dc6f10c532bfee21d41a9c3e30a79492738e0856.jpg", + "image_caption": [ + "Figure 7. The evolution of patch-to-structure correspondences." + ], + "image_footnote": [], + "bbox": [ + 573, + 518, + 846, + 685 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Patch-to-structure correspondences evolution. We visualize feature map trends at different pretraining steps in Fig. 7. As training progresses, feature consistency improves, especially for small structures like bronchioles.", + "bbox": [ + 511, + 710, + 906, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 784, + 633, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose $S^2 DC$ , a medical image SSL framework that improves structure-aware semantic consistency through patch-to-patch and patch-to-structure steps. Experiments and visualizations show its superior performance. Future work could involve larger pre-training datasets or model generalization across modalities for broader applications.", + "bbox": [ + 511, + 809, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20264", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 89, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This paper was supported by the National Natural Science Foundation of China (Grant No.82394432 and 92249302), the Shanghai Municipal Science and Technology Major Project (Grant No.2023SHZDZX02). The computations in this research were performed using the CFFF platform of Fudan University.", + "bbox": [ + 89, + 113, + 483, + 200 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 224, + 187, + 242 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicregl: Self-supervised learning of local visual features. Advances in Neural Information Processing Systems, 35:8799-8810, 2022. 6", + "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 8", + "[3] Anadi Chaman and Ivan Dokmanic. Truly shift-invariant convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3773-3783, 2021. 3", + "[4] Gaoyu Chen, Sheng Liu, Wenxiang Ding, Li Lv, Chen Zhao, Fenghua Weng, Yong Long, Yunlong Zan, and Qiu Huang. A total-body ultra-low dose pet reconstruction method via image space shuffle u-net and body sampling. IEEE Transactions on Radiation and Plasma Medical Sciences, 2023. 1, 5, 6", + "[5] Pengguang Chen, Shu Liu, and Jiaya Jia. Jigsaw clustering for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11526-11535, 2021. 6", + "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 6", + "[7] Zekai Chen, Devansh Agarwal, Kshitij Aggarwal, Wiem Safta, Mariann Micsinai Balan, and Kevin Brown. Masked image modeling advances 3d medical image analysis. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1970-1980, 2023. 6", + "[8] Jiaqi Cui, Pinxian Zeng, Xinyi Zeng, Yuanyuan Xu, Peng Wang, Jiliu Zhou, Yan Wang, and Dinggang Shen. Prior knowledge-guided triple-domain transformer-gan for direct pet reconstruction from low-count sinograms. IEEE Transactions on Medical Imaging, 2024. 7", + "[9] Tugba Akinci D'Antonoli, Lucas K Berger, Ashraya K Indrakanti, Nathan Vishwanathan, Jakob Weiβ, Matthias Jung, Zeynep Berkarda, Alexander Rau, Marco Reisert, Thomas Küstner, et al. Totalsegmentator mri: Sequence-independent segmentation of 59 anatomical structures in mr images. arXiv preprint arXiv:2405.19492, 2024. 1", + "[10] Sergios Gatidis, Marcel Früh, Matthias Fabritius, Sijing Gu, Konstantin Nikolaou, Christian La Fougère, Jin Ye, Junjun He, Yige Peng, Lei Bi, et al. The autopet challenge:" + ], + "bbox": [ + 93, + 250, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "towards fully automated lesion segmentation in oncologic pet/ct imaging. 2023. 1, 6", + "[11] Ali Hatamizadeh, Vishwesh Nath, Yucheng Tang, Dong Yang, Holger R Roth, and Daguang Xu. Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images. In International MICCAI brainlesion workshop, pages 272-284. Springer, 2021. 5", + "[12] Ali Hatamizadeh, Yucheng Tang, Vishwesh Nath, Dong Yang, Andriy Myronenko, Bennett Landman, Holger R Roth, and Daguang Xu. Unetr: Transformers for 3d medical image segmentation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 574-584, 2022. 6", + "[13] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 3, 6", + "[14] Yufan He, Vishwesh Nath, Dong Yang, Yucheng Tang, Andriy Myronenko, and Daguang Xu. Swinunetr-v2: Stronger swim transformers with stagewise convolutions for 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 416-426. Springer, 2023. 1, 3, 5, 6, 7, 8", + "[15] Yuting He, Guanyu Yang, Rongjun Ge, Yang Chen, Jean-Louis Coatieux, Boyu Wang, and Shuo Li. Geometric visual similarity learning in 3d medical image self-supervised pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9538-9547, 2023. 2, 6", + "[16] Junjia Huang, Haofeng Li, Guanbin Li, and Xiang Wan. Attentive symmetric autoencoder for brain mri segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 203-213. Springer, 2022. 2", + "[17] Yankai Jiang, Mingze Sun, Heng Guo, Xiaoyu Bai, Ke Yan, Le Lu, and Minfeng Xu. Anatomical invariance modeling and semantic alignment for self-supervised learning in 3d medical image analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15859-15869, 2023. 1, 2", + "[18] Pamela J LaMontagne, Tammie LS Benzinger, John C Morris, Sarah Keefe, Russ Hornbeck, Chengjie Xiong, Elizabeth Grant, Jason Hassenstab, Krista Moulder, Andrei G Vlassenko, et al. Oasis-3: longitudinal neuroimaging, clinical, and cognitive dataset for normal aging and Alzheimer disease. medrxiv, pages 2019–12, 2019. 5", + "[19] Bennett Landman, Zhoubing Xu, J Igelsias, Martin Styner, Thomas Langerak, and Arno Klein. Miccai multi-atlas labeling beyond the cranial vault-workshop and challenge. In Proc. MICCAI Multi-Atlas Labeling Beyond Cranial Vault—Workshop Challenge, page 12, 2015. 6", + "[20] Haofeng Li, Yiming Ouyang, and Xiang Wan. Self-supervised alignment learning for medical image segmentation. In 2024 IEEE International Symposium on Biomedical Imaging (ISBI), pages 1-5. IEEE, 2024. 2", + "[21] Chengzhi Mao, Lingyu Zhang, Abhishek Vaibhav Joshi, Junfeng Yang, Hao Wang, and Carl Vondrick. Robust perception" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20265", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "through equivariance. In International Conference on Machine Learning, pages 23852-23870. PMLR, 2023. 3", + "[22] Susanne G Mueller, Michael W Weiner, Leon J Thal, Ronald C Petersen, Clifford Jack, William Jagust, John Q Trojanowski, Arthur W Toga, and Laurel Beckett. The alzheimer's disease neuroimaging initiative. Neuroimaging Clinics of North America, 15(4):869, 2005. 5, 6", + "[23] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 1, 3", + "[24] Ignacio Rocco, Mircea Cimpoi, Relja Arandjelović, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Neighbourhood consensus networks. Advances in neural information processing systems, 31, 2018. 4", + "[25] Peter J Rousseeuw. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. Journal of computational and applied mathematics, 20:53-65, 1987. 8", + "[26] Ainkaran Santhirasekaram, Mathias Winkler, Andrea Rockall, and Ben Glocker. A geometric approach to robust medical image segmentation. Medical Image Analysis, 97: 103260, 2024. 2", + "[27] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938-4947, 2020. 4", + "[28] William F Sharpe. Mutual fund performance. The Journal of business, 39(1):119-138, 1966. 4", + "[29] Amber L Simpson, Michela Antonelli, Spyridon Bakas, Michel Bilello, Keyvan Farahani, Bram Van Ginneken, Annette Kopp-Schneider, Bennett A Landman, Geert Litjens, Bjoern Menze, et al. A large annotated medical image dataset for the development and evaluation of segmentation algorithms. arXiv preprint arXiv:1902.09063, 2019. 1, 6", + "[30] Richard Sinkhorn. A relationship between arbitrary positive matrices and doubly stochastic matrices. The annals of mathematical statistics, 35(2):876-879, 1964. 4", + "[31] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 4", + "[32] Aiham Taleb, Winfried Loetzsch, Noel Danz, Julius Severin, Thomas Gaertner, Benjamin Bergner, and Christoph Lippert. 3d self-supervised methods for medical imaging. Advances in neural information processing systems, 33:18158-18172, 2020. 6", + "[33] Zhaorui Tan, Xi Yang, Tan Pan, Tianyi Liu, Chen Jiang, Xin Guo, Qiufeng Wang, Anh Nguyen, Yuan Qi, Kaizhu Huang, et al. Towards a universal 3d medical multi-modality generalization via learning personalized invariant representation. arXiv preprint arXiv:2411.06106, 2024. 1", + "[34] Yucheng Tang, Dong Yang, Wenqi Li, Holger R Roth, Bennett Landman, Daguang Xu, Vishwesh Nath, and Ali Hatamizadeh. Self-supervised pre-training of swim transformers for 3d medical image analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20730-20740, 2022. 2" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Xing Tao, Yuexiang Li, Wenhui Zhou, Kai Ma, and Yefeng Zheng. Revisiting rubik's cube: Self-supervised learning with volume-wise transformation for 3d medical image segmentation. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part IV 23, pages 238-248. Springer, 2020. 2, 6", + "[36] MTCAJ Thomas and A Thomas Joy. Elements of information theory. Wiley-Interscience, 2006. 4", + "[37] Yao-Hung Hubert Tsai, Yue Wu, Ruslan Salakhutdinov, and Louis-Philippe Morency. Self-supervised learning from a multi-view perspective. arXiv preprint arXiv:2006.05576, 2020. 3", + "[38] Chengyao Wang, Li Jiang, Xiaoyang Wu, Zhuotao Tian, Bohao Peng, Hengshuang Zhao, and Jiaya Jia. Groupcontrast: Semantic-aware self-supervised representation learning for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4917-4928, 2024. 3", + "[39] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3024-3033, 2021. 3", + "[40] Yiqing Wang, Zihan Li, Jieru Mei, Zihao Wei, Li Liu, Chen Wang, Shengtian Sang, Alan L Yuille, Cihang Xie, and Yuyin Zhou. Swinmm: masked multi-view with swin transformers for 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 486-496. Springer, 2023. 1, 2, 3, 5, 6, 7, 8", + "[41] Linshan Wu, Jiaxin Zhuang, and Hao Chen. Voco: A simple-yet-effective volume contrastive learning framework for 3d medical image analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22873-22882, 2024. 1, 2, 3, 5, 6, 7, 8", + "[42] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9653-9663, 2022. 6", + "[43] Zhangsihao Yang, Mengwei Ren, Kaize Ding, Guido Gerig, and Yalin Wang. Keypoint-augmented self-supervised learning for medical image segmentation with limited annotation. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[44] Xin Yi, Ekta Walia, and Paul Babyn. Generative adversarial network in medical imaging: A review. Medical image analysis, 58:101552, 2019. 7", + "[45] Chuyan Zhang, Hao Zheng, and Yun Gu. Dive into the details of self-supervised learning for medical image analysis. Medical Image Analysis, 89:102879, 2023. 6", + "[46] Kang Zhang, Xiaohong Liu, Jun Shen, Zhihuan Li, Ye Sang, Xingwang Wu, Yunfei Zha, Wenhua Liang, Chengdi Wang, Ke Wang, et al. Clinically applicable ai system for accurate diagnosis, quantitative measurements, and prognosis of Covid-19 pneumonia using computed tomography. Cell, 181 (6):1423-1433, 2020. 6" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20266", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Zhemin Zhang and Xun Gong. Positional label for self-supervised vision transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3516-3524, 2023. 6", + "[48] Hong-Yu Zhou, Chixiang Lu, Sibei Yang, Xiaoguang Han, and Yizhou Yu. *Preservational learning improves self-supervised medical image models by reconstructing diverse contexts*. In *Proceedings of the IEEE/CVF International Conference on Computer Vision*, pages 3499–3509, 2021. 6", + "[49] Hong-Yu Zhou, Chixiang Lu, Chaoqi Chen, Sibei Yang, and Yizhou Yu. Pcrlv2: A unified visual information preservation framework for self-supervised pre-training in medical image analysis. arXiv preprint arXiv:2301.00772, 2023. 2, 3, 6, 7", + "[50] Zongwei Zhou, Vatsal Sodha, Jiaxuan Pang, Michael B Gotway, and Jianming Liang. Models genesis. Medical image analysis, 67:101840, 2021. 6", + "[51] Jiaxin Zhuang, Linshan Wu, Qiong Wang, Varut Vardhanabhuti, Lin Luo, and Hao Chen. Mim: Mask in mask self-supervised pre-training for 3d medical image analysis. arXiv preprint arXiv:2404.15580, 2024. 2", + "[52] Jia-Xin Zhuang, Luyang Luo, and Hao Chen. Advancing volumetric medical image segmentation via global-local masked autoencoder. arXiv preprint arXiv:2306.08913, 2023. 2, 6" + ], + "bbox": [ + 89, + 90, + 485, + 443 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20267", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Structure-aware Semantic Discrepancy and Consistency for 3D Medical Image Self-supervised Learning/d73adb04-404d-4f0b-9d9c-095f1f245a9b_model.json b/2025/Structure-aware Semantic Discrepancy and Consistency for 3D Medical Image Self-supervised Learning/d73adb04-404d-4f0b-9d9c-095f1f245a9b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d7b1c313d6defa8d0648133b45dc654152b5035d --- /dev/null +++ b/2025/Structure-aware Semantic Discrepancy and Consistency for 3D Medical Image Self-supervised Learning/d73adb04-404d-4f0b-9d9c-095f1f245a9b_model.json @@ -0,0 +1,2400 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.103, + 0.13, + 0.895, + 0.177 + ], + "angle": 0, + "content": "Structure-aware Semantic Discrepancy and Consistency for 3D Medical Image Self-supervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.203, + 0.751, + 0.239 + ], + "angle": 0, + "content": "Tan Pan\\(^{1,2}\\), Zhaorui Tan\\(^{2}\\), Kaiyu Guo\\(^{3,2}\\), Dongli Xu\\(^{2}\\), Weidi Xu, Chen Jiang\\(^{1,2*}\\), Xin Guo\\(^{1,2}\\), Yuan Qi\\(^{1,2,4}\\), Yuan Cheng\\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.177, + 0.239, + 0.821, + 0.276 + ], + "angle": 0, + "content": "\\(^{1}\\) AI\\(^{3}\\), Fudan University \\(^{2}\\) Shanghai Academy of Artificial Intelligence for Science \\(^{3}\\) The University of Queensland \\(^{4}\\) Zhongshan Hospital, Fudan University" + }, + { + "type": "text", + "bbox": [ + 0.191, + 0.278, + 0.799, + 0.292 + ], + "angle": 0, + "content": "pant23@m.fudan.edu.cn, jiangchen@sais.com.cn, cheng-yuan@fudan.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.358, + 0.485, + 0.752 + ], + "angle": 0, + "content": "3D medical image self-supervised learning (mSSL) holds great promise for medical analysis. Effectively supporting broader applications requires considering anatomical structure variations in location, scale, and morphology, which are crucial for capturing meaningful distinctions. However, previous mSSL methods partition images with fixed-size patches, often ignoring the structure variations. In this work, we introduce a novel perspective on 3D medical images with the goal of learning structure-aware representations. We assume that patches within the same structure share the same semantics (semantic consistency) while those from different structures exhibit distinct semantics (semantic discrepancy). Based on this assumption, we propose an mSSL framework named \\( S^2DC \\), achieving Structure-aware Semantic Discrepancy and Consistency in two steps. First, \\( S^2DC \\) enforces distinct representations for different patches to increase semantic discrepancy by leveraging an optimal transport strategy. Second, \\( S^2DC \\) advances semantic consistency at the structural level based on neighborhood similarity distribution. By bridging patch-level and structure-level representations, \\( S^2DC \\) achieves structure-aware representations. Thoroughly evaluated across 10 datasets, 4 tasks, and 3 modalities, our proposed method consistently outperforms the state-of-the-art methods in mSSL. The code is available at https://github.com/Ashespt/S2DC/." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.78, + 0.222, + 0.795 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.805, + 0.484, + 0.867 + ], + "angle": 0, + "content": "Self-supervised learning (SSL) is widely used in medical image analysis to build strong representations that improve performance on various applications and maximize the use of unlabeled data [23, 33, 40, 41]. These applica" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.327, + 0.593, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.383, + 0.593, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.452, + 0.594, + 0.459 + ], + "angle": 0, + "content": "A. Two views of mSSL" + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.327, + 0.898, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.473, + 0.907, + 0.626 + ], + "angle": 0, + "content": "Figure 1. A. Two views of mSSL. B. Similarity heatmaps on CT and MRI images across different methods. We sample an anchor patch (red dots) and compute feature similarities with all other patches. In the first row, the liver anchor should show low similarity with non-liver patches, while in the second row, the white matter anchor should exhibit high similarity with other white matter patches. Current SOTA contrastive-based [41] (b) and reconstruction-based [40] (c) methods struggle with both patch feature discrepancy in different structures and consistency in the same structure. In contrast, our method (d) advances both discrepancy and consistency." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.657, + 0.907, + 0.779 + ], + "angle": 0, + "content": "tions frequently focus on different specific anatomical structures (i.e., organ segmentation), each distinguished across anatomy due to location, scale, morphology, and other variances [4, 9, 10, 29]. For example, the liver is a large, well-defined solid organ with relatively homogeneous intensity and distinct anatomical boundaries. In contrast, veins are thin, tubular structures with varying diameters, lower contrast, and complex branching patterns." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Current medical image SSL (mSSL) methods can be classified into contrastive-based and reconstruction-based methods [14, 17, 40, 41]. For example, Wang et al. utilizes multi-view reconstruction to encourage dense representations, while Wu et al. incorporates overlapped patch ratio as soft contrastive constraint to improve semantic consistency. However, these methods use fixed-size image partition, treating all segmented patches uniformly across struc" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "* Corresponding authors. This research was conducted during an internship at the Shanghai Academy of Artificial Intelligence for Science." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20257" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "tures and overlooking their underlying semantic relationship among structures, as shown in Fig. 1 A. To illustrate these limitations, we visualize similarity heatmaps on CT and MRI images across different methods by overlaying similarity scores onto the original images (Fig. 1 B). Ideally, patches within the same organ should exhibit high similarity, while those from different organs should show low similarity. However, as shown in Fig. 1 (a) and (b), existing methods struggle with maintaining intra-structure patch consistency and distinguishing inter-structure patch differences. A more detailed analysis and metrics of organ-level relationships in the state-of-the-art (SOTA) methods are provided in Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.29, + 0.482, + 0.412 + ], + "angle": 0, + "content": "Based on these empirical analyses, we propose that patches within the same structure should share the same semantics, while those from different structures should exhibit distinct semantics. To achieve this, patch relationships must be structure-aware and adaptively adjusted to account for anatomical variations. Based on this assumption, we define structure-aware learning through two key principles: intra-structure consistency and inter-structure discrepancy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.412, + 0.483, + 0.716 + ], + "angle": 0, + "content": "To achieve these principles, we propose \\( S^2 DC \\), a self-supervised learning framework for 3D medical imaging designed to capture Structure-aware Semantic Consistency and Discrepancy. The core idea of our framework is to explore semantic relevance between patches to enhance structure-level representations, which unfolds in two key steps. First, we increase the discrepancy between patches with different semantics through patch-to-patch discrepancy. By an optimal transport strategy, patches with different semantics are pushed further apart. Second, beyond patch-wise distinction, the framework further explores patch-to-structure semantic consistency, leveraging the observation that patches with shared semantics form peaks in the similarity distribution. These distributional characteristics serve as soft regularization, identifying patches belonging to the same structure and enforcing structural-level consistency. Notably, the principles of inter-structure discrepancy and intra-structure consistency of our framework are also suitable for functional imaging or metabolic imaging." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.718, + 0.466, + 0.732 + ], + "angle": 0, + "content": "This work's contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.735, + 0.483, + 0.839 + ], + "angle": 0, + "content": "- We propose a novel insight of intra-structure consistency and inter-structure discrepancy in the anatomical structure-aware feature learning in 3D medical images. We argue that the patches within the same anatomical structures maintain consistent semantics, while the patches from different anatomical structures reflect distinct semantics." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.841, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- Based on our insight and assumption, we propose \\( S^2 DC \\), a novel training framework that enhances interstructure discrepancy and intra-structure consistency. The framework establishes reliable patch-to-patch correspond" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "dences to reinforce discrepancy while leveraging patch-to-structure semantic connectivity from the similarity distribution to improve consistency." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.137, + 0.905, + 0.198 + ], + "angle": 0, + "content": "- Our method demonstrates superior performance over SOTA medical image SSL methods, evaluated across 10 datasets, 4 tasks, and 3 imaging modalities. The code will be available upon publication." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.211, + 0.651, + 0.227 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.236, + 0.905, + 0.493 + ], + "angle": 0, + "content": "SSL in 3D medical images. Contrastive-based and reconstruction-based SSL are among the most prominent paradigms in mSSL. Some approaches emphasize the fine-grained capabilities of models, incorporating techniques such as multi-scale reconstruction or reconstruction from disarranged images to enhance model performance [35, 51, 52]. Simultaneously, other methods [41, 49] prioritize preserving maximal semantic information by leveraging contextual priors. Recently, several studies have explored the combination of contrastive learning (CL) and reconstruction to further improve model performance [15, 17, 34, 40]. These works aim to learn consistent semantic representations by applying contrastive learning to global features or predicting missing information from available context. However, in mSSL, few approaches enforce constraints to enhance both anatomical structure-aware discrepancy and consistency, which is the central focus of this paper." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.494, + 0.905, + 0.825 + ], + "angle": 0, + "content": "Geometric approaches for medical image analysis. Our method adopts some geometric characteristics for patch-to-patch correspondence. In radiographic imaging, geometric characteristics, such as location and anatomical structure, play a crucial role in organ identification, disease diagnosis, and medical analysis. Several approaches [16, 26] have incorporated consistent and recurrent anatomical patterns as learnable priors to enhance medical image analysis. For example, Huang et al. proposed to enhance segmentation by establishing patch correspondence between two 2D images. Those methods utilizing the geometric or anatomical invariant have shown remarkable improvement in the downstream tasks. Some medical SSL works [15, 20] also explore geometric correspondence for better performance. The SSL method GVSL [15] considers the geometric transformation and reconstructs affine images to original images to learn global and local representations. Yang et al. present a keypoint-augmented fusion layer with the keypoint guidance from SIFT to extract local and global information to improve segmentation results. Those methods leverage the invariance of anatomical structures while not considering the relation between anatomical structures." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Multi-scale features in SSL. Multi-scale features have received much attention in existing SSL research, which is relevant but not sufficient to structure-aware representations. Multi-scale feature learning in SSL can be categorized as reconstruction-based and contrastive-based meth" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20258" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.258 + ], + "angle": 0, + "content": "ods. For instance, the contrastive-based method Wang et al. introduced a dense constraint based on the correspondence of local features for dense prediction tasks. The method [38] developed data-driven contrastive learning methods by utilizing prototypes. Zhou et al. combines multi-scale reconstruction and multi-scale contrastive learning to learn multi-scale features. Although those methods achieve success in learning multi-scale features, they are not aware of anatomical structure variations because of the fixed-size image partition strategy and uniform contribution of different anatomical structures in images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.27, + 0.191, + 0.287 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.296, + 0.484, + 0.402 + ], + "angle": 0, + "content": "We present our two-step approach: patch-to-patch discrepancy and patch-to-structure consistency. The patch-to-patch constraint increases the distance between all patches, ensuring that patches with different semantics are pushed further apart. In contrast, the patch-to-structure constraint keeps patches within the same structure close together, enforcing semantic consistency within the same structure." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.41, + 0.232, + 0.425 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.432, + 0.484, + 0.718 + ], + "angle": 0, + "content": "Contrastive Learning. Contrastive learning in SSL constructs positive sample pairs to capture shared information [23, 37], typically using a source image and its augmented counterpart as two views of the same image. Based on the theory that the features learned from these paired views suffice for downstream tasks [37], we focus solely on positive pairs from the same source image rather than different ones. We adopt vanilla InfoNCE in MoCo [13] as our baseline, leveraging a teacher-student mechanism and a queue storing previously generated features. Following prior SSL methods in 3D medical imaging, images are partitioned into multiple sub-volumes to optimize memory efficiency [14, 40, 41]. Given \\( N \\) non-overlapping subvolumes \\( \\{v_{i} | i = 1, 2, 3, \\dots, N\\} \\) cropped from a full volume \\( \\mathcal{V} \\), we randomly select a subvolume \\( v_{i} \\) and generate an augmented version \\( v_{i}' \\) as a positive pair. The constraint follows a dictionary look-up strategy, where a query feature is matched against candidate features in the dictionary. More details can be found in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.484, + 0.796 + ], + "angle": 0, + "content": "Based on the dictionary \\(\\{q_m^* | m = 1, 2, \\dots, K\\}\\) where \\(q_m^*\\) is the feature encoded by the teacher model \\(\\varphi^*\\) and \\(K\\) is the size of the dictionary and the query \\(q_i\\) encoded by the student model \\(\\varphi\\), the InfoNCE loss is conducted as a contrastive loss \\(\\mathcal{L}_g\\):" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.803, + 0.484, + 0.851 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {g} = - \\log \\frac {\\exp \\left(q _ {i} \\cdot q _ {i} ^ {* ^ {\\prime}} / \\tau\\right)}{\\sum_ {m = 0} ^ {K} \\exp \\left(q _ {i} \\cdot q _ {m} ^ {*}\\right) / \\tau}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.405, + 0.871 + ], + "angle": 0, + "content": "where \\(\\tau\\) is a temperature hyper-parameter [13]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Geometric equivariance. Mao et al. [21] propose that the model satisfies equivariance, a meta-property that can" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.259 + ], + "angle": 0, + "content": "be applied to dense feature maps and generalized to most existing vision tasks. Equivariance implies that geometric transformations applied to the input image are preserved in the output features, which benefits many tasks [3, 21]. Consequently, we expect the token features from the Transformer encoder \\(\\varphi\\) to undergo the same homography transformation as the input images. Based on this assumption, we can establish correspondences between token features in the two input volumes to compute patch-to-patch and patch-to-structure relationships. Remarkably, we consider each token to correspond to a patch in the input volume." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.267, + 0.761, + 0.284 + ], + "angle": 0, + "content": "3.2. Patch-to-patch discrepancy" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.907, + 0.393 + ], + "angle": 0, + "content": "Given a volume \\(\\mathcal{V}_i\\) and its augmented counterpart \\(\\mathcal{V}_i'\\), we build patch-to-patch correspondence between them. A reliable patch-to-patch correspondence refers to a correct one-to-one alignment among two patch sets, which treats every patch as a distinct object. This step aims to enlarge the distribution gap between all semantics and capture distinct characteristics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.906, + 0.5 + ], + "angle": 0, + "content": "In mSSL, augmentations can be broadly classified into geometric and intensity transformations, with only geometric transformations affecting voxel positions. Geometric transformations involve modeling an affine matrix \\( H \\), which is applied to the target image to alter its spatial configuration. Therefore, in this section, we discuss the situation after applying an affine transformation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.5, + 0.906, + 0.593 + ], + "angle": 0, + "content": "Ground truth of patch correspondence. Given two patch centers \\( c_{i} = (x_{i},y_{i},z_{i}) \\) and \\( c_{j} = (x_{j},y_{j},z_{j}) \\) from \\( \\mathcal{V}_{i} \\) and \\( \\mathcal{V}_i^\\prime \\), where \\( x, y, \\) and \\( z \\) are coordinates, the correspondence between two patches is determined based on the re-projection distance of their central positions mapped in the input volume, as described in Eq. (2) and Eq. (3)." + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.614, + 0.907, + 0.656 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} _ {g t} (i, j) \\triangleq \\left\\{ \\begin{array}{l l} 1 & i f \\langle H (c _ {i}), c _ {j} \\rangle \\wedge \\langle H ^ {- 1} (c _ {j}), c _ {i} \\rangle , \\\\ 0 & e l s e. \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.665, + 0.906, + 0.74 + ], + "angle": 0, + "content": "Here, \\( H^{-1}(\\cdot) \\) denotes the inverse of the affine transformation \\( H(\\cdot) \\) in 3D space, yielding the transformed position. The notation \\( \\langle \\cdot, \\cdot \\rangle \\) serves as the metric to determine whether two patches originate from the same patch. Specifically, the metric is:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.748, + 0.907, + 0.816 + ], + "angle": 0, + "content": "\\[\n\\langle H \\left(c _ {i}\\right), c _ {j} \\rangle \\triangleq \\left\\{ \\begin{array}{l} 1, i f (\\bigtriangleup x _ {i j}, \\bigtriangleup y _ {i j}, \\bigtriangleup z _ {i j}) \\\\ \\in \\left[ - \\frac {W}{2}, \\frac {W}{2} \\right] \\times \\left[ - \\frac {H}{2}, \\frac {H}{2} \\right] \\times \\left[ - \\frac {D}{2}, \\frac {D}{2} \\right], \\\\ 0, e l s e, \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.825, + 0.907, + 0.9 + ], + "angle": 0, + "content": "Here, \\(W\\), \\(H\\) and \\(D\\) are the width, height and depth of patch and \\(\\triangle x_{ij}, \\triangle y_{ij}\\) and \\(\\triangle z_{ij}\\) are the coordinate differences of \\(H(c_i)\\) and \\(c_j\\). With Eq. (2) and Eq. (3), the ground-truth patch-to-patch correspondences \\(\\mathcal{M}_{gt}\\) between two images are obtained." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20259" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.166, + 0.092, + 0.837, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.337, + 0.908, + 0.396 + ], + "angle": 0, + "content": "Figure 2. The pipeline of our SSL framework. \\( S^2 DC \\) is established on patch features (i.e., token feature in vision transformer) and incorporates two main steps: (1) Patch-to-patch discrepancy. The model aligns two patch sets, which treats every patch as distinct semantics to enlarge the distribution gap between all semantics. (2) Patch-to-structure consistency. The model applies a soft regularization to align patches within the same structure. By those two steps, the framework captures semantic connectivity between anatomical structures." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.411, + 0.482, + 0.636 + ], + "angle": 0, + "content": "Building reliable patch-to-patch correspondences. Given token features from \\(\\varphi^{*}(\\mathcal{V}_{i}^{\\prime})\\) and \\(\\varphi (\\mathcal{V}_i)\\), we can calculate the similarity map \\(\\mathcal{M}_t\\) between tokens and get the loss between \\(\\mathcal{M}_t\\) and \\(\\mathcal{M}_{gt}\\). However, similar to the issues in image matching [27, 31], reliably identifying patch correspondence in medical images is challenging due to repetitive patterns. This intractable challenge would be optimized by solving a differentiable optimal transport problem, such as the Sinkhorn algorithm [30]. The Sinkhorn algorithm is used to solve entropy-regularized optimal transport, enabling soft assignment of correspondences between feature points by iteratively refining a doubly stochastic matrix, improving robustness to noise and deformations. The efficiency of Sinkhorn is affected by the large number of iterations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.638, + 0.483, + 0.729 + ], + "angle": 0, + "content": "Rocco et al. propose using a dual-softmax operator to establish reliable correspondences based on neighborhood consensus, which brings mutual nearest neighbors closer to learn a more compact similarity distribution. Given the similarity map \\(\\mathcal{M}_t\\) between two images, the similarities after applying the dual-softmax operator are:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.74, + 0.483, + 0.759 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {M}} _ {t} (i, j) = \\operatorname {s o f t m a x} \\left(\\mathcal {M} _ {t} (i, \\cdot)\\right) \\cdot \\operatorname {s o f t m a x} \\left(\\left(\\mathcal {M} _ {t} (\\cdot , j)\\right). \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.771, + 0.483, + 0.851 + ], + "angle": 0, + "content": "Based on the ground-truth correspondences \\(\\mathcal{M}_{gt}\\) and the re-assigned feature similarity matrix \\(\\hat{\\mathcal{M}}_t\\), we use the cross-entropy loss to minimize Kullback-Leibler divergence (KL divergence) between them [36]. This patch-to-patch loss \\(\\mathcal{L}_{p2p}\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.861, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p 2 p} = - \\frac {1}{| \\mathcal {M} _ {g t} |} \\sum_ {i = 0} ^ {N} \\sum_ {j = 0} ^ {N} \\mathcal {M} _ {g t} (i, j) \\times \\log (\\hat {\\mathcal {M}} _ {t} (i, j)), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.411, + 0.907, + 0.44 + ], + "angle": 0, + "content": "where \\(N\\) represents the number of patches (i.e., tokens in the 3D Transformer)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.469, + 0.783, + 0.485 + ], + "angle": 0, + "content": "3.3. Patch-to-structure consistency" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.497, + 0.907, + 0.603 + ], + "angle": 0, + "content": "The correspondence assignment focuses on enhancing reliable patch-to-patch correspondence, which treats every patch as distinct. Patch-to-patch discrepancy limits the exploration of connectivity among patches from the same semantics (i.e., same anatomical structure). Thus, a patch-to-structure correspondence should be explored to improve consistency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.608, + 0.909, + 0.835 + ], + "angle": 0, + "content": "We define neighborhood similarity distribution as \\(\\mathcal{D}_n\\), which represents the similarity vector of an anchor patch feature \\(t_n\\) from \\(\\mathcal{V}_i\\) with all the patch features from \\(\\mathcal{V}_i'\\). As shown in Fig. 3, patch pairs with the same semantics form distinct peaks in the distribution \\(\\mathcal{D}_n\\), which meets the assumption that the same structures have the same semantics. Thus, the central tendency and dispersion of \\(\\mathcal{D}_n\\) can be employed to constrain semantic connectivity. For instance, if \\(t_n\\) is similar to several patches of \\(\\mathcal{V}_i'\\), indicating they may belong to the same anatomical structure, \\(\\mathcal{D}_n\\) tends to be more dispersed (Fig. 3 (b)). In contrast, with no semantic-similar patches, the distribution of \\(t_n\\) is more concentrated. Therefore, we utilize the characteristics of \\(\\mathcal{D}_n\\) as a soft label to regularize the patch-to-patch loss \\(\\mathcal{L}_{p2p}\\) to maintain underlying semantic relevance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.908, + 0.902 + ], + "angle": 0, + "content": "We introduce a simplified version of the Sharpe ratio [28] to reflect the dispersion and concentration of neighborhood similarity distribution \\( \\mathcal{D}_n \\). Given the \\( \\mathcal{D}_n \\) of the patch feature \\( t_n \\), the Sharpe ratio of \\( t_n \\) can be computed as:" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20260" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.161, + 0.106, + 0.49, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.199, + 0.246, + 0.448, + 0.264 + ], + "angle": 0, + "content": "(a) Patches with same semantics" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.101, + 0.837, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.519, + 0.245, + 0.83, + 0.264 + ], + "angle": 0, + "content": "(b) Anchor patch with distinct semantics" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.274, + 0.908, + 0.318 + ], + "angle": 0, + "content": "Figure 3. Illustration of the similarity distribution \\( D_{n} \\). (a) Patches with the same semantics (e.g., bone). Given an anchor patch, patches from the same semantics form several peaks in \\( D_{n} \\). (b) Anchor patch with distinct semantics (e.g., pancreas), \\( D_{n} \\) shows only one large peak (its augmented patch)." + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.34, + 0.483, + 0.378 + ], + "angle": 0, + "content": "\\[\ns r _ {\\mathcal {V} _ {i}} ^ {n} = \\frac {\\operatorname* {m a x} \\left(\\mathcal {D} _ {n}\\right) - \\frac {1}{N} \\sum_ {m = 1} ^ {N} \\mathcal {D} _ {n}}{\\sigma_ {\\mathcal {D} _ {n}}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.383, + 0.484, + 0.472 + ], + "angle": 0, + "content": "where, \\(\\sigma_{\\mathcal{D}_n}\\) is the variance of vector \\(\\mathcal{D}_n\\) and the notation \\(\\max (\\mathcal{D}_n)\\) is the maximal value of \\(\\mathcal{D}_n\\). A low Sharpe ratio of \\(\\mathcal{D}_n\\) indicates that the distribution is more dispersed, which means the patch feature \\(t_i\\) has more than one similar semantic patch feature. More details can be seen in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.473, + 0.484, + 0.596 + ], + "angle": 0, + "content": "\\( sr_{\\mathcal{V}_i}^n \\) is utilized as the soft regularization to adaptively adjust \\( \\mathcal{L}_{p2p} \\), serving as soft evidence for maintaining a semantic connectivity relationship among patch features. The weighted losses for the feature \\( t_n \\) of the volume \\( \\mathcal{V}_i \\) is \\( \\mathcal{L}_c(n,\\cdot) = \\text{softmax}(sr_{\\mathcal{V}_i})_n\\mathcal{L}_{p2p}(n,\\cdot) \\), where \\( sr_{\\mathcal{V}_i} = \\{sr_{\\mathcal{V}_i}^n\\}_n \\). Similarly, we have \\( \\mathcal{L}_c(\\cdot,m) = \\text{softmax}(sr_{\\mathcal{V}_i'})_m\\mathcal{L}_{p2p}(\\cdot,m) \\). Thus, we have the loss of a single pair:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.619, + 0.483, + 0.665 + ], + "angle": 0, + "content": "\\[\nl _ {n m} = \\frac {\\left(\\operatorname {s o f t m a x} \\left(s r _ {\\mathcal {V} _ {i}}\\right) _ {n} + \\operatorname {s o f t m a x} \\left(s r _ {\\mathcal {V} _ {i} ^ {\\prime}}\\right) _ {m}\\right) \\mathcal {L} _ {p 2 p} (n , m)}{2}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.665, + 0.371, + 0.681 + ], + "angle": 0, + "content": "Finally, the patch-to-structure loss \\(\\mathcal{L}_{p2s}\\) is" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.703, + 0.483, + 0.738 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p 2 s} = \\frac {\\sum_ {m = 0} ^ {N} \\sum_ {n = 0} ^ {N} l _ {n , m}}{N \\times N}. \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.748, + 0.288, + 0.763 + ], + "angle": 0, + "content": "3.4. Overall loss function" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.77, + 0.483, + 0.799 + ], + "angle": 0, + "content": "The final combined objective for the proposed framework is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.801, + 0.483, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {g} + \\mathcal {L} _ {p 2 p} + \\mathcal {L} _ {p 2 s}, \\tag {9}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.83, + 0.224, + 0.848 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "This section begins with an overview of the datasets utilized for both pre-training and downstream tasks. Next, we provide a brief summary of the implementation details for the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.33, + 0.907, + 0.376 + ], + "angle": 0, + "content": "method. Lastly, we present a comprehensive comparison of the experimental results for our proposed method against other SOTA SSL approaches in 3D medical imaging." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.389, + 0.829, + 0.406 + ], + "angle": 0, + "content": "4.1. Datasets and implementation details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.413, + 0.906, + 0.518 + ], + "angle": 0, + "content": "Pre-training and downstream datasets. Although we emphasize the anatomical structure-specific semantic connectivity, the training process of our mining semantic connective strategy is also suitable for positron emission tomography (PET). In our experiments, we include computed tomography (CT), magnetic resonance imaging (MRI), and PET modalities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.906, + 0.672 + ], + "angle": 0, + "content": "We use the same dataset settings as VoCo [41] on CT, i.e., 10k data collection. BTCV and TCIA COVID-19 are used as pre-training datasets (1k data) for the BTCV segmentation task, while all other CT segmentation tasks are pre-trained on 10k data. Additionally, we collect 6605 (6k) unlabeled MRI scans from OASIS3 [18] and 3519 (3k) PET scans from UDPET [4] and ADNI [22] for pre-training. We evaluate methods on 10 datasets in 3 modalities, as shown in Tab. 2. More details about dataset splits are provided in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.906, + 0.884 + ], + "angle": 0, + "content": "Training setup. For pre-training stage, \\( S^2 DC \\) adopts Swin-B [11] as the default backbones in three modalities, following previous works [11, 40, 41]. We set 100k iterations for pre-training in all modalities and crop subvolumes with size \\( 96 \\times 96 \\times 96 \\). In our patch-to-patch loss \\( \\mathcal{L}_{p2p} \\), we adopt the dual-softmax operator for efficiency, while the results of the Sinkhorn algorithm can be found in the Appendix. For downstream tasks, we adopt Swin-B and its derivative architectures. The weights of models are initialized by the weights of pre-training models. We adopt the Swin-UNETR [14] framework in segmentation, image-to-image translation, and reconstruction tasks. Swin-B [11] is used as the backbone of classification tasks. More details on pre-training and training settings are in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.885, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Comparison methods. We compare \\( S^2 DC \\) with SOTA" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20261" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.907, + 0.423 + ], + "angle": 0, + "content": "
MethodDice Score(%)
SplRKidLKidGallEsoLivStoAorIVCVeinsPanRAGLAGAVG
From Scratch
UNETR [12]93.0294.1394.1266.9970.8796.1177.2789.2282.170.1676.6565.3259.2179.82
Swin-UNETR [14]94.0693.5493.8065.6174.6097.0975.9491.882.3673.6375.1968.0061.1180.53
With General SSL
MAE3D [7]93.9894.3794.1869.8674.6496.6680.4090.3083.1372.6577.1167.3460.5481.33
SimCLR [6]92.7993.0491.4149.6550.9998.4977.9285.5680.5864.3767.1659.0448.9973.85
SimMIM [42]95.5695.8294.1452.0653.5298.9880.2588.1182.9866.4969.1660.8850.4576.03
MoCo v3 [13]91.9692.8592.4268.2572.7794.9178.8288.2181.5971.1575.7666.4858.8179.54
Jigsaw [5]94.6293.4193.5575.6373.2195.7180.889.4184.7871.0279.5765.6860.2281.35
PositionLabel [47]94.3593.1593.2175.3973.2495.7680.6988.884.0471.1879.0265.1160.0781.09
With medical image SSL
MG [50]91.9993.5291.8165.1176.4195.9886.8889.2983.5971.7981.6267.9763.1881.45
ROT [32]91.7593.1391.6265.0976.5594.2186.1689.7483.0871.1381.5567.9063.7281.20
Vicregl [1]90.3294.1591.3065.1275.4194.7686.0089.1382.5471.2681.0167.6663.0880.89
Rubik++ [35]96.2190.4189.3375.2272.6497.4479.2589.6583.7674.7478.3567.1461.9781.38
PCRLv1 [48]95.7389.6688.5375.4172.3396.2078.9989.1183.0674.4777.8867.0261.8580.78
PCRLv2 [49]95.5091.4389.5276.1573.5497.2879.6490.1684.1775.2078.7168.7462.9381.74
Swin-UNETR [14]95.2593.1692.9763.6273.9696.2179.3289.9883.1976.1182.2568.9965.1181.54
SwinMM [40]94.3394.1894.1672.9774.7596.3783.2389.5682.9170.6575.5269.1762.9081.81
GL-MAE [52]94.5494.3994.3773.1974.9396.5183.4989.7483.1170.8075.7169.3963.1282.01
GVSL [15]95.2791.2292.2572.6973.5696.4482.4088.9084.2270.8476.4267.4863.2581.87
VoCo [41]95.7396.5394.4876.0275.6097.4178.4391.2186.1278.1980.8871.4767.8883.85
\\( S^2DC \\)96.2496.5194.5274.2777.6996.9583.3390.4486.6675.2284.2468.2269.7384.14
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.433, + 0.907, + 0.463 + ], + "angle": 0, + "content": "Table 1. Results on the segmentation task BTCV fine-tuning from pre-training models trained on 1k CT data. Most of the results are drawn from [41, 45]. Best and second best are highlighted" + }, + { + "type": "table", + "bbox": [ + 0.141, + 0.484, + 0.435, + 0.634 + ], + "angle": 0, + "content": "
TaskDatasetModality
SegmentationBTCV [19]CT
MSD-Liver [29]CT
MSD-Lung [29]CT
MSD-Spleen [29]CT
BraTs 21 [29]MRI
AUTOPET [10]PET
ClassificationCC-CCII [46]CT
ADNI-cls [22]PET
ReconstructionUDPET [4]PET
I2I translationBraTs 23 [29]MRI
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.673 + ], + "angle": 0, + "content": "Table 2. The downstream tasks and modalities. I2I translation represents image-to-image translation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.685, + 0.483, + 0.748 + ], + "angle": 0, + "content": "SSL methods in general [5-7, 13, 42, 47] and medical [1, 14, 15, 32, 35, 40, 41, 48-50, 52] images. Meanwhile, we report results of training from scratch (i.e., without initialized weights of pre-training models)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.757, + 0.355, + 0.772 + ], + "angle": 0, + "content": "4.2. Results on downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.779, + 0.483, + 0.854 + ], + "angle": 0, + "content": "We report the results of 4 downstream tasks on 3 modalities. Following previous work, we use the metrics Dice score for segmentation, accuracy for classification, PSNR, SSIM, and NMSE for I2I translation, and SSIM and PSNR for reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Segmentation tasks. Compared to volume-level and voxel-level methods, \\( S^2 DC \\) performs well both in small targets (AUTOPET, MSD-Lung) and large organs (BTCV," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.487, + 0.907, + 0.594 + ], + "angle": 0, + "content": "MSD-Liver, MSD-Spleen and BraTs 21) segmentation, as shown in Tab. 1 and Tab. 3. In BTCV datasets, compared to general SSL methods, medical SSL methods show superior performance. Notably, with \\( S^2DC \\) pre-training, we achieve an average improvement of \\( 4.32\\% \\), reaching a Dice score of \\( 75.47\\% \\), which surpasses other SOTA medical SSL methods by \\( 0.81\\% \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.909, + 0.701 + ], + "angle": 0, + "content": "Classification tasks. We conduct experiments on the CC-CCII dataset for COVID-19 classification and ADNI for AD (Alzheimer's Disease), MCI (Mild Cognitive Impairment), and CN (Cognitively Normal) classification. With the pre-trained model, \\( S^2 DC \\) achieves an average improvement of \\( 3.93\\% \\) and outperforms other medical SSL methods." + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.712, + 0.898, + 0.837 + ], + "angle": 0, + "content": "
MethodNetworkCC-CCIIADNI-clsAVG
Acc(%)Acc(%)Acc(%)
From Scratch
Swin-UNETR [14]Swin-B88.0456.1772.10
With 3D medical image SSL
PCRLv2 [49]Swin-B93.0756.0574.56
Swin-UNETR [14]Swin-B94.1555.9575.05
SwinMM [40]Swin-B94.8056.3675.58
VoCo [41]Swin-B94.6056.3275.46
\\( S^2DC \\)Swin-B95.3456.7276.03
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.842, + 0.907, + 0.885 + ], + "angle": 0, + "content": "Table 4. Experimental results on the classification task CC-CCII and ADNI-cls fine-tuning from pre-training models trained on 6k MRI scans. Best and second best are highlighted." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20262" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.133, + 0.089, + 0.868, + 0.24 + ], + "angle": 0, + "content": "
MethodDice(%)
MSD LiverMSD LungMSD SpleenAUTOPETBraTs21Total
WTTCETAVGAVG
From scratch
Swin-UNETR [14]81.1859.7394.2034.4992.8988.2877.3486.1771.15
With 3D medical image SSL
Swin-UNETR [14]81.5862.4694.4943.8592.7088.4278.8286.6473.95
Swin-MM [40]82.3261.1694.6044.3792.8988.5479.1886.8773.86
PCRLv2 [49]82.1262.3995.2342.6492.8188.4978.7486.6873.81
VoCo [41]82.5063.3495.7045.1092.9088.4478.6286.6574.66
\\( S^2DC \\)83.4364.4095.7346.4792.7688.7179.7987.0975.47
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.251, + 0.905, + 0.294 + ], + "angle": 0, + "content": "Table 3. Results on the segmentation tasks MSD-Liver, MSD-Lung, MSD-Spleen, AUTOPET, and BraTs21. WT, TC, and ET represent the whole tumor, tumor core, and enhancing tumor, respectively. We retrained all methods using \\(10\\mathrm{k}\\), \\(6\\mathrm{k}\\), and \\(3\\mathrm{k}\\) data in three modalities separately, except for VoCo in CT, whose weight is provided in the official code. Best and second best are highlighted." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.303, + 0.485, + 0.413 + ], + "angle": 0, + "content": "
MethodT1->T2T1->T1ce
PSNR↑NMSE↓SSIM↑PSNR↑NMSE↓SSIM↑
From Scratch
Swin-UNETR [14]26.570.13480.880726.910.0820.8871
With 3D medical image SSL
PCRLv2 [49]26.790.16400.882829.620.1230.9179
Swin-UNETR [14]24.810.29380.844130.420.1080.9117
SwinMM [40]24.050.31580.818230.700.0990.9059
VoCo [41]25.810.19160.859831.920.0790.9315
S2DC25.960.11610.883932.650.0410.9412
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.418, + 0.483, + 0.46 + ], + "angle": 0, + "content": "Table 5. Experimental results on the image-to-image translation task BraTs23 fine-tuning from pre-training models trained on 6k MRI scans. Best and second best are highlighted." + }, + { + "type": "table", + "bbox": [ + 0.121, + 0.465, + 0.455, + 0.582 + ], + "angle": 0, + "content": "
MethodNetworkSSIM↑PSNR↑
From Scratch
Swin-UNETR [14]-0.9991758.4100
With 3D medical image SSL
PCRLv2 [49]Swin-UNETR0.9991658.2865
Swin-UNETR [14]Swin-UNETR0.9991758.3409
SwinMM [40]Swin-UNETR0.9991758.4002
VoCo [41]Swin-UNETR0.9991658.2561
\\( S^2DC \\)Swin-UNETR0.9991758.4649
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.587, + 0.483, + 0.63 + ], + "angle": 0, + "content": "Table 6. Experimental results on low-dose to full-dose reconstruction task UDPET fine-tuning from pre-training models trained on 3k PET scans. Best and second best are highlighted." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Image-to-image translation task. We evaluate in BraTs23, which includes four structural MRI modalities for each individual. The translation quality for \\(\\mathrm{T1} \\rightarrow \\mathrm{T2}\\) and \\(\\mathrm{T1} \\rightarrow \\mathrm{T1ce}\\) are compared by peak signal-to-noise ratio (PSNR), normalized mean squared error (NMSE), and structural similarity index (SSIM) [44]. From Table 5, our method excels in two tasks. Remarkably, with \\(S^2 DC\\) pretraining, we gain \\(5.41\\%\\) and \\(0.32\\%\\) improvements in SSIM separately. We notice that other methods fail to get improvement in the task \\(\\mathrm{T1} \\rightarrow \\mathrm{T2}\\) compared to From Scratch. That may be due to the deep domain gap between the two sequences, while pre-trained models provide biased initializations. In contrast, \\(S^2 DC\\) achieves improvement in NMSE and SSIM, showing better ability in this task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Reconstruction task. We evaluate the capabilities of reconstruction in different methods on the UDPET dataset. Following the previous work [8], whole-body data at" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.306, + 0.905, + 0.366 + ], + "angle": 0, + "content": "DFR=50 (i.e., 1/50 doses) were selected to explore the reconstruction ability of all methods. This task has been extensively studied, yet our method, \\( S^2 DC \\), still achieves superior PSNR results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.368, + 0.907, + 0.518 + ], + "angle": 0, + "content": "Overall comparisons on 10 downstream datasets. The results demonstrate that \\( S^2 DC \\) performs effectively across 10 datasets and 4 tasks. Compared to training from scratch, which achieves an average score of \\( 77.93\\% \\), \\( S^2 DC \\) pre-training delivers a \\( 3.5\\% \\) improvement, reaching \\( 81.43\\% \\). Additionally, \\( S^2 DC \\) outperforms the second-best SSL method (VoCo [41], average score \\( 80.65\\% \\)) for all tasks, with an average gain of \\( 0.78\\% \\). These findings indicate that \\( S^2 DC \\) is both reliable and adaptable across various downstream tasks and modalities." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.531, + 0.665, + 0.547 + ], + "angle": 0, + "content": "4.3. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.907, + 0.735 + ], + "angle": 0, + "content": "Influence of different constraints. We use contrastive learning loss \\(\\mathcal{L}_g\\) in Sec. 3.1 as our baseline. The impact of different constraints in pre-training is presented in Tab. 7. The results indicate that combining patch-to-patch loss \\(\\mathcal{L}_{p2p}\\) and patch-to-structure loss \\(\\mathcal{L}_{p2s}\\) yields SOTA performance compared to Tab. 1. Meanwhile, we visualize the t-SNE clustering results of 13 organs under different constraints, as shown in Fig. 4. The \\(\\mathcal{L}_{p2p}\\) constraint leads to a sparser inter-structure distribution, while the \\(\\mathcal{L}_{p2s}\\) constraint results in more compact intra-structure clusters. Combining \\(\\mathcal{L}_{p2p}\\) and \\(\\mathcal{L}_{p2s}\\) achieves the best overall clustering performance." + }, + { + "type": "table", + "bbox": [ + 0.544, + 0.748, + 0.878, + 0.831 + ], + "angle": 0, + "content": "
Baseline(Lg)+Lp2p+Lp2sBTCV DICE(%)AUTOPET
•83.4345.72
••83.9845.85
••84.0246.23
•••84.1446.47
" + }, + { + "type": "table_caption", + "bbox": [ + 0.552, + 0.837, + 0.866, + 0.85 + ], + "angle": 0, + "content": "Table 7. The ablation results of different constraints." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Intra-structure consistency and inter-structure discrepancy. We visualize the t-SNE of features extracted by four medical pre-trained models and calculate their silhouette" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20263" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.473, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.174, + 0.196, + 0.203, + 0.206 + ], + "angle": 0, + "content": "(a) \\(\\mathcal{L}_g\\)" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.207, + 0.468, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.159, + 0.317, + 0.217, + 0.329 + ], + "angle": 0, + "content": "(c) \\(\\mathcal{L}_g + \\mathcal{L}_{p2s}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.195, + 0.408, + 0.207 + ], + "angle": 0, + "content": "(b) \\(\\mathcal{L}_g + \\mathcal{L}_{p2p}\\)" + }, + { + "type": "image", + "bbox": [ + 0.3, + 0.208, + 0.466, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.331, + 0.317, + 0.422, + 0.329 + ], + "angle": 0, + "content": "(d) \\(\\mathcal{L}_g + \\mathcal{L}_{p2p} + \\mathcal{L}_{p2s}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.343, + 0.483, + 0.371 + ], + "angle": 0, + "content": "Figure 4. The t-SNE feature visualization of different losses on 13 organs on the BTCV dataset." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.379, + 0.473, + 0.498 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.165, + 0.499, + 0.207, + 0.509 + ], + "angle": 0, + "content": "(a)Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.51, + 0.249, + 0.52 + ], + "angle": 0, + "content": "Silhouette Score: 0.0381" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.522, + 0.261, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.169, + 0.628, + 0.213, + 0.638 + ], + "angle": 0, + "content": "(c) VoCo" + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.639, + 0.257, + 0.648 + ], + "angle": 0, + "content": "Silhouette Score: -0.1749" + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.542, + 0.283, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.379, + 0.47, + 0.498 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.343, + 0.499, + 0.408, + 0.509 + ], + "angle": 0, + "content": "(b) SwinMM" + }, + { + "type": "image_caption", + "bbox": [ + 0.312, + 0.51, + 0.438, + 0.52 + ], + "angle": 0, + "content": "Silhouette Score:0.0027" + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.522, + 0.472, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.333, + 0.627, + 0.419, + 0.637 + ], + "angle": 0, + "content": "(d) Swin-UNETR" + }, + { + "type": "image_caption", + "bbox": [ + 0.307, + 0.638, + 0.44, + 0.647 + ], + "angle": 0, + "content": "Silhouette Score: -0.0477" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.665, + 0.483, + 0.693 + ], + "angle": 0, + "content": "Figure 5. The t-SNE visualization of features from 13 organs on the BTCV dataset and the silhouette scores of four methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.901 + ], + "angle": 0, + "content": "scores [25], which evaluate the quality of clustering by measuring how similar a data point is to its cluster (consistency) compared to other clusters (discrepancy). Fig. 5 shows the clustering capabilities of 13 organs on the BTCV dataset. \\( S^2 DC \\) shows excellent performance on organ discrimination and consistency, which gets the highest silhouette scores. As shown in Tab. 1, most organs that cluster well also exhibit strong segmentation performance, highlighting the importance of structure-specific consistency and discrepancy. Additionally, we visualize the feature maps of 4 SSL models on 3 modalities in Fig. 6. \\( S^2 DC \\) performs semantic consistency in the background, tissues (CT), functional regions (MRI), and metabolic regions (PET)." + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.089, + 0.891, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.25, + 0.907, + 0.293 + ], + "angle": 0, + "content": "Figure 6. We visualize the first principal components after applying PCA to token features [2] within volumes of CT, MRI, and PET. The figure displays the central slices of feature maps." + }, + { + "type": "table", + "bbox": [ + 0.577, + 0.298, + 0.846, + 0.377 + ], + "angle": 0, + "content": "
MethodAccuracy(%)
10%50%100%
Swin-UNETR [14]77.1592.2894.15
SwinMM [40]87.8793.5094.80
VoCo [41]86.7392.4394.60
\\( S^2DC \\)88.2993.8995.34
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.383, + 0.907, + 0.411 + ], + "angle": 0, + "content": "Table 8. Experiment results on CC-CCII with various ratios of the training data. \\( {10}\\% ,{50}\\% \\) ,and \\( {100}\\% \\) represent ratios." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.418, + 0.907, + 0.508 + ], + "angle": 0, + "content": "Varying ratios of the training dataset. We evaluate the performances of models that were trained with varying ratios of CC-CCII. From Tab. 8, \\( S^2 DC \\) outperforms other methods under different ratios of training data. Notably, \\( S^2 DC \\) achieves \\( 88.29\\% \\) accuracy using only \\( 10\\% \\) of the data, which is effective with limited dataset sizes." + }, + { + "type": "image", + "bbox": [ + 0.574, + 0.52, + 0.847, + 0.686 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.69, + 0.896, + 0.705 + ], + "angle": 0, + "content": "Figure 7. The evolution of patch-to-structure correspondences." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.711, + 0.907, + 0.773 + ], + "angle": 0, + "content": "Patch-to-structure correspondences evolution. We visualize feature map trends at different pretraining steps in Fig. 7. As training progresses, feature consistency improves, especially for small structures like bronchioles." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.785, + 0.634, + 0.8 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We propose \\( S^2 DC \\), a medical image SSL framework that improves structure-aware semantic consistency through patch-to-patch and patch-to-structure steps. Experiments and visualizations show its superior performance. Future work could involve larger pre-training datasets or model generalization across modalities for broader applications." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20264" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.201 + ], + "angle": 0, + "content": "This paper was supported by the National Natural Science Foundation of China (Grant No.82394432 and 92249302), the Shanghai Municipal Science and Technology Major Project (Grant No.2023SHZDZX02). The computations in this research were performed using the CFFF platform of Fudan University." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.226, + 0.188, + 0.243 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.251, + 0.484, + 0.307 + ], + "angle": 0, + "content": "[1] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicregl: Self-supervised learning of local visual features. Advances in Neural Information Processing Systems, 35:8799-8810, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.309, + 0.484, + 0.378 + ], + "angle": 0, + "content": "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.38, + 0.483, + 0.435 + ], + "angle": 0, + "content": "[3] Anadi Chaman and Ivan Dokmanic. Truly shift-invariant convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3773-3783, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.436, + 0.484, + 0.519 + ], + "angle": 0, + "content": "[4] Gaoyu Chen, Sheng Liu, Wenxiang Ding, Li Lv, Chen Zhao, Fenghua Weng, Yong Long, Yunlong Zan, and Qiu Huang. A total-body ultra-low dose pet reconstruction method via image space shuffle u-net and body sampling. IEEE Transactions on Radiation and Plasma Medical Sciences, 2023. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.521, + 0.484, + 0.577 + ], + "angle": 0, + "content": "[5] Pengguang Chen, Shu Liu, and Jiaya Jia. Jigsaw clustering for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11526-11535, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.578, + 0.484, + 0.633 + ], + "angle": 0, + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.634, + 0.484, + 0.704 + ], + "angle": 0, + "content": "[7] Zekai Chen, Devansh Agarwal, Kshitij Aggarwal, Wiem Safta, Mariann Micsinai Balan, and Kevin Brown. Masked image modeling advances 3d medical image analysis. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1970-1980, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.705, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[8] Jiaqi Cui, Pinxian Zeng, Xinyi Zeng, Yuanyuan Xu, Peng Wang, Jiliu Zhou, Yan Wang, and Dinggang Shen. Prior knowledge-guided triple-domain transformer-gan for direct pet reconstruction from low-count sinograms. IEEE Transactions on Medical Imaging, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.775, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[9] Tugba Akinci D'Antonoli, Lucas K Berger, Ashraya K Indrakanti, Nathan Vishwanathan, Jakob Weiβ, Matthias Jung, Zeynep Berkarda, Alexander Rau, Marco Reisert, Thomas Küstner, et al. Totalsegmentator mri: Sequence-independent segmentation of 59 anatomical structures in mr images. arXiv preprint arXiv:2405.19492, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[10] Sergios Gatidis, Marcel Früh, Matthias Fabritius, Sijing Gu, Konstantin Nikolaou, Christian La Fougère, Jin Ye, Junjun He, Yige Peng, Lei Bi, et al. The autopet challenge:" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.251, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "towards fully automated lesion segmentation in oncologic pet/ct imaging. 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[11] Ali Hatamizadeh, Vishwesh Nath, Yucheng Tang, Dong Yang, Holger R Roth, and Daguang Xu. Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images. In International MICCAI brainlesion workshop, pages 272-284. Springer, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[12] Ali Hatamizadeh, Yucheng Tang, Vishwesh Nath, Dong Yang, Andriy Myronenko, Bennett Landman, Holger R Roth, and Daguang Xu. Unetr: Transformers for 3d medical image segmentation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 574-584, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[13] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.344, + 0.905, + 0.427 + ], + "angle": 0, + "content": "[14] Yufan He, Vishwesh Nath, Dong Yang, Yucheng Tang, Andriy Myronenko, and Daguang Xu. Swinunetr-v2: Stronger swim transformers with stagewise convolutions for 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 416-426. Springer, 2023. 1, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.428, + 0.905, + 0.509 + ], + "angle": 0, + "content": "[15] Yuting He, Guanyu Yang, Rongjun Ge, Yang Chen, Jean-Louis Coatieux, Boyu Wang, and Shuo Li. Geometric visual similarity learning in 3d medical image self-supervised pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9538-9547, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.511, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[16] Junjia Huang, Haofeng Li, Guanbin Li, and Xiang Wan. Attentive symmetric autoencoder for brain mri segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 203-213. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[17] Yankai Jiang, Mingze Sun, Heng Guo, Xiaoyu Bai, Ke Yan, Le Lu, and Minfeng Xu. Anatomical invariance modeling and semantic alignment for self-supervised learning in 3d medical image analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15859-15869, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[18] Pamela J LaMontagne, Tammie LS Benzinger, John C Morris, Sarah Keefe, Russ Hornbeck, Chengjie Xiong, Elizabeth Grant, Jason Hassenstab, Krista Moulder, Andrei G Vlassenko, et al. Oasis-3: longitudinal neuroimaging, clinical, and cognitive dataset for normal aging and Alzheimer disease. medrxiv, pages 2019–12, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.748, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[19] Bennett Landman, Zhoubing Xu, J Igelsias, Martin Styner, Thomas Langerak, and Arno Klein. Miccai multi-atlas labeling beyond the cranial vault-workshop and challenge. In Proc. MICCAI Multi-Atlas Labeling Beyond Cranial Vault—Workshop Challenge, page 12, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.905, + 0.873 + ], + "angle": 0, + "content": "[20] Haofeng Li, Yiming Ouyang, and Xiang Wan. Self-supervised alignment learning for medical image segmentation. In 2024 IEEE International Symposium on Biomedical Imaging (ISBI), pages 1-5. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[21] Chengzhi Mao, Lingyu Zhang, Abhishek Vaibhav Joshi, Junfeng Yang, Hao Wang, and Carl Vondrick. Robust perception" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20265" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "through equivariance. In International Conference on Machine Learning, pages 23852-23870. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.121, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[22] Susanne G Mueller, Michael W Weiner, Leon J Thal, Ronald C Petersen, Clifford Jack, William Jagust, John Q Trojanowski, Arthur W Toga, and Laurel Beckett. The alzheimer's disease neuroimaging initiative. Neuroimaging Clinics of North America, 15(4):869, 2005. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.191, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[23] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.287 + ], + "angle": 0, + "content": "[24] Ignacio Rocco, Mircea Cimpoi, Relja Arandjelović, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Neighbourhood consensus networks. Advances in neural information processing systems, 31, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.289, + 0.482, + 0.329 + ], + "angle": 0, + "content": "[25] Peter J Rousseeuw. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. Journal of computational and applied mathematics, 20:53-65, 1987. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.331, + 0.482, + 0.384 + ], + "angle": 0, + "content": "[26] Ainkaran Santhirasekaram, Mathias Winkler, Andrea Rockall, and Ben Glocker. A geometric approach to robust medical image segmentation. Medical Image Analysis, 97: 103260, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.483, + 0.455 + ], + "angle": 0, + "content": "[27] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938-4947, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.457, + 0.483, + 0.482 + ], + "angle": 0, + "content": "[28] William F Sharpe. Mutual fund performance. The Journal of business, 39(1):119-138, 1966. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.566 + ], + "angle": 0, + "content": "[29] Amber L Simpson, Michela Antonelli, Spyridon Bakas, Michel Bilello, Keyvan Farahani, Bram Van Ginneken, Annette Kopp-Schneider, Bennett A Landman, Geert Litjens, Bjoern Menze, et al. A large annotated medical image dataset for the development and evaluation of segmentation algorithms. arXiv preprint arXiv:1902.09063, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.568, + 0.482, + 0.608 + ], + "angle": 0, + "content": "[30] Richard Sinkhorn. A relationship between arbitrary positive matrices and doubly stochastic matrices. The annals of mathematical statistics, 35(2):876-879, 1964. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.61, + 0.482, + 0.677 + ], + "angle": 0, + "content": "[31] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[32] Aiham Taleb, Winfried Loetzsch, Noel Danz, Julius Severin, Thomas Gaertner, Benjamin Bergner, and Christoph Lippert. 3d self-supervised methods for medical imaging. Advances in neural information processing systems, 33:18158-18172, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.749, + 0.482, + 0.817 + ], + "angle": 0, + "content": "[33] Zhaorui Tan, Xi Yang, Tan Pan, Tianyi Liu, Chen Jiang, Xin Guo, Qiufeng Wang, Anh Nguyen, Yuan Qi, Kaizhu Huang, et al. Towards a universal 3d medical multi-modality generalization via learning personalized invariant representation. arXiv preprint arXiv:2411.06106, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[34] Yucheng Tang, Dong Yang, Wenqi Li, Holger R Roth, Bennett Landman, Daguang Xu, Vishwesh Nath, and Ali Hatamizadeh. Self-supervised pre-training of swim transformers for 3d medical image analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20730-20740, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[35] Xing Tao, Yuexiang Li, Wenhui Zhou, Kai Ma, and Yefeng Zheng. Revisiting rubik's cube: Self-supervised learning with volume-wise transformation for 3d medical image segmentation. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part IV 23, pages 238-248. Springer, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[36] MTCAJ Thomas and A Thomas Joy. Elements of information theory. Wiley-Interscience, 2006. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[37] Yao-Hung Hubert Tsai, Yue Wu, Ruslan Salakhutdinov, and Louis-Philippe Morency. Self-supervised learning from a multi-view perspective. arXiv preprint arXiv:2006.05576, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.355 + ], + "angle": 0, + "content": "[38] Chengyao Wang, Li Jiang, Xiaoyang Wu, Zhuotao Tian, Bohao Peng, Hengshuang Zhao, and Jiaya Jia. Groupcontrast: Semantic-aware self-supervised representation learning for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4917-4928, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.357, + 0.905, + 0.425 + ], + "angle": 0, + "content": "[39] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3024-3033, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.428, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[40] Yiqing Wang, Zihan Li, Jieru Mei, Zihao Wei, Li Liu, Chen Wang, Shengtian Sang, Alan L Yuille, Cihang Xie, and Yuyin Zhou. Swinmm: masked multi-view with swin transformers for 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 486-496. Springer, 2023. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[41] Linshan Wu, Jiaxin Zhuang, and Hao Chen. Voco: A simple-yet-effective volume contrastive learning framework for 3d medical image analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22873-22882, 2024. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[42] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9653-9663, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.732 + ], + "angle": 0, + "content": "[43] Zhangsihao Yang, Mengwei Ren, Kaize Ding, Guido Gerig, and Yalin Wang. Keypoint-augmented self-supervised learning for medical image segmentation with limited annotation. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.734, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[44] Xin Yi, Ekta Walia, and Paul Babyn. Generative adversarial network in medical imaging: A review. Medical image analysis, 58:101552, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.904, + 0.817 + ], + "angle": 0, + "content": "[45] Chuyan Zhang, Hao Zheng, and Yun Gu. Dive into the details of self-supervised learning for medical image analysis. Medical Image Analysis, 89:102879, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[46] Kang Zhang, Xiaohong Liu, Jun Shen, Zhihuan Li, Ye Sang, Xingwang Wu, Yunfei Zha, Wenhua Liang, Chengdi Wang, Ke Wang, et al. Clinically applicable ai system for accurate diagnosis, quantitative measurements, and prognosis of Covid-19 pneumonia using computed tomography. Cell, 181 (6):1423-1433, 2020. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20266" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.092, + 0.486, + 0.146 + ], + "angle": 0, + "content": "[47] Zhemin Zhang and Xun Gong. Positional label for self-supervised vision transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3516-3524, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.149, + 0.485, + 0.219 + ], + "angle": 0, + "content": "[48] Hong-Yu Zhou, Chixiang Lu, Sibei Yang, Xiaoguang Han, and Yizhou Yu. *Preservational learning improves self-supervised medical image models by reconstructing diverse contexts*. In *Proceedings of the IEEE/CVF International Conference on Computer Vision*, pages 3499–3509, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.483, + 0.288 + ], + "angle": 0, + "content": "[49] Hong-Yu Zhou, Chixiang Lu, Chaoqi Chen, Sibei Yang, and Yizhou Yu. Pcrlv2: A unified visual information preservation framework for self-supervised pre-training in medical image analysis. arXiv preprint arXiv:2301.00772, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.291, + 0.483, + 0.331 + ], + "angle": 0, + "content": "[50] Zongwei Zhou, Vatsal Sodha, Jiaxuan Pang, Michael B Gotway, and Jianming Liang. Models genesis. Medical image analysis, 67:101840, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.333, + 0.483, + 0.388 + ], + "angle": 0, + "content": "[51] Jiaxin Zhuang, Linshan Wu, Qiong Wang, Varut Vardhanabhuti, Lin Luo, and Hao Chen. Mim: Mask in mask self-supervised pre-training for 3d medical image analysis. arXiv preprint arXiv:2404.15580, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.39, + 0.483, + 0.444 + ], + "angle": 0, + "content": "[52] Jia-Xin Zhuang, Luyang Luo, and Hao Chen. Advancing volumetric medical image segmentation via global-local masked autoencoder. arXiv preprint arXiv:2306.08913, 2023. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "20267" + } + ] +] \ No newline at end of file diff --git a/2025/Structure-aware Semantic Discrepancy and Consistency for 3D Medical Image Self-supervised Learning/d73adb04-404d-4f0b-9d9c-095f1f245a9b_origin.pdf b/2025/Structure-aware Semantic Discrepancy and Consistency for 3D Medical Image Self-supervised Learning/d73adb04-404d-4f0b-9d9c-095f1f245a9b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..10412200837f63fc37cea64ce57a86021fba60b1 --- /dev/null +++ b/2025/Structure-aware Semantic Discrepancy and Consistency for 3D Medical Image Self-supervised Learning/d73adb04-404d-4f0b-9d9c-095f1f245a9b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:641948389a1dd5cdb30925316f9ba6ba2987551203d1ebf1c98e7c87525b53fa +size 17861172 diff --git a/2025/Structure-aware Semantic Discrepancy and Consistency for 3D Medical Image Self-supervised Learning/full.md b/2025/Structure-aware Semantic Discrepancy and Consistency for 3D Medical Image Self-supervised Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a9d0c81d2966280351fc54ba60e7086f09cf257e --- /dev/null +++ b/2025/Structure-aware Semantic Discrepancy and Consistency for 3D Medical Image Self-supervised Learning/full.md @@ -0,0 +1,335 @@ +# Structure-aware Semantic Discrepancy and Consistency for 3D Medical Image Self-supervised Learning + +Tan Pan $^{1,2}$ , Zhaorui Tan $^{2}$ , Kaiyu Guo $^{3,2}$ , Dongli Xu $^{2}$ , Weidi Xu, Chen Jiang $^{1,2*}$ , Xin Guo $^{1,2}$ , Yuan Qi $^{1,2,4}$ , Yuan Cheng $^{1,2*}$ + +$^{1}$ AI $^{3}$ , Fudan University $^{2}$ Shanghai Academy of Artificial Intelligence for Science $^{3}$ The University of Queensland $^{4}$ Zhongshan Hospital, Fudan University + +pant23@m.fudan.edu.cn, jiangchen@sais.com.cn, cheng-yuan@fudan.edu.cn + +# Abstract + +3D medical image self-supervised learning (mSSL) holds great promise for medical analysis. Effectively supporting broader applications requires considering anatomical structure variations in location, scale, and morphology, which are crucial for capturing meaningful distinctions. However, previous mSSL methods partition images with fixed-size patches, often ignoring the structure variations. In this work, we introduce a novel perspective on 3D medical images with the goal of learning structure-aware representations. We assume that patches within the same structure share the same semantics (semantic consistency) while those from different structures exhibit distinct semantics (semantic discrepancy). Based on this assumption, we propose an mSSL framework named $S^2DC$ , achieving Structure-aware Semantic Discrepancy and Consistency in two steps. First, $S^2DC$ enforces distinct representations for different patches to increase semantic discrepancy by leveraging an optimal transport strategy. Second, $S^2DC$ advances semantic consistency at the structural level based on neighborhood similarity distribution. By bridging patch-level and structure-level representations, $S^2DC$ achieves structure-aware representations. Thoroughly evaluated across 10 datasets, 4 tasks, and 3 modalities, our proposed method consistently outperforms the state-of-the-art methods in mSSL. The code is available at https://github.com/Ashespt/S2DC/. + +# 1. Introduction + +Self-supervised learning (SSL) is widely used in medical image analysis to build strong representations that improve performance on various applications and maximize the use of unlabeled data [23, 33, 40, 41]. These applica + +![](images/98d082471e3a49804679308836c0ec15e106963488061adb787739266abea581.jpg) + +![](images/007fe4fd6055ab11529d337e318f0dc8ea566271ed8892065b14269e3385fd75.jpg) +A. Two views of mSSL + +![](images/9d3b6ed803d593e8a5e4236c766658744ae3109dc48675fb32d4ba87d627815f.jpg) +Figure 1. A. Two views of mSSL. B. Similarity heatmaps on CT and MRI images across different methods. We sample an anchor patch (red dots) and compute feature similarities with all other patches. In the first row, the liver anchor should show low similarity with non-liver patches, while in the second row, the white matter anchor should exhibit high similarity with other white matter patches. Current SOTA contrastive-based [41] (b) and reconstruction-based [40] (c) methods struggle with both patch feature discrepancy in different structures and consistency in the same structure. In contrast, our method (d) advances both discrepancy and consistency. + +tions frequently focus on different specific anatomical structures (i.e., organ segmentation), each distinguished across anatomy due to location, scale, morphology, and other variances [4, 9, 10, 29]. For example, the liver is a large, well-defined solid organ with relatively homogeneous intensity and distinct anatomical boundaries. In contrast, veins are thin, tubular structures with varying diameters, lower contrast, and complex branching patterns. + +Current medical image SSL (mSSL) methods can be classified into contrastive-based and reconstruction-based methods [14, 17, 40, 41]. For example, Wang et al. utilizes multi-view reconstruction to encourage dense representations, while Wu et al. incorporates overlapped patch ratio as soft contrastive constraint to improve semantic consistency. However, these methods use fixed-size image partition, treating all segmented patches uniformly across struc + +tures and overlooking their underlying semantic relationship among structures, as shown in Fig. 1 A. To illustrate these limitations, we visualize similarity heatmaps on CT and MRI images across different methods by overlaying similarity scores onto the original images (Fig. 1 B). Ideally, patches within the same organ should exhibit high similarity, while those from different organs should show low similarity. However, as shown in Fig. 1 (a) and (b), existing methods struggle with maintaining intra-structure patch consistency and distinguishing inter-structure patch differences. A more detailed analysis and metrics of organ-level relationships in the state-of-the-art (SOTA) methods are provided in Fig. 5. + +Based on these empirical analyses, we propose that patches within the same structure should share the same semantics, while those from different structures should exhibit distinct semantics. To achieve this, patch relationships must be structure-aware and adaptively adjusted to account for anatomical variations. Based on this assumption, we define structure-aware learning through two key principles: intra-structure consistency and inter-structure discrepancy. + +To achieve these principles, we propose $S^2 DC$ , a self-supervised learning framework for 3D medical imaging designed to capture Structure-aware Semantic Consistency and Discrepancy. The core idea of our framework is to explore semantic relevance between patches to enhance structure-level representations, which unfolds in two key steps. First, we increase the discrepancy between patches with different semantics through patch-to-patch discrepancy. By an optimal transport strategy, patches with different semantics are pushed further apart. Second, beyond patch-wise distinction, the framework further explores patch-to-structure semantic consistency, leveraging the observation that patches with shared semantics form peaks in the similarity distribution. These distributional characteristics serve as soft regularization, identifying patches belonging to the same structure and enforcing structural-level consistency. Notably, the principles of inter-structure discrepancy and intra-structure consistency of our framework are also suitable for functional imaging or metabolic imaging. + +This work's contributions are summarized as follows: + +- We propose a novel insight of intra-structure consistency and inter-structure discrepancy in the anatomical structure-aware feature learning in 3D medical images. We argue that the patches within the same anatomical structures maintain consistent semantics, while the patches from different anatomical structures reflect distinct semantics. +- Based on our insight and assumption, we propose $S^2 DC$ , a novel training framework that enhances interstructure discrepancy and intra-structure consistency. The framework establishes reliable patch-to-patch correspond + +dences to reinforce discrepancy while leveraging patch-to-structure semantic connectivity from the similarity distribution to improve consistency. + +- Our method demonstrates superior performance over SOTA medical image SSL methods, evaluated across 10 datasets, 4 tasks, and 3 imaging modalities. The code will be available upon publication. + +# 2. Related work + +SSL in 3D medical images. Contrastive-based and reconstruction-based SSL are among the most prominent paradigms in mSSL. Some approaches emphasize the fine-grained capabilities of models, incorporating techniques such as multi-scale reconstruction or reconstruction from disarranged images to enhance model performance [35, 51, 52]. Simultaneously, other methods [41, 49] prioritize preserving maximal semantic information by leveraging contextual priors. Recently, several studies have explored the combination of contrastive learning (CL) and reconstruction to further improve model performance [15, 17, 34, 40]. These works aim to learn consistent semantic representations by applying contrastive learning to global features or predicting missing information from available context. However, in mSSL, few approaches enforce constraints to enhance both anatomical structure-aware discrepancy and consistency, which is the central focus of this paper. + +Geometric approaches for medical image analysis. Our method adopts some geometric characteristics for patch-to-patch correspondence. In radiographic imaging, geometric characteristics, such as location and anatomical structure, play a crucial role in organ identification, disease diagnosis, and medical analysis. Several approaches [16, 26] have incorporated consistent and recurrent anatomical patterns as learnable priors to enhance medical image analysis. For example, Huang et al. proposed to enhance segmentation by establishing patch correspondence between two 2D images. Those methods utilizing the geometric or anatomical invariant have shown remarkable improvement in the downstream tasks. Some medical SSL works [15, 20] also explore geometric correspondence for better performance. The SSL method GVSL [15] considers the geometric transformation and reconstructs affine images to original images to learn global and local representations. Yang et al. present a keypoint-augmented fusion layer with the keypoint guidance from SIFT to extract local and global information to improve segmentation results. Those methods leverage the invariance of anatomical structures while not considering the relation between anatomical structures. + +Multi-scale features in SSL. Multi-scale features have received much attention in existing SSL research, which is relevant but not sufficient to structure-aware representations. Multi-scale feature learning in SSL can be categorized as reconstruction-based and contrastive-based meth + +ods. For instance, the contrastive-based method Wang et al. introduced a dense constraint based on the correspondence of local features for dense prediction tasks. The method [38] developed data-driven contrastive learning methods by utilizing prototypes. Zhou et al. combines multi-scale reconstruction and multi-scale contrastive learning to learn multi-scale features. Although those methods achieve success in learning multi-scale features, they are not aware of anatomical structure variations because of the fixed-size image partition strategy and uniform contribution of different anatomical structures in images. + +# 3. Methods + +We present our two-step approach: patch-to-patch discrepancy and patch-to-structure consistency. The patch-to-patch constraint increases the distance between all patches, ensuring that patches with different semantics are pushed further apart. In contrast, the patch-to-structure constraint keeps patches within the same structure close together, enforcing semantic consistency within the same structure. + +# 3.1. Preliminaries + +Contrastive Learning. Contrastive learning in SSL constructs positive sample pairs to capture shared information [23, 37], typically using a source image and its augmented counterpart as two views of the same image. Based on the theory that the features learned from these paired views suffice for downstream tasks [37], we focus solely on positive pairs from the same source image rather than different ones. We adopt vanilla InfoNCE in MoCo [13] as our baseline, leveraging a teacher-student mechanism and a queue storing previously generated features. Following prior SSL methods in 3D medical imaging, images are partitioned into multiple sub-volumes to optimize memory efficiency [14, 40, 41]. Given $N$ non-overlapping subvolumes $\{v_{i} | i = 1, 2, 3, \dots, N\}$ cropped from a full volume $\mathcal{V}$ , we randomly select a subvolume $v_{i}$ and generate an augmented version $v_{i}'$ as a positive pair. The constraint follows a dictionary look-up strategy, where a query feature is matched against candidate features in the dictionary. More details can be found in Appendix. + +Based on the dictionary $\{q_m^* | m = 1, 2, \dots, K\}$ where $q_m^*$ is the feature encoded by the teacher model $\varphi^*$ and $K$ is the size of the dictionary and the query $q_i$ encoded by the student model $\varphi$ , the InfoNCE loss is conducted as a contrastive loss $\mathcal{L}_g$ : + +$$ +\mathcal {L} _ {g} = - \log \frac {\exp \left(q _ {i} \cdot q _ {i} ^ {* ^ {\prime}} / \tau\right)}{\sum_ {m = 0} ^ {K} \exp \left(q _ {i} \cdot q _ {m} ^ {*}\right) / \tau}, \tag {1} +$$ + +where $\tau$ is a temperature hyper-parameter [13]. + +Geometric equivariance. Mao et al. [21] propose that the model satisfies equivariance, a meta-property that can + +be applied to dense feature maps and generalized to most existing vision tasks. Equivariance implies that geometric transformations applied to the input image are preserved in the output features, which benefits many tasks [3, 21]. Consequently, we expect the token features from the Transformer encoder $\varphi$ to undergo the same homography transformation as the input images. Based on this assumption, we can establish correspondences between token features in the two input volumes to compute patch-to-patch and patch-to-structure relationships. Remarkably, we consider each token to correspond to a patch in the input volume. + +# 3.2. Patch-to-patch discrepancy + +Given a volume $\mathcal{V}_i$ and its augmented counterpart $\mathcal{V}_i'$ , we build patch-to-patch correspondence between them. A reliable patch-to-patch correspondence refers to a correct one-to-one alignment among two patch sets, which treats every patch as a distinct object. This step aims to enlarge the distribution gap between all semantics and capture distinct characteristics. + +In mSSL, augmentations can be broadly classified into geometric and intensity transformations, with only geometric transformations affecting voxel positions. Geometric transformations involve modeling an affine matrix $H$ , which is applied to the target image to alter its spatial configuration. Therefore, in this section, we discuss the situation after applying an affine transformation. + +Ground truth of patch correspondence. Given two patch centers $c_{i} = (x_{i},y_{i},z_{i})$ and $c_{j} = (x_{j},y_{j},z_{j})$ from $\mathcal{V}_{i}$ and $\mathcal{V}_i^\prime$ , where $x, y,$ and $z$ are coordinates, the correspondence between two patches is determined based on the re-projection distance of their central positions mapped in the input volume, as described in Eq. (2) and Eq. (3). + +$$ +\mathcal {M} _ {g t} (i, j) \triangleq \left\{ \begin{array}{l l} 1 & i f \langle H (c _ {i}), c _ {j} \rangle \wedge \langle H ^ {- 1} (c _ {j}), c _ {i} \rangle , \\ 0 & e l s e. \end{array} \right. \tag {2} +$$ + +Here, $H^{-1}(\cdot)$ denotes the inverse of the affine transformation $H(\cdot)$ in 3D space, yielding the transformed position. The notation $\langle \cdot, \cdot \rangle$ serves as the metric to determine whether two patches originate from the same patch. Specifically, the metric is: + +$$ +\langle H \left(c _ {i}\right), c _ {j} \rangle \triangleq \left\{ \begin{array}{l} 1, i f (\bigtriangleup x _ {i j}, \bigtriangleup y _ {i j}, \bigtriangleup z _ {i j}) \\ \in \left[ - \frac {W}{2}, \frac {W}{2} \right] \times \left[ - \frac {H}{2}, \frac {H}{2} \right] \times \left[ - \frac {D}{2}, \frac {D}{2} \right], \\ 0, e l s e, \end{array} \right. \tag {3} +$$ + +Here, $W$ , $H$ and $D$ are the width, height and depth of patch and $\triangle x_{ij}, \triangle y_{ij}$ and $\triangle z_{ij}$ are the coordinate differences of $H(c_i)$ and $c_j$ . With Eq. (2) and Eq. (3), the ground-truth patch-to-patch correspondences $\mathcal{M}_{gt}$ between two images are obtained. + +![](images/38530a1509f775f782fa63d9617831187a116249d5e6bcd1caa0cbddb9ae059d.jpg) +Figure 2. The pipeline of our SSL framework. $S^2 DC$ is established on patch features (i.e., token feature in vision transformer) and incorporates two main steps: (1) Patch-to-patch discrepancy. The model aligns two patch sets, which treats every patch as distinct semantics to enlarge the distribution gap between all semantics. (2) Patch-to-structure consistency. The model applies a soft regularization to align patches within the same structure. By those two steps, the framework captures semantic connectivity between anatomical structures. + +Building reliable patch-to-patch correspondences. Given token features from $\varphi^{*}(\mathcal{V}_{i}^{\prime})$ and $\varphi (\mathcal{V}_i)$ , we can calculate the similarity map $\mathcal{M}_t$ between tokens and get the loss between $\mathcal{M}_t$ and $\mathcal{M}_{gt}$ . However, similar to the issues in image matching [27, 31], reliably identifying patch correspondence in medical images is challenging due to repetitive patterns. This intractable challenge would be optimized by solving a differentiable optimal transport problem, such as the Sinkhorn algorithm [30]. The Sinkhorn algorithm is used to solve entropy-regularized optimal transport, enabling soft assignment of correspondences between feature points by iteratively refining a doubly stochastic matrix, improving robustness to noise and deformations. The efficiency of Sinkhorn is affected by the large number of iterations. + +Rocco et al. propose using a dual-softmax operator to establish reliable correspondences based on neighborhood consensus, which brings mutual nearest neighbors closer to learn a more compact similarity distribution. Given the similarity map $\mathcal{M}_t$ between two images, the similarities after applying the dual-softmax operator are: + +$$ +\hat {\mathcal {M}} _ {t} (i, j) = \operatorname {s o f t m a x} \left(\mathcal {M} _ {t} (i, \cdot)\right) \cdot \operatorname {s o f t m a x} \left(\left(\mathcal {M} _ {t} (\cdot , j)\right). \right. \tag {4} +$$ + +Based on the ground-truth correspondences $\mathcal{M}_{gt}$ and the re-assigned feature similarity matrix $\hat{\mathcal{M}}_t$ , we use the cross-entropy loss to minimize Kullback-Leibler divergence (KL divergence) between them [36]. This patch-to-patch loss $\mathcal{L}_{p2p}$ is: + +$$ +\mathcal {L} _ {p 2 p} = - \frac {1}{| \mathcal {M} _ {g t} |} \sum_ {i = 0} ^ {N} \sum_ {j = 0} ^ {N} \mathcal {M} _ {g t} (i, j) \times \log (\hat {\mathcal {M}} _ {t} (i, j)), \tag {5} +$$ + +where $N$ represents the number of patches (i.e., tokens in the 3D Transformer). + +# 3.3. Patch-to-structure consistency + +The correspondence assignment focuses on enhancing reliable patch-to-patch correspondence, which treats every patch as distinct. Patch-to-patch discrepancy limits the exploration of connectivity among patches from the same semantics (i.e., same anatomical structure). Thus, a patch-to-structure correspondence should be explored to improve consistency. + +We define neighborhood similarity distribution as $\mathcal{D}_n$ , which represents the similarity vector of an anchor patch feature $t_n$ from $\mathcal{V}_i$ with all the patch features from $\mathcal{V}_i'$ . As shown in Fig. 3, patch pairs with the same semantics form distinct peaks in the distribution $\mathcal{D}_n$ , which meets the assumption that the same structures have the same semantics. Thus, the central tendency and dispersion of $\mathcal{D}_n$ can be employed to constrain semantic connectivity. For instance, if $t_n$ is similar to several patches of $\mathcal{V}_i'$ , indicating they may belong to the same anatomical structure, $\mathcal{D}_n$ tends to be more dispersed (Fig. 3 (b)). In contrast, with no semantic-similar patches, the distribution of $t_n$ is more concentrated. Therefore, we utilize the characteristics of $\mathcal{D}_n$ as a soft label to regularize the patch-to-patch loss $\mathcal{L}_{p2p}$ to maintain underlying semantic relevance. + +We introduce a simplified version of the Sharpe ratio [28] to reflect the dispersion and concentration of neighborhood similarity distribution $\mathcal{D}_n$ . Given the $\mathcal{D}_n$ of the patch feature $t_n$ , the Sharpe ratio of $t_n$ can be computed as: + +![](images/fca96db32f5aa2436152494b128a1d03d66a25befbf68266629f88604755d534.jpg) +(a) Patches with same semantics + +![](images/842e885f379ce6edf569888115f8a5aa41fe301895d8d8bdacd848d24e8a5f36.jpg) +(b) Anchor patch with distinct semantics +Figure 3. Illustration of the similarity distribution $D_{n}$ . (a) Patches with the same semantics (e.g., bone). Given an anchor patch, patches from the same semantics form several peaks in $D_{n}$ . (b) Anchor patch with distinct semantics (e.g., pancreas), $D_{n}$ shows only one large peak (its augmented patch). + +$$ +s r _ {\mathcal {V} _ {i}} ^ {n} = \frac {\operatorname* {m a x} \left(\mathcal {D} _ {n}\right) - \frac {1}{N} \sum_ {m = 1} ^ {N} \mathcal {D} _ {n}}{\sigma_ {\mathcal {D} _ {n}}}, \tag {6} +$$ + +where, $\sigma_{\mathcal{D}_n}$ is the variance of vector $\mathcal{D}_n$ and the notation $\max (\mathcal{D}_n)$ is the maximal value of $\mathcal{D}_n$ . A low Sharpe ratio of $\mathcal{D}_n$ indicates that the distribution is more dispersed, which means the patch feature $t_i$ has more than one similar semantic patch feature. More details can be seen in Appendix. + +$sr_{\mathcal{V}_i}^n$ is utilized as the soft regularization to adaptively adjust $\mathcal{L}_{p2p}$ , serving as soft evidence for maintaining a semantic connectivity relationship among patch features. The weighted losses for the feature $t_n$ of the volume $\mathcal{V}_i$ is $\mathcal{L}_c(n,\cdot) = \text{softmax}(sr_{\mathcal{V}_i})_n\mathcal{L}_{p2p}(n,\cdot)$ , where $sr_{\mathcal{V}_i} = \{sr_{\mathcal{V}_i}^n\}_n$ . Similarly, we have $\mathcal{L}_c(\cdot,m) = \text{softmax}(sr_{\mathcal{V}_i'})_m\mathcal{L}_{p2p}(\cdot,m)$ . Thus, we have the loss of a single pair: + +$$ +l _ {n m} = \frac {\left(\operatorname {s o f t m a x} \left(s r _ {\mathcal {V} _ {i}}\right) _ {n} + \operatorname {s o f t m a x} \left(s r _ {\mathcal {V} _ {i} ^ {\prime}}\right) _ {m}\right) \mathcal {L} _ {p 2 p} (n , m)}{2}. \tag {7} +$$ + +Finally, the patch-to-structure loss $\mathcal{L}_{p2s}$ is + +$$ +\mathcal {L} _ {p 2 s} = \frac {\sum_ {m = 0} ^ {N} \sum_ {n = 0} ^ {N} l _ {n , m}}{N \times N}. \tag {8} +$$ + +# 3.4. Overall loss function + +The final combined objective for the proposed framework is defined as: + +$$ +\mathcal {L} = \mathcal {L} _ {g} + \mathcal {L} _ {p 2 p} + \mathcal {L} _ {p 2 s}, \tag {9} +$$ + +# 4. Experiments + +This section begins with an overview of the datasets utilized for both pre-training and downstream tasks. Next, we provide a brief summary of the implementation details for the + +method. Lastly, we present a comprehensive comparison of the experimental results for our proposed method against other SOTA SSL approaches in 3D medical imaging. + +# 4.1. Datasets and implementation details + +Pre-training and downstream datasets. Although we emphasize the anatomical structure-specific semantic connectivity, the training process of our mining semantic connective strategy is also suitable for positron emission tomography (PET). In our experiments, we include computed tomography (CT), magnetic resonance imaging (MRI), and PET modalities. + +We use the same dataset settings as VoCo [41] on CT, i.e., 10k data collection. BTCV and TCIA COVID-19 are used as pre-training datasets (1k data) for the BTCV segmentation task, while all other CT segmentation tasks are pre-trained on 10k data. Additionally, we collect 6605 (6k) unlabeled MRI scans from OASIS3 [18] and 3519 (3k) PET scans from UDPET [4] and ADNI [22] for pre-training. We evaluate methods on 10 datasets in 3 modalities, as shown in Tab. 2. More details about dataset splits are provided in Appendix. + +Training setup. For pre-training stage, $S^2 DC$ adopts Swin-B [11] as the default backbones in three modalities, following previous works [11, 40, 41]. We set 100k iterations for pre-training in all modalities and crop subvolumes with size $96 \times 96 \times 96$ . In our patch-to-patch loss $\mathcal{L}_{p2p}$ , we adopt the dual-softmax operator for efficiency, while the results of the Sinkhorn algorithm can be found in the Appendix. For downstream tasks, we adopt Swin-B and its derivative architectures. The weights of models are initialized by the weights of pre-training models. We adopt the Swin-UNETR [14] framework in segmentation, image-to-image translation, and reconstruction tasks. Swin-B [11] is used as the backbone of classification tasks. More details on pre-training and training settings are in the Appendix. + +Comparison methods. We compare $S^2 DC$ with SOTA + +
MethodDice Score(%)
SplRKidLKidGallEsoLivStoAorIVCVeinsPanRAGLAGAVG
From Scratch
UNETR [12]93.0294.1394.1266.9970.8796.1177.2789.2282.170.1676.6565.3259.2179.82
Swin-UNETR [14]94.0693.5493.8065.6174.6097.0975.9491.882.3673.6375.1968.0061.1180.53
With General SSL
MAE3D [7]93.9894.3794.1869.8674.6496.6680.4090.3083.1372.6577.1167.3460.5481.33
SimCLR [6]92.7993.0491.4149.6550.9998.4977.9285.5680.5864.3767.1659.0448.9973.85
SimMIM [42]95.5695.8294.1452.0653.5298.9880.2588.1182.9866.4969.1660.8850.4576.03
MoCo v3 [13]91.9692.8592.4268.2572.7794.9178.8288.2181.5971.1575.7666.4858.8179.54
Jigsaw [5]94.6293.4193.5575.6373.2195.7180.889.4184.7871.0279.5765.6860.2281.35
PositionLabel [47]94.3593.1593.2175.3973.2495.7680.6988.884.0471.1879.0265.1160.0781.09
With medical image SSL
MG [50]91.9993.5291.8165.1176.4195.9886.8889.2983.5971.7981.6267.9763.1881.45
ROT [32]91.7593.1391.6265.0976.5594.2186.1689.7483.0871.1381.5567.9063.7281.20
Vicregl [1]90.3294.1591.3065.1275.4194.7686.0089.1382.5471.2681.0167.6663.0880.89
Rubik++ [35]96.2190.4189.3375.2272.6497.4479.2589.6583.7674.7478.3567.1461.9781.38
PCRLv1 [48]95.7389.6688.5375.4172.3396.2078.9989.1183.0674.4777.8867.0261.8580.78
PCRLv2 [49]95.5091.4389.5276.1573.5497.2879.6490.1684.1775.2078.7168.7462.9381.74
Swin-UNETR [14]95.2593.1692.9763.6273.9696.2179.3289.9883.1976.1182.2568.9965.1181.54
SwinMM [40]94.3394.1894.1672.9774.7596.3783.2389.5682.9170.6575.5269.1762.9081.81
GL-MAE [52]94.5494.3994.3773.1974.9396.5183.4989.7483.1170.8075.7169.3963.1282.01
GVSL [15]95.2791.2292.2572.6973.5696.4482.4088.9084.2270.8476.4267.4863.2581.87
VoCo [41]95.7396.5394.4876.0275.6097.4178.4391.2186.1278.1980.8871.4767.8883.85
\( S^2DC \)96.2496.5194.5274.2777.6996.9583.3390.4486.6675.2284.2468.2269.7384.14
+ +Table 1. Results on the segmentation task BTCV fine-tuning from pre-training models trained on 1k CT data. Most of the results are drawn from [41, 45]. Best and second best are highlighted + +
TaskDatasetModality
SegmentationBTCV [19]CT
MSD-Liver [29]CT
MSD-Lung [29]CT
MSD-Spleen [29]CT
BraTs 21 [29]MRI
AUTOPET [10]PET
ClassificationCC-CCII [46]CT
ADNI-cls [22]PET
ReconstructionUDPET [4]PET
I2I translationBraTs 23 [29]MRI
+ +SSL methods in general [5-7, 13, 42, 47] and medical [1, 14, 15, 32, 35, 40, 41, 48-50, 52] images. Meanwhile, we report results of training from scratch (i.e., without initialized weights of pre-training models). + +# 4.2. Results on downstream tasks. + +We report the results of 4 downstream tasks on 3 modalities. Following previous work, we use the metrics Dice score for segmentation, accuracy for classification, PSNR, SSIM, and NMSE for I2I translation, and SSIM and PSNR for reconstruction. + +Segmentation tasks. Compared to volume-level and voxel-level methods, $S^2 DC$ performs well both in small targets (AUTOPET, MSD-Lung) and large organs (BTCV, + +MSD-Liver, MSD-Spleen and BraTs 21) segmentation, as shown in Tab. 1 and Tab. 3. In BTCV datasets, compared to general SSL methods, medical SSL methods show superior performance. Notably, with $S^2DC$ pre-training, we achieve an average improvement of $4.32\%$ , reaching a Dice score of $75.47\%$ , which surpasses other SOTA medical SSL methods by $0.81\%$ . + +Classification tasks. We conduct experiments on the CC-CCII dataset for COVID-19 classification and ADNI for AD (Alzheimer's Disease), MCI (Mild Cognitive Impairment), and CN (Cognitively Normal) classification. With the pre-trained model, $S^2 DC$ achieves an average improvement of $3.93\%$ and outperforms other medical SSL methods. + +Table 2. The downstream tasks and modalities. I2I translation represents image-to-image translation. + +
MethodNetworkCC-CCIIADNI-clsAVG
Acc(%)Acc(%)Acc(%)
From Scratch
Swin-UNETR [14]Swin-B88.0456.1772.10
With 3D medical image SSL
PCRLv2 [49]Swin-B93.0756.0574.56
Swin-UNETR [14]Swin-B94.1555.9575.05
SwinMM [40]Swin-B94.8056.3675.58
VoCo [41]Swin-B94.6056.3275.46
\( S^2DC \)Swin-B95.3456.7276.03
+ +Table 4. Experimental results on the classification task CC-CCII and ADNI-cls fine-tuning from pre-training models trained on 6k MRI scans. Best and second best are highlighted. + +
MethodDice(%)
MSD LiverMSD LungMSD SpleenAUTOPETBraTs21Total
WTTCETAVGAVG
From scratch
Swin-UNETR [14]81.1859.7394.2034.4992.8988.2877.3486.1771.15
With 3D medical image SSL
Swin-UNETR [14]81.5862.4694.4943.8592.7088.4278.8286.6473.95
Swin-MM [40]82.3261.1694.6044.3792.8988.5479.1886.8773.86
PCRLv2 [49]82.1262.3995.2342.6492.8188.4978.7486.6873.81
VoCo [41]82.5063.3495.7045.1092.9088.4478.6286.6574.66
\( S^2DC \)83.4364.4095.7346.4792.7688.7179.7987.0975.47
+ +Table 3. Results on the segmentation tasks MSD-Liver, MSD-Lung, MSD-Spleen, AUTOPET, and BraTs21. WT, TC, and ET represent the whole tumor, tumor core, and enhancing tumor, respectively. We retrained all methods using $10\mathrm{k}$ , $6\mathrm{k}$ , and $3\mathrm{k}$ data in three modalities separately, except for VoCo in CT, whose weight is provided in the official code. Best and second best are highlighted. + +
MethodT1->T2T1->T1ce
PSNR↑NMSE↓SSIM↑PSNR↑NMSE↓SSIM↑
From Scratch
Swin-UNETR [14]26.570.13480.880726.910.0820.8871
With 3D medical image SSL
PCRLv2 [49]26.790.16400.882829.620.1230.9179
Swin-UNETR [14]24.810.29380.844130.420.1080.9117
SwinMM [40]24.050.31580.818230.700.0990.9059
VoCo [41]25.810.19160.859831.920.0790.9315
S2DC25.960.11610.883932.650.0410.9412
+ +Table 5. Experimental results on the image-to-image translation task BraTs23 fine-tuning from pre-training models trained on 6k MRI scans. Best and second best are highlighted. + +
MethodNetworkSSIM↑PSNR↑
From Scratch
Swin-UNETR [14]-0.9991758.4100
With 3D medical image SSL
PCRLv2 [49]Swin-UNETR0.9991658.2865
Swin-UNETR [14]Swin-UNETR0.9991758.3409
SwinMM [40]Swin-UNETR0.9991758.4002
VoCo [41]Swin-UNETR0.9991658.2561
\( S^2DC \)Swin-UNETR0.9991758.4649
+ +Image-to-image translation task. We evaluate in BraTs23, which includes four structural MRI modalities for each individual. The translation quality for $\mathrm{T1} \rightarrow \mathrm{T2}$ and $\mathrm{T1} \rightarrow \mathrm{T1ce}$ are compared by peak signal-to-noise ratio (PSNR), normalized mean squared error (NMSE), and structural similarity index (SSIM) [44]. From Table 5, our method excels in two tasks. Remarkably, with $S^2 DC$ pretraining, we gain $5.41\%$ and $0.32\%$ improvements in SSIM separately. We notice that other methods fail to get improvement in the task $\mathrm{T1} \rightarrow \mathrm{T2}$ compared to From Scratch. That may be due to the deep domain gap between the two sequences, while pre-trained models provide biased initializations. In contrast, $S^2 DC$ achieves improvement in NMSE and SSIM, showing better ability in this task. + +Reconstruction task. We evaluate the capabilities of reconstruction in different methods on the UDPET dataset. Following the previous work [8], whole-body data at + +DFR=50 (i.e., 1/50 doses) were selected to explore the reconstruction ability of all methods. This task has been extensively studied, yet our method, $S^2 DC$ , still achieves superior PSNR results. + +Overall comparisons on 10 downstream datasets. The results demonstrate that $S^2 DC$ performs effectively across 10 datasets and 4 tasks. Compared to training from scratch, which achieves an average score of $77.93\%$ , $S^2 DC$ pre-training delivers a $3.5\%$ improvement, reaching $81.43\%$ . Additionally, $S^2 DC$ outperforms the second-best SSL method (VoCo [41], average score $80.65\%$ ) for all tasks, with an average gain of $0.78\%$ . These findings indicate that $S^2 DC$ is both reliable and adaptable across various downstream tasks and modalities. + +# 4.3. Ablation study + +Influence of different constraints. We use contrastive learning loss $\mathcal{L}_g$ in Sec. 3.1 as our baseline. The impact of different constraints in pre-training is presented in Tab. 7. The results indicate that combining patch-to-patch loss $\mathcal{L}_{p2p}$ and patch-to-structure loss $\mathcal{L}_{p2s}$ yields SOTA performance compared to Tab. 1. Meanwhile, we visualize the t-SNE clustering results of 13 organs under different constraints, as shown in Fig. 4. The $\mathcal{L}_{p2p}$ constraint leads to a sparser inter-structure distribution, while the $\mathcal{L}_{p2s}$ constraint results in more compact intra-structure clusters. Combining $\mathcal{L}_{p2p}$ and $\mathcal{L}_{p2s}$ achieves the best overall clustering performance. + +Table 6. Experimental results on low-dose to full-dose reconstruction task UDPET fine-tuning from pre-training models trained on 3k PET scans. Best and second best are highlighted. + +
Baseline(Lg)+Lp2p+Lp2sBTCV DICE(%)AUTOPET
•83.4345.72
••83.9845.85
••84.0246.23
•••84.1446.47
+ +Table 7. The ablation results of different constraints. + +Intra-structure consistency and inter-structure discrepancy. We visualize the t-SNE of features extracted by four medical pre-trained models and calculate their silhouette + +![](images/d47ced4e1db314b13b8117c4ba47efcc95fbbbf9be78e5d7940ff15fac26a183.jpg) + +![](images/9b9e96e73c6be48eacaaa298d528f1cf9b6a441120bc5b51c0229f3fe6742841.jpg) +(a) $\mathcal{L}_g$ +(c) $\mathcal{L}_g + \mathcal{L}_{p2s}$ +(b) $\mathcal{L}_g + \mathcal{L}_{p2p}$ + +![](images/322cb4db0aa9699f6d9708ed844f699e2b2bfd7f464a57678823e64e12245dd7.jpg) +(d) $\mathcal{L}_g + \mathcal{L}_{p2p} + \mathcal{L}_{p2s}$ + +![](images/18882e36428500c73bdc3f359f9b9625fd632b24afcc7e67b34ec9a0f0cfe3d0.jpg) +Figure 4. The t-SNE feature visualization of different losses on 13 organs on the BTCV dataset. +(a)Ours +(b) SwinMM + +![](images/643730787308b356cf81c4b51a96985b5aaf4bc4f7aebf049fa084ea18bb77d4.jpg) +Silhouette Score: 0.0381 +(c) VoCo +Silhouette Score: -0.1749 +Figure 5. The t-SNE visualization of features from 13 organs on the BTCV dataset and the silhouette scores of four methods. + +![](images/25e70f9bea063311b9f06ba2ea63f127e09dd2100a31b3bf58c155fba9ac99d7.jpg) + +![](images/ce1d4e767b692bf86b41e6f3ea9cbd06495565090b217b871d11eed313101271.jpg) +Silhouette Score:0.0027 +(d) Swin-UNETR +Silhouette Score: -0.0477 + +scores [25], which evaluate the quality of clustering by measuring how similar a data point is to its cluster (consistency) compared to other clusters (discrepancy). Fig. 5 shows the clustering capabilities of 13 organs on the BTCV dataset. $S^2 DC$ shows excellent performance on organ discrimination and consistency, which gets the highest silhouette scores. As shown in Tab. 1, most organs that cluster well also exhibit strong segmentation performance, highlighting the importance of structure-specific consistency and discrepancy. Additionally, we visualize the feature maps of 4 SSL models on 3 modalities in Fig. 6. $S^2 DC$ performs semantic consistency in the background, tissues (CT), functional regions (MRI), and metabolic regions (PET). + +![](images/bcec55db2721be3b180e82fe43727cf66bde4f036859a4383e6ed642bde83ced.jpg) +Figure 6. We visualize the first principal components after applying PCA to token features [2] within volumes of CT, MRI, and PET. The figure displays the central slices of feature maps. + +
MethodAccuracy(%)
10%50%100%
Swin-UNETR [14]77.1592.2894.15
SwinMM [40]87.8793.5094.80
VoCo [41]86.7392.4394.60
\( S^2DC \)88.2993.8995.34
+ +Table 8. Experiment results on CC-CCII with various ratios of the training data. ${10}\% ,{50}\%$ ,and ${100}\%$ represent ratios. + +Varying ratios of the training dataset. We evaluate the performances of models that were trained with varying ratios of CC-CCII. From Tab. 8, $S^2 DC$ outperforms other methods under different ratios of training data. Notably, $S^2 DC$ achieves $88.29\%$ accuracy using only $10\%$ of the data, which is effective with limited dataset sizes. + +![](images/6d0e7126fed46aaa38e3acf9dc6f10c532bfee21d41a9c3e30a79492738e0856.jpg) +Figure 7. The evolution of patch-to-structure correspondences. + +Patch-to-structure correspondences evolution. We visualize feature map trends at different pretraining steps in Fig. 7. As training progresses, feature consistency improves, especially for small structures like bronchioles. + +# 5. Conclusion + +We propose $S^2 DC$ , a medical image SSL framework that improves structure-aware semantic consistency through patch-to-patch and patch-to-structure steps. Experiments and visualizations show its superior performance. Future work could involve larger pre-training datasets or model generalization across modalities for broader applications. + +# 6. Acknowledgments + +This paper was supported by the National Natural Science Foundation of China (Grant No.82394432 and 92249302), the Shanghai Municipal Science and Technology Major Project (Grant No.2023SHZDZX02). The computations in this research were performed using the CFFF platform of Fudan University. + +# References + +[1] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicregl: Self-supervised learning of local visual features. Advances in Neural Information Processing Systems, 35:8799-8810, 2022. 6 +[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 8 +[3] Anadi Chaman and Ivan Dokmanic. Truly shift-invariant convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3773-3783, 2021. 3 +[4] Gaoyu Chen, Sheng Liu, Wenxiang Ding, Li Lv, Chen Zhao, Fenghua Weng, Yong Long, Yunlong Zan, and Qiu Huang. A total-body ultra-low dose pet reconstruction method via image space shuffle u-net and body sampling. IEEE Transactions on Radiation and Plasma Medical Sciences, 2023. 1, 5, 6 +[5] Pengguang Chen, Shu Liu, and Jiaya Jia. Jigsaw clustering for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11526-11535, 2021. 6 +[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 6 +[7] Zekai Chen, Devansh Agarwal, Kshitij Aggarwal, Wiem Safta, Mariann Micsinai Balan, and Kevin Brown. Masked image modeling advances 3d medical image analysis. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1970-1980, 2023. 6 +[8] Jiaqi Cui, Pinxian Zeng, Xinyi Zeng, Yuanyuan Xu, Peng Wang, Jiliu Zhou, Yan Wang, and Dinggang Shen. Prior knowledge-guided triple-domain transformer-gan for direct pet reconstruction from low-count sinograms. IEEE Transactions on Medical Imaging, 2024. 7 +[9] Tugba Akinci D'Antonoli, Lucas K Berger, Ashraya K Indrakanti, Nathan Vishwanathan, Jakob Weiβ, Matthias Jung, Zeynep Berkarda, Alexander Rau, Marco Reisert, Thomas Küstner, et al. Totalsegmentator mri: Sequence-independent segmentation of 59 anatomical structures in mr images. arXiv preprint arXiv:2405.19492, 2024. 1 +[10] Sergios Gatidis, Marcel Früh, Matthias Fabritius, Sijing Gu, Konstantin Nikolaou, Christian La Fougère, Jin Ye, Junjun He, Yige Peng, Lei Bi, et al. The autopet challenge: + +towards fully automated lesion segmentation in oncologic pet/ct imaging. 2023. 1, 6 +[11] Ali Hatamizadeh, Vishwesh Nath, Yucheng Tang, Dong Yang, Holger R Roth, and Daguang Xu. Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images. In International MICCAI brainlesion workshop, pages 272-284. Springer, 2021. 5 +[12] Ali Hatamizadeh, Yucheng Tang, Vishwesh Nath, Dong Yang, Andriy Myronenko, Bennett Landman, Holger R Roth, and Daguang Xu. Unetr: Transformers for 3d medical image segmentation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 574-584, 2022. 6 +[13] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 3, 6 +[14] Yufan He, Vishwesh Nath, Dong Yang, Yucheng Tang, Andriy Myronenko, and Daguang Xu. Swinunetr-v2: Stronger swim transformers with stagewise convolutions for 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 416-426. Springer, 2023. 1, 3, 5, 6, 7, 8 +[15] Yuting He, Guanyu Yang, Rongjun Ge, Yang Chen, Jean-Louis Coatieux, Boyu Wang, and Shuo Li. Geometric visual similarity learning in 3d medical image self-supervised pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9538-9547, 2023. 2, 6 +[16] Junjia Huang, Haofeng Li, Guanbin Li, and Xiang Wan. Attentive symmetric autoencoder for brain mri segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 203-213. Springer, 2022. 2 +[17] Yankai Jiang, Mingze Sun, Heng Guo, Xiaoyu Bai, Ke Yan, Le Lu, and Minfeng Xu. Anatomical invariance modeling and semantic alignment for self-supervised learning in 3d medical image analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15859-15869, 2023. 1, 2 +[18] Pamela J LaMontagne, Tammie LS Benzinger, John C Morris, Sarah Keefe, Russ Hornbeck, Chengjie Xiong, Elizabeth Grant, Jason Hassenstab, Krista Moulder, Andrei G Vlassenko, et al. Oasis-3: longitudinal neuroimaging, clinical, and cognitive dataset for normal aging and Alzheimer disease. medrxiv, pages 2019–12, 2019. 5 +[19] Bennett Landman, Zhoubing Xu, J Igelsias, Martin Styner, Thomas Langerak, and Arno Klein. Miccai multi-atlas labeling beyond the cranial vault-workshop and challenge. In Proc. MICCAI Multi-Atlas Labeling Beyond Cranial Vault—Workshop Challenge, page 12, 2015. 6 +[20] Haofeng Li, Yiming Ouyang, and Xiang Wan. Self-supervised alignment learning for medical image segmentation. In 2024 IEEE International Symposium on Biomedical Imaging (ISBI), pages 1-5. IEEE, 2024. 2 +[21] Chengzhi Mao, Lingyu Zhang, Abhishek Vaibhav Joshi, Junfeng Yang, Hao Wang, and Carl Vondrick. Robust perception + +through equivariance. In International Conference on Machine Learning, pages 23852-23870. PMLR, 2023. 3 +[22] Susanne G Mueller, Michael W Weiner, Leon J Thal, Ronald C Petersen, Clifford Jack, William Jagust, John Q Trojanowski, Arthur W Toga, and Laurel Beckett. The alzheimer's disease neuroimaging initiative. Neuroimaging Clinics of North America, 15(4):869, 2005. 5, 6 +[23] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 1, 3 +[24] Ignacio Rocco, Mircea Cimpoi, Relja Arandjelović, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Neighbourhood consensus networks. Advances in neural information processing systems, 31, 2018. 4 +[25] Peter J Rousseeuw. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. Journal of computational and applied mathematics, 20:53-65, 1987. 8 +[26] Ainkaran Santhirasekaram, Mathias Winkler, Andrea Rockall, and Ben Glocker. A geometric approach to robust medical image segmentation. Medical Image Analysis, 97: 103260, 2024. 2 +[27] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938-4947, 2020. 4 +[28] William F Sharpe. Mutual fund performance. The Journal of business, 39(1):119-138, 1966. 4 +[29] Amber L Simpson, Michela Antonelli, Spyridon Bakas, Michel Bilello, Keyvan Farahani, Bram Van Ginneken, Annette Kopp-Schneider, Bennett A Landman, Geert Litjens, Bjoern Menze, et al. A large annotated medical image dataset for the development and evaluation of segmentation algorithms. arXiv preprint arXiv:1902.09063, 2019. 1, 6 +[30] Richard Sinkhorn. A relationship between arbitrary positive matrices and doubly stochastic matrices. The annals of mathematical statistics, 35(2):876-879, 1964. 4 +[31] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 4 +[32] Aiham Taleb, Winfried Loetzsch, Noel Danz, Julius Severin, Thomas Gaertner, Benjamin Bergner, and Christoph Lippert. 3d self-supervised methods for medical imaging. Advances in neural information processing systems, 33:18158-18172, 2020. 6 +[33] Zhaorui Tan, Xi Yang, Tan Pan, Tianyi Liu, Chen Jiang, Xin Guo, Qiufeng Wang, Anh Nguyen, Yuan Qi, Kaizhu Huang, et al. Towards a universal 3d medical multi-modality generalization via learning personalized invariant representation. arXiv preprint arXiv:2411.06106, 2024. 1 +[34] Yucheng Tang, Dong Yang, Wenqi Li, Holger R Roth, Bennett Landman, Daguang Xu, Vishwesh Nath, and Ali Hatamizadeh. Self-supervised pre-training of swim transformers for 3d medical image analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20730-20740, 2022. 2 + +[35] Xing Tao, Yuexiang Li, Wenhui Zhou, Kai Ma, and Yefeng Zheng. Revisiting rubik's cube: Self-supervised learning with volume-wise transformation for 3d medical image segmentation. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part IV 23, pages 238-248. Springer, 2020. 2, 6 +[36] MTCAJ Thomas and A Thomas Joy. Elements of information theory. Wiley-Interscience, 2006. 4 +[37] Yao-Hung Hubert Tsai, Yue Wu, Ruslan Salakhutdinov, and Louis-Philippe Morency. Self-supervised learning from a multi-view perspective. arXiv preprint arXiv:2006.05576, 2020. 3 +[38] Chengyao Wang, Li Jiang, Xiaoyang Wu, Zhuotao Tian, Bohao Peng, Hengshuang Zhao, and Jiaya Jia. Groupcontrast: Semantic-aware self-supervised representation learning for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4917-4928, 2024. 3 +[39] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3024-3033, 2021. 3 +[40] Yiqing Wang, Zihan Li, Jieru Mei, Zihao Wei, Li Liu, Chen Wang, Shengtian Sang, Alan L Yuille, Cihang Xie, and Yuyin Zhou. Swinmm: masked multi-view with swin transformers for 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 486-496. Springer, 2023. 1, 2, 3, 5, 6, 7, 8 +[41] Linshan Wu, Jiaxin Zhuang, and Hao Chen. Voco: A simple-yet-effective volume contrastive learning framework for 3d medical image analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22873-22882, 2024. 1, 2, 3, 5, 6, 7, 8 +[42] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9653-9663, 2022. 6 +[43] Zhangsihao Yang, Mengwei Ren, Kaize Ding, Guido Gerig, and Yalin Wang. Keypoint-augmented self-supervised learning for medical image segmentation with limited annotation. Advances in Neural Information Processing Systems, 36, 2024. 2 +[44] Xin Yi, Ekta Walia, and Paul Babyn. Generative adversarial network in medical imaging: A review. Medical image analysis, 58:101552, 2019. 7 +[45] Chuyan Zhang, Hao Zheng, and Yun Gu. Dive into the details of self-supervised learning for medical image analysis. Medical Image Analysis, 89:102879, 2023. 6 +[46] Kang Zhang, Xiaohong Liu, Jun Shen, Zhihuan Li, Ye Sang, Xingwang Wu, Yunfei Zha, Wenhua Liang, Chengdi Wang, Ke Wang, et al. Clinically applicable ai system for accurate diagnosis, quantitative measurements, and prognosis of Covid-19 pneumonia using computed tomography. Cell, 181 (6):1423-1433, 2020. 6 + +[47] Zhemin Zhang and Xun Gong. Positional label for self-supervised vision transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3516-3524, 2023. 6 +[48] Hong-Yu Zhou, Chixiang Lu, Sibei Yang, Xiaoguang Han, and Yizhou Yu. *Preservational learning improves self-supervised medical image models by reconstructing diverse contexts*. In *Proceedings of the IEEE/CVF International Conference on Computer Vision*, pages 3499–3509, 2021. 6 +[49] Hong-Yu Zhou, Chixiang Lu, Chaoqi Chen, Sibei Yang, and Yizhou Yu. Pcrlv2: A unified visual information preservation framework for self-supervised pre-training in medical image analysis. arXiv preprint arXiv:2301.00772, 2023. 2, 3, 6, 7 +[50] Zongwei Zhou, Vatsal Sodha, Jiaxuan Pang, Michael B Gotway, and Jianming Liang. Models genesis. Medical image analysis, 67:101840, 2021. 6 +[51] Jiaxin Zhuang, Linshan Wu, Qiong Wang, Varut Vardhanabhuti, Lin Luo, and Hao Chen. Mim: Mask in mask self-supervised pre-training for 3d medical image analysis. arXiv preprint arXiv:2404.15580, 2024. 2 +[52] Jia-Xin Zhuang, Luyang Luo, and Hao Chen. Advancing volumetric medical image segmentation via global-local masked autoencoder. arXiv preprint arXiv:2306.08913, 2023. 2, 6 \ No newline at end of file diff --git a/2025/Structure-aware Semantic Discrepancy and Consistency for 3D Medical Image Self-supervised Learning/images.zip b/2025/Structure-aware Semantic Discrepancy and Consistency for 3D Medical Image Self-supervised Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f51a3c42fcfebefc9ad1629994dfd3cd00b1d5f0 --- /dev/null +++ b/2025/Structure-aware Semantic Discrepancy and Consistency for 3D Medical Image Self-supervised Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:644fb5107752973ba75f7798236280542b12541e142e8906280882d7801bc887 +size 888348 diff --git a/2025/Structure-aware Semantic Discrepancy and Consistency for 3D Medical Image Self-supervised Learning/layout.json b/2025/Structure-aware Semantic Discrepancy and Consistency for 3D Medical Image Self-supervised Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f2e53cee88c1d0077cfc84ab7e24cbd488f79df9 --- /dev/null +++ b/2025/Structure-aware Semantic Discrepancy and Consistency for 3D Medical Image Self-supervised Learning/layout.json @@ -0,0 +1,10324 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 63, + 102, + 547, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 102, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 63, + 102, + 547, + 140 + ], + "type": "text", + "content": "Structure-aware Semantic Discrepancy and Consistency for 3D Medical Image Self-supervised Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 149, + 160, + 459, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 160, + 459, + 189 + ], + "spans": [ + { + "bbox": [ + 149, + 160, + 459, + 189 + ], + "type": "text", + "content": "Tan Pan" + }, + { + "bbox": [ + 149, + 160, + 459, + 189 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 149, + 160, + 459, + 189 + ], + "type": "text", + "content": ", Zhaorui Tan" + }, + { + "bbox": [ + 149, + 160, + 459, + 189 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 149, + 160, + 459, + 189 + ], + "type": "text", + "content": ", Kaiyu Guo" + }, + { + "bbox": [ + 149, + 160, + 459, + 189 + ], + "type": "inline_equation", + "content": "^{3,2}" + }, + { + "bbox": [ + 149, + 160, + 459, + 189 + ], + "type": "text", + "content": ", Dongli Xu" + }, + { + "bbox": [ + 149, + 160, + 459, + 189 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 149, + 160, + 459, + 189 + ], + "type": "text", + "content": ", Weidi Xu, Chen Jiang" + }, + { + "bbox": [ + 149, + 160, + 459, + 189 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 149, + 160, + 459, + 189 + ], + "type": "text", + "content": ", Xin Guo" + }, + { + "bbox": [ + 149, + 160, + 459, + 189 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 149, + 160, + 459, + 189 + ], + "type": "text", + "content": ", Yuan Qi" + }, + { + "bbox": [ + 149, + 160, + 459, + 189 + ], + "type": "inline_equation", + "content": "^{1,2,4}" + }, + { + "bbox": [ + 149, + 160, + 459, + 189 + ], + "type": "text", + "content": ", Yuan Cheng" + }, + { + "bbox": [ + 149, + 160, + 459, + 189 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 108, + 189, + 502, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 189, + 502, + 218 + ], + "spans": [ + { + "bbox": [ + 108, + 189, + 502, + 218 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 189, + 502, + 218 + ], + "type": "text", + "content": " AI" + }, + { + "bbox": [ + 108, + 189, + 502, + 218 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 108, + 189, + 502, + 218 + ], + "type": "text", + "content": ", Fudan University " + }, + { + "bbox": [ + 108, + 189, + 502, + 218 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 108, + 189, + 502, + 218 + ], + "type": "text", + "content": " Shanghai Academy of Artificial Intelligence for Science " + }, + { + "bbox": [ + 108, + 189, + 502, + 218 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 108, + 189, + 502, + 218 + ], + "type": "text", + "content": " The University of Queensland " + }, + { + "bbox": [ + 108, + 189, + 502, + 218 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 108, + 189, + 502, + 218 + ], + "type": "text", + "content": " Zhongshan Hospital, Fudan University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 220, + 488, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 220, + 488, + 231 + ], + "spans": [ + { + "bbox": [ + 116, + 220, + 488, + 231 + ], + "type": "text", + "content": "pant23@m.fudan.edu.cn, jiangchen@sais.com.cn, cheng-yuan@fudan.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 283, + 296, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 283, + 296, + 595 + ], + "spans": [ + { + "bbox": [ + 54, + 283, + 296, + 595 + ], + "type": "text", + "content": "3D medical image self-supervised learning (mSSL) holds great promise for medical analysis. Effectively supporting broader applications requires considering anatomical structure variations in location, scale, and morphology, which are crucial for capturing meaningful distinctions. However, previous mSSL methods partition images with fixed-size patches, often ignoring the structure variations. In this work, we introduce a novel perspective on 3D medical images with the goal of learning structure-aware representations. We assume that patches within the same structure share the same semantics (semantic consistency) while those from different structures exhibit distinct semantics (semantic discrepancy). Based on this assumption, we propose an mSSL framework named " + }, + { + "bbox": [ + 54, + 283, + 296, + 595 + ], + "type": "inline_equation", + "content": "S^2DC" + }, + { + "bbox": [ + 54, + 283, + 296, + 595 + ], + "type": "text", + "content": ", achieving Structure-aware Semantic Discrepancy and Consistency in two steps. First, " + }, + { + "bbox": [ + 54, + 283, + 296, + 595 + ], + "type": "inline_equation", + "content": "S^2DC" + }, + { + "bbox": [ + 54, + 283, + 296, + 595 + ], + "type": "text", + "content": " enforces distinct representations for different patches to increase semantic discrepancy by leveraging an optimal transport strategy. Second, " + }, + { + "bbox": [ + 54, + 283, + 296, + 595 + ], + "type": "inline_equation", + "content": "S^2DC" + }, + { + "bbox": [ + 54, + 283, + 296, + 595 + ], + "type": "text", + "content": " advances semantic consistency at the structural level based on neighborhood similarity distribution. By bridging patch-level and structure-level representations, " + }, + { + "bbox": [ + 54, + 283, + 296, + 595 + ], + "type": "inline_equation", + "content": "S^2DC" + }, + { + "bbox": [ + 54, + 283, + 296, + 595 + ], + "type": "text", + "content": " achieves structure-aware representations. Thoroughly evaluated across 10 datasets, 4 tasks, and 3 modalities, our proposed method consistently outperforms the state-of-the-art methods in mSSL. The code is available at https://github.com/Ashespt/S2DC/." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 617, + 135, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 135, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 135, + 629 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 637, + 296, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 637, + 296, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 637, + 296, + 686 + ], + "type": "text", + "content": "Self-supervised learning (SSL) is widely used in medical image analysis to build strong representations that improve performance on various applications and maximize the use of unlabeled data [23, 33, 40, 41]. These applica" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 258, + 362, + 302 + ], + "blocks": [ + { + "bbox": [ + 316, + 258, + 362, + 302 + ], + "lines": [ + { + "bbox": [ + 316, + 258, + 362, + 302 + ], + "spans": [ + { + "bbox": [ + 316, + 258, + 362, + 302 + ], + "type": "image", + "image_path": "98d082471e3a49804679308836c0ec15e106963488061adb787739266abea581.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 303, + 362, + 352 + ], + "blocks": [ + { + "bbox": [ + 316, + 303, + 362, + 352 + ], + "lines": [ + { + "bbox": [ + 316, + 303, + 362, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 303, + 362, + 352 + ], + "type": "image", + "image_path": "007fe4fd6055ab11529d337e318f0dc8ea566271ed8892065b14269e3385fd75.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 319, + 357, + 363, + 363 + ], + "lines": [ + { + "bbox": [ + 319, + 357, + 363, + 363 + ], + "spans": [ + { + "bbox": [ + 319, + 357, + 363, + 363 + ], + "type": "text", + "content": "A. Two views of mSSL" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 369, + 258, + 549, + 365 + ], + "blocks": [ + { + "bbox": [ + 369, + 258, + 549, + 365 + ], + "lines": [ + { + "bbox": [ + 369, + 258, + 549, + 365 + ], + "spans": [ + { + "bbox": [ + 369, + 258, + 549, + 365 + ], + "type": "image", + "image_path": "9d3b6ed803d593e8a5e4236c766658744ae3109dc48675fb32d4ba87d627815f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 374, + 555, + 495 + ], + "lines": [ + { + "bbox": [ + 313, + 374, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 555, + 495 + ], + "type": "text", + "content": "Figure 1. A. Two views of mSSL. B. Similarity heatmaps on CT and MRI images across different methods. We sample an anchor patch (red dots) and compute feature similarities with all other patches. In the first row, the liver anchor should show low similarity with non-liver patches, while in the second row, the white matter anchor should exhibit high similarity with other white matter patches. Current SOTA contrastive-based [41] (b) and reconstruction-based [40] (c) methods struggle with both patch feature discrepancy in different structures and consistency in the same structure. In contrast, our method (d) advances both discrepancy and consistency." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 520, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 520, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 312, + 520, + 555, + 616 + ], + "type": "text", + "content": "tions frequently focus on different specific anatomical structures (i.e., organ segmentation), each distinguished across anatomy due to location, scale, morphology, and other variances [4, 9, 10, 29]. For example, the liver is a large, well-defined solid organ with relatively homogeneous intensity and distinct anatomical boundaries. In contrast, veins are thin, tubular structures with varying diameters, lower contrast, and complex branching patterns." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 617, + 556, + 714 + ], + "type": "text", + "content": "Current medical image SSL (mSSL) methods can be classified into contrastive-based and reconstruction-based methods [14, 17, 40, 41]. For example, Wang et al. utilizes multi-view reconstruction to encourage dense representations, while Wu et al. incorporates overlapped patch ratio as soft contrastive constraint to improve semantic consistency. However, these methods use fixed-size image partition, treating all segmented patches uniformly across struc" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "* Corresponding authors. This research was conducted during an internship at the Shanghai Academy of Artificial Intelligence for Science." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20257" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "tures and overlooking their underlying semantic relationship among structures, as shown in Fig. 1 A. To illustrate these limitations, we visualize similarity heatmaps on CT and MRI images across different methods by overlaying similarity scores onto the original images (Fig. 1 B). Ideally, patches within the same organ should exhibit high similarity, while those from different organs should show low similarity. However, as shown in Fig. 1 (a) and (b), existing methods struggle with maintaining intra-structure patch consistency and distinguishing inter-structure patch differences. A more detailed analysis and metrics of organ-level relationships in the state-of-the-art (SOTA) methods are provided in Fig. 5." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 229, + 294, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 229, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 294, + 326 + ], + "type": "text", + "content": "Based on these empirical analyses, we propose that patches within the same structure should share the same semantics, while those from different structures should exhibit distinct semantics. To achieve this, patch relationships must be structure-aware and adaptively adjusted to account for anatomical variations. Based on this assumption, we define structure-aware learning through two key principles: intra-structure consistency and inter-structure discrepancy." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 326, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 295, + 567 + ], + "type": "text", + "content": "To achieve these principles, we propose " + }, + { + "bbox": [ + 55, + 326, + 295, + 567 + ], + "type": "inline_equation", + "content": "S^2 DC" + }, + { + "bbox": [ + 55, + 326, + 295, + 567 + ], + "type": "text", + "content": ", a self-supervised learning framework for 3D medical imaging designed to capture Structure-aware Semantic Consistency and Discrepancy. The core idea of our framework is to explore semantic relevance between patches to enhance structure-level representations, which unfolds in two key steps. First, we increase the discrepancy between patches with different semantics through patch-to-patch discrepancy. By an optimal transport strategy, patches with different semantics are pushed further apart. Second, beyond patch-wise distinction, the framework further explores patch-to-structure semantic consistency, leveraging the observation that patches with shared semantics form peaks in the similarity distribution. These distributional characteristics serve as soft regularization, identifying patches belonging to the same structure and enforcing structural-level consistency. Notably, the principles of inter-structure discrepancy and intra-structure consistency of our framework are also suitable for functional imaging or metabolic imaging." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 568, + 285, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 568, + 285, + 579 + ], + "spans": [ + { + "bbox": [ + 67, + 568, + 285, + 579 + ], + "type": "text", + "content": "This work's contributions are summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 582, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 664 + ], + "type": "text", + "content": "- We propose a novel insight of intra-structure consistency and inter-structure discrepancy in the anatomical structure-aware feature learning in 3D medical images. We argue that the patches within the same anatomical structures maintain consistent semantics, while the patches from different anatomical structures reflect distinct semantics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 666, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 295, + 713 + ], + "type": "text", + "content": "- Based on our insight and assumption, we propose " + }, + { + "bbox": [ + 55, + 666, + 295, + 713 + ], + "type": "inline_equation", + "content": "S^2 DC" + }, + { + "bbox": [ + 55, + 666, + 295, + 713 + ], + "type": "text", + "content": ", a novel training framework that enhances interstructure discrepancy and intra-structure consistency. The framework establishes reliable patch-to-patch correspond" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 108 + ], + "type": "text", + "content": "dences to reinforce discrepancy while leveraging patch-to-structure semantic connectivity from the similarity distribution to improve consistency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 108, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 108, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 314, + 108, + 553, + 156 + ], + "type": "text", + "content": "- Our method demonstrates superior performance over SOTA medical image SSL methods, evaluated across 10 datasets, 4 tasks, and 3 imaging modalities. The code will be available upon publication." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 167, + 398, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 167, + 398, + 179 + ], + "spans": [ + { + "bbox": [ + 314, + 167, + 398, + 179 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 186, + 553, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 553, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 553, + 390 + ], + "type": "text", + "content": "SSL in 3D medical images. Contrastive-based and reconstruction-based SSL are among the most prominent paradigms in mSSL. Some approaches emphasize the fine-grained capabilities of models, incorporating techniques such as multi-scale reconstruction or reconstruction from disarranged images to enhance model performance [35, 51, 52]. Simultaneously, other methods [41, 49] prioritize preserving maximal semantic information by leveraging contextual priors. Recently, several studies have explored the combination of contrastive learning (CL) and reconstruction to further improve model performance [15, 17, 34, 40]. These works aim to learn consistent semantic representations by applying contrastive learning to global features or predicting missing information from available context. However, in mSSL, few approaches enforce constraints to enhance both anatomical structure-aware discrepancy and consistency, which is the central focus of this paper." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 391, + 553, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 553, + 653 + ], + "type": "text", + "content": "Geometric approaches for medical image analysis. Our method adopts some geometric characteristics for patch-to-patch correspondence. In radiographic imaging, geometric characteristics, such as location and anatomical structure, play a crucial role in organ identification, disease diagnosis, and medical analysis. Several approaches [16, 26] have incorporated consistent and recurrent anatomical patterns as learnable priors to enhance medical image analysis. For example, Huang et al. proposed to enhance segmentation by establishing patch correspondence between two 2D images. Those methods utilizing the geometric or anatomical invariant have shown remarkable improvement in the downstream tasks. Some medical SSL works [15, 20] also explore geometric correspondence for better performance. The SSL method GVSL [15] considers the geometric transformation and reconstructs affine images to original images to learn global and local representations. Yang et al. present a keypoint-augmented fusion layer with the keypoint guidance from SIFT to extract local and global information to improve segmentation results. Those methods leverage the invariance of anatomical structures while not considering the relation between anatomical structures." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 654, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 712 + ], + "type": "text", + "content": "Multi-scale features in SSL. Multi-scale features have received much attention in existing SSL research, which is relevant but not sufficient to structure-aware representations. Multi-scale feature learning in SSL can be categorized as reconstruction-based and contrastive-based meth" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20258" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": "ods. For instance, the contrastive-based method Wang et al. introduced a dense constraint based on the correspondence of local features for dense prediction tasks. The method [38] developed data-driven contrastive learning methods by utilizing prototypes. Zhou et al. combines multi-scale reconstruction and multi-scale contrastive learning to learn multi-scale features. Although those methods achieve success in learning multi-scale features, they are not aware of anatomical structure variations because of the fixed-size image partition strategy and uniform contribution of different anatomical structures in images." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 213, + 116, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 213, + 116, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 116, + 227 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 234, + 296, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 234, + 296, + 318 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 296, + 318 + ], + "type": "text", + "content": "We present our two-step approach: patch-to-patch discrepancy and patch-to-structure consistency. The patch-to-patch constraint increases the distance between all patches, ensuring that patches with different semantics are pushed further apart. In contrast, the patch-to-structure constraint keeps patches within the same structure close together, enforcing semantic consistency within the same structure." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 324, + 141, + 336 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 141, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 141, + 336 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 342, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 568 + ], + "type": "text", + "content": "Contrastive Learning. Contrastive learning in SSL constructs positive sample pairs to capture shared information [23, 37], typically using a source image and its augmented counterpart as two views of the same image. Based on the theory that the features learned from these paired views suffice for downstream tasks [37], we focus solely on positive pairs from the same source image rather than different ones. We adopt vanilla InfoNCE in MoCo [13] as our baseline, leveraging a teacher-student mechanism and a queue storing previously generated features. Following prior SSL methods in 3D medical imaging, images are partitioned into multiple sub-volumes to optimize memory efficiency [14, 40, 41]. Given " + }, + { + "bbox": [ + 55, + 342, + 296, + 568 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 342, + 296, + 568 + ], + "type": "text", + "content": " non-overlapping subvolumes " + }, + { + "bbox": [ + 55, + 342, + 296, + 568 + ], + "type": "inline_equation", + "content": "\\{v_{i} | i = 1, 2, 3, \\dots, N\\}" + }, + { + "bbox": [ + 55, + 342, + 296, + 568 + ], + "type": "text", + "content": " cropped from a full volume " + }, + { + "bbox": [ + 55, + 342, + 296, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 342, + 296, + 568 + ], + "type": "text", + "content": ", we randomly select a subvolume " + }, + { + "bbox": [ + 55, + 342, + 296, + 568 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 55, + 342, + 296, + 568 + ], + "type": "text", + "content": " and generate an augmented version " + }, + { + "bbox": [ + 55, + 342, + 296, + 568 + ], + "type": "inline_equation", + "content": "v_{i}'" + }, + { + "bbox": [ + 55, + 342, + 296, + 568 + ], + "type": "text", + "content": " as a positive pair. The constraint follows a dictionary look-up strategy, where a query feature is matched against candidate features in the dictionary. More details can be found in Appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 569, + 296, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 296, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 296, + 630 + ], + "type": "text", + "content": "Based on the dictionary " + }, + { + "bbox": [ + 55, + 569, + 296, + 630 + ], + "type": "inline_equation", + "content": "\\{q_m^* | m = 1, 2, \\dots, K\\}" + }, + { + "bbox": [ + 55, + 569, + 296, + 630 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 569, + 296, + 630 + ], + "type": "inline_equation", + "content": "q_m^*" + }, + { + "bbox": [ + 55, + 569, + 296, + 630 + ], + "type": "text", + "content": " is the feature encoded by the teacher model " + }, + { + "bbox": [ + 55, + 569, + 296, + 630 + ], + "type": "inline_equation", + "content": "\\varphi^*" + }, + { + "bbox": [ + 55, + 569, + 296, + 630 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 569, + 296, + 630 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 569, + 296, + 630 + ], + "type": "text", + "content": " is the size of the dictionary and the query " + }, + { + "bbox": [ + 55, + 569, + 296, + 630 + ], + "type": "inline_equation", + "content": "q_i" + }, + { + "bbox": [ + 55, + 569, + 296, + 630 + ], + "type": "text", + "content": " encoded by the student model " + }, + { + "bbox": [ + 55, + 569, + 296, + 630 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 55, + 569, + 296, + 630 + ], + "type": "text", + "content": ", the InfoNCE loss is conducted as a contrastive loss " + }, + { + "bbox": [ + 55, + 569, + 296, + 630 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_g" + }, + { + "bbox": [ + 55, + 569, + 296, + 630 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 99, + 635, + 296, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 635, + 296, + 673 + ], + "spans": [ + { + "bbox": [ + 99, + 635, + 296, + 673 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {g} = - \\log \\frac {\\exp \\left(q _ {i} \\cdot q _ {i} ^ {* ^ {\\prime}} / \\tau\\right)}{\\sum_ {m = 0} ^ {K} \\exp \\left(q _ {i} \\cdot q _ {m} ^ {*}\\right) / \\tau}, \\tag {1}", + "image_path": "a6cc40422263bbb854a80c8dd99370ba58b06c96d765675d9937178d101ed97e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 247, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 247, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 247, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 247, + 689 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 677, + 247, + 689 + ], + "type": "text", + "content": " is a temperature hyper-parameter [13]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Geometric equivariance. Mao et al. [21] propose that the model satisfies equivariance, a meta-property that can" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "type": "text", + "content": "be applied to dense feature maps and generalized to most existing vision tasks. Equivariance implies that geometric transformations applied to the input image are preserved in the output features, which benefits many tasks [3, 21]. Consequently, we expect the token features from the Transformer encoder " + }, + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "type": "text", + "content": " to undergo the same homography transformation as the input images. Based on this assumption, we can establish correspondences between token features in the two input volumes to compute patch-to-patch and patch-to-structure relationships. Remarkably, we consider each token to correspond to a patch in the input volume." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 211, + 465, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 465, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 465, + 224 + ], + "type": "text", + "content": "3.2. Patch-to-patch discrepancy" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 228, + 555, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 311 + ], + "type": "text", + "content": "Given a volume " + }, + { + "bbox": [ + 313, + 228, + 555, + 311 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_i" + }, + { + "bbox": [ + 313, + 228, + 555, + 311 + ], + "type": "text", + "content": " and its augmented counterpart " + }, + { + "bbox": [ + 313, + 228, + 555, + 311 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_i'" + }, + { + "bbox": [ + 313, + 228, + 555, + 311 + ], + "type": "text", + "content": ", we build patch-to-patch correspondence between them. A reliable patch-to-patch correspondence refers to a correct one-to-one alignment among two patch sets, which treats every patch as a distinct object. This step aims to enlarge the distribution gap between all semantics and capture distinct characteristics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 312, + 554, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 554, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 554, + 396 + ], + "type": "text", + "content": "In mSSL, augmentations can be broadly classified into geometric and intensity transformations, with only geometric transformations affecting voxel positions. Geometric transformations involve modeling an affine matrix " + }, + { + "bbox": [ + 313, + 312, + 554, + 396 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 312, + 554, + 396 + ], + "type": "text", + "content": ", which is applied to the target image to alter its spatial configuration. Therefore, in this section, we discuss the situation after applying an affine transformation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 396, + 554, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 554, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 554, + 469 + ], + "type": "text", + "content": "Ground truth of patch correspondence. Given two patch centers " + }, + { + "bbox": [ + 313, + 396, + 554, + 469 + ], + "type": "inline_equation", + "content": "c_{i} = (x_{i},y_{i},z_{i})" + }, + { + "bbox": [ + 313, + 396, + 554, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 396, + 554, + 469 + ], + "type": "inline_equation", + "content": "c_{j} = (x_{j},y_{j},z_{j})" + }, + { + "bbox": [ + 313, + 396, + 554, + 469 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 396, + 554, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{i}" + }, + { + "bbox": [ + 313, + 396, + 554, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 396, + 554, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_i^\\prime" + }, + { + "bbox": [ + 313, + 396, + 554, + 469 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 396, + 554, + 469 + ], + "type": "inline_equation", + "content": "x, y," + }, + { + "bbox": [ + 313, + 396, + 554, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 396, + 554, + 469 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 396, + 554, + 469 + ], + "type": "text", + "content": " are coordinates, the correspondence between two patches is determined based on the re-projection distance of their central positions mapped in the input volume, as described in Eq. (2) and Eq. (3)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 323, + 486, + 555, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 486, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 323, + 486, + 555, + 519 + ], + "type": "interline_equation", + "content": "\\mathcal {M} _ {g t} (i, j) \\triangleq \\left\\{ \\begin{array}{l l} 1 & i f \\langle H (c _ {i}), c _ {j} \\rangle \\wedge \\langle H ^ {- 1} (c _ {j}), c _ {i} \\rangle , \\\\ 0 & e l s e. \\end{array} \\right. \\tag {2}", + "image_path": "68f39b80f948d15467d3222728abc453429b9a16ece304eb2ecb1f27788685de.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 526, + 554, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 554, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 554, + 586 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 526, + 554, + 586 + ], + "type": "inline_equation", + "content": "H^{-1}(\\cdot)" + }, + { + "bbox": [ + 313, + 526, + 554, + 586 + ], + "type": "text", + "content": " denotes the inverse of the affine transformation " + }, + { + "bbox": [ + 313, + 526, + 554, + 586 + ], + "type": "inline_equation", + "content": "H(\\cdot)" + }, + { + "bbox": [ + 313, + 526, + 554, + 586 + ], + "type": "text", + "content": " in 3D space, yielding the transformed position. The notation " + }, + { + "bbox": [ + 313, + 526, + 554, + 586 + ], + "type": "inline_equation", + "content": "\\langle \\cdot, \\cdot \\rangle" + }, + { + "bbox": [ + 313, + 526, + 554, + 586 + ], + "type": "text", + "content": " serves as the metric to determine whether two patches originate from the same patch. Specifically, the metric is:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 592, + 555, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 592, + 555, + 646 + ], + "spans": [ + { + "bbox": [ + 320, + 592, + 555, + 646 + ], + "type": "interline_equation", + "content": "\\langle H \\left(c _ {i}\\right), c _ {j} \\rangle \\triangleq \\left\\{ \\begin{array}{l} 1, i f (\\bigtriangleup x _ {i j}, \\bigtriangleup y _ {i j}, \\bigtriangleup z _ {i j}) \\\\ \\in \\left[ - \\frac {W}{2}, \\frac {W}{2} \\right] \\times \\left[ - \\frac {H}{2}, \\frac {H}{2} \\right] \\times \\left[ - \\frac {D}{2}, \\frac {D}{2} \\right], \\\\ 0, e l s e, \\end{array} \\right. \\tag {3}", + "image_path": "2d7580c2b3ce588ee1a93893a6400e7f09dc3ba0b1db04d5802976d699978b8c.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "type": "text", + "content": " are the width, height and depth of patch and " + }, + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "type": "inline_equation", + "content": "\\triangle x_{ij}, \\triangle y_{ij}" + }, + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "type": "inline_equation", + "content": "\\triangle z_{ij}" + }, + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "type": "text", + "content": " are the coordinate differences of " + }, + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "type": "inline_equation", + "content": "H(c_i)" + }, + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "type": "inline_equation", + "content": "c_j" + }, + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "type": "text", + "content": ". With Eq. (2) and Eq. (3), the ground-truth patch-to-patch correspondences " + }, + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{gt}" + }, + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "type": "text", + "content": " between two images are obtained." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20259" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 101, + 72, + 512, + 261 + ], + "blocks": [ + { + "bbox": [ + 101, + 72, + 512, + 261 + ], + "lines": [ + { + "bbox": [ + 101, + 72, + 512, + 261 + ], + "spans": [ + { + "bbox": [ + 101, + 72, + 512, + 261 + ], + "type": "image", + "image_path": "38530a1509f775f782fa63d9617831187a116249d5e6bcd1caa0cbddb9ae059d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 266, + 555, + 313 + ], + "lines": [ + { + "bbox": [ + 54, + 266, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 555, + 313 + ], + "type": "text", + "content": "Figure 2. The pipeline of our SSL framework. " + }, + { + "bbox": [ + 54, + 266, + 555, + 313 + ], + "type": "inline_equation", + "content": "S^2 DC" + }, + { + "bbox": [ + 54, + 266, + 555, + 313 + ], + "type": "text", + "content": " is established on patch features (i.e., token feature in vision transformer) and incorporates two main steps: (1) Patch-to-patch discrepancy. The model aligns two patch sets, which treats every patch as distinct semantics to enlarge the distribution gap between all semantics. (2) Patch-to-structure consistency. The model applies a soft regularization to align patches within the same structure. By those two steps, the framework captures semantic connectivity between anatomical structures." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 325, + 294, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 325, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 54, + 325, + 294, + 503 + ], + "type": "text", + "content": "Building reliable patch-to-patch correspondences. Given token features from " + }, + { + "bbox": [ + 54, + 325, + 294, + 503 + ], + "type": "inline_equation", + "content": "\\varphi^{*}(\\mathcal{V}_{i}^{\\prime})" + }, + { + "bbox": [ + 54, + 325, + 294, + 503 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 325, + 294, + 503 + ], + "type": "inline_equation", + "content": "\\varphi (\\mathcal{V}_i)" + }, + { + "bbox": [ + 54, + 325, + 294, + 503 + ], + "type": "text", + "content": ", we can calculate the similarity map " + }, + { + "bbox": [ + 54, + 325, + 294, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_t" + }, + { + "bbox": [ + 54, + 325, + 294, + 503 + ], + "type": "text", + "content": " between tokens and get the loss between " + }, + { + "bbox": [ + 54, + 325, + 294, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_t" + }, + { + "bbox": [ + 54, + 325, + 294, + 503 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 325, + 294, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{gt}" + }, + { + "bbox": [ + 54, + 325, + 294, + 503 + ], + "type": "text", + "content": ". However, similar to the issues in image matching [27, 31], reliably identifying patch correspondence in medical images is challenging due to repetitive patterns. This intractable challenge would be optimized by solving a differentiable optimal transport problem, such as the Sinkhorn algorithm [30]. The Sinkhorn algorithm is used to solve entropy-regularized optimal transport, enabling soft assignment of correspondences between feature points by iteratively refining a doubly stochastic matrix, improving robustness to noise and deformations. The efficiency of Sinkhorn is affected by the large number of iterations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 505, + 295, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 295, + 577 + ], + "type": "text", + "content": "Rocco et al. propose using a dual-softmax operator to establish reliable correspondences based on neighborhood consensus, which brings mutual nearest neighbors closer to learn a more compact similarity distribution. Given the similarity map " + }, + { + "bbox": [ + 55, + 505, + 295, + 577 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_t" + }, + { + "bbox": [ + 55, + 505, + 295, + 577 + ], + "type": "text", + "content": " between two images, the similarities after applying the dual-softmax operator are:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 60, + 586, + 295, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 586, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 60, + 586, + 295, + 601 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {M}} _ {t} (i, j) = \\operatorname {s o f t m a x} \\left(\\mathcal {M} _ {t} (i, \\cdot)\\right) \\cdot \\operatorname {s o f t m a x} \\left(\\left(\\mathcal {M} _ {t} (\\cdot , j)\\right). \\right. \\tag {4}", + "image_path": "cc8a8d769351c047334336f47d40b8b0ebeb47295ccd47902ee616376e9b24b0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 610, + 295, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 610, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 610, + 295, + 673 + ], + "type": "text", + "content": "Based on the ground-truth correspondences " + }, + { + "bbox": [ + 55, + 610, + 295, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{gt}" + }, + { + "bbox": [ + 55, + 610, + 295, + 673 + ], + "type": "text", + "content": " and the re-assigned feature similarity matrix " + }, + { + "bbox": [ + 55, + 610, + 295, + 673 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{M}}_t" + }, + { + "bbox": [ + 55, + 610, + 295, + 673 + ], + "type": "text", + "content": ", we use the cross-entropy loss to minimize Kullback-Leibler divergence (KL divergence) between them [36]. This patch-to-patch loss " + }, + { + "bbox": [ + 55, + 610, + 295, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{p2p}" + }, + { + "bbox": [ + 55, + 610, + 295, + 673 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 681, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 681, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 61, + 681, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p 2 p} = - \\frac {1}{| \\mathcal {M} _ {g t} |} \\sum_ {i = 0} ^ {N} \\sum_ {j = 0} ^ {N} \\mathcal {M} _ {g t} (i, j) \\times \\log (\\hat {\\mathcal {M}} _ {t} (i, j)), \\tag {5}", + "image_path": "4ac57a775b13bd361af09ac0c3e0f9b770486395dbd75a0fb076b46514162914.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 325, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 555, + 348 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 325, + 555, + 348 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 325, + 555, + 348 + ], + "type": "text", + "content": " represents the number of patches (i.e., tokens in the 3D Transformer)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 371, + 479, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 479, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 479, + 384 + ], + "type": "text", + "content": "3.3. Patch-to-structure consistency" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 393, + 555, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 555, + 477 + ], + "type": "text", + "content": "The correspondence assignment focuses on enhancing reliable patch-to-patch correspondence, which treats every patch as distinct. Patch-to-patch discrepancy limits the exploration of connectivity among patches from the same semantics (i.e., same anatomical structure). Thus, a patch-to-structure correspondence should be explored to improve consistency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 481, + 556, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 481, + 556, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 481, + 556, + 661 + ], + "type": "text", + "content": "We define neighborhood similarity distribution as " + }, + { + "bbox": [ + 313, + 481, + 556, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_n" + }, + { + "bbox": [ + 313, + 481, + 556, + 661 + ], + "type": "text", + "content": ", which represents the similarity vector of an anchor patch feature " + }, + { + "bbox": [ + 313, + 481, + 556, + 661 + ], + "type": "inline_equation", + "content": "t_n" + }, + { + "bbox": [ + 313, + 481, + 556, + 661 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 481, + 556, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_i" + }, + { + "bbox": [ + 313, + 481, + 556, + 661 + ], + "type": "text", + "content": " with all the patch features from " + }, + { + "bbox": [ + 313, + 481, + 556, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_i'" + }, + { + "bbox": [ + 313, + 481, + 556, + 661 + ], + "type": "text", + "content": ". As shown in Fig. 3, patch pairs with the same semantics form distinct peaks in the distribution " + }, + { + "bbox": [ + 313, + 481, + 556, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_n" + }, + { + "bbox": [ + 313, + 481, + 556, + 661 + ], + "type": "text", + "content": ", which meets the assumption that the same structures have the same semantics. Thus, the central tendency and dispersion of " + }, + { + "bbox": [ + 313, + 481, + 556, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_n" + }, + { + "bbox": [ + 313, + 481, + 556, + 661 + ], + "type": "text", + "content": " can be employed to constrain semantic connectivity. For instance, if " + }, + { + "bbox": [ + 313, + 481, + 556, + 661 + ], + "type": "inline_equation", + "content": "t_n" + }, + { + "bbox": [ + 313, + 481, + 556, + 661 + ], + "type": "text", + "content": " is similar to several patches of " + }, + { + "bbox": [ + 313, + 481, + 556, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_i'" + }, + { + "bbox": [ + 313, + 481, + 556, + 661 + ], + "type": "text", + "content": ", indicating they may belong to the same anatomical structure, " + }, + { + "bbox": [ + 313, + 481, + 556, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_n" + }, + { + "bbox": [ + 313, + 481, + 556, + 661 + ], + "type": "text", + "content": " tends to be more dispersed (Fig. 3 (b)). In contrast, with no semantic-similar patches, the distribution of " + }, + { + "bbox": [ + 313, + 481, + 556, + 661 + ], + "type": "inline_equation", + "content": "t_n" + }, + { + "bbox": [ + 313, + 481, + 556, + 661 + ], + "type": "text", + "content": " is more concentrated. Therefore, we utilize the characteristics of " + }, + { + "bbox": [ + 313, + 481, + 556, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_n" + }, + { + "bbox": [ + 313, + 481, + 556, + 661 + ], + "type": "text", + "content": " as a soft label to regularize the patch-to-patch loss " + }, + { + "bbox": [ + 313, + 481, + 556, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{p2p}" + }, + { + "bbox": [ + 313, + 481, + 556, + 661 + ], + "type": "text", + "content": " to maintain underlying semantic relevance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "We introduce a simplified version of the Sharpe ratio [28] to reflect the dispersion and concentration of neighborhood similarity distribution " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_n" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": ". Given the " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_n" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " of the patch feature " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "t_n" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": ", the Sharpe ratio of " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "t_n" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " can be computed as:" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20260" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 98, + 83, + 299, + 188 + ], + "blocks": [ + { + "bbox": [ + 98, + 83, + 299, + 188 + ], + "lines": [ + { + "bbox": [ + 98, + 83, + 299, + 188 + ], + "spans": [ + { + "bbox": [ + 98, + 83, + 299, + 188 + ], + "type": "image", + "image_path": "fca96db32f5aa2436152494b128a1d03d66a25befbf68266629f88604755d534.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 121, + 194, + 274, + 209 + ], + "lines": [ + { + "bbox": [ + 121, + 194, + 274, + 209 + ], + "spans": [ + { + "bbox": [ + 121, + 194, + 274, + 209 + ], + "type": "text", + "content": "(a) Patches with same semantics" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 310, + 79, + 512, + 187 + ], + "blocks": [ + { + "bbox": [ + 310, + 79, + 512, + 187 + ], + "lines": [ + { + "bbox": [ + 310, + 79, + 512, + 187 + ], + "spans": [ + { + "bbox": [ + 310, + 79, + 512, + 187 + ], + "type": "image", + "image_path": "842e885f379ce6edf569888115f8a5aa41fe301895d8d8bdacd848d24e8a5f36.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 194, + 507, + 209 + ], + "lines": [ + { + "bbox": [ + 317, + 194, + 507, + 209 + ], + "spans": [ + { + "bbox": [ + 317, + 194, + 507, + 209 + ], + "type": "text", + "content": "(b) Anchor patch with distinct semantics" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 217, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 54, + 217, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 54, + 217, + 555, + 251 + ], + "type": "text", + "content": "Figure 3. Illustration of the similarity distribution " + }, + { + "bbox": [ + 54, + 217, + 555, + 251 + ], + "type": "inline_equation", + "content": "D_{n}" + }, + { + "bbox": [ + 54, + 217, + 555, + 251 + ], + "type": "text", + "content": ". (a) Patches with the same semantics (e.g., bone). Given an anchor patch, patches from the same semantics form several peaks in " + }, + { + "bbox": [ + 54, + 217, + 555, + 251 + ], + "type": "inline_equation", + "content": "D_{n}" + }, + { + "bbox": [ + 54, + 217, + 555, + 251 + ], + "type": "text", + "content": ". (b) Anchor patch with distinct semantics (e.g., pancreas), " + }, + { + "bbox": [ + 54, + 217, + 555, + 251 + ], + "type": "inline_equation", + "content": "D_{n}" + }, + { + "bbox": [ + 54, + 217, + 555, + 251 + ], + "type": "text", + "content": " shows only one large peak (its augmented patch)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 269, + 295, + 299 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 269, + 295, + 299 + ], + "spans": [ + { + "bbox": [ + 102, + 269, + 295, + 299 + ], + "type": "interline_equation", + "content": "s r _ {\\mathcal {V} _ {i}} ^ {n} = \\frac {\\operatorname* {m a x} \\left(\\mathcal {D} _ {n}\\right) - \\frac {1}{N} \\sum_ {m = 1} ^ {N} \\mathcal {D} _ {n}}{\\sigma_ {\\mathcal {D} _ {n}}}, \\tag {6}", + "image_path": "24eb6ef9b6b85dd2b80b1a7a025a92856e531cb6e762152d5c83483713006e20.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 303, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 296, + 373 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 55, + 303, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathcal{D}_n}" + }, + { + "bbox": [ + 55, + 303, + 296, + 373 + ], + "type": "text", + "content": " is the variance of vector " + }, + { + "bbox": [ + 55, + 303, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_n" + }, + { + "bbox": [ + 55, + 303, + 296, + 373 + ], + "type": "text", + "content": " and the notation " + }, + { + "bbox": [ + 55, + 303, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\max (\\mathcal{D}_n)" + }, + { + "bbox": [ + 55, + 303, + 296, + 373 + ], + "type": "text", + "content": " is the maximal value of " + }, + { + "bbox": [ + 55, + 303, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_n" + }, + { + "bbox": [ + 55, + 303, + 296, + 373 + ], + "type": "text", + "content": ". A low Sharpe ratio of " + }, + { + "bbox": [ + 55, + 303, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_n" + }, + { + "bbox": [ + 55, + 303, + 296, + 373 + ], + "type": "text", + "content": " indicates that the distribution is more dispersed, which means the patch feature " + }, + { + "bbox": [ + 55, + 303, + 296, + 373 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 55, + 303, + 296, + 373 + ], + "type": "text", + "content": " has more than one similar semantic patch feature. More details can be seen in Appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 374, + 296, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 296, + 472 + ], + "type": "inline_equation", + "content": "sr_{\\mathcal{V}_i}^n" + }, + { + "bbox": [ + 55, + 374, + 296, + 472 + ], + "type": "text", + "content": " is utilized as the soft regularization to adaptively adjust " + }, + { + "bbox": [ + 55, + 374, + 296, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{p2p}" + }, + { + "bbox": [ + 55, + 374, + 296, + 472 + ], + "type": "text", + "content": ", serving as soft evidence for maintaining a semantic connectivity relationship among patch features. The weighted losses for the feature " + }, + { + "bbox": [ + 55, + 374, + 296, + 472 + ], + "type": "inline_equation", + "content": "t_n" + }, + { + "bbox": [ + 55, + 374, + 296, + 472 + ], + "type": "text", + "content": " of the volume " + }, + { + "bbox": [ + 55, + 374, + 296, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_i" + }, + { + "bbox": [ + 55, + 374, + 296, + 472 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 55, + 374, + 296, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c(n,\\cdot) = \\text{softmax}(sr_{\\mathcal{V}_i})_n\\mathcal{L}_{p2p}(n,\\cdot)" + }, + { + "bbox": [ + 55, + 374, + 296, + 472 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 374, + 296, + 472 + ], + "type": "inline_equation", + "content": "sr_{\\mathcal{V}_i} = \\{sr_{\\mathcal{V}_i}^n\\}_n" + }, + { + "bbox": [ + 55, + 374, + 296, + 472 + ], + "type": "text", + "content": ". Similarly, we have " + }, + { + "bbox": [ + 55, + 374, + 296, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c(\\cdot,m) = \\text{softmax}(sr_{\\mathcal{V}_i'})_m\\mathcal{L}_{p2p}(\\cdot,m)" + }, + { + "bbox": [ + 55, + 374, + 296, + 472 + ], + "type": "text", + "content": ". Thus, we have the loss of a single pair:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 490, + 295, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 295, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 295, + 526 + ], + "type": "interline_equation", + "content": "l _ {n m} = \\frac {\\left(\\operatorname {s o f t m a x} \\left(s r _ {\\mathcal {V} _ {i}}\\right) _ {n} + \\operatorname {s o f t m a x} \\left(s r _ {\\mathcal {V} _ {i} ^ {\\prime}}\\right) _ {m}\\right) \\mathcal {L} _ {p 2 p} (n , m)}{2}. \\tag {7}", + "image_path": "cd9015b12b429a5f8772c40e2eeff5a5f2431fb082c63e8d730e2921140a081f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 526, + 227, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 227, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 227, + 539 + ], + "type": "text", + "content": "Finally, the patch-to-structure loss " + }, + { + "bbox": [ + 55, + 526, + 227, + 539 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{p2s}" + }, + { + "bbox": [ + 55, + 526, + 227, + 539 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 556, + 295, + 584 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 556, + 295, + 584 + ], + "spans": [ + { + "bbox": [ + 116, + 556, + 295, + 584 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p 2 s} = \\frac {\\sum_ {m = 0} ^ {N} \\sum_ {n = 0} ^ {N} l _ {n , m}}{N \\times N}. \\tag {8}", + "image_path": "279549058db949b8677c00e1f4ce19e3f332bb4cc111bfb048f41826aa235a6d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 592, + 176, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 176, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 176, + 604 + ], + "type": "text", + "content": "3.4. Overall loss function" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 609, + 295, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 295, + 632 + ], + "type": "text", + "content": "The final combined objective for the proposed framework is defined as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 125, + 634, + 295, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 634, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 125, + 634, + 295, + 647 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {g} + \\mathcal {L} _ {p 2 p} + \\mathcal {L} _ {p 2 s}, \\tag {9}", + "image_path": "72296434dd90d2a1d5cb1288342a9bd737895b5604d121d33dbfb7586212ed1f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 657, + 137, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 137, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 137, + 671 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "This section begins with an overview of the datasets utilized for both pre-training and downstream tasks. Next, we provide a brief summary of the implementation details for the" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 261, + 555, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 555, + 297 + ], + "type": "text", + "content": "method. Lastly, we present a comprehensive comparison of the experimental results for our proposed method against other SOTA SSL approaches in 3D medical imaging." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 308, + 507, + 321 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 507, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 507, + 321 + ], + "type": "text", + "content": "4.1. Datasets and implementation details" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 327, + 554, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 554, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 554, + 410 + ], + "type": "text", + "content": "Pre-training and downstream datasets. Although we emphasize the anatomical structure-specific semantic connectivity, the training process of our mining semantic connective strategy is also suitable for positron emission tomography (PET). In our experiments, we include computed tomography (CT), magnetic resonance imaging (MRI), and PET modalities." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 411, + 554, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 554, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 554, + 532 + ], + "type": "text", + "content": "We use the same dataset settings as VoCo [41] on CT, i.e., 10k data collection. BTCV and TCIA COVID-19 are used as pre-training datasets (1k data) for the BTCV segmentation task, while all other CT segmentation tasks are pre-trained on 10k data. Additionally, we collect 6605 (6k) unlabeled MRI scans from OASIS3 [18] and 3519 (3k) PET scans from UDPET [4] and ADNI [22] for pre-training. We evaluate methods on 10 datasets in 3 modalities, as shown in Tab. 2. More details about dataset splits are provided in Appendix." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 532, + 554, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 554, + 700 + ], + "type": "text", + "content": "Training setup. For pre-training stage, " + }, + { + "bbox": [ + 313, + 532, + 554, + 700 + ], + "type": "inline_equation", + "content": "S^2 DC" + }, + { + "bbox": [ + 313, + 532, + 554, + 700 + ], + "type": "text", + "content": " adopts Swin-B [11] as the default backbones in three modalities, following previous works [11, 40, 41]. We set 100k iterations for pre-training in all modalities and crop subvolumes with size " + }, + { + "bbox": [ + 313, + 532, + 554, + 700 + ], + "type": "inline_equation", + "content": "96 \\times 96 \\times 96" + }, + { + "bbox": [ + 313, + 532, + 554, + 700 + ], + "type": "text", + "content": ". In our patch-to-patch loss " + }, + { + "bbox": [ + 313, + 532, + 554, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{p2p}" + }, + { + "bbox": [ + 313, + 532, + 554, + 700 + ], + "type": "text", + "content": ", we adopt the dual-softmax operator for efficiency, while the results of the Sinkhorn algorithm can be found in the Appendix. For downstream tasks, we adopt Swin-B and its derivative architectures. The weights of models are initialized by the weights of pre-training models. We adopt the Swin-UNETR [14] framework in segmentation, image-to-image translation, and reconstruction tasks. Swin-B [11] is used as the backbone of classification tasks. More details on pre-training and training settings are in the Appendix." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 325, + 700, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 700, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 700, + 554, + 713 + ], + "type": "text", + "content": "Comparison methods. We compare " + }, + { + "bbox": [ + 325, + 700, + 554, + 713 + ], + "type": "inline_equation", + "content": "S^2 DC" + }, + { + "bbox": [ + 325, + 700, + 554, + 713 + ], + "type": "text", + "content": " with SOTA" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20261" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 555, + 335 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 335 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 335 + ], + "type": "table", + "html": "
MethodDice Score(%)
SplRKidLKidGallEsoLivStoAorIVCVeinsPanRAGLAGAVG
From Scratch
UNETR [12]93.0294.1394.1266.9970.8796.1177.2789.2282.170.1676.6565.3259.2179.82
Swin-UNETR [14]94.0693.5493.8065.6174.6097.0975.9491.882.3673.6375.1968.0061.1180.53
With General SSL
MAE3D [7]93.9894.3794.1869.8674.6496.6680.4090.3083.1372.6577.1167.3460.5481.33
SimCLR [6]92.7993.0491.4149.6550.9998.4977.9285.5680.5864.3767.1659.0448.9973.85
SimMIM [42]95.5695.8294.1452.0653.5298.9880.2588.1182.9866.4969.1660.8850.4576.03
MoCo v3 [13]91.9692.8592.4268.2572.7794.9178.8288.2181.5971.1575.7666.4858.8179.54
Jigsaw [5]94.6293.4193.5575.6373.2195.7180.889.4184.7871.0279.5765.6860.2281.35
PositionLabel [47]94.3593.1593.2175.3973.2495.7680.6988.884.0471.1879.0265.1160.0781.09
With medical image SSL
MG [50]91.9993.5291.8165.1176.4195.9886.8889.2983.5971.7981.6267.9763.1881.45
ROT [32]91.7593.1391.6265.0976.5594.2186.1689.7483.0871.1381.5567.9063.7281.20
Vicregl [1]90.3294.1591.3065.1275.4194.7686.0089.1382.5471.2681.0167.6663.0880.89
Rubik++ [35]96.2190.4189.3375.2272.6497.4479.2589.6583.7674.7478.3567.1461.9781.38
PCRLv1 [48]95.7389.6688.5375.4172.3396.2078.9989.1183.0674.4777.8867.0261.8580.78
PCRLv2 [49]95.5091.4389.5276.1573.5497.2879.6490.1684.1775.2078.7168.7462.9381.74
Swin-UNETR [14]95.2593.1692.9763.6273.9696.2179.3289.9883.1976.1182.2568.9965.1181.54
SwinMM [40]94.3394.1894.1672.9774.7596.3783.2389.5682.9170.6575.5269.1762.9081.81
GL-MAE [52]94.5494.3994.3773.1974.9396.5183.4989.7483.1170.8075.7169.3963.1282.01
GVSL [15]95.2791.2292.2572.6973.5696.4482.4088.9084.2270.8476.4267.4863.2581.87
VoCo [41]95.7396.5394.4876.0275.6097.4178.4391.2186.1278.1980.8871.4767.8883.85
\\( S^2DC \\)96.2496.5194.5274.2777.6996.9583.3390.4486.6675.2284.2468.2269.7384.14
", + "image_path": "d94ffb0ccb96d412475c386cba030d878b6f75f3dd3a80ba0c0040d3ea99c4eb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 86, + 383, + 266, + 502 + ], + "blocks": [ + { + "bbox": [ + 55, + 342, + 555, + 366 + ], + "lines": [ + { + "bbox": [ + 55, + 342, + 555, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 555, + 366 + ], + "type": "text", + "content": "Table 1. Results on the segmentation task BTCV fine-tuning from pre-training models trained on 1k CT data. Most of the results are drawn from [41, 45]. Best and second best are highlighted" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 86, + 383, + 266, + 502 + ], + "lines": [ + { + "bbox": [ + 86, + 383, + 266, + 502 + ], + "spans": [ + { + "bbox": [ + 86, + 383, + 266, + 502 + ], + "type": "table", + "html": "
TaskDatasetModality
SegmentationBTCV [19]CT
MSD-Liver [29]CT
MSD-Lung [29]CT
MSD-Spleen [29]CT
BraTs 21 [29]MRI
AUTOPET [10]PET
ClassificationCC-CCII [46]CT
ADNI-cls [22]PET
ReconstructionUDPET [4]PET
I2I translationBraTs 23 [29]MRI
", + "image_path": "c3b46e87689df02e3fc8bce4ddf78b300c378a488881b9930f0492952524f1ba.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 542, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 295, + 592 + ], + "type": "text", + "content": "SSL methods in general [5-7, 13, 42, 47] and medical [1, 14, 15, 32, 35, 40, 41, 48-50, 52] images. Meanwhile, we report results of training from scratch (i.e., without initialized weights of pre-training models)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 599, + 217, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 217, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 217, + 611 + ], + "type": "text", + "content": "4.2. Results on downstream tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 616, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 676 + ], + "type": "text", + "content": "We report the results of 4 downstream tasks on 3 modalities. Following previous work, we use the metrics Dice score for segmentation, accuracy for classification, PSNR, SSIM, and NMSE for I2I translation, and SSIM and PSNR for reconstruction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Segmentation tasks. Compared to volume-level and voxel-level methods, " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "S^2 DC" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " performs well both in small targets (AUTOPET, MSD-Lung) and large organs (BTCV," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 385, + 555, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 555, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 555, + 470 + ], + "type": "text", + "content": "MSD-Liver, MSD-Spleen and BraTs 21) segmentation, as shown in Tab. 1 and Tab. 3. In BTCV datasets, compared to general SSL methods, medical SSL methods show superior performance. Notably, with " + }, + { + "bbox": [ + 313, + 385, + 555, + 470 + ], + "type": "inline_equation", + "content": "S^2DC" + }, + { + "bbox": [ + 313, + 385, + 555, + 470 + ], + "type": "text", + "content": " pre-training, we achieve an average improvement of " + }, + { + "bbox": [ + 313, + 385, + 555, + 470 + ], + "type": "inline_equation", + "content": "4.32\\%" + }, + { + "bbox": [ + 313, + 385, + 555, + 470 + ], + "type": "text", + "content": ", reaching a Dice score of " + }, + { + "bbox": [ + 313, + 385, + 555, + 470 + ], + "type": "inline_equation", + "content": "75.47\\%" + }, + { + "bbox": [ + 313, + 385, + 555, + 470 + ], + "type": "text", + "content": ", which surpasses other SOTA medical SSL methods by " + }, + { + "bbox": [ + 313, + 385, + 555, + 470 + ], + "type": "inline_equation", + "content": "0.81\\%" + }, + { + "bbox": [ + 313, + 385, + 555, + 470 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 470, + 556, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 556, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 556, + 555 + ], + "type": "text", + "content": "Classification tasks. We conduct experiments on the CC-CCII dataset for COVID-19 classification and ADNI for AD (Alzheimer's Disease), MCI (Mild Cognitive Impairment), and CN (Cognitively Normal) classification. With the pre-trained model, " + }, + { + "bbox": [ + 313, + 470, + 556, + 555 + ], + "type": "inline_equation", + "content": "S^2 DC" + }, + { + "bbox": [ + 313, + 470, + 556, + 555 + ], + "type": "text", + "content": " achieves an average improvement of " + }, + { + "bbox": [ + 313, + 470, + 556, + 555 + ], + "type": "inline_equation", + "content": "3.93\\%" + }, + { + "bbox": [ + 313, + 470, + 556, + 555 + ], + "type": "text", + "content": " and outperforms other medical SSL methods." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 321, + 563, + 549, + 662 + ], + "blocks": [ + { + "bbox": [ + 55, + 509, + 295, + 533 + ], + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 533 + ], + "type": "text", + "content": "Table 2. The downstream tasks and modalities. I2I translation represents image-to-image translation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 563, + 549, + 662 + ], + "lines": [ + { + "bbox": [ + 321, + 563, + 549, + 662 + ], + "spans": [ + { + "bbox": [ + 321, + 563, + 549, + 662 + ], + "type": "table", + "html": "
MethodNetworkCC-CCIIADNI-clsAVG
Acc(%)Acc(%)Acc(%)
From Scratch
Swin-UNETR [14]Swin-B88.0456.1772.10
With 3D medical image SSL
PCRLv2 [49]Swin-B93.0756.0574.56
Swin-UNETR [14]Swin-B94.1555.9575.05
SwinMM [40]Swin-B94.8056.3675.58
VoCo [41]Swin-B94.6056.3275.46
\\( S^2DC \\)Swin-B95.3456.7276.03
", + "image_path": "8e2c562ae83e3455da6357ab121e6a3c0be097d823159f19c53ecee67111225b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 666, + 555, + 700 + ], + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 700 + ], + "type": "text", + "content": "Table 4. Experimental results on the classification task CC-CCII and ADNI-cls fine-tuning from pre-training models trained on 6k MRI scans. Best and second best are highlighted." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20262" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 81, + 70, + 531, + 190 + ], + "blocks": [ + { + "bbox": [ + 81, + 70, + 531, + 190 + ], + "lines": [ + { + "bbox": [ + 81, + 70, + 531, + 190 + ], + "spans": [ + { + "bbox": [ + 81, + 70, + 531, + 190 + ], + "type": "table", + "html": "
MethodDice(%)
MSD LiverMSD LungMSD SpleenAUTOPETBraTs21Total
WTTCETAVGAVG
From scratch
Swin-UNETR [14]81.1859.7394.2034.4992.8988.2877.3486.1771.15
With 3D medical image SSL
Swin-UNETR [14]81.5862.4694.4943.8592.7088.4278.8286.6473.95
Swin-MM [40]82.3261.1694.6044.3792.8988.5479.1886.8773.86
PCRLv2 [49]82.1262.3995.2342.6492.8188.4978.7486.6873.81
VoCo [41]82.5063.3495.7045.1092.9088.4478.6286.6574.66
\\( S^2DC \\)83.4364.4095.7346.4792.7688.7179.7987.0975.47
", + "image_path": "5b6334f3f5707e19aebeb070af190f74bfb0d85226cba7b426449c28f8955505.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 239, + 296, + 327 + ], + "blocks": [ + { + "bbox": [ + 55, + 198, + 553, + 232 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 553, + 232 + ], + "type": "text", + "content": "Table 3. Results on the segmentation tasks MSD-Liver, MSD-Lung, MSD-Spleen, AUTOPET, and BraTs21. WT, TC, and ET represent the whole tumor, tumor core, and enhancing tumor, respectively. We retrained all methods using " + }, + { + "bbox": [ + 55, + 198, + 553, + 232 + ], + "type": "inline_equation", + "content": "10\\mathrm{k}" + }, + { + "bbox": [ + 55, + 198, + 553, + 232 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 198, + 553, + 232 + ], + "type": "inline_equation", + "content": "6\\mathrm{k}" + }, + { + "bbox": [ + 55, + 198, + 553, + 232 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 198, + 553, + 232 + ], + "type": "inline_equation", + "content": "3\\mathrm{k}" + }, + { + "bbox": [ + 55, + 198, + 553, + 232 + ], + "type": "text", + "content": " data in three modalities separately, except for VoCo in CT, whose weight is provided in the official code. Best and second best are highlighted." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 239, + 296, + 327 + ], + "lines": [ + { + "bbox": [ + 56, + 239, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 296, + 327 + ], + "type": "table", + "html": "
MethodT1->T2T1->T1ce
PSNR↑NMSE↓SSIM↑PSNR↑NMSE↓SSIM↑
From Scratch
Swin-UNETR [14]26.570.13480.880726.910.0820.8871
With 3D medical image SSL
PCRLv2 [49]26.790.16400.882829.620.1230.9179
Swin-UNETR [14]24.810.29380.844130.420.1080.9117
SwinMM [40]24.050.31580.818230.700.0990.9059
VoCo [41]25.810.19160.859831.920.0790.9315
S2DC25.960.11610.883932.650.0410.9412
", + "image_path": "a4a2fc53a1aa22dbf19444f5657ac463880a4071b415750f73c7419df01be819.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 74, + 368, + 278, + 460 + ], + "blocks": [ + { + "bbox": [ + 55, + 331, + 295, + 364 + ], + "lines": [ + { + "bbox": [ + 55, + 331, + 295, + 364 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 295, + 364 + ], + "type": "text", + "content": "Table 5. Experimental results on the image-to-image translation task BraTs23 fine-tuning from pre-training models trained on 6k MRI scans. Best and second best are highlighted." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 368, + 278, + 460 + ], + "lines": [ + { + "bbox": [ + 74, + 368, + 278, + 460 + ], + "spans": [ + { + "bbox": [ + 74, + 368, + 278, + 460 + ], + "type": "table", + "html": "
MethodNetworkSSIM↑PSNR↑
From Scratch
Swin-UNETR [14]-0.9991758.4100
With 3D medical image SSL
PCRLv2 [49]Swin-UNETR0.9991658.2865
Swin-UNETR [14]Swin-UNETR0.9991758.3409
SwinMM [40]Swin-UNETR0.9991758.4002
VoCo [41]Swin-UNETR0.9991658.2561
\\( S^2DC \\)Swin-UNETR0.9991758.4649
", + "image_path": "8d9ebea531fbf8011fcee87a291cc4343676d9ae350a0ee23ce065f61610ac40.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 510, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 676 + ], + "type": "text", + "content": "Image-to-image translation task. We evaluate in BraTs23, which includes four structural MRI modalities for each individual. The translation quality for " + }, + { + "bbox": [ + 55, + 510, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\mathrm{T1} \\rightarrow \\mathrm{T2}" + }, + { + "bbox": [ + 55, + 510, + 295, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 510, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\mathrm{T1} \\rightarrow \\mathrm{T1ce}" + }, + { + "bbox": [ + 55, + 510, + 295, + 676 + ], + "type": "text", + "content": " are compared by peak signal-to-noise ratio (PSNR), normalized mean squared error (NMSE), and structural similarity index (SSIM) [44]. From Table 5, our method excels in two tasks. Remarkably, with " + }, + { + "bbox": [ + 55, + 510, + 295, + 676 + ], + "type": "inline_equation", + "content": "S^2 DC" + }, + { + "bbox": [ + 55, + 510, + 295, + 676 + ], + "type": "text", + "content": " pretraining, we gain " + }, + { + "bbox": [ + 55, + 510, + 295, + 676 + ], + "type": "inline_equation", + "content": "5.41\\%" + }, + { + "bbox": [ + 55, + 510, + 295, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 510, + 295, + 676 + ], + "type": "inline_equation", + "content": "0.32\\%" + }, + { + "bbox": [ + 55, + 510, + 295, + 676 + ], + "type": "text", + "content": " improvements in SSIM separately. We notice that other methods fail to get improvement in the task " + }, + { + "bbox": [ + 55, + 510, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\mathrm{T1} \\rightarrow \\mathrm{T2}" + }, + { + "bbox": [ + 55, + 510, + 295, + 676 + ], + "type": "text", + "content": " compared to From Scratch. That may be due to the deep domain gap between the two sequences, while pre-trained models provide biased initializations. In contrast, " + }, + { + "bbox": [ + 55, + 510, + 295, + 676 + ], + "type": "inline_equation", + "content": "S^2 DC" + }, + { + "bbox": [ + 55, + 510, + 295, + 676 + ], + "type": "text", + "content": " achieves improvement in NMSE and SSIM, showing better ability in this task." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Reconstruction task. We evaluate the capabilities of reconstruction in different methods on the UDPET dataset. Following the previous work [8], whole-body data at" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 242, + 553, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 242, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 553, + 289 + ], + "type": "text", + "content": "DFR=50 (i.e., 1/50 doses) were selected to explore the reconstruction ability of all methods. This task has been extensively studied, yet our method, " + }, + { + "bbox": [ + 313, + 242, + 553, + 289 + ], + "type": "inline_equation", + "content": "S^2 DC" + }, + { + "bbox": [ + 313, + 242, + 553, + 289 + ], + "type": "text", + "content": ", still achieves superior PSNR results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 291, + 555, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 555, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 555, + 410 + ], + "type": "text", + "content": "Overall comparisons on 10 downstream datasets. The results demonstrate that " + }, + { + "bbox": [ + 313, + 291, + 555, + 410 + ], + "type": "inline_equation", + "content": "S^2 DC" + }, + { + "bbox": [ + 313, + 291, + 555, + 410 + ], + "type": "text", + "content": " performs effectively across 10 datasets and 4 tasks. Compared to training from scratch, which achieves an average score of " + }, + { + "bbox": [ + 313, + 291, + 555, + 410 + ], + "type": "inline_equation", + "content": "77.93\\%" + }, + { + "bbox": [ + 313, + 291, + 555, + 410 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 291, + 555, + 410 + ], + "type": "inline_equation", + "content": "S^2 DC" + }, + { + "bbox": [ + 313, + 291, + 555, + 410 + ], + "type": "text", + "content": " pre-training delivers a " + }, + { + "bbox": [ + 313, + 291, + 555, + 410 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 313, + 291, + 555, + 410 + ], + "type": "text", + "content": " improvement, reaching " + }, + { + "bbox": [ + 313, + 291, + 555, + 410 + ], + "type": "inline_equation", + "content": "81.43\\%" + }, + { + "bbox": [ + 313, + 291, + 555, + 410 + ], + "type": "text", + "content": ". Additionally, " + }, + { + "bbox": [ + 313, + 291, + 555, + 410 + ], + "type": "inline_equation", + "content": "S^2 DC" + }, + { + "bbox": [ + 313, + 291, + 555, + 410 + ], + "type": "text", + "content": " outperforms the second-best SSL method (VoCo [41], average score " + }, + { + "bbox": [ + 313, + 291, + 555, + 410 + ], + "type": "inline_equation", + "content": "80.65\\%" + }, + { + "bbox": [ + 313, + 291, + 555, + 410 + ], + "type": "text", + "content": ") for all tasks, with an average gain of " + }, + { + "bbox": [ + 313, + 291, + 555, + 410 + ], + "type": "inline_equation", + "content": "0.78\\%" + }, + { + "bbox": [ + 313, + 291, + 555, + 410 + ], + "type": "text", + "content": ". These findings indicate that " + }, + { + "bbox": [ + 313, + 291, + 555, + 410 + ], + "type": "inline_equation", + "content": "S^2 DC" + }, + { + "bbox": [ + 313, + 291, + 555, + 410 + ], + "type": "text", + "content": " is both reliable and adaptable across various downstream tasks and modalities." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 420, + 406, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 406, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 406, + 433 + ], + "type": "text", + "content": "4.3. Ablation study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 438, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 555, + 582 + ], + "type": "text", + "content": "Influence of different constraints. We use contrastive learning loss " + }, + { + "bbox": [ + 313, + 438, + 555, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_g" + }, + { + "bbox": [ + 313, + 438, + 555, + 582 + ], + "type": "text", + "content": " in Sec. 3.1 as our baseline. The impact of different constraints in pre-training is presented in Tab. 7. The results indicate that combining patch-to-patch loss " + }, + { + "bbox": [ + 313, + 438, + 555, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{p2p}" + }, + { + "bbox": [ + 313, + 438, + 555, + 582 + ], + "type": "text", + "content": " and patch-to-structure loss " + }, + { + "bbox": [ + 313, + 438, + 555, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{p2s}" + }, + { + "bbox": [ + 313, + 438, + 555, + 582 + ], + "type": "text", + "content": " yields SOTA performance compared to Tab. 1. Meanwhile, we visualize the t-SNE clustering results of 13 organs under different constraints, as shown in Fig. 4. The " + }, + { + "bbox": [ + 313, + 438, + 555, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{p2p}" + }, + { + "bbox": [ + 313, + 438, + 555, + 582 + ], + "type": "text", + "content": " constraint leads to a sparser inter-structure distribution, while the " + }, + { + "bbox": [ + 313, + 438, + 555, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{p2s}" + }, + { + "bbox": [ + 313, + 438, + 555, + 582 + ], + "type": "text", + "content": " constraint results in more compact intra-structure clusters. Combining " + }, + { + "bbox": [ + 313, + 438, + 555, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{p2p}" + }, + { + "bbox": [ + 313, + 438, + 555, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 438, + 555, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{p2s}" + }, + { + "bbox": [ + 313, + 438, + 555, + 582 + ], + "type": "text", + "content": " achieves the best overall clustering performance." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 332, + 592, + 537, + 658 + ], + "blocks": [ + { + "bbox": [ + 55, + 464, + 295, + 498 + ], + "lines": [ + { + "bbox": [ + 55, + 464, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 295, + 498 + ], + "type": "text", + "content": "Table 6. Experimental results on low-dose to full-dose reconstruction task UDPET fine-tuning from pre-training models trained on 3k PET scans. Best and second best are highlighted." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 592, + 537, + 658 + ], + "lines": [ + { + "bbox": [ + 332, + 592, + 537, + 658 + ], + "spans": [ + { + "bbox": [ + 332, + 592, + 537, + 658 + ], + "type": "table", + "html": "
Baseline(Lg)+Lp2p+Lp2sBTCV DICE(%)AUTOPET
•83.4345.72
••83.9845.85
••84.0246.23
•••84.1446.47
", + "image_path": "8b8191f9ccd2a6ef977c0b4f0f3edc803efae3103e5c555b2cd8126713bb8946.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 337, + 662, + 529, + 673 + ], + "lines": [ + { + "bbox": [ + 337, + 662, + 529, + 673 + ], + "spans": [ + { + "bbox": [ + 337, + 662, + 529, + 673 + ], + "type": "text", + "content": "Table 7. The ablation results of different constraints." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Intra-structure consistency and inter-structure discrepancy. We visualize the t-SNE of features extracted by four medical pre-trained models and calculate their silhouette" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20263" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 289, + 153 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 289, + 153 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 289, + 153 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 289, + 153 + ], + "type": "image", + "image_path": "d47ced4e1db314b13b8117c4ba47efcc95fbbbf9be78e5d7940ff15fac26a183.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 62, + 163, + 286, + 250 + ], + "blocks": [ + { + "bbox": [ + 106, + 155, + 124, + 163 + ], + "lines": [ + { + "bbox": [ + 106, + 155, + 124, + 163 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 124, + 163 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 106, + 155, + 124, + 163 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_g" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 163, + 286, + 250 + ], + "lines": [ + { + "bbox": [ + 62, + 163, + 286, + 250 + ], + "spans": [ + { + "bbox": [ + 62, + 163, + 286, + 250 + ], + "type": "image", + "image_path": "9b9e96e73c6be48eacaaa298d528f1cf9b6a441120bc5b51c0229f3fe6742841.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 251, + 132, + 260 + ], + "lines": [ + { + "bbox": [ + 97, + 251, + 132, + 260 + ], + "spans": [ + { + "bbox": [ + 97, + 251, + 132, + 260 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 97, + 251, + 132, + 260 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_g + \\mathcal{L}_{p2s}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 211, + 154, + 249, + 163 + ], + "lines": [ + { + "bbox": [ + 211, + 154, + 249, + 163 + ], + "spans": [ + { + "bbox": [ + 211, + 154, + 249, + 163 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 211, + 154, + 249, + 163 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_g + \\mathcal{L}_{p2p}" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 183, + 164, + 285, + 251 + ], + "blocks": [ + { + "bbox": [ + 183, + 164, + 285, + 251 + ], + "lines": [ + { + "bbox": [ + 183, + 164, + 285, + 251 + ], + "spans": [ + { + "bbox": [ + 183, + 164, + 285, + 251 + ], + "type": "image", + "image_path": "322cb4db0aa9699f6d9708ed844f699e2b2bfd7f464a57678823e64e12245dd7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 202, + 251, + 258, + 260 + ], + "lines": [ + { + "bbox": [ + 202, + 251, + 258, + 260 + ], + "spans": [ + { + "bbox": [ + 202, + 251, + 258, + 260 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 202, + 251, + 258, + 260 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_g + \\mathcal{L}_{p2p} + \\mathcal{L}_{p2s}" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 62, + 300, + 289, + 394 + ], + "blocks": [ + { + "bbox": [ + 55, + 271, + 295, + 293 + ], + "lines": [ + { + "bbox": [ + 55, + 271, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 295, + 293 + ], + "type": "text", + "content": "Figure 4. The t-SNE feature visualization of different losses on 13 organs on the BTCV dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 300, + 289, + 394 + ], + "lines": [ + { + "bbox": [ + 62, + 300, + 289, + 394 + ], + "spans": [ + { + "bbox": [ + 62, + 300, + 289, + 394 + ], + "type": "image", + "image_path": "18882e36428500c73bdc3f359f9b9625fd632b24afcc7e67b34ec9a0f0cfe3d0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 100, + 395, + 126, + 403 + ], + "lines": [ + { + "bbox": [ + 100, + 395, + 126, + 403 + ], + "spans": [ + { + "bbox": [ + 100, + 395, + 126, + 403 + ], + "type": "text", + "content": "(a)Ours" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 209, + 395, + 249, + 403 + ], + "lines": [ + { + "bbox": [ + 209, + 395, + 249, + 403 + ], + "spans": [ + { + "bbox": [ + 209, + 395, + 249, + 403 + ], + "type": "text", + "content": "(b) SwinMM" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 65, + 413, + 159, + 495 + ], + "blocks": [ + { + "bbox": [ + 74, + 403, + 152, + 411 + ], + "lines": [ + { + "bbox": [ + 74, + 403, + 152, + 411 + ], + "spans": [ + { + "bbox": [ + 74, + 403, + 152, + 411 + ], + "type": "text", + "content": "Silhouette Score: 0.0381" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 413, + 159, + 495 + ], + "lines": [ + { + "bbox": [ + 65, + 413, + 159, + 495 + ], + "spans": [ + { + "bbox": [ + 65, + 413, + 159, + 495 + ], + "type": "image", + "image_path": "643730787308b356cf81c4b51a96985b5aaf4bc4f7aebf049fa084ea18bb77d4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 103, + 497, + 130, + 505 + ], + "lines": [ + { + "bbox": [ + 103, + 497, + 130, + 505 + ], + "spans": [ + { + "bbox": [ + 103, + 497, + 130, + 505 + ], + "type": "text", + "content": "(c) VoCo" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 76, + 506, + 157, + 513 + ], + "lines": [ + { + "bbox": [ + 76, + 506, + 157, + 513 + ], + "spans": [ + { + "bbox": [ + 76, + 506, + 157, + 513 + ], + "type": "text", + "content": "Silhouette Score: -0.1749" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 526, + 295, + 548 + ], + "lines": [ + { + "bbox": [ + 55, + 526, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 295, + 548 + ], + "type": "text", + "content": "Figure 5. The t-SNE visualization of features from 13 organs on the BTCV dataset and the silhouette scores of four methods." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 159, + 429, + 173, + 478 + ], + "blocks": [ + { + "bbox": [ + 159, + 429, + 173, + 478 + ], + "lines": [ + { + "bbox": [ + 159, + 429, + 173, + 478 + ], + "spans": [ + { + "bbox": [ + 159, + 429, + 173, + 478 + ], + "type": "image", + "image_path": "25e70f9bea063311b9f06ba2ea63f127e09dd2100a31b3bf58c155fba9ac99d7.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 175, + 300, + 287, + 394 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 300, + 287, + 394 + ], + "spans": [ + { + "bbox": [ + 175, + 300, + 287, + 394 + ], + "type": "image", + "image_path": "90c167113e1acebd79715334dcad76bd3b0c14b0102e4885ec62e82818e37415.jpg" + } + ] + } + ], + "index": 15, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 160, + 413, + 288, + 495 + ], + "blocks": [ + { + "bbox": [ + 190, + 403, + 268, + 411 + ], + "lines": [ + { + "bbox": [ + 190, + 403, + 268, + 411 + ], + "spans": [ + { + "bbox": [ + 190, + 403, + 268, + 411 + ], + "type": "text", + "content": "Silhouette Score:0.0027" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 160, + 413, + 288, + 495 + ], + "lines": [ + { + "bbox": [ + 160, + 413, + 288, + 495 + ], + "spans": [ + { + "bbox": [ + 160, + 413, + 288, + 495 + ], + "type": "image", + "image_path": "ce1d4e767b692bf86b41e6f3ea9cbd06495565090b217b871d11eed313101271.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 496, + 256, + 504 + ], + "lines": [ + { + "bbox": [ + 203, + 496, + 256, + 504 + ], + "spans": [ + { + "bbox": [ + 203, + 496, + 256, + 504 + ], + "type": "text", + "content": "(d) Swin-UNETR" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 187, + 505, + 269, + 512 + ], + "lines": [ + { + "bbox": [ + 187, + 505, + 269, + 512 + ], + "spans": [ + { + "bbox": [ + 187, + 505, + 269, + 512 + ], + "type": "text", + "content": "Silhouette Score: -0.0477" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": "scores [25], which evaluate the quality of clustering by measuring how similar a data point is to its cluster (consistency) compared to other clusters (discrepancy). Fig. 5 shows the clustering capabilities of 13 organs on the BTCV dataset. " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "S^2 DC" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": " shows excellent performance on organ discrimination and consistency, which gets the highest silhouette scores. As shown in Tab. 1, most organs that cluster well also exhibit strong segmentation performance, highlighting the importance of structure-specific consistency and discrepancy. Additionally, we visualize the feature maps of 4 SSL models on 3 modalities in Fig. 6. " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "S^2 DC" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": " performs semantic consistency in the background, tissues (CT), functional regions (MRI), and metabolic regions (PET)." + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 328, + 70, + 545, + 194 + ], + "blocks": [ + { + "bbox": [ + 328, + 70, + 545, + 194 + ], + "lines": [ + { + "bbox": [ + 328, + 70, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 328, + 70, + 545, + 194 + ], + "type": "image", + "image_path": "bcec55db2721be3b180e82fe43727cf66bde4f036859a4383e6ed642bde83ced.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 198, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 313, + 198, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 555, + 232 + ], + "type": "text", + "content": "Figure 6. We visualize the first principal components after applying PCA to token features [2] within volumes of CT, MRI, and PET. The figure displays the central slices of feature maps." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "table", + "bbox": [ + 353, + 236, + 517, + 298 + ], + "blocks": [ + { + "bbox": [ + 353, + 236, + 517, + 298 + ], + "lines": [ + { + "bbox": [ + 353, + 236, + 517, + 298 + ], + "spans": [ + { + "bbox": [ + 353, + 236, + 517, + 298 + ], + "type": "table", + "html": "
MethodAccuracy(%)
10%50%100%
Swin-UNETR [14]77.1592.2894.15
SwinMM [40]87.8793.5094.80
VoCo [41]86.7392.4394.60
\\( S^2DC \\)88.2993.8995.34
", + "image_path": "024fc64c789ad5f6328f1450f7bfc55d2f2295d8cfc5283838e54097abff34db.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 303, + 555, + 325 + ], + "lines": [ + { + "bbox": [ + 313, + 303, + 555, + 325 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 555, + 325 + ], + "type": "text", + "content": "Table 8. Experiment results on CC-CCII with various ratios of the training data. " + }, + { + "bbox": [ + 313, + 303, + 555, + 325 + ], + "type": "inline_equation", + "content": "{10}\\% ,{50}\\%" + }, + { + "bbox": [ + 313, + 303, + 555, + 325 + ], + "type": "text", + "content": " ,and " + }, + { + "bbox": [ + 313, + 303, + 555, + 325 + ], + "type": "inline_equation", + "content": "{100}\\%" + }, + { + "bbox": [ + 313, + 303, + 555, + 325 + ], + "type": "text", + "content": " represent ratios." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 331, + 555, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 402 + ], + "type": "text", + "content": "Varying ratios of the training dataset. We evaluate the performances of models that were trained with varying ratios of CC-CCII. From Tab. 8, " + }, + { + "bbox": [ + 313, + 331, + 555, + 402 + ], + "type": "inline_equation", + "content": "S^2 DC" + }, + { + "bbox": [ + 313, + 331, + 555, + 402 + ], + "type": "text", + "content": " outperforms other methods under different ratios of training data. Notably, " + }, + { + "bbox": [ + 313, + 331, + 555, + 402 + ], + "type": "inline_equation", + "content": "S^2 DC" + }, + { + "bbox": [ + 313, + 331, + 555, + 402 + ], + "type": "text", + "content": " achieves " + }, + { + "bbox": [ + 313, + 331, + 555, + 402 + ], + "type": "inline_equation", + "content": "88.29\\%" + }, + { + "bbox": [ + 313, + 331, + 555, + 402 + ], + "type": "text", + "content": " accuracy using only " + }, + { + "bbox": [ + 313, + 331, + 555, + 402 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 331, + 555, + 402 + ], + "type": "text", + "content": " of the data, which is effective with limited dataset sizes." + } + ] + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 351, + 411, + 518, + 543 + ], + "blocks": [ + { + "bbox": [ + 351, + 411, + 518, + 543 + ], + "lines": [ + { + "bbox": [ + 351, + 411, + 518, + 543 + ], + "spans": [ + { + "bbox": [ + 351, + 411, + 518, + 543 + ], + "type": "image", + "image_path": "6d0e7126fed46aaa38e3acf9dc6f10c532bfee21d41a9c3e30a79492738e0856.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 546, + 548, + 558 + ], + "lines": [ + { + "bbox": [ + 320, + 546, + 548, + 558 + ], + "spans": [ + { + "bbox": [ + 320, + 546, + 548, + 558 + ], + "type": "text", + "content": "Figure 7. The evolution of patch-to-structure correspondences." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 563, + 555, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 555, + 612 + ], + "type": "text", + "content": "Patch-to-structure correspondences evolution. We visualize feature map trends at different pretraining steps in Fig. 7. As training progresses, feature consistency improves, especially for small structures like bronchioles." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 621, + 388, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 388, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 388, + 633 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": "We propose " + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "inline_equation", + "content": "S^2 DC" + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": ", a medical image SSL framework that improves structure-aware semantic consistency through patch-to-patch and patch-to-structure steps. Experiments and visualizations show its superior performance. Future work could involve larger pre-training datasets or model generalization across modalities for broader applications." + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20264" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 159 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 159 + ], + "type": "text", + "content": "This paper was supported by the National Natural Science Foundation of China (Grant No.82394432 and 92249302), the Shanghai Municipal Science and Technology Major Project (Grant No.2023SHZDZX02). The computations in this research were performed using the CFFF platform of Fudan University." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 178, + 115, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 178, + 115, + 192 + ], + "spans": [ + { + "bbox": [ + 56, + 178, + 115, + 192 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 198, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 198, + 296, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 198, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 61, + 198, + 296, + 243 + ], + "type": "text", + "content": "[1] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicregl: Self-supervised learning of local visual features. Advances in Neural Information Processing Systems, 35:8799-8810, 2022. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 244, + 296, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 244, + 296, + 299 + ], + "spans": [ + { + "bbox": [ + 61, + 244, + 296, + 299 + ], + "type": "text", + "content": "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 300, + 295, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 300, + 295, + 344 + ], + "spans": [ + { + "bbox": [ + 61, + 300, + 295, + 344 + ], + "type": "text", + "content": "[3] Anadi Chaman and Ivan Dokmanic. Truly shift-invariant convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3773-3783, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 345, + 296, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 345, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 61, + 345, + 296, + 411 + ], + "type": "text", + "content": "[4] Gaoyu Chen, Sheng Liu, Wenxiang Ding, Li Lv, Chen Zhao, Fenghua Weng, Yong Long, Yunlong Zan, and Qiu Huang. A total-body ultra-low dose pet reconstruction method via image space shuffle u-net and body sampling. IEEE Transactions on Radiation and Plasma Medical Sciences, 2023. 1, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 412, + 296, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 412, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 412, + 296, + 456 + ], + "type": "text", + "content": "[5] Pengguang Chen, Shu Liu, and Jiaya Jia. Jigsaw clustering for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11526-11535, 2021. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 457, + 296, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 457, + 296, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 457, + 296, + 501 + ], + "type": "text", + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 502, + 296, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 296, + 557 + ], + "type": "text", + "content": "[7] Zekai Chen, Devansh Agarwal, Kshitij Aggarwal, Wiem Safta, Mariann Micsinai Balan, and Kevin Brown. Masked image modeling advances 3d medical image analysis. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1970-1980, 2023. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 558, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 558, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 558, + 296, + 612 + ], + "type": "text", + "content": "[8] Jiaqi Cui, Pinxian Zeng, Xinyi Zeng, Yuanyuan Xu, Peng Wang, Jiliu Zhou, Yan Wang, and Dinggang Shen. Prior knowledge-guided triple-domain transformer-gan for direct pet reconstruction from low-count sinograms. IEEE Transactions on Medical Imaging, 2024. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 613, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 613, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 62, + 613, + 296, + 679 + ], + "type": "text", + "content": "[9] Tugba Akinci D'Antonoli, Lucas K Berger, Ashraya K Indrakanti, Nathan Vishwanathan, Jakob Weiβ, Matthias Jung, Zeynep Berkarda, Alexander Rau, Marco Reisert, Thomas Küstner, et al. Totalsegmentator mri: Sequence-independent segmentation of 59 anatomical structures in mr images. arXiv preprint arXiv:2405.19492, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 681, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 714 + ], + "type": "text", + "content": "[10] Sergios Gatidis, Marcel Früh, Matthias Fabritius, Sijing Gu, Konstantin Nikolaou, Christian La Fougère, Jin Ye, Junjun He, Yige Peng, Lei Bi, et al. The autopet challenge:" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "towards fully automated lesion segmentation in oncologic pet/ct imaging. 2023. 1, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[11] Ali Hatamizadeh, Vishwesh Nath, Yucheng Tang, Dong Yang, Holger R Roth, and Daguang Xu. Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images. In International MICCAI brainlesion workshop, pages 272-284. Springer, 2021. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 151, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 215 + ], + "type": "text", + "content": "[12] Ali Hatamizadeh, Yucheng Tang, Vishwesh Nath, Dong Yang, Andriy Myronenko, Bennett Landman, Holger R Roth, and Daguang Xu. Unetr: Transformers for 3d medical image segmentation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 574-584, 2022. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 217, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 270 + ], + "type": "text", + "content": "[13] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 272, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 272, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 317, + 272, + 553, + 338 + ], + "type": "text", + "content": "[14] Yufan He, Vishwesh Nath, Dong Yang, Yucheng Tang, Andriy Myronenko, and Daguang Xu. Swinunetr-v2: Stronger swim transformers with stagewise convolutions for 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 416-426. Springer, 2023. 1, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 338, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 553, + 403 + ], + "type": "text", + "content": "[15] Yuting He, Guanyu Yang, Rongjun Ge, Yang Chen, Jean-Louis Coatieux, Boyu Wang, and Shuo Li. Geometric visual similarity learning in 3d medical image self-supervised pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9538-9547, 2023. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 404, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 404, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 317, + 404, + 553, + 459 + ], + "type": "text", + "content": "[16] Junjia Huang, Haofeng Li, Guanbin Li, and Xiang Wan. Attentive symmetric autoencoder for brain mri segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 203-213. Springer, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 460, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 525 + ], + "type": "text", + "content": "[17] Yankai Jiang, Mingze Sun, Heng Guo, Xiaoyu Bai, Ke Yan, Le Lu, and Minfeng Xu. Anatomical invariance modeling and semantic alignment for self-supervised learning in 3d medical image analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15859-15869, 2023. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 526, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 592 + ], + "type": "text", + "content": "[18] Pamela J LaMontagne, Tammie LS Benzinger, John C Morris, Sarah Keefe, Russ Hornbeck, Chengjie Xiong, Elizabeth Grant, Jason Hassenstab, Krista Moulder, Andrei G Vlassenko, et al. Oasis-3: longitudinal neuroimaging, clinical, and cognitive dataset for normal aging and Alzheimer disease. medrxiv, pages 2019–12, 2019. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 592, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 592, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 317, + 592, + 553, + 647 + ], + "type": "text", + "content": "[19] Bennett Landman, Zhoubing Xu, J Igelsias, Martin Styner, Thomas Langerak, and Arno Klein. Miccai multi-atlas labeling beyond the cranial vault-workshop and challenge. In Proc. MICCAI Multi-Atlas Labeling Beyond Cranial Vault—Workshop Challenge, page 12, 2015. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 647, + 553, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 691 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 691 + ], + "type": "text", + "content": "[20] Haofeng Li, Yiming Ouyang, and Xiang Wan. Self-supervised alignment learning for medical image segmentation. In 2024 IEEE International Symposium on Biomedical Imaging (ISBI), pages 1-5. IEEE, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "type": "text", + "content": "[21] Chengzhi Mao, Lingyu Zhang, Abhishek Vaibhav Joshi, Junfeng Yang, Hao Wang, and Carl Vondrick. Robust perception" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20265" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "through equivariance. In International Conference on Machine Learning, pages 23852-23870. PMLR, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 95, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 95, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 57, + 95, + 295, + 149 + ], + "type": "text", + "content": "[22] Susanne G Mueller, Michael W Weiner, Leon J Thal, Ronald C Petersen, Clifford Jack, William Jagust, John Q Trojanowski, Arthur W Toga, and Laurel Beckett. The alzheimer's disease neuroimaging initiative. Neuroimaging Clinics of North America, 15(4):869, 2005. 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 151, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 151, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 57, + 151, + 294, + 183 + ], + "type": "text", + "content": "[23] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "type": "text", + "content": "[24] Ignacio Rocco, Mircea Cimpoi, Relja Arandjelović, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Neighbourhood consensus networks. Advances in neural information processing systems, 31, 2018. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 228, + 294, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 228, + 294, + 260 + ], + "spans": [ + { + "bbox": [ + 57, + 228, + 294, + 260 + ], + "type": "text", + "content": "[25] Peter J Rousseeuw. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. Journal of computational and applied mathematics, 20:53-65, 1987. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 262, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 262, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 262, + 294, + 304 + ], + "type": "text", + "content": "[26] Ainkaran Santhirasekaram, Mathias Winkler, Andrea Rockall, and Ben Glocker. A geometric approach to robust medical image segmentation. Medical Image Analysis, 97: 103260, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 305, + 295, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 295, + 360 + ], + "type": "text", + "content": "[27] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938-4947, 2020. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 361, + 295, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 361, + 295, + 381 + ], + "spans": [ + { + "bbox": [ + 57, + 361, + 295, + 381 + ], + "type": "text", + "content": "[28] William F Sharpe. Mutual fund performance. The Journal of business, 39(1):119-138, 1966. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 383, + 294, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 448 + ], + "type": "text", + "content": "[29] Amber L Simpson, Michela Antonelli, Spyridon Bakas, Michel Bilello, Keyvan Farahani, Bram Van Ginneken, Annette Kopp-Schneider, Bennett A Landman, Geert Litjens, Bjoern Menze, et al. A large annotated medical image dataset for the development and evaluation of segmentation algorithms. arXiv preprint arXiv:1902.09063, 2019. 1, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 449, + 294, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 449, + 294, + 481 + ], + "spans": [ + { + "bbox": [ + 57, + 449, + 294, + 481 + ], + "type": "text", + "content": "[30] Richard Sinkhorn. A relationship between arbitrary positive matrices and doubly stochastic matrices. The annals of mathematical statistics, 35(2):876-879, 1964. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 483, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 483, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 483, + 294, + 536 + ], + "type": "text", + "content": "[31] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 537, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 590 + ], + "type": "text", + "content": "[32] Aiham Taleb, Winfried Loetzsch, Noel Danz, Julius Severin, Thomas Gaertner, Benjamin Bergner, and Christoph Lippert. 3d self-supervised methods for medical imaging. Advances in neural information processing systems, 33:18158-18172, 2020. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 593, + 294, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 593, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 593, + 294, + 647 + ], + "type": "text", + "content": "[33] Zhaorui Tan, Xi Yang, Tan Pan, Tianyi Liu, Chen Jiang, Xin Guo, Qiufeng Wang, Anh Nguyen, Yuan Qi, Kaizhu Huang, et al. Towards a universal 3d medical multi-modality generalization via learning personalized invariant representation. arXiv preprint arXiv:2411.06106, 2024. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 712 + ], + "type": "text", + "content": "[34] Yucheng Tang, Dong Yang, Wenqi Li, Holger R Roth, Bennett Landman, Daguang Xu, Vishwesh Nath, and Ali Hatamizadeh. Self-supervised pre-training of swim transformers for 3d medical image analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20730-20740, 2022. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 149 + ], + "type": "text", + "content": "[35] Xing Tao, Yuexiang Li, Wenhui Zhou, Kai Ma, and Yefeng Zheng. Revisiting rubik's cube: Self-supervised learning with volume-wise transformation for 3d medical image segmentation. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part IV 23, pages 238-248. Springer, 2020. 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 151, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 171 + ], + "type": "text", + "content": "[36] MTCAJ Thomas and A Thomas Joy. Elements of information theory. Wiley-Interscience, 2006. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 173, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 215 + ], + "type": "text", + "content": "[37] Yao-Hung Hubert Tsai, Yue Wu, Ruslan Salakhutdinov, and Louis-Philippe Morency. Self-supervised learning from a multi-view perspective. arXiv preprint arXiv:2006.05576, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 217, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 281 + ], + "type": "text", + "content": "[38] Chengyao Wang, Li Jiang, Xiaoyang Wu, Zhuotao Tian, Bohao Peng, Hengshuang Zhao, and Jiaya Jia. Groupcontrast: Semantic-aware self-supervised representation learning for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4917-4928, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 282, + 553, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 282, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 316, + 282, + 553, + 336 + ], + "type": "text", + "content": "[39] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3024-3033, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 338, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 553, + 414 + ], + "type": "text", + "content": "[40] Yiqing Wang, Zihan Li, Jieru Mei, Zihao Wei, Li Liu, Chen Wang, Shengtian Sang, Alan L Yuille, Cihang Xie, and Yuyin Zhou. Swinmm: masked multi-view with swin transformers for 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 486-496. Springer, 2023. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 415, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 470 + ], + "type": "text", + "content": "[41] Linshan Wu, Jiaxin Zhuang, and Hao Chen. Voco: A simple-yet-effective volume contrastive learning framework for 3d medical image analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22873-22882, 2024. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 472, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 553, + 525 + ], + "type": "text", + "content": "[42] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9653-9663, 2022. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 526, + 553, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 579 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 579 + ], + "type": "text", + "content": "[43] Zhangsihao Yang, Mengwei Ren, Kaize Ding, Guido Gerig, and Yalin Wang. Keypoint-augmented self-supervised learning for medical image segmentation with limited annotation. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 581, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 581, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 581, + 553, + 613 + ], + "type": "text", + "content": "[44] Xin Yi, Ekta Walia, and Paul Babyn. Generative adversarial network in medical imaging: A review. Medical image analysis, 58:101552, 2019. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 614, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 647 + ], + "type": "text", + "content": "[45] Chuyan Zhang, Hao Zheng, and Yun Gu. Dive into the details of self-supervised learning for medical image analysis. Medical Image Analysis, 89:102879, 2023. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "text", + "content": "[46] Kang Zhang, Xiaohong Liu, Jun Shen, Zhihuan Li, Ye Sang, Xingwang Wu, Yunfei Zha, Wenhua Liang, Chengdi Wang, Ke Wang, et al. Clinically applicable ai system for accurate diagnosis, quantitative measurements, and prognosis of Covid-19 pneumonia using computed tomography. Cell, 181 (6):1423-1433, 2020. 6" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "20266" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 351 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 115 + ], + "type": "text", + "content": "[47] Zhemin Zhang and Xun Gong. Positional label for self-supervised vision transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3516-3524, 2023. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 118, + 296, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 118, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 296, + 173 + ], + "type": "text", + "content": "[48] Hong-Yu Zhou, Chixiang Lu, Sibei Yang, Xiaoguang Han, and Yizhou Yu. *Preservational learning improves self-supervised medical image models by reconstructing diverse contexts*. In *Proceedings of the IEEE/CVF International Conference on Computer Vision*, pages 3499–3509, 2021. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 228 + ], + "type": "text", + "content": "[49] Hong-Yu Zhou, Chixiang Lu, Chaoqi Chen, Sibei Yang, and Yizhou Yu. Pcrlv2: A unified visual information preservation framework for self-supervised pre-training in medical image analysis. arXiv preprint arXiv:2301.00772, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 230, + 295, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 230, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 57, + 230, + 295, + 262 + ], + "type": "text", + "content": "[50] Zongwei Zhou, Vatsal Sodha, Jiaxuan Pang, Michael B Gotway, and Jianming Liang. Models genesis. Medical image analysis, 67:101840, 2021. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 263, + 295, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 263, + 295, + 307 + ], + "spans": [ + { + "bbox": [ + 57, + 263, + 295, + 307 + ], + "type": "text", + "content": "[51] Jiaxin Zhuang, Linshan Wu, Qiong Wang, Varut Vardhanabhuti, Lin Luo, and Hao Chen. Mim: Mask in mask self-supervised pre-training for 3d medical image analysis. arXiv preprint arXiv:2404.15580, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 308, + 295, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 308, + 295, + 351 + ], + "spans": [ + { + "bbox": [ + 57, + 308, + 295, + 351 + ], + "type": "text", + "content": "[52] Jia-Xin Zhuang, Luyang Luo, and Hao Chen. Advancing volumetric medical image segmentation via global-local masked autoencoder. arXiv preprint arXiv:2306.08913, 2023. 2, 6" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20267" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Structured Policy Optimization_ Enhance Large Vision-Language Model via Self-referenced Dialogue/5161fa7a-2cd7-4a41-959a-e1ddb02c88dc_content_list.json b/2025/Structured Policy Optimization_ Enhance Large Vision-Language Model via Self-referenced Dialogue/5161fa7a-2cd7-4a41-959a-e1ddb02c88dc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..abda38a8d3ab15ec860114bfe99abf0fef11830e --- /dev/null +++ b/2025/Structured Policy Optimization_ Enhance Large Vision-Language Model via Self-referenced Dialogue/5161fa7a-2cd7-4a41-959a-e1ddb02c88dc_content_list.json @@ -0,0 +1,1626 @@ +[ + { + "type": "text", + "text": "Structured Policy Optimization: Enhance Large Vision-Language Model via Self-Referenced Dialogue", + "text_level": 1, + "bbox": [ + 114, + 128, + 883, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Guohao Sun1*, Can Qin2, Yihao Feng2, Zeyuan Chen2, Ran Xu2, Sohail Dianat1, Majid Rabbani1, Raghuveer Rao3, and Zhiqiang Tao1*", + "bbox": [ + 197, + 202, + 805, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Rochester Institute of Technology \n $^{2}$ Salesforce AI Research $^{3}$ DEVCOM Army Research Laboratory", + "bbox": [ + 230, + 243, + 759, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 313, + 326, + 329 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Preference optimization algorithms typically enhance LLM response quality by leveraging human feedback on multiple answers given a fixed instruction. However, these methods often lack capturing the dynamic nature of conversational exchanges. For large vision-language models (LVLMs), direct preference optimization (DPO) can over-emphasize linguistic nuances while overlooking visual context. To address this challenge, we introduce structured policy optimization (SPO) - a novel preference optimization method that simultaneously aligns preference instructions, responses, and dialogue interactions to improve multi-modal understanding and reasoning capabilities. The efficacy of SPO is attributed to one key design: treating the questioning and answering as a sequential action and binding them through a trajectory reward. This reward formulation better aligns with real-world dialogue studies and eliminates the need for fixed instructions. We evaluate our models on interleaved benchmarks, including image, multi-image, and video-based understanding and reasoning tasks. Experimental results show that the proposed SPO finetuning LVLM with multi-modal preference data can align with human preference more efficiently than DPO. The code is available at https://github.com/heliossun/Structure-Policy-Optimization.", + "bbox": [ + 89, + 347, + 483, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 742, + 220, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reinforcement learning from human feedback (RLHF) [8] has demonstrated remarkable success in aligning large language/vision-language models (LLMs/LVLMs) with human preferences, enabling models to adapt their behavior based on success or failure in increasingly complex tasks. Typically, RLHF trains a reward model and optimizes the", + "bbox": [ + 89, + 768, + 482, + 861 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/28ef7d25bcde55ac0d4bf2ed9baa241070cc4861884b8828d0b52146a6288640.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 315, + 767, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ad1d1f1bef0d013fb89038fb4681c4eea37330537c2d1a9c71e75d4d85654f1d.jpg", + "image_caption": [ + "Figure 1. Comparison between DPO and SPO in optimizing an LVLM policy. DPO employs a single reward that optimizes the preferred answer given the fixed visual & question input, $p(y_w \\succ y_l|x, q)$ , which could over-emphasize language nuance and overlook visual clues [38]. By contrast, SPO introduces a structured reward by $p(\\mathcal{Y}_w \\succ \\mathcal{Y}_l|x)$ , conditioning on the visual dialogue, to optimize the preferred trajectories spanning over questions and answers." + ], + "image_footnote": [], + "bbox": [ + 522, + 369, + 756, + 498 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e9a5b624afff8f00463fd23650074968d0cf4410631b8b4ac0882894f1177bed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 314, + 893, + 380 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4a691cd52fd51a7243c09ab7c443d480f16e0c59e0a8f8621a230b62240094fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 776, + 393, + 893, + 496 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "LLM/LVLM policy to maximize the reward, which is applied in both language and vision-language (VL) domains.", + "bbox": [ + 511, + 654, + 906, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, the offline preference optimization methods, such as DPO [29] and its variants [12, 17, 30, 33], have proposed various reward objectives to satisfy different human expectations. Owing to its simplicity and stability, DPO has been widely used in multi-modal preference optimization [11, 19, 22, 38] by aligning LVLM to human-preferred responses conditioned on fixed visual and text instructions (see Fig. 1). However, unlike purely text dialogue, the visual input can lead to far more enriched information interactions, which require varying degrees of visual dependence [13, 25, 37]. Directly applying DPO to multi-modal preference data (e.g., images/videos and texts) will over-emphasize language nuance and overlook the visual condition [38]. Specifically, DPO applies a single, un", + "bbox": [ + 511, + 689, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors: Guohao Sun (gs4288@rit.edu) and Zhiqiang Tao (zhiqiang.tao@rit.edu)", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "741", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/62500da39ba19fe7e3f00edccc19d9cdc47c01c3e09e2d88141948b140e5de27.jpg", + "image_caption": [ + "(a) Question vs Answer." + ], + "image_footnote": [], + "bbox": [ + 102, + 90, + 282, + 175 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ac2b3fef8156fe58734ef89f170e5d8052903d30085bb4c5149f4acce882455d.jpg", + "image_caption": [ + "(b) Score distribution.", + "Figure 2. Scoring visual dialogue with GPT-4o as judge. (a) Questions with higher visual dependency (x-axis, scoring from 4 to 10, $\\uparrow$ ) tend to produce answers with higher quality (y-axis, scoring from 0 to 9, $\\uparrow$ ). (b) Question and answer score distributions." + ], + "image_footnote": [], + "bbox": [ + 287, + 90, + 473, + 176 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "structured reward function to maximize the gap between the preferred and dispreferred answer upon the same input sequence; however, this gap might be due to the lack of visual understanding [5], which is generally ignored by existing DPO methods, especially for LVLMs.", + "bbox": [ + 89, + 292, + 482, + 367 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by the dialogue studies [9, 46], we empirically observe that asking a good visual-related question can turn passive learning into active exploration, allowing for deeper visual understanding and broader knowledge exploration before answering. For instance, as shown in Fig. 1, people prefer the question \"What emotions are conveyed through the shots in the video?\" to \"Which city did the skateboarder go to in the video?\" since the former is more visually dependent and the answer will cover more visual details. To quantitatively validate our observation, we apply GPT-4o to evaluate the visual relevance (scoring from 4 to 10) of questions and score the quality of answers (higher is better) in our curated preference data based on the instruction dataset in [18]. Fig. 2a shows that human-preferred questions with higher visual relevance tend to produce responses with higher quality, which is further demonstrated by the consistent question/answer score distributions in Fig. 2b. We introduce this correlation between visual questions and answer quality as a preference dialogue interaction. By aligning the LVLM policy to generate questions and answers with human preference, we build a more in-depth alignment between vision and dialogues, enabling a better VL understanding for broad tasks.", + "bbox": [ + 89, + 369, + 482, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this study, we propose Structured Policy Optimization (SPO), a novel structured DPO method designed to enhance general vision-language understanding and reasoning. As shown in Fig. 1, SPO utilizes a binary tree structure to track visual dialogue changes, alleviating the limitation of using fixed questions in DPO. Unlike DPO, which only optimizes preference at leaf nodes, the proposed SPO optimizes preferences across different trajectories among sub-trees and develops a new structured reward model by leveraging trajectories to capture the dialogue interaction between the visual input, questions, and answers. Additionally, SPO controls the deviation of question reward and an-", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "swer reward by imposing importance weights on each step within the trajectory, which is substantially different from a single-level joint reward optimization [3]. To realize SPO training and increase data diversity, we utilize supervised fine-turning (SFT) to instruct an LVLM to self-generate the preference question-answer pairs, mitigating the distribution shift between the policy model and the preference optimization process [29]. We summarize the contributions of this work as follows.", + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- For the first time in multi-modal preference optimization, we introduce SPO, a method designed for optimizing structured reward for aligning the policy model to human preferences over visual dialogues.", + "- We present high-quality multi-modal preference data consisting of scoring over visual questions and answers. This dataset reflects the discrepancy between chosen and rejected dialogue, bringing insights into real-world conversational exchange and dialogue studies.", + "- Extensive experiments show that SPO aligns with human preference effectively and overcomes the low visual dependency issue in DPO, leading to better performance on the visual chat benchmarks with free-form instructions." + ], + "bbox": [ + 513, + 227, + 903, + 421 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 436, + 663, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generative Large Vision-Language Model. Recent general-purpose multi-modal models (LVLMs) have shown impressive capability. The commercial proprietary LVLMs, such as GPT-4V [26], GPT-4o [27], and Gemini [36] exhibit excellent generalize ability in various domains [48] such as design, business, science, medical, and science etc. In the open research community, the evolutionary trajectory of LVLMs has also led to subsequent advancements in different scenarios. The pioneering work of LLaVA [24] improves the multi-modal reasoning ability through visual instruction tuning. [4, 10, 24, 34, 35, 55] pushing the performance limits in single-image scenarios. [16, 21] explores multi-image tasks. While video LVLMs [23, 39, 47, 53] aim to alleviate significant challenges due to the long context issue. In contrast to these single-modality models, the latest LLaVA models [18, 20] demonstrate emerging capabilities across all scenarios after training on interleaved data. Following this path, this work aims to improve the general ability of multiple modalities by primarily focusing on improving the model's reasoning ability.", + "bbox": [ + 511, + 462, + 906, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-generated Synthetic Data. Instruction tuning is widely used in supervised fine-tuning (SFT) to improve general-purpose reasoning. In recent research, synthetic instructional data has become increasingly popular and considered a proxy for human data. This approach primarily leverages the ability of GPT family, [19, 24, 52] utilizes GPT-4 [1] to transform image description into conversations, [4, 18, 20] leverages GPT-4V [26] to process the multi-modal input and generate question-answer pairs.", + "bbox": [ + 511, + 765, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "742", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1abbdc8b960da4819ee8a025be615452d45451d361a9de3b173279516748f4f5.jpg", + "image_caption": [ + "Stage 1: Reasoning and Learning to Question", + "Stage2: Preference revelation", + "Figure 3. Illustration of LVLM model architecture and training pipeline in the proposed SPO. In Stage 1, we implement supervised fine-tuning through learning to question [35], resulting in an SFT model that enables better visual reasoning and self-generated questions. In Stage 2, we first prompt the SFT model to generate visual dialogues labeled by GPT-4o with human preference. Then, we apply the proposed SPO approach to optimize the LVLM policy through a structured reward, leading to better VL understanding." + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 903, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, large-scale synthetic data generation comes with unexpected API costs. To overcome this, [34] train the model with a visual-questioning objective function [35] and generate the question-answer pairs efficiently. While most works mentioned above consider one single optimization procedure, self-play fine-tuning methods [6, 31, 32, 42] generate new synthetic data within the training loop. This work focuses primarily on single-turn optimization and avoids iterative processes.", + "bbox": [ + 88, + 386, + 483, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preference Optimization Objectives. Reinforcement Learning from Human Feedback (RLHF) is a technique used to align large language models with human preferences and values [2, 8, 28, 56]. To eliminate the complexity and challenges of optimizing online preference algorithms, Direct Preference Optimization (DPO) [29] is a prominent example. To extend two-instance classification, ranking objectives enable comparisons across multiple instances [12, 17, 30, 33]. Recent efforts in multi-modal preference learning mainly focus on preference data design [11, 19, 22, 38], while utilizing the formal DPO objective function. This work proposes a novel structured DPO by maximizing the structured rewards to improve vision-language alignment.", + "bbox": [ + 89, + 523, + 483, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. SPO: Structured Policy Optimization", + "text_level": 1, + "bbox": [ + 89, + 752, + 428, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preference optimization uses explicit feedback (e.g., human ratings, comparisons, or rewards) to fine-tune the SFT model, making it more responsive to real-world user needs. As shown in Fig. 3, this work employs a two-stage training strategy: 1) Utilizing supervised fine-tuning (SFT) to train a SFT model by reasoning and learning to question [35], which is able to generate diverse visual dialogues that align with the training data's specific patterns and content. 2)", + "bbox": [ + 89, + 779, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fine-tune the LVML policy (given the SFT model as a reference) using the proposed SPO approach by maximizing a novel structured reward function. In the following, we will first provide preliminaries about preference optimization and then elaborate on our proposed SPO approach.", + "bbox": [ + 511, + 386, + 906, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Backgrounds: RLHF and DPO", + "text_level": 1, + "bbox": [ + 511, + 468, + 790, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\mathbf{RLHF}[8]$ first trains a reward model $r_{\\phi}(x,y)$ parametrized by $\\phi$ to predict the human preference distribution $p$ over two responses $y_{1}$ and $y_{2}$ (e.g., answers given by an LVLM) with condition $x$ , simulated by the Bradely-Terry (BT) model as", + "bbox": [ + 511, + 489, + 906, + 551 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} p \\left(y _ {1} \\succ y _ {2} \\mid x\\right) = \\frac {\\exp \\left(r _ {\\phi} \\left(x , y _ {1}\\right)\\right)}{\\exp \\left(r _ {\\phi} \\left(x , y _ {1}\\right)\\right) + \\exp \\left(r _ {\\phi} \\left(x , y _ {2}\\right)\\right)} \\tag {1} \\\\ = \\sigma \\left(r _ {\\phi} (x, y _ {1}) - r _ {\\phi} (x, y _ {2})\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 556, + 903, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $x$ denotes visual input and questions (instructions) in this study. After training, the reward model $r_{\\phi}$ provides feedback for each response generated by policy model $\\pi_{\\theta}$ , which is optimized by the following objective function:", + "bbox": [ + 511, + 613, + 905, + 674 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\max _ {\\pi_ {\\theta}} \\mathbb {E} _ {x \\sim D, y \\sim \\pi_ {\\theta} (y | x)} [ r _ {\\phi} (x, y) ] \\tag {2} \\\\ - \\beta \\mathbb {D} _ {K L} \\left[ \\pi_ {\\theta} (y | x) \\| \\pi_ {r e f} (y | x) \\right], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 680, + 903, + 722 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{D} = \\{x^{(i)},y_w^{(i)},y_l^{(i)}\\}_{i = 1}^N$ is the preference dataset, $\\pi_{\\theta}$ is the policy model that need to be updated, and $\\pi_{ref}$ is a reference model (e.g., an SFT model) that avoids overshifting to preferences. $y_{w}$ and $y_{l}$ refer to preferred and dispreferred answers such as $y_{1}$ and $y_{2}$ in Eq. (1).", + "bbox": [ + 511, + 729, + 905, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DPO [29] reformulates the objective function of RLHF by changing from loss over reward function to optimal policy. Previous work [29] derives the optimal policy $\\pi_{\\theta}$ of Eq. (2) as", + "bbox": [ + 511, + 806, + 906, + 868 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\pi_ {\\theta} (y | x) = \\frac {1}{Z (x)} \\pi_ {r e f} (y | x) \\exp \\left(\\frac {1}{\\beta} r (x, y)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 872, + 903, + 904 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "743", + "bbox": [ + 485, + 944, + 511, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $Z(x) = \\sum_{y} \\pi_{\\text{ref}}(y|x) \\exp \\left( \\frac{1}{\\beta} r(x,y) \\right)$ is the marginal likelihood over all possible model responses. To simplify it, DPO reforms Eq. (3) as", + "bbox": [ + 89, + 90, + 485, + 137 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nr (x, y) = \\beta \\left(\\log \\frac {\\pi_ {\\theta} (y \\mid x)}{\\pi_ {r e f} (y \\mid x)} + \\log Z (x)\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 143, + 482, + 179 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The BT model for preference estimation depends only on the difference of rewards between two responses $y_{1}$ and $y_{2}$ [29]. By substituting Eq. (4) into Eq. (1), DPO expresses the probability of human preference data in terms of the optimal policy and formulating the objective function for a parameterized policy $\\pi_{\\theta}$ as", + "bbox": [ + 89, + 184, + 483, + 270 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n- \\mathbb {E} _ {(x, y _ {w}, y _ {l}) \\sim \\mathcal {D}} \\left[ \\log \\sigma \\left(\\beta \\log \\frac {\\pi_ {\\theta} (y _ {w} | x)}{\\pi_ {r e f} (y _ {w} | x)} - \\beta \\log \\frac {\\pi_ {\\theta} (y _ {l} | x)}{\\pi_ {r e f} (y _ {l} | x)}\\right) \\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 275, + 478, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sigma$ represents the sigmoid function and $\\beta$ is a scalable value to control the deviation from the reference model.", + "bbox": [ + 89, + 311, + 482, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motivation. Nevertheless, DPO methods mainly focus on optimizing the LLM policy by comparing multiple generations $y$ based on a fixed input $x$ , falling short in capturing the dynamics of visual dialogue interactions. Specifically, high visual-dependent instructions require better visual understanding and reasoning ability to respond than easy-to-answer questions [9, 46]. Such discrepancy leads to responses with different amounts of visual clues. To quantitatively assess the visual dependency of each dialogue interaction, we utilized GPT-4o to review visual dialogues according to a set of manually designed rules. As shown in Fig. 2b, instructions with higher visual dependency tend to result in responses with more relevance to the visual input, leading to a preference dialogue interaction.", + "bbox": [ + 88, + 340, + 482, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Empirical evidence [35] suggests that training the policy to align with preference interaction can enhance visual understanding by maximizing the likelihood of preferred questions and answers. To this end, we propose Structured Policy Optimization (SPO). In SPO, we consider question and answer as two independent actions, each leading to a different state, simulating the conversational exchange. As shown in Fig. 1, the policy can be naturally modeled as a binary tree with depth 2. Unlike DPO, which only optimizes preference at the leaf nodes, SPO optimizes preferences across the entire sub-tree.", + "bbox": [ + 88, + 551, + 482, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. SPO Aligns With Dialogue Interaction", + "text_level": 1, + "bbox": [ + 89, + 727, + 421, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Preference Data Curation. To implement SPO, we first prompt the SFT model (given in Stage 1 in Fig. 3) to generate visual instruction-answer pairs, aiming to mitigate the distribution shift between SFT and the preference optimization step [3, 29]. Notably, similar to DPO, the SFT model also works as a reference policy $\\pi_{\\mathrm{ref}}$ in SPO. However, the difference lies in our $\\pi_{\\mathrm{ref}}$ being capable of self-generating questions by implementing SFT through learning to question [35]. As shown in Fig. 4, we represent different visual dialogues as trajectories in a tree structure - the", + "bbox": [ + 88, + 750, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1ba21977451391ad8cea23ef9581256a7a7ecb2d2635a6040255d715911e0625.jpg", + "image_caption": [ + "Figure 4. Illustration of a tree structure that represents different visual dialogues, which are denoted as trajectory, e.g., $\\tau_{1} = (x, q_{w}, y_{w}^{1})$ . For preference labeling, we utilize GPT-4o as a judge to perform pair-wise scoring over trajectory pairs." + ], + "image_footnote": [], + "bbox": [ + 596, + 89, + 823, + 218 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "input $x$ denotes the root, questions occupy the first level, and answers are leaf nodes - and collect preference data $\\mathcal{D} = \\{x^{(i)},\\mathcal{Y}_w^{(i)},\\mathcal{Y}_l^{(i)}\\}_{i = 1}^N$ by sub-trees*, where the preferred sub-tree $\\mathcal{V}_w = (q_w,y_w^1,y_l^1)$ and dispreferred subtree $\\mathcal{V}_l = (q_l,y_w^2,y_l^2)$ . We sample questions $q_{w},q_{l}\\sim \\pi_{\\mathrm{ref}}$ for each visual input $x$ , followed by answers $y_w^1,y_l^1\\sim$ $\\pi_{\\mathrm{ref}}(\\cdot |x,q_w))]$ and $y_w^2,y_l^2\\sim \\pi_{\\mathrm{ref}}(\\cdot |x,q_l))]$ , where the superscripts 1 and 2 indicate conditioning on $q_{w} / q_{l}$ , respectively.", + "bbox": [ + 511, + 310, + 903, + 433 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Empirically, the preference label accuracy is critical to increasing training stability. Following [45], AI-as-judge can offer scalable solutions to reduce human labor in complex task assessments and generate effective reward signals in reinforcement learning. Thus, to evaluate the synthetic data quality, we compared the distribution of 200 instruction-answer scores judged by GPT-4o [27] and an open-sourced LVLM (LLaVA-critic [45]). Fig. 5 shows that GPT-4o tends to judge with a higher standard than LLaVA-critic as its mean score is lower. We find that GPT-4o can provide scores with higher diversity than LLaVA-critic, acting more like a human annotator. Therefore, this work utilizes GPT-4o for all the preference labeling, and we propose a simple rule to filter out low-quality samples after collecting all the reviews. Specifically, we first remove samples with scores lower than a threshold, then remove the samples with repeated questions or answers, such as $q_{w} == q_{l}$ or $y_{w}^{1} == y_{l}^{1}$ . By this means, our collected preference data maintains high diversity and reliability.", + "bbox": [ + 511, + 433, + 906, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Structured Reward Formulation. Given the initial state $x$ , the preferred sub-tree $\\mathcal{V}_w$ should has higher reward since we assume the preference instruction dominates the dialogue interaction. We apply the same Bradley-Terry probability model to ensure that the reward of the preferred sub-tree $\\mathcal{V}_w$ exceeds the reward of the dispreferred sub-tree $\\mathcal{V}_l$ by", + "bbox": [ + 511, + 719, + 908, + 825 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\mathcal {Y} _ {w} \\succ \\mathcal {Y} _ {l} \\mid x\\right) = \\sigma \\left(r \\left(x, \\mathcal {Y} _ {w}\\right) - r \\left(x, \\mathcal {Y} _ {l}\\right)\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 835, + 903, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, directly deriving $r(x,\\mathcal{Y})$ into Eq. (2) will intro", + "bbox": [ + 511, + 862, + 903, + 878 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "*For convenience, we omit the superscript $i$ when no confusion occurs.", + "bbox": [ + 531, + 887, + 903, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "744", + "bbox": [ + 485, + 944, + 511, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "duce a complex term $Z(x)$ (please refer to the supplementary material for a complete derivation). To overcome this issue, we decompose sub-trees $\\mathcal{V}$ to trajectories $\\tau$ and show how it connects to state-action trajectories in RLHF [8]. Let $\\tau = (s_1, a_1, s_2, a_2, \\ldots, s_T, a_T)$ be a trajectory sampled from the policy $\\pi_{\\theta}$ , where $s, a, T$ denote states, actions, and total steps. The total reward of a trajectory is defined as $r(\\tau) = \\sum_{t=1}^{T} r(s_t, a_t)$ . Thus, following this reward formulation, we can define a trajectory to represent one visual dialogue by setting $T = 2$ , where the initial state and action are input $x$ and questioning $s_1 = x, a_1 = q$ , and the following state and action are $s_2 = [x, q], a_2 = y$ , e.g., $\\tau_1 = (x, q_w, [x, q_w], y_w^1), \\tau_2 = (x, q_w, [x, q_w], y_l^1)$ , etc. Finally, we implement Eq. (5) by a trajectory reward as", + "bbox": [ + 89, + 90, + 483, + 303 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr (x, \\mathcal {Y}) \\sim r (\\tau) = \\lambda \\cdot r (x, q) + \\delta \\cdot r ([ x, q ], y), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 314, + 482, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ and $\\delta$ are hyperparameters used to balance the question and answer rewards. By setting $\\lambda = \\delta$ , our objective function could be equivalent to $\\sigma(r(\\tau_1) - r(\\tau_2))$ , direct preference optimization (DPO) over two trajectory (i.e., question-answer pairs), rather than on how it is split across the binary tree. In the next section, we derive the closed-form expression of $r(x, q)$ and $r([x, q], y)$ .", + "bbox": [ + 89, + 342, + 483, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. The SPO Objective Function", + "text_level": 1, + "bbox": [ + 89, + 457, + 348, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following Eq. (2) and our trajectory-based reward model, the general RL objective of SPO is given as", + "bbox": [ + 89, + 479, + 482, + 510 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nJ (\\pi_ {\\theta}) = \\mathbb {E} _ {\\tau \\sim \\pi_ {\\theta}} [ r (\\tau) ] - \\beta \\mathrm {K L} [ \\pi_ {\\theta} (\\tau) | | \\pi_ {\\text {r e f}} (\\tau) ]. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 521, + 482, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By substituting Eq. (6) into Eq. (7), we have $J(\\pi_{\\theta})$ as", + "bbox": [ + 89, + 547, + 447, + 565 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbb {E} _ {(x, q, y) \\sim \\tau} [ \\lambda r (x, q) + \\delta r ([ x, q ], y) ] - \\beta (\\mathrm {K L} [ \\pi_ {\\theta} (\\cdot | x) | | \\pi_ {\\operatorname {r e f}} (\\cdot | x) ]) \\\\ \\left. + \\mathrm {K L} \\left[ \\pi_ {\\theta} (\\cdot | [ x, q ]) \\mid \\mid \\pi_ {\\text {r e f}} (\\cdot | [ x, q ])) \\right]\\right), \\\\ = \\mathbb {E} _ {(x, q) \\sim \\tau} [ \\lambda r (x, q) ] - \\beta \\mathrm {K L} [ \\pi_ {\\theta} (\\cdot | x) | | \\pi_ {\\operatorname {r e f}} (\\cdot | x)) ] \\\\ + \\mathbb {E} _ {(x, q, y) \\sim \\tau} [ \\delta r ([ x, q ], y) ] - \\beta \\mathrm {K L} [ \\pi_ {\\theta} (\\cdot | [ x, q ]) | | \\pi_ {\\text {r e f}} (\\cdot | [ x, q ]) ] , \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 571, + 485, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $q\\sim \\pi_{\\theta}(\\cdot |x)$ and $y\\sim \\pi_{\\theta}(\\cdot |[q,x])$ . Following [41], the optimal solution of $J(\\pi_{\\theta})$ can be derived as", + "bbox": [ + 89, + 650, + 482, + 681 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\pi_ {\\theta} (q | x) = \\frac {1}{Z (x)} \\pi_ {\\mathrm {r e f}} (q | x) \\exp (\\frac {\\lambda}{\\beta} r (x, q)),\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 691, + 400, + 724 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\pi_ {\\theta} (y | [ x, q ]) = \\frac {1}{Z ([ x , q ])} \\pi_ {\\mathrm {r e f}} (y | [ x, q ]) \\exp (\\frac {\\delta}{\\beta} r ([ x, q ], y)).\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 726, + 473, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Thus, we can obtain the reward functions as", + "bbox": [ + 89, + 768, + 382, + 781 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr (x, q) = \\frac {\\beta}{\\lambda} \\log \\frac {\\pi_ {\\theta} (q | x)}{\\pi_ {\\operatorname {r e f}} (q | x)} + \\frac {\\beta}{\\lambda} \\log Z (x), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 792, + 482, + 825 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr ([ x, q ], y) = \\frac {\\beta}{\\delta} \\log \\frac {\\pi_ {\\theta} (y | [ x , q ])}{\\pi_ {\\mathrm {r e f}} (y | [ x , q ])} + \\frac {\\beta}{\\delta} \\log Z ([ x, q ]).\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 827, + 460, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Now, given the definition of each step reward, the total trajectory reward defined in Eq. (6) in terms of the optimal", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3a8e025debf3d9dbf9038d3c4f82ff12d52958bdc234868b3ed24a46d566e4e8.jpg", + "image_caption": [ + "Figure 5. The score distribution of questions and answers, and labeling agreement with 3 LVLMS as judges. GPT-4o provides lower mean score and higher diversity than LLaVA-critic, acting more like a human annotator." + ], + "image_footnote": [], + "bbox": [ + 516, + 90, + 903, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "policy can be directly substituted into the preference model in Eq. (1), where the $Z(x)$ term will be canceled just as the original DPO derivation. Assuming $\\tau_w = (x, q_1, [x, q_1], y_1)$ and $\\tau_l = (x, q_2, [x, q_2], y_2)$ , the preference probability of preferred $\\tau_w$ over dispreferred $\\tau_l$ can be defined as", + "bbox": [ + 511, + 364, + 906, + 439 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} p \\left(\\tau_ {w} \\succ \\tau_ {l}\\right) := \\sigma \\left(\\frac {\\beta}{\\lambda} \\log \\frac {\\pi_ {\\theta} \\left(q _ {1} \\mid x\\right)}{\\pi_ {\\operatorname {r e f}} \\left(q _ {1} \\mid x\\right)} + \\frac {\\beta}{\\delta} \\log \\frac {\\pi_ {\\theta} \\left(y _ {1} \\mid [ x , q _ {1} ]\\right)}{\\pi_ {\\operatorname {r e f}} \\left(y _ {1} \\mid [ x , q _ {1} ]\\right)} \\right. \\\\ \\left. - \\frac {\\beta}{\\lambda} \\log \\frac {\\pi_ {\\theta} \\left(q _ {2} \\mid x\\right)}{\\pi_ {\\text {r e f}} \\left(q _ {2} \\mid x\\right)} - \\frac {\\beta}{\\delta} \\log \\frac {\\pi_ {\\theta} \\left(y _ {2} \\mid [ x , q _ {2} ]\\right)}{\\pi_ {\\text {r e f}} \\left(y _ {2} \\mid [ x , q _ {2} ]\\right)}\\right). \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 455, + 908, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "More details about the derivation from Eq. (8) to Eq. (10) could be found in the supplementary.", + "bbox": [ + 511, + 537, + 905, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 4, we reformulate the preference data based on trajectories as $\\mathcal{D} = \\{\\tau_1^{(i)},\\tau_2^{(i)},\\tau_3^{(i)},\\tau_4^{(i)}\\}_{i = 1}^N$ . By substituting Eq. (8) into Eq. (9), the loss function of classifying between two trajectories is given as", + "bbox": [ + 511, + 571, + 905, + 633 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} \\left(\\tau_ {w}, \\tau_ {l}\\right) = - \\mathbb {E} _ {\\left(\\tau_ {w}, \\tau_ {l}\\right) \\sim \\mathcal {D}} \\left[ \\log \\sigma \\left(r \\left(\\tau_ {w}\\right) - r \\left(\\tau_ {l}\\right)\\right) \\right] \\tag {10} \\\\ \\leq \\mathbb {E} _ {(\\tau_ {w}, \\tau_ {l}) \\sim \\mathcal {D}} [ - \\log \\sigma ((r (x, q _ {1}) - r (x, q _ {2}))) ] \\\\ + \\mathbb {E} _ {\\left(\\tau_ {w}, \\tau_ {t}\\right) \\sim \\mathcal {D}} [ - \\log \\sigma \\left(\\left(r ([ x, q _ {1} ], y _ {1}) - r ([ x, q _ {2} ], y _ {2}))\\right) \\right]. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 652, + 916, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Eventually, we define the preference probability between sub-tree $\\mathcal{V}_w$ and $\\mathcal{V}_l$ as a joint of multiple trajectory-based preference probability, improving the comprehension of policy distribution. The final SPO loss is formulated as", + "bbox": [ + 511, + 726, + 906, + 786 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S P O} = \\mathcal {L} _ {\\left(\\mathcal {Y} _ {w}, \\mathcal {Y} _ {i}\\right)} := \\sum_ {m, n} \\mathcal {L} \\left(\\tau_ {m}, \\tau_ {n}\\right): m < n, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 804, + 903, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}(\\tau_m,\\tau_n)$ is given in Eq. (10), and $m\\in \\Phi_w = \\{1,3\\}$ , $n\\in \\Phi_l = \\{2,4\\}$ are the index of preferred and dispreferred trajectories.", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "745", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ad659c9df4e1cd1e3456f8b4375c9bf3ae16dad412d2dc49fbd3af1206e0da49.jpg", + "table_caption": [ + "Table 1. Evaluation results of our model trained with different preference optimization methods based on the SFT model on a scale of 0.5B and 7B. We report accuracy on all the benchmarks except MME. The best results are bold, and the second-best results are underlined." + ], + "table_footnote": [], + "table_body": "
ModelSingle-Image
MMStarMMMUMMEL-WilderLLaVA-WMMVetRealWorldQAPOPEMMBench
testvaltestsmalltesttesttesttestMulti-discip
GPT-4V [26]57.156.8140981.098.049.961.4-75.0
GPT-4o [27]-69.1-85.9106.176.258.6--
Gemini-1.5-Pro [36]-62.2----70.4--
InternVL-2-8B [7]59.449.3221062.584.560.064.4-81.7
IXC-2.5-7B [50]59.942.9222961.478.151.767.8-82.2
LLaVA-OV-72B [18]66.156.8168272.093.563.771.9-85.9
LLaVA-OV-0.5B [18]37.531.4123855.074.229.155.688.052.1
SFT-0.5B [35]40.731.9124659.472.826.755.788.053.1
SFT-0.5B + DPO [29]40.932.1125058.573.525.856.488.154.0
SFT-0.5B + SPO (ours)41.532.8125161.374.130.056.988.254.3
LLaVA-OV-7B [18]61.748.8158067.890.757.566.389.280.8
SFT-7B [35]59.648.0156569.590.557.666.889.079.7
SFT-7B + DPO [29]61.949.0157969.991.258.267.289.180.6
SFT-7B + SPO (ours)62.149.2158570.292.358.867.389.381.0
", + "bbox": [ + 140, + 128, + 859, + 373 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0de40dfe89f906415724b99fce5bfbb44ff4b2980b683f96401b5fad223e261d.jpg", + "table_caption": [ + "Table 2. Evaluation results of different methods on multi-image and video benchmarks. #Frames is the number of frames used for video encoding. We report accuracy on all the benchmarks. The best results are bold, and the second-best results are underlined." + ], + "table_footnote": [], + "table_body": "
Model#FramesMulti-ImageEgoSchemaVideo
\\(LLaVA^I\\)MuirBenchMLVUNExT-QALVid-BenchVideoMME
out-domainout-domaintestm-avgmcvalw/o subs
GPT-4V [26]1060.362.3-49.2-61.359.9
GPT-4o [27]384-68.0-64.6-66.771.9
Gemini-1.5-Pro [36]1/s-49.472.2--64.075.0
LongVA-7B [51]128---56.368.3-52.6
IXC-2.5-7B [50]64---37.371.0-55.8
LLaVA-Video-7B [53]64--57.370.883.258.263.3
LLaVA-Video-72B [53]64--65.674.485.461.970.5
LLaVA-OV-72B [18]3279.954.862.068.080.261.366.3
LLaVA-OV-0.5B [18]3233.325.526.850.357.245.844.0
SFT-0.5B [35]3232.726.128.851.759.044.644.3
SFT-0.5B + DPO [29]3232.927.928.952.059.246.144.8
SFT-0.5B + SPO (ours)3233.428.029.252.460.447.745.4
LLaVA-OV-7B [18]3264.241.860.164.779.456.458.1
SFT-7B [35]3266.845.560.064.977.257.557.8
SFT-7B + DPO [29]3266.946.259.565.179.858.158.3
SFT-7B + SPO (ours)3267.946.760.565.379.657.858.5
", + "bbox": [ + 142, + 425, + 857, + 690 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 712, + 223, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Training Recipe", + "text_level": 1, + "bbox": [ + 89, + 739, + 250, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We follow [18] to construct our model architecture, including the instructional language model, vision encoder, and a 2-layer MLP projector. This work adopts LoRA [15] for an efficient and lightweight fine-tuning, where we set low-rank adapters with rank $= 64$ and $\\alpha = 128$ for the LLM. We denote $\\theta$ as all the trainable parameters, including LoRA and projector layer. As shown in Fig. 3, the training pipeline has two stages: 1) Utilizing supervised fine-tuning (SFT) to train an SFT model by reasoning and learning to ques", + "bbox": [ + 89, + 763, + 485, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tion [35]. 2) Fine-tune the LVML policy (given the SFT model as a reference) using the proposed SPO approach by maximizing a novel structured reward function. We show the training details for each stage in supplementary.", + "bbox": [ + 511, + 713, + 906, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluation Benchmarks", + "text_level": 1, + "bbox": [ + 511, + 787, + 736, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We primarily assess our models using vision-language benchmarks across different visual modalities (e.g., image, multi-image, and video) that are not included in the training data. For a fair comparison with other LVLMs, we primarily report results from original papers and utilize Immseval [49] for test setting consistency. All our results are", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "746", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f13afe7ae10049180636b63708a58286d0782643ca3b3857014a87f59b4fd5b2.jpg", + "image_caption": [ + "Figure 6. Qualitative results of LLaVA-Bench-Wilder. When trained with DPO, the model becomes over-confident when indicating low visual dependency content. In contrast, SPO improves the model's visual reasoning ability and prevents over-confident prediction when explaining uncertain situations, acting similarly to the reference model, GPT-4V." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 241, + 172 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Are the otters in the image reading a book, and if so, can you provide a detailed description of what might be happening in this whimsical scene?", + "bbox": [ + 99, + 188, + 227, + 292 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "SPO-7B Based on the image, it appears that the otters are interacting with a book in an unusual and anthropomorphic manner. The otters seem to be holding the book open as if they are reading it, which is a humorous and imaginative interpretation of their behavior. The book's placement between them suggests a shared interest or activity. The scene is whimsical because otters do not engage in such human-like behaviors naturally; this image is likely staged for artistic or entertainment purposes.", + "DPO-7B The otters in the image are likely to be sea otters, as they are shown floating on their backs in water and holding a book between them. The fact that they are holding a book also indicates that they have been exposed to human activities or objects, which could mean they live in an area where humans frequently visit or interact with them.", + "GPT-4V Yes, the otters in the image appear to be reading a book, which is a whimsical and anthropomorphized activity for animals. The scene is set in a body of water, with the two otters floating and seemingly fully absorbed in the book. One otter holds the book with one hand and points to a line or page with the other, as if highlighting something or guiding the reading process. The second otter is focused on the book, with an expression of curiosity or concentration. This playful image brings a charming and imaginative element, showcasing animals engaging in human-like activities and suggesting themes of education, curiosity, and companionship. Reference answer" + ], + "bbox": [ + 246, + 90, + 895, + 304 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "reported with greedy sampling and 0-shot settings.", + "bbox": [ + 89, + 385, + 424, + 400 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Single-image Benchmarks. We consider the LVLM as a general-purpose assistant in the wild as the most critical metric. To validate the capabilities in real-world scenarios with open-form instructions, we utilize MMStar [5], LLaVA-Bench-Wilder [18], and RealWorldQA [43]. Besides chat ability, visual perception evaluates the model's reasoning ability, so we adopt MME [13] and MMMU [48].", + "bbox": [ + 89, + 401, + 482, + 505 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-image Benchmarks. Multi-image tasks test the model's ability to connect and reason across images, making them generally more challenging as they require higher-level contextual integration. This work utilizes LLaVA-Interleave Bench [20] under the out-of-domain setting that does not overlap with training scenarios. This setting covers 13 challenging real-world tasks such as spotting the difference and visual storytelling, etc. We also provide a comprehensive and holistic evaluation of MuirBench [37], which contains 12 diverse multi-image tasks.", + "bbox": [ + 89, + 506, + 482, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Video Benchmarks. Video understanding is a critical ability to build a general-purpose vision-language assistant. This work conducts several open-ended and multi-choice video benchmarks. EgoSchema [25] and MLVU [54] focus on long video understanding, NExTQA [44], LongVideoBench [40] and VideoMME [14] contains diverse video domains and duration.", + "bbox": [ + 89, + 657, + 482, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Main Results", + "text_level": 1, + "bbox": [ + 89, + 772, + 228, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 1, while both preference optimization algorithms enhance performance over the SFT model, SPO achieves the best overall performance across all the image benchmarks under 0.5B and 7B scale. The consistency highlights the robustness and effectiveness of SPO. Notably, SPO outperforms its baseline model LLaVA-OV by 6.3 (0.5B) and 2.4 (7B) points on LLaVA-Bench-Wilder, and the improve", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ment on RealWorldQA indicates that SPO enhances the model's visual reasoning capability while handling freeform visual chat and real-world scenario. Such improvement follows our expectations since we precisely optimize the model to capture the alignment between visual input and dialogue interaction. Fig. 6 shows a qualitative example of real-world visual chat from three models.", + "bbox": [ + 511, + 385, + 905, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We have observed an improvement of SPO on the multi-image benchmarks, demonstrating the significance of our method for improving comprehensive and complex visual reasoning. Specifically, SPO-0.5B achieves a $4.4\\%$ and $1.0\\%$ improvement, and SPO-7B achieves an $8.1\\%$ and $2.0\\%$ improvement over LLaVA-OV and our SFT model. Table 2 also provides evaluation results on video benchmarks. The consistent improvement of SPO-0.5b/ 7B on EgoSchema, MLVU, and L-VideoBench indicates that SPO also improves consecutive visual reasoning, especially for medium and long video understanding.", + "bbox": [ + 511, + 491, + 905, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SPO has better visual-question understanding ability than DPO since it encourages visually informative questions by utilizing the structured preference tree that captures diverse question-answer pairs. Regarding the low-quality questions, SPO also learns their preferred answers through a subtree within the structured preferences. To validate our hypothesis, we prompt GPT-4o to score 216 questions in the MMVet dataset and split them into high-dependency (High-Q) and low-dependency (Low-Q) visual questions. Table 4 shows that more visually focused questions consistently improve the answering accuracy given images. As expected, without looking at images (w/o images), Low-Q reports a smaller performance drop than High-Q, and SPO has lower performance drop than DPO. Empirical evidence shows that SPO is more robust to question quality than DPO.", + "bbox": [ + 511, + 657, + 905, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Limitation. While SPO enhances the SFT model, we", + "bbox": [ + 532, + 885, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "747", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/98cb268d47fd151d0ee78be872e018ae739fa8031f259f461263ae6da2d6126a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
r(x,q)r([x,q],y)ImageMulti-ImageVideo
MMELLaVA-WilderLLaVA1MuirBenchEgoSchemaMLVU
✓✓158570.267.946.760.565.3
✓X156969.967.346.260.065.2
X✓157169.767.246.360.165.0
XX156569.566.845.560.064.9
", + "bbox": [ + 99, + 88, + 527, + 170 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bc1a891cb3414ad7f76c576ea366720b1c39daec88bd0b8787d4c406dec59de1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 92, + 903, + 170 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/40109b037f7ca30cf62057e6f51b0d93216cbfda5af806593650eed4fc844511.jpg", + "table_caption": [ + "Table 3. Ablation study under SFT-7B-SPO. We ablate the trajectory re- Figure 7. Left. SPO achieves higher preference classification reward design of SPO. The formal trajectory reward is $r(\\tau) = r(x,q) +$ accuracy than DPO throughout training. Right. SPO maxi $r([x,q],y)$ , where $r(x,q)$ and $r([x,q],y)$ are question and answer reward. mizes the reward margin faster effectiveness than DPO.", + "Table 4. Test accuracy drop in two question splits on MMVet." + ], + "table_footnote": [], + "table_body": "
ModelsHigh-Q (w/o image)Low-Q (w/o image)
SFT-0.5B + DPO27.0 (6.0) ↓ 77.8%17.0 (6.0) ↓ 64.7%
SFT-0.5B + SPO29.0 (8.0) ↓ 72.4%23.0 (12.0) ↓ 47.8%
", + "bbox": [ + 96, + 262, + 478, + 314 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "want to emphasize the critical role of the SFT model as the foundational stage for all preference optimization methods. Notably, the limited computational resources primarily drive the performance decline after SFT on several benchmarks. Our SFT-7B model was fine-tuned with 8 GPUs — only 1/16 of training GPUs in LLaVA-OV — resulting in a much smaller batch size for our approach. Specifically, the performance degradation of SFT-7B on MMStar, MMU, and MME sets a lower bound for the subsequent gains achievable by SPO. A similar trend is observed in video benchmarks, as shown in Table 2. Additionally, we decrease the anyres [18] to $60\\%$ of the regular setting for visual encoding to fit in our computational resource. Due to limited computational resources, SPO mainly focuses on single-turn conversations. We leave multi-turn conversations study in future work.", + "bbox": [ + 88, + 339, + 482, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Model Discussion", + "text_level": 1, + "bbox": [ + 89, + 590, + 261, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Structured Reward Speeds Up Policy Learning. The Bradley-Terry objective function essentially aims to optimize the reward margin of the preferred and dispreferred action $\\Delta r(x, \\rho_w, \\rho_l) = r(x, \\rho_w) - r(x, \\rho_l)$ . In Eq. (10), SPO optimizes a joint reward margin including $\\Delta r(x, q_1, q_2)$ and $\\Delta r([x, q_1], y_1, y_2)$ . While DPO only optimizes the reward margin between responses $\\Delta r([x, q_1], y_1, y_2)$ . We use the SPO and DPO models to investigate the relationship between the learned reward margin and the preference classification accuracy. Fig. 7 shows that SPO maximizes the reward margin more effectively than DPO, and SPO achieves higher preference classification accuracy than DPO along all the training steps. By precisely adjusting $\\beta$ for different preference probabilities, we find that optimizing multiple trajectory pairs together can explicitly increase the data diversity, leading to better generalizability and performance.", + "bbox": [ + 91, + 613, + 482, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Influence of Each Step Reward. SPO optimizes a policy to satisfy multiple preferences in a single objective function. In Eq. (10), we use $\\lambda$ and $\\delta$ to control the deviation of", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5a3cab762a511ca0a63c8e0eac88353e36e36104e3da3ccc286416ab3cdd85d0.jpg", + "table_caption": [ + "Table 5. Latency vs Accuracy on the LLaVA-W dataset" + ], + "table_footnote": [], + "table_body": "
MetricSFT-7BSFT-7B + DPOSFT-7B + SPO
Accuracy (%)72.873.574.1
Inference time (s)523547254
", + "bbox": [ + 524, + 262, + 897, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "preference alignment over two action groups, questioning and answering. Following the same training strategy and preference data, we evaluate the performance of SPO w/o question reward $r(x,q) = 0$ and SPO w/o answer reward $r([x,q] = y) = 0$ . As shown in Table. 3, learning question policy improves real-world QA ability such as LLaVA-Wilder and LLaVA; learning answer policy brings more improvement over fixed-form QA such as MME. Overall, the joint policy improves the model from different aspects.", + "bbox": [ + 511, + 327, + 903, + 464 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Computational Resource Analysis. For training, optimizing the whole dialogue will not burden the computation cost when comparing SPO with DPO (e.g., SPO: 24h vs DPO: 22h using $4^{*}\\mathrm{A}100$ GPUs) since only the question tokens need to be unmasked. For inference, Table 5 evaluates latency by reporting the total inference time of three models on 60 open-ended questions from the LLaVA-W dataset, and the test accuracy is measured by GPT-4o. Empirically, SPO generates more concise and accurate content than the SFT and DPO models.", + "bbox": [ + 511, + 465, + 905, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 632, + 633, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we have proposed Structured Policy Optimization (SPO), a novel approach to enhancing multi-modal preference optimization for large vision-language models (LVLMs). SPO alleviates the limitation of existing DPO methods, which adopt a single, unstructured reward and tend to prioritize linguistic nuances over visual context. Specifically, SPO introduces a new structured reward represented in trajectories to improve the visual dependency by maximizing the reward between preferred and dispreferred questions/answers. Our experimental results demonstrate that SPO improves vision-language understanding and reasoning across diverse tasks, showcasing superior alignment with human-preferred responses compared to traditional DPO methods. Overall, this approach better captures the dynamic interplay between visual inputs, questions, and responses, aligning more effectively with human preferences.", + "bbox": [ + 511, + 657, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "748", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research was supported by the DEVCOM Army Research Laboratory under Contract W911QX-21-D-0001.", + "bbox": [ + 89, + 113, + 483, + 143 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 170, + 187, + 186 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] OpenAI Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, et al. Gpt-4 technical report. 2023. 2", + "[2] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022. 3", + "[3] Hritik Bansal, Ashima Suvarna, Gantavya Bhatt, Nanyun Peng, Kai-Wei Chang, and Aditya Grover. Comparing bad apples to good oranges: Aligning large language models via joint preference optimization. ArXiv, 2024. 2, 4", + "[4] Lin Chen, Jinsong Li, Xiao wen Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. In ArXiv, 2023. 2", + "[5] Lin Chen, Jinsong Li, Xiao wen Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, and Feng Zhao. Are we on the right way for evaluating large vision-language models? In NeurIPS, 2024. 2, 7", + "[6] Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji, and Quanquan Gu. Self-play fine-tuning converts weak language models to strong language models. ArXiv, 2024. 3", + "[7] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 6", + "[8] Paul F. Christiano, Jan Leike, Tom B. Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences. In Proceedings of the 31st International Conference on Neural Information Processing Systems, 2017. 1, 3, 5", + "[9] Ruiguo Cui and Peter Teo. Thinking through talk: Using dialogue to develop students' critical thinking. Teaching and Teacher Education, 2023. 2, 4", + "[10] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Albert Li, Pascale Fung, and Steven C. H. Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, 2024. 2", + "[11] Yihe Deng, Pan Lu, Fan Yin, Ziniu Hu, Sheng Shen, James Zou, Kai-Wei Chang, and Wei Wang. Enhancing large vision language models with self-training on image comprehension. In NeurIPS, 2024. 1, 3", + "[12] Hanze Dong, Wei Xiong, Deepanshu Goyal, Rui Pan, Shizhe Diao, Jipeng Zhang, Kashun Shum, and T. Zhang. Raft:" + ], + "bbox": [ + 93, + 195, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Reward ranked finetuning for generative foundation model alignment. ArXiv, 2023. 1, 3", + "[13] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. ArXiv, 2023. 1, 7", + "[14] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, Peixian Chen, Yanwei Li, Shaohui Lin, Sirui Zhao, Ke Li, Tong Xu, Xiawu Zheng, Enhong Chen, Rongrong Ji, and Xing Sun. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. ArXiv, 2024. 7", + "[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 6", + "[16] Dongfu Jiang, Xuan He, Huaye Zeng, Cong Wei, Max W.F. Ku, Qian Liu, and Wenhu Chen. Mantis: Interleaved multi-image instruction tuning. ArXiv, 2024. 2", + "[17] Xin Lai, Zhuotao Tian, Yukang Chen, Senqiao Yang, Xiangru Peng, and Jiaya Jia. Step-dpo: Step-wise preference optimization for long-chain reasoning of llms. ArXiv, 2024. 1, 3", + "[18] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. ArXiv, 2024. 2, 6, 7, 8", + "[19] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. In NeurIPS, 2023. 1, 2, 3", + "[20] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava-last-interleave: Tackling multi-image, video, and 3d in large multimodal models. ArXiv, 2024. 2, 7", + "[21] Juncheng Li, Kaihang Pan, Zhiqi Ge, Minghe Gao, Hanwang Zhang, Wei Ji, Wenqiao Zhang, Tat seng Chua, Siliang Tang, and Yueting Zhuang. Fine-tuning multimodal llms to follow zero-shot demonstrative instructions. In International Conference on Learning Representations, 2024. 2", + "[22] Lei Li, Zhihui Xie, Mukai Li, Shunian Chen, Peiyi Wang, Liang Chen, Yazheng Yang, Benyou Wang, and Lingpeng Kong. Silkie: Preference distillation for large visual language models. ArXiv, 2023. 1, 3", + "[23] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. ArXiv, 2023. 2", + "[24] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In ArXiv, 2023. 2", + "[25] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. ArXiv, 2023. 1, 7", + "[26] OpenAI. Gpt-4v(ision) system card. 2023. 2, 6" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "749", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] OpenAI. Gpt-4o system card. 2024. 2, 4, 6", + "[28] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke E. Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Francis Christiano, Jan Leike, and Ryan J. Lowe. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 3", + "[29] Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. In NeurIPS, 2023. 1, 2, 3, 4, 6", + "[30] Rafael Rafailov, Joey Hejna, Ryan Park, and Chelsea Finn. From r to q: Your language model is secretly a q-function. In First Conference on Language Modeling, 2024. 1, 3", + "[31] Corby Rosset, Ching-An Cheng, Arindam Mitra, Michael Santacroce, Ahmed Awadallah, and Tengyang Xie. Direct nash optimization: Teaching language models to self-improve with general preferences. ArXiv, 2024. 3", + "[32] Avi Singh, John D. Co-Reyes, Rishabh Agarwal, Ankesh Anand, Piyush Patil, Peter J. Liu, James Harrison, Jaehoon Lee, Kelvin Xu, Aaron T Parisi, Abhishek Kumar, Alex Alemi, Alex Rizkowsky, Azade Nova, Ben Adlam, Bernd Bohnet, Hanie Sedghi, Igor Mordatch, Isabelle Simpson, Izzeddin Gur, Jasper Snoek, Jeffrey Pennington, Jiri Hron, Kathleen Kenealy, Kevin Swersky, Kshitteej Mahajan, Laura Culp, Lechao Xiao, Maxwell Bileschi, Noah Constant, Roman Novak, Rosanne Liu, Tris Brian Warkentin, Yundi Qian, Ethan Dyer, Behnam Neyshabur, Jascha Narain Sohl-Dickstein, and Noah Fiedel. Beyond human data: Scaling self-training for problem-solving with language models. Trans. Mach. Learn. Res., 2023. 3", + "[33] Feifan Song, Yu Bowen, Minghao Li, Haiyang Yu, Fei Huang, Yongbin Li, and Houfeng Wang. Preference ranking optimization for human alignment. ArXiv, 2023. 1, 3", + "[34] Guohao Sun, Can Qin, Huazhu Fu, Linwei Wang, and Zhiqiang Tao. Stllava-med: Self-training large language and vision assistant for medical. In EMNLP, 2024. 2, 3", + "[35] Guohao Sun, Can Qin, Jiamian Wang, Zeyuan Chen, Ran Xu, and Zhiqiang Tao. Sq-llava: Self-questioning for large vision-language assistant. In ECCV, 2024. 2, 3, 4, 6", + "[36] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2, 6", + "[37] Fei Wang, Xingyu Fu, James Y. Huang, Zekun Li, Qin Liu, Xiaogeng Liu, Mingyu Derek Ma, Nan Xu, Wenxuan Zhou, Kai Zhang, Tianyi Yan, Wenjie Jacky Mo, Hsiang-Hui Liu, Pan Lu, Chunyuan Li, Chaowei Xiao, Kai-Wei Chang, Dan Roth, Sheng Zhang, Hoifung Poon, and Muhao Chen. Muir-bench: A comprehensive benchmark for robust multi-image understanding. ArXiv, 2024. 1, 7", + "[38] Fei Wang, Wenxuan Zhou, James Y. Huang, Nan Xu, Sheng Zhang, Hoifung Poon, and Muhao Chen. mdpo: Conditional preference optimization for multimodal large language models. In EMNLP, 2024. 1, 3" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Yuetian Weng, Mingfei Han, Haoyu He, Xiaojun Chang, and Bohan Zhuang. Longvlm: Efficient long video understanding via large language models. ArXiv, 2024. 2", + "[40] Haoning Wu, Dongxu Li, Bei Chen, and Junnan Li. Longvideobench: A benchmark for long-context interleaved video-language understanding. ArXiv, 2024. 7", + "[41] Junkang Wu, Yuexiang Xie, Zhengyi Yang, Jiancan Wu, Jinyang Gao, Bolin Ding, Xiang Wang, and Xiangnan He. $\\beta$ -dpo: Direct preference optimization with dynamic $\\beta$ . ArXiv, 2024. 5", + "[42] Yue Wu, Zhiqing Sun, Huizhuo Yuan, Kaixuan Ji, Yiming Yang, and Quanquan Gu. Self-play preference optimization for language model alignment. ArXiv, 2024. 3", + "[43] x.ai. Grok-1.5 vision preview. Technical report, x.ai, 2024. 7", + "[44] Junbin Xiao, Xindi Shang, Angela Yao, and Tat seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 7", + "[45] Tianyi Xiong, Tianyi Xiong, Xiyao Wang, Dong Guo, Qinghao Ye, Haoqi Fan, Quanquan Gu, Heng Huang, and Chunyuan Li. Llava-critic: Learning to evaluate multimodal models. ArXiv, 2024. 4", + "[46] Enwei Xu, Wei Wang, and Qingxia Wang. The effectiveness of collaborative problem solving in promoting students' critical thinking: A meta-analysis based on empirical literature. *Humanities and Social Sciences Communications*, 2023. 2, 4", + "[47] Xubing Ye, Yukang Gan, Xiaoke Huang, Yixiao Ge, Ying Shan, and Yansong Tang. Voco-llama: Towards vision compression with large language models. 2024. 2", + "[48] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Ren-liang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 7", + "[49] Kaichen Zhang, Bo Li, Peiyuan Zhang, Fanyi Pu, Joshua Adrian Cahyono, Kairui Hu, Shuai Liu, Yuanhan Zhang, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Lmmseval: Reality check on the evaluation of large multimodal models. ArXiv, 2024. 6", + "[50] Pan Zhang, Xiao wen Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, Songyang Zhang, Wenwei Zhang, Yining Li, Yang Gao, Peng Sun, Xinyue Zhang, Wei Li, Jingwen Li, Wenhai Wang, Hang Yan, Conghui He, Xingcheng Zhang, Kai Chen, Jifeng Dai, Yu Qiao, Dahua Lin, and Jiaqi Wang. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. ArXiv, 2024. 6", + "[51] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. In ArXiv, 2024. 6" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "750", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Yanzhe Zhang, Ruiyi Zhang, Jiumiang Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tongfei Sun. Llavar: Enhanced visual instruction tuning for text-rich image understanding. In ArXiv, 2023. 2", + "[53] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. ArXiv, 2024. 2, 6", + "[54] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. ArXiv, 2024. 7", + "[55] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. ArXiv, 2023. 2", + "[56] Daniel M. Ziegler, Nisan Stiennon, Jeff Wu, Tom B. Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. ArXiv, 2019. 3" + ], + "bbox": [ + 91, + 92, + 480, + 358 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "751", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Structured Policy Optimization_ Enhance Large Vision-Language Model via Self-referenced Dialogue/5161fa7a-2cd7-4a41-959a-e1ddb02c88dc_model.json b/2025/Structured Policy Optimization_ Enhance Large Vision-Language Model via Self-referenced Dialogue/5161fa7a-2cd7-4a41-959a-e1ddb02c88dc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b53b23b58ca723fc994771f8a0c64e08bbc10d65 --- /dev/null +++ b/2025/Structured Policy Optimization_ Enhance Large Vision-Language Model via Self-referenced Dialogue/5161fa7a-2cd7-4a41-959a-e1ddb02c88dc_model.json @@ -0,0 +1,2334 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.13, + 0.885, + 0.177 + ], + "angle": 0, + "content": "Structured Policy Optimization: Enhance Large Vision-Language Model via Self-Referenced Dialogue" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.203, + 0.806, + 0.241 + ], + "angle": 0, + "content": "Guohao Sun1*, Can Qin2, Yihao Feng2, Zeyuan Chen2, Ran Xu2, Sohail Dianat1, Majid Rabbani1, Raghuveer Rao3, and Zhiqiang Tao1*" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.244, + 0.761, + 0.282 + ], + "angle": 0, + "content": "\\(^{1}\\)Rochester Institute of Technology \n\\(^{2}\\)Salesforce AI Research \\(^{3}\\)DEVCOM Army Research Laboratory" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.314, + 0.327, + 0.33 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.348, + 0.485, + 0.71 + ], + "angle": 0, + "content": "Preference optimization algorithms typically enhance LLM response quality by leveraging human feedback on multiple answers given a fixed instruction. However, these methods often lack capturing the dynamic nature of conversational exchanges. For large vision-language models (LVLMs), direct preference optimization (DPO) can over-emphasize linguistic nuances while overlooking visual context. To address this challenge, we introduce structured policy optimization (SPO) - a novel preference optimization method that simultaneously aligns preference instructions, responses, and dialogue interactions to improve multi-modal understanding and reasoning capabilities. The efficacy of SPO is attributed to one key design: treating the questioning and answering as a sequential action and binding them through a trajectory reward. This reward formulation better aligns with real-world dialogue studies and eliminates the need for fixed instructions. We evaluate our models on interleaved benchmarks, including image, multi-image, and video-based understanding and reasoning tasks. Experimental results show that the proposed SPO finetuning LVLM with multi-modal preference data can align with human preference more efficiently than DPO. The code is available at https://github.com/heliossun/Structure-Policy-Optimization." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.743, + 0.222, + 0.759 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.77, + 0.483, + 0.862 + ], + "angle": 0, + "content": "Reinforcement learning from human feedback (RLHF) [8] has demonstrated remarkable success in aligning large language/vision-language models (LLMs/LVLMs) with human preferences, enabling models to adapt their behavior based on success or failure in increasingly complex tasks. Typically, RLHF trains a reward model and optimizes the" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.316, + 0.768, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.37, + 0.758, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.315, + 0.895, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.394, + 0.895, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.51, + 0.907, + 0.621 + ], + "angle": 0, + "content": "Figure 1. Comparison between DPO and SPO in optimizing an LVLM policy. DPO employs a single reward that optimizes the preferred answer given the fixed visual & question input, \\( p(y_w \\succ y_l|x, q) \\), which could over-emphasize language nuance and overlook visual clues [38]. By contrast, SPO introduces a structured reward by \\( p(\\mathcal{Y}_w \\succ \\mathcal{Y}_l|x) \\), conditioning on the visual dialogue, to optimize the preferred trajectories spanning over questions and answers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.655, + 0.907, + 0.687 + ], + "angle": 0, + "content": "LLM/LVLM policy to maximize the reward, which is applied in both language and vision-language (VL) domains." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Recently, the offline preference optimization methods, such as DPO [29] and its variants [12, 17, 30, 33], have proposed various reward objectives to satisfy different human expectations. Owing to its simplicity and stability, DPO has been widely used in multi-modal preference optimization [11, 19, 22, 38] by aligning LVLM to human-preferred responses conditioned on fixed visual and text instructions (see Fig. 1). However, unlike purely text dialogue, the visual input can lead to far more enriched information interactions, which require varying degrees of visual dependence [13, 25, 37]. Directly applying DPO to multi-modal preference data (e.g., images/videos and texts) will over-emphasize language nuance and overlook the visual condition [38]. Specifically, DPO applies a single, un" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "*Corresponding authors: Guohao Sun (gs4288@rit.edu) and Zhiqiang Tao (zhiqiang.tao@rit.edu)" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "741" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.103, + 0.092, + 0.283, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.183, + 0.259, + 0.194 + ], + "angle": 0, + "content": "(a) Question vs Answer." + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.091, + 0.474, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.323, + 0.183, + 0.438, + 0.194 + ], + "angle": 0, + "content": "(b) Score distribution." + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.207, + 0.483, + 0.262 + ], + "angle": 0, + "content": "Figure 2. Scoring visual dialogue with GPT-4o as judge. (a) Questions with higher visual dependency (x-axis, scoring from 4 to 10, \\(\\uparrow\\)) tend to produce answers with higher quality (y-axis, scoring from 0 to 9, \\(\\uparrow\\)). (b) Question and answer score distributions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.293, + 0.483, + 0.368 + ], + "angle": 0, + "content": "structured reward function to maximize the gap between the preferred and dispreferred answer upon the same input sequence; however, this gap might be due to the lack of visual understanding [5], which is generally ignored by existing DPO methods, especially for LVLMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.371, + 0.483, + 0.718 + ], + "angle": 0, + "content": "Inspired by the dialogue studies [9, 46], we empirically observe that asking a good visual-related question can turn passive learning into active exploration, allowing for deeper visual understanding and broader knowledge exploration before answering. For instance, as shown in Fig. 1, people prefer the question \"What emotions are conveyed through the shots in the video?\" to \"Which city did the skateboarder go to in the video?\" since the former is more visually dependent and the answer will cover more visual details. To quantitatively validate our observation, we apply GPT-4o to evaluate the visual relevance (scoring from 4 to 10) of questions and score the quality of answers (higher is better) in our curated preference data based on the instruction dataset in [18]. Fig. 2a shows that human-preferred questions with higher visual relevance tend to produce responses with higher quality, which is further demonstrated by the consistent question/answer score distributions in Fig. 2b. We introduce this correlation between visual questions and answer quality as a preference dialogue interaction. By aligning the LVLM policy to generate questions and answers with human preference, we build a more in-depth alignment between vision and dialogues, enabling a better VL understanding for broad tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this study, we propose Structured Policy Optimization (SPO), a novel structured DPO method designed to enhance general vision-language understanding and reasoning. As shown in Fig. 1, SPO utilizes a binary tree structure to track visual dialogue changes, alleviating the limitation of using fixed questions in DPO. Unlike DPO, which only optimizes preference at leaf nodes, the proposed SPO optimizes preferences across different trajectories among sub-trees and develops a new structured reward model by leveraging trajectories to capture the dialogue interaction between the visual input, questions, and answers. Additionally, SPO controls the deviation of question reward and an-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": "swer reward by imposing importance weights on each step within the trajectory, which is substantially different from a single-level joint reward optimization [3]. To realize SPO training and increase data diversity, we utilize supervised fine-turning (SFT) to instruct an LVLM to self-generate the preference question-answer pairs, mitigating the distribution shift between the policy model and the preference optimization process [29]. We summarize the contributions of this work as follows." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.228, + 0.905, + 0.288 + ], + "angle": 0, + "content": "- For the first time in multi-modal preference optimization, we introduce SPO, a method designed for optimizing structured reward for aligning the policy model to human preferences over visual dialogues." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.289, + 0.905, + 0.363 + ], + "angle": 0, + "content": "- We present high-quality multi-modal preference data consisting of scoring over visual questions and answers. This dataset reflects the discrepancy between chosen and rejected dialogue, bringing insights into real-world conversational exchange and dialogue studies." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.364, + 0.905, + 0.422 + ], + "angle": 0, + "content": "- Extensive experiments show that SPO aligns with human preference effectively and overcomes the low visual dependency issue in DPO, leading to better performance on the visual chat benchmarks with free-form instructions." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.228, + 0.905, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.438, + 0.664, + 0.453 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.907, + 0.765 + ], + "angle": 0, + "content": "Generative Large Vision-Language Model. Recent general-purpose multi-modal models (LVLMs) have shown impressive capability. The commercial proprietary LVLMs, such as GPT-4V [26], GPT-4o [27], and Gemini [36] exhibit excellent generalize ability in various domains [48] such as design, business, science, medical, and science etc. In the open research community, the evolutionary trajectory of LVLMs has also led to subsequent advancements in different scenarios. The pioneering work of LLaVA [24] improves the multi-modal reasoning ability through visual instruction tuning. [4, 10, 24, 34, 35, 55] pushing the performance limits in single-image scenarios. [16, 21] explores multi-image tasks. While video LVLMs [23, 39, 47, 53] aim to alleviate significant challenges due to the long context issue. In contrast to these single-modality models, the latest LLaVA models [18, 20] demonstrate emerging capabilities across all scenarios after training on interleaved data. Following this path, this work aims to improve the general ability of multiple modalities by primarily focusing on improving the model's reasoning ability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Self-generated Synthetic Data. Instruction tuning is widely used in supervised fine-tuning (SFT) to improve general-purpose reasoning. In recent research, synthetic instructional data has become increasingly popular and considered a proxy for human data. This approach primarily leverages the ability of GPT family, [19, 24, 52] utilizes GPT-4 [1] to transform image description into conversations, [4, 18, 20] leverages GPT-4V [26] to process the multi-modal input and generate question-answer pairs." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "742" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.905, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.279, + 0.373, + 0.294 + ], + "angle": 0, + "content": "Stage 1: Reasoning and Learning to Question" + }, + { + "type": "image_caption", + "bbox": [ + 0.649, + 0.279, + 0.832, + 0.293 + ], + "angle": 0, + "content": "Stage2: Preference revelation" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.303, + 0.908, + 0.364 + ], + "angle": 0, + "content": "Figure 3. Illustration of LVLM model architecture and training pipeline in the proposed SPO. In Stage 1, we implement supervised fine-tuning through learning to question [35], resulting in an SFT model that enables better visual reasoning and self-generated questions. In Stage 2, we first prompt the SFT model to generate visual dialogues labeled by GPT-4o with human preference. Then, we apply the proposed SPO approach to optimize the LVLM policy through a structured reward, leading to better VL understanding." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.387, + 0.484, + 0.522 + ], + "angle": 0, + "content": "However, large-scale synthetic data generation comes with unexpected API costs. To overcome this, [34] train the model with a visual-questioning objective function [35] and generate the question-answer pairs efficiently. While most works mentioned above consider one single optimization procedure, self-play fine-tuning methods [6, 31, 32, 42] generate new synthetic data within the training loop. This work focuses primarily on single-turn optimization and avoids iterative processes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.484, + 0.736 + ], + "angle": 0, + "content": "Preference Optimization Objectives. Reinforcement Learning from Human Feedback (RLHF) is a technique used to align large language models with human preferences and values [2, 8, 28, 56]. To eliminate the complexity and challenges of optimizing online preference algorithms, Direct Preference Optimization (DPO) [29] is a prominent example. To extend two-instance classification, ranking objectives enable comparisons across multiple instances [12, 17, 30, 33]. Recent efforts in multi-modal preference learning mainly focus on preference data design [11, 19, 22, 38], while utilizing the formal DPO objective function. This work proposes a novel structured DPO by maximizing the structured rewards to improve vision-language alignment." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.753, + 0.429, + 0.772 + ], + "angle": 0, + "content": "3. SPO: Structured Policy Optimization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Preference optimization uses explicit feedback (e.g., human ratings, comparisons, or rewards) to fine-tune the SFT model, making it more responsive to real-world user needs. As shown in Fig. 3, this work employs a two-stage training strategy: 1) Utilizing supervised fine-tuning (SFT) to train a SFT model by reasoning and learning to question [35], which is able to generate diverse visual dialogues that align with the training data's specific patterns and content. 2)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.907, + 0.465 + ], + "angle": 0, + "content": "Fine-tune the LVML policy (given the SFT model as a reference) using the proposed SPO approach by maximizing a novel structured reward function. In the following, we will first provide preliminaries about preference optimization and then elaborate on our proposed SPO approach." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.469, + 0.791, + 0.485 + ], + "angle": 0, + "content": "3.1. Backgrounds: RLHF and DPO" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.907, + 0.553 + ], + "angle": 0, + "content": "\\(\\mathbf{RLHF}[8]\\) first trains a reward model \\(r_{\\phi}(x,y)\\) parametrized by \\(\\phi\\) to predict the human preference distribution \\(p\\) over two responses \\(y_{1}\\) and \\(y_{2}\\) (e.g., answers given by an LVLM) with condition \\(x\\), simulated by the Bradely-Terry (BT) model as" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.557, + 0.905, + 0.61 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} p \\left(y _ {1} \\succ y _ {2} \\mid x\\right) = \\frac {\\exp \\left(r _ {\\phi} \\left(x , y _ {1}\\right)\\right)}{\\exp \\left(r _ {\\phi} \\left(x , y _ {1}\\right)\\right) + \\exp \\left(r _ {\\phi} \\left(x , y _ {2}\\right)\\right)} \\tag {1} \\\\ = \\sigma \\left(r _ {\\phi} (x, y _ {1}) - r _ {\\phi} (x, y _ {2})\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.906, + 0.675 + ], + "angle": 0, + "content": "where \\( x \\) denotes visual input and questions (instructions) in this study. After training, the reward model \\( r_{\\phi} \\) provides feedback for each response generated by policy model \\( \\pi_{\\theta} \\), which is optimized by the following objective function:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.681, + 0.905, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\max _ {\\pi_ {\\theta}} \\mathbb {E} _ {x \\sim D, y \\sim \\pi_ {\\theta} (y | x)} [ r _ {\\phi} (x, y) ] \\tag {2} \\\\ - \\beta \\mathbb {D} _ {K L} \\left[ \\pi_ {\\theta} (y | x) \\| \\pi_ {r e f} (y | x) \\right], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.73, + 0.906, + 0.807 + ], + "angle": 0, + "content": "where \\(\\mathcal{D} = \\{x^{(i)},y_w^{(i)},y_l^{(i)}\\}_{i = 1}^N\\) is the preference dataset, \\(\\pi_{\\theta}\\) is the policy model that need to be updated, and \\(\\pi_{ref}\\) is a reference model (e.g., an SFT model) that avoids overshifting to preferences. \\(y_{w}\\) and \\(y_{l}\\) refer to preferred and dispreferred answers such as \\(y_{1}\\) and \\(y_{2}\\) in Eq. (1)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.808, + 0.907, + 0.869 + ], + "angle": 0, + "content": "DPO [29] reformulates the objective function of RLHF by changing from loss over reward function to optimal policy. Previous work [29] derives the optimal policy \\(\\pi_{\\theta}\\) of Eq. (2) as" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.873, + 0.905, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\pi_ {\\theta} (y | x) = \\frac {1}{Z (x)} \\pi_ {r e f} (y | x) \\exp \\left(\\frac {1}{\\beta} r (x, y)\\right), \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.512, + 0.957 + ], + "angle": 0, + "content": "743" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.486, + 0.138 + ], + "angle": 0, + "content": "where \\( Z(x) = \\sum_{y} \\pi_{\\text{ref}}(y|x) \\exp \\left( \\frac{1}{\\beta} r(x,y) \\right) \\) is the marginal likelihood over all possible model responses. To simplify it, DPO reforms Eq. (3) as" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.145, + 0.483, + 0.18 + ], + "angle": 0, + "content": "\\[\nr (x, y) = \\beta \\left(\\log \\frac {\\pi_ {\\theta} (y \\mid x)}{\\pi_ {r e f} (y \\mid x)} + \\log Z (x)\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.185, + 0.485, + 0.271 + ], + "angle": 0, + "content": "The BT model for preference estimation depends only on the difference of rewards between two responses \\( y_{1} \\) and \\( y_{2} \\) [29]. By substituting Eq. (4) into Eq. (1), DPO expresses the probability of human preference data in terms of the optimal policy and formulating the objective function for a parameterized policy \\( \\pi_{\\theta} \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.276, + 0.479, + 0.307 + ], + "angle": 0, + "content": "\\[\n- \\mathbb {E} _ {(x, y _ {w}, y _ {l}) \\sim \\mathcal {D}} \\left[ \\log \\sigma \\left(\\beta \\log \\frac {\\pi_ {\\theta} (y _ {w} | x)}{\\pi_ {r e f} (y _ {w} | x)} - \\beta \\log \\frac {\\pi_ {\\theta} (y _ {l} | x)}{\\pi_ {r e f} (y _ {l} | x)}\\right) \\right],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.312, + 0.483, + 0.341 + ], + "angle": 0, + "content": "where \\(\\sigma\\) represents the sigmoid function and \\(\\beta\\) is a scalable value to control the deviation from the reference model." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.342, + 0.483, + 0.553 + ], + "angle": 0, + "content": "Motivation. Nevertheless, DPO methods mainly focus on optimizing the LLM policy by comparing multiple generations \\( y \\) based on a fixed input \\( x \\), falling short in capturing the dynamics of visual dialogue interactions. Specifically, high visual-dependent instructions require better visual understanding and reasoning ability to respond than easy-to-answer questions [9, 46]. Such discrepancy leads to responses with different amounts of visual clues. To quantitatively assess the visual dependency of each dialogue interaction, we utilized GPT-4o to review visual dialogues according to a set of manually designed rules. As shown in Fig. 2b, instructions with higher visual dependency tend to result in responses with more relevance to the visual input, leading to a preference dialogue interaction." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.553, + 0.483, + 0.719 + ], + "angle": 0, + "content": "Empirical evidence [35] suggests that training the policy to align with preference interaction can enhance visual understanding by maximizing the likelihood of preferred questions and answers. To this end, we propose Structured Policy Optimization (SPO). In SPO, we consider question and answer as two independent actions, each leading to a different state, simulating the conversational exchange. As shown in Fig. 1, the policy can be naturally modeled as a binary tree with depth 2. Unlike DPO, which only optimizes preference at the leaf nodes, SPO optimizes preferences across the entire sub-tree." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.422, + 0.745 + ], + "angle": 0, + "content": "3.2. SPO Aligns With Dialogue Interaction" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Preference Data Curation. To implement SPO, we first prompt the SFT model (given in Stage 1 in Fig. 3) to generate visual instruction-answer pairs, aiming to mitigate the distribution shift between SFT and the preference optimization step [3, 29]. Notably, similar to DPO, the SFT model also works as a reference policy \\(\\pi_{\\mathrm{ref}}\\) in SPO. However, the difference lies in our \\(\\pi_{\\mathrm{ref}}\\) being capable of self-generating questions by implementing SFT through learning to question [35]. As shown in Fig. 4, we represent different visual dialogues as trajectories in a tree structure - the" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.09, + 0.825, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.23, + 0.907, + 0.287 + ], + "angle": 0, + "content": "Figure 4. Illustration of a tree structure that represents different visual dialogues, which are denoted as trajectory, e.g., \\(\\tau_{1} = (x, q_{w}, y_{w}^{1})\\). For preference labeling, we utilize GPT-4o as a judge to perform pair-wise scoring over trajectory pairs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.311, + 0.905, + 0.434 + ], + "angle": 0, + "content": "input \\(x\\) denotes the root, questions occupy the first level, and answers are leaf nodes - and collect preference data \\(\\mathcal{D} = \\{x^{(i)},\\mathcal{Y}_w^{(i)},\\mathcal{Y}_l^{(i)}\\}_{i = 1}^N\\) by sub-trees*, where the preferred sub-tree \\(\\mathcal{V}_w = (q_w,y_w^1,y_l^1)\\) and dispreferred subtree \\(\\mathcal{V}_l = (q_l,y_w^2,y_l^2)\\). We sample questions \\(q_{w},q_{l}\\sim \\pi_{\\mathrm{ref}}\\) for each visual input \\(x\\), followed by answers \\(y_w^1,y_l^1\\sim\\) \\(\\pi_{\\mathrm{ref}}(\\cdot |x,q_w))]\\) and \\(y_w^2,y_l^2\\sim \\pi_{\\mathrm{ref}}(\\cdot |x,q_l))]\\), where the superscripts 1 and 2 indicate conditioning on \\(q_{w} / q_{l}\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.434, + 0.907, + 0.72 + ], + "angle": 0, + "content": "Empirically, the preference label accuracy is critical to increasing training stability. Following [45], AI-as-judge can offer scalable solutions to reduce human labor in complex task assessments and generate effective reward signals in reinforcement learning. Thus, to evaluate the synthetic data quality, we compared the distribution of 200 instruction-answer scores judged by GPT-4o [27] and an open-sourced LVLM (LLaVA-critic [45]). Fig. 5 shows that GPT-4o tends to judge with a higher standard than LLaVA-critic as its mean score is lower. We find that GPT-4o can provide scores with higher diversity than LLaVA-critic, acting more like a human annotator. Therefore, this work utilizes GPT-4o for all the preference labeling, and we propose a simple rule to filter out low-quality samples after collecting all the reviews. Specifically, we first remove samples with scores lower than a threshold, then remove the samples with repeated questions or answers, such as \\( q_{w} == q_{l} \\) or \\( y_{w}^{1} == y_{l}^{1} \\). By this means, our collected preference data maintains high diversity and reliability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.827 + ], + "angle": 0, + "content": "Structured Reward Formulation. Given the initial state \\( x \\), the preferred sub-tree \\( \\mathcal{V}_w \\) should has higher reward since we assume the preference instruction dominates the dialogue interaction. We apply the same Bradley-Terry probability model to ensure that the reward of the preferred sub-tree \\( \\mathcal{V}_w \\) exceeds the reward of the dispreferred sub-tree \\( \\mathcal{V}_l \\) by" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.837, + 0.905, + 0.854 + ], + "angle": 0, + "content": "\\[\np \\left(\\mathcal {Y} _ {w} \\succ \\mathcal {Y} _ {l} \\mid x\\right) = \\sigma \\left(r \\left(x, \\mathcal {Y} _ {w}\\right) - r \\left(x, \\mathcal {Y} _ {l}\\right)\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.863, + 0.905, + 0.88 + ], + "angle": 0, + "content": "However, directly deriving \\(r(x,\\mathcal{Y})\\) into Eq. (2) will intro" + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.888, + 0.905, + 0.9 + ], + "angle": 0, + "content": "*For convenience, we omit the superscript \\(i\\) when no confusion occurs." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.512, + 0.957 + ], + "angle": 0, + "content": "744" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.304 + ], + "angle": 0, + "content": "duce a complex term \\( Z(x) \\) (please refer to the supplementary material for a complete derivation). To overcome this issue, we decompose sub-trees \\( \\mathcal{V} \\) to trajectories \\( \\tau \\) and show how it connects to state-action trajectories in RLHF [8]. Let \\( \\tau = (s_1, a_1, s_2, a_2, \\ldots, s_T, a_T) \\) be a trajectory sampled from the policy \\( \\pi_{\\theta} \\), where \\( s, a, T \\) denote states, actions, and total steps. The total reward of a trajectory is defined as \\( r(\\tau) = \\sum_{t=1}^{T} r(s_t, a_t) \\). Thus, following this reward formulation, we can define a trajectory to represent one visual dialogue by setting \\( T = 2 \\), where the initial state and action are input \\( x \\) and questioning \\( s_1 = x, a_1 = q \\), and the following state and action are \\( s_2 = [x, q], a_2 = y \\), e.g., \\( \\tau_1 = (x, q_w, [x, q_w], y_w^1), \\tau_2 = (x, q_w, [x, q_w], y_l^1) \\), etc. Finally, we implement Eq. (5) by a trajectory reward as" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.315, + 0.483, + 0.332 + ], + "angle": 0, + "content": "\\[\nr (x, \\mathcal {Y}) \\sim r (\\tau) = \\lambda \\cdot r (x, q) + \\delta \\cdot r ([ x, q ], y), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.343, + 0.484, + 0.451 + ], + "angle": 0, + "content": "where \\(\\lambda\\) and \\(\\delta\\) are hyperparameters used to balance the question and answer rewards. By setting \\(\\lambda = \\delta\\), our objective function could be equivalent to \\(\\sigma(r(\\tau_1) - r(\\tau_2))\\), direct preference optimization (DPO) over two trajectory (i.e., question-answer pairs), rather than on how it is split across the binary tree. In the next section, we derive the closed-form expression of \\(r(x, q)\\) and \\(r([x, q], y)\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.458, + 0.349, + 0.475 + ], + "angle": 0, + "content": "3.3. The SPO Objective Function" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.48, + 0.483, + 0.511 + ], + "angle": 0, + "content": "Following Eq. (2) and our trajectory-based reward model, the general RL objective of SPO is given as" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.522, + 0.483, + 0.54 + ], + "angle": 0, + "content": "\\[\nJ (\\pi_ {\\theta}) = \\mathbb {E} _ {\\tau \\sim \\pi_ {\\theta}} [ r (\\tau) ] - \\beta \\mathrm {K L} [ \\pi_ {\\theta} (\\tau) | | \\pi_ {\\text {r e f}} (\\tau) ]. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.549, + 0.449, + 0.566 + ], + "angle": 0, + "content": "By substituting Eq. (6) into Eq. (7), we have \\( J(\\pi_{\\theta}) \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.573, + 0.486, + 0.644 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbb {E} _ {(x, q, y) \\sim \\tau} [ \\lambda r (x, q) + \\delta r ([ x, q ], y) ] - \\beta (\\mathrm {K L} [ \\pi_ {\\theta} (\\cdot | x) | | \\pi_ {\\operatorname {r e f}} (\\cdot | x) ]) \\\\ \\left. + \\mathrm {K L} \\left[ \\pi_ {\\theta} (\\cdot | [ x, q ]) \\mid \\mid \\pi_ {\\text {r e f}} (\\cdot | [ x, q ])) \\right]\\right), \\\\ = \\mathbb {E} _ {(x, q) \\sim \\tau} [ \\lambda r (x, q) ] - \\beta \\mathrm {K L} [ \\pi_ {\\theta} (\\cdot | x) | | \\pi_ {\\operatorname {r e f}} (\\cdot | x)) ] \\\\ + \\mathbb {E} _ {(x, q, y) \\sim \\tau} [ \\delta r ([ x, q ], y) ] - \\beta \\mathrm {K L} [ \\pi_ {\\theta} (\\cdot | [ x, q ]) | | \\pi_ {\\text {r e f}} (\\cdot | [ x, q ]) ] , \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.651, + 0.483, + 0.682 + ], + "angle": 0, + "content": "where \\(q\\sim \\pi_{\\theta}(\\cdot |x)\\) and \\(y\\sim \\pi_{\\theta}(\\cdot |[q,x])\\) . Following [41], the optimal solution of \\(J(\\pi_{\\theta})\\) can be derived as" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.693, + 0.401, + 0.725 + ], + "angle": 0, + "content": "\\[\n\\pi_ {\\theta} (q | x) = \\frac {1}{Z (x)} \\pi_ {\\mathrm {r e f}} (q | x) \\exp (\\frac {\\lambda}{\\beta} r (x, q)),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.727, + 0.475, + 0.76 + ], + "angle": 0, + "content": "\\[\n\\pi_ {\\theta} (y | [ x, q ]) = \\frac {1}{Z ([ x , q ])} \\pi_ {\\mathrm {r e f}} (y | [ x, q ]) \\exp (\\frac {\\delta}{\\beta} r ([ x, q ], y)).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.769, + 0.383, + 0.782 + ], + "angle": 0, + "content": "Thus, we can obtain the reward functions as" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.793, + 0.483, + 0.827 + ], + "angle": 0, + "content": "\\[\nr (x, q) = \\frac {\\beta}{\\lambda} \\log \\frac {\\pi_ {\\theta} (q | x)}{\\pi_ {\\operatorname {r e f}} (q | x)} + \\frac {\\beta}{\\lambda} \\log Z (x), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.828, + 0.462, + 0.862 + ], + "angle": 0, + "content": "\\[\nr ([ x, q ], y) = \\frac {\\beta}{\\delta} \\log \\frac {\\pi_ {\\theta} (y | [ x , q ])}{\\pi_ {\\mathrm {r e f}} (y | [ x , q ])} + \\frac {\\beta}{\\delta} \\log Z ([ x, q ]).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Now, given the definition of each step reward, the total trajectory reward defined in Eq. (6) in terms of the optimal" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.091, + 0.905, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.275, + 0.908, + 0.33 + ], + "angle": 0, + "content": "Figure 5. The score distribution of questions and answers, and labeling agreement with 3 LVLMS as judges. GPT-4o provides lower mean score and higher diversity than LLaVA-critic, acting more like a human annotator." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.365, + 0.907, + 0.44 + ], + "angle": 0, + "content": "policy can be directly substituted into the preference model in Eq. (1), where the \\( Z(x) \\) term will be canceled just as the original DPO derivation. Assuming \\( \\tau_w = (x, q_1, [x, q_1], y_1) \\) and \\( \\tau_l = (x, q_2, [x, q_2], y_2) \\), the preference probability of preferred \\( \\tau_w \\) over dispreferred \\( \\tau_l \\) can be defined as" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.457, + 0.909, + 0.538 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} p \\left(\\tau_ {w} \\succ \\tau_ {l}\\right) := \\sigma \\left(\\frac {\\beta}{\\lambda} \\log \\frac {\\pi_ {\\theta} \\left(q _ {1} \\mid x\\right)}{\\pi_ {\\operatorname {r e f}} \\left(q _ {1} \\mid x\\right)} + \\frac {\\beta}{\\delta} \\log \\frac {\\pi_ {\\theta} \\left(y _ {1} \\mid [ x , q _ {1} ]\\right)}{\\pi_ {\\operatorname {r e f}} \\left(y _ {1} \\mid [ x , q _ {1} ]\\right)} \\right. \\\\ \\left. - \\frac {\\beta}{\\lambda} \\log \\frac {\\pi_ {\\theta} \\left(q _ {2} \\mid x\\right)}{\\pi_ {\\text {r e f}} \\left(q _ {2} \\mid x\\right)} - \\frac {\\beta}{\\delta} \\log \\frac {\\pi_ {\\theta} \\left(y _ {2} \\mid [ x , q _ {2} ]\\right)}{\\pi_ {\\text {r e f}} \\left(y _ {2} \\mid [ x , q _ {2} ]\\right)}\\right). \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.906, + 0.569 + ], + "angle": 0, + "content": "More details about the derivation from Eq. (8) to Eq. (10) could be found in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.573, + 0.906, + 0.635 + ], + "angle": 0, + "content": "As shown in Fig. 4, we reformulate the preference data based on trajectories as \\(\\mathcal{D} = \\{\\tau_1^{(i)},\\tau_2^{(i)},\\tau_3^{(i)},\\tau_4^{(i)}\\}_{i = 1}^N\\). By substituting Eq. (8) into Eq. (9), the loss function of classifying between two trajectories is given as" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.654, + 0.918, + 0.71 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} \\left(\\tau_ {w}, \\tau_ {l}\\right) = - \\mathbb {E} _ {\\left(\\tau_ {w}, \\tau_ {l}\\right) \\sim \\mathcal {D}} \\left[ \\log \\sigma \\left(r \\left(\\tau_ {w}\\right) - r \\left(\\tau_ {l}\\right)\\right) \\right] \\tag {10} \\\\ \\leq \\mathbb {E} _ {(\\tau_ {w}, \\tau_ {l}) \\sim \\mathcal {D}} [ - \\log \\sigma ((r (x, q _ {1}) - r (x, q _ {2}))) ] \\\\ + \\mathbb {E} _ {\\left(\\tau_ {w}, \\tau_ {t}\\right) \\sim \\mathcal {D}} [ - \\log \\sigma \\left(\\left(r ([ x, q _ {1} ], y _ {1}) - r ([ x, q _ {2} ], y _ {2}))\\right) \\right]. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.727, + 0.908, + 0.787 + ], + "angle": 0, + "content": "Eventually, we define the preference probability between sub-tree \\(\\mathcal{V}_w\\) and \\(\\mathcal{V}_l\\) as a joint of multiple trajectory-based preference probability, improving the comprehension of policy distribution. The final SPO loss is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.805, + 0.905, + 0.837 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S P O} = \\mathcal {L} _ {\\left(\\mathcal {Y} _ {w}, \\mathcal {Y} _ {i}\\right)} := \\sum_ {m, n} \\mathcal {L} \\left(\\tau_ {m}, \\tau_ {n}\\right): m < n, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}(\\tau_m,\\tau_n)\\) is given in Eq. (10), and \\(m\\in \\Phi_w = \\{1,3\\}\\), \\(n\\in \\Phi_l = \\{2,4\\}\\) are the index of preferred and dispreferred trajectories." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "745" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Evaluation results of our model trained with different preference optimization methods based on the SFT model on a scale of 0.5B and 7B. We report accuracy on all the benchmarks except MME. The best results are bold, and the second-best results are underlined." + }, + { + "type": "table", + "bbox": [ + 0.141, + 0.129, + 0.861, + 0.374 + ], + "angle": 0, + "content": "
ModelSingle-Image
MMStarMMMUMMEL-WilderLLaVA-WMMVetRealWorldQAPOPEMMBench
testvaltestsmalltesttesttesttestMulti-discip
GPT-4V [26]57.156.8140981.098.049.961.4-75.0
GPT-4o [27]-69.1-85.9106.176.258.6--
Gemini-1.5-Pro [36]-62.2----70.4--
InternVL-2-8B [7]59.449.3221062.584.560.064.4-81.7
IXC-2.5-7B [50]59.942.9222961.478.151.767.8-82.2
LLaVA-OV-72B [18]66.156.8168272.093.563.771.9-85.9
LLaVA-OV-0.5B [18]37.531.4123855.074.229.155.688.052.1
SFT-0.5B [35]40.731.9124659.472.826.755.788.053.1
SFT-0.5B + DPO [29]40.932.1125058.573.525.856.488.154.0
SFT-0.5B + SPO (ours)41.532.8125161.374.130.056.988.254.3
LLaVA-OV-7B [18]61.748.8158067.890.757.566.389.280.8
SFT-7B [35]59.648.0156569.590.557.666.889.079.7
SFT-7B + DPO [29]61.949.0157969.991.258.267.289.180.6
SFT-7B + SPO (ours)62.149.2158570.292.358.867.389.381.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.386, + 0.908, + 0.416 + ], + "angle": 0, + "content": "Table 2. Evaluation results of different methods on multi-image and video benchmarks. #Frames is the number of frames used for video encoding. We report accuracy on all the benchmarks. The best results are bold, and the second-best results are underlined." + }, + { + "type": "table", + "bbox": [ + 0.143, + 0.426, + 0.859, + 0.691 + ], + "angle": 0, + "content": "
Model#FramesMulti-ImageEgoSchemaVideo
\\(LLaVA^I\\)MuirBenchMLVUNExT-QALVid-BenchVideoMME
out-domainout-domaintestm-avgmcvalw/o subs
GPT-4V [26]1060.362.3-49.2-61.359.9
GPT-4o [27]384-68.0-64.6-66.771.9
Gemini-1.5-Pro [36]1/s-49.472.2--64.075.0
LongVA-7B [51]128---56.368.3-52.6
IXC-2.5-7B [50]64---37.371.0-55.8
LLaVA-Video-7B [53]64--57.370.883.258.263.3
LLaVA-Video-72B [53]64--65.674.485.461.970.5
LLaVA-OV-72B [18]3279.954.862.068.080.261.366.3
LLaVA-OV-0.5B [18]3233.325.526.850.357.245.844.0
SFT-0.5B [35]3232.726.128.851.759.044.644.3
SFT-0.5B + DPO [29]3232.927.928.952.059.246.144.8
SFT-0.5B + SPO (ours)3233.428.029.252.460.447.745.4
LLaVA-OV-7B [18]3264.241.860.164.779.456.458.1
SFT-7B [35]3266.845.560.064.977.257.557.8
SFT-7B + DPO [29]3266.946.259.565.179.858.158.3
SFT-7B + SPO (ours)3267.946.760.565.379.657.858.5
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.224, + 0.73 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.74, + 0.251, + 0.758 + ], + "angle": 0, + "content": "4.1. Training Recipe" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.486, + 0.903 + ], + "angle": 0, + "content": "We follow [18] to construct our model architecture, including the instructional language model, vision encoder, and a 2-layer MLP projector. This work adopts LoRA [15] for an efficient and lightweight fine-tuning, where we set low-rank adapters with rank \\(= 64\\) and \\(\\alpha = 128\\) for the LLM. We denote \\(\\theta\\) as all the trainable parameters, including LoRA and projector layer. As shown in Fig. 3, the training pipeline has two stages: 1) Utilizing supervised fine-tuning (SFT) to train an SFT model by reasoning and learning to ques" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.714, + 0.907, + 0.777 + ], + "angle": 0, + "content": "tion [35]. 2) Fine-tune the LVML policy (given the SFT model as a reference) using the proposed SPO approach by maximizing a novel structured reward function. We show the training details for each stage in supplementary." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.737, + 0.803 + ], + "angle": 0, + "content": "4.2. Evaluation Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We primarily assess our models using vision-language benchmarks across different visual modalities (e.g., image, multi-image, and video) that are not included in the training data. For a fair comparison with other LVLMs, we primarily report results from original papers and utilize Immseval [49] for test setting consistency. All our results are" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "746" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.242, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.189, + 0.228, + 0.293 + ], + "angle": 0, + "content": "Are the otters in the image reading a book, and if so, can you provide a detailed description of what might be happening in this whimsical scene?" + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.092, + 0.892, + 0.154 + ], + "angle": 0, + "content": "SPO-7B Based on the image, it appears that the otters are interacting with a book in an unusual and anthropomorphic manner. The otters seem to be holding the book open as if they are reading it, which is a humorous and imaginative interpretation of their behavior. The book's placement between them suggests a shared interest or activity. The scene is whimsical because otters do not engage in such human-like behaviors naturally; this image is likely staged for artistic or entertainment purposes." + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.163, + 0.894, + 0.212 + ], + "angle": 0, + "content": "DPO-7B The otters in the image are likely to be sea otters, as they are shown floating on their backs in water and holding a book between them. The fact that they are holding a book also indicates that they have been exposed to human activities or objects, which could mean they live in an area where humans frequently visit or interact with them." + }, + { + "type": "text", + "bbox": [ + 0.249, + 0.218, + 0.896, + 0.305 + ], + "angle": 0, + "content": "GPT-4V Yes, the otters in the image appear to be reading a book, which is a whimsical and anthropomorphized activity for animals. The scene is set in a body of water, with the two otters floating and seemingly fully absorbed in the book. One otter holds the book with one hand and points to a line or page with the other, as if highlighting something or guiding the reading process. The second otter is focused on the book, with an expression of curiosity or concentration. This playful image brings a charming and imaginative element, showcasing animals engaging in human-like activities and suggesting themes of education, curiosity, and companionship. Reference answer" + }, + { + "type": "list", + "bbox": [ + 0.248, + 0.092, + 0.896, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.317, + 0.908, + 0.36 + ], + "angle": 0, + "content": "Figure 6. Qualitative results of LLaVA-Bench-Wilder. When trained with DPO, the model becomes over-confident when indicating low visual dependency content. In contrast, SPO improves the model's visual reasoning ability and prevents over-confident prediction when explaining uncertain situations, acting similarly to the reference model, GPT-4V." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.386, + 0.425, + 0.401 + ], + "angle": 0, + "content": "reported with greedy sampling and 0-shot settings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.402, + 0.483, + 0.506 + ], + "angle": 0, + "content": "Single-image Benchmarks. We consider the LVLM as a general-purpose assistant in the wild as the most critical metric. To validate the capabilities in real-world scenarios with open-form instructions, we utilize MMStar [5], LLaVA-Bench-Wilder [18], and RealWorldQA [43]. Besides chat ability, visual perception evaluates the model's reasoning ability, so we adopt MME [13] and MMMU [48]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.483, + 0.657 + ], + "angle": 0, + "content": "Multi-image Benchmarks. Multi-image tasks test the model's ability to connect and reason across images, making them generally more challenging as they require higher-level contextual integration. This work utilizes LLaVA-Interleave Bench [20] under the out-of-domain setting that does not overlap with training scenarios. This setting covers 13 challenging real-world tasks such as spotting the difference and visual storytelling, etc. We also provide a comprehensive and holistic evaluation of MuirBench [37], which contains 12 diverse multi-image tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Video Benchmarks. Video understanding is a critical ability to build a general-purpose vision-language assistant. This work conducts several open-ended and multi-choice video benchmarks. EgoSchema [25] and MLVU [54] focus on long video understanding, NExTQA [44], LongVideoBench [40] and VideoMME [14] contains diverse video domains and duration." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.229, + 0.787 + ], + "angle": 0, + "content": "4.3. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In Table 1, while both preference optimization algorithms enhance performance over the SFT model, SPO achieves the best overall performance across all the image benchmarks under 0.5B and 7B scale. The consistency highlights the robustness and effectiveness of SPO. Notably, SPO outperforms its baseline model LLaVA-OV by 6.3 (0.5B) and 2.4 (7B) points on LLaVA-Bench-Wilder, and the improve" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.386, + 0.906, + 0.49 + ], + "angle": 0, + "content": "ment on RealWorldQA indicates that SPO enhances the model's visual reasoning capability while handling freeform visual chat and real-world scenario. Such improvement follows our expectations since we precisely optimize the model to capture the alignment between visual input and dialogue interaction. Fig. 6 shows a qualitative example of real-world visual chat from three models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.906, + 0.658 + ], + "angle": 0, + "content": "We have observed an improvement of SPO on the multi-image benchmarks, demonstrating the significance of our method for improving comprehensive and complex visual reasoning. Specifically, SPO-0.5B achieves a \\(4.4\\%\\) and \\(1.0\\%\\) improvement, and SPO-7B achieves an \\(8.1\\%\\) and \\(2.0\\%\\) improvement over LLaVA-OV and our SFT model. Table 2 also provides evaluation results on video benchmarks. The consistent improvement of SPO-0.5b/ 7B on EgoSchema, MLVU, and L-VideoBench indicates that SPO also improves consecutive visual reasoning, especially for medium and long video understanding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.906, + 0.885 + ], + "angle": 0, + "content": "SPO has better visual-question understanding ability than DPO since it encourages visually informative questions by utilizing the structured preference tree that captures diverse question-answer pairs. Regarding the low-quality questions, SPO also learns their preferred answers through a subtree within the structured preferences. To validate our hypothesis, we prompt GPT-4o to score 216 questions in the MMVet dataset and split them into high-dependency (High-Q) and low-dependency (Low-Q) visual questions. Table 4 shows that more visually focused questions consistently improve the answering accuracy given images. As expected, without looking at images (w/o images), Low-Q reports a smaller performance drop than High-Q, and SPO has lower performance drop than DPO. Empirical evidence shows that SPO is more robust to question quality than DPO." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Limitation. While SPO enhances the SFT model, we" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "747" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.529, + 0.171 + ], + "angle": 0, + "content": "
r(x,q)r([x,q],y)ImageMulti-ImageVideo
MMELLaVA-WilderLLaVA1MuirBenchEgoSchemaMLVU
✓✓158570.267.946.760.565.3
✓X156969.967.346.260.065.2
X✓157169.767.246.360.165.0
XX156569.566.845.560.064.9
" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.093, + 0.904, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.182, + 0.908, + 0.227 + ], + "angle": 0, + "content": "Table 3. Ablation study under SFT-7B-SPO. We ablate the trajectory re- Figure 7. Left. SPO achieves higher preference classification reward design of SPO. The formal trajectory reward is \\(r(\\tau) = r(x,q) +\\) accuracy than DPO throughout training. Right. SPO maxi\\(r([x,q],y)\\), where \\(r(x,q)\\) and \\(r([x,q],y)\\) are question and answer reward. mizes the reward margin faster effectiveness than DPO." + }, + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.249, + 0.472, + 0.262 + ], + "angle": 0, + "content": "Table 4. Test accuracy drop in two question splits on MMVet." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.263, + 0.479, + 0.315 + ], + "angle": 0, + "content": "
ModelsHigh-Q (w/o image)Low-Q (w/o image)
SFT-0.5B + DPO27.0 (6.0) ↓ 77.8%17.0 (6.0) ↓ 64.7%
SFT-0.5B + SPO29.0 (8.0) ↓ 72.4%23.0 (12.0) ↓ 47.8%
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.34, + 0.483, + 0.581 + ], + "angle": 0, + "content": "want to emphasize the critical role of the SFT model as the foundational stage for all preference optimization methods. Notably, the limited computational resources primarily drive the performance decline after SFT on several benchmarks. Our SFT-7B model was fine-tuned with 8 GPUs — only 1/16 of training GPUs in LLaVA-OV — resulting in a much smaller batch size for our approach. Specifically, the performance degradation of SFT-7B on MMStar, MMU, and MME sets a lower bound for the subsequent gains achievable by SPO. A similar trend is observed in video benchmarks, as shown in Table 2. Additionally, we decrease the anyres [18] to \\(60\\%\\) of the regular setting for visual encoding to fit in our computational resource. Due to limited computational resources, SPO mainly focuses on single-turn conversations. We leave multi-turn conversations study in future work." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.592, + 0.263, + 0.606 + ], + "angle": 0, + "content": "4.4. Model Discussion" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.614, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Structured Reward Speeds Up Policy Learning. The Bradley-Terry objective function essentially aims to optimize the reward margin of the preferred and dispreferred action \\(\\Delta r(x, \\rho_w, \\rho_l) = r(x, \\rho_w) - r(x, \\rho_l)\\). In Eq. (10), SPO optimizes a joint reward margin including \\(\\Delta r(x, q_1, q_2)\\) and \\(\\Delta r([x, q_1], y_1, y_2)\\). While DPO only optimizes the reward margin between responses \\(\\Delta r([x, q_1], y_1, y_2)\\). We use the SPO and DPO models to investigate the relationship between the learned reward margin and the preference classification accuracy. Fig. 7 shows that SPO maximizes the reward margin more effectively than DPO, and SPO achieves higher preference classification accuracy than DPO along all the training steps. By precisely adjusting \\(\\beta\\) for different preference probabilities, we find that optimizing multiple trajectory pairs together can explicitly increase the data diversity, leading to better generalizability and performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Influence of Each Step Reward. SPO optimizes a policy to satisfy multiple preferences in a single objective function. In Eq. (10), we use \\(\\lambda\\) and \\(\\delta\\) to control the deviation of" + }, + { + "type": "table_caption", + "bbox": [ + 0.545, + 0.249, + 0.875, + 0.262 + ], + "angle": 0, + "content": "Table 5. Latency vs Accuracy on the LLaVA-W dataset" + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.263, + 0.898, + 0.314 + ], + "angle": 0, + "content": "
MetricSFT-7BSFT-7B + DPOSFT-7B + SPO
Accuracy (%)72.873.574.1
Inference time (s)523547254
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.328, + 0.905, + 0.465 + ], + "angle": 0, + "content": "preference alignment over two action groups, questioning and answering. Following the same training strategy and preference data, we evaluate the performance of SPO w/o question reward \\( r(x,q) = 0 \\) and SPO w/o answer reward \\( r([x,q] = y) = 0 \\). As shown in Table. 3, learning question policy improves real-world QA ability such as LLaVA-Wilder and LLaVA; learning answer policy brings more improvement over fixed-form QA such as MME. Overall, the joint policy improves the model from different aspects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.466, + 0.906, + 0.616 + ], + "angle": 0, + "content": "Computational Resource Analysis. For training, optimizing the whole dialogue will not burden the computation cost when comparing SPO with DPO (e.g., SPO: 24h vs DPO: 22h using \\(4^{*}\\mathrm{A}100\\) GPUs) since only the question tokens need to be unmasked. For inference, Table 5 evaluates latency by reporting the total inference time of three models on 60 open-ended questions from the LLaVA-W dataset, and the test accuracy is measured by GPT-4o. Empirically, SPO generates more concise and accurate content than the SFT and DPO models." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.633, + 0.634, + 0.649 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this work, we have proposed Structured Policy Optimization (SPO), a novel approach to enhancing multi-modal preference optimization for large vision-language models (LVLMs). SPO alleviates the limitation of existing DPO methods, which adopt a single, unstructured reward and tend to prioritize linguistic nuances over visual context. Specifically, SPO introduces a new structured reward represented in trajectories to improve the visual dependency by maximizing the reward between preferred and dispreferred questions/answers. Our experimental results demonstrate that SPO improves vision-language understanding and reasoning across diverse tasks, showcasing superior alignment with human-preferred responses compared to traditional DPO methods. Overall, this approach better captures the dynamic interplay between visual inputs, questions, and responses, aligning more effectively with human preferences." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "748" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.271, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.114, + 0.484, + 0.144 + ], + "angle": 0, + "content": "This research was supported by the DEVCOM Army Research Laboratory under Contract W911QX-21-D-0001." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.171, + 0.188, + 0.187 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.196, + 0.484, + 0.25 + ], + "angle": 0, + "content": "[1] OpenAI Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, et al. Gpt-4 technical report. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.253, + 0.484, + 0.322 + ], + "angle": 0, + "content": "[2] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.324, + 0.483, + 0.378 + ], + "angle": 0, + "content": "[3] Hritik Bansal, Ashima Suvarna, Gantavya Bhatt, Nanyun Peng, Kai-Wei Chang, and Aditya Grover. Comparing bad apples to good oranges: Aligning large language models via joint preference optimization. ArXiv, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.38, + 0.482, + 0.434 + ], + "angle": 0, + "content": "[4] Lin Chen, Jinsong Li, Xiao wen Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. In ArXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.436, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[5] Lin Chen, Jinsong Li, Xiao wen Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, and Feng Zhao. Are we on the right way for evaluating large vision-language models? In NeurIPS, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.507, + 0.482, + 0.548 + ], + "angle": 0, + "content": "[6] Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji, and Quanquan Gu. Self-play fine-tuning converts weak language models to strong language models. ArXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.55, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[7] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.634, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[8] Paul F. Christiano, Jan Leike, Tom B. Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences. In Proceedings of the 31st International Conference on Neural Information Processing Systems, 2017. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[9] Ruiguo Cui and Peter Teo. Thinking through talk: Using dialogue to develop students' critical thinking. Teaching and Teacher Education, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[10] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Albert Li, Pascale Fung, and Steven C. H. Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[11] Yihe Deng, Pan Lu, Fan Yin, Ziniu Hu, Sheng Shen, James Zou, Kai-Wei Chang, and Wei Wang. Enhancing large vision language models with self-training on image comprehension. In NeurIPS, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[12] Hanze Dong, Wei Xiong, Deepanshu Goyal, Rui Pan, Shizhe Diao, Jipeng Zhang, Kashun Shum, and T. Zhang. Raft:" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.196, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Reward ranked finetuning for generative foundation model alignment. ArXiv, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[13] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. ArXiv, 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[14] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, Peixian Chen, Yanwei Li, Shaohui Lin, Sirui Zhao, Ke Li, Tong Xu, Xiawu Zheng, Enhong Chen, Rongrong Ji, and Xing Sun. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. ArXiv, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.291, + 0.905, + 0.345 + ], + "angle": 0, + "content": "[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.348, + 0.904, + 0.389 + ], + "angle": 0, + "content": "[16] Dongfu Jiang, Xuan He, Huaye Zeng, Cong Wei, Max W.F. Ku, Qian Liu, and Wenhu Chen. Mantis: Interleaved multi-image instruction tuning. ArXiv, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.391, + 0.905, + 0.444 + ], + "angle": 0, + "content": "[17] Xin Lai, Zhuotao Tian, Yukang Chen, Senqiao Yang, Xiangru Peng, and Jiaya Jia. Step-dpo: Step-wise preference optimization for long-chain reasoning of llms. ArXiv, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.447, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[18] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. ArXiv, 2024. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.504, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[19] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. In NeurIPS, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.575, + 0.905, + 0.628 + ], + "angle": 0, + "content": "[20] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava-last-interleave: Tackling multi-image, video, and 3d in large multimodal models. ArXiv, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.631, + 0.905, + 0.7 + ], + "angle": 0, + "content": "[21] Juncheng Li, Kaihang Pan, Zhiqi Ge, Minghe Gao, Hanwang Zhang, Wei Ji, Wenqiao Zhang, Tat seng Chua, Siliang Tang, and Yueting Zhuang. Fine-tuning multimodal llms to follow zero-shot demonstrative instructions. In International Conference on Learning Representations, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.702, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[22] Lei Li, Zhihui Xie, Mukai Li, Shunian Chen, Peiyi Wang, Liang Chen, Yazheng Yang, Benyou Wang, and Lingpeng Kong. Silkie: Preference distillation for large visual language models. ArXiv, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.759, + 0.905, + 0.799 + ], + "angle": 0, + "content": "[23] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. ArXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.802, + 0.905, + 0.841 + ], + "angle": 0, + "content": "[24] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In ArXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.844, + 0.905, + 0.886 + ], + "angle": 0, + "content": "[25] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. ArXiv, 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.888, + 0.83, + 0.901 + ], + "angle": 0, + "content": "[26] OpenAI. Gpt-4v(ision) system card. 2023. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "749" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.384, + 0.106 + ], + "angle": 0, + "content": "[27] OpenAI. Gpt-4o system card. 2024. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.108, + 0.482, + 0.216 + ], + "angle": 0, + "content": "[28] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke E. Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Francis Christiano, Jan Leike, and Ryan J. Lowe. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[29] Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. In NeurIPS, 2023. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.275, + 0.482, + 0.316 + ], + "angle": 0, + "content": "[30] Rafael Rafailov, Joey Hejna, Ryan Park, and Chelsea Finn. From r to q: Your language model is secretly a q-function. In First Conference on Language Modeling, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.317, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[31] Corby Rosset, Ching-An Cheng, Arindam Mitra, Michael Santacroce, Ahmed Awadallah, and Tengyang Xie. Direct nash optimization: Teaching language models to self-improve with general preferences. ArXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.551 + ], + "angle": 0, + "content": "[32] Avi Singh, John D. Co-Reyes, Rishabh Agarwal, Ankesh Anand, Piyush Patil, Peter J. Liu, James Harrison, Jaehoon Lee, Kelvin Xu, Aaron T Parisi, Abhishek Kumar, Alex Alemi, Alex Rizkowsky, Azade Nova, Ben Adlam, Bernd Bohnet, Hanie Sedghi, Igor Mordatch, Isabelle Simpson, Izzeddin Gur, Jasper Snoek, Jeffrey Pennington, Jiri Hron, Kathleen Kenealy, Kevin Swersky, Kshitteej Mahajan, Laura Culp, Lechao Xiao, Maxwell Bileschi, Noah Constant, Roman Novak, Rosanne Liu, Tris Brian Warkentin, Yundi Qian, Ethan Dyer, Behnam Neyshabur, Jascha Narain Sohl-Dickstein, and Noah Fiedel. Beyond human data: Scaling self-training for problem-solving with language models. Trans. Mach. Learn. Res., 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[33] Feifan Song, Yu Bowen, Minghao Li, Haiyang Yu, Fei Huang, Yongbin Li, and Houfeng Wang. Preference ranking optimization for human alignment. ArXiv, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[34] Guohao Sun, Can Qin, Huazhu Fu, Linwei Wang, and Zhiqiang Tao. Stllava-med: Self-training large language and vision assistant for medical. In EMNLP, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[35] Guohao Sun, Can Qin, Jiamian Wang, Zeyuan Chen, Ran Xu, and Zhiqiang Tao. Sq-llava: Self-questioning for large vision-language assistant. In ECCV, 2024. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[36] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[37] Fei Wang, Xingyu Fu, James Y. Huang, Zekun Li, Qin Liu, Xiaogeng Liu, Mingyu Derek Ma, Nan Xu, Wenxuan Zhou, Kai Zhang, Tianyi Yan, Wenjie Jacky Mo, Hsiang-Hui Liu, Pan Lu, Chunyuan Li, Chaowei Xiao, Kai-Wei Chang, Dan Roth, Sheng Zhang, Hoifung Poon, and Muhao Chen. Muir-bench: A comprehensive benchmark for robust multi-image understanding. ArXiv, 2024. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[38] Fei Wang, Wenxuan Zhou, James Y. Huang, Nan Xu, Sheng Zhang, Hoifung Poon, and Muhao Chen. mdpo: Conditional preference optimization for multimodal large language models. In EMNLP, 2024. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[39] Yuetian Weng, Mingfei Han, Haoyu He, Xiaojun Chang, and Bohan Zhuang. Longvlm: Efficient long video understanding via large language models. ArXiv, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[40] Haoning Wu, Dongxu Li, Bei Chen, and Junnan Li. Longvideobench: A benchmark for long-context interleaved video-language understanding. ArXiv, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[41] Junkang Wu, Yuexiang Xie, Zhengyi Yang, Jiancan Wu, Jinyang Gao, Bolin Ding, Xiang Wang, and Xiangnan He. \\(\\beta\\)-dpo: Direct preference optimization with dynamic \\(\\beta\\). ArXiv, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[42] Yue Wu, Zhiqing Sun, Huizhuo Yuan, Kaixuan Ji, Yiming Yang, and Quanquan Gu. Self-play preference optimization for language model alignment. ArXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.905, + 0.299 + ], + "angle": 0, + "content": "[43] x.ai. Grok-1.5 vision preview. Technical report, x.ai, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.303, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[44] Junbin Xiao, Xindi Shang, Angela Yao, and Tat seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[45] Tianyi Xiong, Tianyi Xiong, Xiyao Wang, Dong Guo, Qinghao Ye, Haoqi Fan, Quanquan Gu, Heng Huang, and Chunyuan Li. Llava-critic: Learning to evaluate multimodal models. ArXiv, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.415, + 0.905, + 0.481 + ], + "angle": 0, + "content": "[46] Enwei Xu, Wei Wang, and Qingxia Wang. The effectiveness of collaborative problem solving in promoting students' critical thinking: A meta-analysis based on empirical literature. *Humanities and Social Sciences Communications*, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[47] Xubing Ye, Yukang Gan, Xiaoke Huang, Yixiao Ge, Ying Shan, and Yansong Tang. Voco-llama: Towards vision compression with large language models. 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.527, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[48] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Ren-liang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.652, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[49] Kaichen Zhang, Bo Li, Peiyuan Zhang, Fanyi Pu, Joshua Adrian Cahyono, Kairui Hu, Shuai Liu, Yuanhan Zhang, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Lmmseval: Reality check on the evaluation of large multimodal models. ArXiv, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[50] Pan Zhang, Xiao wen Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, Songyang Zhang, Wenwei Zhang, Yining Li, Yang Gao, Peng Sun, Xinyue Zhang, Wei Li, Jingwen Li, Wenhai Wang, Hang Yan, Conghui He, Xingcheng Zhang, Kai Chen, Jifeng Dai, Yu Qiao, Dahua Lin, and Jiaqi Wang. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. ArXiv, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[51] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. In ArXiv, 2024. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "750" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[52] Yanzhe Zhang, Ruiyi Zhang, Jiumiang Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tongfei Sun. Llavar: Enhanced visual instruction tuning for text-rich image understanding. In ArXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[53] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. ArXiv, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[54] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. ArXiv, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[55] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. ArXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.359 + ], + "angle": 0, + "content": "[56] Daniel M. Ziegler, Nisan Stiennon, Jeff Wu, Tom B. Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. ArXiv, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.51, + 0.957 + ], + "angle": 0, + "content": "751" + } + ] +] \ No newline at end of file diff --git a/2025/Structured Policy Optimization_ Enhance Large Vision-Language Model via Self-referenced Dialogue/5161fa7a-2cd7-4a41-959a-e1ddb02c88dc_origin.pdf b/2025/Structured Policy Optimization_ Enhance Large Vision-Language Model via Self-referenced Dialogue/5161fa7a-2cd7-4a41-959a-e1ddb02c88dc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ffe787a9300f050950d7706a63c8c4a7ed01dfc6 --- /dev/null +++ b/2025/Structured Policy Optimization_ Enhance Large Vision-Language Model via Self-referenced Dialogue/5161fa7a-2cd7-4a41-959a-e1ddb02c88dc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b334992998da0abd883a68a20d0b26359d0f367e0e8d9125a34d39f5d9bd99de +size 1363181 diff --git a/2025/Structured Policy Optimization_ Enhance Large Vision-Language Model via Self-referenced Dialogue/full.md b/2025/Structured Policy Optimization_ Enhance Large Vision-Language Model via Self-referenced Dialogue/full.md new file mode 100644 index 0000000000000000000000000000000000000000..35573f5ced8ea82cb03436de4bfd4a1b0905d5c1 --- /dev/null +++ b/2025/Structured Policy Optimization_ Enhance Large Vision-Language Model via Self-referenced Dialogue/full.md @@ -0,0 +1,341 @@ +# Structured Policy Optimization: Enhance Large Vision-Language Model via Self-Referenced Dialogue + +Guohao Sun1*, Can Qin2, Yihao Feng2, Zeyuan Chen2, Ran Xu2, Sohail Dianat1, Majid Rabbani1, Raghuveer Rao3, and Zhiqiang Tao1* + +$^{1}$ Rochester Institute of Technology + $^{2}$ Salesforce AI Research $^{3}$ DEVCOM Army Research Laboratory + +# Abstract + +Preference optimization algorithms typically enhance LLM response quality by leveraging human feedback on multiple answers given a fixed instruction. However, these methods often lack capturing the dynamic nature of conversational exchanges. For large vision-language models (LVLMs), direct preference optimization (DPO) can over-emphasize linguistic nuances while overlooking visual context. To address this challenge, we introduce structured policy optimization (SPO) - a novel preference optimization method that simultaneously aligns preference instructions, responses, and dialogue interactions to improve multi-modal understanding and reasoning capabilities. The efficacy of SPO is attributed to one key design: treating the questioning and answering as a sequential action and binding them through a trajectory reward. This reward formulation better aligns with real-world dialogue studies and eliminates the need for fixed instructions. We evaluate our models on interleaved benchmarks, including image, multi-image, and video-based understanding and reasoning tasks. Experimental results show that the proposed SPO finetuning LVLM with multi-modal preference data can align with human preference more efficiently than DPO. The code is available at https://github.com/heliossun/Structure-Policy-Optimization. + +# 1. Introduction + +Reinforcement learning from human feedback (RLHF) [8] has demonstrated remarkable success in aligning large language/vision-language models (LLMs/LVLMs) with human preferences, enabling models to adapt their behavior based on success or failure in increasingly complex tasks. Typically, RLHF trains a reward model and optimizes the + +![](images/28ef7d25bcde55ac0d4bf2ed9baa241070cc4861884b8828d0b52146a6288640.jpg) + +![](images/ad1d1f1bef0d013fb89038fb4681c4eea37330537c2d1a9c71e75d4d85654f1d.jpg) +Figure 1. Comparison between DPO and SPO in optimizing an LVLM policy. DPO employs a single reward that optimizes the preferred answer given the fixed visual & question input, $p(y_w \succ y_l|x, q)$ , which could over-emphasize language nuance and overlook visual clues [38]. By contrast, SPO introduces a structured reward by $p(\mathcal{Y}_w \succ \mathcal{Y}_l|x)$ , conditioning on the visual dialogue, to optimize the preferred trajectories spanning over questions and answers. + +![](images/e9a5b624afff8f00463fd23650074968d0cf4410631b8b4ac0882894f1177bed.jpg) + +![](images/4a691cd52fd51a7243c09ab7c443d480f16e0c59e0a8f8621a230b62240094fd.jpg) + +LLM/LVLM policy to maximize the reward, which is applied in both language and vision-language (VL) domains. + +Recently, the offline preference optimization methods, such as DPO [29] and its variants [12, 17, 30, 33], have proposed various reward objectives to satisfy different human expectations. Owing to its simplicity and stability, DPO has been widely used in multi-modal preference optimization [11, 19, 22, 38] by aligning LVLM to human-preferred responses conditioned on fixed visual and text instructions (see Fig. 1). However, unlike purely text dialogue, the visual input can lead to far more enriched information interactions, which require varying degrees of visual dependence [13, 25, 37]. Directly applying DPO to multi-modal preference data (e.g., images/videos and texts) will over-emphasize language nuance and overlook the visual condition [38]. Specifically, DPO applies a single, un + +![](images/62500da39ba19fe7e3f00edccc19d9cdc47c01c3e09e2d88141948b140e5de27.jpg) +(a) Question vs Answer. + +![](images/ac2b3fef8156fe58734ef89f170e5d8052903d30085bb4c5149f4acce882455d.jpg) +(b) Score distribution. +Figure 2. Scoring visual dialogue with GPT-4o as judge. (a) Questions with higher visual dependency (x-axis, scoring from 4 to 10, $\uparrow$ ) tend to produce answers with higher quality (y-axis, scoring from 0 to 9, $\uparrow$ ). (b) Question and answer score distributions. + +structured reward function to maximize the gap between the preferred and dispreferred answer upon the same input sequence; however, this gap might be due to the lack of visual understanding [5], which is generally ignored by existing DPO methods, especially for LVLMs. + +Inspired by the dialogue studies [9, 46], we empirically observe that asking a good visual-related question can turn passive learning into active exploration, allowing for deeper visual understanding and broader knowledge exploration before answering. For instance, as shown in Fig. 1, people prefer the question "What emotions are conveyed through the shots in the video?" to "Which city did the skateboarder go to in the video?" since the former is more visually dependent and the answer will cover more visual details. To quantitatively validate our observation, we apply GPT-4o to evaluate the visual relevance (scoring from 4 to 10) of questions and score the quality of answers (higher is better) in our curated preference data based on the instruction dataset in [18]. Fig. 2a shows that human-preferred questions with higher visual relevance tend to produce responses with higher quality, which is further demonstrated by the consistent question/answer score distributions in Fig. 2b. We introduce this correlation between visual questions and answer quality as a preference dialogue interaction. By aligning the LVLM policy to generate questions and answers with human preference, we build a more in-depth alignment between vision and dialogues, enabling a better VL understanding for broad tasks. + +In this study, we propose Structured Policy Optimization (SPO), a novel structured DPO method designed to enhance general vision-language understanding and reasoning. As shown in Fig. 1, SPO utilizes a binary tree structure to track visual dialogue changes, alleviating the limitation of using fixed questions in DPO. Unlike DPO, which only optimizes preference at leaf nodes, the proposed SPO optimizes preferences across different trajectories among sub-trees and develops a new structured reward model by leveraging trajectories to capture the dialogue interaction between the visual input, questions, and answers. Additionally, SPO controls the deviation of question reward and an- + +swer reward by imposing importance weights on each step within the trajectory, which is substantially different from a single-level joint reward optimization [3]. To realize SPO training and increase data diversity, we utilize supervised fine-turning (SFT) to instruct an LVLM to self-generate the preference question-answer pairs, mitigating the distribution shift between the policy model and the preference optimization process [29]. We summarize the contributions of this work as follows. + +- For the first time in multi-modal preference optimization, we introduce SPO, a method designed for optimizing structured reward for aligning the policy model to human preferences over visual dialogues. +- We present high-quality multi-modal preference data consisting of scoring over visual questions and answers. This dataset reflects the discrepancy between chosen and rejected dialogue, bringing insights into real-world conversational exchange and dialogue studies. +- Extensive experiments show that SPO aligns with human preference effectively and overcomes the low visual dependency issue in DPO, leading to better performance on the visual chat benchmarks with free-form instructions. + +# 2. Related Works + +Generative Large Vision-Language Model. Recent general-purpose multi-modal models (LVLMs) have shown impressive capability. The commercial proprietary LVLMs, such as GPT-4V [26], GPT-4o [27], and Gemini [36] exhibit excellent generalize ability in various domains [48] such as design, business, science, medical, and science etc. In the open research community, the evolutionary trajectory of LVLMs has also led to subsequent advancements in different scenarios. The pioneering work of LLaVA [24] improves the multi-modal reasoning ability through visual instruction tuning. [4, 10, 24, 34, 35, 55] pushing the performance limits in single-image scenarios. [16, 21] explores multi-image tasks. While video LVLMs [23, 39, 47, 53] aim to alleviate significant challenges due to the long context issue. In contrast to these single-modality models, the latest LLaVA models [18, 20] demonstrate emerging capabilities across all scenarios after training on interleaved data. Following this path, this work aims to improve the general ability of multiple modalities by primarily focusing on improving the model's reasoning ability. + +Self-generated Synthetic Data. Instruction tuning is widely used in supervised fine-tuning (SFT) to improve general-purpose reasoning. In recent research, synthetic instructional data has become increasingly popular and considered a proxy for human data. This approach primarily leverages the ability of GPT family, [19, 24, 52] utilizes GPT-4 [1] to transform image description into conversations, [4, 18, 20] leverages GPT-4V [26] to process the multi-modal input and generate question-answer pairs. + +![](images/1abbdc8b960da4819ee8a025be615452d45451d361a9de3b173279516748f4f5.jpg) +Stage 1: Reasoning and Learning to Question +Stage2: Preference revelation +Figure 3. Illustration of LVLM model architecture and training pipeline in the proposed SPO. In Stage 1, we implement supervised fine-tuning through learning to question [35], resulting in an SFT model that enables better visual reasoning and self-generated questions. In Stage 2, we first prompt the SFT model to generate visual dialogues labeled by GPT-4o with human preference. Then, we apply the proposed SPO approach to optimize the LVLM policy through a structured reward, leading to better VL understanding. + +However, large-scale synthetic data generation comes with unexpected API costs. To overcome this, [34] train the model with a visual-questioning objective function [35] and generate the question-answer pairs efficiently. While most works mentioned above consider one single optimization procedure, self-play fine-tuning methods [6, 31, 32, 42] generate new synthetic data within the training loop. This work focuses primarily on single-turn optimization and avoids iterative processes. + +Preference Optimization Objectives. Reinforcement Learning from Human Feedback (RLHF) is a technique used to align large language models with human preferences and values [2, 8, 28, 56]. To eliminate the complexity and challenges of optimizing online preference algorithms, Direct Preference Optimization (DPO) [29] is a prominent example. To extend two-instance classification, ranking objectives enable comparisons across multiple instances [12, 17, 30, 33]. Recent efforts in multi-modal preference learning mainly focus on preference data design [11, 19, 22, 38], while utilizing the formal DPO objective function. This work proposes a novel structured DPO by maximizing the structured rewards to improve vision-language alignment. + +# 3. SPO: Structured Policy Optimization + +Preference optimization uses explicit feedback (e.g., human ratings, comparisons, or rewards) to fine-tune the SFT model, making it more responsive to real-world user needs. As shown in Fig. 3, this work employs a two-stage training strategy: 1) Utilizing supervised fine-tuning (SFT) to train a SFT model by reasoning and learning to question [35], which is able to generate diverse visual dialogues that align with the training data's specific patterns and content. 2) + +Fine-tune the LVML policy (given the SFT model as a reference) using the proposed SPO approach by maximizing a novel structured reward function. In the following, we will first provide preliminaries about preference optimization and then elaborate on our proposed SPO approach. + +# 3.1. Backgrounds: RLHF and DPO + +$\mathbf{RLHF}[8]$ first trains a reward model $r_{\phi}(x,y)$ parametrized by $\phi$ to predict the human preference distribution $p$ over two responses $y_{1}$ and $y_{2}$ (e.g., answers given by an LVLM) with condition $x$ , simulated by the Bradely-Terry (BT) model as + +$$ +\begin{array}{l} p \left(y _ {1} \succ y _ {2} \mid x\right) = \frac {\exp \left(r _ {\phi} \left(x , y _ {1}\right)\right)}{\exp \left(r _ {\phi} \left(x , y _ {1}\right)\right) + \exp \left(r _ {\phi} \left(x , y _ {2}\right)\right)} \tag {1} \\ = \sigma \left(r _ {\phi} (x, y _ {1}) - r _ {\phi} (x, y _ {2})\right), \\ \end{array} +$$ + +where $x$ denotes visual input and questions (instructions) in this study. After training, the reward model $r_{\phi}$ provides feedback for each response generated by policy model $\pi_{\theta}$ , which is optimized by the following objective function: + +$$ +\begin{array}{l} \max _ {\pi_ {\theta}} \mathbb {E} _ {x \sim D, y \sim \pi_ {\theta} (y | x)} [ r _ {\phi} (x, y) ] \tag {2} \\ - \beta \mathbb {D} _ {K L} \left[ \pi_ {\theta} (y | x) \| \pi_ {r e f} (y | x) \right], \\ \end{array} +$$ + +where $\mathcal{D} = \{x^{(i)},y_w^{(i)},y_l^{(i)}\}_{i = 1}^N$ is the preference dataset, $\pi_{\theta}$ is the policy model that need to be updated, and $\pi_{ref}$ is a reference model (e.g., an SFT model) that avoids overshifting to preferences. $y_{w}$ and $y_{l}$ refer to preferred and dispreferred answers such as $y_{1}$ and $y_{2}$ in Eq. (1). + +DPO [29] reformulates the objective function of RLHF by changing from loss over reward function to optimal policy. Previous work [29] derives the optimal policy $\pi_{\theta}$ of Eq. (2) as + +$$ +\pi_ {\theta} (y | x) = \frac {1}{Z (x)} \pi_ {r e f} (y | x) \exp \left(\frac {1}{\beta} r (x, y)\right), \tag {3} +$$ + +where $Z(x) = \sum_{y} \pi_{\text{ref}}(y|x) \exp \left( \frac{1}{\beta} r(x,y) \right)$ is the marginal likelihood over all possible model responses. To simplify it, DPO reforms Eq. (3) as + +$$ +r (x, y) = \beta \left(\log \frac {\pi_ {\theta} (y \mid x)}{\pi_ {r e f} (y \mid x)} + \log Z (x)\right). \tag {4} +$$ + +The BT model for preference estimation depends only on the difference of rewards between two responses $y_{1}$ and $y_{2}$ [29]. By substituting Eq. (4) into Eq. (1), DPO expresses the probability of human preference data in terms of the optimal policy and formulating the objective function for a parameterized policy $\pi_{\theta}$ as + +$$ +- \mathbb {E} _ {(x, y _ {w}, y _ {l}) \sim \mathcal {D}} \left[ \log \sigma \left(\beta \log \frac {\pi_ {\theta} (y _ {w} | x)}{\pi_ {r e f} (y _ {w} | x)} - \beta \log \frac {\pi_ {\theta} (y _ {l} | x)}{\pi_ {r e f} (y _ {l} | x)}\right) \right], +$$ + +where $\sigma$ represents the sigmoid function and $\beta$ is a scalable value to control the deviation from the reference model. + +Motivation. Nevertheless, DPO methods mainly focus on optimizing the LLM policy by comparing multiple generations $y$ based on a fixed input $x$ , falling short in capturing the dynamics of visual dialogue interactions. Specifically, high visual-dependent instructions require better visual understanding and reasoning ability to respond than easy-to-answer questions [9, 46]. Such discrepancy leads to responses with different amounts of visual clues. To quantitatively assess the visual dependency of each dialogue interaction, we utilized GPT-4o to review visual dialogues according to a set of manually designed rules. As shown in Fig. 2b, instructions with higher visual dependency tend to result in responses with more relevance to the visual input, leading to a preference dialogue interaction. + +Empirical evidence [35] suggests that training the policy to align with preference interaction can enhance visual understanding by maximizing the likelihood of preferred questions and answers. To this end, we propose Structured Policy Optimization (SPO). In SPO, we consider question and answer as two independent actions, each leading to a different state, simulating the conversational exchange. As shown in Fig. 1, the policy can be naturally modeled as a binary tree with depth 2. Unlike DPO, which only optimizes preference at the leaf nodes, SPO optimizes preferences across the entire sub-tree. + +# 3.2. SPO Aligns With Dialogue Interaction + +Preference Data Curation. To implement SPO, we first prompt the SFT model (given in Stage 1 in Fig. 3) to generate visual instruction-answer pairs, aiming to mitigate the distribution shift between SFT and the preference optimization step [3, 29]. Notably, similar to DPO, the SFT model also works as a reference policy $\pi_{\mathrm{ref}}$ in SPO. However, the difference lies in our $\pi_{\mathrm{ref}}$ being capable of self-generating questions by implementing SFT through learning to question [35]. As shown in Fig. 4, we represent different visual dialogues as trajectories in a tree structure - the + +![](images/1ba21977451391ad8cea23ef9581256a7a7ecb2d2635a6040255d715911e0625.jpg) +Figure 4. Illustration of a tree structure that represents different visual dialogues, which are denoted as trajectory, e.g., $\tau_{1} = (x, q_{w}, y_{w}^{1})$ . For preference labeling, we utilize GPT-4o as a judge to perform pair-wise scoring over trajectory pairs. + +input $x$ denotes the root, questions occupy the first level, and answers are leaf nodes - and collect preference data $\mathcal{D} = \{x^{(i)},\mathcal{Y}_w^{(i)},\mathcal{Y}_l^{(i)}\}_{i = 1}^N$ by sub-trees*, where the preferred sub-tree $\mathcal{V}_w = (q_w,y_w^1,y_l^1)$ and dispreferred subtree $\mathcal{V}_l = (q_l,y_w^2,y_l^2)$ . We sample questions $q_{w},q_{l}\sim \pi_{\mathrm{ref}}$ for each visual input $x$ , followed by answers $y_w^1,y_l^1\sim$ $\pi_{\mathrm{ref}}(\cdot |x,q_w))]$ and $y_w^2,y_l^2\sim \pi_{\mathrm{ref}}(\cdot |x,q_l))]$ , where the superscripts 1 and 2 indicate conditioning on $q_{w} / q_{l}$ , respectively. + +Empirically, the preference label accuracy is critical to increasing training stability. Following [45], AI-as-judge can offer scalable solutions to reduce human labor in complex task assessments and generate effective reward signals in reinforcement learning. Thus, to evaluate the synthetic data quality, we compared the distribution of 200 instruction-answer scores judged by GPT-4o [27] and an open-sourced LVLM (LLaVA-critic [45]). Fig. 5 shows that GPT-4o tends to judge with a higher standard than LLaVA-critic as its mean score is lower. We find that GPT-4o can provide scores with higher diversity than LLaVA-critic, acting more like a human annotator. Therefore, this work utilizes GPT-4o for all the preference labeling, and we propose a simple rule to filter out low-quality samples after collecting all the reviews. Specifically, we first remove samples with scores lower than a threshold, then remove the samples with repeated questions or answers, such as $q_{w} == q_{l}$ or $y_{w}^{1} == y_{l}^{1}$ . By this means, our collected preference data maintains high diversity and reliability. + +Structured Reward Formulation. Given the initial state $x$ , the preferred sub-tree $\mathcal{V}_w$ should has higher reward since we assume the preference instruction dominates the dialogue interaction. We apply the same Bradley-Terry probability model to ensure that the reward of the preferred sub-tree $\mathcal{V}_w$ exceeds the reward of the dispreferred sub-tree $\mathcal{V}_l$ by + +$$ +p \left(\mathcal {Y} _ {w} \succ \mathcal {Y} _ {l} \mid x\right) = \sigma \left(r \left(x, \mathcal {Y} _ {w}\right) - r \left(x, \mathcal {Y} _ {l}\right)\right). \tag {5} +$$ + +However, directly deriving $r(x,\mathcal{Y})$ into Eq. (2) will intro + +duce a complex term $Z(x)$ (please refer to the supplementary material for a complete derivation). To overcome this issue, we decompose sub-trees $\mathcal{V}$ to trajectories $\tau$ and show how it connects to state-action trajectories in RLHF [8]. Let $\tau = (s_1, a_1, s_2, a_2, \ldots, s_T, a_T)$ be a trajectory sampled from the policy $\pi_{\theta}$ , where $s, a, T$ denote states, actions, and total steps. The total reward of a trajectory is defined as $r(\tau) = \sum_{t=1}^{T} r(s_t, a_t)$ . Thus, following this reward formulation, we can define a trajectory to represent one visual dialogue by setting $T = 2$ , where the initial state and action are input $x$ and questioning $s_1 = x, a_1 = q$ , and the following state and action are $s_2 = [x, q], a_2 = y$ , e.g., $\tau_1 = (x, q_w, [x, q_w], y_w^1), \tau_2 = (x, q_w, [x, q_w], y_l^1)$ , etc. Finally, we implement Eq. (5) by a trajectory reward as + +$$ +r (x, \mathcal {Y}) \sim r (\tau) = \lambda \cdot r (x, q) + \delta \cdot r ([ x, q ], y), \tag {6} +$$ + +where $\lambda$ and $\delta$ are hyperparameters used to balance the question and answer rewards. By setting $\lambda = \delta$ , our objective function could be equivalent to $\sigma(r(\tau_1) - r(\tau_2))$ , direct preference optimization (DPO) over two trajectory (i.e., question-answer pairs), rather than on how it is split across the binary tree. In the next section, we derive the closed-form expression of $r(x, q)$ and $r([x, q], y)$ . + +# 3.3. The SPO Objective Function + +Following Eq. (2) and our trajectory-based reward model, the general RL objective of SPO is given as + +$$ +J (\pi_ {\theta}) = \mathbb {E} _ {\tau \sim \pi_ {\theta}} [ r (\tau) ] - \beta \mathrm {K L} [ \pi_ {\theta} (\tau) | | \pi_ {\text {r e f}} (\tau) ]. \tag {7} +$$ + +By substituting Eq. (6) into Eq. (7), we have $J(\pi_{\theta})$ as + +$$ +\begin{array}{l} \mathbb {E} _ {(x, q, y) \sim \tau} [ \lambda r (x, q) + \delta r ([ x, q ], y) ] - \beta (\mathrm {K L} [ \pi_ {\theta} (\cdot | x) | | \pi_ {\operatorname {r e f}} (\cdot | x) ]) \\ \left. + \mathrm {K L} \left[ \pi_ {\theta} (\cdot | [ x, q ]) \mid \mid \pi_ {\text {r e f}} (\cdot | [ x, q ])) \right]\right), \\ = \mathbb {E} _ {(x, q) \sim \tau} [ \lambda r (x, q) ] - \beta \mathrm {K L} [ \pi_ {\theta} (\cdot | x) | | \pi_ {\operatorname {r e f}} (\cdot | x)) ] \\ + \mathbb {E} _ {(x, q, y) \sim \tau} [ \delta r ([ x, q ], y) ] - \beta \mathrm {K L} [ \pi_ {\theta} (\cdot | [ x, q ]) | | \pi_ {\text {r e f}} (\cdot | [ x, q ]) ] , \\ \end{array} +$$ + +where $q\sim \pi_{\theta}(\cdot |x)$ and $y\sim \pi_{\theta}(\cdot |[q,x])$ . Following [41], the optimal solution of $J(\pi_{\theta})$ can be derived as + +$$ +\pi_ {\theta} (q | x) = \frac {1}{Z (x)} \pi_ {\mathrm {r e f}} (q | x) \exp (\frac {\lambda}{\beta} r (x, q)), +$$ + +$$ +\pi_ {\theta} (y | [ x, q ]) = \frac {1}{Z ([ x , q ])} \pi_ {\mathrm {r e f}} (y | [ x, q ]) \exp (\frac {\delta}{\beta} r ([ x, q ], y)). +$$ + +Thus, we can obtain the reward functions as + +$$ +r (x, q) = \frac {\beta}{\lambda} \log \frac {\pi_ {\theta} (q | x)}{\pi_ {\operatorname {r e f}} (q | x)} + \frac {\beta}{\lambda} \log Z (x), \tag {8} +$$ + +$$ +r ([ x, q ], y) = \frac {\beta}{\delta} \log \frac {\pi_ {\theta} (y | [ x , q ])}{\pi_ {\mathrm {r e f}} (y | [ x , q ])} + \frac {\beta}{\delta} \log Z ([ x, q ]). +$$ + +Now, given the definition of each step reward, the total trajectory reward defined in Eq. (6) in terms of the optimal + +![](images/3a8e025debf3d9dbf9038d3c4f82ff12d52958bdc234868b3ed24a46d566e4e8.jpg) +Figure 5. The score distribution of questions and answers, and labeling agreement with 3 LVLMS as judges. GPT-4o provides lower mean score and higher diversity than LLaVA-critic, acting more like a human annotator. + +policy can be directly substituted into the preference model in Eq. (1), where the $Z(x)$ term will be canceled just as the original DPO derivation. Assuming $\tau_w = (x, q_1, [x, q_1], y_1)$ and $\tau_l = (x, q_2, [x, q_2], y_2)$ , the preference probability of preferred $\tau_w$ over dispreferred $\tau_l$ can be defined as + +$$ +\begin{array}{l} p \left(\tau_ {w} \succ \tau_ {l}\right) := \sigma \left(\frac {\beta}{\lambda} \log \frac {\pi_ {\theta} \left(q _ {1} \mid x\right)}{\pi_ {\operatorname {r e f}} \left(q _ {1} \mid x\right)} + \frac {\beta}{\delta} \log \frac {\pi_ {\theta} \left(y _ {1} \mid [ x , q _ {1} ]\right)}{\pi_ {\operatorname {r e f}} \left(y _ {1} \mid [ x , q _ {1} ]\right)} \right. \\ \left. - \frac {\beta}{\lambda} \log \frac {\pi_ {\theta} \left(q _ {2} \mid x\right)}{\pi_ {\text {r e f}} \left(q _ {2} \mid x\right)} - \frac {\beta}{\delta} \log \frac {\pi_ {\theta} \left(y _ {2} \mid [ x , q _ {2} ]\right)}{\pi_ {\text {r e f}} \left(y _ {2} \mid [ x , q _ {2} ]\right)}\right). \tag {9} \\ \end{array} +$$ + +More details about the derivation from Eq. (8) to Eq. (10) could be found in the supplementary. + +As shown in Fig. 4, we reformulate the preference data based on trajectories as $\mathcal{D} = \{\tau_1^{(i)},\tau_2^{(i)},\tau_3^{(i)},\tau_4^{(i)}\}_{i = 1}^N$ . By substituting Eq. (8) into Eq. (9), the loss function of classifying between two trajectories is given as + +$$ +\begin{array}{l} \mathcal {L} \left(\tau_ {w}, \tau_ {l}\right) = - \mathbb {E} _ {\left(\tau_ {w}, \tau_ {l}\right) \sim \mathcal {D}} \left[ \log \sigma \left(r \left(\tau_ {w}\right) - r \left(\tau_ {l}\right)\right) \right] \tag {10} \\ \leq \mathbb {E} _ {(\tau_ {w}, \tau_ {l}) \sim \mathcal {D}} [ - \log \sigma ((r (x, q _ {1}) - r (x, q _ {2}))) ] \\ + \mathbb {E} _ {\left(\tau_ {w}, \tau_ {t}\right) \sim \mathcal {D}} [ - \log \sigma \left(\left(r ([ x, q _ {1} ], y _ {1}) - r ([ x, q _ {2} ], y _ {2}))\right) \right]. \\ \end{array} +$$ + +Eventually, we define the preference probability between sub-tree $\mathcal{V}_w$ and $\mathcal{V}_l$ as a joint of multiple trajectory-based preference probability, improving the comprehension of policy distribution. The final SPO loss is formulated as + +$$ +\mathcal {L} _ {S P O} = \mathcal {L} _ {\left(\mathcal {Y} _ {w}, \mathcal {Y} _ {i}\right)} := \sum_ {m, n} \mathcal {L} \left(\tau_ {m}, \tau_ {n}\right): m < n, \tag {11} +$$ + +where $\mathcal{L}(\tau_m,\tau_n)$ is given in Eq. (10), and $m\in \Phi_w = \{1,3\}$ , $n\in \Phi_l = \{2,4\}$ are the index of preferred and dispreferred trajectories. + +Table 1. Evaluation results of our model trained with different preference optimization methods based on the SFT model on a scale of 0.5B and 7B. We report accuracy on all the benchmarks except MME. The best results are bold, and the second-best results are underlined. + +
ModelSingle-Image
MMStarMMMUMMEL-WilderLLaVA-WMMVetRealWorldQAPOPEMMBench
testvaltestsmalltesttesttesttestMulti-discip
GPT-4V [26]57.156.8140981.098.049.961.4-75.0
GPT-4o [27]-69.1-85.9106.176.258.6--
Gemini-1.5-Pro [36]-62.2----70.4--
InternVL-2-8B [7]59.449.3221062.584.560.064.4-81.7
IXC-2.5-7B [50]59.942.9222961.478.151.767.8-82.2
LLaVA-OV-72B [18]66.156.8168272.093.563.771.9-85.9
LLaVA-OV-0.5B [18]37.531.4123855.074.229.155.688.052.1
SFT-0.5B [35]40.731.9124659.472.826.755.788.053.1
SFT-0.5B + DPO [29]40.932.1125058.573.525.856.488.154.0
SFT-0.5B + SPO (ours)41.532.8125161.374.130.056.988.254.3
LLaVA-OV-7B [18]61.748.8158067.890.757.566.389.280.8
SFT-7B [35]59.648.0156569.590.557.666.889.079.7
SFT-7B + DPO [29]61.949.0157969.991.258.267.289.180.6
SFT-7B + SPO (ours)62.149.2158570.292.358.867.389.381.0
+ +Table 2. Evaluation results of different methods on multi-image and video benchmarks. #Frames is the number of frames used for video encoding. We report accuracy on all the benchmarks. The best results are bold, and the second-best results are underlined. + +
Model#FramesMulti-ImageEgoSchemaVideo
\(LLaVA^I\)MuirBenchMLVUNExT-QALVid-BenchVideoMME
out-domainout-domaintestm-avgmcvalw/o subs
GPT-4V [26]1060.362.3-49.2-61.359.9
GPT-4o [27]384-68.0-64.6-66.771.9
Gemini-1.5-Pro [36]1/s-49.472.2--64.075.0
LongVA-7B [51]128---56.368.3-52.6
IXC-2.5-7B [50]64---37.371.0-55.8
LLaVA-Video-7B [53]64--57.370.883.258.263.3
LLaVA-Video-72B [53]64--65.674.485.461.970.5
LLaVA-OV-72B [18]3279.954.862.068.080.261.366.3
LLaVA-OV-0.5B [18]3233.325.526.850.357.245.844.0
SFT-0.5B [35]3232.726.128.851.759.044.644.3
SFT-0.5B + DPO [29]3232.927.928.952.059.246.144.8
SFT-0.5B + SPO (ours)3233.428.029.252.460.447.745.4
LLaVA-OV-7B [18]3264.241.860.164.779.456.458.1
SFT-7B [35]3266.845.560.064.977.257.557.8
SFT-7B + DPO [29]3266.946.259.565.179.858.158.3
SFT-7B + SPO (ours)3267.946.760.565.379.657.858.5
+ +# 4. Experiments + +# 4.1. Training Recipe + +We follow [18] to construct our model architecture, including the instructional language model, vision encoder, and a 2-layer MLP projector. This work adopts LoRA [15] for an efficient and lightweight fine-tuning, where we set low-rank adapters with rank $= 64$ and $\alpha = 128$ for the LLM. We denote $\theta$ as all the trainable parameters, including LoRA and projector layer. As shown in Fig. 3, the training pipeline has two stages: 1) Utilizing supervised fine-tuning (SFT) to train an SFT model by reasoning and learning to ques + +tion [35]. 2) Fine-tune the LVML policy (given the SFT model as a reference) using the proposed SPO approach by maximizing a novel structured reward function. We show the training details for each stage in supplementary. + +# 4.2. Evaluation Benchmarks + +We primarily assess our models using vision-language benchmarks across different visual modalities (e.g., image, multi-image, and video) that are not included in the training data. For a fair comparison with other LVLMs, we primarily report results from original papers and utilize Immseval [49] for test setting consistency. All our results are + +![](images/f13afe7ae10049180636b63708a58286d0782643ca3b3857014a87f59b4fd5b2.jpg) +Figure 6. Qualitative results of LLaVA-Bench-Wilder. When trained with DPO, the model becomes over-confident when indicating low visual dependency content. In contrast, SPO improves the model's visual reasoning ability and prevents over-confident prediction when explaining uncertain situations, acting similarly to the reference model, GPT-4V. + +Are the otters in the image reading a book, and if so, can you provide a detailed description of what might be happening in this whimsical scene? + +SPO-7B Based on the image, it appears that the otters are interacting with a book in an unusual and anthropomorphic manner. The otters seem to be holding the book open as if they are reading it, which is a humorous and imaginative interpretation of their behavior. The book's placement between them suggests a shared interest or activity. The scene is whimsical because otters do not engage in such human-like behaviors naturally; this image is likely staged for artistic or entertainment purposes. +DPO-7B The otters in the image are likely to be sea otters, as they are shown floating on their backs in water and holding a book between them. The fact that they are holding a book also indicates that they have been exposed to human activities or objects, which could mean they live in an area where humans frequently visit or interact with them. +GPT-4V Yes, the otters in the image appear to be reading a book, which is a whimsical and anthropomorphized activity for animals. The scene is set in a body of water, with the two otters floating and seemingly fully absorbed in the book. One otter holds the book with one hand and points to a line or page with the other, as if highlighting something or guiding the reading process. The second otter is focused on the book, with an expression of curiosity or concentration. This playful image brings a charming and imaginative element, showcasing animals engaging in human-like activities and suggesting themes of education, curiosity, and companionship. Reference answer + +reported with greedy sampling and 0-shot settings. + +Single-image Benchmarks. We consider the LVLM as a general-purpose assistant in the wild as the most critical metric. To validate the capabilities in real-world scenarios with open-form instructions, we utilize MMStar [5], LLaVA-Bench-Wilder [18], and RealWorldQA [43]. Besides chat ability, visual perception evaluates the model's reasoning ability, so we adopt MME [13] and MMMU [48]. + +Multi-image Benchmarks. Multi-image tasks test the model's ability to connect and reason across images, making them generally more challenging as they require higher-level contextual integration. This work utilizes LLaVA-Interleave Bench [20] under the out-of-domain setting that does not overlap with training scenarios. This setting covers 13 challenging real-world tasks such as spotting the difference and visual storytelling, etc. We also provide a comprehensive and holistic evaluation of MuirBench [37], which contains 12 diverse multi-image tasks. + +Video Benchmarks. Video understanding is a critical ability to build a general-purpose vision-language assistant. This work conducts several open-ended and multi-choice video benchmarks. EgoSchema [25] and MLVU [54] focus on long video understanding, NExTQA [44], LongVideoBench [40] and VideoMME [14] contains diverse video domains and duration. + +# 4.3. Main Results + +In Table 1, while both preference optimization algorithms enhance performance over the SFT model, SPO achieves the best overall performance across all the image benchmarks under 0.5B and 7B scale. The consistency highlights the robustness and effectiveness of SPO. Notably, SPO outperforms its baseline model LLaVA-OV by 6.3 (0.5B) and 2.4 (7B) points on LLaVA-Bench-Wilder, and the improve + +ment on RealWorldQA indicates that SPO enhances the model's visual reasoning capability while handling freeform visual chat and real-world scenario. Such improvement follows our expectations since we precisely optimize the model to capture the alignment between visual input and dialogue interaction. Fig. 6 shows a qualitative example of real-world visual chat from three models. + +We have observed an improvement of SPO on the multi-image benchmarks, demonstrating the significance of our method for improving comprehensive and complex visual reasoning. Specifically, SPO-0.5B achieves a $4.4\%$ and $1.0\%$ improvement, and SPO-7B achieves an $8.1\%$ and $2.0\%$ improvement over LLaVA-OV and our SFT model. Table 2 also provides evaluation results on video benchmarks. The consistent improvement of SPO-0.5b/ 7B on EgoSchema, MLVU, and L-VideoBench indicates that SPO also improves consecutive visual reasoning, especially for medium and long video understanding. + +SPO has better visual-question understanding ability than DPO since it encourages visually informative questions by utilizing the structured preference tree that captures diverse question-answer pairs. Regarding the low-quality questions, SPO also learns their preferred answers through a subtree within the structured preferences. To validate our hypothesis, we prompt GPT-4o to score 216 questions in the MMVet dataset and split them into high-dependency (High-Q) and low-dependency (Low-Q) visual questions. Table 4 shows that more visually focused questions consistently improve the answering accuracy given images. As expected, without looking at images (w/o images), Low-Q reports a smaller performance drop than High-Q, and SPO has lower performance drop than DPO. Empirical evidence shows that SPO is more robust to question quality than DPO. + +Limitation. While SPO enhances the SFT model, we + +
r(x,q)r([x,q],y)ImageMulti-ImageVideo
MMELLaVA-WilderLLaVA1MuirBenchEgoSchemaMLVU
✓✓158570.267.946.760.565.3
✓X156969.967.346.260.065.2
X✓157169.767.246.360.165.0
XX156569.566.845.560.064.9
+ +![](images/bc1a891cb3414ad7f76c576ea366720b1c39daec88bd0b8787d4c406dec59de1.jpg) + +Table 3. Ablation study under SFT-7B-SPO. We ablate the trajectory re- Figure 7. Left. SPO achieves higher preference classification reward design of SPO. The formal trajectory reward is $r(\tau) = r(x,q) +$ accuracy than DPO throughout training. Right. SPO maxi $r([x,q],y)$ , where $r(x,q)$ and $r([x,q],y)$ are question and answer reward. mizes the reward margin faster effectiveness than DPO. +Table 4. Test accuracy drop in two question splits on MMVet. + +
ModelsHigh-Q (w/o image)Low-Q (w/o image)
SFT-0.5B + DPO27.0 (6.0) ↓ 77.8%17.0 (6.0) ↓ 64.7%
SFT-0.5B + SPO29.0 (8.0) ↓ 72.4%23.0 (12.0) ↓ 47.8%
+ +want to emphasize the critical role of the SFT model as the foundational stage for all preference optimization methods. Notably, the limited computational resources primarily drive the performance decline after SFT on several benchmarks. Our SFT-7B model was fine-tuned with 8 GPUs — only 1/16 of training GPUs in LLaVA-OV — resulting in a much smaller batch size for our approach. Specifically, the performance degradation of SFT-7B on MMStar, MMU, and MME sets a lower bound for the subsequent gains achievable by SPO. A similar trend is observed in video benchmarks, as shown in Table 2. Additionally, we decrease the anyres [18] to $60\%$ of the regular setting for visual encoding to fit in our computational resource. Due to limited computational resources, SPO mainly focuses on single-turn conversations. We leave multi-turn conversations study in future work. + +# 4.4. Model Discussion + +Structured Reward Speeds Up Policy Learning. The Bradley-Terry objective function essentially aims to optimize the reward margin of the preferred and dispreferred action $\Delta r(x, \rho_w, \rho_l) = r(x, \rho_w) - r(x, \rho_l)$ . In Eq. (10), SPO optimizes a joint reward margin including $\Delta r(x, q_1, q_2)$ and $\Delta r([x, q_1], y_1, y_2)$ . While DPO only optimizes the reward margin between responses $\Delta r([x, q_1], y_1, y_2)$ . We use the SPO and DPO models to investigate the relationship between the learned reward margin and the preference classification accuracy. Fig. 7 shows that SPO maximizes the reward margin more effectively than DPO, and SPO achieves higher preference classification accuracy than DPO along all the training steps. By precisely adjusting $\beta$ for different preference probabilities, we find that optimizing multiple trajectory pairs together can explicitly increase the data diversity, leading to better generalizability and performance. + +Influence of Each Step Reward. SPO optimizes a policy to satisfy multiple preferences in a single objective function. In Eq. (10), we use $\lambda$ and $\delta$ to control the deviation of + +Table 5. Latency vs Accuracy on the LLaVA-W dataset + +
MetricSFT-7BSFT-7B + DPOSFT-7B + SPO
Accuracy (%)72.873.574.1
Inference time (s)523547254
+ +preference alignment over two action groups, questioning and answering. Following the same training strategy and preference data, we evaluate the performance of SPO w/o question reward $r(x,q) = 0$ and SPO w/o answer reward $r([x,q] = y) = 0$ . As shown in Table. 3, learning question policy improves real-world QA ability such as LLaVA-Wilder and LLaVA; learning answer policy brings more improvement over fixed-form QA such as MME. Overall, the joint policy improves the model from different aspects. + +Computational Resource Analysis. For training, optimizing the whole dialogue will not burden the computation cost when comparing SPO with DPO (e.g., SPO: 24h vs DPO: 22h using $4^{*}\mathrm{A}100$ GPUs) since only the question tokens need to be unmasked. For inference, Table 5 evaluates latency by reporting the total inference time of three models on 60 open-ended questions from the LLaVA-W dataset, and the test accuracy is measured by GPT-4o. Empirically, SPO generates more concise and accurate content than the SFT and DPO models. + +# 5. Conclusion + +In this work, we have proposed Structured Policy Optimization (SPO), a novel approach to enhancing multi-modal preference optimization for large vision-language models (LVLMs). SPO alleviates the limitation of existing DPO methods, which adopt a single, unstructured reward and tend to prioritize linguistic nuances over visual context. Specifically, SPO introduces a new structured reward represented in trajectories to improve the visual dependency by maximizing the reward between preferred and dispreferred questions/answers. Our experimental results demonstrate that SPO improves vision-language understanding and reasoning across diverse tasks, showcasing superior alignment with human-preferred responses compared to traditional DPO methods. Overall, this approach better captures the dynamic interplay between visual inputs, questions, and responses, aligning more effectively with human preferences. + +# 6. Acknowledgement + +This research was supported by the DEVCOM Army Research Laboratory under Contract W911QX-21-D-0001. + +# References + +[1] OpenAI Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, et al. Gpt-4 technical report. 2023. 2 +[2] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022. 3 +[3] Hritik Bansal, Ashima Suvarna, Gantavya Bhatt, Nanyun Peng, Kai-Wei Chang, and Aditya Grover. Comparing bad apples to good oranges: Aligning large language models via joint preference optimization. ArXiv, 2024. 2, 4 +[4] Lin Chen, Jinsong Li, Xiao wen Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. In ArXiv, 2023. 2 +[5] Lin Chen, Jinsong Li, Xiao wen Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, and Feng Zhao. Are we on the right way for evaluating large vision-language models? In NeurIPS, 2024. 2, 7 +[6] Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji, and Quanquan Gu. Self-play fine-tuning converts weak language models to strong language models. ArXiv, 2024. 3 +[7] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 6 +[8] Paul F. Christiano, Jan Leike, Tom B. Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences. In Proceedings of the 31st International Conference on Neural Information Processing Systems, 2017. 1, 3, 5 +[9] Ruiguo Cui and Peter Teo. Thinking through talk: Using dialogue to develop students' critical thinking. Teaching and Teacher Education, 2023. 2, 4 +[10] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Albert Li, Pascale Fung, and Steven C. H. Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, 2024. 2 +[11] Yihe Deng, Pan Lu, Fan Yin, Ziniu Hu, Sheng Shen, James Zou, Kai-Wei Chang, and Wei Wang. Enhancing large vision language models with self-training on image comprehension. In NeurIPS, 2024. 1, 3 +[12] Hanze Dong, Wei Xiong, Deepanshu Goyal, Rui Pan, Shizhe Diao, Jipeng Zhang, Kashun Shum, and T. Zhang. Raft: + +Reward ranked finetuning for generative foundation model alignment. ArXiv, 2023. 1, 3 +[13] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. ArXiv, 2023. 1, 7 +[14] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, Peixian Chen, Yanwei Li, Shaohui Lin, Sirui Zhao, Ke Li, Tong Xu, Xiawu Zheng, Enhong Chen, Rongrong Ji, and Xing Sun. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. ArXiv, 2024. 7 +[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 6 +[16] Dongfu Jiang, Xuan He, Huaye Zeng, Cong Wei, Max W.F. Ku, Qian Liu, and Wenhu Chen. Mantis: Interleaved multi-image instruction tuning. ArXiv, 2024. 2 +[17] Xin Lai, Zhuotao Tian, Yukang Chen, Senqiao Yang, Xiangru Peng, and Jiaya Jia. Step-dpo: Step-wise preference optimization for long-chain reasoning of llms. ArXiv, 2024. 1, 3 +[18] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. ArXiv, 2024. 2, 6, 7, 8 +[19] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. In NeurIPS, 2023. 1, 2, 3 +[20] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava-last-interleave: Tackling multi-image, video, and 3d in large multimodal models. ArXiv, 2024. 2, 7 +[21] Juncheng Li, Kaihang Pan, Zhiqi Ge, Minghe Gao, Hanwang Zhang, Wei Ji, Wenqiao Zhang, Tat seng Chua, Siliang Tang, and Yueting Zhuang. Fine-tuning multimodal llms to follow zero-shot demonstrative instructions. In International Conference on Learning Representations, 2024. 2 +[22] Lei Li, Zhihui Xie, Mukai Li, Shunian Chen, Peiyi Wang, Liang Chen, Yazheng Yang, Benyou Wang, and Lingpeng Kong. Silkie: Preference distillation for large visual language models. ArXiv, 2023. 1, 3 +[23] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. ArXiv, 2023. 2 +[24] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In ArXiv, 2023. 2 +[25] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. ArXiv, 2023. 1, 7 +[26] OpenAI. Gpt-4v(ision) system card. 2023. 2, 6 + +[27] OpenAI. Gpt-4o system card. 2024. 2, 4, 6 +[28] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke E. Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Francis Christiano, Jan Leike, and Ryan J. Lowe. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 3 +[29] Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. In NeurIPS, 2023. 1, 2, 3, 4, 6 +[30] Rafael Rafailov, Joey Hejna, Ryan Park, and Chelsea Finn. From r to q: Your language model is secretly a q-function. In First Conference on Language Modeling, 2024. 1, 3 +[31] Corby Rosset, Ching-An Cheng, Arindam Mitra, Michael Santacroce, Ahmed Awadallah, and Tengyang Xie. Direct nash optimization: Teaching language models to self-improve with general preferences. ArXiv, 2024. 3 +[32] Avi Singh, John D. Co-Reyes, Rishabh Agarwal, Ankesh Anand, Piyush Patil, Peter J. Liu, James Harrison, Jaehoon Lee, Kelvin Xu, Aaron T Parisi, Abhishek Kumar, Alex Alemi, Alex Rizkowsky, Azade Nova, Ben Adlam, Bernd Bohnet, Hanie Sedghi, Igor Mordatch, Isabelle Simpson, Izzeddin Gur, Jasper Snoek, Jeffrey Pennington, Jiri Hron, Kathleen Kenealy, Kevin Swersky, Kshitteej Mahajan, Laura Culp, Lechao Xiao, Maxwell Bileschi, Noah Constant, Roman Novak, Rosanne Liu, Tris Brian Warkentin, Yundi Qian, Ethan Dyer, Behnam Neyshabur, Jascha Narain Sohl-Dickstein, and Noah Fiedel. Beyond human data: Scaling self-training for problem-solving with language models. Trans. Mach. Learn. Res., 2023. 3 +[33] Feifan Song, Yu Bowen, Minghao Li, Haiyang Yu, Fei Huang, Yongbin Li, and Houfeng Wang. Preference ranking optimization for human alignment. ArXiv, 2023. 1, 3 +[34] Guohao Sun, Can Qin, Huazhu Fu, Linwei Wang, and Zhiqiang Tao. Stllava-med: Self-training large language and vision assistant for medical. In EMNLP, 2024. 2, 3 +[35] Guohao Sun, Can Qin, Jiamian Wang, Zeyuan Chen, Ran Xu, and Zhiqiang Tao. Sq-llava: Self-questioning for large vision-language assistant. In ECCV, 2024. 2, 3, 4, 6 +[36] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2, 6 +[37] Fei Wang, Xingyu Fu, James Y. Huang, Zekun Li, Qin Liu, Xiaogeng Liu, Mingyu Derek Ma, Nan Xu, Wenxuan Zhou, Kai Zhang, Tianyi Yan, Wenjie Jacky Mo, Hsiang-Hui Liu, Pan Lu, Chunyuan Li, Chaowei Xiao, Kai-Wei Chang, Dan Roth, Sheng Zhang, Hoifung Poon, and Muhao Chen. Muir-bench: A comprehensive benchmark for robust multi-image understanding. ArXiv, 2024. 1, 7 +[38] Fei Wang, Wenxuan Zhou, James Y. Huang, Nan Xu, Sheng Zhang, Hoifung Poon, and Muhao Chen. mdpo: Conditional preference optimization for multimodal large language models. In EMNLP, 2024. 1, 3 + +[39] Yuetian Weng, Mingfei Han, Haoyu He, Xiaojun Chang, and Bohan Zhuang. Longvlm: Efficient long video understanding via large language models. ArXiv, 2024. 2 +[40] Haoning Wu, Dongxu Li, Bei Chen, and Junnan Li. Longvideobench: A benchmark for long-context interleaved video-language understanding. ArXiv, 2024. 7 +[41] Junkang Wu, Yuexiang Xie, Zhengyi Yang, Jiancan Wu, Jinyang Gao, Bolin Ding, Xiang Wang, and Xiangnan He. $\beta$ -dpo: Direct preference optimization with dynamic $\beta$ . ArXiv, 2024. 5 +[42] Yue Wu, Zhiqing Sun, Huizhuo Yuan, Kaixuan Ji, Yiming Yang, and Quanquan Gu. Self-play preference optimization for language model alignment. ArXiv, 2024. 3 +[43] x.ai. Grok-1.5 vision preview. Technical report, x.ai, 2024. 7 +[44] Junbin Xiao, Xindi Shang, Angela Yao, and Tat seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 7 +[45] Tianyi Xiong, Tianyi Xiong, Xiyao Wang, Dong Guo, Qinghao Ye, Haoqi Fan, Quanquan Gu, Heng Huang, and Chunyuan Li. Llava-critic: Learning to evaluate multimodal models. ArXiv, 2024. 4 +[46] Enwei Xu, Wei Wang, and Qingxia Wang. The effectiveness of collaborative problem solving in promoting students' critical thinking: A meta-analysis based on empirical literature. *Humanities and Social Sciences Communications*, 2023. 2, 4 +[47] Xubing Ye, Yukang Gan, Xiaoke Huang, Yixiao Ge, Ying Shan, and Yansong Tang. Voco-llama: Towards vision compression with large language models. 2024. 2 +[48] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Ren-liang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 7 +[49] Kaichen Zhang, Bo Li, Peiyuan Zhang, Fanyi Pu, Joshua Adrian Cahyono, Kairui Hu, Shuai Liu, Yuanhan Zhang, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Lmmseval: Reality check on the evaluation of large multimodal models. ArXiv, 2024. 6 +[50] Pan Zhang, Xiao wen Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, Songyang Zhang, Wenwei Zhang, Yining Li, Yang Gao, Peng Sun, Xinyue Zhang, Wei Li, Jingwen Li, Wenhai Wang, Hang Yan, Conghui He, Xingcheng Zhang, Kai Chen, Jifeng Dai, Yu Qiao, Dahua Lin, and Jiaqi Wang. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. ArXiv, 2024. 6 +[51] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. In ArXiv, 2024. 6 + +[52] Yanzhe Zhang, Ruiyi Zhang, Jiumiang Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tongfei Sun. Llavar: Enhanced visual instruction tuning for text-rich image understanding. In ArXiv, 2023. 2 +[53] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. ArXiv, 2024. 2, 6 +[54] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. ArXiv, 2024. 7 +[55] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. ArXiv, 2023. 2 +[56] Daniel M. Ziegler, Nisan Stiennon, Jeff Wu, Tom B. Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. ArXiv, 2019. 3 \ No newline at end of file diff --git a/2025/Structured Policy Optimization_ Enhance Large Vision-Language Model via Self-referenced Dialogue/images.zip b/2025/Structured Policy Optimization_ Enhance Large Vision-Language Model via Self-referenced Dialogue/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e6b1c319a105025fdc082fa301b05d95e2121f3e --- /dev/null +++ b/2025/Structured Policy Optimization_ Enhance Large Vision-Language Model via Self-referenced Dialogue/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:987ee2591ccc7b2c764709b55d34e8c0223bf41dd7504fbb6f1f57c889e275e7 +size 636137 diff --git a/2025/Structured Policy Optimization_ Enhance Large Vision-Language Model via Self-referenced Dialogue/layout.json b/2025/Structured Policy Optimization_ Enhance Large Vision-Language Model via Self-referenced Dialogue/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a192584776b6d8e2d3f09e8ee267a3efc34b7de8 --- /dev/null +++ b/2025/Structured Policy Optimization_ Enhance Large Vision-Language Model via Self-referenced Dialogue/layout.json @@ -0,0 +1,9327 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 102, + 541, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 102, + 541, + 140 + ], + "spans": [ + { + "bbox": [ + 70, + 102, + 541, + 140 + ], + "type": "text", + "content": "Structured Policy Optimization: Enhance Large Vision-Language Model via Self-Referenced Dialogue" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 121, + 160, + 493, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 160, + 493, + 190 + ], + "spans": [ + { + "bbox": [ + 121, + 160, + 493, + 190 + ], + "type": "text", + "content": "Guohao Sun1*, Can Qin2, Yihao Feng2, Zeyuan Chen2, Ran Xu2, Sohail Dianat1, Majid Rabbani1, Raghuveer Rao3, and Zhiqiang Tao1*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 141, + 193, + 465, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 193, + 465, + 223 + ], + "spans": [ + { + "bbox": [ + 141, + 193, + 465, + 223 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 141, + 193, + 465, + 223 + ], + "type": "text", + "content": "Rochester Institute of Technology \n" + }, + { + "bbox": [ + 141, + 193, + 465, + 223 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 141, + 193, + 465, + 223 + ], + "type": "text", + "content": "Salesforce AI Research " + }, + { + "bbox": [ + 141, + 193, + 465, + 223 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 141, + 193, + 465, + 223 + ], + "type": "text", + "content": "DEVCOM Army Research Laboratory" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 248, + 200, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 248, + 200, + 261 + ], + "spans": [ + { + "bbox": [ + 151, + 248, + 200, + 261 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 275, + 296, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 296, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 296, + 562 + ], + "type": "text", + "content": "Preference optimization algorithms typically enhance LLM response quality by leveraging human feedback on multiple answers given a fixed instruction. However, these methods often lack capturing the dynamic nature of conversational exchanges. For large vision-language models (LVLMs), direct preference optimization (DPO) can over-emphasize linguistic nuances while overlooking visual context. To address this challenge, we introduce structured policy optimization (SPO) - a novel preference optimization method that simultaneously aligns preference instructions, responses, and dialogue interactions to improve multi-modal understanding and reasoning capabilities. The efficacy of SPO is attributed to one key design: treating the questioning and answering as a sequential action and binding them through a trajectory reward. This reward formulation better aligns with real-world dialogue studies and eliminates the need for fixed instructions. We evaluate our models on interleaved benchmarks, including image, multi-image, and video-based understanding and reasoning tasks. Experimental results show that the proposed SPO finetuning LVLM with multi-modal preference data can align with human preference more efficiently than DPO. The code is available at https://github.com/heliossun/Structure-Policy-Optimization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 588, + 135, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 588, + 135, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 588, + 135, + 601 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 609, + 295, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 295, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 295, + 682 + ], + "type": "text", + "content": "Reinforcement learning from human feedback (RLHF) [8] has demonstrated remarkable success in aligning large language/vision-language models (LLMs/LVLMs) with human preferences, enabling models to adapt their behavior based on success or failure in increasingly complex tasks. Typically, RLHF trains a reward model and optimizes the" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 319, + 250, + 470, + 291 + ], + "blocks": [ + { + "bbox": [ + 319, + 250, + 470, + 291 + ], + "lines": [ + { + "bbox": [ + 319, + 250, + 470, + 291 + ], + "spans": [ + { + "bbox": [ + 319, + 250, + 470, + 291 + ], + "type": "image", + "image_path": "28ef7d25bcde55ac0d4bf2ed9baa241070cc4861884b8828d0b52146a6288640.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 320, + 293, + 463, + 395 + ], + "blocks": [ + { + "bbox": [ + 320, + 293, + 463, + 395 + ], + "lines": [ + { + "bbox": [ + 320, + 293, + 463, + 395 + ], + "spans": [ + { + "bbox": [ + 320, + 293, + 463, + 395 + ], + "type": "image", + "image_path": "ad1d1f1bef0d013fb89038fb4681c4eea37330537c2d1a9c71e75d4d85654f1d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 403, + 555, + 491 + ], + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 491 + ], + "type": "text", + "content": "Figure 1. Comparison between DPO and SPO in optimizing an LVLM policy. DPO employs a single reward that optimizes the preferred answer given the fixed visual & question input, " + }, + { + "bbox": [ + 313, + 403, + 555, + 491 + ], + "type": "inline_equation", + "content": "p(y_w \\succ y_l|x, q)" + }, + { + "bbox": [ + 313, + 403, + 555, + 491 + ], + "type": "text", + "content": ", which could over-emphasize language nuance and overlook visual clues [38]. By contrast, SPO introduces a structured reward by " + }, + { + "bbox": [ + 313, + 403, + 555, + 491 + ], + "type": "inline_equation", + "content": "p(\\mathcal{Y}_w \\succ \\mathcal{Y}_l|x)" + }, + { + "bbox": [ + 313, + 403, + 555, + 491 + ], + "type": "text", + "content": ", conditioning on the visual dialogue, to optimize the preferred trajectories spanning over questions and answers." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 474, + 249, + 547, + 301 + ], + "blocks": [ + { + "bbox": [ + 474, + 249, + 547, + 301 + ], + "lines": [ + { + "bbox": [ + 474, + 249, + 547, + 301 + ], + "spans": [ + { + "bbox": [ + 474, + 249, + 547, + 301 + ], + "type": "image", + "image_path": "e9a5b624afff8f00463fd23650074968d0cf4410631b8b4ac0882894f1177bed.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 475, + 312, + 547, + 393 + ], + "blocks": [ + { + "bbox": [ + 475, + 312, + 547, + 393 + ], + "lines": [ + { + "bbox": [ + 475, + 312, + 547, + 393 + ], + "spans": [ + { + "bbox": [ + 475, + 312, + 547, + 393 + ], + "type": "image", + "image_path": "4a691cd52fd51a7243c09ab7c443d480f16e0c59e0a8f8621a230b62240094fd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 518, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 555, + 544 + ], + "type": "text", + "content": "LLM/LVLM policy to maximize the reward, which is applied in both language and vision-language (VL) domains." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "type": "text", + "content": "Recently, the offline preference optimization methods, such as DPO [29] and its variants [12, 17, 30, 33], have proposed various reward objectives to satisfy different human expectations. Owing to its simplicity and stability, DPO has been widely used in multi-modal preference optimization [11, 19, 22, 38] by aligning LVLM to human-preferred responses conditioned on fixed visual and text instructions (see Fig. 1). However, unlike purely text dialogue, the visual input can lead to far more enriched information interactions, which require varying degrees of visual dependence [13, 25, 37]. Directly applying DPO to multi-modal preference data (e.g., images/videos and texts) will over-emphasize language nuance and overlook the visual condition [38]. Specifically, DPO applies a single, un" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "*Corresponding authors: Guohao Sun (gs4288@rit.edu) and Zhiqiang Tao (zhiqiang.tao@rit.edu)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "741" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 72, + 173, + 139 + ], + "blocks": [ + { + "bbox": [ + 63, + 72, + 173, + 139 + ], + "lines": [ + { + "bbox": [ + 63, + 72, + 173, + 139 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 173, + 139 + ], + "type": "image", + "image_path": "62500da39ba19fe7e3f00edccc19d9cdc47c01c3e09e2d88141948b140e5de27.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 79, + 144, + 158, + 153 + ], + "lines": [ + { + "bbox": [ + 79, + 144, + 158, + 153 + ], + "spans": [ + { + "bbox": [ + 79, + 144, + 158, + 153 + ], + "type": "text", + "content": "(a) Question vs Answer." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 176, + 72, + 290, + 140 + ], + "blocks": [ + { + "bbox": [ + 176, + 72, + 290, + 140 + ], + "lines": [ + { + "bbox": [ + 176, + 72, + 290, + 140 + ], + "spans": [ + { + "bbox": [ + 176, + 72, + 290, + 140 + ], + "type": "image", + "image_path": "ac2b3fef8156fe58734ef89f170e5d8052903d30085bb4c5149f4acce882455d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 197, + 144, + 268, + 153 + ], + "lines": [ + { + "bbox": [ + 197, + 144, + 268, + 153 + ], + "spans": [ + { + "bbox": [ + 197, + 144, + 268, + 153 + ], + "type": "text", + "content": "(b) Score distribution." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 163, + 295, + 207 + ], + "lines": [ + { + "bbox": [ + 55, + 163, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 295, + 207 + ], + "type": "text", + "content": "Figure 2. Scoring visual dialogue with GPT-4o as judge. (a) Questions with higher visual dependency (x-axis, scoring from 4 to 10, " + }, + { + "bbox": [ + 55, + 163, + 295, + 207 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 55, + 163, + 295, + 207 + ], + "type": "text", + "content": ") tend to produce answers with higher quality (y-axis, scoring from 0 to 9, " + }, + { + "bbox": [ + 55, + 163, + 295, + 207 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 55, + 163, + 295, + 207 + ], + "type": "text", + "content": "). (b) Question and answer score distributions." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 232, + 295, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 295, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 295, + 291 + ], + "type": "text", + "content": "structured reward function to maximize the gap between the preferred and dispreferred answer upon the same input sequence; however, this gap might be due to the lack of visual understanding [5], which is generally ignored by existing DPO methods, especially for LVLMs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 293, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 295, + 568 + ], + "type": "text", + "content": "Inspired by the dialogue studies [9, 46], we empirically observe that asking a good visual-related question can turn passive learning into active exploration, allowing for deeper visual understanding and broader knowledge exploration before answering. For instance, as shown in Fig. 1, people prefer the question \"What emotions are conveyed through the shots in the video?\" to \"Which city did the skateboarder go to in the video?\" since the former is more visually dependent and the answer will cover more visual details. To quantitatively validate our observation, we apply GPT-4o to evaluate the visual relevance (scoring from 4 to 10) of questions and score the quality of answers (higher is better) in our curated preference data based on the instruction dataset in [18]. Fig. 2a shows that human-preferred questions with higher visual relevance tend to produce responses with higher quality, which is further demonstrated by the consistent question/answer score distributions in Fig. 2b. We introduce this correlation between visual questions and answer quality as a preference dialogue interaction. By aligning the LVLM policy to generate questions and answers with human preference, we build a more in-depth alignment between vision and dialogues, enabling a better VL understanding for broad tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "In this study, we propose Structured Policy Optimization (SPO), a novel structured DPO method designed to enhance general vision-language understanding and reasoning. As shown in Fig. 1, SPO utilizes a binary tree structure to track visual dialogue changes, alleviating the limitation of using fixed questions in DPO. Unlike DPO, which only optimizes preference at leaf nodes, the proposed SPO optimizes preferences across different trajectories among sub-trees and develops a new structured reward model by leveraging trajectories to capture the dialogue interaction between the visual input, questions, and answers. Additionally, SPO controls the deviation of question reward and an-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": "swer reward by imposing importance weights on each step within the trajectory, which is substantially different from a single-level joint reward optimization [3]. To realize SPO training and increase data diversity, we utilize supervised fine-turning (SFT) to instruct an LVLM to self-generate the preference question-answer pairs, mitigating the distribution shift between the policy model and the preference optimization process [29]. We summarize the contributions of this work as follows." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 180, + 553, + 334 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 314, + 180, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 180, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 314, + 180, + 553, + 228 + ], + "type": "text", + "content": "- For the first time in multi-modal preference optimization, we introduce SPO, a method designed for optimizing structured reward for aligning the policy model to human preferences over visual dialogues." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 228, + 553, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 228, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 314, + 228, + 553, + 287 + ], + "type": "text", + "content": "- We present high-quality multi-modal preference data consisting of scoring over visual questions and answers. This dataset reflects the discrepancy between chosen and rejected dialogue, bringing insights into real-world conversational exchange and dialogue studies." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 288, + 553, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 288, + 553, + 334 + ], + "spans": [ + { + "bbox": [ + 314, + 288, + 553, + 334 + ], + "type": "text", + "content": "- Extensive experiments show that SPO aligns with human preference effectively and overcomes the low visual dependency issue in DPO, leading to better performance on the visual chat benchmarks with free-form instructions." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 346, + 406, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 346, + 406, + 358 + ], + "spans": [ + { + "bbox": [ + 314, + 346, + 406, + 358 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 366, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 555, + 605 + ], + "type": "text", + "content": "Generative Large Vision-Language Model. Recent general-purpose multi-modal models (LVLMs) have shown impressive capability. The commercial proprietary LVLMs, such as GPT-4V [26], GPT-4o [27], and Gemini [36] exhibit excellent generalize ability in various domains [48] such as design, business, science, medical, and science etc. In the open research community, the evolutionary trajectory of LVLMs has also led to subsequent advancements in different scenarios. The pioneering work of LLaVA [24] improves the multi-modal reasoning ability through visual instruction tuning. [4, 10, 24, 34, 35, 55] pushing the performance limits in single-image scenarios. [16, 21] explores multi-image tasks. While video LVLMs [23, 39, 47, 53] aim to alleviate significant challenges due to the long context issue. In contrast to these single-modality models, the latest LLaVA models [18, 20] demonstrate emerging capabilities across all scenarios after training on interleaved data. Following this path, this work aims to improve the general ability of multiple modalities by primarily focusing on improving the model's reasoning ability." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": "Self-generated Synthetic Data. Instruction tuning is widely used in supervised fine-tuning (SFT) to improve general-purpose reasoning. In recent research, synthetic instructional data has become increasingly popular and considered a proxy for human data. This approach primarily leverages the ability of GPT family, [19, 24, 52] utilizes GPT-4 [1] to transform image description into conversations, [4, 18, 20] leverages GPT-4V [26] to process the multi-modal input and generate question-answer pairs." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "742" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 553, + 220 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 553, + 220 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 553, + 220 + ], + "type": "image", + "image_path": "1abbdc8b960da4819ee8a025be615452d45451d361a9de3b173279516748f4f5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 220, + 228, + 232 + ], + "lines": [ + { + "bbox": [ + 58, + 220, + 228, + 232 + ], + "spans": [ + { + "bbox": [ + 58, + 220, + 228, + 232 + ], + "type": "text", + "content": "Stage 1: Reasoning and Learning to Question" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 397, + 220, + 509, + 232 + ], + "lines": [ + { + "bbox": [ + 397, + 220, + 509, + 232 + ], + "spans": [ + { + "bbox": [ + 397, + 220, + 509, + 232 + ], + "type": "text", + "content": "Stage2: Preference revelation" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 239, + 555, + 288 + ], + "lines": [ + { + "bbox": [ + 54, + 239, + 555, + 288 + ], + "spans": [ + { + "bbox": [ + 54, + 239, + 555, + 288 + ], + "type": "text", + "content": "Figure 3. Illustration of LVLM model architecture and training pipeline in the proposed SPO. In Stage 1, we implement supervised fine-tuning through learning to question [35], resulting in an SFT model that enables better visual reasoning and self-generated questions. In Stage 2, we first prompt the SFT model to generate visual dialogues labeled by GPT-4o with human preference. Then, we apply the proposed SPO approach to optimize the LVLM policy through a structured reward, leading to better VL understanding." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 306, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 306, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 54, + 306, + 296, + 413 + ], + "type": "text", + "content": "However, large-scale synthetic data generation comes with unexpected API costs. To overcome this, [34] train the model with a visual-questioning objective function [35] and generate the question-answer pairs efficiently. While most works mentioned above consider one single optimization procedure, self-play fine-tuning methods [6, 31, 32, 42] generate new synthetic data within the training loop. This work focuses primarily on single-turn optimization and avoids iterative processes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 415, + 296, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 582 + ], + "type": "text", + "content": "Preference Optimization Objectives. Reinforcement Learning from Human Feedback (RLHF) is a technique used to align large language models with human preferences and values [2, 8, 28, 56]. To eliminate the complexity and challenges of optimizing online preference algorithms, Direct Preference Optimization (DPO) [29] is a prominent example. To extend two-instance classification, ranking objectives enable comparisons across multiple instances [12, 17, 30, 33]. Recent efforts in multi-modal preference learning mainly focus on preference data design [11, 19, 22, 38], while utilizing the formal DPO objective function. This work proposes a novel structured DPO by maximizing the structured rewards to improve vision-language alignment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 596, + 262, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 596, + 262, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 596, + 262, + 611 + ], + "type": "text", + "content": "3. SPO: Structured Policy Optimization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": "Preference optimization uses explicit feedback (e.g., human ratings, comparisons, or rewards) to fine-tune the SFT model, making it more responsive to real-world user needs. As shown in Fig. 3, this work employs a two-stage training strategy: 1) Utilizing supervised fine-tuning (SFT) to train a SFT model by reasoning and learning to question [35], which is able to generate diverse visual dialogues that align with the training data's specific patterns and content. 2)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 306, + 555, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 555, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 555, + 368 + ], + "type": "text", + "content": "Fine-tune the LVML policy (given the SFT model as a reference) using the proposed SPO approach by maximizing a novel structured reward function. In the following, we will first provide preliminaries about preference optimization and then elaborate on our proposed SPO approach." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 371, + 484, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 484, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 484, + 384 + ], + "type": "text", + "content": "3.1. Backgrounds: RLHF and DPO" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 388, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{RLHF}[8]" + }, + { + "bbox": [ + 313, + 388, + 555, + 437 + ], + "type": "text", + "content": " first trains a reward model " + }, + { + "bbox": [ + 313, + 388, + 555, + 437 + ], + "type": "inline_equation", + "content": "r_{\\phi}(x,y)" + }, + { + "bbox": [ + 313, + 388, + 555, + 437 + ], + "type": "text", + "content": " parametrized by " + }, + { + "bbox": [ + 313, + 388, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 388, + 555, + 437 + ], + "type": "text", + "content": " to predict the human preference distribution " + }, + { + "bbox": [ + 313, + 388, + 555, + 437 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 388, + 555, + 437 + ], + "type": "text", + "content": " over two responses " + }, + { + "bbox": [ + 313, + 388, + 555, + 437 + ], + "type": "inline_equation", + "content": "y_{1}" + }, + { + "bbox": [ + 313, + 388, + 555, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 388, + 555, + 437 + ], + "type": "inline_equation", + "content": "y_{2}" + }, + { + "bbox": [ + 313, + 388, + 555, + 437 + ], + "type": "text", + "content": " (e.g., answers given by an LVLM) with condition " + }, + { + "bbox": [ + 313, + 388, + 555, + 437 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 388, + 555, + 437 + ], + "type": "text", + "content": ", simulated by the Bradely-Terry (BT) model as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 324, + 441, + 553, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 441, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 324, + 441, + 553, + 483 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} p \\left(y _ {1} \\succ y _ {2} \\mid x\\right) = \\frac {\\exp \\left(r _ {\\phi} \\left(x , y _ {1}\\right)\\right)}{\\exp \\left(r _ {\\phi} \\left(x , y _ {1}\\right)\\right) + \\exp \\left(r _ {\\phi} \\left(x , y _ {2}\\right)\\right)} \\tag {1} \\\\ = \\sigma \\left(r _ {\\phi} (x, y _ {1}) - r _ {\\phi} (x, y _ {2})\\right), \\\\ \\end{array}", + "image_path": "a86b357185b9d6701fba57b0c23f7be12a828789c1dc5cc3a5993c09aa7d2c63.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 486, + 554, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 554, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 554, + 534 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 486, + 554, + 534 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 486, + 554, + 534 + ], + "type": "text", + "content": " denotes visual input and questions (instructions) in this study. After training, the reward model " + }, + { + "bbox": [ + 313, + 486, + 554, + 534 + ], + "type": "inline_equation", + "content": "r_{\\phi}" + }, + { + "bbox": [ + 313, + 486, + 554, + 534 + ], + "type": "text", + "content": " provides feedback for each response generated by policy model " + }, + { + "bbox": [ + 313, + 486, + 554, + 534 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 313, + 486, + 554, + 534 + ], + "type": "text", + "content": ", which is optimized by the following objective function:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 344, + 539, + 553, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 539, + 553, + 572 + ], + "spans": [ + { + "bbox": [ + 344, + 539, + 553, + 572 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\max _ {\\pi_ {\\theta}} \\mathbb {E} _ {x \\sim D, y \\sim \\pi_ {\\theta} (y | x)} [ r _ {\\phi} (x, y) ] \\tag {2} \\\\ - \\beta \\mathbb {D} _ {K L} \\left[ \\pi_ {\\theta} (y | x) \\| \\pi_ {r e f} (y | x) \\right], \\\\ \\end{array}", + "image_path": "1bc7733c4a6db0cf53688ad128d0979ccabbbc60f201196c7c38857cce07c991.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 578, + 554, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 554, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 554, + 639 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 578, + 554, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{x^{(i)},y_w^{(i)},y_l^{(i)}\\}_{i = 1}^N" + }, + { + "bbox": [ + 313, + 578, + 554, + 639 + ], + "type": "text", + "content": " is the preference dataset, " + }, + { + "bbox": [ + 313, + 578, + 554, + 639 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 313, + 578, + 554, + 639 + ], + "type": "text", + "content": " is the policy model that need to be updated, and " + }, + { + "bbox": [ + 313, + 578, + 554, + 639 + ], + "type": "inline_equation", + "content": "\\pi_{ref}" + }, + { + "bbox": [ + 313, + 578, + 554, + 639 + ], + "type": "text", + "content": " is a reference model (e.g., an SFT model) that avoids overshifting to preferences. " + }, + { + "bbox": [ + 313, + 578, + 554, + 639 + ], + "type": "inline_equation", + "content": "y_{w}" + }, + { + "bbox": [ + 313, + 578, + 554, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 578, + 554, + 639 + ], + "type": "inline_equation", + "content": "y_{l}" + }, + { + "bbox": [ + 313, + 578, + 554, + 639 + ], + "type": "text", + "content": " refer to preferred and dispreferred answers such as " + }, + { + "bbox": [ + 313, + 578, + 554, + 639 + ], + "type": "inline_equation", + "content": "y_{1}" + }, + { + "bbox": [ + 313, + 578, + 554, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 578, + 554, + 639 + ], + "type": "inline_equation", + "content": "y_{2}" + }, + { + "bbox": [ + 313, + 578, + 554, + 639 + ], + "type": "text", + "content": " in Eq. (1)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 639, + 555, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 555, + 688 + ], + "type": "text", + "content": "DPO [29] reformulates the objective function of RLHF by changing from loss over reward function to optimal policy. Previous work [29] derives the optimal policy " + }, + { + "bbox": [ + 313, + 639, + 555, + 688 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 313, + 639, + 555, + 688 + ], + "type": "text", + "content": " of Eq. (2) as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 346, + 691, + 553, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 691, + 553, + 716 + ], + "spans": [ + { + "bbox": [ + 346, + 691, + 553, + 716 + ], + "type": "interline_equation", + "content": "\\pi_ {\\theta} (y | x) = \\frac {1}{Z (x)} \\pi_ {r e f} (y | x) \\exp \\left(\\frac {1}{\\beta} r (x, y)\\right), \\tag {3}", + "image_path": "cd22004c93325147d95c731b15f360ae8a7fbd1ce29137076479c003c356f3cf.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "text", + "content": "743" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "inline_equation", + "content": "Z(x) = \\sum_{y} \\pi_{\\text{ref}}(y|x) \\exp \\left( \\frac{1}{\\beta} r(x,y) \\right)" + }, + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": " is the marginal likelihood over all possible model responses. To simplify it, DPO reforms Eq. (3) as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 94, + 114, + 295, + 142 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 114, + 295, + 142 + ], + "spans": [ + { + "bbox": [ + 94, + 114, + 295, + 142 + ], + "type": "interline_equation", + "content": "r (x, y) = \\beta \\left(\\log \\frac {\\pi_ {\\theta} (y \\mid x)}{\\pi_ {r e f} (y \\mid x)} + \\log Z (x)\\right). \\tag {4}", + "image_path": "a9923a9931dd2975689346dd124cf5ad94fdc164a3e9b3acc8195b092a1866e7.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 146, + 296, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 146, + 296, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 146, + 296, + 214 + ], + "type": "text", + "content": "The BT model for preference estimation depends only on the difference of rewards between two responses " + }, + { + "bbox": [ + 55, + 146, + 296, + 214 + ], + "type": "inline_equation", + "content": "y_{1}" + }, + { + "bbox": [ + 55, + 146, + 296, + 214 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 146, + 296, + 214 + ], + "type": "inline_equation", + "content": "y_{2}" + }, + { + "bbox": [ + 55, + 146, + 296, + 214 + ], + "type": "text", + "content": " [29]. By substituting Eq. (4) into Eq. (1), DPO expresses the probability of human preference data in terms of the optimal policy and formulating the objective function for a parameterized policy " + }, + { + "bbox": [ + 55, + 146, + 296, + 214 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 55, + 146, + 296, + 214 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 218, + 293, + 243 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 218, + 293, + 243 + ], + "spans": [ + { + "bbox": [ + 59, + 218, + 293, + 243 + ], + "type": "interline_equation", + "content": "- \\mathbb {E} _ {(x, y _ {w}, y _ {l}) \\sim \\mathcal {D}} \\left[ \\log \\sigma \\left(\\beta \\log \\frac {\\pi_ {\\theta} (y _ {w} | x)}{\\pi_ {r e f} (y _ {w} | x)} - \\beta \\log \\frac {\\pi_ {\\theta} (y _ {l} | x)}{\\pi_ {r e f} (y _ {l} | x)}\\right) \\right],", + "image_path": "cbdada3059ce8bd17684a21911ed9a159282ad439d9dadaa1ff28db422af40b3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 247, + 295, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 295, + 270 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 247, + 295, + 270 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 247, + 295, + 270 + ], + "type": "text", + "content": " represents the sigmoid function and " + }, + { + "bbox": [ + 55, + 247, + 295, + 270 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 247, + 295, + 270 + ], + "type": "text", + "content": " is a scalable value to control the deviation from the reference model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 270, + 295, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 295, + 437 + ], + "type": "text", + "content": "Motivation. Nevertheless, DPO methods mainly focus on optimizing the LLM policy by comparing multiple generations " + }, + { + "bbox": [ + 54, + 270, + 295, + 437 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 54, + 270, + 295, + 437 + ], + "type": "text", + "content": " based on a fixed input " + }, + { + "bbox": [ + 54, + 270, + 295, + 437 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 54, + 270, + 295, + 437 + ], + "type": "text", + "content": ", falling short in capturing the dynamics of visual dialogue interactions. Specifically, high visual-dependent instructions require better visual understanding and reasoning ability to respond than easy-to-answer questions [9, 46]. Such discrepancy leads to responses with different amounts of visual clues. To quantitatively assess the visual dependency of each dialogue interaction, we utilized GPT-4o to review visual dialogues according to a set of manually designed rules. As shown in Fig. 2b, instructions with higher visual dependency tend to result in responses with more relevance to the visual input, leading to a preference dialogue interaction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 437, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 437, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 54, + 437, + 295, + 569 + ], + "type": "text", + "content": "Empirical evidence [35] suggests that training the policy to align with preference interaction can enhance visual understanding by maximizing the likelihood of preferred questions and answers. To this end, we propose Structured Policy Optimization (SPO). In SPO, we consider question and answer as two independent actions, each leading to a different state, simulating the conversational exchange. As shown in Fig. 1, the policy can be naturally modeled as a binary tree with depth 2. Unlike DPO, which only optimizes preference at the leaf nodes, SPO optimizes preferences across the entire sub-tree." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 576, + 258, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 258, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 258, + 590 + ], + "type": "text", + "content": "3.2. SPO Aligns With Dialogue Interaction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "text", + "content": "Preference Data Curation. To implement SPO, we first prompt the SFT model (given in Stage 1 in Fig. 3) to generate visual instruction-answer pairs, aiming to mitigate the distribution shift between SFT and the preference optimization step [3, 29]. Notably, similar to DPO, the SFT model also works as a reference policy " + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{ref}}" + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "text", + "content": " in SPO. However, the difference lies in our " + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{ref}}" + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "text", + "content": " being capable of self-generating questions by implementing SFT through learning to question [35]. As shown in Fig. 4, we represent different visual dialogues as trajectories in a tree structure - the" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 365, + 71, + 504, + 173 + ], + "blocks": [ + { + "bbox": [ + 365, + 71, + 504, + 173 + ], + "lines": [ + { + "bbox": [ + 365, + 71, + 504, + 173 + ], + "spans": [ + { + "bbox": [ + 365, + 71, + 504, + 173 + ], + "type": "image", + "image_path": "1ba21977451391ad8cea23ef9581256a7a7ecb2d2635a6040255d715911e0625.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 182, + 555, + 227 + ], + "lines": [ + { + "bbox": [ + 313, + 182, + 555, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 555, + 227 + ], + "type": "text", + "content": "Figure 4. Illustration of a tree structure that represents different visual dialogues, which are denoted as trajectory, e.g., " + }, + { + "bbox": [ + 313, + 182, + 555, + 227 + ], + "type": "inline_equation", + "content": "\\tau_{1} = (x, q_{w}, y_{w}^{1})" + }, + { + "bbox": [ + 313, + 182, + 555, + 227 + ], + "type": "text", + "content": ". For preference labeling, we utilize GPT-4o as a judge to perform pair-wise scoring over trajectory pairs." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 246, + 553, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 553, + 343 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 553, + 343 + ], + "type": "text", + "content": "input " + }, + { + "bbox": [ + 313, + 246, + 553, + 343 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 246, + 553, + 343 + ], + "type": "text", + "content": " denotes the root, questions occupy the first level, and answers are leaf nodes - and collect preference data " + }, + { + "bbox": [ + 313, + 246, + 553, + 343 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{x^{(i)},\\mathcal{Y}_w^{(i)},\\mathcal{Y}_l^{(i)}\\}_{i = 1}^N" + }, + { + "bbox": [ + 313, + 246, + 553, + 343 + ], + "type": "text", + "content": " by sub-trees*, where the preferred sub-tree " + }, + { + "bbox": [ + 313, + 246, + 553, + 343 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_w = (q_w,y_w^1,y_l^1)" + }, + { + "bbox": [ + 313, + 246, + 553, + 343 + ], + "type": "text", + "content": " and dispreferred subtree " + }, + { + "bbox": [ + 313, + 246, + 553, + 343 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_l = (q_l,y_w^2,y_l^2)" + }, + { + "bbox": [ + 313, + 246, + 553, + 343 + ], + "type": "text", + "content": ". We sample questions " + }, + { + "bbox": [ + 313, + 246, + 553, + 343 + ], + "type": "inline_equation", + "content": "q_{w},q_{l}\\sim \\pi_{\\mathrm{ref}}" + }, + { + "bbox": [ + 313, + 246, + 553, + 343 + ], + "type": "text", + "content": " for each visual input " + }, + { + "bbox": [ + 313, + 246, + 553, + 343 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 246, + 553, + 343 + ], + "type": "text", + "content": ", followed by answers " + }, + { + "bbox": [ + 313, + 246, + 553, + 343 + ], + "type": "inline_equation", + "content": "y_w^1,y_l^1\\sim" + }, + { + "bbox": [ + 313, + 246, + 553, + 343 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{ref}}(\\cdot |x,q_w))]" + }, + { + "bbox": [ + 313, + 246, + 553, + 343 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 246, + 553, + 343 + ], + "type": "inline_equation", + "content": "y_w^2,y_l^2\\sim \\pi_{\\mathrm{ref}}(\\cdot |x,q_l))]" + }, + { + "bbox": [ + 313, + 246, + 553, + 343 + ], + "type": "text", + "content": ", where the superscripts 1 and 2 indicate conditioning on " + }, + { + "bbox": [ + 313, + 246, + 553, + 343 + ], + "type": "inline_equation", + "content": "q_{w} / q_{l}" + }, + { + "bbox": [ + 313, + 246, + 553, + 343 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 343, + 555, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 343, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 555, + 570 + ], + "type": "text", + "content": "Empirically, the preference label accuracy is critical to increasing training stability. Following [45], AI-as-judge can offer scalable solutions to reduce human labor in complex task assessments and generate effective reward signals in reinforcement learning. Thus, to evaluate the synthetic data quality, we compared the distribution of 200 instruction-answer scores judged by GPT-4o [27] and an open-sourced LVLM (LLaVA-critic [45]). Fig. 5 shows that GPT-4o tends to judge with a higher standard than LLaVA-critic as its mean score is lower. We find that GPT-4o can provide scores with higher diversity than LLaVA-critic, acting more like a human annotator. Therefore, this work utilizes GPT-4o for all the preference labeling, and we propose a simple rule to filter out low-quality samples after collecting all the reviews. Specifically, we first remove samples with scores lower than a threshold, then remove the samples with repeated questions or answers, such as " + }, + { + "bbox": [ + 313, + 343, + 555, + 570 + ], + "type": "inline_equation", + "content": "q_{w} == q_{l}" + }, + { + "bbox": [ + 313, + 343, + 555, + 570 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 343, + 555, + 570 + ], + "type": "inline_equation", + "content": "y_{w}^{1} == y_{l}^{1}" + }, + { + "bbox": [ + 313, + 343, + 555, + 570 + ], + "type": "text", + "content": ". By this means, our collected preference data maintains high diversity and reliability." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 570, + 556, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 654 + ], + "type": "text", + "content": "Structured Reward Formulation. Given the initial state " + }, + { + "bbox": [ + 313, + 570, + 556, + 654 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 570, + 556, + 654 + ], + "type": "text", + "content": ", the preferred sub-tree " + }, + { + "bbox": [ + 313, + 570, + 556, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_w" + }, + { + "bbox": [ + 313, + 570, + 556, + 654 + ], + "type": "text", + "content": " should has higher reward since we assume the preference instruction dominates the dialogue interaction. We apply the same Bradley-Terry probability model to ensure that the reward of the preferred sub-tree " + }, + { + "bbox": [ + 313, + 570, + 556, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_w" + }, + { + "bbox": [ + 313, + 570, + 556, + 654 + ], + "type": "text", + "content": " exceeds the reward of the dispreferred sub-tree " + }, + { + "bbox": [ + 313, + 570, + 556, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_l" + }, + { + "bbox": [ + 313, + 570, + 556, + 654 + ], + "type": "text", + "content": " by" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 345, + 662, + 553, + 676 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 662, + 553, + 676 + ], + "spans": [ + { + "bbox": [ + 345, + 662, + 553, + 676 + ], + "type": "interline_equation", + "content": "p \\left(\\mathcal {Y} _ {w} \\succ \\mathcal {Y} _ {l} \\mid x\\right) = \\sigma \\left(r \\left(x, \\mathcal {Y} _ {w}\\right) - r \\left(x, \\mathcal {Y} _ {l}\\right)\\right). \\tag {5}", + "image_path": "d2a45271bf6933e7789046c07f5316ab7e378ff67d84c9e852efd87e03eacd93.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 683, + 553, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 683, + 553, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 683, + 553, + 696 + ], + "type": "text", + "content": "However, directly deriving " + }, + { + "bbox": [ + 313, + 683, + 553, + 696 + ], + "type": "inline_equation", + "content": "r(x,\\mathcal{Y})" + }, + { + "bbox": [ + 313, + 683, + 553, + 696 + ], + "type": "text", + "content": " into Eq. (2) will intro" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 703, + 553, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 703, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 325, + 703, + 553, + 712 + ], + "type": "text", + "content": "*For convenience, we omit the superscript " + }, + { + "bbox": [ + 325, + 703, + 553, + 712 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 325, + 703, + 553, + 712 + ], + "type": "text", + "content": " when no confusion occurs." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "text", + "content": "744" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": "duce a complex term " + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "Z(x)" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": " (please refer to the supplementary material for a complete derivation). To overcome this issue, we decompose sub-trees " + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": " to trajectories " + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": " and show how it connects to state-action trajectories in RLHF [8]. Let " + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\tau = (s_1, a_1, s_2, a_2, \\ldots, s_T, a_T)" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": " be a trajectory sampled from the policy " + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "s, a, T" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": " denote states, actions, and total steps. The total reward of a trajectory is defined as " + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "r(\\tau) = \\sum_{t=1}^{T} r(s_t, a_t)" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": ". Thus, following this reward formulation, we can define a trajectory to represent one visual dialogue by setting " + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "T = 2" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": ", where the initial state and action are input " + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": " and questioning " + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "s_1 = x, a_1 = q" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": ", and the following state and action are " + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "s_2 = [x, q], a_2 = y" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": ", e.g., " + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\tau_1 = (x, q_w, [x, q_w], y_w^1), \\tau_2 = (x, q_w, [x, q_w], y_l^1)" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": ", etc. Finally, we implement Eq. (5) by a trajectory reward as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 81, + 249, + 295, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 249, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 81, + 249, + 295, + 262 + ], + "type": "interline_equation", + "content": "r (x, \\mathcal {Y}) \\sim r (\\tau) = \\lambda \\cdot r (x, q) + \\delta \\cdot r ([ x, q ], y), \\tag {6}", + "image_path": "8bd460957b88a216598a7aa0d0f560132e19f27ad9e8ad1be2ddb22fd05c74a0.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 271, + 296, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 296, + 357 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 271, + 296, + 357 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 271, + 296, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 271, + 296, + 357 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 271, + 296, + 357 + ], + "type": "text", + "content": " are hyperparameters used to balance the question and answer rewards. By setting " + }, + { + "bbox": [ + 55, + 271, + 296, + 357 + ], + "type": "inline_equation", + "content": "\\lambda = \\delta" + }, + { + "bbox": [ + 55, + 271, + 296, + 357 + ], + "type": "text", + "content": ", our objective function could be equivalent to " + }, + { + "bbox": [ + 55, + 271, + 296, + 357 + ], + "type": "inline_equation", + "content": "\\sigma(r(\\tau_1) - r(\\tau_2))" + }, + { + "bbox": [ + 55, + 271, + 296, + 357 + ], + "type": "text", + "content": ", direct preference optimization (DPO) over two trajectory (i.e., question-answer pairs), rather than on how it is split across the binary tree. In the next section, we derive the closed-form expression of " + }, + { + "bbox": [ + 55, + 271, + 296, + 357 + ], + "type": "inline_equation", + "content": "r(x, q)" + }, + { + "bbox": [ + 55, + 271, + 296, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 271, + 296, + 357 + ], + "type": "inline_equation", + "content": "r([x, q], y)" + }, + { + "bbox": [ + 55, + 271, + 296, + 357 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 362, + 213, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 362, + 213, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 362, + 213, + 376 + ], + "type": "text", + "content": "3.3. The SPO Objective Function" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 380, + 295, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 295, + 404 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 295, + 404 + ], + "type": "text", + "content": "Following Eq. (2) and our trajectory-based reward model, the general RL objective of SPO is given as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 82, + 413, + 295, + 427 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 413, + 295, + 427 + ], + "spans": [ + { + "bbox": [ + 82, + 413, + 295, + 427 + ], + "type": "interline_equation", + "content": "J (\\pi_ {\\theta}) = \\mathbb {E} _ {\\tau \\sim \\pi_ {\\theta}} [ r (\\tau) ] - \\beta \\mathrm {K L} [ \\pi_ {\\theta} (\\tau) | | \\pi_ {\\text {r e f}} (\\tau) ]. \\tag {7}", + "image_path": "a84d1b1372c67938bde342ba0d78d526522c53d1c5e9b58f9cd09cf0c5e25556.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 434, + 274, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 274, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 274, + 448 + ], + "type": "text", + "content": "By substituting Eq. (6) into Eq. (7), we have " + }, + { + "bbox": [ + 55, + 434, + 274, + 448 + ], + "type": "inline_equation", + "content": "J(\\pi_{\\theta})" + }, + { + "bbox": [ + 55, + 434, + 274, + 448 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 453, + 297, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 453, + 297, + 510 + ], + "spans": [ + { + "bbox": [ + 58, + 453, + 297, + 510 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbb {E} _ {(x, q, y) \\sim \\tau} [ \\lambda r (x, q) + \\delta r ([ x, q ], y) ] - \\beta (\\mathrm {K L} [ \\pi_ {\\theta} (\\cdot | x) | | \\pi_ {\\operatorname {r e f}} (\\cdot | x) ]) \\\\ \\left. + \\mathrm {K L} \\left[ \\pi_ {\\theta} (\\cdot | [ x, q ]) \\mid \\mid \\pi_ {\\text {r e f}} (\\cdot | [ x, q ])) \\right]\\right), \\\\ = \\mathbb {E} _ {(x, q) \\sim \\tau} [ \\lambda r (x, q) ] - \\beta \\mathrm {K L} [ \\pi_ {\\theta} (\\cdot | x) | | \\pi_ {\\operatorname {r e f}} (\\cdot | x)) ] \\\\ + \\mathbb {E} _ {(x, q, y) \\sim \\tau} [ \\delta r ([ x, q ], y) ] - \\beta \\mathrm {K L} [ \\pi_ {\\theta} (\\cdot | [ x, q ]) | | \\pi_ {\\text {r e f}} (\\cdot | [ x, q ]) ] , \\\\ \\end{array}", + "image_path": "68f8460620294ede9f3e42b6d6e764b07c4c2f91e3fd2b750a7f0bfe30974477.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 515, + 295, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 295, + 540 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 515, + 295, + 540 + ], + "type": "inline_equation", + "content": "q\\sim \\pi_{\\theta}(\\cdot |x)" + }, + { + "bbox": [ + 55, + 515, + 295, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 515, + 295, + 540 + ], + "type": "inline_equation", + "content": "y\\sim \\pi_{\\theta}(\\cdot |[q,x])" + }, + { + "bbox": [ + 55, + 515, + 295, + 540 + ], + "type": "text", + "content": " . Following [41], the optimal solution of " + }, + { + "bbox": [ + 55, + 515, + 295, + 540 + ], + "type": "inline_equation", + "content": "J(\\pi_{\\theta})" + }, + { + "bbox": [ + 55, + 515, + 295, + 540 + ], + "type": "text", + "content": " can be derived as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 77, + 548, + 245, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 548, + 245, + 574 + ], + "spans": [ + { + "bbox": [ + 77, + 548, + 245, + 574 + ], + "type": "interline_equation", + "content": "\\pi_ {\\theta} (q | x) = \\frac {1}{Z (x)} \\pi_ {\\mathrm {r e f}} (q | x) \\exp (\\frac {\\lambda}{\\beta} r (x, q)),", + "image_path": "d25fed2b57fe75c943cb9ff3a8a9dd5648e3231f48cfda464dbaf69442fa9bd7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 575, + 290, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 575, + 290, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 575, + 290, + 601 + ], + "type": "interline_equation", + "content": "\\pi_ {\\theta} (y | [ x, q ]) = \\frac {1}{Z ([ x , q ])} \\pi_ {\\mathrm {r e f}} (y | [ x, q ]) \\exp (\\frac {\\delta}{\\beta} r ([ x, q ], y)).", + "image_path": "ef72bad7eb54bde01d31062c9c86013bf9da5dd9c28a649f5cc6159340e6438c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 609, + 234, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 234, + 619 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 234, + 619 + ], + "type": "text", + "content": "Thus, we can obtain the reward functions as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 84, + 628, + 295, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 628, + 295, + 654 + ], + "spans": [ + { + "bbox": [ + 84, + 628, + 295, + 654 + ], + "type": "interline_equation", + "content": "r (x, q) = \\frac {\\beta}{\\lambda} \\log \\frac {\\pi_ {\\theta} (q | x)}{\\pi_ {\\operatorname {r e f}} (q | x)} + \\frac {\\beta}{\\lambda} \\log Z (x), \\tag {8}", + "image_path": "9a86c5eeeceb9d12752cee91c13a61ea8fad16381040c3e3219300f76dbc8438.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 69, + 655, + 282, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 655, + 282, + 682 + ], + "spans": [ + { + "bbox": [ + 69, + 655, + 282, + 682 + ], + "type": "interline_equation", + "content": "r ([ x, q ], y) = \\frac {\\beta}{\\delta} \\log \\frac {\\pi_ {\\theta} (y | [ x , q ])}{\\pi_ {\\mathrm {r e f}} (y | [ x , q ])} + \\frac {\\beta}{\\delta} \\log Z ([ x, q ]).", + "image_path": "0a7aafd9b1fa42319adb09a73980a989d7151896e295c8fcfd4065fcfe24026f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Now, given the definition of each step reward, the total trajectory reward defined in Eq. (6) in terms of the optimal" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 316, + 72, + 553, + 200 + ], + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 200 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 200 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 200 + ], + "type": "image", + "image_path": "3a8e025debf3d9dbf9038d3c4f82ff12d52958bdc234868b3ed24a46d566e4e8.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 217, + 555, + 261 + ], + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 261 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 261 + ], + "type": "text", + "content": "Figure 5. The score distribution of questions and answers, and labeling agreement with 3 LVLMS as judges. GPT-4o provides lower mean score and higher diversity than LLaVA-critic, acting more like a human annotator." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 289, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 555, + 348 + ], + "type": "text", + "content": "policy can be directly substituted into the preference model in Eq. (1), where the " + }, + { + "bbox": [ + 313, + 289, + 555, + 348 + ], + "type": "inline_equation", + "content": "Z(x)" + }, + { + "bbox": [ + 313, + 289, + 555, + 348 + ], + "type": "text", + "content": " term will be canceled just as the original DPO derivation. Assuming " + }, + { + "bbox": [ + 313, + 289, + 555, + 348 + ], + "type": "inline_equation", + "content": "\\tau_w = (x, q_1, [x, q_1], y_1)" + }, + { + "bbox": [ + 313, + 289, + 555, + 348 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 289, + 555, + 348 + ], + "type": "inline_equation", + "content": "\\tau_l = (x, q_2, [x, q_2], y_2)" + }, + { + "bbox": [ + 313, + 289, + 555, + 348 + ], + "type": "text", + "content": ", the preference probability of preferred " + }, + { + "bbox": [ + 313, + 289, + 555, + 348 + ], + "type": "inline_equation", + "content": "\\tau_w" + }, + { + "bbox": [ + 313, + 289, + 555, + 348 + ], + "type": "text", + "content": " over dispreferred " + }, + { + "bbox": [ + 313, + 289, + 555, + 348 + ], + "type": "inline_equation", + "content": "\\tau_l" + }, + { + "bbox": [ + 313, + 289, + 555, + 348 + ], + "type": "text", + "content": " can be defined as" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 361, + 556, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 361, + 556, + 426 + ], + "spans": [ + { + "bbox": [ + 315, + 361, + 556, + 426 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} p \\left(\\tau_ {w} \\succ \\tau_ {l}\\right) := \\sigma \\left(\\frac {\\beta}{\\lambda} \\log \\frac {\\pi_ {\\theta} \\left(q _ {1} \\mid x\\right)}{\\pi_ {\\operatorname {r e f}} \\left(q _ {1} \\mid x\\right)} + \\frac {\\beta}{\\delta} \\log \\frac {\\pi_ {\\theta} \\left(y _ {1} \\mid [ x , q _ {1} ]\\right)}{\\pi_ {\\operatorname {r e f}} \\left(y _ {1} \\mid [ x , q _ {1} ]\\right)} \\right. \\\\ \\left. - \\frac {\\beta}{\\lambda} \\log \\frac {\\pi_ {\\theta} \\left(q _ {2} \\mid x\\right)}{\\pi_ {\\text {r e f}} \\left(q _ {2} \\mid x\\right)} - \\frac {\\beta}{\\delta} \\log \\frac {\\pi_ {\\theta} \\left(y _ {2} \\mid [ x , q _ {2} ]\\right)}{\\pi_ {\\text {r e f}} \\left(y _ {2} \\mid [ x , q _ {2} ]\\right)}\\right). \\tag {9} \\\\ \\end{array}", + "image_path": "570213cbc4ddcba4dbc8e9767da1212e9891a74691f8a16daf26be743f30e78d.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 426, + 554, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 554, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 554, + 450 + ], + "type": "text", + "content": "More details about the derivation from Eq. (8) to Eq. (10) could be found in the supplementary." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 453, + 554, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 554, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 554, + 502 + ], + "type": "text", + "content": "As shown in Fig. 4, we reformulate the preference data based on trajectories as " + }, + { + "bbox": [ + 313, + 453, + 554, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{\\tau_1^{(i)},\\tau_2^{(i)},\\tau_3^{(i)},\\tau_4^{(i)}\\}_{i = 1}^N" + }, + { + "bbox": [ + 313, + 453, + 554, + 502 + ], + "type": "text", + "content": ". By substituting Eq. (8) into Eq. (9), the loss function of classifying between two trajectories is given as" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 517, + 561, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 517, + 561, + 562 + ], + "spans": [ + { + "bbox": [ + 316, + 517, + 561, + 562 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} \\left(\\tau_ {w}, \\tau_ {l}\\right) = - \\mathbb {E} _ {\\left(\\tau_ {w}, \\tau_ {l}\\right) \\sim \\mathcal {D}} \\left[ \\log \\sigma \\left(r \\left(\\tau_ {w}\\right) - r \\left(\\tau_ {l}\\right)\\right) \\right] \\tag {10} \\\\ \\leq \\mathbb {E} _ {(\\tau_ {w}, \\tau_ {l}) \\sim \\mathcal {D}} [ - \\log \\sigma ((r (x, q _ {1}) - r (x, q _ {2}))) ] \\\\ + \\mathbb {E} _ {\\left(\\tau_ {w}, \\tau_ {t}\\right) \\sim \\mathcal {D}} [ - \\log \\sigma \\left(\\left(r ([ x, q _ {1} ], y _ {1}) - r ([ x, q _ {2} ], y _ {2}))\\right) \\right]. \\\\ \\end{array}", + "image_path": "67cf24f8ae1cb4b9bac2d30d286bf1fd7e497e4db642edf16c0dec22beb6569d.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 575, + 555, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 555, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 555, + 623 + ], + "type": "text", + "content": "Eventually, we define the preference probability between sub-tree " + }, + { + "bbox": [ + 313, + 575, + 555, + 623 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_w" + }, + { + "bbox": [ + 313, + 575, + 555, + 623 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 575, + 555, + 623 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_l" + }, + { + "bbox": [ + 313, + 575, + 555, + 623 + ], + "type": "text", + "content": " as a joint of multiple trajectory-based preference probability, improving the comprehension of policy distribution. The final SPO loss is formulated as" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 331, + 637, + 553, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 637, + 553, + 662 + ], + "spans": [ + { + "bbox": [ + 331, + 637, + 553, + 662 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S P O} = \\mathcal {L} _ {\\left(\\mathcal {Y} _ {w}, \\mathcal {Y} _ {i}\\right)} := \\sum_ {m, n} \\mathcal {L} \\left(\\tau_ {m}, \\tau_ {n}\\right): m < n, \\tag {11}", + "image_path": "d9336d837673c64bb664ba2dd4897119d6f412e60f280bb41b2802c480a0f068.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\tau_m,\\tau_n)" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": " is given in Eq. (10), and " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "m\\in \\Phi_w = \\{1,3\\}" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "n\\in \\Phi_l = \\{2,4\\}" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": " are the index of preferred and dispreferred trajectories." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "745" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 86, + 102, + 526, + 296 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Evaluation results of our model trained with different preference optimization methods based on the SFT model on a scale of 0.5B and 7B. We report accuracy on all the benchmarks except MME. The best results are bold, and the second-best results are underlined." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 86, + 102, + 526, + 296 + ], + "lines": [ + { + "bbox": [ + 86, + 102, + 526, + 296 + ], + "spans": [ + { + "bbox": [ + 86, + 102, + 526, + 296 + ], + "type": "table", + "html": "
ModelSingle-Image
MMStarMMMUMMEL-WilderLLaVA-WMMVetRealWorldQAPOPEMMBench
testvaltestsmalltesttesttesttestMulti-discip
GPT-4V [26]57.156.8140981.098.049.961.4-75.0
GPT-4o [27]-69.1-85.9106.176.258.6--
Gemini-1.5-Pro [36]-62.2----70.4--
InternVL-2-8B [7]59.449.3221062.584.560.064.4-81.7
IXC-2.5-7B [50]59.942.9222961.478.151.767.8-82.2
LLaVA-OV-72B [18]66.156.8168272.093.563.771.9-85.9
LLaVA-OV-0.5B [18]37.531.4123855.074.229.155.688.052.1
SFT-0.5B [35]40.731.9124659.472.826.755.788.053.1
SFT-0.5B + DPO [29]40.932.1125058.573.525.856.488.154.0
SFT-0.5B + SPO (ours)41.532.8125161.374.130.056.988.254.3
LLaVA-OV-7B [18]61.748.8158067.890.757.566.389.280.8
SFT-7B [35]59.648.0156569.590.557.666.889.079.7
SFT-7B + DPO [29]61.949.0157969.991.258.267.289.180.6
SFT-7B + SPO (ours)62.149.2158570.292.358.867.389.381.0
", + "image_path": "ad659c9df4e1cd1e3456f8b4375c9bf3ae16dad412d2dc49fbd3af1206e0da49.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 87, + 337, + 525, + 547 + ], + "blocks": [ + { + "bbox": [ + 55, + 305, + 555, + 329 + ], + "lines": [ + { + "bbox": [ + 55, + 305, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 55, + 305, + 555, + 329 + ], + "type": "text", + "content": "Table 2. Evaluation results of different methods on multi-image and video benchmarks. #Frames is the number of frames used for video encoding. We report accuracy on all the benchmarks. The best results are bold, and the second-best results are underlined." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 87, + 337, + 525, + 547 + ], + "lines": [ + { + "bbox": [ + 87, + 337, + 525, + 547 + ], + "spans": [ + { + "bbox": [ + 87, + 337, + 525, + 547 + ], + "type": "table", + "html": "
Model#FramesMulti-ImageEgoSchemaVideo
\\(LLaVA^I\\)MuirBenchMLVUNExT-QALVid-BenchVideoMME
out-domainout-domaintestm-avgmcvalw/o subs
GPT-4V [26]1060.362.3-49.2-61.359.9
GPT-4o [27]384-68.0-64.6-66.771.9
Gemini-1.5-Pro [36]1/s-49.472.2--64.075.0
LongVA-7B [51]128---56.368.3-52.6
IXC-2.5-7B [50]64---37.371.0-55.8
LLaVA-Video-7B [53]64--57.370.883.258.263.3
LLaVA-Video-72B [53]64--65.674.485.461.970.5
LLaVA-OV-72B [18]3279.954.862.068.080.261.366.3
LLaVA-OV-0.5B [18]3233.325.526.850.357.245.844.0
SFT-0.5B [35]3232.726.128.851.759.044.644.3
SFT-0.5B + DPO [29]3232.927.928.952.059.246.144.8
SFT-0.5B + SPO (ours)3233.428.029.252.460.447.745.4
LLaVA-OV-7B [18]3264.241.860.164.779.456.458.1
SFT-7B [35]3266.845.560.064.977.257.557.8
SFT-7B + DPO [29]3266.946.259.565.179.858.158.3
SFT-7B + SPO (ours)3267.946.760.565.379.657.858.5
", + "image_path": "0de40dfe89f906415724b99fce5bfbb44ff4b2980b683f96401b5fad223e261d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 564, + 137, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 137, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 137, + 578 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 586, + 153, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 153, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 153, + 600 + ], + "type": "text", + "content": "4.1. Training Recipe" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "text", + "content": "We follow [18] to construct our model architecture, including the instructional language model, vision encoder, and a 2-layer MLP projector. This work adopts LoRA [15] for an efficient and lightweight fine-tuning, where we set low-rank adapters with rank " + }, + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "inline_equation", + "content": "= 64" + }, + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\alpha = 128" + }, + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "text", + "content": " for the LLM. We denote " + }, + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "text", + "content": " as all the trainable parameters, including LoRA and projector layer. As shown in Fig. 3, the training pipeline has two stages: 1) Utilizing supervised fine-tuning (SFT) to train an SFT model by reasoning and learning to ques" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 565, + 555, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 565, + 555, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 565, + 555, + 615 + ], + "type": "text", + "content": "tion [35]. 2) Fine-tune the LVML policy (given the SFT model as a reference) using the proposed SPO approach by maximizing a novel structured reward function. We show the training details for each stage in supplementary." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 624, + 451, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 451, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 451, + 635 + ], + "type": "text", + "content": "4.2. Evaluation Benchmarks" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "We primarily assess our models using vision-language benchmarks across different visual modalities (e.g., image, multi-image, and video) that are not included in the training data. For a fair comparison with other LVLMs, we primarily report results from original papers and utilize Immseval [49] for test setting consistency. All our results are" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "746" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 148, + 137 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 148, + 137 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 148, + 137 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 148, + 137 + ], + "type": "image", + "image_path": "f13afe7ae10049180636b63708a58286d0782643ca3b3857014a87f59b4fd5b2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 251, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 555, + 285 + ], + "type": "text", + "content": "Figure 6. Qualitative results of LLaVA-Bench-Wilder. When trained with DPO, the model becomes over-confident when indicating low visual dependency content. In contrast, SPO improves the model's visual reasoning ability and prevents over-confident prediction when explaining uncertain situations, acting similarly to the reference model, GPT-4V." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 149, + 139, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 149, + 139, + 232 + ], + "spans": [ + { + "bbox": [ + 61, + 149, + 139, + 232 + ], + "type": "text", + "content": "Are the otters in the image reading a book, and if so, can you provide a detailed description of what might be happening in this whimsical scene?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 151, + 72, + 548, + 241 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 151, + 72, + 545, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 72, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 151, + 72, + 545, + 121 + ], + "type": "text", + "content": "SPO-7B Based on the image, it appears that the otters are interacting with a book in an unusual and anthropomorphic manner. The otters seem to be holding the book open as if they are reading it, which is a humorous and imaginative interpretation of their behavior. The book's placement between them suggests a shared interest or activity. The scene is whimsical because otters do not engage in such human-like behaviors naturally; this image is likely staged for artistic or entertainment purposes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 151, + 129, + 547, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 129, + 547, + 167 + ], + "spans": [ + { + "bbox": [ + 151, + 129, + 547, + 167 + ], + "type": "text", + "content": "DPO-7B The otters in the image are likely to be sea otters, as they are shown floating on their backs in water and holding a book between them. The fact that they are holding a book also indicates that they have been exposed to human activities or objects, which could mean they live in an area where humans frequently visit or interact with them." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 172, + 548, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 172, + 548, + 241 + ], + "spans": [ + { + "bbox": [ + 152, + 172, + 548, + 241 + ], + "type": "text", + "content": "GPT-4V Yes, the otters in the image appear to be reading a book, which is a whimsical and anthropomorphized activity for animals. The scene is set in a body of water, with the two otters floating and seemingly fully absorbed in the book. One otter holds the book with one hand and points to a line or page with the other, as if highlighting something or guiding the reading process. The second otter is focused on the book, with an expression of curiosity or concentration. This playful image brings a charming and imaginative element, showcasing animals engaging in human-like activities and suggesting themes of education, curiosity, and companionship. Reference answer" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 305, + 260, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 305, + 260, + 317 + ], + "spans": [ + { + "bbox": [ + 55, + 305, + 260, + 317 + ], + "type": "text", + "content": "reported with greedy sampling and 0-shot settings." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 318, + 295, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 295, + 400 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 295, + 400 + ], + "type": "text", + "content": "Single-image Benchmarks. We consider the LVLM as a general-purpose assistant in the wild as the most critical metric. To validate the capabilities in real-world scenarios with open-form instructions, we utilize MMStar [5], LLaVA-Bench-Wilder [18], and RealWorldQA [43]. Besides chat ability, visual perception evaluates the model's reasoning ability, so we adopt MME [13] and MMMU [48]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 401, + 295, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 295, + 520 + ], + "type": "text", + "content": "Multi-image Benchmarks. Multi-image tasks test the model's ability to connect and reason across images, making them generally more challenging as they require higher-level contextual integration. This work utilizes LLaVA-Interleave Bench [20] under the out-of-domain setting that does not overlap with training scenarios. This setting covers 13 challenging real-world tasks such as spotting the difference and visual storytelling, etc. We also provide a comprehensive and holistic evaluation of MuirBench [37], which contains 12 diverse multi-image tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 521, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 604 + ], + "type": "text", + "content": "Video Benchmarks. Video understanding is a critical ability to build a general-purpose vision-language assistant. This work conducts several open-ended and multi-choice video benchmarks. EgoSchema [25] and MLVU [54] focus on long video understanding, NExTQA [44], LongVideoBench [40] and VideoMME [14] contains diverse video domains and duration." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 612, + 140, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 140, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 140, + 623 + ], + "type": "text", + "content": "4.3. Main Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "In Table 1, while both preference optimization algorithms enhance performance over the SFT model, SPO achieves the best overall performance across all the image benchmarks under 0.5B and 7B scale. The consistency highlights the robustness and effectiveness of SPO. Notably, SPO outperforms its baseline model LLaVA-OV by 6.3 (0.5B) and 2.4 (7B) points on LLaVA-Bench-Wilder, and the improve" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 305, + 554, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 305, + 554, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 554, + 388 + ], + "type": "text", + "content": "ment on RealWorldQA indicates that SPO enhances the model's visual reasoning capability while handling freeform visual chat and real-world scenario. Such improvement follows our expectations since we precisely optimize the model to capture the alignment between visual input and dialogue interaction. Fig. 6 shows a qualitative example of real-world visual chat from three models." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 389, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 554, + 521 + ], + "type": "text", + "content": "We have observed an improvement of SPO on the multi-image benchmarks, demonstrating the significance of our method for improving comprehensive and complex visual reasoning. Specifically, SPO-0.5B achieves a " + }, + { + "bbox": [ + 313, + 389, + 554, + 521 + ], + "type": "inline_equation", + "content": "4.4\\%" + }, + { + "bbox": [ + 313, + 389, + 554, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 389, + 554, + 521 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 313, + 389, + 554, + 521 + ], + "type": "text", + "content": " improvement, and SPO-7B achieves an " + }, + { + "bbox": [ + 313, + 389, + 554, + 521 + ], + "type": "inline_equation", + "content": "8.1\\%" + }, + { + "bbox": [ + 313, + 389, + 554, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 389, + 554, + 521 + ], + "type": "inline_equation", + "content": "2.0\\%" + }, + { + "bbox": [ + 313, + 389, + 554, + 521 + ], + "type": "text", + "content": " improvement over LLaVA-OV and our SFT model. Table 2 also provides evaluation results on video benchmarks. The consistent improvement of SPO-0.5b/ 7B on EgoSchema, MLVU, and L-VideoBench indicates that SPO also improves consecutive visual reasoning, especially for medium and long video understanding." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 521, + 554, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 700 + ], + "type": "text", + "content": "SPO has better visual-question understanding ability than DPO since it encourages visually informative questions by utilizing the structured preference tree that captures diverse question-answer pairs. Regarding the low-quality questions, SPO also learns their preferred answers through a subtree within the structured preferences. To validate our hypothesis, we prompt GPT-4o to score 216 questions in the MMVet dataset and split them into high-dependency (High-Q) and low-dependency (Low-Q) visual questions. Table 4 shows that more visually focused questions consistently improve the answering accuracy given images. As expected, without looking at images (w/o images), Low-Q reports a smaller performance drop than High-Q, and SPO has lower performance drop than DPO. Empirical evidence shows that SPO is more robust to question quality than DPO." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "text", + "content": "Limitation. While SPO enhances the SFT model, we" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "747" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 323, + 135 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 323, + 135 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 323, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 323, + 135 + ], + "type": "table", + "html": "
r(x,q)r([x,q],y)ImageMulti-ImageVideo
MMELLaVA-WilderLLaVA1MuirBenchEgoSchemaMLVU
✓✓158570.267.946.760.565.3
✓X156969.967.346.260.065.2
X✓157169.767.246.360.165.0
XX156569.566.845.560.064.9
", + "image_path": "98cb268d47fd151d0ee78be872e018ae739fa8031f259f461263ae6da2d6126a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 331, + 73, + 553, + 135 + ], + "blocks": [ + { + "bbox": [ + 331, + 73, + 553, + 135 + ], + "lines": [ + { + "bbox": [ + 331, + 73, + 553, + 135 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 553, + 135 + ], + "type": "image", + "image_path": "bc1a891cb3414ad7f76c576ea366720b1c39daec88bd0b8787d4c406dec59de1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 59, + 208, + 293, + 249 + ], + "blocks": [ + { + "bbox": [ + 55, + 144, + 555, + 179 + ], + "lines": [ + { + "bbox": [ + 55, + 144, + 555, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 555, + 179 + ], + "type": "text", + "content": "Table 3. Ablation study under SFT-7B-SPO. We ablate the trajectory re- Figure 7. Left. SPO achieves higher preference classification reward design of SPO. The formal trajectory reward is " + }, + { + "bbox": [ + 55, + 144, + 555, + 179 + ], + "type": "inline_equation", + "content": "r(\\tau) = r(x,q) +" + }, + { + "bbox": [ + 55, + 144, + 555, + 179 + ], + "type": "text", + "content": " accuracy than DPO throughout training. Right. SPO maxi" + }, + { + "bbox": [ + 55, + 144, + 555, + 179 + ], + "type": "inline_equation", + "content": "r([x,q],y)" + }, + { + "bbox": [ + 55, + 144, + 555, + 179 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 144, + 555, + 179 + ], + "type": "inline_equation", + "content": "r(x,q)" + }, + { + "bbox": [ + 55, + 144, + 555, + 179 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 144, + 555, + 179 + ], + "type": "inline_equation", + "content": "r([x,q],y)" + }, + { + "bbox": [ + 55, + 144, + 555, + 179 + ], + "type": "text", + "content": " are question and answer reward. mizes the reward margin faster effectiveness than DPO." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 197, + 288, + 207 + ], + "lines": [ + { + "bbox": [ + 61, + 197, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 61, + 197, + 288, + 207 + ], + "type": "text", + "content": "Table 4. Test accuracy drop in two question splits on MMVet." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 208, + 293, + 249 + ], + "lines": [ + { + "bbox": [ + 59, + 208, + 293, + 249 + ], + "spans": [ + { + "bbox": [ + 59, + 208, + 293, + 249 + ], + "type": "table", + "html": "
ModelsHigh-Q (w/o image)Low-Q (w/o image)
SFT-0.5B + DPO27.0 (6.0) ↓ 77.8%17.0 (6.0) ↓ 64.7%
SFT-0.5B + SPO29.0 (8.0) ↓ 72.4%23.0 (12.0) ↓ 47.8%
", + "image_path": "40109b037f7ca30cf62057e6f51b0d93216cbfda5af806593650eed4fc844511.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 269, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 269, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 54, + 269, + 295, + 460 + ], + "type": "text", + "content": "want to emphasize the critical role of the SFT model as the foundational stage for all preference optimization methods. Notably, the limited computational resources primarily drive the performance decline after SFT on several benchmarks. Our SFT-7B model was fine-tuned with 8 GPUs — only 1/16 of training GPUs in LLaVA-OV — resulting in a much smaller batch size for our approach. Specifically, the performance degradation of SFT-7B on MMStar, MMU, and MME sets a lower bound for the subsequent gains achievable by SPO. A similar trend is observed in video benchmarks, as shown in Table 2. Additionally, we decrease the anyres [18] to " + }, + { + "bbox": [ + 54, + 269, + 295, + 460 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 54, + 269, + 295, + 460 + ], + "type": "text", + "content": " of the regular setting for visual encoding to fit in our computational resource. Due to limited computational resources, SPO mainly focuses on single-turn conversations. We leave multi-turn conversations study in future work." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 468, + 160, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 160, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 160, + 479 + ], + "type": "text", + "content": "4.4. Model Discussion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 486, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 486, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 486, + 295, + 677 + ], + "type": "text", + "content": "Structured Reward Speeds Up Policy Learning. The Bradley-Terry objective function essentially aims to optimize the reward margin of the preferred and dispreferred action " + }, + { + "bbox": [ + 56, + 486, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\Delta r(x, \\rho_w, \\rho_l) = r(x, \\rho_w) - r(x, \\rho_l)" + }, + { + "bbox": [ + 56, + 486, + 295, + 677 + ], + "type": "text", + "content": ". In Eq. (10), SPO optimizes a joint reward margin including " + }, + { + "bbox": [ + 56, + 486, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\Delta r(x, q_1, q_2)" + }, + { + "bbox": [ + 56, + 486, + 295, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 486, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\Delta r([x, q_1], y_1, y_2)" + }, + { + "bbox": [ + 56, + 486, + 295, + 677 + ], + "type": "text", + "content": ". While DPO only optimizes the reward margin between responses " + }, + { + "bbox": [ + 56, + 486, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\Delta r([x, q_1], y_1, y_2)" + }, + { + "bbox": [ + 56, + 486, + 295, + 677 + ], + "type": "text", + "content": ". We use the SPO and DPO models to investigate the relationship between the learned reward margin and the preference classification accuracy. Fig. 7 shows that SPO maximizes the reward margin more effectively than DPO, and SPO achieves higher preference classification accuracy than DPO along all the training steps. By precisely adjusting " + }, + { + "bbox": [ + 56, + 486, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 56, + 486, + 295, + 677 + ], + "type": "text", + "content": " for different preference probabilities, we find that optimizing multiple trajectory pairs together can explicitly increase the data diversity, leading to better generalizability and performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Influence of Each Step Reward. SPO optimizes a policy to satisfy multiple preferences in a single objective function. In Eq. (10), we use " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " to control the deviation of" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 321, + 208, + 549, + 248 + ], + "blocks": [ + { + "bbox": [ + 333, + 197, + 535, + 207 + ], + "lines": [ + { + "bbox": [ + 333, + 197, + 535, + 207 + ], + "spans": [ + { + "bbox": [ + 333, + 197, + 535, + 207 + ], + "type": "text", + "content": "Table 5. Latency vs Accuracy on the LLaVA-W dataset" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 208, + 549, + 248 + ], + "lines": [ + { + "bbox": [ + 321, + 208, + 549, + 248 + ], + "spans": [ + { + "bbox": [ + 321, + 208, + 549, + 248 + ], + "type": "table", + "html": "
MetricSFT-7BSFT-7B + DPOSFT-7B + SPO
Accuracy (%)72.873.574.1
Inference time (s)523547254
", + "image_path": "5a3cab762a511ca0a63c8e0eac88353e36e36104e3da3ccc286416ab3cdd85d0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 259, + 553, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 259, + 553, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 553, + 368 + ], + "type": "text", + "content": "preference alignment over two action groups, questioning and answering. Following the same training strategy and preference data, we evaluate the performance of SPO w/o question reward " + }, + { + "bbox": [ + 313, + 259, + 553, + 368 + ], + "type": "inline_equation", + "content": "r(x,q) = 0" + }, + { + "bbox": [ + 313, + 259, + 553, + 368 + ], + "type": "text", + "content": " and SPO w/o answer reward " + }, + { + "bbox": [ + 313, + 259, + 553, + 368 + ], + "type": "inline_equation", + "content": "r([x,q] = y) = 0" + }, + { + "bbox": [ + 313, + 259, + 553, + 368 + ], + "type": "text", + "content": ". As shown in Table. 3, learning question policy improves real-world QA ability such as LLaVA-Wilder and LLaVA; learning answer policy brings more improvement over fixed-form QA such as MME. Overall, the joint policy improves the model from different aspects." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 369, + 554, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 554, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 554, + 487 + ], + "type": "text", + "content": "Computational Resource Analysis. For training, optimizing the whole dialogue will not burden the computation cost when comparing SPO with DPO (e.g., SPO: 24h vs DPO: 22h using " + }, + { + "bbox": [ + 313, + 369, + 554, + 487 + ], + "type": "inline_equation", + "content": "4^{*}\\mathrm{A}100" + }, + { + "bbox": [ + 313, + 369, + 554, + 487 + ], + "type": "text", + "content": " GPUs) since only the question tokens need to be unmasked. For inference, Table 5 evaluates latency by reporting the total inference time of three models on 60 open-ended questions from the LLaVA-W dataset, and the test accuracy is measured by GPT-4o. Empirically, SPO generates more concise and accurate content than the SFT and DPO models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 501, + 388, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 501, + 388, + 514 + ], + "spans": [ + { + "bbox": [ + 314, + 501, + 388, + 514 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "content": "In this work, we have proposed Structured Policy Optimization (SPO), a novel approach to enhancing multi-modal preference optimization for large vision-language models (LVLMs). SPO alleviates the limitation of existing DPO methods, which adopt a single, unstructured reward and tend to prioritize linguistic nuances over visual context. Specifically, SPO introduces a new structured reward represented in trajectories to improve the visual dependency by maximizing the reward between preferred and dispreferred questions/answers. Our experimental results demonstrate that SPO improves vision-language understanding and reasoning across diverse tasks, showcasing superior alignment with human-preferred responses compared to traditional DPO methods. Overall, this approach better captures the dynamic interplay between visual inputs, questions, and responses, aligning more effectively with human preferences." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "748" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 114 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 114 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 114 + ], + "type": "text", + "content": "This research was supported by the DEVCOM Army Research Laboratory under Contract W911QX-21-D-0001." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 135, + 115, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 135, + 115, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 135, + 115, + 148 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 155, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 155, + 296, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 155, + 296, + 198 + ], + "spans": [ + { + "bbox": [ + 61, + 155, + 296, + 198 + ], + "type": "text", + "content": "[1] OpenAI Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, et al. Gpt-4 technical report. 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 200, + 296, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 200, + 296, + 255 + ], + "spans": [ + { + "bbox": [ + 61, + 200, + 296, + 255 + ], + "type": "text", + "content": "[2] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 256, + 295, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 256, + 295, + 299 + ], + "spans": [ + { + "bbox": [ + 62, + 256, + 295, + 299 + ], + "type": "text", + "content": "[3] Hritik Bansal, Ashima Suvarna, Gantavya Bhatt, Nanyun Peng, Kai-Wei Chang, and Aditya Grover. Comparing bad apples to good oranges: Aligning large language models via joint preference optimization. ArXiv, 2024. 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 300, + 294, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 300, + 294, + 343 + ], + "spans": [ + { + "bbox": [ + 62, + 300, + 294, + 343 + ], + "type": "text", + "content": "[4] Lin Chen, Jinsong Li, Xiao wen Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. In ArXiv, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 345, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 345, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 62, + 345, + 294, + 398 + ], + "type": "text", + "content": "[5] Lin Chen, Jinsong Li, Xiao wen Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, and Feng Zhao. Are we on the right way for evaluating large vision-language models? In NeurIPS, 2024. 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 401, + 294, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 401, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 401, + 294, + 434 + ], + "type": "text", + "content": "[6] Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji, and Quanquan Gu. Self-play fine-tuning converts weak language models to strong language models. ArXiv, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 435, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 435, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 62, + 435, + 294, + 499 + ], + "type": "text", + "content": "[7] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 502, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 294, + 555 + ], + "type": "text", + "content": "[8] Paul F. Christiano, Jan Leike, Tom B. Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences. In Proceedings of the 31st International Conference on Neural Information Processing Systems, 2017. 1, 3, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 557, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 294, + 589 + ], + "type": "text", + "content": "[9] Ruiguo Cui and Peter Teo. Thinking through talk: Using dialogue to develop students' critical thinking. Teaching and Teacher Education, 2023. 2, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 591, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 294, + 645 + ], + "type": "text", + "content": "[10] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Albert Li, Pascale Fung, and Steven C. H. Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "type": "text", + "content": "[11] Yihe Deng, Pan Lu, Fan Yin, Ziniu Hu, Sheng Shen, James Zou, Kai-Wei Chang, and Wei Wang. Enhancing large vision language models with self-training on image comprehension. In NeurIPS, 2024. 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 714 + ], + "type": "text", + "content": "[12] Hanze Dong, Wei Xiong, Deepanshu Goyal, Rui Pan, Shizhe Diao, Jipeng Zhang, Kashun Shum, and T. Zhang. Raft:" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Reward ranked finetuning for generative foundation model alignment. ArXiv, 2023. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 97, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 150 + ], + "type": "text", + "content": "[13] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. ArXiv, 2023. 1, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 152, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 228 + ], + "type": "text", + "content": "[14] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, Peixian Chen, Yanwei Li, Shaohui Lin, Sirui Zhao, Ke Li, Tong Xu, Xiawu Zheng, Enhong Chen, Rongrong Ji, and Xing Sun. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. ArXiv, 2024. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 230, + 553, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 230, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 317, + 230, + 553, + 273 + ], + "type": "text", + "content": "[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 275, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 275, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 317, + 275, + 553, + 308 + ], + "type": "text", + "content": "[16] Dongfu Jiang, Xuan He, Huaye Zeng, Cong Wei, Max W.F. Ku, Qian Liu, and Wenhu Chen. Mantis: Interleaved multi-image instruction tuning. ArXiv, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 309, + 553, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 309, + 553, + 351 + ], + "spans": [ + { + "bbox": [ + 317, + 309, + 553, + 351 + ], + "type": "text", + "content": "[17] Xin Lai, Zhuotao Tian, Yukang Chen, Senqiao Yang, Xiangru Peng, and Jiaya Jia. Step-dpo: Step-wise preference optimization for long-chain reasoning of llms. ArXiv, 2024. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 354, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 553, + 397 + ], + "type": "text", + "content": "[18] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. ArXiv, 2024. 2, 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 399, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 453 + ], + "type": "text", + "content": "[19] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. In NeurIPS, 2023. 1, 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 455, + 553, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 455, + 553, + 497 + ], + "spans": [ + { + "bbox": [ + 317, + 455, + 553, + 497 + ], + "type": "text", + "content": "[20] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava-last-interleave: Tackling multi-image, video, and 3d in large multimodal models. ArXiv, 2024. 2, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 499, + 553, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 499, + 553, + 554 + ], + "spans": [ + { + "bbox": [ + 317, + 499, + 553, + 554 + ], + "type": "text", + "content": "[21] Juncheng Li, Kaihang Pan, Zhiqi Ge, Minghe Gao, Hanwang Zhang, Wei Ji, Wenqiao Zhang, Tat seng Chua, Siliang Tang, and Yueting Zhuang. Fine-tuning multimodal llms to follow zero-shot demonstrative instructions. In International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 555, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 555, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 317, + 555, + 553, + 599 + ], + "type": "text", + "content": "[22] Lei Li, Zhihui Xie, Mukai Li, Shunian Chen, Peiyi Wang, Liang Chen, Yazheng Yang, Benyou Wang, and Lingpeng Kong. Silkie: Preference distillation for large visual language models. ArXiv, 2023. 1, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 601, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 601, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 317, + 601, + 553, + 632 + ], + "type": "text", + "content": "[23] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. ArXiv, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 635, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 666 + ], + "type": "text", + "content": "[24] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In ArXiv, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 668, + 553, + 701 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 668, + 553, + 701 + ], + "spans": [ + { + "bbox": [ + 316, + 668, + 553, + 701 + ], + "type": "text", + "content": "[25] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. ArXiv, 2023. 1, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 703, + 507, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 703, + 507, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 703, + 507, + 713 + ], + "type": "text", + "content": "[26] OpenAI. Gpt-4v(ision) system card. 2023. 2, 6" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "749" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 235, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 235, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 235, + 83 + ], + "type": "text", + "content": "[27] OpenAI. Gpt-4o system card. 2024. 2, 4, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 85, + 294, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 85, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 85, + 294, + 171 + ], + "type": "text", + "content": "[28] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke E. Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Francis Christiano, Jan Leike, and Ryan J. Lowe. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 294, + 216 + ], + "type": "text", + "content": "[29] Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. In NeurIPS, 2023. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 250 + ], + "type": "text", + "content": "[30] Rafael Rafailov, Joey Hejna, Ryan Park, and Chelsea Finn. From r to q: Your language model is secretly a q-function. In First Conference on Language Modeling, 2024. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 251, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 294, + 293 + ], + "type": "text", + "content": "[31] Corby Rosset, Ching-An Cheng, Arindam Mitra, Michael Santacroce, Ahmed Awadallah, and Tengyang Xie. Direct nash optimization: Teaching language models to self-improve with general preferences. ArXiv, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 294, + 294, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 436 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 436 + ], + "type": "text", + "content": "[32] Avi Singh, John D. Co-Reyes, Rishabh Agarwal, Ankesh Anand, Piyush Patil, Peter J. Liu, James Harrison, Jaehoon Lee, Kelvin Xu, Aaron T Parisi, Abhishek Kumar, Alex Alemi, Alex Rizkowsky, Azade Nova, Ben Adlam, Bernd Bohnet, Hanie Sedghi, Igor Mordatch, Isabelle Simpson, Izzeddin Gur, Jasper Snoek, Jeffrey Pennington, Jiri Hron, Kathleen Kenealy, Kevin Swersky, Kshitteej Mahajan, Laura Culp, Lechao Xiao, Maxwell Bileschi, Noah Constant, Roman Novak, Rosanne Liu, Tris Brian Warkentin, Yundi Qian, Ethan Dyer, Behnam Neyshabur, Jascha Narain Sohl-Dickstein, and Noah Fiedel. Beyond human data: Scaling self-training for problem-solving with language models. Trans. Mach. Learn. Res., 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 437, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 470 + ], + "type": "text", + "content": "[33] Feifan Song, Yu Bowen, Minghao Li, Haiyang Yu, Fei Huang, Yongbin Li, and Houfeng Wang. Preference ranking optimization for human alignment. ArXiv, 2023. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 471, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 503 + ], + "type": "text", + "content": "[34] Guohao Sun, Can Qin, Huazhu Fu, Linwei Wang, and Zhiqiang Tao. Stllava-med: Self-training large language and vision assistant for medical. In EMNLP, 2024. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 504, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 536 + ], + "type": "text", + "content": "[35] Guohao Sun, Can Qin, Jiamian Wang, Zeyuan Chen, Ran Xu, and Zhiqiang Tao. Sq-llava: Self-questioning for large vision-language assistant. In ECCV, 2024. 2, 3, 4, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 537, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 591 + ], + "type": "text", + "content": "[36] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 668 + ], + "type": "text", + "content": "[37] Fei Wang, Xingyu Fu, James Y. Huang, Zekun Li, Qin Liu, Xiaogeng Liu, Mingyu Derek Ma, Nan Xu, Wenxuan Zhou, Kai Zhang, Tianyi Yan, Wenjie Jacky Mo, Hsiang-Hui Liu, Pan Lu, Chunyuan Li, Chaowei Xiao, Kai-Wei Chang, Dan Roth, Sheng Zhang, Hoifung Poon, and Muhao Chen. Muir-bench: A comprehensive benchmark for robust multi-image understanding. ArXiv, 2024. 1, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[38] Fei Wang, Wenxuan Zhou, James Y. Huang, Nan Xu, Sheng Zhang, Hoifung Poon, and Muhao Chen. mdpo: Conditional preference optimization for multimodal large language models. In EMNLP, 2024. 1, 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[39] Yuetian Weng, Mingfei Han, Haoyu He, Xiaojun Chang, and Bohan Zhuang. Longvlm: Efficient long video understanding via large language models. ArXiv, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "type": "text", + "content": "[40] Haoning Wu, Dongxu Li, Bei Chen, and Junnan Li. Longvideobench: A benchmark for long-context interleaved video-language understanding. ArXiv, 2024. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "type": "text", + "content": "[41] Junkang Wu, Yuexiang Xie, Zhengyi Yang, Jiancan Wu, Jinyang Gao, Bolin Ding, Xiang Wang, and Xiangnan He. " + }, + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "type": "text", + "content": "-dpo: Direct preference optimization with dynamic " + }, + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "type": "text", + "content": ". ArXiv, 2024. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 184, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 217 + ], + "type": "text", + "content": "[42] Yue Wu, Zhiqing Sun, Huizhuo Yuan, Kaixuan Ji, Yiming Yang, and Quanquan Gu. Self-play preference optimization for language model alignment. ArXiv, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 217, + 553, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 236 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 236 + ], + "type": "text", + "content": "[43] x.ai. Grok-1.5 vision preview. Technical report, x.ai, 2024. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "text", + "content": "[44] Junbin Xiao, Xindi Shang, Angela Yao, and Tat seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "type": "text", + "content": "[45] Tianyi Xiong, Tianyi Xiong, Xiyao Wang, Dong Guo, Qinghao Ye, Haoqi Fan, Quanquan Gu, Heng Huang, and Chunyuan Li. Llava-critic: Learning to evaluate multimodal models. ArXiv, 2024. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 328, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 328, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 328, + 553, + 380 + ], + "type": "text", + "content": "[46] Enwei Xu, Wei Wang, and Qingxia Wang. The effectiveness of collaborative problem solving in promoting students' critical thinking: A meta-analysis based on empirical literature. *Humanities and Social Sciences Communications*, 2023. 2, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 383, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 415 + ], + "type": "text", + "content": "[47] Xubing Ye, Yukang Gan, Xiaoke Huang, Yixiao Ge, Ying Shan, and Yansong Tang. Voco-llama: Towards vision compression with large language models. 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 417, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 417, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 417, + 553, + 514 + ], + "type": "text", + "content": "[48] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Ren-liang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 516, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 516, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 516, + 553, + 568 + ], + "type": "text", + "content": "[49] Kaichen Zhang, Bo Li, Peiyuan Zhang, Fanyi Pu, Joshua Adrian Cahyono, Kairui Hu, Shuai Liu, Yuanhan Zhang, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Lmmseval: Reality check on the evaluation of large multimodal models. ArXiv, 2024. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 571, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 553, + 667 + ], + "type": "text", + "content": "[50] Pan Zhang, Xiao wen Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, Songyang Zhang, Wenwei Zhang, Yining Li, Yang Gao, Peng Sun, Xinyue Zhang, Wei Li, Jingwen Li, Wenhai Wang, Hang Yan, Conghui He, Xingcheng Zhang, Kai Chen, Jifeng Dai, Yu Qiao, Dahua Lin, and Jiaqi Wang. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. ArXiv, 2024. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[51] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. In ArXiv, 2024. 6" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "750" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 284 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[52] Yanzhe Zhang, Ruiyi Zhang, Jiumiang Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tongfei Sun. Llavar: Enhanced visual instruction tuning for text-rich image understanding. In ArXiv, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "text", + "content": "[53] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. ArXiv, 2024. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[54] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. ArXiv, 2024. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "type": "text", + "content": "[55] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. ArXiv, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 294, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 284 + ], + "type": "text", + "content": "[56] Daniel M. Ziegler, Nisan Stiennon, Jeff Wu, Tom B. Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. ArXiv, 2019. 3" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "751" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/StyleKeeper_ Prevent Content Leakage using Negative Visual Query Guidance/45fcd356-f226-4ed7-980b-85432724bf8a_content_list.json b/2025/StyleKeeper_ Prevent Content Leakage using Negative Visual Query Guidance/45fcd356-f226-4ed7-980b-85432724bf8a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..29e7aca0dac21dec83a9fb15196605db1ab97762 --- /dev/null +++ b/2025/StyleKeeper_ Prevent Content Leakage using Negative Visual Query Guidance/45fcd356-f226-4ed7-980b-85432724bf8a_content_list.json @@ -0,0 +1,1611 @@ +[ + { + "type": "text", + "text": "StyleKeeper: Prevent Content Leakage using Negative Visual Query Guidance", + "text_level": 1, + "bbox": [ + 102, + 130, + 893, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jaeseok Jeong\\*† \nYonsei University \njete-jeong@yonsei.ac.kr", + "bbox": [ + 143, + 180, + 348, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junho Kim* \nNAVER AI Lab \njhkim.ai@navercorp.com", + "bbox": [ + 390, + 181, + 586, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Gayoung Lee \nNAVER AI Lab \ngayoung.lee@navercorp.com", + "bbox": [ + 629, + 181, + 852, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yunjey Choi \nNAVER AI Lab", + "text_level": 1, + "bbox": [ + 315, + 241, + 446, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yunjay.choi@gmail.com", + "bbox": [ + 287, + 280, + 475, + 294 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Youngjung Uh‡ Yonsei University yj.uh@yonsei.ac.ki", + "bbox": [ + 542, + 241, + 697, + 294 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 329, + 326, + 344 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the domain of text-to-image generation, diffusion models have emerged as powerful tools. Recently, studies on visual prompting, where images are used as prompts, have enabled more precise control over style and content. However, existing methods often suffer from content leakage, where undesired elements of the visual style prompt are transferred along with the intended style. To address this issue, we 1) extend classifier-free guidance (CFG) to utilize swapping self-attention and propose 2) negative visual query guidance (NVQG) to reduce the transfer of unwanted contents. NVQG employs negative score by intentionally simulating content leakage scenarios that swap queries instead of key and values of self-attention layers from visual style prompts. This simple yet effective method significantly reduces content leakage. Furthermore, we provide careful solutions for using a real image as visual style prompts. Through extensive evaluation across various styles and text prompts, our method demonstrates superiority over existing approaches, reflecting the style of the references, and ensuring that resulting images match the text prompts.", + "bbox": [ + 89, + 359, + 483, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 691, + 220, + 707 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-image diffusion models (T2I DMs) excel at synthesizing images that correspond to given text prompts [37, 38]. However, relying solely on text prompts may not allow precise control over the desired output. Even with highly detailed text prompts, controlling the exact style of the resulting images remains a challenge because of the modality gap between text and image. For example, from \"a red car\", different people imagine different variants of red. Instead, \"A dinosaur\" drawn in the style of the reference", + "bbox": [ + 89, + 715, + 483, + 853 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4484b300f1bbcb13260020d8d5c7810353cf61467e5f891e92bb3fabe0e52705.jpg", + "image_caption": [ + "Figure 1. Qualitative comparison with the same style. Competitors face challenges in generating images with diverse layouts and compositions, i.e., content leakage from the reference." + ], + "image_footnote": [], + "bbox": [ + 516, + 327, + 906, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "image in Figure 1 provides a more concrete specification of the desired style.", + "bbox": [ + 511, + 625, + 906, + 655 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, using reference images as visual style prompts comes with a fundamental challenge: a dilemma between better reflecting style and preventing content leakage. Figure 1 shows that the pose of the reference horse appears in the results of previous methods. The content leakage may include poses, specific shapes, layouts or entire objects. Furthermore, content leakage reduces diversity and text alignments as shown in Figure 1.", + "bbox": [ + 511, + 657, + 906, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Approaches that involve training the model, such as fine-tuning a diffusion model on a set of themed images [21, 39], learning new text embeddings [2, 10, 13], or adapting cross-attention modules to incorporate image features [35, 52, 55, 59], come with significant drawbacks. These methods not only incur high computational costs, but also face an inherent trade-off between style and content. Achieving a balance where the model faithfully captures the", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 107, + 862, + 217, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work was done during his internship at Naver AI Lab.", + "bbox": [ + 109, + 875, + 436, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\ddagger$ Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15760", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "desired style without unintentionally preserving content remains a challenge.", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this, recent papers [1, 6, 16] have proposed training-free methods by modifying feature swapping mechanisms within the self-attention layers. Specifically, they replace the key and value features in the original process with those extracted from the visual style prompt. This allows the query to preserve content while the key-value pair carries the style, enabling promising style transfer performance.", + "bbox": [ + 89, + 122, + 482, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, these methods struggle to fully eliminate content leakage [42]. Additionally, Hertz et al. [16] does not support real images as reference inputs, while Chung et al. [6] and Alaluf et al. [1] primarily focus on image-to-image (I2I) tasks. When using real images as style references, they often suffer from either poor style transfer or excessive content leakage, significantly limiting their effectiveness.", + "bbox": [ + 89, + 243, + 483, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this study, we propose a method to separately control the strength of style and content from visual prompts such that the results align with the text prompts. We design a novel variation of classifier-free guidance (CFG) [17], emulating images with undesirable content leakage as a negative guidance. To address content leakage, we introduce negative visual query guidance, ensuring a clear separation between content and style. We also show that stochastic encoding achieves better style reflection with less artifacts and computation, and incorporate color calibration to precisely match the final output to the reference image's color statistics leading to better style reflection.", + "bbox": [ + 89, + 349, + 483, + 530 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We analyze where to apply swapping self-attention, identifying the optimal layers for balancing style transfer and content fidelity. Additionally, our method can effectively remove content that is difficult to eliminate with key and value swapping alone, working successfully even in cases with significant structural gaps between the style image and content text, as shown in Figure 1 (e.g., a complex scene of \"a horse with the vibrant colors of a Chinese ink painting\" and a single object \"Dinosaur\"). Our method is well compatible with ControlNet for I2I style transfer, further enhancing its flexibility. Qualitative and quantitative evaluations show our method outperforms state-of-the-art approaches, providing precise control over content and style without content leakage. Our approach is both robust and efficient, ideal for visual style prompting tasks. $^{1}$", + "bbox": [ + 89, + 532, + 483, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Visual style prompting", + "text_level": 1, + "bbox": [ + 89, + 773, + 305, + 791 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose StyleKeeper which receives a text prompt and a visual style prompt to generate new images without content leakage. The results contain the content and style specified by the text prompt and the visual style prompt, respectively, with variations due to initial noises. The overview", + "bbox": [ + 89, + 799, + 482, + 875 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of our method is illustrated in Figure 2. First, we explain the swapping self-attention in the aspect of style transfer literature. StyleKeeper consists of classifier-free guidance (CFG) with swapping self-attention, negative visual query guidance (NVQG), optimal layer choice, stochastic encoding of real visual style prompts, and generalization to ControlNet for real content images. We explain the first three components in text-to-image (T2I) scenario with generated visual style prompt. Then we proceed to T2I with real visual style prompt and image-to-image (I2I) scenario.", + "bbox": [ + 511, + 90, + 906, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Swapping self-attention in style transfer literature", + "text_level": 1, + "bbox": [ + 511, + 252, + 906, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Modern diffusion models consist of a number of self-attention and cross-attention blocks [47]. Both of them employ the attention mechanism, which first obtains an attention map using similarity between query features $Q$ and key features $K$ , then aggregates value features $V$ using the attention map as weights: $Attention(Q, K, V) = \\text{Softmax}(\\frac{QK^T}{\\sqrt{d}})V$ . Opposed to the cross-attention layer, self-attention layer receives key and values coming from the main denoising process, which has spatial dimensions with more freedom to represent spatially varying visual elements. As our goal is to reflect style elements from a reference image that are not easily represented by textual description, we opt to borrow key and values of self-attention layers in the reference process to the original process, namely swapping self-attention (Figure 2). In addition, swapping self-attention has a strong connection with style transfer literature [8, 25, 32, 41, 58]. where the attention mechanism reassembles visual features of a style image (key, value) on a content image (query). However, methods [3, 16] utilizing self-attention operation often exhibit artifacts and content leakage since style elements are entangled with content elements in the visual feature.", + "bbox": [ + 511, + 289, + 906, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. CFG sampling with swapping self-attention for T2I", + "text_level": 1, + "bbox": [ + 511, + 636, + 906, + 665 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous works [1, 6] have not taken CFG into account while manipulating features. We propose CFG combined with swapping self-attention to reflect a visual style prompt in the T2I results without artifacts. CFG [17] is a key technique that steers generated images toward specified text prompts. Given a score $\\epsilon_{\\theta}(x_t,c)$ conditioned on $c$ and an unconditional score $\\epsilon_{\\theta}(x_t,\\emptyset)$ , the CFG score is defined as:", + "bbox": [ + 511, + 672, + 908, + 779 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\epsilon} _ {\\theta} = (1 + w) \\epsilon_ {\\theta} \\left(x _ {t}, c\\right) - w \\epsilon_ {\\theta} \\left(x _ {t}, \\emptyset\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 791, + 906, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $w$ controls guidance strength. CFG with $w > 1$ enhances image quality and text alignment, but excessive $w$ may lead to mode collapse [5]. Notably, CFG has not been", + "bbox": [ + 511, + 818, + 906, + 866 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "2We omit the diffusion timestep $t$ in the arguments and abuse $x_{t}$ instead of $z_{t}$ from latent diffusion model.", + "bbox": [ + 511, + 875, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1Our code will be released for reproducibility.", + "bbox": [ + 107, + 886, + 354, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15761", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0fadae4ffc93a3f1c143b083f110b1aff2a3b49c913dc2a0c7a46e3ba07ded46.jpg", + "image_caption": [ + "Figure 2. Overveiw of StyleKeeper. Our proposed method includes 4 key components, highlighted in red boxes. First, stochastic encoding (Section 3) converts reference images into suitable latents for the visual style prompting task. Second, swapping self-attention (Section 2.2, 2.4) ensures the reference image's style is accurately reflected. Third, negative visual query guidance (Section 2.3) reduces content leakage from the reference image, allowing the desired text content (e.g., \"Moose\") to be better represented. Lastly, color calibration (Section 3) minimizes errors during the denoising process, helping to produce a cleaner final image." + ], + "image_footnote": [], + "bbox": [ + 137, + 90, + 861, + 344 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "explored in the context of reflecting denoising process with modified features.", + "bbox": [ + 89, + 448, + 482, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Assuming a given condition $c$ has hidden content $h^{\\mathrm{content}}$ and style $h^{\\mathrm{style}}$ , we model $p_{\\theta}(x_0|h_{\\mathrm{text}}^{\\mathrm{content}},h_{\\mathrm{visual}}^{\\mathrm{style}})$ using:", + "bbox": [ + 89, + 478, + 482, + 510 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The original T2I score: $\\epsilon_{\\theta}(x_t, c_{\\text{text}})$ , leading to $x_0^{\\text{ori}} \\sim p_{\\theta}(x_0 | c_{\\text{text}}) = p_{\\theta}(x_0 | h_{\\text{text}}^{\\text{content}}, h_{\\text{text}}^{\\text{style}})$ .", + "- The reference score: $\\epsilon_{\\theta}(x_t, c_{\\mathrm{visual}})$ , leading to $x_0^{\\mathrm{visual}} \\sim p_{\\theta}(x_0 | c_{\\mathrm{visual}}) = p_{\\theta}(x_0 | h_{\\mathrm{visual}}^{\\mathrm{content}}, h_{\\mathrm{visual}}^{\\mathrm{style}})$ ." + ], + "bbox": [ + 89, + 512, + 482, + 578 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To ensure proper content and style control, we design the CFG score as:", + "bbox": [ + 89, + 579, + 482, + 608 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde {\\epsilon} _ {\\theta} \\left(x _ {t}, h _ {\\text {t e x t}} ^ {\\text {c o n t e n t}}, h _ {\\text {v i s u a l}} ^ {\\text {s t y l e}}\\right) = (1 + w) \\\\ \\ddot {\\epsilon} _ {\\theta} \\left(x _ {t}, Q _ {\\text {t e x t}}, K V _ {\\text {v i s u a l}}\\right) \\tag {2} \\\\ - w \\epsilon_ {\\theta} (x _ {t}, \\emptyset). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 618, + 482, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\ddot{\\epsilon}_{\\theta}(x_{t},Q_{\\mathrm{text}},KV_{\\mathrm{visual}})$ represents a KV-injected denoising score, indicating feature manipulation rather than direct conditioning.", + "bbox": [ + 89, + 686, + 482, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For given two denoising processes, one as original and another as a reference, borrowing the key-value in self-attention from the reference to the original process, i.e., key-value (KV) injection, tends to produces results with the content from the original process and the style from the reference process with limited control [1, 6, 56].", + "bbox": [ + 89, + 732, + 482, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We define the KV-injected score by $\\ddot{\\epsilon}_{\\theta}(x_{t},Q_{\\mathrm{text}},KV_{\\mathrm{visual}})$ where $Q_{\\mathrm{text}}$ and $KV_{\\mathrm{visual}}$ denote the query from the original score $\\epsilon_{\\theta}(x_t,c_{\\mathrm{text}})$ and the", + "bbox": [ + 89, + 821, + 482, + 866 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "key, value from the reference score $\\epsilon_{\\theta}(x_t,c_{\\mathrm{visual}})$ . Then KV-injected-Attention becomes:", + "bbox": [ + 511, + 448, + 906, + 477 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t e n t i o n} \\left(Q _ {\\text {t e x t}}, K _ {\\text {v i s u a l}}, V _ {\\text {v i s u a l}}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {Q _ {\\text {t e x t}} K _ {\\text {v i s u a l}} ^ {\\intercal}}{\\sqrt {d}}\\right) V _ {\\text {v i s u a l}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 492, + 905, + 534 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We omit the layer index for brevity.", + "bbox": [ + 513, + 535, + 751, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Naive denoising process with $\\ddot{e}_{\\theta}(x_{t},Q_{\\mathrm{text}},KV_{\\mathrm{visual}})$ provides limited control in generating images with content and style specified by a text $c_{\\mathrm{text}}$ and a visual style prompt, respectively. Moving forward, our CFG with swapping self-attention in Eq. (2) enjoys higher image quality and more accurate text alignment than the naive denoising process as in the original CFG for T2I generation. The results are deferred to Section 4.1. However, content leakage still hinders the creation of diverse images, and disrupts text alignment.", + "bbox": [ + 511, + 551, + 906, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Negative visual query guidance", + "text_level": 1, + "bbox": [ + 513, + 719, + 790, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose negative visual query guidance (NVQG) to mitigate content leakage by reducing the content $h_{\\mathrm{visual}}^{\\mathrm{content}}$ from the visual style prompt appearing in the results. Briefly, NVQG negates the CFG of a score $\\ddot{\\epsilon}(x_t, Q_{\\mathrm{visual}}, KV_{\\mathrm{text}})$ .", + "bbox": [ + 511, + 743, + 905, + 804 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Liu et al. [24], a complex text condition $c$ is factorized as a set of conditions $\\{c_0, c_1, \\ldots\\}$ and Bayes' rule induces $p_{\\theta}(x_t | c_0, c_1, \\ldots) \\propto \\Pi \\frac{p_{\\theta}(x_t | c_i)}{p_{\\theta}(x_t)}$ . Then, the score of the complex text condition $c$ becomes $\\epsilon_{\\theta}(x_t, c) = \\epsilon_{\\theta}(x_t, \\emptyset) + \\Pi (\\epsilon_{\\theta}(x_t, c_i) - \\epsilon_{\\theta}(x_t, \\emptyset))$ . It allows reducing a specific concept $\\tilde{c}$ with composition by negating its guidance with scale", + "bbox": [ + 511, + 805, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "3The original score and its query within are recursively altered by KV injection along the denoising process.", + "bbox": [ + 89, + 875, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15762", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8ea22fbcecac27f5c3edf52615eb626e68a3a16748453ac49dfcae68d79ac58d.jpg", + "image_caption": [ + "Figure 3. The effect of swapping self-attention across different blocks. Swapping self-attention on the bottleneck and downblocks causes content leakage, resulting in cat-like images despite a dog prompt, while swapping on downblocks disrupts resulting images. We only apply swapping self-attention in the upblocks to reflect style elements effectively." + ], + "image_footnote": [], + "bbox": [ + 99, + 89, + 480, + 261 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$w_{\\mathrm{neg}}$", + "bbox": [ + 89, + 382, + 130, + 395 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {\\theta} \\left(x _ {t}, c, \\text {n o t} \\tilde {c}\\right) = \\epsilon_ {\\theta} \\left(x _ {t}, c\\right) - w _ {\\text {n e g}} \\left(\\epsilon_ {\\theta} \\left(x _ {t}, \\tilde {c}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, \\emptyset\\right)\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 402, + 482, + 433 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although we design $\\ddot{\\epsilon}_{\\theta}(x_{t},Q_{\\mathrm{text}},KV_{\\mathrm{visual}})$ to predict the score toward $p_{\\theta}(x_0|h_{\\mathrm{text}}^{\\mathrm{content}},h_{\\mathrm{visual}}^{\\mathrm{style}})$ , $\\ddot{\\epsilon} (x_{t},Q_{\\mathrm{text}},KV_{\\mathrm{visual}})$ still contain $h_\\mathrm{visual}$ . Assuming a hidden factorization $KV_{\\mathrm{visual}} = \\{KV_{\\mathrm{visual}}^{\\mathrm{content}},KV_{\\mathrm{visual}}^{\\mathrm{style}}\\}$ , Bayes' rule induces", + "bbox": [ + 89, + 433, + 483, + 497 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} p _ {\\theta} \\left(x _ {t} \\mid Q _ {\\text {t e x t}}, K V _ {\\text {v i s u a l}} ^ {\\text {s t y l e}}, K V _ {\\text {v i s u a l}} ^ {\\text {c o n t e n t}}\\right) \\propto p _ {\\theta} \\left(x _ {t}\\right) \\frac {p _ {\\theta} \\left(x _ {t} \\mid Q _ {\\text {t e x t}} , K V _ {\\text {v i s u a l}} ^ {\\text {s t y l e}}\\right)}{p _ {\\theta} \\left(x _ {t}\\right)} \\\\ \\times \\frac {p _ {\\theta} \\left(x _ {t} \\mid Q _ {\\emptyset} , K V _ {\\text {v i s u a l}} ^ {\\text {c o n t e n t}}\\right)}{p _ {\\theta} \\left(x _ {t}\\right)}. \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 503, + 480, + 577 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since $\\epsilon_{\\theta}(x_t) = \\epsilon_{\\theta}(x_t, Q_{\\emptyset}, KV_{\\emptyset})$ , we obtain the desired conditional score of $\\hat{c} = \\{Q_{\\mathrm{text}}, KV_{\\mathrm{visual}}^{\\mathrm{style}}\\}$ :", + "bbox": [ + 89, + 579, + 482, + 612 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\epsilon_ {\\theta} \\left(x _ {t}, \\hat {c}\\right) \\leftarrow w _ {\\text {v i s u a l}} \\left(\\ddot {\\epsilon} _ {\\theta} \\left(x _ {t}, Q _ {\\text {t e x t}}, K V _ {\\text {v i s u a l}}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}\\right)\\right) \\\\ - w _ {\\text {c o n t e n t}} \\left(\\ddot {\\epsilon} _ {\\theta} \\left(x _ {t}, Q _ {\\emptyset}, K V _ {\\text {v i s u a l}} ^ {\\text {c o n t e n t}}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}\\right)\\right) \\\\ + \\epsilon_ {\\theta} \\left(x _ {t}\\right). \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 619, + 480, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w_{\\mathrm{visual}}$ and $w_{\\mathrm{content}}$ regulate the strength of each guidance. By leveraging query injection [1, 6, 46, 56], we approximate $\\ddot{\\epsilon}_{\\theta}(x_t, Q_{\\emptyset}, KV_{\\mathrm{visual}}^{\\mathrm{content}}) \\approx \\ddot{\\epsilon}_{\\theta}(x_t, Q_{\\mathrm{visual}}, KV_{\\emptyset})$ . Finally, inserting (6) into (2), and empirically replacing $\\epsilon_{\\theta}(x_t)$ with $\\ddot{\\epsilon}\\theta(x_t, Q_{\\mathrm{visual}})$ results in similar performance. This allows a simplified reformulation of diffusion sampling using $w' = w_{\\mathrm{visual}}(w + 1)$ : Since we can highly relate the $w' \\ddot{\\epsilon}_{\\theta}(x_t, Q_{\\mathrm{visual}})$ to the concept negation in (4) which guides the negation of concept with a scale $w_{\\mathrm{neg}}$ , we named the query term as negative visual query guidance.", + "bbox": [ + 89, + 690, + 482, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.4. Choosing blocks for swapping self-attention", + "text_level": 1, + "bbox": [ + 89, + 848, + 460, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to maximize the effect of NVQG in preventing content leakage, we explore the depth of the self-attention", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f3312850a2d60f836f692e9f0785c66f700290c48ffd93a6f5b42094dc45ad80.jpg", + "image_caption": [ + "Figure 4. The effect of CFG and the proposed negative visual query guidance on image generation. The reference images provide the style for each generated output. Without NVQG, content leakage occurs, and the generated images fail to fully capture the intended content. In contrast, using NVQG ensures better alignment with both the reference style and the \"Cat\" prompt, reducing content distortion and improving quality." + ], + "image_footnote": [], + "bbox": [ + 544, + 90, + 885, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "blocks to be swapped in the sense of granularity of visual elements. Modern architecture of diffusion models roughly consists of three sections in a sequence: downblocks, bottleneck blocks, and upblocks. Given that the bottleneck of diffusion models contains content elements of the image [20, 22, 33], we opt not to apply swapping self-attention to bottleneck blocks to prevent transferring contents in a reference image. Figure 3 shows that not swapping the bottleneck blocks prevents content leaking from the reference image. However, the synthesized images show disrupted results with seriously scattered objects. Furthermore, while swapping self-attention implements the reassembling operation, simply applying to all self-attention layers exposes a content leakage problem, where the content of the reference image influences the resulting image, as shown in the first row of Figure 3. I.e., the results contain cats even though the prompts specify \"a dog\". We conjecture that this phenomenon happens because feature maps of downblocks have unclear content layout [3, 27], so substituting features based on this inaccurate layout causes the disrupted results. To avoid injecting unnecessary features, we choose to swap the key and value of self-attention only in upblocks.", + "bbox": [ + 511, + 424, + 906, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We note that Hertz et al. [16] applies self-attention operation at all blocks and suffers content leakage. The last row of Figure 3 shows the success of our strategy. Following this, we proceed to search for a specific optimal layer in the upblock in Section 4.2", + "bbox": [ + 511, + 757, + 905, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Real images as references", + "text_level": 1, + "bbox": [ + 511, + 844, + 751, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "So far, we have assumed a generated visual style prompt $x_0^{\\mathrm{visual}} \\sim p_\\theta(x_0 | c_{\\mathrm{visual}}}$ . Here, we allow real visual style", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15763", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/742e55a6753a13dceac66efd180e4fa851ff786a6abde0262353f7cdf9f6b28f.jpg", + "image_caption": [ + "Figure 5. Analysis on the optimal range of upblocks for swapping self-attention. We find the optimal range of upblocks for a balanced trade-off between different aspects. The images on the right illustrate the visual results for different upblock layer indices, with the red cross indicating poor diversity and misalignment to the text prompt, the red triangle indicating a lack of style similarity, and the yellow star indicating the optimal results. Please refer to Section 4.2 for details." + ], + "image_footnote": [], + "bbox": [ + 120, + 88, + 468, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "prompts by 1) stochastic encoding and 2) color calibration.", + "bbox": [ + 89, + 662, + 480, + 678 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose stochastic encoding to obtain $x_{t}^{\\mathrm{visual}} \\sim q(x_{t}|x_{0}^{\\mathrm{visual}})$ by adding a random noise on $x_{0}^{\\mathrm{visual}}$ following the forward process of DMs [18]:", + "bbox": [ + 89, + 678, + 482, + 724 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {t} \\sim \\mathcal {N} (0, I), x _ {t} ^ {\\text {v i s u a l}} = \\sqrt {\\alpha_ {t}} \\cdot x _ {0} ^ {\\text {v i s u a l}} + \\sqrt {1 - \\alpha_ {t}} \\cdot \\epsilon_ {t} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 734, + 482, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At each timestep, we sample $\\epsilon_{t}$ to encode $x_{t}^{\\mathrm{visual}}$ . It ensures that $x_{t}^{\\mathrm{visual}}$ lies on the learned trajectory and does not suffer from accumulative numerical error due to iterative process of DDIM inversion used by previous methods [6, 16] as shown in Figure A1. Furthermore, it does not need to store the intermediate latents as opposed to DDPM inversion used by Alaluf et al. [1].", + "bbox": [ + 89, + 763, + 482, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although stochastic encoding performs better than DDIM inversion, a subtle color discrepancy occurs between", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the resulting images and the visual style prompts. We introduce color calibration at $x_{t}^{\\mathrm{ori}}$ in the original process to match the statistics of predicted $x_{0}^{\\mathrm{ori}}$ to predicted $x_{0}^{\\mathrm{visual}}$ . In Gatys [12], distance between channel-wise statistics is employed as a style loss for style transfer. In Song et al. [43], predicted $x_{0} \\left( = \\frac{x_{t}^{\\mathrm{visual}} - \\sqrt{1 - \\alpha_{t}} \\cdot \\epsilon_{\\theta}(\\hat{x}_{t})}{\\sqrt{\\alpha_{t}}} \\right)$ allows to estimate $x_{0}$ with high probability at intermediate timesteps using deterministic sampling. Inheriting the advantage, we execute adain operation to match mean&std of predicted $x_{0}^{\\mathrm{ori}}$ with those of $x_{0}^{\\mathrm{visual}}$ . It allows precise color calibration rather than directly matching $x_{t}^{\\mathrm{ori}}$ to $x_{t}^{\\mathrm{visual}}$ in Alaluf et al. [1], Chung et al. [6]. Furthermore, our approach differs from Chung et al. [6] in that using predicted $x_{0}$ at intermediate timestep $t \\in (0,T)$ other than $x_{T}$ by reducing cumulative sampling error after the operation. Furthermore, Chung et al. [6] executes AdaIN at timestep T, inducing lengthy cumulative error.", + "bbox": [ + 511, + 90, + 903, + 337 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We provide supportive experiments that show the effectiveness of the proposed method in Section 4.4 and a detailed Algorithm in Appendix 3.", + "bbox": [ + 511, + 338, + 903, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 400, + 645, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we describe the effects of our proposed methods: CFG with swapping self-attention, Negative visual query guidance (NVQG), stochastic encoding, and color calibration. For swapping self-attention, we provide a detailed analysis through experiments to determine the optimal layers for swapping. The impact of NVQG is demonstrated through qualitative results. Additionally, we show why stochastic encoding outperforms DDIM inversion when inverting real images, and we highlight the benefits of color calibration through experimental results.", + "bbox": [ + 511, + 426, + 903, + 577 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also conducted both quantitative and qualitative comparisons of our method against various competitors, including StyleAligned [16], IP-Adapter [59], Dreambooth-LoRA [39, 40], StyleDrop [42], DEADiff [35], CSGO [55] and InstantStyle [50, 51]. The details of these comparisons, along with the experimental setup and metrics, are described in the Appendix 2.", + "bbox": [ + 511, + 579, + 903, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Effectiveness of CFG and NVQG", + "text_level": 1, + "bbox": [ + 511, + 696, + 802, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section analyzes the effects of Classifier-Free Guidance (CFG) and Negative Visual Query Guidance (NVQG) on image generation, with a focus on text alignment and content leakage. Figure 4 shows the results of the three configurations. In the $1^{\\text{st}}$ row, without CFG and NVQG, the generated images suffer from severe artifacts. The absence of CFG causes poor image quality resulting in significant misalignment with the prompt. In the $2^{\\text{nd}}$ row, CFG with swapping self-attention improves the text misalignment by boosting image quality. Here, the \"cat\" in target text prompt becomes clearer in the generated images. However, content leakage from the reference image remains where unwanted", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15764", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7db0b67b31668fdea91bc83886ab0a678a0195591dcc4347a54a5a0cb310cc17.jpg", + "image_caption": [ + "Figure 6. Attention map visualization over late and early upblock layers. The late upblock better focuses on the style-corresponding region than the early upblock, leading to more freedom to reassemble small parts. The early upblock attends a larger region leading to content leakage." + ], + "image_footnote": [], + "bbox": [ + 81, + 23, + 480, + 243 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "elements (layouts, structure, and composition) from the reference image affect the results.", + "bbox": [ + 89, + 338, + 482, + 367 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the $3^{\\mathrm{rd}}$ row, NVQG releases the content leakage and produces the best results closely matching the text prompt while reflecting style reference.", + "bbox": [ + 89, + 368, + 482, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overall, Figure 4 demonstrates the critical role of NVQG in reducing content leakage, enjoying the quality boosting of CFG. Together, they ensure that the generated images align to both the target text prompt and the visual style prompts, resulting in high-quality, coherent outputs.", + "bbox": [ + 89, + 412, + 482, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We qualitatively showcase diversity of results within a text prompt in Figure A9 and text alignment with complex text prompts in Figure A22.", + "bbox": [ + 89, + 489, + 483, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Analysis for Swapping Self-Attention", + "text_level": 1, + "bbox": [ + 89, + 542, + 411, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Optimal layers in upblocks Since recent large T2I DMs consist of many layers, we further analyze the behavior by changing the start of the swapping while the end of the swapping is fixed at the end. We use four key metrics as shown in Figure 5, there is a layer where all four metrics abruptly change (red line). Notably, this point remains consistent regardless of the reference image. All the results in the paper use only one fixed optimal layer.", + "bbox": [ + 89, + 565, + 482, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We choose this layer as the optimal start of the swapping for a balanced trade-off among all aspects. We provide qualitative results with detailed split of layers in Figure A3 and A4.", + "bbox": [ + 89, + 686, + 482, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visualizing Attention maps Figure 6 compares average attention maps from the late upblock and the early upblock applying swapping self-attention. Using late upblock has more freedom to reassemble the reference style elements leading to more doggy results than early upblock which produces some cat-like attributes. The right two columns visualize the attention weight of query points marked as red stars and yellow dots. Swapping self-attention on late upblock reassemblies features from a style correspondence,", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a02de94a8bac9e60c8baf74bb9bf83e37f85cf6f6611f760605ce6d222548fd3.jpg", + "image_caption": [ + "(a) Synthesizing the content specified by the text prompts" + ], + "image_footnote": [], + "bbox": [ + 535, + 87, + 877, + 170 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/294659934ab125c51d9e03757c7be199ff88ee10e90372955153aa97745e125c.jpg", + "image_caption": [ + "(b) Synthesizing the pose specified by the text prompt" + ], + "image_footnote": [], + "bbox": [ + 535, + 186, + 874, + 266 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1345f86ee2b06b053f2c2d4e887274eb2a4922b71a756c564fb4d704dd3df9d7.jpg", + "image_caption": [ + "(c) Synthesizing a single object specified by the text prompt", + "Figure 7. Comparison for content leakage. While StyleAligned suffers from content leakage from the reference, our results clearly align with the text prompts" + ], + "image_footnote": [], + "bbox": [ + 535, + 282, + 874, + 363 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "e.g., texture and color. On the other hand, swapping self-attention on early upblock reassembles features from a wider area with different styles. This comparison clarifies the reasons for using only late upblock. Please refer to Figure A6, A7 for a more visualization and a detailed analysis.", + "bbox": [ + 511, + 460, + 906, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison against competitors", + "text_level": 1, + "bbox": [ + 511, + 564, + 799, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Style & content control We provide a qualitative comparison in Figure 8, focusing on controlling style and content. Our method faithfully synthesizes content from the text prompt with the style of the reference image. In contrast, other methods add elements like color or texture not in the reference (e.g., feathers, bricks, iron, skin) and often suffer from content leakage (e.g., layout, screaming person, castle), which compromises text prompt faithfulness. Quantitative results in Figure 9 support these findings: IP-Adapter shows higher style similarity but neglects text prompts significantly.", + "bbox": [ + 511, + 588, + 906, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Diversity within a text specification Starting from different initial noises, the diffusion models trained on a large dataset produce diverse results within the specification of a text prompt. Figure 1, A19 shows that ours can synthesize various poses and viewpoints while others barely change, i.e., other methods limit the diversity of the pretrained model due to content leakage. We further support our diversity in Figure A9.", + "bbox": [ + 511, + 779, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15765", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/16642bdb6d4b6d938e40598a25670587e9b25d0d5d8af6b3303ae87ea9408cc0.jpg", + "image_caption": [ + "Figure 8. Qualitative comparison across various styles and text prompts. StyleKeeper faithfully reflects style elements in reference images without causing content leakage from the reference images." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 500, + 333 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ec3c6df998fac707c65c563de5a5b534b103beac9e4305e1a01f54a42aea5858.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 88, + 903, + 330 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d6b0e853896fd98ea49de2ea556527d55f664585c2779fd4e027add15c533bf2.jpg", + "image_caption": [ + "Figure 9. Quantitative comparison. Ours achieves the best style similarity while not compromising text alignment and being free from content leakage." + ], + "image_footnote": [], + "bbox": [ + 91, + 386, + 486, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Content leakage Content leakage refers to the phenomenon where the content of a reference image appears in a result. As evident in Figure A12, our model exhibits significantly less content leakage when compared to other models. We focus on a comparison against a runner-up method, StyleAligned, particularly in terms of how content leakage can be an obstacle to controlling the content using a text prompt. Figure 7a compares examples where strong content leakage prevents the object from the text from appearing in StyleAligned. We often observe famous paintings in the reference easily leak into the results of StyleAligned while ours does not struggle. Figure 7b compares examples where the pose in the reference leaks into the results of StyleAligned preventing the reflection of specified pose in the text prompt. On the other hand, our method reflects the correct poses. Figure 7c compares examples where the number of small instances in the reference leaks into the result of StyleAligned. Contrarily, our method correctly synthesizes a single instance of the content specified by the text. In Figure 9, we provide quantitative comparison results. It supports that StyleKeeper achieves the best style", + "bbox": [ + 88, + 583, + 486, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "similarity while not suffering content leakage. We can also observe that models with high content leakage exhibit poor text alignment.", + "bbox": [ + 511, + 397, + 908, + 444 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Preserving the content of the original denoising process Figure A13 shows the results of ours and other methods using the same initial noise in each column. Our method precisely reflects the style in the reference with minimal changes of contents in the original denoising process. On the other hand, the other methods severely alter pose, shape, or layouts due to content leakage. It is an important virtue of controlling style to keep the rest intact. This also supports that ours is free from content leakage during the denoising process. We provide more results in Figure A14.", + "bbox": [ + 511, + 479, + 908, + 632 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. DDIM inversion vs. Our strategy", + "text_level": 1, + "bbox": [ + 511, + 648, + 805, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 11 shows that StyleKeeper can take real images as style reference with our strategies. As shown in Figure 11, stochastic encoding outperforms DDIM inversion and color calibration improves text alignment and style similarity. We provide more results in Figure A15. Moreover, we show that our strategy can boost the performance of the other self-attention variants [3, 16] in Figure A16, A17 and color calibration can be used for generation settings in Figure A18. We provide an ablation study in Figure 12 for each configuration (swapping self-attention, NVQG, and color calibration) in both real reference and generated reference settings. In both settings, swapping self-attention and NVQG improve style similarity and text alignment, while color calibration helps improve style similarity. Additionally, stochastic encoding outperforms DDIM inversion.", + "bbox": [ + 509, + 674, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15766", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3bbc89c10d184627277529b4f7860917d936649cb65c4c1d43b37cd9d94d73e3.jpg", + "image_caption": [ + "Figure 10. Qualitative comparison in I2I style transfer task. We compare our method for I2I style transfer task where the content image is given. Compared to the previous methods, ours achieves the best style similarity." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 903, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/225153d76822477e8f057c560c9e7bcce824ff3119ab63fcd62ba3b42f4b1080.jpg", + "image_caption": [ + "Figure 11. Comparison of DDIM inversion vs. stochastic encoding and the effect of color calibration. Stochastic encoding reduces artifacts in the resulting images, while color calibration better reflects the colors of the reference image." + ], + "image_footnote": [], + "bbox": [ + 120, + 285, + 455, + 458 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a43cf8f9591b5a70d31b8a6bf4a3076f541ce48d11840ad333c486cecae3bc7e.jpg", + "image_caption": [ + "Figure 12. Ablation study of four proposed methods with a real reference and a generated reference setting.. We evaluated performance by individually removing each of the four proposed methods. In all cases, both generated and real images showed similar improvements in style similarity and text alignment. Moreover, stochastic encoding outperformed DDIM inversion in all aspects, comparing purple and blue points." + ], + "image_footnote": [], + "bbox": [ + 96, + 536, + 480, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Comparison for I2I style transfer", + "text_level": 1, + "bbox": [ + 89, + 768, + 382, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Figure 10, we present a qualitative comparison between our method using ControlNet and existing state-of-the-art methods, CrossAttn [1] and StyleID [6], for the I2I style transfer task where a content image is provided. Both CrossAttn and StyleID inject the key & value obtained by inversion from the style images.", + "bbox": [ + 89, + 792, + 483, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "CrossAttn often fails to transfer the detailed representa", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tion of the reference style, leading to a rough, blocky appearance. For instance, the center examples in Figure 10 show that the style details are absent or inaccurate compared to our results. Similarly, StyleID suffers poor style reflection and color deficiency in the center and the rightmost example. In contrast, ours effectively reflects the details of the reference style image, with no noticeable transfer of the color values from the content images.", + "bbox": [ + 511, + 292, + 906, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Limitation", + "text_level": 1, + "bbox": [ + 511, + 431, + 764, + 448 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce StyleKeeper which faithfully applies the style of real reference images without content leakage in a training-free manner. CFG with swapping self-attention captures the reference image's style accurately and allows for direct content generation from the text. This integration with CFG enhances performance by balancing content generation and style transfer. To address content leakage, we propose Negative visual query guidance that ensures the reference image's content does not interfere with the text-specified content, enhancing diversity and improving text alignment & style similarity. Additionally, Stochastic encoding efficiently maps real images to suitable latents, removing artifacts. Color calibration aligns the final output to the reference's color statistics.", + "bbox": [ + 511, + 459, + 906, + 671 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method shows both qualitative and quantitative improvements over existing approaches, providing a robust solution for visual style prompting without complex training. We also analyze where to apply swapping self-attention, identifying optimal layers to balance style transfer and content fidelity. StyleKeeper outperforms prior methods and works well with algorithms like ControlNet [60] and DBLoRA [40], as shown in Figure A20.", + "bbox": [ + 511, + 672, + 906, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "However, StyleKeeper is limited by the capabilities of the pretrained diffusion models and cannot generate elements outside their scope (e.g., \"stone golem\" in Figure A24a). Also, if visual and textual styles conflict, the visual style dominates (Figure A24b). Future work could explore extending our method to other domains such as video, expanding its applicability and research potential.", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15767", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements We thank Jin-Hwa Kim for helpful comments. This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS-2024-00439762, Developing Techniques for Analyzing and Assessing Vulnerabilities, and Tools for Confidentiality Evaluation in Generative AI Models).", + "bbox": [ + 89, + 90, + 485, + 196 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 210, + 187, + 224 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer. In ACM SIGGRAPH 2024 Conference Papers, pages 1–12, 2024. 2, 3, 4, 5, 8, 1", + "[2] Omri Avrahami, Kfir Aberman, Ohad Fried, Daniel Cohen-Or, and Dani Lischinski. Break-a-scene: Extracting multiple concepts from a single image. In SIGGRAPH Asia, 2023. 1", + "[3] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In ICCV), 2023. 2, 4, 7, 1, 9, 15", + "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 1", + "[5] Hyungjin Chung, Jeongsol Kim, Geon Yeong Park, Hyelin Nam, and Jong Chul Ye. Cfg++: Manifold-constrained classifier free guidance for diffusion models. arXiv preprint arXiv:2406.08070, 2024. 2", + "[6] Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8795-8805, 2024. 2, 3, 4, 5, 8, 1", + "[7] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance, 2022. 1", + "[8] Yingying Deng, Fan Tang, Weiming Dong, Chongyang Ma, Xingjia Pan, Lei Wang, and Changsheng Xu. Stytr2: Image style transfer with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11326-11336, 2022. 2", + "[9] Martin Nicolas Everaert, Marco Bocchio, Sami Arpa, Sabine Süssstrunk, and Radhakrishna Achanta. Diffusion in style. In ICCV, 2023. 1", + "[10] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2022. 1", + "[11] Daniel Garibi, Or Patashnik, Andrey Voynov, Hadar Averbuch-Elor, and Daniel Cohen-Or. Renoise: Real image inversion through iterative noising, 2024. 1", + "[12] Leon A Gatys. A neural algorithm of artistic style. arXiv preprint arXiv:1508.06576, 2015. 5, 1", + "[13] Inhwa Han, Serin Yang, Taesung Kwon, and Jong Chul Ye. Highly personalized text embedding for image manipulation by stable diffusion. arXiv, 2023. 1" + ], + "bbox": [ + 93, + 234, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Ligong Han, Song Wen, Qi Chen, Zhixing Zhang, Kunpeng Song, Mengwei Ren, Ruijiang Gao, Anastasis Stathopoulos, Xiaoxiao He, Yuxiao Chen, Di Liu, Qilong Zhangli, Jindong Jiang, Zhaoyang Xia, Akash Srivastava, and Dimitris Metaxas. Improving tuning-free real image editing with proximal guidance, 2023. 1", + "[15] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control, 2022. 1", + "[16] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. arXiv, 2023. 2, 4, 5, 7, 1, 15", + "[17] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2", + "[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 5, 1", + "[19] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations, 2024. 1", + "[20] Jaeseok Jeong, Mingi Kwon, and Youngjung Uh. Training-free content injection using h-space in diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5151-5161, 2024. 4", + "[21] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In CVPR, 2023. 1", + "[22] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. In ICLR, 2023. 4", + "[23] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In CVPR, 2023. 1", + "[24] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, pages 423-439. Springer, 2022. 3", + "[25] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, and Errui Ding. Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In ICCV, 2021. 2", + "[26] Frank J. Massey. The kolmogorov-smirnov test for goodness of fit. Journal of the American Statistical Association, 46 (253):68-78, 1951. 2", + "[27] Benyuan Meng, Qianqian Xu, Zitai Wang, Xiaochun Cao, and Qingming Huang. Not all diffusion model activations have been evaluated as discriminative features. arXiv preprint arXiv:2410.03558, 2024. 4", + "[28] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 1", + "[29] Daiki Miyake, Akihiro Iohara, Yu Saito, and Toshiyuki Tanaka. Negative-prompt inversion: Fast image inversion for editing with text-guided diffusion models, 2023. 1", + "[30] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, 2023. 1" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15768", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Thao Nguyen, Yuheng Li, Utkarsh Ojha, and Yong Jae Lee. Visual instruction inversion: Image editing via visual prompting. NeurIPS, 2023. 1", + "[32] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In CVPR, 2019. 2", + "[33] Yong-Hyun Park, Mingi Kwon, Jaewoong Choi, Junghyo Jo, and Youngjung Uh. Understanding the latent space of diffusion models through the lens of riemannian geometry. Advances in Neural Information Processing Systems, 36: 24129-24142, 2023. 4", + "[34] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv, 2023. 1", + "[35] Tianhao Qi, Shancheng Fang, Yanze Wu, Hongtao Xie, Jiawei Liu, Lang Chen, Qian He, and Yongdong Zhang. Deadiff: An efficient stylization diffusion model with disentangled representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8693-8702, 2024. 1, 5, 16", + "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2", + "[37] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, pages 8821-8831. PMLR, 2021. 1", + "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1", + "[39] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 1, 5", + "[40] Simo Ryu. Low-rank adaptation for fast text-to-image diffusion fine-tuning. https://github.com/cloneofsimo/lora, 2023.5,8,1", + "[41] Lu Sheng, Ziyi Lin, Jing Shao, and Xiaogang Wang. Avatarnet: Multi-scale zero-shot style transfer by feature decoration. In CVPR, 2018. 2", + "[42] Kihyuk Sohn, Lu Jiang, Jarred Barber, Kimin Lee, Nataniel Ruiz, Dilip Krishnan, Huiwen Chang, Yuzhzen Li, Irfan Essa, Michael Rubinstein, Yuan Hao, Glenn Entis, Irina Blok, and Daniel Castro Chin. Styledrop: Text-to-image synthesis of any style. In NeurIPS, 2023. 2, 5, 1", + "[43] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2020. 5, 1", + "[44] Adéla Šubrtová, Michal Lukáč, Jan Čech, David Futschik, Eli Shechtman, and Daniel Sykora. Diffusion image analogies. In SIGGRAPH, 2023. 1", + "[45] Yasheng Sun, Yifan Yang, Houwen Peng, Yifei Shen, Yuqing Yang, Han Hu, Lili Qiu, and Hideki Koike. Imagebrush: Learning visual in-context instructions for exemplar-based image manipulation. NeurlPS, 2023. 1" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In CVPR, 2023. 4, 1", + "[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 2", + "[48] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. $p+$ : Extended textual conditioning in text-to-image generation. arXiv, 2023. 1", + "[49] Bram Wallace, Akash Gokul, and Nikhil Naik. Edict: Exact diffusion inversion via coupled transformations, 2022. 1", + "[50] Haofan Wang, Matteo Spinelli, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv preprint arXiv:2404.02733, 2024. 5", + "[51] Haofan Wang, Peng Xing, Renyuan Huang, Hao Ai, Qixun Wang, and Xu Bai. Instantstyle-plus: Style transfer with content-preserving in text-to-image generation. arXiv preprint arXiv:2407.00788, 2024. 5", + "[52] Zhouxia Wang, Xintao Wang, Liangbin Xie, Zhongang Qi, Ying Shan, Wenping Wang, and Ping Luo. Styleadapter: A single-pass lora-free model for stylized image generation. arXiv, 2023. 1", + "[53] Chen Henry Wu and Fernando De la Torre. A latent space of stochastic diffusion models for zero-shot image editing and guidance. In ICCV, 2023. 1", + "[54] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, 2023. 1", + "[55] Peng Xing, Haofan Wang, Yanpeng Sun, Qixun Wang, Xu Bai, Hao Ai, Renyuan Huang, and Zechao Li. Csgo: Content-style composition in text-to-image generation. arXiv preprint arXiv:2408.16766, 2024. 1, 5", + "[56] Sihan Xu, Yidong Huang, Jiayi Pan, Ziqiao Ma, and Joyce Chai. Inversion-free image editing with natural language. arXiv, 2023. 3, 4", + "[57] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. In SIGGRAPH Asia, 2023. 1", + "[58] Yuan Yao, Jianqiang Ren, Xuansong Xie, Weidong Liu, Yong-Jin Liu, and Jun Wang. Attention-aware multi-stroke style transfer, 2019. 2", + "[59] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv, 2023. 1, 5", + "[60] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 8, 1", + "[61] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 2" + ], + "bbox": [ + 516, + 92, + 903, + 844 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15769", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/StyleKeeper_ Prevent Content Leakage using Negative Visual Query Guidance/45fcd356-f226-4ed7-980b-85432724bf8a_model.json b/2025/StyleKeeper_ Prevent Content Leakage using Negative Visual Query Guidance/45fcd356-f226-4ed7-980b-85432724bf8a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..52aff2e77ee9972f0173b375bf5d597eaf0d67af --- /dev/null +++ b/2025/StyleKeeper_ Prevent Content Leakage using Negative Visual Query Guidance/45fcd356-f226-4ed7-980b-85432724bf8a_model.json @@ -0,0 +1,2343 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.131, + 0.895, + 0.154 + ], + "angle": 0, + "content": "StyleKeeper: Prevent Content Leakage using Negative Visual Query Guidance" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.181, + 0.349, + 0.234 + ], + "angle": 0, + "content": "Jaeseok Jeong\\*† \nYonsei University \njete-jeong@yonsei.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.391, + 0.182, + 0.588, + 0.234 + ], + "angle": 0, + "content": "Junho Kim* \nNAVER AI Lab \njhkim.ai@navercorp.com" + }, + { + "type": "text", + "bbox": [ + 0.63, + 0.182, + 0.854, + 0.234 + ], + "angle": 0, + "content": "Gayoung Lee \nNAVER AI Lab \ngayoung.lee@navercorp.com" + }, + { + "type": "title", + "bbox": [ + 0.316, + 0.242, + 0.447, + 0.276 + ], + "angle": 0, + "content": "Yunjey Choi \nNAVER AI Lab" + }, + { + "type": "text", + "bbox": [ + 0.288, + 0.281, + 0.477, + 0.295 + ], + "angle": 0, + "content": "yunjay.choi@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.543, + 0.242, + 0.699, + 0.295 + ], + "angle": 0, + "content": "Youngjung Uh‡ Yonsei University yj.uh@yonsei.ac.ki" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.33, + 0.327, + 0.345 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.361, + 0.485, + 0.664 + ], + "angle": 0, + "content": "In the domain of text-to-image generation, diffusion models have emerged as powerful tools. Recently, studies on visual prompting, where images are used as prompts, have enabled more precise control over style and content. However, existing methods often suffer from content leakage, where undesired elements of the visual style prompt are transferred along with the intended style. To address this issue, we 1) extend classifier-free guidance (CFG) to utilize swapping self-attention and propose 2) negative visual query guidance (NVQG) to reduce the transfer of unwanted contents. NVQG employs negative score by intentionally simulating content leakage scenarios that swap queries instead of key and values of self-attention layers from visual style prompts. This simple yet effective method significantly reduces content leakage. Furthermore, we provide careful solutions for using a real image as visual style prompts. Through extensive evaluation across various styles and text prompts, our method demonstrates superiority over existing approaches, reflecting the style of the references, and ensuring that resulting images match the text prompts." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.692, + 0.222, + 0.708 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.717, + 0.484, + 0.854 + ], + "angle": 0, + "content": "Text-to-image diffusion models (T2I DMs) excel at synthesizing images that correspond to given text prompts [37, 38]. However, relying solely on text prompts may not allow precise control over the desired output. Even with highly detailed text prompts, controlling the exact style of the resulting images remains a challenge because of the modality gap between text and image. For example, from \"a red car\", different people imagine different variants of red. Instead, \"A dinosaur\" drawn in the style of the reference" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.328, + 0.907, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.555, + 0.907, + 0.597 + ], + "angle": 0, + "content": "Figure 1. Qualitative comparison with the same style. Competitors face challenges in generating images with diverse layouts and compositions, i.e., content leakage from the reference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.626, + 0.907, + 0.656 + ], + "angle": 0, + "content": "image in Figure 1 provides a more concrete specification of the desired style." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.907, + 0.779 + ], + "angle": 0, + "content": "However, using reference images as visual style prompts comes with a fundamental challenge: a dilemma between better reflecting style and preventing content leakage. Figure 1 shows that the pose of the reference horse appears in the results of previous methods. The content leakage may include poses, specific shapes, layouts or entire objects. Furthermore, content leakage reduces diversity and text alignments as shown in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Approaches that involve training the model, such as fine-tuning a diffusion model on a set of themed images [21, 39], learning new text embeddings [2, 10, 13], or adapting cross-attention modules to incorporate image features [35, 52, 55, 59], come with significant drawbacks. These methods not only incur high computational costs, but also face an inherent trade-off between style and content. Achieving a balance where the model faithfully captures the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.863, + 0.218, + 0.875 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.876, + 0.437, + 0.888 + ], + "angle": 0, + "content": "This work was done during his internship at Naver AI Lab." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "\\(\\ddagger\\) Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.863, + 0.437, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15760" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "desired style without unintentionally preserving content remains a challenge." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.123, + 0.483, + 0.242 + ], + "angle": 0, + "content": "To address this, recent papers [1, 6, 16] have proposed training-free methods by modifying feature swapping mechanisms within the self-attention layers. Specifically, they replace the key and value features in the original process with those extracted from the visual style prompt. This allows the query to preserve content while the key-value pair carries the style, enabling promising style transfer performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.244, + 0.484, + 0.349 + ], + "angle": 0, + "content": "However, these methods struggle to fully eliminate content leakage [42]. Additionally, Hertz et al. [16] does not support real images as reference inputs, while Chung et al. [6] and Alaluf et al. [1] primarily focus on image-to-image (I2I) tasks. When using real images as style references, they often suffer from either poor style transfer or excessive content leakage, significantly limiting their effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.351, + 0.484, + 0.531 + ], + "angle": 0, + "content": "In this study, we propose a method to separately control the strength of style and content from visual prompts such that the results align with the text prompts. We design a novel variation of classifier-free guidance (CFG) [17], emulating images with undesirable content leakage as a negative guidance. To address content leakage, we introduce negative visual query guidance, ensuring a clear separation between content and style. We also show that stochastic encoding achieves better style reflection with less artifacts and computation, and incorporate color calibration to precisely match the final output to the reference image's color statistics leading to better style reflection." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.533, + 0.484, + 0.76 + ], + "angle": 0, + "content": "We analyze where to apply swapping self-attention, identifying the optimal layers for balancing style transfer and content fidelity. Additionally, our method can effectively remove content that is difficult to eliminate with key and value swapping alone, working successfully even in cases with significant structural gaps between the style image and content text, as shown in Figure 1 (e.g., a complex scene of \"a horse with the vibrant colors of a Chinese ink painting\" and a single object \"Dinosaur\"). Our method is well compatible with ControlNet for I2I style transfer, further enhancing its flexibility. Qualitative and quantitative evaluations show our method outperforms state-of-the-art approaches, providing precise control over content and style without content leakage. Our approach is both robust and efficient, ideal for visual style prompting tasks.\\(^{1}\\)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.774, + 0.307, + 0.792 + ], + "angle": 0, + "content": "2. Visual style prompting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.8, + 0.483, + 0.876 + ], + "angle": 0, + "content": "We propose StyleKeeper which receives a text prompt and a visual style prompt to generate new images without content leakage. The results contain the content and style specified by the text prompt and the visual style prompt, respectively, with variations due to initial noises. The overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.244 + ], + "angle": 0, + "content": "of our method is illustrated in Figure 2. First, we explain the swapping self-attention in the aspect of style transfer literature. StyleKeeper consists of classifier-free guidance (CFG) with swapping self-attention, negative visual query guidance (NVQG), optimal layer choice, stochastic encoding of real visual style prompts, and generalization to ControlNet for real content images. We explain the first three components in text-to-image (T2I) scenario with generated visual style prompt. Then we proceed to T2I with real visual style prompt and image-to-image (I2I) scenario." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.253, + 0.907, + 0.283 + ], + "angle": 0, + "content": "2.1. Swapping self-attention in style transfer literature" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.29, + 0.908, + 0.627 + ], + "angle": 0, + "content": "Modern diffusion models consist of a number of self-attention and cross-attention blocks [47]. Both of them employ the attention mechanism, which first obtains an attention map using similarity between query features \\( Q \\) and key features \\( K \\), then aggregates value features \\( V \\) using the attention map as weights: \\( Attention(Q, K, V) = \\text{Softmax}(\\frac{QK^T}{\\sqrt{d}})V \\). Opposed to the cross-attention layer, self-attention layer receives key and values coming from the main denoising process, which has spatial dimensions with more freedom to represent spatially varying visual elements. As our goal is to reflect style elements from a reference image that are not easily represented by textual description, we opt to borrow key and values of self-attention layers in the reference process to the original process, namely swapping self-attention (Figure 2). In addition, swapping self-attention has a strong connection with style transfer literature [8, 25, 32, 41, 58]. where the attention mechanism reassembles visual features of a style image (key, value) on a content image (query). However, methods [3, 16] utilizing self-attention operation often exhibit artifacts and content leakage since style elements are entangled with content elements in the visual feature." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.637, + 0.907, + 0.666 + ], + "angle": 0, + "content": "2.2. CFG sampling with swapping self-attention for T2I" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.909, + 0.78 + ], + "angle": 0, + "content": "Previous works [1, 6] have not taken CFG into account while manipulating features. We propose CFG combined with swapping self-attention to reflect a visual style prompt in the T2I results without artifacts. CFG [17] is a key technique that steers generated images toward specified text prompts. Given a score \\(\\epsilon_{\\theta}(x_t,c)\\) conditioned on \\(c\\) and an unconditional score \\(\\epsilon_{\\theta}(x_t,\\emptyset)\\), the CFG score is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.792, + 0.907, + 0.809 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\epsilon} _ {\\theta} = (1 + w) \\epsilon_ {\\theta} \\left(x _ {t}, c\\right) - w \\epsilon_ {\\theta} \\left(x _ {t}, \\emptyset\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.819, + 0.907, + 0.867 + ], + "angle": 0, + "content": "where \\( w \\) controls guidance strength. CFG with \\( w > 1 \\) enhances image quality and text alignment, but excessive \\( w \\) may lead to mode collapse [5]. Notably, CFG has not been" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.907, + 0.901 + ], + "angle": 0, + "content": "2We omit the diffusion timestep \\(t\\) in the arguments and abuse \\(x_{t}\\) instead of \\(z_{t}\\) from latent diffusion model." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.355, + 0.901 + ], + "angle": 0, + "content": "1Our code will be released for reproducibility." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15761" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.138, + 0.092, + 0.862, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.352, + 0.908, + 0.425 + ], + "angle": 0, + "content": "Figure 2. Overveiw of StyleKeeper. Our proposed method includes 4 key components, highlighted in red boxes. First, stochastic encoding (Section 3) converts reference images into suitable latents for the visual style prompting task. Second, swapping self-attention (Section 2.2, 2.4) ensures the reference image's style is accurately reflected. Third, negative visual query guidance (Section 2.3) reduces content leakage from the reference image, allowing the desired text content (e.g., \"Moose\") to be better represented. Lastly, color calibration (Section 3) minimizes errors during the denoising process, helping to produce a cleaner final image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.449, + 0.483, + 0.478 + ], + "angle": 0, + "content": "explored in the context of reflecting denoising process with modified features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.479, + 0.483, + 0.511 + ], + "angle": 0, + "content": "Assuming a given condition \\( c \\) has hidden content \\( h^{\\mathrm{content}} \\) and style \\( h^{\\mathrm{style}} \\), we model \\( p_{\\theta}(x_0|h_{\\mathrm{text}}^{\\mathrm{content}},h_{\\mathrm{visual}}^{\\mathrm{style}}) \\) using:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.513, + 0.483, + 0.546 + ], + "angle": 0, + "content": "- The original T2I score: \\(\\epsilon_{\\theta}(x_t, c_{\\text{text}})\\), leading to \\(x_0^{\\text{ori}} \\sim p_{\\theta}(x_0 | c_{\\text{text}}) = p_{\\theta}(x_0 | h_{\\text{text}}^{\\text{content}}, h_{\\text{text}}^{\\text{style}})\\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.546, + 0.483, + 0.579 + ], + "angle": 0, + "content": "- The reference score: \\(\\epsilon_{\\theta}(x_t, c_{\\mathrm{visual}})\\), leading to \\(x_0^{\\mathrm{visual}} \\sim p_{\\theta}(x_0 | c_{\\mathrm{visual}}) = p_{\\theta}(x_0 | h_{\\mathrm{visual}}^{\\mathrm{content}}, h_{\\mathrm{visual}}^{\\mathrm{style}})\\)." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.513, + 0.483, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.483, + 0.609 + ], + "angle": 0, + "content": "To ensure proper content and style control, we design the CFG score as:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.619, + 0.483, + 0.674 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tilde {\\epsilon} _ {\\theta} \\left(x _ {t}, h _ {\\text {t e x t}} ^ {\\text {c o n t e n t}}, h _ {\\text {v i s u a l}} ^ {\\text {s t y l e}}\\right) = (1 + w) \\\\ \\ddot {\\epsilon} _ {\\theta} \\left(x _ {t}, Q _ {\\text {t e x t}}, K V _ {\\text {v i s u a l}}\\right) \\tag {2} \\\\ - w \\epsilon_ {\\theta} (x _ {t}, \\emptyset). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.483, + 0.732 + ], + "angle": 0, + "content": "where \\(\\ddot{\\epsilon}_{\\theta}(x_{t},Q_{\\mathrm{text}},KV_{\\mathrm{visual}})\\) represents a KV-injected denoising score, indicating feature manipulation rather than direct conditioning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.483, + 0.822 + ], + "angle": 0, + "content": "For given two denoising processes, one as original and another as a reference, borrowing the key-value in self-attention from the reference to the original process, i.e., key-value (KV) injection, tends to produces results with the content from the original process and the style from the reference process with limited control [1, 6, 56]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.822, + 0.483, + 0.867 + ], + "angle": 0, + "content": "We define the KV-injected score by \\(\\ddot{\\epsilon}_{\\theta}(x_{t},Q_{\\mathrm{text}},KV_{\\mathrm{visual}})\\) where \\(Q_{\\mathrm{text}}\\) and \\(KV_{\\mathrm{visual}}\\) denote the query from the original score \\(\\epsilon_{\\theta}(x_t,c_{\\mathrm{text}})\\) and the" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.449, + 0.907, + 0.478 + ], + "angle": 0, + "content": "key, value from the reference score \\(\\epsilon_{\\theta}(x_t,c_{\\mathrm{visual}})\\). Then KV-injected-Attention becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.493, + 0.906, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t e n t i o n} \\left(Q _ {\\text {t e x t}}, K _ {\\text {v i s u a l}}, V _ {\\text {v i s u a l}}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {Q _ {\\text {t e x t}} K _ {\\text {v i s u a l}} ^ {\\intercal}}{\\sqrt {d}}\\right) V _ {\\text {v i s u a l}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.536, + 0.753, + 0.551 + ], + "angle": 0, + "content": "We omit the layer index for brevity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.704 + ], + "angle": 0, + "content": "Naive denoising process with \\(\\ddot{e}_{\\theta}(x_{t},Q_{\\mathrm{text}},KV_{\\mathrm{visual}})\\) provides limited control in generating images with content and style specified by a text \\(c_{\\mathrm{text}}\\) and a visual style prompt, respectively. Moving forward, our CFG with swapping self-attention in Eq. (2) enjoys higher image quality and more accurate text alignment than the naive denoising process as in the original CFG for T2I generation. The results are deferred to Section 4.1. However, content leakage still hinders the creation of diverse images, and disrupts text alignment." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.72, + 0.792, + 0.736 + ], + "angle": 0, + "content": "2.3. Negative visual query guidance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.744, + 0.906, + 0.805 + ], + "angle": 0, + "content": "We propose negative visual query guidance (NVQG) to mitigate content leakage by reducing the content \\( h_{\\mathrm{visual}}^{\\mathrm{content}} \\) from the visual style prompt appearing in the results. Briefly, NVQG negates the CFG of a score \\( \\ddot{\\epsilon}(x_t, Q_{\\mathrm{visual}}, KV_{\\mathrm{text}}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.806, + 0.907, + 0.902 + ], + "angle": 0, + "content": "In Liu et al. [24], a complex text condition \\(c\\) is factorized as a set of conditions \\(\\{c_0, c_1, \\ldots\\}\\) and Bayes' rule induces \\(p_{\\theta}(x_t | c_0, c_1, \\ldots) \\propto \\Pi \\frac{p_{\\theta}(x_t | c_i)}{p_{\\theta}(x_t)}\\). Then, the score of the complex text condition \\(c\\) becomes \\(\\epsilon_{\\theta}(x_t, c) = \\epsilon_{\\theta}(x_t, \\emptyset) + \\Pi (\\epsilon_{\\theta}(x_t, c_i) - \\epsilon_{\\theta}(x_t, \\emptyset))\\). It allows reducing a specific concept \\(\\tilde{c}\\) with composition by negating its guidance with scale" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.902 + ], + "angle": 0, + "content": "3The original score and its query within are recursively altered by KV injection along the denoising process." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15762" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.09, + 0.482, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.272, + 0.485, + 0.356 + ], + "angle": 0, + "content": "Figure 3. The effect of swapping self-attention across different blocks. Swapping self-attention on the bottleneck and downblocks causes content leakage, resulting in cat-like images despite a dog prompt, while swapping on downblocks disrupts resulting images. We only apply swapping self-attention in the upblocks to reflect style elements effectively." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.383, + 0.131, + 0.396 + ], + "angle": 0, + "content": "\\(w_{\\mathrm{neg}}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.404, + 0.483, + 0.434 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {\\theta} \\left(x _ {t}, c, \\text {n o t} \\tilde {c}\\right) = \\epsilon_ {\\theta} \\left(x _ {t}, c\\right) - w _ {\\text {n e g}} \\left(\\epsilon_ {\\theta} \\left(x _ {t}, \\tilde {c}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, \\emptyset\\right)\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.434, + 0.484, + 0.498 + ], + "angle": 0, + "content": "Although we design \\(\\ddot{\\epsilon}_{\\theta}(x_{t},Q_{\\mathrm{text}},KV_{\\mathrm{visual}})\\) to predict the score toward \\(p_{\\theta}(x_0|h_{\\mathrm{text}}^{\\mathrm{content}},h_{\\mathrm{visual}}^{\\mathrm{style}})\\), \\(\\ddot{\\epsilon} (x_{t},Q_{\\mathrm{text}},KV_{\\mathrm{visual}})\\) still contain \\(h_\\mathrm{visual}\\). Assuming a hidden factorization \\(KV_{\\mathrm{visual}} = \\{KV_{\\mathrm{visual}}^{\\mathrm{content}},KV_{\\mathrm{visual}}^{\\mathrm{style}}\\}\\), Bayes' rule induces" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.504, + 0.482, + 0.578 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} p _ {\\theta} \\left(x _ {t} \\mid Q _ {\\text {t e x t}}, K V _ {\\text {v i s u a l}} ^ {\\text {s t y l e}}, K V _ {\\text {v i s u a l}} ^ {\\text {c o n t e n t}}\\right) \\propto p _ {\\theta} \\left(x _ {t}\\right) \\frac {p _ {\\theta} \\left(x _ {t} \\mid Q _ {\\text {t e x t}} , K V _ {\\text {v i s u a l}} ^ {\\text {s t y l e}}\\right)}{p _ {\\theta} \\left(x _ {t}\\right)} \\\\ \\times \\frac {p _ {\\theta} \\left(x _ {t} \\mid Q _ {\\emptyset} , K V _ {\\text {v i s u a l}} ^ {\\text {c o n t e n t}}\\right)}{p _ {\\theta} \\left(x _ {t}\\right)}. \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.483, + 0.613 + ], + "angle": 0, + "content": "Since \\(\\epsilon_{\\theta}(x_t) = \\epsilon_{\\theta}(x_t, Q_{\\emptyset}, KV_{\\emptyset})\\), we obtain the desired conditional score of \\(\\hat{c} = \\{Q_{\\mathrm{text}}, KV_{\\mathrm{visual}}^{\\mathrm{style}}\\}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.621, + 0.482, + 0.689 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\epsilon_ {\\theta} \\left(x _ {t}, \\hat {c}\\right) \\leftarrow w _ {\\text {v i s u a l}} \\left(\\ddot {\\epsilon} _ {\\theta} \\left(x _ {t}, Q _ {\\text {t e x t}}, K V _ {\\text {v i s u a l}}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}\\right)\\right) \\\\ - w _ {\\text {c o n t e n t}} \\left(\\ddot {\\epsilon} _ {\\theta} \\left(x _ {t}, Q _ {\\emptyset}, K V _ {\\text {v i s u a l}} ^ {\\text {c o n t e n t}}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}\\right)\\right) \\\\ + \\epsilon_ {\\theta} \\left(x _ {t}\\right). \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.691, + 0.483, + 0.841 + ], + "angle": 0, + "content": "where \\( w_{\\mathrm{visual}} \\) and \\( w_{\\mathrm{content}} \\) regulate the strength of each guidance. By leveraging query injection [1, 6, 46, 56], we approximate \\( \\ddot{\\epsilon}_{\\theta}(x_t, Q_{\\emptyset}, KV_{\\mathrm{visual}}^{\\mathrm{content}}) \\approx \\ddot{\\epsilon}_{\\theta}(x_t, Q_{\\mathrm{visual}}, KV_{\\emptyset}) \\). Finally, inserting (6) into (2), and empirically replacing \\( \\epsilon_{\\theta}(x_t) \\) with \\( \\ddot{\\epsilon}\\theta(x_t, Q_{\\mathrm{visual}}) \\) results in similar performance. This allows a simplified reformulation of diffusion sampling using \\( w' = w_{\\mathrm{visual}}(w + 1) \\): Since we can highly relate the \\( w' \\ddot{\\epsilon}_{\\theta}(x_t, Q_{\\mathrm{visual}}) \\) to the concept negation in (4) which guides the negation of concept with a scale \\( w_{\\mathrm{neg}} \\), we named the query term as negative visual query guidance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.462, + 0.866 + ], + "angle": 0, + "content": "2.4. Choosing blocks for swapping self-attention" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In order to maximize the effect of NVQG in preventing content leakage, we explore the depth of the self-attention" + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.092, + 0.887, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.317, + 0.907, + 0.415 + ], + "angle": 0, + "content": "Figure 4. The effect of CFG and the proposed negative visual query guidance on image generation. The reference images provide the style for each generated output. Without NVQG, content leakage occurs, and the generated images fail to fully capture the intended content. In contrast, using NVQG ensures better alignment with both the reference style and the \"Cat\" prompt, reducing content distortion and improving quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.425, + 0.907, + 0.756 + ], + "angle": 0, + "content": "blocks to be swapped in the sense of granularity of visual elements. Modern architecture of diffusion models roughly consists of three sections in a sequence: downblocks, bottleneck blocks, and upblocks. Given that the bottleneck of diffusion models contains content elements of the image [20, 22, 33], we opt not to apply swapping self-attention to bottleneck blocks to prevent transferring contents in a reference image. Figure 3 shows that not swapping the bottleneck blocks prevents content leaking from the reference image. However, the synthesized images show disrupted results with seriously scattered objects. Furthermore, while swapping self-attention implements the reassembling operation, simply applying to all self-attention layers exposes a content leakage problem, where the content of the reference image influences the resulting image, as shown in the first row of Figure 3. I.e., the results contain cats even though the prompts specify \"a dog\". We conjecture that this phenomenon happens because feature maps of downblocks have unclear content layout [3, 27], so substituting features based on this inaccurate layout causes the disrupted results. To avoid injecting unnecessary features, we choose to swap the key and value of self-attention only in upblocks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.758, + 0.906, + 0.832 + ], + "angle": 0, + "content": "We note that Hertz et al. [16] applies self-attention operation at all blocks and suffers content leakage. The last row of Figure 3 shows the success of our strategy. Following this, we proceed to search for a specific optimal layer in the upblock in Section 4.2" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.845, + 0.753, + 0.862 + ], + "angle": 0, + "content": "3. Real images as references" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.903 + ], + "angle": 0, + "content": "So far, we have assumed a generated visual style prompt \\( x_0^{\\mathrm{visual}} \\sim p_\\theta(x_0 | c_{\\mathrm{visual}}} \\). Here, we allow real visual style" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15763" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.121, + 0.089, + 0.47, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.525, + 0.483, + 0.637 + ], + "angle": 0, + "content": "Figure 5. Analysis on the optimal range of upblocks for swapping self-attention. We find the optimal range of upblocks for a balanced trade-off between different aspects. The images on the right illustrate the visual results for different upblock layer indices, with the red cross indicating poor diversity and misalignment to the text prompt, the red triangle indicating a lack of style similarity, and the yellow star indicating the optimal results. Please refer to Section 4.2 for details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.664, + 0.481, + 0.679 + ], + "angle": 0, + "content": "prompts by 1) stochastic encoding and 2) color calibration." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.679, + 0.483, + 0.725 + ], + "angle": 0, + "content": "We propose stochastic encoding to obtain \\( x_{t}^{\\mathrm{visual}} \\sim q(x_{t}|x_{0}^{\\mathrm{visual}}) \\) by adding a random noise on \\( x_{0}^{\\mathrm{visual}} \\) following the forward process of DMs [18]:" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.735, + 0.483, + 0.755 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {t} \\sim \\mathcal {N} (0, I), x _ {t} ^ {\\text {v i s u a l}} = \\sqrt {\\alpha_ {t}} \\cdot x _ {0} ^ {\\text {v i s u a l}} + \\sqrt {1 - \\alpha_ {t}} \\cdot \\epsilon_ {t} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.87 + ], + "angle": 0, + "content": "At each timestep, we sample \\(\\epsilon_{t}\\) to encode \\(x_{t}^{\\mathrm{visual}}\\). It ensures that \\(x_{t}^{\\mathrm{visual}}\\) lies on the learned trajectory and does not suffer from accumulative numerical error due to iterative process of DDIM inversion used by previous methods [6, 16] as shown in Figure A1. Furthermore, it does not need to store the intermediate latents as opposed to DDPM inversion used by Alaluf et al. [1]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Although stochastic encoding performs better than DDIM inversion, a subtle color discrepancy occurs between" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.338 + ], + "angle": 0, + "content": "the resulting images and the visual style prompts. We introduce color calibration at \\( x_{t}^{\\mathrm{ori}} \\) in the original process to match the statistics of predicted \\( x_{0}^{\\mathrm{ori}} \\) to predicted \\( x_{0}^{\\mathrm{visual}} \\). In Gatys [12], distance between channel-wise statistics is employed as a style loss for style transfer. In Song et al. [43], predicted \\( x_{0} \\left( = \\frac{x_{t}^{\\mathrm{visual}} - \\sqrt{1 - \\alpha_{t}} \\cdot \\epsilon_{\\theta}(\\hat{x}_{t})}{\\sqrt{\\alpha_{t}}} \\right) \\) allows to estimate \\( x_{0} \\) with high probability at intermediate timesteps using deterministic sampling. Inheriting the advantage, we execute adain operation to match mean&std of predicted \\( x_{0}^{\\mathrm{ori}} \\) with those of \\( x_{0}^{\\mathrm{visual}} \\). It allows precise color calibration rather than directly matching \\( x_{t}^{\\mathrm{ori}} \\) to \\( x_{t}^{\\mathrm{visual}} \\) in Alaluf et al. [1], Chung et al. [6]. Furthermore, our approach differs from Chung et al. [6] in that using predicted \\( x_{0} \\) at intermediate timestep \\( t \\in (0,T) \\) other than \\( x_{T} \\) by reducing cumulative sampling error after the operation. Furthermore, Chung et al. [6] executes AdaIN at timestep T, inducing lengthy cumulative error." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.339, + 0.905, + 0.385 + ], + "angle": 0, + "content": "We provide supportive experiments that show the effectiveness of the proposed method in Section 4.4 and a detailed Algorithm in Appendix 3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.401, + 0.646, + 0.418 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.427, + 0.905, + 0.578 + ], + "angle": 0, + "content": "In this section, we describe the effects of our proposed methods: CFG with swapping self-attention, Negative visual query guidance (NVQG), stochastic encoding, and color calibration. For swapping self-attention, we provide a detailed analysis through experiments to determine the optimal layers for swapping. The impact of NVQG is demonstrated through qualitative results. Additionally, we show why stochastic encoding outperforms DDIM inversion when inverting real images, and we highlight the benefits of color calibration through experimental results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.905, + 0.685 + ], + "angle": 0, + "content": "We also conducted both quantitative and qualitative comparisons of our method against various competitors, including StyleAligned [16], IP-Adapter [59], Dreambooth-LoRA [39, 40], StyleDrop [42], DEADiff [35], CSGO [55] and InstantStyle [50, 51]. The details of these comparisons, along with the experimental setup and metrics, are described in the Appendix 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.697, + 0.803, + 0.714 + ], + "angle": 0, + "content": "4.1. Effectiveness of CFG and NVQG" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This section analyzes the effects of Classifier-Free Guidance (CFG) and Negative Visual Query Guidance (NVQG) on image generation, with a focus on text alignment and content leakage. Figure 4 shows the results of the three configurations. In the \\(1^{\\text{st}}\\) row, without CFG and NVQG, the generated images suffer from severe artifacts. The absence of CFG causes poor image quality resulting in significant misalignment with the prompt. In the \\(2^{\\text{nd}}\\) row, CFG with swapping self-attention improves the text misalignment by boosting image quality. Here, the \"cat\" in target text prompt becomes clearer in the generated images. However, content leakage from the reference image remains where unwanted" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15764" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.024, + 0.482, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.245, + 0.483, + 0.314 + ], + "angle": 0, + "content": "Figure 6. Attention map visualization over late and early upblock layers. The late upblock better focuses on the style-corresponding region than the early upblock, leading to more freedom to reassemble small parts. The early upblock attends a larger region leading to content leakage." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.339, + 0.483, + 0.368 + ], + "angle": 0, + "content": "elements (layouts, structure, and composition) from the reference image affect the results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.369, + 0.483, + 0.414 + ], + "angle": 0, + "content": "In the \\(3^{\\mathrm{rd}}\\) row, NVQG releases the content leakage and produces the best results closely matching the text prompt while reflecting style reference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.414, + 0.483, + 0.49 + ], + "angle": 0, + "content": "Overall, Figure 4 demonstrates the critical role of NVQG in reducing content leakage, enjoying the quality boosting of CFG. Together, they ensure that the generated images align to both the target text prompt and the visual style prompts, resulting in high-quality, coherent outputs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.49, + 0.484, + 0.535 + ], + "angle": 0, + "content": "We qualitatively showcase diversity of results within a text prompt in Figure A9 and text alignment with complex text prompts in Figure A22." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.544, + 0.413, + 0.56 + ], + "angle": 0, + "content": "4.2. Analysis for Swapping Self-Attention" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.566, + 0.483, + 0.687 + ], + "angle": 0, + "content": "Optimal layers in upblocks Since recent large T2I DMs consist of many layers, we further analyze the behavior by changing the start of the swapping while the end of the swapping is fixed at the end. We use four key metrics as shown in Figure 5, there is a layer where all four metrics abruptly change (red line). Notably, this point remains consistent regardless of the reference image. All the results in the paper use only one fixed optimal layer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.483, + 0.746 + ], + "angle": 0, + "content": "We choose this layer as the optimal start of the swapping for a balanced trade-off among all aspects. We provide qualitative results with detailed split of layers in Figure A3 and A4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Visualizing Attention maps Figure 6 compares average attention maps from the late upblock and the early upblock applying swapping self-attention. Using late upblock has more freedom to reassemble the reference style elements leading to more doggy results than early upblock which produces some cat-like attributes. The right two columns visualize the attention weight of query points marked as red stars and yellow dots. Swapping self-attention on late upblock reassemblies features from a style correspondence," + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.088, + 0.878, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.172, + 0.877, + 0.184 + ], + "angle": 0, + "content": "(a) Synthesizing the content specified by the text prompts" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.187, + 0.875, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.268, + 0.866, + 0.28 + ], + "angle": 0, + "content": "(b) Synthesizing the pose specified by the text prompt" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.284, + 0.875, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.59, + 0.365, + 0.88, + 0.377 + ], + "angle": 0, + "content": "(c) Synthesizing a single object specified by the text prompt" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.39, + 0.907, + 0.432 + ], + "angle": 0, + "content": "Figure 7. Comparison for content leakage. While StyleAligned suffers from content leakage from the reference, our results clearly align with the text prompts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.907, + 0.537 + ], + "angle": 0, + "content": "e.g., texture and color. On the other hand, swapping self-attention on early upblock reassembles features from a wider area with different styles. This comparison clarifies the reasons for using only late upblock. Please refer to Figure A6, A7 for a more visualization and a detailed analysis." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.565, + 0.8, + 0.581 + ], + "angle": 0, + "content": "4.3. Comparison against competitors" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.589, + 0.907, + 0.755 + ], + "angle": 0, + "content": "Style & content control We provide a qualitative comparison in Figure 8, focusing on controlling style and content. Our method faithfully synthesizes content from the text prompt with the style of the reference image. In contrast, other methods add elements like color or texture not in the reference (e.g., feathers, bricks, iron, skin) and often suffer from content leakage (e.g., layout, screaming person, castle), which compromises text prompt faithfulness. Quantitative results in Figure 9 support these findings: IP-Adapter shows higher style similarity but neglects text prompts significantly." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Diversity within a text specification Starting from different initial noises, the diffusion models trained on a large dataset produce diverse results within the specification of a text prompt. Figure 1, A19 shows that ours can synthesize various poses and viewpoints while others barely change, i.e., other methods limit the diversity of the pretrained model due to content leakage. We further support our diversity in Figure A9." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15765" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.5, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.089, + 0.905, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.343, + 0.908, + 0.374 + ], + "angle": 0, + "content": "Figure 8. Qualitative comparison across various styles and text prompts. StyleKeeper faithfully reflects style elements in reference images without causing content leakage from the reference images." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.387, + 0.487, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.524, + 0.487, + 0.568 + ], + "angle": 0, + "content": "Figure 9. Quantitative comparison. Ours achieves the best style similarity while not compromising text alignment and being free from content leakage." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.584, + 0.487, + 0.903 + ], + "angle": 0, + "content": "Content leakage Content leakage refers to the phenomenon where the content of a reference image appears in a result. As evident in Figure A12, our model exhibits significantly less content leakage when compared to other models. We focus on a comparison against a runner-up method, StyleAligned, particularly in terms of how content leakage can be an obstacle to controlling the content using a text prompt. Figure 7a compares examples where strong content leakage prevents the object from the text from appearing in StyleAligned. We often observe famous paintings in the reference easily leak into the results of StyleAligned while ours does not struggle. Figure 7b compares examples where the pose in the reference leaks into the results of StyleAligned preventing the reflection of specified pose in the text prompt. On the other hand, our method reflects the correct poses. Figure 7c compares examples where the number of small instances in the reference leaks into the result of StyleAligned. Contrarily, our method correctly synthesizes a single instance of the content specified by the text. In Figure 9, we provide quantitative comparison results. It supports that StyleKeeper achieves the best style" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.909, + 0.445 + ], + "angle": 0, + "content": "similarity while not suffering content leakage. We can also observe that models with high content leakage exhibit poor text alignment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.48, + 0.909, + 0.633 + ], + "angle": 0, + "content": "Preserving the content of the original denoising process Figure A13 shows the results of ours and other methods using the same initial noise in each column. Our method precisely reflects the style in the reference with minimal changes of contents in the original denoising process. On the other hand, the other methods severely alter pose, shape, or layouts due to content leakage. It is an important virtue of controlling style to keep the rest intact. This also supports that ours is free from content leakage during the denoising process. We provide more results in Figure A14." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.649, + 0.807, + 0.667 + ], + "angle": 0, + "content": "4.4. DDIM inversion vs. Our strategy" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Figure 11 shows that StyleKeeper can take real images as style reference with our strategies. As shown in Figure 11, stochastic encoding outperforms DDIM inversion and color calibration improves text alignment and style similarity. We provide more results in Figure A15. Moreover, we show that our strategy can boost the performance of the other self-attention variants [3, 16] in Figure A16, A17 and color calibration can be used for generation settings in Figure A18. We provide an ablation study in Figure 12 for each configuration (swapping self-attention, NVQG, and color calibration) in both real reference and generated reference settings. In both settings, swapping self-attention and NVQG improve style similarity and text alignment, while color calibration helps improve style similarity. Additionally, stochastic encoding outperforms DDIM inversion." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15766" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.905, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.245, + 0.908, + 0.274 + ], + "angle": 0, + "content": "Figure 10. Qualitative comparison in I2I style transfer task. We compare our method for I2I style transfer task where the content image is given. Compared to the previous methods, ours achieves the best style similarity." + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.286, + 0.456, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.469, + 0.484, + 0.525 + ], + "angle": 0, + "content": "Figure 11. Comparison of DDIM inversion vs. stochastic encoding and the effect of color calibration. Stochastic encoding reduces artifacts in the resulting images, while color calibration better reflects the colors of the reference image." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.537, + 0.482, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.641, + 0.484, + 0.74 + ], + "angle": 0, + "content": "Figure 12. Ablation study of four proposed methods with a real reference and a generated reference setting.. We evaluated performance by individually removing each of the four proposed methods. In all cases, both generated and real images showed similar improvements in style similarity and text alignment. Moreover, stochastic encoding outperformed DDIM inversion in all aspects, comparing purple and blue points." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.769, + 0.383, + 0.785 + ], + "angle": 0, + "content": "4.5. Comparison for I2I style transfer" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.793, + 0.484, + 0.884 + ], + "angle": 0, + "content": "In Figure 10, we present a qualitative comparison between our method using ControlNet and existing state-of-the-art methods, CrossAttn [1] and StyleID [6], for the I2I style transfer task where a content image is provided. Both CrossAttn and StyleID inject the key & value obtained by inversion from the style images." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "CrossAttn often fails to transfer the detailed representa" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.907, + 0.414 + ], + "angle": 0, + "content": "tion of the reference style, leading to a rough, blocky appearance. For instance, the center examples in Figure 10 show that the style details are absent or inaccurate compared to our results. Similarly, StyleID suffers poor style reflection and color deficiency in the center and the rightmost example. In contrast, ours effectively reflects the details of the reference style image, with no noticeable transfer of the color values from the content images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.433, + 0.766, + 0.449 + ], + "angle": 0, + "content": "5. Conclusion and Limitation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.46, + 0.907, + 0.672 + ], + "angle": 0, + "content": "In this paper, we introduce StyleKeeper which faithfully applies the style of real reference images without content leakage in a training-free manner. CFG with swapping self-attention captures the reference image's style accurately and allows for direct content generation from the text. This integration with CFG enhances performance by balancing content generation and style transfer. To address content leakage, we propose Negative visual query guidance that ensures the reference image's content does not interfere with the text-specified content, enhancing diversity and improving text alignment & style similarity. Additionally, Stochastic encoding efficiently maps real images to suitable latents, removing artifacts. Color calibration aligns the final output to the reference's color statistics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.907, + 0.794 + ], + "angle": 0, + "content": "Our method shows both qualitative and quantitative improvements over existing approaches, providing a robust solution for visual style prompting without complex training. We also analyze where to apply swapping self-attention, identifying optimal layers to balance style transfer and content fidelity. StyleKeeper outperforms prior methods and works well with algorithms like ControlNet [60] and DBLoRA [40], as shown in Figure A20." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "However, StyleKeeper is limited by the capabilities of the pretrained diffusion models and cannot generate elements outside their scope (e.g., \"stone golem\" in Figure A24a). Also, if visual and textual styles conflict, the visual style dominates (Figure A24b). Future work could explore extending our method to other domains such as video, expanding its applicability and research potential." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15767" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.198 + ], + "angle": 0, + "content": "Acknowledgements We thank Jin-Hwa Kim for helpful comments. This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS-2024-00439762, Developing Techniques for Analyzing and Assessing Vulnerabilities, and Tools for Confidentiality Evaluation in Generative AI Models)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.211, + 0.188, + 0.226 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.236, + 0.484, + 0.291 + ], + "angle": 0, + "content": "[1] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer. In ACM SIGGRAPH 2024 Conference Papers, pages 1–12, 2024. 2, 3, 4, 5, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.293, + 0.484, + 0.334 + ], + "angle": 0, + "content": "[2] Omri Avrahami, Kfir Aberman, Ohad Fried, Daniel Cohen-Or, and Dani Lischinski. Break-a-scene: Extracting multiple concepts from a single image. In SIGGRAPH Asia, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.336, + 0.483, + 0.39 + ], + "angle": 0, + "content": "[3] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In ICCV), 2023. 2, 4, 7, 1, 9, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.392, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.449, + 0.483, + 0.503 + ], + "angle": 0, + "content": "[5] Hyungjin Chung, Jeongsol Kim, Geon Yeong Park, Hyelin Nam, and Jong Chul Ye. Cfg++: Manifold-constrained classifier free guidance for diffusion models. arXiv preprint arXiv:2406.08070, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.506, + 0.484, + 0.574 + ], + "angle": 0, + "content": "[6] Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8795-8805, 2024. 2, 3, 4, 5, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.576, + 0.482, + 0.617 + ], + "angle": 0, + "content": "[7] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.619, + 0.484, + 0.687 + ], + "angle": 0, + "content": "[8] Yingying Deng, Fan Tang, Weiming Dong, Chongyang Ma, Xingjia Pan, Lei Wang, and Changsheng Xu. Stytr2: Image style transfer with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11326-11336, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.689, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[9] Martin Nicolas Everaert, Marco Bocchio, Sami Arpa, Sabine Süssstrunk, and Radhakrishna Achanta. Diffusion in style. In ICCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.732, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[10] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[11] Daniel Garibi, Or Patashnik, Andrey Voynov, Hadar Averbuch-Elor, and Daniel Cohen-Or. Renoise: Real image inversion through iterative noising, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[12] Leon A Gatys. A neural algorithm of artistic style. arXiv preprint arXiv:1508.06576, 2015. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[13] Inhwa Han, Serin Yang, Taesung Kwon, and Jong Chul Ye. Highly personalized text embedding for image manipulation by stable diffusion. arXiv, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.236, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[14] Ligong Han, Song Wen, Qi Chen, Zhixing Zhang, Kunpeng Song, Mengwei Ren, Ruijiang Gao, Anastasis Stathopoulos, Xiaoxiao He, Yuxiao Chen, Di Liu, Qilong Zhangli, Jindong Jiang, Zhaoyang Xia, Akash Srivastava, and Dimitris Metaxas. Improving tuning-free real image editing with proximal guidance, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[15] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[16] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. arXiv, 2023. 2, 4, 5, 7, 1, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.904, + 0.287 + ], + "angle": 0, + "content": "[17] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.904, + 0.315 + ], + "angle": 0, + "content": "[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[19] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[20] Jaeseok Jeong, Mingi Kwon, and Youngjung Uh. Training-free content injection using h-space in diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5151-5161, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.415, + 0.905, + 0.454 + ], + "angle": 0, + "content": "[21] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.905, + 0.494 + ], + "angle": 0, + "content": "[22] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. In ICLR, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[23] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.608 + ], + "angle": 0, + "content": "[24] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, pages 423-439. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.61, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[25] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, and Errui Ding. Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[26] Frank J. Massey. The kolmogorov-smirnov test for goodness of fit. Journal of the American Statistical Association, 46 (253):68-78, 1951. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[27] Benyuan Meng, Qianqian Xu, Zitai Wang, Xiaochun Cao, and Qingming Huang. Not all diffusion model activations have been evaluated as discriminative features. arXiv preprint arXiv:2410.03558, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[28] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[29] Daiki Miyake, Akihiro Iohara, Yu Saito, and Toshiyuki Tanaka. Negative-prompt inversion: Fast image inversion for editing with text-guided diffusion models, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[30] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "15768" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.135 + ], + "angle": 0, + "content": "[31] Thao Nguyen, Yuheng Li, Utkarsh Ojha, and Yong Jae Lee. Visual instruction inversion: Image editing via visual prompting. NeurIPS, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.137, + 0.483, + 0.164 + ], + "angle": 0, + "content": "[32] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.167, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[33] Yong-Hyun Park, Mingi Kwon, Jaewoong Choi, Junghyo Jo, and Youngjung Uh. Understanding the latent space of diffusion models through the lens of riemannian geometry. Advances in Neural Information Processing Systems, 36: 24129-24142, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.239, + 0.482, + 0.293 + ], + "angle": 0, + "content": "[34] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.296, + 0.482, + 0.378 + ], + "angle": 0, + "content": "[35] Tianhao Qi, Shancheng Fang, Yanze Wu, Hongtao Xie, Jiawei Liu, Lang Chen, Qian He, and Yongdong Zhang. Deadiff: An efficient stylization diffusion model with disentangled representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8693-8702, 2024. 1, 5, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.382, + 0.482, + 0.449 + ], + "angle": 0, + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.453, + 0.482, + 0.507 + ], + "angle": 0, + "content": "[37] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, pages 8821-8831. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.551 + ], + "angle": 0, + "content": "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.555, + 0.482, + 0.609 + ], + "angle": 0, + "content": "[39] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.613, + 0.482, + 0.653 + ], + "angle": 0, + "content": "[40] Simo Ryu. Low-rank adaptation for fast text-to-image diffusion fine-tuning. https://github.com/cloneofsimo/lora, 2023.5,8,1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.482, + 0.696 + ], + "angle": 0, + "content": "[41] Lu Sheng, Ziyi Lin, Jing Shao, and Xiaogang Wang. Avatarnet: Multi-scale zero-shot style transfer by feature decoration. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.7, + 0.482, + 0.768 + ], + "angle": 0, + "content": "[42] Kihyuk Sohn, Lu Jiang, Jarred Barber, Kimin Lee, Nataniel Ruiz, Dilip Krishnan, Huiwen Chang, Yuzhzen Li, Irfan Essa, Michael Rubinstein, Yuan Hao, Glenn Entis, Irina Blok, and Daniel Castro Chin. Styledrop: Text-to-image synthesis of any style. In NeurIPS, 2023. 2, 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.772, + 0.482, + 0.799 + ], + "angle": 0, + "content": "[43] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2020. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[44] Adéla Šubrtová, Michal Lukáč, Jan Čech, David Futschik, Eli Shechtman, and Daniel Sykora. Diffusion image analogies. In SIGGRAPH, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[45] Yasheng Sun, Yifan Yang, Houwen Peng, Yifei Shen, Yuqing Yang, Han Hu, Lili Qiu, and Hideki Koike. Imagebrush: Learning visual in-context instructions for exemplar-based image manipulation. NeurlPS, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[46] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In CVPR, 2023. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[48] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. \\(p+\\): Extended textual conditioning in text-to-image generation. arXiv, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[49] Bram Wallace, Akash Gokul, and Nikhil Naik. Edict: Exact diffusion inversion via coupled transformations, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.251, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[50] Haofan Wang, Matteo Spinelli, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv preprint arXiv:2404.02733, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[51] Haofan Wang, Peng Xing, Renyuan Huang, Hao Ai, Qixun Wang, and Xu Bai. Instantstyle-plus: Style transfer with content-preserving in text-to-image generation. arXiv preprint arXiv:2407.00788, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[52] Zhouxia Wang, Xintao Wang, Liangbin Xie, Zhongang Qi, Ying Shan, Wenping Wang, and Ping Luo. Styleadapter: A single-pass lora-free model for stylized image generation. arXiv, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.421, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[53] Chen Henry Wu and Fernando De la Torre. A latent space of stochastic diffusion models for zero-shot image editing and guidance. In ICCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[54] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[55] Peng Xing, Haofan Wang, Yanpeng Sun, Qixun Wang, Xu Bai, Hao Ai, Renyuan Huang, and Zechao Li. Csgo: Content-style composition in text-to-image generation. arXiv preprint arXiv:2408.16766, 2024. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[56] Sihan Xu, Yidong Huang, Jiayi Pan, Ziqiao Ma, and Joyce Chai. Inversion-free image editing with natural language. arXiv, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[57] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. In SIGGRAPH Asia, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[58] Yuan Yao, Jianqiang Ren, Xuansong Xie, Weidong Liu, Yong-Jin Liu, and Jun Wang. Attention-aware multi-stroke style transfer, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.719, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[59] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[60] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[61] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.845 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15769" + } + ] +] \ No newline at end of file diff --git a/2025/StyleKeeper_ Prevent Content Leakage using Negative Visual Query Guidance/45fcd356-f226-4ed7-980b-85432724bf8a_origin.pdf b/2025/StyleKeeper_ Prevent Content Leakage using Negative Visual Query Guidance/45fcd356-f226-4ed7-980b-85432724bf8a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4fd22439b1a5a154735add9949af8d52f7658b2f --- /dev/null +++ b/2025/StyleKeeper_ Prevent Content Leakage using Negative Visual Query Guidance/45fcd356-f226-4ed7-980b-85432724bf8a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:23f9a7de78d4212ac42f83958a0d172925d15f8a5c3081d4734b52c594b76580 +size 14428376 diff --git a/2025/StyleKeeper_ Prevent Content Leakage using Negative Visual Query Guidance/full.md b/2025/StyleKeeper_ Prevent Content Leakage using Negative Visual Query Guidance/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8abcdf7a26aa802ccb5af70077ae49bc35fc04c3 --- /dev/null +++ b/2025/StyleKeeper_ Prevent Content Leakage using Negative Visual Query Guidance/full.md @@ -0,0 +1,321 @@ +# StyleKeeper: Prevent Content Leakage using Negative Visual Query Guidance + +Jaeseok Jeong\*† +Yonsei University +jete-jeong@yonsei.ac.kr + +Junho Kim* +NAVER AI Lab +jhkim.ai@navercorp.com + +Gayoung Lee +NAVER AI Lab +gayoung.lee@navercorp.com + +# Yunjey Choi +NAVER AI Lab + +yunjay.choi@gmail.com + +Youngjung Uh‡ Yonsei University yj.uh@yonsei.ac.ki + +# Abstract + +In the domain of text-to-image generation, diffusion models have emerged as powerful tools. Recently, studies on visual prompting, where images are used as prompts, have enabled more precise control over style and content. However, existing methods often suffer from content leakage, where undesired elements of the visual style prompt are transferred along with the intended style. To address this issue, we 1) extend classifier-free guidance (CFG) to utilize swapping self-attention and propose 2) negative visual query guidance (NVQG) to reduce the transfer of unwanted contents. NVQG employs negative score by intentionally simulating content leakage scenarios that swap queries instead of key and values of self-attention layers from visual style prompts. This simple yet effective method significantly reduces content leakage. Furthermore, we provide careful solutions for using a real image as visual style prompts. Through extensive evaluation across various styles and text prompts, our method demonstrates superiority over existing approaches, reflecting the style of the references, and ensuring that resulting images match the text prompts. + +# 1. Introduction + +Text-to-image diffusion models (T2I DMs) excel at synthesizing images that correspond to given text prompts [37, 38]. However, relying solely on text prompts may not allow precise control over the desired output. Even with highly detailed text prompts, controlling the exact style of the resulting images remains a challenge because of the modality gap between text and image. For example, from "a red car", different people imagine different variants of red. Instead, "A dinosaur" drawn in the style of the reference + +![](images/4484b300f1bbcb13260020d8d5c7810353cf61467e5f891e92bb3fabe0e52705.jpg) +Figure 1. Qualitative comparison with the same style. Competitors face challenges in generating images with diverse layouts and compositions, i.e., content leakage from the reference. + +image in Figure 1 provides a more concrete specification of the desired style. + +However, using reference images as visual style prompts comes with a fundamental challenge: a dilemma between better reflecting style and preventing content leakage. Figure 1 shows that the pose of the reference horse appears in the results of previous methods. The content leakage may include poses, specific shapes, layouts or entire objects. Furthermore, content leakage reduces diversity and text alignments as shown in Figure 1. + +Approaches that involve training the model, such as fine-tuning a diffusion model on a set of themed images [21, 39], learning new text embeddings [2, 10, 13], or adapting cross-attention modules to incorporate image features [35, 52, 55, 59], come with significant drawbacks. These methods not only incur high computational costs, but also face an inherent trade-off between style and content. Achieving a balance where the model faithfully captures the + +desired style without unintentionally preserving content remains a challenge. + +To address this, recent papers [1, 6, 16] have proposed training-free methods by modifying feature swapping mechanisms within the self-attention layers. Specifically, they replace the key and value features in the original process with those extracted from the visual style prompt. This allows the query to preserve content while the key-value pair carries the style, enabling promising style transfer performance. + +However, these methods struggle to fully eliminate content leakage [42]. Additionally, Hertz et al. [16] does not support real images as reference inputs, while Chung et al. [6] and Alaluf et al. [1] primarily focus on image-to-image (I2I) tasks. When using real images as style references, they often suffer from either poor style transfer or excessive content leakage, significantly limiting their effectiveness. + +In this study, we propose a method to separately control the strength of style and content from visual prompts such that the results align with the text prompts. We design a novel variation of classifier-free guidance (CFG) [17], emulating images with undesirable content leakage as a negative guidance. To address content leakage, we introduce negative visual query guidance, ensuring a clear separation between content and style. We also show that stochastic encoding achieves better style reflection with less artifacts and computation, and incorporate color calibration to precisely match the final output to the reference image's color statistics leading to better style reflection. + +We analyze where to apply swapping self-attention, identifying the optimal layers for balancing style transfer and content fidelity. Additionally, our method can effectively remove content that is difficult to eliminate with key and value swapping alone, working successfully even in cases with significant structural gaps between the style image and content text, as shown in Figure 1 (e.g., a complex scene of "a horse with the vibrant colors of a Chinese ink painting" and a single object "Dinosaur"). Our method is well compatible with ControlNet for I2I style transfer, further enhancing its flexibility. Qualitative and quantitative evaluations show our method outperforms state-of-the-art approaches, providing precise control over content and style without content leakage. Our approach is both robust and efficient, ideal for visual style prompting tasks. $^{1}$ + +# 2. Visual style prompting + +We propose StyleKeeper which receives a text prompt and a visual style prompt to generate new images without content leakage. The results contain the content and style specified by the text prompt and the visual style prompt, respectively, with variations due to initial noises. The overview + +of our method is illustrated in Figure 2. First, we explain the swapping self-attention in the aspect of style transfer literature. StyleKeeper consists of classifier-free guidance (CFG) with swapping self-attention, negative visual query guidance (NVQG), optimal layer choice, stochastic encoding of real visual style prompts, and generalization to ControlNet for real content images. We explain the first three components in text-to-image (T2I) scenario with generated visual style prompt. Then we proceed to T2I with real visual style prompt and image-to-image (I2I) scenario. + +# 2.1. Swapping self-attention in style transfer literature + +Modern diffusion models consist of a number of self-attention and cross-attention blocks [47]. Both of them employ the attention mechanism, which first obtains an attention map using similarity between query features $Q$ and key features $K$ , then aggregates value features $V$ using the attention map as weights: $Attention(Q, K, V) = \text{Softmax}(\frac{QK^T}{\sqrt{d}})V$ . Opposed to the cross-attention layer, self-attention layer receives key and values coming from the main denoising process, which has spatial dimensions with more freedom to represent spatially varying visual elements. As our goal is to reflect style elements from a reference image that are not easily represented by textual description, we opt to borrow key and values of self-attention layers in the reference process to the original process, namely swapping self-attention (Figure 2). In addition, swapping self-attention has a strong connection with style transfer literature [8, 25, 32, 41, 58]. where the attention mechanism reassembles visual features of a style image (key, value) on a content image (query). However, methods [3, 16] utilizing self-attention operation often exhibit artifacts and content leakage since style elements are entangled with content elements in the visual feature. + +# 2.2. CFG sampling with swapping self-attention for T2I + +Previous works [1, 6] have not taken CFG into account while manipulating features. We propose CFG combined with swapping self-attention to reflect a visual style prompt in the T2I results without artifacts. CFG [17] is a key technique that steers generated images toward specified text prompts. Given a score $\epsilon_{\theta}(x_t,c)$ conditioned on $c$ and an unconditional score $\epsilon_{\theta}(x_t,\emptyset)$ , the CFG score is defined as: + +$$ +\tilde {\epsilon} _ {\theta} = (1 + w) \epsilon_ {\theta} \left(x _ {t}, c\right) - w \epsilon_ {\theta} \left(x _ {t}, \emptyset\right), \tag {1} +$$ + +where $w$ controls guidance strength. CFG with $w > 1$ enhances image quality and text alignment, but excessive $w$ may lead to mode collapse [5]. Notably, CFG has not been + +![](images/0fadae4ffc93a3f1c143b083f110b1aff2a3b49c913dc2a0c7a46e3ba07ded46.jpg) +Figure 2. Overveiw of StyleKeeper. Our proposed method includes 4 key components, highlighted in red boxes. First, stochastic encoding (Section 3) converts reference images into suitable latents for the visual style prompting task. Second, swapping self-attention (Section 2.2, 2.4) ensures the reference image's style is accurately reflected. Third, negative visual query guidance (Section 2.3) reduces content leakage from the reference image, allowing the desired text content (e.g., "Moose") to be better represented. Lastly, color calibration (Section 3) minimizes errors during the denoising process, helping to produce a cleaner final image. + +explored in the context of reflecting denoising process with modified features. + +Assuming a given condition $c$ has hidden content $h^{\mathrm{content}}$ and style $h^{\mathrm{style}}$ , we model $p_{\theta}(x_0|h_{\mathrm{text}}^{\mathrm{content}},h_{\mathrm{visual}}^{\mathrm{style}})$ using: + +- The original T2I score: $\epsilon_{\theta}(x_t, c_{\text{text}})$ , leading to $x_0^{\text{ori}} \sim p_{\theta}(x_0 | c_{\text{text}}) = p_{\theta}(x_0 | h_{\text{text}}^{\text{content}}, h_{\text{text}}^{\text{style}})$ . +- The reference score: $\epsilon_{\theta}(x_t, c_{\mathrm{visual}})$ , leading to $x_0^{\mathrm{visual}} \sim p_{\theta}(x_0 | c_{\mathrm{visual}}) = p_{\theta}(x_0 | h_{\mathrm{visual}}^{\mathrm{content}}, h_{\mathrm{visual}}^{\mathrm{style}})$ . + +To ensure proper content and style control, we design the CFG score as: + +$$ +\begin{array}{l} \tilde {\epsilon} _ {\theta} \left(x _ {t}, h _ {\text {t e x t}} ^ {\text {c o n t e n t}}, h _ {\text {v i s u a l}} ^ {\text {s t y l e}}\right) = (1 + w) \\ \ddot {\epsilon} _ {\theta} \left(x _ {t}, Q _ {\text {t e x t}}, K V _ {\text {v i s u a l}}\right) \tag {2} \\ - w \epsilon_ {\theta} (x _ {t}, \emptyset). \\ \end{array} +$$ + +where $\ddot{\epsilon}_{\theta}(x_{t},Q_{\mathrm{text}},KV_{\mathrm{visual}})$ represents a KV-injected denoising score, indicating feature manipulation rather than direct conditioning. + +For given two denoising processes, one as original and another as a reference, borrowing the key-value in self-attention from the reference to the original process, i.e., key-value (KV) injection, tends to produces results with the content from the original process and the style from the reference process with limited control [1, 6, 56]. + +We define the KV-injected score by $\ddot{\epsilon}_{\theta}(x_{t},Q_{\mathrm{text}},KV_{\mathrm{visual}})$ where $Q_{\mathrm{text}}$ and $KV_{\mathrm{visual}}$ denote the query from the original score $\epsilon_{\theta}(x_t,c_{\mathrm{text}})$ and the + +key, value from the reference score $\epsilon_{\theta}(x_t,c_{\mathrm{visual}})$ . Then KV-injected-Attention becomes: + +$$ +\operatorname {A t t e n t i o n} \left(Q _ {\text {t e x t}}, K _ {\text {v i s u a l}}, V _ {\text {v i s u a l}}\right) = \operatorname {S o f t m a x} \left(\frac {Q _ {\text {t e x t}} K _ {\text {v i s u a l}} ^ {\intercal}}{\sqrt {d}}\right) V _ {\text {v i s u a l}}. \tag {3} +$$ + +We omit the layer index for brevity. + +Naive denoising process with $\ddot{e}_{\theta}(x_{t},Q_{\mathrm{text}},KV_{\mathrm{visual}})$ provides limited control in generating images with content and style specified by a text $c_{\mathrm{text}}$ and a visual style prompt, respectively. Moving forward, our CFG with swapping self-attention in Eq. (2) enjoys higher image quality and more accurate text alignment than the naive denoising process as in the original CFG for T2I generation. The results are deferred to Section 4.1. However, content leakage still hinders the creation of diverse images, and disrupts text alignment. + +# 2.3. Negative visual query guidance + +We propose negative visual query guidance (NVQG) to mitigate content leakage by reducing the content $h_{\mathrm{visual}}^{\mathrm{content}}$ from the visual style prompt appearing in the results. Briefly, NVQG negates the CFG of a score $\ddot{\epsilon}(x_t, Q_{\mathrm{visual}}, KV_{\mathrm{text}})$ . + +In Liu et al. [24], a complex text condition $c$ is factorized as a set of conditions $\{c_0, c_1, \ldots\}$ and Bayes' rule induces $p_{\theta}(x_t | c_0, c_1, \ldots) \propto \Pi \frac{p_{\theta}(x_t | c_i)}{p_{\theta}(x_t)}$ . Then, the score of the complex text condition $c$ becomes $\epsilon_{\theta}(x_t, c) = \epsilon_{\theta}(x_t, \emptyset) + \Pi (\epsilon_{\theta}(x_t, c_i) - \epsilon_{\theta}(x_t, \emptyset))$ . It allows reducing a specific concept $\tilde{c}$ with composition by negating its guidance with scale + +![](images/8ea22fbcecac27f5c3edf52615eb626e68a3a16748453ac49dfcae68d79ac58d.jpg) +Figure 3. The effect of swapping self-attention across different blocks. Swapping self-attention on the bottleneck and downblocks causes content leakage, resulting in cat-like images despite a dog prompt, while swapping on downblocks disrupts resulting images. We only apply swapping self-attention in the upblocks to reflect style elements effectively. + +$w_{\mathrm{neg}}$ + +$$ +\epsilon_ {\theta} \left(x _ {t}, c, \text {n o t} \tilde {c}\right) = \epsilon_ {\theta} \left(x _ {t}, c\right) - w _ {\text {n e g}} \left(\epsilon_ {\theta} \left(x _ {t}, \tilde {c}\right) - \epsilon_ {\theta} \left(x _ {t}, \emptyset\right)\right) \tag {4} +$$ + +Although we design $\ddot{\epsilon}_{\theta}(x_{t},Q_{\mathrm{text}},KV_{\mathrm{visual}})$ to predict the score toward $p_{\theta}(x_0|h_{\mathrm{text}}^{\mathrm{content}},h_{\mathrm{visual}}^{\mathrm{style}})$ , $\ddot{\epsilon} (x_{t},Q_{\mathrm{text}},KV_{\mathrm{visual}})$ still contain $h_\mathrm{visual}$ . Assuming a hidden factorization $KV_{\mathrm{visual}} = \{KV_{\mathrm{visual}}^{\mathrm{content}},KV_{\mathrm{visual}}^{\mathrm{style}}\}$ , Bayes' rule induces + +$$ +\begin{array}{l} p _ {\theta} \left(x _ {t} \mid Q _ {\text {t e x t}}, K V _ {\text {v i s u a l}} ^ {\text {s t y l e}}, K V _ {\text {v i s u a l}} ^ {\text {c o n t e n t}}\right) \propto p _ {\theta} \left(x _ {t}\right) \frac {p _ {\theta} \left(x _ {t} \mid Q _ {\text {t e x t}} , K V _ {\text {v i s u a l}} ^ {\text {s t y l e}}\right)}{p _ {\theta} \left(x _ {t}\right)} \\ \times \frac {p _ {\theta} \left(x _ {t} \mid Q _ {\emptyset} , K V _ {\text {v i s u a l}} ^ {\text {c o n t e n t}}\right)}{p _ {\theta} \left(x _ {t}\right)}. \tag {5} \\ \end{array} +$$ + +Since $\epsilon_{\theta}(x_t) = \epsilon_{\theta}(x_t, Q_{\emptyset}, KV_{\emptyset})$ , we obtain the desired conditional score of $\hat{c} = \{Q_{\mathrm{text}}, KV_{\mathrm{visual}}^{\mathrm{style}}\}$ : + +$$ +\begin{array}{l} \epsilon_ {\theta} \left(x _ {t}, \hat {c}\right) \leftarrow w _ {\text {v i s u a l}} \left(\ddot {\epsilon} _ {\theta} \left(x _ {t}, Q _ {\text {t e x t}}, K V _ {\text {v i s u a l}}\right) - \epsilon_ {\theta} \left(x _ {t}\right)\right) \\ - w _ {\text {c o n t e n t}} \left(\ddot {\epsilon} _ {\theta} \left(x _ {t}, Q _ {\emptyset}, K V _ {\text {v i s u a l}} ^ {\text {c o n t e n t}}\right) - \epsilon_ {\theta} \left(x _ {t}\right)\right) \\ + \epsilon_ {\theta} \left(x _ {t}\right). \tag {6} \\ \end{array} +$$ + +where $w_{\mathrm{visual}}$ and $w_{\mathrm{content}}$ regulate the strength of each guidance. By leveraging query injection [1, 6, 46, 56], we approximate $\ddot{\epsilon}_{\theta}(x_t, Q_{\emptyset}, KV_{\mathrm{visual}}^{\mathrm{content}}) \approx \ddot{\epsilon}_{\theta}(x_t, Q_{\mathrm{visual}}, KV_{\emptyset})$ . Finally, inserting (6) into (2), and empirically replacing $\epsilon_{\theta}(x_t)$ with $\ddot{\epsilon}\theta(x_t, Q_{\mathrm{visual}})$ results in similar performance. This allows a simplified reformulation of diffusion sampling using $w' = w_{\mathrm{visual}}(w + 1)$ : Since we can highly relate the $w' \ddot{\epsilon}_{\theta}(x_t, Q_{\mathrm{visual}})$ to the concept negation in (4) which guides the negation of concept with a scale $w_{\mathrm{neg}}$ , we named the query term as negative visual query guidance. + +# 2.4. Choosing blocks for swapping self-attention + +In order to maximize the effect of NVQG in preventing content leakage, we explore the depth of the self-attention + +![](images/f3312850a2d60f836f692e9f0785c66f700290c48ffd93a6f5b42094dc45ad80.jpg) +Figure 4. The effect of CFG and the proposed negative visual query guidance on image generation. The reference images provide the style for each generated output. Without NVQG, content leakage occurs, and the generated images fail to fully capture the intended content. In contrast, using NVQG ensures better alignment with both the reference style and the "Cat" prompt, reducing content distortion and improving quality. + +blocks to be swapped in the sense of granularity of visual elements. Modern architecture of diffusion models roughly consists of three sections in a sequence: downblocks, bottleneck blocks, and upblocks. Given that the bottleneck of diffusion models contains content elements of the image [20, 22, 33], we opt not to apply swapping self-attention to bottleneck blocks to prevent transferring contents in a reference image. Figure 3 shows that not swapping the bottleneck blocks prevents content leaking from the reference image. However, the synthesized images show disrupted results with seriously scattered objects. Furthermore, while swapping self-attention implements the reassembling operation, simply applying to all self-attention layers exposes a content leakage problem, where the content of the reference image influences the resulting image, as shown in the first row of Figure 3. I.e., the results contain cats even though the prompts specify "a dog". We conjecture that this phenomenon happens because feature maps of downblocks have unclear content layout [3, 27], so substituting features based on this inaccurate layout causes the disrupted results. To avoid injecting unnecessary features, we choose to swap the key and value of self-attention only in upblocks. + +We note that Hertz et al. [16] applies self-attention operation at all blocks and suffers content leakage. The last row of Figure 3 shows the success of our strategy. Following this, we proceed to search for a specific optimal layer in the upblock in Section 4.2 + +# 3. Real images as references + +So far, we have assumed a generated visual style prompt $x_0^{\mathrm{visual}} \sim p_\theta(x_0 | c_{\mathrm{visual}}}$ . Here, we allow real visual style + +![](images/742e55a6753a13dceac66efd180e4fa851ff786a6abde0262353f7cdf9f6b28f.jpg) +Figure 5. Analysis on the optimal range of upblocks for swapping self-attention. We find the optimal range of upblocks for a balanced trade-off between different aspects. The images on the right illustrate the visual results for different upblock layer indices, with the red cross indicating poor diversity and misalignment to the text prompt, the red triangle indicating a lack of style similarity, and the yellow star indicating the optimal results. Please refer to Section 4.2 for details. + +prompts by 1) stochastic encoding and 2) color calibration. + +We propose stochastic encoding to obtain $x_{t}^{\mathrm{visual}} \sim q(x_{t}|x_{0}^{\mathrm{visual}})$ by adding a random noise on $x_{0}^{\mathrm{visual}}$ following the forward process of DMs [18]: + +$$ +\epsilon_ {t} \sim \mathcal {N} (0, I), x _ {t} ^ {\text {v i s u a l}} = \sqrt {\alpha_ {t}} \cdot x _ {0} ^ {\text {v i s u a l}} + \sqrt {1 - \alpha_ {t}} \cdot \epsilon_ {t} \tag {7} +$$ + +At each timestep, we sample $\epsilon_{t}$ to encode $x_{t}^{\mathrm{visual}}$ . It ensures that $x_{t}^{\mathrm{visual}}$ lies on the learned trajectory and does not suffer from accumulative numerical error due to iterative process of DDIM inversion used by previous methods [6, 16] as shown in Figure A1. Furthermore, it does not need to store the intermediate latents as opposed to DDPM inversion used by Alaluf et al. [1]. + +Although stochastic encoding performs better than DDIM inversion, a subtle color discrepancy occurs between + +the resulting images and the visual style prompts. We introduce color calibration at $x_{t}^{\mathrm{ori}}$ in the original process to match the statistics of predicted $x_{0}^{\mathrm{ori}}$ to predicted $x_{0}^{\mathrm{visual}}$ . In Gatys [12], distance between channel-wise statistics is employed as a style loss for style transfer. In Song et al. [43], predicted $x_{0} \left( = \frac{x_{t}^{\mathrm{visual}} - \sqrt{1 - \alpha_{t}} \cdot \epsilon_{\theta}(\hat{x}_{t})}{\sqrt{\alpha_{t}}} \right)$ allows to estimate $x_{0}$ with high probability at intermediate timesteps using deterministic sampling. Inheriting the advantage, we execute adain operation to match mean&std of predicted $x_{0}^{\mathrm{ori}}$ with those of $x_{0}^{\mathrm{visual}}$ . It allows precise color calibration rather than directly matching $x_{t}^{\mathrm{ori}}$ to $x_{t}^{\mathrm{visual}}$ in Alaluf et al. [1], Chung et al. [6]. Furthermore, our approach differs from Chung et al. [6] in that using predicted $x_{0}$ at intermediate timestep $t \in (0,T)$ other than $x_{T}$ by reducing cumulative sampling error after the operation. Furthermore, Chung et al. [6] executes AdaIN at timestep T, inducing lengthy cumulative error. + +We provide supportive experiments that show the effectiveness of the proposed method in Section 4.4 and a detailed Algorithm in Appendix 3. + +# 4. Experiments + +In this section, we describe the effects of our proposed methods: CFG with swapping self-attention, Negative visual query guidance (NVQG), stochastic encoding, and color calibration. For swapping self-attention, we provide a detailed analysis through experiments to determine the optimal layers for swapping. The impact of NVQG is demonstrated through qualitative results. Additionally, we show why stochastic encoding outperforms DDIM inversion when inverting real images, and we highlight the benefits of color calibration through experimental results. + +We also conducted both quantitative and qualitative comparisons of our method against various competitors, including StyleAligned [16], IP-Adapter [59], Dreambooth-LoRA [39, 40], StyleDrop [42], DEADiff [35], CSGO [55] and InstantStyle [50, 51]. The details of these comparisons, along with the experimental setup and metrics, are described in the Appendix 2. + +# 4.1. Effectiveness of CFG and NVQG + +This section analyzes the effects of Classifier-Free Guidance (CFG) and Negative Visual Query Guidance (NVQG) on image generation, with a focus on text alignment and content leakage. Figure 4 shows the results of the three configurations. In the $1^{\text{st}}$ row, without CFG and NVQG, the generated images suffer from severe artifacts. The absence of CFG causes poor image quality resulting in significant misalignment with the prompt. In the $2^{\text{nd}}$ row, CFG with swapping self-attention improves the text misalignment by boosting image quality. Here, the "cat" in target text prompt becomes clearer in the generated images. However, content leakage from the reference image remains where unwanted + +![](images/7db0b67b31668fdea91bc83886ab0a678a0195591dcc4347a54a5a0cb310cc17.jpg) +Figure 6. Attention map visualization over late and early upblock layers. The late upblock better focuses on the style-corresponding region than the early upblock, leading to more freedom to reassemble small parts. The early upblock attends a larger region leading to content leakage. + +elements (layouts, structure, and composition) from the reference image affect the results. + +In the $3^{\mathrm{rd}}$ row, NVQG releases the content leakage and produces the best results closely matching the text prompt while reflecting style reference. + +Overall, Figure 4 demonstrates the critical role of NVQG in reducing content leakage, enjoying the quality boosting of CFG. Together, they ensure that the generated images align to both the target text prompt and the visual style prompts, resulting in high-quality, coherent outputs. + +We qualitatively showcase diversity of results within a text prompt in Figure A9 and text alignment with complex text prompts in Figure A22. + +# 4.2. Analysis for Swapping Self-Attention + +Optimal layers in upblocks Since recent large T2I DMs consist of many layers, we further analyze the behavior by changing the start of the swapping while the end of the swapping is fixed at the end. We use four key metrics as shown in Figure 5, there is a layer where all four metrics abruptly change (red line). Notably, this point remains consistent regardless of the reference image. All the results in the paper use only one fixed optimal layer. + +We choose this layer as the optimal start of the swapping for a balanced trade-off among all aspects. We provide qualitative results with detailed split of layers in Figure A3 and A4. + +Visualizing Attention maps Figure 6 compares average attention maps from the late upblock and the early upblock applying swapping self-attention. Using late upblock has more freedom to reassemble the reference style elements leading to more doggy results than early upblock which produces some cat-like attributes. The right two columns visualize the attention weight of query points marked as red stars and yellow dots. Swapping self-attention on late upblock reassemblies features from a style correspondence, + +![](images/a02de94a8bac9e60c8baf74bb9bf83e37f85cf6f6611f760605ce6d222548fd3.jpg) +(a) Synthesizing the content specified by the text prompts + +![](images/294659934ab125c51d9e03757c7be199ff88ee10e90372955153aa97745e125c.jpg) +(b) Synthesizing the pose specified by the text prompt + +![](images/1345f86ee2b06b053f2c2d4e887274eb2a4922b71a756c564fb4d704dd3df9d7.jpg) +(c) Synthesizing a single object specified by the text prompt +Figure 7. Comparison for content leakage. While StyleAligned suffers from content leakage from the reference, our results clearly align with the text prompts + +e.g., texture and color. On the other hand, swapping self-attention on early upblock reassembles features from a wider area with different styles. This comparison clarifies the reasons for using only late upblock. Please refer to Figure A6, A7 for a more visualization and a detailed analysis. + +# 4.3. Comparison against competitors + +Style & content control We provide a qualitative comparison in Figure 8, focusing on controlling style and content. Our method faithfully synthesizes content from the text prompt with the style of the reference image. In contrast, other methods add elements like color or texture not in the reference (e.g., feathers, bricks, iron, skin) and often suffer from content leakage (e.g., layout, screaming person, castle), which compromises text prompt faithfulness. Quantitative results in Figure 9 support these findings: IP-Adapter shows higher style similarity but neglects text prompts significantly. + +Diversity within a text specification Starting from different initial noises, the diffusion models trained on a large dataset produce diverse results within the specification of a text prompt. Figure 1, A19 shows that ours can synthesize various poses and viewpoints while others barely change, i.e., other methods limit the diversity of the pretrained model due to content leakage. We further support our diversity in Figure A9. + +![](images/16642bdb6d4b6d938e40598a25670587e9b25d0d5d8af6b3303ae87ea9408cc0.jpg) +Figure 8. Qualitative comparison across various styles and text prompts. StyleKeeper faithfully reflects style elements in reference images without causing content leakage from the reference images. + +![](images/ec3c6df998fac707c65c563de5a5b534b103beac9e4305e1a01f54a42aea5858.jpg) + +![](images/d6b0e853896fd98ea49de2ea556527d55f664585c2779fd4e027add15c533bf2.jpg) +Figure 9. Quantitative comparison. Ours achieves the best style similarity while not compromising text alignment and being free from content leakage. + +Content leakage Content leakage refers to the phenomenon where the content of a reference image appears in a result. As evident in Figure A12, our model exhibits significantly less content leakage when compared to other models. We focus on a comparison against a runner-up method, StyleAligned, particularly in terms of how content leakage can be an obstacle to controlling the content using a text prompt. Figure 7a compares examples where strong content leakage prevents the object from the text from appearing in StyleAligned. We often observe famous paintings in the reference easily leak into the results of StyleAligned while ours does not struggle. Figure 7b compares examples where the pose in the reference leaks into the results of StyleAligned preventing the reflection of specified pose in the text prompt. On the other hand, our method reflects the correct poses. Figure 7c compares examples where the number of small instances in the reference leaks into the result of StyleAligned. Contrarily, our method correctly synthesizes a single instance of the content specified by the text. In Figure 9, we provide quantitative comparison results. It supports that StyleKeeper achieves the best style + +similarity while not suffering content leakage. We can also observe that models with high content leakage exhibit poor text alignment. + +Preserving the content of the original denoising process Figure A13 shows the results of ours and other methods using the same initial noise in each column. Our method precisely reflects the style in the reference with minimal changes of contents in the original denoising process. On the other hand, the other methods severely alter pose, shape, or layouts due to content leakage. It is an important virtue of controlling style to keep the rest intact. This also supports that ours is free from content leakage during the denoising process. We provide more results in Figure A14. + +# 4.4. DDIM inversion vs. Our strategy + +Figure 11 shows that StyleKeeper can take real images as style reference with our strategies. As shown in Figure 11, stochastic encoding outperforms DDIM inversion and color calibration improves text alignment and style similarity. We provide more results in Figure A15. Moreover, we show that our strategy can boost the performance of the other self-attention variants [3, 16] in Figure A16, A17 and color calibration can be used for generation settings in Figure A18. We provide an ablation study in Figure 12 for each configuration (swapping self-attention, NVQG, and color calibration) in both real reference and generated reference settings. In both settings, swapping self-attention and NVQG improve style similarity and text alignment, while color calibration helps improve style similarity. Additionally, stochastic encoding outperforms DDIM inversion. + +![](images/3bbc89c10d184627277529b4f7860917d936649cb65c4c1d43b37cd9d94d73e3.jpg) +Figure 10. Qualitative comparison in I2I style transfer task. We compare our method for I2I style transfer task where the content image is given. Compared to the previous methods, ours achieves the best style similarity. + +![](images/225153d76822477e8f057c560c9e7bcce824ff3119ab63fcd62ba3b42f4b1080.jpg) +Figure 11. Comparison of DDIM inversion vs. stochastic encoding and the effect of color calibration. Stochastic encoding reduces artifacts in the resulting images, while color calibration better reflects the colors of the reference image. + +![](images/a43cf8f9591b5a70d31b8a6bf4a3076f541ce48d11840ad333c486cecae3bc7e.jpg) +Figure 12. Ablation study of four proposed methods with a real reference and a generated reference setting.. We evaluated performance by individually removing each of the four proposed methods. In all cases, both generated and real images showed similar improvements in style similarity and text alignment. Moreover, stochastic encoding outperformed DDIM inversion in all aspects, comparing purple and blue points. + +# 4.5. Comparison for I2I style transfer + +In Figure 10, we present a qualitative comparison between our method using ControlNet and existing state-of-the-art methods, CrossAttn [1] and StyleID [6], for the I2I style transfer task where a content image is provided. Both CrossAttn and StyleID inject the key & value obtained by inversion from the style images. + +CrossAttn often fails to transfer the detailed representa + +tion of the reference style, leading to a rough, blocky appearance. For instance, the center examples in Figure 10 show that the style details are absent or inaccurate compared to our results. Similarly, StyleID suffers poor style reflection and color deficiency in the center and the rightmost example. In contrast, ours effectively reflects the details of the reference style image, with no noticeable transfer of the color values from the content images. + +# 5. Conclusion and Limitation + +In this paper, we introduce StyleKeeper which faithfully applies the style of real reference images without content leakage in a training-free manner. CFG with swapping self-attention captures the reference image's style accurately and allows for direct content generation from the text. This integration with CFG enhances performance by balancing content generation and style transfer. To address content leakage, we propose Negative visual query guidance that ensures the reference image's content does not interfere with the text-specified content, enhancing diversity and improving text alignment & style similarity. Additionally, Stochastic encoding efficiently maps real images to suitable latents, removing artifacts. Color calibration aligns the final output to the reference's color statistics. + +Our method shows both qualitative and quantitative improvements over existing approaches, providing a robust solution for visual style prompting without complex training. We also analyze where to apply swapping self-attention, identifying optimal layers to balance style transfer and content fidelity. StyleKeeper outperforms prior methods and works well with algorithms like ControlNet [60] and DBLoRA [40], as shown in Figure A20. + +However, StyleKeeper is limited by the capabilities of the pretrained diffusion models and cannot generate elements outside their scope (e.g., "stone golem" in Figure A24a). Also, if visual and textual styles conflict, the visual style dominates (Figure A24b). Future work could explore extending our method to other domains such as video, expanding its applicability and research potential. + +Acknowledgements We thank Jin-Hwa Kim for helpful comments. This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS-2024-00439762, Developing Techniques for Analyzing and Assessing Vulnerabilities, and Tools for Confidentiality Evaluation in Generative AI Models). + +# References + +[1] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer. In ACM SIGGRAPH 2024 Conference Papers, pages 1–12, 2024. 2, 3, 4, 5, 8, 1 +[2] Omri Avrahami, Kfir Aberman, Ohad Fried, Daniel Cohen-Or, and Dani Lischinski. Break-a-scene: Extracting multiple concepts from a single image. In SIGGRAPH Asia, 2023. 1 +[3] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In ICCV), 2023. 2, 4, 7, 1, 9, 15 +[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 1 +[5] Hyungjin Chung, Jeongsol Kim, Geon Yeong Park, Hyelin Nam, and Jong Chul Ye. Cfg++: Manifold-constrained classifier free guidance for diffusion models. arXiv preprint arXiv:2406.08070, 2024. 2 +[6] Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8795-8805, 2024. 2, 3, 4, 5, 8, 1 +[7] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance, 2022. 1 +[8] Yingying Deng, Fan Tang, Weiming Dong, Chongyang Ma, Xingjia Pan, Lei Wang, and Changsheng Xu. Stytr2: Image style transfer with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11326-11336, 2022. 2 +[9] Martin Nicolas Everaert, Marco Bocchio, Sami Arpa, Sabine Süssstrunk, and Radhakrishna Achanta. Diffusion in style. In ICCV, 2023. 1 +[10] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2022. 1 +[11] Daniel Garibi, Or Patashnik, Andrey Voynov, Hadar Averbuch-Elor, and Daniel Cohen-Or. Renoise: Real image inversion through iterative noising, 2024. 1 +[12] Leon A Gatys. A neural algorithm of artistic style. arXiv preprint arXiv:1508.06576, 2015. 5, 1 +[13] Inhwa Han, Serin Yang, Taesung Kwon, and Jong Chul Ye. Highly personalized text embedding for image manipulation by stable diffusion. arXiv, 2023. 1 + +[14] Ligong Han, Song Wen, Qi Chen, Zhixing Zhang, Kunpeng Song, Mengwei Ren, Ruijiang Gao, Anastasis Stathopoulos, Xiaoxiao He, Yuxiao Chen, Di Liu, Qilong Zhangli, Jindong Jiang, Zhaoyang Xia, Akash Srivastava, and Dimitris Metaxas. Improving tuning-free real image editing with proximal guidance, 2023. 1 +[15] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control, 2022. 1 +[16] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. arXiv, 2023. 2, 4, 5, 7, 1, 15 +[17] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2 +[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 5, 1 +[19] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations, 2024. 1 +[20] Jaeseok Jeong, Mingi Kwon, and Youngjung Uh. Training-free content injection using h-space in diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5151-5161, 2024. 4 +[21] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In CVPR, 2023. 1 +[22] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. In ICLR, 2023. 4 +[23] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In CVPR, 2023. 1 +[24] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, pages 423-439. Springer, 2022. 3 +[25] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, and Errui Ding. Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In ICCV, 2021. 2 +[26] Frank J. Massey. The kolmogorov-smirnov test for goodness of fit. Journal of the American Statistical Association, 46 (253):68-78, 1951. 2 +[27] Benyuan Meng, Qianqian Xu, Zitai Wang, Xiaochun Cao, and Qingming Huang. Not all diffusion model activations have been evaluated as discriminative features. arXiv preprint arXiv:2410.03558, 2024. 4 +[28] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 1 +[29] Daiki Miyake, Akihiro Iohara, Yu Saito, and Toshiyuki Tanaka. Negative-prompt inversion: Fast image inversion for editing with text-guided diffusion models, 2023. 1 +[30] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, 2023. 1 + +[31] Thao Nguyen, Yuheng Li, Utkarsh Ojha, and Yong Jae Lee. Visual instruction inversion: Image editing via visual prompting. NeurIPS, 2023. 1 +[32] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In CVPR, 2019. 2 +[33] Yong-Hyun Park, Mingi Kwon, Jaewoong Choi, Junghyo Jo, and Youngjung Uh. Understanding the latent space of diffusion models through the lens of riemannian geometry. Advances in Neural Information Processing Systems, 36: 24129-24142, 2023. 4 +[34] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv, 2023. 1 +[35] Tianhao Qi, Shancheng Fang, Yanze Wu, Hongtao Xie, Jiawei Liu, Lang Chen, Qian He, and Yongdong Zhang. Deadiff: An efficient stylization diffusion model with disentangled representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8693-8702, 2024. 1, 5, 16 +[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2 +[37] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, pages 8821-8831. PMLR, 2021. 1 +[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1 +[39] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 1, 5 +[40] Simo Ryu. Low-rank adaptation for fast text-to-image diffusion fine-tuning. https://github.com/cloneofsimo/lora, 2023.5,8,1 +[41] Lu Sheng, Ziyi Lin, Jing Shao, and Xiaogang Wang. Avatarnet: Multi-scale zero-shot style transfer by feature decoration. In CVPR, 2018. 2 +[42] Kihyuk Sohn, Lu Jiang, Jarred Barber, Kimin Lee, Nataniel Ruiz, Dilip Krishnan, Huiwen Chang, Yuzhzen Li, Irfan Essa, Michael Rubinstein, Yuan Hao, Glenn Entis, Irina Blok, and Daniel Castro Chin. Styledrop: Text-to-image synthesis of any style. In NeurIPS, 2023. 2, 5, 1 +[43] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2020. 5, 1 +[44] Adéla Šubrtová, Michal Lukáč, Jan Čech, David Futschik, Eli Shechtman, and Daniel Sykora. Diffusion image analogies. In SIGGRAPH, 2023. 1 +[45] Yasheng Sun, Yifan Yang, Houwen Peng, Yifei Shen, Yuqing Yang, Han Hu, Lili Qiu, and Hideki Koike. Imagebrush: Learning visual in-context instructions for exemplar-based image manipulation. NeurlPS, 2023. 1 + +[46] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In CVPR, 2023. 4, 1 +[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 2 +[48] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. $p+$ : Extended textual conditioning in text-to-image generation. arXiv, 2023. 1 +[49] Bram Wallace, Akash Gokul, and Nikhil Naik. Edict: Exact diffusion inversion via coupled transformations, 2022. 1 +[50] Haofan Wang, Matteo Spinelli, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv preprint arXiv:2404.02733, 2024. 5 +[51] Haofan Wang, Peng Xing, Renyuan Huang, Hao Ai, Qixun Wang, and Xu Bai. Instantstyle-plus: Style transfer with content-preserving in text-to-image generation. arXiv preprint arXiv:2407.00788, 2024. 5 +[52] Zhouxia Wang, Xintao Wang, Liangbin Xie, Zhongang Qi, Ying Shan, Wenping Wang, and Ping Luo. Styleadapter: A single-pass lora-free model for stylized image generation. arXiv, 2023. 1 +[53] Chen Henry Wu and Fernando De la Torre. A latent space of stochastic diffusion models for zero-shot image editing and guidance. In ICCV, 2023. 1 +[54] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, 2023. 1 +[55] Peng Xing, Haofan Wang, Yanpeng Sun, Qixun Wang, Xu Bai, Hao Ai, Renyuan Huang, and Zechao Li. Csgo: Content-style composition in text-to-image generation. arXiv preprint arXiv:2408.16766, 2024. 1, 5 +[56] Sihan Xu, Yidong Huang, Jiayi Pan, Ziqiao Ma, and Joyce Chai. Inversion-free image editing with natural language. arXiv, 2023. 3, 4 +[57] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. In SIGGRAPH Asia, 2023. 1 +[58] Yuan Yao, Jianqiang Ren, Xuansong Xie, Weidong Liu, Yong-Jin Liu, and Jun Wang. Attention-aware multi-stroke style transfer, 2019. 2 +[59] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv, 2023. 1, 5 +[60] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 8, 1 +[61] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 2 \ No newline at end of file diff --git a/2025/StyleKeeper_ Prevent Content Leakage using Negative Visual Query Guidance/images.zip b/2025/StyleKeeper_ Prevent Content Leakage using Negative Visual Query Guidance/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c54adb76e4ff494293d1060aa37af6f54ea40b92 --- /dev/null +++ b/2025/StyleKeeper_ Prevent Content Leakage using Negative Visual Query Guidance/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:612b2f2fe958456aefffc78af35a845f7e1bd53eb877d5a6d330b28f1d6e961b +size 898547 diff --git a/2025/StyleKeeper_ Prevent Content Leakage using Negative Visual Query Guidance/layout.json b/2025/StyleKeeper_ Prevent Content Leakage using Negative Visual Query Guidance/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0b8f01f0eb6991c03598f852567b0272c0daaabe --- /dev/null +++ b/2025/StyleKeeper_ Prevent Content Leakage using Negative Visual Query Guidance/layout.json @@ -0,0 +1,8638 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 63, + 103, + 547, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 103, + 547, + 121 + ], + "spans": [ + { + "bbox": [ + 63, + 103, + 547, + 121 + ], + "type": "text", + "content": "StyleKeeper: Prevent Content Leakage using Negative Visual Query Guidance" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 143, + 213, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 143, + 213, + 185 + ], + "spans": [ + { + "bbox": [ + 88, + 143, + 213, + 185 + ], + "type": "text", + "content": "Jaeseok Jeong\\*† \nYonsei University \njete-jeong@yonsei.ac.kr" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 239, + 144, + 359, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 144, + 359, + 185 + ], + "spans": [ + { + "bbox": [ + 239, + 144, + 359, + 185 + ], + "type": "text", + "content": "Junho Kim* \nNAVER AI Lab \njhkim.ai@navercorp.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 385, + 144, + 522, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 144, + 522, + 185 + ], + "spans": [ + { + "bbox": [ + 385, + 144, + 522, + 185 + ], + "type": "text", + "content": "Gayoung Lee \nNAVER AI Lab \ngayoung.lee@navercorp.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 193, + 191, + 273, + 218 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 191, + 273, + 218 + ], + "spans": [ + { + "bbox": [ + 193, + 191, + 273, + 218 + ], + "type": "text", + "content": "Yunjey Choi \nNAVER AI Lab" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 176, + 222, + 291, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 222, + 291, + 233 + ], + "spans": [ + { + "bbox": [ + 176, + 222, + 291, + 233 + ], + "type": "text", + "content": "yunjay.choi@gmail.com" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 332, + 191, + 427, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 191, + 427, + 233 + ], + "spans": [ + { + "bbox": [ + 332, + 191, + 427, + 233 + ], + "type": "text", + "content": "Youngjung Uh‡ Yonsei University yj.uh@yonsei.ac.ki" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 152, + 261, + 200, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 261, + 200, + 273 + ], + "spans": [ + { + "bbox": [ + 152, + 261, + 200, + 273 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 285, + 296, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 296, + 525 + ], + "type": "text", + "content": "In the domain of text-to-image generation, diffusion models have emerged as powerful tools. Recently, studies on visual prompting, where images are used as prompts, have enabled more precise control over style and content. However, existing methods often suffer from content leakage, where undesired elements of the visual style prompt are transferred along with the intended style. To address this issue, we 1) extend classifier-free guidance (CFG) to utilize swapping self-attention and propose 2) negative visual query guidance (NVQG) to reduce the transfer of unwanted contents. NVQG employs negative score by intentionally simulating content leakage scenarios that swap queries instead of key and values of self-attention layers from visual style prompts. This simple yet effective method significantly reduces content leakage. Furthermore, we provide careful solutions for using a real image as visual style prompts. Through extensive evaluation across various styles and text prompts, our method demonstrates superiority over existing approaches, reflecting the style of the references, and ensuring that resulting images match the text prompts." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 548, + 135, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 135, + 560 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 135, + 560 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 567, + 296, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 296, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 296, + 676 + ], + "type": "text", + "content": "Text-to-image diffusion models (T2I DMs) excel at synthesizing images that correspond to given text prompts [37, 38]. However, relying solely on text prompts may not allow precise control over the desired output. Even with highly detailed text prompts, controlling the exact style of the resulting images remains a challenge because of the modality gap between text and image. For example, from \"a red car\", different people imagine different variants of red. Instead, \"A dinosaur\" drawn in the style of the reference" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 316, + 259, + 555, + 430 + ], + "blocks": [ + { + "bbox": [ + 316, + 259, + 555, + 430 + ], + "lines": [ + { + "bbox": [ + 316, + 259, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 259, + 555, + 430 + ], + "type": "image", + "image_path": "4484b300f1bbcb13260020d8d5c7810353cf61467e5f891e92bb3fabe0e52705.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 439, + 555, + 472 + ], + "lines": [ + { + "bbox": [ + 313, + 439, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 555, + 472 + ], + "type": "text", + "content": "Figure 1. Qualitative comparison with the same style. Competitors face challenges in generating images with diverse layouts and compositions, i.e., content leakage from the reference." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 495, + 555, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 555, + 519 + ], + "type": "text", + "content": "image in Figure 1 provides a more concrete specification of the desired style." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 521, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 616 + ], + "type": "text", + "content": "However, using reference images as visual style prompts comes with a fundamental challenge: a dilemma between better reflecting style and preventing content leakage. Figure 1 shows that the pose of the reference horse appears in the results of previous methods. The content leakage may include poses, specific shapes, layouts or entire objects. Furthermore, content leakage reduces diversity and text alignments as shown in Figure 1." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "Approaches that involve training the model, such as fine-tuning a diffusion model on a set of themed images [21, 39], learning new text embeddings [2, 10, 13], or adapting cross-attention modules to incorporate image features [35, 52, 55, 59], come with significant drawbacks. These methods not only incur high computational costs, but also face an inherent trade-off between style and content. Achieving a balance where the model faithfully captures the" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 683, + 133, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 683, + 133, + 693 + ], + "spans": [ + { + "bbox": [ + 66, + 683, + 133, + 693 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 67, + 693, + 267, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 693, + 267, + 703 + ], + "spans": [ + { + "bbox": [ + 67, + 693, + 267, + 703 + ], + "type": "text", + "content": "This work was done during his internship at Naver AI Lab." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": " Corresponding author." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15760" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "desired style without unintentionally preserving content remains a challenge." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 97, + 295, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 97, + 295, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 97, + 295, + 191 + ], + "type": "text", + "content": "To address this, recent papers [1, 6, 16] have proposed training-free methods by modifying feature swapping mechanisms within the self-attention layers. Specifically, they replace the key and value features in the original process with those extracted from the visual style prompt. This allows the query to preserve content while the key-value pair carries the style, enabling promising style transfer performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 193, + 296, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 193, + 296, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 296, + 276 + ], + "type": "text", + "content": "However, these methods struggle to fully eliminate content leakage [42]. Additionally, Hertz et al. [16] does not support real images as reference inputs, while Chung et al. [6] and Alaluf et al. [1] primarily focus on image-to-image (I2I) tasks. When using real images as style references, they often suffer from either poor style transfer or excessive content leakage, significantly limiting their effectiveness." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 277, + 296, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 296, + 420 + ], + "type": "text", + "content": "In this study, we propose a method to separately control the strength of style and content from visual prompts such that the results align with the text prompts. We design a novel variation of classifier-free guidance (CFG) [17], emulating images with undesirable content leakage as a negative guidance. To address content leakage, we introduce negative visual query guidance, ensuring a clear separation between content and style. We also show that stochastic encoding achieves better style reflection with less artifacts and computation, and incorporate color calibration to precisely match the final output to the reference image's color statistics leading to better style reflection." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 422, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 601 + ], + "type": "text", + "content": "We analyze where to apply swapping self-attention, identifying the optimal layers for balancing style transfer and content fidelity. Additionally, our method can effectively remove content that is difficult to eliminate with key and value swapping alone, working successfully even in cases with significant structural gaps between the style image and content text, as shown in Figure 1 (e.g., a complex scene of \"a horse with the vibrant colors of a Chinese ink painting\" and a single object \"Dinosaur\"). Our method is well compatible with ControlNet for I2I style transfer, further enhancing its flexibility. Qualitative and quantitative evaluations show our method outperforms state-of-the-art approaches, providing precise control over content and style without content leakage. Our approach is both robust and efficient, ideal for visual style prompting tasks." + }, + { + "bbox": [ + 55, + 422, + 296, + 601 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 613, + 187, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 187, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 187, + 627 + ], + "type": "text", + "content": "2. Visual style prompting" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 633, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 295, + 693 + ], + "type": "text", + "content": "We propose StyleKeeper which receives a text prompt and a visual style prompt to generate new images without content leakage. The results contain the content and style specified by the text prompt and the visual style prompt, respectively, with variations due to initial noises. The overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": "of our method is illustrated in Figure 2. First, we explain the swapping self-attention in the aspect of style transfer literature. StyleKeeper consists of classifier-free guidance (CFG) with swapping self-attention, negative visual query guidance (NVQG), optimal layer choice, stochastic encoding of real visual style prompts, and generalization to ControlNet for real content images. We explain the first three components in text-to-image (T2I) scenario with generated visual style prompt. Then we proceed to T2I with real visual style prompt and image-to-image (I2I) scenario." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 200, + 555, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 200, + 555, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 200, + 555, + 224 + ], + "type": "text", + "content": "2.1. Swapping self-attention in style transfer literature" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 229, + 555, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 229, + 555, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 555, + 496 + ], + "type": "text", + "content": "Modern diffusion models consist of a number of self-attention and cross-attention blocks [47]. Both of them employ the attention mechanism, which first obtains an attention map using similarity between query features " + }, + { + "bbox": [ + 313, + 229, + 555, + 496 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 229, + 555, + 496 + ], + "type": "text", + "content": " and key features " + }, + { + "bbox": [ + 313, + 229, + 555, + 496 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 229, + 555, + 496 + ], + "type": "text", + "content": ", then aggregates value features " + }, + { + "bbox": [ + 313, + 229, + 555, + 496 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 229, + 555, + 496 + ], + "type": "text", + "content": " using the attention map as weights: " + }, + { + "bbox": [ + 313, + 229, + 555, + 496 + ], + "type": "inline_equation", + "content": "Attention(Q, K, V) = \\text{Softmax}(\\frac{QK^T}{\\sqrt{d}})V" + }, + { + "bbox": [ + 313, + 229, + 555, + 496 + ], + "type": "text", + "content": ". Opposed to the cross-attention layer, self-attention layer receives key and values coming from the main denoising process, which has spatial dimensions with more freedom to represent spatially varying visual elements. As our goal is to reflect style elements from a reference image that are not easily represented by textual description, we opt to borrow key and values of self-attention layers in the reference process to the original process, namely swapping self-attention (Figure 2). In addition, swapping self-attention has a strong connection with style transfer literature [8, 25, 32, 41, 58]. where the attention mechanism reassembles visual features of a style image (key, value) on a content image (query). However, methods [3, 16] utilizing self-attention operation often exhibit artifacts and content leakage since style elements are entangled with content elements in the visual feature." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 504, + 555, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 555, + 527 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 555, + 527 + ], + "type": "text", + "content": "2.2. CFG sampling with swapping self-attention for T2I" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 533, + 556, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 617 + ], + "type": "text", + "content": "Previous works [1, 6] have not taken CFG into account while manipulating features. We propose CFG combined with swapping self-attention to reflect a visual style prompt in the T2I results without artifacts. CFG [17] is a key technique that steers generated images toward specified text prompts. Given a score " + }, + { + "bbox": [ + 313, + 533, + 556, + 617 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,c)" + }, + { + "bbox": [ + 313, + 533, + 556, + 617 + ], + "type": "text", + "content": " conditioned on " + }, + { + "bbox": [ + 313, + 533, + 556, + 617 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 533, + 556, + 617 + ], + "type": "text", + "content": " and an unconditional score " + }, + { + "bbox": [ + 313, + 533, + 556, + 617 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,\\emptyset)" + }, + { + "bbox": [ + 313, + 533, + 556, + 617 + ], + "type": "text", + "content": ", the CFG score is defined as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 359, + 627, + 555, + 640 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 627, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 359, + 627, + 555, + 640 + ], + "type": "interline_equation", + "content": "\\tilde {\\epsilon} _ {\\theta} = (1 + w) \\epsilon_ {\\theta} \\left(x _ {t}, c\\right) - w \\epsilon_ {\\theta} \\left(x _ {t}, \\emptyset\\right), \\tag {1}", + "image_path": "d726258ae39a8893741277ef8a96d7d6322e74b32002873ee3a63f61eead8ae1.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 648, + 555, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 555, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 555, + 686 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 648, + 555, + 686 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 648, + 555, + 686 + ], + "type": "text", + "content": " controls guidance strength. CFG with " + }, + { + "bbox": [ + 313, + 648, + 555, + 686 + ], + "type": "inline_equation", + "content": "w > 1" + }, + { + "bbox": [ + 313, + 648, + 555, + 686 + ], + "type": "text", + "content": " enhances image quality and text alignment, but excessive " + }, + { + "bbox": [ + 313, + 648, + 555, + 686 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 648, + 555, + 686 + ], + "type": "text", + "content": " may lead to mode collapse [5]. Notably, CFG has not been" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "type": "text", + "content": "2We omit the diffusion timestep " + }, + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "type": "text", + "content": " in the arguments and abuse " + }, + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "type": "text", + "content": " from latent diffusion model." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 66, + 702, + 217, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 217, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 217, + 713 + ], + "type": "text", + "content": "1Our code will be released for reproducibility." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15761" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 84, + 72, + 527, + 273 + ], + "blocks": [ + { + "bbox": [ + 84, + 72, + 527, + 273 + ], + "lines": [ + { + "bbox": [ + 84, + 72, + 527, + 273 + ], + "spans": [ + { + "bbox": [ + 84, + 72, + 527, + 273 + ], + "type": "image", + "image_path": "0fadae4ffc93a3f1c143b083f110b1aff2a3b49c913dc2a0c7a46e3ba07ded46.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 278, + 555, + 336 + ], + "lines": [ + { + "bbox": [ + 54, + 278, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 54, + 278, + 555, + 336 + ], + "type": "text", + "content": "Figure 2. Overveiw of StyleKeeper. Our proposed method includes 4 key components, highlighted in red boxes. First, stochastic encoding (Section 3) converts reference images into suitable latents for the visual style prompting task. Second, swapping self-attention (Section 2.2, 2.4) ensures the reference image's style is accurately reflected. Third, negative visual query guidance (Section 2.3) reduces content leakage from the reference image, allowing the desired text content (e.g., \"Moose\") to be better represented. Lastly, color calibration (Section 3) minimizes errors during the denoising process, helping to produce a cleaner final image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 355, + 295, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 355, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 355, + 295, + 378 + ], + "type": "text", + "content": "explored in the context of reflecting denoising process with modified features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 379, + 295, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 295, + 404 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 295, + 404 + ], + "type": "text", + "content": "Assuming a given condition " + }, + { + "bbox": [ + 55, + 379, + 295, + 404 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 379, + 295, + 404 + ], + "type": "text", + "content": " has hidden content " + }, + { + "bbox": [ + 55, + 379, + 295, + 404 + ], + "type": "inline_equation", + "content": "h^{\\mathrm{content}}" + }, + { + "bbox": [ + 55, + 379, + 295, + 404 + ], + "type": "text", + "content": " and style " + }, + { + "bbox": [ + 55, + 379, + 295, + 404 + ], + "type": "inline_equation", + "content": "h^{\\mathrm{style}}" + }, + { + "bbox": [ + 55, + 379, + 295, + 404 + ], + "type": "text", + "content": ", we model " + }, + { + "bbox": [ + 55, + 379, + 295, + 404 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_0|h_{\\mathrm{text}}^{\\mathrm{content}},h_{\\mathrm{visual}}^{\\mathrm{style}})" + }, + { + "bbox": [ + 55, + 379, + 295, + 404 + ], + "type": "text", + "content": " using:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 406, + 295, + 458 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 406, + 295, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 406, + 295, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 295, + 432 + ], + "type": "text", + "content": "- The original T2I score: " + }, + { + "bbox": [ + 55, + 406, + 295, + 432 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t, c_{\\text{text}})" + }, + { + "bbox": [ + 55, + 406, + 295, + 432 + ], + "type": "text", + "content": ", leading to " + }, + { + "bbox": [ + 55, + 406, + 295, + 432 + ], + "type": "inline_equation", + "content": "x_0^{\\text{ori}} \\sim p_{\\theta}(x_0 | c_{\\text{text}}) = p_{\\theta}(x_0 | h_{\\text{text}}^{\\text{content}}, h_{\\text{text}}^{\\text{style}})" + }, + { + "bbox": [ + 55, + 406, + 295, + 432 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 432, + 295, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 295, + 458 + ], + "type": "text", + "content": "- The reference score: " + }, + { + "bbox": [ + 55, + 432, + 295, + 458 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t, c_{\\mathrm{visual}})" + }, + { + "bbox": [ + 55, + 432, + 295, + 458 + ], + "type": "text", + "content": ", leading to " + }, + { + "bbox": [ + 55, + 432, + 295, + 458 + ], + "type": "inline_equation", + "content": "x_0^{\\mathrm{visual}} \\sim p_{\\theta}(x_0 | c_{\\mathrm{visual}}) = p_{\\theta}(x_0 | h_{\\mathrm{visual}}^{\\mathrm{content}}, h_{\\mathrm{visual}}^{\\mathrm{style}})" + }, + { + "bbox": [ + 55, + 432, + 295, + 458 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 459, + 295, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 295, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 295, + 482 + ], + "type": "text", + "content": "To ensure proper content and style control, we design the CFG score as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 82, + 490, + 295, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 490, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 82, + 490, + 295, + 533 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tilde {\\epsilon} _ {\\theta} \\left(x _ {t}, h _ {\\text {t e x t}} ^ {\\text {c o n t e n t}}, h _ {\\text {v i s u a l}} ^ {\\text {s t y l e}}\\right) = (1 + w) \\\\ \\ddot {\\epsilon} _ {\\theta} \\left(x _ {t}, Q _ {\\text {t e x t}}, K V _ {\\text {v i s u a l}}\\right) \\tag {2} \\\\ - w \\epsilon_ {\\theta} (x _ {t}, \\emptyset). \\\\ \\end{array}", + "image_path": "a9d98a9b463b298b264f38688142be64cb2c33faa08bf56af18e138ee802063a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 544, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 579 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 544, + 295, + 579 + ], + "type": "inline_equation", + "content": "\\ddot{\\epsilon}_{\\theta}(x_{t},Q_{\\mathrm{text}},KV_{\\mathrm{visual}})" + }, + { + "bbox": [ + 55, + 544, + 295, + 579 + ], + "type": "text", + "content": " represents a KV-injected denoising score, indicating feature manipulation rather than direct conditioning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 580, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 295, + 651 + ], + "type": "text", + "content": "For given two denoising processes, one as original and another as a reference, borrowing the key-value in self-attention from the reference to the original process, i.e., key-value (KV) injection, tends to produces results with the content from the original process and the style from the reference process with limited control [1, 6, 56]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 651, + 295, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 295, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 295, + 686 + ], + "type": "text", + "content": "We define the KV-injected score by " + }, + { + "bbox": [ + 55, + 651, + 295, + 686 + ], + "type": "inline_equation", + "content": "\\ddot{\\epsilon}_{\\theta}(x_{t},Q_{\\mathrm{text}},KV_{\\mathrm{visual}})" + }, + { + "bbox": [ + 55, + 651, + 295, + 686 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 651, + 295, + 686 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{text}}" + }, + { + "bbox": [ + 55, + 651, + 295, + 686 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 651, + 295, + 686 + ], + "type": "inline_equation", + "content": "KV_{\\mathrm{visual}}" + }, + { + "bbox": [ + 55, + 651, + 295, + 686 + ], + "type": "text", + "content": " denote the query from the original score " + }, + { + "bbox": [ + 55, + 651, + 295, + 686 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,c_{\\mathrm{text}})" + }, + { + "bbox": [ + 55, + 651, + 295, + 686 + ], + "type": "text", + "content": " and the" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 355, + 555, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 555, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 555, + 378 + ], + "type": "text", + "content": "key, value from the reference score " + }, + { + "bbox": [ + 313, + 355, + 555, + 378 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,c_{\\mathrm{visual}})" + }, + { + "bbox": [ + 313, + 355, + 555, + 378 + ], + "type": "text", + "content": ". Then KV-injected-Attention becomes:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 390, + 554, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 390, + 554, + 423 + ], + "spans": [ + { + "bbox": [ + 314, + 390, + 554, + 423 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t e n t i o n} \\left(Q _ {\\text {t e x t}}, K _ {\\text {v i s u a l}}, V _ {\\text {v i s u a l}}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {Q _ {\\text {t e x t}} K _ {\\text {v i s u a l}} ^ {\\intercal}}{\\sqrt {d}}\\right) V _ {\\text {v i s u a l}}. \\tag {3}", + "image_path": "81ca0b665df35f9d4f133daf410f9359a7ed48edd7e884a518c83f1a1197b156.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 424, + 460, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 424, + 460, + 436 + ], + "spans": [ + { + "bbox": [ + 314, + 424, + 460, + 436 + ], + "type": "text", + "content": "We omit the layer index for brevity." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 437, + 555, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 557 + ], + "type": "text", + "content": "Naive denoising process with " + }, + { + "bbox": [ + 313, + 437, + 555, + 557 + ], + "type": "inline_equation", + "content": "\\ddot{e}_{\\theta}(x_{t},Q_{\\mathrm{text}},KV_{\\mathrm{visual}})" + }, + { + "bbox": [ + 313, + 437, + 555, + 557 + ], + "type": "text", + "content": " provides limited control in generating images with content and style specified by a text " + }, + { + "bbox": [ + 313, + 437, + 555, + 557 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{text}}" + }, + { + "bbox": [ + 313, + 437, + 555, + 557 + ], + "type": "text", + "content": " and a visual style prompt, respectively. Moving forward, our CFG with swapping self-attention in Eq. (2) enjoys higher image quality and more accurate text alignment than the naive denoising process as in the original CFG for T2I generation. The results are deferred to Section 4.1. However, content leakage still hinders the creation of diverse images, and disrupts text alignment." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 570, + 484, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 570, + 484, + 582 + ], + "spans": [ + { + "bbox": [ + 314, + 570, + 484, + 582 + ], + "type": "text", + "content": "2.3. Negative visual query guidance" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 589, + 554, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 589, + 554, + 637 + ], + "spans": [ + { + "bbox": [ + 313, + 589, + 554, + 637 + ], + "type": "text", + "content": "We propose negative visual query guidance (NVQG) to mitigate content leakage by reducing the content " + }, + { + "bbox": [ + 313, + 589, + 554, + 637 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{visual}}^{\\mathrm{content}}" + }, + { + "bbox": [ + 313, + 589, + 554, + 637 + ], + "type": "text", + "content": " from the visual style prompt appearing in the results. Briefly, NVQG negates the CFG of a score " + }, + { + "bbox": [ + 313, + 589, + 554, + 637 + ], + "type": "inline_equation", + "content": "\\ddot{\\epsilon}(x_t, Q_{\\mathrm{visual}}, KV_{\\mathrm{text}})" + }, + { + "bbox": [ + 313, + 589, + 554, + 637 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 638, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 638, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 638, + 555, + 714 + ], + "type": "text", + "content": "In Liu et al. [24], a complex text condition " + }, + { + "bbox": [ + 313, + 638, + 555, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 638, + 555, + 714 + ], + "type": "text", + "content": " is factorized as a set of conditions " + }, + { + "bbox": [ + 313, + 638, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\{c_0, c_1, \\ldots\\}" + }, + { + "bbox": [ + 313, + 638, + 555, + 714 + ], + "type": "text", + "content": " and Bayes' rule induces " + }, + { + "bbox": [ + 313, + 638, + 555, + 714 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_t | c_0, c_1, \\ldots) \\propto \\Pi \\frac{p_{\\theta}(x_t | c_i)}{p_{\\theta}(x_t)}" + }, + { + "bbox": [ + 313, + 638, + 555, + 714 + ], + "type": "text", + "content": ". Then, the score of the complex text condition " + }, + { + "bbox": [ + 313, + 638, + 555, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 638, + 555, + 714 + ], + "type": "text", + "content": " becomes " + }, + { + "bbox": [ + 313, + 638, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t, c) = \\epsilon_{\\theta}(x_t, \\emptyset) + \\Pi (\\epsilon_{\\theta}(x_t, c_i) - \\epsilon_{\\theta}(x_t, \\emptyset))" + }, + { + "bbox": [ + 313, + 638, + 555, + 714 + ], + "type": "text", + "content": ". It allows reducing a specific concept " + }, + { + "bbox": [ + 313, + 638, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{c}" + }, + { + "bbox": [ + 313, + 638, + 555, + 714 + ], + "type": "text", + "content": " with composition by negating its guidance with scale" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 714 + ], + "type": "text", + "content": "3The original score and its query within are recursively altered by KV injection along the denoising process." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15762" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 71, + 294, + 207 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 294, + 207 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 294, + 207 + ], + "type": "image", + "image_path": "8ea22fbcecac27f5c3edf52615eb626e68a3a16748453ac49dfcae68d79ac58d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 215, + 296, + 281 + ], + "lines": [ + { + "bbox": [ + 55, + 215, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 296, + 281 + ], + "type": "text", + "content": "Figure 3. The effect of swapping self-attention across different blocks. Swapping self-attention on the bottleneck and downblocks causes content leakage, resulting in cat-like images despite a dog prompt, while swapping on downblocks disrupts resulting images. We only apply swapping self-attention in the upblocks to reflect style elements effectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 303, + 80, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 80, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 80, + 313 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{neg}}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 319, + 295, + 343 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 319, + 295, + 343 + ], + "spans": [ + { + "bbox": [ + 61, + 319, + 295, + 343 + ], + "type": "interline_equation", + "content": "\\epsilon_ {\\theta} \\left(x _ {t}, c, \\text {n o t} \\tilde {c}\\right) = \\epsilon_ {\\theta} \\left(x _ {t}, c\\right) - w _ {\\text {n e g}} \\left(\\epsilon_ {\\theta} \\left(x _ {t}, \\tilde {c}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, \\emptyset\\right)\\right) \\tag {4}", + "image_path": "8a1d4ca5c0a1b1b3de6cc674440c33663c20f2db481535ce47fc571735ff74d2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 343, + 296, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 296, + 394 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 296, + 394 + ], + "type": "text", + "content": "Although we design " + }, + { + "bbox": [ + 55, + 343, + 296, + 394 + ], + "type": "inline_equation", + "content": "\\ddot{\\epsilon}_{\\theta}(x_{t},Q_{\\mathrm{text}},KV_{\\mathrm{visual}})" + }, + { + "bbox": [ + 55, + 343, + 296, + 394 + ], + "type": "text", + "content": " to predict the score toward " + }, + { + "bbox": [ + 55, + 343, + 296, + 394 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_0|h_{\\mathrm{text}}^{\\mathrm{content}},h_{\\mathrm{visual}}^{\\mathrm{style}})" + }, + { + "bbox": [ + 55, + 343, + 296, + 394 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 343, + 296, + 394 + ], + "type": "inline_equation", + "content": "\\ddot{\\epsilon} (x_{t},Q_{\\mathrm{text}},KV_{\\mathrm{visual}})" + }, + { + "bbox": [ + 55, + 343, + 296, + 394 + ], + "type": "text", + "content": " still contain " + }, + { + "bbox": [ + 55, + 343, + 296, + 394 + ], + "type": "inline_equation", + "content": "h_\\mathrm{visual}" + }, + { + "bbox": [ + 55, + 343, + 296, + 394 + ], + "type": "text", + "content": ". Assuming a hidden factorization " + }, + { + "bbox": [ + 55, + 343, + 296, + 394 + ], + "type": "inline_equation", + "content": "KV_{\\mathrm{visual}} = \\{KV_{\\mathrm{visual}}^{\\mathrm{content}},KV_{\\mathrm{visual}}^{\\mathrm{style}}\\}" + }, + { + "bbox": [ + 55, + 343, + 296, + 394 + ], + "type": "text", + "content": ", Bayes' rule induces" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 399, + 294, + 457 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 457 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 457 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} p _ {\\theta} \\left(x _ {t} \\mid Q _ {\\text {t e x t}}, K V _ {\\text {v i s u a l}} ^ {\\text {s t y l e}}, K V _ {\\text {v i s u a l}} ^ {\\text {c o n t e n t}}\\right) \\propto p _ {\\theta} \\left(x _ {t}\\right) \\frac {p _ {\\theta} \\left(x _ {t} \\mid Q _ {\\text {t e x t}} , K V _ {\\text {v i s u a l}} ^ {\\text {s t y l e}}\\right)}{p _ {\\theta} \\left(x _ {t}\\right)} \\\\ \\times \\frac {p _ {\\theta} \\left(x _ {t} \\mid Q _ {\\emptyset} , K V _ {\\text {v i s u a l}} ^ {\\text {c o n t e n t}}\\right)}{p _ {\\theta} \\left(x _ {t}\\right)}. \\tag {5} \\\\ \\end{array}", + "image_path": "008b4c5f52a6b1b206f3369312be5b07ad0cbe5bc5b4449b50dfeebb6fad9cc8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 459, + 295, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 295, + 485 + ], + "type": "text", + "content": "Since " + }, + { + "bbox": [ + 55, + 459, + 295, + 485 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t) = \\epsilon_{\\theta}(x_t, Q_{\\emptyset}, KV_{\\emptyset})" + }, + { + "bbox": [ + 55, + 459, + 295, + 485 + ], + "type": "text", + "content": ", we obtain the desired conditional score of " + }, + { + "bbox": [ + 55, + 459, + 295, + 485 + ], + "type": "inline_equation", + "content": "\\hat{c} = \\{Q_{\\mathrm{text}}, KV_{\\mathrm{visual}}^{\\mathrm{style}}\\}" + }, + { + "bbox": [ + 55, + 459, + 295, + 485 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 64, + 491, + 294, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 491, + 294, + 545 + ], + "spans": [ + { + "bbox": [ + 64, + 491, + 294, + 545 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\epsilon_ {\\theta} \\left(x _ {t}, \\hat {c}\\right) \\leftarrow w _ {\\text {v i s u a l}} \\left(\\ddot {\\epsilon} _ {\\theta} \\left(x _ {t}, Q _ {\\text {t e x t}}, K V _ {\\text {v i s u a l}}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}\\right)\\right) \\\\ - w _ {\\text {c o n t e n t}} \\left(\\ddot {\\epsilon} _ {\\theta} \\left(x _ {t}, Q _ {\\emptyset}, K V _ {\\text {v i s u a l}} ^ {\\text {c o n t e n t}}\\right) - \\epsilon_ {\\theta} \\left(x _ {t}\\right)\\right) \\\\ + \\epsilon_ {\\theta} \\left(x _ {t}\\right). \\tag {6} \\\\ \\end{array}", + "image_path": "54a2947f07c128f1cf5dfa7c4d1a1f669ceb568985152cabc31c1cdfb96fe88f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 547, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 547, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 547, + 295, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 547, + 295, + 666 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{visual}}" + }, + { + "bbox": [ + 55, + 547, + 295, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 547, + 295, + 666 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{content}}" + }, + { + "bbox": [ + 55, + 547, + 295, + 666 + ], + "type": "text", + "content": " regulate the strength of each guidance. By leveraging query injection [1, 6, 46, 56], we approximate " + }, + { + "bbox": [ + 55, + 547, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\ddot{\\epsilon}_{\\theta}(x_t, Q_{\\emptyset}, KV_{\\mathrm{visual}}^{\\mathrm{content}}) \\approx \\ddot{\\epsilon}_{\\theta}(x_t, Q_{\\mathrm{visual}}, KV_{\\emptyset})" + }, + { + "bbox": [ + 55, + 547, + 295, + 666 + ], + "type": "text", + "content": ". Finally, inserting (6) into (2), and empirically replacing " + }, + { + "bbox": [ + 55, + 547, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t)" + }, + { + "bbox": [ + 55, + 547, + 295, + 666 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 547, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\ddot{\\epsilon}\\theta(x_t, Q_{\\mathrm{visual}})" + }, + { + "bbox": [ + 55, + 547, + 295, + 666 + ], + "type": "text", + "content": " results in similar performance. This allows a simplified reformulation of diffusion sampling using " + }, + { + "bbox": [ + 55, + 547, + 295, + 666 + ], + "type": "inline_equation", + "content": "w' = w_{\\mathrm{visual}}(w + 1)" + }, + { + "bbox": [ + 55, + 547, + 295, + 666 + ], + "type": "text", + "content": ": Since we can highly relate the " + }, + { + "bbox": [ + 55, + 547, + 295, + 666 + ], + "type": "inline_equation", + "content": "w' \\ddot{\\epsilon}_{\\theta}(x_t, Q_{\\mathrm{visual}})" + }, + { + "bbox": [ + 55, + 547, + 295, + 666 + ], + "type": "text", + "content": " to the concept negation in (4) which guides the negation of concept with a scale " + }, + { + "bbox": [ + 55, + 547, + 295, + 666 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{neg}}" + }, + { + "bbox": [ + 55, + 547, + 295, + 666 + ], + "type": "text", + "content": ", we named the query term as negative visual query guidance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 672, + 282, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 282, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 282, + 685 + ], + "type": "text", + "content": "2.4. Choosing blocks for swapping self-attention" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "In order to maximize the effect of NVQG in preventing content leakage, we explore the depth of the self-attention" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 333, + 72, + 542, + 242 + ], + "blocks": [ + { + "bbox": [ + 333, + 72, + 542, + 242 + ], + "lines": [ + { + "bbox": [ + 333, + 72, + 542, + 242 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 542, + 242 + ], + "type": "image", + "image_path": "f3312850a2d60f836f692e9f0785c66f700290c48ffd93a6f5b42094dc45ad80.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 251, + 555, + 328 + ], + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 328 + ], + "type": "text", + "content": "Figure 4. The effect of CFG and the proposed negative visual query guidance on image generation. The reference images provide the style for each generated output. Without NVQG, content leakage occurs, and the generated images fail to fully capture the intended content. In contrast, using NVQG ensures better alignment with both the reference style and the \"Cat\" prompt, reducing content distortion and improving quality." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 336, + 555, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 555, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 555, + 598 + ], + "type": "text", + "content": "blocks to be swapped in the sense of granularity of visual elements. Modern architecture of diffusion models roughly consists of three sections in a sequence: downblocks, bottleneck blocks, and upblocks. Given that the bottleneck of diffusion models contains content elements of the image [20, 22, 33], we opt not to apply swapping self-attention to bottleneck blocks to prevent transferring contents in a reference image. Figure 3 shows that not swapping the bottleneck blocks prevents content leaking from the reference image. However, the synthesized images show disrupted results with seriously scattered objects. Furthermore, while swapping self-attention implements the reassembling operation, simply applying to all self-attention layers exposes a content leakage problem, where the content of the reference image influences the resulting image, as shown in the first row of Figure 3. I.e., the results contain cats even though the prompts specify \"a dog\". We conjecture that this phenomenon happens because feature maps of downblocks have unclear content layout [3, 27], so substituting features based on this inaccurate layout causes the disrupted results. To avoid injecting unnecessary features, we choose to swap the key and value of self-attention only in upblocks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 600, + 554, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 554, + 658 + ], + "type": "text", + "content": "We note that Hertz et al. [16] applies self-attention operation at all blocks and suffers content leakage. The last row of Figure 3 shows the success of our strategy. Following this, we proceed to search for a specific optimal layer in the upblock in Section 4.2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 669, + 460, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 669, + 460, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 669, + 460, + 682 + ], + "type": "text", + "content": "3. Real images as references" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "So far, we have assumed a generated visual style prompt " + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "inline_equation", + "content": "x_0^{\\mathrm{visual}} \\sim p_\\theta(x_0 | c_{\\mathrm{visual}}}" + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": ". Here, we allow real visual style" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15763" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 74, + 70, + 287, + 408 + ], + "blocks": [ + { + "bbox": [ + 74, + 70, + 287, + 408 + ], + "lines": [ + { + "bbox": [ + 74, + 70, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 74, + 70, + 287, + 408 + ], + "type": "image", + "image_path": "742e55a6753a13dceac66efd180e4fa851ff786a6abde0262353f7cdf9f6b28f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 415, + 295, + 504 + ], + "lines": [ + { + "bbox": [ + 55, + 415, + 295, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 295, + 504 + ], + "type": "text", + "content": "Figure 5. Analysis on the optimal range of upblocks for swapping self-attention. We find the optimal range of upblocks for a balanced trade-off between different aspects. The images on the right illustrate the visual results for different upblock layer indices, with the red cross indicating poor diversity and misalignment to the text prompt, the red triangle indicating a lack of style similarity, and the yellow star indicating the optimal results. Please refer to Section 4.2 for details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 525, + 294, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 294, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 294, + 537 + ], + "type": "text", + "content": "prompts by 1) stochastic encoding and 2) color calibration." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 537, + 295, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 537, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 537, + 295, + 574 + ], + "type": "text", + "content": "We propose stochastic encoding to obtain " + }, + { + "bbox": [ + 55, + 537, + 295, + 574 + ], + "type": "inline_equation", + "content": "x_{t}^{\\mathrm{visual}} \\sim q(x_{t}|x_{0}^{\\mathrm{visual}})" + }, + { + "bbox": [ + 55, + 537, + 295, + 574 + ], + "type": "text", + "content": " by adding a random noise on " + }, + { + "bbox": [ + 55, + 537, + 295, + 574 + ], + "type": "inline_equation", + "content": "x_{0}^{\\mathrm{visual}}" + }, + { + "bbox": [ + 55, + 537, + 295, + 574 + ], + "type": "text", + "content": " following the forward process of DMs [18]:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 582, + 295, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 582, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 64, + 582, + 295, + 597 + ], + "type": "interline_equation", + "content": "\\epsilon_ {t} \\sim \\mathcal {N} (0, I), x _ {t} ^ {\\text {v i s u a l}} = \\sqrt {\\alpha_ {t}} \\cdot x _ {0} ^ {\\text {v i s u a l}} + \\sqrt {1 - \\alpha_ {t}} \\cdot \\epsilon_ {t} \\tag {7}", + "image_path": "c21be29ae3ba906e410f236529b0d0355db73ff19530301ea500321c9a092632.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "content": "At each timestep, we sample " + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\epsilon_{t}" + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "content": " to encode " + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "inline_equation", + "content": "x_{t}^{\\mathrm{visual}}" + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "content": ". It ensures that " + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "inline_equation", + "content": "x_{t}^{\\mathrm{visual}}" + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "content": " lies on the learned trajectory and does not suffer from accumulative numerical error due to iterative process of DDIM inversion used by previous methods [6, 16] as shown in Figure A1. Furthermore, it does not need to store the intermediate latents as opposed to DDPM inversion used by Alaluf et al. [1]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Although stochastic encoding performs better than DDIM inversion, a subtle color discrepancy occurs between" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 267 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 267 + ], + "type": "text", + "content": "the resulting images and the visual style prompts. We introduce color calibration at " + }, + { + "bbox": [ + 313, + 72, + 553, + 267 + ], + "type": "inline_equation", + "content": "x_{t}^{\\mathrm{ori}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 267 + ], + "type": "text", + "content": " in the original process to match the statistics of predicted " + }, + { + "bbox": [ + 313, + 72, + 553, + 267 + ], + "type": "inline_equation", + "content": "x_{0}^{\\mathrm{ori}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 267 + ], + "type": "text", + "content": " to predicted " + }, + { + "bbox": [ + 313, + 72, + 553, + 267 + ], + "type": "inline_equation", + "content": "x_{0}^{\\mathrm{visual}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 267 + ], + "type": "text", + "content": ". In Gatys [12], distance between channel-wise statistics is employed as a style loss for style transfer. In Song et al. [43], predicted " + }, + { + "bbox": [ + 313, + 72, + 553, + 267 + ], + "type": "inline_equation", + "content": "x_{0} \\left( = \\frac{x_{t}^{\\mathrm{visual}} - \\sqrt{1 - \\alpha_{t}} \\cdot \\epsilon_{\\theta}(\\hat{x}_{t})}{\\sqrt{\\alpha_{t}}} \\right)" + }, + { + "bbox": [ + 313, + 72, + 553, + 267 + ], + "type": "text", + "content": " allows to estimate " + }, + { + "bbox": [ + 313, + 72, + 553, + 267 + ], + "type": "inline_equation", + "content": "x_{0}" + }, + { + "bbox": [ + 313, + 72, + 553, + 267 + ], + "type": "text", + "content": " with high probability at intermediate timesteps using deterministic sampling. Inheriting the advantage, we execute adain operation to match mean&std of predicted " + }, + { + "bbox": [ + 313, + 72, + 553, + 267 + ], + "type": "inline_equation", + "content": "x_{0}^{\\mathrm{ori}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 267 + ], + "type": "text", + "content": " with those of " + }, + { + "bbox": [ + 313, + 72, + 553, + 267 + ], + "type": "inline_equation", + "content": "x_{0}^{\\mathrm{visual}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 267 + ], + "type": "text", + "content": ". It allows precise color calibration rather than directly matching " + }, + { + "bbox": [ + 313, + 72, + 553, + 267 + ], + "type": "inline_equation", + "content": "x_{t}^{\\mathrm{ori}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 267 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 72, + 553, + 267 + ], + "type": "inline_equation", + "content": "x_{t}^{\\mathrm{visual}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 267 + ], + "type": "text", + "content": " in Alaluf et al. [1], Chung et al. [6]. Furthermore, our approach differs from Chung et al. [6] in that using predicted " + }, + { + "bbox": [ + 313, + 72, + 553, + 267 + ], + "type": "inline_equation", + "content": "x_{0}" + }, + { + "bbox": [ + 313, + 72, + 553, + 267 + ], + "type": "text", + "content": " at intermediate timestep " + }, + { + "bbox": [ + 313, + 72, + 553, + 267 + ], + "type": "inline_equation", + "content": "t \\in (0,T)" + }, + { + "bbox": [ + 313, + 72, + 553, + 267 + ], + "type": "text", + "content": " other than " + }, + { + "bbox": [ + 313, + 72, + 553, + 267 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 313, + 72, + 553, + 267 + ], + "type": "text", + "content": " by reducing cumulative sampling error after the operation. Furthermore, Chung et al. [6] executes AdaIN at timestep T, inducing lengthy cumulative error." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 268, + 553, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 553, + 304 + ], + "type": "text", + "content": "We provide supportive experiments that show the effectiveness of the proposed method in Section 4.4 and a detailed Algorithm in Appendix 3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 317, + 395, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 395, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 395, + 331 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 338, + 553, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 553, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 553, + 457 + ], + "type": "text", + "content": "In this section, we describe the effects of our proposed methods: CFG with swapping self-attention, Negative visual query guidance (NVQG), stochastic encoding, and color calibration. For swapping self-attention, we provide a detailed analysis through experiments to determine the optimal layers for swapping. The impact of NVQG is demonstrated through qualitative results. Additionally, we show why stochastic encoding outperforms DDIM inversion when inverting real images, and we highlight the benefits of color calibration through experimental results." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 459, + 553, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 553, + 542 + ], + "type": "text", + "content": "We also conducted both quantitative and qualitative comparisons of our method against various competitors, including StyleAligned [16], IP-Adapter [59], Dreambooth-LoRA [39, 40], StyleDrop [42], DEADiff [35], CSGO [55] and InstantStyle [50, 51]. The details of these comparisons, along with the experimental setup and metrics, are described in the Appendix 2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 552, + 491, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 491, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 491, + 565 + ], + "type": "text", + "content": "4.1. Effectiveness of CFG and NVQG" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "This section analyzes the effects of Classifier-Free Guidance (CFG) and Negative Visual Query Guidance (NVQG) on image generation, with a focus on text alignment and content leakage. Figure 4 shows the results of the three configurations. In the " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "1^{\\text{st}}" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " row, without CFG and NVQG, the generated images suffer from severe artifacts. The absence of CFG causes poor image quality resulting in significant misalignment with the prompt. In the " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "2^{\\text{nd}}" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " row, CFG with swapping self-attention improves the text misalignment by boosting image quality. Here, the \"cat\" in target text prompt becomes clearer in the generated images. However, content leakage from the reference image remains where unwanted" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15764" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 19, + 294, + 193 + ], + "blocks": [ + { + "bbox": [ + 50, + 19, + 294, + 193 + ], + "lines": [ + { + "bbox": [ + 50, + 19, + 294, + 193 + ], + "spans": [ + { + "bbox": [ + 50, + 19, + 294, + 193 + ], + "type": "image", + "image_path": "7db0b67b31668fdea91bc83886ab0a678a0195591dcc4347a54a5a0cb310cc17.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 194, + 295, + 248 + ], + "lines": [ + { + "bbox": [ + 55, + 194, + 295, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 295, + 248 + ], + "type": "text", + "content": "Figure 6. Attention map visualization over late and early upblock layers. The late upblock better focuses on the style-corresponding region than the early upblock, leading to more freedom to reassemble small parts. The early upblock attends a larger region leading to content leakage." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 268, + 295, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 268, + 295, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 268, + 295, + 291 + ], + "type": "text", + "content": "elements (layouts, structure, and composition) from the reference image affect the results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 292, + 295, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 292, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 295, + 327 + ], + "type": "text", + "content": "In the " + }, + { + "bbox": [ + 55, + 292, + 295, + 327 + ], + "type": "inline_equation", + "content": "3^{\\mathrm{rd}}" + }, + { + "bbox": [ + 55, + 292, + 295, + 327 + ], + "type": "text", + "content": " row, NVQG releases the content leakage and produces the best results closely matching the text prompt while reflecting style reference." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 327, + 295, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 295, + 388 + ], + "type": "text", + "content": "Overall, Figure 4 demonstrates the critical role of NVQG in reducing content leakage, enjoying the quality boosting of CFG. Together, they ensure that the generated images align to both the target text prompt and the visual style prompts, resulting in high-quality, coherent outputs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 388, + 296, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 296, + 423 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 296, + 423 + ], + "type": "text", + "content": "We qualitatively showcase diversity of results within a text prompt in Figure A9 and text alignment with complex text prompts in Figure A22." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 430, + 252, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 252, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 252, + 443 + ], + "type": "text", + "content": "4.2. Analysis for Swapping Self-Attention" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 448, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 295, + 544 + ], + "type": "text", + "content": "Optimal layers in upblocks Since recent large T2I DMs consist of many layers, we further analyze the behavior by changing the start of the swapping while the end of the swapping is fixed at the end. We use four key metrics as shown in Figure 5, there is a layer where all four metrics abruptly change (red line). Notably, this point remains consistent regardless of the reference image. All the results in the paper use only one fixed optimal layer." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 544, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 590 + ], + "type": "text", + "content": "We choose this layer as the optimal start of the swapping for a balanced trade-off among all aspects. We provide qualitative results with detailed split of layers in Figure A3 and A4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "Visualizing Attention maps Figure 6 compares average attention maps from the late upblock and the early upblock applying swapping self-attention. Using late upblock has more freedom to reassemble the reference style elements leading to more doggy results than early upblock which produces some cat-like attributes. The right two columns visualize the attention weight of query points marked as red stars and yellow dots. Swapping self-attention on late upblock reassemblies features from a style correspondence," + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 328, + 69, + 537, + 135 + ], + "blocks": [ + { + "bbox": [ + 328, + 69, + 537, + 135 + ], + "lines": [ + { + "bbox": [ + 328, + 69, + 537, + 135 + ], + "spans": [ + { + "bbox": [ + 328, + 69, + 537, + 135 + ], + "type": "image", + "image_path": "a02de94a8bac9e60c8baf74bb9bf83e37f85cf6f6611f760605ce6d222548fd3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 136, + 536, + 145 + ], + "lines": [ + { + "bbox": [ + 362, + 136, + 536, + 145 + ], + "spans": [ + { + "bbox": [ + 362, + 136, + 536, + 145 + ], + "type": "text", + "content": "(a) Synthesizing the content specified by the text prompts" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 328, + 148, + 535, + 211 + ], + "blocks": [ + { + "bbox": [ + 328, + 148, + 535, + 211 + ], + "lines": [ + { + "bbox": [ + 328, + 148, + 535, + 211 + ], + "spans": [ + { + "bbox": [ + 328, + 148, + 535, + 211 + ], + "type": "image", + "image_path": "294659934ab125c51d9e03757c7be199ff88ee10e90372955153aa97745e125c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 212, + 529, + 221 + ], + "lines": [ + { + "bbox": [ + 369, + 212, + 529, + 221 + ], + "spans": [ + { + "bbox": [ + 369, + 212, + 529, + 221 + ], + "type": "text", + "content": "(b) Synthesizing the pose specified by the text prompt" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 328, + 224, + 535, + 288 + ], + "blocks": [ + { + "bbox": [ + 328, + 224, + 535, + 288 + ], + "lines": [ + { + "bbox": [ + 328, + 224, + 535, + 288 + ], + "spans": [ + { + "bbox": [ + 328, + 224, + 535, + 288 + ], + "type": "image", + "image_path": "1345f86ee2b06b053f2c2d4e887274eb2a4922b71a756c564fb4d704dd3df9d7.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 289, + 538, + 298 + ], + "lines": [ + { + "bbox": [ + 361, + 289, + 538, + 298 + ], + "spans": [ + { + "bbox": [ + 361, + 289, + 538, + 298 + ], + "type": "text", + "content": "(c) Synthesizing a single object specified by the text prompt" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 308, + 555, + 342 + ], + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 342 + ], + "type": "text", + "content": "Figure 7. Comparison for content leakage. While StyleAligned suffers from content leakage from the reference, our results clearly align with the text prompts" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 365, + 555, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 425 + ], + "type": "text", + "content": "e.g., texture and color. On the other hand, swapping self-attention on early upblock reassembles features from a wider area with different styles. This comparison clarifies the reasons for using only late upblock. Please refer to Figure A6, A7 for a more visualization and a detailed analysis." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 447, + 489, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 489, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 489, + 460 + ], + "type": "text", + "content": "4.3. Comparison against competitors" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 466, + 555, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 555, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 555, + 597 + ], + "type": "text", + "content": "Style & content control We provide a qualitative comparison in Figure 8, focusing on controlling style and content. Our method faithfully synthesizes content from the text prompt with the style of the reference image. In contrast, other methods add elements like color or texture not in the reference (e.g., feathers, bricks, iron, skin) and often suffer from content leakage (e.g., layout, screaming person, castle), which compromises text prompt faithfulness. Quantitative results in Figure 9 support these findings: IP-Adapter shows higher style similarity but neglects text prompts significantly." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": "Diversity within a text specification Starting from different initial noises, the diffusion models trained on a large dataset produce diverse results within the specification of a text prompt. Figure 1, A19 shows that ours can synthesize various poses and viewpoints while others barely change, i.e., other methods limit the diversity of the pretrained model due to content leakage. We further support our diversity in Figure A9." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15765" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 306, + 264 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 306, + 264 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 306, + 264 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 306, + 264 + ], + "type": "image", + "image_path": "16642bdb6d4b6d938e40598a25670587e9b25d0d5d8af6b3303ae87ea9408cc0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 271, + 555, + 296 + ], + "lines": [ + { + "bbox": [ + 55, + 271, + 555, + 296 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 555, + 296 + ], + "type": "text", + "content": "Figure 8. Qualitative comparison across various styles and text prompts. StyleKeeper faithfully reflects style elements in reference images without causing content leakage from the reference images." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 306, + 70, + 553, + 262 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 553, + 262 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 553, + 262 + ], + "type": "image", + "image_path": "ec3c6df998fac707c65c563de5a5b534b103beac9e4305e1a01f54a42aea5858.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 56, + 306, + 298, + 407 + ], + "blocks": [ + { + "bbox": [ + 56, + 306, + 298, + 407 + ], + "lines": [ + { + "bbox": [ + 56, + 306, + 298, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 306, + 298, + 407 + ], + "type": "image", + "image_path": "d6b0e853896fd98ea49de2ea556527d55f664585c2779fd4e027add15c533bf2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 415, + 298, + 449 + ], + "lines": [ + { + "bbox": [ + 55, + 415, + 298, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 298, + 449 + ], + "type": "text", + "content": "Figure 9. Quantitative comparison. Ours achieves the best style similarity while not compromising text alignment and being free from content leakage." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 462, + 298, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 462, + 298, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 462, + 298, + 715 + ], + "type": "text", + "content": "Content leakage Content leakage refers to the phenomenon where the content of a reference image appears in a result. As evident in Figure A12, our model exhibits significantly less content leakage when compared to other models. We focus on a comparison against a runner-up method, StyleAligned, particularly in terms of how content leakage can be an obstacle to controlling the content using a text prompt. Figure 7a compares examples where strong content leakage prevents the object from the text from appearing in StyleAligned. We often observe famous paintings in the reference easily leak into the results of StyleAligned while ours does not struggle. Figure 7b compares examples where the pose in the reference leaks into the results of StyleAligned preventing the reflection of specified pose in the text prompt. On the other hand, our method reflects the correct poses. Figure 7c compares examples where the number of small instances in the reference leaks into the result of StyleAligned. Contrarily, our method correctly synthesizes a single instance of the content specified by the text. In Figure 9, we provide quantitative comparison results. It supports that StyleKeeper achieves the best style" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 315, + 556, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 556, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 556, + 352 + ], + "type": "text", + "content": "similarity while not suffering content leakage. We can also observe that models with high content leakage exhibit poor text alignment." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 380, + 556, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 556, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 556, + 501 + ], + "type": "text", + "content": "Preserving the content of the original denoising process Figure A13 shows the results of ours and other methods using the same initial noise in each column. Our method precisely reflects the style in the reference with minimal changes of contents in the original denoising process. On the other hand, the other methods severely alter pose, shape, or layouts due to content leakage. It is an important virtue of controlling style to keep the rest intact. This also supports that ours is free from content leakage during the denoising process. We provide more results in Figure A14." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 514, + 493, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 493, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 493, + 528 + ], + "type": "text", + "content": "4.4. DDIM inversion vs. Our strategy" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 534, + 556, + 715 + ], + "type": "text", + "content": "Figure 11 shows that StyleKeeper can take real images as style reference with our strategies. As shown in Figure 11, stochastic encoding outperforms DDIM inversion and color calibration improves text alignment and style similarity. We provide more results in Figure A15. Moreover, we show that our strategy can boost the performance of the other self-attention variants [3, 16] in Figure A16, A17 and color calibration can be used for generation settings in Figure A18. We provide an ablation study in Figure 12 for each configuration (swapping self-attention, NVQG, and color calibration) in both real reference and generated reference settings. In both settings, swapping self-attention and NVQG improve style similarity and text alignment, while color calibration helps improve style similarity. Additionally, stochastic encoding outperforms DDIM inversion." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15766" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 553, + 178 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 178 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 178 + ], + "type": "image", + "image_path": "3bbc89c10d184627277529b4f7860917d936649cb65c4c1d43b37cd9d94d73e3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 194, + 555, + 217 + ], + "lines": [ + { + "bbox": [ + 55, + 194, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 555, + 217 + ], + "type": "text", + "content": "Figure 10. Qualitative comparison in I2I style transfer task. We compare our method for I2I style transfer task where the content image is given. Compared to the previous methods, ours achieves the best style similarity." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 74, + 226, + 279, + 363 + ], + "blocks": [ + { + "bbox": [ + 74, + 226, + 279, + 363 + ], + "lines": [ + { + "bbox": [ + 74, + 226, + 279, + 363 + ], + "spans": [ + { + "bbox": [ + 74, + 226, + 279, + 363 + ], + "type": "image", + "image_path": "225153d76822477e8f057c560c9e7bcce824ff3119ab63fcd62ba3b42f4b1080.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 371, + 296, + 415 + ], + "lines": [ + { + "bbox": [ + 55, + 371, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 296, + 415 + ], + "type": "text", + "content": "Figure 11. Comparison of DDIM inversion vs. stochastic encoding and the effect of color calibration. Stochastic encoding reduces artifacts in the resulting images, while color calibration better reflects the colors of the reference image." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 59, + 425, + 294, + 495 + ], + "blocks": [ + { + "bbox": [ + 59, + 425, + 294, + 495 + ], + "lines": [ + { + "bbox": [ + 59, + 425, + 294, + 495 + ], + "spans": [ + { + "bbox": [ + 59, + 425, + 294, + 495 + ], + "type": "image", + "image_path": "a43cf8f9591b5a70d31b8a6bf4a3076f541ce48d11840ad333c486cecae3bc7e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 507, + 296, + 586 + ], + "lines": [ + { + "bbox": [ + 55, + 507, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 296, + 586 + ], + "type": "text", + "content": "Figure 12. Ablation study of four proposed methods with a real reference and a generated reference setting.. We evaluated performance by individually removing each of the four proposed methods. In all cases, both generated and real images showed similar improvements in style similarity and text alignment. Moreover, stochastic encoding outperformed DDIM inversion in all aspects, comparing purple and blue points." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 609, + 234, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 234, + 621 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 234, + 621 + ], + "type": "text", + "content": "4.5. Comparison for I2I style transfer" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 628, + 296, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 296, + 700 + ], + "type": "text", + "content": "In Figure 10, we present a qualitative comparison between our method using ControlNet and existing state-of-the-art methods, CrossAttn [1] and StyleID [6], for the I2I style transfer task where a content image is provided. Both CrossAttn and StyleID inject the key & value obtained by inversion from the style images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "CrossAttn often fails to transfer the detailed representa" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 232, + 555, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 327 + ], + "type": "text", + "content": "tion of the reference style, leading to a rough, blocky appearance. For instance, the center examples in Figure 10 show that the style details are absent or inaccurate compared to our results. Similarly, StyleID suffers poor style reflection and color deficiency in the center and the rightmost example. In contrast, ours effectively reflects the details of the reference style image, with no noticeable transfer of the color values from the content images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 342, + 468, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 468, + 355 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 468, + 355 + ], + "type": "text", + "content": "5. Conclusion and Limitation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 364, + 555, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 555, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 555, + 532 + ], + "type": "text", + "content": "In this paper, we introduce StyleKeeper which faithfully applies the style of real reference images without content leakage in a training-free manner. CFG with swapping self-attention captures the reference image's style accurately and allows for direct content generation from the text. This integration with CFG enhances performance by balancing content generation and style transfer. To address content leakage, we propose Negative visual query guidance that ensures the reference image's content does not interfere with the text-specified content, enhancing diversity and improving text alignment & style similarity. Additionally, Stochastic encoding efficiently maps real images to suitable latents, removing artifacts. Color calibration aligns the final output to the reference's color statistics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "content": "Our method shows both qualitative and quantitative improvements over existing approaches, providing a robust solution for visual style prompting without complex training. We also analyze where to apply swapping self-attention, identifying optimal layers to balance style transfer and content fidelity. StyleKeeper outperforms prior methods and works well with algorithms like ControlNet [60] and DBLoRA [40], as shown in Figure A20." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "However, StyleKeeper is limited by the capabilities of the pretrained diffusion models and cannot generate elements outside their scope (e.g., \"stone golem\" in Figure A24a). Also, if visual and textual styles conflict, the visual style dominates (Figure A24b). Future work could explore extending our method to other domains such as video, expanding its applicability and research potential." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15767" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "type": "text", + "content": "Acknowledgements We thank Jin-Hwa Kim for helpful comments. This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS-2024-00439762, Developing Techniques for Analyzing and Assessing Vulnerabilities, and Tools for Confidentiality Evaluation in Generative AI Models)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 167, + 115, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 167, + 115, + 178 + ], + "spans": [ + { + "bbox": [ + 56, + 167, + 115, + 178 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 186, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 186, + 296, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 186, + 296, + 230 + ], + "spans": [ + { + "bbox": [ + 61, + 186, + 296, + 230 + ], + "type": "text", + "content": "[1] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer. In ACM SIGGRAPH 2024 Conference Papers, pages 1–12, 2024. 2, 3, 4, 5, 8, 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 232, + 296, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 232, + 296, + 264 + ], + "spans": [ + { + "bbox": [ + 61, + 232, + 296, + 264 + ], + "type": "text", + "content": "[2] Omri Avrahami, Kfir Aberman, Ohad Fried, Daniel Cohen-Or, and Dani Lischinski. Break-a-scene: Extracting multiple concepts from a single image. In SIGGRAPH Asia, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 266, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 266, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 62, + 266, + 295, + 308 + ], + "type": "text", + "content": "[3] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In ICCV), 2023. 2, 4, 7, 1, 9, 15" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 310, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 310, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 62, + 310, + 294, + 353 + ], + "type": "text", + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 355, + 295, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 355, + 295, + 398 + ], + "spans": [ + { + "bbox": [ + 62, + 355, + 295, + 398 + ], + "type": "text", + "content": "[5] Hyungjin Chung, Jeongsol Kim, Geon Yeong Park, Hyelin Nam, and Jong Chul Ye. Cfg++: Manifold-constrained classifier free guidance for diffusion models. arXiv preprint arXiv:2406.08070, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 400, + 296, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 400, + 296, + 454 + ], + "spans": [ + { + "bbox": [ + 62, + 400, + 296, + 454 + ], + "type": "text", + "content": "[6] Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8795-8805, 2024. 2, 3, 4, 5, 8, 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 456, + 294, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 456, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 62, + 456, + 294, + 488 + ], + "type": "text", + "content": "[7] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 490, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 490, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 490, + 296, + 544 + ], + "type": "text", + "content": "[8] Yingying Deng, Fan Tang, Weiming Dong, Chongyang Ma, Xingjia Pan, Lei Wang, and Changsheng Xu. Stytr2: Image style transfer with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11326-11336, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 545, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 545, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 62, + 545, + 294, + 577 + ], + "type": "text", + "content": "[9] Martin Nicolas Everaert, Marco Bocchio, Sami Arpa, Sabine Süssstrunk, and Radhakrishna Achanta. Diffusion in style. In ICCV, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 579, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 579, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 57, + 579, + 294, + 623 + ], + "type": "text", + "content": "[10] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 624, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 294, + 656 + ], + "type": "text", + "content": "[11] Daniel Garibi, Or Patashnik, Andrey Voynov, Hadar Averbuch-Elor, and Daniel Cohen-Or. Renoise: Real image inversion through iterative noising, 2024. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 679 + ], + "type": "text", + "content": "[12] Leon A Gatys. A neural algorithm of artistic style. arXiv preprint arXiv:1508.06576, 2015. 5, 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "text", + "content": "[13] Inhwa Han, Serin Yang, Taesung Kwon, and Jong Chul Ye. Highly personalized text embedding for image manipulation by stable diffusion. arXiv, 2023. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "type": "text", + "content": "[14] Ligong Han, Song Wen, Qi Chen, Zhixing Zhang, Kunpeng Song, Mengwei Ren, Ruijiang Gao, Anastasis Stathopoulos, Xiaoxiao He, Yuxiao Chen, Di Liu, Qilong Zhangli, Jindong Jiang, Zhaoyang Xia, Akash Srivastava, and Dimitris Metaxas. Improving tuning-free real image editing with proximal guidance, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 140, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 171 + ], + "type": "text", + "content": "[15] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 173, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 205 + ], + "type": "text", + "content": "[16] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. arXiv, 2023. 2, 4, 5, 7, 1, 15" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 206, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 227 + ], + "type": "text", + "content": "[17] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 228, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 249 + ], + "type": "text", + "content": "[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 5, 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 251, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 282 + ], + "type": "text", + "content": "[19] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "type": "text", + "content": "[20] Jaeseok Jeong, Mingi Kwon, and Youngjung Uh. Training-free content injection using h-space in diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5151-5161, 2024. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 328, + 553, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 328, + 553, + 359 + ], + "spans": [ + { + "bbox": [ + 316, + 328, + 553, + 359 + ], + "type": "text", + "content": "[21] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In CVPR, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 361, + 553, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 391 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 391 + ], + "type": "text", + "content": "[22] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. In ICLR, 2023. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 393, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 437 + ], + "type": "text", + "content": "[23] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In CVPR, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 438, + 553, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 481 + ], + "type": "text", + "content": "[24] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, pages 423-439. Springer, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 483, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 483, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 483, + 553, + 525 + ], + "type": "text", + "content": "[25] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, and Errui Ding. Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In ICCV, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 526, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 558 + ], + "type": "text", + "content": "[26] Frank J. Massey. The kolmogorov-smirnov test for goodness of fit. Journal of the American Statistical Association, 46 (253):68-78, 1951. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 559, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 602 + ], + "type": "text", + "content": "[27] Benyuan Meng, Qianqian Xu, Zitai Wang, Xiaochun Cao, and Qingming Huang. Not all diffusion model activations have been evaluated as discriminative features. arXiv preprint arXiv:2410.03558, 2024. 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 604, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 553, + 646 + ], + "type": "text", + "content": "[28] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "type": "text", + "content": "[29] Daiki Miyake, Akihiro Iohara, Yu Saito, and Toshiyuki Tanaka. Negative-prompt inversion: Fast image inversion for editing with text-guided diffusion models, 2023. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[30] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, 2023. 1" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15768" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[31] Thao Nguyen, Yuheng Li, Utkarsh Ojha, and Yong Jae Lee. Visual instruction inversion: Image editing via visual prompting. NeurIPS, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 108, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 108, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 108, + 295, + 129 + ], + "type": "text", + "content": "[32] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In CVPR, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 132, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 132, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 132, + 294, + 185 + ], + "type": "text", + "content": "[33] Yong-Hyun Park, Mingi Kwon, Jaewoong Choi, Junghyo Jo, and Youngjung Uh. Understanding the latent space of diffusion models through the lens of riemannian geometry. Advances in Neural Information Processing Systems, 36: 24129-24142, 2023. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 189, + 294, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 189, + 294, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 189, + 294, + 232 + ], + "type": "text", + "content": "[34] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 234, + 294, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 234, + 294, + 299 + ], + "spans": [ + { + "bbox": [ + 56, + 234, + 294, + 299 + ], + "type": "text", + "content": "[35] Tianhao Qi, Shancheng Fang, Yanze Wu, Hongtao Xie, Jiawei Liu, Lang Chen, Qian He, and Yongdong Zhang. Deadiff: An efficient stylization diffusion model with disentangled representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8693-8702, 2024. 1, 5, 16" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 302, + 294, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 302, + 294, + 355 + ], + "spans": [ + { + "bbox": [ + 56, + 302, + 294, + 355 + ], + "type": "text", + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 358, + 294, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 358, + 294, + 401 + ], + "spans": [ + { + "bbox": [ + 56, + 358, + 294, + 401 + ], + "type": "text", + "content": "[37] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, pages 8821-8831. PMLR, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 294, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 436 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 436 + ], + "type": "text", + "content": "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 439, + 294, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 439, + 294, + 482 + ], + "spans": [ + { + "bbox": [ + 56, + 439, + 294, + 482 + ], + "type": "text", + "content": "[39] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 1, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 485, + 294, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 485, + 294, + 517 + ], + "spans": [ + { + "bbox": [ + 56, + 485, + 294, + 517 + ], + "type": "text", + "content": "[40] Simo Ryu. Low-rank adaptation for fast text-to-image diffusion fine-tuning. https://github.com/cloneofsimo/lora, 2023.5,8,1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 520, + 294, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 294, + 551 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 294, + 551 + ], + "type": "text", + "content": "[41] Lu Sheng, Ziyi Lin, Jing Shao, and Xiaogang Wang. Avatarnet: Multi-scale zero-shot style transfer by feature decoration. In CVPR, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 554, + 294, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 554, + 294, + 608 + ], + "spans": [ + { + "bbox": [ + 56, + 554, + 294, + 608 + ], + "type": "text", + "content": "[42] Kihyuk Sohn, Lu Jiang, Jarred Barber, Kimin Lee, Nataniel Ruiz, Dilip Krishnan, Huiwen Chang, Yuzhzen Li, Irfan Essa, Michael Rubinstein, Yuan Hao, Glenn Entis, Irina Blok, and Daniel Castro Chin. Styledrop: Text-to-image synthesis of any style. In NeurIPS, 2023. 2, 5, 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 611, + 294, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 611, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 611, + 294, + 632 + ], + "type": "text", + "content": "[43] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2020. 5, 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 666 + ], + "type": "text", + "content": "[44] Adéla Šubrtová, Michal Lukáč, Jan Čech, David Futschik, Eli Shechtman, and Daniel Sykora. Diffusion image analogies. In SIGGRAPH, 2023. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[45] Yasheng Sun, Yifan Yang, Houwen Peng, Yifei Shen, Yuqing Yang, Han Hu, Lili Qiu, and Hideki Koike. Imagebrush: Learning visual in-context instructions for exemplar-based image manipulation. NeurlPS, 2023. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 669 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[46] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In CVPR, 2023. 4, 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 108, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 140 + ], + "type": "text", + "content": "[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "type": "text", + "content": "[48] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. " + }, + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "type": "inline_equation", + "content": "p+" + }, + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "type": "text", + "content": ": Extended textual conditioning in text-to-image generation. arXiv, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 175, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 196 + ], + "type": "text", + "content": "[49] Bram Wallace, Akash Gokul, and Nikhil Naik. Edict: Exact diffusion inversion via coupled transformations, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 198, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 240 + ], + "type": "text", + "content": "[50] Haofan Wang, Matteo Spinelli, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv preprint arXiv:2404.02733, 2024. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "text", + "content": "[51] Haofan Wang, Peng Xing, Renyuan Huang, Hao Ai, Qixun Wang, and Xu Bai. Instantstyle-plus: Style transfer with content-preserving in text-to-image generation. arXiv preprint arXiv:2407.00788, 2024. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 288, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 553, + 331 + ], + "type": "text", + "content": "[52] Zhouxia Wang, Xintao Wang, Liangbin Xie, Zhongang Qi, Ying Shan, Wenping Wang, and Ping Luo. Styleadapter: A single-pass lora-free model for stylized image generation. arXiv, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 333, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 333, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 333, + 553, + 364 + ], + "type": "text", + "content": "[53] Chen Henry Wu and Fernando De la Torre. A latent space of stochastic diffusion models for zero-shot image editing and guidance. In ICCV, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 366, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 553, + 420 + ], + "type": "text", + "content": "[54] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 422, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 465 + ], + "type": "text", + "content": "[55] Peng Xing, Haofan Wang, Yanpeng Sun, Qixun Wang, Xu Bai, Hao Ai, Renyuan Huang, and Zechao Li. Csgo: Content-style composition in text-to-image generation. arXiv preprint arXiv:2408.16766, 2024. 1, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 468, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 553, + 498 + ], + "type": "text", + "content": "[56] Sihan Xu, Yidong Huang, Jiayi Pan, Ziqiao Ma, and Joyce Chai. Inversion-free image editing with natural language. arXiv, 2023. 3, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 501, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 533 + ], + "type": "text", + "content": "[57] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. In SIGGRAPH Asia, 2023. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 535, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 553, + 567 + ], + "type": "text", + "content": "[58] Yuan Yao, Jianqiang Ren, Xuansong Xie, Weidong Liu, Yong-Jin Liu, and Jun Wang. Attention-aware multi-stroke style transfer, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 569, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 569, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 569, + 553, + 601 + ], + "type": "text", + "content": "[59] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv, 2023. 1, 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 603, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 634 + ], + "type": "text", + "content": "[60] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 8, 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "type": "text", + "content": "[61] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15769" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/StyleMotif_ Multi-Modal Motion Stylization using Style-Content Cross Fusion/ffae00f3-5394-4dd8-b9a4-bc3713aa96e3_content_list.json b/2025/StyleMotif_ Multi-Modal Motion Stylization using Style-Content Cross Fusion/ffae00f3-5394-4dd8-b9a4-bc3713aa96e3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1118d445a6fd9dc54e280608e06d7f399066b8b1 --- /dev/null +++ b/2025/StyleMotif_ Multi-Modal Motion Stylization using Style-Content Cross Fusion/ffae00f3-5394-4dd8-b9a4-bc3713aa96e3_content_list.json @@ -0,0 +1,1439 @@ +[ + { + "type": "text", + "text": "STYLEMOtIF: Multi-Modal Motion Stylization using Style-Content Cross Fusion", + "text_level": 1, + "bbox": [ + 91, + 130, + 906, + 154 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/2375eaaa2a709112399a1158ca8f1cc8f04570c312cf87b6077ac50dcd0ded49.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Ziyu Guo†Young Yoon LeeJoseph Liu
The Chinese University of Hong KongRobloxRoblox
ziyuguo@link.cuhk.edu.hkylee@roblox.comjosephliu@roblox.com
Yizhak Ben-ShabatVictor ZordanMubbasir Kapadia
RobloxRobloxRoblox
ibenshabat@roblox.comvbzordan@roblox.commkapadia@roblox.com
", + "bbox": [ + 156, + 178, + 843, + 300 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 333, + 326, + 349 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present STYLEMOTIF, a novel Stylized Motion Latent Diffusion model, generating motion conditioned on both content and style from multiple modalities. Unlike existing approaches that either focus on generating diverse motion content or transferring style from sequences, STYLEMOTIF seamlessly synthesizes motion across a wide range of content while incorporating stylistic cues from multi-modal inputs, including motion, text, image, video, and audio. To achieve this, we introduce a style-content cross fusion mechanism and align a style encoder with a pre-trained multi-modal model, ensuring that the generated motion accurately captures the reference style while preserving realism. Extensive experiments demonstrate that our framework surpasses existing methods in stylized motion generation and exhibits emergent capabilities for multi-modal motion stylization, enabling more nuanced motion synthesis. Project Page: https://stylemotif.github.io.", + "bbox": [ + 89, + 366, + 486, + 623 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 670, + 222, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human motion generation is a fundamental task in computer graphics and animation, enabling the synthesis of realistic and expressive human movements. Broadly, human motion can be characterized by two complementary aspects: content, which defines the underlying action (e.g., walking, jumping), and style, which encodes variations such as personal flair, emotional expression, or cultural influences (e.g., jubilant, aggressive). This separation allows for greater control and flexibility in generating motion, making it particularly valuable in creative industries like game development, film production, and virtual reality. However, traditional approaches to stylized motion generation often depend on manual pro", + "bbox": [ + 89, + 693, + 486, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a2c6ee5da755e8bdbcefd5d585a697bc8bbc4745f64d4ec6b582eee470f116b3.jpg", + "image_caption": [ + "Figure 1. Comparison of Our Proposed STYLEMOTIF with SMooDi. SMooDi duplicates and separates two diffusion backbones for content and style, increasing complexity. In contrast, STYLEMOTIF adopts a single diffusion backbone with direct style injection, enabling efficient and realistic motion stylization." + ], + "image_footnote": [], + "bbox": [ + 517, + 318, + 903, + 530 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cesses such as motion capture or keyframe animation, which are costly, time-consuming, and labor-intensive.", + "bbox": [ + 511, + 627, + 906, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent progress in text-to-motion (T2M) diffusion frameworks [6, 27, 54] has greatly advanced the ability to translate natural-language prompts into realistic human motions. By leveraging powerful denoising diffusion models, these approaches capture intricate spatiotemporal dependencies in the data, enabling coherent sequences of human movements to be generated directly from brief textual descriptions. Despite their success in content fidelity and diversity, most current T2M diffusion methods concentrate primarily on what action is performed, while overlooking how it is performed, namely, its stylistic details. Simply appending a separate style-transfer module to text-driven motion diffusion pipelines can introduce additional complexity and risk of compounding errors in the final output.", + "bbox": [ + 509, + 657, + 908, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In parallel, motion style transfer has been actively studied to infuse stylistic cues from a reference motion (or style", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Work done as an intern at Roblox.", + "bbox": [ + 114, + 886, + 307, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13349", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "data) into another sequence [1, 24, 37, 53]. Although many of these methods effectively disentangle content and style for small-scale tasks, the pipeline becomes cumbersome when a large variety of content motions need to be stylized. Moreover, they commonly assume that the input or target sequences are high-quality motion data. In scenarios where content motions are synthetically generated, or partially noisy, the transfer process can deteriorate, leading to undesirable motion artifacts or compromised style fidelity.", + "bbox": [ + 89, + 90, + 480, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the need for simultaneously controlling both content and style, some recent works have merged style encoding with diffusion-based motion generation. Among these, the most recent and representative approach [73] augments a pre-trained latent diffusion model [6] with a style adaptor and classifier-based style guidance, achieving stylized motion from textual prompts and motion-style references. While effective, this method relies on additional training diffusion backbones, which shares structural similarities with ControlNet [66] as shown in Figure 1, which increases model complexity and training overhead. It is also constrained to motion as the primary style input. Concurrent work [33] proposes a bidirectional control flow mechanism to mitigate conflicts between style and content, extending style control to multiple modalities, but this approach also duplicates two separate diffusion backbones with substantial training overhead and limited applicability across modalities.", + "bbox": [ + 91, + 229, + 485, + 486 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we propose STYLEMOTIF, a new framework for multi-modal motion stylization that unifies text-to-motion diffusion with style conditioning in a single-diffusion structure. Specifically, we leverage a pre-trained motion latent diffusion model (MLD) [6] to preserve strong content generation capabilities, and seamlessly integrate style features extracted from a dedicated encoder, which is aligned with a multi-modal foundation model [12]. In contrast to previous works, STYLEMOTIF avoids duplicating large portions of the network or relying on specialized style diffusion network. Instead, we introduce a style-content cross fusion mechanism, which injects stylistic cues into the diffusion process while maintaining motion realism. As a result, our STYLEMOTIF not only yields more robust stylized outputs but also supports diverse style signals, such as motion, text, images, audio, or video clips, via the alignment in multi-modal feature space. We summarize our main contributions as follows:", + "bbox": [ + 89, + 488, + 483, + 744 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present STYLEMOTIF, a stylized latent motion diffusion framework that unifies diverse motion content and multi-modal styles within a compact single-diffusion design.", + "- We propose a style-content cross fusion module that injects stylistic cues into the diffusion denoising process, achieving faithful stylization without compromising motion realism and ensuring efficiency.", + "- We achieve a unified multi-modal style feature space and unveil new emergent capabilities through multi-modal" + ], + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "alignment, which accommodate various sources including motion, text, images, audio, and video, for flexible and versatile multi-modal style control.", + "bbox": [ + 526, + 90, + 903, + 135 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Extensive experiments demonstrate that STYLEMOTIF consistently outperforms existing methods regarding style expressiveness, content preservation, and efficiency.", + "bbox": [ + 513, + 136, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 193, + 653, + 208 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human Motion Generation. Recent progress in human motion generation [3, 5, 8, 13, 14, 41-43, 46, 55, 57, 58, 62] has been driven by transformer [14, 34] and diffusion models [2, 8, 49, 66], showing great potential in producing realistic and diverse motions. These approaches have shown great potential in producing realistic and diverse motion sequences. For example, Momask [14] improves motion generation using a residual VQ-VAE. Similarly, LaMP [34] introduces a motion-aware text encoder and a motion-to-text language model to enhance motion quality through text conditioning. Diffusion models, in particular, have become a key approach in motion generation [6, 22, 27, 48, 54, 61, 65, 67]. MDM [54] introduces a motion diffusion model that operates directly on raw motion data to capture the relationship between motions and input text conditions. MLD [6] improve efficiency by embedding the diffusion process in latent space, reducing computational cost. They also allow conditioning on specific constraints, such as predefined trajectories [27, 55, 59] or human-object interactions [7, 40, 57], enabling greater control and diversity. Our work leverages pretrained motion latent diffusion model [6] and multi-modal foundation models [12] to achieve stylized human motion generation while maintaining high motion quality.", + "bbox": [ + 511, + 218, + 906, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motion Stylization. Motion stylization [1, 24, 28, 37, 39, 44, 45, 50, 51, 53, 56, 60] involves transferring stylistic features from a reference motion to a source motion/textual prompt, enabling creative transformations while preserving the original motion content. Early methods, like those in motion style transfer [1, 24], typically separate motion content and style for recombination. For instance, Aberman et al. [1] used a generative adversarial network to decouple style from content without paired data, while Motion Puzzle [24] allows style control for individual body parts, and Guo et al. [15] utilized pretrained motion models for better style integration. Recent approaches, like SMooDi [73], generate stylized motion from text and style sequences using style guidance. However, these models face two main challenges: (1) inefficient architectures that duplicate diffusion backbones for content and style, and (2) limited to style motion for guidance. Another concurrent approach, MulSMo [33], attempts to address some of these by introducing a bidirectional control flow between style and content networks, reducing conflicts. Still, it remains restricted to a few modalities and duplicates diffusion networks. In this", + "bbox": [ + 511, + 583, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13350", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e43eacda52e40f7d1fbd364185ec40d421eb98e999aea47c2f37b9f95de4d56e.jpg", + "image_caption": [ + "Figure 2. Overall Pipeline of STYLEMOTIF, a diffusion framework for multi-modal motion stylization. Given a text prompt and a reference style from various modalities, our model extract style features and fuse them with content by style-content cross fusion. Through multi-modal alignment with contrastive learning, we enable seamless multi-modal conditioning and flexible stylization across motion, text, images, audio, and video." + ], + "image_footnote": [], + "bbox": [ + 98, + 66, + 908, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "work we eliminate redundant diffusion backbone, resulting in a simpler, more effective, and efficient approach to stylized motion generation. Our method also enables efficient multi-modal motion stylization, supporting text, image and multi-modal inputs while maintaining both high stylization quality and motion realism.", + "bbox": [ + 88, + 450, + 485, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-modality Learning. Recent advancements in multimodality learning [9, 10, 38] have revolutionized various domains by enabling joint understanding across diverse data types. Foundational models like CLIP [47] and its extension [16, 68, 69, 75] establish robust vision-language alignment, while ImageBind [12], Point-Bind [17], and Language-Bind [74] expand this paradigm to diverse modalities, demonstrating emergent cross-modal capabilities. The rise of multi-modal large language models [4, 19, 21, 30, 31, 64, 72] further enhances semantic understanding through unified text-visual processing, achieving superior performance across 2D images [11, 29, 70], 3D point clouds [17, 18, 52], and complex reasoning scenarios [20, 26, 71]. For human motion, some approaches [5, 14, 25] incorporate multiple modalities into human motion understanding and generation tasks. Our work introduces the first framework for multimodal guided motion stylization, achieving seamless style injection from diverse modalities through a single diffusion architecture with style-content cross fusion.", + "bbox": [ + 88, + 613, + 486, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. STYLEMOTIF: Multi-Modal Motion Stylization with Style-Content Cross Fusion", + "text_level": 1, + "bbox": [ + 511, + 449, + 908, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose STYLEMOTIF, a novel framework for stylized motion synthesis that combines style-content cross fusion and multi-modal motion stylization, as illustrated in Figure 2. Our approach integrates a style-content cross fusion mechanism that allows for coherent feature blending, ensuring the generated motion accurately reflects the reference style while maintaining the content's realism (§3.2). This mechanism is complemented by a multi-modal motion stylization strategy, which leverages inputs from diverse modalities such as image, video, audio, and text to provide control over the stylization process (§3.3). This framework enables highly flexible and realistic stylized motion synthesis, which offers greater control and diversity.", + "bbox": [ + 511, + 492, + 908, + 690 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 511, + 696, + 624, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preliminaries. Motion Latent Diffusion (MLD) [6] formulates a conditional latent diffusion model by training $\\epsilon_{\\theta}(z_t,t,c)$ to denoise a sequence of latents $\\{z_t\\}_{t = 0}^T$ , where $z_{t}\\in \\mathbb{R}^{n\\times d}$ represents the motion latent at timestep $t$ , conditioned on the distribution $p(z_{t}|c)$ . A conditional domain encoder $\\tau_{\\theta}(c)$ , such as CLIP [47], enables text-to-motion tasks, and the model is trained as $\\epsilon_{\\theta}(z_t,t,\\tau_{\\theta}(c))$ . To incorporate additional style conditioning, a style condition $s$ can be introduced with its own encoder $\\psi_{\\theta}(s)$ , which may take the form of a motion or text encoder. This extends the model to $\\epsilon_{\\theta}(z_t,t,\\tau_{\\theta}(c),\\psi_{\\theta_s}(s))$ . SMooDi [73] adopts a ControlNet [66]-style approach for style conditioning by", + "bbox": [ + 511, + 719, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13351", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "creating a trainable copy of the neural network weights $\\theta_{s}$ from the original MLD model $\\theta_{c}$ . This copy includes zero-initialized linear layers, denoted as $\\mathcal{Z}(\\cdot ;\\cdot)$ . The output $\\mathcal{F}^i (z_t,t,\\tau_\\theta (c),\\psi_\\theta (s);\\theta_c)$ of the $i$ -th MLD block, now conditioned on style, is computed as:", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {F} ^ {i} \\left(z _ {t}, t, \\tau_ {\\theta} (c), \\psi_ {\\theta_ {s}} (s); \\theta_ {c}\\right) = \\mathcal {F} ^ {i} \\left(z _ {t}, t, \\tau_ {\\theta} (c); \\theta_ {c}\\right) \\\\ + \\mathcal {Z} \\left(\\mathcal {F} ^ {i} \\left(z _ {t}, t, \\tau_ {\\theta} (c), \\psi_ {\\theta_ {s}} (s); \\theta_ {s}\\right); \\theta_ {z _ {i}}\\right) \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 178, + 482, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A key property of this formulation is that since $\\theta_{z_i}$ is initialized to zero, thus $\\mathcal{F}^i (z_t,t,\\tau_\\theta (c),\\psi_\\theta (s)) = \\mathcal{F}^i (z_t,t,\\tau_\\theta (c))$ at the beginning. However, the tradeoff of that method is that it needs to maintain the additional parameters $\\theta_{s}$ and $\\theta_{z_i}$ in order for style transfer.", + "bbox": [ + 89, + 227, + 483, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "STYLEMOTIF. We propose STYLEMOTIF, eschewing zero linear layers in favor of injecting it directly via statistical manipulation, with the pipeline shown in Figure 2 describing our approach. Our STYLEMOTIF utilizes latent space diffusion within a single generative diffusion backbone, building upon a pretrained MLD model [6]. Instead of perturbing the outputs of each block $i$ via zero initialized layer $\\mathcal{Z}(\\mathcal{F}^i (z_t,t,\\tau_\\theta (c),\\psi_\\theta (s);\\theta_s);\\theta_{z_i})$ , STYLEMOTIF replaces this with a statistically transformed style embedding that is injected into the original MLD branch. This simplifies the model while ensuring high-quality stylization results, whose details are presented in § 3.2. Also, we follow SMooDi's generation guidance and training scheme to ensure high-quality stylized motion synthesis. We use a hybrid guidance strategy that balances content fidelity and style adherence by combining classifier-free and classifier-based techniques during diffusion sampling.", + "bbox": [ + 89, + 321, + 483, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Style-Content Cross Fusion", + "text_level": 1, + "bbox": [ + 89, + 588, + 336, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Style Encoder Pre-training. To establish a robust foundation for style-content fusion, we combine the content knowledge from the pre-trained MLD's [6] VAE with the style knowledge from the 100STYLE [37] dataset. The MLD's VAE is pre-trained on the HumanML3D [13] dataset, which provides extensive understanding of content motion. Building on this, we further fine-tune the model on the 100STYLE dataset in a variational autoencoding manner to initially align the content and style data distributions in the latent space. After training, we discard the decoder and retain only the encoder as the motion style encoder. This reconstruction task enables the encoder to learn robust motion feature representations, which are essential for supporting the subsequent stylization process. By integrating content and style knowledge in this way, we ensure a strong foundation for seamless style-content fusion.", + "bbox": [ + 89, + 609, + 483, + 851 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Style-Content Cross Normalization. To train a stylized diffusion model $\\epsilon_{\\theta}(z_t, t, \\tau_{\\theta}(c), \\psi_{\\theta}(s))$ , we effectively fuse", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "content and style features directly within the latent space diffusion process instead of duplicating two pre-trained diffusion backbones for content and style or separate control mechanisms [66]. Prior work [23, 32] has proven that feature statistics (mean, variance) effectively encode style. Given the output features $\\mathcal{F}^i$ of the $i$ -th block, we derive the content features $\\mathcal{F}_c^i = \\mathcal{F}^i(z_t, t, \\tau_\\theta(c); \\theta_c)$ from the input text $c$ , and the style features $\\mathcal{F}_s = \\psi_{\\theta_s}(s)$ are extracted from the reference motion sequence $s$ using the pre-trained style encoder. To perform the fusion, we first compute the mean $\\mu_c$ and variance $\\sigma_c^2$ of the content features across the feature dimension:", + "bbox": [ + 511, + 90, + 906, + 272 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {c} = \\frac {1}{D} \\sum_ {j = 1} ^ {D} \\mathcal {F} _ {c} ^ {i, j}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 296, + 906, + 339 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {c} ^ {2} = \\frac {1}{D} \\sum_ {j = 1} ^ {D} \\left(\\mathcal {F} _ {c} ^ {i, j} - \\mu_ {c}\\right) ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 343, + 906, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{F}_c^{i,j}$ denotes the content features of the $i$ -th block and the $j$ -th feature element. Next, we normalize the style features $\\mathcal{F}_s$ using these content statistics, ensuring that the style features are adapted to the content's statistical properties:", + "bbox": [ + 511, + 396, + 906, + 457 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {\\mathcal {F}} _ {s, c} = \\frac {\\mathcal {F} _ {s} - \\mu_ {c}}{\\sqrt {\\sigma_ {c} ^ {2} + \\eta}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 645, + 468, + 906, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\eta$ is a constant added for numerical stability. This style-content cross normalization ensures that the style features are smoothly integrated with the content features while maintaining the content's original structure. After that, we add the normalized style features to the content features, formulating our final cross-normalization as", + "bbox": [ + 511, + 512, + 908, + 602 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} ^ {i} \\left(z _ {t}, t, \\tau_ {\\theta} (c), \\psi_ {\\theta_ {s}} (s); \\theta_ {c}\\right) = \\mathcal {F} _ {c} ^ {i} + \\gamma \\cdot \\tilde {\\mathcal {F}} _ {s, c}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 613, + 906, + 631 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\gamma$ is a parameter used for scaling the normalized value. $\\widetilde{\\mathcal{F}}_{s,c}$ can be thought of as a perturbation of $\\mathcal{F}_c^i$ as it is scaled within its range. Notably, the fusion process is performed only once after the $m$ -th block during the denoising process to avoid distorting the content while effectively introducing the style. This efficient fusion method eliminates the need for additional learnable parameters, as it is based solely on statistical transformations, achieving high-quality results with minimal computational overhead.", + "bbox": [ + 511, + 642, + 906, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Multi-Modal Motion Stylization", + "text_level": 1, + "bbox": [ + 511, + 787, + 795, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our STYLEMOTIF extends to support multi-modal motion stylization. To achieve this, we integrate a pre-trained multimodal foundation model [12], which provides a unified, multi-modal aligned feature space, enabling effective cross-modal alignment between motion and other modalities. With this alignment, our model can flexibly combine multiple", + "bbox": [ + 511, + 809, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13352", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "input modalities, such as text, image, audio, and video, to guide the stylization process in a comprehensive manner. This capability results in emergent abilities for multi-modal motion stylization, allowing for more nuanced outputs.", + "bbox": [ + 89, + 90, + 485, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Motion-Text Pair Curation. To align motion with other modalities, we process a set of motion-text pairs using the curated 100STYLE subset [37], which is carefully selected to avoid conflicts between content and style motions, so that the model can effectively learn the relationships between motion and text. Each motion sequence is paired with a corresponding single textual label, which serves as the text prompt. The curated motion-text pairs will be used for the following cross-model alignment.", + "bbox": [ + 89, + 170, + 483, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-Modal Alignment. To maintain the alignment within the multi-modal space, we freeze the text encoder of ImageBind [12] and introduce a lightweight projection layer after the encoder. This projection layer aligns the feature dimensions of the text and motion encoders. Since the multi-modal model represents each modality with a global feature, our alignment process focuses on these global features to achieve robust alignment between text and motion representations, formalized as:", + "bbox": [ + 89, + 325, + 485, + 460 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {t} = \\pi \\left(\\mathcal {E} _ {\\text {t e x t}} (l)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 473, + 483, + 489 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {s} = \\psi_ {\\theta_ {s}} (s), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 492, + 483, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $l$ and $s$ are the paired input text label and style motion, $\\mathcal{F}_t$ and $\\mathcal{F}_s$ denote the feature representations from the text and style motion encoders, $\\mathcal{E}_{\\mathrm{text}}$ and $\\psi_{\\theta_s}$ , respectively, and $\\pi$ represents the projection operation that maps text features into the motion feature space. We employ a contrastive learning loss [63] to align the feature spaces of motion and text, bringing them closer together in the shared multi-modal space. Thus, we obtain a unified space between motion and all the modalities. We formulated it as", + "bbox": [ + 89, + 521, + 483, + 656 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l i g n}} = - \\frac {1}{2} \\sum_ {(i, j)} \\log \\frac {\\exp \\frac {\\mathcal {F} _ {t} ^ {i} \\cdot \\mathcal {F} _ {s} ^ {j}}{\\tau_ {0}}}{\\sum_ {k} \\exp \\frac {\\mathcal {F} _ {t} ^ {i} \\cdot \\mathcal {F} _ {s} ^ {k}}{\\tau_ {0}}} + \\log \\frac {\\exp \\frac {\\mathcal {F} _ {t} ^ {i} \\cdot \\mathcal {F} _ {s} ^ {j}}{\\tau_ {0}}}{\\sum_ {k} \\exp \\frac {\\mathcal {F} _ {t} ^ {k} \\cdot \\mathcal {F} _ {s} ^ {j}}{\\tau_ {0}}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 665, + 483, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $t$ and $s$ represent two modalities (text and style motion) and $(i,j)$ indicates a positive pair in each training batch, $k$ indexes all samples in the batch, including both positive and negative ones, and $\\tau_0$ is a temperature parameter. During inference, we obtain the multi-modal (text, image, video, or audio) features from the multi-modal input $m$ ,", + "bbox": [ + 89, + 739, + 485, + 829 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {m} = \\mathcal {E} _ {\\text {I m a g e B i n d}} (m), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 842, + 483, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{E}_{\\mathrm{ImageBind}}$ denotes ImageBind [12]. We use the multimodal features $\\mathcal{F}_m$ to retrieve the most semantically similar", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "motion features from the unified multi-modal space. The retrieved motion features are then used to guide the stylization process, ensuring that fine-grained style details are preserved and accurately reflected in the final output. This approach enhances the model's ability to generate stylized motions that are both contextually relevant and visually coherent.", + "bbox": [ + 511, + 90, + 906, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 511, + 196, + 638, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We provide extensive quantitative and qualitative analyses across multiple tasks in this section.", + "bbox": [ + 511, + 220, + 905, + 251 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 511, + 261, + 720, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We adopt the pre-trained MLD [6] as the foundation for motion generation. The style encoder of our model derives from the encoder of MLD's VAE, while the projection layer after text encoder is a single Linear layer. During diffusion training, we only enable the style encoder to be trainable while freezing other parameters. The model is optimized using the AdamW optimizer [35] with a constant learning rate of $10^{-5}$ .", + "bbox": [ + 511, + 284, + 906, + 404 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset Settings. We use the HumanML3D [13] dataset as our primary motion content dataset, which collects 14,616 motion sequences from AMASS [36] and annotates 44,970 sequence-level textual description. To train the style network, we utilize the 100STYLE [37] dataset, containing 45,303 style motions. We also adopt the text annotations for 100STYLE from previous work [73], which are pseudo text descriptions generated from MotionGPT [25]. We utilize the consistent root-velocity motion representations from HumanML3D for both the content and style data.", + "bbox": [ + 511, + 422, + 908, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We mainly evaluate our method against SMooDi [73] on motion-guided stylization and motion style transfer tasks. We also compare our text-guided stylcification with 'ChatGPT+MLD' approach, which utilize ChatGPT to combine style text and content text, and functions as a straightforward text-to-motion model without control capabilities [73]. Additionally, we qualitatively compare our model with SMooDi on motion-guided stylization.", + "bbox": [ + 511, + 594, + 908, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. To evaluate our model, we use several metrics following previous works [73]. We measure R-precision, Multi-modal Distance (MM Dist), Diversity, and Frechet Inception Distance (FID) to assess how well the content from text is preserved and how realistic the generated motions are. We use Style Recognition Accuracy (SRA) to evaluate how accurately the style of the reference motion is reflected in the generated motion. Also, we adopt the Foot Skate Ratio, which helps assess the realism of the generated motion, reducing artifacts like foot sliding. During evaluation, following previous works [73], we randomly select", + "bbox": [ + 511, + 734, + 910, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13353", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2df0e4fdcf7720c903310ca0a11e82d081ce87f75f4d1c449d6771fcdd07fb94.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSRA ↑FID ↓MM Dist ↓R-Precision ↑ (Top-3)Diversity →Foot Skate Ratio ↓
Motion-Guided Stylization
MLD + Aberman [1]54.373.3095.9830.4068.8160.347
MLD + Motion Puzzle [24]63.776.1276.4670.2906.4760.185
SMooDi [73]72.421.6094.4770.5719.2350.124
STYLEMOTIF (Ours)77.651.5514.3540.5867.5670.097
Text-Guided Stylization
MLD + ChatGPT [73]4.820.6144.3130.6058.8360.131
STYLEMOTIF (Ours)56.710.6033.6840.6909.1010.101
", + "bbox": [ + 98, + 80, + 580, + 210 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6a35f697f0454caa9794f9e526c24280359fa8fc7f3efff1bf3899a2418650a1.jpg", + "table_caption": [ + "Table 1. Quantitative Results for Motion-Guided and Text-Guided Stylization. Bold values denote the best performance. As there is no ground-truth reference for Diversity, no value is highlighted in bold; but the metric is provided for reference." + ], + "table_footnote": [], + "table_body": "
MethodSRA ↑FID ↓Foot Skate Ratio ↓
MLD + Aberman [1]61.013.8920.338
MLD + Motion Puzzle [24]67.236.8710.197
SMooDi [73]65.151.5820.095
STYLEMOTIF (Ours)68.811.3750.094
", + "bbox": [ + 598, + 79, + 888, + 157 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative Results of Motion Style Transfer on HumanML3D [13] dataset. Our method outperforms previous works in all metrics, which demonstrates effective style-content fusion for high-quality motion style transfer, providing significant advantages for downstream tasks besides motion stylization.", + "bbox": [ + 588, + 166, + 897, + 263 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/20b2c436f31ed23d4ea56d7967d06c6d7253e3f3a8fa14e2bf3162f090148f20.jpg", + "image_caption": [ + "Figure 3. Qualitative Results of Motion-Guided Stylization. Our model generates cohesive and realistic motions that effectively align style and content, such as preserving the 'circular' trajectory (first column) and 'hop' content (third column). In contrast, SMooDi [73] struggles to maintain content fidelity and sometimes fails to reflect the specified style (e.g., 'phone on the left' in the second column)." + ], + "image_footnote": [], + "bbox": [ + 94, + 273, + 906, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "a content text from HumanML3D [13] and a style motion from 100STYLE [37], and compute SRA for the generated motion using a pre-trained classifier [73].", + "bbox": [ + 89, + 684, + 482, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Motion-guided Stylization", + "text_level": 1, + "bbox": [ + 89, + 741, + 328, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Analysis. In Table 1, we report the quantitative results for motion-guided stylization, where the motion serves as style input while text prompt as content one. Our method outperforms three baseline approaches [1, 24, 73] in all metrics. Specifically, we achieve a $5.23\\%$ improvement in SRA while maintaining competitive performance in FID compared to the best baseline [73]. It demonstrates that our method, by effectively utilizing style-content cross fusion, generates motions that better align with the style reference", + "bbox": [ + 88, + 763, + 485, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "while maintaining content integrity, showcasing the strength of our design in balancing style and content preservation.", + "bbox": [ + 511, + 684, + 906, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Analysis. In Figure 3, we show some qualitative results of motion-guided stylization from our model and baseline [73]. Compared with the baseline, our model produces more cohesive stylized motions, with better alignment of style and content. For instance, in the provided examples, SMooDi fails to maintain the 'circular' trajectory or 'hop' action specified by the content or cannot accurately reflect the intended style, 'phone on the left'. This indicates that our approach more effectively integrates style and content, resulting in more realistic and consistent stylized motions.", + "bbox": [ + 511, + 750, + 910, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13354", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a54fce626f941968a92bdb3eb80e90a89f5fe555534ab4bbc9641437e19dfd07.jpg", + "image_caption": [ + "Figure 4. Qualitative Results of Text-Guided Stylization. Our model seamlessly integrates textual style descriptions with content, producing visually coherent and stylistically consistent results.", + "Figure 5. Qualitative Results of Style Interpolation. Our model blends multiple style inputs while preserving content integrity, demonstrating effective style-content fusion of our model." + ], + "image_footnote": [], + "bbox": [ + 94, + 78, + 919, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e9bc3135ae16e288144a830654b13ddd14b9d2787f6006ea4a9e418075dfdc81.jpg", + "image_caption": [ + "Figure 6. Qualitative Results of Multi-Modal Motion Stylization. Our model generates stylized motions guided by diverse modalities (e.g., text, image, video, audio), effectively transferring style while preserving content integrity." + ], + "image_footnote": [], + "bbox": [ + 94, + 311, + 906, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Text-guided Stylization", + "text_level": 1, + "bbox": [ + 89, + 580, + 305, + 598 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Analysis. For the text-guided stylization task, we use the HumanML3D [13] dataset for motion content and employ the single text label [37] as the style control. As reported in Table 1 (Bottom), our method significantly outperforms the baseline, achieving $56.71\\%$ in SRA, compared to $4.82\\%$ for 'ChatGPT + MLD'. Moreover, our method maintains competitive FID, balancing style reflection with content preservation. This indicates that our design, which combines multi-modal alignment, helps utilize the style signal in the motion-text shared spaces and achieve significant effectiveness in text-guided stylization.", + "bbox": [ + 88, + 603, + 485, + 770 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Analysis. In Figure 4, we showcase qualitative results for text-guided stylization, where our model also demonstrates strong capability in harmonizing style and content, generating high-quality and visually coherent results.", + "bbox": [ + 88, + 780, + 485, + 842 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Motion Style Transfer", + "text_level": 1, + "bbox": [ + 89, + 848, + 297, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the motion style transfer task, we use the HumanML3D [13] dataset for motion content and the", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "100STYLE dataset [37] for motion styles. As shown in Table 2, our method outperforms the baseline models across all key metrics. Specifically, we achieve a $3.66\\%$ improvement in SRA, indicating better style reflection, and a 0.207 reduction in FID reflecting improved realism. These demonstrate that our framework not only improves the alignment between style and content but also provides significant advantages for downstream tasks like motion style transfer with efficient style-content cross fusion. By maintaining content integrity while seamlessly integrating style, STYLEMOTIF ensures more realistic and dynamically consistent stylized motions, which is critical for high-quality motion style transfer.", + "bbox": [ + 511, + 582, + 908, + 765 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Multi-Modal Motion Stylization", + "text_level": 1, + "bbox": [ + 511, + 772, + 797, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With the power of the aligned multi-modal space, our model supports stylization guided by a variety of modalities, including motion, text, image, video, audio. Here, we utilize the input style feature to retrieve the corresponding motion features as the style condition, while the content input is provided as text prompt. The model then generates a stylized motion that incorporates the style from the input modality", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13355", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ef6e357332c4ece566d7503d051fb462663a343173f66097cd6cbb344ff33fac.jpg", + "image_caption": [ + "Figure 7. Ablation Study on Scaling Ratio $\\gamma$ in Eq. 5. We report both SRA and FID to show the impact of the scaling ratio on both stylization and content preservation." + ], + "image_footnote": [], + "bbox": [ + 101, + 74, + 316, + 193 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/163f0ce532f297f2de08cd6880d54fc409f279a3ab7e7a5abe9eed4611c3cf63.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Style Encoder Pre-training Strategy
HumanML3D100STYLESRA ↑FID ↓MM Dist ↓R-Precision ↑Diversity →Foot Skate Ratio ↓
-✓76.731.7884.3490.5717.5050.101
✓-76.581.6354.4580.5728.5340.109
✓✓77.651.5514.3540.5867.5670.097
Text in Multi-modal Alignment
Textual ExpressionSRA ↑FID ↓MM Dist ↓R-Precision ↑Diversity →Foot Skate Ratio ↓
Brief Description76.841.5804.3780.5787.2510.099
Detailed Description75.251.6224.4190.5637.7640.102
Single Text Label77.651.5514.3540.5867.5670.097
", + "bbox": [ + 344, + 71, + 890, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation Study on Style Encoder Pre-training Strategies and Text Expression for Multi-modal Alignment. 'w. HumanML3D' and 'w. 100STYLE' denote pre-training with HumanML3D [13] and 100STYLE [37] data respectively.", + "bbox": [ + 341, + 234, + 890, + 277 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "while maintaining the content integrity as specified by the text prompt. We showcase several examples of multi-modal motion stylization in Figure 6, where different modalities guide the motion generation. For instance, when a text content \"A person is walking.\" is provided alongside an image of a duck as style input, the model retrieves a relevant motion feature and blends the content with the style of 'Duckfoot'. As shown, the style from various inputs (text, image, video, audio) is effectively transferred to the generated motion.", + "bbox": [ + 88, + 295, + 485, + 434 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Style Interpolation", + "text_level": 1, + "bbox": [ + 89, + 443, + 272, + 459 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Leveraging the aligned multi-modal space, our model enables text-guided style interpolation. Given one content text along with at least two style style texts, our model generates a motion that combines the characteristics of all input styles. When style texts are provided, we retrieve the most similar style motion features from the shared space. These features are then combined by weighted summation. The combined style features are fused with the content features using the proposed style-content cross fusion. In Figure 5, we showcase some qualitative results of style interpolation in, where the generated motions successfully blend the characteristics of both styles while maintaining the content's integrity.", + "bbox": [ + 88, + 465, + 483, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.7. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 657, + 243, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Style Encoder Pre-training. In Table 3, we investigate the impact of different pre-training strategies for the style encoder. Our results show that pre-training on both HumanML3D [13] and 100STYLE [37] yields the best performance. Compared to pre-training solely on either one, the combined training on both datasets provides the style encoder with a richer set of prior knowledge. This enables the model to effectively balance content preservation and style reflection, leveraging the diverse characteristics of both datasets to enhance the overall stylization quality.", + "bbox": [ + 88, + 681, + 485, + 834 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cross Normalization Scaling Ratio. In Figure 7, we examine the effect of different scaling ratios $\\gamma$ in Eq. 5 on the stylization performance. As shown, the model achieves", + "bbox": [ + 88, + 854, + 485, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the best performance with $\\gamma = 0.6$ , striking an optimal balance between style reflection and content preservation. The choice of scaling ratio influences both SRA and FID, where the optimal value improves stylization without sacrificing content integrity.", + "bbox": [ + 511, + 295, + 908, + 372 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Text Expression for Multi-modal Alignment. In Table 3, we also explored different text representations for contrastive learning alignment. Specifically, we tested single labels (e.g., \"Old\"), brief descriptions (e.g., \"An old person\"), and more detailed descriptions (e.g., \"An old person is moving slow and stiff\"). Our experiments show that using a single label produced the best results, providing a clearer and more concise style signal for the model while avoiding unnecessary complexity in the text representation.", + "bbox": [ + 509, + 400, + 908, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 556, + 633, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduce STYLEMOTIF, a novel stylized motion diffusion model capable of generating motion conditioned on both content and style from multiple modalities. Unlike prior approaches that either focus on motion generation across various content types or style transfer between sequences, STYLEMOTIF effectively synthesizes motion while incorporating stylistic cues from multi-modal inputs, including text, image, video, and audio. To achieve this, we introduce a style-content cross fusion mechanism and align a style encoder with a pre-trained multi-modal model. Through extensive experiments across diverse applications, we demonstrate that STYLEMOTIF outperforms existing methods in stylized motion generation, producing high-quality, realistic results that faithfully adhere to the given style references. Moreover, our model exhibits emergent capabilities for multi-modal motion stylization, enabling richer and more nuanced motion synthesis. These findings indicate the potential of STYLEMOTIF in advancing stylized motion generation and open new avenues for future research in multi-modal-driven motion synthesis and style-aware generative models.", + "bbox": [ + 511, + 583, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13356", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We would like to extend our sincere thanks to the Core AI and Avatar Movement team at Roblox for their guidance and support throughout this project. In particular, we are grateful to Haomiao Jiang, Vincent Petrella, Sean Palmer, and Simone Guggiari for their valuable feedback and insights. We also thank Lei Zhong and Zhe Li for their technical advice.", + "bbox": [ + 89, + 116, + 485, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 224, + 187, + 239 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] K. Aberman, Y. Weng, D. Lischinski, D. Cohen-Or, and B. Chen. Unpaired motion style transfer from video to animation. TOG, 2020. 2, 6", + "[2] Nefeli Andreou, Xi Wang, Victoria Fernández Abrevaya, Marie-Paule Cani, Yiorgos Chrysanthou, and Vicky Kalogeiton. Lead: Latent realignment for human motion diffusion. arXiv preprint arXiv:2410.14508, 2024. 2", + "[3] Z. Cen, H. Pi, S. Peng, Z. Shen, M. Yang, S. Zhu, H. Bao, and X. Zhou. Generating human motion in 3d scenes from text descriptions. In CVPR, 2024. 2", + "[4] Kexin Chen, Yuyang Du, Tao You, Mobarakol Islam, Ziyu Guo, Yueming Jin, Guangyong Chen, and Pheng-Ann Heng. Llm-assisted multi-teacher continual learning for visual question answering in robotic surgery. ICRA 2024, 2024. 3", + "[5] L.H. Chen, S. Lu, A. Zeng, H. Zhang, B. Wang, R. Zhang, and L. Zhang. Motionllm: Understanding human behaviors from human motions and videos. ArXiv, 2024. 2, 3", + "[6] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In CVPR, 2023. 1, 2, 3, 4, 5", + "[7] Sammy Christen, Shreyas Hampali, Fadime Sener, Edoardo Remelli, Tomas Hodan, Eric Sauser, Shugao Ma, and Bugra Tekin. Diffh2o: Diffusion-based synthesis of hand-object interactions from textual descriptions. In SIGGRAPH Asia 2024 Conference Papers, 2024. 2", + "[8] Wenxun Dai, Ling-Hao Chen, Jingbo Wang, Jinpeng Liu, Bo Dai, and Yansong Tang. Motionlcm: Real-time controllable motion generation via latent consistency model. arXiv preprint arXiv:2404.19759, 2024. 2", + "[9] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11162-11173, 2021. 3", + "[10] Han Fang, Pengfei Xiong, Luhui Xu, and Yu Chen. Clip2video: Mastering video-text retrieval via image clip. arXiv preprint arXiv:2106.11097, 2021. 3", + "[11] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xiangyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023. 3", + "[12] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15180-15190, 2023. 2, 3, 4, 5" + ], + "bbox": [ + 93, + 250, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, 2022. 2, 4, 5, 6, 7, 8", + "[14] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In CVPR, 2024. 2, 3", + "[15] Chuan Guo, Yuxuan Mu, Xinxin Zuo, Peng Dai, Youliang Yan, Juwei Lu, and Li Cheng. Generative human motion stylization in latent space. arXiv preprint arXiv:2401.13505, 2024. 2", + "[16] Ziyu Guo, Renrui Zhang, Longtian Qiu, Xianzheng Ma, Xu-peng Miao, Xuming He, and Bin Cui. Calip: Zero-shot enhancement of clip with parameter-free attention. AAAI 2023 Oral, 2022. 3", + "[17] Ziyu Guo, Renrui Zhang, Xiangyang Zhu, Yiwen Tang, Xianzheng Ma, Jiaming Han, Kexin Chen, Peng Gao, Xianzhi Li, Hongsheng Li, et al. Point-bind & point-llm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. arXiv preprint arXiv:2309.00615, 2023. 3", + "[18] Zilu Guo, Hongbin Lin, Zhihao Yuan, Chaoda Zheng, Pengshuo Qiu, Dongzhi Jiang, Renrui Zhang, Chun-Mei Feng, and Zhen Li. Pisa: A self-augmented data engine and training strategy for 3d understanding with large models. arXiv preprint arXiv:2503.10529, 2025. 3", + "[19] Ziyu Guo, Ray Zhang, Hao Chen, Jialin Gao, Dongzhi Jiang, Jiaze Wang, and Pheng-Ann Heng. Sciverse: Unveiling the knowledge comprehension and visual reasoning of lmms on multi-modal scientific problems. arXiv preprint arXiv:2503.10627, 2025. 3", + "[20] Ziyu Guo, Renrui Zhang, Chengzhuo Tong, Zhizheng Zhao, Peng Gao, Hongsheng Li, and Pheng-Ann Heng. Can we generate images with cot? let's verify and reinforce image generation step by step. arXiv preprint arXiv:2501.13926, 2025.3", + "[21] Jiaming Han, Renrui Zhang, Wenqi Shao, Peng Gao, Peng Xu, Han Xiao, Kaipeng Zhang, Chris Liu, Song Wen, Ziyu Guo, et al. Imagebind-llm: Multi-modality instruction tuning. arXiv preprint arXiv:2309.03905, 2023. 3", + "[22] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In CVPR, 2023. 2", + "[23] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 4", + "[24] Deok-Kyeong Jang, Soomin Park, and Sung-Hee Lee. Motion puzzle: Arbitrary motion style transfer by body part. ACM TOG, 2022. 2, 6", + "[25] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. Advances in Neural Information Processing Systems, 36:20067-20079, 2023. 3, 5", + "[26] Dongzhi Jiang, Renrui Zhang, Ziyu Guo, Yanwei Li, Yu Qi, Xinyan Chen, Liuhui Wang, Jianhan Jin, Claire Guo, Shen Yan, et al. Mme-cot: Benchmarking chain-of-thought in large" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13357", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "multimodal models for reasoning quality, robustness, and efficiency. arXiv preprint arXiv:2502.09621, 2025.3", + "[27] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Gmd: Controllable human motion synthesis via guided diffusion models. In ICCV, 2023. 1, 2", + "[28] Boeun Kim, Jungho Kim, Hyung Jin Chang, and Jin Young Choi. Most: Motion style transformer between diverse action contents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1705-1714, 2024. 2", + "[29] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer, 2024a. URL https://arxiv.org/abs/2408.03326. 3", + "[30] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. ICLR 2025 Spotlight, 2024. 3", + "[31] Xiaoqi Li, Mingxu Zhang, Yiran Geng, Haoran Geng, Yuxing Long, Yan Shen, Renrui Zhang, Jiaming Liu, and Hao Dong. Manipllm: Embodied multimodal large language model for object-centric robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18061-18070, 2024. 3", + "[32] Yanghao Li, Naiyan Wang, Jiaying Liu, and Xiaodi Hou. Demystifying neural style transfer. arXiv preprint arXiv:1701.01036, 2017. 4", + "[33] Zhe Li, Yisheng He, Lei Zhong, Weichao Shen, Qi Zuo, Lingteng Qiu, Zilong Dong, Laurence Tianruo Yang, and Weihao Yuan. Mulsmo: Multimodal stylized motion generation by bidirectional control flow. arXiv preprint arXiv:2412.09901, 2024. 2", + "[34] Zhe Li, Weihao Yuan, Yisheng He, Lingteng Qiu, Shenhao Zhu, Xiaodong Gu, Weichao Shen, Yuan Dong, Zilong Dong, and Laurence T Yang. Lamp: Language-motion pretraining for motion generation, retrieval, and captioning. arXiv preprint arXiv:2410.07093, 2024. 2", + "[35] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5", + "[36] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In ICCV, 2019. 5", + "[37] Ian Mason, Sebastian Starke, and Taku Komura. Real-time style modelling of human locomotion via feature-wise transformations and local motion phases. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 2022. 2, 4, 5, 6, 7, 8", + "[38] Arsha Nagrani, Paul Hongsuck Seo, Bryan Seybold, Anja Hauth, Santiago Manen, Chen Sun, and Cordelia Schmid. Learning audio-video modalities from image captions. In European Conference on Computer Vision, pages 407-426. Springer, 2022. 3", + "[39] Soomin Park, Deok-Kyeong Jang, and Sung-Hee Lee. *Diverse motion stylization for multiple style domains via spatial-temporal graph-based generative model.* Proceedings of the ACM on Computer Graphics and Interactive Techniques, 2021. 2" + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Xiaogang Peng, Yiming Xie, Zizhao Wu, Varun Jampani, Deqing Sun, and Huaizu Jiang. Hoi-diff: Text-driven synthesis of 3d human-object interactions using diffusion models. arXiv preprint arXiv:2312.06553, 2023. 2", + "[41] Mathis Petrovich, Or Litany, Umar Iqbal, Michael J Black, Gul Varol, Xue Bin Peng, and Davis Rempe. Multi-track timeline control for text-driven 3d human motion generation. In CVPR, 2024. 2", + "[42] Ekkasit Pinyoanuntapong, Muhammad Usama Saleem, Pu Wang, Minwoo Lee, Srijan Das, and Chen Chen. Bamm: Bidirectional autoregressive motion model. arXiv preprint arXiv:2403.19435, 2024.", + "[43] Ekkasit Pinyoanuntapong, Pu Wang, Minwoo Lee, and Chen Chen. Mmm: Generative masked motion model. In CVPR, 2024. 2", + "[44] Ziyun Qian, Zeyu Xiao, Zhenyi Wu, Dingkang Yang, Mingcheng Li, Shunli Wang, Shuaibing Wang, Dongliang Kou, and Lihua Zhang. Smcd: High realism motion style transfer via mamba-based diffusion. arXiv preprint arXiv:2405.02844, 2024. 2", + "[45] Sigal Raab, Inbal Leibovitch, Guy Tevet, Moab Arar, Amit H Bermano, and Daniel Cohen-Or. Single motion diffusion. arXiv preprint arXiv:2302.05905, 2023. 2", + "[46] Sigal Raab, Inbar Gat, Nathan Sala, Guy Tevet, Rotem Shalev-Arkushin, Ohad Fried, Amit H Bermano, and Daniel Cohen-Or. Monkey see, monkey do: Harnessing self-attention in motion diffusion for zero-shot motion transfer. arXiv preprint arXiv:2406.06508, 2024. 2", + "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. 2021. 3", + "[48] Davis Rempe, Zhengyi Luo, Xue Bin Peng, Ye Yuan, Kris Kitani, Karsten Kreis, Sanja Fidler, and Or Litany. Trace and pace: Controllable pedestrian animation via guided trajectory diffusion. In CVPR, 2023. 2", + "[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2", + "[50] Wenfeng Song, Xingliang Jin, Shuai Li, Chenglizhao Chen, Aimin Hao, Xia Hou, Ning Li, and Hong Qin. Arbitrary motion style transfer with multi-condition motion latent diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 821-830, 2024. 2", + "[51] Xiangjun Tang, Linjun Wu, He Wang, Bo Hu, Xu Gong, Yuchen Liao, Songnan Li, Qilong Kou, and Xiaogang Jin. Rsmt: Real-time stylized motion transition for characters. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 2", + "[52] Yiwen Tang, Zoey Guo, Zhuhao Wang, Ray Zhang, Qizhi Chen, Junli Liu, Delin Qu, Zhigang Wang, Dong Wang, Xuelong Li, et al. Exploring the potential of encoder-free architectures in 3d Imms. arXiv preprint arXiv:2502.09620, 2025. 3" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13358", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Tianxin Tao, Xiaohang Zhan, Zhongquan Chen, and Michiel van de Panne. Style-erd: Responsive and coherent online motion style transfer. 2022. 2", + "[54] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2023. 1, 2", + "[55] Weilin Wan, Zhiyang Dou, Taku Komura, Wenping Wang, Dinesh Jayaraman, and Lingjie Liu. Tlcontrol: Trajectory and language control for human motion synthesis. arXiv preprint arXiv:2311.17135, 2023. 2", + "[56] Yu-Hui Wen, Zhipeng Yang, Hongbo Fu, Lin Gao, Yanan Sun, and Yong-Jin Liu. Autoregressive stylized motion synthesis with generative flow. In CVPR, 2021. 2", + "[57] Qianyang Wu, Ye Shi, Xiaoshui Huang, Jingyi Yu, Lan Xu, and Jingya Wang. Thor: Text to human-object interaction diffusion via relation intervention. arXiv preprint arXiv:2403.11208, 2024. 2", + "[58] Qi Wu, Yubo Zhao, Yifan Wang, Yu-Wing Tai, and Chi-Keung Tang. Motionllm: Multimodal motion-language learning with large language models. arXiv preprint arXiv:2405.17013, 2024. 2", + "[59] Yiming Xie, Varun Jampani, Lei Zhong, Deqing Sun, and Huaizu Jiang. Omnicontrol: Control any joint at any time for human motion generation. In ICLR, 2024. 2", + "[60] Jingwei Xu, Huazhe Xu, Bingbing Ni, Xiaokang Yang, Xiaolong Wang, and Trevor Darrell. Hierarchical style-based networks for motion synthesis. In ECCV, 2020. 2", + "[61] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. Interdiff: Generating 3d human-object interactions with physics-informed diffusion. In ICCV, 2023. 2", + "[62] Sirui Xu, Ziyin Wang, Yu-Xiong Wang, and Liang-Yan Gui. Interdreamer: Zero-shot text to 3d dynamic human-object interaction. arXiv preprint arXiv:2403.19652, 2024. 2", + "[63] Le Xue, Mingfei Gao, Chen Xing, Roberto Martin-Martin, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1179-1189, 2023. 5", + "[64] Senqiao Yang, Jiaming Liu, Ray Zhang, Mingjie Pan, Ziyu Guo, Xiaqi Li, Zehui Chen, Peng Gao, Yandong Guo, and Shanghai Zhang. Lidar-llm: Exploring the potential of large language models for 3d lidar understanding. AAAI 2025, 2023. 3", + "[65] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In ICCV, 2023. 2", + "[66] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3836-3847, 2023. 2, 3, 4", + "[67] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. PAMI, 2024. 2", + "[68] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Point" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "clip: Point cloud understanding by clip. In CVPR 2022, 2022. 3", + "[69] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In ECCV 2022. Springer Nature Switzerland, 2022. 3", + "[70] Renrui Zhang, Jiaming Han, Chris Liu, Aojun Zhou, Pan Lu, Yu Qiao, Hongsheng Li, and Peng Gao. Llama-adapter: Efficient fine-tuning of large language models with zero-initialized attention. In ICLR 2024, 2024. 3", + "[71] Renrui Zhang, Dongzhi Jiang, Yichi Zhang, Haokun Lin, Ziyu Guo, Pengshuo Qiu, Aojun Zhou, Pan Lu, Kai-Wei Chang, Peng Gao, et al. Mathverse: Does your multi-modal llm truly see the diagrams in visual math problems? ECCV 2024, 2024. 3", + "[72] Renrui Zhang, Xinyu Wei, Dongzhi Jiang, Ziyu Guo, Shicheng Li, Yichi Zhang, Chengzhuo Tong, Jiaming Liu, Aojun Zhou, Bin Wei, et al. Mavis: Mathematical visual instruction tuning with an automatic data engine. arXiv preprint arXiv:2407.08739, 2024. 3", + "[73] Lei Zhong, Yiming Xie, Varun Jampani, Deqing Sun, and Huaizu Jiang. Smoodi: Stylized motion diffusion model. In ECCV, 2024. 2, 3, 5, 6", + "[74] Bin Zhu, Bin Lin, Munan Ning, Yang Yan, Jiaxi Cui, HongFa Wang, Yatian Pang, Wenhao Jiang, Junwu Zhang, Zongwei Li, et al. Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. arXiv preprint arXiv:2310.01852, 2023. 3", + "[75] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. ICCV 2023, 2023. 3" + ], + "bbox": [ + 516, + 90, + 906, + 556 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13359", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/StyleMotif_ Multi-Modal Motion Stylization using Style-Content Cross Fusion/ffae00f3-5394-4dd8-b9a4-bc3713aa96e3_model.json b/2025/StyleMotif_ Multi-Modal Motion Stylization using Style-Content Cross Fusion/ffae00f3-5394-4dd8-b9a4-bc3713aa96e3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..98e4b0d8bda55130f4b8f02d281bb90879e2f84c --- /dev/null +++ b/2025/StyleMotif_ Multi-Modal Motion Stylization using Style-Content Cross Fusion/ffae00f3-5394-4dd8-b9a4-bc3713aa96e3_model.json @@ -0,0 +1,2268 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.131, + 0.908, + 0.155 + ], + "angle": 0, + "content": "STYLEMOtIF: Multi-Modal Motion Stylization using Style-Content Cross Fusion" + }, + { + "type": "table", + "bbox": [ + 0.157, + 0.179, + 0.844, + 0.301 + ], + "angle": 0, + "content": "
Ziyu Guo†Young Yoon LeeJoseph Liu
The Chinese University of Hong KongRobloxRoblox
ziyuguo@link.cuhk.edu.hkylee@roblox.comjosephliu@roblox.com
Yizhak Ben-ShabatVictor ZordanMubbasir Kapadia
RobloxRobloxRoblox
ibenshabat@roblox.comvbzordan@roblox.commkapadia@roblox.com
" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.334, + 0.327, + 0.35 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.367, + 0.487, + 0.624 + ], + "angle": 0, + "content": "We present STYLEMOTIF, a novel Stylized Motion Latent Diffusion model, generating motion conditioned on both content and style from multiple modalities. Unlike existing approaches that either focus on generating diverse motion content or transferring style from sequences, STYLEMOTIF seamlessly synthesizes motion across a wide range of content while incorporating stylistic cues from multi-modal inputs, including motion, text, image, video, and audio. To achieve this, we introduce a style-content cross fusion mechanism and align a style encoder with a pre-trained multi-modal model, ensuring that the generated motion accurately captures the reference style while preserving realism. Extensive experiments demonstrate that our framework surpasses existing methods in stylized motion generation and exhibits emergent capabilities for multi-modal motion stylization, enabling more nuanced motion synthesis. Project Page: https://stylemotif.github.io." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.671, + 0.223, + 0.687 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.694, + 0.487, + 0.875 + ], + "angle": 0, + "content": "Human motion generation is a fundamental task in computer graphics and animation, enabling the synthesis of realistic and expressive human movements. Broadly, human motion can be characterized by two complementary aspects: content, which defines the underlying action (e.g., walking, jumping), and style, which encodes variations such as personal flair, emotional expression, or cultural influences (e.g., jubilant, aggressive). This separation allows for greater control and flexibility in generating motion, making it particularly valuable in creative industries like game development, film production, and virtual reality. However, traditional approaches to stylized motion generation often depend on manual pro" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.319, + 0.905, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.542, + 0.909, + 0.613 + ], + "angle": 0, + "content": "Figure 1. Comparison of Our Proposed STYLEMOTIF with SMooDi. SMooDi duplicates and separates two diffusion backbones for content and style, increasing complexity. In contrast, STYLEMOTIF adopts a single diffusion backbone with direct style injection, enabling efficient and realistic motion stylization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.907, + 0.658 + ], + "angle": 0, + "content": "cesses such as motion capture or keyframe animation, which are costly, time-consuming, and labor-intensive." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.659, + 0.909, + 0.87 + ], + "angle": 0, + "content": "Recent progress in text-to-motion (T2M) diffusion frameworks [6, 27, 54] has greatly advanced the ability to translate natural-language prompts into realistic human motions. By leveraging powerful denoising diffusion models, these approaches capture intricate spatiotemporal dependencies in the data, enabling coherent sequences of human movements to be generated directly from brief textual descriptions. Despite their success in content fidelity and diversity, most current T2M diffusion methods concentrate primarily on what action is performed, while overlooking how it is performed, namely, its stylistic details. Simply appending a separate style-transfer module to text-driven motion diffusion pipelines can introduce additional complexity and risk of compounding errors in the final output." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.901 + ], + "angle": 0, + "content": "In parallel, motion style transfer has been actively studied to infuse stylistic cues from a reference motion (or style" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.308, + 0.9 + ], + "angle": 0, + "content": "† Work done as an intern at Roblox." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13349" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.228 + ], + "angle": 0, + "content": "data) into another sequence [1, 24, 37, 53]. Although many of these methods effectively disentangle content and style for small-scale tasks, the pipeline becomes cumbersome when a large variety of content motions need to be stylized. Moreover, they commonly assume that the input or target sequences are high-quality motion data. In scenarios where content motions are synthetically generated, or partially noisy, the transfer process can deteriorate, leading to undesirable motion artifacts or compromised style fidelity." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.23, + 0.486, + 0.487 + ], + "angle": 0, + "content": "To address the need for simultaneously controlling both content and style, some recent works have merged style encoding with diffusion-based motion generation. Among these, the most recent and representative approach [73] augments a pre-trained latent diffusion model [6] with a style adaptor and classifier-based style guidance, achieving stylized motion from textual prompts and motion-style references. While effective, this method relies on additional training diffusion backbones, which shares structural similarities with ControlNet [66] as shown in Figure 1, which increases model complexity and training overhead. It is also constrained to motion as the primary style input. Concurrent work [33] proposes a bidirectional control flow mechanism to mitigate conflicts between style and content, extending style control to multiple modalities, but this approach also duplicates two separate diffusion backbones with substantial training overhead and limited applicability across modalities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.485, + 0.746 + ], + "angle": 0, + "content": "To this end, we propose STYLEMOTIF, a new framework for multi-modal motion stylization that unifies text-to-motion diffusion with style conditioning in a single-diffusion structure. Specifically, we leverage a pre-trained motion latent diffusion model (MLD) [6] to preserve strong content generation capabilities, and seamlessly integrate style features extracted from a dedicated encoder, which is aligned with a multi-modal foundation model [12]. In contrast to previous works, STYLEMOTIF avoids duplicating large portions of the network or relying on specialized style diffusion network. Instead, we introduce a style-content cross fusion mechanism, which injects stylistic cues into the diffusion process while maintaining motion realism. As a result, our STYLEMOTIF not only yields more robust stylized outputs but also supports diverse style signals, such as motion, text, images, audio, or video clips, via the alignment in multi-modal feature space. We summarize our main contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.75, + 0.484, + 0.81 + ], + "angle": 0, + "content": "- We present STYLEMOTIF, a stylized latent motion diffusion framework that unifies diverse motion content and multi-modal styles within a compact single-diffusion design." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.811, + 0.484, + 0.87 + ], + "angle": 0, + "content": "- We propose a style-content cross fusion module that injects stylistic cues into the diffusion denoising process, achieving faithful stylization without compromising motion realism and ensuring efficiency." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "- We achieve a unified multi-modal style feature space and unveil new emergent capabilities through multi-modal" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.136 + ], + "angle": 0, + "content": "alignment, which accommodate various sources including motion, text, images, audio, and video, for flexible and versatile multi-modal style control." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.137, + 0.905, + 0.183 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate that STYLEMOTIF consistently outperforms existing methods regarding style expressiveness, content preservation, and efficiency." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.194, + 0.655, + 0.209 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.219, + 0.908, + 0.568 + ], + "angle": 0, + "content": "Human Motion Generation. Recent progress in human motion generation [3, 5, 8, 13, 14, 41-43, 46, 55, 57, 58, 62] has been driven by transformer [14, 34] and diffusion models [2, 8, 49, 66], showing great potential in producing realistic and diverse motions. These approaches have shown great potential in producing realistic and diverse motion sequences. For example, Momask [14] improves motion generation using a residual VQ-VAE. Similarly, LaMP [34] introduces a motion-aware text encoder and a motion-to-text language model to enhance motion quality through text conditioning. Diffusion models, in particular, have become a key approach in motion generation [6, 22, 27, 48, 54, 61, 65, 67]. MDM [54] introduces a motion diffusion model that operates directly on raw motion data to capture the relationship between motions and input text conditions. MLD [6] improve efficiency by embedding the diffusion process in latent space, reducing computational cost. They also allow conditioning on specific constraints, such as predefined trajectories [27, 55, 59] or human-object interactions [7, 40, 57], enabling greater control and diversity. Our work leverages pretrained motion latent diffusion model [6] and multi-modal foundation models [12] to achieve stylized human motion generation while maintaining high motion quality." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.584, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Motion Stylization. Motion stylization [1, 24, 28, 37, 39, 44, 45, 50, 51, 53, 56, 60] involves transferring stylistic features from a reference motion to a source motion/textual prompt, enabling creative transformations while preserving the original motion content. Early methods, like those in motion style transfer [1, 24], typically separate motion content and style for recombination. For instance, Aberman et al. [1] used a generative adversarial network to decouple style from content without paired data, while Motion Puzzle [24] allows style control for individual body parts, and Guo et al. [15] utilized pretrained motion models for better style integration. Recent approaches, like SMooDi [73], generate stylized motion from text and style sequences using style guidance. However, these models face two main challenges: (1) inefficient architectures that duplicate diffusion backbones for content and style, and (2) limited to style motion for guidance. Another concurrent approach, MulSMo [33], attempts to address some of these by introducing a bidirectional control flow between style and content networks, reducing conflicts. Still, it remains restricted to a few modalities and duplicates diffusion networks. In this" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13350" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.068, + 0.91, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.376, + 0.908, + 0.435 + ], + "angle": 0, + "content": "Figure 2. Overall Pipeline of STYLEMOTIF, a diffusion framework for multi-modal motion stylization. Given a text prompt and a reference style from various modalities, our model extract style features and fuse them with content by style-content cross fusion. Through multi-modal alignment with contrastive learning, we enable seamless multi-modal conditioning and flexible stylization across motion, text, images, audio, and video." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.452, + 0.486, + 0.543 + ], + "angle": 0, + "content": "work we eliminate redundant diffusion backbone, resulting in a simpler, more effective, and efficient approach to stylized motion generation. Our method also enables efficient multi-modal motion stylization, supporting text, image and multi-modal inputs while maintaining both high stylization quality and motion realism." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.614, + 0.487, + 0.902 + ], + "angle": 0, + "content": "Multi-modality Learning. Recent advancements in multimodality learning [9, 10, 38] have revolutionized various domains by enabling joint understanding across diverse data types. Foundational models like CLIP [47] and its extension [16, 68, 69, 75] establish robust vision-language alignment, while ImageBind [12], Point-Bind [17], and Language-Bind [74] expand this paradigm to diverse modalities, demonstrating emergent cross-modal capabilities. The rise of multi-modal large language models [4, 19, 21, 30, 31, 64, 72] further enhances semantic understanding through unified text-visual processing, achieving superior performance across 2D images [11, 29, 70], 3D point clouds [17, 18, 52], and complex reasoning scenarios [20, 26, 71]. For human motion, some approaches [5, 14, 25] incorporate multiple modalities into human motion understanding and generation tasks. Our work introduces the first framework for multimodal guided motion stylization, achieving seamless style injection from diverse modalities through a single diffusion architecture with style-content cross fusion." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.45, + 0.909, + 0.485 + ], + "angle": 0, + "content": "3. STYLEMOTIF: Multi-Modal Motion Stylization with Style-Content Cross Fusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.909, + 0.691 + ], + "angle": 0, + "content": "We propose STYLEMOTIF, a novel framework for stylized motion synthesis that combines style-content cross fusion and multi-modal motion stylization, as illustrated in Figure 2. Our approach integrates a style-content cross fusion mechanism that allows for coherent feature blending, ensuring the generated motion accurately reflects the reference style while maintaining the content's realism (§3.2). This mechanism is complemented by a multi-modal motion stylization strategy, which leverages inputs from diverse modalities such as image, video, audio, and text to provide control over the stylization process (§3.3). This framework enables highly flexible and realistic stylized motion synthesis, which offers greater control and diversity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.698, + 0.625, + 0.713 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Preliminaries. Motion Latent Diffusion (MLD) [6] formulates a conditional latent diffusion model by training \\(\\epsilon_{\\theta}(z_t,t,c)\\) to denoise a sequence of latents \\(\\{z_t\\}_{t = 0}^T\\) , where \\(z_{t}\\in \\mathbb{R}^{n\\times d}\\) represents the motion latent at timestep \\(t\\) , conditioned on the distribution \\(p(z_{t}|c)\\) . A conditional domain encoder \\(\\tau_{\\theta}(c)\\) , such as CLIP [47], enables text-to-motion tasks, and the model is trained as \\(\\epsilon_{\\theta}(z_t,t,\\tau_{\\theta}(c))\\) . To incorporate additional style conditioning, a style condition \\(s\\) can be introduced with its own encoder \\(\\psi_{\\theta}(s)\\) , which may take the form of a motion or text encoder. This extends the model to \\(\\epsilon_{\\theta}(z_t,t,\\tau_{\\theta}(c),\\psi_{\\theta_s}(s))\\) . SMooDi [73] adopts a ControlNet [66]-style approach for style conditioning by" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13351" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.168 + ], + "angle": 0, + "content": "creating a trainable copy of the neural network weights \\(\\theta_{s}\\) from the original MLD model \\(\\theta_{c}\\). This copy includes zero-initialized linear layers, denoted as \\(\\mathcal{Z}(\\cdot ;\\cdot)\\). The output \\(\\mathcal{F}^i (z_t,t,\\tau_\\theta (c),\\psi_\\theta (s);\\theta_c)\\) of the \\(i\\)-th MLD block, now conditioned on style, is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.179, + 0.483, + 0.217 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {F} ^ {i} \\left(z _ {t}, t, \\tau_ {\\theta} (c), \\psi_ {\\theta_ {s}} (s); \\theta_ {c}\\right) = \\mathcal {F} ^ {i} \\left(z _ {t}, t, \\tau_ {\\theta} (c); \\theta_ {c}\\right) \\\\ + \\mathcal {Z} \\left(\\mathcal {F} ^ {i} \\left(z _ {t}, t, \\tau_ {\\theta} (c), \\psi_ {\\theta_ {s}} (s); \\theta_ {s}\\right); \\theta_ {z _ {i}}\\right) \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.228, + 0.485, + 0.304 + ], + "angle": 0, + "content": "A key property of this formulation is that since \\(\\theta_{z_i}\\) is initialized to zero, thus \\(\\mathcal{F}^i (z_t,t,\\tau_\\theta (c),\\psi_\\theta (s)) = \\mathcal{F}^i (z_t,t,\\tau_\\theta (c))\\) at the beginning. However, the tradeoff of that method is that it needs to maintain the additional parameters \\(\\theta_{s}\\) and \\(\\theta_{z_i}\\) in order for style transfer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.484, + 0.58 + ], + "angle": 0, + "content": "STYLEMOTIF. We propose STYLEMOTIF, eschewing zero linear layers in favor of injecting it directly via statistical manipulation, with the pipeline shown in Figure 2 describing our approach. Our STYLEMOTIF utilizes latent space diffusion within a single generative diffusion backbone, building upon a pretrained MLD model [6]. Instead of perturbing the outputs of each block \\(i\\) via zero initialized layer \\(\\mathcal{Z}(\\mathcal{F}^i (z_t,t,\\tau_\\theta (c),\\psi_\\theta (s);\\theta_s);\\theta_{z_i})\\), STYLEMOTIF replaces this with a statistically transformed style embedding that is injected into the original MLD branch. This simplifies the model while ensuring high-quality stylization results, whose details are presented in § 3.2. Also, we follow SMooDi's generation guidance and training scheme to ensure high-quality stylized motion synthesis. We use a hybrid guidance strategy that balances content fidelity and style adherence by combining classifier-free and classifier-based techniques during diffusion sampling." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.589, + 0.338, + 0.604 + ], + "angle": 0, + "content": "3.2. Style-Content Cross Fusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.484, + 0.852 + ], + "angle": 0, + "content": "Style Encoder Pre-training. To establish a robust foundation for style-content fusion, we combine the content knowledge from the pre-trained MLD's [6] VAE with the style knowledge from the 100STYLE [37] dataset. The MLD's VAE is pre-trained on the HumanML3D [13] dataset, which provides extensive understanding of content motion. Building on this, we further fine-tune the model on the 100STYLE dataset in a variational autoencoding manner to initially align the content and style data distributions in the latent space. After training, we discard the decoder and retain only the encoder as the motion style encoder. This reconstruction task enables the encoder to learn robust motion feature representations, which are essential for supporting the subsequent stylization process. By integrating content and style knowledge in this way, we ensure a strong foundation for seamless style-content fusion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Style-Content Cross Normalization. To train a stylized diffusion model \\(\\epsilon_{\\theta}(z_t, t, \\tau_{\\theta}(c), \\psi_{\\theta}(s))\\), we effectively fuse" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.273 + ], + "angle": 0, + "content": "content and style features directly within the latent space diffusion process instead of duplicating two pre-trained diffusion backbones for content and style or separate control mechanisms [66]. Prior work [23, 32] has proven that feature statistics (mean, variance) effectively encode style. Given the output features \\(\\mathcal{F}^i\\) of the \\(i\\)-th block, we derive the content features \\(\\mathcal{F}_c^i = \\mathcal{F}^i(z_t, t, \\tau_\\theta(c); \\theta_c)\\) from the input text \\(c\\), and the style features \\(\\mathcal{F}_s = \\psi_{\\theta_s}(s)\\) are extracted from the reference motion sequence \\(s\\) using the pre-trained style encoder. To perform the fusion, we first compute the mean \\(\\mu_c\\) and variance \\(\\sigma_c^2\\) of the content features across the feature dimension:" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.297, + 0.907, + 0.34 + ], + "angle": 0, + "content": "\\[\n\\mu_ {c} = \\frac {1}{D} \\sum_ {j = 1} ^ {D} \\mathcal {F} _ {c} ^ {i, j}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.344, + 0.907, + 0.384 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {c} ^ {2} = \\frac {1}{D} \\sum_ {j = 1} ^ {D} \\left(\\mathcal {F} _ {c} ^ {i, j} - \\mu_ {c}\\right) ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.397, + 0.907, + 0.458 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}_c^{i,j}\\) denotes the content features of the \\(i\\)-th block and the \\(j\\)-th feature element. Next, we normalize the style features \\(\\mathcal{F}_s\\) using these content statistics, ensuring that the style features are adapted to the content's statistical properties:" + }, + { + "type": "equation", + "bbox": [ + 0.647, + 0.469, + 0.907, + 0.503 + ], + "angle": 0, + "content": "\\[\n\\widetilde {\\mathcal {F}} _ {s, c} = \\frac {\\mathcal {F} _ {s} - \\mu_ {c}}{\\sqrt {\\sigma_ {c} ^ {2} + \\eta}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.513, + 0.909, + 0.603 + ], + "angle": 0, + "content": "where \\(\\eta\\) is a constant added for numerical stability. This style-content cross normalization ensures that the style features are smoothly integrated with the content features while maintaining the content's original structure. After that, we add the normalized style features to the content features, formulating our final cross-normalization as" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.614, + 0.907, + 0.632 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} ^ {i} \\left(z _ {t}, t, \\tau_ {\\theta} (c), \\psi_ {\\theta_ {s}} (s); \\theta_ {c}\\right) = \\mathcal {F} _ {c} ^ {i} + \\gamma \\cdot \\tilde {\\mathcal {F}} _ {s, c}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.78 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is a parameter used for scaling the normalized value. \\(\\widetilde{\\mathcal{F}}_{s,c}\\) can be thought of as a perturbation of \\(\\mathcal{F}_c^i\\) as it is scaled within its range. Notably, the fusion process is performed only once after the \\(m\\)-th block during the denoising process to avoid distorting the content while effectively introducing the style. This efficient fusion method eliminates the need for additional learnable parameters, as it is based solely on statistical transformations, achieving high-quality results with minimal computational overhead." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.789, + 0.796, + 0.804 + ], + "angle": 0, + "content": "3.3. Multi-Modal Motion Stylization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Our STYLEMOTIF extends to support multi-modal motion stylization. To achieve this, we integrate a pre-trained multimodal foundation model [12], which provides a unified, multi-modal aligned feature space, enabling effective cross-modal alignment between motion and other modalities. With this alignment, our model can flexibly combine multiple" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13352" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.152 + ], + "angle": 0, + "content": "input modalities, such as text, image, audio, and video, to guide the stylization process in a comprehensive manner. This capability results in emergent abilities for multi-modal motion stylization, allowing for more nuanced outputs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.171, + 0.485, + 0.307 + ], + "angle": 0, + "content": "Motion-Text Pair Curation. To align motion with other modalities, we process a set of motion-text pairs using the curated 100STYLE subset [37], which is carefully selected to avoid conflicts between content and style motions, so that the model can effectively learn the relationships between motion and text. Each motion sequence is paired with a corresponding single textual label, which serves as the text prompt. The curated motion-text pairs will be used for the following cross-model alignment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.326, + 0.486, + 0.462 + ], + "angle": 0, + "content": "Multi-Modal Alignment. To maintain the alignment within the multi-modal space, we freeze the text encoder of ImageBind [12] and introduce a lightweight projection layer after the encoder. This projection layer aligns the feature dimensions of the text and motion encoders. Since the multi-modal model represents each modality with a global feature, our alignment process focuses on these global features to achieve robust alignment between text and motion representations, formalized as:" + }, + { + "type": "equation", + "bbox": [ + 0.227, + 0.474, + 0.484, + 0.491 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {t} = \\pi \\left(\\mathcal {E} _ {\\text {t e x t}} (l)\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.229, + 0.493, + 0.484, + 0.51 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {s} = \\psi_ {\\theta_ {s}} (s), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.485, + 0.657 + ], + "angle": 0, + "content": "where \\(l\\) and \\(s\\) are the paired input text label and style motion, \\(\\mathcal{F}_t\\) and \\(\\mathcal{F}_s\\) denote the feature representations from the text and style motion encoders, \\(\\mathcal{E}_{\\mathrm{text}}\\) and \\(\\psi_{\\theta_s}\\), respectively, and \\(\\pi\\) represents the projection operation that maps text features into the motion feature space. We employ a contrastive learning loss [63] to align the feature spaces of motion and text, bringing them closer together in the shared multi-modal space. Thus, we obtain a unified space between motion and all the modalities. We formulated it as" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.666, + 0.484, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l i g n}} = - \\frac {1}{2} \\sum_ {(i, j)} \\log \\frac {\\exp \\frac {\\mathcal {F} _ {t} ^ {i} \\cdot \\mathcal {F} _ {s} ^ {j}}{\\tau_ {0}}}{\\sum_ {k} \\exp \\frac {\\mathcal {F} _ {t} ^ {i} \\cdot \\mathcal {F} _ {s} ^ {k}}{\\tau_ {0}}} + \\log \\frac {\\exp \\frac {\\mathcal {F} _ {t} ^ {i} \\cdot \\mathcal {F} _ {s} ^ {j}}{\\tau_ {0}}}{\\sum_ {k} \\exp \\frac {\\mathcal {F} _ {t} ^ {k} \\cdot \\mathcal {F} _ {s} ^ {j}}{\\tau_ {0}}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.74, + 0.486, + 0.83 + ], + "angle": 0, + "content": "where \\( t \\) and \\( s \\) represent two modalities (text and style motion) and \\( (i,j) \\) indicates a positive pair in each training batch, \\( k \\) indexes all samples in the batch, including both positive and negative ones, and \\( \\tau_0 \\) is a temperature parameter. During inference, we obtain the multi-modal (text, image, video, or audio) features from the multi-modal input \\( m \\)," + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.843, + 0.484, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {m} = \\mathcal {E} _ {\\text {I m a g e B i n d}} (m), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.485, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathcal{E}_{\\mathrm{ImageBind}}\\) denotes ImageBind [12]. We use the multimodal features \\(\\mathcal{F}_m\\) to retrieve the most semantically similar" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.183 + ], + "angle": 0, + "content": "motion features from the unified multi-modal space. The retrieved motion features are then used to guide the stylization process, ensuring that fine-grained style details are preserved and accurately reflected in the final output. This approach enhances the model's ability to generate stylized motions that are both contextually relevant and visually coherent." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.197, + 0.64, + 0.213 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.222, + 0.906, + 0.252 + ], + "angle": 0, + "content": "We provide extensive quantitative and qualitative analyses across multiple tasks in this section." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.262, + 0.722, + 0.278 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.285, + 0.908, + 0.405 + ], + "angle": 0, + "content": "Implementation Details. We adopt the pre-trained MLD [6] as the foundation for motion generation. The style encoder of our model derives from the encoder of MLD's VAE, while the projection layer after text encoder is a single Linear layer. During diffusion training, we only enable the style encoder to be trainable while freezing other parameters. The model is optimized using the AdamW optimizer [35] with a constant learning rate of \\(10^{-5}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.909, + 0.576 + ], + "angle": 0, + "content": "Dataset Settings. We use the HumanML3D [13] dataset as our primary motion content dataset, which collects 14,616 motion sequences from AMASS [36] and annotates 44,970 sequence-level textual description. To train the style network, we utilize the 100STYLE [37] dataset, containing 45,303 style motions. We also adopt the text annotations for 100STYLE from previous work [73], which are pseudo text descriptions generated from MotionGPT [25]. We utilize the consistent root-velocity motion representations from HumanML3D for both the content and style data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.909, + 0.716 + ], + "angle": 0, + "content": "Baselines. We mainly evaluate our method against SMooDi [73] on motion-guided stylization and motion style transfer tasks. We also compare our text-guided stylcification with 'ChatGPT+MLD' approach, which utilize ChatGPT to combine style text and content text, and functions as a straightforward text-to-motion model without control capabilities [73]. Additionally, we qualitatively compare our model with SMooDi on motion-guided stylization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.911, + 0.901 + ], + "angle": 0, + "content": "Evaluation Metrics. To evaluate our model, we use several metrics following previous works [73]. We measure R-precision, Multi-modal Distance (MM Dist), Diversity, and Frechet Inception Distance (FID) to assess how well the content from text is preserved and how realistic the generated motions are. We use Style Recognition Accuracy (SRA) to evaluate how accurately the style of the reference motion is reflected in the generated motion. Also, we adopt the Foot Skate Ratio, which helps assess the realism of the generated motion, reducing artifacts like foot sliding. During evaluation, following previous works [73], we randomly select" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13353" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.099, + 0.082, + 0.581, + 0.212 + ], + "angle": 0, + "content": "
MethodSRA ↑FID ↓MM Dist ↓R-Precision ↑ (Top-3)Diversity →Foot Skate Ratio ↓
Motion-Guided Stylization
MLD + Aberman [1]54.373.3095.9830.4068.8160.347
MLD + Motion Puzzle [24]63.776.1276.4670.2906.4760.185
SMooDi [73]72.421.6094.4770.5719.2350.124
STYLEMOTIF (Ours)77.651.5514.3540.5867.5670.097
Text-Guided Stylization
MLD + ChatGPT [73]4.820.6144.3130.6058.8360.131
STYLEMOTIF (Ours)56.710.6033.6840.6909.1010.101
" + }, + { + "type": "table_caption", + "bbox": [ + 0.096, + 0.221, + 0.584, + 0.264 + ], + "angle": 0, + "content": "Table 1. Quantitative Results for Motion-Guided and Text-Guided Stylization. Bold values denote the best performance. As there is no ground-truth reference for Diversity, no value is highlighted in bold; but the metric is provided for reference." + }, + { + "type": "table", + "bbox": [ + 0.599, + 0.08, + 0.89, + 0.158 + ], + "angle": 0, + "content": "
MethodSRA ↑FID ↓Foot Skate Ratio ↓
MLD + Aberman [1]61.013.8920.338
MLD + Motion Puzzle [24]67.236.8710.197
SMooDi [73]65.151.5820.095
STYLEMOTIF (Ours)68.811.3750.094
" + }, + { + "type": "table_caption", + "bbox": [ + 0.589, + 0.167, + 0.898, + 0.264 + ], + "angle": 0, + "content": "Table 2. Quantitative Results of Motion Style Transfer on HumanML3D [13] dataset. Our method outperforms previous works in all metrics, which demonstrates effective style-content fusion for high-quality motion style transfer, providing significant advantages for downstream tasks besides motion stylization." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.275, + 0.907, + 0.609 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.624, + 0.908, + 0.669 + ], + "angle": 0, + "content": "Figure 3. Qualitative Results of Motion-Guided Stylization. Our model generates cohesive and realistic motions that effectively align style and content, such as preserving the 'circular' trajectory (first column) and 'hop' content (third column). In contrast, SMooDi [73] struggles to maintain content fidelity and sometimes fails to reflect the specified style (e.g., 'phone on the left' in the second column)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.685, + 0.483, + 0.732 + ], + "angle": 0, + "content": "a content text from HumanML3D [13] and a style motion from 100STYLE [37], and compute SRA for the generated motion using a pre-trained classifier [73]." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.742, + 0.33, + 0.759 + ], + "angle": 0, + "content": "4.2. Motion-guided Stylization" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Quantitative Analysis. In Table 1, we report the quantitative results for motion-guided stylization, where the motion serves as style input while text prompt as content one. Our method outperforms three baseline approaches [1, 24, 73] in all metrics. Specifically, we achieve a \\(5.23\\%\\) improvement in SRA while maintaining competitive performance in FID compared to the best baseline [73]. It demonstrates that our method, by effectively utilizing style-content cross fusion, generates motions that better align with the style reference" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.685, + 0.907, + 0.716 + ], + "angle": 0, + "content": "while maintaining content integrity, showcasing the strength of our design in balancing style and content preservation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Qualitative Analysis. In Figure 3, we show some qualitative results of motion-guided stylization from our model and baseline [73]. Compared with the baseline, our model produces more cohesive stylized motions, with better alignment of style and content. For instance, in the provided examples, SMooDi fails to maintain the 'circular' trajectory or 'hop' action specified by the content or cannot accurately reflect the intended style, 'phone on the left'. This indicates that our approach more effectively integrates style and content, resulting in more realistic and consistent stylized motions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13354" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.079, + 0.92, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.258, + 0.51, + 0.3 + ], + "angle": 0, + "content": "Figure 4. Qualitative Results of Text-Guided Stylization. Our model seamlessly integrates textual style descriptions with content, producing visually coherent and stylistically consistent results." + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.258, + 0.92, + 0.301 + ], + "angle": 0, + "content": "Figure 5. Qualitative Results of Style Interpolation. Our model blends multiple style inputs while preserving content integrity, demonstrating effective style-content fusion of our model." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.312, + 0.907, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.535, + 0.908, + 0.566 + ], + "angle": 0, + "content": "Figure 6. Qualitative Results of Multi-Modal Motion Stylization. Our model generates stylized motions guided by diverse modalities (e.g., text, image, video, audio), effectively transferring style while preserving content integrity." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.582, + 0.307, + 0.599 + ], + "angle": 0, + "content": "4.3. Text-guided Stylization" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.604, + 0.486, + 0.771 + ], + "angle": 0, + "content": "Quantitative Analysis. For the text-guided stylization task, we use the HumanML3D [13] dataset for motion content and employ the single text label [37] as the style control. As reported in Table 1 (Bottom), our method significantly outperforms the baseline, achieving \\(56.71\\%\\) in SRA, compared to \\(4.82\\%\\) for 'ChatGPT + MLD'. Moreover, our method maintains competitive FID, balancing style reflection with content preservation. This indicates that our design, which combines multi-modal alignment, helps utilize the style signal in the motion-text shared spaces and achieve significant effectiveness in text-guided stylization." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.781, + 0.486, + 0.843 + ], + "angle": 0, + "content": "Qualitative Analysis. In Figure 4, we showcase qualitative results for text-guided stylization, where our model also demonstrates strong capability in harmonizing style and content, generating high-quality and visually coherent results." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.299, + 0.866 + ], + "angle": 0, + "content": "4.4. Motion Style Transfer" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.485, + 0.901 + ], + "angle": 0, + "content": "For the motion style transfer task, we use the HumanML3D [13] dataset for motion content and the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.909, + 0.766 + ], + "angle": 0, + "content": "100STYLE dataset [37] for motion styles. As shown in Table 2, our method outperforms the baseline models across all key metrics. Specifically, we achieve a \\(3.66\\%\\) improvement in SRA, indicating better style reflection, and a 0.207 reduction in FID reflecting improved realism. These demonstrate that our framework not only improves the alignment between style and content but also provides significant advantages for downstream tasks like motion style transfer with efficient style-content cross fusion. By maintaining content integrity while seamlessly integrating style, STYLEMOTIF ensures more realistic and dynamically consistent stylized motions, which is critical for high-quality motion style transfer." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.798, + 0.79 + ], + "angle": 0, + "content": "4.5. Multi-Modal Motion Stylization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "With the power of the aligned multi-modal space, our model supports stylization guided by a variety of modalities, including motion, text, image, video, audio. Here, we utilize the input style feature to retrieve the corresponding motion features as the style condition, while the content input is provided as text prompt. The model then generates a stylized motion that incorporates the style from the input modality" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13355" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.075, + 0.318, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.204, + 0.321, + 0.275 + ], + "angle": 0, + "content": "Figure 7. Ablation Study on Scaling Ratio \\(\\gamma\\) in Eq. 5. We report both SRA and FID to show the impact of the scaling ratio on both stylization and content preservation." + }, + { + "type": "table", + "bbox": [ + 0.345, + 0.072, + 0.892, + 0.226 + ], + "angle": 0, + "content": "
Style Encoder Pre-training Strategy
HumanML3D100STYLESRA ↑FID ↓MM Dist ↓R-Precision ↑Diversity →Foot Skate Ratio ↓
-✓76.731.7884.3490.5717.5050.101
✓-76.581.6354.4580.5728.5340.109
✓✓77.651.5514.3540.5867.5670.097
Text in Multi-modal Alignment
Textual ExpressionSRA ↑FID ↓MM Dist ↓R-Precision ↑Diversity →Foot Skate Ratio ↓
Brief Description76.841.5804.3780.5787.2510.099
Detailed Description75.251.6224.4190.5637.7640.102
Single Text Label77.651.5514.3540.5867.5670.097
" + }, + { + "type": "table_caption", + "bbox": [ + 0.342, + 0.235, + 0.892, + 0.278 + ], + "angle": 0, + "content": "Table 3. Ablation Study on Style Encoder Pre-training Strategies and Text Expression for Multi-modal Alignment. 'w. HumanML3D' and 'w. 100STYLE' denote pre-training with HumanML3D [13] and 100STYLE [37] data respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.296, + 0.486, + 0.435 + ], + "angle": 0, + "content": "while maintaining the content integrity as specified by the text prompt. We showcase several examples of multi-modal motion stylization in Figure 6, where different modalities guide the motion generation. For instance, when a text content \"A person is walking.\" is provided alongside an image of a duck as style input, the model retrieves a relevant motion feature and blends the content with the style of 'Duckfoot'. As shown, the style from various inputs (text, image, video, audio) is effectively transferred to the generated motion." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.444, + 0.273, + 0.46 + ], + "angle": 0, + "content": "4.6. Style Interpolation" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.467, + 0.485, + 0.649 + ], + "angle": 0, + "content": "Leveraging the aligned multi-modal space, our model enables text-guided style interpolation. Given one content text along with at least two style style texts, our model generates a motion that combines the characteristics of all input styles. When style texts are provided, we retrieve the most similar style motion features from the shared space. These features are then combined by weighted summation. The combined style features are fused with the content features using the proposed style-content cross fusion. In Figure 5, we showcase some qualitative results of style interpolation in, where the generated motions successfully blend the characteristics of both styles while maintaining the content's integrity." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.659, + 0.245, + 0.675 + ], + "angle": 0, + "content": "4.7. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.682, + 0.486, + 0.835 + ], + "angle": 0, + "content": "Style Encoder Pre-training. In Table 3, we investigate the impact of different pre-training strategies for the style encoder. Our results show that pre-training on both HumanML3D [13] and 100STYLE [37] yields the best performance. Compared to pre-training solely on either one, the combined training on both datasets provides the style encoder with a richer set of prior knowledge. This enables the model to effectively balance content preservation and style reflection, leveraging the diverse characteristics of both datasets to enhance the overall stylization quality." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.856, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Cross Normalization Scaling Ratio. In Figure 7, we examine the effect of different scaling ratios \\(\\gamma\\) in Eq. 5 on the stylization performance. As shown, the model achieves" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.296, + 0.909, + 0.373 + ], + "angle": 0, + "content": "the best performance with \\(\\gamma = 0.6\\), striking an optimal balance between style reflection and content preservation. The choice of scaling ratio influences both SRA and FID, where the optimal value improves stylization without sacrificing content integrity." + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.401, + 0.909, + 0.538 + ], + "angle": 0, + "content": "Text Expression for Multi-modal Alignment. In Table 3, we also explored different text representations for contrastive learning alignment. Specifically, we tested single labels (e.g., \"Old\"), brief descriptions (e.g., \"An old person\"), and more detailed descriptions (e.g., \"An old person is moving slow and stiff\"). Our experiments show that using a single label produced the best results, providing a clearer and more concise style signal for the model while avoiding unnecessary complexity in the text representation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.557, + 0.634, + 0.573 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this work, we introduce STYLEMOTIF, a novel stylized motion diffusion model capable of generating motion conditioned on both content and style from multiple modalities. Unlike prior approaches that either focus on motion generation across various content types or style transfer between sequences, STYLEMOTIF effectively synthesizes motion while incorporating stylistic cues from multi-modal inputs, including text, image, video, and audio. To achieve this, we introduce a style-content cross fusion mechanism and align a style encoder with a pre-trained multi-modal model. Through extensive experiments across diverse applications, we demonstrate that STYLEMOTIF outperforms existing methods in stylized motion generation, producing high-quality, realistic results that faithfully adhere to the given style references. Moreover, our model exhibits emergent capabilities for multi-modal motion stylization, enabling richer and more nuanced motion synthesis. These findings indicate the potential of STYLEMOTIF in advancing stylized motion generation and open new avenues for future research in multi-modal-driven motion synthesis and style-aware generative models." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13356" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.117, + 0.486, + 0.207 + ], + "angle": 0, + "content": "We would like to extend our sincere thanks to the Core AI and Avatar Movement team at Roblox for their guidance and support throughout this project. In particular, we are grateful to Haomiao Jiang, Vincent Petrella, Sean Palmer, and Simone Guggiari for their valuable feedback and insights. We also thank Lei Zhong and Zhe Li for their technical advice." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.225, + 0.188, + 0.24 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.25, + 0.486, + 0.291 + ], + "angle": 0, + "content": "[1] K. Aberman, Y. Weng, D. Lischinski, D. Cohen-Or, and B. Chen. Unpaired motion style transfer from video to animation. TOG, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.294, + 0.484, + 0.349 + ], + "angle": 0, + "content": "[2] Nefeli Andreou, Xi Wang, Victoria Fernández Abrevaya, Marie-Paule Cani, Yiorgos Chrysanthou, and Vicky Kalogeiton. Lead: Latent realignment for human motion diffusion. arXiv preprint arXiv:2410.14508, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.352, + 0.484, + 0.393 + ], + "angle": 0, + "content": "[3] Z. Cen, H. Pi, S. Peng, Z. Shen, M. Yang, S. Zhu, H. Bao, and X. Zhou. Generating human motion in 3d scenes from text descriptions. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.396, + 0.485, + 0.451 + ], + "angle": 0, + "content": "[4] Kexin Chen, Yuyang Du, Tao You, Mobarakol Islam, Ziyu Guo, Yueming Jin, Guangyong Chen, and Pheng-Ann Heng. Llm-assisted multi-teacher continual learning for visual question answering in robotic surgery. ICRA 2024, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.454, + 0.484, + 0.495 + ], + "angle": 0, + "content": "[5] L.H. Chen, S. Lu, A. Zeng, H. Zhang, B. Wang, R. Zhang, and L. Zhang. Motionllm: Understanding human behaviors from human motions and videos. ArXiv, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.498, + 0.482, + 0.539 + ], + "angle": 0, + "content": "[6] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In CVPR, 2023. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.542, + 0.484, + 0.611 + ], + "angle": 0, + "content": "[7] Sammy Christen, Shreyas Hampali, Fadime Sener, Edoardo Remelli, Tomas Hodan, Eric Sauser, Shugao Ma, and Bugra Tekin. Diffh2o: Diffusion-based synthesis of hand-object interactions from textual descriptions. In SIGGRAPH Asia 2024 Conference Papers, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.614, + 0.484, + 0.669 + ], + "angle": 0, + "content": "[8] Wenxun Dai, Ling-Hao Chen, Jingbo Wang, Jinpeng Liu, Bo Dai, and Yansong Tang. Motionlcm: Real-time controllable motion generation via latent consistency model. arXiv preprint arXiv:2404.19759, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.672, + 0.484, + 0.727 + ], + "angle": 0, + "content": "[9] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11162-11173, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.73, + 0.484, + 0.771 + ], + "angle": 0, + "content": "[10] Han Fang, Pengfei Xiong, Luhui Xu, and Yu Chen. Clip2video: Mastering video-text retrieval via image clip. arXiv preprint arXiv:2106.11097, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.774, + 0.484, + 0.829 + ], + "angle": 0, + "content": "[11] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xiangyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[12] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15180-15190, 2023. 2, 3, 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.25, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.134 + ], + "angle": 0, + "content": "[13] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, 2022. 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.907, + 0.176 + ], + "angle": 0, + "content": "[14] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In CVPR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.907, + 0.233 + ], + "angle": 0, + "content": "[15] Chuan Guo, Yuxuan Mu, Xinxin Zuo, Peng Dai, Youliang Yan, Juwei Lu, and Li Cheng. Generative human motion stylization in latent space. arXiv preprint arXiv:2401.13505, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.907, + 0.289 + ], + "angle": 0, + "content": "[16] Ziyu Guo, Renrui Zhang, Longtian Qiu, Xianzheng Ma, Xu-peng Miao, Xuming He, and Bin Cui. Calip: Zero-shot enhancement of clip with parameter-free attention. AAAI 2023 Oral, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.293, + 0.907, + 0.374 + ], + "angle": 0, + "content": "[17] Ziyu Guo, Renrui Zhang, Xiangyang Zhu, Yiwen Tang, Xianzheng Ma, Jiaming Han, Kexin Chen, Peng Gao, Xianzhi Li, Hongsheng Li, et al. Point-bind & point-llm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. arXiv preprint arXiv:2309.00615, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.377, + 0.907, + 0.446 + ], + "angle": 0, + "content": "[18] Zilu Guo, Hongbin Lin, Zhihao Yuan, Chaoda Zheng, Pengshuo Qiu, Dongzhi Jiang, Renrui Zhang, Chun-Mei Feng, and Zhen Li. Pisa: A self-augmented data engine and training strategy for 3d understanding with large models. arXiv preprint arXiv:2503.10529, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.448, + 0.907, + 0.516 + ], + "angle": 0, + "content": "[19] Ziyu Guo, Ray Zhang, Hao Chen, Jialin Gao, Dongzhi Jiang, Jiaze Wang, and Pheng-Ann Heng. Sciverse: Unveiling the knowledge comprehension and visual reasoning of lmms on multi-modal scientific problems. arXiv preprint arXiv:2503.10627, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.519, + 0.907, + 0.585 + ], + "angle": 0, + "content": "[20] Ziyu Guo, Renrui Zhang, Chengzhuo Tong, Zhizheng Zhao, Peng Gao, Hongsheng Li, and Pheng-Ann Heng. Can we generate images with cot? let's verify and reinforce image generation step by step. arXiv preprint arXiv:2501.13926, 2025.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.589, + 0.907, + 0.644 + ], + "angle": 0, + "content": "[21] Jiaming Han, Renrui Zhang, Wenqi Shao, Peng Gao, Peng Xu, Han Xiao, Kaipeng Zhang, Chris Liu, Song Wen, Ziyu Guo, et al. Imagebind-llm: Multi-modality instruction tuning. arXiv preprint arXiv:2309.03905, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.647, + 0.907, + 0.7 + ], + "angle": 0, + "content": "[22] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.703, + 0.907, + 0.758 + ], + "angle": 0, + "content": "[23] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.76, + 0.906, + 0.8 + ], + "angle": 0, + "content": "[24] Deok-Kyeong Jang, Soomin Park, and Sung-Hee Lee. Motion puzzle: Arbitrary motion style transfer by body part. ACM TOG, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.907, + 0.857 + ], + "angle": 0, + "content": "[25] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. Advances in Neural Information Processing Systems, 36:20067-20079, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[26] Dongzhi Jiang, Renrui Zhang, Ziyu Guo, Yanwei Li, Yu Qi, Xinyan Chen, Liuhui Wang, Jianhan Jin, Claire Guo, Shen Yan, et al. Mme-cot: Benchmarking chain-of-thought in large" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13357" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.483, + 0.12 + ], + "angle": 0, + "content": "multimodal models for reasoning quality, robustness, and efficiency. arXiv preprint arXiv:2502.09621, 2025.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.484, + 0.164 + ], + "angle": 0, + "content": "[27] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Gmd: Controllable human motion synthesis via guided diffusion models. In ICCV, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.483, + 0.233 + ], + "angle": 0, + "content": "[28] Boeun Kim, Jungho Kim, Hyung Jin Chang, and Jin Young Choi. Most: Motion style transformer between diverse action contents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1705-1714, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.484, + 0.29 + ], + "angle": 0, + "content": "[29] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer, 2024a. URL https://arxiv.org/abs/2408.03326. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.484, + 0.348 + ], + "angle": 0, + "content": "[30] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. ICLR 2025 Spotlight, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.35, + 0.484, + 0.432 + ], + "angle": 0, + "content": "[31] Xiaoqi Li, Mingxu Zhang, Yiran Geng, Haoran Geng, Yuxing Long, Yan Shen, Renrui Zhang, Jiaming Liu, and Hao Dong. Manipllm: Embodied multimodal large language model for object-centric robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18061-18070, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.483, + 0.475 + ], + "angle": 0, + "content": "[32] Yanghao Li, Naiyan Wang, Jiaying Liu, and Xiaodi Hou. Demystifying neural style transfer. arXiv preprint arXiv:1701.01036, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.483, + 0.546 + ], + "angle": 0, + "content": "[33] Zhe Li, Yisheng He, Lei Zhong, Weichao Shen, Qi Zuo, Lingteng Qiu, Zilong Dong, Laurence Tianruo Yang, and Weihao Yuan. Mulsmo: Multimodal stylized motion generation by bidirectional control flow. arXiv preprint arXiv:2412.09901, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.483, + 0.617 + ], + "angle": 0, + "content": "[34] Zhe Li, Weihao Yuan, Yisheng He, Lingteng Qiu, Shenhao Zhu, Xiaodong Gu, Weichao Shen, Yuan Dong, Zilong Dong, and Laurence T Yang. Lamp: Language-motion pretraining for motion generation, retrieval, and captioning. arXiv preprint arXiv:2410.07093, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.483, + 0.646 + ], + "angle": 0, + "content": "[35] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.648, + 0.483, + 0.689 + ], + "angle": 0, + "content": "[36] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In ICCV, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.69, + 0.483, + 0.759 + ], + "angle": 0, + "content": "[37] Ian Mason, Sebastian Starke, and Taku Komura. Real-time style modelling of human locomotion via feature-wise transformations and local motion phases. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 2022. 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[38] Arsha Nagrani, Paul Hongsuck Seo, Bryan Seybold, Anja Hauth, Santiago Manen, Chen Sun, and Cordelia Schmid. Learning audio-video modalities from image captions. In European Conference on Computer Vision, pages 407-426. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[39] Soomin Park, Deok-Kyeong Jang, and Sung-Hee Lee. *Diverse motion stylization for multiple style domains via spatial-temporal graph-based generative model.* Proceedings of the ACM on Computer Graphics and Interactive Techniques, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.147 + ], + "angle": 0, + "content": "[40] Xiaogang Peng, Yiming Xie, Zizhao Wu, Varun Jampani, Deqing Sun, and Huaizu Jiang. Hoi-diff: Text-driven synthesis of 3d human-object interactions using diffusion models. arXiv preprint arXiv:2312.06553, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.151, + 0.907, + 0.205 + ], + "angle": 0, + "content": "[41] Mathis Petrovich, Or Litany, Umar Iqbal, Michael J Black, Gul Varol, Xue Bin Peng, and Davis Rempe. Multi-track timeline control for text-driven 3d human motion generation. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.209, + 0.907, + 0.263 + ], + "angle": 0, + "content": "[42] Ekkasit Pinyoanuntapong, Muhammad Usama Saleem, Pu Wang, Minwoo Lee, Srijan Das, and Chen Chen. Bamm: Bidirectional autoregressive motion model. arXiv preprint arXiv:2403.19435, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.267, + 0.907, + 0.307 + ], + "angle": 0, + "content": "[43] Ekkasit Pinyoanuntapong, Pu Wang, Minwoo Lee, and Chen Chen. Mmm: Generative masked motion model. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.311, + 0.907, + 0.379 + ], + "angle": 0, + "content": "[44] Ziyun Qian, Zeyu Xiao, Zhenyi Wu, Dingkang Yang, Mingcheng Li, Shunli Wang, Shuaibing Wang, Dongliang Kou, and Lihua Zhang. Smcd: High realism motion style transfer via mamba-based diffusion. arXiv preprint arXiv:2405.02844, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.383, + 0.907, + 0.425 + ], + "angle": 0, + "content": "[45] Sigal Raab, Inbal Leibovitch, Guy Tevet, Moab Arar, Amit H Bermano, and Daniel Cohen-Or. Single motion diffusion. arXiv preprint arXiv:2302.05905, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.907, + 0.495 + ], + "angle": 0, + "content": "[46] Sigal Raab, Inbar Gat, Nathan Sala, Guy Tevet, Rotem Shalev-Arkushin, Ohad Fried, Amit H Bermano, and Daniel Cohen-Or. Monkey see, monkey do: Harnessing self-attention in motion diffusion for zero-shot motion transfer. arXiv preprint arXiv:2406.06508, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.499, + 0.907, + 0.568 + ], + "angle": 0, + "content": "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.572, + 0.906, + 0.626 + ], + "angle": 0, + "content": "[48] Davis Rempe, Zhengyi Luo, Xue Bin Peng, Ye Yuan, Kris Kitani, Karsten Kreis, Sanja Fidler, and Or Litany. Trace and pace: Controllable pedestrian animation via guided trajectory diffusion. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.63, + 0.906, + 0.671 + ], + "angle": 0, + "content": "[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.674, + 0.907, + 0.755 + ], + "angle": 0, + "content": "[50] Wenfeng Song, Xingliang Jin, Shuai Li, Chenglizhao Chen, Aimin Hao, Xia Hou, Ning Li, and Hong Qin. Arbitrary motion style transfer with multi-condition motion latent diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 821-830, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.907, + 0.828 + ], + "angle": 0, + "content": "[51] Xiangjun Tang, Linjun Wu, He Wang, Bo Hu, Xu Gong, Yuchen Liao, Songnan Li, Qilong Kou, and Xiaogang Jin. Rsmt: Real-time stylized motion transition for characters. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.907, + 0.899 + ], + "angle": 0, + "content": "[52] Yiwen Tang, Zoey Guo, Zhuhao Wang, Ray Zhang, Qizhi Chen, Junli Liu, Delin Qu, Zhigang Wang, Dong Wang, Xuelong Li, et al. Exploring the potential of encoder-free architectures in 3d Imms. arXiv preprint arXiv:2502.09620, 2025. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13358" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[53] Tianxin Tao, Xiaohang Zhan, Zhongquan Chen, and Michiel van de Panne. Style-erd: Responsive and coherent online motion style transfer. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.177 + ], + "angle": 0, + "content": "[54] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[55] Weilin Wan, Zhiyang Dou, Taku Komura, Wenping Wang, Dinesh Jayaraman, and Lingjie Liu. Tlcontrol: Trajectory and language control for human motion synthesis. arXiv preprint arXiv:2311.17135, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[56] Yu-Hui Wen, Zhipeng Yang, Hongbo Fu, Lin Gao, Yanan Sun, and Yong-Jin Liu. Autoregressive stylized motion synthesis with generative flow. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[57] Qianyang Wu, Ye Shi, Xiaoshui Huang, Jingyi Yu, Lan Xu, and Jingya Wang. Thor: Text to human-object interaction diffusion via relation intervention. arXiv preprint arXiv:2403.11208, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[58] Qi Wu, Yubo Zhao, Yifan Wang, Yu-Wing Tai, and Chi-Keung Tang. Motionllm: Multimodal motion-language learning with large language models. arXiv preprint arXiv:2405.17013, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.433 + ], + "angle": 0, + "content": "[59] Yiming Xie, Varun Jampani, Lei Zhong, Deqing Sun, and Huaizu Jiang. Omnicontrol: Control any joint at any time for human motion generation. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[60] Jingwei Xu, Huazhe Xu, Bingbing Ni, Xiaokang Yang, Xiaolong Wang, and Trevor Darrell. Hierarchical style-based networks for motion synthesis. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.482, + 0.518 + ], + "angle": 0, + "content": "[61] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. Interdiff: Generating 3d human-object interactions with physics-informed diffusion. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.52, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[62] Sirui Xu, Ziyin Wang, Yu-Xiong Wang, and Liang-Yan Gui. Interdreamer: Zero-shot text to 3d dynamic human-object interaction. arXiv preprint arXiv:2403.19652, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.563, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[63] Le Xue, Mingfei Gao, Chen Xing, Roberto Martin-Martin, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1179-1189, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.714 + ], + "angle": 0, + "content": "[64] Senqiao Yang, Jiaming Liu, Ray Zhang, Mingjie Pan, Ziyu Guo, Xiaqi Li, Zehui Chen, Peng Gao, Yandong Guo, and Shanghai Zhang. Lidar-llm: Exploring the potential of large language models for 3d lidar understanding. AAAI 2025, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.717, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[65] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[66] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3836-3847, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[67] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. PAMI, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[68] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Point" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.908, + 0.119 + ], + "angle": 0, + "content": "clip: Point cloud understanding by clip. In CVPR 2022, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.908, + 0.189 + ], + "angle": 0, + "content": "[69] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In ECCV 2022. Springer Nature Switzerland, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.907, + 0.247 + ], + "angle": 0, + "content": "[70] Renrui Zhang, Jiaming Han, Chris Liu, Aojun Zhou, Pan Lu, Yu Qiao, Hongsheng Li, and Peng Gao. Llama-adapter: Efficient fine-tuning of large language models with zero-initialized attention. In ICLR 2024, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.907, + 0.316 + ], + "angle": 0, + "content": "[71] Renrui Zhang, Dongzhi Jiang, Yichi Zhang, Haokun Lin, Ziyu Guo, Pengshuo Qiu, Aojun Zhou, Pan Lu, Kai-Wei Chang, Peng Gao, et al. Mathverse: Does your multi-modal llm truly see the diagrams in visual math problems? ECCV 2024, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.907, + 0.387 + ], + "angle": 0, + "content": "[72] Renrui Zhang, Xinyu Wei, Dongzhi Jiang, Ziyu Guo, Shicheng Li, Yichi Zhang, Chengzhuo Tong, Jiaming Liu, Aojun Zhou, Bin Wei, et al. Mavis: Mathematical visual instruction tuning with an automatic data engine. arXiv preprint arXiv:2407.08739, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.906, + 0.431 + ], + "angle": 0, + "content": "[73] Lei Zhong, Yiming Xie, Varun Jampani, Deqing Sun, and Huaizu Jiang. Smoodi: Stylized motion diffusion model. In ECCV, 2024. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.907, + 0.501 + ], + "angle": 0, + "content": "[74] Bin Zhu, Bin Lin, Munan Ning, Yang Yan, Jiaxi Cui, HongFa Wang, Yatian Pang, Wenhao Jiang, Junwu Zhang, Zongwei Li, et al. Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. arXiv preprint arXiv:2310.01852, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.503, + 0.907, + 0.558 + ], + "angle": 0, + "content": "[75] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. ICCV 2023, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13359" + } + ] +] \ No newline at end of file diff --git a/2025/StyleMotif_ Multi-Modal Motion Stylization using Style-Content Cross Fusion/ffae00f3-5394-4dd8-b9a4-bc3713aa96e3_origin.pdf b/2025/StyleMotif_ Multi-Modal Motion Stylization using Style-Content Cross Fusion/ffae00f3-5394-4dd8-b9a4-bc3713aa96e3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..241858c0fd0c51a0aaf82f4cfeb4a03eb6b2388f --- /dev/null +++ b/2025/StyleMotif_ Multi-Modal Motion Stylization using Style-Content Cross Fusion/ffae00f3-5394-4dd8-b9a4-bc3713aa96e3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7dd8d19fe1770e1670e603c148d14c0dc1f7e626b3f6acf373b9d63665aded20 +size 2200734 diff --git a/2025/StyleMotif_ Multi-Modal Motion Stylization using Style-Content Cross Fusion/full.md b/2025/StyleMotif_ Multi-Modal Motion Stylization using Style-Content Cross Fusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8d8edee0c0e46a3b7c87813ef3ad5a27535ed903 --- /dev/null +++ b/2025/StyleMotif_ Multi-Modal Motion Stylization using Style-Content Cross Fusion/full.md @@ -0,0 +1,303 @@ +# STYLEMOtIF: Multi-Modal Motion Stylization using Style-Content Cross Fusion + +
Ziyu Guo†Young Yoon LeeJoseph Liu
The Chinese University of Hong KongRobloxRoblox
ziyuguo@link.cuhk.edu.hkylee@roblox.comjosephliu@roblox.com
Yizhak Ben-ShabatVictor ZordanMubbasir Kapadia
RobloxRobloxRoblox
ibenshabat@roblox.comvbzordan@roblox.commkapadia@roblox.com
+ +# Abstract + +We present STYLEMOTIF, a novel Stylized Motion Latent Diffusion model, generating motion conditioned on both content and style from multiple modalities. Unlike existing approaches that either focus on generating diverse motion content or transferring style from sequences, STYLEMOTIF seamlessly synthesizes motion across a wide range of content while incorporating stylistic cues from multi-modal inputs, including motion, text, image, video, and audio. To achieve this, we introduce a style-content cross fusion mechanism and align a style encoder with a pre-trained multi-modal model, ensuring that the generated motion accurately captures the reference style while preserving realism. Extensive experiments demonstrate that our framework surpasses existing methods in stylized motion generation and exhibits emergent capabilities for multi-modal motion stylization, enabling more nuanced motion synthesis. Project Page: https://stylemotif.github.io. + +# 1. Introduction + +Human motion generation is a fundamental task in computer graphics and animation, enabling the synthesis of realistic and expressive human movements. Broadly, human motion can be characterized by two complementary aspects: content, which defines the underlying action (e.g., walking, jumping), and style, which encodes variations such as personal flair, emotional expression, or cultural influences (e.g., jubilant, aggressive). This separation allows for greater control and flexibility in generating motion, making it particularly valuable in creative industries like game development, film production, and virtual reality. However, traditional approaches to stylized motion generation often depend on manual pro + +![](images/a2c6ee5da755e8bdbcefd5d585a697bc8bbc4745f64d4ec6b582eee470f116b3.jpg) +Figure 1. Comparison of Our Proposed STYLEMOTIF with SMooDi. SMooDi duplicates and separates two diffusion backbones for content and style, increasing complexity. In contrast, STYLEMOTIF adopts a single diffusion backbone with direct style injection, enabling efficient and realistic motion stylization. + +cesses such as motion capture or keyframe animation, which are costly, time-consuming, and labor-intensive. + +Recent progress in text-to-motion (T2M) diffusion frameworks [6, 27, 54] has greatly advanced the ability to translate natural-language prompts into realistic human motions. By leveraging powerful denoising diffusion models, these approaches capture intricate spatiotemporal dependencies in the data, enabling coherent sequences of human movements to be generated directly from brief textual descriptions. Despite their success in content fidelity and diversity, most current T2M diffusion methods concentrate primarily on what action is performed, while overlooking how it is performed, namely, its stylistic details. Simply appending a separate style-transfer module to text-driven motion diffusion pipelines can introduce additional complexity and risk of compounding errors in the final output. + +In parallel, motion style transfer has been actively studied to infuse stylistic cues from a reference motion (or style + +data) into another sequence [1, 24, 37, 53]. Although many of these methods effectively disentangle content and style for small-scale tasks, the pipeline becomes cumbersome when a large variety of content motions need to be stylized. Moreover, they commonly assume that the input or target sequences are high-quality motion data. In scenarios where content motions are synthetically generated, or partially noisy, the transfer process can deteriorate, leading to undesirable motion artifacts or compromised style fidelity. + +To address the need for simultaneously controlling both content and style, some recent works have merged style encoding with diffusion-based motion generation. Among these, the most recent and representative approach [73] augments a pre-trained latent diffusion model [6] with a style adaptor and classifier-based style guidance, achieving stylized motion from textual prompts and motion-style references. While effective, this method relies on additional training diffusion backbones, which shares structural similarities with ControlNet [66] as shown in Figure 1, which increases model complexity and training overhead. It is also constrained to motion as the primary style input. Concurrent work [33] proposes a bidirectional control flow mechanism to mitigate conflicts between style and content, extending style control to multiple modalities, but this approach also duplicates two separate diffusion backbones with substantial training overhead and limited applicability across modalities. + +To this end, we propose STYLEMOTIF, a new framework for multi-modal motion stylization that unifies text-to-motion diffusion with style conditioning in a single-diffusion structure. Specifically, we leverage a pre-trained motion latent diffusion model (MLD) [6] to preserve strong content generation capabilities, and seamlessly integrate style features extracted from a dedicated encoder, which is aligned with a multi-modal foundation model [12]. In contrast to previous works, STYLEMOTIF avoids duplicating large portions of the network or relying on specialized style diffusion network. Instead, we introduce a style-content cross fusion mechanism, which injects stylistic cues into the diffusion process while maintaining motion realism. As a result, our STYLEMOTIF not only yields more robust stylized outputs but also supports diverse style signals, such as motion, text, images, audio, or video clips, via the alignment in multi-modal feature space. We summarize our main contributions as follows: + +- We present STYLEMOTIF, a stylized latent motion diffusion framework that unifies diverse motion content and multi-modal styles within a compact single-diffusion design. +- We propose a style-content cross fusion module that injects stylistic cues into the diffusion denoising process, achieving faithful stylization without compromising motion realism and ensuring efficiency. +- We achieve a unified multi-modal style feature space and unveil new emergent capabilities through multi-modal + +alignment, which accommodate various sources including motion, text, images, audio, and video, for flexible and versatile multi-modal style control. + +- Extensive experiments demonstrate that STYLEMOTIF consistently outperforms existing methods regarding style expressiveness, content preservation, and efficiency. + +# 2. Related Work + +Human Motion Generation. Recent progress in human motion generation [3, 5, 8, 13, 14, 41-43, 46, 55, 57, 58, 62] has been driven by transformer [14, 34] and diffusion models [2, 8, 49, 66], showing great potential in producing realistic and diverse motions. These approaches have shown great potential in producing realistic and diverse motion sequences. For example, Momask [14] improves motion generation using a residual VQ-VAE. Similarly, LaMP [34] introduces a motion-aware text encoder and a motion-to-text language model to enhance motion quality through text conditioning. Diffusion models, in particular, have become a key approach in motion generation [6, 22, 27, 48, 54, 61, 65, 67]. MDM [54] introduces a motion diffusion model that operates directly on raw motion data to capture the relationship between motions and input text conditions. MLD [6] improve efficiency by embedding the diffusion process in latent space, reducing computational cost. They also allow conditioning on specific constraints, such as predefined trajectories [27, 55, 59] or human-object interactions [7, 40, 57], enabling greater control and diversity. Our work leverages pretrained motion latent diffusion model [6] and multi-modal foundation models [12] to achieve stylized human motion generation while maintaining high motion quality. + +Motion Stylization. Motion stylization [1, 24, 28, 37, 39, 44, 45, 50, 51, 53, 56, 60] involves transferring stylistic features from a reference motion to a source motion/textual prompt, enabling creative transformations while preserving the original motion content. Early methods, like those in motion style transfer [1, 24], typically separate motion content and style for recombination. For instance, Aberman et al. [1] used a generative adversarial network to decouple style from content without paired data, while Motion Puzzle [24] allows style control for individual body parts, and Guo et al. [15] utilized pretrained motion models for better style integration. Recent approaches, like SMooDi [73], generate stylized motion from text and style sequences using style guidance. However, these models face two main challenges: (1) inefficient architectures that duplicate diffusion backbones for content and style, and (2) limited to style motion for guidance. Another concurrent approach, MulSMo [33], attempts to address some of these by introducing a bidirectional control flow between style and content networks, reducing conflicts. Still, it remains restricted to a few modalities and duplicates diffusion networks. In this + +![](images/e43eacda52e40f7d1fbd364185ec40d421eb98e999aea47c2f37b9f95de4d56e.jpg) +Figure 2. Overall Pipeline of STYLEMOTIF, a diffusion framework for multi-modal motion stylization. Given a text prompt and a reference style from various modalities, our model extract style features and fuse them with content by style-content cross fusion. Through multi-modal alignment with contrastive learning, we enable seamless multi-modal conditioning and flexible stylization across motion, text, images, audio, and video. + +work we eliminate redundant diffusion backbone, resulting in a simpler, more effective, and efficient approach to stylized motion generation. Our method also enables efficient multi-modal motion stylization, supporting text, image and multi-modal inputs while maintaining both high stylization quality and motion realism. + +Multi-modality Learning. Recent advancements in multimodality learning [9, 10, 38] have revolutionized various domains by enabling joint understanding across diverse data types. Foundational models like CLIP [47] and its extension [16, 68, 69, 75] establish robust vision-language alignment, while ImageBind [12], Point-Bind [17], and Language-Bind [74] expand this paradigm to diverse modalities, demonstrating emergent cross-modal capabilities. The rise of multi-modal large language models [4, 19, 21, 30, 31, 64, 72] further enhances semantic understanding through unified text-visual processing, achieving superior performance across 2D images [11, 29, 70], 3D point clouds [17, 18, 52], and complex reasoning scenarios [20, 26, 71]. For human motion, some approaches [5, 14, 25] incorporate multiple modalities into human motion understanding and generation tasks. Our work introduces the first framework for multimodal guided motion stylization, achieving seamless style injection from diverse modalities through a single diffusion architecture with style-content cross fusion. + +# 3. STYLEMOTIF: Multi-Modal Motion Stylization with Style-Content Cross Fusion + +We propose STYLEMOTIF, a novel framework for stylized motion synthesis that combines style-content cross fusion and multi-modal motion stylization, as illustrated in Figure 2. Our approach integrates a style-content cross fusion mechanism that allows for coherent feature blending, ensuring the generated motion accurately reflects the reference style while maintaining the content's realism (§3.2). This mechanism is complemented by a multi-modal motion stylization strategy, which leverages inputs from diverse modalities such as image, video, audio, and text to provide control over the stylization process (§3.3). This framework enables highly flexible and realistic stylized motion synthesis, which offers greater control and diversity. + +# 3.1. Overview + +Preliminaries. Motion Latent Diffusion (MLD) [6] formulates a conditional latent diffusion model by training $\epsilon_{\theta}(z_t,t,c)$ to denoise a sequence of latents $\{z_t\}_{t = 0}^T$ , where $z_{t}\in \mathbb{R}^{n\times d}$ represents the motion latent at timestep $t$ , conditioned on the distribution $p(z_{t}|c)$ . A conditional domain encoder $\tau_{\theta}(c)$ , such as CLIP [47], enables text-to-motion tasks, and the model is trained as $\epsilon_{\theta}(z_t,t,\tau_{\theta}(c))$ . To incorporate additional style conditioning, a style condition $s$ can be introduced with its own encoder $\psi_{\theta}(s)$ , which may take the form of a motion or text encoder. This extends the model to $\epsilon_{\theta}(z_t,t,\tau_{\theta}(c),\psi_{\theta_s}(s))$ . SMooDi [73] adopts a ControlNet [66]-style approach for style conditioning by + +creating a trainable copy of the neural network weights $\theta_{s}$ from the original MLD model $\theta_{c}$ . This copy includes zero-initialized linear layers, denoted as $\mathcal{Z}(\cdot ;\cdot)$ . The output $\mathcal{F}^i (z_t,t,\tau_\theta (c),\psi_\theta (s);\theta_c)$ of the $i$ -th MLD block, now conditioned on style, is computed as: + +$$ +\begin{array}{l} \mathcal {F} ^ {i} \left(z _ {t}, t, \tau_ {\theta} (c), \psi_ {\theta_ {s}} (s); \theta_ {c}\right) = \mathcal {F} ^ {i} \left(z _ {t}, t, \tau_ {\theta} (c); \theta_ {c}\right) \\ + \mathcal {Z} \left(\mathcal {F} ^ {i} \left(z _ {t}, t, \tau_ {\theta} (c), \psi_ {\theta_ {s}} (s); \theta_ {s}\right); \theta_ {z _ {i}}\right) \tag {1} \\ \end{array} +$$ + +A key property of this formulation is that since $\theta_{z_i}$ is initialized to zero, thus $\mathcal{F}^i (z_t,t,\tau_\theta (c),\psi_\theta (s)) = \mathcal{F}^i (z_t,t,\tau_\theta (c))$ at the beginning. However, the tradeoff of that method is that it needs to maintain the additional parameters $\theta_{s}$ and $\theta_{z_i}$ in order for style transfer. + +STYLEMOTIF. We propose STYLEMOTIF, eschewing zero linear layers in favor of injecting it directly via statistical manipulation, with the pipeline shown in Figure 2 describing our approach. Our STYLEMOTIF utilizes latent space diffusion within a single generative diffusion backbone, building upon a pretrained MLD model [6]. Instead of perturbing the outputs of each block $i$ via zero initialized layer $\mathcal{Z}(\mathcal{F}^i (z_t,t,\tau_\theta (c),\psi_\theta (s);\theta_s);\theta_{z_i})$ , STYLEMOTIF replaces this with a statistically transformed style embedding that is injected into the original MLD branch. This simplifies the model while ensuring high-quality stylization results, whose details are presented in § 3.2. Also, we follow SMooDi's generation guidance and training scheme to ensure high-quality stylized motion synthesis. We use a hybrid guidance strategy that balances content fidelity and style adherence by combining classifier-free and classifier-based techniques during diffusion sampling. + +# 3.2. Style-Content Cross Fusion + +Style Encoder Pre-training. To establish a robust foundation for style-content fusion, we combine the content knowledge from the pre-trained MLD's [6] VAE with the style knowledge from the 100STYLE [37] dataset. The MLD's VAE is pre-trained on the HumanML3D [13] dataset, which provides extensive understanding of content motion. Building on this, we further fine-tune the model on the 100STYLE dataset in a variational autoencoding manner to initially align the content and style data distributions in the latent space. After training, we discard the decoder and retain only the encoder as the motion style encoder. This reconstruction task enables the encoder to learn robust motion feature representations, which are essential for supporting the subsequent stylization process. By integrating content and style knowledge in this way, we ensure a strong foundation for seamless style-content fusion. + +Style-Content Cross Normalization. To train a stylized diffusion model $\epsilon_{\theta}(z_t, t, \tau_{\theta}(c), \psi_{\theta}(s))$ , we effectively fuse + +content and style features directly within the latent space diffusion process instead of duplicating two pre-trained diffusion backbones for content and style or separate control mechanisms [66]. Prior work [23, 32] has proven that feature statistics (mean, variance) effectively encode style. Given the output features $\mathcal{F}^i$ of the $i$ -th block, we derive the content features $\mathcal{F}_c^i = \mathcal{F}^i(z_t, t, \tau_\theta(c); \theta_c)$ from the input text $c$ , and the style features $\mathcal{F}_s = \psi_{\theta_s}(s)$ are extracted from the reference motion sequence $s$ using the pre-trained style encoder. To perform the fusion, we first compute the mean $\mu_c$ and variance $\sigma_c^2$ of the content features across the feature dimension: + +$$ +\mu_ {c} = \frac {1}{D} \sum_ {j = 1} ^ {D} \mathcal {F} _ {c} ^ {i, j}, \tag {2} +$$ + +$$ +\sigma_ {c} ^ {2} = \frac {1}{D} \sum_ {j = 1} ^ {D} \left(\mathcal {F} _ {c} ^ {i, j} - \mu_ {c}\right) ^ {2}, \tag {3} +$$ + +where $\mathcal{F}_c^{i,j}$ denotes the content features of the $i$ -th block and the $j$ -th feature element. Next, we normalize the style features $\mathcal{F}_s$ using these content statistics, ensuring that the style features are adapted to the content's statistical properties: + +$$ +\widetilde {\mathcal {F}} _ {s, c} = \frac {\mathcal {F} _ {s} - \mu_ {c}}{\sqrt {\sigma_ {c} ^ {2} + \eta}}, \tag {4} +$$ + +where $\eta$ is a constant added for numerical stability. This style-content cross normalization ensures that the style features are smoothly integrated with the content features while maintaining the content's original structure. After that, we add the normalized style features to the content features, formulating our final cross-normalization as + +$$ +\mathcal {F} ^ {i} \left(z _ {t}, t, \tau_ {\theta} (c), \psi_ {\theta_ {s}} (s); \theta_ {c}\right) = \mathcal {F} _ {c} ^ {i} + \gamma \cdot \tilde {\mathcal {F}} _ {s, c}, \tag {5} +$$ + +where $\gamma$ is a parameter used for scaling the normalized value. $\widetilde{\mathcal{F}}_{s,c}$ can be thought of as a perturbation of $\mathcal{F}_c^i$ as it is scaled within its range. Notably, the fusion process is performed only once after the $m$ -th block during the denoising process to avoid distorting the content while effectively introducing the style. This efficient fusion method eliminates the need for additional learnable parameters, as it is based solely on statistical transformations, achieving high-quality results with minimal computational overhead. + +# 3.3. Multi-Modal Motion Stylization + +Our STYLEMOTIF extends to support multi-modal motion stylization. To achieve this, we integrate a pre-trained multimodal foundation model [12], which provides a unified, multi-modal aligned feature space, enabling effective cross-modal alignment between motion and other modalities. With this alignment, our model can flexibly combine multiple + +input modalities, such as text, image, audio, and video, to guide the stylization process in a comprehensive manner. This capability results in emergent abilities for multi-modal motion stylization, allowing for more nuanced outputs. + +Motion-Text Pair Curation. To align motion with other modalities, we process a set of motion-text pairs using the curated 100STYLE subset [37], which is carefully selected to avoid conflicts between content and style motions, so that the model can effectively learn the relationships between motion and text. Each motion sequence is paired with a corresponding single textual label, which serves as the text prompt. The curated motion-text pairs will be used for the following cross-model alignment. + +Multi-Modal Alignment. To maintain the alignment within the multi-modal space, we freeze the text encoder of ImageBind [12] and introduce a lightweight projection layer after the encoder. This projection layer aligns the feature dimensions of the text and motion encoders. Since the multi-modal model represents each modality with a global feature, our alignment process focuses on these global features to achieve robust alignment between text and motion representations, formalized as: + +$$ +\mathcal {F} _ {t} = \pi \left(\mathcal {E} _ {\text {t e x t}} (l)\right), \tag {6} +$$ + +$$ +\mathcal {F} _ {s} = \psi_ {\theta_ {s}} (s), \tag {7} +$$ + +where $l$ and $s$ are the paired input text label and style motion, $\mathcal{F}_t$ and $\mathcal{F}_s$ denote the feature representations from the text and style motion encoders, $\mathcal{E}_{\mathrm{text}}$ and $\psi_{\theta_s}$ , respectively, and $\pi$ represents the projection operation that maps text features into the motion feature space. We employ a contrastive learning loss [63] to align the feature spaces of motion and text, bringing them closer together in the shared multi-modal space. Thus, we obtain a unified space between motion and all the modalities. We formulated it as + +$$ +\mathcal {L} _ {\text {a l i g n}} = - \frac {1}{2} \sum_ {(i, j)} \log \frac {\exp \frac {\mathcal {F} _ {t} ^ {i} \cdot \mathcal {F} _ {s} ^ {j}}{\tau_ {0}}}{\sum_ {k} \exp \frac {\mathcal {F} _ {t} ^ {i} \cdot \mathcal {F} _ {s} ^ {k}}{\tau_ {0}}} + \log \frac {\exp \frac {\mathcal {F} _ {t} ^ {i} \cdot \mathcal {F} _ {s} ^ {j}}{\tau_ {0}}}{\sum_ {k} \exp \frac {\mathcal {F} _ {t} ^ {k} \cdot \mathcal {F} _ {s} ^ {j}}{\tau_ {0}}}, \tag {8} +$$ + +where $t$ and $s$ represent two modalities (text and style motion) and $(i,j)$ indicates a positive pair in each training batch, $k$ indexes all samples in the batch, including both positive and negative ones, and $\tau_0$ is a temperature parameter. During inference, we obtain the multi-modal (text, image, video, or audio) features from the multi-modal input $m$ , + +$$ +\mathcal {F} _ {m} = \mathcal {E} _ {\text {I m a g e B i n d}} (m), \tag {9} +$$ + +where $\mathcal{E}_{\mathrm{ImageBind}}$ denotes ImageBind [12]. We use the multimodal features $\mathcal{F}_m$ to retrieve the most semantically similar + +motion features from the unified multi-modal space. The retrieved motion features are then used to guide the stylization process, ensuring that fine-grained style details are preserved and accurately reflected in the final output. This approach enhances the model's ability to generate stylized motions that are both contextually relevant and visually coherent. + +# 4. Experiment + +We provide extensive quantitative and qualitative analyses across multiple tasks in this section. + +# 4.1. Experimental Settings + +Implementation Details. We adopt the pre-trained MLD [6] as the foundation for motion generation. The style encoder of our model derives from the encoder of MLD's VAE, while the projection layer after text encoder is a single Linear layer. During diffusion training, we only enable the style encoder to be trainable while freezing other parameters. The model is optimized using the AdamW optimizer [35] with a constant learning rate of $10^{-5}$ . + +Dataset Settings. We use the HumanML3D [13] dataset as our primary motion content dataset, which collects 14,616 motion sequences from AMASS [36] and annotates 44,970 sequence-level textual description. To train the style network, we utilize the 100STYLE [37] dataset, containing 45,303 style motions. We also adopt the text annotations for 100STYLE from previous work [73], which are pseudo text descriptions generated from MotionGPT [25]. We utilize the consistent root-velocity motion representations from HumanML3D for both the content and style data. + +Baselines. We mainly evaluate our method against SMooDi [73] on motion-guided stylization and motion style transfer tasks. We also compare our text-guided stylcification with 'ChatGPT+MLD' approach, which utilize ChatGPT to combine style text and content text, and functions as a straightforward text-to-motion model without control capabilities [73]. Additionally, we qualitatively compare our model with SMooDi on motion-guided stylization. + +Evaluation Metrics. To evaluate our model, we use several metrics following previous works [73]. We measure R-precision, Multi-modal Distance (MM Dist), Diversity, and Frechet Inception Distance (FID) to assess how well the content from text is preserved and how realistic the generated motions are. We use Style Recognition Accuracy (SRA) to evaluate how accurately the style of the reference motion is reflected in the generated motion. Also, we adopt the Foot Skate Ratio, which helps assess the realism of the generated motion, reducing artifacts like foot sliding. During evaluation, following previous works [73], we randomly select + +
MethodSRA ↑FID ↓MM Dist ↓R-Precision ↑ (Top-3)Diversity →Foot Skate Ratio ↓
Motion-Guided Stylization
MLD + Aberman [1]54.373.3095.9830.4068.8160.347
MLD + Motion Puzzle [24]63.776.1276.4670.2906.4760.185
SMooDi [73]72.421.6094.4770.5719.2350.124
STYLEMOTIF (Ours)77.651.5514.3540.5867.5670.097
Text-Guided Stylization
MLD + ChatGPT [73]4.820.6144.3130.6058.8360.131
STYLEMOTIF (Ours)56.710.6033.6840.6909.1010.101
+ +Table 1. Quantitative Results for Motion-Guided and Text-Guided Stylization. Bold values denote the best performance. As there is no ground-truth reference for Diversity, no value is highlighted in bold; but the metric is provided for reference. + +
MethodSRA ↑FID ↓Foot Skate Ratio ↓
MLD + Aberman [1]61.013.8920.338
MLD + Motion Puzzle [24]67.236.8710.197
SMooDi [73]65.151.5820.095
STYLEMOTIF (Ours)68.811.3750.094
+ +Table 2. Quantitative Results of Motion Style Transfer on HumanML3D [13] dataset. Our method outperforms previous works in all metrics, which demonstrates effective style-content fusion for high-quality motion style transfer, providing significant advantages for downstream tasks besides motion stylization. + +![](images/20b2c436f31ed23d4ea56d7967d06c6d7253e3f3a8fa14e2bf3162f090148f20.jpg) +Figure 3. Qualitative Results of Motion-Guided Stylization. Our model generates cohesive and realistic motions that effectively align style and content, such as preserving the 'circular' trajectory (first column) and 'hop' content (third column). In contrast, SMooDi [73] struggles to maintain content fidelity and sometimes fails to reflect the specified style (e.g., 'phone on the left' in the second column). + +a content text from HumanML3D [13] and a style motion from 100STYLE [37], and compute SRA for the generated motion using a pre-trained classifier [73]. + +# 4.2. Motion-guided Stylization + +Quantitative Analysis. In Table 1, we report the quantitative results for motion-guided stylization, where the motion serves as style input while text prompt as content one. Our method outperforms three baseline approaches [1, 24, 73] in all metrics. Specifically, we achieve a $5.23\%$ improvement in SRA while maintaining competitive performance in FID compared to the best baseline [73]. It demonstrates that our method, by effectively utilizing style-content cross fusion, generates motions that better align with the style reference + +while maintaining content integrity, showcasing the strength of our design in balancing style and content preservation. + +Qualitative Analysis. In Figure 3, we show some qualitative results of motion-guided stylization from our model and baseline [73]. Compared with the baseline, our model produces more cohesive stylized motions, with better alignment of style and content. For instance, in the provided examples, SMooDi fails to maintain the 'circular' trajectory or 'hop' action specified by the content or cannot accurately reflect the intended style, 'phone on the left'. This indicates that our approach more effectively integrates style and content, resulting in more realistic and consistent stylized motions. + +![](images/a54fce626f941968a92bdb3eb80e90a89f5fe555534ab4bbc9641437e19dfd07.jpg) +Figure 4. Qualitative Results of Text-Guided Stylization. Our model seamlessly integrates textual style descriptions with content, producing visually coherent and stylistically consistent results. +Figure 5. Qualitative Results of Style Interpolation. Our model blends multiple style inputs while preserving content integrity, demonstrating effective style-content fusion of our model. + +![](images/e9bc3135ae16e288144a830654b13ddd14b9d2787f6006ea4a9e418075dfdc81.jpg) +Figure 6. Qualitative Results of Multi-Modal Motion Stylization. Our model generates stylized motions guided by diverse modalities (e.g., text, image, video, audio), effectively transferring style while preserving content integrity. + +# 4.3. Text-guided Stylization + +Quantitative Analysis. For the text-guided stylization task, we use the HumanML3D [13] dataset for motion content and employ the single text label [37] as the style control. As reported in Table 1 (Bottom), our method significantly outperforms the baseline, achieving $56.71\%$ in SRA, compared to $4.82\%$ for 'ChatGPT + MLD'. Moreover, our method maintains competitive FID, balancing style reflection with content preservation. This indicates that our design, which combines multi-modal alignment, helps utilize the style signal in the motion-text shared spaces and achieve significant effectiveness in text-guided stylization. + +Qualitative Analysis. In Figure 4, we showcase qualitative results for text-guided stylization, where our model also demonstrates strong capability in harmonizing style and content, generating high-quality and visually coherent results. + +# 4.4. Motion Style Transfer + +For the motion style transfer task, we use the HumanML3D [13] dataset for motion content and the + +100STYLE dataset [37] for motion styles. As shown in Table 2, our method outperforms the baseline models across all key metrics. Specifically, we achieve a $3.66\%$ improvement in SRA, indicating better style reflection, and a 0.207 reduction in FID reflecting improved realism. These demonstrate that our framework not only improves the alignment between style and content but also provides significant advantages for downstream tasks like motion style transfer with efficient style-content cross fusion. By maintaining content integrity while seamlessly integrating style, STYLEMOTIF ensures more realistic and dynamically consistent stylized motions, which is critical for high-quality motion style transfer. + +# 4.5. Multi-Modal Motion Stylization + +With the power of the aligned multi-modal space, our model supports stylization guided by a variety of modalities, including motion, text, image, video, audio. Here, we utilize the input style feature to retrieve the corresponding motion features as the style condition, while the content input is provided as text prompt. The model then generates a stylized motion that incorporates the style from the input modality + +![](images/ef6e357332c4ece566d7503d051fb462663a343173f66097cd6cbb344ff33fac.jpg) +Figure 7. Ablation Study on Scaling Ratio $\gamma$ in Eq. 5. We report both SRA and FID to show the impact of the scaling ratio on both stylization and content preservation. + +
Style Encoder Pre-training Strategy
HumanML3D100STYLESRA ↑FID ↓MM Dist ↓R-Precision ↑Diversity →Foot Skate Ratio ↓
-✓76.731.7884.3490.5717.5050.101
✓-76.581.6354.4580.5728.5340.109
✓✓77.651.5514.3540.5867.5670.097
Text in Multi-modal Alignment
Textual ExpressionSRA ↑FID ↓MM Dist ↓R-Precision ↑Diversity →Foot Skate Ratio ↓
Brief Description76.841.5804.3780.5787.2510.099
Detailed Description75.251.6224.4190.5637.7640.102
Single Text Label77.651.5514.3540.5867.5670.097
+ +Table 3. Ablation Study on Style Encoder Pre-training Strategies and Text Expression for Multi-modal Alignment. 'w. HumanML3D' and 'w. 100STYLE' denote pre-training with HumanML3D [13] and 100STYLE [37] data respectively. + +while maintaining the content integrity as specified by the text prompt. We showcase several examples of multi-modal motion stylization in Figure 6, where different modalities guide the motion generation. For instance, when a text content "A person is walking." is provided alongside an image of a duck as style input, the model retrieves a relevant motion feature and blends the content with the style of 'Duckfoot'. As shown, the style from various inputs (text, image, video, audio) is effectively transferred to the generated motion. + +# 4.6. Style Interpolation + +Leveraging the aligned multi-modal space, our model enables text-guided style interpolation. Given one content text along with at least two style style texts, our model generates a motion that combines the characteristics of all input styles. When style texts are provided, we retrieve the most similar style motion features from the shared space. These features are then combined by weighted summation. The combined style features are fused with the content features using the proposed style-content cross fusion. In Figure 5, we showcase some qualitative results of style interpolation in, where the generated motions successfully blend the characteristics of both styles while maintaining the content's integrity. + +# 4.7. Ablation Study + +Style Encoder Pre-training. In Table 3, we investigate the impact of different pre-training strategies for the style encoder. Our results show that pre-training on both HumanML3D [13] and 100STYLE [37] yields the best performance. Compared to pre-training solely on either one, the combined training on both datasets provides the style encoder with a richer set of prior knowledge. This enables the model to effectively balance content preservation and style reflection, leveraging the diverse characteristics of both datasets to enhance the overall stylization quality. + +Cross Normalization Scaling Ratio. In Figure 7, we examine the effect of different scaling ratios $\gamma$ in Eq. 5 on the stylization performance. As shown, the model achieves + +the best performance with $\gamma = 0.6$ , striking an optimal balance between style reflection and content preservation. The choice of scaling ratio influences both SRA and FID, where the optimal value improves stylization without sacrificing content integrity. + +Text Expression for Multi-modal Alignment. In Table 3, we also explored different text representations for contrastive learning alignment. Specifically, we tested single labels (e.g., "Old"), brief descriptions (e.g., "An old person"), and more detailed descriptions (e.g., "An old person is moving slow and stiff"). Our experiments show that using a single label produced the best results, providing a clearer and more concise style signal for the model while avoiding unnecessary complexity in the text representation. + +# 5. Conclusion + +In this work, we introduce STYLEMOTIF, a novel stylized motion diffusion model capable of generating motion conditioned on both content and style from multiple modalities. Unlike prior approaches that either focus on motion generation across various content types or style transfer between sequences, STYLEMOTIF effectively synthesizes motion while incorporating stylistic cues from multi-modal inputs, including text, image, video, and audio. To achieve this, we introduce a style-content cross fusion mechanism and align a style encoder with a pre-trained multi-modal model. Through extensive experiments across diverse applications, we demonstrate that STYLEMOTIF outperforms existing methods in stylized motion generation, producing high-quality, realistic results that faithfully adhere to the given style references. Moreover, our model exhibits emergent capabilities for multi-modal motion stylization, enabling richer and more nuanced motion synthesis. These findings indicate the potential of STYLEMOTIF in advancing stylized motion generation and open new avenues for future research in multi-modal-driven motion synthesis and style-aware generative models. + +# Acknowledgements + +We would like to extend our sincere thanks to the Core AI and Avatar Movement team at Roblox for their guidance and support throughout this project. In particular, we are grateful to Haomiao Jiang, Vincent Petrella, Sean Palmer, and Simone Guggiari for their valuable feedback and insights. We also thank Lei Zhong and Zhe Li for their technical advice. + +# References + +[1] K. Aberman, Y. Weng, D. Lischinski, D. Cohen-Or, and B. Chen. Unpaired motion style transfer from video to animation. TOG, 2020. 2, 6 +[2] Nefeli Andreou, Xi Wang, Victoria Fernández Abrevaya, Marie-Paule Cani, Yiorgos Chrysanthou, and Vicky Kalogeiton. Lead: Latent realignment for human motion diffusion. arXiv preprint arXiv:2410.14508, 2024. 2 +[3] Z. Cen, H. Pi, S. Peng, Z. Shen, M. Yang, S. Zhu, H. Bao, and X. Zhou. Generating human motion in 3d scenes from text descriptions. In CVPR, 2024. 2 +[4] Kexin Chen, Yuyang Du, Tao You, Mobarakol Islam, Ziyu Guo, Yueming Jin, Guangyong Chen, and Pheng-Ann Heng. Llm-assisted multi-teacher continual learning for visual question answering in robotic surgery. ICRA 2024, 2024. 3 +[5] L.H. Chen, S. Lu, A. Zeng, H. Zhang, B. Wang, R. Zhang, and L. Zhang. Motionllm: Understanding human behaviors from human motions and videos. ArXiv, 2024. 2, 3 +[6] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In CVPR, 2023. 1, 2, 3, 4, 5 +[7] Sammy Christen, Shreyas Hampali, Fadime Sener, Edoardo Remelli, Tomas Hodan, Eric Sauser, Shugao Ma, and Bugra Tekin. Diffh2o: Diffusion-based synthesis of hand-object interactions from textual descriptions. In SIGGRAPH Asia 2024 Conference Papers, 2024. 2 +[8] Wenxun Dai, Ling-Hao Chen, Jingbo Wang, Jinpeng Liu, Bo Dai, and Yansong Tang. Motionlcm: Real-time controllable motion generation via latent consistency model. arXiv preprint arXiv:2404.19759, 2024. 2 +[9] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11162-11173, 2021. 3 +[10] Han Fang, Pengfei Xiong, Luhui Xu, and Yu Chen. Clip2video: Mastering video-text retrieval via image clip. arXiv preprint arXiv:2106.11097, 2021. 3 +[11] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xiangyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023. 3 +[12] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15180-15190, 2023. 2, 3, 4, 5 + +[13] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, 2022. 2, 4, 5, 6, 7, 8 +[14] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In CVPR, 2024. 2, 3 +[15] Chuan Guo, Yuxuan Mu, Xinxin Zuo, Peng Dai, Youliang Yan, Juwei Lu, and Li Cheng. Generative human motion stylization in latent space. arXiv preprint arXiv:2401.13505, 2024. 2 +[16] Ziyu Guo, Renrui Zhang, Longtian Qiu, Xianzheng Ma, Xu-peng Miao, Xuming He, and Bin Cui. Calip: Zero-shot enhancement of clip with parameter-free attention. AAAI 2023 Oral, 2022. 3 +[17] Ziyu Guo, Renrui Zhang, Xiangyang Zhu, Yiwen Tang, Xianzheng Ma, Jiaming Han, Kexin Chen, Peng Gao, Xianzhi Li, Hongsheng Li, et al. Point-bind & point-llm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. arXiv preprint arXiv:2309.00615, 2023. 3 +[18] Zilu Guo, Hongbin Lin, Zhihao Yuan, Chaoda Zheng, Pengshuo Qiu, Dongzhi Jiang, Renrui Zhang, Chun-Mei Feng, and Zhen Li. Pisa: A self-augmented data engine and training strategy for 3d understanding with large models. arXiv preprint arXiv:2503.10529, 2025. 3 +[19] Ziyu Guo, Ray Zhang, Hao Chen, Jialin Gao, Dongzhi Jiang, Jiaze Wang, and Pheng-Ann Heng. Sciverse: Unveiling the knowledge comprehension and visual reasoning of lmms on multi-modal scientific problems. arXiv preprint arXiv:2503.10627, 2025. 3 +[20] Ziyu Guo, Renrui Zhang, Chengzhuo Tong, Zhizheng Zhao, Peng Gao, Hongsheng Li, and Pheng-Ann Heng. Can we generate images with cot? let's verify and reinforce image generation step by step. arXiv preprint arXiv:2501.13926, 2025.3 +[21] Jiaming Han, Renrui Zhang, Wenqi Shao, Peng Gao, Peng Xu, Han Xiao, Kaipeng Zhang, Chris Liu, Song Wen, Ziyu Guo, et al. Imagebind-llm: Multi-modality instruction tuning. arXiv preprint arXiv:2309.03905, 2023. 3 +[22] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In CVPR, 2023. 2 +[23] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 4 +[24] Deok-Kyeong Jang, Soomin Park, and Sung-Hee Lee. Motion puzzle: Arbitrary motion style transfer by body part. ACM TOG, 2022. 2, 6 +[25] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. Advances in Neural Information Processing Systems, 36:20067-20079, 2023. 3, 5 +[26] Dongzhi Jiang, Renrui Zhang, Ziyu Guo, Yanwei Li, Yu Qi, Xinyan Chen, Liuhui Wang, Jianhan Jin, Claire Guo, Shen Yan, et al. Mme-cot: Benchmarking chain-of-thought in large + +multimodal models for reasoning quality, robustness, and efficiency. arXiv preprint arXiv:2502.09621, 2025.3 +[27] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Gmd: Controllable human motion synthesis via guided diffusion models. In ICCV, 2023. 1, 2 +[28] Boeun Kim, Jungho Kim, Hyung Jin Chang, and Jin Young Choi. Most: Motion style transformer between diverse action contents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1705-1714, 2024. 2 +[29] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer, 2024a. URL https://arxiv.org/abs/2408.03326. 3 +[30] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. ICLR 2025 Spotlight, 2024. 3 +[31] Xiaoqi Li, Mingxu Zhang, Yiran Geng, Haoran Geng, Yuxing Long, Yan Shen, Renrui Zhang, Jiaming Liu, and Hao Dong. Manipllm: Embodied multimodal large language model for object-centric robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18061-18070, 2024. 3 +[32] Yanghao Li, Naiyan Wang, Jiaying Liu, and Xiaodi Hou. Demystifying neural style transfer. arXiv preprint arXiv:1701.01036, 2017. 4 +[33] Zhe Li, Yisheng He, Lei Zhong, Weichao Shen, Qi Zuo, Lingteng Qiu, Zilong Dong, Laurence Tianruo Yang, and Weihao Yuan. Mulsmo: Multimodal stylized motion generation by bidirectional control flow. arXiv preprint arXiv:2412.09901, 2024. 2 +[34] Zhe Li, Weihao Yuan, Yisheng He, Lingteng Qiu, Shenhao Zhu, Xiaodong Gu, Weichao Shen, Yuan Dong, Zilong Dong, and Laurence T Yang. Lamp: Language-motion pretraining for motion generation, retrieval, and captioning. arXiv preprint arXiv:2410.07093, 2024. 2 +[35] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5 +[36] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In ICCV, 2019. 5 +[37] Ian Mason, Sebastian Starke, and Taku Komura. Real-time style modelling of human locomotion via feature-wise transformations and local motion phases. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 2022. 2, 4, 5, 6, 7, 8 +[38] Arsha Nagrani, Paul Hongsuck Seo, Bryan Seybold, Anja Hauth, Santiago Manen, Chen Sun, and Cordelia Schmid. Learning audio-video modalities from image captions. In European Conference on Computer Vision, pages 407-426. Springer, 2022. 3 +[39] Soomin Park, Deok-Kyeong Jang, and Sung-Hee Lee. *Diverse motion stylization for multiple style domains via spatial-temporal graph-based generative model.* Proceedings of the ACM on Computer Graphics and Interactive Techniques, 2021. 2 + +[40] Xiaogang Peng, Yiming Xie, Zizhao Wu, Varun Jampani, Deqing Sun, and Huaizu Jiang. Hoi-diff: Text-driven synthesis of 3d human-object interactions using diffusion models. arXiv preprint arXiv:2312.06553, 2023. 2 +[41] Mathis Petrovich, Or Litany, Umar Iqbal, Michael J Black, Gul Varol, Xue Bin Peng, and Davis Rempe. Multi-track timeline control for text-driven 3d human motion generation. In CVPR, 2024. 2 +[42] Ekkasit Pinyoanuntapong, Muhammad Usama Saleem, Pu Wang, Minwoo Lee, Srijan Das, and Chen Chen. Bamm: Bidirectional autoregressive motion model. arXiv preprint arXiv:2403.19435, 2024. +[43] Ekkasit Pinyoanuntapong, Pu Wang, Minwoo Lee, and Chen Chen. Mmm: Generative masked motion model. In CVPR, 2024. 2 +[44] Ziyun Qian, Zeyu Xiao, Zhenyi Wu, Dingkang Yang, Mingcheng Li, Shunli Wang, Shuaibing Wang, Dongliang Kou, and Lihua Zhang. Smcd: High realism motion style transfer via mamba-based diffusion. arXiv preprint arXiv:2405.02844, 2024. 2 +[45] Sigal Raab, Inbal Leibovitch, Guy Tevet, Moab Arar, Amit H Bermano, and Daniel Cohen-Or. Single motion diffusion. arXiv preprint arXiv:2302.05905, 2023. 2 +[46] Sigal Raab, Inbar Gat, Nathan Sala, Guy Tevet, Rotem Shalev-Arkushin, Ohad Fried, Amit H Bermano, and Daniel Cohen-Or. Monkey see, monkey do: Harnessing self-attention in motion diffusion for zero-shot motion transfer. arXiv preprint arXiv:2406.06508, 2024. 2 +[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. 2021. 3 +[48] Davis Rempe, Zhengyi Luo, Xue Bin Peng, Ye Yuan, Kris Kitani, Karsten Kreis, Sanja Fidler, and Or Litany. Trace and pace: Controllable pedestrian animation via guided trajectory diffusion. In CVPR, 2023. 2 +[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2 +[50] Wenfeng Song, Xingliang Jin, Shuai Li, Chenglizhao Chen, Aimin Hao, Xia Hou, Ning Li, and Hong Qin. Arbitrary motion style transfer with multi-condition motion latent diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 821-830, 2024. 2 +[51] Xiangjun Tang, Linjun Wu, He Wang, Bo Hu, Xu Gong, Yuchen Liao, Songnan Li, Qilong Kou, and Xiaogang Jin. Rsmt: Real-time stylized motion transition for characters. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 2 +[52] Yiwen Tang, Zoey Guo, Zhuhao Wang, Ray Zhang, Qizhi Chen, Junli Liu, Delin Qu, Zhigang Wang, Dong Wang, Xuelong Li, et al. Exploring the potential of encoder-free architectures in 3d Imms. arXiv preprint arXiv:2502.09620, 2025. 3 + +[53] Tianxin Tao, Xiaohang Zhan, Zhongquan Chen, and Michiel van de Panne. Style-erd: Responsive and coherent online motion style transfer. 2022. 2 +[54] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2023. 1, 2 +[55] Weilin Wan, Zhiyang Dou, Taku Komura, Wenping Wang, Dinesh Jayaraman, and Lingjie Liu. Tlcontrol: Trajectory and language control for human motion synthesis. arXiv preprint arXiv:2311.17135, 2023. 2 +[56] Yu-Hui Wen, Zhipeng Yang, Hongbo Fu, Lin Gao, Yanan Sun, and Yong-Jin Liu. Autoregressive stylized motion synthesis with generative flow. In CVPR, 2021. 2 +[57] Qianyang Wu, Ye Shi, Xiaoshui Huang, Jingyi Yu, Lan Xu, and Jingya Wang. Thor: Text to human-object interaction diffusion via relation intervention. arXiv preprint arXiv:2403.11208, 2024. 2 +[58] Qi Wu, Yubo Zhao, Yifan Wang, Yu-Wing Tai, and Chi-Keung Tang. Motionllm: Multimodal motion-language learning with large language models. arXiv preprint arXiv:2405.17013, 2024. 2 +[59] Yiming Xie, Varun Jampani, Lei Zhong, Deqing Sun, and Huaizu Jiang. Omnicontrol: Control any joint at any time for human motion generation. In ICLR, 2024. 2 +[60] Jingwei Xu, Huazhe Xu, Bingbing Ni, Xiaokang Yang, Xiaolong Wang, and Trevor Darrell. Hierarchical style-based networks for motion synthesis. In ECCV, 2020. 2 +[61] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. Interdiff: Generating 3d human-object interactions with physics-informed diffusion. In ICCV, 2023. 2 +[62] Sirui Xu, Ziyin Wang, Yu-Xiong Wang, and Liang-Yan Gui. Interdreamer: Zero-shot text to 3d dynamic human-object interaction. arXiv preprint arXiv:2403.19652, 2024. 2 +[63] Le Xue, Mingfei Gao, Chen Xing, Roberto Martin-Martin, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1179-1189, 2023. 5 +[64] Senqiao Yang, Jiaming Liu, Ray Zhang, Mingjie Pan, Ziyu Guo, Xiaqi Li, Zehui Chen, Peng Gao, Yandong Guo, and Shanghai Zhang. Lidar-llm: Exploring the potential of large language models for 3d lidar understanding. AAAI 2025, 2023. 3 +[65] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In ICCV, 2023. 2 +[66] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3836-3847, 2023. 2, 3, 4 +[67] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. PAMI, 2024. 2 +[68] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Point + +clip: Point cloud understanding by clip. In CVPR 2022, 2022. 3 +[69] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In ECCV 2022. Springer Nature Switzerland, 2022. 3 +[70] Renrui Zhang, Jiaming Han, Chris Liu, Aojun Zhou, Pan Lu, Yu Qiao, Hongsheng Li, and Peng Gao. Llama-adapter: Efficient fine-tuning of large language models with zero-initialized attention. In ICLR 2024, 2024. 3 +[71] Renrui Zhang, Dongzhi Jiang, Yichi Zhang, Haokun Lin, Ziyu Guo, Pengshuo Qiu, Aojun Zhou, Pan Lu, Kai-Wei Chang, Peng Gao, et al. Mathverse: Does your multi-modal llm truly see the diagrams in visual math problems? ECCV 2024, 2024. 3 +[72] Renrui Zhang, Xinyu Wei, Dongzhi Jiang, Ziyu Guo, Shicheng Li, Yichi Zhang, Chengzhuo Tong, Jiaming Liu, Aojun Zhou, Bin Wei, et al. Mavis: Mathematical visual instruction tuning with an automatic data engine. arXiv preprint arXiv:2407.08739, 2024. 3 +[73] Lei Zhong, Yiming Xie, Varun Jampani, Deqing Sun, and Huaizu Jiang. Smoodi: Stylized motion diffusion model. In ECCV, 2024. 2, 3, 5, 6 +[74] Bin Zhu, Bin Lin, Munan Ning, Yang Yan, Jiaxi Cui, HongFa Wang, Yatian Pang, Wenhao Jiang, Junwu Zhang, Zongwei Li, et al. Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. arXiv preprint arXiv:2310.01852, 2023. 3 +[75] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. ICCV 2023, 2023. 3 \ No newline at end of file diff --git a/2025/StyleMotif_ Multi-Modal Motion Stylization using Style-Content Cross Fusion/images.zip b/2025/StyleMotif_ Multi-Modal Motion Stylization using Style-Content Cross Fusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..30859bff2203ee6e1b94516bf35be7045baf7414 --- /dev/null +++ b/2025/StyleMotif_ Multi-Modal Motion Stylization using Style-Content Cross Fusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:666555ada25ca43bacedaee6e37272ab7320991c051e90f921212860c91a0c8f +size 567227 diff --git a/2025/StyleMotif_ Multi-Modal Motion Stylization using Style-Content Cross Fusion/layout.json b/2025/StyleMotif_ Multi-Modal Motion Stylization using Style-Content Cross Fusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7be2896934fdbcfb09a8d6285c9b15a19251801f --- /dev/null +++ b/2025/StyleMotif_ Multi-Modal Motion Stylization using Style-Content Cross Fusion/layout.json @@ -0,0 +1,8099 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 555, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 555, + 122 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 555, + 122 + ], + "type": "text", + "content": "STYLEMOtIF: Multi-Modal Motion Stylization using Style-Content Cross Fusion" + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 96, + 141, + 516, + 238 + ], + "blocks": [ + { + "bbox": [ + 96, + 141, + 516, + 238 + ], + "lines": [ + { + "bbox": [ + 96, + 141, + 516, + 238 + ], + "spans": [ + { + "bbox": [ + 96, + 141, + 516, + 238 + ], + "type": "table", + "html": "
Ziyu Guo†Young Yoon LeeJoseph Liu
The Chinese University of Hong KongRobloxRoblox
ziyuguo@link.cuhk.edu.hkylee@roblox.comjosephliu@roblox.com
Yizhak Ben-ShabatVictor ZordanMubbasir Kapadia
RobloxRobloxRoblox
ibenshabat@roblox.comvbzordan@roblox.commkapadia@roblox.com
", + "image_path": "2375eaaa2a709112399a1158ca8f1cc8f04570c312cf87b6077ac50dcd0ded49.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 264, + 200, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 264, + 200, + 277 + ], + "spans": [ + { + "bbox": [ + 152, + 264, + 200, + 277 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 290, + 298, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 290, + 298, + 494 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 298, + 494 + ], + "type": "text", + "content": "We present STYLEMOTIF, a novel Stylized Motion Latent Diffusion model, generating motion conditioned on both content and style from multiple modalities. Unlike existing approaches that either focus on generating diverse motion content or transferring style from sequences, STYLEMOTIF seamlessly synthesizes motion across a wide range of content while incorporating stylistic cues from multi-modal inputs, including motion, text, image, video, and audio. To achieve this, we introduce a style-content cross fusion mechanism and align a style encoder with a pre-trained multi-modal model, ensuring that the generated motion accurately captures the reference style while preserving realism. Extensive experiments demonstrate that our framework surpasses existing methods in stylized motion generation and exhibits emergent capabilities for multi-modal motion stylization, enabling more nuanced motion synthesis. Project Page: https://stylemotif.github.io." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 531, + 136, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 136, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 136, + 544 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 549, + 298, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 549, + 298, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 549, + 298, + 693 + ], + "type": "text", + "content": "Human motion generation is a fundamental task in computer graphics and animation, enabling the synthesis of realistic and expressive human movements. Broadly, human motion can be characterized by two complementary aspects: content, which defines the underlying action (e.g., walking, jumping), and style, which encodes variations such as personal flair, emotional expression, or cultural influences (e.g., jubilant, aggressive). This separation allows for greater control and flexibility in generating motion, making it particularly valuable in creative industries like game development, film production, and virtual reality. However, traditional approaches to stylized motion generation often depend on manual pro" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 252, + 553, + 420 + ], + "blocks": [ + { + "bbox": [ + 317, + 252, + 553, + 420 + ], + "lines": [ + { + "bbox": [ + 317, + 252, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 317, + 252, + 553, + 420 + ], + "type": "image", + "image_path": "a2c6ee5da755e8bdbcefd5d585a697bc8bbc4745f64d4ec6b582eee470f116b3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 429, + 556, + 485 + ], + "lines": [ + { + "bbox": [ + 313, + 429, + 556, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 556, + 485 + ], + "type": "text", + "content": "Figure 1. Comparison of Our Proposed STYLEMOTIF with SMooDi. SMooDi duplicates and separates two diffusion backbones for content and style, increasing complexity. In contrast, STYLEMOTIF adopts a single diffusion backbone with direct style injection, enabling efficient and realistic motion stylization." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 497, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 555, + 521 + ], + "type": "text", + "content": "cesses such as motion capture or keyframe animation, which are costly, time-consuming, and labor-intensive." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 521, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 521, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 312, + 521, + 556, + 689 + ], + "type": "text", + "content": "Recent progress in text-to-motion (T2M) diffusion frameworks [6, 27, 54] has greatly advanced the ability to translate natural-language prompts into realistic human motions. By leveraging powerful denoising diffusion models, these approaches capture intricate spatiotemporal dependencies in the data, enabling coherent sequences of human movements to be generated directly from brief textual descriptions. Despite their success in content fidelity and diversity, most current T2M diffusion methods concentrate primarily on what action is performed, while overlooking how it is performed, namely, its stylistic details. Simply appending a separate style-transfer module to text-driven motion diffusion pipelines can introduce additional complexity and risk of compounding errors in the final output." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "In parallel, motion style transfer has been actively studied to infuse stylistic cues from a reference motion (or style" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 188, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 188, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 188, + 712 + ], + "type": "text", + "content": "† Work done as an intern at Roblox." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13349" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": "data) into another sequence [1, 24, 37, 53]. Although many of these methods effectively disentangle content and style for small-scale tasks, the pipeline becomes cumbersome when a large variety of content motions need to be stylized. Moreover, they commonly assume that the input or target sequences are high-quality motion data. In scenarios where content motions are synthetically generated, or partially noisy, the transfer process can deteriorate, leading to undesirable motion artifacts or compromised style fidelity." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 182, + 297, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 182, + 297, + 385 + ], + "spans": [ + { + "bbox": [ + 56, + 182, + 297, + 385 + ], + "type": "text", + "content": "To address the need for simultaneously controlling both content and style, some recent works have merged style encoding with diffusion-based motion generation. Among these, the most recent and representative approach [73] augments a pre-trained latent diffusion model [6] with a style adaptor and classifier-based style guidance, achieving stylized motion from textual prompts and motion-style references. While effective, this method relies on additional training diffusion backbones, which shares structural similarities with ControlNet [66] as shown in Figure 1, which increases model complexity and training overhead. It is also constrained to motion as the primary style input. Concurrent work [33] proposes a bidirectional control flow mechanism to mitigate conflicts between style and content, extending style control to multiple modalities, but this approach also duplicates two separate diffusion backbones with substantial training overhead and limited applicability across modalities." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 387, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 296, + 590 + ], + "type": "text", + "content": "To this end, we propose STYLEMOTIF, a new framework for multi-modal motion stylization that unifies text-to-motion diffusion with style conditioning in a single-diffusion structure. Specifically, we leverage a pre-trained motion latent diffusion model (MLD) [6] to preserve strong content generation capabilities, and seamlessly integrate style features extracted from a dedicated encoder, which is aligned with a multi-modal foundation model [12]. In contrast to previous works, STYLEMOTIF avoids duplicating large portions of the network or relying on specialized style diffusion network. Instead, we introduce a style-content cross fusion mechanism, which injects stylistic cues into the diffusion process while maintaining motion realism. As a result, our STYLEMOTIF not only yields more robust stylized outputs but also supports diverse style signals, such as motion, text, images, audio, or video clips, via the alignment in multi-modal feature space. We summarize our main contributions as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 594, + 296, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 641 + ], + "type": "text", + "content": "- We present STYLEMOTIF, a stylized latent motion diffusion framework that unifies diverse motion content and multi-modal styles within a compact single-diffusion design." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 642, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 689 + ], + "type": "text", + "content": "- We propose a style-content cross fusion module that injects stylistic cues into the diffusion denoising process, achieving faithful stylization without compromising motion realism and ensuring efficiency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "- We achieve a unified multi-modal style feature space and unveil new emergent capabilities through multi-modal" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 72, + 553, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 107 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 107 + ], + "type": "text", + "content": "alignment, which accommodate various sources including motion, text, images, audio, and video, for flexible and versatile multi-modal style control." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 108, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 108, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 314, + 108, + 553, + 144 + ], + "type": "text", + "content": "- Extensive experiments demonstrate that STYLEMOTIF consistently outperforms existing methods regarding style expressiveness, content preservation, and efficiency." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 153, + 400, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 153, + 400, + 165 + ], + "spans": [ + { + "bbox": [ + 314, + 153, + 400, + 165 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 173, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 173, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 173, + 555, + 449 + ], + "type": "text", + "content": "Human Motion Generation. Recent progress in human motion generation [3, 5, 8, 13, 14, 41-43, 46, 55, 57, 58, 62] has been driven by transformer [14, 34] and diffusion models [2, 8, 49, 66], showing great potential in producing realistic and diverse motions. These approaches have shown great potential in producing realistic and diverse motion sequences. For example, Momask [14] improves motion generation using a residual VQ-VAE. Similarly, LaMP [34] introduces a motion-aware text encoder and a motion-to-text language model to enhance motion quality through text conditioning. Diffusion models, in particular, have become a key approach in motion generation [6, 22, 27, 48, 54, 61, 65, 67]. MDM [54] introduces a motion diffusion model that operates directly on raw motion data to capture the relationship between motions and input text conditions. MLD [6] improve efficiency by embedding the diffusion process in latent space, reducing computational cost. They also allow conditioning on specific constraints, such as predefined trajectories [27, 55, 59] or human-object interactions [7, 40, 57], enabling greater control and diversity. Our work leverages pretrained motion latent diffusion model [6] and multi-modal foundation models [12] to achieve stylized human motion generation while maintaining high motion quality." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 462, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 556, + 713 + ], + "type": "text", + "content": "Motion Stylization. Motion stylization [1, 24, 28, 37, 39, 44, 45, 50, 51, 53, 56, 60] involves transferring stylistic features from a reference motion to a source motion/textual prompt, enabling creative transformations while preserving the original motion content. Early methods, like those in motion style transfer [1, 24], typically separate motion content and style for recombination. For instance, Aberman et al. [1] used a generative adversarial network to decouple style from content without paired data, while Motion Puzzle [24] allows style control for individual body parts, and Guo et al. [15] utilized pretrained motion models for better style integration. Recent approaches, like SMooDi [73], generate stylized motion from text and style sequences using style guidance. However, these models face two main challenges: (1) inefficient architectures that duplicate diffusion backbones for content and style, and (2) limited to style motion for guidance. Another concurrent approach, MulSMo [33], attempts to address some of these by introducing a bidirectional control flow between style and content networks, reducing conflicts. Still, it remains restricted to a few modalities and duplicates diffusion networks. In this" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13350" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 53, + 556, + 289 + ], + "blocks": [ + { + "bbox": [ + 60, + 53, + 556, + 289 + ], + "lines": [ + { + "bbox": [ + 60, + 53, + 556, + 289 + ], + "spans": [ + { + "bbox": [ + 60, + 53, + 556, + 289 + ], + "type": "image", + "image_path": "e43eacda52e40f7d1fbd364185ec40d421eb98e999aea47c2f37b9f95de4d56e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 297, + 555, + 344 + ], + "lines": [ + { + "bbox": [ + 54, + 297, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 54, + 297, + 555, + 344 + ], + "type": "text", + "content": "Figure 2. Overall Pipeline of STYLEMOTIF, a diffusion framework for multi-modal motion stylization. Given a text prompt and a reference style from various modalities, our model extract style features and fuse them with content by style-content cross fusion. Through multi-modal alignment with contrastive learning, we enable seamless multi-modal conditioning and flexible stylization across motion, text, images, audio, and video." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 357, + 297, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 357, + 297, + 430 + ], + "spans": [ + { + "bbox": [ + 54, + 357, + 297, + 430 + ], + "type": "text", + "content": "work we eliminate redundant diffusion backbone, resulting in a simpler, more effective, and efficient approach to stylized motion generation. Our method also enables efficient multi-modal motion stylization, supporting text, image and multi-modal inputs while maintaining both high stylization quality and motion realism." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 486, + 298, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 486, + 298, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 486, + 298, + 714 + ], + "type": "text", + "content": "Multi-modality Learning. Recent advancements in multimodality learning [9, 10, 38] have revolutionized various domains by enabling joint understanding across diverse data types. Foundational models like CLIP [47] and its extension [16, 68, 69, 75] establish robust vision-language alignment, while ImageBind [12], Point-Bind [17], and Language-Bind [74] expand this paradigm to diverse modalities, demonstrating emergent cross-modal capabilities. The rise of multi-modal large language models [4, 19, 21, 30, 31, 64, 72] further enhances semantic understanding through unified text-visual processing, achieving superior performance across 2D images [11, 29, 70], 3D point clouds [17, 18, 52], and complex reasoning scenarios [20, 26, 71]. For human motion, some approaches [5, 14, 25] incorporate multiple modalities into human motion understanding and generation tasks. Our work introduces the first framework for multimodal guided motion stylization, achieving seamless style injection from diverse modalities through a single diffusion architecture with style-content cross fusion." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 356, + 556, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 556, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 556, + 384 + ], + "type": "text", + "content": "3. STYLEMOTIF: Multi-Modal Motion Stylization with Style-Content Cross Fusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 390, + 556, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 556, + 547 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 556, + 547 + ], + "type": "text", + "content": "We propose STYLEMOTIF, a novel framework for stylized motion synthesis that combines style-content cross fusion and multi-modal motion stylization, as illustrated in Figure 2. Our approach integrates a style-content cross fusion mechanism that allows for coherent feature blending, ensuring the generated motion accurately reflects the reference style while maintaining the content's realism (§3.2). This mechanism is complemented by a multi-modal motion stylization strategy, which leverages inputs from diverse modalities such as image, video, audio, and text to provide control over the stylization process (§3.3). This framework enables highly flexible and realistic stylized motion synthesis, which offers greater control and diversity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 552, + 382, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 382, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 382, + 564 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": "Preliminaries. Motion Latent Diffusion (MLD) [6] formulates a conditional latent diffusion model by training " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_t,t,c)" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " to denoise a sequence of latents " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\{z_t\\}_{t = 0}^T" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "z_{t}\\in \\mathbb{R}^{n\\times d}" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " represents the motion latent at timestep " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " , conditioned on the distribution " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "p(z_{t}|c)" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " . A conditional domain encoder " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\tau_{\\theta}(c)" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " , such as CLIP [47], enables text-to-motion tasks, and the model is trained as " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_t,t,\\tau_{\\theta}(c))" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " . To incorporate additional style conditioning, a style condition " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " can be introduced with its own encoder " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\psi_{\\theta}(s)" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " , which may take the form of a motion or text encoder. This extends the model to " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_t,t,\\tau_{\\theta}(c),\\psi_{\\theta_s}(s))" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " . SMooDi [73] adopts a ControlNet [66]-style approach for style conditioning by" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13351" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "creating a trainable copy of the neural network weights " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\theta_{s}" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " from the original MLD model " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\theta_{c}" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": ". This copy includes zero-initialized linear layers, denoted as " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}(\\cdot ;\\cdot)" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": ". The output " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^i (z_t,t,\\tau_\\theta (c),\\psi_\\theta (s);\\theta_c)" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "-th MLD block, now conditioned on style, is computed as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 141, + 295, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 141, + 295, + 171 + ], + "spans": [ + { + "bbox": [ + 66, + 141, + 295, + 171 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {F} ^ {i} \\left(z _ {t}, t, \\tau_ {\\theta} (c), \\psi_ {\\theta_ {s}} (s); \\theta_ {c}\\right) = \\mathcal {F} ^ {i} \\left(z _ {t}, t, \\tau_ {\\theta} (c); \\theta_ {c}\\right) \\\\ + \\mathcal {Z} \\left(\\mathcal {F} ^ {i} \\left(z _ {t}, t, \\tau_ {\\theta} (c), \\psi_ {\\theta_ {s}} (s); \\theta_ {s}\\right); \\theta_ {z _ {i}}\\right) \\tag {1} \\\\ \\end{array}", + "image_path": "e2fe02b643820b7ea8f9ce1037d91ca6809dfc267fc90279fc5d4090d0dd6bcb.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 180, + 296, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 296, + 240 + ], + "type": "text", + "content": "A key property of this formulation is that since " + }, + { + "bbox": [ + 55, + 180, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\theta_{z_i}" + }, + { + "bbox": [ + 55, + 180, + 296, + 240 + ], + "type": "text", + "content": " is initialized to zero, thus " + }, + { + "bbox": [ + 55, + 180, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^i (z_t,t,\\tau_\\theta (c),\\psi_\\theta (s)) = \\mathcal{F}^i (z_t,t,\\tau_\\theta (c))" + }, + { + "bbox": [ + 55, + 180, + 296, + 240 + ], + "type": "text", + "content": " at the beginning. However, the tradeoff of that method is that it needs to maintain the additional parameters " + }, + { + "bbox": [ + 55, + 180, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\theta_{s}" + }, + { + "bbox": [ + 55, + 180, + 296, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 180, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\theta_{z_i}" + }, + { + "bbox": [ + 55, + 180, + 296, + 240 + ], + "type": "text", + "content": " in order for style transfer." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 255, + 296, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 459 + ], + "type": "text", + "content": "STYLEMOTIF. We propose STYLEMOTIF, eschewing zero linear layers in favor of injecting it directly via statistical manipulation, with the pipeline shown in Figure 2 describing our approach. Our STYLEMOTIF utilizes latent space diffusion within a single generative diffusion backbone, building upon a pretrained MLD model [6]. Instead of perturbing the outputs of each block " + }, + { + "bbox": [ + 55, + 255, + 296, + 459 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 255, + 296, + 459 + ], + "type": "text", + "content": " via zero initialized layer " + }, + { + "bbox": [ + 55, + 255, + 296, + 459 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}(\\mathcal{F}^i (z_t,t,\\tau_\\theta (c),\\psi_\\theta (s);\\theta_s);\\theta_{z_i})" + }, + { + "bbox": [ + 55, + 255, + 296, + 459 + ], + "type": "text", + "content": ", STYLEMOTIF replaces this with a statistically transformed style embedding that is injected into the original MLD branch. This simplifies the model while ensuring high-quality stylization results, whose details are presented in § 3.2. Also, we follow SMooDi's generation guidance and training scheme to ensure high-quality stylized motion synthesis. We use a hybrid guidance strategy that balances content fidelity and style adherence by combining classifier-free and classifier-based techniques during diffusion sampling." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 466, + 206, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 466, + 206, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 206, + 478 + ], + "type": "text", + "content": "3.2. Style-Content Cross Fusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 483, + 296, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 674 + ], + "type": "text", + "content": "Style Encoder Pre-training. To establish a robust foundation for style-content fusion, we combine the content knowledge from the pre-trained MLD's [6] VAE with the style knowledge from the 100STYLE [37] dataset. The MLD's VAE is pre-trained on the HumanML3D [13] dataset, which provides extensive understanding of content motion. Building on this, we further fine-tune the model on the 100STYLE dataset in a variational autoencoding manner to initially align the content and style data distributions in the latent space. After training, we discard the decoder and retain only the encoder as the motion style encoder. This reconstruction task enables the encoder to learn robust motion feature representations, which are essential for supporting the subsequent stylization process. By integrating content and style knowledge in this way, we ensure a strong foundation for seamless style-content fusion." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Style-Content Cross Normalization. To train a stylized diffusion model " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_t, t, \\tau_{\\theta}(c), \\psi_{\\theta}(s))" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": ", we effectively fuse" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": "content and style features directly within the latent space diffusion process instead of duplicating two pre-trained diffusion backbones for content and style or separate control mechanisms [66]. Prior work [23, 32] has proven that feature statistics (mean, variance) effectively encode style. Given the output features " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^i" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": "-th block, we derive the content features " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_c^i = \\mathcal{F}^i(z_t, t, \\tau_\\theta(c); \\theta_c)" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": " from the input text " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": ", and the style features " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s = \\psi_{\\theta_s}(s)" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": " are extracted from the reference motion sequence " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": " using the pre-trained style encoder. To perform the fusion, we first compute the mean " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\mu_c" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": " and variance " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\sigma_c^2" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": " of the content features across the feature dimension:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 378, + 235, + 555, + 269 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 235, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 378, + 235, + 555, + 269 + ], + "type": "interline_equation", + "content": "\\mu_ {c} = \\frac {1}{D} \\sum_ {j = 1} ^ {D} \\mathcal {F} _ {c} ^ {i, j}, \\tag {2}", + "image_path": "e914bc6c913aa68ccadf844f2056523a01fba0781a83181a99ad3ae260ecfffe.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 378, + 272, + 555, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 272, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 378, + 272, + 555, + 304 + ], + "type": "interline_equation", + "content": "\\sigma_ {c} ^ {2} = \\frac {1}{D} \\sum_ {j = 1} ^ {D} \\left(\\mathcal {F} _ {c} ^ {i, j} - \\mu_ {c}\\right) ^ {2}, \\tag {3}", + "image_path": "d7e53d9965ed676651760976a23c2f4e5bc7e3dba76c6ba24aa8b5bd50863e86.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 314, + 555, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 555, + 362 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 555, + 362 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 314, + 555, + 362 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_c^{i,j}" + }, + { + "bbox": [ + 313, + 314, + 555, + 362 + ], + "type": "text", + "content": " denotes the content features of the " + }, + { + "bbox": [ + 313, + 314, + 555, + 362 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 314, + 555, + 362 + ], + "type": "text", + "content": "-th block and the " + }, + { + "bbox": [ + 313, + 314, + 555, + 362 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 314, + 555, + 362 + ], + "type": "text", + "content": "-th feature element. Next, we normalize the style features " + }, + { + "bbox": [ + 313, + 314, + 555, + 362 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s" + }, + { + "bbox": [ + 313, + 314, + 555, + 362 + ], + "type": "text", + "content": " using these content statistics, ensuring that the style features are adapted to the content's statistical properties:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 395, + 371, + 555, + 398 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 371, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 395, + 371, + 555, + 398 + ], + "type": "interline_equation", + "content": "\\widetilde {\\mathcal {F}} _ {s, c} = \\frac {\\mathcal {F} _ {s} - \\mu_ {c}}{\\sqrt {\\sigma_ {c} ^ {2} + \\eta}}, \\tag {4}", + "image_path": "e2101760524d18ae6d39620de96f218968bc88b3230aa0917f68b59d8d281483.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 406, + 556, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 556, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 556, + 477 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 406, + 556, + 477 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 406, + 556, + 477 + ], + "type": "text", + "content": " is a constant added for numerical stability. This style-content cross normalization ensures that the style features are smoothly integrated with the content features while maintaining the content's original structure. After that, we add the normalized style features to the content features, formulating our final cross-normalization as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 347, + 486, + 555, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 486, + 555, + 500 + ], + "spans": [ + { + "bbox": [ + 347, + 486, + 555, + 500 + ], + "type": "interline_equation", + "content": "\\mathcal {F} ^ {i} \\left(z _ {t}, t, \\tau_ {\\theta} (c), \\psi_ {\\theta_ {s}} (s); \\theta_ {c}\\right) = \\mathcal {F} _ {c} ^ {i} + \\gamma \\cdot \\tilde {\\mathcal {F}} _ {s, c}, \\tag {5}", + "image_path": "3705d908c242a39c62aaf5d21729c45ddc38e6d3d687c7924dfe446538bb5f01.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "text", + "content": " is a parameter used for scaling the normalized value. " + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{F}}_{s,c}" + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "text", + "content": " can be thought of as a perturbation of " + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_c^i" + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "text", + "content": " as it is scaled within its range. Notably, the fusion process is performed only once after the " + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "text", + "content": "-th block during the denoising process to avoid distorting the content while effectively introducing the style. This efficient fusion method eliminates the need for additional learnable parameters, as it is based solely on statistical transformations, achieving high-quality results with minimal computational overhead." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 624, + 487, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 487, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 487, + 636 + ], + "type": "text", + "content": "3.3. Multi-Modal Motion Stylization" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": "Our STYLEMOTIF extends to support multi-modal motion stylization. To achieve this, we integrate a pre-trained multimodal foundation model [12], which provides a unified, multi-modal aligned feature space, enabling effective cross-modal alignment between motion and other modalities. With this alignment, our model can flexibly combine multiple" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13352" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "content": "input modalities, such as text, image, audio, and video, to guide the stylization process in a comprehensive manner. This capability results in emergent abilities for multi-modal motion stylization, allowing for more nuanced outputs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 135, + 296, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 135, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 135, + 296, + 243 + ], + "type": "text", + "content": "Motion-Text Pair Curation. To align motion with other modalities, we process a set of motion-text pairs using the curated 100STYLE subset [37], which is carefully selected to avoid conflicts between content and style motions, so that the model can effectively learn the relationships between motion and text. Each motion sequence is paired with a corresponding single textual label, which serves as the text prompt. The curated motion-text pairs will be used for the following cross-model alignment." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 258, + 297, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 297, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 297, + 365 + ], + "type": "text", + "content": "Multi-Modal Alignment. To maintain the alignment within the multi-modal space, we freeze the text encoder of ImageBind [12] and introduce a lightweight projection layer after the encoder. This projection layer aligns the feature dimensions of the text and motion encoders. Since the multi-modal model represents each modality with a global feature, our alignment process focuses on these global features to achieve robust alignment between text and motion representations, formalized as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 375, + 296, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 375, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 138, + 375, + 296, + 388 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {t} = \\pi \\left(\\mathcal {E} _ {\\text {t e x t}} (l)\\right), \\tag {6}", + "image_path": "a49a81b804c07ab9a4fd89313e23f45fd4394d484275d5e20dad1461560e717c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 390, + 296, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 390, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 140, + 390, + 296, + 403 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {s} = \\psi_ {\\theta_ {s}} (s), \\tag {7}", + "image_path": "931223a307c079b986a43cc07a8496b014eb6d385e1402e1eda755c869b2edf4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 413, + 296, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 296, + 520 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 413, + 296, + 520 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 413, + 296, + 520 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 413, + 296, + 520 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 413, + 296, + 520 + ], + "type": "text", + "content": " are the paired input text label and style motion, " + }, + { + "bbox": [ + 55, + 413, + 296, + 520 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_t" + }, + { + "bbox": [ + 55, + 413, + 296, + 520 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 413, + 296, + 520 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s" + }, + { + "bbox": [ + 55, + 413, + 296, + 520 + ], + "type": "text", + "content": " denote the feature representations from the text and style motion encoders, " + }, + { + "bbox": [ + 55, + 413, + 296, + 520 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{text}}" + }, + { + "bbox": [ + 55, + 413, + 296, + 520 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 413, + 296, + 520 + ], + "type": "inline_equation", + "content": "\\psi_{\\theta_s}" + }, + { + "bbox": [ + 55, + 413, + 296, + 520 + ], + "type": "text", + "content": ", respectively, and " + }, + { + "bbox": [ + 55, + 413, + 296, + 520 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 55, + 413, + 296, + 520 + ], + "type": "text", + "content": " represents the projection operation that maps text features into the motion feature space. We employ a contrastive learning loss [63] to align the feature spaces of motion and text, bringing them closer together in the shared multi-modal space. Thus, we obtain a unified space between motion and all the modalities. We formulated it as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 527, + 296, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 527, + 296, + 576 + ], + "spans": [ + { + "bbox": [ + 56, + 527, + 296, + 576 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l i g n}} = - \\frac {1}{2} \\sum_ {(i, j)} \\log \\frac {\\exp \\frac {\\mathcal {F} _ {t} ^ {i} \\cdot \\mathcal {F} _ {s} ^ {j}}{\\tau_ {0}}}{\\sum_ {k} \\exp \\frac {\\mathcal {F} _ {t} ^ {i} \\cdot \\mathcal {F} _ {s} ^ {k}}{\\tau_ {0}}} + \\log \\frac {\\exp \\frac {\\mathcal {F} _ {t} ^ {i} \\cdot \\mathcal {F} _ {s} ^ {j}}{\\tau_ {0}}}{\\sum_ {k} \\exp \\frac {\\mathcal {F} _ {t} ^ {k} \\cdot \\mathcal {F} _ {s} ^ {j}}{\\tau_ {0}}}, \\tag {8}", + "image_path": "0586d3df351f81d69cdca83471404ed6fa063d3408c84fbf15cbc877a62fce2b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 586, + 297, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 297, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 297, + 657 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 586, + 297, + 657 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 586, + 297, + 657 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 586, + 297, + 657 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 586, + 297, + 657 + ], + "type": "text", + "content": " represent two modalities (text and style motion) and " + }, + { + "bbox": [ + 55, + 586, + 297, + 657 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 55, + 586, + 297, + 657 + ], + "type": "text", + "content": " indicates a positive pair in each training batch, " + }, + { + "bbox": [ + 55, + 586, + 297, + 657 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 586, + 297, + 657 + ], + "type": "text", + "content": " indexes all samples in the batch, including both positive and negative ones, and " + }, + { + "bbox": [ + 55, + 586, + 297, + 657 + ], + "type": "inline_equation", + "content": "\\tau_0" + }, + { + "bbox": [ + 55, + 586, + 297, + 657 + ], + "type": "text", + "content": " is a temperature parameter. During inference, we obtain the multi-modal (text, image, video, or audio) features from the multi-modal input " + }, + { + "bbox": [ + 55, + 586, + 297, + 657 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 586, + 297, + 657 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 667, + 296, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 667, + 296, + 681 + ], + "spans": [ + { + "bbox": [ + 132, + 667, + 296, + 681 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {m} = \\mathcal {E} _ {\\text {I m a g e B i n d}} (m), \\tag {9}", + "image_path": "5a49e0aa1a28ca0c43d39ece9ec494bfe5a818c5121c338d0016500979fbb12b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{ImageBind}}" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": " denotes ImageBind [12]. We use the multimodal features " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_m" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": " to retrieve the most semantically similar" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": "motion features from the unified multi-modal space. The retrieved motion features are then used to guide the stylization process, ensuring that fine-grained style details are preserved and accurately reflected in the final output. This approach enhances the model's ability to generate stylized motions that are both contextually relevant and visually coherent." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 156, + 391, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 391, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 391, + 168 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 175, + 554, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 554, + 199 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 554, + 199 + ], + "type": "text", + "content": "We provide extensive quantitative and qualitative analyses across multiple tasks in this section." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 207, + 441, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 207, + 441, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 207, + 441, + 220 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 225, + 555, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 225, + 555, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 225, + 555, + 320 + ], + "type": "text", + "content": "Implementation Details. We adopt the pre-trained MLD [6] as the foundation for motion generation. The style encoder of our model derives from the encoder of MLD's VAE, while the projection layer after text encoder is a single Linear layer. During diffusion training, we only enable the style encoder to be trainable while freezing other parameters. The model is optimized using the AdamW optimizer [35] with a constant learning rate of " + }, + { + "bbox": [ + 313, + 225, + 555, + 320 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 313, + 225, + 555, + 320 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 335, + 556, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 556, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 556, + 456 + ], + "type": "text", + "content": "Dataset Settings. We use the HumanML3D [13] dataset as our primary motion content dataset, which collects 14,616 motion sequences from AMASS [36] and annotates 44,970 sequence-level textual description. To train the style network, we utilize the 100STYLE [37] dataset, containing 45,303 style motions. We also adopt the text annotations for 100STYLE from previous work [73], which are pseudo text descriptions generated from MotionGPT [25]. We utilize the consistent root-velocity motion representations from HumanML3D for both the content and style data." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 471, + 556, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 556, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 556, + 567 + ], + "type": "text", + "content": "Baselines. We mainly evaluate our method against SMooDi [73] on motion-guided stylization and motion style transfer tasks. We also compare our text-guided stylcification with 'ChatGPT+MLD' approach, which utilize ChatGPT to combine style text and content text, and functions as a straightforward text-to-motion model without control capabilities [73]. Additionally, we qualitatively compare our model with SMooDi on motion-guided stylization." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 582, + 557, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 557, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 557, + 713 + ], + "type": "text", + "content": "Evaluation Metrics. To evaluate our model, we use several metrics following previous works [73]. We measure R-precision, Multi-modal Distance (MM Dist), Diversity, and Frechet Inception Distance (FID) to assess how well the content from text is preserved and how realistic the generated motions are. We use Style Recognition Accuracy (SRA) to evaluate how accurately the style of the reference motion is reflected in the generated motion. Also, we adopt the Foot Skate Ratio, which helps assess the realism of the generated motion, reducing artifacts like foot sliding. During evaluation, following previous works [73], we randomly select" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13353" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 64, + 355, + 167 + ], + "blocks": [ + { + "bbox": [ + 60, + 64, + 355, + 167 + ], + "lines": [ + { + "bbox": [ + 60, + 64, + 355, + 167 + ], + "spans": [ + { + "bbox": [ + 60, + 64, + 355, + 167 + ], + "type": "table", + "html": "
MethodSRA ↑FID ↓MM Dist ↓R-Precision ↑ (Top-3)Diversity →Foot Skate Ratio ↓
Motion-Guided Stylization
MLD + Aberman [1]54.373.3095.9830.4068.8160.347
MLD + Motion Puzzle [24]63.776.1276.4670.2906.4760.185
SMooDi [73]72.421.6094.4770.5719.2350.124
STYLEMOTIF (Ours)77.651.5514.3540.5867.5670.097
Text-Guided Stylization
MLD + ChatGPT [73]4.820.6144.3130.6058.8360.131
STYLEMOTIF (Ours)56.710.6033.6840.6909.1010.101
", + "image_path": "2df0e4fdcf7720c903310ca0a11e82d081ce87f75f4d1c449d6771fcdd07fb94.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 366, + 63, + 544, + 125 + ], + "blocks": [ + { + "bbox": [ + 58, + 175, + 357, + 209 + ], + "lines": [ + { + "bbox": [ + 58, + 175, + 357, + 209 + ], + "spans": [ + { + "bbox": [ + 58, + 175, + 357, + 209 + ], + "type": "text", + "content": "Table 1. Quantitative Results for Motion-Guided and Text-Guided Stylization. Bold values denote the best performance. As there is no ground-truth reference for Diversity, no value is highlighted in bold; but the metric is provided for reference." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 366, + 63, + 544, + 125 + ], + "lines": [ + { + "bbox": [ + 366, + 63, + 544, + 125 + ], + "spans": [ + { + "bbox": [ + 366, + 63, + 544, + 125 + ], + "type": "table", + "html": "
MethodSRA ↑FID ↓Foot Skate Ratio ↓
MLD + Aberman [1]61.013.8920.338
MLD + Motion Puzzle [24]67.236.8710.197
SMooDi [73]65.151.5820.095
STYLEMOTIF (Ours)68.811.3750.094
", + "image_path": "6a35f697f0454caa9794f9e526c24280359fa8fc7f3efff1bf3899a2418650a1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 360, + 132, + 549, + 209 + ], + "lines": [ + { + "bbox": [ + 360, + 132, + 549, + 209 + ], + "spans": [ + { + "bbox": [ + 360, + 132, + 549, + 209 + ], + "type": "text", + "content": "Table 2. Quantitative Results of Motion Style Transfer on HumanML3D [13] dataset. Our method outperforms previous works in all metrics, which demonstrates effective style-content fusion for high-quality motion style transfer, providing significant advantages for downstream tasks besides motion stylization." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 58, + 217, + 555, + 482 + ], + "blocks": [ + { + "bbox": [ + 58, + 217, + 555, + 482 + ], + "lines": [ + { + "bbox": [ + 58, + 217, + 555, + 482 + ], + "spans": [ + { + "bbox": [ + 58, + 217, + 555, + 482 + ], + "type": "image", + "image_path": "20b2c436f31ed23d4ea56d7967d06c6d7253e3f3a8fa14e2bf3162f090148f20.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 494, + 555, + 529 + ], + "lines": [ + { + "bbox": [ + 54, + 494, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 54, + 494, + 555, + 529 + ], + "type": "text", + "content": "Figure 3. Qualitative Results of Motion-Guided Stylization. Our model generates cohesive and realistic motions that effectively align style and content, such as preserving the 'circular' trajectory (first column) and 'hop' content (third column). In contrast, SMooDi [73] struggles to maintain content fidelity and sometimes fails to reflect the specified style (e.g., 'phone on the left' in the second column)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 542, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 295, + 579 + ], + "type": "text", + "content": "a content text from HumanML3D [13] and a style motion from 100STYLE [37], and compute SRA for the generated motion using a pre-trained classifier [73]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 587, + 201, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 201, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 201, + 601 + ], + "type": "text", + "content": "4.2. Motion-guided Stylization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 605, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 297, + 715 + ], + "type": "text", + "content": "Quantitative Analysis. In Table 1, we report the quantitative results for motion-guided stylization, where the motion serves as style input while text prompt as content one. Our method outperforms three baseline approaches [1, 24, 73] in all metrics. Specifically, we achieve a " + }, + { + "bbox": [ + 54, + 605, + 297, + 715 + ], + "type": "inline_equation", + "content": "5.23\\%" + }, + { + "bbox": [ + 54, + 605, + 297, + 715 + ], + "type": "text", + "content": " improvement in SRA while maintaining competitive performance in FID compared to the best baseline [73]. It demonstrates that our method, by effectively utilizing style-content cross fusion, generates motions that better align with the style reference" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 542, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 555, + 567 + ], + "type": "text", + "content": "while maintaining content integrity, showcasing the strength of our design in balancing style and content preservation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "content": "Qualitative Analysis. In Figure 3, we show some qualitative results of motion-guided stylization from our model and baseline [73]. Compared with the baseline, our model produces more cohesive stylized motions, with better alignment of style and content. For instance, in the provided examples, SMooDi fails to maintain the 'circular' trajectory or 'hop' action specified by the content or cannot accurately reflect the intended style, 'phone on the left'. This indicates that our approach more effectively integrates style and content, resulting in more realistic and consistent stylized motions." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13354" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 62, + 563, + 196 + ], + "blocks": [ + { + "bbox": [ + 58, + 62, + 563, + 196 + ], + "lines": [ + { + "bbox": [ + 58, + 62, + 563, + 196 + ], + "spans": [ + { + "bbox": [ + 58, + 62, + 563, + 196 + ], + "type": "image", + "image_path": "a54fce626f941968a92bdb3eb80e90a89f5fe555534ab4bbc9641437e19dfd07.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 204, + 312, + 237 + ], + "lines": [ + { + "bbox": [ + 55, + 204, + 312, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 312, + 237 + ], + "type": "text", + "content": "Figure 4. Qualitative Results of Text-Guided Stylization. Our model seamlessly integrates textual style descriptions with content, producing visually coherent and stylistically consistent results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 314, + 204, + 563, + 238 + ], + "lines": [ + { + "bbox": [ + 314, + 204, + 563, + 238 + ], + "spans": [ + { + "bbox": [ + 314, + 204, + 563, + 238 + ], + "type": "text", + "content": "Figure 5. Qualitative Results of Style Interpolation. Our model blends multiple style inputs while preserving content integrity, demonstrating effective style-content fusion of our model." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 247, + 555, + 415 + ], + "blocks": [ + { + "bbox": [ + 58, + 247, + 555, + 415 + ], + "lines": [ + { + "bbox": [ + 58, + 247, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 58, + 247, + 555, + 415 + ], + "type": "image", + "image_path": "e9bc3135ae16e288144a830654b13ddd14b9d2787f6006ea4a9e418075dfdc81.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 423, + 555, + 448 + ], + "lines": [ + { + "bbox": [ + 54, + 423, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 54, + 423, + 555, + 448 + ], + "type": "text", + "content": "Figure 6. Qualitative Results of Multi-Modal Motion Stylization. Our model generates stylized motions guided by diverse modalities (e.g., text, image, video, audio), effectively transferring style while preserving content integrity." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 460, + 187, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 187, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 187, + 474 + ], + "type": "text", + "content": "4.3. Text-guided Stylization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 478, + 297, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 478, + 297, + 610 + ], + "spans": [ + { + "bbox": [ + 54, + 478, + 297, + 610 + ], + "type": "text", + "content": "Quantitative Analysis. For the text-guided stylization task, we use the HumanML3D [13] dataset for motion content and employ the single text label [37] as the style control. As reported in Table 1 (Bottom), our method significantly outperforms the baseline, achieving " + }, + { + "bbox": [ + 54, + 478, + 297, + 610 + ], + "type": "inline_equation", + "content": "56.71\\%" + }, + { + "bbox": [ + 54, + 478, + 297, + 610 + ], + "type": "text", + "content": " in SRA, compared to " + }, + { + "bbox": [ + 54, + 478, + 297, + 610 + ], + "type": "inline_equation", + "content": "4.82\\%" + }, + { + "bbox": [ + 54, + 478, + 297, + 610 + ], + "type": "text", + "content": " for 'ChatGPT + MLD'. Moreover, our method maintains competitive FID, balancing style reflection with content preservation. This indicates that our design, which combines multi-modal alignment, helps utilize the style signal in the motion-text shared spaces and achieve significant effectiveness in text-guided stylization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 618, + 297, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 618, + 297, + 667 + ], + "spans": [ + { + "bbox": [ + 54, + 618, + 297, + 667 + ], + "type": "text", + "content": "Qualitative Analysis. In Figure 4, we showcase qualitative results for text-guided stylization, where our model also demonstrates strong capability in harmonizing style and content, generating high-quality and visually coherent results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 672, + 182, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 182, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 182, + 685 + ], + "type": "text", + "content": "4.4. Motion Style Transfer" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "For the motion style transfer task, we use the HumanML3D [13] dataset for motion content and the" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 461, + 556, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 556, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 556, + 606 + ], + "type": "text", + "content": "100STYLE dataset [37] for motion styles. As shown in Table 2, our method outperforms the baseline models across all key metrics. Specifically, we achieve a " + }, + { + "bbox": [ + 313, + 461, + 556, + 606 + ], + "type": "inline_equation", + "content": "3.66\\%" + }, + { + "bbox": [ + 313, + 461, + 556, + 606 + ], + "type": "text", + "content": " improvement in SRA, indicating better style reflection, and a 0.207 reduction in FID reflecting improved realism. These demonstrate that our framework not only improves the alignment between style and content but also provides significant advantages for downstream tasks like motion style transfer with efficient style-content cross fusion. By maintaining content integrity while seamlessly integrating style, STYLEMOTIF ensures more realistic and dynamically consistent stylized motions, which is critical for high-quality motion style transfer." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 612, + 488, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 488, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 488, + 625 + ], + "type": "text", + "content": "4.5. Multi-Modal Motion Stylization" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "With the power of the aligned multi-modal space, our model supports stylization guided by a variety of modalities, including motion, text, image, video, audio. Here, we utilize the input style feature to retrieve the corresponding motion features as the style condition, while the content input is provided as text prompt. The model then generates a stylized motion that incorporates the style from the input modality" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13355" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 59, + 194, + 153 + ], + "blocks": [ + { + "bbox": [ + 62, + 59, + 194, + 153 + ], + "lines": [ + { + "bbox": [ + 62, + 59, + 194, + 153 + ], + "spans": [ + { + "bbox": [ + 62, + 59, + 194, + 153 + ], + "type": "image", + "image_path": "ef6e357332c4ece566d7503d051fb462663a343173f66097cd6cbb344ff33fac.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 161, + 196, + 217 + ], + "lines": [ + { + "bbox": [ + 61, + 161, + 196, + 217 + ], + "spans": [ + { + "bbox": [ + 61, + 161, + 196, + 217 + ], + "type": "text", + "content": "Figure 7. Ablation Study on Scaling Ratio " + }, + { + "bbox": [ + 61, + 161, + 196, + 217 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 61, + 161, + 196, + 217 + ], + "type": "text", + "content": " in Eq. 5. We report both SRA and FID to show the impact of the scaling ratio on both stylization and content preservation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 211, + 57, + 545, + 178 + ], + "blocks": [ + { + "bbox": [ + 211, + 57, + 545, + 178 + ], + "lines": [ + { + "bbox": [ + 211, + 57, + 545, + 178 + ], + "spans": [ + { + "bbox": [ + 211, + 57, + 545, + 178 + ], + "type": "table", + "html": "
Style Encoder Pre-training Strategy
HumanML3D100STYLESRA ↑FID ↓MM Dist ↓R-Precision ↑Diversity →Foot Skate Ratio ↓
-✓76.731.7884.3490.5717.5050.101
✓-76.581.6354.4580.5728.5340.109
✓✓77.651.5514.3540.5867.5670.097
Text in Multi-modal Alignment
Textual ExpressionSRA ↑FID ↓MM Dist ↓R-Precision ↑Diversity →Foot Skate Ratio ↓
Brief Description76.841.5804.3780.5787.2510.099
Detailed Description75.251.6224.4190.5637.7640.102
Single Text Label77.651.5514.3540.5867.5670.097
", + "image_path": "163f0ce532f297f2de08cd6880d54fc409f279a3ab7e7a5abe9eed4611c3cf63.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 209, + 186, + 545, + 220 + ], + "lines": [ + { + "bbox": [ + 209, + 186, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 209, + 186, + 545, + 220 + ], + "type": "text", + "content": "Table 3. Ablation Study on Style Encoder Pre-training Strategies and Text Expression for Multi-modal Alignment. 'w. HumanML3D' and 'w. 100STYLE' denote pre-training with HumanML3D [13] and 100STYLE [37] data respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 234, + 297, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 234, + 297, + 344 + ], + "spans": [ + { + "bbox": [ + 54, + 234, + 297, + 344 + ], + "type": "text", + "content": "while maintaining the content integrity as specified by the text prompt. We showcase several examples of multi-modal motion stylization in Figure 6, where different modalities guide the motion generation. For instance, when a text content \"A person is walking.\" is provided alongside an image of a duck as style input, the model retrieves a relevant motion feature and blends the content with the style of 'Duckfoot'. As shown, the style from various inputs (text, image, video, audio) is effectively transferred to the generated motion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 351, + 167, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 167, + 364 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 167, + 364 + ], + "type": "text", + "content": "4.6. Style Interpolation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 369, + 296, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 369, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 54, + 369, + 296, + 514 + ], + "type": "text", + "content": "Leveraging the aligned multi-modal space, our model enables text-guided style interpolation. Given one content text along with at least two style style texts, our model generates a motion that combines the characteristics of all input styles. When style texts are provided, we retrieve the most similar style motion features from the shared space. These features are then combined by weighted summation. The combined style features are fused with the content features using the proposed style-content cross fusion. In Figure 5, we showcase some qualitative results of style interpolation in, where the generated motions successfully blend the characteristics of both styles while maintaining the content's integrity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 521, + 149, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 149, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 149, + 534 + ], + "type": "text", + "content": "4.7. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 540, + 297, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 540, + 297, + 661 + ], + "spans": [ + { + "bbox": [ + 54, + 540, + 297, + 661 + ], + "type": "text", + "content": "Style Encoder Pre-training. In Table 3, we investigate the impact of different pre-training strategies for the style encoder. Our results show that pre-training on both HumanML3D [13] and 100STYLE [37] yields the best performance. Compared to pre-training solely on either one, the combined training on both datasets provides the style encoder with a richer set of prior knowledge. This enables the model to effectively balance content preservation and style reflection, leveraging the diverse characteristics of both datasets to enhance the overall stylization quality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 677, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 677, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 677, + 297, + 715 + ], + "type": "text", + "content": "Cross Normalization Scaling Ratio. In Figure 7, we examine the effect of different scaling ratios " + }, + { + "bbox": [ + 54, + 677, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 54, + 677, + 297, + 715 + ], + "type": "text", + "content": " in Eq. 5 on the stylization performance. As shown, the model achieves" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 234, + 556, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 556, + 295 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 556, + 295 + ], + "type": "text", + "content": "the best performance with " + }, + { + "bbox": [ + 313, + 234, + 556, + 295 + ], + "type": "inline_equation", + "content": "\\gamma = 0.6" + }, + { + "bbox": [ + 313, + 234, + 556, + 295 + ], + "type": "text", + "content": ", striking an optimal balance between style reflection and content preservation. The choice of scaling ratio influences both SRA and FID, where the optimal value improves stylization without sacrificing content integrity." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 317, + 556, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 317, + 556, + 426 + ], + "spans": [ + { + "bbox": [ + 312, + 317, + 556, + 426 + ], + "type": "text", + "content": "Text Expression for Multi-modal Alignment. In Table 3, we also explored different text representations for contrastive learning alignment. Specifically, we tested single labels (e.g., \"Old\"), brief descriptions (e.g., \"An old person\"), and more detailed descriptions (e.g., \"An old person is moving slow and stiff\"). Our experiments show that using a single label produced the best results, providing a clearer and more concise style signal for the model while avoiding unnecessary complexity in the text representation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 441, + 388, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 388, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 388, + 453 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 462, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 556, + 715 + ], + "type": "text", + "content": "In this work, we introduce STYLEMOTIF, a novel stylized motion diffusion model capable of generating motion conditioned on both content and style from multiple modalities. Unlike prior approaches that either focus on motion generation across various content types or style transfer between sequences, STYLEMOTIF effectively synthesizes motion while incorporating stylistic cues from multi-modal inputs, including text, image, video, and audio. To achieve this, we introduce a style-content cross fusion mechanism and align a style encoder with a pre-trained multi-modal model. Through extensive experiments across diverse applications, we demonstrate that STYLEMOTIF outperforms existing methods in stylized motion generation, producing high-quality, realistic results that faithfully adhere to the given style references. Moreover, our model exhibits emergent capabilities for multi-modal motion stylization, enabling richer and more nuanced motion synthesis. These findings indicate the potential of STYLEMOTIF in advancing stylized motion generation and open new avenues for future research in multi-modal-driven motion synthesis and style-aware generative models." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13356" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 92, + 297, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 92, + 297, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 92, + 297, + 163 + ], + "type": "text", + "content": "We would like to extend our sincere thanks to the Core AI and Avatar Movement team at Roblox for their guidance and support throughout this project. In particular, we are grateful to Haomiao Jiang, Vincent Petrella, Sean Palmer, and Simone Guggiari for their valuable feedback and insights. We also thank Lei Zhong and Zhe Li for their technical advice." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 178, + 115, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 178, + 115, + 190 + ], + "spans": [ + { + "bbox": [ + 56, + 178, + 115, + 190 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 198, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 198, + 297, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 198, + 297, + 230 + ], + "spans": [ + { + "bbox": [ + 61, + 198, + 297, + 230 + ], + "type": "text", + "content": "[1] K. Aberman, Y. Weng, D. Lischinski, D. Cohen-Or, and B. Chen. Unpaired motion style transfer from video to animation. TOG, 2020. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 232, + 296, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 232, + 296, + 276 + ], + "spans": [ + { + "bbox": [ + 61, + 232, + 296, + 276 + ], + "type": "text", + "content": "[2] Nefeli Andreou, Xi Wang, Victoria Fernández Abrevaya, Marie-Paule Cani, Yiorgos Chrysanthou, and Vicky Kalogeiton. Lead: Latent realignment for human motion diffusion. arXiv preprint arXiv:2410.14508, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 278, + 296, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 278, + 296, + 311 + ], + "spans": [ + { + "bbox": [ + 62, + 278, + 296, + 311 + ], + "type": "text", + "content": "[3] Z. Cen, H. Pi, S. Peng, Z. Shen, M. Yang, S. Zhu, H. Bao, and X. Zhou. Generating human motion in 3d scenes from text descriptions. In CVPR, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 313, + 296, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 313, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 313, + 296, + 357 + ], + "type": "text", + "content": "[4] Kexin Chen, Yuyang Du, Tao You, Mobarakol Islam, Ziyu Guo, Yueming Jin, Guangyong Chen, and Pheng-Ann Heng. Llm-assisted multi-teacher continual learning for visual question answering in robotic surgery. ICRA 2024, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 359, + 296, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 359, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 62, + 359, + 296, + 392 + ], + "type": "text", + "content": "[5] L.H. Chen, S. Lu, A. Zeng, H. Zhang, B. Wang, R. Zhang, and L. Zhang. Motionllm: Understanding human behaviors from human motions and videos. ArXiv, 2024. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 394, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 394, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 62, + 394, + 294, + 426 + ], + "type": "text", + "content": "[6] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In CVPR, 2023. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 429, + 296, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 429, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 62, + 429, + 296, + 483 + ], + "type": "text", + "content": "[7] Sammy Christen, Shreyas Hampali, Fadime Sener, Edoardo Remelli, Tomas Hodan, Eric Sauser, Shugao Ma, and Bugra Tekin. Diffh2o: Diffusion-based synthesis of hand-object interactions from textual descriptions. In SIGGRAPH Asia 2024 Conference Papers, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 486, + 296, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 486, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 62, + 486, + 296, + 529 + ], + "type": "text", + "content": "[8] Wenxun Dai, Ling-Hao Chen, Jingbo Wang, Jinpeng Liu, Bo Dai, and Yansong Tang. Motionlcm: Real-time controllable motion generation via latent consistency model. arXiv preprint arXiv:2404.19759, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 532, + 296, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 532, + 296, + 575 + ], + "spans": [ + { + "bbox": [ + 62, + 532, + 296, + 575 + ], + "type": "text", + "content": "[9] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11162-11173, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 578, + 296, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 296, + 610 + ], + "type": "text", + "content": "[10] Han Fang, Pengfei Xiong, Luhui Xu, and Yu Chen. Clip2video: Mastering video-text retrieval via image clip. arXiv preprint arXiv:2106.11097, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 613, + 296, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 296, + 656 + ], + "type": "text", + "content": "[11] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xiangyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "text", + "content": "[12] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15180-15190, 2023. 2, 3, 4, 5" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "type": "text", + "content": "[13] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, 2022. 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 108, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 555, + 139 + ], + "type": "text", + "content": "[14] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In CVPR, 2024. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 141, + 555, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 555, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 555, + 184 + ], + "type": "text", + "content": "[15] Chuan Guo, Yuxuan Mu, Xinxin Zuo, Peng Dai, Youliang Yan, Juwei Lu, and Li Cheng. Generative human motion stylization in latent space. arXiv preprint arXiv:2401.13505, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 186, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 555, + 228 + ], + "type": "text", + "content": "[16] Ziyu Guo, Renrui Zhang, Longtian Qiu, Xianzheng Ma, Xu-peng Miao, Xuming He, and Bin Cui. Calip: Zero-shot enhancement of clip with parameter-free attention. AAAI 2023 Oral, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 232, + 555, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 555, + 296 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 555, + 296 + ], + "type": "text", + "content": "[17] Ziyu Guo, Renrui Zhang, Xiangyang Zhu, Yiwen Tang, Xianzheng Ma, Jiaming Han, Kexin Chen, Peng Gao, Xianzhi Li, Hongsheng Li, et al. Point-bind & point-llm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. arXiv preprint arXiv:2309.00615, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 298, + 555, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 298, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 317, + 298, + 555, + 353 + ], + "type": "text", + "content": "[18] Zilu Guo, Hongbin Lin, Zhihao Yuan, Chaoda Zheng, Pengshuo Qiu, Dongzhi Jiang, Renrui Zhang, Chun-Mei Feng, and Zhen Li. Pisa: A self-augmented data engine and training strategy for 3d understanding with large models. arXiv preprint arXiv:2503.10529, 2025. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 354, + 555, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 555, + 408 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 555, + 408 + ], + "type": "text", + "content": "[19] Ziyu Guo, Ray Zhang, Hao Chen, Jialin Gao, Dongzhi Jiang, Jiaze Wang, and Pheng-Ann Heng. Sciverse: Unveiling the knowledge comprehension and visual reasoning of lmms on multi-modal scientific problems. arXiv preprint arXiv:2503.10627, 2025. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 411, + 555, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 411, + 555, + 463 + ], + "spans": [ + { + "bbox": [ + 317, + 411, + 555, + 463 + ], + "type": "text", + "content": "[20] Ziyu Guo, Renrui Zhang, Chengzhuo Tong, Zhizheng Zhao, Peng Gao, Hongsheng Li, and Pheng-Ann Heng. Can we generate images with cot? let's verify and reinforce image generation step by step. arXiv preprint arXiv:2501.13926, 2025.3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 466, + 555, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 466, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 466, + 555, + 510 + ], + "type": "text", + "content": "[21] Jiaming Han, Renrui Zhang, Wenqi Shao, Peng Gao, Peng Xu, Han Xiao, Kaipeng Zhang, Chris Liu, Song Wen, Ziyu Guo, et al. Imagebind-llm: Multi-modality instruction tuning. arXiv preprint arXiv:2309.03905, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 512, + 555, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 512, + 555, + 554 + ], + "spans": [ + { + "bbox": [ + 317, + 512, + 555, + 554 + ], + "type": "text", + "content": "[22] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In CVPR, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 556, + 555, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 556, + 555, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 556, + 555, + 600 + ], + "type": "text", + "content": "[23] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 601, + 554, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 601, + 554, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 601, + 554, + 633 + ], + "type": "text", + "content": "[24] Deok-Kyeong Jang, Soomin Park, and Sung-Hee Lee. Motion puzzle: Arbitrary motion style transfer by body part. ACM TOG, 2022. 2, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 635, + 555, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 555, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 555, + 678 + ], + "type": "text", + "content": "[25] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. Advances in Neural Information Processing Systems, 36:20067-20079, 2023. 3, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 554, + 713 + ], + "type": "text", + "content": "[26] Dongzhi Jiang, Renrui Zhang, Ziyu Guo, Yanwei Li, Yu Qi, Xinyan Chen, Liuhui Wang, Jianhan Jin, Claire Guo, Shen Yan, et al. Mme-cot: Benchmarking chain-of-thought in large" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13357" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 72, + 295, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 295, + 95 + ], + "type": "text", + "content": "multimodal models for reasoning quality, robustness, and efficiency. arXiv preprint arXiv:2502.09621, 2025.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 296, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 296, + 129 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 296, + 129 + ], + "type": "text", + "content": "[27] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Gmd: Controllable human motion synthesis via guided diffusion models. In ICCV, 2023. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 295, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 295, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 295, + 184 + ], + "type": "text", + "content": "[28] Boeun Kim, Jungho Kim, Hyung Jin Chang, and Jin Young Choi. Most: Motion style transformer between diverse action contents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1705-1714, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 296, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 296, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 296, + 229 + ], + "type": "text", + "content": "[29] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer, 2024a. URL https://arxiv.org/abs/2408.03326. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 296, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 296, + 275 + ], + "type": "text", + "content": "[30] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. ICLR 2025 Spotlight, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 277, + 296, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 277, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 277, + 296, + 342 + ], + "type": "text", + "content": "[31] Xiaoqi Li, Mingxu Zhang, Yiran Geng, Haoran Geng, Yuxing Long, Yan Shen, Renrui Zhang, Jiaming Liu, and Hao Dong. Manipllm: Embodied multimodal large language model for object-centric robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18061-18070, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 295, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 295, + 376 + ], + "type": "text", + "content": "[32] Yanghao Li, Naiyan Wang, Jiaying Liu, and Xiaodi Hou. Demystifying neural style transfer. arXiv preprint arXiv:1701.01036, 2017. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 377, + 295, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 295, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 295, + 432 + ], + "type": "text", + "content": "[33] Zhe Li, Yisheng He, Lei Zhong, Weichao Shen, Qi Zuo, Lingteng Qiu, Zilong Dong, Laurence Tianruo Yang, and Weihao Yuan. Mulsmo: Multimodal stylized motion generation by bidirectional control flow. arXiv preprint arXiv:2412.09901, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 295, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 295, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 295, + 488 + ], + "type": "text", + "content": "[34] Zhe Li, Weihao Yuan, Yisheng He, Lingteng Qiu, Shenhao Zhu, Xiaodong Gu, Weichao Shen, Yuan Dong, Zilong Dong, and Laurence T Yang. Lamp: Language-motion pretraining for motion generation, retrieval, and captioning. arXiv preprint arXiv:2410.07093, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 490, + 295, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 295, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 295, + 511 + ], + "type": "text", + "content": "[35] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 513, + 295, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 513, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 56, + 513, + 295, + 545 + ], + "type": "text", + "content": "[36] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In ICCV, 2019. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 546, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 546, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 295, + 601 + ], + "type": "text", + "content": "[37] Ian Mason, Sebastian Starke, and Taku Komura. Real-time style modelling of human locomotion via feature-wise transformations and local motion phases. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 2022. 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 602, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 295, + 657 + ], + "type": "text", + "content": "[38] Arsha Nagrani, Paul Hongsuck Seo, Bryan Seybold, Anja Hauth, Santiago Manen, Chen Sun, and Cordelia Schmid. Learning audio-video modalities from image captions. In European Conference on Computer Vision, pages 407-426. Springer, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "text", + "content": "[39] Soomin Park, Deok-Kyeong Jang, and Sung-Hee Lee. *Diverse motion stylization for multiple style domains via spatial-temporal graph-based generative model.* Proceedings of the ACM on Computer Graphics and Interactive Techniques, 2021. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "text", + "content": "[40] Xiaogang Peng, Yiming Xie, Zizhao Wu, Varun Jampani, Deqing Sun, and Huaizu Jiang. Hoi-diff: Text-driven synthesis of 3d human-object interactions using diffusion models. arXiv preprint arXiv:2312.06553, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 119, + 555, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 119, + 555, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 119, + 555, + 162 + ], + "type": "text", + "content": "[41] Mathis Petrovich, Or Litany, Umar Iqbal, Michael J Black, Gul Varol, Xue Bin Peng, and Davis Rempe. Multi-track timeline control for text-driven 3d human motion generation. In CVPR, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 165, + 555, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 165, + 555, + 208 + ], + "spans": [ + { + "bbox": [ + 316, + 165, + 555, + 208 + ], + "type": "text", + "content": "[42] Ekkasit Pinyoanuntapong, Muhammad Usama Saleem, Pu Wang, Minwoo Lee, Srijan Das, and Chen Chen. Bamm: Bidirectional autoregressive motion model. arXiv preprint arXiv:2403.19435, 2024." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 211, + 555, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 211, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 316, + 211, + 555, + 243 + ], + "type": "text", + "content": "[43] Ekkasit Pinyoanuntapong, Pu Wang, Minwoo Lee, and Chen Chen. Mmm: Generative masked motion model. In CVPR, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 246, + 555, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 246, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 316, + 246, + 555, + 300 + ], + "type": "text", + "content": "[44] Ziyun Qian, Zeyu Xiao, Zhenyi Wu, Dingkang Yang, Mingcheng Li, Shunli Wang, Shuaibing Wang, Dongliang Kou, and Lihua Zhang. Smcd: High realism motion style transfer via mamba-based diffusion. arXiv preprint arXiv:2405.02844, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 303, + 555, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 303, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 316, + 303, + 555, + 336 + ], + "type": "text", + "content": "[45] Sigal Raab, Inbal Leibovitch, Guy Tevet, Moab Arar, Amit H Bermano, and Daniel Cohen-Or. Single motion diffusion. arXiv preprint arXiv:2302.05905, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 338, + 555, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 555, + 392 + ], + "type": "text", + "content": "[46] Sigal Raab, Inbar Gat, Nathan Sala, Guy Tevet, Rotem Shalev-Arkushin, Ohad Fried, Amit H Bermano, and Daniel Cohen-Or. Monkey see, monkey do: Harnessing self-attention in motion diffusion for zero-shot motion transfer. arXiv preprint arXiv:2406.06508, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 395, + 555, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 395, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 395, + 555, + 449 + ], + "type": "text", + "content": "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 453, + 554, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 554, + 495 + ], + "type": "text", + "content": "[48] Davis Rempe, Zhengyi Luo, Xue Bin Peng, Ye Yuan, Kris Kitani, Karsten Kreis, Sanja Fidler, and Or Litany. Trace and pace: Controllable pedestrian animation via guided trajectory diffusion. In CVPR, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 498, + 554, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 554, + 531 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 554, + 531 + ], + "type": "text", + "content": "[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 533, + 555, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 555, + 597 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 555, + 597 + ], + "type": "text", + "content": "[50] Wenfeng Song, Xingliang Jin, Shuai Li, Chenglizhao Chen, Aimin Hao, Xia Hou, Ning Li, and Hong Qin. Arbitrary motion style transfer with multi-condition motion latent diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 821-830, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 601, + 555, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 555, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 555, + 655 + ], + "type": "text", + "content": "[51] Xiangjun Tang, Linjun Wu, He Wang, Bo Hu, Xu Gong, Yuchen Liao, Songnan Li, Qilong Kou, and Xiaogang Jin. Rsmt: Real-time stylized motion transition for characters. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "type": "text", + "content": "[52] Yiwen Tang, Zoey Guo, Zhuhao Wang, Ray Zhang, Qizhi Chen, Junli Liu, Delin Qu, Zhigang Wang, Dong Wang, Xuelong Li, et al. Exploring the potential of encoder-free architectures in 3d Imms. arXiv preprint arXiv:2502.09620, 2025. 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13358" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[53] Tianxin Tao, Xiaohang Zhan, Zhongquan Chen, and Michiel van de Panne. Style-erd: Responsive and coherent online motion style transfer. 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "text", + "content": "[54] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2023. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "text", + "content": "[55] Weilin Wan, Zhiyang Dou, Taku Komura, Wenping Wang, Dinesh Jayaraman, and Lingjie Liu. Tlcontrol: Trajectory and language control for human motion synthesis. arXiv preprint arXiv:2311.17135, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 218 + ], + "type": "text", + "content": "[56] Yu-Hui Wen, Zhipeng Yang, Hongbo Fu, Lin Gao, Yanan Sun, and Yong-Jin Liu. Autoregressive stylized motion synthesis with generative flow. In CVPR, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "text", + "content": "[57] Qianyang Wu, Ye Shi, Xiaoshui Huang, Jingyi Yu, Lan Xu, and Jingya Wang. Thor: Text to human-object interaction diffusion via relation intervention. arXiv preprint arXiv:2403.11208, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "type": "text", + "content": "[58] Qi Wu, Yubo Zhao, Yifan Wang, Yu-Wing Tai, and Chi-Keung Tang. Motionllm: Multimodal motion-language learning with large language models. arXiv preprint arXiv:2405.17013, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 309, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 342 + ], + "type": "text", + "content": "[59] Yiming Xie, Varun Jampani, Lei Zhong, Deqing Sun, and Huaizu Jiang. Omnicontrol: Control any joint at any time for human motion generation. In ICLR, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 344, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 294, + 376 + ], + "type": "text", + "content": "[60] Jingwei Xu, Huazhe Xu, Bingbing Ni, Xiaokang Yang, Xiaolong Wang, and Trevor Darrell. Hierarchical style-based networks for motion synthesis. In ECCV, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 377, + 294, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 294, + 410 + ], + "type": "text", + "content": "[61] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. Interdiff: Generating 3d human-object interactions with physics-informed diffusion. In ICCV, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 411, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 443 + ], + "type": "text", + "content": "[62] Sirui Xu, Ziyin Wang, Yu-Xiong Wang, and Liang-Yan Gui. Interdreamer: Zero-shot text to 3d dynamic human-object interaction. arXiv preprint arXiv:2403.19652, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 445, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 510 + ], + "type": "text", + "content": "[63] Le Xue, Mingfei Gao, Chen Xing, Roberto Martin-Martin, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1179-1189, 2023. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 512, + 294, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 565 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 565 + ], + "type": "text", + "content": "[64] Senqiao Yang, Jiaming Liu, Ray Zhang, Mingjie Pan, Ziyu Guo, Xiaqi Li, Zehui Chen, Peng Gao, Yandong Guo, and Shanghai Zhang. Lidar-llm: Exploring the potential of large language models for 3d lidar understanding. AAAI 2025, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 567, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 294, + 600 + ], + "type": "text", + "content": "[65] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In ICCV, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 601, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 646 + ], + "type": "text", + "content": "[66] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3836-3847, 2023. 2, 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[67] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. PAMI, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[68] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Point" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 441 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 72, + 555, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 555, + 94 + ], + "type": "text", + "content": "clip: Point cloud understanding by clip. In CVPR 2022, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 96, + 555, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 149 + ], + "type": "text", + "content": "[69] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In ECCV 2022. Springer Nature Switzerland, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 152, + 555, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 195 + ], + "type": "text", + "content": "[70] Renrui Zhang, Jiaming Han, Chris Liu, Aojun Zhou, Pan Lu, Yu Qiao, Hongsheng Li, and Peng Gao. Llama-adapter: Efficient fine-tuning of large language models with zero-initialized attention. In ICLR 2024, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 197, + 555, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 555, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 555, + 250 + ], + "type": "text", + "content": "[71] Renrui Zhang, Dongzhi Jiang, Yichi Zhang, Haokun Lin, Ziyu Guo, Pengshuo Qiu, Aojun Zhou, Pan Lu, Kai-Wei Chang, Peng Gao, et al. Mathverse: Does your multi-modal llm truly see the diagrams in visual math problems? ECCV 2024, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 252, + 555, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 555, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 555, + 306 + ], + "type": "text", + "content": "[72] Renrui Zhang, Xinyu Wei, Dongzhi Jiang, Ziyu Guo, Shicheng Li, Yichi Zhang, Chengzhuo Tong, Jiaming Liu, Aojun Zhou, Bin Wei, et al. Mavis: Mathematical visual instruction tuning with an automatic data engine. arXiv preprint arXiv:2407.08739, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 308, + 554, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 554, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 554, + 341 + ], + "type": "text", + "content": "[73] Lei Zhong, Yiming Xie, Varun Jampani, Deqing Sun, and Huaizu Jiang. Smoodi: Stylized motion diffusion model. In ECCV, 2024. 2, 3, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 342, + 555, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 555, + 396 + ], + "type": "text", + "content": "[74] Bin Zhu, Bin Lin, Munan Ning, Yang Yan, Jiaxi Cui, HongFa Wang, Yatian Pang, Wenhao Jiang, Junwu Zhang, Zongwei Li, et al. Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. arXiv preprint arXiv:2310.01852, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 398, + 555, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 398, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 398, + 555, + 441 + ], + "type": "text", + "content": "[75] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. ICCV 2023, 2023. 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13359" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/StyleSRN_ Scene Text Image Super-Resolution with Text Style Embedding/34baee7e-bb1d-4173-bb75-506b85eb6a0f_content_list.json b/2025/StyleSRN_ Scene Text Image Super-Resolution with Text Style Embedding/34baee7e-bb1d-4173-bb75-506b85eb6a0f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..44519a0933fe7e494da96788fb2d5ea3ae0c71ce --- /dev/null +++ b/2025/StyleSRN_ Scene Text Image Super-Resolution with Text Style Embedding/34baee7e-bb1d-4173-bb75-506b85eb6a0f_content_list.json @@ -0,0 +1,1377 @@ +[ + { + "type": "text", + "text": "StyleSRN: Scene Text Image Super-Resolution with Text Style Embedding", + "text_level": 1, + "bbox": [ + 124, + 130, + 872, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shengrong Yuan $^{1}$ , Runmin Wang $^{1,*}$ , Ke Hao $^{1}$ , Xuqi Ma $^{1}$ , Changxin Gao $^{2}$ , Li Liu $^{3}$ , Nong Sang $^{2}$ $^{1}$ School of Information Science and Engineering, Hunan Normal University \n $^{2}$ School of Artificial Intelligence and Automation, Huazhong University of Science and Technology \n $^{3}$ School of Electronic Science, National University of Defense Technology", + "bbox": [ + 104, + 179, + 890, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{shengrongyuan, runminwang, ke, xqma}@hunnu.edu.cn, {cgao, nsang}@hust.edu.cn, liliu@oulu.fi", + "bbox": [ + 130, + 253, + 859, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 304, + 326, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Scene text image super-resolution (STISR) focuses on enhancing the clarity and readability of low-resolution text images. Existing methods often rely on text probability distribution priors derived from text recognizers to guide the super-resolution process. While effective in capturing general structural information of text, these priors lack the ability to preserve specific text style details, such as font, stereoscopic effect and spatial transformation, leading to a loss of visual quality and stylistic consistency. To address these limitations, we propose a Style embedding-based scene text image Super-Resolution Network (StyleSRN), which introduces a text style embedding mechanism to preserve and enhance text style features during the super-resolution process. The proposed architecture includes Style Enhancement Blocks for capturing multi-scale cross-channel dependencies, and Style Content Fusion Blocks that effectively integrates text content with style information, ensuring that the structure and style of the restored text are not distorted. Furthermore, we introduce a Text Style Loss based on the Gram matrix to supervise the reconstruction process at the style level, thereby maintaining the stylistic consistency of the restored text images. Extensive experiments on the TextZoom dataset and five scene text recognition benchmarks demonstrate the superiority of our method. Code is available at: https://github.com/Yuanssr/StyleSRN.", + "bbox": [ + 89, + 335, + 483, + 714 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 743, + 220, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Scene text images contain both visual and linguistic modalities, which can be extracted and analyzed to enhance various applications, including autonomous driving [32], visual question answering [1], and card recognition [13], etc. However, due to the limitations of imaging conditions of sensing devices, such as shooting jitter and low-focus cameras, scene text images often suffer from varying degrad", + "bbox": [ + 89, + 768, + 482, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e9d9b1c05c501673d15bb5d50fdfba13a5fdd98fb7474310be5d15cbf9abb91c.jpg", + "image_caption": [ + "Figure 1. Comparison of super-resolution results between previous methods and ours. Methods without text priors such as TSRN [30] retains less style information, while text priors-guided methods such as TATT [17] and LEMMA [7] almost completely lose the style information." + ], + "image_footnote": [], + "bbox": [ + 527, + 311, + 890, + 468 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion, including blur, distortion, etc. Unlike general object images, scene text images possess more complex character structures and fine-grained stroke-level details, making even slight degradation challenging to recognize. Such low-resolution scene text images significantly hinder the performance of downstream tasks such as scene text detection, optical character recognition, and scene text recognition. Thus, restoring both the visual style quality and character structures of these degraded images is crucial.", + "bbox": [ + 511, + 580, + 906, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, many super-resolution methods specifically designed for scene text images have used the unique properties of text to improve restoration. Early approaches primarily focused on extracting sequential information from text images. Wang et al. [30] proposed a sequential residual block to capture the inherent sequential nature of text images. Chen et al. [2] introduced a text-focus loss function that emphasizes character positions and content, thereby improving recognition accuracy. More recently, methods that incorporate deep character classification information obtained from text recognizers have become the norm. Ma et al. [18] highlighted the benefits of integrating text", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18693", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "prriors into the super-resolution process. Moreover, Ma et al. [17] developed an attention-based text priors interpreter to interactively fuse text probability distributions with image features. Guo et al. [7] demonstrated the effectiveness of using a bidirectional alignment strategy for text and images to obtain high-level guidance information. Most of these methods rely on text recognizers to generate text probability distribution priors, which are then encoded into the super-resolution network to guide the recovery of character structures.", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, the text probability distribution priors derived from text recognizers mainly capture the general character structures but are less sensitive to the variations in text styles across different images, which hampers the super-resolution process. This loss of style details not only affects subsequent scene text recognition tasks but also disrupts the visual consistency of the images. The problem is particularly pronounced in scene text images with distinctive text styles, such as artistic characters or logos, where the loss of stylistic features can lead to more severe distortion in the restored images. As shown in Figure 1, text priors-guided methods such as TATT [17] and LEMMA [7] almost completely lose the style information of the degraded text image, while methods without text priors such as TSRN [30] retain less style information. To address these challenges, we draw inspiration from Style Transfer [6, 8, 12] to explore the potential of preserving image style information to enhance the restoration of low-resolution text images. Unlike content information, which is sensitive to position, style information in an image can be viewed as the correlation between different feature channels and is largely position-insensitive. By effectively encoding and embedding this style information during the super-resolution process, our proposed method aims to preserve both the structural integrity and the stylistic consistency of the restored text images.", + "bbox": [ + 91, + 247, + 483, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a Style embedding-based scene text image Super-Resolution Network (StyleSRN) for STISR, which not only restores the character structure but also preserves the stylistic consistency of the text. Unlike conventional methods that rely solely on text probability priors to guide the super-resolution process, our approach introduces a Text Style Embedding Branch that extracts and enhances style features. Additionally, we design a Style Enhancement Block (SEB), which captures multi-scale cross-channel dependencies, thereby enhancing the representation of style information. To further ensure the integration of style and content, we introduce a Style Content Fuse Block (SCFB) that effectively combines style embeddings with text priors. Finally, we propose a Text Style Loss based on the Gram matrix to supervise the reconstructed images at the style level. As illustrated in Figure 1, our method achieves superior visual quality and stylistic consistency in the restored images. Overall, our contributions can be sum", + "bbox": [ + 89, + 628, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "marized as follows:", + "bbox": [ + 513, + 92, + 645, + 104 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We, for the first time to our best knowledge, successfully embed text style information to guide super-resolution process for STISR task, and validate its effectiveness to ensure both structural integrity and stylistic consistency of the super-resolution process.", + "- We propose a Text Style Loss based on the Gram Matrix which leverages a text recognizer to extract features to offer robust supervision at style level for scene text images, which effectively improve both image quality and text recognition accuracy.", + "- Extensive experiment results demonstrate that the proposed method not only achieves state-of-the-art performance on the TextZoom dataset but also validates exceptional generalization across five scene text recognition datasets." + ], + "bbox": [ + 511, + 109, + 903, + 334 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 352, + 653, + 368 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Scene Text Image Super-resolution", + "text_level": 1, + "bbox": [ + 513, + 377, + 815, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Scene text image super-resolution aims to enhance the visual quality and readability of scene text images. Early studies in this field usually adopt general image super-resolution methods [4, 5, 14, 21, 28] to reconstruct scene text images. However, due to the complex character structure of text images, these methods often fail to produce sufficiently recognizable text images. Therefore, some unique attributes of scene text have been exploited to improve super-resolution performance[2, 3, 7, 17, 18, 30]. For example, some methods focus on the sequential information of scene text. Wang et al. [30] introduced sequential residual block to capture the sequential information in text images. Chen et al. [2] proposed a Transformer-Based SuperResolution Network containing a Self-Attention Module to extract sequential information. Recently, the method of using text recognizers to obtain text probability distribution priors has become mainstream. For example, Ma et al. [18] incorporates categorical text priors into the STISR model training process, while Ma et al. [17] proposed a text prior interpreter to better transform text priors. Guo et al. [7] proposed a bimodal alignment mechanism to better utilize text priors. Additionally, some works use text priors to guide diffusion models to restore low-resolution images [20, 35]. Most of these methods use a pre-trained text recognizer to obtain deep character classification information to guide the super-resolution network. While this approach enables the model to effectively learn the general character structure, it results in the loss of specific style information in each image.", + "bbox": [ + 511, + 398, + 903, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Style Transfer", + "text_level": 1, + "bbox": [ + 513, + 847, + 660, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Style transfer is a technique that involves altering the stylistic attributes of an image while maintaining its core con", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18694", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/76805dc2ee52ed0c4838498beecf290ad5891d41cf4d92e87938a8565ec1ce3d.jpg", + "image_caption": [ + "Figure 2. Overview architecture of the proposed StyleSRN. GAP denotes the Global Average Pooling. $G(\\cdot)$ denotes the Gram matrix, $\\Phi_l(\\cdot)$ represents the feature map obtained from the $l$ -th layer of the text recognizer. $\\mathbf{I}_{\\mathrm{LR}}$ , $\\mathbf{I}_{\\mathrm{SR}}$ and $\\mathbf{I}_{\\mathrm{HR}}$ represent low-resolution images, super-resolution images and high-resolution images, respectively." + ], + "image_footnote": [], + "bbox": [ + 114, + 89, + 890, + 361 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tent. This concept was pioneered by Gatys et al. [6], who utilized convolutional neural networks (CNNs) to transfer artistic styles between images by matching the Gram matrices of feature maps. Subsequent work has expanded on this foundation, exploring various methods to improve both the quality and efficiency of style transfer. One significant advancement is the introduction of adaptive instance normalization (AdaIN) [8], which allows for real-time arbitrary style transfer by aligning the mean and variance of content features to those of style features. Some works employ style transfer technology specifically for scene text style transfer [23, 27]. In our work, we draw inspiration from these style transfer methods, particularly in how they encode and embed style features. By adopting these principles, we aim to restore text style information, which is crucial for maintaining the visual consistency of the super-resolved text images.", + "bbox": [ + 88, + 441, + 486, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 700, + 225, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overall Architecture", + "text_level": 1, + "bbox": [ + 89, + 726, + 287, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall architecture of the proposed StyleSRN is illustrated in Figure 2. Our method consists of a text style embedding branch for extracting and enhancing style features and a super-resolution branch for restoring low-resolution images $\\mathbf{I}_{\\mathrm{LR}}$ . For the text style embedding branch, the low-resolution image will first pass through ResNet18 to obtain the feature map $\\mathbf{X}_f \\in \\mathbb{R}^{C \\times W \\times H}$ , where $C$ represents the number of channels, and $H$ and $W$ represent the height and width of the feature map, respectively. Then several stacked Style Enhancement Blocks will extract the correla", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion information between the feature map channels. After the style enhancement, we perform global average pooling on it to obtain the latent vector $\\mathbf{z}$ . Given a latent vector $\\mathbf{z}$ in the input latent space $\\mathcal{Z}$ , a non-linear mapping network $f: \\mathcal{Z} \\to \\mathcal{W}$ first produces $\\mathbf{w} \\in \\mathcal{W}$ . Learned affine transformations then specialize $\\mathbf{w}$ to styles $\\mathbf{y} = (\\mathbf{y}_s, \\mathbf{y}_b)$ , where $\\mathbf{y}_s$ and $\\mathbf{y}_b$ denote the learned normalization scaling and bias coefficients, respectively. For the Super-Resolution Branch, the low-resolution image will first be processed by the Spatial Transformer Network (STN) [9] to align the spatially deformed texts. The convolution layer will extract features from the image processed by STN to obtain shallow image features. Then the shallow image features, text priors, and style information will be sent to several stacked style-content fusion blocks for style and content fusion. Each block contains a Style Enhancement Block, an adaptive instance normalization (AdaIN) [8], and a content embedding block. Finally, PixelShuffle is applied as an upsampling operator to generate $\\mathbf{I}_{\\mathrm{SR}}$ .", + "bbox": [ + 511, + 441, + 908, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Style Enhancement Block", + "text_level": 1, + "bbox": [ + 511, + 739, + 746, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The style of an image has been traditionally viewed as the correlation between different features, as demonstrated in the advancements of style transfer [6, 8, 12]. While previous approaches, such as [12], directly utilized ResNet features to obtain style vectors, our approach introduces a novel enhancement by applying channel-wise attention to these feature maps, thereby better capturing and preserving the stylistic characteristics in scene text images.", + "bbox": [ + 511, + 763, + 906, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by the Efficient Channel Attention (ECA)", + "bbox": [ + 531, + 885, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18695", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mechanism [29], we designed a Style Enhancement Block (SEB) to improve the model's capacity for capturing the correlation between different channel features. Let the input feature map be denoted as $\\mathbf{X}_f\\in \\mathbb{R}^{C\\times H\\times W}$ . We begin by applying Global Average Pooling (GAP) to the input feature map to obtain channel-wise statistics $\\mathbf{m}\\in \\mathbb{R}^C$ . To further improve the model's capacity for capturing local cross-channel interactions at multiple scales, we employ multiple 1D convolution kernels of varying sizes, in contrast to the fixed-size kernels used in [29]. Specifically, let $n$ denote the number of different kernel sizes, each represented by $k_{l}$ , where $l\\in 1,2,\\ldots ,n$ . The convolutional operation at each scale can be expressed as:", + "bbox": [ + 89, + 90, + 483, + 287 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {m} _ {c} ^ {\\prime (l)} = \\sum_ {j = - \\frac {k _ {l}}{2}} ^ {\\frac {k _ {l}}{2}} b _ {j} ^ {l} \\mathbf {m} _ {c + j}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 300, + 483, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{b}^l\\in \\mathbb{R}^{k_l}$ represents the weight vector for the convolution kernel of size $k_{l}$ . This multi-scale approach allows the model to capture a more nuanced set of cross-channel dependencies, enhancing the representation of style information. The outputs from the multiple convolutional scales are aggregated by computing the final channel attention weight $\\mathbf{m}'_c$ as a weighted sum:", + "bbox": [ + 89, + 366, + 483, + 474 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {m} _ {c} ^ {\\prime} = \\sum_ {l = 1} ^ {n} \\alpha_ {l} \\mathbf {m} _ {c} ^ {\\prime (l)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 484, + 483, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha_{l}$ denotes the weighting coefficient for the output of the $l$ -th convolutional scale. These weights are normalized using a Sigmoid function:", + "bbox": [ + 89, + 537, + 483, + 583 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {f} ^ {\\prime c, i, j} = \\sigma \\left(\\mathbf {m} _ {c} ^ {\\prime}\\right) \\cdot \\mathbf {X} _ {f} ^ {c, i, j}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 595, + 482, + 618 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sigma (\\cdot)$ represents the Sigmoid activation function, and $\\mathbf{X}^{\\prime}$ is the final output feature map with the applied channelwise attention.", + "bbox": [ + 89, + 627, + 482, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The enhanced feature map is passed through several SEBs, which refines the correlation between different feature channels, ensuring that the model pays more attention to the features of important channels. This process effectively enhances the model's capability to capture and preserve the intricate stylistic details in scene text images. Subsequently, global average pooling (GAP) is applied to the final enhanced feature map to obtain a latent vector $\\mathbf{z}$ in the latent space $\\mathcal{Z}$ . The latent vector is then processed by a non-linear style mapping network $f: \\mathcal{Z} \\to \\mathcal{W}$ , which produces a vector $\\mathbf{w} \\in \\mathcal{W}$ that decouples the latent space. The style mapping network consists of an 8-layer MLP with LeakyReLU activation functions. Finally, learned affine transformations are applied to transform $\\mathbf{w}$ into style parameters $\\mathbf{y} = (\\mathbf{y}_s, \\mathbf{y}_b)$ .", + "bbox": [ + 89, + 674, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Style Content Fuse Block", + "text_level": 1, + "bbox": [ + 513, + 90, + 743, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the field of scene text image super-resolution, existing approaches primarily focus on embedding content information, including deep character category information derived from text probability priors. While this approach effectively guides the super-resolution network in restoring character structure, it often results in the loss of crucial stylistic features. To address this limitation, we propose a novel Style Content Fuse Block (SCFB) that allows the super-resolution network to restore character structures while preserving the stylistic consistency of the characters.", + "bbox": [ + 511, + 112, + 906, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our innovation lies in embedding style information directly into the super-resolution network alongside content information. This dual embedding approach ensures that both character structure and style are preserved during the super-resolution process. Specifically, for the shallow image features $\\mathbf{X}_s\\in \\mathbb{R}^{C'\\times H'\\times W'}$ obtained through a Spatial Transformer Network (STN) and a convolutional layer, we use Adaptive Instance Normalization (AdaIN) [8] to modulate their style. AdaIN has been shown to be effective in transferring style information across different domains. Given a feature map $\\mathbf{X}_s$ and a style vector $\\mathbf{y}$ , the AdaIN operation can be defined as follows:", + "bbox": [ + 511, + 265, + 908, + 446 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A d a I N} \\left(\\mathbf {X} _ {s}, \\mathbf {y}\\right) = \\mathbf {y} _ {s} \\left(\\frac {\\mathbf {X} _ {s} - \\mu \\left(\\mathbf {X} _ {s}\\right)}{\\sigma \\left(\\mathbf {X} _ {s}\\right)}\\right) + \\mathbf {y} _ {b}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 457, + 906, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mu (\\cdot)$ and $\\sigma (\\cdot)$ represent the mean and variance of $\\mathbf{X}_s$ . Each AdaIN layer within the SCFB normalizes feature maps independently, with the dimensionality of $\\mathbf{y}$ being twice the number of feature map channels $C^\\prime$ in that layer. This embedding ensures that style information is seamlessly integrated into the spatial features of the image, thus preserving the character's stylistic attributes.", + "bbox": [ + 511, + 503, + 906, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to embedding style information, content information, such as text probability distribution priors, should also be embedded into the super-resolution network to guide the learning of character structures. To bridge the domain gap between text priors and image features, we adopt the TP Interpreter architecture from [17] to extract the content embedding features. These interpreted features are then concatenated with the style-embedded features and subsequently passed through a Sequential Residual Block [30] for further refinement.", + "bbox": [ + 511, + 609, + 908, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Text Style Loss", + "text_level": 1, + "bbox": [ + 513, + 771, + 666, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although the proposed StyleSRN network effectively embeds the text's style features, these features are derived from the low-resolution image itself. When the style information in the low-resolution image is degraded, even if the network extracts these style features, it cannot fully reconstruct the style of the corresponding high-resolution image. Therefore, the reconstructed text image requires supervision at", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18696", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the style level to effectively learn the degradation of style information in the low-resolution image.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To supervise the style in the reconstructed image, we introduce a Text Style Loss based on the Gram matrix. Unlike traditional methods [6] that use VGG networks to capture style information, our approach leverages a text recognizer to extract feature maps for Gram matrix computation. This enables better capture of text-specific style features, which are crucial for scene text images.", + "bbox": [ + 89, + 121, + 483, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Gram matrix $\\mathbf{G}_l$ for a given feature map $\\Phi_l(\\mathbf{I})$ at the $l$ -th layer is computed as follows. First, we reshape the feature map $\\Phi_l(\\mathbf{I})$ of dimensions $C_l \\times H_l \\times W_l$ into a matrix $\\mathbf{F}_l$ of size $C_l \\times (H_l \\times W_l)$ , where each row corresponds to a channel and each column corresponds to a pixel. The Gram matrix $\\mathbf{G}_l$ is the inner product between the different channels in layer $l$ :", + "bbox": [ + 89, + 227, + 483, + 335 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {G} _ {i j} ^ {l} = \\sum_ {k = 1} ^ {H _ {l} \\times W _ {l}} \\mathbf {F} _ {i k} ^ {l} \\mathbf {F} _ {j k} ^ {l} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 344, + 482, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Gram matrix captures the correlations between different channels in the feature map, thereby encoding the style information of the image.", + "bbox": [ + 89, + 397, + 483, + 443 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Text Style Loss $\\mathcal{L}_{\\mathrm{style}}$ is defined as the Mean Squared Error (MSE) between the Gram matrix of the high-resolution (HR) ground truth and the super-resolved (SR) output:", + "bbox": [ + 89, + 443, + 483, + 503 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s t y l e}} = \\sum_ {l = 1} ^ {L} \\frac {1}{4 C _ {l} ^ {2} H _ {l} ^ {2} W _ {l} ^ {2}} \\| \\mathbf {G} \\left(\\Phi_ {l} \\left(\\mathbf {I} _ {\\mathrm {H R}}\\right)\\right) - \\mathbf {G} \\left(\\Phi_ {l} \\left(\\mathbf {I} _ {\\mathrm {S R}}\\right)\\right) \\| _ {2} ^ {2} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 515, + 483, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{G}(\\cdot)$ denotes the Gram matrix, $\\Phi_l(\\cdot)$ represents the feature map obtained from the $l$ -th layer of the text recognizer, and $C_l, H_l, W_l$ are the dimensions of the feature map at layer $l$ .", + "bbox": [ + 89, + 569, + 483, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Overall Loss Function", + "text_level": 1, + "bbox": [ + 89, + 638, + 299, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In training, the overall loss function includes a Mean Square Error (MSE) loss $\\mathcal{L}_2$ , a Text-Focus Loss $\\mathcal{L}_{TP}$ [30], and the proposed Text Style Loss $\\mathcal{L}_{style}$ . The $\\mathcal{L}_2$ measures the difference between the super-resolution (SR) output and the ground-truth high-resolution (HR) image. The overall loss function is described as follows:", + "bbox": [ + 89, + 662, + 483, + 752 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {2} + \\alpha \\mathcal {L} _ {T P} + \\beta \\mathcal {L} _ {\\text {s t y l e}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 765, + 482, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the $\\alpha$ and $\\beta$ are the balancing parameters.", + "bbox": [ + 89, + 794, + 413, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 823, + 223, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 89, + 848, + 192, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "TextZoom TextZoom [30] is a scene text super-resolution dataset designed for real-world scenarios, containing", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "21,740 pairs of low-resolution and high-resolution images captured by cameras with various focal lengths in natural environments, with 17,367 samples used for training. The test set is divided into easy, medium, and hard subsets, comprising 1,619, 1,411, and 1,343 LR-HR pairs, respectively, based on the camera focal lengths.", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Scene Text Recognition Datasets To assess the robustness and generalization capability of StyleSRN, we conducted comprehensive evaluations on five benchmark scene text recognition datasets: IC13 [10], IC15 [11], CUTE80 [24], IIIT5K [19], and SVTP [22]. These datasets were manually degraded to create low-resolution counterparts. Detailed descriptions of the datasets and degradation procedures are provided in the supplementary materials.", + "bbox": [ + 511, + 203, + 906, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Metric", + "text_level": 1, + "bbox": [ + 513, + 335, + 602, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following previous works, we adopt two types of metrics: super-resolution image text recognition accuracy and image quality. Specifically, to evaluate the generalization capability of StyleSRN, we employed three scene text recognizers: CRNN [25], MORAN [16], and ASTER [26]. For the image quality metric, we adopt the widely used Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM)[31].", + "bbox": [ + 511, + 358, + 905, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 489, + 730, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We implement our model using PyTorch. All experiments are conducted on a single RTX 3090 GPU. We train our model with a batch size of 64 for 500 epochs using AdamW [15] for optimization. The learning rate is set to 1e-3 and decayed by a factor of 0.5 after 400 epochs. The number of SEBs and SCFBs are set to 3 and 5, respectively. In the SEBs, the sizes of the convolution kernels are $3 \\times 3$ , $5 \\times 5$ , and $7 \\times 7$ , respectively. We use CRNN to generate the features for computing the text style loss and empirically set the number of feature maps L used to calculate the style loss to 5. We set the parameters $\\alpha$ and $\\beta$ to 0.5 and 1.0, respectively.", + "bbox": [ + 511, + 512, + 906, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 705, + 666, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we conduct ablation studies to investigate the effectiveness of SEBs and SCFBs, the effectiveness of training with Text Style Loss, the impact of numbers of SEBs and SCFBs. For STISR methods that do not use text priors, such as TSRN [30] and TBSRN [2], using SCFB means that image features are not used for text prior fusion, but are directly concatenated with style-embedded features.", + "bbox": [ + 511, + 727, + 905, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Effectiveness of SEBs and SCFBs. We compared the performance of different STISR methods using three strategies: No Style Embedding (NSE), only SCFB, and SCFB", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18697", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/62c6d87eefd0af27d741d4506ba7050e28460a2f36f7f5e2a961505de7ff97fe.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodNSESCFBSEBImage QualityAccuracy(%)
PSNRSSIM
TSRN [30]✓--21.420.769041.4
-✓-21.540.769943.5
-✓✓21.560.769544.4
TBSRN [2]✓--20.910.760348.1
-✓-21.040.762350.3
-✓✓21.220.761451.1
TPGSR [18]✓--20.970.771951.8
-✓-21.040.772352.4
-✓✓21.170.772352.7
TATT [17]✓--21.520.793052.6
-✓-21.660.779054.1
-✓✓21.870.778554.9
LEMMA [7]✓--20.880.776056.3
-✓-20.940.777156.7
-✓✓21.050.776557.0
", + "bbox": [ + 94, + 89, + 478, + 292 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9de9154bdec70af2fdfa5535f23f27e13fc1b523de33df9dc5ac888e56c6ba8b.jpg", + "table_caption": [ + "Table 1. Effectiveness of the Style Content Fuse Block (SCFB) and the Style Enhancement Block (SEB). NSE refers to No Style Embedding. Accuracy represents the average accuracy with CRNN [25]." + ], + "table_footnote": [], + "table_body": "
Method\\( \\mathcal{L}_{\\text{style}} \\)PSNRSSIMAccuracy(%)
TSRN [30]-21.420.769041.4
✓21.560.771143.7
TPGSR [18]-20.970.771951.8
✓21.120.772452.4
TATT [17]-21.520.793052.6
✓21.810.793853.8
StyleSRN-21.540.775854.9
✓21.820.777857.4
", + "bbox": [ + 94, + 388, + 480, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "combined with the SEB in Table 2. NSE means removing both SEB and SCFB. SCFB can be used independently of SEB, but when using SEB, SCFB must be used to embed style information. Removing SEB means that the features extracted by Resnet18 are directly fed into the style mapping network through global pooling. Removing SCFB means that the features directly passed through the text interpreter do not need to be cascaded with the style features. As presented in Table 1, using only SCFB generally improves image quality and recognition accuracy over NSE. In addition, SCFB has a greater improvement over methods without text priors such as TSRN [30] and TBSRN [2]. For instance, with TSRN, the introduction of SCFB increases accuracy by $2.1\\%$ , along with notable gains in PSNR and SSIM. We speculate that this may be because these methods do not have any prior guidance, while SCFB introduces the guidance of style information. Using SCFB and SEB at the same time can effectively improve PSNR and text recognition accuracy, but slightly reduce SSIM. This shows that SEB can effectively enhance style information, but slightly reduce structural consistency.", + "bbox": [ + 88, + 582, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7efdac0064330c72787acbd80c757b82bfbdeb595ad21af9fa30b17f43f22689.jpg", + "table_caption": [ + "Table 2. Ablation of the Text Style Loss. Accuracy represents the average accuracy with CRNN [25]." + ], + "table_footnote": [], + "table_body": "
DiscriminatorPSNRSSIMAccuracy (%)
VGG21.310.776156.3
CRNN21.820.777857.4
", + "bbox": [ + 517, + 89, + 903, + 131 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/47f37610fdc5dfce9a78865b669f5cde2e7686d4eb63a1386a2e01ad8aa2c2f8.jpg", + "table_caption": [ + "Table 3. Ablation of the discriminator used in text style loss." + ], + "table_footnote": [], + "table_body": "
NQuality MetricAccuracy(%)
PSNRSSIMCRNN [25]MORAN [16]ASTER [26]
020.760.761854.661.865.0
121.820.777856.363.366.2
221.770.776456.964.066.7
421.770.777057.164.067.0
521.730.776855.963.266.7
321.820.777857.464.167.3
", + "bbox": [ + 517, + 180, + 903, + 286 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/48fb7d9208a9e7794991f42c3dcb69f9925c8e1c28824c53ca0fe2a80e94ba65.jpg", + "table_caption": [ + "Table 4. Ablation of the number of Style Enhancement Blocks." + ], + "table_footnote": [], + "table_body": "
NQuality MetricAccuracy(%)
PSNRSSIMCRNN [25]MORAN [16]ASTER [26]
020.670.743153.261.064.7
321.450.772255.362.366.2
421.430.775156.763.966.9
621.740.776956.964.067.0
721.550.775656.062.866.2
521.820.777857.464.167.3
", + "bbox": [ + 517, + 335, + 903, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 5. Ablation of the number of Style Content Fuse Blocks.", + "bbox": [ + 521, + 453, + 893, + 467 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness of training with Text Style Loss. We show the effectiveness of the Text Style Loss in Table 2. Incorporating the Text Style Loss results in improvements across all metrics. For instance, the PSNR of TSRN increases from 21.42 to 21.56, and its SSIM improves from 0.7690 to 0.7711. In terms of text recognition accuracy, models incorporating the Text Style Loss demonstrate enhanced performance. The accuracy of TSRN improves from $41.4\\%$ to $43.7\\%$ . These results highlight that the Text Style Loss effectively enhances both image quality and text recognition accuracy. Besides, we also compared the impact of using different discriminators in Table 3. It can be seen that using CRNN performs better than VGG.", + "bbox": [ + 511, + 501, + 906, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Numbers of SEBs. We show the effect of the number of SEBs in Table 4. As the number of SEBs increases, there is a noticeable improvement in both the PSNR and SSIM metrics. Specifically, using 3 SEBs achieves the highest PSNR of 21.82 and an SSIM of 0.7778, indicating the optimal enhancement of image quality. In terms of text recognition accuracy, models incorporating 3 SEBs also demonstrate superior performance across all recognition models. These findings suggest that SEBs significantly enhance both image quality and text recognition performance, with 3 SEBs being the optimal configuration.", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18698", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/94c83dbd65ba253d2744496b44701ec1a6bdf20dae78373e964f5f5550bd6ccf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAccuracy of CRNN [25] (%)Accuracy of MORAN [16] (%)Accuracy of ASTER [26] (%)
EasyMediumHardAvgEasyMediumHardAvgEasyMediumhardAvg
BICUBIC36.421.121.126.860.637.930.844.167.442.431.248.2
HR76.475.164.672.491.285.374.284.194.287.776.286.6
TSRN [30]52.538.231.441.470.155.337.955.475.156.340.158.3
TBSRN [2]59.647.135.348.174.157.040.858.475.759.941.660.1
TG [3]61.247.635.548.975.857.841.459.477.960.242.461.3
TPGSR [18]63.152.038.651.874.960.544.160.578.962.744.562.8
TATT [17]62.653.439.852.672.660.243.159.578.963.445.463.6
C3-STISR [34]65.253.639.853.774.261.043.259.579.163.346.864.1
TSAN [37]64.653.338.853.078.461.345.162.779.664.145.364.1
LEMMA [7]67.158.840.656.377.764.444.663.281.166.347.466.0
DPMN [36]64.454.239.253.473.361.543.960.479.364.145.263.9
TCDM [20]67.357.342.755.777.662.945.962.281.365.150.165.5
RTSRN [33]67.059.242.657.077.163.346.563.280.466.149.166.2
StyleSRN68.159.442.457.478.665.145.764.182.767.448.767.3
", + "bbox": [ + 94, + 89, + 906, + 280 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d5b2ac58a64030dc4aa9b6d0267f3eea58579c0bbb46c4f227d6307ab9a2b855.jpg", + "table_caption": [ + "Table 6. Comparison with the existing methods in terms of the recognition accuracy on TextZoom[30]. BICUBIC means LR images are directly upsampled by the bicubic interpolation, and the same definition applies to BICUBIC in subsequent tables." + ], + "table_footnote": [], + "table_body": "
MethodLossPSNRSSIM
EasyMediumHardAvgEasyMediumHardAvg
BICUBIC-22.3518.9819.3920.350.78840.62540.65920.6961
TSRN [30]\\( \\mathcal{L}_2 + \\mathcal{L}_{GP} \\)25.0718.8619.7121.420.88970.66760.73020.7690
TBSRN [2]\\( \\mathcal{L}_{POS} + \\mathcal{L}_{CON} \\)23.4619.1719.6820.910.87290.64550.74520.7603
TPGSR [18]\\( \\mathcal{L}_2 + \\mathcal{L}_{TP} \\)23.7318.6820.0620.970.88050.67380.74400.7719
TATT [17]\\( \\mathcal{L}_2 + \\mathcal{L}_{TP} + \\mathcal{L}_{TSC} \\)24.7219.0220.3121.520.90060.69110.77030.7930
LEMMA [7]\\( \\mathcal{L}_2 + \\mathcal{L}_{TP} + \\mathcal{L}_{finetune} \\)23.7019.3719.8420.880.87530.69050.74910.7760
StyleSRN\\( \\mathcal{L}_2 + \\mathcal{L}_{TP} + \\mathcal{L}_{style} \\)24.6919.7020.5821.820.88440.67910.75310.7778
", + "bbox": [ + 94, + 337, + 906, + 454 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Numbers of SCFBs. We further examined the impact of the numbers of SCFBs, as detailed in Table 5. While increasing the number of SCFBs improves recognition accuracy metrics, diminishing returns are observed beyond a certain point. The optimal performance, consistent with the SEB ablation study, is achieved with 5 SCFBs, yielding a PSNR of 21.82 and an SSIM of 0.7778. Moreover, text recognition accuracy peaks with CRNN (57.40%), MORAN (64.14%), and ASTER (67.31%) when 5 SCFBs are utilized. This suggests that while the SCFB is essential for balancing style and content, its optimal configuration involves a moderate number of blocks, beyond which performance gains begin to plateau.", + "bbox": [ + 88, + 508, + 482, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Comparison with State-of-the-Arts", + "text_level": 1, + "bbox": [ + 89, + 712, + 395, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on TextZoom. We evaluate the recognition accuracy of text images generated by various STISR methods using three text recognizers: CRNN [25], MORAN [16], and ASTER [26]. As shown in Table 7, our proposed StyleSRN consistently outperforms state-of-the-art methods across all recognizers, achieving average accuracy rates of $57.4\\%$ with CRNN, $64.1\\%$ with MORAN, and $67.3\\%$ with ASTER. On the hard subset, StyleSRN performs slightly lower than some methods, such as TCDM and DPMN, likely due to challenges in balancing structure and style preservation in extremely blurred or occluded", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/64528ca3edaa26e18866c013f18ef5688dd20732b1faf36c7416b59ef515ca05.jpg", + "table_caption": [ + "Table 7. Comparison of the image quality on TextZoom [30]. Avg represents the average value of three subsets." + ], + "table_footnote": [], + "table_body": "
MethodAccuracy(%)
IC13IC15CT80SVTPIII5K
BICUBIC76.1638.2155.2135.5059.70
TSRN [30]75.6741.5254.8635.9762.60
TBSRN [2]76.9543.0754.8637.3663.60
TPGSR [18]77.6442.8553.4738.2962.07
TATT [17]76.0648.4851.0442.0262.67
LEMMA [7]79.8049.7051.0443.7265.07
StyleSRN80.9950.4755.5646.5165.63
", + "bbox": [ + 517, + 507, + 901, + 623 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 8. Comparison of the accuracy on manually degrade scene text recognition datasets. Accuracy represents the average accuracy with CRNN [25].", + "bbox": [ + 511, + 637, + 906, + 679 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cases. In terms of image quality, as shown in Table 7, our method achieves the highest average PSNR of 21.82, surpassing other methods like TATT (21.52) and LEMMA (20.93). Our method achieves an SSIM of 0.7778, which is comparable to TATT (0.7930). The visual comparison is presented in Figure 3, which demonstrates that our method achieves superior visual quality. Our method not only restores the structure of the characters but also preserves their style information. For instance, for certain stereoscopic scene texts shown in Figure 3, our method more effectively restores their stereoscopic effect and shadows details.", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18699", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4e0fd2032e03f8e5ea2b61d52496d82f99a05df52b1fa502eb7b905fe010bd94.jpg", + "image_caption": [ + "Figure 3. Visualization of SR images and their recognition result. Each text image below uses CRNN [25] for recognition. Characters marked in red indicate incorrect recognition results." + ], + "image_footnote": [], + "bbox": [ + 114, + 88, + 883, + 334 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/48408a6d7957e9693abe8ca9efcdf347a6e7fce11c38ad672c5cb1affa70758d.jpg", + "image_caption": [ + "Figure 4. Visualization of the STISR results on scene text with uneven illumination and scene text with severe blur." + ], + "image_footnote": [], + "bbox": [ + 101, + 396, + 472, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generalization to recognition datasets. To further evaluate the generalization capability and robustness of the proposed StyleSRN, we also conduct experiments on five scene text recognition datasets. These datasets cover a broad spectrum of image styles and label distributions, presenting substantial challenges for the generalization performance of super-resolution networks. Due to the absence of paired low-resolution and high-resolution images in these datasets, we manually degrad the original images before applying StyleSRN for preprocessing. As illustrated in Table 2, our method consistently surpasses previous approaches across these datasets, confirming its strong generalization ability.", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 513, + 398, + 627, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While StyleSRN effectively improves the structural integrity and style consistency of STISR, as shown in Figure 4, it struggles to recover character structures in images with uneven illumination or severe blur, highlighting the need for further refinement to address degraded visual conditions. Another limitation is the absence of a standardized quantitative metric for evaluating style degradation in STISR methods. While our method restores the style of low-resolution scene text images, no existing metric specifically measures style preservation or degradation. This lack of objective criteria makes it difficult to rigorously assess style restoration performance, as current evaluations are largely subjective and based on visual inspection. Future research should focus on developing standardized, quantitative metrics to objectively evaluate style preservation in STISR, enabling more precise assessments and meaningful comparisons across methods.", + "bbox": [ + 511, + 422, + 906, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 694, + 640, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose StyleSRN to address the limitations of current STISR methods that neglect the stylistic details of scene text images. StyleSRN incorporates Style Enhancement Blocks and Style Content Fuse Blocks to effectively capture and integrate text style with content, ensuring both structural integrity and stylistic consistency. Additionally, the proposed Text Style Loss supervises reconstruction at the style level, which effectively enhances both image quality and text recognition accuracy. Results on TextZoom and five STR benchmarks demonstrate that StyleSRN consistently enhances downstream recognition accuracy and image quality.", + "bbox": [ + 509, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18700", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 277, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China (No.61502164, 62176097), the Natural Science Foundation of Hunan Province (No.2020JJ4057, 2024JJ10034), the Key Research and Development Program of Changsha Science and Technology Bureau (No.kq2004050), and the Scientific Research Foundation of the Education Department of Hunan Province of China (No.21A0052).", + "bbox": [ + 89, + 113, + 483, + 226 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 251, + 187, + 266 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE International Conference on Computer Vision, pages 2425-2433, 2015. 1", + "[2] Jingye Chen, Bin Li, and Xiangyang Xue. Scene text telescope: Text-focused scene image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12026-12035, 2021. 1, 2, 5, 6, 7", + "[3] Jingye Chen, Haiyang Yu, Jianqi Ma, Bin Li, and Xiangyang Xue. Text gestalt: Stroke-aware scene text image superresolution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 285-293, 2022. 2, 7", + "[4] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In Proceedings of the European Conference on Computer Vision, pages 184-199, 2014. 2", + "[5] Chao Dong, Ximei Zhu, Yubin Deng, Chen Change Loy, and Yu Qiao. Boosting optical character recognition: A superresolution approach. arXiv preprint arXiv:1506.02211, 2015.2", + "[6] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2414-2423, 2016. 2, 3, 5", + "[7] Hang Guo, Tao Dai, Guanghao Meng, and Shu-Tao Xia. Towards robust scene text image super-resolution via explicit location enhancement. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 782–790, 2023. 1, 2, 6, 7", + "[8] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE International Conference on Computer Vision, pages 1501-1510, 2017. 2, 3, 4", + "[9] Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. Proceedings of the Advances in Neural Information Processing Systems, 28, 2015. 3", + "[10] Dimosthenis Karatzas, Faisal Shafait, Seiichi Uchida, Masakazu Iwamura, Lluis Gomez i Bigorda, Sergi Robles Mestre, Joan Mas, David Fernandez Mota, Jon Almazan Almazan, and Lluis Pere De Las Heras. Icdar 2013 robust reading competition. In Proceedings of the 12th International Conference on Document Analysis and Recognition, pages 1484-1493, 2013. 5" + ], + "bbox": [ + 93, + 276, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Dimosthenis Karatzas, Lluis Gomez-Bigorda, Anguelos Nicolaou, Suman Ghosh, Andrew Bagdanov, Masakazu Iwamura, Jiri Matas, Lukas Neumann, Vijay Ramaseshan Chandrasekhar, Shijian Lu, et al. Icdar 2015 competition on robust reading. In Proceedings of the 13th International Conference on Document Analysis and Recognition, pages 1156-1160, 2015. 5", + "[12] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks, 2019. 2, 3", + "[13] Vijeta Khare, Palaiahnakote Shivakumara, Chee Seng Chan, Tong Lu, Liang Kim Meng, Hon Hock Woon, and Michael Blumenstein. A novel character segmentation-reconstruction approach for license plate recognition. Expert Systems with Applications, 131:219-239, 2019. 1", + "[14] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4681-4690, 2017. 2", + "[15] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5", + "[16] Canjie Luo, Lianwen Jin, and Zenghui Sun. Moran: A multi-object rectified attention network for scene text recognition. Pattern Recognition, 90:109-118, 2019. 5, 6, 7", + "[17] Jianqi Ma, Zhetong Liang, and Lei Zhang. A text attention network for spatial deformation robust scene text image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5911-5920, 2022. 1, 2, 4, 6, 7", + "[18] Jianqi Ma, Shi Guo, and Lei Zhang. Text prior guided scene text image super-resolution. IEEE Transactions on Image Processing, 32:1341-1353, 2023. 1, 2, 6, 7", + "[19] Anand Mishra, Karteek Alahari, and CV Jawahar. Scene text recognition using higher order language priors. In Proceedings of the British Machine Vision Conference, pages 1-11, 2012. 5", + "[20] Chihiro Noguchi, Shun Fukuda, and Masao Yamanaka. Scene text image super-resolution based on text-conditional diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1474–1484, 2024. 2, 7", + "[21] Ram Krishna Pandey, K Vignesh, AG Ramakrishnan, et al. Binary document image super resolution for improved readability andOCR performance. arXiv preprint arXiv:1812.02475, 2018. 2", + "[22] Trung Quy Phan, Palaiahnakote Shivakumara, Shangxuan Tian, and Chew Lim Tan. Recognizing text with perspective distortion in natural scenes. In Proceedings of the IEEE International Conference on Computer Vision, pages 569-576, 2013. 5", + "[23] Yadong Qu, Qingfeng Tan, Hongtao Xie, Jianjun Xu, Yuxin Wang, and Yongdong Zhang. Exploring stroke-level modifications for scene text editing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2119-2127, 2023. 3" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18701", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Anhar Risnumawan, Palaiahankote Shivakumara, Chee Seng Chan, and Chew Lim Tan. A robust arbitrary text detection system for natural scene images. Expert Systems with Applications, 41(18):8027-8048, 2014. 5", + "[25] Baoguang Shi, Xiang Bai, and Cong Yao. An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(11):2298-2304, 2016. 5, 6, 7, 8", + "[26] Baoguang Shi, Mingkun Yang, Xinggang Wang, Pengyuan Lyu, Cong Yao, and Xiang Bai. Aster: An attentional scene text recognizer with flexible rectification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(9):2035-2048, 2018. 5, 6, 7", + "[27] Tonghua Su, Fuxiang Yang, Xiang Zhou, Donglin Di, Zhongjie Wang, and Songze Li. Scene style text editing, 2023. 3", + "[28] Hanh TM Tran and Tien Ho-Phuoc. Deep laplacian pyramid network for text images super-resolution. In 2019 IEEE-RIVF International Conference on Computing and Communication Technologies (RIVF), pages 1-6, 2019. 2", + "[29] Qilong Wang, Banggu Wu, Pengfei Zhu, Peihua Li, Wangmeng Zuo, and Qinghua Hu. Eca-net: Efficient channel attention for deep convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11534-11542, 2020. 4", + "[30] Wenjia Wang, Enze Xie, Xuebo Liu, Wenhai Wang, Ding Liang, Chunhua Shen, and Xiang Bai. Scene text image super-resolution in the wild. In Proceedings of the European Conference on Computer Vision, pages 650-666, 2020. 1, 2, 4, 5, 6, 7", + "[31] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5", + "[32] Chongsheng Zhang, Weiping Ding, Guowen Peng, Feifei Fu, and Wei Wang. Street view text recognition with deep learning for urban scene understanding in intelligent transportation systems. IEEE Transactions on Intelligent Transportation Systems, 22(7):4727-4743, 2020. 1", + "[33] Wenyu Zhang, Xin Deng, Baojun Jia, Xingtong Yu, Yifan Chen, Jin Ma, Qing Ding, and Xinming Zhang. Pixel adapter: A graph-based post-processing approach for scene text image super-resolution. In Proceedings of the 31st ACM International Conference on Multimedia, pages 2168-2179, 2023. 7", + "[34] Minyi Zhao, Miao Wang, Fan Bai, Bingjia Li, Jie Wang, and Shuigeng Zhou. C3-STISR: scene text image superresolution with triple clues. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, pages 1707–1713, 2022. 7", + "[35] Yuxuan Zhou, Liangcai Gao, Zhi Tang, and Baole Wei. Recognition-guided diffusion model for scene text image super-resolution. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, pages 2940-2944, 2024. 2", + "[36] Shipeng Zhu, Zuoyan Zhao, Pengfei Fang, and Hui Xue. Improving scene text image super-resolution via dual prior" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "modulation network. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3843-3851, 2023. 7", + "[37] Xiangyuan Zhu, Kehua Guo, Hui Fang, Rui Ding, Zheng Wu, and Gerald Schaefer. Gradient-based graph attention for scene text image super-resolution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3861-3869, 2023. 7" + ], + "bbox": [ + 516, + 92, + 906, + 188 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18702", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/StyleSRN_ Scene Text Image Super-Resolution with Text Style Embedding/34baee7e-bb1d-4173-bb75-506b85eb6a0f_model.json b/2025/StyleSRN_ Scene Text Image Super-Resolution with Text Style Embedding/34baee7e-bb1d-4173-bb75-506b85eb6a0f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..50c41120fe7817e46135cc9dac40f78cccdf128e --- /dev/null +++ b/2025/StyleSRN_ Scene Text Image Super-Resolution with Text Style Embedding/34baee7e-bb1d-4173-bb75-506b85eb6a0f_model.json @@ -0,0 +1,1826 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.125, + 0.131, + 0.873, + 0.154 + ], + "angle": 0, + "content": "StyleSRN: Scene Text Image Super-Resolution with Text Style Embedding" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.18, + 0.892, + 0.253 + ], + "angle": 0, + "content": "Shengrong Yuan\\(^{1}\\), Runmin Wang\\(^{1,*}\\), Ke Hao\\(^{1}\\), Xuqi Ma\\(^{1}\\), Changxin Gao\\(^{2}\\), Li Liu\\(^{3}\\), Nong Sang\\(^{2}\\) \n\\(^{1}\\)School of Information Science and Engineering, Hunan Normal University \n\\(^{2}\\)School of Artificial Intelligence and Automation, Huazhong University of Science and Technology \n\\(^{3}\\)School of Electronic Science, National University of Defense Technology" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.255, + 0.861, + 0.27 + ], + "angle": 0, + "content": "{shengrongyuan, runminwang, ke, xqma}@hunnu.edu.cn, {cgao, nsang}@hust.edu.cn, liliu@oulu.fi" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.305, + 0.327, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.337, + 0.485, + 0.715 + ], + "angle": 0, + "content": "Scene text image super-resolution (STISR) focuses on enhancing the clarity and readability of low-resolution text images. Existing methods often rely on text probability distribution priors derived from text recognizers to guide the super-resolution process. While effective in capturing general structural information of text, these priors lack the ability to preserve specific text style details, such as font, stereoscopic effect and spatial transformation, leading to a loss of visual quality and stylistic consistency. To address these limitations, we propose a Style embedding-based scene text image Super-Resolution Network (StyleSRN), which introduces a text style embedding mechanism to preserve and enhance text style features during the super-resolution process. The proposed architecture includes Style Enhancement Blocks for capturing multi-scale cross-channel dependencies, and Style Content Fusion Blocks that effectively integrates text content with style information, ensuring that the structure and style of the restored text are not distorted. Furthermore, we introduce a Text Style Loss based on the Gram matrix to supervise the reconstruction process at the style level, thereby maintaining the stylistic consistency of the restored text images. Extensive experiments on the TextZoom dataset and five scene text recognition benchmarks demonstrate the superiority of our method. Code is available at: https://github.com/Yuanssr/StyleSRN." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.744, + 0.222, + 0.76 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.77, + 0.483, + 0.877 + ], + "angle": 0, + "content": "Scene text images contain both visual and linguistic modalities, which can be extracted and analyzed to enhance various applications, including autonomous driving [32], visual question answering [1], and card recognition [13], etc. However, due to the limitations of imaging conditions of sensing devices, such as shooting jitter and low-focus cameras, scene text images often suffer from varying degrad" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.313, + 0.892, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.482, + 0.908, + 0.552 + ], + "angle": 0, + "content": "Figure 1. Comparison of super-resolution results between previous methods and ours. Methods without text priors such as TSRN [30] retains less style information, while text priors-guided methods such as TATT [17] and LEMMA [7] almost completely lose the style information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.907, + 0.718 + ], + "angle": 0, + "content": "tion, including blur, distortion, etc. Unlike general object images, scene text images possess more complex character structures and fine-grained stroke-level details, making even slight degradation challenging to recognize. Such low-resolution scene text images significantly hinder the performance of downstream tasks such as scene text detection, optical character recognition, and scene text recognition. Thus, restoring both the visual style quality and character structures of these degraded images is crucial." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "In recent years, many super-resolution methods specifically designed for scene text images have used the unique properties of text to improve restoration. Early approaches primarily focused on extracting sequential information from text images. Wang et al. [30] proposed a sequential residual block to capture the inherent sequential nature of text images. Chen et al. [2] introduced a text-focus loss function that emphasizes character positions and content, thereby improving recognition accuracy. More recently, methods that incorporate deep character classification information obtained from text recognizers have become the norm. Ma et al. [18] highlighted the benefits of integrating text" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.235, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18693" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "prriors into the super-resolution process. Moreover, Ma et al. [17] developed an attention-based text priors interpreter to interactively fuse text probability distributions with image features. Guo et al. [7] demonstrated the effectiveness of using a bidirectional alignment strategy for text and images to obtain high-level guidance information. Most of these methods rely on text recognizers to generate text probability distribution priors, which are then encoded into the super-resolution network to guide the recovery of character structures." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.248, + 0.485, + 0.626 + ], + "angle": 0, + "content": "However, the text probability distribution priors derived from text recognizers mainly capture the general character structures but are less sensitive to the variations in text styles across different images, which hampers the super-resolution process. This loss of style details not only affects subsequent scene text recognition tasks but also disrupts the visual consistency of the images. The problem is particularly pronounced in scene text images with distinctive text styles, such as artistic characters or logos, where the loss of stylistic features can lead to more severe distortion in the restored images. As shown in Figure 1, text priors-guided methods such as TATT [17] and LEMMA [7] almost completely lose the style information of the degraded text image, while methods without text priors such as TSRN [30] retain less style information. To address these challenges, we draw inspiration from Style Transfer [6, 8, 12] to explore the potential of preserving image style information to enhance the restoration of low-resolution text images. Unlike content information, which is sensitive to position, style information in an image can be viewed as the correlation between different feature channels and is largely position-insensitive. By effectively encoding and embedding this style information during the super-resolution process, our proposed method aims to preserve both the structural integrity and the stylistic consistency of the restored text images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a Style embedding-based scene text image Super-Resolution Network (StyleSRN) for STISR, which not only restores the character structure but also preserves the stylistic consistency of the text. Unlike conventional methods that rely solely on text probability priors to guide the super-resolution process, our approach introduces a Text Style Embedding Branch that extracts and enhances style features. Additionally, we design a Style Enhancement Block (SEB), which captures multi-scale cross-channel dependencies, thereby enhancing the representation of style information. To further ensure the integration of style and content, we introduce a Style Content Fuse Block (SCFB) that effectively combines style embeddings with text priors. Finally, we propose a Text Style Loss based on the Gram matrix to supervise the reconstructed images at the style level. As illustrated in Figure 1, our method achieves superior visual quality and stylistic consistency in the restored images. Overall, our contributions can be sum" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.093, + 0.646, + 0.106 + ], + "angle": 0, + "content": "marized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.11, + 0.905, + 0.184 + ], + "angle": 0, + "content": "- We, for the first time to our best knowledge, successfully embed text style information to guide super-resolution process for STISR task, and validate its effectiveness to ensure both structural integrity and stylistic consistency of the super-resolution process." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.186, + 0.905, + 0.261 + ], + "angle": 0, + "content": "- We propose a Text Style Loss based on the Gram Matrix which leverages a text recognizer to extract features to offer robust supervision at style level for scene text images, which effectively improve both image quality and text recognition accuracy." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.261, + 0.905, + 0.335 + ], + "angle": 0, + "content": "- Extensive experiment results demonstrate that the proposed method not only achieves state-of-the-art performance on the TextZoom dataset but also validates exceptional generalization across five scene text recognition datasets." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.11, + 0.905, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.353, + 0.655, + 0.369 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.378, + 0.816, + 0.394 + ], + "angle": 0, + "content": "2.1. Scene Text Image Super-resolution" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.4, + 0.905, + 0.839 + ], + "angle": 0, + "content": "Scene text image super-resolution aims to enhance the visual quality and readability of scene text images. Early studies in this field usually adopt general image super-resolution methods [4, 5, 14, 21, 28] to reconstruct scene text images. However, due to the complex character structure of text images, these methods often fail to produce sufficiently recognizable text images. Therefore, some unique attributes of scene text have been exploited to improve super-resolution performance[2, 3, 7, 17, 18, 30]. For example, some methods focus on the sequential information of scene text. Wang et al. [30] introduced sequential residual block to capture the sequential information in text images. Chen et al. [2] proposed a Transformer-Based SuperResolution Network containing a Self-Attention Module to extract sequential information. Recently, the method of using text recognizers to obtain text probability distribution priors has become mainstream. For example, Ma et al. [18] incorporates categorical text priors into the STISR model training process, while Ma et al. [17] proposed a text prior interpreter to better transform text priors. Guo et al. [7] proposed a bimodal alignment mechanism to better utilize text priors. Additionally, some works use text priors to guide diffusion models to restore low-resolution images [20, 35]. Most of these methods use a pre-trained text recognizer to obtain deep character classification information to guide the super-resolution network. While this approach enables the model to effectively learn the general character structure, it results in the loss of specific style information in each image." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.848, + 0.661, + 0.864 + ], + "angle": 0, + "content": "2.2. Style Transfer" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Style transfer is a technique that involves altering the stylistic attributes of an image while maintaining its core con" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18694" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.09, + 0.891, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.373, + 0.908, + 0.418 + ], + "angle": 0, + "content": "Figure 2. Overview architecture of the proposed StyleSRN. GAP denotes the Global Average Pooling. \\( G(\\cdot) \\) denotes the Gram matrix, \\( \\Phi_l(\\cdot) \\) represents the feature map obtained from the \\( l \\)-th layer of the text recognizer. \\( \\mathbf{I}_{\\mathrm{LR}} \\), \\( \\mathbf{I}_{\\mathrm{SR}} \\) and \\( \\mathbf{I}_{\\mathrm{HR}} \\) represent low-resolution images, super-resolution images and high-resolution images, respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.442, + 0.487, + 0.685 + ], + "angle": 0, + "content": "tent. This concept was pioneered by Gatys et al. [6], who utilized convolutional neural networks (CNNs) to transfer artistic styles between images by matching the Gram matrices of feature maps. Subsequent work has expanded on this foundation, exploring various methods to improve both the quality and efficiency of style transfer. One significant advancement is the introduction of adaptive instance normalization (AdaIN) [8], which allows for real-time arbitrary style transfer by aligning the mean and variance of content features to those of style features. Some works employ style transfer technology specifically for scene text style transfer [23, 27]. In our work, we draw inspiration from these style transfer methods, particularly in how they encode and embed style features. By adopting these principles, we aim to restore text style information, which is crucial for maintaining the visual consistency of the super-resolved text images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.701, + 0.226, + 0.718 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.288, + 0.74 + ], + "angle": 0, + "content": "3.1. Overall Architecture" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.485, + 0.901 + ], + "angle": 0, + "content": "The overall architecture of the proposed StyleSRN is illustrated in Figure 2. Our method consists of a text style embedding branch for extracting and enhancing style features and a super-resolution branch for restoring low-resolution images \\(\\mathbf{I}_{\\mathrm{LR}}\\). For the text style embedding branch, the low-resolution image will first pass through ResNet18 to obtain the feature map \\(\\mathbf{X}_f \\in \\mathbb{R}^{C \\times W \\times H}\\), where \\(C\\) represents the number of channels, and \\(H\\) and \\(W\\) represent the height and width of the feature map, respectively. Then several stacked Style Enhancement Blocks will extract the correla" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.442, + 0.909, + 0.73 + ], + "angle": 0, + "content": "tion information between the feature map channels. After the style enhancement, we perform global average pooling on it to obtain the latent vector \\(\\mathbf{z}\\). Given a latent vector \\(\\mathbf{z}\\) in the input latent space \\(\\mathcal{Z}\\), a non-linear mapping network \\(f: \\mathcal{Z} \\to \\mathcal{W}\\) first produces \\(\\mathbf{w} \\in \\mathcal{W}\\). Learned affine transformations then specialize \\(\\mathbf{w}\\) to styles \\(\\mathbf{y} = (\\mathbf{y}_s, \\mathbf{y}_b)\\), where \\(\\mathbf{y}_s\\) and \\(\\mathbf{y}_b\\) denote the learned normalization scaling and bias coefficients, respectively. For the Super-Resolution Branch, the low-resolution image will first be processed by the Spatial Transformer Network (STN) [9] to align the spatially deformed texts. The convolution layer will extract features from the image processed by STN to obtain shallow image features. Then the shallow image features, text priors, and style information will be sent to several stacked style-content fusion blocks for style and content fusion. Each block contains a Style Enhancement Block, an adaptive instance normalization (AdaIN) [8], and a content embedding block. Finally, PixelShuffle is applied as an upsampling operator to generate \\(\\mathbf{I}_{\\mathrm{SR}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.741, + 0.748, + 0.756 + ], + "angle": 0, + "content": "3.2. Style Enhancement Block" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.907, + 0.884 + ], + "angle": 0, + "content": "The style of an image has been traditionally viewed as the correlation between different features, as demonstrated in the advancements of style transfer [6, 8, 12]. While previous approaches, such as [12], directly utilized ResNet features to obtain style vectors, our approach introduces a novel enhancement by applying channel-wise attention to these feature maps, thereby better capturing and preserving the stylistic characteristics in scene text images." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Inspired by the Efficient Channel Attention (ECA)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18695" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.289 + ], + "angle": 0, + "content": "mechanism [29], we designed a Style Enhancement Block (SEB) to improve the model's capacity for capturing the correlation between different channel features. Let the input feature map be denoted as \\(\\mathbf{X}_f\\in \\mathbb{R}^{C\\times H\\times W}\\). We begin by applying Global Average Pooling (GAP) to the input feature map to obtain channel-wise statistics \\(\\mathbf{m}\\in \\mathbb{R}^C\\). To further improve the model's capacity for capturing local cross-channel interactions at multiple scales, we employ multiple 1D convolution kernels of varying sizes, in contrast to the fixed-size kernels used in [29]. Specifically, let \\(n\\) denote the number of different kernel sizes, each represented by \\(k_{l}\\), where \\(l\\in 1,2,\\ldots ,n\\). The convolutional operation at each scale can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.301, + 0.484, + 0.356 + ], + "angle": 0, + "content": "\\[\n\\mathbf {m} _ {c} ^ {\\prime (l)} = \\sum_ {j = - \\frac {k _ {l}}{2}} ^ {\\frac {k _ {l}}{2}} b _ {j} ^ {l} \\mathbf {m} _ {c + j}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.367, + 0.484, + 0.475 + ], + "angle": 0, + "content": "where \\(\\mathbf{b}^l\\in \\mathbb{R}^{k_l}\\) represents the weight vector for the convolution kernel of size \\(k_{l}\\). This multi-scale approach allows the model to capture a more nuanced set of cross-channel dependencies, enhancing the representation of style information. The outputs from the multiple convolutional scales are aggregated by computing the final channel attention weight \\(\\mathbf{m}'_c\\) as a weighted sum:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.486, + 0.484, + 0.527 + ], + "angle": 0, + "content": "\\[\n\\mathbf {m} _ {c} ^ {\\prime} = \\sum_ {l = 1} ^ {n} \\alpha_ {l} \\mathbf {m} _ {c} ^ {\\prime (l)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.484, + 0.584 + ], + "angle": 0, + "content": "where \\(\\alpha_{l}\\) denotes the weighting coefficient for the output of the \\(l\\)-th convolutional scale. These weights are normalized using a Sigmoid function:" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.596, + 0.483, + 0.619 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {f} ^ {\\prime c, i, j} = \\sigma \\left(\\mathbf {m} _ {c} ^ {\\prime}\\right) \\cdot \\mathbf {X} _ {f} ^ {c, i, j}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.483, + 0.673 + ], + "angle": 0, + "content": "where \\(\\sigma (\\cdot)\\) represents the Sigmoid activation function, and \\(\\mathbf{X}^{\\prime}\\) is the final output feature map with the applied channelwise attention." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.903 + ], + "angle": 0, + "content": "The enhanced feature map is passed through several SEBs, which refines the correlation between different feature channels, ensuring that the model pays more attention to the features of important channels. This process effectively enhances the model's capability to capture and preserve the intricate stylistic details in scene text images. Subsequently, global average pooling (GAP) is applied to the final enhanced feature map to obtain a latent vector \\(\\mathbf{z}\\) in the latent space \\(\\mathcal{Z}\\). The latent vector is then processed by a non-linear style mapping network \\(f: \\mathcal{Z} \\to \\mathcal{W}\\), which produces a vector \\(\\mathbf{w} \\in \\mathcal{W}\\) that decouples the latent space. The style mapping network consists of an 8-layer MLP with LeakyReLU activation functions. Finally, learned affine transformations are applied to transform \\(\\mathbf{w}\\) into style parameters \\(\\mathbf{y} = (\\mathbf{y}_s, \\mathbf{y}_b)\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.745, + 0.107 + ], + "angle": 0, + "content": "3.3. Style Content Fuse Block" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.907, + 0.265 + ], + "angle": 0, + "content": "In the field of scene text image super-resolution, existing approaches primarily focus on embedding content information, including deep character category information derived from text probability priors. While this approach effectively guides the super-resolution network in restoring character structure, it often results in the loss of crucial stylistic features. To address this limitation, we propose a novel Style Content Fuse Block (SCFB) that allows the super-resolution network to restore character structures while preserving the stylistic consistency of the characters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.266, + 0.909, + 0.448 + ], + "angle": 0, + "content": "Our innovation lies in embedding style information directly into the super-resolution network alongside content information. This dual embedding approach ensures that both character structure and style are preserved during the super-resolution process. Specifically, for the shallow image features \\(\\mathbf{X}_s\\in \\mathbb{R}^{C'\\times H'\\times W'}\\) obtained through a Spatial Transformer Network (STN) and a convolutional layer, we use Adaptive Instance Normalization (AdaIN) [8] to modulate their style. AdaIN has been shown to be effective in transferring style information across different domains. Given a feature map \\(\\mathbf{X}_s\\) and a style vector \\(\\mathbf{y}\\), the AdaIN operation can be defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.458, + 0.907, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A d a I N} \\left(\\mathbf {X} _ {s}, \\mathbf {y}\\right) = \\mathbf {y} _ {s} \\left(\\frac {\\mathbf {X} _ {s} - \\mu \\left(\\mathbf {X} _ {s}\\right)}{\\sigma \\left(\\mathbf {X} _ {s}\\right)}\\right) + \\mathbf {y} _ {b}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.504, + 0.907, + 0.61 + ], + "angle": 0, + "content": "where \\(\\mu (\\cdot)\\) and \\(\\sigma (\\cdot)\\) represent the mean and variance of \\(\\mathbf{X}_s\\). Each AdaIN layer within the SCFB normalizes feature maps independently, with the dimensionality of \\(\\mathbf{y}\\) being twice the number of feature map channels \\(C^\\prime\\) in that layer. This embedding ensures that style information is seamlessly integrated into the spatial features of the image, thus preserving the character's stylistic attributes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.909, + 0.763 + ], + "angle": 0, + "content": "In addition to embedding style information, content information, such as text probability distribution priors, should also be embedded into the super-resolution network to guide the learning of character structures. To bridge the domain gap between text priors and image features, we adopt the TP Interpreter architecture from [17] to extract the content embedding features. These interpreted features are then concatenated with the style-embedded features and subsequently passed through a Sequential Residual Block [30] for further refinement." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.772, + 0.667, + 0.789 + ], + "angle": 0, + "content": "3.4. Text Style Loss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Although the proposed StyleSRN network effectively embeds the text's style features, these features are derived from the low-resolution image itself. When the style information in the low-resolution image is degraded, even if the network extracts these style features, it cannot fully reconstruct the style of the corresponding high-resolution image. Therefore, the reconstructed text image requires supervision at" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18696" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "the style level to effectively learn the degradation of style information in the low-resolution image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.484, + 0.228 + ], + "angle": 0, + "content": "To supervise the style in the reconstructed image, we introduce a Text Style Loss based on the Gram matrix. Unlike traditional methods [6] that use VGG networks to capture style information, our approach leverages a text recognizer to extract feature maps for Gram matrix computation. This enables better capture of text-specific style features, which are crucial for scene text images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.228, + 0.484, + 0.336 + ], + "angle": 0, + "content": "The Gram matrix \\(\\mathbf{G}_l\\) for a given feature map \\(\\Phi_l(\\mathbf{I})\\) at the \\(l\\)-th layer is computed as follows. First, we reshape the feature map \\(\\Phi_l(\\mathbf{I})\\) of dimensions \\(C_l \\times H_l \\times W_l\\) into a matrix \\(\\mathbf{F}_l\\) of size \\(C_l \\times (H_l \\times W_l)\\), where each row corresponds to a channel and each column corresponds to a pixel. The Gram matrix \\(\\mathbf{G}_l\\) is the inner product between the different channels in layer \\(l\\):" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.345, + 0.483, + 0.387 + ], + "angle": 0, + "content": "\\[\n\\mathbf {G} _ {i j} ^ {l} = \\sum_ {k = 1} ^ {H _ {l} \\times W _ {l}} \\mathbf {F} _ {i k} ^ {l} \\mathbf {F} _ {j k} ^ {l} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.484, + 0.444 + ], + "angle": 0, + "content": "The Gram matrix captures the correlations between different channels in the feature map, thereby encoding the style information of the image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.444, + 0.484, + 0.505 + ], + "angle": 0, + "content": "The Text Style Loss \\(\\mathcal{L}_{\\mathrm{style}}\\) is defined as the Mean Squared Error (MSE) between the Gram matrix of the high-resolution (HR) ground truth and the super-resolved (SR) output:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.516, + 0.484, + 0.569 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s t y l e}} = \\sum_ {l = 1} ^ {L} \\frac {1}{4 C _ {l} ^ {2} H _ {l} ^ {2} W _ {l} ^ {2}} \\| \\mathbf {G} \\left(\\Phi_ {l} \\left(\\mathbf {I} _ {\\mathrm {H R}}\\right)\\right) - \\mathbf {G} \\left(\\Phi_ {l} \\left(\\mathbf {I} _ {\\mathrm {S R}}\\right)\\right) \\| _ {2} ^ {2} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.484, + 0.631 + ], + "angle": 0, + "content": "where \\(\\mathbf{G}(\\cdot)\\) denotes the Gram matrix, \\(\\Phi_l(\\cdot)\\) represents the feature map obtained from the \\(l\\)-th layer of the text recognizer, and \\(C_l, H_l, W_l\\) are the dimensions of the feature map at layer \\(l\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.64, + 0.3, + 0.656 + ], + "angle": 0, + "content": "3.5. Overall Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.663, + 0.484, + 0.753 + ], + "angle": 0, + "content": "In training, the overall loss function includes a Mean Square Error (MSE) loss \\(\\mathcal{L}_2\\), a Text-Focus Loss \\(\\mathcal{L}_{TP}\\) [30], and the proposed Text Style Loss \\(\\mathcal{L}_{style}\\). The \\(\\mathcal{L}_2\\) measures the difference between the super-resolution (SR) output and the ground-truth high-resolution (HR) image. The overall loss function is described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.766, + 0.483, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {2} + \\alpha \\mathcal {L} _ {T P} + \\beta \\mathcal {L} _ {\\text {s t y l e}} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.795, + 0.414, + 0.811 + ], + "angle": 0, + "content": "where the \\(\\alpha\\) and \\(\\beta\\) are the balancing parameters." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.824, + 0.224, + 0.84 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.194, + 0.863 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "TextZoom TextZoom [30] is a scene text super-resolution dataset designed for real-world scenarios, containing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.183 + ], + "angle": 0, + "content": "21,740 pairs of low-resolution and high-resolution images captured by cameras with various focal lengths in natural environments, with 17,367 samples used for training. The test set is divided into easy, medium, and hard subsets, comprising 1,619, 1,411, and 1,343 LR-HR pairs, respectively, based on the camera focal lengths." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.204, + 0.907, + 0.326 + ], + "angle": 0, + "content": "Scene Text Recognition Datasets To assess the robustness and generalization capability of StyleSRN, we conducted comprehensive evaluations on five benchmark scene text recognition datasets: IC13 [10], IC15 [11], CUTE80 [24], IIIT5K [19], and SVTP [22]. These datasets were manually degraded to create low-resolution counterparts. Detailed descriptions of the datasets and degradation procedures are provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.336, + 0.604, + 0.351 + ], + "angle": 0, + "content": "4.2. Metric" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.359, + 0.906, + 0.48 + ], + "angle": 0, + "content": "Following previous works, we adopt two types of metrics: super-resolution image text recognition accuracy and image quality. Specifically, to evaluate the generalization capability of StyleSRN, we employed three scene text recognizers: CRNN [25], MORAN [16], and ASTER [26]. For the image quality metric, we adopt the widely used Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM)[31]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.491, + 0.731, + 0.507 + ], + "angle": 0, + "content": "4.3. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.513, + 0.907, + 0.696 + ], + "angle": 0, + "content": "We implement our model using PyTorch. All experiments are conducted on a single RTX 3090 GPU. We train our model with a batch size of 64 for 500 epochs using AdamW [15] for optimization. The learning rate is set to 1e-3 and decayed by a factor of 0.5 after 400 epochs. The number of SEBs and SCFBs are set to 3 and 5, respectively. In the SEBs, the sizes of the convolution kernels are \\(3 \\times 3\\), \\(5 \\times 5\\), and \\(7 \\times 7\\), respectively. We use CRNN to generate the features for computing the text style loss and empirically set the number of feature maps L used to calculate the style loss to 5. We set the parameters \\(\\alpha\\) and \\(\\beta\\) to 0.5 and 1.0, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.706, + 0.667, + 0.722 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.728, + 0.906, + 0.835 + ], + "angle": 0, + "content": "In this section, we conduct ablation studies to investigate the effectiveness of SEBs and SCFBs, the effectiveness of training with Text Style Loss, the impact of numbers of SEBs and SCFBs. For STISR methods that do not use text priors, such as TSRN [30] and TBSRN [2], using SCFB means that image features are not used for text prior fusion, but are directly concatenated with style-embedded features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Effectiveness of SEBs and SCFBs. We compared the performance of different STISR methods using three strategies: No Style Embedding (NSE), only SCFB, and SCFB" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18697" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.096, + 0.09, + 0.48, + 0.294 + ], + "angle": 0, + "content": "
MethodNSESCFBSEBImage QualityAccuracy(%)
PSNRSSIM
TSRN [30]✓--21.420.769041.4
-✓-21.540.769943.5
-✓✓21.560.769544.4
TBSRN [2]✓--20.910.760348.1
-✓-21.040.762350.3
-✓✓21.220.761451.1
TPGSR [18]✓--20.970.771951.8
-✓-21.040.772352.4
-✓✓21.170.772352.7
TATT [17]✓--21.520.793052.6
-✓-21.660.779054.1
-✓✓21.870.778554.9
LEMMA [7]✓--20.880.776056.3
-✓-20.940.777156.7
-✓✓21.050.776557.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.307, + 0.483, + 0.364 + ], + "angle": 0, + "content": "Table 1. Effectiveness of the Style Content Fuse Block (SCFB) and the Style Enhancement Block (SEB). NSE refers to No Style Embedding. Accuracy represents the average accuracy with CRNN [25]." + }, + { + "type": "table", + "bbox": [ + 0.096, + 0.39, + 0.481, + 0.504 + ], + "angle": 0, + "content": "
Method\\( \\mathcal{L}_{\\text{style}} \\)PSNRSSIMAccuracy(%)
TSRN [30]-21.420.769041.4
✓21.560.771143.7
TPGSR [18]-20.970.771951.8
✓21.120.772452.4
TATT [17]-21.520.793052.6
✓21.810.793853.8
StyleSRN-21.540.775854.9
✓21.820.777857.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.517, + 0.483, + 0.547 + ], + "angle": 0, + "content": "Table 2. Ablation of the Text Style Loss. Accuracy represents the average accuracy with CRNN [25]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.583, + 0.483, + 0.901 + ], + "angle": 0, + "content": "combined with the SEB in Table 2. NSE means removing both SEB and SCFB. SCFB can be used independently of SEB, but when using SEB, SCFB must be used to embed style information. Removing SEB means that the features extracted by Resnet18 are directly fed into the style mapping network through global pooling. Removing SCFB means that the features directly passed through the text interpreter do not need to be cascaded with the style features. As presented in Table 1, using only SCFB generally improves image quality and recognition accuracy over NSE. In addition, SCFB has a greater improvement over methods without text priors such as TSRN [30] and TBSRN [2]. For instance, with TSRN, the introduction of SCFB increases accuracy by \\(2.1\\%\\), along with notable gains in PSNR and SSIM. We speculate that this may be because these methods do not have any prior guidance, while SCFB introduces the guidance of style information. Using SCFB and SEB at the same time can effectively improve PSNR and text recognition accuracy, but slightly reduce SSIM. This shows that SEB can effectively enhance style information, but slightly reduce structural consistency." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.09, + 0.905, + 0.132 + ], + "angle": 0, + "content": "
DiscriminatorPSNRSSIMAccuracy (%)
VGG21.310.776156.3
CRNN21.820.777857.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.53, + 0.145, + 0.89, + 0.16 + ], + "angle": 0, + "content": "Table 3. Ablation of the discriminator used in text style loss." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.181, + 0.905, + 0.287 + ], + "angle": 0, + "content": "
NQuality MetricAccuracy(%)
PSNRSSIMCRNN [25]MORAN [16]ASTER [26]
020.760.761854.661.865.0
121.820.777856.363.366.2
221.770.776456.964.066.7
421.770.777057.164.067.0
521.730.776855.963.266.7
321.820.777857.464.167.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.522, + 0.3, + 0.896, + 0.314 + ], + "angle": 0, + "content": "Table 4. Ablation of the number of Style Enhancement Blocks." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.336, + 0.905, + 0.442 + ], + "angle": 0, + "content": "
NQuality MetricAccuracy(%)
PSNRSSIMCRNN [25]MORAN [16]ASTER [26]
020.670.743153.261.064.7
321.450.772255.362.366.2
421.430.775156.763.966.9
621.740.776956.964.067.0
721.550.775656.062.866.2
521.820.777857.464.167.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.522, + 0.454, + 0.895, + 0.468 + ], + "angle": 0, + "content": "Table 5. Ablation of the number of Style Content Fuse Blocks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.502, + 0.907, + 0.699 + ], + "angle": 0, + "content": "Effectiveness of training with Text Style Loss. We show the effectiveness of the Text Style Loss in Table 2. Incorporating the Text Style Loss results in improvements across all metrics. For instance, the PSNR of TSRN increases from 21.42 to 21.56, and its SSIM improves from 0.7690 to 0.7711. In terms of text recognition accuracy, models incorporating the Text Style Loss demonstrate enhanced performance. The accuracy of TSRN improves from \\(41.4\\%\\) to \\(43.7\\%\\). These results highlight that the Text Style Loss effectively enhances both image quality and text recognition accuracy. Besides, we also compared the impact of using different discriminators in Table 3. It can be seen that using CRNN performs better than VGG." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Numbers of SEBs. We show the effect of the number of SEBs in Table 4. As the number of SEBs increases, there is a noticeable improvement in both the PSNR and SSIM metrics. Specifically, using 3 SEBs achieves the highest PSNR of 21.82 and an SSIM of 0.7778, indicating the optimal enhancement of image quality. In terms of text recognition accuracy, models incorporating 3 SEBs also demonstrate superior performance across all recognition models. These findings suggest that SEBs significantly enhance both image quality and text recognition performance, with 3 SEBs being the optimal configuration." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18698" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.09, + 0.907, + 0.281 + ], + "angle": 0, + "content": "
MethodAccuracy of CRNN [25] (%)Accuracy of MORAN [16] (%)Accuracy of ASTER [26] (%)
EasyMediumHardAvgEasyMediumHardAvgEasyMediumhardAvg
BICUBIC36.421.121.126.860.637.930.844.167.442.431.248.2
HR76.475.164.672.491.285.374.284.194.287.776.286.6
TSRN [30]52.538.231.441.470.155.337.955.475.156.340.158.3
TBSRN [2]59.647.135.348.174.157.040.858.475.759.941.660.1
TG [3]61.247.635.548.975.857.841.459.477.960.242.461.3
TPGSR [18]63.152.038.651.874.960.544.160.578.962.744.562.8
TATT [17]62.653.439.852.672.660.243.159.578.963.445.463.6
C3-STISR [34]65.253.639.853.774.261.043.259.579.163.346.864.1
TSAN [37]64.653.338.853.078.461.345.162.779.664.145.364.1
LEMMA [7]67.158.840.656.377.764.444.663.281.166.347.466.0
DPMN [36]64.454.239.253.473.361.543.960.479.364.145.263.9
TCDM [20]67.357.342.755.777.662.945.962.281.365.150.165.5
RTSRN [33]67.059.242.657.077.163.346.563.280.466.149.166.2
StyleSRN68.159.442.457.478.665.145.764.182.767.448.767.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.294, + 0.907, + 0.325 + ], + "angle": 0, + "content": "Table 6. Comparison with the existing methods in terms of the recognition accuracy on TextZoom[30]. BICUBIC means LR images are directly upsampled by the bicubic interpolation, and the same definition applies to BICUBIC in subsequent tables." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.338, + 0.907, + 0.455 + ], + "angle": 0, + "content": "
MethodLossPSNRSSIM
EasyMediumHardAvgEasyMediumHardAvg
BICUBIC-22.3518.9819.3920.350.78840.62540.65920.6961
TSRN [30]\\( \\mathcal{L}_2 + \\mathcal{L}_{GP} \\)25.0718.8619.7121.420.88970.66760.73020.7690
TBSRN [2]\\( \\mathcal{L}_{POS} + \\mathcal{L}_{CON} \\)23.4619.1719.6820.910.87290.64550.74520.7603
TPGSR [18]\\( \\mathcal{L}_2 + \\mathcal{L}_{TP} \\)23.7318.6820.0620.970.88050.67380.74400.7719
TATT [17]\\( \\mathcal{L}_2 + \\mathcal{L}_{TP} + \\mathcal{L}_{TSC} \\)24.7219.0220.3121.520.90060.69110.77030.7930
LEMMA [7]\\( \\mathcal{L}_2 + \\mathcal{L}_{TP} + \\mathcal{L}_{finetune} \\)23.7019.3719.8420.880.87530.69050.74910.7760
StyleSRN\\( \\mathcal{L}_2 + \\mathcal{L}_{TP} + \\mathcal{L}_{style} \\)24.6919.7020.5821.820.88440.67910.75310.7778
" + }, + { + "type": "table_caption", + "bbox": [ + 0.168, + 0.468, + 0.828, + 0.483 + ], + "angle": 0, + "content": "Table 7. Comparison of the image quality on TextZoom [30]. Avg represents the average value of three subsets." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.509, + 0.483, + 0.706 + ], + "angle": 0, + "content": "Numbers of SCFBs. We further examined the impact of the numbers of SCFBs, as detailed in Table 5. While increasing the number of SCFBs improves recognition accuracy metrics, diminishing returns are observed beyond a certain point. The optimal performance, consistent with the SEB ablation study, is achieved with 5 SCFBs, yielding a PSNR of 21.82 and an SSIM of 0.7778. Moreover, text recognition accuracy peaks with CRNN (57.40%), MORAN (64.14%), and ASTER (67.31%) when 5 SCFBs are utilized. This suggests that while the SCFB is essential for balancing style and content, its optimal configuration involves a moderate number of blocks, beyond which performance gains begin to plateau." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.713, + 0.396, + 0.729 + ], + "angle": 0, + "content": "4.5. Comparison with State-of-the-Arts" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Results on TextZoom. We evaluate the recognition accuracy of text images generated by various STISR methods using three text recognizers: CRNN [25], MORAN [16], and ASTER [26]. As shown in Table 7, our proposed StyleSRN consistently outperforms state-of-the-art methods across all recognizers, achieving average accuracy rates of \\(57.4\\%\\) with CRNN, \\(64.1\\%\\) with MORAN, and \\(67.3\\%\\) with ASTER. On the hard subset, StyleSRN performs slightly lower than some methods, such as TCDM and DPMN, likely due to challenges in balancing structure and style preservation in extremely blurred or occluded" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.508, + 0.902, + 0.624 + ], + "angle": 0, + "content": "
MethodAccuracy(%)
IC13IC15CT80SVTPIII5K
BICUBIC76.1638.2155.2135.5059.70
TSRN [30]75.6741.5254.8635.9762.60
TBSRN [2]76.9543.0754.8637.3663.60
TPGSR [18]77.6442.8553.4738.2962.07
TATT [17]76.0648.4851.0442.0262.67
LEMMA [7]79.8049.7051.0443.7265.07
StyleSRN80.9950.4755.5646.5165.63
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.638, + 0.907, + 0.68 + ], + "angle": 0, + "content": "Table 8. Comparison of the accuracy on manually degrade scene text recognition datasets. Accuracy represents the average accuracy with CRNN [25]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "cases. In terms of image quality, as shown in Table 7, our method achieves the highest average PSNR of 21.82, surpassing other methods like TATT (21.52) and LEMMA (20.93). Our method achieves an SSIM of 0.7778, which is comparable to TATT (0.7930). The visual comparison is presented in Figure 3, which demonstrates that our method achieves superior visual quality. Our method not only restores the structure of the characters but also preserves their style information. For instance, for certain stereoscopic scene texts shown in Figure 3, our method more effectively restores their stereoscopic effect and shadows details." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18699" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.116, + 0.089, + 0.885, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.345, + 0.908, + 0.374 + ], + "angle": 0, + "content": "Figure 3. Visualization of SR images and their recognition result. Each text image below uses CRNN [25] for recognition. Characters marked in red indicate incorrect recognition results." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.397, + 0.473, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.62, + 0.483, + 0.647 + ], + "angle": 0, + "content": "Figure 4. Visualization of the STISR results on scene text with uneven illumination and scene text with severe blur." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Generalization to recognition datasets. To further evaluate the generalization capability and robustness of the proposed StyleSRN, we also conduct experiments on five scene text recognition datasets. These datasets cover a broad spectrum of image styles and label distributions, presenting substantial challenges for the generalization performance of super-resolution networks. Due to the absence of paired low-resolution and high-resolution images in these datasets, we manually degrad the original images before applying StyleSRN for preprocessing. As illustrated in Table 2, our method consistently surpasses previous approaches across these datasets, confirming its strong generalization ability." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.399, + 0.628, + 0.414 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.907, + 0.68 + ], + "angle": 0, + "content": "While StyleSRN effectively improves the structural integrity and style consistency of STISR, as shown in Figure 4, it struggles to recover character structures in images with uneven illumination or severe blur, highlighting the need for further refinement to address degraded visual conditions. Another limitation is the absence of a standardized quantitative metric for evaluating style degradation in STISR methods. While our method restores the style of low-resolution scene text images, no existing metric specifically measures style preservation or degradation. This lack of objective criteria makes it difficult to rigorously assess style restoration performance, as current evaluations are largely subjective and based on visual inspection. Future research should focus on developing standardized, quantitative metrics to objectively evaluate style preservation in STISR, enabling more precise assessments and meaningful comparisons across methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.641, + 0.71 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose StyleSRN to address the limitations of current STISR methods that neglect the stylistic details of scene text images. StyleSRN incorporates Style Enhancement Blocks and Style Content Fuse Blocks to effectively capture and integrate text style with content, ensuring both structural integrity and stylistic consistency. Additionally, the proposed Text Style Loss supervises reconstruction at the style level, which effectively enhances both image quality and text recognition accuracy. Results on TextZoom and five STR benchmarks demonstrate that StyleSRN consistently enhances downstream recognition accuracy and image quality." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18700" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.279, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.227 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China (No.61502164, 62176097), the Natural Science Foundation of Hunan Province (No.2020JJ4057, 2024JJ10034), the Key Research and Development Program of Changsha Science and Technology Bureau (No.kq2004050), and the Scientific Research Foundation of the Education Department of Hunan Province of China (No.21A0052)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.252, + 0.188, + 0.267 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.277, + 0.484, + 0.345 + ], + "angle": 0, + "content": "[1] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE International Conference on Computer Vision, pages 2425-2433, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.347, + 0.484, + 0.415 + ], + "angle": 0, + "content": "[2] Jingye Chen, Bin Li, and Xiangyang Xue. Scene text telescope: Text-focused scene image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12026-12035, 2021. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.417, + 0.483, + 0.47 + ], + "angle": 0, + "content": "[3] Jingye Chen, Haiyang Yu, Jianqi Ma, Bin Li, and Xiangyang Xue. Text gestalt: Stroke-aware scene text image superresolution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 285-293, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.472, + 0.482, + 0.526 + ], + "angle": 0, + "content": "[4] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In Proceedings of the European Conference on Computer Vision, pages 184-199, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.528, + 0.483, + 0.58 + ], + "angle": 0, + "content": "[5] Chao Dong, Ximei Zhu, Yubin Deng, Chen Change Loy, and Yu Qiao. Boosting optical character recognition: A superresolution approach. arXiv preprint arXiv:1506.02211, 2015.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.582, + 0.482, + 0.637 + ], + "angle": 0, + "content": "[6] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2414-2423, 2016. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.639, + 0.482, + 0.707 + ], + "angle": 0, + "content": "[7] Hang Guo, Tao Dai, Guanghao Meng, and Shu-Tao Xia. Towards robust scene text image super-resolution via explicit location enhancement. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 782–790, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.708, + 0.482, + 0.762 + ], + "angle": 0, + "content": "[8] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE International Conference on Computer Vision, pages 1501-1510, 2017. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.763, + 0.482, + 0.804 + ], + "angle": 0, + "content": "[9] Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. Proceedings of the Advances in Neural Information Processing Systems, 28, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.805, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[10] Dimosthenis Karatzas, Faisal Shafait, Seiichi Uchida, Masakazu Iwamura, Lluis Gomez i Bigorda, Sergi Robles Mestre, Joan Mas, David Fernandez Mota, Jon Almazan Almazan, and Lluis Pere De Las Heras. Icdar 2013 robust reading competition. In Proceedings of the 12th International Conference on Document Analysis and Recognition, pages 1484-1493, 2013. 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.277, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.188 + ], + "angle": 0, + "content": "[11] Dimosthenis Karatzas, Lluis Gomez-Bigorda, Anguelos Nicolaou, Suman Ghosh, Andrew Bagdanov, Masakazu Iwamura, Jiri Matas, Lukas Neumann, Vijay Ramaseshan Chandrasekhar, Shijian Lu, et al. Icdar 2015 competition on robust reading. In Proceedings of the 13th International Conference on Document Analysis and Recognition, pages 1156-1160, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.19, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[12] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.905, + 0.301 + ], + "angle": 0, + "content": "[13] Vijeta Khare, Palaiahnakote Shivakumara, Chee Seng Chan, Tong Lu, Liang Kim Meng, Hon Hock Woon, and Michael Blumenstein. A novel character segmentation-reconstruction approach for license plate recognition. Expert Systems with Applications, 131:219-239, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[14] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4681-4690, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.904, + 0.427 + ], + "angle": 0, + "content": "[15] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.904, + 0.468 + ], + "angle": 0, + "content": "[16] Canjie Luo, Lianwen Jin, and Zenghui Sun. Moran: A multi-object rectified attention network for scene text recognition. Pattern Recognition, 90:109-118, 2019. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.904, + 0.537 + ], + "angle": 0, + "content": "[17] Jianqi Ma, Zhetong Liang, and Lei Zhang. A text attention network for spatial deformation robust scene text image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5911-5920, 2022. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[18] Jianqi Ma, Shi Guo, and Lei Zhang. Text prior guided scene text image super-resolution. IEEE Transactions on Image Processing, 32:1341-1353, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[19] Anand Mishra, Karteek Alahari, and CV Jawahar. Scene text recognition using higher order language priors. In Proceedings of the British Machine Vision Conference, pages 1-11, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.637, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[20] Chihiro Noguchi, Shun Fukuda, and Masao Yamanaka. Scene text image super-resolution based on text-conditional diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1474–1484, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.706, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[21] Ram Krishna Pandey, K Vignesh, AG Ramakrishnan, et al. Binary document image super resolution for improved readability andOCR performance. arXiv preprint arXiv:1812.02475, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[22] Trung Quy Phan, Palaiahnakote Shivakumara, Shangxuan Tian, and Chew Lim Tan. Recognizing text with perspective distortion in natural scenes. In Proceedings of the IEEE International Conference on Computer Vision, pages 569-576, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[23] Yadong Qu, Qingfeng Tan, Hongtao Xie, Jianjun Xu, Yuxin Wang, and Yongdong Zhang. Exploring stroke-level modifications for scene text editing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2119-2127, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18701" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[24] Anhar Risnumawan, Palaiahankote Shivakumara, Chee Seng Chan, and Chew Lim Tan. A robust arbitrary text detection system for natural scene images. Expert Systems with Applications, 41(18):8027-8048, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.217 + ], + "angle": 0, + "content": "[25] Baoguang Shi, Xiang Bai, and Cong Yao. An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(11):2298-2304, 2016. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.286 + ], + "angle": 0, + "content": "[26] Baoguang Shi, Mingkun Yang, Xinggang Wang, Pengyuan Lyu, Cong Yao, and Xiang Bai. Aster: An attentional scene text recognizer with flexible rectification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(9):2035-2048, 2018. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[27] Tonghua Su, Fuxiang Yang, Xiang Zhou, Donglin Di, Zhongjie Wang, and Songze Li. Scene style text editing, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.385 + ], + "angle": 0, + "content": "[28] Hanh TM Tran and Tien Ho-Phuoc. Deep laplacian pyramid network for text images super-resolution. In 2019 IEEE-RIVF International Conference on Computing and Communication Technologies (RIVF), pages 1-6, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[29] Qilong Wang, Banggu Wu, Pengfei Zhu, Peihua Li, Wangmeng Zuo, and Qinghua Hu. Eca-net: Efficient channel attention for deep convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11534-11542, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.482, + 0.523 + ], + "angle": 0, + "content": "[30] Wenjia Wang, Enze Xie, Xuebo Liu, Wenhai Wang, Ding Liang, Chunhua Shen, and Xiang Bai. Scene text image super-resolution in the wild. In Proceedings of the European Conference on Computer Vision, pages 650-666, 2020. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[31] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.649 + ], + "angle": 0, + "content": "[32] Chongsheng Zhang, Weiping Ding, Guowen Peng, Feifei Fu, and Wei Wang. Street view text recognition with deep learning for urban scene understanding in intelligent transportation systems. IEEE Transactions on Intelligent Transportation Systems, 22(7):4727-4743, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.482, + 0.732 + ], + "angle": 0, + "content": "[33] Wenyu Zhang, Xin Deng, Baojun Jia, Xingtong Yu, Yifan Chen, Jin Ma, Qing Ding, and Xinming Zhang. Pixel adapter: A graph-based post-processing approach for scene text image super-resolution. In Proceedings of the 31st ACM International Conference on Multimedia, pages 2168-2179, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.734, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[34] Minyi Zhao, Miao Wang, Fan Bai, Bingjia Li, Jie Wang, and Shuigeng Zhou. C3-STISR: scene text image superresolution with triple clues. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, pages 1707–1713, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[35] Yuxuan Zhou, Liangcai Gao, Zhi Tang, and Baole Wei. Recognition-guided diffusion model for scene text image super-resolution. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, pages 2940-2944, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[36] Shipeng Zhu, Zuoyan Zhao, Pengfei Fang, and Hui Xue. Improving scene text image super-resolution via dual prior" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "modulation network. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3843-3851, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.189 + ], + "angle": 0, + "content": "[37] Xiangyuan Zhu, Kehua Guo, Hui Fang, Rui Ding, Zheng Wu, and Gerald Schaefer. Gradient-based graph attention for scene text image super-resolution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3861-3869, 2023. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18702" + } + ] +] \ No newline at end of file diff --git a/2025/StyleSRN_ Scene Text Image Super-Resolution with Text Style Embedding/34baee7e-bb1d-4173-bb75-506b85eb6a0f_origin.pdf b/2025/StyleSRN_ Scene Text Image Super-Resolution with Text Style Embedding/34baee7e-bb1d-4173-bb75-506b85eb6a0f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..01ccbd45f9c5044592e814345b97fb0a2eedd43b --- /dev/null +++ b/2025/StyleSRN_ Scene Text Image Super-Resolution with Text Style Embedding/34baee7e-bb1d-4173-bb75-506b85eb6a0f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9b5c2d5975ef7613ff8abc35d96c67fb1bdaad4217ee441ad175973c14a91ba3 +size 1506625 diff --git a/2025/StyleSRN_ Scene Text Image Super-Resolution with Text Style Embedding/full.md b/2025/StyleSRN_ Scene Text Image Super-Resolution with Text Style Embedding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f737c238b57ac6333577751922f511cf3b16b81c --- /dev/null +++ b/2025/StyleSRN_ Scene Text Image Super-Resolution with Text Style Embedding/full.md @@ -0,0 +1,267 @@ +# StyleSRN: Scene Text Image Super-Resolution with Text Style Embedding + +Shengrong Yuan $^{1}$ , Runmin Wang $^{1,*}$ , Ke Hao $^{1}$ , Xuqi Ma $^{1}$ , Changxin Gao $^{2}$ , Li Liu $^{3}$ , Nong Sang $^{2}$ $^{1}$ School of Information Science and Engineering, Hunan Normal University + $^{2}$ School of Artificial Intelligence and Automation, Huazhong University of Science and Technology + $^{3}$ School of Electronic Science, National University of Defense Technology + +{shengrongyuan, runminwang, ke, xqma}@hunnu.edu.cn, {cgao, nsang}@hust.edu.cn, liliu@oulu.fi + +# Abstract + +Scene text image super-resolution (STISR) focuses on enhancing the clarity and readability of low-resolution text images. Existing methods often rely on text probability distribution priors derived from text recognizers to guide the super-resolution process. While effective in capturing general structural information of text, these priors lack the ability to preserve specific text style details, such as font, stereoscopic effect and spatial transformation, leading to a loss of visual quality and stylistic consistency. To address these limitations, we propose a Style embedding-based scene text image Super-Resolution Network (StyleSRN), which introduces a text style embedding mechanism to preserve and enhance text style features during the super-resolution process. The proposed architecture includes Style Enhancement Blocks for capturing multi-scale cross-channel dependencies, and Style Content Fusion Blocks that effectively integrates text content with style information, ensuring that the structure and style of the restored text are not distorted. Furthermore, we introduce a Text Style Loss based on the Gram matrix to supervise the reconstruction process at the style level, thereby maintaining the stylistic consistency of the restored text images. Extensive experiments on the TextZoom dataset and five scene text recognition benchmarks demonstrate the superiority of our method. Code is available at: https://github.com/Yuanssr/StyleSRN. + +# 1. Introduction + +Scene text images contain both visual and linguistic modalities, which can be extracted and analyzed to enhance various applications, including autonomous driving [32], visual question answering [1], and card recognition [13], etc. However, due to the limitations of imaging conditions of sensing devices, such as shooting jitter and low-focus cameras, scene text images often suffer from varying degrad + +![](images/e9d9b1c05c501673d15bb5d50fdfba13a5fdd98fb7474310be5d15cbf9abb91c.jpg) +Figure 1. Comparison of super-resolution results between previous methods and ours. Methods without text priors such as TSRN [30] retains less style information, while text priors-guided methods such as TATT [17] and LEMMA [7] almost completely lose the style information. + +tion, including blur, distortion, etc. Unlike general object images, scene text images possess more complex character structures and fine-grained stroke-level details, making even slight degradation challenging to recognize. Such low-resolution scene text images significantly hinder the performance of downstream tasks such as scene text detection, optical character recognition, and scene text recognition. Thus, restoring both the visual style quality and character structures of these degraded images is crucial. + +In recent years, many super-resolution methods specifically designed for scene text images have used the unique properties of text to improve restoration. Early approaches primarily focused on extracting sequential information from text images. Wang et al. [30] proposed a sequential residual block to capture the inherent sequential nature of text images. Chen et al. [2] introduced a text-focus loss function that emphasizes character positions and content, thereby improving recognition accuracy. More recently, methods that incorporate deep character classification information obtained from text recognizers have become the norm. Ma et al. [18] highlighted the benefits of integrating text + +prriors into the super-resolution process. Moreover, Ma et al. [17] developed an attention-based text priors interpreter to interactively fuse text probability distributions with image features. Guo et al. [7] demonstrated the effectiveness of using a bidirectional alignment strategy for text and images to obtain high-level guidance information. Most of these methods rely on text recognizers to generate text probability distribution priors, which are then encoded into the super-resolution network to guide the recovery of character structures. + +However, the text probability distribution priors derived from text recognizers mainly capture the general character structures but are less sensitive to the variations in text styles across different images, which hampers the super-resolution process. This loss of style details not only affects subsequent scene text recognition tasks but also disrupts the visual consistency of the images. The problem is particularly pronounced in scene text images with distinctive text styles, such as artistic characters or logos, where the loss of stylistic features can lead to more severe distortion in the restored images. As shown in Figure 1, text priors-guided methods such as TATT [17] and LEMMA [7] almost completely lose the style information of the degraded text image, while methods without text priors such as TSRN [30] retain less style information. To address these challenges, we draw inspiration from Style Transfer [6, 8, 12] to explore the potential of preserving image style information to enhance the restoration of low-resolution text images. Unlike content information, which is sensitive to position, style information in an image can be viewed as the correlation between different feature channels and is largely position-insensitive. By effectively encoding and embedding this style information during the super-resolution process, our proposed method aims to preserve both the structural integrity and the stylistic consistency of the restored text images. + +In this paper, we propose a Style embedding-based scene text image Super-Resolution Network (StyleSRN) for STISR, which not only restores the character structure but also preserves the stylistic consistency of the text. Unlike conventional methods that rely solely on text probability priors to guide the super-resolution process, our approach introduces a Text Style Embedding Branch that extracts and enhances style features. Additionally, we design a Style Enhancement Block (SEB), which captures multi-scale cross-channel dependencies, thereby enhancing the representation of style information. To further ensure the integration of style and content, we introduce a Style Content Fuse Block (SCFB) that effectively combines style embeddings with text priors. Finally, we propose a Text Style Loss based on the Gram matrix to supervise the reconstructed images at the style level. As illustrated in Figure 1, our method achieves superior visual quality and stylistic consistency in the restored images. Overall, our contributions can be sum + +marized as follows: + +- We, for the first time to our best knowledge, successfully embed text style information to guide super-resolution process for STISR task, and validate its effectiveness to ensure both structural integrity and stylistic consistency of the super-resolution process. +- We propose a Text Style Loss based on the Gram Matrix which leverages a text recognizer to extract features to offer robust supervision at style level for scene text images, which effectively improve both image quality and text recognition accuracy. +- Extensive experiment results demonstrate that the proposed method not only achieves state-of-the-art performance on the TextZoom dataset but also validates exceptional generalization across five scene text recognition datasets. + +# 2. Related Work + +# 2.1. Scene Text Image Super-resolution + +Scene text image super-resolution aims to enhance the visual quality and readability of scene text images. Early studies in this field usually adopt general image super-resolution methods [4, 5, 14, 21, 28] to reconstruct scene text images. However, due to the complex character structure of text images, these methods often fail to produce sufficiently recognizable text images. Therefore, some unique attributes of scene text have been exploited to improve super-resolution performance[2, 3, 7, 17, 18, 30]. For example, some methods focus on the sequential information of scene text. Wang et al. [30] introduced sequential residual block to capture the sequential information in text images. Chen et al. [2] proposed a Transformer-Based SuperResolution Network containing a Self-Attention Module to extract sequential information. Recently, the method of using text recognizers to obtain text probability distribution priors has become mainstream. For example, Ma et al. [18] incorporates categorical text priors into the STISR model training process, while Ma et al. [17] proposed a text prior interpreter to better transform text priors. Guo et al. [7] proposed a bimodal alignment mechanism to better utilize text priors. Additionally, some works use text priors to guide diffusion models to restore low-resolution images [20, 35]. Most of these methods use a pre-trained text recognizer to obtain deep character classification information to guide the super-resolution network. While this approach enables the model to effectively learn the general character structure, it results in the loss of specific style information in each image. + +# 2.2. Style Transfer + +Style transfer is a technique that involves altering the stylistic attributes of an image while maintaining its core con + +![](images/76805dc2ee52ed0c4838498beecf290ad5891d41cf4d92e87938a8565ec1ce3d.jpg) +Figure 2. Overview architecture of the proposed StyleSRN. GAP denotes the Global Average Pooling. $G(\cdot)$ denotes the Gram matrix, $\Phi_l(\cdot)$ represents the feature map obtained from the $l$ -th layer of the text recognizer. $\mathbf{I}_{\mathrm{LR}}$ , $\mathbf{I}_{\mathrm{SR}}$ and $\mathbf{I}_{\mathrm{HR}}$ represent low-resolution images, super-resolution images and high-resolution images, respectively. + +tent. This concept was pioneered by Gatys et al. [6], who utilized convolutional neural networks (CNNs) to transfer artistic styles between images by matching the Gram matrices of feature maps. Subsequent work has expanded on this foundation, exploring various methods to improve both the quality and efficiency of style transfer. One significant advancement is the introduction of adaptive instance normalization (AdaIN) [8], which allows for real-time arbitrary style transfer by aligning the mean and variance of content features to those of style features. Some works employ style transfer technology specifically for scene text style transfer [23, 27]. In our work, we draw inspiration from these style transfer methods, particularly in how they encode and embed style features. By adopting these principles, we aim to restore text style information, which is crucial for maintaining the visual consistency of the super-resolved text images. + +# 3. Methodology + +# 3.1. Overall Architecture + +The overall architecture of the proposed StyleSRN is illustrated in Figure 2. Our method consists of a text style embedding branch for extracting and enhancing style features and a super-resolution branch for restoring low-resolution images $\mathbf{I}_{\mathrm{LR}}$ . For the text style embedding branch, the low-resolution image will first pass through ResNet18 to obtain the feature map $\mathbf{X}_f \in \mathbb{R}^{C \times W \times H}$ , where $C$ represents the number of channels, and $H$ and $W$ represent the height and width of the feature map, respectively. Then several stacked Style Enhancement Blocks will extract the correla + +tion information between the feature map channels. After the style enhancement, we perform global average pooling on it to obtain the latent vector $\mathbf{z}$ . Given a latent vector $\mathbf{z}$ in the input latent space $\mathcal{Z}$ , a non-linear mapping network $f: \mathcal{Z} \to \mathcal{W}$ first produces $\mathbf{w} \in \mathcal{W}$ . Learned affine transformations then specialize $\mathbf{w}$ to styles $\mathbf{y} = (\mathbf{y}_s, \mathbf{y}_b)$ , where $\mathbf{y}_s$ and $\mathbf{y}_b$ denote the learned normalization scaling and bias coefficients, respectively. For the Super-Resolution Branch, the low-resolution image will first be processed by the Spatial Transformer Network (STN) [9] to align the spatially deformed texts. The convolution layer will extract features from the image processed by STN to obtain shallow image features. Then the shallow image features, text priors, and style information will be sent to several stacked style-content fusion blocks for style and content fusion. Each block contains a Style Enhancement Block, an adaptive instance normalization (AdaIN) [8], and a content embedding block. Finally, PixelShuffle is applied as an upsampling operator to generate $\mathbf{I}_{\mathrm{SR}}$ . + +# 3.2. Style Enhancement Block + +The style of an image has been traditionally viewed as the correlation between different features, as demonstrated in the advancements of style transfer [6, 8, 12]. While previous approaches, such as [12], directly utilized ResNet features to obtain style vectors, our approach introduces a novel enhancement by applying channel-wise attention to these feature maps, thereby better capturing and preserving the stylistic characteristics in scene text images. + +Inspired by the Efficient Channel Attention (ECA) + +mechanism [29], we designed a Style Enhancement Block (SEB) to improve the model's capacity for capturing the correlation between different channel features. Let the input feature map be denoted as $\mathbf{X}_f\in \mathbb{R}^{C\times H\times W}$ . We begin by applying Global Average Pooling (GAP) to the input feature map to obtain channel-wise statistics $\mathbf{m}\in \mathbb{R}^C$ . To further improve the model's capacity for capturing local cross-channel interactions at multiple scales, we employ multiple 1D convolution kernels of varying sizes, in contrast to the fixed-size kernels used in [29]. Specifically, let $n$ denote the number of different kernel sizes, each represented by $k_{l}$ , where $l\in 1,2,\ldots ,n$ . The convolutional operation at each scale can be expressed as: + +$$ +\mathbf {m} _ {c} ^ {\prime (l)} = \sum_ {j = - \frac {k _ {l}}{2}} ^ {\frac {k _ {l}}{2}} b _ {j} ^ {l} \mathbf {m} _ {c + j}, \tag {1} +$$ + +where $\mathbf{b}^l\in \mathbb{R}^{k_l}$ represents the weight vector for the convolution kernel of size $k_{l}$ . This multi-scale approach allows the model to capture a more nuanced set of cross-channel dependencies, enhancing the representation of style information. The outputs from the multiple convolutional scales are aggregated by computing the final channel attention weight $\mathbf{m}'_c$ as a weighted sum: + +$$ +\mathbf {m} _ {c} ^ {\prime} = \sum_ {l = 1} ^ {n} \alpha_ {l} \mathbf {m} _ {c} ^ {\prime (l)}, \tag {2} +$$ + +where $\alpha_{l}$ denotes the weighting coefficient for the output of the $l$ -th convolutional scale. These weights are normalized using a Sigmoid function: + +$$ +\mathbf {X} _ {f} ^ {\prime c, i, j} = \sigma \left(\mathbf {m} _ {c} ^ {\prime}\right) \cdot \mathbf {X} _ {f} ^ {c, i, j}, \tag {3} +$$ + +where $\sigma (\cdot)$ represents the Sigmoid activation function, and $\mathbf{X}^{\prime}$ is the final output feature map with the applied channelwise attention. + +The enhanced feature map is passed through several SEBs, which refines the correlation between different feature channels, ensuring that the model pays more attention to the features of important channels. This process effectively enhances the model's capability to capture and preserve the intricate stylistic details in scene text images. Subsequently, global average pooling (GAP) is applied to the final enhanced feature map to obtain a latent vector $\mathbf{z}$ in the latent space $\mathcal{Z}$ . The latent vector is then processed by a non-linear style mapping network $f: \mathcal{Z} \to \mathcal{W}$ , which produces a vector $\mathbf{w} \in \mathcal{W}$ that decouples the latent space. The style mapping network consists of an 8-layer MLP with LeakyReLU activation functions. Finally, learned affine transformations are applied to transform $\mathbf{w}$ into style parameters $\mathbf{y} = (\mathbf{y}_s, \mathbf{y}_b)$ . + +# 3.3. Style Content Fuse Block + +In the field of scene text image super-resolution, existing approaches primarily focus on embedding content information, including deep character category information derived from text probability priors. While this approach effectively guides the super-resolution network in restoring character structure, it often results in the loss of crucial stylistic features. To address this limitation, we propose a novel Style Content Fuse Block (SCFB) that allows the super-resolution network to restore character structures while preserving the stylistic consistency of the characters. + +Our innovation lies in embedding style information directly into the super-resolution network alongside content information. This dual embedding approach ensures that both character structure and style are preserved during the super-resolution process. Specifically, for the shallow image features $\mathbf{X}_s\in \mathbb{R}^{C'\times H'\times W'}$ obtained through a Spatial Transformer Network (STN) and a convolutional layer, we use Adaptive Instance Normalization (AdaIN) [8] to modulate their style. AdaIN has been shown to be effective in transferring style information across different domains. Given a feature map $\mathbf{X}_s$ and a style vector $\mathbf{y}$ , the AdaIN operation can be defined as follows: + +$$ +\operatorname {A d a I N} \left(\mathbf {X} _ {s}, \mathbf {y}\right) = \mathbf {y} _ {s} \left(\frac {\mathbf {X} _ {s} - \mu \left(\mathbf {X} _ {s}\right)}{\sigma \left(\mathbf {X} _ {s}\right)}\right) + \mathbf {y} _ {b}, \tag {4} +$$ + +where $\mu (\cdot)$ and $\sigma (\cdot)$ represent the mean and variance of $\mathbf{X}_s$ . Each AdaIN layer within the SCFB normalizes feature maps independently, with the dimensionality of $\mathbf{y}$ being twice the number of feature map channels $C^\prime$ in that layer. This embedding ensures that style information is seamlessly integrated into the spatial features of the image, thus preserving the character's stylistic attributes. + +In addition to embedding style information, content information, such as text probability distribution priors, should also be embedded into the super-resolution network to guide the learning of character structures. To bridge the domain gap between text priors and image features, we adopt the TP Interpreter architecture from [17] to extract the content embedding features. These interpreted features are then concatenated with the style-embedded features and subsequently passed through a Sequential Residual Block [30] for further refinement. + +# 3.4. Text Style Loss + +Although the proposed StyleSRN network effectively embeds the text's style features, these features are derived from the low-resolution image itself. When the style information in the low-resolution image is degraded, even if the network extracts these style features, it cannot fully reconstruct the style of the corresponding high-resolution image. Therefore, the reconstructed text image requires supervision at + +the style level to effectively learn the degradation of style information in the low-resolution image. + +To supervise the style in the reconstructed image, we introduce a Text Style Loss based on the Gram matrix. Unlike traditional methods [6] that use VGG networks to capture style information, our approach leverages a text recognizer to extract feature maps for Gram matrix computation. This enables better capture of text-specific style features, which are crucial for scene text images. + +The Gram matrix $\mathbf{G}_l$ for a given feature map $\Phi_l(\mathbf{I})$ at the $l$ -th layer is computed as follows. First, we reshape the feature map $\Phi_l(\mathbf{I})$ of dimensions $C_l \times H_l \times W_l$ into a matrix $\mathbf{F}_l$ of size $C_l \times (H_l \times W_l)$ , where each row corresponds to a channel and each column corresponds to a pixel. The Gram matrix $\mathbf{G}_l$ is the inner product between the different channels in layer $l$ : + +$$ +\mathbf {G} _ {i j} ^ {l} = \sum_ {k = 1} ^ {H _ {l} \times W _ {l}} \mathbf {F} _ {i k} ^ {l} \mathbf {F} _ {j k} ^ {l} \tag {5} +$$ + +The Gram matrix captures the correlations between different channels in the feature map, thereby encoding the style information of the image. + +The Text Style Loss $\mathcal{L}_{\mathrm{style}}$ is defined as the Mean Squared Error (MSE) between the Gram matrix of the high-resolution (HR) ground truth and the super-resolved (SR) output: + +$$ +\mathcal {L} _ {\text {s t y l e}} = \sum_ {l = 1} ^ {L} \frac {1}{4 C _ {l} ^ {2} H _ {l} ^ {2} W _ {l} ^ {2}} \| \mathbf {G} \left(\Phi_ {l} \left(\mathbf {I} _ {\mathrm {H R}}\right)\right) - \mathbf {G} \left(\Phi_ {l} \left(\mathbf {I} _ {\mathrm {S R}}\right)\right) \| _ {2} ^ {2} \tag {6} +$$ + +where $\mathbf{G}(\cdot)$ denotes the Gram matrix, $\Phi_l(\cdot)$ represents the feature map obtained from the $l$ -th layer of the text recognizer, and $C_l, H_l, W_l$ are the dimensions of the feature map at layer $l$ . + +# 3.5. Overall Loss Function + +In training, the overall loss function includes a Mean Square Error (MSE) loss $\mathcal{L}_2$ , a Text-Focus Loss $\mathcal{L}_{TP}$ [30], and the proposed Text Style Loss $\mathcal{L}_{style}$ . The $\mathcal{L}_2$ measures the difference between the super-resolution (SR) output and the ground-truth high-resolution (HR) image. The overall loss function is described as follows: + +$$ +\mathcal {L} = \mathcal {L} _ {2} + \alpha \mathcal {L} _ {T P} + \beta \mathcal {L} _ {\text {s t y l e}} \tag {7} +$$ + +where the $\alpha$ and $\beta$ are the balancing parameters. + +# 4. Experiments + +# 4.1. Datasets + +TextZoom TextZoom [30] is a scene text super-resolution dataset designed for real-world scenarios, containing + +21,740 pairs of low-resolution and high-resolution images captured by cameras with various focal lengths in natural environments, with 17,367 samples used for training. The test set is divided into easy, medium, and hard subsets, comprising 1,619, 1,411, and 1,343 LR-HR pairs, respectively, based on the camera focal lengths. + +Scene Text Recognition Datasets To assess the robustness and generalization capability of StyleSRN, we conducted comprehensive evaluations on five benchmark scene text recognition datasets: IC13 [10], IC15 [11], CUTE80 [24], IIIT5K [19], and SVTP [22]. These datasets were manually degraded to create low-resolution counterparts. Detailed descriptions of the datasets and degradation procedures are provided in the supplementary materials. + +# 4.2. Metric + +Following previous works, we adopt two types of metrics: super-resolution image text recognition accuracy and image quality. Specifically, to evaluate the generalization capability of StyleSRN, we employed three scene text recognizers: CRNN [25], MORAN [16], and ASTER [26]. For the image quality metric, we adopt the widely used Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM)[31]. + +# 4.3. Implementation Details + +We implement our model using PyTorch. All experiments are conducted on a single RTX 3090 GPU. We train our model with a batch size of 64 for 500 epochs using AdamW [15] for optimization. The learning rate is set to 1e-3 and decayed by a factor of 0.5 after 400 epochs. The number of SEBs and SCFBs are set to 3 and 5, respectively. In the SEBs, the sizes of the convolution kernels are $3 \times 3$ , $5 \times 5$ , and $7 \times 7$ , respectively. We use CRNN to generate the features for computing the text style loss and empirically set the number of feature maps L used to calculate the style loss to 5. We set the parameters $\alpha$ and $\beta$ to 0.5 and 1.0, respectively. + +# 4.4. Ablation Study + +In this section, we conduct ablation studies to investigate the effectiveness of SEBs and SCFBs, the effectiveness of training with Text Style Loss, the impact of numbers of SEBs and SCFBs. For STISR methods that do not use text priors, such as TSRN [30] and TBSRN [2], using SCFB means that image features are not used for text prior fusion, but are directly concatenated with style-embedded features. + +Effectiveness of SEBs and SCFBs. We compared the performance of different STISR methods using three strategies: No Style Embedding (NSE), only SCFB, and SCFB + +
MethodNSESCFBSEBImage QualityAccuracy(%)
PSNRSSIM
TSRN [30]✓--21.420.769041.4
-✓-21.540.769943.5
-✓✓21.560.769544.4
TBSRN [2]✓--20.910.760348.1
-✓-21.040.762350.3
-✓✓21.220.761451.1
TPGSR [18]✓--20.970.771951.8
-✓-21.040.772352.4
-✓✓21.170.772352.7
TATT [17]✓--21.520.793052.6
-✓-21.660.779054.1
-✓✓21.870.778554.9
LEMMA [7]✓--20.880.776056.3
-✓-20.940.777156.7
-✓✓21.050.776557.0
+ +Table 1. Effectiveness of the Style Content Fuse Block (SCFB) and the Style Enhancement Block (SEB). NSE refers to No Style Embedding. Accuracy represents the average accuracy with CRNN [25]. + +
Method\( \mathcal{L}_{\text{style}} \)PSNRSSIMAccuracy(%)
TSRN [30]-21.420.769041.4
✓21.560.771143.7
TPGSR [18]-20.970.771951.8
✓21.120.772452.4
TATT [17]-21.520.793052.6
✓21.810.793853.8
StyleSRN-21.540.775854.9
✓21.820.777857.4
+ +combined with the SEB in Table 2. NSE means removing both SEB and SCFB. SCFB can be used independently of SEB, but when using SEB, SCFB must be used to embed style information. Removing SEB means that the features extracted by Resnet18 are directly fed into the style mapping network through global pooling. Removing SCFB means that the features directly passed through the text interpreter do not need to be cascaded with the style features. As presented in Table 1, using only SCFB generally improves image quality and recognition accuracy over NSE. In addition, SCFB has a greater improvement over methods without text priors such as TSRN [30] and TBSRN [2]. For instance, with TSRN, the introduction of SCFB increases accuracy by $2.1\%$ , along with notable gains in PSNR and SSIM. We speculate that this may be because these methods do not have any prior guidance, while SCFB introduces the guidance of style information. Using SCFB and SEB at the same time can effectively improve PSNR and text recognition accuracy, but slightly reduce SSIM. This shows that SEB can effectively enhance style information, but slightly reduce structural consistency. + +Table 2. Ablation of the Text Style Loss. Accuracy represents the average accuracy with CRNN [25]. + +
DiscriminatorPSNRSSIMAccuracy (%)
VGG21.310.776156.3
CRNN21.820.777857.4
+ +Table 3. Ablation of the discriminator used in text style loss. + +
NQuality MetricAccuracy(%)
PSNRSSIMCRNN [25]MORAN [16]ASTER [26]
020.760.761854.661.865.0
121.820.777856.363.366.2
221.770.776456.964.066.7
421.770.777057.164.067.0
521.730.776855.963.266.7
321.820.777857.464.167.3
+ +Table 4. Ablation of the number of Style Enhancement Blocks. + +
NQuality MetricAccuracy(%)
PSNRSSIMCRNN [25]MORAN [16]ASTER [26]
020.670.743153.261.064.7
321.450.772255.362.366.2
421.430.775156.763.966.9
621.740.776956.964.067.0
721.550.775656.062.866.2
521.820.777857.464.167.3
+ +Table 5. Ablation of the number of Style Content Fuse Blocks. + +Effectiveness of training with Text Style Loss. We show the effectiveness of the Text Style Loss in Table 2. Incorporating the Text Style Loss results in improvements across all metrics. For instance, the PSNR of TSRN increases from 21.42 to 21.56, and its SSIM improves from 0.7690 to 0.7711. In terms of text recognition accuracy, models incorporating the Text Style Loss demonstrate enhanced performance. The accuracy of TSRN improves from $41.4\%$ to $43.7\%$ . These results highlight that the Text Style Loss effectively enhances both image quality and text recognition accuracy. Besides, we also compared the impact of using different discriminators in Table 3. It can be seen that using CRNN performs better than VGG. + +Numbers of SEBs. We show the effect of the number of SEBs in Table 4. As the number of SEBs increases, there is a noticeable improvement in both the PSNR and SSIM metrics. Specifically, using 3 SEBs achieves the highest PSNR of 21.82 and an SSIM of 0.7778, indicating the optimal enhancement of image quality. In terms of text recognition accuracy, models incorporating 3 SEBs also demonstrate superior performance across all recognition models. These findings suggest that SEBs significantly enhance both image quality and text recognition performance, with 3 SEBs being the optimal configuration. + +
MethodAccuracy of CRNN [25] (%)Accuracy of MORAN [16] (%)Accuracy of ASTER [26] (%)
EasyMediumHardAvgEasyMediumHardAvgEasyMediumhardAvg
BICUBIC36.421.121.126.860.637.930.844.167.442.431.248.2
HR76.475.164.672.491.285.374.284.194.287.776.286.6
TSRN [30]52.538.231.441.470.155.337.955.475.156.340.158.3
TBSRN [2]59.647.135.348.174.157.040.858.475.759.941.660.1
TG [3]61.247.635.548.975.857.841.459.477.960.242.461.3
TPGSR [18]63.152.038.651.874.960.544.160.578.962.744.562.8
TATT [17]62.653.439.852.672.660.243.159.578.963.445.463.6
C3-STISR [34]65.253.639.853.774.261.043.259.579.163.346.864.1
TSAN [37]64.653.338.853.078.461.345.162.779.664.145.364.1
LEMMA [7]67.158.840.656.377.764.444.663.281.166.347.466.0
DPMN [36]64.454.239.253.473.361.543.960.479.364.145.263.9
TCDM [20]67.357.342.755.777.662.945.962.281.365.150.165.5
RTSRN [33]67.059.242.657.077.163.346.563.280.466.149.166.2
StyleSRN68.159.442.457.478.665.145.764.182.767.448.767.3
+ +Table 6. Comparison with the existing methods in terms of the recognition accuracy on TextZoom[30]. BICUBIC means LR images are directly upsampled by the bicubic interpolation, and the same definition applies to BICUBIC in subsequent tables. + +
MethodLossPSNRSSIM
EasyMediumHardAvgEasyMediumHardAvg
BICUBIC-22.3518.9819.3920.350.78840.62540.65920.6961
TSRN [30]\( \mathcal{L}_2 + \mathcal{L}_{GP} \)25.0718.8619.7121.420.88970.66760.73020.7690
TBSRN [2]\( \mathcal{L}_{POS} + \mathcal{L}_{CON} \)23.4619.1719.6820.910.87290.64550.74520.7603
TPGSR [18]\( \mathcal{L}_2 + \mathcal{L}_{TP} \)23.7318.6820.0620.970.88050.67380.74400.7719
TATT [17]\( \mathcal{L}_2 + \mathcal{L}_{TP} + \mathcal{L}_{TSC} \)24.7219.0220.3121.520.90060.69110.77030.7930
LEMMA [7]\( \mathcal{L}_2 + \mathcal{L}_{TP} + \mathcal{L}_{finetune} \)23.7019.3719.8420.880.87530.69050.74910.7760
StyleSRN\( \mathcal{L}_2 + \mathcal{L}_{TP} + \mathcal{L}_{style} \)24.6919.7020.5821.820.88440.67910.75310.7778
+ +Numbers of SCFBs. We further examined the impact of the numbers of SCFBs, as detailed in Table 5. While increasing the number of SCFBs improves recognition accuracy metrics, diminishing returns are observed beyond a certain point. The optimal performance, consistent with the SEB ablation study, is achieved with 5 SCFBs, yielding a PSNR of 21.82 and an SSIM of 0.7778. Moreover, text recognition accuracy peaks with CRNN (57.40%), MORAN (64.14%), and ASTER (67.31%) when 5 SCFBs are utilized. This suggests that while the SCFB is essential for balancing style and content, its optimal configuration involves a moderate number of blocks, beyond which performance gains begin to plateau. + +# 4.5. Comparison with State-of-the-Arts + +Results on TextZoom. We evaluate the recognition accuracy of text images generated by various STISR methods using three text recognizers: CRNN [25], MORAN [16], and ASTER [26]. As shown in Table 7, our proposed StyleSRN consistently outperforms state-of-the-art methods across all recognizers, achieving average accuracy rates of $57.4\%$ with CRNN, $64.1\%$ with MORAN, and $67.3\%$ with ASTER. On the hard subset, StyleSRN performs slightly lower than some methods, such as TCDM and DPMN, likely due to challenges in balancing structure and style preservation in extremely blurred or occluded + +Table 7. Comparison of the image quality on TextZoom [30]. Avg represents the average value of three subsets. + +
MethodAccuracy(%)
IC13IC15CT80SVTPIII5K
BICUBIC76.1638.2155.2135.5059.70
TSRN [30]75.6741.5254.8635.9762.60
TBSRN [2]76.9543.0754.8637.3663.60
TPGSR [18]77.6442.8553.4738.2962.07
TATT [17]76.0648.4851.0442.0262.67
LEMMA [7]79.8049.7051.0443.7265.07
StyleSRN80.9950.4755.5646.5165.63
+ +Table 8. Comparison of the accuracy on manually degrade scene text recognition datasets. Accuracy represents the average accuracy with CRNN [25]. + +cases. In terms of image quality, as shown in Table 7, our method achieves the highest average PSNR of 21.82, surpassing other methods like TATT (21.52) and LEMMA (20.93). Our method achieves an SSIM of 0.7778, which is comparable to TATT (0.7930). The visual comparison is presented in Figure 3, which demonstrates that our method achieves superior visual quality. Our method not only restores the structure of the characters but also preserves their style information. For instance, for certain stereoscopic scene texts shown in Figure 3, our method more effectively restores their stereoscopic effect and shadows details. + +![](images/4e0fd2032e03f8e5ea2b61d52496d82f99a05df52b1fa502eb7b905fe010bd94.jpg) +Figure 3. Visualization of SR images and their recognition result. Each text image below uses CRNN [25] for recognition. Characters marked in red indicate incorrect recognition results. + +![](images/48408a6d7957e9693abe8ca9efcdf347a6e7fce11c38ad672c5cb1affa70758d.jpg) +Figure 4. Visualization of the STISR results on scene text with uneven illumination and scene text with severe blur. + +Generalization to recognition datasets. To further evaluate the generalization capability and robustness of the proposed StyleSRN, we also conduct experiments on five scene text recognition datasets. These datasets cover a broad spectrum of image styles and label distributions, presenting substantial challenges for the generalization performance of super-resolution networks. Due to the absence of paired low-resolution and high-resolution images in these datasets, we manually degrad the original images before applying StyleSRN for preprocessing. As illustrated in Table 2, our method consistently surpasses previous approaches across these datasets, confirming its strong generalization ability. + +# 5. Discussion + +While StyleSRN effectively improves the structural integrity and style consistency of STISR, as shown in Figure 4, it struggles to recover character structures in images with uneven illumination or severe blur, highlighting the need for further refinement to address degraded visual conditions. Another limitation is the absence of a standardized quantitative metric for evaluating style degradation in STISR methods. While our method restores the style of low-resolution scene text images, no existing metric specifically measures style preservation or degradation. This lack of objective criteria makes it difficult to rigorously assess style restoration performance, as current evaluations are largely subjective and based on visual inspection. Future research should focus on developing standardized, quantitative metrics to objectively evaluate style preservation in STISR, enabling more precise assessments and meaningful comparisons across methods. + +# 6. Conclusions + +In this work, we propose StyleSRN to address the limitations of current STISR methods that neglect the stylistic details of scene text images. StyleSRN incorporates Style Enhancement Blocks and Style Content Fuse Blocks to effectively capture and integrate text style with content, ensuring both structural integrity and stylistic consistency. Additionally, the proposed Text Style Loss supervises reconstruction at the style level, which effectively enhances both image quality and text recognition accuracy. Results on TextZoom and five STR benchmarks demonstrate that StyleSRN consistently enhances downstream recognition accuracy and image quality. + +# 7. Acknowledgements + +This work was supported in part by the National Natural Science Foundation of China (No.61502164, 62176097), the Natural Science Foundation of Hunan Province (No.2020JJ4057, 2024JJ10034), the Key Research and Development Program of Changsha Science and Technology Bureau (No.kq2004050), and the Scientific Research Foundation of the Education Department of Hunan Province of China (No.21A0052). + +# References + +[1] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE International Conference on Computer Vision, pages 2425-2433, 2015. 1 +[2] Jingye Chen, Bin Li, and Xiangyang Xue. Scene text telescope: Text-focused scene image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12026-12035, 2021. 1, 2, 5, 6, 7 +[3] Jingye Chen, Haiyang Yu, Jianqi Ma, Bin Li, and Xiangyang Xue. Text gestalt: Stroke-aware scene text image superresolution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 285-293, 2022. 2, 7 +[4] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In Proceedings of the European Conference on Computer Vision, pages 184-199, 2014. 2 +[5] Chao Dong, Ximei Zhu, Yubin Deng, Chen Change Loy, and Yu Qiao. Boosting optical character recognition: A superresolution approach. arXiv preprint arXiv:1506.02211, 2015.2 +[6] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2414-2423, 2016. 2, 3, 5 +[7] Hang Guo, Tao Dai, Guanghao Meng, and Shu-Tao Xia. Towards robust scene text image super-resolution via explicit location enhancement. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 782–790, 2023. 1, 2, 6, 7 +[8] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE International Conference on Computer Vision, pages 1501-1510, 2017. 2, 3, 4 +[9] Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. Proceedings of the Advances in Neural Information Processing Systems, 28, 2015. 3 +[10] Dimosthenis Karatzas, Faisal Shafait, Seiichi Uchida, Masakazu Iwamura, Lluis Gomez i Bigorda, Sergi Robles Mestre, Joan Mas, David Fernandez Mota, Jon Almazan Almazan, and Lluis Pere De Las Heras. Icdar 2013 robust reading competition. In Proceedings of the 12th International Conference on Document Analysis and Recognition, pages 1484-1493, 2013. 5 + +[11] Dimosthenis Karatzas, Lluis Gomez-Bigorda, Anguelos Nicolaou, Suman Ghosh, Andrew Bagdanov, Masakazu Iwamura, Jiri Matas, Lukas Neumann, Vijay Ramaseshan Chandrasekhar, Shijian Lu, et al. Icdar 2015 competition on robust reading. In Proceedings of the 13th International Conference on Document Analysis and Recognition, pages 1156-1160, 2015. 5 +[12] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks, 2019. 2, 3 +[13] Vijeta Khare, Palaiahnakote Shivakumara, Chee Seng Chan, Tong Lu, Liang Kim Meng, Hon Hock Woon, and Michael Blumenstein. A novel character segmentation-reconstruction approach for license plate recognition. Expert Systems with Applications, 131:219-239, 2019. 1 +[14] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4681-4690, 2017. 2 +[15] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5 +[16] Canjie Luo, Lianwen Jin, and Zenghui Sun. Moran: A multi-object rectified attention network for scene text recognition. Pattern Recognition, 90:109-118, 2019. 5, 6, 7 +[17] Jianqi Ma, Zhetong Liang, and Lei Zhang. A text attention network for spatial deformation robust scene text image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5911-5920, 2022. 1, 2, 4, 6, 7 +[18] Jianqi Ma, Shi Guo, and Lei Zhang. Text prior guided scene text image super-resolution. IEEE Transactions on Image Processing, 32:1341-1353, 2023. 1, 2, 6, 7 +[19] Anand Mishra, Karteek Alahari, and CV Jawahar. Scene text recognition using higher order language priors. In Proceedings of the British Machine Vision Conference, pages 1-11, 2012. 5 +[20] Chihiro Noguchi, Shun Fukuda, and Masao Yamanaka. Scene text image super-resolution based on text-conditional diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1474–1484, 2024. 2, 7 +[21] Ram Krishna Pandey, K Vignesh, AG Ramakrishnan, et al. Binary document image super resolution for improved readability andOCR performance. arXiv preprint arXiv:1812.02475, 2018. 2 +[22] Trung Quy Phan, Palaiahnakote Shivakumara, Shangxuan Tian, and Chew Lim Tan. Recognizing text with perspective distortion in natural scenes. In Proceedings of the IEEE International Conference on Computer Vision, pages 569-576, 2013. 5 +[23] Yadong Qu, Qingfeng Tan, Hongtao Xie, Jianjun Xu, Yuxin Wang, and Yongdong Zhang. Exploring stroke-level modifications for scene text editing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2119-2127, 2023. 3 + +[24] Anhar Risnumawan, Palaiahankote Shivakumara, Chee Seng Chan, and Chew Lim Tan. A robust arbitrary text detection system for natural scene images. Expert Systems with Applications, 41(18):8027-8048, 2014. 5 +[25] Baoguang Shi, Xiang Bai, and Cong Yao. An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(11):2298-2304, 2016. 5, 6, 7, 8 +[26] Baoguang Shi, Mingkun Yang, Xinggang Wang, Pengyuan Lyu, Cong Yao, and Xiang Bai. Aster: An attentional scene text recognizer with flexible rectification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(9):2035-2048, 2018. 5, 6, 7 +[27] Tonghua Su, Fuxiang Yang, Xiang Zhou, Donglin Di, Zhongjie Wang, and Songze Li. Scene style text editing, 2023. 3 +[28] Hanh TM Tran and Tien Ho-Phuoc. Deep laplacian pyramid network for text images super-resolution. In 2019 IEEE-RIVF International Conference on Computing and Communication Technologies (RIVF), pages 1-6, 2019. 2 +[29] Qilong Wang, Banggu Wu, Pengfei Zhu, Peihua Li, Wangmeng Zuo, and Qinghua Hu. Eca-net: Efficient channel attention for deep convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11534-11542, 2020. 4 +[30] Wenjia Wang, Enze Xie, Xuebo Liu, Wenhai Wang, Ding Liang, Chunhua Shen, and Xiang Bai. Scene text image super-resolution in the wild. In Proceedings of the European Conference on Computer Vision, pages 650-666, 2020. 1, 2, 4, 5, 6, 7 +[31] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5 +[32] Chongsheng Zhang, Weiping Ding, Guowen Peng, Feifei Fu, and Wei Wang. Street view text recognition with deep learning for urban scene understanding in intelligent transportation systems. IEEE Transactions on Intelligent Transportation Systems, 22(7):4727-4743, 2020. 1 +[33] Wenyu Zhang, Xin Deng, Baojun Jia, Xingtong Yu, Yifan Chen, Jin Ma, Qing Ding, and Xinming Zhang. Pixel adapter: A graph-based post-processing approach for scene text image super-resolution. In Proceedings of the 31st ACM International Conference on Multimedia, pages 2168-2179, 2023. 7 +[34] Minyi Zhao, Miao Wang, Fan Bai, Bingjia Li, Jie Wang, and Shuigeng Zhou. C3-STISR: scene text image superresolution with triple clues. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, pages 1707–1713, 2022. 7 +[35] Yuxuan Zhou, Liangcai Gao, Zhi Tang, and Baole Wei. Recognition-guided diffusion model for scene text image super-resolution. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, pages 2940-2944, 2024. 2 +[36] Shipeng Zhu, Zuoyan Zhao, Pengfei Fang, and Hui Xue. Improving scene text image super-resolution via dual prior + +modulation network. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3843-3851, 2023. 7 +[37] Xiangyuan Zhu, Kehua Guo, Hui Fang, Rui Ding, Zheng Wu, and Gerald Schaefer. Gradient-based graph attention for scene text image super-resolution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3861-3869, 2023. 7 \ No newline at end of file diff --git a/2025/StyleSRN_ Scene Text Image Super-Resolution with Text Style Embedding/images.zip b/2025/StyleSRN_ Scene Text Image Super-Resolution with Text Style Embedding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1b08d2dfbdd580a27b9725c3a46d875491e8a8c9 --- /dev/null +++ b/2025/StyleSRN_ Scene Text Image Super-Resolution with Text Style Embedding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d5d4a99c94fa4103f8b968cd92abd591040a34d5d2a94ea629c4fb7dd723c09b +size 632932 diff --git a/2025/StyleSRN_ Scene Text Image Super-Resolution with Text Style Embedding/layout.json b/2025/StyleSRN_ Scene Text Image Super-Resolution with Text Style Embedding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6f4d8f435bc71637fa2c283c3ced531c62fd6b57 --- /dev/null +++ b/2025/StyleSRN_ Scene Text Image Super-Resolution with Text Style Embedding/layout.json @@ -0,0 +1,7362 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 76, + 103, + 534, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 103, + 534, + 121 + ], + "spans": [ + { + "bbox": [ + 76, + 103, + 534, + 121 + ], + "type": "text", + "content": "StyleSRN: Scene Text Image Super-Resolution with Text Style Embedding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 142, + 545, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 142, + 545, + 200 + ], + "spans": [ + { + "bbox": [ + 64, + 142, + 545, + 200 + ], + "type": "text", + "content": "Shengrong Yuan" + }, + { + "bbox": [ + 64, + 142, + 545, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 142, + 545, + 200 + ], + "type": "text", + "content": ", Runmin Wang" + }, + { + "bbox": [ + 64, + 142, + 545, + 200 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 64, + 142, + 545, + 200 + ], + "type": "text", + "content": ", Ke Hao" + }, + { + "bbox": [ + 64, + 142, + 545, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 142, + 545, + 200 + ], + "type": "text", + "content": ", Xuqi Ma" + }, + { + "bbox": [ + 64, + 142, + 545, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 142, + 545, + 200 + ], + "type": "text", + "content": ", Changxin Gao" + }, + { + "bbox": [ + 64, + 142, + 545, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 142, + 545, + 200 + ], + "type": "text", + "content": ", Li Liu" + }, + { + "bbox": [ + 64, + 142, + 545, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 64, + 142, + 545, + 200 + ], + "type": "text", + "content": ", Nong Sang" + }, + { + "bbox": [ + 64, + 142, + 545, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 142, + 545, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 142, + 545, + 200 + ], + "type": "text", + "content": "School of Information Science and Engineering, Hunan Normal University \n" + }, + { + "bbox": [ + 64, + 142, + 545, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 142, + 545, + 200 + ], + "type": "text", + "content": "School of Artificial Intelligence and Automation, Huazhong University of Science and Technology \n" + }, + { + "bbox": [ + 64, + 142, + 545, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 64, + 142, + 545, + 200 + ], + "type": "text", + "content": "School of Electronic Science, National University of Defense Technology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 80, + 201, + 526, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 201, + 526, + 213 + ], + "spans": [ + { + "bbox": [ + 80, + 201, + 526, + 213 + ], + "type": "text", + "content": "{shengrongyuan, runminwang, ke, xqma}@hunnu.edu.cn, {cgao, nsang}@hust.edu.cn, liliu@oulu.fi" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "spans": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 266, + 296, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 296, + 566 + ], + "type": "text", + "content": "Scene text image super-resolution (STISR) focuses on enhancing the clarity and readability of low-resolution text images. Existing methods often rely on text probability distribution priors derived from text recognizers to guide the super-resolution process. While effective in capturing general structural information of text, these priors lack the ability to preserve specific text style details, such as font, stereoscopic effect and spatial transformation, leading to a loss of visual quality and stylistic consistency. To address these limitations, we propose a Style embedding-based scene text image Super-Resolution Network (StyleSRN), which introduces a text style embedding mechanism to preserve and enhance text style features during the super-resolution process. The proposed architecture includes Style Enhancement Blocks for capturing multi-scale cross-channel dependencies, and Style Content Fusion Blocks that effectively integrates text content with style information, ensuring that the structure and style of the restored text are not distorted. Furthermore, we introduce a Text Style Loss based on the Gram matrix to supervise the reconstruction process at the style level, thereby maintaining the stylistic consistency of the restored text images. Extensive experiments on the TextZoom dataset and five scene text recognition benchmarks demonstrate the superiority of our method. Code is available at: https://github.com/Yuanssr/StyleSRN." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 589, + 135, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 589, + 135, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 589, + 135, + 601 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 609, + 295, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 295, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 295, + 694 + ], + "type": "text", + "content": "Scene text images contain both visual and linguistic modalities, which can be extracted and analyzed to enhance various applications, including autonomous driving [32], visual question answering [1], and card recognition [13], etc. However, due to the limitations of imaging conditions of sensing devices, such as shooting jitter and low-focus cameras, scene text images often suffer from varying degrad" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 323, + 247, + 545, + 371 + ], + "blocks": [ + { + "bbox": [ + 323, + 247, + 545, + 371 + ], + "lines": [ + { + "bbox": [ + 323, + 247, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 323, + 247, + 545, + 371 + ], + "type": "image", + "image_path": "e9d9b1c05c501673d15bb5d50fdfba13a5fdd98fb7474310be5d15cbf9abb91c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 381, + 555, + 437 + ], + "lines": [ + { + "bbox": [ + 313, + 381, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 555, + 437 + ], + "type": "text", + "content": "Figure 1. Comparison of super-resolution results between previous methods and ours. Methods without text priors such as TSRN [30] retains less style information, while text priors-guided methods such as TATT [17] and LEMMA [7] almost completely lose the style information." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 460, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 568 + ], + "type": "text", + "content": "tion, including blur, distortion, etc. Unlike general object images, scene text images possess more complex character structures and fine-grained stroke-level details, making even slight degradation challenging to recognize. Such low-resolution scene text images significantly hinder the performance of downstream tasks such as scene text detection, optical character recognition, and scene text recognition. Thus, restoring both the visual style quality and character structures of these degraded images is crucial." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "In recent years, many super-resolution methods specifically designed for scene text images have used the unique properties of text to improve restoration. Early approaches primarily focused on extracting sequential information from text images. Wang et al. [30] proposed a sequential residual block to capture the inherent sequential nature of text images. Chen et al. [2] introduced a text-focus loss function that emphasizes character positions and content, thereby improving recognition accuracy. More recently, methods that incorporate deep character classification information obtained from text recognizers have become the norm. Ma et al. [18] highlighted the benefits of integrating text" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18693" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "prriors into the super-resolution process. Moreover, Ma et al. [17] developed an attention-based text priors interpreter to interactively fuse text probability distributions with image features. Guo et al. [7] demonstrated the effectiveness of using a bidirectional alignment strategy for text and images to obtain high-level guidance information. Most of these methods rely on text recognizers to generate text probability distribution priors, which are then encoded into the super-resolution network to guide the recovery of character structures." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 196, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 196, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 56, + 196, + 296, + 495 + ], + "type": "text", + "content": "However, the text probability distribution priors derived from text recognizers mainly capture the general character structures but are less sensitive to the variations in text styles across different images, which hampers the super-resolution process. This loss of style details not only affects subsequent scene text recognition tasks but also disrupts the visual consistency of the images. The problem is particularly pronounced in scene text images with distinctive text styles, such as artistic characters or logos, where the loss of stylistic features can lead to more severe distortion in the restored images. As shown in Figure 1, text priors-guided methods such as TATT [17] and LEMMA [7] almost completely lose the style information of the degraded text image, while methods without text priors such as TSRN [30] retain less style information. To address these challenges, we draw inspiration from Style Transfer [6, 8, 12] to explore the potential of preserving image style information to enhance the restoration of low-resolution text images. Unlike content information, which is sensitive to position, style information in an image can be viewed as the correlation between different feature channels and is largely position-insensitive. By effectively encoding and embedding this style information during the super-resolution process, our proposed method aims to preserve both the structural integrity and the stylistic consistency of the restored text images." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "content": "In this paper, we propose a Style embedding-based scene text image Super-Resolution Network (StyleSRN) for STISR, which not only restores the character structure but also preserves the stylistic consistency of the text. Unlike conventional methods that rely solely on text probability priors to guide the super-resolution process, our approach introduces a Text Style Embedding Branch that extracts and enhances style features. Additionally, we design a Style Enhancement Block (SEB), which captures multi-scale cross-channel dependencies, thereby enhancing the representation of style information. To further ensure the integration of style and content, we introduce a Style Content Fuse Block (SCFB) that effectively combines style embeddings with text priors. Finally, we propose a Text Style Loss based on the Gram matrix to supervise the reconstructed images at the style level. As illustrated in Figure 1, our method achieves superior visual quality and stylistic consistency in the restored images. Overall, our contributions can be sum" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 314, + 73, + 395, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 73, + 395, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 73, + 395, + 83 + ], + "type": "text", + "content": "marized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 87, + 553, + 265 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 314, + 87, + 553, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 87, + 553, + 145 + ], + "spans": [ + { + "bbox": [ + 314, + 87, + 553, + 145 + ], + "type": "text", + "content": "- We, for the first time to our best knowledge, successfully embed text style information to guide super-resolution process for STISR task, and validate its effectiveness to ensure both structural integrity and stylistic consistency of the super-resolution process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 147, + 553, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 147, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 313, + 147, + 553, + 206 + ], + "type": "text", + "content": "- We propose a Text Style Loss based on the Gram Matrix which leverages a text recognizer to extract features to offer robust supervision at style level for scene text images, which effectively improve both image quality and text recognition accuracy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 206, + 553, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 206, + 553, + 265 + ], + "spans": [ + { + "bbox": [ + 313, + 206, + 553, + 265 + ], + "type": "text", + "content": "- Extensive experiment results demonstrate that the proposed method not only achieves state-of-the-art performance on the TextZoom dataset but also validates exceptional generalization across five scene text recognition datasets." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 279, + 400, + 292 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 279, + 400, + 292 + ], + "spans": [ + { + "bbox": [ + 314, + 279, + 400, + 292 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 299, + 499, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 299, + 499, + 312 + ], + "spans": [ + { + "bbox": [ + 314, + 299, + 499, + 312 + ], + "type": "text", + "content": "2.1. Scene Text Image Super-resolution" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 316, + 553, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 553, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 553, + 664 + ], + "type": "text", + "content": "Scene text image super-resolution aims to enhance the visual quality and readability of scene text images. Early studies in this field usually adopt general image super-resolution methods [4, 5, 14, 21, 28] to reconstruct scene text images. However, due to the complex character structure of text images, these methods often fail to produce sufficiently recognizable text images. Therefore, some unique attributes of scene text have been exploited to improve super-resolution performance[2, 3, 7, 17, 18, 30]. For example, some methods focus on the sequential information of scene text. Wang et al. [30] introduced sequential residual block to capture the sequential information in text images. Chen et al. [2] proposed a Transformer-Based SuperResolution Network containing a Self-Attention Module to extract sequential information. Recently, the method of using text recognizers to obtain text probability distribution priors has become mainstream. For example, Ma et al. [18] incorporates categorical text priors into the STISR model training process, while Ma et al. [17] proposed a text prior interpreter to better transform text priors. Guo et al. [7] proposed a bimodal alignment mechanism to better utilize text priors. Additionally, some works use text priors to guide diffusion models to restore low-resolution images [20, 35]. Most of these methods use a pre-trained text recognizer to obtain deep character classification information to guide the super-resolution network. While this approach enables the model to effectively learn the general character structure, it results in the loss of specific style information in each image." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 671, + 404, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 671, + 404, + 684 + ], + "spans": [ + { + "bbox": [ + 314, + 671, + 404, + 684 + ], + "type": "text", + "content": "2.2. Style Transfer" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Style transfer is a technique that involves altering the stylistic attributes of an image while maintaining its core con" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18694" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 71, + 545, + 286 + ], + "blocks": [ + { + "bbox": [ + 70, + 71, + 545, + 286 + ], + "lines": [ + { + "bbox": [ + 70, + 71, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 70, + 71, + 545, + 286 + ], + "type": "image", + "image_path": "76805dc2ee52ed0c4838498beecf290ad5891d41cf4d92e87938a8565ec1ce3d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 295, + 555, + 331 + ], + "lines": [ + { + "bbox": [ + 54, + 295, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 54, + 295, + 555, + 331 + ], + "type": "text", + "content": "Figure 2. Overview architecture of the proposed StyleSRN. GAP denotes the Global Average Pooling. " + }, + { + "bbox": [ + 54, + 295, + 555, + 331 + ], + "type": "inline_equation", + "content": "G(\\cdot)" + }, + { + "bbox": [ + 54, + 295, + 555, + 331 + ], + "type": "text", + "content": " denotes the Gram matrix, " + }, + { + "bbox": [ + 54, + 295, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\Phi_l(\\cdot)" + }, + { + "bbox": [ + 54, + 295, + 555, + 331 + ], + "type": "text", + "content": " represents the feature map obtained from the " + }, + { + "bbox": [ + 54, + 295, + 555, + 331 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 54, + 295, + 555, + 331 + ], + "type": "text", + "content": "-th layer of the text recognizer. " + }, + { + "bbox": [ + 54, + 295, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{LR}}" + }, + { + "bbox": [ + 54, + 295, + 555, + 331 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 295, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{SR}}" + }, + { + "bbox": [ + 54, + 295, + 555, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 295, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{HR}}" + }, + { + "bbox": [ + 54, + 295, + 555, + 331 + ], + "type": "text", + "content": " represent low-resolution images, super-resolution images and high-resolution images, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 350, + 298, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 350, + 298, + 542 + ], + "spans": [ + { + "bbox": [ + 54, + 350, + 298, + 542 + ], + "type": "text", + "content": "tent. This concept was pioneered by Gatys et al. [6], who utilized convolutional neural networks (CNNs) to transfer artistic styles between images by matching the Gram matrices of feature maps. Subsequent work has expanded on this foundation, exploring various methods to improve both the quality and efficiency of style transfer. One significant advancement is the introduction of adaptive instance normalization (AdaIN) [8], which allows for real-time arbitrary style transfer by aligning the mean and variance of content features to those of style features. Some works employ style transfer technology specifically for scene text style transfer [23, 27]. In our work, we draw inspiration from these style transfer methods, particularly in how they encode and embed style features. By adopting these principles, we aim to restore text style information, which is crucial for maintaining the visual consistency of the super-resolved text images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 555, + 138, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 138, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 138, + 568 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 575, + 176, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 176, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 176, + 586 + ], + "type": "text", + "content": "3.1. Overall Architecture" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "The overall architecture of the proposed StyleSRN is illustrated in Figure 2. Our method consists of a text style embedding branch for extracting and enhancing style features and a super-resolution branch for restoring low-resolution images " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{LR}}" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": ". For the text style embedding branch, the low-resolution image will first pass through ResNet18 to obtain the feature map " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_f \\in \\mathbb{R}^{C \\times W \\times H}" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " represents the number of channels, and " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " represent the height and width of the feature map, respectively. Then several stacked Style Enhancement Blocks will extract the correla" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 350, + 556, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 556, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 556, + 578 + ], + "type": "text", + "content": "tion information between the feature map channels. After the style enhancement, we perform global average pooling on it to obtain the latent vector " + }, + { + "bbox": [ + 313, + 350, + 556, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 313, + 350, + 556, + 578 + ], + "type": "text", + "content": ". Given a latent vector " + }, + { + "bbox": [ + 313, + 350, + 556, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 313, + 350, + 556, + 578 + ], + "type": "text", + "content": " in the input latent space " + }, + { + "bbox": [ + 313, + 350, + 556, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 313, + 350, + 556, + 578 + ], + "type": "text", + "content": ", a non-linear mapping network " + }, + { + "bbox": [ + 313, + 350, + 556, + 578 + ], + "type": "inline_equation", + "content": "f: \\mathcal{Z} \\to \\mathcal{W}" + }, + { + "bbox": [ + 313, + 350, + 556, + 578 + ], + "type": "text", + "content": " first produces " + }, + { + "bbox": [ + 313, + 350, + 556, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{w} \\in \\mathcal{W}" + }, + { + "bbox": [ + 313, + 350, + 556, + 578 + ], + "type": "text", + "content": ". Learned affine transformations then specialize " + }, + { + "bbox": [ + 313, + 350, + 556, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 313, + 350, + 556, + 578 + ], + "type": "text", + "content": " to styles " + }, + { + "bbox": [ + 313, + 350, + 556, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{y} = (\\mathbf{y}_s, \\mathbf{y}_b)" + }, + { + "bbox": [ + 313, + 350, + 556, + 578 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 350, + 556, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_s" + }, + { + "bbox": [ + 313, + 350, + 556, + 578 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 350, + 556, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_b" + }, + { + "bbox": [ + 313, + 350, + 556, + 578 + ], + "type": "text", + "content": " denote the learned normalization scaling and bias coefficients, respectively. For the Super-Resolution Branch, the low-resolution image will first be processed by the Spatial Transformer Network (STN) [9] to align the spatially deformed texts. The convolution layer will extract features from the image processed by STN to obtain shallow image features. Then the shallow image features, text priors, and style information will be sent to several stacked style-content fusion blocks for style and content fusion. Each block contains a Style Enhancement Block, an adaptive instance normalization (AdaIN) [8], and a content embedding block. Finally, PixelShuffle is applied as an upsampling operator to generate " + }, + { + "bbox": [ + 313, + 350, + 556, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{SR}}" + }, + { + "bbox": [ + 313, + 350, + 556, + 578 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 586, + 457, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 457, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 457, + 598 + ], + "type": "text", + "content": "3.2. Style Enhancement Block" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 605, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 700 + ], + "type": "text", + "content": "The style of an image has been traditionally viewed as the correlation between different features, as demonstrated in the advancements of style transfer [6, 8, 12]. While previous approaches, such as [12], directly utilized ResNet features to obtain style vectors, our approach introduces a novel enhancement by applying channel-wise attention to these feature maps, thereby better capturing and preserving the stylistic characteristics in scene text images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "text", + "content": "Inspired by the Efficient Channel Attention (ECA)" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18695" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": "mechanism [29], we designed a Style Enhancement Block (SEB) to improve the model's capacity for capturing the correlation between different channel features. Let the input feature map be denoted as " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_f\\in \\mathbb{R}^{C\\times H\\times W}" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": ". We begin by applying Global Average Pooling (GAP) to the input feature map to obtain channel-wise statistics " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{m}\\in \\mathbb{R}^C" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": ". To further improve the model's capacity for capturing local cross-channel interactions at multiple scales, we employ multiple 1D convolution kernels of varying sizes, in contrast to the fixed-size kernels used in [29]. Specifically, let " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": " denote the number of different kernel sizes, each represented by " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "k_{l}" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "l\\in 1,2,\\ldots ,n" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": ". The convolutional operation at each scale can be expressed as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 125, + 238, + 296, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 238, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 125, + 238, + 296, + 281 + ], + "type": "interline_equation", + "content": "\\mathbf {m} _ {c} ^ {\\prime (l)} = \\sum_ {j = - \\frac {k _ {l}}{2}} ^ {\\frac {k _ {l}}{2}} b _ {j} ^ {l} \\mathbf {m} _ {c + j}, \\tag {1}", + "image_path": "d5dd232a1e7604eaf7163503e69616fe7a63df70fcb25fce905861848beab78e.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 290, + 296, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 290, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 296, + 376 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 290, + 296, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{b}^l\\in \\mathbb{R}^{k_l}" + }, + { + "bbox": [ + 55, + 290, + 296, + 376 + ], + "type": "text", + "content": " represents the weight vector for the convolution kernel of size " + }, + { + "bbox": [ + 55, + 290, + 296, + 376 + ], + "type": "inline_equation", + "content": "k_{l}" + }, + { + "bbox": [ + 55, + 290, + 296, + 376 + ], + "type": "text", + "content": ". This multi-scale approach allows the model to capture a more nuanced set of cross-channel dependencies, enhancing the representation of style information. The outputs from the multiple convolutional scales are aggregated by computing the final channel attention weight " + }, + { + "bbox": [ + 55, + 290, + 296, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{m}'_c" + }, + { + "bbox": [ + 55, + 290, + 296, + 376 + ], + "type": "text", + "content": " as a weighted sum:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 384, + 296, + 417 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 384, + 296, + 417 + ], + "spans": [ + { + "bbox": [ + 133, + 384, + 296, + 417 + ], + "type": "interline_equation", + "content": "\\mathbf {m} _ {c} ^ {\\prime} = \\sum_ {l = 1} ^ {n} \\alpha_ {l} \\mathbf {m} _ {c} ^ {\\prime (l)}, \\tag {2}", + "image_path": "4d03e309dabcec32b26998f80289208348dbfd3b2e0d6b338d691ef29fc5b231.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 426, + 296, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 462 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 426, + 296, + 462 + ], + "type": "inline_equation", + "content": "\\alpha_{l}" + }, + { + "bbox": [ + 55, + 426, + 296, + 462 + ], + "type": "text", + "content": " denotes the weighting coefficient for the output of the " + }, + { + "bbox": [ + 55, + 426, + 296, + 462 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 426, + 296, + 462 + ], + "type": "text", + "content": "-th convolutional scale. These weights are normalized using a Sigmoid function:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 121, + 472, + 295, + 490 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 472, + 295, + 490 + ], + "spans": [ + { + "bbox": [ + 121, + 472, + 295, + 490 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {f} ^ {\\prime c, i, j} = \\sigma \\left(\\mathbf {m} _ {c} ^ {\\prime}\\right) \\cdot \\mathbf {X} _ {f} ^ {c, i, j}, \\tag {3}", + "image_path": "134b63ed6be125f0357d97ae33b46ba1fc6c486ffdda3701ec9916d64035a0fe.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 497, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 295, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 497, + 295, + 533 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 55, + 497, + 295, + 533 + ], + "type": "text", + "content": " represents the Sigmoid activation function, and " + }, + { + "bbox": [ + 55, + 497, + 295, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{\\prime}" + }, + { + "bbox": [ + 55, + 497, + 295, + 533 + ], + "type": "text", + "content": " is the final output feature map with the applied channelwise attention." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 534, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 715 + ], + "type": "text", + "content": "The enhanced feature map is passed through several SEBs, which refines the correlation between different feature channels, ensuring that the model pays more attention to the features of important channels. This process effectively enhances the model's capability to capture and preserve the intricate stylistic details in scene text images. Subsequently, global average pooling (GAP) is applied to the final enhanced feature map to obtain a latent vector " + }, + { + "bbox": [ + 55, + 534, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 55, + 534, + 296, + 715 + ], + "type": "text", + "content": " in the latent space " + }, + { + "bbox": [ + 55, + 534, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 55, + 534, + 296, + 715 + ], + "type": "text", + "content": ". The latent vector is then processed by a non-linear style mapping network " + }, + { + "bbox": [ + 55, + 534, + 296, + 715 + ], + "type": "inline_equation", + "content": "f: \\mathcal{Z} \\to \\mathcal{W}" + }, + { + "bbox": [ + 55, + 534, + 296, + 715 + ], + "type": "text", + "content": ", which produces a vector " + }, + { + "bbox": [ + 55, + 534, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{w} \\in \\mathcal{W}" + }, + { + "bbox": [ + 55, + 534, + 296, + 715 + ], + "type": "text", + "content": " that decouples the latent space. The style mapping network consists of an 8-layer MLP with LeakyReLU activation functions. Finally, learned affine transformations are applied to transform " + }, + { + "bbox": [ + 55, + 534, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 55, + 534, + 296, + 715 + ], + "type": "text", + "content": " into style parameters " + }, + { + "bbox": [ + 55, + 534, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{y} = (\\mathbf{y}_s, \\mathbf{y}_b)" + }, + { + "bbox": [ + 55, + 534, + 296, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 72, + 455, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 455, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 455, + 84 + ], + "type": "text", + "content": "3.3. Style Content Fuse Block" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 89, + 555, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 209 + ], + "type": "text", + "content": "In the field of scene text image super-resolution, existing approaches primarily focus on embedding content information, including deep character category information derived from text probability priors. While this approach effectively guides the super-resolution network in restoring character structure, it often results in the loss of crucial stylistic features. To address this limitation, we propose a novel Style Content Fuse Block (SCFB) that allows the super-resolution network to restore character structures while preserving the stylistic consistency of the characters." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 210, + 556, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 210, + 556, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 210, + 556, + 354 + ], + "type": "text", + "content": "Our innovation lies in embedding style information directly into the super-resolution network alongside content information. This dual embedding approach ensures that both character structure and style are preserved during the super-resolution process. Specifically, for the shallow image features " + }, + { + "bbox": [ + 313, + 210, + 556, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_s\\in \\mathbb{R}^{C'\\times H'\\times W'}" + }, + { + "bbox": [ + 313, + 210, + 556, + 354 + ], + "type": "text", + "content": " obtained through a Spatial Transformer Network (STN) and a convolutional layer, we use Adaptive Instance Normalization (AdaIN) [8] to modulate their style. AdaIN has been shown to be effective in transferring style information across different domains. Given a feature map " + }, + { + "bbox": [ + 313, + 210, + 556, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_s" + }, + { + "bbox": [ + 313, + 210, + 556, + 354 + ], + "type": "text", + "content": " and a style vector " + }, + { + "bbox": [ + 313, + 210, + 556, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 313, + 210, + 556, + 354 + ], + "type": "text", + "content": ", the AdaIN operation can be defined as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 343, + 362, + 555, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 362, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 343, + 362, + 555, + 389 + ], + "type": "interline_equation", + "content": "\\operatorname {A d a I N} \\left(\\mathbf {X} _ {s}, \\mathbf {y}\\right) = \\mathbf {y} _ {s} \\left(\\frac {\\mathbf {X} _ {s} - \\mu \\left(\\mathbf {X} _ {s}\\right)}{\\sigma \\left(\\mathbf {X} _ {s}\\right)}\\right) + \\mathbf {y} _ {b}, \\tag {4}", + "image_path": "47b2d8a482e14f05914fae68f2616dbcc917bc6b57690593b19416c9b82460bd.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 399, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 555, + 483 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 399, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\mu (\\cdot)" + }, + { + "bbox": [ + 313, + 399, + 555, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 399, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 313, + 399, + 555, + 483 + ], + "type": "text", + "content": " represent the mean and variance of " + }, + { + "bbox": [ + 313, + 399, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_s" + }, + { + "bbox": [ + 313, + 399, + 555, + 483 + ], + "type": "text", + "content": ". Each AdaIN layer within the SCFB normalizes feature maps independently, with the dimensionality of " + }, + { + "bbox": [ + 313, + 399, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 313, + 399, + 555, + 483 + ], + "type": "text", + "content": " being twice the number of feature map channels " + }, + { + "bbox": [ + 313, + 399, + 555, + 483 + ], + "type": "inline_equation", + "content": "C^\\prime" + }, + { + "bbox": [ + 313, + 399, + 555, + 483 + ], + "type": "text", + "content": " in that layer. This embedding ensures that style information is seamlessly integrated into the spatial features of the image, thus preserving the character's stylistic attributes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 483, + 556, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 556, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 556, + 604 + ], + "type": "text", + "content": "In addition to embedding style information, content information, such as text probability distribution priors, should also be embedded into the super-resolution network to guide the learning of character structures. To bridge the domain gap between text priors and image features, we adopt the TP Interpreter architecture from [17] to extract the content embedding features. These interpreted features are then concatenated with the style-embedded features and subsequently passed through a Sequential Residual Block [30] for further refinement." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 611, + 408, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 611, + 408, + 624 + ], + "spans": [ + { + "bbox": [ + 314, + 611, + 408, + 624 + ], + "type": "text", + "content": "3.4. Text Style Loss" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "Although the proposed StyleSRN network effectively embeds the text's style features, these features are derived from the low-resolution image itself. When the style information in the low-resolution image is degraded, even if the network extracts these style features, it cannot fully reconstruct the style of the corresponding high-resolution image. Therefore, the reconstructed text image requires supervision at" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18696" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "the style level to effectively learn the degradation of style information in the low-resolution image." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "content": "To supervise the style in the reconstructed image, we introduce a Text Style Loss based on the Gram matrix. Unlike traditional methods [6] that use VGG networks to capture style information, our approach leverages a text recognizer to extract feature maps for Gram matrix computation. This enables better capture of text-specific style features, which are crucial for scene text images." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 180, + 296, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 296, + 266 + ], + "type": "text", + "content": "The Gram matrix " + }, + { + "bbox": [ + 55, + 180, + 296, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_l" + }, + { + "bbox": [ + 55, + 180, + 296, + 266 + ], + "type": "text", + "content": " for a given feature map " + }, + { + "bbox": [ + 55, + 180, + 296, + 266 + ], + "type": "inline_equation", + "content": "\\Phi_l(\\mathbf{I})" + }, + { + "bbox": [ + 55, + 180, + 296, + 266 + ], + "type": "text", + "content": " at the " + }, + { + "bbox": [ + 55, + 180, + 296, + 266 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 180, + 296, + 266 + ], + "type": "text", + "content": "-th layer is computed as follows. First, we reshape the feature map " + }, + { + "bbox": [ + 55, + 180, + 296, + 266 + ], + "type": "inline_equation", + "content": "\\Phi_l(\\mathbf{I})" + }, + { + "bbox": [ + 55, + 180, + 296, + 266 + ], + "type": "text", + "content": " of dimensions " + }, + { + "bbox": [ + 55, + 180, + 296, + 266 + ], + "type": "inline_equation", + "content": "C_l \\times H_l \\times W_l" + }, + { + "bbox": [ + 55, + 180, + 296, + 266 + ], + "type": "text", + "content": " into a matrix " + }, + { + "bbox": [ + 55, + 180, + 296, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_l" + }, + { + "bbox": [ + 55, + 180, + 296, + 266 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 55, + 180, + 296, + 266 + ], + "type": "inline_equation", + "content": "C_l \\times (H_l \\times W_l)" + }, + { + "bbox": [ + 55, + 180, + 296, + 266 + ], + "type": "text", + "content": ", where each row corresponds to a channel and each column corresponds to a pixel. The Gram matrix " + }, + { + "bbox": [ + 55, + 180, + 296, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_l" + }, + { + "bbox": [ + 55, + 180, + 296, + 266 + ], + "type": "text", + "content": " is the inner product between the different channels in layer " + }, + { + "bbox": [ + 55, + 180, + 296, + 266 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 180, + 296, + 266 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 129, + 273, + 295, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 273, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 129, + 273, + 295, + 306 + ], + "type": "interline_equation", + "content": "\\mathbf {G} _ {i j} ^ {l} = \\sum_ {k = 1} ^ {H _ {l} \\times W _ {l}} \\mathbf {F} _ {i k} ^ {l} \\mathbf {F} _ {j k} ^ {l} \\tag {5}", + "image_path": "0a86942600b3266f0a0249ea69a88369c220e50fdc042dc774386397a724ed19.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 315, + 296, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 296, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 296, + 351 + ], + "type": "text", + "content": "The Gram matrix captures the correlations between different channels in the feature map, thereby encoding the style information of the image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 351, + 296, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 296, + 399 + ], + "type": "text", + "content": "The Text Style Loss " + }, + { + "bbox": [ + 55, + 351, + 296, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{style}}" + }, + { + "bbox": [ + 55, + 351, + 296, + 399 + ], + "type": "text", + "content": " is defined as the Mean Squared Error (MSE) between the Gram matrix of the high-resolution (HR) ground truth and the super-resolved (SR) output:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 408, + 296, + 450 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 408, + 296, + 450 + ], + "spans": [ + { + "bbox": [ + 59, + 408, + 296, + 450 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s t y l e}} = \\sum_ {l = 1} ^ {L} \\frac {1}{4 C _ {l} ^ {2} H _ {l} ^ {2} W _ {l} ^ {2}} \\| \\mathbf {G} \\left(\\Phi_ {l} \\left(\\mathbf {I} _ {\\mathrm {H R}}\\right)\\right) - \\mathbf {G} \\left(\\Phi_ {l} \\left(\\mathbf {I} _ {\\mathrm {S R}}\\right)\\right) \\| _ {2} ^ {2} \\tag {6}", + "image_path": "7db69edb598e0c815c9c33efcecd3d453fc42ad98121d7598f8162bd58a4e629.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 451, + 296, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 296, + 499 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 451, + 296, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{G}(\\cdot)" + }, + { + "bbox": [ + 55, + 451, + 296, + 499 + ], + "type": "text", + "content": " denotes the Gram matrix, " + }, + { + "bbox": [ + 55, + 451, + 296, + 499 + ], + "type": "inline_equation", + "content": "\\Phi_l(\\cdot)" + }, + { + "bbox": [ + 55, + 451, + 296, + 499 + ], + "type": "text", + "content": " represents the feature map obtained from the " + }, + { + "bbox": [ + 55, + 451, + 296, + 499 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 451, + 296, + 499 + ], + "type": "text", + "content": "-th layer of the text recognizer, and " + }, + { + "bbox": [ + 55, + 451, + 296, + 499 + ], + "type": "inline_equation", + "content": "C_l, H_l, W_l" + }, + { + "bbox": [ + 55, + 451, + 296, + 499 + ], + "type": "text", + "content": " are the dimensions of the feature map at layer " + }, + { + "bbox": [ + 55, + 451, + 296, + 499 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 451, + 296, + 499 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 506, + 183, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 183, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 183, + 519 + ], + "type": "text", + "content": "3.5. Overall Loss Function" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 525, + 296, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 296, + 596 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 296, + 596 + ], + "type": "text", + "content": "In training, the overall loss function includes a Mean Square Error (MSE) loss " + }, + { + "bbox": [ + 55, + 525, + 296, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 55, + 525, + 296, + 596 + ], + "type": "text", + "content": ", a Text-Focus Loss " + }, + { + "bbox": [ + 55, + 525, + 296, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{TP}" + }, + { + "bbox": [ + 55, + 525, + 296, + 596 + ], + "type": "text", + "content": " [30], and the proposed Text Style Loss " + }, + { + "bbox": [ + 55, + 525, + 296, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{style}" + }, + { + "bbox": [ + 55, + 525, + 296, + 596 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 55, + 525, + 296, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 55, + 525, + 296, + 596 + ], + "type": "text", + "content": " measures the difference between the super-resolution (SR) output and the ground-truth high-resolution (HR) image. The overall loss function is described as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 117, + 606, + 295, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 606, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 117, + 606, + 295, + 620 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {2} + \\alpha \\mathcal {L} _ {T P} + \\beta \\mathcal {L} _ {\\text {s t y l e}} \\tag {7}", + "image_path": "a1cd767c4fdeed308624018c111d1eb205398c8e469453a1701880315dd2c78e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 629, + 253, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 253, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 253, + 642 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 55, + 629, + 253, + 642 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 629, + 253, + 642 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 629, + 253, + 642 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 629, + 253, + 642 + ], + "type": "text", + "content": " are the balancing parameters." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 652, + 137, + 665 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 652, + 137, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 652, + 137, + 665 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 672, + 118, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 118, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 118, + 683 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "TextZoom TextZoom [30] is a scene text super-resolution dataset designed for real-world scenarios, containing" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "21,740 pairs of low-resolution and high-resolution images captured by cameras with various focal lengths in natural environments, with 17,367 samples used for training. The test set is divided into easy, medium, and hard subsets, comprising 1,619, 1,411, and 1,343 LR-HR pairs, respectively, based on the camera focal lengths." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 161, + 555, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 161, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 161, + 555, + 258 + ], + "type": "text", + "content": "Scene Text Recognition Datasets To assess the robustness and generalization capability of StyleSRN, we conducted comprehensive evaluations on five benchmark scene text recognition datasets: IC13 [10], IC15 [11], CUTE80 [24], IIIT5K [19], and SVTP [22]. These datasets were manually degraded to create low-resolution counterparts. Detailed descriptions of the datasets and degradation procedures are provided in the supplementary materials." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 266, + 369, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 266, + 369, + 277 + ], + "spans": [ + { + "bbox": [ + 314, + 266, + 369, + 277 + ], + "type": "text", + "content": "4.2. Metric" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 284, + 554, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 284, + 554, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 554, + 380 + ], + "type": "text", + "content": "Following previous works, we adopt two types of metrics: super-resolution image text recognition accuracy and image quality. Specifically, to evaluate the generalization capability of StyleSRN, we employed three scene text recognizers: CRNN [25], MORAN [16], and ASTER [26]. For the image quality metric, we adopt the widely used Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM)[31]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 388, + 447, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 447, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 447, + 401 + ], + "type": "text", + "content": "4.3. Implementation Details" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 406, + 555, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 555, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 555, + 551 + ], + "type": "text", + "content": "We implement our model using PyTorch. All experiments are conducted on a single RTX 3090 GPU. We train our model with a batch size of 64 for 500 epochs using AdamW [15] for optimization. The learning rate is set to 1e-3 and decayed by a factor of 0.5 after 400 epochs. The number of SEBs and SCFBs are set to 3 and 5, respectively. In the SEBs, the sizes of the convolution kernels are " + }, + { + "bbox": [ + 313, + 406, + 555, + 551 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 313, + 406, + 555, + 551 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 406, + 555, + 551 + ], + "type": "inline_equation", + "content": "5 \\times 5" + }, + { + "bbox": [ + 313, + 406, + 555, + 551 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 406, + 555, + 551 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 313, + 406, + 555, + 551 + ], + "type": "text", + "content": ", respectively. We use CRNN to generate the features for computing the text style loss and empirically set the number of feature maps L used to calculate the style loss to 5. We set the parameters " + }, + { + "bbox": [ + 313, + 406, + 555, + 551 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 406, + 555, + 551 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 406, + 555, + 551 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 406, + 555, + 551 + ], + "type": "text", + "content": " to 0.5 and 1.0, respectively." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 559, + 408, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 408, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 408, + 571 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 576, + 554, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 554, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 554, + 661 + ], + "type": "text", + "content": "In this section, we conduct ablation studies to investigate the effectiveness of SEBs and SCFBs, the effectiveness of training with Text Style Loss, the impact of numbers of SEBs and SCFBs. For STISR methods that do not use text priors, such as TSRN [30] and TBSRN [2], using SCFB means that image features are not used for text prior fusion, but are directly concatenated with style-embedded features." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "Effectiveness of SEBs and SCFBs. We compared the performance of different STISR methods using three strategies: No Style Embedding (NSE), only SCFB, and SCFB" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18697" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 71, + 293, + 232 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 293, + 232 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 293, + 232 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 293, + 232 + ], + "type": "table", + "html": "
MethodNSESCFBSEBImage QualityAccuracy(%)
PSNRSSIM
TSRN [30]✓--21.420.769041.4
-✓-21.540.769943.5
-✓✓21.560.769544.4
TBSRN [2]✓--20.910.760348.1
-✓-21.040.762350.3
-✓✓21.220.761451.1
TPGSR [18]✓--20.970.771951.8
-✓-21.040.772352.4
-✓✓21.170.772352.7
TATT [17]✓--21.520.793052.6
-✓-21.660.779054.1
-✓✓21.870.778554.9
LEMMA [7]✓--20.880.776056.3
-✓-20.940.777156.7
-✓✓21.050.776557.0
", + "image_path": "62c6d87eefd0af27d741d4506ba7050e28460a2f36f7f5e2a961505de7ff97fe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 308, + 294, + 399 + ], + "blocks": [ + { + "bbox": [ + 55, + 243, + 295, + 288 + ], + "lines": [ + { + "bbox": [ + 55, + 243, + 295, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 295, + 288 + ], + "type": "text", + "content": "Table 1. Effectiveness of the Style Content Fuse Block (SCFB) and the Style Enhancement Block (SEB). NSE refers to No Style Embedding. Accuracy represents the average accuracy with CRNN [25]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 308, + 294, + 399 + ], + "lines": [ + { + "bbox": [ + 58, + 308, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 58, + 308, + 294, + 399 + ], + "type": "table", + "html": "
Method\\( \\mathcal{L}_{\\text{style}} \\)PSNRSSIMAccuracy(%)
TSRN [30]-21.420.769041.4
✓21.560.771143.7
TPGSR [18]-20.970.771951.8
✓21.120.772452.4
TATT [17]-21.520.793052.6
✓21.810.793853.8
StyleSRN-21.540.775854.9
✓21.820.777857.4
", + "image_path": "9de9154bdec70af2fdfa5535f23f27e13fc1b523de33df9dc5ac888e56c6ba8b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 461, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 461, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 461, + 295, + 713 + ], + "type": "text", + "content": "combined with the SEB in Table 2. NSE means removing both SEB and SCFB. SCFB can be used independently of SEB, but when using SEB, SCFB must be used to embed style information. Removing SEB means that the features extracted by Resnet18 are directly fed into the style mapping network through global pooling. Removing SCFB means that the features directly passed through the text interpreter do not need to be cascaded with the style features. As presented in Table 1, using only SCFB generally improves image quality and recognition accuracy over NSE. In addition, SCFB has a greater improvement over methods without text priors such as TSRN [30] and TBSRN [2]. For instance, with TSRN, the introduction of SCFB increases accuracy by " + }, + { + "bbox": [ + 54, + 461, + 295, + 713 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 54, + 461, + 295, + 713 + ], + "type": "text", + "content": ", along with notable gains in PSNR and SSIM. We speculate that this may be because these methods do not have any prior guidance, while SCFB introduces the guidance of style information. Using SCFB and SEB at the same time can effectively improve PSNR and text recognition accuracy, but slightly reduce SSIM. This shows that SEB can effectively enhance style information, but slightly reduce structural consistency." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 317, + 71, + 553, + 104 + ], + "blocks": [ + { + "bbox": [ + 55, + 409, + 295, + 433 + ], + "lines": [ + { + "bbox": [ + 55, + 409, + 295, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 295, + 433 + ], + "type": "text", + "content": "Table 2. Ablation of the Text Style Loss. Accuracy represents the average accuracy with CRNN [25]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 71, + 553, + 104 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 553, + 104 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 553, + 104 + ], + "type": "table", + "html": "
DiscriminatorPSNRSSIMAccuracy (%)
VGG21.310.776156.3
CRNN21.820.777857.4
", + "image_path": "7efdac0064330c72787acbd80c757b82bfbdeb595ad21af9fa30b17f43f22689.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 143, + 553, + 227 + ], + "blocks": [ + { + "bbox": [ + 324, + 114, + 544, + 126 + ], + "lines": [ + { + "bbox": [ + 324, + 114, + 544, + 126 + ], + "spans": [ + { + "bbox": [ + 324, + 114, + 544, + 126 + ], + "type": "text", + "content": "Table 3. Ablation of the discriminator used in text style loss." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 143, + 553, + 227 + ], + "lines": [ + { + "bbox": [ + 317, + 143, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 143, + 553, + 227 + ], + "type": "table", + "html": "
NQuality MetricAccuracy(%)
PSNRSSIMCRNN [25]MORAN [16]ASTER [26]
020.760.761854.661.865.0
121.820.777856.363.366.2
221.770.776456.964.066.7
421.770.777057.164.067.0
521.730.776855.963.266.7
321.820.777857.464.167.3
", + "image_path": "47f37610fdc5dfce9a78865b669f5cde2e7686d4eb63a1386a2e01ad8aa2c2f8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 266, + 553, + 350 + ], + "blocks": [ + { + "bbox": [ + 319, + 237, + 548, + 248 + ], + "lines": [ + { + "bbox": [ + 319, + 237, + 548, + 248 + ], + "spans": [ + { + "bbox": [ + 319, + 237, + 548, + 248 + ], + "type": "text", + "content": "Table 4. Ablation of the number of Style Enhancement Blocks." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 266, + 553, + 350 + ], + "lines": [ + { + "bbox": [ + 317, + 266, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 317, + 266, + 553, + 350 + ], + "type": "table", + "html": "
NQuality MetricAccuracy(%)
PSNRSSIMCRNN [25]MORAN [16]ASTER [26]
020.670.743153.261.064.7
321.450.772255.362.366.2
421.430.775156.763.966.9
621.740.776956.964.067.0
721.550.775656.062.866.2
521.820.777857.464.167.3
", + "image_path": "48fb7d9208a9e7794991f42c3dcb69f9925c8e1c28824c53ca0fe2a80e94ba65.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 319, + 359, + 547, + 370 + ], + "lines": [ + { + "bbox": [ + 319, + 359, + 547, + 370 + ], + "spans": [ + { + "bbox": [ + 319, + 359, + 547, + 370 + ], + "type": "text", + "content": "Table 5. Ablation of the number of Style Content Fuse Blocks." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 397, + 555, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 555, + 553 + ], + "type": "text", + "content": "Effectiveness of training with Text Style Loss. We show the effectiveness of the Text Style Loss in Table 2. Incorporating the Text Style Loss results in improvements across all metrics. For instance, the PSNR of TSRN increases from 21.42 to 21.56, and its SSIM improves from 0.7690 to 0.7711. In terms of text recognition accuracy, models incorporating the Text Style Loss demonstrate enhanced performance. The accuracy of TSRN improves from " + }, + { + "bbox": [ + 313, + 397, + 555, + 553 + ], + "type": "inline_equation", + "content": "41.4\\%" + }, + { + "bbox": [ + 313, + 397, + 555, + 553 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 397, + 555, + 553 + ], + "type": "inline_equation", + "content": "43.7\\%" + }, + { + "bbox": [ + 313, + 397, + 555, + 553 + ], + "type": "text", + "content": ". These results highlight that the Text Style Loss effectively enhances both image quality and text recognition accuracy. Besides, we also compared the impact of using different discriminators in Table 3. It can be seen that using CRNN performs better than VGG." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "Numbers of SEBs. We show the effect of the number of SEBs in Table 4. As the number of SEBs increases, there is a noticeable improvement in both the PSNR and SSIM metrics. Specifically, using 3 SEBs achieves the highest PSNR of 21.82 and an SSIM of 0.7778, indicating the optimal enhancement of image quality. In terms of text recognition accuracy, models incorporating 3 SEBs also demonstrate superior performance across all recognition models. These findings suggest that SEBs significantly enhance both image quality and text recognition performance, with 3 SEBs being the optimal configuration." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18698" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 71, + 555, + 222 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 555, + 222 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 555, + 222 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 555, + 222 + ], + "type": "table", + "html": "
MethodAccuracy of CRNN [25] (%)Accuracy of MORAN [16] (%)Accuracy of ASTER [26] (%)
EasyMediumHardAvgEasyMediumHardAvgEasyMediumhardAvg
BICUBIC36.421.121.126.860.637.930.844.167.442.431.248.2
HR76.475.164.672.491.285.374.284.194.287.776.286.6
TSRN [30]52.538.231.441.470.155.337.955.475.156.340.158.3
TBSRN [2]59.647.135.348.174.157.040.858.475.759.941.660.1
TG [3]61.247.635.548.975.857.841.459.477.960.242.461.3
TPGSR [18]63.152.038.651.874.960.544.160.578.962.744.562.8
TATT [17]62.653.439.852.672.660.243.159.578.963.445.463.6
C3-STISR [34]65.253.639.853.774.261.043.259.579.163.346.864.1
TSAN [37]64.653.338.853.078.461.345.162.779.664.145.364.1
LEMMA [7]67.158.840.656.377.764.444.663.281.166.347.466.0
DPMN [36]64.454.239.253.473.361.543.960.479.364.145.263.9
TCDM [20]67.357.342.755.777.662.945.962.281.365.150.165.5
RTSRN [33]67.059.242.657.077.163.346.563.280.466.149.166.2
StyleSRN68.159.442.457.478.665.145.764.182.767.448.767.3
", + "image_path": "94c83dbd65ba253d2744496b44701ec1a6bdf20dae78373e964f5f5550bd6ccf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 267, + 555, + 360 + ], + "blocks": [ + { + "bbox": [ + 55, + 232, + 555, + 257 + ], + "lines": [ + { + "bbox": [ + 55, + 232, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 555, + 257 + ], + "type": "text", + "content": "Table 6. Comparison with the existing methods in terms of the recognition accuracy on TextZoom[30]. BICUBIC means LR images are directly upsampled by the bicubic interpolation, and the same definition applies to BICUBIC in subsequent tables." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 267, + 555, + 360 + ], + "lines": [ + { + "bbox": [ + 58, + 267, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 58, + 267, + 555, + 360 + ], + "type": "table", + "html": "
MethodLossPSNRSSIM
EasyMediumHardAvgEasyMediumHardAvg
BICUBIC-22.3518.9819.3920.350.78840.62540.65920.6961
TSRN [30]\\( \\mathcal{L}_2 + \\mathcal{L}_{GP} \\)25.0718.8619.7121.420.88970.66760.73020.7690
TBSRN [2]\\( \\mathcal{L}_{POS} + \\mathcal{L}_{CON} \\)23.4619.1719.6820.910.87290.64550.74520.7603
TPGSR [18]\\( \\mathcal{L}_2 + \\mathcal{L}_{TP} \\)23.7318.6820.0620.970.88050.67380.74400.7719
TATT [17]\\( \\mathcal{L}_2 + \\mathcal{L}_{TP} + \\mathcal{L}_{TSC} \\)24.7219.0220.3121.520.90060.69110.77030.7930
LEMMA [7]\\( \\mathcal{L}_2 + \\mathcal{L}_{TP} + \\mathcal{L}_{finetune} \\)23.7019.3719.8420.880.87530.69050.74910.7760
StyleSRN\\( \\mathcal{L}_2 + \\mathcal{L}_{TP} + \\mathcal{L}_{style} \\)24.6919.7020.5821.820.88440.67910.75310.7778
", + "image_path": "d5b2ac58a64030dc4aa9b6d0267f3eea58579c0bbb46c4f227d6307ab9a2b855.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 403, + 295, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 403, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 54, + 403, + 295, + 559 + ], + "type": "text", + "content": "Numbers of SCFBs. We further examined the impact of the numbers of SCFBs, as detailed in Table 5. While increasing the number of SCFBs improves recognition accuracy metrics, diminishing returns are observed beyond a certain point. The optimal performance, consistent with the SEB ablation study, is achieved with 5 SCFBs, yielding a PSNR of 21.82 and an SSIM of 0.7778. Moreover, text recognition accuracy peaks with CRNN (57.40%), MORAN (64.14%), and ASTER (67.31%) when 5 SCFBs are utilized. This suggests that while the SCFB is essential for balancing style and content, its optimal configuration involves a moderate number of blocks, beyond which performance gains begin to plateau." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 564, + 242, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 242, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 242, + 577 + ], + "type": "text", + "content": "4.5. Comparison with State-of-the-Arts" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "Results on TextZoom. We evaluate the recognition accuracy of text images generated by various STISR methods using three text recognizers: CRNN [25], MORAN [16], and ASTER [26]. As shown in Table 7, our proposed StyleSRN consistently outperforms state-of-the-art methods across all recognizers, achieving average accuracy rates of " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "57.4\\%" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " with CRNN, " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "64.1\\%" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " with MORAN, and " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "67.3\\%" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " with ASTER. On the hard subset, StyleSRN performs slightly lower than some methods, such as TCDM and DPMN, likely due to challenges in balancing structure and style preservation in extremely blurred or occluded" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 402, + 552, + 494 + ], + "blocks": [ + { + "bbox": [ + 102, + 370, + 506, + 382 + ], + "lines": [ + { + "bbox": [ + 102, + 370, + 506, + 382 + ], + "spans": [ + { + "bbox": [ + 102, + 370, + 506, + 382 + ], + "type": "text", + "content": "Table 7. Comparison of the image quality on TextZoom [30]. Avg represents the average value of three subsets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 402, + 552, + 494 + ], + "lines": [ + { + "bbox": [ + 317, + 402, + 552, + 494 + ], + "spans": [ + { + "bbox": [ + 317, + 402, + 552, + 494 + ], + "type": "table", + "html": "
MethodAccuracy(%)
IC13IC15CT80SVTPIII5K
BICUBIC76.1638.2155.2135.5059.70
TSRN [30]75.6741.5254.8635.9762.60
TBSRN [2]76.9543.0754.8637.3663.60
TPGSR [18]77.6442.8553.4738.2962.07
TATT [17]76.0648.4851.0442.0262.67
LEMMA [7]79.8049.7051.0443.7265.07
StyleSRN80.9950.4755.5646.5165.63
", + "image_path": "64528ca3edaa26e18866c013f18ef5688dd20732b1faf36c7416b59ef515ca05.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 505, + 555, + 538 + ], + "lines": [ + { + "bbox": [ + 313, + 505, + 555, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 555, + 538 + ], + "type": "text", + "content": "Table 8. Comparison of the accuracy on manually degrade scene text recognition datasets. Accuracy represents the average accuracy with CRNN [25]." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "cases. In terms of image quality, as shown in Table 7, our method achieves the highest average PSNR of 21.82, surpassing other methods like TATT (21.52) and LEMMA (20.93). Our method achieves an SSIM of 0.7778, which is comparable to TATT (0.7930). The visual comparison is presented in Figure 3, which demonstrates that our method achieves superior visual quality. Our method not only restores the structure of the characters but also preserves their style information. For instance, for certain stereoscopic scene texts shown in Figure 3, our method more effectively restores their stereoscopic effect and shadows details." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18699" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 70, + 541, + 265 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 541, + 265 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 541, + 265 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 541, + 265 + ], + "type": "image", + "image_path": "4e0fd2032e03f8e5ea2b61d52496d82f99a05df52b1fa502eb7b905fe010bd94.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 273, + 555, + 296 + ], + "lines": [ + { + "bbox": [ + 55, + 273, + 555, + 296 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 555, + 296 + ], + "type": "text", + "content": "Figure 3. Visualization of SR images and their recognition result. Each text image below uses CRNN [25] for recognition. Characters marked in red indicate incorrect recognition results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 62, + 314, + 289, + 479 + ], + "blocks": [ + { + "bbox": [ + 62, + 314, + 289, + 479 + ], + "lines": [ + { + "bbox": [ + 62, + 314, + 289, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 314, + 289, + 479 + ], + "type": "image", + "image_path": "48408a6d7957e9693abe8ca9efcdf347a6e7fce11c38ad672c5cb1affa70758d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 491, + 295, + 512 + ], + "lines": [ + { + "bbox": [ + 55, + 491, + 295, + 512 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 295, + 512 + ], + "type": "text", + "content": "Figure 4. Visualization of the STISR results on scene text with uneven illumination and scene text with severe blur." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "Generalization to recognition datasets. To further evaluate the generalization capability and robustness of the proposed StyleSRN, we also conduct experiments on five scene text recognition datasets. These datasets cover a broad spectrum of image styles and label distributions, presenting substantial challenges for the generalization performance of super-resolution networks. Due to the absence of paired low-resolution and high-resolution images in these datasets, we manually degrad the original images before applying StyleSRN for preprocessing. As illustrated in Table 2, our method consistently surpasses previous approaches across these datasets, confirming its strong generalization ability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 316, + 384, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 316, + 384, + 327 + ], + "spans": [ + { + "bbox": [ + 314, + 316, + 384, + 327 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 335, + 555, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 555, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 555, + 538 + ], + "type": "text", + "content": "While StyleSRN effectively improves the structural integrity and style consistency of STISR, as shown in Figure 4, it struggles to recover character structures in images with uneven illumination or severe blur, highlighting the need for further refinement to address degraded visual conditions. Another limitation is the absence of a standardized quantitative metric for evaluating style degradation in STISR methods. While our method restores the style of low-resolution scene text images, no existing metric specifically measures style preservation or degradation. This lack of objective criteria makes it difficult to rigorously assess style restoration performance, as current evaluations are largely subjective and based on visual inspection. Future research should focus on developing standardized, quantitative metrics to objectively evaluate style preservation in STISR, enabling more precise assessments and meaningful comparisons across methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 550, + 392, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 392, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 392, + 562 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 570, + 555, + 713 + ], + "type": "text", + "content": "In this work, we propose StyleSRN to address the limitations of current STISR methods that neglect the stylistic details of scene text images. StyleSRN incorporates Style Enhancement Blocks and Style Content Fuse Blocks to effectively capture and integrate text style with content, ensuring both structural integrity and stylistic consistency. Additionally, the proposed Text Style Loss supervises reconstruction at the style level, which effectively enhances both image quality and text recognition accuracy. Results on TextZoom and five STR benchmarks demonstrate that StyleSRN consistently enhances downstream recognition accuracy and image quality." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18700" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 170, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 170, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 170, + 85 + ], + "type": "text", + "content": "7. Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 179 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China (No.61502164, 62176097), the Natural Science Foundation of Hunan Province (No.2020JJ4057, 2024JJ10034), the Key Research and Development Program of Changsha Science and Technology Bureau (No.kq2004050), and the Scientific Research Foundation of the Education Department of Hunan Province of China (No.21A0052)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 199, + 115, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 199, + 115, + 211 + ], + "spans": [ + { + "bbox": [ + 56, + 199, + 115, + 211 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 219, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 219, + 296, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 219, + 296, + 273 + ], + "spans": [ + { + "bbox": [ + 61, + 219, + 296, + 273 + ], + "type": "text", + "content": "[1] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE International Conference on Computer Vision, pages 2425-2433, 2015. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 274, + 296, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 274, + 296, + 328 + ], + "spans": [ + { + "bbox": [ + 61, + 274, + 296, + 328 + ], + "type": "text", + "content": "[2] Jingye Chen, Bin Li, and Xiangyang Xue. Scene text telescope: Text-focused scene image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12026-12035, 2021. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 330, + 295, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 330, + 295, + 372 + ], + "spans": [ + { + "bbox": [ + 62, + 330, + 295, + 372 + ], + "type": "text", + "content": "[3] Jingye Chen, Haiyang Yu, Jianqi Ma, Bin Li, and Xiangyang Xue. Text gestalt: Stroke-aware scene text image superresolution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 285-293, 2022. 2, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 373, + 294, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 373, + 294, + 416 + ], + "spans": [ + { + "bbox": [ + 62, + 373, + 294, + 416 + ], + "type": "text", + "content": "[4] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In Proceedings of the European Conference on Computer Vision, pages 184-199, 2014. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 418, + 295, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 418, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 62, + 418, + 295, + 459 + ], + "type": "text", + "content": "[5] Chao Dong, Ximei Zhu, Yubin Deng, Chen Change Loy, and Yu Qiao. Boosting optical character recognition: A superresolution approach. arXiv preprint arXiv:1506.02211, 2015.2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 460, + 294, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 460, + 294, + 504 + ], + "spans": [ + { + "bbox": [ + 62, + 460, + 294, + 504 + ], + "type": "text", + "content": "[6] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2414-2423, 2016. 2, 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 506, + 294, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 506, + 294, + 559 + ], + "spans": [ + { + "bbox": [ + 62, + 506, + 294, + 559 + ], + "type": "text", + "content": "[7] Hang Guo, Tao Dai, Guanghao Meng, and Shu-Tao Xia. Towards robust scene text image super-resolution via explicit location enhancement. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 782–790, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 560, + 294, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 560, + 294, + 603 + ], + "spans": [ + { + "bbox": [ + 62, + 560, + 294, + 603 + ], + "type": "text", + "content": "[8] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE International Conference on Computer Vision, pages 1501-1510, 2017. 2, 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 604, + 294, + 636 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 604, + 294, + 636 + ], + "spans": [ + { + "bbox": [ + 62, + 604, + 294, + 636 + ], + "type": "text", + "content": "[9] Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. Proceedings of the Advances in Neural Information Processing Systems, 28, 2015. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 637, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 637, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 637, + 295, + 712 + ], + "type": "text", + "content": "[10] Dimosthenis Karatzas, Faisal Shafait, Seiichi Uchida, Masakazu Iwamura, Lluis Gomez i Bigorda, Sergi Robles Mestre, Joan Mas, David Fernandez Mota, Jon Almazan Almazan, and Lluis Pere De Las Heras. Icdar 2013 robust reading competition. In Proceedings of the 12th International Conference on Document Analysis and Recognition, pages 1484-1493, 2013. 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 148 + ], + "type": "text", + "content": "[11] Dimosthenis Karatzas, Lluis Gomez-Bigorda, Anguelos Nicolaou, Suman Ghosh, Andrew Bagdanov, Masakazu Iwamura, Jiri Matas, Lukas Neumann, Vijay Ramaseshan Chandrasekhar, Shijian Lu, et al. Icdar 2015 competition on robust reading. In Proceedings of the 13th International Conference on Document Analysis and Recognition, pages 1156-1160, 2015. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 150, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 150, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 150, + 553, + 183 + ], + "type": "text", + "content": "[12] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks, 2019. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 184, + 553, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 553, + 238 + ], + "type": "text", + "content": "[13] Vijeta Khare, Palaiahnakote Shivakumara, Chee Seng Chan, Tong Lu, Liang Kim Meng, Hon Hock Woon, and Michael Blumenstein. A novel character segmentation-reconstruction approach for license plate recognition. Expert Systems with Applications, 131:219-239, 2019. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 239, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 315 + ], + "type": "text", + "content": "[14] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4681-4690, 2017. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 316, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 338 + ], + "type": "text", + "content": "[15] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 338, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 370 + ], + "type": "text", + "content": "[16] Canjie Luo, Lianwen Jin, and Zenghui Sun. Moran: A multi-object rectified attention network for scene text recognition. Pattern Recognition, 90:109-118, 2019. 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 372, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 425 + ], + "type": "text", + "content": "[17] Jianqi Ma, Zhetong Liang, and Lei Zhang. A text attention network for spatial deformation robust scene text image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5911-5920, 2022. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 426, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 459 + ], + "type": "text", + "content": "[18] Jianqi Ma, Shi Guo, and Lei Zhang. Text prior guided scene text image super-resolution. IEEE Transactions on Image Processing, 32:1341-1353, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 460, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 502 + ], + "type": "text", + "content": "[19] Anand Mishra, Karteek Alahari, and CV Jawahar. Scene text recognition using higher order language priors. In Proceedings of the British Machine Vision Conference, pages 1-11, 2012. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 504, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 504, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 317, + 504, + 553, + 557 + ], + "type": "text", + "content": "[20] Chihiro Noguchi, Shun Fukuda, and Masao Yamanaka. Scene text image super-resolution based on text-conditional diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1474–1484, 2024. 2, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 559, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 602 + ], + "type": "text", + "content": "[21] Ram Krishna Pandey, K Vignesh, AG Ramakrishnan, et al. Binary document image super resolution for improved readability andOCR performance. arXiv preprint arXiv:1812.02475, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 603, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 657 + ], + "type": "text", + "content": "[22] Trung Quy Phan, Palaiahnakote Shivakumara, Shangxuan Tian, and Chew Lim Tan. Recognizing text with perspective distortion in natural scenes. In Proceedings of the IEEE International Conference on Computer Vision, pages 569-576, 2013. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[23] Yadong Qu, Qingfeng Tan, Hongtao Xie, Jianjun Xu, Yuxin Wang, and Yongdong Zhang. Exploring stroke-level modifications for scene text editing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2119-2127, 2023. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18701" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[24] Anhar Risnumawan, Palaiahankote Shivakumara, Chee Seng Chan, and Chew Lim Tan. A robust arbitrary text detection system for natural scene images. Expert Systems with Applications, 41(18):8027-8048, 2014. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "type": "text", + "content": "[25] Baoguang Shi, Xiang Bai, and Cong Yao. An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(11):2298-2304, 2016. 5, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 294, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 226 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 226 + ], + "type": "text", + "content": "[26] Baoguang Shi, Mingkun Yang, Xinggang Wang, Pengyuan Lyu, Cong Yao, and Xiang Bai. Aster: An attentional scene text recognizer with flexible rectification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(9):2035-2048, 2018. 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 259 + ], + "type": "text", + "content": "[27] Tonghua Su, Fuxiang Yang, Xiang Zhou, Donglin Di, Zhongjie Wang, and Songze Li. Scene style text editing, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "type": "text", + "content": "[28] Hanh TM Tran and Tien Ho-Phuoc. Deep laplacian pyramid network for text images super-resolution. In 2019 IEEE-RIVF International Conference on Computing and Communication Technologies (RIVF), pages 1-6, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 305, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 294, + 360 + ], + "type": "text", + "content": "[29] Qilong Wang, Banggu Wu, Pengfei Zhu, Peihua Li, Wangmeng Zuo, and Qinghua Hu. Eca-net: Efficient channel attention for deep convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11534-11542, 2020. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 361, + 294, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 294, + 414 + ], + "type": "text", + "content": "[30] Wenjia Wang, Enze Xie, Xuebo Liu, Wenhai Wang, Ding Liang, Chunhua Shen, and Xiang Bai. Scene text image super-resolution in the wild. In Proceedings of the European Conference on Computer Vision, pages 650-666, 2020. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 415, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 458 + ], + "type": "text", + "content": "[31] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "type": "text", + "content": "[32] Chongsheng Zhang, Weiping Ding, Guowen Peng, Feifei Fu, and Wei Wang. Street view text recognition with deep learning for urban scene understanding in intelligent transportation systems. IEEE Transactions on Intelligent Transportation Systems, 22(7):4727-4743, 2020. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 514, + 294, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 294, + 579 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 294, + 579 + ], + "type": "text", + "content": "[33] Wenyu Zhang, Xin Deng, Baojun Jia, Xingtong Yu, Yifan Chen, Jin Ma, Qing Ding, and Xinming Zhang. Pixel adapter: A graph-based post-processing approach for scene text image super-resolution. In Proceedings of the 31st ACM International Conference on Multimedia, pages 2168-2179, 2023. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 581, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 294, + 635 + ], + "type": "text", + "content": "[34] Minyi Zhao, Miao Wang, Fan Bai, Bingjia Li, Jie Wang, and Shuigeng Zhou. C3-STISR: scene text image superresolution with triple clues. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, pages 1707–1713, 2022. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "type": "text", + "content": "[35] Yuxuan Zhou, Liangcai Gao, Zhi Tang, and Baole Wei. Recognition-guided diffusion model for scene text image super-resolution. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, pages 2940-2944, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[36] Shipeng Zhu, Zuoyan Zhao, Pengfei Fang, and Hui Xue. Improving scene text image super-resolution via dual prior" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 149 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "modulation network. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3843-3851, 2023. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 555, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 149 + ], + "type": "text", + "content": "[37] Xiangyuan Zhu, Kehua Guo, Hui Fang, Rui Ding, Zheng Wu, and Gerald Schaefer. Gradient-based graph attention for scene text image super-resolution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3861-3869, 2023. 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18702" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Stylized-Face_ A Million-level Stylized Face Dataset for Face Recognition/fc351eb8-6a0f-465e-ac2f-6fa6da65b294_content_list.json b/2025/Stylized-Face_ A Million-level Stylized Face Dataset for Face Recognition/fc351eb8-6a0f-465e-ac2f-6fa6da65b294_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4b1fa7e0b26fca7ebacefd173e5e5c7fd5eb6f58 --- /dev/null +++ b/2025/Stylized-Face_ A Million-level Stylized Face Dataset for Face Recognition/fc351eb8-6a0f-465e-ac2f-6fa6da65b294_content_list.json @@ -0,0 +1,2250 @@ +[ + { + "type": "text", + "text": "Stylized-Face: A Million-level Stylized Face Dataset for Face Recognition", + "text_level": 1, + "bbox": [ + 130, + 130, + 866, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhengyuan Peng $^{1*}$ , Jianqing Xu $^{2*}$ , Yuge Huang $^{2\\dagger}$ , Jinkun Hao $^{1}$ ,", + "bbox": [ + 243, + 179, + 764, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shouhong Ding $^{2}$ , Zhizhong Zhang $^{3}$ , Xin Tan $^{3}$ , Lizhuang Ma $^{1,3\\dagger}$", + "bbox": [ + 243, + 198, + 764, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Shanghai Jiao Tong University, $^{2}$ Youtu Lab, Tencent, $^{3}$ East China Normal University", + "bbox": [ + 161, + 215, + 834, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Stylized face recognition is the task of recognizing generated faces with the same ID across diverse stylistic domains (e.g., anime, painting, cyberpunk styles). This emerging field plays a vital role in the governance of generative image, serving the primary objective: Recognize the ID information of stylized faces to detect potential infringements of portrait rights. Despite its importance, progress in stylized face recognition has been hindered by the lack of large-scale, stylistically diverse datasets. To address this gap, we introduce the Stylized-Face dataset, which is the first dataset specifically designed for stylized face recognition. Stylized-Face dataset includes 4.6 million images across 62k IDs, specifically curated to enhance model performance in stylized face recognition tasks. To ensure data quality (i.e., ID preservation) at this massive scale, we implement a semi-automated pipeline for large-scale data cleaning. Based on the Stylized-Face dataset, we establish three benchmarks to evaluate the robustness and generalization of recognition models across various scenarios, including within-distribution performance, cross-prompt generalization, and cross-method generalization, which target key challenges in stylized face recognition. Experimental results demonstrate that models trained on Stylized-Face achieve remarkable improvements in both stylized face recognition performance (a $15.9\\%$ improvement in TAR at $\\mathrm{FAR} = 1\\mathrm{e - }4$ ) and generalization (a $13.3\\%$ improvement in TAR at $\\mathrm{FAR} = 1\\mathrm{e - }3$ in cross-method generalization).", + "bbox": [ + 88, + 303, + 485, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 743, + 220, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, the rapid development of personalized diffusion methods has been notable. Their low barrier to entry enables users to easily generate a wide array of images, but this same accessibility also facilitates the production of content that may infringe on personal portrait rights, raising significant privacy concerns. Traditionally, face recognition", + "bbox": [ + 89, + 768, + 483, + 861 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/465156445b391cb165cb770b5bad8de3b85776091d196f6ea7e5aff19a698da3.jpg", + "image_caption": [ + "Figure 1. Illustration of the Stylized-Face dataset across diverse styles. The dataset features 80 distinct styles organized into four primary groups: Artistic Living, Classical Art Styles, Fantasy Styles, and ACG Styles (Animation, Comics, and Games)." + ], + "image_footnote": [], + "bbox": [ + 519, + 267, + 901, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "models have served as invaluable tools for detecting potential portrait infringements within images. However, the advent of these new technologies presents a more complex issue. Personalized diffusion methods can generate images in various styles (e.g., anime, painting, cyberpunk styles) simply by altering prompts. This stylistic diversity poses a challenge for recognition models in recognizing faces with the same ID across a broad spectrum of stylized images.", + "bbox": [ + 511, + 566, + 906, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous methods [29, 67] addressing this challenge primarily rely on two models: CLIP-based models [45] and real face recognition models [13, 23, 56]. To systematically evaluate these models, we present a large-scale and diverse analysis across four different personalized diffusion models [29, 58, 64, 67] in Sec. 3.1. The results indicate that both exhibit significant limitations. Notably, CLIP-based models demonstrate considerable recognition biases. Although real face recognition models more closely mirror human perceptual judgments, their overall accuracy remains suboptimal. We illustrate failure cases in Fig. 2.", + "bbox": [ + 511, + 688, + 908, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Due to the limitations of the current recognition model, filling in the gaps in the relevant training data is an effective approach. We introduce a novel stylized face dataset,", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution", + "bbox": [ + 107, + 875, + 220, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding Author", + "bbox": [ + 109, + 887, + 236, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13053", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fcb2fca84b4e6a97360b8a0c5776a3d32d5f3db8b17e5b850f5eba89fc04de75.jpg", + "image_caption": [ + "Reference Image" + ], + "image_footnote": [], + "bbox": [ + 166, + 107, + 251, + 167 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4453d9c85933a961fdfac719b27ca4d8d324424106acc644c3dfdb764a634c6b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 171, + 179, + 248, + 237 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d8e5c4853dfbca0c0239a8baa07061819411b573585dafdca845bcc2526ffb28.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 173, + 251, + 248, + 308 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ac3343c2399de544f214e3c65fc8bd335212ad6045f61d6e11d6232675d0a0a0.jpg", + "image_caption": [ + "Line Art" + ], + "image_footnote": [], + "bbox": [ + 171, + 323, + 250, + 380 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2b2b12cf7055007de7eaf68b0c6f714c7b8510e50f3102e38096f3322ed277b2.jpg", + "image_caption": [ + "0.31" + ], + "image_footnote": [], + "bbox": [ + 267, + 109, + 341, + 166 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1d02d7d4c7bed319b7cc7456a87e3aea1f401ab1ae98fadae8e697b55ae39e6e.jpg", + "image_caption": [ + "0.26" + ], + "image_footnote": [], + "bbox": [ + 348, + 109, + 419, + 166 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a93b25614bfeb52a18410cbf308eeba75668ff4628b7a88c16a55da07ccf5a38.jpg", + "image_caption": [ + "Stylized Image" + ], + "image_footnote": [], + "bbox": [ + 426, + 109, + 500, + 166 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b794ddef1b526ecab4a5abcfe31d1d8e6ac7cb4498a49aeb489c5173a6ccb1e5.jpg", + "image_caption": [ + "0.42" + ], + "image_footnote": [], + "bbox": [ + 504, + 109, + 578, + 166 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dc939fe1a3dae834e5db12bdb19a8897fa5faeeabb3799c324f02043951dacc4.jpg", + "image_caption": [ + "0.19" + ], + "image_footnote": [], + "bbox": [ + 584, + 109, + 656, + 166 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a1339114f21d345dc6b2c9bd1085c4908b000d1a546a611a15e9d3caa1c9669c.jpg", + "image_caption": [ + "0.43" + ], + "image_footnote": [], + "bbox": [ + 267, + 180, + 341, + 236 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4c695cdc7a8703d1a9aee3dfc7370b7e8f118b9654ef04856c2b3a9399091b2f.jpg", + "image_caption": [ + "0.21" + ], + "image_footnote": [], + "bbox": [ + 348, + 180, + 421, + 236 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/744f6d7afe503442fe284c32859dadf023610e3b41e05bf50c9938f712b51192.jpg", + "image_caption": [ + "0.36" + ], + "image_footnote": [], + "bbox": [ + 426, + 180, + 500, + 236 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/729f60c52b1bbde85f3578452fe0a49e17c5be8ff72bd182f5bbadacc3d572e4.jpg", + "image_caption": [ + "0.25" + ], + "image_footnote": [], + "bbox": [ + 504, + 180, + 578, + 236 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/79f4f682b09261715a1a4203f8df8968cfe341b698f0d4821aa00f2b18e079df.jpg", + "image_caption": [ + "0.31" + ], + "image_footnote": [], + "bbox": [ + 584, + 180, + 656, + 236 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dd7c1457891b2b166d2d5862bd876fdbe30c32697d6c25217aac5cfecb674bb7.jpg", + "image_caption": [ + "0.36" + ], + "image_footnote": [], + "bbox": [ + 267, + 251, + 341, + 308 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/04958b94f7574a545b2270975a7a54ef961e6c3271e5af00d78a298b9caefab7.jpg", + "image_caption": [ + "0.46" + ], + "image_footnote": [], + "bbox": [ + 348, + 251, + 421, + 308 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d1b9378d7c76e1afd600b34f173ec0f5bf85a59df7d325718f54d01c65ae00bc.jpg", + "image_caption": [ + "0.44", + "0.45" + ], + "image_footnote": [], + "bbox": [ + 426, + 251, + 500, + 308 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/50d4769982fb7009ef4f894f81186ab9a3fa8a31b38d26752107105f021c206b.jpg", + "image_caption": [ + "0.32" + ], + "image_footnote": [], + "bbox": [ + 504, + 251, + 578, + 308 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/340c8a6d813370222e3b78aca8e90ee03ac839abbd9324e475ae6ff1b9483528.jpg", + "image_caption": [ + "0.36" + ], + "image_footnote": [], + "bbox": [ + 584, + 251, + 656, + 308 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6d48a6eaf8a4800e3d36eef9f553fe8390a69f08073e42a697720156c4211cfb.jpg", + "image_caption": [ + "0.42" + ], + "image_footnote": [], + "bbox": [ + 267, + 323, + 341, + 378 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/958d861fcc247708faf9b3250c8d593be9b053119453849ea10e1b3c0cde54e7.jpg", + "image_caption": [ + "0.16", + "Neonpunk" + ], + "image_footnote": [], + "bbox": [ + 348, + 323, + 421, + 378 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/92eacb424526efb71e274f924f20929e69ff8f8a3c39a79b5b6ee30ddcf74119.jpg", + "image_caption": [ + "0.32", + "Pop Art", + "Figure 2. Illustration of the Stylized-Face dataset and Comparison with Real Images. We randomly select a frontal face for each identity from the CASIA-WebFace [65] as a reference image. These reference images are compared with stylized images from our dataset and other images of the same identity from CASIA-WebFace [65], analyzing their feature similarity using a face recognition model trained on real faces. The results, recorded beneath each image, show that while the model performs well on real face images, it struggles to accurately identify the stylized images." + ], + "image_footnote": [], + "bbox": [ + 426, + 323, + 500, + 378 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4ab50453c2e7cc0c7e513b581f2986d5821d58b013575f2a96d9cbb63ab59d7f.jpg", + "image_caption": [ + "0.23", + "Film Noir" + ], + "image_footnote": [], + "bbox": [ + 504, + 323, + 578, + 378 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a08ea658c1ffc893e99e2d6207a76e0bc68457c45eccccaf2d11365fbb5c2d51.jpg", + "image_caption": [ + "0.24", + "Retro Arcade" + ], + "image_footnote": [], + "bbox": [ + 584, + 323, + 656, + 378 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f315b5db148131e68ceeb1f8c38c70d3299a22feb7c2f6bd334b6e16def57a47.jpg", + "image_caption": [ + "Real Image", + "0.78" + ], + "image_footnote": [], + "bbox": [ + 676, + 109, + 750, + 166 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ceb6aa35fd87d80f5bb4d00475d258c1ec6c57a55a924e2934af5036a0fba825.jpg", + "image_caption": [ + "0.81" + ], + "image_footnote": [], + "bbox": [ + 754, + 109, + 826, + 166 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a2afffa86c4f145c513e05321e622274cd1624ef511871c7029305a7f1865a4d.jpg", + "image_caption": [ + "0.76" + ], + "image_footnote": [], + "bbox": [ + 676, + 180, + 750, + 236 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7fb41f536dc415a8f9da97e7d21f40fe179a240789ad820c21ca873ecd2f1594.jpg", + "image_caption": [ + "0.84" + ], + "image_footnote": [], + "bbox": [ + 754, + 180, + 826, + 236 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7a01dcebbb8a210b671380025b1ef5e7317dc801239316a98809d751d2c3eb3c.jpg", + "image_caption": [ + "0.71" + ], + "image_footnote": [], + "bbox": [ + 676, + 251, + 750, + 308 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/36c635950a9ca2bc9b187b43b7d8126c064813d99eedcb27e38129f60d1b75cc.jpg", + "image_caption": [ + "0.79" + ], + "image_footnote": [], + "bbox": [ + 754, + 251, + 826, + 308 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e59a0ef3095de3fd786f64d32453f3430d5a249612d2de7bc8099378156ba19b.jpg", + "image_caption": [ + "0.79" + ], + "image_footnote": [], + "bbox": [ + 676, + 323, + 750, + 378 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0798fff24f9169c0322c13bc888519ef0531c653855e699d4961ae02f26378a2.jpg", + "image_caption": [ + "0.87" + ], + "image_footnote": [], + "bbox": [ + 754, + 323, + 826, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Stylized-Face, generated from the CASIA-WebFace [65] and Glint360k [6] datasets, as illustrated in Fig. 1 and Fig. 2. Stylized-Face is specifically designed to enhance model performance and generalization on stylized face recognition, particularly those that significantly deviate from real faces. It includes 4.6 million images across 62k IDs and 80 distinct styles. We employ InstantID[58] (the best-performing method in our evaluation) as the primary generative method, supplemented with three additional methods (PhotoMaker[29], Flashface[67], IPAdapter[64]) for dataset diversity. Owing to the inherent instability of generative model outputs, some images may feature faces that do not match the assigned IDs or might not contain faces at all. To maintain data quality at this scale, we use a semi-automatic process for data cleaning, which combines manual and model verification. To our knowledge, Stylized-Face is the first dataset dedicated to stylized face recognition.", + "bbox": [ + 89, + 520, + 483, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To further refine the setup for stylized face recognition, we construct three benchmarks: 1) Standard Stylized Face Recognition. Models are trained and evaluated on multiple styles from the same distribution, assessing their ability to recognize stylized faces. 2) Cross-prompt stylized face recognition. Different prompts generate images with distribution differences. We utilize varied prompts for both training and evaluation to assess the generalization perfor", + "bbox": [ + 89, + 779, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mance. 3) Cross-method stylized face recognition. Apart from the prompt, different generation methods yield varying distributions. We evaluate the generalization capabilities across the generation models by training and testing them on data produced through various techniques.", + "bbox": [ + 511, + 521, + 906, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate the baseline models using our benchmark. Experimental results demonstrate that models trained on Stylized-Face significantly enhance stylized face recognition performance, achieving a $15.9\\%$ increase in TAR at $\\mathrm{FAR} = 1\\mathrm{e - }4$ . Additionally, they improve cross-prompt generalization by $12.2\\%$ in TAR at $\\mathrm{FAR} = 1\\mathrm{e - }4$ and cross-method generalization by $13.3\\%$ in TAR at $\\mathrm{FAR} = 1\\mathrm{e - }3$ . These improvements underscore the value of our dataset for advancing stylized face recognition.", + "bbox": [ + 511, + 597, + 906, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To sum up, our major contributions are listed as follows:", + "bbox": [ + 531, + 733, + 903, + 748 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We evaluate previous models for stylized face recognition and mainstream personalized diffusion methods for the ability to preserve ID information. Our findings reveal significant shortcomings of previous recognition models.", + "- We introduce a stylized face dataset, Stylized-Face. Stylized-Face dataset includes 4.6 million images across 62k IDs, created by advanced generative techniques and semi-automatic process for data cleaning. Notably, it is the first dataset for stylized face recognition.", + "- We introduce three novel benchmarks designed to eval" + ], + "bbox": [ + 511, + 750, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13054", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "uate the performance of stylized face recognition models. We conduct a comprehensive evaluation of existing face recognition methods and our method using the Stylized-Face dataset.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 167, + 232, + 183 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Face Recognition", + "text_level": 1, + "bbox": [ + 89, + 193, + 259, + 209 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Face recognition technology [5, 7, 12, 13, 23, 25, 31, 32, 35, 36, 50, 56, 62, 71] extract and analyze condensed face features, commonly referred to as facial templates. It employs advanced loss functions based on angular margins to enhance recognition accuracy. Angular margin-based approach [13, 23, 56] aims to maximize inter-identity discrepancies and minimize intra-identity discrepancies (similarities across various images of the same person). With the introduction of the Stylized-Face dataset, there is an increased demand for discriminative capabilities in recognition models. This demands models that can extract more crucial and generalized features to maintain identity consistency across diverse styles, a challenge inherent to Domain Generalization (DG) [14-16, 19, 33, 40-42, 46, 49].", + "bbox": [ + 89, + 215, + 483, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Face Recognition Datasets", + "text_level": 1, + "bbox": [ + 89, + 438, + 328, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Real Face Dataset: Existing real face recognition models predominantly rely on datasets like Webface42M[72], CASIA-WebFace [65], MS1Mv2 [17], MS1Mv3 [13], LFW [21], CPLFW[68], CALFW [69], IJB-B [60] and IJB-C [34]. These datasets comprise real-world face images and are designed to train and evaluate models to improve accuracy in real face applications. Stylized-Face dataset differs from the above as it explores diverse styles of dataset specific to generated face images, which significantly deviates from real face.", + "bbox": [ + 89, + 460, + 482, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Synthetic Face Dataset: Earlier Synthetic datasets [8-10, 44] aim to enhance model performance on real face recognition and serve as substitutes for real face images. There are the main differences: (1) Authenticity of identity. We generate corresponding stylized images of real individuals; therefore, the IDs in our dataset are real, whereas existing datasets use virtual IDs. (2) Style. Our dataset includes various styles for the same ID, while existing datasets primarily focus on generating images in realistic style. (3) Objective. Our goal is designed for stylized face recognition, whereas existing datasets aims to enhance model performance on real face recognition and serve as substitutes for real face images.", + "bbox": [ + 89, + 613, + 482, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Some other works[4] attempt to generate stylized faces, but they only perform style transfer on thousands of images without checking the ID information in the stylized images, lacking the fundamental reliable ID information essential for face recognition tasks. Therefore, these works are not suitable for stylized face recognition tasks.", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cartoon Face Dataset: Some datasets[24, 37, 70] aims to explore different face style, such as cartoon face images. These datasets usually feature a single style and lack corresponding real human faces for comparison. Moreover, they are often sourced from animated films, which limits the data volume. In contrast, Stylized-Face dataset generates a large volume of face images across a wide range of styles, each paired with the ID of their corresponding real face.", + "bbox": [ + 511, + 90, + 903, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Text-to-Image Generation", + "text_level": 1, + "bbox": [ + 511, + 224, + 750, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-to-image diffusion models [22, 26, 30, 38, 43, 47, 52-54, 59, 66] are at the forefront of image generation, renowned for their ability to transform text prompts into high-quality images with diverse styles. These models utilize a pre-trained language encoder to convert text into a latent representation, guiding the diffusion process to precisely adjust pixels. For instance, GLIDE [38] features a cascaded text-guided diffusion architecture for image generation. DisCo [59] introduces clip[45] encoder to guide the image generation. Stable Diffusion [47] optimizes the process by performing diffusion in a compressed latent space, which lowers computational costs. Further evolving this technology, Stable Diffusion XL [43] employs a larger network and additional text encoder to improve image fidelity and textual influence. Additionally, there are some GAN-based methods[11, 27, 28, 55, 57, 63], but they usually have weak generalization capabilities and are computationally expensive to train.", + "bbox": [ + 511, + 246, + 906, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.4. Personalized diffusion models", + "text_level": 1, + "bbox": [ + 511, + 530, + 784, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In image generation, ID-preserving techniques prioritize key facial attributes to ensure identity remains consistent. These methods vary in their need for fine-tuning during testing. For instance, techniques such as DreamBooth [48] and LoRA [20] require extensive fine-tuning for each character, making them impractical for large-scale ID image generation. Recent developments aim to eliminate the need for fine-tuning altogether. Two fundamental methods include directly embedding the image features into the text embedding and using a cross-attention module to integrate the image encoding with the text encoding. The former includes methods such as Photomaker [29], FastComposer [61], and Face2Diffusion [51]. The latter comprises methods such as IP-Adapter-FaceID [64], InstantID [58], and Flashface [67]. In terms of specific designs, PhotoMaker [29] enhance identity preservation by modifying image encoders. IP-Adapter-FaceID [64] specifically utilizes face ID embeddings to ensure consistent identity retention. Moreover, InstantID [58] replaces the standard clip[45] encoder with face features and employs an additional IdentityNet, supervised by key facial landmarks, to guide face generation. FlashFace [67] encodes IDs into a series of feature maps to preserve finer face details. Our dataset utilizes these ID-preserving tech", + "bbox": [ + 511, + 553, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13055", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/6dda9b7af4e165739da8a1b71fee6638fa83da10f9b8b05ead58570f5bae1e41.jpg", + "table_caption": [ + "Table 1. Styles Overview." + ], + "table_footnote": [], + "table_body": "
CategoryExample Style Prompts
1. Artistic Living
1.1 Photography Styles:film noir, analog film photo, long exposure photo, et al.
1.2 Fashion Styles:glamorous photo, low contrast realistic photo, minimalist style, et al.
2. Fantasy Styles
2.1 Sci-Fi and Futurism:sci-fi style, futuristic style, retro-futuristic, space-themed, et al.
2.2 Cyberpunk Styles:vaporwave style, neonpunk style, neon noir, dystopian style, et al.
2.3 Fantasy and Themed Styles:fairy tale, gothic style, nautical-themed, disco-themed, et al.
3. Classical Art Styles
3.1 Genre Art Styles:impressionist painting, cubist artwork, expressionist, renaissance style, et al.
3.2 Painting Arts:watercolor painting, line art drawing, graffiti style, urban art, et al.
3.3 Paper Art and Handicrafts:origami style, kirigami representation, 3D papercut shadow box, et al.
4. ACG Styles
4.1 Anime and Cartoon Styles:animework, manga style, kawaii style, et al.
4.2 Comics and Game Art:comic, role-playing game style fantasy, retro game art, pixel art, et al.
", + "bbox": [ + 106, + 114, + 890, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "niques and enhances them by multiple prompt styles and pre-trained diffusion models.", + "bbox": [ + 89, + 375, + 482, + 405 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Dataset Construction and Method", + "text_level": 1, + "bbox": [ + 89, + 422, + 400, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Evaluation of Recognition Models and Personalized Diffusion Methods", + "text_level": 1, + "bbox": [ + 89, + 450, + 482, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We begin by evaluating the effectiveness of existing evaluation models and personalized diffusion methods. For the evaluation metrics, we select two well-established metrics: CLIP-I, which measures identity preservation through CLIP features, and Face Sim., which is based on real face recognition models. For comparison, we also incorporate human evaluations.", + "bbox": [ + 89, + 489, + 483, + 593 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In selecting personalized diffusion methods, we choose four of the latest ID-preserving image generation techniques: PhotoMaker [29], InstanceID [58], Flashface [67], and IPAdapter [64]. All methods are implemented using the officially recommended settings. We analyze additional methods that are excluded from our study in Appendix. To ensure a comprehensive assessment, we additionally utilize the FID metric [18] and SDD-FIQA [39] to evaluate the overall image quality. We employ CLIP-T to evaluate the consistency between the stylized images and their corresponding textual prompts.", + "bbox": [ + 89, + 597, + 482, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To comprehensively evaluate the different methods, we incorporate a diverse array of artistic styles. We first classify these common styles into 4 major categories and 10 sub-categories, as shown in Tab. 1. Our classification is based on people's usual perception of art and common forms of expression, ensuring that the grouping reflects intuitive and widely recognized distinctions in artistic presentation. For each sub-category, we curate a collection of specific prompts that together cover most common scenarios.", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f0c954e0a182b7ecb678241fb1a7d70788eeda937bb61d7effbd8ce023721e5d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 372, + 696, + 464 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d94eaf62b28e07f65a97c77d88177fcedeee56e047e1b228ca46ce6078d73e07.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 372, + 903, + 464 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/87a5da8a13c6d56a6381cf14c8405289f048d3a7fbb8b4a39faea1d7d19dafbe.jpg", + "image_caption": [ + "IPAdapter", + "InstantID", + "Figure 3. Quality Evaluation. We evaluate the generation quality of different methods using SDD-FIQA [39]." + ], + "image_footnote": [], + "bbox": [ + 517, + 479, + 694, + 571 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8b4a658085449901092217fd1b433a465c02d7f29eea4c3ed14240b06da41d7e.jpg", + "image_caption": [ + "Flashface", + "Photomaker" + ], + "image_footnote": [], + "bbox": [ + 727, + 479, + 903, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our collection is both diverse and extensive, comprising: 1) we use official SDXL prompts [43], carefully selected for their high quality and distinctive style; 2) we also collect popular repository resources [3]. After deduplicating and removing prompts that conflict with identity information, we retain 80 distinct style prompts that cover a broad spectrum of common scenarios. We use 1,000 IDs from the CASIA-WebFace[65] dataset, and each method generates 80,000 images.", + "bbox": [ + 511, + 657, + 906, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We present the results of our analysis in Tab. 2 and Fig. 3. Our findings indicate that the similarity scores derived from CLIP features lack sufficient discriminative power and exhibit bias when compared to human evaluations. This limitation stems from CLIP's focus on extracting a generalized representation of identity. In contrast, real face recognition models align more closely with human judgment. For fur", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13056", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/c423711dbb367c3fed03b8e884ab85732053fe40d95cc0b299b5b16e2d4ab964.jpg", + "table_caption": [ + "Table 2. Comparison of Different Personalized Diffusion Methods Based on Various Evaluation Metrics." + ], + "table_footnote": [], + "table_body": "
MethodPreserve ID InformationText ConsistencyGeneration Quality
CLIP-IFace Sim.Human EvalCLIP-TFID
PhotoMaker [29]54.5%12.3%21.9%20.5%313
InstanceID [58]65.2%70.9%90.0%26.1%285
Flashface [67]65.6%35.9%87.1%22.9%244
IPAdapter [67]60.3%25.3%64.9%25.1%255
", + "bbox": [ + 187, + 114, + 810, + 214 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ther study, a detailed assessment of real face recognition models is provided in the experiment section. The evaluation under classification metrics indicates that real face recognition models are also inaccurate in assessing identity consistency.", + "bbox": [ + 89, + 239, + 480, + 315 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In terms of the specific method, InstantID achieves the best performance in preserving ID information, text preservation, and quality distribution, demonstrating strong performance. In terms of FID score, Flashface performs better.", + "bbox": [ + 89, + 316, + 482, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Setup and Data Cleaning", + "text_level": 1, + "bbox": [ + 89, + 391, + 320, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on Sec. 3.1, InstantID [58] provides stable performance, making it an ideal choice for our task. As a result, we choose InstantID as the primary method for generation. To further enhance the diversity of the dataset, we use three Pre-trained Diffusion Models:SDXL [43], YamerMIX [2], Anything-XL [1]. To further enrich the dataset, we incorporate three additional methods: PhotoMaker [29], Flashface [67] and IPAdapter [64].", + "bbox": [ + 89, + 414, + 483, + 535 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use 10k IDs from CISIA-Webface and 52k IDs from Glint360k as the original images. To enhance generation quality, we apply a series of filtering steps to the original dataset. These steps include: 1) discarding images with low similarity to others within the same ID; 2) removing low-quality images based on SDD-FIQA results, specifically those with scores below 60, which eliminates blurry or heavily tilted images; and 3) filtering out images with face angles exceeding 30 degrees, based on human pose detection. After applying these filtering steps, $78\\%$ of the face images are retained. We generate the dataset by iterating over all combinations of methods, pre-trained diffusion models, and prompts across the entire dataset, producing more than 10 million images. We directly skip combinations where over $50\\%$ of the images are identified by all annotators as not matching the original ID will be entirely excluded.", + "bbox": [ + 89, + 536, + 483, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the stylized images, we begin with preprocessing. We use the RetinaFace model for face detection, adjusting the detected faces based on five face landmarks. Notably, $8.4\\%$ of the images fail to detect any faces. In such cases, we resize the original image directly to maintain a consistent image size. We also compared additional face detection models, with more details provided in Appendix.", + "bbox": [ + 89, + 795, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3fd2d0e242c099542ecad6e1d7d457bff8717d5826576adaa7de6f9ec16ec084.jpg", + "image_caption": [ + "Figure 4. Pipeline of Semi-automatic Process. D. indicates delete, M. indicates manual verification, Sim. BSI measures the similarity between stylized images, Sim. BOI measures the similarity between original images, and Sim. BIP measures the similarity between a stylized image and its corresponding original image. We alternate between two phases: model verification and manual verification. In the model image pairs, a trained face recognition model performs two types of filtering: 1) The model compares the similarity between a query image and other stylized images generated from the highly similar original face. If the similarity score (TAR@FAR 1e-3) falls below the threshold, the query image is discarded. 2) It removes image pairs (stylized image and its origin image) with low similarity (TAR@FAR 1e-2). During the manual verification phase, these confusion image pairs (between TAR@FAR 1e-2 and TAR@FAR 1e-3) are labeled." + ], + "image_footnote": [], + "bbox": [ + 521, + 237, + 901, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to the instability of generative methods, we need to verify whether each image truly preserves the ID information. However, given the large scale of the data, comprehensive manual verification becomes impractical. To alleviate the annotation burden, we employ a semi-automatic process for data cleaning, as shown in Fig. 4. In the semi-automated process, we alternate between two main phases: model validation and manual verification.", + "bbox": [ + 511, + 715, + 905, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the model validation phase, we first train a face recognition model using the remaining dataset. We use two methods for rapid filtering by trained model: 1) Images with a similarity to original real faces below a threshold", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13057", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(TAR@FAR at 1e-2) are deleted. 2) Generally, images generated from similar original images tend to have higher similarity; if there is a significant deviation, it is highly likely to be a bad case. Images with similarity to other stylized images from the same ID below the threshold (TAR@FAR at 1e-3), while similarity between their corresponding original images greater than 0.8, are also deleted.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the manual validation phase, annotators compare the confusion images (between TAR@FAR at 1e-2 and TAR@FAR at 1e-3) to verify the ID information. During the annotation process, we develop a tool that provides yes and no options for each image pair, which are confirmed by the annotators. If more than $50\\%$ of the annotators believe that the two images belong to the same ID, they are retained. Any images below this threshold are excluded from the dataset.", + "bbox": [ + 88, + 198, + 482, + 332 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Dataset statistics", + "text_level": 1, + "bbox": [ + 89, + 340, + 256, + 357 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Large scale images. The Stylized-Face dataset consists of 4,660,285 images of 62,446 individuals. The generation size of all images is determined by the default parameters of their methods, and ultimately, they are sized at $112 \\times 112$ pixels.", + "bbox": [ + 89, + 363, + 482, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Utilized Models. Tab. 3 illustrates the proportions of various methods employed in our dataset. The higher proportion of the InstantID(SD) method is due to its more common use in real-world applications. We provide a detailed comparison of these methods in Appendix.", + "bbox": [ + 89, + 439, + 483, + 513 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d1ffa45b2ce33e1a6875d748ff8bbce1651e802b4a80c73cddbbb4ff1b231b9d.jpg", + "table_caption": [ + "Table 3. The proportions of various methods" + ], + "table_footnote": [], + "table_body": "
MethodsProportion
InstantID(SD) [58]69%
InstantID(YamerMIX) [58]14%
InstantID(Anything-XL) [58]14%
PhotoMaker [29]1%
Flashface [67]1%
IPAdapter [67]1%
", + "bbox": [ + 135, + 551, + 439, + 675 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Mean Faces. Fig. 5 displays the average faces for individual IDs from our dataset compared with those from CASIA-WebFace. Notably, while our dataset exhibit considerable style variations, the core facial features remain relatively stable.", + "bbox": [ + 89, + 689, + 482, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments of Stylized Face Recognition", + "text_level": 1, + "bbox": [ + 89, + 777, + 464, + 796 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Evaluation Protocol", + "text_level": 1, + "bbox": [ + 89, + 803, + 281, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Standard Stylized Face Recognition. The training set consists of 4,453,814 images generated by the InstantID method, encompassing 61k IDs. Additionally, we also use the real face images corresponding origin dataset for training. The test set includes 1k independent IDs, with 56,644", + "bbox": [ + 89, + 824, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/67c2cc1f17e39ac7f64e38380b84262ee469cc785cde1bfcecd8a8c44ef326e8.jpg", + "image_caption": [ + "CISIA-WebFace II" + ], + "image_footnote": [], + "bbox": [ + 526, + 90, + 612, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/355814a46a3c471a3698697306813e813181212f81cb46d93d41c6d0f508beeb.jpg", + "image_caption": [ + "Stylized-Face" + ], + "image_footnote": [], + "bbox": [ + 616, + 90, + 700, + 156 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/314d148384962168cedb61bcc40c019027408045aa400583827e2b427dc778e0.jpg", + "image_caption": [ + "Figure 5. Display of average faces for single ID. We randomly select two IDs from the CISIA-WebFace dataset and compute the average face for each ID using the images corresponding to those IDs." + ], + "image_footnote": [], + "bbox": [ + 722, + 90, + 805, + 156 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c8545101ca8549dcbb266a189e41c5c5bc5a88db91a1c81429fe45a15be7cdfd.jpg", + "image_caption": [ + "CISIA-WebFace Stylized-Face ID 2" + ], + "image_footnote": [], + "bbox": [ + 813, + 90, + 895, + 156 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Stylized-Face", + "bbox": [ + 820, + 159, + 890, + 180 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "images also generated by InstantID. Both the training and test sets utilize the full set of 80 prompts. The primary goal of this evaluation is to assess the model's performance in open-set stylized face recognition.", + "bbox": [ + 511, + 276, + 905, + 337 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cross-prompt Stylized Face Recognition. The training and test sets for Cross-prompt Stylized Face Recognition are subsets of the corresponding training and test sets for Standard Stylized Face Recognition. The training set only includes images generated using 60 prompts from the Standard Stylized Face Recognition training set. It contains a total of 3,811,707 images. Additionally, the real face images corresponding to the IDs is also used for training. The test set consists of images generated using the last 20 prompts from the Standard Stylized Face Recognition test set, totaling 42,361 images. This setting ensures that the prompts and IDs used in the training and test sets are distinct. More details are shown in Appendix.", + "bbox": [ + 511, + 337, + 905, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cross-method Stylized Face Recognition. In this setup, we aim to assess cross-method generalization. The training set is the same as that used in Standard Stylized Face Recognition. The test set consists of 1k independent IDs, with 149,827 images generated by Flashface, IP-Adapter, and Photomaker. In detail, Flashface contributed 70,912 images, IP-Adapter contributed 51,915 images, and Photomaker contributed 27,000 images.", + "bbox": [ + 511, + 534, + 905, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 511, + 662, + 696, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In all experiments, we follow the standard evaluation metrics for face recognition. We evaluate the performance of the model on face verification. The gallery consists of all images in the test set, excluding the test images themselves. We report the results at multiple points, including TAR@FAR at 1e-2, TAR@FAR at 1e-3, and TAR@FAR at 1e-4. The metrics are defined as follows:", + "bbox": [ + 511, + 685, + 905, + 790 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- True Acceptance Rate (TAR) (complementary to False Non-Match Rate (FNMR)): Proportion of correctly identified valid identities:", + "bbox": [ + 511, + 792, + 905, + 838 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nT A R = \\frac {M}{N}\n$$\n", + "text_format": "latex", + "bbox": [ + 671, + 845, + 758, + 876 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $M$ is the number of correctly identified identities,", + "bbox": [ + 526, + 885, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13058", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and $N$ is the total number of valid identity attempts.", + "bbox": [ + 102, + 90, + 447, + 104 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note: ${FNMR} = 1 - {TAR}$", + "bbox": [ + 102, + 106, + 295, + 119 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- False Acceptance Rate (FAR) (equivalent to False Match Rate (FMR)): Proportion of non-target identities incorrectly accepted:", + "bbox": [ + 91, + 121, + 482, + 167 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nF A R = \\frac {L}{K}\n$$\n", + "text_format": "latex", + "bbox": [ + 251, + 170, + 336, + 200 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $L$ is the number of incorrect acceptances, and $K$ is the number of non-target identity attempts.", + "bbox": [ + 102, + 204, + 482, + 234 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note: ${FMR} = {FAR}$", + "bbox": [ + 102, + 234, + 256, + 248 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- TAR@FAR: Represents the True Acceptance Rate (TAR) at a specific False Acceptance Rate (FAR), used to evaluate system accuracy at a fixed security level.", + "bbox": [ + 91, + 250, + 482, + 296 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 301, + 307, + 318 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We maintain the same setup for all experiments. Our backbone networks are IResNet-50 and IResNet-101. We apply Cosface [56] framework and use an angular margin of 0.3 for Cosface. We set the batch size to 512 and train for a total of 26 epochs. The models are subsequently trained using Stochastic Gradient Descent (SGD), with a momentum coefficient of 0.9 and a weight decay parameter of $5 \\times 10^{-4}$ . The initial learning rate is set at $1 \\times 10^{-1}$ and is decreased by a factor of 8 at the midpoint of the epochs.", + "bbox": [ + 89, + 324, + 483, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Result", + "text_level": 1, + "bbox": [ + 89, + 467, + 176, + 482 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.1. Standard Stylized Face Recognition.", + "text_level": 1, + "bbox": [ + 89, + 489, + 388, + 505 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We assess the model's performance trained on the large-scale MS1MV3 [13] and Glint360k [6] dataset. As presented in Tab. 4, the results indicate that despite the substantial size of the training set, there is a significant performance degradation compared to real face scenarios. Although the model trained on Glint360k performs worse than ours, it shows significant improvement over the model trained on MS1MV3, highlighting the positive impact of dataset size on generalization.", + "bbox": [ + 89, + 508, + 482, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The model trained on our dataset significantly outperforms the one trained on real face data in open-set face recognition scenarios, achieving a face verification TAR@FAR at 1e-4 of $90.6\\%$ , as detailed in Tab. 4. This highlights the crucial role of our dataset in effectively recognizing stylized faces.", + "bbox": [ + 89, + 645, + 482, + 734 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/dc5aee8b9869bcf0441ce0410355b6e956885bd4f5750b2f9b54098d876d9311.jpg", + "table_caption": [ + "Table 4. Results on Standard Stylized Face Recognition. Backbone is IResNet-50. Values in bold indicate the top results." + ], + "table_footnote": [], + "table_body": "
Training sets1e-21e-31e-4
MS1MV3 [13]85.168.049.7
Gint360k [6]95.087.074.7
Stylized-Face97.995.590.6
", + "bbox": [ + 163, + 785, + 411, + 857 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Fig. 6, the model trained on our dataset significantly outperforms the one trained on real face data in", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "open-set face recognition scenarios. It highlights the crucial role of our dataset in effectively recognizing stylized faces.", + "bbox": [ + 511, + 90, + 905, + 135 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a0fee25d5cbf56b5dbe7d2f868628be37cfe3e5fd28f308214db990acaeb9103.jpg", + "image_caption": [ + "Figure 6. the FMR-FNMR plot for All pairs verification. (lower FNMR is better)." + ], + "image_footnote": [], + "bbox": [ + 539, + 161, + 890, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.2. Cross-prompt stylized face recognition.", + "text_level": 1, + "bbox": [ + 511, + 409, + 831, + 424 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5 displays the performance across various prompts using the InstanceID [58], highlighting its generalization capabilities in stylized face recognition across different prompts. It also underscores our dataset's effectiveness in enhancing the cross-prompt generalization of stylized face recognition.", + "bbox": [ + 511, + 428, + 905, + 518 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1c5473515d28e89a55c186be97a9fa1b234acbb6fd6df821c5f37ae2fb28d2dd.jpg", + "table_caption": [ + "Table 5. Results on Cross-prompt Stylized Face Recognition. Backbone is IResNet-50. Values in bold indicate the top results." + ], + "table_footnote": [], + "table_body": "
Training sets1e-21e-31e-4
MS1MV3 [13]85.767.749.3
Gint360k [6]94.786.573.9
Stylized-Face97.995.490.3
", + "bbox": [ + 586, + 571, + 834, + 645 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.3. Cross-method Stylized Face Recognition", + "text_level": 1, + "bbox": [ + 511, + 669, + 841, + 684 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cross-method experiments in Table 6 evaluate the ability of models to generalize feature extraction across various generation methods. The results indicate that, despite our method showing clear superiority over real-face models, current methods struggle with cross-method generalization across different methods, emphasizing the need for more in-depth research to enhance cross-method robustness.", + "bbox": [ + 511, + 688, + 905, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further analyze cross-method generalization using the Flashface dataset. Prompts are divided into easy and hard categories based on the distribution shift, measured by the difference between Flashface and the InstantID method on the same data. Results show that performance on easy prompts is similar to within-distribution performance, while generalization challenges are more pronounced for hard", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13059", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0854cb7d9932feb3f070adc15174ec2a267f5ac459f8c7de0fb1534621f4ffc4.jpg", + "table_caption": [ + "Table 6. Results on Cross-method Stylized Face Recognition. Backbone is IResNet-50. All refers to Flashface, IP-Adapter, and Photomaker. Due to the limited number of images, the results exclude TAR@FAR at 1e-4. Values in bold indicate the top results." + ], + "table_footnote": [], + "table_body": "
Training setsTest Parts1e-21e-3
MS1MV3 [13]Flashface82.171.0
Gint360k [6]Flashface92.386.7
Stylized-FaceFlashface93.187.3
MS1MV3 [13]IP-Adapter68.262.2
Gint360k [6]IP-Adapter91.474.3
Stylized-FaceIP-Adapter93.678.3
MS1MV3 [13]Photomaker62.636.4
Gint360k [6]Photomaker86.562.2
Stylized-FacePhotomaker90.171.2
MS1MV3 [13]All59.031.8
Gint360k [6]All78.758.6
Stylized-FaceAll87.271.9
", + "bbox": [ + 140, + 155, + 436, + 373 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "prompts with significant distributional shifts. Notably, for these hard prompts, our model demonstrates particularly remarkable performance improvements compared to the real face recognition model. Results are shown in Tab. 7.", + "bbox": [ + 89, + 393, + 483, + 455 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/86ef15ab03947bcd70f9f587ae188efafa009c98851b52e192a1fb1966740071.jpg", + "table_caption": [ + "Table 7. Analysis of Cross-method generalization. The evaluation metrics is TAR@FAR at 1e-2 on Flashface. Backbone is IResNet-50." + ], + "table_footnote": [], + "table_body": "
MethodsEasy PromptsHard PromptsAll
Glint360k96.984.992.3
Stylized-Face96.987.193.1
", + "bbox": [ + 109, + 518, + 464, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 599, + 243, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Performance on Real Face Datasets. We evaluate the model's performance on real face datasets, with the results presented in Fig. 8. The comparison demonstrates that while our model achieves comparable performance, there exists a trade-off between real face recognition and stylized face recognition.", + "bbox": [ + 89, + 622, + 485, + 715 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e7194a974665c45e179726ffa584251a580a615a0721fec5b37b691b47d6e18f.jpg", + "table_caption": [ + "Table 8. Results on Real Face Recognition. Backbone is IResNet-50." + ], + "table_footnote": [], + "table_body": "
Training setLFWCFP_FPAgeDBAvg.
CASIA99.595.294.996.5
MS1MV299.898.198.198.7
MS1MV399.898.598.398.9
Glint360k99.899.198.599.1
Stylized-Face99.798.497.998.7
", + "bbox": [ + 117, + 763, + 454, + 872 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "More Methods To comprehensively assess the perfor", + "bbox": [ + 89, + 885, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mance of stylized face recognition dataset, we compare advanced face recognition methods, including Cosface [56], Arcface [13], and CurricularFace [23]. These methods represent leading techniques in the field of face recognition. The experimental results are summarized in Tab. 9. The performance of the different training methods is similar.", + "bbox": [ + 511, + 90, + 905, + 183 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ec3880e0d0577c711eb01366f43040e10f42f0e47760237394d2b6634e9c2564.jpg", + "table_caption": [ + "Table 9. Performance Comparison of Different Training Methods on Standard Stylized Face Recognition." + ], + "table_footnote": [], + "table_body": "
Training Methods1e-21e-31e-4
Cosface [56]97.995.590.6
CurricularFace [23]97.995.590.5
Arcface [13]97.995.489.9
", + "bbox": [ + 571, + 233, + 849, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "More Backbones We evaluate the performance of various backbones and Tab. 10 shows that scaling up the model size can lead to further improvements in performance.", + "bbox": [ + 511, + 320, + 906, + 367 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/23c2eaa8b2ea357c327d6765338a47a64151cdcafb432e3b018fd3a49f455cd9.jpg", + "table_caption": [ + "Table 10. Performance Comparison of Different Backbones on Standard Stylized Face Recognition." + ], + "table_footnote": [], + "table_body": "
Backbone1e-21e-31e-4
IResNet-5097.995.590.6
IResNet-10198.396.392.0
", + "bbox": [ + 591, + 417, + 828, + 477 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 503, + 633, + 520 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In addressing the challenge of recognizing stylized face images, we introduce the Stylized-Face dataset based on the CASIA-WebFace and Glint360k. Our dataset includes 4.6 million images spanning 62k IDs and incorporate a wide range of styles and identity-preserving methods. To ensure data quality (i.e., ID preservation) at this massive scale, we implement a semi-automated pipeline for large-scale data cleaning. Stylized-Face is the first dataset for stylized face recognition. Furthermore, we establish benchmarks to evaluate the robustness and generalization of stylized face recognition methods. Experimental results demonstrate that models trained on Stylized-Face achieve remarkable improvements in stylized face recognition.", + "bbox": [ + 511, + 529, + 906, + 727 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Limitation and Future Work", + "text_level": 1, + "bbox": [ + 511, + 739, + 779, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This task continues to present challenges and opportunities for improvement. (1) Cross-method Stylized Face Recognition requires the adoption of additional designs to improve its generalization capabilities. (2) Explore approaches that can simultaneously achieve state-of-the-art performance in both real and stylized face recognition (3) Furthermore, considering the various long-tail distribution styles in real-world scenarios, which is also a direction for future research.", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13060", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the National Natural Science Foundation of China (62302167, U23A20343, 62472282, 72192821, 62476090), Young Elite Scientists Sponsorship Program by CAST (YESS20240780), Shanghai Sailing Program (23YF1410500), Chenguang Program of Shanghai Education Development Foundation, CCF-Tencent Rhino-Bird Young Faculty Open Research Fund (CCF-Tencent RAGR20230121, RAGR20240122), Shanghai Municipal Education Commission (23CGA34) and the Open Research Fund of Key Laboratory of Advanced Theory and Application in Statistics and Data Science-MOE, ECNU.", + "bbox": [ + 89, + 114, + 485, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 294, + 187, + 310 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anything-xl. https://civitai.com/models/9409/or-anything-xl.5", + "[2] Yamermix. https://civitai.com/models/84040/sdxl-unstable-diffusers-yamermix. 5", + "[3] Sdsl prompt styler, a custom node for comfyui. https://github.com/twri/sdsl_prompt_styler. 4", + "[4] Reverse style transfer. https://devpost.com/software/reverse-style-transfer. 3", + "[5] Xiang An, Xuhan Zhu, Yuan Gao, Yang Xiao, Yongle Zhao, Ziyong Feng, Lan Wu, Bin Qin, Ming Zhang, Debing Zhang, and Ying Fu. Partial fc: Training 10 million identities on a single machine. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2021. 3", + "[6] Xiang An, Xuhan Zhu, Yuan Gao, Yang Xiao, Yongle Zhao, Ziyong Feng, Lan Wu, Bin Qin, Ming Zhang, Debing Zhang, et al. Partial fc: Training 10 million identities on a single machine. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1445-1449, 2021. 2, 7, 8", + "[7] Xiang An, Jiangkang Deng, Jia Guo, Ziyong Feng, Xuhan Zhu, Yang Jing, and Liu Tongliang. Killing two birds with one stone: Efficient and robust training of face recognitionCNNs by partial fc. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 3", + "[8] Gwangbin Bae, Martin de La Gorce, Tadas Baltrusaitis, Charlie Hewitt, Dong Chen, Julien Valentin, Roberto Cipolla, and Jingjing Shen. Digiface-1m: 1 million digital face images for face recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3526-3535, 2023. 3", + "[9] Fadi Boutros, Marco Huber, Patrick Siebke, Tim Rieber, and Naser Damer. Sface: Privacy-friendly and accurate face recognition using synthetic data. In 2022 IEEE International Joint Conference on Biometrics (IJCB), pages 1-11. IEEE, 2022.", + "[10] Fadi Boutros, Marcel Klemt, Meiling Fang, Arjan Kuijper, and Naser Damer. Unsupervised face recognition using unlabeled synthetic data. In 2023 IEEE 17th International Conference on Automatic Face and Gesture Recognition (FG), pages 1-8. IEEE, 2023. 3" + ], + "bbox": [ + 94, + 320, + 482, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] PAN Chuanyu, YANG Guowei, MU Taijiang, and LAI Yu-Kun. Generating animatable 3d cartoon faces from single portraits. Virtual Reality & Intelligent Hardware, 6(4):292-307, 2024. 3", + "[12] Jun Dan, Yang Liu, Haoyu Xie, Jiankang Deng, Haoran Xie, Xuansong Xie, and Baigui Sun. Transface: Calibrating transformer training for face recognition from a data-centric perspective. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20642-20653, 2023. 3", + "[13] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 1, 3, 7, 8", + "[14] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pages 1180-1189. PMLR, 2015. 3", + "[15] Jingyu Gong, Fengqi Liu, Jiachen Xu, Min Wang, Xin Tan, Zhizhong Zhang, Ran Yi, Haichuan Song, Yuan Xie, and Lizhuang Ma. Optimization over disentangled encoding: Unsupervised cross-domain point cloud completion via occlusion factor manipulation. In European Conference on Computer Vision, pages 517-533. Springer, 2022.", + "[16] Jingyu Gong, Yujing Lou, Fengqi Liu, Zhiwei Zhang, Haoming Chen, Zhizhong Zhang, Xin Tan, Yuan Xie, and Lizhuang Ma. Scene point cloud understanding and reconstruction technologies in 3d space. Journal of Image and Graphics, 28(06):1741-1766, 2023. 3", + "[17] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 87-102. Springer, 2016. 3", + "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 4", + "[19] Chengyang Hu, Ke-Yue Zhang, Taiping Yao, Shice Liu, Shouhong Ding, Xin Tan, and Lizhuang Ma. Domainhallucinated updating for multi-domain face anti-spoofing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2193-2201, 2024. 3", + "[20] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3", + "[21] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008. 3", + "[22] Lianghua Huang, Di Chen, Yu Liu, Yujun Shen, Deli Zhao, and Jingren Zhou. Composer: Creative and controllable image synthesis with composable conditions. arXiv preprint arXiv:2302.09778, 2023. 3" + ], + "bbox": [ + 516, + 92, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13061", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: adaptive curriculum learning loss for deep face recognition. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5901-5910, 2020. 1, 3, 8", + "[24] Jing Huoa, Wenbin Lia, Yinghuan Shia, Yang Gaoa, and Hujun Yinb. Webcaricature: a benchmark for caricature face recognition. arXiv preprint arXiv:1703.03230, 2017. 3", + "[25] Shen Li, Jianqing Xu, Xiaqing Xu, Pengcheng Shen, Shaoxin Li, and Bryan Hooi. Spherical confidence learning for face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15629-15637, 2021. 3", + "[26] Xiaofan Li, Xin Tan, Zhuo Chen, Zhizhong Zhang, Ruixin Zhang, Rizen Guo, Guanna Jiang, Yulong Chen, Yanyun Qu, Lizhuang Ma, et al. One-for-more: Continual diffusion model for anomaly detection. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 4766-4775, 2025. 3", + "[27] Yidi Li, Jun Xiao, Yiqun Wang, and Zhengda Lu. Depthgan: Gan-based depth generation from semantic layouts. Computational Visual Media, 10(3):505-522, 2024. 3", + "[28] Yuxuan Li, Lingfeng Yang, and Xiang Li. Apf-gan: Exploring asymmetric pre-training and fine-tuning strategy for conditional generative adversarial network. Computational Visual Media, 10(1):187-192, 2024. 3", + "[29] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. arXiv preprint arXiv:2312.04461, 2023. 1, 2, 3, 4, 5, 6", + "[30] Zonglin Li, Xiaogian Lv, Wei Yu, Qinglin Liu, Jingbo Lin, and Shengping Zhang. Face shape transfer via semantic warping. Visual Intelligence, 2(1):26, 2024. 3", + "[31] Xiao Lin, Yuge Huang, Jianqing Xu, Yuxi Mi, Shuigeng Zhou, and Shouhong Ding. Uface: Unleashing inherent model capabilities to enhance intra-class diversity in synthetic face recognition. arXiv preprint arXiv:2502.19803, 2025.3", + "[32] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherical face: Deep hypersphere embedding for face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 212-220, 2017. 3", + "[33] Wang Lu, Jindong Wang, Haoliang Li, Yiqiang Chen, and Xing Xie. Domain-invariant feature exploration for domain generalization. arXiv preprint arXiv:2207.12020, 2022. 3", + "[34] Brianna Maze, Jocelyn Adams, James A Duncan, Nathan Kalka, Tim Miller, Charles Otto, Anil K Jain, W Tyler Niggel, Janet Anderson, Jordan Cheney, et al. Iarpa janus benchmark-c: Face dataset and protocol. In 2018 international conference on biometrics (ICB), pages 158-165. IEEE, 2018. 3", + "[35] Yuxi Mi, Zhizhou Zhong, Yuge Huang, Jiazhen Ji, Jianqing Xu, Jun Wang, Shaoming Wang, Shouhong Ding, and Shuigeng Zhou. Privacy-preserving face recognition using trainable feature subtraction. In Proceedings of the" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 297-307, 2024. 3", + "[36] Yuxi Mi, Zhizhou Zhong, Yuge Huang, Qiuyang Yuan, Xuan Zhao, Jianqing Xu, Shouhong Ding, Shaoming Wang, Rizen Guo, and Shuigeng Zhou. Data synthesis with diverse styles for face recognition via 3dmm-guided diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21203-21214, 2025. 3", + "[37] Ashutosh Mishra, Shyam Nandan Rai, Anand Mishra, and CV Jawahar. Iit-cfw: A benchmark database of cartoon faces in the wild. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part I 14, pages 35-47. Springer, 2016. 3", + "[38] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3", + "[39] Fu-Zhao Ou, Xingyu Chen, Ruixin Zhang, Yuge Huang, Shaoxin Li, Jilin Li, Yong Li, Liujuan Cao, and Yuan-Gen Wang. Sdd-fiaq: Unsupervised face image quality assessment with similarity distribution distance. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7670-7679, 2021. 4", + "[40] Zhengyuan Peng, Qijian Tian, Jianqing Xu, Yizhang Jin, Xuequan Lu, Xin Tan, Yuan Xie, and Lizhuang Ma. Generalized category discovery in semantic segmentation. arXiv preprint arXiv:2311.11525, 2023. 3", + "[41] Zhengyuan Peng, Jinpeng Ma, Zhimin Sun, Ran Yi, Haichuan Song, Xin Tan, and Lizhuang Ma. Mos: Modeling object-scene associations in generalized category discovery. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 15118-15128, 2025.", + "[42] Zhengyuan Peng, Jianqing Xu, Shen Li, Jiazhen Ji, Yuge Huang, Jingyun Zhang, Jinmin Li, Shouhong Ding, Rizen Guo, Xin Tan, et al. Eyeseg: An uncertainty-aware eye segmentation framework for ar/vr. arXiv preprint arXiv:2507.09649, 2025.3", + "[43] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 3, 4, 5", + "[44] Haibo Qiu, Baosheng Yu, Dihong Gong, Zhifeng Li, Wei Liu, and Dacheng Tao. Synface: Face recognition with synthetic data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10880-10890, 2021. 3", + "[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3", + "[46] Qibing Ren, Chang Gao, Jing Shao, Junchi Yan, Xin Tan, Wai Lam, and Lizhuang Ma. Codeattack: Revealing safety" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13062", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "generalization challenges of large language models via code completion. arXiv preprint arXiv:2403.07865, 2024. 3", + "[47] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3", + "[48] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 3", + "[49] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731, 2019. 3", + "[50] Yichen Shi, Yuhao Gao, Yingxin Lai, Hongyang Wang, Jun Feng, Lei He, Jun Wan, Changsheng Chen, Zitong Yu, and Xiaochun Cao. Shield: An evaluation benchmark for face spoofing and forgery detection with multimodal large language models. Visual Intelligence, 3(1):9, 2025. 3", + "[51] Kaede Shiohara and Toshihiko Yamasaki. Face2diffusion for fast and editable face personalization. arXiv preprint arXiv:2403.05094, 2024. 3", + "[52] Shuai Tan, Bin Ji, and Ye Pan. Style2talker: High-resolution talking head generation with emotion style and art style. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5079-5087, 2024. 3", + "[53] Shuai Tan, Biao Gong, Yutong Feng, Kecheng Zheng, Dandan Zheng, Shuwei Shi, Yujun Shen, Jingdong Chen, and Ming Yang. Mimir: Improving video diffusion models for precise text understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025.", + "[54] Shuai Tan, Biao Gong, Yujie Wei, Shiwei Zhang, Zhuoxin Liu, Dandan Zheng, Jingdong Chen, Yan Wang, Hao Ouyang, Kecheng Zheng, and Yujun Shen. Synmotion: Semantic-visual adaptation for motion customized video generation. arXiv preprint arXiv:2506.23690, 2025. 3", + "[55] Ming Tao, Hao Tang, Songsong Wu, Nicu Sebe, Xiao-Yuan Jing, Fei Wu, and Bingkun Bao. Df-gan: Deep fusion generative adversarial networks for text-to-image synthesis. arXiv preprint arXiv:2008.05865, 2(6), 2020. 3", + "[56] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5265-5274, 2018. 1, 3, 7, 8", + "[57] Qian Wang, Cai Guo, Hong-Ning Dai, and Ping Li. Stroke-gan painter: Learning to paint artworks using stroke-style generative adversarial networks. Computational Visual Media, 9(4):787-806, 2023. 3", + "[58] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 1, 2, 3, 4, 5, 6, 7" + ], + "bbox": [ + 91, + 92, + 483, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[59] Tan Wang, Linjie Li, Kevin Lin, Chung-Ching Lin, Zhengyuan Yang, Hanwang Zhang, Zicheng Liu, and Lijuan Wang. disco: Disentangled control for referring human dance generation in real world. arXiv e-prints, pages arXiv-2307, 2023. 3", + "[60] Cameron Whitelam, Emma Taborsky, Austin Blanton, Brianna Maze, Jocelyn Adams, Tim Miller, Nathan Kalka, Anil K Jain, James A Duncan, Kristen Allen, et al. Iarpa janus benchmark-b face dataset. In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 90-98, 2017. 3", + "[61] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multisubject image generation with localized attention. arXiv preprint arXiv:2305.10431, 2023. 3", + "[62] Jianqing Xu, Shen Li, Jiaying Wu, Miao Xiong, Ailin Deng, Jiazhen Ji, Yuge Huang, Guodong Mu, Wenjie Feng, Shouhong Ding, et al. $\\text{Id}^3$ : Identity-preserving-yet-diversified diffusion models for synthetic face recognition. Advances in Neural Information Processing Systems, 37: 77777-77798, 2024. 3", + "[63] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Finegrained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 3", + "[64] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 2, 3, 4, 5", + "[65] Dong Yi, Zhen Lei, Shengcai Liao, and Stan Z Li. Learning face representation from scratch. arXiv preprint arXiv:1411.7923, 2014. 2, 3, 4", + "[66] Ran Yi, Teng Hu, Mengfei Xia, Yizhe Tang, and Yong-Jin Liu. Feditnet++: Few-shot editing of latent semantics in gan spaces with correlated attribute disentanglement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3", + "[67] Shilong Zhang, Lianghua Huang, Xi Chen, Yifei Zhang, Zhi-Fan Wu, Yutong Feng, Wei Wang, Yujun Shen, Yu Liu, and Ping Luo. Flashface: Human image personalization with high-fidelity identity preservation. arXiv preprint arXiv:2403.17008, 2024. 1, 2, 3, 4, 5, 6", + "[68] Tianyue Zheng and Weihong Deng. Cross-pose lfw: A database for studying cross-posed face recognition in unconstrained environments. Beijing University of Posts and Telecommunications, Tech. Rep, 5(7):5, 2018. 3", + "[69] Tianyue Zheng, Weihong Deng, and Jiani Hu. Cross-age lfw: A database for studying cross-age face recognition in unconstrained environments. arXiv preprint arXiv:1708.08197, 2017.3", + "[70] Yi Zheng, Yifan Zhao, Mengyuan Ren, He Yan, Xiangju Lu, Junhui Liu, and Jia Li. Cartoon face recognition: A benchmark dataset. In Proceedings of the 28th ACM international conference on multimedia, pages 2264-2272, 2020. 3", + "[71] Zhizhou Zhong, Yuxi Mi, Yuge Huang, Jianqing Xu, Guodong Mu, Shouhong Ding, Jingyun Zhang, Rizen Guo," + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13063", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yunsheng Wu, and Shuigeng Zhou. Slerpface: Face template protection via spherical linear interpolation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10698-10706, 2025. 3", + "[72] Zheng Zhu, Guan Huang, Jiankang Deng, Yun Ye, Junjie Huang, Xinze Chen, Jiagang Zhu, Tian Yang, Jiwen Lu, Dalong Du, and Jie Zhou. Webface260m: A benchmark unveiling the power of million-scale deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10492-10502, 2021. 3" + ], + "bbox": [ + 91, + 90, + 480, + 244 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "13064", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Stylized-Face_ A Million-level Stylized Face Dataset for Face Recognition/fc351eb8-6a0f-465e-ac2f-6fa6da65b294_model.json b/2025/Stylized-Face_ A Million-level Stylized Face Dataset for Face Recognition/fc351eb8-6a0f-465e-ac2f-6fa6da65b294_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6ccc929899e2aad06b8373a4a4f48a09433418e1 --- /dev/null +++ b/2025/Stylized-Face_ A Million-level Stylized Face Dataset for Face Recognition/fc351eb8-6a0f-465e-ac2f-6fa6da65b294_model.json @@ -0,0 +1,3447 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.131, + 0.131, + 0.867, + 0.153 + ], + "angle": 0, + "content": "Stylized-Face: A Million-level Stylized Face Dataset for Face Recognition" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.18, + 0.766, + 0.199 + ], + "angle": 0, + "content": "Zhengyuan Peng\\(^{1*}\\), Jianqing Xu\\(^{2*}\\), Yuge Huang\\(^{2\\dagger}\\), Jinkun Hao\\(^{1}\\)," + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.199, + 0.765, + 0.217 + ], + "angle": 0, + "content": "Shouhong Ding\\(^{2}\\), Zhizhong Zhang\\(^{3}\\), Xin Tan\\(^{3}\\), Lizhuang Ma\\(^{1,3\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.163, + 0.217, + 0.836, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\)Shanghai Jiao Tong University, \\(^{2}\\)Youtu Lab, Tencent, \\(^{3}\\)East China Normal University" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.304, + 0.486, + 0.712 + ], + "angle": 0, + "content": "Stylized face recognition is the task of recognizing generated faces with the same ID across diverse stylistic domains (e.g., anime, painting, cyberpunk styles). This emerging field plays a vital role in the governance of generative image, serving the primary objective: Recognize the ID information of stylized faces to detect potential infringements of portrait rights. Despite its importance, progress in stylized face recognition has been hindered by the lack of large-scale, stylistically diverse datasets. To address this gap, we introduce the Stylized-Face dataset, which is the first dataset specifically designed for stylized face recognition. Stylized-Face dataset includes 4.6 million images across 62k IDs, specifically curated to enhance model performance in stylized face recognition tasks. To ensure data quality (i.e., ID preservation) at this massive scale, we implement a semi-automated pipeline for large-scale data cleaning. Based on the Stylized-Face dataset, we establish three benchmarks to evaluate the robustness and generalization of recognition models across various scenarios, including within-distribution performance, cross-prompt generalization, and cross-method generalization, which target key challenges in stylized face recognition. Experimental results demonstrate that models trained on Stylized-Face achieve remarkable improvements in both stylized face recognition performance (a \\(15.9\\%\\) improvement in TAR at \\(\\mathrm{FAR} = 1\\mathrm{e - }4\\)) and generalization (a \\(13.3\\%\\) improvement in TAR at \\(\\mathrm{FAR} = 1\\mathrm{e - }3\\) in cross-method generalization)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.744, + 0.222, + 0.759 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.77, + 0.484, + 0.862 + ], + "angle": 0, + "content": "In recent years, the rapid development of personalized diffusion methods has been notable. Their low barrier to entry enables users to easily generate a wide array of images, but this same accessibility also facilitates the production of content that may infringe on personal portrait rights, raising significant privacy concerns. Traditionally, face recognition" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.268, + 0.902, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.482, + 0.907, + 0.538 + ], + "angle": 0, + "content": "Figure 1. Illustration of the Stylized-Face dataset across diverse styles. The dataset features 80 distinct styles organized into four primary groups: Artistic Living, Classical Art Styles, Fantasy Styles, and ACG Styles (Animation, Comics, and Games)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.907, + 0.688 + ], + "angle": 0, + "content": "models have served as invaluable tools for detecting potential portrait infringements within images. However, the advent of these new technologies presents a more complex issue. Personalized diffusion methods can generate images in various styles (e.g., anime, painting, cyberpunk styles) simply by altering prompts. This stylistic diversity poses a challenge for recognition models in recognizing faces with the same ID across a broad spectrum of stylized images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.854 + ], + "angle": 0, + "content": "Previous methods [29, 67] addressing this challenge primarily rely on two models: CLIP-based models [45] and real face recognition models [13, 23, 56]. To systematically evaluate these models, we present a large-scale and diverse analysis across four different personalized diffusion models [29, 58, 64, 67] in Sec. 3.1. The results indicate that both exhibit significant limitations. Notably, CLIP-based models demonstrate considerable recognition biases. Although real face recognition models more closely mirror human perceptual judgments, their overall accuracy remains suboptimal. We illustrate failure cases in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Due to the limitations of the current recognition model, filling in the gaps in the relevant training data is an effective approach. We introduce a novel stylized face dataset," + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.221, + 0.888 + ], + "angle": 0, + "content": "*Equal Contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.237, + 0.9 + ], + "angle": 0, + "content": "† Corresponding Author" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.237, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13053" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.167, + 0.09, + 0.276, + 0.104 + ], + "angle": 0, + "content": "Reference Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.419, + 0.091, + 0.516, + 0.105 + ], + "angle": 0, + "content": "Stylized Image" + }, + { + "type": "image", + "bbox": [ + 0.168, + 0.108, + 0.252, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.173, + 0.18, + 0.249, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.174, + 0.252, + 0.249, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.173, + 0.324, + 0.25, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.268, + 0.111, + 0.343, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.297, + 0.169, + 0.32, + 0.179 + ], + "angle": 0, + "content": "0.31" + }, + { + "type": "image", + "bbox": [ + 0.349, + 0.111, + 0.421, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.376, + 0.17, + 0.4, + 0.179 + ], + "angle": 0, + "content": "0.26" + }, + { + "type": "image", + "bbox": [ + 0.428, + 0.111, + 0.501, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.454, + 0.17, + 0.477, + 0.179 + ], + "angle": 0, + "content": "0.36" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.111, + 0.58, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.17, + 0.555, + 0.179 + ], + "angle": 0, + "content": "0.42" + }, + { + "type": "image", + "bbox": [ + 0.585, + 0.111, + 0.658, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.613, + 0.17, + 0.635, + 0.179 + ], + "angle": 0, + "content": "0.19" + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.181, + 0.343, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.239, + 0.32, + 0.248 + ], + "angle": 0, + "content": "0.43" + }, + { + "type": "image", + "bbox": [ + 0.349, + 0.181, + 0.423, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.376, + 0.239, + 0.398, + 0.249 + ], + "angle": 0, + "content": "0.21" + }, + { + "type": "image", + "bbox": [ + 0.428, + 0.181, + 0.501, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.451, + 0.239, + 0.474, + 0.249 + ], + "angle": 0, + "content": "0.44" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.181, + 0.58, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.239, + 0.557, + 0.249 + ], + "angle": 0, + "content": "0.25" + }, + { + "type": "image", + "bbox": [ + 0.585, + 0.181, + 0.658, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.239, + 0.634, + 0.249 + ], + "angle": 0, + "content": "0.31" + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.252, + 0.343, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.311, + 0.319, + 0.32 + ], + "angle": 0, + "content": "0.36" + }, + { + "type": "image", + "bbox": [ + 0.349, + 0.252, + 0.423, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.311, + 0.398, + 0.32 + ], + "angle": 0, + "content": "0.46" + }, + { + "type": "image", + "bbox": [ + 0.428, + 0.252, + 0.501, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.453, + 0.311, + 0.475, + 0.32 + ], + "angle": 0, + "content": "0.45" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.252, + 0.58, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.311, + 0.558, + 0.32 + ], + "angle": 0, + "content": "0.32" + }, + { + "type": "image", + "bbox": [ + 0.585, + 0.252, + 0.658, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.311, + 0.635, + 0.32 + ], + "angle": 0, + "content": "0.36" + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.324, + 0.343, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.383, + 0.32, + 0.392 + ], + "angle": 0, + "content": "0.42" + }, + { + "type": "image", + "bbox": [ + 0.349, + 0.324, + 0.423, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.376, + 0.383, + 0.4, + 0.392 + ], + "angle": 0, + "content": "0.16" + }, + { + "type": "image", + "bbox": [ + 0.428, + 0.324, + 0.501, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.454, + 0.383, + 0.477, + 0.392 + ], + "angle": 0, + "content": "0.32" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.324, + 0.58, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.383, + 0.558, + 0.392 + ], + "angle": 0, + "content": "0.23" + }, + { + "type": "image", + "bbox": [ + 0.585, + 0.324, + 0.658, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.383, + 0.635, + 0.392 + ], + "angle": 0, + "content": "0.24" + }, + { + "type": "image_caption", + "bbox": [ + 0.278, + 0.399, + 0.33, + 0.41 + ], + "angle": 0, + "content": "Line Art" + }, + { + "type": "image_caption", + "bbox": [ + 0.359, + 0.399, + 0.421, + 0.411 + ], + "angle": 0, + "content": "Neonpunk" + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.399, + 0.489, + 0.411 + ], + "angle": 0, + "content": "Pop Art" + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.399, + 0.573, + 0.41 + ], + "angle": 0, + "content": "Film Noir" + }, + { + "type": "image_caption", + "bbox": [ + 0.588, + 0.399, + 0.665, + 0.41 + ], + "angle": 0, + "content": "Retro Arcade" + }, + { + "type": "image_caption", + "bbox": [ + 0.719, + 0.092, + 0.793, + 0.105 + ], + "angle": 0, + "content": "Real Image" + }, + { + "type": "image", + "bbox": [ + 0.677, + 0.111, + 0.75, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.701, + 0.17, + 0.725, + 0.179 + ], + "angle": 0, + "content": "0.78" + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.111, + 0.828, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.78, + 0.17, + 0.802, + 0.179 + ], + "angle": 0, + "content": "0.81" + }, + { + "type": "image", + "bbox": [ + 0.677, + 0.181, + 0.75, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.702, + 0.239, + 0.726, + 0.249 + ], + "angle": 0, + "content": "0.76" + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.181, + 0.828, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.78, + 0.239, + 0.802, + 0.249 + ], + "angle": 0, + "content": "0.84" + }, + { + "type": "image", + "bbox": [ + 0.677, + 0.252, + 0.75, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.703, + 0.311, + 0.726, + 0.32 + ], + "angle": 0, + "content": "0.71" + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.252, + 0.828, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.78, + 0.311, + 0.803, + 0.319 + ], + "angle": 0, + "content": "0.79" + }, + { + "type": "image", + "bbox": [ + 0.677, + 0.324, + 0.75, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.706, + 0.383, + 0.728, + 0.392 + ], + "angle": 0, + "content": "0.79" + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.324, + 0.828, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.782, + 0.383, + 0.805, + 0.392 + ], + "angle": 0, + "content": "0.87" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.425, + 0.907, + 0.495 + ], + "angle": 0, + "content": "Figure 2. Illustration of the Stylized-Face dataset and Comparison with Real Images. We randomly select a frontal face for each identity from the CASIA-WebFace [65] as a reference image. These reference images are compared with stylized images from our dataset and other images of the same identity from CASIA-WebFace [65], analyzing their feature similarity using a face recognition model trained on real faces. The results, recorded beneath each image, show that while the model performs well on real face images, it struggles to accurately identify the stylized images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.521, + 0.484, + 0.779 + ], + "angle": 0, + "content": "Stylized-Face, generated from the CASIA-WebFace [65] and Glint360k [6] datasets, as illustrated in Fig. 1 and Fig. 2. Stylized-Face is specifically designed to enhance model performance and generalization on stylized face recognition, particularly those that significantly deviate from real faces. It includes 4.6 million images across 62k IDs and 80 distinct styles. We employ InstantID[58] (the best-performing method in our evaluation) as the primary generative method, supplemented with three additional methods (PhotoMaker[29], Flashface[67], IPAdapter[64]) for dataset diversity. Owing to the inherent instability of generative model outputs, some images may feature faces that do not match the assigned IDs or might not contain faces at all. To maintain data quality at this scale, we use a semi-automatic process for data cleaning, which combines manual and model verification. To our knowledge, Stylized-Face is the first dataset dedicated to stylized face recognition." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To further refine the setup for stylized face recognition, we construct three benchmarks: 1) Standard Stylized Face Recognition. Models are trained and evaluated on multiple styles from the same distribution, assessing their ability to recognize stylized faces. 2) Cross-prompt stylized face recognition. Different prompts generate images with distribution differences. We utilize varied prompts for both training and evaluation to assess the generalization perfor" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.907, + 0.597 + ], + "angle": 0, + "content": "mance. 3) Cross-method stylized face recognition. Apart from the prompt, different generation methods yield varying distributions. We evaluate the generalization capabilities across the generation models by training and testing them on data produced through various techniques." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.907, + 0.734 + ], + "angle": 0, + "content": "We evaluate the baseline models using our benchmark. Experimental results demonstrate that models trained on Stylized-Face significantly enhance stylized face recognition performance, achieving a \\(15.9\\%\\) increase in TAR at \\(\\mathrm{FAR} = 1\\mathrm{e - }4\\). Additionally, they improve cross-prompt generalization by \\(12.2\\%\\) in TAR at \\(\\mathrm{FAR} = 1\\mathrm{e - }4\\) and cross-method generalization by \\(13.3\\%\\) in TAR at \\(\\mathrm{FAR} = 1\\mathrm{e - }3\\). These improvements underscore the value of our dataset for advancing stylized face recognition." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.734, + 0.905, + 0.749 + ], + "angle": 0, + "content": "To sum up, our major contributions are listed as follows:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.905, + 0.81 + ], + "angle": 0, + "content": "- We evaluate previous models for stylized face recognition and mainstream personalized diffusion methods for the ability to preserve ID information. Our findings reveal significant shortcomings of previous recognition models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.905, + 0.885 + ], + "angle": 0, + "content": "- We introduce a stylized face dataset, Stylized-Face. Stylized-Face dataset includes 4.6 million images across 62k IDs, created by advanced generative techniques and semi-automatic process for data cleaning. Notably, it is the first dataset for stylized face recognition." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "- We introduce three novel benchmarks designed to eval" + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.75, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13054" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "uate the performance of stylized face recognition models. We conduct a comprehensive evaluation of existing face recognition methods and our method using the Stylized-Face dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.168, + 0.233, + 0.184 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.194, + 0.26, + 0.21 + ], + "angle": 0, + "content": "2.1. Face Recognition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.216, + 0.484, + 0.428 + ], + "angle": 0, + "content": "Face recognition technology [5, 7, 12, 13, 23, 25, 31, 32, 35, 36, 50, 56, 62, 71] extract and analyze condensed face features, commonly referred to as facial templates. It employs advanced loss functions based on angular margins to enhance recognition accuracy. Angular margin-based approach [13, 23, 56] aims to maximize inter-identity discrepancies and minimize intra-identity discrepancies (similarities across various images of the same person). With the introduction of the Stylized-Face dataset, there is an increased demand for discriminative capabilities in recognition models. This demands models that can extract more crucial and generalized features to maintain identity consistency across diverse styles, a challenge inherent to Domain Generalization (DG) [14-16, 19, 33, 40-42, 46, 49]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.439, + 0.33, + 0.455 + ], + "angle": 0, + "content": "2.2. Face Recognition Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.483, + 0.612 + ], + "angle": 0, + "content": "Real Face Dataset: Existing real face recognition models predominantly rely on datasets like Webface42M[72], CASIA-WebFace [65], MS1Mv2 [17], MS1Mv3 [13], LFW [21], CPLFW[68], CALFW [69], IJB-B [60] and IJB-C [34]. These datasets comprise real-world face images and are designed to train and evaluate models to improve accuracy in real face applications. Stylized-Face dataset differs from the above as it explores diverse styles of dataset specific to generated face images, which significantly deviates from real face." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Synthetic Face Dataset: Earlier Synthetic datasets [8-10, 44] aim to enhance model performance on real face recognition and serve as substitutes for real face images. There are the main differences: (1) Authenticity of identity. We generate corresponding stylized images of real individuals; therefore, the IDs in our dataset are real, whereas existing datasets use virtual IDs. (2) Style. Our dataset includes various styles for the same ID, while existing datasets primarily focus on generating images in realistic style. (3) Objective. Our goal is designed for stylized face recognition, whereas existing datasets aims to enhance model performance on real face recognition and serve as substitutes for real face images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Some other works[4] attempt to generate stylized faces, but they only perform style transfer on thousands of images without checking the ID information in the stylized images, lacking the fundamental reliable ID information essential for face recognition tasks. Therefore, these works are not suitable for stylized face recognition tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.213 + ], + "angle": 0, + "content": "Cartoon Face Dataset: Some datasets[24, 37, 70] aims to explore different face style, such as cartoon face images. These datasets usually feature a single style and lack corresponding real human faces for comparison. Moreover, they are often sourced from animated films, which limits the data volume. In contrast, Stylized-Face dataset generates a large volume of face images across a wide range of styles, each paired with the ID of their corresponding real face." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.225, + 0.75, + 0.24 + ], + "angle": 0, + "content": "2.3. Text-to-Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.247, + 0.907, + 0.52 + ], + "angle": 0, + "content": "Text-to-image diffusion models [22, 26, 30, 38, 43, 47, 52-54, 59, 66] are at the forefront of image generation, renowned for their ability to transform text prompts into high-quality images with diverse styles. These models utilize a pre-trained language encoder to convert text into a latent representation, guiding the diffusion process to precisely adjust pixels. For instance, GLIDE [38] features a cascaded text-guided diffusion architecture for image generation. DisCo [59] introduces clip[45] encoder to guide the image generation. Stable Diffusion [47] optimizes the process by performing diffusion in a compressed latent space, which lowers computational costs. Further evolving this technology, Stable Diffusion XL [43] employs a larger network and additional text encoder to improve image fidelity and textual influence. Additionally, there are some GAN-based methods[11, 27, 28, 55, 57, 63], but they usually have weak generalization capabilities and are computationally expensive to train." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.531, + 0.785, + 0.545 + ], + "angle": 0, + "content": "2.4. Personalized diffusion models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In image generation, ID-preserving techniques prioritize key facial attributes to ensure identity remains consistent. These methods vary in their need for fine-tuning during testing. For instance, techniques such as DreamBooth [48] and LoRA [20] require extensive fine-tuning for each character, making them impractical for large-scale ID image generation. Recent developments aim to eliminate the need for fine-tuning altogether. Two fundamental methods include directly embedding the image features into the text embedding and using a cross-attention module to integrate the image encoding with the text encoding. The former includes methods such as Photomaker [29], FastComposer [61], and Face2Diffusion [51]. The latter comprises methods such as IP-Adapter-FaceID [64], InstantID [58], and Flashface [67]. In terms of specific designs, PhotoMaker [29] enhance identity preservation by modifying image encoders. IP-Adapter-FaceID [64] specifically utilizes face ID embeddings to ensure consistent identity retention. Moreover, InstantID [58] replaces the standard clip[45] encoder with face features and employs an additional IdentityNet, supervised by key facial landmarks, to guide face generation. FlashFace [67] encodes IDs into a series of feature maps to preserve finer face details. Our dataset utilizes these ID-preserving tech" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13055" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.419, + 0.09, + 0.579, + 0.103 + ], + "angle": 0, + "content": "Table 1. Styles Overview." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.115, + 0.891, + 0.35 + ], + "angle": 0, + "content": "
CategoryExample Style Prompts
1. Artistic Living
1.1 Photography Styles:film noir, analog film photo, long exposure photo, et al.
1.2 Fashion Styles:glamorous photo, low contrast realistic photo, minimalist style, et al.
2. Fantasy Styles
2.1 Sci-Fi and Futurism:sci-fi style, futuristic style, retro-futuristic, space-themed, et al.
2.2 Cyberpunk Styles:vaporwave style, neonpunk style, neon noir, dystopian style, et al.
2.3 Fantasy and Themed Styles:fairy tale, gothic style, nautical-themed, disco-themed, et al.
3. Classical Art Styles
3.1 Genre Art Styles:impressionist painting, cubist artwork, expressionist, renaissance style, et al.
3.2 Painting Arts:watercolor painting, line art drawing, graffiti style, urban art, et al.
3.3 Paper Art and Handicrafts:origami style, kirigami representation, 3D papercut shadow box, et al.
4. ACG Styles
4.1 Anime and Cartoon Styles:animework, manga style, kawaii style, et al.
4.2 Comics and Game Art:comic, role-playing game style fantasy, retro game art, pixel art, et al.
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.375, + 0.483, + 0.406 + ], + "angle": 0, + "content": "niques and enhances them by multiple prompt styles and pre-trained diffusion models." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.424, + 0.401, + 0.44 + ], + "angle": 0, + "content": "3. Dataset Construction and Method" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.451, + 0.483, + 0.481 + ], + "angle": 0, + "content": "3.1. Evaluation of Recognition Models and Personalized Diffusion Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.49, + 0.484, + 0.594 + ], + "angle": 0, + "content": "We begin by evaluating the effectiveness of existing evaluation models and personalized diffusion methods. For the evaluation metrics, we select two well-established metrics: CLIP-I, which measures identity preservation through CLIP features, and Face Sim., which is based on real face recognition models. For comparison, we also incorporate human evaluations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.763 + ], + "angle": 0, + "content": "In selecting personalized diffusion methods, we choose four of the latest ID-preserving image generation techniques: PhotoMaker [29], InstanceID [58], Flashface [67], and IPAdapter [64]. All methods are implemented using the officially recommended settings. We analyze additional methods that are excluded from our study in Appendix. To ensure a comprehensive assessment, we additionally utilize the FID metric [18] and SDD-FIQA [39] to evaluate the overall image quality. We employ CLIP-T to evaluate the consistency between the stylized images and their corresponding textual prompts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To comprehensively evaluate the different methods, we incorporate a diverse array of artistic styles. We first classify these common styles into 4 major categories and 10 sub-categories, as shown in Tab. 1. Our classification is based on people's usual perception of art and common forms of expression, ensuring that the grouping reflects intuitive and widely recognized distinctions in artistic presentation. For each sub-category, we curate a collection of specific prompts that together cover most common scenarios." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.373, + 0.697, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.588, + 0.469, + 0.648, + 0.481 + ], + "angle": 0, + "content": "IPAdapter" + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.373, + 0.904, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.8, + 0.469, + 0.856, + 0.479 + ], + "angle": 0, + "content": "Flashface" + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.481, + 0.696, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.577, + 0.643, + 0.588 + ], + "angle": 0, + "content": "InstantID" + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.481, + 0.905, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.795, + 0.577, + 0.865, + 0.588 + ], + "angle": 0, + "content": "Photomaker" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.628 + ], + "angle": 0, + "content": "Figure 3. Quality Evaluation. We evaluate the generation quality of different methods using SDD-FIQA [39]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.907, + 0.794 + ], + "angle": 0, + "content": "Our collection is both diverse and extensive, comprising: 1) we use official SDXL prompts [43], carefully selected for their high quality and distinctive style; 2) we also collect popular repository resources [3]. After deduplicating and removing prompts that conflict with identity information, we retain 80 distinct style prompts that cover a broad spectrum of common scenarios. We use 1,000 IDs from the CASIA-WebFace[65] dataset, and each method generates 80,000 images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We present the results of our analysis in Tab. 2 and Fig. 3. Our findings indicate that the similarity scores derived from CLIP features lack sufficient discriminative power and exhibit bias when compared to human evaluations. This limitation stems from CLIP's focus on extracting a generalized representation of identity. In contrast, real face recognition models align more closely with human judgment. For fur" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13056" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.172, + 0.089, + 0.825, + 0.104 + ], + "angle": 0, + "content": "Table 2. Comparison of Different Personalized Diffusion Methods Based on Various Evaluation Metrics." + }, + { + "type": "table", + "bbox": [ + 0.188, + 0.115, + 0.812, + 0.215 + ], + "angle": 0, + "content": "
MethodPreserve ID InformationText ConsistencyGeneration Quality
CLIP-IFace Sim.Human EvalCLIP-TFID
PhotoMaker [29]54.5%12.3%21.9%20.5%313
InstanceID [58]65.2%70.9%90.0%26.1%285
Flashface [67]65.6%35.9%87.1%22.9%244
IPAdapter [67]60.3%25.3%64.9%25.1%255
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.24, + 0.482, + 0.316 + ], + "angle": 0, + "content": "ther study, a detailed assessment of real face recognition models is provided in the experiment section. The evaluation under classification metrics indicates that real face recognition models are also inaccurate in assessing identity consistency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.317, + 0.483, + 0.379 + ], + "angle": 0, + "content": "In terms of the specific method, InstantID achieves the best performance in preserving ID information, text preservation, and quality distribution, demonstrating strong performance. In terms of FID score, Flashface performs better." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.392, + 0.321, + 0.409 + ], + "angle": 0, + "content": "3.2. Setup and Data Cleaning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.484, + 0.536 + ], + "angle": 0, + "content": "Based on Sec. 3.1, InstantID [58] provides stable performance, making it an ideal choice for our task. As a result, we choose InstantID as the primary method for generation. To further enhance the diversity of the dataset, we use three Pre-trained Diffusion Models:SDXL [43], YamerMIX [2], Anything-XL [1]. To further enrich the dataset, we incorporate three additional methods: PhotoMaker [29], Flashface [67] and IPAdapter [64]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.537, + 0.484, + 0.793 + ], + "angle": 0, + "content": "We use 10k IDs from CISIA-Webface and 52k IDs from Glint360k as the original images. To enhance generation quality, we apply a series of filtering steps to the original dataset. These steps include: 1) discarding images with low similarity to others within the same ID; 2) removing low-quality images based on SDD-FIQA results, specifically those with scores below 60, which eliminates blurry or heavily tilted images; and 3) filtering out images with face angles exceeding 30 degrees, based on human pose detection. After applying these filtering steps, \\(78\\%\\) of the face images are retained. We generate the dataset by iterating over all combinations of methods, pre-trained diffusion models, and prompts across the entire dataset, producing more than 10 million images. We directly skip combinations where over \\(50\\%\\) of the images are identified by all annotators as not matching the original ID will be entirely excluded." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.903 + ], + "angle": 0, + "content": "For the stylized images, we begin with preprocessing. We use the RetinaFace model for face detection, adjusting the detected faces based on five face landmarks. Notably, \\(8.4\\%\\) of the images fail to detect any faces. In such cases, we resize the original image directly to maintain a consistent image size. We also compared additional face detection models, with more details provided in Appendix." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.238, + 0.903, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.475, + 0.907, + 0.683 + ], + "angle": 0, + "content": "Figure 4. Pipeline of Semi-automatic Process. D. indicates delete, M. indicates manual verification, Sim. BSI measures the similarity between stylized images, Sim. BOI measures the similarity between original images, and Sim. BIP measures the similarity between a stylized image and its corresponding original image. We alternate between two phases: model verification and manual verification. In the model image pairs, a trained face recognition model performs two types of filtering: 1) The model compares the similarity between a query image and other stylized images generated from the highly similar original face. If the similarity score (TAR@FAR 1e-3) falls below the threshold, the query image is discarded. 2) It removes image pairs (stylized image and its origin image) with low similarity (TAR@FAR 1e-2). During the manual verification phase, these confusion image pairs (between TAR@FAR 1e-2 and TAR@FAR 1e-3) are labeled." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.717, + 0.906, + 0.837 + ], + "angle": 0, + "content": "Due to the instability of generative methods, we need to verify whether each image truly preserves the ID information. However, given the large scale of the data, comprehensive manual verification becomes impractical. To alleviate the annotation burden, we employ a semi-automatic process for data cleaning, as shown in Fig. 4. In the semi-automated process, we alternate between two main phases: model validation and manual verification." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In the model validation phase, we first train a face recognition model using the remaining dataset. We use two methods for rapid filtering by trained model: 1) Images with a similarity to original real faces below a threshold" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13057" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "(TAR@FAR at 1e-2) are deleted. 2) Generally, images generated from similar original images tend to have higher similarity; if there is a significant deviation, it is highly likely to be a bad case. Images with similarity to other stylized images from the same ID below the threshold (TAR@FAR at 1e-3), while similarity between their corresponding original images greater than 0.8, are also deleted." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.199, + 0.483, + 0.333 + ], + "angle": 0, + "content": "In the manual validation phase, annotators compare the confusion images (between TAR@FAR at 1e-2 and TAR@FAR at 1e-3) to verify the ID information. During the annotation process, we develop a tool that provides yes and no options for each image pair, which are confirmed by the annotators. If more than \\(50\\%\\) of the annotators believe that the two images belong to the same ID, they are retained. Any images below this threshold are excluded from the dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.342, + 0.258, + 0.358 + ], + "angle": 0, + "content": "3.3. Dataset statistics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.483, + 0.438 + ], + "angle": 0, + "content": "Large scale images. The Stylized-Face dataset consists of 4,660,285 images of 62,446 individuals. The generation size of all images is determined by the default parameters of their methods, and ultimately, they are sized at \\(112 \\times 112\\) pixels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.44, + 0.484, + 0.515 + ], + "angle": 0, + "content": "Utilized Models. Tab. 3 illustrates the proportions of various methods employed in our dataset. The higher proportion of the InstantID(SD) method is due to its more common use in real-world applications. We provide a detailed comparison of these methods in Appendix." + }, + { + "type": "table_caption", + "bbox": [ + 0.147, + 0.528, + 0.428, + 0.542 + ], + "angle": 0, + "content": "Table 3. The proportions of various methods" + }, + { + "type": "table", + "bbox": [ + 0.136, + 0.553, + 0.44, + 0.676 + ], + "angle": 0, + "content": "
MethodsProportion
InstantID(SD) [58]69%
InstantID(YamerMIX) [58]14%
InstantID(Anything-XL) [58]14%
PhotoMaker [29]1%
Flashface [67]1%
IPAdapter [67]1%
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.765 + ], + "angle": 0, + "content": "Mean Faces. Fig. 5 displays the average faces for individual IDs from our dataset compared with those from CASIA-WebFace. Notably, while our dataset exhibit considerable style variations, the core facial features remain relatively stable." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.779, + 0.465, + 0.797 + ], + "angle": 0, + "content": "4. Experiments of Stylized Face Recognition" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.804, + 0.282, + 0.818 + ], + "angle": 0, + "content": "4.1. Evaluation Protocol" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Standard Stylized Face Recognition. The training set consists of 4,453,814 images generated by the InstantID method, encompassing 61k IDs. Additionally, we also use the real face images corresponding origin dataset for training. The test set includes 1k independent IDs, with 56,644" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.091, + 0.613, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.16, + 0.613, + 0.182 + ], + "angle": 0, + "content": "CISIA-WebFace II" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.092, + 0.701, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.627, + 0.16, + 0.698, + 0.181 + ], + "angle": 0, + "content": "Stylized-Face" + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.092, + 0.807, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.725, + 0.16, + 0.892, + 0.182 + ], + "angle": 0, + "content": "CISIA-WebFace Stylized-Face ID 2" + }, + { + "type": "image", + "bbox": [ + 0.814, + 0.092, + 0.897, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.821, + 0.16, + 0.892, + 0.181 + ], + "angle": 0, + "content": "Stylized-Face" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.195, + 0.906, + 0.249 + ], + "angle": 0, + "content": "Figure 5. Display of average faces for single ID. We randomly select two IDs from the CISIA-WebFace dataset and compute the average face for each ID using the images corresponding to those IDs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.277, + 0.906, + 0.338 + ], + "angle": 0, + "content": "images also generated by InstantID. Both the training and test sets utilize the full set of 80 prompts. The primary goal of this evaluation is to assess the model's performance in open-set stylized face recognition." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.338, + 0.906, + 0.534 + ], + "angle": 0, + "content": "Cross-prompt Stylized Face Recognition. The training and test sets for Cross-prompt Stylized Face Recognition are subsets of the corresponding training and test sets for Standard Stylized Face Recognition. The training set only includes images generated using 60 prompts from the Standard Stylized Face Recognition training set. It contains a total of 3,811,707 images. Additionally, the real face images corresponding to the IDs is also used for training. The test set consists of images generated using the last 20 prompts from the Standard Stylized Face Recognition test set, totaling 42,361 images. This setting ensures that the prompts and IDs used in the training and test sets are distinct. More details are shown in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.906, + 0.655 + ], + "angle": 0, + "content": "Cross-method Stylized Face Recognition. In this setup, we aim to assess cross-method generalization. The training set is the same as that used in Standard Stylized Face Recognition. The test set consists of 1k independent IDs, with 149,827 images generated by Flashface, IP-Adapter, and Photomaker. In detail, Flashface contributed 70,912 images, IP-Adapter contributed 51,915 images, and Photomaker contributed 27,000 images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.664, + 0.697, + 0.679 + ], + "angle": 0, + "content": "4.2. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.686, + 0.906, + 0.791 + ], + "angle": 0, + "content": "In all experiments, we follow the standard evaluation metrics for face recognition. We evaluate the performance of the model on face verification. The gallery consists of all images in the test set, excluding the test images themselves. We report the results at multiple points, including TAR@FAR at 1e-2, TAR@FAR at 1e-3, and TAR@FAR at 1e-4. The metrics are defined as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.794, + 0.906, + 0.839 + ], + "angle": 0, + "content": "- True Acceptance Rate (TAR) (complementary to False Non-Match Rate (FNMR)): Proportion of correctly identified valid identities:" + }, + { + "type": "equation", + "bbox": [ + 0.673, + 0.847, + 0.759, + 0.877 + ], + "angle": 0, + "content": "\\[\nT A R = \\frac {M}{N}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "where \\(M\\) is the number of correctly identified identities," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13058" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.104, + 0.092, + 0.448, + 0.106 + ], + "angle": 0, + "content": "and \\(N\\) is the total number of valid identity attempts." + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.107, + 0.297, + 0.12 + ], + "angle": 0, + "content": "Note: \\( {FNMR} = 1 - {TAR} \\)" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.122, + 0.483, + 0.168 + ], + "angle": 0, + "content": "- False Acceptance Rate (FAR) (equivalent to False Match Rate (FMR)): Proportion of non-target identities incorrectly accepted:" + }, + { + "type": "equation", + "bbox": [ + 0.252, + 0.171, + 0.337, + 0.202 + ], + "angle": 0, + "content": "\\[\nF A R = \\frac {L}{K}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.205, + 0.483, + 0.235 + ], + "angle": 0, + "content": "where \\( L \\) is the number of incorrect acceptances, and \\( K \\) is the number of non-target identity attempts." + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.236, + 0.258, + 0.249 + ], + "angle": 0, + "content": "Note: \\( {FMR} = {FAR} \\)" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.251, + 0.483, + 0.297 + ], + "angle": 0, + "content": "- TAR@FAR: Represents the True Acceptance Rate (TAR) at a specific False Acceptance Rate (FAR), used to evaluate system accuracy at a fixed security level." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.303, + 0.308, + 0.319 + ], + "angle": 0, + "content": "4.3. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.325, + 0.484, + 0.462 + ], + "angle": 0, + "content": "We maintain the same setup for all experiments. Our backbone networks are IResNet-50 and IResNet-101. We apply Cosface [56] framework and use an angular margin of 0.3 for Cosface. We set the batch size to 512 and train for a total of 26 epochs. The models are subsequently trained using Stochastic Gradient Descent (SGD), with a momentum coefficient of 0.9 and a weight decay parameter of \\(5 \\times 10^{-4}\\). The initial learning rate is set at \\(1 \\times 10^{-1}\\) and is decreased by a factor of 8 at the midpoint of the epochs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.468, + 0.178, + 0.483 + ], + "angle": 0, + "content": "4.4. Result" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.491, + 0.389, + 0.506 + ], + "angle": 0, + "content": "4.4.1. Standard Stylized Face Recognition." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.483, + 0.644 + ], + "angle": 0, + "content": "We assess the model's performance trained on the large-scale MS1MV3 [13] and Glint360k [6] dataset. As presented in Tab. 4, the results indicate that despite the substantial size of the training set, there is a significant performance degradation compared to real face scenarios. Although the model trained on Glint360k performs worse than ours, it shows significant improvement over the model trained on MS1MV3, highlighting the positive impact of dataset size on generalization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.646, + 0.483, + 0.736 + ], + "angle": 0, + "content": "The model trained on our dataset significantly outperforms the one trained on real face data in open-set face recognition scenarios, achieving a face verification TAR@FAR at 1e-4 of \\(90.6\\%\\), as detailed in Tab. 4. This highlights the crucial role of our dataset in effectively recognizing stylized faces." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.746, + 0.483, + 0.775 + ], + "angle": 0, + "content": "Table 4. Results on Standard Stylized Face Recognition. Backbone is IResNet-50. Values in bold indicate the top results." + }, + { + "type": "table", + "bbox": [ + 0.164, + 0.786, + 0.413, + 0.858 + ], + "angle": 0, + "content": "
Training sets1e-21e-31e-4
MS1MV3 [13]85.168.049.7
Gint360k [6]95.087.074.7
Stylized-Face97.995.590.6
" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig. 6, the model trained on our dataset significantly outperforms the one trained on real face data in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.136 + ], + "angle": 0, + "content": "open-set face recognition scenarios. It highlights the crucial role of our dataset in effectively recognizing stylized faces." + }, + { + "type": "image", + "bbox": [ + 0.54, + 0.162, + 0.892, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.352, + 0.906, + 0.38 + ], + "angle": 0, + "content": "Figure 6. the FMR-FNMR plot for All pairs verification. (lower FNMR is better)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.41, + 0.833, + 0.425 + ], + "angle": 0, + "content": "4.4.2. Cross-prompt stylized face recognition." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.906, + 0.519 + ], + "angle": 0, + "content": "Table 5 displays the performance across various prompts using the InstanceID [58], highlighting its generalization capabilities in stylized face recognition across different prompts. It also underscores our dataset's effectiveness in enhancing the cross-prompt generalization of stylized face recognition." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.534, + 0.906, + 0.562 + ], + "angle": 0, + "content": "Table 5. Results on Cross-prompt Stylized Face Recognition. Backbone is IResNet-50. Values in bold indicate the top results." + }, + { + "type": "table", + "bbox": [ + 0.587, + 0.573, + 0.835, + 0.646 + ], + "angle": 0, + "content": "
Training sets1e-21e-31e-4
MS1MV3 [13]85.767.749.3
Gint360k [6]94.786.573.9
Stylized-Face97.995.490.3
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.67, + 0.842, + 0.685 + ], + "angle": 0, + "content": "4.4.3. Cross-method Stylized Face Recognition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.794 + ], + "angle": 0, + "content": "Cross-method experiments in Table 6 evaluate the ability of models to generalize feature extraction across various generation methods. The results indicate that, despite our method showing clear superiority over real-face models, current methods struggle with cross-method generalization across different methods, emphasizing the need for more in-depth research to enhance cross-method robustness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We further analyze cross-method generalization using the Flashface dataset. Prompts are divided into easy and hard categories based on the distribution shift, measured by the difference between Flashface and the InstantID method on the same data. Results show that performance on easy prompts is similar to within-distribution performance, while generalization challenges are more pronounced for hard" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13059" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.147 + ], + "angle": 0, + "content": "Table 6. Results on Cross-method Stylized Face Recognition. Backbone is IResNet-50. All refers to Flashface, IP-Adapter, and Photomaker. Due to the limited number of images, the results exclude TAR@FAR at 1e-4. Values in bold indicate the top results." + }, + { + "type": "table", + "bbox": [ + 0.141, + 0.156, + 0.437, + 0.374 + ], + "angle": 0, + "content": "
Training setsTest Parts1e-21e-3
MS1MV3 [13]Flashface82.171.0
Gint360k [6]Flashface92.386.7
Stylized-FaceFlashface93.187.3
MS1MV3 [13]IP-Adapter68.262.2
Gint360k [6]IP-Adapter91.474.3
Stylized-FaceIP-Adapter93.678.3
MS1MV3 [13]Photomaker62.636.4
Gint360k [6]Photomaker86.562.2
Stylized-FacePhotomaker90.171.2
MS1MV3 [13]All59.031.8
Gint360k [6]All78.758.6
Stylized-FaceAll87.271.9
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.395, + 0.485, + 0.457 + ], + "angle": 0, + "content": "prompts with significant distributional shifts. Notably, for these hard prompts, our model demonstrates particularly remarkable performance improvements compared to the real face recognition model. Results are shown in Tab. 7." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.468, + 0.485, + 0.51 + ], + "angle": 0, + "content": "Table 7. Analysis of Cross-method generalization. The evaluation metrics is TAR@FAR at 1e-2 on Flashface. Backbone is IResNet-50." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.52, + 0.465, + 0.581 + ], + "angle": 0, + "content": "
MethodsEasy PromptsHard PromptsAll
Glint360k96.984.992.3
Stylized-Face96.987.193.1
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.601, + 0.245, + 0.618 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.623, + 0.486, + 0.716 + ], + "angle": 0, + "content": "Performance on Real Face Datasets. We evaluate the model's performance on real face datasets, with the results presented in Fig. 8. The comparison demonstrates that while our model achieves comparable performance, there exists a trade-off between real face recognition and stylized face recognition." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.726, + 0.485, + 0.755 + ], + "angle": 0, + "content": "Table 8. Results on Real Face Recognition. Backbone is IResNet-50." + }, + { + "type": "table", + "bbox": [ + 0.119, + 0.765, + 0.455, + 0.873 + ], + "angle": 0, + "content": "
Training setLFWCFP_FPAgeDBAvg.
CASIA99.595.294.996.5
MS1MV299.898.198.198.7
MS1MV399.898.598.398.9
Glint360k99.899.198.599.1
Stylized-Face99.798.497.998.7
" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.485, + 0.903 + ], + "angle": 0, + "content": "More Methods To comprehensively assess the perfor" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.184 + ], + "angle": 0, + "content": "mance of stylized face recognition dataset, we compare advanced face recognition methods, including Cosface [56], Arcface [13], and CurricularFace [23]. These methods represent leading techniques in the field of face recognition. The experimental results are summarized in Tab. 9. The performance of the different training methods is similar." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.195, + 0.907, + 0.224 + ], + "angle": 0, + "content": "Table 9. Performance Comparison of Different Training Methods on Standard Stylized Face Recognition." + }, + { + "type": "table", + "bbox": [ + 0.572, + 0.234, + 0.851, + 0.307 + ], + "angle": 0, + "content": "
Training Methods1e-21e-31e-4
Cosface [56]97.995.590.6
CurricularFace [23]97.995.590.5
Arcface [13]97.995.489.9
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.321, + 0.907, + 0.368 + ], + "angle": 0, + "content": "More Backbones We evaluate the performance of various backbones and Tab. 10 shows that scaling up the model size can lead to further improvements in performance." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.379, + 0.907, + 0.408 + ], + "angle": 0, + "content": "Table 10. Performance Comparison of Different Backbones on Standard Stylized Face Recognition." + }, + { + "type": "table", + "bbox": [ + 0.593, + 0.419, + 0.83, + 0.478 + ], + "angle": 0, + "content": "
Backbone1e-21e-31e-4
IResNet-5097.995.590.6
IResNet-10198.396.392.0
" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.505, + 0.634, + 0.521 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.907, + 0.728 + ], + "angle": 0, + "content": "In addressing the challenge of recognizing stylized face images, we introduce the Stylized-Face dataset based on the CASIA-WebFace and Glint360k. Our dataset includes 4.6 million images spanning 62k IDs and incorporate a wide range of styles and identity-preserving methods. To ensure data quality (i.e., ID preservation) at this massive scale, we implement a semi-automated pipeline for large-scale data cleaning. Stylized-Face is the first dataset for stylized face recognition. Furthermore, we establish benchmarks to evaluate the robustness and generalization of stylized face recognition methods. Experimental results demonstrate that models trained on Stylized-Face achieve remarkable improvements in stylized face recognition." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.74, + 0.78, + 0.756 + ], + "angle": 0, + "content": "6. Limitation and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This task continues to present challenges and opportunities for improvement. (1) Cross-method Stylized Face Recognition requires the adoption of additional designs to improve its generalization capabilities. (2) Explore approaches that can simultaneously achieve state-of-the-art performance in both real and stylized face recognition (3) Furthermore, considering the various long-tail distribution styles in real-world scenarios, which is also a direction for future research." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13060" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.283 + ], + "angle": 0, + "content": "This work is supported by the National Natural Science Foundation of China (62302167, U23A20343, 62472282, 72192821, 62476090), Young Elite Scientists Sponsorship Program by CAST (YESS20240780), Shanghai Sailing Program (23YF1410500), Chenguang Program of Shanghai Education Development Foundation, CCF-Tencent Rhino-Bird Young Faculty Open Research Fund (CCF-Tencent RAGR20230121, RAGR20240122), Shanghai Municipal Education Commission (23CGA34) and the Open Research Fund of Key Laboratory of Advanced Theory and Application in Statistics and Data Science-MOE, ECNU." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.295, + 0.188, + 0.311 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.321, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[1] Anything-xl. https://civitai.com/models/9409/or-anything-xl.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.35, + 0.483, + 0.377 + ], + "angle": 0, + "content": "[2] Yamermix. https://civitai.com/models/84040/sdxl-unstable-diffusers-yamermix. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.38, + 0.482, + 0.407 + ], + "angle": 0, + "content": "[3] Sdsl prompt styler, a custom node for comfyui. https://github.com/twri/sdsl_prompt_styler. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.409, + 0.482, + 0.435 + ], + "angle": 0, + "content": "[4] Reverse style transfer. https://devpost.com/software/reverse-style-transfer. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.483, + 0.519 + ], + "angle": 0, + "content": "[5] Xiang An, Xuhan Zhu, Yuan Gao, Yang Xiao, Yongle Zhao, Ziyong Feng, Lan Wu, Bin Qin, Ming Zhang, Debing Zhang, and Ying Fu. Partial fc: Training 10 million identities on a single machine. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.522, + 0.483, + 0.603 + ], + "angle": 0, + "content": "[6] Xiang An, Xuhan Zhu, Yuan Gao, Yang Xiao, Yongle Zhao, Ziyong Feng, Lan Wu, Bin Qin, Ming Zhang, Debing Zhang, et al. Partial fc: Training 10 million identities on a single machine. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1445-1449, 2021. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.483, + 0.675 + ], + "angle": 0, + "content": "[7] Xiang An, Jiangkang Deng, Jia Guo, Ziyong Feng, Xuhan Zhu, Yang Jing, and Liu Tongliang. Killing two birds with one stone: Efficient and robust training of face recognitionCNNs by partial fc. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.677, + 0.483, + 0.759 + ], + "angle": 0, + "content": "[8] Gwangbin Bae, Martin de La Gorce, Tadas Baltrusaitis, Charlie Hewitt, Dong Chen, Julien Valentin, Roberto Cipolla, and Jingjing Shen. Digiface-1m: 1 million digital face images for face recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3526-3535, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.762, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[9] Fadi Boutros, Marco Huber, Patrick Siebke, Tim Rieber, and Naser Damer. Sface: Privacy-friendly and accurate face recognition using synthetic data. In 2022 IEEE International Joint Conference on Biometrics (IJCB), pages 1-11. IEEE, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.832, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[10] Fadi Boutros, Marcel Klemt, Meiling Fang, Arjan Kuijper, and Naser Damer. Unsupervised face recognition using unlabeled synthetic data. In 2023 IEEE 17th International Conference on Automatic Face and Gesture Recognition (FG), pages 1-8. IEEE, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.321, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[11] PAN Chuanyu, YANG Guowei, MU Taijiang, and LAI Yu-Kun. Generating animatable 3d cartoon faces from single portraits. Virtual Reality & Intelligent Hardware, 6(4):292-307, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.906, + 0.232 + ], + "angle": 0, + "content": "[12] Jun Dan, Yang Liu, Haoyu Xie, Jiankang Deng, Haoran Xie, Xuansong Xie, and Baigui Sun. Transface: Calibrating transformer training for face recognition from a data-centric perspective. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20642-20653, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[13] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 1, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.344 + ], + "angle": 0, + "content": "[14] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pages 1180-1189. PMLR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.345, + 0.905, + 0.427 + ], + "angle": 0, + "content": "[15] Jingyu Gong, Fengqi Liu, Jiachen Xu, Min Wang, Xin Tan, Zhizhong Zhang, Ran Yi, Haichuan Song, Yuan Xie, and Lizhuang Ma. Optimization over disentangled encoding: Unsupervised cross-domain point cloud completion via occlusion factor manipulation. In European Conference on Computer Vision, pages 517-533. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.428, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[16] Jingyu Gong, Yujing Lou, Fengqi Liu, Zhiwei Zhang, Haoming Chen, Zhizhong Zhang, Xin Tan, Yuan Xie, and Lizhuang Ma. Scene point cloud understanding and reconstruction technologies in 3d space. Journal of Image and Graphics, 28(06):1741-1766, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.498, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[17] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 87-102. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.651, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[19] Chengyang Hu, Ke-Yue Zhang, Taiping Yao, Shice Liu, Shouhong Ding, Xin Tan, and Lizhuang Ma. Domainhallucinated updating for multi-domain face anti-spoofing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2193-2201, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[20] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.777, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[21] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[22] Lianghua Huang, Di Chen, Yu Liu, Yujun Shen, Deli Zhao, and Jingren Zhou. Composer: Creative and controllable image synthesis with composable conditions. arXiv preprint arXiv:2302.09778, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13061" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[23] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: adaptive curriculum learning loss for deep face recognition. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5901-5910, 2020. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[24] Jing Huoa, Wenbin Lia, Yinghuan Shia, Yang Gaoa, and Hujun Yinb. Webcaricature: a benchmark for caricature face recognition. arXiv preprint arXiv:1703.03230, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[25] Shen Li, Jianqing Xu, Xiaqing Xu, Pengcheng Shen, Shaoxin Li, and Bryan Hooi. Spherical confidence learning for face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15629-15637, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[26] Xiaofan Li, Xin Tan, Zhuo Chen, Zhizhong Zhang, Ruixin Zhang, Rizen Guo, Guanna Jiang, Yulong Chen, Yanyun Qu, Lizhuang Ma, et al. One-for-more: Continual diffusion model for anomaly detection. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 4766-4775, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[27] Yidi Li, Jun Xiao, Yiqun Wang, and Zhengda Lu. Depthgan: Gan-based depth generation from semantic layouts. Computational Visual Media, 10(3):505-522, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[28] Yuxuan Li, Lingfeng Yang, and Xiang Li. Apf-gan: Exploring asymmetric pre-training and fine-tuning strategy for conditional generative adversarial network. Computational Visual Media, 10(1):187-192, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[29] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. arXiv preprint arXiv:2312.04461, 2023. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.482, + 0.575 + ], + "angle": 0, + "content": "[30] Zonglin Li, Xiaogian Lv, Wei Yu, Qinglin Liu, Jingbo Lin, and Shengping Zhang. Face shape transfer via semantic warping. Visual Intelligence, 2(1):26, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.577, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[31] Xiao Lin, Yuge Huang, Jianqing Xu, Yuxi Mi, Shuigeng Zhou, and Shouhong Ding. Uface: Unleashing inherent model capabilities to enhance intra-class diversity in synthetic face recognition. arXiv preprint arXiv:2502.19803, 2025.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[32] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherical face: Deep hypersphere embedding for face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 212-220, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[33] Wang Lu, Jindong Wang, Haoliang Li, Yiqiang Chen, and Xing Xie. Domain-invariant feature exploration for domain generalization. arXiv preprint arXiv:2207.12020, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[34] Brianna Maze, Jocelyn Adams, James A Duncan, Nathan Kalka, Tim Miller, Charles Otto, Anil K Jain, W Tyler Niggel, Janet Anderson, Jordan Cheney, et al. Iarpa janus benchmark-c: Face dataset and protocol. In 2018 international conference on biometrics (ICB), pages 158-165. IEEE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[35] Yuxi Mi, Zhizhou Zhong, Yuge Huang, Jiazhen Ji, Jianqing Xu, Jun Wang, Shaoming Wang, Shouhong Ding, and Shuigeng Zhou. Privacy-preserving face recognition using trainable feature subtraction. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 297-307, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.206 + ], + "angle": 0, + "content": "[36] Yuxi Mi, Zhizhou Zhong, Yuge Huang, Qiuyang Yuan, Xuan Zhao, Jianqing Xu, Shouhong Ding, Shaoming Wang, Rizen Guo, and Shuigeng Zhou. Data synthesis with diverse styles for face recognition via 3dmm-guided diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21203-21214, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[37] Ashutosh Mishra, Shyam Nandan Rai, Anand Mishra, and CV Jawahar. Iit-cfw: A benchmark database of cartoon faces in the wild. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part I 14, pages 35-47. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[38] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[39] Fu-Zhao Ou, Xingyu Chen, Ruixin Zhang, Yuge Huang, Shaoxin Li, Jilin Li, Yong Li, Liujuan Cao, and Yuan-Gen Wang. Sdd-fiaq: Unsupervised face image quality assessment with similarity distribution distance. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7670-7679, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[40] Zhengyuan Peng, Qijian Tian, Jianqing Xu, Yizhang Jin, Xuequan Lu, Xin Tan, Yuan Xie, and Lizhuang Ma. Generalized category discovery in semantic segmentation. arXiv preprint arXiv:2311.11525, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[41] Zhengyuan Peng, Jinpeng Ma, Zhimin Sun, Ran Yi, Haichuan Song, Xin Tan, and Lizhuang Ma. Mos: Modeling object-scene associations in generalized category discovery. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 15118-15128, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[42] Zhengyuan Peng, Jianqing Xu, Shen Li, Jiazhen Ji, Yuge Huang, Jingyun Zhang, Jinmin Li, Shouhong Ding, Rizen Guo, Xin Tan, et al. Eyeseg: An uncertainty-aware eye segmentation framework for ar/vr. arXiv preprint arXiv:2507.09649, 2025.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[43] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[44] Haibo Qiu, Baosheng Yu, Dihong Gong, Zhifeng Li, Wei Liu, and Dacheng Tao. Synface: Face recognition with synthetic data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10880-10890, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[46] Qibing Ren, Chang Gao, Jing Shao, Junchi Yan, Xin Tan, Wai Lam, and Lizhuang Ma. Codeattack: Revealing safety" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13062" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "generalization challenges of large language models via code completion. arXiv preprint arXiv:2403.07865, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.189 + ], + "angle": 0, + "content": "[47] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.482, + 0.272 + ], + "angle": 0, + "content": "[48] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[49] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[50] Yichen Shi, Yuhao Gao, Yingxin Lai, Hongyang Wang, Jun Feng, Lei He, Jun Wan, Changsheng Chen, Zitong Yu, and Xiaochun Cao. Shield: An evaluation benchmark for face spoofing and forgery detection with multimodal large language models. Visual Intelligence, 3(1):9, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[51] Kaede Shiohara and Toshihiko Yamasaki. Face2diffusion for fast and editable face personalization. arXiv preprint arXiv:2403.05094, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[52] Shuai Tan, Bin Ji, and Ye Pan. Style2talker: High-resolution talking head generation with emotion style and art style. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5079-5087, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.592 + ], + "angle": 0, + "content": "[53] Shuai Tan, Biao Gong, Yutong Feng, Kecheng Zheng, Dandan Zheng, Shuwei Shi, Yujun Shen, Jingdong Chen, and Ming Yang. Mimir: Improving video diffusion models for precise text understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[54] Shuai Tan, Biao Gong, Yujie Wei, Shiwei Zhang, Zhuoxin Liu, Dandan Zheng, Jingdong Chen, Yan Wang, Hao Ouyang, Kecheng Zheng, and Yujun Shen. Synmotion: Semantic-visual adaptation for motion customized video generation. arXiv preprint arXiv:2506.23690, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[55] Ming Tao, Hao Tang, Songsong Wu, Nicu Sebe, Xiao-Yuan Jing, Fei Wu, and Bingkun Bao. Df-gan: Deep fusion generative adversarial networks for text-to-image synthesis. arXiv preprint arXiv:2008.05865, 2(6), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.721, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[56] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5265-5274, 2018. 1, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[57] Qian Wang, Cai Guo, Hong-Ning Dai, and Ping Li. Stroke-gan painter: Learning to paint artworks using stroke-style generative adversarial networks. Computational Visual Media, 9(4):787-806, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[58] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.16 + ], + "angle": 0, + "content": "[59] Tan Wang, Linjie Li, Kevin Lin, Chung-Ching Lin, Zhengyuan Yang, Hanwang Zhang, Zicheng Liu, and Lijuan Wang. disco: Disentangled control for referring human dance generation in real world. arXiv e-prints, pages arXiv-2307, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[60] Cameron Whitelam, Emma Taborsky, Austin Blanton, Brianna Maze, Jocelyn Adams, Tim Miller, Nathan Kalka, Anil K Jain, James A Duncan, Kristen Allen, et al. Iarpa janus benchmark-b face dataset. In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 90-98, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[61] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multisubject image generation with localized attention. arXiv preprint arXiv:2305.10431, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.383 + ], + "angle": 0, + "content": "[62] Jianqing Xu, Shen Li, Jiaying Wu, Miao Xiong, Ailin Deng, Jiazhen Ji, Yuge Huang, Guodong Mu, Wenjie Feng, Shouhong Ding, et al. \\( \\text{Id}^3 \\): Identity-preserving-yet-diversified diffusion models for synthetic face recognition. Advances in Neural Information Processing Systems, 37: 77777-77798, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.905, + 0.467 + ], + "angle": 0, + "content": "[63] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Finegrained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[64] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.564 + ], + "angle": 0, + "content": "[65] Dong Yi, Zhen Lei, Shengcai Liao, and Stan Z Li. Learning face representation from scratch. arXiv preprint arXiv:1411.7923, 2014. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.634 + ], + "angle": 0, + "content": "[66] Ran Yi, Teng Hu, Mengfei Xia, Yizhe Tang, and Yong-Jin Liu. Feditnet++: Few-shot editing of latent semantics in gan spaces with correlated attribute disentanglement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[67] Shilong Zhang, Lianghua Huang, Xi Chen, Yifei Zhang, Zhi-Fan Wu, Yutong Feng, Wei Wang, Yujun Shen, Yu Liu, and Ping Luo. Flashface: Human image personalization with high-fidelity identity preservation. arXiv preprint arXiv:2403.17008, 2024. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[68] Tianyue Zheng and Weihong Deng. Cross-pose lfw: A database for studying cross-posed face recognition in unconstrained environments. Beijing University of Posts and Telecommunications, Tech. Rep, 5(7):5, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[69] Tianyue Zheng, Weihong Deng, and Jiani Hu. Cross-age lfw: A database for studying cross-age face recognition in unconstrained environments. arXiv preprint arXiv:1708.08197, 2017.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[70] Yi Zheng, Yifan Zhao, Mengyuan Ren, He Yan, Xiangju Lu, Junhui Liu, and Jia Li. Cartoon face recognition: A benchmark dataset. In Proceedings of the 28th ACM international conference on multimedia, pages 2264-2272, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[71] Zhizhou Zhong, Yuxi Mi, Yuge Huang, Jianqing Xu, Guodong Mu, Shouhong Ding, Jingyun Zhang, Rizen Guo," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13063" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "Yunsheng Wu, and Shuigeng Zhou. Slerpface: Face template protection via spherical linear interpolation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10698-10706, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[72] Zheng Zhu, Guan Huang, Jiankang Deng, Yun Ye, Junjie Huang, Xinze Chen, Jiagang Zhu, Tian Yang, Jiwen Lu, Dalong Du, and Jie Zhou. Webface260m: A benchmark unveiling the power of million-scale deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10492-10502, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "13064" + } + ] +] \ No newline at end of file diff --git a/2025/Stylized-Face_ A Million-level Stylized Face Dataset for Face Recognition/fc351eb8-6a0f-465e-ac2f-6fa6da65b294_origin.pdf b/2025/Stylized-Face_ A Million-level Stylized Face Dataset for Face Recognition/fc351eb8-6a0f-465e-ac2f-6fa6da65b294_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..615edb2f90fc22f4b08c770c82ee936a78a9936a --- /dev/null +++ b/2025/Stylized-Face_ A Million-level Stylized Face Dataset for Face Recognition/fc351eb8-6a0f-465e-ac2f-6fa6da65b294_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f501f5d7e6f3b0fc87f4f36566274e34a7ac3135b0bb587da95acfac3745502d +size 1599364 diff --git a/2025/Stylized-Face_ A Million-level Stylized Face Dataset for Face Recognition/full.md b/2025/Stylized-Face_ A Million-level Stylized Face Dataset for Face Recognition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fff2e4ec0dd1ce414885846acb526d8477344fd8 --- /dev/null +++ b/2025/Stylized-Face_ A Million-level Stylized Face Dataset for Face Recognition/full.md @@ -0,0 +1,456 @@ +# Stylized-Face: A Million-level Stylized Face Dataset for Face Recognition + +Zhengyuan Peng $^{1*}$ , Jianqing Xu $^{2*}$ , Yuge Huang $^{2\dagger}$ , Jinkun Hao $^{1}$ , + +Shouhong Ding $^{2}$ , Zhizhong Zhang $^{3}$ , Xin Tan $^{3}$ , Lizhuang Ma $^{1,3\dagger}$ + +$^{1}$ Shanghai Jiao Tong University, $^{2}$ Youtu Lab, Tencent, $^{3}$ East China Normal University + +# Abstract + +Stylized face recognition is the task of recognizing generated faces with the same ID across diverse stylistic domains (e.g., anime, painting, cyberpunk styles). This emerging field plays a vital role in the governance of generative image, serving the primary objective: Recognize the ID information of stylized faces to detect potential infringements of portrait rights. Despite its importance, progress in stylized face recognition has been hindered by the lack of large-scale, stylistically diverse datasets. To address this gap, we introduce the Stylized-Face dataset, which is the first dataset specifically designed for stylized face recognition. Stylized-Face dataset includes 4.6 million images across 62k IDs, specifically curated to enhance model performance in stylized face recognition tasks. To ensure data quality (i.e., ID preservation) at this massive scale, we implement a semi-automated pipeline for large-scale data cleaning. Based on the Stylized-Face dataset, we establish three benchmarks to evaluate the robustness and generalization of recognition models across various scenarios, including within-distribution performance, cross-prompt generalization, and cross-method generalization, which target key challenges in stylized face recognition. Experimental results demonstrate that models trained on Stylized-Face achieve remarkable improvements in both stylized face recognition performance (a $15.9\%$ improvement in TAR at $\mathrm{FAR} = 1\mathrm{e - }4$ ) and generalization (a $13.3\%$ improvement in TAR at $\mathrm{FAR} = 1\mathrm{e - }3$ in cross-method generalization). + +# 1. Introduction + +In recent years, the rapid development of personalized diffusion methods has been notable. Their low barrier to entry enables users to easily generate a wide array of images, but this same accessibility also facilitates the production of content that may infringe on personal portrait rights, raising significant privacy concerns. Traditionally, face recognition + +![](images/465156445b391cb165cb770b5bad8de3b85776091d196f6ea7e5aff19a698da3.jpg) +Figure 1. Illustration of the Stylized-Face dataset across diverse styles. The dataset features 80 distinct styles organized into four primary groups: Artistic Living, Classical Art Styles, Fantasy Styles, and ACG Styles (Animation, Comics, and Games). + +models have served as invaluable tools for detecting potential portrait infringements within images. However, the advent of these new technologies presents a more complex issue. Personalized diffusion methods can generate images in various styles (e.g., anime, painting, cyberpunk styles) simply by altering prompts. This stylistic diversity poses a challenge for recognition models in recognizing faces with the same ID across a broad spectrum of stylized images. + +Previous methods [29, 67] addressing this challenge primarily rely on two models: CLIP-based models [45] and real face recognition models [13, 23, 56]. To systematically evaluate these models, we present a large-scale and diverse analysis across four different personalized diffusion models [29, 58, 64, 67] in Sec. 3.1. The results indicate that both exhibit significant limitations. Notably, CLIP-based models demonstrate considerable recognition biases. Although real face recognition models more closely mirror human perceptual judgments, their overall accuracy remains suboptimal. We illustrate failure cases in Fig. 2. + +Due to the limitations of the current recognition model, filling in the gaps in the relevant training data is an effective approach. We introduce a novel stylized face dataset, + +![](images/fcb2fca84b4e6a97360b8a0c5776a3d32d5f3db8b17e5b850f5eba89fc04de75.jpg) +Reference Image + +![](images/4453d9c85933a961fdfac719b27ca4d8d324424106acc644c3dfdb764a634c6b.jpg) + +![](images/d8e5c4853dfbca0c0239a8baa07061819411b573585dafdca845bcc2526ffb28.jpg) + +![](images/ac3343c2399de544f214e3c65fc8bd335212ad6045f61d6e11d6232675d0a0a0.jpg) +Line Art + +![](images/2b2b12cf7055007de7eaf68b0c6f714c7b8510e50f3102e38096f3322ed277b2.jpg) +0.31 + +![](images/1d02d7d4c7bed319b7cc7456a87e3aea1f401ab1ae98fadae8e697b55ae39e6e.jpg) +0.26 + +![](images/a93b25614bfeb52a18410cbf308eeba75668ff4628b7a88c16a55da07ccf5a38.jpg) +Stylized Image + +![](images/b794ddef1b526ecab4a5abcfe31d1d8e6ac7cb4498a49aeb489c5173a6ccb1e5.jpg) +0.42 + +![](images/dc939fe1a3dae834e5db12bdb19a8897fa5faeeabb3799c324f02043951dacc4.jpg) +0.19 + +![](images/a1339114f21d345dc6b2c9bd1085c4908b000d1a546a611a15e9d3caa1c9669c.jpg) +0.43 + +![](images/4c695cdc7a8703d1a9aee3dfc7370b7e8f118b9654ef04856c2b3a9399091b2f.jpg) +0.21 + +![](images/744f6d7afe503442fe284c32859dadf023610e3b41e05bf50c9938f712b51192.jpg) +0.36 + +![](images/729f60c52b1bbde85f3578452fe0a49e17c5be8ff72bd182f5bbadacc3d572e4.jpg) +0.25 + +![](images/79f4f682b09261715a1a4203f8df8968cfe341b698f0d4821aa00f2b18e079df.jpg) +0.31 + +![](images/dd7c1457891b2b166d2d5862bd876fdbe30c32697d6c25217aac5cfecb674bb7.jpg) +0.36 + +![](images/04958b94f7574a545b2270975a7a54ef961e6c3271e5af00d78a298b9caefab7.jpg) +0.46 + +![](images/d1b9378d7c76e1afd600b34f173ec0f5bf85a59df7d325718f54d01c65ae00bc.jpg) +0.44 +0.45 + +![](images/50d4769982fb7009ef4f894f81186ab9a3fa8a31b38d26752107105f021c206b.jpg) +0.32 + +![](images/340c8a6d813370222e3b78aca8e90ee03ac839abbd9324e475ae6ff1b9483528.jpg) +0.36 + +![](images/6d48a6eaf8a4800e3d36eef9f553fe8390a69f08073e42a697720156c4211cfb.jpg) +0.42 + +![](images/958d861fcc247708faf9b3250c8d593be9b053119453849ea10e1b3c0cde54e7.jpg) +0.16 +Neonpunk + +![](images/92eacb424526efb71e274f924f20929e69ff8f8a3c39a79b5b6ee30ddcf74119.jpg) +0.32 +Pop Art +Figure 2. Illustration of the Stylized-Face dataset and Comparison with Real Images. We randomly select a frontal face for each identity from the CASIA-WebFace [65] as a reference image. These reference images are compared with stylized images from our dataset and other images of the same identity from CASIA-WebFace [65], analyzing their feature similarity using a face recognition model trained on real faces. The results, recorded beneath each image, show that while the model performs well on real face images, it struggles to accurately identify the stylized images. + +![](images/4ab50453c2e7cc0c7e513b581f2986d5821d58b013575f2a96d9cbb63ab59d7f.jpg) +0.23 +Film Noir + +![](images/a08ea658c1ffc893e99e2d6207a76e0bc68457c45eccccaf2d11365fbb5c2d51.jpg) +0.24 +Retro Arcade + +![](images/f315b5db148131e68ceeb1f8c38c70d3299a22feb7c2f6bd334b6e16def57a47.jpg) +Real Image +0.78 + +![](images/ceb6aa35fd87d80f5bb4d00475d258c1ec6c57a55a924e2934af5036a0fba825.jpg) +0.81 + +![](images/a2afffa86c4f145c513e05321e622274cd1624ef511871c7029305a7f1865a4d.jpg) +0.76 + +![](images/7fb41f536dc415a8f9da97e7d21f40fe179a240789ad820c21ca873ecd2f1594.jpg) +0.84 + +![](images/7a01dcebbb8a210b671380025b1ef5e7317dc801239316a98809d751d2c3eb3c.jpg) +0.71 + +![](images/36c635950a9ca2bc9b187b43b7d8126c064813d99eedcb27e38129f60d1b75cc.jpg) +0.79 + +![](images/e59a0ef3095de3fd786f64d32453f3430d5a249612d2de7bc8099378156ba19b.jpg) +0.79 + +![](images/0798fff24f9169c0322c13bc888519ef0531c653855e699d4961ae02f26378a2.jpg) +0.87 + +Stylized-Face, generated from the CASIA-WebFace [65] and Glint360k [6] datasets, as illustrated in Fig. 1 and Fig. 2. Stylized-Face is specifically designed to enhance model performance and generalization on stylized face recognition, particularly those that significantly deviate from real faces. It includes 4.6 million images across 62k IDs and 80 distinct styles. We employ InstantID[58] (the best-performing method in our evaluation) as the primary generative method, supplemented with three additional methods (PhotoMaker[29], Flashface[67], IPAdapter[64]) for dataset diversity. Owing to the inherent instability of generative model outputs, some images may feature faces that do not match the assigned IDs or might not contain faces at all. To maintain data quality at this scale, we use a semi-automatic process for data cleaning, which combines manual and model verification. To our knowledge, Stylized-Face is the first dataset dedicated to stylized face recognition. + +To further refine the setup for stylized face recognition, we construct three benchmarks: 1) Standard Stylized Face Recognition. Models are trained and evaluated on multiple styles from the same distribution, assessing their ability to recognize stylized faces. 2) Cross-prompt stylized face recognition. Different prompts generate images with distribution differences. We utilize varied prompts for both training and evaluation to assess the generalization perfor + +mance. 3) Cross-method stylized face recognition. Apart from the prompt, different generation methods yield varying distributions. We evaluate the generalization capabilities across the generation models by training and testing them on data produced through various techniques. + +We evaluate the baseline models using our benchmark. Experimental results demonstrate that models trained on Stylized-Face significantly enhance stylized face recognition performance, achieving a $15.9\%$ increase in TAR at $\mathrm{FAR} = 1\mathrm{e - }4$ . Additionally, they improve cross-prompt generalization by $12.2\%$ in TAR at $\mathrm{FAR} = 1\mathrm{e - }4$ and cross-method generalization by $13.3\%$ in TAR at $\mathrm{FAR} = 1\mathrm{e - }3$ . These improvements underscore the value of our dataset for advancing stylized face recognition. + +To sum up, our major contributions are listed as follows: + +- We evaluate previous models for stylized face recognition and mainstream personalized diffusion methods for the ability to preserve ID information. Our findings reveal significant shortcomings of previous recognition models. +- We introduce a stylized face dataset, Stylized-Face. Stylized-Face dataset includes 4.6 million images across 62k IDs, created by advanced generative techniques and semi-automatic process for data cleaning. Notably, it is the first dataset for stylized face recognition. +- We introduce three novel benchmarks designed to eval + +uate the performance of stylized face recognition models. We conduct a comprehensive evaluation of existing face recognition methods and our method using the Stylized-Face dataset. + +# 2. Related Work + +# 2.1. Face Recognition + +Face recognition technology [5, 7, 12, 13, 23, 25, 31, 32, 35, 36, 50, 56, 62, 71] extract and analyze condensed face features, commonly referred to as facial templates. It employs advanced loss functions based on angular margins to enhance recognition accuracy. Angular margin-based approach [13, 23, 56] aims to maximize inter-identity discrepancies and minimize intra-identity discrepancies (similarities across various images of the same person). With the introduction of the Stylized-Face dataset, there is an increased demand for discriminative capabilities in recognition models. This demands models that can extract more crucial and generalized features to maintain identity consistency across diverse styles, a challenge inherent to Domain Generalization (DG) [14-16, 19, 33, 40-42, 46, 49]. + +# 2.2. Face Recognition Datasets + +Real Face Dataset: Existing real face recognition models predominantly rely on datasets like Webface42M[72], CASIA-WebFace [65], MS1Mv2 [17], MS1Mv3 [13], LFW [21], CPLFW[68], CALFW [69], IJB-B [60] and IJB-C [34]. These datasets comprise real-world face images and are designed to train and evaluate models to improve accuracy in real face applications. Stylized-Face dataset differs from the above as it explores diverse styles of dataset specific to generated face images, which significantly deviates from real face. + +Synthetic Face Dataset: Earlier Synthetic datasets [8-10, 44] aim to enhance model performance on real face recognition and serve as substitutes for real face images. There are the main differences: (1) Authenticity of identity. We generate corresponding stylized images of real individuals; therefore, the IDs in our dataset are real, whereas existing datasets use virtual IDs. (2) Style. Our dataset includes various styles for the same ID, while existing datasets primarily focus on generating images in realistic style. (3) Objective. Our goal is designed for stylized face recognition, whereas existing datasets aims to enhance model performance on real face recognition and serve as substitutes for real face images. + +Some other works[4] attempt to generate stylized faces, but they only perform style transfer on thousands of images without checking the ID information in the stylized images, lacking the fundamental reliable ID information essential for face recognition tasks. Therefore, these works are not suitable for stylized face recognition tasks. + +Cartoon Face Dataset: Some datasets[24, 37, 70] aims to explore different face style, such as cartoon face images. These datasets usually feature a single style and lack corresponding real human faces for comparison. Moreover, they are often sourced from animated films, which limits the data volume. In contrast, Stylized-Face dataset generates a large volume of face images across a wide range of styles, each paired with the ID of their corresponding real face. + +# 2.3. Text-to-Image Generation + +Text-to-image diffusion models [22, 26, 30, 38, 43, 47, 52-54, 59, 66] are at the forefront of image generation, renowned for their ability to transform text prompts into high-quality images with diverse styles. These models utilize a pre-trained language encoder to convert text into a latent representation, guiding the diffusion process to precisely adjust pixels. For instance, GLIDE [38] features a cascaded text-guided diffusion architecture for image generation. DisCo [59] introduces clip[45] encoder to guide the image generation. Stable Diffusion [47] optimizes the process by performing diffusion in a compressed latent space, which lowers computational costs. Further evolving this technology, Stable Diffusion XL [43] employs a larger network and additional text encoder to improve image fidelity and textual influence. Additionally, there are some GAN-based methods[11, 27, 28, 55, 57, 63], but they usually have weak generalization capabilities and are computationally expensive to train. + +# 2.4. Personalized diffusion models + +In image generation, ID-preserving techniques prioritize key facial attributes to ensure identity remains consistent. These methods vary in their need for fine-tuning during testing. For instance, techniques such as DreamBooth [48] and LoRA [20] require extensive fine-tuning for each character, making them impractical for large-scale ID image generation. Recent developments aim to eliminate the need for fine-tuning altogether. Two fundamental methods include directly embedding the image features into the text embedding and using a cross-attention module to integrate the image encoding with the text encoding. The former includes methods such as Photomaker [29], FastComposer [61], and Face2Diffusion [51]. The latter comprises methods such as IP-Adapter-FaceID [64], InstantID [58], and Flashface [67]. In terms of specific designs, PhotoMaker [29] enhance identity preservation by modifying image encoders. IP-Adapter-FaceID [64] specifically utilizes face ID embeddings to ensure consistent identity retention. Moreover, InstantID [58] replaces the standard clip[45] encoder with face features and employs an additional IdentityNet, supervised by key facial landmarks, to guide face generation. FlashFace [67] encodes IDs into a series of feature maps to preserve finer face details. Our dataset utilizes these ID-preserving tech + +Table 1. Styles Overview. + +
CategoryExample Style Prompts
1. Artistic Living
1.1 Photography Styles:film noir, analog film photo, long exposure photo, et al.
1.2 Fashion Styles:glamorous photo, low contrast realistic photo, minimalist style, et al.
2. Fantasy Styles
2.1 Sci-Fi and Futurism:sci-fi style, futuristic style, retro-futuristic, space-themed, et al.
2.2 Cyberpunk Styles:vaporwave style, neonpunk style, neon noir, dystopian style, et al.
2.3 Fantasy and Themed Styles:fairy tale, gothic style, nautical-themed, disco-themed, et al.
3. Classical Art Styles
3.1 Genre Art Styles:impressionist painting, cubist artwork, expressionist, renaissance style, et al.
3.2 Painting Arts:watercolor painting, line art drawing, graffiti style, urban art, et al.
3.3 Paper Art and Handicrafts:origami style, kirigami representation, 3D papercut shadow box, et al.
4. ACG Styles
4.1 Anime and Cartoon Styles:animework, manga style, kawaii style, et al.
4.2 Comics and Game Art:comic, role-playing game style fantasy, retro game art, pixel art, et al.
+ +niques and enhances them by multiple prompt styles and pre-trained diffusion models. + +# 3. Dataset Construction and Method + +# 3.1. Evaluation of Recognition Models and Personalized Diffusion Methods + +We begin by evaluating the effectiveness of existing evaluation models and personalized diffusion methods. For the evaluation metrics, we select two well-established metrics: CLIP-I, which measures identity preservation through CLIP features, and Face Sim., which is based on real face recognition models. For comparison, we also incorporate human evaluations. + +In selecting personalized diffusion methods, we choose four of the latest ID-preserving image generation techniques: PhotoMaker [29], InstanceID [58], Flashface [67], and IPAdapter [64]. All methods are implemented using the officially recommended settings. We analyze additional methods that are excluded from our study in Appendix. To ensure a comprehensive assessment, we additionally utilize the FID metric [18] and SDD-FIQA [39] to evaluate the overall image quality. We employ CLIP-T to evaluate the consistency between the stylized images and their corresponding textual prompts. + +To comprehensively evaluate the different methods, we incorporate a diverse array of artistic styles. We first classify these common styles into 4 major categories and 10 sub-categories, as shown in Tab. 1. Our classification is based on people's usual perception of art and common forms of expression, ensuring that the grouping reflects intuitive and widely recognized distinctions in artistic presentation. For each sub-category, we curate a collection of specific prompts that together cover most common scenarios. + +![](images/f0c954e0a182b7ecb678241fb1a7d70788eeda937bb61d7effbd8ce023721e5d.jpg) + +![](images/d94eaf62b28e07f65a97c77d88177fcedeee56e047e1b228ca46ce6078d73e07.jpg) + +![](images/87a5da8a13c6d56a6381cf14c8405289f048d3a7fbb8b4a39faea1d7d19dafbe.jpg) +IPAdapter +InstantID +Figure 3. Quality Evaluation. We evaluate the generation quality of different methods using SDD-FIQA [39]. + +![](images/8b4a658085449901092217fd1b433a465c02d7f29eea4c3ed14240b06da41d7e.jpg) +Flashface +Photomaker + +Our collection is both diverse and extensive, comprising: 1) we use official SDXL prompts [43], carefully selected for their high quality and distinctive style; 2) we also collect popular repository resources [3]. After deduplicating and removing prompts that conflict with identity information, we retain 80 distinct style prompts that cover a broad spectrum of common scenarios. We use 1,000 IDs from the CASIA-WebFace[65] dataset, and each method generates 80,000 images. + +We present the results of our analysis in Tab. 2 and Fig. 3. Our findings indicate that the similarity scores derived from CLIP features lack sufficient discriminative power and exhibit bias when compared to human evaluations. This limitation stems from CLIP's focus on extracting a generalized representation of identity. In contrast, real face recognition models align more closely with human judgment. For fur + +Table 2. Comparison of Different Personalized Diffusion Methods Based on Various Evaluation Metrics. + +
MethodPreserve ID InformationText ConsistencyGeneration Quality
CLIP-IFace Sim.Human EvalCLIP-TFID
PhotoMaker [29]54.5%12.3%21.9%20.5%313
InstanceID [58]65.2%70.9%90.0%26.1%285
Flashface [67]65.6%35.9%87.1%22.9%244
IPAdapter [67]60.3%25.3%64.9%25.1%255
+ +ther study, a detailed assessment of real face recognition models is provided in the experiment section. The evaluation under classification metrics indicates that real face recognition models are also inaccurate in assessing identity consistency. + +In terms of the specific method, InstantID achieves the best performance in preserving ID information, text preservation, and quality distribution, demonstrating strong performance. In terms of FID score, Flashface performs better. + +# 3.2. Setup and Data Cleaning + +Based on Sec. 3.1, InstantID [58] provides stable performance, making it an ideal choice for our task. As a result, we choose InstantID as the primary method for generation. To further enhance the diversity of the dataset, we use three Pre-trained Diffusion Models:SDXL [43], YamerMIX [2], Anything-XL [1]. To further enrich the dataset, we incorporate three additional methods: PhotoMaker [29], Flashface [67] and IPAdapter [64]. + +We use 10k IDs from CISIA-Webface and 52k IDs from Glint360k as the original images. To enhance generation quality, we apply a series of filtering steps to the original dataset. These steps include: 1) discarding images with low similarity to others within the same ID; 2) removing low-quality images based on SDD-FIQA results, specifically those with scores below 60, which eliminates blurry or heavily tilted images; and 3) filtering out images with face angles exceeding 30 degrees, based on human pose detection. After applying these filtering steps, $78\%$ of the face images are retained. We generate the dataset by iterating over all combinations of methods, pre-trained diffusion models, and prompts across the entire dataset, producing more than 10 million images. We directly skip combinations where over $50\%$ of the images are identified by all annotators as not matching the original ID will be entirely excluded. + +For the stylized images, we begin with preprocessing. We use the RetinaFace model for face detection, adjusting the detected faces based on five face landmarks. Notably, $8.4\%$ of the images fail to detect any faces. In such cases, we resize the original image directly to maintain a consistent image size. We also compared additional face detection models, with more details provided in Appendix. + +![](images/3fd2d0e242c099542ecad6e1d7d457bff8717d5826576adaa7de6f9ec16ec084.jpg) +Figure 4. Pipeline of Semi-automatic Process. D. indicates delete, M. indicates manual verification, Sim. BSI measures the similarity between stylized images, Sim. BOI measures the similarity between original images, and Sim. BIP measures the similarity between a stylized image and its corresponding original image. We alternate between two phases: model verification and manual verification. In the model image pairs, a trained face recognition model performs two types of filtering: 1) The model compares the similarity between a query image and other stylized images generated from the highly similar original face. If the similarity score (TAR@FAR 1e-3) falls below the threshold, the query image is discarded. 2) It removes image pairs (stylized image and its origin image) with low similarity (TAR@FAR 1e-2). During the manual verification phase, these confusion image pairs (between TAR@FAR 1e-2 and TAR@FAR 1e-3) are labeled. + +Due to the instability of generative methods, we need to verify whether each image truly preserves the ID information. However, given the large scale of the data, comprehensive manual verification becomes impractical. To alleviate the annotation burden, we employ a semi-automatic process for data cleaning, as shown in Fig. 4. In the semi-automated process, we alternate between two main phases: model validation and manual verification. + +In the model validation phase, we first train a face recognition model using the remaining dataset. We use two methods for rapid filtering by trained model: 1) Images with a similarity to original real faces below a threshold + +(TAR@FAR at 1e-2) are deleted. 2) Generally, images generated from similar original images tend to have higher similarity; if there is a significant deviation, it is highly likely to be a bad case. Images with similarity to other stylized images from the same ID below the threshold (TAR@FAR at 1e-3), while similarity between their corresponding original images greater than 0.8, are also deleted. + +In the manual validation phase, annotators compare the confusion images (between TAR@FAR at 1e-2 and TAR@FAR at 1e-3) to verify the ID information. During the annotation process, we develop a tool that provides yes and no options for each image pair, which are confirmed by the annotators. If more than $50\%$ of the annotators believe that the two images belong to the same ID, they are retained. Any images below this threshold are excluded from the dataset. + +# 3.3. Dataset statistics + +Large scale images. The Stylized-Face dataset consists of 4,660,285 images of 62,446 individuals. The generation size of all images is determined by the default parameters of their methods, and ultimately, they are sized at $112 \times 112$ pixels. + +Utilized Models. Tab. 3 illustrates the proportions of various methods employed in our dataset. The higher proportion of the InstantID(SD) method is due to its more common use in real-world applications. We provide a detailed comparison of these methods in Appendix. + +Table 3. The proportions of various methods + +
MethodsProportion
InstantID(SD) [58]69%
InstantID(YamerMIX) [58]14%
InstantID(Anything-XL) [58]14%
PhotoMaker [29]1%
Flashface [67]1%
IPAdapter [67]1%
+ +Mean Faces. Fig. 5 displays the average faces for individual IDs from our dataset compared with those from CASIA-WebFace. Notably, while our dataset exhibit considerable style variations, the core facial features remain relatively stable. + +# 4. Experiments of Stylized Face Recognition + +# 4.1. Evaluation Protocol + +Standard Stylized Face Recognition. The training set consists of 4,453,814 images generated by the InstantID method, encompassing 61k IDs. Additionally, we also use the real face images corresponding origin dataset for training. The test set includes 1k independent IDs, with 56,644 + +![](images/67c2cc1f17e39ac7f64e38380b84262ee469cc785cde1bfcecd8a8c44ef326e8.jpg) +CISIA-WebFace II + +![](images/355814a46a3c471a3698697306813e813181212f81cb46d93d41c6d0f508beeb.jpg) +Stylized-Face + +![](images/314d148384962168cedb61bcc40c019027408045aa400583827e2b427dc778e0.jpg) +Figure 5. Display of average faces for single ID. We randomly select two IDs from the CISIA-WebFace dataset and compute the average face for each ID using the images corresponding to those IDs. + +![](images/c8545101ca8549dcbb266a189e41c5c5bc5a88db91a1c81429fe45a15be7cdfd.jpg) +CISIA-WebFace Stylized-Face ID 2 + +Stylized-Face + +images also generated by InstantID. Both the training and test sets utilize the full set of 80 prompts. The primary goal of this evaluation is to assess the model's performance in open-set stylized face recognition. + +Cross-prompt Stylized Face Recognition. The training and test sets for Cross-prompt Stylized Face Recognition are subsets of the corresponding training and test sets for Standard Stylized Face Recognition. The training set only includes images generated using 60 prompts from the Standard Stylized Face Recognition training set. It contains a total of 3,811,707 images. Additionally, the real face images corresponding to the IDs is also used for training. The test set consists of images generated using the last 20 prompts from the Standard Stylized Face Recognition test set, totaling 42,361 images. This setting ensures that the prompts and IDs used in the training and test sets are distinct. More details are shown in Appendix. + +Cross-method Stylized Face Recognition. In this setup, we aim to assess cross-method generalization. The training set is the same as that used in Standard Stylized Face Recognition. The test set consists of 1k independent IDs, with 149,827 images generated by Flashface, IP-Adapter, and Photomaker. In detail, Flashface contributed 70,912 images, IP-Adapter contributed 51,915 images, and Photomaker contributed 27,000 images. + +# 4.2. Evaluation Metrics + +In all experiments, we follow the standard evaluation metrics for face recognition. We evaluate the performance of the model on face verification. The gallery consists of all images in the test set, excluding the test images themselves. We report the results at multiple points, including TAR@FAR at 1e-2, TAR@FAR at 1e-3, and TAR@FAR at 1e-4. The metrics are defined as follows: + +- True Acceptance Rate (TAR) (complementary to False Non-Match Rate (FNMR)): Proportion of correctly identified valid identities: + +$$ +T A R = \frac {M}{N} +$$ + +where $M$ is the number of correctly identified identities, + +and $N$ is the total number of valid identity attempts. + +Note: ${FNMR} = 1 - {TAR}$ + +- False Acceptance Rate (FAR) (equivalent to False Match Rate (FMR)): Proportion of non-target identities incorrectly accepted: + +$$ +F A R = \frac {L}{K} +$$ + +where $L$ is the number of incorrect acceptances, and $K$ is the number of non-target identity attempts. + +Note: ${FMR} = {FAR}$ + +- TAR@FAR: Represents the True Acceptance Rate (TAR) at a specific False Acceptance Rate (FAR), used to evaluate system accuracy at a fixed security level. + +# 4.3. Implementation Details + +We maintain the same setup for all experiments. Our backbone networks are IResNet-50 and IResNet-101. We apply Cosface [56] framework and use an angular margin of 0.3 for Cosface. We set the batch size to 512 and train for a total of 26 epochs. The models are subsequently trained using Stochastic Gradient Descent (SGD), with a momentum coefficient of 0.9 and a weight decay parameter of $5 \times 10^{-4}$ . The initial learning rate is set at $1 \times 10^{-1}$ and is decreased by a factor of 8 at the midpoint of the epochs. + +# 4.4. Result + +# 4.4.1. Standard Stylized Face Recognition. + +We assess the model's performance trained on the large-scale MS1MV3 [13] and Glint360k [6] dataset. As presented in Tab. 4, the results indicate that despite the substantial size of the training set, there is a significant performance degradation compared to real face scenarios. Although the model trained on Glint360k performs worse than ours, it shows significant improvement over the model trained on MS1MV3, highlighting the positive impact of dataset size on generalization. + +The model trained on our dataset significantly outperforms the one trained on real face data in open-set face recognition scenarios, achieving a face verification TAR@FAR at 1e-4 of $90.6\%$ , as detailed in Tab. 4. This highlights the crucial role of our dataset in effectively recognizing stylized faces. + +Table 4. Results on Standard Stylized Face Recognition. Backbone is IResNet-50. Values in bold indicate the top results. + +
Training sets1e-21e-31e-4
MS1MV3 [13]85.168.049.7
Gint360k [6]95.087.074.7
Stylized-Face97.995.590.6
+ +As shown in Fig. 6, the model trained on our dataset significantly outperforms the one trained on real face data in + +open-set face recognition scenarios. It highlights the crucial role of our dataset in effectively recognizing stylized faces. + +![](images/a0fee25d5cbf56b5dbe7d2f868628be37cfe3e5fd28f308214db990acaeb9103.jpg) +Figure 6. the FMR-FNMR plot for All pairs verification. (lower FNMR is better). + +# 4.4.2. Cross-prompt stylized face recognition. + +Table 5 displays the performance across various prompts using the InstanceID [58], highlighting its generalization capabilities in stylized face recognition across different prompts. It also underscores our dataset's effectiveness in enhancing the cross-prompt generalization of stylized face recognition. + +Table 5. Results on Cross-prompt Stylized Face Recognition. Backbone is IResNet-50. Values in bold indicate the top results. + +
Training sets1e-21e-31e-4
MS1MV3 [13]85.767.749.3
Gint360k [6]94.786.573.9
Stylized-Face97.995.490.3
+ +# 4.4.3. Cross-method Stylized Face Recognition + +Cross-method experiments in Table 6 evaluate the ability of models to generalize feature extraction across various generation methods. The results indicate that, despite our method showing clear superiority over real-face models, current methods struggle with cross-method generalization across different methods, emphasizing the need for more in-depth research to enhance cross-method robustness. + +We further analyze cross-method generalization using the Flashface dataset. Prompts are divided into easy and hard categories based on the distribution shift, measured by the difference between Flashface and the InstantID method on the same data. Results show that performance on easy prompts is similar to within-distribution performance, while generalization challenges are more pronounced for hard + +Table 6. Results on Cross-method Stylized Face Recognition. Backbone is IResNet-50. All refers to Flashface, IP-Adapter, and Photomaker. Due to the limited number of images, the results exclude TAR@FAR at 1e-4. Values in bold indicate the top results. + +
Training setsTest Parts1e-21e-3
MS1MV3 [13]Flashface82.171.0
Gint360k [6]Flashface92.386.7
Stylized-FaceFlashface93.187.3
MS1MV3 [13]IP-Adapter68.262.2
Gint360k [6]IP-Adapter91.474.3
Stylized-FaceIP-Adapter93.678.3
MS1MV3 [13]Photomaker62.636.4
Gint360k [6]Photomaker86.562.2
Stylized-FacePhotomaker90.171.2
MS1MV3 [13]All59.031.8
Gint360k [6]All78.758.6
Stylized-FaceAll87.271.9
+ +prompts with significant distributional shifts. Notably, for these hard prompts, our model demonstrates particularly remarkable performance improvements compared to the real face recognition model. Results are shown in Tab. 7. + +Table 7. Analysis of Cross-method generalization. The evaluation metrics is TAR@FAR at 1e-2 on Flashface. Backbone is IResNet-50. + +
MethodsEasy PromptsHard PromptsAll
Glint360k96.984.992.3
Stylized-Face96.987.193.1
+ +# 4.5. Ablation Study + +Performance on Real Face Datasets. We evaluate the model's performance on real face datasets, with the results presented in Fig. 8. The comparison demonstrates that while our model achieves comparable performance, there exists a trade-off between real face recognition and stylized face recognition. + +Table 8. Results on Real Face Recognition. Backbone is IResNet-50. + +
Training setLFWCFP_FPAgeDBAvg.
CASIA99.595.294.996.5
MS1MV299.898.198.198.7
MS1MV399.898.598.398.9
Glint360k99.899.198.599.1
Stylized-Face99.798.497.998.7
+ +More Methods To comprehensively assess the perfor + +mance of stylized face recognition dataset, we compare advanced face recognition methods, including Cosface [56], Arcface [13], and CurricularFace [23]. These methods represent leading techniques in the field of face recognition. The experimental results are summarized in Tab. 9. The performance of the different training methods is similar. + +Table 9. Performance Comparison of Different Training Methods on Standard Stylized Face Recognition. + +
Training Methods1e-21e-31e-4
Cosface [56]97.995.590.6
CurricularFace [23]97.995.590.5
Arcface [13]97.995.489.9
+ +More Backbones We evaluate the performance of various backbones and Tab. 10 shows that scaling up the model size can lead to further improvements in performance. + +Table 10. Performance Comparison of Different Backbones on Standard Stylized Face Recognition. + +
Backbone1e-21e-31e-4
IResNet-5097.995.590.6
IResNet-10198.396.392.0
+ +# 5. Conclusion + +In addressing the challenge of recognizing stylized face images, we introduce the Stylized-Face dataset based on the CASIA-WebFace and Glint360k. Our dataset includes 4.6 million images spanning 62k IDs and incorporate a wide range of styles and identity-preserving methods. To ensure data quality (i.e., ID preservation) at this massive scale, we implement a semi-automated pipeline for large-scale data cleaning. Stylized-Face is the first dataset for stylized face recognition. Furthermore, we establish benchmarks to evaluate the robustness and generalization of stylized face recognition methods. Experimental results demonstrate that models trained on Stylized-Face achieve remarkable improvements in stylized face recognition. + +# 6. Limitation and Future Work + +This task continues to present challenges and opportunities for improvement. (1) Cross-method Stylized Face Recognition requires the adoption of additional designs to improve its generalization capabilities. (2) Explore approaches that can simultaneously achieve state-of-the-art performance in both real and stylized face recognition (3) Furthermore, considering the various long-tail distribution styles in real-world scenarios, which is also a direction for future research. + +# Acknowledgments + +This work is supported by the National Natural Science Foundation of China (62302167, U23A20343, 62472282, 72192821, 62476090), Young Elite Scientists Sponsorship Program by CAST (YESS20240780), Shanghai Sailing Program (23YF1410500), Chenguang Program of Shanghai Education Development Foundation, CCF-Tencent Rhino-Bird Young Faculty Open Research Fund (CCF-Tencent RAGR20230121, RAGR20240122), Shanghai Municipal Education Commission (23CGA34) and the Open Research Fund of Key Laboratory of Advanced Theory and Application in Statistics and Data Science-MOE, ECNU. + +# References + +[1] Anything-xl. https://civitai.com/models/9409/or-anything-xl.5 +[2] Yamermix. https://civitai.com/models/84040/sdxl-unstable-diffusers-yamermix. 5 +[3] Sdsl prompt styler, a custom node for comfyui. https://github.com/twri/sdsl_prompt_styler. 4 +[4] Reverse style transfer. https://devpost.com/software/reverse-style-transfer. 3 +[5] Xiang An, Xuhan Zhu, Yuan Gao, Yang Xiao, Yongle Zhao, Ziyong Feng, Lan Wu, Bin Qin, Ming Zhang, Debing Zhang, and Ying Fu. Partial fc: Training 10 million identities on a single machine. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2021. 3 +[6] Xiang An, Xuhan Zhu, Yuan Gao, Yang Xiao, Yongle Zhao, Ziyong Feng, Lan Wu, Bin Qin, Ming Zhang, Debing Zhang, et al. Partial fc: Training 10 million identities on a single machine. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1445-1449, 2021. 2, 7, 8 +[7] Xiang An, Jiangkang Deng, Jia Guo, Ziyong Feng, Xuhan Zhu, Yang Jing, and Liu Tongliang. Killing two birds with one stone: Efficient and robust training of face recognitionCNNs by partial fc. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 3 +[8] Gwangbin Bae, Martin de La Gorce, Tadas Baltrusaitis, Charlie Hewitt, Dong Chen, Julien Valentin, Roberto Cipolla, and Jingjing Shen. Digiface-1m: 1 million digital face images for face recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3526-3535, 2023. 3 +[9] Fadi Boutros, Marco Huber, Patrick Siebke, Tim Rieber, and Naser Damer. Sface: Privacy-friendly and accurate face recognition using synthetic data. In 2022 IEEE International Joint Conference on Biometrics (IJCB), pages 1-11. IEEE, 2022. +[10] Fadi Boutros, Marcel Klemt, Meiling Fang, Arjan Kuijper, and Naser Damer. Unsupervised face recognition using unlabeled synthetic data. In 2023 IEEE 17th International Conference on Automatic Face and Gesture Recognition (FG), pages 1-8. IEEE, 2023. 3 + +[11] PAN Chuanyu, YANG Guowei, MU Taijiang, and LAI Yu-Kun. Generating animatable 3d cartoon faces from single portraits. Virtual Reality & Intelligent Hardware, 6(4):292-307, 2024. 3 +[12] Jun Dan, Yang Liu, Haoyu Xie, Jiankang Deng, Haoran Xie, Xuansong Xie, and Baigui Sun. Transface: Calibrating transformer training for face recognition from a data-centric perspective. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20642-20653, 2023. 3 +[13] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 1, 3, 7, 8 +[14] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pages 1180-1189. PMLR, 2015. 3 +[15] Jingyu Gong, Fengqi Liu, Jiachen Xu, Min Wang, Xin Tan, Zhizhong Zhang, Ran Yi, Haichuan Song, Yuan Xie, and Lizhuang Ma. Optimization over disentangled encoding: Unsupervised cross-domain point cloud completion via occlusion factor manipulation. In European Conference on Computer Vision, pages 517-533. Springer, 2022. +[16] Jingyu Gong, Yujing Lou, Fengqi Liu, Zhiwei Zhang, Haoming Chen, Zhizhong Zhang, Xin Tan, Yuan Xie, and Lizhuang Ma. Scene point cloud understanding and reconstruction technologies in 3d space. Journal of Image and Graphics, 28(06):1741-1766, 2023. 3 +[17] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 87-102. Springer, 2016. 3 +[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 4 +[19] Chengyang Hu, Ke-Yue Zhang, Taiping Yao, Shice Liu, Shouhong Ding, Xin Tan, and Lizhuang Ma. Domainhallucinated updating for multi-domain face anti-spoofing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2193-2201, 2024. 3 +[20] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3 +[21] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008. 3 +[22] Lianghua Huang, Di Chen, Yu Liu, Yujun Shen, Deli Zhao, and Jingren Zhou. Composer: Creative and controllable image synthesis with composable conditions. arXiv preprint arXiv:2302.09778, 2023. 3 + +[23] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: adaptive curriculum learning loss for deep face recognition. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5901-5910, 2020. 1, 3, 8 +[24] Jing Huoa, Wenbin Lia, Yinghuan Shia, Yang Gaoa, and Hujun Yinb. Webcaricature: a benchmark for caricature face recognition. arXiv preprint arXiv:1703.03230, 2017. 3 +[25] Shen Li, Jianqing Xu, Xiaqing Xu, Pengcheng Shen, Shaoxin Li, and Bryan Hooi. Spherical confidence learning for face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15629-15637, 2021. 3 +[26] Xiaofan Li, Xin Tan, Zhuo Chen, Zhizhong Zhang, Ruixin Zhang, Rizen Guo, Guanna Jiang, Yulong Chen, Yanyun Qu, Lizhuang Ma, et al. One-for-more: Continual diffusion model for anomaly detection. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 4766-4775, 2025. 3 +[27] Yidi Li, Jun Xiao, Yiqun Wang, and Zhengda Lu. Depthgan: Gan-based depth generation from semantic layouts. Computational Visual Media, 10(3):505-522, 2024. 3 +[28] Yuxuan Li, Lingfeng Yang, and Xiang Li. Apf-gan: Exploring asymmetric pre-training and fine-tuning strategy for conditional generative adversarial network. Computational Visual Media, 10(1):187-192, 2024. 3 +[29] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. arXiv preprint arXiv:2312.04461, 2023. 1, 2, 3, 4, 5, 6 +[30] Zonglin Li, Xiaogian Lv, Wei Yu, Qinglin Liu, Jingbo Lin, and Shengping Zhang. Face shape transfer via semantic warping. Visual Intelligence, 2(1):26, 2024. 3 +[31] Xiao Lin, Yuge Huang, Jianqing Xu, Yuxi Mi, Shuigeng Zhou, and Shouhong Ding. Uface: Unleashing inherent model capabilities to enhance intra-class diversity in synthetic face recognition. arXiv preprint arXiv:2502.19803, 2025.3 +[32] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherical face: Deep hypersphere embedding for face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 212-220, 2017. 3 +[33] Wang Lu, Jindong Wang, Haoliang Li, Yiqiang Chen, and Xing Xie. Domain-invariant feature exploration for domain generalization. arXiv preprint arXiv:2207.12020, 2022. 3 +[34] Brianna Maze, Jocelyn Adams, James A Duncan, Nathan Kalka, Tim Miller, Charles Otto, Anil K Jain, W Tyler Niggel, Janet Anderson, Jordan Cheney, et al. Iarpa janus benchmark-c: Face dataset and protocol. In 2018 international conference on biometrics (ICB), pages 158-165. IEEE, 2018. 3 +[35] Yuxi Mi, Zhizhou Zhong, Yuge Huang, Jiazhen Ji, Jianqing Xu, Jun Wang, Shaoming Wang, Shouhong Ding, and Shuigeng Zhou. Privacy-preserving face recognition using trainable feature subtraction. In Proceedings of the + +IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 297-307, 2024. 3 +[36] Yuxi Mi, Zhizhou Zhong, Yuge Huang, Qiuyang Yuan, Xuan Zhao, Jianqing Xu, Shouhong Ding, Shaoming Wang, Rizen Guo, and Shuigeng Zhou. Data synthesis with diverse styles for face recognition via 3dmm-guided diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21203-21214, 2025. 3 +[37] Ashutosh Mishra, Shyam Nandan Rai, Anand Mishra, and CV Jawahar. Iit-cfw: A benchmark database of cartoon faces in the wild. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part I 14, pages 35-47. Springer, 2016. 3 +[38] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3 +[39] Fu-Zhao Ou, Xingyu Chen, Ruixin Zhang, Yuge Huang, Shaoxin Li, Jilin Li, Yong Li, Liujuan Cao, and Yuan-Gen Wang. Sdd-fiaq: Unsupervised face image quality assessment with similarity distribution distance. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7670-7679, 2021. 4 +[40] Zhengyuan Peng, Qijian Tian, Jianqing Xu, Yizhang Jin, Xuequan Lu, Xin Tan, Yuan Xie, and Lizhuang Ma. Generalized category discovery in semantic segmentation. arXiv preprint arXiv:2311.11525, 2023. 3 +[41] Zhengyuan Peng, Jinpeng Ma, Zhimin Sun, Ran Yi, Haichuan Song, Xin Tan, and Lizhuang Ma. Mos: Modeling object-scene associations in generalized category discovery. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 15118-15128, 2025. +[42] Zhengyuan Peng, Jianqing Xu, Shen Li, Jiazhen Ji, Yuge Huang, Jingyun Zhang, Jinmin Li, Shouhong Ding, Rizen Guo, Xin Tan, et al. Eyeseg: An uncertainty-aware eye segmentation framework for ar/vr. arXiv preprint arXiv:2507.09649, 2025.3 +[43] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 3, 4, 5 +[44] Haibo Qiu, Baosheng Yu, Dihong Gong, Zhifeng Li, Wei Liu, and Dacheng Tao. Synface: Face recognition with synthetic data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10880-10890, 2021. 3 +[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3 +[46] Qibing Ren, Chang Gao, Jing Shao, Junchi Yan, Xin Tan, Wai Lam, and Lizhuang Ma. Codeattack: Revealing safety + +generalization challenges of large language models via code completion. arXiv preprint arXiv:2403.07865, 2024. 3 +[47] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3 +[48] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 3 +[49] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731, 2019. 3 +[50] Yichen Shi, Yuhao Gao, Yingxin Lai, Hongyang Wang, Jun Feng, Lei He, Jun Wan, Changsheng Chen, Zitong Yu, and Xiaochun Cao. Shield: An evaluation benchmark for face spoofing and forgery detection with multimodal large language models. Visual Intelligence, 3(1):9, 2025. 3 +[51] Kaede Shiohara and Toshihiko Yamasaki. Face2diffusion for fast and editable face personalization. arXiv preprint arXiv:2403.05094, 2024. 3 +[52] Shuai Tan, Bin Ji, and Ye Pan. Style2talker: High-resolution talking head generation with emotion style and art style. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5079-5087, 2024. 3 +[53] Shuai Tan, Biao Gong, Yutong Feng, Kecheng Zheng, Dandan Zheng, Shuwei Shi, Yujun Shen, Jingdong Chen, and Ming Yang. Mimir: Improving video diffusion models for precise text understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. +[54] Shuai Tan, Biao Gong, Yujie Wei, Shiwei Zhang, Zhuoxin Liu, Dandan Zheng, Jingdong Chen, Yan Wang, Hao Ouyang, Kecheng Zheng, and Yujun Shen. Synmotion: Semantic-visual adaptation for motion customized video generation. arXiv preprint arXiv:2506.23690, 2025. 3 +[55] Ming Tao, Hao Tang, Songsong Wu, Nicu Sebe, Xiao-Yuan Jing, Fei Wu, and Bingkun Bao. Df-gan: Deep fusion generative adversarial networks for text-to-image synthesis. arXiv preprint arXiv:2008.05865, 2(6), 2020. 3 +[56] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5265-5274, 2018. 1, 3, 7, 8 +[57] Qian Wang, Cai Guo, Hong-Ning Dai, and Ping Li. Stroke-gan painter: Learning to paint artworks using stroke-style generative adversarial networks. Computational Visual Media, 9(4):787-806, 2023. 3 +[58] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 1, 2, 3, 4, 5, 6, 7 + +[59] Tan Wang, Linjie Li, Kevin Lin, Chung-Ching Lin, Zhengyuan Yang, Hanwang Zhang, Zicheng Liu, and Lijuan Wang. disco: Disentangled control for referring human dance generation in real world. arXiv e-prints, pages arXiv-2307, 2023. 3 +[60] Cameron Whitelam, Emma Taborsky, Austin Blanton, Brianna Maze, Jocelyn Adams, Tim Miller, Nathan Kalka, Anil K Jain, James A Duncan, Kristen Allen, et al. Iarpa janus benchmark-b face dataset. In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 90-98, 2017. 3 +[61] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multisubject image generation with localized attention. arXiv preprint arXiv:2305.10431, 2023. 3 +[62] Jianqing Xu, Shen Li, Jiaying Wu, Miao Xiong, Ailin Deng, Jiazhen Ji, Yuge Huang, Guodong Mu, Wenjie Feng, Shouhong Ding, et al. $\text{Id}^3$ : Identity-preserving-yet-diversified diffusion models for synthetic face recognition. Advances in Neural Information Processing Systems, 37: 77777-77798, 2024. 3 +[63] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Finegrained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 3 +[64] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 2, 3, 4, 5 +[65] Dong Yi, Zhen Lei, Shengcai Liao, and Stan Z Li. Learning face representation from scratch. arXiv preprint arXiv:1411.7923, 2014. 2, 3, 4 +[66] Ran Yi, Teng Hu, Mengfei Xia, Yizhe Tang, and Yong-Jin Liu. Feditnet++: Few-shot editing of latent semantics in gan spaces with correlated attribute disentanglement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3 +[67] Shilong Zhang, Lianghua Huang, Xi Chen, Yifei Zhang, Zhi-Fan Wu, Yutong Feng, Wei Wang, Yujun Shen, Yu Liu, and Ping Luo. Flashface: Human image personalization with high-fidelity identity preservation. arXiv preprint arXiv:2403.17008, 2024. 1, 2, 3, 4, 5, 6 +[68] Tianyue Zheng and Weihong Deng. Cross-pose lfw: A database for studying cross-posed face recognition in unconstrained environments. Beijing University of Posts and Telecommunications, Tech. Rep, 5(7):5, 2018. 3 +[69] Tianyue Zheng, Weihong Deng, and Jiani Hu. Cross-age lfw: A database for studying cross-age face recognition in unconstrained environments. arXiv preprint arXiv:1708.08197, 2017.3 +[70] Yi Zheng, Yifan Zhao, Mengyuan Ren, He Yan, Xiangju Lu, Junhui Liu, and Jia Li. Cartoon face recognition: A benchmark dataset. In Proceedings of the 28th ACM international conference on multimedia, pages 2264-2272, 2020. 3 +[71] Zhizhou Zhong, Yuxi Mi, Yuge Huang, Jianqing Xu, Guodong Mu, Shouhong Ding, Jingyun Zhang, Rizen Guo, + +Yunsheng Wu, and Shuigeng Zhou. Slerpface: Face template protection via spherical linear interpolation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10698-10706, 2025. 3 +[72] Zheng Zhu, Guan Huang, Jiankang Deng, Yun Ye, Junjie Huang, Xinze Chen, Jiagang Zhu, Tian Yang, Jiwen Lu, Dalong Du, and Jie Zhou. Webface260m: A benchmark unveiling the power of million-scale deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10492-10502, 2021. 3 \ No newline at end of file diff --git a/2025/Stylized-Face_ A Million-level Stylized Face Dataset for Face Recognition/images.zip b/2025/Stylized-Face_ A Million-level Stylized Face Dataset for Face Recognition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7134eba05db1e71dd7ae73479cf46b6bfbb942b8 --- /dev/null +++ b/2025/Stylized-Face_ A Million-level Stylized Face Dataset for Face Recognition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:99fbaa8da1949ce8af186ee73a706524580e2d4ac5391f9cf863e4ca548cb912 +size 660509 diff --git a/2025/Stylized-Face_ A Million-level Stylized Face Dataset for Face Recognition/layout.json b/2025/Stylized-Face_ A Million-level Stylized Face Dataset for Face Recognition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..35aa53407c70247a14d6ed5b942356d57a25c0f9 --- /dev/null +++ b/2025/Stylized-Face_ A Million-level Stylized Face Dataset for Face Recognition/layout.json @@ -0,0 +1,11544 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 80, + 103, + 530, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 103, + 530, + 121 + ], + "spans": [ + { + "bbox": [ + 80, + 103, + 530, + 121 + ], + "type": "text", + "content": "Stylized-Face: A Million-level Stylized Face Dataset for Face Recognition" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 149, + 142, + 468, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 142, + 468, + 157 + ], + "spans": [ + { + "bbox": [ + 149, + 142, + 468, + 157 + ], + "type": "text", + "content": "Zhengyuan Peng" + }, + { + "bbox": [ + 149, + 142, + 468, + 157 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 149, + 142, + 468, + 157 + ], + "type": "text", + "content": ", Jianqing Xu" + }, + { + "bbox": [ + 149, + 142, + 468, + 157 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 149, + 142, + 468, + 157 + ], + "type": "text", + "content": ", Yuge Huang" + }, + { + "bbox": [ + 149, + 142, + 468, + 157 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 149, + 142, + 468, + 157 + ], + "type": "text", + "content": ", Jinkun Hao" + }, + { + "bbox": [ + 149, + 142, + 468, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 142, + 468, + 157 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 149, + 157, + 468, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 157, + 468, + 171 + ], + "spans": [ + { + "bbox": [ + 149, + 157, + 468, + 171 + ], + "type": "text", + "content": "Shouhong Ding" + }, + { + "bbox": [ + 149, + 157, + 468, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 149, + 157, + 468, + 171 + ], + "type": "text", + "content": ", Zhizhong Zhang" + }, + { + "bbox": [ + 149, + 157, + 468, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 149, + 157, + 468, + 171 + ], + "type": "text", + "content": ", Xin Tan" + }, + { + "bbox": [ + 149, + 157, + 468, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 149, + 157, + 468, + 171 + ], + "type": "text", + "content": ", Lizhuang Ma" + }, + { + "bbox": [ + 149, + 157, + 468, + 171 + ], + "type": "inline_equation", + "content": "^{1,3\\dagger}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 99, + 171, + 511, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 171, + 511, + 186 + ], + "spans": [ + { + "bbox": [ + 99, + 171, + 511, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 99, + 171, + 511, + 186 + ], + "type": "text", + "content": "Shanghai Jiao Tong University, " + }, + { + "bbox": [ + 99, + 171, + 511, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 99, + 171, + 511, + 186 + ], + "type": "text", + "content": "Youtu Lab, Tencent, " + }, + { + "bbox": [ + 99, + 171, + 511, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 99, + 171, + 511, + 186 + ], + "type": "text", + "content": "East China Normal University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 240, + 297, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 240, + 297, + 563 + ], + "spans": [ + { + "bbox": [ + 54, + 240, + 297, + 563 + ], + "type": "text", + "content": "Stylized face recognition is the task of recognizing generated faces with the same ID across diverse stylistic domains (e.g., anime, painting, cyberpunk styles). This emerging field plays a vital role in the governance of generative image, serving the primary objective: Recognize the ID information of stylized faces to detect potential infringements of portrait rights. Despite its importance, progress in stylized face recognition has been hindered by the lack of large-scale, stylistically diverse datasets. To address this gap, we introduce the Stylized-Face dataset, which is the first dataset specifically designed for stylized face recognition. Stylized-Face dataset includes 4.6 million images across 62k IDs, specifically curated to enhance model performance in stylized face recognition tasks. To ensure data quality (i.e., ID preservation) at this massive scale, we implement a semi-automated pipeline for large-scale data cleaning. Based on the Stylized-Face dataset, we establish three benchmarks to evaluate the robustness and generalization of recognition models across various scenarios, including within-distribution performance, cross-prompt generalization, and cross-method generalization, which target key challenges in stylized face recognition. Experimental results demonstrate that models trained on Stylized-Face achieve remarkable improvements in both stylized face recognition performance (a " + }, + { + "bbox": [ + 54, + 240, + 297, + 563 + ], + "type": "inline_equation", + "content": "15.9\\%" + }, + { + "bbox": [ + 54, + 240, + 297, + 563 + ], + "type": "text", + "content": " improvement in TAR at " + }, + { + "bbox": [ + 54, + 240, + 297, + 563 + ], + "type": "inline_equation", + "content": "\\mathrm{FAR} = 1\\mathrm{e - }4" + }, + { + "bbox": [ + 54, + 240, + 297, + 563 + ], + "type": "text", + "content": ") and generalization (a " + }, + { + "bbox": [ + 54, + 240, + 297, + 563 + ], + "type": "inline_equation", + "content": "13.3\\%" + }, + { + "bbox": [ + 54, + 240, + 297, + 563 + ], + "type": "text", + "content": " improvement in TAR at " + }, + { + "bbox": [ + 54, + 240, + 297, + 563 + ], + "type": "inline_equation", + "content": "\\mathrm{FAR} = 1\\mathrm{e - }3" + }, + { + "bbox": [ + 54, + 240, + 297, + 563 + ], + "type": "text", + "content": " in cross-method generalization)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 589, + 135, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 589, + 135, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 589, + 135, + 601 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 609, + 296, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 296, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 296, + 682 + ], + "type": "text", + "content": "In recent years, the rapid development of personalized diffusion methods has been notable. Their low barrier to entry enables users to easily generate a wide array of images, but this same accessibility also facilitates the production of content that may infringe on personal portrait rights, raising significant privacy concerns. Traditionally, face recognition" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 212, + 552, + 373 + ], + "blocks": [ + { + "bbox": [ + 318, + 212, + 552, + 373 + ], + "lines": [ + { + "bbox": [ + 318, + 212, + 552, + 373 + ], + "spans": [ + { + "bbox": [ + 318, + 212, + 552, + 373 + ], + "type": "image", + "image_path": "465156445b391cb165cb770b5bad8de3b85776091d196f6ea7e5aff19a698da3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 381, + 555, + 426 + ], + "lines": [ + { + "bbox": [ + 313, + 381, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 555, + 426 + ], + "type": "text", + "content": "Figure 1. Illustration of the Stylized-Face dataset across diverse styles. The dataset features 80 distinct styles organized into four primary groups: Artistic Living, Classical Art Styles, Fantasy Styles, and ACG Styles (Animation, Comics, and Games)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 449, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 544 + ], + "type": "text", + "content": "models have served as invaluable tools for detecting potential portrait infringements within images. However, the advent of these new technologies presents a more complex issue. Personalized diffusion methods can generate images in various styles (e.g., anime, painting, cyberpunk styles) simply by altering prompts. This stylistic diversity poses a challenge for recognition models in recognizing faces with the same ID across a broad spectrum of stylized images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 545, + 556, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 676 + ], + "type": "text", + "content": "Previous methods [29, 67] addressing this challenge primarily rely on two models: CLIP-based models [45] and real face recognition models [13, 23, 56]. To systematically evaluate these models, we present a large-scale and diverse analysis across four different personalized diffusion models [29, 58, 64, 67] in Sec. 3.1. The results indicate that both exhibit significant limitations. Notably, CLIP-based models demonstrate considerable recognition biases. Although real face recognition models more closely mirror human perceptual judgments, their overall accuracy remains suboptimal. We illustrate failure cases in Fig. 2." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Due to the limitations of the current recognition model, filling in the gaps in the relevant training data is an effective approach. We introduce a novel stylized face dataset," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "text", + "content": "*Equal Contribution" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 67, + 703, + 145, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 145, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 145, + 712 + ], + "type": "text", + "content": "† Corresponding Author" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13053" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 102, + 85, + 154, + 133 + ], + "blocks": [ + { + "bbox": [ + 102, + 71, + 168, + 82 + ], + "lines": [ + { + "bbox": [ + 102, + 71, + 168, + 82 + ], + "spans": [ + { + "bbox": [ + 102, + 71, + 168, + 82 + ], + "type": "text", + "content": "Reference Image" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 102, + 85, + 154, + 133 + ], + "lines": [ + { + "bbox": [ + 102, + 85, + 154, + 133 + ], + "spans": [ + { + "bbox": [ + 102, + 85, + 154, + 133 + ], + "type": "image", + "image_path": "fcb2fca84b4e6a97360b8a0c5776a3d32d5f3db8b17e5b850f5eba89fc04de75.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 105, + 142, + 152, + 188 + ], + "blocks": [ + { + "bbox": [ + 105, + 142, + 152, + 188 + ], + "lines": [ + { + "bbox": [ + 105, + 142, + 152, + 188 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 152, + 188 + ], + "type": "image", + "image_path": "4453d9c85933a961fdfac719b27ca4d8d324424106acc644c3dfdb764a634c6b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 106, + 199, + 152, + 244 + ], + "blocks": [ + { + "bbox": [ + 106, + 199, + 152, + 244 + ], + "lines": [ + { + "bbox": [ + 106, + 199, + 152, + 244 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 152, + 244 + ], + "type": "image", + "image_path": "d8e5c4853dfbca0c0239a8baa07061819411b573585dafdca845bcc2526ffb28.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 105, + 256, + 153, + 301 + ], + "blocks": [ + { + "bbox": [ + 105, + 256, + 153, + 301 + ], + "lines": [ + { + "bbox": [ + 105, + 256, + 153, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 256, + 153, + 301 + ], + "type": "image", + "image_path": "ac3343c2399de544f214e3c65fc8bd335212ad6045f61d6e11d6232675d0a0a0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 316, + 201, + 324 + ], + "lines": [ + { + "bbox": [ + 170, + 316, + 201, + 324 + ], + "spans": [ + { + "bbox": [ + 170, + 316, + 201, + 324 + ], + "type": "text", + "content": "Line Art" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 164, + 87, + 209, + 132 + ], + "blocks": [ + { + "bbox": [ + 164, + 87, + 209, + 132 + ], + "lines": [ + { + "bbox": [ + 164, + 87, + 209, + 132 + ], + "spans": [ + { + "bbox": [ + 164, + 87, + 209, + 132 + ], + "type": "image", + "image_path": "2b2b12cf7055007de7eaf68b0c6f714c7b8510e50f3102e38096f3322ed277b2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 133, + 195, + 141 + ], + "lines": [ + { + "bbox": [ + 181, + 133, + 195, + 141 + ], + "spans": [ + { + "bbox": [ + 181, + 133, + 195, + 141 + ], + "type": "text", + "content": "0.31" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 213, + 87, + 257, + 132 + ], + "blocks": [ + { + "bbox": [ + 213, + 87, + 257, + 132 + ], + "lines": [ + { + "bbox": [ + 213, + 87, + 257, + 132 + ], + "spans": [ + { + "bbox": [ + 213, + 87, + 257, + 132 + ], + "type": "image", + "image_path": "1d02d7d4c7bed319b7cc7456a87e3aea1f401ab1ae98fadae8e697b55ae39e6e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 134, + 244, + 141 + ], + "lines": [ + { + "bbox": [ + 230, + 134, + 244, + 141 + ], + "spans": [ + { + "bbox": [ + 230, + 134, + 244, + 141 + ], + "type": "text", + "content": "0.26" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 261, + 87, + 306, + 132 + ], + "blocks": [ + { + "bbox": [ + 256, + 72, + 315, + 83 + ], + "lines": [ + { + "bbox": [ + 256, + 72, + 315, + 83 + ], + "spans": [ + { + "bbox": [ + 256, + 72, + 315, + 83 + ], + "type": "text", + "content": "Stylized Image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 261, + 87, + 306, + 132 + ], + "lines": [ + { + "bbox": [ + 261, + 87, + 306, + 132 + ], + "spans": [ + { + "bbox": [ + 261, + 87, + 306, + 132 + ], + "type": "image", + "image_path": "a93b25614bfeb52a18410cbf308eeba75668ff4628b7a88c16a55da07ccf5a38.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 87, + 354, + 132 + ], + "blocks": [ + { + "bbox": [ + 309, + 87, + 354, + 132 + ], + "lines": [ + { + "bbox": [ + 309, + 87, + 354, + 132 + ], + "spans": [ + { + "bbox": [ + 309, + 87, + 354, + 132 + ], + "type": "image", + "image_path": "b794ddef1b526ecab4a5abcfe31d1d8e6ac7cb4498a49aeb489c5173a6ccb1e5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 134, + 339, + 141 + ], + "lines": [ + { + "bbox": [ + 325, + 134, + 339, + 141 + ], + "spans": [ + { + "bbox": [ + 325, + 134, + 339, + 141 + ], + "type": "text", + "content": "0.42" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 358, + 87, + 402, + 132 + ], + "blocks": [ + { + "bbox": [ + 358, + 87, + 402, + 132 + ], + "lines": [ + { + "bbox": [ + 358, + 87, + 402, + 132 + ], + "spans": [ + { + "bbox": [ + 358, + 87, + 402, + 132 + ], + "type": "image", + "image_path": "dc939fe1a3dae834e5db12bdb19a8897fa5faeeabb3799c324f02043951dacc4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 134, + 388, + 141 + ], + "lines": [ + { + "bbox": [ + 375, + 134, + 388, + 141 + ], + "spans": [ + { + "bbox": [ + 375, + 134, + 388, + 141 + ], + "type": "text", + "content": "0.19" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 164, + 143, + 209, + 187 + ], + "blocks": [ + { + "bbox": [ + 164, + 143, + 209, + 187 + ], + "lines": [ + { + "bbox": [ + 164, + 143, + 209, + 187 + ], + "spans": [ + { + "bbox": [ + 164, + 143, + 209, + 187 + ], + "type": "image", + "image_path": "a1339114f21d345dc6b2c9bd1085c4908b000d1a546a611a15e9d3caa1c9669c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 189, + 195, + 196 + ], + "lines": [ + { + "bbox": [ + 181, + 189, + 195, + 196 + ], + "spans": [ + { + "bbox": [ + 181, + 189, + 195, + 196 + ], + "type": "text", + "content": "0.43" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 213, + 143, + 258, + 187 + ], + "blocks": [ + { + "bbox": [ + 213, + 143, + 258, + 187 + ], + "lines": [ + { + "bbox": [ + 213, + 143, + 258, + 187 + ], + "spans": [ + { + "bbox": [ + 213, + 143, + 258, + 187 + ], + "type": "image", + "image_path": "4c695cdc7a8703d1a9aee3dfc7370b7e8f118b9654ef04856c2b3a9399091b2f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 189, + 243, + 197 + ], + "lines": [ + { + "bbox": [ + 230, + 189, + 243, + 197 + ], + "spans": [ + { + "bbox": [ + 230, + 189, + 243, + 197 + ], + "type": "text", + "content": "0.21" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 261, + 143, + 306, + 187 + ], + "blocks": [ + { + "bbox": [ + 277, + 134, + 291, + 141 + ], + "lines": [ + { + "bbox": [ + 277, + 134, + 291, + 141 + ], + "spans": [ + { + "bbox": [ + 277, + 134, + 291, + 141 + ], + "type": "text", + "content": "0.36" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 261, + 143, + 306, + 187 + ], + "lines": [ + { + "bbox": [ + 261, + 143, + 306, + 187 + ], + "spans": [ + { + "bbox": [ + 261, + 143, + 306, + 187 + ], + "type": "image", + "image_path": "744f6d7afe503442fe284c32859dadf023610e3b41e05bf50c9938f712b51192.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 309, + 143, + 354, + 187 + ], + "blocks": [ + { + "bbox": [ + 309, + 143, + 354, + 187 + ], + "lines": [ + { + "bbox": [ + 309, + 143, + 354, + 187 + ], + "spans": [ + { + "bbox": [ + 309, + 143, + 354, + 187 + ], + "type": "image", + "image_path": "729f60c52b1bbde85f3578452fe0a49e17c5be8ff72bd182f5bbadacc3d572e4.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 327, + 189, + 340, + 197 + ], + "lines": [ + { + "bbox": [ + 327, + 189, + 340, + 197 + ], + "spans": [ + { + "bbox": [ + 327, + 189, + 340, + 197 + ], + "type": "text", + "content": "0.25" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 358, + 143, + 402, + 187 + ], + "blocks": [ + { + "bbox": [ + 358, + 143, + 402, + 187 + ], + "lines": [ + { + "bbox": [ + 358, + 143, + 402, + 187 + ], + "spans": [ + { + "bbox": [ + 358, + 143, + 402, + 187 + ], + "type": "image", + "image_path": "79f4f682b09261715a1a4203f8df8968cfe341b698f0d4821aa00f2b18e079df.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 189, + 388, + 197 + ], + "lines": [ + { + "bbox": [ + 374, + 189, + 388, + 197 + ], + "spans": [ + { + "bbox": [ + 374, + 189, + 388, + 197 + ], + "type": "text", + "content": "0.31" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 164, + 199, + 209, + 244 + ], + "blocks": [ + { + "bbox": [ + 164, + 199, + 209, + 244 + ], + "lines": [ + { + "bbox": [ + 164, + 199, + 209, + 244 + ], + "spans": [ + { + "bbox": [ + 164, + 199, + 209, + 244 + ], + "type": "image", + "image_path": "dd7c1457891b2b166d2d5862bd876fdbe30c32697d6c25217aac5cfecb674bb7.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 246, + 195, + 253 + ], + "lines": [ + { + "bbox": [ + 181, + 246, + 195, + 253 + ], + "spans": [ + { + "bbox": [ + 181, + 246, + 195, + 253 + ], + "type": "text", + "content": "0.36" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 213, + 199, + 258, + 244 + ], + "blocks": [ + { + "bbox": [ + 213, + 199, + 258, + 244 + ], + "lines": [ + { + "bbox": [ + 213, + 199, + 258, + 244 + ], + "spans": [ + { + "bbox": [ + 213, + 199, + 258, + 244 + ], + "type": "image", + "image_path": "04958b94f7574a545b2270975a7a54ef961e6c3271e5af00d78a298b9caefab7.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 246, + 243, + 253 + ], + "lines": [ + { + "bbox": [ + 229, + 246, + 243, + 253 + ], + "spans": [ + { + "bbox": [ + 229, + 246, + 243, + 253 + ], + "type": "text", + "content": "0.46" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 261, + 199, + 306, + 244 + ], + "blocks": [ + { + "bbox": [ + 276, + 189, + 290, + 197 + ], + "lines": [ + { + "bbox": [ + 276, + 189, + 290, + 197 + ], + "spans": [ + { + "bbox": [ + 276, + 189, + 290, + 197 + ], + "type": "text", + "content": "0.44" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 261, + 199, + 306, + 244 + ], + "lines": [ + { + "bbox": [ + 261, + 199, + 306, + 244 + ], + "spans": [ + { + "bbox": [ + 261, + 199, + 306, + 244 + ], + "type": "image", + "image_path": "d1b9378d7c76e1afd600b34f173ec0f5bf85a59df7d325718f54d01c65ae00bc.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 246, + 290, + 253 + ], + "lines": [ + { + "bbox": [ + 277, + 246, + 290, + 253 + ], + "spans": [ + { + "bbox": [ + 277, + 246, + 290, + 253 + ], + "type": "text", + "content": "0.45" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 309, + 199, + 354, + 244 + ], + "blocks": [ + { + "bbox": [ + 309, + 199, + 354, + 244 + ], + "lines": [ + { + "bbox": [ + 309, + 199, + 354, + 244 + ], + "spans": [ + { + "bbox": [ + 309, + 199, + 354, + 244 + ], + "type": "image", + "image_path": "50d4769982fb7009ef4f894f81186ab9a3fa8a31b38d26752107105f021c206b.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 246, + 341, + 253 + ], + "lines": [ + { + "bbox": [ + 328, + 246, + 341, + 253 + ], + "spans": [ + { + "bbox": [ + 328, + 246, + 341, + 253 + ], + "type": "text", + "content": "0.32" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 358, + 199, + 402, + 244 + ], + "blocks": [ + { + "bbox": [ + 358, + 199, + 402, + 244 + ], + "lines": [ + { + "bbox": [ + 358, + 199, + 402, + 244 + ], + "spans": [ + { + "bbox": [ + 358, + 199, + 402, + 244 + ], + "type": "image", + "image_path": "340c8a6d813370222e3b78aca8e90ee03ac839abbd9324e475ae6ff1b9483528.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 246, + 388, + 253 + ], + "lines": [ + { + "bbox": [ + 374, + 246, + 388, + 253 + ], + "spans": [ + { + "bbox": [ + 374, + 246, + 388, + 253 + ], + "type": "text", + "content": "0.36" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 164, + 256, + 209, + 300 + ], + "blocks": [ + { + "bbox": [ + 164, + 256, + 209, + 300 + ], + "lines": [ + { + "bbox": [ + 164, + 256, + 209, + 300 + ], + "spans": [ + { + "bbox": [ + 164, + 256, + 209, + 300 + ], + "type": "image", + "image_path": "6d48a6eaf8a4800e3d36eef9f553fe8390a69f08073e42a697720156c4211cfb.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 303, + 195, + 310 + ], + "lines": [ + { + "bbox": [ + 181, + 303, + 195, + 310 + ], + "spans": [ + { + "bbox": [ + 181, + 303, + 195, + 310 + ], + "type": "text", + "content": "0.42" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 213, + 256, + 258, + 300 + ], + "blocks": [ + { + "bbox": [ + 213, + 256, + 258, + 300 + ], + "lines": [ + { + "bbox": [ + 213, + 256, + 258, + 300 + ], + "spans": [ + { + "bbox": [ + 213, + 256, + 258, + 300 + ], + "type": "image", + "image_path": "958d861fcc247708faf9b3250c8d593be9b053119453849ea10e1b3c0cde54e7.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 303, + 244, + 310 + ], + "lines": [ + { + "bbox": [ + 230, + 303, + 244, + 310 + ], + "spans": [ + { + "bbox": [ + 230, + 303, + 244, + 310 + ], + "type": "text", + "content": "0.16" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 219, + 316, + 257, + 325 + ], + "lines": [ + { + "bbox": [ + 219, + 316, + 257, + 325 + ], + "spans": [ + { + "bbox": [ + 219, + 316, + 257, + 325 + ], + "type": "text", + "content": "Neonpunk" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 261, + 256, + 306, + 300 + ], + "blocks": [ + { + "bbox": [ + 261, + 256, + 306, + 300 + ], + "lines": [ + { + "bbox": [ + 261, + 256, + 306, + 300 + ], + "spans": [ + { + "bbox": [ + 261, + 256, + 306, + 300 + ], + "type": "image", + "image_path": "92eacb424526efb71e274f924f20929e69ff8f8a3c39a79b5b6ee30ddcf74119.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 303, + 291, + 310 + ], + "lines": [ + { + "bbox": [ + 277, + 303, + 291, + 310 + ], + "spans": [ + { + "bbox": [ + 277, + 303, + 291, + 310 + ], + "type": "text", + "content": "0.32" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 270, + 316, + 299, + 325 + ], + "lines": [ + { + "bbox": [ + 270, + 316, + 299, + 325 + ], + "spans": [ + { + "bbox": [ + 270, + 316, + 299, + 325 + ], + "type": "text", + "content": "Pop Art" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 336, + 555, + 392 + ], + "lines": [ + { + "bbox": [ + 55, + 336, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 555, + 392 + ], + "type": "text", + "content": "Figure 2. Illustration of the Stylized-Face dataset and Comparison with Real Images. We randomly select a frontal face for each identity from the CASIA-WebFace [65] as a reference image. These reference images are compared with stylized images from our dataset and other images of the same identity from CASIA-WebFace [65], analyzing their feature similarity using a face recognition model trained on real faces. The results, recorded beneath each image, show that while the model performs well on real face images, it struggles to accurately identify the stylized images." + } + ] + } + ], + "index": 68, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 309, + 256, + 354, + 300 + ], + "blocks": [ + { + "bbox": [ + 309, + 256, + 354, + 300 + ], + "lines": [ + { + "bbox": [ + 309, + 256, + 354, + 300 + ], + "spans": [ + { + "bbox": [ + 309, + 256, + 354, + 300 + ], + "type": "image", + "image_path": "4ab50453c2e7cc0c7e513b581f2986d5821d58b013575f2a96d9cbb63ab59d7f.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 303, + 341, + 310 + ], + "lines": [ + { + "bbox": [ + 328, + 303, + 341, + 310 + ], + "spans": [ + { + "bbox": [ + 328, + 303, + 341, + 310 + ], + "type": "text", + "content": "0.23" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 314, + 316, + 350, + 324 + ], + "lines": [ + { + "bbox": [ + 314, + 316, + 350, + 324 + ], + "spans": [ + { + "bbox": [ + 314, + 316, + 350, + 324 + ], + "type": "text", + "content": "Film Noir" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 358, + 256, + 402, + 300 + ], + "blocks": [ + { + "bbox": [ + 358, + 256, + 402, + 300 + ], + "lines": [ + { + "bbox": [ + 358, + 256, + 402, + 300 + ], + "spans": [ + { + "bbox": [ + 358, + 256, + 402, + 300 + ], + "type": "image", + "image_path": "a08ea658c1ffc893e99e2d6207a76e0bc68457c45eccccaf2d11365fbb5c2d51.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 303, + 388, + 310 + ], + "lines": [ + { + "bbox": [ + 374, + 303, + 388, + 310 + ], + "spans": [ + { + "bbox": [ + 374, + 303, + 388, + 310 + ], + "type": "text", + "content": "0.24" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 359, + 316, + 406, + 324 + ], + "lines": [ + { + "bbox": [ + 359, + 316, + 406, + 324 + ], + "spans": [ + { + "bbox": [ + 359, + 316, + 406, + 324 + ], + "type": "text", + "content": "Retro Arcade" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 414, + 87, + 459, + 132 + ], + "blocks": [ + { + "bbox": [ + 440, + 72, + 485, + 83 + ], + "lines": [ + { + "bbox": [ + 440, + 72, + 485, + 83 + ], + "spans": [ + { + "bbox": [ + 440, + 72, + 485, + 83 + ], + "type": "text", + "content": "Real Image" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 414, + 87, + 459, + 132 + ], + "lines": [ + { + "bbox": [ + 414, + 87, + 459, + 132 + ], + "spans": [ + { + "bbox": [ + 414, + 87, + 459, + 132 + ], + "type": "image", + "image_path": "f315b5db148131e68ceeb1f8c38c70d3299a22feb7c2f6bd334b6e16def57a47.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 134, + 443, + 141 + ], + "lines": [ + { + "bbox": [ + 429, + 134, + 443, + 141 + ], + "spans": [ + { + "bbox": [ + 429, + 134, + 443, + 141 + ], + "type": "text", + "content": "0.78" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 462, + 87, + 506, + 132 + ], + "blocks": [ + { + "bbox": [ + 462, + 87, + 506, + 132 + ], + "lines": [ + { + "bbox": [ + 462, + 87, + 506, + 132 + ], + "spans": [ + { + "bbox": [ + 462, + 87, + 506, + 132 + ], + "type": "image", + "image_path": "ceb6aa35fd87d80f5bb4d00475d258c1ec6c57a55a924e2934af5036a0fba825.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 477, + 134, + 490, + 141 + ], + "lines": [ + { + "bbox": [ + 477, + 134, + 490, + 141 + ], + "spans": [ + { + "bbox": [ + 477, + 134, + 490, + 141 + ], + "type": "text", + "content": "0.81" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_caption" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 414, + 143, + 459, + 187 + ], + "blocks": [ + { + "bbox": [ + 414, + 143, + 459, + 187 + ], + "lines": [ + { + "bbox": [ + 414, + 143, + 459, + 187 + ], + "spans": [ + { + "bbox": [ + 414, + 143, + 459, + 187 + ], + "type": "image", + "image_path": "a2afffa86c4f145c513e05321e622274cd1624ef511871c7029305a7f1865a4d.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 189, + 444, + 197 + ], + "lines": [ + { + "bbox": [ + 429, + 189, + 444, + 197 + ], + "spans": [ + { + "bbox": [ + 429, + 189, + 444, + 197 + ], + "type": "text", + "content": "0.76" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_caption" + } + ], + "index": 56 + }, + { + "type": "image", + "bbox": [ + 462, + 143, + 506, + 187 + ], + "blocks": [ + { + "bbox": [ + 462, + 143, + 506, + 187 + ], + "lines": [ + { + "bbox": [ + 462, + 143, + 506, + 187 + ], + "spans": [ + { + "bbox": [ + 462, + 143, + 506, + 187 + ], + "type": "image", + "image_path": "7fb41f536dc415a8f9da97e7d21f40fe179a240789ad820c21ca873ecd2f1594.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 477, + 189, + 490, + 197 + ], + "lines": [ + { + "bbox": [ + 477, + 189, + 490, + 197 + ], + "spans": [ + { + "bbox": [ + 477, + 189, + 490, + 197 + ], + "type": "text", + "content": "0.84" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_caption" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 414, + 199, + 459, + 244 + ], + "blocks": [ + { + "bbox": [ + 414, + 199, + 459, + 244 + ], + "lines": [ + { + "bbox": [ + 414, + 199, + 459, + 244 + ], + "spans": [ + { + "bbox": [ + 414, + 199, + 459, + 244 + ], + "type": "image", + "image_path": "7a01dcebbb8a210b671380025b1ef5e7317dc801239316a98809d751d2c3eb3c.jpg" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 246, + 444, + 253 + ], + "lines": [ + { + "bbox": [ + 430, + 246, + 444, + 253 + ], + "spans": [ + { + "bbox": [ + 430, + 246, + 444, + 253 + ], + "type": "text", + "content": "0.71" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_caption" + } + ], + "index": 60 + }, + { + "type": "image", + "bbox": [ + 462, + 199, + 506, + 244 + ], + "blocks": [ + { + "bbox": [ + 462, + 199, + 506, + 244 + ], + "lines": [ + { + "bbox": [ + 462, + 199, + 506, + 244 + ], + "spans": [ + { + "bbox": [ + 462, + 199, + 506, + 244 + ], + "type": "image", + "image_path": "36c635950a9ca2bc9b187b43b7d8126c064813d99eedcb27e38129f60d1b75cc.jpg" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 477, + 246, + 491, + 252 + ], + "lines": [ + { + "bbox": [ + 477, + 246, + 491, + 252 + ], + "spans": [ + { + "bbox": [ + 477, + 246, + 491, + 252 + ], + "type": "text", + "content": "0.79" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_caption" + } + ], + "index": 62 + }, + { + "type": "image", + "bbox": [ + 414, + 256, + 459, + 300 + ], + "blocks": [ + { + "bbox": [ + 414, + 256, + 459, + 300 + ], + "lines": [ + { + "bbox": [ + 414, + 256, + 459, + 300 + ], + "spans": [ + { + "bbox": [ + 414, + 256, + 459, + 300 + ], + "type": "image", + "image_path": "e59a0ef3095de3fd786f64d32453f3430d5a249612d2de7bc8099378156ba19b.jpg" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 432, + 303, + 445, + 310 + ], + "lines": [ + { + "bbox": [ + 432, + 303, + 445, + 310 + ], + "spans": [ + { + "bbox": [ + 432, + 303, + 445, + 310 + ], + "type": "text", + "content": "0.79" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_caption" + } + ], + "index": 64 + }, + { + "type": "image", + "bbox": [ + 462, + 256, + 506, + 300 + ], + "blocks": [ + { + "bbox": [ + 462, + 256, + 506, + 300 + ], + "lines": [ + { + "bbox": [ + 462, + 256, + 506, + 300 + ], + "spans": [ + { + "bbox": [ + 462, + 256, + 506, + 300 + ], + "type": "image", + "image_path": "0798fff24f9169c0322c13bc888519ef0531c653855e699d4961ae02f26378a2.jpg" + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 478, + 303, + 492, + 310 + ], + "lines": [ + { + "bbox": [ + 478, + 303, + 492, + 310 + ], + "spans": [ + { + "bbox": [ + 478, + 303, + 492, + 310 + ], + "type": "text", + "content": "0.87" + } + ] + } + ], + "index": 67, + "angle": 0, + "type": "image_caption" + } + ], + "index": 66 + }, + { + "bbox": [ + 55, + 412, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 296, + 616 + ], + "type": "text", + "content": "Stylized-Face, generated from the CASIA-WebFace [65] and Glint360k [6] datasets, as illustrated in Fig. 1 and Fig. 2. Stylized-Face is specifically designed to enhance model performance and generalization on stylized face recognition, particularly those that significantly deviate from real faces. It includes 4.6 million images across 62k IDs and 80 distinct styles. We employ InstantID[58] (the best-performing method in our evaluation) as the primary generative method, supplemented with three additional methods (PhotoMaker[29], Flashface[67], IPAdapter[64]) for dataset diversity. Owing to the inherent instability of generative model outputs, some images may feature faces that do not match the assigned IDs or might not contain faces at all. To maintain data quality at this scale, we use a semi-automatic process for data cleaning, which combines manual and model verification. To our knowledge, Stylized-Face is the first dataset dedicated to stylized face recognition." + } + ] + } + ], + "index": 69 + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": "To further refine the setup for stylized face recognition, we construct three benchmarks: 1) Standard Stylized Face Recognition. Models are trained and evaluated on multiple styles from the same distribution, assessing their ability to recognize stylized faces. 2) Cross-prompt stylized face recognition. Different prompts generate images with distribution differences. We utilize varied prompts for both training and evaluation to assess the generalization perfor" + } + ] + } + ], + "index": 70 + }, + { + "bbox": [ + 313, + 413, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 555, + 472 + ], + "type": "text", + "content": "mance. 3) Cross-method stylized face recognition. Apart from the prompt, different generation methods yield varying distributions. We evaluate the generalization capabilities across the generation models by training and testing them on data produced through various techniques." + } + ] + } + ], + "index": 71 + }, + { + "bbox": [ + 313, + 473, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 555, + 581 + ], + "type": "text", + "content": "We evaluate the baseline models using our benchmark. Experimental results demonstrate that models trained on Stylized-Face significantly enhance stylized face recognition performance, achieving a " + }, + { + "bbox": [ + 313, + 473, + 555, + 581 + ], + "type": "inline_equation", + "content": "15.9\\%" + }, + { + "bbox": [ + 313, + 473, + 555, + 581 + ], + "type": "text", + "content": " increase in TAR at " + }, + { + "bbox": [ + 313, + 473, + 555, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{FAR} = 1\\mathrm{e - }4" + }, + { + "bbox": [ + 313, + 473, + 555, + 581 + ], + "type": "text", + "content": ". Additionally, they improve cross-prompt generalization by " + }, + { + "bbox": [ + 313, + 473, + 555, + 581 + ], + "type": "inline_equation", + "content": "12.2\\%" + }, + { + "bbox": [ + 313, + 473, + 555, + 581 + ], + "type": "text", + "content": " in TAR at " + }, + { + "bbox": [ + 313, + 473, + 555, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{FAR} = 1\\mathrm{e - }4" + }, + { + "bbox": [ + 313, + 473, + 555, + 581 + ], + "type": "text", + "content": " and cross-method generalization by " + }, + { + "bbox": [ + 313, + 473, + 555, + 581 + ], + "type": "inline_equation", + "content": "13.3\\%" + }, + { + "bbox": [ + 313, + 473, + 555, + 581 + ], + "type": "text", + "content": " in TAR at " + }, + { + "bbox": [ + 313, + 473, + 555, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{FAR} = 1\\mathrm{e - }3" + }, + { + "bbox": [ + 313, + 473, + 555, + 581 + ], + "type": "text", + "content": ". These improvements underscore the value of our dataset for advancing stylized face recognition." + } + ] + } + ], + "index": 72 + }, + { + "bbox": [ + 325, + 581, + 553, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 581, + 553, + 593 + ], + "spans": [ + { + "bbox": [ + 325, + 581, + 553, + 593 + ], + "type": "text", + "content": "To sum up, our major contributions are listed as follows:" + } + ] + } + ], + "index": 73 + }, + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 77, + "blocks": [ + { + "bbox": [ + 313, + 594, + 553, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 641 + ], + "type": "text", + "content": "- We evaluate previous models for stylized face recognition and mainstream personalized diffusion methods for the ability to preserve ID information. Our findings reveal significant shortcomings of previous recognition models." + } + ] + } + ], + "index": 74 + }, + { + "bbox": [ + 313, + 642, + 553, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 700 + ], + "type": "text", + "content": "- We introduce a stylized face dataset, Stylized-Face. Stylized-Face dataset includes 4.6 million images across 62k IDs, created by advanced generative techniques and semi-automatic process for data cleaning. Notably, it is the first dataset for stylized face recognition." + } + ] + } + ], + "index": 75 + }, + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "text", + "content": "- We introduce three novel benchmarks designed to eval" + } + ] + } + ], + "index": 76 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13054" + } + ] + } + ], + "index": 78 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "uate the performance of stylized face recognition models. We conduct a comprehensive evaluation of existing face recognition methods and our method using the Stylized-Face dataset." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 142, + 145 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 142, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 142, + 145 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 153, + 159, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 153, + 159, + 166 + ], + "spans": [ + { + "bbox": [ + 55, + 153, + 159, + 166 + ], + "type": "text", + "content": "2.1. Face Recognition" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 171, + 296, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 296, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 296, + 338 + ], + "type": "text", + "content": "Face recognition technology [5, 7, 12, 13, 23, 25, 31, 32, 35, 36, 50, 56, 62, 71] extract and analyze condensed face features, commonly referred to as facial templates. It employs advanced loss functions based on angular margins to enhance recognition accuracy. Angular margin-based approach [13, 23, 56] aims to maximize inter-identity discrepancies and minimize intra-identity discrepancies (similarities across various images of the same person). With the introduction of the Stylized-Face dataset, there is an increased demand for discriminative capabilities in recognition models. This demands models that can extract more crucial and generalized features to maintain identity consistency across diverse styles, a challenge inherent to Domain Generalization (DG) [14-16, 19, 33, 40-42, 46, 49]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 347, + 201, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 201, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 201, + 360 + ], + "type": "text", + "content": "2.2. Face Recognition Datasets" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 365, + 295, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 484 + ], + "type": "text", + "content": "Real Face Dataset: Existing real face recognition models predominantly rely on datasets like Webface42M[72], CASIA-WebFace [65], MS1Mv2 [17], MS1Mv3 [13], LFW [21], CPLFW[68], CALFW [69], IJB-B [60] and IJB-C [34]. These datasets comprise real-world face images and are designed to train and evaluate models to improve accuracy in real face applications. Stylized-Face dataset differs from the above as it explores diverse styles of dataset specific to generated face images, which significantly deviates from real face." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 486, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 640 + ], + "type": "text", + "content": "Synthetic Face Dataset: Earlier Synthetic datasets [8-10, 44] aim to enhance model performance on real face recognition and serve as substitutes for real face images. There are the main differences: (1) Authenticity of identity. We generate corresponding stylized images of real individuals; therefore, the IDs in our dataset are real, whereas existing datasets use virtual IDs. (2) Style. Our dataset includes various styles for the same ID, while existing datasets primarily focus on generating images in realistic style. (3) Objective. Our goal is designed for stylized face recognition, whereas existing datasets aims to enhance model performance on real face recognition and serve as substitutes for real face images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "Some other works[4] attempt to generate stylized faces, but they only perform style transfer on thousands of images without checking the ID information in the stylized images, lacking the fundamental reliable ID information essential for face recognition tasks. Therefore, these works are not suitable for stylized face recognition tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "content": "Cartoon Face Dataset: Some datasets[24, 37, 70] aims to explore different face style, such as cartoon face images. These datasets usually feature a single style and lack corresponding real human faces for comparison. Moreover, they are often sourced from animated films, which limits the data volume. In contrast, Stylized-Face dataset generates a large volume of face images across a wide range of styles, each paired with the ID of their corresponding real face." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 178, + 459, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 459, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 459, + 190 + ], + "type": "text", + "content": "2.3. Text-to-Image Generation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 195, + 555, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 195, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 195, + 555, + 411 + ], + "type": "text", + "content": "Text-to-image diffusion models [22, 26, 30, 38, 43, 47, 52-54, 59, 66] are at the forefront of image generation, renowned for their ability to transform text prompts into high-quality images with diverse styles. These models utilize a pre-trained language encoder to convert text into a latent representation, guiding the diffusion process to precisely adjust pixels. For instance, GLIDE [38] features a cascaded text-guided diffusion architecture for image generation. DisCo [59] introduces clip[45] encoder to guide the image generation. Stable Diffusion [47] optimizes the process by performing diffusion in a compressed latent space, which lowers computational costs. Further evolving this technology, Stable Diffusion XL [43] employs a larger network and additional text encoder to improve image fidelity and textual influence. Additionally, there are some GAN-based methods[11, 27, 28, 55, 57, 63], but they usually have weak generalization capabilities and are computationally expensive to train." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 420, + 480, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 480, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 480, + 431 + ], + "type": "text", + "content": "2.4. Personalized diffusion models" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "text", + "content": "In image generation, ID-preserving techniques prioritize key facial attributes to ensure identity remains consistent. These methods vary in their need for fine-tuning during testing. For instance, techniques such as DreamBooth [48] and LoRA [20] require extensive fine-tuning for each character, making them impractical for large-scale ID image generation. Recent developments aim to eliminate the need for fine-tuning altogether. Two fundamental methods include directly embedding the image features into the text embedding and using a cross-attention module to integrate the image encoding with the text encoding. The former includes methods such as Photomaker [29], FastComposer [61], and Face2Diffusion [51]. The latter comprises methods such as IP-Adapter-FaceID [64], InstantID [58], and Flashface [67]. In terms of specific designs, PhotoMaker [29] enhance identity preservation by modifying image encoders. IP-Adapter-FaceID [64] specifically utilizes face ID embeddings to ensure consistent identity retention. Moreover, InstantID [58] replaces the standard clip[45] encoder with face features and employs an additional IdentityNet, supervised by key facial landmarks, to guide face generation. FlashFace [67] encodes IDs into a series of feature maps to preserve finer face details. Our dataset utilizes these ID-preserving tech" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13055" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 91, + 545, + 277 + ], + "blocks": [ + { + "bbox": [ + 256, + 71, + 354, + 81 + ], + "lines": [ + { + "bbox": [ + 256, + 71, + 354, + 81 + ], + "spans": [ + { + "bbox": [ + 256, + 71, + 354, + 81 + ], + "type": "text", + "content": "Table 1. Styles Overview." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 91, + 545, + 277 + ], + "lines": [ + { + "bbox": [ + 65, + 91, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 65, + 91, + 545, + 277 + ], + "type": "table", + "html": "
CategoryExample Style Prompts
1. Artistic Living
1.1 Photography Styles:film noir, analog film photo, long exposure photo, et al.
1.2 Fashion Styles:glamorous photo, low contrast realistic photo, minimalist style, et al.
2. Fantasy Styles
2.1 Sci-Fi and Futurism:sci-fi style, futuristic style, retro-futuristic, space-themed, et al.
2.2 Cyberpunk Styles:vaporwave style, neonpunk style, neon noir, dystopian style, et al.
2.3 Fantasy and Themed Styles:fairy tale, gothic style, nautical-themed, disco-themed, et al.
3. Classical Art Styles
3.1 Genre Art Styles:impressionist painting, cubist artwork, expressionist, renaissance style, et al.
3.2 Painting Arts:watercolor painting, line art drawing, graffiti style, urban art, et al.
3.3 Paper Art and Handicrafts:origami style, kirigami representation, 3D papercut shadow box, et al.
4. ACG Styles
4.1 Anime and Cartoon Styles:animework, manga style, kawaii style, et al.
4.2 Comics and Game Art:comic, role-playing game style fantasy, retro game art, pixel art, et al.
", + "image_path": "6dda9b7af4e165739da8a1b71fee6638fa83da10f9b8b05ead58570f5bae1e41.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 297, + 295, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 295, + 321 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 295, + 321 + ], + "type": "text", + "content": "niques and enhances them by multiple prompt styles and pre-trained diffusion models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 335, + 245, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 245, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 245, + 348 + ], + "type": "text", + "content": "3. Dataset Construction and Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 357, + 295, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 295, + 380 + ], + "type": "text", + "content": "3.1. Evaluation of Recognition Models and Personalized Diffusion Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 388, + 296, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 296, + 470 + ], + "type": "text", + "content": "We begin by evaluating the effectiveness of existing evaluation models and personalized diffusion methods. For the evaluation metrics, we select two well-established metrics: CLIP-I, which measures identity preservation through CLIP features, and Face Sim., which is based on real face recognition models. For comparison, we also incorporate human evaluations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 473, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 604 + ], + "type": "text", + "content": "In selecting personalized diffusion methods, we choose four of the latest ID-preserving image generation techniques: PhotoMaker [29], InstanceID [58], Flashface [67], and IPAdapter [64]. All methods are implemented using the officially recommended settings. We analyze additional methods that are excluded from our study in Appendix. To ensure a comprehensive assessment, we additionally utilize the FID metric [18] and SDD-FIQA [39] to evaluate the overall image quality. We employ CLIP-T to evaluate the consistency between the stylized images and their corresponding textual prompts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "To comprehensively evaluate the different methods, we incorporate a diverse array of artistic styles. We first classify these common styles into 4 major categories and 10 sub-categories, as shown in Tab. 1. Our classification is based on people's usual perception of art and common forms of expression, ensuring that the grouping reflects intuitive and widely recognized distinctions in artistic presentation. For each sub-category, we curate a collection of specific prompts that together cover most common scenarios." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 295, + 426, + 368 + ], + "blocks": [ + { + "bbox": [ + 317, + 295, + 426, + 368 + ], + "lines": [ + { + "bbox": [ + 317, + 295, + 426, + 368 + ], + "spans": [ + { + "bbox": [ + 317, + 295, + 426, + 368 + ], + "type": "image", + "image_path": "f0c954e0a182b7ecb678241fb1a7d70788eeda937bb61d7effbd8ce023721e5d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 446, + 295, + 553, + 368 + ], + "blocks": [ + { + "bbox": [ + 446, + 295, + 553, + 368 + ], + "lines": [ + { + "bbox": [ + 446, + 295, + 553, + 368 + ], + "spans": [ + { + "bbox": [ + 446, + 295, + 553, + 368 + ], + "type": "image", + "image_path": "d94eaf62b28e07f65a97c77d88177fcedeee56e047e1b228ca46ce6078d73e07.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 380, + 425, + 453 + ], + "blocks": [ + { + "bbox": [ + 359, + 371, + 396, + 380 + ], + "lines": [ + { + "bbox": [ + 359, + 371, + 396, + 380 + ], + "spans": [ + { + "bbox": [ + 359, + 371, + 396, + 380 + ], + "type": "text", + "content": "IPAdapter" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 380, + 425, + 453 + ], + "lines": [ + { + "bbox": [ + 317, + 380, + 425, + 453 + ], + "spans": [ + { + "bbox": [ + 317, + 380, + 425, + 453 + ], + "type": "image", + "image_path": "87a5da8a13c6d56a6381cf14c8405289f048d3a7fbb8b4a39faea1d7d19dafbe.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 360, + 456, + 393, + 465 + ], + "lines": [ + { + "bbox": [ + 360, + 456, + 393, + 465 + ], + "spans": [ + { + "bbox": [ + 360, + 456, + 393, + 465 + ], + "type": "text", + "content": "InstantID" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 474, + 555, + 497 + ], + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 497 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 497 + ], + "type": "text", + "content": "Figure 3. Quality Evaluation. We evaluate the generation quality of different methods using SDD-FIQA [39]." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 445, + 380, + 553, + 453 + ], + "blocks": [ + { + "bbox": [ + 489, + 371, + 523, + 379 + ], + "lines": [ + { + "bbox": [ + 489, + 371, + 523, + 379 + ], + "spans": [ + { + "bbox": [ + 489, + 371, + 523, + 379 + ], + "type": "text", + "content": "Flashface" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 445, + 380, + 553, + 453 + ], + "lines": [ + { + "bbox": [ + 445, + 380, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 445, + 380, + 553, + 453 + ], + "type": "image", + "image_path": "8b4a658085449901092217fd1b433a465c02d7f29eea4c3ed14240b06da41d7e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 486, + 456, + 529, + 465 + ], + "lines": [ + { + "bbox": [ + 486, + 456, + 529, + 465 + ], + "spans": [ + { + "bbox": [ + 486, + 456, + 529, + 465 + ], + "type": "text", + "content": "Photomaker" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "text", + "content": "Our collection is both diverse and extensive, comprising: 1) we use official SDXL prompts [43], carefully selected for their high quality and distinctive style; 2) we also collect popular repository resources [3]. After deduplicating and removing prompts that conflict with identity information, we retain 80 distinct style prompts that cover a broad spectrum of common scenarios. We use 1,000 IDs from the CASIA-WebFace[65] dataset, and each method generates 80,000 images." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "We present the results of our analysis in Tab. 2 and Fig. 3. Our findings indicate that the similarity scores derived from CLIP features lack sufficient discriminative power and exhibit bias when compared to human evaluations. This limitation stems from CLIP's focus on extracting a generalized representation of identity. In contrast, real face recognition models align more closely with human judgment. For fur" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13056" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 115, + 91, + 496, + 170 + ], + "blocks": [ + { + "bbox": [ + 105, + 70, + 504, + 82 + ], + "lines": [ + { + "bbox": [ + 105, + 70, + 504, + 82 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 504, + 82 + ], + "type": "text", + "content": "Table 2. Comparison of Different Personalized Diffusion Methods Based on Various Evaluation Metrics." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 115, + 91, + 496, + 170 + ], + "lines": [ + { + "bbox": [ + 115, + 91, + 496, + 170 + ], + "spans": [ + { + "bbox": [ + 115, + 91, + 496, + 170 + ], + "type": "table", + "html": "
MethodPreserve ID InformationText ConsistencyGeneration Quality
CLIP-IFace Sim.Human EvalCLIP-TFID
PhotoMaker [29]54.5%12.3%21.9%20.5%313
InstanceID [58]65.2%70.9%90.0%26.1%285
Flashface [67]65.6%35.9%87.1%22.9%244
IPAdapter [67]60.3%25.3%64.9%25.1%255
", + "image_path": "c423711dbb367c3fed03b8e884ab85732053fe40d95cc0b299b5b16e2d4ab964.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 190, + 294, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 294, + 250 + ], + "type": "text", + "content": "ther study, a detailed assessment of real face recognition models is provided in the experiment section. The evaluation under classification metrics indicates that real face recognition models are also inaccurate in assessing identity consistency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 251, + 295, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 295, + 300 + ], + "type": "text", + "content": "In terms of the specific method, InstantID achieves the best performance in preserving ID information, text preservation, and quality distribution, demonstrating strong performance. In terms of FID score, Flashface performs better." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 310, + 196, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 310, + 196, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 196, + 323 + ], + "type": "text", + "content": "3.2. Setup and Data Cleaning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 328, + 296, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 296, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 296, + 424 + ], + "type": "text", + "content": "Based on Sec. 3.1, InstantID [58] provides stable performance, making it an ideal choice for our task. As a result, we choose InstantID as the primary method for generation. To further enhance the diversity of the dataset, we use three Pre-trained Diffusion Models:SDXL [43], YamerMIX [2], Anything-XL [1]. To further enrich the dataset, we incorporate three additional methods: PhotoMaker [29], Flashface [67] and IPAdapter [64]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 425, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 296, + 628 + ], + "type": "text", + "content": "We use 10k IDs from CISIA-Webface and 52k IDs from Glint360k as the original images. To enhance generation quality, we apply a series of filtering steps to the original dataset. These steps include: 1) discarding images with low similarity to others within the same ID; 2) removing low-quality images based on SDD-FIQA results, specifically those with scores below 60, which eliminates blurry or heavily tilted images; and 3) filtering out images with face angles exceeding 30 degrees, based on human pose detection. After applying these filtering steps, " + }, + { + "bbox": [ + 55, + 425, + 296, + 628 + ], + "type": "inline_equation", + "content": "78\\%" + }, + { + "bbox": [ + 55, + 425, + 296, + 628 + ], + "type": "text", + "content": " of the face images are retained. We generate the dataset by iterating over all combinations of methods, pre-trained diffusion models, and prompts across the entire dataset, producing more than 10 million images. We directly skip combinations where over " + }, + { + "bbox": [ + 55, + 425, + 296, + 628 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 425, + 296, + 628 + ], + "type": "text", + "content": " of the images are identified by all annotators as not matching the original ID will be entirely excluded." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": "For the stylized images, we begin with preprocessing. We use the RetinaFace model for face detection, adjusting the detected faces based on five face landmarks. Notably, " + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "8.4\\%" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": " of the images fail to detect any faces. In such cases, we resize the original image directly to maintain a consistent image size. We also compared additional face detection models, with more details provided in Appendix." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 319, + 188, + 552, + 366 + ], + "blocks": [ + { + "bbox": [ + 319, + 188, + 552, + 366 + ], + "lines": [ + { + "bbox": [ + 319, + 188, + 552, + 366 + ], + "spans": [ + { + "bbox": [ + 319, + 188, + 552, + 366 + ], + "type": "image", + "image_path": "3fd2d0e242c099542ecad6e1d7d457bff8717d5826576adaa7de6f9ec16ec084.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 376, + 555, + 540 + ], + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 540 + ], + "type": "text", + "content": "Figure 4. Pipeline of Semi-automatic Process. D. indicates delete, M. indicates manual verification, Sim. BSI measures the similarity between stylized images, Sim. BOI measures the similarity between original images, and Sim. BIP measures the similarity between a stylized image and its corresponding original image. We alternate between two phases: model verification and manual verification. In the model image pairs, a trained face recognition model performs two types of filtering: 1) The model compares the similarity between a query image and other stylized images generated from the highly similar original face. If the similarity score (TAR@FAR 1e-3) falls below the threshold, the query image is discarded. 2) It removes image pairs (stylized image and its origin image) with low similarity (TAR@FAR 1e-2). During the manual verification phase, these confusion image pairs (between TAR@FAR 1e-2 and TAR@FAR 1e-3) are labeled." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 567, + 554, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 554, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 554, + 662 + ], + "type": "text", + "content": "Due to the instability of generative methods, we need to verify whether each image truly preserves the ID information. However, given the large scale of the data, comprehensive manual verification becomes impractical. To alleviate the annotation burden, we employ a semi-automatic process for data cleaning, as shown in Fig. 4. In the semi-automated process, we alternate between two main phases: model validation and manual verification." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "In the model validation phase, we first train a face recognition model using the remaining dataset. We use two methods for rapid filtering by trained model: 1) Images with a similarity to original real faces below a threshold" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13057" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "(TAR@FAR at 1e-2) are deleted. 2) Generally, images generated from similar original images tend to have higher similarity; if there is a significant deviation, it is highly likely to be a bad case. Images with similarity to other stylized images from the same ID below the threshold (TAR@FAR at 1e-3), while similarity between their corresponding original images greater than 0.8, are also deleted." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 157, + 295, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 157, + 295, + 263 + ], + "spans": [ + { + "bbox": [ + 54, + 157, + 295, + 263 + ], + "type": "text", + "content": "In the manual validation phase, annotators compare the confusion images (between TAR@FAR at 1e-2 and TAR@FAR at 1e-3) to verify the ID information. During the annotation process, we develop a tool that provides yes and no options for each image pair, which are confirmed by the annotators. If more than " + }, + { + "bbox": [ + 54, + 157, + 295, + 263 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 54, + 157, + 295, + 263 + ], + "type": "text", + "content": " of the annotators believe that the two images belong to the same ID, they are retained. Any images below this threshold are excluded from the dataset." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 270, + 157, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 157, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 157, + 283 + ], + "type": "text", + "content": "3.3. Dataset statistics" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 288, + 295, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 295, + 346 + ], + "type": "text", + "content": "Large scale images. The Stylized-Face dataset consists of 4,660,285 images of 62,446 individuals. The generation size of all images is determined by the default parameters of their methods, and ultimately, they are sized at " + }, + { + "bbox": [ + 55, + 288, + 295, + 346 + ], + "type": "inline_equation", + "content": "112 \\times 112" + }, + { + "bbox": [ + 55, + 288, + 295, + 346 + ], + "type": "text", + "content": " pixels." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 348, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 296, + 407 + ], + "type": "text", + "content": "Utilized Models. Tab. 3 illustrates the proportions of various methods employed in our dataset. The higher proportion of the InstantID(SD) method is due to its more common use in real-world applications. We provide a detailed comparison of these methods in Appendix." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 83, + 437, + 269, + 535 + ], + "blocks": [ + { + "bbox": [ + 89, + 418, + 261, + 429 + ], + "lines": [ + { + "bbox": [ + 89, + 418, + 261, + 429 + ], + "spans": [ + { + "bbox": [ + 89, + 418, + 261, + 429 + ], + "type": "text", + "content": "Table 3. The proportions of various methods" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 437, + 269, + 535 + ], + "lines": [ + { + "bbox": [ + 83, + 437, + 269, + 535 + ], + "spans": [ + { + "bbox": [ + 83, + 437, + 269, + 535 + ], + "type": "table", + "html": "
MethodsProportion
InstantID(SD) [58]69%
InstantID(YamerMIX) [58]14%
InstantID(Anything-XL) [58]14%
PhotoMaker [29]1%
Flashface [67]1%
IPAdapter [67]1%
", + "image_path": "d1ffa45b2ce33e1a6875d748ff8bbce1651e802b4a80c73cddbbb4ff1b231b9d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 546, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 605 + ], + "type": "text", + "content": "Mean Faces. Fig. 5 displays the average faces for individual IDs from our dataset compared with those from CASIA-WebFace. Notably, while our dataset exhibit considerable style variations, the core facial features remain relatively stable." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 616, + 284, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 284, + 631 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 284, + 631 + ], + "type": "text", + "content": "4. Experiments of Stylized Face Recognition" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 636, + 172, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 172, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 172, + 647 + ], + "type": "text", + "content": "4.1. Evaluation Protocol" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": "Standard Stylized Face Recognition. The training set consists of 4,453,814 images generated by the InstantID method, encompassing 61k IDs. Additionally, we also use the real face images corresponding origin dataset for training. The test set includes 1k independent IDs, with 56,644" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 322, + 72, + 375, + 125 + ], + "blocks": [ + { + "bbox": [ + 322, + 72, + 375, + 125 + ], + "lines": [ + { + "bbox": [ + 322, + 72, + 375, + 125 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 375, + 125 + ], + "type": "image", + "image_path": "67c2cc1f17e39ac7f64e38380b84262ee469cc785cde1bfcecd8a8c44ef326e8.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 126, + 375, + 144 + ], + "lines": [ + { + "bbox": [ + 322, + 126, + 375, + 144 + ], + "spans": [ + { + "bbox": [ + 322, + 126, + 375, + 144 + ], + "type": "text", + "content": "CISIA-WebFace II" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 377, + 72, + 429, + 124 + ], + "blocks": [ + { + "bbox": [ + 377, + 72, + 429, + 124 + ], + "lines": [ + { + "bbox": [ + 377, + 72, + 429, + 124 + ], + "spans": [ + { + "bbox": [ + 377, + 72, + 429, + 124 + ], + "type": "image", + "image_path": "355814a46a3c471a3698697306813e813181212f81cb46d93d41c6d0f508beeb.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 126, + 427, + 143 + ], + "lines": [ + { + "bbox": [ + 383, + 126, + 427, + 143 + ], + "spans": [ + { + "bbox": [ + 383, + 126, + 427, + 143 + ], + "type": "text", + "content": "Stylized-Face" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 442, + 72, + 493, + 124 + ], + "blocks": [ + { + "bbox": [ + 442, + 72, + 493, + 124 + ], + "lines": [ + { + "bbox": [ + 442, + 72, + 493, + 124 + ], + "spans": [ + { + "bbox": [ + 442, + 72, + 493, + 124 + ], + "type": "image", + "image_path": "314d148384962168cedb61bcc40c019027408045aa400583827e2b427dc778e0.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 154, + 554, + 197 + ], + "lines": [ + { + "bbox": [ + 313, + 154, + 554, + 197 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 554, + 197 + ], + "type": "text", + "content": "Figure 5. Display of average faces for single ID. We randomly select two IDs from the CISIA-WebFace dataset and compute the average face for each ID using the images corresponding to those IDs." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 498, + 72, + 548, + 124 + ], + "blocks": [ + { + "bbox": [ + 443, + 126, + 545, + 144 + ], + "lines": [ + { + "bbox": [ + 443, + 126, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 443, + 126, + 545, + 144 + ], + "type": "text", + "content": "CISIA-WebFace Stylized-Face ID 2" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 498, + 72, + 548, + 124 + ], + "lines": [ + { + "bbox": [ + 498, + 72, + 548, + 124 + ], + "spans": [ + { + "bbox": [ + 498, + 72, + 548, + 124 + ], + "type": "image", + "image_path": "c8545101ca8549dcbb266a189e41c5c5bc5a88db91a1c81429fe45a15be7cdfd.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 502, + 126, + 545, + 143 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 502, + 126, + 545, + 143 + ], + "spans": [ + { + "bbox": [ + 502, + 126, + 545, + 143 + ], + "type": "text", + "content": "Stylized-Face" + } + ] + } + ], + "index": 18, + "type": "text" + }, + { + "bbox": [ + 313, + 219, + 554, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 219, + 554, + 267 + ], + "spans": [ + { + "bbox": [ + 313, + 219, + 554, + 267 + ], + "type": "text", + "content": "images also generated by InstantID. Both the training and test sets utilize the full set of 80 prompts. The primary goal of this evaluation is to assess the model's performance in open-set stylized face recognition." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 267, + 554, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 267, + 554, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 267, + 554, + 422 + ], + "type": "text", + "content": "Cross-prompt Stylized Face Recognition. The training and test sets for Cross-prompt Stylized Face Recognition are subsets of the corresponding training and test sets for Standard Stylized Face Recognition. The training set only includes images generated using 60 prompts from the Standard Stylized Face Recognition training set. It contains a total of 3,811,707 images. Additionally, the real face images corresponding to the IDs is also used for training. The test set consists of images generated using the last 20 prompts from the Standard Stylized Face Recognition test set, totaling 42,361 images. This setting ensures that the prompts and IDs used in the training and test sets are distinct. More details are shown in Appendix." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 423, + 554, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 554, + 518 + ], + "type": "text", + "content": "Cross-method Stylized Face Recognition. In this setup, we aim to assess cross-method generalization. The training set is the same as that used in Standard Stylized Face Recognition. The test set consists of 1k independent IDs, with 149,827 images generated by Flashface, IP-Adapter, and Photomaker. In detail, Flashface contributed 70,912 images, IP-Adapter contributed 51,915 images, and Photomaker contributed 27,000 images." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 525, + 426, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 426, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 426, + 537 + ], + "type": "text", + "content": "4.2. Evaluation Metrics" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 543, + 554, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 554, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 554, + 626 + ], + "type": "text", + "content": "In all experiments, we follow the standard evaluation metrics for face recognition. We evaluate the performance of the model on face verification. The gallery consists of all images in the test set, excluding the test images themselves. We report the results at multiple points, including TAR@FAR at 1e-2, TAR@FAR at 1e-3, and TAR@FAR at 1e-4. The metrics are defined as follows:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 628, + 554, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 554, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 554, + 664 + ], + "type": "text", + "content": "- True Acceptance Rate (TAR) (complementary to False Non-Match Rate (FNMR)): Proportion of correctly identified valid identities:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 411, + 670, + 464, + 694 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 670, + 464, + 694 + ], + "spans": [ + { + "bbox": [ + 411, + 670, + 464, + 694 + ], + "type": "interline_equation", + "content": "T A R = \\frac {M}{N}", + "image_path": "a58a2e2208c1b56ee2afc1dd0bf5028acc2b2fbc316533c1aba86cc636cfad70.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 322, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 322, + 701, + 553, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 322, + 701, + 553, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 322, + 701, + 553, + 713 + ], + "type": "text", + "content": " is the number of correctly identified identities," + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13058" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 72, + 274, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 72, + 274, + 83 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 274, + 83 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 63, + 72, + 274, + 83 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 63, + 72, + 274, + 83 + ], + "type": "text", + "content": " is the total number of valid identity attempts." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 84, + 181, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 84, + 181, + 95 + ], + "spans": [ + { + "bbox": [ + 63, + 84, + 181, + 95 + ], + "type": "text", + "content": "Note: " + }, + { + "bbox": [ + 63, + 84, + 181, + 95 + ], + "type": "inline_equation", + "content": "{FNMR} = 1 - {TAR}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 96, + 295, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 133 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 133 + ], + "type": "text", + "content": "- False Acceptance Rate (FAR) (equivalent to False Match Rate (FMR)): Proportion of non-target identities incorrectly accepted:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 154, + 135, + 206, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 135, + 206, + 159 + ], + "spans": [ + { + "bbox": [ + 154, + 135, + 206, + 159 + ], + "type": "interline_equation", + "content": "F A R = \\frac {L}{K}", + "image_path": "780a03a3c548d064ddaa84ed5a659971b59fef7934800c9177c4d78e000bd5cf.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 162, + 295, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 162, + 295, + 186 + ], + "spans": [ + { + "bbox": [ + 63, + 162, + 295, + 186 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 63, + 162, + 295, + 186 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 63, + 162, + 295, + 186 + ], + "type": "text", + "content": " is the number of incorrect acceptances, and " + }, + { + "bbox": [ + 63, + 162, + 295, + 186 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 63, + 162, + 295, + 186 + ], + "type": "text", + "content": " is the number of non-target identity attempts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 186, + 157, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 186, + 157, + 197 + ], + "spans": [ + { + "bbox": [ + 63, + 186, + 157, + 197 + ], + "type": "text", + "content": "Note: " + }, + { + "bbox": [ + 63, + 186, + 157, + 197 + ], + "type": "inline_equation", + "content": "{FMR} = {FAR}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 198, + 295, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 295, + 235 + ], + "type": "text", + "content": "- TAR@FAR: Represents the True Acceptance Rate (TAR) at a specific False Acceptance Rate (FAR), used to evaluate system accuracy at a fixed security level." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 239, + 188, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 188, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 188, + 252 + ], + "type": "text", + "content": "4.3. Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 257, + 296, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 296, + 365 + ], + "type": "text", + "content": "We maintain the same setup for all experiments. Our backbone networks are IResNet-50 and IResNet-101. We apply Cosface [56] framework and use an angular margin of 0.3 for Cosface. We set the batch size to 512 and train for a total of 26 epochs. The models are subsequently trained using Stochastic Gradient Descent (SGD), with a momentum coefficient of 0.9 and a weight decay parameter of " + }, + { + "bbox": [ + 55, + 257, + 296, + 365 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 55, + 257, + 296, + 365 + ], + "type": "text", + "content": ". The initial learning rate is set at " + }, + { + "bbox": [ + 55, + 257, + 296, + 365 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-1}" + }, + { + "bbox": [ + 55, + 257, + 296, + 365 + ], + "type": "text", + "content": " and is decreased by a factor of 8 at the midpoint of the epochs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 370, + 108, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 108, + 382 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 108, + 382 + ], + "type": "text", + "content": "4.4. Result" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 388, + 238, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 238, + 400 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 238, + 400 + ], + "type": "text", + "content": "4.4.1. Standard Stylized Face Recognition." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 403, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 295, + 510 + ], + "type": "text", + "content": "We assess the model's performance trained on the large-scale MS1MV3 [13] and Glint360k [6] dataset. As presented in Tab. 4, the results indicate that despite the substantial size of the training set, there is a significant performance degradation compared to real face scenarios. Although the model trained on Glint360k performs worse than ours, it shows significant improvement over the model trained on MS1MV3, highlighting the positive impact of dataset size on generalization." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 511, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 511, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 511, + 295, + 582 + ], + "type": "text", + "content": "The model trained on our dataset significantly outperforms the one trained on real face data in open-set face recognition scenarios, achieving a face verification TAR@FAR at 1e-4 of " + }, + { + "bbox": [ + 55, + 511, + 295, + 582 + ], + "type": "inline_equation", + "content": "90.6\\%" + }, + { + "bbox": [ + 55, + 511, + 295, + 582 + ], + "type": "text", + "content": ", as detailed in Tab. 4. This highlights the crucial role of our dataset in effectively recognizing stylized faces." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 100, + 622, + 252, + 679 + ], + "blocks": [ + { + "bbox": [ + 55, + 590, + 295, + 613 + ], + "lines": [ + { + "bbox": [ + 55, + 590, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 295, + 613 + ], + "type": "text", + "content": "Table 4. Results on Standard Stylized Face Recognition. Backbone is IResNet-50. Values in bold indicate the top results." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 100, + 622, + 252, + 679 + ], + "lines": [ + { + "bbox": [ + 100, + 622, + 252, + 679 + ], + "spans": [ + { + "bbox": [ + 100, + 622, + 252, + 679 + ], + "type": "table", + "html": "
Training sets1e-21e-31e-4
MS1MV3 [13]85.168.049.7
Gint360k [6]95.087.074.7
Stylized-Face97.995.590.6
", + "image_path": "dc5aee8b9869bcf0441ce0410355b6e956885bd4f5750b2f9b54098d876d9311.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "As shown in Fig. 6, the model trained on our dataset significantly outperforms the one trained on real face data in" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 72, + 554, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 107 + ], + "type": "text", + "content": "open-set face recognition scenarios. It highlights the crucial role of our dataset in effectively recognizing stylized faces." + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 330, + 128, + 545, + 266 + ], + "blocks": [ + { + "bbox": [ + 330, + 128, + 545, + 266 + ], + "lines": [ + { + "bbox": [ + 330, + 128, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 330, + 128, + 545, + 266 + ], + "type": "image", + "image_path": "a0fee25d5cbf56b5dbe7d2f868628be37cfe3e5fd28f308214db990acaeb9103.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 278, + 554, + 300 + ], + "lines": [ + { + "bbox": [ + 313, + 278, + 554, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 554, + 300 + ], + "type": "text", + "content": "Figure 6. the FMR-FNMR plot for All pairs verification. (lower FNMR is better)." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 324, + 509, + 336 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 509, + 336 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 509, + 336 + ], + "type": "text", + "content": "4.4.2. Cross-prompt stylized face recognition." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "type": "text", + "content": "Table 5 displays the performance across various prompts using the InstanceID [58], highlighting its generalization capabilities in stylized face recognition across different prompts. It also underscores our dataset's effectiveness in enhancing the cross-prompt generalization of stylized face recognition." + } + ] + } + ], + "index": 20 + }, + { + "type": "table", + "bbox": [ + 359, + 453, + 511, + 511 + ], + "blocks": [ + { + "bbox": [ + 313, + 422, + 554, + 445 + ], + "lines": [ + { + "bbox": [ + 313, + 422, + 554, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 554, + 445 + ], + "type": "text", + "content": "Table 5. Results on Cross-prompt Stylized Face Recognition. Backbone is IResNet-50. Values in bold indicate the top results." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 359, + 453, + 511, + 511 + ], + "lines": [ + { + "bbox": [ + 359, + 453, + 511, + 511 + ], + "spans": [ + { + "bbox": [ + 359, + 453, + 511, + 511 + ], + "type": "table", + "html": "
Training sets1e-21e-31e-4
MS1MV3 [13]85.767.749.3
Gint360k [6]94.786.573.9
Stylized-Face97.995.490.3
", + "image_path": "1c5473515d28e89a55c186be97a9fa1b234acbb6fd6df821c5f37ae2fb28d2dd.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 530, + 515, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 515, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 515, + 542 + ], + "type": "text", + "content": "4.4.3. Cross-method Stylized Face Recognition" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 545, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 628 + ], + "type": "text", + "content": "Cross-method experiments in Table 6 evaluate the ability of models to generalize feature extraction across various generation methods. The results indicate that, despite our method showing clear superiority over real-face models, current methods struggle with cross-method generalization across different methods, emphasizing the need for more in-depth research to enhance cross-method robustness." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "We further analyze cross-method generalization using the Flashface dataset. Prompts are divided into easy and hard categories based on the distribution shift, measured by the difference between Flashface and the InstantID method on the same data. Results show that performance on easy prompts is similar to within-distribution performance, while generalization challenges are more pronounced for hard" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13059" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 86, + 123, + 267, + 296 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 116 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 116 + ], + "type": "text", + "content": "Table 6. Results on Cross-method Stylized Face Recognition. Backbone is IResNet-50. All refers to Flashface, IP-Adapter, and Photomaker. Due to the limited number of images, the results exclude TAR@FAR at 1e-4. Values in bold indicate the top results." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 86, + 123, + 267, + 296 + ], + "lines": [ + { + "bbox": [ + 86, + 123, + 267, + 296 + ], + "spans": [ + { + "bbox": [ + 86, + 123, + 267, + 296 + ], + "type": "table", + "html": "
Training setsTest Parts1e-21e-3
MS1MV3 [13]Flashface82.171.0
Gint360k [6]Flashface92.386.7
Stylized-FaceFlashface93.187.3
MS1MV3 [13]IP-Adapter68.262.2
Gint360k [6]IP-Adapter91.474.3
Stylized-FaceIP-Adapter93.678.3
MS1MV3 [13]Photomaker62.636.4
Gint360k [6]Photomaker86.562.2
Stylized-FacePhotomaker90.171.2
MS1MV3 [13]All59.031.8
Gint360k [6]All78.758.6
Stylized-FaceAll87.271.9
", + "image_path": "0854cb7d9932feb3f070adc15174ec2a267f5ac459f8c7de0fb1534621f4ffc4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 312, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 296, + 361 + ], + "type": "text", + "content": "prompts with significant distributional shifts. Notably, for these hard prompts, our model demonstrates particularly remarkable performance improvements compared to the real face recognition model. Results are shown in Tab. 7." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 67, + 411, + 284, + 460 + ], + "blocks": [ + { + "bbox": [ + 55, + 370, + 296, + 403 + ], + "lines": [ + { + "bbox": [ + 55, + 370, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 296, + 403 + ], + "type": "text", + "content": "Table 7. Analysis of Cross-method generalization. The evaluation metrics is TAR@FAR at 1e-2 on Flashface. Backbone is IResNet-50." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 411, + 284, + 460 + ], + "lines": [ + { + "bbox": [ + 67, + 411, + 284, + 460 + ], + "spans": [ + { + "bbox": [ + 67, + 411, + 284, + 460 + ], + "type": "table", + "html": "
MethodsEasy PromptsHard PromptsAll
Glint360k96.984.992.3
Stylized-Face96.987.193.1
", + "image_path": "86ef15ab03947bcd70f9f587ae188efafa009c98851b52e192a1fb1966740071.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 475, + 149, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 149, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 149, + 489 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 493, + 297, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 297, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 297, + 567 + ], + "type": "text", + "content": "Performance on Real Face Datasets. We evaluate the model's performance on real face datasets, with the results presented in Fig. 8. The comparison demonstrates that while our model achieves comparable performance, there exists a trade-off between real face recognition and stylized face recognition." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 72, + 605, + 278, + 691 + ], + "blocks": [ + { + "bbox": [ + 55, + 574, + 296, + 597 + ], + "lines": [ + { + "bbox": [ + 55, + 574, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 296, + 597 + ], + "type": "text", + "content": "Table 8. Results on Real Face Recognition. Backbone is IResNet-50." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 605, + 278, + 691 + ], + "lines": [ + { + "bbox": [ + 72, + 605, + 278, + 691 + ], + "spans": [ + { + "bbox": [ + 72, + 605, + 278, + 691 + ], + "type": "table", + "html": "
Training setLFWCFP_FPAgeDBAvg.
CASIA99.595.294.996.5
MS1MV299.898.198.198.7
MS1MV399.898.598.398.9
Glint360k99.899.198.599.1
Stylized-Face99.798.497.998.7
", + "image_path": "e7194a974665c45e179726ffa584251a580a615a0721fec5b37b691b47d6e18f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 701, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 715 + ], + "type": "text", + "content": "More Methods To comprehensively assess the perfor" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "text", + "content": "mance of stylized face recognition dataset, we compare advanced face recognition methods, including Cosface [56], Arcface [13], and CurricularFace [23]. These methods represent leading techniques in the field of face recognition. The experimental results are summarized in Tab. 9. The performance of the different training methods is similar." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 350, + 185, + 520, + 243 + ], + "blocks": [ + { + "bbox": [ + 313, + 154, + 555, + 177 + ], + "lines": [ + { + "bbox": [ + 313, + 154, + 555, + 177 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 555, + 177 + ], + "type": "text", + "content": "Table 9. Performance Comparison of Different Training Methods on Standard Stylized Face Recognition." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 350, + 185, + 520, + 243 + ], + "lines": [ + { + "bbox": [ + 350, + 185, + 520, + 243 + ], + "spans": [ + { + "bbox": [ + 350, + 185, + 520, + 243 + ], + "type": "table", + "html": "
Training Methods1e-21e-31e-4
Cosface [56]97.995.590.6
CurricularFace [23]97.995.590.5
Arcface [13]97.995.489.9
", + "image_path": "ec3880e0d0577c711eb01366f43040e10f42f0e47760237394d2b6634e9c2564.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 254, + 555, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 254, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 555, + 291 + ], + "type": "text", + "content": "More Backbones We evaluate the performance of various backbones and Tab. 10 shows that scaling up the model size can lead to further improvements in performance." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 362, + 331, + 507, + 378 + ], + "blocks": [ + { + "bbox": [ + 313, + 300, + 555, + 323 + ], + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 323 + ], + "type": "text", + "content": "Table 10. Performance Comparison of Different Backbones on Standard Stylized Face Recognition." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 362, + 331, + 507, + 378 + ], + "lines": [ + { + "bbox": [ + 362, + 331, + 507, + 378 + ], + "spans": [ + { + "bbox": [ + 362, + 331, + 507, + 378 + ], + "type": "table", + "html": "
Backbone1e-21e-31e-4
IResNet-5097.995.590.6
IResNet-10198.396.392.0
", + "image_path": "23c2eaa8b2ea357c327d6765338a47a64151cdcafb432e3b018fd3a49f455cd9.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 399, + 388, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 399, + 388, + 412 + ], + "spans": [ + { + "bbox": [ + 314, + 399, + 388, + 412 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 419, + 555, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 555, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 555, + 576 + ], + "type": "text", + "content": "In addressing the challenge of recognizing stylized face images, we introduce the Stylized-Face dataset based on the CASIA-WebFace and Glint360k. Our dataset includes 4.6 million images spanning 62k IDs and incorporate a wide range of styles and identity-preserving methods. To ensure data quality (i.e., ID preservation) at this massive scale, we implement a semi-automated pipeline for large-scale data cleaning. Stylized-Face is the first dataset for stylized face recognition. Furthermore, we establish benchmarks to evaluate the robustness and generalization of stylized face recognition methods. Experimental results demonstrate that models trained on Stylized-Face achieve remarkable improvements in stylized face recognition." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 586, + 477, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 477, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 477, + 598 + ], + "type": "text", + "content": "6. Limitation and Future Work" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "This task continues to present challenges and opportunities for improvement. (1) Cross-method Stylized Face Recognition requires the adoption of additional designs to improve its generalization capabilities. (2) Explore approaches that can simultaneously achieve state-of-the-art performance in both real and stylized face recognition (3) Furthermore, considering the various long-tail distribution styles in real-world scenarios, which is also a direction for future research." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13060" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 224 + ], + "type": "text", + "content": "This work is supported by the National Natural Science Foundation of China (62302167, U23A20343, 62472282, 72192821, 62476090), Young Elite Scientists Sponsorship Program by CAST (YESS20240780), Shanghai Sailing Program (23YF1410500), Chenguang Program of Shanghai Education Development Foundation, CCF-Tencent Rhino-Bird Young Faculty Open Research Fund (CCF-Tencent RAGR20230121, RAGR20240122), Shanghai Municipal Education Commission (23CGA34) and the Open Research Fund of Key Laboratory of Advanced Theory and Application in Statistics and Data Science-MOE, ECNU." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 233, + 115, + 246 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 233, + 115, + 246 + ], + "spans": [ + { + "bbox": [ + 56, + 233, + 115, + 246 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 254, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 254, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 254, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 61, + 254, + 294, + 274 + ], + "type": "text", + "content": "[1] Anything-xl. https://civitai.com/models/9409/or-anything-xl.5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 277, + 295, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 277, + 295, + 298 + ], + "spans": [ + { + "bbox": [ + 61, + 277, + 295, + 298 + ], + "type": "text", + "content": "[2] Yamermix. https://civitai.com/models/84040/sdxl-unstable-diffusers-yamermix. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 300, + 294, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 300, + 294, + 322 + ], + "spans": [ + { + "bbox": [ + 62, + 300, + 294, + 322 + ], + "type": "text", + "content": "[3] Sdsl prompt styler, a custom node for comfyui. https://github.com/twri/sdsl_prompt_styler. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 323, + 294, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 323, + 294, + 344 + ], + "spans": [ + { + "bbox": [ + 62, + 323, + 294, + 344 + ], + "type": "text", + "content": "[4] Reverse style transfer. https://devpost.com/software/reverse-style-transfer. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 346, + 295, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 295, + 411 + ], + "type": "text", + "content": "[5] Xiang An, Xuhan Zhu, Yuan Gao, Yang Xiao, Yongle Zhao, Ziyong Feng, Lan Wu, Bin Qin, Ming Zhang, Debing Zhang, and Ying Fu. Partial fc: Training 10 million identities on a single machine. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 413, + 295, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 413, + 295, + 477 + ], + "spans": [ + { + "bbox": [ + 62, + 413, + 295, + 477 + ], + "type": "text", + "content": "[6] Xiang An, Xuhan Zhu, Yuan Gao, Yang Xiao, Yongle Zhao, Ziyong Feng, Lan Wu, Bin Qin, Ming Zhang, Debing Zhang, et al. Partial fc: Training 10 million identities on a single machine. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1445-1449, 2021. 2, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 480, + 295, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 295, + 534 + ], + "type": "text", + "content": "[7] Xiang An, Jiangkang Deng, Jia Guo, Ziyong Feng, Xuhan Zhu, Yang Jing, and Liu Tongliang. Killing two birds with one stone: Efficient and robust training of face recognitionCNNs by partial fc. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 536, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 295, + 601 + ], + "type": "text", + "content": "[8] Gwangbin Bae, Martin de La Gorce, Tadas Baltrusaitis, Charlie Hewitt, Dong Chen, Julien Valentin, Roberto Cipolla, and Jingjing Shen. Digiface-1m: 1 million digital face images for face recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3526-3535, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 603, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 603, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 62, + 603, + 295, + 656 + ], + "type": "text", + "content": "[9] Fadi Boutros, Marco Huber, Patrick Siebke, Tim Rieber, and Naser Damer. Sface: Privacy-friendly and accurate face recognition using synthetic data. In 2022 IEEE International Joint Conference on Biometrics (IJCB), pages 1-11. IEEE, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 58, + 658, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 658, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 658, + 295, + 714 + ], + "type": "text", + "content": "[10] Fadi Boutros, Marcel Klemt, Meiling Fang, Arjan Kuijper, and Naser Damer. Unsupervised face recognition using unlabeled synthetic data. In 2023 IEEE 17th International Conference on Automatic Face and Gesture Recognition (FG), pages 1-8. IEEE, 2023. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[11] PAN Chuanyu, YANG Guowei, MU Taijiang, and LAI Yu-Kun. Generating animatable 3d cartoon faces from single portraits. Virtual Reality & Intelligent Hardware, 6(4):292-307, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 118, + 554, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 183 + ], + "type": "text", + "content": "[12] Jun Dan, Yang Liu, Haoyu Xie, Jiankang Deng, Haoran Xie, Xuansong Xie, and Baigui Sun. Transface: Calibrating transformer training for face recognition from a data-centric perspective. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20642-20653, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 184, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 553, + 239 + ], + "type": "text", + "content": "[13] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 1, 3, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 240, + 553, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 272 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 272 + ], + "type": "text", + "content": "[14] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pages 1180-1189. PMLR, 2015. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 273, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 273, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 317, + 273, + 553, + 338 + ], + "type": "text", + "content": "[15] Jingyu Gong, Fengqi Liu, Jiachen Xu, Min Wang, Xin Tan, Zhizhong Zhang, Ran Yi, Haichuan Song, Yuan Xie, and Lizhuang Ma. Optimization over disentangled encoding: Unsupervised cross-domain point cloud completion via occlusion factor manipulation. In European Conference on Computer Vision, pages 517-533. Springer, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 338, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 553, + 392 + ], + "type": "text", + "content": "[16] Jingyu Gong, Yujing Lou, Fengqi Liu, Zhiwei Zhang, Haoming Chen, Zhizhong Zhang, Xin Tan, Yuan Xie, and Lizhuang Ma. Scene point cloud understanding and reconstruction technologies in 3d space. Journal of Image and Graphics, 28(06):1741-1766, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 394, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 394, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 317, + 394, + 553, + 459 + ], + "type": "text", + "content": "[17] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 87-102. Springer, 2016. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 460, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 514 + ], + "type": "text", + "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 515, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 515, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 553, + 569 + ], + "type": "text", + "content": "[19] Chengyang Hu, Ke-Yue Zhang, Taiping Yao, Shice Liu, Shouhong Ding, Xin Tan, and Lizhuang Ma. Domainhallucinated updating for multi-domain face anti-spoofing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2193-2201, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 571, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 553, + 613 + ], + "type": "text", + "content": "[20] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 615, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 615, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 317, + 615, + 553, + 669 + ], + "type": "text", + "content": "[21] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[22] Lianghua Huang, Di Chen, Yu Liu, Yujun Shen, Deli Zhao, and Jingren Zhou. Composer: Creative and controllable image synthesis with composable conditions. arXiv preprint arXiv:2302.09778, 2023. 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13061" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "type": "text", + "content": "[23] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: adaptive curriculum learning loss for deep face recognition. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5901-5910, 2020. 1, 3, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 173 + ], + "type": "text", + "content": "[24] Jing Huoa, Wenbin Lia, Yinghuan Shia, Yang Gaoa, and Hujun Yinb. Webcaricature: a benchmark for caricature face recognition. arXiv preprint arXiv:1703.03230, 2017. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "text", + "content": "[25] Shen Li, Jianqing Xu, Xiaqing Xu, Pengcheng Shen, Shaoxin Li, and Bryan Hooi. Spherical confidence learning for face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15629-15637, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 296 + ], + "type": "text", + "content": "[26] Xiaofan Li, Xin Tan, Zhuo Chen, Zhizhong Zhang, Ruixin Zhang, Rizen Guo, Guanna Jiang, Yulong Chen, Yanyun Qu, Lizhuang Ma, et al. One-for-more: Continual diffusion model for anomaly detection. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 4766-4775, 2025. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 297, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 330 + ], + "type": "text", + "content": "[27] Yidi Li, Jun Xiao, Yiqun Wang, and Zhengda Lu. Depthgan: Gan-based depth generation from semantic layouts. Computational Visual Media, 10(3):505-522, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 374 + ], + "type": "text", + "content": "[28] Yuxuan Li, Lingfeng Yang, and Xiang Li. Apf-gan: Exploring asymmetric pre-training and fine-tuning strategy for conditional generative adversarial network. Computational Visual Media, 10(1):187-192, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 421 + ], + "type": "text", + "content": "[29] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. arXiv preprint arXiv:2312.04461, 2023. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 294, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 455 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 455 + ], + "type": "text", + "content": "[30] Zonglin Li, Xiaogian Lv, Wei Yu, Qinglin Liu, Jingbo Lin, and Shengping Zhang. Face shape transfer via semantic warping. Visual Intelligence, 2(1):26, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "type": "text", + "content": "[31] Xiao Lin, Yuge Huang, Jianqing Xu, Yuxi Mi, Shuigeng Zhou, and Shouhong Ding. Uface: Unleashing inherent model capabilities to enhance intra-class diversity in synthetic face recognition. arXiv preprint arXiv:2502.19803, 2025.3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 566 + ], + "type": "text", + "content": "[32] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherical face: Deep hypersphere embedding for face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 212-220, 2017. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 601 + ], + "type": "text", + "content": "[33] Wang Lu, Jindong Wang, Haoliang Li, Yiqiang Chen, and Xing Xie. Domain-invariant feature exploration for domain generalization. arXiv preprint arXiv:2207.12020, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 602, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 667 + ], + "type": "text", + "content": "[34] Brianna Maze, Jocelyn Adams, James A Duncan, Nathan Kalka, Tim Miller, Charles Otto, Anil K Jain, W Tyler Niggel, Janet Anderson, Jordan Cheney, et al. Iarpa janus benchmark-c: Face dataset and protocol. In 2018 international conference on biometrics (ICB), pages 158-165. IEEE, 2018. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[35] Yuxi Mi, Zhizhou Zhong, Yuge Huang, Jiazhen Ji, Jianqing Xu, Jun Wang, Shaoming Wang, Shouhong Ding, and Shuigeng Zhou. Privacy-preserving face recognition using trainable feature subtraction. In Proceedings of the" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 297-307, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 97, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 163 + ], + "type": "text", + "content": "[36] Yuxi Mi, Zhizhou Zhong, Yuge Huang, Qiuyang Yuan, Xuan Zhao, Jianqing Xu, Shouhong Ding, Shaoming Wang, Rizen Guo, and Shuigeng Zhou. Data synthesis with diverse styles for face recognition via 3dmm-guided diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21203-21214, 2025. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 164, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 553, + 228 + ], + "type": "text", + "content": "[37] Ashutosh Mishra, Shyam Nandan Rai, Anand Mishra, and CV Jawahar. Iit-cfw: A benchmark database of cartoon faces in the wild. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part I 14, pages 35-47. Springer, 2016. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 230, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 285 + ], + "type": "text", + "content": "[38] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 287, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 353 + ], + "type": "text", + "content": "[39] Fu-Zhao Ou, Xingyu Chen, Ruixin Zhang, Yuge Huang, Shaoxin Li, Jilin Li, Yong Li, Liujuan Cao, and Yuan-Gen Wang. Sdd-fiaq: Unsupervised face image quality assessment with similarity distribution distance. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7670-7679, 2021. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 354, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 398 + ], + "type": "text", + "content": "[40] Zhengyuan Peng, Qijian Tian, Jianqing Xu, Yizhang Jin, Xuequan Lu, Xin Tan, Yuan Xie, and Lizhuang Ma. Generalized category discovery in semantic segmentation. arXiv preprint arXiv:2311.11525, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 399, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 454 + ], + "type": "text", + "content": "[41] Zhengyuan Peng, Jinpeng Ma, Zhimin Sun, Ran Yi, Haichuan Song, Xin Tan, and Lizhuang Ma. Mos: Modeling object-scene associations in generalized category discovery. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 15118-15128, 2025." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "type": "text", + "content": "[42] Zhengyuan Peng, Jianqing Xu, Shen Li, Jiazhen Ji, Yuge Huang, Jingyun Zhang, Jinmin Li, Shouhong Ding, Rizen Guo, Xin Tan, et al. Eyeseg: An uncertainty-aware eye segmentation framework for ar/vr. arXiv preprint arXiv:2507.09649, 2025.3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 512, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 566 + ], + "type": "text", + "content": "[43] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 3, 4, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 568, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 621 + ], + "type": "text", + "content": "[44] Haibo Qiu, Baosheng Yu, Dihong Gong, Zhifeng Li, Wei Liu, and Dacheng Tao. Synface: Face recognition with synthetic data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10880-10890, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 624, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 689 + ], + "type": "text", + "content": "[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[46] Qibing Ren, Chang Gao, Jing Shao, Junchi Yan, Xin Tan, Wai Lam, and Lizhuang Ma. Codeattack: Revealing safety" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13062" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "generalization challenges of large language models via code completion. arXiv preprint arXiv:2403.07865, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 149 + ], + "type": "text", + "content": "[47] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 294, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 294, + 215 + ], + "type": "text", + "content": "[48] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "type": "text", + "content": "[49] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 326 + ], + "type": "text", + "content": "[50] Yichen Shi, Yuhao Gao, Yingxin Lai, Hongyang Wang, Jun Feng, Lei He, Jun Wan, Changsheng Chen, Zitong Yu, and Xiaochun Cao. Shield: An evaluation benchmark for face spoofing and forgery detection with multimodal large language models. Visual Intelligence, 3(1):9, 2025. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 358 + ], + "type": "text", + "content": "[51] Kaede Shiohara and Toshihiko Yamasaki. Face2diffusion for fast and editable face personalization. arXiv preprint arXiv:2403.05094, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 360, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 403 + ], + "type": "text", + "content": "[52] Shuai Tan, Bin Ji, and Ye Pan. Style2talker: High-resolution talking head generation with emotion style and art style. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5079-5087, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 294, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 468 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 468 + ], + "type": "text", + "content": "[53] Shuai Tan, Biao Gong, Yutong Feng, Kecheng Zheng, Dandan Zheng, Shuwei Shi, Yujun Shen, Jingdong Chen, and Ming Yang. Mimir: Improving video diffusion models for precise text understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "text", + "content": "[54] Shuai Tan, Biao Gong, Yujie Wei, Shiwei Zhang, Zhuoxin Liu, Dandan Zheng, Jingdong Chen, Yan Wang, Hao Ouyang, Kecheng Zheng, and Yujun Shen. Synmotion: Semantic-visual adaptation for motion customized video generation. arXiv preprint arXiv:2506.23690, 2025. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 526, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 569 + ], + "type": "text", + "content": "[55] Ming Tao, Hao Tang, Songsong Wu, Nicu Sebe, Xiao-Yuan Jing, Fei Wu, and Bingkun Bao. Df-gan: Deep fusion generative adversarial networks for text-to-image synthesis. arXiv preprint arXiv:2008.05865, 2(6), 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 571, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 571, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 571, + 294, + 624 + ], + "type": "text", + "content": "[56] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5265-5274, 2018. 1, 3, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 625, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 668 + ], + "type": "text", + "content": "[57] Qian Wang, Cai Guo, Hong-Ning Dai, and Ping Li. Stroke-gan painter: Learning to paint artworks using stroke-style generative adversarial networks. Computational Visual Media, 9(4):787-806, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[58] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "type": "text", + "content": "[59] Tan Wang, Linjie Li, Kevin Lin, Chung-Ching Lin, Zhengyuan Yang, Hanwang Zhang, Zicheng Liu, and Lijuan Wang. disco: Disentangled control for referring human dance generation in real world. arXiv e-prints, pages arXiv-2307, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 194 + ], + "type": "text", + "content": "[60] Cameron Whitelam, Emma Taborsky, Austin Blanton, Brianna Maze, Jocelyn Adams, Tim Miller, Nathan Kalka, Anil K Jain, James A Duncan, Kristen Allen, et al. Iarpa janus benchmark-b face dataset. In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 90-98, 2017. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 195, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 237 + ], + "type": "text", + "content": "[61] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multisubject image generation with localized attention. arXiv preprint arXiv:2305.10431, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 239, + 553, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 303 + ], + "type": "text", + "content": "[62] Jianqing Xu, Shen Li, Jiaying Wu, Miao Xiong, Ailin Deng, Jiazhen Ji, Yuge Huang, Guodong Mu, Wenjie Feng, Shouhong Ding, et al. " + }, + { + "bbox": [ + 316, + 239, + 553, + 303 + ], + "type": "inline_equation", + "content": "\\text{Id}^3" + }, + { + "bbox": [ + 316, + 239, + 553, + 303 + ], + "type": "text", + "content": ": Identity-preserving-yet-diversified diffusion models for synthetic face recognition. Advances in Neural Information Processing Systems, 37: 77777-77798, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 304, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 553, + 369 + ], + "type": "text", + "content": "[63] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Finegrained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 371, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 414 + ], + "type": "text", + "content": "[64] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 415, + 553, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 446 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 446 + ], + "type": "text", + "content": "[65] Dong Yi, Zhen Lei, Shengcai Liao, and Stan Z Li. Learning face representation from scratch. arXiv preprint arXiv:1411.7923, 2014. 2, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "type": "text", + "content": "[66] Ran Yi, Teng Hu, Mengfei Xia, Yizhe Tang, and Yong-Jin Liu. Feditnet++: Few-shot editing of latent semantics in gan spaces with correlated attribute disentanglement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 504, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 557 + ], + "type": "text", + "content": "[67] Shilong Zhang, Lianghua Huang, Xi Chen, Yifei Zhang, Zhi-Fan Wu, Yutong Feng, Wei Wang, Yujun Shen, Yu Liu, and Ping Luo. Flashface: Human image personalization with high-fidelity identity preservation. arXiv preprint arXiv:2403.17008, 2024. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 559, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 602 + ], + "type": "text", + "content": "[68] Tianyue Zheng and Weihong Deng. Cross-pose lfw: A database for studying cross-posed face recognition in unconstrained environments. Beijing University of Posts and Telecommunications, Tech. Rep, 5(7):5, 2018. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 604, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 553, + 645 + ], + "type": "text", + "content": "[69] Tianyue Zheng, Weihong Deng, and Jiani Hu. Cross-age lfw: A database for studying cross-age face recognition in unconstrained environments. arXiv preprint arXiv:1708.08197, 2017.3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "text", + "content": "[70] Yi Zheng, Yifan Zhao, Mengyuan Ren, He Yan, Xiangju Lu, Junhui Liu, and Jia Li. Cartoon face recognition: A benchmark dataset. In Proceedings of the 28th ACM international conference on multimedia, pages 2264-2272, 2020. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 692, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 712 + ], + "type": "text", + "content": "[71] Zhizhou Zhong, Yuxi Mi, Yuge Huang, Jianqing Xu, Guodong Mu, Shouhong Ding, Jingyun Zhang, Rizen Guo," + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13063" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 194 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 75, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 294, + 116 + ], + "type": "text", + "content": "Yunsheng Wu, and Shuigeng Zhou. Slerpface: Face template protection via spherical linear interpolation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10698-10706, 2025. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 194 + ], + "type": "text", + "content": "[72] Zheng Zhu, Guan Huang, Jiankang Deng, Yun Ye, Junjie Huang, Xinze Chen, Jiagang Zhu, Tian Yang, Jiwen Lu, Dalong Du, and Jie Zhou. Webface260m: A benchmark unveiling the power of million-scale deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10492-10502, 2021. 3" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13064" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SuMa_ A Subspace Mapping Approach for Robust and Effective Concept Erasure in Text-to-Image Diffusion Models/51080fa1-f759-4df5-8927-63f76b232e8f_content_list.json b/2025/SuMa_ A Subspace Mapping Approach for Robust and Effective Concept Erasure in Text-to-Image Diffusion Models/51080fa1-f759-4df5-8927-63f76b232e8f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2d72d8c41c014792e23dc68445389f1a448d442b --- /dev/null +++ b/2025/SuMa_ A Subspace Mapping Approach for Robust and Effective Concept Erasure in Text-to-Image Diffusion Models/51080fa1-f759-4df5-8927-63f76b232e8f_content_list.json @@ -0,0 +1,1453 @@ +[ + { + "type": "text", + "text": "SuMa: A Subspace Mapping Approach for Robust and Effective Concept Erasure in Text-to-Image Diffusion Models", + "text_level": 1, + "bbox": [ + 129, + 128, + 869, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kien Nguyen1 Anh Tran1 Cuong Pham2,1", + "bbox": [ + 285, + 202, + 709, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Qualcomm AI Research", + "bbox": [ + 398, + 220, + 598, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2Posts and Telecommunications Institute of Technology (PTIT)", + "bbox": [ + 250, + 239, + 746, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{kienn, anhtra, pcuong}@qti.qualcomm.com", + "bbox": [ + 320, + 258, + 671, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The rapid growth of text-to-image diffusion models has raised concerns about their potential misuse in generating harmful or unauthorized contents. To address these issues, several Concept Erasure methods have been proposed. However, most of them fail to achieve both robustness, i.e., the ability to robustly remove the target concept, and effectiveness, i.e., maintaining image quality. While few recent techniques successfully achieve these goals for NSFW concepts, none could handle narrow concepts such as copyrighted characters or celebrities. Erasing these narrow concepts is critical in addressing copyright and legal concerns. However, erasing them is challenging due to their close distances to non-target neighboring concepts, requiring finer-grained manipulation. In this paper, we introduce Subspace Mapping (SuMa), a novel method specifically designed to achieve both robustness and effectiveness in easing these narrow concepts. SuMa first derives a target subspace representing the concept to be erased and then neutralizes it by mapping it to a reference subspace that minimizes the distance between the two. This mapping ensures the target concept is robustly erased while preserving image quality. We conduct extensive experiments with SuMa across four tasks: subclass erasure, celebrity erasure, artistic style erasure, and instance erasure and compare the results with current state-of-the-art methods. Our method achieves image quality comparable to approaches focused on effectiveness, while also yielding results that are on par with methods targeting completeness.", + "bbox": [ + 91, + 342, + 482, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 797, + 220, + 813 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-image models (T2I) at a large scale have achieved notable success in producing lifelike images [4, 20, 23, 26, 28, 37]. This progress is not only a result of enhanced algorithms and computing resources but also the extensive datasets gathered from the Internet [30]. Unfortunately,", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "these datasets often contain copyrighted works, artistic creations [14, 31, 32], NSFW content, and personal images, raising ethical and legal concerns [18].", + "bbox": [ + 511, + 310, + 903, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Some publicly available models, like Stable Diffusion (SD) [24], trained on vast web-scraped datasets, can generate inappropriate content, including copyrighted images, offensive material, and deepfakes, raising serious community concerns. To address this, researchers have attempted to filter out such content, resulting in the release of Stable Diffusion 2.0 [25]. However, this version, despite the filtering efforts, delivers lower image quality than its predecessor, with some explicit content still slipping through. Another approach involves post-processing filters, such as NSFW detectors, which flag inappropriate outputs and replace them with blank images. Yet, these safeguards can be easily bypassed once the model is publicly accessible [18].", + "bbox": [ + 509, + 356, + 906, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent research has concentrated on fine-tuning text-to-image (T2I) diffusion models to 'forget' inappropriate content, giving rise to Concept Erased Models (CEMs) as a practical approach to mitigate such issues. The first generation of these models, which we term Effective Concept Erased Models (E-CEMs), demonstrates success in removing inappropriate content while maintaining image quality comparable to the original outputs [2, 7, 16, 18, 19]. However, recent studies have uncovered significant vulnerabilities in these methods [22, 35, 39]. For instance, Unlearn-Diff [39] identifies combinations of existing tokens within current models capable of reconstructing the target concept, effectively bypassing the erasing mechanism. Additionally, CCE [22] leverages Textual Inversion [6] to derive a new token that regenerates the erased concept. Collectively, these findings suggest that E-CEMs fail to robustly eliminate target concepts from the model's latent space, leaving them susceptible to regeneration through adversarial techniques.", + "bbox": [ + 509, + 551, + 905, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The second generation, which we call Robust Concept Erased Models (R-CEMs), addresses vulnerabilities revealed by adversarial attacks. Early approaches, like RACE [15], use an iterative framework based on E-CEM, identifying adversarial tokens and erasing them. AdvUnlearn", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19587", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "[38] improves RACE by adding a prior preservation loss to retain model utility. However, while these methods mitigate UnlearnDiff attacks, they struggle against CCE. This is because CCE generates new tokens, while UnlearnDiff relies on existing token combinations, limiting its versatility compared to CCE's broader search space. We provide a discussion about these two attacking methods in Appendix A. More recent methods, such as STEREO [33], refine this process by identifying and erasing all adversarial tokens detected by CCE simultaneously. While it demonstrates notable success in terms of robustness and resilience against current attack methods, it incurs a significant reduction in image quality. In contrast, DUO [21] first applies an image editing approach to modify inappropriate content, then leverages the Diffusion-DPO framework [36] to assign high scores to appropriate images and low scores to inappropriate ones while also introducing a novel prior preservation loss. This method successfully protects the model from existing attacks while maintaining image quality. However, it is currently limited to erasing NSFW content.", + "bbox": [ + 89, + 90, + 483, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While NSFW content is a common target of concept erasure, narrow concepts are not less important targets. Here, we use the terms \"narrow concepts\" to refer to fine-grained categories or specific instances within, such as \"English Springer\", \"Grumpy Cat\", or \"Elon Musk\". The ability to erase narrow concepts is critical, as they may involve copyright issues or contribute to the spread of misleading information. While DUO is the only robust and effective concept erasure technique, it struggles to handle narrow concepts. We discuss in more detail the reason in the Appendix A. In short, this method relies on an editing approach to erase concepts. For instance, it edits the \"nudity\" concept to \"dressed in\", where the difference between the two is clear. However, for narrow concepts, such as editing \"Grumpy Cat\" to \"Cat\", it is harder to classify these as distinct classes, requiring a finer-grained method to address.", + "bbox": [ + 89, + 397, + 483, + 640 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given the aforementioned challenges, our paper presents complementary work to DUO [21], focusing on achieving a balance between robust and effective erasure for narrow concepts. We first investigate why adversarial tokens can bypass E-CEMs and analyze how STEREO achieves robustness against such attacks. We then introduce our method, Subspace Mapping Concept Erasure (SuMa). Our approach constructs both a target subspace containing the concept to be erased and a reference subspace that is independent yet close to the target subspace by identifying their basis vectors. We eliminate the target subspace by fine-tuning the model to optimize a novel objective function, aiming to project the basis of the target subspace onto the reference subspace. This method is integrated with E-CEM techniques, making the model robust against adversarial attacks while still preserving image quality. As a result, SuMa achieves image quality on par with E-CEMs in", + "bbox": [ + 89, + 643, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/277078d12c1231060298c8585bf5d7f5a46762c3cfaa1c2c78092f860fc90230.jpg", + "image_caption": [ + "AdvUnlearn" + ], + "image_footnote": [], + "bbox": [ + 522, + 89, + 606, + 152 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/be61c490d799cfcb718ca6bf6d9c101fc4368d3b8b2c249bb307993d3fd0fb4e.jpg", + "image_caption": [ + "RACE" + ], + "image_footnote": [], + "bbox": [ + 614, + 89, + 697, + 152 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a8a405c84350eda90f0268bf14bbcfff8e0aba93b3ea9c34af118779fd5b25a6.jpg", + "image_caption": [ + "STEREO", + "Figure 1. Illustration of current C-CEMs methods trained to erase garbage truck against the CCE Attack. We can see that AdvUnlearn and RACE are still vulnerable to the CCE Attack, while STEREO successfully defends against it, though at the cost of image quality. Our method successfully balances Completeness and Effectiveness, changing garbage truck into Semi-trailer truck." + ], + "image_footnote": [], + "bbox": [ + 712, + 89, + 797, + 154 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/335b486836d33d5aa32c3b5238c712da9740f8cfe5c09d4f19349c15e7f93258.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 813, + 90, + 897, + 152 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "terms of FID and CLIP-score, and an Attack Success Rate (ASR) comparable to the state-of-the-art R-CEM method, STEREO. Our main contributions are as follows:", + "bbox": [ + 511, + 276, + 903, + 320 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We analyze why adversarial tokens can bypass E-CEMs and how STEREO remains resilient against these attacks.", + "- We propose the Subspace Mapping Concept Erasure (SuMa) framework, which balances robustness and effectiveness in erasing narrow concepts. Our framework first constructs a target subspace that captures the concept to be erased and a reference subspace that is distinct yet proximate to the target. We then introduce a novel objective function that removes the target concept by mapping the target subspace onto the reference subspace.", + "- Extensive experiments across four tasks—object erasure, celebrity erasure, artistic style erasure, and subclass erasure—demonstrate that SuMa offers the best trade-off between completeness and effectiveness. SuMa completely removes the target concept while preserving model utility. To the best of our knowledge, we are the first to achieve these dual objectives for narrow concepts." + ], + "bbox": [ + 511, + 321, + 903, + 579 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 595, + 653, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Effective Concept Erasure (E-CEMs)", + "text_level": 1, + "bbox": [ + 513, + 619, + 834, + 636 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The first generation of the Concept Erasure Models (CEMs) was designed to fine-tune original models to forget inappropriate content, successfully removing unwanted elements while closely maintaining image quality. Concept Ablation (CA) [16] finetunes the model to redirect the target prompt, e.g., \"grumpy cat\", to be similar as a general term, e.g., \"cat\", by aligning the noise predictions on these prompts. ESD [7] pushed the noise predictions on the target concept away from the original noise predictions by reversing the direction of conditional guidance [11]. SPM [19] focused on preserving non-target concepts during erasure by applying the CA method with added regularization that maintained noise predictions on the closet concept in the CLIP Embedding Space. MACE [18] proposes a LORA-Finetuning [13] method to erase multiple concepts simultaneously. Despite these advancements, recent adversarial attacks [22, 35, 39] successfully recovered the \"erased\" concepts, exposing vul", + "bbox": [ + 511, + 643, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19588", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/165ae80020bdf0b85fa1d0f09a70dfdf69a858ae647698a5420c6fff71bd863a.jpg", + "image_caption": [ + "(a) Cosine similarity of TIs from SD1.4." + ], + "image_footnote": [], + "bbox": [ + 96, + 94, + 356, + 213 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4474b9c98fc92f7b550a34fb29b445e6eb8801e5663dcb6eb19ece8d6db33e42.jpg", + "image_caption": [ + "Figure 2. Distribution of cosine similarity scores among TIs in three cases: from the same SD1.4 but different TI training steps, different erasure levels of CA, and between a TI at erasure level $t + 1$ and the subspace created by the $t$ previous TIs." + ], + "image_footnote": [], + "bbox": [ + 366, + 93, + 630, + 213 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/090dd577857f4242ee2266246ba8b5825c876b252cbc7c8136663f35b22cb4c5.jpg", + "image_caption": [ + "(b) Cosine similarity of TIs at different CA levels.", + "(c) Cosine similarity of TIs with subspace." + ], + "image_footnote": [], + "bbox": [ + 638, + 90, + 875, + 213 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "nerabilities in all these methods and indicating a need for further improvements in concept erasure techniques.", + "bbox": [ + 89, + 282, + 482, + 314 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Circumventing Concept Erasure", + "text_level": 1, + "bbox": [ + 89, + 320, + 377, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Several methods have been proposed to attack E-CEMs [22, 35, 39]. P4D [5] enhances adversarial prompts by ensuring that noise estimates between pre-trained and concept-erased models are similar. UnlearnDiff [39] is similar to P4D but leverages the classification capabilities of diffusion models, removing the reliance on pre-trained models. Ring-A-Bell [35] creates adversarial prompts by guiding a \"safe\" prompt toward an \"unsafe\" one in the text vector space. Among these, CCE [22] is the simplest and most effective, employing Textual Inversion (TI) [6] to derive a new token from the E-CEM, which regenerates the erased concept. Unlike UnlearnDiff, which relies on combinations of existing tokens, CCE searches for a new token in the latent text space, greatly widening its search space. Hence, defending against CCE is far more challenging than against UnlearnDiff.", + "bbox": [ + 88, + 342, + 482, + 569 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Robust Concept Erasure (R-CEMs)", + "text_level": 1, + "bbox": [ + 89, + 575, + 401, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motivated by the need to protect CEMs from adversarial attacks, several methods have emerged as part of the second generation of CEMs [15, 33, 38]. RACE [15] proposes an iterative framework that first identifies adversarial tokens using the UnlearnDiff framework but with faster optimization. It then erases the found tokens using a loss function from ESD, providing protection against UnlearnDiff. However, as shown in Fig. 1, it remains vulnerable to more advanced CCE attacks. AdvUnlearn [38] is similar to RACE, but it introduces a prior preservation loss during the erasing phase to maintain the model's utility. Still, it is susceptible to CCE attacks, as illustrated in Fig. 1. STEREO [33] successfully defends against all existing attack methods by iteratively identifying all Textual Inversions (TIs) from the concept-erased model, and erasing all of them simultaneously from the original model. However, this method leads to a significant degradation in image quality, resulting in an FID nearly double that of E-CEMs. DUO [21] addresses both robustness and effectiveness by leveraging Diffusion-DPO to assign high scores to concept-edited images and", + "bbox": [ + 89, + 598, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "low scores to inexact ones. It also introduces a novel prior preservation loss to preserve the model's utility. However, this approach is currently limited to NSFW concepts.", + "bbox": [ + 511, + 282, + 906, + 329 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 342, + 604, + 358 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we begin by exploring whether E-CEMs, trained to erase a single token, can also erase tokens that differ in form but convey the same concept (Sec. 3.1). This analysis sets the stage for understanding why CCE [22] can attack E-CEMs and why STEREO [33] can robustly erase a target concept. We then discuss how to find effective reference subspaces for erasing narrow concepts (Sec. 3.2), as well as the motivation for this subspace mapping approach, rather than simply pushing the target subspace away. Finally, we present our proposed method, the Subspace Mapping Approach for Concept Erasure (SuMa) (Sec. 3.3).", + "bbox": [ + 511, + 368, + 906, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Observations", + "text_level": 1, + "bbox": [ + 511, + 545, + 651, + 560 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We conducted a toy experiment using Concept-Ablation (CA) [16] to investigate whether it can erase tokens that differ in form but convey the same underlying concept. We began by applying Textual Inversion (TI) on SD1.4 for 500 steps using five concepts reported in CA [16]. For each concept, we collected learned tokens from steps 300 to 500, denoted as set $\\mathcal{A}$ . At this stage, all tokens in $\\mathcal{A}$ could generate the target concept. We then trained CA to erase one randomly selected token from $\\mathcal{A}$ and observed that the remaining tokens, though differing in form, were also erased.", + "bbox": [ + 511, + 568, + 906, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Next, we calculated the cosine similarity between the erased token and all tokens from $\\mathcal{A}$ , resulting in a distribution shown in Fig. 2a. We noted that the cosine similarity values ranged from 0.6 to 1.00. We proceeded by running TI on the CA-based concept-erased model above. Upon reaching step 300, TI successfully recovered the erased concept. We then computed the cosine similarity between the new TI tokens and all tokens from $\\mathcal{A}$ again and showed the result in Fig. 2b. At this time, the cosine similarity ranged between 0.3 and 0.4, significantly lower than that of the previously erased token. Based on these observations, we hypothesize that we can't represent the target concept as a single", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19589", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9c1ae6d555ebb5978227a619110cd1478e986a8e0afc46c43245971476aecd9f.jpg", + "image_caption": [ + "(a) Subspace Construction: Following the STEREO approach, we iteratively remove the Target Concept from model $\\theta$ using method $\\mathcal{M}$ , then run Textual Inversion (TI) and extract the corresponding tokens. At each iteration, we take the token from the final TI step and one from an early step to form the basis of the Target Subspace and Reference Subspace, respectively." + ], + "image_footnote": [], + "bbox": [ + 96, + 114, + 483, + 233 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/57a7c040693777596ac41c14a2b2c630eecd0a71c97341a5d11a4ea10b8307bf.jpg", + "image_caption": [ + "(b) Subspace Elimination: In this stage, our objective is to map the Target Subspace $\\mathcal{U}$ onto the Reference Subspace $\\mathcal{V}$ , and to maintain the Reference Subspace in its fine-tuned state $\\mathcal{V}'$ unchanged. We let $\\mathcal{T}$ represent the Text-encoder and $L_{sub}$ denote the cost of mapping $\\mathcal{U}$ to $\\mathcal{V}$ and $\\mathcal{V}'$ to $\\mathcal{V}$ . The parameter $W_k'$ is initialized as $W_k$ and will be updated during fine-tuning.", + "Figure 3. Our proposed framework consists of two main stages. The first stage, Subspace Construction, is used to build Target Subspace and Reference Subspace. The second stage, Subspace Elimination, aims to eliminate the Target Subspace." + ], + "image_footnote": [], + "bbox": [ + 501, + 114, + 893, + 233 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "token but rather as a subspace; erasing a concept could be viewed as eliminating a subspace, and if the cosine similarity between this subspace and tokens that produce a similar concept exceeds 0.6, those tokens are erased along with the target subspace.", + "bbox": [ + 88, + 351, + 482, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To verify this hypothesis, we follow the procedure in STEREO [33] by iteratively erasing tokens found by TI using CA. The process is illustrated in Fig. 3a. For each iteration, we collect the token from the final training step of TI. We aggregate tokens over $n$ iterations, denoted as $x_{1},x_{2},\\ldots ,x_{n}\\in \\mathbb{R}^{d}$ , ensuring that each token $x_{i}$ for $i\\in [1,n]$ can generate the target concept. We then construct the matrix $\\mathbf{S}_t = [x_1,\\dots,x_t]\\in \\mathbb{R}^{d\\times t}$ for $t\\in [1,n - 1]$ . Next, we compute the angle between each token $x_{j}$ for $j\\in [t + 1,n]$ and the column space of $\\mathbf{S}_t$ , denoted by $c_{tj}$ , using the formula:", + "bbox": [ + 89, + 426, + 483, + 592 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nc _ {t j} = \\cos \\left(x _ {j}, \\mathbf {S} _ {t} \\left(\\mathbf {S} _ {t} ^ {T} \\mathbf {S} _ {t}\\right) ^ {- 1} \\mathbf {S} _ {t} ^ {T} x _ {j}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 606, + 482, + 632 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{S}_t\\left(\\mathbf{S}_t^T\\mathbf{S}_t^T\\right)^{-1}\\mathbf{S}_t^T$ is the projection matrix onto the column space of $\\mathbf{S}_t$ . We compute the average value of $c_{tj}$ for different values of $j$ , denoted as $c_{t}$ , and report them in Fig. 2c. Notably, when $t \\geq 3$ , $c_{t}$ exceeds 0.6—the threshold we hypothesize is sufficient to cover the target concept. Then, we employ the second stage of STEREO to erase two sets of tokens: $\\mathcal{S}_1 = \\{x_1, x_2, x_3\\}$ and $\\mathcal{S}_2 = \\{x_1, x_2\\}$ . We calculated the average Attack Success Rate (ASR) across all concepts, yielding $\\mathcal{S}_1$ at $3.21\\%$ and $\\mathcal{S}_2$ at $85.32\\%$ . This indicates STEREO effectively defends against CCE when trained to erase $\\mathcal{S}_1$ , but not when only $\\mathcal{S}_2$ is erased. An additional experiment on token similarity's impact on erasure is detailed in Appendix B.", + "bbox": [ + 88, + 641, + 482, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We found that the adversarial tokens identified by TI encompass those detected by UnlearnDiff. Specifically, we evaluated a model trained to erase $S_{1}$ using UnlearnDiff, which yielded an average Attack Success Rate of $1.52\\%$ , in", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "dicating robustness against UnlearnDiff. To explain this result, we provide an additional justification. Since UnlearnDiff identifies combinations of existing tokens, we leveraged the attention mechanism of the text encoder by using the $\\langle \\mathrm{EOT} \\rangle$ token to represent the target concept, and applied the same technique to all tokens from $S_{1}$ . After this processing, we computed a cosine similarity of 0.78 between the subspace formed by the $\\langle \\mathrm{EOT} \\rangle$ token from $S_{1}$ and the one identified by UnlearnDiff, exceeding 0.6 and supporting our observations.", + "bbox": [ + 511, + 351, + 903, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In summary, our analysis explains why STEREO can robustly remove the target concept. By applying TI on the concept-erased model iteratively in at least 3 iterations, it can collect a sufficient set of tokens spanning the target concept space. When erasing these tokens together, it successfully eradicates the model knowledge of the concept.", + "bbox": [ + 511, + 503, + 905, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Reference Subspace Selection for Narrow Concepts", + "text_level": 1, + "bbox": [ + 511, + 604, + 903, + 637 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CA [16] and ESD [7] were the first methods proposed for concept erasure, but they present contradictory approaches. CA maps noise conditioned on the target concept to a more general representation, while ESD pushes the target noise as far as possible from the original. Studies [18, 33] show that CA outperforms ESD in terms of image quality. It is because mapping approach offers a clear direction to erase target concepts, whereas pushing approach can inadvertently affect unrelated concepts. The second stage of STEREO [33], also a pushing approach, significantly degrades image quality. With this motivation, our paper proposes a mapping approach that erases tokens by mapping them to nearby non-target concept subspaces, preserving model utility. The key challenge is how to automatically select these reference subspaces. We also provide experimental results for Subspace Pushing, where the target subspace is simply pushed as far as possible from the original one, in Appendix C.", + "bbox": [ + 511, + 643, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19590", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/44304fbfbecb48842e4507ad5ca7d6b8bc7fa96598f2988d7804eb1552cc5a04.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ConceptTI Early TokenTextual Token
R2D2 Robot0.641.23
Garbage Truck1.051.28
Grumpy Cat0.981.22
English Springer Dog0.720.96
", + "bbox": [ + 94, + 88, + 475, + 180 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. The Euclid distance of the target token to the tokens from early TI training steps and the textual one.", + "bbox": [ + 89, + 184, + 482, + 213 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We observe that constructing the reference subspace using tokens from the early training steps of TI is effective for narrow concepts, as early tokens typically converge to more general forms (e.g., grumpy cat to cat, or R2D2 robot to robot; see Section 4.3 for details). Moreover, these tokens are closer to the target token learned from TI because they align along the same gradient direction. To validate this, we compare the Euclidean distance between the target token and early TI training tokens, as well as the target token and the textual token, as shown in Table 1. The results indicate that while early tokens may represent concepts similar to the textual token, their distance from the target token is much smaller. This suggests that mapping the target token to an earlier token minimizes weight modifications, thereby better preserving the model's utility.", + "bbox": [ + 89, + 227, + 483, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Proposed Method", + "text_level": 1, + "bbox": [ + 89, + 463, + 264, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our framework, $SuMa$ , consists of two stages, as illustrated in Fig. 3. The first stage, termed Subspace Construction, aims to identify Textual Inversion tokens to establish both sets of Target Subspace $\\{\\mathcal{U}^i\\}$ and Reference Subspace $\\{\\mathcal{V}^i\\}$ . Here, $\\mathcal{U}^i$ represents $i$ -th subspace (corresponding to $i$ -th Unet layer) of the concept to be erased, while $\\mathcal{V}^i$ is $i$ -th the subspace into which we intend to map our $i$ -th target subspace. In the second stage, Subspace Elimination, we refine the to- $K$ and to- $V$ weights of the Cross-Attention Layer in U-Net, denoted as $W_{k}$ and $W_{v}$ , allowing us to project the basis of $\\mathcal{U}^i$ onto $\\mathcal{V}^i$ . Due to the identical functionality of $W_{k}$ and $W_{v}$ in our work, we denote both as $W_{k}$ for simplicity and readability.", + "bbox": [ + 89, + 484, + 483, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Subspace Construction. Starting with the original model, denoted as $\\theta$ , and a concept to be removed, $c$ , we apply a Concept Ablation $(\\mathcal{M})$ to eliminate $c$ from $\\theta$ , resulting in a modified model $\\theta'$ . This process is repeated over $l$ steps. At $j$ -th iteration, we apply Textual Inversion (TI) for the concept to be erased from $\\theta'$ , yielding two tokens: one from the final step of TI training, denoted as $\\langle u\\rangle_{j}$ , and another from an earlier step, denoted as $\\langle v\\rangle_{j}$ . Subsequently, we construct prompts in the formats \"A photo of $\\langle u\\rangle_{j}$ \" and \"A photo of $\\langle v\\rangle_{j}$ \", which are then passed through the Text Encoder $\\mathcal{T}$ to extract embeddings for the tokens $\\langle u_j\\rangle$ and $\\langle v_j\\rangle$ . We then apply the transformation $W_k^i$ , where $i$ represent $i$ -th layer in Unet, to these extracted embeddings, treating the outputs as $\\langle u'\\rangle_{j}^{i}$ and $\\langle v'\\rangle_{j}^{i}$ as basis vectors for the subspaces $\\mathcal{U}^i$ and $\\mathcal{V}^i$", + "bbox": [ + 89, + 688, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "respectively. Formally, this process is defined as follows:", + "bbox": [ + 511, + 90, + 888, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\langle u ^ {\\prime} \\rangle_ {j} ^ {i} = W _ {k} ^ {i} \\mathcal {T} (p + \\langle u \\rangle_ {j}), \\tag {2} \\\\ \\langle v ^ {\\prime} \\rangle_ {j} ^ {i} = W _ {k} ^ {i} \\mathcal {T} (p + \\langle v \\rangle_ {j}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 111, + 903, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p$ represents the prompt template. Finally, we fine-tune $\\theta'$ using $\\mathcal{M}$ to remove $\\langle u_j\\rangle$ , resulting in an updated model $\\theta'$ at the end of each iteration (Fig. 3a).", + "bbox": [ + 511, + 154, + 905, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Subspace Elimination. Fig. 3b illustrates this process. After constructing sets of the subspaces $\\{\\mathcal{U}^i\\}_{i=1}^M$ and $\\{\\mathcal{V}^i\\}_{i=1}^M$ , with $M$ is number of layer in Unet, our goal is to eliminate all $\\mathcal{U}^i$ by fine-tuning $W_k$ so that the basis vectors of $\\mathcal{U}^i$ are projected onto $\\mathcal{V}^i$ . To construct the projection matrix onto $\\mathcal{V}^i$ , we first define the matrix $\\mathbf{V}^i = [\\langle v' \\rangle_1^i, \\langle v' \\rangle_2^i, \\dots, \\langle v' \\rangle_l^i] \\in \\mathbb{R}^{d \\times l}$ , with $d$ is the dimension of token embedding vector, and the projection matrix $\\mathbf{P}_{\\mathbf{V}^i}$ can then be formulated as:", + "bbox": [ + 511, + 207, + 906, + 342 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {\\mathbf {V} ^ {i}} = \\mathbf {V} ^ {i} \\left(\\left(\\mathbf {V} ^ {i}\\right) ^ {T} \\mathbf {V} ^ {i}\\right) ^ {- 1} \\left(\\mathbf {V} ^ {i}\\right) ^ {T} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 353, + 906, + 382 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We define $L_{proj}$ as the objective to project all basis of $\\mathcal{U}^i$ onto $\\mathcal{V}^i$ by following formula:", + "bbox": [ + 511, + 385, + 905, + 415 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {p r o j} = \\sum_ {i = 1} ^ {M} \\sum_ {j = 1} ^ {l} \\left\\| W _ {k} ^ {i} \\mathcal {T} (p + \\langle u \\rangle_ {j}) - \\mathbf {P} _ {\\mathbf {V} ^ {i}} W _ {k} ^ {i} \\mathcal {T} (p + \\langle u \\rangle_ {j}) \\right\\| _ {2} ^ {2} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 420, + 923, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As per [3], we add a regularization term to keep the Reference Subspace $\\mathcal{V}^i$ unchanged during fine-tuning:", + "bbox": [ + 511, + 477, + 906, + 507 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {r e g} = \\sum_ {i = 1} ^ {M} \\sum_ {j = 1} ^ {l} \\left\\| W _ {k} ^ {i} \\mathcal {T} (p + \\langle v \\rangle_ {j}) - \\mathbf {P} _ {\\mathbf {V} ^ {i}} W _ {k} ^ {i} \\mathcal {T} (p + \\langle v \\rangle_ {j}) \\right\\| \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 523, + 906, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The objective function for subspace mapping, $L_{sub}$ , is then defined as the weighted sum of these two terms:", + "bbox": [ + 511, + 580, + 905, + 609 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s u b} = L _ {p r o j} + \\lambda_ {r e g} L _ {r e g}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 617, + 903, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{reg}$ is a hyperparameter (default 1). If we use $L_{sub}$ to fine-tune $\\theta$ to obtain $\\theta'$ and use CCE to find a new target token, this token still generates the target concept when conditioned on $\\theta'$ , but not on $\\theta$ (see Appendix D). We assume modifying only the cross-attention layer is effective for erasing a token, as in [18], but affects the SDE trajectory when erasing multiple tokens. To mitigate this, we use the CA loss function and discuss our choice in Section 4.3. In short, their objective is:", + "bbox": [ + 511, + 638, + 905, + 773 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {C A}} = \\mathbb {E} _ {z _ {t}, c, t} \\left[ \\| \\Phi^ {\\prime} \\left(z _ {t}, \\mathcal {T} (p + c ^ {\\prime}), t\\right) - \\Phi \\left(z _ {t}, \\mathcal {T} (p + c), t\\right) \\| _ {2} ^ {2} \\right] \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 779, + 913, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Phi'$ is the model to learn, $\\Phi$ is pretrained, $\\mathcal{T}$ is TextEncoder, $p$ is prompt template, $c$ is target concept, and $c'$ is its general term (e.g., grumpy cat vs. cat). Our objective is:", + "bbox": [ + 511, + 811, + 905, + 857 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nW _ {k} ^ {*} = \\underset {W _ {k}} {\\operatorname {a r g m i n}} \\left(L _ {\\text {f i n a l}}\\right) = \\underset {W _ {k}} {\\operatorname {a r g m i n}} \\left(L _ {\\mathrm {C A}} + L _ {\\text {s u b}}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 877, + 903, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19591", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7b4628f46847a4d261e799125cc900f15c76736d7157e4d9f447c7b24cd804ab.jpg", + "image_caption": [ + "(a) Qualitative Comparison across 6 methods for the CCE attack. Both E-CEM methods, CA and MACE, are vulnerable to the CCE attack. STEREO's erasure is too aggressive, leaving only Van Gogh-style elements, with this concept being easier to erase. Our method defends against CCE by mapping the target concept to its reference concept." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 495, + 248 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2aa6741d1bcd3c2a72907200ba9aadbafa3627df1845f08481d8a36ba91da338.jpg", + "image_caption": [ + "(b) Qualitative Comparison across 6 methods for image generation. CA struggles with \"English Springer\" and \"Elon Musk.\" MACE performs well, preserving image quality. RACE shows slight degradation, and STEREO significantly reduces quality. Our method strikes a balance between E-CEMs and C-CEMs in terms of quality.", + "Figure 4. Qualitative Results. We present qualitative results comparing five methods against the CCE attack and textual erasure." + ], + "image_footnote": [], + "bbox": [ + 500, + 88, + 901, + 250 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 349, + 223, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setting", + "text_level": 1, + "bbox": [ + 89, + 373, + 290, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. We evaluate SuMa across four tasks: subclass erasure, celebrity erasure, artistic style erasure, and instance erasure. For subclass erasure, we focus on a subset of ImageNette [12] known for its distinctive features. Three subclasses are selected for the test, including English Springer Dog, Garbage Truck, and Golf Ball. For celebrity erasure, we select 3 celebrities in the Celebrity-1000 dataset [34], including David Beckham, Elon Musk, and Adam Lambert. For artistic style erasure, following the approach in [7, 15, 33], we focus on the style of Van Gogh, which is both famous and straightforward to verify in outputs. For instance erasure, we conduct experiments on three copyrighted concepts: Grumpy Cat, Mickey Mouse, and the R2D2 robot. Note that our method is specifically designed to address the problem with narrow concepts, so it does not work as well for NSFW concepts. However, as mentioned earlier, our work is complementary to DUO, so it can be easily integrated with this method to completely erase all kinds of concepts, demonstrated in the Appendix E.", + "bbox": [ + 89, + 402, + 483, + 690 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline. We compare SuMa with four baselines: two E-CEMs (CA [16] and MACE [18]) and four C-CEMs (RACE [15], AdvUnlearn (AU) [38], STEREO [33], DUO [21]). For CA, MACE, RACE, and AU, we use their public code and recommended settings to erase concepts. For STEREO and DUO, we re-implement their methods using paper recommendations and verify consistency with reported results.", + "bbox": [ + 89, + 696, + 483, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Circumventing Methods. We evaluate the robustness of our method under CCE [22] and UnlearnDiff (UD) [39]. For CCE, we use a different set of images from the Subspace Construction stage to ensure the model didn't overfit to this image set. For both attack methods, we use publicly available code to generate adversarial prompts.", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. First, to measure effectiveness, for each erased model, we generate 5,000 images using prompts from the MS-COCO validation set. We then evaluate the model quality when generating such non-target prompts using CLIP scores (higher is better) and FID scores (lower is better). To measure completeness, in the CCE attack, for each erased model, we execute the circumventing method to acquire the adversarial token $c_{adv}$ . We then generate 200 images using a template prompt: \"A photo of a $\\{c_{adv}\\}$ .\" For style erasure, we use the template \"A painting in the style of $\\{c_{adv}\\}$ \". In the UD attack, we execute its code on our erased model to get the adversarial prompt and pass the prompt to the erased model. We report the attack success rate (ASR), where a lower score indicates better performance, following the method in [15, 33]. For subclass erasure tasks, we utilize ResNet-50 [10], pretrained on ImageNet [27], to measure ASR. For celebrity detection, we use the GIPHY Celebrity Detector (GCD) [9] as the classifier. To evaluate artistic style, we follow the approach in [15] by fine-tuning a ViT-based model [1] on the WikiArt dataset [29]. Finally, for instance detection, due to the lack of pretrained models or datasets, we use LLAVA-7B [17] with a template prompt: \"Does this image contain {object}? Answer Yes or No.\" for evaluation.", + "bbox": [ + 511, + 349, + 906, + 714 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We use SD1.4 and SD1.5 as the pretrained model. In all experiments, we fine-tune all cross-attention layers. In the Subspace Construction stage, we use the same setting as CA [16] for concept erasion. Training TI is set for 500 steps, with the early step to get token $\\langle v\\rangle$ set to 50, the initial token is toy, with Adam optimization with learning rate, is 2e-6. For the Subspace Elimination stage, we fine-tune most objects for 500 steps, except for subclass objects, which require 750 steps to ensure model safety. We use the Adam optimizer with a learning rate of 2e-5. All experiments are done on 2 NVIDIA A100 GPUs.", + "bbox": [ + 511, + 734, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19592", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/538436298f35c4ef5ee1862de3032cc5402c846796676cdf6dfb09eee177012a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCCE↓UD↓FID↓CLIP↑
Subclass
SD [24]0.84 / 0.830.74 / 0.7617.04 / 16.950.33 / 0.32
CA [16]0.78 / 0.750.67 / 0.6219.28 / 19.260.32 / 0.32
MACE [18]0.86 / 0.840.73 / 0.6817.15 / 16.980.29 / 0.31
RACE [15]0.74 / 0.730.17 / 0.1826.34 / 26.220.27 / 0.28
AdvUnlearn [38]0.76 / 0.740.18 / 0.1618.47 / 18.390.30 / 0.31
DUO [21]0.65 / 0.630.62 / 0.6517.09 / 16.990.29 / 0.30
STEREO [33]0.05 / 0.030.05 / 0.0427.33 / 27.200.30 / 0.28
Ours0.12 / 0.100.08 / 0.0618.33 / 18.150.31 / 0.30
Identity
SD [24]0.91 / 0.900.86 / 0.8417.04 / 16.950.33 / 0.32
CA [16]0.87 / 0.890.79 / 0.7618.21 / 18.270.31 / 0.30
MACE [18]0.89 / 0.870.65 / 0.6316.74 / 16.690.29 / 0.28
RACE [15]0.78 / 0.790.47 / 0.4624.32 / 24.170.27 / 0.29
AdvUnlearn [38]0.77 / 0.690.52 / 0.5317.53 / 17.690.31 / 0.31
DUO [21]0.73 / 0.710.67 / 0.7317.32 / 17.110.29 / 0.30
STEREO [33]0.02 / 0.010.03 / 0.0126.33 / 26.150.29 / 0.28
Ours0.09 / 0.070.18 / 0.1417.94 / 17.890.31 / 0.29
", + "bbox": [ + 119, + 88, + 504, + 364 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/60148edea511fdaae62eaae61df1bc24bf2884d8c38b03e4ab182cb65a4ef554.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCCE↓UD↓FID↓CLIP↑
Artistic
SD [24]0.86 / 0.830.82 / 0.8517.04 / 16.950.33 / 0.32
CA [16]0.79 / 0.780.74 / 0.7618.02 / 17.850.34 / 0.31
MACE [18]0.84 / 0.820.79 / 0.8316.57 / 16.580.32 / 0.33
RACE [15]0.82 / 0.790.06 / 0.0421.32 / 21.100.30 / 0.31
AdvUnlearn [38]0.83 / 0.760.03 / 0.0517.73 / 17.600.32 / 0.33
DUO [21]0.74 / 0.720.67 / 0.7317.74 / 17.890.30 / 0.31
STEREO [33]0.00 / 0.020.00 / 0.0128.04 / 27.900.30 / 0.31
Ours0.05 / 0.040.07 / 0.0517.18 / 17.210.33 / 0.31
Instance
SD [24]0.97 / 0.940.94 / 0.9317.04 / 16.950.33 / 0.32
CA [16]0.97 / 0.930.92 / 0.8918.43 / 18.300.30 / 0.31
MACE [18]0.97 / 0.920.93 / 0.8916.93 / 16.800.30 / 0.31
RACE [15]0.95 / 0.920.06 / 0.0724.32 / 24.150.28 / 0.29
AdvUnlearn [38]0.95 / 0.930.03 / 0.0418.37 / 18.200.32 / 0.33
DUO [21]0.92 / 0.910.89 / 0.8718.31 / 18.920.30 / 0.31
STEREO [33]0.03 / 0.020.05 / 0.0327.04 / 26.900.31 / 0.32
Ours0.17 / 0.150.19 / 0.1622.34 / 22.200.29 / 0.30
", + "bbox": [ + 509, + 88, + 893, + 362 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f7d4025e6d31d7eb2959699f67836833cbf15beb927ea1d48be36fba8eeae83a.jpg", + "table_caption": [ + "Table 2. Evaluation of Erasing across 4 Categories. We average the results of all target concepts in each category across three metrics and compare our framework, SuMa, with state-of-the-art methods. For ASRs, successful attacks (ASR ≥ 20%) are in red, and defeated attacks (ASR < 20%) in green. The data in each cell show the result of erasing a concept from pretrained SD1.4 / SD1.5." + ], + "table_footnote": [], + "table_body": "
Reference SubspaceSubclassIdentityArtisticInstance
ASR ↓FID ↓CLIP ↑ASR ↓FID ↓CLIP ↑ASR ↓FID ↓CLIP ↑ASR ↓FID ↓CLIP ↑
Step-300.0919.110.300.0818.340.290.0518.820.310.1723.960.27
Step-500.1218.330.310.0917.940.310.0517.180.330.1722.340.29
Step-700.5617.320.330.6717.920.320.3717.050.330.8317.740.32
General Term0.1818.270.320.1718.020.320.1118.210.330.1523.340.27
", + "bbox": [ + 114, + 414, + 879, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation Study for the selection of Reference subspace.", + "bbox": [ + 299, + 531, + 694, + 545 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dd3e639ca00e25b5659003905736140f3dce21d76f1ff1da0abb65506adfa83a.jpg", + "image_caption": [ + "Figure 5. Visualization of Textual Inversion tokens at different numbers of TI training steps." + ], + "image_footnote": [], + "bbox": [ + 127, + 554, + 442, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Results", + "text_level": 1, + "bbox": [ + 89, + 821, + 184, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Results. We average each metric from Section 4.1 for each category and report the results in Table 2. For all baselines, we ensure the erased model works", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "well for erasing the textual prompt, and we report results for this trivial task in the Appendix F. For completeness evaluation, our method achieves an ASR score below 0.2 for all categories under both CCE and UD attacks. Compared to STEREO, our method has a slightly higher ASR score (0.02 to 0.14 higher), but within an acceptable range. RACE and AdvUnlearn show safety against the UD attack in the Subclass, Artistic, and Instance categories but fail under Identity. This is due to the subtle facial features of human identity, which make this category more fine-grained and challenging. However, both fail under the CCE attack, showing that protecting the model against CCE is harder than UD. The main reason is that RACE and AdvUnlearn erase adversarial tokens based on UnlearnDiff, but these tokens don't fully represent the target concept, protecting against UD but not CCE. DUO works well for NSFW concepts (results in Appendix E) but fails for narrow concepts. For effectiveness, although STEREO performs well under both CCE and UD, image quality is significantly affected, resulting in an FID score roughly double that of other E-CEMs, making it less effective. Our method, while not achieving a perfect ASR, achieves an FID score comparable", + "bbox": [ + 511, + 558, + 906, + 891 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19593", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1b0defb8e6523aef785ed64195ffacfaaa9f6ad964b3665f3d56b629920eb3b1.jpg", + "image_caption": [ + "Figure 6. Multiple Concept Erasure Result." + ], + "image_footnote": [], + "bbox": [ + 99, + 93, + 468, + 231 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6492bd105a2b98f4d13a6a37d889f3e4881e61be8f03f46250388bd3f68d1f48.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
RegulizationSubclassIdentityArtisticInstance
×ASR: 0.09ASR: 0.06ASR: 0.05ASR: 0.15
FID: 19.21FID: 20.31FID: 19.32FID: 24.57
CLIP: 0.29CLIP: 0.29CLIP: 0.30CLIP: 0.26
✓ASR: 0.12ASR: 0.09ASR: 0.05ASR: 0.17
FID: 18.33FID: 17.94FID: 17.18FID: 22.34
CLIP: 0.31CLIP: 0.31CLIP: 0.33CLIP: 0.29
", + "bbox": [ + 94, + 252, + 478, + 353 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/786c489e395da048f0a2275ee82609085de63c5b21ce6803d69d75934a5920db.jpg", + "table_caption": [ + "Table 4. Ablation Study for the need of Subspace Regulization." + ], + "table_footnote": [], + "table_body": "
EntityFID ↓CCE ↓UD ↓
English Springer20.12 / 18.140.12 / 0.470.07 / 0.34
Van Gogh20.16 / 17.540.07 / 0.520.04 / 0.46
Elon Musk20.15 / 18.220.07 / 0.490.05 / 0.57
Grumpy Cat24.57 / 20.320.18 / 0.540.14 / 0.46
", + "bbox": [ + 93, + 378, + 478, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation Study for the base method (ESD / UCE).", + "bbox": [ + 102, + 468, + 470, + 482 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to other E-CEMs in the Subclass, Artistic, and Identity categories, outperforming CA with an FID around 0.85 lower. We found that Instance category is harder to be erased, so we fine-tuned it for 750 steps, resulting in a higher FID compared to the others, fine-tuned for 500 steps.", + "bbox": [ + 89, + 503, + 482, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Results are provided in Fig. 4. Fig. 4a shows results of these models against the CCE attack, and we include qualitative results for UD in the Appendix J. CA, MACE, RACE, and AdvUnlearn are highly vulnerable to this attack. STEREO completely removes these concepts, but at the cost of image quality ( $6^{th}$ column in Fig. 4b). Our method successfully converts \"Springer Dog\" to \"Dog,\" \"Elon Musk\" to a general person, and \"Grumpy Cat\" to \"Cat toy,\" while maintaining image quality comparable to other E-CEM methods (last column in Fig. 4b).", + "bbox": [ + 89, + 588, + 482, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multiple Concepts Erasure. We provide results for erasing multiple concepts in Fig. 6. We test the ASR with the CCE attack and evaluate two settings: Iterative Erase (IE) and Simultaneous Erase (SE). IE means we fine-tune the weights after erasing one concept to erase others. SE sums all $L_{sub}$ to erase all concepts in one fine-tuning, with 1000 steps fine-tuning for this experiment. In both settings, our method defends against CCE and maintains model utility. With SE, we achieve better FID scores, as summing $L_{sub}$ preserves a broader range of knowledge from $L_{reg}$ .", + "bbox": [ + 89, + 750, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 513, + 90, + 679, + 106 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of selecting the Reference Subspace. We present results for four ways of constructing the Reference Subspace, including TI tokens from Steps 30, 50, 70 of TI training, and the concept's general term, as shown in Table 3. Visualizations of each token are in Fig. 5. At Step 30, the token is far from the target, with little relevant information. At Step 50, the token shares features with the target but is not an exact match, enabling minimal transformation. At Step 70, the token is close to the target, making full erasure difficult. These observations align with Table 3, where Step 50 balances ASR, FID, and CLIP scores, while Step 30 improves ASR but slightly reduces image quality. Step 70 shows results similar to CA, indicating SuMa's inefficacy when the Reference Subspace and Target Subspace are too close.", + "bbox": [ + 511, + 119, + 906, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of choosing the base method. As noted in Section 3.3, we adopt CA loss for our method. Results using ESD [7] and UCE [8] as base methods are shown in Table 5. While ESD resembles CA, it uses a pushing strategy, whereas CA employs a mapping strategy. UCE erases concepts by altering the cross-attention map. Our experiments show that modifying only the attention layer causes the token learned by CCE to express the target concept only under the erased model as shown in Appendix D. ESD provides some defense but increases FID moderately, while UCE fails to defend. We conclude that modifying the model's noise prediction is more effective, as it redirects the target concept's trajectory toward another in the original model. The mapping approach is also favored over the pushing one, as discussed in Section 3.2 and supported by our results.", + "bbox": [ + 511, + 338, + 908, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of adding Subspace Regularization. We present the effect of adding Subspace Regularization and report the results in Table 4. We observe that without regularization, our method performs slightly better against the CCE attack, with an average $2\\%$ reduction in attack rate, but at the cost of a gradual increase in the FID score. This phenomenon was studied in [3], where it was shown that the concept closest to the target is the most affected. Therefore, we conclude that adding Subspace Regularization loss is necessary to maintain the model's effectiveness.", + "bbox": [ + 511, + 573, + 908, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Limitations", + "text_level": 1, + "bbox": [ + 511, + 739, + 769, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduced SuMa, a method that robustly and effectively erases narrow concepts in Text-to-Image Diffusion models. SuMa constructs a Target Subspace for the concept to be erased and maps it to a Reference Subspace. Experimental results show our method is the first to meet these dual objectives for narrow concepts. A key limitation is its time cost, as it may require many iterations to construct both subspaces. Future work will aim to accelerate this process and improve target concept representation.", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19594", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Dosovitskiy Alexey. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv: 2010.11929, 2020. 6", + "[2] Anh Bui, Khanh Doan, Trung Le, Paul Montague, Tamas Abraham, and Dinh Phung. Removing undesirable concepts in text-to-image generative models with learnable prompts. arXiv preprint arXiv:2403.12326, 2024. 1", + "[3] Anh Bui, Long Vuong, Khanh Doan, Trung Le, Paul Montague, Tamas Abraham, and Dinh Phung. Erasing undesirable concepts in diffusion models with adversarial preservation. arXiv preprint arXiv:2410.15618, 2024. 5, 8", + "[4] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, José Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. In Proceedings of the 40th International Conference on Machine Learning, pages 4055-4075, 2023. 1", + "[5] Zhi-Yi Chin, Chieh-Ming Jiang, Ching-Chun Huang, PinYu Chen, and Wei-Chen Chiu. Prompting4debugging: Redteaming text-to-image diffusion models by finding problematic prompts. arXiv preprint arXiv:2309.06135, 2023. 3", + "[6] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel CohenOr. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 1, 3", + "[7] Rohit Gandikota, Joanna Materzynska, Jaden Fiitto-Kaufman, and David Bau. Erasing concepts from diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2426-2436, 2023. 1, 2, 4, 6, 8", + "[8] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5111-5120, 2024. 8", + "[9] Nick Hasty, Ihor Kroosh, Dmitry Voitekh, and Dmytro Korduban. Giphy celebrity detector, 2022. 6", + "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6", + "[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2", + "[12] Jeremy Howard and Sylvain Gugger. Fastai: a layered api for deep learning. Information, 11(2):108, 2020. 6", + "[13] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2", + "[14] Harry H Jiang, Lauren Brown, Jessica Cheng, Mehtab Khan, Abhishek Gupta, Deja Workman, Alex Hanna, Johnathan Flowers, and Timnit Gebru. Ai art and its impact on artists. In Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society, pages 363-374, 2023. 1" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Changhoon Kim, Kyle Min, and Yezhou Yang. Race: Robust adversarial concept erasure for secure text-to-image diffusion model. arXiv preprint arXiv:2405.16341, 2024. 1, 3, 6, 7, 4, 5", + "[16] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22691-22702, 2023. 1, 2, 3, 4, 6, 7, 5", + "[17] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 6", + "[18] Shilin Lu, Zilan Wang, Leyang Li, Yanzhu Liu, and Adams Wai-Kin Kong. Mace: Mass concept erasure in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6430-6440, 2024. 1, 2, 4, 5, 6, 7", + "[19] Mengyao Lyu, Yuhong Yang, Haiwen Hong, Hui Chen, Xuan Jin, Yuan He, Hui Xue, Jungong Han, and Guiguang Ding. One-dimensional adapter to rule them all: Concepts diffusion models and erasing applications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7559-7568, 2024. 1, 2", + "[20] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 1", + "[21] Yong-Hyun Park, Sangdoo Yun, Jin-Hwa Kim, Junho Kim, Geonhui Jang, Yonghyun Jeong, Junghyo Jo, and Gayoung Lee. Direct unlearning optimization for robust and safe text-to-image models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2, 3, 6, 7, 4, 5", + "[22] Minh Pham, Kelly O Marshall, Niv Cohen, Govind Mittal, and Chinmay Hegde. Circumventing concept erasure methods for text-to-image generative models. In The Twelfth International Conference on Learning Representations, 2023. 1, 2, 3, 6", + "[23] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1", + "[24] Robin Rombach. Stable diffusion v1-4 model card., 2022. 1, 7, 4, 5", + "[25] Robin Rombach. Stable diffusion 2.0 release., 2022. 1", + "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1", + "[27] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 6" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19595", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1", + "[29] Babak Saleh and Ahmed Elgammal. Large-scale classification of fine-art paintings: Learning the right metric on the right feature. arxiv 2015. arXiv preprint arXiv:1505.00855, 2015.6", + "[30] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 1", + "[31] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by {Text-to-Image} models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 2187–2204, 2023. 1", + "[32] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6048-6058, 2023. 1", + "[33] Koushik Srivatsan, Fahad Shamshad, Muzammal Naseer, Vishal M Patel, and Karthik Nandakumar. Stereo: A two-stage framework for adversarially robust concept erasing from text-to-image diffusion models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 23765-23774, 2025. 2, 3, 4, 6, 7, 5", + "[34] Tonyassi. Celebrity-1000 datasets, 2022. 6", + "[35] Yu-Lin Tsai, Chia-Yi Hsu, Chulin Xie, Chih-Hsun Lin, Jia-You Chen, Bo Li, Pin-Yu Chen, Chia-Mu Yu, and Chun-Ying Huang. Ring-a-bell! how reliable are concept removal methods for diffusion models? arXiv preprint arXiv:2310.10012, 2023. 1, 2, 3", + "[36] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228-8238, 2024. 2", + "[37] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research, 2021. 1", + "[38] Yimeng Zhang, Xin Chen, Jinghan Jia, Yihua Zhang, Chongyu Fan, Jiancheng Liu, Mingyi Hong, Ke Ding, and Sijia Liu. Defensive unlearning with adversarial training for robust concept erasure in diffusion models. arXiv preprint arXiv:2405.15234, 2024. 2, 3, 6, 7, 4, 5", + "[39] Yimeng Zhang, Jinghan Jia, Xin Chen, Aochuan Chen, Yihua Zhang, Jiancheng Liu, Ke Ding, and Sijia Liu. To generate or not? safety-driven unlearned diffusion models are still" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "easy to generate unsafe images... for now. In European Conference on Computer Vision, pages 385-403. Springer, 2025. 1, 2, 3, 6", + "bbox": [ + 545, + 92, + 903, + 133 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19596", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SuMa_ A Subspace Mapping Approach for Robust and Effective Concept Erasure in Text-to-Image Diffusion Models/51080fa1-f759-4df5-8927-63f76b232e8f_model.json b/2025/SuMa_ A Subspace Mapping Approach for Robust and Effective Concept Erasure in Text-to-Image Diffusion Models/51080fa1-f759-4df5-8927-63f76b232e8f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5434272dd24b8953ecdecda6ee220b252fb687aa --- /dev/null +++ b/2025/SuMa_ A Subspace Mapping Approach for Robust and Effective Concept Erasure in Text-to-Image Diffusion Models/51080fa1-f759-4df5-8927-63f76b232e8f_model.json @@ -0,0 +1,1991 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.13, + 0.13, + 0.87, + 0.175 + ], + "angle": 0, + "content": "SuMa: A Subspace Mapping Approach for Robust and Effective Concept Erasure in Text-to-Image Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.203, + 0.71, + 0.221 + ], + "angle": 0, + "content": "Kien Nguyen1 Anh Tran1 Cuong Pham2,1" + }, + { + "type": "text", + "bbox": [ + 0.399, + 0.222, + 0.599, + 0.239 + ], + "angle": 0, + "content": "\\(^{1}\\)Qualcomm AI Research" + }, + { + "type": "text", + "bbox": [ + 0.25, + 0.24, + 0.748, + 0.257 + ], + "angle": 0, + "content": "2Posts and Telecommunications Institute of Technology (PTIT)" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.26, + 0.673, + 0.275 + ], + "angle": 0, + "content": "{kienn, anhtra, pcuong}@qti.qualcomm.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.343, + 0.483, + 0.765 + ], + "angle": 0, + "content": "The rapid growth of text-to-image diffusion models has raised concerns about their potential misuse in generating harmful or unauthorized contents. To address these issues, several Concept Erasure methods have been proposed. However, most of them fail to achieve both robustness, i.e., the ability to robustly remove the target concept, and effectiveness, i.e., maintaining image quality. While few recent techniques successfully achieve these goals for NSFW concepts, none could handle narrow concepts such as copyrighted characters or celebrities. Erasing these narrow concepts is critical in addressing copyright and legal concerns. However, erasing them is challenging due to their close distances to non-target neighboring concepts, requiring finer-grained manipulation. In this paper, we introduce Subspace Mapping (SuMa), a novel method specifically designed to achieve both robustness and effectiveness in easing these narrow concepts. SuMa first derives a target subspace representing the concept to be erased and then neutralizes it by mapping it to a reference subspace that minimizes the distance between the two. This mapping ensures the target concept is robustly erased while preserving image quality. We conduct extensive experiments with SuMa across four tasks: subclass erasure, celebrity erasure, artistic style erasure, and instance erasure and compare the results with current state-of-the-art methods. Our method achieves image quality comparable to approaches focused on effectiveness, while also yielding results that are on par with methods targeting completeness." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.798, + 0.222, + 0.814 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Text-to-image models (T2I) at a large scale have achieved notable success in producing lifelike images [4, 20, 23, 26, 28, 37]. This progress is not only a result of enhanced algorithms and computing resources but also the extensive datasets gathered from the Internet [30]. Unfortunately," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.311, + 0.905, + 0.356 + ], + "angle": 0, + "content": "these datasets often contain copyrighted works, artistic creations [14, 31, 32], NSFW content, and personal images, raising ethical and legal concerns [18]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.357, + 0.907, + 0.553 + ], + "angle": 0, + "content": "Some publicly available models, like Stable Diffusion (SD) [24], trained on vast web-scraped datasets, can generate inappropriate content, including copyrighted images, offensive material, and deepfakes, raising serious community concerns. To address this, researchers have attempted to filter out such content, resulting in the release of Stable Diffusion 2.0 [25]. However, this version, despite the filtering efforts, delivers lower image quality than its predecessor, with some explicit content still slipping through. Another approach involves post-processing filters, such as NSFW detectors, which flag inappropriate outputs and replace them with blank images. Yet, these safeguards can be easily bypassed once the model is publicly accessible [18]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.553, + 0.906, + 0.825 + ], + "angle": 0, + "content": "Recent research has concentrated on fine-tuning text-to-image (T2I) diffusion models to 'forget' inappropriate content, giving rise to Concept Erased Models (CEMs) as a practical approach to mitigate such issues. The first generation of these models, which we term Effective Concept Erased Models (E-CEMs), demonstrates success in removing inappropriate content while maintaining image quality comparable to the original outputs [2, 7, 16, 18, 19]. However, recent studies have uncovered significant vulnerabilities in these methods [22, 35, 39]. For instance, Unlearn-Diff [39] identifies combinations of existing tokens within current models capable of reconstructing the target concept, effectively bypassing the erasing mechanism. Additionally, CCE [22] leverages Textual Inversion [6] to derive a new token that regenerates the erased concept. Collectively, these findings suggest that E-CEMs fail to robustly eliminate target concepts from the model's latent space, leaving them susceptible to regeneration through adversarial techniques." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "The second generation, which we call Robust Concept Erased Models (R-CEMs), addresses vulnerabilities revealed by adversarial attacks. Early approaches, like RACE [15], use an iterative framework based on E-CEM, identifying adversarial tokens and erasing them. AdvUnlearn" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19587" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.395 + ], + "angle": 0, + "content": "[38] improves RACE by adding a prior preservation loss to retain model utility. However, while these methods mitigate UnlearnDiff attacks, they struggle against CCE. This is because CCE generates new tokens, while UnlearnDiff relies on existing token combinations, limiting its versatility compared to CCE's broader search space. We provide a discussion about these two attacking methods in Appendix A. More recent methods, such as STEREO [33], refine this process by identifying and erasing all adversarial tokens detected by CCE simultaneously. While it demonstrates notable success in terms of robustness and resilience against current attack methods, it incurs a significant reduction in image quality. In contrast, DUO [21] first applies an image editing approach to modify inappropriate content, then leverages the Diffusion-DPO framework [36] to assign high scores to appropriate images and low scores to inappropriate ones while also introducing a novel prior preservation loss. This method successfully protects the model from existing attacks while maintaining image quality. However, it is currently limited to erasing NSFW content." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.484, + 0.641 + ], + "angle": 0, + "content": "While NSFW content is a common target of concept erasure, narrow concepts are not less important targets. Here, we use the terms \"narrow concepts\" to refer to fine-grained categories or specific instances within, such as \"English Springer\", \"Grumpy Cat\", or \"Elon Musk\". The ability to erase narrow concepts is critical, as they may involve copyright issues or contribute to the spread of misleading information. While DUO is the only robust and effective concept erasure technique, it struggles to handle narrow concepts. We discuss in more detail the reason in the Appendix A. In short, this method relies on an editing approach to erase concepts. For instance, it edits the \"nudity\" concept to \"dressed in\", where the difference between the two is clear. However, for narrow concepts, such as editing \"Grumpy Cat\" to \"Cat\", it is harder to classify these as distinct classes, requiring a finer-grained method to address." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Given the aforementioned challenges, our paper presents complementary work to DUO [21], focusing on achieving a balance between robust and effective erasure for narrow concepts. We first investigate why adversarial tokens can bypass E-CEMs and analyze how STEREO achieves robustness against such attacks. We then introduce our method, Subspace Mapping Concept Erasure (SuMa). Our approach constructs both a target subspace containing the concept to be erased and a reference subspace that is independent yet close to the target subspace by identifying their basis vectors. We eliminate the target subspace by fine-tuning the model to optimize a novel objective function, aiming to project the basis of the target subspace onto the reference subspace. This method is integrated with E-CEM techniques, making the model robust against adversarial attacks while still preserving image quality. As a result, SuMa achieves image quality on par with E-CEMs in" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.09, + 0.607, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.157, + 0.595, + 0.164 + ], + "angle": 0, + "content": "AdvUnlearn" + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.09, + 0.699, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.642, + 0.157, + 0.668, + 0.164 + ], + "angle": 0, + "content": "RACE" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.09, + 0.798, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.735, + 0.157, + 0.771, + 0.164 + ], + "angle": 0, + "content": "STEREO" + }, + { + "type": "image", + "bbox": [ + 0.815, + 0.091, + 0.898, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.842, + 0.157, + 0.866, + 0.166 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.177, + 0.907, + 0.261 + ], + "angle": 0, + "content": "Figure 1. Illustration of current C-CEMs methods trained to erase garbage truck against the CCE Attack. We can see that AdvUnlearn and RACE are still vulnerable to the CCE Attack, while STEREO successfully defends against it, though at the cost of image quality. Our method successfully balances Completeness and Effectiveness, changing garbage truck into Semi-trailer truck." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.277, + 0.905, + 0.321 + ], + "angle": 0, + "content": "terms of FID and CLIP-score, and an Attack Success Rate (ASR) comparable to the state-of-the-art R-CEM method, STEREO. Our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.323, + 0.905, + 0.351 + ], + "angle": 0, + "content": "- We analyze why adversarial tokens can bypass E-CEMs and how STEREO remains resilient against these attacks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.353, + 0.905, + 0.473 + ], + "angle": 0, + "content": "- We propose the Subspace Mapping Concept Erasure (SuMa) framework, which balances robustness and effectiveness in erasing narrow concepts. Our framework first constructs a target subspace that captures the concept to be erased and a reference subspace that is distinct yet proximate to the target. We then introduce a novel objective function that removes the target concept by mapping the target subspace onto the reference subspace." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.474, + 0.905, + 0.58 + ], + "angle": 0, + "content": "- Extensive experiments across four tasks—object erasure, celebrity erasure, artistic style erasure, and subclass erasure—demonstrate that SuMa offers the best trade-off between completeness and effectiveness. SuMa completely removes the target concept while preserving model utility. To the best of our knowledge, we are the first to achieve these dual objectives for narrow concepts." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.323, + 0.905, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.596, + 0.655, + 0.611 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.621, + 0.836, + 0.637 + ], + "angle": 0, + "content": "2.1. Effective Concept Erasure (E-CEMs)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.903 + ], + "angle": 0, + "content": "The first generation of the Concept Erasure Models (CEMs) was designed to fine-tune original models to forget inappropriate content, successfully removing unwanted elements while closely maintaining image quality. Concept Ablation (CA) [16] finetunes the model to redirect the target prompt, e.g., \"grumpy cat\", to be similar as a general term, e.g., \"cat\", by aligning the noise predictions on these prompts. ESD [7] pushed the noise predictions on the target concept away from the original noise predictions by reversing the direction of conditional guidance [11]. SPM [19] focused on preserving non-target concepts during erasure by applying the CA method with added regularization that maintained noise predictions on the closet concept in the CLIP Embedding Space. MACE [18] proposes a LORA-Finetuning [13] method to erase multiple concepts simultaneously. Despite these advancements, recent adversarial attacks [22, 35, 39] successfully recovered the \"erased\" concepts, exposing vul" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19588" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.095, + 0.357, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.218, + 0.331, + 0.23 + ], + "angle": 0, + "content": "(a) Cosine similarity of TIs from SD1.4." + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.094, + 0.631, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.366, + 0.218, + 0.628, + 0.23 + ], + "angle": 0, + "content": "(b) Cosine similarity of TIs at different CA levels." + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.091, + 0.876, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.658, + 0.218, + 0.882, + 0.23 + ], + "angle": 0, + "content": "(c) Cosine similarity of TIs with subspace." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.235, + 0.908, + 0.264 + ], + "angle": 0, + "content": "Figure 2. Distribution of cosine similarity scores among TIs in three cases: from the same SD1.4 but different TI training steps, different erasure levels of CA, and between a TI at erasure level \\( t + 1 \\) and the subspace created by the \\( t \\) previous TIs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.283, + 0.483, + 0.315 + ], + "angle": 0, + "content": "nerabilities in all these methods and indicating a need for further improvements in concept erasure techniques." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.321, + 0.379, + 0.337 + ], + "angle": 0, + "content": "2.2. Circumventing Concept Erasure" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.343, + 0.483, + 0.57 + ], + "angle": 0, + "content": "Several methods have been proposed to attack E-CEMs [22, 35, 39]. P4D [5] enhances adversarial prompts by ensuring that noise estimates between pre-trained and concept-erased models are similar. UnlearnDiff [39] is similar to P4D but leverages the classification capabilities of diffusion models, removing the reliance on pre-trained models. Ring-A-Bell [35] creates adversarial prompts by guiding a \"safe\" prompt toward an \"unsafe\" one in the text vector space. Among these, CCE [22] is the simplest and most effective, employing Textual Inversion (TI) [6] to derive a new token from the E-CEM, which regenerates the erased concept. Unlike UnlearnDiff, which relies on combinations of existing tokens, CCE searches for a new token in the latent text space, greatly widening its search space. Hence, defending against CCE is far more challenging than against UnlearnDiff." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.577, + 0.402, + 0.593 + ], + "angle": 0, + "content": "2.3. Robust Concept Erasure (R-CEMs)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Motivated by the need to protect CEMs from adversarial attacks, several methods have emerged as part of the second generation of CEMs [15, 33, 38]. RACE [15] proposes an iterative framework that first identifies adversarial tokens using the UnlearnDiff framework but with faster optimization. It then erases the found tokens using a loss function from ESD, providing protection against UnlearnDiff. However, as shown in Fig. 1, it remains vulnerable to more advanced CCE attacks. AdvUnlearn [38] is similar to RACE, but it introduces a prior preservation loss during the erasing phase to maintain the model's utility. Still, it is susceptible to CCE attacks, as illustrated in Fig. 1. STEREO [33] successfully defends against all existing attack methods by iteratively identifying all Textual Inversions (TIs) from the concept-erased model, and erasing all of them simultaneously from the original model. However, this method leads to a significant degradation in image quality, resulting in an FID nearly double that of E-CEMs. DUO [21] addresses both robustness and effectiveness by leveraging Diffusion-DPO to assign high scores to concept-edited images and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.283, + 0.907, + 0.33 + ], + "angle": 0, + "content": "low scores to inexact ones. It also introduces a novel prior preservation loss to preserve the model's utility. However, this approach is currently limited to NSFW concepts." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.343, + 0.606, + 0.359 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.369, + 0.907, + 0.537 + ], + "angle": 0, + "content": "In this section, we begin by exploring whether E-CEMs, trained to erase a single token, can also erase tokens that differ in form but convey the same concept (Sec. 3.1). This analysis sets the stage for understanding why CCE [22] can attack E-CEMs and why STEREO [33] can robustly erase a target concept. We then discuss how to find effective reference subspaces for erasing narrow concepts (Sec. 3.2), as well as the motivation for this subspace mapping approach, rather than simply pushing the target subspace away. Finally, we present our proposed method, the Subspace Mapping Approach for Concept Erasure (SuMa) (Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.546, + 0.652, + 0.561 + ], + "angle": 0, + "content": "3.1. Observations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.719 + ], + "angle": 0, + "content": "We conducted a toy experiment using Concept-Ablation (CA) [16] to investigate whether it can erase tokens that differ in form but convey the same underlying concept. We began by applying Textual Inversion (TI) on SD1.4 for 500 steps using five concepts reported in CA [16]. For each concept, we collected learned tokens from steps 300 to 500, denoted as set \\(\\mathcal{A}\\). At this stage, all tokens in \\(\\mathcal{A}\\) could generate the target concept. We then trained CA to erase one randomly selected token from \\(\\mathcal{A}\\) and observed that the remaining tokens, though differing in form, were also erased." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Next, we calculated the cosine similarity between the erased token and all tokens from \\(\\mathcal{A}\\), resulting in a distribution shown in Fig. 2a. We noted that the cosine similarity values ranged from 0.6 to 1.00. We proceeded by running TI on the CA-based concept-erased model above. Upon reaching step 300, TI successfully recovered the erased concept. We then computed the cosine similarity between the new TI tokens and all tokens from \\(\\mathcal{A}\\) again and showed the result in Fig. 2b. At this time, the cosine similarity ranged between 0.3 and 0.4, significantly lower than that of the previously erased token. Based on these observations, we hypothesize that we can't represent the target concept as a single" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19589" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.116, + 0.485, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.236, + 0.496, + 0.298 + ], + "angle": 0, + "content": "(a) Subspace Construction: Following the STEREO approach, we iteratively remove the Target Concept from model \\(\\theta\\) using method \\(\\mathcal{M}\\), then run Textual Inversion (TI) and extract the corresponding tokens. At each iteration, we take the token from the final TI step and one from an early step to form the basis of the Target Subspace and Reference Subspace, respectively." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.116, + 0.895, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.236, + 0.905, + 0.3 + ], + "angle": 0, + "content": "(b) Subspace Elimination: In this stage, our objective is to map the Target Subspace \\(\\mathcal{U}\\) onto the Reference Subspace \\(\\mathcal{V}\\), and to maintain the Reference Subspace in its fine-tuned state \\(\\mathcal{V}'\\) unchanged. We let \\(\\mathcal{T}\\) represent the Text-encoder and \\(L_{sub}\\) denote the cost of mapping \\(\\mathcal{U}\\) to \\(\\mathcal{V}\\) and \\(\\mathcal{V}'\\) to \\(\\mathcal{V}\\). The parameter \\(W_k'\\) is initialized as \\(W_k\\) and will be updated during fine-tuning." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.309, + 0.908, + 0.339 + ], + "angle": 0, + "content": "Figure 3. Our proposed framework consists of two main stages. The first stage, Subspace Construction, is used to build Target Subspace and Reference Subspace. The second stage, Subspace Elimination, aims to eliminate the Target Subspace." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.352, + 0.483, + 0.427 + ], + "angle": 0, + "content": "token but rather as a subspace; erasing a concept could be viewed as eliminating a subspace, and if the cosine similarity between this subspace and tokens that produce a similar concept exceeds 0.6, those tokens are erased along with the target subspace." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.428, + 0.484, + 0.593 + ], + "angle": 0, + "content": "To verify this hypothesis, we follow the procedure in STEREO [33] by iteratively erasing tokens found by TI using CA. The process is illustrated in Fig. 3a. For each iteration, we collect the token from the final training step of TI. We aggregate tokens over \\( n \\) iterations, denoted as \\( x_{1},x_{2},\\ldots ,x_{n}\\in \\mathbb{R}^{d} \\), ensuring that each token \\( x_{i} \\) for \\( i\\in [1,n] \\) can generate the target concept. We then construct the matrix \\( \\mathbf{S}_t = [x_1,\\dots,x_t]\\in \\mathbb{R}^{d\\times t} \\) for \\( t\\in [1,n - 1] \\). Next, we compute the angle between each token \\( x_{j} \\) for \\( j\\in [t + 1,n] \\) and the column space of \\( \\mathbf{S}_t \\), denoted by \\( c_{tj} \\), using the formula:" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.607, + 0.483, + 0.633 + ], + "angle": 0, + "content": "\\[\nc _ {t j} = \\cos \\left(x _ {j}, \\mathbf {S} _ {t} \\left(\\mathbf {S} _ {t} ^ {T} \\mathbf {S} _ {t}\\right) ^ {- 1} \\mathbf {S} _ {t} ^ {T} x _ {j}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.642, + 0.483, + 0.839 + ], + "angle": 0, + "content": "where \\(\\mathbf{S}_t\\left(\\mathbf{S}_t^T\\mathbf{S}_t^T\\right)^{-1}\\mathbf{S}_t^T\\) is the projection matrix onto the column space of \\(\\mathbf{S}_t\\). We compute the average value of \\(c_{tj}\\) for different values of \\(j\\), denoted as \\(c_{t}\\), and report them in Fig. 2c. Notably, when \\(t \\geq 3\\), \\(c_{t}\\) exceeds 0.6—the threshold we hypothesize is sufficient to cover the target concept. Then, we employ the second stage of STEREO to erase two sets of tokens: \\(\\mathcal{S}_1 = \\{x_1, x_2, x_3\\}\\) and \\(\\mathcal{S}_2 = \\{x_1, x_2\\}\\). We calculated the average Attack Success Rate (ASR) across all concepts, yielding \\(\\mathcal{S}_1\\) at \\(3.21\\%\\) and \\(\\mathcal{S}_2\\) at \\(85.32\\%\\). This indicates STEREO effectively defends against CCE when trained to erase \\(\\mathcal{S}_1\\), but not when only \\(\\mathcal{S}_2\\) is erased. An additional experiment on token similarity's impact on erasure is detailed in Appendix B." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We found that the adversarial tokens identified by TI encompass those detected by UnlearnDiff. Specifically, we evaluated a model trained to erase \\( S_{1} \\) using UnlearnDiff, which yielded an average Attack Success Rate of \\( 1.52\\% \\), in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.352, + 0.905, + 0.502 + ], + "angle": 0, + "content": "dicating robustness against UnlearnDiff. To explain this result, we provide an additional justification. Since UnlearnDiff identifies combinations of existing tokens, we leveraged the attention mechanism of the text encoder by using the \\(\\langle \\mathrm{EOT} \\rangle\\) token to represent the target concept, and applied the same technique to all tokens from \\(S_{1}\\). After this processing, we computed a cosine similarity of 0.78 between the subspace formed by the \\(\\langle \\mathrm{EOT} \\rangle\\) token from \\(S_{1}\\) and the one identified by UnlearnDiff, exceeding 0.6 and supporting our observations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.504, + 0.906, + 0.595 + ], + "angle": 0, + "content": "In summary, our analysis explains why STEREO can robustly remove the target concept. By applying TI on the concept-erased model iteratively in at least 3 iterations, it can collect a sufficient set of tokens spanning the target concept space. When erasing these tokens together, it successfully eradicates the model knowledge of the concept." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.606, + 0.905, + 0.638 + ], + "angle": 0, + "content": "3.2. Reference Subspace Selection for Narrow Concepts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.901 + ], + "angle": 0, + "content": "CA [16] and ESD [7] were the first methods proposed for concept erasure, but they present contradictory approaches. CA maps noise conditioned on the target concept to a more general representation, while ESD pushes the target noise as far as possible from the original. Studies [18, 33] show that CA outperforms ESD in terms of image quality. It is because mapping approach offers a clear direction to erase target concepts, whereas pushing approach can inadvertently affect unrelated concepts. The second stage of STEREO [33], also a pushing approach, significantly degrades image quality. With this motivation, our paper proposes a mapping approach that erases tokens by mapping them to nearby non-target concept subspaces, preserving model utility. The key challenge is how to automatically select these reference subspaces. We also provide experimental results for Subspace Pushing, where the target subspace is simply pushed as far as possible from the original one, in Appendix C." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19590" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.477, + 0.181 + ], + "angle": 0, + "content": "
ConceptTI Early TokenTextual Token
R2D2 Robot0.641.23
Garbage Truck1.051.28
Grumpy Cat0.981.22
English Springer Dog0.720.96
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.185, + 0.483, + 0.214 + ], + "angle": 0, + "content": "Table 1. The Euclid distance of the target token to the tokens from early TI training steps and the textual one." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.228, + 0.484, + 0.455 + ], + "angle": 0, + "content": "We observe that constructing the reference subspace using tokens from the early training steps of TI is effective for narrow concepts, as early tokens typically converge to more general forms (e.g., grumpy cat to cat, or R2D2 robot to robot; see Section 4.3 for details). Moreover, these tokens are closer to the target token learned from TI because they align along the same gradient direction. To validate this, we compare the Euclidean distance between the target token and early TI training tokens, as well as the target token and the textual token, as shown in Table 1. The results indicate that while early tokens may represent concepts similar to the textual token, their distance from the target token is much smaller. This suggests that mapping the target token to an earlier token minimizes weight modifications, thereby better preserving the model's utility." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.464, + 0.265, + 0.48 + ], + "angle": 0, + "content": "3.3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.486, + 0.484, + 0.683 + ], + "angle": 0, + "content": "Our framework, \\( SuMa \\), consists of two stages, as illustrated in Fig. 3. The first stage, termed Subspace Construction, aims to identify Textual Inversion tokens to establish both sets of Target Subspace \\( \\{\\mathcal{U}^i\\} \\) and Reference Subspace \\( \\{\\mathcal{V}^i\\} \\). Here, \\( \\mathcal{U}^i \\) represents \\( i \\)-th subspace (corresponding to \\( i \\)-th Unet layer) of the concept to be erased, while \\( \\mathcal{V}^i \\) is \\( i \\)-th the subspace into which we intend to map our \\( i \\)-th target subspace. In the second stage, Subspace Elimination, we refine the to- \\( K \\) and to- \\( V \\) weights of the Cross-Attention Layer in U-Net, denoted as \\( W_{k} \\) and \\( W_{v} \\), allowing us to project the basis of \\( \\mathcal{U}^i \\) onto \\( \\mathcal{V}^i \\). Due to the identical functionality of \\( W_{k} \\) and \\( W_{v} \\) in our work, we denote both as \\( W_{k} \\) for simplicity and readability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Subspace Construction. Starting with the original model, denoted as \\(\\theta\\), and a concept to be removed, \\(c\\), we apply a Concept Ablation \\((\\mathcal{M})\\) to eliminate \\(c\\) from \\(\\theta\\), resulting in a modified model \\(\\theta'\\). This process is repeated over \\(l\\) steps. At \\(j\\)-th iteration, we apply Textual Inversion (TI) for the concept to be erased from \\(\\theta'\\), yielding two tokens: one from the final step of TI training, denoted as \\(\\langle u\\rangle_{j}\\), and another from an earlier step, denoted as \\(\\langle v\\rangle_{j}\\). Subsequently, we construct prompts in the formats \"A photo of \\(\\langle u\\rangle_{j}\\)\" and \"A photo of \\(\\langle v\\rangle_{j}\\)\", which are then passed through the Text Encoder \\(\\mathcal{T}\\) to extract embeddings for the tokens \\(\\langle u_j\\rangle\\) and \\(\\langle v_j\\rangle\\). We then apply the transformation \\(W_k^i\\), where \\(i\\) represent \\(i\\)-th layer in Unet, to these extracted embeddings, treating the outputs as \\(\\langle u'\\rangle_{j}^{i}\\) and \\(\\langle v'\\rangle_{j}^{i}\\) as basis vectors for the subspaces \\(\\mathcal{U}^i\\) and \\(\\mathcal{V}^i\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.89, + 0.107 + ], + "angle": 0, + "content": "respectively. Formally, this process is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.112, + 0.905, + 0.151 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\langle u ^ {\\prime} \\rangle_ {j} ^ {i} = W _ {k} ^ {i} \\mathcal {T} (p + \\langle u \\rangle_ {j}), \\tag {2} \\\\ \\langle v ^ {\\prime} \\rangle_ {j} ^ {i} = W _ {k} ^ {i} \\mathcal {T} (p + \\langle v \\rangle_ {j}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.155, + 0.906, + 0.201 + ], + "angle": 0, + "content": "where \\(p\\) represents the prompt template. Finally, we fine-tune \\(\\theta'\\) using \\(\\mathcal{M}\\) to remove \\(\\langle u_j\\rangle\\), resulting in an updated model \\(\\theta'\\) at the end of each iteration (Fig. 3a)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.208, + 0.907, + 0.343 + ], + "angle": 0, + "content": "Subspace Elimination. Fig. 3b illustrates this process. After constructing sets of the subspaces \\(\\{\\mathcal{U}^i\\}_{i=1}^M\\) and \\(\\{\\mathcal{V}^i\\}_{i=1}^M\\), with \\(M\\) is number of layer in Unet, our goal is to eliminate all \\(\\mathcal{U}^i\\) by fine-tuning \\(W_k\\) so that the basis vectors of \\(\\mathcal{U}^i\\) are projected onto \\(\\mathcal{V}^i\\). To construct the projection matrix onto \\(\\mathcal{V}^i\\), we first define the matrix \\(\\mathbf{V}^i = [\\langle v' \\rangle_1^i, \\langle v' \\rangle_2^i, \\dots, \\langle v' \\rangle_l^i] \\in \\mathbb{R}^{d \\times l}\\), with \\(d\\) is the dimension of token embedding vector, and the projection matrix \\(\\mathbf{P}_{\\mathbf{V}^i}\\) can then be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.354, + 0.907, + 0.383 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {\\mathbf {V} ^ {i}} = \\mathbf {V} ^ {i} \\left(\\left(\\mathbf {V} ^ {i}\\right) ^ {T} \\mathbf {V} ^ {i}\\right) ^ {- 1} \\left(\\mathbf {V} ^ {i}\\right) ^ {T} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.386, + 0.906, + 0.416 + ], + "angle": 0, + "content": "We define \\(L_{proj}\\) as the objective to project all basis of \\(\\mathcal{U}^i\\) onto \\(\\mathcal{V}^i\\) by following formula:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.421, + 0.924, + 0.476 + ], + "angle": 0, + "content": "\\[\nL _ {p r o j} = \\sum_ {i = 1} ^ {M} \\sum_ {j = 1} ^ {l} \\left\\| W _ {k} ^ {i} \\mathcal {T} (p + \\langle u \\rangle_ {j}) - \\mathbf {P} _ {\\mathbf {V} ^ {i}} W _ {k} ^ {i} \\mathcal {T} (p + \\langle u \\rangle_ {j}) \\right\\| _ {2} ^ {2} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.478, + 0.907, + 0.508 + ], + "angle": 0, + "content": "As per [3], we add a regularization term to keep the Reference Subspace \\(\\mathcal{V}^i\\) unchanged during fine-tuning:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.525, + 0.907, + 0.58 + ], + "angle": 0, + "content": "\\[\nL _ {r e g} = \\sum_ {i = 1} ^ {M} \\sum_ {j = 1} ^ {l} \\left\\| W _ {k} ^ {i} \\mathcal {T} (p + \\langle v \\rangle_ {j}) - \\mathbf {P} _ {\\mathbf {V} ^ {i}} W _ {k} ^ {i} \\mathcal {T} (p + \\langle v \\rangle_ {j}) \\right\\| \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.906, + 0.611 + ], + "angle": 0, + "content": "The objective function for subspace mapping, \\(L_{sub}\\), is then defined as the weighted sum of these two terms:" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.618, + 0.905, + 0.634 + ], + "angle": 0, + "content": "\\[\nL _ {s u b} = L _ {p r o j} + \\lambda_ {r e g} L _ {r e g}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.906, + 0.775 + ], + "angle": 0, + "content": "where \\(\\lambda_{reg}\\) is a hyperparameter (default 1). If we use \\(L_{sub}\\) to fine-tune \\(\\theta\\) to obtain \\(\\theta'\\) and use CCE to find a new target token, this token still generates the target concept when conditioned on \\(\\theta'\\), but not on \\(\\theta\\) (see Appendix D). We assume modifying only the cross-attention layer is effective for erasing a token, as in [18], but affects the SDE trajectory when erasing multiple tokens. To mitigate this, we use the CA loss function and discuss our choice in Section 4.3. In short, their objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.78, + 0.915, + 0.811 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {C A}} = \\mathbb {E} _ {z _ {t}, c, t} \\left[ \\| \\Phi^ {\\prime} \\left(z _ {t}, \\mathcal {T} (p + c ^ {\\prime}), t\\right) - \\Phi \\left(z _ {t}, \\mathcal {T} (p + c), t\\right) \\| _ {2} ^ {2} \\right] \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.812, + 0.906, + 0.858 + ], + "angle": 0, + "content": "where \\(\\Phi'\\) is the model to learn, \\(\\Phi\\) is pretrained, \\(\\mathcal{T}\\) is TextEncoder, \\(p\\) is prompt template, \\(c\\) is target concept, and \\(c'\\) is its general term (e.g., grumpy cat vs. cat). Our objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.878, + 0.905, + 0.904 + ], + "angle": 0, + "content": "\\[\nW _ {k} ^ {*} = \\underset {W _ {k}} {\\operatorname {a r g m i n}} \\left(L _ {\\text {f i n a l}}\\right) = \\underset {W _ {k}} {\\operatorname {a r g m i n}} \\left(L _ {\\mathrm {C A}} + L _ {\\text {s u b}}\\right). \\tag {8}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19591" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.496, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.251, + 0.495, + 0.314 + ], + "angle": 0, + "content": "(a) Qualitative Comparison across 6 methods for the CCE attack. Both E-CEM methods, CA and MACE, are vulnerable to the CCE attack. STEREO's erasure is too aggressive, leaving only Van Gogh-style elements, with this concept being easier to erase. Our method defends against CCE by mapping the target concept to its reference concept." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.089, + 0.902, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.252, + 0.905, + 0.313 + ], + "angle": 0, + "content": "(b) Qualitative Comparison across 6 methods for image generation. CA struggles with \"English Springer\" and \"Elon Musk.\" MACE performs well, preserving image quality. RACE shows slight degradation, and STEREO significantly reduces quality. Our method strikes a balance between E-CEMs and C-CEMs in terms of quality." + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.318, + 0.88, + 0.332 + ], + "angle": 0, + "content": "Figure 4. Qualitative Results. We present qualitative results comparing five methods against the CCE attack and textual erasure." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.35, + 0.224, + 0.366 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.374, + 0.292, + 0.391 + ], + "angle": 0, + "content": "4.1. Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.485, + 0.691 + ], + "angle": 0, + "content": "Dataset. We evaluate SuMa across four tasks: subclass erasure, celebrity erasure, artistic style erasure, and instance erasure. For subclass erasure, we focus on a subset of ImageNette [12] known for its distinctive features. Three subclasses are selected for the test, including English Springer Dog, Garbage Truck, and Golf Ball. For celebrity erasure, we select 3 celebrities in the Celebrity-1000 dataset [34], including David Beckham, Elon Musk, and Adam Lambert. For artistic style erasure, following the approach in [7, 15, 33], we focus on the style of Van Gogh, which is both famous and straightforward to verify in outputs. For instance erasure, we conduct experiments on three copyrighted concepts: Grumpy Cat, Mickey Mouse, and the R2D2 robot. Note that our method is specifically designed to address the problem with narrow concepts, so it does not work as well for NSFW concepts. However, as mentioned earlier, our work is complementary to DUO, so it can be easily integrated with this method to completely erase all kinds of concepts, demonstrated in the Appendix E." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.698, + 0.484, + 0.804 + ], + "angle": 0, + "content": "Baseline. We compare SuMa with four baselines: two E-CEMs (CA [16] and MACE [18]) and four C-CEMs (RACE [15], AdvUnlearn (AU) [38], STEREO [33], DUO [21]). For CA, MACE, RACE, and AU, we use their public code and recommended settings to erase concepts. For STEREO and DUO, we re-implement their methods using paper recommendations and verify consistency with reported results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Circumventing Methods. We evaluate the robustness of our method under CCE [22] and UnlearnDiff (UD) [39]. For CCE, we use a different set of images from the Subspace Construction stage to ensure the model didn't overfit to this image set. For both attack methods, we use publicly available code to generate adversarial prompts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.351, + 0.908, + 0.715 + ], + "angle": 0, + "content": "Evaluation Metrics. First, to measure effectiveness, for each erased model, we generate 5,000 images using prompts from the MS-COCO validation set. We then evaluate the model quality when generating such non-target prompts using CLIP scores (higher is better) and FID scores (lower is better). To measure completeness, in the CCE attack, for each erased model, we execute the circumventing method to acquire the adversarial token \\( c_{adv} \\). We then generate 200 images using a template prompt: \"A photo of a \\( \\{c_{adv}\\} \\).\" For style erasure, we use the template \"A painting in the style of \\( \\{c_{adv}\\} \\)\". In the UD attack, we execute its code on our erased model to get the adversarial prompt and pass the prompt to the erased model. We report the attack success rate (ASR), where a lower score indicates better performance, following the method in [15, 33]. For subclass erasure tasks, we utilize ResNet-50 [10], pretrained on ImageNet [27], to measure ASR. For celebrity detection, we use the GIPHY Celebrity Detector (GCD) [9] as the classifier. To evaluate artistic style, we follow the approach in [15] by fine-tuning a ViT-based model [1] on the WikiArt dataset [29]. Finally, for instance detection, due to the lack of pretrained models or datasets, we use LLAVA-7B [17] with a template prompt: \"Does this image contain {object}? Answer Yes or No.\" for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. We use SD1.4 and SD1.5 as the pretrained model. In all experiments, we fine-tune all cross-attention layers. In the Subspace Construction stage, we use the same setting as CA [16] for concept erasion. Training TI is set for 500 steps, with the early step to get token \\(\\langle v\\rangle\\) set to 50, the initial token is toy, with Adam optimization with learning rate, is 2e-6. For the Subspace Elimination stage, we fine-tune most objects for 500 steps, except for subclass objects, which require 750 steps to ensure model safety. We use the Adam optimizer with a learning rate of 2e-5. All experiments are done on 2 NVIDIA A100 GPUs." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19592" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.089, + 0.505, + 0.365 + ], + "angle": 0, + "content": "
MethodCCE↓UD↓FID↓CLIP↑
Subclass
SD [24]0.84 / 0.830.74 / 0.7617.04 / 16.950.33 / 0.32
CA [16]0.78 / 0.750.67 / 0.6219.28 / 19.260.32 / 0.32
MACE [18]0.86 / 0.840.73 / 0.6817.15 / 16.980.29 / 0.31
RACE [15]0.74 / 0.730.17 / 0.1826.34 / 26.220.27 / 0.28
AdvUnlearn [38]0.76 / 0.740.18 / 0.1618.47 / 18.390.30 / 0.31
DUO [21]0.65 / 0.630.62 / 0.6517.09 / 16.990.29 / 0.30
STEREO [33]0.05 / 0.030.05 / 0.0427.33 / 27.200.30 / 0.28
Ours0.12 / 0.100.08 / 0.0618.33 / 18.150.31 / 0.30
Identity
SD [24]0.91 / 0.900.86 / 0.8417.04 / 16.950.33 / 0.32
CA [16]0.87 / 0.890.79 / 0.7618.21 / 18.270.31 / 0.30
MACE [18]0.89 / 0.870.65 / 0.6316.74 / 16.690.29 / 0.28
RACE [15]0.78 / 0.790.47 / 0.4624.32 / 24.170.27 / 0.29
AdvUnlearn [38]0.77 / 0.690.52 / 0.5317.53 / 17.690.31 / 0.31
DUO [21]0.73 / 0.710.67 / 0.7317.32 / 17.110.29 / 0.30
STEREO [33]0.02 / 0.010.03 / 0.0126.33 / 26.150.29 / 0.28
Ours0.09 / 0.070.18 / 0.1417.94 / 17.890.31 / 0.29
" + }, + { + "type": "table", + "bbox": [ + 0.51, + 0.089, + 0.895, + 0.363 + ], + "angle": 0, + "content": "
MethodCCE↓UD↓FID↓CLIP↑
Artistic
SD [24]0.86 / 0.830.82 / 0.8517.04 / 16.950.33 / 0.32
CA [16]0.79 / 0.780.74 / 0.7618.02 / 17.850.34 / 0.31
MACE [18]0.84 / 0.820.79 / 0.8316.57 / 16.580.32 / 0.33
RACE [15]0.82 / 0.790.06 / 0.0421.32 / 21.100.30 / 0.31
AdvUnlearn [38]0.83 / 0.760.03 / 0.0517.73 / 17.600.32 / 0.33
DUO [21]0.74 / 0.720.67 / 0.7317.74 / 17.890.30 / 0.31
STEREO [33]0.00 / 0.020.00 / 0.0128.04 / 27.900.30 / 0.31
Ours0.05 / 0.040.07 / 0.0517.18 / 17.210.33 / 0.31
Instance
SD [24]0.97 / 0.940.94 / 0.9317.04 / 16.950.33 / 0.32
CA [16]0.97 / 0.930.92 / 0.8918.43 / 18.300.30 / 0.31
MACE [18]0.97 / 0.920.93 / 0.8916.93 / 16.800.30 / 0.31
RACE [15]0.95 / 0.920.06 / 0.0724.32 / 24.150.28 / 0.29
AdvUnlearn [38]0.95 / 0.930.03 / 0.0418.37 / 18.200.32 / 0.33
DUO [21]0.92 / 0.910.89 / 0.8718.31 / 18.920.30 / 0.31
STEREO [33]0.03 / 0.020.05 / 0.0327.04 / 26.900.31 / 0.32
Ours0.17 / 0.150.19 / 0.1622.34 / 22.200.29 / 0.30
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.368, + 0.908, + 0.41 + ], + "angle": 0, + "content": "Table 2. Evaluation of Erasing across 4 Categories. We average the results of all target concepts in each category across three metrics and compare our framework, SuMa, with state-of-the-art methods. For ASRs, successful attacks (ASR ≥ 20%) are in red, and defeated attacks (ASR < 20%) in green. The data in each cell show the result of erasing a concept from pretrained SD1.4 / SD1.5." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.415, + 0.88, + 0.528 + ], + "angle": 0, + "content": "
Reference SubspaceSubclassIdentityArtisticInstance
ASR ↓FID ↓CLIP ↑ASR ↓FID ↓CLIP ↑ASR ↓FID ↓CLIP ↑ASR ↓FID ↓CLIP ↑
Step-300.0919.110.300.0818.340.290.0518.820.310.1723.960.27
Step-500.1218.330.310.0917.940.310.0517.180.330.1722.340.29
Step-700.5617.320.330.6717.920.320.3717.050.330.8317.740.32
General Term0.1818.270.320.1718.020.320.1118.210.330.1523.340.27
" + }, + { + "type": "table_caption", + "bbox": [ + 0.3, + 0.532, + 0.695, + 0.546 + ], + "angle": 0, + "content": "Table 3. Ablation Study for the selection of Reference subspace." + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.555, + 0.443, + 0.778 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.782, + 0.483, + 0.811 + ], + "angle": 0, + "content": "Figure 5. Visualization of Textual Inversion tokens at different numbers of TI training steps." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.823, + 0.185, + 0.837 + ], + "angle": 0, + "content": "4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Quantitative Results. We average each metric from Section 4.1 for each category and report the results in Table 2. For all baselines, we ensure the erased model works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.559, + 0.907, + 0.892 + ], + "angle": 0, + "content": "well for erasing the textual prompt, and we report results for this trivial task in the Appendix F. For completeness evaluation, our method achieves an ASR score below 0.2 for all categories under both CCE and UD attacks. Compared to STEREO, our method has a slightly higher ASR score (0.02 to 0.14 higher), but within an acceptable range. RACE and AdvUnlearn show safety against the UD attack in the Subclass, Artistic, and Instance categories but fail under Identity. This is due to the subtle facial features of human identity, which make this category more fine-grained and challenging. However, both fail under the CCE attack, showing that protecting the model against CCE is harder than UD. The main reason is that RACE and AdvUnlearn erase adversarial tokens based on UnlearnDiff, but these tokens don't fully represent the target concept, protecting against UD but not CCE. DUO works well for NSFW concepts (results in Appendix E) but fails for narrow concepts. For effectiveness, although STEREO performs well under both CCE and UD, image quality is significantly affected, resulting in an FID score roughly double that of other E-CEMs, making it less effective. Our method, while not achieving a perfect ASR, achieves an FID score comparable" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19593" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.094, + 0.47, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.158, + 0.234, + 0.416, + 0.248 + ], + "angle": 0, + "content": "Figure 6. Multiple Concept Erasure Result." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.253, + 0.479, + 0.354 + ], + "angle": 0, + "content": "
RegulizationSubclassIdentityArtisticInstance
×ASR: 0.09ASR: 0.06ASR: 0.05ASR: 0.15
FID: 19.21FID: 20.31FID: 19.32FID: 24.57
CLIP: 0.29CLIP: 0.29CLIP: 0.30CLIP: 0.26
✓ASR: 0.12ASR: 0.09ASR: 0.05ASR: 0.17
FID: 18.33FID: 17.94FID: 17.18FID: 22.34
CLIP: 0.31CLIP: 0.31CLIP: 0.33CLIP: 0.29
" + }, + { + "type": "table_caption", + "bbox": [ + 0.094, + 0.357, + 0.482, + 0.372 + ], + "angle": 0, + "content": "Table 4. Ablation Study for the need of Subspace Regulization." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.38, + 0.479, + 0.466 + ], + "angle": 0, + "content": "
EntityFID ↓CCE ↓UD ↓
English Springer20.12 / 18.140.12 / 0.470.07 / 0.34
Van Gogh20.16 / 17.540.07 / 0.520.04 / 0.46
Elon Musk20.15 / 18.220.07 / 0.490.05 / 0.57
Grumpy Cat24.57 / 20.320.18 / 0.540.14 / 0.46
" + }, + { + "type": "table_caption", + "bbox": [ + 0.103, + 0.469, + 0.471, + 0.483 + ], + "angle": 0, + "content": "Table 5. Ablation Study for the base method (ESD / UCE)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.504, + 0.483, + 0.581 + ], + "angle": 0, + "content": "to other E-CEMs in the Subclass, Artistic, and Identity categories, outperforming CA with an FID around 0.85 lower. We found that Instance category is harder to be erased, so we fine-tuned it for 750 steps, resulting in a higher FID compared to the others, fine-tuned for 500 steps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.589, + 0.483, + 0.741 + ], + "angle": 0, + "content": "Qualitative Results are provided in Fig. 4. Fig. 4a shows results of these models against the CCE attack, and we include qualitative results for UD in the Appendix J. CA, MACE, RACE, and AdvUnlearn are highly vulnerable to this attack. STEREO completely removes these concepts, but at the cost of image quality (\\(6^{th}\\) column in Fig. 4b). Our method successfully converts \"Springer Dog\" to \"Dog,\" \"Elon Musk\" to a general person, and \"Grumpy Cat\" to \"Cat toy,\" while maintaining image quality comparable to other E-CEM methods (last column in Fig. 4b)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Multiple Concepts Erasure. We provide results for erasing multiple concepts in Fig. 6. We test the ASR with the CCE attack and evaluate two settings: Iterative Erase (IE) and Simultaneous Erase (SE). IE means we fine-tune the weights after erasing one concept to erase others. SE sums all \\( L_{sub} \\) to erase all concepts in one fine-tuning, with 1000 steps fine-tuning for this experiment. In both settings, our method defends against CCE and maintains model utility. With SE, we achieve better FID scores, as summing \\( L_{sub} \\) preserves a broader range of knowledge from \\( L_{reg} \\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.68, + 0.107 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.12, + 0.907, + 0.332 + ], + "angle": 0, + "content": "Effect of selecting the Reference Subspace. We present results for four ways of constructing the Reference Subspace, including TI tokens from Steps 30, 50, 70 of TI training, and the concept's general term, as shown in Table 3. Visualizations of each token are in Fig. 5. At Step 30, the token is far from the target, with little relevant information. At Step 50, the token shares features with the target but is not an exact match, enabling minimal transformation. At Step 70, the token is close to the target, making full erasure difficult. These observations align with Table 3, where Step 50 balances ASR, FID, and CLIP scores, while Step 30 improves ASR but slightly reduces image quality. Step 70 shows results similar to CA, indicating SuMa's inefficacy when the Reference Subspace and Target Subspace are too close." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.339, + 0.909, + 0.568 + ], + "angle": 0, + "content": "Effect of choosing the base method. As noted in Section 3.3, we adopt CA loss for our method. Results using ESD [7] and UCE [8] as base methods are shown in Table 5. While ESD resembles CA, it uses a pushing strategy, whereas CA employs a mapping strategy. UCE erases concepts by altering the cross-attention map. Our experiments show that modifying only the attention layer causes the token learned by CCE to express the target concept only under the erased model as shown in Appendix D. ESD provides some defense but increases FID moderately, while UCE fails to defend. We conclude that modifying the model's noise prediction is more effective, as it redirects the target concept's trajectory toward another in the original model. The mapping approach is also favored over the pushing one, as discussed in Section 3.2 and supported by our results." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.574, + 0.909, + 0.725 + ], + "angle": 0, + "content": "Effect of adding Subspace Regularization. We present the effect of adding Subspace Regularization and report the results in Table 4. We observe that without regularization, our method performs slightly better against the CCE attack, with an average \\(2\\%\\) reduction in attack rate, but at the cost of a gradual increase in the FID score. This phenomenon was studied in [3], where it was shown that the concept closest to the target is the most affected. Therefore, we conclude that adding Subspace Regularization loss is necessary to maintain the model's effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.74, + 0.771, + 0.755 + ], + "angle": 0, + "content": "5. Conclusion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this paper, we introduced SuMa, a method that robustly and effectively erases narrow concepts in Text-to-Image Diffusion models. SuMa constructs a Target Subspace for the concept to be erased and maps it to a Reference Subspace. Experimental results show our method is the first to meet these dual objectives for narrow concepts. A key limitation is its time cost, as it may require many iterations to construct both subspaces. Future work will aim to accelerate this process and improve target concept representation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19594" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.156 + ], + "angle": 0, + "content": "[1] Dosovitskiy Alexey. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv: 2010.11929, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.484, + 0.214 + ], + "angle": 0, + "content": "[2] Anh Bui, Khanh Doan, Trung Le, Paul Montague, Tamas Abraham, and Dinh Phung. Removing undesirable concepts in text-to-image generative models with learnable prompts. arXiv preprint arXiv:2403.12326, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.217, + 0.483, + 0.271 + ], + "angle": 0, + "content": "[3] Anh Bui, Long Vuong, Khanh Doan, Trung Le, Paul Montague, Tamas Abraham, and Dinh Phung. Erasing undesirable concepts in diffusion models with adversarial preservation. arXiv preprint arXiv:2410.15618, 2024. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.273, + 0.483, + 0.356 + ], + "angle": 0, + "content": "[4] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, José Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. In Proceedings of the 40th International Conference on Machine Learning, pages 4055-4075, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.359, + 0.483, + 0.413 + ], + "angle": 0, + "content": "[5] Zhi-Yi Chin, Chieh-Ming Jiang, Ching-Chun Huang, PinYu Chen, and Wei-Chen Chiu. Prompting4debugging: Redteaming text-to-image diffusion models by finding problematic prompts. arXiv preprint arXiv:2309.06135, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.416, + 0.483, + 0.483 + ], + "angle": 0, + "content": "[6] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel CohenOr. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.487, + 0.483, + 0.555 + ], + "angle": 0, + "content": "[7] Rohit Gandikota, Joanna Materzynska, Jaden Fiitto-Kaufman, and David Bau. Erasing concepts from diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2426-2436, 2023. 1, 2, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.558, + 0.483, + 0.626 + ], + "angle": 0, + "content": "[8] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5111-5120, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.629, + 0.483, + 0.656 + ], + "angle": 0, + "content": "[9] Nick Hasty, Ihor Kroosh, Dmitry Voitekh, and Dmytro Korduban. Giphy celebrity detector, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.659, + 0.483, + 0.713 + ], + "angle": 0, + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.716, + 0.483, + 0.743 + ], + "angle": 0, + "content": "[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.483, + 0.772 + ], + "angle": 0, + "content": "[12] Jeremy Howard and Sylvain Gugger. Fastai: a layered api for deep learning. Information, 11(2):108, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[13] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Harry H Jiang, Lauren Brown, Jessica Cheng, Mehtab Khan, Abhishek Gupta, Deja Workman, Alex Hanna, Johnathan Flowers, and Timnit Gebru. Ai art and its impact on artists. In Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society, pages 363-374, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[15] Changhoon Kim, Kyle Min, and Yezhou Yang. Race: Robust adversarial concept erasure for secure text-to-image diffusion model. arXiv preprint arXiv:2405.16341, 2024. 1, 3, 6, 7, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.907, + 0.219 + ], + "angle": 0, + "content": "[16] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22691-22702, 2023. 1, 2, 3, 4, 6, 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.906, + 0.262 + ], + "angle": 0, + "content": "[17] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.906, + 0.331 + ], + "angle": 0, + "content": "[18] Shilin Lu, Zilan Wang, Leyang Li, Yanzhu Liu, and Adams Wai-Kin Kong. Mace: Mass concept erasure in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6430-6440, 2024. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.334, + 0.907, + 0.417 + ], + "angle": 0, + "content": "[19] Mengyao Lyu, Yuhong Yang, Haiwen Hong, Hui Chen, Xuan Jin, Yuan He, Hui Xue, Jungong Han, and Guiguang Ding. One-dimensional adapter to rule them all: Concepts diffusion models and erasing applications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7559-7568, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.42, + 0.906, + 0.501 + ], + "angle": 0, + "content": "[20] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.504, + 0.906, + 0.585 + ], + "angle": 0, + "content": "[21] Yong-Hyun Park, Sangdoo Yun, Jin-Hwa Kim, Junho Kim, Geonhui Jang, Yonghyun Jeong, Junghyo Jo, and Gayoung Lee. Direct unlearning optimization for robust and safe text-to-image models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2, 3, 6, 7, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.589, + 0.906, + 0.656 + ], + "angle": 0, + "content": "[22] Minh Pham, Kelly O Marshall, Niv Cohen, Govind Mittal, and Chinmay Hegde. Circumventing concept erasure methods for text-to-image generative models. In The Twelfth International Conference on Learning Representations, 2023. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.66, + 0.906, + 0.714 + ], + "angle": 0, + "content": "[23] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.717, + 0.906, + 0.743 + ], + "angle": 0, + "content": "[24] Robin Rombach. Stable diffusion v1-4 model card., 2022. 1, 7, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.871, + 0.758 + ], + "angle": 0, + "content": "[25] Robin Rombach. Stable diffusion 2.0 release., 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.761, + 0.907, + 0.829 + ], + "angle": 0, + "content": "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[27] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19595" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[28] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[29] Babak Saleh and Ahmed Elgammal. Large-scale classification of fine-art paintings: Learning the right metric on the right feature. arxiv 2015. arXiv preprint arXiv:1505.00855, 2015.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[30] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.319, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[31] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by {Text-to-Image} models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 2187–2204, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[32] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6048-6058, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.461, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[33] Koushik Srivatsan, Fahad Shamshad, Muzammal Naseer, Vishal M Patel, and Karthik Nandakumar. Stereo: A two-stage framework for adversarially robust concept erasing from text-to-image diffusion models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 23765-23774, 2025. 2, 3, 4, 6, 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.546, + 0.381, + 0.56 + ], + "angle": 0, + "content": "[34] Tonyassi. Celebrity-1000 datasets, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[35] Yu-Lin Tsai, Chia-Yi Hsu, Chulin Xie, Chih-Hsun Lin, Jia-You Chen, Bo Li, Pin-Yu Chen, Chia-Mu Yu, and Chun-Ying Huang. Ring-a-bell! how reliable are concept removal methods for diffusion models? arXiv preprint arXiv:2310.10012, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[36] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228-8238, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[37] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[38] Yimeng Zhang, Xin Chen, Jinghan Jia, Yihua Zhang, Chongyu Fan, Jiancheng Liu, Mingyi Hong, Ke Ding, and Sijia Liu. Defensive unlearning with adversarial training for robust concept erasure in diffusion models. arXiv preprint arXiv:2405.15234, 2024. 2, 3, 6, 7, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[39] Yimeng Zhang, Jinghan Jia, Xin Chen, Aochuan Chen, Yihua Zhang, Jiancheng Liu, Ke Ding, and Sijia Liu. To generate or not? safety-driven unlearned diffusion models are still" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.547, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "easy to generate unsafe images... for now. In European Conference on Computer Vision, pages 385-403. Springer, 2025. 1, 2, 3, 6" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19596" + } + ] +] \ No newline at end of file diff --git a/2025/SuMa_ A Subspace Mapping Approach for Robust and Effective Concept Erasure in Text-to-Image Diffusion Models/51080fa1-f759-4df5-8927-63f76b232e8f_origin.pdf b/2025/SuMa_ A Subspace Mapping Approach for Robust and Effective Concept Erasure in Text-to-Image Diffusion Models/51080fa1-f759-4df5-8927-63f76b232e8f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..559f5286b4b5e589046166b7abf0a75707d73c6f --- /dev/null +++ b/2025/SuMa_ A Subspace Mapping Approach for Robust and Effective Concept Erasure in Text-to-Image Diffusion Models/51080fa1-f759-4df5-8927-63f76b232e8f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c3bef74212747bfa7d89ce0ec92c666abe13aec71cea353acb6982f71f81ae7 +size 8212583 diff --git a/2025/SuMa_ A Subspace Mapping Approach for Robust and Effective Concept Erasure in Text-to-Image Diffusion Models/full.md b/2025/SuMa_ A Subspace Mapping Approach for Robust and Effective Concept Erasure in Text-to-Image Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5e05af57387e454d90b690bff1248d8fb4d836a2 --- /dev/null +++ b/2025/SuMa_ A Subspace Mapping Approach for Robust and Effective Concept Erasure in Text-to-Image Diffusion Models/full.md @@ -0,0 +1,286 @@ +# SuMa: A Subspace Mapping Approach for Robust and Effective Concept Erasure in Text-to-Image Diffusion Models + +Kien Nguyen1 Anh Tran1 Cuong Pham2,1 + +$^{1}$ Qualcomm AI Research + +2Posts and Telecommunications Institute of Technology (PTIT) + +{kienn, anhtra, pcuong}@qti.qualcomm.com + +# Abstract + +The rapid growth of text-to-image diffusion models has raised concerns about their potential misuse in generating harmful or unauthorized contents. To address these issues, several Concept Erasure methods have been proposed. However, most of them fail to achieve both robustness, i.e., the ability to robustly remove the target concept, and effectiveness, i.e., maintaining image quality. While few recent techniques successfully achieve these goals for NSFW concepts, none could handle narrow concepts such as copyrighted characters or celebrities. Erasing these narrow concepts is critical in addressing copyright and legal concerns. However, erasing them is challenging due to their close distances to non-target neighboring concepts, requiring finer-grained manipulation. In this paper, we introduce Subspace Mapping (SuMa), a novel method specifically designed to achieve both robustness and effectiveness in easing these narrow concepts. SuMa first derives a target subspace representing the concept to be erased and then neutralizes it by mapping it to a reference subspace that minimizes the distance between the two. This mapping ensures the target concept is robustly erased while preserving image quality. We conduct extensive experiments with SuMa across four tasks: subclass erasure, celebrity erasure, artistic style erasure, and instance erasure and compare the results with current state-of-the-art methods. Our method achieves image quality comparable to approaches focused on effectiveness, while also yielding results that are on par with methods targeting completeness. + +# 1. Introduction + +Text-to-image models (T2I) at a large scale have achieved notable success in producing lifelike images [4, 20, 23, 26, 28, 37]. This progress is not only a result of enhanced algorithms and computing resources but also the extensive datasets gathered from the Internet [30]. Unfortunately, + +these datasets often contain copyrighted works, artistic creations [14, 31, 32], NSFW content, and personal images, raising ethical and legal concerns [18]. + +Some publicly available models, like Stable Diffusion (SD) [24], trained on vast web-scraped datasets, can generate inappropriate content, including copyrighted images, offensive material, and deepfakes, raising serious community concerns. To address this, researchers have attempted to filter out such content, resulting in the release of Stable Diffusion 2.0 [25]. However, this version, despite the filtering efforts, delivers lower image quality than its predecessor, with some explicit content still slipping through. Another approach involves post-processing filters, such as NSFW detectors, which flag inappropriate outputs and replace them with blank images. Yet, these safeguards can be easily bypassed once the model is publicly accessible [18]. + +Recent research has concentrated on fine-tuning text-to-image (T2I) diffusion models to 'forget' inappropriate content, giving rise to Concept Erased Models (CEMs) as a practical approach to mitigate such issues. The first generation of these models, which we term Effective Concept Erased Models (E-CEMs), demonstrates success in removing inappropriate content while maintaining image quality comparable to the original outputs [2, 7, 16, 18, 19]. However, recent studies have uncovered significant vulnerabilities in these methods [22, 35, 39]. For instance, Unlearn-Diff [39] identifies combinations of existing tokens within current models capable of reconstructing the target concept, effectively bypassing the erasing mechanism. Additionally, CCE [22] leverages Textual Inversion [6] to derive a new token that regenerates the erased concept. Collectively, these findings suggest that E-CEMs fail to robustly eliminate target concepts from the model's latent space, leaving them susceptible to regeneration through adversarial techniques. + +The second generation, which we call Robust Concept Erased Models (R-CEMs), addresses vulnerabilities revealed by adversarial attacks. Early approaches, like RACE [15], use an iterative framework based on E-CEM, identifying adversarial tokens and erasing them. AdvUnlearn + +[38] improves RACE by adding a prior preservation loss to retain model utility. However, while these methods mitigate UnlearnDiff attacks, they struggle against CCE. This is because CCE generates new tokens, while UnlearnDiff relies on existing token combinations, limiting its versatility compared to CCE's broader search space. We provide a discussion about these two attacking methods in Appendix A. More recent methods, such as STEREO [33], refine this process by identifying and erasing all adversarial tokens detected by CCE simultaneously. While it demonstrates notable success in terms of robustness and resilience against current attack methods, it incurs a significant reduction in image quality. In contrast, DUO [21] first applies an image editing approach to modify inappropriate content, then leverages the Diffusion-DPO framework [36] to assign high scores to appropriate images and low scores to inappropriate ones while also introducing a novel prior preservation loss. This method successfully protects the model from existing attacks while maintaining image quality. However, it is currently limited to erasing NSFW content. + +While NSFW content is a common target of concept erasure, narrow concepts are not less important targets. Here, we use the terms "narrow concepts" to refer to fine-grained categories or specific instances within, such as "English Springer", "Grumpy Cat", or "Elon Musk". The ability to erase narrow concepts is critical, as they may involve copyright issues or contribute to the spread of misleading information. While DUO is the only robust and effective concept erasure technique, it struggles to handle narrow concepts. We discuss in more detail the reason in the Appendix A. In short, this method relies on an editing approach to erase concepts. For instance, it edits the "nudity" concept to "dressed in", where the difference between the two is clear. However, for narrow concepts, such as editing "Grumpy Cat" to "Cat", it is harder to classify these as distinct classes, requiring a finer-grained method to address. + +Given the aforementioned challenges, our paper presents complementary work to DUO [21], focusing on achieving a balance between robust and effective erasure for narrow concepts. We first investigate why adversarial tokens can bypass E-CEMs and analyze how STEREO achieves robustness against such attacks. We then introduce our method, Subspace Mapping Concept Erasure (SuMa). Our approach constructs both a target subspace containing the concept to be erased and a reference subspace that is independent yet close to the target subspace by identifying their basis vectors. We eliminate the target subspace by fine-tuning the model to optimize a novel objective function, aiming to project the basis of the target subspace onto the reference subspace. This method is integrated with E-CEM techniques, making the model robust against adversarial attacks while still preserving image quality. As a result, SuMa achieves image quality on par with E-CEMs in + +![](images/277078d12c1231060298c8585bf5d7f5a46762c3cfaa1c2c78092f860fc90230.jpg) +AdvUnlearn + +![](images/be61c490d799cfcb718ca6bf6d9c101fc4368d3b8b2c249bb307993d3fd0fb4e.jpg) +RACE + +![](images/a8a405c84350eda90f0268bf14bbcfff8e0aba93b3ea9c34af118779fd5b25a6.jpg) +STEREO +Figure 1. Illustration of current C-CEMs methods trained to erase garbage truck against the CCE Attack. We can see that AdvUnlearn and RACE are still vulnerable to the CCE Attack, while STEREO successfully defends against it, though at the cost of image quality. Our method successfully balances Completeness and Effectiveness, changing garbage truck into Semi-trailer truck. + +![](images/335b486836d33d5aa32c3b5238c712da9740f8cfe5c09d4f19349c15e7f93258.jpg) +Ours + +terms of FID and CLIP-score, and an Attack Success Rate (ASR) comparable to the state-of-the-art R-CEM method, STEREO. Our main contributions are as follows: + +- We analyze why adversarial tokens can bypass E-CEMs and how STEREO remains resilient against these attacks. +- We propose the Subspace Mapping Concept Erasure (SuMa) framework, which balances robustness and effectiveness in erasing narrow concepts. Our framework first constructs a target subspace that captures the concept to be erased and a reference subspace that is distinct yet proximate to the target. We then introduce a novel objective function that removes the target concept by mapping the target subspace onto the reference subspace. +- Extensive experiments across four tasks—object erasure, celebrity erasure, artistic style erasure, and subclass erasure—demonstrate that SuMa offers the best trade-off between completeness and effectiveness. SuMa completely removes the target concept while preserving model utility. To the best of our knowledge, we are the first to achieve these dual objectives for narrow concepts. + +# 2. Related Work + +# 2.1. Effective Concept Erasure (E-CEMs) + +The first generation of the Concept Erasure Models (CEMs) was designed to fine-tune original models to forget inappropriate content, successfully removing unwanted elements while closely maintaining image quality. Concept Ablation (CA) [16] finetunes the model to redirect the target prompt, e.g., "grumpy cat", to be similar as a general term, e.g., "cat", by aligning the noise predictions on these prompts. ESD [7] pushed the noise predictions on the target concept away from the original noise predictions by reversing the direction of conditional guidance [11]. SPM [19] focused on preserving non-target concepts during erasure by applying the CA method with added regularization that maintained noise predictions on the closet concept in the CLIP Embedding Space. MACE [18] proposes a LORA-Finetuning [13] method to erase multiple concepts simultaneously. Despite these advancements, recent adversarial attacks [22, 35, 39] successfully recovered the "erased" concepts, exposing vul + +![](images/165ae80020bdf0b85fa1d0f09a70dfdf69a858ae647698a5420c6fff71bd863a.jpg) +(a) Cosine similarity of TIs from SD1.4. + +![](images/4474b9c98fc92f7b550a34fb29b445e6eb8801e5663dcb6eb19ece8d6db33e42.jpg) +Figure 2. Distribution of cosine similarity scores among TIs in three cases: from the same SD1.4 but different TI training steps, different erasure levels of CA, and between a TI at erasure level $t + 1$ and the subspace created by the $t$ previous TIs. + +![](images/090dd577857f4242ee2266246ba8b5825c876b252cbc7c8136663f35b22cb4c5.jpg) +(b) Cosine similarity of TIs at different CA levels. +(c) Cosine similarity of TIs with subspace. + +nerabilities in all these methods and indicating a need for further improvements in concept erasure techniques. + +# 2.2. Circumventing Concept Erasure + +Several methods have been proposed to attack E-CEMs [22, 35, 39]. P4D [5] enhances adversarial prompts by ensuring that noise estimates between pre-trained and concept-erased models are similar. UnlearnDiff [39] is similar to P4D but leverages the classification capabilities of diffusion models, removing the reliance on pre-trained models. Ring-A-Bell [35] creates adversarial prompts by guiding a "safe" prompt toward an "unsafe" one in the text vector space. Among these, CCE [22] is the simplest and most effective, employing Textual Inversion (TI) [6] to derive a new token from the E-CEM, which regenerates the erased concept. Unlike UnlearnDiff, which relies on combinations of existing tokens, CCE searches for a new token in the latent text space, greatly widening its search space. Hence, defending against CCE is far more challenging than against UnlearnDiff. + +# 2.3. Robust Concept Erasure (R-CEMs) + +Motivated by the need to protect CEMs from adversarial attacks, several methods have emerged as part of the second generation of CEMs [15, 33, 38]. RACE [15] proposes an iterative framework that first identifies adversarial tokens using the UnlearnDiff framework but with faster optimization. It then erases the found tokens using a loss function from ESD, providing protection against UnlearnDiff. However, as shown in Fig. 1, it remains vulnerable to more advanced CCE attacks. AdvUnlearn [38] is similar to RACE, but it introduces a prior preservation loss during the erasing phase to maintain the model's utility. Still, it is susceptible to CCE attacks, as illustrated in Fig. 1. STEREO [33] successfully defends against all existing attack methods by iteratively identifying all Textual Inversions (TIs) from the concept-erased model, and erasing all of them simultaneously from the original model. However, this method leads to a significant degradation in image quality, resulting in an FID nearly double that of E-CEMs. DUO [21] addresses both robustness and effectiveness by leveraging Diffusion-DPO to assign high scores to concept-edited images and + +low scores to inexact ones. It also introduces a novel prior preservation loss to preserve the model's utility. However, this approach is currently limited to NSFW concepts. + +# 3. Method + +In this section, we begin by exploring whether E-CEMs, trained to erase a single token, can also erase tokens that differ in form but convey the same concept (Sec. 3.1). This analysis sets the stage for understanding why CCE [22] can attack E-CEMs and why STEREO [33] can robustly erase a target concept. We then discuss how to find effective reference subspaces for erasing narrow concepts (Sec. 3.2), as well as the motivation for this subspace mapping approach, rather than simply pushing the target subspace away. Finally, we present our proposed method, the Subspace Mapping Approach for Concept Erasure (SuMa) (Sec. 3.3). + +# 3.1. Observations + +We conducted a toy experiment using Concept-Ablation (CA) [16] to investigate whether it can erase tokens that differ in form but convey the same underlying concept. We began by applying Textual Inversion (TI) on SD1.4 for 500 steps using five concepts reported in CA [16]. For each concept, we collected learned tokens from steps 300 to 500, denoted as set $\mathcal{A}$ . At this stage, all tokens in $\mathcal{A}$ could generate the target concept. We then trained CA to erase one randomly selected token from $\mathcal{A}$ and observed that the remaining tokens, though differing in form, were also erased. + +Next, we calculated the cosine similarity between the erased token and all tokens from $\mathcal{A}$ , resulting in a distribution shown in Fig. 2a. We noted that the cosine similarity values ranged from 0.6 to 1.00. We proceeded by running TI on the CA-based concept-erased model above. Upon reaching step 300, TI successfully recovered the erased concept. We then computed the cosine similarity between the new TI tokens and all tokens from $\mathcal{A}$ again and showed the result in Fig. 2b. At this time, the cosine similarity ranged between 0.3 and 0.4, significantly lower than that of the previously erased token. Based on these observations, we hypothesize that we can't represent the target concept as a single + +![](images/9c1ae6d555ebb5978227a619110cd1478e986a8e0afc46c43245971476aecd9f.jpg) +(a) Subspace Construction: Following the STEREO approach, we iteratively remove the Target Concept from model $\theta$ using method $\mathcal{M}$ , then run Textual Inversion (TI) and extract the corresponding tokens. At each iteration, we take the token from the final TI step and one from an early step to form the basis of the Target Subspace and Reference Subspace, respectively. + +![](images/57a7c040693777596ac41c14a2b2c630eecd0a71c97341a5d11a4ea10b8307bf.jpg) +(b) Subspace Elimination: In this stage, our objective is to map the Target Subspace $\mathcal{U}$ onto the Reference Subspace $\mathcal{V}$ , and to maintain the Reference Subspace in its fine-tuned state $\mathcal{V}'$ unchanged. We let $\mathcal{T}$ represent the Text-encoder and $L_{sub}$ denote the cost of mapping $\mathcal{U}$ to $\mathcal{V}$ and $\mathcal{V}'$ to $\mathcal{V}$ . The parameter $W_k'$ is initialized as $W_k$ and will be updated during fine-tuning. +Figure 3. Our proposed framework consists of two main stages. The first stage, Subspace Construction, is used to build Target Subspace and Reference Subspace. The second stage, Subspace Elimination, aims to eliminate the Target Subspace. + +token but rather as a subspace; erasing a concept could be viewed as eliminating a subspace, and if the cosine similarity between this subspace and tokens that produce a similar concept exceeds 0.6, those tokens are erased along with the target subspace. + +To verify this hypothesis, we follow the procedure in STEREO [33] by iteratively erasing tokens found by TI using CA. The process is illustrated in Fig. 3a. For each iteration, we collect the token from the final training step of TI. We aggregate tokens over $n$ iterations, denoted as $x_{1},x_{2},\ldots ,x_{n}\in \mathbb{R}^{d}$ , ensuring that each token $x_{i}$ for $i\in [1,n]$ can generate the target concept. We then construct the matrix $\mathbf{S}_t = [x_1,\dots,x_t]\in \mathbb{R}^{d\times t}$ for $t\in [1,n - 1]$ . Next, we compute the angle between each token $x_{j}$ for $j\in [t + 1,n]$ and the column space of $\mathbf{S}_t$ , denoted by $c_{tj}$ , using the formula: + +$$ +c _ {t j} = \cos \left(x _ {j}, \mathbf {S} _ {t} \left(\mathbf {S} _ {t} ^ {T} \mathbf {S} _ {t}\right) ^ {- 1} \mathbf {S} _ {t} ^ {T} x _ {j}\right), \tag {1} +$$ + +where $\mathbf{S}_t\left(\mathbf{S}_t^T\mathbf{S}_t^T\right)^{-1}\mathbf{S}_t^T$ is the projection matrix onto the column space of $\mathbf{S}_t$ . We compute the average value of $c_{tj}$ for different values of $j$ , denoted as $c_{t}$ , and report them in Fig. 2c. Notably, when $t \geq 3$ , $c_{t}$ exceeds 0.6—the threshold we hypothesize is sufficient to cover the target concept. Then, we employ the second stage of STEREO to erase two sets of tokens: $\mathcal{S}_1 = \{x_1, x_2, x_3\}$ and $\mathcal{S}_2 = \{x_1, x_2\}$ . We calculated the average Attack Success Rate (ASR) across all concepts, yielding $\mathcal{S}_1$ at $3.21\%$ and $\mathcal{S}_2$ at $85.32\%$ . This indicates STEREO effectively defends against CCE when trained to erase $\mathcal{S}_1$ , but not when only $\mathcal{S}_2$ is erased. An additional experiment on token similarity's impact on erasure is detailed in Appendix B. + +We found that the adversarial tokens identified by TI encompass those detected by UnlearnDiff. Specifically, we evaluated a model trained to erase $S_{1}$ using UnlearnDiff, which yielded an average Attack Success Rate of $1.52\%$ , in + +dicating robustness against UnlearnDiff. To explain this result, we provide an additional justification. Since UnlearnDiff identifies combinations of existing tokens, we leveraged the attention mechanism of the text encoder by using the $\langle \mathrm{EOT} \rangle$ token to represent the target concept, and applied the same technique to all tokens from $S_{1}$ . After this processing, we computed a cosine similarity of 0.78 between the subspace formed by the $\langle \mathrm{EOT} \rangle$ token from $S_{1}$ and the one identified by UnlearnDiff, exceeding 0.6 and supporting our observations. + +In summary, our analysis explains why STEREO can robustly remove the target concept. By applying TI on the concept-erased model iteratively in at least 3 iterations, it can collect a sufficient set of tokens spanning the target concept space. When erasing these tokens together, it successfully eradicates the model knowledge of the concept. + +# 3.2. Reference Subspace Selection for Narrow Concepts + +CA [16] and ESD [7] were the first methods proposed for concept erasure, but they present contradictory approaches. CA maps noise conditioned on the target concept to a more general representation, while ESD pushes the target noise as far as possible from the original. Studies [18, 33] show that CA outperforms ESD in terms of image quality. It is because mapping approach offers a clear direction to erase target concepts, whereas pushing approach can inadvertently affect unrelated concepts. The second stage of STEREO [33], also a pushing approach, significantly degrades image quality. With this motivation, our paper proposes a mapping approach that erases tokens by mapping them to nearby non-target concept subspaces, preserving model utility. The key challenge is how to automatically select these reference subspaces. We also provide experimental results for Subspace Pushing, where the target subspace is simply pushed as far as possible from the original one, in Appendix C. + +
ConceptTI Early TokenTextual Token
R2D2 Robot0.641.23
Garbage Truck1.051.28
Grumpy Cat0.981.22
English Springer Dog0.720.96
+ +Table 1. The Euclid distance of the target token to the tokens from early TI training steps and the textual one. + +We observe that constructing the reference subspace using tokens from the early training steps of TI is effective for narrow concepts, as early tokens typically converge to more general forms (e.g., grumpy cat to cat, or R2D2 robot to robot; see Section 4.3 for details). Moreover, these tokens are closer to the target token learned from TI because they align along the same gradient direction. To validate this, we compare the Euclidean distance between the target token and early TI training tokens, as well as the target token and the textual token, as shown in Table 1. The results indicate that while early tokens may represent concepts similar to the textual token, their distance from the target token is much smaller. This suggests that mapping the target token to an earlier token minimizes weight modifications, thereby better preserving the model's utility. + +# 3.3. Proposed Method + +Our framework, $SuMa$ , consists of two stages, as illustrated in Fig. 3. The first stage, termed Subspace Construction, aims to identify Textual Inversion tokens to establish both sets of Target Subspace $\{\mathcal{U}^i\}$ and Reference Subspace $\{\mathcal{V}^i\}$ . Here, $\mathcal{U}^i$ represents $i$ -th subspace (corresponding to $i$ -th Unet layer) of the concept to be erased, while $\mathcal{V}^i$ is $i$ -th the subspace into which we intend to map our $i$ -th target subspace. In the second stage, Subspace Elimination, we refine the to- $K$ and to- $V$ weights of the Cross-Attention Layer in U-Net, denoted as $W_{k}$ and $W_{v}$ , allowing us to project the basis of $\mathcal{U}^i$ onto $\mathcal{V}^i$ . Due to the identical functionality of $W_{k}$ and $W_{v}$ in our work, we denote both as $W_{k}$ for simplicity and readability. + +Subspace Construction. Starting with the original model, denoted as $\theta$ , and a concept to be removed, $c$ , we apply a Concept Ablation $(\mathcal{M})$ to eliminate $c$ from $\theta$ , resulting in a modified model $\theta'$ . This process is repeated over $l$ steps. At $j$ -th iteration, we apply Textual Inversion (TI) for the concept to be erased from $\theta'$ , yielding two tokens: one from the final step of TI training, denoted as $\langle u\rangle_{j}$ , and another from an earlier step, denoted as $\langle v\rangle_{j}$ . Subsequently, we construct prompts in the formats "A photo of $\langle u\rangle_{j}$ " and "A photo of $\langle v\rangle_{j}$ ", which are then passed through the Text Encoder $\mathcal{T}$ to extract embeddings for the tokens $\langle u_j\rangle$ and $\langle v_j\rangle$ . We then apply the transformation $W_k^i$ , where $i$ represent $i$ -th layer in Unet, to these extracted embeddings, treating the outputs as $\langle u'\rangle_{j}^{i}$ and $\langle v'\rangle_{j}^{i}$ as basis vectors for the subspaces $\mathcal{U}^i$ and $\mathcal{V}^i$ + +respectively. Formally, this process is defined as follows: + +$$ +\begin{array}{l} \langle u ^ {\prime} \rangle_ {j} ^ {i} = W _ {k} ^ {i} \mathcal {T} (p + \langle u \rangle_ {j}), \tag {2} \\ \langle v ^ {\prime} \rangle_ {j} ^ {i} = W _ {k} ^ {i} \mathcal {T} (p + \langle v \rangle_ {j}), \\ \end{array} +$$ + +where $p$ represents the prompt template. Finally, we fine-tune $\theta'$ using $\mathcal{M}$ to remove $\langle u_j\rangle$ , resulting in an updated model $\theta'$ at the end of each iteration (Fig. 3a). + +Subspace Elimination. Fig. 3b illustrates this process. After constructing sets of the subspaces $\{\mathcal{U}^i\}_{i=1}^M$ and $\{\mathcal{V}^i\}_{i=1}^M$ , with $M$ is number of layer in Unet, our goal is to eliminate all $\mathcal{U}^i$ by fine-tuning $W_k$ so that the basis vectors of $\mathcal{U}^i$ are projected onto $\mathcal{V}^i$ . To construct the projection matrix onto $\mathcal{V}^i$ , we first define the matrix $\mathbf{V}^i = [\langle v' \rangle_1^i, \langle v' \rangle_2^i, \dots, \langle v' \rangle_l^i] \in \mathbb{R}^{d \times l}$ , with $d$ is the dimension of token embedding vector, and the projection matrix $\mathbf{P}_{\mathbf{V}^i}$ can then be formulated as: + +$$ +\mathbf {P} _ {\mathbf {V} ^ {i}} = \mathbf {V} ^ {i} \left(\left(\mathbf {V} ^ {i}\right) ^ {T} \mathbf {V} ^ {i}\right) ^ {- 1} \left(\mathbf {V} ^ {i}\right) ^ {T} \tag {3} +$$ + +We define $L_{proj}$ as the objective to project all basis of $\mathcal{U}^i$ onto $\mathcal{V}^i$ by following formula: + +$$ +L _ {p r o j} = \sum_ {i = 1} ^ {M} \sum_ {j = 1} ^ {l} \left\| W _ {k} ^ {i} \mathcal {T} (p + \langle u \rangle_ {j}) - \mathbf {P} _ {\mathbf {V} ^ {i}} W _ {k} ^ {i} \mathcal {T} (p + \langle u \rangle_ {j}) \right\| _ {2} ^ {2} \tag {4} +$$ + +As per [3], we add a regularization term to keep the Reference Subspace $\mathcal{V}^i$ unchanged during fine-tuning: + +$$ +L _ {r e g} = \sum_ {i = 1} ^ {M} \sum_ {j = 1} ^ {l} \left\| W _ {k} ^ {i} \mathcal {T} (p + \langle v \rangle_ {j}) - \mathbf {P} _ {\mathbf {V} ^ {i}} W _ {k} ^ {i} \mathcal {T} (p + \langle v \rangle_ {j}) \right\| \tag {5} +$$ + +The objective function for subspace mapping, $L_{sub}$ , is then defined as the weighted sum of these two terms: + +$$ +L _ {s u b} = L _ {p r o j} + \lambda_ {r e g} L _ {r e g}, \tag {6} +$$ + +where $\lambda_{reg}$ is a hyperparameter (default 1). If we use $L_{sub}$ to fine-tune $\theta$ to obtain $\theta'$ and use CCE to find a new target token, this token still generates the target concept when conditioned on $\theta'$ , but not on $\theta$ (see Appendix D). We assume modifying only the cross-attention layer is effective for erasing a token, as in [18], but affects the SDE trajectory when erasing multiple tokens. To mitigate this, we use the CA loss function and discuss our choice in Section 4.3. In short, their objective is: + +$$ +L _ {\mathrm {C A}} = \mathbb {E} _ {z _ {t}, c, t} \left[ \| \Phi^ {\prime} \left(z _ {t}, \mathcal {T} (p + c ^ {\prime}), t\right) - \Phi \left(z _ {t}, \mathcal {T} (p + c), t\right) \| _ {2} ^ {2} \right] \tag {7} +$$ + +where $\Phi'$ is the model to learn, $\Phi$ is pretrained, $\mathcal{T}$ is TextEncoder, $p$ is prompt template, $c$ is target concept, and $c'$ is its general term (e.g., grumpy cat vs. cat). Our objective is: + +$$ +W _ {k} ^ {*} = \underset {W _ {k}} {\operatorname {a r g m i n}} \left(L _ {\text {f i n a l}}\right) = \underset {W _ {k}} {\operatorname {a r g m i n}} \left(L _ {\mathrm {C A}} + L _ {\text {s u b}}\right). \tag {8} +$$ + +![](images/7b4628f46847a4d261e799125cc900f15c76736d7157e4d9f447c7b24cd804ab.jpg) +(a) Qualitative Comparison across 6 methods for the CCE attack. Both E-CEM methods, CA and MACE, are vulnerable to the CCE attack. STEREO's erasure is too aggressive, leaving only Van Gogh-style elements, with this concept being easier to erase. Our method defends against CCE by mapping the target concept to its reference concept. + +![](images/2aa6741d1bcd3c2a72907200ba9aadbafa3627df1845f08481d8a36ba91da338.jpg) +(b) Qualitative Comparison across 6 methods for image generation. CA struggles with "English Springer" and "Elon Musk." MACE performs well, preserving image quality. RACE shows slight degradation, and STEREO significantly reduces quality. Our method strikes a balance between E-CEMs and C-CEMs in terms of quality. +Figure 4. Qualitative Results. We present qualitative results comparing five methods against the CCE attack and textual erasure. + +# 4. Experiments + +# 4.1. Experimental Setting + +Dataset. We evaluate SuMa across four tasks: subclass erasure, celebrity erasure, artistic style erasure, and instance erasure. For subclass erasure, we focus on a subset of ImageNette [12] known for its distinctive features. Three subclasses are selected for the test, including English Springer Dog, Garbage Truck, and Golf Ball. For celebrity erasure, we select 3 celebrities in the Celebrity-1000 dataset [34], including David Beckham, Elon Musk, and Adam Lambert. For artistic style erasure, following the approach in [7, 15, 33], we focus on the style of Van Gogh, which is both famous and straightforward to verify in outputs. For instance erasure, we conduct experiments on three copyrighted concepts: Grumpy Cat, Mickey Mouse, and the R2D2 robot. Note that our method is specifically designed to address the problem with narrow concepts, so it does not work as well for NSFW concepts. However, as mentioned earlier, our work is complementary to DUO, so it can be easily integrated with this method to completely erase all kinds of concepts, demonstrated in the Appendix E. + +Baseline. We compare SuMa with four baselines: two E-CEMs (CA [16] and MACE [18]) and four C-CEMs (RACE [15], AdvUnlearn (AU) [38], STEREO [33], DUO [21]). For CA, MACE, RACE, and AU, we use their public code and recommended settings to erase concepts. For STEREO and DUO, we re-implement their methods using paper recommendations and verify consistency with reported results. + +Circumventing Methods. We evaluate the robustness of our method under CCE [22] and UnlearnDiff (UD) [39]. For CCE, we use a different set of images from the Subspace Construction stage to ensure the model didn't overfit to this image set. For both attack methods, we use publicly available code to generate adversarial prompts. + +Evaluation Metrics. First, to measure effectiveness, for each erased model, we generate 5,000 images using prompts from the MS-COCO validation set. We then evaluate the model quality when generating such non-target prompts using CLIP scores (higher is better) and FID scores (lower is better). To measure completeness, in the CCE attack, for each erased model, we execute the circumventing method to acquire the adversarial token $c_{adv}$ . We then generate 200 images using a template prompt: "A photo of a $\{c_{adv}\}$ ." For style erasure, we use the template "A painting in the style of $\{c_{adv}\}$ ". In the UD attack, we execute its code on our erased model to get the adversarial prompt and pass the prompt to the erased model. We report the attack success rate (ASR), where a lower score indicates better performance, following the method in [15, 33]. For subclass erasure tasks, we utilize ResNet-50 [10], pretrained on ImageNet [27], to measure ASR. For celebrity detection, we use the GIPHY Celebrity Detector (GCD) [9] as the classifier. To evaluate artistic style, we follow the approach in [15] by fine-tuning a ViT-based model [1] on the WikiArt dataset [29]. Finally, for instance detection, due to the lack of pretrained models or datasets, we use LLAVA-7B [17] with a template prompt: "Does this image contain {object}? Answer Yes or No." for evaluation. + +Implementation Details. We use SD1.4 and SD1.5 as the pretrained model. In all experiments, we fine-tune all cross-attention layers. In the Subspace Construction stage, we use the same setting as CA [16] for concept erasion. Training TI is set for 500 steps, with the early step to get token $\langle v\rangle$ set to 50, the initial token is toy, with Adam optimization with learning rate, is 2e-6. For the Subspace Elimination stage, we fine-tune most objects for 500 steps, except for subclass objects, which require 750 steps to ensure model safety. We use the Adam optimizer with a learning rate of 2e-5. All experiments are done on 2 NVIDIA A100 GPUs. + +
MethodCCE↓UD↓FID↓CLIP↑
Subclass
SD [24]0.84 / 0.830.74 / 0.7617.04 / 16.950.33 / 0.32
CA [16]0.78 / 0.750.67 / 0.6219.28 / 19.260.32 / 0.32
MACE [18]0.86 / 0.840.73 / 0.6817.15 / 16.980.29 / 0.31
RACE [15]0.74 / 0.730.17 / 0.1826.34 / 26.220.27 / 0.28
AdvUnlearn [38]0.76 / 0.740.18 / 0.1618.47 / 18.390.30 / 0.31
DUO [21]0.65 / 0.630.62 / 0.6517.09 / 16.990.29 / 0.30
STEREO [33]0.05 / 0.030.05 / 0.0427.33 / 27.200.30 / 0.28
Ours0.12 / 0.100.08 / 0.0618.33 / 18.150.31 / 0.30
Identity
SD [24]0.91 / 0.900.86 / 0.8417.04 / 16.950.33 / 0.32
CA [16]0.87 / 0.890.79 / 0.7618.21 / 18.270.31 / 0.30
MACE [18]0.89 / 0.870.65 / 0.6316.74 / 16.690.29 / 0.28
RACE [15]0.78 / 0.790.47 / 0.4624.32 / 24.170.27 / 0.29
AdvUnlearn [38]0.77 / 0.690.52 / 0.5317.53 / 17.690.31 / 0.31
DUO [21]0.73 / 0.710.67 / 0.7317.32 / 17.110.29 / 0.30
STEREO [33]0.02 / 0.010.03 / 0.0126.33 / 26.150.29 / 0.28
Ours0.09 / 0.070.18 / 0.1417.94 / 17.890.31 / 0.29
+ +
MethodCCE↓UD↓FID↓CLIP↑
Artistic
SD [24]0.86 / 0.830.82 / 0.8517.04 / 16.950.33 / 0.32
CA [16]0.79 / 0.780.74 / 0.7618.02 / 17.850.34 / 0.31
MACE [18]0.84 / 0.820.79 / 0.8316.57 / 16.580.32 / 0.33
RACE [15]0.82 / 0.790.06 / 0.0421.32 / 21.100.30 / 0.31
AdvUnlearn [38]0.83 / 0.760.03 / 0.0517.73 / 17.600.32 / 0.33
DUO [21]0.74 / 0.720.67 / 0.7317.74 / 17.890.30 / 0.31
STEREO [33]0.00 / 0.020.00 / 0.0128.04 / 27.900.30 / 0.31
Ours0.05 / 0.040.07 / 0.0517.18 / 17.210.33 / 0.31
Instance
SD [24]0.97 / 0.940.94 / 0.9317.04 / 16.950.33 / 0.32
CA [16]0.97 / 0.930.92 / 0.8918.43 / 18.300.30 / 0.31
MACE [18]0.97 / 0.920.93 / 0.8916.93 / 16.800.30 / 0.31
RACE [15]0.95 / 0.920.06 / 0.0724.32 / 24.150.28 / 0.29
AdvUnlearn [38]0.95 / 0.930.03 / 0.0418.37 / 18.200.32 / 0.33
DUO [21]0.92 / 0.910.89 / 0.8718.31 / 18.920.30 / 0.31
STEREO [33]0.03 / 0.020.05 / 0.0327.04 / 26.900.31 / 0.32
Ours0.17 / 0.150.19 / 0.1622.34 / 22.200.29 / 0.30
+ +Table 2. Evaluation of Erasing across 4 Categories. We average the results of all target concepts in each category across three metrics and compare our framework, SuMa, with state-of-the-art methods. For ASRs, successful attacks (ASR ≥ 20%) are in red, and defeated attacks (ASR < 20%) in green. The data in each cell show the result of erasing a concept from pretrained SD1.4 / SD1.5. + +
Reference SubspaceSubclassIdentityArtisticInstance
ASR ↓FID ↓CLIP ↑ASR ↓FID ↓CLIP ↑ASR ↓FID ↓CLIP ↑ASR ↓FID ↓CLIP ↑
Step-300.0919.110.300.0818.340.290.0518.820.310.1723.960.27
Step-500.1218.330.310.0917.940.310.0517.180.330.1722.340.29
Step-700.5617.320.330.6717.920.320.3717.050.330.8317.740.32
General Term0.1818.270.320.1718.020.320.1118.210.330.1523.340.27
+ +Table 3. Ablation Study for the selection of Reference subspace. + +![](images/dd3e639ca00e25b5659003905736140f3dce21d76f1ff1da0abb65506adfa83a.jpg) +Figure 5. Visualization of Textual Inversion tokens at different numbers of TI training steps. + +# 4.2. Results + +Quantitative Results. We average each metric from Section 4.1 for each category and report the results in Table 2. For all baselines, we ensure the erased model works + +well for erasing the textual prompt, and we report results for this trivial task in the Appendix F. For completeness evaluation, our method achieves an ASR score below 0.2 for all categories under both CCE and UD attacks. Compared to STEREO, our method has a slightly higher ASR score (0.02 to 0.14 higher), but within an acceptable range. RACE and AdvUnlearn show safety against the UD attack in the Subclass, Artistic, and Instance categories but fail under Identity. This is due to the subtle facial features of human identity, which make this category more fine-grained and challenging. However, both fail under the CCE attack, showing that protecting the model against CCE is harder than UD. The main reason is that RACE and AdvUnlearn erase adversarial tokens based on UnlearnDiff, but these tokens don't fully represent the target concept, protecting against UD but not CCE. DUO works well for NSFW concepts (results in Appendix E) but fails for narrow concepts. For effectiveness, although STEREO performs well under both CCE and UD, image quality is significantly affected, resulting in an FID score roughly double that of other E-CEMs, making it less effective. Our method, while not achieving a perfect ASR, achieves an FID score comparable + +![](images/1b0defb8e6523aef785ed64195ffacfaaa9f6ad964b3665f3d56b629920eb3b1.jpg) +Figure 6. Multiple Concept Erasure Result. + +
RegulizationSubclassIdentityArtisticInstance
×ASR: 0.09ASR: 0.06ASR: 0.05ASR: 0.15
FID: 19.21FID: 20.31FID: 19.32FID: 24.57
CLIP: 0.29CLIP: 0.29CLIP: 0.30CLIP: 0.26
✓ASR: 0.12ASR: 0.09ASR: 0.05ASR: 0.17
FID: 18.33FID: 17.94FID: 17.18FID: 22.34
CLIP: 0.31CLIP: 0.31CLIP: 0.33CLIP: 0.29
+ +Table 4. Ablation Study for the need of Subspace Regulization. + +
EntityFID ↓CCE ↓UD ↓
English Springer20.12 / 18.140.12 / 0.470.07 / 0.34
Van Gogh20.16 / 17.540.07 / 0.520.04 / 0.46
Elon Musk20.15 / 18.220.07 / 0.490.05 / 0.57
Grumpy Cat24.57 / 20.320.18 / 0.540.14 / 0.46
+ +Table 5. Ablation Study for the base method (ESD / UCE). + +to other E-CEMs in the Subclass, Artistic, and Identity categories, outperforming CA with an FID around 0.85 lower. We found that Instance category is harder to be erased, so we fine-tuned it for 750 steps, resulting in a higher FID compared to the others, fine-tuned for 500 steps. + +Qualitative Results are provided in Fig. 4. Fig. 4a shows results of these models against the CCE attack, and we include qualitative results for UD in the Appendix J. CA, MACE, RACE, and AdvUnlearn are highly vulnerable to this attack. STEREO completely removes these concepts, but at the cost of image quality ( $6^{th}$ column in Fig. 4b). Our method successfully converts "Springer Dog" to "Dog," "Elon Musk" to a general person, and "Grumpy Cat" to "Cat toy," while maintaining image quality comparable to other E-CEM methods (last column in Fig. 4b). + +Multiple Concepts Erasure. We provide results for erasing multiple concepts in Fig. 6. We test the ASR with the CCE attack and evaluate two settings: Iterative Erase (IE) and Simultaneous Erase (SE). IE means we fine-tune the weights after erasing one concept to erase others. SE sums all $L_{sub}$ to erase all concepts in one fine-tuning, with 1000 steps fine-tuning for this experiment. In both settings, our method defends against CCE and maintains model utility. With SE, we achieve better FID scores, as summing $L_{sub}$ preserves a broader range of knowledge from $L_{reg}$ . + +# 4.3. Ablation Studies + +Effect of selecting the Reference Subspace. We present results for four ways of constructing the Reference Subspace, including TI tokens from Steps 30, 50, 70 of TI training, and the concept's general term, as shown in Table 3. Visualizations of each token are in Fig. 5. At Step 30, the token is far from the target, with little relevant information. At Step 50, the token shares features with the target but is not an exact match, enabling minimal transformation. At Step 70, the token is close to the target, making full erasure difficult. These observations align with Table 3, where Step 50 balances ASR, FID, and CLIP scores, while Step 30 improves ASR but slightly reduces image quality. Step 70 shows results similar to CA, indicating SuMa's inefficacy when the Reference Subspace and Target Subspace are too close. + +Effect of choosing the base method. As noted in Section 3.3, we adopt CA loss for our method. Results using ESD [7] and UCE [8] as base methods are shown in Table 5. While ESD resembles CA, it uses a pushing strategy, whereas CA employs a mapping strategy. UCE erases concepts by altering the cross-attention map. Our experiments show that modifying only the attention layer causes the token learned by CCE to express the target concept only under the erased model as shown in Appendix D. ESD provides some defense but increases FID moderately, while UCE fails to defend. We conclude that modifying the model's noise prediction is more effective, as it redirects the target concept's trajectory toward another in the original model. The mapping approach is also favored over the pushing one, as discussed in Section 3.2 and supported by our results. + +Effect of adding Subspace Regularization. We present the effect of adding Subspace Regularization and report the results in Table 4. We observe that without regularization, our method performs slightly better against the CCE attack, with an average $2\%$ reduction in attack rate, but at the cost of a gradual increase in the FID score. This phenomenon was studied in [3], where it was shown that the concept closest to the target is the most affected. Therefore, we conclude that adding Subspace Regularization loss is necessary to maintain the model's effectiveness. + +# 5. Conclusion and Limitations + +In this paper, we introduced SuMa, a method that robustly and effectively erases narrow concepts in Text-to-Image Diffusion models. SuMa constructs a Target Subspace for the concept to be erased and maps it to a Reference Subspace. Experimental results show our method is the first to meet these dual objectives for narrow concepts. A key limitation is its time cost, as it may require many iterations to construct both subspaces. Future work will aim to accelerate this process and improve target concept representation. + +# References + +[1] Dosovitskiy Alexey. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv: 2010.11929, 2020. 6 +[2] Anh Bui, Khanh Doan, Trung Le, Paul Montague, Tamas Abraham, and Dinh Phung. Removing undesirable concepts in text-to-image generative models with learnable prompts. arXiv preprint arXiv:2403.12326, 2024. 1 +[3] Anh Bui, Long Vuong, Khanh Doan, Trung Le, Paul Montague, Tamas Abraham, and Dinh Phung. Erasing undesirable concepts in diffusion models with adversarial preservation. arXiv preprint arXiv:2410.15618, 2024. 5, 8 +[4] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, José Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. In Proceedings of the 40th International Conference on Machine Learning, pages 4055-4075, 2023. 1 +[5] Zhi-Yi Chin, Chieh-Ming Jiang, Ching-Chun Huang, PinYu Chen, and Wei-Chen Chiu. Prompting4debugging: Redteaming text-to-image diffusion models by finding problematic prompts. arXiv preprint arXiv:2309.06135, 2023. 3 +[6] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel CohenOr. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 1, 3 +[7] Rohit Gandikota, Joanna Materzynska, Jaden Fiitto-Kaufman, and David Bau. Erasing concepts from diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2426-2436, 2023. 1, 2, 4, 6, 8 +[8] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5111-5120, 2024. 8 +[9] Nick Hasty, Ihor Kroosh, Dmitry Voitekh, and Dmytro Korduban. Giphy celebrity detector, 2022. 6 +[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6 +[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2 +[12] Jeremy Howard and Sylvain Gugger. Fastai: a layered api for deep learning. Information, 11(2):108, 2020. 6 +[13] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2 +[14] Harry H Jiang, Lauren Brown, Jessica Cheng, Mehtab Khan, Abhishek Gupta, Deja Workman, Alex Hanna, Johnathan Flowers, and Timnit Gebru. Ai art and its impact on artists. In Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society, pages 363-374, 2023. 1 + +[15] Changhoon Kim, Kyle Min, and Yezhou Yang. Race: Robust adversarial concept erasure for secure text-to-image diffusion model. arXiv preprint arXiv:2405.16341, 2024. 1, 3, 6, 7, 4, 5 +[16] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22691-22702, 2023. 1, 2, 3, 4, 6, 7, 5 +[17] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 6 +[18] Shilin Lu, Zilan Wang, Leyang Li, Yanzhu Liu, and Adams Wai-Kin Kong. Mace: Mass concept erasure in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6430-6440, 2024. 1, 2, 4, 5, 6, 7 +[19] Mengyao Lyu, Yuhong Yang, Haiwen Hong, Hui Chen, Xuan Jin, Yuan He, Hui Xue, Jungong Han, and Guiguang Ding. One-dimensional adapter to rule them all: Concepts diffusion models and erasing applications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7559-7568, 2024. 1, 2 +[20] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 1 +[21] Yong-Hyun Park, Sangdoo Yun, Jin-Hwa Kim, Junho Kim, Geonhui Jang, Yonghyun Jeong, Junghyo Jo, and Gayoung Lee. Direct unlearning optimization for robust and safe text-to-image models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2, 3, 6, 7, 4, 5 +[22] Minh Pham, Kelly O Marshall, Niv Cohen, Govind Mittal, and Chinmay Hegde. Circumventing concept erasure methods for text-to-image generative models. In The Twelfth International Conference on Learning Representations, 2023. 1, 2, 3, 6 +[23] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1 +[24] Robin Rombach. Stable diffusion v1-4 model card., 2022. 1, 7, 4, 5 +[25] Robin Rombach. Stable diffusion 2.0 release., 2022. 1 +[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1 +[27] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 6 + +[28] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1 +[29] Babak Saleh and Ahmed Elgammal. Large-scale classification of fine-art paintings: Learning the right metric on the right feature. arxiv 2015. arXiv preprint arXiv:1505.00855, 2015.6 +[30] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 1 +[31] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by {Text-to-Image} models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 2187–2204, 2023. 1 +[32] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6048-6058, 2023. 1 +[33] Koushik Srivatsan, Fahad Shamshad, Muzammal Naseer, Vishal M Patel, and Karthik Nandakumar. Stereo: A two-stage framework for adversarially robust concept erasing from text-to-image diffusion models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 23765-23774, 2025. 2, 3, 4, 6, 7, 5 +[34] Tonyassi. Celebrity-1000 datasets, 2022. 6 +[35] Yu-Lin Tsai, Chia-Yi Hsu, Chulin Xie, Chih-Hsun Lin, Jia-You Chen, Bo Li, Pin-Yu Chen, Chia-Mu Yu, and Chun-Ying Huang. Ring-a-bell! how reliable are concept removal methods for diffusion models? arXiv preprint arXiv:2310.10012, 2023. 1, 2, 3 +[36] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228-8238, 2024. 2 +[37] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research, 2021. 1 +[38] Yimeng Zhang, Xin Chen, Jinghan Jia, Yihua Zhang, Chongyu Fan, Jiancheng Liu, Mingyi Hong, Ke Ding, and Sijia Liu. Defensive unlearning with adversarial training for robust concept erasure in diffusion models. arXiv preprint arXiv:2405.15234, 2024. 2, 3, 6, 7, 4, 5 +[39] Yimeng Zhang, Jinghan Jia, Xin Chen, Aochuan Chen, Yihua Zhang, Jiancheng Liu, Ke Ding, and Sijia Liu. To generate or not? safety-driven unlearned diffusion models are still + +easy to generate unsafe images... for now. In European Conference on Computer Vision, pages 385-403. Springer, 2025. 1, 2, 3, 6 \ No newline at end of file diff --git a/2025/SuMa_ A Subspace Mapping Approach for Robust and Effective Concept Erasure in Text-to-Image Diffusion Models/images.zip b/2025/SuMa_ A Subspace Mapping Approach for Robust and Effective Concept Erasure in Text-to-Image Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6d4450f4ef1a67e621d16341087ec9910c638c68 --- /dev/null +++ b/2025/SuMa_ A Subspace Mapping Approach for Robust and Effective Concept Erasure in Text-to-Image Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b0fbcb08f50d0aee663459c06a63e2cb138775aab0854e3fac52df9e3f0db4c3 +size 686574 diff --git a/2025/SuMa_ A Subspace Mapping Approach for Robust and Effective Concept Erasure in Text-to-Image Diffusion Models/layout.json b/2025/SuMa_ A Subspace Mapping Approach for Robust and Effective Concept Erasure in Text-to-Image Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5749b57a439989b4e1dcdbe25e460bcaaf221a83 --- /dev/null +++ b/2025/SuMa_ A Subspace Mapping Approach for Robust and Effective Concept Erasure in Text-to-Image Diffusion Models/layout.json @@ -0,0 +1,8869 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 79, + 102, + 532, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 102, + 532, + 138 + ], + "spans": [ + { + "bbox": [ + 79, + 102, + 532, + 138 + ], + "type": "text", + "content": "SuMa: A Subspace Mapping Approach for Robust and Effective Concept Erasure in Text-to-Image Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 175, + 160, + 434, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 160, + 434, + 175 + ], + "spans": [ + { + "bbox": [ + 175, + 160, + 434, + 175 + ], + "type": "text", + "content": "Kien Nguyen1 Anh Tran1 Cuong Pham2,1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 244, + 175, + 366, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 175, + 366, + 189 + ], + "spans": [ + { + "bbox": [ + 244, + 175, + 366, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 244, + 175, + 366, + 189 + ], + "type": "text", + "content": "Qualcomm AI Research" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 153, + 190, + 457, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 190, + 457, + 203 + ], + "spans": [ + { + "bbox": [ + 153, + 190, + 457, + 203 + ], + "type": "text", + "content": "2Posts and Telecommunications Institute of Technology (PTIT)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 196, + 205, + 411, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 205, + 411, + 217 + ], + "spans": [ + { + "bbox": [ + 196, + 205, + 411, + 217 + ], + "type": "text", + "content": "{kienn, anhtra, pcuong}@qti.qualcomm.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 271, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 271, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 56, + 271, + 295, + 605 + ], + "type": "text", + "content": "The rapid growth of text-to-image diffusion models has raised concerns about their potential misuse in generating harmful or unauthorized contents. To address these issues, several Concept Erasure methods have been proposed. However, most of them fail to achieve both robustness, i.e., the ability to robustly remove the target concept, and effectiveness, i.e., maintaining image quality. While few recent techniques successfully achieve these goals for NSFW concepts, none could handle narrow concepts such as copyrighted characters or celebrities. Erasing these narrow concepts is critical in addressing copyright and legal concerns. However, erasing them is challenging due to their close distances to non-target neighboring concepts, requiring finer-grained manipulation. In this paper, we introduce Subspace Mapping (SuMa), a novel method specifically designed to achieve both robustness and effectiveness in easing these narrow concepts. SuMa first derives a target subspace representing the concept to be erased and then neutralizes it by mapping it to a reference subspace that minimizes the distance between the two. This mapping ensures the target concept is robustly erased while preserving image quality. We conduct extensive experiments with SuMa across four tasks: subclass erasure, celebrity erasure, artistic style erasure, and instance erasure and compare the results with current state-of-the-art methods. Our method achieves image quality comparable to approaches focused on effectiveness, while also yielding results that are on par with methods targeting completeness." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 632, + 135, + 644 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 632, + 135, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 632, + 135, + 644 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Text-to-image models (T2I) at a large scale have achieved notable success in producing lifelike images [4, 20, 23, 26, 28, 37]. This progress is not only a result of enhanced algorithms and computing resources but also the extensive datasets gathered from the Internet [30]. Unfortunately," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 246, + 553, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 553, + 281 + ], + "type": "text", + "content": "these datasets often contain copyrighted works, artistic creations [14, 31, 32], NSFW content, and personal images, raising ethical and legal concerns [18]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 282, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 282, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 312, + 282, + 555, + 437 + ], + "type": "text", + "content": "Some publicly available models, like Stable Diffusion (SD) [24], trained on vast web-scraped datasets, can generate inappropriate content, including copyrighted images, offensive material, and deepfakes, raising serious community concerns. To address this, researchers have attempted to filter out such content, resulting in the release of Stable Diffusion 2.0 [25]. However, this version, despite the filtering efforts, delivers lower image quality than its predecessor, with some explicit content still slipping through. Another approach involves post-processing filters, such as NSFW detectors, which flag inappropriate outputs and replace them with blank images. Yet, these safeguards can be easily bypassed once the model is publicly accessible [18]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 437, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 437, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 312, + 437, + 554, + 653 + ], + "type": "text", + "content": "Recent research has concentrated on fine-tuning text-to-image (T2I) diffusion models to 'forget' inappropriate content, giving rise to Concept Erased Models (CEMs) as a practical approach to mitigate such issues. The first generation of these models, which we term Effective Concept Erased Models (E-CEMs), demonstrates success in removing inappropriate content while maintaining image quality comparable to the original outputs [2, 7, 16, 18, 19]. However, recent studies have uncovered significant vulnerabilities in these methods [22, 35, 39]. For instance, Unlearn-Diff [39] identifies combinations of existing tokens within current models capable of reconstructing the target concept, effectively bypassing the erasing mechanism. Additionally, CCE [22] leverages Textual Inversion [6] to derive a new token that regenerates the erased concept. Collectively, these findings suggest that E-CEMs fail to robustly eliminate target concepts from the model's latent space, leaving them susceptible to regeneration through adversarial techniques." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "The second generation, which we call Robust Concept Erased Models (R-CEMs), addresses vulnerabilities revealed by adversarial attacks. Early approaches, like RACE [15], use an iterative framework based on E-CEM, identifying adversarial tokens and erasing them. AdvUnlearn" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19587" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "text", + "content": "[38] improves RACE by adding a prior preservation loss to retain model utility. However, while these methods mitigate UnlearnDiff attacks, they struggle against CCE. This is because CCE generates new tokens, while UnlearnDiff relies on existing token combinations, limiting its versatility compared to CCE's broader search space. We provide a discussion about these two attacking methods in Appendix A. More recent methods, such as STEREO [33], refine this process by identifying and erasing all adversarial tokens detected by CCE simultaneously. While it demonstrates notable success in terms of robustness and resilience against current attack methods, it incurs a significant reduction in image quality. In contrast, DUO [21] first applies an image editing approach to modify inappropriate content, then leverages the Diffusion-DPO framework [36] to assign high scores to appropriate images and low scores to inappropriate ones while also introducing a novel prior preservation loss. This method successfully protects the model from existing attacks while maintaining image quality. However, it is currently limited to erasing NSFW content." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 315, + 296, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 296, + 507 + ], + "type": "text", + "content": "While NSFW content is a common target of concept erasure, narrow concepts are not less important targets. Here, we use the terms \"narrow concepts\" to refer to fine-grained categories or specific instances within, such as \"English Springer\", \"Grumpy Cat\", or \"Elon Musk\". The ability to erase narrow concepts is critical, as they may involve copyright issues or contribute to the spread of misleading information. While DUO is the only robust and effective concept erasure technique, it struggles to handle narrow concepts. We discuss in more detail the reason in the Appendix A. In short, this method relies on an editing approach to erase concepts. For instance, it edits the \"nudity\" concept to \"dressed in\", where the difference between the two is clear. However, for narrow concepts, such as editing \"Grumpy Cat\" to \"Cat\", it is harder to classify these as distinct classes, requiring a finer-grained method to address." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": "Given the aforementioned challenges, our paper presents complementary work to DUO [21], focusing on achieving a balance between robust and effective erasure for narrow concepts. We first investigate why adversarial tokens can bypass E-CEMs and analyze how STEREO achieves robustness against such attacks. We then introduce our method, Subspace Mapping Concept Erasure (SuMa). Our approach constructs both a target subspace containing the concept to be erased and a reference subspace that is independent yet close to the target subspace by identifying their basis vectors. We eliminate the target subspace by fine-tuning the model to optimize a novel objective function, aiming to project the basis of the target subspace onto the reference subspace. This method is integrated with E-CEM techniques, making the model robust against adversarial attacks while still preserving image quality. As a result, SuMa achieves image quality on par with E-CEMs in" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 320, + 71, + 371, + 121 + ], + "blocks": [ + { + "bbox": [ + 320, + 71, + 371, + 121 + ], + "lines": [ + { + "bbox": [ + 320, + 71, + 371, + 121 + ], + "spans": [ + { + "bbox": [ + 320, + 71, + 371, + 121 + ], + "type": "image", + "image_path": "277078d12c1231060298c8585bf5d7f5a46762c3cfaa1c2c78092f860fc90230.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 327, + 124, + 364, + 129 + ], + "lines": [ + { + "bbox": [ + 327, + 124, + 364, + 129 + ], + "spans": [ + { + "bbox": [ + 327, + 124, + 364, + 129 + ], + "type": "text", + "content": "AdvUnlearn" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 376, + 71, + 427, + 121 + ], + "blocks": [ + { + "bbox": [ + 376, + 71, + 427, + 121 + ], + "lines": [ + { + "bbox": [ + 376, + 71, + 427, + 121 + ], + "spans": [ + { + "bbox": [ + 376, + 71, + 427, + 121 + ], + "type": "image", + "image_path": "be61c490d799cfcb718ca6bf6d9c101fc4368d3b8b2c249bb307993d3fd0fb4e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 124, + 408, + 129 + ], + "lines": [ + { + "bbox": [ + 392, + 124, + 408, + 129 + ], + "spans": [ + { + "bbox": [ + 392, + 124, + 408, + 129 + ], + "type": "text", + "content": "RACE" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 436, + 71, + 488, + 122 + ], + "blocks": [ + { + "bbox": [ + 436, + 71, + 488, + 122 + ], + "lines": [ + { + "bbox": [ + 436, + 71, + 488, + 122 + ], + "spans": [ + { + "bbox": [ + 436, + 71, + 488, + 122 + ], + "type": "image", + "image_path": "a8a405c84350eda90f0268bf14bbcfff8e0aba93b3ea9c34af118779fd5b25a6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 124, + 471, + 129 + ], + "lines": [ + { + "bbox": [ + 449, + 124, + 471, + 129 + ], + "spans": [ + { + "bbox": [ + 449, + 124, + 471, + 129 + ], + "type": "text", + "content": "STEREO" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 140, + 555, + 206 + ], + "lines": [ + { + "bbox": [ + 313, + 140, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 555, + 206 + ], + "type": "text", + "content": "Figure 1. Illustration of current C-CEMs methods trained to erase garbage truck against the CCE Attack. We can see that AdvUnlearn and RACE are still vulnerable to the CCE Attack, while STEREO successfully defends against it, though at the cost of image quality. Our method successfully balances Completeness and Effectiveness, changing garbage truck into Semi-trailer truck." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 498, + 72, + 549, + 121 + ], + "blocks": [ + { + "bbox": [ + 498, + 72, + 549, + 121 + ], + "lines": [ + { + "bbox": [ + 498, + 72, + 549, + 121 + ], + "spans": [ + { + "bbox": [ + 498, + 72, + 549, + 121 + ], + "type": "image", + "image_path": "335b486836d33d5aa32c3b5238c712da9740f8cfe5c09d4f19349c15e7f93258.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 515, + 124, + 529, + 131 + ], + "lines": [ + { + "bbox": [ + 515, + 124, + 529, + 131 + ], + "spans": [ + { + "bbox": [ + 515, + 124, + 529, + 131 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 219, + 553, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 219, + 553, + 254 + ], + "spans": [ + { + "bbox": [ + 313, + 219, + 553, + 254 + ], + "type": "text", + "content": "terms of FID and CLIP-score, and an Attack Success Rate (ASR) comparable to the state-of-the-art R-CEM method, STEREO. Our main contributions are as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 255, + 553, + 459 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 314, + 255, + 553, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 255, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 314, + 255, + 553, + 277 + ], + "type": "text", + "content": "- We analyze why adversarial tokens can bypass E-CEMs and how STEREO remains resilient against these attacks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 279, + 553, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 553, + 374 + ], + "type": "text", + "content": "- We propose the Subspace Mapping Concept Erasure (SuMa) framework, which balances robustness and effectiveness in erasing narrow concepts. Our framework first constructs a target subspace that captures the concept to be erased and a reference subspace that is distinct yet proximate to the target. We then introduce a novel objective function that removes the target concept by mapping the target subspace onto the reference subspace." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 375, + 553, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 375, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 314, + 375, + 553, + 459 + ], + "type": "text", + "content": "- Extensive experiments across four tasks—object erasure, celebrity erasure, artistic style erasure, and subclass erasure—demonstrate that SuMa offers the best trade-off between completeness and effectiveness. SuMa completely removes the target concept while preserving model utility. To the best of our knowledge, we are the first to achieve these dual objectives for narrow concepts." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 472, + 400, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 472, + 400, + 483 + ], + "spans": [ + { + "bbox": [ + 314, + 472, + 400, + 483 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 491, + 511, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 491, + 511, + 504 + ], + "spans": [ + { + "bbox": [ + 314, + 491, + 511, + 504 + ], + "type": "text", + "content": "2.1. Effective Concept Erasure (E-CEMs)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 510, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 715 + ], + "type": "text", + "content": "The first generation of the Concept Erasure Models (CEMs) was designed to fine-tune original models to forget inappropriate content, successfully removing unwanted elements while closely maintaining image quality. Concept Ablation (CA) [16] finetunes the model to redirect the target prompt, e.g., \"grumpy cat\", to be similar as a general term, e.g., \"cat\", by aligning the noise predictions on these prompts. ESD [7] pushed the noise predictions on the target concept away from the original noise predictions by reversing the direction of conditional guidance [11]. SPM [19] focused on preserving non-target concepts during erasure by applying the CA method with added regularization that maintained noise predictions on the closet concept in the CLIP Embedding Space. MACE [18] proposes a LORA-Finetuning [13] method to erase multiple concepts simultaneously. Despite these advancements, recent adversarial attacks [22, 35, 39] successfully recovered the \"erased\" concepts, exposing vul" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19588" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 75, + 218, + 169 + ], + "blocks": [ + { + "bbox": [ + 59, + 75, + 218, + 169 + ], + "lines": [ + { + "bbox": [ + 59, + 75, + 218, + 169 + ], + "spans": [ + { + "bbox": [ + 59, + 75, + 218, + 169 + ], + "type": "image", + "image_path": "165ae80020bdf0b85fa1d0f09a70dfdf69a858ae647698a5420c6fff71bd863a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 172, + 202, + 182 + ], + "lines": [ + { + "bbox": [ + 71, + 172, + 202, + 182 + ], + "spans": [ + { + "bbox": [ + 71, + 172, + 202, + 182 + ], + "type": "text", + "content": "(a) Cosine similarity of TIs from SD1.4." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 224, + 74, + 386, + 169 + ], + "blocks": [ + { + "bbox": [ + 224, + 74, + 386, + 169 + ], + "lines": [ + { + "bbox": [ + 224, + 74, + 386, + 169 + ], + "spans": [ + { + "bbox": [ + 224, + 74, + 386, + 169 + ], + "type": "image", + "image_path": "4474b9c98fc92f7b550a34fb29b445e6eb8801e5663dcb6eb19ece8d6db33e42.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 186, + 555, + 209 + ], + "lines": [ + { + "bbox": [ + 55, + 186, + 555, + 209 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 555, + 209 + ], + "type": "text", + "content": "Figure 2. Distribution of cosine similarity scores among TIs in three cases: from the same SD1.4 but different TI training steps, different erasure levels of CA, and between a TI at erasure level " + }, + { + "bbox": [ + 55, + 186, + 555, + 209 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 55, + 186, + 555, + 209 + ], + "type": "text", + "content": " and the subspace created by the " + }, + { + "bbox": [ + 55, + 186, + 555, + 209 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 186, + 555, + 209 + ], + "type": "text", + "content": " previous TIs." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 391, + 72, + 536, + 169 + ], + "blocks": [ + { + "bbox": [ + 223, + 172, + 384, + 182 + ], + "lines": [ + { + "bbox": [ + 223, + 172, + 384, + 182 + ], + "spans": [ + { + "bbox": [ + 223, + 172, + 384, + 182 + ], + "type": "text", + "content": "(b) Cosine similarity of TIs at different CA levels." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 391, + 72, + 536, + 169 + ], + "lines": [ + { + "bbox": [ + 391, + 72, + 536, + 169 + ], + "spans": [ + { + "bbox": [ + 391, + 72, + 536, + 169 + ], + "type": "image", + "image_path": "090dd577857f4242ee2266246ba8b5825c876b252cbc7c8136663f35b22cb4c5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 172, + 539, + 182 + ], + "lines": [ + { + "bbox": [ + 402, + 172, + 539, + 182 + ], + "spans": [ + { + "bbox": [ + 402, + 172, + 539, + 182 + ], + "type": "text", + "content": "(c) Cosine similarity of TIs with subspace." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 224, + 295, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 295, + 249 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 295, + 249 + ], + "type": "text", + "content": "nerabilities in all these methods and indicating a need for further improvements in concept erasure techniques." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 254, + 231, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 254, + 231, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 231, + 266 + ], + "type": "text", + "content": "2.2. Circumventing Concept Erasure" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 271, + 295, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 271, + 295, + 451 + ], + "spans": [ + { + "bbox": [ + 54, + 271, + 295, + 451 + ], + "type": "text", + "content": "Several methods have been proposed to attack E-CEMs [22, 35, 39]. P4D [5] enhances adversarial prompts by ensuring that noise estimates between pre-trained and concept-erased models are similar. UnlearnDiff [39] is similar to P4D but leverages the classification capabilities of diffusion models, removing the reliance on pre-trained models. Ring-A-Bell [35] creates adversarial prompts by guiding a \"safe\" prompt toward an \"unsafe\" one in the text vector space. Among these, CCE [22] is the simplest and most effective, employing Textual Inversion (TI) [6] to derive a new token from the E-CEM, which regenerates the erased concept. Unlike UnlearnDiff, which relies on combinations of existing tokens, CCE searches for a new token in the latent text space, greatly widening its search space. Hence, defending against CCE is far more challenging than against UnlearnDiff." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 456, + 246, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 246, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 246, + 469 + ], + "type": "text", + "content": "2.3. Robust Concept Erasure (R-CEMs)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "text", + "content": "Motivated by the need to protect CEMs from adversarial attacks, several methods have emerged as part of the second generation of CEMs [15, 33, 38]. RACE [15] proposes an iterative framework that first identifies adversarial tokens using the UnlearnDiff framework but with faster optimization. It then erases the found tokens using a loss function from ESD, providing protection against UnlearnDiff. However, as shown in Fig. 1, it remains vulnerable to more advanced CCE attacks. AdvUnlearn [38] is similar to RACE, but it introduces a prior preservation loss during the erasing phase to maintain the model's utility. Still, it is susceptible to CCE attacks, as illustrated in Fig. 1. STEREO [33] successfully defends against all existing attack methods by iteratively identifying all Textual Inversions (TIs) from the concept-erased model, and erasing all of them simultaneously from the original model. However, this method leads to a significant degradation in image quality, resulting in an FID nearly double that of E-CEMs. DUO [21] addresses both robustness and effectiveness by leveraging Diffusion-DPO to assign high scores to concept-edited images and" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 224, + 555, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 555, + 261 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 555, + 261 + ], + "type": "text", + "content": "low scores to inexact ones. It also introduces a novel prior preservation loss to preserve the model's utility. However, this approach is currently limited to NSFW concepts." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 271, + 370, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 271, + 370, + 284 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 370, + 284 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 292, + 555, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 292, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 555, + 425 + ], + "type": "text", + "content": "In this section, we begin by exploring whether E-CEMs, trained to erase a single token, can also erase tokens that differ in form but convey the same concept (Sec. 3.1). This analysis sets the stage for understanding why CCE [22] can attack E-CEMs and why STEREO [33] can robustly erase a target concept. We then discuss how to find effective reference subspaces for erasing narrow concepts (Sec. 3.2), as well as the motivation for this subspace mapping approach, rather than simply pushing the target subspace away. Finally, we present our proposed method, the Subspace Mapping Approach for Concept Erasure (SuMa) (Sec. 3.3)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 432, + 399, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 399, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 399, + 444 + ], + "type": "text", + "content": "3.1. Observations" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 450, + 555, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 569 + ], + "type": "text", + "content": "We conducted a toy experiment using Concept-Ablation (CA) [16] to investigate whether it can erase tokens that differ in form but convey the same underlying concept. We began by applying Textual Inversion (TI) on SD1.4 for 500 steps using five concepts reported in CA [16]. For each concept, we collected learned tokens from steps 300 to 500, denoted as set " + }, + { + "bbox": [ + 313, + 450, + 555, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 450, + 555, + 569 + ], + "type": "text", + "content": ". At this stage, all tokens in " + }, + { + "bbox": [ + 313, + 450, + 555, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 450, + 555, + 569 + ], + "type": "text", + "content": " could generate the target concept. We then trained CA to erase one randomly selected token from " + }, + { + "bbox": [ + 313, + 450, + 555, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 450, + 555, + 569 + ], + "type": "text", + "content": " and observed that the remaining tokens, though differing in form, were also erased." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "Next, we calculated the cosine similarity between the erased token and all tokens from " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": ", resulting in a distribution shown in Fig. 2a. We noted that the cosine similarity values ranged from 0.6 to 1.00. We proceeded by running TI on the CA-based concept-erased model above. Upon reaching step 300, TI successfully recovered the erased concept. We then computed the cosine similarity between the new TI tokens and all tokens from " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " again and showed the result in Fig. 2b. At this time, the cosine similarity ranged between 0.3 and 0.4, significantly lower than that of the previously erased token. Based on these observations, we hypothesize that we can't represent the target concept as a single" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19589" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 91, + 296, + 185 + ], + "blocks": [ + { + "bbox": [ + 59, + 91, + 296, + 185 + ], + "lines": [ + { + "bbox": [ + 59, + 91, + 296, + 185 + ], + "spans": [ + { + "bbox": [ + 59, + 91, + 296, + 185 + ], + "type": "image", + "image_path": "9c1ae6d555ebb5978227a619110cd1478e986a8e0afc46c43245971476aecd9f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 186, + 303, + 236 + ], + "lines": [ + { + "bbox": [ + 56, + 186, + 303, + 236 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 303, + 236 + ], + "type": "text", + "content": "(a) Subspace Construction: Following the STEREO approach, we iteratively remove the Target Concept from model " + }, + { + "bbox": [ + 56, + 186, + 303, + 236 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 56, + 186, + 303, + 236 + ], + "type": "text", + "content": " using method " + }, + { + "bbox": [ + 56, + 186, + 303, + 236 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 56, + 186, + 303, + 236 + ], + "type": "text", + "content": ", then run Textual Inversion (TI) and extract the corresponding tokens. At each iteration, we take the token from the final TI step and one from an early step to form the basis of the Target Subspace and Reference Subspace, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 307, + 91, + 547, + 185 + ], + "blocks": [ + { + "bbox": [ + 307, + 91, + 547, + 185 + ], + "lines": [ + { + "bbox": [ + 307, + 91, + 547, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 91, + 547, + 185 + ], + "type": "image", + "image_path": "57a7c040693777596ac41c14a2b2c630eecd0a71c97341a5d11a4ea10b8307bf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 186, + 553, + 237 + ], + "lines": [ + { + "bbox": [ + 306, + 186, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 306, + 186, + 553, + 237 + ], + "type": "text", + "content": "(b) Subspace Elimination: In this stage, our objective is to map the Target Subspace " + }, + { + "bbox": [ + 306, + 186, + 553, + 237 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 306, + 186, + 553, + 237 + ], + "type": "text", + "content": " onto the Reference Subspace " + }, + { + "bbox": [ + 306, + 186, + 553, + 237 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 306, + 186, + 553, + 237 + ], + "type": "text", + "content": ", and to maintain the Reference Subspace in its fine-tuned state " + }, + { + "bbox": [ + 306, + 186, + 553, + 237 + ], + "type": "inline_equation", + "content": "\\mathcal{V}'" + }, + { + "bbox": [ + 306, + 186, + 553, + 237 + ], + "type": "text", + "content": " unchanged. We let " + }, + { + "bbox": [ + 306, + 186, + 553, + 237 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 306, + 186, + 553, + 237 + ], + "type": "text", + "content": " represent the Text-encoder and " + }, + { + "bbox": [ + 306, + 186, + 553, + 237 + ], + "type": "inline_equation", + "content": "L_{sub}" + }, + { + "bbox": [ + 306, + 186, + 553, + 237 + ], + "type": "text", + "content": " denote the cost of mapping " + }, + { + "bbox": [ + 306, + 186, + 553, + 237 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 306, + 186, + 553, + 237 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 306, + 186, + 553, + 237 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 306, + 186, + 553, + 237 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 186, + 553, + 237 + ], + "type": "inline_equation", + "content": "\\mathcal{V}'" + }, + { + "bbox": [ + 306, + 186, + 553, + 237 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 306, + 186, + 553, + 237 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 306, + 186, + 553, + 237 + ], + "type": "text", + "content": ". The parameter " + }, + { + "bbox": [ + 306, + 186, + 553, + 237 + ], + "type": "inline_equation", + "content": "W_k'" + }, + { + "bbox": [ + 306, + 186, + 553, + 237 + ], + "type": "text", + "content": " is initialized as " + }, + { + "bbox": [ + 306, + 186, + 553, + 237 + ], + "type": "inline_equation", + "content": "W_k" + }, + { + "bbox": [ + 306, + 186, + 553, + 237 + ], + "type": "text", + "content": " and will be updated during fine-tuning." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 244, + 555, + 268 + ], + "lines": [ + { + "bbox": [ + 55, + 244, + 555, + 268 + ], + "spans": [ + { + "bbox": [ + 55, + 244, + 555, + 268 + ], + "type": "text", + "content": "Figure 3. Our proposed framework consists of two main stages. The first stage, Subspace Construction, is used to build Target Subspace and Reference Subspace. The second stage, Subspace Elimination, aims to eliminate the Target Subspace." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 278, + 295, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 278, + 295, + 338 + ], + "spans": [ + { + "bbox": [ + 54, + 278, + 295, + 338 + ], + "type": "text", + "content": "token but rather as a subspace; erasing a concept could be viewed as eliminating a subspace, and if the cosine similarity between this subspace and tokens that produce a similar concept exceeds 0.6, those tokens are erased along with the target subspace." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 338, + 296, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 296, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 296, + 469 + ], + "type": "text", + "content": "To verify this hypothesis, we follow the procedure in STEREO [33] by iteratively erasing tokens found by TI using CA. The process is illustrated in Fig. 3a. For each iteration, we collect the token from the final training step of TI. We aggregate tokens over " + }, + { + "bbox": [ + 55, + 338, + 296, + 469 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 338, + 296, + 469 + ], + "type": "text", + "content": " iterations, denoted as " + }, + { + "bbox": [ + 55, + 338, + 296, + 469 + ], + "type": "inline_equation", + "content": "x_{1},x_{2},\\ldots ,x_{n}\\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 55, + 338, + 296, + 469 + ], + "type": "text", + "content": ", ensuring that each token " + }, + { + "bbox": [ + 55, + 338, + 296, + 469 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 55, + 338, + 296, + 469 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 338, + 296, + 469 + ], + "type": "inline_equation", + "content": "i\\in [1,n]" + }, + { + "bbox": [ + 55, + 338, + 296, + 469 + ], + "type": "text", + "content": " can generate the target concept. We then construct the matrix " + }, + { + "bbox": [ + 55, + 338, + 296, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_t = [x_1,\\dots,x_t]\\in \\mathbb{R}^{d\\times t}" + }, + { + "bbox": [ + 55, + 338, + 296, + 469 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 338, + 296, + 469 + ], + "type": "inline_equation", + "content": "t\\in [1,n - 1]" + }, + { + "bbox": [ + 55, + 338, + 296, + 469 + ], + "type": "text", + "content": ". Next, we compute the angle between each token " + }, + { + "bbox": [ + 55, + 338, + 296, + 469 + ], + "type": "inline_equation", + "content": "x_{j}" + }, + { + "bbox": [ + 55, + 338, + 296, + 469 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 338, + 296, + 469 + ], + "type": "inline_equation", + "content": "j\\in [t + 1,n]" + }, + { + "bbox": [ + 55, + 338, + 296, + 469 + ], + "type": "text", + "content": " and the column space of " + }, + { + "bbox": [ + 55, + 338, + 296, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_t" + }, + { + "bbox": [ + 55, + 338, + 296, + 469 + ], + "type": "text", + "content": ", denoted by " + }, + { + "bbox": [ + 55, + 338, + 296, + 469 + ], + "type": "inline_equation", + "content": "c_{tj}" + }, + { + "bbox": [ + 55, + 338, + 296, + 469 + ], + "type": "text", + "content": ", using the formula:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 99, + 480, + 295, + 501 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 480, + 295, + 501 + ], + "spans": [ + { + "bbox": [ + 99, + 480, + 295, + 501 + ], + "type": "interline_equation", + "content": "c _ {t j} = \\cos \\left(x _ {j}, \\mathbf {S} _ {t} \\left(\\mathbf {S} _ {t} ^ {T} \\mathbf {S} _ {t}\\right) ^ {- 1} \\mathbf {S} _ {t} ^ {T} x _ {j}\\right), \\tag {1}", + "image_path": "bfa2aebaec5bfe66b8dfad6c30bb436b2aa4e980ebb11dcdbcae40dfa5431685.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 508, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 508, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 54, + 508, + 295, + 664 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 54, + 508, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_t\\left(\\mathbf{S}_t^T\\mathbf{S}_t^T\\right)^{-1}\\mathbf{S}_t^T" + }, + { + "bbox": [ + 54, + 508, + 295, + 664 + ], + "type": "text", + "content": " is the projection matrix onto the column space of " + }, + { + "bbox": [ + 54, + 508, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_t" + }, + { + "bbox": [ + 54, + 508, + 295, + 664 + ], + "type": "text", + "content": ". We compute the average value of " + }, + { + "bbox": [ + 54, + 508, + 295, + 664 + ], + "type": "inline_equation", + "content": "c_{tj}" + }, + { + "bbox": [ + 54, + 508, + 295, + 664 + ], + "type": "text", + "content": " for different values of " + }, + { + "bbox": [ + 54, + 508, + 295, + 664 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 54, + 508, + 295, + 664 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 54, + 508, + 295, + 664 + ], + "type": "inline_equation", + "content": "c_{t}" + }, + { + "bbox": [ + 54, + 508, + 295, + 664 + ], + "type": "text", + "content": ", and report them in Fig. 2c. Notably, when " + }, + { + "bbox": [ + 54, + 508, + 295, + 664 + ], + "type": "inline_equation", + "content": "t \\geq 3" + }, + { + "bbox": [ + 54, + 508, + 295, + 664 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 508, + 295, + 664 + ], + "type": "inline_equation", + "content": "c_{t}" + }, + { + "bbox": [ + 54, + 508, + 295, + 664 + ], + "type": "text", + "content": " exceeds 0.6—the threshold we hypothesize is sufficient to cover the target concept. Then, we employ the second stage of STEREO to erase two sets of tokens: " + }, + { + "bbox": [ + 54, + 508, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_1 = \\{x_1, x_2, x_3\\}" + }, + { + "bbox": [ + 54, + 508, + 295, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 508, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_2 = \\{x_1, x_2\\}" + }, + { + "bbox": [ + 54, + 508, + 295, + 664 + ], + "type": "text", + "content": ". We calculated the average Attack Success Rate (ASR) across all concepts, yielding " + }, + { + "bbox": [ + 54, + 508, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_1" + }, + { + "bbox": [ + 54, + 508, + 295, + 664 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 54, + 508, + 295, + 664 + ], + "type": "inline_equation", + "content": "3.21\\%" + }, + { + "bbox": [ + 54, + 508, + 295, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 508, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_2" + }, + { + "bbox": [ + 54, + 508, + 295, + 664 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 54, + 508, + 295, + 664 + ], + "type": "inline_equation", + "content": "85.32\\%" + }, + { + "bbox": [ + 54, + 508, + 295, + 664 + ], + "type": "text", + "content": ". This indicates STEREO effectively defends against CCE when trained to erase " + }, + { + "bbox": [ + 54, + 508, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_1" + }, + { + "bbox": [ + 54, + 508, + 295, + 664 + ], + "type": "text", + "content": ", but not when only " + }, + { + "bbox": [ + 54, + 508, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_2" + }, + { + "bbox": [ + 54, + 508, + 295, + 664 + ], + "type": "text", + "content": " is erased. An additional experiment on token similarity's impact on erasure is detailed in Appendix B." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "We found that the adversarial tokens identified by TI encompass those detected by UnlearnDiff. Specifically, we evaluated a model trained to erase " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "S_{1}" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": " using UnlearnDiff, which yielded an average Attack Success Rate of " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "1.52\\%" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": ", in" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 278, + 553, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 553, + 397 + ], + "type": "text", + "content": "dicating robustness against UnlearnDiff. To explain this result, we provide an additional justification. Since UnlearnDiff identifies combinations of existing tokens, we leveraged the attention mechanism of the text encoder by using the " + }, + { + "bbox": [ + 313, + 278, + 553, + 397 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{EOT} \\rangle" + }, + { + "bbox": [ + 313, + 278, + 553, + 397 + ], + "type": "text", + "content": " token to represent the target concept, and applied the same technique to all tokens from " + }, + { + "bbox": [ + 313, + 278, + 553, + 397 + ], + "type": "inline_equation", + "content": "S_{1}" + }, + { + "bbox": [ + 313, + 278, + 553, + 397 + ], + "type": "text", + "content": ". After this processing, we computed a cosine similarity of 0.78 between the subspace formed by the " + }, + { + "bbox": [ + 313, + 278, + 553, + 397 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{EOT} \\rangle" + }, + { + "bbox": [ + 313, + 278, + 553, + 397 + ], + "type": "text", + "content": " token from " + }, + { + "bbox": [ + 313, + 278, + 553, + 397 + ], + "type": "inline_equation", + "content": "S_{1}" + }, + { + "bbox": [ + 313, + 278, + 553, + 397 + ], + "type": "text", + "content": " and the one identified by UnlearnDiff, exceeding 0.6 and supporting our observations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 399, + 554, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 554, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 554, + 471 + ], + "type": "text", + "content": "In summary, our analysis explains why STEREO can robustly remove the target concept. By applying TI on the concept-erased model iteratively in at least 3 iterations, it can collect a sufficient set of tokens spanning the target concept space. When erasing these tokens together, it successfully eradicates the model knowledge of the concept." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 479, + 553, + 505 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 553, + 505 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 553, + 505 + ], + "type": "text", + "content": "3.2. Reference Subspace Selection for Narrow Concepts" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": "CA [16] and ESD [7] were the first methods proposed for concept erasure, but they present contradictory approaches. CA maps noise conditioned on the target concept to a more general representation, while ESD pushes the target noise as far as possible from the original. Studies [18, 33] show that CA outperforms ESD in terms of image quality. It is because mapping approach offers a clear direction to erase target concepts, whereas pushing approach can inadvertently affect unrelated concepts. The second stage of STEREO [33], also a pushing approach, significantly degrades image quality. With this motivation, our paper proposes a mapping approach that erases tokens by mapping them to nearby non-target concept subspaces, preserving model utility. The key challenge is how to automatically select these reference subspaces. We also provide experimental results for Subspace Pushing, where the target subspace is simply pushed as far as possible from the original one, in Appendix C." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19590" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 291, + 143 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 291, + 143 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 291, + 143 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 291, + 143 + ], + "type": "table", + "html": "
ConceptTI Early TokenTextual Token
R2D2 Robot0.641.23
Garbage Truck1.051.28
Grumpy Cat0.981.22
English Springer Dog0.720.96
", + "image_path": "44304fbfbecb48842e4507ad5ca7d6b8bc7fa96598f2988d7804eb1552cc5a04.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 146, + 295, + 169 + ], + "lines": [ + { + "bbox": [ + 55, + 146, + 295, + 169 + ], + "spans": [ + { + "bbox": [ + 55, + 146, + 295, + 169 + ], + "type": "text", + "content": "Table 1. The Euclid distance of the target token to the tokens from early TI training steps and the textual one." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 180, + 296, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 296, + 360 + ], + "type": "text", + "content": "We observe that constructing the reference subspace using tokens from the early training steps of TI is effective for narrow concepts, as early tokens typically converge to more general forms (e.g., grumpy cat to cat, or R2D2 robot to robot; see Section 4.3 for details). Moreover, these tokens are closer to the target token learned from TI because they align along the same gradient direction. To validate this, we compare the Euclidean distance between the target token and early TI training tokens, as well as the target token and the textual token, as shown in Table 1. The results indicate that while early tokens may represent concepts similar to the textual token, their distance from the target token is much smaller. This suggests that mapping the target token to an earlier token minimizes weight modifications, thereby better preserving the model's utility." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 367, + 162, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 162, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 162, + 380 + ], + "type": "text", + "content": "3.3. Proposed Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "text", + "content": "Our framework, " + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "inline_equation", + "content": "SuMa" + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "text", + "content": ", consists of two stages, as illustrated in Fig. 3. The first stage, termed Subspace Construction, aims to identify Textual Inversion tokens to establish both sets of Target Subspace " + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{U}^i\\}" + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "text", + "content": " and Reference Subspace " + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{V}^i\\}" + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{U}^i" + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "text", + "content": " represents " + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "text", + "content": "-th subspace (corresponding to " + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "text", + "content": "-th Unet layer) of the concept to be erased, while " + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^i" + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "text", + "content": "-th the subspace into which we intend to map our " + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "text", + "content": "-th target subspace. In the second stage, Subspace Elimination, we refine the to- " + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "text", + "content": " and to- " + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "text", + "content": " weights of the Cross-Attention Layer in U-Net, denoted as " + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "inline_equation", + "content": "W_{k}" + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "inline_equation", + "content": "W_{v}" + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "text", + "content": ", allowing us to project the basis of " + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{U}^i" + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "text", + "content": " onto " + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^i" + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "text", + "content": ". Due to the identical functionality of " + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "inline_equation", + "content": "W_{k}" + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "inline_equation", + "content": "W_{v}" + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "text", + "content": " in our work, we denote both as " + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "inline_equation", + "content": "W_{k}" + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "text", + "content": " for simplicity and readability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "text", + "content": "Subspace Construction. Starting with the original model, denoted as " + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "text", + "content": ", and a concept to be removed, " + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "text", + "content": ", we apply a Concept Ablation " + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "inline_equation", + "content": "(\\mathcal{M})" + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "text", + "content": " to eliminate " + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "text", + "content": ", resulting in a modified model " + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\theta'" + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "text", + "content": ". This process is repeated over " + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "text", + "content": " steps. At " + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "text", + "content": "-th iteration, we apply Textual Inversion (TI) for the concept to be erased from " + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\theta'" + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "text", + "content": ", yielding two tokens: one from the final step of TI training, denoted as " + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\langle u\\rangle_{j}" + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "text", + "content": ", and another from an earlier step, denoted as " + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\langle v\\rangle_{j}" + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "text", + "content": ". Subsequently, we construct prompts in the formats \"A photo of " + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\langle u\\rangle_{j}" + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "text", + "content": "\" and \"A photo of " + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\langle v\\rangle_{j}" + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "text", + "content": "\", which are then passed through the Text Encoder " + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "text", + "content": " to extract embeddings for the tokens " + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\langle u_j\\rangle" + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\langle v_j\\rangle" + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "text", + "content": ". We then apply the transformation " + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "inline_equation", + "content": "W_k^i" + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "text", + "content": " represent " + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "text", + "content": "-th layer in Unet, to these extracted embeddings, treating the outputs as " + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\langle u'\\rangle_{j}^{i}" + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\langle v'\\rangle_{j}^{i}" + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "text", + "content": " as basis vectors for the subspaces " + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{U}^i" + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^i" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 544, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 544, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 544, + 84 + ], + "type": "text", + "content": "respectively. Formally, this process is defined as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 381, + 88, + 553, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 88, + 553, + 119 + ], + "spans": [ + { + "bbox": [ + 381, + 88, + 553, + 119 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\langle u ^ {\\prime} \\rangle_ {j} ^ {i} = W _ {k} ^ {i} \\mathcal {T} (p + \\langle u \\rangle_ {j}), \\tag {2} \\\\ \\langle v ^ {\\prime} \\rangle_ {j} ^ {i} = W _ {k} ^ {i} \\mathcal {T} (p + \\langle v \\rangle_ {j}), \\\\ \\end{array}", + "image_path": "6dd8dadd416243201d8276060a90a8a0bdd8b7a049558ade448c722d485ba845.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 122, + 554, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 122, + 554, + 159 + ], + "spans": [ + { + "bbox": [ + 313, + 122, + 554, + 159 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 122, + 554, + 159 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 122, + 554, + 159 + ], + "type": "text", + "content": " represents the prompt template. Finally, we fine-tune " + }, + { + "bbox": [ + 313, + 122, + 554, + 159 + ], + "type": "inline_equation", + "content": "\\theta'" + }, + { + "bbox": [ + 313, + 122, + 554, + 159 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 313, + 122, + 554, + 159 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 122, + 554, + 159 + ], + "type": "text", + "content": " to remove " + }, + { + "bbox": [ + 313, + 122, + 554, + 159 + ], + "type": "inline_equation", + "content": "\\langle u_j\\rangle" + }, + { + "bbox": [ + 313, + 122, + 554, + 159 + ], + "type": "text", + "content": ", resulting in an updated model " + }, + { + "bbox": [ + 313, + 122, + 554, + 159 + ], + "type": "inline_equation", + "content": "\\theta'" + }, + { + "bbox": [ + 313, + 122, + 554, + 159 + ], + "type": "text", + "content": " at the end of each iteration (Fig. 3a)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 164, + 555, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 164, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 313, + 164, + 555, + 271 + ], + "type": "text", + "content": "Subspace Elimination. Fig. 3b illustrates this process. After constructing sets of the subspaces " + }, + { + "bbox": [ + 313, + 164, + 555, + 271 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{U}^i\\}_{i=1}^M" + }, + { + "bbox": [ + 313, + 164, + 555, + 271 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 164, + 555, + 271 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{V}^i\\}_{i=1}^M" + }, + { + "bbox": [ + 313, + 164, + 555, + 271 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 164, + 555, + 271 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 164, + 555, + 271 + ], + "type": "text", + "content": " is number of layer in Unet, our goal is to eliminate all " + }, + { + "bbox": [ + 313, + 164, + 555, + 271 + ], + "type": "inline_equation", + "content": "\\mathcal{U}^i" + }, + { + "bbox": [ + 313, + 164, + 555, + 271 + ], + "type": "text", + "content": " by fine-tuning " + }, + { + "bbox": [ + 313, + 164, + 555, + 271 + ], + "type": "inline_equation", + "content": "W_k" + }, + { + "bbox": [ + 313, + 164, + 555, + 271 + ], + "type": "text", + "content": " so that the basis vectors of " + }, + { + "bbox": [ + 313, + 164, + 555, + 271 + ], + "type": "inline_equation", + "content": "\\mathcal{U}^i" + }, + { + "bbox": [ + 313, + 164, + 555, + 271 + ], + "type": "text", + "content": " are projected onto " + }, + { + "bbox": [ + 313, + 164, + 555, + 271 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^i" + }, + { + "bbox": [ + 313, + 164, + 555, + 271 + ], + "type": "text", + "content": ". To construct the projection matrix onto " + }, + { + "bbox": [ + 313, + 164, + 555, + 271 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^i" + }, + { + "bbox": [ + 313, + 164, + 555, + 271 + ], + "type": "text", + "content": ", we first define the matrix " + }, + { + "bbox": [ + 313, + 164, + 555, + 271 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^i = [\\langle v' \\rangle_1^i, \\langle v' \\rangle_2^i, \\dots, \\langle v' \\rangle_l^i] \\in \\mathbb{R}^{d \\times l}" + }, + { + "bbox": [ + 313, + 164, + 555, + 271 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 164, + 555, + 271 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 164, + 555, + 271 + ], + "type": "text", + "content": " is the dimension of token embedding vector, and the projection matrix " + }, + { + "bbox": [ + 313, + 164, + 555, + 271 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathbf{V}^i}" + }, + { + "bbox": [ + 313, + 164, + 555, + 271 + ], + "type": "text", + "content": " can then be formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 362, + 280, + 555, + 303 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 280, + 555, + 303 + ], + "spans": [ + { + "bbox": [ + 362, + 280, + 555, + 303 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {\\mathbf {V} ^ {i}} = \\mathbf {V} ^ {i} \\left(\\left(\\mathbf {V} ^ {i}\\right) ^ {T} \\mathbf {V} ^ {i}\\right) ^ {- 1} \\left(\\mathbf {V} ^ {i}\\right) ^ {T} \\tag {3}", + "image_path": "ec8c3ae87973a3d53c5424bce6d51baf5738832afb2fc5ff669948da4547a5ba.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 305, + 554, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 305, + 554, + 329 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 554, + 329 + ], + "type": "text", + "content": "We define " + }, + { + "bbox": [ + 313, + 305, + 554, + 329 + ], + "type": "inline_equation", + "content": "L_{proj}" + }, + { + "bbox": [ + 313, + 305, + 554, + 329 + ], + "type": "text", + "content": " as the objective to project all basis of " + }, + { + "bbox": [ + 313, + 305, + 554, + 329 + ], + "type": "inline_equation", + "content": "\\mathcal{U}^i" + }, + { + "bbox": [ + 313, + 305, + 554, + 329 + ], + "type": "text", + "content": " onto " + }, + { + "bbox": [ + 313, + 305, + 554, + 329 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^i" + }, + { + "bbox": [ + 313, + 305, + 554, + 329 + ], + "type": "text", + "content": " by following formula:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 333, + 565, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 333, + 565, + 376 + ], + "spans": [ + { + "bbox": [ + 314, + 333, + 565, + 376 + ], + "type": "interline_equation", + "content": "L _ {p r o j} = \\sum_ {i = 1} ^ {M} \\sum_ {j = 1} ^ {l} \\left\\| W _ {k} ^ {i} \\mathcal {T} (p + \\langle u \\rangle_ {j}) - \\mathbf {P} _ {\\mathbf {V} ^ {i}} W _ {k} ^ {i} \\mathcal {T} (p + \\langle u \\rangle_ {j}) \\right\\| _ {2} ^ {2} \\tag {4}", + "image_path": "4ddafc21c5d6637bcdc041cfd848e30d3ead1bddadbabcb9976ebb034eafdc1b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 378, + 555, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 555, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 555, + 402 + ], + "type": "text", + "content": "As per [3], we add a regularization term to keep the Reference Subspace " + }, + { + "bbox": [ + 313, + 378, + 555, + 402 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^i" + }, + { + "bbox": [ + 313, + 378, + 555, + 402 + ], + "type": "text", + "content": " unchanged during fine-tuning:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 415, + 555, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 415, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 314, + 415, + 555, + 459 + ], + "type": "interline_equation", + "content": "L _ {r e g} = \\sum_ {i = 1} ^ {M} \\sum_ {j = 1} ^ {l} \\left\\| W _ {k} ^ {i} \\mathcal {T} (p + \\langle v \\rangle_ {j}) - \\mathbf {P} _ {\\mathbf {V} ^ {i}} W _ {k} ^ {i} \\mathcal {T} (p + \\langle v \\rangle_ {j}) \\right\\| \\tag {5}", + "image_path": "d11019817a527db39e2e7e58f7586e9aaa0a1d7c64b99fd6425810df18bae612.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 460, + 554, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 483 + ], + "type": "text", + "content": "The objective function for subspace mapping, " + }, + { + "bbox": [ + 313, + 460, + 554, + 483 + ], + "type": "inline_equation", + "content": "L_{sub}" + }, + { + "bbox": [ + 313, + 460, + 554, + 483 + ], + "type": "text", + "content": ", is then defined as the weighted sum of these two terms:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 378, + 489, + 553, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 489, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 378, + 489, + 553, + 502 + ], + "type": "interline_equation", + "content": "L _ {s u b} = L _ {p r o j} + \\lambda_ {r e g} L _ {r e g}, \\tag {6}", + "image_path": "8c6460032ae28ab5ddaea4e4f3883ac26b91e64da91227aa05faf14555ab0271.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 506, + 554, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 554, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 554, + 613 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 506, + 554, + 613 + ], + "type": "inline_equation", + "content": "\\lambda_{reg}" + }, + { + "bbox": [ + 313, + 506, + 554, + 613 + ], + "type": "text", + "content": " is a hyperparameter (default 1). If we use " + }, + { + "bbox": [ + 313, + 506, + 554, + 613 + ], + "type": "inline_equation", + "content": "L_{sub}" + }, + { + "bbox": [ + 313, + 506, + 554, + 613 + ], + "type": "text", + "content": " to fine-tune " + }, + { + "bbox": [ + 313, + 506, + 554, + 613 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 506, + 554, + 613 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 313, + 506, + 554, + 613 + ], + "type": "inline_equation", + "content": "\\theta'" + }, + { + "bbox": [ + 313, + 506, + 554, + 613 + ], + "type": "text", + "content": " and use CCE to find a new target token, this token still generates the target concept when conditioned on " + }, + { + "bbox": [ + 313, + 506, + 554, + 613 + ], + "type": "inline_equation", + "content": "\\theta'" + }, + { + "bbox": [ + 313, + 506, + 554, + 613 + ], + "type": "text", + "content": ", but not on " + }, + { + "bbox": [ + 313, + 506, + 554, + 613 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 506, + 554, + 613 + ], + "type": "text", + "content": " (see Appendix D). We assume modifying only the cross-attention layer is effective for erasing a token, as in [18], but affects the SDE trajectory when erasing multiple tokens. To mitigate this, we use the CA loss function and discuss our choice in Section 4.3. In short, their objective is:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 617, + 559, + 642 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 617, + 559, + 642 + ], + "spans": [ + { + "bbox": [ + 314, + 617, + 559, + 642 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {C A}} = \\mathbb {E} _ {z _ {t}, c, t} \\left[ \\| \\Phi^ {\\prime} \\left(z _ {t}, \\mathcal {T} (p + c ^ {\\prime}), t\\right) - \\Phi \\left(z _ {t}, \\mathcal {T} (p + c), t\\right) \\| _ {2} ^ {2} \\right] \\tag {7}", + "image_path": "6b1e3a60e5898c53ba813d15aca3ea29da4fd274a9e9e61737413e68c292f1e3.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 643, + 554, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 643, + 554, + 679 + ], + "spans": [ + { + "bbox": [ + 313, + 643, + 554, + 679 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 643, + 554, + 679 + ], + "type": "inline_equation", + "content": "\\Phi'" + }, + { + "bbox": [ + 313, + 643, + 554, + 679 + ], + "type": "text", + "content": " is the model to learn, " + }, + { + "bbox": [ + 313, + 643, + 554, + 679 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 313, + 643, + 554, + 679 + ], + "type": "text", + "content": " is pretrained, " + }, + { + "bbox": [ + 313, + 643, + 554, + 679 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 643, + 554, + 679 + ], + "type": "text", + "content": " is TextEncoder, " + }, + { + "bbox": [ + 313, + 643, + 554, + 679 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 643, + 554, + 679 + ], + "type": "text", + "content": " is prompt template, " + }, + { + "bbox": [ + 313, + 643, + 554, + 679 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 643, + 554, + 679 + ], + "type": "text", + "content": " is target concept, and " + }, + { + "bbox": [ + 313, + 643, + 554, + 679 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 313, + 643, + 554, + 679 + ], + "type": "text", + "content": " is its general term (e.g., grumpy cat vs. cat). Our objective is:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 331, + 695, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 695, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 331, + 695, + 553, + 715 + ], + "type": "interline_equation", + "content": "W _ {k} ^ {*} = \\underset {W _ {k}} {\\operatorname {a r g m i n}} \\left(L _ {\\text {f i n a l}}\\right) = \\underset {W _ {k}} {\\operatorname {a r g m i n}} \\left(L _ {\\mathrm {C A}} + L _ {\\text {s u b}}\\right). \\tag {8}", + "image_path": "b8d81bcf36c35a6e0cdcf7778684cd016f94a28cd3987990fbbf22f3ad4f42fc.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19591" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 303, + 197 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 303, + 197 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 303, + 197 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 303, + 197 + ], + "type": "image", + "image_path": "7b4628f46847a4d261e799125cc900f15c76736d7157e4d9f447c7b24cd804ab.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 198, + 302, + 248 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 302, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 302, + 248 + ], + "type": "text", + "content": "(a) Qualitative Comparison across 6 methods for the CCE attack. Both E-CEM methods, CA and MACE, are vulnerable to the CCE attack. STEREO's erasure is too aggressive, leaving only Van Gogh-style elements, with this concept being easier to erase. Our method defends against CCE by mapping the target concept to its reference concept." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 306, + 70, + 552, + 198 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 552, + 198 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 552, + 198 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 552, + 198 + ], + "type": "image", + "image_path": "2aa6741d1bcd3c2a72907200ba9aadbafa3627df1845f08481d8a36ba91da338.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 199, + 553, + 247 + ], + "lines": [ + { + "bbox": [ + 306, + 199, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 306, + 199, + 553, + 247 + ], + "type": "text", + "content": "(b) Qualitative Comparison across 6 methods for image generation. CA struggles with \"English Springer\" and \"Elon Musk.\" MACE performs well, preserving image quality. RACE shows slight degradation, and STEREO significantly reduces quality. Our method strikes a balance between E-CEMs and C-CEMs in terms of quality." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 251, + 538, + 262 + ], + "lines": [ + { + "bbox": [ + 70, + 251, + 538, + 262 + ], + "spans": [ + { + "bbox": [ + 70, + 251, + 538, + 262 + ], + "type": "text", + "content": "Figure 4. Qualitative Results. We present qualitative results comparing five methods against the CCE attack and textual erasure." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 277, + 137, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 137, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 137, + 289 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 296, + 178, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 296, + 178, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 178, + 309 + ], + "type": "text", + "content": "4.1. Experimental Setting" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 319, + 296, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 547 + ], + "type": "text", + "content": "Dataset. We evaluate SuMa across four tasks: subclass erasure, celebrity erasure, artistic style erasure, and instance erasure. For subclass erasure, we focus on a subset of ImageNette [12] known for its distinctive features. Three subclasses are selected for the test, including English Springer Dog, Garbage Truck, and Golf Ball. For celebrity erasure, we select 3 celebrities in the Celebrity-1000 dataset [34], including David Beckham, Elon Musk, and Adam Lambert. For artistic style erasure, following the approach in [7, 15, 33], we focus on the style of Van Gogh, which is both famous and straightforward to verify in outputs. For instance erasure, we conduct experiments on three copyrighted concepts: Grumpy Cat, Mickey Mouse, and the R2D2 robot. Note that our method is specifically designed to address the problem with narrow concepts, so it does not work as well for NSFW concepts. However, as mentioned earlier, our work is complementary to DUO, so it can be easily integrated with this method to completely erase all kinds of concepts, demonstrated in the Appendix E." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 552, + 296, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 296, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 296, + 636 + ], + "type": "text", + "content": "Baseline. We compare SuMa with four baselines: two E-CEMs (CA [16] and MACE [18]) and four C-CEMs (RACE [15], AdvUnlearn (AU) [38], STEREO [33], DUO [21]). For CA, MACE, RACE, and AU, we use their public code and recommended settings to erase concepts. For STEREO and DUO, we re-implement their methods using paper recommendations and verify consistency with reported results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "Circumventing Methods. We evaluate the robustness of our method under CCE [22] and UnlearnDiff (UD) [39]. For CCE, we use a different set of images from the Subspace Construction stage to ensure the model didn't overfit to this image set. For both attack methods, we use publicly available code to generate adversarial prompts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 277, + 555, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 566 + ], + "type": "text", + "content": "Evaluation Metrics. First, to measure effectiveness, for each erased model, we generate 5,000 images using prompts from the MS-COCO validation set. We then evaluate the model quality when generating such non-target prompts using CLIP scores (higher is better) and FID scores (lower is better). To measure completeness, in the CCE attack, for each erased model, we execute the circumventing method to acquire the adversarial token " + }, + { + "bbox": [ + 313, + 277, + 555, + 566 + ], + "type": "inline_equation", + "content": "c_{adv}" + }, + { + "bbox": [ + 313, + 277, + 555, + 566 + ], + "type": "text", + "content": ". We then generate 200 images using a template prompt: \"A photo of a " + }, + { + "bbox": [ + 313, + 277, + 555, + 566 + ], + "type": "inline_equation", + "content": "\\{c_{adv}\\}" + }, + { + "bbox": [ + 313, + 277, + 555, + 566 + ], + "type": "text", + "content": ".\" For style erasure, we use the template \"A painting in the style of " + }, + { + "bbox": [ + 313, + 277, + 555, + 566 + ], + "type": "inline_equation", + "content": "\\{c_{adv}\\}" + }, + { + "bbox": [ + 313, + 277, + 555, + 566 + ], + "type": "text", + "content": "\". In the UD attack, we execute its code on our erased model to get the adversarial prompt and pass the prompt to the erased model. We report the attack success rate (ASR), where a lower score indicates better performance, following the method in [15, 33]. For subclass erasure tasks, we utilize ResNet-50 [10], pretrained on ImageNet [27], to measure ASR. For celebrity detection, we use the GIPHY Celebrity Detector (GCD) [9] as the classifier. To evaluate artistic style, we follow the approach in [15] by fine-tuning a ViT-based model [1] on the WikiArt dataset [29]. Finally, for instance detection, due to the lack of pretrained models or datasets, we use LLAVA-7B [17] with a template prompt: \"Does this image contain {object}? Answer Yes or No.\" for evaluation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": "Implementation Details. We use SD1.4 and SD1.5 as the pretrained model. In all experiments, we fine-tune all cross-attention layers. In the Subspace Construction stage, we use the same setting as CA [16] for concept erasion. Training TI is set for 500 steps, with the early step to get token " + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\langle v\\rangle" + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": " set to 50, the initial token is toy, with Adam optimization with learning rate, is 2e-6. For the Subspace Elimination stage, we fine-tune most objects for 500 steps, except for subclass objects, which require 750 steps to ensure model safety. We use the Adam optimizer with a learning rate of 2e-5. All experiments are done on 2 NVIDIA A100 GPUs." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19592" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 70, + 309, + 289 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 309, + 289 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 309, + 289 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 309, + 289 + ], + "type": "table", + "html": "
MethodCCE↓UD↓FID↓CLIP↑
Subclass
SD [24]0.84 / 0.830.74 / 0.7617.04 / 16.950.33 / 0.32
CA [16]0.78 / 0.750.67 / 0.6219.28 / 19.260.32 / 0.32
MACE [18]0.86 / 0.840.73 / 0.6817.15 / 16.980.29 / 0.31
RACE [15]0.74 / 0.730.17 / 0.1826.34 / 26.220.27 / 0.28
AdvUnlearn [38]0.76 / 0.740.18 / 0.1618.47 / 18.390.30 / 0.31
DUO [21]0.65 / 0.630.62 / 0.6517.09 / 16.990.29 / 0.30
STEREO [33]0.05 / 0.030.05 / 0.0427.33 / 27.200.30 / 0.28
Ours0.12 / 0.100.08 / 0.0618.33 / 18.150.31 / 0.30
Identity
SD [24]0.91 / 0.900.86 / 0.8417.04 / 16.950.33 / 0.32
CA [16]0.87 / 0.890.79 / 0.7618.21 / 18.270.31 / 0.30
MACE [18]0.89 / 0.870.65 / 0.6316.74 / 16.690.29 / 0.28
RACE [15]0.78 / 0.790.47 / 0.4624.32 / 24.170.27 / 0.29
AdvUnlearn [38]0.77 / 0.690.52 / 0.5317.53 / 17.690.31 / 0.31
DUO [21]0.73 / 0.710.67 / 0.7317.32 / 17.110.29 / 0.30
STEREO [33]0.02 / 0.010.03 / 0.0126.33 / 26.150.29 / 0.28
Ours0.09 / 0.070.18 / 0.1417.94 / 17.890.31 / 0.29
", + "image_path": "538436298f35c4ef5ee1862de3032cc5402c846796676cdf6dfb09eee177012a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 312, + 70, + 547, + 287 + ], + "blocks": [ + { + "bbox": [ + 312, + 70, + 547, + 287 + ], + "lines": [ + { + "bbox": [ + 312, + 70, + 547, + 287 + ], + "spans": [ + { + "bbox": [ + 312, + 70, + 547, + 287 + ], + "type": "table", + "html": "
MethodCCE↓UD↓FID↓CLIP↑
Artistic
SD [24]0.86 / 0.830.82 / 0.8517.04 / 16.950.33 / 0.32
CA [16]0.79 / 0.780.74 / 0.7618.02 / 17.850.34 / 0.31
MACE [18]0.84 / 0.820.79 / 0.8316.57 / 16.580.32 / 0.33
RACE [15]0.82 / 0.790.06 / 0.0421.32 / 21.100.30 / 0.31
AdvUnlearn [38]0.83 / 0.760.03 / 0.0517.73 / 17.600.32 / 0.33
DUO [21]0.74 / 0.720.67 / 0.7317.74 / 17.890.30 / 0.31
STEREO [33]0.00 / 0.020.00 / 0.0128.04 / 27.900.30 / 0.31
Ours0.05 / 0.040.07 / 0.0517.18 / 17.210.33 / 0.31
Instance
SD [24]0.97 / 0.940.94 / 0.9317.04 / 16.950.33 / 0.32
CA [16]0.97 / 0.930.92 / 0.8918.43 / 18.300.30 / 0.31
MACE [18]0.97 / 0.920.93 / 0.8916.93 / 16.800.30 / 0.31
RACE [15]0.95 / 0.920.06 / 0.0724.32 / 24.150.28 / 0.29
AdvUnlearn [38]0.95 / 0.930.03 / 0.0418.37 / 18.200.32 / 0.33
DUO [21]0.92 / 0.910.89 / 0.8718.31 / 18.920.30 / 0.31
STEREO [33]0.03 / 0.020.05 / 0.0327.04 / 26.900.31 / 0.32
Ours0.17 / 0.150.19 / 0.1622.34 / 22.200.29 / 0.30
", + "image_path": "60148edea511fdaae62eaae61df1bc24bf2884d8c38b03e4ab182cb65a4ef554.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 70, + 328, + 538, + 418 + ], + "blocks": [ + { + "bbox": [ + 55, + 291, + 555, + 324 + ], + "lines": [ + { + "bbox": [ + 55, + 291, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 555, + 324 + ], + "type": "text", + "content": "Table 2. Evaluation of Erasing across 4 Categories. We average the results of all target concepts in each category across three metrics and compare our framework, SuMa, with state-of-the-art methods. For ASRs, successful attacks (ASR ≥ 20%) are in red, and defeated attacks (ASR < 20%) in green. The data in each cell show the result of erasing a concept from pretrained SD1.4 / SD1.5." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 328, + 538, + 418 + ], + "lines": [ + { + "bbox": [ + 70, + 328, + 538, + 418 + ], + "spans": [ + { + "bbox": [ + 70, + 328, + 538, + 418 + ], + "type": "table", + "html": "
Reference SubspaceSubclassIdentityArtisticInstance
ASR ↓FID ↓CLIP ↑ASR ↓FID ↓CLIP ↑ASR ↓FID ↓CLIP ↑ASR ↓FID ↓CLIP ↑
Step-300.0919.110.300.0818.340.290.0518.820.310.1723.960.27
Step-500.1218.330.310.0917.940.310.0517.180.330.1722.340.29
Step-700.5617.320.330.6717.920.320.3717.050.330.8317.740.32
General Term0.1818.270.320.1718.020.320.1118.210.330.1523.340.27
", + "image_path": "f7d4025e6d31d7eb2959699f67836833cbf15beb927ea1d48be36fba8eeae83a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 183, + 421, + 425, + 432 + ], + "lines": [ + { + "bbox": [ + 183, + 421, + 425, + 432 + ], + "spans": [ + { + "bbox": [ + 183, + 421, + 425, + 432 + ], + "type": "text", + "content": "Table 3. Ablation Study for the selection of Reference subspace." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 78, + 439, + 271, + 616 + ], + "blocks": [ + { + "bbox": [ + 78, + 439, + 271, + 616 + ], + "lines": [ + { + "bbox": [ + 78, + 439, + 271, + 616 + ], + "spans": [ + { + "bbox": [ + 78, + 439, + 271, + 616 + ], + "type": "image", + "image_path": "dd3e639ca00e25b5659003905736140f3dce21d76f1ff1da0abb65506adfa83a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 619, + 295, + 642 + ], + "lines": [ + { + "bbox": [ + 55, + 619, + 295, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 619, + 295, + 642 + ], + "type": "text", + "content": "Figure 5. Visualization of Textual Inversion tokens at different numbers of TI training steps." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 651, + 113, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 113, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 113, + 662 + ], + "type": "text", + "content": "4.2. Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Quantitative Results. We average each metric from Section 4.1 for each category and report the results in Table 2. For all baselines, we ensure the erased model works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 442, + 555, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 555, + 706 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 555, + 706 + ], + "type": "text", + "content": "well for erasing the textual prompt, and we report results for this trivial task in the Appendix F. For completeness evaluation, our method achieves an ASR score below 0.2 for all categories under both CCE and UD attacks. Compared to STEREO, our method has a slightly higher ASR score (0.02 to 0.14 higher), but within an acceptable range. RACE and AdvUnlearn show safety against the UD attack in the Subclass, Artistic, and Instance categories but fail under Identity. This is due to the subtle facial features of human identity, which make this category more fine-grained and challenging. However, both fail under the CCE attack, showing that protecting the model against CCE is harder than UD. The main reason is that RACE and AdvUnlearn erase adversarial tokens based on UnlearnDiff, but these tokens don't fully represent the target concept, protecting against UD but not CCE. DUO works well for NSFW concepts (results in Appendix E) but fails for narrow concepts. For effectiveness, although STEREO performs well under both CCE and UD, image quality is significantly affected, resulting in an FID score roughly double that of other E-CEMs, making it less effective. Our method, while not achieving a perfect ASR, achieves an FID score comparable" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19593" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 74, + 287, + 183 + ], + "blocks": [ + { + "bbox": [ + 61, + 74, + 287, + 183 + ], + "lines": [ + { + "bbox": [ + 61, + 74, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 61, + 74, + 287, + 183 + ], + "type": "image", + "image_path": "1b0defb8e6523aef785ed64195ffacfaaa9f6ad964b3665f3d56b629920eb3b1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 96, + 185, + 254, + 196 + ], + "lines": [ + { + "bbox": [ + 96, + 185, + 254, + 196 + ], + "spans": [ + { + "bbox": [ + 96, + 185, + 254, + 196 + ], + "type": "text", + "content": "Figure 6. Multiple Concept Erasure Result." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 200, + 293, + 280 + ], + "blocks": [ + { + "bbox": [ + 58, + 200, + 293, + 280 + ], + "lines": [ + { + "bbox": [ + 58, + 200, + 293, + 280 + ], + "spans": [ + { + "bbox": [ + 58, + 200, + 293, + 280 + ], + "type": "table", + "html": "
RegulizationSubclassIdentityArtisticInstance
×ASR: 0.09ASR: 0.06ASR: 0.05ASR: 0.15
FID: 19.21FID: 20.31FID: 19.32FID: 24.57
CLIP: 0.29CLIP: 0.29CLIP: 0.30CLIP: 0.26
✓ASR: 0.12ASR: 0.09ASR: 0.05ASR: 0.17
FID: 18.33FID: 17.94FID: 17.18FID: 22.34
CLIP: 0.31CLIP: 0.31CLIP: 0.33CLIP: 0.29
", + "image_path": "6492bd105a2b98f4d13a6a37d889f3e4881e61be8f03f46250388bd3f68d1f48.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 57, + 300, + 293, + 369 + ], + "blocks": [ + { + "bbox": [ + 57, + 282, + 294, + 294 + ], + "lines": [ + { + "bbox": [ + 57, + 282, + 294, + 294 + ], + "spans": [ + { + "bbox": [ + 57, + 282, + 294, + 294 + ], + "type": "text", + "content": "Table 4. Ablation Study for the need of Subspace Regulization." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 300, + 293, + 369 + ], + "lines": [ + { + "bbox": [ + 57, + 300, + 293, + 369 + ], + "spans": [ + { + "bbox": [ + 57, + 300, + 293, + 369 + ], + "type": "table", + "html": "
EntityFID ↓CCE ↓UD ↓
English Springer20.12 / 18.140.12 / 0.470.07 / 0.34
Van Gogh20.16 / 17.540.07 / 0.520.04 / 0.46
Elon Musk20.15 / 18.220.07 / 0.490.05 / 0.57
Grumpy Cat24.57 / 20.320.18 / 0.540.14 / 0.46
", + "image_path": "786c489e395da048f0a2275ee82609085de63c5b21ce6803d69d75934a5920db.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 371, + 288, + 382 + ], + "lines": [ + { + "bbox": [ + 63, + 371, + 288, + 382 + ], + "spans": [ + { + "bbox": [ + 63, + 371, + 288, + 382 + ], + "type": "text", + "content": "Table 5. Ablation Study for the base method (ESD / UCE)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 399, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 295, + 460 + ], + "type": "text", + "content": "to other E-CEMs in the Subclass, Artistic, and Identity categories, outperforming CA with an FID around 0.85 lower. We found that Instance category is harder to be erased, so we fine-tuned it for 750 steps, resulting in a higher FID compared to the others, fine-tuned for 500 steps." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 466, + 295, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 466, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 295, + 586 + ], + "type": "text", + "content": "Qualitative Results are provided in Fig. 4. Fig. 4a shows results of these models against the CCE attack, and we include qualitative results for UD in the Appendix J. CA, MACE, RACE, and AdvUnlearn are highly vulnerable to this attack. STEREO completely removes these concepts, but at the cost of image quality (" + }, + { + "bbox": [ + 55, + 466, + 295, + 586 + ], + "type": "inline_equation", + "content": "6^{th}" + }, + { + "bbox": [ + 55, + 466, + 295, + 586 + ], + "type": "text", + "content": " column in Fig. 4b). Our method successfully converts \"Springer Dog\" to \"Dog,\" \"Elon Musk\" to a general person, and \"Grumpy Cat\" to \"Cat toy,\" while maintaining image quality comparable to other E-CEM methods (last column in Fig. 4b)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 715 + ], + "type": "text", + "content": "Multiple Concepts Erasure. We provide results for erasing multiple concepts in Fig. 6. We test the ASR with the CCE attack and evaluate two settings: Iterative Erase (IE) and Simultaneous Erase (SE). IE means we fine-tune the weights after erasing one concept to erase others. SE sums all " + }, + { + "bbox": [ + 55, + 594, + 295, + 715 + ], + "type": "inline_equation", + "content": "L_{sub}" + }, + { + "bbox": [ + 55, + 594, + 295, + 715 + ], + "type": "text", + "content": " to erase all concepts in one fine-tuning, with 1000 steps fine-tuning for this experiment. In both settings, our method defends against CCE and maintains model utility. With SE, we achieve better FID scores, as summing " + }, + { + "bbox": [ + 55, + 594, + 295, + 715 + ], + "type": "inline_equation", + "content": "L_{sub}" + }, + { + "bbox": [ + 55, + 594, + 295, + 715 + ], + "type": "text", + "content": " preserves a broader range of knowledge from " + }, + { + "bbox": [ + 55, + 594, + 295, + 715 + ], + "type": "inline_equation", + "content": "L_{reg}" + }, + { + "bbox": [ + 55, + 594, + 295, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 72, + 416, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 416, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 416, + 84 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 95, + 555, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 95, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 95, + 555, + 262 + ], + "type": "text", + "content": "Effect of selecting the Reference Subspace. We present results for four ways of constructing the Reference Subspace, including TI tokens from Steps 30, 50, 70 of TI training, and the concept's general term, as shown in Table 3. Visualizations of each token are in Fig. 5. At Step 30, the token is far from the target, with little relevant information. At Step 50, the token shares features with the target but is not an exact match, enabling minimal transformation. At Step 70, the token is close to the target, making full erasure difficult. These observations align with Table 3, where Step 50 balances ASR, FID, and CLIP scores, while Step 30 improves ASR but slightly reduces image quality. Step 70 shows results similar to CA, indicating SuMa's inefficacy when the Reference Subspace and Target Subspace are too close." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 268, + 556, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 556, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 556, + 449 + ], + "type": "text", + "content": "Effect of choosing the base method. As noted in Section 3.3, we adopt CA loss for our method. Results using ESD [7] and UCE [8] as base methods are shown in Table 5. While ESD resembles CA, it uses a pushing strategy, whereas CA employs a mapping strategy. UCE erases concepts by altering the cross-attention map. Our experiments show that modifying only the attention layer causes the token learned by CCE to express the target concept only under the erased model as shown in Appendix D. ESD provides some defense but increases FID moderately, while UCE fails to defend. We conclude that modifying the model's noise prediction is more effective, as it redirects the target concept's trajectory toward another in the original model. The mapping approach is also favored over the pushing one, as discussed in Section 3.2 and supported by our results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 454, + 556, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 556, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 556, + 574 + ], + "type": "text", + "content": "Effect of adding Subspace Regularization. We present the effect of adding Subspace Regularization and report the results in Table 4. We observe that without regularization, our method performs slightly better against the CCE attack, with an average " + }, + { + "bbox": [ + 313, + 454, + 556, + 574 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 313, + 454, + 556, + 574 + ], + "type": "text", + "content": " reduction in attack rate, but at the cost of a gradual increase in the FID score. This phenomenon was studied in [3], where it was shown that the concept closest to the target is the most affected. Therefore, we conclude that adding Subspace Regularization loss is necessary to maintain the model's effectiveness." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 586, + 471, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 471, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 471, + 597 + ], + "type": "text", + "content": "5. Conclusion and Limitations" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "In this paper, we introduced SuMa, a method that robustly and effectively erases narrow concepts in Text-to-Image Diffusion models. SuMa constructs a Target Subspace for the concept to be erased and maps it to a Reference Subspace. Experimental results show our method is the first to meet these dual objectives for narrow concepts. A key limitation is its time cost, as it may require many iterations to construct both subspaces. Future work will aim to accelerate this process and improve target concept representation." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19594" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "text", + "content": "[1] Dosovitskiy Alexey. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv: 2010.11929, 2020. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 169 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 169 + ], + "type": "text", + "content": "[2] Anh Bui, Khanh Doan, Trung Le, Paul Montague, Tamas Abraham, and Dinh Phung. Removing undesirable concepts in text-to-image generative models with learnable prompts. arXiv preprint arXiv:2403.12326, 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 171, + 295, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 171, + 295, + 214 + ], + "spans": [ + { + "bbox": [ + 62, + 171, + 295, + 214 + ], + "type": "text", + "content": "[3] Anh Bui, Long Vuong, Khanh Doan, Trung Le, Paul Montague, Tamas Abraham, and Dinh Phung. Erasing undesirable concepts in diffusion models with adversarial preservation. arXiv preprint arXiv:2410.15618, 2024. 5, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 216, + 295, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 216, + 295, + 281 + ], + "spans": [ + { + "bbox": [ + 62, + 216, + 295, + 281 + ], + "type": "text", + "content": "[4] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, José Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. In Proceedings of the 40th International Conference on Machine Learning, pages 4055-4075, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 284, + 295, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 284, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 284, + 295, + 327 + ], + "type": "text", + "content": "[5] Zhi-Yi Chin, Chieh-Ming Jiang, Ching-Chun Huang, PinYu Chen, and Wei-Chen Chiu. Prompting4debugging: Redteaming text-to-image diffusion models by finding problematic prompts. arXiv preprint arXiv:2309.06135, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 329, + 295, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 329, + 295, + 382 + ], + "spans": [ + { + "bbox": [ + 62, + 329, + 295, + 382 + ], + "type": "text", + "content": "[6] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel CohenOr. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 385, + 295, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 385, + 295, + 439 + ], + "spans": [ + { + "bbox": [ + 62, + 385, + 295, + 439 + ], + "type": "text", + "content": "[7] Rohit Gandikota, Joanna Materzynska, Jaden Fiitto-Kaufman, and David Bau. Erasing concepts from diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2426-2436, 2023. 1, 2, 4, 6, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 441, + 295, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 441, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 62, + 441, + 295, + 495 + ], + "type": "text", + "content": "[8] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5111-5120, 2024. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 498, + 295, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 498, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 62, + 498, + 295, + 519 + ], + "type": "text", + "content": "[9] Nick Hasty, Ihor Kroosh, Dmitry Voitekh, and Dmytro Korduban. Giphy celebrity detector, 2022. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 521, + 295, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 521, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 57, + 521, + 295, + 564 + ], + "type": "text", + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 567, + 295, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 295, + 588 + ], + "type": "text", + "content": "[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 590, + 295, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 295, + 611 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 295, + 611 + ], + "type": "text", + "content": "[12] Jeremy Howard and Sylvain Gugger. Fastai: a layered api for deep learning. Information, 11(2):108, 2020. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 613, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 656 + ], + "type": "text", + "content": "[13] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "text", + "content": "[14] Harry H Jiang, Lauren Brown, Jessica Cheng, Mehtab Khan, Abhishek Gupta, Deja Workman, Alex Hanna, Johnathan Flowers, and Timnit Gebru. Ai art and its impact on artists. In Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society, pages 363-374, 2023. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[15] Changhoon Kim, Kyle Min, and Yezhou Yang. Race: Robust adversarial concept erasure for secure text-to-image diffusion model. arXiv preprint arXiv:2405.16341, 2024. 1, 3, 6, 7, 4, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 119, + 555, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 555, + 173 + ], + "type": "text", + "content": "[16] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22691-22702, 2023. 1, 2, 3, 4, 6, 7, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 175, + 554, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 554, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 554, + 207 + ], + "type": "text", + "content": "[17] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 209, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 554, + 262 + ], + "type": "text", + "content": "[18] Shilin Lu, Zilan Wang, Leyang Li, Yanzhu Liu, and Adams Wai-Kin Kong. Mace: Mass concept erasure in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6430-6440, 2024. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 264, + 555, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 264, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 317, + 264, + 555, + 330 + ], + "type": "text", + "content": "[19] Mengyao Lyu, Yuhong Yang, Haiwen Hong, Hui Chen, Xuan Jin, Yuan He, Hui Xue, Jungong Han, and Guiguang Ding. One-dimensional adapter to rule them all: Concepts diffusion models and erasing applications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7559-7568, 2024. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 332, + 554, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 332, + 554, + 396 + ], + "spans": [ + { + "bbox": [ + 317, + 332, + 554, + 396 + ], + "type": "text", + "content": "[20] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 399, + 554, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 554, + 463 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 554, + 463 + ], + "type": "text", + "content": "[21] Yong-Hyun Park, Sangdoo Yun, Jin-Hwa Kim, Junho Kim, Geonhui Jang, Yonghyun Jeong, Junghyo Jo, and Gayoung Lee. Direct unlearning optimization for robust and safe text-to-image models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2, 3, 6, 7, 4, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 466, + 554, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 466, + 554, + 519 + ], + "spans": [ + { + "bbox": [ + 317, + 466, + 554, + 519 + ], + "type": "text", + "content": "[22] Minh Pham, Kelly O Marshall, Niv Cohen, Govind Mittal, and Chinmay Hegde. Circumventing concept erasure methods for text-to-image generative models. In The Twelfth International Conference on Learning Representations, 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 522, + 554, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 522, + 554, + 565 + ], + "spans": [ + { + "bbox": [ + 317, + 522, + 554, + 565 + ], + "type": "text", + "content": "[23] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 567, + 554, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 554, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 554, + 588 + ], + "type": "text", + "content": "[24] Robin Rombach. Stable diffusion v1-4 model card., 2022. 1, 7, 4, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 590, + 533, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 533, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 533, + 600 + ], + "type": "text", + "content": "[25] Robin Rombach. Stable diffusion 2.0 release., 2022. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 602, + 555, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 602, + 555, + 656 + ], + "spans": [ + { + "bbox": [ + 317, + 602, + 555, + 656 + ], + "type": "text", + "content": "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 658, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 555, + 712 + ], + "type": "text", + "content": "[27] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 6" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19595" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "type": "text", + "content": "[28] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "text", + "content": "[29] Babak Saleh and Ahmed Elgammal. Large-scale classification of fine-art paintings: Learning the right metric on the right feature. arxiv 2015. arXiv preprint arXiv:1505.00855, 2015.6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 251 + ], + "type": "text", + "content": "[30] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 252, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 252, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 252, + 294, + 306 + ], + "type": "text", + "content": "[31] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by {Text-to-Image} models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 2187–2204, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 308, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 363 + ], + "type": "text", + "content": "[32] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6048-6058, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 365, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 430 + ], + "type": "text", + "content": "[33] Koushik Srivatsan, Fahad Shamshad, Muzammal Naseer, Vishal M Patel, and Karthik Nandakumar. Stereo: A two-stage framework for adversarially robust concept erasing from text-to-image diffusion models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 23765-23774, 2025. 2, 3, 4, 6, 7, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 432, + 233, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 432, + 233, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 432, + 233, + 443 + ], + "type": "text", + "content": "[34] Tonyassi. Celebrity-1000 datasets, 2022. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "type": "text", + "content": "[35] Yu-Lin Tsai, Chia-Yi Hsu, Chulin Xie, Chih-Hsun Lin, Jia-You Chen, Bo Li, Pin-Yu Chen, Chia-Mu Yu, and Chun-Ying Huang. Ring-a-bell! how reliable are concept removal methods for diffusion models? arXiv preprint arXiv:2310.10012, 2023. 1, 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 501, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 566 + ], + "type": "text", + "content": "[36] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228-8238, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 568, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 622 + ], + "type": "text", + "content": "[37] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "type": "text", + "content": "[38] Yimeng Zhang, Xin Chen, Jinghan Jia, Yihua Zhang, Chongyu Fan, Jiancheng Liu, Mingyi Hong, Ke Ding, and Sijia Liu. Defensive unlearning with adversarial training for robust concept erasure in diffusion models. arXiv preprint arXiv:2405.15234, 2024. 2, 3, 6, 7, 4, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "text", + "content": "[39] Yimeng Zhang, Jinghan Jia, Xin Chen, Aochuan Chen, Yihua Zhang, Jiancheng Liu, Ke Ding, and Sijia Liu. To generate or not? safety-driven unlearned diffusion models are still" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 334, + 73, + 553, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 334, + 73, + 553, + 106 + ], + "type": "text", + "content": "easy to generate unsafe images... for now. In European Conference on Computer Vision, pages 385-403. Springer, 2025. 1, 2, 3, 6" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "19596" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Subjective Camera 1.0_ Bridging Human Cognition and Visual Reconstruction through Sequence-Aware Sketch-Guided Diffusion/cae6ed8f-de57-4610-b479-1652f9260751_content_list.json b/2025/Subjective Camera 1.0_ Bridging Human Cognition and Visual Reconstruction through Sequence-Aware Sketch-Guided Diffusion/cae6ed8f-de57-4610-b479-1652f9260751_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..da7b74f5d1488a9a7c494f7d71d608d5da2f4d95 --- /dev/null +++ b/2025/Subjective Camera 1.0_ Bridging Human Cognition and Visual Reconstruction through Sequence-Aware Sketch-Guided Diffusion/cae6ed8f-de57-4610-b479-1652f9260751_content_list.json @@ -0,0 +1,1740 @@ +[ + { + "type": "image", + "img_path": "images/036f49fca5b1146f0dd58d8ee5fc660c76717546e46a77730b006441659d3a92.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 158, + 123, + 202, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Subjective Camera 1.0: Bridging Human Cognition and Visual", + "text_level": 1, + "bbox": [ + 202, + 132, + 836, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reconstruction through Sequence-Aware Sketch-Guided Diffusion", + "text_level": 1, + "bbox": [ + 163, + 156, + 833, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haoyang Chen $^{1,2,3}$ Dongfang Sun $^{1,2}$ Caoyuan Ma $^{1,2}$ Shiqin Wang $^{1,2}$ Kewei Zhang $^{1,2}$ Zheng Wang $^{1,2,3}$ Zhixiang Wang $^{4}$", + "bbox": [ + 133, + 203, + 867, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ National Engineering Research Center for Multimedia Software, Institute of Artificial Intelligence, School of Computer Science, Wuhan University $^{2}$ Hubei Key Laboratory of Multimedia and Network Communication Engineering $^{3}$ Zhongguancun Academy, Beijing, China. 100094 $^{4}$ CyberAgent AI Lab, Japan", + "bbox": [ + 98, + 243, + 903, + 297 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f3172a01beebfc96259e00a3d02b0f8b0756a128e359a6431d8dd59d4d807e32.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 117, + 333, + 140, + 352 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Physical Camera", + "text_level": 1, + "bbox": [ + 145, + 337, + 269, + 351 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Photo after Processing", + "bbox": [ + 133, + 354, + 259, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/198dea298b5766e5660194fc5b560b65a5d8808d85811a9a3a5e5487d22e5e9a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 127, + 371, + 251, + 469 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2e44cc24f8fb45f8fe7385acfcd4dba41e09513702fd2e31e90b0373617f18b0.jpg", + "image_caption": [ + "Figure 1. What is a subjective camera? A subjective camera transforms a person's memory into photorealistic images. We take a small yet crucial step toward this vision by leveraging generative models to sequentially decode textual and sketch-based readouts in their natural order, paving the way for reconstructing moments without physical cameras." + ], + "image_footnote": [], + "bbox": [ + 127, + 484, + 254, + 580 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/781beaaf605b41f5a82caaef17a296da083e4925311b0e88c7a89a5664ab39e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 332, + 446, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The Concept of Subjective Camera", + "text_level": 1, + "bbox": [ + 452, + 337, + 705, + 351 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Readout (Text and Sketch Sequences)", + "bbox": [ + 351, + 354, + 542, + 367 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There is a black car", + "bbox": [ + 294, + 380, + 419, + 396 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "towing a white trailer on the road near mountains.", + "bbox": [ + 292, + 400, + 421, + 452 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There is a motorbike and a tent on the grass.", + "bbox": [ + 290, + 501, + 416, + 554 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6e76ebb9c9b187a325fc221c175ddff7dec200ef986cabc99ef19e3f06bced32.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 375, + 563, + 462 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/da44a85605c00f461abc1335f1f9da5d685896f9f5a65a0d08839818d8f46c1c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 563, + 465, + 591, + 484 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/46beac0003d7a4dfedbf29fd69a6b3bee621f52ed784ce88f46ee68655b6fedd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 354, + 879, + 469 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d870bcf8c119e317df29b4087a76c625145f021f658294d0322cec026198e1df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 428, + 479, + 571, + 580 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1ef267c53bf9908bd17d6b36f1ddc725503d068b70310c85e24991f33458b465.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 593, + 479, + 879, + 579 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 672, + 326, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce the concept of a subjective camera to reconstruct meaningful moments that physical cameras fail to capture. We propose Subjective Camera 1.0, a framework for reconstructing real-world scenes from readily accessible subjective readouts, i.e., textual descriptions and progressively drawn rough sketches. Built on optimization-based alignment of diffusion models, our approach avoids large-scale paired training data and mitigates generalization issues. To address the challenge of integrating multiple abstract concepts in real-world scenar", + "bbox": [ + 86, + 707, + 485, + 858 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ios, we design a Sequence-Aware Sketch-Guided Diffusion framework with three loss terms for concept-wise sequential optimization, following the natural order of subjective readouts. Experiments on two datasets demonstrate that our method achieves state-of-the-art performance in image quality as well as spatial and semantic alignment with target scenes. User studies with 40 participants further confirm that our approach is consistently preferred. Our project page is at: subjective-camera.github.io", + "bbox": [ + 511, + 672, + 908, + 811 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 843, + 645, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cameras have long been the most popular devices for preserving memorable moments. However, a physical camera", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "These authors contributed equally to this work.", + "bbox": [ + 102, + 873, + 364, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding authors.", + "bbox": [ + 102, + 887, + 241, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17838", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cannot always be present to capture every moment. This raises a question of how such unrecorded experiences can be preserved as pixels in the absence of a camera.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper proposes the concept of a subjective camera1, where humans function as \"imaging devices\", aiming to faithfully reconstruct real-world scenes from memory. Specifically, we define a subjective camera as a system through which individuals encode sensory inputs into memory based on personal salience and emotional context, and later read out these stored representations and decode them into pixels that reconstruct scenes they experienced. We draw an analogy to physical cameras: just as a camera records, reads out, and reconstructs the visual world through optical, electronic, and computational processes, the human mind selectively captures (\"records\") perceptual details, which are organized into a retrievable format (\"read out\") and later reconstructs them as pixels.", + "bbox": [ + 88, + 137, + 482, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "It is generally challenging for humans without professional training to directly reconstruct a dense, pixel-wise image. Instead, the most practical approach is to read out the mental scene through textual descriptions and/or free-hand sketches and leverage computational tools to decode these readouts. However, text alone often fails to fully convey a mental image, particularly in terms of object layout and fine details. Freehand sketches provide complementary information beyond text but remain sparse and often contain noise and uncertainty [11, 15]. With the advancement of generative models [20, 23], this long-standing vision is becoming increasingly feasible by decoding such subjective readouts into photorealistic imagery.", + "bbox": [ + 89, + 349, + 482, + 545 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although not intended to reconstruct real-world scenes, several works have explored generating images from combined text and sketch inputs [4, 25, 30, 35-37]. These approaches either adapt model weights using large-scale paired data [19, 25, 28, 30, 33, 36, 37] or modify the latent variables through per-scene optimization [4, 28, 35]. Training-based methods directly learn the mapping from sketches to pixels. However, they require extensive paired data and struggle with user-specific biases, often generalizing poorly to abstract sketches beyond the training distribution. In contrast, training-free methods avoid the heavy data requirements, high computational costs, and generalization challenges of training-based approaches.", + "bbox": [ + 89, + 546, + 482, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, optimization-based methods often fail to reconstruct multiple concepts from idiosyncratic sketches, as would be required for recreating real-world scenes. This often leads to missing concepts or misaligned spatial relationships (see Figure 2a). The main obstacle lies in simultaneously interpreting multi-concept sketches, which inherently combine varying levels of abstraction and randomness across different concepts. For instance, freehand sketches", + "bbox": [ + 89, + 742, + 482, + 863 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a35ba4a51f5908149de79a0d33288e1b6f0e75abf38588ca6171cdc6f24acf02.jpg", + "image_caption": [ + "a: Limitation of \"snapshot\" generation" + ], + "image_footnote": [], + "bbox": [ + 555, + 89, + 864, + 160 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2c41e85fa300491fce409300694f16caa528e5f6e7f20fb0cd20939030c53796.jpg", + "image_caption": [ + "b: Inspiration from human readout order", + "Figure 2. Motivation for concept-wise sequential optimization. (a) Attempting to \"snap\" all concepts into a complete image at once using generative diffusion priors often leads to suboptimal results, such as distortions and misaligned semantics or layouts. (b) In contrast, humans reconstruct mental scenes step by step. This natural readout inspires our design2, making it better suited to leverage current generative diffusion priors." + ], + "image_footnote": [], + "bbox": [ + 555, + 181, + 861, + 252 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "may only roughly indicate object boundaries, omit fine details such as texture and geometry, or distort proportions in non-standard ways. While generative priors can fill sparse regions and mitigate randomness during the transformation from sketches to pixels, attempting to integrate all concepts at once often reduces the problem to a least-squares solution [1, 4], ultimately blurring their distinct characteristics.", + "bbox": [ + 511, + 407, + 903, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our key idea is simple — to apply optimizations to each concept individually and do so sequentially. This design considers not only the limitations of the \"snapshot\" generation process but also human cognition. Humans tend to describe their mental images step-by-step [13], typically starting with a textual description of abstract concepts or the overall scene structure, and then progressively refining individual elements with increasing detail (see Figure 2b).", + "bbox": [ + 511, + 513, + 905, + 635 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building upon this idea, we introduce a sequence-aware, sketch-guided diffusion framework that sequentially optimizes the latent noise of a pre-trained text-to-image (T2I) diffusion model [23] to align individual subjective readouts. The order of these optimizations mirrors the natural process of reading out subjective impressions: it begins with a text-reward optimization (Section 3.2) to align the generated image with the initial textual description, followed by concept-wise sequential optimizations (Section 3.5) to incorporate progressively provided sketches that may encode shapes, poses, and spatial arrangements. The sequential optimization also incorporates a loss to ensure that previously optimized concepts are not significantly altered.", + "bbox": [ + 511, + 637, + 905, + 833 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Two additional components are proposed to improve", + "bbox": [ + 532, + 834, + 903, + 849 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1 While the term \"subjective camera\" appears in cinematic contexts, our usage here differs significantly in meaning.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "2This design philosophy is reminiscent of the progressive decoding used in rolling-shutter image sensors, which arises from the need to accommodate the limited processing capacity of the circuits.", + "bbox": [ + 511, + 862, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17839", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the concept-wise sequential optimization. First, each input sketch is individually encoded into the latent space to provide spatial guidance. To bridge the gap between rough sketch shapes and their corresponding real-world forms, we introduce an optimization-based inversion strategy for refining this spatial guidance (Section 3.3). Second, because input sketches often lack appearance details, directly optimizing with their guidance over multiple iterations may lead to unnatural textures and degraded visual fidelity. We therefore leverage the latent representation obtained from text-reward optimization to define an appearance loss (Section 3.4), thereby preserving visual style quality.", + "bbox": [ + 89, + 90, + 480, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By processing sketches individually and sequentially, our approach naturally aligns with the human cognition process, enabling users to incrementally construct mental images while ensuring each addition respects and complements previous elements. This sequence-aware generation strategy provides fine-grained control over complex multi-concept scenes while preserving the user's subjective intent encoded in the drawing order. Evaluations on two datasets demonstrate that our method achieves state-of-the-art performance in recreating a real-world scene from human memory, demonstrating the potential of subjective cameras. Notably, our framework is entirely training-free, eliminating the need for large-scale paired datasets and avoiding generalization.", + "bbox": [ + 89, + 273, + 480, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To sum up, our contributions are twofold:", + "bbox": [ + 109, + 487, + 385, + 501 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "We propose the concept of a subjective camera, which functions the human as an \"imaging device\" and leverages computational tools to transform cognitive impressions into photographs.", + "We propose a sequence-aware diffusion-guided generation framework that enables faithful reconstruction of complex, multi-concept scenes. This framework aligns with human cognition and surpasses \"snapshot\" generation methods, thereby establishing a new paradigm for cognition-driven image generation." + ], + "bbox": [ + 99, + 503, + 478, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 672, + 232, + 689 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Aligning Image Synthesis Models Despite their success, T2I generative models [20, 22, 23] often fail to reproduce the fine-grained semantics and compositional details described in complex prompts. Reward-based alignment has emerged as a promising direction [6, 7, 12, 14, 26, 31, 32], using human preference models [12, 29, 32] to guide generation. The human preference models were trained on paired human preference data. Early works [12, 32] fine-tune diffusion models with the reward models, whereas later methods [6, 14] sidestep costly fine-tuning by directly optimizing the latent noise [6, 31]. However, these techniques largely improve global prompt adherence while neglecting spatial control. Our model not only aligns the T2I genera", + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tive models with the text prompt, but also ensures the spatial information to respect to user-provided sketches.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Controllable Image Synthesis T2I models inherently rely on highly compressed textual tokens, which limits their controllability and often prevents them from meeting user expectations [20, 22, 23]. To overcome these limitations, a growing body of research has explored controllable generation by incorporating fine-grained conditioning signals, such as reference images [2], edge maps or contours or skeletons [33], semantic layouts [1, 3, 38], and lighting specifications [34]. One line of work fine-tunes pretrained diffusion models by adding trainable modules [19, 33, 37], while others directly retrain the diffusion backbone by minimizing reconstruction objectives [2]. In parallel, training-free techniques [5, 7, 9, 26] have been developed to enhance controllability without modifying model weights. These include attention injection or latent optimization strategies, such as Prompt-to-Prompt editing [9] and diffusion self-guidance [5]. Despite these advances, existing methods often fail in complex scenes involving multiple guidance given by freehand sketches, where interactions between elements lead to interference and degraded synthesis quality. Building on these insights, our method incorporates attention-based guidance extraction but extends it to better handle multi-concept freehand sketches.", + "bbox": [ + 511, + 125, + 903, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sketch-to-Image Synthesis It is the task most relevant to our concept of the subjective camera. It aims to generate images from freehand sketches along with textual prompts. Several recent efforts have attempted to adapt T2I diffusion models for sketch-to-image synthesis [4, 17, 25, 28, 35, 36]. Training-based approaches [25, 36] fine-tune the T2I models with paired data. However, this training-based method incurs substantial computational cost and suffers from poor generalization. In contrast, training-free approaches [4, 17, 35] avoid fine-tuning by guiding the diffusion denoising process with sketches. However, these methods still jointly create multiple concepts, causing interference between their differing levels of randomness and abstraction. More critically, both training-based and training-free approaches fail to address the subjective biases inherent in user-provided sketches and are particularly vulnerable to highly abstract inputs, often leading to appearance distortions and semantic inconsistencies in the generated outputs. Furthermore, unlike creative generation methods, our focus is on reconstructing real-world scenes.", + "bbox": [ + 511, + 476, + 903, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Sequence-Aware Sketch-Guided Diffusion", + "text_level": 1, + "bbox": [ + 511, + 792, + 885, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 511, + 816, + 624, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to render photorealistic images from subjective readouts, consisting of a textual description $p$ and a sequence $S = \\{s_i\\}_{i=1}^N$ of $N$ freehand sketches, where each sketch represents a distinct concept. As illustrated", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17840", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e491c29db16566f6abc2ac861ee480133a7e2c9045a62d4a5b2605baff667363.jpg", + "image_caption": [ + "Figure 3. Sequence-aware sketch-guided diffusion. (Left) User-provided textual description undergoes text-to-image generation and text-reward optimization to obtain the initial latent. Then, the following generation progresses under three constraints — spatial layout conforming to sketch topology while appearance details adhere to the initial image from text-reward optimization, with consistency maintained through ordered latent propagation and three loss terms, ensuring coherent integration of emerging and established scene elements. (Right) The concept-wise sequential optimization." + ], + "image_footnote": [], + "bbox": [ + 94, + 90, + 903, + 265 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in Figure 3, we design a Sequence-Aware Sketch-Guided Diffusion framework to accomplish this challenging task. Our method processes the subjective readouts individually and sequentially, following the order in which users describe them. It employs a sequence-aware optimization strategy to align a pre-trained generative diffusion model, $\\mathcal{G} := \\epsilon_{\\theta}(\\epsilon_{\\theta}(\\dots \\epsilon_{\\theta}(\\epsilon ,p)))$ , and $\\epsilon_{\\theta}$ denotes the pre-trained U-Net, with these readouts, naturally mirroring the human cognitive process of step-by-step scene construction. The process begins with a text-reward optimization, which establishes an initial latent representation and appearance prior for subsequent concept-wise optimizations. The appearance prior is extracted via style encoding, while sketch encoding, implemented using diffusion inversion, provides spatial guidance for individual sketches. In addition, a preserving loss is introduced to maintain previously optimized concepts throughout the sequential process. These components collectively enable our framework to decode subjective readouts into coherent, photorealistic scenes.", + "bbox": [ + 91, + 375, + 483, + 661 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Text-Reward Optimization", + "text_level": 1, + "bbox": [ + 89, + 670, + 334, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We randomly sample a latent noise $\\epsilon$ and feed it, together with the textual description $p$ , into a pre-trained T2I diffusion model, which renders an image through the full reverse diffusion process, $\\mathcal{G}_{\\theta}$ . To ensure that the generated image contains the concepts in the given textual information, without concept omission or significant misalignment, we formulate a text-reward optimization procedure as existing practice [6, 14] that iteratively refines the latent noise to maximize a text-image alignment objective:", + "bbox": [ + 89, + 691, + 483, + 829 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon^ {0 \\star} = \\underset {\\epsilon} {\\arg \\max } \\mathcal {C} _ {F} \\left(\\mathcal {G} _ {\\theta} (\\epsilon , p), p\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 837, + 482, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{C}_F$ represents a differentiable reward model that measures alignment between the generated image and the", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "text prompt $p$ . The optimization process iteratively refines $\\epsilon$ through gradient ascent:", + "bbox": [ + 511, + 375, + 906, + 405 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {t + 1} = \\epsilon_ {t} + \\eta \\nabla_ {\\epsilon} \\mathcal {C} _ {F} (\\mathcal {G} _ {\\theta} (\\epsilon_ {t}, p)), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 417, + 906, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\eta$ is the learning rate.", + "bbox": [ + 511, + 445, + 700, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Sketch Encoding", + "text_level": 1, + "bbox": [ + 511, + 469, + 679, + 486 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Before proceeding to the sketch-guided optimization, we first encode the sketch to extract high-level semantic guidance, rather than relying on low-level line details. Given a list of input sketches $\\{s_i\\}_{i=1}^N$ , each sketch $s_i$ is encoded into a latent representation $z_i \\in \\mathbb{R}^{h \\times w \\times d}$ using the variational autoencoder (VAE) encoder $E_{\\phi}$ of the diffusion model: $z_i = E_{\\phi}(s_i)$ . Then, we can leverage arbitrary diffusion inversion algorithms, such as DDIM inversion [18] or SDEdit [16], to map the sketch into latent noise. To address the significant domain gap between abstract planar sketches and realistic images, we introduce an optimization procedure that iteratively refines the latent representation $z_i$ by aligning it with the rich prior knowledge embedded in the diffusion model.", + "bbox": [ + 511, + 492, + 906, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At each optimization step $k$ , we perturb the latent sketch representation with a Gaussian noise $\\epsilon_{k} \\sim \\mathcal{N}(0, \\mathbf{I})$ to obtain a noisy latent representation", + "bbox": [ + 511, + 703, + 906, + 750 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t} ^ {(k)} = \\alpha_ {t _ {k}} z _ {i} ^ {(k)} + \\sigma_ {t _ {k}} \\epsilon_ {k}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 758, + 906, + 780 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha_{t_k}$ and $\\sigma_{t_k}$ are time-dependent scaling factors following the noise schedule of the diffusion process. The noisy latent $z_{t}^{(k)}$ is then fed into the pre-trained U-Net $\\epsilon_{\\theta}$ which predicts the noise component $\\hat{\\epsilon}_{\\theta}(z_{t}^{(k)}, t_{k})$ . The noise prediction error $\\Delta \\epsilon_{k}$ is computed as:", + "bbox": [ + 511, + 790, + 905, + 871 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\epsilon_ {k} = \\hat {\\epsilon} _ {\\theta} \\left(z _ {t} ^ {(k)}, t _ {k}\\right) - \\epsilon_ {k}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 882, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17841", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ba965996a6b33632add424df6e05c58ac886742064647239aecf7781edf65720.jpg", + "image_caption": [ + "Figure 4. The effectiveness of sketch optimization. Without optimization (left), the generated image fails to accurately respect the wheel placement and proportions indicated in the sketch. With our optimization (right), the generated result better aligns with the sketch structure, particularly in the wheel positions (green arrows) and overall shape, demonstrating improved spatial coherence." + ], + "image_footnote": [], + "bbox": [ + 117, + 88, + 459, + 175 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This noise gap serves as a gradient signal to iteratively refine the latent sketch features. Specifically, we update the latent representation $z_{i}^{(k)}$ as", + "bbox": [ + 89, + 294, + 483, + 340 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {i} ^ {(k + 1)} = z _ {i} ^ {(k)} - \\eta_ {k} \\left(\\Delta \\epsilon_ {k} + \\lambda \\left(z _ {i} ^ {(k)} - z _ {i} ^ {(0)}\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 347, + 482, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is a regularization term that anchors the optimization to the initial latent $z_{i}^{(0)}$ , preventing semantic drift.", + "bbox": [ + 89, + 378, + 482, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Through this iterative optimization in Eq. (3-5), the latent $\\dot{z}_i$ is pushed to align with the rich prior knowledge embedded in the diffusion model. This alignment effectively infuses the sketch with realistic details, textures, and structural coherence, thereby establishing a robust mapping from human hand-drawn sketches to physically grounded visual representations (see Figure 4).", + "bbox": [ + 89, + 411, + 483, + 516 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, we observed that the optimization method is less effective when the sketch is drawn relatively small. In such cases, the generated result often fails to align with the sketch and instead produces a shape-independent object in a central position. To address this, we suggest scaling the original sketch using linear interpolation or, alternatively, discarding the optimization for that step.", + "bbox": [ + 89, + 517, + 483, + 623 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Appearance Encoding", + "text_level": 1, + "bbox": [ + 89, + 630, + 299, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While the sketch provides spatial information, additional appearance cues are required to ensure visual quality during optimization. We first define a operator $\\mathcal{A}(\\cdot)$ to obtain the appearance descriptor:", + "bbox": [ + 89, + 651, + 483, + 714 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} (x) = \\left\\{\\frac {\\sum_ {u , v} \\operatorname {s i g m o i d} ([ x ] _ {u v}) [ f _ {x} ] _ {u v}}{\\sum_ {u , v} \\operatorname {s i g m o i d} ([ x ] _ {u v})} \\mid f _ {x} \\right\\}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 718, + 482, + 760 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which is a collection of weighted spatial means of diffusion features $f_{x}$ and $f_{x}$ are from different layer and time steps. Then, we treat the descriptor extracted from the optimized latent $z^{0\\star}$ , derived from $\\epsilon^{0\\star}$ in Eq. (1), as our appearance prior. We will compare the appearance descriptor of the current latent and the appearance prior during optimization.", + "bbox": [ + 89, + 763, + 483, + 856 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Concept-wise Sequential Optimization", + "text_level": 1, + "bbox": [ + 513, + 90, + 844, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we leverage this order information of subjective readouts to progressively construct the scene while maintaining coherence between newly introduced and previously established elements.", + "bbox": [ + 511, + 112, + 905, + 172 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Spatial Attention-guided Control We leverage the spatial attention mechanisms within the diffusion model's U-Net architecture. For each sketch concept $s_i$ , we extract cross-attention maps that highlight regions in the latent space corresponding to specific semantic concepts:", + "bbox": [ + 511, + 176, + 905, + 252 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} (x) = \\operatorname {s o f t m a x} \\left(\\frac {Q (x) K (x) ^ {T}}{\\sqrt {d}}\\right) > \\tau , \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 258, + 906, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Q(x)$ and $K(x)$ are query and key matrices, $d$ is the dimensionality of the key vectors, and $\\tau$ is a threshold parameter. The attention maps $\\mathcal{M}(\\dot{z}_i)$ extracted from $\\dot{z}_i$ , the latent of the sketch $s_i$ (Section 3.3), serve as spatial indicators to guide the concept's layout. The control is given through the spatial loss, defined as:", + "bbox": [ + 511, + 300, + 905, + 391 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s p a t i a l}} \\left(z _ {i}, \\dot {z} _ {i}\\right) = \\frac {\\sum_ {u , v} \\left[ \\mathcal {M} \\left(\\dot {z} _ {i}\\right) \\right] _ {u v} \\left\\| \\left[ \\dot {z} _ {i} \\right] _ {u v} - \\left[ z _ {i} \\right] _ {u v} \\right\\| _ {2} ^ {2}}{\\sum_ {u , v} \\left[ \\mathcal {M} \\left(\\dot {z} _ {i}\\right) \\right] _ {u v}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 398, + 905, + 438 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $u, v$ are spatial positions indices.", + "bbox": [ + 513, + 444, + 777, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Appearance Control We define the appearance loss as:", + "bbox": [ + 513, + 463, + 885, + 479 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {a p p}} \\left(z _ {i}, z ^ {0 \\star}\\right) = \\left\\| \\mathcal {A} \\left(z _ {i}\\right) - \\mathcal {A} \\left(z ^ {0 \\star}\\right) \\right\\| _ {2} ^ {2}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 484, + 905, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{A}$ is the operator of exacting appearance descriptors.", + "bbox": [ + 511, + 513, + 905, + 530 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sequential Concept Integration To maintain coherence across the progressive addition of concepts, we implement a sequential integration process that updates the latent noise representation while preserving previously incorporated elements. For each sketch $s_i$ in the sequence, we compute:", + "bbox": [ + 511, + 532, + 905, + 608 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {i} = z _ {i - 1} + \\Delta z _ {i}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 648, + 617, + 903, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $z_{i-1}$ is the optimized latent from the previous concept, and $\\Delta z_i$ represents the incremental update required to incorporate the current sketch concept $s_i$ . This update is derived from our loss function:", + "bbox": [ + 511, + 641, + 905, + 700 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i} = \\alpha_ {i} \\mathcal {L} _ {\\text {n e w}} \\left(z _ {i}, \\dot {z} _ {i}, z ^ {0 \\star}\\right) + \\beta_ {i} \\mathcal {L} _ {\\text {p r e s e r v e}} \\left(z _ {i}, \\dot {z} _ {0, \\dots , i - 1}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 709, + 903, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha_{i}$ and $\\beta_{i}$ are balancing coefficients. $\\mathcal{L}_{\\mathrm{new}}$ ensures proper integration of the current $z_{i}$ and is defined as:", + "bbox": [ + 511, + 734, + 905, + 765 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {n e w}} \\left(z _ {i}, \\dot {z} _ {i}, z ^ {0 \\star}\\right) = \\mathcal {L} _ {\\text {s p a t i a l}} \\left(z _ {i}, \\dot {z} _ {i}\\right) + \\mathcal {L} _ {\\text {a p p}} \\left(z _ {i}, z ^ {0 \\star}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 772, + 903, + 791 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "while $\\mathcal{L}_{\\mathrm{preserve}}$ maintains the integrity of previously optimized concepts, given by:", + "bbox": [ + 511, + 797, + 903, + 829 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p r e s e r v e}} \\left(z _ {i}, \\dot {z} _ {0, \\dots , i - 1}\\right) = \\sum_ {j = 1} ^ {i - 1} \\gamma^ {i - j} \\mathcal {L} _ {\\text {s p a t i a l}} \\left(z _ {i}, \\dot {z} _ {j}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 835, + 903, + 878 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "3To avoid ambiguity, we use $\\dot{z}_i$ to denote the optimized latent for the $i$ -th sketch concept, distinguishing it from latent $z_i$ would be used in the concept-wise sequential optimization.", + "bbox": [ + 89, + 862, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "4In fact, an inner-loop optimization is omitted here for simplicity.", + "bbox": [ + 529, + 886, + 877, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17842", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/44bc7404717272fe3bbed7925a95892c497e415a622eefe6a972f57bdf68e924.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison of existing sketch-to-image generation methods, and our method on FMC and CMC datasets. Existing training-based methods tend to overfit to sketch lines, often failing to reconstruct scenes realistically, while training-free methods frequently misinterpret or overlook key concepts. In comparison, our approach more faithfully represents real-world physical scenes. The reading and processing order of the sketches is indicated by colors as follows: $\\bullet$ $\\bullet$ $\\bullet$" + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 898, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $j$ means preserving established concepts, and $\\gamma^{i - j}$ is a decay factor that weights the importance of earlier concepts based on their step gap from the current step.", + "bbox": [ + 89, + 561, + 482, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments and Results", + "text_level": 1, + "bbox": [ + 89, + 622, + 323, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 89, + 648, + 266, + 664 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets To rigorously evaluate our method, we have undertaken extensive efforts to develop two specialized datasets: the Composed Multi-Concept (CMC) dataset and the Freehand Multi-Concept (FMC) dataset. Both datasets are designed to comprehensively assess the capabilities of our framework in handling multi-concept scene understanding and generation. CMC dataset comprises 142 scene sketches, each containing 2-4 distinct concepts sampled from the 150 semantic classes in the Sketchy dataset [24]. Each is accompanied by annotated textual prompts with the help of multimodal large language models. The CMC dataset does not provide real-world images captured by physical cameras for reference. FMC dataset is constructed from freehand sketches drawn by volunteers, based on real-world images sourced from publicly available outdoor", + "bbox": [ + 89, + 674, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "scene datasets. It comprises 42 hand-drawn sketches, each containing 2-3 distinct concepts. Additionally, each sketch is paired with a corresponding real-world image as a reference for validation and evaluation purposes.", + "bbox": [ + 511, + 561, + 906, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics To conduct a comprehensive quantitative evaluation of the quality of generated images, four widely-used evaluation metrics were employed, including Frechet Inception Distance (FID) [10], CLIP-T Score (CLIP-T) [21], Human Preference Score (HPS) [29], and CLIP-I distance (CLIP-I) [21]. FID measures the feature distribution similarity between generated and real images. Lower FID values indicate better distribution matching $(\\downarrow)$ . CLIP-T quantifies the image-text alignment degree within the CLIP-T embedding space. A higher CLIP-T score reflects greater semantic alignment between the generated image and text prompt $(\\uparrow)$ . HPS measures the subjective quality of the generated images based on human evaluations. Higher HPS values suggest that the generated images are more favorable or preferred by human evaluators $(\\uparrow)$ . CLIP-I measures the alignment between two images. A higher value suggests that the two images are more consistent in terms of their semantic content $(\\uparrow)$ .", + "bbox": [ + 511, + 628, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17843", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d895d59adab5fa97f378c744bd46e240af54896a72d2c4e7bd4b9c5f38a90c6f.jpg", + "image_caption": [ + "Figure 6. Ablation study on FMC and CMC datasets. Column headers are defined as follows: T2I denotes the baseline text-to-image diffusion model; OP (Overall Perception) indicates the single-step guidance approach utilizing complete sketch as a holistic conditioning signal; CSO refers to our proposed Concept-wise Sequential Optimization (Section 3.5); SE represents the Sketch Encoding module (Section 3.3) that enhances structural plausibility by injecting diffusion priors into imperfect sketches. The qualitative comparison demonstrates that CSO achieves finer layout control than OP. While SE effectively bridges the domain gap between freehand sketches and physical reality through latent space regularization, ultimately elevating generation quality." + ], + "image_footnote": [], + "bbox": [ + 130, + 83, + 861, + 234 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4a7a71b0db0b6b949857a860536601479a106e5e76e23a85f2d5a475469d0e21.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Metric MethodCMCFMC
FID ↓CLIP-T ↑HPS ↑FID ↓CLIP-T ↑HPS ↑CLIP-I ↑
Training-based methods
ControlNet [33]15.2861.710.79414.4566.800.8020.622
T2I-Adapter [19]18.1361.610.74515.3761.610.7450.689
Uni-ControlNet [37]19.6062.580.76118.5362.580.7610.712
Training-free methods
Multi-Sketch [4]13.0564.110.74011.5972.020.7750.690
PNP [27]17.0163.270.84016.2671.240.8520.593
FreeControl [17]14.5264.740.75711.1766.670.7590.775
Ours11.5666.640.85010.1874.250.8620.797
", + "bbox": [ + 91, + 358, + 478, + 469 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative results on CMC and FMC datasets. Our method consistently surpasses all training-free methods in distribution matching, image-text alignment, subjective quality, and appearance details as measured by FID, CLIP-T, HPS, and CLIP-I. Our method achieves superior image-text alignment compared to training-based methods ( $\\downarrow$ indicates lower values are better, while $\\uparrow$ is opposite).", + "bbox": [ + 88, + 481, + 482, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details In the implementation of our proposed method, we set the learning rate $\\eta$ in Eq. (2) and the threshold parameter $\\tau$ in Eq. (7) is 5.0 and 0.3, respectively. Meanwhile, hyperparameters $\\alpha_{i}$ and $\\beta_{i}$ in Eq. (11) are set as 0.8 and 0.2, respectively. $\\gamma$ in Eq. (13) is set as 0.9. Our method is performed on one NVIDIA GeForce RTX 4090, and input sketch images are $512 \\times 512$ in resolution.", + "bbox": [ + 88, + 607, + 483, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparisons", + "text_level": 1, + "bbox": [ + 89, + 724, + 230, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conducted a comprehensive quantitative and qualitative evaluation of the proposed method and competing methods on the CMC and FMC datasets. Our competing methods include the training-based approaches ControlNet [33], T2I-Adapter [19], and Uni-ControlNet [37], as well as the training-free approaches Multi-Sketch [4], PNP [27], and FreeControl [17]. The quantitative results are presented in Table 1 and qualitative results are shown in Figure 5.", + "bbox": [ + 89, + 748, + 482, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We observe that training-based methods [19, 33, 37] are highly susceptible to subjective biases in sketches and of", + "bbox": [ + 89, + 872, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ten fail to adhere to text prompts, producing distorted and unrealistic images with lower CLIP-T scores. Training-free methods [4, 8, 17] suffer from concept omissions, conspicuous artifacts, and inconsistencies both among concepts and between concepts and their backgrounds, resulting in poor image realism. These issues stem from mutual interference during the \"snapshot\" generation of multiple concepts. In contrast, our proposed method faithfully reconstructs real-world scenes and achieves the best performance across all evaluation metrics on both the CMC and FMC datasets, demonstrating its effectiveness.", + "bbox": [ + 511, + 363, + 906, + 529 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bb3ee8fc8f5cbc78041af524b18487d10b76b38b4a5760795f2a00c1a23c44c8.jpg", + "image_caption": [ + "Figure 7. User preference experiment. We asked participants to answer four questions based on the reference images, subjective readouts, and generated images, which included one produced by our method and one by a competing method. Results show that users significantly preferred our outputs over those of competing methods in most cases." + ], + "image_footnote": [], + "bbox": [ + 514, + 539, + 903, + 621 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 742, + 666, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To validate the effectiveness of each module in our method, we conducted ablation experiments on the FMC and CMC datasets, with the visualization results shown in Figure 6.", + "bbox": [ + 511, + 763, + 905, + 810 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "It is observed that traditional T2I diffusion models (T2I) generate images that are close to text prompts. While incorporating sketches as guidance further constrains the consistency of concept positioning, deviations in appearance and spatial alignment persist (T2I, OP). The sequential concept understanding facilitated by Concept-wise Sequential", + "bbox": [ + 511, + 811, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17844", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1725d49d543ff881a7f5b1e56687215f6fcc6a05644b4cbc5156804a37dadc4b.jpg", + "image_caption": [ + "Figure 8. Control experiments on user preferences. We controlled three factors to examine their influence: whether the rater drew the photo, the availability timing of the reference image ( $\\infty$ vs. 5s), and the types of questions asked. In general, our method is consistently better than the competing method, FreeControl." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 480, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Optimization (CSO) in Section 3.5 enhances the alignment of concept positioning and appearance with input sketches (CSO). Simultaneously, the Sketch Encoding (SE) module in Section 3.3 aligns content latent variables with real-world objects, mitigating geometric misalignment (T2I, OP, SE). When integrating all modules, our method faithfully reconstructs real-world scenes and achieves optimal results (Ours), demonstrating the effectiveness of each module and underscoring the indispensable nature of their design.", + "bbox": [ + 89, + 292, + 483, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. User Preference Experiments", + "text_level": 1, + "bbox": [ + 89, + 440, + 352, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We invited 12 participants (group A) to provide subjective readouts and an additional 28 participants (group B) to evaluate the reconstructed images. Participants in group A were first asked to select a deeply impressive scene from their photo gallery and then provide a textual description along with sketches within two minutes. Our method and competing approaches, including two baselines (text-reward optimization and text-reward optimization with sketch-guided optimization based on all concepts) as well as an optimal method, FreeControl [17], were used to generate images from these readouts. Participants in both groups were then shown the subjective readouts, the reference image, and the generated images for the same scene. They were asked to answer four single-choice questions in a single-blind manner: Q1: Which image is closer to the original image (both style and content)? Q2: Which image has higher visual quality (e.g., more realistic, more natural)? Q3: Which image better matches the structure of the sketch? Q4: Which image better matches the textual description?", + "bbox": [ + 89, + 462, + 482, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 7 presents the statistics of the collected results. Participants judged our reconstructions as being closer to the corresponding real-world scenes compared to competing methods. Similar conclusions were observed for the other three questions. However, there was an exception in the comparison between our method and text-reward optimization regarding image quality (Q2). We observed that text-reward optimization without additional control tends to hallucinate details that enhance perceived realism, leading some users to prefer it for Q2. This also explains why our", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "advantage over text-reward optimization in Q1 was not statistically significant and why some participants expressed hesitation in this case.", + "bbox": [ + 511, + 90, + 903, + 135 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further conducted controlled experiments to examine the effects of three factors: whether the rater provided the subjective readouts, the availability timing of the reference image, and the types of questions asked. Results in Figure 8 show that our method consistently outperformed FreeControl [17], regardless of these variations. Notably: 1) Participants who provided subjective readouts showed slightly lower preference for our method in terms of similarity to the reference image compared to those who did not, while showing a higher preference in terms of alignment with the readouts and perceived image quality. We suppose this is mainly because of the subjectivity in reading out. 2) Reducing the time participants could view the reference image decreased the preference for our method; when allowed to compare freely, participants preferred our results more strongly.", + "bbox": [ + 511, + 136, + 906, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 392, + 638, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presented Subjective Camera, a paradigm for reconstructing real-world scenes from humans' mental impressions. Our framework relies on the most accessible readouts of mental imagery, i.e., textual descriptions and sketches, and employs a training-free, sequence-aware optimization process. This approach eliminates the need for costly fine-tuning of text-to-image diffusion models on large-scale data while mitigating challenges such as dependence on sketch quality and geometric misalignment. Extensive experiments on both synthetic and real-world sketch datasets demonstrate that Subjective Camera achieves state-of-the-art performance, outperforming both training-free and training-based sketch-to-image generation methods across all evaluation metrics. Furthermore, user preference studies confirm that participants consistently favor our method.", + "bbox": [ + 511, + 417, + 903, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In closing, we address a natural question: \"With the rapid advancement of generative AI, will physical cameras ever become unnecessary\" For now, the answer is no. First, the reconstructed images still fall short of perfectly matching actual scenes. This limitation largely stems from the fact that current T2I models are trained on highly diverse data, creating gaps between the prior and the specific scene to be reconstructed. One promising direction to mitigate this is to personalize the prior models using user-specific galleries. Second, we observe that variations in the order and quality of sketches have a significant impact on reconstruction results. To address this, we plan to enhance our framework with computational tools that guide users in providing more precise and suitable subjective readouts. For these reasons, we designate this system as Subjective Camera 1.0, leaving these improvements for future iterations.", + "bbox": [ + 511, + 659, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17845", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements This work was supported in part by the National Natural Science Foundation of China (62225113, 624B2109, 623B2079) and the Zhongguancun Academy Project (20240308). We thank Po-Hsun Huang for providing the questionnaire system for our user study, Kaijin Zheng for assistance with data processing, and all participants who took part in the studies.", + "bbox": [ + 89, + 90, + 485, + 198 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 210, + 187, + 226 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. Proceedings of Machine Learning Research, 202:1737-1752, 2023. 2, 3", + "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18392-18402, 2023. 3", + "[3] Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5343-5353, 2024. 3", + "[4] Zhenwei Cheng, Lei Wu, Changshuo Wang, and Xiangxu Meng. Scene sketch-to-image synthesis based on multi-object control. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3775-3779. IEEE, 2024. 2, 3, 7", + "[5] Dave Epstein, Allan Jabri, Ben Poole, Alexei Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. Advances in Neural Information Processing Systems, 36:16222-16239, 2023. 3", + "[6] Luca Eyring, Shyamgopal Karthik, Karsten Roth, Alexey Dosovitskiy, and Zeynep Akata. Reno: Enhancing one-step text-to-image models through reward-based noise optimization. Advances in Neural Information Processing Systems, 37:125487-125519, 2025. 3, 4", + "[7] Xiefan Guo, Jinlin Liu, Miaomiao Cui, Jiankai Li, Hongyu Yang, and Di Huang. Initno: Boosting text-to-image diffusion models via initial noise optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9380-9389, 2024. 3", + "[8] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In The Eleventh International Conference on Learning Representations. 7", + "[9] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control.(2022). URL https://arxiv.org/abs/2208.01626, 3, 2022. 3", + "[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6", + "[11] Zechao Hu, Zhengwei Yang, Hao Li, Yixiong Zou, Fengbin Zhu, and Zheng Wang. Unified category and style generalization for instance-level sketch retrieval. In Proceedings of" + ], + "bbox": [ + 93, + 234, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, page 864-873. Association for Computing Machinery, 2025. 2", + "[12] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Martiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36: 36652-36663, 2023. 3", + "[13] Stephen M Kosslyn. Aspects of a cognitive neuroscience of mental imagery. Science, 240(4859):1621-1626, 1988. 2", + "[14] Seung Hyun Lee, Yinxiao Li, Junjie Ke, Innfarn Yoo, Han Zhang, Jiahui Yu, Qifei Wang, Fei Deng, Glenn Entis, Junfeng He, et al. Parrot: Pareto-optimal multi-reward reinforcement learning framework for text-to-image generation. In European Conference on Computer Vision, pages 462-478. Springer, 2024. 3, 4", + "[15] Kejun Lin, Zhixiang Wang, Zheng Wang, Yinqiang Zheng, and Shin'ichi Satoh. Beyond domain gap: Exploiting subjectivity in sketch-based person retrieval. In Proceedings of the 31st ACM international conference on multimedia, pages 2078-2089, 2023. 2", + "[16] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 4", + "[17] Sicheng Mo, Fangzhou Mu, Kuan Heng Lin, Yanli Liu, Bochen Guan, Yin Li, and Bolei Zhou. Freecontrol: Training-free spatial control of any text-to-image diffusion model with any condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7465-7475, 2024. 3, 7, 8", + "[18] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6038–6047, 2023. 4", + "[19] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296-4304, 2024. 2, 3, 7", + "[20] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195-4205, 2023. 2, 3", + "[21] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 6", + "[22] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 3", + "[23] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17846", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3", + "[24] Patsorn Sangkloy, Nathan Burnell, Cusuh Ham, and James Hays. The sketchy database: learning to retrieve badly drawn bunnies. ACM Transactions on Graphics (TOG), 35(4):1-12, 2016. 6", + "[25] Vishnu Sarukkai, Lu Yuan, Mia Tang, Maneesh Agrawala, and Kayvon Fatahalian. Block and detail: Scaffolding sketch-to-image generation. In Proceedings of the 37th Annual ACM Symposium on User Interface Software and Technology, pages 1-13, 2024. 2, 3", + "[26] Aravindan Sundaram, Ujjayan Pal, Abhimanyu Chauhan, Aishwarya Agarwal, and Srikrishna Karanam. Cocono: Attention contrast-and-complete for initial noise optimization in text-to-image synthesis. arXiv preprint arXiv:2411.16783, 2024. 3", + "[27] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 7", + "[28] Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. Sketch-guided text-to-image diffusion models. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2, 3", + "[29] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score: Better aligning text-to-image models with human preference. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2096-2105, 2023. 3, 6", + "[30] Zhenbei Wu, Haoge Deng, Qiang Wang, Di Kong, Jie Yang, and Yonggang Qi. Sketchscene: Scene sketch to image generation with diffusion models. In 2023 IEEE International Conference on Multimedia and Expo (ICME), pages 2087-2092. IEEE, 2023. 2", + "[31] Xin Xie and Dong Gong. Dymo: Training-free diffusion model alignment with dynamic multi-objective scheduling. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 13220-13230, 2025. 3", + "[32] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36:15903-15935, 2023. 3", + "[33] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3, 7", + "[34] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Scaling in-the-wild training for diffusion-based illumination harmonization and editing by imposing consistent light transport. In The Thirteenth International Conference on Learning Representations, 2025. 3", + "[35] Tianyu Zhang and Haoran Xie. Sketch-guided text-to-image generation with spatial control. In 2024 2nd International Conference on Computer Graphics and Image Processing (CGIP), pages 153-159. IEEE, 2024. 2, 3" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Tianyu Zhang, Xiaoxuan Xie, Xusheng Du, and Haoran Xie. Sketch-guided scene image generation. arXiv preprint arXiv:2407.06469, 2024. 2, 3", + "[37] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. Advances in Neural Information Processing Systems, 36:11127-11150, 2023. 2, 3, 7", + "[38] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22490-22499, 2023. 3" + ], + "bbox": [ + 516, + 90, + 903, + 273 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17847", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Subjective Camera 1.0_ Bridging Human Cognition and Visual Reconstruction through Sequence-Aware Sketch-Guided Diffusion/cae6ed8f-de57-4610-b479-1652f9260751_model.json b/2025/Subjective Camera 1.0_ Bridging Human Cognition and Visual Reconstruction through Sequence-Aware Sketch-Guided Diffusion/cae6ed8f-de57-4610-b479-1652f9260751_model.json new file mode 100644 index 0000000000000000000000000000000000000000..28695c6215c72f4853a5fefa088aeda875e2c5dc --- /dev/null +++ b/2025/Subjective Camera 1.0_ Bridging Human Cognition and Visual Reconstruction through Sequence-Aware Sketch-Guided Diffusion/cae6ed8f-de57-4610-b479-1652f9260751_model.json @@ -0,0 +1,2200 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "image", + "bbox": [ + 0.16, + 0.124, + 0.203, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.204, + 0.133, + 0.837, + 0.155 + ], + "angle": 0, + "content": "Subjective Camera 1.0: Bridging Human Cognition and Visual" + }, + { + "type": "title", + "bbox": [ + 0.165, + 0.157, + 0.834, + 0.178 + ], + "angle": 0, + "content": "Reconstruction through Sequence-Aware Sketch-Guided Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.204, + 0.869, + 0.242 + ], + "angle": 0, + "content": "Haoyang Chen\\(^{1,2,3}\\) Dongfang Sun\\(^{1,2}\\) Caoyuan Ma\\(^{1,2}\\) Shiqin Wang\\(^{1,2}\\) Kewei Zhang\\(^{1,2}\\) Zheng Wang\\(^{1,2,3}\\) Zhixiang Wang\\(^{4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.244, + 0.905, + 0.298 + ], + "angle": 0, + "content": "\\(^{1}\\)National Engineering Research Center for Multimedia Software, Institute of Artificial Intelligence, School of Computer Science, Wuhan University \\(^{2}\\)Hubei Key Laboratory of Multimedia and Network Communication Engineering \\(^{3}\\)Zhongguancun Academy, Beijing, China. 100094 \\(^{4}\\)CyberAgent AI Lab, Japan" + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.334, + 0.141, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.146, + 0.338, + 0.27, + 0.352 + ], + "angle": 0, + "content": "Physical Camera" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.356, + 0.261, + 0.368 + ], + "angle": 0, + "content": "Photo after Processing" + }, + { + "type": "image", + "bbox": [ + 0.128, + 0.372, + 0.253, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.485, + 0.256, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.417, + 0.333, + 0.447, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.453, + 0.338, + 0.706, + 0.352 + ], + "angle": 0, + "content": "The Concept of Subjective Camera" + }, + { + "type": "text", + "bbox": [ + 0.352, + 0.356, + 0.543, + 0.368 + ], + "angle": 0, + "content": "Readout (Text and Sketch Sequences)" + }, + { + "type": "text", + "bbox": [ + 0.295, + 0.381, + 0.421, + 0.397 + ], + "angle": 0, + "content": "There is a black car" + }, + { + "type": "text", + "bbox": [ + 0.294, + 0.401, + 0.422, + 0.453 + ], + "angle": 0, + "content": "towing a white trailer on the road near mountains." + }, + { + "type": "text", + "bbox": [ + 0.292, + 0.502, + 0.418, + 0.555 + ], + "angle": 0, + "content": "There is a motorbike and a tent on the grass." + }, + { + "type": "image", + "bbox": [ + 0.452, + 0.376, + 0.564, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.466, + 0.593, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.356, + 0.88, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.481, + 0.572, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.594, + 0.481, + 0.88, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.595, + 0.908, + 0.639 + ], + "angle": 0, + "content": "Figure 1. What is a subjective camera? A subjective camera transforms a person's memory into photorealistic images. We take a small yet crucial step toward this vision by leveraging generative models to sequentially decode textual and sketch-based readouts in their natural order, paving the way for reconstructing moments without physical cameras." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.673, + 0.327, + 0.689 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.708, + 0.486, + 0.859 + ], + "angle": 0, + "content": "We introduce the concept of a subjective camera to reconstruct meaningful moments that physical cameras fail to capture. We propose Subjective Camera 1.0, a framework for reconstructing real-world scenes from readily accessible subjective readouts, i.e., textual descriptions and progressively drawn rough sketches. Built on optimization-based alignment of diffusion models, our approach avoids large-scale paired training data and mitigates generalization issues. To address the challenge of integrating multiple abstract concepts in real-world scenar" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.909, + 0.812 + ], + "angle": 0, + "content": "ios, we design a Sequence-Aware Sketch-Guided Diffusion framework with three loss terms for concept-wise sequential optimization, following the natural order of subjective readouts. Experiments on two datasets demonstrate that our method achieves state-of-the-art performance in image quality as well as spatial and semantic alignment with target scenes. User studies with 40 participants further confirm that our approach is consistently preferred. Our project page is at: subjective-camera.github.io" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.844, + 0.646, + 0.86 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Cameras have long been the most popular devices for preserving memorable moments. However, a physical camera" + }, + { + "type": "page_footnote", + "bbox": [ + 0.103, + 0.874, + 0.365, + 0.888 + ], + "angle": 0, + "content": "These authors contributed equally to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.104, + 0.888, + 0.242, + 0.9 + ], + "angle": 0, + "content": "Corresponding authors." + }, + { + "type": "list", + "bbox": [ + 0.103, + 0.874, + 0.365, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17838" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "cannot always be present to capture every moment. This raises a question of how such unrecorded experiences can be preserved as pixels in the absence of a camera." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.138, + 0.483, + 0.349 + ], + "angle": 0, + "content": "This paper proposes the concept of a subjective camera1, where humans function as \"imaging devices\", aiming to faithfully reconstruct real-world scenes from memory. Specifically, we define a subjective camera as a system through which individuals encode sensory inputs into memory based on personal salience and emotional context, and later read out these stored representations and decode them into pixels that reconstruct scenes they experienced. We draw an analogy to physical cameras: just as a camera records, reads out, and reconstructs the visual world through optical, electronic, and computational processes, the human mind selectively captures (\"records\") perceptual details, which are organized into a retrievable format (\"read out\") and later reconstructs them as pixels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.35, + 0.483, + 0.546 + ], + "angle": 0, + "content": "It is generally challenging for humans without professional training to directly reconstruct a dense, pixel-wise image. Instead, the most practical approach is to read out the mental scene through textual descriptions and/or free-hand sketches and leverage computational tools to decode these readouts. However, text alone often fails to fully convey a mental image, particularly in terms of object layout and fine details. Freehand sketches provide complementary information beyond text but remain sparse and often contain noise and uncertainty [11, 15]. With the advancement of generative models [20, 23], this long-standing vision is becoming increasingly feasible by decoding such subjective readouts into photorealistic imagery." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.547, + 0.483, + 0.743 + ], + "angle": 0, + "content": "Although not intended to reconstruct real-world scenes, several works have explored generating images from combined text and sketch inputs [4, 25, 30, 35-37]. These approaches either adapt model weights using large-scale paired data [19, 25, 28, 30, 33, 36, 37] or modify the latent variables through per-scene optimization [4, 28, 35]. Training-based methods directly learn the mapping from sketches to pixels. However, they require extensive paired data and struggle with user-specific biases, often generalizing poorly to abstract sketches beyond the training distribution. In contrast, training-free methods avoid the heavy data requirements, high computational costs, and generalization challenges of training-based approaches." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.743, + 0.483, + 0.864 + ], + "angle": 0, + "content": "However, optimization-based methods often fail to reconstruct multiple concepts from idiosyncratic sketches, as would be required for recreating real-world scenes. This often leads to missing concepts or misaligned spatial relationships (see Figure 2a). The main obstacle lies in simultaneously interpreting multi-concept sketches, which inherently combine varying levels of abstraction and randomness across different concepts. For instance, freehand sketches" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.09, + 0.865, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.602, + 0.163, + 0.824, + 0.175 + ], + "angle": 0, + "content": "a: Limitation of \"snapshot\" generation" + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.182, + 0.862, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.591, + 0.257, + 0.829, + 0.269 + ], + "angle": 0, + "content": "b: Inspiration from human readout order" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.282, + 0.907, + 0.379 + ], + "angle": 0, + "content": "Figure 2. Motivation for concept-wise sequential optimization. (a) Attempting to \"snap\" all concepts into a complete image at once using generative diffusion priors often leads to suboptimal results, such as distortions and misaligned semantics or layouts. (b) In contrast, humans reconstruct mental scenes step by step. This natural readout inspires our design2, making it better suited to leverage current generative diffusion priors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.408, + 0.905, + 0.514 + ], + "angle": 0, + "content": "may only roughly indicate object boundaries, omit fine details such as texture and geometry, or distort proportions in non-standard ways. While generative priors can fill sparse regions and mitigate randomness during the transformation from sketches to pixels, attempting to integrate all concepts at once often reduces the problem to a least-squares solution [1, 4], ultimately blurring their distinct characteristics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.515, + 0.906, + 0.636 + ], + "angle": 0, + "content": "Our key idea is simple — to apply optimizations to each concept individually and do so sequentially. This design considers not only the limitations of the \"snapshot\" generation process but also human cognition. Humans tend to describe their mental images step-by-step [13], typically starting with a textual description of abstract concepts or the overall scene structure, and then progressively refining individual elements with increasing detail (see Figure 2b)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.638, + 0.906, + 0.834 + ], + "angle": 0, + "content": "Building upon this idea, we introduce a sequence-aware, sketch-guided diffusion framework that sequentially optimizes the latent noise of a pre-trained text-to-image (T2I) diffusion model [23] to align individual subjective readouts. The order of these optimizations mirrors the natural process of reading out subjective impressions: it begins with a text-reward optimization (Section 3.2) to align the generated image with the initial textual description, followed by concept-wise sequential optimizations (Section 3.5) to incorporate progressively provided sketches that may encode shapes, poses, and spatial arrangements. The sequential optimization also incorporates a loss to ensure that previously optimized concepts are not significantly altered." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.835, + 0.905, + 0.851 + ], + "angle": 0, + "content": "Two additional components are proposed to improve" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1 While the term \"subjective camera\" appears in cinematic contexts, our usage here differs significantly in meaning." + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.863, + 0.905, + 0.901 + ], + "angle": 0, + "content": "2This design philosophy is reminiscent of the progressive decoding used in rolling-shutter image sensors, which arises from the need to accommodate the limited processing capacity of the circuits." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17839" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.273 + ], + "angle": 0, + "content": "the concept-wise sequential optimization. First, each input sketch is individually encoded into the latent space to provide spatial guidance. To bridge the gap between rough sketch shapes and their corresponding real-world forms, we introduce an optimization-based inversion strategy for refining this spatial guidance (Section 3.3). Second, because input sketches often lack appearance details, directly optimizing with their guidance over multiple iterations may lead to unnatural textures and degraded visual fidelity. We therefore leverage the latent representation obtained from text-reward optimization to define an appearance loss (Section 3.4), thereby preserving visual style quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.275, + 0.482, + 0.486 + ], + "angle": 0, + "content": "By processing sketches individually and sequentially, our approach naturally aligns with the human cognition process, enabling users to incrementally construct mental images while ensuring each addition respects and complements previous elements. This sequence-aware generation strategy provides fine-grained control over complex multi-concept scenes while preserving the user's subjective intent encoded in the drawing order. Evaluations on two datasets demonstrate that our method achieves state-of-the-art performance in recreating a real-world scene from human memory, demonstrating the potential of subjective cameras. Notably, our framework is entirely training-free, eliminating the need for large-scale paired datasets and avoiding generalization." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.488, + 0.386, + 0.502 + ], + "angle": 0, + "content": "To sum up, our contributions are twofold:" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.504, + 0.48, + 0.564 + ], + "angle": 0, + "content": "We propose the concept of a subjective camera, which functions the human as an \"imaging device\" and leverages computational tools to transform cognitive impressions into photographs." + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.565, + 0.48, + 0.654 + ], + "angle": 0, + "content": "We propose a sequence-aware diffusion-guided generation framework that enables faithful reconstruction of complex, multi-concept scenes. This framework aligns with human cognition and surpasses \"snapshot\" generation methods, thereby establishing a new paradigm for cognition-driven image generation." + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.504, + 0.48, + 0.654 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.674, + 0.233, + 0.69 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Aligning Image Synthesis Models Despite their success, T2I generative models [20, 22, 23] often fail to reproduce the fine-grained semantics and compositional details described in complex prompts. Reward-based alignment has emerged as a promising direction [6, 7, 12, 14, 26, 31, 32], using human preference models [12, 29, 32] to guide generation. The human preference models were trained on paired human preference data. Early works [12, 32] fine-tune diffusion models with the reward models, whereas later methods [6, 14] sidestep costly fine-tuning by directly optimizing the latent noise [6, 31]. However, these techniques largely improve global prompt adherence while neglecting spatial control. Our model not only aligns the T2I genera" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "tive models with the text prompt, but also ensures the spatial information to respect to user-provided sketches." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.126, + 0.905, + 0.473 + ], + "angle": 0, + "content": "Controllable Image Synthesis T2I models inherently rely on highly compressed textual tokens, which limits their controllability and often prevents them from meeting user expectations [20, 22, 23]. To overcome these limitations, a growing body of research has explored controllable generation by incorporating fine-grained conditioning signals, such as reference images [2], edge maps or contours or skeletons [33], semantic layouts [1, 3, 38], and lighting specifications [34]. One line of work fine-tunes pretrained diffusion models by adding trainable modules [19, 33, 37], while others directly retrain the diffusion backbone by minimizing reconstruction objectives [2]. In parallel, training-free techniques [5, 7, 9, 26] have been developed to enhance controllability without modifying model weights. These include attention injection or latent optimization strategies, such as Prompt-to-Prompt editing [9] and diffusion self-guidance [5]. Despite these advances, existing methods often fail in complex scenes involving multiple guidance given by freehand sketches, where interactions between elements lead to interference and degraded synthesis quality. Building on these insights, our method incorporates attention-based guidance extraction but extends it to better handle multi-concept freehand sketches." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.905, + 0.78 + ], + "angle": 0, + "content": "Sketch-to-Image Synthesis It is the task most relevant to our concept of the subjective camera. It aims to generate images from freehand sketches along with textual prompts. Several recent efforts have attempted to adapt T2I diffusion models for sketch-to-image synthesis [4, 17, 25, 28, 35, 36]. Training-based approaches [25, 36] fine-tune the T2I models with paired data. However, this training-based method incurs substantial computational cost and suffers from poor generalization. In contrast, training-free approaches [4, 17, 35] avoid fine-tuning by guiding the diffusion denoising process with sketches. However, these methods still jointly create multiple concepts, causing interference between their differing levels of randomness and abstraction. More critically, both training-based and training-free approaches fail to address the subjective biases inherent in user-provided sketches and are particularly vulnerable to highly abstract inputs, often leading to appearance distortions and semantic inconsistencies in the generated outputs. Furthermore, unlike creative generation methods, our focus is on reconstructing real-world scenes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.793, + 0.887, + 0.81 + ], + "angle": 0, + "content": "3. Sequence-Aware Sketch-Guided Diffusion" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.625, + 0.833 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Our goal is to render photorealistic images from subjective readouts, consisting of a textual description \\( p \\) and a sequence \\( S = \\{s_i\\}_{i=1}^N \\) of \\( N \\) freehand sketches, where each sketch represents a distinct concept. As illustrated" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17840" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.092, + 0.904, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.279, + 0.908, + 0.351 + ], + "angle": 0, + "content": "Figure 3. Sequence-aware sketch-guided diffusion. (Left) User-provided textual description undergoes text-to-image generation and text-reward optimization to obtain the initial latent. Then, the following generation progresses under three constraints — spatial layout conforming to sketch topology while appearance details adhere to the initial image from text-reward optimization, with consistency maintained through ordered latent propagation and three loss terms, ensuring coherent integration of emerging and established scene elements. (Right) The concept-wise sequential optimization." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.376, + 0.485, + 0.662 + ], + "angle": 0, + "content": "in Figure 3, we design a Sequence-Aware Sketch-Guided Diffusion framework to accomplish this challenging task. Our method processes the subjective readouts individually and sequentially, following the order in which users describe them. It employs a sequence-aware optimization strategy to align a pre-trained generative diffusion model, \\(\\mathcal{G} := \\epsilon_{\\theta}(\\epsilon_{\\theta}(\\dots \\epsilon_{\\theta}(\\epsilon ,p)))\\), and \\(\\epsilon_{\\theta}\\) denotes the pre-trained U-Net, with these readouts, naturally mirroring the human cognitive process of step-by-step scene construction. The process begins with a text-reward optimization, which establishes an initial latent representation and appearance prior for subsequent concept-wise optimizations. The appearance prior is extracted via style encoding, while sketch encoding, implemented using diffusion inversion, provides spatial guidance for individual sketches. In addition, a preserving loss is introduced to maintain previously optimized concepts throughout the sequential process. These components collectively enable our framework to decode subjective readouts into coherent, photorealistic scenes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.671, + 0.336, + 0.687 + ], + "angle": 0, + "content": "3.2. Text-Reward Optimization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.693, + 0.484, + 0.83 + ], + "angle": 0, + "content": "We randomly sample a latent noise \\(\\epsilon\\) and feed it, together with the textual description \\(p\\), into a pre-trained T2I diffusion model, which renders an image through the full reverse diffusion process, \\(\\mathcal{G}_{\\theta}\\). To ensure that the generated image contains the concepts in the given textual information, without concept omission or significant misalignment, we formulate a text-reward optimization procedure as existing practice [6, 14] that iteratively refines the latent noise to maximize a text-image alignment objective:" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.838, + 0.483, + 0.862 + ], + "angle": 0, + "content": "\\[\n\\epsilon^ {0 \\star} = \\underset {\\epsilon} {\\arg \\max } \\mathcal {C} _ {F} \\left(\\mathcal {G} _ {\\theta} (\\epsilon , p), p\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{C}_F\\) represents a differentiable reward model that measures alignment between the generated image and the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.376, + 0.907, + 0.406 + ], + "angle": 0, + "content": "text prompt \\( p \\). The optimization process iteratively refines \\( \\epsilon \\) through gradient ascent:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.418, + 0.907, + 0.435 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {t + 1} = \\epsilon_ {t} + \\eta \\nabla_ {\\epsilon} \\mathcal {C} _ {F} (\\mathcal {G} _ {\\theta} (\\epsilon_ {t}, p)), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.446, + 0.702, + 0.461 + ], + "angle": 0, + "content": "where \\(\\eta\\) is the learning rate." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.47, + 0.681, + 0.487 + ], + "angle": 0, + "content": "3.3. Sketch Encoding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.907, + 0.703 + ], + "angle": 0, + "content": "Before proceeding to the sketch-guided optimization, we first encode the sketch to extract high-level semantic guidance, rather than relying on low-level line details. Given a list of input sketches \\(\\{s_i\\}_{i=1}^N\\), each sketch \\(s_i\\) is encoded into a latent representation \\(z_i \\in \\mathbb{R}^{h \\times w \\times d}\\) using the variational autoencoder (VAE) encoder \\(E_{\\phi}\\) of the diffusion model: \\(z_i = E_{\\phi}(s_i)\\). Then, we can leverage arbitrary diffusion inversion algorithms, such as DDIM inversion [18] or SDEdit [16], to map the sketch into latent noise. To address the significant domain gap between abstract planar sketches and realistic images, we introduce an optimization procedure that iteratively refines the latent representation \\(z_i\\) by aligning it with the rich prior knowledge embedded in the diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.75 + ], + "angle": 0, + "content": "At each optimization step \\(k\\), we perturb the latent sketch representation with a Gaussian noise \\(\\epsilon_{k} \\sim \\mathcal{N}(0, \\mathbf{I})\\) to obtain a noisy latent representation" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.759, + 0.907, + 0.781 + ], + "angle": 0, + "content": "\\[\nz _ {t} ^ {(k)} = \\alpha_ {t _ {k}} z _ {i} ^ {(k)} + \\sigma_ {t _ {k}} \\epsilon_ {k}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.791, + 0.906, + 0.872 + ], + "angle": 0, + "content": "where \\(\\alpha_{t_k}\\) and \\(\\sigma_{t_k}\\) are time-dependent scaling factors following the noise schedule of the diffusion process. The noisy latent \\(z_{t}^{(k)}\\) is then fed into the pre-trained U-Net \\(\\epsilon_{\\theta}\\) which predicts the noise component \\(\\hat{\\epsilon}_{\\theta}(z_{t}^{(k)}, t_{k})\\). The noise prediction error \\(\\Delta \\epsilon_{k}\\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.883, + 0.907, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\Delta \\epsilon_ {k} = \\hat {\\epsilon} _ {\\theta} \\left(z _ {t} ^ {(k)}, t _ {k}\\right) - \\epsilon_ {k}. \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17841" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.089, + 0.46, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.187, + 0.485, + 0.274 + ], + "angle": 0, + "content": "Figure 4. The effectiveness of sketch optimization. Without optimization (left), the generated image fails to accurately respect the wheel placement and proportions indicated in the sketch. With our optimization (right), the generated result better aligns with the sketch structure, particularly in the wheel positions (green arrows) and overall shape, demonstrating improved spatial coherence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.295, + 0.484, + 0.342 + ], + "angle": 0, + "content": "This noise gap serves as a gradient signal to iteratively refine the latent sketch features. Specifically, we update the latent representation \\( z_{i}^{(k)} \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.348, + 0.483, + 0.375 + ], + "angle": 0, + "content": "\\[\nz _ {i} ^ {(k + 1)} = z _ {i} ^ {(k)} - \\eta_ {k} \\left(\\Delta \\epsilon_ {k} + \\lambda \\left(z _ {i} ^ {(k)} - z _ {i} ^ {(0)}\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.483, + 0.412 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a regularization term that anchors the optimization to the initial latent \\(z_{i}^{(0)}\\), preventing semantic drift." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.412, + 0.484, + 0.517 + ], + "angle": 0, + "content": "Through this iterative optimization in Eq. (3-5), the latent \\(\\dot{z}_i\\) is pushed to align with the rich prior knowledge embedded in the diffusion model. This alignment effectively infuses the sketch with realistic details, textures, and structural coherence, thereby establishing a robust mapping from human hand-drawn sketches to physically grounded visual representations (see Figure 4)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.484, + 0.624 + ], + "angle": 0, + "content": "However, we observed that the optimization method is less effective when the sketch is drawn relatively small. In such cases, the generated result often fails to align with the sketch and instead produces a shape-independent object in a central position. To address this, we suggest scaling the original sketch using linear interpolation or, alternatively, discarding the optimization for that step." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.631, + 0.3, + 0.648 + ], + "angle": 0, + "content": "3.4. Appearance Encoding" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.652, + 0.484, + 0.715 + ], + "angle": 0, + "content": "While the sketch provides spatial information, additional appearance cues are required to ensure visual quality during optimization. We first define a operator \\(\\mathcal{A}(\\cdot)\\) to obtain the appearance descriptor:" + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.719, + 0.483, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} (x) = \\left\\{\\frac {\\sum_ {u , v} \\operatorname {s i g m o i d} ([ x ] _ {u v}) [ f _ {x} ] _ {u v}}{\\sum_ {u , v} \\operatorname {s i g m o i d} ([ x ] _ {u v})} \\mid f _ {x} \\right\\}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.857 + ], + "angle": 0, + "content": "which is a collection of weighted spatial means of diffusion features \\( f_{x} \\) and \\( f_{x} \\) are from different layer and time steps. Then, we treat the descriptor extracted from the optimized latent \\( z^{0\\star} \\), derived from \\( \\epsilon^{0\\star} \\) in Eq. (1), as our appearance prior. We will compare the appearance descriptor of the current latent and the appearance prior during optimization." + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.484, + 0.902 + ], + "angle": 0, + "content": "3To avoid ambiguity, we use \\(\\dot{z}_i\\) to denote the optimized latent for the \\(i\\)-th sketch concept, distinguishing it from latent \\(z_i\\) would be used in the concept-wise sequential optimization." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.845, + 0.108 + ], + "angle": 0, + "content": "3.5. Concept-wise Sequential Optimization" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.113, + 0.906, + 0.173 + ], + "angle": 0, + "content": "Next, we leverage this order information of subjective readouts to progressively construct the scene while maintaining coherence between newly introduced and previously established elements." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.177, + 0.906, + 0.253 + ], + "angle": 0, + "content": "Spatial Attention-guided Control We leverage the spatial attention mechanisms within the diffusion model's U-Net architecture. For each sketch concept \\( s_i \\), we extract cross-attention maps that highlight regions in the latent space corresponding to specific semantic concepts:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.259, + 0.907, + 0.295 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} (x) = \\operatorname {s o f t m a x} \\left(\\frac {Q (x) K (x) ^ {T}}{\\sqrt {d}}\\right) > \\tau , \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.301, + 0.906, + 0.392 + ], + "angle": 0, + "content": "where \\( Q(x) \\) and \\( K(x) \\) are query and key matrices, \\( d \\) is the dimensionality of the key vectors, and \\( \\tau \\) is a threshold parameter. The attention maps \\( \\mathcal{M}(\\dot{z}_i) \\) extracted from \\( \\dot{z}_i \\), the latent of the sketch \\( s_i \\) (Section 3.3), serve as spatial indicators to guide the concept's layout. The control is given through the spatial loss, defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.399, + 0.906, + 0.439 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s p a t i a l}} \\left(z _ {i}, \\dot {z} _ {i}\\right) = \\frac {\\sum_ {u , v} \\left[ \\mathcal {M} \\left(\\dot {z} _ {i}\\right) \\right] _ {u v} \\left\\| \\left[ \\dot {z} _ {i} \\right] _ {u v} - \\left[ z _ {i} \\right] _ {u v} \\right\\| _ {2} ^ {2}}{\\sum_ {u , v} \\left[ \\mathcal {M} \\left(\\dot {z} _ {i}\\right) \\right] _ {u v}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.445, + 0.779, + 0.461 + ], + "angle": 0, + "content": "where \\( u, v \\) are spatial positions indices." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.464, + 0.887, + 0.48 + ], + "angle": 0, + "content": "Appearance Control We define the appearance loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.486, + 0.906, + 0.509 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {a p p}} \\left(z _ {i}, z ^ {0 \\star}\\right) = \\left\\| \\mathcal {A} \\left(z _ {i}\\right) - \\mathcal {A} \\left(z ^ {0 \\star}\\right) \\right\\| _ {2} ^ {2}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.515, + 0.906, + 0.531 + ], + "angle": 0, + "content": "where \\(\\mathcal{A}\\) is the operator of exacting appearance descriptors." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.533, + 0.906, + 0.609 + ], + "angle": 0, + "content": "Sequential Concept Integration To maintain coherence across the progressive addition of concepts, we implement a sequential integration process that updates the latent noise representation while preserving previously incorporated elements. For each sketch \\( s_i \\) in the sequence, we compute:" + }, + { + "type": "equation", + "bbox": [ + 0.649, + 0.618, + 0.905, + 0.634 + ], + "angle": 0, + "content": "\\[\nz _ {i} = z _ {i - 1} + \\Delta z _ {i}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.642, + 0.906, + 0.702 + ], + "angle": 0, + "content": "where \\( z_{i-1} \\) is the optimized latent from the previous concept, and \\( \\Delta z_i \\) represents the incremental update required to incorporate the current sketch concept \\( s_i \\). This update is derived from our loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.71, + 0.905, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i} = \\alpha_ {i} \\mathcal {L} _ {\\text {n e w}} \\left(z _ {i}, \\dot {z} _ {i}, z ^ {0 \\star}\\right) + \\beta_ {i} \\mathcal {L} _ {\\text {p r e s e r v e}} \\left(z _ {i}, \\dot {z} _ {0, \\dots , i - 1}\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.735, + 0.906, + 0.766 + ], + "angle": 0, + "content": "where \\(\\alpha_{i}\\) and \\(\\beta_{i}\\) are balancing coefficients. \\(\\mathcal{L}_{\\mathrm{new}}\\) ensures proper integration of the current \\(z_{i}\\) and is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.773, + 0.905, + 0.792 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {n e w}} \\left(z _ {i}, \\dot {z} _ {i}, z ^ {0 \\star}\\right) = \\mathcal {L} _ {\\text {s p a t i a l}} \\left(z _ {i}, \\dot {z} _ {i}\\right) + \\mathcal {L} _ {\\text {a p p}} \\left(z _ {i}, z ^ {0 \\star}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.799, + 0.905, + 0.83 + ], + "angle": 0, + "content": "while \\(\\mathcal{L}_{\\mathrm{preserve}}\\) maintains the integrity of previously optimized concepts, given by:" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.837, + 0.905, + 0.879 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p r e s e r v e}} \\left(z _ {i}, \\dot {z} _ {0, \\dots , i - 1}\\right) = \\sum_ {j = 1} ^ {i - 1} \\gamma^ {i - j} \\mathcal {L} _ {\\text {s p a t i a l}} \\left(z _ {i}, \\dot {z} _ {j}\\right), \\tag {13}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.887, + 0.879, + 0.901 + ], + "angle": 0, + "content": "4In fact, an inner-loop optimization is omitted here for simplicity." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17842" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.09, + 0.9, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.48, + 0.908, + 0.536 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison of existing sketch-to-image generation methods, and our method on FMC and CMC datasets. Existing training-based methods tend to overfit to sketch lines, often failing to reconstruct scenes realistically, while training-free methods frequently misinterpret or overlook key concepts. In comparison, our approach more faithfully represents real-world physical scenes. The reading and processing order of the sketches is indicated by colors as follows: \\(\\bullet\\) \\(\\bullet\\) \\(\\bullet\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.562, + 0.483, + 0.609 + ], + "angle": 0, + "content": "where \\( j \\) means preserving established concepts, and \\( \\gamma^{i - j} \\) is a decay factor that weights the importance of earlier concepts based on their step gap from the current step." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.623, + 0.325, + 0.64 + ], + "angle": 0, + "content": "4. Experiments and Results" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.649, + 0.267, + 0.665 + ], + "angle": 0, + "content": "4.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Datasets To rigorously evaluate our method, we have undertaken extensive efforts to develop two specialized datasets: the Composed Multi-Concept (CMC) dataset and the Freehand Multi-Concept (FMC) dataset. Both datasets are designed to comprehensively assess the capabilities of our framework in handling multi-concept scene understanding and generation. CMC dataset comprises 142 scene sketches, each containing 2-4 distinct concepts sampled from the 150 semantic classes in the Sketchy dataset [24]. Each is accompanied by annotated textual prompts with the help of multimodal large language models. The CMC dataset does not provide real-world images captured by physical cameras for reference. FMC dataset is constructed from freehand sketches drawn by volunteers, based on real-world images sourced from publicly available outdoor" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.562, + 0.907, + 0.624 + ], + "angle": 0, + "content": "scene datasets. It comprises 42 hand-drawn sketches, each containing 2-3 distinct concepts. Additionally, each sketch is paired with a corresponding real-world image as a reference for validation and evaluation purposes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Evaluation Metrics To conduct a comprehensive quantitative evaluation of the quality of generated images, four widely-used evaluation metrics were employed, including Frechet Inception Distance (FID) [10], CLIP-T Score (CLIP-T) [21], Human Preference Score (HPS) [29], and CLIP-I distance (CLIP-I) [21]. FID measures the feature distribution similarity between generated and real images. Lower FID values indicate better distribution matching \\((\\downarrow)\\). CLIP-T quantifies the image-text alignment degree within the CLIP-T embedding space. A higher CLIP-T score reflects greater semantic alignment between the generated image and text prompt \\((\\uparrow)\\). HPS measures the subjective quality of the generated images based on human evaluations. Higher HPS values suggest that the generated images are more favorable or preferred by human evaluators \\((\\uparrow)\\). CLIP-I measures the alignment between two images. A higher value suggests that the two images are more consistent in terms of their semantic content \\((\\uparrow)\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17843" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.132, + 0.084, + 0.862, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.253, + 0.907, + 0.338 + ], + "angle": 0, + "content": "Figure 6. Ablation study on FMC and CMC datasets. Column headers are defined as follows: T2I denotes the baseline text-to-image diffusion model; OP (Overall Perception) indicates the single-step guidance approach utilizing complete sketch as a holistic conditioning signal; CSO refers to our proposed Concept-wise Sequential Optimization (Section 3.5); SE represents the Sketch Encoding module (Section 3.3) that enhances structural plausibility by injecting diffusion priors into imperfect sketches. The qualitative comparison demonstrates that CSO achieves finer layout control than OP. While SE effectively bridges the domain gap between freehand sketches and physical reality through latent space regularization, ultimately elevating generation quality." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.359, + 0.479, + 0.47 + ], + "angle": 0, + "content": "
Metric MethodCMCFMC
FID ↓CLIP-T ↑HPS ↑FID ↓CLIP-T ↑HPS ↑CLIP-I ↑
Training-based methods
ControlNet [33]15.2861.710.79414.4566.800.8020.622
T2I-Adapter [19]18.1361.610.74515.3761.610.7450.689
Uni-ControlNet [37]19.6062.580.76118.5362.580.7610.712
Training-free methods
Multi-Sketch [4]13.0564.110.74011.5972.020.7750.690
PNP [27]17.0163.270.84016.2671.240.8520.593
FreeControl [17]14.5264.740.75711.1766.670.7590.775
Ours11.5666.640.85010.1874.250.8620.797
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.482, + 0.483, + 0.581 + ], + "angle": 0, + "content": "Table 1. Quantitative results on CMC and FMC datasets. Our method consistently surpasses all training-free methods in distribution matching, image-text alignment, subjective quality, and appearance details as measured by FID, CLIP-T, HPS, and CLIP-I. Our method achieves superior image-text alignment compared to training-based methods (\\( \\downarrow \\) indicates lower values are better, while \\( \\uparrow \\) is opposite)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.608, + 0.484, + 0.715 + ], + "angle": 0, + "content": "Implementation Details In the implementation of our proposed method, we set the learning rate \\(\\eta\\) in Eq. (2) and the threshold parameter \\(\\tau\\) in Eq. (7) is 5.0 and 0.3, respectively. Meanwhile, hyperparameters \\(\\alpha_{i}\\) and \\(\\beta_{i}\\) in Eq. (11) are set as 0.8 and 0.2, respectively. \\(\\gamma\\) in Eq. (13) is set as 0.9. Our method is performed on one NVIDIA GeForce RTX 4090, and input sketch images are \\(512 \\times 512\\) in resolution." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.726, + 0.231, + 0.743 + ], + "angle": 0, + "content": "4.2. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.749, + 0.483, + 0.869 + ], + "angle": 0, + "content": "We conducted a comprehensive quantitative and qualitative evaluation of the proposed method and competing methods on the CMC and FMC datasets. Our competing methods include the training-based approaches ControlNet [33], T2I-Adapter [19], and Uni-ControlNet [37], as well as the training-free approaches Multi-Sketch [4], PNP [27], and FreeControl [17]. The quantitative results are presented in Table 1 and qualitative results are shown in Figure 5." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.873, + 0.484, + 0.904 + ], + "angle": 0, + "content": "We observe that training-based methods [19, 33, 37] are highly susceptible to subjective biases in sketches and of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.907, + 0.53 + ], + "angle": 0, + "content": "ten fail to adhere to text prompts, producing distorted and unrealistic images with lower CLIP-T scores. Training-free methods [4, 8, 17] suffer from concept omissions, conspicuous artifacts, and inconsistencies both among concepts and between concepts and their backgrounds, resulting in poor image realism. These issues stem from mutual interference during the \"snapshot\" generation of multiple concepts. In contrast, our proposed method faithfully reconstructs real-world scenes and achieves the best performance across all evaluation metrics on both the CMC and FMC datasets, demonstrating its effectiveness." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.54, + 0.905, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.636, + 0.907, + 0.718 + ], + "angle": 0, + "content": "Figure 7. User preference experiment. We asked participants to answer four questions based on the reference images, subjective readouts, and generated images, which included one produced by our method and one by a competing method. Results show that users significantly preferred our outputs over those of competing methods in most cases." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.667, + 0.76 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.811 + ], + "angle": 0, + "content": "To validate the effectiveness of each module in our method, we conducted ablation experiments on the FMC and CMC datasets, with the visualization results shown in Figure 6." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.812, + 0.907, + 0.902 + ], + "angle": 0, + "content": "It is observed that traditional T2I diffusion models (T2I) generate images that are close to text prompts. While incorporating sketches as guidance further constrains the consistency of concept positioning, deviations in appearance and spatial alignment persist (T2I, OP). The sequential concept understanding facilitated by Concept-wise Sequential" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17844" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.482, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.198, + 0.483, + 0.268 + ], + "angle": 0, + "content": "Figure 8. Control experiments on user preferences. We controlled three factors to examine their influence: whether the rater drew the photo, the availability timing of the reference image (\\(\\infty\\) vs. 5s), and the types of questions asked. In general, our method is consistently better than the competing method, FreeControl." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.294, + 0.484, + 0.432 + ], + "angle": 0, + "content": "Optimization (CSO) in Section 3.5 enhances the alignment of concept positioning and appearance with input sketches (CSO). Simultaneously, the Sketch Encoding (SE) module in Section 3.3 aligns content latent variables with real-world objects, mitigating geometric misalignment (T2I, OP, SE). When integrating all modules, our method faithfully reconstructs real-world scenes and achieves optimal results (Ours), demonstrating the effectiveness of each module and underscoring the indispensable nature of their design." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.441, + 0.353, + 0.457 + ], + "angle": 0, + "content": "4.4. User Preference Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.463, + 0.483, + 0.75 + ], + "angle": 0, + "content": "We invited 12 participants (group A) to provide subjective readouts and an additional 28 participants (group B) to evaluate the reconstructed images. Participants in group A were first asked to select a deeply impressive scene from their photo gallery and then provide a textual description along with sketches within two minutes. Our method and competing approaches, including two baselines (text-reward optimization and text-reward optimization with sketch-guided optimization based on all concepts) as well as an optimal method, FreeControl [17], were used to generate images from these readouts. Participants in both groups were then shown the subjective readouts, the reference image, and the generated images for the same scene. They were asked to answer four single-choice questions in a single-blind manner: Q1: Which image is closer to the original image (both style and content)? Q2: Which image has higher visual quality (e.g., more realistic, more natural)? Q3: Which image better matches the structure of the sketch? Q4: Which image better matches the textual description?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Figure 7 presents the statistics of the collected results. Participants judged our reconstructions as being closer to the corresponding real-world scenes compared to competing methods. Similar conclusions were observed for the other three questions. However, there was an exception in the comparison between our method and text-reward optimization regarding image quality (Q2). We observed that text-reward optimization without additional control tends to hallucinate details that enhance perceived realism, leading some users to prefer it for Q2. This also explains why our" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.136 + ], + "angle": 0, + "content": "advantage over text-reward optimization in Q1 was not statistically significant and why some participants expressed hesitation in this case." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.137, + 0.907, + 0.38 + ], + "angle": 0, + "content": "We further conducted controlled experiments to examine the effects of three factors: whether the rater provided the subjective readouts, the availability timing of the reference image, and the types of questions asked. Results in Figure 8 show that our method consistently outperformed FreeControl [17], regardless of these variations. Notably: 1) Participants who provided subjective readouts showed slightly lower preference for our method in terms of similarity to the reference image compared to those who did not, while showing a higher preference in terms of alignment with the readouts and perceived image quality. We suppose this is mainly because of the subjectivity in reading out. 2) Reducing the time participants could view the reference image decreased the preference for our method; when allowed to compare freely, participants preferred our results more strongly." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.393, + 0.64, + 0.408 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.905, + 0.658 + ], + "angle": 0, + "content": "This paper presented Subjective Camera, a paradigm for reconstructing real-world scenes from humans' mental impressions. Our framework relies on the most accessible readouts of mental imagery, i.e., textual descriptions and sketches, and employs a training-free, sequence-aware optimization process. This approach eliminates the need for costly fine-tuning of text-to-image diffusion models on large-scale data while mitigating challenges such as dependence on sketch quality and geometric misalignment. Extensive experiments on both synthetic and real-world sketch datasets demonstrate that Subjective Camera achieves state-of-the-art performance, outperforming both training-free and training-based sketch-to-image generation methods across all evaluation metrics. Furthermore, user preference studies confirm that participants consistently favor our method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In closing, we address a natural question: \"With the rapid advancement of generative AI, will physical cameras ever become unnecessary\" For now, the answer is no. First, the reconstructed images still fall short of perfectly matching actual scenes. This limitation largely stems from the fact that current T2I models are trained on highly diverse data, creating gaps between the prior and the specific scene to be reconstructed. One promising direction to mitigate this is to personalize the prior models using user-specific galleries. Second, we observe that variations in the order and quality of sketches have a significant impact on reconstruction results. To address this, we plan to enhance our framework with computational tools that guide users in providing more precise and suitable subjective readouts. For these reasons, we designate this system as Subjective Camera 1.0, leaving these improvements for future iterations." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17845" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.199 + ], + "angle": 0, + "content": "Acknowledgements This work was supported in part by the National Natural Science Foundation of China (62225113, 624B2109, 623B2079) and the Zhongguancun Academy Project (20240308). We thank Po-Hsun Huang for providing the questionnaire system for our user study, Kaijin Zheng for assistance with data processing, and all participants who took part in the studies." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.211, + 0.188, + 0.227 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.236, + 0.482, + 0.291 + ], + "angle": 0, + "content": "[1] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. Proceedings of Machine Learning Research, 202:1737-1752, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.293, + 0.482, + 0.348 + ], + "angle": 0, + "content": "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18392-18402, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.351, + 0.484, + 0.405 + ], + "angle": 0, + "content": "[3] Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5343-5353, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.407, + 0.482, + 0.476 + ], + "angle": 0, + "content": "[4] Zhenwei Cheng, Lei Wu, Changshuo Wang, and Xiangxu Meng. Scene sketch-to-image synthesis based on multi-object control. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3775-3779. IEEE, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.478, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[5] Dave Epstein, Allan Jabri, Ben Poole, Alexei Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. Advances in Neural Information Processing Systems, 36:16222-16239, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.535, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[6] Luca Eyring, Shyamgopal Karthik, Karsten Roth, Alexey Dosovitskiy, and Zeynep Akata. Reno: Enhancing one-step text-to-image models through reward-based noise optimization. Advances in Neural Information Processing Systems, 37:125487-125519, 2025. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.605, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[7] Xiefan Guo, Jinlin Liu, Miaomiao Cui, Jiankai Li, Hongyu Yang, and Di Huang. Initno: Boosting text-to-image diffusion models via initial noise optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9380-9389, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.676, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[8] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In The Eleventh International Conference on Learning Representations. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.733, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[9] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control.(2022). URL https://arxiv.org/abs/2208.01626, 3, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.789, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[11] Zechao Hu, Zhengwei Yang, Hao Li, Yixiong Zou, Fengbin Zhu, and Zheng Wang. Unified category and style generalization for instance-level sketch retrieval. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.236, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, page 864-873. Association for Computing Machinery, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[12] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Martiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36: 36652-36663, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[13] Stephen M Kosslyn. Aspects of a cognitive neuroscience of mental imagery. Science, 240(4859):1621-1626, 1988. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.519, + 0.236, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[14] Seung Hyun Lee, Yinxiao Li, Junjie Ke, Innfarn Yoo, Han Zhang, Jiahui Yu, Qifei Wang, Fei Deng, Glenn Entis, Junfeng He, et al. Parrot: Pareto-optimal multi-reward reinforcement learning framework for text-to-image generation. In European Conference on Computer Vision, pages 462-478. Springer, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.907, + 0.389 + ], + "angle": 0, + "content": "[15] Kejun Lin, Zhixiang Wang, Zheng Wang, Yinqiang Zheng, and Shin'ichi Satoh. Beyond domain gap: Exploiting subjectivity in sketch-based person retrieval. In Proceedings of the 31st ACM international conference on multimedia, pages 2078-2089, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.391, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[16] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.449, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[17] Sicheng Mo, Fangzhou Mu, Kuan Heng Lin, Yanli Liu, Bochen Guan, Yin Li, and Bolei Zhou. Freecontrol: Training-free spatial control of any text-to-image diffusion model with any condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7465-7475, 2024. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.533, + 0.907, + 0.602 + ], + "angle": 0, + "content": "[18] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6038–6047, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[19] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296-4304, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.675, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[20] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195-4205, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.732, + 0.905, + 0.813 + ], + "angle": 0, + "content": "[21] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[22] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17846" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.188 + ], + "angle": 0, + "content": "[24] Patsorn Sangkloy, Nathan Burnell, Cusuh Ham, and James Hays. The sketchy database: learning to retrieve badly drawn bunnies. ACM Transactions on Graphics (TOG), 35(4):1-12, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[25] Vishnu Sarukkai, Lu Yuan, Mia Tang, Maneesh Agrawala, and Kayvon Fatahalian. Block and detail: Scaffolding sketch-to-image generation. In Proceedings of the 37th Annual ACM Symposium on User Interface Software and Technology, pages 1-13, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[26] Aravindan Sundaram, Ujjayan Pal, Abhimanyu Chauhan, Aishwarya Agarwal, and Srikrishna Karanam. Cocono: Attention contrast-and-complete for initial noise optimization in text-to-image synthesis. arXiv preprint arXiv:2411.16783, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[27] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[28] Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. Sketch-guided text-to-image diffusion models. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[29] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score: Better aligning text-to-image models with human preference. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2096-2105, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.593 + ], + "angle": 0, + "content": "[30] Zhenbei Wu, Haoge Deng, Qiang Wang, Di Kong, Jie Yang, and Yonggang Qi. Sketchscene: Scene sketch to image generation with diffusion models. In 2023 IEEE International Conference on Multimedia and Expo (ICME), pages 2087-2092. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.65 + ], + "angle": 0, + "content": "[31] Xin Xie and Dong Gong. Dymo: Training-free diffusion model alignment with dynamic multi-objective scheduling. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 13220-13230, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[32] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36:15903-15935, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.721, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[33] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[34] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Scaling in-the-wild training for diffusion-based illumination harmonization and editing by imposing consistent light transport. In The Thirteenth International Conference on Learning Representations, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[35] Tianyu Zhang and Haoran Xie. Sketch-guided text-to-image generation with spatial control. In 2024 2nd International Conference on Computer Graphics and Image Processing (CGIP), pages 153-159. IEEE, 2024. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[36] Tianyu Zhang, Xiaoxuan Xie, Xusheng Du, and Haoran Xie. Sketch-guided scene image generation. arXiv preprint arXiv:2407.06469, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[37] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. Advances in Neural Information Processing Systems, 36:11127-11150, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[38] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22490-22499, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17847" + } + ] +] \ No newline at end of file diff --git a/2025/Subjective Camera 1.0_ Bridging Human Cognition and Visual Reconstruction through Sequence-Aware Sketch-Guided Diffusion/cae6ed8f-de57-4610-b479-1652f9260751_origin.pdf b/2025/Subjective Camera 1.0_ Bridging Human Cognition and Visual Reconstruction through Sequence-Aware Sketch-Guided Diffusion/cae6ed8f-de57-4610-b479-1652f9260751_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..745115eae30d912fd002e506b5c35cba19865030 --- /dev/null +++ b/2025/Subjective Camera 1.0_ Bridging Human Cognition and Visual Reconstruction through Sequence-Aware Sketch-Guided Diffusion/cae6ed8f-de57-4610-b479-1652f9260751_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4955c74f7fa5fb2ae46aaa861d63ab26585bf4a355ddbad677d3a50d49208105 +size 4348418 diff --git a/2025/Subjective Camera 1.0_ Bridging Human Cognition and Visual Reconstruction through Sequence-Aware Sketch-Guided Diffusion/full.md b/2025/Subjective Camera 1.0_ Bridging Human Cognition and Visual Reconstruction through Sequence-Aware Sketch-Guided Diffusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cf2209f05525eb1b5288b218319951abace8ddba --- /dev/null +++ b/2025/Subjective Camera 1.0_ Bridging Human Cognition and Visual Reconstruction through Sequence-Aware Sketch-Guided Diffusion/full.md @@ -0,0 +1,326 @@ +![](images/036f49fca5b1146f0dd58d8ee5fc660c76717546e46a77730b006441659d3a92.jpg) + +# Subjective Camera 1.0: Bridging Human Cognition and Visual + +# Reconstruction through Sequence-Aware Sketch-Guided Diffusion + +Haoyang Chen $^{1,2,3}$ Dongfang Sun $^{1,2}$ Caoyuan Ma $^{1,2}$ Shiqin Wang $^{1,2}$ Kewei Zhang $^{1,2}$ Zheng Wang $^{1,2,3}$ Zhixiang Wang $^{4}$ + +$^{1}$ National Engineering Research Center for Multimedia Software, Institute of Artificial Intelligence, School of Computer Science, Wuhan University $^{2}$ Hubei Key Laboratory of Multimedia and Network Communication Engineering $^{3}$ Zhongguancun Academy, Beijing, China. 100094 $^{4}$ CyberAgent AI Lab, Japan + +![](images/f3172a01beebfc96259e00a3d02b0f8b0756a128e359a6431d8dd59d4d807e32.jpg) + +# Physical Camera + +Photo after Processing + +![](images/198dea298b5766e5660194fc5b560b65a5d8808d85811a9a3a5e5487d22e5e9a.jpg) + +![](images/2e44cc24f8fb45f8fe7385acfcd4dba41e09513702fd2e31e90b0373617f18b0.jpg) +Figure 1. What is a subjective camera? A subjective camera transforms a person's memory into photorealistic images. We take a small yet crucial step toward this vision by leveraging generative models to sequentially decode textual and sketch-based readouts in their natural order, paving the way for reconstructing moments without physical cameras. + +![](images/781beaaf605b41f5a82caaef17a296da083e4925311b0e88c7a89a5664ab39e5.jpg) + +# The Concept of Subjective Camera + +Readout (Text and Sketch Sequences) + +There is a black car + +towing a white trailer on the road near mountains. + +There is a motorbike and a tent on the grass. + +![](images/6e76ebb9c9b187a325fc221c175ddff7dec200ef986cabc99ef19e3f06bced32.jpg) + +![](images/da44a85605c00f461abc1335f1f9da5d685896f9f5a65a0d08839818d8f46c1c.jpg) + +![](images/46beac0003d7a4dfedbf29fd69a6b3bee621f52ed784ce88f46ee68655b6fedd.jpg) + +![](images/d870bcf8c119e317df29b4087a76c625145f021f658294d0322cec026198e1df.jpg) + +![](images/1ef267c53bf9908bd17d6b36f1ddc725503d068b70310c85e24991f33458b465.jpg) + +# Abstract + +We introduce the concept of a subjective camera to reconstruct meaningful moments that physical cameras fail to capture. We propose Subjective Camera 1.0, a framework for reconstructing real-world scenes from readily accessible subjective readouts, i.e., textual descriptions and progressively drawn rough sketches. Built on optimization-based alignment of diffusion models, our approach avoids large-scale paired training data and mitigates generalization issues. To address the challenge of integrating multiple abstract concepts in real-world scenar + +ios, we design a Sequence-Aware Sketch-Guided Diffusion framework with three loss terms for concept-wise sequential optimization, following the natural order of subjective readouts. Experiments on two datasets demonstrate that our method achieves state-of-the-art performance in image quality as well as spatial and semantic alignment with target scenes. User studies with 40 participants further confirm that our approach is consistently preferred. Our project page is at: subjective-camera.github.io + +# 1. Introduction + +Cameras have long been the most popular devices for preserving memorable moments. However, a physical camera + +cannot always be present to capture every moment. This raises a question of how such unrecorded experiences can be preserved as pixels in the absence of a camera. + +This paper proposes the concept of a subjective camera1, where humans function as "imaging devices", aiming to faithfully reconstruct real-world scenes from memory. Specifically, we define a subjective camera as a system through which individuals encode sensory inputs into memory based on personal salience and emotional context, and later read out these stored representations and decode them into pixels that reconstruct scenes they experienced. We draw an analogy to physical cameras: just as a camera records, reads out, and reconstructs the visual world through optical, electronic, and computational processes, the human mind selectively captures ("records") perceptual details, which are organized into a retrievable format ("read out") and later reconstructs them as pixels. + +It is generally challenging for humans without professional training to directly reconstruct a dense, pixel-wise image. Instead, the most practical approach is to read out the mental scene through textual descriptions and/or free-hand sketches and leverage computational tools to decode these readouts. However, text alone often fails to fully convey a mental image, particularly in terms of object layout and fine details. Freehand sketches provide complementary information beyond text but remain sparse and often contain noise and uncertainty [11, 15]. With the advancement of generative models [20, 23], this long-standing vision is becoming increasingly feasible by decoding such subjective readouts into photorealistic imagery. + +Although not intended to reconstruct real-world scenes, several works have explored generating images from combined text and sketch inputs [4, 25, 30, 35-37]. These approaches either adapt model weights using large-scale paired data [19, 25, 28, 30, 33, 36, 37] or modify the latent variables through per-scene optimization [4, 28, 35]. Training-based methods directly learn the mapping from sketches to pixels. However, they require extensive paired data and struggle with user-specific biases, often generalizing poorly to abstract sketches beyond the training distribution. In contrast, training-free methods avoid the heavy data requirements, high computational costs, and generalization challenges of training-based approaches. + +However, optimization-based methods often fail to reconstruct multiple concepts from idiosyncratic sketches, as would be required for recreating real-world scenes. This often leads to missing concepts or misaligned spatial relationships (see Figure 2a). The main obstacle lies in simultaneously interpreting multi-concept sketches, which inherently combine varying levels of abstraction and randomness across different concepts. For instance, freehand sketches + +![](images/a35ba4a51f5908149de79a0d33288e1b6f0e75abf38588ca6171cdc6f24acf02.jpg) +a: Limitation of "snapshot" generation + +![](images/2c41e85fa300491fce409300694f16caa528e5f6e7f20fb0cd20939030c53796.jpg) +b: Inspiration from human readout order +Figure 2. Motivation for concept-wise sequential optimization. (a) Attempting to "snap" all concepts into a complete image at once using generative diffusion priors often leads to suboptimal results, such as distortions and misaligned semantics or layouts. (b) In contrast, humans reconstruct mental scenes step by step. This natural readout inspires our design2, making it better suited to leverage current generative diffusion priors. + +may only roughly indicate object boundaries, omit fine details such as texture and geometry, or distort proportions in non-standard ways. While generative priors can fill sparse regions and mitigate randomness during the transformation from sketches to pixels, attempting to integrate all concepts at once often reduces the problem to a least-squares solution [1, 4], ultimately blurring their distinct characteristics. + +Our key idea is simple — to apply optimizations to each concept individually and do so sequentially. This design considers not only the limitations of the "snapshot" generation process but also human cognition. Humans tend to describe their mental images step-by-step [13], typically starting with a textual description of abstract concepts or the overall scene structure, and then progressively refining individual elements with increasing detail (see Figure 2b). + +Building upon this idea, we introduce a sequence-aware, sketch-guided diffusion framework that sequentially optimizes the latent noise of a pre-trained text-to-image (T2I) diffusion model [23] to align individual subjective readouts. The order of these optimizations mirrors the natural process of reading out subjective impressions: it begins with a text-reward optimization (Section 3.2) to align the generated image with the initial textual description, followed by concept-wise sequential optimizations (Section 3.5) to incorporate progressively provided sketches that may encode shapes, poses, and spatial arrangements. The sequential optimization also incorporates a loss to ensure that previously optimized concepts are not significantly altered. + +Two additional components are proposed to improve + +the concept-wise sequential optimization. First, each input sketch is individually encoded into the latent space to provide spatial guidance. To bridge the gap between rough sketch shapes and their corresponding real-world forms, we introduce an optimization-based inversion strategy for refining this spatial guidance (Section 3.3). Second, because input sketches often lack appearance details, directly optimizing with their guidance over multiple iterations may lead to unnatural textures and degraded visual fidelity. We therefore leverage the latent representation obtained from text-reward optimization to define an appearance loss (Section 3.4), thereby preserving visual style quality. + +By processing sketches individually and sequentially, our approach naturally aligns with the human cognition process, enabling users to incrementally construct mental images while ensuring each addition respects and complements previous elements. This sequence-aware generation strategy provides fine-grained control over complex multi-concept scenes while preserving the user's subjective intent encoded in the drawing order. Evaluations on two datasets demonstrate that our method achieves state-of-the-art performance in recreating a real-world scene from human memory, demonstrating the potential of subjective cameras. Notably, our framework is entirely training-free, eliminating the need for large-scale paired datasets and avoiding generalization. + +To sum up, our contributions are twofold: + +We propose the concept of a subjective camera, which functions the human as an "imaging device" and leverages computational tools to transform cognitive impressions into photographs. +We propose a sequence-aware diffusion-guided generation framework that enables faithful reconstruction of complex, multi-concept scenes. This framework aligns with human cognition and surpasses "snapshot" generation methods, thereby establishing a new paradigm for cognition-driven image generation. + +# 2. Related Work + +Aligning Image Synthesis Models Despite their success, T2I generative models [20, 22, 23] often fail to reproduce the fine-grained semantics and compositional details described in complex prompts. Reward-based alignment has emerged as a promising direction [6, 7, 12, 14, 26, 31, 32], using human preference models [12, 29, 32] to guide generation. The human preference models were trained on paired human preference data. Early works [12, 32] fine-tune diffusion models with the reward models, whereas later methods [6, 14] sidestep costly fine-tuning by directly optimizing the latent noise [6, 31]. However, these techniques largely improve global prompt adherence while neglecting spatial control. Our model not only aligns the T2I genera + +tive models with the text prompt, but also ensures the spatial information to respect to user-provided sketches. + +Controllable Image Synthesis T2I models inherently rely on highly compressed textual tokens, which limits their controllability and often prevents them from meeting user expectations [20, 22, 23]. To overcome these limitations, a growing body of research has explored controllable generation by incorporating fine-grained conditioning signals, such as reference images [2], edge maps or contours or skeletons [33], semantic layouts [1, 3, 38], and lighting specifications [34]. One line of work fine-tunes pretrained diffusion models by adding trainable modules [19, 33, 37], while others directly retrain the diffusion backbone by minimizing reconstruction objectives [2]. In parallel, training-free techniques [5, 7, 9, 26] have been developed to enhance controllability without modifying model weights. These include attention injection or latent optimization strategies, such as Prompt-to-Prompt editing [9] and diffusion self-guidance [5]. Despite these advances, existing methods often fail in complex scenes involving multiple guidance given by freehand sketches, where interactions between elements lead to interference and degraded synthesis quality. Building on these insights, our method incorporates attention-based guidance extraction but extends it to better handle multi-concept freehand sketches. + +Sketch-to-Image Synthesis It is the task most relevant to our concept of the subjective camera. It aims to generate images from freehand sketches along with textual prompts. Several recent efforts have attempted to adapt T2I diffusion models for sketch-to-image synthesis [4, 17, 25, 28, 35, 36]. Training-based approaches [25, 36] fine-tune the T2I models with paired data. However, this training-based method incurs substantial computational cost and suffers from poor generalization. In contrast, training-free approaches [4, 17, 35] avoid fine-tuning by guiding the diffusion denoising process with sketches. However, these methods still jointly create multiple concepts, causing interference between their differing levels of randomness and abstraction. More critically, both training-based and training-free approaches fail to address the subjective biases inherent in user-provided sketches and are particularly vulnerable to highly abstract inputs, often leading to appearance distortions and semantic inconsistencies in the generated outputs. Furthermore, unlike creative generation methods, our focus is on reconstructing real-world scenes. + +# 3. Sequence-Aware Sketch-Guided Diffusion + +# 3.1. Overview + +Our goal is to render photorealistic images from subjective readouts, consisting of a textual description $p$ and a sequence $S = \{s_i\}_{i=1}^N$ of $N$ freehand sketches, where each sketch represents a distinct concept. As illustrated + +![](images/e491c29db16566f6abc2ac861ee480133a7e2c9045a62d4a5b2605baff667363.jpg) +Figure 3. Sequence-aware sketch-guided diffusion. (Left) User-provided textual description undergoes text-to-image generation and text-reward optimization to obtain the initial latent. Then, the following generation progresses under three constraints — spatial layout conforming to sketch topology while appearance details adhere to the initial image from text-reward optimization, with consistency maintained through ordered latent propagation and three loss terms, ensuring coherent integration of emerging and established scene elements. (Right) The concept-wise sequential optimization. + +in Figure 3, we design a Sequence-Aware Sketch-Guided Diffusion framework to accomplish this challenging task. Our method processes the subjective readouts individually and sequentially, following the order in which users describe them. It employs a sequence-aware optimization strategy to align a pre-trained generative diffusion model, $\mathcal{G} := \epsilon_{\theta}(\epsilon_{\theta}(\dots \epsilon_{\theta}(\epsilon ,p)))$ , and $\epsilon_{\theta}$ denotes the pre-trained U-Net, with these readouts, naturally mirroring the human cognitive process of step-by-step scene construction. The process begins with a text-reward optimization, which establishes an initial latent representation and appearance prior for subsequent concept-wise optimizations. The appearance prior is extracted via style encoding, while sketch encoding, implemented using diffusion inversion, provides spatial guidance for individual sketches. In addition, a preserving loss is introduced to maintain previously optimized concepts throughout the sequential process. These components collectively enable our framework to decode subjective readouts into coherent, photorealistic scenes. + +# 3.2. Text-Reward Optimization + +We randomly sample a latent noise $\epsilon$ and feed it, together with the textual description $p$ , into a pre-trained T2I diffusion model, which renders an image through the full reverse diffusion process, $\mathcal{G}_{\theta}$ . To ensure that the generated image contains the concepts in the given textual information, without concept omission or significant misalignment, we formulate a text-reward optimization procedure as existing practice [6, 14] that iteratively refines the latent noise to maximize a text-image alignment objective: + +$$ +\epsilon^ {0 \star} = \underset {\epsilon} {\arg \max } \mathcal {C} _ {F} \left(\mathcal {G} _ {\theta} (\epsilon , p), p\right), \tag {1} +$$ + +where $\mathcal{C}_F$ represents a differentiable reward model that measures alignment between the generated image and the + +text prompt $p$ . The optimization process iteratively refines $\epsilon$ through gradient ascent: + +$$ +\epsilon_ {t + 1} = \epsilon_ {t} + \eta \nabla_ {\epsilon} \mathcal {C} _ {F} (\mathcal {G} _ {\theta} (\epsilon_ {t}, p)), \tag {2} +$$ + +where $\eta$ is the learning rate. + +# 3.3. Sketch Encoding + +Before proceeding to the sketch-guided optimization, we first encode the sketch to extract high-level semantic guidance, rather than relying on low-level line details. Given a list of input sketches $\{s_i\}_{i=1}^N$ , each sketch $s_i$ is encoded into a latent representation $z_i \in \mathbb{R}^{h \times w \times d}$ using the variational autoencoder (VAE) encoder $E_{\phi}$ of the diffusion model: $z_i = E_{\phi}(s_i)$ . Then, we can leverage arbitrary diffusion inversion algorithms, such as DDIM inversion [18] or SDEdit [16], to map the sketch into latent noise. To address the significant domain gap between abstract planar sketches and realistic images, we introduce an optimization procedure that iteratively refines the latent representation $z_i$ by aligning it with the rich prior knowledge embedded in the diffusion model. + +At each optimization step $k$ , we perturb the latent sketch representation with a Gaussian noise $\epsilon_{k} \sim \mathcal{N}(0, \mathbf{I})$ to obtain a noisy latent representation + +$$ +z _ {t} ^ {(k)} = \alpha_ {t _ {k}} z _ {i} ^ {(k)} + \sigma_ {t _ {k}} \epsilon_ {k}, \tag {3} +$$ + +where $\alpha_{t_k}$ and $\sigma_{t_k}$ are time-dependent scaling factors following the noise schedule of the diffusion process. The noisy latent $z_{t}^{(k)}$ is then fed into the pre-trained U-Net $\epsilon_{\theta}$ which predicts the noise component $\hat{\epsilon}_{\theta}(z_{t}^{(k)}, t_{k})$ . The noise prediction error $\Delta \epsilon_{k}$ is computed as: + +$$ +\Delta \epsilon_ {k} = \hat {\epsilon} _ {\theta} \left(z _ {t} ^ {(k)}, t _ {k}\right) - \epsilon_ {k}. \tag {4} +$$ + +![](images/ba965996a6b33632add424df6e05c58ac886742064647239aecf7781edf65720.jpg) +Figure 4. The effectiveness of sketch optimization. Without optimization (left), the generated image fails to accurately respect the wheel placement and proportions indicated in the sketch. With our optimization (right), the generated result better aligns with the sketch structure, particularly in the wheel positions (green arrows) and overall shape, demonstrating improved spatial coherence. + +This noise gap serves as a gradient signal to iteratively refine the latent sketch features. Specifically, we update the latent representation $z_{i}^{(k)}$ as + +$$ +z _ {i} ^ {(k + 1)} = z _ {i} ^ {(k)} - \eta_ {k} \left(\Delta \epsilon_ {k} + \lambda \left(z _ {i} ^ {(k)} - z _ {i} ^ {(0)}\right)\right), \tag {5} +$$ + +where $\lambda$ is a regularization term that anchors the optimization to the initial latent $z_{i}^{(0)}$ , preventing semantic drift. + +Through this iterative optimization in Eq. (3-5), the latent $\dot{z}_i$ is pushed to align with the rich prior knowledge embedded in the diffusion model. This alignment effectively infuses the sketch with realistic details, textures, and structural coherence, thereby establishing a robust mapping from human hand-drawn sketches to physically grounded visual representations (see Figure 4). + +However, we observed that the optimization method is less effective when the sketch is drawn relatively small. In such cases, the generated result often fails to align with the sketch and instead produces a shape-independent object in a central position. To address this, we suggest scaling the original sketch using linear interpolation or, alternatively, discarding the optimization for that step. + +# 3.4. Appearance Encoding + +While the sketch provides spatial information, additional appearance cues are required to ensure visual quality during optimization. We first define a operator $\mathcal{A}(\cdot)$ to obtain the appearance descriptor: + +$$ +\mathcal {A} (x) = \left\{\frac {\sum_ {u , v} \operatorname {s i g m o i d} ([ x ] _ {u v}) [ f _ {x} ] _ {u v}}{\sum_ {u , v} \operatorname {s i g m o i d} ([ x ] _ {u v})} \mid f _ {x} \right\}, \tag {6} +$$ + +which is a collection of weighted spatial means of diffusion features $f_{x}$ and $f_{x}$ are from different layer and time steps. Then, we treat the descriptor extracted from the optimized latent $z^{0\star}$ , derived from $\epsilon^{0\star}$ in Eq. (1), as our appearance prior. We will compare the appearance descriptor of the current latent and the appearance prior during optimization. + +# 3.5. Concept-wise Sequential Optimization + +Next, we leverage this order information of subjective readouts to progressively construct the scene while maintaining coherence between newly introduced and previously established elements. + +Spatial Attention-guided Control We leverage the spatial attention mechanisms within the diffusion model's U-Net architecture. For each sketch concept $s_i$ , we extract cross-attention maps that highlight regions in the latent space corresponding to specific semantic concepts: + +$$ +\mathcal {M} (x) = \operatorname {s o f t m a x} \left(\frac {Q (x) K (x) ^ {T}}{\sqrt {d}}\right) > \tau , \tag {7} +$$ + +where $Q(x)$ and $K(x)$ are query and key matrices, $d$ is the dimensionality of the key vectors, and $\tau$ is a threshold parameter. The attention maps $\mathcal{M}(\dot{z}_i)$ extracted from $\dot{z}_i$ , the latent of the sketch $s_i$ (Section 3.3), serve as spatial indicators to guide the concept's layout. The control is given through the spatial loss, defined as: + +$$ +\mathcal {L} _ {\text {s p a t i a l}} \left(z _ {i}, \dot {z} _ {i}\right) = \frac {\sum_ {u , v} \left[ \mathcal {M} \left(\dot {z} _ {i}\right) \right] _ {u v} \left\| \left[ \dot {z} _ {i} \right] _ {u v} - \left[ z _ {i} \right] _ {u v} \right\| _ {2} ^ {2}}{\sum_ {u , v} \left[ \mathcal {M} \left(\dot {z} _ {i}\right) \right] _ {u v}}, \tag {8} +$$ + +where $u, v$ are spatial positions indices. + +Appearance Control We define the appearance loss as: + +$$ +\mathcal {L} _ {\mathrm {a p p}} \left(z _ {i}, z ^ {0 \star}\right) = \left\| \mathcal {A} \left(z _ {i}\right) - \mathcal {A} \left(z ^ {0 \star}\right) \right\| _ {2} ^ {2}, \tag {9} +$$ + +where $\mathcal{A}$ is the operator of exacting appearance descriptors. + +Sequential Concept Integration To maintain coherence across the progressive addition of concepts, we implement a sequential integration process that updates the latent noise representation while preserving previously incorporated elements. For each sketch $s_i$ in the sequence, we compute: + +$$ +z _ {i} = z _ {i - 1} + \Delta z _ {i}, \tag {10} +$$ + +where $z_{i-1}$ is the optimized latent from the previous concept, and $\Delta z_i$ represents the incremental update required to incorporate the current sketch concept $s_i$ . This update is derived from our loss function: + +$$ +\mathcal {L} _ {i} = \alpha_ {i} \mathcal {L} _ {\text {n e w}} \left(z _ {i}, \dot {z} _ {i}, z ^ {0 \star}\right) + \beta_ {i} \mathcal {L} _ {\text {p r e s e r v e}} \left(z _ {i}, \dot {z} _ {0, \dots , i - 1}\right), \tag {11} +$$ + +where $\alpha_{i}$ and $\beta_{i}$ are balancing coefficients. $\mathcal{L}_{\mathrm{new}}$ ensures proper integration of the current $z_{i}$ and is defined as: + +$$ +\mathcal {L} _ {\text {n e w}} \left(z _ {i}, \dot {z} _ {i}, z ^ {0 \star}\right) = \mathcal {L} _ {\text {s p a t i a l}} \left(z _ {i}, \dot {z} _ {i}\right) + \mathcal {L} _ {\text {a p p}} \left(z _ {i}, z ^ {0 \star}\right), \tag {12} +$$ + +while $\mathcal{L}_{\mathrm{preserve}}$ maintains the integrity of previously optimized concepts, given by: + +$$ +\mathcal {L} _ {\text {p r e s e r v e}} \left(z _ {i}, \dot {z} _ {0, \dots , i - 1}\right) = \sum_ {j = 1} ^ {i - 1} \gamma^ {i - j} \mathcal {L} _ {\text {s p a t i a l}} \left(z _ {i}, \dot {z} _ {j}\right), \tag {13} +$$ + +![](images/44bc7404717272fe3bbed7925a95892c497e415a622eefe6a972f57bdf68e924.jpg) +Figure 5. Qualitative comparison of existing sketch-to-image generation methods, and our method on FMC and CMC datasets. Existing training-based methods tend to overfit to sketch lines, often failing to reconstruct scenes realistically, while training-free methods frequently misinterpret or overlook key concepts. In comparison, our approach more faithfully represents real-world physical scenes. The reading and processing order of the sketches is indicated by colors as follows: $\bullet$ $\bullet$ $\bullet$ + +where $j$ means preserving established concepts, and $\gamma^{i - j}$ is a decay factor that weights the importance of earlier concepts based on their step gap from the current step. + +# 4. Experiments and Results + +# 4.1. Experiment Setup + +Datasets To rigorously evaluate our method, we have undertaken extensive efforts to develop two specialized datasets: the Composed Multi-Concept (CMC) dataset and the Freehand Multi-Concept (FMC) dataset. Both datasets are designed to comprehensively assess the capabilities of our framework in handling multi-concept scene understanding and generation. CMC dataset comprises 142 scene sketches, each containing 2-4 distinct concepts sampled from the 150 semantic classes in the Sketchy dataset [24]. Each is accompanied by annotated textual prompts with the help of multimodal large language models. The CMC dataset does not provide real-world images captured by physical cameras for reference. FMC dataset is constructed from freehand sketches drawn by volunteers, based on real-world images sourced from publicly available outdoor + +scene datasets. It comprises 42 hand-drawn sketches, each containing 2-3 distinct concepts. Additionally, each sketch is paired with a corresponding real-world image as a reference for validation and evaluation purposes. + +Evaluation Metrics To conduct a comprehensive quantitative evaluation of the quality of generated images, four widely-used evaluation metrics were employed, including Frechet Inception Distance (FID) [10], CLIP-T Score (CLIP-T) [21], Human Preference Score (HPS) [29], and CLIP-I distance (CLIP-I) [21]. FID measures the feature distribution similarity between generated and real images. Lower FID values indicate better distribution matching $(\downarrow)$ . CLIP-T quantifies the image-text alignment degree within the CLIP-T embedding space. A higher CLIP-T score reflects greater semantic alignment between the generated image and text prompt $(\uparrow)$ . HPS measures the subjective quality of the generated images based on human evaluations. Higher HPS values suggest that the generated images are more favorable or preferred by human evaluators $(\uparrow)$ . CLIP-I measures the alignment between two images. A higher value suggests that the two images are more consistent in terms of their semantic content $(\uparrow)$ . + +![](images/d895d59adab5fa97f378c744bd46e240af54896a72d2c4e7bd4b9c5f38a90c6f.jpg) +Figure 6. Ablation study on FMC and CMC datasets. Column headers are defined as follows: T2I denotes the baseline text-to-image diffusion model; OP (Overall Perception) indicates the single-step guidance approach utilizing complete sketch as a holistic conditioning signal; CSO refers to our proposed Concept-wise Sequential Optimization (Section 3.5); SE represents the Sketch Encoding module (Section 3.3) that enhances structural plausibility by injecting diffusion priors into imperfect sketches. The qualitative comparison demonstrates that CSO achieves finer layout control than OP. While SE effectively bridges the domain gap between freehand sketches and physical reality through latent space regularization, ultimately elevating generation quality. + +
Metric MethodCMCFMC
FID ↓CLIP-T ↑HPS ↑FID ↓CLIP-T ↑HPS ↑CLIP-I ↑
Training-based methods
ControlNet [33]15.2861.710.79414.4566.800.8020.622
T2I-Adapter [19]18.1361.610.74515.3761.610.7450.689
Uni-ControlNet [37]19.6062.580.76118.5362.580.7610.712
Training-free methods
Multi-Sketch [4]13.0564.110.74011.5972.020.7750.690
PNP [27]17.0163.270.84016.2671.240.8520.593
FreeControl [17]14.5264.740.75711.1766.670.7590.775
Ours11.5666.640.85010.1874.250.8620.797
+ +Table 1. Quantitative results on CMC and FMC datasets. Our method consistently surpasses all training-free methods in distribution matching, image-text alignment, subjective quality, and appearance details as measured by FID, CLIP-T, HPS, and CLIP-I. Our method achieves superior image-text alignment compared to training-based methods ( $\downarrow$ indicates lower values are better, while $\uparrow$ is opposite). + +Implementation Details In the implementation of our proposed method, we set the learning rate $\eta$ in Eq. (2) and the threshold parameter $\tau$ in Eq. (7) is 5.0 and 0.3, respectively. Meanwhile, hyperparameters $\alpha_{i}$ and $\beta_{i}$ in Eq. (11) are set as 0.8 and 0.2, respectively. $\gamma$ in Eq. (13) is set as 0.9. Our method is performed on one NVIDIA GeForce RTX 4090, and input sketch images are $512 \times 512$ in resolution. + +# 4.2. Comparisons + +We conducted a comprehensive quantitative and qualitative evaluation of the proposed method and competing methods on the CMC and FMC datasets. Our competing methods include the training-based approaches ControlNet [33], T2I-Adapter [19], and Uni-ControlNet [37], as well as the training-free approaches Multi-Sketch [4], PNP [27], and FreeControl [17]. The quantitative results are presented in Table 1 and qualitative results are shown in Figure 5. + +We observe that training-based methods [19, 33, 37] are highly susceptible to subjective biases in sketches and of + +ten fail to adhere to text prompts, producing distorted and unrealistic images with lower CLIP-T scores. Training-free methods [4, 8, 17] suffer from concept omissions, conspicuous artifacts, and inconsistencies both among concepts and between concepts and their backgrounds, resulting in poor image realism. These issues stem from mutual interference during the "snapshot" generation of multiple concepts. In contrast, our proposed method faithfully reconstructs real-world scenes and achieves the best performance across all evaluation metrics on both the CMC and FMC datasets, demonstrating its effectiveness. + +![](images/bb3ee8fc8f5cbc78041af524b18487d10b76b38b4a5760795f2a00c1a23c44c8.jpg) +Figure 7. User preference experiment. We asked participants to answer four questions based on the reference images, subjective readouts, and generated images, which included one produced by our method and one by a competing method. Results show that users significantly preferred our outputs over those of competing methods in most cases. + +# 4.3. Ablation Study + +To validate the effectiveness of each module in our method, we conducted ablation experiments on the FMC and CMC datasets, with the visualization results shown in Figure 6. + +It is observed that traditional T2I diffusion models (T2I) generate images that are close to text prompts. While incorporating sketches as guidance further constrains the consistency of concept positioning, deviations in appearance and spatial alignment persist (T2I, OP). The sequential concept understanding facilitated by Concept-wise Sequential + +![](images/1725d49d543ff881a7f5b1e56687215f6fcc6a05644b4cbc5156804a37dadc4b.jpg) +Figure 8. Control experiments on user preferences. We controlled three factors to examine their influence: whether the rater drew the photo, the availability timing of the reference image ( $\infty$ vs. 5s), and the types of questions asked. In general, our method is consistently better than the competing method, FreeControl. + +Optimization (CSO) in Section 3.5 enhances the alignment of concept positioning and appearance with input sketches (CSO). Simultaneously, the Sketch Encoding (SE) module in Section 3.3 aligns content latent variables with real-world objects, mitigating geometric misalignment (T2I, OP, SE). When integrating all modules, our method faithfully reconstructs real-world scenes and achieves optimal results (Ours), demonstrating the effectiveness of each module and underscoring the indispensable nature of their design. + +# 4.4. User Preference Experiments + +We invited 12 participants (group A) to provide subjective readouts and an additional 28 participants (group B) to evaluate the reconstructed images. Participants in group A were first asked to select a deeply impressive scene from their photo gallery and then provide a textual description along with sketches within two minutes. Our method and competing approaches, including two baselines (text-reward optimization and text-reward optimization with sketch-guided optimization based on all concepts) as well as an optimal method, FreeControl [17], were used to generate images from these readouts. Participants in both groups were then shown the subjective readouts, the reference image, and the generated images for the same scene. They were asked to answer four single-choice questions in a single-blind manner: Q1: Which image is closer to the original image (both style and content)? Q2: Which image has higher visual quality (e.g., more realistic, more natural)? Q3: Which image better matches the structure of the sketch? Q4: Which image better matches the textual description? + +Figure 7 presents the statistics of the collected results. Participants judged our reconstructions as being closer to the corresponding real-world scenes compared to competing methods. Similar conclusions were observed for the other three questions. However, there was an exception in the comparison between our method and text-reward optimization regarding image quality (Q2). We observed that text-reward optimization without additional control tends to hallucinate details that enhance perceived realism, leading some users to prefer it for Q2. This also explains why our + +advantage over text-reward optimization in Q1 was not statistically significant and why some participants expressed hesitation in this case. + +We further conducted controlled experiments to examine the effects of three factors: whether the rater provided the subjective readouts, the availability timing of the reference image, and the types of questions asked. Results in Figure 8 show that our method consistently outperformed FreeControl [17], regardless of these variations. Notably: 1) Participants who provided subjective readouts showed slightly lower preference for our method in terms of similarity to the reference image compared to those who did not, while showing a higher preference in terms of alignment with the readouts and perceived image quality. We suppose this is mainly because of the subjectivity in reading out. 2) Reducing the time participants could view the reference image decreased the preference for our method; when allowed to compare freely, participants preferred our results more strongly. + +# 5. Conclusions + +This paper presented Subjective Camera, a paradigm for reconstructing real-world scenes from humans' mental impressions. Our framework relies on the most accessible readouts of mental imagery, i.e., textual descriptions and sketches, and employs a training-free, sequence-aware optimization process. This approach eliminates the need for costly fine-tuning of text-to-image diffusion models on large-scale data while mitigating challenges such as dependence on sketch quality and geometric misalignment. Extensive experiments on both synthetic and real-world sketch datasets demonstrate that Subjective Camera achieves state-of-the-art performance, outperforming both training-free and training-based sketch-to-image generation methods across all evaluation metrics. Furthermore, user preference studies confirm that participants consistently favor our method. + +In closing, we address a natural question: "With the rapid advancement of generative AI, will physical cameras ever become unnecessary" For now, the answer is no. First, the reconstructed images still fall short of perfectly matching actual scenes. This limitation largely stems from the fact that current T2I models are trained on highly diverse data, creating gaps between the prior and the specific scene to be reconstructed. One promising direction to mitigate this is to personalize the prior models using user-specific galleries. Second, we observe that variations in the order and quality of sketches have a significant impact on reconstruction results. To address this, we plan to enhance our framework with computational tools that guide users in providing more precise and suitable subjective readouts. For these reasons, we designate this system as Subjective Camera 1.0, leaving these improvements for future iterations. + +Acknowledgements This work was supported in part by the National Natural Science Foundation of China (62225113, 624B2109, 623B2079) and the Zhongguancun Academy Project (20240308). We thank Po-Hsun Huang for providing the questionnaire system for our user study, Kaijin Zheng for assistance with data processing, and all participants who took part in the studies. + +# References + +[1] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. Proceedings of Machine Learning Research, 202:1737-1752, 2023. 2, 3 +[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18392-18402, 2023. 3 +[3] Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5343-5353, 2024. 3 +[4] Zhenwei Cheng, Lei Wu, Changshuo Wang, and Xiangxu Meng. Scene sketch-to-image synthesis based on multi-object control. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3775-3779. IEEE, 2024. 2, 3, 7 +[5] Dave Epstein, Allan Jabri, Ben Poole, Alexei Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. Advances in Neural Information Processing Systems, 36:16222-16239, 2023. 3 +[6] Luca Eyring, Shyamgopal Karthik, Karsten Roth, Alexey Dosovitskiy, and Zeynep Akata. Reno: Enhancing one-step text-to-image models through reward-based noise optimization. Advances in Neural Information Processing Systems, 37:125487-125519, 2025. 3, 4 +[7] Xiefan Guo, Jinlin Liu, Miaomiao Cui, Jiankai Li, Hongyu Yang, and Di Huang. Initno: Boosting text-to-image diffusion models via initial noise optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9380-9389, 2024. 3 +[8] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In The Eleventh International Conference on Learning Representations. 7 +[9] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control.(2022). URL https://arxiv.org/abs/2208.01626, 3, 2022. 3 +[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[11] Zechao Hu, Zhengwei Yang, Hao Li, Yixiong Zou, Fengbin Zhu, and Zheng Wang. Unified category and style generalization for instance-level sketch retrieval. In Proceedings of + +the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, page 864-873. Association for Computing Machinery, 2025. 2 +[12] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Martiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36: 36652-36663, 2023. 3 +[13] Stephen M Kosslyn. Aspects of a cognitive neuroscience of mental imagery. Science, 240(4859):1621-1626, 1988. 2 +[14] Seung Hyun Lee, Yinxiao Li, Junjie Ke, Innfarn Yoo, Han Zhang, Jiahui Yu, Qifei Wang, Fei Deng, Glenn Entis, Junfeng He, et al. Parrot: Pareto-optimal multi-reward reinforcement learning framework for text-to-image generation. In European Conference on Computer Vision, pages 462-478. Springer, 2024. 3, 4 +[15] Kejun Lin, Zhixiang Wang, Zheng Wang, Yinqiang Zheng, and Shin'ichi Satoh. Beyond domain gap: Exploiting subjectivity in sketch-based person retrieval. In Proceedings of the 31st ACM international conference on multimedia, pages 2078-2089, 2023. 2 +[16] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 4 +[17] Sicheng Mo, Fangzhou Mu, Kuan Heng Lin, Yanli Liu, Bochen Guan, Yin Li, and Bolei Zhou. Freecontrol: Training-free spatial control of any text-to-image diffusion model with any condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7465-7475, 2024. 3, 7, 8 +[18] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6038–6047, 2023. 4 +[19] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296-4304, 2024. 2, 3, 7 +[20] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195-4205, 2023. 2, 3 +[21] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 6 +[22] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 3 +[23] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image + +synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3 +[24] Patsorn Sangkloy, Nathan Burnell, Cusuh Ham, and James Hays. The sketchy database: learning to retrieve badly drawn bunnies. ACM Transactions on Graphics (TOG), 35(4):1-12, 2016. 6 +[25] Vishnu Sarukkai, Lu Yuan, Mia Tang, Maneesh Agrawala, and Kayvon Fatahalian. Block and detail: Scaffolding sketch-to-image generation. In Proceedings of the 37th Annual ACM Symposium on User Interface Software and Technology, pages 1-13, 2024. 2, 3 +[26] Aravindan Sundaram, Ujjayan Pal, Abhimanyu Chauhan, Aishwarya Agarwal, and Srikrishna Karanam. Cocono: Attention contrast-and-complete for initial noise optimization in text-to-image synthesis. arXiv preprint arXiv:2411.16783, 2024. 3 +[27] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 7 +[28] Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. Sketch-guided text-to-image diffusion models. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2, 3 +[29] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score: Better aligning text-to-image models with human preference. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2096-2105, 2023. 3, 6 +[30] Zhenbei Wu, Haoge Deng, Qiang Wang, Di Kong, Jie Yang, and Yonggang Qi. Sketchscene: Scene sketch to image generation with diffusion models. In 2023 IEEE International Conference on Multimedia and Expo (ICME), pages 2087-2092. IEEE, 2023. 2 +[31] Xin Xie and Dong Gong. Dymo: Training-free diffusion model alignment with dynamic multi-objective scheduling. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 13220-13230, 2025. 3 +[32] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36:15903-15935, 2023. 3 +[33] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3, 7 +[34] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Scaling in-the-wild training for diffusion-based illumination harmonization and editing by imposing consistent light transport. In The Thirteenth International Conference on Learning Representations, 2025. 3 +[35] Tianyu Zhang and Haoran Xie. Sketch-guided text-to-image generation with spatial control. In 2024 2nd International Conference on Computer Graphics and Image Processing (CGIP), pages 153-159. IEEE, 2024. 2, 3 + +[36] Tianyu Zhang, Xiaoxuan Xie, Xusheng Du, and Haoran Xie. Sketch-guided scene image generation. arXiv preprint arXiv:2407.06469, 2024. 2, 3 +[37] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. Advances in Neural Information Processing Systems, 36:11127-11150, 2023. 2, 3, 7 +[38] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22490-22499, 2023. 3 \ No newline at end of file diff --git a/2025/Subjective Camera 1.0_ Bridging Human Cognition and Visual Reconstruction through Sequence-Aware Sketch-Guided Diffusion/images.zip b/2025/Subjective Camera 1.0_ Bridging Human Cognition and Visual Reconstruction through Sequence-Aware Sketch-Guided Diffusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..992ce3689d896ce06adf1ae7e087075deb40e2fc --- /dev/null +++ b/2025/Subjective Camera 1.0_ Bridging Human Cognition and Visual Reconstruction through Sequence-Aware Sketch-Guided Diffusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fa6cabbf7b3b6f0c4e8413f69a83382153047a1897fca166cf3bc5e57c3e62c4 +size 631832 diff --git a/2025/Subjective Camera 1.0_ Bridging Human Cognition and Visual Reconstruction through Sequence-Aware Sketch-Guided Diffusion/layout.json b/2025/Subjective Camera 1.0_ Bridging Human Cognition and Visual Reconstruction through Sequence-Aware Sketch-Guided Diffusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6f229eb881a54937ac4c3273c78792718faee2bf --- /dev/null +++ b/2025/Subjective Camera 1.0_ Bridging Human Cognition and Visual Reconstruction through Sequence-Aware Sketch-Guided Diffusion/layout.json @@ -0,0 +1,8434 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 98, + 124, + 120 + ], + "blocks": [ + { + "bbox": [ + 97, + 98, + 124, + 120 + ], + "lines": [ + { + "bbox": [ + 97, + 98, + 124, + 120 + ], + "spans": [ + { + "bbox": [ + 97, + 98, + 124, + 120 + ], + "type": "image", + "image_path": "036f49fca5b1146f0dd58d8ee5fc660c76717546e46a77730b006441659d3a92.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 124, + 105, + 512, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 105, + 512, + 122 + ], + "spans": [ + { + "bbox": [ + 124, + 105, + 512, + 122 + ], + "type": "text", + "content": "Subjective Camera 1.0: Bridging Human Cognition and Visual" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 100, + 124, + 510, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 124, + 510, + 140 + ], + "spans": [ + { + "bbox": [ + 100, + 124, + 510, + 140 + ], + "type": "text", + "content": "Reconstruction through Sequence-Aware Sketch-Guided Diffusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 82, + 161, + 531, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 161, + 531, + 191 + ], + "spans": [ + { + "bbox": [ + 82, + 161, + 531, + 191 + ], + "type": "text", + "content": "Haoyang Chen" + }, + { + "bbox": [ + 82, + 161, + 531, + 191 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 82, + 161, + 531, + 191 + ], + "type": "text", + "content": " Dongfang Sun" + }, + { + "bbox": [ + 82, + 161, + 531, + 191 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 82, + 161, + 531, + 191 + ], + "type": "text", + "content": " Caoyuan Ma" + }, + { + "bbox": [ + 82, + 161, + 531, + 191 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 82, + 161, + 531, + 191 + ], + "type": "text", + "content": " Shiqin Wang" + }, + { + "bbox": [ + 82, + 161, + 531, + 191 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 82, + 161, + 531, + 191 + ], + "type": "text", + "content": " Kewei Zhang" + }, + { + "bbox": [ + 82, + 161, + 531, + 191 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 82, + 161, + 531, + 191 + ], + "type": "text", + "content": " Zheng Wang" + }, + { + "bbox": [ + 82, + 161, + 531, + 191 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 82, + 161, + 531, + 191 + ], + "type": "text", + "content": " Zhixiang Wang" + }, + { + "bbox": [ + 82, + 161, + 531, + 191 + ], + "type": "inline_equation", + "content": "^{4}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 60, + 193, + 553, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 193, + 553, + 236 + ], + "spans": [ + { + "bbox": [ + 60, + 193, + 553, + 236 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 60, + 193, + 553, + 236 + ], + "type": "text", + "content": "National Engineering Research Center for Multimedia Software, Institute of Artificial Intelligence, School of Computer Science, Wuhan University " + }, + { + "bbox": [ + 60, + 193, + 553, + 236 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 60, + 193, + 553, + 236 + ], + "type": "text", + "content": "Hubei Key Laboratory of Multimedia and Network Communication Engineering " + }, + { + "bbox": [ + 60, + 193, + 553, + 236 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 60, + 193, + 553, + 236 + ], + "type": "text", + "content": "Zhongguancun Academy, Beijing, China. 100094 " + }, + { + "bbox": [ + 60, + 193, + 553, + 236 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 60, + 193, + 553, + 236 + ], + "type": "text", + "content": "CyberAgent AI Lab, Japan" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 72, + 264, + 86, + 279 + ], + "blocks": [ + { + "bbox": [ + 72, + 264, + 86, + 279 + ], + "lines": [ + { + "bbox": [ + 72, + 264, + 86, + 279 + ], + "spans": [ + { + "bbox": [ + 72, + 264, + 86, + 279 + ], + "type": "image", + "image_path": "f3172a01beebfc96259e00a3d02b0f8b0756a128e359a6431d8dd59d4d807e32.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 89, + 267, + 165, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 267, + 165, + 278 + ], + "spans": [ + { + "bbox": [ + 89, + 267, + 165, + 278 + ], + "type": "text", + "content": "Physical Camera" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 82, + 281, + 159, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 281, + 159, + 291 + ], + "spans": [ + { + "bbox": [ + 82, + 281, + 159, + 291 + ], + "type": "text", + "content": "Photo after Processing" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 78, + 294, + 154, + 372 + ], + "blocks": [ + { + "bbox": [ + 78, + 294, + 154, + 372 + ], + "lines": [ + { + "bbox": [ + 78, + 294, + 154, + 372 + ], + "spans": [ + { + "bbox": [ + 78, + 294, + 154, + 372 + ], + "type": "image", + "image_path": "198dea298b5766e5660194fc5b560b65a5d8808d85811a9a3a5e5487d22e5e9a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 78, + 384, + 156, + 460 + ], + "blocks": [ + { + "bbox": [ + 78, + 384, + 156, + 460 + ], + "lines": [ + { + "bbox": [ + 78, + 384, + 156, + 460 + ], + "spans": [ + { + "bbox": [ + 78, + 384, + 156, + 460 + ], + "type": "image", + "image_path": "2e44cc24f8fb45f8fe7385acfcd4dba41e09513702fd2e31e90b0373617f18b0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 471, + 555, + 506 + ], + "lines": [ + { + "bbox": [ + 54, + 471, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 54, + 471, + 555, + 506 + ], + "type": "text", + "content": "Figure 1. What is a subjective camera? A subjective camera transforms a person's memory into photorealistic images. We take a small yet crucial step toward this vision by leveraging generative models to sequentially decode textual and sketch-based readouts in their natural order, paving the way for reconstructing moments without physical cameras." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 255, + 263, + 273, + 280 + ], + "blocks": [ + { + "bbox": [ + 255, + 263, + 273, + 280 + ], + "lines": [ + { + "bbox": [ + 255, + 263, + 273, + 280 + ], + "spans": [ + { + "bbox": [ + 255, + 263, + 273, + 280 + ], + "type": "image", + "image_path": "781beaaf605b41f5a82caaef17a296da083e4925311b0e88c7a89a5664ab39e5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 277, + 267, + 432, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 267, + 432, + 278 + ], + "spans": [ + { + "bbox": [ + 277, + 267, + 432, + 278 + ], + "type": "text", + "content": "The Concept of Subjective Camera" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 215, + 281, + 332, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 281, + 332, + 291 + ], + "spans": [ + { + "bbox": [ + 215, + 281, + 332, + 291 + ], + "type": "text", + "content": "Readout (Text and Sketch Sequences)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 180, + 301, + 257, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 301, + 257, + 314 + ], + "spans": [ + { + "bbox": [ + 180, + 301, + 257, + 314 + ], + "type": "text", + "content": "There is a black car" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 179, + 317, + 258, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 317, + 258, + 358 + ], + "spans": [ + { + "bbox": [ + 179, + 317, + 258, + 358 + ], + "type": "text", + "content": "towing a white trailer on the road near mountains." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 178, + 397, + 255, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 397, + 255, + 439 + ], + "spans": [ + { + "bbox": [ + 178, + 397, + 255, + 439 + ], + "type": "text", + "content": "There is a motorbike and a tent on the grass." + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 276, + 297, + 345, + 366 + ], + "blocks": [ + { + "bbox": [ + 276, + 297, + 345, + 366 + ], + "lines": [ + { + "bbox": [ + 276, + 297, + 345, + 366 + ], + "spans": [ + { + "bbox": [ + 276, + 297, + 345, + 366 + ], + "type": "image", + "image_path": "6e76ebb9c9b187a325fc221c175ddff7dec200ef986cabc99ef19e3f06bced32.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 345, + 369, + 362, + 384 + ], + "blocks": [ + { + "bbox": [ + 345, + 369, + 362, + 384 + ], + "lines": [ + { + "bbox": [ + 345, + 369, + 362, + 384 + ], + "spans": [ + { + "bbox": [ + 345, + 369, + 362, + 384 + ], + "type": "image", + "image_path": "da44a85605c00f461abc1335f1f9da5d685896f9f5a65a0d08839818d8f46c1c.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 361, + 281, + 538, + 372 + ], + "blocks": [ + { + "bbox": [ + 361, + 281, + 538, + 372 + ], + "lines": [ + { + "bbox": [ + 361, + 281, + 538, + 372 + ], + "spans": [ + { + "bbox": [ + 361, + 281, + 538, + 372 + ], + "type": "image", + "image_path": "46beac0003d7a4dfedbf29fd69a6b3bee621f52ed784ce88f46ee68655b6fedd.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 262, + 380, + 350, + 460 + ], + "blocks": [ + { + "bbox": [ + 262, + 380, + 350, + 460 + ], + "lines": [ + { + "bbox": [ + 262, + 380, + 350, + 460 + ], + "spans": [ + { + "bbox": [ + 262, + 380, + 350, + 460 + ], + "type": "image", + "image_path": "d870bcf8c119e317df29b4087a76c625145f021f658294d0322cec026198e1df.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 363, + 380, + 538, + 459 + ], + "blocks": [ + { + "bbox": [ + 363, + 380, + 538, + 459 + ], + "lines": [ + { + "bbox": [ + 363, + 380, + 538, + 459 + ], + "spans": [ + { + "bbox": [ + 363, + 380, + 538, + 459 + ], + "type": "image", + "image_path": "1ef267c53bf9908bd17d6b36f1ddc725503d068b70310c85e24991f33458b465.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 152, + 533, + 200, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 533, + 200, + 545 + ], + "spans": [ + { + "bbox": [ + 152, + 533, + 200, + 545 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 53, + 560, + 297, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 560, + 297, + 680 + ], + "spans": [ + { + "bbox": [ + 53, + 560, + 297, + 680 + ], + "type": "text", + "content": "We introduce the concept of a subjective camera to reconstruct meaningful moments that physical cameras fail to capture. We propose Subjective Camera 1.0, a framework for reconstructing real-world scenes from readily accessible subjective readouts, i.e., textual descriptions and progressively drawn rough sketches. Built on optimization-based alignment of diffusion models, our approach avoids large-scale paired training data and mitigates generalization issues. To address the challenge of integrating multiple abstract concepts in real-world scenar" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 533, + 556, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 643 + ], + "type": "text", + "content": "ios, we design a Sequence-Aware Sketch-Guided Diffusion framework with three loss terms for concept-wise sequential optimization, following the natural order of subjective readouts. Experiments on two datasets demonstrate that our method achieves state-of-the-art performance in image quality as well as spatial and semantic alignment with target scenes. User studies with 40 participants further confirm that our approach is consistently preferred. Our project page is at: subjective-camera.github.io" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 314, + 668, + 395, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 668, + 395, + 681 + ], + "spans": [ + { + "bbox": [ + 314, + 668, + 395, + 681 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "Cameras have long been the most popular devices for preserving memorable moments. However, a physical camera" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 63, + 692, + 223, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 692, + 223, + 703 + ], + "spans": [ + { + "bbox": [ + 63, + 692, + 223, + 703 + ], + "type": "text", + "content": "These authors contributed equally to this work." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 63, + 703, + 148, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 703, + 148, + 712 + ], + "spans": [ + { + "bbox": [ + 63, + 703, + 148, + 712 + ], + "type": "text", + "content": "Corresponding authors." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17838" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "cannot always be present to capture every moment. This raises a question of how such unrecorded experiences can be preserved as pixels in the absence of a camera." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 109, + 295, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 109, + 295, + 276 + ], + "spans": [ + { + "bbox": [ + 54, + 109, + 295, + 276 + ], + "type": "text", + "content": "This paper proposes the concept of a subjective camera1, where humans function as \"imaging devices\", aiming to faithfully reconstruct real-world scenes from memory. Specifically, we define a subjective camera as a system through which individuals encode sensory inputs into memory based on personal salience and emotional context, and later read out these stored representations and decode them into pixels that reconstruct scenes they experienced. We draw an analogy to physical cameras: just as a camera records, reads out, and reconstructs the visual world through optical, electronic, and computational processes, the human mind selectively captures (\"records\") perceptual details, which are organized into a retrievable format (\"read out\") and later reconstructs them as pixels." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 277, + 295, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 295, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 295, + 432 + ], + "type": "text", + "content": "It is generally challenging for humans without professional training to directly reconstruct a dense, pixel-wise image. Instead, the most practical approach is to read out the mental scene through textual descriptions and/or free-hand sketches and leverage computational tools to decode these readouts. However, text alone often fails to fully convey a mental image, particularly in terms of object layout and fine details. Freehand sketches provide complementary information beyond text but remain sparse and often contain noise and uncertainty [11, 15]. With the advancement of generative models [20, 23], this long-standing vision is becoming increasingly feasible by decoding such subjective readouts into photorealistic imagery." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 433, + 295, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 295, + 588 + ], + "type": "text", + "content": "Although not intended to reconstruct real-world scenes, several works have explored generating images from combined text and sketch inputs [4, 25, 30, 35-37]. These approaches either adapt model weights using large-scale paired data [19, 25, 28, 30, 33, 36, 37] or modify the latent variables through per-scene optimization [4, 28, 35]. Training-based methods directly learn the mapping from sketches to pixels. However, they require extensive paired data and struggle with user-specific biases, often generalizing poorly to abstract sketches beyond the training distribution. In contrast, training-free methods avoid the heavy data requirements, high computational costs, and generalization challenges of training-based approaches." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 588, + 295, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 295, + 684 + ], + "type": "text", + "content": "However, optimization-based methods often fail to reconstruct multiple concepts from idiosyncratic sketches, as would be required for recreating real-world scenes. This often leads to missing concepts or misaligned spatial relationships (see Figure 2a). The main obstacle lies in simultaneously interpreting multi-concept sketches, which inherently combine varying levels of abstraction and randomness across different concepts. For instance, freehand sketches" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 340, + 71, + 529, + 127 + ], + "blocks": [ + { + "bbox": [ + 340, + 71, + 529, + 127 + ], + "lines": [ + { + "bbox": [ + 340, + 71, + 529, + 127 + ], + "spans": [ + { + "bbox": [ + 340, + 71, + 529, + 127 + ], + "type": "image", + "image_path": "a35ba4a51f5908149de79a0d33288e1b6f0e75abf38588ca6171cdc6f24acf02.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 368, + 129, + 504, + 138 + ], + "lines": [ + { + "bbox": [ + 368, + 129, + 504, + 138 + ], + "spans": [ + { + "bbox": [ + 368, + 129, + 504, + 138 + ], + "type": "text", + "content": "a: Limitation of \"snapshot\" generation" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 340, + 144, + 527, + 200 + ], + "blocks": [ + { + "bbox": [ + 340, + 144, + 527, + 200 + ], + "lines": [ + { + "bbox": [ + 340, + 144, + 527, + 200 + ], + "spans": [ + { + "bbox": [ + 340, + 144, + 527, + 200 + ], + "type": "image", + "image_path": "2c41e85fa300491fce409300694f16caa528e5f6e7f20fb0cd20939030c53796.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 203, + 507, + 213 + ], + "lines": [ + { + "bbox": [ + 361, + 203, + 507, + 213 + ], + "spans": [ + { + "bbox": [ + 361, + 203, + 507, + 213 + ], + "type": "text", + "content": "b: Inspiration from human readout order" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 223, + 555, + 300 + ], + "lines": [ + { + "bbox": [ + 313, + 223, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 555, + 300 + ], + "type": "text", + "content": "Figure 2. Motivation for concept-wise sequential optimization. (a) Attempting to \"snap\" all concepts into a complete image at once using generative diffusion priors often leads to suboptimal results, such as distortions and misaligned semantics or layouts. (b) In contrast, humans reconstruct mental scenes step by step. This natural readout inspires our design2, making it better suited to leverage current generative diffusion priors." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 323, + 553, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 553, + 407 + ], + "type": "text", + "content": "may only roughly indicate object boundaries, omit fine details such as texture and geometry, or distort proportions in non-standard ways. While generative priors can fill sparse regions and mitigate randomness during the transformation from sketches to pixels, attempting to integrate all concepts at once often reduces the problem to a least-squares solution [1, 4], ultimately blurring their distinct characteristics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 407, + 554, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 554, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 554, + 503 + ], + "type": "text", + "content": "Our key idea is simple — to apply optimizations to each concept individually and do so sequentially. This design considers not only the limitations of the \"snapshot\" generation process but also human cognition. Humans tend to describe their mental images step-by-step [13], typically starting with a textual description of abstract concepts or the overall scene structure, and then progressively refining individual elements with increasing detail (see Figure 2b)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 505, + 554, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 554, + 660 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 554, + 660 + ], + "type": "text", + "content": "Building upon this idea, we introduce a sequence-aware, sketch-guided diffusion framework that sequentially optimizes the latent noise of a pre-trained text-to-image (T2I) diffusion model [23] to align individual subjective readouts. The order of these optimizations mirrors the natural process of reading out subjective impressions: it begins with a text-reward optimization (Section 3.2) to align the generated image with the initial textual description, followed by concept-wise sequential optimizations (Section 3.5) to incorporate progressively provided sketches that may encode shapes, poses, and spatial arrangements. The sequential optimization also incorporates a loss to ensure that previously optimized concepts are not significantly altered." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 326, + 661, + 553, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 661, + 553, + 673 + ], + "spans": [ + { + "bbox": [ + 326, + 661, + 553, + 673 + ], + "type": "text", + "content": "Two additional components are proposed to improve" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "1 While the term \"subjective camera\" appears in cinematic contexts, our usage here differs significantly in meaning." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 683, + 553, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 683, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 683, + 553, + 713 + ], + "type": "text", + "content": "2This design philosophy is reminiscent of the progressive decoding used in rolling-shutter image sensors, which arises from the need to accommodate the limited processing capacity of the circuits." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17839" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": "the concept-wise sequential optimization. First, each input sketch is individually encoded into the latent space to provide spatial guidance. To bridge the gap between rough sketch shapes and their corresponding real-world forms, we introduce an optimization-based inversion strategy for refining this spatial guidance (Section 3.3). Second, because input sketches often lack appearance details, directly optimizing with their guidance over multiple iterations may lead to unnatural textures and degraded visual fidelity. We therefore leverage the latent representation obtained from text-reward optimization to define an appearance loss (Section 3.4), thereby preserving visual style quality." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 217, + 294, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 294, + 384 + ], + "type": "text", + "content": "By processing sketches individually and sequentially, our approach naturally aligns with the human cognition process, enabling users to incrementally construct mental images while ensuring each addition respects and complements previous elements. This sequence-aware generation strategy provides fine-grained control over complex multi-concept scenes while preserving the user's subjective intent encoded in the drawing order. Evaluations on two datasets demonstrate that our method achieves state-of-the-art performance in recreating a real-world scene from human memory, demonstrating the potential of subjective cameras. Notably, our framework is entirely training-free, eliminating the need for large-scale paired datasets and avoiding generalization." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 386, + 236, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 386, + 236, + 397 + ], + "spans": [ + { + "bbox": [ + 67, + 386, + 236, + 397 + ], + "type": "text", + "content": "To sum up, our contributions are twofold:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 399, + 293, + 517 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 61, + 399, + 293, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 399, + 293, + 446 + ], + "spans": [ + { + "bbox": [ + 61, + 399, + 293, + 446 + ], + "type": "text", + "content": "We propose the concept of a subjective camera, which functions the human as an \"imaging device\" and leverages computational tools to transform cognitive impressions into photographs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 447, + 293, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 447, + 293, + 517 + ], + "spans": [ + { + "bbox": [ + 61, + 447, + 293, + 517 + ], + "type": "text", + "content": "We propose a sequence-aware diffusion-guided generation framework that enables faithful reconstruction of complex, multi-concept scenes. This framework aligns with human cognition and surpasses \"snapshot\" generation methods, thereby establishing a new paradigm for cognition-driven image generation." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 533, + 142, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 142, + 546 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 142, + 546 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": "Aligning Image Synthesis Models Despite their success, T2I generative models [20, 22, 23] often fail to reproduce the fine-grained semantics and compositional details described in complex prompts. Reward-based alignment has emerged as a promising direction [6, 7, 12, 14, 26, 31, 32], using human preference models [12, 29, 32] to guide generation. The human preference models were trained on paired human preference data. Early works [12, 32] fine-tune diffusion models with the reward models, whereas later methods [6, 14] sidestep costly fine-tuning by directly optimizing the latent noise [6, 31]. However, these techniques largely improve global prompt adherence while neglecting spatial control. Our model not only aligns the T2I genera" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "tive models with the text prompt, but also ensures the spatial information to respect to user-provided sketches." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 99, + 553, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 99, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 99, + 553, + 374 + ], + "type": "text", + "content": "Controllable Image Synthesis T2I models inherently rely on highly compressed textual tokens, which limits their controllability and often prevents them from meeting user expectations [20, 22, 23]. To overcome these limitations, a growing body of research has explored controllable generation by incorporating fine-grained conditioning signals, such as reference images [2], edge maps or contours or skeletons [33], semantic layouts [1, 3, 38], and lighting specifications [34]. One line of work fine-tunes pretrained diffusion models by adding trainable modules [19, 33, 37], while others directly retrain the diffusion backbone by minimizing reconstruction objectives [2]. In parallel, training-free techniques [5, 7, 9, 26] have been developed to enhance controllability without modifying model weights. These include attention injection or latent optimization strategies, such as Prompt-to-Prompt editing [9] and diffusion self-guidance [5]. Despite these advances, existing methods often fail in complex scenes involving multiple guidance given by freehand sketches, where interactions between elements lead to interference and degraded synthesis quality. Building on these insights, our method incorporates attention-based guidance extraction but extends it to better handle multi-concept freehand sketches." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 377, + 553, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 553, + 617 + ], + "type": "text", + "content": "Sketch-to-Image Synthesis It is the task most relevant to our concept of the subjective camera. It aims to generate images from freehand sketches along with textual prompts. Several recent efforts have attempted to adapt T2I diffusion models for sketch-to-image synthesis [4, 17, 25, 28, 35, 36]. Training-based approaches [25, 36] fine-tune the T2I models with paired data. However, this training-based method incurs substantial computational cost and suffers from poor generalization. In contrast, training-free approaches [4, 17, 35] avoid fine-tuning by guiding the diffusion denoising process with sketches. However, these methods still jointly create multiple concepts, causing interference between their differing levels of randomness and abstraction. More critically, both training-based and training-free approaches fail to address the subjective biases inherent in user-provided sketches and are particularly vulnerable to highly abstract inputs, often leading to appearance distortions and semantic inconsistencies in the generated outputs. Furthermore, unlike creative generation methods, our focus is on reconstructing real-world scenes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 628, + 542, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 542, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 542, + 641 + ], + "type": "text", + "content": "3. Sequence-Aware Sketch-Guided Diffusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 647, + 382, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 382, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 382, + 659 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "Our goal is to render photorealistic images from subjective readouts, consisting of a textual description " + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": " and a sequence " + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "inline_equation", + "content": "S = \\{s_i\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": " freehand sketches, where each sketch represents a distinct concept. As illustrated" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17840" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 72, + 553, + 210 + ], + "blocks": [ + { + "bbox": [ + 58, + 72, + 553, + 210 + ], + "lines": [ + { + "bbox": [ + 58, + 72, + 553, + 210 + ], + "spans": [ + { + "bbox": [ + 58, + 72, + 553, + 210 + ], + "type": "image", + "image_path": "e491c29db16566f6abc2ac861ee480133a7e2c9045a62d4a5b2605baff667363.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 220, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 54, + 220, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 54, + 220, + 555, + 277 + ], + "type": "text", + "content": "Figure 3. Sequence-aware sketch-guided diffusion. (Left) User-provided textual description undergoes text-to-image generation and text-reward optimization to obtain the initial latent. Then, the following generation progresses under three constraints — spatial layout conforming to sketch topology while appearance details adhere to the initial image from text-reward optimization, with consistency maintained through ordered latent propagation and three loss terms, ensuring coherent integration of emerging and established scene elements. (Right) The concept-wise sequential optimization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 297, + 296, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 296, + 524 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 296, + 524 + ], + "type": "text", + "content": "in Figure 3, we design a Sequence-Aware Sketch-Guided Diffusion framework to accomplish this challenging task. Our method processes the subjective readouts individually and sequentially, following the order in which users describe them. It employs a sequence-aware optimization strategy to align a pre-trained generative diffusion model, " + }, + { + "bbox": [ + 56, + 297, + 296, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{G} := \\epsilon_{\\theta}(\\epsilon_{\\theta}(\\dots \\epsilon_{\\theta}(\\epsilon ,p)))" + }, + { + "bbox": [ + 56, + 297, + 296, + 524 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 56, + 297, + 296, + 524 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 56, + 297, + 296, + 524 + ], + "type": "text", + "content": " denotes the pre-trained U-Net, with these readouts, naturally mirroring the human cognitive process of step-by-step scene construction. The process begins with a text-reward optimization, which establishes an initial latent representation and appearance prior for subsequent concept-wise optimizations. The appearance prior is extracted via style encoding, while sketch encoding, implemented using diffusion inversion, provides spatial guidance for individual sketches. In addition, a preserving loss is introduced to maintain previously optimized concepts throughout the sequential process. These components collectively enable our framework to decode subjective readouts into coherent, photorealistic scenes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 531, + 205, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 205, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 205, + 544 + ], + "type": "text", + "content": "3.2. Text-Reward Optimization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 548, + 296, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 296, + 657 + ], + "type": "text", + "content": "We randomly sample a latent noise " + }, + { + "bbox": [ + 55, + 548, + 296, + 657 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 548, + 296, + 657 + ], + "type": "text", + "content": " and feed it, together with the textual description " + }, + { + "bbox": [ + 55, + 548, + 296, + 657 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 548, + 296, + 657 + ], + "type": "text", + "content": ", into a pre-trained T2I diffusion model, which renders an image through the full reverse diffusion process, " + }, + { + "bbox": [ + 55, + 548, + 296, + 657 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\theta}" + }, + { + "bbox": [ + 55, + 548, + 296, + 657 + ], + "type": "text", + "content": ". To ensure that the generated image contains the concepts in the given textual information, without concept omission or significant misalignment, we formulate a text-reward optimization procedure as existing practice [6, 14] that iteratively refines the latent noise to maximize a text-image alignment objective:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 108, + 663, + 295, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 663, + 295, + 682 + ], + "spans": [ + { + "bbox": [ + 108, + 663, + 295, + 682 + ], + "type": "interline_equation", + "content": "\\epsilon^ {0 \\star} = \\underset {\\epsilon} {\\arg \\max } \\mathcal {C} _ {F} \\left(\\mathcal {G} _ {\\theta} (\\epsilon , p), p\\right), \\tag {1}", + "image_path": "6f7791998e2e94c3bf9feaf0837ce0e4630a06b54d9f77c94f5052c1e518e740.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_F" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " represents a differentiable reward model that measures alignment between the generated image and the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 297, + 555, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 321 + ], + "type": "text", + "content": "text prompt " + }, + { + "bbox": [ + 313, + 297, + 555, + 321 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 297, + 555, + 321 + ], + "type": "text", + "content": ". The optimization process iteratively refines " + }, + { + "bbox": [ + 313, + 297, + 555, + 321 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 297, + 555, + 321 + ], + "type": "text", + "content": " through gradient ascent:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 369, + 331, + 555, + 344 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 331, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 369, + 331, + 555, + 344 + ], + "type": "interline_equation", + "content": "\\epsilon_ {t + 1} = \\epsilon_ {t} + \\eta \\nabla_ {\\epsilon} \\mathcal {C} _ {F} (\\mathcal {G} _ {\\theta} (\\epsilon_ {t}, p)), \\tag {2}", + "image_path": "48777601e6d51f1eb443dea88ec8af688d0e094515edbb186608bcedcc1b8112.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 353, + 429, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 429, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 429, + 365 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 353, + 429, + 365 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 353, + 429, + 365 + ], + "type": "text", + "content": " is the learning rate." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 372, + 416, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 416, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 416, + 385 + ], + "type": "text", + "content": "3.3. Sketch Encoding" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 390, + 555, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 555, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 555, + 556 + ], + "type": "text", + "content": "Before proceeding to the sketch-guided optimization, we first encode the sketch to extract high-level semantic guidance, rather than relying on low-level line details. Given a list of input sketches " + }, + { + "bbox": [ + 313, + 390, + 555, + 556 + ], + "type": "inline_equation", + "content": "\\{s_i\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 390, + 555, + 556 + ], + "type": "text", + "content": ", each sketch " + }, + { + "bbox": [ + 313, + 390, + 555, + 556 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 313, + 390, + 555, + 556 + ], + "type": "text", + "content": " is encoded into a latent representation " + }, + { + "bbox": [ + 313, + 390, + 555, + 556 + ], + "type": "inline_equation", + "content": "z_i \\in \\mathbb{R}^{h \\times w \\times d}" + }, + { + "bbox": [ + 313, + 390, + 555, + 556 + ], + "type": "text", + "content": " using the variational autoencoder (VAE) encoder " + }, + { + "bbox": [ + 313, + 390, + 555, + 556 + ], + "type": "inline_equation", + "content": "E_{\\phi}" + }, + { + "bbox": [ + 313, + 390, + 555, + 556 + ], + "type": "text", + "content": " of the diffusion model: " + }, + { + "bbox": [ + 313, + 390, + 555, + 556 + ], + "type": "inline_equation", + "content": "z_i = E_{\\phi}(s_i)" + }, + { + "bbox": [ + 313, + 390, + 555, + 556 + ], + "type": "text", + "content": ". Then, we can leverage arbitrary diffusion inversion algorithms, such as DDIM inversion [18] or SDEdit [16], to map the sketch into latent noise. To address the significant domain gap between abstract planar sketches and realistic images, we introduce an optimization procedure that iteratively refines the latent representation " + }, + { + "bbox": [ + 313, + 390, + 555, + 556 + ], + "type": "inline_equation", + "content": "z_i" + }, + { + "bbox": [ + 313, + 390, + 555, + 556 + ], + "type": "text", + "content": " by aligning it with the rich prior knowledge embedded in the diffusion model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 557, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 594 + ], + "type": "text", + "content": "At each optimization step " + }, + { + "bbox": [ + 313, + 557, + 555, + 594 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 557, + 555, + 594 + ], + "type": "text", + "content": ", we perturb the latent sketch representation with a Gaussian noise " + }, + { + "bbox": [ + 313, + 557, + 555, + 594 + ], + "type": "inline_equation", + "content": "\\epsilon_{k} \\sim \\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 313, + 557, + 555, + 594 + ], + "type": "text", + "content": " to obtain a noisy latent representation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 383, + 601, + 555, + 618 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 601, + 555, + 618 + ], + "spans": [ + { + "bbox": [ + 383, + 601, + 555, + 618 + ], + "type": "interline_equation", + "content": "z _ {t} ^ {(k)} = \\alpha_ {t _ {k}} z _ {i} ^ {(k)} + \\sigma_ {t _ {k}} \\epsilon_ {k}, \\tag {3}", + "image_path": "2629d0ec94c9705a0bc94d1725ef88590a3e244037591a9c8648fd4fec8ddaf5.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 626, + 554, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 626, + 554, + 690 + ], + "spans": [ + { + "bbox": [ + 313, + 626, + 554, + 690 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 626, + 554, + 690 + ], + "type": "inline_equation", + "content": "\\alpha_{t_k}" + }, + { + "bbox": [ + 313, + 626, + 554, + 690 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 626, + 554, + 690 + ], + "type": "inline_equation", + "content": "\\sigma_{t_k}" + }, + { + "bbox": [ + 313, + 626, + 554, + 690 + ], + "type": "text", + "content": " are time-dependent scaling factors following the noise schedule of the diffusion process. The noisy latent " + }, + { + "bbox": [ + 313, + 626, + 554, + 690 + ], + "type": "inline_equation", + "content": "z_{t}^{(k)}" + }, + { + "bbox": [ + 313, + 626, + 554, + 690 + ], + "type": "text", + "content": " is then fed into the pre-trained U-Net " + }, + { + "bbox": [ + 313, + 626, + 554, + 690 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 626, + 554, + 690 + ], + "type": "text", + "content": " which predicts the noise component " + }, + { + "bbox": [ + 313, + 626, + 554, + 690 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\theta}(z_{t}^{(k)}, t_{k})" + }, + { + "bbox": [ + 313, + 626, + 554, + 690 + ], + "type": "text", + "content": ". The noise prediction error " + }, + { + "bbox": [ + 313, + 626, + 554, + 690 + ], + "type": "inline_equation", + "content": "\\Delta \\epsilon_{k}" + }, + { + "bbox": [ + 313, + 626, + 554, + 690 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 382, + 699, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 699, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 382, + 699, + 555, + 715 + ], + "type": "interline_equation", + "content": "\\Delta \\epsilon_ {k} = \\hat {\\epsilon} _ {\\theta} \\left(z _ {t} ^ {(k)}, t _ {k}\\right) - \\epsilon_ {k}. \\tag {4}", + "image_path": "46367193ce8245b616e0ff1cfa69c3355dc75d33a6338a663280580bcb31c7e3.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17841" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 70, + 281, + 139 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 281, + 139 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 281, + 139 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 281, + 139 + ], + "type": "image", + "image_path": "ba965996a6b33632add424df6e05c58ac886742064647239aecf7781edf65720.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 148, + 296, + 217 + ], + "lines": [ + { + "bbox": [ + 55, + 148, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 148, + 296, + 217 + ], + "type": "text", + "content": "Figure 4. The effectiveness of sketch optimization. Without optimization (left), the generated image fails to accurately respect the wheel placement and proportions indicated in the sketch. With our optimization (right), the generated result better aligns with the sketch structure, particularly in the wheel positions (green arrows) and overall shape, demonstrating improved spatial coherence." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 233, + 296, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 233, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 233, + 296, + 270 + ], + "type": "text", + "content": "This noise gap serves as a gradient signal to iteratively refine the latent sketch features. Specifically, we update the latent representation " + }, + { + "bbox": [ + 55, + 233, + 296, + 270 + ], + "type": "inline_equation", + "content": "z_{i}^{(k)}" + }, + { + "bbox": [ + 55, + 233, + 296, + 270 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 275, + 295, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 275, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 82, + 275, + 295, + 297 + ], + "type": "interline_equation", + "content": "z _ {i} ^ {(k + 1)} = z _ {i} ^ {(k)} - \\eta_ {k} \\left(\\Delta \\epsilon_ {k} + \\lambda \\left(z _ {i} ^ {(k)} - z _ {i} ^ {(0)}\\right)\\right), \\tag {5}", + "image_path": "288aaf45a22e0b3f4cd968dd7452a740fe18b7891f9f65bfca9e21160ca3573d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 300, + 295, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 295, + 326 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 300, + 295, + 326 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 300, + 295, + 326 + ], + "type": "text", + "content": " is a regularization term that anchors the optimization to the initial latent " + }, + { + "bbox": [ + 55, + 300, + 295, + 326 + ], + "type": "inline_equation", + "content": "z_{i}^{(0)}" + }, + { + "bbox": [ + 55, + 300, + 295, + 326 + ], + "type": "text", + "content": ", preventing semantic drift." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 326, + 296, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 296, + 409 + ], + "type": "text", + "content": "Through this iterative optimization in Eq. (3-5), the latent " + }, + { + "bbox": [ + 55, + 326, + 296, + 409 + ], + "type": "inline_equation", + "content": "\\dot{z}_i" + }, + { + "bbox": [ + 55, + 326, + 296, + 409 + ], + "type": "text", + "content": " is pushed to align with the rich prior knowledge embedded in the diffusion model. This alignment effectively infuses the sketch with realistic details, textures, and structural coherence, thereby establishing a robust mapping from human hand-drawn sketches to physically grounded visual representations (see Figure 4)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 410, + 296, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 296, + 494 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 296, + 494 + ], + "type": "text", + "content": "However, we observed that the optimization method is less effective when the sketch is drawn relatively small. In such cases, the generated result often fails to align with the sketch and instead produces a shape-independent object in a central position. To address this, we suggest scaling the original sketch using linear interpolation or, alternatively, discarding the optimization for that step." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 499, + 183, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 499, + 183, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 499, + 183, + 513 + ], + "type": "text", + "content": "3.4. Appearance Encoding" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 516, + 296, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 296, + 566 + ], + "type": "text", + "content": "While the sketch provides spatial information, additional appearance cues are required to ensure visual quality during optimization. We first define a operator " + }, + { + "bbox": [ + 55, + 516, + 296, + 566 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(\\cdot)" + }, + { + "bbox": [ + 55, + 516, + 296, + 566 + ], + "type": "text", + "content": " to obtain the appearance descriptor:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 72, + 569, + 295, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 569, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 72, + 569, + 295, + 602 + ], + "type": "interline_equation", + "content": "\\mathcal {A} (x) = \\left\\{\\frac {\\sum_ {u , v} \\operatorname {s i g m o i d} ([ x ] _ {u v}) [ f _ {x} ] _ {u v}}{\\sum_ {u , v} \\operatorname {s i g m o i d} ([ x ] _ {u v})} \\mid f _ {x} \\right\\}, \\tag {6}", + "image_path": "e5a6e8f154497da29698791c2a8be2594d9f6c685f8651d90c4951e977186605.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 605, + 296, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 678 + ], + "type": "text", + "content": "which is a collection of weighted spatial means of diffusion features " + }, + { + "bbox": [ + 55, + 605, + 296, + 678 + ], + "type": "inline_equation", + "content": "f_{x}" + }, + { + "bbox": [ + 55, + 605, + 296, + 678 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 605, + 296, + 678 + ], + "type": "inline_equation", + "content": "f_{x}" + }, + { + "bbox": [ + 55, + 605, + 296, + 678 + ], + "type": "text", + "content": " are from different layer and time steps. Then, we treat the descriptor extracted from the optimized latent " + }, + { + "bbox": [ + 55, + 605, + 296, + 678 + ], + "type": "inline_equation", + "content": "z^{0\\star}" + }, + { + "bbox": [ + 55, + 605, + 296, + 678 + ], + "type": "text", + "content": ", derived from " + }, + { + "bbox": [ + 55, + 605, + 296, + 678 + ], + "type": "inline_equation", + "content": "\\epsilon^{0\\star}" + }, + { + "bbox": [ + 55, + 605, + 296, + 678 + ], + "type": "text", + "content": " in Eq. (1), as our appearance prior. We will compare the appearance descriptor of the current latent and the appearance prior during optimization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 72, + 517, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 517, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 517, + 85 + ], + "type": "text", + "content": "3.5. Concept-wise Sequential Optimization" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 89, + 554, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 554, + 137 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 554, + 137 + ], + "type": "text", + "content": "Next, we leverage this order information of subjective readouts to progressively construct the scene while maintaining coherence between newly introduced and previously established elements." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 140, + 554, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 140, + 554, + 200 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 554, + 200 + ], + "type": "text", + "content": "Spatial Attention-guided Control We leverage the spatial attention mechanisms within the diffusion model's U-Net architecture. For each sketch concept " + }, + { + "bbox": [ + 313, + 140, + 554, + 200 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 313, + 140, + 554, + 200 + ], + "type": "text", + "content": ", we extract cross-attention maps that highlight regions in the latent space corresponding to specific semantic concepts:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 351, + 205, + 555, + 233 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 205, + 555, + 233 + ], + "spans": [ + { + "bbox": [ + 351, + 205, + 555, + 233 + ], + "type": "interline_equation", + "content": "\\mathcal {M} (x) = \\operatorname {s o f t m a x} \\left(\\frac {Q (x) K (x) ^ {T}}{\\sqrt {d}}\\right) > \\tau , \\tag {7}", + "image_path": "53eb5d7bad08783404ec03e3925ad3567bc4742b3e0af0afc810496ac506634d.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 238, + 554, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 238, + 554, + 310 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 554, + 310 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 238, + 554, + 310 + ], + "type": "inline_equation", + "content": "Q(x)" + }, + { + "bbox": [ + 313, + 238, + 554, + 310 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 238, + 554, + 310 + ], + "type": "inline_equation", + "content": "K(x)" + }, + { + "bbox": [ + 313, + 238, + 554, + 310 + ], + "type": "text", + "content": " are query and key matrices, " + }, + { + "bbox": [ + 313, + 238, + 554, + 310 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 238, + 554, + 310 + ], + "type": "text", + "content": " is the dimensionality of the key vectors, and " + }, + { + "bbox": [ + 313, + 238, + 554, + 310 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 238, + 554, + 310 + ], + "type": "text", + "content": " is a threshold parameter. The attention maps " + }, + { + "bbox": [ + 313, + 238, + 554, + 310 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(\\dot{z}_i)" + }, + { + "bbox": [ + 313, + 238, + 554, + 310 + ], + "type": "text", + "content": " extracted from " + }, + { + "bbox": [ + 313, + 238, + 554, + 310 + ], + "type": "inline_equation", + "content": "\\dot{z}_i" + }, + { + "bbox": [ + 313, + 238, + 554, + 310 + ], + "type": "text", + "content": ", the latent of the sketch " + }, + { + "bbox": [ + 313, + 238, + 554, + 310 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 313, + 238, + 554, + 310 + ], + "type": "text", + "content": " (Section 3.3), serve as spatial indicators to guide the concept's layout. The control is given through the spatial loss, defined as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 324, + 316, + 554, + 347 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 316, + 554, + 347 + ], + "spans": [ + { + "bbox": [ + 324, + 316, + 554, + 347 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s p a t i a l}} \\left(z _ {i}, \\dot {z} _ {i}\\right) = \\frac {\\sum_ {u , v} \\left[ \\mathcal {M} \\left(\\dot {z} _ {i}\\right) \\right] _ {u v} \\left\\| \\left[ \\dot {z} _ {i} \\right] _ {u v} - \\left[ z _ {i} \\right] _ {u v} \\right\\| _ {2} ^ {2}}{\\sum_ {u , v} \\left[ \\mathcal {M} \\left(\\dot {z} _ {i}\\right) \\right] _ {u v}}, \\tag {8}", + "image_path": "525c971ea519711d91d770431f67e4d309c578d06341a8c91c17542941f516cc.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 352, + 476, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 352, + 476, + 365 + ], + "spans": [ + { + "bbox": [ + 314, + 352, + 476, + 365 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 314, + 352, + 476, + 365 + ], + "type": "inline_equation", + "content": "u, v" + }, + { + "bbox": [ + 314, + 352, + 476, + 365 + ], + "type": "text", + "content": " are spatial positions indices." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 367, + 542, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 367, + 542, + 380 + ], + "spans": [ + { + "bbox": [ + 314, + 367, + 542, + 380 + ], + "type": "text", + "content": "Appearance Control We define the appearance loss as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 358, + 384, + 554, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 384, + 554, + 403 + ], + "spans": [ + { + "bbox": [ + 358, + 384, + 554, + 403 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {a p p}} \\left(z _ {i}, z ^ {0 \\star}\\right) = \\left\\| \\mathcal {A} \\left(z _ {i}\\right) - \\mathcal {A} \\left(z ^ {0 \\star}\\right) \\right\\| _ {2} ^ {2}, \\tag {9}", + "image_path": "bc661055e23441ea225cddf84f7307695e1a4a8951bdae512219a88744de482f.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 407, + 554, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 554, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 554, + 420 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 407, + 554, + 420 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 407, + 554, + 420 + ], + "type": "text", + "content": " is the operator of exacting appearance descriptors." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 422, + 554, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 554, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 554, + 482 + ], + "type": "text", + "content": "Sequential Concept Integration To maintain coherence across the progressive addition of concepts, we implement a sequential integration process that updates the latent noise representation while preserving previously incorporated elements. For each sketch " + }, + { + "bbox": [ + 313, + 422, + 554, + 482 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 313, + 422, + 554, + 482 + ], + "type": "text", + "content": " in the sequence, we compute:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 397, + 489, + 553, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 489, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 397, + 489, + 553, + 502 + ], + "type": "interline_equation", + "content": "z _ {i} = z _ {i - 1} + \\Delta z _ {i}, \\tag {10}", + "image_path": "11f3c91aeeb6a03e1195badd778095c78b4274605ebdf1e90115e448ac0eb054.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 508, + 554, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 554, + 555 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 508, + 554, + 555 + ], + "type": "inline_equation", + "content": "z_{i-1}" + }, + { + "bbox": [ + 313, + 508, + 554, + 555 + ], + "type": "text", + "content": " is the optimized latent from the previous concept, and " + }, + { + "bbox": [ + 313, + 508, + 554, + 555 + ], + "type": "inline_equation", + "content": "\\Delta z_i" + }, + { + "bbox": [ + 313, + 508, + 554, + 555 + ], + "type": "text", + "content": " represents the incremental update required to incorporate the current sketch concept " + }, + { + "bbox": [ + 313, + 508, + 554, + 555 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 313, + 508, + 554, + 555 + ], + "type": "text", + "content": ". This update is derived from our loss function:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 562, + 553, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 562, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 320, + 562, + 553, + 576 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i} = \\alpha_ {i} \\mathcal {L} _ {\\text {n e w}} \\left(z _ {i}, \\dot {z} _ {i}, z ^ {0 \\star}\\right) + \\beta_ {i} \\mathcal {L} _ {\\text {p r e s e r v e}} \\left(z _ {i}, \\dot {z} _ {0, \\dots , i - 1}\\right), \\tag {11}", + "image_path": "447057eab7a9c4470c1587f9ccf14384f1f9604218ea612ace7f783ce29a510a.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 582, + 554, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 606 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 582, + 554, + 606 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 313, + 582, + 554, + 606 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 582, + 554, + 606 + ], + "type": "inline_equation", + "content": "\\beta_{i}" + }, + { + "bbox": [ + 313, + 582, + 554, + 606 + ], + "type": "text", + "content": " are balancing coefficients. " + }, + { + "bbox": [ + 313, + 582, + 554, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{new}}" + }, + { + "bbox": [ + 313, + 582, + 554, + 606 + ], + "type": "text", + "content": " ensures proper integration of the current " + }, + { + "bbox": [ + 313, + 582, + 554, + 606 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 313, + 582, + 554, + 606 + ], + "type": "text", + "content": " and is defined as:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 324, + 612, + 553, + 627 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 612, + 553, + 627 + ], + "spans": [ + { + "bbox": [ + 324, + 612, + 553, + 627 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {n e w}} \\left(z _ {i}, \\dot {z} _ {i}, z ^ {0 \\star}\\right) = \\mathcal {L} _ {\\text {s p a t i a l}} \\left(z _ {i}, \\dot {z} _ {i}\\right) + \\mathcal {L} _ {\\text {a p p}} \\left(z _ {i}, z ^ {0 \\star}\\right), \\tag {12}", + "image_path": "354dcb7fdeea7c1668f26d92eb04fb11bc1b5df72ff603b477b755c6a69eb945.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 632, + 553, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 632, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 313, + 632, + 553, + 657 + ], + "type": "text", + "content": "while " + }, + { + "bbox": [ + 313, + 632, + 553, + 657 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{preserve}}" + }, + { + "bbox": [ + 313, + 632, + 553, + 657 + ], + "type": "text", + "content": " maintains the integrity of previously optimized concepts, given by:" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 328, + 662, + 553, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 662, + 553, + 696 + ], + "spans": [ + { + "bbox": [ + 328, + 662, + 553, + 696 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p r e s e r v e}} \\left(z _ {i}, \\dot {z} _ {0, \\dots , i - 1}\\right) = \\sum_ {j = 1} ^ {i - 1} \\gamma^ {i - j} \\mathcal {L} _ {\\text {s p a t i a l}} \\left(z _ {i}, \\dot {z} _ {j}\\right), \\tag {13}", + "image_path": "f55318c84dc18312d9cc34a5da174d1855373cc49ae36ad53a3b3720f3069b27.jpg" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 683, + 296, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 296, + 714 + ], + "type": "text", + "content": "3To avoid ambiguity, we use " + }, + { + "bbox": [ + 55, + 683, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\dot{z}_i" + }, + { + "bbox": [ + 55, + 683, + 296, + 714 + ], + "type": "text", + "content": " to denote the optimized latent for the " + }, + { + "bbox": [ + 55, + 683, + 296, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 683, + 296, + 714 + ], + "type": "text", + "content": "-th sketch concept, distinguishing it from latent " + }, + { + "bbox": [ + 55, + 683, + 296, + 714 + ], + "type": "inline_equation", + "content": "z_i" + }, + { + "bbox": [ + 55, + 683, + 296, + 714 + ], + "type": "text", + "content": " would be used in the concept-wise sequential optimization." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 324, + 702, + 537, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 702, + 537, + 713 + ], + "spans": [ + { + "bbox": [ + 324, + 702, + 537, + 713 + ], + "type": "text", + "content": "4In fact, an inner-loop optimization is omitted here for simplicity." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17842" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 550, + 369 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 550, + 369 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 550, + 369 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 550, + 369 + ], + "type": "image", + "image_path": "44bc7404717272fe3bbed7925a95892c497e415a622eefe6a972f57bdf68e924.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 380, + 555, + 424 + ], + "lines": [ + { + "bbox": [ + 55, + 380, + 555, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 555, + 424 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison of existing sketch-to-image generation methods, and our method on FMC and CMC datasets. Existing training-based methods tend to overfit to sketch lines, often failing to reconstruct scenes realistically, while training-free methods frequently misinterpret or overlook key concepts. In comparison, our approach more faithfully represents real-world physical scenes. The reading and processing order of the sketches is indicated by colors as follows: " + }, + { + "bbox": [ + 55, + 380, + 555, + 424 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 55, + 380, + 555, + 424 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 55, + 380, + 555, + 424 + ], + "type": "inline_equation", + "content": "\\bullet" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 445, + 295, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 295, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 295, + 482 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 445, + 295, + 482 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 445, + 295, + 482 + ], + "type": "text", + "content": " means preserving established concepts, and " + }, + { + "bbox": [ + 55, + 445, + 295, + 482 + ], + "type": "inline_equation", + "content": "\\gamma^{i - j}" + }, + { + "bbox": [ + 55, + 445, + 295, + 482 + ], + "type": "text", + "content": " is a decay factor that weights the importance of earlier concepts based on their step gap from the current step." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 493, + 198, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 198, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 198, + 506 + ], + "type": "text", + "content": "4. Experiments and Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 514, + 163, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 163, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 163, + 526 + ], + "type": "text", + "content": "4.1. Experiment Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": "Datasets To rigorously evaluate our method, we have undertaken extensive efforts to develop two specialized datasets: the Composed Multi-Concept (CMC) dataset and the Freehand Multi-Concept (FMC) dataset. Both datasets are designed to comprehensively assess the capabilities of our framework in handling multi-concept scene understanding and generation. CMC dataset comprises 142 scene sketches, each containing 2-4 distinct concepts sampled from the 150 semantic classes in the Sketchy dataset [24]. Each is accompanied by annotated textual prompts with the help of multimodal large language models. The CMC dataset does not provide real-world images captured by physical cameras for reference. FMC dataset is constructed from freehand sketches drawn by volunteers, based on real-world images sourced from publicly available outdoor" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 445, + 555, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 494 + ], + "type": "text", + "content": "scene datasets. It comprises 42 hand-drawn sketches, each containing 2-3 distinct concepts. Additionally, each sketch is paired with a corresponding real-world image as a reference for validation and evaluation purposes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "type": "text", + "content": "Evaluation Metrics To conduct a comprehensive quantitative evaluation of the quality of generated images, four widely-used evaluation metrics were employed, including Frechet Inception Distance (FID) [10], CLIP-T Score (CLIP-T) [21], Human Preference Score (HPS) [29], and CLIP-I distance (CLIP-I) [21]. FID measures the feature distribution similarity between generated and real images. Lower FID values indicate better distribution matching " + }, + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "type": "inline_equation", + "content": "(\\downarrow)" + }, + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "type": "text", + "content": ". CLIP-T quantifies the image-text alignment degree within the CLIP-T embedding space. A higher CLIP-T score reflects greater semantic alignment between the generated image and text prompt " + }, + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "type": "inline_equation", + "content": "(\\uparrow)" + }, + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "type": "text", + "content": ". HPS measures the subjective quality of the generated images based on human evaluations. Higher HPS values suggest that the generated images are more favorable or preferred by human evaluators " + }, + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "type": "inline_equation", + "content": "(\\uparrow)" + }, + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "type": "text", + "content": ". CLIP-I measures the alignment between two images. A higher value suggests that the two images are more consistent in terms of their semantic content " + }, + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "type": "inline_equation", + "content": "(\\uparrow)" + }, + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17843" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 66, + 527, + 186 + ], + "blocks": [ + { + "bbox": [ + 80, + 66, + 527, + 186 + ], + "lines": [ + { + "bbox": [ + 80, + 66, + 527, + 186 + ], + "spans": [ + { + "bbox": [ + 80, + 66, + 527, + 186 + ], + "type": "image", + "image_path": "d895d59adab5fa97f378c744bd46e240af54896a72d2c4e7bd4b9c5f38a90c6f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 200, + 555, + 267 + ], + "lines": [ + { + "bbox": [ + 54, + 200, + 555, + 267 + ], + "spans": [ + { + "bbox": [ + 54, + 200, + 555, + 267 + ], + "type": "text", + "content": "Figure 6. Ablation study on FMC and CMC datasets. Column headers are defined as follows: T2I denotes the baseline text-to-image diffusion model; OP (Overall Perception) indicates the single-step guidance approach utilizing complete sketch as a holistic conditioning signal; CSO refers to our proposed Concept-wise Sequential Optimization (Section 3.5); SE represents the Sketch Encoding module (Section 3.3) that enhances structural plausibility by injecting diffusion priors into imperfect sketches. The qualitative comparison demonstrates that CSO achieves finer layout control than OP. While SE effectively bridges the domain gap between freehand sketches and physical reality through latent space regularization, ultimately elevating generation quality." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 284, + 293, + 372 + ], + "blocks": [ + { + "bbox": [ + 56, + 284, + 293, + 372 + ], + "lines": [ + { + "bbox": [ + 56, + 284, + 293, + 372 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 293, + 372 + ], + "type": "table", + "html": "
Metric MethodCMCFMC
FID ↓CLIP-T ↑HPS ↑FID ↓CLIP-T ↑HPS ↑CLIP-I ↑
Training-based methods
ControlNet [33]15.2861.710.79414.4566.800.8020.622
T2I-Adapter [19]18.1361.610.74515.3761.610.7450.689
Uni-ControlNet [37]19.6062.580.76118.5362.580.7610.712
Training-free methods
Multi-Sketch [4]13.0564.110.74011.5972.020.7750.690
PNP [27]17.0163.270.84016.2671.240.8520.593
FreeControl [17]14.5264.740.75711.1766.670.7590.775
Ours11.5666.640.85010.1874.250.8620.797
", + "image_path": "4a7a71b0db0b6b949857a860536601479a106e5e76e23a85f2d5a475469d0e21.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 381, + 295, + 460 + ], + "lines": [ + { + "bbox": [ + 54, + 381, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 54, + 381, + 295, + 460 + ], + "type": "text", + "content": "Table 1. Quantitative results on CMC and FMC datasets. Our method consistently surpasses all training-free methods in distribution matching, image-text alignment, subjective quality, and appearance details as measured by FID, CLIP-T, HPS, and CLIP-I. Our method achieves superior image-text alignment compared to training-based methods (" + }, + { + "bbox": [ + 54, + 381, + 295, + 460 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 54, + 381, + 295, + 460 + ], + "type": "text", + "content": " indicates lower values are better, while " + }, + { + "bbox": [ + 54, + 381, + 295, + 460 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 54, + 381, + 295, + 460 + ], + "type": "text", + "content": " is opposite)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 481, + 296, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 481, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 54, + 481, + 296, + 566 + ], + "type": "text", + "content": "Implementation Details In the implementation of our proposed method, we set the learning rate " + }, + { + "bbox": [ + 54, + 481, + 296, + 566 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 54, + 481, + 296, + 566 + ], + "type": "text", + "content": " in Eq. (2) and the threshold parameter " + }, + { + "bbox": [ + 54, + 481, + 296, + 566 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 54, + 481, + 296, + 566 + ], + "type": "text", + "content": " in Eq. (7) is 5.0 and 0.3, respectively. Meanwhile, hyperparameters " + }, + { + "bbox": [ + 54, + 481, + 296, + 566 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 54, + 481, + 296, + 566 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 481, + 296, + 566 + ], + "type": "inline_equation", + "content": "\\beta_{i}" + }, + { + "bbox": [ + 54, + 481, + 296, + 566 + ], + "type": "text", + "content": " in Eq. (11) are set as 0.8 and 0.2, respectively. " + }, + { + "bbox": [ + 54, + 481, + 296, + 566 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 54, + 481, + 296, + 566 + ], + "type": "text", + "content": " in Eq. (13) is set as 0.9. Our method is performed on one NVIDIA GeForce RTX 4090, and input sketch images are " + }, + { + "bbox": [ + 54, + 481, + 296, + 566 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 54, + 481, + 296, + 566 + ], + "type": "text", + "content": " in resolution." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 574, + 141, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 141, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 141, + 588 + ], + "type": "text", + "content": "4.2. Comparisons" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "text", + "content": "We conducted a comprehensive quantitative and qualitative evaluation of the proposed method and competing methods on the CMC and FMC datasets. Our competing methods include the training-based approaches ControlNet [33], T2I-Adapter [19], and Uni-ControlNet [37], as well as the training-free approaches Multi-Sketch [4], PNP [27], and FreeControl [17]. The quantitative results are presented in Table 1 and qualitative results are shown in Figure 5." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 691, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 691, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 691, + 296, + 715 + ], + "type": "text", + "content": "We observe that training-based methods [19, 33, 37] are highly susceptible to subjective biases in sketches and of" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 288, + 555, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 555, + 419 + ], + "type": "text", + "content": "ten fail to adhere to text prompts, producing distorted and unrealistic images with lower CLIP-T scores. Training-free methods [4, 8, 17] suffer from concept omissions, conspicuous artifacts, and inconsistencies both among concepts and between concepts and their backgrounds, resulting in poor image realism. These issues stem from mutual interference during the \"snapshot\" generation of multiple concepts. In contrast, our proposed method faithfully reconstructs real-world scenes and achieves the best performance across all evaluation metrics on both the CMC and FMC datasets, demonstrating its effectiveness." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 427, + 553, + 492 + ], + "blocks": [ + { + "bbox": [ + 315, + 427, + 553, + 492 + ], + "lines": [ + { + "bbox": [ + 315, + 427, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 315, + 427, + 553, + 492 + ], + "type": "image", + "image_path": "bb3ee8fc8f5cbc78041af524b18487d10b76b38b4a5760795f2a00c1a23c44c8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 503, + 555, + 568 + ], + "lines": [ + { + "bbox": [ + 313, + 503, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 555, + 568 + ], + "type": "text", + "content": "Figure 7. User preference experiment. We asked participants to answer four questions based on the reference images, subjective readouts, and generated images, which included one produced by our method and one by a competing method. Results show that users significantly preferred our outputs over those of competing methods in most cases." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 588, + 408, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 408, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 408, + 601 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 605, + 554, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 642 + ], + "type": "text", + "content": "To validate the effectiveness of each module in our method, we conducted ablation experiments on the FMC and CMC datasets, with the visualization results shown in Figure 6." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 643, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 643, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 643, + 555, + 714 + ], + "type": "text", + "content": "It is observed that traditional T2I diffusion models (T2I) generate images that are close to text prompts. While incorporating sketches as guidance further constrains the consistency of concept positioning, deviations in appearance and spatial alignment persist (T2I, OP). The sequential concept understanding facilitated by Concept-wise Sequential" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "17844" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 294, + 148 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 148 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 148 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 148 + ], + "type": "image", + "image_path": "1725d49d543ff881a7f5b1e56687215f6fcc6a05644b4cbc5156804a37dadc4b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 156, + 295, + 212 + ], + "lines": [ + { + "bbox": [ + 55, + 156, + 295, + 212 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 295, + 212 + ], + "type": "text", + "content": "Figure 8. Control experiments on user preferences. We controlled three factors to examine their influence: whether the rater drew the photo, the availability timing of the reference image (" + }, + { + "bbox": [ + 55, + 156, + 295, + 212 + ], + "type": "inline_equation", + "content": "\\infty" + }, + { + "bbox": [ + 55, + 156, + 295, + 212 + ], + "type": "text", + "content": " vs. 5s), and the types of questions asked. In general, our method is consistently better than the competing method, FreeControl." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 232, + 296, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 296, + 342 + ], + "type": "text", + "content": "Optimization (CSO) in Section 3.5 enhances the alignment of concept positioning and appearance with input sketches (CSO). Simultaneously, the Sketch Encoding (SE) module in Section 3.3 aligns content latent variables with real-world objects, mitigating geometric misalignment (T2I, OP, SE). When integrating all modules, our method faithfully reconstructs real-world scenes and achieves optimal results (Ours), demonstrating the effectiveness of each module and underscoring the indispensable nature of their design." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 349, + 216, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 216, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 216, + 361 + ], + "type": "text", + "content": "4.4. User Preference Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 366, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 295, + 594 + ], + "type": "text", + "content": "We invited 12 participants (group A) to provide subjective readouts and an additional 28 participants (group B) to evaluate the reconstructed images. Participants in group A were first asked to select a deeply impressive scene from their photo gallery and then provide a textual description along with sketches within two minutes. Our method and competing approaches, including two baselines (text-reward optimization and text-reward optimization with sketch-guided optimization based on all concepts) as well as an optimal method, FreeControl [17], were used to generate images from these readouts. Participants in both groups were then shown the subjective readouts, the reference image, and the generated images for the same scene. They were asked to answer four single-choice questions in a single-blind manner: Q1: Which image is closer to the original image (both style and content)? Q2: Which image has higher visual quality (e.g., more realistic, more natural)? Q3: Which image better matches the structure of the sketch? Q4: Which image better matches the textual description?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "Figure 7 presents the statistics of the collected results. Participants judged our reconstructions as being closer to the corresponding real-world scenes compared to competing methods. Similar conclusions were observed for the other three questions. However, there was an exception in the comparison between our method and text-reward optimization regarding image quality (Q2). We observed that text-reward optimization without additional control tends to hallucinate details that enhance perceived realism, leading some users to prefer it for Q2. This also explains why our" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "content": "advantage over text-reward optimization in Q1 was not statistically significant and why some participants expressed hesitation in this case." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 108, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 555, + 300 + ], + "type": "text", + "content": "We further conducted controlled experiments to examine the effects of three factors: whether the rater provided the subjective readouts, the availability timing of the reference image, and the types of questions asked. Results in Figure 8 show that our method consistently outperformed FreeControl [17], regardless of these variations. Notably: 1) Participants who provided subjective readouts showed slightly lower preference for our method in terms of similarity to the reference image compared to those who did not, while showing a higher preference in terms of alignment with the readouts and perceived image quality. We suppose this is mainly because of the subjectivity in reading out. 2) Reducing the time participants could view the reference image decreased the preference for our method; when allowed to compare freely, participants preferred our results more strongly." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 311, + 391, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 391, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 391, + 323 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 331, + 553, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 553, + 521 + ], + "type": "text", + "content": "This paper presented Subjective Camera, a paradigm for reconstructing real-world scenes from humans' mental impressions. Our framework relies on the most accessible readouts of mental imagery, i.e., textual descriptions and sketches, and employs a training-free, sequence-aware optimization process. This approach eliminates the need for costly fine-tuning of text-to-image diffusion models on large-scale data while mitigating challenges such as dependence on sketch quality and geometric misalignment. Extensive experiments on both synthetic and real-world sketch datasets demonstrate that Subjective Camera achieves state-of-the-art performance, outperforming both training-free and training-based sketch-to-image generation methods across all evaluation metrics. Furthermore, user preference studies confirm that participants consistently favor our method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "content": "In closing, we address a natural question: \"With the rapid advancement of generative AI, will physical cameras ever become unnecessary\" For now, the answer is no. First, the reconstructed images still fall short of perfectly matching actual scenes. This limitation largely stems from the fact that current T2I models are trained on highly diverse data, creating gaps between the prior and the specific scene to be reconstructed. One promising direction to mitigate this is to personalize the prior models using user-specific galleries. Second, we observe that variations in the order and quality of sketches have a significant impact on reconstruction results. To address this, we plan to enhance our framework with computational tools that guide users in providing more precise and suitable subjective readouts. For these reasons, we designate this system as Subjective Camera 1.0, leaving these improvements for future iterations." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17845" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 157 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 157 + ], + "type": "text", + "content": "Acknowledgements This work was supported in part by the National Natural Science Foundation of China (62225113, 624B2109, 623B2079) and the Zhongguancun Academy Project (20240308). We thank Po-Hsun Huang for providing the questionnaire system for our user study, Kaijin Zheng for assistance with data processing, and all participants who took part in the studies." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 167, + 115, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 167, + 115, + 179 + ], + "spans": [ + { + "bbox": [ + 56, + 167, + 115, + 179 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 186, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 186, + 294, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 186, + 294, + 230 + ], + "spans": [ + { + "bbox": [ + 61, + 186, + 294, + 230 + ], + "type": "text", + "content": "[1] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. Proceedings of Machine Learning Research, 202:1737-1752, 2023. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 232, + 294, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 232, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 61, + 232, + 294, + 275 + ], + "type": "text", + "content": "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18392-18402, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 277, + 296, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 277, + 296, + 320 + ], + "spans": [ + { + "bbox": [ + 62, + 277, + 296, + 320 + ], + "type": "text", + "content": "[3] Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5343-5353, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 322, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 322, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 62, + 322, + 294, + 376 + ], + "type": "text", + "content": "[4] Zhenwei Cheng, Lei Wu, Changshuo Wang, and Xiangxu Meng. Scene sketch-to-image synthesis based on multi-object control. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3775-3779. IEEE, 2024. 2, 3, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 378, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 378, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 62, + 378, + 294, + 421 + ], + "type": "text", + "content": "[5] Dave Epstein, Allan Jabri, Ben Poole, Alexei Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. Advances in Neural Information Processing Systems, 36:16222-16239, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 423, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 423, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 62, + 423, + 294, + 476 + ], + "type": "text", + "content": "[6] Luca Eyring, Shyamgopal Karthik, Karsten Roth, Alexey Dosovitskiy, and Zeynep Akata. Reno: Enhancing one-step text-to-image models through reward-based noise optimization. Advances in Neural Information Processing Systems, 37:125487-125519, 2025. 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 479, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 479, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 479, + 294, + 533 + ], + "type": "text", + "content": "[7] Xiefan Guo, Jinlin Liu, Miaomiao Cui, Jiankai Li, Hongyu Yang, and Di Huang. Initno: Boosting text-to-image diffusion models via initial noise optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9380-9389, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 535, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 535, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 535, + 294, + 578 + ], + "type": "text", + "content": "[8] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In The Eleventh International Conference on Learning Representations. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 580, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 580, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 62, + 580, + 294, + 623 + ], + "type": "text", + "content": "[9] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control.(2022). URL https://arxiv.org/abs/2208.01626, 3, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 58, + 624, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 624, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 58, + 624, + 294, + 678 + ], + "type": "text", + "content": "[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "type": "text", + "content": "[11] Zechao Hu, Zhengwei Yang, Hao Li, Yixiong Zou, Fengbin Zhu, and Zheng Wang. Unified category and style generalization for instance-level sketch retrieval. In Proceedings of" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, page 864-873. Association for Computing Machinery, 2025. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 108, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 161 + ], + "type": "text", + "content": "[12] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Martiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36: 36652-36663, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 163, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 185 + ], + "type": "text", + "content": "[13] Stephen M Kosslyn. Aspects of a cognitive neuroscience of mental imagery. Science, 240(4859):1621-1626, 1988. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 186, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 252 + ], + "type": "text", + "content": "[14] Seung Hyun Lee, Yinxiao Li, Junjie Ke, Innfarn Yoo, Han Zhang, Jiahui Yu, Qifei Wang, Fei Deng, Glenn Entis, Junfeng He, et al. Parrot: Pareto-optimal multi-reward reinforcement learning framework for text-to-image generation. In European Conference on Computer Vision, pages 462-478. Springer, 2024. 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 254, + 555, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 555, + 308 + ], + "type": "text", + "content": "[15] Kejun Lin, Zhixiang Wang, Zheng Wang, Yinqiang Zheng, and Shin'ichi Satoh. Beyond domain gap: Exploiting subjectivity in sketch-based person retrieval. In Proceedings of the 31st ACM international conference on multimedia, pages 2078-2089, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 309, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 309, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 317, + 309, + 553, + 354 + ], + "type": "text", + "content": "[16] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 355, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 355, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 317, + 355, + 553, + 420 + ], + "type": "text", + "content": "[17] Sicheng Mo, Fangzhou Mu, Kuan Heng Lin, Yanli Liu, Bochen Guan, Yin Li, and Bolei Zhou. Freecontrol: Training-free spatial control of any text-to-image diffusion model with any condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7465-7475, 2024. 3, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 422, + 555, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 555, + 476 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 555, + 476 + ], + "type": "text", + "content": "[18] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6038–6047, 2023. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 478, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 553, + 533 + ], + "type": "text", + "content": "[19] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296-4304, 2024. 2, 3, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 534, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 534, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 317, + 534, + 553, + 577 + ], + "type": "text", + "content": "[20] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195-4205, 2023. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 579, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 579, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 317, + 579, + 553, + 643 + ], + "type": "text", + "content": "[21] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 646, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 553, + 690 + ], + "type": "text", + "content": "[22] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[23] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17846" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 148 + ], + "type": "text", + "content": "[24] Patsorn Sangkloy, Nathan Burnell, Cusuh Ham, and James Hays. The sketchy database: learning to retrieve badly drawn bunnies. ACM Transactions on Graphics (TOG), 35(4):1-12, 2016. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "type": "text", + "content": "[25] Vishnu Sarukkai, Lu Yuan, Mia Tang, Maneesh Agrawala, and Kayvon Fatahalian. Block and detail: Scaffolding sketch-to-image generation. In Proceedings of the 37th Annual ACM Symposium on User Interface Software and Technology, pages 1-13, 2024. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 259 + ], + "type": "text", + "content": "[26] Aravindan Sundaram, Ujjayan Pal, Abhimanyu Chauhan, Aishwarya Agarwal, and Srikrishna Karanam. Cocono: Attention contrast-and-complete for initial noise optimization in text-to-image synthesis. arXiv preprint arXiv:2411.16783, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "type": "text", + "content": "[27] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 358 + ], + "type": "text", + "content": "[28] Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. Sketch-guided text-to-image diffusion models. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 360, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 415 + ], + "type": "text", + "content": "[29] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score: Better aligning text-to-image models with human preference. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2096-2105, 2023. 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 415, + 294, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 469 + ], + "type": "text", + "content": "[30] Zhenbei Wu, Haoge Deng, Qiang Wang, Di Kong, Jie Yang, and Yonggang Qi. Sketchscene: Scene sketch to image generation with diffusion models. In 2023 IEEE International Conference on Multimedia and Expo (ICME), pages 2087-2092. IEEE, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "type": "text", + "content": "[31] Xin Xie and Dong Gong. Dymo: Training-free diffusion model alignment with dynamic multi-objective scheduling. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 13220-13230, 2025. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 514, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 294, + 569 + ], + "type": "text", + "content": "[32] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36:15903-15935, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 571, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 571, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 571, + 294, + 613 + ], + "type": "text", + "content": "[33] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 614, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 669 + ], + "type": "text", + "content": "[34] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Scaling in-the-wild training for diffusion-based illumination harmonization and editing by imposing consistent light transport. In The Thirteenth International Conference on Learning Representations, 2025. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[35] Tianyu Zhang and Haoran Xie. Sketch-guided text-to-image generation with spatial control. In 2024 2nd International Conference on Computer Graphics and Image Processing (CGIP), pages 153-159. IEEE, 2024. 2, 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 217 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "text", + "content": "[36] Tianyu Zhang, Xiaoxuan Xie, Xusheng Du, and Haoran Xie. Sketch-guided scene image generation. arXiv preprint arXiv:2407.06469, 2024. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "type": "text", + "content": "[37] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. Advances in Neural Information Processing Systems, 36:11127-11150, 2023. 2, 3, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "text", + "content": "[38] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22490-22499, 2023. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17847" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SummDiff_ Generative Modeling of Video Summarization with Diffusion/314d20e1-b991-4676-b961-e90cddf3f9d6_content_list.json b/2025/SummDiff_ Generative Modeling of Video Summarization with Diffusion/314d20e1-b991-4676-b961-e90cddf3f9d6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..548aabe1a278fbca3224fe072225db3c56482948 --- /dev/null +++ b/2025/SummDiff_ Generative Modeling of Video Summarization with Diffusion/314d20e1-b991-4676-b961-e90cddf3f9d6_content_list.json @@ -0,0 +1,1809 @@ +[ + { + "type": "text", + "text": "SummDiff: Generative Modeling of Video Summarization with Diffusion", + "text_level": 1, + "bbox": [ + 130, + 130, + 867, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kwanseok Kim $^{1*}$ , Jaehoon Hahm $^{2*}$ , Sumin Kim $^{1}$ , Jinhwan Sul $^{3}$ , Byunghak Kim $^{4}$ , Joonseok Lee $^{1\\dagger}$ , Seoul National University, 2UIUC, Georgia Institute of Technology, Hyundai Card", + "bbox": [ + 122, + 179, + 879, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "kjvd1009@snu.ac.kr, jh141@illinois.edu, sumink@snu.ac.kr, jsul7@gatech.edu, byunghak.kim@hcs.com, joonseok@snu.ac.kr", + "bbox": [ + 240, + 219, + 750, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video summarization is a task of shortening a video by choosing a subset of frames while preserving its essential moments. Despite the innate subjectivity of the task, previous works have deterministically regressed to an averaged frame score over multiple raters, ignoring the inherent subjectivity of what constitutes a \"good\" summary. We propose a novel problem formulation by framing video summarization as a conditional generation task, allowing a model to learn the distribution of good summaries and to generate multiple plausible summaries that better reflect varying human perspectives. Adopting diffusion models for the first time in video summarization, our proposed method, SummDiff, dynamically adapts to visual contexts and generates multiple candidate summaries conditioned on the input video. Extensive experiments demonstrate that SummDiff not only achieves the state-of-the-art performance on various benchmarks but also produces summaries that closely align with individual annotator preferences. Moreover, we provide a deeper insight with novel metrics from an analysis of the knapsack, which is an important last step of generating summaries but has been overlooked in evaluation.", + "bbox": [ + 88, + 319, + 485, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 667, + 220, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, short-form videos draw significant attention on video sharing platforms, with a trend that consumers increasingly prefer to quickly grasp the content. They often compressively convey contents that are originally in a longer form, summarizing the core contents into a shorter one; e.g., sport games highlights or movie summarization. This task of selecting core parts of a long video to construct a shorter one is called video summarization. This task is inherently subjective, since there can be multiple criteria for a 'good summary'; e.g., comprehensively covering the entire storyline or subjectively selecting impressive parts of", + "bbox": [ + 89, + 695, + 483, + 861 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the video (highlight detection). Due to this inherent subjectivity, most video summarization or highlight detection datasets [20, 71] offer annotations by multiple raters to reflect various perspectives.", + "bbox": [ + 511, + 287, + 906, + 348 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Since each annotator may have different opinion on the importance of a frame, most existing methods [3, 17, 22, 34, 67, 82] take the frame-level importance scores averaged across multiple annotators as their target label, and are trained to predict them. This frame-level score aggregation looks reasonable in some sense, but in fact it loses the various perspectives to summarize each video. For instance, suppose half of the annotators select clips from the first quarter of the video while the other half select clips from the last quarter. If one simply averages their frame scores, both the first and last quarters end up with similar importance, obscuring the two distinct valid summaries. That is, this simple regression to the averaged frame-level importance scores fail to preserve multiple viewpoints to summarize the video.", + "bbox": [ + 511, + 351, + 908, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In order to preserve and reflect various views to summarize a video, we pose a distribution of its good summaries and let the model to learn it, instead of giving an already-aggregated single ground truth importance score. Then, the video summarization task can be seen from a generative perspective; i.e., a process of learning and inferring the distribution of good summaries conditioned on the input video. Specifically, the model is now in charge of estimating the distribution of plausible summaries for the given video. Once trained, it allows us to sample multiple summaries from the estimated conditional data manifold. To the best of our knowledge, this generative approach to the video summarization has not been extensively studied, except for a few works [2, 50] that applied adversarial losses to construct a summary that looks like the original video.", + "bbox": [ + 511, + 580, + 908, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Formulating the video summarization problem as a conditional generation task, we propose to adopt the generative diffusion [25, 72] mechanism, which has been successfully applied to various conditional generation tasks [24, 54, 61]. Specifically, conditioned on the input video, our proposed SummDiff model learns to denoise a random importance", + "bbox": [ + 511, + 810, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 107, + 875, + 220, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 109, + 887, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15096", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b84fa467829d5ca6a0793d03fe6eec5da4442d119ef6b672ccfffe9d9e9054db.jpg", + "image_caption": [ + "Figure 1. Overview of SummDiff. Given an input video, SummDiff generates importance scores conditioned on video frames. $T$ denotes the number of DDIM steps." + ], + "image_footnote": [], + "bbox": [ + 93, + 85, + 614, + 247 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9f842a7c643f7740f4f686512caadea0c71a5be9b0b95427cdab875affe6f21b.jpg", + "image_caption": [ + "Figure 2. Video Importance Score Denoiser. We use AdaLN layer to inject time and positional conditions, following [58]." + ], + "image_footnote": [], + "bbox": [ + 627, + 70, + 901, + 237 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "score vector over the given video into an importance score vector sampled from the distribution that corresponds to a good summary of the video. In contrast to the previous deterministic methods, our approach allows to sample multiple plausible summaries for a given video starting from a different random noise, better aligned with the subjective nature of summarization where we usually have various true labels reflecting multiple views.", + "bbox": [ + 88, + 301, + 483, + 421 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments demonstrate that our model outperforms existing baselines across multiple datasets. Also, we revisit the knapsack, a relatively unexplored step in the summarization evaluation in spite of its nontrivial impact on the performance, and propose additional novel metrics based on this analysis to provide deeper insights.", + "bbox": [ + 89, + 422, + 483, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 109, + 513, + 434, + 527 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel generative viewpoint of video summarization, better suited for the subjective nature of the task, allowing multiple plausible summaries for a video.", + "- We innovatively apply diffusion to the video summarization for the first time, integrating learning the distribution of good summaries into the model.", + "- We analyze the knapsack optimization process and propose additional metrics to quantify the optimality of the predicted importance scores." + ], + "bbox": [ + 99, + 529, + 483, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Problem Formulation", + "text_level": 1, + "bbox": [ + 89, + 678, + 295, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given a video of $N$ frames, the objective of video summarization is to identify and select $S < N$ frames that effectively encapsulate the essence of the video content. Let $\\mathbf{X} \\equiv \\{\\mathbf{X}_i \\in \\mathbb{R}^{H \\times W \\times 3} : i = 0, \\dots, N - 1\\} \\in \\mathbb{R}^{N \\times H \\times W \\times 3}$ be a video, where $H, W$ denote the size of the frames.", + "bbox": [ + 89, + 703, + 483, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A video summary, $\\mathbf{y} \\equiv \\{y_i \\in \\{0,1\\} : i = 0,\\dots,N-1, \\sum_i y_i \\leq S\\} \\in \\{0,1\\}^N$ , indicates inclusion (1) or exclusion (0) of each frame. When a video is provided with multiple ground truth annotations, we denote each individual score by $\\mathbf{s}^{(r)} \\in [0,1]^N$ and the corresponding binary summary by $\\mathbf{y}^{(r)} \\in \\{0,1\\}^N$ which is obtained following the procedure explained in Sec. 3.3. The predicted summary is denoted by $\\hat{\\mathbf{y}} \\equiv \\{\\hat{y}_i \\in \\mathbb{R} : i = 0,\\dots,N-1\\} \\in \\mathbb{R}^N$ .", + "bbox": [ + 89, + 779, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous models [3, 17, 67] have approached video summarization as a regression task, aiming to predict the importance score $\\mathbf{s} \\in [0,1]^N$ for each video, often the average of multiple importance scores, $\\frac{1}{|R|} \\sum_{r \\in \\mathcal{R}} \\mathbf{s}^{(r)}$ . In contrast, our approach allows multiple summaries for each video, aiming to learn the distribution of its plausible summaries. The previous scheme can be seen as a special case of ours, where every video has a single golden way of summarization, pointed at $\\mathbf{s}$ with zero variance. Under our extended setting, multiple importance scores $\\{\\mathbf{s}^{(r)} | r \\in \\mathcal{R}\\}$ can be given to the same conditioning video $\\mathbf{X}$ , forming a probability distribution of plausible importance scores.", + "bbox": [ + 511, + 301, + 906, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Diffusion-based Video Summarization", + "text_level": 1, + "bbox": [ + 511, + 500, + 857, + 515 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Posing the video summarization as a conditional generation task, we introduce our SummDiff model, designed to adapt the distribution of individual importance scores for a given video by learning to denoise.", + "bbox": [ + 511, + 525, + 906, + 585 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Overall Flow of the Proposed Method", + "text_level": 1, + "bbox": [ + 511, + 597, + 836, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fig. 1 illustrates the overall flow of our SummDiff model.", + "bbox": [ + 511, + 619, + 893, + 635 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Encoding Process. We first encode each frame $\\mathbf{X}_i$ for $i = 1, \\dots, N$ from the input video $\\mathbf{X}$ using a pre-trained image encoder. The extracted features are further contextualized through self-attention [80], as seen in Fig. 1. We denote the encoded feature for each individual frame by $\\mathbf{z}_i \\in \\mathbb{R}^D$ , and collectively the entire feature matrix by $\\mathbf{Z} \\in \\mathbb{R}^{N \\times D}$ .", + "bbox": [ + 511, + 638, + 905, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Denoising Process. We then learn to denoise an individual importance score vector from a random noise, conditioned on the visual embeddings. First, the forward process adds noise to the ground truth individual importance score $\\mathbf{s}_0 \\equiv \\mathbf{s}^{(r)}$ and sample a noised one $\\mathbf{s}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{s}_0 + \\sqrt{1 - \\bar{\\alpha}_t}\\boldsymbol{\\epsilon}_t$ , where $\\boldsymbol{\\epsilon}_t \\sim \\mathcal{N}(0,\\mathbf{I}), \\bar{\\alpha}_t = \\prod_{\\tau=1}^t (1 - \\beta_\\tau)$ , and $t$ is the diffusion time step. The perturbation kernel at $t$ , defined as $q(\\mathbf{s}_t|\\mathbf{s}_{t-1}) = \\mathcal{N}(\\mathbf{s}_t; \\sqrt{1 - \\beta_t}\\mathbf{s}_{t-1}, \\beta_t\\mathbf{I})$ , where $\\beta_t$ is defined by the variance schedule. Since $\\mathbf{s}_0$ is bounded within $[0,1]$ , it is not straightforward to add Gaussian noise directly to it so we first transform it to its logit, $\\mathbf{u}_0 = \\log \\frac{\\mathbf{s}_0}{1 - \\mathbf{s}_0} \\in$", + "bbox": [ + 511, + 734, + 906, + 904 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15097", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "$\\mathbb{R}^N$ , and perform the noising process in this logit space. For numerical stability, we clip $\\mathbf{s}_0$ to $[\\epsilon, 1 - \\epsilon]$ , where $\\epsilon$ is a small constant.", + "bbox": [ + 89, + 90, + 480, + 133 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then, the reverse process progressively removes noise from $\\mathbf{s}_t$ to $\\mathbf{s}_0$ , formulated by $p_{\\theta}(\\mathbf{s}_{t - 1}|\\mathbf{s}_t) = \\mathcal{N}(\\mathbf{s}_{t - 1};\\boldsymbol {\\mu}_\\theta (\\mathbf{s}_t,t),\\sigma_t^2\\mathbf{I})$ , where $\\sigma_t^2$ is determined in relation to $\\beta_{t}$ , and the posterior mean $\\pmb{\\mu}_{\\theta}(\\mathbf{s}_t,t)$ is modeled with a trainable neural network. In this paper, we extend this estimator of posterior mean to be conditioned on a video, similarly to the conditional generation on diffusion models.", + "bbox": [ + 89, + 136, + 480, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Once trained, our denoiser is able to recover a plausible importance score for a video from a random noise, which is used to generate a summary (Sec. 3.3). Repeated generation across the noise distribution would converge to the true distribution of plausible scores for the video.", + "bbox": [ + 89, + 242, + 480, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Video Importance Score Denoiser", + "text_level": 1, + "bbox": [ + 89, + 327, + 383, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We formulate the key component of our method, the video importance score denoiser in Fig. 2. Starting from a noised logit score vector $\\mathbf{u}_t \\in \\mathbb{R}^N$ , it predicts a plausible importance score, conditioned on the given video. Denoted by $f_{\\theta}(\\mathbf{u}_t, t, \\mathbf{Z})$ , it learns to denoise the given score vector $\\mathbf{u}_t$ at the diffusion time step $t$ under the visual condition $\\mathbf{Z}$ , producing the denoised score $\\hat{\\mathbf{u}}_{t-1}$ by one time step, where $\\theta$ is the set of its learnable parameters.", + "bbox": [ + 89, + 349, + 483, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transformer-based Diffusion. We denoise the logit-transformed score $\\mathbf{u}_t$ given $t$ and $\\mathbf{Z}$ , predicting $\\hat{\\mathbf{u}}_{t-1}$ , using a transformer-based cross-attention [80]. The logit-transformed score $\\mathbf{u}_t$ acts as the query, and the visual condition $\\mathbf{Z}$ is used as key/value. This setup allows the model to effectively denoise $\\mathbf{u}_t$ conditioned on the information from the input video, ensuring consistency between the predicted logit score $\\hat{\\mathbf{u}}_{t-1}$ and the condition $\\mathbf{Z}$ .", + "bbox": [ + 89, + 474, + 483, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To apply dot-product cross-attention, the dimensionality should match for queries and keys. To this end, we quantize the importance scores into a predefined number $(K)$ of uniform segments. Specifically, we convert the noised logit $\\mathbf{u}_t$ back to its original bounded range $[0,1]$ by applying sigmoid $1 / (1 + e^{-\\mathbf{u}_t})$ , and divide them into $K$ equally-binned segments. Each score range is associated with a learnable embedding of size $D$ (codebook in Fig. 1). Based on the codebook $\\mathcal{C}$ , we map the scores $\\mathbf{u}_t \\in \\mathbb{R}^N$ to their corresponding quantized embedding $\\mathcal{C}(\\mathbf{u}_t) \\in \\mathbb{R}^{N \\times D}$ . Our denoiser $f_{\\theta}(\\mathcal{C}(\\mathbf{u}_t), t, \\mathbf{Z})$ takes this $\\mathcal{C}(\\mathbf{u}_t)$ as input, instead of $\\mathbf{u}_t$ . See Sec. 4.5 for analysis on the size of codebook.", + "bbox": [ + 89, + 595, + 483, + 775 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Query Formation in Cross-Attention. Following [25], the time step $t$ is embedded as $\\pmb{\\tau} \\in \\mathbb{R}^{D}$ using sinusoidal functions and an MLP. Considering the sequential nature of videos, we also introduce temporal positional embeddings $\\Phi \\in \\mathbb{R}^{N \\times D}$ using sinusoidal functions.", + "bbox": [ + 89, + 779, + 483, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The simplest way to set the query in the cross-attention would be to add all inputs except for the condition $\\mathbf{Z}$ ; that is, $\\mathbf{Q}_t = \\mathcal{C}(\\mathbf{u}_t) + \\boldsymbol{\\tau} + \\boldsymbol{\\Phi}$ . However, we leverage AdaLN-Zero", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "block [58] to integrate the time embedding $\\pmb{\\tau}$ and temporal positional encoding $\\Phi$ more effectively by separating them from the query $\\mathbf{Q}_t$ , preventing information mixing. Hence, the query becomes $\\mathbf{Q}_t = \\mathcal{C}(\\mathbf{u}_t)$ and $\\pmb{\\tau}, \\Phi$ are conditioned via scale-shift operation. See Sec. 4.5 for ablation studies.", + "bbox": [ + 511, + 90, + 903, + 166 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We conduct cross-attention with $\\mathbf{Q}_t$ as queries and visual conditions $\\mathbf{Z} \\in \\mathbb{R}^{N \\times D}$ as keys $\\mathbf{K}$ and values $\\mathbf{V}$ . We take an MLP from the AdaLN output to regress the scale and shift parameters, denoted by $\\mathbf{A}_1$ , $\\mathbf{B}_1$ , $\\Gamma_1$ , $\\mathbf{A}_2$ , $\\mathbf{B}_2$ , and $\\Gamma_2 \\in \\mathbb{R}^{N \\times D}$ . As depicted in Fig. 2, they scale and shift $\\mathbf{Q}_t$ , $\\mathbf{K}$ , and $\\mathbf{V}$ . Then, they are passed through cross-attention with skip connections and a subsequent rescaling. Formally,", + "bbox": [ + 511, + 167, + 905, + 273 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} ^ {i ^ {\\prime}} = \\boldsymbol {\\Gamma} _ {1} \\odot \\mathbf {H} ^ {i} + \\mathbf {H} ^ {i} + \\mathbf {B} _ {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 286, + 761, + 301 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {1} = \\mathbf {A} _ {1} \\odot \\operatorname {s o f t m a x} \\left(\\mathbf {Q} _ {t} ^ {i ^ {\\prime}} \\mathbf {K} ^ {i ^ {\\prime} \\top}\\right) \\mathbf {V} ^ {i ^ {\\prime}} + \\mathbf {Q} _ {t} ^ {i ^ {\\prime}}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 306, + 857, + 325 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {2} = \\boldsymbol {\\Gamma} _ {2} \\odot \\mathbf {X} _ {1} + \\mathbf {X} _ {1} + \\mathbf {B} _ {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 330, + 764, + 344 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} _ {t} ^ {i + 1} = \\mathbf {A} _ {2} \\odot \\operatorname {M L P} (\\mathbf {X} _ {2}) + \\mathbf {X} _ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 349, + 777, + 366 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{H}^{i^{(t)}}\\in \\{\\mathbf{Q}_t^{i^{(t)}},\\mathbf{K}^{i^{(t)}},\\mathbf{V}^{i^{(t)}}\\}$ denotes the matrices used for attention, $\\odot$ is the (broadcasted) Hadamard product, and $i = 1,\\ldots ,L$ is the layer index.", + "bbox": [ + 511, + 378, + 903, + 430 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training. We train the denoiser $f_{\\theta}$ to estimate the true importance score $\\hat{\\mathbf{s}}_0$ after acting upon a fully-connected layer and a sigmoid function $\\sigma$ . We minimize the following loss [12] on each annotator's individual importance score:", + "bbox": [ + 511, + 434, + 905, + 494 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} \\left(\\mathbf {s} _ {0}, \\hat {\\mathbf {s}} _ {0}\\right) = \\left\\| \\mathbf {s} _ {0} - \\hat {\\mathbf {s}} _ {0} \\right\\| _ {2} ^ {2} = \\left\\| \\mathbf {s} _ {0} - \\sigma \\left(\\operatorname {F C} \\left(f _ {\\theta} \\left(\\mathcal {C} \\left(\\mathbf {u} _ {t}\\right), t, \\mathbf {Z}\\right)\\right)\\right) \\right\\| _ {2} ^ {2}.\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 507, + 901, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The learnable embeddings in the codebook $\\mathcal{C}$ are compositely optimized, finding an effective representation for $\\mathbf{u}_t$ during training.", + "bbox": [ + 511, + 537, + 903, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inference. Our model generates a logit-transformed importance score from a random noise $\\mathbf{u}_T \\sim \\mathcal{N}(0, \\mathbf{I})$ for a given video. Employing the reverse diffusion process [68], it iteratively refines the logit score towards cleaner estimations:", + "bbox": [ + 511, + 588, + 903, + 648 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {\\mathbf {u}} _ {t - 1} = \\sqrt {1 - \\bar {\\alpha} _ {t - 1} - \\sigma_ {t} ^ {2}} \\frac {\\hat {\\mathbf {u}} _ {t} - \\sqrt {\\bar {\\alpha} _ {t}} f _ {\\theta} (\\mathcal {C} (\\mathbf {u} _ {t}) , t , \\mathbf {Z}))}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\\\ + \\sqrt {\\bar {\\alpha} _ {t - 1}} f _ {\\theta} (\\mathcal {C} (\\mathbf {u} _ {t}), t, \\mathbf {Z})) + \\sigma_ {t} \\epsilon_ {t} \\in \\mathbb {R} ^ {N}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 661, + 880, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\sigma_{t} = \\sqrt{(1 - \\bar{\\alpha}_{t - 1}) / (1 - \\bar{\\alpha}_{t})}\\sqrt{1 - \\bar{\\alpha}_{t} / \\bar{\\alpha}_{t - 1}}$ (DDPM reverse process). When $t = 1$ at the last step, $\\hat{\\mathbf{u}}_0 = f_\\theta (\\mathcal{C}(\\mathbf{u}_1),1,\\mathbf{Z})$ is directly used to remove stochasticity at the end of inference process. Finally, we take sigmoid to $\\hat{\\mathbf{u}}_0$ to derive the final importance score $\\hat{\\mathbf{s}}_0\\in (0,1)^N$", + "bbox": [ + 511, + 728, + 905, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Summary Generation", + "text_level": 1, + "bbox": [ + 511, + 816, + 718, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "From the raw importance score $\\hat{\\mathbf{s}}_0$ , we apply the standard knapsack-based approach to decide which frames to be included in the final summary. To make this summary video more realistic, it is common to choose at a semantic clip", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15098", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "level instead of individual frame level. We adopt a widely-used Kernel Temporal Segmentation (KTS) [59, 84] to partition a video into disjoint temporal intervals. We take the average score among the corresponding frames as the clip-level importance; i.e., $v_{i} = \\sum_{j = t_{i}}^{t_{i + 1}}\\hat{\\mathbf{s}}_{0,j} / (t_{i + 1} - t_{i})$ for the $i$ -th clip composed of frames from $t_{i}$ to $t_{i + 1}$ . Subsequently, we select the clips based on $v_{i}$ by solving the binary knapsack problem (KP) [23] with dynamic programming [71]:", + "bbox": [ + 89, + 90, + 483, + 212 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {K P} (\\boldsymbol {v}, \\boldsymbol {w}, \\rho) \\equiv \\underset {\\boldsymbol {s} \\in \\{0, 1 \\} ^ {M}} {\\operatorname {a r g m a x}} \\sum_ {i = 1} ^ {M} v _ {i} s _ {i} \\quad \\text {s . t .} \\sum_ {i = 1} ^ {M} w _ {i} s _ {i} \\leq \\rho N,\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 223, + 473, + 265 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{v},\\pmb{w},\\rho$ denotes the values of each clip $(v_{i}\\in [0,1])$ , costs of selecting each clip, and the budget constraint ratio (e.g., $\\rho = 0.15$ ; $15\\%$ of the video length), respectively. If the argmax has multiple solutions, we will abuse the notation and denote $\\mathcal{KP}(\\pmb {v},\\pmb {w},\\rho)$ as an arbitrary element of the solution set.", + "bbox": [ + 89, + 276, + 483, + 367 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 381, + 223, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 406, + 297, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets. We evaluate our method on three benchmarks. TVSum [71] and SumMe [20] are traditional datasets, containing 50 and 25 videos respectively and manually labeled annotations by up to 20 raters. Mr. HiSum [74] is a large-scale dataset, composed of 31,892 videos and importance scores derived from YouTube Most Replayed statistics. It provides an importance score $\\mathbf{s}_0$ and a corresponding summary $\\mathbf{y}_0$ for each video, averaged over 50,000+ viewers. We use the default features for each dataset, GoogLeNet [75] for TVSum and SumMe, and Inception-v3 [76] PCA-ed to 1024D [1] for Mr. HiSum.", + "bbox": [ + 89, + 428, + 483, + 593 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We adopt two dataset splits. First, we randomly split TVSum and SumMe into 7:1:2 for train, validation, and test set, and report averaged test scores for 5 random splits (namely, TVT). Also, following prior works [67, 77], we include five-fold cross-validation results (5FCV), although these may suffer from overfitting to the test set due to the lack of explicit validation set, and thus may not generalize well on unseen videos. For Mr. HiSum, we use its original train/validation/test split. We run all experiments 5 times.", + "bbox": [ + 89, + 595, + 483, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following recent practice [22, 40, 55, 67], we report rank order statistics, Kendall's $\\tau$ [37] and Spearman's $\\rho$ [96], to assess how well the model ranks frame importance. F1 score had been widely used in video summarization, but recent literature [67, 77] reports that it is excessively sensitive to video segmentation and unreasonably favors summaries composed of many short shots [55], while disregarding longer key shots [67]. As an alternative to F1, we propose additional metrics that represent the influence of importance score through an analysis of the knapsack algorithm in Sec. 4.4.", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/df654f5d1cb98b267b5ef5e03974cf7e0ff2176ef07e2718947dd01c963023cd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSumMeTVSum
τρτρ
Random0.0000.0000.0000.000
Human0.2050.2130.1770.204
A2Summ [22]0.0880.0960.1570.206
VASNet [17]0.0890.0990.1530.205
PGL-SUM [3]0.1040.1160.1410.186
CSTA [67]0.1080.1200.1680.221
SummDiff (Ours)0.1330.1480.1730.226
", + "bbox": [ + 544, + 88, + 874, + 231 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/e6b4bb92ef07c2a69120ab8ffc13102d9402d1a78e87b3a48b3d0c8bf87f1114.jpg", + "table_caption": [ + "Table 1. Comparison of models trained under TVT (train/val/test split) on SumMe [20] and TVSum [71]. Best and second-best results are boldfaced and underlined, respectively." + ], + "table_footnote": [], + "table_body": "
MethodSumMeTVSum
τρτρ
Random0.0000.0000.0000.000
Human0.2050.2130.1770.204
DSNet-AF [93]0.0370.0460.1130.138
DSNet-AB [93]0.0510.0590.1080.129
SUM-GAN [50]0.0490.0660.0240.031
AC-SUM-GAN [2]0.1020.0880.0310.041
CLIP-It [52]--0.1080.147
iPTNet [34]0.1010.1190.1340.163
A2Summ [22]0.1080.1290.1370.165
VASNet [17]0.1600.1700.1600.170
PGL-SUM [3]--0.1570.206
AAAM [77]--0.1690.223
MAAM [77]--0.1790.236
VSS-Net [85]--0.1900.249
DMASum [81]0.0630.0890.2030.267
SSPVS [40]0.1920.2570.1810.238
CSTA [67]0.2460.2740.1940.255
SummDiff (Ours)0.2560.2850.1950.255
", + "bbox": [ + 542, + 282, + 874, + 556 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/2b9557a29c7a75350871861a7141eb95acfacd494bd9ea976b44ca6d9b6c228f.jpg", + "table_caption": [ + "Table 2. Comparison of models trained under 5FCV (5-Fold Cross Validation) on SumMe [20] and TVSum [71]. Training under the 5FCV setting tends to overfit the test set." + ], + "table_footnote": [], + "table_body": "
Modelτ↑ρ↑MAPρ=50%↑MAPρ=15%↑
SUM-GAN [50]0.0670.09556.6223.56
VASNet [17]0.0690.10258.6925.28
AC-SUM-GAN [2]0.0120.01855.3521.88
SL-module [82]0.0600.08858.6324.95
PGL-SUM [3]0.0970.14161.6027.45
iPTNet [34]0.0200.02955.5322.74
A2Summ [22]0.1210.17263.2032.34
CSTA [67]0.1280.18563.3830.42
SummDiff0.1750.23865.4433.83
", + "bbox": [ + 524, + 608, + 890, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 3. Evaluation on Mr. HiSum [74]. See Appendix D for the full table including standard deviations.", + "bbox": [ + 511, + 753, + 906, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We further evaluate our method on video highlight detection following [74]. First, we uniformly divide the input video into 5-second-long shots and calculate the average frame scores for each shot. The top $\\rho \\in \\{0.15, 0.5\\}$ of these shots are designated as ground truth highlights. We report Mean Average Precision (MAP), following [28, 56, 84]. See Appendix E for more implementation details.", + "bbox": [ + 511, + 794, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15099", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Results and Analysis", + "text_level": 1, + "bbox": [ + 89, + 90, + 285, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Considering the innate subjectivity of video summarization, we train our model to learn from each individual score, allowing it to capture multiple ways of summarization for each video. Specifically, we use 20 individual importance scores per video in TVSum [71] and 15-18 individual binary summaries in SumMe [20], since SumMe does not provide individual importance scores. Tab. 1 and 2 show that SummDiff achieves the best performance on SumMe and TVSum, with the sole exception of TVSum in 5FCV. While DMA-SUM [81] performs the best on TVSum in 5FCV, it significantly underperforms on SumMe, indicating that its high performance on TVSum does not generalize well. Comparing TVT and 5FCV, most models exhibit a considerable performance gap. This demonstrates how much existing models have overfitted to the test set under the conventional 5FCV protocol. We adopt TVT as a remedy to this, but due to the extremely small size of these datasets, the reliability of evaluation is still limited [74].", + "bbox": [ + 89, + 112, + 483, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "On the larger Mr. HiSum, we train our model using the aggregated annotation, as it does not provide individual annotations. This evaluation mainly aims to verify scalability of SummDiff on a large-scale dataset. Tab. 3 presents the performance on Mr. HiSum, where SummDiff consistently outperforms all baselines across all metrics, significantly surpassing the strongest competitor, CSTA [67] with a large margin. These results highlight scalability and effectiveness of our method even under the single-label setting.", + "bbox": [ + 89, + 385, + 483, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Qualitative Demonstration", + "text_level": 1, + "bbox": [ + 89, + 531, + 331, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We demonstrate how our method discovers a variety of possible summaries for a video. Specifically, we generate video summaries for the 50 videos in TVSum using CSTA, PGL-SUM, VASNet, and ours. For each generated summary, we measure the Kendall's $\\tau$ with all 20 ground truth annotations, respectively. Considering $\\tau \\geq 0.25$ as the threshold for the summary to be matched with the annotator, Fig. 3 shows if the generated summary for a video (row) matches with each annotator (column). As the baselines (CSTA, PGL-SUM, and VASNet) summarize a video deterministically, each cell is either completely matched (1) or not (0). If the summary matches with majority of annotators (e.g., video 29 for CSTA), it means the deterministic summary matches with the dominant way of summarization for the video. If it matches only with a few annotators (e.g., video 24 for CSTA), it means the generated summary matches with a minor opinion. For our method, we generate 100 summaries per video from different Gaussian noise vectors, and mark each cell with the ratio of summaries that match with each annotator. The darker the cell colors are for each row, the more various ways of summaries have been discovered by our method. Comparing the heat maps, we clearly observe that SummDiff generates significantly more various", + "bbox": [ + 89, + 553, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f6cca2b7b476a0a7178d63f359fccc011c31a9fb1f76beb79ca54083ab5574d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 82, + 705, + 239 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ed7787ef43b384971c21f8eb6d490f9d15b74c6150f5154fa81abe62b82c9536.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 82, + 880, + 239 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c77634b845e23c2160ffb7c7cd3817622f8142f6232c9fb7d822ba099c7450fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 241, + 705, + 404 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bc2e7b4fbbc547903332dbb2ce3759e013d81c4480865824204160c0bb4986e8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 241, + 880, + 404 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9df82d52124190e43a3d21ab5eb0b84ba032c661885ddca815c6f2cd07ebfdec.jpg", + "image_caption": [ + "Figure 3. Ratio of summaries with $\\tau \\geq 0.25$ for each video-annotator pair. The heatmap illustrates how closely each method's summary matches with individual annotations. This comparison reveals the extent to which each method captures the varied human summaries. SummDiff covers a larger area of the heatmap, which indicates better coverage over various summaries." + ], + "image_footnote": [], + "bbox": [ + 540, + 491, + 709, + 614 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a883bc0e6af6c6b10d15d29b157630f6feedc0d4f7a0f108e61f696362092211.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 491, + 880, + 614 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/18a5f862fe166f908f6ba15ecdaa5b03d9cde469cf7c05c78d9f499218bd91f7.jpg", + "image_caption": [ + "Figure 4. Distribution of true and predicted summaries of given videos. SummDiff can generate various summaries and cover the targeted distribution of summaries, while baselines deterministically predicts a single summary." + ], + "image_footnote": [ + "SummDiff CSTA PGL_SUM VASNet Annotator" + ], + "bbox": [ + 539, + 614, + 710, + 739 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3ae161d9bbdee32928283589715ed8b7ad9d1dd34952dabb9e1f4ab851441c37.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 614, + 880, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "summaries covering multiple viewpoints, better reflecting the distribution of annotated summaries.", + "bbox": [ + 511, + 824, + 903, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We further illustrate the variation and quality of generated summaries in Fig. 4. The summaries annotated by human raters ( $\\bullet$ Annotator) and those generated by ours ( $\\bullet$", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15100", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 5. Demonstration of video summaries generated by competing methods on a TVSum video. Shaded parts indicate the segments selected by each method, and the two rows of edged boxes within each method indicate the two distinct true annotations. The results clearly demonstrate the effectiveness of SummDiff in capturing multiple plausible summaries for a video. See another example in Appendix F.", + "bbox": [ + 89, + 282, + 906, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SummDiff) and baselines ( • CSTA, • PGL-SUM, and • VASNet) are projected onto the first two principal components, which are computed using only the human-annotated summaries. The distribution of true summaries are displayed using a contour map, obtained by Gaussian kernel density estimation. It demonstrates that our SummDiff generates multiple summaries closely aligned with the distribution of true annotations. For example, video 22 (top-left) roughly has three ways to summarize, according to the human raters. SummDiff generates summaries dominantly covering the two modes in the lower-left region, and one case in the upper-right mode. In contrast, baselines produce a single, less accurate summary, failing to account for the variety inherent in human-annotated summaries. We observe similar patterns in the other three plots as well. Fig. II provides more examples in Appendix C.", + "bbox": [ + 88, + 338, + 480, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 5 illustrates example summaries generated by CSTA, A2Summ, PGL-SUM, VASNet, and our SummDiff, for a video selected from TVSum test set. Shaded parts indicate the segments selected by each method (note that SummDiff can generate two different summaries while baselines always produce a single summary), and the two rows of edged boxes within each method indicate two different true annotations. These results demonstrate that SummDiff produces more accurate and various summaries, effectively capturing multiple plausible summarizations for each video.", + "bbox": [ + 89, + 580, + 482, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Metrics Inspired from Knapsack Optimality", + "text_level": 1, + "bbox": [ + 89, + 741, + 467, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In spite of nontrivial impact of the knapsack at the end of the summary generation, previous evaluation metrics have focused only on the accuracy of the importance scores. Through a thorough analysis on the knapsack problem (KP), we provide additional metrics that measure the contribution of importance scores. By accounting for knapsack optimality and clip-level weights, our new metrics are capable of assessing the predicted importance score more accurately than existing metrics such as F1 or ranking-based ones.", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Confidence of Importance Score. First, we analyze the conditions under which the same optimal KP solution remains valid despite some perturbations to the profits $\\pmb{v}$ . These perturbations can be a modeling to the imperfect estimation of the importance scores in video summarization.", + "bbox": [ + 511, + 338, + 903, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Let $\\pmb{y}^{*}$ be the optimal solution to the original KP (i.e., $\\pmb{y}^{*} = \\mathcal{KP}(\\pmb{v},\\pmb{w},\\rho)$ ; see definition in Sec. 3.3) and $\\Gamma \\subseteq [N]\\equiv \\{n\\in \\mathbb{N}\\mid 1\\leq n\\leq N\\}$ be an arbitrary subset of items associated with the perturbed profits $v_{i}^{\\prime} = v_{i} + \\Delta v_{i}$ , where $\\Delta v_{i} = 0$ if $i\\notin \\Gamma$ and $N$ is the number of items. In our case, $\\pmb{v}\\in \\mathbb{R}^{N}$ is the ground truth importance scores, $\\pmb{y}^{*}\\in [0,1]^{N}$ is the ground truth summary, and the perturbed profit $\\pmb{v}^{\\prime}\\in \\mathbb{R}^{N}$ corresponds to the imperfectly predicted importance scores. We further define two disjoint subsets of $\\Gamma$ . First, $\\Gamma_0^+ \\equiv \\{i\\in \\Gamma \\mid y_i^* = 0,\\Delta p_i > 0\\}$ refers to the set of items excluded in the original optimal solution but undergoing increased profit, and $\\Gamma_1^- \\equiv \\{i\\in \\Gamma \\mid y_i^* = 1,\\Delta p_i < 0\\}$ is the opposite, the set of items included in the optimal solution but experiencing a decreased profit. Then, Hifi et al. [23] claims that the solution $\\pmb{y}^{*}$ remains optimal for the perturbed KP, i.e., $\\pmb{y}^{*} = \\mathcal{KP}(\\pmb{v},\\pmb{w},\\rho) = \\mathcal{KP}(\\pmb{v}^{\\prime},\\pmb{w},\\rho)$ , if", + "bbox": [ + 511, + 412, + 906, + 672 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\sum_ {i \\in \\Gamma_ {0} ^ {+}} \\Delta v _ {i} - \\sum_ {i \\in \\Gamma_ {1} ^ {-}} \\Delta v _ {i} \\leq \\boldsymbol {v} \\cdot \\mathcal {K P} (\\boldsymbol {v}, \\boldsymbol {w}, \\rho) - \\max (\\Psi^ {+}, \\Psi^ {-}),\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 681, + 903, + 719 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\Psi^{+} \\equiv \\max_{i \\in \\Gamma_{0}^{+}} [\\pmb{v} \\cdot \\mathcal{KP}(\\pmb{v} - v_{i}\\hat{e}_{i}, \\pmb{w}, \\rho - w_{i}/N)]$ is the maximum value of the KP when removing any item from $\\Gamma_{0}^{+}$ and adjusting the constraint, and similarly, $\\Psi^{-} \\equiv \\max_{i \\in \\Gamma_{1}^{-}} [\\pmb{v} \\cdot \\mathcal{KP}(\\pmb{v} - v_{i}\\hat{e}_{i}, \\pmb{w}, \\rho)]$ is the maximum value of the KP when removing any item from $\\Gamma_{1}^{-}$ .", + "bbox": [ + 511, + 729, + 906, + 810 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Based on this theorem we propose a new metric, Confidence of Importance Score (CIS), to estimate how close the predicted scores are to the ground truth:", + "bbox": [ + 511, + 811, + 903, + 856 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {CIS} = \\sum_ {i \\in \\Gamma_ {0} ^ {+}} \\Delta v _ {i} - \\sum_ {i \\in \\Gamma_ {1} ^ {-}} \\Delta v _ {i} - \\boldsymbol {v} \\cdot \\mathcal {K P} (\\boldsymbol {v}, \\boldsymbol {w}, \\rho) + \\max \\left(\\Psi^ {+}, \\Psi^ {-}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 866, + 924, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15101", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "If CIS $\\leq 0$ , the theorem indicates that the predicted importance score $\\hat{\\pmb{v}}$ is guaranteed to induce the true KP solution $\\pmb{y}^{*}$ . If CIS $>0$ , $\\hat{\\pmb{v}}$ does not guarantee to induce the KP solution, but a lower CIS would indicate a solution closer to $\\pmb{y}^{*}$ than a higher one. That is, a smaller CIS indicates a stronger confidence in satisfying the inequality, indicating a higher chance for the solution of the perturbed KP to be identical to that of the original KP.", + "bbox": [ + 88, + 90, + 480, + 210 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Unlike F1 that simply measures how close the predicted summary is to the ground truth, our CIS accounts for the confidence that the predicted score will exactly induce the true summary. This is particularly important in video summarization, since the importance score itself is subjective and often noisy, requiring robust selection of frames.", + "bbox": [ + 89, + 210, + 480, + 301 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Weighted Inclusion Ratios. Analyzing the sensitivity of the KP optimum to the perturbation of the importance score, Belgacem and Hifi [5] established bounds of the perturbed profits to retain the original optimum. These intervals are", + "bbox": [ + 89, + 305, + 480, + 367 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {y} _ {i} ^ {*} = 1 \\rightarrow \\Delta v _ {i} \\in I _ {i} ^ {1} = \\left[ \\right. \\max \\left(\\Delta_ {i} ^ {-}, w _ {i} \\max _ {k \\in \\Gamma} \\mu_ {k} - p _ {i}\\right), + \\infty\\left. \\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 373, + 503, + 407 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {y} _ {i} ^ {*} = 0 \\rightarrow \\Delta v _ {i} \\in I _ {i} ^ {0} = \\left( \\right.- \\infty , \\min \\left(\\Delta_ {i} ^ {+}, p _ {i} - w _ {i} \\min _ {k \\in \\Gamma} \\mu_ {k}\\right)\\left. \\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 407, + 503, + 444 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\Delta_i^+ = \\pmb{v} \\cdot (\\mathcal{KP}(\\pmb{v}, \\pmb{w}, \\rho) - \\mathcal{KP}(\\pmb{v} - v_i \\hat{e}_i, \\pmb{w}, \\rho - w_i / N), \\Delta_i^- = \\pmb{v} \\cdot (\\mathcal{KP}(\\pmb{v} - v_i \\hat{e}_i, \\pmb{w}, \\rho) - \\mathcal{KP}(\\pmb{v}, \\pmb{w}, \\rho))$ , and $\\mu_k$ denotes the critical ratio of $k \\in \\Gamma$ : $\\mu_k = \\mathcal{CR}([N] \\setminus \\{k\\}, \\rho - 1_{\\{y_k^* = 0\\}} \\cdot w_k / N)$ , with $\\mathcal{CR}(A, \\gamma) = v_s / w_s$ and $s = \\min\\{l | \\sum_{i=1}^{l} v_s / w_s > \\rho\\}$ is the index of the critical item. The items in $A$ are considered in descending order of $v_s / w_s$ . To sum up, if the perturbed profit resides in the interval $v_i' \\in I_i^{y_i^*}$ , it is guaranteed that the solution of perturbed KP is identical with that of the original KP.", + "bbox": [ + 89, + 450, + 480, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We propose Inclusion Ratio (IR) to estimate how much the predicted scores fall into the safe bounds; for the $i$ -th importance score, $\\mathrm{IR}_i \\equiv |\\{\\Delta v_i \\in I_i^{y_i^*} | i \\in [N]\\}| / N$ . This ratio reflects the proportion of the predicted scores that retain the KP solution. We propose the Weighted average of the Inclusion Ratios (WIR) as a metric to measure the proximity of the predicted score to the ground truth: $\\mathrm{WIR} \\equiv \\sum_{i=1}^{n} \\left( \\frac{w_i}{\\sum_{j=1}^{n} w_j} \\mathrm{IR}_i \\right)$ , where $w_i$ denotes the segment length of each shot.", + "bbox": [ + 89, + 594, + 480, + 737 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Weighted Sum of Errors (WSE). Additionally, we compare a weighted sum of the errors $\\sum_{i} w_{i} \\Delta p_{i}$ , where $w_{i}$ is the segment length of each clip.", + "bbox": [ + 89, + 742, + 480, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison. Tab. 4 compares several competitive methods using these new metrics. Our SummDiff achieves the best performance in all metrics, indicating that the summaries generated by our method are closer to the optimal knapsack result, under the theoretical analysis in literature.", + "bbox": [ + 89, + 791, + 480, + 866 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Discussion. Our newly proposed metrics, WIR and CIS, evaluate the quality of the generated importance scores", + "bbox": [ + 89, + 869, + 480, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/dce8e748cc878bed82f909da2f2278731459d5b937dcfc57ffc44b37773dc4af.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methodτ↑ρ↑CIS ↓WIR ↑WSE ↓
Uniform-Random0.0000.0009.270.4546.82
SL-module [82]0.0600.0886.830.5630.09
CSTA [67]0.1280.1856.230.5725.76
PGL-SUM [3]0.0970.1416.140.5826.22
VASNet [17]0.0690.1026.250.5926.79
A2Summ [22]0.1210.1726.650.5530.55
SummDiff0.1750.2385.960.6125.24
", + "bbox": [ + 519, + 84, + 895, + 196 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8b7bb23fb86b45a2ea0becd10db4eeec9655c1064737e1f504c00934ecbda8fa.jpg", + "table_caption": [ + "Table 4. Evaluation using our new metrics proposed in Sec. 4.4." + ], + "table_footnote": [], + "table_body": "
Moduleτ↑ρ↑
Encoder Only0.0710.104
(+) Video Importance Denoiser0.0790.116
(+) Learnable Embedding0.0860.124
(+) AdaLN (τ)0.1450.204
(+) AdaLN (τ, Φ) 1 layer0.1710.232
(-) Quantization0.1250.174
(+) Classifier-free Guidance (CFG) [24]0.1750.238
(+) Self-attention Guidance (SAG) [30]0.1770.239
", + "bbox": [ + 542, + 218, + 874, + 342 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Effect of individual components", + "bbox": [ + 586, + 345, + 831, + 359 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in two aspects. WIR measures how many importance scores are individually trustworthy, weighted by the duration. Specifically, WIR first computes the safe interval $I_{i}$ for each importance score such that the true summary remains unchanged. Then, given the predicted importance scores $\\hat{\\mathbf{v}}$ , it measures how many of them fall within their corresponding safe intervals $\\hat{v}_{i} \\in I_{i}$ .", + "bbox": [ + 511, + 369, + 903, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On the other hand, CIS quantifies the risk that the predicted importance scores will lead to a different summary than the one generated using the true importance scores. In other words, CIS evaluates the overall chance how likely the predicted score vector $\\hat{\\mathbf{v}}$ , as a whole, is to give a summary different from the true one. With the analysis of knapsack optimality, these metrics would better measure the accuracy of predicted importance score, considering the fidelity of final summary after knapsack to the ground truth, than F1 or ranking-based metrics.", + "bbox": [ + 511, + 476, + 905, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 636, + 666, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We report ablation study on Mr. HiSum [74] in Tab. 5, comparing performance adding a component step-by-step. Starting from a two-layer transformer encoder predicting the importance score, adding our denoiser with a naively summed query $\\mathbf{Q}_t = \\mathcal{C}(\\mathbf{u}_t) + \\pmb{\\tau} + \\pmb{\\Phi}$ improves the Kendall's $\\tau$ $(0.071\\rightarrow 0.079)$ . We observe further slight improvement $(0.079\\to 0.086)$ when we replace the fixed random score embeddings with learnable ones (Learnable Embedding). Introducing AdaLN transformation [58] on the time $(\\tau)$ and positional $(\\Phi)$ encodings further improves $\\tau$ to 0.145 and 0.171, respectively. As we adopt the quantization idea, grouping nearby scores into a unified embedding, in Sec. 3.2 with this change, we experiment without it (using a sinusoidal function [80] instead), and observe significant performance drop $(0.171\\rightarrow 0.125)$ . This confirms that quantization is superior to treating the score", + "bbox": [ + 511, + 657, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15102", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3818b010d77493871d9c18871f7317518d166e3cc18edb2154e74fe5aaa650a2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Modelτ↑ρ↑
CSTA [67]0.1280.185
SummDiff (1 step)0.1700.234
SummDiff (10 steps)0.1750.238
SummDiff (100 steps)0.1820.245
", + "bbox": [ + 174, + 88, + 395, + 165 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Performance with various numbers of DDIM steps", + "bbox": [ + 109, + 167, + 464, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d05a072a6fd027b8eff8f686d74093d6cca3b5b0e967d64ba25169f53789094a.jpg", + "image_caption": [ + "Figure 6. Measured $\\mathbb{E}|\\{\\pmb{y}^*\\}|$ and $\\mathbb{E}[\\Delta \\pmb{y}^*]$ with respect to the quantization strength $K$ . A larger $K$ is associated with a smaller number of allowed solutions for the given KP." + ], + "image_footnote": [], + "bbox": [ + 127, + 186, + 444, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "as a continuous scalar. On top of this, we combine our model with two widely-known ideas, classifier-free guidance (CFG) [24] and self-Attention guidance (SAG) [30], for further improvement. See Appendix B for details.", + "bbox": [ + 89, + 378, + 482, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Number of DDIM Steps. Tab. 6 suggests that SummDiff outperforms CSTA [67], the best-performing baseline, even with a single DDIM [68] step. Beyond this, SummDiff further improves its performance with more DDIM steps.", + "bbox": [ + 89, + 441, + 483, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Quantization Strength. Solving a KP usually has a unique solution if the profits $v_{i}$ are continuous. As quantization unifies importance score values within the same interval to a single embedding, it may lead to multiple solutions to $\\mathcal{KP}(\\tilde{\\boldsymbol{v}},\\boldsymbol {w},\\rho)$ , where $\\tilde{\\boldsymbol{v}}$ is a quantized version of $\\boldsymbol{v}$ .", + "bbox": [ + 89, + 506, + 482, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To analyze this effect, we measure the average number of solutions $\\mathbb{E}_{\\pmb{v}_1\\sim U(0,1)\\otimes N}[\\{\\pmb{y}^*\\in \\{0,1\\} ^N |\\pmb{y}^* = \\mathcal{KP}(\\tilde{\\pmb{v}},\\pmb {w},\\rho)\\} ]$ while varying $K$ , taking expectation over $\\pmb{v}$ , which is a value vector sampled from a joint uniform distribution. We also consider the average $L_{1}$ difference between two summaries, $\\mathbb{E}_{\\pmb{v}_1\\sim U(0,1)\\otimes N},\\pmb{v}_2\\sim \\mathcal{N}(\\pmb{v}_1,\\frac{1}{K}\\mathbf{I}_N)[\\| \\mathcal{KP}(\\tilde{\\pmb{v}}_1,\\pmb {w},\\rho) - \\mathcal{KP}(\\tilde{\\pmb{v}}_2,\\pmb {w},\\rho)\\| _1]$ , obtained by solving the KP with two similar but distinct value vectors, i.e., $\\pmb{v}_2 = \\pmb{v}_1 + \\Delta \\pmb{v}$ , where $\\Delta \\pmb{v}\\sim \\mathcal{N}(0,\\frac{1}{K}\\mathbf{I}_N)$ .", + "bbox": [ + 89, + 582, + 482, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We numerically analyze the relation between these two values, which indicate the scale of multiplicity of the solutions to the KP, and the quantization strength $K$ . As shown in Fig. 6, both the number of solutions $|\\{\\pmb{y}^*\\}|$ and the summary deviation $\\Delta \\pmb{y}^*$ decrease with larger $K$ , suggesting more accurate summaries.", + "bbox": [ + 89, + 734, + 482, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To sum up, increasing the quantization strength $K$ reduces the number of multiple optimal solutions, leading to a more accurate summary. However, an excessively higher $K$ leads to insufficient samples per bin, degrading the overall performance. See Appendix A for more ablation studies.", + "bbox": [ + 89, + 825, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Related Work", + "text_level": 1, + "bbox": [ + 513, + 89, + 656, + 106 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Video Summarization. Early models primarily rely on heuristic unsupervised learning [13, 15, 29, 38, 47, 49-51, 53, 71, 83] to select important or diverse frames, struggling with generalization. Benefiting from annotated datasets, supervised methods [8, 38, 57, 59, 60] have emerged. DSNet [93] and iPTNet [34] improve keyframe selection using frame-level annotations, and SL-Module [82] captures high-level features. With the advent of deep learning, RNNs [84, 88-90] and attention mechanism [4, 6, 33, 35, 39, 42, 77, 85, 94] have been adopted to capture temporal dependencies in videos. Particularly, VASNet [17] and PGL-SUM [3] capture both local and global frame dependencies using self-attention. DMASum [81] introduces mixture of attention layer mitigate the key Softmax Bottleneck. SUM-GAN [50] and AC-SUM-GAN [2] leverage generative adversarial networks.", + "bbox": [ + 511, + 114, + 906, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multimodal approaches [32, 91] integrate audio or textual data to video summarization. Multimodal transformers are adopted to link frames with corresponding captions, improving context-aware summaries, by CLIP-It [52], MSVA [18], SSPVS [40], and A2Summ [22]. CSTA [67] addresses computational complexity by using CNN-based spatiotemporal attention for efficient frame selection. Unlike these deterministic summarizers, our approach generates multiple plausible summaries by capturing the distribution of video summaries with various perspectives.", + "bbox": [ + 511, + 357, + 908, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Diffusion for Video Tasks. Diffusion [66] have emerged as a groundbreaking tool in high-quality image generation [21, 25, 54, 63, 64, 69, 70] and super-resolution [14, 36, 61, 65]. Recently, they are extended to video generation [7, 11, 16, 19, 26, 27, 44], joint video and audio generation [41, 62, 73], video editing [9, 10, 79], video inpainting [87] and prediction [31, 86]. Furthermore, they are applied to video understanding tasks like moment retrieval [43, 48], video object segmentation [92, 95], action segmentation [45].", + "bbox": [ + 511, + 512, + 908, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Summary", + "text_level": 1, + "bbox": [ + 511, + 662, + 624, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Video summarization is inherently subjective since people have different perspectives of a good summary. We suggest a generative viewpoint of this task where the model learns the distribution of good summaries, in contrast to the traditional approach of aggregated importance score regression. Our proposed SummDiff model, adopting diffusion models for video summarization for the first time, is able to generate multiple good summaries conditioned on the input video. Our model not only outperforms other baselines but also demonstrates the ability to generate accurate summaries customized to the individual annotators. We further propose additional metrics that measure the quality of the predicted importance scores through an insight from the actual summary generation using knapsack.", + "bbox": [ + 511, + 688, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15103", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by Samsung Electronics (IO240512-09881-01), Youlchon Foundation, NRF grants (RS-2021-NR05515, RS-2024-00336576, RS-2023-0022663) and IITP grants (RS-2022-II220264, RS-2024-00353131) by the government of Korea.", + "bbox": [ + 89, + 114, + 483, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 205, + 187, + 220 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sami Abu-El-Haija, Nisarg Kothari, Joonseok Lee, Paul Natsev, George Toderici, Balakrishnan Varadarajan, and Sudheendra Vijayanarasimhan. Youtube-8M: A large-scale video classification benchmark. arXiv:1609.08675, 2016. 4", + "[2] Evlampios Apostolidis, Eleni Adamantidou, Alexandros I Metsai, Vasileios Mezaris, and Ioannis Patras. AC-SUMGAN: Connecting actor-critic and generative adversarial networks for unsupervised video summarization. IEEE Transactions on Circuits and Systems for Video Technology, 31(8): 3278-3292, 2020. 1, 4, 8, iii", + "[3] Evlampios Apostolidis, Georgios Balaouras, Vasileios Mezaris, and Ioannis Patras. Combining global and local attention with positional encoding for video summarization. In IEEE international symposium on multimedia (ISM), 2021. 1, 2, 4, 7, 8, i, iii", + "[4] Evlampios Apostolidis, Georgios Balaouras, Vasileios Mezaris, and Ioannis Patras. Summarizing videos using concentrated attention and considering the uniqueness and diversity of the video frames. In ICMR, 2022. 8", + "[5] Tarik Belgacem and Mhand Hifi. Sensitivity analysis of the knapsack problem: Tighter lower and upper bound limits. Journal of systems science and systems engineering, 17:156-170, 2008. 7", + "[6] Manjot Bilkhu, Siyang Wang, and Tushar Dobhal. Attention is all you need for videos: Self-attention based video summarization using universal transformers. arXiv:1906.02792, 2019.8", + "[7] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 8", + "[8] Sijia Cai, Wangmeng Zuo, Larry S Davis, and Lei Zhang. Weakly-supervised video summarization using variational encoder-decoder and web prior. In ECCV, 2018. 8", + "[9] Duygu Ceylan, Chun-Hao P Huang, and Niloy J Mitra. Pix2video: Video editing using image diffusion. In ICCV, 2023. 8", + "[10] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. StableVideo: Text-driven consistency-aware diffusion video editing. In ICCV, 2023. 8", + "[11] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. VideoCrafter2: Overcoming data limitations for high-quality video diffusion models. In CVPR, 2024. 8", + "[12] Ting Chen, Ruixiang Zhang, and Geoffrey Hinton. Analog bits: Generating discrete data using diffusion models with self-conditioning. In ICLR, 2022. 3" + ], + "bbox": [ + 93, + 231, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Wen-Sheng Chu, Yale Song, and Alejandro Jaimes. Video co-summarization: Video summarization by visual co-occurrence. In CVPR, 2015. 8", + "[14] Hyungjin Chung, Byeongsu Sim, and Jong Chul Ye. Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In CVPR, 2022. 8", + "[15] Sandra Eliza Fontes De Avila, Ana Paula Brandao Lopes, Antonio da Luz Jr, and Arnaldo de Albuquerque Araujo. Vsumm: A mechanism designed to produce static video summaries and a novel evaluation method. Pattern recognition letters, 32(1):56-68, 2011. 8", + "[16] Patrick Esser, Johnathan Chiu, Parmida Atighechian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In ICCV, 2023. 8", + "[17] Jiri Fajtl, Hajar Sadeghi Sokeh, Vasileios Argyriou, Dorothy Monekosso, and Paolo Remagnino. Summarizing videos with attention. In ACCV Workshop, 2019. 1, 2, 4, 7, 8, i, iii", + "[18] Junaid Ahmed Ghauri, Sherzod Hakimov, and Ralph Ewerth. Supervised video summarization via multiple feature sets with parallel attention. In ICME, 2021. 8, ii", + "[19] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Fei-Fei Li, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. In ECCV, 2025. 8", + "[20] Michael Gygli, Helmut Grabner, Hayko Riemenschneider, and Luc Van Gool. Creating summaries from user videos. In ECCV, 2014. 1, 4, 5", + "[21] Jaehoon Hahm, Junho Lee, Sunghyun Kim, and Joonseok Lee. Isometric representation learning for disentangled latent space of diffusion models. In ICML, 2024. 8", + "[22] Bo He, Jun Wang, Jielin Qiu, Trung Bui, Abhinav Shrivastava, and Zhaowen Wang. Align and attend: Multimodal summarization with dual contrastive losses. In CVPR, 2023. 1, 4, 7, 8, i, iii", + "[23] Mhand Hifi, Hedi Mhalla, and Slim Sadfi. Sensitivity of the optimum to perturbations of the profit or weight of an item in the binary knapsack problem. Journal of Combinatorial Optimization, 10:239-260, 2005. 4, 6", + "[24] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshop on Deep Generative Models and Downstream Applications, 2021. 1, 7, 8, i", + "[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 3, 8", + "[26] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv:2210.02303, 2022. 8", + "[27] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In NeurIPS, 2022. 8", + "[28] Fa-Ting Hong, Xuanteng Huang, Wei-Hong Li, and Wei-Shi Zheng. Mini-net: Multiple instance ranking network for video highlight detection. In ECCV, 2020. 4" + ], + "bbox": [ + 516, + 90, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15104", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Richang Hong, Jinhui Tang, Hung-Khoon Tan, Shuicheng Yan, Chongwah Ngo, and Tat-Seng Chua. Event driven summarization for web videos. In SIGMM workshop on Social media, 2009. 8", + "[30] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seunglyong Kim. Improving sample quality of diffusion models using self-attention guidance. In ICCV, 2023. 7, 8, i", + "[31] Tobias Hörppe, Arash Mehrjou, Stefan Bauer, Didrik Nielsen, and Andrea Dittadi. Diffusion models for video prediction and infilling. arXiv:2206.07696, 2022. 8", + "[32] Tzu-Chun Hsu, Yi-Sheng Liao, and Chun-Rong Huang. Video summarization with spatiotemporal vision transformer. IEEE Transactions on Image Processing, 2023. 8, ii", + "[33] Jia-Hong Huang, Luka Murn, Marta Mrak, and Marcel Worring. GPT2MVS: Generative pre-trained transformer-2 for multi-modal video summarization. In ICMR, 2021. 8", + "[34] Hao Jiang and Yadong Mu. Joint video summarization and moment localization by cross-task sample transfer. In CVPR, 2022. 1, 4, 8, iii", + "[35] Yunjae Jung, Donghyeon Cho, Sanghyun Woo, and In So Kweon. Global-and-local relative position embedding for unsupervised video summarization. In ECCV, 2020. 8", + "[36] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. In NeurIPS, 2022. 8", + "[37] Maurice G Kendall. The treatment of ties in ranking problems. Biometrika, 33(3):239-251, 1945. 4", + "[38] Aditya Khosla, Raffay Hamid, Chih-Jen Lin, and Neel Sundaresan. Large-scale video summarization using web-image priors. In CVPR, 2013. 8", + "[39] Haopeng Li, Qiuhong Ke, Mingming Gong, and Rui Zhang. Video joint modelling based on hierarchical transformer for co-summulation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3904–3917, 2022. 8", + "[40] Haopeng Li, Qiuhong Ke, Mingming Gong, and Tom Drummond. Progressive video summarization via multimodal self-supervised learning. In WACV, 2023. 4, 8, ii", + "[41] Judith Li, Xuchan Bao, Zhong Yi Wan, Kun Su, Timo Denk, Dima Kuzmin, Joonseok Lee, and Fei Sha. Diff4steer: Steerable diffusion prior for generative music retrieval with semantic guidance. In ICASSP, 2024. 8", + "[42] Ping Li, Qinghao Ye, Luming Zhang, Li Yuan, Xianghua Xu, and Ling Shao. Exploring global diverse attention via pairwise temporal relation for video summarization. Pattern Recognition, 111:107677, 2021. 8", + "[43] Pandeng Li, Chen-Wei Xie, Hongtao Xie, Liming Zhao, Lei Zhang, Yun Zheng, Deli Zhao, and Yongdong Zhang. MomentDiff: Generative video moment retrieval from random to real. In NeurIPS, 2024. 8", + "[44] Jingyun Liang, Yuchen Fan, Kai Zhang, Radu Timofte, Luc Van Gool, and Rakesh Ranjan. MoVideo: Motion-aware video generation with diffusion model. In ECCV, 2025. 8", + "[45] Daochang Liu, Qiyue Li, Anh-Dung Dinh, Tingting Jiang, Mubarak Shah, and Chang Xu. Diffusion action segmentation. In ICCV, 2023. 8" + ], + "bbox": [ + 91, + 90, + 485, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2017. ii", + "[47] Zheng Lu and Kristen Grauman. Story-driven summarization for egocentric video. In CVPR, 2013. 8", + "[48] Dezhao Luo, Shaogang Gong, Jiabo Huang, Hailin Jin, and Yang Liu. Generative video diffusion for unseen cross-domain video moment retrieval. arXiv:2401.13329, 2024. 8", + "[49] Yu-Fei Ma, Lie Lu, Hong-Jiang Zhang, and Mingjing Li. A user attention model for video summarization. In ACM MM, 2002. 8", + "[50] Behrooz Mahasseni, Michael Lam, and Sinisa Todorovic. Unsupervised video summarization with adversarial LSTM networks. In CVPR, 2017. 1, 4, 8, ii, iii", + "[51] Padmavathi Mundur, Yong Rao, and Yelena Yesha. Keyframe-based video summarization using delaunay clustering. International Journal on Digital Libraries, 6:219-232, 2006. 8", + "[52] Medhini Narasimhan, Anna Rohrbach, and Trevor Darrell. Clip-it! language-guided video summarization. In NeurIPS, 2021. 4, 8", + "[53] Chong-Wah Ngo, Yu-Fei Ma, and Hong-Jiang Zhang. Automatic video summarization by graph modeling. In ICCV, 2003. 8", + "[54] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 1, 8", + "[55] Mayu Otani, Yuta Nakashima, Esa Rahtu, and Janne Heikkila. Rethinking the evaluation of video summaries. In CVPR, 2019. 4", + "[56] Rameswar Panda and Amit K Roy-Chowdhury. Collaborative summarization of topic-related videos. In CVPR, 2017. 4", + "[57] Rameswar Panda, Abir Das, Ziyan Wu, Jan Ernst, and Amit K Roy-Chowdhury. Weakly supervised summarization of web videos. In ICCV, 2017. 8", + "[58] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 2, 3, 7", + "[59] Danila Potapov, Matthijs Douze, Zaid Harchaoui, and Cordelia Schmid. Category-specific video summarization. In ECCV, 2014. 4, 8", + "[60] Mrigank Rochan and Yang Wang. Video summarization by learning from unpaired data. In CVPR, 2019. 8", + "[61] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 8", + "[62] Ludan Ruan, Yiyang Ma, Huan Yang, Huiguo He, Bei Liu, Jianlong Fu, Nicholas Jing Yuan, Qin Jin, and Baining Guo. MM-diffusion: Learning multi-modal diffusion models for joint audio and video generation. In CVPR, 2023. 8", + "[63] Suho Ryu, Kihyun Kim, Eugene Baek, Dongsoo Shin, and Joonseok Lee. Towards scalable human-aligned benchmark for text-guided image editing. In CVPR, 2025. 8", + "[64] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour," + ], + "bbox": [ + 514, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15105", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 8", + "[65] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE transactions on pattern analysis and machine intelligence, 45(4):4713-4726, 2022. 8", + "[66] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 8", + "[67] Jaewon Son, Jaehun Park, and Kwangsu Kim. CSTA: CNN-based spatiotemporal attention for video summarization. In CVPR, 2024. 1, 2, 4, 5, 7, 8, i, ii, iii", + "[68] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In *ICLR*, 2020. 3, 8, ii", + "[69] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In NeurIPS, 2019. 8", + "[70] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. In NeurIPS, 2020. 8", + "[71] Yale Song, Jordi Vallmitjana, Amanda Stent, and Alejandro Jaimes. TVSum: Summarizing web videos using titles. In CVPR, 2015. 1, 4, 5, 8", + "[72] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2020. 1", + "[73] Kun Su, Judith Li, Qingqing Huang, Dima Kuzmin, Joon-seok Lee, Chris Donahue, Fei Sha, Aren Jansen, Yu Wang, Mauro Verzetti, and Timo Denk. V2Meow: Meowing to the visual beat via video-to-music generation. In AAAI, 2024. 8", + "[74] Jinhwan Sul, Jihoon Han, and Joonseok Lee. Mr. HiSum: A large-scale dataset for video highlight detection and summarization. In NeurlPS, 2024. 4, 5, 7", + "[75] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In CVPR, 2015. 4", + "[76] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016. 4", + "[77] Hacene Terbouche, Maryan Morel, Mariano Rodriguez, and Alice Othmani. Multi-annotation attention model for video summarization. In CVPR, 2023. 4, 8", + "[78] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. ii", + "[79] Shuyuan Tu, Qi Dai, Zhi-Qi Cheng, Han Hu, Xintong Han, Zuxuan Wu, and Yu-Gang Jiang. MotionEditor: Editing video motion via content-aware diffusion. In CVPR, 2024. 8", + "[80] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, 2017. 2, 3, 7" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[81] Junyan Wang, Yang Bai, Yang Long, Bingzhang Hu, Zhenhua Chai, Yu Guan, and Xiaolin Wei. Query twice: Dual mixture attention meta learning for video summarization. In ACM MM, 2020. 4, 5, 8", + "[82] Minghao Xu, Hang Wang, Bingbing Ni, Riheng Zhu, Zhenbang Sun, and Changhu Wang. Cross-category video highlight detection via set-based learning. In ICCV, 2021. 1, 4, 7, 8, i, iii", + "[83] Hong Jiang Zhang, Jianhua Wu, Di Zhong, and Stephen W Smoliar. An integrated system for content-based video retrieval and browsing. Pattern recognition, 30(4):643-658, 1997. 8", + "[84] Ke Zhang, Wei-Lun Chao, Fei Sha, and Kristen Grauman. Video summarization with long short-term memory. In ECCV, 2016. 4, 8, ii", + "[85] Yunzuo Zhang, Yameng Liu, Weili Kang, and Ran Tao. VSS-Net: visual semantic self-mining network for video summarization. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 4, 8", + "[86] Zhicheng Zhang, Junyao Hu, Wentao Cheng, Danda Paudel, and Jufeng Yang. ExtDM: Distribution extrapolation diffusion model for video prediction. In CVPR, 2024. 8", + "[87] Zhixing Zhang, Bichen Wu, Xiaoyan Wang, Yaqiao Luo, Luxin Zhang, Yinan Zhao, Peter Vajda, Dimitris Metaxas, and Licheng Yu. AVID: Any-length video inpainting with diffusion model. In CVPR, 2024. 8", + "[88] Bin Zhao, Xuelong Li, and Xiaoqiang Lu. Hierarchical recurrent neural network for video summarization. In ACM MM, 2017. 8", + "[89] Bin Zhao, Xuelong Li, and Xiaogiang Lu. HSA-RNN: Hierarchical structure-adaptive rnn for video summarization. In CVPR, 2018.", + "[90] Bin Zhao, Xuelong Li, and Xiaogiang Lu. TTH-RNN: Tensor-train hierarchical recurrent neural network for video summarization. IEEE Transactions on Industrial Electronics, 68(4):3629-3637, 2020. 8", + "[91] Bin Zhao, Maoguo Gong, and Xuelong Li. Hierarchical multimodal transformer to summarize videos. Neurocomputing, 468:360-369, 2022. 8", + "[92] Chuanwei Zhou, Chunyan Xu, Jun Li, Zhen Cui, and Jian Yang. Quality-aware pattern diffusion for video object segmentation. Neurocomputing, 528:148-159, 2023. 8", + "[93] Wencheng Zhu, Jiwen Lu, Jiahao Li, and Jie Zhou. DSNet: A flexible detect-to-summarize network for video summarization. IEEE Transactions on Image Processing, 30:948–962, 2020. 4, 8", + "[94] Wencheng Zhu, Yucheng Han, Jiwen Lu, and Jie Zhou. Relational reasoning over spatial-temporal graphs for video summarization. IEEE Transactions on Image Processing, 31: 3017-3031, 2022. 8", + "[95] Zixin Zhu, Xuelu Feng, Dongdong Chen, Junsong Yuan, Chunming Qiao, and Gang Hua. Exploring pre-trained text-to-video diffusion models for referring video object segmentation. arXiv:2403.12042, 2024. 8", + "[96] Daniel Zwillinger and Stephen Kokoska. CRC standard probability and statistics tables and formulae. CRC Press, 1999. 4" + ], + "bbox": [ + 516, + 90, + 903, + 897 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15106", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SummDiff_ Generative Modeling of Video Summarization with Diffusion/314d20e1-b991-4676-b961-e90cddf3f9d6_model.json b/2025/SummDiff_ Generative Modeling of Video Summarization with Diffusion/314d20e1-b991-4676-b961-e90cddf3f9d6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7ad0179d3b4b1909087192a378d09c9f7e45b0ff --- /dev/null +++ b/2025/SummDiff_ Generative Modeling of Video Summarization with Diffusion/314d20e1-b991-4676-b961-e90cddf3f9d6_model.json @@ -0,0 +1,2829 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.131, + 0.131, + 0.868, + 0.152 + ], + "angle": 0, + "content": "SummDiff: Generative Modeling of Video Summarization with Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.18, + 0.88, + 0.217 + ], + "angle": 0, + "content": "Kwanseok Kim\\(^{1*}\\), Jaehoon Hahm\\(^{2*}\\), Sumin Kim\\(^{1}\\), Jinhwan Sul\\(^{3}\\), Byunghak Kim\\(^{4}\\), Joonseok Lee\\(^{1\\dagger}\\), Seoul National University, 2UIUC, Georgia Institute of Technology, Hyundai Card" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.22, + 0.751, + 0.252 + ], + "angle": 0, + "content": "kjvd1009@snu.ac.kr, jh141@illinois.edu, sumink@snu.ac.kr, jsul7@gatech.edu, byunghak.kim@hcs.com, joonseok@snu.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.32, + 0.486, + 0.637 + ], + "angle": 0, + "content": "Video summarization is a task of shortening a video by choosing a subset of frames while preserving its essential moments. Despite the innate subjectivity of the task, previous works have deterministically regressed to an averaged frame score over multiple raters, ignoring the inherent subjectivity of what constitutes a \"good\" summary. We propose a novel problem formulation by framing video summarization as a conditional generation task, allowing a model to learn the distribution of good summaries and to generate multiple plausible summaries that better reflect varying human perspectives. Adopting diffusion models for the first time in video summarization, our proposed method, SummDiff, dynamically adapts to visual contexts and generates multiple candidate summaries conditioned on the input video. Extensive experiments demonstrate that SummDiff not only achieves the state-of-the-art performance on various benchmarks but also produces summaries that closely align with individual annotator preferences. Moreover, we provide a deeper insight with novel metrics from an analysis of the knapsack, which is an important last step of generating summaries but has been overlooked in evaluation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.669, + 0.222, + 0.685 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.696, + 0.485, + 0.862 + ], + "angle": 0, + "content": "Recently, short-form videos draw significant attention on video sharing platforms, with a trend that consumers increasingly prefer to quickly grasp the content. They often compressively convey contents that are originally in a longer form, summarizing the core contents into a shorter one; e.g., sport games highlights or movie summarization. This task of selecting core parts of a long video to construct a shorter one is called video summarization. This task is inherently subjective, since there can be multiple criteria for a 'good summary'; e.g., comprehensively covering the entire storyline or subjectively selecting impressive parts of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.907, + 0.349 + ], + "angle": 0, + "content": "the video (highlight detection). Due to this inherent subjectivity, most video summarization or highlight detection datasets [20, 71] offer annotations by multiple raters to reflect various perspectives." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.352, + 0.909, + 0.577 + ], + "angle": 0, + "content": "Since each annotator may have different opinion on the importance of a frame, most existing methods [3, 17, 22, 34, 67, 82] take the frame-level importance scores averaged across multiple annotators as their target label, and are trained to predict them. This frame-level score aggregation looks reasonable in some sense, but in fact it loses the various perspectives to summarize each video. For instance, suppose half of the annotators select clips from the first quarter of the video while the other half select clips from the last quarter. If one simply averages their frame scores, both the first and last quarters end up with similar importance, obscuring the two distinct valid summaries. That is, this simple regression to the averaged frame-level importance scores fail to preserve multiple viewpoints to summarize the video." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.909, + 0.807 + ], + "angle": 0, + "content": "In order to preserve and reflect various views to summarize a video, we pose a distribution of its good summaries and let the model to learn it, instead of giving an already-aggregated single ground truth importance score. Then, the video summarization task can be seen from a generative perspective; i.e., a process of learning and inferring the distribution of good summaries conditioned on the input video. Specifically, the model is now in charge of estimating the distribution of plausible summaries for the given video. Once trained, it allows us to sample multiple summaries from the estimated conditional data manifold. To the best of our knowledge, this generative approach to the video summarization has not been extensively studied, except for a few works [2, 50] that applied adversarial losses to construct a summary that looks like the original video." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Formulating the video summarization problem as a conditional generation task, we propose to adopt the generative diffusion [25, 72] mechanism, which has been successfully applied to various conditional generation tasks [24, 54, 61]. Specifically, conditioned on the input video, our proposed SummDiff model learns to denoise a random importance" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.221, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.235, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15096" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.086, + 0.616, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.258, + 0.616, + 0.287 + ], + "angle": 0, + "content": "Figure 1. Overview of SummDiff. Given an input video, SummDiff generates importance scores conditioned on video frames. \\( T \\) denotes the number of DDIM steps." + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.071, + 0.903, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.628, + 0.244, + 0.903, + 0.287 + ], + "angle": 0, + "content": "Figure 2. Video Importance Score Denoiser. We use AdaLN layer to inject time and positional conditions, following [58]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.303, + 0.484, + 0.422 + ], + "angle": 0, + "content": "score vector over the given video into an importance score vector sampled from the distribution that corresponds to a good summary of the video. In contrast to the previous deterministic methods, our approach allows to sample multiple plausible summaries for a given video starting from a different random noise, better aligned with the subjective nature of summarization where we usually have various true labels reflecting multiple views." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.484, + 0.514 + ], + "angle": 0, + "content": "Extensive experiments demonstrate that our model outperforms existing baselines across multiple datasets. Also, we revisit the knapsack, a relatively unexplored step in the summarization evaluation in spite of its nontrivial impact on the performance, and propose additional novel metrics based on this analysis to provide deeper insights." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.515, + 0.435, + 0.528 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.53, + 0.483, + 0.575 + ], + "angle": 0, + "content": "- We propose a novel generative viewpoint of video summarization, better suited for the subjective nature of the task, allowing multiple plausible summaries for a video." + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.576, + 0.484, + 0.62 + ], + "angle": 0, + "content": "- We innovatively apply diffusion to the video summarization for the first time, integrating learning the distribution of good summaries into the model." + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.62, + 0.484, + 0.665 + ], + "angle": 0, + "content": "- We analyze the knapsack optimization process and propose additional metrics to quantify the optimality of the predicted importance scores." + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.53, + 0.484, + 0.665 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.679, + 0.296, + 0.695 + ], + "angle": 0, + "content": "2. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.484, + 0.779 + ], + "angle": 0, + "content": "Given a video of \\(N\\) frames, the objective of video summarization is to identify and select \\(S < N\\) frames that effectively encapsulate the essence of the video content. Let \\(\\mathbf{X} \\equiv \\{\\mathbf{X}_i \\in \\mathbb{R}^{H \\times W \\times 3} : i = 0, \\dots, N - 1\\} \\in \\mathbb{R}^{N \\times H \\times W \\times 3}\\) be a video, where \\(H, W\\) denote the size of the frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.902 + ], + "angle": 0, + "content": "A video summary, \\(\\mathbf{y} \\equiv \\{y_i \\in \\{0,1\\} : i = 0,\\dots,N-1, \\sum_i y_i \\leq S\\} \\in \\{0,1\\}^N\\), indicates inclusion (1) or exclusion (0) of each frame. When a video is provided with multiple ground truth annotations, we denote each individual score by \\(\\mathbf{s}^{(r)} \\in [0,1]^N\\) and the corresponding binary summary by \\(\\mathbf{y}^{(r)} \\in \\{0,1\\}^N\\) which is obtained following the procedure explained in Sec. 3.3. The predicted summary is denoted by \\(\\hat{\\mathbf{y}} \\equiv \\{\\hat{y}_i \\in \\mathbb{R} : i = 0,\\dots,N-1\\} \\in \\mathbb{R}^N\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.303, + 0.907, + 0.486 + ], + "angle": 0, + "content": "Previous models [3, 17, 67] have approached video summarization as a regression task, aiming to predict the importance score \\(\\mathbf{s} \\in [0,1]^N\\) for each video, often the average of multiple importance scores, \\(\\frac{1}{|R|} \\sum_{r \\in \\mathcal{R}} \\mathbf{s}^{(r)}\\). In contrast, our approach allows multiple summaries for each video, aiming to learn the distribution of its plausible summaries. The previous scheme can be seen as a special case of ours, where every video has a single golden way of summarization, pointed at \\(\\mathbf{s}\\) with zero variance. Under our extended setting, multiple importance scores \\(\\{\\mathbf{s}^{(r)} | r \\in \\mathcal{R}\\}\\) can be given to the same conditioning video \\(\\mathbf{X}\\), forming a probability distribution of plausible importance scores." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.5, + 0.859, + 0.516 + ], + "angle": 0, + "content": "3. Diffusion-based Video Summarization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.526, + 0.907, + 0.587 + ], + "angle": 0, + "content": "Posing the video summarization as a conditional generation task, we introduce our SummDiff model, designed to adapt the distribution of individual importance scores for a given video by learning to denoise." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.598, + 0.838, + 0.614 + ], + "angle": 0, + "content": "3.1. Overall Flow of the Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.621, + 0.895, + 0.636 + ], + "angle": 0, + "content": "Fig. 1 illustrates the overall flow of our SummDiff model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.906, + 0.731 + ], + "angle": 0, + "content": "Encoding Process. We first encode each frame \\(\\mathbf{X}_i\\) for \\(i = 1, \\dots, N\\) from the input video \\(\\mathbf{X}\\) using a pre-trained image encoder. The extracted features are further contextualized through self-attention [80], as seen in Fig. 1. We denote the encoded feature for each individual frame by \\(\\mathbf{z}_i \\in \\mathbb{R}^D\\), and collectively the entire feature matrix by \\(\\mathbf{Z} \\in \\mathbb{R}^{N \\times D}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.905 + ], + "angle": 0, + "content": "Denoising Process. We then learn to denoise an individual importance score vector from a random noise, conditioned on the visual embeddings. First, the forward process adds noise to the ground truth individual importance score \\(\\mathbf{s}_0 \\equiv \\mathbf{s}^{(r)}\\) and sample a noised one \\(\\mathbf{s}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{s}_0 + \\sqrt{1 - \\bar{\\alpha}_t}\\boldsymbol{\\epsilon}_t\\), where \\(\\boldsymbol{\\epsilon}_t \\sim \\mathcal{N}(0,\\mathbf{I}), \\bar{\\alpha}_t = \\prod_{\\tau=1}^t (1 - \\beta_\\tau)\\), and \\(t\\) is the diffusion time step. The perturbation kernel at \\(t\\), defined as \\(q(\\mathbf{s}_t|\\mathbf{s}_{t-1}) = \\mathcal{N}(\\mathbf{s}_t; \\sqrt{1 - \\beta_t}\\mathbf{s}_{t-1}, \\beta_t\\mathbf{I})\\), where \\(\\beta_t\\) is defined by the variance schedule. Since \\(\\mathbf{s}_0\\) is bounded within \\([0,1]\\), it is not straightforward to add Gaussian noise directly to it so we first transform it to its logit, \\(\\mathbf{u}_0 = \\log \\frac{\\mathbf{s}_0}{1 - \\mathbf{s}_0} \\in\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15097" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.135 + ], + "angle": 0, + "content": "\\(\\mathbb{R}^N\\), and perform the noising process in this logit space. For numerical stability, we clip \\(\\mathbf{s}_0\\) to \\([\\epsilon, 1 - \\epsilon]\\), where \\(\\epsilon\\) is a small constant." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.137, + 0.482, + 0.242 + ], + "angle": 0, + "content": "Then, the reverse process progressively removes noise from \\(\\mathbf{s}_t\\) to \\(\\mathbf{s}_0\\), formulated by \\(p_{\\theta}(\\mathbf{s}_{t - 1}|\\mathbf{s}_t) = \\mathcal{N}(\\mathbf{s}_{t - 1};\\boldsymbol {\\mu}_\\theta (\\mathbf{s}_t,t),\\sigma_t^2\\mathbf{I})\\), where \\(\\sigma_t^2\\) is determined in relation to \\(\\beta_{t}\\), and the posterior mean \\(\\pmb{\\mu}_{\\theta}(\\mathbf{s}_t,t)\\) is modeled with a trainable neural network. In this paper, we extend this estimator of posterior mean to be conditioned on a video, similarly to the conditional generation on diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.482, + 0.318 + ], + "angle": 0, + "content": "Once trained, our denoiser is able to recover a plausible importance score for a video from a random noise, which is used to generate a summary (Sec. 3.3). Repeated generation across the noise distribution would converge to the true distribution of plausible scores for the video." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.328, + 0.385, + 0.344 + ], + "angle": 0, + "content": "3.2. Video Importance Score Denoiser" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.35, + 0.484, + 0.472 + ], + "angle": 0, + "content": "We formulate the key component of our method, the video importance score denoiser in Fig. 2. Starting from a noised logit score vector \\(\\mathbf{u}_t \\in \\mathbb{R}^N\\), it predicts a plausible importance score, conditioned on the given video. Denoted by \\(f_{\\theta}(\\mathbf{u}_t, t, \\mathbf{Z})\\), it learns to denoise the given score vector \\(\\mathbf{u}_t\\) at the diffusion time step \\(t\\) under the visual condition \\(\\mathbf{Z}\\), producing the denoised score \\(\\hat{\\mathbf{u}}_{t-1}\\) by one time step, where \\(\\theta\\) is the set of its learnable parameters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.475, + 0.484, + 0.595 + ], + "angle": 0, + "content": "Transformer-based Diffusion. We denoise the logit-transformed score \\(\\mathbf{u}_t\\) given \\(t\\) and \\(\\mathbf{Z}\\), predicting \\(\\hat{\\mathbf{u}}_{t-1}\\), using a transformer-based cross-attention [80]. The logit-transformed score \\(\\mathbf{u}_t\\) acts as the query, and the visual condition \\(\\mathbf{Z}\\) is used as key/value. This setup allows the model to effectively denoise \\(\\mathbf{u}_t\\) conditioned on the information from the input video, ensuring consistency between the predicted logit score \\(\\hat{\\mathbf{u}}_{t-1}\\) and the condition \\(\\mathbf{Z}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.596, + 0.484, + 0.776 + ], + "angle": 0, + "content": "To apply dot-product cross-attention, the dimensionality should match for queries and keys. To this end, we quantize the importance scores into a predefined number \\((K)\\) of uniform segments. Specifically, we convert the noised logit \\(\\mathbf{u}_t\\) back to its original bounded range \\([0,1]\\) by applying sigmoid \\(1 / (1 + e^{-\\mathbf{u}_t})\\), and divide them into \\(K\\) equally-binned segments. Each score range is associated with a learnable embedding of size \\(D\\) (codebook in Fig. 1). Based on the codebook \\(\\mathcal{C}\\), we map the scores \\(\\mathbf{u}_t \\in \\mathbb{R}^N\\) to their corresponding quantized embedding \\(\\mathcal{C}(\\mathbf{u}_t) \\in \\mathbb{R}^{N \\times D}\\). Our denoiser \\(f_{\\theta}(\\mathcal{C}(\\mathbf{u}_t), t, \\mathbf{Z})\\) takes this \\(\\mathcal{C}(\\mathbf{u}_t)\\) as input, instead of \\(\\mathbf{u}_t\\). See Sec. 4.5 for analysis on the size of codebook." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.855 + ], + "angle": 0, + "content": "Query Formation in Cross-Attention. Following [25], the time step \\(t\\) is embedded as \\(\\pmb{\\tau} \\in \\mathbb{R}^{D}\\) using sinusoidal functions and an MLP. Considering the sequential nature of videos, we also introduce temporal positional embeddings \\(\\Phi \\in \\mathbb{R}^{N \\times D}\\) using sinusoidal functions." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The simplest way to set the query in the cross-attention would be to add all inputs except for the condition \\(\\mathbf{Z}\\); that is, \\(\\mathbf{Q}_t = \\mathcal{C}(\\mathbf{u}_t) + \\boldsymbol{\\tau} + \\boldsymbol{\\Phi}\\). However, we leverage AdaLN-Zero" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.167 + ], + "angle": 0, + "content": "block [58] to integrate the time embedding \\(\\pmb{\\tau}\\) and temporal positional encoding \\(\\Phi\\) more effectively by separating them from the query \\(\\mathbf{Q}_t\\), preventing information mixing. Hence, the query becomes \\(\\mathbf{Q}_t = \\mathcal{C}(\\mathbf{u}_t)\\) and \\(\\pmb{\\tau}, \\Phi\\) are conditioned via scale-shift operation. See Sec. 4.5 for ablation studies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.906, + 0.275 + ], + "angle": 0, + "content": "We conduct cross-attention with \\(\\mathbf{Q}_t\\) as queries and visual conditions \\(\\mathbf{Z} \\in \\mathbb{R}^{N \\times D}\\) as keys \\(\\mathbf{K}\\) and values \\(\\mathbf{V}\\). We take an MLP from the AdaLN output to regress the scale and shift parameters, denoted by \\(\\mathbf{A}_1\\), \\(\\mathbf{B}_1\\), \\(\\Gamma_1\\), \\(\\mathbf{A}_2\\), \\(\\mathbf{B}_2\\), and \\(\\Gamma_2 \\in \\mathbb{R}^{N \\times D}\\). As depicted in Fig. 2, they scale and shift \\(\\mathbf{Q}_t\\), \\(\\mathbf{K}\\), and \\(\\mathbf{V}\\). Then, they are passed through cross-attention with skip connections and a subsequent rescaling. Formally," + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.287, + 0.762, + 0.302 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} ^ {i ^ {\\prime}} = \\boldsymbol {\\Gamma} _ {1} \\odot \\mathbf {H} ^ {i} + \\mathbf {H} ^ {i} + \\mathbf {B} _ {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.308, + 0.858, + 0.327 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {1} = \\mathbf {A} _ {1} \\odot \\operatorname {s o f t m a x} \\left(\\mathbf {Q} _ {t} ^ {i ^ {\\prime}} \\mathbf {K} ^ {i ^ {\\prime} \\top}\\right) \\mathbf {V} ^ {i ^ {\\prime}} + \\mathbf {Q} _ {t} ^ {i ^ {\\prime}}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.331, + 0.765, + 0.345 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {2} = \\boldsymbol {\\Gamma} _ {2} \\odot \\mathbf {X} _ {1} + \\mathbf {X} _ {1} + \\mathbf {B} _ {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.35, + 0.778, + 0.367 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} _ {t} ^ {i + 1} = \\mathbf {A} _ {2} \\odot \\operatorname {M L P} (\\mathbf {X} _ {2}) + \\mathbf {X} _ {2},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.905, + 0.431 + ], + "angle": 0, + "content": "where \\(\\mathbf{H}^{i^{(t)}}\\in \\{\\mathbf{Q}_t^{i^{(t)}},\\mathbf{K}^{i^{(t)}},\\mathbf{V}^{i^{(t)}}\\}\\) denotes the matrices used for attention, \\(\\odot\\) is the (broadcasted) Hadamard product, and \\(i = 1,\\ldots ,L\\) is the layer index." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.435, + 0.906, + 0.495 + ], + "angle": 0, + "content": "Training. We train the denoiser \\( f_{\\theta} \\) to estimate the true importance score \\( \\hat{\\mathbf{s}}_0 \\) after acting upon a fully-connected layer and a sigmoid function \\( \\sigma \\). We minimize the following loss [12] on each annotator's individual importance score:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.508, + 0.903, + 0.527 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} \\left(\\mathbf {s} _ {0}, \\hat {\\mathbf {s}} _ {0}\\right) = \\left\\| \\mathbf {s} _ {0} - \\hat {\\mathbf {s}} _ {0} \\right\\| _ {2} ^ {2} = \\left\\| \\mathbf {s} _ {0} - \\sigma \\left(\\operatorname {F C} \\left(f _ {\\theta} \\left(\\mathcal {C} \\left(\\mathbf {u} _ {t}\\right), t, \\mathbf {Z}\\right)\\right)\\right) \\right\\| _ {2} ^ {2}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.905, + 0.584 + ], + "angle": 0, + "content": "The learnable embeddings in the codebook \\(\\mathcal{C}\\) are compositely optimized, finding an effective representation for \\(\\mathbf{u}_t\\) during training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.589, + 0.905, + 0.649 + ], + "angle": 0, + "content": "Inference. Our model generates a logit-transformed importance score from a random noise \\(\\mathbf{u}_T \\sim \\mathcal{N}(0, \\mathbf{I})\\) for a given video. Employing the reverse diffusion process [68], it iteratively refines the logit score towards cleaner estimations:" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.662, + 0.882, + 0.715 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {\\mathbf {u}} _ {t - 1} = \\sqrt {1 - \\bar {\\alpha} _ {t - 1} - \\sigma_ {t} ^ {2}} \\frac {\\hat {\\mathbf {u}} _ {t} - \\sqrt {\\bar {\\alpha} _ {t}} f _ {\\theta} (\\mathcal {C} (\\mathbf {u} _ {t}) , t , \\mathbf {Z}))}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\\\ + \\sqrt {\\bar {\\alpha} _ {t - 1}} f _ {\\theta} (\\mathcal {C} (\\mathbf {u} _ {t}), t, \\mathbf {Z})) + \\sigma_ {t} \\epsilon_ {t} \\in \\mathbb {R} ^ {N}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.729, + 0.906, + 0.807 + ], + "angle": 0, + "content": "where \\(\\sigma_{t} = \\sqrt{(1 - \\bar{\\alpha}_{t - 1}) / (1 - \\bar{\\alpha}_{t})}\\sqrt{1 - \\bar{\\alpha}_{t} / \\bar{\\alpha}_{t - 1}}\\) (DDPM reverse process). When \\(t = 1\\) at the last step, \\(\\hat{\\mathbf{u}}_0 = f_\\theta (\\mathcal{C}(\\mathbf{u}_1),1,\\mathbf{Z})\\) is directly used to remove stochasticity at the end of inference process. Finally, we take sigmoid to \\(\\hat{\\mathbf{u}}_0\\) to derive the final importance score \\(\\hat{\\mathbf{s}}_0\\in (0,1)^N\\)" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.719, + 0.833 + ], + "angle": 0, + "content": "3.3. Summary Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "From the raw importance score \\(\\hat{\\mathbf{s}}_0\\), we apply the standard knapsack-based approach to decide which frames to be included in the final summary. To make this summary video more realistic, it is common to choose at a semantic clip" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15098" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.213 + ], + "angle": 0, + "content": "level instead of individual frame level. We adopt a widely-used Kernel Temporal Segmentation (KTS) [59, 84] to partition a video into disjoint temporal intervals. We take the average score among the corresponding frames as the clip-level importance; i.e., \\( v_{i} = \\sum_{j = t_{i}}^{t_{i + 1}}\\hat{\\mathbf{s}}_{0,j} / (t_{i + 1} - t_{i}) \\) for the \\( i \\)-th clip composed of frames from \\( t_{i} \\) to \\( t_{i + 1} \\). Subsequently, we select the clips based on \\( v_{i} \\) by solving the binary knapsack problem (KP) [23] with dynamic programming [71]:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.224, + 0.474, + 0.266 + ], + "angle": 0, + "content": "\\[\n\\mathcal {K P} (\\boldsymbol {v}, \\boldsymbol {w}, \\rho) \\equiv \\underset {\\boldsymbol {s} \\in \\{0, 1 \\} ^ {M}} {\\operatorname {a r g m a x}} \\sum_ {i = 1} ^ {M} v _ {i} s _ {i} \\quad \\text {s . t .} \\sum_ {i = 1} ^ {M} w _ {i} s _ {i} \\leq \\rho N,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.277, + 0.484, + 0.368 + ], + "angle": 0, + "content": "where \\(\\pmb{v},\\pmb{w},\\rho\\) denotes the values of each clip \\((v_{i}\\in [0,1])\\), costs of selecting each clip, and the budget constraint ratio (e.g., \\(\\rho = 0.15\\); \\(15\\%\\) of the video length), respectively. If the argmax has multiple solutions, we will abuse the notation and denote \\(\\mathcal{KP}(\\pmb {v},\\pmb {w},\\rho)\\) as an arbitrary element of the solution set." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.382, + 0.224, + 0.399 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.407, + 0.299, + 0.423 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.429, + 0.484, + 0.594 + ], + "angle": 0, + "content": "Datasets. We evaluate our method on three benchmarks. TVSum [71] and SumMe [20] are traditional datasets, containing 50 and 25 videos respectively and manually labeled annotations by up to 20 raters. Mr. HiSum [74] is a large-scale dataset, composed of 31,892 videos and importance scores derived from YouTube Most Replayed statistics. It provides an importance score \\( \\mathbf{s}_0 \\) and a corresponding summary \\( \\mathbf{y}_0 \\) for each video, averaged over 50,000+ viewers. We use the default features for each dataset, GoogLeNet [75] for TVSum and SumMe, and Inception-v3 [76] PCA-ed to 1024D [1] for Mr. HiSum." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.596, + 0.484, + 0.732 + ], + "angle": 0, + "content": "We adopt two dataset splits. First, we randomly split TVSum and SumMe into 7:1:2 for train, validation, and test set, and report averaged test scores for 5 random splits (namely, TVT). Also, following prior works [67, 77], we include five-fold cross-validation results (5FCV), although these may suffer from overfitting to the test set due to the lack of explicit validation set, and thus may not generalize well on unseen videos. For Mr. HiSum, we use its original train/validation/test split. We run all experiments 5 times." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Evaluation Metrics. Following recent practice [22, 40, 55, 67], we report rank order statistics, Kendall's \\(\\tau\\) [37] and Spearman's \\(\\rho\\) [96], to assess how well the model ranks frame importance. F1 score had been widely used in video summarization, but recent literature [67, 77] reports that it is excessively sensitive to video segmentation and unreasonably favors summaries composed of many short shots [55], while disregarding longer key shots [67]. As an alternative to F1, we propose additional metrics that represent the influence of importance score through an analysis of the knapsack algorithm in Sec. 4.4." + }, + { + "type": "table", + "bbox": [ + 0.545, + 0.089, + 0.875, + 0.232 + ], + "angle": 0, + "content": "
MethodSumMeTVSum
τρτρ
Random0.0000.0000.0000.000
Human0.2050.2130.1770.204
A2Summ [22]0.0880.0960.1570.206
VASNet [17]0.0890.0990.1530.205
PGL-SUM [3]0.1040.1160.1410.186
CSTA [67]0.1080.1200.1680.221
SummDiff (Ours)0.1330.1480.1730.226
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.235, + 0.907, + 0.278 + ], + "angle": 0, + "content": "Table 1. Comparison of models trained under TVT (train/val/test split) on SumMe [20] and TVSum [71]. Best and second-best results are boldfaced and underlined, respectively." + }, + { + "type": "table", + "bbox": [ + 0.543, + 0.283, + 0.875, + 0.557 + ], + "angle": 0, + "content": "
MethodSumMeTVSum
τρτρ
Random0.0000.0000.0000.000
Human0.2050.2130.1770.204
DSNet-AF [93]0.0370.0460.1130.138
DSNet-AB [93]0.0510.0590.1080.129
SUM-GAN [50]0.0490.0660.0240.031
AC-SUM-GAN [2]0.1020.0880.0310.041
CLIP-It [52]--0.1080.147
iPTNet [34]0.1010.1190.1340.163
A2Summ [22]0.1080.1290.1370.165
VASNet [17]0.1600.1700.1600.170
PGL-SUM [3]--0.1570.206
AAAM [77]--0.1690.223
MAAM [77]--0.1790.236
VSS-Net [85]--0.1900.249
DMASum [81]0.0630.0890.2030.267
SSPVS [40]0.1920.2570.1810.238
CSTA [67]0.2460.2740.1940.255
SummDiff (Ours)0.2560.2850.1950.255
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.561, + 0.907, + 0.604 + ], + "angle": 0, + "content": "Table 2. Comparison of models trained under 5FCV (5-Fold Cross Validation) on SumMe [20] and TVSum [71]. Training under the 5FCV setting tends to overfit the test set." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.609, + 0.892, + 0.751 + ], + "angle": 0, + "content": "
Modelτ↑ρ↑MAPρ=50%↑MAPρ=15%↑
SUM-GAN [50]0.0670.09556.6223.56
VASNet [17]0.0690.10258.6925.28
AC-SUM-GAN [2]0.0120.01855.3521.88
SL-module [82]0.0600.08858.6324.95
PGL-SUM [3]0.0970.14161.6027.45
iPTNet [34]0.0200.02955.5322.74
A2Summ [22]0.1210.17263.2032.34
CSTA [67]0.1280.18563.3830.42
SummDiff0.1750.23865.4433.83
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.755, + 0.907, + 0.784 + ], + "angle": 0, + "content": "Table 3. Evaluation on Mr. HiSum [74]. See Appendix D for the full table including standard deviations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We further evaluate our method on video highlight detection following [74]. First, we uniformly divide the input video into 5-second-long shots and calculate the average frame scores for each shot. The top \\(\\rho \\in \\{0.15, 0.5\\}\\) of these shots are designated as ground truth highlights. We report Mean Average Precision (MAP), following [28, 56, 84]. See Appendix E for more implementation details." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15099" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.286, + 0.106 + ], + "angle": 0, + "content": "4.2. Results and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.385 + ], + "angle": 0, + "content": "Considering the innate subjectivity of video summarization, we train our model to learn from each individual score, allowing it to capture multiple ways of summarization for each video. Specifically, we use 20 individual importance scores per video in TVSum [71] and 15-18 individual binary summaries in SumMe [20], since SumMe does not provide individual importance scores. Tab. 1 and 2 show that SummDiff achieves the best performance on SumMe and TVSum, with the sole exception of TVSum in 5FCV. While DMA-SUM [81] performs the best on TVSum in 5FCV, it significantly underperforms on SumMe, indicating that its high performance on TVSum does not generalize well. Comparing TVT and 5FCV, most models exhibit a considerable performance gap. This demonstrates how much existing models have overfitted to the test set under the conventional 5FCV protocol. We adopt TVT as a remedy to this, but due to the extremely small size of these datasets, the reliability of evaluation is still limited [74]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.386, + 0.484, + 0.522 + ], + "angle": 0, + "content": "On the larger Mr. HiSum, we train our model using the aggregated annotation, as it does not provide individual annotations. This evaluation mainly aims to verify scalability of SummDiff on a large-scale dataset. Tab. 3 presents the performance on Mr. HiSum, where SummDiff consistently outperforms all baselines across all metrics, significantly surpassing the strongest competitor, CSTA [67] with a large margin. These results highlight scalability and effectiveness of our method even under the single-label setting." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.532, + 0.333, + 0.547 + ], + "angle": 0, + "content": "4.3. Qualitative Demonstration" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We demonstrate how our method discovers a variety of possible summaries for a video. Specifically, we generate video summaries for the 50 videos in TVSum using CSTA, PGL-SUM, VASNet, and ours. For each generated summary, we measure the Kendall's \\(\\tau\\) with all 20 ground truth annotations, respectively. Considering \\(\\tau \\geq 0.25\\) as the threshold for the summary to be matched with the annotator, Fig. 3 shows if the generated summary for a video (row) matches with each annotator (column). As the baselines (CSTA, PGL-SUM, and VASNet) summarize a video deterministically, each cell is either completely matched (1) or not (0). If the summary matches with majority of annotators (e.g., video 29 for CSTA), it means the deterministic summary matches with the dominant way of summarization for the video. If it matches only with a few annotators (e.g., video 24 for CSTA), it means the generated summary matches with a minor opinion. For our method, we generate 100 summaries per video from different Gaussian noise vectors, and mark each cell with the ratio of summaries that match with each annotator. The darker the cell colors are for each row, the more various ways of summaries have been discovered by our method. Comparing the heat maps, we clearly observe that SummDiff generates significantly more various" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.083, + 0.706, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.083, + 0.882, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.242, + 0.706, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.242, + 0.882, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.409, + 0.907, + 0.492 + ], + "angle": 0, + "content": "Figure 3. Ratio of summaries with \\(\\tau \\geq 0.25\\) for each video-annotator pair. The heatmap illustrates how closely each method's summary matches with individual annotations. This comparison reveals the extent to which each method captures the varied human summaries. SummDiff covers a larger area of the heatmap, which indicates better coverage over various summaries." + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.492, + 0.71, + 0.615 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.492, + 0.882, + 0.615 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.54, + 0.616, + 0.712, + 0.74 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.616, + 0.882, + 0.74 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.576, + 0.744, + 0.85, + 0.753 + ], + "angle": 0, + "content": "SummDiff CSTA PGL_SUM VASNet Annotator" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.758, + 0.907, + 0.815 + ], + "angle": 0, + "content": "Figure 4. Distribution of true and predicted summaries of given videos. SummDiff can generate various summaries and cover the targeted distribution of summaries, while baselines deterministically predicts a single summary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.905, + 0.854 + ], + "angle": 0, + "content": "summaries covering multiple viewpoints, better reflecting the distribution of annotated summaries." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We further illustrate the variation and quality of generated summaries in Fig. 4. The summaries annotated by human raters ( \\(\\bullet\\) Annotator) and those generated by ours ( \\(\\bullet\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15100" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.284, + 0.907, + 0.327 + ], + "angle": 0, + "content": "Figure 5. Demonstration of video summaries generated by competing methods on a TVSum video. Shaded parts indicate the segments selected by each method, and the two rows of edged boxes within each method indicate the two distinct true annotations. The results clearly demonstrate the effectiveness of SummDiff in capturing multiple plausible summaries for a video. See another example in Appendix F." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.339, + 0.482, + 0.58 + ], + "angle": 0, + "content": "SummDiff) and baselines ( • CSTA, • PGL-SUM, and • VASNet) are projected onto the first two principal components, which are computed using only the human-annotated summaries. The distribution of true summaries are displayed using a contour map, obtained by Gaussian kernel density estimation. It demonstrates that our SummDiff generates multiple summaries closely aligned with the distribution of true annotations. For example, video 22 (top-left) roughly has three ways to summarize, according to the human raters. SummDiff generates summaries dominantly covering the two modes in the lower-left region, and one case in the upper-right mode. In contrast, baselines produce a single, less accurate summary, failing to account for the variety inherent in human-annotated summaries. We observe similar patterns in the other three plots as well. Fig. II provides more examples in Appendix C." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Fig. 5 illustrates example summaries generated by CSTA, A2Summ, PGL-SUM, VASNet, and our SummDiff, for a video selected from TVSum test set. Shaded parts indicate the segments selected by each method (note that SummDiff can generate two different summaries while baselines always produce a single summary), and the two rows of edged boxes within each method indicate two different true annotations. These results demonstrate that SummDiff produces more accurate and various summaries, effectively capturing multiple plausible summarizations for each video." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.742, + 0.468, + 0.76 + ], + "angle": 0, + "content": "4.4. Metrics Inspired from Knapsack Optimality" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In spite of nontrivial impact of the knapsack at the end of the summary generation, previous evaluation metrics have focused only on the accuracy of the importance scores. Through a thorough analysis on the knapsack problem (KP), we provide additional metrics that measure the contribution of importance scores. By accounting for knapsack optimality and clip-level weights, our new metrics are capable of assessing the predicted importance score more accurately than existing metrics such as F1 or ranking-based ones." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.339, + 0.905, + 0.413 + ], + "angle": 0, + "content": "Confidence of Importance Score. First, we analyze the conditions under which the same optimal KP solution remains valid despite some perturbations to the profits \\( \\pmb{v} \\). These perturbations can be a modeling to the imperfect estimation of the importance scores in video summarization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.414, + 0.907, + 0.673 + ], + "angle": 0, + "content": "Let \\(\\pmb{y}^{*}\\) be the optimal solution to the original KP (i.e., \\(\\pmb{y}^{*} = \\mathcal{KP}(\\pmb{v},\\pmb{w},\\rho)\\); see definition in Sec. 3.3) and \\(\\Gamma \\subseteq [N]\\equiv \\{n\\in \\mathbb{N}\\mid 1\\leq n\\leq N\\}\\) be an arbitrary subset of items associated with the perturbed profits \\(v_{i}^{\\prime} = v_{i} + \\Delta v_{i}\\), where \\(\\Delta v_{i} = 0\\) if \\(i\\notin \\Gamma\\) and \\(N\\) is the number of items. In our case, \\(\\pmb{v}\\in \\mathbb{R}^{N}\\) is the ground truth importance scores, \\(\\pmb{y}^{*}\\in [0,1]^{N}\\) is the ground truth summary, and the perturbed profit \\(\\pmb{v}^{\\prime}\\in \\mathbb{R}^{N}\\) corresponds to the imperfectly predicted importance scores. We further define two disjoint subsets of \\(\\Gamma\\). First, \\(\\Gamma_0^+ \\equiv \\{i\\in \\Gamma \\mid y_i^* = 0,\\Delta p_i > 0\\}\\) refers to the set of items excluded in the original optimal solution but undergoing increased profit, and \\(\\Gamma_1^- \\equiv \\{i\\in \\Gamma \\mid y_i^* = 1,\\Delta p_i < 0\\}\\) is the opposite, the set of items included in the optimal solution but experiencing a decreased profit. Then, Hifi et al. [23] claims that the solution \\(\\pmb{y}^{*}\\) remains optimal for the perturbed KP, i.e., \\(\\pmb{y}^{*} = \\mathcal{KP}(\\pmb{v},\\pmb{w},\\rho) = \\mathcal{KP}(\\pmb{v}^{\\prime},\\pmb{w},\\rho)\\), if" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.682, + 0.905, + 0.72 + ], + "angle": 0, + "content": "\\[\n\\sum_ {i \\in \\Gamma_ {0} ^ {+}} \\Delta v _ {i} - \\sum_ {i \\in \\Gamma_ {1} ^ {-}} \\Delta v _ {i} \\leq \\boldsymbol {v} \\cdot \\mathcal {K P} (\\boldsymbol {v}, \\boldsymbol {w}, \\rho) - \\max (\\Psi^ {+}, \\Psi^ {-}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.73, + 0.907, + 0.811 + ], + "angle": 0, + "content": "where \\(\\Psi^{+} \\equiv \\max_{i \\in \\Gamma_{0}^{+}} [\\pmb{v} \\cdot \\mathcal{KP}(\\pmb{v} - v_{i}\\hat{e}_{i}, \\pmb{w}, \\rho - w_{i}/N)]\\) is the maximum value of the KP when removing any item from \\(\\Gamma_{0}^{+}\\) and adjusting the constraint, and similarly, \\(\\Psi^{-} \\equiv \\max_{i \\in \\Gamma_{1}^{-}} [\\pmb{v} \\cdot \\mathcal{KP}(\\pmb{v} - v_{i}\\hat{e}_{i}, \\pmb{w}, \\rho)]\\) is the maximum value of the KP when removing any item from \\(\\Gamma_{1}^{-}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.812, + 0.905, + 0.857 + ], + "angle": 0, + "content": "Based on this theorem we propose a new metric, Confidence of Importance Score (CIS), to estimate how close the predicted scores are to the ground truth:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.867, + 0.925, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\operatorname {CIS} = \\sum_ {i \\in \\Gamma_ {0} ^ {+}} \\Delta v _ {i} - \\sum_ {i \\in \\Gamma_ {1} ^ {-}} \\Delta v _ {i} - \\boldsymbol {v} \\cdot \\mathcal {K P} (\\boldsymbol {v}, \\boldsymbol {w}, \\rho) + \\max \\left(\\Psi^ {+}, \\Psi^ {-}\\right).\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15101" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.091, + 0.482, + 0.212 + ], + "angle": 0, + "content": "If CIS \\(\\leq 0\\), the theorem indicates that the predicted importance score \\(\\hat{\\pmb{v}}\\) is guaranteed to induce the true KP solution \\(\\pmb{y}^{*}\\). If CIS \\(>0\\), \\(\\hat{\\pmb{v}}\\) does not guarantee to induce the KP solution, but a lower CIS would indicate a solution closer to \\(\\pmb{y}^{*}\\) than a higher one. That is, a smaller CIS indicates a stronger confidence in satisfying the inequality, indicating a higher chance for the solution of the perturbed KP to be identical to that of the original KP." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.212, + 0.482, + 0.303 + ], + "angle": 0, + "content": "Unlike F1 that simply measures how close the predicted summary is to the ground truth, our CIS accounts for the confidence that the predicted score will exactly induce the true summary. This is particularly important in video summarization, since the importance score itself is subjective and often noisy, requiring robust selection of frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.306, + 0.482, + 0.368 + ], + "angle": 0, + "content": "Weighted Inclusion Ratios. Analyzing the sensitivity of the KP optimum to the perturbation of the importance score, Belgacem and Hifi [5] established bounds of the perturbed profits to retain the original optimum. These intervals are" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.374, + 0.504, + 0.409 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {y} _ {i} ^ {*} = 1 \\rightarrow \\Delta v _ {i} \\in I _ {i} ^ {1} = \\left[ \\right. \\max \\left(\\Delta_ {i} ^ {-}, w _ {i} \\max _ {k \\in \\Gamma} \\mu_ {k} - p _ {i}\\right), + \\infty\\left. \\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.409, + 0.504, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {y} _ {i} ^ {*} = 0 \\rightarrow \\Delta v _ {i} \\in I _ {i} ^ {0} = \\left( \\right.- \\infty , \\min \\left(\\Delta_ {i} ^ {+}, p _ {i} - w _ {i} \\min _ {k \\in \\Gamma} \\mu_ {k}\\right)\\left. \\right],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.451, + 0.482, + 0.594 + ], + "angle": 0, + "content": "where \\(\\Delta_i^+ = \\pmb{v} \\cdot (\\mathcal{KP}(\\pmb{v}, \\pmb{w}, \\rho) - \\mathcal{KP}(\\pmb{v} - v_i \\hat{e}_i, \\pmb{w}, \\rho - w_i / N), \\Delta_i^- = \\pmb{v} \\cdot (\\mathcal{KP}(\\pmb{v} - v_i \\hat{e}_i, \\pmb{w}, \\rho) - \\mathcal{KP}(\\pmb{v}, \\pmb{w}, \\rho))\\), and \\(\\mu_k\\) denotes the critical ratio of \\(k \\in \\Gamma\\): \\(\\mu_k = \\mathcal{CR}([N] \\setminus \\{k\\}, \\rho - 1_{\\{y_k^* = 0\\}} \\cdot w_k / N)\\), with \\(\\mathcal{CR}(A, \\gamma) = v_s / w_s\\) and \\(s = \\min\\{l | \\sum_{i=1}^{l} v_s / w_s > \\rho\\}\\) is the index of the critical item. The items in \\(A\\) are considered in descending order of \\(v_s / w_s\\). To sum up, if the perturbed profit resides in the interval \\(v_i' \\in I_i^{y_i^*}\\), it is guaranteed that the solution of perturbed KP is identical with that of the original KP." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.595, + 0.482, + 0.738 + ], + "angle": 0, + "content": "We propose Inclusion Ratio (IR) to estimate how much the predicted scores fall into the safe bounds; for the \\(i\\)-th importance score, \\(\\mathrm{IR}_i \\equiv |\\{\\Delta v_i \\in I_i^{y_i^*} | i \\in [N]\\}| / N\\). This ratio reflects the proportion of the predicted scores that retain the KP solution. We propose the Weighted average of the Inclusion Ratios (WIR) as a metric to measure the proximity of the predicted score to the ground truth: \\(\\mathrm{WIR} \\equiv \\sum_{i=1}^{n} \\left( \\frac{w_i}{\\sum_{j=1}^{n} w_j} \\mathrm{IR}_i \\right)\\), where \\(w_i\\) denotes the segment length of each shot." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.743, + 0.482, + 0.789 + ], + "angle": 0, + "content": "Weighted Sum of Errors (WSE). Additionally, we compare a weighted sum of the errors \\(\\sum_{i} w_{i} \\Delta p_{i}\\), where \\(w_{i}\\) is the segment length of each clip." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.792, + 0.482, + 0.867 + ], + "angle": 0, + "content": "Comparison. Tab. 4 compares several competitive methods using these new metrics. Our SummDiff achieves the best performance in all metrics, indicating that the summaries generated by our method are closer to the optimal knapsack result, under the theoretical analysis in literature." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Discussion. Our newly proposed metrics, WIR and CIS, evaluate the quality of the generated importance scores" + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.085, + 0.897, + 0.198 + ], + "angle": 0, + "content": "
Methodτ↑ρ↑CIS ↓WIR ↑WSE ↓
Uniform-Random0.0000.0009.270.4546.82
SL-module [82]0.0600.0886.830.5630.09
CSTA [67]0.1280.1856.230.5725.76
PGL-SUM [3]0.0970.1416.140.5826.22
VASNet [17]0.0690.1026.250.5926.79
A2Summ [22]0.1210.1726.650.5530.55
SummDiff0.1750.2385.960.6125.24
" + }, + { + "type": "table_caption", + "bbox": [ + 0.521, + 0.202, + 0.897, + 0.215 + ], + "angle": 0, + "content": "Table 4. Evaluation using our new metrics proposed in Sec. 4.4." + }, + { + "type": "table", + "bbox": [ + 0.543, + 0.219, + 0.875, + 0.343 + ], + "angle": 0, + "content": "
Moduleτ↑ρ↑
Encoder Only0.0710.104
(+) Video Importance Denoiser0.0790.116
(+) Learnable Embedding0.0860.124
(+) AdaLN (τ)0.1450.204
(+) AdaLN (τ, Φ) 1 layer0.1710.232
(-) Quantization0.1250.174
(+) Classifier-free Guidance (CFG) [24]0.1750.238
(+) Self-attention Guidance (SAG) [30]0.1770.239
" + }, + { + "type": "table_caption", + "bbox": [ + 0.587, + 0.347, + 0.832, + 0.361 + ], + "angle": 0, + "content": "Table 5. Effect of individual components" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.37, + 0.905, + 0.476 + ], + "angle": 0, + "content": "in two aspects. WIR measures how many importance scores are individually trustworthy, weighted by the duration. Specifically, WIR first computes the safe interval \\( I_{i} \\) for each importance score such that the true summary remains unchanged. Then, given the predicted importance scores \\( \\hat{\\mathbf{v}} \\), it measures how many of them fall within their corresponding safe intervals \\( \\hat{v}_{i} \\in I_{i} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.906, + 0.627 + ], + "angle": 0, + "content": "On the other hand, CIS quantifies the risk that the predicted importance scores will lead to a different summary than the one generated using the true importance scores. In other words, CIS evaluates the overall chance how likely the predicted score vector \\(\\hat{\\mathbf{v}}\\), as a whole, is to give a summary different from the true one. With the analysis of knapsack optimality, these metrics would better measure the accuracy of predicted importance score, considering the fidelity of final summary after knapsack to the ground truth, than F1 or ranking-based metrics." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.637, + 0.667, + 0.654 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We report ablation study on Mr. HiSum [74] in Tab. 5, comparing performance adding a component step-by-step. Starting from a two-layer transformer encoder predicting the importance score, adding our denoiser with a naively summed query \\(\\mathbf{Q}_t = \\mathcal{C}(\\mathbf{u}_t) + \\pmb{\\tau} + \\pmb{\\Phi}\\) improves the Kendall's \\(\\tau\\) \\((0.071\\rightarrow 0.079)\\). We observe further slight improvement \\((0.079\\to 0.086)\\) when we replace the fixed random score embeddings with learnable ones (Learnable Embedding). Introducing AdaLN transformation [58] on the time \\((\\tau)\\) and positional \\((\\Phi)\\) encodings further improves \\(\\tau\\) to 0.145 and 0.171, respectively. As we adopt the quantization idea, grouping nearby scores into a unified embedding, in Sec. 3.2 with this change, we experiment without it (using a sinusoidal function [80] instead), and observe significant performance drop \\((0.171\\rightarrow 0.125)\\). This confirms that quantization is superior to treating the score" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15102" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.176, + 0.089, + 0.396, + 0.166 + ], + "angle": 0, + "content": "
Modelτ↑ρ↑
CSTA [67]0.1280.185
SummDiff (1 step)0.1700.234
SummDiff (10 steps)0.1750.238
SummDiff (100 steps)0.1820.245
" + }, + { + "type": "table_caption", + "bbox": [ + 0.11, + 0.169, + 0.465, + 0.184 + ], + "angle": 0, + "content": "Table 6. Performance with various numbers of DDIM steps" + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.188, + 0.445, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.331, + 0.483, + 0.373 + ], + "angle": 0, + "content": "Figure 6. Measured \\(\\mathbb{E}|\\{\\pmb{y}^*\\}|\\) and \\(\\mathbb{E}[\\Delta \\pmb{y}^*]\\) with respect to the quantization strength \\(K\\). A larger \\(K\\) is associated with a smaller number of allowed solutions for the given KP." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.379, + 0.483, + 0.44 + ], + "angle": 0, + "content": "as a continuous scalar. On top of this, we combine our model with two widely-known ideas, classifier-free guidance (CFG) [24] and self-Attention guidance (SAG) [30], for further improvement. See Appendix B for details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.443, + 0.484, + 0.504 + ], + "angle": 0, + "content": "Number of DDIM Steps. Tab. 6 suggests that SummDiff outperforms CSTA [67], the best-performing baseline, even with a single DDIM [68] step. Beyond this, SummDiff further improves its performance with more DDIM steps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.483, + 0.581 + ], + "angle": 0, + "content": "Effect of Quantization Strength. Solving a KP usually has a unique solution if the profits \\( v_{i} \\) are continuous. As quantization unifies importance score values within the same interval to a single embedding, it may lead to multiple solutions to \\( \\mathcal{KP}(\\tilde{\\boldsymbol{v}},\\boldsymbol {w},\\rho) \\), where \\( \\tilde{\\boldsymbol{v}} \\) is a quantized version of \\( \\boldsymbol{v} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.483, + 0.734 + ], + "angle": 0, + "content": "To analyze this effect, we measure the average number of solutions \\(\\mathbb{E}_{\\pmb{v}_1\\sim U(0,1)\\otimes N}[\\{\\pmb{y}^*\\in \\{0,1\\} ^N |\\pmb{y}^* = \\mathcal{KP}(\\tilde{\\pmb{v}},\\pmb {w},\\rho)\\} ]\\) while varying \\(K\\), taking expectation over \\(\\pmb{v}\\), which is a value vector sampled from a joint uniform distribution. We also consider the average \\(L_{1}\\) difference between two summaries, \\(\\mathbb{E}_{\\pmb{v}_1\\sim U(0,1)\\otimes N},\\pmb{v}_2\\sim \\mathcal{N}(\\pmb{v}_1,\\frac{1}{K}\\mathbf{I}_N)[\\| \\mathcal{KP}(\\tilde{\\pmb{v}}_1,\\pmb {w},\\rho) - \\mathcal{KP}(\\tilde{\\pmb{v}}_2,\\pmb {w},\\rho)\\| _1]\\), obtained by solving the KP with two similar but distinct value vectors, i.e., \\(\\pmb{v}_2 = \\pmb{v}_1 + \\Delta \\pmb{v}\\), where \\(\\Delta \\pmb{v}\\sim \\mathcal{N}(0,\\frac{1}{K}\\mathbf{I}_N)\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.736, + 0.483, + 0.825 + ], + "angle": 0, + "content": "We numerically analyze the relation between these two values, which indicate the scale of multiplicity of the solutions to the KP, and the quantization strength \\(K\\). As shown in Fig. 6, both the number of solutions \\(|\\{\\pmb{y}^*\\}|\\) and the summary deviation \\(\\Delta \\pmb{y}^*\\) decrease with larger \\(K\\), suggesting more accurate summaries." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.903 + ], + "angle": 0, + "content": "To sum up, increasing the quantization strength \\( K \\) reduces the number of multiple optimal solutions, leading to a more accurate summary. However, an excessively higher \\( K \\) leads to insufficient samples per bin, degrading the overall performance. See Appendix A for more ablation studies." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.657, + 0.107 + ], + "angle": 0, + "content": "5. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.907, + 0.356 + ], + "angle": 0, + "content": "Video Summarization. Early models primarily rely on heuristic unsupervised learning [13, 15, 29, 38, 47, 49-51, 53, 71, 83] to select important or diverse frames, struggling with generalization. Benefiting from annotated datasets, supervised methods [8, 38, 57, 59, 60] have emerged. DSNet [93] and iPTNet [34] improve keyframe selection using frame-level annotations, and SL-Module [82] captures high-level features. With the advent of deep learning, RNNs [84, 88-90] and attention mechanism [4, 6, 33, 35, 39, 42, 77, 85, 94] have been adopted to capture temporal dependencies in videos. Particularly, VASNet [17] and PGL-SUM [3] capture both local and global frame dependencies using self-attention. DMASum [81] introduces mixture of attention layer mitigate the key Softmax Bottleneck. SUM-GAN [50] and AC-SUM-GAN [2] leverage generative adversarial networks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.358, + 0.909, + 0.51 + ], + "angle": 0, + "content": "Multimodal approaches [32, 91] integrate audio or textual data to video summarization. Multimodal transformers are adopted to link frames with corresponding captions, improving context-aware summaries, by CLIP-It [52], MSVA [18], SSPVS [40], and A2Summ [22]. CSTA [67] addresses computational complexity by using CNN-based spatiotemporal attention for efficient frame selection. Unlike these deterministic summarizers, our approach generates multiple plausible summaries by capturing the distribution of video summaries with various perspectives." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.513, + 0.909, + 0.651 + ], + "angle": 0, + "content": "Diffusion for Video Tasks. Diffusion [66] have emerged as a groundbreaking tool in high-quality image generation [21, 25, 54, 63, 64, 69, 70] and super-resolution [14, 36, 61, 65]. Recently, they are extended to video generation [7, 11, 16, 19, 26, 27, 44], joint video and audio generation [41, 62, 73], video editing [9, 10, 79], video inpainting [87] and prediction [31, 86]. Furthermore, they are applied to video understanding tasks like moment retrieval [43, 48], video object segmentation [92, 95], action segmentation [45]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.664, + 0.625, + 0.681 + ], + "angle": 0, + "content": "6. Summary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Video summarization is inherently subjective since people have different perspectives of a good summary. We suggest a generative viewpoint of this task where the model learns the distribution of good summaries, in contrast to the traditional approach of aggregated importance score regression. Our proposed SummDiff model, adopting diffusion models for video summarization for the first time, is able to generate multiple good summaries conditioned on the input video. Our model not only outperforms other baselines but also demonstrates the ability to generate accurate summaries customized to the individual annotators. We further propose additional metrics that measure the quality of the predicted importance scores through an insight from the actual summary generation using knapsack." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15103" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.484, + 0.192 + ], + "angle": 0, + "content": "This work was supported by Samsung Electronics (IO240512-09881-01), Youlchon Foundation, NRF grants (RS-2021-NR05515, RS-2024-00336576, RS-2023-0022663) and IITP grants (RS-2022-II220264, RS-2024-00353131) by the government of Korea." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.206, + 0.188, + 0.222 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.232, + 0.484, + 0.288 + ], + "angle": 0, + "content": "[1] Sami Abu-El-Haija, Nisarg Kothari, Joonseok Lee, Paul Natsev, George Toderici, Balakrishnan Varadarajan, and Sudheendra Vijayanarasimhan. Youtube-8M: A large-scale video classification benchmark. arXiv:1609.08675, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.289, + 0.484, + 0.371 + ], + "angle": 0, + "content": "[2] Evlampios Apostolidis, Eleni Adamantidou, Alexandros I Metsai, Vasileios Mezaris, and Ioannis Patras. AC-SUMGAN: Connecting actor-critic and generative adversarial networks for unsupervised video summarization. IEEE Transactions on Circuits and Systems for Video Technology, 31(8): 3278-3292, 2020. 1, 4, 8, iii" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.373, + 0.484, + 0.442 + ], + "angle": 0, + "content": "[3] Evlampios Apostolidis, Georgios Balaouras, Vasileios Mezaris, and Ioannis Patras. Combining global and local attention with positional encoding for video summarization. In IEEE international symposium on multimedia (ISM), 2021. 1, 2, 4, 7, 8, i, iii" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.444, + 0.483, + 0.5 + ], + "angle": 0, + "content": "[4] Evlampios Apostolidis, Georgios Balaouras, Vasileios Mezaris, and Ioannis Patras. Summarizing videos using concentrated attention and considering the uniqueness and diversity of the video frames. In ICMR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.501, + 0.484, + 0.557 + ], + "angle": 0, + "content": "[5] Tarik Belgacem and Mhand Hifi. Sensitivity analysis of the knapsack problem: Tighter lower and upper bound limits. Journal of systems science and systems engineering, 17:156-170, 2008. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.558, + 0.484, + 0.613 + ], + "angle": 0, + "content": "[6] Manjot Bilkhu, Siyang Wang, and Tushar Dobhal. Attention is all you need for videos: Self-attention based video summarization using universal transformers. arXiv:1906.02792, 2019.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.615, + 0.484, + 0.671 + ], + "angle": 0, + "content": "[7] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.673, + 0.484, + 0.715 + ], + "angle": 0, + "content": "[8] Sijia Cai, Wangmeng Zuo, Larry S Davis, and Lei Zhang. Weakly-supervised video summarization using variational encoder-decoder and web prior. In ECCV, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.716, + 0.484, + 0.757 + ], + "angle": 0, + "content": "[9] Duygu Ceylan, Chun-Hao P Huang, and Niloy J Mitra. Pix2video: Video editing using image diffusion. In ICCV, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.758, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[10] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. StableVideo: Text-driven consistency-aware diffusion video editing. In ICCV, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[11] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. VideoCrafter2: Overcoming data limitations for high-quality video diffusion models. In CVPR, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.858, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Ting Chen, Ruixiang Zhang, and Geoffrey Hinton. Analog bits: Generating discrete data using diffusion models with self-conditioning. In ICLR, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.232, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.134 + ], + "angle": 0, + "content": "[13] Wen-Sheng Chu, Yale Song, and Alejandro Jaimes. Video co-summarization: Video summarization by visual co-occurrence. In CVPR, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.906, + 0.19 + ], + "angle": 0, + "content": "[14] Hyungjin Chung, Byeongsu Sim, and Jong Chul Ye. Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In CVPR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.906, + 0.261 + ], + "angle": 0, + "content": "[15] Sandra Eliza Fontes De Avila, Ana Paula Brandao Lopes, Antonio da Luz Jr, and Arnaldo de Albuquerque Araujo. Vsumm: A mechanism designed to produce static video summaries and a novel evaluation method. Pattern recognition letters, 32(1):56-68, 2011. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.906, + 0.317 + ], + "angle": 0, + "content": "[16] Patrick Esser, Johnathan Chiu, Parmida Atighechian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In ICCV, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.906, + 0.373 + ], + "angle": 0, + "content": "[17] Jiri Fajtl, Hajar Sadeghi Sokeh, Vasileios Argyriou, Dorothy Monekosso, and Paolo Remagnino. Summarizing videos with attention. In ACCV Workshop, 2019. 1, 2, 4, 7, 8, i, iii" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.906, + 0.417 + ], + "angle": 0, + "content": "[18] Junaid Ahmed Ghauri, Sherzod Hakimov, and Ralph Ewerth. Supervised video summarization via multiple feature sets with parallel attention. In ICME, 2021. 8, ii" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.906, + 0.473 + ], + "angle": 0, + "content": "[19] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Fei-Fei Li, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. In ECCV, 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.475, + 0.906, + 0.516 + ], + "angle": 0, + "content": "[20] Michael Gygli, Helmut Grabner, Hayko Riemenschneider, and Luc Van Gool. Creating summaries from user videos. In ECCV, 2014. 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.906, + 0.559 + ], + "angle": 0, + "content": "[21] Jaehoon Hahm, Junho Lee, Sunghyun Kim, and Joonseok Lee. Isometric representation learning for disentangled latent space of diffusion models. In ICML, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.561, + 0.906, + 0.615 + ], + "angle": 0, + "content": "[22] Bo He, Jun Wang, Jielin Qiu, Trung Bui, Abhinav Shrivastava, and Zhaowen Wang. Align and attend: Multimodal summarization with dual contrastive losses. In CVPR, 2023. 1, 4, 7, 8, i, iii" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.617, + 0.906, + 0.673 + ], + "angle": 0, + "content": "[23] Mhand Hifi, Hedi Mhalla, and Slim Sadfi. Sensitivity of the optimum to perturbations of the profit or weight of an item in the binary knapsack problem. Journal of Combinatorial Optimization, 10:239-260, 2005. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.674, + 0.906, + 0.716 + ], + "angle": 0, + "content": "[24] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshop on Deep Generative Models and Downstream Applications, 2021. 1, 7, 8, i" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.717, + 0.906, + 0.744 + ], + "angle": 0, + "content": "[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.745, + 0.906, + 0.814 + ], + "angle": 0, + "content": "[26] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv:2210.02303, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.906, + 0.857 + ], + "angle": 0, + "content": "[27] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In NeurIPS, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.858, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[28] Fa-Ting Hong, Xuanteng Huang, Wei-Hong Li, and Wei-Shi Zheng. Mini-net: Multiple instance ranking network for video highlight detection. In ECCV, 2020. 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15104" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[29] Richang Hong, Jinhui Tang, Hung-Khoon Tan, Shuicheng Yan, Chongwah Ngo, and Tat-Seng Chua. Event driven summarization for web videos. In SIGMM workshop on Social media, 2009. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.483, + 0.192 + ], + "angle": 0, + "content": "[30] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seunglyong Kim. Improving sample quality of diffusion models using self-attention guidance. In ICCV, 2023. 7, 8, i" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.194, + 0.483, + 0.235 + ], + "angle": 0, + "content": "[31] Tobias Hörppe, Arash Mehrjou, Stefan Bauer, Didrik Nielsen, and Andrea Dittadi. Diffusion models for video prediction and infilling. arXiv:2206.07696, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.238, + 0.483, + 0.292 + ], + "angle": 0, + "content": "[32] Tzu-Chun Hsu, Yi-Sheng Liao, and Chun-Rong Huang. Video summarization with spatiotemporal vision transformer. IEEE Transactions on Image Processing, 2023. 8, ii" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.295, + 0.483, + 0.336 + ], + "angle": 0, + "content": "[33] Jia-Hong Huang, Luka Murn, Marta Mrak, and Marcel Worring. GPT2MVS: Generative pre-trained transformer-2 for multi-modal video summarization. In ICMR, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.339, + 0.483, + 0.379 + ], + "angle": 0, + "content": "[34] Hao Jiang and Yadong Mu. Joint video summarization and moment localization by cross-task sample transfer. In CVPR, 2022. 1, 4, 8, iii" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.382, + 0.483, + 0.424 + ], + "angle": 0, + "content": "[35] Yunjae Jung, Donghyeon Cho, Sanghyun Woo, and In So Kweon. Global-and-local relative position embedding for unsupervised video summarization. In ECCV, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.426, + 0.483, + 0.466 + ], + "angle": 0, + "content": "[36] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. In NeurIPS, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.483, + 0.497 + ], + "angle": 0, + "content": "[37] Maurice G Kendall. The treatment of ties in ranking problems. Biometrika, 33(3):239-251, 1945. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.499, + 0.483, + 0.54 + ], + "angle": 0, + "content": "[38] Aditya Khosla, Raffay Hamid, Chih-Jen Lin, and Neel Sundaresan. Large-scale video summarization using web-image priors. In CVPR, 2013. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.543, + 0.483, + 0.598 + ], + "angle": 0, + "content": "[39] Haopeng Li, Qiuhong Ke, Mingming Gong, and Rui Zhang. Video joint modelling based on hierarchical transformer for co-summulation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3904–3917, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.601, + 0.483, + 0.641 + ], + "angle": 0, + "content": "[40] Haopeng Li, Qiuhong Ke, Mingming Gong, and Tom Drummond. Progressive video summarization via multimodal self-supervised learning. In WACV, 2023. 4, 8, ii" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.644, + 0.483, + 0.698 + ], + "angle": 0, + "content": "[41] Judith Li, Xuchan Bao, Zhong Yi Wan, Kun Su, Timo Denk, Dima Kuzmin, Joonseok Lee, and Fei Sha. Diff4steer: Steerable diffusion prior for generative music retrieval with semantic guidance. In ICASSP, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.701, + 0.483, + 0.756 + ], + "angle": 0, + "content": "[42] Ping Li, Qinghao Ye, Luming Zhang, Li Yuan, Xianghua Xu, and Ling Shao. Exploring global diverse attention via pairwise temporal relation for video summarization. Pattern Recognition, 111:107677, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.759, + 0.483, + 0.813 + ], + "angle": 0, + "content": "[43] Pandeng Li, Chen-Wei Xie, Hongtao Xie, Liming Zhao, Lei Zhang, Yun Zheng, Deli Zhao, and Yongdong Zhang. MomentDiff: Generative video moment retrieval from random to real. In NeurIPS, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.816, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[44] Jingyun Liang, Yuchen Fan, Kai Zhang, Radu Timofte, Luc Van Gool, and Rakesh Ranjan. MoVideo: Motion-aware video generation with diffusion model. In ECCV, 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[45] Daochang Liu, Qiyue Li, Anh-Dung Dinh, Tingting Jiang, Mubarak Shah, and Chang Xu. Diffusion action segmentation. In ICCV, 2023. 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.12 + ], + "angle": 0, + "content": "[46] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2017. ii" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.149 + ], + "angle": 0, + "content": "[47] Zheng Lu and Kristen Grauman. Story-driven summarization for egocentric video. In CVPR, 2013. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[48] Dezhao Luo, Shaogang Gong, Jiabo Huang, Hailin Jin, and Yang Liu. Generative video diffusion for unseen cross-domain video moment retrieval. arXiv:2401.13329, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[49] Yu-Fei Ma, Lie Lu, Hong-Jiang Zhang, and Mingjing Li. A user attention model for video summarization. In ACM MM, 2002. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[50] Behrooz Mahasseni, Michael Lam, and Sinisa Todorovic. Unsupervised video summarization with adversarial LSTM networks. In CVPR, 2017. 1, 4, 8, ii, iii" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.292, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[51] Padmavathi Mundur, Yong Rao, and Yelena Yesha. Keyframe-based video summarization using delaunay clustering. International Journal on Digital Libraries, 6:219-232, 2006. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[52] Medhini Narasimhan, Anna Rohrbach, and Trevor Darrell. Clip-it! language-guided video summarization. In NeurIPS, 2021. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[53] Chong-Wah Ngo, Yu-Fei Ma, and Hong-Jiang Zhang. Automatic video summarization by graph modeling. In ICCV, 2003. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.434, + 0.905, + 0.501 + ], + "angle": 0, + "content": "[54] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.544 + ], + "angle": 0, + "content": "[55] Mayu Otani, Yuta Nakashima, Esa Rahtu, and Janne Heikkila. Rethinking the evaluation of video summaries. In CVPR, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.585 + ], + "angle": 0, + "content": "[56] Rameswar Panda and Amit K Roy-Chowdhury. Collaborative summarization of topic-related videos. In CVPR, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.589, + 0.905, + 0.629 + ], + "angle": 0, + "content": "[57] Rameswar Panda, Abir Das, Ziyan Wu, Jan Ernst, and Amit K Roy-Chowdhury. Weakly supervised summarization of web videos. In ICCV, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[58] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[59] Danila Potapov, Matthijs Douze, Zaid Harchaoui, and Cordelia Schmid. Category-specific video summarization. In ECCV, 2014. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[60] Mrigank Rochan and Yang Wang. Video summarization by learning from unpaired data. In CVPR, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.733, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[61] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.775, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[62] Ludan Ruan, Yiyang Ma, Huan Yang, Huiguo He, Bei Liu, Jianlong Fu, Nicholas Jing Yuan, Qin Jin, and Baining Guo. MM-diffusion: Learning multi-modal diffusion models for joint audio and video generation. In CVPR, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[63] Suho Ryu, Kihyun Kim, Eugene Baek, Dongsoo Shin, and Joonseok Lee. Towards scalable human-aligned benchmark for text-guided image editing. In CVPR, 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[64] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour," + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15105" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[65] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE transactions on pattern analysis and machine intelligence, 45(4):4713-4726, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.483, + 0.248 + ], + "angle": 0, + "content": "[66] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.251, + 0.483, + 0.292 + ], + "angle": 0, + "content": "[67] Jaewon Son, Jaehun Park, and Kwangsu Kim. CSTA: CNN-based spatiotemporal attention for video summarization. In CVPR, 2024. 1, 2, 4, 5, 7, 8, i, ii, iii" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.295, + 0.483, + 0.322 + ], + "angle": 0, + "content": "[68] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In *ICLR*, 2020. 3, 8, ii" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.324, + 0.483, + 0.364 + ], + "angle": 0, + "content": "[69] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In NeurIPS, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.368, + 0.483, + 0.408 + ], + "angle": 0, + "content": "[70] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. In NeurIPS, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.412, + 0.483, + 0.452 + ], + "angle": 0, + "content": "[71] Yale Song, Jordi Vallmitjana, Amanda Stent, and Alejandro Jaimes. TVSum: Summarizing web videos using titles. In CVPR, 2015. 1, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.483, + 0.509 + ], + "angle": 0, + "content": "[72] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.513, + 0.483, + 0.568 + ], + "angle": 0, + "content": "[73] Kun Su, Judith Li, Qingqing Huang, Dima Kuzmin, Joon-seok Lee, Chris Donahue, Fei Sha, Aren Jansen, Yu Wang, Mauro Verzetti, and Timo Denk. V2Meow: Meowing to the visual beat via video-to-music generation. In AAAI, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.571, + 0.483, + 0.611 + ], + "angle": 0, + "content": "[74] Jinhwan Sul, Jihoon Han, and Joonseok Lee. Mr. HiSum: A large-scale dataset for video highlight detection and summarization. In NeurlPS, 2024. 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.614, + 0.483, + 0.668 + ], + "angle": 0, + "content": "[75] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In CVPR, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.671, + 0.483, + 0.712 + ], + "angle": 0, + "content": "[76] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.715, + 0.483, + 0.755 + ], + "angle": 0, + "content": "[77] Hacene Terbouche, Maryan Morel, Mariano Rodriguez, and Alice Othmani. Multi-annotation attention model for video summarization. In CVPR, 2023. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.758, + 0.483, + 0.813 + ], + "angle": 0, + "content": "[78] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. ii" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.816, + 0.483, + 0.856 + ], + "angle": 0, + "content": "[79] Shuyuan Tu, Qi Dai, Zhi-Qi Cheng, Han Hu, Xintong Han, Zuxuan Wu, and Yu-Gang Jiang. MotionEditor: Editing video motion via content-aware diffusion. In CVPR, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[80] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, 2017. 2, 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[81] Junyan Wang, Yang Bai, Yang Long, Bingzhang Hu, Zhenhua Chai, Yu Guan, and Xiaolin Wei. Query twice: Dual mixture attention meta learning for video summarization. In ACM MM, 2020. 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[82] Minghao Xu, Hang Wang, Bingbing Ni, Riheng Zhu, Zhenbang Sun, and Changhu Wang. Cross-category video highlight detection via set-based learning. In ICCV, 2021. 1, 4, 7, 8, i, iii" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[83] Hong Jiang Zhang, Jianhua Wu, Di Zhong, and Stephen W Smoliar. An integrated system for content-based video retrieval and browsing. Pattern recognition, 30(4):643-658, 1997. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[84] Ke Zhang, Wei-Lun Chao, Fei Sha, and Kristen Grauman. Video summarization with long short-term memory. In ECCV, 2016. 4, 8, ii" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[85] Yunzuo Zhang, Yameng Liu, Weili Kang, and Ran Tao. VSS-Net: visual semantic self-mining network for video summarization. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[86] Zhicheng Zhang, Junyao Hu, Wentao Cheng, Danda Paudel, and Jufeng Yang. ExtDM: Distribution extrapolation diffusion model for video prediction. In CVPR, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[87] Zhixing Zhang, Bichen Wu, Xiaoyan Wang, Yaqiao Luo, Luxin Zhang, Yinan Zhao, Peter Vajda, Dimitris Metaxas, and Licheng Yu. AVID: Any-length video inpainting with diffusion model. In CVPR, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[88] Bin Zhao, Xuelong Li, and Xiaoqiang Lu. Hierarchical recurrent neural network for video summarization. In ACM MM, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.544 + ], + "angle": 0, + "content": "[89] Bin Zhao, Xuelong Li, and Xiaogiang Lu. HSA-RNN: Hierarchical structure-adaptive rnn for video summarization. In CVPR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[90] Bin Zhao, Xuelong Li, and Xiaogiang Lu. TTH-RNN: Tensor-train hierarchical recurrent neural network for video summarization. IEEE Transactions on Industrial Electronics, 68(4):3629-3637, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[91] Bin Zhao, Maoguo Gong, and Xuelong Li. Hierarchical multimodal transformer to summarize videos. Neurocomputing, 468:360-369, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[92] Chuanwei Zhou, Chunyan Xu, Jun Li, Zhen Cui, and Jian Yang. Quality-aware pattern diffusion for video object segmentation. Neurocomputing, 528:148-159, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[93] Wencheng Zhu, Jiwen Lu, Jiahao Li, and Jie Zhou. DSNet: A flexible detect-to-summarize network for video summarization. IEEE Transactions on Image Processing, 30:948–962, 2020. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[94] Wencheng Zhu, Yucheng Han, Jiwen Lu, and Jie Zhou. Relational reasoning over spatial-temporal graphs for video summarization. IEEE Transactions on Image Processing, 31: 3017-3031, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.856 + ], + "angle": 0, + "content": "[95] Zixin Zhu, Xuelu Feng, Dongdong Chen, Junsong Yuan, Chunming Qiao, and Gang Hua. Exploring pre-trained text-to-video diffusion models for referring video object segmentation. arXiv:2403.12042, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.898 + ], + "angle": 0, + "content": "[96] Daniel Zwillinger and Stephen Kokoska. CRC standard probability and statistics tables and formulae. CRC Press, 1999. 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15106" + } + ] +] \ No newline at end of file diff --git a/2025/SummDiff_ Generative Modeling of Video Summarization with Diffusion/314d20e1-b991-4676-b961-e90cddf3f9d6_origin.pdf b/2025/SummDiff_ Generative Modeling of Video Summarization with Diffusion/314d20e1-b991-4676-b961-e90cddf3f9d6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..26bf191908de922b6acf3341ca1b97cd211ea5ee --- /dev/null +++ b/2025/SummDiff_ Generative Modeling of Video Summarization with Diffusion/314d20e1-b991-4676-b961-e90cddf3f9d6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8ebe6e7f4f6122b5f4a359ce00cd5913d995fd9740828467cc9d3ff82a7c7c16 +size 5010794 diff --git a/2025/SummDiff_ Generative Modeling of Video Summarization with Diffusion/full.md b/2025/SummDiff_ Generative Modeling of Video Summarization with Diffusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a6749c7b1ff54b346534b9f1f15c790e3e0a4c58 --- /dev/null +++ b/2025/SummDiff_ Generative Modeling of Video Summarization with Diffusion/full.md @@ -0,0 +1,388 @@ +# SummDiff: Generative Modeling of Video Summarization with Diffusion + +Kwanseok Kim $^{1*}$ , Jaehoon Hahm $^{2*}$ , Sumin Kim $^{1}$ , Jinhwan Sul $^{3}$ , Byunghak Kim $^{4}$ , Joonseok Lee $^{1\dagger}$ , Seoul National University, 2UIUC, Georgia Institute of Technology, Hyundai Card + +kjvd1009@snu.ac.kr, jh141@illinois.edu, sumink@snu.ac.kr, jsul7@gatech.edu, byunghak.kim@hcs.com, joonseok@snu.ac.kr + +# Abstract + +Video summarization is a task of shortening a video by choosing a subset of frames while preserving its essential moments. Despite the innate subjectivity of the task, previous works have deterministically regressed to an averaged frame score over multiple raters, ignoring the inherent subjectivity of what constitutes a "good" summary. We propose a novel problem formulation by framing video summarization as a conditional generation task, allowing a model to learn the distribution of good summaries and to generate multiple plausible summaries that better reflect varying human perspectives. Adopting diffusion models for the first time in video summarization, our proposed method, SummDiff, dynamically adapts to visual contexts and generates multiple candidate summaries conditioned on the input video. Extensive experiments demonstrate that SummDiff not only achieves the state-of-the-art performance on various benchmarks but also produces summaries that closely align with individual annotator preferences. Moreover, we provide a deeper insight with novel metrics from an analysis of the knapsack, which is an important last step of generating summaries but has been overlooked in evaluation. + +# 1. Introduction + +Recently, short-form videos draw significant attention on video sharing platforms, with a trend that consumers increasingly prefer to quickly grasp the content. They often compressively convey contents that are originally in a longer form, summarizing the core contents into a shorter one; e.g., sport games highlights or movie summarization. This task of selecting core parts of a long video to construct a shorter one is called video summarization. This task is inherently subjective, since there can be multiple criteria for a 'good summary'; e.g., comprehensively covering the entire storyline or subjectively selecting impressive parts of + +the video (highlight detection). Due to this inherent subjectivity, most video summarization or highlight detection datasets [20, 71] offer annotations by multiple raters to reflect various perspectives. + +Since each annotator may have different opinion on the importance of a frame, most existing methods [3, 17, 22, 34, 67, 82] take the frame-level importance scores averaged across multiple annotators as their target label, and are trained to predict them. This frame-level score aggregation looks reasonable in some sense, but in fact it loses the various perspectives to summarize each video. For instance, suppose half of the annotators select clips from the first quarter of the video while the other half select clips from the last quarter. If one simply averages their frame scores, both the first and last quarters end up with similar importance, obscuring the two distinct valid summaries. That is, this simple regression to the averaged frame-level importance scores fail to preserve multiple viewpoints to summarize the video. + +In order to preserve and reflect various views to summarize a video, we pose a distribution of its good summaries and let the model to learn it, instead of giving an already-aggregated single ground truth importance score. Then, the video summarization task can be seen from a generative perspective; i.e., a process of learning and inferring the distribution of good summaries conditioned on the input video. Specifically, the model is now in charge of estimating the distribution of plausible summaries for the given video. Once trained, it allows us to sample multiple summaries from the estimated conditional data manifold. To the best of our knowledge, this generative approach to the video summarization has not been extensively studied, except for a few works [2, 50] that applied adversarial losses to construct a summary that looks like the original video. + +Formulating the video summarization problem as a conditional generation task, we propose to adopt the generative diffusion [25, 72] mechanism, which has been successfully applied to various conditional generation tasks [24, 54, 61]. Specifically, conditioned on the input video, our proposed SummDiff model learns to denoise a random importance + +![](images/b84fa467829d5ca6a0793d03fe6eec5da4442d119ef6b672ccfffe9d9e9054db.jpg) +Figure 1. Overview of SummDiff. Given an input video, SummDiff generates importance scores conditioned on video frames. $T$ denotes the number of DDIM steps. + +![](images/9f842a7c643f7740f4f686512caadea0c71a5be9b0b95427cdab875affe6f21b.jpg) +Figure 2. Video Importance Score Denoiser. We use AdaLN layer to inject time and positional conditions, following [58]. + +score vector over the given video into an importance score vector sampled from the distribution that corresponds to a good summary of the video. In contrast to the previous deterministic methods, our approach allows to sample multiple plausible summaries for a given video starting from a different random noise, better aligned with the subjective nature of summarization where we usually have various true labels reflecting multiple views. + +Extensive experiments demonstrate that our model outperforms existing baselines across multiple datasets. Also, we revisit the knapsack, a relatively unexplored step in the summarization evaluation in spite of its nontrivial impact on the performance, and propose additional novel metrics based on this analysis to provide deeper insights. + +Our contributions can be summarized as follows: + +- We propose a novel generative viewpoint of video summarization, better suited for the subjective nature of the task, allowing multiple plausible summaries for a video. +- We innovatively apply diffusion to the video summarization for the first time, integrating learning the distribution of good summaries into the model. +- We analyze the knapsack optimization process and propose additional metrics to quantify the optimality of the predicted importance scores. + +# 2. Problem Formulation + +Given a video of $N$ frames, the objective of video summarization is to identify and select $S < N$ frames that effectively encapsulate the essence of the video content. Let $\mathbf{X} \equiv \{\mathbf{X}_i \in \mathbb{R}^{H \times W \times 3} : i = 0, \dots, N - 1\} \in \mathbb{R}^{N \times H \times W \times 3}$ be a video, where $H, W$ denote the size of the frames. + +A video summary, $\mathbf{y} \equiv \{y_i \in \{0,1\} : i = 0,\dots,N-1, \sum_i y_i \leq S\} \in \{0,1\}^N$ , indicates inclusion (1) or exclusion (0) of each frame. When a video is provided with multiple ground truth annotations, we denote each individual score by $\mathbf{s}^{(r)} \in [0,1]^N$ and the corresponding binary summary by $\mathbf{y}^{(r)} \in \{0,1\}^N$ which is obtained following the procedure explained in Sec. 3.3. The predicted summary is denoted by $\hat{\mathbf{y}} \equiv \{\hat{y}_i \in \mathbb{R} : i = 0,\dots,N-1\} \in \mathbb{R}^N$ . + +Previous models [3, 17, 67] have approached video summarization as a regression task, aiming to predict the importance score $\mathbf{s} \in [0,1]^N$ for each video, often the average of multiple importance scores, $\frac{1}{|R|} \sum_{r \in \mathcal{R}} \mathbf{s}^{(r)}$ . In contrast, our approach allows multiple summaries for each video, aiming to learn the distribution of its plausible summaries. The previous scheme can be seen as a special case of ours, where every video has a single golden way of summarization, pointed at $\mathbf{s}$ with zero variance. Under our extended setting, multiple importance scores $\{\mathbf{s}^{(r)} | r \in \mathcal{R}\}$ can be given to the same conditioning video $\mathbf{X}$ , forming a probability distribution of plausible importance scores. + +# 3. Diffusion-based Video Summarization + +Posing the video summarization as a conditional generation task, we introduce our SummDiff model, designed to adapt the distribution of individual importance scores for a given video by learning to denoise. + +# 3.1. Overall Flow of the Proposed Method + +Fig. 1 illustrates the overall flow of our SummDiff model. + +Encoding Process. We first encode each frame $\mathbf{X}_i$ for $i = 1, \dots, N$ from the input video $\mathbf{X}$ using a pre-trained image encoder. The extracted features are further contextualized through self-attention [80], as seen in Fig. 1. We denote the encoded feature for each individual frame by $\mathbf{z}_i \in \mathbb{R}^D$ , and collectively the entire feature matrix by $\mathbf{Z} \in \mathbb{R}^{N \times D}$ . + +Denoising Process. We then learn to denoise an individual importance score vector from a random noise, conditioned on the visual embeddings. First, the forward process adds noise to the ground truth individual importance score $\mathbf{s}_0 \equiv \mathbf{s}^{(r)}$ and sample a noised one $\mathbf{s}_t = \sqrt{\bar{\alpha}_t}\mathbf{s}_0 + \sqrt{1 - \bar{\alpha}_t}\boldsymbol{\epsilon}_t$ , where $\boldsymbol{\epsilon}_t \sim \mathcal{N}(0,\mathbf{I}), \bar{\alpha}_t = \prod_{\tau=1}^t (1 - \beta_\tau)$ , and $t$ is the diffusion time step. The perturbation kernel at $t$ , defined as $q(\mathbf{s}_t|\mathbf{s}_{t-1}) = \mathcal{N}(\mathbf{s}_t; \sqrt{1 - \beta_t}\mathbf{s}_{t-1}, \beta_t\mathbf{I})$ , where $\beta_t$ is defined by the variance schedule. Since $\mathbf{s}_0$ is bounded within $[0,1]$ , it is not straightforward to add Gaussian noise directly to it so we first transform it to its logit, $\mathbf{u}_0 = \log \frac{\mathbf{s}_0}{1 - \mathbf{s}_0} \in$ + +$\mathbb{R}^N$ , and perform the noising process in this logit space. For numerical stability, we clip $\mathbf{s}_0$ to $[\epsilon, 1 - \epsilon]$ , where $\epsilon$ is a small constant. + +Then, the reverse process progressively removes noise from $\mathbf{s}_t$ to $\mathbf{s}_0$ , formulated by $p_{\theta}(\mathbf{s}_{t - 1}|\mathbf{s}_t) = \mathcal{N}(\mathbf{s}_{t - 1};\boldsymbol {\mu}_\theta (\mathbf{s}_t,t),\sigma_t^2\mathbf{I})$ , where $\sigma_t^2$ is determined in relation to $\beta_{t}$ , and the posterior mean $\pmb{\mu}_{\theta}(\mathbf{s}_t,t)$ is modeled with a trainable neural network. In this paper, we extend this estimator of posterior mean to be conditioned on a video, similarly to the conditional generation on diffusion models. + +Once trained, our denoiser is able to recover a plausible importance score for a video from a random noise, which is used to generate a summary (Sec. 3.3). Repeated generation across the noise distribution would converge to the true distribution of plausible scores for the video. + +# 3.2. Video Importance Score Denoiser + +We formulate the key component of our method, the video importance score denoiser in Fig. 2. Starting from a noised logit score vector $\mathbf{u}_t \in \mathbb{R}^N$ , it predicts a plausible importance score, conditioned on the given video. Denoted by $f_{\theta}(\mathbf{u}_t, t, \mathbf{Z})$ , it learns to denoise the given score vector $\mathbf{u}_t$ at the diffusion time step $t$ under the visual condition $\mathbf{Z}$ , producing the denoised score $\hat{\mathbf{u}}_{t-1}$ by one time step, where $\theta$ is the set of its learnable parameters. + +Transformer-based Diffusion. We denoise the logit-transformed score $\mathbf{u}_t$ given $t$ and $\mathbf{Z}$ , predicting $\hat{\mathbf{u}}_{t-1}$ , using a transformer-based cross-attention [80]. The logit-transformed score $\mathbf{u}_t$ acts as the query, and the visual condition $\mathbf{Z}$ is used as key/value. This setup allows the model to effectively denoise $\mathbf{u}_t$ conditioned on the information from the input video, ensuring consistency between the predicted logit score $\hat{\mathbf{u}}_{t-1}$ and the condition $\mathbf{Z}$ . + +To apply dot-product cross-attention, the dimensionality should match for queries and keys. To this end, we quantize the importance scores into a predefined number $(K)$ of uniform segments. Specifically, we convert the noised logit $\mathbf{u}_t$ back to its original bounded range $[0,1]$ by applying sigmoid $1 / (1 + e^{-\mathbf{u}_t})$ , and divide them into $K$ equally-binned segments. Each score range is associated with a learnable embedding of size $D$ (codebook in Fig. 1). Based on the codebook $\mathcal{C}$ , we map the scores $\mathbf{u}_t \in \mathbb{R}^N$ to their corresponding quantized embedding $\mathcal{C}(\mathbf{u}_t) \in \mathbb{R}^{N \times D}$ . Our denoiser $f_{\theta}(\mathcal{C}(\mathbf{u}_t), t, \mathbf{Z})$ takes this $\mathcal{C}(\mathbf{u}_t)$ as input, instead of $\mathbf{u}_t$ . See Sec. 4.5 for analysis on the size of codebook. + +Query Formation in Cross-Attention. Following [25], the time step $t$ is embedded as $\pmb{\tau} \in \mathbb{R}^{D}$ using sinusoidal functions and an MLP. Considering the sequential nature of videos, we also introduce temporal positional embeddings $\Phi \in \mathbb{R}^{N \times D}$ using sinusoidal functions. + +The simplest way to set the query in the cross-attention would be to add all inputs except for the condition $\mathbf{Z}$ ; that is, $\mathbf{Q}_t = \mathcal{C}(\mathbf{u}_t) + \boldsymbol{\tau} + \boldsymbol{\Phi}$ . However, we leverage AdaLN-Zero + +block [58] to integrate the time embedding $\pmb{\tau}$ and temporal positional encoding $\Phi$ more effectively by separating them from the query $\mathbf{Q}_t$ , preventing information mixing. Hence, the query becomes $\mathbf{Q}_t = \mathcal{C}(\mathbf{u}_t)$ and $\pmb{\tau}, \Phi$ are conditioned via scale-shift operation. See Sec. 4.5 for ablation studies. + +We conduct cross-attention with $\mathbf{Q}_t$ as queries and visual conditions $\mathbf{Z} \in \mathbb{R}^{N \times D}$ as keys $\mathbf{K}$ and values $\mathbf{V}$ . We take an MLP from the AdaLN output to regress the scale and shift parameters, denoted by $\mathbf{A}_1$ , $\mathbf{B}_1$ , $\Gamma_1$ , $\mathbf{A}_2$ , $\mathbf{B}_2$ , and $\Gamma_2 \in \mathbb{R}^{N \times D}$ . As depicted in Fig. 2, they scale and shift $\mathbf{Q}_t$ , $\mathbf{K}$ , and $\mathbf{V}$ . Then, they are passed through cross-attention with skip connections and a subsequent rescaling. Formally, + +$$ +\mathbf {H} ^ {i ^ {\prime}} = \boldsymbol {\Gamma} _ {1} \odot \mathbf {H} ^ {i} + \mathbf {H} ^ {i} + \mathbf {B} _ {1} +$$ + +$$ +\mathbf {X} _ {1} = \mathbf {A} _ {1} \odot \operatorname {s o f t m a x} \left(\mathbf {Q} _ {t} ^ {i ^ {\prime}} \mathbf {K} ^ {i ^ {\prime} \top}\right) \mathbf {V} ^ {i ^ {\prime}} + \mathbf {Q} _ {t} ^ {i ^ {\prime}} +$$ + +$$ +\mathbf {X} _ {2} = \boldsymbol {\Gamma} _ {2} \odot \mathbf {X} _ {1} + \mathbf {X} _ {1} + \mathbf {B} _ {2} +$$ + +$$ +\mathbf {Q} _ {t} ^ {i + 1} = \mathbf {A} _ {2} \odot \operatorname {M L P} (\mathbf {X} _ {2}) + \mathbf {X} _ {2}, +$$ + +where $\mathbf{H}^{i^{(t)}}\in \{\mathbf{Q}_t^{i^{(t)}},\mathbf{K}^{i^{(t)}},\mathbf{V}^{i^{(t)}}\}$ denotes the matrices used for attention, $\odot$ is the (broadcasted) Hadamard product, and $i = 1,\ldots ,L$ is the layer index. + +Training. We train the denoiser $f_{\theta}$ to estimate the true importance score $\hat{\mathbf{s}}_0$ after acting upon a fully-connected layer and a sigmoid function $\sigma$ . We minimize the following loss [12] on each annotator's individual importance score: + +$$ +\mathcal {L} \left(\mathbf {s} _ {0}, \hat {\mathbf {s}} _ {0}\right) = \left\| \mathbf {s} _ {0} - \hat {\mathbf {s}} _ {0} \right\| _ {2} ^ {2} = \left\| \mathbf {s} _ {0} - \sigma \left(\operatorname {F C} \left(f _ {\theta} \left(\mathcal {C} \left(\mathbf {u} _ {t}\right), t, \mathbf {Z}\right)\right)\right) \right\| _ {2} ^ {2}. +$$ + +The learnable embeddings in the codebook $\mathcal{C}$ are compositely optimized, finding an effective representation for $\mathbf{u}_t$ during training. + +Inference. Our model generates a logit-transformed importance score from a random noise $\mathbf{u}_T \sim \mathcal{N}(0, \mathbf{I})$ for a given video. Employing the reverse diffusion process [68], it iteratively refines the logit score towards cleaner estimations: + +$$ +\begin{array}{l} \hat {\mathbf {u}} _ {t - 1} = \sqrt {1 - \bar {\alpha} _ {t - 1} - \sigma_ {t} ^ {2}} \frac {\hat {\mathbf {u}} _ {t} - \sqrt {\bar {\alpha} _ {t}} f _ {\theta} (\mathcal {C} (\mathbf {u} _ {t}) , t , \mathbf {Z}))}{\sqrt {1 - \bar {\alpha} _ {t}}} \\ + \sqrt {\bar {\alpha} _ {t - 1}} f _ {\theta} (\mathcal {C} (\mathbf {u} _ {t}), t, \mathbf {Z})) + \sigma_ {t} \epsilon_ {t} \in \mathbb {R} ^ {N}, \\ \end{array} +$$ + +where $\sigma_{t} = \sqrt{(1 - \bar{\alpha}_{t - 1}) / (1 - \bar{\alpha}_{t})}\sqrt{1 - \bar{\alpha}_{t} / \bar{\alpha}_{t - 1}}$ (DDPM reverse process). When $t = 1$ at the last step, $\hat{\mathbf{u}}_0 = f_\theta (\mathcal{C}(\mathbf{u}_1),1,\mathbf{Z})$ is directly used to remove stochasticity at the end of inference process. Finally, we take sigmoid to $\hat{\mathbf{u}}_0$ to derive the final importance score $\hat{\mathbf{s}}_0\in (0,1)^N$ + +# 3.3. Summary Generation + +From the raw importance score $\hat{\mathbf{s}}_0$ , we apply the standard knapsack-based approach to decide which frames to be included in the final summary. To make this summary video more realistic, it is common to choose at a semantic clip + +level instead of individual frame level. We adopt a widely-used Kernel Temporal Segmentation (KTS) [59, 84] to partition a video into disjoint temporal intervals. We take the average score among the corresponding frames as the clip-level importance; i.e., $v_{i} = \sum_{j = t_{i}}^{t_{i + 1}}\hat{\mathbf{s}}_{0,j} / (t_{i + 1} - t_{i})$ for the $i$ -th clip composed of frames from $t_{i}$ to $t_{i + 1}$ . Subsequently, we select the clips based on $v_{i}$ by solving the binary knapsack problem (KP) [23] with dynamic programming [71]: + +$$ +\mathcal {K P} (\boldsymbol {v}, \boldsymbol {w}, \rho) \equiv \underset {\boldsymbol {s} \in \{0, 1 \} ^ {M}} {\operatorname {a r g m a x}} \sum_ {i = 1} ^ {M} v _ {i} s _ {i} \quad \text {s . t .} \sum_ {i = 1} ^ {M} w _ {i} s _ {i} \leq \rho N, +$$ + +where $\pmb{v},\pmb{w},\rho$ denotes the values of each clip $(v_{i}\in [0,1])$ , costs of selecting each clip, and the budget constraint ratio (e.g., $\rho = 0.15$ ; $15\%$ of the video length), respectively. If the argmax has multiple solutions, we will abuse the notation and denote $\mathcal{KP}(\pmb {v},\pmb {w},\rho)$ as an arbitrary element of the solution set. + +# 4. Experiments + +# 4.1. Experimental Settings + +Datasets. We evaluate our method on three benchmarks. TVSum [71] and SumMe [20] are traditional datasets, containing 50 and 25 videos respectively and manually labeled annotations by up to 20 raters. Mr. HiSum [74] is a large-scale dataset, composed of 31,892 videos and importance scores derived from YouTube Most Replayed statistics. It provides an importance score $\mathbf{s}_0$ and a corresponding summary $\mathbf{y}_0$ for each video, averaged over 50,000+ viewers. We use the default features for each dataset, GoogLeNet [75] for TVSum and SumMe, and Inception-v3 [76] PCA-ed to 1024D [1] for Mr. HiSum. + +We adopt two dataset splits. First, we randomly split TVSum and SumMe into 7:1:2 for train, validation, and test set, and report averaged test scores for 5 random splits (namely, TVT). Also, following prior works [67, 77], we include five-fold cross-validation results (5FCV), although these may suffer from overfitting to the test set due to the lack of explicit validation set, and thus may not generalize well on unseen videos. For Mr. HiSum, we use its original train/validation/test split. We run all experiments 5 times. + +Evaluation Metrics. Following recent practice [22, 40, 55, 67], we report rank order statistics, Kendall's $\tau$ [37] and Spearman's $\rho$ [96], to assess how well the model ranks frame importance. F1 score had been widely used in video summarization, but recent literature [67, 77] reports that it is excessively sensitive to video segmentation and unreasonably favors summaries composed of many short shots [55], while disregarding longer key shots [67]. As an alternative to F1, we propose additional metrics that represent the influence of importance score through an analysis of the knapsack algorithm in Sec. 4.4. + +
MethodSumMeTVSum
τρτρ
Random0.0000.0000.0000.000
Human0.2050.2130.1770.204
A2Summ [22]0.0880.0960.1570.206
VASNet [17]0.0890.0990.1530.205
PGL-SUM [3]0.1040.1160.1410.186
CSTA [67]0.1080.1200.1680.221
SummDiff (Ours)0.1330.1480.1730.226
+ +Table 1. Comparison of models trained under TVT (train/val/test split) on SumMe [20] and TVSum [71]. Best and second-best results are boldfaced and underlined, respectively. + +
MethodSumMeTVSum
τρτρ
Random0.0000.0000.0000.000
Human0.2050.2130.1770.204
DSNet-AF [93]0.0370.0460.1130.138
DSNet-AB [93]0.0510.0590.1080.129
SUM-GAN [50]0.0490.0660.0240.031
AC-SUM-GAN [2]0.1020.0880.0310.041
CLIP-It [52]--0.1080.147
iPTNet [34]0.1010.1190.1340.163
A2Summ [22]0.1080.1290.1370.165
VASNet [17]0.1600.1700.1600.170
PGL-SUM [3]--0.1570.206
AAAM [77]--0.1690.223
MAAM [77]--0.1790.236
VSS-Net [85]--0.1900.249
DMASum [81]0.0630.0890.2030.267
SSPVS [40]0.1920.2570.1810.238
CSTA [67]0.2460.2740.1940.255
SummDiff (Ours)0.2560.2850.1950.255
+ +Table 2. Comparison of models trained under 5FCV (5-Fold Cross Validation) on SumMe [20] and TVSum [71]. Training under the 5FCV setting tends to overfit the test set. + +
Modelτ↑ρ↑MAPρ=50%↑MAPρ=15%↑
SUM-GAN [50]0.0670.09556.6223.56
VASNet [17]0.0690.10258.6925.28
AC-SUM-GAN [2]0.0120.01855.3521.88
SL-module [82]0.0600.08858.6324.95
PGL-SUM [3]0.0970.14161.6027.45
iPTNet [34]0.0200.02955.5322.74
A2Summ [22]0.1210.17263.2032.34
CSTA [67]0.1280.18563.3830.42
SummDiff0.1750.23865.4433.83
+ +Table 3. Evaluation on Mr. HiSum [74]. See Appendix D for the full table including standard deviations. + +We further evaluate our method on video highlight detection following [74]. First, we uniformly divide the input video into 5-second-long shots and calculate the average frame scores for each shot. The top $\rho \in \{0.15, 0.5\}$ of these shots are designated as ground truth highlights. We report Mean Average Precision (MAP), following [28, 56, 84]. See Appendix E for more implementation details. + +# 4.2. Results and Analysis + +Considering the innate subjectivity of video summarization, we train our model to learn from each individual score, allowing it to capture multiple ways of summarization for each video. Specifically, we use 20 individual importance scores per video in TVSum [71] and 15-18 individual binary summaries in SumMe [20], since SumMe does not provide individual importance scores. Tab. 1 and 2 show that SummDiff achieves the best performance on SumMe and TVSum, with the sole exception of TVSum in 5FCV. While DMA-SUM [81] performs the best on TVSum in 5FCV, it significantly underperforms on SumMe, indicating that its high performance on TVSum does not generalize well. Comparing TVT and 5FCV, most models exhibit a considerable performance gap. This demonstrates how much existing models have overfitted to the test set under the conventional 5FCV protocol. We adopt TVT as a remedy to this, but due to the extremely small size of these datasets, the reliability of evaluation is still limited [74]. + +On the larger Mr. HiSum, we train our model using the aggregated annotation, as it does not provide individual annotations. This evaluation mainly aims to verify scalability of SummDiff on a large-scale dataset. Tab. 3 presents the performance on Mr. HiSum, where SummDiff consistently outperforms all baselines across all metrics, significantly surpassing the strongest competitor, CSTA [67] with a large margin. These results highlight scalability and effectiveness of our method even under the single-label setting. + +# 4.3. Qualitative Demonstration + +We demonstrate how our method discovers a variety of possible summaries for a video. Specifically, we generate video summaries for the 50 videos in TVSum using CSTA, PGL-SUM, VASNet, and ours. For each generated summary, we measure the Kendall's $\tau$ with all 20 ground truth annotations, respectively. Considering $\tau \geq 0.25$ as the threshold for the summary to be matched with the annotator, Fig. 3 shows if the generated summary for a video (row) matches with each annotator (column). As the baselines (CSTA, PGL-SUM, and VASNet) summarize a video deterministically, each cell is either completely matched (1) or not (0). If the summary matches with majority of annotators (e.g., video 29 for CSTA), it means the deterministic summary matches with the dominant way of summarization for the video. If it matches only with a few annotators (e.g., video 24 for CSTA), it means the generated summary matches with a minor opinion. For our method, we generate 100 summaries per video from different Gaussian noise vectors, and mark each cell with the ratio of summaries that match with each annotator. The darker the cell colors are for each row, the more various ways of summaries have been discovered by our method. Comparing the heat maps, we clearly observe that SummDiff generates significantly more various + +![](images/f6cca2b7b476a0a7178d63f359fccc011c31a9fb1f76beb79ca54083ab5574d8.jpg) + +![](images/ed7787ef43b384971c21f8eb6d490f9d15b74c6150f5154fa81abe62b82c9536.jpg) + +![](images/c77634b845e23c2160ffb7c7cd3817622f8142f6232c9fb7d822ba099c7450fd.jpg) + +![](images/bc2e7b4fbbc547903332dbb2ce3759e013d81c4480865824204160c0bb4986e8.jpg) + +![](images/9df82d52124190e43a3d21ab5eb0b84ba032c661885ddca815c6f2cd07ebfdec.jpg) +Figure 3. Ratio of summaries with $\tau \geq 0.25$ for each video-annotator pair. The heatmap illustrates how closely each method's summary matches with individual annotations. This comparison reveals the extent to which each method captures the varied human summaries. SummDiff covers a larger area of the heatmap, which indicates better coverage over various summaries. + +![](images/a883bc0e6af6c6b10d15d29b157630f6feedc0d4f7a0f108e61f696362092211.jpg) + +Figure 4. Distribution of true and predicted summaries of given videos. SummDiff can generate various summaries and cover the targeted distribution of summaries, while baselines deterministically predicts a single summary. +![](images/18a5f862fe166f908f6ba15ecdaa5b03d9cde469cf7c05c78d9f499218bd91f7.jpg) +SummDiff CSTA PGL_SUM VASNet Annotator + +![](images/3ae161d9bbdee32928283589715ed8b7ad9d1dd34952dabb9e1f4ab851441c37.jpg) + +summaries covering multiple viewpoints, better reflecting the distribution of annotated summaries. + +We further illustrate the variation and quality of generated summaries in Fig. 4. The summaries annotated by human raters ( $\bullet$ Annotator) and those generated by ours ( $\bullet$ + +Figure 5. Demonstration of video summaries generated by competing methods on a TVSum video. Shaded parts indicate the segments selected by each method, and the two rows of edged boxes within each method indicate the two distinct true annotations. The results clearly demonstrate the effectiveness of SummDiff in capturing multiple plausible summaries for a video. See another example in Appendix F. + +SummDiff) and baselines ( • CSTA, • PGL-SUM, and • VASNet) are projected onto the first two principal components, which are computed using only the human-annotated summaries. The distribution of true summaries are displayed using a contour map, obtained by Gaussian kernel density estimation. It demonstrates that our SummDiff generates multiple summaries closely aligned with the distribution of true annotations. For example, video 22 (top-left) roughly has three ways to summarize, according to the human raters. SummDiff generates summaries dominantly covering the two modes in the lower-left region, and one case in the upper-right mode. In contrast, baselines produce a single, less accurate summary, failing to account for the variety inherent in human-annotated summaries. We observe similar patterns in the other three plots as well. Fig. II provides more examples in Appendix C. + +Fig. 5 illustrates example summaries generated by CSTA, A2Summ, PGL-SUM, VASNet, and our SummDiff, for a video selected from TVSum test set. Shaded parts indicate the segments selected by each method (note that SummDiff can generate two different summaries while baselines always produce a single summary), and the two rows of edged boxes within each method indicate two different true annotations. These results demonstrate that SummDiff produces more accurate and various summaries, effectively capturing multiple plausible summarizations for each video. + +# 4.4. Metrics Inspired from Knapsack Optimality + +In spite of nontrivial impact of the knapsack at the end of the summary generation, previous evaluation metrics have focused only on the accuracy of the importance scores. Through a thorough analysis on the knapsack problem (KP), we provide additional metrics that measure the contribution of importance scores. By accounting for knapsack optimality and clip-level weights, our new metrics are capable of assessing the predicted importance score more accurately than existing metrics such as F1 or ranking-based ones. + +Confidence of Importance Score. First, we analyze the conditions under which the same optimal KP solution remains valid despite some perturbations to the profits $\pmb{v}$ . These perturbations can be a modeling to the imperfect estimation of the importance scores in video summarization. + +Let $\pmb{y}^{*}$ be the optimal solution to the original KP (i.e., $\pmb{y}^{*} = \mathcal{KP}(\pmb{v},\pmb{w},\rho)$ ; see definition in Sec. 3.3) and $\Gamma \subseteq [N]\equiv \{n\in \mathbb{N}\mid 1\leq n\leq N\}$ be an arbitrary subset of items associated with the perturbed profits $v_{i}^{\prime} = v_{i} + \Delta v_{i}$ , where $\Delta v_{i} = 0$ if $i\notin \Gamma$ and $N$ is the number of items. In our case, $\pmb{v}\in \mathbb{R}^{N}$ is the ground truth importance scores, $\pmb{y}^{*}\in [0,1]^{N}$ is the ground truth summary, and the perturbed profit $\pmb{v}^{\prime}\in \mathbb{R}^{N}$ corresponds to the imperfectly predicted importance scores. We further define two disjoint subsets of $\Gamma$ . First, $\Gamma_0^+ \equiv \{i\in \Gamma \mid y_i^* = 0,\Delta p_i > 0\}$ refers to the set of items excluded in the original optimal solution but undergoing increased profit, and $\Gamma_1^- \equiv \{i\in \Gamma \mid y_i^* = 1,\Delta p_i < 0\}$ is the opposite, the set of items included in the optimal solution but experiencing a decreased profit. Then, Hifi et al. [23] claims that the solution $\pmb{y}^{*}$ remains optimal for the perturbed KP, i.e., $\pmb{y}^{*} = \mathcal{KP}(\pmb{v},\pmb{w},\rho) = \mathcal{KP}(\pmb{v}^{\prime},\pmb{w},\rho)$ , if + +$$ +\sum_ {i \in \Gamma_ {0} ^ {+}} \Delta v _ {i} - \sum_ {i \in \Gamma_ {1} ^ {-}} \Delta v _ {i} \leq \boldsymbol {v} \cdot \mathcal {K P} (\boldsymbol {v}, \boldsymbol {w}, \rho) - \max (\Psi^ {+}, \Psi^ {-}), +$$ + +where $\Psi^{+} \equiv \max_{i \in \Gamma_{0}^{+}} [\pmb{v} \cdot \mathcal{KP}(\pmb{v} - v_{i}\hat{e}_{i}, \pmb{w}, \rho - w_{i}/N)]$ is the maximum value of the KP when removing any item from $\Gamma_{0}^{+}$ and adjusting the constraint, and similarly, $\Psi^{-} \equiv \max_{i \in \Gamma_{1}^{-}} [\pmb{v} \cdot \mathcal{KP}(\pmb{v} - v_{i}\hat{e}_{i}, \pmb{w}, \rho)]$ is the maximum value of the KP when removing any item from $\Gamma_{1}^{-}$ . + +Based on this theorem we propose a new metric, Confidence of Importance Score (CIS), to estimate how close the predicted scores are to the ground truth: + +$$ +\operatorname {CIS} = \sum_ {i \in \Gamma_ {0} ^ {+}} \Delta v _ {i} - \sum_ {i \in \Gamma_ {1} ^ {-}} \Delta v _ {i} - \boldsymbol {v} \cdot \mathcal {K P} (\boldsymbol {v}, \boldsymbol {w}, \rho) + \max \left(\Psi^ {+}, \Psi^ {-}\right). +$$ + +If CIS $\leq 0$ , the theorem indicates that the predicted importance score $\hat{\pmb{v}}$ is guaranteed to induce the true KP solution $\pmb{y}^{*}$ . If CIS $>0$ , $\hat{\pmb{v}}$ does not guarantee to induce the KP solution, but a lower CIS would indicate a solution closer to $\pmb{y}^{*}$ than a higher one. That is, a smaller CIS indicates a stronger confidence in satisfying the inequality, indicating a higher chance for the solution of the perturbed KP to be identical to that of the original KP. + +Unlike F1 that simply measures how close the predicted summary is to the ground truth, our CIS accounts for the confidence that the predicted score will exactly induce the true summary. This is particularly important in video summarization, since the importance score itself is subjective and often noisy, requiring robust selection of frames. + +Weighted Inclusion Ratios. Analyzing the sensitivity of the KP optimum to the perturbation of the importance score, Belgacem and Hifi [5] established bounds of the perturbed profits to retain the original optimum. These intervals are + +$$ +\boldsymbol {y} _ {i} ^ {*} = 1 \rightarrow \Delta v _ {i} \in I _ {i} ^ {1} = \left[ \right. \max \left(\Delta_ {i} ^ {-}, w _ {i} \max _ {k \in \Gamma} \mu_ {k} - p _ {i}\right), + \infty\left. \right) +$$ + +$$ +\boldsymbol {y} _ {i} ^ {*} = 0 \rightarrow \Delta v _ {i} \in I _ {i} ^ {0} = \left( \right.- \infty , \min \left(\Delta_ {i} ^ {+}, p _ {i} - w _ {i} \min _ {k \in \Gamma} \mu_ {k}\right)\left. \right], +$$ + +where $\Delta_i^+ = \pmb{v} \cdot (\mathcal{KP}(\pmb{v}, \pmb{w}, \rho) - \mathcal{KP}(\pmb{v} - v_i \hat{e}_i, \pmb{w}, \rho - w_i / N), \Delta_i^- = \pmb{v} \cdot (\mathcal{KP}(\pmb{v} - v_i \hat{e}_i, \pmb{w}, \rho) - \mathcal{KP}(\pmb{v}, \pmb{w}, \rho))$ , and $\mu_k$ denotes the critical ratio of $k \in \Gamma$ : $\mu_k = \mathcal{CR}([N] \setminus \{k\}, \rho - 1_{\{y_k^* = 0\}} \cdot w_k / N)$ , with $\mathcal{CR}(A, \gamma) = v_s / w_s$ and $s = \min\{l | \sum_{i=1}^{l} v_s / w_s > \rho\}$ is the index of the critical item. The items in $A$ are considered in descending order of $v_s / w_s$ . To sum up, if the perturbed profit resides in the interval $v_i' \in I_i^{y_i^*}$ , it is guaranteed that the solution of perturbed KP is identical with that of the original KP. + +We propose Inclusion Ratio (IR) to estimate how much the predicted scores fall into the safe bounds; for the $i$ -th importance score, $\mathrm{IR}_i \equiv |\{\Delta v_i \in I_i^{y_i^*} | i \in [N]\}| / N$ . This ratio reflects the proportion of the predicted scores that retain the KP solution. We propose the Weighted average of the Inclusion Ratios (WIR) as a metric to measure the proximity of the predicted score to the ground truth: $\mathrm{WIR} \equiv \sum_{i=1}^{n} \left( \frac{w_i}{\sum_{j=1}^{n} w_j} \mathrm{IR}_i \right)$ , where $w_i$ denotes the segment length of each shot. + +Weighted Sum of Errors (WSE). Additionally, we compare a weighted sum of the errors $\sum_{i} w_{i} \Delta p_{i}$ , where $w_{i}$ is the segment length of each clip. + +Comparison. Tab. 4 compares several competitive methods using these new metrics. Our SummDiff achieves the best performance in all metrics, indicating that the summaries generated by our method are closer to the optimal knapsack result, under the theoretical analysis in literature. + +Discussion. Our newly proposed metrics, WIR and CIS, evaluate the quality of the generated importance scores + +
Methodτ↑ρ↑CIS ↓WIR ↑WSE ↓
Uniform-Random0.0000.0009.270.4546.82
SL-module [82]0.0600.0886.830.5630.09
CSTA [67]0.1280.1856.230.5725.76
PGL-SUM [3]0.0970.1416.140.5826.22
VASNet [17]0.0690.1026.250.5926.79
A2Summ [22]0.1210.1726.650.5530.55
SummDiff0.1750.2385.960.6125.24
+ +Table 4. Evaluation using our new metrics proposed in Sec. 4.4. + +
Moduleτ↑ρ↑
Encoder Only0.0710.104
(+) Video Importance Denoiser0.0790.116
(+) Learnable Embedding0.0860.124
(+) AdaLN (τ)0.1450.204
(+) AdaLN (τ, Φ) 1 layer0.1710.232
(-) Quantization0.1250.174
(+) Classifier-free Guidance (CFG) [24]0.1750.238
(+) Self-attention Guidance (SAG) [30]0.1770.239
+ +Table 5. Effect of individual components + +in two aspects. WIR measures how many importance scores are individually trustworthy, weighted by the duration. Specifically, WIR first computes the safe interval $I_{i}$ for each importance score such that the true summary remains unchanged. Then, given the predicted importance scores $\hat{\mathbf{v}}$ , it measures how many of them fall within their corresponding safe intervals $\hat{v}_{i} \in I_{i}$ . + +On the other hand, CIS quantifies the risk that the predicted importance scores will lead to a different summary than the one generated using the true importance scores. In other words, CIS evaluates the overall chance how likely the predicted score vector $\hat{\mathbf{v}}$ , as a whole, is to give a summary different from the true one. With the analysis of knapsack optimality, these metrics would better measure the accuracy of predicted importance score, considering the fidelity of final summary after knapsack to the ground truth, than F1 or ranking-based metrics. + +# 4.5. Ablation Study + +We report ablation study on Mr. HiSum [74] in Tab. 5, comparing performance adding a component step-by-step. Starting from a two-layer transformer encoder predicting the importance score, adding our denoiser with a naively summed query $\mathbf{Q}_t = \mathcal{C}(\mathbf{u}_t) + \pmb{\tau} + \pmb{\Phi}$ improves the Kendall's $\tau$ $(0.071\rightarrow 0.079)$ . We observe further slight improvement $(0.079\to 0.086)$ when we replace the fixed random score embeddings with learnable ones (Learnable Embedding). Introducing AdaLN transformation [58] on the time $(\tau)$ and positional $(\Phi)$ encodings further improves $\tau$ to 0.145 and 0.171, respectively. As we adopt the quantization idea, grouping nearby scores into a unified embedding, in Sec. 3.2 with this change, we experiment without it (using a sinusoidal function [80] instead), and observe significant performance drop $(0.171\rightarrow 0.125)$ . This confirms that quantization is superior to treating the score + +
Modelτ↑ρ↑
CSTA [67]0.1280.185
SummDiff (1 step)0.1700.234
SummDiff (10 steps)0.1750.238
SummDiff (100 steps)0.1820.245
+ +Table 6. Performance with various numbers of DDIM steps + +![](images/d05a072a6fd027b8eff8f686d74093d6cca3b5b0e967d64ba25169f53789094a.jpg) +Figure 6. Measured $\mathbb{E}|\{\pmb{y}^*\}|$ and $\mathbb{E}[\Delta \pmb{y}^*]$ with respect to the quantization strength $K$ . A larger $K$ is associated with a smaller number of allowed solutions for the given KP. + +as a continuous scalar. On top of this, we combine our model with two widely-known ideas, classifier-free guidance (CFG) [24] and self-Attention guidance (SAG) [30], for further improvement. See Appendix B for details. + +Number of DDIM Steps. Tab. 6 suggests that SummDiff outperforms CSTA [67], the best-performing baseline, even with a single DDIM [68] step. Beyond this, SummDiff further improves its performance with more DDIM steps. + +Effect of Quantization Strength. Solving a KP usually has a unique solution if the profits $v_{i}$ are continuous. As quantization unifies importance score values within the same interval to a single embedding, it may lead to multiple solutions to $\mathcal{KP}(\tilde{\boldsymbol{v}},\boldsymbol {w},\rho)$ , where $\tilde{\boldsymbol{v}}$ is a quantized version of $\boldsymbol{v}$ . + +To analyze this effect, we measure the average number of solutions $\mathbb{E}_{\pmb{v}_1\sim U(0,1)\otimes N}[\{\pmb{y}^*\in \{0,1\} ^N |\pmb{y}^* = \mathcal{KP}(\tilde{\pmb{v}},\pmb {w},\rho)\} ]$ while varying $K$ , taking expectation over $\pmb{v}$ , which is a value vector sampled from a joint uniform distribution. We also consider the average $L_{1}$ difference between two summaries, $\mathbb{E}_{\pmb{v}_1\sim U(0,1)\otimes N},\pmb{v}_2\sim \mathcal{N}(\pmb{v}_1,\frac{1}{K}\mathbf{I}_N)[\| \mathcal{KP}(\tilde{\pmb{v}}_1,\pmb {w},\rho) - \mathcal{KP}(\tilde{\pmb{v}}_2,\pmb {w},\rho)\| _1]$ , obtained by solving the KP with two similar but distinct value vectors, i.e., $\pmb{v}_2 = \pmb{v}_1 + \Delta \pmb{v}$ , where $\Delta \pmb{v}\sim \mathcal{N}(0,\frac{1}{K}\mathbf{I}_N)$ . + +We numerically analyze the relation between these two values, which indicate the scale of multiplicity of the solutions to the KP, and the quantization strength $K$ . As shown in Fig. 6, both the number of solutions $|\{\pmb{y}^*\}|$ and the summary deviation $\Delta \pmb{y}^*$ decrease with larger $K$ , suggesting more accurate summaries. + +To sum up, increasing the quantization strength $K$ reduces the number of multiple optimal solutions, leading to a more accurate summary. However, an excessively higher $K$ leads to insufficient samples per bin, degrading the overall performance. See Appendix A for more ablation studies. + +# 5. Related Work + +Video Summarization. Early models primarily rely on heuristic unsupervised learning [13, 15, 29, 38, 47, 49-51, 53, 71, 83] to select important or diverse frames, struggling with generalization. Benefiting from annotated datasets, supervised methods [8, 38, 57, 59, 60] have emerged. DSNet [93] and iPTNet [34] improve keyframe selection using frame-level annotations, and SL-Module [82] captures high-level features. With the advent of deep learning, RNNs [84, 88-90] and attention mechanism [4, 6, 33, 35, 39, 42, 77, 85, 94] have been adopted to capture temporal dependencies in videos. Particularly, VASNet [17] and PGL-SUM [3] capture both local and global frame dependencies using self-attention. DMASum [81] introduces mixture of attention layer mitigate the key Softmax Bottleneck. SUM-GAN [50] and AC-SUM-GAN [2] leverage generative adversarial networks. + +Multimodal approaches [32, 91] integrate audio or textual data to video summarization. Multimodal transformers are adopted to link frames with corresponding captions, improving context-aware summaries, by CLIP-It [52], MSVA [18], SSPVS [40], and A2Summ [22]. CSTA [67] addresses computational complexity by using CNN-based spatiotemporal attention for efficient frame selection. Unlike these deterministic summarizers, our approach generates multiple plausible summaries by capturing the distribution of video summaries with various perspectives. + +Diffusion for Video Tasks. Diffusion [66] have emerged as a groundbreaking tool in high-quality image generation [21, 25, 54, 63, 64, 69, 70] and super-resolution [14, 36, 61, 65]. Recently, they are extended to video generation [7, 11, 16, 19, 26, 27, 44], joint video and audio generation [41, 62, 73], video editing [9, 10, 79], video inpainting [87] and prediction [31, 86]. Furthermore, they are applied to video understanding tasks like moment retrieval [43, 48], video object segmentation [92, 95], action segmentation [45]. + +# 6. Summary + +Video summarization is inherently subjective since people have different perspectives of a good summary. We suggest a generative viewpoint of this task where the model learns the distribution of good summaries, in contrast to the traditional approach of aggregated importance score regression. Our proposed SummDiff model, adopting diffusion models for video summarization for the first time, is able to generate multiple good summaries conditioned on the input video. Our model not only outperforms other baselines but also demonstrates the ability to generate accurate summaries customized to the individual annotators. We further propose additional metrics that measure the quality of the predicted importance scores through an insight from the actual summary generation using knapsack. + +# Acknowledgments + +This work was supported by Samsung Electronics (IO240512-09881-01), Youlchon Foundation, NRF grants (RS-2021-NR05515, RS-2024-00336576, RS-2023-0022663) and IITP grants (RS-2022-II220264, RS-2024-00353131) by the government of Korea. + +# References + +[1] Sami Abu-El-Haija, Nisarg Kothari, Joonseok Lee, Paul Natsev, George Toderici, Balakrishnan Varadarajan, and Sudheendra Vijayanarasimhan. Youtube-8M: A large-scale video classification benchmark. arXiv:1609.08675, 2016. 4 +[2] Evlampios Apostolidis, Eleni Adamantidou, Alexandros I Metsai, Vasileios Mezaris, and Ioannis Patras. AC-SUMGAN: Connecting actor-critic and generative adversarial networks for unsupervised video summarization. IEEE Transactions on Circuits and Systems for Video Technology, 31(8): 3278-3292, 2020. 1, 4, 8, iii +[3] Evlampios Apostolidis, Georgios Balaouras, Vasileios Mezaris, and Ioannis Patras. Combining global and local attention with positional encoding for video summarization. In IEEE international symposium on multimedia (ISM), 2021. 1, 2, 4, 7, 8, i, iii +[4] Evlampios Apostolidis, Georgios Balaouras, Vasileios Mezaris, and Ioannis Patras. Summarizing videos using concentrated attention and considering the uniqueness and diversity of the video frames. In ICMR, 2022. 8 +[5] Tarik Belgacem and Mhand Hifi. Sensitivity analysis of the knapsack problem: Tighter lower and upper bound limits. Journal of systems science and systems engineering, 17:156-170, 2008. 7 +[6] Manjot Bilkhu, Siyang Wang, and Tushar Dobhal. Attention is all you need for videos: Self-attention based video summarization using universal transformers. arXiv:1906.02792, 2019.8 +[7] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 8 +[8] Sijia Cai, Wangmeng Zuo, Larry S Davis, and Lei Zhang. Weakly-supervised video summarization using variational encoder-decoder and web prior. In ECCV, 2018. 8 +[9] Duygu Ceylan, Chun-Hao P Huang, and Niloy J Mitra. Pix2video: Video editing using image diffusion. In ICCV, 2023. 8 +[10] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. StableVideo: Text-driven consistency-aware diffusion video editing. In ICCV, 2023. 8 +[11] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. VideoCrafter2: Overcoming data limitations for high-quality video diffusion models. In CVPR, 2024. 8 +[12] Ting Chen, Ruixiang Zhang, and Geoffrey Hinton. Analog bits: Generating discrete data using diffusion models with self-conditioning. In ICLR, 2022. 3 + +[13] Wen-Sheng Chu, Yale Song, and Alejandro Jaimes. Video co-summarization: Video summarization by visual co-occurrence. In CVPR, 2015. 8 +[14] Hyungjin Chung, Byeongsu Sim, and Jong Chul Ye. Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In CVPR, 2022. 8 +[15] Sandra Eliza Fontes De Avila, Ana Paula Brandao Lopes, Antonio da Luz Jr, and Arnaldo de Albuquerque Araujo. Vsumm: A mechanism designed to produce static video summaries and a novel evaluation method. Pattern recognition letters, 32(1):56-68, 2011. 8 +[16] Patrick Esser, Johnathan Chiu, Parmida Atighechian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In ICCV, 2023. 8 +[17] Jiri Fajtl, Hajar Sadeghi Sokeh, Vasileios Argyriou, Dorothy Monekosso, and Paolo Remagnino. Summarizing videos with attention. In ACCV Workshop, 2019. 1, 2, 4, 7, 8, i, iii +[18] Junaid Ahmed Ghauri, Sherzod Hakimov, and Ralph Ewerth. Supervised video summarization via multiple feature sets with parallel attention. In ICME, 2021. 8, ii +[19] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Fei-Fei Li, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. In ECCV, 2025. 8 +[20] Michael Gygli, Helmut Grabner, Hayko Riemenschneider, and Luc Van Gool. Creating summaries from user videos. In ECCV, 2014. 1, 4, 5 +[21] Jaehoon Hahm, Junho Lee, Sunghyun Kim, and Joonseok Lee. Isometric representation learning for disentangled latent space of diffusion models. In ICML, 2024. 8 +[22] Bo He, Jun Wang, Jielin Qiu, Trung Bui, Abhinav Shrivastava, and Zhaowen Wang. Align and attend: Multimodal summarization with dual contrastive losses. In CVPR, 2023. 1, 4, 7, 8, i, iii +[23] Mhand Hifi, Hedi Mhalla, and Slim Sadfi. Sensitivity of the optimum to perturbations of the profit or weight of an item in the binary knapsack problem. Journal of Combinatorial Optimization, 10:239-260, 2005. 4, 6 +[24] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshop on Deep Generative Models and Downstream Applications, 2021. 1, 7, 8, i +[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 3, 8 +[26] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv:2210.02303, 2022. 8 +[27] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In NeurIPS, 2022. 8 +[28] Fa-Ting Hong, Xuanteng Huang, Wei-Hong Li, and Wei-Shi Zheng. Mini-net: Multiple instance ranking network for video highlight detection. In ECCV, 2020. 4 + +[29] Richang Hong, Jinhui Tang, Hung-Khoon Tan, Shuicheng Yan, Chongwah Ngo, and Tat-Seng Chua. Event driven summarization for web videos. In SIGMM workshop on Social media, 2009. 8 +[30] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seunglyong Kim. Improving sample quality of diffusion models using self-attention guidance. In ICCV, 2023. 7, 8, i +[31] Tobias Hörppe, Arash Mehrjou, Stefan Bauer, Didrik Nielsen, and Andrea Dittadi. Diffusion models for video prediction and infilling. arXiv:2206.07696, 2022. 8 +[32] Tzu-Chun Hsu, Yi-Sheng Liao, and Chun-Rong Huang. Video summarization with spatiotemporal vision transformer. IEEE Transactions on Image Processing, 2023. 8, ii +[33] Jia-Hong Huang, Luka Murn, Marta Mrak, and Marcel Worring. GPT2MVS: Generative pre-trained transformer-2 for multi-modal video summarization. In ICMR, 2021. 8 +[34] Hao Jiang and Yadong Mu. Joint video summarization and moment localization by cross-task sample transfer. In CVPR, 2022. 1, 4, 8, iii +[35] Yunjae Jung, Donghyeon Cho, Sanghyun Woo, and In So Kweon. Global-and-local relative position embedding for unsupervised video summarization. In ECCV, 2020. 8 +[36] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. In NeurIPS, 2022. 8 +[37] Maurice G Kendall. The treatment of ties in ranking problems. Biometrika, 33(3):239-251, 1945. 4 +[38] Aditya Khosla, Raffay Hamid, Chih-Jen Lin, and Neel Sundaresan. Large-scale video summarization using web-image priors. In CVPR, 2013. 8 +[39] Haopeng Li, Qiuhong Ke, Mingming Gong, and Rui Zhang. Video joint modelling based on hierarchical transformer for co-summulation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3904–3917, 2022. 8 +[40] Haopeng Li, Qiuhong Ke, Mingming Gong, and Tom Drummond. Progressive video summarization via multimodal self-supervised learning. In WACV, 2023. 4, 8, ii +[41] Judith Li, Xuchan Bao, Zhong Yi Wan, Kun Su, Timo Denk, Dima Kuzmin, Joonseok Lee, and Fei Sha. Diff4steer: Steerable diffusion prior for generative music retrieval with semantic guidance. In ICASSP, 2024. 8 +[42] Ping Li, Qinghao Ye, Luming Zhang, Li Yuan, Xianghua Xu, and Ling Shao. Exploring global diverse attention via pairwise temporal relation for video summarization. Pattern Recognition, 111:107677, 2021. 8 +[43] Pandeng Li, Chen-Wei Xie, Hongtao Xie, Liming Zhao, Lei Zhang, Yun Zheng, Deli Zhao, and Yongdong Zhang. MomentDiff: Generative video moment retrieval from random to real. In NeurIPS, 2024. 8 +[44] Jingyun Liang, Yuchen Fan, Kai Zhang, Radu Timofte, Luc Van Gool, and Rakesh Ranjan. MoVideo: Motion-aware video generation with diffusion model. In ECCV, 2025. 8 +[45] Daochang Liu, Qiyue Li, Anh-Dung Dinh, Tingting Jiang, Mubarak Shah, and Chang Xu. Diffusion action segmentation. In ICCV, 2023. 8 + +[46] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2017. ii +[47] Zheng Lu and Kristen Grauman. Story-driven summarization for egocentric video. In CVPR, 2013. 8 +[48] Dezhao Luo, Shaogang Gong, Jiabo Huang, Hailin Jin, and Yang Liu. Generative video diffusion for unseen cross-domain video moment retrieval. arXiv:2401.13329, 2024. 8 +[49] Yu-Fei Ma, Lie Lu, Hong-Jiang Zhang, and Mingjing Li. A user attention model for video summarization. In ACM MM, 2002. 8 +[50] Behrooz Mahasseni, Michael Lam, and Sinisa Todorovic. Unsupervised video summarization with adversarial LSTM networks. In CVPR, 2017. 1, 4, 8, ii, iii +[51] Padmavathi Mundur, Yong Rao, and Yelena Yesha. Keyframe-based video summarization using delaunay clustering. International Journal on Digital Libraries, 6:219-232, 2006. 8 +[52] Medhini Narasimhan, Anna Rohrbach, and Trevor Darrell. Clip-it! language-guided video summarization. In NeurIPS, 2021. 4, 8 +[53] Chong-Wah Ngo, Yu-Fei Ma, and Hong-Jiang Zhang. Automatic video summarization by graph modeling. In ICCV, 2003. 8 +[54] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 1, 8 +[55] Mayu Otani, Yuta Nakashima, Esa Rahtu, and Janne Heikkila. Rethinking the evaluation of video summaries. In CVPR, 2019. 4 +[56] Rameswar Panda and Amit K Roy-Chowdhury. Collaborative summarization of topic-related videos. In CVPR, 2017. 4 +[57] Rameswar Panda, Abir Das, Ziyan Wu, Jan Ernst, and Amit K Roy-Chowdhury. Weakly supervised summarization of web videos. In ICCV, 2017. 8 +[58] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 2, 3, 7 +[59] Danila Potapov, Matthijs Douze, Zaid Harchaoui, and Cordelia Schmid. Category-specific video summarization. In ECCV, 2014. 4, 8 +[60] Mrigank Rochan and Yang Wang. Video summarization by learning from unpaired data. In CVPR, 2019. 8 +[61] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 8 +[62] Ludan Ruan, Yiyang Ma, Huan Yang, Huiguo He, Bei Liu, Jianlong Fu, Nicholas Jing Yuan, Qin Jin, and Baining Guo. MM-diffusion: Learning multi-modal diffusion models for joint audio and video generation. In CVPR, 2023. 8 +[63] Suho Ryu, Kihyun Kim, Eugene Baek, Dongsoo Shin, and Joonseok Lee. Towards scalable human-aligned benchmark for text-guided image editing. In CVPR, 2025. 8 +[64] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, + +Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 8 +[65] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE transactions on pattern analysis and machine intelligence, 45(4):4713-4726, 2022. 8 +[66] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 8 +[67] Jaewon Son, Jaehun Park, and Kwangsu Kim. CSTA: CNN-based spatiotemporal attention for video summarization. In CVPR, 2024. 1, 2, 4, 5, 7, 8, i, ii, iii +[68] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In *ICLR*, 2020. 3, 8, ii +[69] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In NeurIPS, 2019. 8 +[70] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. In NeurIPS, 2020. 8 +[71] Yale Song, Jordi Vallmitjana, Amanda Stent, and Alejandro Jaimes. TVSum: Summarizing web videos using titles. In CVPR, 2015. 1, 4, 5, 8 +[72] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2020. 1 +[73] Kun Su, Judith Li, Qingqing Huang, Dima Kuzmin, Joon-seok Lee, Chris Donahue, Fei Sha, Aren Jansen, Yu Wang, Mauro Verzetti, and Timo Denk. V2Meow: Meowing to the visual beat via video-to-music generation. In AAAI, 2024. 8 +[74] Jinhwan Sul, Jihoon Han, and Joonseok Lee. Mr. HiSum: A large-scale dataset for video highlight detection and summarization. In NeurlPS, 2024. 4, 5, 7 +[75] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In CVPR, 2015. 4 +[76] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016. 4 +[77] Hacene Terbouche, Maryan Morel, Mariano Rodriguez, and Alice Othmani. Multi-annotation attention model for video summarization. In CVPR, 2023. 4, 8 +[78] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. ii +[79] Shuyuan Tu, Qi Dai, Zhi-Qi Cheng, Han Hu, Xintong Han, Zuxuan Wu, and Yu-Gang Jiang. MotionEditor: Editing video motion via content-aware diffusion. In CVPR, 2024. 8 +[80] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, 2017. 2, 3, 7 + +[81] Junyan Wang, Yang Bai, Yang Long, Bingzhang Hu, Zhenhua Chai, Yu Guan, and Xiaolin Wei. Query twice: Dual mixture attention meta learning for video summarization. In ACM MM, 2020. 4, 5, 8 +[82] Minghao Xu, Hang Wang, Bingbing Ni, Riheng Zhu, Zhenbang Sun, and Changhu Wang. Cross-category video highlight detection via set-based learning. In ICCV, 2021. 1, 4, 7, 8, i, iii +[83] Hong Jiang Zhang, Jianhua Wu, Di Zhong, and Stephen W Smoliar. An integrated system for content-based video retrieval and browsing. Pattern recognition, 30(4):643-658, 1997. 8 +[84] Ke Zhang, Wei-Lun Chao, Fei Sha, and Kristen Grauman. Video summarization with long short-term memory. In ECCV, 2016. 4, 8, ii +[85] Yunzuo Zhang, Yameng Liu, Weili Kang, and Ran Tao. VSS-Net: visual semantic self-mining network for video summarization. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 4, 8 +[86] Zhicheng Zhang, Junyao Hu, Wentao Cheng, Danda Paudel, and Jufeng Yang. ExtDM: Distribution extrapolation diffusion model for video prediction. In CVPR, 2024. 8 +[87] Zhixing Zhang, Bichen Wu, Xiaoyan Wang, Yaqiao Luo, Luxin Zhang, Yinan Zhao, Peter Vajda, Dimitris Metaxas, and Licheng Yu. AVID: Any-length video inpainting with diffusion model. In CVPR, 2024. 8 +[88] Bin Zhao, Xuelong Li, and Xiaoqiang Lu. Hierarchical recurrent neural network for video summarization. In ACM MM, 2017. 8 +[89] Bin Zhao, Xuelong Li, and Xiaogiang Lu. HSA-RNN: Hierarchical structure-adaptive rnn for video summarization. In CVPR, 2018. +[90] Bin Zhao, Xuelong Li, and Xiaogiang Lu. TTH-RNN: Tensor-train hierarchical recurrent neural network for video summarization. IEEE Transactions on Industrial Electronics, 68(4):3629-3637, 2020. 8 +[91] Bin Zhao, Maoguo Gong, and Xuelong Li. Hierarchical multimodal transformer to summarize videos. Neurocomputing, 468:360-369, 2022. 8 +[92] Chuanwei Zhou, Chunyan Xu, Jun Li, Zhen Cui, and Jian Yang. Quality-aware pattern diffusion for video object segmentation. Neurocomputing, 528:148-159, 2023. 8 +[93] Wencheng Zhu, Jiwen Lu, Jiahao Li, and Jie Zhou. DSNet: A flexible detect-to-summarize network for video summarization. IEEE Transactions on Image Processing, 30:948–962, 2020. 4, 8 +[94] Wencheng Zhu, Yucheng Han, Jiwen Lu, and Jie Zhou. Relational reasoning over spatial-temporal graphs for video summarization. IEEE Transactions on Image Processing, 31: 3017-3031, 2022. 8 +[95] Zixin Zhu, Xuelu Feng, Dongdong Chen, Junsong Yuan, Chunming Qiao, and Gang Hua. Exploring pre-trained text-to-video diffusion models for referring video object segmentation. arXiv:2403.12042, 2024. 8 +[96] Daniel Zwillinger and Stephen Kokoska. CRC standard probability and statistics tables and formulae. CRC Press, 1999. 4 \ No newline at end of file diff --git a/2025/SummDiff_ Generative Modeling of Video Summarization with Diffusion/images.zip b/2025/SummDiff_ Generative Modeling of Video Summarization with Diffusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8e6c79aa776b8cc5931704c6cc653b2c3cde1b29 --- /dev/null +++ b/2025/SummDiff_ Generative Modeling of Video Summarization with Diffusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:36d18ea97d03c234712e0594f5c280f9983c943de6ce11dfd22e08c28249a686 +size 542681 diff --git a/2025/SummDiff_ Generative Modeling of Video Summarization with Diffusion/layout.json b/2025/SummDiff_ Generative Modeling of Video Summarization with Diffusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0602758309f6390bb043f719d1034239e82ed202 --- /dev/null +++ b/2025/SummDiff_ Generative Modeling of Video Summarization with Diffusion/layout.json @@ -0,0 +1,12680 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 80, + 103, + 531, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 103, + 531, + 120 + ], + "spans": [ + { + "bbox": [ + 80, + 103, + 531, + 120 + ], + "type": "text", + "content": "SummDiff: Generative Modeling of Video Summarization with Diffusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 142, + 538, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 142, + 538, + 171 + ], + "spans": [ + { + "bbox": [ + 75, + 142, + 538, + 171 + ], + "type": "text", + "content": "Kwanseok Kim" + }, + { + "bbox": [ + 75, + 142, + 538, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 75, + 142, + 538, + 171 + ], + "type": "text", + "content": ", Jaehoon Hahm" + }, + { + "bbox": [ + 75, + 142, + 538, + 171 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 75, + 142, + 538, + 171 + ], + "type": "text", + "content": ", Sumin Kim" + }, + { + "bbox": [ + 75, + 142, + 538, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 142, + 538, + 171 + ], + "type": "text", + "content": ", Jinhwan Sul" + }, + { + "bbox": [ + 75, + 142, + 538, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 75, + 142, + 538, + 171 + ], + "type": "text", + "content": ", Byunghak Kim" + }, + { + "bbox": [ + 75, + 142, + 538, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 75, + 142, + 538, + 171 + ], + "type": "text", + "content": ", Joonseok Lee" + }, + { + "bbox": [ + 75, + 142, + 538, + 171 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 75, + 142, + 538, + 171 + ], + "type": "text", + "content": ", Seoul National University, 2UIUC, Georgia Institute of Technology, Hyundai Card" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 147, + 174, + 459, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 174, + 459, + 199 + ], + "spans": [ + { + "bbox": [ + 147, + 174, + 459, + 199 + ], + "type": "text", + "content": "kjvd1009@snu.ac.kr, jh141@illinois.edu, sumink@snu.ac.kr, jsul7@gatech.edu, byunghak.kim@hcs.com, joonseok@snu.ac.kr" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 253, + 297, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 253, + 297, + 504 + ], + "spans": [ + { + "bbox": [ + 54, + 253, + 297, + 504 + ], + "type": "text", + "content": "Video summarization is a task of shortening a video by choosing a subset of frames while preserving its essential moments. Despite the innate subjectivity of the task, previous works have deterministically regressed to an averaged frame score over multiple raters, ignoring the inherent subjectivity of what constitutes a \"good\" summary. We propose a novel problem formulation by framing video summarization as a conditional generation task, allowing a model to learn the distribution of good summaries and to generate multiple plausible summaries that better reflect varying human perspectives. Adopting diffusion models for the first time in video summarization, our proposed method, SummDiff, dynamically adapts to visual contexts and generates multiple candidate summaries conditioned on the input video. Extensive experiments demonstrate that SummDiff not only achieves the state-of-the-art performance on various benchmarks but also produces summaries that closely align with individual annotator preferences. Moreover, we provide a deeper insight with novel metrics from an analysis of the knapsack, which is an important last step of generating summaries but has been overlooked in evaluation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 529, + 135, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 135, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 135, + 542 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 551, + 296, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 296, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 296, + 682 + ], + "type": "text", + "content": "Recently, short-form videos draw significant attention on video sharing platforms, with a trend that consumers increasingly prefer to quickly grasp the content. They often compressively convey contents that are originally in a longer form, summarizing the core contents into a shorter one; e.g., sport games highlights or movie summarization. This task of selecting core parts of a long video to construct a shorter one is called video summarization. This task is inherently subjective, since there can be multiple criteria for a 'good summary'; e.g., comprehensively covering the entire storyline or subjectively selecting impressive parts of" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 228, + 555, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 276 + ], + "type": "text", + "content": "the video (highlight detection). Due to this inherent subjectivity, most video summarization or highlight detection datasets [20, 71] offer annotations by multiple raters to reflect various perspectives." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 278, + 556, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 556, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 556, + 456 + ], + "type": "text", + "content": "Since each annotator may have different opinion on the importance of a frame, most existing methods [3, 17, 22, 34, 67, 82] take the frame-level importance scores averaged across multiple annotators as their target label, and are trained to predict them. This frame-level score aggregation looks reasonable in some sense, but in fact it loses the various perspectives to summarize each video. For instance, suppose half of the annotators select clips from the first quarter of the video while the other half select clips from the last quarter. If one simply averages their frame scores, both the first and last quarters end up with similar importance, obscuring the two distinct valid summaries. That is, this simple regression to the averaged frame-level importance scores fail to preserve multiple viewpoints to summarize the video." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 460, + 556, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 556, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 556, + 639 + ], + "type": "text", + "content": "In order to preserve and reflect various views to summarize a video, we pose a distribution of its good summaries and let the model to learn it, instead of giving an already-aggregated single ground truth importance score. Then, the video summarization task can be seen from a generative perspective; i.e., a process of learning and inferring the distribution of good summaries conditioned on the input video. Specifically, the model is now in charge of estimating the distribution of plausible summaries for the given video. Once trained, it allows us to sample multiple summaries from the estimated conditional data manifold. To the best of our knowledge, this generative approach to the video summarization has not been extensively studied, except for a few works [2, 50] that applied adversarial losses to construct a summary that looks like the original video." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "text", + "content": "Formulating the video summarization problem as a conditional generation task, we propose to adopt the generative diffusion [25, 72] mechanism, which has been successfully applied to various conditional generation tasks [24, 54, 61]. Specifically, conditioned on the input video, our proposed SummDiff model learns to denoise a random importance" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15096" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 68, + 376, + 196 + ], + "blocks": [ + { + "bbox": [ + 57, + 68, + 376, + 196 + ], + "lines": [ + { + "bbox": [ + 57, + 68, + 376, + 196 + ], + "spans": [ + { + "bbox": [ + 57, + 68, + 376, + 196 + ], + "type": "image", + "image_path": "b84fa467829d5ca6a0793d03fe6eec5da4442d119ef6b672ccfffe9d9e9054db.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 204, + 376, + 227 + ], + "lines": [ + { + "bbox": [ + 55, + 204, + 376, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 376, + 227 + ], + "type": "text", + "content": "Figure 1. Overview of SummDiff. Given an input video, SummDiff generates importance scores conditioned on video frames. " + }, + { + "bbox": [ + 55, + 204, + 376, + 227 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 204, + 376, + 227 + ], + "type": "text", + "content": " denotes the number of DDIM steps." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 384, + 56, + 552, + 188 + ], + "blocks": [ + { + "bbox": [ + 384, + 56, + 552, + 188 + ], + "lines": [ + { + "bbox": [ + 384, + 56, + 552, + 188 + ], + "spans": [ + { + "bbox": [ + 384, + 56, + 552, + 188 + ], + "type": "image", + "image_path": "9f842a7c643f7740f4f686512caadea0c71a5be9b0b95427cdab875affe6f21b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 193, + 552, + 227 + ], + "lines": [ + { + "bbox": [ + 384, + 193, + 552, + 227 + ], + "spans": [ + { + "bbox": [ + 384, + 193, + 552, + 227 + ], + "type": "text", + "content": "Figure 2. Video Importance Score Denoiser. We use AdaLN layer to inject time and positional conditions, following [58]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 239, + 296, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 239, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 54, + 239, + 296, + 334 + ], + "type": "text", + "content": "score vector over the given video into an importance score vector sampled from the distribution that corresponds to a good summary of the video. In contrast to the previous deterministic methods, our approach allows to sample multiple plausible summaries for a given video starting from a different random noise, better aligned with the subjective nature of summarization where we usually have various true labels reflecting multiple views." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 335, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 296, + 407 + ], + "type": "text", + "content": "Extensive experiments demonstrate that our model outperforms existing baselines across multiple datasets. Also, we revisit the knapsack, a relatively unexplored step in the summarization evaluation in spite of its nontrivial impact on the performance, and propose additional novel metrics based on this analysis to provide deeper insights." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 407, + 266, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 407, + 266, + 418 + ], + "spans": [ + { + "bbox": [ + 67, + 407, + 266, + 418 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 419, + 296, + 526 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 61, + 419, + 295, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 419, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 61, + 419, + 295, + 455 + ], + "type": "text", + "content": "- We propose a novel generative viewpoint of video summarization, better suited for the subjective nature of the task, allowing multiple plausible summaries for a video." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 456, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 456, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 61, + 456, + 296, + 491 + ], + "type": "text", + "content": "- We innovatively apply diffusion to the video summarization for the first time, integrating learning the distribution of good summaries into the model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 491, + 296, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 491, + 296, + 526 + ], + "spans": [ + { + "bbox": [ + 61, + 491, + 296, + 526 + ], + "type": "text", + "content": "- We analyze the knapsack optimization process and propose additional metrics to quantify the optimality of the predicted importance scores." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 537, + 181, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 537, + 181, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 537, + 181, + 550 + ], + "type": "text", + "content": "2. Problem Formulation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 557, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 296, + 616 + ], + "type": "text", + "content": "Given a video of " + }, + { + "bbox": [ + 55, + 557, + 296, + 616 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 557, + 296, + 616 + ], + "type": "text", + "content": " frames, the objective of video summarization is to identify and select " + }, + { + "bbox": [ + 55, + 557, + 296, + 616 + ], + "type": "inline_equation", + "content": "S < N" + }, + { + "bbox": [ + 55, + 557, + 296, + 616 + ], + "type": "text", + "content": " frames that effectively encapsulate the essence of the video content. Let " + }, + { + "bbox": [ + 55, + 557, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\equiv \\{\\mathbf{X}_i \\in \\mathbb{R}^{H \\times W \\times 3} : i = 0, \\dots, N - 1\\} \\in \\mathbb{R}^{N \\times H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 557, + 296, + 616 + ], + "type": "text", + "content": " be a video, where " + }, + { + "bbox": [ + 55, + 557, + 296, + 616 + ], + "type": "inline_equation", + "content": "H, W" + }, + { + "bbox": [ + 55, + 557, + 296, + 616 + ], + "type": "text", + "content": " denote the size of the frames." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": "A video summary, " + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{y} \\equiv \\{y_i \\in \\{0,1\\} : i = 0,\\dots,N-1, \\sum_i y_i \\leq S\\} \\in \\{0,1\\}^N" + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": ", indicates inclusion (1) or exclusion (0) of each frame. When a video is provided with multiple ground truth annotations, we denote each individual score by " + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{s}^{(r)} \\in [0,1]^N" + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": " and the corresponding binary summary by " + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^{(r)} \\in \\{0,1\\}^N" + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": " which is obtained following the procedure explained in Sec. 3.3. The predicted summary is denoted by " + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}} \\equiv \\{\\hat{y}_i \\in \\mathbb{R} : i = 0,\\dots,N-1\\} \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 239, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 555, + 384 + ], + "type": "text", + "content": "Previous models [3, 17, 67] have approached video summarization as a regression task, aiming to predict the importance score " + }, + { + "bbox": [ + 313, + 239, + 555, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{s} \\in [0,1]^N" + }, + { + "bbox": [ + 313, + 239, + 555, + 384 + ], + "type": "text", + "content": " for each video, often the average of multiple importance scores, " + }, + { + "bbox": [ + 313, + 239, + 555, + 384 + ], + "type": "inline_equation", + "content": "\\frac{1}{|R|} \\sum_{r \\in \\mathcal{R}} \\mathbf{s}^{(r)}" + }, + { + "bbox": [ + 313, + 239, + 555, + 384 + ], + "type": "text", + "content": ". In contrast, our approach allows multiple summaries for each video, aiming to learn the distribution of its plausible summaries. The previous scheme can be seen as a special case of ours, where every video has a single golden way of summarization, pointed at " + }, + { + "bbox": [ + 313, + 239, + 555, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 313, + 239, + 555, + 384 + ], + "type": "text", + "content": " with zero variance. Under our extended setting, multiple importance scores " + }, + { + "bbox": [ + 313, + 239, + 555, + 384 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{s}^{(r)} | r \\in \\mathcal{R}\\}" + }, + { + "bbox": [ + 313, + 239, + 555, + 384 + ], + "type": "text", + "content": " can be given to the same conditioning video " + }, + { + "bbox": [ + 313, + 239, + 555, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 239, + 555, + 384 + ], + "type": "text", + "content": ", forming a probability distribution of plausible importance scores." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 396, + 525, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 525, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 525, + 408 + ], + "type": "text", + "content": "3. Diffusion-based Video Summarization" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 416, + 555, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 416, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 555, + 464 + ], + "type": "text", + "content": "Posing the video summarization as a conditional generation task, we introduce our SummDiff model, designed to adapt the distribution of individual importance scores for a given video by learning to denoise." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 473, + 512, + 486 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 512, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 512, + 486 + ], + "type": "text", + "content": "3.1. Overall Flow of the Proposed Method" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 491, + 547, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 547, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 547, + 503 + ], + "type": "text", + "content": "Fig. 1 illustrates the overall flow of our SummDiff model." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 506, + 554, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 554, + 578 + ], + "type": "text", + "content": "Encoding Process. We first encode each frame " + }, + { + "bbox": [ + 313, + 506, + 554, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i" + }, + { + "bbox": [ + 313, + 506, + 554, + 578 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 506, + 554, + 578 + ], + "type": "inline_equation", + "content": "i = 1, \\dots, N" + }, + { + "bbox": [ + 313, + 506, + 554, + 578 + ], + "type": "text", + "content": " from the input video " + }, + { + "bbox": [ + 313, + 506, + 554, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 506, + 554, + 578 + ], + "type": "text", + "content": " using a pre-trained image encoder. The extracted features are further contextualized through self-attention [80], as seen in Fig. 1. We denote the encoded feature for each individual frame by " + }, + { + "bbox": [ + 313, + 506, + 554, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 313, + 506, + 554, + 578 + ], + "type": "text", + "content": ", and collectively the entire feature matrix by " + }, + { + "bbox": [ + 313, + 506, + 554, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 313, + 506, + 554, + 578 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 582, + 555, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 716 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 716 + ], + "type": "text", + "content": "Denoising Process. We then learn to denoise an individual importance score vector from a random noise, conditioned on the visual embeddings. First, the forward process adds noise to the ground truth individual importance score " + }, + { + "bbox": [ + 313, + 582, + 555, + 716 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_0 \\equiv \\mathbf{s}^{(r)}" + }, + { + "bbox": [ + 313, + 582, + 555, + 716 + ], + "type": "text", + "content": " and sample a noised one " + }, + { + "bbox": [ + 313, + 582, + 555, + 716 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{s}_0 + \\sqrt{1 - \\bar{\\alpha}_t}\\boldsymbol{\\epsilon}_t" + }, + { + "bbox": [ + 313, + 582, + 555, + 716 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 582, + 555, + 716 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\epsilon}_t \\sim \\mathcal{N}(0,\\mathbf{I}), \\bar{\\alpha}_t = \\prod_{\\tau=1}^t (1 - \\beta_\\tau)" + }, + { + "bbox": [ + 313, + 582, + 555, + 716 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 582, + 555, + 716 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 582, + 555, + 716 + ], + "type": "text", + "content": " is the diffusion time step. The perturbation kernel at " + }, + { + "bbox": [ + 313, + 582, + 555, + 716 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 582, + 555, + 716 + ], + "type": "text", + "content": ", defined as " + }, + { + "bbox": [ + 313, + 582, + 555, + 716 + ], + "type": "inline_equation", + "content": "q(\\mathbf{s}_t|\\mathbf{s}_{t-1}) = \\mathcal{N}(\\mathbf{s}_t; \\sqrt{1 - \\beta_t}\\mathbf{s}_{t-1}, \\beta_t\\mathbf{I})" + }, + { + "bbox": [ + 313, + 582, + 555, + 716 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 582, + 555, + 716 + ], + "type": "inline_equation", + "content": "\\beta_t" + }, + { + "bbox": [ + 313, + 582, + 555, + 716 + ], + "type": "text", + "content": " is defined by the variance schedule. Since " + }, + { + "bbox": [ + 313, + 582, + 555, + 716 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_0" + }, + { + "bbox": [ + 313, + 582, + 555, + 716 + ], + "type": "text", + "content": " is bounded within " + }, + { + "bbox": [ + 313, + 582, + 555, + 716 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 313, + 582, + 555, + 716 + ], + "type": "text", + "content": ", it is not straightforward to add Gaussian noise directly to it so we first transform it to its logit, " + }, + { + "bbox": [ + 313, + 582, + 555, + 716 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_0 = \\log \\frac{\\mathbf{s}_0}{1 - \\mathbf{s}_0} \\in" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15097" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^N" + }, + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "type": "text", + "content": ", and perform the noising process in this logit space. For numerical stability, we clip " + }, + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_0" + }, + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "type": "inline_equation", + "content": "[\\epsilon, 1 - \\epsilon]" + }, + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "type": "text", + "content": " is a small constant." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "type": "text", + "content": "Then, the reverse process progressively removes noise from " + }, + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_t" + }, + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_0" + }, + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "type": "text", + "content": ", formulated by " + }, + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{s}_{t - 1}|\\mathbf{s}_t) = \\mathcal{N}(\\mathbf{s}_{t - 1};\\boldsymbol {\\mu}_\\theta (\\mathbf{s}_t,t),\\sigma_t^2\\mathbf{I})" + }, + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "type": "inline_equation", + "content": "\\sigma_t^2" + }, + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "type": "text", + "content": " is determined in relation to " + }, + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "type": "text", + "content": ", and the posterior mean " + }, + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_{\\theta}(\\mathbf{s}_t,t)" + }, + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "type": "text", + "content": " is modeled with a trainable neural network. In this paper, we extend this estimator of posterior mean to be conditioned on a video, similarly to the conditional generation on diffusion models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 192, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 294, + 251 + ], + "type": "text", + "content": "Once trained, our denoiser is able to recover a plausible importance score for a video from a random noise, which is used to generate a summary (Sec. 3.3). Repeated generation across the noise distribution would converge to the true distribution of plausible scores for the video." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 259, + 235, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 259, + 235, + 272 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 235, + 272 + ], + "type": "text", + "content": "3.2. Video Importance Score Denoiser" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "text", + "content": "We formulate the key component of our method, the video importance score denoiser in Fig. 2. Starting from a noised logit score vector " + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_t \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "text", + "content": ", it predicts a plausible importance score, conditioned on the given video. Denoted by " + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\mathbf{u}_t, t, \\mathbf{Z})" + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "text", + "content": ", it learns to denoise the given score vector " + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_t" + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "text", + "content": " at the diffusion time step " + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "text", + "content": " under the visual condition " + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "text", + "content": ", producing the denoised score " + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{u}}_{t-1}" + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "text", + "content": " by one time step, where " + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "text", + "content": " is the set of its learnable parameters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 376, + 296, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 296, + 471 + ], + "type": "text", + "content": "Transformer-based Diffusion. We denoise the logit-transformed score " + }, + { + "bbox": [ + 55, + 376, + 296, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_t" + }, + { + "bbox": [ + 55, + 376, + 296, + 471 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 55, + 376, + 296, + 471 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 376, + 296, + 471 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 376, + 296, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 55, + 376, + 296, + 471 + ], + "type": "text", + "content": ", predicting " + }, + { + "bbox": [ + 55, + 376, + 296, + 471 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{u}}_{t-1}" + }, + { + "bbox": [ + 55, + 376, + 296, + 471 + ], + "type": "text", + "content": ", using a transformer-based cross-attention [80]. The logit-transformed score " + }, + { + "bbox": [ + 55, + 376, + 296, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_t" + }, + { + "bbox": [ + 55, + 376, + 296, + 471 + ], + "type": "text", + "content": " acts as the query, and the visual condition " + }, + { + "bbox": [ + 55, + 376, + 296, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 55, + 376, + 296, + 471 + ], + "type": "text", + "content": " is used as key/value. This setup allows the model to effectively denoise " + }, + { + "bbox": [ + 55, + 376, + 296, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_t" + }, + { + "bbox": [ + 55, + 376, + 296, + 471 + ], + "type": "text", + "content": " conditioned on the information from the input video, ensuring consistency between the predicted logit score " + }, + { + "bbox": [ + 55, + 376, + 296, + 471 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{u}}_{t-1}" + }, + { + "bbox": [ + 55, + 376, + 296, + 471 + ], + "type": "text", + "content": " and the condition " + }, + { + "bbox": [ + 55, + 376, + 296, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 55, + 376, + 296, + 471 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 472, + 296, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 614 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 614 + ], + "type": "text", + "content": "To apply dot-product cross-attention, the dimensionality should match for queries and keys. To this end, we quantize the importance scores into a predefined number " + }, + { + "bbox": [ + 55, + 472, + 296, + 614 + ], + "type": "inline_equation", + "content": "(K)" + }, + { + "bbox": [ + 55, + 472, + 296, + 614 + ], + "type": "text", + "content": " of uniform segments. Specifically, we convert the noised logit " + }, + { + "bbox": [ + 55, + 472, + 296, + 614 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_t" + }, + { + "bbox": [ + 55, + 472, + 296, + 614 + ], + "type": "text", + "content": " back to its original bounded range " + }, + { + "bbox": [ + 55, + 472, + 296, + 614 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 55, + 472, + 296, + 614 + ], + "type": "text", + "content": " by applying sigmoid " + }, + { + "bbox": [ + 55, + 472, + 296, + 614 + ], + "type": "inline_equation", + "content": "1 / (1 + e^{-\\mathbf{u}_t})" + }, + { + "bbox": [ + 55, + 472, + 296, + 614 + ], + "type": "text", + "content": ", and divide them into " + }, + { + "bbox": [ + 55, + 472, + 296, + 614 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 472, + 296, + 614 + ], + "type": "text", + "content": " equally-binned segments. Each score range is associated with a learnable embedding of size " + }, + { + "bbox": [ + 55, + 472, + 296, + 614 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 472, + 296, + 614 + ], + "type": "text", + "content": " (codebook in Fig. 1). Based on the codebook " + }, + { + "bbox": [ + 55, + 472, + 296, + 614 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 472, + 296, + 614 + ], + "type": "text", + "content": ", we map the scores " + }, + { + "bbox": [ + 55, + 472, + 296, + 614 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_t \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 55, + 472, + 296, + 614 + ], + "type": "text", + "content": " to their corresponding quantized embedding " + }, + { + "bbox": [ + 55, + 472, + 296, + 614 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(\\mathbf{u}_t) \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 55, + 472, + 296, + 614 + ], + "type": "text", + "content": ". Our denoiser " + }, + { + "bbox": [ + 55, + 472, + 296, + 614 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\mathcal{C}(\\mathbf{u}_t), t, \\mathbf{Z})" + }, + { + "bbox": [ + 55, + 472, + 296, + 614 + ], + "type": "text", + "content": " takes this " + }, + { + "bbox": [ + 55, + 472, + 296, + 614 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(\\mathbf{u}_t)" + }, + { + "bbox": [ + 55, + 472, + 296, + 614 + ], + "type": "text", + "content": " as input, instead of " + }, + { + "bbox": [ + 55, + 472, + 296, + 614 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_t" + }, + { + "bbox": [ + 55, + 472, + 296, + 614 + ], + "type": "text", + "content": ". See Sec. 4.5 for analysis on the size of codebook." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 617, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 677 + ], + "type": "text", + "content": "Query Formation in Cross-Attention. Following [25], the time step " + }, + { + "bbox": [ + 55, + 617, + 296, + 677 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 617, + 296, + 677 + ], + "type": "text", + "content": " is embedded as " + }, + { + "bbox": [ + 55, + 617, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\pmb{\\tau} \\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 55, + 617, + 296, + 677 + ], + "type": "text", + "content": " using sinusoidal functions and an MLP. Considering the sequential nature of videos, we also introduce temporal positional embeddings " + }, + { + "bbox": [ + 55, + 617, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\Phi \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 55, + 617, + 296, + 677 + ], + "type": "text", + "content": " using sinusoidal functions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "The simplest way to set the query in the cross-attention would be to add all inputs except for the condition " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "; that is, " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_t = \\mathcal{C}(\\mathbf{u}_t) + \\boldsymbol{\\tau} + \\boldsymbol{\\Phi}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": ". However, we leverage AdaLN-Zero" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "text", + "content": "block [58] to integrate the time embedding " + }, + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "inline_equation", + "content": "\\pmb{\\tau}" + }, + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "text", + "content": " and temporal positional encoding " + }, + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "text", + "content": " more effectively by separating them from the query " + }, + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_t" + }, + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "text", + "content": ", preventing information mixing. Hence, the query becomes " + }, + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_t = \\mathcal{C}(\\mathbf{u}_t)" + }, + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "inline_equation", + "content": "\\pmb{\\tau}, \\Phi" + }, + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "text", + "content": " are conditioned via scale-shift operation. See Sec. 4.5 for ablation studies." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 133, + 554, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 554, + 217 + ], + "type": "text", + "content": "We conduct cross-attention with " + }, + { + "bbox": [ + 313, + 133, + 554, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_t" + }, + { + "bbox": [ + 313, + 133, + 554, + 217 + ], + "type": "text", + "content": " as queries and visual conditions " + }, + { + "bbox": [ + 313, + 133, + 554, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 313, + 133, + 554, + 217 + ], + "type": "text", + "content": " as keys " + }, + { + "bbox": [ + 313, + 133, + 554, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 313, + 133, + 554, + 217 + ], + "type": "text", + "content": " and values " + }, + { + "bbox": [ + 313, + 133, + 554, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 313, + 133, + 554, + 217 + ], + "type": "text", + "content": ". We take an MLP from the AdaLN output to regress the scale and shift parameters, denoted by " + }, + { + "bbox": [ + 313, + 133, + 554, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_1" + }, + { + "bbox": [ + 313, + 133, + 554, + 217 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 133, + 554, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_1" + }, + { + "bbox": [ + 313, + 133, + 554, + 217 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 133, + 554, + 217 + ], + "type": "inline_equation", + "content": "\\Gamma_1" + }, + { + "bbox": [ + 313, + 133, + 554, + 217 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 133, + 554, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_2" + }, + { + "bbox": [ + 313, + 133, + 554, + 217 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 133, + 554, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_2" + }, + { + "bbox": [ + 313, + 133, + 554, + 217 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 133, + 554, + 217 + ], + "type": "inline_equation", + "content": "\\Gamma_2 \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 313, + 133, + 554, + 217 + ], + "type": "text", + "content": ". As depicted in Fig. 2, they scale and shift " + }, + { + "bbox": [ + 313, + 133, + 554, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_t" + }, + { + "bbox": [ + 313, + 133, + 554, + 217 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 133, + 554, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 313, + 133, + 554, + 217 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 133, + 554, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 313, + 133, + 554, + 217 + ], + "type": "text", + "content": ". Then, they are passed through cross-attention with skip connections and a subsequent rescaling. Formally," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 351, + 227, + 466, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 227, + 466, + 239 + ], + "spans": [ + { + "bbox": [ + 351, + 227, + 466, + 239 + ], + "type": "interline_equation", + "content": "\\mathbf {H} ^ {i ^ {\\prime}} = \\boldsymbol {\\Gamma} _ {1} \\odot \\mathbf {H} ^ {i} + \\mathbf {H} ^ {i} + \\mathbf {B} _ {1}", + "image_path": "bec50874e546262a788aa36b7f4fff4359199c57310abe91647677e5dc9cd4fc.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 354, + 243, + 525, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 243, + 525, + 258 + ], + "spans": [ + { + "bbox": [ + 354, + 243, + 525, + 258 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {1} = \\mathbf {A} _ {1} \\odot \\operatorname {s o f t m a x} \\left(\\mathbf {Q} _ {t} ^ {i ^ {\\prime}} \\mathbf {K} ^ {i ^ {\\prime} \\top}\\right) \\mathbf {V} ^ {i ^ {\\prime}} + \\mathbf {Q} _ {t} ^ {i ^ {\\prime}}", + "image_path": "070b39345235f67878e515ba4886ce1e6445b5deafff2987878b797bb1932e93.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 354, + 262, + 468, + 273 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 262, + 468, + 273 + ], + "spans": [ + { + "bbox": [ + 354, + 262, + 468, + 273 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {2} = \\boldsymbol {\\Gamma} _ {2} \\odot \\mathbf {X} _ {1} + \\mathbf {X} _ {1} + \\mathbf {B} _ {2}", + "image_path": "8a2dd833f225bfbc2d1779457c4c14dc6528099d73b4bc061123892533f950aa.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 345, + 277, + 476, + 290 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 277, + 476, + 290 + ], + "spans": [ + { + "bbox": [ + 345, + 277, + 476, + 290 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} _ {t} ^ {i + 1} = \\mathbf {A} _ {2} \\odot \\operatorname {M L P} (\\mathbf {X} _ {2}) + \\mathbf {X} _ {2},", + "image_path": "23820feb15ab088795c4ce864b2d1b316de52e5fc9479b3278470a6db3ce0b73.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 300, + 553, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 553, + 341 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 300, + 553, + 341 + ], + "type": "inline_equation", + "content": "\\mathbf{H}^{i^{(t)}}\\in \\{\\mathbf{Q}_t^{i^{(t)}},\\mathbf{K}^{i^{(t)}},\\mathbf{V}^{i^{(t)}}\\}" + }, + { + "bbox": [ + 313, + 300, + 553, + 341 + ], + "type": "text", + "content": " denotes the matrices used for attention, " + }, + { + "bbox": [ + 313, + 300, + 553, + 341 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 300, + 553, + 341 + ], + "type": "text", + "content": " is the (broadcasted) Hadamard product, and " + }, + { + "bbox": [ + 313, + 300, + 553, + 341 + ], + "type": "inline_equation", + "content": "i = 1,\\ldots ,L" + }, + { + "bbox": [ + 313, + 300, + 553, + 341 + ], + "type": "text", + "content": " is the layer index." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 344, + 554, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 344, + 554, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 344, + 554, + 392 + ], + "type": "text", + "content": "Training. We train the denoiser " + }, + { + "bbox": [ + 313, + 344, + 554, + 392 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 313, + 344, + 554, + 392 + ], + "type": "text", + "content": " to estimate the true importance score " + }, + { + "bbox": [ + 313, + 344, + 554, + 392 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{s}}_0" + }, + { + "bbox": [ + 313, + 344, + 554, + 392 + ], + "type": "text", + "content": " after acting upon a fully-connected layer and a sigmoid function " + }, + { + "bbox": [ + 313, + 344, + 554, + 392 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 344, + 554, + 392 + ], + "type": "text", + "content": ". We minimize the following loss [12] on each annotator's individual importance score:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 402, + 552, + 417 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 402, + 552, + 417 + ], + "spans": [ + { + "bbox": [ + 314, + 402, + 552, + 417 + ], + "type": "interline_equation", + "content": "\\mathcal {L} \\left(\\mathbf {s} _ {0}, \\hat {\\mathbf {s}} _ {0}\\right) = \\left\\| \\mathbf {s} _ {0} - \\hat {\\mathbf {s}} _ {0} \\right\\| _ {2} ^ {2} = \\left\\| \\mathbf {s} _ {0} - \\sigma \\left(\\operatorname {F C} \\left(f _ {\\theta} \\left(\\mathcal {C} \\left(\\mathbf {u} _ {t}\\right), t, \\mathbf {Z}\\right)\\right)\\right) \\right\\| _ {2} ^ {2}.", + "image_path": "17dcb324130d7e13cc56387780f90757b5a9ce537b23dedec483d6b1efe3e648.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 426, + 553, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 553, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 553, + 462 + ], + "type": "text", + "content": "The learnable embeddings in the codebook " + }, + { + "bbox": [ + 313, + 426, + 553, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 426, + 553, + 462 + ], + "type": "text", + "content": " are compositely optimized, finding an effective representation for " + }, + { + "bbox": [ + 313, + 426, + 553, + 462 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_t" + }, + { + "bbox": [ + 313, + 426, + 553, + 462 + ], + "type": "text", + "content": " during training." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 466, + 553, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 553, + 514 + ], + "type": "text", + "content": "Inference. Our model generates a logit-transformed importance score from a random noise " + }, + { + "bbox": [ + 313, + 466, + 553, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_T \\sim \\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 313, + 466, + 553, + 514 + ], + "type": "text", + "content": " for a given video. Employing the reverse diffusion process [68], it iteratively refines the logit score towards cleaner estimations:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 329, + 524, + 539, + 566 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 524, + 539, + 566 + ], + "spans": [ + { + "bbox": [ + 329, + 524, + 539, + 566 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {\\mathbf {u}} _ {t - 1} = \\sqrt {1 - \\bar {\\alpha} _ {t - 1} - \\sigma_ {t} ^ {2}} \\frac {\\hat {\\mathbf {u}} _ {t} - \\sqrt {\\bar {\\alpha} _ {t}} f _ {\\theta} (\\mathcal {C} (\\mathbf {u} _ {t}) , t , \\mathbf {Z}))}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\\\ + \\sqrt {\\bar {\\alpha} _ {t - 1}} f _ {\\theta} (\\mathcal {C} (\\mathbf {u} _ {t}), t, \\mathbf {Z})) + \\sigma_ {t} \\epsilon_ {t} \\in \\mathbb {R} ^ {N}, \\\\ \\end{array}", + "image_path": "5e9448513153b39b9c1c75f5f3cae3c63ca103d129df4d3b4cd9316eba14811b.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 577, + 554, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 577, + 554, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 577, + 554, + 639 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 577, + 554, + 639 + ], + "type": "inline_equation", + "content": "\\sigma_{t} = \\sqrt{(1 - \\bar{\\alpha}_{t - 1}) / (1 - \\bar{\\alpha}_{t})}\\sqrt{1 - \\bar{\\alpha}_{t} / \\bar{\\alpha}_{t - 1}}" + }, + { + "bbox": [ + 313, + 577, + 554, + 639 + ], + "type": "text", + "content": " (DDPM reverse process). When " + }, + { + "bbox": [ + 313, + 577, + 554, + 639 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 313, + 577, + 554, + 639 + ], + "type": "text", + "content": " at the last step, " + }, + { + "bbox": [ + 313, + 577, + 554, + 639 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{u}}_0 = f_\\theta (\\mathcal{C}(\\mathbf{u}_1),1,\\mathbf{Z})" + }, + { + "bbox": [ + 313, + 577, + 554, + 639 + ], + "type": "text", + "content": " is directly used to remove stochasticity at the end of inference process. Finally, we take sigmoid to " + }, + { + "bbox": [ + 313, + 577, + 554, + 639 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{u}}_0" + }, + { + "bbox": [ + 313, + 577, + 554, + 639 + ], + "type": "text", + "content": " to derive the final importance score " + }, + { + "bbox": [ + 313, + 577, + 554, + 639 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{s}}_0\\in (0,1)^N" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 647, + 440, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 440, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 440, + 659 + ], + "type": "text", + "content": "3.3. Summary Generation" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "From the raw importance score " + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{s}}_0" + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": ", we apply the standard knapsack-based approach to decide which frames to be included in the final summary. To make this summary video more realistic, it is common to choose at a semantic clip" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15098" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": "level instead of individual frame level. We adopt a widely-used Kernel Temporal Segmentation (KTS) [59, 84] to partition a video into disjoint temporal intervals. We take the average score among the corresponding frames as the clip-level importance; i.e., " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "v_{i} = \\sum_{j = t_{i}}^{t_{i + 1}}\\hat{\\mathbf{s}}_{0,j} / (t_{i + 1} - t_{i})" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": "-th clip composed of frames from " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "t_{i}" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "t_{i + 1}" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": ". Subsequently, we select the clips based on " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " by solving the binary knapsack problem (KP) [23] with dynamic programming [71]:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 177, + 290, + 210 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 177, + 290, + 210 + ], + "spans": [ + { + "bbox": [ + 61, + 177, + 290, + 210 + ], + "type": "interline_equation", + "content": "\\mathcal {K P} (\\boldsymbol {v}, \\boldsymbol {w}, \\rho) \\equiv \\underset {\\boldsymbol {s} \\in \\{0, 1 \\} ^ {M}} {\\operatorname {a r g m a x}} \\sum_ {i = 1} ^ {M} v _ {i} s _ {i} \\quad \\text {s . t .} \\sum_ {i = 1} ^ {M} w _ {i} s _ {i} \\leq \\rho N,", + "image_path": "7a18ada69a3a6834b2cdd1bb7829ee8acc270d35bf5ee072b6ec0ee6570b97cc.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 219, + 296, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 219, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 219, + 296, + 291 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 219, + 296, + 291 + ], + "type": "inline_equation", + "content": "\\pmb{v},\\pmb{w},\\rho" + }, + { + "bbox": [ + 55, + 219, + 296, + 291 + ], + "type": "text", + "content": " denotes the values of each clip " + }, + { + "bbox": [ + 55, + 219, + 296, + 291 + ], + "type": "inline_equation", + "content": "(v_{i}\\in [0,1])" + }, + { + "bbox": [ + 55, + 219, + 296, + 291 + ], + "type": "text", + "content": ", costs of selecting each clip, and the budget constraint ratio (e.g., " + }, + { + "bbox": [ + 55, + 219, + 296, + 291 + ], + "type": "inline_equation", + "content": "\\rho = 0.15" + }, + { + "bbox": [ + 55, + 219, + 296, + 291 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 55, + 219, + 296, + 291 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 55, + 219, + 296, + 291 + ], + "type": "text", + "content": " of the video length), respectively. If the argmax has multiple solutions, we will abuse the notation and denote " + }, + { + "bbox": [ + 55, + 219, + 296, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{KP}(\\pmb {v},\\pmb {w},\\rho)" + }, + { + "bbox": [ + 55, + 219, + 296, + 291 + ], + "type": "text", + "content": " as an arbitrary element of the solution set." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 302, + 137, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 302, + 137, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 302, + 137, + 316 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 322, + 182, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 322, + 182, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 182, + 335 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 339, + 296, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 296, + 470 + ], + "type": "text", + "content": "Datasets. We evaluate our method on three benchmarks. TVSum [71] and SumMe [20] are traditional datasets, containing 50 and 25 videos respectively and manually labeled annotations by up to 20 raters. Mr. HiSum [74] is a large-scale dataset, composed of 31,892 videos and importance scores derived from YouTube Most Replayed statistics. It provides an importance score " + }, + { + "bbox": [ + 55, + 339, + 296, + 470 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_0" + }, + { + "bbox": [ + 55, + 339, + 296, + 470 + ], + "type": "text", + "content": " and a corresponding summary " + }, + { + "bbox": [ + 55, + 339, + 296, + 470 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_0" + }, + { + "bbox": [ + 55, + 339, + 296, + 470 + ], + "type": "text", + "content": " for each video, averaged over 50,000+ viewers. We use the default features for each dataset, GoogLeNet [75] for TVSum and SumMe, and Inception-v3 [76] PCA-ed to 1024D [1] for Mr. HiSum." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 472, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 579 + ], + "type": "text", + "content": "We adopt two dataset splits. First, we randomly split TVSum and SumMe into 7:1:2 for train, validation, and test set, and report averaged test scores for 5 random splits (namely, TVT). Also, following prior works [67, 77], we include five-fold cross-validation results (5FCV), although these may suffer from overfitting to the test set due to the lack of explicit validation set, and thus may not generalize well on unseen videos. For Mr. HiSum, we use its original train/validation/test split. We run all experiments 5 times." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "Evaluation Metrics. Following recent practice [22, 40, 55, 67], we report rank order statistics, Kendall's " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " [37] and Spearman's " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " [96], to assess how well the model ranks frame importance. F1 score had been widely used in video summarization, but recent literature [67, 77] reports that it is excessively sensitive to video segmentation and unreasonably favors summaries composed of many short shots [55], while disregarding longer key shots [67]. As an alternative to F1, we propose additional metrics that represent the influence of importance score through an analysis of the knapsack algorithm in Sec. 4.4." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 333, + 70, + 535, + 183 + ], + "blocks": [ + { + "bbox": [ + 333, + 70, + 535, + 183 + ], + "lines": [ + { + "bbox": [ + 333, + 70, + 535, + 183 + ], + "spans": [ + { + "bbox": [ + 333, + 70, + 535, + 183 + ], + "type": "table", + "html": "
MethodSumMeTVSum
τρτρ
Random0.0000.0000.0000.000
Human0.2050.2130.1770.204
A2Summ [22]0.0880.0960.1570.206
VASNet [17]0.0890.0990.1530.205
PGL-SUM [3]0.1040.1160.1410.186
CSTA [67]0.1080.1200.1680.221
SummDiff (Ours)0.1330.1480.1730.226
", + "image_path": "df654f5d1cb98b267b5ef5e03974cf7e0ff2176ef07e2718947dd01c963023cd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 332, + 224, + 535, + 441 + ], + "blocks": [ + { + "bbox": [ + 313, + 186, + 555, + 220 + ], + "lines": [ + { + "bbox": [ + 313, + 186, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 555, + 220 + ], + "type": "text", + "content": "Table 1. Comparison of models trained under TVT (train/val/test split) on SumMe [20] and TVSum [71]. Best and second-best results are boldfaced and underlined, respectively." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 224, + 535, + 441 + ], + "lines": [ + { + "bbox": [ + 332, + 224, + 535, + 441 + ], + "spans": [ + { + "bbox": [ + 332, + 224, + 535, + 441 + ], + "type": "table", + "html": "
MethodSumMeTVSum
τρτρ
Random0.0000.0000.0000.000
Human0.2050.2130.1770.204
DSNet-AF [93]0.0370.0460.1130.138
DSNet-AB [93]0.0510.0590.1080.129
SUM-GAN [50]0.0490.0660.0240.031
AC-SUM-GAN [2]0.1020.0880.0310.041
CLIP-It [52]--0.1080.147
iPTNet [34]0.1010.1190.1340.163
A2Summ [22]0.1080.1290.1370.165
VASNet [17]0.1600.1700.1600.170
PGL-SUM [3]--0.1570.206
AAAM [77]--0.1690.223
MAAM [77]--0.1790.236
VSS-Net [85]--0.1900.249
DMASum [81]0.0630.0890.2030.267
SSPVS [40]0.1920.2570.1810.238
CSTA [67]0.2460.2740.1940.255
SummDiff (Ours)0.2560.2850.1950.255
", + "image_path": "e6b4bb92ef07c2a69120ab8ffc13102d9402d1a78e87b3a48b3d0c8bf87f1114.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 321, + 482, + 545, + 594 + ], + "blocks": [ + { + "bbox": [ + 313, + 444, + 555, + 478 + ], + "lines": [ + { + "bbox": [ + 313, + 444, + 555, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 555, + 478 + ], + "type": "text", + "content": "Table 2. Comparison of models trained under 5FCV (5-Fold Cross Validation) on SumMe [20] and TVSum [71]. Training under the 5FCV setting tends to overfit the test set." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 482, + 545, + 594 + ], + "lines": [ + { + "bbox": [ + 321, + 482, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 321, + 482, + 545, + 594 + ], + "type": "table", + "html": "
Modelτ↑ρ↑MAPρ=50%↑MAPρ=15%↑
SUM-GAN [50]0.0670.09556.6223.56
VASNet [17]0.0690.10258.6925.28
AC-SUM-GAN [2]0.0120.01855.3521.88
SL-module [82]0.0600.08858.6324.95
PGL-SUM [3]0.0970.14161.6027.45
iPTNet [34]0.0200.02955.5322.74
A2Summ [22]0.1210.17263.2032.34
CSTA [67]0.1280.18563.3830.42
SummDiff0.1750.23865.4433.83
", + "image_path": "2b9557a29c7a75350871861a7141eb95acfacd494bd9ea976b44ca6d9b6c228f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 597, + 555, + 620 + ], + "lines": [ + { + "bbox": [ + 313, + 597, + 555, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 555, + 620 + ], + "type": "text", + "content": "Table 3. Evaluation on Mr. HiSum [74]. See Appendix D for the full table including standard deviations." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": "We further evaluate our method on video highlight detection following [74]. First, we uniformly divide the input video into 5-second-long shots and calculate the average frame scores for each shot. The top " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\rho \\in \\{0.15, 0.5\\}" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": " of these shots are designated as ground truth highlights. We report Mean Average Precision (MAP), following [28, 56, 84]. See Appendix E for more implementation details." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15099" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 175, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 175, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 175, + 83 + ], + "type": "text", + "content": "4.2. Results and Analysis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 304 + ], + "type": "text", + "content": "Considering the innate subjectivity of video summarization, we train our model to learn from each individual score, allowing it to capture multiple ways of summarization for each video. Specifically, we use 20 individual importance scores per video in TVSum [71] and 15-18 individual binary summaries in SumMe [20], since SumMe does not provide individual importance scores. Tab. 1 and 2 show that SummDiff achieves the best performance on SumMe and TVSum, with the sole exception of TVSum in 5FCV. While DMA-SUM [81] performs the best on TVSum in 5FCV, it significantly underperforms on SumMe, indicating that its high performance on TVSum does not generalize well. Comparing TVT and 5FCV, most models exhibit a considerable performance gap. This demonstrates how much existing models have overfitted to the test set under the conventional 5FCV protocol. We adopt TVT as a remedy to this, but due to the extremely small size of these datasets, the reliability of evaluation is still limited [74]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 305, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 305, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 305, + 296, + 413 + ], + "type": "text", + "content": "On the larger Mr. HiSum, we train our model using the aggregated annotation, as it does not provide individual annotations. This evaluation mainly aims to verify scalability of SummDiff on a large-scale dataset. Tab. 3 presents the performance on Mr. HiSum, where SummDiff consistently outperforms all baselines across all metrics, significantly surpassing the strongest competitor, CSTA [67] with a large margin. These results highlight scalability and effectiveness of our method even under the single-label setting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 421, + 203, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 203, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 203, + 433 + ], + "type": "text", + "content": "4.3. Qualitative Demonstration" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 438, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 296, + 714 + ], + "type": "text", + "content": "We demonstrate how our method discovers a variety of possible summaries for a video. Specifically, we generate video summaries for the 50 videos in TVSum using CSTA, PGL-SUM, VASNet, and ours. For each generated summary, we measure the Kendall's " + }, + { + "bbox": [ + 55, + 438, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 438, + 296, + 714 + ], + "type": "text", + "content": " with all 20 ground truth annotations, respectively. Considering " + }, + { + "bbox": [ + 55, + 438, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\tau \\geq 0.25" + }, + { + "bbox": [ + 55, + 438, + 296, + 714 + ], + "type": "text", + "content": " as the threshold for the summary to be matched with the annotator, Fig. 3 shows if the generated summary for a video (row) matches with each annotator (column). As the baselines (CSTA, PGL-SUM, and VASNet) summarize a video deterministically, each cell is either completely matched (1) or not (0). If the summary matches with majority of annotators (e.g., video 29 for CSTA), it means the deterministic summary matches with the dominant way of summarization for the video. If it matches only with a few annotators (e.g., video 24 for CSTA), it means the generated summary matches with a minor opinion. For our method, we generate 100 summaries per video from different Gaussian noise vectors, and mark each cell with the ratio of summaries that match with each annotator. The darker the cell colors are for each row, the more various ways of summaries have been discovered by our method. Comparing the heat maps, we clearly observe that SummDiff generates significantly more various" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 331, + 65, + 432, + 190 + ], + "blocks": [ + { + "bbox": [ + 331, + 65, + 432, + 190 + ], + "lines": [ + { + "bbox": [ + 331, + 65, + 432, + 190 + ], + "spans": [ + { + "bbox": [ + 331, + 65, + 432, + 190 + ], + "type": "image", + "image_path": "f6cca2b7b476a0a7178d63f359fccc011c31a9fb1f76beb79ca54083ab5574d8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 432, + 65, + 539, + 190 + ], + "blocks": [ + { + "bbox": [ + 432, + 65, + 539, + 190 + ], + "lines": [ + { + "bbox": [ + 432, + 65, + 539, + 190 + ], + "spans": [ + { + "bbox": [ + 432, + 65, + 539, + 190 + ], + "type": "image", + "image_path": "ed7787ef43b384971c21f8eb6d490f9d15b74c6150f5154fa81abe62b82c9536.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 331, + 191, + 432, + 320 + ], + "blocks": [ + { + "bbox": [ + 331, + 191, + 432, + 320 + ], + "lines": [ + { + "bbox": [ + 331, + 191, + 432, + 320 + ], + "spans": [ + { + "bbox": [ + 331, + 191, + 432, + 320 + ], + "type": "image", + "image_path": "c77634b845e23c2160ffb7c7cd3817622f8142f6232c9fb7d822ba099c7450fd.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 432, + 191, + 539, + 320 + ], + "blocks": [ + { + "bbox": [ + 432, + 191, + 539, + 320 + ], + "lines": [ + { + "bbox": [ + 432, + 191, + 539, + 320 + ], + "spans": [ + { + "bbox": [ + 432, + 191, + 539, + 320 + ], + "type": "image", + "image_path": "bc2e7b4fbbc547903332dbb2ce3759e013d81c4480865824204160c0bb4986e8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 331, + 389, + 434, + 487 + ], + "blocks": [ + { + "bbox": [ + 313, + 323, + 555, + 389 + ], + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 389 + ], + "type": "text", + "content": "Figure 3. Ratio of summaries with " + }, + { + "bbox": [ + 313, + 323, + 555, + 389 + ], + "type": "inline_equation", + "content": "\\tau \\geq 0.25" + }, + { + "bbox": [ + 313, + 323, + 555, + 389 + ], + "type": "text", + "content": " for each video-annotator pair. The heatmap illustrates how closely each method's summary matches with individual annotations. This comparison reveals the extent to which each method captures the varied human summaries. SummDiff covers a larger area of the heatmap, which indicates better coverage over various summaries." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 331, + 389, + 434, + 487 + ], + "lines": [ + { + "bbox": [ + 331, + 389, + 434, + 487 + ], + "spans": [ + { + "bbox": [ + 331, + 389, + 434, + 487 + ], + "type": "image", + "image_path": "9df82d52124190e43a3d21ab5eb0b84ba032c661885ddca815c6f2cd07ebfdec.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 434, + 389, + 539, + 487 + ], + "blocks": [ + { + "bbox": [ + 434, + 389, + 539, + 487 + ], + "lines": [ + { + "bbox": [ + 434, + 389, + 539, + 487 + ], + "spans": [ + { + "bbox": [ + 434, + 389, + 539, + 487 + ], + "type": "image", + "image_path": "a883bc0e6af6c6b10d15d29b157630f6feedc0d4f7a0f108e61f696362092211.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 330, + 487, + 435, + 586 + ], + "blocks": [ + { + "bbox": [ + 330, + 487, + 435, + 586 + ], + "lines": [ + { + "bbox": [ + 330, + 487, + 435, + 586 + ], + "spans": [ + { + "bbox": [ + 330, + 487, + 435, + 586 + ], + "type": "image", + "image_path": "18a5f862fe166f908f6ba15ecdaa5b03d9cde469cf7c05c78d9f499218bd91f7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 589, + 520, + 596 + ], + "lines": [ + { + "bbox": [ + 352, + 589, + 520, + 596 + ], + "spans": [ + { + "bbox": [ + 352, + 589, + 520, + 596 + ], + "type": "text", + "content": "SummDiff CSTA PGL_SUM VASNet Annotator" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 313, + 600, + 555, + 645 + ], + "lines": [ + { + "bbox": [ + 313, + 600, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 555, + 645 + ], + "type": "text", + "content": "Figure 4. Distribution of true and predicted summaries of given videos. SummDiff can generate various summaries and cover the targeted distribution of summaries, while baselines deterministically predicts a single summary." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 435, + 487, + 539, + 586 + ], + "blocks": [ + { + "bbox": [ + 435, + 487, + 539, + 586 + ], + "lines": [ + { + "bbox": [ + 435, + 487, + 539, + 586 + ], + "spans": [ + { + "bbox": [ + 435, + 487, + 539, + 586 + ], + "type": "image", + "image_path": "3ae161d9bbdee32928283589715ed8b7ad9d1dd34952dabb9e1f4ab851441c37.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 653, + 553, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 553, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 553, + 676 + ], + "type": "text", + "content": "summaries covering multiple viewpoints, better reflecting the distribution of annotated summaries." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "We further illustrate the variation and quality of generated summaries in Fig. 4. The summaries annotated by human raters ( " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": " Annotator) and those generated by ours ( " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\bullet" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15100" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 224, + 555, + 258 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 555, + 258 + ], + "type": "text", + "content": "Figure 5. Demonstration of video summaries generated by competing methods on a TVSum video. Shaded parts indicate the segments selected by each method, and the two rows of edged boxes within each method indicate the two distinct true annotations. The results clearly demonstrate the effectiveness of SummDiff in capturing multiple plausible summaries for a video. See another example in Appendix F." + } + ] + } + ], + "index": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 268, + 294, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 268, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 54, + 268, + 294, + 459 + ], + "type": "text", + "content": "SummDiff) and baselines ( • CSTA, • PGL-SUM, and • VASNet) are projected onto the first two principal components, which are computed using only the human-annotated summaries. The distribution of true summaries are displayed using a contour map, obtained by Gaussian kernel density estimation. It demonstrates that our SummDiff generates multiple summaries closely aligned with the distribution of true annotations. For example, video 22 (top-left) roughly has three ways to summarize, according to the human raters. SummDiff generates summaries dominantly covering the two modes in the lower-left region, and one case in the upper-right mode. In contrast, baselines produce a single, less accurate summary, failing to account for the variety inherent in human-annotated summaries. We observe similar patterns in the other three plots as well. Fig. II provides more examples in Appendix C." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 460, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 581 + ], + "type": "text", + "content": "Fig. 5 illustrates example summaries generated by CSTA, A2Summ, PGL-SUM, VASNet, and our SummDiff, for a video selected from TVSum test set. Shaded parts indicate the segments selected by each method (note that SummDiff can generate two different summaries while baselines always produce a single summary), and the two rows of edged boxes within each method indicate two different true annotations. These results demonstrate that SummDiff produces more accurate and various summaries, effectively capturing multiple plausible summarizations for each video." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 587, + 286, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 286, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 286, + 601 + ], + "type": "text", + "content": "4.4. Metrics Inspired from Knapsack Optimality" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "In spite of nontrivial impact of the knapsack at the end of the summary generation, previous evaluation metrics have focused only on the accuracy of the importance scores. Through a thorough analysis on the knapsack problem (KP), we provide additional metrics that measure the contribution of importance scores. By accounting for knapsack optimality and clip-level weights, our new metrics are capable of assessing the predicted importance score more accurately than existing metrics such as F1 or ranking-based ones." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 268, + 553, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 553, + 327 + ], + "type": "text", + "content": "Confidence of Importance Score. First, we analyze the conditions under which the same optimal KP solution remains valid despite some perturbations to the profits " + }, + { + "bbox": [ + 313, + 268, + 553, + 327 + ], + "type": "inline_equation", + "content": "\\pmb{v}" + }, + { + "bbox": [ + 313, + 268, + 553, + 327 + ], + "type": "text", + "content": ". These perturbations can be a modeling to the imperfect estimation of the importance scores in video summarization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 327, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 533 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 327, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\pmb{y}^{*}" + }, + { + "bbox": [ + 313, + 327, + 555, + 533 + ], + "type": "text", + "content": " be the optimal solution to the original KP (i.e., " + }, + { + "bbox": [ + 313, + 327, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\pmb{y}^{*} = \\mathcal{KP}(\\pmb{v},\\pmb{w},\\rho)" + }, + { + "bbox": [ + 313, + 327, + 555, + 533 + ], + "type": "text", + "content": "; see definition in Sec. 3.3) and " + }, + { + "bbox": [ + 313, + 327, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\Gamma \\subseteq [N]\\equiv \\{n\\in \\mathbb{N}\\mid 1\\leq n\\leq N\\}" + }, + { + "bbox": [ + 313, + 327, + 555, + 533 + ], + "type": "text", + "content": " be an arbitrary subset of items associated with the perturbed profits " + }, + { + "bbox": [ + 313, + 327, + 555, + 533 + ], + "type": "inline_equation", + "content": "v_{i}^{\\prime} = v_{i} + \\Delta v_{i}" + }, + { + "bbox": [ + 313, + 327, + 555, + 533 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 327, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\Delta v_{i} = 0" + }, + { + "bbox": [ + 313, + 327, + 555, + 533 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 313, + 327, + 555, + 533 + ], + "type": "inline_equation", + "content": "i\\notin \\Gamma" + }, + { + "bbox": [ + 313, + 327, + 555, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 327, + 555, + 533 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 327, + 555, + 533 + ], + "type": "text", + "content": " is the number of items. In our case, " + }, + { + "bbox": [ + 313, + 327, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\pmb{v}\\in \\mathbb{R}^{N}" + }, + { + "bbox": [ + 313, + 327, + 555, + 533 + ], + "type": "text", + "content": " is the ground truth importance scores, " + }, + { + "bbox": [ + 313, + 327, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\pmb{y}^{*}\\in [0,1]^{N}" + }, + { + "bbox": [ + 313, + 327, + 555, + 533 + ], + "type": "text", + "content": " is the ground truth summary, and the perturbed profit " + }, + { + "bbox": [ + 313, + 327, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\pmb{v}^{\\prime}\\in \\mathbb{R}^{N}" + }, + { + "bbox": [ + 313, + 327, + 555, + 533 + ], + "type": "text", + "content": " corresponds to the imperfectly predicted importance scores. We further define two disjoint subsets of " + }, + { + "bbox": [ + 313, + 327, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 313, + 327, + 555, + 533 + ], + "type": "text", + "content": ". First, " + }, + { + "bbox": [ + 313, + 327, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\Gamma_0^+ \\equiv \\{i\\in \\Gamma \\mid y_i^* = 0,\\Delta p_i > 0\\}" + }, + { + "bbox": [ + 313, + 327, + 555, + 533 + ], + "type": "text", + "content": " refers to the set of items excluded in the original optimal solution but undergoing increased profit, and " + }, + { + "bbox": [ + 313, + 327, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\Gamma_1^- \\equiv \\{i\\in \\Gamma \\mid y_i^* = 1,\\Delta p_i < 0\\}" + }, + { + "bbox": [ + 313, + 327, + 555, + 533 + ], + "type": "text", + "content": " is the opposite, the set of items included in the optimal solution but experiencing a decreased profit. Then, Hifi et al. [23] claims that the solution " + }, + { + "bbox": [ + 313, + 327, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\pmb{y}^{*}" + }, + { + "bbox": [ + 313, + 327, + 555, + 533 + ], + "type": "text", + "content": " remains optimal for the perturbed KP, i.e., " + }, + { + "bbox": [ + 313, + 327, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\pmb{y}^{*} = \\mathcal{KP}(\\pmb{v},\\pmb{w},\\rho) = \\mathcal{KP}(\\pmb{v}^{\\prime},\\pmb{w},\\rho)" + }, + { + "bbox": [ + 313, + 327, + 555, + 533 + ], + "type": "text", + "content": ", if" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 540, + 553, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 540, + 553, + 570 + ], + "spans": [ + { + "bbox": [ + 314, + 540, + 553, + 570 + ], + "type": "interline_equation", + "content": "\\sum_ {i \\in \\Gamma_ {0} ^ {+}} \\Delta v _ {i} - \\sum_ {i \\in \\Gamma_ {1} ^ {-}} \\Delta v _ {i} \\leq \\boldsymbol {v} \\cdot \\mathcal {K P} (\\boldsymbol {v}, \\boldsymbol {w}, \\rho) - \\max (\\Psi^ {+}, \\Psi^ {-}),", + "image_path": "8170db4681d6a86875a244a4234e773e255ecc513efb676a07bcad8d62d2f6c7.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 578, + 555, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 555, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 555, + 642 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 578, + 555, + 642 + ], + "type": "inline_equation", + "content": "\\Psi^{+} \\equiv \\max_{i \\in \\Gamma_{0}^{+}} [\\pmb{v} \\cdot \\mathcal{KP}(\\pmb{v} - v_{i}\\hat{e}_{i}, \\pmb{w}, \\rho - w_{i}/N)]" + }, + { + "bbox": [ + 313, + 578, + 555, + 642 + ], + "type": "text", + "content": " is the maximum value of the KP when removing any item from " + }, + { + "bbox": [ + 313, + 578, + 555, + 642 + ], + "type": "inline_equation", + "content": "\\Gamma_{0}^{+}" + }, + { + "bbox": [ + 313, + 578, + 555, + 642 + ], + "type": "text", + "content": " and adjusting the constraint, and similarly, " + }, + { + "bbox": [ + 313, + 578, + 555, + 642 + ], + "type": "inline_equation", + "content": "\\Psi^{-} \\equiv \\max_{i \\in \\Gamma_{1}^{-}} [\\pmb{v} \\cdot \\mathcal{KP}(\\pmb{v} - v_{i}\\hat{e}_{i}, \\pmb{w}, \\rho)]" + }, + { + "bbox": [ + 313, + 578, + 555, + 642 + ], + "type": "text", + "content": " is the maximum value of the KP when removing any item from " + }, + { + "bbox": [ + 313, + 578, + 555, + 642 + ], + "type": "inline_equation", + "content": "\\Gamma_{1}^{-}" + }, + { + "bbox": [ + 313, + 578, + 555, + 642 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 643, + 553, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 643, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 313, + 643, + 553, + 678 + ], + "type": "text", + "content": "Based on this theorem we propose a new metric, Confidence of Importance Score (CIS), to estimate how close the predicted scores are to the ground truth:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 686, + 566, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 686, + 566, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 686, + 566, + 715 + ], + "type": "interline_equation", + "content": "\\operatorname {CIS} = \\sum_ {i \\in \\Gamma_ {0} ^ {+}} \\Delta v _ {i} - \\sum_ {i \\in \\Gamma_ {1} ^ {-}} \\Delta v _ {i} - \\boldsymbol {v} \\cdot \\mathcal {K P} (\\boldsymbol {v}, \\boldsymbol {w}, \\rho) + \\max \\left(\\Psi^ {+}, \\Psi^ {-}\\right).", + "image_path": "cf918ec843da7462f0aa34985bd856f4590c2a1e3e26b912804fb4c271ee2e7f.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15101" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "text", + "content": "If CIS " + }, + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "\\leq 0" + }, + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "text", + "content": ", the theorem indicates that the predicted importance score " + }, + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{v}}" + }, + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "text", + "content": " is guaranteed to induce the true KP solution " + }, + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "\\pmb{y}^{*}" + }, + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "text", + "content": ". If CIS " + }, + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": ">0" + }, + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{v}}" + }, + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "text", + "content": " does not guarantee to induce the KP solution, but a lower CIS would indicate a solution closer to " + }, + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "\\pmb{y}^{*}" + }, + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "text", + "content": " than a higher one. That is, a smaller CIS indicates a stronger confidence in satisfying the inequality, indicating a higher chance for the solution of the perturbed KP to be identical to that of the original KP." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 167, + 294, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 167, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 294, + 239 + ], + "type": "text", + "content": "Unlike F1 that simply measures how close the predicted summary is to the ground truth, our CIS accounts for the confidence that the predicted score will exactly induce the true summary. This is particularly important in video summarization, since the importance score itself is subjective and often noisy, requiring robust selection of frames." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 242, + 294, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 242, + 294, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 294, + 291 + ], + "type": "text", + "content": "Weighted Inclusion Ratios. Analyzing the sensitivity of the KP optimum to the perturbation of the importance score, Belgacem and Hifi [5] established bounds of the perturbed profits to retain the original optimum. These intervals are" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 296, + 308, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 296, + 308, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 308, + 323 + ], + "type": "interline_equation", + "content": "\\boldsymbol {y} _ {i} ^ {*} = 1 \\rightarrow \\Delta v _ {i} \\in I _ {i} ^ {1} = \\left[ \\right. \\max \\left(\\Delta_ {i} ^ {-}, w _ {i} \\max _ {k \\in \\Gamma} \\mu_ {k} - p _ {i}\\right), + \\infty\\left. \\right)", + "image_path": "a5327aae69071ade3a173f24577a40d69913bdb7b81148ae6171202ea2edea26.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 323, + 308, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 308, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 308, + 352 + ], + "type": "interline_equation", + "content": "\\boldsymbol {y} _ {i} ^ {*} = 0 \\rightarrow \\Delta v _ {i} \\in I _ {i} ^ {0} = \\left( \\right.- \\infty , \\min \\left(\\Delta_ {i} ^ {+}, p _ {i} - w _ {i} \\min _ {k \\in \\Gamma} \\mu_ {k}\\right)\\left. \\right],", + "image_path": "c357d15be2111416baeb3993c114aefcadc268d9bcc0a3a8d6dfcbe7cd0ca186.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 357, + 294, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 294, + 470 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 357, + 294, + 470 + ], + "type": "inline_equation", + "content": "\\Delta_i^+ = \\pmb{v} \\cdot (\\mathcal{KP}(\\pmb{v}, \\pmb{w}, \\rho) - \\mathcal{KP}(\\pmb{v} - v_i \\hat{e}_i, \\pmb{w}, \\rho - w_i / N), \\Delta_i^- = \\pmb{v} \\cdot (\\mathcal{KP}(\\pmb{v} - v_i \\hat{e}_i, \\pmb{w}, \\rho) - \\mathcal{KP}(\\pmb{v}, \\pmb{w}, \\rho))" + }, + { + "bbox": [ + 55, + 357, + 294, + 470 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 357, + 294, + 470 + ], + "type": "inline_equation", + "content": "\\mu_k" + }, + { + "bbox": [ + 55, + 357, + 294, + 470 + ], + "type": "text", + "content": " denotes the critical ratio of " + }, + { + "bbox": [ + 55, + 357, + 294, + 470 + ], + "type": "inline_equation", + "content": "k \\in \\Gamma" + }, + { + "bbox": [ + 55, + 357, + 294, + 470 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 55, + 357, + 294, + 470 + ], + "type": "inline_equation", + "content": "\\mu_k = \\mathcal{CR}([N] \\setminus \\{k\\}, \\rho - 1_{\\{y_k^* = 0\\}} \\cdot w_k / N)" + }, + { + "bbox": [ + 55, + 357, + 294, + 470 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 55, + 357, + 294, + 470 + ], + "type": "inline_equation", + "content": "\\mathcal{CR}(A, \\gamma) = v_s / w_s" + }, + { + "bbox": [ + 55, + 357, + 294, + 470 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 357, + 294, + 470 + ], + "type": "inline_equation", + "content": "s = \\min\\{l | \\sum_{i=1}^{l} v_s / w_s > \\rho\\}" + }, + { + "bbox": [ + 55, + 357, + 294, + 470 + ], + "type": "text", + "content": " is the index of the critical item. The items in " + }, + { + "bbox": [ + 55, + 357, + 294, + 470 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 357, + 294, + 470 + ], + "type": "text", + "content": " are considered in descending order of " + }, + { + "bbox": [ + 55, + 357, + 294, + 470 + ], + "type": "inline_equation", + "content": "v_s / w_s" + }, + { + "bbox": [ + 55, + 357, + 294, + 470 + ], + "type": "text", + "content": ". To sum up, if the perturbed profit resides in the interval " + }, + { + "bbox": [ + 55, + 357, + 294, + 470 + ], + "type": "inline_equation", + "content": "v_i' \\in I_i^{y_i^*}" + }, + { + "bbox": [ + 55, + 357, + 294, + 470 + ], + "type": "text", + "content": ", it is guaranteed that the solution of perturbed KP is identical with that of the original KP." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 471, + 294, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 294, + 584 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 294, + 584 + ], + "type": "text", + "content": "We propose Inclusion Ratio (IR) to estimate how much the predicted scores fall into the safe bounds; for the " + }, + { + "bbox": [ + 55, + 471, + 294, + 584 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 471, + 294, + 584 + ], + "type": "text", + "content": "-th importance score, " + }, + { + "bbox": [ + 55, + 471, + 294, + 584 + ], + "type": "inline_equation", + "content": "\\mathrm{IR}_i \\equiv |\\{\\Delta v_i \\in I_i^{y_i^*} | i \\in [N]\\}| / N" + }, + { + "bbox": [ + 55, + 471, + 294, + 584 + ], + "type": "text", + "content": ". This ratio reflects the proportion of the predicted scores that retain the KP solution. We propose the Weighted average of the Inclusion Ratios (WIR) as a metric to measure the proximity of the predicted score to the ground truth: " + }, + { + "bbox": [ + 55, + 471, + 294, + 584 + ], + "type": "inline_equation", + "content": "\\mathrm{WIR} \\equiv \\sum_{i=1}^{n} \\left( \\frac{w_i}{\\sum_{j=1}^{n} w_j} \\mathrm{IR}_i \\right)" + }, + { + "bbox": [ + 55, + 471, + 294, + 584 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 471, + 294, + 584 + ], + "type": "inline_equation", + "content": "w_i" + }, + { + "bbox": [ + 55, + 471, + 294, + 584 + ], + "type": "text", + "content": " denotes the segment length of each shot." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 588, + 294, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 294, + 624 + ], + "type": "text", + "content": "Weighted Sum of Errors (WSE). Additionally, we compare a weighted sum of the errors " + }, + { + "bbox": [ + 55, + 588, + 294, + 624 + ], + "type": "inline_equation", + "content": "\\sum_{i} w_{i} \\Delta p_{i}" + }, + { + "bbox": [ + 55, + 588, + 294, + 624 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 588, + 294, + 624 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 55, + 588, + 294, + 624 + ], + "type": "text", + "content": " is the segment length of each clip." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 627, + 294, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 627, + 294, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 627, + 294, + 686 + ], + "type": "text", + "content": "Comparison. Tab. 4 compares several competitive methods using these new metrics. Our SummDiff achieves the best performance in all metrics, indicating that the summaries generated by our method are closer to the optimal knapsack result, under the theoretical analysis in literature." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "type": "text", + "content": "Discussion. Our newly proposed metrics, WIR and CIS, evaluate the quality of the generated importance scores" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 318, + 67, + 548, + 156 + ], + "blocks": [ + { + "bbox": [ + 318, + 67, + 548, + 156 + ], + "lines": [ + { + "bbox": [ + 318, + 67, + 548, + 156 + ], + "spans": [ + { + "bbox": [ + 318, + 67, + 548, + 156 + ], + "type": "table", + "html": "
Methodτ↑ρ↑CIS ↓WIR ↑WSE ↓
Uniform-Random0.0000.0009.270.4546.82
SL-module [82]0.0600.0886.830.5630.09
CSTA [67]0.1280.1856.230.5725.76
PGL-SUM [3]0.0970.1416.140.5826.22
VASNet [17]0.0690.1026.250.5926.79
A2Summ [22]0.1210.1726.650.5530.55
SummDiff0.1750.2385.960.6125.24
", + "image_path": "dce8e748cc878bed82f909da2f2278731459d5b937dcfc57ffc44b37773dc4af.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 332, + 173, + 535, + 271 + ], + "blocks": [ + { + "bbox": [ + 318, + 159, + 548, + 170 + ], + "lines": [ + { + "bbox": [ + 318, + 159, + 548, + 170 + ], + "spans": [ + { + "bbox": [ + 318, + 159, + 548, + 170 + ], + "type": "text", + "content": "Table 4. Evaluation using our new metrics proposed in Sec. 4.4." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 173, + 535, + 271 + ], + "lines": [ + { + "bbox": [ + 332, + 173, + 535, + 271 + ], + "spans": [ + { + "bbox": [ + 332, + 173, + 535, + 271 + ], + "type": "table", + "html": "
Moduleτ↑ρ↑
Encoder Only0.0710.104
(+) Video Importance Denoiser0.0790.116
(+) Learnable Embedding0.0860.124
(+) AdaLN (τ)0.1450.204
(+) AdaLN (τ, Φ) 1 layer0.1710.232
(-) Quantization0.1250.174
(+) Classifier-free Guidance (CFG) [24]0.1750.238
(+) Self-attention Guidance (SAG) [30]0.1770.239
", + "image_path": "8b7bb23fb86b45a2ea0becd10db4eeec9655c1064737e1f504c00934ecbda8fa.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 359, + 274, + 509, + 285 + ], + "lines": [ + { + "bbox": [ + 359, + 274, + 509, + 285 + ], + "spans": [ + { + "bbox": [ + 359, + 274, + 509, + 285 + ], + "type": "text", + "content": "Table 5. Effect of individual components" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 293, + 553, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 553, + 376 + ], + "type": "text", + "content": "in two aspects. WIR measures how many importance scores are individually trustworthy, weighted by the duration. Specifically, WIR first computes the safe interval " + }, + { + "bbox": [ + 313, + 293, + 553, + 376 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 313, + 293, + 553, + 376 + ], + "type": "text", + "content": " for each importance score such that the true summary remains unchanged. Then, given the predicted importance scores " + }, + { + "bbox": [ + 313, + 293, + 553, + 376 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{v}}" + }, + { + "bbox": [ + 313, + 293, + 553, + 376 + ], + "type": "text", + "content": ", it measures how many of them fall within their corresponding safe intervals " + }, + { + "bbox": [ + 313, + 293, + 553, + 376 + ], + "type": "inline_equation", + "content": "\\hat{v}_{i} \\in I_{i}" + }, + { + "bbox": [ + 313, + 293, + 553, + 376 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 377, + 554, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 554, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 554, + 496 + ], + "type": "text", + "content": "On the other hand, CIS quantifies the risk that the predicted importance scores will lead to a different summary than the one generated using the true importance scores. In other words, CIS evaluates the overall chance how likely the predicted score vector " + }, + { + "bbox": [ + 313, + 377, + 554, + 496 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{v}}" + }, + { + "bbox": [ + 313, + 377, + 554, + 496 + ], + "type": "text", + "content": ", as a whole, is to give a summary different from the true one. With the analysis of knapsack optimality, these metrics would better measure the accuracy of predicted importance score, considering the fidelity of final summary after knapsack to the ground truth, than F1 or ranking-based metrics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 504, + 408, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 408, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 408, + 517 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": "We report ablation study on Mr. HiSum [74] in Tab. 5, comparing performance adding a component step-by-step. Starting from a two-layer transformer encoder predicting the importance score, adding our denoiser with a naively summed query " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_t = \\mathcal{C}(\\mathbf{u}_t) + \\pmb{\\tau} + \\pmb{\\Phi}" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " improves the Kendall's " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "(0.071\\rightarrow 0.079)" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": ". We observe further slight improvement " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "(0.079\\to 0.086)" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " when we replace the fixed random score embeddings with learnable ones (Learnable Embedding). Introducing AdaLN transformation [58] on the time " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "(\\tau)" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " and positional " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "(\\Phi)" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " encodings further improves " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " to 0.145 and 0.171, respectively. As we adopt the quantization idea, grouping nearby scores into a unified embedding, in Sec. 3.2 with this change, we experiment without it (using a sinusoidal function [80] instead), and observe significant performance drop " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "(0.171\\rightarrow 0.125)" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": ". This confirms that quantization is superior to treating the score" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15102" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 107, + 70, + 242, + 131 + ], + "blocks": [ + { + "bbox": [ + 107, + 70, + 242, + 131 + ], + "lines": [ + { + "bbox": [ + 107, + 70, + 242, + 131 + ], + "spans": [ + { + "bbox": [ + 107, + 70, + 242, + 131 + ], + "type": "table", + "html": "
Modelτ↑ρ↑
CSTA [67]0.1280.185
SummDiff (1 step)0.1700.234
SummDiff (10 steps)0.1750.238
SummDiff (100 steps)0.1820.245
", + "image_path": "3818b010d77493871d9c18871f7317518d166e3cc18edb2154e74fe5aaa650a2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 133, + 284, + 145 + ], + "lines": [ + { + "bbox": [ + 67, + 133, + 284, + 145 + ], + "spans": [ + { + "bbox": [ + 67, + 133, + 284, + 145 + ], + "type": "text", + "content": "Table 6. Performance with various numbers of DDIM steps" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 78, + 148, + 272, + 259 + ], + "blocks": [ + { + "bbox": [ + 78, + 148, + 272, + 259 + ], + "lines": [ + { + "bbox": [ + 78, + 148, + 272, + 259 + ], + "spans": [ + { + "bbox": [ + 78, + 148, + 272, + 259 + ], + "type": "image", + "image_path": "d05a072a6fd027b8eff8f686d74093d6cca3b5b0e967d64ba25169f53789094a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 262, + 295, + 295 + ], + "lines": [ + { + "bbox": [ + 55, + 262, + 295, + 295 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 295, + 295 + ], + "type": "text", + "content": "Figure 6. Measured " + }, + { + "bbox": [ + 55, + 262, + 295, + 295 + ], + "type": "inline_equation", + "content": "\\mathbb{E}|\\{\\pmb{y}^*\\}|" + }, + { + "bbox": [ + 55, + 262, + 295, + 295 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 262, + 295, + 295 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[\\Delta \\pmb{y}^*]" + }, + { + "bbox": [ + 55, + 262, + 295, + 295 + ], + "type": "text", + "content": " with respect to the quantization strength " + }, + { + "bbox": [ + 55, + 262, + 295, + 295 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 262, + 295, + 295 + ], + "type": "text", + "content": ". A larger " + }, + { + "bbox": [ + 55, + 262, + 295, + 295 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 262, + 295, + 295 + ], + "type": "text", + "content": " is associated with a smaller number of allowed solutions for the given KP." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 300, + 295, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 295, + 348 + ], + "type": "text", + "content": "as a continuous scalar. On top of this, we combine our model with two widely-known ideas, classifier-free guidance (CFG) [24] and self-Attention guidance (SAG) [30], for further improvement. See Appendix B for details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 350, + 296, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 296, + 399 + ], + "type": "text", + "content": "Number of DDIM Steps. Tab. 6 suggests that SummDiff outperforms CSTA [67], the best-performing baseline, even with a single DDIM [68] step. Beyond this, SummDiff further improves its performance with more DDIM steps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 401, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 295, + 460 + ], + "type": "text", + "content": "Effect of Quantization Strength. Solving a KP usually has a unique solution if the profits " + }, + { + "bbox": [ + 55, + 401, + 295, + 460 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 55, + 401, + 295, + 460 + ], + "type": "text", + "content": " are continuous. As quantization unifies importance score values within the same interval to a single embedding, it may lead to multiple solutions to " + }, + { + "bbox": [ + 55, + 401, + 295, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{KP}(\\tilde{\\boldsymbol{v}},\\boldsymbol {w},\\rho)" + }, + { + "bbox": [ + 55, + 401, + 295, + 460 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 401, + 295, + 460 + ], + "type": "inline_equation", + "content": "\\tilde{\\boldsymbol{v}}" + }, + { + "bbox": [ + 55, + 401, + 295, + 460 + ], + "type": "text", + "content": " is a quantized version of " + }, + { + "bbox": [ + 55, + 401, + 295, + 460 + ], + "type": "inline_equation", + "content": "\\boldsymbol{v}" + }, + { + "bbox": [ + 55, + 401, + 295, + 460 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 461, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 295, + 581 + ], + "type": "text", + "content": "To analyze this effect, we measure the average number of solutions " + }, + { + "bbox": [ + 55, + 461, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{\\pmb{v}_1\\sim U(0,1)\\otimes N}[\\{\\pmb{y}^*\\in \\{0,1\\} ^N |\\pmb{y}^* = \\mathcal{KP}(\\tilde{\\pmb{v}},\\pmb {w},\\rho)\\} ]" + }, + { + "bbox": [ + 55, + 461, + 295, + 581 + ], + "type": "text", + "content": " while varying " + }, + { + "bbox": [ + 55, + 461, + 295, + 581 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 461, + 295, + 581 + ], + "type": "text", + "content": ", taking expectation over " + }, + { + "bbox": [ + 55, + 461, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\pmb{v}" + }, + { + "bbox": [ + 55, + 461, + 295, + 581 + ], + "type": "text", + "content": ", which is a value vector sampled from a joint uniform distribution. We also consider the average " + }, + { + "bbox": [ + 55, + 461, + 295, + 581 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 55, + 461, + 295, + 581 + ], + "type": "text", + "content": " difference between two summaries, " + }, + { + "bbox": [ + 55, + 461, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{\\pmb{v}_1\\sim U(0,1)\\otimes N},\\pmb{v}_2\\sim \\mathcal{N}(\\pmb{v}_1,\\frac{1}{K}\\mathbf{I}_N)[\\| \\mathcal{KP}(\\tilde{\\pmb{v}}_1,\\pmb {w},\\rho) - \\mathcal{KP}(\\tilde{\\pmb{v}}_2,\\pmb {w},\\rho)\\| _1]" + }, + { + "bbox": [ + 55, + 461, + 295, + 581 + ], + "type": "text", + "content": ", obtained by solving the KP with two similar but distinct value vectors, i.e., " + }, + { + "bbox": [ + 55, + 461, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\pmb{v}_2 = \\pmb{v}_1 + \\Delta \\pmb{v}" + }, + { + "bbox": [ + 55, + 461, + 295, + 581 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 461, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\Delta \\pmb{v}\\sim \\mathcal{N}(0,\\frac{1}{K}\\mathbf{I}_N)" + }, + { + "bbox": [ + 55, + 461, + 295, + 581 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 653 + ], + "type": "text", + "content": "We numerically analyze the relation between these two values, which indicate the scale of multiplicity of the solutions to the KP, and the quantization strength " + }, + { + "bbox": [ + 55, + 582, + 295, + 653 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 582, + 295, + 653 + ], + "type": "text", + "content": ". As shown in Fig. 6, both the number of solutions " + }, + { + "bbox": [ + 55, + 582, + 295, + 653 + ], + "type": "inline_equation", + "content": "|\\{\\pmb{y}^*\\}|" + }, + { + "bbox": [ + 55, + 582, + 295, + 653 + ], + "type": "text", + "content": " and the summary deviation " + }, + { + "bbox": [ + 55, + 582, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\Delta \\pmb{y}^*" + }, + { + "bbox": [ + 55, + 582, + 295, + 653 + ], + "type": "text", + "content": " decrease with larger " + }, + { + "bbox": [ + 55, + 582, + 295, + 653 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 582, + 295, + 653 + ], + "type": "text", + "content": ", suggesting more accurate summaries." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": "To sum up, increasing the quantization strength " + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": " reduces the number of multiple optimal solutions, leading to a more accurate summary. However, an excessively higher " + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": " leads to insufficient samples per bin, degrading the overall performance. See Appendix A for more ablation studies." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 71, + 402, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 402, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 402, + 84 + ], + "type": "text", + "content": "5. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 91, + 555, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 281 + ], + "type": "text", + "content": "Video Summarization. Early models primarily rely on heuristic unsupervised learning [13, 15, 29, 38, 47, 49-51, 53, 71, 83] to select important or diverse frames, struggling with generalization. Benefiting from annotated datasets, supervised methods [8, 38, 57, 59, 60] have emerged. DSNet [93] and iPTNet [34] improve keyframe selection using frame-level annotations, and SL-Module [82] captures high-level features. With the advent of deep learning, RNNs [84, 88-90] and attention mechanism [4, 6, 33, 35, 39, 42, 77, 85, 94] have been adopted to capture temporal dependencies in videos. Particularly, VASNet [17] and PGL-SUM [3] capture both local and global frame dependencies using self-attention. DMASum [81] introduces mixture of attention layer mitigate the key Softmax Bottleneck. SUM-GAN [50] and AC-SUM-GAN [2] leverage generative adversarial networks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 283, + 556, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 556, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 556, + 403 + ], + "type": "text", + "content": "Multimodal approaches [32, 91] integrate audio or textual data to video summarization. Multimodal transformers are adopted to link frames with corresponding captions, improving context-aware summaries, by CLIP-It [52], MSVA [18], SSPVS [40], and A2Summ [22]. CSTA [67] addresses computational complexity by using CNN-based spatiotemporal attention for efficient frame selection. Unlike these deterministic summarizers, our approach generates multiple plausible summaries by capturing the distribution of video summaries with various perspectives." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 406, + 556, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 556, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 556, + 515 + ], + "type": "text", + "content": "Diffusion for Video Tasks. Diffusion [66] have emerged as a groundbreaking tool in high-quality image generation [21, 25, 54, 63, 64, 69, 70] and super-resolution [14, 36, 61, 65]. Recently, they are extended to video generation [7, 11, 16, 19, 26, 27, 44], joint video and audio generation [41, 62, 73], video editing [9, 10, 79], video inpainting [87] and prediction [31, 86]. Furthermore, they are applied to video understanding tasks like moment retrieval [43, 48], video object segmentation [92, 95], action segmentation [45]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 525, + 382, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 382, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 382, + 539 + ], + "type": "text", + "content": "6. Summary" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": "Video summarization is inherently subjective since people have different perspectives of a good summary. We suggest a generative viewpoint of this task where the model learns the distribution of good summaries, in contrast to the traditional approach of aggregated importance score regression. Our proposed SummDiff model, adopting diffusion models for video summarization for the first time, is able to generate multiple good summaries conditioned on the input video. Our model not only outperforms other baselines but also demonstrates the ability to generate accurate summaries customized to the individual annotators. We further propose additional metrics that measure the quality of the predicted importance scores through an insight from the actual summary generation using knapsack." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15103" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "content": "This work was supported by Samsung Electronics (IO240512-09881-01), Youlchon Foundation, NRF grants (RS-2021-NR05515, RS-2024-00336576, RS-2023-0022663) and IITP grants (RS-2022-II220264, RS-2024-00353131) by the government of Korea." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 115, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 115, + 175 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 115, + 175 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 183, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 183, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 183, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 61, + 183, + 296, + 228 + ], + "type": "text", + "content": "[1] Sami Abu-El-Haija, Nisarg Kothari, Joonseok Lee, Paul Natsev, George Toderici, Balakrishnan Varadarajan, and Sudheendra Vijayanarasimhan. Youtube-8M: A large-scale video classification benchmark. arXiv:1609.08675, 2016. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 228, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 228, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 61, + 228, + 296, + 293 + ], + "type": "text", + "content": "[2] Evlampios Apostolidis, Eleni Adamantidou, Alexandros I Metsai, Vasileios Mezaris, and Ioannis Patras. AC-SUMGAN: Connecting actor-critic and generative adversarial networks for unsupervised video summarization. IEEE Transactions on Circuits and Systems for Video Technology, 31(8): 3278-3292, 2020. 1, 4, 8, iii" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 295, + 296, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 295, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 61, + 295, + 296, + 350 + ], + "type": "text", + "content": "[3] Evlampios Apostolidis, Georgios Balaouras, Vasileios Mezaris, and Ioannis Patras. Combining global and local attention with positional encoding for video summarization. In IEEE international symposium on multimedia (ISM), 2021. 1, 2, 4, 7, 8, i, iii" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 351, + 295, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 351, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 351, + 295, + 396 + ], + "type": "text", + "content": "[4] Evlampios Apostolidis, Georgios Balaouras, Vasileios Mezaris, and Ioannis Patras. Summarizing videos using concentrated attention and considering the uniqueness and diversity of the video frames. In ICMR, 2022. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 396, + 296, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 396, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 396, + 296, + 441 + ], + "type": "text", + "content": "[5] Tarik Belgacem and Mhand Hifi. Sensitivity analysis of the knapsack problem: Tighter lower and upper bound limits. Journal of systems science and systems engineering, 17:156-170, 2008. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 441, + 296, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 441, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 62, + 441, + 296, + 485 + ], + "type": "text", + "content": "[6] Manjot Bilkhu, Siyang Wang, and Tushar Dobhal. Attention is all you need for videos: Self-attention based video summarization using universal transformers. arXiv:1906.02792, 2019.8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 487, + 296, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 487, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 487, + 296, + 531 + ], + "type": "text", + "content": "[7] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 533, + 296, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 296, + 566 + ], + "type": "text", + "content": "[8] Sijia Cai, Wangmeng Zuo, Larry S Davis, and Lei Zhang. Weakly-supervised video summarization using variational encoder-decoder and web prior. In ECCV, 2018. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 567, + 296, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 296, + 599 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 296, + 599 + ], + "type": "text", + "content": "[9] Duygu Ceylan, Chun-Hao P Huang, and Niloy J Mitra. Pix2video: Video editing using image diffusion. In ICCV, 2023. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 600, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 600, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 600, + 295, + 634 + ], + "type": "text", + "content": "[10] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. StableVideo: Text-driven consistency-aware diffusion video editing. In ICCV, 2023. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 635, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 678 + ], + "type": "text", + "content": "[11] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. VideoCrafter2: Overcoming data limitations for high-quality video diffusion models. In CVPR, 2024. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 679, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 679, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 679, + 295, + 713 + ], + "type": "text", + "content": "[12] Ting Chen, Ruixiang Zhang, and Geoffrey Hinton. Analog bits: Generating discrete data using diffusion models with self-conditioning. In ICLR, 2022. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 106 + ], + "type": "text", + "content": "[13] Wen-Sheng Chu, Yale Song, and Alejandro Jaimes. Video co-summarization: Video summarization by visual co-occurrence. In CVPR, 2015. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 107, + 554, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 554, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 554, + 150 + ], + "type": "text", + "content": "[14] Hyungjin Chung, Byeongsu Sim, and Jong Chul Ye. Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In CVPR, 2022. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 152, + 554, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 554, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 554, + 206 + ], + "type": "text", + "content": "[15] Sandra Eliza Fontes De Avila, Ana Paula Brandao Lopes, Antonio da Luz Jr, and Arnaldo de Albuquerque Araujo. Vsumm: A mechanism designed to produce static video summaries and a novel evaluation method. Pattern recognition letters, 32(1):56-68, 2011. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 207, + 554, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 554, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 554, + 251 + ], + "type": "text", + "content": "[16] Patrick Esser, Johnathan Chiu, Parmida Atighechian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In ICCV, 2023. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 252, + 554, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 554, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 554, + 295 + ], + "type": "text", + "content": "[17] Jiri Fajtl, Hajar Sadeghi Sokeh, Vasileios Argyriou, Dorothy Monekosso, and Paolo Remagnino. Summarizing videos with attention. In ACCV Workshop, 2019. 1, 2, 4, 7, 8, i, iii" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 297, + 554, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 554, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 554, + 330 + ], + "type": "text", + "content": "[18] Junaid Ahmed Ghauri, Sherzod Hakimov, and Ralph Ewerth. Supervised video summarization via multiple feature sets with parallel attention. In ICME, 2021. 8, ii" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 331, + 554, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 554, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 554, + 374 + ], + "type": "text", + "content": "[19] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Fei-Fei Li, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. In ECCV, 2025. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 376, + 554, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 554, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 554, + 408 + ], + "type": "text", + "content": "[20] Michael Gygli, Helmut Grabner, Hayko Riemenschneider, and Luc Van Gool. Creating summaries from user videos. In ECCV, 2014. 1, 4, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 410, + 554, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 554, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 554, + 442 + ], + "type": "text", + "content": "[21] Jaehoon Hahm, Junho Lee, Sunghyun Kim, and Joonseok Lee. Isometric representation learning for disentangled latent space of diffusion models. In ICML, 2024. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 444, + 554, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 444, + 554, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 444, + 554, + 487 + ], + "type": "text", + "content": "[22] Bo He, Jun Wang, Jielin Qiu, Trung Bui, Abhinav Shrivastava, and Zhaowen Wang. Align and attend: Multimodal summarization with dual contrastive losses. In CVPR, 2023. 1, 4, 7, 8, i, iii" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 488, + 554, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 488, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 488, + 554, + 533 + ], + "type": "text", + "content": "[23] Mhand Hifi, Hedi Mhalla, and Slim Sadfi. Sensitivity of the optimum to perturbations of the profit or weight of an item in the binary knapsack problem. Journal of Combinatorial Optimization, 10:239-260, 2005. 4, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 533, + 554, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 554, + 567 + ], + "type": "text", + "content": "[24] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshop on Deep Generative Models and Downstream Applications, 2021. 1, 7, 8, i" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 567, + 554, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 554, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 554, + 589 + ], + "type": "text", + "content": "[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 3, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 590, + 554, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 554, + 644 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 554, + 644 + ], + "type": "text", + "content": "[26] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv:2210.02303, 2022. 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 646, + 554, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 554, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 554, + 678 + ], + "type": "text", + "content": "[27] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In NeurIPS, 2022. 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 679, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 679, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 679, + 554, + 713 + ], + "type": "text", + "content": "[28] Fa-Ting Hong, Xuanteng Huang, Wei-Hong Li, and Wei-Shi Zheng. Mini-net: Multiple instance ranking network for video highlight detection. In ECCV, 2020. 4" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15104" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "text", + "content": "[29] Richang Hong, Jinhui Tang, Hung-Khoon Tan, Shuicheng Yan, Chongwah Ngo, and Tat-Seng Chua. Event driven summarization for web videos. In SIGMM workshop on Social media, 2009. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 152 + ], + "type": "text", + "content": "[30] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seunglyong Kim. Improving sample quality of diffusion models using self-attention guidance. In ICCV, 2023. 7, 8, i" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 153, + 295, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 153, + 295, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 153, + 295, + 186 + ], + "type": "text", + "content": "[31] Tobias Hörppe, Arash Mehrjou, Stefan Bauer, Didrik Nielsen, and Andrea Dittadi. Diffusion models for video prediction and infilling. arXiv:2206.07696, 2022. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 188, + 295, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 188, + 295, + 231 + ], + "spans": [ + { + "bbox": [ + 56, + 188, + 295, + 231 + ], + "type": "text", + "content": "[32] Tzu-Chun Hsu, Yi-Sheng Liao, and Chun-Rong Huang. Video summarization with spatiotemporal vision transformer. IEEE Transactions on Image Processing, 2023. 8, ii" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 233, + 295, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 233, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 56, + 233, + 295, + 266 + ], + "type": "text", + "content": "[33] Jia-Hong Huang, Luka Murn, Marta Mrak, and Marcel Worring. GPT2MVS: Generative pre-trained transformer-2 for multi-modal video summarization. In ICMR, 2021. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 268, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 268, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 56, + 268, + 295, + 300 + ], + "type": "text", + "content": "[34] Hao Jiang and Yadong Mu. Joint video summarization and moment localization by cross-task sample transfer. In CVPR, 2022. 1, 4, 8, iii" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 302, + 295, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 302, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 56, + 302, + 295, + 335 + ], + "type": "text", + "content": "[35] Yunjae Jung, Donghyeon Cho, Sanghyun Woo, and In So Kweon. Global-and-local relative position embedding for unsupervised video summarization. In ECCV, 2020. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 337, + 295, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 337, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 337, + 295, + 369 + ], + "type": "text", + "content": "[36] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. In NeurIPS, 2022. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 371, + 295, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 295, + 393 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 295, + 393 + ], + "type": "text", + "content": "[37] Maurice G Kendall. The treatment of ties in ranking problems. Biometrika, 33(3):239-251, 1945. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 395, + 295, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 395, + 295, + 427 + ], + "spans": [ + { + "bbox": [ + 56, + 395, + 295, + 427 + ], + "type": "text", + "content": "[38] Aditya Khosla, Raffay Hamid, Chih-Jen Lin, and Neel Sundaresan. Large-scale video summarization using web-image priors. In CVPR, 2013. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 430, + 295, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 430, + 295, + 473 + ], + "spans": [ + { + "bbox": [ + 56, + 430, + 295, + 473 + ], + "type": "text", + "content": "[39] Haopeng Li, Qiuhong Ke, Mingming Gong, and Rui Zhang. Video joint modelling based on hierarchical transformer for co-summulation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3904–3917, 2022. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 475, + 295, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 295, + 507 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 295, + 507 + ], + "type": "text", + "content": "[40] Haopeng Li, Qiuhong Ke, Mingming Gong, and Tom Drummond. Progressive video summarization via multimodal self-supervised learning. In WACV, 2023. 4, 8, ii" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 510, + 295, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 295, + 552 + ], + "type": "text", + "content": "[41] Judith Li, Xuchan Bao, Zhong Yi Wan, Kun Su, Timo Denk, Dima Kuzmin, Joonseok Lee, and Fei Sha. Diff4steer: Steerable diffusion prior for generative music retrieval with semantic guidance. In ICASSP, 2024. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 555, + 295, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 295, + 598 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 295, + 598 + ], + "type": "text", + "content": "[42] Ping Li, Qinghao Ye, Luming Zhang, Li Yuan, Xianghua Xu, and Ling Shao. Exploring global diverse attention via pairwise temporal relation for video summarization. Pattern Recognition, 111:107677, 2021. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 601, + 295, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 295, + 643 + ], + "type": "text", + "content": "[43] Pandeng Li, Chen-Wei Xie, Hongtao Xie, Liming Zhao, Lei Zhang, Yun Zheng, Deli Zhao, and Yongdong Zhang. MomentDiff: Generative video moment retrieval from random to real. In NeurIPS, 2024. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 646, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 646, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 646, + 295, + 678 + ], + "type": "text", + "content": "[44] Jingyun Liang, Yuchen Fan, Kai Zhang, Radu Timofte, Luc Van Gool, and Rakesh Ranjan. MoVideo: Motion-aware video generation with diffusion model. In ECCV, 2025. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 681, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 295, + 712 + ], + "type": "text", + "content": "[45] Daochang Liu, Qiyue Li, Anh-Dung Dinh, Tingting Jiang, Mubarak Shah, and Chang Xu. Diffusion action segmentation. In ICCV, 2023. 8" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 95 + ], + "type": "text", + "content": "[46] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2017. ii" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 96, + 553, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 118 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 118 + ], + "type": "text", + "content": "[47] Zheng Lu and Kristen Grauman. Story-driven summarization for egocentric video. In CVPR, 2013. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 119, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 162 + ], + "type": "text", + "content": "[48] Dezhao Luo, Shaogang Gong, Jiabo Huang, Hailin Jin, and Yang Liu. Generative video diffusion for unseen cross-domain video moment retrieval. arXiv:2401.13329, 2024. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 163, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 195 + ], + "type": "text", + "content": "[49] Yu-Fei Ma, Lie Lu, Hong-Jiang Zhang, and Mingjing Li. A user attention model for video summarization. In ACM MM, 2002. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 197, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 229 + ], + "type": "text", + "content": "[50] Behrooz Mahasseni, Michael Lam, and Sinisa Todorovic. Unsupervised video summarization with adversarial LSTM networks. In CVPR, 2017. 1, 4, 8, ii, iii" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 231, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 231, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 231, + 553, + 274 + ], + "type": "text", + "content": "[51] Padmavathi Mundur, Yong Rao, and Yelena Yesha. Keyframe-based video summarization using delaunay clustering. International Journal on Digital Libraries, 6:219-232, 2006. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 276, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 307 + ], + "type": "text", + "content": "[52] Medhini Narasimhan, Anna Rohrbach, and Trevor Darrell. Clip-it! language-guided video summarization. In NeurIPS, 2021. 4, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 309, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 341 + ], + "type": "text", + "content": "[53] Chong-Wah Ngo, Yu-Fei Ma, and Hong-Jiang Zhang. Automatic video summarization by graph modeling. In ICCV, 2003. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 343, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 343, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 317, + 343, + 553, + 396 + ], + "type": "text", + "content": "[54] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 1, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 399, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 430 + ], + "type": "text", + "content": "[55] Mayu Otani, Yuta Nakashima, Esa Rahtu, and Janne Heikkila. Rethinking the evaluation of video summaries. In CVPR, 2019. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 433, + 553, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 463 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 463 + ], + "type": "text", + "content": "[56] Rameswar Panda and Amit K Roy-Chowdhury. Collaborative summarization of topic-related videos. In CVPR, 2017. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 466, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 466, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 466, + 553, + 498 + ], + "type": "text", + "content": "[57] Rameswar Panda, Abir Das, Ziyan Wu, Jan Ernst, and Amit K Roy-Chowdhury. Weakly supervised summarization of web videos. In ICCV, 2017. 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 500, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 520 + ], + "type": "text", + "content": "[58] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 2, 3, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "type": "text", + "content": "[59] Danila Potapov, Matthijs Douze, Zaid Harchaoui, and Cordelia Schmid. Category-specific video summarization. In ECCV, 2014. 4, 8" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 557, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 578 + ], + "type": "text", + "content": "[60] Mrigank Rochan and Yang Wang. Video summarization by learning from unpaired data. In CVPR, 2019. 8" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 580, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 580, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 580, + 553, + 611 + ], + "type": "text", + "content": "[61] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 8" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 317, + 613, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 656 + ], + "type": "text", + "content": "[62] Ludan Ruan, Yiyang Ma, Huan Yang, Huiguo He, Bei Liu, Jianlong Fu, Nicholas Jing Yuan, Qin Jin, and Baining Guo. MM-diffusion: Learning multi-modal diffusion models for joint audio and video generation. In CVPR, 2023. 8" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "text", + "content": "[63] Suho Ryu, Kihyun Kim, Eugene Baek, Dongsoo Shin, and Joonseok Lee. Towards scalable human-aligned benchmark for text-guided image editing. In CVPR, 2025. 8" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[64] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour," + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15105" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "text", + "content": "[65] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE transactions on pattern analysis and machine intelligence, 45(4):4713-4726, 2022. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 295, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 295, + 196 + ], + "type": "text", + "content": "[66] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 295, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 295, + 231 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 295, + 231 + ], + "type": "text", + "content": "[67] Jaewon Son, Jaehun Park, and Kwangsu Kim. CSTA: CNN-based spatiotemporal attention for video summarization. In CVPR, 2024. 1, 2, 4, 5, 7, 8, i, ii, iii" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 233, + 295, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 233, + 295, + 255 + ], + "spans": [ + { + "bbox": [ + 56, + 233, + 295, + 255 + ], + "type": "text", + "content": "[68] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In *ICLR*, 2020. 3, 8, ii" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 256, + 295, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 256, + 295, + 288 + ], + "spans": [ + { + "bbox": [ + 56, + 256, + 295, + 288 + ], + "type": "text", + "content": "[69] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In NeurIPS, 2019. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 291, + 295, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 291, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 56, + 291, + 295, + 323 + ], + "type": "text", + "content": "[70] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. In NeurIPS, 2020. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 326, + 295, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 326, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 56, + 326, + 295, + 357 + ], + "type": "text", + "content": "[71] Yale Song, Jordi Vallmitjana, Amanda Stent, and Alejandro Jaimes. TVSum: Summarizing web videos using titles. In CVPR, 2015. 1, 4, 5, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 360, + 295, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 295, + 403 + ], + "type": "text", + "content": "[72] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2020. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 406, + 295, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 406, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 56, + 406, + 295, + 449 + ], + "type": "text", + "content": "[73] Kun Su, Judith Li, Qingqing Huang, Dima Kuzmin, Joon-seok Lee, Chris Donahue, Fei Sha, Aren Jansen, Yu Wang, Mauro Verzetti, and Timo Denk. V2Meow: Meowing to the visual beat via video-to-music generation. In AAAI, 2024. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 452, + 295, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 452, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 56, + 452, + 295, + 483 + ], + "type": "text", + "content": "[74] Jinhwan Sul, Jihoon Han, and Joonseok Lee. Mr. HiSum: A large-scale dataset for video highlight detection and summarization. In NeurlPS, 2024. 4, 5, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 486, + 295, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 486, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 486, + 295, + 529 + ], + "type": "text", + "content": "[75] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In CVPR, 2015. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 531, + 295, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 531, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 531, + 295, + 563 + ], + "type": "text", + "content": "[76] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 566, + 295, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 566, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 56, + 566, + 295, + 597 + ], + "type": "text", + "content": "[77] Hacene Terbouche, Maryan Morel, Mariano Rodriguez, and Alice Othmani. Multi-annotation attention model for video summarization. In CVPR, 2023. 4, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 600, + 295, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 600, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 56, + 600, + 295, + 643 + ], + "type": "text", + "content": "[78] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. ii" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 646, + 295, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 646, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 646, + 295, + 677 + ], + "type": "text", + "content": "[79] Shuyuan Tu, Qi Dai, Zhi-Qi Cheng, Han Hu, Xintong Han, Zuxuan Wu, and Yu-Gang Jiang. MotionEditor: Editing video motion via content-aware diffusion. In CVPR, 2024. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 680, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 295, + 712 + ], + "type": "text", + "content": "[80] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, 2017. 2, 3, 7" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 711 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "text", + "content": "[81] Junyan Wang, Yang Bai, Yang Long, Bingzhang Hu, Zhenhua Chai, Yu Guan, and Xiaolin Wei. Query twice: Dual mixture attention meta learning for video summarization. In ACM MM, 2020. 4, 5, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "text", + "content": "[82] Minghao Xu, Hang Wang, Bingbing Ni, Riheng Zhu, Zhenbang Sun, and Changhu Wang. Cross-category video highlight detection via set-based learning. In ICCV, 2021. 1, 4, 7, 8, i, iii" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 163, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 205 + ], + "type": "text", + "content": "[83] Hong Jiang Zhang, Jianhua Wu, Di Zhong, and Stephen W Smoliar. An integrated system for content-based video retrieval and browsing. Pattern recognition, 30(4):643-658, 1997. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 207, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 553, + 239 + ], + "type": "text", + "content": "[84] Ke Zhang, Wei-Lun Chao, Fei Sha, and Kristen Grauman. Video summarization with long short-term memory. In ECCV, 2016. 4, 8, ii" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "text", + "content": "[85] Yunzuo Zhang, Yameng Liu, Weili Kang, and Ran Tao. VSS-Net: visual semantic self-mining network for video summarization. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 4, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "type": "text", + "content": "[86] Zhicheng Zhang, Junyao Hu, Wentao Cheng, Danda Paudel, and Jufeng Yang. ExtDM: Distribution extrapolation diffusion model for video prediction. In CVPR, 2024. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 320, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 363 + ], + "type": "text", + "content": "[87] Zhixing Zhang, Bichen Wu, Xiaoyan Wang, Yaqiao Luo, Luxin Zhang, Yinan Zhao, Peter Vajda, Dimitris Metaxas, and Licheng Yu. AVID: Any-length video inpainting with diffusion model. In CVPR, 2024. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 365, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 397 + ], + "type": "text", + "content": "[88] Bin Zhao, Xuelong Li, and Xiaoqiang Lu. Hierarchical recurrent neural network for video summarization. In ACM MM, 2017. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 399, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 430 + ], + "type": "text", + "content": "[89] Bin Zhao, Xuelong Li, and Xiaogiang Lu. HSA-RNN: Hierarchical structure-adaptive rnn for video summarization. In CVPR, 2018." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 433, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 476 + ], + "type": "text", + "content": "[90] Bin Zhao, Xuelong Li, and Xiaogiang Lu. TTH-RNN: Tensor-train hierarchical recurrent neural network for video summarization. IEEE Transactions on Industrial Electronics, 68(4):3629-3637, 2020. 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "type": "text", + "content": "[91] Bin Zhao, Maoguo Gong, and Xuelong Li. Hierarchical multimodal transformer to summarize videos. Neurocomputing, 468:360-369, 2022. 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 512, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 544 + ], + "type": "text", + "content": "[92] Chuanwei Zhou, Chunyan Xu, Jun Li, Zhen Cui, and Jian Yang. Quality-aware pattern diffusion for video object segmentation. Neurocomputing, 528:148-159, 2023. 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 545, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 588 + ], + "type": "text", + "content": "[93] Wencheng Zhu, Jiwen Lu, Jiahao Li, and Jie Zhou. DSNet: A flexible detect-to-summarize network for video summarization. IEEE Transactions on Image Processing, 30:948–962, 2020. 4, 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "type": "text", + "content": "[94] Wencheng Zhu, Yucheng Han, Jiwen Lu, and Jie Zhou. Relational reasoning over spatial-temporal graphs for video summarization. IEEE Transactions on Image Processing, 31: 3017-3031, 2022. 8" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 635, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 677 + ], + "type": "text", + "content": "[95] Zixin Zhu, Xuelu Feng, Dongdong Chen, Junsong Yuan, Chunming Qiao, and Gang Hua. Exploring pre-trained text-to-video diffusion models for referring video object segmentation. arXiv:2403.12042, 2024. 8" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 680, + 553, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 711 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 711 + ], + "type": "text", + "content": "[96] Daniel Zwillinger and Stephen Kokoska. CRC standard probability and statistics tables and formulae. CRC Press, 1999. 4" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "15106" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Super Resolved Imaging with Adaptive Optics/330a7bf8-492d-4b73-b28f-8e96f1e5081a_content_list.json b/2025/Super Resolved Imaging with Adaptive Optics/330a7bf8-492d-4b73-b28f-8e96f1e5081a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f00489f75475adf8919578bca070ee89dbba5462 --- /dev/null +++ b/2025/Super Resolved Imaging with Adaptive Optics/330a7bf8-492d-4b73-b28f-8e96f1e5081a_content_list.json @@ -0,0 +1,1802 @@ +[ + { + "type": "text", + "text": "Super Resolved Imaging with Adaptive Optics", + "text_level": 1, + "bbox": [ + 264, + 130, + 733, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Robin Swanson $^{1,2}$ Esther Y. H. Lin $^{1}$ Masen Lamb $^{3,4}$ Suresh Sivanandam $^{1,2}$ Kiriakos N. Kutulakos $^{1}$", + "bbox": [ + 99, + 191, + 937, + 210 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of Toronto 2 Dunlap Institute for Astronomy & Astrophysics 3 International Gemini Observatory 4 University of Victoria", + "bbox": [ + 225, + 224, + 813, + 265 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fb9a3cfe2feafcf49849fb82be20353e7180bc7520a4af8f625c8cf9e0434eed.jpg", + "image_caption": [ + "Figure 1. Super Resolved Imaging with Adaptive Optics is a computational imaging method to enable multi-image super-resolution in scientific telescopes. Without any additional hardware, existing adaptive mirrors are used to inject diversity into each sub-exposure. When jointly reconstructed, we are able to recover higher spatial resolution from an otherwise undersampled signal. Each exposure's phase profiles is jointly optimized along with the image reconstruction algorithm—resulting in greatly improved performance." + ], + "image_footnote": [], + "bbox": [ + 91, + 276, + 903, + 500 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 598, + 326, + 614 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Astronomical telescopes suffer from a tradeoff between field of view (FoV) and image resolution: increasing the FoV leads to an optical field that is under-sampled by the science camera. This work presents a novel computational imaging approach to overcome this tradeoff by leveraging the existing adaptive optics (AO) systems in modern ground-based telescopes. Our key idea is to use the AO system's deformable mirror to apply a series of learned, precisely controlled distortions to the optical wavefront, producing a sequence of images that exhibit distinct, high-frequency, sub-pixel shifts. These images can then be jointly upsampled to yield the final super-resolved image. Crucially, we show this can be done while simultaneously maintaining the core AO operation—correcting for the unknown and rapidly changing wavefront distortions caused by Earth's atmosphere. To achieve this, we incorporate end-to-end optimization of both the induced mirror distortions and the", + "bbox": [ + 88, + 632, + 485, + 891 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "upsampling algorithm, such that telescope-specific optics and temporal statistics of atmospheric wavefront distortions are accounted for. Our experimental results with a hardware prototype, as well as simulations, demonstrate significant SNR improvements of up to 12 dB over non-AO super-resolution baselines, using only existing telescope optics and no hardware modifications. Moreover, by using a precise bench-top replica of a complete telescope and AO system, we show that our methodology can be readily transferred to an operational telescope.", + "bbox": [ + 511, + 599, + 908, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 789, + 645, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Super-resolution (SR) techniques have been integral to astronomical imaging due to inherent telescope design constraints [23]. Telescopes can have many science instruments", + "bbox": [ + 511, + 814, + 906, + 859 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding author: robin@cs.toronto.edu \nProject website: www.cs.toronto.edu/~robin/aosr", + "bbox": [ + 535, + 875, + 864, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "29142", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and in certain cases optical resolution is not their priority [67]. A large field of view (FoV), for example, necessitates undersampling to capture a wider view of the night sky. Another common case in photon-starved conditions is to increase signal-to-noise ratio by focusing the light onto fewer, larger, pixels, thereby incurring less read noise [43]. In these, and many other, cases although the telescope is capable of providing a diffraction-limited signal, the science instrument can be undersampled. Thus, the use of SR methods in astronomical imaging has a long history in both satellite and terrestrial telescopes [34].", + "bbox": [ + 89, + 90, + 480, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unlike telescopes stationed in orbit, ground telescopes face an additional challenge: the Earth's atmosphere [5]. Our atmosphere is a source of dynamic optical aberrations which introduce additional difficulties regardless of a telescope's location, size, or design [58]. As the atmosphere evolves, small changes in index of refraction occur across the sky. This causes incoming light to distort as it transitions from outer space and through our telescope, blurring the images. Therefore, traditional multi-image SR methods which combine multiple sharp images with small geometric distortions—whether from telescope movements [23] or Earth's rotation [4] are not well suited for these conditions.", + "bbox": [ + 89, + 267, + 482, + 448 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We present Super Resolved Imaging with Adaptive Optics, an end-to-end SR method for terrestrial telescopes that leverages existing adaptive optics (AO) systems—designed to correct for atmospheric distortions—to simultaneously correct for aberrations and achieve SR. The crux of our approach is to take advantage of the existing AO system in ground telescopes, which applies a phase mask in the telescope's Fourier domain to counteract atmospheric aberrations. We use the AO system by optimizing and injecting additional phase displacements that produce sub-pixel shifts in the image plane. After capturing several images, these shifts serve as inputs to an end-to-end trained, multi-image SR (MISR) method. These additional phases are optimized to minimally alter the telescope's point spread function (PSF), thus subtly modifying each exposure to best facilitate MISR. The phase displacements can be co-optimized with deep learning-based (or any differentiable) MISR to enhance overall performance. This optimization is performed entirely offline, requiring no modifications to the telescope or use of on-sky time.", + "bbox": [ + 89, + 458, + 482, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions include:", + "bbox": [ + 109, + 770, + 369, + 785 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A novel MISR method for telescopes that leverages existing hardware, requires no modifications, and does not impact the science payload.", + "- Demonstration that combining AO with MISR outperforms either technique alone, achieving performance greater than the sum of its parts." + ], + "bbox": [ + 106, + 795, + 478, + 882 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Showing that phase displacements and image reconstruction can be co-optimized end-to-end offline using simple telescope telemetry as a proxy for real systems, enabling unobtrusive deployment where telescope time is extremely costly [31].", + "bbox": [ + 527, + 90, + 901, + 162 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- A real-hardware setup demonstrating the feasibility of integrating our method into existing telescope systems.", + "bbox": [ + 527, + 167, + 901, + 194 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Principles of Adaptive Optics", + "text_level": 1, + "bbox": [ + 513, + 205, + 782, + 223 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The Earth's atmosphere fundamentally limits the resolving power of any telescope by introducing optical aberrations—one of the reasons why space-based telescopes such as the Hubble Space Telescope have outperformed many larger ground-based observatories. This results in the final telescope PSF being several factors worse than the diffraction limit—the best achievable resolution. As such, AO systems are now widely used in modern ground-based optical and near-infrared telescopes to reverse these effects.", + "bbox": [ + 511, + 231, + 903, + 366 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Components of an AO system", + "text_level": 1, + "bbox": [ + 513, + 372, + 774, + 387 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "AO systems correct atmospheric turbulence by introducing a phase equal in magnitude but opposite in sign to that of the atmosphere. A prototypical AO setup (Fig. 2) consists of a deformable mirror (DM), wavefront sensor (WFS), and control system. The DM, placed on the Fourier plane of the telescope, takes on the opposite phase of the atmosphere, resulting in corrected wavefronts of light which are then passed on to the science camera and back to the WFS. Because the atmosphere evolves over time, the WFS must continuously measure the residual, uncorrected wavefront, and send updated commands to the DM. This loop, implemented via the control system, typically operates on the order of $1 - 2\\mathrm{kHz}$ for modern AO systems.", + "bbox": [ + 511, + 397, + 903, + 592 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The most common types of WFSs are based on Shack-Hartmann [50] and pyramid [51] designs. To operate they require a \"guide-star\"—a distant light source which can be approximated as a point. Usually a bright star near the science object is used, but laser beacons can also be used to create an artificial star image in the upper atmosphere. The guide star samples a similar turbulence path as the science image without reducing its signal-to-noise ratio (SNR). By comparing the measured signal to a reference, we can calculate how the atmosphere has affected the wavefront across the telescope pupil. Pre-calibrated responses are then used to update the DM shape and negate the distortions [58].", + "bbox": [ + 511, + 603, + 903, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The DM can be constructed in several ways. Many of today's telescopes use a continuous design with a single flexible reflective membrane supported by a grid of discrete actuators [20]. These actuators can be individually pushed or pulled to collectively generate complex phase patterns. In contrast, many upcoming AO systems use grids of individual mirrors, enabling much larger DMs and corrections for", + "bbox": [ + 511, + 794, + 903, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "29143", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "even higher spatial frequencies [26].", + "bbox": [ + 89, + 90, + 331, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Once converged, the effects of the AO loop, shown in Fig. 2, are visible and have enabled many recent breakthroughs in modern astronomy. Direct imaging of exoplanets [42] and the first direct evidence of a super-massive black hole within our galaxy [25] would not have been possible without the increased resolution and contrast provided by AO. For additional details on how AO has been recently used in both astronomy and microscopy see [28].", + "bbox": [ + 89, + 116, + 483, + 238 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/29c1c3ffd47a923f8b0c3c045ab02c18de25cecbd978d2fd9516f1cb75198b51.jpg", + "image_caption": [ + "Figure 2. Starlight is aberrated by Earth's atmosphere $\\mathcal{A}$ , blurring the telescope PSF. An AO system (a) measures the atmosphere with a wavefront sensor and applies a corrective phase, $\\mathcal{D}$ , on its deformable mirror to cancel it out. When AO is off, resolution is directly limited by the seeing conditions (i.e., turbulence) at the observatory for both short and long exposures (b). When AO is enabled the effects of the atmosphere are removed. If sampled at or above the Nyquist rate, the resolution is limited only by the mirror diameter and its PSF forms a diffraction-limited Airy disk (c). Our method enables under-sampled instruments with AO to recover spatial resolution." + ], + "image_footnote": [], + "bbox": [ + 91, + 247, + 483, + 498 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. The AO Image Formation Model", + "text_level": 1, + "bbox": [ + 89, + 669, + 380, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Suppose we are observing a distant, constant object (e.g., a natural guide star) over an exposure time $T$ using a telescope without AO. The recorded image $I(x,y)$ in pixel coordinates $(x,y)$ is modeled as the convolution of the object $O(x,y)$ with an atmosphere-degraded PSF $\\mathcal{P}_{\\mathrm{TEL}}$ . Here, $\\mathcal{P}_{\\mathrm{TEL}}$ corresponds to a diffraction-limited PSF $\\mathcal{P}_{\\mathrm{IDEAL}}$ (i.e., an Airy Disk [59]) degraded by the time-varying turbulence of the Earth's atmosphere. In Fourier space, this turbulence can be represented by a phase distribution $\\mathcal{A}(u,v,t)$ with unit amplitude where $(u,v)$ are its spatial frequency coordinates. By the convolution theorem, this can be described in Fourier space with a point-wise multiplication:", + "bbox": [ + 89, + 694, + 483, + 876 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} \\left\\{\\mathcal {P} _ {\\mathrm {T E L}} \\right\\} (u, v, t) = e ^ {j \\mathcal {A} (u, v, t)} \\mathcal {F} \\left\\{\\mathcal {P} _ {\\text {I D E A L}} \\right\\} (u, v, t), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 111, + 883, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{F}\\{\\cdot\\}$ denotes the Fourier Transform operator. While the AO is in operation, a DM continuously applies an additional phase distortion $\\mathcal{D}(u,v,t)$ in the Fourier plane of the telescope to counteract the effects of the atmosphere. The resulting corrected PSF, $\\mathcal{P}_{\\mathrm{CORR}}$ , can be described in the Fourier domain as", + "bbox": [ + 511, + 90, + 906, + 181 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} \\left\\{\\mathcal {P} _ {\\text {C O R R}} \\right\\} (u, v, t) = e ^ {j \\mathcal {D} (u, v, t)} e ^ {j \\mathcal {A} (u, v, t)} \\mathcal {F} \\left\\{\\mathcal {P} _ {\\text {I D E A L}} \\right\\} (u, v, t). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 190, + 908, + 223 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In an ideal scenario, the atmospheric phase $\\mathcal{A}$ and the deformable mirror phase $\\mathcal{D}$ would perfectly cancel at every point and time; i.e., $\\mathcal{P}_{\\mathrm{CORR}} \\approx \\mathcal{P}_{\\mathrm{IDEAL}}$ .", + "bbox": [ + 511, + 223, + 905, + 270 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In practice the DM is limited by fitting errors and the finite range of its actuators which restricts the range of spatial frequencies it can correct. Additionally, the effects of non-circular mirror geometries [29] (e.g., hexagonal apertures) and mirror support structures cause the DM to deviate from ideal performance. Consequently, the final effective PSF of the telescope never truly reaches $\\mathcal{P}_{\\mathrm{IDEAL}}$ . We model these effects on $\\mathcal{P}_{\\mathrm{EFF}}$ in the Fourier domain as", + "bbox": [ + 511, + 279, + 906, + 400 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} \\left\\{\\mathcal {P} _ {\\text {E F F}} \\right\\} (u, v, t) = e ^ {j \\delta (u, v, t)} \\mathcal {F} \\left\\{\\mathcal {P} _ {\\text {C O R R}} \\right\\} (u, v, t), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 409, + 906, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\delta(u, v, t)$ encapsulates the hardware and optical constraints described above.", + "bbox": [ + 511, + 438, + 905, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that these contributions are predominantly static and can be approximated as $\\delta(u,v,t) \\approx \\delta(u,v)$ . For a well-performing, diffraction-limited AO system, variations will be short because the AO re-converges quickly. Therefore, over long exposures, they average out over time. Finally, we consider incoherent image formation, as astronomical sources emit light independently with random phases. Our image formation becomes", + "bbox": [ + 511, + 479, + 905, + 599 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI (x, y) = \\int_ {0} ^ {T} | O (x, y) | ^ {2} * \\left| \\mathcal {P} _ {\\text {E F F}} (x, y, t) \\right| ^ {2} \\mathrm {d} t + \\eta \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 609, + 905, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\ast$ denotes the convolution operator, $T$ is the total exposure time, and $\\eta$ is cumulative noise.", + "bbox": [ + 511, + 652, + 905, + 683 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Previous Work", + "text_level": 1, + "bbox": [ + 511, + 694, + 663, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image Super-Resolution (SR). SR techniques aim to reconstruct a high-resolution (HR) image from one or more observed low-resolution (LR) images [65]. However, the problem is ill-posed in nature, and many plausible solutions exist [41]. SR methods are distinguished between single-image (SISR) and multi-image (MISR) super-resolution depending on the number of images used. Deep-learning SR has emerged as the state-of-the-art for both SISR and MISR tasks [2, 41, 41, 56, 64], thanks to the growth of convolutional networks and large datasets [18, 35]. However, deep SR networks thrive on natural images but struggle with out-of-distribution data due to limited training data [2]. In", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "29144", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "such cases, incorporating priors based on sensor, object, and acquisition conditions can improve performance [2]. Astronomical imaging is a regime with unique conditions, including extreme low-light, non-static aberrations, and image features unlike those typically seen in terrestrial images. Hence, additional modifications to existing deep-learning SR techniques are required.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SR in Astronomical Imaging: Drizzle—a SR technique commonly used in astronomical imaging and astrophotography—aligns and linearly blends multiple deliberately dithered frames to produce a super-resolved image [23]. The technique has been widely adopted, including by the Hubble Space Telescope, and remains in use today [27]. However, the success of drizzling and its contemporaries hinges on highly accurate image registration and precise knowledge of the offsets [3, 8], a challenging task when the atmosphere adds time-varying aberrations and distortions [24]. In contrast, our method leverages the telescope's built-in AO system to inject optimized phasedomain \"jitters\" and does not use any type of image registration. SR techniques have also been leveraged to improve the resolution and quality of WFS measurements [46]. Our method, building on our previous work [54], is agnostic to the underlying WFS type and can remain applicable as WFS technology advances.", + "bbox": [ + 91, + 205, + 486, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SR in Computational Imaging: Camera system design often incorporates super-resolution. One early example jointly optimized both the optics and SISR algorithms for simple, refractive lenses [52]. Burst photography from handheld cameras—where small distortions are induced by hand shake—are extremely successful at recovering high-resolution information [7, 62] even in low-light conditions [38]. Deep learning methods have also been explored in the scientific imaging community for both microscopy [21] and satellite imaging [45]. MISR has also long played a role in classic camera development; many digital cameras are able to programmatically shift the imaging sensor between subsequent images which can then be combined into one higher resolution image (e.g., Sony's Pixel Shift Multi Shooting [15]). This work draws inspiration from previous works by incorporating end-to-end optimization of telescope mirror distortions with SR algorithms.", + "bbox": [ + 91, + 488, + 483, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Computational Wavefront Sensing and Control: CNNs have been used to improve WFS reconstruction [19, 47], as well as to predict atmospheric conditions to reduce servolag in the system [33, 53]. Similar techniques have shown to be effective in microscopy AO systems as well [17, 66]. More general-purpose methods have been proposed such as [60] which used coded wavefront sensors and spatial light modulators (SLMs) to achieve much higher order spatial frequency correction than is capable from typical AO sys", + "bbox": [ + 89, + 755, + 480, + 891 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tems in astronomy or microscopy. However, SLMs, unlike DMs, may not be suitable for correcting rapidly changing atmospheric conditions due to their slower speeds [10].", + "bbox": [ + 511, + 90, + 903, + 137 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Optical Phase and PSF Optimization: End-to-end optimization of optical systems and their image reconstruction algorithms have been used in thin lens design [57], holography [9, 49], and active illumination [12]. These methods typically focus on optimization of critical optical elements of the system, requiring either direct access to the optical system [63] or robust, differentiable simulations during optimization. Due to the complexity and operating costs of modern telescopes, modification to the underlying system or real-time access for network training would be infeasible. Instead, we focus on optimization of optical elements available to us in such a way that would not affect its operation, allowing us to side step these limitations.", + "bbox": [ + 511, + 146, + 906, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Super Resolving Adaptive Optics", + "text_level": 1, + "bbox": [ + 511, + 353, + 815, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This work augments existing MISR methods by learning the optimal way to enhance each individual image captured over a single exposure. Using the DM already present in a telescope's AO system, we inject optimized phase profiles, thereby enabling multi-image super-resolution.", + "bbox": [ + 511, + 378, + 905, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we present an end-to-end approach to determine optimal phase profiles while accounting for AO hardware constraints and atmospheric effects on final images. We discuss our approach's assumptions (Sec. 4.1), the framework for optimizing phase profiles (Sec. 4.2), compatibility with various image reconstruction methods for combining sub-exposures (Sec. 4.3), and the training details for obtaining the optimal phase profiles (Sec. 4.3).", + "bbox": [ + 511, + 464, + 905, + 584 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. The \"Small Phase Shift\" Assumption", + "text_level": 1, + "bbox": [ + 511, + 590, + 834, + 607 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To include the image formation model in our learning infrastructure we must make some reasonable assumptions about the process. Importantly, we assume that sub-pixel shifts negligibly impact the AO loop or science image. In our case this is reasonable due to the fact that we consider only under-sampled systems. If our phase shifts keep $\\mathcal{P}_{\\mathrm{TEL}}$ within the range of its under-sampled area, the aggregated effect over the entire exposure will be negligible1. The shifts define the DM \"flat\" (a standard feature of AO real-time controllers), and can be updated easily between exposures. Therefore, the AO system will run entirely as usual—unaware of the phase offsets applied by our method.", + "bbox": [ + 511, + 612, + 905, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Phase Shifting and Optimization", + "text_level": 1, + "bbox": [ + 511, + 799, + 802, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Because our DM is designed to create highly complex shapes, we are not restricted to planar movements like traditional methods [23]. We describe a given DM shape $S$", + "bbox": [ + 511, + 820, + 903, + 867 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1If necessary, we can also impose constraints on the magnitude of our learned phase shifts to guarantee science performance", + "bbox": [ + 511, + 875, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "29145", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "using either actuator positions or, as detailed below, a more general modal basis.", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, one approach would be to use a combination of Zernike polynomial modes, $\\mathcal{Z}$ . The Zernike modes are well suited for this problem because of their ability to represent rich phases (including planar shifts) with minimal coefficients and are well-suited for circular apertures, such as those of telescopes [32]. They can also provide some insight into the nature of a given phase due to the relationship between Zernike modes and classical optics aberrations (e.g., focus, coma, and astigmatism [13]). Besides planar shifts (tip and tilt Zernike modes), there is no trivial or well known choice of Zernike modes for super-resolution available. We therefore optimize their values in conjunction with the MISR algorithm of choice to estimate the best possible choice of modes (i.e., an end-to-end optimization).", + "bbox": [ + 89, + 131, + 483, + 343 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We can now update our image formation model from a single exposure $I_{i}$ of time $T$ to include $N$ exposures of time $T / N$ with individually estimable phases $S_{i = 1\\dots N}$", + "bbox": [ + 89, + 352, + 483, + 398 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI _ {i} = | O | ^ {2} \\circledast | \\mathcal {P} _ {i} \\circledast e ^ {j S _ {i}} | ^ {2} + \\eta . \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 407, + 482, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This same model can be equally applied to a simple shifting or Zernike model by replacing $\\mathcal{S}$ with $\\mathcal{Z}$ and choosing an appropriate number of Zernike coefficients.", + "bbox": [ + 89, + 435, + 483, + 481 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Now, let $f_{\\uparrow}(x \\mid \\theta)$ be a learned MISR function with parameters $\\theta$ , and $f_{\\Downarrow}(x)$ be a downsampling method which reduces an image's spatial dimensions by a fixed factor. Our end-to-end optical optimization problem is then,", + "bbox": [ + 89, + 491, + 483, + 551 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\{\\theta , \\mathcal {S} \\}} \\left| \\right.\\left| \\right. f _ {\\uparrow} \\left( \\right.f _ {\\downarrow} \\left( \\right.| O | ^ {2} * \\mid \\mathcal {P} _ {1 \\dots N} * e ^ {j \\mathcal {S} _ {1 \\dots N}} \\left. \\right| ^ {2} + \\eta \\mid \\theta\\left. \\right)\\left. \\right) - O \\left. \\right|\\left. \\right| _ {1}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 561, + 486, + 603 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $O$ is an image randomly sampled from a set of high-resolution images, $\\mathcal{P}_i$ is randomly chosen from a set of precomputed AO telescope PSFs, $S_{i}$ are our trainable phase profiles, and $||\\cdot ||_1$ denotes the $\\ell_1$ norm. As described earlier, $S$ can be represented in many ways. To explore its effects, we compare the most common methods in our experiments. These include:", + "bbox": [ + 89, + 603, + 483, + 708 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- No Learned Phase: The base case for CNN methods. The same number of exposures are given to the CNN during training—and still affected by $\\mathcal{P}_i$ —but without optimizing any per-exposure $S_i$ .", + "- Tip/Tilt: A common MISR technique where the image is translated by a sub-pixel shift [22]. The Classic case uses a sequence of exposures whose translational shifts form a regular sub-pixel raster scan. The Optimized variant initializes with the same shifts as Classic and then optimizes the values of those shifts.", + "- Zernike Modes: Each phase is represented using the Zernike basis and its coefficients are optimized." + ], + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Non-Modal: Instead of a basis function, the DM shape is directly optimized. This requires precise knowledge of its influence functions and calibration with respect to the WFS [6]. While simple in simulation, this proves to be difficult to use in practice.", + "bbox": [ + 513, + 90, + 905, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Super-Resolution Reconstruction Methods", + "text_level": 1, + "bbox": [ + 511, + 172, + 875, + 189 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although our method could work equally well with any super-resolution reconstruction method, for the purposes of this paper we focus on only two:", + "bbox": [ + 511, + 195, + 905, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, we consider a linear-reconstruction method similar to Drizzle [23], also known as the shift-and-add method [22]. This may be of interest in science cases where such methods are already widely used, albeit for systems without adaptive optics and with un-optimized phases.", + "bbox": [ + 511, + 250, + 905, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Second, we consider a state of the art deep learning-based SR algorithm. Specifically, we based our model on the widely known EDSR [39] single-image SR CNN. Our network was modified to incorporate the forward image formation model (Eq. (4)) as well as to accept any number of input channels, representing the phase-shifted exposures. The EDSR model was chosen for its speed, widespread adoption, and its avoidance of adversarial training (i.e., it does not rely on a GAN [16]) which can complicate training and introduce priors that may not be appropriate for scientific purposes2.", + "bbox": [ + 511, + 335, + 906, + 502 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Phase and Network Training Details", + "text_level": 1, + "bbox": [ + 511, + 508, + 828, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition to a training set of high-resolution images, our phase and upsampling parameters require a collection of system PSFs under the influence of the AO system and the expected turbulence. These can be generated by simulations of the system or collected from telescope telemetry.", + "bbox": [ + 511, + 530, + 905, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At training time, for a model expecting $N$ exposures, we provide the network a random batch of high-resolution images as well as a random sample of $N$ system PSFs. The low-resolution images are generated for each PSF, as detailed in Eq. (5), using a trainable phase profile for each subexposure. We now have low-resolution images affected by the system PSF in addition to our optimized phases. We also have the ground-truth high-resolution image, which can be used to compute the loss function in Eq. (6).", + "bbox": [ + 511, + 616, + 905, + 752 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results shown in this work were trained on a combination of the DIV2K [1] and Microsoft ImagePairs [30] datasets. Both CNN and linear reconstructors were implemented and trained with pyTorch [48] following the learning procedure described in the EDSR paper [39]. Unless stated elsewhere (e.g., Sec. 5.1), the following experi", + "bbox": [ + 511, + 762, + 905, + 853 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2Our outlined procedure can be easily adapted to any SR method including diffusion [37] or transformer [40] models, or traditional convex optimization methods [11].", + "bbox": [ + 511, + 862, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "29146", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "iments are all performed with $N = 4$ exposures. While our method can be used with any number of exposures, this value was chosen to easily compare to the classic \"tip/tilt\" methods when solving for a $2 \\times$ upscaling factor. Restricting the number of exposures to four highlights the performance differences between linear and deep learning-based reconstruction methods, as discussed in Secs. 5 to 7.", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Simulated Results", + "text_level": 1, + "bbox": [ + 89, + 210, + 266, + 226 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first evaluate our method entirely in simulation. We train SR parameters and phase profiles solely on synthetic PSFs generated by the discretized image-formation model in Eq. (5). After training, we run a new AO simulation, setting the optimized phase profiles as the deformable mirror's \"flat\" position. This imposes the optimized phase on the system, replicating our convolution in the simulated hardware. The PSF generated from this new simulation is then used as the PSF for evaluating test images.", + "bbox": [ + 89, + 234, + 483, + 371 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. AO Simulation Environment", + "text_level": 1, + "bbox": [ + 89, + 381, + 346, + 395 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To create a realistic AO simulation we use the OOMAO [14] toolbox. We implemented a realistic AO simulation with multiple layers of atmosphere to generate realistic PSFs for training. Our simulated AO system was set up to match our experimental hardware (detailed in Sec. 6) with an $8\\mathrm{m}$ diameter telescope and 397-actuator deformable mirror operating at $1000\\mathrm{Hz}$ . Each PSF was generated by performing a full AO simulation with an evolving atmosphere over a 2-second exposure. The simulated $\\mathcal{P}_{\\mathrm{TEL}}$ was therefore imaged on a simulated science camera under the effects of AO correction and realistic sensor noise. To cover a wide range of conditions, the wind direction, wind speed, and guide star magnitude were randomized for each simulation.", + "bbox": [ + 89, + 401, + 483, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect of Joint Optimization: We investigate how reconstruction results are affected by jointly optimizing phase profiles and the CNN reconstruction method. As shown in Fig. 4, the jointly-optimized approach yields better resolution enhancement than a naive approach that trains the CNN without phase optimization, both qualitatively and quantitatively.", + "bbox": [ + 89, + 608, + 483, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect of Super-Resolution Factor: We compare different reconstruction methods and phase representations for the case where $N = 4$ exposures. We compare our different phase representations against a baseline CNN case where the input to the neural network are four fixed unshifted PSFs with no estimizable phase and simple bilinear interpolation. Experimental results are shown in Sec. 5.1. Sample phase profiles and output can be found in Fig. 3.", + "bbox": [ + 89, + 724, + 483, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect of Number of Exposures: We consider the effect of number of exposures on each method, freezing the super-resolution factor. The results are shown in Table 1.", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/079f406041aafa267b210c2cff954d0bc7569e1451891d2e9c52bf3cbc5a5376.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodScale Factor
2×4×8×
CNNNo Learned Phase33.7227.9824.04
Tip/Tilt (Classic)39.5633.4625.47
Tip/Tilt (Optimized)39.6733.8826.32
Zernike Modes41.5134.0927.74
Non-Modal43.1334.2427.79
LinearTip/Tilt (Classic)28.5626.0423.32
Tip/Tilt (Optimized)29.1126.1223.26
Zernike Modes30.8426.1723.36
Non-Modal31.0826.6723.33
MethodNumber of Exposures
2468
CNNTip/Tilt (Classic)30.7733.8936.0336.31
Zernike Modes31.1334.1136.2937.44
Non-Modal31.1734.2436.2437.48
LinearTip/Tilt26.1126.1226.1226.11
Zernike Modes26.1426.1726.1626.17
Non-Modal26.6626.6726.7126.69
", + "bbox": [ + 545, + 87, + 874, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Effect of number of exposures and super-resolution factor on SR methods. Values presented in PSNR (dB) $\\uparrow$ . In general the more exposures, and the more expressive the phase basis, the better the results. Details on each method can be found in Sec. 4.2.", + "bbox": [ + 511, + 366, + 906, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 450, + 718, + 467 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To demonstrate our method's ability to work in realistic settings, we also implemented the experimental setup described in Sec. 5 on an optical table. Our setup, shown in Fig. 5, consists of a 97-actuator ALPAO deformable mirror and a custom-built $25 \\times 25$ sub-aperture Shack-Hartmann wavefront sensor. Our guide star was created with an $850\\mathrm{~nm}$ near-infrared laser whose pupil and beam size were matched to the DM's size $(13.5\\mathrm{~mm})$ . Realistic atmospheric turbulence was induced by a rotating Lexitek phase screen driven by a stepping motor. The sensor's native pixel pitch was $5.86\\mu \\mathrm{m}$ and the diffraction-limited spot was $22.95\\mu \\mathrm{m}$ ( $f = 400\\mathrm{~mm}$ , pupil size $13.5\\mathrm{~mm}$ , $\\lambda = 635\\mathrm{~nm}$ ). Note that even though our experimental hardware was not undersampled by design, we used $12 \\times$ on-sensor binning to capture individual images. This resulted in an effective pixel pitch of $70.3\\mu \\mathrm{m}$ .", + "bbox": [ + 511, + 474, + 906, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The AO loop was controlled with the open source pyRTC software4 which allows full control of the DM and science camera. This allowed us to define the optimized phase shifts to be the reference mirror \"flat\", from which the residual wavefront was measured by the AO WFS. Each image was then captured as usual, changing the DM's \"flat\" between exposures as dictated by the optimized phases, and then combining these images to restore our final SR image.", + "bbox": [ + 511, + 726, + 908, + 849 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "3Lexitek Inc. Phase Plates - https://lexitek.com", + "bbox": [ + 529, + 875, + 836, + 887 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "4pyRTC on Github: https://github.com/jacotay7/pyRTC", + "bbox": [ + 532, + 887, + 900, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "29147", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c6519791bbdc34218af30f4e09771ec1365aabc167ad5395c2b462392a8af882.jpg", + "image_caption": [ + "INPUT", + "Figure 3. Simulated SR results. Comparisons of reconstructions with super resolution factors of $2 \\times$ and $8 \\times$ (inset regions marked in red). Images were generated with $N = 4$ exposures and phases were optimized for each method and SR factor. For a full view of the results, please refer to the supplemental material [55]" + ], + "image_footnote": [], + "bbox": [ + 107, + 88, + 176, + 284 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/82558cc619ce16844a7134d5b5b6065eae9b4a3073a4c3e214ab07eb12cf7d1d.jpg", + "image_caption": [ + "OPTIMIZED PHASE" + ], + "image_footnote": [], + "bbox": [ + 205, + 98, + 352, + 141 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3bc147c196dc618ad92155a2ab49eff7b480b20fd4858f21e182951f700f5a60.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 377, + 94, + 524, + 141 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/767144f58486218e662454fabacfe7c13be9ec4c69e2a664b4b5bcd2c52e7bb4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 95, + 707, + 141 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f2b722cb5929d4d2c093f697ccd33fc2c1a55949d43f197826a8be3ecfe55a23.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 94, + 880, + 141 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c5eee0e14b1419f122ce471da2e9168cb85c21f209a75f88f1bdfd9062cf77d9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 191, + 143, + 364, + 208 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/64ba72ab7d3d9ce8a5e539a95faf5be13e8d601f65e2d44902c9c78131767d53.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 191, + 212, + 364, + 246 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/54f5df6e228dff27794b8bf05847001b0d472e32a75b1562bdc07eb0dbe9a5c6.jpg", + "image_caption": [ + "LINEAR $2 \\times$" + ], + "image_footnote": [], + "bbox": [ + 191, + 246, + 364, + 281 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4ef192cb17ad1bee79bc862559fb6969eda26ee1ccad688521640a307fe297d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 366, + 143, + 539, + 208 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eb8fa4e353775931a31f4d599041398335d90a75d7244d4ebad4f099b8614249.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 366, + 212, + 539, + 246 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c6be9a541ba11c0e6c81e8d5694ed9d5d48025058dc72085af484cce1db07936.jpg", + "image_caption": [ + "CNN 2×" + ], + "image_footnote": [], + "bbox": [ + 366, + 246, + 539, + 281 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/116d9a959290107b696da56766587aa6b1e38bf41c85e8e20eb34aaebca4d3a1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 545, + 143, + 720, + 208 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bdb19dbd4380011122d5f6c8fa530c3c8d726725d6eaf7f7bfd6ed203f3521d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 545, + 212, + 720, + 246 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a3f7e380236cbe19b423fb7efaad3a7a1876d6c9f18d7072184fcb6b3f3a387e.jpg", + "image_caption": [ + "LINEAR $8\\times$" + ], + "image_footnote": [], + "bbox": [ + 545, + 246, + 720, + 281 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d0fedd2c4b13b70a28065dda9d350127c889f7ba6ba189c68d1fc7e266c4a748.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 143, + 893, + 208 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6cef5f6300fcf7aabad55a9c482a99800af1a0e4e7349495279b1d5690dc5a88.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 212, + 893, + 246 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/366dfb1dc26e768494389fce14bd3f0d5cb584918e59aa77e6a9116bea76e996.jpg", + "image_caption": [ + "CNN 8×" + ], + "image_footnote": [], + "bbox": [ + 720, + 246, + 893, + 281 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cdfbe552a399a3fc8ce88659b043376a270c73db1af32818f55455d1e486dfbf.jpg", + "image_caption": [ + "USAF Target", + "Figure 4. Improved SR with jointly-optimized phase and CNN on simulated USAF resolution target. End-to-end training of both phase and CNN yields superior resolution improvements compared to a naive approach which does not optimize phases. The input to both CNNs is the inset region of the USAF target, marked in red. Qualitatively, by jointly optimizing the phase and the CNN, we can resolve smaller features and, quantitatively, achieve lower RMSE (displayed above each image)." + ], + "image_footnote": [], + "bbox": [ + 94, + 397, + 171, + 460 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c0852707515971bcc50aa711e47d138dbdf1300515d7b7110efdf21977fd36f4.jpg", + "image_caption": [ + "Input (USAF Insert)" + ], + "image_footnote": [], + "bbox": [ + 178, + 397, + 266, + 460 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/28166676520f708c71388ae5675c2edb153a11b53076a73653d33d8e72c6e517.jpg", + "image_caption": [ + "Naive CNN", + "RMSE $= 0.0544$" + ], + "image_footnote": [], + "bbox": [ + 276, + 398, + 364, + 459 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/df367d91ef4f47282f9e04dfea08556a6bec14ecc255545a06c80990c57de081.jpg", + "image_caption": [ + "CNN (Optimized Phase)", + "RMSE = 0.0185" + ], + "image_footnote": [], + "bbox": [ + 370, + 398, + 465, + 459 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1. Results", + "text_level": 1, + "bbox": [ + 89, + 613, + 184, + 627 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A green LED (centered at $530~\\mathrm{nm}$ ) was used to project photographic slides on the optical bench detailed above. To demonstrate our method's resolving power, we imaged a 1\" US Air Force resolution chart (Thorlabs Positive 1951 USAF Test Target Groups 2-7, $\\varnothing 1^{\\prime \\prime}$ ). Results are shown for both simulation (Fig. 4) and experiment (Fig. 6). We also show our method's performance on more general scenes using a variety of real-world photos in Fig. 6. Because our AO system was designed specifically for sensing in infrared wavelengths, the quality of our images (captured at $635~\\mathrm{nm}$ ) is limited by chromatic aberration and vignetting. As such, results in this paper are cropped and masked for clarity; raw images are provided in the supplemental materials [55].", + "bbox": [ + 89, + 633, + 483, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7. Discussion", + "text_level": 1, + "bbox": [ + 89, + 845, + 204, + 859 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "MTF Analysis: The modulation transfer function (MTF) is a measure of the achievable contrast for an optical system", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cb365a0e87768ea3fb3d082032a858ef70a78bcd1f020b4dac4388241295343b.jpg", + "image_caption": [ + "Figure 5. Optical setup. We present a top-down view of the optical bench where our experiments were conducted, labeling both the wavefront sensor path and the science (image) paths." + ], + "image_footnote": [], + "bbox": [ + 563, + 380, + 851, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "over a range of spatial frequencies [44]. An ideal lens will linearly decrease from an MTF of 1 down to 0 as the spa", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "29148", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/13caa61d6c4b9cb1d678eec7e2934cc1fceeae7f62bafc85d1966af2be352b5e.jpg", + "image_caption": [ + "Figure 6. Experimental results. We reconstruct $8 \\times$ upsampled images of the USFA resolution target and three natural images (Grapes, Bookshelf, Trees) from low resolution images (LR Input) captured with the optical setup depicted in Fig. 5. We present results using both linear and CNN method, each using their own optimized phase profiles. Both result in higher resolutions than the original input, as seen quantitatively on the USFA chart." + ], + "image_footnote": [], + "bbox": [ + 99, + 87, + 472, + 296 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tial frequency increases and, in general, the closer a measured MTF behaves to this the better its resolving power. The MTF plots in Fig. 7 correspond to the USAF simulation of Fig. 4 and the real experiment of Fig. 6 computed using a knife-edge method [36]. Both show an increase of MTF across all spatial resolutions and thus demonstrate our method's superiority (i.e., higher preserved contrast).", + "bbox": [ + 89, + 426, + 483, + 532 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c4c1c23bd24ae26a710620a60101adce5fa80642e0c075ebd6eb0fb478f7b405.jpg", + "image_caption": [ + "Figure 7. Comparison of simulated and experimental MTFs." + ], + "image_footnote": [], + "bbox": [ + 91, + 544, + 483, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Phase Analysis: One advantage of representing phase profiles using Zernike modes is the ability to infer exactly which optical aberrations are most important to our reconstruction algorithms. In Fig. 8, we plot the total absolute power of the first nine basis modes identified by our method for both the CNN and the linear reconstructor, using a fixed number of exposures while varying the magnification factor. The optimized phases reveal insights. First, total modal power increases with the magnification order. Intuitively, this makes sense as more movement within a pixel is needed to sample higher resolutions. Second, CNNs apply roughly the same power across all modes for a given magnification while the linear method has more variance. Perhaps this is because linear solvers cannot leverage higher-frequency spatial information in the image as well as CNNs can.", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/72a42af0d1574fdfa027b525e70c226b1a9810708ea68f1746d5e608a4721528.jpg", + "image_caption": [ + "Figure 8. Comparison of modal powers used to induce SR. We show modal power in the Zernike basis to explore the optimized phases (corresponding to the phases shown in Fig. 3). The linear method uses low-order, symmetric modes while the CNN spreads across all modes. Comparing $8 \\times$ vs. $2 \\times$ magnifications, they take opposite approaches: the CNN prefers low modes at higher resolutions, while the linear uses lower modes at low resolutions." + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 903, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations & Future Work: Currently, our work is limited to telescopes with a single-conjugate adaptive optics (SCAO) system, which treats the atmosphere as a single column. Many of today's observatories instead use multiple WFS and DMs. This leads to more accurate corrections across a wider FoV and can even be tuned to different atmospheric layers. Such systems do not preclude the use of our method, since our method can be extended to multiple DMs and potentially enhanced by their extra degrees of freedom.", + "bbox": [ + 511, + 353, + 906, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Other directions for future work include real-time viewing of video from the science camera by cycling through each phase multiple times to create individually super-resolved frames. Once extended to multiple DMs, additional multiplexing could also be achieved by varying each DM at different rates.", + "bbox": [ + 511, + 500, + 905, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Most importantly, now that we have demonstrated our methodology with an optical bench prototype, our current priority is to show its effectiveness on sky with a science class telescope and AO system.", + "bbox": [ + 511, + 599, + 905, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusion: This work shows that science images can encode sufficient MISR information using the AO systems already present in state-of-the-art telescope observatories. We achieve resolution gains without affecting the scientific data in both simulations and optical-bench experiments. With this method, we aim to unlock new scientific capabilities at existing observatories without altering the optical system.", + "bbox": [ + 511, + 670, + 905, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 511, + 790, + 681, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "RS acknowledges NVIDIA for GPU hardware support. EYHL acknowledges the support of NSERC under CGSD. SS acknowledges the support of the Canada Foundation for Innovation and the Ontario Research Fund. KNK acknowledges the support of NSERC under the RGPIN and RTI programs.", + "bbox": [ + 511, + 813, + 905, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "29149", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 126-135, 2017. 5", + "[2] Saeed Anwar, Salman Khan, and Nick Barnes. A deep journey into super-resolution: A survey. ACM Computing Surveys (CSUR), 53(3):1-34, 2020. 3, 4", + "[3] Roberto J. Avila, Warren J. Hack, and Stsci Team. Astrodrizzle: Aligning images from multiple instruments. In American Astronomical Society, 2012. 4", + "[4] Thilo Bauer. Super-resolution imaging: The use case of optical astronomy. In Proceedings of the IADIS International Conference Computer Graphics, Visualization, Computer Vision and Image Processing, Rome, pages 49-59, 2011. 2", + "[5] Jacques M Beckers. Adaptive optics for astronomy-principles, performance, and applications. In: Annual Review of Astronomy and Astrophysics. Vol. 31 (A94-12726 02-90), p. 13-62., 31:13-62, 1993. 2", + "[6] Anthony Berdeu, Michel Tallon, Éric Thiébaut, and Maud Langlois. Analytical modelling of adaptive optics systems: Role of the influence function. *Astronomy & Astrophysics*, 674:A112, 2023. 5", + "[7] Goutam Bhat, Martin Danelljan, Luc Van Gool, and Radu Timofte. Deep burst super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9209-9218, 2021. 4", + "[8] David Borncamp, J. M. Anderson, Norman A. Grogin, and Warren J. Hack. Acs/wfc revised geometric distortion for drizzlepac. In Instrument Science Report ACS/WFC, 2015. 4", + "[9] Praneeth Chakravarthula, Ethan Tseng, Tarun Srivastava, Henry Fuchs, and Felix Heide. Learned hardware-in-the-loop phase retrieval for holographic near-eye displays. ACM Transactions on Graphics (TOG), 39(6):1-18, 2020. 4", + "[10] Dorian Chan, Srinivas Narasimhan, and Matthew O'Toole. Holocurtains: Programming light curtains via binary holography. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 4", + "[11] Stanley H Chan, Xiran Wang, and Omar A Elgendy. Plug-and-play admm for image restoration: Fixed-point convergence and applications. IEEE Transactions on Computational Imaging, 3(1):84-98, 2016. 5", + "[12] Wenzheng Chen, Parsa Mirdehghan, Sanja Fidler, and Kiriakos N Kutulakos. Auto-tuning structured light by optical stochastic gradient descent. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5970-5980, 2020. 4", + "[13] Xu Cheng, Arthur Bradley, Sowmya Ravikumar, and Larry N Thibos. Visual impact of zernike and seidel forms of monochromatic aberrations. Optometry and Vision Science, 87(5):300-312, 2010. 5", + "[14] Rodolphe Conan and C Correia. Object-oriented matlab adaptive optics toolbox. In Adaptive Optics Systems IV, page 91486C. International Society for Optics and Photonics, 2014. 6" + ], + "bbox": [ + 93, + 116, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Sony Corporation. Pixel Shift Multi Shooting | SONY — support.d-imaging.sony.co.jp. https://support. d-imaging.sony.co.jp/support/ilm/cpsms/ ilce7rm3/en/index.html, 2025. [Accessed 30-03- 2025]. 4", + "[16] Antonia Creswell, Tom White, Vincent Dumoulin, Kai Arulkumaran, Biswa Sengupta, and Anil A Bharath. Generative adversarial networks: An overview. IEEE Signal Processing Magazine, 35(1):53-65, 2018. 5", + "[17] Benjamin P Cumming and Min Gu. Direct determination of aberration functions in microscopy by an artificial neural network. Optics Express, 28(10):14511-14521, 2020. 4", + "[18] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In Computer Vision - ECCV 2014, pages 184-199, Cham, 2014. Springer International Publishing. 3", + "[19] Theodore B DuBose, Dennis F Gardner, and Abbie T Watnik. Intensity-enhanced deep network wavefront reconstruction in shack-hartmann sensors. Optics Letters, 45(7):1699-1702, 2020. 4", + "[20] Mark A Ealey and John F Washeba. Continuous facesheet low voltage deformable mirrors. Optical Engineering, 29 (10):1191-1198, 1990. 2", + "[21] Linjing Fang, Fred Monroe, Sammy Weiser Novak, Lyndsey Kirk, Cara R Schiavon, Seungyoon B Yu, Tong Zhang, Melissa Wu, Kyle Kastner, Alaa Abdel Latif, et al. Deep learning-based point-scanning super-resolution imaging. Nature Methods, 18(4):406-416, 2021. 4", + "[22] Sina Farsiu, Dirk Robinson, Michael Elad, and Peyman Milanfar. Robust shift and add approach to superresolution. In Applications of Digital Image Processing XXVI, pages 121-130. SPIE, 2003. 5", + "[23] AS Fruchter and RN Hook. Drizzle: A method for the linear reconstruction of undersampled images. *Publications of the Astronomical Society of the Pacific*, 114(792):144, 2002. 1, 2, 4, 5", + "[24] Andrew S. Fruchter. A new method for band-limited imaging with undersampled detectors. *Publications of the Astronomical Society of the Pacific*, 123(902):497, 2011. 4", + "[25] Andrea M Ghez, BL Klein, M Morris, and EE Becklin. High proper-motion stars in the vicinity of sagittarius a*: Evidence for a supermassive black hole at the center of our galaxy. The Astrophysical Journal, 509(2):678, 1998. 3", + "[26] Roberto Gilmotzzi and Jason Spyromilio. The european extremely large telescope (e-elt). The Messenger, 127(11):3, 2007. 3", + "[27] S Gonzaga, W Hack, A Fruchter, and J Mack. The drizzlepac handbook. The DrizzlePac Handbook, 2012. 4", + "[28] Karen M Hampson, Raphaël Turcotte, Donald T Miller, Kazuhiro Kurokawa, Jared R Males, Na Ji, and Martin J Booth. Adaptive optics for high-resolution imaging. Nature Reviews Methods Primers, 1(1):1-26, 2021. 3", + "[29] James E Harvey and Christ Ftaclas. Diffraction effects of telescope secondary mirror spiders on various image-quality criteria. Applied Optics, 34(28):6337-6349, 1995. 3", + "[30] Hamid Reza Vaezi Joze, Ilya Zharkov, Karlton Powell, Carl Ringler, Luming Liang, Andy Roulston, Moshe Lutz," + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "29150", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Vivek Pradeep. Imagepairs: Realistic super resolution dataset via beam splitter camera rig. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 518-519, 2020. 5", + "[31] SR Kulkarni. Instruments on large optical telescopes-a case study. arXiv preprint arXiv:1606.06674, 2016. 2", + "[32] Vasudevan Lakshminarayanan and Andre Fleck. Zernike polynomials: a guide. Journal of Modern Optics, 58(7):545-561, 2011. 5", + "[33] Rico Landman, Sebastiaan Y Haffert, Vikram M Radhakrishnan, and Christoph U Keller. Self-optimizing adaptive optics control with reinforcement learning. In Adaptive Optics Systems VII, page 1144849. International Society for Optics and Photonics, 2020. 4", + "[34] Tod R Lauer. Combining undersampled dithered images. *Publications of the Astronomical Society of the Pacific*, 111 (756):227, 1999. 2", + "[35] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, and Wenzhe Shi. Photo-realistic single image super-resolution using a generative adversarial network, 2017. 3", + "[36] Hang Li, Changxiang Yan, and Jianbing Shao. Measurement of the modulation transfer function of infrared imaging system by modified slant edge method. Journal of the Optical Society of Korea, 20(3):381-388, 2016. 8", + "[37] Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479:47-59, 2022. 5", + "[38] Orly Liba, Kiran Murthy, Yun-Ta Tsai, Tim Brooks, Tianfan Xue, Nikhil Karnad, Qiurui He, Jonathan T Barron, Dillon Sharlet, Ryan Geiss, et al. Handheld mobile photography in very low light. ACM Transactions on Graphics (TOG), 38 (6):1-16, 2019. 4", + "[39] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 136-144, 2017. 5", + "[40] Zhisheng Lu, Juncheng Li, Hong Liu, Chaoyan Huang, Linlin Zhang, and Tieyong Zeng. Transformer for single image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 457-466, 2022. 5", + "[41] Bahattin Can Maral. Single image super-resolution methods: A survey. arXiv preprint arXiv:2202.11763, 2022. 3", + "[42] Christian Marois, Bruce Macintosh, Travis Barman, B Zuckerman, Inseok Song, Jennifer Patience, David Lafreniere, and René Doyon. Direct imaging of multiple planets orbiting the star hr 8799. science, 322(5906):1348-1352, 2008. 3", + "[43] Y Mellier, Abdurrouf Abdurrouf, JA Acevedo Barroso, A Achucarro, J Adamek, R Adam, GE Addison, N Aghanim, M Aguena, V Ajani, et al. Euclid. i. overview of the euclid mission. *Astronomy & Astrophysics*, 2024. 2", + "[44] HH Nasse. How to read mtf curves. Carl Zeiss, Camera Lens Division, page 33, 2008. 7" + ], + "bbox": [ + 91, + 90, + 483, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Ngoc Long Nguyen, Jérémy Anger, Axel Davy, Pablo Arias, and Gabriele Facciolo. Self-supervised multi-image superresolution for push-frame satellite images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1121-1131, 2021. 4", + "[46] Sylvain Oberti, Carlos Correia, Thierry Fusco, Benoit Neichel, and Pierre Guiraud. Super-resolution wavefront reconstruction. *Astronomy & Astrophysics*, 667:A48, 2022. 4", + "[47] Scott W Paine and James R Fienup. Machine learning for improved image-based wavefront sensing. Optics Letters, 43(6):1235-1238, 2018. 4", + "[48] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5", + "[49] Yifan Peng, Suyeon Choi, Nitish Padmanaban, and Gordon Wetzstein. Neural holography with camera-in-the-loop training. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 4", + "[50] Ben C Platt and Roland Shack. History and principles of shack-hartmann wavefront sensing, 2001. 2", + "[51] Roberto Ragazzoni, Emiliano Diolaiti, and Elise Vernet. A pyramid wavefront sensor with no dynamic modulation. Optics Communications, 208(1-3):51-60, 2002. 2", + "[52] Vincent Sitzmann, Steven Diamond, Yifan Peng, Xiong Dun, Stephen Boyd, Wolfgang Heidrich, Felix Heide, and Gordon Wetzstein. End-to-end optimization of optics and image processing for achromatic extended depth of field and superresolution imaging. ACM Transactions on Graphics (TOG), 37(4):1-13, 2018. 4", + "[53] Robin Swanson, Masen Lamb, Carlos M Correia, Suresh Sivanandam, and Kiriakos Kutulakos. Closed loop predictive control of adaptive optics systems with convolutional neural networks. Monthly Notices of the Royal Astronomical Society, 503(2):2944-2954, 2021. 4", + "[54] Robin Swanson, Suresh Sivanandam, and Kiriakos N Kutulakos. Adaptive optics mediated sub-pixel super resolution. In Adaptive Optics Systems VIII, pages 2594-2600. SPIE, 2022. 4", + "[55] Robin Swanson, Esther Y. H. Lin, Masen Lamb, Suresh Sivanandam, and Kiriakos N. Kutulakos. Super resolved imaging with adaptive optics: Supplemental document. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2025. 7", + "[56] Jing Tian and Kai-Kuang Ma. A survey on super-resolution imaging. Signal, Image and Video Processing, 5(3):329-342, 2011. 3", + "[57] Ethan Tseng, Shane Colburn, James Whitehead, Luocheng Huang, Seung-Hwan Baek, Arka Majumdar, and Felix Heide. Neural nano-optics for high-quality thin lens imaging. Nature Communications, 12(1):1-7, 2021. 4", + "[58] Robert K Tyson. Adaptive ptics engineering handbook. Marcel Dekker New York, 2000. 2", + "[59] Robert K Tyson and Benjamin West Frazier. Principles of adaptive optics. CRC press, 2022. 3" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "29151", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[60] Congli Wang, Qiang Fu, Xiong Dun, and Wolfgang Heidrich. Megapixel adaptive optics: towards correcting large-scale distortions in computational cameras. ACM Transactions on Graphics (TOG), 37(4):1-12, 2018. 4", + "[61] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020. 2", + "[62] Bartlomiej Wronski, Ignacio Garcia-Dorado, Manfred Ernst, Damien Kelly, Michael Krainin, Chia-Kai Liang, Marc Levoy, and Peyman Milanfar. Handheld multi-frame superresolution. ACM Transactions on Graphics (TOG), 38(4): 1-18, 2019. 4", + "[63] Haitang Yang, Esther Y. H. Lin, Kiriakos N. Kutulakos, and George V. Eleftheriades. Sub-wavelength passive single-shot computational super-oscillatory imaging. Optica, 9(12): 1444–1447, 2022. 4", + "[64] Wenming Yang, Xuechen Zhang, Yapeng Tian, Wei Wang, Jing-Hao Xue, and Qingmin Liao. Deep learning for single image super-resolution: A brief review. IEEE Transactions on Multimedia, 21(12):3106-3121, 2019. 3", + "[65] Linwei Yue, Huanfeng Shen, Jie Li, Qiangqiang Yuan, Hongyan Zhang, and Liangpei Zhang. Image superresolution: The techniques, applications, and future. Signal Processing, 128:389-408, 2016. 3", + "[66] Biwei Zhang, Jiazhu Zhu, Ke Si, and Wei Gong. Deep learning assisted zonal adaptive aberration correction. Frontiers in Physics, page 634, 2021. 4", + "[67] Hao Zhu, Mantang Guo, Hongdong Li, Qing Wang, and Antonio Robles-Kelly. Revisiting spatio-angular trade-off in light field cameras and extended applications in superresolution. IEEE Transactions on Visualization and Computer Graphics, 27(6):3019-3033, 2019. 2" + ], + "bbox": [ + 91, + 90, + 480, + 556 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "29152", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Super Resolved Imaging with Adaptive Optics/330a7bf8-492d-4b73-b28f-8e96f1e5081a_model.json b/2025/Super Resolved Imaging with Adaptive Optics/330a7bf8-492d-4b73-b28f-8e96f1e5081a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d42698a4749d4aba9b10d794e82999dbe8a3d8d1 --- /dev/null +++ b/2025/Super Resolved Imaging with Adaptive Optics/330a7bf8-492d-4b73-b28f-8e96f1e5081a_model.json @@ -0,0 +1,2653 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.265, + 0.131, + 0.734, + 0.154 + ], + "angle": 0, + "content": "Super Resolved Imaging with Adaptive Optics" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.192, + 0.938, + 0.211 + ], + "angle": 0, + "content": "Robin Swanson\\(^{1,2}\\) Esther Y. H. Lin\\(^{1}\\) Masen Lamb\\(^{3,4}\\) Suresh Sivanandam\\(^{1,2}\\) Kiriakos N. Kutulakos\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.226, + 0.815, + 0.266 + ], + "angle": 0, + "content": "1 University of Toronto 2 Dunlap Institute for Astronomy & Astrophysics 3 International Gemini Observatory 4 University of Victoria" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.277, + 0.905, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.509, + 0.908, + 0.568 + ], + "angle": 0, + "content": "Figure 1. Super Resolved Imaging with Adaptive Optics is a computational imaging method to enable multi-image super-resolution in scientific telescopes. Without any additional hardware, existing adaptive mirrors are used to inject diversity into each sub-exposure. When jointly reconstructed, we are able to recover higher spatial resolution from an otherwise undersampled signal. Each exposure's phase profiles is jointly optimized along with the image reconstruction algorithm—resulting in greatly improved performance." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.599, + 0.328, + 0.615 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.633, + 0.486, + 0.892 + ], + "angle": 0, + "content": "Astronomical telescopes suffer from a tradeoff between field of view (FoV) and image resolution: increasing the FoV leads to an optical field that is under-sampled by the science camera. This work presents a novel computational imaging approach to overcome this tradeoff by leveraging the existing adaptive optics (AO) systems in modern ground-based telescopes. Our key idea is to use the AO system's deformable mirror to apply a series of learned, precisely controlled distortions to the optical wavefront, producing a sequence of images that exhibit distinct, high-frequency, sub-pixel shifts. These images can then be jointly upsampled to yield the final super-resolved image. Crucially, we show this can be done while simultaneously maintaining the core AO operation—correcting for the unknown and rapidly changing wavefront distortions caused by Earth's atmosphere. To achieve this, we incorporate end-to-end optimization of both the induced mirror distortions and the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.601, + 0.909, + 0.753 + ], + "angle": 0, + "content": "upsampling algorithm, such that telescope-specific optics and temporal statistics of atmospheric wavefront distortions are accounted for. Our experimental results with a hardware prototype, as well as simulations, demonstrate significant SNR improvements of up to 12 dB over non-AO super-resolution baselines, using only existing telescope optics and no hardware modifications. Moreover, by using a precise bench-top replica of a complete telescope and AO system, we show that our methodology can be readily transferred to an operational telescope." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.79, + 0.646, + 0.807 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.815, + 0.908, + 0.861 + ], + "angle": 0, + "content": "Super-resolution (SR) techniques have been integral to astronomical imaging due to inherent telescope design constraints [23]. Telescopes can have many science instruments" + }, + { + "type": "page_footnote", + "bbox": [ + 0.537, + 0.876, + 0.865, + 0.901 + ], + "angle": 0, + "content": "Corresponding author: robin@cs.toronto.edu \nProject website: www.cs.toronto.edu/~robin/aosr" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29142" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.258 + ], + "angle": 0, + "content": "and in certain cases optical resolution is not their priority [67]. A large field of view (FoV), for example, necessitates undersampling to capture a wider view of the night sky. Another common case in photon-starved conditions is to increase signal-to-noise ratio by focusing the light onto fewer, larger, pixels, thereby incurring less read noise [43]. In these, and many other, cases although the telescope is capable of providing a diffraction-limited signal, the science instrument can be undersampled. Thus, the use of SR methods in astronomical imaging has a long history in both satellite and terrestrial telescopes [34]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.268, + 0.483, + 0.449 + ], + "angle": 0, + "content": "Unlike telescopes stationed in orbit, ground telescopes face an additional challenge: the Earth's atmosphere [5]. Our atmosphere is a source of dynamic optical aberrations which introduce additional difficulties regardless of a telescope's location, size, or design [58]. As the atmosphere evolves, small changes in index of refraction occur across the sky. This causes incoming light to distort as it transitions from outer space and through our telescope, blurring the images. Therefore, traditional multi-image SR methods which combine multiple sharp images with small geometric distortions—whether from telescope movements [23] or Earth's rotation [4] are not well suited for these conditions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.483, + 0.761 + ], + "angle": 0, + "content": "We present Super Resolved Imaging with Adaptive Optics, an end-to-end SR method for terrestrial telescopes that leverages existing adaptive optics (AO) systems—designed to correct for atmospheric distortions—to simultaneously correct for aberrations and achieve SR. The crux of our approach is to take advantage of the existing AO system in ground telescopes, which applies a phase mask in the telescope's Fourier domain to counteract atmospheric aberrations. We use the AO system by optimizing and injecting additional phase displacements that produce sub-pixel shifts in the image plane. After capturing several images, these shifts serve as inputs to an end-to-end trained, multi-image SR (MISR) method. These additional phases are optimized to minimally alter the telescope's point spread function (PSF), thus subtly modifying each exposure to best facilitate MISR. The phase displacements can be co-optimized with deep learning-based (or any differentiable) MISR to enhance overall performance. This optimization is performed entirely offline, requiring no modifications to the telescope or use of on-sky time." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.771, + 0.37, + 0.786 + ], + "angle": 0, + "content": "In summary, our contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.796, + 0.48, + 0.838 + ], + "angle": 0, + "content": "- A novel MISR method for telescopes that leverages existing hardware, requires no modifications, and does not impact the science payload." + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.842, + 0.48, + 0.883 + ], + "angle": 0, + "content": "- Demonstration that combining AO with MISR outperforms either technique alone, achieving performance greater than the sum of its parts." + }, + { + "type": "list", + "bbox": [ + 0.107, + 0.796, + 0.48, + 0.883 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.529, + 0.092, + 0.902, + 0.164 + ], + "angle": 0, + "content": "- Showing that phase displacements and image reconstruction can be co-optimized end-to-end offline using simple telescope telemetry as a proxy for real systems, enabling unobtrusive deployment where telescope time is extremely costly [31]." + }, + { + "type": "text", + "bbox": [ + 0.529, + 0.168, + 0.903, + 0.195 + ], + "angle": 0, + "content": "- A real-hardware setup demonstrating the feasibility of integrating our method into existing telescope systems." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.207, + 0.783, + 0.224 + ], + "angle": 0, + "content": "2. Principles of Adaptive Optics" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.232, + 0.905, + 0.367 + ], + "angle": 0, + "content": "The Earth's atmosphere fundamentally limits the resolving power of any telescope by introducing optical aberrations—one of the reasons why space-based telescopes such as the Hubble Space Telescope have outperformed many larger ground-based observatories. This results in the final telescope PSF being several factors worse than the diffraction limit—the best achievable resolution. As such, AO systems are now widely used in modern ground-based optical and near-infrared telescopes to reverse these effects." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.373, + 0.776, + 0.388 + ], + "angle": 0, + "content": "2.1. Components of an AO system" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.905, + 0.593 + ], + "angle": 0, + "content": "AO systems correct atmospheric turbulence by introducing a phase equal in magnitude but opposite in sign to that of the atmosphere. A prototypical AO setup (Fig. 2) consists of a deformable mirror (DM), wavefront sensor (WFS), and control system. The DM, placed on the Fourier plane of the telescope, takes on the opposite phase of the atmosphere, resulting in corrected wavefronts of light which are then passed on to the science camera and back to the WFS. Because the atmosphere evolves over time, the WFS must continuously measure the residual, uncorrected wavefront, and send updated commands to the DM. This loop, implemented via the control system, typically operates on the order of \\(1 - 2\\mathrm{kHz}\\) for modern AO systems." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.604, + 0.905, + 0.785 + ], + "angle": 0, + "content": "The most common types of WFSs are based on Shack-Hartmann [50] and pyramid [51] designs. To operate they require a \"guide-star\"—a distant light source which can be approximated as a point. Usually a bright star near the science object is used, but laser beacons can also be used to create an artificial star image in the upper atmosphere. The guide star samples a similar turbulence path as the science image without reducing its signal-to-noise ratio (SNR). By comparing the measured signal to a reference, we can calculate how the atmosphere has affected the wavefront across the telescope pupil. Pre-calibrated responses are then used to update the DM shape and negate the distortions [58]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.905, + 0.9 + ], + "angle": 0, + "content": "The DM can be constructed in several ways. Many of today's telescopes use a continuous design with a single flexible reflective membrane supported by a grid of discrete actuators [20]. These actuators can be individually pushed or pulled to collectively generate complex phase patterns. In contrast, many upcoming AO systems use grids of individual mirrors, enabling much larger DMs and corrections for" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29143" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.333, + 0.107 + ], + "angle": 0, + "content": "even higher spatial frequencies [26]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.117, + 0.484, + 0.239 + ], + "angle": 0, + "content": "Once converged, the effects of the AO loop, shown in Fig. 2, are visible and have enabled many recent breakthroughs in modern astronomy. Direct imaging of exoplanets [42] and the first direct evidence of a super-massive black hole within our galaxy [25] would not have been possible without the increased resolution and contrast provided by AO. For additional details on how AO has been recently used in both astronomy and microscopy see [28]." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.248, + 0.484, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.507, + 0.484, + 0.659 + ], + "angle": 0, + "content": "Figure 2. Starlight is aberrated by Earth's atmosphere \\(\\mathcal{A}\\), blurring the telescope PSF. An AO system (a) measures the atmosphere with a wavefront sensor and applies a corrective phase, \\(\\mathcal{D}\\), on its deformable mirror to cancel it out. When AO is off, resolution is directly limited by the seeing conditions (i.e., turbulence) at the observatory for both short and long exposures (b). When AO is enabled the effects of the atmosphere are removed. If sampled at or above the Nyquist rate, the resolution is limited only by the mirror diameter and its PSF forms a diffraction-limited Airy disk (c). Our method enables under-sampled instruments with AO to recover spatial resolution." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.67, + 0.381, + 0.687 + ], + "angle": 0, + "content": "2.2. The AO Image Formation Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.695, + 0.484, + 0.877 + ], + "angle": 0, + "content": "Suppose we are observing a distant, constant object (e.g., a natural guide star) over an exposure time \\( T \\) using a telescope without AO. The recorded image \\( I(x,y) \\) in pixel coordinates \\( (x,y) \\) is modeled as the convolution of the object \\( O(x,y) \\) with an atmosphere-degraded PSF \\( \\mathcal{P}_{\\mathrm{TEL}} \\). Here, \\( \\mathcal{P}_{\\mathrm{TEL}} \\) corresponds to a diffraction-limited PSF \\( \\mathcal{P}_{\\mathrm{IDEAL}} \\) (i.e., an Airy Disk [59]) degraded by the time-varying turbulence of the Earth's atmosphere. In Fourier space, this turbulence can be represented by a phase distribution \\( \\mathcal{A}(u,v,t) \\) with unit amplitude where \\( (u,v) \\) are its spatial frequency coordinates. By the convolution theorem, this can be described in Fourier space with a point-wise multiplication:" + }, + { + "type": "equation", + "bbox": [ + 0.112, + 0.884, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} \\left\\{\\mathcal {P} _ {\\mathrm {T E L}} \\right\\} (u, v, t) = e ^ {j \\mathcal {A} (u, v, t)} \\mathcal {F} \\left\\{\\mathcal {P} _ {\\text {I D E A L}} \\right\\} (u, v, t), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.182 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}\\{\\cdot\\}\\) denotes the Fourier Transform operator. While the AO is in operation, a DM continuously applies an additional phase distortion \\(\\mathcal{D}(u,v,t)\\) in the Fourier plane of the telescope to counteract the effects of the atmosphere. The resulting corrected PSF, \\(\\mathcal{P}_{\\mathrm{CORR}}\\), can be described in the Fourier domain as" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.191, + 0.909, + 0.224 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} \\left\\{\\mathcal {P} _ {\\text {C O R R}} \\right\\} (u, v, t) = e ^ {j \\mathcal {D} (u, v, t)} e ^ {j \\mathcal {A} (u, v, t)} \\mathcal {F} \\left\\{\\mathcal {P} _ {\\text {I D E A L}} \\right\\} (u, v, t). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.224, + 0.906, + 0.271 + ], + "angle": 0, + "content": "In an ideal scenario, the atmospheric phase \\(\\mathcal{A}\\) and the deformable mirror phase \\(\\mathcal{D}\\) would perfectly cancel at every point and time; i.e., \\(\\mathcal{P}_{\\mathrm{CORR}} \\approx \\mathcal{P}_{\\mathrm{IDEAL}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.28, + 0.907, + 0.401 + ], + "angle": 0, + "content": "In practice the DM is limited by fitting errors and the finite range of its actuators which restricts the range of spatial frequencies it can correct. Additionally, the effects of non-circular mirror geometries [29] (e.g., hexagonal apertures) and mirror support structures cause the DM to deviate from ideal performance. Consequently, the final effective PSF of the telescope never truly reaches \\(\\mathcal{P}_{\\mathrm{IDEAL}}\\). We model these effects on \\(\\mathcal{P}_{\\mathrm{EFF}}\\) in the Fourier domain as" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.41, + 0.907, + 0.43 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} \\left\\{\\mathcal {P} _ {\\text {E F F}} \\right\\} (u, v, t) = e ^ {j \\delta (u, v, t)} \\mathcal {F} \\left\\{\\mathcal {P} _ {\\text {C O R R}} \\right\\} (u, v, t), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.439, + 0.906, + 0.468 + ], + "angle": 0, + "content": "where \\(\\delta(u, v, t)\\) encapsulates the hardware and optical constraints described above." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.48, + 0.906, + 0.6 + ], + "angle": 0, + "content": "Note that these contributions are predominantly static and can be approximated as \\(\\delta(u,v,t) \\approx \\delta(u,v)\\). For a well-performing, diffraction-limited AO system, variations will be short because the AO re-converges quickly. Therefore, over long exposures, they average out over time. Finally, we consider incoherent image formation, as astronomical sources emit light independently with random phases. Our image formation becomes" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.61, + 0.906, + 0.645 + ], + "angle": 0, + "content": "\\[\nI (x, y) = \\int_ {0} ^ {T} | O (x, y) | ^ {2} * \\left| \\mathcal {P} _ {\\text {E F F}} (x, y, t) \\right| ^ {2} \\mathrm {d} t + \\eta \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.654, + 0.906, + 0.684 + ], + "angle": 0, + "content": "where \\(\\ast\\) denotes the convolution operator, \\(T\\) is the total exposure time, and \\(\\eta\\) is cumulative noise." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.665, + 0.71 + ], + "angle": 0, + "content": "3. Previous Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Image Super-Resolution (SR). SR techniques aim to reconstruct a high-resolution (HR) image from one or more observed low-resolution (LR) images [65]. However, the problem is ill-posed in nature, and many plausible solutions exist [41]. SR methods are distinguished between single-image (SISR) and multi-image (MISR) super-resolution depending on the number of images used. Deep-learning SR has emerged as the state-of-the-art for both SISR and MISR tasks [2, 41, 41, 56, 64], thanks to the growth of convolutional networks and large datasets [18, 35]. However, deep SR networks thrive on natural images but struggle with out-of-distribution data due to limited training data [2]. In" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29144" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "such cases, incorporating priors based on sensor, object, and acquisition conditions can improve performance [2]. Astronomical imaging is a regime with unique conditions, including extreme low-light, non-static aberrations, and image features unlike those typically seen in terrestrial images. Hence, additional modifications to existing deep-learning SR techniques are required." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.207, + 0.487, + 0.48 + ], + "angle": 0, + "content": "SR in Astronomical Imaging: Drizzle—a SR technique commonly used in astronomical imaging and astrophotography—aligns and linearly blends multiple deliberately dithered frames to produce a super-resolved image [23]. The technique has been widely adopted, including by the Hubble Space Telescope, and remains in use today [27]. However, the success of drizzling and its contemporaries hinges on highly accurate image registration and precise knowledge of the offsets [3, 8], a challenging task when the atmosphere adds time-varying aberrations and distortions [24]. In contrast, our method leverages the telescope's built-in AO system to inject optimized phasedomain \"jitters\" and does not use any type of image registration. SR techniques have also been leveraged to improve the resolution and quality of WFS measurements [46]. Our method, building on our previous work [54], is agnostic to the underlying WFS type and can remain applicable as WFS technology advances." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.489, + 0.485, + 0.747 + ], + "angle": 0, + "content": "SR in Computational Imaging: Camera system design often incorporates super-resolution. One early example jointly optimized both the optics and SISR algorithms for simple, refractive lenses [52]. Burst photography from handheld cameras—where small distortions are induced by hand shake—are extremely successful at recovering high-resolution information [7, 62] even in low-light conditions [38]. Deep learning methods have also been explored in the scientific imaging community for both microscopy [21] and satellite imaging [45]. MISR has also long played a role in classic camera development; many digital cameras are able to programmatically shift the imaging sensor between subsequent images which can then be combined into one higher resolution image (e.g., Sony's Pixel Shift Multi Shooting [15]). This work draws inspiration from previous works by incorporating end-to-end optimization of telescope mirror distortions with SR algorithms." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.756, + 0.482, + 0.892 + ], + "angle": 0, + "content": "Computational Wavefront Sensing and Control: CNNs have been used to improve WFS reconstruction [19, 47], as well as to predict atmospheric conditions to reduce servolag in the system [33, 53]. Similar techniques have shown to be effective in microscopy AO systems as well [17, 66]. More general-purpose methods have been proposed such as [60] which used coded wavefront sensors and spatial light modulators (SLMs) to achieve much higher order spatial frequency correction than is capable from typical AO sys" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.138 + ], + "angle": 0, + "content": "tems in astronomy or microscopy. However, SLMs, unlike DMs, may not be suitable for correcting rapidly changing atmospheric conditions due to their slower speeds [10]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.147, + 0.907, + 0.344 + ], + "angle": 0, + "content": "Optical Phase and PSF Optimization: End-to-end optimization of optical systems and their image reconstruction algorithms have been used in thin lens design [57], holography [9, 49], and active illumination [12]. These methods typically focus on optimization of critical optical elements of the system, requiring either direct access to the optical system [63] or robust, differentiable simulations during optimization. Due to the complexity and operating costs of modern telescopes, modification to the underlying system or real-time access for network training would be infeasible. Instead, we focus on optimization of optical elements available to us in such a way that would not affect its operation, allowing us to side step these limitations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.354, + 0.816, + 0.373 + ], + "angle": 0, + "content": "4. Super Resolving Adaptive Optics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.379, + 0.906, + 0.456 + ], + "angle": 0, + "content": "This work augments existing MISR methods by learning the optimal way to enhance each individual image captured over a single exposure. Using the DM already present in a telescope's AO system, we inject optimized phase profiles, thereby enabling multi-image super-resolution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.465, + 0.906, + 0.585 + ], + "angle": 0, + "content": "In this section, we present an end-to-end approach to determine optimal phase profiles while accounting for AO hardware constraints and atmospheric effects on final images. We discuss our approach's assumptions (Sec. 4.1), the framework for optimizing phase profiles (Sec. 4.2), compatibility with various image reconstruction methods for combining sub-exposures (Sec. 4.3), and the training details for obtaining the optimal phase profiles (Sec. 4.3)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.591, + 0.835, + 0.608 + ], + "angle": 0, + "content": "4.1. The \"Small Phase Shift\" Assumption" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.906, + 0.794 + ], + "angle": 0, + "content": "To include the image formation model in our learning infrastructure we must make some reasonable assumptions about the process. Importantly, we assume that sub-pixel shifts negligibly impact the AO loop or science image. In our case this is reasonable due to the fact that we consider only under-sampled systems. If our phase shifts keep \\(\\mathcal{P}_{\\mathrm{TEL}}\\) within the range of its under-sampled area, the aggregated effect over the entire exposure will be negligible1. The shifts define the DM \"flat\" (a standard feature of AO real-time controllers), and can be updated easily between exposures. Therefore, the AO system will run entirely as usual—unaware of the phase offsets applied by our method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.8, + 0.803, + 0.817 + ], + "angle": 0, + "content": "4.2. Phase Shifting and Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.821, + 0.905, + 0.868 + ], + "angle": 0, + "content": "Because our DM is designed to create highly complex shapes, we are not restricted to planar movements like traditional methods [23]. We describe a given DM shape \\(S\\)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.907, + 0.901 + ], + "angle": 0, + "content": "1If necessary, we can also impose constraints on the magnitude of our learned phase shifts to guarantee science performance" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29145" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "using either actuator positions or, as detailed below, a more general modal basis." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.132, + 0.484, + 0.344 + ], + "angle": 0, + "content": "Specifically, one approach would be to use a combination of Zernike polynomial modes, \\(\\mathcal{Z}\\). The Zernike modes are well suited for this problem because of their ability to represent rich phases (including planar shifts) with minimal coefficients and are well-suited for circular apertures, such as those of telescopes [32]. They can also provide some insight into the nature of a given phase due to the relationship between Zernike modes and classical optics aberrations (e.g., focus, coma, and astigmatism [13]). Besides planar shifts (tip and tilt Zernike modes), there is no trivial or well known choice of Zernike modes for super-resolution available. We therefore optimize their values in conjunction with the MISR algorithm of choice to estimate the best possible choice of modes (i.e., an end-to-end optimization)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.353, + 0.484, + 0.4 + ], + "angle": 0, + "content": "We can now update our image formation model from a single exposure \\( I_{i} \\) of time \\( T \\) to include \\( N \\) exposures of time \\( T / N \\) with individually estimable phases \\( S_{i = 1\\dots N} \\)" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.408, + 0.483, + 0.427 + ], + "angle": 0, + "content": "\\[\nI _ {i} = | O | ^ {2} \\circledast | \\mathcal {P} _ {i} \\circledast e ^ {j S _ {i}} | ^ {2} + \\eta . \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.436, + 0.484, + 0.482 + ], + "angle": 0, + "content": "This same model can be equally applied to a simple shifting or Zernike model by replacing \\( \\mathcal{S} \\) with \\( \\mathcal{Z} \\) and choosing an appropriate number of Zernike coefficients." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.484, + 0.553 + ], + "angle": 0, + "content": "Now, let \\( f_{\\uparrow}(x \\mid \\theta) \\) be a learned MISR function with parameters \\( \\theta \\), and \\( f_{\\Downarrow}(x) \\) be a downsampling method which reduces an image's spatial dimensions by a fixed factor. Our end-to-end optical optimization problem is then," + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.562, + 0.487, + 0.604 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\{\\theta , \\mathcal {S} \\}} \\left| \\right.\\left| \\right. f _ {\\uparrow} \\left( \\right.f _ {\\downarrow} \\left( \\right.| O | ^ {2} * \\mid \\mathcal {P} _ {1 \\dots N} * e ^ {j \\mathcal {S} _ {1 \\dots N}} \\left. \\right| ^ {2} + \\eta \\mid \\theta\\left. \\right)\\left. \\right) - O \\left. \\right|\\left. \\right| _ {1}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.604, + 0.484, + 0.709 + ], + "angle": 0, + "content": "where \\(O\\) is an image randomly sampled from a set of high-resolution images, \\(\\mathcal{P}_i\\) is randomly chosen from a set of precomputed AO telescope PSFs, \\(S_{i}\\) are our trainable phase profiles, and \\(||\\cdot ||_1\\) denotes the \\(\\ell_1\\) norm. As described earlier, \\(S\\) can be represented in many ways. To explore its effects, we compare the most common methods in our experiments. These include:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.72, + 0.483, + 0.78 + ], + "angle": 0, + "content": "- No Learned Phase: The base case for CNN methods. The same number of exposures are given to the CNN during training—and still affected by \\(\\mathcal{P}_i\\)—but without optimizing any per-exposure \\(S_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.781, + 0.483, + 0.869 + ], + "angle": 0, + "content": "- Tip/Tilt: A common MISR technique where the image is translated by a sub-pixel shift [22]. The Classic case uses a sequence of exposures whose translational shifts form a regular sub-pixel raster scan. The Optimized variant initializes with the same shifts as Classic and then optimizes the values of those shifts." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- Zernike Modes: Each phase is represented using the Zernike basis and its coefficients are optimized." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "- Non-Modal: Instead of a basis function, the DM shape is directly optimized. This requires precise knowledge of its influence functions and calibration with respect to the WFS [6]. While simple in simulation, this proves to be difficult to use in practice." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.174, + 0.877, + 0.19 + ], + "angle": 0, + "content": "4.3. Super-Resolution Reconstruction Methods" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.196, + 0.906, + 0.241 + ], + "angle": 0, + "content": "Although our method could work equally well with any super-resolution reconstruction method, for the purposes of this paper we focus on only two:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.251, + 0.906, + 0.327 + ], + "angle": 0, + "content": "First, we consider a linear-reconstruction method similar to Drizzle [23], also known as the shift-and-add method [22]. This may be of interest in science cases where such methods are already widely used, albeit for systems without adaptive optics and with un-optimized phases." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.337, + 0.907, + 0.503 + ], + "angle": 0, + "content": "Second, we consider a state of the art deep learning-based SR algorithm. Specifically, we based our model on the widely known EDSR [39] single-image SR CNN. Our network was modified to incorporate the forward image formation model (Eq. (4)) as well as to accept any number of input channels, representing the phase-shifted exposures. The EDSR model was chosen for its speed, widespread adoption, and its avoidance of adversarial training (i.e., it does not rely on a GAN [16]) which can complicate training and introduce priors that may not be appropriate for scientific purposes2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.51, + 0.829, + 0.527 + ], + "angle": 0, + "content": "4.4. Phase and Network Training Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.531, + 0.906, + 0.607 + ], + "angle": 0, + "content": "In addition to a training set of high-resolution images, our phase and upsampling parameters require a collection of system PSFs under the influence of the AO system and the expected turbulence. These can be generated by simulations of the system or collected from telescope telemetry." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.617, + 0.906, + 0.753 + ], + "angle": 0, + "content": "At training time, for a model expecting \\( N \\) exposures, we provide the network a random batch of high-resolution images as well as a random sample of \\( N \\) system PSFs. The low-resolution images are generated for each PSF, as detailed in Eq. (5), using a trainable phase profile for each subexposure. We now have low-resolution images affected by the system PSF in addition to our optimized phases. We also have the ground-truth high-resolution image, which can be used to compute the loss function in Eq. (6)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.763, + 0.906, + 0.854 + ], + "angle": 0, + "content": "Results shown in this work were trained on a combination of the DIV2K [1] and Microsoft ImagePairs [30] datasets. Both CNN and linear reconstructors were implemented and trained with pyTorch [48] following the learning procedure described in the EDSR paper [39]. Unless stated elsewhere (e.g., Sec. 5.1), the following experi" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.863, + 0.906, + 0.901 + ], + "angle": 0, + "content": "2Our outlined procedure can be easily adapted to any SR method including diffusion [37] or transformer [40] models, or traditional convex optimization methods [11]." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29146" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.198 + ], + "angle": 0, + "content": "iments are all performed with \\( N = 4 \\) exposures. While our method can be used with any number of exposures, this value was chosen to easily compare to the classic \"tip/tilt\" methods when solving for a \\( 2 \\times \\) upscaling factor. Restricting the number of exposures to four highlights the performance differences between linear and deep learning-based reconstruction methods, as discussed in Secs. 5 to 7." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.211, + 0.267, + 0.227 + ], + "angle": 0, + "content": "5. Simulated Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.235, + 0.485, + 0.372 + ], + "angle": 0, + "content": "We first evaluate our method entirely in simulation. We train SR parameters and phase profiles solely on synthetic PSFs generated by the discretized image-formation model in Eq. (5). After training, we run a new AO simulation, setting the optimized phase profiles as the deformable mirror's \"flat\" position. This imposes the optimized phase on the system, replicating our convolution in the simulated hardware. The PSF generated from this new simulation is then used as the PSF for evaluating test images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.382, + 0.347, + 0.396 + ], + "angle": 0, + "content": "5.1. AO Simulation Environment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.402, + 0.484, + 0.599 + ], + "angle": 0, + "content": "To create a realistic AO simulation we use the OOMAO [14] toolbox. We implemented a realistic AO simulation with multiple layers of atmosphere to generate realistic PSFs for training. Our simulated AO system was set up to match our experimental hardware (detailed in Sec. 6) with an \\(8\\mathrm{m}\\) diameter telescope and 397-actuator deformable mirror operating at \\(1000\\mathrm{Hz}\\). Each PSF was generated by performing a full AO simulation with an evolving atmosphere over a 2-second exposure. The simulated \\(\\mathcal{P}_{\\mathrm{TEL}}\\) was therefore imaged on a simulated science camera under the effects of AO correction and realistic sensor noise. To cover a wide range of conditions, the wind direction, wind speed, and guide star magnitude were randomized for each simulation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.609, + 0.484, + 0.716 + ], + "angle": 0, + "content": "Effect of Joint Optimization: We investigate how reconstruction results are affected by jointly optimizing phase profiles and the CNN reconstruction method. As shown in Fig. 4, the jointly-optimized approach yields better resolution enhancement than a naive approach that trains the CNN without phase optimization, both qualitatively and quantitatively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.725, + 0.484, + 0.846 + ], + "angle": 0, + "content": "Effect of Super-Resolution Factor: We compare different reconstruction methods and phase representations for the case where \\( N = 4 \\) exposures. We compare our different phase representations against a baseline CNN case where the input to the neural network are four fixed unshifted PSFs with no estimizable phase and simple bilinear interpolation. Experimental results are shown in Sec. 5.1. Sample phase profiles and output can be found in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Effect of Number of Exposures: We consider the effect of number of exposures on each method, freezing the super-resolution factor. The results are shown in Table 1." + }, + { + "type": "table", + "bbox": [ + 0.547, + 0.088, + 0.875, + 0.356 + ], + "angle": 0, + "content": "
MethodScale Factor
2×4×8×
CNNNo Learned Phase33.7227.9824.04
Tip/Tilt (Classic)39.5633.4625.47
Tip/Tilt (Optimized)39.6733.8826.32
Zernike Modes41.5134.0927.74
Non-Modal43.1334.2427.79
LinearTip/Tilt (Classic)28.5626.0423.32
Tip/Tilt (Optimized)29.1126.1223.26
Zernike Modes30.8426.1723.36
Non-Modal31.0826.6723.33
MethodNumber of Exposures
2468
CNNTip/Tilt (Classic)30.7733.8936.0336.31
Zernike Modes31.1334.1136.2937.44
Non-Modal31.1734.2436.2437.48
LinearTip/Tilt26.1126.1226.1226.11
Zernike Modes26.1426.1726.1626.17
Non-Modal26.6626.6726.7126.69
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.367, + 0.907, + 0.424 + ], + "angle": 0, + "content": "Table 1. Effect of number of exposures and super-resolution factor on SR methods. Values presented in PSNR (dB) \\(\\uparrow\\). In general the more exposures, and the more expressive the phase basis, the better the results. Details on each method can be found in Sec. 4.2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.451, + 0.72, + 0.468 + ], + "angle": 0, + "content": "6. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.907, + 0.718 + ], + "angle": 0, + "content": "To demonstrate our method's ability to work in realistic settings, we also implemented the experimental setup described in Sec. 5 on an optical table. Our setup, shown in Fig. 5, consists of a 97-actuator ALPAO deformable mirror and a custom-built \\(25 \\times 25\\) sub-aperture Shack-Hartmann wavefront sensor. Our guide star was created with an \\(850\\mathrm{~nm}\\) near-infrared laser whose pupil and beam size were matched to the DM's size \\((13.5\\mathrm{~mm})\\). Realistic atmospheric turbulence was induced by a rotating Lexitek phase screen driven by a stepping motor. The sensor's native pixel pitch was \\(5.86\\mu \\mathrm{m}\\) and the diffraction-limited spot was \\(22.95\\mu \\mathrm{m}\\) (\\(f = 400\\mathrm{~mm}\\), pupil size \\(13.5\\mathrm{~mm}\\), \\(\\lambda = 635\\mathrm{~nm}\\)). Note that even though our experimental hardware was not undersampled by design, we used \\(12 \\times\\) on-sensor binning to capture individual images. This resulted in an effective pixel pitch of \\(70.3\\mu \\mathrm{m}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.727, + 0.909, + 0.85 + ], + "angle": 0, + "content": "The AO loop was controlled with the open source pyRTC software4 which allows full control of the DM and science camera. This allowed us to define the optimized phase shifts to be the reference mirror \"flat\", from which the residual wavefront was measured by the AO WFS. Each image was then captured as usual, changing the DM's \"flat\" between exposures as dictated by the optimized phases, and then combining these images to restore our final SR image." + }, + { + "type": "page_footnote", + "bbox": [ + 0.53, + 0.875, + 0.838, + 0.888 + ], + "angle": 0, + "content": "3Lexitek Inc. Phase Plates - https://lexitek.com" + }, + { + "type": "page_footnote", + "bbox": [ + 0.533, + 0.888, + 0.901, + 0.901 + ], + "angle": 0, + "content": "4pyRTC on Github: https://github.com/jacotay7/pyRTC" + }, + { + "type": "list", + "bbox": [ + 0.53, + 0.875, + 0.901, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29147" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.089, + 0.177, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.287, + 0.162, + 0.297 + ], + "angle": 0, + "content": "INPUT" + }, + { + "type": "image_caption", + "bbox": [ + 0.207, + 0.09, + 0.277, + 0.099 + ], + "angle": 0, + "content": "OPTIMIZED PHASE" + }, + { + "type": "image", + "bbox": [ + 0.207, + 0.099, + 0.353, + 0.142 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.379, + 0.095, + 0.526, + 0.142 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.096, + 0.708, + 0.142 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.095, + 0.882, + 0.142 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.144, + 0.366, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.213, + 0.365, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.247, + 0.365, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.246, + 0.286, + 0.312, + 0.297 + ], + "angle": 0, + "content": "LINEAR \\(2 \\times\\)" + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.144, + 0.54, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.213, + 0.54, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.247, + 0.54, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.426, + 0.286, + 0.478, + 0.297 + ], + "angle": 0, + "content": "CNN 2×" + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.144, + 0.721, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.213, + 0.721, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.247, + 0.721, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.601, + 0.287, + 0.667, + 0.297 + ], + "angle": 0, + "content": "LINEAR \\(8\\times\\)" + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.144, + 0.895, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.213, + 0.895, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.247, + 0.895, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.781, + 0.286, + 0.832, + 0.297 + ], + "angle": 0, + "content": "CNN 8×" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.304, + 0.908, + 0.348 + ], + "angle": 0, + "content": "Figure 3. Simulated SR results. Comparisons of reconstructions with super resolution factors of \\(2 \\times\\) and \\(8 \\times\\) (inset regions marked in red). Images were generated with \\(N = 4\\) exposures and phases were optimized for each method and SR factor. For a full view of the results, please refer to the supplemental material [55]" + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.381, + 0.166, + 0.391 + ], + "angle": 0, + "content": "USAF Target" + }, + { + "type": "image_caption", + "bbox": [ + 0.179, + 0.381, + 0.266, + 0.391 + ], + "angle": 0, + "content": "Input (USAF Insert)" + }, + { + "type": "image_caption", + "bbox": [ + 0.297, + 0.382, + 0.351, + 0.391 + ], + "angle": 0, + "content": "Naive CNN" + }, + { + "type": "image_caption", + "bbox": [ + 0.297, + 0.391, + 0.365, + 0.4 + ], + "angle": 0, + "content": "RMSE \\(= 0.0544\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.382, + 0.475, + 0.391 + ], + "angle": 0, + "content": "CNN (Optimized Phase)" + }, + { + "type": "image_caption", + "bbox": [ + 0.392, + 0.391, + 0.45, + 0.398 + ], + "angle": 0, + "content": "RMSE = 0.0185" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.398, + 0.172, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.398, + 0.267, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.4, + 0.366, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.399, + 0.467, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.473, + 0.484, + 0.584 + ], + "angle": 0, + "content": "Figure 4. Improved SR with jointly-optimized phase and CNN on simulated USAF resolution target. End-to-end training of both phase and CNN yields superior resolution improvements compared to a naive approach which does not optimize phases. The input to both CNNs is the inset region of the USAF target, marked in red. Qualitatively, by jointly optimizing the phase and the CNN, we can resolve smaller features and, quantitatively, achieve lower RMSE (displayed above each image)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.614, + 0.185, + 0.628 + ], + "angle": 0, + "content": "6.1. Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.635, + 0.484, + 0.832 + ], + "angle": 0, + "content": "A green LED (centered at \\(530~\\mathrm{nm}\\)) was used to project photographic slides on the optical bench detailed above. To demonstrate our method's resolving power, we imaged a 1\" US Air Force resolution chart (Thorlabs Positive 1951 USAF Test Target Groups 2-7, \\(\\varnothing 1^{\\prime \\prime}\\)). Results are shown for both simulation (Fig. 4) and experiment (Fig. 6). We also show our method's performance on more general scenes using a variety of real-world photos in Fig. 6. Because our AO system was designed specifically for sensing in infrared wavelengths, the quality of our images (captured at \\(635~\\mathrm{nm}\\)) is limited by chromatic aberration and vignetting. As such, results in this paper are cropped and masked for clarity; raw images are provided in the supplemental materials [55]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.846, + 0.205, + 0.861 + ], + "angle": 0, + "content": "7. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "MTF Analysis: The modulation transfer function (MTF) is a measure of the achievable contrast for an optical system" + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.381, + 0.852, + 0.791 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.795, + 0.908, + 0.837 + ], + "angle": 0, + "content": "Figure 5. Optical setup. We present a top-down view of the optical bench where our experiments were conducted, labeling both the wavefront sensor path and the science (image) paths." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "over a range of spatial frequencies [44]. An ideal lens will linearly decrease from an MTF of 1 down to 0 as the spa" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29148" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.088, + 0.473, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.303, + 0.484, + 0.402 + ], + "angle": 0, + "content": "Figure 6. Experimental results. We reconstruct \\(8 \\times\\) upsampled images of the USFA resolution target and three natural images (Grapes, Bookshelf, Trees) from low resolution images (LR Input) captured with the optical setup depicted in Fig. 5. We present results using both linear and CNN method, each using their own optimized phase profiles. Both result in higher resolutions than the original input, as seen quantitatively on the USFA chart." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.428, + 0.484, + 0.534 + ], + "angle": 0, + "content": "tial frequency increases and, in general, the closer a measured MTF behaves to this the better its resolving power. The MTF plots in Fig. 7 correspond to the USAF simulation of Fig. 4 and the real experiment of Fig. 6 computed using a knife-edge method [36]. Both show an increase of MTF across all spatial resolutions and thus demonstrate our method's superiority (i.e., higher preserved contrast)." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.545, + 0.485, + 0.639 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.644, + 0.466, + 0.658 + ], + "angle": 0, + "content": "Figure 7. Comparison of simulated and experimental MTFs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Phase Analysis: One advantage of representing phase profiles using Zernike modes is the ability to infer exactly which optical aberrations are most important to our reconstruction algorithms. In Fig. 8, we plot the total absolute power of the first nine basis modes identified by our method for both the CNN and the linear reconstructor, using a fixed number of exposures while varying the magnification factor. The optimized phases reveal insights. First, total modal power increases with the magnification order. Intuitively, this makes sense as more movement within a pixel is needed to sample higher resolutions. Second, CNNs apply roughly the same power across all modes for a given magnification while the linear method has more variance. Perhaps this is because linear solvers cannot leverage higher-frequency spatial information in the image as well as CNNs can." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.905, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.227, + 0.907, + 0.325 + ], + "angle": 0, + "content": "Figure 8. Comparison of modal powers used to induce SR. We show modal power in the Zernike basis to explore the optimized phases (corresponding to the phases shown in Fig. 3). The linear method uses low-order, symmetric modes while the CNN spreads across all modes. Comparing \\(8 \\times\\) vs. \\(2 \\times\\) magnifications, they take opposite approaches: the CNN prefers low modes at higher resolutions, while the linear uses lower modes at low resolutions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.354, + 0.907, + 0.492 + ], + "angle": 0, + "content": "Limitations & Future Work: Currently, our work is limited to telescopes with a single-conjugate adaptive optics (SCAO) system, which treats the atmosphere as a single column. Many of today's observatories instead use multiple WFS and DMs. This leads to more accurate corrections across a wider FoV and can even be tuned to different atmospheric layers. Such systems do not preclude the use of our method, since our method can be extended to multiple DMs and potentially enhanced by their extra degrees of freedom." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.5, + 0.906, + 0.59 + ], + "angle": 0, + "content": "Other directions for future work include real-time viewing of video from the science camera by cycling through each phase multiple times to create individually super-resolved frames. Once extended to multiple DMs, additional multiplexing could also be achieved by varying each DM at different rates." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.601, + 0.906, + 0.661 + ], + "angle": 0, + "content": "Most importantly, now that we have demonstrated our methodology with an optical bench prototype, our current priority is to show its effectiveness on sky with a science class telescope and AO system." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.671, + 0.906, + 0.779 + ], + "angle": 0, + "content": "Conclusion: This work shows that science images can encode sufficient MISR information using the AO systems already present in state-of-the-art telescope observatories. We achieve resolution gains without affecting the scientific data in both simulations and optical-bench experiments. With this method, we aim to unlock new scientific capabilities at existing observatories without altering the optical system." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.791, + 0.682, + 0.808 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.814, + 0.906, + 0.899 + ], + "angle": 0, + "content": "RS acknowledges NVIDIA for GPU hardware support. EYHL acknowledges the support of NSERC under CGSD. SS acknowledges the support of the Canada Foundation for Innovation and the Ontario Research Fund. KNK acknowledges the support of NSERC under the RGPIN and RTI programs." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29149" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.117, + 0.484, + 0.185 + ], + "angle": 0, + "content": "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 126-135, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.229 + ], + "angle": 0, + "content": "[2] Saeed Anwar, Salman Khan, and Nick Barnes. A deep journey into super-resolution: A survey. ACM Computing Surveys (CSUR), 53(3):1-34, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.231, + 0.483, + 0.273 + ], + "angle": 0, + "content": "[3] Roberto J. Avila, Warren J. Hack, and Stsci Team. Astrodrizzle: Aligning images from multiple instruments. In American Astronomical Society, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.274, + 0.483, + 0.33 + ], + "angle": 0, + "content": "[4] Thilo Bauer. Super-resolution imaging: The use case of optical astronomy. In Proceedings of the IADIS International Conference Computer Graphics, Visualization, Computer Vision and Image Processing, Rome, pages 49-59, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.332, + 0.483, + 0.387 + ], + "angle": 0, + "content": "[5] Jacques M Beckers. Adaptive optics for astronomy-principles, performance, and applications. In: Annual Review of Astronomy and Astrophysics. Vol. 31 (A94-12726 02-90), p. 13-62., 31:13-62, 1993. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.388, + 0.483, + 0.443 + ], + "angle": 0, + "content": "[6] Anthony Berdeu, Michel Tallon, Éric Thiébaut, and Maud Langlois. Analytical modelling of adaptive optics systems: Role of the influence function. *Astronomy & Astrophysics*, 674:A112, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.446, + 0.484, + 0.5 + ], + "angle": 0, + "content": "[7] Goutam Bhat, Martin Danelljan, Luc Van Gool, and Radu Timofte. Deep burst super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9209-9218, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.502, + 0.483, + 0.544 + ], + "angle": 0, + "content": "[8] David Borncamp, J. M. Anderson, Norman A. Grogin, and Warren J. Hack. Acs/wfc revised geometric distortion for drizzlepac. In Instrument Science Report ACS/WFC, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.546, + 0.483, + 0.601 + ], + "angle": 0, + "content": "[9] Praneeth Chakravarthula, Ethan Tseng, Tarun Srivastava, Henry Fuchs, and Felix Heide. Learned hardware-in-the-loop phase retrieval for holographic near-eye displays. ACM Transactions on Graphics (TOG), 39(6):1-18, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.603, + 0.483, + 0.659 + ], + "angle": 0, + "content": "[10] Dorian Chan, Srinivas Narasimhan, and Matthew O'Toole. Holocurtains: Programming light curtains via binary holography. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.66, + 0.483, + 0.716 + ], + "angle": 0, + "content": "[11] Stanley H Chan, Xiran Wang, and Omar A Elgendy. Plug-and-play admm for image restoration: Fixed-point convergence and applications. IEEE Transactions on Computational Imaging, 3(1):84-98, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.718, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[12] Wenzheng Chen, Parsa Mirdehghan, Sanja Fidler, and Kiriakos N Kutulakos. Auto-tuning structured light by optical stochastic gradient descent. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5970-5980, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.788, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[13] Xu Cheng, Arthur Bradley, Sowmya Ravikumar, and Larry N Thibos. Visual impact of zernike and seidel forms of monochromatic aberrations. Optometry and Vision Science, 87(5):300-312, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.845, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[14] Rodolphe Conan and C Correia. Object-oriented matlab adaptive optics toolbox. In Adaptive Optics Systems IV, page 91486C. International Society for Optics and Photonics, 2014. 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.117, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[15] Sony Corporation. Pixel Shift Multi Shooting | SONY — support.d-imaging.sony.co.jp. https://support. d-imaging.sony.co.jp/support/ilm/cpsms/ ilce7rm3/en/index.html, 2025. [Accessed 30-03- 2025]. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.906, + 0.219 + ], + "angle": 0, + "content": "[16] Antonia Creswell, Tom White, Vincent Dumoulin, Kai Arulkumaran, Biswa Sengupta, and Anil A Bharath. Generative adversarial networks: An overview. IEEE Signal Processing Magazine, 35(1):53-65, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.906, + 0.261 + ], + "angle": 0, + "content": "[17] Benjamin P Cumming and Min Gu. Direct determination of aberration functions in microscopy by an artificial neural network. Optics Express, 28(10):14511-14521, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.906, + 0.318 + ], + "angle": 0, + "content": "[18] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In Computer Vision - ECCV 2014, pages 184-199, Cham, 2014. Springer International Publishing. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.906, + 0.373 + ], + "angle": 0, + "content": "[19] Theodore B DuBose, Dennis F Gardner, and Abbie T Watnik. Intensity-enhanced deep network wavefront reconstruction in shack-hartmann sensors. Optics Letters, 45(7):1699-1702, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.906, + 0.417 + ], + "angle": 0, + "content": "[20] Mark A Ealey and John F Washeba. Continuous facesheet low voltage deformable mirrors. Optical Engineering, 29 (10):1191-1198, 1990. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.906, + 0.488 + ], + "angle": 0, + "content": "[21] Linjing Fang, Fred Monroe, Sammy Weiser Novak, Lyndsey Kirk, Cara R Schiavon, Seungyoon B Yu, Tong Zhang, Melissa Wu, Kyle Kastner, Alaa Abdel Latif, et al. Deep learning-based point-scanning super-resolution imaging. Nature Methods, 18(4):406-416, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.906, + 0.543 + ], + "angle": 0, + "content": "[22] Sina Farsiu, Dirk Robinson, Michael Elad, and Peyman Milanfar. Robust shift and add approach to superresolution. In Applications of Digital Image Processing XXVI, pages 121-130. SPIE, 2003. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.546, + 0.906, + 0.6 + ], + "angle": 0, + "content": "[23] AS Fruchter and RN Hook. Drizzle: A method for the linear reconstruction of undersampled images. *Publications of the Astronomical Society of the Pacific*, 114(792):144, 2002. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.603, + 0.906, + 0.644 + ], + "angle": 0, + "content": "[24] Andrew S. Fruchter. A new method for band-limited imaging with undersampled detectors. *Publications of the Astronomical Society of the Pacific*, 123(902):497, 2011. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.646, + 0.906, + 0.701 + ], + "angle": 0, + "content": "[25] Andrea M Ghez, BL Klein, M Morris, and EE Becklin. High proper-motion stars in the vicinity of sagittarius a*: Evidence for a supermassive black hole at the center of our galaxy. The Astrophysical Journal, 509(2):678, 1998. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.703, + 0.906, + 0.743 + ], + "angle": 0, + "content": "[26] Roberto Gilmotzzi and Jason Spyromilio. The european extremely large telescope (e-elt). The Messenger, 127(11):3, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.906, + 0.772 + ], + "angle": 0, + "content": "[27] S Gonzaga, W Hack, A Fruchter, and J Mack. The drizzlepac handbook. The DrizzlePac Handbook, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.774, + 0.906, + 0.829 + ], + "angle": 0, + "content": "[28] Karen M Hampson, Raphaël Turcotte, Donald T Miller, Kazuhiro Kurokawa, Jared R Males, Na Ji, and Martin J Booth. Adaptive optics for high-resolution imaging. Nature Reviews Methods Primers, 1(1):1-26, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.906, + 0.871 + ], + "angle": 0, + "content": "[29] James E Harvey and Christ Ftaclas. Diffraction effects of telescope secondary mirror spiders on various image-quality criteria. Applied Optics, 34(28):6337-6349, 1995. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[30] Hamid Reza Vaezi Joze, Ilya Zharkov, Karlton Powell, Carl Ringler, Luming Liang, Andy Roulston, Moshe Lutz," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "29150" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.148 + ], + "angle": 0, + "content": "and Vivek Pradeep. Imagepairs: Realistic super resolution dataset via beam splitter camera rig. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 518-519, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.149, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[31] SR Kulkarni. Instruments on large optical telescopes-a case study. arXiv preprint arXiv:1606.06674, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[32] Vasudevan Lakshminarayanan and Andre Fleck. Zernike polynomials: a guide. Journal of Modern Optics, 58(7):545-561, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[33] Rico Landman, Sebastiaan Y Haffert, Vikram M Radhakrishnan, and Christoph U Keller. Self-optimizing adaptive optics control with reinforcement learning. In Adaptive Optics Systems VII, page 1144849. International Society for Optics and Photonics, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.333 + ], + "angle": 0, + "content": "[34] Tod R Lauer. Combining undersampled dithered images. *Publications of the Astronomical Society of the Pacific*, 111 (756):227, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.335, + 0.483, + 0.404 + ], + "angle": 0, + "content": "[35] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, and Wenzhe Shi. Photo-realistic single image super-resolution using a generative adversarial network, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.483, + 0.461 + ], + "angle": 0, + "content": "[36] Hang Li, Changxiang Yan, and Jianbing Shao. Measurement of the modulation transfer function of infrared imaging system by modified slant edge method. Journal of the Optical Society of Korea, 20(3):381-388, 2016. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.483, + 0.517 + ], + "angle": 0, + "content": "[37] Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479:47-59, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.483, + 0.587 + ], + "angle": 0, + "content": "[38] Orly Liba, Kiran Murthy, Yun-Ta Tsai, Tim Brooks, Tianfan Xue, Nikhil Karnad, Qiurui He, Jonathan T Barron, Dillon Sharlet, Ryan Geiss, et al. Handheld mobile photography in very low light. ACM Transactions on Graphics (TOG), 38 (6):1-16, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.483, + 0.659 + ], + "angle": 0, + "content": "[39] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 136-144, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.483, + 0.729 + ], + "angle": 0, + "content": "[40] Zhisheng Lu, Juncheng Li, Hong Liu, Chaoyan Huang, Linlin Zhang, and Tieyong Zeng. Transformer for single image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 457-466, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.731, + 0.483, + 0.759 + ], + "angle": 0, + "content": "[41] Bahattin Can Maral. Single image super-resolution methods: A survey. arXiv preprint arXiv:2202.11763, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[42] Christian Marois, Bruce Macintosh, Travis Barman, B Zuckerman, Inseok Song, Jennifer Patience, David Lafreniere, and René Doyon. Direct imaging of multiple planets orbiting the star hr 8799. science, 322(5906):1348-1352, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[43] Y Mellier, Abdurrouf Abdurrouf, JA Acevedo Barroso, A Achucarro, J Adamek, R Adam, GE Addison, N Aghanim, M Aguena, V Ajani, et al. Euclid. i. overview of the euclid mission. *Astronomy & Astrophysics*, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[44] HH Nasse. How to read mtf curves. Carl Zeiss, Camera Lens Division, page 33, 2008. 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.162 + ], + "angle": 0, + "content": "[45] Ngoc Long Nguyen, Jérémy Anger, Axel Davy, Pablo Arias, and Gabriele Facciolo. Self-supervised multi-image superresolution for push-frame satellite images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1121-1131, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.906, + 0.205 + ], + "angle": 0, + "content": "[46] Sylvain Oberti, Carlos Correia, Thierry Fusco, Benoit Neichel, and Pierre Guiraud. Super-resolution wavefront reconstruction. *Astronomy & Astrophysics*, 667:A48, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.906, + 0.247 + ], + "angle": 0, + "content": "[47] Scott W Paine and James R Fienup. Machine learning for improved image-based wavefront sensing. Optics Letters, 43(6):1235-1238, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.906, + 0.331 + ], + "angle": 0, + "content": "[48] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.906, + 0.387 + ], + "angle": 0, + "content": "[49] Yifan Peng, Suyeon Choi, Nitish Padmanaban, and Gordon Wetzstein. Neural holography with camera-in-the-loop training. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.906, + 0.418 + ], + "angle": 0, + "content": "[50] Ben C Platt and Roland Shack. History and principles of shack-hartmann wavefront sensing, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.906, + 0.46 + ], + "angle": 0, + "content": "[51] Roberto Ragazzoni, Emiliano Diolaiti, and Elise Vernet. A pyramid wavefront sensor with no dynamic modulation. Optics Communications, 208(1-3):51-60, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.906, + 0.545 + ], + "angle": 0, + "content": "[52] Vincent Sitzmann, Steven Diamond, Yifan Peng, Xiong Dun, Stephen Boyd, Wolfgang Heidrich, Felix Heide, and Gordon Wetzstein. End-to-end optimization of optics and image processing for achromatic extended depth of field and superresolution imaging. ACM Transactions on Graphics (TOG), 37(4):1-13, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.906, + 0.616 + ], + "angle": 0, + "content": "[53] Robin Swanson, Masen Lamb, Carlos M Correia, Suresh Sivanandam, and Kiriakos Kutulakos. Closed loop predictive control of adaptive optics systems with convolutional neural networks. Monthly Notices of the Royal Astronomical Society, 503(2):2944-2954, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.617, + 0.906, + 0.671 + ], + "angle": 0, + "content": "[54] Robin Swanson, Suresh Sivanandam, and Kiriakos N Kutulakos. Adaptive optics mediated sub-pixel super resolution. In Adaptive Optics Systems VIII, pages 2594-2600. SPIE, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.673, + 0.906, + 0.743 + ], + "angle": 0, + "content": "[55] Robin Swanson, Esther Y. H. Lin, Masen Lamb, Suresh Sivanandam, and Kiriakos N. Kutulakos. Super resolved imaging with adaptive optics: Supplemental document. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2025. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.745, + 0.906, + 0.785 + ], + "angle": 0, + "content": "[56] Jing Tian and Kai-Kuang Ma. A survey on super-resolution imaging. Signal, Image and Video Processing, 5(3):329-342, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.906, + 0.843 + ], + "angle": 0, + "content": "[57] Ethan Tseng, Shane Colburn, James Whitehead, Luocheng Huang, Seung-Hwan Baek, Arka Majumdar, and Felix Heide. Neural nano-optics for high-quality thin lens imaging. Nature Communications, 12(1):1-7, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.906, + 0.871 + ], + "angle": 0, + "content": "[58] Robert K Tyson. Adaptive ptics engineering handbook. Marcel Dekker New York, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.908, + 0.902 + ], + "angle": 0, + "content": "[59] Robert K Tyson and Benjamin West Frazier. Principles of adaptive optics. CRC press, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29151" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[60] Congli Wang, Qiang Fu, Xiong Dun, and Wolfgang Heidrich. Megapixel adaptive optics: towards correcting large-scale distortions in computational cameras. ACM Transactions on Graphics (TOG), 37(4):1-12, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[61] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[62] Bartlomiej Wronski, Ignacio Garcia-Dorado, Manfred Ernst, Damien Kelly, Michael Krainin, Chia-Kai Liang, Marc Levoy, and Peyman Milanfar. Handheld multi-frame superresolution. ACM Transactions on Graphics (TOG), 38(4): 1-18, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.33 + ], + "angle": 0, + "content": "[63] Haitang Yang, Esther Y. H. Lin, Kiriakos N. Kutulakos, and George V. Eleftheriades. Sub-wavelength passive single-shot computational super-oscillatory imaging. Optica, 9(12): 1444–1447, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[64] Wenming Yang, Xuechen Zhang, Yapeng Tian, Wei Wang, Jing-Hao Xue, and Qingmin Liao. Deep learning for single image super-resolution: A brief review. IEEE Transactions on Multimedia, 21(12):3106-3121, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.482, + 0.444 + ], + "angle": 0, + "content": "[65] Linwei Yue, Huanfeng Shen, Jie Li, Qiangqiang Yuan, Hongyan Zhang, and Liangpei Zhang. Image superresolution: The techniques, applications, and future. Signal Processing, 128:389-408, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.446, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[66] Biwei Zhang, Jiazhu Zhu, Ke Si, and Wei Gong. Deep learning assisted zonal adaptive aberration correction. Frontiers in Physics, page 634, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.489, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[67] Hao Zhu, Mantang Guo, Hongdong Li, Qing Wang, and Antonio Robles-Kelly. Revisiting spatio-angular trade-off in light field cameras and extended applications in superresolution. IEEE Transactions on Visualization and Computer Graphics, 27(6):3019-3033, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29152" + } + ] +] \ No newline at end of file diff --git a/2025/Super Resolved Imaging with Adaptive Optics/330a7bf8-492d-4b73-b28f-8e96f1e5081a_origin.pdf b/2025/Super Resolved Imaging with Adaptive Optics/330a7bf8-492d-4b73-b28f-8e96f1e5081a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7a9be05e8e06d69aa697b4e133ff7e3c1d5b887c --- /dev/null +++ b/2025/Super Resolved Imaging with Adaptive Optics/330a7bf8-492d-4b73-b28f-8e96f1e5081a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:30c12da5659e24a1ea0ae04fd0a1e09996fcf9534be94e0705a8b3c37216f254 +size 8823236 diff --git a/2025/Super Resolved Imaging with Adaptive Optics/full.md b/2025/Super Resolved Imaging with Adaptive Optics/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8b85b238b52d7958ee3c0ac00ac22b788b797b4e --- /dev/null +++ b/2025/Super Resolved Imaging with Adaptive Optics/full.md @@ -0,0 +1,351 @@ +# Super Resolved Imaging with Adaptive Optics + +Robin Swanson $^{1,2}$ Esther Y. H. Lin $^{1}$ Masen Lamb $^{3,4}$ Suresh Sivanandam $^{1,2}$ Kiriakos N. Kutulakos $^{1}$ + +1 University of Toronto 2 Dunlap Institute for Astronomy & Astrophysics 3 International Gemini Observatory 4 University of Victoria + +![](images/fb9a3cfe2feafcf49849fb82be20353e7180bc7520a4af8f625c8cf9e0434eed.jpg) +Figure 1. Super Resolved Imaging with Adaptive Optics is a computational imaging method to enable multi-image super-resolution in scientific telescopes. Without any additional hardware, existing adaptive mirrors are used to inject diversity into each sub-exposure. When jointly reconstructed, we are able to recover higher spatial resolution from an otherwise undersampled signal. Each exposure's phase profiles is jointly optimized along with the image reconstruction algorithm—resulting in greatly improved performance. + +# Abstract + +Astronomical telescopes suffer from a tradeoff between field of view (FoV) and image resolution: increasing the FoV leads to an optical field that is under-sampled by the science camera. This work presents a novel computational imaging approach to overcome this tradeoff by leveraging the existing adaptive optics (AO) systems in modern ground-based telescopes. Our key idea is to use the AO system's deformable mirror to apply a series of learned, precisely controlled distortions to the optical wavefront, producing a sequence of images that exhibit distinct, high-frequency, sub-pixel shifts. These images can then be jointly upsampled to yield the final super-resolved image. Crucially, we show this can be done while simultaneously maintaining the core AO operation—correcting for the unknown and rapidly changing wavefront distortions caused by Earth's atmosphere. To achieve this, we incorporate end-to-end optimization of both the induced mirror distortions and the + +upsampling algorithm, such that telescope-specific optics and temporal statistics of atmospheric wavefront distortions are accounted for. Our experimental results with a hardware prototype, as well as simulations, demonstrate significant SNR improvements of up to 12 dB over non-AO super-resolution baselines, using only existing telescope optics and no hardware modifications. Moreover, by using a precise bench-top replica of a complete telescope and AO system, we show that our methodology can be readily transferred to an operational telescope. + +# 1. Introduction + +Super-resolution (SR) techniques have been integral to astronomical imaging due to inherent telescope design constraints [23]. Telescopes can have many science instruments + +and in certain cases optical resolution is not their priority [67]. A large field of view (FoV), for example, necessitates undersampling to capture a wider view of the night sky. Another common case in photon-starved conditions is to increase signal-to-noise ratio by focusing the light onto fewer, larger, pixels, thereby incurring less read noise [43]. In these, and many other, cases although the telescope is capable of providing a diffraction-limited signal, the science instrument can be undersampled. Thus, the use of SR methods in astronomical imaging has a long history in both satellite and terrestrial telescopes [34]. + +Unlike telescopes stationed in orbit, ground telescopes face an additional challenge: the Earth's atmosphere [5]. Our atmosphere is a source of dynamic optical aberrations which introduce additional difficulties regardless of a telescope's location, size, or design [58]. As the atmosphere evolves, small changes in index of refraction occur across the sky. This causes incoming light to distort as it transitions from outer space and through our telescope, blurring the images. Therefore, traditional multi-image SR methods which combine multiple sharp images with small geometric distortions—whether from telescope movements [23] or Earth's rotation [4] are not well suited for these conditions. + +We present Super Resolved Imaging with Adaptive Optics, an end-to-end SR method for terrestrial telescopes that leverages existing adaptive optics (AO) systems—designed to correct for atmospheric distortions—to simultaneously correct for aberrations and achieve SR. The crux of our approach is to take advantage of the existing AO system in ground telescopes, which applies a phase mask in the telescope's Fourier domain to counteract atmospheric aberrations. We use the AO system by optimizing and injecting additional phase displacements that produce sub-pixel shifts in the image plane. After capturing several images, these shifts serve as inputs to an end-to-end trained, multi-image SR (MISR) method. These additional phases are optimized to minimally alter the telescope's point spread function (PSF), thus subtly modifying each exposure to best facilitate MISR. The phase displacements can be co-optimized with deep learning-based (or any differentiable) MISR to enhance overall performance. This optimization is performed entirely offline, requiring no modifications to the telescope or use of on-sky time. + +In summary, our contributions include: + +- A novel MISR method for telescopes that leverages existing hardware, requires no modifications, and does not impact the science payload. +- Demonstration that combining AO with MISR outperforms either technique alone, achieving performance greater than the sum of its parts. + +- Showing that phase displacements and image reconstruction can be co-optimized end-to-end offline using simple telescope telemetry as a proxy for real systems, enabling unobtrusive deployment where telescope time is extremely costly [31]. + +- A real-hardware setup demonstrating the feasibility of integrating our method into existing telescope systems. + +# 2. Principles of Adaptive Optics + +The Earth's atmosphere fundamentally limits the resolving power of any telescope by introducing optical aberrations—one of the reasons why space-based telescopes such as the Hubble Space Telescope have outperformed many larger ground-based observatories. This results in the final telescope PSF being several factors worse than the diffraction limit—the best achievable resolution. As such, AO systems are now widely used in modern ground-based optical and near-infrared telescopes to reverse these effects. + +# 2.1. Components of an AO system + +AO systems correct atmospheric turbulence by introducing a phase equal in magnitude but opposite in sign to that of the atmosphere. A prototypical AO setup (Fig. 2) consists of a deformable mirror (DM), wavefront sensor (WFS), and control system. The DM, placed on the Fourier plane of the telescope, takes on the opposite phase of the atmosphere, resulting in corrected wavefronts of light which are then passed on to the science camera and back to the WFS. Because the atmosphere evolves over time, the WFS must continuously measure the residual, uncorrected wavefront, and send updated commands to the DM. This loop, implemented via the control system, typically operates on the order of $1 - 2\mathrm{kHz}$ for modern AO systems. + +The most common types of WFSs are based on Shack-Hartmann [50] and pyramid [51] designs. To operate they require a "guide-star"—a distant light source which can be approximated as a point. Usually a bright star near the science object is used, but laser beacons can also be used to create an artificial star image in the upper atmosphere. The guide star samples a similar turbulence path as the science image without reducing its signal-to-noise ratio (SNR). By comparing the measured signal to a reference, we can calculate how the atmosphere has affected the wavefront across the telescope pupil. Pre-calibrated responses are then used to update the DM shape and negate the distortions [58]. + +The DM can be constructed in several ways. Many of today's telescopes use a continuous design with a single flexible reflective membrane supported by a grid of discrete actuators [20]. These actuators can be individually pushed or pulled to collectively generate complex phase patterns. In contrast, many upcoming AO systems use grids of individual mirrors, enabling much larger DMs and corrections for + +even higher spatial frequencies [26]. + +Once converged, the effects of the AO loop, shown in Fig. 2, are visible and have enabled many recent breakthroughs in modern astronomy. Direct imaging of exoplanets [42] and the first direct evidence of a super-massive black hole within our galaxy [25] would not have been possible without the increased resolution and contrast provided by AO. For additional details on how AO has been recently used in both astronomy and microscopy see [28]. + +![](images/29c1c3ffd47a923f8b0c3c045ab02c18de25cecbd978d2fd9516f1cb75198b51.jpg) +Figure 2. Starlight is aberrated by Earth's atmosphere $\mathcal{A}$ , blurring the telescope PSF. An AO system (a) measures the atmosphere with a wavefront sensor and applies a corrective phase, $\mathcal{D}$ , on its deformable mirror to cancel it out. When AO is off, resolution is directly limited by the seeing conditions (i.e., turbulence) at the observatory for both short and long exposures (b). When AO is enabled the effects of the atmosphere are removed. If sampled at or above the Nyquist rate, the resolution is limited only by the mirror diameter and its PSF forms a diffraction-limited Airy disk (c). Our method enables under-sampled instruments with AO to recover spatial resolution. + +# 2.2. The AO Image Formation Model + +Suppose we are observing a distant, constant object (e.g., a natural guide star) over an exposure time $T$ using a telescope without AO. The recorded image $I(x,y)$ in pixel coordinates $(x,y)$ is modeled as the convolution of the object $O(x,y)$ with an atmosphere-degraded PSF $\mathcal{P}_{\mathrm{TEL}}$ . Here, $\mathcal{P}_{\mathrm{TEL}}$ corresponds to a diffraction-limited PSF $\mathcal{P}_{\mathrm{IDEAL}}$ (i.e., an Airy Disk [59]) degraded by the time-varying turbulence of the Earth's atmosphere. In Fourier space, this turbulence can be represented by a phase distribution $\mathcal{A}(u,v,t)$ with unit amplitude where $(u,v)$ are its spatial frequency coordinates. By the convolution theorem, this can be described in Fourier space with a point-wise multiplication: + +$$ +\mathcal {F} \left\{\mathcal {P} _ {\mathrm {T E L}} \right\} (u, v, t) = e ^ {j \mathcal {A} (u, v, t)} \mathcal {F} \left\{\mathcal {P} _ {\text {I D E A L}} \right\} (u, v, t), \tag {1} +$$ + +where $\mathcal{F}\{\cdot\}$ denotes the Fourier Transform operator. While the AO is in operation, a DM continuously applies an additional phase distortion $\mathcal{D}(u,v,t)$ in the Fourier plane of the telescope to counteract the effects of the atmosphere. The resulting corrected PSF, $\mathcal{P}_{\mathrm{CORR}}$ , can be described in the Fourier domain as + +$$ +\mathcal {F} \left\{\mathcal {P} _ {\text {C O R R}} \right\} (u, v, t) = e ^ {j \mathcal {D} (u, v, t)} e ^ {j \mathcal {A} (u, v, t)} \mathcal {F} \left\{\mathcal {P} _ {\text {I D E A L}} \right\} (u, v, t). \tag {2} +$$ + +In an ideal scenario, the atmospheric phase $\mathcal{A}$ and the deformable mirror phase $\mathcal{D}$ would perfectly cancel at every point and time; i.e., $\mathcal{P}_{\mathrm{CORR}} \approx \mathcal{P}_{\mathrm{IDEAL}}$ . + +In practice the DM is limited by fitting errors and the finite range of its actuators which restricts the range of spatial frequencies it can correct. Additionally, the effects of non-circular mirror geometries [29] (e.g., hexagonal apertures) and mirror support structures cause the DM to deviate from ideal performance. Consequently, the final effective PSF of the telescope never truly reaches $\mathcal{P}_{\mathrm{IDEAL}}$ . We model these effects on $\mathcal{P}_{\mathrm{EFF}}$ in the Fourier domain as + +$$ +\mathcal {F} \left\{\mathcal {P} _ {\text {E F F}} \right\} (u, v, t) = e ^ {j \delta (u, v, t)} \mathcal {F} \left\{\mathcal {P} _ {\text {C O R R}} \right\} (u, v, t), \tag {3} +$$ + +where $\delta(u, v, t)$ encapsulates the hardware and optical constraints described above. + +Note that these contributions are predominantly static and can be approximated as $\delta(u,v,t) \approx \delta(u,v)$ . For a well-performing, diffraction-limited AO system, variations will be short because the AO re-converges quickly. Therefore, over long exposures, they average out over time. Finally, we consider incoherent image formation, as astronomical sources emit light independently with random phases. Our image formation becomes + +$$ +I (x, y) = \int_ {0} ^ {T} | O (x, y) | ^ {2} * \left| \mathcal {P} _ {\text {E F F}} (x, y, t) \right| ^ {2} \mathrm {d} t + \eta \tag {4} +$$ + +where $\ast$ denotes the convolution operator, $T$ is the total exposure time, and $\eta$ is cumulative noise. + +# 3. Previous Work + +Image Super-Resolution (SR). SR techniques aim to reconstruct a high-resolution (HR) image from one or more observed low-resolution (LR) images [65]. However, the problem is ill-posed in nature, and many plausible solutions exist [41]. SR methods are distinguished between single-image (SISR) and multi-image (MISR) super-resolution depending on the number of images used. Deep-learning SR has emerged as the state-of-the-art for both SISR and MISR tasks [2, 41, 41, 56, 64], thanks to the growth of convolutional networks and large datasets [18, 35]. However, deep SR networks thrive on natural images but struggle with out-of-distribution data due to limited training data [2]. In + +such cases, incorporating priors based on sensor, object, and acquisition conditions can improve performance [2]. Astronomical imaging is a regime with unique conditions, including extreme low-light, non-static aberrations, and image features unlike those typically seen in terrestrial images. Hence, additional modifications to existing deep-learning SR techniques are required. + +SR in Astronomical Imaging: Drizzle—a SR technique commonly used in astronomical imaging and astrophotography—aligns and linearly blends multiple deliberately dithered frames to produce a super-resolved image [23]. The technique has been widely adopted, including by the Hubble Space Telescope, and remains in use today [27]. However, the success of drizzling and its contemporaries hinges on highly accurate image registration and precise knowledge of the offsets [3, 8], a challenging task when the atmosphere adds time-varying aberrations and distortions [24]. In contrast, our method leverages the telescope's built-in AO system to inject optimized phasedomain "jitters" and does not use any type of image registration. SR techniques have also been leveraged to improve the resolution and quality of WFS measurements [46]. Our method, building on our previous work [54], is agnostic to the underlying WFS type and can remain applicable as WFS technology advances. + +SR in Computational Imaging: Camera system design often incorporates super-resolution. One early example jointly optimized both the optics and SISR algorithms for simple, refractive lenses [52]. Burst photography from handheld cameras—where small distortions are induced by hand shake—are extremely successful at recovering high-resolution information [7, 62] even in low-light conditions [38]. Deep learning methods have also been explored in the scientific imaging community for both microscopy [21] and satellite imaging [45]. MISR has also long played a role in classic camera development; many digital cameras are able to programmatically shift the imaging sensor between subsequent images which can then be combined into one higher resolution image (e.g., Sony's Pixel Shift Multi Shooting [15]). This work draws inspiration from previous works by incorporating end-to-end optimization of telescope mirror distortions with SR algorithms. + +Computational Wavefront Sensing and Control: CNNs have been used to improve WFS reconstruction [19, 47], as well as to predict atmospheric conditions to reduce servolag in the system [33, 53]. Similar techniques have shown to be effective in microscopy AO systems as well [17, 66]. More general-purpose methods have been proposed such as [60] which used coded wavefront sensors and spatial light modulators (SLMs) to achieve much higher order spatial frequency correction than is capable from typical AO sys + +tems in astronomy or microscopy. However, SLMs, unlike DMs, may not be suitable for correcting rapidly changing atmospheric conditions due to their slower speeds [10]. + +Optical Phase and PSF Optimization: End-to-end optimization of optical systems and their image reconstruction algorithms have been used in thin lens design [57], holography [9, 49], and active illumination [12]. These methods typically focus on optimization of critical optical elements of the system, requiring either direct access to the optical system [63] or robust, differentiable simulations during optimization. Due to the complexity and operating costs of modern telescopes, modification to the underlying system or real-time access for network training would be infeasible. Instead, we focus on optimization of optical elements available to us in such a way that would not affect its operation, allowing us to side step these limitations. + +# 4. Super Resolving Adaptive Optics + +This work augments existing MISR methods by learning the optimal way to enhance each individual image captured over a single exposure. Using the DM already present in a telescope's AO system, we inject optimized phase profiles, thereby enabling multi-image super-resolution. + +In this section, we present an end-to-end approach to determine optimal phase profiles while accounting for AO hardware constraints and atmospheric effects on final images. We discuss our approach's assumptions (Sec. 4.1), the framework for optimizing phase profiles (Sec. 4.2), compatibility with various image reconstruction methods for combining sub-exposures (Sec. 4.3), and the training details for obtaining the optimal phase profiles (Sec. 4.3). + +# 4.1. The "Small Phase Shift" Assumption + +To include the image formation model in our learning infrastructure we must make some reasonable assumptions about the process. Importantly, we assume that sub-pixel shifts negligibly impact the AO loop or science image. In our case this is reasonable due to the fact that we consider only under-sampled systems. If our phase shifts keep $\mathcal{P}_{\mathrm{TEL}}$ within the range of its under-sampled area, the aggregated effect over the entire exposure will be negligible1. The shifts define the DM "flat" (a standard feature of AO real-time controllers), and can be updated easily between exposures. Therefore, the AO system will run entirely as usual—unaware of the phase offsets applied by our method. + +# 4.2. Phase Shifting and Optimization + +Because our DM is designed to create highly complex shapes, we are not restricted to planar movements like traditional methods [23]. We describe a given DM shape $S$ + +using either actuator positions or, as detailed below, a more general modal basis. + +Specifically, one approach would be to use a combination of Zernike polynomial modes, $\mathcal{Z}$ . The Zernike modes are well suited for this problem because of their ability to represent rich phases (including planar shifts) with minimal coefficients and are well-suited for circular apertures, such as those of telescopes [32]. They can also provide some insight into the nature of a given phase due to the relationship between Zernike modes and classical optics aberrations (e.g., focus, coma, and astigmatism [13]). Besides planar shifts (tip and tilt Zernike modes), there is no trivial or well known choice of Zernike modes for super-resolution available. We therefore optimize their values in conjunction with the MISR algorithm of choice to estimate the best possible choice of modes (i.e., an end-to-end optimization). + +We can now update our image formation model from a single exposure $I_{i}$ of time $T$ to include $N$ exposures of time $T / N$ with individually estimable phases $S_{i = 1\dots N}$ + +$$ +I _ {i} = | O | ^ {2} \circledast | \mathcal {P} _ {i} \circledast e ^ {j S _ {i}} | ^ {2} + \eta . \tag {5} +$$ + +This same model can be equally applied to a simple shifting or Zernike model by replacing $\mathcal{S}$ with $\mathcal{Z}$ and choosing an appropriate number of Zernike coefficients. + +Now, let $f_{\uparrow}(x \mid \theta)$ be a learned MISR function with parameters $\theta$ , and $f_{\Downarrow}(x)$ be a downsampling method which reduces an image's spatial dimensions by a fixed factor. Our end-to-end optical optimization problem is then, + +$$ +\min _ {\{\theta , \mathcal {S} \}} \left| \right.\left| \right. f _ {\uparrow} \left( \right.f _ {\downarrow} \left( \right.| O | ^ {2} * \mid \mathcal {P} _ {1 \dots N} * e ^ {j \mathcal {S} _ {1 \dots N}} \left. \right| ^ {2} + \eta \mid \theta\left. \right)\left. \right) - O \left. \right|\left. \right| _ {1}, \tag {6} +$$ + +where $O$ is an image randomly sampled from a set of high-resolution images, $\mathcal{P}_i$ is randomly chosen from a set of precomputed AO telescope PSFs, $S_{i}$ are our trainable phase profiles, and $||\cdot ||_1$ denotes the $\ell_1$ norm. As described earlier, $S$ can be represented in many ways. To explore its effects, we compare the most common methods in our experiments. These include: + +- No Learned Phase: The base case for CNN methods. The same number of exposures are given to the CNN during training—and still affected by $\mathcal{P}_i$ —but without optimizing any per-exposure $S_i$ . +- Tip/Tilt: A common MISR technique where the image is translated by a sub-pixel shift [22]. The Classic case uses a sequence of exposures whose translational shifts form a regular sub-pixel raster scan. The Optimized variant initializes with the same shifts as Classic and then optimizes the values of those shifts. +- Zernike Modes: Each phase is represented using the Zernike basis and its coefficients are optimized. + +- Non-Modal: Instead of a basis function, the DM shape is directly optimized. This requires precise knowledge of its influence functions and calibration with respect to the WFS [6]. While simple in simulation, this proves to be difficult to use in practice. + +# 4.3. Super-Resolution Reconstruction Methods + +Although our method could work equally well with any super-resolution reconstruction method, for the purposes of this paper we focus on only two: + +First, we consider a linear-reconstruction method similar to Drizzle [23], also known as the shift-and-add method [22]. This may be of interest in science cases where such methods are already widely used, albeit for systems without adaptive optics and with un-optimized phases. + +Second, we consider a state of the art deep learning-based SR algorithm. Specifically, we based our model on the widely known EDSR [39] single-image SR CNN. Our network was modified to incorporate the forward image formation model (Eq. (4)) as well as to accept any number of input channels, representing the phase-shifted exposures. The EDSR model was chosen for its speed, widespread adoption, and its avoidance of adversarial training (i.e., it does not rely on a GAN [16]) which can complicate training and introduce priors that may not be appropriate for scientific purposes2. + +# 4.4. Phase and Network Training Details + +In addition to a training set of high-resolution images, our phase and upsampling parameters require a collection of system PSFs under the influence of the AO system and the expected turbulence. These can be generated by simulations of the system or collected from telescope telemetry. + +At training time, for a model expecting $N$ exposures, we provide the network a random batch of high-resolution images as well as a random sample of $N$ system PSFs. The low-resolution images are generated for each PSF, as detailed in Eq. (5), using a trainable phase profile for each subexposure. We now have low-resolution images affected by the system PSF in addition to our optimized phases. We also have the ground-truth high-resolution image, which can be used to compute the loss function in Eq. (6). + +Results shown in this work were trained on a combination of the DIV2K [1] and Microsoft ImagePairs [30] datasets. Both CNN and linear reconstructors were implemented and trained with pyTorch [48] following the learning procedure described in the EDSR paper [39]. Unless stated elsewhere (e.g., Sec. 5.1), the following experi + +iments are all performed with $N = 4$ exposures. While our method can be used with any number of exposures, this value was chosen to easily compare to the classic "tip/tilt" methods when solving for a $2 \times$ upscaling factor. Restricting the number of exposures to four highlights the performance differences between linear and deep learning-based reconstruction methods, as discussed in Secs. 5 to 7. + +# 5. Simulated Results + +We first evaluate our method entirely in simulation. We train SR parameters and phase profiles solely on synthetic PSFs generated by the discretized image-formation model in Eq. (5). After training, we run a new AO simulation, setting the optimized phase profiles as the deformable mirror's "flat" position. This imposes the optimized phase on the system, replicating our convolution in the simulated hardware. The PSF generated from this new simulation is then used as the PSF for evaluating test images. + +# 5.1. AO Simulation Environment + +To create a realistic AO simulation we use the OOMAO [14] toolbox. We implemented a realistic AO simulation with multiple layers of atmosphere to generate realistic PSFs for training. Our simulated AO system was set up to match our experimental hardware (detailed in Sec. 6) with an $8\mathrm{m}$ diameter telescope and 397-actuator deformable mirror operating at $1000\mathrm{Hz}$ . Each PSF was generated by performing a full AO simulation with an evolving atmosphere over a 2-second exposure. The simulated $\mathcal{P}_{\mathrm{TEL}}$ was therefore imaged on a simulated science camera under the effects of AO correction and realistic sensor noise. To cover a wide range of conditions, the wind direction, wind speed, and guide star magnitude were randomized for each simulation. + +Effect of Joint Optimization: We investigate how reconstruction results are affected by jointly optimizing phase profiles and the CNN reconstruction method. As shown in Fig. 4, the jointly-optimized approach yields better resolution enhancement than a naive approach that trains the CNN without phase optimization, both qualitatively and quantitatively. + +Effect of Super-Resolution Factor: We compare different reconstruction methods and phase representations for the case where $N = 4$ exposures. We compare our different phase representations against a baseline CNN case where the input to the neural network are four fixed unshifted PSFs with no estimizable phase and simple bilinear interpolation. Experimental results are shown in Sec. 5.1. Sample phase profiles and output can be found in Fig. 3. + +Effect of Number of Exposures: We consider the effect of number of exposures on each method, freezing the super-resolution factor. The results are shown in Table 1. + +
MethodScale Factor
2×4×8×
CNNNo Learned Phase33.7227.9824.04
Tip/Tilt (Classic)39.5633.4625.47
Tip/Tilt (Optimized)39.6733.8826.32
Zernike Modes41.5134.0927.74
Non-Modal43.1334.2427.79
LinearTip/Tilt (Classic)28.5626.0423.32
Tip/Tilt (Optimized)29.1126.1223.26
Zernike Modes30.8426.1723.36
Non-Modal31.0826.6723.33
MethodNumber of Exposures
2468
CNNTip/Tilt (Classic)30.7733.8936.0336.31
Zernike Modes31.1334.1136.2937.44
Non-Modal31.1734.2436.2437.48
LinearTip/Tilt26.1126.1226.1226.11
Zernike Modes26.1426.1726.1626.17
Non-Modal26.6626.6726.7126.69
+ +Table 1. Effect of number of exposures and super-resolution factor on SR methods. Values presented in PSNR (dB) $\uparrow$ . In general the more exposures, and the more expressive the phase basis, the better the results. Details on each method can be found in Sec. 4.2. + +# 6. Experimental Results + +To demonstrate our method's ability to work in realistic settings, we also implemented the experimental setup described in Sec. 5 on an optical table. Our setup, shown in Fig. 5, consists of a 97-actuator ALPAO deformable mirror and a custom-built $25 \times 25$ sub-aperture Shack-Hartmann wavefront sensor. Our guide star was created with an $850\mathrm{~nm}$ near-infrared laser whose pupil and beam size were matched to the DM's size $(13.5\mathrm{~mm})$ . Realistic atmospheric turbulence was induced by a rotating Lexitek phase screen driven by a stepping motor. The sensor's native pixel pitch was $5.86\mu \mathrm{m}$ and the diffraction-limited spot was $22.95\mu \mathrm{m}$ ( $f = 400\mathrm{~mm}$ , pupil size $13.5\mathrm{~mm}$ , $\lambda = 635\mathrm{~nm}$ ). Note that even though our experimental hardware was not undersampled by design, we used $12 \times$ on-sensor binning to capture individual images. This resulted in an effective pixel pitch of $70.3\mu \mathrm{m}$ . + +The AO loop was controlled with the open source pyRTC software4 which allows full control of the DM and science camera. This allowed us to define the optimized phase shifts to be the reference mirror "flat", from which the residual wavefront was measured by the AO WFS. Each image was then captured as usual, changing the DM's "flat" between exposures as dictated by the optimized phases, and then combining these images to restore our final SR image. + +![](images/c6519791bbdc34218af30f4e09771ec1365aabc167ad5395c2b462392a8af882.jpg) +INPUT +Figure 3. Simulated SR results. Comparisons of reconstructions with super resolution factors of $2 \times$ and $8 \times$ (inset regions marked in red). Images were generated with $N = 4$ exposures and phases were optimized for each method and SR factor. For a full view of the results, please refer to the supplemental material [55] + +![](images/82558cc619ce16844a7134d5b5b6065eae9b4a3073a4c3e214ab07eb12cf7d1d.jpg) +OPTIMIZED PHASE + +![](images/3bc147c196dc618ad92155a2ab49eff7b480b20fd4858f21e182951f700f5a60.jpg) + +![](images/767144f58486218e662454fabacfe7c13be9ec4c69e2a664b4b5bcd2c52e7bb4.jpg) + +![](images/f2b722cb5929d4d2c093f697ccd33fc2c1a55949d43f197826a8be3ecfe55a23.jpg) + +![](images/c5eee0e14b1419f122ce471da2e9168cb85c21f209a75f88f1bdfd9062cf77d9.jpg) + +![](images/64ba72ab7d3d9ce8a5e539a95faf5be13e8d601f65e2d44902c9c78131767d53.jpg) + +![](images/54f5df6e228dff27794b8bf05847001b0d472e32a75b1562bdc07eb0dbe9a5c6.jpg) +LINEAR $2 \times$ + +![](images/4ef192cb17ad1bee79bc862559fb6969eda26ee1ccad688521640a307fe297d7.jpg) + +![](images/eb8fa4e353775931a31f4d599041398335d90a75d7244d4ebad4f099b8614249.jpg) + +![](images/c6be9a541ba11c0e6c81e8d5694ed9d5d48025058dc72085af484cce1db07936.jpg) +CNN 2× + +![](images/116d9a959290107b696da56766587aa6b1e38bf41c85e8e20eb34aaebca4d3a1.jpg) + +![](images/bdb19dbd4380011122d5f6c8fa530c3c8d726725d6eaf7f7bfd6ed203f3521d5.jpg) + +![](images/a3f7e380236cbe19b423fb7efaad3a7a1876d6c9f18d7072184fcb6b3f3a387e.jpg) +LINEAR $8\times$ + +![](images/d0fedd2c4b13b70a28065dda9d350127c889f7ba6ba189c68d1fc7e266c4a748.jpg) + +![](images/6cef5f6300fcf7aabad55a9c482a99800af1a0e4e7349495279b1d5690dc5a88.jpg) + +![](images/366dfb1dc26e768494389fce14bd3f0d5cb584918e59aa77e6a9116bea76e996.jpg) +CNN 8× + +![](images/cdfbe552a399a3fc8ce88659b043376a270c73db1af32818f55455d1e486dfbf.jpg) +USAF Target +Figure 4. Improved SR with jointly-optimized phase and CNN on simulated USAF resolution target. End-to-end training of both phase and CNN yields superior resolution improvements compared to a naive approach which does not optimize phases. The input to both CNNs is the inset region of the USAF target, marked in red. Qualitatively, by jointly optimizing the phase and the CNN, we can resolve smaller features and, quantitatively, achieve lower RMSE (displayed above each image). + +![](images/c0852707515971bcc50aa711e47d138dbdf1300515d7b7110efdf21977fd36f4.jpg) +Input (USAF Insert) + +![](images/28166676520f708c71388ae5675c2edb153a11b53076a73653d33d8e72c6e517.jpg) +Naive CNN +RMSE $= 0.0544$ + +![](images/df367d91ef4f47282f9e04dfea08556a6bec14ecc255545a06c80990c57de081.jpg) +CNN (Optimized Phase) +RMSE = 0.0185 + +# 6.1. Results + +A green LED (centered at $530~\mathrm{nm}$ ) was used to project photographic slides on the optical bench detailed above. To demonstrate our method's resolving power, we imaged a 1" US Air Force resolution chart (Thorlabs Positive 1951 USAF Test Target Groups 2-7, $\varnothing 1^{\prime \prime}$ ). Results are shown for both simulation (Fig. 4) and experiment (Fig. 6). We also show our method's performance on more general scenes using a variety of real-world photos in Fig. 6. Because our AO system was designed specifically for sensing in infrared wavelengths, the quality of our images (captured at $635~\mathrm{nm}$ ) is limited by chromatic aberration and vignetting. As such, results in this paper are cropped and masked for clarity; raw images are provided in the supplemental materials [55]. + +# 7. Discussion + +MTF Analysis: The modulation transfer function (MTF) is a measure of the achievable contrast for an optical system + +![](images/cb365a0e87768ea3fb3d082032a858ef70a78bcd1f020b4dac4388241295343b.jpg) +Figure 5. Optical setup. We present a top-down view of the optical bench where our experiments were conducted, labeling both the wavefront sensor path and the science (image) paths. + +over a range of spatial frequencies [44]. An ideal lens will linearly decrease from an MTF of 1 down to 0 as the spa + +![](images/13caa61d6c4b9cb1d678eec7e2934cc1fceeae7f62bafc85d1966af2be352b5e.jpg) +Figure 6. Experimental results. We reconstruct $8 \times$ upsampled images of the USFA resolution target and three natural images (Grapes, Bookshelf, Trees) from low resolution images (LR Input) captured with the optical setup depicted in Fig. 5. We present results using both linear and CNN method, each using their own optimized phase profiles. Both result in higher resolutions than the original input, as seen quantitatively on the USFA chart. + +tial frequency increases and, in general, the closer a measured MTF behaves to this the better its resolving power. The MTF plots in Fig. 7 correspond to the USAF simulation of Fig. 4 and the real experiment of Fig. 6 computed using a knife-edge method [36]. Both show an increase of MTF across all spatial resolutions and thus demonstrate our method's superiority (i.e., higher preserved contrast). + +![](images/c4c1c23bd24ae26a710620a60101adce5fa80642e0c075ebd6eb0fb478f7b405.jpg) +Figure 7. Comparison of simulated and experimental MTFs. + +Phase Analysis: One advantage of representing phase profiles using Zernike modes is the ability to infer exactly which optical aberrations are most important to our reconstruction algorithms. In Fig. 8, we plot the total absolute power of the first nine basis modes identified by our method for both the CNN and the linear reconstructor, using a fixed number of exposures while varying the magnification factor. The optimized phases reveal insights. First, total modal power increases with the magnification order. Intuitively, this makes sense as more movement within a pixel is needed to sample higher resolutions. Second, CNNs apply roughly the same power across all modes for a given magnification while the linear method has more variance. Perhaps this is because linear solvers cannot leverage higher-frequency spatial information in the image as well as CNNs can. + +![](images/72a42af0d1574fdfa027b525e70c226b1a9810708ea68f1746d5e608a4721528.jpg) +Figure 8. Comparison of modal powers used to induce SR. We show modal power in the Zernike basis to explore the optimized phases (corresponding to the phases shown in Fig. 3). The linear method uses low-order, symmetric modes while the CNN spreads across all modes. Comparing $8 \times$ vs. $2 \times$ magnifications, they take opposite approaches: the CNN prefers low modes at higher resolutions, while the linear uses lower modes at low resolutions. + +Limitations & Future Work: Currently, our work is limited to telescopes with a single-conjugate adaptive optics (SCAO) system, which treats the atmosphere as a single column. Many of today's observatories instead use multiple WFS and DMs. This leads to more accurate corrections across a wider FoV and can even be tuned to different atmospheric layers. Such systems do not preclude the use of our method, since our method can be extended to multiple DMs and potentially enhanced by their extra degrees of freedom. + +Other directions for future work include real-time viewing of video from the science camera by cycling through each phase multiple times to create individually super-resolved frames. Once extended to multiple DMs, additional multiplexing could also be achieved by varying each DM at different rates. + +Most importantly, now that we have demonstrated our methodology with an optical bench prototype, our current priority is to show its effectiveness on sky with a science class telescope and AO system. + +Conclusion: This work shows that science images can encode sufficient MISR information using the AO systems already present in state-of-the-art telescope observatories. We achieve resolution gains without affecting the scientific data in both simulations and optical-bench experiments. With this method, we aim to unlock new scientific capabilities at existing observatories without altering the optical system. + +# Acknowledgements + +RS acknowledges NVIDIA for GPU hardware support. EYHL acknowledges the support of NSERC under CGSD. SS acknowledges the support of the Canada Foundation for Innovation and the Ontario Research Fund. KNK acknowledges the support of NSERC under the RGPIN and RTI programs. + +# References + +[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 126-135, 2017. 5 +[2] Saeed Anwar, Salman Khan, and Nick Barnes. A deep journey into super-resolution: A survey. ACM Computing Surveys (CSUR), 53(3):1-34, 2020. 3, 4 +[3] Roberto J. Avila, Warren J. Hack, and Stsci Team. Astrodrizzle: Aligning images from multiple instruments. In American Astronomical Society, 2012. 4 +[4] Thilo Bauer. Super-resolution imaging: The use case of optical astronomy. In Proceedings of the IADIS International Conference Computer Graphics, Visualization, Computer Vision and Image Processing, Rome, pages 49-59, 2011. 2 +[5] Jacques M Beckers. Adaptive optics for astronomy-principles, performance, and applications. In: Annual Review of Astronomy and Astrophysics. Vol. 31 (A94-12726 02-90), p. 13-62., 31:13-62, 1993. 2 +[6] Anthony Berdeu, Michel Tallon, Éric Thiébaut, and Maud Langlois. Analytical modelling of adaptive optics systems: Role of the influence function. *Astronomy & Astrophysics*, 674:A112, 2023. 5 +[7] Goutam Bhat, Martin Danelljan, Luc Van Gool, and Radu Timofte. Deep burst super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9209-9218, 2021. 4 +[8] David Borncamp, J. M. Anderson, Norman A. Grogin, and Warren J. Hack. Acs/wfc revised geometric distortion for drizzlepac. In Instrument Science Report ACS/WFC, 2015. 4 +[9] Praneeth Chakravarthula, Ethan Tseng, Tarun Srivastava, Henry Fuchs, and Felix Heide. Learned hardware-in-the-loop phase retrieval for holographic near-eye displays. ACM Transactions on Graphics (TOG), 39(6):1-18, 2020. 4 +[10] Dorian Chan, Srinivas Narasimhan, and Matthew O'Toole. Holocurtains: Programming light curtains via binary holography. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 4 +[11] Stanley H Chan, Xiran Wang, and Omar A Elgendy. Plug-and-play admm for image restoration: Fixed-point convergence and applications. IEEE Transactions on Computational Imaging, 3(1):84-98, 2016. 5 +[12] Wenzheng Chen, Parsa Mirdehghan, Sanja Fidler, and Kiriakos N Kutulakos. Auto-tuning structured light by optical stochastic gradient descent. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5970-5980, 2020. 4 +[13] Xu Cheng, Arthur Bradley, Sowmya Ravikumar, and Larry N Thibos. Visual impact of zernike and seidel forms of monochromatic aberrations. Optometry and Vision Science, 87(5):300-312, 2010. 5 +[14] Rodolphe Conan and C Correia. Object-oriented matlab adaptive optics toolbox. In Adaptive Optics Systems IV, page 91486C. International Society for Optics and Photonics, 2014. 6 + +[15] Sony Corporation. Pixel Shift Multi Shooting | SONY — support.d-imaging.sony.co.jp. https://support. d-imaging.sony.co.jp/support/ilm/cpsms/ ilce7rm3/en/index.html, 2025. [Accessed 30-03- 2025]. 4 +[16] Antonia Creswell, Tom White, Vincent Dumoulin, Kai Arulkumaran, Biswa Sengupta, and Anil A Bharath. Generative adversarial networks: An overview. IEEE Signal Processing Magazine, 35(1):53-65, 2018. 5 +[17] Benjamin P Cumming and Min Gu. Direct determination of aberration functions in microscopy by an artificial neural network. Optics Express, 28(10):14511-14521, 2020. 4 +[18] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In Computer Vision - ECCV 2014, pages 184-199, Cham, 2014. Springer International Publishing. 3 +[19] Theodore B DuBose, Dennis F Gardner, and Abbie T Watnik. Intensity-enhanced deep network wavefront reconstruction in shack-hartmann sensors. Optics Letters, 45(7):1699-1702, 2020. 4 +[20] Mark A Ealey and John F Washeba. Continuous facesheet low voltage deformable mirrors. Optical Engineering, 29 (10):1191-1198, 1990. 2 +[21] Linjing Fang, Fred Monroe, Sammy Weiser Novak, Lyndsey Kirk, Cara R Schiavon, Seungyoon B Yu, Tong Zhang, Melissa Wu, Kyle Kastner, Alaa Abdel Latif, et al. Deep learning-based point-scanning super-resolution imaging. Nature Methods, 18(4):406-416, 2021. 4 +[22] Sina Farsiu, Dirk Robinson, Michael Elad, and Peyman Milanfar. Robust shift and add approach to superresolution. In Applications of Digital Image Processing XXVI, pages 121-130. SPIE, 2003. 5 +[23] AS Fruchter and RN Hook. Drizzle: A method for the linear reconstruction of undersampled images. *Publications of the Astronomical Society of the Pacific*, 114(792):144, 2002. 1, 2, 4, 5 +[24] Andrew S. Fruchter. A new method for band-limited imaging with undersampled detectors. *Publications of the Astronomical Society of the Pacific*, 123(902):497, 2011. 4 +[25] Andrea M Ghez, BL Klein, M Morris, and EE Becklin. High proper-motion stars in the vicinity of sagittarius a*: Evidence for a supermassive black hole at the center of our galaxy. The Astrophysical Journal, 509(2):678, 1998. 3 +[26] Roberto Gilmotzzi and Jason Spyromilio. The european extremely large telescope (e-elt). The Messenger, 127(11):3, 2007. 3 +[27] S Gonzaga, W Hack, A Fruchter, and J Mack. The drizzlepac handbook. The DrizzlePac Handbook, 2012. 4 +[28] Karen M Hampson, Raphaël Turcotte, Donald T Miller, Kazuhiro Kurokawa, Jared R Males, Na Ji, and Martin J Booth. Adaptive optics for high-resolution imaging. Nature Reviews Methods Primers, 1(1):1-26, 2021. 3 +[29] James E Harvey and Christ Ftaclas. Diffraction effects of telescope secondary mirror spiders on various image-quality criteria. Applied Optics, 34(28):6337-6349, 1995. 3 +[30] Hamid Reza Vaezi Joze, Ilya Zharkov, Karlton Powell, Carl Ringler, Luming Liang, Andy Roulston, Moshe Lutz, + +and Vivek Pradeep. Imagepairs: Realistic super resolution dataset via beam splitter camera rig. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 518-519, 2020. 5 +[31] SR Kulkarni. Instruments on large optical telescopes-a case study. arXiv preprint arXiv:1606.06674, 2016. 2 +[32] Vasudevan Lakshminarayanan and Andre Fleck. Zernike polynomials: a guide. Journal of Modern Optics, 58(7):545-561, 2011. 5 +[33] Rico Landman, Sebastiaan Y Haffert, Vikram M Radhakrishnan, and Christoph U Keller. Self-optimizing adaptive optics control with reinforcement learning. In Adaptive Optics Systems VII, page 1144849. International Society for Optics and Photonics, 2020. 4 +[34] Tod R Lauer. Combining undersampled dithered images. *Publications of the Astronomical Society of the Pacific*, 111 (756):227, 1999. 2 +[35] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, and Wenzhe Shi. Photo-realistic single image super-resolution using a generative adversarial network, 2017. 3 +[36] Hang Li, Changxiang Yan, and Jianbing Shao. Measurement of the modulation transfer function of infrared imaging system by modified slant edge method. Journal of the Optical Society of Korea, 20(3):381-388, 2016. 8 +[37] Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479:47-59, 2022. 5 +[38] Orly Liba, Kiran Murthy, Yun-Ta Tsai, Tim Brooks, Tianfan Xue, Nikhil Karnad, Qiurui He, Jonathan T Barron, Dillon Sharlet, Ryan Geiss, et al. Handheld mobile photography in very low light. ACM Transactions on Graphics (TOG), 38 (6):1-16, 2019. 4 +[39] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 136-144, 2017. 5 +[40] Zhisheng Lu, Juncheng Li, Hong Liu, Chaoyan Huang, Linlin Zhang, and Tieyong Zeng. Transformer for single image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 457-466, 2022. 5 +[41] Bahattin Can Maral. Single image super-resolution methods: A survey. arXiv preprint arXiv:2202.11763, 2022. 3 +[42] Christian Marois, Bruce Macintosh, Travis Barman, B Zuckerman, Inseok Song, Jennifer Patience, David Lafreniere, and René Doyon. Direct imaging of multiple planets orbiting the star hr 8799. science, 322(5906):1348-1352, 2008. 3 +[43] Y Mellier, Abdurrouf Abdurrouf, JA Acevedo Barroso, A Achucarro, J Adamek, R Adam, GE Addison, N Aghanim, M Aguena, V Ajani, et al. Euclid. i. overview of the euclid mission. *Astronomy & Astrophysics*, 2024. 2 +[44] HH Nasse. How to read mtf curves. Carl Zeiss, Camera Lens Division, page 33, 2008. 7 + +[45] Ngoc Long Nguyen, Jérémy Anger, Axel Davy, Pablo Arias, and Gabriele Facciolo. Self-supervised multi-image superresolution for push-frame satellite images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1121-1131, 2021. 4 +[46] Sylvain Oberti, Carlos Correia, Thierry Fusco, Benoit Neichel, and Pierre Guiraud. Super-resolution wavefront reconstruction. *Astronomy & Astrophysics*, 667:A48, 2022. 4 +[47] Scott W Paine and James R Fienup. Machine learning for improved image-based wavefront sensing. Optics Letters, 43(6):1235-1238, 2018. 4 +[48] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5 +[49] Yifan Peng, Suyeon Choi, Nitish Padmanaban, and Gordon Wetzstein. Neural holography with camera-in-the-loop training. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 4 +[50] Ben C Platt and Roland Shack. History and principles of shack-hartmann wavefront sensing, 2001. 2 +[51] Roberto Ragazzoni, Emiliano Diolaiti, and Elise Vernet. A pyramid wavefront sensor with no dynamic modulation. Optics Communications, 208(1-3):51-60, 2002. 2 +[52] Vincent Sitzmann, Steven Diamond, Yifan Peng, Xiong Dun, Stephen Boyd, Wolfgang Heidrich, Felix Heide, and Gordon Wetzstein. End-to-end optimization of optics and image processing for achromatic extended depth of field and superresolution imaging. ACM Transactions on Graphics (TOG), 37(4):1-13, 2018. 4 +[53] Robin Swanson, Masen Lamb, Carlos M Correia, Suresh Sivanandam, and Kiriakos Kutulakos. Closed loop predictive control of adaptive optics systems with convolutional neural networks. Monthly Notices of the Royal Astronomical Society, 503(2):2944-2954, 2021. 4 +[54] Robin Swanson, Suresh Sivanandam, and Kiriakos N Kutulakos. Adaptive optics mediated sub-pixel super resolution. In Adaptive Optics Systems VIII, pages 2594-2600. SPIE, 2022. 4 +[55] Robin Swanson, Esther Y. H. Lin, Masen Lamb, Suresh Sivanandam, and Kiriakos N. Kutulakos. Super resolved imaging with adaptive optics: Supplemental document. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2025. 7 +[56] Jing Tian and Kai-Kuang Ma. A survey on super-resolution imaging. Signal, Image and Video Processing, 5(3):329-342, 2011. 3 +[57] Ethan Tseng, Shane Colburn, James Whitehead, Luocheng Huang, Seung-Hwan Baek, Arka Majumdar, and Felix Heide. Neural nano-optics for high-quality thin lens imaging. Nature Communications, 12(1):1-7, 2021. 4 +[58] Robert K Tyson. Adaptive ptics engineering handbook. Marcel Dekker New York, 2000. 2 +[59] Robert K Tyson and Benjamin West Frazier. Principles of adaptive optics. CRC press, 2022. 3 + +[60] Congli Wang, Qiang Fu, Xiong Dun, and Wolfgang Heidrich. Megapixel adaptive optics: towards correcting large-scale distortions in computational cameras. ACM Transactions on Graphics (TOG), 37(4):1-12, 2018. 4 +[61] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020. 2 +[62] Bartlomiej Wronski, Ignacio Garcia-Dorado, Manfred Ernst, Damien Kelly, Michael Krainin, Chia-Kai Liang, Marc Levoy, and Peyman Milanfar. Handheld multi-frame superresolution. ACM Transactions on Graphics (TOG), 38(4): 1-18, 2019. 4 +[63] Haitang Yang, Esther Y. H. Lin, Kiriakos N. Kutulakos, and George V. Eleftheriades. Sub-wavelength passive single-shot computational super-oscillatory imaging. Optica, 9(12): 1444–1447, 2022. 4 +[64] Wenming Yang, Xuechen Zhang, Yapeng Tian, Wei Wang, Jing-Hao Xue, and Qingmin Liao. Deep learning for single image super-resolution: A brief review. IEEE Transactions on Multimedia, 21(12):3106-3121, 2019. 3 +[65] Linwei Yue, Huanfeng Shen, Jie Li, Qiangqiang Yuan, Hongyan Zhang, and Liangpei Zhang. Image superresolution: The techniques, applications, and future. Signal Processing, 128:389-408, 2016. 3 +[66] Biwei Zhang, Jiazhu Zhu, Ke Si, and Wei Gong. Deep learning assisted zonal adaptive aberration correction. Frontiers in Physics, page 634, 2021. 4 +[67] Hao Zhu, Mantang Guo, Hongdong Li, Qing Wang, and Antonio Robles-Kelly. Revisiting spatio-angular trade-off in light field cameras and extended applications in superresolution. IEEE Transactions on Visualization and Computer Graphics, 27(6):3019-3033, 2019. 2 \ No newline at end of file diff --git a/2025/Super Resolved Imaging with Adaptive Optics/images.zip b/2025/Super Resolved Imaging with Adaptive Optics/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..faca47953ead98c5b81cc4b6515bb4fa46a2964d --- /dev/null +++ b/2025/Super Resolved Imaging with Adaptive Optics/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ecfa84562e904c6df5d357123c45e61f710c224dac1698d7a775bb0f74310a6b +size 594337 diff --git a/2025/Super Resolved Imaging with Adaptive Optics/layout.json b/2025/Super Resolved Imaging with Adaptive Optics/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9365465e517022758d1d90d134c443d737db83f1 --- /dev/null +++ b/2025/Super Resolved Imaging with Adaptive Optics/layout.json @@ -0,0 +1,9771 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 162, + 103, + 449, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 103, + 449, + 121 + ], + "spans": [ + { + "bbox": [ + 162, + 103, + 449, + 121 + ], + "type": "text", + "content": "Super Resolved Imaging with Adaptive Optics" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 152, + 574, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 152, + 574, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 152, + 574, + 167 + ], + "type": "text", + "content": "Robin Swanson" + }, + { + "bbox": [ + 61, + 152, + 574, + 167 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 61, + 152, + 574, + 167 + ], + "type": "text", + "content": " Esther Y. H. Lin" + }, + { + "bbox": [ + 61, + 152, + 574, + 167 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 152, + 574, + 167 + ], + "type": "text", + "content": " Masen Lamb" + }, + { + "bbox": [ + 61, + 152, + 574, + 167 + ], + "type": "inline_equation", + "content": "^{3,4}" + }, + { + "bbox": [ + 61, + 152, + 574, + 167 + ], + "type": "text", + "content": " Suresh Sivanandam" + }, + { + "bbox": [ + 61, + 152, + 574, + 167 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 61, + 152, + 574, + 167 + ], + "type": "text", + "content": " Kiriakos N. Kutulakos" + }, + { + "bbox": [ + 61, + 152, + 574, + 167 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 178, + 498, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 178, + 498, + 210 + ], + "spans": [ + { + "bbox": [ + 138, + 178, + 498, + 210 + ], + "type": "text", + "content": "1 University of Toronto 2 Dunlap Institute for Astronomy & Astrophysics 3 International Gemini Observatory 4 University of Victoria" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 56, + 219, + 553, + 396 + ], + "blocks": [ + { + "bbox": [ + 56, + 219, + 553, + 396 + ], + "lines": [ + { + "bbox": [ + 56, + 219, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 553, + 396 + ], + "type": "image", + "image_path": "fb9a3cfe2feafcf49849fb82be20353e7180bc7520a4af8f625c8cf9e0434eed.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 403, + 555, + 449 + ], + "lines": [ + { + "bbox": [ + 54, + 403, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 54, + 403, + 555, + 449 + ], + "type": "text", + "content": "Figure 1. Super Resolved Imaging with Adaptive Optics is a computational imaging method to enable multi-image super-resolution in scientific telescopes. Without any additional hardware, existing adaptive mirrors are used to inject diversity into each sub-exposure. When jointly reconstructed, we are able to recover higher spatial resolution from an otherwise undersampled signal. Each exposure's phase profiles is jointly optimized along with the image reconstruction algorithm—resulting in greatly improved performance." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 474, + 200, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 474, + 200, + 487 + ], + "spans": [ + { + "bbox": [ + 151, + 474, + 200, + 487 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 501, + 297, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 501, + 297, + 706 + ], + "spans": [ + { + "bbox": [ + 54, + 501, + 297, + 706 + ], + "type": "text", + "content": "Astronomical telescopes suffer from a tradeoff between field of view (FoV) and image resolution: increasing the FoV leads to an optical field that is under-sampled by the science camera. This work presents a novel computational imaging approach to overcome this tradeoff by leveraging the existing adaptive optics (AO) systems in modern ground-based telescopes. Our key idea is to use the AO system's deformable mirror to apply a series of learned, precisely controlled distortions to the optical wavefront, producing a sequence of images that exhibit distinct, high-frequency, sub-pixel shifts. These images can then be jointly upsampled to yield the final super-resolved image. Crucially, we show this can be done while simultaneously maintaining the core AO operation—correcting for the unknown and rapidly changing wavefront distortions caused by Earth's atmosphere. To achieve this, we incorporate end-to-end optimization of both the induced mirror distortions and the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 475, + 556, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 556, + 596 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 556, + 596 + ], + "type": "text", + "content": "upsampling algorithm, such that telescope-specific optics and temporal statistics of atmospheric wavefront distortions are accounted for. Our experimental results with a hardware prototype, as well as simulations, demonstrate significant SNR improvements of up to 12 dB over non-AO super-resolution baselines, using only existing telescope optics and no hardware modifications. Moreover, by using a precise bench-top replica of a complete telescope and AO system, we show that our methodology can be readily transferred to an operational telescope." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 625, + 395, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 625, + 395, + 639 + ], + "spans": [ + { + "bbox": [ + 314, + 625, + 395, + 639 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 645, + 555, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 645, + 555, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 645, + 555, + 681 + ], + "type": "text", + "content": "Super-resolution (SR) techniques have been integral to astronomical imaging due to inherent telescope design constraints [23]. Telescopes can have many science instruments" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 328, + 693, + 529, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 693, + 529, + 713 + ], + "spans": [ + { + "bbox": [ + 328, + 693, + 529, + 713 + ], + "type": "text", + "content": "Corresponding author: robin@cs.toronto.edu \nProject website: www.cs.toronto.edu/~robin/aosr" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29142" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": "and in certain cases optical resolution is not their priority [67]. A large field of view (FoV), for example, necessitates undersampling to capture a wider view of the night sky. Another common case in photon-starved conditions is to increase signal-to-noise ratio by focusing the light onto fewer, larger, pixels, thereby incurring less read noise [43]. In these, and many other, cases although the telescope is capable of providing a diffraction-limited signal, the science instrument can be undersampled. Thus, the use of SR methods in astronomical imaging has a long history in both satellite and terrestrial telescopes [34]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 212, + 295, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 212, + 295, + 355 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 295, + 355 + ], + "type": "text", + "content": "Unlike telescopes stationed in orbit, ground telescopes face an additional challenge: the Earth's atmosphere [5]. Our atmosphere is a source of dynamic optical aberrations which introduce additional difficulties regardless of a telescope's location, size, or design [58]. As the atmosphere evolves, small changes in index of refraction occur across the sky. This causes incoming light to distort as it transitions from outer space and through our telescope, blurring the images. Therefore, traditional multi-image SR methods which combine multiple sharp images with small geometric distortions—whether from telescope movements [23] or Earth's rotation [4] are not well suited for these conditions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 363, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 295, + 602 + ], + "type": "text", + "content": "We present Super Resolved Imaging with Adaptive Optics, an end-to-end SR method for terrestrial telescopes that leverages existing adaptive optics (AO) systems—designed to correct for atmospheric distortions—to simultaneously correct for aberrations and achieve SR. The crux of our approach is to take advantage of the existing AO system in ground telescopes, which applies a phase mask in the telescope's Fourier domain to counteract atmospheric aberrations. We use the AO system by optimizing and injecting additional phase displacements that produce sub-pixel shifts in the image plane. After capturing several images, these shifts serve as inputs to an end-to-end trained, multi-image SR (MISR) method. These additional phases are optimized to minimally alter the telescope's point spread function (PSF), thus subtly modifying each exposure to best facilitate MISR. The phase displacements can be co-optimized with deep learning-based (or any differentiable) MISR to enhance overall performance. This optimization is performed entirely offline, requiring no modifications to the telescope or use of on-sky time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 610, + 226, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 610, + 226, + 622 + ], + "spans": [ + { + "bbox": [ + 67, + 610, + 226, + 622 + ], + "type": "text", + "content": "In summary, our contributions include:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 65, + 630, + 293, + 699 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 65, + 630, + 293, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 630, + 293, + 663 + ], + "spans": [ + { + "bbox": [ + 65, + 630, + 293, + 663 + ], + "type": "text", + "content": "- A novel MISR method for telescopes that leverages existing hardware, requires no modifications, and does not impact the science payload." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 65, + 666, + 293, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 666, + 293, + 699 + ], + "spans": [ + { + "bbox": [ + 65, + 666, + 293, + 699 + ], + "type": "text", + "content": "- Demonstration that combining AO with MISR outperforms either technique alone, achieving performance greater than the sum of its parts." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 323, + 72, + 552, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 72, + 552, + 129 + ], + "spans": [ + { + "bbox": [ + 323, + 72, + 552, + 129 + ], + "type": "text", + "content": "- Showing that phase displacements and image reconstruction can be co-optimized end-to-end offline using simple telescope telemetry as a proxy for real systems, enabling unobtrusive deployment where telescope time is extremely costly [31]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 323, + 133, + 552, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 133, + 552, + 154 + ], + "spans": [ + { + "bbox": [ + 323, + 133, + 552, + 154 + ], + "type": "text", + "content": "- A real-hardware setup demonstrating the feasibility of integrating our method into existing telescope systems." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 163, + 479, + 177 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 163, + 479, + 177 + ], + "spans": [ + { + "bbox": [ + 314, + 163, + 479, + 177 + ], + "type": "text", + "content": "2. Principles of Adaptive Optics" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 183, + 553, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 183, + 553, + 290 + ], + "spans": [ + { + "bbox": [ + 313, + 183, + 553, + 290 + ], + "type": "text", + "content": "The Earth's atmosphere fundamentally limits the resolving power of any telescope by introducing optical aberrations—one of the reasons why space-based telescopes such as the Hubble Space Telescope have outperformed many larger ground-based observatories. This results in the final telescope PSF being several factors worse than the diffraction limit—the best achievable resolution. As such, AO systems are now widely used in modern ground-based optical and near-infrared telescopes to reverse these effects." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 295, + 474, + 307 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 295, + 474, + 307 + ], + "spans": [ + { + "bbox": [ + 314, + 295, + 474, + 307 + ], + "type": "text", + "content": "2.1. Components of an AO system" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 315, + 553, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 553, + 469 + ], + "type": "text", + "content": "AO systems correct atmospheric turbulence by introducing a phase equal in magnitude but opposite in sign to that of the atmosphere. A prototypical AO setup (Fig. 2) consists of a deformable mirror (DM), wavefront sensor (WFS), and control system. The DM, placed on the Fourier plane of the telescope, takes on the opposite phase of the atmosphere, resulting in corrected wavefronts of light which are then passed on to the science camera and back to the WFS. Because the atmosphere evolves over time, the WFS must continuously measure the residual, uncorrected wavefront, and send updated commands to the DM. This loop, implemented via the control system, typically operates on the order of " + }, + { + "bbox": [ + 313, + 315, + 553, + 469 + ], + "type": "inline_equation", + "content": "1 - 2\\mathrm{kHz}" + }, + { + "bbox": [ + 313, + 315, + 553, + 469 + ], + "type": "text", + "content": " for modern AO systems." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 478, + 553, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 553, + 621 + ], + "type": "text", + "content": "The most common types of WFSs are based on Shack-Hartmann [50] and pyramid [51] designs. To operate they require a \"guide-star\"—a distant light source which can be approximated as a point. Usually a bright star near the science object is used, but laser beacons can also be used to create an artificial star image in the upper atmosphere. The guide star samples a similar turbulence path as the science image without reducing its signal-to-noise ratio (SNR). By comparing the measured signal to a reference, we can calculate how the atmosphere has affected the wavefront across the telescope pupil. Pre-calibrated responses are then used to update the DM shape and negate the distortions [58]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 629, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 553, + 712 + ], + "type": "text", + "content": "The DM can be constructed in several ways. Many of today's telescopes use a continuous design with a single flexible reflective membrane supported by a grid of discrete actuators [20]. These actuators can be individually pushed or pulled to collectively generate complex phase patterns. In contrast, many upcoming AO systems use grids of individual mirrors, enabling much larger DMs and corrections for" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29143" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 203, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 203, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 203, + 84 + ], + "type": "text", + "content": "even higher spatial frequencies [26]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 92, + 296, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 92, + 296, + 189 + ], + "spans": [ + { + "bbox": [ + 55, + 92, + 296, + 189 + ], + "type": "text", + "content": "Once converged, the effects of the AO loop, shown in Fig. 2, are visible and have enabled many recent breakthroughs in modern astronomy. Direct imaging of exoplanets [42] and the first direct evidence of a super-massive black hole within our galaxy [25] would not have been possible without the increased resolution and contrast provided by AO. For additional details on how AO has been recently used in both astronomy and microscopy see [28]." + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 56, + 196, + 296, + 395 + ], + "blocks": [ + { + "bbox": [ + 56, + 196, + 296, + 395 + ], + "lines": [ + { + "bbox": [ + 56, + 196, + 296, + 395 + ], + "spans": [ + { + "bbox": [ + 56, + 196, + 296, + 395 + ], + "type": "image", + "image_path": "29c1c3ffd47a923f8b0c3c045ab02c18de25cecbd978d2fd9516f1cb75198b51.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 401, + 296, + 521 + ], + "lines": [ + { + "bbox": [ + 55, + 401, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 296, + 521 + ], + "type": "text", + "content": "Figure 2. Starlight is aberrated by Earth's atmosphere " + }, + { + "bbox": [ + 55, + 401, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 401, + 296, + 521 + ], + "type": "text", + "content": ", blurring the telescope PSF. An AO system (a) measures the atmosphere with a wavefront sensor and applies a corrective phase, " + }, + { + "bbox": [ + 55, + 401, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 401, + 296, + 521 + ], + "type": "text", + "content": ", on its deformable mirror to cancel it out. When AO is off, resolution is directly limited by the seeing conditions (i.e., turbulence) at the observatory for both short and long exposures (b). When AO is enabled the effects of the atmosphere are removed. If sampled at or above the Nyquist rate, the resolution is limited only by the mirror diameter and its PSF forms a diffraction-limited Airy disk (c). Our method enables under-sampled instruments with AO to recover spatial resolution." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 530, + 233, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 530, + 233, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 233, + 544 + ], + "type": "text", + "content": "2.2. The AO Image Formation Model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 550, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 296, + 694 + ], + "type": "text", + "content": "Suppose we are observing a distant, constant object (e.g., a natural guide star) over an exposure time " + }, + { + "bbox": [ + 55, + 550, + 296, + 694 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 550, + 296, + 694 + ], + "type": "text", + "content": " using a telescope without AO. The recorded image " + }, + { + "bbox": [ + 55, + 550, + 296, + 694 + ], + "type": "inline_equation", + "content": "I(x,y)" + }, + { + "bbox": [ + 55, + 550, + 296, + 694 + ], + "type": "text", + "content": " in pixel coordinates " + }, + { + "bbox": [ + 55, + 550, + 296, + 694 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 55, + 550, + 296, + 694 + ], + "type": "text", + "content": " is modeled as the convolution of the object " + }, + { + "bbox": [ + 55, + 550, + 296, + 694 + ], + "type": "inline_equation", + "content": "O(x,y)" + }, + { + "bbox": [ + 55, + 550, + 296, + 694 + ], + "type": "text", + "content": " with an atmosphere-degraded PSF " + }, + { + "bbox": [ + 55, + 550, + 296, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{TEL}}" + }, + { + "bbox": [ + 55, + 550, + 296, + 694 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 55, + 550, + 296, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{TEL}}" + }, + { + "bbox": [ + 55, + 550, + 296, + 694 + ], + "type": "text", + "content": " corresponds to a diffraction-limited PSF " + }, + { + "bbox": [ + 55, + 550, + 296, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{IDEAL}}" + }, + { + "bbox": [ + 55, + 550, + 296, + 694 + ], + "type": "text", + "content": " (i.e., an Airy Disk [59]) degraded by the time-varying turbulence of the Earth's atmosphere. In Fourier space, this turbulence can be represented by a phase distribution " + }, + { + "bbox": [ + 55, + 550, + 296, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(u,v,t)" + }, + { + "bbox": [ + 55, + 550, + 296, + 694 + ], + "type": "text", + "content": " with unit amplitude where " + }, + { + "bbox": [ + 55, + 550, + 296, + 694 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 55, + 550, + 296, + 694 + ], + "type": "text", + "content": " are its spatial frequency coordinates. By the convolution theorem, this can be described in Fourier space with a point-wise multiplication:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 68, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {F} \\left\\{\\mathcal {P} _ {\\mathrm {T E L}} \\right\\} (u, v, t) = e ^ {j \\mathcal {A} (u, v, t)} \\mathcal {F} \\left\\{\\mathcal {P} _ {\\text {I D E A L}} \\right\\} (u, v, t), \\tag {1}", + "image_path": "9d980cb225ade6ee3de2594844497ea3b6225f6dca3d1c14d186e94f15ae812a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{F}\\{\\cdot\\}" + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": " denotes the Fourier Transform operator. While the AO is in operation, a DM continuously applies an additional phase distortion " + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(u,v,t)" + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": " in the Fourier plane of the telescope to counteract the effects of the atmosphere. The resulting corrected PSF, " + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{CORR}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": ", can be described in the Fourier domain as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 151, + 556, + 177 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 151, + 556, + 177 + ], + "spans": [ + { + "bbox": [ + 313, + 151, + 556, + 177 + ], + "type": "interline_equation", + "content": "\\mathcal {F} \\left\\{\\mathcal {P} _ {\\text {C O R R}} \\right\\} (u, v, t) = e ^ {j \\mathcal {D} (u, v, t)} e ^ {j \\mathcal {A} (u, v, t)} \\mathcal {F} \\left\\{\\mathcal {P} _ {\\text {I D E A L}} \\right\\} (u, v, t). \\tag {2}", + "image_path": "ae98759707c37685075a683b3397d5d922b8a40469d92152f5e2bb01bcf4c12e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 177, + 554, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 177, + 554, + 214 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 554, + 214 + ], + "type": "text", + "content": "In an ideal scenario, the atmospheric phase " + }, + { + "bbox": [ + 313, + 177, + 554, + 214 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 177, + 554, + 214 + ], + "type": "text", + "content": " and the deformable mirror phase " + }, + { + "bbox": [ + 313, + 177, + 554, + 214 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 177, + 554, + 214 + ], + "type": "text", + "content": " would perfectly cancel at every point and time; i.e., " + }, + { + "bbox": [ + 313, + 177, + 554, + 214 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{CORR}} \\approx \\mathcal{P}_{\\mathrm{IDEAL}}" + }, + { + "bbox": [ + 313, + 177, + 554, + 214 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 221, + 555, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 221, + 555, + 317 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 555, + 317 + ], + "type": "text", + "content": "In practice the DM is limited by fitting errors and the finite range of its actuators which restricts the range of spatial frequencies it can correct. Additionally, the effects of non-circular mirror geometries [29] (e.g., hexagonal apertures) and mirror support structures cause the DM to deviate from ideal performance. Consequently, the final effective PSF of the telescope never truly reaches " + }, + { + "bbox": [ + 313, + 221, + 555, + 317 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{IDEAL}}" + }, + { + "bbox": [ + 313, + 221, + 555, + 317 + ], + "type": "text", + "content": ". We model these effects on " + }, + { + "bbox": [ + 313, + 221, + 555, + 317 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{EFF}}" + }, + { + "bbox": [ + 313, + 221, + 555, + 317 + ], + "type": "text", + "content": " in the Fourier domain as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 330, + 324, + 555, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 324, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 330, + 324, + 555, + 340 + ], + "type": "interline_equation", + "content": "\\mathcal {F} \\left\\{\\mathcal {P} _ {\\text {E F F}} \\right\\} (u, v, t) = e ^ {j \\delta (u, v, t)} \\mathcal {F} \\left\\{\\mathcal {P} _ {\\text {C O R R}} \\right\\} (u, v, t), \\tag {3}", + "image_path": "9629d7d588e4609e0e7ec59a789f8ca3385c9494baba22e268ea1ff65fa2e065.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 347, + 554, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 554, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 554, + 370 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 347, + 554, + 370 + ], + "type": "inline_equation", + "content": "\\delta(u, v, t)" + }, + { + "bbox": [ + 313, + 347, + 554, + 370 + ], + "type": "text", + "content": " encapsulates the hardware and optical constraints described above." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 380, + 554, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 554, + 475 + ], + "type": "text", + "content": "Note that these contributions are predominantly static and can be approximated as " + }, + { + "bbox": [ + 313, + 380, + 554, + 475 + ], + "type": "inline_equation", + "content": "\\delta(u,v,t) \\approx \\delta(u,v)" + }, + { + "bbox": [ + 313, + 380, + 554, + 475 + ], + "type": "text", + "content": ". For a well-performing, diffraction-limited AO system, variations will be short because the AO re-converges quickly. Therefore, over long exposures, they average out over time. Finally, we consider incoherent image formation, as astronomical sources emit light independently with random phases. Our image formation becomes" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 323, + 483, + 554, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 483, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 323, + 483, + 554, + 510 + ], + "type": "interline_equation", + "content": "I (x, y) = \\int_ {0} ^ {T} | O (x, y) | ^ {2} * \\left| \\mathcal {P} _ {\\text {E F F}} (x, y, t) \\right| ^ {2} \\mathrm {d} t + \\eta \\tag {4}", + "image_path": "d94ce5c95c4ea4cbe4c53fe43de86e84728003b7ba78df9d1025ec6dd9e6323e.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 517, + 554, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 554, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 554, + 541 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 517, + 554, + 541 + ], + "type": "inline_equation", + "content": "\\ast" + }, + { + "bbox": [ + 313, + 517, + 554, + 541 + ], + "type": "text", + "content": " denotes the convolution operator, " + }, + { + "bbox": [ + 313, + 517, + 554, + 541 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 517, + 554, + 541 + ], + "type": "text", + "content": " is the total exposure time, and " + }, + { + "bbox": [ + 313, + 517, + 554, + 541 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 517, + 554, + 541 + ], + "type": "text", + "content": " is cumulative noise." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 550, + 406, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 406, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 406, + 562 + ], + "type": "text", + "content": "3. Previous Work" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "Image Super-Resolution (SR). SR techniques aim to reconstruct a high-resolution (HR) image from one or more observed low-resolution (LR) images [65]. However, the problem is ill-posed in nature, and many plausible solutions exist [41]. SR methods are distinguished between single-image (SISR) and multi-image (MISR) super-resolution depending on the number of images used. Deep-learning SR has emerged as the state-of-the-art for both SISR and MISR tasks [2, 41, 41, 56, 64], thanks to the growth of convolutional networks and large datasets [18, 35]. However, deep SR networks thrive on natural images but struggle with out-of-distribution data due to limited training data [2]. In" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29144" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "such cases, incorporating priors based on sensor, object, and acquisition conditions can improve performance [2]. Astronomical imaging is a regime with unique conditions, including extreme low-light, non-static aberrations, and image features unlike those typically seen in terrestrial images. Hence, additional modifications to existing deep-learning SR techniques are required." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 163, + 298, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 298, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 298, + 380 + ], + "type": "text", + "content": "SR in Astronomical Imaging: Drizzle—a SR technique commonly used in astronomical imaging and astrophotography—aligns and linearly blends multiple deliberately dithered frames to produce a super-resolved image [23]. The technique has been widely adopted, including by the Hubble Space Telescope, and remains in use today [27]. However, the success of drizzling and its contemporaries hinges on highly accurate image registration and precise knowledge of the offsets [3, 8], a challenging task when the atmosphere adds time-varying aberrations and distortions [24]. In contrast, our method leverages the telescope's built-in AO system to inject optimized phasedomain \"jitters\" and does not use any type of image registration. SR techniques have also been leveraged to improve the resolution and quality of WFS measurements [46]. Our method, building on our previous work [54], is agnostic to the underlying WFS type and can remain applicable as WFS technology advances." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 387, + 296, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 387, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 387, + 296, + 591 + ], + "type": "text", + "content": "SR in Computational Imaging: Camera system design often incorporates super-resolution. One early example jointly optimized both the optics and SISR algorithms for simple, refractive lenses [52]. Burst photography from handheld cameras—where small distortions are induced by hand shake—are extremely successful at recovering high-resolution information [7, 62] even in low-light conditions [38]. Deep learning methods have also been explored in the scientific imaging community for both microscopy [21] and satellite imaging [45]. MISR has also long played a role in classic camera development; many digital cameras are able to programmatically shift the imaging sensor between subsequent images which can then be combined into one higher resolution image (e.g., Sony's Pixel Shift Multi Shooting [15]). This work draws inspiration from previous works by incorporating end-to-end optimization of telescope mirror distortions with SR algorithms." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 598, + 294, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 598, + 294, + 706 + ], + "spans": [ + { + "bbox": [ + 55, + 598, + 294, + 706 + ], + "type": "text", + "content": "Computational Wavefront Sensing and Control: CNNs have been used to improve WFS reconstruction [19, 47], as well as to predict atmospheric conditions to reduce servolag in the system [33, 53]. Similar techniques have shown to be effective in microscopy AO systems as well [17, 66]. More general-purpose methods have been proposed such as [60] which used coded wavefront sensors and spatial light modulators (SLMs) to achieve much higher order spatial frequency correction than is capable from typical AO sys" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "content": "tems in astronomy or microscopy. However, SLMs, unlike DMs, may not be suitable for correcting rapidly changing atmospheric conditions due to their slower speeds [10]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 116, + 555, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 116, + 555, + 272 + ], + "spans": [ + { + "bbox": [ + 313, + 116, + 555, + 272 + ], + "type": "text", + "content": "Optical Phase and PSF Optimization: End-to-end optimization of optical systems and their image reconstruction algorithms have been used in thin lens design [57], holography [9, 49], and active illumination [12]. These methods typically focus on optimization of critical optical elements of the system, requiring either direct access to the optical system [63] or robust, differentiable simulations during optimization. Due to the complexity and operating costs of modern telescopes, modification to the underlying system or real-time access for network training would be infeasible. Instead, we focus on optimization of optical elements available to us in such a way that would not affect its operation, allowing us to side step these limitations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 280, + 499, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 280, + 499, + 295 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 499, + 295 + ], + "type": "text", + "content": "4. Super Resolving Adaptive Optics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 300, + 554, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 554, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 554, + 361 + ], + "type": "text", + "content": "This work augments existing MISR methods by learning the optimal way to enhance each individual image captured over a single exposure. Using the DM already present in a telescope's AO system, we inject optimized phase profiles, thereby enabling multi-image super-resolution." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 368, + 554, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 554, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 554, + 463 + ], + "type": "text", + "content": "In this section, we present an end-to-end approach to determine optimal phase profiles while accounting for AO hardware constraints and atmospheric effects on final images. We discuss our approach's assumptions (Sec. 4.1), the framework for optimizing phase profiles (Sec. 4.2), compatibility with various image reconstruction methods for combining sub-exposures (Sec. 4.3), and the training details for obtaining the optimal phase profiles (Sec. 4.3)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 468, + 511, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 511, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 511, + 481 + ], + "type": "text", + "content": "4.1. The \"Small Phase Shift\" Assumption" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 485, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 554, + 628 + ], + "type": "text", + "content": "To include the image formation model in our learning infrastructure we must make some reasonable assumptions about the process. Importantly, we assume that sub-pixel shifts negligibly impact the AO loop or science image. In our case this is reasonable due to the fact that we consider only under-sampled systems. If our phase shifts keep " + }, + { + "bbox": [ + 313, + 485, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{TEL}}" + }, + { + "bbox": [ + 313, + 485, + 554, + 628 + ], + "type": "text", + "content": " within the range of its under-sampled area, the aggregated effect over the entire exposure will be negligible1. The shifts define the DM \"flat\" (a standard feature of AO real-time controllers), and can be updated easily between exposures. Therefore, the AO system will run entirely as usual—unaware of the phase offsets applied by our method." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 633, + 491, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 633, + 491, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 633, + 491, + 647 + ], + "type": "text", + "content": "4.2. Phase Shifting and Optimization" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 650, + 553, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 650, + 553, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 650, + 553, + 687 + ], + "type": "text", + "content": "Because our DM is designed to create highly complex shapes, we are not restricted to planar movements like traditional methods [23]. We describe a given DM shape " + }, + { + "bbox": [ + 313, + 650, + 553, + 687 + ], + "type": "inline_equation", + "content": "S" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "type": "text", + "content": "1If necessary, we can also impose constraints on the magnitude of our learned phase shifts to guarantee science performance" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29145" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "using either actuator positions or, as detailed below, a more general modal basis." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 104, + 296, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 104, + 296, + 272 + ], + "spans": [ + { + "bbox": [ + 55, + 104, + 296, + 272 + ], + "type": "text", + "content": "Specifically, one approach would be to use a combination of Zernike polynomial modes, " + }, + { + "bbox": [ + 55, + 104, + 296, + 272 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 55, + 104, + 296, + 272 + ], + "type": "text", + "content": ". The Zernike modes are well suited for this problem because of their ability to represent rich phases (including planar shifts) with minimal coefficients and are well-suited for circular apertures, such as those of telescopes [32]. They can also provide some insight into the nature of a given phase due to the relationship between Zernike modes and classical optics aberrations (e.g., focus, coma, and astigmatism [13]). Besides planar shifts (tip and tilt Zernike modes), there is no trivial or well known choice of Zernike modes for super-resolution available. We therefore optimize their values in conjunction with the MISR algorithm of choice to estimate the best possible choice of modes (i.e., an end-to-end optimization)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 279, + 296, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 296, + 316 + ], + "type": "text", + "content": "We can now update our image formation model from a single exposure " + }, + { + "bbox": [ + 55, + 279, + 296, + 316 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 55, + 279, + 296, + 316 + ], + "type": "text", + "content": " of time " + }, + { + "bbox": [ + 55, + 279, + 296, + 316 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 279, + 296, + 316 + ], + "type": "text", + "content": " to include " + }, + { + "bbox": [ + 55, + 279, + 296, + 316 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 279, + 296, + 316 + ], + "type": "text", + "content": " exposures of time " + }, + { + "bbox": [ + 55, + 279, + 296, + 316 + ], + "type": "inline_equation", + "content": "T / N" + }, + { + "bbox": [ + 55, + 279, + 296, + 316 + ], + "type": "text", + "content": " with individually estimable phases " + }, + { + "bbox": [ + 55, + 279, + 296, + 316 + ], + "type": "inline_equation", + "content": "S_{i = 1\\dots N}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 323, + 295, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 323, + 295, + 338 + ], + "spans": [ + { + "bbox": [ + 113, + 323, + 295, + 338 + ], + "type": "interline_equation", + "content": "I _ {i} = | O | ^ {2} \\circledast | \\mathcal {P} _ {i} \\circledast e ^ {j S _ {i}} | ^ {2} + \\eta . \\tag {5}", + "image_path": "0ea2e00bd28d8e6c03172014202732265d18124dee7dde443a9e3fe95ec8d69b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "type": "text", + "content": "This same model can be equally applied to a simple shifting or Zernike model by replacing " + }, + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "type": "text", + "content": " and choosing an appropriate number of Zernike coefficients." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 389, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 437 + ], + "type": "text", + "content": "Now, let " + }, + { + "bbox": [ + 55, + 389, + 296, + 437 + ], + "type": "inline_equation", + "content": "f_{\\uparrow}(x \\mid \\theta)" + }, + { + "bbox": [ + 55, + 389, + 296, + 437 + ], + "type": "text", + "content": " be a learned MISR function with parameters " + }, + { + "bbox": [ + 55, + 389, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 389, + 296, + 437 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 389, + 296, + 437 + ], + "type": "inline_equation", + "content": "f_{\\Downarrow}(x)" + }, + { + "bbox": [ + 55, + 389, + 296, + 437 + ], + "type": "text", + "content": " be a downsampling method which reduces an image's spatial dimensions by a fixed factor. Our end-to-end optical optimization problem is then," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 445, + 298, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 298, + 478 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 298, + 478 + ], + "type": "interline_equation", + "content": "\\min _ {\\{\\theta , \\mathcal {S} \\}} \\left| \\right.\\left| \\right. f _ {\\uparrow} \\left( \\right.f _ {\\downarrow} \\left( \\right.| O | ^ {2} * \\mid \\mathcal {P} _ {1 \\dots N} * e ^ {j \\mathcal {S} _ {1 \\dots N}} \\left. \\right| ^ {2} + \\eta \\mid \\theta\\left. \\right)\\left. \\right) - O \\left. \\right|\\left. \\right| _ {1}, \\tag {6}", + "image_path": "947da1cadcf87fd5508747f70b8a7baa6f300a13862d1352097e5f91b29cd096.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 478, + 296, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 296, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 296, + 561 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 478, + 296, + 561 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 55, + 478, + 296, + 561 + ], + "type": "text", + "content": " is an image randomly sampled from a set of high-resolution images, " + }, + { + "bbox": [ + 55, + 478, + 296, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 55, + 478, + 296, + 561 + ], + "type": "text", + "content": " is randomly chosen from a set of precomputed AO telescope PSFs, " + }, + { + "bbox": [ + 55, + 478, + 296, + 561 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 55, + 478, + 296, + 561 + ], + "type": "text", + "content": " are our trainable phase profiles, and " + }, + { + "bbox": [ + 55, + 478, + 296, + 561 + ], + "type": "inline_equation", + "content": "||\\cdot ||_1" + }, + { + "bbox": [ + 55, + 478, + 296, + 561 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 55, + 478, + 296, + 561 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 55, + 478, + 296, + 561 + ], + "type": "text", + "content": " norm. As described earlier, " + }, + { + "bbox": [ + 55, + 478, + 296, + 561 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 478, + 296, + 561 + ], + "type": "text", + "content": " can be represented in many ways. To explore its effects, we compare the most common methods in our experiments. These include:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "text", + "content": "- No Learned Phase: The base case for CNN methods. The same number of exposures are given to the CNN during training—and still affected by " + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "text", + "content": "—but without optimizing any per-exposure " + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "inline_equation", + "content": "S_i" + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 618, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 688 + ], + "type": "text", + "content": "- Tip/Tilt: A common MISR technique where the image is translated by a sub-pixel shift [22]. The Classic case uses a sequence of exposures whose translational shifts form a regular sub-pixel raster scan. The Optimized variant initializes with the same shifts as Classic and then optimizes the values of those shifts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "- Zernike Modes: Each phase is represented using the Zernike basis and its coefficients are optimized." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 554, + 133 + ], + "type": "text", + "content": "- Non-Modal: Instead of a basis function, the DM shape is directly optimized. This requires precise knowledge of its influence functions and calibration with respect to the WFS [6]. While simple in simulation, this proves to be difficult to use in practice." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 137, + 536, + 150 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 137, + 536, + 150 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 536, + 150 + ], + "type": "text", + "content": "4.3. Super-Resolution Reconstruction Methods" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 155, + 554, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 155, + 554, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 155, + 554, + 190 + ], + "type": "text", + "content": "Although our method could work equally well with any super-resolution reconstruction method, for the purposes of this paper we focus on only two:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 198, + 554, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 554, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 554, + 258 + ], + "type": "text", + "content": "First, we consider a linear-reconstruction method similar to Drizzle [23], also known as the shift-and-add method [22]. This may be of interest in science cases where such methods are already widely used, albeit for systems without adaptive optics and with un-optimized phases." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 266, + 555, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 555, + 398 + ], + "type": "text", + "content": "Second, we consider a state of the art deep learning-based SR algorithm. Specifically, we based our model on the widely known EDSR [39] single-image SR CNN. Our network was modified to incorporate the forward image formation model (Eq. (4)) as well as to accept any number of input channels, representing the phase-shifted exposures. The EDSR model was chosen for its speed, widespread adoption, and its avoidance of adversarial training (i.e., it does not rely on a GAN [16]) which can complicate training and introduce priors that may not be appropriate for scientific purposes2." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 403, + 507, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 507, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 507, + 417 + ], + "type": "text", + "content": "4.4. Phase and Network Training Details" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 420, + 554, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 554, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 554, + 480 + ], + "type": "text", + "content": "In addition to a training set of high-resolution images, our phase and upsampling parameters require a collection of system PSFs under the influence of the AO system and the expected turbulence. These can be generated by simulations of the system or collected from telescope telemetry." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 488, + 554, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 488, + 554, + 596 + ], + "spans": [ + { + "bbox": [ + 313, + 488, + 554, + 596 + ], + "type": "text", + "content": "At training time, for a model expecting " + }, + { + "bbox": [ + 313, + 488, + 554, + 596 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 488, + 554, + 596 + ], + "type": "text", + "content": " exposures, we provide the network a random batch of high-resolution images as well as a random sample of " + }, + { + "bbox": [ + 313, + 488, + 554, + 596 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 488, + 554, + 596 + ], + "type": "text", + "content": " system PSFs. The low-resolution images are generated for each PSF, as detailed in Eq. (5), using a trainable phase profile for each subexposure. We now have low-resolution images affected by the system PSF in addition to our optimized phases. We also have the ground-truth high-resolution image, which can be used to compute the loss function in Eq. (6)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 604, + 554, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 554, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 554, + 676 + ], + "type": "text", + "content": "Results shown in this work were trained on a combination of the DIV2K [1] and Microsoft ImagePairs [30] datasets. Both CNN and linear reconstructors were implemented and trained with pyTorch [48] following the learning procedure described in the EDSR paper [39]. Unless stated elsewhere (e.g., Sec. 5.1), the following experi" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 683, + 554, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 683, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 683, + 554, + 713 + ], + "type": "text", + "content": "2Our outlined procedure can be easily adapted to any SR method including diffusion [37] or transformer [40] models, or traditional convex optimization methods [11]." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29146" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "iments are all performed with " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "N = 4" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " exposures. While our method can be used with any number of exposures, this value was chosen to easily compare to the classic \"tip/tilt\" methods when solving for a " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " upscaling factor. Restricting the number of exposures to four highlights the performance differences between linear and deep learning-based reconstruction methods, as discussed in Secs. 5 to 7." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 167, + 163, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 167, + 163, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 163, + 179 + ], + "type": "text", + "content": "5. Simulated Results" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 186, + 296, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 186, + 296, + 294 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 296, + 294 + ], + "type": "text", + "content": "We first evaluate our method entirely in simulation. We train SR parameters and phase profiles solely on synthetic PSFs generated by the discretized image-formation model in Eq. (5). After training, we run a new AO simulation, setting the optimized phase profiles as the deformable mirror's \"flat\" position. This imposes the optimized phase on the system, replicating our convolution in the simulated hardware. The PSF generated from this new simulation is then used as the PSF for evaluating test images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 302, + 212, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 302, + 212, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 302, + 212, + 313 + ], + "type": "text", + "content": "5.1. AO Simulation Environment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 318, + 296, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 296, + 474 + ], + "type": "text", + "content": "To create a realistic AO simulation we use the OOMAO [14] toolbox. We implemented a realistic AO simulation with multiple layers of atmosphere to generate realistic PSFs for training. Our simulated AO system was set up to match our experimental hardware (detailed in Sec. 6) with an " + }, + { + "bbox": [ + 55, + 318, + 296, + 474 + ], + "type": "inline_equation", + "content": "8\\mathrm{m}" + }, + { + "bbox": [ + 55, + 318, + 296, + 474 + ], + "type": "text", + "content": " diameter telescope and 397-actuator deformable mirror operating at " + }, + { + "bbox": [ + 55, + 318, + 296, + 474 + ], + "type": "inline_equation", + "content": "1000\\mathrm{Hz}" + }, + { + "bbox": [ + 55, + 318, + 296, + 474 + ], + "type": "text", + "content": ". Each PSF was generated by performing a full AO simulation with an evolving atmosphere over a 2-second exposure. The simulated " + }, + { + "bbox": [ + 55, + 318, + 296, + 474 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{TEL}}" + }, + { + "bbox": [ + 55, + 318, + 296, + 474 + ], + "type": "text", + "content": " was therefore imaged on a simulated science camera under the effects of AO correction and realistic sensor noise. To cover a wide range of conditions, the wind direction, wind speed, and guide star magnitude were randomized for each simulation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 482, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 482, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 296, + 567 + ], + "type": "text", + "content": "Effect of Joint Optimization: We investigate how reconstruction results are affected by jointly optimizing phase profiles and the CNN reconstruction method. As shown in Fig. 4, the jointly-optimized approach yields better resolution enhancement than a naive approach that trains the CNN without phase optimization, both qualitatively and quantitatively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 574, + 296, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 296, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 296, + 670 + ], + "type": "text", + "content": "Effect of Super-Resolution Factor: We compare different reconstruction methods and phase representations for the case where " + }, + { + "bbox": [ + 55, + 574, + 296, + 670 + ], + "type": "inline_equation", + "content": "N = 4" + }, + { + "bbox": [ + 55, + 574, + 296, + 670 + ], + "type": "text", + "content": " exposures. We compare our different phase representations against a baseline CNN case where the input to the neural network are four fixed unshifted PSFs with no estimizable phase and simple bilinear interpolation. Experimental results are shown in Sec. 5.1. Sample phase profiles and output can be found in Fig. 3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Effect of Number of Exposures: We consider the effect of number of exposures on each method, freezing the super-resolution factor. The results are shown in Table 1." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 334, + 69, + 535, + 281 + ], + "blocks": [ + { + "bbox": [ + 334, + 69, + 535, + 281 + ], + "lines": [ + { + "bbox": [ + 334, + 69, + 535, + 281 + ], + "spans": [ + { + "bbox": [ + 334, + 69, + 535, + 281 + ], + "type": "table", + "html": "
MethodScale Factor
2×4×8×
CNNNo Learned Phase33.7227.9824.04
Tip/Tilt (Classic)39.5633.4625.47
Tip/Tilt (Optimized)39.6733.8826.32
Zernike Modes41.5134.0927.74
Non-Modal43.1334.2427.79
LinearTip/Tilt (Classic)28.5626.0423.32
Tip/Tilt (Optimized)29.1126.1223.26
Zernike Modes30.8426.1723.36
Non-Modal31.0826.6723.33
MethodNumber of Exposures
2468
CNNTip/Tilt (Classic)30.7733.8936.0336.31
Zernike Modes31.1334.1136.2937.44
Non-Modal31.1734.2436.2437.48
LinearTip/Tilt26.1126.1226.1226.11
Zernike Modes26.1426.1726.1626.17
Non-Modal26.6626.6726.7126.69
", + "image_path": "079f406041aafa267b210c2cff954d0bc7569e1451891d2e9c52bf3cbc5a5376.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 290, + 555, + 335 + ], + "lines": [ + { + "bbox": [ + 313, + 290, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 555, + 335 + ], + "type": "text", + "content": "Table 1. Effect of number of exposures and super-resolution factor on SR methods. Values presented in PSNR (dB) " + }, + { + "bbox": [ + 313, + 290, + 555, + 335 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 313, + 290, + 555, + 335 + ], + "type": "text", + "content": ". In general the more exposures, and the more expressive the phase basis, the better the results. Details on each method can be found in Sec. 4.2." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 357, + 440, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 440, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 440, + 370 + ], + "type": "text", + "content": "6. Experimental Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 376, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 568 + ], + "type": "text", + "content": "To demonstrate our method's ability to work in realistic settings, we also implemented the experimental setup described in Sec. 5 on an optical table. Our setup, shown in Fig. 5, consists of a 97-actuator ALPAO deformable mirror and a custom-built " + }, + { + "bbox": [ + 313, + 376, + 555, + 568 + ], + "type": "inline_equation", + "content": "25 \\times 25" + }, + { + "bbox": [ + 313, + 376, + 555, + 568 + ], + "type": "text", + "content": " sub-aperture Shack-Hartmann wavefront sensor. Our guide star was created with an " + }, + { + "bbox": [ + 313, + 376, + 555, + 568 + ], + "type": "inline_equation", + "content": "850\\mathrm{~nm}" + }, + { + "bbox": [ + 313, + 376, + 555, + 568 + ], + "type": "text", + "content": " near-infrared laser whose pupil and beam size were matched to the DM's size " + }, + { + "bbox": [ + 313, + 376, + 555, + 568 + ], + "type": "inline_equation", + "content": "(13.5\\mathrm{~mm})" + }, + { + "bbox": [ + 313, + 376, + 555, + 568 + ], + "type": "text", + "content": ". Realistic atmospheric turbulence was induced by a rotating Lexitek phase screen driven by a stepping motor. The sensor's native pixel pitch was " + }, + { + "bbox": [ + 313, + 376, + 555, + 568 + ], + "type": "inline_equation", + "content": "5.86\\mu \\mathrm{m}" + }, + { + "bbox": [ + 313, + 376, + 555, + 568 + ], + "type": "text", + "content": " and the diffraction-limited spot was " + }, + { + "bbox": [ + 313, + 376, + 555, + 568 + ], + "type": "inline_equation", + "content": "22.95\\mu \\mathrm{m}" + }, + { + "bbox": [ + 313, + 376, + 555, + 568 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 376, + 555, + 568 + ], + "type": "inline_equation", + "content": "f = 400\\mathrm{~mm}" + }, + { + "bbox": [ + 313, + 376, + 555, + 568 + ], + "type": "text", + "content": ", pupil size " + }, + { + "bbox": [ + 313, + 376, + 555, + 568 + ], + "type": "inline_equation", + "content": "13.5\\mathrm{~mm}" + }, + { + "bbox": [ + 313, + 376, + 555, + 568 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 376, + 555, + 568 + ], + "type": "inline_equation", + "content": "\\lambda = 635\\mathrm{~nm}" + }, + { + "bbox": [ + 313, + 376, + 555, + 568 + ], + "type": "text", + "content": "). Note that even though our experimental hardware was not undersampled by design, we used " + }, + { + "bbox": [ + 313, + 376, + 555, + 568 + ], + "type": "inline_equation", + "content": "12 \\times" + }, + { + "bbox": [ + 313, + 376, + 555, + 568 + ], + "type": "text", + "content": " on-sensor binning to capture individual images. This resulted in an effective pixel pitch of " + }, + { + "bbox": [ + 313, + 376, + 555, + 568 + ], + "type": "inline_equation", + "content": "70.3\\mu \\mathrm{m}" + }, + { + "bbox": [ + 313, + 376, + 555, + 568 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 575, + 556, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 556, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 556, + 673 + ], + "type": "text", + "content": "The AO loop was controlled with the open source pyRTC software4 which allows full control of the DM and science camera. This allowed us to define the optimized phase shifts to be the reference mirror \"flat\", from which the residual wavefront was measured by the AO WFS. Each image was then captured as usual, changing the DM's \"flat\" between exposures as dictated by the optimized phases, and then combining these images to restore our final SR image." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 693, + 512, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 693, + 512, + 703 + ], + "spans": [ + { + "bbox": [ + 324, + 693, + 512, + 703 + ], + "type": "text", + "content": "3Lexitek Inc. Phase Plates - https://lexitek.com" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 326, + 703, + 551, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 703, + 551, + 713 + ], + "spans": [ + { + "bbox": [ + 326, + 703, + 551, + 713 + ], + "type": "text", + "content": "4pyRTC on Github: https://github.com/jacotay7/pyRTC" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29147" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 70, + 108, + 225 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 108, + 225 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 108, + 225 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 108, + 225 + ], + "type": "image", + "image_path": "c6519791bbdc34218af30f4e09771ec1365aabc167ad5395c2b462392a8af882.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 227, + 99, + 235 + ], + "lines": [ + { + "bbox": [ + 75, + 227, + 99, + 235 + ], + "spans": [ + { + "bbox": [ + 75, + 227, + 99, + 235 + ], + "type": "text", + "content": "INPUT" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 240, + 555, + 275 + ], + "lines": [ + { + "bbox": [ + 55, + 240, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 555, + 275 + ], + "type": "text", + "content": "Figure 3. Simulated SR results. Comparisons of reconstructions with super resolution factors of " + }, + { + "bbox": [ + 55, + 240, + 555, + 275 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 55, + 240, + 555, + 275 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 240, + 555, + 275 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 55, + 240, + 555, + 275 + ], + "type": "text", + "content": " (inset regions marked in red). Images were generated with " + }, + { + "bbox": [ + 55, + 240, + 555, + 275 + ], + "type": "inline_equation", + "content": "N = 4" + }, + { + "bbox": [ + 55, + 240, + 555, + 275 + ], + "type": "text", + "content": " exposures and phases were optimized for each method and SR factor. For a full view of the results, please refer to the supplemental material [55]" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 126, + 78, + 216, + 112 + ], + "blocks": [ + { + "bbox": [ + 126, + 71, + 169, + 78 + ], + "lines": [ + { + "bbox": [ + 126, + 71, + 169, + 78 + ], + "spans": [ + { + "bbox": [ + 126, + 71, + 169, + 78 + ], + "type": "text", + "content": "OPTIMIZED PHASE" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 126, + 78, + 216, + 112 + ], + "lines": [ + { + "bbox": [ + 126, + 78, + 216, + 112 + ], + "spans": [ + { + "bbox": [ + 126, + 78, + 216, + 112 + ], + "type": "image", + "image_path": "82558cc619ce16844a7134d5b5b6065eae9b4a3073a4c3e214ab07eb12cf7d1d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 231, + 75, + 321, + 112 + ], + "blocks": [ + { + "bbox": [ + 231, + 75, + 321, + 112 + ], + "lines": [ + { + "bbox": [ + 231, + 75, + 321, + 112 + ], + "spans": [ + { + "bbox": [ + 231, + 75, + 321, + 112 + ], + "type": "image", + "image_path": "3bc147c196dc618ad92155a2ab49eff7b480b20fd4858f21e182951f700f5a60.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 343, + 76, + 433, + 112 + ], + "blocks": [ + { + "bbox": [ + 343, + 76, + 433, + 112 + ], + "lines": [ + { + "bbox": [ + 343, + 76, + 433, + 112 + ], + "spans": [ + { + "bbox": [ + 343, + 76, + 433, + 112 + ], + "type": "image", + "image_path": "767144f58486218e662454fabacfe7c13be9ec4c69e2a664b4b5bcd2c52e7bb4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 449, + 75, + 539, + 112 + ], + "blocks": [ + { + "bbox": [ + 449, + 75, + 539, + 112 + ], + "lines": [ + { + "bbox": [ + 449, + 75, + 539, + 112 + ], + "spans": [ + { + "bbox": [ + 449, + 75, + 539, + 112 + ], + "type": "image", + "image_path": "f2b722cb5929d4d2c093f697ccd33fc2c1a55949d43f197826a8be3ecfe55a23.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 117, + 114, + 223, + 165 + ], + "blocks": [ + { + "bbox": [ + 117, + 114, + 223, + 165 + ], + "lines": [ + { + "bbox": [ + 117, + 114, + 223, + 165 + ], + "spans": [ + { + "bbox": [ + 117, + 114, + 223, + 165 + ], + "type": "image", + "image_path": "c5eee0e14b1419f122ce471da2e9168cb85c21f209a75f88f1bdfd9062cf77d9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 117, + 168, + 223, + 195 + ], + "blocks": [ + { + "bbox": [ + 117, + 168, + 223, + 195 + ], + "lines": [ + { + "bbox": [ + 117, + 168, + 223, + 195 + ], + "spans": [ + { + "bbox": [ + 117, + 168, + 223, + 195 + ], + "type": "image", + "image_path": "64ba72ab7d3d9ce8a5e539a95faf5be13e8d601f65e2d44902c9c78131767d53.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 117, + 195, + 223, + 223 + ], + "blocks": [ + { + "bbox": [ + 117, + 195, + 223, + 223 + ], + "lines": [ + { + "bbox": [ + 117, + 195, + 223, + 223 + ], + "spans": [ + { + "bbox": [ + 117, + 195, + 223, + 223 + ], + "type": "image", + "image_path": "54f5df6e228dff27794b8bf05847001b0d472e32a75b1562bdc07eb0dbe9a5c6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 150, + 226, + 190, + 235 + ], + "lines": [ + { + "bbox": [ + 150, + 226, + 190, + 235 + ], + "spans": [ + { + "bbox": [ + 150, + 226, + 190, + 235 + ], + "type": "text", + "content": "LINEAR " + }, + { + "bbox": [ + 150, + 226, + 190, + 235 + ], + "type": "inline_equation", + "content": "2 \\times" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 224, + 114, + 330, + 165 + ], + "blocks": [ + { + "bbox": [ + 224, + 114, + 330, + 165 + ], + "lines": [ + { + "bbox": [ + 224, + 114, + 330, + 165 + ], + "spans": [ + { + "bbox": [ + 224, + 114, + 330, + 165 + ], + "type": "image", + "image_path": "4ef192cb17ad1bee79bc862559fb6969eda26ee1ccad688521640a307fe297d7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 224, + 168, + 330, + 195 + ], + "blocks": [ + { + "bbox": [ + 224, + 168, + 330, + 195 + ], + "lines": [ + { + "bbox": [ + 224, + 168, + 330, + 195 + ], + "spans": [ + { + "bbox": [ + 224, + 168, + 330, + 195 + ], + "type": "image", + "image_path": "eb8fa4e353775931a31f4d599041398335d90a75d7244d4ebad4f099b8614249.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 224, + 195, + 330, + 223 + ], + "blocks": [ + { + "bbox": [ + 224, + 195, + 330, + 223 + ], + "lines": [ + { + "bbox": [ + 224, + 195, + 330, + 223 + ], + "spans": [ + { + "bbox": [ + 224, + 195, + 330, + 223 + ], + "type": "image", + "image_path": "c6be9a541ba11c0e6c81e8d5694ed9d5d48025058dc72085af484cce1db07936.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 226, + 292, + 235 + ], + "lines": [ + { + "bbox": [ + 260, + 226, + 292, + 235 + ], + "spans": [ + { + "bbox": [ + 260, + 226, + 292, + 235 + ], + "type": "text", + "content": "CNN 2×" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 334, + 114, + 441, + 165 + ], + "blocks": [ + { + "bbox": [ + 334, + 114, + 441, + 165 + ], + "lines": [ + { + "bbox": [ + 334, + 114, + 441, + 165 + ], + "spans": [ + { + "bbox": [ + 334, + 114, + 441, + 165 + ], + "type": "image", + "image_path": "116d9a959290107b696da56766587aa6b1e38bf41c85e8e20eb34aaebca4d3a1.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 334, + 168, + 441, + 195 + ], + "blocks": [ + { + "bbox": [ + 334, + 168, + 441, + 195 + ], + "lines": [ + { + "bbox": [ + 334, + 168, + 441, + 195 + ], + "spans": [ + { + "bbox": [ + 334, + 168, + 441, + 195 + ], + "type": "image", + "image_path": "bdb19dbd4380011122d5f6c8fa530c3c8d726725d6eaf7f7bfd6ed203f3521d5.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 334, + 195, + 441, + 223 + ], + "blocks": [ + { + "bbox": [ + 334, + 195, + 441, + 223 + ], + "lines": [ + { + "bbox": [ + 334, + 195, + 441, + 223 + ], + "spans": [ + { + "bbox": [ + 334, + 195, + 441, + 223 + ], + "type": "image", + "image_path": "a3f7e380236cbe19b423fb7efaad3a7a1876d6c9f18d7072184fcb6b3f3a387e.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 227, + 408, + 235 + ], + "lines": [ + { + "bbox": [ + 367, + 227, + 408, + 235 + ], + "spans": [ + { + "bbox": [ + 367, + 227, + 408, + 235 + ], + "type": "text", + "content": "LINEAR " + }, + { + "bbox": [ + 367, + 227, + 408, + 235 + ], + "type": "inline_equation", + "content": "8\\times" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 441, + 114, + 547, + 165 + ], + "blocks": [ + { + "bbox": [ + 441, + 114, + 547, + 165 + ], + "lines": [ + { + "bbox": [ + 441, + 114, + 547, + 165 + ], + "spans": [ + { + "bbox": [ + 441, + 114, + 547, + 165 + ], + "type": "image", + "image_path": "d0fedd2c4b13b70a28065dda9d350127c889f7ba6ba189c68d1fc7e266c4a748.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 441, + 168, + 547, + 195 + ], + "blocks": [ + { + "bbox": [ + 441, + 168, + 547, + 195 + ], + "lines": [ + { + "bbox": [ + 441, + 168, + 547, + 195 + ], + "spans": [ + { + "bbox": [ + 441, + 168, + 547, + 195 + ], + "type": "image", + "image_path": "6cef5f6300fcf7aabad55a9c482a99800af1a0e4e7349495279b1d5690dc5a88.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 441, + 195, + 547, + 223 + ], + "blocks": [ + { + "bbox": [ + 441, + 195, + 547, + 223 + ], + "lines": [ + { + "bbox": [ + 441, + 195, + 547, + 223 + ], + "spans": [ + { + "bbox": [ + 441, + 195, + 547, + 223 + ], + "type": "image", + "image_path": "366dfb1dc26e768494389fce14bd3f0d5cb584918e59aa77e6a9116bea76e996.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 477, + 226, + 509, + 235 + ], + "lines": [ + { + "bbox": [ + 477, + 226, + 509, + 235 + ], + "spans": [ + { + "bbox": [ + 477, + 226, + 509, + 235 + ], + "type": "text", + "content": "CNN 8×" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 58, + 315, + 105, + 365 + ], + "blocks": [ + { + "bbox": [ + 63, + 301, + 101, + 309 + ], + "lines": [ + { + "bbox": [ + 63, + 301, + 101, + 309 + ], + "spans": [ + { + "bbox": [ + 63, + 301, + 101, + 309 + ], + "type": "text", + "content": "USAF Target" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 315, + 105, + 365 + ], + "lines": [ + { + "bbox": [ + 58, + 315, + 105, + 365 + ], + "spans": [ + { + "bbox": [ + 58, + 315, + 105, + 365 + ], + "type": "image", + "image_path": "cdfbe552a399a3fc8ce88659b043376a270c73db1af32818f55455d1e486dfbf.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 374, + 296, + 462 + ], + "lines": [ + { + "bbox": [ + 55, + 374, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 296, + 462 + ], + "type": "text", + "content": "Figure 4. Improved SR with jointly-optimized phase and CNN on simulated USAF resolution target. End-to-end training of both phase and CNN yields superior resolution improvements compared to a naive approach which does not optimize phases. The input to both CNNs is the inset region of the USAF target, marked in red. Qualitatively, by jointly optimizing the phase and the CNN, we can resolve smaller features and, quantitatively, achieve lower RMSE (displayed above each image)." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 109, + 315, + 163, + 365 + ], + "blocks": [ + { + "bbox": [ + 109, + 301, + 162, + 309 + ], + "lines": [ + { + "bbox": [ + 109, + 301, + 162, + 309 + ], + "spans": [ + { + "bbox": [ + 109, + 301, + 162, + 309 + ], + "type": "text", + "content": "Input (USAF Insert)" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 109, + 315, + 163, + 365 + ], + "lines": [ + { + "bbox": [ + 109, + 315, + 163, + 365 + ], + "spans": [ + { + "bbox": [ + 109, + 315, + 163, + 365 + ], + "type": "image", + "image_path": "c0852707515971bcc50aa711e47d138dbdf1300515d7b7110efdf21977fd36f4.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 169, + 316, + 223, + 364 + ], + "blocks": [ + { + "bbox": [ + 181, + 302, + 214, + 309 + ], + "lines": [ + { + "bbox": [ + 181, + 302, + 214, + 309 + ], + "spans": [ + { + "bbox": [ + 181, + 302, + 214, + 309 + ], + "type": "text", + "content": "Naive CNN" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 181, + 309, + 223, + 316 + ], + "lines": [ + { + "bbox": [ + 181, + 309, + 223, + 316 + ], + "spans": [ + { + "bbox": [ + 181, + 309, + 223, + 316 + ], + "type": "text", + "content": "RMSE " + }, + { + "bbox": [ + 181, + 309, + 223, + 316 + ], + "type": "inline_equation", + "content": "= 0.0544" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 169, + 316, + 223, + 364 + ], + "lines": [ + { + "bbox": [ + 169, + 316, + 223, + 364 + ], + "spans": [ + { + "bbox": [ + 169, + 316, + 223, + 364 + ], + "type": "image", + "image_path": "28166676520f708c71388ae5675c2edb153a11b53076a73653d33d8e72c6e517.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 227, + 316, + 285, + 364 + ], + "blocks": [ + { + "bbox": [ + 225, + 302, + 290, + 309 + ], + "lines": [ + { + "bbox": [ + 225, + 302, + 290, + 309 + ], + "spans": [ + { + "bbox": [ + 225, + 302, + 290, + 309 + ], + "type": "text", + "content": "CNN (Optimized Phase)" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 239, + 309, + 275, + 315 + ], + "lines": [ + { + "bbox": [ + 239, + 309, + 275, + 315 + ], + "spans": [ + { + "bbox": [ + 239, + 309, + 275, + 315 + ], + "type": "text", + "content": "RMSE = 0.0185" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 227, + 316, + 285, + 364 + ], + "lines": [ + { + "bbox": [ + 227, + 316, + 285, + 364 + ], + "spans": [ + { + "bbox": [ + 227, + 316, + 285, + 364 + ], + "type": "image", + "image_path": "df367d91ef4f47282f9e04dfea08556a6bec14ecc255545a06c80990c57de081.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "bbox": [ + 55, + 486, + 113, + 497 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 113, + 497 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 113, + 497 + ], + "type": "text", + "content": "6.1. Results" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 55, + 502, + 296, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 296, + 658 + ], + "type": "text", + "content": "A green LED (centered at " + }, + { + "bbox": [ + 55, + 502, + 296, + 658 + ], + "type": "inline_equation", + "content": "530~\\mathrm{nm}" + }, + { + "bbox": [ + 55, + 502, + 296, + 658 + ], + "type": "text", + "content": ") was used to project photographic slides on the optical bench detailed above. To demonstrate our method's resolving power, we imaged a 1\" US Air Force resolution chart (Thorlabs Positive 1951 USAF Test Target Groups 2-7, " + }, + { + "bbox": [ + 55, + 502, + 296, + 658 + ], + "type": "inline_equation", + "content": "\\varnothing 1^{\\prime \\prime}" + }, + { + "bbox": [ + 55, + 502, + 296, + 658 + ], + "type": "text", + "content": "). Results are shown for both simulation (Fig. 4) and experiment (Fig. 6). We also show our method's performance on more general scenes using a variety of real-world photos in Fig. 6. Because our AO system was designed specifically for sensing in infrared wavelengths, the quality of our images (captured at " + }, + { + "bbox": [ + 55, + 502, + 296, + 658 + ], + "type": "inline_equation", + "content": "635~\\mathrm{nm}" + }, + { + "bbox": [ + 55, + 502, + 296, + 658 + ], + "type": "text", + "content": ") is limited by chromatic aberration and vignetting. As such, results in this paper are cropped and masked for clarity; raw images are provided in the supplemental materials [55]." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 55, + 670, + 125, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 125, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 125, + 681 + ], + "type": "text", + "content": "7. Discussion" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "MTF Analysis: The modulation transfer function (MTF) is a measure of the achievable contrast for an optical system" + } + ] + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 345, + 301, + 521, + 626 + ], + "blocks": [ + { + "bbox": [ + 345, + 301, + 521, + 626 + ], + "lines": [ + { + "bbox": [ + 345, + 301, + 521, + 626 + ], + "spans": [ + { + "bbox": [ + 345, + 301, + 521, + 626 + ], + "type": "image", + "image_path": "cb365a0e87768ea3fb3d082032a858ef70a78bcd1f020b4dac4388241295343b.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 629, + 555, + 662 + ], + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 662 + ], + "type": "text", + "content": "Figure 5. Optical setup. We present a top-down view of the optical bench where our experiments were conducted, labeling both the wavefront sensor path and the science (image) paths." + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "over a range of spatial frequencies [44]. An ideal lens will linearly decrease from an MTF of 1 down to 0 as the spa" + } + ] + } + ], + "index": 41 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29148" + } + ] + } + ], + "index": 42 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 69, + 289, + 235 + ], + "blocks": [ + { + "bbox": [ + 61, + 69, + 289, + 235 + ], + "lines": [ + { + "bbox": [ + 61, + 69, + 289, + 235 + ], + "spans": [ + { + "bbox": [ + 61, + 69, + 289, + 235 + ], + "type": "image", + "image_path": "13caa61d6c4b9cb1d678eec7e2934cc1fceeae7f62bafc85d1966af2be352b5e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 239, + 296, + 318 + ], + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 318 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 318 + ], + "type": "text", + "content": "Figure 6. Experimental results. We reconstruct " + }, + { + "bbox": [ + 55, + 239, + 296, + 318 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 55, + 239, + 296, + 318 + ], + "type": "text", + "content": " upsampled images of the USFA resolution target and three natural images (Grapes, Bookshelf, Trees) from low resolution images (LR Input) captured with the optical setup depicted in Fig. 5. We present results using both linear and CNN method, each using their own optimized phase profiles. Both result in higher resolutions than the original input, as seen quantitatively on the USFA chart." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 338, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 296, + 422 + ], + "type": "text", + "content": "tial frequency increases and, in general, the closer a measured MTF behaves to this the better its resolving power. The MTF plots in Fig. 7 correspond to the USAF simulation of Fig. 4 and the real experiment of Fig. 6 computed using a knife-edge method [36]. Both show an increase of MTF across all spatial resolutions and thus demonstrate our method's superiority (i.e., higher preserved contrast)." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 56, + 431, + 296, + 506 + ], + "blocks": [ + { + "bbox": [ + 56, + 431, + 296, + 506 + ], + "lines": [ + { + "bbox": [ + 56, + 431, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 56, + 431, + 296, + 506 + ], + "type": "image", + "image_path": "c4c1c23bd24ae26a710620a60101adce5fa80642e0c075ebd6eb0fb478f7b405.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 65, + 510, + 285, + 521 + ], + "lines": [ + { + "bbox": [ + 65, + 510, + 285, + 521 + ], + "spans": [ + { + "bbox": [ + 65, + 510, + 285, + 521 + ], + "type": "text", + "content": "Figure 7. Comparison of simulated and experimental MTFs." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "Phase Analysis: One advantage of representing phase profiles using Zernike modes is the ability to infer exactly which optical aberrations are most important to our reconstruction algorithms. In Fig. 8, we plot the total absolute power of the first nine basis modes identified by our method for both the CNN and the linear reconstructor, using a fixed number of exposures while varying the magnification factor. The optimized phases reveal insights. First, total modal power increases with the magnification order. Intuitively, this makes sense as more movement within a pixel is needed to sample higher resolutions. Second, CNNs apply roughly the same power across all modes for a given magnification while the linear method has more variance. Perhaps this is because linear solvers cannot leverage higher-frequency spatial information in the image as well as CNNs can." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 553, + 178 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 553, + 178 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 553, + 178 + ], + "type": "image", + "image_path": "72a42af0d1574fdfa027b525e70c226b1a9810708ea68f1746d5e608a4721528.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 179, + 555, + 257 + ], + "lines": [ + { + "bbox": [ + 313, + 179, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 313, + 179, + 555, + 257 + ], + "type": "text", + "content": "Figure 8. Comparison of modal powers used to induce SR. We show modal power in the Zernike basis to explore the optimized phases (corresponding to the phases shown in Fig. 3). The linear method uses low-order, symmetric modes while the CNN spreads across all modes. Comparing " + }, + { + "bbox": [ + 313, + 179, + 555, + 257 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 313, + 179, + 555, + 257 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 313, + 179, + 555, + 257 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 313, + 179, + 555, + 257 + ], + "type": "text", + "content": " magnifications, they take opposite approaches: the CNN prefers low modes at higher resolutions, while the linear uses lower modes at low resolutions." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 280, + 555, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 280, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 555, + 389 + ], + "type": "text", + "content": "Limitations & Future Work: Currently, our work is limited to telescopes with a single-conjugate adaptive optics (SCAO) system, which treats the atmosphere as a single column. Many of today's observatories instead use multiple WFS and DMs. This leads to more accurate corrections across a wider FoV and can even be tuned to different atmospheric layers. Such systems do not preclude the use of our method, since our method can be extended to multiple DMs and potentially enhanced by their extra degrees of freedom." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 396, + 554, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 554, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 554, + 467 + ], + "type": "text", + "content": "Other directions for future work include real-time viewing of video from the science camera by cycling through each phase multiple times to create individually super-resolved frames. Once extended to multiple DMs, additional multiplexing could also be achieved by varying each DM at different rates." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 475, + 554, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 554, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 554, + 523 + ], + "type": "text", + "content": "Most importantly, now that we have demonstrated our methodology with an optical bench prototype, our current priority is to show its effectiveness on sky with a science class telescope and AO system." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 531, + 554, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 554, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 554, + 616 + ], + "type": "text", + "content": "Conclusion: This work shows that science images can encode sufficient MISR information using the AO systems already present in state-of-the-art telescope observatories. We achieve resolution gains without affecting the scientific data in both simulations and optical-bench experiments. With this method, we aim to unlock new scientific capabilities at existing observatories without altering the optical system." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 626, + 417, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 626, + 417, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 626, + 417, + 639 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 644, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 644, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 644, + 554, + 712 + ], + "type": "text", + "content": "RS acknowledges NVIDIA for GPU hardware support. EYHL acknowledges the support of NSERC under CGSD. SS acknowledges the support of the Canada Foundation for Innovation and the Ontario Research Fund. KNK acknowledges the support of NSERC under the RGPIN and RTI programs." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29149" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 92, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 92, + 296, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 92, + 296, + 146 + ], + "spans": [ + { + "bbox": [ + 61, + 92, + 296, + 146 + ], + "type": "text", + "content": "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 126-135, 2017. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 181 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 181 + ], + "type": "text", + "content": "[2] Saeed Anwar, Salman Khan, and Nick Barnes. A deep journey into super-resolution: A survey. ACM Computing Surveys (CSUR), 53(3):1-34, 2020. 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 295, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 295, + 216 + ], + "type": "text", + "content": "[3] Roberto J. Avila, Warren J. Hack, and Stsci Team. Astrodrizzle: Aligning images from multiple instruments. In American Astronomical Society, 2012. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 217, + 295, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 217, + 295, + 261 + ], + "spans": [ + { + "bbox": [ + 62, + 217, + 295, + 261 + ], + "type": "text", + "content": "[4] Thilo Bauer. Super-resolution imaging: The use case of optical astronomy. In Proceedings of the IADIS International Conference Computer Graphics, Visualization, Computer Vision and Image Processing, Rome, pages 49-59, 2011. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 262, + 295, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 262, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 62, + 262, + 295, + 306 + ], + "type": "text", + "content": "[5] Jacques M Beckers. Adaptive optics for astronomy-principles, performance, and applications. In: Annual Review of Astronomy and Astrophysics. Vol. 31 (A94-12726 02-90), p. 13-62., 31:13-62, 1993. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 307, + 295, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 307, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 307, + 295, + 350 + ], + "type": "text", + "content": "[6] Anthony Berdeu, Michel Tallon, Éric Thiébaut, and Maud Langlois. Analytical modelling of adaptive optics systems: Role of the influence function. *Astronomy & Astrophysics*, 674:A112, 2023. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 353, + 296, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 353, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 353, + 296, + 396 + ], + "type": "text", + "content": "[7] Goutam Bhat, Martin Danelljan, Luc Van Gool, and Radu Timofte. Deep burst super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9209-9218, 2021. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 397, + 295, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 397, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 397, + 295, + 430 + ], + "type": "text", + "content": "[8] David Borncamp, J. M. Anderson, Norman A. Grogin, and Warren J. Hack. Acs/wfc revised geometric distortion for drizzlepac. In Instrument Science Report ACS/WFC, 2015. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 432, + 295, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 432, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 432, + 295, + 475 + ], + "type": "text", + "content": "[9] Praneeth Chakravarthula, Ethan Tseng, Tarun Srivastava, Henry Fuchs, and Felix Heide. Learned hardware-in-the-loop phase retrieval for holographic near-eye displays. ACM Transactions on Graphics (TOG), 39(6):1-18, 2020. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 477, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 477, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 57, + 477, + 295, + 521 + ], + "type": "text", + "content": "[10] Dorian Chan, Srinivas Narasimhan, and Matthew O'Toole. Holocurtains: Programming light curtains via binary holography. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 522, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 522, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 57, + 522, + 295, + 567 + ], + "type": "text", + "content": "[11] Stanley H Chan, Xiran Wang, and Omar A Elgendy. Plug-and-play admm for image restoration: Fixed-point convergence and applications. IEEE Transactions on Computational Imaging, 3(1):84-98, 2016. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 568, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 568, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 57, + 568, + 295, + 623 + ], + "type": "text", + "content": "[12] Wenzheng Chen, Parsa Mirdehghan, Sanja Fidler, and Kiriakos N Kutulakos. Auto-tuning structured light by optical stochastic gradient descent. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5970-5980, 2020. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 624, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 667 + ], + "type": "text", + "content": "[13] Xu Cheng, Arthur Bradley, Sowmya Ravikumar, and Larry N Thibos. Visual impact of zernike and seidel forms of monochromatic aberrations. Optometry and Vision Science, 87(5):300-312, 2010. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 669, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 669, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 669, + 295, + 712 + ], + "type": "text", + "content": "[14] Rodolphe Conan and C Correia. Object-oriented matlab adaptive optics toolbox. In Adaptive Optics Systems IV, page 91486C. International Society for Optics and Photonics, 2014. 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "text", + "content": "[15] Sony Corporation. Pixel Shift Multi Shooting | SONY — support.d-imaging.sony.co.jp. https://support. d-imaging.sony.co.jp/support/ilm/cpsms/ ilce7rm3/en/index.html, 2025. [Accessed 30-03- 2025]. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 129, + 554, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 554, + 173 + ], + "type": "text", + "content": "[16] Antonia Creswell, Tom White, Vincent Dumoulin, Kai Arulkumaran, Biswa Sengupta, and Anil A Bharath. Generative adversarial networks: An overview. IEEE Signal Processing Magazine, 35(1):53-65, 2018. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 175, + 554, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 554, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 554, + 206 + ], + "type": "text", + "content": "[17] Benjamin P Cumming and Min Gu. Direct determination of aberration functions in microscopy by an artificial neural network. Optics Express, 28(10):14511-14521, 2020. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 208, + 554, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 554, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 554, + 251 + ], + "type": "text", + "content": "[18] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In Computer Vision - ECCV 2014, pages 184-199, Cham, 2014. Springer International Publishing. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 253, + 554, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 554, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 554, + 295 + ], + "type": "text", + "content": "[19] Theodore B DuBose, Dennis F Gardner, and Abbie T Watnik. Intensity-enhanced deep network wavefront reconstruction in shack-hartmann sensors. Optics Letters, 45(7):1699-1702, 2020. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 297, + 554, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 554, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 554, + 330 + ], + "type": "text", + "content": "[20] Mark A Ealey and John F Washeba. Continuous facesheet low voltage deformable mirrors. Optical Engineering, 29 (10):1191-1198, 1990. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 332, + 554, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 554, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 554, + 386 + ], + "type": "text", + "content": "[21] Linjing Fang, Fred Monroe, Sammy Weiser Novak, Lyndsey Kirk, Cara R Schiavon, Seungyoon B Yu, Tong Zhang, Melissa Wu, Kyle Kastner, Alaa Abdel Latif, et al. Deep learning-based point-scanning super-resolution imaging. Nature Methods, 18(4):406-416, 2021. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 388, + 554, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 554, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 554, + 430 + ], + "type": "text", + "content": "[22] Sina Farsiu, Dirk Robinson, Michael Elad, and Peyman Milanfar. Robust shift and add approach to superresolution. In Applications of Digital Image Processing XXVI, pages 121-130. SPIE, 2003. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 432, + 554, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 432, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 432, + 554, + 475 + ], + "type": "text", + "content": "[23] AS Fruchter and RN Hook. Drizzle: A method for the linear reconstruction of undersampled images. *Publications of the Astronomical Society of the Pacific*, 114(792):144, 2002. 1, 2, 4, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 477, + 554, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 477, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 477, + 554, + 510 + ], + "type": "text", + "content": "[24] Andrew S. Fruchter. A new method for band-limited imaging with undersampled detectors. *Publications of the Astronomical Society of the Pacific*, 123(902):497, 2011. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 511, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 511, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 511, + 554, + 555 + ], + "type": "text", + "content": "[25] Andrea M Ghez, BL Klein, M Morris, and EE Becklin. High proper-motion stars in the vicinity of sagittarius a*: Evidence for a supermassive black hole at the center of our galaxy. The Astrophysical Journal, 509(2):678, 1998. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 556, + 554, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 556, + 554, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 556, + 554, + 588 + ], + "type": "text", + "content": "[26] Roberto Gilmotzzi and Jason Spyromilio. The european extremely large telescope (e-elt). The Messenger, 127(11):3, 2007. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 590, + 554, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 554, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 554, + 611 + ], + "type": "text", + "content": "[27] S Gonzaga, W Hack, A Fruchter, and J Mack. The drizzlepac handbook. The DrizzlePac Handbook, 2012. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 613, + 554, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 554, + 656 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 554, + 656 + ], + "type": "text", + "content": "[28] Karen M Hampson, Raphaël Turcotte, Donald T Miller, Kazuhiro Kurokawa, Jared R Males, Na Ji, and Martin J Booth. Adaptive optics for high-resolution imaging. Nature Reviews Methods Primers, 1(1):1-26, 2021. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 658, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 689 + ], + "type": "text", + "content": "[29] James E Harvey and Christ Ftaclas. Diffraction effects of telescope secondary mirror spiders on various image-quality criteria. Applied Optics, 34(28):6337-6349, 1995. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "text", + "content": "[30] Hamid Reza Vaezi Joze, Ilya Zharkov, Karlton Powell, Carl Ringler, Luming Liang, Andy Roulston, Moshe Lutz," + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29150" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 117 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 117 + ], + "type": "text", + "content": "and Vivek Pradeep. Imagepairs: Realistic super resolution dataset via beam splitter camera rig. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 518-519, 2020. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 118, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 118, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 57, + 118, + 295, + 140 + ], + "type": "text", + "content": "[31] SR Kulkarni. Instruments on large optical telescopes-a case study. arXiv preprint arXiv:1606.06674, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 295, + 173 + ], + "type": "text", + "content": "[32] Vasudevan Lakshminarayanan and Andre Fleck. Zernike polynomials: a guide. Journal of Modern Optics, 58(7):545-561, 2011. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 229 + ], + "type": "text", + "content": "[33] Rico Landman, Sebastiaan Y Haffert, Vikram M Radhakrishnan, and Christoph U Keller. Self-optimizing adaptive optics control with reinforcement learning. In Adaptive Optics Systems VII, page 1144849. International Society for Optics and Photonics, 2020. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 294, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 263 + ], + "type": "text", + "content": "[34] Tod R Lauer. Combining undersampled dithered images. *Publications of the Astronomical Society of the Pacific*, 111 (756):227, 1999. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 265, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 265, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 265, + 295, + 319 + ], + "type": "text", + "content": "[35] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, and Wenzhe Shi. Photo-realistic single image super-resolution using a generative adversarial network, 2017. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 321, + 295, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 295, + 365 + ], + "type": "text", + "content": "[36] Hang Li, Changxiang Yan, and Jianbing Shao. Measurement of the modulation transfer function of infrared imaging system by modified slant edge method. Journal of the Optical Society of Korea, 20(3):381-388, 2016. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 365, + 295, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 295, + 409 + ], + "type": "text", + "content": "[37] Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479:47-59, 2022. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 411, + 295, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 295, + 464 + ], + "type": "text", + "content": "[38] Orly Liba, Kiran Murthy, Yun-Ta Tsai, Tim Brooks, Tianfan Xue, Nikhil Karnad, Qiurui He, Jonathan T Barron, Dillon Sharlet, Ryan Geiss, et al. Handheld mobile photography in very low light. ACM Transactions on Graphics (TOG), 38 (6):1-16, 2019. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 467, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 295, + 521 + ], + "type": "text", + "content": "[39] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 136-144, 2017. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 523, + 295, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 295, + 577 + ], + "type": "text", + "content": "[40] Zhisheng Lu, Juncheng Li, Hong Liu, Chaoyan Huang, Linlin Zhang, and Tieyong Zeng. Transformer for single image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 457-466, 2022. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 578, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 295, + 601 + ], + "type": "text", + "content": "[41] Bahattin Can Maral. Single image super-resolution methods: A survey. arXiv preprint arXiv:2202.11763, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 601, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 295, + 645 + ], + "type": "text", + "content": "[42] Christian Marois, Bruce Macintosh, Travis Barman, B Zuckerman, Inseok Song, Jennifer Patience, David Lafreniere, and René Doyon. Direct imaging of multiple planets orbiting the star hr 8799. science, 322(5906):1348-1352, 2008. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 690 + ], + "type": "text", + "content": "[43] Y Mellier, Abdurrouf Abdurrouf, JA Acevedo Barroso, A Achucarro, J Adamek, R Adam, GE Addison, N Aghanim, M Aguena, V Ajani, et al. Euclid. i. overview of the euclid mission. *Astronomy & Astrophysics*, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 714 + ], + "type": "text", + "content": "[44] HH Nasse. How to read mtf curves. Carl Zeiss, Camera Lens Division, page 33, 2008. 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "type": "text", + "content": "[45] Ngoc Long Nguyen, Jérémy Anger, Axel Davy, Pablo Arias, and Gabriele Facciolo. Self-supervised multi-image superresolution for push-frame satellite images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1121-1131, 2021. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 129, + 554, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 162 + ], + "type": "text", + "content": "[46] Sylvain Oberti, Carlos Correia, Thierry Fusco, Benoit Neichel, and Pierre Guiraud. Super-resolution wavefront reconstruction. *Astronomy & Astrophysics*, 667:A48, 2022. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 163, + 554, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 554, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 554, + 195 + ], + "type": "text", + "content": "[47] Scott W Paine and James R Fienup. Machine learning for improved image-based wavefront sensing. Optics Letters, 43(6):1235-1238, 2018. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 197, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 554, + 262 + ], + "type": "text", + "content": "[48] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 263, + 554, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 554, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 554, + 306 + ], + "type": "text", + "content": "[49] Yifan Peng, Suyeon Choi, Nitish Padmanaban, and Gordon Wetzstein. Neural holography with camera-in-the-loop training. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 308, + 554, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 554, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 554, + 331 + ], + "type": "text", + "content": "[50] Ben C Platt and Roland Shack. History and principles of shack-hartmann wavefront sensing, 2001. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 332, + 554, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 554, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 554, + 364 + ], + "type": "text", + "content": "[51] Roberto Ragazzoni, Emiliano Diolaiti, and Elise Vernet. A pyramid wavefront sensor with no dynamic modulation. Optics Communications, 208(1-3):51-60, 2002. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 365, + 554, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 554, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 554, + 431 + ], + "type": "text", + "content": "[52] Vincent Sitzmann, Steven Diamond, Yifan Peng, Xiong Dun, Stephen Boyd, Wolfgang Heidrich, Felix Heide, and Gordon Wetzstein. End-to-end optimization of optics and image processing for achromatic extended depth of field and superresolution imaging. ACM Transactions on Graphics (TOG), 37(4):1-13, 2018. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 433, + 554, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 554, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 554, + 487 + ], + "type": "text", + "content": "[53] Robin Swanson, Masen Lamb, Carlos M Correia, Suresh Sivanandam, and Kiriakos Kutulakos. Closed loop predictive control of adaptive optics systems with convolutional neural networks. Monthly Notices of the Royal Astronomical Society, 503(2):2944-2954, 2021. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 488, + 554, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 488, + 554, + 531 + ], + "spans": [ + { + "bbox": [ + 316, + 488, + 554, + 531 + ], + "type": "text", + "content": "[54] Robin Swanson, Suresh Sivanandam, and Kiriakos N Kutulakos. Adaptive optics mediated sub-pixel super resolution. In Adaptive Optics Systems VIII, pages 2594-2600. SPIE, 2022. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 533, + 554, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 554, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 554, + 588 + ], + "type": "text", + "content": "[55] Robin Swanson, Esther Y. H. Lin, Masen Lamb, Suresh Sivanandam, and Kiriakos N. Kutulakos. Super resolved imaging with adaptive optics: Supplemental document. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2025. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 590, + 554, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 554, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 554, + 621 + ], + "type": "text", + "content": "[56] Jing Tian and Kai-Kuang Ma. A survey on super-resolution imaging. Signal, Image and Video Processing, 5(3):329-342, 2011. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 624, + 554, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 554, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 554, + 667 + ], + "type": "text", + "content": "[57] Ethan Tseng, Shane Colburn, James Whitehead, Luocheng Huang, Seung-Hwan Baek, Arka Majumdar, and Felix Heide. Neural nano-optics for high-quality thin lens imaging. Nature Communications, 12(1):1-7, 2021. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 669, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 554, + 689 + ], + "type": "text", + "content": "[58] Robert K Tyson. Adaptive ptics engineering handbook. Marcel Dekker New York, 2000. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 691, + 555, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 555, + 714 + ], + "type": "text", + "content": "[59] Robert K Tyson and Benjamin West Frazier. Principles of adaptive optics. CRC press, 2022. 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29151" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 441 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[60] Congli Wang, Qiang Fu, Xiong Dun, and Wolfgang Heidrich. Megapixel adaptive optics: towards correcting large-scale distortions in computational cameras. ACM Transactions on Graphics (TOG), 37(4):1-12, 2018. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "text", + "content": "[61] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "text", + "content": "[62] Bartlomiej Wronski, Ignacio Garcia-Dorado, Manfred Ernst, Damien Kelly, Michael Krainin, Chia-Kai Liang, Marc Levoy, and Peyman Milanfar. Handheld multi-frame superresolution. ACM Transactions on Graphics (TOG), 38(4): 1-18, 2019. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 261 + ], + "type": "text", + "content": "[63] Haitang Yang, Esther Y. H. Lin, Kiriakos N. Kutulakos, and George V. Eleftheriades. Sub-wavelength passive single-shot computational super-oscillatory imaging. Optica, 9(12): 1444–1447, 2022. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 306 + ], + "type": "text", + "content": "[64] Wenming Yang, Xuechen Zhang, Yapeng Tian, Wei Wang, Jing-Hao Xue, and Qingmin Liao. Deep learning for single image super-resolution: A brief review. IEEE Transactions on Multimedia, 21(12):3106-3121, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 294, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 351 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 351 + ], + "type": "text", + "content": "[65] Linwei Yue, Huanfeng Shen, Jie Li, Qiangqiang Yuan, Hongyan Zhang, and Liangpei Zhang. Image superresolution: The techniques, applications, and future. Signal Processing, 128:389-408, 2016. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 353, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 294, + 386 + ], + "type": "text", + "content": "[66] Biwei Zhang, Jiazhu Zhu, Ke Si, and Wei Gong. Deep learning assisted zonal adaptive aberration correction. Frontiers in Physics, page 634, 2021. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 387, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 387, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 387, + 294, + 441 + ], + "type": "text", + "content": "[67] Hao Zhu, Mantang Guo, Hongdong Li, Qing Wang, and Antonio Robles-Kelly. Revisiting spatio-angular trade-off in light field cameras and extended applications in superresolution. IEEE Transactions on Visualization and Computer Graphics, 27(6):3019-3033, 2019. 2" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29152" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SuperDec_ 3D Scene Decomposition with Superquadrics Primitives/ed2eb095-495d-4dd3-9f36-9c1449d1ce9d_content_list.json b/2025/SuperDec_ 3D Scene Decomposition with Superquadrics Primitives/ed2eb095-495d-4dd3-9f36-9c1449d1ce9d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e93abd517cb41b569e27f825d367864824f1ea11 --- /dev/null +++ b/2025/SuperDec_ 3D Scene Decomposition with Superquadrics Primitives/ed2eb095-495d-4dd3-9f36-9c1449d1ce9d_content_list.json @@ -0,0 +1,1749 @@ +[ + { + "type": "text", + "text": "SUPERDEC: 3D Scene Decomposition with Superquadric Primitives", + "text_level": 1, + "bbox": [ + 155, + 130, + 841, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Elisabetta Fedele $^{1,2}$", + "bbox": [ + 99, + 164, + 259, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Boyang Sun", + "bbox": [ + 279, + 165, + 387, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Leonidas Guibas", + "bbox": [ + 408, + 165, + 553, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Marc Pollefeys1,3", + "bbox": [ + 571, + 165, + 710, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Francis Engelmann", + "bbox": [ + 733, + 165, + 897, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ ETH Zurich", + "bbox": [ + 330, + 198, + 413, + 213 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Stanford University", + "bbox": [ + 441, + 198, + 566, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Microsoft", + "bbox": [ + 594, + 198, + 663, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 250, + 326, + 265 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present SUPERDEC, an approach for creating compact 3D scene representations via decomposition into superquadric primitives. While most recent methods use geometric primitives to obtain photorealistic 3D reconstructions, we instead leverage them to obtain a compact yet expressive representation. To this end, we design a novel architecture that efficiently decomposes point clouds of arbitrary objects into a compact set of superquadrics. We train our model on ShapeNet and demonstrate its generalization capabilities on object instances from ScanNet++ as well as on full Replica scenes. Finally, we show that our compact superquadric-based representation supports a wide range of downstream applications, including robotic manipulation and controllable visual content generation. Project page: https://super-dec.github.io.", + "bbox": [ + 89, + 282, + 483, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 540, + 220, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D scene representations play a central role in computer vision and robotics, enabling tasks such as 3D scene understanding [33, 48, 49], scene generation [40, 42, 43], functional reasoning [19, 59], and scene interaction [9, 13, 21, 38]. Recent work [18] utilizes 3D Gaussians as geometric primitives to produce high-quality, photorealistic reconstructions. However, such representations are often memory-intensive. In contrast, we propose a more lightweight yet geometrically faithful 3D scene representation by decomposing the input point cloud into a compact set of explicit primitives, namely superquadrics (Fig. 1).", + "bbox": [ + 89, + 566, + 482, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Representations for 3D scenes include well-established formats such as point clouds, meshes, signed distance functions, and voxel grids, each offering different trade-offs among geometric detail, computational cost, resolution, performance, interpretability, and editability. Recently, multi-view approaches such as Neural Radiance Fields (NeRF) [29] and Gaussian Splatting (GS) [18] have gained popularity as 3D scene representations. These methods optimize photometric losses to ensure that their underlying representations (implicit in the case of NeRF, and explicit in the case of GS) are consistent with observed images. While", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/84cec3249a90f605c4bf77d716a779cfe529f568e00cfca34074d5699173ef81.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 248, + 709, + 359 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5c1752c06b242dd1a2f062f39b7cb30a65b456fc79c0fdfc7528387eaf83a4ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 248, + 903, + 359 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fee4ab809c7cb660769b3b3c811e49a4e109eb29d2763c96cbc2e1b9a54c9e29.jpg", + "image_caption": [ + "Figure 1. 3D Scene Decomposition with Superquadrics. Given a 3D point cloud of an indoor scene, SUPERDEC decomposes all scene objects into a compact set of superquadric primitives. Different object instances are visualized with different colors." + ], + "image_footnote": [], + "bbox": [ + 517, + 361, + 709, + 473 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6ec755e3bb8d3a158ef44d63135f557a4062c39c52b501b2163819ec69e8be30.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 361, + 903, + 473 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "these approaches excel at achieving photorealism, they lack explicit control over compactness, often resulting in large, non-modular scene encodings that are unsuitable for tasks requiring explicit spatial reasoning.", + "bbox": [ + 511, + 561, + 906, + 622 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While optimizing compactness on a scene level remains a challenging task, prior work has shown that geometric primitives such as cuboids [51, 57] or superquadrics [1, 31, 32] enable compact and interpretable decompositions of individual objects. These methods are generally either learning-based [32, 57], favoring speed at the cost of geometric accuracy, or optimization-based [1, 23, 31], offering improved accuracy at the expense of higher computational cost. Although both types of approaches can be effective for certain object classes, they often struggle to generalize across datasets with diverse object geometries. Learning-based methods typically require category-specific training, while optimization-based ones rely on hand-crafted heuristics, limiting their scalability in open-world settings.", + "bbox": [ + 511, + 625, + 908, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Motivated by the abstraction capabilities of geometric primitives for individual object categories, we propose to represent complex 3D scenes using a compact set of superquadrics. To this end, we learn class-agnostic object", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24625", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "level shape priors to optimize compactness and leverage an off-the-shelf 3D instance segmentation method (Mask3D [41]) to scale our approach to full 3D scenes.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We choose superquadrics as building blocks for our representation, as they offer more accurate shape modeling than cuboids while incurring minimal additional parameter overhead (9 v.s. 11, including 6-DoF pose parameters). To develop a model capable of generalizing across diverse object types, we draw inspiration from supervised segmentation [5, 6, 41], and approach the problem from the perspective of unsupervised geometric-based segmentation, using local point-based features to iteratively refine predicted geometric primitives. Our model is trained on ShapeNet [4] and evaluated on three challenging and diverse 3D datasets: ShapeNet [4], ScanNet++ [58], and Replica [45]. On the 3D object dataset ShapeNet, our approach achieves an L2 error six times smaller than prior state-of-the-art methods [32], while requiring only half the number of primitives. On ScanNet++ and Replica, we demonstrate that our method generalizes well to real-world scene-level settings, despite being trained solely on ShapeNet. Finally, we show the practical utility of our method as a scene representation for robotic tasks, including path planning and object grasping, as well as an editable 3D scene representation for controllable image generation.", + "bbox": [ + 91, + 137, + 482, + 467 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are the following:", + "bbox": [ + 89, + 468, + 413, + 483 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) We introduce SUPERDEC, a novel method for decomposing 3D scenes using superquadric primitives.", + "2) SUPERDEC achieves state-of-the-art object decomposition scores on ShapeNet trained jointly on multiple classes.", + "3) We demonstrate the effectiveness of 3D superquadric scene representations for robotic tasks and controllable generative content creation." + ], + "bbox": [ + 89, + 483, + 482, + 588 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 603, + 232, + 618 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning-based methods have shown that neural networks, when equipped with suitable reconstruction losses, can directly predict geometric primitive parameters to decompose point clouds into a minimal set of primitives for specific object categories. Tulsiani [51] introduced a CNN-based method for cuboid decomposition, which was later extended to more expressive primitives such as superquadrics by Paschalidou et al. [32]. CSA [57] further enhanced interpretability by employing a stronger point encoder and jointly predicting cuboid parameters and part segmentations. However, these methods remain constrained by their reliance on category-specific training. We attribute this limitation to their model design, which encodes only global shape features, sufficient for intra-category generalization but ineffective for decomposing out-of-category objects.", + "bbox": [ + 89, + 622, + 482, + 849 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Optimization-based methods largely originate from the literature on superquadric fitting. EMS [23] revisited this line of work by introducing a probabilistic formulation that", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "enables the decomposition of arbitrary objects into multiple superquadrics. Given an input point cloud, the method first fits a superquadric to the main structure and identifies unfitted outlier clusters, which are then recursively processed in a hierarchical fashion up to a predefined depth level. However, as noted in their paper and confirmed by our experiments, this approach implicitly assumes that objects exhibit a hierarchical geometric structure, limiting its applicability to many real-world objects such as tables and chairs. Other methods, such as Marching Primitives [24], require Signed Distance Functions (SDFs) as input, which are generally not easily available in real-world scenes. More fundamentally, since these approaches optimize from scratch for each object, they cannot leverage generalizable point features or learned shape priors, both of which are critical for abstraction and robustness under partial observations, a common challenge in practical 3D capture scenarios.", + "bbox": [ + 511, + 90, + 903, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Scene-level decomposition. With the emergence of 3DGS [18], an increasing number of works have explored representing 3D scenes using various geometric primitives as generalized ellipsoids [15] and convexes [14]. While heuristics can control the number of Gaussians, achieving truly compact representations remains challenging. DBW [31] addresses this by fitting a small set of textured superquadrics to 3D scenes, building on the principles of 3DGS. Given a set of scene images, it performs test-time optimization with a photometric loss and renders the primitives using a differentiable rasterizer [22]. To model the environment, DBW adds a meshed ground plane and a meshed icosphere for the background. However, it is restricted to scenes with fewer than 10 primitives and requires objects to be aligned to a ground plane. Furthermore, the optimization is computationally expensive, taking around three hours even on simple DTU [17] scenes. Our method differs significantly in terms of input requirements, generality of application, and computational efficiency.", + "bbox": [ + 511, + 354, + 906, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Superquadrics are a parametric family of shapes introduced by Barr et al. [2] in 1981 and have since been widely adopted in both computer vision and graphics [7, 34, 44]. Their popularity stems from their ability to represent a diverse range of shapes with a highly compact parameterization. A superquadric in its canonical pose is defined by just five parameters: $(s_x, s_y, s_z)$ for the scales along the three principal semi-axes and $(\\epsilon_1, \\epsilon_2)$ for the shape-defining exponents. Given those parameters, their surface is described by the implicit equation:", + "bbox": [ + 511, + 650, + 908, + 801 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nf (\\mathbf {x}) = \\left(\\left(\\frac {x}{s _ {x}}\\right) ^ {\\frac {2}{\\epsilon_ {2}}} + \\left(\\frac {y}{s _ {y}}\\right) ^ {\\frac {2}{\\epsilon_ {2}}}\\right) ^ {\\frac {\\epsilon_ {2}}{\\epsilon_ {1}}} + \\left(\\frac {z}{s _ {z}}\\right) ^ {\\frac {2}{\\epsilon_ {1}}} = 1. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 813, + 903, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extending this representation to a global coordinate system requires 6 additional parameters (3 for translation and 3", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24626", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "for rotation), resulting in a total of 11 parameters per superquadric. Another key property of superquadrics is the ability to compute the radial distance from any point in 3D space to the superquadric surface, i.e., the distance between a point and the superquadric's surface along the line connecting that point to the center of the superquadric. Specifically, given a point $\\mathbf{x} \\in \\mathbb{R}^3$ , its radial distance to the surface of a canonically oriented superquadric is defined as:", + "bbox": [ + 89, + 90, + 483, + 212 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nd _ {r} = | \\mathbf {x} | \\cdot | 1 - f (\\mathbf {x}) ^ {- \\epsilon_ {1} / 2} |, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 224, + 482, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $f(\\mathbf{x})$ is given in Eq. 1. We refer the reader to [53] for the derivation of Eq. 2 and to [16] for a more comprehensive overview on superquadrics.", + "bbox": [ + 89, + 255, + 483, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 316, + 181, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our ultimate goal is a 3D scene decomposition using superquadric primitives. To this end, we first focus on single-object decomposition and then show how our method, combined with 3D instance segmentation [41], can be applied to full 3D scenes. We detail the single-object approach in Sec. 3.1 and its extension to full scenes in Sec. 3.2.", + "bbox": [ + 89, + 342, + 483, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Single Object Decomposition", + "text_level": 1, + "bbox": [ + 89, + 444, + 348, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 2 illustrates our model for single-object decomposition. It consists of two main components: a self-supervised feed-forward neural network that jointly predicts superquadric parameters and a segmentation matrix associating points to superquadrics, followed by a lightweight Levenberg-Marquardt (LM) optimization [20, 28].", + "bbox": [ + 89, + 467, + 483, + 558 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1. Feed-forward Neural Network", + "text_level": 1, + "bbox": [ + 89, + 568, + 348, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our deep learning model draws inspiration from recent fully-supervised Transformer-based [54] segmentation models [5, 6, 41]. These models iteratively decode a sequence of queries, each representing a segmentation mask, by cross-attending to input pixels or points. In our case, the queries represent superquadrics. Next, we show how such an architecture can be adapted to unsupervisedly segment superquadrics, instead of supervisedly segment objects.", + "bbox": [ + 89, + 587, + 483, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Model Details. Given an input point cloud $\\mathcal{P} \\in \\mathbb{R}^{N \\times 3}$ , where each of the $N$ points has a 3D coordinate, we first extract rich point features $\\mathcal{F}_{PC} \\in \\mathbb{R}^{N \\times H}$ using the PVCNN [25] point encoder. At the same time, we initialize $P$ superquadrics features $\\mathcal{F}_{SQ} \\in \\mathbb{R}^{P \\times H}$ with sinusoidal positional encodings. We feed these features in a Transformer decoder [54] which leverages self-attention, cross-attention to the point features, and feed-forward layers to refine them. Once refined, the superquadric features $\\mathcal{F}_{SQ}$ and the point features $\\mathcal{F}_{PC}$ are fed into two prediction heads: The segmentation head takes as input $\\mathcal{F}_{SQ}$ and $\\mathcal{F}_{PC}$ and predicts a soft assignment matrix $M \\in \\mathbb{R}^{N \\times P}$ associating points to", + "bbox": [ + 89, + 717, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "superquadrics and whose elements are defined as:", + "bbox": [ + 513, + 90, + 841, + 106 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nm _ {i j} = \\sigma (\\phi (\\mathcal {F} _ {P C}) \\cdot \\mathcal {F} _ {S Q}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 113, + 903, + 131 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\phi (\\mathcal{F}_{PC})\\in \\mathbb{R}^{N\\times H}$ is a learned projection of the point features to match the dimensionality of the superquadric features, and $\\sigma$ is the softmax function. The second head, the superquadric head, takes the superquadric features $\\mathcal{F}_{SQ}$ as input and predicts 12 parameters for each superquadric: 11 encoding its 5-DoF shape and 6-DoF pose, and one modeling its existence probability $\\alpha$ , enabling a variable number of superquadrics per object.", + "bbox": [ + 511, + 136, + 906, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Losses. We train our model in a self-supervised manner, without requiring any ground truth annotation. Specifically, the total loss is defined as:", + "bbox": [ + 511, + 263, + 905, + 306 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e c}} + \\lambda_ {\\text {p a r}} \\mathcal {L} _ {\\text {p a r}} + \\lambda_ {\\text {e x i s t}} \\mathcal {L} _ {\\text {e x i s t}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 316, + 903, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{rec}$ is the reconstruction loss aligning the predicted superquadrics to the input point cloud $\\mathcal{P}$ , $\\mathcal{L}_{par}$ is the parsimony loss encouraging a small number of primitives, $\\mathcal{L}_{exist}$ is the existence loss, and $\\lambda_{par}, \\lambda_{exist}$ are weighting coefficients. The reconstruction loss $\\mathcal{L}_{rec}$ consists of three terms:", + "bbox": [ + 511, + 340, + 905, + 415 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e c} = \\mathcal {L} _ {\\mathcal {P} \\rightarrow S Q} + \\mathcal {L} _ {S Q \\rightarrow \\mathcal {P}} + \\mathcal {L} _ {N}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 422, + 903, + 440 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The first two terms correspond to the bi-directional Chamfer distance between the input point cloud and the superquadric surfaces, while the third term serves as a regularizer incorporating normal information to improve convergence during training. To compute the Chamfer distance, we approximate each superquadric surface by uniformly sampling $S$ points, following the method of Pilu et al. [35]. Denoting by $d(\\mathbf{x}_i,\\mathbf{x}_{js}^{\\prime})$ the euclidean distance between the $i$ -th point in the input point cloud and the $s$ -th point sampled on the surface of the $j$ -th superquadric, we define $\\mathcal{L}_{\\mathcal{P}\\to SQ}$ as:", + "bbox": [ + 511, + 446, + 906, + 599 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {P} \\rightarrow S Q} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {P} m _ {i j} \\min _ {s \\in [ S ]} d \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {j s} ^ {\\prime}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 604, + 903, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and $\\mathcal{L}_{SQ\\to \\mathcal{P}}$ as:", + "bbox": [ + 513, + 654, + 624, + 670 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S Q \\rightarrow \\mathcal {P}} = \\frac {1}{S \\sum_ {j = 1} ^ {P} \\alpha_ {j}} \\sum_ {j = 1} ^ {P} \\alpha_ {j} \\sum_ {s = 1} ^ {S} \\min _ {i \\in [ N ]} d \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {j s} ^ {\\prime}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 676, + 903, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The last term of Eq. 5, i.e., $\\mathcal{L}_N$ is defined as the reconstruction loss from Yang et al. [57], and is used to incorporate normal information during training which leads to accelerated convergence. Additionally, since we seek not only accuracy but also compactness, we introduce a parsimony loss to encourage the use of fewer primitives. To do that, we optimize the 0.5-norm of $m_j \\coloneqq \\sum_{i=1}^{N} \\frac{m_{ij}}{N}$ and define the parsimony loss as:", + "bbox": [ + 511, + 724, + 905, + 847 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p a r} = \\left(\\frac {1}{P} \\sum_ {j = 1} ^ {P} \\frac {\\sqrt {m _ {j}}}{P}\\right) ^ {2}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 852, + 903, + 904 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24627", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/189e12d07aae7438d31a40affef30c862ea9e9b29fb2f7506ab4e5590c043883.jpg", + "image_caption": [ + "Figure 2. Illustration of the SUPERDEC Model. Given a point cloud of an object with $N$ points, a Transformer-based neural network predicts parameters for $P$ superquadrics, as well as a soft segmentation matrix that assigns points to superquadrics. The predicted parameters include the 11 superquadric parameters and an objectness score. These predictions provide an effective initialization for the subsequent Levenberg-Marquardt (LM) optimization, which refines the superquadrics." + ], + "image_footnote": [], + "bbox": [ + 99, + 85, + 903, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Lastly, we employ an existence loss $\\mathcal{L}_{\\text {exist }}$ which uses the predicted segmentation as a teacher for the linear head in charge of predicting the existence probability. More specifically, given a threshold $\\epsilon_{\\text {exist }}$ , we define the ground-truth existence of the $j$ th superquadric as $\\hat{\\alpha}_j \\coloneqq m_j > \\epsilon_{\\text {exist }}$ and define $\\mathcal{L}_{\\text {exist }}$ as:", + "bbox": [ + 88, + 359, + 483, + 454 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {e x i s t}} = \\sum_ {j = 1} ^ {P} \\frac {B C E \\left(\\alpha_ {j} , \\hat {\\alpha} _ {j}\\right)}{P}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 464, + 482, + 507 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where BCE is the binary cross entropy and $\\alpha_{j}$ is the predicted existence probability for the $j$ th superquadric.", + "bbox": [ + 89, + 518, + 483, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.2. Optimization", + "text_level": 1, + "bbox": [ + 89, + 560, + 230, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our optimization module takes as input the predicted soft segmentation matrix $M$ as well as the superquadric parameters $\\Theta$ , and further refines the superquadric parameters using the Levenberg-Marquardt (LM) [20, 28] algorithm. Specifically, given a point cloud of $N$ points, it iteratively refines the parameters $\\Theta_{j}$ of the $j$ th superquadric, by computing two sets of residuals: The first set of residuals $r_{ij}$ with $i \\in [1, N]$ and $j \\in [1, P]$ is defined as:", + "bbox": [ + 89, + 580, + 483, + 702 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nr _ {i j} = m _ {i j} \\tilde {d} _ {j} \\left(\\mathbf {x} _ {i}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 714, + 482, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tilde{d}_j(\\mathbf{x}_i)$ denotes the radial distance of point $\\mathbf{x}_i$ from the $j$ th superquadric, computed according to Eq. 2. The second set of residuals is used for normalization and is obtained by sampling a set of $K$ points $\\mathbf{p_1},\\ldots ,\\mathbf{p_K}$ on the surface of the given superquadric and then computing the distance of each of them from the point cloud. Specifically, for $i\\in [N + 1,N + K]$ and $j\\in [1,P]$ we compute $r_{ij}$ as:", + "bbox": [ + 89, + 746, + 483, + 854 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nr _ {i j} = \\min _ {k} \\left\\| \\mathbf {p} _ {\\mathbf {i} - \\mathbf {N}} - \\Pi_ {j} \\left(\\mathbf {x} _ {\\mathbf {k}}\\right) \\right\\| _ {2}, \\quad \\text {w i t h} k \\in [ 1, N ]. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 867, + 483, + 888 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Decomposition of Full 3D Scenes", + "text_level": 1, + "bbox": [ + 511, + 359, + 802, + 376 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After training on single objects, extending SUPERDEC to full 3D scenes is straightforward. Given a scene-level point cloud, we extract 3D object instance masks using Mask3D [41]. Each object is centered and uniformly rescaled to a sphere of radius 0.5. We then predict the superquadric primitives for each object individually using our model. We found our model trained on ShapeNet [4] to generalize well on real-world 3D scenes from ScanNet++ [58] and Replica [45] without additional fine-tuning.", + "bbox": [ + 511, + 383, + 906, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 536, + 645, + 553 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first compare our SUPERDEC with previous state-of-the-art methods on individual objects and full 3D scenes (Sec. 4.1). We then demonstrate the usefulness of our representation on down-stream applications for robotics and controllable image generation (Sec. 4.2). Finally, in Sec. 4.3, we present additional analyses on part segmentation and the implicit learning of shape categories, followed by a study of the compactness-accuracy trade-off and runtime.", + "bbox": [ + 511, + 561, + 906, + 684 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Comparing with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 511, + 696, + 870, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets. We compare on three different datasets: ShapeNet [4]: We use the 13 classes of the ShapeNet subset and train-val-test splits as defined in Choy et al. [8]. All objects are pre-aligned in a canonical orientation. For each object we sample 4096 points using Farthest Point Sampling (FPS) [36]. ShapeNet is a widely used dataset and is well-suited for comparison with existing baselines.", + "bbox": [ + 511, + 719, + 905, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ScanNet++ [58]: We further evaluate our model on real-world object scans from the ScanNet++ validation set. Each object is extracted using ground truth mask annotations, and 4096 points per object are sampled with FPS. In contrast to ShapeNet, these object point clouds are noisier, partially ob", + "bbox": [ + 511, + 825, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24628", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/478549b1dd75314af136d61fa0c1a14474e84396a4e494dd4168e8eba492c3d6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelPrimitive TypeSegmentationIn-categoryOut-of-category
L1 ↓L2 ↓# Prim.↓L1 ↓L2 ↓# Prim.↓
EMS (Liu et al.) [23]SuperquadricsX5.7711.3455.685.4101.2115.68
CSA (Yang et al.) [57]Cuboids✓5.1570.5279.214.8970.42711.75
SQ (Paschalidou et al.) [32]SuperquadricsX3.6680.279104.1930.3549
SUPERDEC (Ours)Superquadrics✓1.6980.0475.81.8470.0615.26
", + "bbox": [ + 91, + 88, + 903, + 188 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Quantitative Results on ShapeNet [4]. We show scores for in-category and out-of-category experiments and are scaled by $10^{2}$ .", + "bbox": [ + 91, + 190, + 901, + 205 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4dec65a8552e92cd98987fe6a1aa90b6ed09ba9c3a8f3067ec3907771b35d527.jpg", + "image_caption": [ + "Figure 3. Qualitative Results on ShapeNet [4]. We show results on test samples for in-category (four first columns) classes and out-of-category classes (two last columns). The latter were not seen during training and illustrate how well models generalize to novel classes." + ], + "image_footnote": [], + "bbox": [ + 93, + 210, + 897, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "served, and subject to random orientation and translation, providing a more realistic and challenging evaluation setting for our method.", + "bbox": [ + 88, + 635, + 482, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Replica [45]: Finally, we present both quantitative and qualitative results on full 3D scenes from Replica. Object instances are extracted using either ground truth instance annotations or the pre-trained 3D instance segmentation model Mask3D [41], enabling us to evaluate our approach in a fully realistic setting, including scenarios where no ground truth annotations are available.", + "bbox": [ + 88, + 680, + 482, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Methods in comparison. We compare to learning- and optimization-based prior works using both cuboids and superquadrics as geometric primitives. $SQ$ [32] is a learning-based approach for object-level decomposition using superquadrics; it takes a voxel grid as input and predicts superquadric primitives via a CNN. CSA [57] is another learning-based method but uses cuboids as geometric prim", + "bbox": [ + 88, + 794, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ities. It takes a point cloud as input and predicts cuboid parameters from a global latent code. Lastly, EMS [23] is an optimization-based approach that decomposes objects by hierarchically fitting superquadrics to parts of a point cloud.", + "bbox": [ + 511, + 635, + 903, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Details. Our goal is to develop a general-purpose, class-agnostic model capable of representing arbitrary objects as superquadrics. Existing methods typically train separate models for each object class, assuming that all classes are known in advance and that sufficient training data is available for each. These assumptions, however, often fail in real-world scenarios. To address this, we jointly train a single model on all 13 ShapeNet classes using the publicly available code of prior methods, moving towards a more realistic class-agnostic solution. In our model we set the following hyper-parameters $P = 16$ , $S = 4096$ , $K = 25$ , $H = 128$ , $D = 3$ , $\\epsilon_{\\text{exist}} = 24$ , $\\lambda_{\\text{exist}} = 0.01$ , $\\lambda_{\\text{par}} = 0.06$ . We refer to the supplementary for additional details.", + "bbox": [ + 511, + 703, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24629", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8dc5d127fb1af53e239f115a86ab8ed9cd381d24c83f9f7c72f1eb67c56228c8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ScanNet++ [58]Replica [45]
MethodL1 ↓L2 ↓#Prim.↓L1 ↓L2 ↓#Prim.↓
SQ [32]10.454.2610.011.845.9110
EMS [23]5.512.114.255.402.123.61
CSA [57]2.910.4111.643.680.709.63
SUPERDEC (Ours)1.700.115.181.790.196.58
", + "bbox": [ + 93, + 88, + 478, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. We assess reconstruction accuracy using L1 and L2 Chamfer distances and compactness by the average number of geometric primitives.", + "bbox": [ + 89, + 229, + 483, + 276 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.1. Results on ShapeNet", + "text_level": 1, + "bbox": [ + 89, + 294, + 282, + 309 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We show scores in Tab. 1 and qualitative results in Fig. 3. To evaluate both accuracy and generalization, we conduct two experiments: in-category and out-of-category. In the in-category setting, all learning-based methods are jointly trained on the 13 classes of the ShapeNet training set and evaluated on the corresponding test set. In the out-of-category setting, models are trained on half of the categories (airplane, bench, chair, lamp, rifle, table) and tested on the remaining ones (car, sofa, loudspeaker, cabinet, display, telephone, watercraft). Our SUPERDEC model significantly outperforms both learned and non-learned baselines. Compared to learned baselines, we reduce the L2 loss by a factor of six while using nearly half the number of primitives, supporting our hypothesis that leveraging local point features as opposed to a single global descriptor improves 3D decomposition in both accuracy and compactness. Compared to the non-learned baseline, we predict a similar number of primitives but achieve an L2 loss approximately 20 times smaller, validating the benefit of learning shape priors to avoid local minima that often hinder purely optimization-based approaches.", + "bbox": [ + 89, + 316, + 483, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.2. Results on 3D Scenes", + "text_level": 1, + "bbox": [ + 89, + 652, + 284, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, models are evaluated on real-world, out-of-category objects, which appear in arbitrary orientations and often exhibit incomplete point clouds due to reconstruction artifacts and occlusions. Tab. 2 shows the quantitative results. Despite never being trained on real-world objects, our method outperforms both the optimization- and the learning-based baselines by a large margin. Lastly, we qualitatively evaluate our pipeline on full 3D scenes from Replica, where our object-level model is applied on top of class-agnostic instance segmentation predictions from Mask3D [41]. As shown in Fig. 1, our method effectively reconstructs object shapes, even under noisy segmentation masks and geometries that differ substantially from those seen during training. We refer to the supplementary for additional qualitative results.", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c1c4c529791fd4d774efcc77fa5ec4c35fabedfa5e9f67f567e73929e97d8e32.jpg", + "table_caption": [ + "Table 2. Quantitative Results on Scenes. Evaluation on objects from 3D scene datasets [45, 58]. Scores are scaled by $10^{2}$ ." + ], + "table_footnote": [], + "table_body": "
MethodTime (ms)Suc. (%)Mem. (MB)
Occupancy0.056100.000.873
PointCloud0.06389.5719.286
Voxels0.03098.780.101
Cuboids [37]0.12061.230.024
SUPERDEC0.15091.710.042
", + "bbox": [ + 526, + 88, + 893, + 188 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Path Planning Scores. Mean on 15 ScanNet++ scenes.", + "bbox": [ + 517, + 191, + 900, + 205 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/43864e64c65f7c33f8e0a90b6592b8790c4f5fc401010af062544c0af5d85ecc.jpg", + "image_caption": [ + "Figure 4. Grasping Result. Visualization of computed grasp poses for a milk bottle, some flowers, a side table, and a plant." + ], + "image_footnote": [], + "bbox": [ + 517, + 210, + 903, + 347 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Down-stream Applications", + "text_level": 1, + "bbox": [ + 511, + 386, + 756, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next, we show the versatility of the SUPERDEC representation for downstream applications, including robotics tasks such as path planning and object grasping (Sec. 4.2.1), and controllable image generation (Sec. 4.2.2).", + "bbox": [ + 511, + 409, + 906, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1. Robotics", + "text_level": 1, + "bbox": [ + 511, + 479, + 620, + 492 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Path planning seeks to compute a collision-free shortest path between a given start and end point in 3D space, enabling efficient robot navigation. Although essential for traversing large environments, it typically demands storing large-scale 3D representations. Here, we assess whether our compact representation can perform this task effectively while reducing memory requirements. We conduct experiments on 15 ScanNet++ [58] scenes, comparing SUPERDEC to common 3D representations, including dense occupancy grids, point clouds, voxel grids, and cuboids [37]. As shown in Tab. 6, SUPERDEC not only reduces memory consumption compared to traditional representations but also achieves a higher success rate than dense point clouds. Further details about experiment setup, metrics and analysis are provided in the Appendix.", + "bbox": [ + 511, + 498, + 906, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Object Grasping enables robots to grasp real-world objects by computing suitable grasping poses. Existing methods fall into two categories, each with complementary limitations. Geometry-based approaches [3, 12, 30] require precise 3D object models, which are often unavailable in real-world scenarios. Learning-based approaches [27, 47, 56] operate directly on raw sensor data but tend to be biased towards training data, which typically consists of tabletop scenes with small, convex, or low-genus objects [11, 46]. To overcome these limitations, SuperQ-GRASP [50] explored decomposing objects into explicit primitives. How", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24630", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/96ccec016ec636580169833e21ee26e396184942cbfe518eac6d358d802342e4.jpg", + "image_caption": [ + "Input Point Cloud" + ], + "image_footnote": [], + "bbox": [ + 93, + 97, + 287, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3b73dddd69eb9d8ff84dcabf84c9d6171d5ca57bfb6c698c1b2cb323c63503e7.jpg", + "image_caption": [ + "Superquadrics" + ], + "image_footnote": [], + "bbox": [ + 287, + 97, + 480, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/872c27960f652358d2624d27a8fbb67cf6b9d439573594ac947f32f90445e908.jpg", + "image_caption": [ + "Figure 5. Real-world robot experiment. The top row shows the input scan (left) and the representation from SUPERDEC with the computed path and grasping pose (right). The bottom row illustrates the robot following the planned path. We denote the starting point of the path with a green sphere, and the target location with a red sphere. The target object (a milk bottle) is circled in red." + ], + "image_footnote": [], + "bbox": [ + 93, + 183, + 480, + 308 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ever, its reliance on Marching Primitives [24] to obtain superquadrics from the object's Signed Distance Function (SDF) makes it unsuitable for most real-world cases where only point clouds are available. In contrast, our approach directly processes point clouds of entire scenes and, when combined with the class-agnostic segmentations from Mask3D [41], extracts superquadrics for all objects. Given the superquadric parameters, we employ a superquadric-based geometric method [55] to compute grasping poses for selected objects. Fig. 4 shows predicted grasping poses on objects from a real-world 3D scan of a room. In practice, our method eliminates the need for data-driven grasping models while remaining adaptable to diverse object shapes and producing high-quality grasping poses.", + "bbox": [ + 89, + 411, + 483, + 623 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Real-world Experiment. Finally, we demonstrate the real-world applicability of our superquadric-based representation by deploying it on a legged robot (Boston Dynamics Spot) equipped with an arm, supporting both motion planning and object grasping in an indoor environment. We scan a scene using a 3D scanning application on an iPad, extract a dense point cloud, and run SUPERDEC on it. Given the robot's starting position and a specified target object (a milk bottle), we compute both the path and the grasping pose as described earlier (see Fig. 4), enabling the robot to approach and successfully grasp the object. Fig. 5 shows the computed representation, the planned trajectory, the grasping pose, and a frame from the real-world demonstration. This experiment suggests that integrating SUPERDEC with open-vocabulary segmentation methods such as OpenMask3D [48] could allow robots to navigate to and grasp arbitrary objects specified via natural-language prompts.", + "bbox": [ + 89, + 643, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e67dd82e63110d293547f0ef7d56641338114490b24f5f21c59e377da3a618b1.jpg", + "image_caption": [ + "Original", + "Editing", + "Addition", + "Deletion" + ], + "image_footnote": [], + "bbox": [ + 516, + 99, + 903, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2143ccb01f4229cb1edb7587b0fe971d16e0e17f835fcd8eec0480af3115e7cb.jpg", + "image_caption": [ + "Figure 6. Spatial control using SUPERDEC. Top row shows superquadrics generated by SUPERDEC, bottom row shows generated images using the prompt 'A corner of a room with a plant'." + ], + "image_footnote": [], + "bbox": [ + 516, + 183, + 903, + 268 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/549a00a9f8ad1734a9a270a4598af061adff061ce951c65e7b9e10375d1407f7.jpg", + "image_caption": [ + "Superquadrics", + "\"Pink living room\"" + ], + "image_footnote": [], + "bbox": [ + 516, + 353, + 709, + 450 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2fa33d0f2387e0deaa6bfa7c02eebbd4b83976db59e761379758333d07597226.jpg", + "image_caption": [ + "Depth Prompt", + "\"Modern living room\"" + ], + "image_footnote": [], + "bbox": [ + 712, + 353, + 901, + 450 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3e59075a06027fbc2bd7003279a192ddc6c52423476fa057dd559f77108d379c.jpg", + "image_caption": [ + "Figure 7. Semantic control using SUPERDEC. Top row: superquadrics created by our SUPERDEC, and depth map to prompt the generation of text-to-image diffusion model. Bottom row: generations with two different textual prompts." + ], + "image_footnote": [], + "bbox": [ + 517, + 465, + 709, + 563 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5eb20ce3b4c02010a3963f79146e31a91393b7ab52a8e40fda7ec8b8044582cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 465, + 901, + 563 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2. Controllable Generation and Editing", + "text_level": 1, + "bbox": [ + 511, + 647, + 815, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We investigate how the SUPERDEC representation can be used to introduce joint spatial and semantic control in text-to-image diffusion models [39]. Specifically, we generate images by conditioning ControlNet [60] on depth maps rendered from the superquadrics extracted from Replica [45] scenes. Qualitative results are shown in Fig. 6 and Fig. 7. Fig. 6 demonstrates spatial control: by moving, duplicating, or removing superquadrics corresponding to a plant, we coherently influence the generated images. Fig. 7 highlights semantic control: we can vary the room's style while preserving its semantic and geometric structure, and observe that object semantics naturally emerge from the spatial arrangement of superquadrics without explicit conditioning, e.g., pillows appear on couches, and a plant is placed on a central table, reflecting plausible real-world arrangements.", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24631", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4ee67679311c87aefdbf97b48165d576960955d6dace3c71d86dd1c6f9f1a42c.jpg", + "image_caption": [ + "Figure 8. Qualitative Results on ShapeNet [4] segmentation. We show the resulting segmentation matrices on test samples for in-category (four first columns) classes and out-of-category classes (two last columns). The latter were not seen during training." + ], + "image_footnote": [], + "bbox": [ + 94, + 122, + 465, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e2ed7ba8ef746fe51ffd49bb3304a34ba9e753e7338d1f8610a96291247c4f26.jpg", + "image_caption": [ + "Input 3D Scan" + ], + "image_footnote": [], + "bbox": [ + 91, + 243, + 223, + 311 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/685b7d9f74bb6c0449748e930066efc7dd478d39a43152c54b9839facfbbf6e2.jpg", + "image_caption": [ + "$1^{\\mathrm{st}}$ Hierarchy", + "$2^{\\mathrm{nd}}$ Hierarchy", + "Figure 9. Hierarchical decomposition with SUPERDEC. From left to right we visualize: the input point cloud, the superquadrics predicted in the first hierarchy, the superquadrics predicted in the second hierarchy." + ], + "image_footnote": [], + "bbox": [ + 225, + 243, + 480, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Analysis Experiments", + "text_level": 1, + "bbox": [ + 89, + 414, + 294, + 429 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Unsupervised part segmentation. Besides superquadric parameters, our method also predicts a segmentation matrix which segments the initial point cloud into parts that are fitted to the predicted superquadrics. In Fig. 8, we visualize the predicted segmentation masks for the same examples shown in Fig. 3. We observe that segmentation masks, especially in the in-category experiments, appear very sharp. This suggests that our method, especially if trained at a larger scale, can be leveraged for different applications as geometry-based part segmentation or as pretraining for supervised semantic part segmentation.", + "bbox": [ + 89, + 436, + 483, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Can we do hierarchical decomposition? We run SUPERDEC in a hierarchical fashion on a whole simple scene, without using any instance segmentation method to segment the scene into different objects. We observed that by applying SUPERDEC on the whole point cloud it can coarsely segment the scene into components such as grass, table, and chairs. By applying the model again on the points segmented during the first prediction, we obtained a more accurate decomposition of the individual object instances. Given that SUPERDEC was trained only on single object instances from ShapeNet [4] we found this results very interesting. We see future potential in the use of our model to obtain unsupervised hierarchical decompositions.", + "bbox": [ + 89, + 621, + 483, + 818 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "What does our network learn? Since our network performs unsupervised part segmentation, we analyze the features learned by the Transformer decoder across object classes. Inspired by BERT [10]'s [CLS] token, we append a learnable embedding to the sequence of embedded", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/927a6ac7e941d7c142444a279e78b7f93a0eda48241105913c330a7785d765b1.jpg", + "image_caption": [ + "Figure 10. t-SNE Visualization of Primitive Embeddings across different ShapeNet classes." + ], + "image_footnote": [], + "bbox": [ + 560, + 75, + 864, + 265 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "superquadrics; although never explicitly decoded, this embedding is refined through self- and cross-attention. After training, we extract and visualize these embeddings using t-SNE [52] for ShapeNet [4] categories (Fig. 10). We observe that categories with consistent shapes, such as chairs, airplanes, and cars, form clear clusters, while categories with high intra-class variability, such as watercraft, are more dispersed. This indicates that our model organizes objects by geometric structure without requiring class annotations.", + "bbox": [ + 511, + 324, + 906, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "How fast is our method? Our model is highly parallelizable, allowing multiple objects to be batched and processed simultaneously in a single forward pass. On an RTX 4090 (24 GB), we can process up to 256 objects in parallel. On average, the forward pass takes $0.13\\mathrm{s}$ for a complete Replica [45] scene, 3D instance segmentation with Mask3D [41] requires $0.3\\mathrm{s}$ , and each LM optimization step takes less than 1 s, see supplementary for more details.", + "bbox": [ + 511, + 468, + 905, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 599, + 633, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We proposed SUPERDEC, a method for deriving compact yet expressive 3D scene representations based on simple geometric primitives – specifically, superquadrics. Our model outperforms prior primitive-based methods and generalizes well to out-of-category classes. We further demonstrated the potential of the resulting 3D scene representation for various applications in robotics, and as a geometric prompt for diffusion-based image generation. While this is only a first step towards more compact, geometry-aware 3D scene representations, we anticipate broader applications and expect to see further research in this direction.", + "bbox": [ + 511, + 619, + 906, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. Elisabetta Fedele is supported by the ETH AI Center doctoral fellowship, by the Swiss National Science Foundation (SNSF) Advanced Grant 216260 (Beyond Frozen Worlds: Capturing Functional 3D Digital Twins from the Real World), and an SNSF Mobility Grant. Francis Engelmann is supported by an SNSF PostDoc.mobility grant. We also gratefully acknowledge a GPU grant from NVIDIA.", + "bbox": [ + 511, + 795, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "In-category", + "bbox": [ + 181, + 104, + 254, + 116 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Out-of-category", + "bbox": [ + 359, + 104, + 457, + 116 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24632", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Stephan Alaniz, Massimiliano Mancini, and Zeynep Akata. Iterative superquadric decomposition of 3d objects from multiple viewsd. In International Conference on Computer Vision (ICCV), 2023. 1", + "[2] Alan H. Barr. Superquadrics and angle-preserving transformations. IEEE Computer Graphics and Applications, 1981. 2", + "[3] Junhao Cai, Jingcheng Su, Zida Zhou, Hui Cheng, Qifeng Chen, and Michael Yu Wang. Volumetric-based contact point detection for 7-dof grasping. In Conference on Robot Learning (CoRL), 2022. 6", + "[4] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical report, 2015. 2, 4, 5, 8", + "[5] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention Mask Transformer for Universal Image Segmentation. International Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3", + "[6] Bowen Cheng, Alexander G. Schwing, and Alexander Kirillov. Per-Pixel Classification is Not All You Need for Semantic Segmentation. In International Conference on Neural Information Processing Systems (NeurIPS), 2021. 2, 3", + "[7] Laurent Chevalier, Fabrice Jaillet, and Atilla Baskurt. Segmentation and superquadric modeling of 3d objects. In International Conference in Central Europe on Computer Graphics and Visualization, 2003. 2", + "[8] Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3D-R2N2: A Unified Approach for Single and Multi-view 3D Object Reconstruction. In European Conference on Computer Vision (ECCV), 2016. 4", + "[9] Alexandros Delitzas, Ayca Takmaz, Federico Tombari, Robert Sumner, Marc Pollefeys, and Francis Engelmann. SceneFun3D: Fine-Grained Functionality and Affordance Understanding in 3D Scenes. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1", + "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies., 2019. 8", + "[11] Hao-Shu Fang, Chenxi Wang, Hongjie Fang, Minghao Gou, Jirong Liu, Hengxu Yan, Wenhai Liu, Yichen Xie, and Cewu Lu. Anygrasp: Robust and efficient grasp perception in spatial and temporal domains. IEEE Transactions on Robotics, 2023. 6", + "[12] B. Faverjon and J. Ponce. On Computing Two-finger Force-closure Grasps of Curved 2D Objects. In International Conference on Robotics and Automation (ICRA), 1991. 6", + "[13] Qiao Gu, Ali Kuwajerwala, Sacha Morin, Krishna Murthy Jatavallabhula, Bipasha Sen, Aditya Agarwal, Corban Rivera, William Paul, Kirsty Ellis, Rama Chellappa, et al." + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conceptgraphs: Open-vocabulary 3d scene graphs for perception and planning. In International Conference on Robotics and Automation (ICRA), 2024. 1", + "[14] Abdullah Hamdi, Luke Melas-Kyriazi, Jinjie Mai, Guocheng Qian, Ruoshi Liu, Carl Vondrick, Bernard Ghanem, and Andrea Vedaldi. Ges: Generalized exponential splatting for efficient radiance field rendering. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[15] Jan Held, Renaud Vandeghen, Abdullah Hamdi, Adrien Deli'ege, Anthony Cioppa, Silvio Giancola, Andrea Vedaldi, Bernard Ghanem, and Marc Van Droogenbroeck. 3D convex splatting: Radiance field rendering with 3D smooth convexes. ArXiv, 2024. 2", + "[16] Ales Jaklic, Ales Leonardis, and Franc Solina. Segmentation and recovery of superquadrics. Springer Science & Business Media, 2000. 3, 1", + "[17] Rasmus Ramsbøl Jensen, A. Dahl, George Vogiatzis, Engil Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. International Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 2", + "[18] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions On Graphics (TOG), 2023. 1, 2", + "[19] Sebastian Koch, Narunas Vaskevicius, Mirco Colosi, Pedro Hermosilla, and Timo Ropinski. Open3DSG: Open-vocabulary 3D Scene Graphs from Point Clouds with Queryable Objects and Open-set Relationships. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1", + "[20] Kenneth Levenberg. A method for the solution of certain non-linear problems in least squares. Quarterly of Applied Mathematics, 1944. 3, 4", + "[21] Lei Li and Angela Dai. Genzi: Zero-shot 3d human-scene interaction generation. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1", + "[22] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. International Conference on Computer Vision (ICCV), 2019. 2", + "[23] Weixiao Liu, Yuwei Wu, Sipu Ruan, and Gregory S Chirikjian. Robust and Accurate Superquadric Recovery: A Probabilistic Approach. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 5, 6", + "[24] Weixiao Liu, Yuwei Wu, Sipu Ruan, and Gregory S Chirikjian. Marching-primitives: Shape abstraction from signed distance function. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 7", + "[25] Zhijian Liu, Haotian Tang, Yujun Lin, and Song Han. PointVoxel CNN for Efficient 3D Deep Learning. In International Conference on Neural Information Processing Systems (NeurIPS), 2019. 3", + "[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. International Conference on Learning Representations (ICLR), 2019. 1" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24633", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Jeffrey Mahler, Matthew Matl, Vishal Satish, Michael Danielczuk, Bill DeRose, Stephen McKinley, and Ken Goldberg. Learning ambidextrous robot grasping policies. Science Robotics, 2019. 6", + "[28] Donald W. Marquardt. An algorithm for least-squares estimation of nonlinear parameters. Journal of The Society for Industrial and Applied Mathematics, 1963. 3, 4", + "[29] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 1", + "[30] A.T. Miller, S. Knoop, H.I. Christensen, and P.K. Allen. Automatic Grasp Planning Using Shape Primitives. In International Conference on Robotics and Automation (ICRA), 2003. 6", + "[31] Tom Monnier, Jake Austin, Angjoo Kanazawa, Alexei A. Efros, and Mathieu Aubry. Differentiable blocks world: Qualitative 3d decomposition by rendering primitives. International Conference on Neural Information Processing Systems (NeurIPS), 2023. 1, 2", + "[32] Despoina Paschalidou, Ali Osman Ulusoy, and Andreas Geiger. Superquadrics revisited: Learning 3d shape parsing beyond cuboids. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2, 5, 6", + "[33] Songyou Peng, Kyle Genova, Chiyu Max Jiang, Andrea Tagliasacchi, Marc Pollefeys, and Thomas Funkhouser. OpenScene: 3D Scene Understanding with Open Vocabularies. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1", + "[34] Alex P Pentland. Parts: structured descriptions of shape. In Proceedings of the Fifth AAAI National Conference on Artificial Intelligence, 1986. 2", + "[35] Maurizio Pilu and Robert B Fisher. Equal-distance sampling of superellipse models. In British Machine Vision Conference (BMVC), 1995. 3", + "[36] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space. International Conference on Neural Information Processing Systems (NeurIPS), 2017. 4", + "[37] Michael Ramamonjisoa, Sinisa Stekovic, and Vincent Lepetit. MonteBoxFinder: Detecting and Filtering Primitives to Fit a Noisy Point Cloud. In European Conference on Computer Vision (ECCV), 2022. 6, 5", + "[38] Aaron Ray, Christopher Bradley, Luca Carlone, and Nicholas Roy. Task and motion planning in hierarchical 3d scene graphs. arXiv preprint arXiv:2403.08094, 2024. 1", + "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. International Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 7", + "[40] Zuria Bauer Marc Pollefeys Federico Tombari Leonidas Guibas Gao Huang Francis Engelmann Rui Huang, Guangyao Zhai. VIPs: Video Perception Model for 3D Scene Generation. arXiv preprint, 2025. 1" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask transformer for 3d semantic instance segmentation. International Conference on Robotics and Automation (ICRA), 2022. 2, 3, 4, 5, 6, 7, 8", + "[42] Jonas Schult, Sam Tsai, Lukas Hollein, Bichen Wu, Jialiang Wang, Chih-Yao Ma, Kunpeng Li, Xiaofang Wang, Felix Wimbauer, Zijian He, et al. Controlroom3d: Room generation using semantic proxy rooms. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1", + "[43] Jaidev Shriram, Alex Trevithick, Lingjie Liu, and Ravi Ramamoorthi. Realmdreamer: Text-driven 3d scene generation with inpainting and depth diffusion. arXiv preprint arXiv:2404.07199, 2024. 1", + "[44] Franc Solina and Ruzena Bajcsy. Recovery of parametric models from range images: The case for superquadrics with global deformations. Transactions on Pattern Analysis and Machine Intelligence (PAMI), 1990. 2", + "[45] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Yuheng Ren, Shobhit Verma, Anton Clarkson, Ming Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke Malte Strasdat, Renzo De Nardi, Michael Goesele, S. Lovegrove, and Richard A. Newcombe. The replica dataset: A digital replica of indoor spaces. ArXiv, 2019. 2, 4, 5, 6, 7, 8, 1, 3", + "[46] Matan Sudry, Tom Jurgenson, Aviv Tamar, and Erez Karpas. Hierarchical planning for rope manipulation using knot theory and a learned inverse model. In Conference on Robot Learning, 2023. 6", + "[47] Martin Sundermeyer, Arsalan Mousavian, Rudolph Triebel, and Dieter Fox. Contact-grasnet: Efficient 6-dof grasp generation in cluttered scenes. In International Conference on Robotics and Automation (ICRA), 2021. 6", + "[48] Ayca Takmaz, Elisabetta Fedele, Robert W. Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. OpenMask3D: Open-Vocabulary 3D Instance Segmentation. In International Conference on Neural Information Processing Systems (NeurIPS), 2023. 1, 7", + "[49] Ayca Takmaz, Alexandros Delitzas, Robert W. Sumner, Francis Engelmann, Johanna Wald, and Federico Tombari. Search3D: Hierarchical Open-Vocabulary 3D Segmentation. IEEE Robotics and Automation Letters (RA-L), 2025. 1", + "[50] Xun Tu and Karthik Desingh. Superq-grasp: Superquadrics-based grasp pose estimation on larger objects for mobile-manipulation. arXiv, 2024. 6", + "[51] Shubham Tulsiani, Hao Su, Leonidas J. Guibas, Alexei A. Efros, and Jitendra Malik. Learning shape abstractions by assembling volumetric primitives. International Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1, 2", + "[52] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. Journal of machine learning research, 2008. 8" + ], + "bbox": [ + 516, + 90, + 905, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24634", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Erik Roeland van Dop and Paulus P.L. Regtien. Fitting undeformed superquadrics to range data: improving model recovery and classification. International Conference on Computer Vision and Pattern Recognition (CVPR), 1998. 3", + "[54] Ashish Vaswani, Noam M. Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is All you Need. In International Conference on Neural Information Processing Systems (NeurIPS), 2017. 3", + "[55] Giulia Vezzani, Ugo Pattacini, and Lorenzo Natale. A grasping approach based on superquadric models. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 1579-1586. IEEE, 2017. 7", + "[56] Chenxi Wang, Hao-Shu Fang, Minghao Gou, Hongjie Fang, Jin Gao, and Cewu Lu. Graspness Discovery in Clutters for Fast and Accurate Grasp Detection. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6", + "[57] Kaizhi Yang and Xuejin Chen. Unsupervised learning for cuboid shape abstraction via joint segmentation from point clouds. ACM Transactions On Graphics (TOG), 2021. 1, 2, 3, 5, 6", + "[58] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. International Conference on Computer Vision (ICCV), 2023. 2, 4, 6, 1, 3, 5", + "[59] Chenyangguang Zhang, Alexandros Delitzas, Fangjinhua Wang, Ruida Zhang, Xiangyang Ji, Marc Pollefeys, and Francis Engelmann. Open-Vocabulary Functional 3D Scene Graphs for Real-World Indoor Spaces. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 1", + "[60] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. International Conference on Computer Vision (ICCV), 2023. 7" + ], + "bbox": [ + 91, + 92, + 480, + 584 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24635", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SuperDec_ 3D Scene Decomposition with Superquadrics Primitives/ed2eb095-495d-4dd3-9f36-9c1449d1ce9d_model.json b/2025/SuperDec_ 3D Scene Decomposition with Superquadrics Primitives/ed2eb095-495d-4dd3-9f36-9c1449d1ce9d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5ce8c3248967b86c0b654d3debc986eb8c67a690 --- /dev/null +++ b/2025/SuperDec_ 3D Scene Decomposition with Superquadrics Primitives/ed2eb095-495d-4dd3-9f36-9c1449d1ce9d_model.json @@ -0,0 +1,2543 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.156, + 0.131, + 0.843, + 0.154 + ], + "angle": 0, + "content": "SUPERDEC: 3D Scene Decomposition with Superquadric Primitives" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.165, + 0.26, + 0.185 + ], + "angle": 0, + "content": "Elisabetta Fedele\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.281, + 0.166, + 0.388, + 0.186 + ], + "angle": 0, + "content": "Boyang Sun" + }, + { + "type": "text", + "bbox": [ + 0.41, + 0.166, + 0.555, + 0.185 + ], + "angle": 0, + "content": "Leonidas Guibas" + }, + { + "type": "text", + "bbox": [ + 0.572, + 0.166, + 0.712, + 0.186 + ], + "angle": 0, + "content": "Marc Pollefeys1,3" + }, + { + "type": "text", + "bbox": [ + 0.734, + 0.166, + 0.898, + 0.186 + ], + "angle": 0, + "content": "Francis Engelmann" + }, + { + "type": "text", + "bbox": [ + 0.331, + 0.199, + 0.414, + 0.214 + ], + "angle": 0, + "content": "\\(^{1}\\)ETH Zurich" + }, + { + "type": "text", + "bbox": [ + 0.442, + 0.199, + 0.567, + 0.215 + ], + "angle": 0, + "content": "\\(^{2}\\)Stanford University" + }, + { + "type": "text", + "bbox": [ + 0.595, + 0.199, + 0.665, + 0.215 + ], + "angle": 0, + "content": "3Microsoft" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.25, + 0.327, + 0.266 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.283, + 0.484, + 0.51 + ], + "angle": 0, + "content": "We present SUPERDEC, an approach for creating compact 3D scene representations via decomposition into superquadric primitives. While most recent methods use geometric primitives to obtain photorealistic 3D reconstructions, we instead leverage them to obtain a compact yet expressive representation. To this end, we design a novel architecture that efficiently decomposes point clouds of arbitrary objects into a compact set of superquadrics. We train our model on ShapeNet and demonstrate its generalization capabilities on object instances from ScanNet++ as well as on full Replica scenes. Finally, we show that our compact superquadric-based representation supports a wide range of downstream applications, including robotic manipulation and controllable visual content generation. Project page: https://super-dec.github.io." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.541, + 0.222, + 0.557 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.483, + 0.734 + ], + "angle": 0, + "content": "3D scene representations play a central role in computer vision and robotics, enabling tasks such as 3D scene understanding [33, 48, 49], scene generation [40, 42, 43], functional reasoning [19, 59], and scene interaction [9, 13, 21, 38]. Recent work [18] utilizes 3D Gaussians as geometric primitives to produce high-quality, photorealistic reconstructions. However, such representations are often memory-intensive. In contrast, we propose a more lightweight yet geometrically faithful 3D scene representation by decomposing the input point cloud into a compact set of explicit primitives, namely superquadrics (Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Representations for 3D scenes include well-established formats such as point clouds, meshes, signed distance functions, and voxel grids, each offering different trade-offs among geometric detail, computational cost, resolution, performance, interpretability, and editability. Recently, multi-view approaches such as Neural Radiance Fields (NeRF) [29] and Gaussian Splatting (GS) [18] have gained popularity as 3D scene representations. These methods optimize photometric losses to ensure that their underlying representations (implicit in the case of NeRF, and explicit in the case of GS) are consistent with observed images. While" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.249, + 0.71, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.249, + 0.904, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.362, + 0.71, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.362, + 0.904, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.478, + 0.907, + 0.534 + ], + "angle": 0, + "content": "Figure 1. 3D Scene Decomposition with Superquadrics. Given a 3D point cloud of an indoor scene, SUPERDEC decomposes all scene objects into a compact set of superquadric primitives. Different object instances are visualized with different colors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.563, + 0.907, + 0.623 + ], + "angle": 0, + "content": "these approaches excel at achieving photorealism, they lack explicit control over compactness, often resulting in large, non-modular scene encodings that are unsuitable for tasks requiring explicit spatial reasoning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.626, + 0.909, + 0.838 + ], + "angle": 0, + "content": "While optimizing compactness on a scene level remains a challenging task, prior work has shown that geometric primitives such as cuboids [51, 57] or superquadrics [1, 31, 32] enable compact and interpretable decompositions of individual objects. These methods are generally either learning-based [32, 57], favoring speed at the cost of geometric accuracy, or optimization-based [1, 23, 31], offering improved accuracy at the expense of higher computational cost. Although both types of approaches can be effective for certain object classes, they often struggle to generalize across datasets with diverse object geometries. Learning-based methods typically require category-specific training, while optimization-based ones rely on hand-crafted heuristics, limiting their scalability in open-world settings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Motivated by the abstraction capabilities of geometric primitives for individual object categories, we propose to represent complex 3D scenes using a compact set of superquadrics. To this end, we learn class-agnostic object" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24625" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "level shape priors to optimize compactness and leverage an off-the-shelf 3D instance segmentation method (Mask3D [41]) to scale our approach to full 3D scenes." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.138, + 0.483, + 0.468 + ], + "angle": 0, + "content": "We choose superquadrics as building blocks for our representation, as they offer more accurate shape modeling than cuboids while incurring minimal additional parameter overhead (9 v.s. 11, including 6-DoF pose parameters). To develop a model capable of generalizing across diverse object types, we draw inspiration from supervised segmentation [5, 6, 41], and approach the problem from the perspective of unsupervised geometric-based segmentation, using local point-based features to iteratively refine predicted geometric primitives. Our model is trained on ShapeNet [4] and evaluated on three challenging and diverse 3D datasets: ShapeNet [4], ScanNet++ [58], and Replica [45]. On the 3D object dataset ShapeNet, our approach achieves an L2 error six times smaller than prior state-of-the-art methods [32], while requiring only half the number of primitives. On ScanNet++ and Replica, we demonstrate that our method generalizes well to real-world scene-level settings, despite being trained solely on ShapeNet. Finally, we show the practical utility of our method as a scene representation for robotic tasks, including path planning and object grasping, as well as an editable 3D scene representation for controllable image generation." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.469, + 0.414, + 0.484 + ], + "angle": 0, + "content": "In summary, our contributions are the following:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.484, + 0.483, + 0.514 + ], + "angle": 0, + "content": "1) We introduce SUPERDEC, a novel method for decomposing 3D scenes using superquadric primitives." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.515, + 0.482, + 0.544 + ], + "angle": 0, + "content": "2) SUPERDEC achieves state-of-the-art object decomposition scores on ShapeNet trained jointly on multiple classes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.545, + 0.482, + 0.589 + ], + "angle": 0, + "content": "3) We demonstrate the effectiveness of 3D superquadric scene representations for robotic tasks and controllable generative content creation." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.484, + 0.483, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.604, + 0.233, + 0.619 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.623, + 0.483, + 0.851 + ], + "angle": 0, + "content": "Learning-based methods have shown that neural networks, when equipped with suitable reconstruction losses, can directly predict geometric primitive parameters to decompose point clouds into a minimal set of primitives for specific object categories. Tulsiani [51] introduced a CNN-based method for cuboid decomposition, which was later extended to more expressive primitives such as superquadrics by Paschalidou et al. [32]. CSA [57] further enhanced interpretability by employing a stronger point encoder and jointly predicting cuboid parameters and part segmentations. However, these methods remain constrained by their reliance on category-specific training. We attribute this limitation to their model design, which encodes only global shape features, sufficient for intra-category generalization but ineffective for decomposing out-of-category objects." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Optimization-based methods largely originate from the literature on superquadric fitting. EMS [23] revisited this line of work by introducing a probabilistic formulation that" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.349 + ], + "angle": 0, + "content": "enables the decomposition of arbitrary objects into multiple superquadrics. Given an input point cloud, the method first fits a superquadric to the main structure and identifies unfitted outlier clusters, which are then recursively processed in a hierarchical fashion up to a predefined depth level. However, as noted in their paper and confirmed by our experiments, this approach implicitly assumes that objects exhibit a hierarchical geometric structure, limiting its applicability to many real-world objects such as tables and chairs. Other methods, such as Marching Primitives [24], require Signed Distance Functions (SDFs) as input, which are generally not easily available in real-world scenes. More fundamentally, since these approaches optimize from scratch for each object, they cannot leverage generalizable point features or learned shape priors, both of which are critical for abstraction and robustness under partial observations, a common challenge in practical 3D capture scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.356, + 0.907, + 0.644 + ], + "angle": 0, + "content": "Scene-level decomposition. With the emergence of 3DGS [18], an increasing number of works have explored representing 3D scenes using various geometric primitives as generalized ellipsoids [15] and convexes [14]. While heuristics can control the number of Gaussians, achieving truly compact representations remains challenging. DBW [31] addresses this by fitting a small set of textured superquadrics to 3D scenes, building on the principles of 3DGS. Given a set of scene images, it performs test-time optimization with a photometric loss and renders the primitives using a differentiable rasterizer [22]. To model the environment, DBW adds a meshed ground plane and a meshed icosphere for the background. However, it is restricted to scenes with fewer than 10 primitives and requires objects to be aligned to a ground plane. Furthermore, the optimization is computationally expensive, taking around three hours even on simple DTU [17] scenes. Our method differs significantly in terms of input requirements, generality of application, and computational efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.651, + 0.909, + 0.803 + ], + "angle": 0, + "content": "Superquadrics are a parametric family of shapes introduced by Barr et al. [2] in 1981 and have since been widely adopted in both computer vision and graphics [7, 34, 44]. Their popularity stems from their ability to represent a diverse range of shapes with a highly compact parameterization. A superquadric in its canonical pose is defined by just five parameters: \\((s_x, s_y, s_z)\\) for the scales along the three principal semi-axes and \\((\\epsilon_1, \\epsilon_2)\\) for the shape-defining exponents. Given those parameters, their surface is described by the implicit equation:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.814, + 0.905, + 0.86 + ], + "angle": 0, + "content": "\\[\nf (\\mathbf {x}) = \\left(\\left(\\frac {x}{s _ {x}}\\right) ^ {\\frac {2}{\\epsilon_ {2}}} + \\left(\\frac {y}{s _ {y}}\\right) ^ {\\frac {2}{\\epsilon_ {2}}}\\right) ^ {\\frac {\\epsilon_ {2}}{\\epsilon_ {1}}} + \\left(\\frac {z}{s _ {z}}\\right) ^ {\\frac {2}{\\epsilon_ {1}}} = 1. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Extending this representation to a global coordinate system requires 6 additional parameters (3 for translation and 3" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24626" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.213 + ], + "angle": 0, + "content": "for rotation), resulting in a total of 11 parameters per superquadric. Another key property of superquadrics is the ability to compute the radial distance from any point in 3D space to the superquadric surface, i.e., the distance between a point and the superquadric's surface along the line connecting that point to the center of the superquadric. Specifically, given a point \\(\\mathbf{x} \\in \\mathbb{R}^3\\), its radial distance to the surface of a canonically oriented superquadric is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.225, + 0.483, + 0.243 + ], + "angle": 0, + "content": "\\[\nd _ {r} = | \\mathbf {x} | \\cdot | 1 - f (\\mathbf {x}) ^ {- \\epsilon_ {1} / 2} |, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.256, + 0.484, + 0.302 + ], + "angle": 0, + "content": "where \\( f(\\mathbf{x}) \\) is given in Eq. 1. We refer the reader to [53] for the derivation of Eq. 2 and to [16] for a more comprehensive overview on superquadrics." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.317, + 0.182, + 0.332 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.343, + 0.484, + 0.433 + ], + "angle": 0, + "content": "Our ultimate goal is a 3D scene decomposition using superquadric primitives. To this end, we first focus on single-object decomposition and then show how our method, combined with 3D instance segmentation [41], can be applied to full 3D scenes. We detail the single-object approach in Sec. 3.1 and its extension to full scenes in Sec. 3.2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.445, + 0.349, + 0.461 + ], + "angle": 0, + "content": "3.1. Single Object Decomposition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.468, + 0.484, + 0.559 + ], + "angle": 0, + "content": "Fig. 2 illustrates our model for single-object decomposition. It consists of two main components: a self-supervised feed-forward neural network that jointly predicts superquadric parameters and a segmentation matrix associating points to superquadrics, followed by a lightweight Levenberg-Marquardt (LM) optimization [20, 28]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.569, + 0.349, + 0.582 + ], + "angle": 0, + "content": "3.1.1. Feed-forward Neural Network" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.588, + 0.484, + 0.71 + ], + "angle": 0, + "content": "Our deep learning model draws inspiration from recent fully-supervised Transformer-based [54] segmentation models [5, 6, 41]. These models iteratively decode a sequence of queries, each representing a segmentation mask, by cross-attending to input pixels or points. In our case, the queries represent superquadrics. Next, we show how such an architecture can be adapted to unsupervisedly segment superquadrics, instead of supervisedly segment objects." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.718, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Model Details. Given an input point cloud \\(\\mathcal{P} \\in \\mathbb{R}^{N \\times 3}\\), where each of the \\(N\\) points has a 3D coordinate, we first extract rich point features \\(\\mathcal{F}_{PC} \\in \\mathbb{R}^{N \\times H}\\) using the PVCNN [25] point encoder. At the same time, we initialize \\(P\\) superquadrics features \\(\\mathcal{F}_{SQ} \\in \\mathbb{R}^{P \\times H}\\) with sinusoidal positional encodings. We feed these features in a Transformer decoder [54] which leverages self-attention, cross-attention to the point features, and feed-forward layers to refine them. Once refined, the superquadric features \\(\\mathcal{F}_{SQ}\\) and the point features \\(\\mathcal{F}_{PC}\\) are fed into two prediction heads: The segmentation head takes as input \\(\\mathcal{F}_{SQ}\\) and \\(\\mathcal{F}_{PC}\\) and predicts a soft assignment matrix \\(M \\in \\mathbb{R}^{N \\times P}\\) associating points to" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.843, + 0.107 + ], + "angle": 0, + "content": "superquadrics and whose elements are defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.114, + 0.905, + 0.132 + ], + "angle": 0, + "content": "\\[\nm _ {i j} = \\sigma (\\phi (\\mathcal {F} _ {P C}) \\cdot \\mathcal {F} _ {S Q}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.137, + 0.907, + 0.259 + ], + "angle": 0, + "content": "where \\(\\phi (\\mathcal{F}_{PC})\\in \\mathbb{R}^{N\\times H}\\) is a learned projection of the point features to match the dimensionality of the superquadric features, and \\(\\sigma\\) is the softmax function. The second head, the superquadric head, takes the superquadric features \\(\\mathcal{F}_{SQ}\\) as input and predicts 12 parameters for each superquadric: 11 encoding its 5-DoF shape and 6-DoF pose, and one modeling its existence probability \\(\\alpha\\), enabling a variable number of superquadrics per object." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.264, + 0.906, + 0.308 + ], + "angle": 0, + "content": "Losses. We train our model in a self-supervised manner, without requiring any ground truth annotation. Specifically, the total loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.317, + 0.905, + 0.334 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e c}} + \\lambda_ {\\text {p a r}} \\mathcal {L} _ {\\text {p a r}} + \\lambda_ {\\text {e x i s t}} \\mathcal {L} _ {\\text {e x i s t}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.341, + 0.906, + 0.416 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{rec}\\) is the reconstruction loss aligning the predicted superquadrics to the input point cloud \\(\\mathcal{P}\\), \\(\\mathcal{L}_{par}\\) is the parsimony loss encouraging a small number of primitives, \\(\\mathcal{L}_{exist}\\) is the existence loss, and \\(\\lambda_{par}, \\lambda_{exist}\\) are weighting coefficients. The reconstruction loss \\(\\mathcal{L}_{rec}\\) consists of three terms:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.424, + 0.905, + 0.441 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e c} = \\mathcal {L} _ {\\mathcal {P} \\rightarrow S Q} + \\mathcal {L} _ {S Q \\rightarrow \\mathcal {P}} + \\mathcal {L} _ {N}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.907, + 0.6 + ], + "angle": 0, + "content": "The first two terms correspond to the bi-directional Chamfer distance between the input point cloud and the superquadric surfaces, while the third term serves as a regularizer incorporating normal information to improve convergence during training. To compute the Chamfer distance, we approximate each superquadric surface by uniformly sampling \\(S\\) points, following the method of Pilu et al. [35]. Denoting by \\(d(\\mathbf{x}_i,\\mathbf{x}_{js}^{\\prime})\\) the euclidean distance between the \\(i\\)-th point in the input point cloud and the \\(s\\)-th point sampled on the surface of the \\(j\\)-th superquadric, we define \\(\\mathcal{L}_{\\mathcal{P}\\to SQ}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.606, + 0.905, + 0.649 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {P} \\rightarrow S Q} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {P} m _ {i j} \\min _ {s \\in [ S ]} d \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {j s} ^ {\\prime}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.655, + 0.625, + 0.671 + ], + "angle": 0, + "content": "and \\(\\mathcal{L}_{SQ\\to \\mathcal{P}}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.677, + 0.905, + 0.72 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S Q \\rightarrow \\mathcal {P}} = \\frac {1}{S \\sum_ {j = 1} ^ {P} \\alpha_ {j}} \\sum_ {j = 1} ^ {P} \\alpha_ {j} \\sum_ {s = 1} ^ {S} \\min _ {i \\in [ N ]} d \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {j s} ^ {\\prime}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.726, + 0.906, + 0.848 + ], + "angle": 0, + "content": "The last term of Eq. 5, i.e., \\(\\mathcal{L}_N\\) is defined as the reconstruction loss from Yang et al. [57], and is used to incorporate normal information during training which leads to accelerated convergence. Additionally, since we seek not only accuracy but also compactness, we introduce a parsimony loss to encourage the use of fewer primitives. To do that, we optimize the 0.5-norm of \\(m_j \\coloneqq \\sum_{i=1}^{N} \\frac{m_{ij}}{N}\\) and define the parsimony loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.853, + 0.905, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p a r} = \\left(\\frac {1}{P} \\sum_ {j = 1} ^ {P} \\frac {\\sqrt {m _ {j}}}{P}\\right) ^ {2}. \\tag {8}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24627" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.087, + 0.905, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.282, + 0.908, + 0.34 + ], + "angle": 0, + "content": "Figure 2. Illustration of the SUPERDEC Model. Given a point cloud of an object with \\(N\\) points, a Transformer-based neural network predicts parameters for \\(P\\) superquadrics, as well as a soft segmentation matrix that assigns points to superquadrics. The predicted parameters include the 11 superquadric parameters and an objectness score. These predictions provide an effective initialization for the subsequent Levenberg-Marquardt (LM) optimization, which refines the superquadrics." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.361, + 0.484, + 0.455 + ], + "angle": 0, + "content": "Lastly, we employ an existence loss \\(\\mathcal{L}_{\\text {exist }}\\) which uses the predicted segmentation as a teacher for the linear head in charge of predicting the existence probability. More specifically, given a threshold \\(\\epsilon_{\\text {exist }}\\), we define the ground-truth existence of the \\(j\\)th superquadric as \\(\\hat{\\alpha}_j \\coloneqq m_j > \\epsilon_{\\text {exist }}\\) and define \\(\\mathcal{L}_{\\text {exist }}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.465, + 0.483, + 0.508 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {e x i s t}} = \\sum_ {j = 1} ^ {P} \\frac {B C E \\left(\\alpha_ {j} , \\hat {\\alpha} _ {j}\\right)}{P}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.484, + 0.552 + ], + "angle": 0, + "content": "where BCE is the binary cross entropy and \\(\\alpha_{j}\\) is the predicted existence probability for the \\(j\\)th superquadric." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.561, + 0.232, + 0.577 + ], + "angle": 0, + "content": "3.1.2. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.484, + 0.703 + ], + "angle": 0, + "content": "Our optimization module takes as input the predicted soft segmentation matrix \\(M\\) as well as the superquadric parameters \\(\\Theta\\), and further refines the superquadric parameters using the Levenberg-Marquardt (LM) [20, 28] algorithm. Specifically, given a point cloud of \\(N\\) points, it iteratively refines the parameters \\(\\Theta_{j}\\) of the \\(j\\)th superquadric, by computing two sets of residuals: The first set of residuals \\(r_{ij}\\) with \\(i \\in [1, N]\\) and \\(j \\in [1, P]\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.224, + 0.715, + 0.483, + 0.734 + ], + "angle": 0, + "content": "\\[\nr _ {i j} = m _ {i j} \\tilde {d} _ {j} \\left(\\mathbf {x} _ {i}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.747, + 0.484, + 0.856 + ], + "angle": 0, + "content": "where \\(\\tilde{d}_j(\\mathbf{x}_i)\\) denotes the radial distance of point \\(\\mathbf{x}_i\\) from the \\(j\\)th superquadric, computed according to Eq. 2. The second set of residuals is used for normalization and is obtained by sampling a set of \\(K\\) points \\(\\mathbf{p_1},\\ldots ,\\mathbf{p_K}\\) on the surface of the given superquadric and then computing the distance of each of them from the point cloud. Specifically, for \\(i\\in [N + 1,N + K]\\) and \\(j\\in [1,P]\\) we compute \\(r_{ij}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.868, + 0.484, + 0.89 + ], + "angle": 0, + "content": "\\[\nr _ {i j} = \\min _ {k} \\left\\| \\mathbf {p} _ {\\mathbf {i} - \\mathbf {N}} - \\Pi_ {j} \\left(\\mathbf {x} _ {\\mathbf {k}}\\right) \\right\\| _ {2}, \\quad \\text {w i t h} k \\in [ 1, N ]. \\tag {11}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.36, + 0.803, + 0.377 + ], + "angle": 0, + "content": "3.2. Decomposition of Full 3D Scenes" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.384, + 0.907, + 0.522 + ], + "angle": 0, + "content": "After training on single objects, extending SUPERDEC to full 3D scenes is straightforward. Given a scene-level point cloud, we extract 3D object instance masks using Mask3D [41]. Each object is centered and uniformly rescaled to a sphere of radius 0.5. We then predict the superquadric primitives for each object individually using our model. We found our model trained on ShapeNet [4] to generalize well on real-world 3D scenes from ScanNet++ [58] and Replica [45] without additional fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.537, + 0.646, + 0.554 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.563, + 0.907, + 0.685 + ], + "angle": 0, + "content": "We first compare our SUPERDEC with previous state-of-the-art methods on individual objects and full 3D scenes (Sec. 4.1). We then demonstrate the usefulness of our representation on down-stream applications for robotics and controllable image generation (Sec. 4.2). Finally, in Sec. 4.3, we present additional analyses on part segmentation and the implicit learning of shape categories, followed by a study of the compactness-accuracy trade-off and runtime." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.697, + 0.871, + 0.713 + ], + "angle": 0, + "content": "4.1. Comparing with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.825 + ], + "angle": 0, + "content": "Datasets. We compare on three different datasets: ShapeNet [4]: We use the 13 classes of the ShapeNet subset and train-val-test splits as defined in Choy et al. [8]. All objects are pre-aligned in a canonical orientation. For each object we sample 4096 points using Farthest Point Sampling (FPS) [36]. ShapeNet is a widely used dataset and is well-suited for comparison with existing baselines." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.902 + ], + "angle": 0, + "content": "ScanNet++ [58]: We further evaluate our model on real-world object scans from the ScanNet++ validation set. Each object is extracted using ground truth mask annotations, and 4096 points per object are sampled with FPS. In contrast to ShapeNet, these object point clouds are noisier, partially ob" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24628" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.905, + 0.189 + ], + "angle": 0, + "content": "
ModelPrimitive TypeSegmentationIn-categoryOut-of-category
L1 ↓L2 ↓# Prim.↓L1 ↓L2 ↓# Prim.↓
EMS (Liu et al.) [23]SuperquadricsX5.7711.3455.685.4101.2115.68
CSA (Yang et al.) [57]Cuboids✓5.1570.5279.214.8970.42711.75
SQ (Paschalidou et al.) [32]SuperquadricsX3.6680.279104.1930.3549
SUPERDEC (Ours)Superquadrics✓1.6980.0475.81.8470.0615.26
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.191, + 0.903, + 0.207 + ], + "angle": 0, + "content": "Table 1. Quantitative Results on ShapeNet [4]. We show scores for in-category and out-of-category experiments and are scaled by \\(10^{2}\\)." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.212, + 0.898, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.584, + 0.908, + 0.613 + ], + "angle": 0, + "content": "Figure 3. Qualitative Results on ShapeNet [4]. We show results on test samples for in-category (four first columns) classes and out-of-category classes (two last columns). The latter were not seen during training and illustrate how well models generalize to novel classes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.636, + 0.483, + 0.68 + ], + "angle": 0, + "content": "served, and subject to random orientation and translation, providing a more realistic and challenging evaluation setting for our method." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.681, + 0.483, + 0.786 + ], + "angle": 0, + "content": "Replica [45]: Finally, we present both quantitative and qualitative results on full 3D scenes from Replica. Object instances are extracted using either ground truth instance annotations or the pre-trained 3D instance segmentation model Mask3D [41], enabling us to evaluate our approach in a fully realistic setting, including scenarios where no ground truth annotations are available." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Methods in comparison. We compare to learning- and optimization-based prior works using both cuboids and superquadrics as geometric primitives. \\( SQ \\) [32] is a learning-based approach for object-level decomposition using superquadrics; it takes a voxel grid as input and predicts superquadric primitives via a CNN. CSA [57] is another learning-based method but uses cuboids as geometric prim" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.636, + 0.905, + 0.697 + ], + "angle": 0, + "content": "ities. It takes a point cloud as input and predicts cuboid parameters from a global latent code. Lastly, EMS [23] is an optimization-based approach that decomposes objects by hierarchically fitting superquadrics to parts of a point cloud." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Training Details. Our goal is to develop a general-purpose, class-agnostic model capable of representing arbitrary objects as superquadrics. Existing methods typically train separate models for each object class, assuming that all classes are known in advance and that sufficient training data is available for each. These assumptions, however, often fail in real-world scenarios. To address this, we jointly train a single model on all 13 ShapeNet classes using the publicly available code of prior methods, moving towards a more realistic class-agnostic solution. In our model we set the following hyper-parameters \\( P = 16 \\), \\( S = 4096 \\), \\( K = 25 \\), \\( H = 128 \\), \\( D = 3 \\), \\( \\epsilon_{\\text{exist}} = 24 \\), \\( \\lambda_{\\text{exist}} = 0.01 \\), \\( \\lambda_{\\text{par}} = 0.06 \\). We refer to the supplementary for additional details." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24629" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.48, + 0.169 + ], + "angle": 0, + "content": "
ScanNet++ [58]Replica [45]
MethodL1 ↓L2 ↓#Prim.↓L1 ↓L2 ↓#Prim.↓
SQ [32]10.454.2610.011.845.9110
EMS [23]5.512.114.255.402.123.61
CSA [57]2.910.4111.643.680.709.63
SUPERDEC (Ours)1.700.115.181.790.196.58
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.173, + 0.483, + 0.203 + ], + "angle": 0, + "content": "Table 2. Quantitative Results on Scenes. Evaluation on objects from 3D scene datasets [45, 58]. Scores are scaled by \\(10^{2}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.231, + 0.484, + 0.277 + ], + "angle": 0, + "content": "Metrics. We assess reconstruction accuracy using L1 and L2 Chamfer distances and compactness by the average number of geometric primitives." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.295, + 0.283, + 0.31 + ], + "angle": 0, + "content": "4.1.1. Results on ShapeNet" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.317, + 0.484, + 0.635 + ], + "angle": 0, + "content": "We show scores in Tab. 1 and qualitative results in Fig. 3. To evaluate both accuracy and generalization, we conduct two experiments: in-category and out-of-category. In the in-category setting, all learning-based methods are jointly trained on the 13 classes of the ShapeNet training set and evaluated on the corresponding test set. In the out-of-category setting, models are trained on half of the categories (airplane, bench, chair, lamp, rifle, table) and tested on the remaining ones (car, sofa, loudspeaker, cabinet, display, telephone, watercraft). Our SUPERDEC model significantly outperforms both learned and non-learned baselines. Compared to learned baselines, we reduce the L2 loss by a factor of six while using nearly half the number of primitives, supporting our hypothesis that leveraging local point features as opposed to a single global descriptor improves 3D decomposition in both accuracy and compactness. Compared to the non-learned baseline, we predict a similar number of primitives but achieve an L2 loss approximately 20 times smaller, validating the benefit of learning shape priors to avoid local minima that often hinder purely optimization-based approaches." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.653, + 0.285, + 0.667 + ], + "angle": 0, + "content": "4.1.2. Results on 3D Scenes" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this section, models are evaluated on real-world, out-of-category objects, which appear in arbitrary orientations and often exhibit incomplete point clouds due to reconstruction artifacts and occlusions. Tab. 2 shows the quantitative results. Despite never being trained on real-world objects, our method outperforms both the optimization- and the learning-based baselines by a large margin. Lastly, we qualitatively evaluate our pipeline on full 3D scenes from Replica, where our object-level model is applied on top of class-agnostic instance segmentation predictions from Mask3D [41]. As shown in Fig. 1, our method effectively reconstructs object shapes, even under noisy segmentation masks and geometries that differ substantially from those seen during training. We refer to the supplementary for additional qualitative results." + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.089, + 0.895, + 0.189 + ], + "angle": 0, + "content": "
MethodTime (ms)Suc. (%)Mem. (MB)
Occupancy0.056100.000.873
PointCloud0.06389.5719.286
Voxels0.03098.780.101
Cuboids [37]0.12061.230.024
SUPERDEC0.15091.710.042
" + }, + { + "type": "table_caption", + "bbox": [ + 0.518, + 0.192, + 0.901, + 0.206 + ], + "angle": 0, + "content": "Table 3. Path Planning Scores. Mean on 15 ScanNet++ scenes." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.212, + 0.905, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.35, + 0.906, + 0.379 + ], + "angle": 0, + "content": "Figure 4. Grasping Result. Visualization of computed grasp poses for a milk bottle, some flowers, a side table, and a plant." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.387, + 0.757, + 0.403 + ], + "angle": 0, + "content": "4.2. Down-stream Applications" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.907, + 0.47 + ], + "angle": 0, + "content": "Next, we show the versatility of the SUPERDEC representation for downstream applications, including robotics tasks such as path planning and object grasping (Sec. 4.2.1), and controllable image generation (Sec. 4.2.2)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.48, + 0.622, + 0.493 + ], + "angle": 0, + "content": "4.2.1. Robotics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.499, + 0.907, + 0.726 + ], + "angle": 0, + "content": "Path planning seeks to compute a collision-free shortest path between a given start and end point in 3D space, enabling efficient robot navigation. Although essential for traversing large environments, it typically demands storing large-scale 3D representations. Here, we assess whether our compact representation can perform this task effectively while reducing memory requirements. We conduct experiments on 15 ScanNet++ [58] scenes, comparing SUPERDEC to common 3D representations, including dense occupancy grids, point clouds, voxel grids, and cuboids [37]. As shown in Tab. 6, SUPERDEC not only reduces memory consumption compared to traditional representations but also achieves a higher success rate than dense point clouds. Further details about experiment setup, metrics and analysis are provided in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Object Grasping enables robots to grasp real-world objects by computing suitable grasping poses. Existing methods fall into two categories, each with complementary limitations. Geometry-based approaches [3, 12, 30] require precise 3D object models, which are often unavailable in real-world scenarios. Learning-based approaches [27, 47, 56] operate directly on raw sensor data but tend to be biased towards training data, which typically consists of tabletop scenes with small, convex, or low-genus objects [11, 46]. To overcome these limitations, SuperQ-GRASP [50] explored decomposing objects into explicit primitives. How" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24630" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.142, + 0.085, + 0.241, + 0.097 + ], + "angle": 0, + "content": "Input Point Cloud" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.098, + 0.288, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.347, + 0.085, + 0.426, + 0.097 + ], + "angle": 0, + "content": "Superquadrics" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.098, + 0.482, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.184, + 0.482, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.312, + 0.483, + 0.397 + ], + "angle": 0, + "content": "Figure 5. Real-world robot experiment. The top row shows the input scan (left) and the representation from SUPERDEC with the computed path and grasping pose (right). The bottom row illustrates the robot following the planned path. We denote the starting point of the path with a green sphere, and the target location with a red sphere. The target object (a milk bottle) is circled in red." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.412, + 0.484, + 0.624 + ], + "angle": 0, + "content": "ever, its reliance on Marching Primitives [24] to obtain superquadrics from the object's Signed Distance Function (SDF) makes it unsuitable for most real-world cases where only point clouds are available. In contrast, our approach directly processes point clouds of entire scenes and, when combined with the class-agnostic segmentations from Mask3D [41], extracts superquadrics for all objects. Given the superquadric parameters, we employ a superquadric-based geometric method [55] to compute grasping poses for selected objects. Fig. 4 shows predicted grasping poses on objects from a real-world 3D scan of a room. In practice, our method eliminates the need for data-driven grasping models while remaining adaptable to diverse object shapes and producing high-quality grasping poses." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Real-world Experiment. Finally, we demonstrate the real-world applicability of our superquadric-based representation by deploying it on a legged robot (Boston Dynamics Spot) equipped with an arm, supporting both motion planning and object grasping in an indoor environment. We scan a scene using a 3D scanning application on an iPad, extract a dense point cloud, and run SUPERDEC on it. Given the robot's starting position and a specified target object (a milk bottle), we compute both the path and the grasping pose as described earlier (see Fig. 4), enabling the robot to approach and successfully grasp the object. Fig. 5 shows the computed representation, the planned trajectory, the grasping pose, and a frame from the real-world demonstration. This experiment suggests that integrating SUPERDEC with open-vocabulary segmentation methods such as OpenMask3D [48] could allow robots to navigate to and grasp arbitrary objects specified via natural-language prompts." + }, + { + "type": "image_caption", + "bbox": [ + 0.54, + 0.086, + 0.593, + 0.098 + ], + "angle": 0, + "content": "Original" + }, + { + "type": "image_caption", + "bbox": [ + 0.64, + 0.086, + 0.686, + 0.098 + ], + "angle": 0, + "content": "Editing" + }, + { + "type": "image_caption", + "bbox": [ + 0.732, + 0.086, + 0.786, + 0.097 + ], + "angle": 0, + "content": "Addition" + }, + { + "type": "image_caption", + "bbox": [ + 0.829, + 0.086, + 0.882, + 0.097 + ], + "angle": 0, + "content": "Deletion" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.1, + 0.905, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.184, + 0.905, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.272, + 0.905, + 0.315 + ], + "angle": 0, + "content": "Figure 6. Spatial control using SUPERDEC. Top row shows superquadrics generated by SUPERDEC, bottom row shows generated images using the prompt 'A corner of a room with a plant'." + }, + { + "type": "image_caption", + "bbox": [ + 0.57, + 0.338, + 0.658, + 0.351 + ], + "angle": 0, + "content": "Superquadrics" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.354, + 0.71, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.453, + 0.674, + 0.465 + ], + "angle": 0, + "content": "\"Pink living room\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.762, + 0.338, + 0.848, + 0.351 + ], + "angle": 0, + "content": "Depth Prompt" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.354, + 0.903, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.738, + 0.453, + 0.872, + 0.465 + ], + "angle": 0, + "content": "\"Modern living room\"" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.467, + 0.71, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.467, + 0.903, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.568, + 0.906, + 0.624 + ], + "angle": 0, + "content": "Figure 7. Semantic control using SUPERDEC. Top row: superquadrics created by our SUPERDEC, and depth map to prompt the generation of text-to-image diffusion model. Bottom row: generations with two different textual prompts." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.648, + 0.816, + 0.664 + ], + "angle": 0, + "content": "4.2.2. Controllable Generation and Editing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We investigate how the SUPERDEC representation can be used to introduce joint spatial and semantic control in text-to-image diffusion models [39]. Specifically, we generate images by conditioning ControlNet [60] on depth maps rendered from the superquadrics extracted from Replica [45] scenes. Qualitative results are shown in Fig. 6 and Fig. 7. Fig. 6 demonstrates spatial control: by moving, duplicating, or removing superquadrics corresponding to a plant, we coherently influence the generated images. Fig. 7 highlights semantic control: we can vary the room's style while preserving its semantic and geometric structure, and observe that object semantics naturally emerge from the spatial arrangement of superquadrics without explicit conditioning, e.g., pillows appear on couches, and a plant is placed on a central table, reflecting plausible real-world arrangements." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24631" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.183, + 0.105, + 0.255, + 0.117 + ], + "angle": 0, + "content": "In-category" + }, + { + "type": "header", + "bbox": [ + 0.36, + 0.105, + 0.458, + 0.117 + ], + "angle": 0, + "content": "Out-of-category" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.123, + 0.466, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.18, + 0.483, + 0.236 + ], + "angle": 0, + "content": "Figure 8. Qualitative Results on ShapeNet [4] segmentation. We show the resulting segmentation matrices on test samples for in-category (four first columns) classes and out-of-category classes (two last columns). The latter were not seen during training." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.244, + 0.225, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.316, + 0.192, + 0.328 + ], + "angle": 0, + "content": "Input 3D Scan" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.244, + 0.482, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.242, + 0.316, + 0.313, + 0.328 + ], + "angle": 0, + "content": "\\(1^{\\mathrm{st}}\\) Hierarchy" + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.315, + 0.447, + 0.328 + ], + "angle": 0, + "content": "\\(2^{\\mathrm{nd}}\\) Hierarchy" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.336, + 0.484, + 0.392 + ], + "angle": 0, + "content": "Figure 9. Hierarchical decomposition with SUPERDEC. From left to right we visualize: the input point cloud, the superquadrics predicted in the first hierarchy, the superquadrics predicted in the second hierarchy." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.415, + 0.295, + 0.43 + ], + "angle": 0, + "content": "4.3. Analysis Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.437, + 0.484, + 0.603 + ], + "angle": 0, + "content": "Unsupervised part segmentation. Besides superquadric parameters, our method also predicts a segmentation matrix which segments the initial point cloud into parts that are fitted to the predicted superquadrics. In Fig. 8, we visualize the predicted segmentation masks for the same examples shown in Fig. 3. We observe that segmentation masks, especially in the in-category experiments, appear very sharp. This suggests that our method, especially if trained at a larger scale, can be leveraged for different applications as geometry-based part segmentation or as pretraining for supervised semantic part segmentation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.622, + 0.484, + 0.819 + ], + "angle": 0, + "content": "Can we do hierarchical decomposition? We run SUPERDEC in a hierarchical fashion on a whole simple scene, without using any instance segmentation method to segment the scene into different objects. We observed that by applying SUPERDEC on the whole point cloud it can coarsely segment the scene into components such as grass, table, and chairs. By applying the model again on the points segmented during the first prediction, we obtained a more accurate decomposition of the individual object instances. Given that SUPERDEC was trained only on single object instances from ShapeNet [4] we found this results very interesting. We see future potential in the use of our model to obtain unsupervised hierarchical decompositions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "What does our network learn? Since our network performs unsupervised part segmentation, we analyze the features learned by the Transformer decoder across object classes. Inspired by BERT [10]'s [CLS] token, we append a learnable embedding to the sequence of embedded" + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.077, + 0.865, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.27, + 0.905, + 0.299 + ], + "angle": 0, + "content": "Figure 10. t-SNE Visualization of Primitive Embeddings across different ShapeNet classes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.325, + 0.907, + 0.462 + ], + "angle": 0, + "content": "superquadrics; although never explicitly decoded, this embedding is refined through self- and cross-attention. After training, we extract and visualize these embeddings using t-SNE [52] for ShapeNet [4] categories (Fig. 10). We observe that categories with consistent shapes, such as chairs, airplanes, and cars, form clear clusters, while categories with high intra-class variability, such as watercraft, are more dispersed. This indicates that our model organizes objects by geometric structure without requiring class annotations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.906, + 0.59 + ], + "angle": 0, + "content": "How fast is our method? Our model is highly parallelizable, allowing multiple objects to be batched and processed simultaneously in a single forward pass. On an RTX 4090 (24 GB), we can process up to 256 objects in parallel. On average, the forward pass takes \\(0.13\\mathrm{s}\\) for a complete Replica [45] scene, 3D instance segmentation with Mask3D [41] requires \\(0.3\\mathrm{s}\\), and each LM optimization step takes less than 1 s, see supplementary for more details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.601, + 0.634, + 0.615 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.62, + 0.907, + 0.786 + ], + "angle": 0, + "content": "We proposed SUPERDEC, a method for deriving compact yet expressive 3D scene representations based on simple geometric primitives – specifically, superquadrics. Our model outperforms prior primitive-based methods and generalizes well to out-of-category classes. We further demonstrated the potential of the resulting 3D scene representation for various applications in robotics, and as a geometric prompt for diffusion-based image generation. While this is only a first step towards more compact, geometry-aware 3D scene representations, we anticipate broader applications and expect to see further research in this direction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments. Elisabetta Fedele is supported by the ETH AI Center doctoral fellowship, by the Swiss National Science Foundation (SNSF) Advanced Grant 216260 (Beyond Frozen Worlds: Capturing Functional 3D Digital Twins from the Real World), and an SNSF Mobility Grant. Francis Engelmann is supported by an SNSF PostDoc.mobility grant. We also gratefully acknowledge a GPU grant from NVIDIA." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24632" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Stephan Alaniz, Massimiliano Mancini, and Zeynep Akata. Iterative superquadric decomposition of 3d objects from multiple viewsd. In International Conference on Computer Vision (ICCV), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.171, + 0.484, + 0.212 + ], + "angle": 0, + "content": "[2] Alan H. Barr. Superquadrics and angle-preserving transformations. IEEE Computer Graphics and Applications, 1981. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.213, + 0.484, + 0.269 + ], + "angle": 0, + "content": "[3] Junhao Cai, Jingcheng Su, Zida Zhou, Hui Cheng, Qifeng Chen, and Michael Yu Wang. Volumetric-based contact point detection for 7-dof grasping. In Conference on Robot Learning (CoRL), 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.27, + 0.483, + 0.34 + ], + "angle": 0, + "content": "[4] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical report, 2015. 2, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.34, + 0.483, + 0.408 + ], + "angle": 0, + "content": "[5] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention Mask Transformer for Universal Image Segmentation. International Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.41, + 0.483, + 0.466 + ], + "angle": 0, + "content": "[6] Bowen Cheng, Alexander G. Schwing, and Alexander Kirillov. Per-Pixel Classification is Not All You Need for Semantic Segmentation. In International Conference on Neural Information Processing Systems (NeurIPS), 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.467, + 0.483, + 0.522 + ], + "angle": 0, + "content": "[7] Laurent Chevalier, Fabrice Jaillet, and Atilla Baskurt. Segmentation and superquadric modeling of 3d objects. In International Conference in Central Europe on Computer Graphics and Visualization, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.523, + 0.483, + 0.579 + ], + "angle": 0, + "content": "[8] Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3D-R2N2: A Unified Approach for Single and Multi-view 3D Object Reconstruction. In European Conference on Computer Vision (ECCV), 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.58, + 0.483, + 0.661 + ], + "angle": 0, + "content": "[9] Alexandros Delitzas, Ayca Takmaz, Federico Tombari, Robert Sumner, Marc Pollefeys, and Francis Engelmann. SceneFun3D: Fine-Grained Functionality and Affordance Understanding in 3D Scenes. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies., 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[11] Hao-Shu Fang, Chenxi Wang, Hongjie Fang, Minghao Gou, Jirong Liu, Hengxu Yan, Wenhai Liu, Yichen Xie, and Cewu Lu. Anygrasp: Robust and efficient grasp perception in spatial and temporal domains. IEEE Transactions on Robotics, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.859 + ], + "angle": 0, + "content": "[12] B. Faverjon and J. Ponce. On Computing Two-finger Force-closure Grasps of Curved 2D Objects. In International Conference on Robotics and Automation (ICRA), 1991. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Qiao Gu, Ali Kuwajerwala, Sacha Morin, Krishna Murthy Jatavallabhula, Bipasha Sen, Aditya Agarwal, Corban Rivera, William Paul, Kirsty Ellis, Rama Chellappa, et al." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Conceptgraphs: Open-vocabulary 3d scene graphs for perception and planning. In International Conference on Robotics and Automation (ICRA), 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.135, + 0.905, + 0.216 + ], + "angle": 0, + "content": "[14] Abdullah Hamdi, Luke Melas-Kyriazi, Jinjie Mai, Guocheng Qian, Ruoshi Liu, Carl Vondrick, Bernard Ghanem, and Andrea Vedaldi. Ges: Generalized exponential splatting for efficient radiance field rendering. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.218, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[15] Jan Held, Renaud Vandeghen, Abdullah Hamdi, Adrien Deli'ege, Anthony Cioppa, Silvio Giancola, Andrea Vedaldi, Bernard Ghanem, and Marc Van Droogenbroeck. 3D convex splatting: Radiance field rendering with 3D smooth convexes. ArXiv, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.288, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[16] Ales Jaklic, Ales Leonardis, and Franc Solina. Segmentation and recovery of superquadrics. Springer Science & Business Media, 2000. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.329, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[17] Rasmus Ramsbøl Jensen, A. Dahl, George Vogiatzis, Engil Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. International Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.905, + 0.441 + ], + "angle": 0, + "content": "[18] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions On Graphics (TOG), 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[19] Sebastian Koch, Narunas Vaskevicius, Mirco Colosi, Pedro Hermosilla, and Timo Ropinski. Open3DSG: Open-vocabulary 3D Scene Graphs from Point Clouds with Queryable Objects and Open-set Relationships. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.565 + ], + "angle": 0, + "content": "[20] Kenneth Levenberg. A method for the solution of certain non-linear problems in least squares. Quarterly of Applied Mathematics, 1944. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.608 + ], + "angle": 0, + "content": "[21] Lei Li and Angela Dai. Genzi: Zero-shot 3d human-scene interaction generation. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.905, + 0.662 + ], + "angle": 0, + "content": "[22] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. International Conference on Computer Vision (ICCV), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.664, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[23] Weixiao Liu, Yuwei Wu, Sipu Ruan, and Gregory S Chirikjian. Robust and Accurate Superquadric Recovery: A Probabilistic Approach. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[24] Weixiao Liu, Yuwei Wu, Sipu Ruan, and Gregory S Chirikjian. Marching-primitives: Shape abstraction from signed distance function. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.859 + ], + "angle": 0, + "content": "[25] Zhijian Liu, Haotian Tang, Yujun Lin, and Song Han. PointVoxel CNN for Efficient 3D Deep Learning. In International Conference on Neural Information Processing Systems (NeurIPS), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. International Conference on Learning Representations (ICLR), 2019. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24633" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[27] Jeffrey Mahler, Matthew Matl, Vishal Satish, Michael Danielczuk, Bill DeRose, Stephen McKinley, and Ken Goldberg. Learning ambidextrous robot grasping policies. Science Robotics, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.151, + 0.483, + 0.192 + ], + "angle": 0, + "content": "[28] Donald W. Marquardt. An algorithm for least-squares estimation of nonlinear parameters. Journal of The Society for Industrial and Applied Mathematics, 1963. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.195, + 0.482, + 0.263 + ], + "angle": 0, + "content": "[29] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.267, + 0.482, + 0.32 + ], + "angle": 0, + "content": "[30] A.T. Miller, S. Knoop, H.I. Christensen, and P.K. Allen. Automatic Grasp Planning Using Shape Primitives. In International Conference on Robotics and Automation (ICRA), 2003. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.325, + 0.482, + 0.393 + ], + "angle": 0, + "content": "[31] Tom Monnier, Jake Austin, Angjoo Kanazawa, Alexei A. Efros, and Mathieu Aubry. Differentiable blocks world: Qualitative 3d decomposition by rendering primitives. International Conference on Neural Information Processing Systems (NeurIPS), 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.396, + 0.482, + 0.451 + ], + "angle": 0, + "content": "[32] Despoina Paschalidou, Ali Osman Ulusoy, and Andreas Geiger. Superquadrics revisited: Learning 3d shape parsing beyond cuboids. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.454, + 0.482, + 0.522 + ], + "angle": 0, + "content": "[33] Songyou Peng, Kyle Genova, Chiyu Max Jiang, Andrea Tagliasacchi, Marc Pollefeys, and Thomas Funkhouser. OpenScene: 3D Scene Understanding with Open Vocabularies. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.482, + 0.567 + ], + "angle": 0, + "content": "[34] Alex P Pentland. Parts: structured descriptions of shape. In Proceedings of the Fifth AAAI National Conference on Artificial Intelligence, 1986. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.57, + 0.482, + 0.611 + ], + "angle": 0, + "content": "[35] Maurizio Pilu and Robert B Fisher. Equal-distance sampling of superellipse models. In British Machine Vision Conference (BMVC), 1995. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.614, + 0.482, + 0.668 + ], + "angle": 0, + "content": "[36] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space. International Conference on Neural Information Processing Systems (NeurIPS), 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.672, + 0.482, + 0.727 + ], + "angle": 0, + "content": "[37] Michael Ramamonjisoa, Sinisa Stekovic, and Vincent Lepetit. MonteBoxFinder: Detecting and Filtering Primitives to Fit a Noisy Point Cloud. In European Conference on Computer Vision (ECCV), 2022. 6, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.73, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[38] Aaron Ray, Christopher Bradley, Luca Carlone, and Nicholas Roy. Task and motion planning in hierarchical 3d scene graphs. arXiv preprint arXiv:2403.08094, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.774, + 0.482, + 0.841 + ], + "angle": 0, + "content": "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. International Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[40] Zuria Bauer Marc Pollefeys Federico Tombari Leonidas Guibas Gao Huang Francis Engelmann Rui Huang, Guangyao Zhai. VIPs: Video Perception Model for 3D Scene Generation. arXiv preprint, 2025. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.16 + ], + "angle": 0, + "content": "[41] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask transformer for 3d semantic instance segmentation. International Conference on Robotics and Automation (ICRA), 2022. 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[42] Jonas Schult, Sam Tsai, Lukas Hollein, Bichen Wu, Jialiang Wang, Chih-Yao Ma, Kunpeng Li, Xiaofang Wang, Felix Wimbauer, Zijian He, et al. Controlroom3d: Room generation using semantic proxy rooms. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[43] Jaidev Shriram, Alex Trevithick, Lingjie Liu, and Ravi Ramamoorthi. Realmdreamer: Text-driven 3d scene generation with inpainting and depth diffusion. arXiv preprint arXiv:2404.07199, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.306, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[44] Franc Solina and Ruzena Bajcsy. Recovery of parametric models from range images: The case for superquadrics with global deformations. Transactions on Pattern Analysis and Machine Intelligence (PAMI), 1990. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.5 + ], + "angle": 0, + "content": "[45] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Yuheng Ren, Shobhit Verma, Anton Clarkson, Ming Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke Malte Strasdat, Renzo De Nardi, Michael Goesele, S. Lovegrove, and Richard A. Newcombe. The replica dataset: A digital replica of indoor spaces. ArXiv, 2019. 2, 4, 5, 6, 7, 8, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.503, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[46] Matan Sudry, Tom Jurgenson, Aviv Tamar, and Erez Karpas. Hierarchical planning for rope manipulation using knot theory and a learned inverse model. In Conference on Robot Learning, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.56, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[47] Martin Sundermeyer, Arsalan Mousavian, Rudolph Triebel, and Dieter Fox. Contact-grasnet: Efficient 6-dof grasp generation in cluttered scenes. In International Conference on Robotics and Automation (ICRA), 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[48] Ayca Takmaz, Elisabetta Fedele, Robert W. Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. OpenMask3D: Open-Vocabulary 3D Instance Segmentation. In International Conference on Neural Information Processing Systems (NeurIPS), 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.688, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[49] Ayca Takmaz, Alexandros Delitzas, Robert W. Sumner, Francis Engelmann, Johanna Wald, and Federico Tombari. Search3D: Hierarchical Open-Vocabulary 3D Segmentation. IEEE Robotics and Automation Letters (RA-L), 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[50] Xun Tu and Karthik Desingh. Superq-grasp: Superquadrics-based grasp pose estimation on larger objects for mobile-manipulation. arXiv, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.856 + ], + "angle": 0, + "content": "[51] Shubham Tulsiani, Hao Su, Leonidas J. Guibas, Alexei A. Efros, and Jitendra Malik. Learning shape abstractions by assembling volumetric primitives. International Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[52] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. Journal of machine learning research, 2008. 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24634" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[53] Erik Roeland van Dop and Paulus P.L. Regtien. Fitting undeformed superquadrics to range data: improving model recovery and classification. International Conference on Computer Vision and Pattern Recognition (CVPR), 1998. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[54] Ashish Vaswani, Noam M. Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is All you Need. In International Conference on Neural Information Processing Systems (NeurIPS), 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[55] Giulia Vezzani, Ugo Pattacini, and Lorenzo Natale. A grasping approach based on superquadric models. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 1579-1586. IEEE, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.343 + ], + "angle": 0, + "content": "[56] Chenxi Wang, Hao-Shu Fang, Minghao Gou, Hongjie Fang, Jin Gao, and Cewu Lu. Graspness Discovery in Clutters for Fast and Accurate Grasp Detection. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.346, + 0.482, + 0.4 + ], + "angle": 0, + "content": "[57] Kaizhi Yang and Xuejin Chen. Unsupervised learning for cuboid shape abstraction via joint segmentation from point clouds. ACM Transactions On Graphics (TOG), 2021. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.403, + 0.482, + 0.457 + ], + "angle": 0, + "content": "[58] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. International Conference on Computer Vision (ICCV), 2023. 2, 4, 6, 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.46, + 0.482, + 0.542 + ], + "angle": 0, + "content": "[59] Chenyangguang Zhang, Alexandros Delitzas, Fangjinhua Wang, Ruida Zhang, Xiangyang Ji, Marc Pollefeys, and Francis Engelmann. Open-Vocabulary Functional 3D Scene Graphs for Real-World Indoor Spaces. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.544, + 0.482, + 0.585 + ], + "angle": 0, + "content": "[60] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. International Conference on Computer Vision (ICCV), 2023. 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "24635" + } + ] +] \ No newline at end of file diff --git a/2025/SuperDec_ 3D Scene Decomposition with Superquadrics Primitives/ed2eb095-495d-4dd3-9f36-9c1449d1ce9d_origin.pdf b/2025/SuperDec_ 3D Scene Decomposition with Superquadrics Primitives/ed2eb095-495d-4dd3-9f36-9c1449d1ce9d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b3d740cf9598ac7c8d754c648f3c4a5b7e494034 --- /dev/null +++ b/2025/SuperDec_ 3D Scene Decomposition with Superquadrics Primitives/ed2eb095-495d-4dd3-9f36-9c1449d1ce9d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b555a35e417535a5d0e3bddbac9340832a170fe12c65b5d3dbd07090809b96fb +size 8060498 diff --git a/2025/SuperDec_ 3D Scene Decomposition with Superquadrics Primitives/full.md b/2025/SuperDec_ 3D Scene Decomposition with Superquadrics Primitives/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dc7801a1295d6449046a9f0fb531ac8321c6226c --- /dev/null +++ b/2025/SuperDec_ 3D Scene Decomposition with Superquadrics Primitives/full.md @@ -0,0 +1,356 @@ +# SUPERDEC: 3D Scene Decomposition with Superquadric Primitives + +Elisabetta Fedele $^{1,2}$ + +Boyang Sun + +Leonidas Guibas + +Marc Pollefeys1,3 + +Francis Engelmann + +$^{1}$ ETH Zurich + +$^{2}$ Stanford University + +3Microsoft + +# Abstract + +We present SUPERDEC, an approach for creating compact 3D scene representations via decomposition into superquadric primitives. While most recent methods use geometric primitives to obtain photorealistic 3D reconstructions, we instead leverage them to obtain a compact yet expressive representation. To this end, we design a novel architecture that efficiently decomposes point clouds of arbitrary objects into a compact set of superquadrics. We train our model on ShapeNet and demonstrate its generalization capabilities on object instances from ScanNet++ as well as on full Replica scenes. Finally, we show that our compact superquadric-based representation supports a wide range of downstream applications, including robotic manipulation and controllable visual content generation. Project page: https://super-dec.github.io. + +# 1. Introduction + +3D scene representations play a central role in computer vision and robotics, enabling tasks such as 3D scene understanding [33, 48, 49], scene generation [40, 42, 43], functional reasoning [19, 59], and scene interaction [9, 13, 21, 38]. Recent work [18] utilizes 3D Gaussians as geometric primitives to produce high-quality, photorealistic reconstructions. However, such representations are often memory-intensive. In contrast, we propose a more lightweight yet geometrically faithful 3D scene representation by decomposing the input point cloud into a compact set of explicit primitives, namely superquadrics (Fig. 1). + +Representations for 3D scenes include well-established formats such as point clouds, meshes, signed distance functions, and voxel grids, each offering different trade-offs among geometric detail, computational cost, resolution, performance, interpretability, and editability. Recently, multi-view approaches such as Neural Radiance Fields (NeRF) [29] and Gaussian Splatting (GS) [18] have gained popularity as 3D scene representations. These methods optimize photometric losses to ensure that their underlying representations (implicit in the case of NeRF, and explicit in the case of GS) are consistent with observed images. While + +![](images/84cec3249a90f605c4bf77d716a779cfe529f568e00cfca34074d5699173ef81.jpg) + +![](images/5c1752c06b242dd1a2f062f39b7cb30a65b456fc79c0fdfc7528387eaf83a4ed.jpg) + +![](images/fee4ab809c7cb660769b3b3c811e49a4e109eb29d2763c96cbc2e1b9a54c9e29.jpg) +Figure 1. 3D Scene Decomposition with Superquadrics. Given a 3D point cloud of an indoor scene, SUPERDEC decomposes all scene objects into a compact set of superquadric primitives. Different object instances are visualized with different colors. + +![](images/6ec755e3bb8d3a158ef44d63135f557a4062c39c52b501b2163819ec69e8be30.jpg) + +these approaches excel at achieving photorealism, they lack explicit control over compactness, often resulting in large, non-modular scene encodings that are unsuitable for tasks requiring explicit spatial reasoning. + +While optimizing compactness on a scene level remains a challenging task, prior work has shown that geometric primitives such as cuboids [51, 57] or superquadrics [1, 31, 32] enable compact and interpretable decompositions of individual objects. These methods are generally either learning-based [32, 57], favoring speed at the cost of geometric accuracy, or optimization-based [1, 23, 31], offering improved accuracy at the expense of higher computational cost. Although both types of approaches can be effective for certain object classes, they often struggle to generalize across datasets with diverse object geometries. Learning-based methods typically require category-specific training, while optimization-based ones rely on hand-crafted heuristics, limiting their scalability in open-world settings. + +Motivated by the abstraction capabilities of geometric primitives for individual object categories, we propose to represent complex 3D scenes using a compact set of superquadrics. To this end, we learn class-agnostic object + +level shape priors to optimize compactness and leverage an off-the-shelf 3D instance segmentation method (Mask3D [41]) to scale our approach to full 3D scenes. + +We choose superquadrics as building blocks for our representation, as they offer more accurate shape modeling than cuboids while incurring minimal additional parameter overhead (9 v.s. 11, including 6-DoF pose parameters). To develop a model capable of generalizing across diverse object types, we draw inspiration from supervised segmentation [5, 6, 41], and approach the problem from the perspective of unsupervised geometric-based segmentation, using local point-based features to iteratively refine predicted geometric primitives. Our model is trained on ShapeNet [4] and evaluated on three challenging and diverse 3D datasets: ShapeNet [4], ScanNet++ [58], and Replica [45]. On the 3D object dataset ShapeNet, our approach achieves an L2 error six times smaller than prior state-of-the-art methods [32], while requiring only half the number of primitives. On ScanNet++ and Replica, we demonstrate that our method generalizes well to real-world scene-level settings, despite being trained solely on ShapeNet. Finally, we show the practical utility of our method as a scene representation for robotic tasks, including path planning and object grasping, as well as an editable 3D scene representation for controllable image generation. + +In summary, our contributions are the following: + +1) We introduce SUPERDEC, a novel method for decomposing 3D scenes using superquadric primitives. +2) SUPERDEC achieves state-of-the-art object decomposition scores on ShapeNet trained jointly on multiple classes. +3) We demonstrate the effectiveness of 3D superquadric scene representations for robotic tasks and controllable generative content creation. + +# 2. Related Work + +Learning-based methods have shown that neural networks, when equipped with suitable reconstruction losses, can directly predict geometric primitive parameters to decompose point clouds into a minimal set of primitives for specific object categories. Tulsiani [51] introduced a CNN-based method for cuboid decomposition, which was later extended to more expressive primitives such as superquadrics by Paschalidou et al. [32]. CSA [57] further enhanced interpretability by employing a stronger point encoder and jointly predicting cuboid parameters and part segmentations. However, these methods remain constrained by their reliance on category-specific training. We attribute this limitation to their model design, which encodes only global shape features, sufficient for intra-category generalization but ineffective for decomposing out-of-category objects. + +Optimization-based methods largely originate from the literature on superquadric fitting. EMS [23] revisited this line of work by introducing a probabilistic formulation that + +enables the decomposition of arbitrary objects into multiple superquadrics. Given an input point cloud, the method first fits a superquadric to the main structure and identifies unfitted outlier clusters, which are then recursively processed in a hierarchical fashion up to a predefined depth level. However, as noted in their paper and confirmed by our experiments, this approach implicitly assumes that objects exhibit a hierarchical geometric structure, limiting its applicability to many real-world objects such as tables and chairs. Other methods, such as Marching Primitives [24], require Signed Distance Functions (SDFs) as input, which are generally not easily available in real-world scenes. More fundamentally, since these approaches optimize from scratch for each object, they cannot leverage generalizable point features or learned shape priors, both of which are critical for abstraction and robustness under partial observations, a common challenge in practical 3D capture scenarios. + +Scene-level decomposition. With the emergence of 3DGS [18], an increasing number of works have explored representing 3D scenes using various geometric primitives as generalized ellipsoids [15] and convexes [14]. While heuristics can control the number of Gaussians, achieving truly compact representations remains challenging. DBW [31] addresses this by fitting a small set of textured superquadrics to 3D scenes, building on the principles of 3DGS. Given a set of scene images, it performs test-time optimization with a photometric loss and renders the primitives using a differentiable rasterizer [22]. To model the environment, DBW adds a meshed ground plane and a meshed icosphere for the background. However, it is restricted to scenes with fewer than 10 primitives and requires objects to be aligned to a ground plane. Furthermore, the optimization is computationally expensive, taking around three hours even on simple DTU [17] scenes. Our method differs significantly in terms of input requirements, generality of application, and computational efficiency. + +Superquadrics are a parametric family of shapes introduced by Barr et al. [2] in 1981 and have since been widely adopted in both computer vision and graphics [7, 34, 44]. Their popularity stems from their ability to represent a diverse range of shapes with a highly compact parameterization. A superquadric in its canonical pose is defined by just five parameters: $(s_x, s_y, s_z)$ for the scales along the three principal semi-axes and $(\epsilon_1, \epsilon_2)$ for the shape-defining exponents. Given those parameters, their surface is described by the implicit equation: + +$$ +f (\mathbf {x}) = \left(\left(\frac {x}{s _ {x}}\right) ^ {\frac {2}{\epsilon_ {2}}} + \left(\frac {y}{s _ {y}}\right) ^ {\frac {2}{\epsilon_ {2}}}\right) ^ {\frac {\epsilon_ {2}}{\epsilon_ {1}}} + \left(\frac {z}{s _ {z}}\right) ^ {\frac {2}{\epsilon_ {1}}} = 1. \tag {1} +$$ + +Extending this representation to a global coordinate system requires 6 additional parameters (3 for translation and 3 + +for rotation), resulting in a total of 11 parameters per superquadric. Another key property of superquadrics is the ability to compute the radial distance from any point in 3D space to the superquadric surface, i.e., the distance between a point and the superquadric's surface along the line connecting that point to the center of the superquadric. Specifically, given a point $\mathbf{x} \in \mathbb{R}^3$ , its radial distance to the surface of a canonically oriented superquadric is defined as: + +$$ +d _ {r} = | \mathbf {x} | \cdot | 1 - f (\mathbf {x}) ^ {- \epsilon_ {1} / 2} |, \tag {2} +$$ + +where $f(\mathbf{x})$ is given in Eq. 1. We refer the reader to [53] for the derivation of Eq. 2 and to [16] for a more comprehensive overview on superquadrics. + +# 3. Method + +Our ultimate goal is a 3D scene decomposition using superquadric primitives. To this end, we first focus on single-object decomposition and then show how our method, combined with 3D instance segmentation [41], can be applied to full 3D scenes. We detail the single-object approach in Sec. 3.1 and its extension to full scenes in Sec. 3.2. + +# 3.1. Single Object Decomposition + +Fig. 2 illustrates our model for single-object decomposition. It consists of two main components: a self-supervised feed-forward neural network that jointly predicts superquadric parameters and a segmentation matrix associating points to superquadrics, followed by a lightweight Levenberg-Marquardt (LM) optimization [20, 28]. + +# 3.1.1. Feed-forward Neural Network + +Our deep learning model draws inspiration from recent fully-supervised Transformer-based [54] segmentation models [5, 6, 41]. These models iteratively decode a sequence of queries, each representing a segmentation mask, by cross-attending to input pixels or points. In our case, the queries represent superquadrics. Next, we show how such an architecture can be adapted to unsupervisedly segment superquadrics, instead of supervisedly segment objects. + +Model Details. Given an input point cloud $\mathcal{P} \in \mathbb{R}^{N \times 3}$ , where each of the $N$ points has a 3D coordinate, we first extract rich point features $\mathcal{F}_{PC} \in \mathbb{R}^{N \times H}$ using the PVCNN [25] point encoder. At the same time, we initialize $P$ superquadrics features $\mathcal{F}_{SQ} \in \mathbb{R}^{P \times H}$ with sinusoidal positional encodings. We feed these features in a Transformer decoder [54] which leverages self-attention, cross-attention to the point features, and feed-forward layers to refine them. Once refined, the superquadric features $\mathcal{F}_{SQ}$ and the point features $\mathcal{F}_{PC}$ are fed into two prediction heads: The segmentation head takes as input $\mathcal{F}_{SQ}$ and $\mathcal{F}_{PC}$ and predicts a soft assignment matrix $M \in \mathbb{R}^{N \times P}$ associating points to + +superquadrics and whose elements are defined as: + +$$ +m _ {i j} = \sigma (\phi (\mathcal {F} _ {P C}) \cdot \mathcal {F} _ {S Q}), \tag {3} +$$ + +where $\phi (\mathcal{F}_{PC})\in \mathbb{R}^{N\times H}$ is a learned projection of the point features to match the dimensionality of the superquadric features, and $\sigma$ is the softmax function. The second head, the superquadric head, takes the superquadric features $\mathcal{F}_{SQ}$ as input and predicts 12 parameters for each superquadric: 11 encoding its 5-DoF shape and 6-DoF pose, and one modeling its existence probability $\alpha$ , enabling a variable number of superquadrics per object. + +Losses. We train our model in a self-supervised manner, without requiring any ground truth annotation. Specifically, the total loss is defined as: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {r e c}} + \lambda_ {\text {p a r}} \mathcal {L} _ {\text {p a r}} + \lambda_ {\text {e x i s t}} \mathcal {L} _ {\text {e x i s t}}, \tag {4} +$$ + +where $\mathcal{L}_{rec}$ is the reconstruction loss aligning the predicted superquadrics to the input point cloud $\mathcal{P}$ , $\mathcal{L}_{par}$ is the parsimony loss encouraging a small number of primitives, $\mathcal{L}_{exist}$ is the existence loss, and $\lambda_{par}, \lambda_{exist}$ are weighting coefficients. The reconstruction loss $\mathcal{L}_{rec}$ consists of three terms: + +$$ +\mathcal {L} _ {r e c} = \mathcal {L} _ {\mathcal {P} \rightarrow S Q} + \mathcal {L} _ {S Q \rightarrow \mathcal {P}} + \mathcal {L} _ {N}. \tag {5} +$$ + +The first two terms correspond to the bi-directional Chamfer distance between the input point cloud and the superquadric surfaces, while the third term serves as a regularizer incorporating normal information to improve convergence during training. To compute the Chamfer distance, we approximate each superquadric surface by uniformly sampling $S$ points, following the method of Pilu et al. [35]. Denoting by $d(\mathbf{x}_i,\mathbf{x}_{js}^{\prime})$ the euclidean distance between the $i$ -th point in the input point cloud and the $s$ -th point sampled on the surface of the $j$ -th superquadric, we define $\mathcal{L}_{\mathcal{P}\to SQ}$ as: + +$$ +\mathcal {L} _ {\mathcal {P} \rightarrow S Q} = \frac {1}{N} \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {P} m _ {i j} \min _ {s \in [ S ]} d \left(\mathbf {x} _ {i}, \mathbf {x} _ {j s} ^ {\prime}\right), \tag {6} +$$ + +and $\mathcal{L}_{SQ\to \mathcal{P}}$ as: + +$$ +\mathcal {L} _ {S Q \rightarrow \mathcal {P}} = \frac {1}{S \sum_ {j = 1} ^ {P} \alpha_ {j}} \sum_ {j = 1} ^ {P} \alpha_ {j} \sum_ {s = 1} ^ {S} \min _ {i \in [ N ]} d \left(\mathbf {x} _ {i}, \mathbf {x} _ {j s} ^ {\prime}\right). \tag {7} +$$ + +The last term of Eq. 5, i.e., $\mathcal{L}_N$ is defined as the reconstruction loss from Yang et al. [57], and is used to incorporate normal information during training which leads to accelerated convergence. Additionally, since we seek not only accuracy but also compactness, we introduce a parsimony loss to encourage the use of fewer primitives. To do that, we optimize the 0.5-norm of $m_j \coloneqq \sum_{i=1}^{N} \frac{m_{ij}}{N}$ and define the parsimony loss as: + +$$ +\mathcal {L} _ {p a r} = \left(\frac {1}{P} \sum_ {j = 1} ^ {P} \frac {\sqrt {m _ {j}}}{P}\right) ^ {2}. \tag {8} +$$ + +![](images/189e12d07aae7438d31a40affef30c862ea9e9b29fb2f7506ab4e5590c043883.jpg) +Figure 2. Illustration of the SUPERDEC Model. Given a point cloud of an object with $N$ points, a Transformer-based neural network predicts parameters for $P$ superquadrics, as well as a soft segmentation matrix that assigns points to superquadrics. The predicted parameters include the 11 superquadric parameters and an objectness score. These predictions provide an effective initialization for the subsequent Levenberg-Marquardt (LM) optimization, which refines the superquadrics. + +Lastly, we employ an existence loss $\mathcal{L}_{\text {exist }}$ which uses the predicted segmentation as a teacher for the linear head in charge of predicting the existence probability. More specifically, given a threshold $\epsilon_{\text {exist }}$ , we define the ground-truth existence of the $j$ th superquadric as $\hat{\alpha}_j \coloneqq m_j > \epsilon_{\text {exist }}$ and define $\mathcal{L}_{\text {exist }}$ as: + +$$ +\mathcal {L} _ {\text {e x i s t}} = \sum_ {j = 1} ^ {P} \frac {B C E \left(\alpha_ {j} , \hat {\alpha} _ {j}\right)}{P}, \tag {9} +$$ + +where BCE is the binary cross entropy and $\alpha_{j}$ is the predicted existence probability for the $j$ th superquadric. + +# 3.1.2. Optimization + +Our optimization module takes as input the predicted soft segmentation matrix $M$ as well as the superquadric parameters $\Theta$ , and further refines the superquadric parameters using the Levenberg-Marquardt (LM) [20, 28] algorithm. Specifically, given a point cloud of $N$ points, it iteratively refines the parameters $\Theta_{j}$ of the $j$ th superquadric, by computing two sets of residuals: The first set of residuals $r_{ij}$ with $i \in [1, N]$ and $j \in [1, P]$ is defined as: + +$$ +r _ {i j} = m _ {i j} \tilde {d} _ {j} \left(\mathbf {x} _ {i}\right), \tag {10} +$$ + +where $\tilde{d}_j(\mathbf{x}_i)$ denotes the radial distance of point $\mathbf{x}_i$ from the $j$ th superquadric, computed according to Eq. 2. The second set of residuals is used for normalization and is obtained by sampling a set of $K$ points $\mathbf{p_1},\ldots ,\mathbf{p_K}$ on the surface of the given superquadric and then computing the distance of each of them from the point cloud. Specifically, for $i\in [N + 1,N + K]$ and $j\in [1,P]$ we compute $r_{ij}$ as: + +$$ +r _ {i j} = \min _ {k} \left\| \mathbf {p} _ {\mathbf {i} - \mathbf {N}} - \Pi_ {j} \left(\mathbf {x} _ {\mathbf {k}}\right) \right\| _ {2}, \quad \text {w i t h} k \in [ 1, N ]. \tag {11} +$$ + +# 3.2. Decomposition of Full 3D Scenes + +After training on single objects, extending SUPERDEC to full 3D scenes is straightforward. Given a scene-level point cloud, we extract 3D object instance masks using Mask3D [41]. Each object is centered and uniformly rescaled to a sphere of radius 0.5. We then predict the superquadric primitives for each object individually using our model. We found our model trained on ShapeNet [4] to generalize well on real-world 3D scenes from ScanNet++ [58] and Replica [45] without additional fine-tuning. + +# 4. Experiments + +We first compare our SUPERDEC with previous state-of-the-art methods on individual objects and full 3D scenes (Sec. 4.1). We then demonstrate the usefulness of our representation on down-stream applications for robotics and controllable image generation (Sec. 4.2). Finally, in Sec. 4.3, we present additional analyses on part segmentation and the implicit learning of shape categories, followed by a study of the compactness-accuracy trade-off and runtime. + +# 4.1. Comparing with State-of-the-art Methods + +Datasets. We compare on three different datasets: ShapeNet [4]: We use the 13 classes of the ShapeNet subset and train-val-test splits as defined in Choy et al. [8]. All objects are pre-aligned in a canonical orientation. For each object we sample 4096 points using Farthest Point Sampling (FPS) [36]. ShapeNet is a widely used dataset and is well-suited for comparison with existing baselines. + +ScanNet++ [58]: We further evaluate our model on real-world object scans from the ScanNet++ validation set. Each object is extracted using ground truth mask annotations, and 4096 points per object are sampled with FPS. In contrast to ShapeNet, these object point clouds are noisier, partially ob + +
ModelPrimitive TypeSegmentationIn-categoryOut-of-category
L1 ↓L2 ↓# Prim.↓L1 ↓L2 ↓# Prim.↓
EMS (Liu et al.) [23]SuperquadricsX5.7711.3455.685.4101.2115.68
CSA (Yang et al.) [57]Cuboids✓5.1570.5279.214.8970.42711.75
SQ (Paschalidou et al.) [32]SuperquadricsX3.6680.279104.1930.3549
SUPERDEC (Ours)Superquadrics✓1.6980.0475.81.8470.0615.26
+ +Table 1. Quantitative Results on ShapeNet [4]. We show scores for in-category and out-of-category experiments and are scaled by $10^{2}$ . + +![](images/4dec65a8552e92cd98987fe6a1aa90b6ed09ba9c3a8f3067ec3907771b35d527.jpg) +Figure 3. Qualitative Results on ShapeNet [4]. We show results on test samples for in-category (four first columns) classes and out-of-category classes (two last columns). The latter were not seen during training and illustrate how well models generalize to novel classes. + +served, and subject to random orientation and translation, providing a more realistic and challenging evaluation setting for our method. + +Replica [45]: Finally, we present both quantitative and qualitative results on full 3D scenes from Replica. Object instances are extracted using either ground truth instance annotations or the pre-trained 3D instance segmentation model Mask3D [41], enabling us to evaluate our approach in a fully realistic setting, including scenarios where no ground truth annotations are available. + +Methods in comparison. We compare to learning- and optimization-based prior works using both cuboids and superquadrics as geometric primitives. $SQ$ [32] is a learning-based approach for object-level decomposition using superquadrics; it takes a voxel grid as input and predicts superquadric primitives via a CNN. CSA [57] is another learning-based method but uses cuboids as geometric prim + +ities. It takes a point cloud as input and predicts cuboid parameters from a global latent code. Lastly, EMS [23] is an optimization-based approach that decomposes objects by hierarchically fitting superquadrics to parts of a point cloud. + +Training Details. Our goal is to develop a general-purpose, class-agnostic model capable of representing arbitrary objects as superquadrics. Existing methods typically train separate models for each object class, assuming that all classes are known in advance and that sufficient training data is available for each. These assumptions, however, often fail in real-world scenarios. To address this, we jointly train a single model on all 13 ShapeNet classes using the publicly available code of prior methods, moving towards a more realistic class-agnostic solution. In our model we set the following hyper-parameters $P = 16$ , $S = 4096$ , $K = 25$ , $H = 128$ , $D = 3$ , $\epsilon_{\text{exist}} = 24$ , $\lambda_{\text{exist}} = 0.01$ , $\lambda_{\text{par}} = 0.06$ . We refer to the supplementary for additional details. + +
ScanNet++ [58]Replica [45]
MethodL1 ↓L2 ↓#Prim.↓L1 ↓L2 ↓#Prim.↓
SQ [32]10.454.2610.011.845.9110
EMS [23]5.512.114.255.402.123.61
CSA [57]2.910.4111.643.680.709.63
SUPERDEC (Ours)1.700.115.181.790.196.58
+ +Metrics. We assess reconstruction accuracy using L1 and L2 Chamfer distances and compactness by the average number of geometric primitives. + +# 4.1.1. Results on ShapeNet + +We show scores in Tab. 1 and qualitative results in Fig. 3. To evaluate both accuracy and generalization, we conduct two experiments: in-category and out-of-category. In the in-category setting, all learning-based methods are jointly trained on the 13 classes of the ShapeNet training set and evaluated on the corresponding test set. In the out-of-category setting, models are trained on half of the categories (airplane, bench, chair, lamp, rifle, table) and tested on the remaining ones (car, sofa, loudspeaker, cabinet, display, telephone, watercraft). Our SUPERDEC model significantly outperforms both learned and non-learned baselines. Compared to learned baselines, we reduce the L2 loss by a factor of six while using nearly half the number of primitives, supporting our hypothesis that leveraging local point features as opposed to a single global descriptor improves 3D decomposition in both accuracy and compactness. Compared to the non-learned baseline, we predict a similar number of primitives but achieve an L2 loss approximately 20 times smaller, validating the benefit of learning shape priors to avoid local minima that often hinder purely optimization-based approaches. + +# 4.1.2. Results on 3D Scenes + +In this section, models are evaluated on real-world, out-of-category objects, which appear in arbitrary orientations and often exhibit incomplete point clouds due to reconstruction artifacts and occlusions. Tab. 2 shows the quantitative results. Despite never being trained on real-world objects, our method outperforms both the optimization- and the learning-based baselines by a large margin. Lastly, we qualitatively evaluate our pipeline on full 3D scenes from Replica, where our object-level model is applied on top of class-agnostic instance segmentation predictions from Mask3D [41]. As shown in Fig. 1, our method effectively reconstructs object shapes, even under noisy segmentation masks and geometries that differ substantially from those seen during training. We refer to the supplementary for additional qualitative results. + +Table 2. Quantitative Results on Scenes. Evaluation on objects from 3D scene datasets [45, 58]. Scores are scaled by $10^{2}$ . + +
MethodTime (ms)Suc. (%)Mem. (MB)
Occupancy0.056100.000.873
PointCloud0.06389.5719.286
Voxels0.03098.780.101
Cuboids [37]0.12061.230.024
SUPERDEC0.15091.710.042
+ +Table 3. Path Planning Scores. Mean on 15 ScanNet++ scenes. + +![](images/43864e64c65f7c33f8e0a90b6592b8790c4f5fc401010af062544c0af5d85ecc.jpg) +Figure 4. Grasping Result. Visualization of computed grasp poses for a milk bottle, some flowers, a side table, and a plant. + +# 4.2. Down-stream Applications + +Next, we show the versatility of the SUPERDEC representation for downstream applications, including robotics tasks such as path planning and object grasping (Sec. 4.2.1), and controllable image generation (Sec. 4.2.2). + +# 4.2.1. Robotics + +Path planning seeks to compute a collision-free shortest path between a given start and end point in 3D space, enabling efficient robot navigation. Although essential for traversing large environments, it typically demands storing large-scale 3D representations. Here, we assess whether our compact representation can perform this task effectively while reducing memory requirements. We conduct experiments on 15 ScanNet++ [58] scenes, comparing SUPERDEC to common 3D representations, including dense occupancy grids, point clouds, voxel grids, and cuboids [37]. As shown in Tab. 6, SUPERDEC not only reduces memory consumption compared to traditional representations but also achieves a higher success rate than dense point clouds. Further details about experiment setup, metrics and analysis are provided in the Appendix. + +Object Grasping enables robots to grasp real-world objects by computing suitable grasping poses. Existing methods fall into two categories, each with complementary limitations. Geometry-based approaches [3, 12, 30] require precise 3D object models, which are often unavailable in real-world scenarios. Learning-based approaches [27, 47, 56] operate directly on raw sensor data but tend to be biased towards training data, which typically consists of tabletop scenes with small, convex, or low-genus objects [11, 46]. To overcome these limitations, SuperQ-GRASP [50] explored decomposing objects into explicit primitives. How + +![](images/96ccec016ec636580169833e21ee26e396184942cbfe518eac6d358d802342e4.jpg) +Input Point Cloud + +![](images/3b73dddd69eb9d8ff84dcabf84c9d6171d5ca57bfb6c698c1b2cb323c63503e7.jpg) +Superquadrics + +![](images/872c27960f652358d2624d27a8fbb67cf6b9d439573594ac947f32f90445e908.jpg) +Figure 5. Real-world robot experiment. The top row shows the input scan (left) and the representation from SUPERDEC with the computed path and grasping pose (right). The bottom row illustrates the robot following the planned path. We denote the starting point of the path with a green sphere, and the target location with a red sphere. The target object (a milk bottle) is circled in red. + +ever, its reliance on Marching Primitives [24] to obtain superquadrics from the object's Signed Distance Function (SDF) makes it unsuitable for most real-world cases where only point clouds are available. In contrast, our approach directly processes point clouds of entire scenes and, when combined with the class-agnostic segmentations from Mask3D [41], extracts superquadrics for all objects. Given the superquadric parameters, we employ a superquadric-based geometric method [55] to compute grasping poses for selected objects. Fig. 4 shows predicted grasping poses on objects from a real-world 3D scan of a room. In practice, our method eliminates the need for data-driven grasping models while remaining adaptable to diverse object shapes and producing high-quality grasping poses. + +Real-world Experiment. Finally, we demonstrate the real-world applicability of our superquadric-based representation by deploying it on a legged robot (Boston Dynamics Spot) equipped with an arm, supporting both motion planning and object grasping in an indoor environment. We scan a scene using a 3D scanning application on an iPad, extract a dense point cloud, and run SUPERDEC on it. Given the robot's starting position and a specified target object (a milk bottle), we compute both the path and the grasping pose as described earlier (see Fig. 4), enabling the robot to approach and successfully grasp the object. Fig. 5 shows the computed representation, the planned trajectory, the grasping pose, and a frame from the real-world demonstration. This experiment suggests that integrating SUPERDEC with open-vocabulary segmentation methods such as OpenMask3D [48] could allow robots to navigate to and grasp arbitrary objects specified via natural-language prompts. + +![](images/e67dd82e63110d293547f0ef7d56641338114490b24f5f21c59e377da3a618b1.jpg) +Original +Editing +Addition +Deletion + +![](images/2143ccb01f4229cb1edb7587b0fe971d16e0e17f835fcd8eec0480af3115e7cb.jpg) +Figure 6. Spatial control using SUPERDEC. Top row shows superquadrics generated by SUPERDEC, bottom row shows generated images using the prompt 'A corner of a room with a plant'. + +![](images/549a00a9f8ad1734a9a270a4598af061adff061ce951c65e7b9e10375d1407f7.jpg) +Superquadrics +"Pink living room" + +![](images/2fa33d0f2387e0deaa6bfa7c02eebbd4b83976db59e761379758333d07597226.jpg) +Depth Prompt +"Modern living room" + +![](images/3e59075a06027fbc2bd7003279a192ddc6c52423476fa057dd559f77108d379c.jpg) +Figure 7. Semantic control using SUPERDEC. Top row: superquadrics created by our SUPERDEC, and depth map to prompt the generation of text-to-image diffusion model. Bottom row: generations with two different textual prompts. + +![](images/5eb20ce3b4c02010a3963f79146e31a91393b7ab52a8e40fda7ec8b8044582cd.jpg) + +# 4.2.2. Controllable Generation and Editing + +We investigate how the SUPERDEC representation can be used to introduce joint spatial and semantic control in text-to-image diffusion models [39]. Specifically, we generate images by conditioning ControlNet [60] on depth maps rendered from the superquadrics extracted from Replica [45] scenes. Qualitative results are shown in Fig. 6 and Fig. 7. Fig. 6 demonstrates spatial control: by moving, duplicating, or removing superquadrics corresponding to a plant, we coherently influence the generated images. Fig. 7 highlights semantic control: we can vary the room's style while preserving its semantic and geometric structure, and observe that object semantics naturally emerge from the spatial arrangement of superquadrics without explicit conditioning, e.g., pillows appear on couches, and a plant is placed on a central table, reflecting plausible real-world arrangements. + +![](images/4ee67679311c87aefdbf97b48165d576960955d6dace3c71d86dd1c6f9f1a42c.jpg) +Figure 8. Qualitative Results on ShapeNet [4] segmentation. We show the resulting segmentation matrices on test samples for in-category (four first columns) classes and out-of-category classes (two last columns). The latter were not seen during training. + +![](images/e2ed7ba8ef746fe51ffd49bb3304a34ba9e753e7338d1f8610a96291247c4f26.jpg) +Input 3D Scan + +![](images/685b7d9f74bb6c0449748e930066efc7dd478d39a43152c54b9839facfbbf6e2.jpg) +$1^{\mathrm{st}}$ Hierarchy +$2^{\mathrm{nd}}$ Hierarchy +Figure 9. Hierarchical decomposition with SUPERDEC. From left to right we visualize: the input point cloud, the superquadrics predicted in the first hierarchy, the superquadrics predicted in the second hierarchy. + +# 4.3. Analysis Experiments + +Unsupervised part segmentation. Besides superquadric parameters, our method also predicts a segmentation matrix which segments the initial point cloud into parts that are fitted to the predicted superquadrics. In Fig. 8, we visualize the predicted segmentation masks for the same examples shown in Fig. 3. We observe that segmentation masks, especially in the in-category experiments, appear very sharp. This suggests that our method, especially if trained at a larger scale, can be leveraged for different applications as geometry-based part segmentation or as pretraining for supervised semantic part segmentation. + +Can we do hierarchical decomposition? We run SUPERDEC in a hierarchical fashion on a whole simple scene, without using any instance segmentation method to segment the scene into different objects. We observed that by applying SUPERDEC on the whole point cloud it can coarsely segment the scene into components such as grass, table, and chairs. By applying the model again on the points segmented during the first prediction, we obtained a more accurate decomposition of the individual object instances. Given that SUPERDEC was trained only on single object instances from ShapeNet [4] we found this results very interesting. We see future potential in the use of our model to obtain unsupervised hierarchical decompositions. + +What does our network learn? Since our network performs unsupervised part segmentation, we analyze the features learned by the Transformer decoder across object classes. Inspired by BERT [10]'s [CLS] token, we append a learnable embedding to the sequence of embedded + +![](images/927a6ac7e941d7c142444a279e78b7f93a0eda48241105913c330a7785d765b1.jpg) +Figure 10. t-SNE Visualization of Primitive Embeddings across different ShapeNet classes. + +superquadrics; although never explicitly decoded, this embedding is refined through self- and cross-attention. After training, we extract and visualize these embeddings using t-SNE [52] for ShapeNet [4] categories (Fig. 10). We observe that categories with consistent shapes, such as chairs, airplanes, and cars, form clear clusters, while categories with high intra-class variability, such as watercraft, are more dispersed. This indicates that our model organizes objects by geometric structure without requiring class annotations. + +How fast is our method? Our model is highly parallelizable, allowing multiple objects to be batched and processed simultaneously in a single forward pass. On an RTX 4090 (24 GB), we can process up to 256 objects in parallel. On average, the forward pass takes $0.13\mathrm{s}$ for a complete Replica [45] scene, 3D instance segmentation with Mask3D [41] requires $0.3\mathrm{s}$ , and each LM optimization step takes less than 1 s, see supplementary for more details. + +# 5. Conclusion + +We proposed SUPERDEC, a method for deriving compact yet expressive 3D scene representations based on simple geometric primitives – specifically, superquadrics. Our model outperforms prior primitive-based methods and generalizes well to out-of-category classes. We further demonstrated the potential of the resulting 3D scene representation for various applications in robotics, and as a geometric prompt for diffusion-based image generation. While this is only a first step towards more compact, geometry-aware 3D scene representations, we anticipate broader applications and expect to see further research in this direction. + +Acknowledgments. Elisabetta Fedele is supported by the ETH AI Center doctoral fellowship, by the Swiss National Science Foundation (SNSF) Advanced Grant 216260 (Beyond Frozen Worlds: Capturing Functional 3D Digital Twins from the Real World), and an SNSF Mobility Grant. Francis Engelmann is supported by an SNSF PostDoc.mobility grant. We also gratefully acknowledge a GPU grant from NVIDIA. + +# References + +[1] Stephan Alaniz, Massimiliano Mancini, and Zeynep Akata. Iterative superquadric decomposition of 3d objects from multiple viewsd. In International Conference on Computer Vision (ICCV), 2023. 1 +[2] Alan H. Barr. Superquadrics and angle-preserving transformations. IEEE Computer Graphics and Applications, 1981. 2 +[3] Junhao Cai, Jingcheng Su, Zida Zhou, Hui Cheng, Qifeng Chen, and Michael Yu Wang. Volumetric-based contact point detection for 7-dof grasping. In Conference on Robot Learning (CoRL), 2022. 6 +[4] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical report, 2015. 2, 4, 5, 8 +[5] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention Mask Transformer for Universal Image Segmentation. International Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3 +[6] Bowen Cheng, Alexander G. Schwing, and Alexander Kirillov. Per-Pixel Classification is Not All You Need for Semantic Segmentation. In International Conference on Neural Information Processing Systems (NeurIPS), 2021. 2, 3 +[7] Laurent Chevalier, Fabrice Jaillet, and Atilla Baskurt. Segmentation and superquadric modeling of 3d objects. In International Conference in Central Europe on Computer Graphics and Visualization, 2003. 2 +[8] Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3D-R2N2: A Unified Approach for Single and Multi-view 3D Object Reconstruction. In European Conference on Computer Vision (ECCV), 2016. 4 +[9] Alexandros Delitzas, Ayca Takmaz, Federico Tombari, Robert Sumner, Marc Pollefeys, and Francis Engelmann. SceneFun3D: Fine-Grained Functionality and Affordance Understanding in 3D Scenes. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1 +[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies., 2019. 8 +[11] Hao-Shu Fang, Chenxi Wang, Hongjie Fang, Minghao Gou, Jirong Liu, Hengxu Yan, Wenhai Liu, Yichen Xie, and Cewu Lu. Anygrasp: Robust and efficient grasp perception in spatial and temporal domains. IEEE Transactions on Robotics, 2023. 6 +[12] B. Faverjon and J. Ponce. On Computing Two-finger Force-closure Grasps of Curved 2D Objects. In International Conference on Robotics and Automation (ICRA), 1991. 6 +[13] Qiao Gu, Ali Kuwajerwala, Sacha Morin, Krishna Murthy Jatavallabhula, Bipasha Sen, Aditya Agarwal, Corban Rivera, William Paul, Kirsty Ellis, Rama Chellappa, et al. + +Conceptgraphs: Open-vocabulary 3d scene graphs for perception and planning. In International Conference on Robotics and Automation (ICRA), 2024. 1 +[14] Abdullah Hamdi, Luke Melas-Kyriazi, Jinjie Mai, Guocheng Qian, Ruoshi Liu, Carl Vondrick, Bernard Ghanem, and Andrea Vedaldi. Ges: Generalized exponential splatting for efficient radiance field rendering. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[15] Jan Held, Renaud Vandeghen, Abdullah Hamdi, Adrien Deli'ege, Anthony Cioppa, Silvio Giancola, Andrea Vedaldi, Bernard Ghanem, and Marc Van Droogenbroeck. 3D convex splatting: Radiance field rendering with 3D smooth convexes. ArXiv, 2024. 2 +[16] Ales Jaklic, Ales Leonardis, and Franc Solina. Segmentation and recovery of superquadrics. Springer Science & Business Media, 2000. 3, 1 +[17] Rasmus Ramsbøl Jensen, A. Dahl, George Vogiatzis, Engil Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. International Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 2 +[18] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions On Graphics (TOG), 2023. 1, 2 +[19] Sebastian Koch, Narunas Vaskevicius, Mirco Colosi, Pedro Hermosilla, and Timo Ropinski. Open3DSG: Open-vocabulary 3D Scene Graphs from Point Clouds with Queryable Objects and Open-set Relationships. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1 +[20] Kenneth Levenberg. A method for the solution of certain non-linear problems in least squares. Quarterly of Applied Mathematics, 1944. 3, 4 +[21] Lei Li and Angela Dai. Genzi: Zero-shot 3d human-scene interaction generation. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1 +[22] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. International Conference on Computer Vision (ICCV), 2019. 2 +[23] Weixiao Liu, Yuwei Wu, Sipu Ruan, and Gregory S Chirikjian. Robust and Accurate Superquadric Recovery: A Probabilistic Approach. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 5, 6 +[24] Weixiao Liu, Yuwei Wu, Sipu Ruan, and Gregory S Chirikjian. Marching-primitives: Shape abstraction from signed distance function. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 7 +[25] Zhijian Liu, Haotian Tang, Yujun Lin, and Song Han. PointVoxel CNN for Efficient 3D Deep Learning. In International Conference on Neural Information Processing Systems (NeurIPS), 2019. 3 +[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. International Conference on Learning Representations (ICLR), 2019. 1 + +[27] Jeffrey Mahler, Matthew Matl, Vishal Satish, Michael Danielczuk, Bill DeRose, Stephen McKinley, and Ken Goldberg. Learning ambidextrous robot grasping policies. Science Robotics, 2019. 6 +[28] Donald W. Marquardt. An algorithm for least-squares estimation of nonlinear parameters. Journal of The Society for Industrial and Applied Mathematics, 1963. 3, 4 +[29] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 1 +[30] A.T. Miller, S. Knoop, H.I. Christensen, and P.K. Allen. Automatic Grasp Planning Using Shape Primitives. In International Conference on Robotics and Automation (ICRA), 2003. 6 +[31] Tom Monnier, Jake Austin, Angjoo Kanazawa, Alexei A. Efros, and Mathieu Aubry. Differentiable blocks world: Qualitative 3d decomposition by rendering primitives. International Conference on Neural Information Processing Systems (NeurIPS), 2023. 1, 2 +[32] Despoina Paschalidou, Ali Osman Ulusoy, and Andreas Geiger. Superquadrics revisited: Learning 3d shape parsing beyond cuboids. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2, 5, 6 +[33] Songyou Peng, Kyle Genova, Chiyu Max Jiang, Andrea Tagliasacchi, Marc Pollefeys, and Thomas Funkhouser. OpenScene: 3D Scene Understanding with Open Vocabularies. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1 +[34] Alex P Pentland. Parts: structured descriptions of shape. In Proceedings of the Fifth AAAI National Conference on Artificial Intelligence, 1986. 2 +[35] Maurizio Pilu and Robert B Fisher. Equal-distance sampling of superellipse models. In British Machine Vision Conference (BMVC), 1995. 3 +[36] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space. International Conference on Neural Information Processing Systems (NeurIPS), 2017. 4 +[37] Michael Ramamonjisoa, Sinisa Stekovic, and Vincent Lepetit. MonteBoxFinder: Detecting and Filtering Primitives to Fit a Noisy Point Cloud. In European Conference on Computer Vision (ECCV), 2022. 6, 5 +[38] Aaron Ray, Christopher Bradley, Luca Carlone, and Nicholas Roy. Task and motion planning in hierarchical 3d scene graphs. arXiv preprint arXiv:2403.08094, 2024. 1 +[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. International Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 7 +[40] Zuria Bauer Marc Pollefeys Federico Tombari Leonidas Guibas Gao Huang Francis Engelmann Rui Huang, Guangyao Zhai. VIPs: Video Perception Model for 3D Scene Generation. arXiv preprint, 2025. 1 + +[41] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask transformer for 3d semantic instance segmentation. International Conference on Robotics and Automation (ICRA), 2022. 2, 3, 4, 5, 6, 7, 8 +[42] Jonas Schult, Sam Tsai, Lukas Hollein, Bichen Wu, Jialiang Wang, Chih-Yao Ma, Kunpeng Li, Xiaofang Wang, Felix Wimbauer, Zijian He, et al. Controlroom3d: Room generation using semantic proxy rooms. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1 +[43] Jaidev Shriram, Alex Trevithick, Lingjie Liu, and Ravi Ramamoorthi. Realmdreamer: Text-driven 3d scene generation with inpainting and depth diffusion. arXiv preprint arXiv:2404.07199, 2024. 1 +[44] Franc Solina and Ruzena Bajcsy. Recovery of parametric models from range images: The case for superquadrics with global deformations. Transactions on Pattern Analysis and Machine Intelligence (PAMI), 1990. 2 +[45] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Yuheng Ren, Shobhit Verma, Anton Clarkson, Ming Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke Malte Strasdat, Renzo De Nardi, Michael Goesele, S. Lovegrove, and Richard A. Newcombe. The replica dataset: A digital replica of indoor spaces. ArXiv, 2019. 2, 4, 5, 6, 7, 8, 1, 3 +[46] Matan Sudry, Tom Jurgenson, Aviv Tamar, and Erez Karpas. Hierarchical planning for rope manipulation using knot theory and a learned inverse model. In Conference on Robot Learning, 2023. 6 +[47] Martin Sundermeyer, Arsalan Mousavian, Rudolph Triebel, and Dieter Fox. Contact-grasnet: Efficient 6-dof grasp generation in cluttered scenes. In International Conference on Robotics and Automation (ICRA), 2021. 6 +[48] Ayca Takmaz, Elisabetta Fedele, Robert W. Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. OpenMask3D: Open-Vocabulary 3D Instance Segmentation. In International Conference on Neural Information Processing Systems (NeurIPS), 2023. 1, 7 +[49] Ayca Takmaz, Alexandros Delitzas, Robert W. Sumner, Francis Engelmann, Johanna Wald, and Federico Tombari. Search3D: Hierarchical Open-Vocabulary 3D Segmentation. IEEE Robotics and Automation Letters (RA-L), 2025. 1 +[50] Xun Tu and Karthik Desingh. Superq-grasp: Superquadrics-based grasp pose estimation on larger objects for mobile-manipulation. arXiv, 2024. 6 +[51] Shubham Tulsiani, Hao Su, Leonidas J. Guibas, Alexei A. Efros, and Jitendra Malik. Learning shape abstractions by assembling volumetric primitives. International Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1, 2 +[52] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. Journal of machine learning research, 2008. 8 + +[53] Erik Roeland van Dop and Paulus P.L. Regtien. Fitting undeformed superquadrics to range data: improving model recovery and classification. International Conference on Computer Vision and Pattern Recognition (CVPR), 1998. 3 +[54] Ashish Vaswani, Noam M. Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is All you Need. In International Conference on Neural Information Processing Systems (NeurIPS), 2017. 3 +[55] Giulia Vezzani, Ugo Pattacini, and Lorenzo Natale. A grasping approach based on superquadric models. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 1579-1586. IEEE, 2017. 7 +[56] Chenxi Wang, Hao-Shu Fang, Minghao Gou, Hongjie Fang, Jin Gao, and Cewu Lu. Graspness Discovery in Clutters for Fast and Accurate Grasp Detection. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6 +[57] Kaizhi Yang and Xuejin Chen. Unsupervised learning for cuboid shape abstraction via joint segmentation from point clouds. ACM Transactions On Graphics (TOG), 2021. 1, 2, 3, 5, 6 +[58] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. International Conference on Computer Vision (ICCV), 2023. 2, 4, 6, 1, 3, 5 +[59] Chenyangguang Zhang, Alexandros Delitzas, Fangjinhua Wang, Ruida Zhang, Xiangyang Ji, Marc Pollefeys, and Francis Engelmann. Open-Vocabulary Functional 3D Scene Graphs for Real-World Indoor Spaces. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 1 +[60] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. International Conference on Computer Vision (ICCV), 2023. 7 \ No newline at end of file diff --git a/2025/SuperDec_ 3D Scene Decomposition with Superquadrics Primitives/images.zip b/2025/SuperDec_ 3D Scene Decomposition with Superquadrics Primitives/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8559126b5d5bf31a9b3acda6cfb6a8be31b22b06 --- /dev/null +++ b/2025/SuperDec_ 3D Scene Decomposition with Superquadrics Primitives/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aae94e09ec8955c192f34a92e44564ea160dd53547d1b262261b52f211590378 +size 594021 diff --git a/2025/SuperDec_ 3D Scene Decomposition with Superquadrics Primitives/layout.json b/2025/SuperDec_ 3D Scene Decomposition with Superquadrics Primitives/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..847658a541b1a0b1bc9fb25f752787d2d442b8a7 --- /dev/null +++ b/2025/SuperDec_ 3D Scene Decomposition with Superquadrics Primitives/layout.json @@ -0,0 +1,9330 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 95, + 103, + 515, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 103, + 515, + 121 + ], + "spans": [ + { + "bbox": [ + 95, + 103, + 515, + 121 + ], + "type": "text", + "content": "SUPERDEC: 3D Scene Decomposition with Superquadric Primitives" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 130, + 159, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 130, + 159, + 146 + ], + "spans": [ + { + "bbox": [ + 61, + 130, + 159, + 146 + ], + "type": "text", + "content": "Elisabetta Fedele" + }, + { + "bbox": [ + 61, + 130, + 159, + 146 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 171, + 131, + 237, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 131, + 237, + 147 + ], + "spans": [ + { + "bbox": [ + 171, + 131, + 237, + 147 + ], + "type": "text", + "content": "Boyang Sun" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 250, + 131, + 339, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 131, + 339, + 146 + ], + "spans": [ + { + "bbox": [ + 250, + 131, + 339, + 146 + ], + "type": "text", + "content": "Leonidas Guibas" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 350, + 131, + 435, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 131, + 435, + 147 + ], + "spans": [ + { + "bbox": [ + 350, + 131, + 435, + 147 + ], + "type": "text", + "content": "Marc Pollefeys1,3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 449, + 131, + 549, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 449, + 131, + 549, + 147 + ], + "spans": [ + { + "bbox": [ + 449, + 131, + 549, + 147 + ], + "type": "text", + "content": "Francis Engelmann" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 202, + 157, + 253, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 157, + 253, + 169 + ], + "spans": [ + { + "bbox": [ + 202, + 157, + 253, + 169 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 202, + 157, + 253, + 169 + ], + "type": "text", + "content": "ETH Zurich" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 270, + 157, + 347, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 157, + 347, + 170 + ], + "spans": [ + { + "bbox": [ + 270, + 157, + 347, + 170 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 270, + 157, + 347, + 170 + ], + "type": "text", + "content": "Stanford University" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 364, + 157, + 406, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 157, + 406, + 170 + ], + "spans": [ + { + "bbox": [ + 364, + 157, + 406, + 170 + ], + "type": "text", + "content": "3Microsoft" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 151, + 198, + 200, + 210 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 198, + 200, + 210 + ], + "spans": [ + { + "bbox": [ + 151, + 198, + 200, + 210 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 224, + 296, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 296, + 403 + ], + "type": "text", + "content": "We present SUPERDEC, an approach for creating compact 3D scene representations via decomposition into superquadric primitives. While most recent methods use geometric primitives to obtain photorealistic 3D reconstructions, we instead leverage them to obtain a compact yet expressive representation. To this end, we design a novel architecture that efficiently decomposes point clouds of arbitrary objects into a compact set of superquadrics. We train our model on ShapeNet and demonstrate its generalization capabilities on object instances from ScanNet++ as well as on full Replica scenes. Finally, we show that our compact superquadric-based representation supports a wide range of downstream applications, including robotic manipulation and controllable visual content generation. Project page: https://super-dec.github.io." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 428, + 135, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 428, + 135, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 428, + 135, + 441 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 449, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 581 + ], + "type": "text", + "content": "3D scene representations play a central role in computer vision and robotics, enabling tasks such as 3D scene understanding [33, 48, 49], scene generation [40, 42, 43], functional reasoning [19, 59], and scene interaction [9, 13, 21, 38]. Recent work [18] utilizes 3D Gaussians as geometric primitives to produce high-quality, photorealistic reconstructions. However, such representations are often memory-intensive. In contrast, we propose a more lightweight yet geometrically faithful 3D scene representation by decomposing the input point cloud into a compact set of explicit primitives, namely superquadrics (Fig. 1)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "Representations for 3D scenes include well-established formats such as point clouds, meshes, signed distance functions, and voxel grids, each offering different trade-offs among geometric detail, computational cost, resolution, performance, interpretability, and editability. Recently, multi-view approaches such as Neural Radiance Fields (NeRF) [29] and Gaussian Splatting (GS) [18] have gained popularity as 3D scene representations. These methods optimize photometric losses to ensure that their underlying representations (implicit in the case of NeRF, and explicit in the case of GS) are consistent with observed images. While" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 317, + 197, + 434, + 285 + ], + "blocks": [ + { + "bbox": [ + 317, + 197, + 434, + 285 + ], + "lines": [ + { + "bbox": [ + 317, + 197, + 434, + 285 + ], + "spans": [ + { + "bbox": [ + 317, + 197, + 434, + 285 + ], + "type": "image", + "image_path": "84cec3249a90f605c4bf77d716a779cfe529f568e00cfca34074d5699173ef81.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 434, + 197, + 553, + 285 + ], + "blocks": [ + { + "bbox": [ + 434, + 197, + 553, + 285 + ], + "lines": [ + { + "bbox": [ + 434, + 197, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 434, + 197, + 553, + 285 + ], + "type": "image", + "image_path": "5c1752c06b242dd1a2f062f39b7cb30a65b456fc79c0fdfc7528387eaf83a4ed.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 317, + 286, + 434, + 375 + ], + "blocks": [ + { + "bbox": [ + 317, + 286, + 434, + 375 + ], + "lines": [ + { + "bbox": [ + 317, + 286, + 434, + 375 + ], + "spans": [ + { + "bbox": [ + 317, + 286, + 434, + 375 + ], + "type": "image", + "image_path": "fee4ab809c7cb660769b3b3c811e49a4e109eb29d2763c96cbc2e1b9a54c9e29.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 378, + 555, + 422 + ], + "lines": [ + { + "bbox": [ + 313, + 378, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 555, + 422 + ], + "type": "text", + "content": "Figure 1. 3D Scene Decomposition with Superquadrics. Given a 3D point cloud of an indoor scene, SUPERDEC decomposes all scene objects into a compact set of superquadric primitives. Different object instances are visualized with different colors." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 434, + 286, + 553, + 375 + ], + "blocks": [ + { + "bbox": [ + 434, + 286, + 553, + 375 + ], + "lines": [ + { + "bbox": [ + 434, + 286, + 553, + 375 + ], + "spans": [ + { + "bbox": [ + 434, + 286, + 553, + 375 + ], + "type": "image", + "image_path": "6ec755e3bb8d3a158ef44d63135f557a4062c39c52b501b2163819ec69e8be30.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 445, + 555, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 493 + ], + "type": "text", + "content": "these approaches excel at achieving photorealism, they lack explicit control over compactness, often resulting in large, non-modular scene encodings that are unsuitable for tasks requiring explicit spatial reasoning." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 495, + 556, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 556, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 556, + 663 + ], + "type": "text", + "content": "While optimizing compactness on a scene level remains a challenging task, prior work has shown that geometric primitives such as cuboids [51, 57] or superquadrics [1, 31, 32] enable compact and interpretable decompositions of individual objects. These methods are generally either learning-based [32, 57], favoring speed at the cost of geometric accuracy, or optimization-based [1, 23, 31], offering improved accuracy at the expense of higher computational cost. Although both types of approaches can be effective for certain object classes, they often struggle to generalize across datasets with diverse object geometries. Learning-based methods typically require category-specific training, while optimization-based ones rely on hand-crafted heuristics, limiting their scalability in open-world settings." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Motivated by the abstraction capabilities of geometric primitives for individual object categories, we propose to represent complex 3D scenes using a compact set of superquadrics. To this end, we learn class-agnostic object" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24625" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "level shape priors to optimize compactness and leverage an off-the-shelf 3D instance segmentation method (Mask3D [41]) to scale our approach to full 3D scenes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 109, + 295, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 109, + 295, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 109, + 295, + 370 + ], + "type": "text", + "content": "We choose superquadrics as building blocks for our representation, as they offer more accurate shape modeling than cuboids while incurring minimal additional parameter overhead (9 v.s. 11, including 6-DoF pose parameters). To develop a model capable of generalizing across diverse object types, we draw inspiration from supervised segmentation [5, 6, 41], and approach the problem from the perspective of unsupervised geometric-based segmentation, using local point-based features to iteratively refine predicted geometric primitives. Our model is trained on ShapeNet [4] and evaluated on three challenging and diverse 3D datasets: ShapeNet [4], ScanNet++ [58], and Replica [45]. On the 3D object dataset ShapeNet, our approach achieves an L2 error six times smaller than prior state-of-the-art methods [32], while requiring only half the number of primitives. On ScanNet++ and Replica, we demonstrate that our method generalizes well to real-world scene-level settings, despite being trained solely on ShapeNet. Finally, we show the practical utility of our method as a scene representation for robotic tasks, including path planning and object grasping, as well as an editable 3D scene representation for controllable image generation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 371, + 253, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 253, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 253, + 383 + ], + "type": "text", + "content": "In summary, our contributions are the following:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 383, + 295, + 466 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 383, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 383, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 383, + 295, + 407 + ], + "type": "text", + "content": "1) We introduce SUPERDEC, a novel method for decomposing 3D scenes using superquadric primitives." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 407, + 294, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 294, + 430 + ], + "type": "text", + "content": "2) SUPERDEC achieves state-of-the-art object decomposition scores on ShapeNet trained jointly on multiple classes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 431, + 294, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 294, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 294, + 466 + ], + "type": "text", + "content": "3) We demonstrate the effectiveness of 3D superquadric scene representations for robotic tasks and controllable generative content creation." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 478, + 142, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 142, + 490 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 142, + 490 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 493, + 295, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 295, + 673 + ], + "type": "text", + "content": "Learning-based methods have shown that neural networks, when equipped with suitable reconstruction losses, can directly predict geometric primitive parameters to decompose point clouds into a minimal set of primitives for specific object categories. Tulsiani [51] introduced a CNN-based method for cuboid decomposition, which was later extended to more expressive primitives such as superquadrics by Paschalidou et al. [32]. CSA [57] further enhanced interpretability by employing a stronger point encoder and jointly predicting cuboid parameters and part segmentations. However, these methods remain constrained by their reliance on category-specific training. We attribute this limitation to their model design, which encodes only global shape features, sufficient for intra-category generalization but ineffective for decomposing out-of-category objects." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Optimization-based methods largely originate from the literature on superquadric fitting. EMS [23] revisited this line of work by introducing a probabilistic formulation that" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "text", + "content": "enables the decomposition of arbitrary objects into multiple superquadrics. Given an input point cloud, the method first fits a superquadric to the main structure and identifies unfitted outlier clusters, which are then recursively processed in a hierarchical fashion up to a predefined depth level. However, as noted in their paper and confirmed by our experiments, this approach implicitly assumes that objects exhibit a hierarchical geometric structure, limiting its applicability to many real-world objects such as tables and chairs. Other methods, such as Marching Primitives [24], require Signed Distance Functions (SDFs) as input, which are generally not easily available in real-world scenes. More fundamentally, since these approaches optimize from scratch for each object, they cannot leverage generalizable point features or learned shape priors, both of which are critical for abstraction and robustness under partial observations, a common challenge in practical 3D capture scenarios." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 281, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 510 + ], + "type": "text", + "content": "Scene-level decomposition. With the emergence of 3DGS [18], an increasing number of works have explored representing 3D scenes using various geometric primitives as generalized ellipsoids [15] and convexes [14]. While heuristics can control the number of Gaussians, achieving truly compact representations remains challenging. DBW [31] addresses this by fitting a small set of textured superquadrics to 3D scenes, building on the principles of 3DGS. Given a set of scene images, it performs test-time optimization with a photometric loss and renders the primitives using a differentiable rasterizer [22]. To model the environment, DBW adds a meshed ground plane and a meshed icosphere for the background. However, it is restricted to scenes with fewer than 10 primitives and requires objects to be aligned to a ground plane. Furthermore, the optimization is computationally expensive, taking around three hours even on simple DTU [17] scenes. Our method differs significantly in terms of input requirements, generality of application, and computational efficiency." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 515, + 556, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 556, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 556, + 635 + ], + "type": "text", + "content": "Superquadrics are a parametric family of shapes introduced by Barr et al. [2] in 1981 and have since been widely adopted in both computer vision and graphics [7, 34, 44]. Their popularity stems from their ability to represent a diverse range of shapes with a highly compact parameterization. A superquadric in its canonical pose is defined by just five parameters: " + }, + { + "bbox": [ + 313, + 515, + 556, + 635 + ], + "type": "inline_equation", + "content": "(s_x, s_y, s_z)" + }, + { + "bbox": [ + 313, + 515, + 556, + 635 + ], + "type": "text", + "content": " for the scales along the three principal semi-axes and " + }, + { + "bbox": [ + 313, + 515, + 556, + 635 + ], + "type": "inline_equation", + "content": "(\\epsilon_1, \\epsilon_2)" + }, + { + "bbox": [ + 313, + 515, + 556, + 635 + ], + "type": "text", + "content": " for the shape-defining exponents. Given those parameters, their surface is described by the implicit equation:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 318, + 644, + 553, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 644, + 553, + 681 + ], + "spans": [ + { + "bbox": [ + 318, + 644, + 553, + 681 + ], + "type": "interline_equation", + "content": "f (\\mathbf {x}) = \\left(\\left(\\frac {x}{s _ {x}}\\right) ^ {\\frac {2}{\\epsilon_ {2}}} + \\left(\\frac {y}{s _ {y}}\\right) ^ {\\frac {2}{\\epsilon_ {2}}}\\right) ^ {\\frac {\\epsilon_ {2}}{\\epsilon_ {1}}} + \\left(\\frac {z}{s _ {z}}\\right) ^ {\\frac {2}{\\epsilon_ {1}}} = 1. \\tag {1}", + "image_path": "aecb7ec34626ddfa81062dcd7ef4475b01b9130c925442c8fad43d5662c929c6.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Extending this representation to a global coordinate system requires 6 additional parameters (3 for translation and 3" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24626" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": "for rotation), resulting in a total of 11 parameters per superquadric. Another key property of superquadrics is the ability to compute the radial distance from any point in 3D space to the superquadric surface, i.e., the distance between a point and the superquadric's surface along the line connecting that point to the center of the superquadric. Specifically, given a point " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": ", its radial distance to the surface of a canonically oriented superquadric is defined as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 117, + 178, + 295, + 192 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 178, + 295, + 192 + ], + "spans": [ + { + "bbox": [ + 117, + 178, + 295, + 192 + ], + "type": "interline_equation", + "content": "d _ {r} = | \\mathbf {x} | \\cdot | 1 - f (\\mathbf {x}) ^ {- \\epsilon_ {1} / 2} |, \\tag {2}", + "image_path": "a1b83290a0a405b4730d68bd15176b5d20f3a8e62ee5b3dc8827e70aa352d491.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 202, + 296, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 202, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 296, + 239 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 202, + 296, + 239 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 202, + 296, + 239 + ], + "type": "text", + "content": " is given in Eq. 1. We refer the reader to [53] for the derivation of Eq. 2 and to [16] for a more comprehensive overview on superquadrics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 251, + 111, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 111, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 111, + 262 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 271, + 296, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 296, + 342 + ], + "type": "text", + "content": "Our ultimate goal is a 3D scene decomposition using superquadric primitives. To this end, we first focus on single-object decomposition and then show how our method, combined with 3D instance segmentation [41], can be applied to full 3D scenes. We detail the single-object approach in Sec. 3.1 and its extension to full scenes in Sec. 3.2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 352, + 213, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 213, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 213, + 365 + ], + "type": "text", + "content": "3.1. Single Object Decomposition" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 370, + 296, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 296, + 442 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 296, + 442 + ], + "type": "text", + "content": "Fig. 2 illustrates our model for single-object decomposition. It consists of two main components: a self-supervised feed-forward neural network that jointly predicts superquadric parameters and a segmentation matrix associating points to superquadrics, followed by a lightweight Levenberg-Marquardt (LM) optimization [20, 28]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 450, + 213, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 213, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 213, + 460 + ], + "type": "text", + "content": "3.1.1. Feed-forward Neural Network" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 465, + 296, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 465, + 296, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 465, + 296, + 562 + ], + "type": "text", + "content": "Our deep learning model draws inspiration from recent fully-supervised Transformer-based [54] segmentation models [5, 6, 41]. These models iteratively decode a sequence of queries, each representing a segmentation mask, by cross-attending to input pixels or points. In our case, the queries represent superquadrics. Next, we show how such an architecture can be adapted to unsupervisedly segment superquadrics, instead of supervisedly segment objects." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "text", + "content": "Model Details. Given an input point cloud " + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P} \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "text", + "content": ", where each of the " + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "text", + "content": " points has a 3D coordinate, we first extract rich point features " + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{PC} \\in \\mathbb{R}^{N \\times H}" + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "text", + "content": " using the PVCNN [25] point encoder. At the same time, we initialize " + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "text", + "content": " superquadrics features " + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{SQ} \\in \\mathbb{R}^{P \\times H}" + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "text", + "content": " with sinusoidal positional encodings. We feed these features in a Transformer decoder [54] which leverages self-attention, cross-attention to the point features, and feed-forward layers to refine them. Once refined, the superquadric features " + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{SQ}" + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "text", + "content": " and the point features " + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{PC}" + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "text", + "content": " are fed into two prediction heads: The segmentation head takes as input " + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{SQ}" + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{PC}" + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "text", + "content": " and predicts a soft assignment matrix " + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{R}^{N \\times P}" + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "text", + "content": " associating points to" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 72, + 515, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 515, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 515, + 84 + ], + "type": "text", + "content": "superquadrics and whose elements are defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 377, + 90, + 553, + 104 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 90, + 553, + 104 + ], + "spans": [ + { + "bbox": [ + 377, + 90, + 553, + 104 + ], + "type": "interline_equation", + "content": "m _ {i j} = \\sigma (\\phi (\\mathcal {F} _ {P C}) \\cdot \\mathcal {F} _ {S Q}), \\tag {3}", + "image_path": "6760394c53991695274c1ae8ac8993c248e305566567d96d0e7a6808c4b73e60.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 108, + 555, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 555, + 205 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 108, + 555, + 205 + ], + "type": "inline_equation", + "content": "\\phi (\\mathcal{F}_{PC})\\in \\mathbb{R}^{N\\times H}" + }, + { + "bbox": [ + 313, + 108, + 555, + 205 + ], + "type": "text", + "content": " is a learned projection of the point features to match the dimensionality of the superquadric features, and " + }, + { + "bbox": [ + 313, + 108, + 555, + 205 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 108, + 555, + 205 + ], + "type": "text", + "content": " is the softmax function. The second head, the superquadric head, takes the superquadric features " + }, + { + "bbox": [ + 313, + 108, + 555, + 205 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{SQ}" + }, + { + "bbox": [ + 313, + 108, + 555, + 205 + ], + "type": "text", + "content": " as input and predicts 12 parameters for each superquadric: 11 encoding its 5-DoF shape and 6-DoF pose, and one modeling its existence probability " + }, + { + "bbox": [ + 313, + 108, + 555, + 205 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 108, + 555, + 205 + ], + "type": "text", + "content": ", enabling a variable number of superquadrics per object." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 209, + 554, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 209, + 554, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 554, + 243 + ], + "type": "text", + "content": "Losses. We train our model in a self-supervised manner, without requiring any ground truth annotation. Specifically, the total loss is defined as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 353, + 251, + 553, + 264 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 251, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 353, + 251, + 553, + 264 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {r e c}} + \\lambda_ {\\text {p a r}} \\mathcal {L} _ {\\text {p a r}} + \\lambda_ {\\text {e x i s t}} \\mathcal {L} _ {\\text {e x i s t}}, \\tag {4}", + "image_path": "2fa774b9b6b2e43716e6bd1544d15ff81c65fb0ca14396fc7faa120f1f1f53f8.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 270, + 554, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 554, + 329 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 554, + 329 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 270, + 554, + 329 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 313, + 270, + 554, + 329 + ], + "type": "text", + "content": " is the reconstruction loss aligning the predicted superquadrics to the input point cloud " + }, + { + "bbox": [ + 313, + 270, + 554, + 329 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 270, + 554, + 329 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 270, + 554, + 329 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{par}" + }, + { + "bbox": [ + 313, + 270, + 554, + 329 + ], + "type": "text", + "content": " is the parsimony loss encouraging a small number of primitives, " + }, + { + "bbox": [ + 313, + 270, + 554, + 329 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{exist}" + }, + { + "bbox": [ + 313, + 270, + 554, + 329 + ], + "type": "text", + "content": " is the existence loss, and " + }, + { + "bbox": [ + 313, + 270, + 554, + 329 + ], + "type": "inline_equation", + "content": "\\lambda_{par}, \\lambda_{exist}" + }, + { + "bbox": [ + 313, + 270, + 554, + 329 + ], + "type": "text", + "content": " are weighting coefficients. The reconstruction loss " + }, + { + "bbox": [ + 313, + 270, + 554, + 329 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 313, + 270, + 554, + 329 + ], + "type": "text", + "content": " consists of three terms:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 361, + 335, + 553, + 349 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 335, + 553, + 349 + ], + "spans": [ + { + "bbox": [ + 361, + 335, + 553, + 349 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e c} = \\mathcal {L} _ {\\mathcal {P} \\rightarrow S Q} + \\mathcal {L} _ {S Q \\rightarrow \\mathcal {P}} + \\mathcal {L} _ {N}. \\tag {5}", + "image_path": "97ae91f39d55bc018aea6a606f5f8f4834886bb16fa0c74bffe6f540513382f1.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 354, + 555, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 475 + ], + "type": "text", + "content": "The first two terms correspond to the bi-directional Chamfer distance between the input point cloud and the superquadric surfaces, while the third term serves as a regularizer incorporating normal information to improve convergence during training. To compute the Chamfer distance, we approximate each superquadric surface by uniformly sampling " + }, + { + "bbox": [ + 313, + 354, + 555, + 475 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 354, + 555, + 475 + ], + "type": "text", + "content": " points, following the method of Pilu et al. [35]. Denoting by " + }, + { + "bbox": [ + 313, + 354, + 555, + 475 + ], + "type": "inline_equation", + "content": "d(\\mathbf{x}_i,\\mathbf{x}_{js}^{\\prime})" + }, + { + "bbox": [ + 313, + 354, + 555, + 475 + ], + "type": "text", + "content": " the euclidean distance between the " + }, + { + "bbox": [ + 313, + 354, + 555, + 475 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 354, + 555, + 475 + ], + "type": "text", + "content": "-th point in the input point cloud and the " + }, + { + "bbox": [ + 313, + 354, + 555, + 475 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 354, + 555, + 475 + ], + "type": "text", + "content": "-th point sampled on the surface of the " + }, + { + "bbox": [ + 313, + 354, + 555, + 475 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 354, + 555, + 475 + ], + "type": "text", + "content": "-th superquadric, we define " + }, + { + "bbox": [ + 313, + 354, + 555, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{P}\\to SQ}" + }, + { + "bbox": [ + 313, + 354, + 555, + 475 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 344, + 479, + 553, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 479, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 344, + 479, + 553, + 514 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {P} \\rightarrow S Q} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {P} m _ {i j} \\min _ {s \\in [ S ]} d \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {j s} ^ {\\prime}\\right), \\tag {6}", + "image_path": "e8a0b634abc6e5bac10975fd89dd3fd87a71b0f615a93a74af25401956625122.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 518, + 382, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 518, + 382, + 531 + ], + "spans": [ + { + "bbox": [ + 314, + 518, + 382, + 531 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 314, + 518, + 382, + 531 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SQ\\to \\mathcal{P}}" + }, + { + "bbox": [ + 314, + 518, + 382, + 531 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 322, + 536, + 553, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 536, + 553, + 570 + ], + "spans": [ + { + "bbox": [ + 322, + 536, + 553, + 570 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S Q \\rightarrow \\mathcal {P}} = \\frac {1}{S \\sum_ {j = 1} ^ {P} \\alpha_ {j}} \\sum_ {j = 1} ^ {P} \\alpha_ {j} \\sum_ {s = 1} ^ {S} \\min _ {i \\in [ N ]} d \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {j s} ^ {\\prime}\\right). \\tag {7}", + "image_path": "1b74ee42ce85ee65d1a1edc83e914841d3830b2239b8890baa92ad1538277c0b.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 574, + 554, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 554, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 554, + 671 + ], + "type": "text", + "content": "The last term of Eq. 5, i.e., " + }, + { + "bbox": [ + 313, + 574, + 554, + 671 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_N" + }, + { + "bbox": [ + 313, + 574, + 554, + 671 + ], + "type": "text", + "content": " is defined as the reconstruction loss from Yang et al. [57], and is used to incorporate normal information during training which leads to accelerated convergence. Additionally, since we seek not only accuracy but also compactness, we introduce a parsimony loss to encourage the use of fewer primitives. To do that, we optimize the 0.5-norm of " + }, + { + "bbox": [ + 313, + 574, + 554, + 671 + ], + "type": "inline_equation", + "content": "m_j \\coloneqq \\sum_{i=1}^{N} \\frac{m_{ij}}{N}" + }, + { + "bbox": [ + 313, + 574, + 554, + 671 + ], + "type": "text", + "content": " and define the parsimony loss as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 375, + 675, + 553, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 675, + 553, + 716 + ], + "spans": [ + { + "bbox": [ + 375, + 675, + 553, + 716 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p a r} = \\left(\\frac {1}{P} \\sum_ {j = 1} ^ {P} \\frac {\\sqrt {m _ {j}}}{P}\\right) ^ {2}. \\tag {8}", + "image_path": "f9b8e1bfd9c534eea9afcedcc2f3f6f9d803cecdff4c9563006d932f3360f744.jpg" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24627" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 68, + 553, + 220 + ], + "blocks": [ + { + "bbox": [ + 61, + 68, + 553, + 220 + ], + "lines": [ + { + "bbox": [ + 61, + 68, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 61, + 68, + 553, + 220 + ], + "type": "image", + "image_path": "189e12d07aae7438d31a40affef30c862ea9e9b29fb2f7506ab4e5590c043883.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 223, + 555, + 269 + ], + "lines": [ + { + "bbox": [ + 54, + 223, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 54, + 223, + 555, + 269 + ], + "type": "text", + "content": "Figure 2. Illustration of the SUPERDEC Model. Given a point cloud of an object with " + }, + { + "bbox": [ + 54, + 223, + 555, + 269 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 223, + 555, + 269 + ], + "type": "text", + "content": " points, a Transformer-based neural network predicts parameters for " + }, + { + "bbox": [ + 54, + 223, + 555, + 269 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 54, + 223, + 555, + 269 + ], + "type": "text", + "content": " superquadrics, as well as a soft segmentation matrix that assigns points to superquadrics. The predicted parameters include the 11 superquadric parameters and an objectness score. These predictions provide an effective initialization for the subsequent Levenberg-Marquardt (LM) optimization, which refines the superquadrics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 285, + 296, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 285, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 54, + 285, + 296, + 360 + ], + "type": "text", + "content": "Lastly, we employ an existence loss " + }, + { + "bbox": [ + 54, + 285, + 296, + 360 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {exist }}" + }, + { + "bbox": [ + 54, + 285, + 296, + 360 + ], + "type": "text", + "content": " which uses the predicted segmentation as a teacher for the linear head in charge of predicting the existence probability. More specifically, given a threshold " + }, + { + "bbox": [ + 54, + 285, + 296, + 360 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\text {exist }}" + }, + { + "bbox": [ + 54, + 285, + 296, + 360 + ], + "type": "text", + "content": ", we define the ground-truth existence of the " + }, + { + "bbox": [ + 54, + 285, + 296, + 360 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 54, + 285, + 296, + 360 + ], + "type": "text", + "content": "th superquadric as " + }, + { + "bbox": [ + 54, + 285, + 296, + 360 + ], + "type": "inline_equation", + "content": "\\hat{\\alpha}_j \\coloneqq m_j > \\epsilon_{\\text {exist }}" + }, + { + "bbox": [ + 54, + 285, + 296, + 360 + ], + "type": "text", + "content": " and define " + }, + { + "bbox": [ + 54, + 285, + 296, + 360 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {exist }}" + }, + { + "bbox": [ + 54, + 285, + 296, + 360 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 368, + 295, + 402 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 368, + 295, + 402 + ], + "spans": [ + { + "bbox": [ + 113, + 368, + 295, + 402 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {e x i s t}} = \\sum_ {j = 1} ^ {P} \\frac {B C E \\left(\\alpha_ {j} , \\hat {\\alpha} _ {j}\\right)}{P}, \\tag {9}", + "image_path": "ea2ef5d4772af7045db68a06a93f2fc9e19e751b7942901dccaa02e8e803ff86.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 411, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 296, + 437 + ], + "type": "text", + "content": "where BCE is the binary cross entropy and " + }, + { + "bbox": [ + 55, + 411, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\alpha_{j}" + }, + { + "bbox": [ + 55, + 411, + 296, + 437 + ], + "type": "text", + "content": " is the predicted existence probability for the " + }, + { + "bbox": [ + 55, + 411, + 296, + 437 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 411, + 296, + 437 + ], + "type": "text", + "content": "th superquadric." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 444, + 141, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 141, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 141, + 456 + ], + "type": "text", + "content": "3.1.2. Optimization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 460, + 296, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 556 + ], + "type": "text", + "content": "Our optimization module takes as input the predicted soft segmentation matrix " + }, + { + "bbox": [ + 55, + 460, + 296, + 556 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 460, + 296, + 556 + ], + "type": "text", + "content": " as well as the superquadric parameters " + }, + { + "bbox": [ + 55, + 460, + 296, + 556 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 55, + 460, + 296, + 556 + ], + "type": "text", + "content": ", and further refines the superquadric parameters using the Levenberg-Marquardt (LM) [20, 28] algorithm. Specifically, given a point cloud of " + }, + { + "bbox": [ + 55, + 460, + 296, + 556 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 460, + 296, + 556 + ], + "type": "text", + "content": " points, it iteratively refines the parameters " + }, + { + "bbox": [ + 55, + 460, + 296, + 556 + ], + "type": "inline_equation", + "content": "\\Theta_{j}" + }, + { + "bbox": [ + 55, + 460, + 296, + 556 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 55, + 460, + 296, + 556 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 460, + 296, + 556 + ], + "type": "text", + "content": "th superquadric, by computing two sets of residuals: The first set of residuals " + }, + { + "bbox": [ + 55, + 460, + 296, + 556 + ], + "type": "inline_equation", + "content": "r_{ij}" + }, + { + "bbox": [ + 55, + 460, + 296, + 556 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 460, + 296, + 556 + ], + "type": "inline_equation", + "content": "i \\in [1, N]" + }, + { + "bbox": [ + 55, + 460, + 296, + 556 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 460, + 296, + 556 + ], + "type": "inline_equation", + "content": "j \\in [1, P]" + }, + { + "bbox": [ + 55, + 460, + 296, + 556 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 566, + 295, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 566, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 137, + 566, + 295, + 581 + ], + "type": "interline_equation", + "content": "r _ {i j} = m _ {i j} \\tilde {d} _ {j} \\left(\\mathbf {x} _ {i}\\right), \\tag {10}", + "image_path": "de257fc5731b858c0937d85b40d5448f54e2751be3c916eab9091714536e3008.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 591, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 296, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 591, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\tilde{d}_j(\\mathbf{x}_i)" + }, + { + "bbox": [ + 55, + 591, + 296, + 677 + ], + "type": "text", + "content": " denotes the radial distance of point " + }, + { + "bbox": [ + 55, + 591, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 55, + 591, + 296, + 677 + ], + "type": "text", + "content": " from the " + }, + { + "bbox": [ + 55, + 591, + 296, + 677 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 591, + 296, + 677 + ], + "type": "text", + "content": "th superquadric, computed according to Eq. 2. The second set of residuals is used for normalization and is obtained by sampling a set of " + }, + { + "bbox": [ + 55, + 591, + 296, + 677 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 591, + 296, + 677 + ], + "type": "text", + "content": " points " + }, + { + "bbox": [ + 55, + 591, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{p_1},\\ldots ,\\mathbf{p_K}" + }, + { + "bbox": [ + 55, + 591, + 296, + 677 + ], + "type": "text", + "content": " on the surface of the given superquadric and then computing the distance of each of them from the point cloud. Specifically, for " + }, + { + "bbox": [ + 55, + 591, + 296, + 677 + ], + "type": "inline_equation", + "content": "i\\in [N + 1,N + K]" + }, + { + "bbox": [ + 55, + 591, + 296, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 591, + 296, + 677 + ], + "type": "inline_equation", + "content": "j\\in [1,P]" + }, + { + "bbox": [ + 55, + 591, + 296, + 677 + ], + "type": "text", + "content": " we compute " + }, + { + "bbox": [ + 55, + 591, + 296, + 677 + ], + "type": "inline_equation", + "content": "r_{ij}" + }, + { + "bbox": [ + 55, + 591, + 296, + 677 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 687, + 296, + 704 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 687, + 296, + 704 + ], + "spans": [ + { + "bbox": [ + 62, + 687, + 296, + 704 + ], + "type": "interline_equation", + "content": "r _ {i j} = \\min _ {k} \\left\\| \\mathbf {p} _ {\\mathbf {i} - \\mathbf {N}} - \\Pi_ {j} \\left(\\mathbf {x} _ {\\mathbf {k}}\\right) \\right\\| _ {2}, \\quad \\text {w i t h} k \\in [ 1, N ]. \\tag {11}", + "image_path": "dfd086c64a49b2c548bf23cc5533755461e84c954735fd4b33c84148f8064672.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 285, + 491, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 491, + 298 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 491, + 298 + ], + "type": "text", + "content": "3.2. Decomposition of Full 3D Scenes" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 304, + 555, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 555, + 413 + ], + "type": "text", + "content": "After training on single objects, extending SUPERDEC to full 3D scenes is straightforward. Given a scene-level point cloud, we extract 3D object instance masks using Mask3D [41]. Each object is centered and uniformly rescaled to a sphere of radius 0.5. We then predict the superquadric primitives for each object individually using our model. We found our model trained on ShapeNet [4] to generalize well on real-world 3D scenes from ScanNet++ [58] and Replica [45] without additional fine-tuning." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 425, + 395, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 395, + 438 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 395, + 438 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 445, + 555, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 542 + ], + "type": "text", + "content": "We first compare our SUPERDEC with previous state-of-the-art methods on individual objects and full 3D scenes (Sec. 4.1). We then demonstrate the usefulness of our representation on down-stream applications for robotics and controllable image generation (Sec. 4.2). Finally, in Sec. 4.3, we present additional analyses on part segmentation and the implicit learning of shape categories, followed by a study of the compactness-accuracy trade-off and runtime." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 552, + 533, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 533, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 533, + 564 + ], + "type": "text", + "content": "4.1. Comparing with State-of-the-art Methods" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 570, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 653 + ], + "type": "text", + "content": "Datasets. We compare on three different datasets: ShapeNet [4]: We use the 13 classes of the ShapeNet subset and train-val-test splits as defined in Choy et al. [8]. All objects are pre-aligned in a canonical orientation. For each object we sample 4096 points using Farthest Point Sampling (FPS) [36]. ShapeNet is a widely used dataset and is well-suited for comparison with existing baselines." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": "ScanNet++ [58]: We further evaluate our model on real-world object scans from the ScanNet++ validation set. Each object is extracted using ground truth mask annotations, and 4096 points per object are sampled with FPS. In contrast to ShapeNet, these object point clouds are noisier, partially ob" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24628" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 553, + 149 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 553, + 149 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 553, + 149 + ], + "type": "table", + "html": "
ModelPrimitive TypeSegmentationIn-categoryOut-of-category
L1 ↓L2 ↓# Prim.↓L1 ↓L2 ↓# Prim.↓
EMS (Liu et al.) [23]SuperquadricsX5.7711.3455.685.4101.2115.68
CSA (Yang et al.) [57]Cuboids✓5.1570.5279.214.8970.42711.75
SQ (Paschalidou et al.) [32]SuperquadricsX3.6680.279104.1930.3549
SUPERDEC (Ours)Superquadrics✓1.6980.0475.81.8470.0615.26
", + "image_path": "478549b1dd75314af136d61fa0c1a14474e84396a4e494dd4168e8eba492c3d6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 151, + 552, + 163 + ], + "lines": [ + { + "bbox": [ + 56, + 151, + 552, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 552, + 163 + ], + "type": "text", + "content": "Table 1. Quantitative Results on ShapeNet [4]. We show scores for in-category and out-of-category experiments and are scaled by " + }, + { + "bbox": [ + 56, + 151, + 552, + 163 + ], + "type": "inline_equation", + "content": "10^{2}" + }, + { + "bbox": [ + 56, + 151, + 552, + 163 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 57, + 167, + 549, + 459 + ], + "blocks": [ + { + "bbox": [ + 57, + 167, + 549, + 459 + ], + "lines": [ + { + "bbox": [ + 57, + 167, + 549, + 459 + ], + "spans": [ + { + "bbox": [ + 57, + 167, + 549, + 459 + ], + "type": "image", + "image_path": "4dec65a8552e92cd98987fe6a1aa90b6ed09ba9c3a8f3067ec3907771b35d527.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 462, + 555, + 485 + ], + "lines": [ + { + "bbox": [ + 55, + 462, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 555, + 485 + ], + "type": "text", + "content": "Figure 3. Qualitative Results on ShapeNet [4]. We show results on test samples for in-category (four first columns) classes and out-of-category classes (two last columns). The latter were not seen during training and illustrate how well models generalize to novel classes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 503, + 295, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 503, + 295, + 538 + ], + "spans": [ + { + "bbox": [ + 54, + 503, + 295, + 538 + ], + "type": "text", + "content": "served, and subject to random orientation and translation, providing a more realistic and challenging evaluation setting for our method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 539, + 295, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 539, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 54, + 539, + 295, + 622 + ], + "type": "text", + "content": "Replica [45]: Finally, we present both quantitative and qualitative results on full 3D scenes from Replica. Object instances are extracted using either ground truth instance annotations or the pre-trained 3D instance segmentation model Mask3D [41], enabling us to evaluate our approach in a fully realistic setting, including scenarios where no ground truth annotations are available." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 629, + 295, + 713 + ], + "type": "text", + "content": "Methods in comparison. We compare to learning- and optimization-based prior works using both cuboids and superquadrics as geometric primitives. " + }, + { + "bbox": [ + 54, + 629, + 295, + 713 + ], + "type": "inline_equation", + "content": "SQ" + }, + { + "bbox": [ + 54, + 629, + 295, + 713 + ], + "type": "text", + "content": " [32] is a learning-based approach for object-level decomposition using superquadrics; it takes a voxel grid as input and predicts superquadric primitives via a CNN. CSA [57] is another learning-based method but uses cuboids as geometric prim" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 503, + 553, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 553, + 552 + ], + "type": "text", + "content": "ities. It takes a point cloud as input and predicts cuboid parameters from a global latent code. Lastly, EMS [23] is an optimization-based approach that decomposes objects by hierarchically fitting superquadrics to parts of a point cloud." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "content": "Training Details. Our goal is to develop a general-purpose, class-agnostic model capable of representing arbitrary objects as superquadrics. Existing methods typically train separate models for each object class, assuming that all classes are known in advance and that sufficient training data is available for each. These assumptions, however, often fail in real-world scenarios. To address this, we jointly train a single model on all 13 ShapeNet classes using the publicly available code of prior methods, moving towards a more realistic class-agnostic solution. In our model we set the following hyper-parameters " + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "inline_equation", + "content": "P = 16" + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "inline_equation", + "content": "S = 4096" + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "inline_equation", + "content": "K = 25" + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "inline_equation", + "content": "H = 128" + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "inline_equation", + "content": "D = 3" + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\text{exist}} = 24" + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{\\text{exist}} = 0.01" + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{\\text{par}} = 0.06" + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "content": ". We refer to the supplementary for additional details." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24629" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 293, + 133 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 293, + 133 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 293, + 133 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 293, + 133 + ], + "type": "table", + "html": "
ScanNet++ [58]Replica [45]
MethodL1 ↓L2 ↓#Prim.↓L1 ↓L2 ↓#Prim.↓
SQ [32]10.454.2610.011.845.9110
EMS [23]5.512.114.255.402.123.61
CSA [57]2.910.4111.643.680.709.63
SUPERDEC (Ours)1.700.115.181.790.196.58
", + "image_path": "8dc5d127fb1af53e239f115a86ab8ed9cd381d24c83f9f7c72f1eb67c56228c8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 182, + 296, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 296, + 219 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 296, + 219 + ], + "type": "text", + "content": "Metrics. We assess reconstruction accuracy using L1 and L2 Chamfer distances and compactness by the average number of geometric primitives." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 233, + 173, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 233, + 173, + 245 + ], + "spans": [ + { + "bbox": [ + 55, + 233, + 173, + 245 + ], + "type": "text", + "content": "4.1.1. Results on ShapeNet" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 251, + 296, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 502 + ], + "type": "text", + "content": "We show scores in Tab. 1 and qualitative results in Fig. 3. To evaluate both accuracy and generalization, we conduct two experiments: in-category and out-of-category. In the in-category setting, all learning-based methods are jointly trained on the 13 classes of the ShapeNet training set and evaluated on the corresponding test set. In the out-of-category setting, models are trained on half of the categories (airplane, bench, chair, lamp, rifle, table) and tested on the remaining ones (car, sofa, loudspeaker, cabinet, display, telephone, watercraft). Our SUPERDEC model significantly outperforms both learned and non-learned baselines. Compared to learned baselines, we reduce the L2 loss by a factor of six while using nearly half the number of primitives, supporting our hypothesis that leveraging local point features as opposed to a single global descriptor improves 3D decomposition in both accuracy and compactness. Compared to the non-learned baseline, we predict a similar number of primitives but achieve an L2 loss approximately 20 times smaller, validating the benefit of learning shape priors to avoid local minima that often hinder purely optimization-based approaches." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 517, + 174, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 174, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 174, + 528 + ], + "type": "text", + "content": "4.1.2. Results on 3D Scenes" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "In this section, models are evaluated on real-world, out-of-category objects, which appear in arbitrary orientations and often exhibit incomplete point clouds due to reconstruction artifacts and occlusions. Tab. 2 shows the quantitative results. Despite never being trained on real-world objects, our method outperforms both the optimization- and the learning-based baselines by a large margin. Lastly, we qualitatively evaluate our pipeline on full 3D scenes from Replica, where our object-level model is applied on top of class-agnostic instance segmentation predictions from Mask3D [41]. As shown in Fig. 1, our method effectively reconstructs object shapes, even under noisy segmentation masks and geometries that differ substantially from those seen during training. We refer to the supplementary for additional qualitative results." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 322, + 70, + 547, + 149 + ], + "blocks": [ + { + "bbox": [ + 55, + 137, + 295, + 160 + ], + "lines": [ + { + "bbox": [ + 55, + 137, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 55, + 137, + 295, + 160 + ], + "type": "text", + "content": "Table 2. Quantitative Results on Scenes. Evaluation on objects from 3D scene datasets [45, 58]. Scores are scaled by " + }, + { + "bbox": [ + 55, + 137, + 295, + 160 + ], + "type": "inline_equation", + "content": "10^{2}" + }, + { + "bbox": [ + 55, + 137, + 295, + 160 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 70, + 547, + 149 + ], + "lines": [ + { + "bbox": [ + 322, + 70, + 547, + 149 + ], + "spans": [ + { + "bbox": [ + 322, + 70, + 547, + 149 + ], + "type": "table", + "html": "
MethodTime (ms)Suc. (%)Mem. (MB)
Occupancy0.056100.000.873
PointCloud0.06389.5719.286
Voxels0.03098.780.101
Cuboids [37]0.12061.230.024
SUPERDEC0.15091.710.042
", + "image_path": "c1c4c529791fd4d774efcc77fa5ec4c35fabedfa5e9f67f567e73929e97d8e32.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 152, + 551, + 163 + ], + "lines": [ + { + "bbox": [ + 317, + 152, + 551, + 163 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 551, + 163 + ], + "type": "text", + "content": "Table 3. Path Planning Scores. Mean on 15 ScanNet++ scenes." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 317, + 167, + 553, + 275 + ], + "blocks": [ + { + "bbox": [ + 317, + 167, + 553, + 275 + ], + "lines": [ + { + "bbox": [ + 317, + 167, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 317, + 167, + 553, + 275 + ], + "type": "image", + "image_path": "43864e64c65f7c33f8e0a90b6592b8790c4f5fc401010af062544c0af5d85ecc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 277, + 554, + 300 + ], + "lines": [ + { + "bbox": [ + 313, + 277, + 554, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 554, + 300 + ], + "type": "text", + "content": "Figure 4. Grasping Result. Visualization of computed grasp poses for a milk bottle, some flowers, a side table, and a plant." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 306, + 463, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 463, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 463, + 319 + ], + "type": "text", + "content": "4.2. Down-stream Applications" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 324, + 555, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 555, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 555, + 372 + ], + "type": "text", + "content": "Next, we show the versatility of the SUPERDEC representation for downstream applications, including robotics tasks such as path planning and object grasping (Sec. 4.2.1), and controllable image generation (Sec. 4.2.2)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 380, + 380, + 390 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 380, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 380, + 390 + ], + "type": "text", + "content": "4.2.1. Robotics" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 395, + 555, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 555, + 574 + ], + "type": "text", + "content": "Path planning seeks to compute a collision-free shortest path between a given start and end point in 3D space, enabling efficient robot navigation. Although essential for traversing large environments, it typically demands storing large-scale 3D representations. Here, we assess whether our compact representation can perform this task effectively while reducing memory requirements. We conduct experiments on 15 ScanNet++ [58] scenes, comparing SUPERDEC to common 3D representations, including dense occupancy grids, point clouds, voxel grids, and cuboids [37]. As shown in Tab. 6, SUPERDEC not only reduces memory consumption compared to traditional representations but also achieves a higher success rate than dense point clouds. Further details about experiment setup, metrics and analysis are provided in the Appendix." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "Object Grasping enables robots to grasp real-world objects by computing suitable grasping poses. Existing methods fall into two categories, each with complementary limitations. Geometry-based approaches [3, 12, 30] require precise 3D object models, which are often unavailable in real-world scenarios. Learning-based approaches [27, 47, 56] operate directly on raw sensor data but tend to be biased towards training data, which typically consists of tabletop scenes with small, convex, or low-genus objects [11, 46]. To overcome these limitations, SuperQ-GRASP [50] explored decomposing objects into explicit primitives. How" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24630" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 77, + 176, + 144 + ], + "blocks": [ + { + "bbox": [ + 86, + 67, + 147, + 76 + ], + "lines": [ + { + "bbox": [ + 86, + 67, + 147, + 76 + ], + "spans": [ + { + "bbox": [ + 86, + 67, + 147, + 76 + ], + "type": "text", + "content": "Input Point Cloud" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 77, + 176, + 144 + ], + "lines": [ + { + "bbox": [ + 57, + 77, + 176, + 144 + ], + "spans": [ + { + "bbox": [ + 57, + 77, + 176, + 144 + ], + "type": "image", + "image_path": "96ccec016ec636580169833e21ee26e396184942cbfe518eac6d358d802342e4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 176, + 77, + 294, + 144 + ], + "blocks": [ + { + "bbox": [ + 212, + 67, + 260, + 76 + ], + "lines": [ + { + "bbox": [ + 212, + 67, + 260, + 76 + ], + "spans": [ + { + "bbox": [ + 212, + 67, + 260, + 76 + ], + "type": "text", + "content": "Superquadrics" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 176, + 77, + 294, + 144 + ], + "lines": [ + { + "bbox": [ + 176, + 77, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 176, + 77, + 294, + 144 + ], + "type": "image", + "image_path": "3b73dddd69eb9d8ff84dcabf84c9d6171d5ca57bfb6c698c1b2cb323c63503e7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 57, + 145, + 294, + 244 + ], + "blocks": [ + { + "bbox": [ + 57, + 145, + 294, + 244 + ], + "lines": [ + { + "bbox": [ + 57, + 145, + 294, + 244 + ], + "spans": [ + { + "bbox": [ + 57, + 145, + 294, + 244 + ], + "type": "image", + "image_path": "872c27960f652358d2624d27a8fbb67cf6b9d439573594ac947f32f90445e908.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 247, + 295, + 314 + ], + "lines": [ + { + "bbox": [ + 55, + 247, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 295, + 314 + ], + "type": "text", + "content": "Figure 5. Real-world robot experiment. The top row shows the input scan (left) and the representation from SUPERDEC with the computed path and grasping pose (right). The bottom row illustrates the robot following the planned path. We denote the starting point of the path with a green sphere, and the target location with a red sphere. The target object (a milk bottle) is circled in red." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 326, + 296, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 296, + 494 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 296, + 494 + ], + "type": "text", + "content": "ever, its reliance on Marching Primitives [24] to obtain superquadrics from the object's Signed Distance Function (SDF) makes it unsuitable for most real-world cases where only point clouds are available. In contrast, our approach directly processes point clouds of entire scenes and, when combined with the class-agnostic segmentations from Mask3D [41], extracts superquadrics for all objects. Given the superquadric parameters, we employ a superquadric-based geometric method [55] to compute grasping poses for selected objects. Fig. 4 shows predicted grasping poses on objects from a real-world 3D scan of a room. In practice, our method eliminates the need for data-driven grasping models while remaining adaptable to diverse object shapes and producing high-quality grasping poses." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": "Real-world Experiment. Finally, we demonstrate the real-world applicability of our superquadric-based representation by deploying it on a legged robot (Boston Dynamics Spot) equipped with an arm, supporting both motion planning and object grasping in an indoor environment. We scan a scene using a 3D scanning application on an iPad, extract a dense point cloud, and run SUPERDEC on it. Given the robot's starting position and a specified target object (a milk bottle), we compute both the path and the grasping pose as described earlier (see Fig. 4), enabling the robot to approach and successfully grasp the object. Fig. 5 shows the computed representation, the planned trajectory, the grasping pose, and a frame from the real-world demonstration. This experiment suggests that integrating SUPERDEC with open-vocabulary segmentation methods such as OpenMask3D [48] could allow robots to navigate to and grasp arbitrary objects specified via natural-language prompts." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 79, + 553, + 144 + ], + "blocks": [ + { + "bbox": [ + 330, + 68, + 362, + 77 + ], + "lines": [ + { + "bbox": [ + 330, + 68, + 362, + 77 + ], + "spans": [ + { + "bbox": [ + 330, + 68, + 362, + 77 + ], + "type": "text", + "content": "Original" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 391, + 68, + 419, + 77 + ], + "lines": [ + { + "bbox": [ + 391, + 68, + 419, + 77 + ], + "spans": [ + { + "bbox": [ + 391, + 68, + 419, + 77 + ], + "type": "text", + "content": "Editing" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 447, + 68, + 481, + 76 + ], + "lines": [ + { + "bbox": [ + 447, + 68, + 481, + 76 + ], + "spans": [ + { + "bbox": [ + 447, + 68, + 481, + 76 + ], + "type": "text", + "content": "Addition" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 507, + 68, + 539, + 76 + ], + "lines": [ + { + "bbox": [ + 507, + 68, + 539, + 76 + ], + "spans": [ + { + "bbox": [ + 507, + 68, + 539, + 76 + ], + "type": "text", + "content": "Deletion" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 79, + 553, + 144 + ], + "lines": [ + { + "bbox": [ + 316, + 79, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 316, + 79, + 553, + 144 + ], + "type": "image", + "image_path": "e67dd82e63110d293547f0ef7d56641338114490b24f5f21c59e377da3a618b1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 316, + 145, + 553, + 213 + ], + "blocks": [ + { + "bbox": [ + 316, + 145, + 553, + 213 + ], + "lines": [ + { + "bbox": [ + 316, + 145, + 553, + 213 + ], + "spans": [ + { + "bbox": [ + 316, + 145, + 553, + 213 + ], + "type": "image", + "image_path": "2143ccb01f4229cb1edb7587b0fe971d16e0e17f835fcd8eec0480af3115e7cb.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 215, + 553, + 249 + ], + "lines": [ + { + "bbox": [ + 313, + 215, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 313, + 215, + 553, + 249 + ], + "type": "text", + "content": "Figure 6. Spatial control using SUPERDEC. Top row shows superquadrics generated by SUPERDEC, bottom row shows generated images using the prompt 'A corner of a room with a plant'." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 316, + 280, + 434, + 357 + ], + "blocks": [ + { + "bbox": [ + 348, + 267, + 402, + 277 + ], + "lines": [ + { + "bbox": [ + 348, + 267, + 402, + 277 + ], + "spans": [ + { + "bbox": [ + 348, + 267, + 402, + 277 + ], + "type": "text", + "content": "Superquadrics" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 280, + 434, + 357 + ], + "lines": [ + { + "bbox": [ + 316, + 280, + 434, + 357 + ], + "spans": [ + { + "bbox": [ + 316, + 280, + 434, + 357 + ], + "type": "image", + "image_path": "549a00a9f8ad1734a9a270a4598af061adff061ce951c65e7b9e10375d1407f7.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 358, + 412, + 368 + ], + "lines": [ + { + "bbox": [ + 342, + 358, + 412, + 368 + ], + "spans": [ + { + "bbox": [ + 342, + 358, + 412, + 368 + ], + "type": "text", + "content": "\"Pink living room\"" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 436, + 280, + 552, + 357 + ], + "blocks": [ + { + "bbox": [ + 466, + 267, + 518, + 277 + ], + "lines": [ + { + "bbox": [ + 466, + 267, + 518, + 277 + ], + "spans": [ + { + "bbox": [ + 466, + 267, + 518, + 277 + ], + "type": "text", + "content": "Depth Prompt" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 436, + 280, + 552, + 357 + ], + "lines": [ + { + "bbox": [ + 436, + 280, + 552, + 357 + ], + "spans": [ + { + "bbox": [ + 436, + 280, + 552, + 357 + ], + "type": "image", + "image_path": "2fa33d0f2387e0deaa6bfa7c02eebbd4b83976db59e761379758333d07597226.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 358, + 533, + 368 + ], + "lines": [ + { + "bbox": [ + 451, + 358, + 533, + 368 + ], + "spans": [ + { + "bbox": [ + 451, + 358, + 533, + 368 + ], + "type": "text", + "content": "\"Modern living room\"" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 317, + 369, + 434, + 446 + ], + "blocks": [ + { + "bbox": [ + 317, + 369, + 434, + 446 + ], + "lines": [ + { + "bbox": [ + 317, + 369, + 434, + 446 + ], + "spans": [ + { + "bbox": [ + 317, + 369, + 434, + 446 + ], + "type": "image", + "image_path": "3e59075a06027fbc2bd7003279a192ddc6c52423476fa057dd559f77108d379c.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 449, + 554, + 494 + ], + "lines": [ + { + "bbox": [ + 313, + 449, + 554, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 554, + 494 + ], + "type": "text", + "content": "Figure 7. Semantic control using SUPERDEC. Top row: superquadrics created by our SUPERDEC, and depth map to prompt the generation of text-to-image diffusion model. Bottom row: generations with two different textual prompts." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 435, + 369, + 552, + 446 + ], + "blocks": [ + { + "bbox": [ + 435, + 369, + 552, + 446 + ], + "lines": [ + { + "bbox": [ + 435, + 369, + 552, + 446 + ], + "spans": [ + { + "bbox": [ + 435, + 369, + 552, + 446 + ], + "type": "image", + "image_path": "5eb20ce3b4c02010a3963f79146e31a91393b7ab52a8e40fda7ec8b8044582cd.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 513, + 499, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 499, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 499, + 525 + ], + "type": "text", + "content": "4.2.2. Controllable Generation and Editing" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "We investigate how the SUPERDEC representation can be used to introduce joint spatial and semantic control in text-to-image diffusion models [39]. Specifically, we generate images by conditioning ControlNet [60] on depth maps rendered from the superquadrics extracted from Replica [45] scenes. Qualitative results are shown in Fig. 6 and Fig. 7. Fig. 6 demonstrates spatial control: by moving, duplicating, or removing superquadrics corresponding to a plant, we coherently influence the generated images. Fig. 7 highlights semantic control: we can vary the room's style while preserving its semantic and geometric structure, and observe that object semantics naturally emerge from the spatial arrangement of superquadrics without explicit conditioning, e.g., pillows appear on couches, and a plant is placed on a central table, reflecting plausible real-world arrangements." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24631" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 97, + 285, + 140 + ], + "blocks": [ + { + "bbox": [ + 58, + 97, + 285, + 140 + ], + "lines": [ + { + "bbox": [ + 58, + 97, + 285, + 140 + ], + "spans": [ + { + "bbox": [ + 58, + 97, + 285, + 140 + ], + "type": "image", + "image_path": "4ee67679311c87aefdbf97b48165d576960955d6dace3c71d86dd1c6f9f1a42c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 142, + 295, + 186 + ], + "lines": [ + { + "bbox": [ + 55, + 142, + 295, + 186 + ], + "spans": [ + { + "bbox": [ + 55, + 142, + 295, + 186 + ], + "type": "text", + "content": "Figure 8. Qualitative Results on ShapeNet [4] segmentation. We show the resulting segmentation matrices on test samples for in-category (four first columns) classes and out-of-category classes (two last columns). The latter were not seen during training." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 56, + 193, + 137, + 247 + ], + "blocks": [ + { + "bbox": [ + 56, + 193, + 137, + 247 + ], + "lines": [ + { + "bbox": [ + 56, + 193, + 137, + 247 + ], + "spans": [ + { + "bbox": [ + 56, + 193, + 137, + 247 + ], + "type": "image", + "image_path": "e2ed7ba8ef746fe51ffd49bb3304a34ba9e753e7338d1f8610a96291247c4f26.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 250, + 117, + 259 + ], + "lines": [ + { + "bbox": [ + 69, + 250, + 117, + 259 + ], + "spans": [ + { + "bbox": [ + 69, + 250, + 117, + 259 + ], + "type": "text", + "content": "Input 3D Scan" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 138, + 193, + 294, + 248 + ], + "blocks": [ + { + "bbox": [ + 138, + 193, + 294, + 248 + ], + "lines": [ + { + "bbox": [ + 138, + 193, + 294, + 248 + ], + "spans": [ + { + "bbox": [ + 138, + 193, + 294, + 248 + ], + "type": "image", + "image_path": "685b7d9f74bb6c0449748e930066efc7dd478d39a43152c54b9839facfbbf6e2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 250, + 191, + 259 + ], + "lines": [ + { + "bbox": [ + 148, + 250, + 191, + 259 + ], + "spans": [ + { + "bbox": [ + 148, + 250, + 191, + 259 + ], + "type": "inline_equation", + "content": "1^{\\mathrm{st}}" + }, + { + "bbox": [ + 148, + 250, + 191, + 259 + ], + "type": "text", + "content": " Hierarchy" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 228, + 249, + 273, + 259 + ], + "lines": [ + { + "bbox": [ + 228, + 249, + 273, + 259 + ], + "spans": [ + { + "bbox": [ + 228, + 249, + 273, + 259 + ], + "type": "inline_equation", + "content": "2^{\\mathrm{nd}}" + }, + { + "bbox": [ + 228, + 249, + 273, + 259 + ], + "type": "text", + "content": " Hierarchy" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 266, + 296, + 310 + ], + "lines": [ + { + "bbox": [ + 55, + 266, + 296, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 296, + 310 + ], + "type": "text", + "content": "Figure 9. Hierarchical decomposition with SUPERDEC. From left to right we visualize: the input point cloud, the superquadrics predicted in the first hierarchy, the superquadrics predicted in the second hierarchy." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 328, + 180, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 180, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 180, + 340 + ], + "type": "text", + "content": "4.3. Analysis Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 346, + 296, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 296, + 477 + ], + "type": "text", + "content": "Unsupervised part segmentation. Besides superquadric parameters, our method also predicts a segmentation matrix which segments the initial point cloud into parts that are fitted to the predicted superquadrics. In Fig. 8, we visualize the predicted segmentation masks for the same examples shown in Fig. 3. We observe that segmentation masks, especially in the in-category experiments, appear very sharp. This suggests that our method, especially if trained at a larger scale, can be leveraged for different applications as geometry-based part segmentation or as pretraining for supervised semantic part segmentation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 492, + 296, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 296, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 296, + 648 + ], + "type": "text", + "content": "Can we do hierarchical decomposition? We run SUPERDEC in a hierarchical fashion on a whole simple scene, without using any instance segmentation method to segment the scene into different objects. We observed that by applying SUPERDEC on the whole point cloud it can coarsely segment the scene into components such as grass, table, and chairs. By applying the model again on the points segmented during the first prediction, we obtained a more accurate decomposition of the individual object instances. Given that SUPERDEC was trained only on single object instances from ShapeNet [4] we found this results very interesting. We see future potential in the use of our model to obtain unsupervised hierarchical decompositions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "What does our network learn? Since our network performs unsupervised part segmentation, we analyze the features learned by the Transformer decoder across object classes. Inspired by BERT [10]'s [CLS] token, we append a learnable embedding to the sequence of embedded" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 343, + 60, + 529, + 210 + ], + "blocks": [ + { + "bbox": [ + 343, + 60, + 529, + 210 + ], + "lines": [ + { + "bbox": [ + 343, + 60, + 529, + 210 + ], + "spans": [ + { + "bbox": [ + 343, + 60, + 529, + 210 + ], + "type": "image", + "image_path": "927a6ac7e941d7c142444a279e78b7f93a0eda48241105913c330a7785d765b1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 213, + 553, + 236 + ], + "lines": [ + { + "bbox": [ + 313, + 213, + 553, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 553, + 236 + ], + "type": "text", + "content": "Figure 10. t-SNE Visualization of Primitive Embeddings across different ShapeNet classes." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 257, + 555, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 257, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 555, + 365 + ], + "type": "text", + "content": "superquadrics; although never explicitly decoded, this embedding is refined through self- and cross-attention. After training, we extract and visualize these embeddings using t-SNE [52] for ShapeNet [4] categories (Fig. 10). We observe that categories with consistent shapes, such as chairs, airplanes, and cars, form clear clusters, while categories with high intra-class variability, such as watercraft, are more dispersed. This indicates that our model organizes objects by geometric structure without requiring class annotations." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 371, + 554, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 554, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 554, + 467 + ], + "type": "text", + "content": "How fast is our method? Our model is highly parallelizable, allowing multiple objects to be batched and processed simultaneously in a single forward pass. On an RTX 4090 (24 GB), we can process up to 256 objects in parallel. On average, the forward pass takes " + }, + { + "bbox": [ + 313, + 371, + 554, + 467 + ], + "type": "inline_equation", + "content": "0.13\\mathrm{s}" + }, + { + "bbox": [ + 313, + 371, + 554, + 467 + ], + "type": "text", + "content": " for a complete Replica [45] scene, 3D instance segmentation with Mask3D [41] requires " + }, + { + "bbox": [ + 313, + 371, + 554, + 467 + ], + "type": "inline_equation", + "content": "0.3\\mathrm{s}" + }, + { + "bbox": [ + 313, + 371, + 554, + 467 + ], + "type": "text", + "content": ", and each LM optimization step takes less than 1 s, see supplementary for more details." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 475, + 388, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 388, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 388, + 487 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 491, + 555, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 555, + 622 + ], + "type": "text", + "content": "We proposed SUPERDEC, a method for deriving compact yet expressive 3D scene representations based on simple geometric primitives – specifically, superquadrics. Our model outperforms prior primitive-based methods and generalizes well to out-of-category classes. We further demonstrated the potential of the resulting 3D scene representation for various applications in robotics, and as a geometric prompt for diffusion-based image generation. While this is only a first step towards more compact, geometry-aware 3D scene representations, we anticipate broader applications and expect to see further research in this direction." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "content": "Acknowledgments. Elisabetta Fedele is supported by the ETH AI Center doctoral fellowship, by the Swiss National Science Foundation (SNSF) Advanced Grant 216260 (Beyond Frozen Worlds: Capturing Functional 3D Digital Twins from the Real World), and an SNSF Mobility Grant. Francis Engelmann is supported by an SNSF PostDoc.mobility grant. We also gratefully acknowledge a GPU grant from NVIDIA." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 111, + 83, + 156, + 92 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 83, + 156, + 92 + ], + "spans": [ + { + "bbox": [ + 111, + 83, + 156, + 92 + ], + "type": "text", + "content": "In-category" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 220, + 83, + 280, + 92 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 83, + 280, + 92 + ], + "spans": [ + { + "bbox": [ + 220, + 83, + 280, + 92 + ], + "type": "text", + "content": "Out-of-category" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24632" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Stephan Alaniz, Massimiliano Mancini, and Zeynep Akata. Iterative superquadric decomposition of 3d objects from multiple viewsd. In International Conference on Computer Vision (ICCV), 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 135, + 296, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 135, + 296, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 135, + 296, + 167 + ], + "type": "text", + "content": "[2] Alan H. Barr. Superquadrics and angle-preserving transformations. IEEE Computer Graphics and Applications, 1981. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 168, + 296, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 168, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 168, + 296, + 213 + ], + "type": "text", + "content": "[3] Junhao Cai, Jingcheng Su, Zida Zhou, Hui Cheng, Qifeng Chen, and Michael Yu Wang. Volumetric-based contact point detection for 7-dof grasping. In Conference on Robot Learning (CoRL), 2022. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 213, + 295, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 213, + 295, + 269 + ], + "spans": [ + { + "bbox": [ + 62, + 213, + 295, + 269 + ], + "type": "text", + "content": "[4] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical report, 2015. 2, 4, 5, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 269, + 295, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 269, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 269, + 295, + 323 + ], + "type": "text", + "content": "[5] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention Mask Transformer for Universal Image Segmentation. International Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 324, + 295, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 324, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 62, + 324, + 295, + 369 + ], + "type": "text", + "content": "[6] Bowen Cheng, Alexander G. Schwing, and Alexander Kirillov. Per-Pixel Classification is Not All You Need for Semantic Segmentation. In International Conference on Neural Information Processing Systems (NeurIPS), 2021. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 369, + 295, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 295, + 413 + ], + "type": "text", + "content": "[7] Laurent Chevalier, Fabrice Jaillet, and Atilla Baskurt. Segmentation and superquadric modeling of 3d objects. In International Conference in Central Europe on Computer Graphics and Visualization, 2003. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 414, + 295, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 295, + 458 + ], + "type": "text", + "content": "[8] Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3D-R2N2: A Unified Approach for Single and Multi-view 3D Object Reconstruction. In European Conference on Computer Vision (ECCV), 2016. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 459, + 295, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 459, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 459, + 295, + 523 + ], + "type": "text", + "content": "[9] Alexandros Delitzas, Ayca Takmaz, Federico Tombari, Robert Sumner, Marc Pollefeys, and Francis Engelmann. SceneFun3D: Fine-Grained Functionality and Affordance Understanding in 3D Scenes. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 525, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 295, + 590 + ], + "type": "text", + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies., 2019. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 591, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 645 + ], + "type": "text", + "content": "[11] Hao-Shu Fang, Chenxi Wang, Hongjie Fang, Minghao Gou, Jirong Liu, Hengxu Yan, Wenhai Liu, Yichen Xie, and Cewu Lu. Anygrasp: Robust and efficient grasp perception in spatial and temporal domains. IEEE Transactions on Robotics, 2023. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 647, + 295, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 680 + ], + "type": "text", + "content": "[12] B. Faverjon and J. Ponce. On Computing Two-finger Force-closure Grasps of Curved 2D Objects. In International Conference on Robotics and Automation (ICRA), 1991. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[13] Qiao Gu, Ali Kuwajerwala, Sacha Morin, Krishna Murthy Jatavallabhula, Bipasha Sen, Aditya Agarwal, Corban Rivera, William Paul, Kirsty Ellis, Rama Chellappa, et al." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "Conceptgraphs: Open-vocabulary 3d scene graphs for perception and planning. In International Conference on Robotics and Automation (ICRA), 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 106, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 106, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 106, + 553, + 171 + ], + "type": "text", + "content": "[14] Abdullah Hamdi, Luke Melas-Kyriazi, Jinjie Mai, Guocheng Qian, Ruoshi Liu, Carl Vondrick, Bernard Ghanem, and Andrea Vedaldi. Ges: Generalized exponential splatting for efficient radiance field rendering. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 172, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 172, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 172, + 553, + 227 + ], + "type": "text", + "content": "[15] Jan Held, Renaud Vandeghen, Abdullah Hamdi, Adrien Deli'ege, Anthony Cioppa, Silvio Giancola, Andrea Vedaldi, Bernard Ghanem, and Marc Van Droogenbroeck. 3D convex splatting: Radiance field rendering with 3D smooth convexes. ArXiv, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 228, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 553, + 259 + ], + "type": "text", + "content": "[16] Ales Jaklic, Ales Leonardis, and Franc Solina. Segmentation and recovery of superquadrics. Springer Science & Business Media, 2000. 3, 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 260, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 260, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 260, + 553, + 304 + ], + "type": "text", + "content": "[17] Rasmus Ramsbøl Jensen, A. Dahl, George Vogiatzis, Engil Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. International Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 304, + 553, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 553, + 349 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 553, + 349 + ], + "type": "text", + "content": "[18] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions On Graphics (TOG), 2023. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 350, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 415 + ], + "type": "text", + "content": "[19] Sebastian Koch, Narunas Vaskevicius, Mirco Colosi, Pedro Hermosilla, and Timo Ropinski. Open3DSG: Open-vocabulary 3D Scene Graphs from Point Clouds with Queryable Objects and Open-set Relationships. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 415, + 553, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 447 + ], + "type": "text", + "content": "[20] Kenneth Levenberg. A method for the solution of certain non-linear problems in least squares. Quarterly of Applied Mathematics, 1944. 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 449, + 553, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 481 + ], + "type": "text", + "content": "[21] Lei Li and Angela Dai. Genzi: Zero-shot 3d human-scene interaction generation. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 482, + 553, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 553, + 524 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 553, + 524 + ], + "type": "text", + "content": "[22] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. International Conference on Computer Vision (ICCV), 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 525, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 525, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 317, + 525, + 553, + 581 + ], + "type": "text", + "content": "[23] Weixiao Liu, Yuwei Wu, Sipu Ruan, and Gregory S Chirikjian. Robust and Accurate Superquadric Recovery: A Probabilistic Approach. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 582, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 553, + 634 + ], + "type": "text", + "content": "[24] Weixiao Liu, Yuwei Wu, Sipu Ruan, and Gregory S Chirikjian. Marching-primitives: Shape abstraction from signed distance function. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 635, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 680 + ], + "type": "text", + "content": "[25] Zhijian Liu, Haotian Tang, Yujun Lin, and Song Han. PointVoxel CNN for Efficient 3D Deep Learning. In International Conference on Neural Information Processing Systems (NeurIPS), 2019. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. International Conference on Learning Representations (ICLR), 2019. 1" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24633" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[27] Jeffrey Mahler, Matthew Matl, Vishal Satish, Michael Danielczuk, Bill DeRose, Stephen McKinley, and Ken Goldberg. Learning ambidextrous robot grasping policies. Science Robotics, 2019. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 119, + 295, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 119, + 295, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 295, + 152 + ], + "type": "text", + "content": "[28] Donald W. Marquardt. An algorithm for least-squares estimation of nonlinear parameters. Journal of The Society for Industrial and Applied Mathematics, 1963. 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 154, + 294, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 154, + 294, + 208 + ], + "spans": [ + { + "bbox": [ + 56, + 154, + 294, + 208 + ], + "type": "text", + "content": "[29] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 211, + 294, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 211, + 294, + 253 + ], + "spans": [ + { + "bbox": [ + 56, + 211, + 294, + 253 + ], + "type": "text", + "content": "[30] A.T. Miller, S. Knoop, H.I. Christensen, and P.K. Allen. Automatic Grasp Planning Using Shape Primitives. In International Conference on Robotics and Automation (ICRA), 2003. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 257, + 294, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 257, + 294, + 311 + ], + "spans": [ + { + "bbox": [ + 56, + 257, + 294, + 311 + ], + "type": "text", + "content": "[31] Tom Monnier, Jake Austin, Angjoo Kanazawa, Alexei A. Efros, and Mathieu Aubry. Differentiable blocks world: Qualitative 3d decomposition by rendering primitives. International Conference on Neural Information Processing Systems (NeurIPS), 2023. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 313, + 294, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 313, + 294, + 357 + ], + "spans": [ + { + "bbox": [ + 56, + 313, + 294, + 357 + ], + "type": "text", + "content": "[32] Despoina Paschalidou, Ali Osman Ulusoy, and Andreas Geiger. Superquadrics revisited: Learning 3d shape parsing beyond cuboids. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 359, + 294, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 359, + 294, + 413 + ], + "spans": [ + { + "bbox": [ + 56, + 359, + 294, + 413 + ], + "type": "text", + "content": "[33] Songyou Peng, Kyle Genova, Chiyu Max Jiang, Andrea Tagliasacchi, Marc Pollefeys, and Thomas Funkhouser. OpenScene: 3D Scene Understanding with Open Vocabularies. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 416, + 294, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 294, + 449 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 294, + 449 + ], + "type": "text", + "content": "[34] Alex P Pentland. Parts: structured descriptions of shape. In Proceedings of the Fifth AAAI National Conference on Artificial Intelligence, 1986. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 451, + 294, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 451, + 294, + 483 + ], + "spans": [ + { + "bbox": [ + 56, + 451, + 294, + 483 + ], + "type": "text", + "content": "[35] Maurizio Pilu and Robert B Fisher. Equal-distance sampling of superellipse models. In British Machine Vision Conference (BMVC), 1995. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 486, + 294, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 486, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 486, + 294, + 529 + ], + "type": "text", + "content": "[36] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space. International Conference on Neural Information Processing Systems (NeurIPS), 2017. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 532, + 294, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 532, + 294, + 575 + ], + "spans": [ + { + "bbox": [ + 56, + 532, + 294, + 575 + ], + "type": "text", + "content": "[37] Michael Ramamonjisoa, Sinisa Stekovic, and Vincent Lepetit. MonteBoxFinder: Detecting and Filtering Primitives to Fit a Noisy Point Cloud. In European Conference on Computer Vision (ECCV), 2022. 6, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 578, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 294, + 610 + ], + "type": "text", + "content": "[38] Aaron Ray, Christopher Bradley, Luca Carlone, and Nicholas Roy. Task and motion planning in hierarchical 3d scene graphs. arXiv preprint arXiv:2403.08094, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 613, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 666 + ], + "type": "text", + "content": "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. International Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[40] Zuria Bauer Marc Pollefeys Federico Tombari Leonidas Guibas Gao Huang Francis Engelmann Rui Huang, Guangyao Zhai. VIPs: Video Perception Model for 3D Scene Generation. arXiv preprint, 2025. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 126 + ], + "type": "text", + "content": "[41] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask transformer for 3d semantic instance segmentation. International Conference on Robotics and Automation (ICRA), 2022. 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 194 + ], + "type": "text", + "content": "[42] Jonas Schult, Sam Tsai, Lukas Hollein, Bichen Wu, Jialiang Wang, Chih-Yao Ma, Kunpeng Li, Xiaofang Wang, Felix Wimbauer, Zijian He, et al. Controlroom3d: Room generation using semantic proxy rooms. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "type": "text", + "content": "[43] Jaidev Shriram, Alex Trevithick, Lingjie Liu, and Ravi Ramamoorthi. Realmdreamer: Text-driven 3d scene generation with inpainting and depth diffusion. arXiv preprint arXiv:2404.07199, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 242, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 242, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 242, + 553, + 285 + ], + "type": "text", + "content": "[44] Franc Solina and Ruzena Bajcsy. Recovery of parametric models from range images: The case for superquadrics with global deformations. Transactions on Pattern Analysis and Machine Intelligence (PAMI), 1990. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 287, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 396 + ], + "type": "text", + "content": "[45] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Yuheng Ren, Shobhit Verma, Anton Clarkson, Ming Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke Malte Strasdat, Renzo De Nardi, Michael Goesele, S. Lovegrove, and Richard A. Newcombe. The replica dataset: A digital replica of indoor spaces. ArXiv, 2019. 2, 4, 5, 6, 7, 8, 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 398, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 398, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 398, + 553, + 441 + ], + "type": "text", + "content": "[46] Matan Sudry, Tom Jurgenson, Aviv Tamar, and Erez Karpas. Hierarchical planning for rope manipulation using knot theory and a learned inverse model. In Conference on Robot Learning, 2023. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 443, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 443, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 443, + 553, + 487 + ], + "type": "text", + "content": "[47] Martin Sundermeyer, Arsalan Mousavian, Rudolph Triebel, and Dieter Fox. Contact-grasnet: Efficient 6-dof grasp generation in cluttered scenes. In International Conference on Robotics and Automation (ICRA), 2021. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 489, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 553, + 542 + ], + "type": "text", + "content": "[48] Ayca Takmaz, Elisabetta Fedele, Robert W. Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. OpenMask3D: Open-Vocabulary 3D Instance Segmentation. In International Conference on Neural Information Processing Systems (NeurIPS), 2023. 1, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 544, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 588 + ], + "type": "text", + "content": "[49] Ayca Takmaz, Alexandros Delitzas, Robert W. Sumner, Francis Engelmann, Johanna Wald, and Federico Tombari. Search3D: Hierarchical Open-Vocabulary 3D Segmentation. IEEE Robotics and Automation Letters (RA-L), 2025. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 590, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 622 + ], + "type": "text", + "content": "[50] Xun Tu and Karthik Desingh. Superq-grasp: Superquadrics-based grasp pose estimation on larger objects for mobile-manipulation. arXiv, 2024. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 624, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 677 + ], + "type": "text", + "content": "[51] Shubham Tulsiani, Hao Su, Leonidas J. Guibas, Alexei A. Efros, and Jitendra Malik. Learning shape abstractions by assembling volumetric primitives. International Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "type": "text", + "content": "[52] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. Journal of machine learning research, 2008. 8" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24634" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 463 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[53] Erik Roeland van Dop and Paulus P.L. Regtien. Fitting undeformed superquadrics to range data: improving model recovery and classification. International Conference on Computer Vision and Pattern Recognition (CVPR), 1998. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "text", + "content": "[54] Ashish Vaswani, Noam M. Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is All you Need. In International Conference on Neural Information Processing Systems (NeurIPS), 2017. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 217 + ], + "type": "text", + "content": "[55] Giulia Vezzani, Ugo Pattacini, and Lorenzo Natale. A grasping approach based on superquadric models. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 1579-1586. IEEE, 2017. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 271 + ], + "type": "text", + "content": "[56] Chenxi Wang, Hao-Shu Fang, Minghao Gou, Hongjie Fang, Jin Gao, and Cewu Lu. Graspness Discovery in Clutters for Fast and Accurate Grasp Detection. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 274, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 294, + 316 + ], + "type": "text", + "content": "[57] Kaizhi Yang and Xuejin Chen. Unsupervised learning for cuboid shape abstraction via joint segmentation from point clouds. ACM Transactions On Graphics (TOG), 2021. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 294, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 361 + ], + "type": "text", + "content": "[58] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. International Conference on Computer Vision (ICCV), 2023. 2, 4, 6, 1, 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 364, + 294, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 364, + 294, + 429 + ], + "spans": [ + { + "bbox": [ + 56, + 364, + 294, + 429 + ], + "type": "text", + "content": "[59] Chenyangguang Zhang, Alexandros Delitzas, Fangjinhua Wang, Ruida Zhang, Xiangyang Ji, Marc Pollefeys, and Francis Engelmann. Open-Vocabulary Functional 3D Scene Graphs for Real-World Indoor Spaces. In International Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 430, + 294, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 430, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 56, + 430, + 294, + 463 + ], + "type": "text", + "content": "[60] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. International Conference on Computer Vision (ICCV), 2023. 7" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24635" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SuperEdit_ Rectifying and Facilitating Supervision for Instruction-Based Image Editing/395fe9d1-b3f1-4c45-afff-e5eaba8e26ad_content_list.json b/2025/SuperEdit_ Rectifying and Facilitating Supervision for Instruction-Based Image Editing/395fe9d1-b3f1-4c45-afff-e5eaba8e26ad_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ababcaa216666ecec75c749dbd4d1dd7429701db --- /dev/null +++ b/2025/SuperEdit_ Rectifying and Facilitating Supervision for Instruction-Based Image Editing/395fe9d1-b3f1-4c45-afff-e5eaba8e26ad_content_list.json @@ -0,0 +1,1495 @@ +[ + { + "type": "text", + "text": "SuperEdit: Rectifying and Facilitating Supervision for Instruction-Based Image Editing", + "text_level": 1, + "bbox": [ + 89, + 130, + 906, + 152 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c2ca5a6587dbdec196e1bd2af034b0cb961ad511d7ba97e98087c6cf70d746f1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 161, + 383, + 178 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Website", + "bbox": [ + 385, + 162, + 450, + 178 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1dfe022054d2bc31d98f7b4664ca6b2ef1ac3ad32b5b48f00734411f4ab16e0b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 473, + 161, + 496, + 178 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Code", + "bbox": [ + 498, + 162, + 540, + 178 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d3e90959e8e22205dbda6a58db2876dbbb512e05de6fbd9dbf9dfa236f9492a2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 568, + 162, + 594, + 179 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Data", + "bbox": [ + 596, + 162, + 637, + 178 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ming Li $^{1,2,\\ddagger}$ , Xin Gu $^{1}$ , Fan Chen $^{1}$ , Xiaoying Xing $^{1}$ , Longyin Wen $^{1}$ , Chen Chen $^{2}$ , Sijie Zhu $^{1,*}$ $^{1}$ ByteDance Intelligent Creation Center for Research in Computer Vision, University of Central Florida", + "bbox": [ + 99, + 189, + 897, + 224 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c0340ed1d5247b3e0f8c9581efff1a4ec5797b1263767e54034fb4c2cf301c49.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 234, + 379, + 314 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ec81b6f49f3c3d93a85a8e1d7c0df06b602a80611dc0aebba05c0b9564ed0e6a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 379, + 234, + 665, + 315 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7fe340698e5085dde31385bc7521f02a98d61a1cc9a8db24be8c7fb40023202d.jpg", + "image_caption": [ + "Global: Transform the scene into a snowy winter scene" + ], + "image_footnote": [], + "bbox": [ + 94, + 315, + 377, + 398 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/025353ddf90f7754c2b5b16880540198ad39bcc3750ba52e82369de167e7db07.jpg", + "image_caption": [ + "Replace: Replace the bow tie with a butterfly." + ], + "image_footnote": [], + "bbox": [ + 377, + 315, + 663, + 398 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b2e53b28238c2b7a879ed25964505f0ddd0abfc293eab28fe26a386f1f573554.jpg", + "image_caption": [ + "Style: Convert the image to a watercolor style Remove: Remove some clouds in the sky", + "(a) Examples of Our Method on Real & High-resolution Image" + ], + "image_footnote": [], + "bbox": [ + 94, + 405, + 377, + 482 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dd8e57b09606df98dbb654124ffa8c514c0b3eefc2dd82a2c69beed868c003fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 377, + 405, + 663, + 482 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1a709fe36d69a2d9f4630a351d54b3ecfa99c371e4339b663c3435495418bfa4.jpg", + "image_caption": [ + "Comparison with Existing Methods", + "(b) Comparison with GPT-4o Evaluation" + ], + "image_footnote": [], + "bbox": [ + 665, + 244, + 903, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e9026c6f1f07d055b479d701aa26a475843a7322b08da88f137dbb67b5d2360b.jpg", + "image_caption": [ + "(c) Comparison with Human Evaluation", + "Figure 1. (a) Our editing method works well with real and high-resolution images, handling various free-form edits (left) and local edits (right); (b) Compared to the current state-of-the-art SmartEdit, our method achieves a $9.19\\%$ performance improvement with $30\\times$ less training data and $13\\times$ fewer model parameters; (c) Our method achieves better overall scores on the human evaluation results, indicating more precise editing capabilities." + ], + "image_footnote": [], + "bbox": [ + 666, + 407, + 903, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 250, + 560, + 323, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Due to the challenges of manually collecting accurate editing data, existing datasets are typically constructed using various automated methods, leading to noisy supervision signals caused by the mismatch between editing instructions and original-edited image pairs. Recent efforts attempt to improve editing models through generating higher-quality edited images, pre-training on recognition tasks, or introducing vision-language models (VLMs) but fail to resolve this fundamental issue. In this paper, we offer a novel solution by constructing more effective editing instructions for given image pairs. This includes rectifying the editing instructions to better align with the original-edited image pairs and using contrastive editing instructions to further enhance their effectiveness. Specifically, we find that editing models exhibit specific generation attributes at different inference steps, independent of the text. Based on these prior attributes, we define a unified guide for VLMs to rectify editing instructions. However, there are some challenging editing scenarios that cannot be resolved solely with rectified instructions.", + "bbox": [ + 88, + 590, + 485, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To this end, we further construct contrastive supervision signals with positive and negative instructions and introduce them into the model training using triplet loss, thereby further facilitating supervision effectiveness. Our method does not require the VLM modules or pre-training tasks used in previous work, offering a more direct and efficient way to provide better supervision signals, and providing a novel, simple, and effective solution for instruction-based image editing. Results on multiple benchmarks demonstrate that our method significantly outperforms existing approaches. Compared with previous SOTA SmartEdit, we achieve $9.19\\%$ improvements on the Real-Edit benchmark with $30\\times$ less training data and $13\\times$ smaller model size. All data and models are open-sourced on Github for future research.", + "bbox": [ + 511, + 561, + 908, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 784, + 637, + 799 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, significant progress has been made in text-to-image (T2I) generation [10, 37, 40-42] due to the development of diffusion models [9, 21, 48, 49]. These T2I diffusion models can generate images that align with natural language descriptions while satisfying human perception and preferences. Consequently, numerous image editing", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author, sijiezhu@bytedance.com", + "bbox": [ + 112, + 875, + 364, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\ddagger$ This work was done during the first author's internship at ByteDance USA", + "bbox": [ + 112, + 887, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19206", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "methods [4, 6, 19, 31] based on these models have been proposed to achieve various editing effects. Instruction-based methods [4, 12, 24] have become increasingly popular as they allow users to conveniently and easily modify images using language instructions without the need to provide masks, as required by mask-based methods [18, 28, 47, 53].", + "bbox": [ + 89, + 90, + 482, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The training of instruction-based editing models requires the original-edited image pairs and corresponding editing instruction, making it difficult to manually create or collect a large amount of relevant data [58]. To address the issue of scarce training data, existing efforts [13, 19, 25, 62] have attempted to develop various automated pipelines to synthesize large datasets. Specifically, most methods first use large language models (LLMs) to modify the text descriptions of original images. The original images and modified texts are then input into various pre-trained diffusion models to automatically generate edited images. However, current text-to-image diffusion models struggle to fully correspond to input text prompts [15, 59]. This often changes parts of the original images that do not require editing, leading to misaligned editing instructions and original-edited image pairs, thus resulting in noisy supervision signals. To mitigate the potential issues of noisy supervision in image editing models, existing work has attempted to introduce additional recognition pre-training tasks for U-Net [43] such as semantic segmentation [14, 45], or replace CLIP [38] text encoder with vision-language models (VLMs) [12, 24] to better understand editing instructions from noisy supervision signals. However, these methods not only introduce significant computational overhead but also overlook the issue of noisy supervision signals.", + "bbox": [ + 91, + 185, + 486, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we focus on addressing the fundamental challenge by introducing more effective editing instructions, as demonstrated in Fig. 2. Our data-oriented method explores a different research question: how much performance improvement can be achieved solely by focusing on supervision signal quality and optimization in image editing? Surprisingly, SuperEdit outperforms existing methods in both GPT-4o and human evaluations, despite using less data and requiring no additional modules or pretraining as shown in Fig.1. This demonstrates that high-quality supervision signals can significantly compensate for architectural simplicity, achieving results comparable to or better than methods with more complex requirements.", + "bbox": [ + 89, + 535, + 483, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, to enhance the effectiveness of supervision signals for instruction-based image editing methods, we propose using VLMs to rectify editing instructions, creating better-aligned instructions for the original-edited image pairs. However, determining which VLM to use for this task and how to establish a unified rectification method for various editing instructions remain unexplored problems. To address this, we first analyze the ability of different VLMs to understand the differences between original and edited images, showing that GPT-4o [1] is the most capable of rectifying editing instructions. Additionally, we observe that both editing models and text-to-image diffusion models share a similar prior, as shown in Fig. 4: different inference", + "bbox": [ + 89, + 719, + 486, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/588b8b2639c6ce4ed4512fc8e5f3bd26b3b1ce7fe508459c2ed64d91aa84f442.jpg", + "image_caption": [ + "(c) InstructDiffusion; Emu-Edit", + "Figure 2. Unlike existing efforts that attempt to (a) scale up edited images with noisy supervision [4, 25], (b) introduce massive VLMs into editing model architecture [12, 24], and (c) perform additional pre-training tasks [14, 45], (d) we focus on improving the effectiveness of supervision signals, which is the fundamental issue of image editing." + ], + "image_footnote": [], + "bbox": [ + 535, + 87, + 704, + 282 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/39c8d4fcbb602f3d1c7c3d7dbf54a8dd3d16cdeb40b2d534de7497da4668aa2a.jpg", + "image_caption": [ + "(d) Ours" + ], + "image_footnote": [], + "bbox": [ + 707, + 88, + 880, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "stages correspond to the generation of different image attributes, independent of the input text prompt [2, 17, 19, 36, 46, 56, 61] or editing instructions. Inspired by this, we guide VLMs based on these attributes to establish a unified rectification guideline for various editing instructions, as demonstrated in Fig. 3.", + "bbox": [ + 511, + 371, + 906, + 446 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "When training with only rectified editing instructions, we find that the editing model can better understand the editing commands but still faces challenges in handling complex scenarios. For example, when the original image contains multiple objects, the edit model struggles to perform an accurate editing function if the instructions modify only one of these objects. Additionally, inherent issues present in pre-trained text-to-image diffusion models [15, 22-24], such as difficulty in understanding quantity, position, or object relationships, persist in the editing models. To address these issues, we propose using contrastive supervision signals to further optimize the editing model. Specifically, we first construct incorrect editing instructions based on the rectified instructions to generate positive and negative samples. We then introduce a triplet loss to guide the model, thereby enhancing the effectiveness of supervision, as shown in Fig. 5.", + "bbox": [ + 511, + 448, + 908, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are summarized as follows:", + "bbox": [ + 537, + 675, + 898, + 688 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- New Insight: We aim to address the noisy supervision problem that arises from the misalignment between editing instructions and original-edited image pairs, which is a fundamental issue overlooked by previous work, as shown in Fig. 2.", + "- Rectifying Supervision: We leverage diffusion generation priors to guide the vision-language model to generate better-aligned editing instructions for original-edited image pairs.", + "- Facilitating Supervision: We introduce contrastive supervision using triplet loss, enabling the editing model to learn from both positive and negative editing instructions.", + "- Promising Results: We achieve significant improvements on multiple benchmarks without additional pre-training or VLM. Compared to SmartEdit [24], we achieved a $9.19\\%$ improvement while reducing $30 \\times$ data and $13 \\times$ model parameters." + ], + "bbox": [ + 513, + 689, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19207", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 91, + 89, + 225, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Image Editing with Diffusion Models", + "text_level": 1, + "bbox": [ + 89, + 114, + 410, + 131 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Building on advancements in text-to-image (T2I) diffusion models [10, 37, 40-42, 44], recent research has explored them for image editing [4, 19]. Training-free methods [5, 19, 29, 31, 34, 51] typically achieve this by adjusting attentions in pre-trained T2I models, but have limited performance and generalization capabilities on various editing tasks.", + "bbox": [ + 89, + 137, + 483, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training-based methods address these limitations with specialized editing models, which can be categorized into mask-based and instruction-based approaches. Mask-based methods [6, 18, 28, 47, 53, 57] enable fine-grained local edits with user-provided or predicted masks and corresponding text descriptions. However, it struggles with global image editing and is constrained by the lack of mask-based editing data [24].", + "bbox": [ + 89, + 228, + 483, + 334 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instruction-based methods directly accept textual commands, such as \"add a dog\", offering better editing flexibility and generalization. InstructPix2Pix [4] pioneered this paradigm by generating instruction-based editing data and modifying the conditions of T2I diffusion models. Building on this framework, subsequent work introduces vision-language models [12, 24, 33] or additional pre-training tasks for the denoising U-Net [14, 24, 43, 45] to enhance the understanding and reasoning of input conditions. However, these methods not only introduce substantial computational overhead but also overlook the fundamental noisy supervision issue.", + "bbox": [ + 89, + 335, + 483, + 501 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Generating and Improving Editing Supervision", + "text_level": 1, + "bbox": [ + 89, + 511, + 482, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Due to the difficulty of scaling instruction-based image editing data through manual collection, existing efforts [4, 13, 25, 58, 62] aim to automatically modify text descriptions of original images and generate edited images with T2I diffusion models. However, this approach often produces synthesized images that do not align with the editing instructions, as shown in Figure 3, resulting in noisy editing supervision signals [58, 62]. To address this, MagicBrush [58] manually filters out incorrect editing data, but it is hard to scale. Unlike existing methods focusing on edited image quality, we leverage diffusion prior and vision-language model (i.e., GPT-40 [1]) to create better-aligned instructions with original-edited image pairs, providing more accurate supervision.", + "bbox": [ + 89, + 534, + 483, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Alignment of Diffusion Models", + "text_level": 1, + "bbox": [ + 89, + 742, + 362, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The success of alignment training in large language models (LLMs) [26, 32, 39] has been applied to diffusion models for better image generation. This is achieved by maximizing reward scores [8, 27, 54] or the generation probability of the winner image in a pair [11, 52, 55]. In image editing, HIVE [60] and MultiReward [16] attempt to incorporate reward information into the text condition to align the editing model. In contrast, we guide the editing model by rectifying and constructing contrastive editing instructions, achieving more effective alignment.", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 89, + 599, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first introduce the most general image editing framework in Sec. 3.1. Then, we explain how to use diffusion priors to rectify editing instructions with the multimodal model (i.e., GPT-4o) in Sec. 3.2, thereby enhancing the accuracy of supervision signals. Finally, we describe how to construct contrastive supervision with both correct and incorrect editing instructions and integrate it into the editing model training using triplet loss in Sec. 3.3.", + "bbox": [ + 511, + 114, + 906, + 236 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Instruction-based Image Editing Framework", + "text_level": 1, + "bbox": [ + 511, + 246, + 893, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "InstructPix2Pix [4] pioneered instruction-based image editing, performing editing tasks by simultaneously taking the original image $C^I$ and editing instructions $C^T$ as input conditions to generate the edited image $x$ from random noise $\\epsilon$ . Following the definition of DDPM [21], we randomly sample a timestep $t \\in T$ during training, and then add corresponding noise $\\epsilon_t$ to the edited image $x$ :", + "bbox": [ + 511, + 268, + 906, + 375 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t}, \\quad \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, I), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 385, + 906, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon$ is a noise map sampled from a Gaussian distribution, and $\\bar{\\alpha}_t\\coloneqq \\prod_{s = 0}^t\\alpha_s$ , $\\alpha_{t} = 1 - \\beta_{t}$ is a differentiable function of timestep $t$ , which is determined by the denoising sampler such as DDPM [21]. Then the training objective of the editing model $\\epsilon_{\\theta}$ is predicting the added noise at timestep $t$ , which can be written as:", + "bbox": [ + 511, + 414, + 905, + 502 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t r a i n}} = \\mathbb {E} _ {x _ {t}, t, C ^ {I}, C ^ {T}, \\epsilon} \\left[ \\left\\| \\epsilon_ {\\theta} \\left(\\operatorname {c o n c a t} \\left(x _ {t}, C ^ {I}\\right), t, C ^ {T}\\right) - \\epsilon_ {t} \\right\\| _ {2} ^ {2} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 508, + 906, + 535 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where concat refers to concatenating the image latents of noised edited image $x_{t}$ and original image $c_{I}$ in the channel dimension.", + "bbox": [ + 511, + 545, + 906, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Rectifying Supervision with Diffusion Priors", + "text_level": 1, + "bbox": [ + 511, + 584, + 888, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 3, existing image editing datasets [4, 13, 58] typically use only Steps 1 and 2: LLMs construct editing prompts and captions, and then text-to-image diffusion models synthesize edited images. However, diffusion models often fail to accurately follow prompts while maintaining image layout, creating mismatches between original-edited pairs and editing instructions, resulting in inaccurate supervision. While better supervision signals for text-to-image diffusion models are common in image generation [3, 50], this approach remains underexplored in image editing due to two challenges: (1) VLMs trained on single-image data struggle with multi-image inputs, and (2) editing instructions vary widely, making unified rectification guidelines difficult. To address these issues, we: (1) analyzed different VLMs' capabilities with multi-image inputs, finding GPT-4o most effective, and (2) discovered that timestep-specific roles in image generation also apply to editing, providing a foundation for a unified rectification method across various instructions (Fig. 3 and 4). Due to page limitations, our VLM analysis is in the Supplementary Material, while this section focuses on Diffusion Prior and Editing Instruction Rectification.", + "bbox": [ + 511, + 606, + 908, + 909 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19208", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fdde02f13f77e9774b108c3b58180ea004f09ab08d1e62c594fabfdff486e951.jpg", + "image_caption": [ + "Figure 3. (a) Existing work primarily uses LLMs and diffusion models to automatically generate edited images. However, current diffusion models often fail to accurately follow text prompts while maintaining the input image's layout, resulting in mismatches between the original-edited image pairs and the editing instructions. (b) We perform instruction rectification (Step 3) based on the images constructed in Steps 1 and 2. We show VLMs can understand the differences between the images, enabling them to rectify editing instructions to be better aligned with original-edited image pairs." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 401, + 273 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f23c3efcdc9fefb30d130072c56bac743446b7a251c68c1d7ceb3b08c5514d80.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 88, + 727, + 273 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1645bbc98dc6936728b9ff771ba16479949c73578c30bf58456f7f36312800d6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 723, + 95, + 898, + 258 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5e45d22c74c34dc577fc4e7888b3561686cb3741c730bf1df9aaf04e785d8fd3.jpg", + "image_caption": [ + "Figure 4. We show that the editing model follows consistent generation attributes at different sampling stages, independent of the editing instructions. The early, middle, and late sampling stages correspond to global, local, and detail changes, respectively, while style changes occur at all stages. All the generated images here are DDIM 30-step sampled final images. The orange progress bar and the grid progress bar represent the sampling stages with and without the editing instructions, respectively." + ], + "image_footnote": [], + "bbox": [ + 91, + 358, + 478, + 661 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion Generation Priors. Previous work has shown that different timesteps play distinct roles in image generation for text-to-image diffusion models, regardless of the text prompt [2, 17, 19, 36, 46, 56, 61]. We find that this phenomenon also exists in instruction-based editing models and present examples based on pre-trained InstructPix2Pix [4], as shown in Fig. 4. Specifically, diffusion models focus on global layout in the early stages, local object attributes in the mid stages, and image details", + "bbox": [ + 89, + 779, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in the late stages of sampling. This finding inspires us to guide VLMs based on these four generation attributes, establishing a unified rectification method for various editing instructions. We provide more analysis and details in the Supplementary Material.", + "bbox": [ + 511, + 359, + 908, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Editing Instruction Rectification. As demonstrated in Fig. 3, we extend the existing editing data generation pipeline by introducing our instruction rectification (Step 3). This process relies on the original edited image pairs obtained through Steps 1 and 2 from previous work. Specifically, we input original-edited image pairs into the vision-language model (i.e., GPT-4o) and instruct it to describe the changes in the edited image compared to the original image according to the above diffusion prior generation attributes. Finally, we use VLM to summarize the instructions and ensure that its length is less than the maximum length of CLIP text encoder, which is 77 tokens.", + "bbox": [ + 511, + 440, + 908, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Facilitating Supervision with Contrastive Instructions", + "text_level": 1, + "bbox": [ + 511, + 619, + 906, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although using rectified editing instructions can significantly improve performance across various editing tasks, we find that editing models still struggle with closely related text instructions. For example, \"add a cat on the left side of the image\" and \"add two cats on the right side of the image\" might produce the same edited image. This indicates that inherent biases in pre-trained text-to-image diffusion models [15, 22], such as difficulties in understanding quantity, position, and spatial relationships, persist in editing models. More importantly, our experiments show that training models with rectified editing instructions does not resolve these challenges. To further facilitate supervision signal effectiveness, we drew on successful alignment experiences from large language models [1, 32, 39] and text-to-image diffusion models [7, 52, 54]: constructing positive and negative sample pairs and guiding the model to assign a higher generation probability to positive samples compared to negative ones.", + "bbox": [ + 511, + 659, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19209", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/162a52a8a1dcc266331549d208f7f177bbf2aa21ad63ea7dbd4fdce26317e909.jpg", + "image_caption": [ + "Figure 5. (a) Based on the rectified editing instruction and original-edited image pair, we utilize the Vision-Language Models (VLM) to generate various image-related wrong instructions. These involve random substitutions of quantities, spatial locations, and objects within the rectified editing instructions according to the original-edited images context; (b) During each training iteration, we randomly select one wrong instruction $c_{\\text{neg}}^T$ and input it along with the rectified instruction $c_{pos}^T$ into the editing model to obtain predicted noises. The goal is to make the rectified instruction's predicted noise $\\epsilon_{pos}$ closer to the sampled training diffusion noise $\\epsilon$ , while ensuring the noise from incorrect instructions $\\epsilon_{\\text{neg}}$ is farther. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 413, + 277 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d2df2c9f86bef1789db5a2f08610bce45bd67df87c7d4970f19964fcf01d8296.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 92, + 906, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Constructing Contrastive Instructions. Unlike the standard alignment process for large language models or text-to-image diffusion models, it is challenging to generate different editing results from the same instruction to create positive and negative sample pairs for image editing tasks. To address this, we construct positive and negative editing instructions for alignment, thereby generating relatively positive and negative edited images. As shown in Fig. 5 (a), we use the original image, edited image, and rectified editing instruction as input. The VLM (i.e., GPT-4o) is used to modify attributes in the rectified editing instruction, such as quantity, spatial relationships, and object types, to create different wrong instructions. Here, we require VLM to modify only a single attribute from the rectified editing instruction in each wrong instruction, keeping most of the editing text unchanged. Since only a few words are replaced between the rectified instruction and the wrong instruction, the text embeddings produced by the CLIP text encoder that serve as input to the denoising model will also be similar. This ensures the task's learning difficulty, helping the model understand how subtle differences between the two editing instructions result in significantly different editing results.", + "bbox": [ + 91, + 372, + 483, + 690 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Facilitating Editing Models with Contrastive Instructions. Our key insight is that enhancing the effectiveness of supervision signals can improve various editing tasks without introducing additional model architectures or pre-training tasks. Therefore, we adhere strictly to the InstructPix2Pix [4] model architecture and training pipeline. To be specific, the inputs including the original image $c^I$ , edited image $x$ , the rectified instruction $c_{pos}^T$ , and wrong editing instruction $c_{neg}^T$ . During training, we will add a sampled timestep $t \\in T$ to obtain the noised edited image $x_t$ with Equation 1. Both the rectified and wrong editing instructions are fed into the denoising model to predict the final noises $\\epsilon_{pos}$ and $\\epsilon_{neg}$ , which are then used to", + "bbox": [ + 89, + 719, + 485, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "construct positive and negative samples, respectively:", + "bbox": [ + 511, + 372, + 846, + 387 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {p o s} = \\epsilon_ {\\theta} \\left(\\operatorname {c o n c a t} \\left(x _ {t}, c ^ {I}\\right), t, c _ {p o s} ^ {T}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 396, + 906, + 415 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {n e g} = \\epsilon_ {\\theta} \\big (\\operatorname {c o n c a t} \\left(x _ {t}, c ^ {I}\\right), t, c _ {n e g} ^ {T} \\big). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 417, + 906, + 438 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After constructing the positive and negative sample pairs, we aim for the noise predicted by the positive editing instruction $\\epsilon_{pos}$ to be closer to the true noise $\\epsilon_t$ sampled during training, compared to the noise predicted by the wrong editing instruction $\\epsilon_{neg}$ . This goal can be achieved through a triplet loss function:", + "bbox": [ + 511, + 446, + 906, + 523 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t r i p l e t}} = \\max \\left\\{d \\left(\\epsilon_ {t}, \\epsilon_ {\\text {p o s}}\\right) - d \\left(\\epsilon_ {t}, \\epsilon_ {\\text {n e g}}\\right) + m, 0 \\right\\}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 532, + 906, + 550 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d(x,y) = \\|\\mathbf{x} - \\mathbf{y}\\|_2^2$ and margin $m$ is a hyper-parameter. The final training loss is the combination of the original diffusion training loss and the triplet loss:", + "bbox": [ + 511, + 560, + 906, + 607 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {t r a i n}} + \\lambda \\cdot \\mathcal {L} _ {\\text {t r i p l e t}}, \\text {w h e r e} \\mathcal {L} _ {\\text {t r a i n}} = d \\left(\\epsilon_ {t}, \\epsilon_ {\\text {p o s}}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 618, + 906, + 635 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Please note that the contrastive supervision signals are only used during the training phase. During inference, the editing model only requires one single input editing instruction.", + "bbox": [ + 511, + 643, + 906, + 690 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 511, + 702, + 633, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Data Collection and Construction", + "text_level": 1, + "bbox": [ + 511, + 727, + 808, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To build a diverse dataset with various types of editing instructions, we need original and edited images from different data domains, as well as a wide variety of editing instructions. To achieve this, we sampled data from different public editing datasets to construct rectified and contrastive supervision signals. Specifically, we extracted 10,177, 8,807, and 21,016 editing pairs from InstructPix2Pix [4], MagicBrush [58], and Seed-Data-Edit [13], respectively, resulting in a total of 40,000 training samples. During extraction, we strive to ensure that the data for different types of editing tasks is as balanced as possible.", + "bbox": [ + 511, + 750, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19210", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/35dbcee4f093c470ccb517a3cce40883521fa0057500e611f1e7679b3bc1a05b.jpg", + "image_caption": [ + "Figure 6. Visual comparison with existing methods and the corresponding human-aligned GPT-4o evaluation scores (Following, Preserving, Quality Scores from left to right). We achieve better results while preserving the layout, quality, and details of the original image. Please note that we do not claim that our editing results are flawless. We provide more visual comparison results in the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 116, + 87, + 880, + 851 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19211", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0a37ff70ed3a6ccb3cbd08df08ba02fedc7c167a4cb8c5294066674c27a0a950.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodExtra ModulePretrain TasksEdit DataModel SizeFollowing ↑Preserving ↑Quality ↑Overall ↑
AccScoreAccScoreAccScoreAccScore
KOSMOS-G [33]✓✓9.0M1.9B51%2.829%1.4327%3.2029.0%2.48
MGIE [12]✓✓1.0M8.1B40%2.4345%2.7938%3.3541.0%2.86
SmartEdit [24]✓✓1.2M14.1B64%3.5066%3.7045%3.5658.3%3.59
MultiReward [16]✓✓320K1.2B63%3.3958%3.4354%3.8058.3%3.54
InstructDiffusion [14]X✓860K1.1B52%2.8754%3.1745%3.5850.3%3.21
InstructPix2Pix [4]XX300K1.1B52%2.9453%3.3150%3.6951.7%3.31
MagicBrush [58]XX310K1.1B51%2.9070%3.8550%3.6757.0%3.47
HIVE [60]XX1.1M1.1B54%2.9356%3.3653%3.7254.3%3.34
HQ-Edit [25]XX500K1.1B51%2.8416%1.6354%3.8440.3%2.77
SuperEdit (Ours)XX40K1.1B67%3.5977%4.1465%4.0169.7%3.91
", + "bbox": [ + 133, + 80, + 859, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/508db1f1a9247e288ae28c88a641cb6d98e7eb900f9d1be3a63b0af34aa08f3a.jpg", + "image_caption": [ + "(a) Following Score" + ], + "image_footnote": [], + "bbox": [ + 93, + 301, + 359, + 378 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e1c861c97676cba85a10ded11521df6c0aaeb41c54fcf90de3ec9dd7ae60237d.jpg", + "image_caption": [ + "(b) Preserving Score", + "Figure 7. Human evaluation on three evaluation criteria for image editing effects. (a) Following: whether the edited image adheres to the editing instructions; (b) Preserving: whether the image structure outside of the editing instructions has been preserved; (c) Quality: whether the overall quality/aesthetics of the edited image has been degraded compared to the input image. Our SuperEdit achieves the best results on all of these metrics." + ], + "image_footnote": [], + "bbox": [ + 364, + 301, + 630, + 377 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d9cd5943e7f01b703b88967325315bdf68d820d296c8d04fdbef5ecaae428227.jpg", + "image_caption": [ + "(c) Quality Score" + ], + "image_footnote": [], + "bbox": [ + 635, + 301, + 898, + 377 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9b93bcc462f1625d125bf5a8f2dca2f5ebaac0d69bb164fe37a768d9e0bb4580.jpg", + "table_caption": [ + "Table 1. Comparison with instruction-based image editing methods on Real-Edit benchmark [16]. Compared to existing work, our method achieves state-of-the-art performance across all metrics using a small amount of high-quality editing data without introducing additional models or pre-training tasks. Please note that the scores range from 0 to 5. $\\uparrow$ denotes a higher result is better. All baseline results are cited from the MultiReward [16] paper." + ], + "table_footnote": [], + "table_body": "
Following ↑Preserving ↑Quality ↑Overall ↑
InstructPix2Pix [4]2.412.622.442.49
SmartEdit-13B [24]3.093.062.632.93
SuperEdit3.18+1.80%3.86+16.00%3.37+14.80%3.47+10.80%
", + "bbox": [ + 94, + 448, + 478, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Human evaluation results on Real-Edit [16] benchmark. All the human-evaluated scores range from 0 to 5. Overall represents the average score of Following, Preserving, and Quality scores.", + "bbox": [ + 89, + 500, + 482, + 544 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We then applied our proposed methods in Sec. 3 to rectify and construct contrastive editing instructions for these training samples. Since the MagicBrush data has been manually verified, we skip the rectification step for this dataset and directly construct contrastive supervision based on the original editing instructions. For Seed-Data-Edit dataset, we only sample images from the first part of data without human editing instructions.", + "bbox": [ + 89, + 556, + 483, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 679, + 297, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Benchmarks and Metrics. To more accurately assess the effectiveness of various editing models, we conducted assessments on the Real-Edit benchmark [16], which is a human-aligned evaluation benchmark with GPT-4o scoring. Specifically, MultiReward [16] uses high-resolution images from the Unsplash community as a test dataset and combines them with GPT-4o [1] to create an automated evaluation method for single-turn editing. It assesses edited images in terms of accuracy (\\%) and scores (from 0 to 5), evaluating whether they adhere to the editing instructions (Following), whether the image structure outside of the editing instructions has been preserved (Preserving), and whether the overall quality/aesthetics of the edited image has been degraded compared to the original one (Quality).", + "bbox": [ + 89, + 703, + 485, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 450, + 718, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison on Real-Edit Benchmark. In Tab. 1, we present the quantitative results of editing effectiveness on the Real-Edit benchmark [16]. Without introducing additional parameters or pre-training stages, our method achieves the best results in the three GPT-4o automated evaluation metrics: Following, Preserving, and Quality, each of which includes percentage accuracy (Acc) and scores (from 0 to 5). For example, compared to SmartEdit [24], which introduces an additional 13B vision-language model (i.e., LLaVA [30]) to the 1.1B InstructPix2Pix [4] framework, we achieved improvements of $11.4\\%$ Overall Score. This suggests that given accurate and effective supervision signals, the trained editing model can understand and successfully execute the editing instructions, without the need for additional vision-language models.", + "bbox": [ + 511, + 473, + 906, + 685 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "It is worth noting that unlike existing image editing methods, which often show improvement in a single metric while others remain unchanged or worsen, our method achieves comprehensive and significant advancements across all three metrics. This indicates that improving the effectiveness of supervision signals can accurately execute editing instructions while reducing disruption to other non-edited parts of the image, and preserving the quality and aesthetics of the original images. Specifically, we surpassed the current best methods by $3\\%$ , $7\\%$ , and $11\\%$ Acc results in Following, Preserving, and Quality, respectively.", + "bbox": [ + 511, + 686, + 908, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Human Evaluation We also conduct a comprehensive human evaluation on Real-Edit benchmarks [16] in Tab. 2 and Fig. 7. The assessment involved 15 experienced evaluators", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19212", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "who rated edited images based on three critical metrics: instruction faithfulness (Following), preservation of irrelevant content (Preserving), and visual quality (Quality). The results of this manual evaluation demonstrate strong consistency with the GPT-4o scoring results shown in Tab. 1, with Cohen's Kappa=0.89, Krippendorff's Alpha=0.87. This high alignment thoroughly validates that our proposed SuperEdit significantly outperforms existing methods across all evaluation criteria. Specifically, our SuperEdit surpasses the previous state-of-the-art method SmartEdit [24] by $1.8\\%$ , $16\\%$ , $14.8\\%$ , and $10.8\\%$ on Following, Preserving, Quality, and Overall scores, respectively. These substantial improvements not only confirm the effectiveness of our approach but also establish SuperEdit as a new benchmark in instruction-guided image editing, achieving superior performance while requiring significantly less training data and cost.", + "bbox": [ + 89, + 90, + 485, + 318 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visual Comparison with State-of-the-art Methods. We show the visual comparison with existing image editing methods in Fig. 6. Compared to existing instruction-based editing methods, our approach not only better understands and executes editing instructions but also preserves the original image's layout and quality more effectively, thereby significantly outperforming previous methods. For example, with the instruction \"Replace the tiger with a lion, maintaining the same position in the water\" our SuperEdit method achieved superior results (4.8/4.8/4.8) compared to SmartEdit (4.8/4.8/2.5) and other methods. Additionally, our method improves the model's comprehension of editing instructions. For the instruction \"Change the background to a sandy beach with the ocean in the distance\" our method received perfect scores (4.8/5.0/5.0) while SmartEdit only achieved (5.0/4.8/4.0). Similarly, for style transformation instructions like \"Change the image style to look like an impressionist painting style\" SuperEdit significantly outperformed SmartEdit with scores of (4.8/4.8/4.8) versus (1.0/4.8/4.8), demonstrating our method's superior ability to handle complex artistic transformations. Even more impressively, for scene transformation tasks like \"Transform the entire scene to a winter setting with snow covering the houses, trees, and boat\", our SuperEdit achieved (5.0/4.8/4.8) while SmartEdit only obtained (2.0/4.5/4.5). We provide more visual comparisons with other instruction-based image editing methods in the Supplementary Material.", + "bbox": [ + 91, + 327, + 486, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 712, + 243, + 728 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on the Rectified and Contrastive Instructions. Considering that the Real-Edit [16] benchmark is evaluated by GPT-4o [1], and its evaluation results closely align with human ratings [16], we choose this benchmark to conduct ablation experiments in Tab. 3. Compared to the original 40K training data, our training data with rectified editing instructions significantly improves all the performance of the editing model. Specifically, our approach improves scores by 0.95, 0.79, and 0.11, and accuracy by $21\\%$ , $22\\%$ , and $4\\%$ in these three metrics, respectively. In addition, editing performance is further enhanced by incorporating contrastive supervision signals. Compared to using only", + "bbox": [ + 89, + 734, + 485, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7d06eb71ef8ba6b2484c13c9d645cf740a18b4e0589373fcb60be2732f31430d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Rectified InstructionContrastive InstructionsFollowing↑Preserving↑Quality↑
AccScoreAccScoreAccScore
XX41%2.4553%3.2761%3.90
✓X62%3.4075%4.0665%4.01
✓✓67%3.5977%4.1465%4.01
", + "bbox": [ + 517, + 89, + 903, + 152 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "rectified editing instructions, the introduction of contrastive supervision signals improves the following and preserving scores by 0.19 and 0.08, and accuracy by $5\\%$ and $2\\%$ , while maintaining the quality accuracy and score. In summary, both the introduction of rectified editing instructions and contrastive editing instructions improve the overall performance of the editing model.", + "bbox": [ + 511, + 194, + 908, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on Data Scaling. We investigated the impact of training data volume on model performance by experimenting with datasets ranging from 5k to 40k samples. Tab. 4 shows consistent improvements across all metrics as training data increases. With just 5k samples, our model achieves reasonable performance (54.7% accuracy, 3.42 overall score), but scaling to 40k samples yields substantial gains (69.7% accuracy, 3.91 overall score). The most significant improvements appear in the Preserving and Quality metrics, with 10% and 15%, respectively. This upward trend across all data points demonstrates that SuperEdit effectively leverages additional training examples without performance saturation, suggesting potential for further gains with larger datasets.", + "bbox": [ + 511, + 295, + 908, + 492 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b17783e3234fd5a6fa44150a75cb6b7d3b8e3f988e3eaaccd0fe8d4a0e1cc5fc.jpg", + "table_caption": [ + "Table 3. Ablation study on our methods. Both rectified and contrastive editing instructions achieved improvements across all metrics." + ], + "table_footnote": [], + "table_body": "
Data SizeFollowing ↑Preserving ↑Quality ↑Overall ↑
AccScoreAccScoreAccScoreAccScore
5k49%2.8760%3.7155%3.6954.7%3.42
10k57%3.2671%3.7658%3.8762.0%3.63
20k64%3.4072%4.0263%3.9466.3%3.79
40k67%3.5977%4.1465%4.0169.7%3.91
", + "bbox": [ + 517, + 497, + 901, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation study on data scaling results on Real-Edit [16].", + "bbox": [ + 524, + 575, + 890, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 603, + 627, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we re-examine image editing models from the perspective of enhancing supervision signals, finding that existing models have not adequately addressed this challenge, resulting in suboptimal performance. We introduce a unified editing instruction rectification guideline based on diffusion priors that better aligns instructions with original-edited image pairs, thereby enhancing supervision effectiveness. We also construct contrastive editing instructions allowing models to learn from both positive and negative examples. Our data-oriented approach explores an important but overlooked research question: What level of performance can be achieved with minimal architectural modifications by primarily focusing on supervision quality and optimization? Remarkably, under both GPT-4o and human evaluation, our method outperforms existing approaches despite using less data and requiring no architectural modifications or additional pretraining. This shows high-quality supervision signals can effectively compensate for architectural simplicity, offering valuable new perspectives for image editing research.", + "bbox": [ + 511, + 628, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19213", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 183, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 3, 4, 7, 8", + "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2, 4", + "[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2023. 3", + "[4] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 2, 3, 4, 5, 7, 1", + "[5] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In ICCV, 2023. 3", + "[6] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. In ICLR, 2023. 2, 3", + "[7] Yusuf Dalva and Pinar Yanardag. Noiseclar: A contrastive learning approach for unsupervised discovery of interpretable directions in diffusion models. In CVPR, 2024. 4", + "[8] Fei Deng, Qifei Wang, Wei Wei, Tingbo Hou, and Matthias Grundmann. Prdp: Proximal reward difference prediction for large-scale reward finetuning of diffusion models. In CVPR, 2024. 3", + "[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 2021. 1", + "[10] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 1, 3", + "[11] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Reinforcement learning for fine-tuning text-to-image diffusion models. NeurIPS, 2024. 3", + "[12] Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding instruction-based image editing via multimodal large language models. In ICLR, 2024. 2, 3, 7", + "[13] Yuying Ge, Sijie Zhao, Chen Li, Yixiao Ge, and Ying Shan. Seed-data-edit technical report: A hybrid dataset for instructional image editing. arXiv preprint arXiv:2405.04007, 2024. 2, 3, 5", + "[14] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, et al. Instructdiffusion: A generalist modeling interface for vision tasks. In CVPR, 2024. 2, 3, 7, 1", + "[15] Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. NeurIPS, 2024. 2, 4", + "[16] Xin Gu, Ming Li, Libo Zhang, Fan Chen, Longyin Wen, Tiejian Luo, and Sijie Zhu. Multi-reward as condition for instruction-based image editing. In ICLR, 2025. 3, 7, 8, 1" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Qin Guo and Tianwei Lin. Focus on your instruction: Fine-grained and multi-instruction image editing by attention modulation. In CVPR, 2024. 2, 4", + "[18] Qin Guo and Tianwei Lin. Focus on your instruction: Fine-grained and multi-instruction image editing by attention modulation. In CVPR, 2024. 2, 3", + "[19] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In ICLR, 2023. 2, 3, 4", + "[20] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 1", + "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 1, 3", + "[22] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv preprint arXiv:2403.05135, 2024. 2, 4", + "[23] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. NeurIPS, 2023.", + "[24] Yuzhou Huang, Liangbin Xie, Xintao Wang, Ziyang Yuan, Xiaodong Cun, Yixiao Ge, Jiantao Zhou, Chao Dong, Rui Huang, Ruimao Zhang, et al. Smartedit: Exploring complex instruction-based image editing with multimodal large language models. In CVPR, 2024. 2, 3, 7, 8, 1, 4", + "[25] Mude Hui, Siwei Yang, Bingchen Zhao, Yichun Shi, Heng Wang, Peng Wang, Yuyin Zhou, and Cihang Xie. HQ-edit: A high-quality dataset for instruction-based image editing. arXiv preprint arXiv:2404.09990, 2024. 2, 3, 7", + "[26] Harrison Lee, Samrat Phatale, Hassan Mansoor, Thomas Mesnard, Johan Ferret, Kellie Lu, Colton Bishop, Ethan Hall, Victor Carbune, Abhinav Rastogi, et al. Rlaif: Scaling reinforcement learning from human feedback with ai feedback. arXiv preprint arXiv:2309.00267, 2023. 3", + "[27] Ming Li, Taojiannan Yang, Huafeng Kuang, Jie Wu, Zhaoning Wang, Xuefeng Xiao, and Chen Chen. Controlnet++: Improving conditional controls with efficient consistency feedback. In ECCV, 2024. 3", + "[28] Yuanze Lin, Yi-Wen Chen, Yi-Hsuan Tsai, Lu Jiang, and Ming-Hsuan Yang. Text-driven image editing via learnable regions. In CVPR, 2024. 2, 3", + "[29] Bingyan Liu, Chengyu Wang, Tingfeng Cao, Kui Jia, and Jun Huang. Towards understanding cross and self-attention in stable diffusion for text-guided image editing. In CVPR, 2024. 3", + "[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2024. 7", + "[31] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 2, 3", + "[32] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. NeurIPS, 2022. 3, 4", + "[33] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-g: Generating images in context with multimodal large language models. In ICLR, 2024. 3, 7" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19214", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH, 2023. 3", + "[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 1", + "[36] Or Patashnik, Daniel Garibi, Idan Azuri, Hadar Averbuch-Elor, and Daniel Cohen-Or. Localizing object-level shape variations with text-to-image diffusion models. In ICCV, 2023. 2, 4", + "[37] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2023. 1, 3", + "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3", + "[39] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. NeurIPS, 2024. 3, 4", + "[40] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 1, 3", + "[41] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.", + "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 3", + "[43] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 2, 3", + "[44] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurlPS, 2022. 3", + "[45] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and generation tasks. In CVPR, 2024. 2, 3, 1", + "[46] Enis Simsar, Alessio Tonioni, Yongqin Xian, Thomas Hofmann, and Federico Tombari. Lime: localized image editing via attention regularization in diffusion models. arXiv preprint arXiv:2312.09256, 2023. 2, 4", + "[47] Jaskirat Singh, Jianming Zhang, Qing Liu, Cameron Smith, Zhe Lin, and Liang Zheng. Smartmask: Context aware high-fidelity mask generation for fine-grained object insertion and layout control. In CVPR, 2024. 2, 3", + "[48] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 1", + "[49] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 1", + "[50] Kcolors Team. Kcolors: Effective training of diffusion model for photorealistic text-to-image synthesis. arXiv preprint, 2024. 3" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In CVPR, 2023. 3", + "[52] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In CVPR, 2024. 3, 4", + "[53] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In CVPR, 2023. 2, 3", + "[54] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. NeurlPS, 2023. 3, 4", + "[55] Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. In CVPR, 2024. 3", + "[56] Mingyang Yi, Aoxue Li, Yi Xin, and Zhenguo Li. Towards understanding the working mechanism of text-to-image diffusion model. arXiv preprint arXiv:2405.15330, 2024. 2, 4", + "[57] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. 3", + "[58] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. NeurIPS, 2024. 2, 3, 5, 7, 1", + "[59] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2", + "[60] Shu Zhang, Xinyi Yang, Yihao Feng, Can Qin, Chia-Chih Chen, Ning Yu, Zeyuan Chen, Huan Wang, Silvio Savarese, Stefano Ermon, et al. Hive: Harnessing human feedback for instructional visual editing. In CVPR, 2024. 3, 7", + "[61] Wentian Zhang, Haozhe Liu, Jinheng Xie, Francesco Faccio, Mike Zheng Shou, and Jurgen Schmidhuber. Cross-attention makes inference cumbersome in text-to-image diffusion models. arXiv preprint arXiv:2404.02747, 2024. 2, 4", + "[62] Haozhe Zhao, Xiaojian Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. Ultraedit: Instruction-based fine-grained image editing at scale. arXiv preprint arXiv:2407.05282, 2024. 2, 3" + ], + "bbox": [ + 514, + 92, + 906, + 686 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19215", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SuperEdit_ Rectifying and Facilitating Supervision for Instruction-Based Image Editing/395fe9d1-b3f1-4c45-afff-e5eaba8e26ad_model.json b/2025/SuperEdit_ Rectifying and Facilitating Supervision for Instruction-Based Image Editing/395fe9d1-b3f1-4c45-afff-e5eaba8e26ad_model.json new file mode 100644 index 0000000000000000000000000000000000000000..586b709d1d3df2ef36d688e0d8f8263dea8f9fb0 --- /dev/null +++ b/2025/SuperEdit_ Rectifying and Facilitating Supervision for Instruction-Based Image Editing/395fe9d1-b3f1-4c45-afff-e5eaba8e26ad_model.json @@ -0,0 +1,2288 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.131, + 0.908, + 0.153 + ], + "angle": 0, + "content": "SuperEdit: Rectifying and Facilitating Supervision for Instruction-Based Image Editing" + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.162, + 0.385, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.387, + 0.163, + 0.451, + 0.179 + ], + "angle": 0, + "content": "Website" + }, + { + "type": "image", + "bbox": [ + 0.475, + 0.162, + 0.498, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.163, + 0.542, + 0.179 + ], + "angle": 0, + "content": "Code" + }, + { + "type": "image", + "bbox": [ + 0.569, + 0.163, + 0.596, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.597, + 0.164, + 0.638, + 0.179 + ], + "angle": 0, + "content": "Data" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.19, + 0.898, + 0.226 + ], + "angle": 0, + "content": "Ming Li\\(^{1,2,\\ddagger}\\), Xin Gu\\(^{1}\\), Fan Chen\\(^{1}\\), Xiaoying Xing\\(^{1}\\), Longyin Wen\\(^{1}\\), Chen Chen\\(^{2}\\), Sijie Zhu\\(^{1,*}\\) \n\\(^{1}\\)ByteDance Intelligent Creation Center for Research in Computer Vision, University of Central Florida" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.236, + 0.38, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.236, + 0.666, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.316, + 0.379, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.379, + 0.316, + 0.665, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.14, + 0.399, + 0.335, + 0.406 + ], + "angle": 0, + "content": "Global: Transform the scene into a snowy winter scene" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.406, + 0.379, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.439, + 0.399, + 0.6, + 0.406 + ], + "angle": 0, + "content": "Replace: Replace the bow tie with a butterfly." + }, + { + "type": "image", + "bbox": [ + 0.379, + 0.406, + 0.664, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.483, + 0.592, + 0.49 + ], + "angle": 0, + "content": "Style: Convert the image to a watercolor style Remove: Remove some clouds in the sky" + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.491, + 0.536, + 0.501 + ], + "angle": 0, + "content": "(a) Examples of Our Method on Real & High-resolution Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.722, + 0.237, + 0.853, + 0.245 + ], + "angle": 0, + "content": "Comparison with Existing Methods" + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.245, + 0.905, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.685, + 0.388, + 0.89, + 0.397 + ], + "angle": 0, + "content": "(b) Comparison with GPT-4o Evaluation" + }, + { + "type": "image", + "bbox": [ + 0.667, + 0.408, + 0.905, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.684, + 0.49, + 0.887, + 0.5 + ], + "angle": 0, + "content": "(c) Comparison with Human Evaluation" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.503, + 0.908, + 0.545 + ], + "angle": 0, + "content": "Figure 1. (a) Our editing method works well with real and high-resolution images, handling various free-form edits (left) and local edits (right); (b) Compared to the current state-of-the-art SmartEdit, our method achieves a \\(9.19\\%\\) performance improvement with \\(30\\times\\) less training data and \\(13\\times\\) fewer model parameters; (c) Our method achieves better overall scores on the human evaluation results, indicating more precise editing capabilities." + }, + { + "type": "title", + "bbox": [ + 0.25, + 0.561, + 0.325, + 0.576 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.592, + 0.486, + 0.865 + ], + "angle": 0, + "content": "Due to the challenges of manually collecting accurate editing data, existing datasets are typically constructed using various automated methods, leading to noisy supervision signals caused by the mismatch between editing instructions and original-edited image pairs. Recent efforts attempt to improve editing models through generating higher-quality edited images, pre-training on recognition tasks, or introducing vision-language models (VLMs) but fail to resolve this fundamental issue. In this paper, we offer a novel solution by constructing more effective editing instructions for given image pairs. This includes rectifying the editing instructions to better align with the original-edited image pairs and using contrastive editing instructions to further enhance their effectiveness. Specifically, we find that editing models exhibit specific generation attributes at different inference steps, independent of the text. Based on these prior attributes, we define a unified guide for VLMs to rectify editing instructions. However, there are some challenging editing scenarios that cannot be resolved solely with rectified instructions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.562, + 0.909, + 0.76 + ], + "angle": 0, + "content": "To this end, we further construct contrastive supervision signals with positive and negative instructions and introduce them into the model training using triplet loss, thereby further facilitating supervision effectiveness. Our method does not require the VLM modules or pre-training tasks used in previous work, offering a more direct and efficient way to provide better supervision signals, and providing a novel, simple, and effective solution for instruction-based image editing. Results on multiple benchmarks demonstrate that our method significantly outperforms existing approaches. Compared with previous SOTA SmartEdit, we achieve \\(9.19\\%\\) improvements on the Real-Edit benchmark with \\(30\\times\\) less training data and \\(13\\times\\) smaller model size. All data and models are open-sourced on Github for future research." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.785, + 0.638, + 0.8 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In recent years, significant progress has been made in text-to-image (T2I) generation [10, 37, 40-42] due to the development of diffusion models [9, 21, 48, 49]. These T2I diffusion models can generate images that align with natural language descriptions while satisfying human perception and preferences. Consequently, numerous image editing" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.876, + 0.366, + 0.888 + ], + "angle": 0, + "content": "* Corresponding author, sijiezhu@bytedance.com" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.888, + 0.483, + 0.901 + ], + "angle": 0, + "content": "\\(\\ddagger\\) This work was done during the first author's internship at ByteDance USA" + }, + { + "type": "list", + "bbox": [ + 0.114, + 0.876, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19206" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.182 + ], + "angle": 0, + "content": "methods [4, 6, 19, 31] based on these models have been proposed to achieve various editing effects. Instruction-based methods [4, 12, 24] have become increasingly popular as they allow users to conveniently and easily modify images using language instructions without the need to provide masks, as required by mask-based methods [18, 28, 47, 53]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.186, + 0.487, + 0.533 + ], + "angle": 0, + "content": "The training of instruction-based editing models requires the original-edited image pairs and corresponding editing instruction, making it difficult to manually create or collect a large amount of relevant data [58]. To address the issue of scarce training data, existing efforts [13, 19, 25, 62] have attempted to develop various automated pipelines to synthesize large datasets. Specifically, most methods first use large language models (LLMs) to modify the text descriptions of original images. The original images and modified texts are then input into various pre-trained diffusion models to automatically generate edited images. However, current text-to-image diffusion models struggle to fully correspond to input text prompts [15, 59]. This often changes parts of the original images that do not require editing, leading to misaligned editing instructions and original-edited image pairs, thus resulting in noisy supervision signals. To mitigate the potential issues of noisy supervision in image editing models, existing work has attempted to introduce additional recognition pre-training tasks for U-Net [43] such as semantic segmentation [14, 45], or replace CLIP [38] text encoder with vision-language models (VLMs) [12, 24] to better understand editing instructions from noisy supervision signals. However, these methods not only introduce significant computational overhead but also overlook the issue of noisy supervision signals." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.536, + 0.484, + 0.718 + ], + "angle": 0, + "content": "In this paper, we focus on addressing the fundamental challenge by introducing more effective editing instructions, as demonstrated in Fig. 2. Our data-oriented method explores a different research question: how much performance improvement can be achieved solely by focusing on supervision signal quality and optimization in image editing? Surprisingly, SuperEdit outperforms existing methods in both GPT-4o and human evaluations, despite using less data and requiring no additional modules or pretraining as shown in Fig.1. This demonstrates that high-quality supervision signals can significantly compensate for architectural simplicity, achieving results comparable to or better than methods with more complex requirements." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.487, + 0.902 + ], + "angle": 0, + "content": "Specifically, to enhance the effectiveness of supervision signals for instruction-based image editing methods, we propose using VLMs to rectify editing instructions, creating better-aligned instructions for the original-edited image pairs. However, determining which VLM to use for this task and how to establish a unified rectification method for various editing instructions remain unexplored problems. To address this, we first analyze the ability of different VLMs to understand the differences between original and edited images, showing that GPT-4o [1] is the most capable of rectifying editing instructions. Additionally, we observe that both editing models and text-to-image diffusion models share a similar prior, as shown in Fig. 4: different inference" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.088, + 0.705, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.285, + 0.706, + 0.296 + ], + "angle": 0, + "content": "(c) InstructDiffusion; Emu-Edit" + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.089, + 0.882, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.766, + 0.285, + 0.817, + 0.296 + ], + "angle": 0, + "content": "(d) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.298, + 0.907, + 0.368 + ], + "angle": 0, + "content": "Figure 2. Unlike existing efforts that attempt to (a) scale up edited images with noisy supervision [4, 25], (b) introduce massive VLMs into editing model architecture [12, 24], and (c) perform additional pre-training tasks [14, 45], (d) we focus on improving the effectiveness of supervision signals, which is the fundamental issue of image editing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.907, + 0.447 + ], + "angle": 0, + "content": "stages correspond to the generation of different image attributes, independent of the input text prompt [2, 17, 19, 36, 46, 56, 61] or editing instructions. Inspired by this, we guide VLMs based on these attributes to establish a unified rectification guideline for various editing instructions, as demonstrated in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.449, + 0.909, + 0.674 + ], + "angle": 0, + "content": "When training with only rectified editing instructions, we find that the editing model can better understand the editing commands but still faces challenges in handling complex scenarios. For example, when the original image contains multiple objects, the edit model struggles to perform an accurate editing function if the instructions modify only one of these objects. Additionally, inherent issues present in pre-trained text-to-image diffusion models [15, 22-24], such as difficulty in understanding quantity, position, or object relationships, persist in the editing models. To address these issues, we propose using contrastive supervision signals to further optimize the editing model. Specifically, we first construct incorrect editing instructions based on the rectified instructions to generate positive and negative samples. We then introduce a triplet loss to guide the model, thereby enhancing the effectiveness of supervision, as shown in Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.538, + 0.676, + 0.899, + 0.689 + ], + "angle": 0, + "content": "In summary, our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.69, + 0.908, + 0.75 + ], + "angle": 0, + "content": "- New Insight: We aim to address the noisy supervision problem that arises from the misalignment between editing instructions and original-edited image pairs, which is a fundamental issue overlooked by previous work, as shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.751, + 0.908, + 0.796 + ], + "angle": 0, + "content": "- Rectifying Supervision: We leverage diffusion generation priors to guide the vision-language model to generate better-aligned editing instructions for original-edited image pairs." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.797, + 0.908, + 0.841 + ], + "angle": 0, + "content": "- Facilitating Supervision: We introduce contrastive supervision using triplet loss, enabling the editing model to learn from both positive and negative editing instructions." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.841, + 0.909, + 0.902 + ], + "angle": 0, + "content": "- Promising Results: We achieve significant improvements on multiple benchmarks without additional pre-training or VLM. Compared to SmartEdit [24], we achieved a \\(9.19\\%\\) improvement while reducing \\(30 \\times\\) data and \\(13 \\times\\) model parameters." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.69, + 0.909, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19207" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.09, + 0.226, + 0.105 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.115, + 0.411, + 0.132 + ], + "angle": 0, + "content": "2.1. Image Editing with Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.138, + 0.484, + 0.228 + ], + "angle": 0, + "content": "Building on advancements in text-to-image (T2I) diffusion models [10, 37, 40-42, 44], recent research has explored them for image editing [4, 19]. Training-free methods [5, 19, 29, 31, 34, 51] typically achieve this by adjusting attentions in pre-trained T2I models, but have limited performance and generalization capabilities on various editing tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.229, + 0.484, + 0.335 + ], + "angle": 0, + "content": "Training-based methods address these limitations with specialized editing models, which can be categorized into mask-based and instruction-based approaches. Mask-based methods [6, 18, 28, 47, 53, 57] enable fine-grained local edits with user-provided or predicted masks and corresponding text descriptions. However, it struggles with global image editing and is constrained by the lack of mask-based editing data [24]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.336, + 0.484, + 0.502 + ], + "angle": 0, + "content": "Instruction-based methods directly accept textual commands, such as \"add a dog\", offering better editing flexibility and generalization. InstructPix2Pix [4] pioneered this paradigm by generating instruction-based editing data and modifying the conditions of T2I diffusion models. Building on this framework, subsequent work introduces vision-language models [12, 24, 33] or additional pre-training tasks for the denoising U-Net [14, 24, 43, 45] to enhance the understanding and reasoning of input conditions. However, these methods not only introduce substantial computational overhead but also overlook the fundamental noisy supervision issue." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.512, + 0.483, + 0.529 + ], + "angle": 0, + "content": "2.2. Generating and Improving Editing Supervision" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.484, + 0.732 + ], + "angle": 0, + "content": "Due to the difficulty of scaling instruction-based image editing data through manual collection, existing efforts [4, 13, 25, 58, 62] aim to automatically modify text descriptions of original images and generate edited images with T2I diffusion models. However, this approach often produces synthesized images that do not align with the editing instructions, as shown in Figure 3, resulting in noisy editing supervision signals [58, 62]. To address this, MagicBrush [58] manually filters out incorrect editing data, but it is hard to scale. Unlike existing methods focusing on edited image quality, we leverage diffusion prior and vision-language model (i.e., GPT-40 [1]) to create better-aligned instructions with original-edited image pairs, providing more accurate supervision." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.364, + 0.758 + ], + "angle": 0, + "content": "2.3. Alignment of Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The success of alignment training in large language models (LLMs) [26, 32, 39] has been applied to diffusion models for better image generation. This is achieved by maximizing reward scores [8, 27, 54] or the generation probability of the winner image in a pair [11, 52, 55]. In image editing, HIVE [60] and MultiReward [16] attempt to incorporate reward information into the text condition to align the editing model. In contrast, we guide the editing model by rectifying and constructing contrastive editing instructions, achieving more effective alignment." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.6, + 0.105 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.907, + 0.237 + ], + "angle": 0, + "content": "In this section, we first introduce the most general image editing framework in Sec. 3.1. Then, we explain how to use diffusion priors to rectify editing instructions with the multimodal model (i.e., GPT-4o) in Sec. 3.2, thereby enhancing the accuracy of supervision signals. Finally, we describe how to construct contrastive supervision with both correct and incorrect editing instructions and integrate it into the editing model training using triplet loss in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.247, + 0.895, + 0.264 + ], + "angle": 0, + "content": "3.1. Instruction-based Image Editing Framework" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.269, + 0.907, + 0.376 + ], + "angle": 0, + "content": "InstructPix2Pix [4] pioneered instruction-based image editing, performing editing tasks by simultaneously taking the original image \\( C^I \\) and editing instructions \\( C^T \\) as input conditions to generate the edited image \\( x \\) from random noise \\( \\epsilon \\). Following the definition of DDPM [21], we randomly sample a timestep \\( t \\in T \\) during training, and then add corresponding noise \\( \\epsilon_t \\) to the edited image \\( x \\):" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.386, + 0.907, + 0.403 + ], + "angle": 0, + "content": "\\[\nx _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t}, \\quad \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, I), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.415, + 0.906, + 0.503 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) is a noise map sampled from a Gaussian distribution, and \\(\\bar{\\alpha}_t\\coloneqq \\prod_{s = 0}^t\\alpha_s\\), \\(\\alpha_{t} = 1 - \\beta_{t}\\) is a differentiable function of timestep \\(t\\), which is determined by the denoising sampler such as DDPM [21]. Then the training objective of the editing model \\(\\epsilon_{\\theta}\\) is predicting the added noise at timestep \\(t\\), which can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.51, + 0.907, + 0.536 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t r a i n}} = \\mathbb {E} _ {x _ {t}, t, C ^ {I}, C ^ {T}, \\epsilon} \\left[ \\left\\| \\epsilon_ {\\theta} \\left(\\operatorname {c o n c a t} \\left(x _ {t}, C ^ {I}\\right), t, C ^ {T}\\right) - \\epsilon_ {t} \\right\\| _ {2} ^ {2} \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.546, + 0.907, + 0.575 + ], + "angle": 0, + "content": "where concat refers to concatenating the image latents of noised edited image \\( x_{t} \\) and original image \\( c_{I} \\) in the channel dimension." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.585, + 0.89, + 0.601 + ], + "angle": 0, + "content": "3.2. Rectifying Supervision with Diffusion Priors" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.909, + 0.91 + ], + "angle": 0, + "content": "As shown in Fig. 3, existing image editing datasets [4, 13, 58] typically use only Steps 1 and 2: LLMs construct editing prompts and captions, and then text-to-image diffusion models synthesize edited images. However, diffusion models often fail to accurately follow prompts while maintaining image layout, creating mismatches between original-edited pairs and editing instructions, resulting in inaccurate supervision. While better supervision signals for text-to-image diffusion models are common in image generation [3, 50], this approach remains underexplored in image editing due to two challenges: (1) VLMs trained on single-image data struggle with multi-image inputs, and (2) editing instructions vary widely, making unified rectification guidelines difficult. To address these issues, we: (1) analyzed different VLMs' capabilities with multi-image inputs, finding GPT-4o most effective, and (2) discovered that timestep-specific roles in image generation also apply to editing, providing a foundation for a unified rectification method across various instructions (Fig. 3 and 4). Due to page limitations, our VLM analysis is in the Supplementary Material, while this section focuses on Diffusion Prior and Editing Instruction Rectification." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19208" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.402, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.089, + 0.728, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.097, + 0.9, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.278, + 0.907, + 0.335 + ], + "angle": 0, + "content": "Figure 3. (a) Existing work primarily uses LLMs and diffusion models to automatically generate edited images. However, current diffusion models often fail to accurately follow text prompts while maintaining the input image's layout, resulting in mismatches between the original-edited image pairs and the editing instructions. (b) We perform instruction rectification (Step 3) based on the images constructed in Steps 1 and 2. We show VLMs can understand the differences between the images, enabling them to rectify editing instructions to be better aligned with original-edited image pairs." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.359, + 0.479, + 0.662 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.664, + 0.486, + 0.763 + ], + "angle": 0, + "content": "Figure 4. We show that the editing model follows consistent generation attributes at different sampling stages, independent of the editing instructions. The early, middle, and late sampling stages correspond to global, local, and detail changes, respectively, while style changes occur at all stages. All the generated images here are DDIM 30-step sampled final images. The orange progress bar and the grid progress bar represent the sampling stages with and without the editing instructions, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Diffusion Generation Priors. Previous work has shown that different timesteps play distinct roles in image generation for text-to-image diffusion models, regardless of the text prompt [2, 17, 19, 36, 46, 56, 61]. We find that this phenomenon also exists in instruction-based editing models and present examples based on pre-trained InstructPix2Pix [4], as shown in Fig. 4. Specifically, diffusion models focus on global layout in the early stages, local object attributes in the mid stages, and image details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.36, + 0.909, + 0.422 + ], + "angle": 0, + "content": "in the late stages of sampling. This finding inspires us to guide VLMs based on these four generation attributes, establishing a unified rectification method for various editing instructions. We provide more analysis and details in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.909, + 0.609 + ], + "angle": 0, + "content": "Editing Instruction Rectification. As demonstrated in Fig. 3, we extend the existing editing data generation pipeline by introducing our instruction rectification (Step 3). This process relies on the original edited image pairs obtained through Steps 1 and 2 from previous work. Specifically, we input original-edited image pairs into the vision-language model (i.e., GPT-4o) and instruct it to describe the changes in the edited image compared to the original image according to the above diffusion prior generation attributes. Finally, we use VLM to summarize the instructions and ensure that its length is less than the maximum length of CLIP text encoder, which is 77 tokens." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.621, + 0.907, + 0.652 + ], + "angle": 0, + "content": "3.3. Facilitating Supervision with Contrastive Instructions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Although using rectified editing instructions can significantly improve performance across various editing tasks, we find that editing models still struggle with closely related text instructions. For example, \"add a cat on the left side of the image\" and \"add two cats on the right side of the image\" might produce the same edited image. This indicates that inherent biases in pre-trained text-to-image diffusion models [15, 22], such as difficulties in understanding quantity, position, and spatial relationships, persist in editing models. More importantly, our experiments show that training models with rectified editing instructions does not resolve these challenges. To further facilitate supervision signal effectiveness, we drew on successful alignment experiences from large language models [1, 32, 39] and text-to-image diffusion models [7, 52, 54]: constructing positive and negative sample pairs and guiding the model to assign a higher generation probability to positive samples compared to negative ones." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19209" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.415, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.093, + 0.907, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.283, + 0.908, + 0.355 + ], + "angle": 0, + "content": "Figure 5. (a) Based on the rectified editing instruction and original-edited image pair, we utilize the Vision-Language Models (VLM) to generate various image-related wrong instructions. These involve random substitutions of quantities, spatial locations, and objects within the rectified editing instructions according to the original-edited images context; (b) During each training iteration, we randomly select one wrong instruction \\( c_{\\text{neg}}^T \\) and input it along with the rectified instruction \\( c_{pos}^T \\) into the editing model to obtain predicted noises. The goal is to make the rectified instruction's predicted noise \\( \\epsilon_{pos} \\) closer to the sampled training diffusion noise \\( \\epsilon \\), while ensuring the noise from incorrect instructions \\( \\epsilon_{\\text{neg}} \\) is farther. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.373, + 0.485, + 0.691 + ], + "angle": 0, + "content": "Constructing Contrastive Instructions. Unlike the standard alignment process for large language models or text-to-image diffusion models, it is challenging to generate different editing results from the same instruction to create positive and negative sample pairs for image editing tasks. To address this, we construct positive and negative editing instructions for alignment, thereby generating relatively positive and negative edited images. As shown in Fig. 5 (a), we use the original image, edited image, and rectified editing instruction as input. The VLM (i.e., GPT-4o) is used to modify attributes in the rectified editing instruction, such as quantity, spatial relationships, and object types, to create different wrong instructions. Here, we require VLM to modify only a single attribute from the rectified editing instruction in each wrong instruction, keeping most of the editing text unchanged. Since only a few words are replaced between the rectified instruction and the wrong instruction, the text embeddings produced by the CLIP text encoder that serve as input to the denoising model will also be similar. This ensures the task's learning difficulty, helping the model understand how subtle differences between the two editing instructions result in significantly different editing results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Facilitating Editing Models with Contrastive Instructions. Our key insight is that enhancing the effectiveness of supervision signals can improve various editing tasks without introducing additional model architectures or pre-training tasks. Therefore, we adhere strictly to the InstructPix2Pix [4] model architecture and training pipeline. To be specific, the inputs including the original image \\( c^I \\), edited image \\( x \\), the rectified instruction \\( c_{pos}^T \\), and wrong editing instruction \\( c_{neg}^T \\). During training, we will add a sampled timestep \\( t \\in T \\) to obtain the noised edited image \\( x_t \\) with Equation 1. Both the rectified and wrong editing instructions are fed into the denoising model to predict the final noises \\( \\epsilon_{pos} \\) and \\( \\epsilon_{neg} \\), which are then used to" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.373, + 0.848, + 0.388 + ], + "angle": 0, + "content": "construct positive and negative samples, respectively:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.397, + 0.907, + 0.416 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {p o s} = \\epsilon_ {\\theta} \\left(\\operatorname {c o n c a t} \\left(x _ {t}, c ^ {I}\\right), t, c _ {p o s} ^ {T}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.419, + 0.907, + 0.439 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {n e g} = \\epsilon_ {\\theta} \\big (\\operatorname {c o n c a t} \\left(x _ {t}, c ^ {I}\\right), t, c _ {n e g} ^ {T} \\big). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.447, + 0.907, + 0.524 + ], + "angle": 0, + "content": "After constructing the positive and negative sample pairs, we aim for the noise predicted by the positive editing instruction \\(\\epsilon_{pos}\\) to be closer to the true noise \\(\\epsilon_t\\) sampled during training, compared to the noise predicted by the wrong editing instruction \\(\\epsilon_{neg}\\). This goal can be achieved through a triplet loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.534, + 0.907, + 0.551 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t r i p l e t}} = \\max \\left\\{d \\left(\\epsilon_ {t}, \\epsilon_ {\\text {p o s}}\\right) - d \\left(\\epsilon_ {t}, \\epsilon_ {\\text {n e g}}\\right) + m, 0 \\right\\}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.561, + 0.908, + 0.608 + ], + "angle": 0, + "content": "where \\(d(x,y) = \\|\\mathbf{x} - \\mathbf{y}\\|_2^2\\) and margin \\(m\\) is a hyper-parameter. The final training loss is the combination of the original diffusion training loss and the triplet loss:" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.619, + 0.907, + 0.636 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {t r a i n}} + \\lambda \\cdot \\mathcal {L} _ {\\text {t r i p l e t}}, \\text {w h e r e} \\mathcal {L} _ {\\text {t r a i n}} = d \\left(\\epsilon_ {t}, \\epsilon_ {\\text {p o s}}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.908, + 0.691 + ], + "angle": 0, + "content": "Please note that the contrastive supervision signals are only used during the training phase. During inference, the editing model only requires one single input editing instruction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.703, + 0.634, + 0.72 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.728, + 0.809, + 0.743 + ], + "angle": 0, + "content": "4.1. Data Collection and Construction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To build a diverse dataset with various types of editing instructions, we need original and edited images from different data domains, as well as a wide variety of editing instructions. To achieve this, we sampled data from different public editing datasets to construct rectified and contrastive supervision signals. Specifically, we extracted 10,177, 8,807, and 21,016 editing pairs from InstructPix2Pix [4], MagicBrush [58], and Seed-Data-Edit [13], respectively, resulting in a total of 40,000 training samples. During extraction, we strive to ensure that the data for different types of editing tasks is as balanced as possible." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19210" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.088, + 0.881, + 0.852 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.855, + 0.907, + 0.898 + ], + "angle": 0, + "content": "Figure 6. Visual comparison with existing methods and the corresponding human-aligned GPT-4o evaluation scores (Following, Preserving, Quality Scores from left to right). We achieve better results while preserving the layout, quality, and details of the original image. Please note that we do not claim that our editing results are flawless. We provide more visual comparison results in the supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19211" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.135, + 0.082, + 0.861, + 0.243 + ], + "angle": 0, + "content": "
MethodExtra ModulePretrain TasksEdit DataModel SizeFollowing ↑Preserving ↑Quality ↑Overall ↑
AccScoreAccScoreAccScoreAccScore
KOSMOS-G [33]✓✓9.0M1.9B51%2.829%1.4327%3.2029.0%2.48
MGIE [12]✓✓1.0M8.1B40%2.4345%2.7938%3.3541.0%2.86
SmartEdit [24]✓✓1.2M14.1B64%3.5066%3.7045%3.5658.3%3.59
MultiReward [16]✓✓320K1.2B63%3.3958%3.4354%3.8058.3%3.54
InstructDiffusion [14]X✓860K1.1B52%2.8754%3.1745%3.5850.3%3.21
InstructPix2Pix [4]XX300K1.1B52%2.9453%3.3150%3.6951.7%3.31
MagicBrush [58]XX310K1.1B51%2.9070%3.8550%3.6757.0%3.47
HIVE [60]XX1.1M1.1B54%2.9356%3.3653%3.7254.3%3.34
HQ-Edit [25]XX500K1.1B51%2.8416%1.6354%3.8440.3%2.77
SuperEdit (Ours)XX40K1.1B67%3.5977%4.1465%4.0169.7%3.91
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.247, + 0.907, + 0.292 + ], + "angle": 0, + "content": "Table 1. Comparison with instruction-based image editing methods on Real-Edit benchmark [16]. Compared to existing work, our method achieves state-of-the-art performance across all metrics using a small amount of high-quality editing data without introducing additional models or pre-training tasks. Please note that the scores range from 0 to 5. \\(\\uparrow\\) denotes a higher result is better. All baseline results are cited from the MultiReward [16] paper." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.302, + 0.361, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.189, + 0.38, + 0.267, + 0.389 + ], + "angle": 0, + "content": "(a) Following Score" + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.302, + 0.631, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.456, + 0.38, + 0.538, + 0.389 + ], + "angle": 0, + "content": "(b) Preserving Score" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.302, + 0.9, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.729, + 0.38, + 0.797, + 0.389 + ], + "angle": 0, + "content": "(c) Quality Score" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.391, + 0.907, + 0.434 + ], + "angle": 0, + "content": "Figure 7. Human evaluation on three evaluation criteria for image editing effects. (a) Following: whether the edited image adheres to the editing instructions; (b) Preserving: whether the image structure outside of the editing instructions has been preserved; (c) Quality: whether the overall quality/aesthetics of the edited image has been degraded compared to the input image. Our SuperEdit achieves the best results on all of these metrics." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.449, + 0.48, + 0.499 + ], + "angle": 0, + "content": "
Following ↑Preserving ↑Quality ↑Overall ↑
InstructPix2Pix [4]2.412.622.442.49
SmartEdit-13B [24]3.093.062.632.93
SuperEdit3.18+1.80%3.86+16.00%3.37+14.80%3.47+10.80%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.501, + 0.483, + 0.545 + ], + "angle": 0, + "content": "Table 2. Human evaluation results on Real-Edit [16] benchmark. All the human-evaluated scores range from 0 to 5. Overall represents the average score of Following, Preserving, and Quality scores." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.557, + 0.484, + 0.664 + ], + "angle": 0, + "content": "We then applied our proposed methods in Sec. 3 to rectify and construct contrastive editing instructions for these training samples. Since the MagicBrush data has been manually verified, we skip the rectification step for this dataset and directly construct contrastive supervision based on the original editing instructions. For Seed-Data-Edit dataset, we only sample images from the first part of data without human editing instructions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.68, + 0.299, + 0.697 + ], + "angle": 0, + "content": "4.2. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Evaluation Benchmarks and Metrics. To more accurately assess the effectiveness of various editing models, we conducted assessments on the Real-Edit benchmark [16], which is a human-aligned evaluation benchmark with GPT-4o scoring. Specifically, MultiReward [16] uses high-resolution images from the Unsplash community as a test dataset and combines them with GPT-4o [1] to create an automated evaluation method for single-turn editing. It assesses edited images in terms of accuracy (\\%) and scores (from 0 to 5), evaluating whether they adhere to the editing instructions (Following), whether the image structure outside of the editing instructions has been preserved (Preserving), and whether the overall quality/aesthetics of the edited image has been degraded compared to the original one (Quality)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.451, + 0.719, + 0.468 + ], + "angle": 0, + "content": "4.3. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.474, + 0.907, + 0.686 + ], + "angle": 0, + "content": "Comparison on Real-Edit Benchmark. In Tab. 1, we present the quantitative results of editing effectiveness on the Real-Edit benchmark [16]. Without introducing additional parameters or pre-training stages, our method achieves the best results in the three GPT-4o automated evaluation metrics: Following, Preserving, and Quality, each of which includes percentage accuracy (Acc) and scores (from 0 to 5). For example, compared to SmartEdit [24], which introduces an additional 13B vision-language model (i.e., LLaVA [30]) to the 1.1B InstructPix2Pix [4] framework, we achieved improvements of \\(11.4\\%\\) Overall Score. This suggests that given accurate and effective supervision signals, the trained editing model can understand and successfully execute the editing instructions, without the need for additional vision-language models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.909, + 0.84 + ], + "angle": 0, + "content": "It is worth noting that unlike existing image editing methods, which often show improvement in a single metric while others remain unchanged or worsen, our method achieves comprehensive and significant advancements across all three metrics. This indicates that improving the effectiveness of supervision signals can accurately execute editing instructions while reducing disruption to other non-edited parts of the image, and preserving the quality and aesthetics of the original images. Specifically, we surpassed the current best methods by \\(3\\%\\), \\(7\\%\\), and \\(11\\%\\) Acc results in Following, Preserving, and Quality, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Human Evaluation We also conduct a comprehensive human evaluation on Real-Edit benchmarks [16] in Tab. 2 and Fig. 7. The assessment involved 15 experienced evaluators" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19212" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.319 + ], + "angle": 0, + "content": "who rated edited images based on three critical metrics: instruction faithfulness (Following), preservation of irrelevant content (Preserving), and visual quality (Quality). The results of this manual evaluation demonstrate strong consistency with the GPT-4o scoring results shown in Tab. 1, with Cohen's Kappa=0.89, Krippendorff's Alpha=0.87. This high alignment thoroughly validates that our proposed SuperEdit significantly outperforms existing methods across all evaluation criteria. Specifically, our SuperEdit surpasses the previous state-of-the-art method SmartEdit [24] by \\(1.8\\%\\), \\(16\\%\\), \\(14.8\\%\\), and \\(10.8\\%\\) on Following, Preserving, Quality, and Overall scores, respectively. These substantial improvements not only confirm the effectiveness of our approach but also establish SuperEdit as a new benchmark in instruction-guided image editing, achieving superior performance while requiring significantly less training data and cost." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.328, + 0.487, + 0.707 + ], + "angle": 0, + "content": "Visual Comparison with State-of-the-art Methods. We show the visual comparison with existing image editing methods in Fig. 6. Compared to existing instruction-based editing methods, our approach not only better understands and executes editing instructions but also preserves the original image's layout and quality more effectively, thereby significantly outperforming previous methods. For example, with the instruction \"Replace the tiger with a lion, maintaining the same position in the water\" our SuperEdit method achieved superior results (4.8/4.8/4.8) compared to SmartEdit (4.8/4.8/2.5) and other methods. Additionally, our method improves the model's comprehension of editing instructions. For the instruction \"Change the background to a sandy beach with the ocean in the distance\" our method received perfect scores (4.8/5.0/5.0) while SmartEdit only achieved (5.0/4.8/4.0). Similarly, for style transformation instructions like \"Change the image style to look like an impressionist painting style\" SuperEdit significantly outperformed SmartEdit with scores of (4.8/4.8/4.8) versus (1.0/4.8/4.8), demonstrating our method's superior ability to handle complex artistic transformations. Even more impressively, for scene transformation tasks like \"Transform the entire scene to a winter setting with snow covering the houses, trees, and boat\", our SuperEdit achieved (5.0/4.8/4.8) while SmartEdit only obtained (2.0/4.5/4.5). We provide more visual comparisons with other instruction-based image editing methods in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.245, + 0.729 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Ablation on the Rectified and Contrastive Instructions. Considering that the Real-Edit [16] benchmark is evaluated by GPT-4o [1], and its evaluation results closely align with human ratings [16], we choose this benchmark to conduct ablation experiments in Tab. 3. Compared to the original 40K training data, our training data with rectified editing instructions significantly improves all the performance of the editing model. Specifically, our approach improves scores by 0.95, 0.79, and 0.11, and accuracy by \\(21\\%\\), \\(22\\%\\), and \\(4\\%\\) in these three metrics, respectively. In addition, editing performance is further enhanced by incorporating contrastive supervision signals. Compared to using only" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.09, + 0.905, + 0.153 + ], + "angle": 0, + "content": "
Rectified InstructionContrastive InstructionsFollowing↑Preserving↑Quality↑
AccScoreAccScoreAccScore
XX41%2.4553%3.2761%3.90
✓X62%3.4075%4.0665%4.01
✓✓67%3.5977%4.1465%4.01
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.158, + 0.907, + 0.186 + ], + "angle": 0, + "content": "Table 3. Ablation study on our methods. Both rectified and contrastive editing instructions achieved improvements across all metrics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.195, + 0.909, + 0.288 + ], + "angle": 0, + "content": "rectified editing instructions, the introduction of contrastive supervision signals improves the following and preserving scores by 0.19 and 0.08, and accuracy by \\(5\\%\\) and \\(2\\%\\), while maintaining the quality accuracy and score. In summary, both the introduction of rectified editing instructions and contrastive editing instructions improve the overall performance of the editing model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.296, + 0.909, + 0.493 + ], + "angle": 0, + "content": "Ablation on Data Scaling. We investigated the impact of training data volume on model performance by experimenting with datasets ranging from 5k to 40k samples. Tab. 4 shows consistent improvements across all metrics as training data increases. With just 5k samples, our model achieves reasonable performance (54.7% accuracy, 3.42 overall score), but scaling to 40k samples yields substantial gains (69.7% accuracy, 3.91 overall score). The most significant improvements appear in the Preserving and Quality metrics, with 10% and 15%, respectively. This upward trend across all data points demonstrates that SuperEdit effectively leverages additional training examples without performance saturation, suggesting potential for further gains with larger datasets." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.498, + 0.903, + 0.576 + ], + "angle": 0, + "content": "
Data SizeFollowing ↑Preserving ↑Quality ↑Overall ↑
AccScoreAccScoreAccScoreAccScore
5k49%2.8760%3.7155%3.6954.7%3.42
10k57%3.2671%3.7658%3.8762.0%3.63
20k64%3.4072%4.0263%3.9466.3%3.79
40k67%3.5977%4.1465%4.0169.7%3.91
" + }, + { + "type": "table_caption", + "bbox": [ + 0.526, + 0.576, + 0.892, + 0.589 + ], + "angle": 0, + "content": "Table 4. Ablation study on data scaling results on Real-Edit [16]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.604, + 0.628, + 0.62 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In this paper, we re-examine image editing models from the perspective of enhancing supervision signals, finding that existing models have not adequately addressed this challenge, resulting in suboptimal performance. We introduce a unified editing instruction rectification guideline based on diffusion priors that better aligns instructions with original-edited image pairs, thereby enhancing supervision effectiveness. We also construct contrastive editing instructions allowing models to learn from both positive and negative examples. Our data-oriented approach explores an important but overlooked research question: What level of performance can be achieved with minimal architectural modifications by primarily focusing on supervision quality and optimization? Remarkably, under both GPT-4o and human evaluation, our method outperforms existing approaches despite using less data and requiring no architectural modifications or additional pretraining. This shows high-quality supervision signals can effectively compensate for architectural simplicity, offering valuable new perspectives for image editing research." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19213" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.184, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.24 + ], + "angle": 0, + "content": "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.242, + 0.484, + 0.297 + ], + "angle": 0, + "content": "[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.484, + 0.338 + ], + "angle": 0, + "content": "[4] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 2, 3, 4, 5, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.341, + 0.483, + 0.394 + ], + "angle": 0, + "content": "[5] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.396, + 0.483, + 0.437 + ], + "angle": 0, + "content": "[6] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. In ICLR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.483, + 0.478 + ], + "angle": 0, + "content": "[7] Yusuf Dalva and Pinar Yanardag. Noiseclar: A contrastive learning approach for unsupervised discovery of interpretable directions in diffusion models. In CVPR, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.483, + 0.534 + ], + "angle": 0, + "content": "[8] Fei Deng, Qifei Wang, Wei Wei, Tingbo Hou, and Matthias Grundmann. Prdp: Proximal reward difference prediction for large-scale reward finetuning of diffusion models. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.537, + 0.483, + 0.564 + ], + "angle": 0, + "content": "[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.565, + 0.483, + 0.62 + ], + "angle": 0, + "content": "[10] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.621, + 0.483, + 0.676 + ], + "angle": 0, + "content": "[11] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Reinforcement learning for fine-tuning text-to-image diffusion models. NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.677, + 0.483, + 0.718 + ], + "angle": 0, + "content": "[12] Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding instruction-based image editing via multimodal large language models. In ICLR, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.719, + 0.483, + 0.761 + ], + "angle": 0, + "content": "[13] Yuying Ge, Sijie Zhao, Chen Li, Yixiao Ge, and Ying Shan. Seed-data-edit technical report: A hybrid dataset for instructional image editing. arXiv preprint arXiv:2405.04007, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[14] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, et al. Instructdiffusion: A generalist modeling interface for vision tasks. In CVPR, 2024. 2, 3, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.483, + 0.859 + ], + "angle": 0, + "content": "[15] Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. NeurIPS, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[16] Xin Gu, Ming Li, Libo Zhang, Fan Chen, Longyin Wen, Tiejian Luo, and Sijie Zhu. Multi-reward as condition for instruction-based image editing. In ICLR, 2025. 3, 7, 8, 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.134 + ], + "angle": 0, + "content": "[17] Qin Guo and Tianwei Lin. Focus on your instruction: Fine-grained and multi-instruction image editing by attention modulation. In CVPR, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.907, + 0.177 + ], + "angle": 0, + "content": "[18] Qin Guo and Tianwei Lin. Focus on your instruction: Fine-grained and multi-instruction image editing by attention modulation. In CVPR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.18, + 0.906, + 0.221 + ], + "angle": 0, + "content": "[19] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In ICLR, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.224, + 0.906, + 0.251 + ], + "angle": 0, + "content": "[20] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.254, + 0.906, + 0.28 + ], + "angle": 0, + "content": "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.283, + 0.906, + 0.325 + ], + "angle": 0, + "content": "[22] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv preprint arXiv:2403.05135, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.327, + 0.907, + 0.369 + ], + "angle": 0, + "content": "[23] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. NeurIPS, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.371, + 0.906, + 0.438 + ], + "angle": 0, + "content": "[24] Yuzhou Huang, Liangbin Xie, Xintao Wang, Ziyang Yuan, Xiaodong Cun, Yixiao Ge, Jiantao Zhou, Chao Dong, Rui Huang, Ruimao Zhang, et al. Smartedit: Exploring complex instruction-based image editing with multimodal large language models. In CVPR, 2024. 2, 3, 7, 8, 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.906, + 0.496 + ], + "angle": 0, + "content": "[25] Mude Hui, Siwei Yang, Bingchen Zhao, Yichun Shi, Heng Wang, Peng Wang, Yuyin Zhou, and Cihang Xie. HQ-edit: A high-quality dataset for instruction-based image editing. arXiv preprint arXiv:2404.09990, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.499, + 0.907, + 0.568 + ], + "angle": 0, + "content": "[26] Harrison Lee, Samrat Phatale, Hassan Mansoor, Thomas Mesnard, Johan Ferret, Kellie Lu, Colton Bishop, Ethan Hall, Victor Carbune, Abhinav Rastogi, et al. Rlaif: Scaling reinforcement learning from human feedback with ai feedback. arXiv preprint arXiv:2309.00267, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.571, + 0.906, + 0.624 + ], + "angle": 0, + "content": "[27] Ming Li, Taojiannan Yang, Huafeng Kuang, Jie Wu, Zhaoning Wang, Xuefeng Xiao, and Chen Chen. Controlnet++: Improving conditional controls with efficient consistency feedback. In ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.628, + 0.906, + 0.669 + ], + "angle": 0, + "content": "[28] Yuanze Lin, Yi-Wen Chen, Yi-Hsuan Tsai, Lu Jiang, and Ming-Hsuan Yang. Text-driven image editing via learnable regions. In CVPR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.672, + 0.906, + 0.713 + ], + "angle": 0, + "content": "[29] Bingyan Liu, Chengyu Wang, Tingfeng Cao, Kui Jia, and Jun Huang. Towards understanding cross and self-attention in stable diffusion for text-guided image editing. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.715, + 0.907, + 0.742 + ], + "angle": 0, + "content": "[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.745, + 0.906, + 0.799 + ], + "angle": 0, + "content": "[31] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.906, + 0.857 + ], + "angle": 0, + "content": "[32] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. NeurIPS, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[33] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-g: Generating images in context with multimodal large language models. In ICLR, 2024. 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19214" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[34] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.483, + 0.232 + ], + "angle": 0, + "content": "[36] Or Patashnik, Daniel Garibi, Idan Azuri, Hadar Averbuch-Elor, and Daniel Cohen-Or. Localizing object-level shape variations with text-to-image diffusion models. In ICCV, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.483, + 0.287 + ], + "angle": 0, + "content": "[37] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.483, + 0.343 + ], + "angle": 0, + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.483, + 0.398 + ], + "angle": 0, + "content": "[39] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. NeurIPS, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.483, + 0.44 + ], + "angle": 0, + "content": "[40] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.483, + 0.482 + ], + "angle": 0, + "content": "[41] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.483, + 0.523 + ], + "angle": 0, + "content": "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.483, + 0.564 + ], + "angle": 0, + "content": "[43] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.483, + 0.635 + ], + "angle": 0, + "content": "[44] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurlPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.483, + 0.691 + ], + "angle": 0, + "content": "[45] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and generation tasks. In CVPR, 2024. 2, 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[46] Enis Simsar, Alessio Tonioni, Yongqin Xian, Thomas Hofmann, and Federico Tombari. Lime: localized image editing via attention regularization in diffusion models. arXiv preprint arXiv:2312.09256, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[47] Jaskirat Singh, Jianming Zhang, Qing Liu, Cameron Smith, Zhe Lin, and Liang Zheng. Smartmask: Context aware high-fidelity mask generation for fine-grained object insertion and layout control. In CVPR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[48] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[49] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[50] Kcolors Team. Kcolors: Effective training of diffusion model for photorealistic text-to-image synthesis. arXiv preprint, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.907, + 0.134 + ], + "angle": 0, + "content": "[51] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.136, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[52] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In CVPR, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.192, + 0.906, + 0.233 + ], + "angle": 0, + "content": "[53] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In CVPR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.235, + 0.907, + 0.289 + ], + "angle": 0, + "content": "[54] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. NeurlPS, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.291, + 0.907, + 0.345 + ], + "angle": 0, + "content": "[55] Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.347, + 0.906, + 0.389 + ], + "angle": 0, + "content": "[56] Mingyang Yi, Aoxue Li, Yi Xin, and Zhenguo Li. Towards understanding the working mechanism of text-to-image diffusion model. arXiv preprint arXiv:2405.15330, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.391, + 0.907, + 0.445 + ], + "angle": 0, + "content": "[57] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.447, + 0.906, + 0.489 + ], + "angle": 0, + "content": "[58] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. NeurIPS, 2024. 2, 3, 5, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.49, + 0.907, + 0.518 + ], + "angle": 0, + "content": "[59] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.519, + 0.907, + 0.574 + ], + "angle": 0, + "content": "[60] Shu Zhang, Xinyi Yang, Yihao Feng, Can Qin, Chia-Chih Chen, Ning Yu, Zeyuan Chen, Huan Wang, Silvio Savarese, Stefano Ermon, et al. Hive: Harnessing human feedback for instructional visual editing. In CVPR, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.576, + 0.907, + 0.631 + ], + "angle": 0, + "content": "[61] Wentian Zhang, Haozhe Liu, Jinheng Xie, Francesco Faccio, Mike Zheng Shou, and Jurgen Schmidhuber. Cross-attention makes inference cumbersome in text-to-image diffusion models. arXiv preprint arXiv:2404.02747, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.633, + 0.907, + 0.688 + ], + "angle": 0, + "content": "[62] Haozhe Zhao, Xiaojian Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. Ultraedit: Instruction-based fine-grained image editing at scale. arXiv preprint arXiv:2407.05282, 2024. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.907, + 0.688 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19215" + } + ] +] \ No newline at end of file diff --git a/2025/SuperEdit_ Rectifying and Facilitating Supervision for Instruction-Based Image Editing/395fe9d1-b3f1-4c45-afff-e5eaba8e26ad_origin.pdf b/2025/SuperEdit_ Rectifying and Facilitating Supervision for Instruction-Based Image Editing/395fe9d1-b3f1-4c45-afff-e5eaba8e26ad_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..470c075c49469e1433e39e4330121e8be51a1b77 --- /dev/null +++ b/2025/SuperEdit_ Rectifying and Facilitating Supervision for Instruction-Based Image Editing/395fe9d1-b3f1-4c45-afff-e5eaba8e26ad_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ef4b41216c222450314959e1bb45fa8be8a39a9b18a912bd7c2420a67424de1 +size 9945847 diff --git a/2025/SuperEdit_ Rectifying and Facilitating Supervision for Instruction-Based Image Editing/full.md b/2025/SuperEdit_ Rectifying and Facilitating Supervision for Instruction-Based Image Editing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..64a8f6cda06f7f784094a016b732abe784b1065b --- /dev/null +++ b/2025/SuperEdit_ Rectifying and Facilitating Supervision for Instruction-Based Image Editing/full.md @@ -0,0 +1,303 @@ +# SuperEdit: Rectifying and Facilitating Supervision for Instruction-Based Image Editing + +![](images/c2ca5a6587dbdec196e1bd2af034b0cb961ad511d7ba97e98087c6cf70d746f1.jpg) + +Website + +![](images/1dfe022054d2bc31d98f7b4664ca6b2ef1ac3ad32b5b48f00734411f4ab16e0b.jpg) + +Code + +![](images/d3e90959e8e22205dbda6a58db2876dbbb512e05de6fbd9dbf9dfa236f9492a2.jpg) + +Data + +Ming Li $^{1,2,\ddagger}$ , Xin Gu $^{1}$ , Fan Chen $^{1}$ , Xiaoying Xing $^{1}$ , Longyin Wen $^{1}$ , Chen Chen $^{2}$ , Sijie Zhu $^{1,*}$ $^{1}$ ByteDance Intelligent Creation Center for Research in Computer Vision, University of Central Florida + +![](images/c0340ed1d5247b3e0f8c9581efff1a4ec5797b1263767e54034fb4c2cf301c49.jpg) + +![](images/ec81b6f49f3c3d93a85a8e1d7c0df06b602a80611dc0aebba05c0b9564ed0e6a.jpg) + +![](images/7fe340698e5085dde31385bc7521f02a98d61a1cc9a8db24be8c7fb40023202d.jpg) +Global: Transform the scene into a snowy winter scene + +![](images/025353ddf90f7754c2b5b16880540198ad39bcc3750ba52e82369de167e7db07.jpg) +Replace: Replace the bow tie with a butterfly. + +![](images/b2e53b28238c2b7a879ed25964505f0ddd0abfc293eab28fe26a386f1f573554.jpg) +Style: Convert the image to a watercolor style Remove: Remove some clouds in the sky +(a) Examples of Our Method on Real & High-resolution Image + +![](images/dd8e57b09606df98dbb654124ffa8c514c0b3eefc2dd82a2c69beed868c003fc.jpg) + +![](images/1a709fe36d69a2d9f4630a351d54b3ecfa99c371e4339b663c3435495418bfa4.jpg) +Comparison with Existing Methods +(b) Comparison with GPT-4o Evaluation + +![](images/e9026c6f1f07d055b479d701aa26a475843a7322b08da88f137dbb67b5d2360b.jpg) +(c) Comparison with Human Evaluation +Figure 1. (a) Our editing method works well with real and high-resolution images, handling various free-form edits (left) and local edits (right); (b) Compared to the current state-of-the-art SmartEdit, our method achieves a $9.19\%$ performance improvement with $30\times$ less training data and $13\times$ fewer model parameters; (c) Our method achieves better overall scores on the human evaluation results, indicating more precise editing capabilities. + +# Abstract + +Due to the challenges of manually collecting accurate editing data, existing datasets are typically constructed using various automated methods, leading to noisy supervision signals caused by the mismatch between editing instructions and original-edited image pairs. Recent efforts attempt to improve editing models through generating higher-quality edited images, pre-training on recognition tasks, or introducing vision-language models (VLMs) but fail to resolve this fundamental issue. In this paper, we offer a novel solution by constructing more effective editing instructions for given image pairs. This includes rectifying the editing instructions to better align with the original-edited image pairs and using contrastive editing instructions to further enhance their effectiveness. Specifically, we find that editing models exhibit specific generation attributes at different inference steps, independent of the text. Based on these prior attributes, we define a unified guide for VLMs to rectify editing instructions. However, there are some challenging editing scenarios that cannot be resolved solely with rectified instructions. + +To this end, we further construct contrastive supervision signals with positive and negative instructions and introduce them into the model training using triplet loss, thereby further facilitating supervision effectiveness. Our method does not require the VLM modules or pre-training tasks used in previous work, offering a more direct and efficient way to provide better supervision signals, and providing a novel, simple, and effective solution for instruction-based image editing. Results on multiple benchmarks demonstrate that our method significantly outperforms existing approaches. Compared with previous SOTA SmartEdit, we achieve $9.19\%$ improvements on the Real-Edit benchmark with $30\times$ less training data and $13\times$ smaller model size. All data and models are open-sourced on Github for future research. + +# 1. Introduction + +In recent years, significant progress has been made in text-to-image (T2I) generation [10, 37, 40-42] due to the development of diffusion models [9, 21, 48, 49]. These T2I diffusion models can generate images that align with natural language descriptions while satisfying human perception and preferences. Consequently, numerous image editing + +methods [4, 6, 19, 31] based on these models have been proposed to achieve various editing effects. Instruction-based methods [4, 12, 24] have become increasingly popular as they allow users to conveniently and easily modify images using language instructions without the need to provide masks, as required by mask-based methods [18, 28, 47, 53]. + +The training of instruction-based editing models requires the original-edited image pairs and corresponding editing instruction, making it difficult to manually create or collect a large amount of relevant data [58]. To address the issue of scarce training data, existing efforts [13, 19, 25, 62] have attempted to develop various automated pipelines to synthesize large datasets. Specifically, most methods first use large language models (LLMs) to modify the text descriptions of original images. The original images and modified texts are then input into various pre-trained diffusion models to automatically generate edited images. However, current text-to-image diffusion models struggle to fully correspond to input text prompts [15, 59]. This often changes parts of the original images that do not require editing, leading to misaligned editing instructions and original-edited image pairs, thus resulting in noisy supervision signals. To mitigate the potential issues of noisy supervision in image editing models, existing work has attempted to introduce additional recognition pre-training tasks for U-Net [43] such as semantic segmentation [14, 45], or replace CLIP [38] text encoder with vision-language models (VLMs) [12, 24] to better understand editing instructions from noisy supervision signals. However, these methods not only introduce significant computational overhead but also overlook the issue of noisy supervision signals. + +In this paper, we focus on addressing the fundamental challenge by introducing more effective editing instructions, as demonstrated in Fig. 2. Our data-oriented method explores a different research question: how much performance improvement can be achieved solely by focusing on supervision signal quality and optimization in image editing? Surprisingly, SuperEdit outperforms existing methods in both GPT-4o and human evaluations, despite using less data and requiring no additional modules or pretraining as shown in Fig.1. This demonstrates that high-quality supervision signals can significantly compensate for architectural simplicity, achieving results comparable to or better than methods with more complex requirements. + +Specifically, to enhance the effectiveness of supervision signals for instruction-based image editing methods, we propose using VLMs to rectify editing instructions, creating better-aligned instructions for the original-edited image pairs. However, determining which VLM to use for this task and how to establish a unified rectification method for various editing instructions remain unexplored problems. To address this, we first analyze the ability of different VLMs to understand the differences between original and edited images, showing that GPT-4o [1] is the most capable of rectifying editing instructions. Additionally, we observe that both editing models and text-to-image diffusion models share a similar prior, as shown in Fig. 4: different inference + +![](images/588b8b2639c6ce4ed4512fc8e5f3bd26b3b1ce7fe508459c2ed64d91aa84f442.jpg) +(c) InstructDiffusion; Emu-Edit +Figure 2. Unlike existing efforts that attempt to (a) scale up edited images with noisy supervision [4, 25], (b) introduce massive VLMs into editing model architecture [12, 24], and (c) perform additional pre-training tasks [14, 45], (d) we focus on improving the effectiveness of supervision signals, which is the fundamental issue of image editing. + +![](images/39c8d4fcbb602f3d1c7c3d7dbf54a8dd3d16cdeb40b2d534de7497da4668aa2a.jpg) +(d) Ours + +stages correspond to the generation of different image attributes, independent of the input text prompt [2, 17, 19, 36, 46, 56, 61] or editing instructions. Inspired by this, we guide VLMs based on these attributes to establish a unified rectification guideline for various editing instructions, as demonstrated in Fig. 3. + +When training with only rectified editing instructions, we find that the editing model can better understand the editing commands but still faces challenges in handling complex scenarios. For example, when the original image contains multiple objects, the edit model struggles to perform an accurate editing function if the instructions modify only one of these objects. Additionally, inherent issues present in pre-trained text-to-image diffusion models [15, 22-24], such as difficulty in understanding quantity, position, or object relationships, persist in the editing models. To address these issues, we propose using contrastive supervision signals to further optimize the editing model. Specifically, we first construct incorrect editing instructions based on the rectified instructions to generate positive and negative samples. We then introduce a triplet loss to guide the model, thereby enhancing the effectiveness of supervision, as shown in Fig. 5. + +In summary, our contributions are summarized as follows: + +- New Insight: We aim to address the noisy supervision problem that arises from the misalignment between editing instructions and original-edited image pairs, which is a fundamental issue overlooked by previous work, as shown in Fig. 2. +- Rectifying Supervision: We leverage diffusion generation priors to guide the vision-language model to generate better-aligned editing instructions for original-edited image pairs. +- Facilitating Supervision: We introduce contrastive supervision using triplet loss, enabling the editing model to learn from both positive and negative editing instructions. +- Promising Results: We achieve significant improvements on multiple benchmarks without additional pre-training or VLM. Compared to SmartEdit [24], we achieved a $9.19\%$ improvement while reducing $30 \times$ data and $13 \times$ model parameters. + +# 2. Related Work + +# 2.1. Image Editing with Diffusion Models + +Building on advancements in text-to-image (T2I) diffusion models [10, 37, 40-42, 44], recent research has explored them for image editing [4, 19]. Training-free methods [5, 19, 29, 31, 34, 51] typically achieve this by adjusting attentions in pre-trained T2I models, but have limited performance and generalization capabilities on various editing tasks. + +Training-based methods address these limitations with specialized editing models, which can be categorized into mask-based and instruction-based approaches. Mask-based methods [6, 18, 28, 47, 53, 57] enable fine-grained local edits with user-provided or predicted masks and corresponding text descriptions. However, it struggles with global image editing and is constrained by the lack of mask-based editing data [24]. + +Instruction-based methods directly accept textual commands, such as "add a dog", offering better editing flexibility and generalization. InstructPix2Pix [4] pioneered this paradigm by generating instruction-based editing data and modifying the conditions of T2I diffusion models. Building on this framework, subsequent work introduces vision-language models [12, 24, 33] or additional pre-training tasks for the denoising U-Net [14, 24, 43, 45] to enhance the understanding and reasoning of input conditions. However, these methods not only introduce substantial computational overhead but also overlook the fundamental noisy supervision issue. + +# 2.2. Generating and Improving Editing Supervision + +Due to the difficulty of scaling instruction-based image editing data through manual collection, existing efforts [4, 13, 25, 58, 62] aim to automatically modify text descriptions of original images and generate edited images with T2I diffusion models. However, this approach often produces synthesized images that do not align with the editing instructions, as shown in Figure 3, resulting in noisy editing supervision signals [58, 62]. To address this, MagicBrush [58] manually filters out incorrect editing data, but it is hard to scale. Unlike existing methods focusing on edited image quality, we leverage diffusion prior and vision-language model (i.e., GPT-40 [1]) to create better-aligned instructions with original-edited image pairs, providing more accurate supervision. + +# 2.3. Alignment of Diffusion Models + +The success of alignment training in large language models (LLMs) [26, 32, 39] has been applied to diffusion models for better image generation. This is achieved by maximizing reward scores [8, 27, 54] or the generation probability of the winner image in a pair [11, 52, 55]. In image editing, HIVE [60] and MultiReward [16] attempt to incorporate reward information into the text condition to align the editing model. In contrast, we guide the editing model by rectifying and constructing contrastive editing instructions, achieving more effective alignment. + +# 3. Method + +In this section, we first introduce the most general image editing framework in Sec. 3.1. Then, we explain how to use diffusion priors to rectify editing instructions with the multimodal model (i.e., GPT-4o) in Sec. 3.2, thereby enhancing the accuracy of supervision signals. Finally, we describe how to construct contrastive supervision with both correct and incorrect editing instructions and integrate it into the editing model training using triplet loss in Sec. 3.3. + +# 3.1. Instruction-based Image Editing Framework + +InstructPix2Pix [4] pioneered instruction-based image editing, performing editing tasks by simultaneously taking the original image $C^I$ and editing instructions $C^T$ as input conditions to generate the edited image $x$ from random noise $\epsilon$ . Following the definition of DDPM [21], we randomly sample a timestep $t \in T$ during training, and then add corresponding noise $\epsilon_t$ to the edited image $x$ : + +$$ +x _ {t} = \sqrt {\bar {\alpha} _ {t}} x + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {t}, \quad \epsilon \sim \mathcal {N} (\mathbf {0}, I), \tag {1} +$$ + +where $\epsilon$ is a noise map sampled from a Gaussian distribution, and $\bar{\alpha}_t\coloneqq \prod_{s = 0}^t\alpha_s$ , $\alpha_{t} = 1 - \beta_{t}$ is a differentiable function of timestep $t$ , which is determined by the denoising sampler such as DDPM [21]. Then the training objective of the editing model $\epsilon_{\theta}$ is predicting the added noise at timestep $t$ , which can be written as: + +$$ +\mathcal {L} _ {\text {t r a i n}} = \mathbb {E} _ {x _ {t}, t, C ^ {I}, C ^ {T}, \epsilon} \left[ \left\| \epsilon_ {\theta} \left(\operatorname {c o n c a t} \left(x _ {t}, C ^ {I}\right), t, C ^ {T}\right) - \epsilon_ {t} \right\| _ {2} ^ {2} \right], \tag {2} +$$ + +where concat refers to concatenating the image latents of noised edited image $x_{t}$ and original image $c_{I}$ in the channel dimension. + +# 3.2. Rectifying Supervision with Diffusion Priors + +As shown in Fig. 3, existing image editing datasets [4, 13, 58] typically use only Steps 1 and 2: LLMs construct editing prompts and captions, and then text-to-image diffusion models synthesize edited images. However, diffusion models often fail to accurately follow prompts while maintaining image layout, creating mismatches between original-edited pairs and editing instructions, resulting in inaccurate supervision. While better supervision signals for text-to-image diffusion models are common in image generation [3, 50], this approach remains underexplored in image editing due to two challenges: (1) VLMs trained on single-image data struggle with multi-image inputs, and (2) editing instructions vary widely, making unified rectification guidelines difficult. To address these issues, we: (1) analyzed different VLMs' capabilities with multi-image inputs, finding GPT-4o most effective, and (2) discovered that timestep-specific roles in image generation also apply to editing, providing a foundation for a unified rectification method across various instructions (Fig. 3 and 4). Due to page limitations, our VLM analysis is in the Supplementary Material, while this section focuses on Diffusion Prior and Editing Instruction Rectification. + +![](images/fdde02f13f77e9774b108c3b58180ea004f09ab08d1e62c594fabfdff486e951.jpg) +Figure 3. (a) Existing work primarily uses LLMs and diffusion models to automatically generate edited images. However, current diffusion models often fail to accurately follow text prompts while maintaining the input image's layout, resulting in mismatches between the original-edited image pairs and the editing instructions. (b) We perform instruction rectification (Step 3) based on the images constructed in Steps 1 and 2. We show VLMs can understand the differences between the images, enabling them to rectify editing instructions to be better aligned with original-edited image pairs. + +![](images/f23c3efcdc9fefb30d130072c56bac743446b7a251c68c1d7ceb3b08c5514d80.jpg) + +![](images/1645bbc98dc6936728b9ff771ba16479949c73578c30bf58456f7f36312800d6.jpg) + +![](images/5e45d22c74c34dc577fc4e7888b3561686cb3741c730bf1df9aaf04e785d8fd3.jpg) +Figure 4. We show that the editing model follows consistent generation attributes at different sampling stages, independent of the editing instructions. The early, middle, and late sampling stages correspond to global, local, and detail changes, respectively, while style changes occur at all stages. All the generated images here are DDIM 30-step sampled final images. The orange progress bar and the grid progress bar represent the sampling stages with and without the editing instructions, respectively. + +Diffusion Generation Priors. Previous work has shown that different timesteps play distinct roles in image generation for text-to-image diffusion models, regardless of the text prompt [2, 17, 19, 36, 46, 56, 61]. We find that this phenomenon also exists in instruction-based editing models and present examples based on pre-trained InstructPix2Pix [4], as shown in Fig. 4. Specifically, diffusion models focus on global layout in the early stages, local object attributes in the mid stages, and image details + +in the late stages of sampling. This finding inspires us to guide VLMs based on these four generation attributes, establishing a unified rectification method for various editing instructions. We provide more analysis and details in the Supplementary Material. + +Editing Instruction Rectification. As demonstrated in Fig. 3, we extend the existing editing data generation pipeline by introducing our instruction rectification (Step 3). This process relies on the original edited image pairs obtained through Steps 1 and 2 from previous work. Specifically, we input original-edited image pairs into the vision-language model (i.e., GPT-4o) and instruct it to describe the changes in the edited image compared to the original image according to the above diffusion prior generation attributes. Finally, we use VLM to summarize the instructions and ensure that its length is less than the maximum length of CLIP text encoder, which is 77 tokens. + +# 3.3. Facilitating Supervision with Contrastive Instructions + +Although using rectified editing instructions can significantly improve performance across various editing tasks, we find that editing models still struggle with closely related text instructions. For example, "add a cat on the left side of the image" and "add two cats on the right side of the image" might produce the same edited image. This indicates that inherent biases in pre-trained text-to-image diffusion models [15, 22], such as difficulties in understanding quantity, position, and spatial relationships, persist in editing models. More importantly, our experiments show that training models with rectified editing instructions does not resolve these challenges. To further facilitate supervision signal effectiveness, we drew on successful alignment experiences from large language models [1, 32, 39] and text-to-image diffusion models [7, 52, 54]: constructing positive and negative sample pairs and guiding the model to assign a higher generation probability to positive samples compared to negative ones. + +![](images/162a52a8a1dcc266331549d208f7f177bbf2aa21ad63ea7dbd4fdce26317e909.jpg) +Figure 5. (a) Based on the rectified editing instruction and original-edited image pair, we utilize the Vision-Language Models (VLM) to generate various image-related wrong instructions. These involve random substitutions of quantities, spatial locations, and objects within the rectified editing instructions according to the original-edited images context; (b) During each training iteration, we randomly select one wrong instruction $c_{\text{neg}}^T$ and input it along with the rectified instruction $c_{pos}^T$ into the editing model to obtain predicted noises. The goal is to make the rectified instruction's predicted noise $\epsilon_{pos}$ closer to the sampled training diffusion noise $\epsilon$ , while ensuring the noise from incorrect instructions $\epsilon_{\text{neg}}$ is farther. Best viewed in color. + +![](images/d2df2c9f86bef1789db5a2f08610bce45bd67df87c7d4970f19964fcf01d8296.jpg) + +Constructing Contrastive Instructions. Unlike the standard alignment process for large language models or text-to-image diffusion models, it is challenging to generate different editing results from the same instruction to create positive and negative sample pairs for image editing tasks. To address this, we construct positive and negative editing instructions for alignment, thereby generating relatively positive and negative edited images. As shown in Fig. 5 (a), we use the original image, edited image, and rectified editing instruction as input. The VLM (i.e., GPT-4o) is used to modify attributes in the rectified editing instruction, such as quantity, spatial relationships, and object types, to create different wrong instructions. Here, we require VLM to modify only a single attribute from the rectified editing instruction in each wrong instruction, keeping most of the editing text unchanged. Since only a few words are replaced between the rectified instruction and the wrong instruction, the text embeddings produced by the CLIP text encoder that serve as input to the denoising model will also be similar. This ensures the task's learning difficulty, helping the model understand how subtle differences between the two editing instructions result in significantly different editing results. + +Facilitating Editing Models with Contrastive Instructions. Our key insight is that enhancing the effectiveness of supervision signals can improve various editing tasks without introducing additional model architectures or pre-training tasks. Therefore, we adhere strictly to the InstructPix2Pix [4] model architecture and training pipeline. To be specific, the inputs including the original image $c^I$ , edited image $x$ , the rectified instruction $c_{pos}^T$ , and wrong editing instruction $c_{neg}^T$ . During training, we will add a sampled timestep $t \in T$ to obtain the noised edited image $x_t$ with Equation 1. Both the rectified and wrong editing instructions are fed into the denoising model to predict the final noises $\epsilon_{pos}$ and $\epsilon_{neg}$ , which are then used to + +construct positive and negative samples, respectively: + +$$ +\epsilon_ {p o s} = \epsilon_ {\theta} \left(\operatorname {c o n c a t} \left(x _ {t}, c ^ {I}\right), t, c _ {p o s} ^ {T}\right), \tag {3} +$$ + +$$ +\epsilon_ {n e g} = \epsilon_ {\theta} \big (\operatorname {c o n c a t} \left(x _ {t}, c ^ {I}\right), t, c _ {n e g} ^ {T} \big). \tag {4} +$$ + +After constructing the positive and negative sample pairs, we aim for the noise predicted by the positive editing instruction $\epsilon_{pos}$ to be closer to the true noise $\epsilon_t$ sampled during training, compared to the noise predicted by the wrong editing instruction $\epsilon_{neg}$ . This goal can be achieved through a triplet loss function: + +$$ +\mathcal {L} _ {\text {t r i p l e t}} = \max \left\{d \left(\epsilon_ {t}, \epsilon_ {\text {p o s}}\right) - d \left(\epsilon_ {t}, \epsilon_ {\text {n e g}}\right) + m, 0 \right\}, \tag {5} +$$ + +where $d(x,y) = \|\mathbf{x} - \mathbf{y}\|_2^2$ and margin $m$ is a hyper-parameter. The final training loss is the combination of the original diffusion training loss and the triplet loss: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {t r a i n}} + \lambda \cdot \mathcal {L} _ {\text {t r i p l e t}}, \text {w h e r e} \mathcal {L} _ {\text {t r a i n}} = d \left(\epsilon_ {t}, \epsilon_ {\text {p o s}}\right). \tag {6} +$$ + +Please note that the contrastive supervision signals are only used during the training phase. During inference, the editing model only requires one single input editing instruction. + +# 4. Experiment + +# 4.1. Data Collection and Construction + +To build a diverse dataset with various types of editing instructions, we need original and edited images from different data domains, as well as a wide variety of editing instructions. To achieve this, we sampled data from different public editing datasets to construct rectified and contrastive supervision signals. Specifically, we extracted 10,177, 8,807, and 21,016 editing pairs from InstructPix2Pix [4], MagicBrush [58], and Seed-Data-Edit [13], respectively, resulting in a total of 40,000 training samples. During extraction, we strive to ensure that the data for different types of editing tasks is as balanced as possible. + +![](images/35dbcee4f093c470ccb517a3cce40883521fa0057500e611f1e7679b3bc1a05b.jpg) +Figure 6. Visual comparison with existing methods and the corresponding human-aligned GPT-4o evaluation scores (Following, Preserving, Quality Scores from left to right). We achieve better results while preserving the layout, quality, and details of the original image. Please note that we do not claim that our editing results are flawless. We provide more visual comparison results in the supplementary material. + +
MethodExtra ModulePretrain TasksEdit DataModel SizeFollowing ↑Preserving ↑Quality ↑Overall ↑
AccScoreAccScoreAccScoreAccScore
KOSMOS-G [33]✓✓9.0M1.9B51%2.829%1.4327%3.2029.0%2.48
MGIE [12]✓✓1.0M8.1B40%2.4345%2.7938%3.3541.0%2.86
SmartEdit [24]✓✓1.2M14.1B64%3.5066%3.7045%3.5658.3%3.59
MultiReward [16]✓✓320K1.2B63%3.3958%3.4354%3.8058.3%3.54
InstructDiffusion [14]X✓860K1.1B52%2.8754%3.1745%3.5850.3%3.21
InstructPix2Pix [4]XX300K1.1B52%2.9453%3.3150%3.6951.7%3.31
MagicBrush [58]XX310K1.1B51%2.9070%3.8550%3.6757.0%3.47
HIVE [60]XX1.1M1.1B54%2.9356%3.3653%3.7254.3%3.34
HQ-Edit [25]XX500K1.1B51%2.8416%1.6354%3.8440.3%2.77
SuperEdit (Ours)XX40K1.1B67%3.5977%4.1465%4.0169.7%3.91
+ +![](images/508db1f1a9247e288ae28c88a641cb6d98e7eb900f9d1be3a63b0af34aa08f3a.jpg) +(a) Following Score + +![](images/e1c861c97676cba85a10ded11521df6c0aaeb41c54fcf90de3ec9dd7ae60237d.jpg) +(b) Preserving Score +Figure 7. Human evaluation on three evaluation criteria for image editing effects. (a) Following: whether the edited image adheres to the editing instructions; (b) Preserving: whether the image structure outside of the editing instructions has been preserved; (c) Quality: whether the overall quality/aesthetics of the edited image has been degraded compared to the input image. Our SuperEdit achieves the best results on all of these metrics. + +![](images/d9cd5943e7f01b703b88967325315bdf68d820d296c8d04fdbef5ecaae428227.jpg) +(c) Quality Score + +Table 1. Comparison with instruction-based image editing methods on Real-Edit benchmark [16]. Compared to existing work, our method achieves state-of-the-art performance across all metrics using a small amount of high-quality editing data without introducing additional models or pre-training tasks. Please note that the scores range from 0 to 5. $\uparrow$ denotes a higher result is better. All baseline results are cited from the MultiReward [16] paper. + +
Following ↑Preserving ↑Quality ↑Overall ↑
InstructPix2Pix [4]2.412.622.442.49
SmartEdit-13B [24]3.093.062.632.93
SuperEdit3.18+1.80%3.86+16.00%3.37+14.80%3.47+10.80%
+ +Table 2. Human evaluation results on Real-Edit [16] benchmark. All the human-evaluated scores range from 0 to 5. Overall represents the average score of Following, Preserving, and Quality scores. + +We then applied our proposed methods in Sec. 3 to rectify and construct contrastive editing instructions for these training samples. Since the MagicBrush data has been manually verified, we skip the rectification step for this dataset and directly construct contrastive supervision based on the original editing instructions. For Seed-Data-Edit dataset, we only sample images from the first part of data without human editing instructions. + +# 4.2. Experimental Settings + +Evaluation Benchmarks and Metrics. To more accurately assess the effectiveness of various editing models, we conducted assessments on the Real-Edit benchmark [16], which is a human-aligned evaluation benchmark with GPT-4o scoring. Specifically, MultiReward [16] uses high-resolution images from the Unsplash community as a test dataset and combines them with GPT-4o [1] to create an automated evaluation method for single-turn editing. It assesses edited images in terms of accuracy (\%) and scores (from 0 to 5), evaluating whether they adhere to the editing instructions (Following), whether the image structure outside of the editing instructions has been preserved (Preserving), and whether the overall quality/aesthetics of the edited image has been degraded compared to the original one (Quality). + +# 4.3. Experimental Results + +Comparison on Real-Edit Benchmark. In Tab. 1, we present the quantitative results of editing effectiveness on the Real-Edit benchmark [16]. Without introducing additional parameters or pre-training stages, our method achieves the best results in the three GPT-4o automated evaluation metrics: Following, Preserving, and Quality, each of which includes percentage accuracy (Acc) and scores (from 0 to 5). For example, compared to SmartEdit [24], which introduces an additional 13B vision-language model (i.e., LLaVA [30]) to the 1.1B InstructPix2Pix [4] framework, we achieved improvements of $11.4\%$ Overall Score. This suggests that given accurate and effective supervision signals, the trained editing model can understand and successfully execute the editing instructions, without the need for additional vision-language models. + +It is worth noting that unlike existing image editing methods, which often show improvement in a single metric while others remain unchanged or worsen, our method achieves comprehensive and significant advancements across all three metrics. This indicates that improving the effectiveness of supervision signals can accurately execute editing instructions while reducing disruption to other non-edited parts of the image, and preserving the quality and aesthetics of the original images. Specifically, we surpassed the current best methods by $3\%$ , $7\%$ , and $11\%$ Acc results in Following, Preserving, and Quality, respectively. + +Human Evaluation We also conduct a comprehensive human evaluation on Real-Edit benchmarks [16] in Tab. 2 and Fig. 7. The assessment involved 15 experienced evaluators + +who rated edited images based on three critical metrics: instruction faithfulness (Following), preservation of irrelevant content (Preserving), and visual quality (Quality). The results of this manual evaluation demonstrate strong consistency with the GPT-4o scoring results shown in Tab. 1, with Cohen's Kappa=0.89, Krippendorff's Alpha=0.87. This high alignment thoroughly validates that our proposed SuperEdit significantly outperforms existing methods across all evaluation criteria. Specifically, our SuperEdit surpasses the previous state-of-the-art method SmartEdit [24] by $1.8\%$ , $16\%$ , $14.8\%$ , and $10.8\%$ on Following, Preserving, Quality, and Overall scores, respectively. These substantial improvements not only confirm the effectiveness of our approach but also establish SuperEdit as a new benchmark in instruction-guided image editing, achieving superior performance while requiring significantly less training data and cost. + +Visual Comparison with State-of-the-art Methods. We show the visual comparison with existing image editing methods in Fig. 6. Compared to existing instruction-based editing methods, our approach not only better understands and executes editing instructions but also preserves the original image's layout and quality more effectively, thereby significantly outperforming previous methods. For example, with the instruction "Replace the tiger with a lion, maintaining the same position in the water" our SuperEdit method achieved superior results (4.8/4.8/4.8) compared to SmartEdit (4.8/4.8/2.5) and other methods. Additionally, our method improves the model's comprehension of editing instructions. For the instruction "Change the background to a sandy beach with the ocean in the distance" our method received perfect scores (4.8/5.0/5.0) while SmartEdit only achieved (5.0/4.8/4.0). Similarly, for style transformation instructions like "Change the image style to look like an impressionist painting style" SuperEdit significantly outperformed SmartEdit with scores of (4.8/4.8/4.8) versus (1.0/4.8/4.8), demonstrating our method's superior ability to handle complex artistic transformations. Even more impressively, for scene transformation tasks like "Transform the entire scene to a winter setting with snow covering the houses, trees, and boat", our SuperEdit achieved (5.0/4.8/4.8) while SmartEdit only obtained (2.0/4.5/4.5). We provide more visual comparisons with other instruction-based image editing methods in the Supplementary Material. + +# 4.4. Ablation Study + +Ablation on the Rectified and Contrastive Instructions. Considering that the Real-Edit [16] benchmark is evaluated by GPT-4o [1], and its evaluation results closely align with human ratings [16], we choose this benchmark to conduct ablation experiments in Tab. 3. Compared to the original 40K training data, our training data with rectified editing instructions significantly improves all the performance of the editing model. Specifically, our approach improves scores by 0.95, 0.79, and 0.11, and accuracy by $21\%$ , $22\%$ , and $4\%$ in these three metrics, respectively. In addition, editing performance is further enhanced by incorporating contrastive supervision signals. Compared to using only + +
Rectified InstructionContrastive InstructionsFollowing↑Preserving↑Quality↑
AccScoreAccScoreAccScore
XX41%2.4553%3.2761%3.90
✓X62%3.4075%4.0665%4.01
✓✓67%3.5977%4.1465%4.01
+ +rectified editing instructions, the introduction of contrastive supervision signals improves the following and preserving scores by 0.19 and 0.08, and accuracy by $5\%$ and $2\%$ , while maintaining the quality accuracy and score. In summary, both the introduction of rectified editing instructions and contrastive editing instructions improve the overall performance of the editing model. + +Ablation on Data Scaling. We investigated the impact of training data volume on model performance by experimenting with datasets ranging from 5k to 40k samples. Tab. 4 shows consistent improvements across all metrics as training data increases. With just 5k samples, our model achieves reasonable performance (54.7% accuracy, 3.42 overall score), but scaling to 40k samples yields substantial gains (69.7% accuracy, 3.91 overall score). The most significant improvements appear in the Preserving and Quality metrics, with 10% and 15%, respectively. This upward trend across all data points demonstrates that SuperEdit effectively leverages additional training examples without performance saturation, suggesting potential for further gains with larger datasets. + +Table 3. Ablation study on our methods. Both rectified and contrastive editing instructions achieved improvements across all metrics. + +
Data SizeFollowing ↑Preserving ↑Quality ↑Overall ↑
AccScoreAccScoreAccScoreAccScore
5k49%2.8760%3.7155%3.6954.7%3.42
10k57%3.2671%3.7658%3.8762.0%3.63
20k64%3.4072%4.0263%3.9466.3%3.79
40k67%3.5977%4.1465%4.0169.7%3.91
+ +Table 4. Ablation study on data scaling results on Real-Edit [16]. + +# 5. Conclusion + +In this paper, we re-examine image editing models from the perspective of enhancing supervision signals, finding that existing models have not adequately addressed this challenge, resulting in suboptimal performance. We introduce a unified editing instruction rectification guideline based on diffusion priors that better aligns instructions with original-edited image pairs, thereby enhancing supervision effectiveness. We also construct contrastive editing instructions allowing models to learn from both positive and negative examples. Our data-oriented approach explores an important but overlooked research question: What level of performance can be achieved with minimal architectural modifications by primarily focusing on supervision quality and optimization? Remarkably, under both GPT-4o and human evaluation, our method outperforms existing approaches despite using less data and requiring no architectural modifications or additional pretraining. This shows high-quality supervision signals can effectively compensate for architectural simplicity, offering valuable new perspectives for image editing research. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 3, 4, 7, 8 +[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2, 4 +[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2023. 3 +[4] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 2, 3, 4, 5, 7, 1 +[5] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In ICCV, 2023. 3 +[6] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. In ICLR, 2023. 2, 3 +[7] Yusuf Dalva and Pinar Yanardag. Noiseclar: A contrastive learning approach for unsupervised discovery of interpretable directions in diffusion models. In CVPR, 2024. 4 +[8] Fei Deng, Qifei Wang, Wei Wei, Tingbo Hou, and Matthias Grundmann. Prdp: Proximal reward difference prediction for large-scale reward finetuning of diffusion models. In CVPR, 2024. 3 +[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 2021. 1 +[10] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 1, 3 +[11] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Reinforcement learning for fine-tuning text-to-image diffusion models. NeurIPS, 2024. 3 +[12] Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding instruction-based image editing via multimodal large language models. In ICLR, 2024. 2, 3, 7 +[13] Yuying Ge, Sijie Zhao, Chen Li, Yixiao Ge, and Ying Shan. Seed-data-edit technical report: A hybrid dataset for instructional image editing. arXiv preprint arXiv:2405.04007, 2024. 2, 3, 5 +[14] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, et al. Instructdiffusion: A generalist modeling interface for vision tasks. In CVPR, 2024. 2, 3, 7, 1 +[15] Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. NeurIPS, 2024. 2, 4 +[16] Xin Gu, Ming Li, Libo Zhang, Fan Chen, Longyin Wen, Tiejian Luo, and Sijie Zhu. Multi-reward as condition for instruction-based image editing. In ICLR, 2025. 3, 7, 8, 1 + +[17] Qin Guo and Tianwei Lin. Focus on your instruction: Fine-grained and multi-instruction image editing by attention modulation. In CVPR, 2024. 2, 4 +[18] Qin Guo and Tianwei Lin. Focus on your instruction: Fine-grained and multi-instruction image editing by attention modulation. In CVPR, 2024. 2, 3 +[19] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In ICLR, 2023. 2, 3, 4 +[20] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 1 +[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 1, 3 +[22] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv preprint arXiv:2403.05135, 2024. 2, 4 +[23] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. NeurIPS, 2023. +[24] Yuzhou Huang, Liangbin Xie, Xintao Wang, Ziyang Yuan, Xiaodong Cun, Yixiao Ge, Jiantao Zhou, Chao Dong, Rui Huang, Ruimao Zhang, et al. Smartedit: Exploring complex instruction-based image editing with multimodal large language models. In CVPR, 2024. 2, 3, 7, 8, 1, 4 +[25] Mude Hui, Siwei Yang, Bingchen Zhao, Yichun Shi, Heng Wang, Peng Wang, Yuyin Zhou, and Cihang Xie. HQ-edit: A high-quality dataset for instruction-based image editing. arXiv preprint arXiv:2404.09990, 2024. 2, 3, 7 +[26] Harrison Lee, Samrat Phatale, Hassan Mansoor, Thomas Mesnard, Johan Ferret, Kellie Lu, Colton Bishop, Ethan Hall, Victor Carbune, Abhinav Rastogi, et al. Rlaif: Scaling reinforcement learning from human feedback with ai feedback. arXiv preprint arXiv:2309.00267, 2023. 3 +[27] Ming Li, Taojiannan Yang, Huafeng Kuang, Jie Wu, Zhaoning Wang, Xuefeng Xiao, and Chen Chen. Controlnet++: Improving conditional controls with efficient consistency feedback. In ECCV, 2024. 3 +[28] Yuanze Lin, Yi-Wen Chen, Yi-Hsuan Tsai, Lu Jiang, and Ming-Hsuan Yang. Text-driven image editing via learnable regions. In CVPR, 2024. 2, 3 +[29] Bingyan Liu, Chengyu Wang, Tingfeng Cao, Kui Jia, and Jun Huang. Towards understanding cross and self-attention in stable diffusion for text-guided image editing. In CVPR, 2024. 3 +[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2024. 7 +[31] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 2, 3 +[32] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. NeurIPS, 2022. 3, 4 +[33] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-g: Generating images in context with multimodal large language models. In ICLR, 2024. 3, 7 + +[34] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH, 2023. 3 +[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 1 +[36] Or Patashnik, Daniel Garibi, Idan Azuri, Hadar Averbuch-Elor, and Daniel Cohen-Or. Localizing object-level shape variations with text-to-image diffusion models. In ICCV, 2023. 2, 4 +[37] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2023. 1, 3 +[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3 +[39] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. NeurIPS, 2024. 3, 4 +[40] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 1, 3 +[41] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. +[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 3 +[43] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 2, 3 +[44] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurlPS, 2022. 3 +[45] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and generation tasks. In CVPR, 2024. 2, 3, 1 +[46] Enis Simsar, Alessio Tonioni, Yongqin Xian, Thomas Hofmann, and Federico Tombari. Lime: localized image editing via attention regularization in diffusion models. arXiv preprint arXiv:2312.09256, 2023. 2, 4 +[47] Jaskirat Singh, Jianming Zhang, Qing Liu, Cameron Smith, Zhe Lin, and Liang Zheng. Smartmask: Context aware high-fidelity mask generation for fine-grained object insertion and layout control. In CVPR, 2024. 2, 3 +[48] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 1 +[49] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 1 +[50] Kcolors Team. Kcolors: Effective training of diffusion model for photorealistic text-to-image synthesis. arXiv preprint, 2024. 3 + +[51] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In CVPR, 2023. 3 +[52] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In CVPR, 2024. 3, 4 +[53] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In CVPR, 2023. 2, 3 +[54] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. NeurlPS, 2023. 3, 4 +[55] Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. In CVPR, 2024. 3 +[56] Mingyang Yi, Aoxue Li, Yi Xin, and Zhenguo Li. Towards understanding the working mechanism of text-to-image diffusion model. arXiv preprint arXiv:2405.15330, 2024. 2, 4 +[57] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. 3 +[58] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. NeurIPS, 2024. 2, 3, 5, 7, 1 +[59] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2 +[60] Shu Zhang, Xinyi Yang, Yihao Feng, Can Qin, Chia-Chih Chen, Ning Yu, Zeyuan Chen, Huan Wang, Silvio Savarese, Stefano Ermon, et al. Hive: Harnessing human feedback for instructional visual editing. In CVPR, 2024. 3, 7 +[61] Wentian Zhang, Haozhe Liu, Jinheng Xie, Francesco Faccio, Mike Zheng Shou, and Jurgen Schmidhuber. Cross-attention makes inference cumbersome in text-to-image diffusion models. arXiv preprint arXiv:2404.02747, 2024. 2, 4 +[62] Haozhe Zhao, Xiaojian Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. Ultraedit: Instruction-based fine-grained image editing at scale. arXiv preprint arXiv:2407.05282, 2024. 2, 3 \ No newline at end of file diff --git a/2025/SuperEdit_ Rectifying and Facilitating Supervision for Instruction-Based Image Editing/images.zip b/2025/SuperEdit_ Rectifying and Facilitating Supervision for Instruction-Based Image Editing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5f134d4ae7150a4e82178ed0f6888e5be1267eb1 --- /dev/null +++ b/2025/SuperEdit_ Rectifying and Facilitating Supervision for Instruction-Based Image Editing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:016ba0b27308c59b9c57696a4e50d6701ee65922f77a8b8f03bf8d4e75944f4a +size 1211466 diff --git a/2025/SuperEdit_ Rectifying and Facilitating Supervision for Instruction-Based Image Editing/layout.json b/2025/SuperEdit_ Rectifying and Facilitating Supervision for Instruction-Based Image Editing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4ecccf45db278ef197da295e7e2e7a192102f521 --- /dev/null +++ b/2025/SuperEdit_ Rectifying and Facilitating Supervision for Instruction-Based Image Editing/layout.json @@ -0,0 +1,8400 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 55, + 103, + 555, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 103, + 555, + 121 + ], + "spans": [ + { + "bbox": [ + 55, + 103, + 555, + 121 + ], + "type": "text", + "content": "SuperEdit: Rectifying and Facilitating Supervision for Instruction-Based Image Editing" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 220, + 128, + 235, + 141 + ], + "blocks": [ + { + "bbox": [ + 220, + 128, + 235, + 141 + ], + "lines": [ + { + "bbox": [ + 220, + 128, + 235, + 141 + ], + "spans": [ + { + "bbox": [ + 220, + 128, + 235, + 141 + ], + "type": "image", + "image_path": "c2ca5a6587dbdec196e1bd2af034b0cb961ad511d7ba97e98087c6cf70d746f1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 236, + 129, + 276, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 129, + 276, + 141 + ], + "spans": [ + { + "bbox": [ + 236, + 129, + 276, + 141 + ], + "type": "text", + "content": "Website" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 290, + 128, + 304, + 141 + ], + "blocks": [ + { + "bbox": [ + 290, + 128, + 304, + 141 + ], + "lines": [ + { + "bbox": [ + 290, + 128, + 304, + 141 + ], + "spans": [ + { + "bbox": [ + 290, + 128, + 304, + 141 + ], + "type": "image", + "image_path": "1dfe022054d2bc31d98f7b4664ca6b2ef1ac3ad32b5b48f00734411f4ab16e0b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 129, + 331, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 129, + 331, + 141 + ], + "spans": [ + { + "bbox": [ + 305, + 129, + 331, + 141 + ], + "type": "text", + "content": "Code" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 348, + 129, + 364, + 142 + ], + "blocks": [ + { + "bbox": [ + 348, + 129, + 364, + 142 + ], + "lines": [ + { + "bbox": [ + 348, + 129, + 364, + 142 + ], + "spans": [ + { + "bbox": [ + 348, + 129, + 364, + 142 + ], + "type": "image", + "image_path": "d3e90959e8e22205dbda6a58db2876dbbb512e05de6fbd9dbf9dfa236f9492a2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 365, + 129, + 390, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 129, + 390, + 141 + ], + "spans": [ + { + "bbox": [ + 365, + 129, + 390, + 141 + ], + "type": "text", + "content": "Data" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 150, + 549, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 150, + 549, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 150, + 549, + 178 + ], + "type": "text", + "content": "Ming Li" + }, + { + "bbox": [ + 61, + 150, + 549, + 178 + ], + "type": "inline_equation", + "content": "^{1,2,\\ddagger}" + }, + { + "bbox": [ + 61, + 150, + 549, + 178 + ], + "type": "text", + "content": ", Xin Gu" + }, + { + "bbox": [ + 61, + 150, + 549, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 150, + 549, + 178 + ], + "type": "text", + "content": ", Fan Chen" + }, + { + "bbox": [ + 61, + 150, + 549, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 150, + 549, + 178 + ], + "type": "text", + "content": ", Xiaoying Xing" + }, + { + "bbox": [ + 61, + 150, + 549, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 150, + 549, + 178 + ], + "type": "text", + "content": ", Longyin Wen" + }, + { + "bbox": [ + 61, + 150, + 549, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 150, + 549, + 178 + ], + "type": "text", + "content": ", Chen Chen" + }, + { + "bbox": [ + 61, + 150, + 549, + 178 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 150, + 549, + 178 + ], + "type": "text", + "content": ", Sijie Zhu" + }, + { + "bbox": [ + 61, + 150, + 549, + 178 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 61, + 150, + 549, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 150, + 549, + 178 + ], + "type": "text", + "content": "ByteDance Intelligent Creation Center for Research in Computer Vision, University of Central Florida" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 58, + 186, + 232, + 249 + ], + "blocks": [ + { + "bbox": [ + 58, + 186, + 232, + 249 + ], + "lines": [ + { + "bbox": [ + 58, + 186, + 232, + 249 + ], + "spans": [ + { + "bbox": [ + 58, + 186, + 232, + 249 + ], + "type": "image", + "image_path": "c0340ed1d5247b3e0f8c9581efff1a4ec5797b1263767e54034fb4c2cf301c49.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 232, + 186, + 407, + 250 + ], + "blocks": [ + { + "bbox": [ + 232, + 186, + 407, + 250 + ], + "lines": [ + { + "bbox": [ + 232, + 186, + 407, + 250 + ], + "spans": [ + { + "bbox": [ + 232, + 186, + 407, + 250 + ], + "type": "image", + "image_path": "ec81b6f49f3c3d93a85a8e1d7c0df06b602a80611dc0aebba05c0b9564ed0e6a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 58, + 250, + 231, + 316 + ], + "blocks": [ + { + "bbox": [ + 58, + 250, + 231, + 316 + ], + "lines": [ + { + "bbox": [ + 58, + 250, + 231, + 316 + ], + "spans": [ + { + "bbox": [ + 58, + 250, + 231, + 316 + ], + "type": "image", + "image_path": "7fe340698e5085dde31385bc7521f02a98d61a1cc9a8db24be8c7fb40023202d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 316, + 205, + 321 + ], + "lines": [ + { + "bbox": [ + 85, + 316, + 205, + 321 + ], + "spans": [ + { + "bbox": [ + 85, + 316, + 205, + 321 + ], + "type": "text", + "content": "Global: Transform the scene into a snowy winter scene" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 231, + 250, + 406, + 316 + ], + "blocks": [ + { + "bbox": [ + 231, + 250, + 406, + 316 + ], + "lines": [ + { + "bbox": [ + 231, + 250, + 406, + 316 + ], + "spans": [ + { + "bbox": [ + 231, + 250, + 406, + 316 + ], + "type": "image", + "image_path": "025353ddf90f7754c2b5b16880540198ad39bcc3750ba52e82369de167e7db07.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 268, + 316, + 367, + 321 + ], + "lines": [ + { + "bbox": [ + 268, + 316, + 367, + 321 + ], + "spans": [ + { + "bbox": [ + 268, + 316, + 367, + 321 + ], + "type": "text", + "content": "Replace: Replace the bow tie with a butterfly." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 58, + 321, + 231, + 382 + ], + "blocks": [ + { + "bbox": [ + 58, + 321, + 231, + 382 + ], + "lines": [ + { + "bbox": [ + 58, + 321, + 231, + 382 + ], + "spans": [ + { + "bbox": [ + 58, + 321, + 231, + 382 + ], + "type": "image", + "image_path": "b2e53b28238c2b7a879ed25964505f0ddd0abfc293eab28fe26a386f1f573554.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 382, + 362, + 388 + ], + "lines": [ + { + "bbox": [ + 94, + 382, + 362, + 388 + ], + "spans": [ + { + "bbox": [ + 94, + 382, + 362, + 388 + ], + "type": "text", + "content": "Style: Convert the image to a watercolor style Remove: Remove some clouds in the sky" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 135, + 388, + 328, + 396 + ], + "lines": [ + { + "bbox": [ + 135, + 388, + 328, + 396 + ], + "spans": [ + { + "bbox": [ + 135, + 388, + 328, + 396 + ], + "type": "text", + "content": "(a) Examples of Our Method on Real & High-resolution Image" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 231, + 321, + 406, + 382 + ], + "blocks": [ + { + "bbox": [ + 231, + 321, + 406, + 382 + ], + "lines": [ + { + "bbox": [ + 231, + 321, + 406, + 382 + ], + "spans": [ + { + "bbox": [ + 231, + 321, + 406, + 382 + ], + "type": "image", + "image_path": "dd8e57b09606df98dbb654124ffa8c514c0b3eefc2dd82a2c69beed868c003fc.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 407, + 194, + 553, + 304 + ], + "blocks": [ + { + "bbox": [ + 441, + 187, + 522, + 194 + ], + "lines": [ + { + "bbox": [ + 441, + 187, + 522, + 194 + ], + "spans": [ + { + "bbox": [ + 441, + 187, + 522, + 194 + ], + "type": "text", + "content": "Comparison with Existing Methods" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 407, + 194, + 553, + 304 + ], + "lines": [ + { + "bbox": [ + 407, + 194, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 407, + 194, + 553, + 304 + ], + "type": "image", + "image_path": "1a709fe36d69a2d9f4630a351d54b3ecfa99c371e4339b663c3435495418bfa4.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 307, + 544, + 314 + ], + "lines": [ + { + "bbox": [ + 419, + 307, + 544, + 314 + ], + "spans": [ + { + "bbox": [ + 419, + 307, + 544, + 314 + ], + "type": "text", + "content": "(b) Comparison with GPT-4o Evaluation" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 408, + 323, + 553, + 384 + ], + "blocks": [ + { + "bbox": [ + 408, + 323, + 553, + 384 + ], + "lines": [ + { + "bbox": [ + 408, + 323, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 408, + 323, + 553, + 384 + ], + "type": "image", + "image_path": "e9026c6f1f07d055b479d701aa26a475843a7322b08da88f137dbb67b5d2360b.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 418, + 388, + 542, + 396 + ], + "lines": [ + { + "bbox": [ + 418, + 388, + 542, + 396 + ], + "spans": [ + { + "bbox": [ + 418, + 388, + 542, + 396 + ], + "type": "text", + "content": "(c) Comparison with Human Evaluation" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 398, + 555, + 431 + ], + "lines": [ + { + "bbox": [ + 55, + 398, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 555, + 431 + ], + "type": "text", + "content": "Figure 1. (a) Our editing method works well with real and high-resolution images, handling various free-form edits (left) and local edits (right); (b) Compared to the current state-of-the-art SmartEdit, our method achieves a " + }, + { + "bbox": [ + 55, + 398, + 555, + 431 + ], + "type": "inline_equation", + "content": "9.19\\%" + }, + { + "bbox": [ + 55, + 398, + 555, + 431 + ], + "type": "text", + "content": " performance improvement with " + }, + { + "bbox": [ + 55, + 398, + 555, + 431 + ], + "type": "inline_equation", + "content": "30\\times" + }, + { + "bbox": [ + 55, + 398, + 555, + 431 + ], + "type": "text", + "content": " less training data and " + }, + { + "bbox": [ + 55, + 398, + 555, + 431 + ], + "type": "inline_equation", + "content": "13\\times" + }, + { + "bbox": [ + 55, + 398, + 555, + 431 + ], + "type": "text", + "content": " fewer model parameters; (c) Our method achieves better overall scores on the human evaluation results, indicating more precise editing capabilities." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 153, + 444, + 198, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 444, + 198, + 456 + ], + "spans": [ + { + "bbox": [ + 153, + 444, + 198, + 456 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 54, + 468, + 297, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 468, + 297, + 685 + ], + "spans": [ + { + "bbox": [ + 54, + 468, + 297, + 685 + ], + "type": "text", + "content": "Due to the challenges of manually collecting accurate editing data, existing datasets are typically constructed using various automated methods, leading to noisy supervision signals caused by the mismatch between editing instructions and original-edited image pairs. Recent efforts attempt to improve editing models through generating higher-quality edited images, pre-training on recognition tasks, or introducing vision-language models (VLMs) but fail to resolve this fundamental issue. In this paper, we offer a novel solution by constructing more effective editing instructions for given image pairs. This includes rectifying the editing instructions to better align with the original-edited image pairs and using contrastive editing instructions to further enhance their effectiveness. Specifically, we find that editing models exhibit specific generation attributes at different inference steps, independent of the text. Based on these prior attributes, we define a unified guide for VLMs to rectify editing instructions. However, there are some challenging editing scenarios that cannot be resolved solely with rectified instructions." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 445, + 556, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 556, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 556, + 601 + ], + "type": "text", + "content": "To this end, we further construct contrastive supervision signals with positive and negative instructions and introduce them into the model training using triplet loss, thereby further facilitating supervision effectiveness. Our method does not require the VLM modules or pre-training tasks used in previous work, offering a more direct and efficient way to provide better supervision signals, and providing a novel, simple, and effective solution for instruction-based image editing. Results on multiple benchmarks demonstrate that our method significantly outperforms existing approaches. Compared with previous SOTA SmartEdit, we achieve " + }, + { + "bbox": [ + 313, + 445, + 556, + 601 + ], + "type": "inline_equation", + "content": "9.19\\%" + }, + { + "bbox": [ + 313, + 445, + 556, + 601 + ], + "type": "text", + "content": " improvements on the Real-Edit benchmark with " + }, + { + "bbox": [ + 313, + 445, + 556, + 601 + ], + "type": "inline_equation", + "content": "30\\times" + }, + { + "bbox": [ + 313, + 445, + 556, + 601 + ], + "type": "text", + "content": " less training data and " + }, + { + "bbox": [ + 313, + 445, + 556, + 601 + ], + "type": "inline_equation", + "content": "13\\times" + }, + { + "bbox": [ + 313, + 445, + 556, + 601 + ], + "type": "text", + "content": " smaller model size. All data and models are open-sourced on Github for future research." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 314, + 621, + 390, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 621, + 390, + 633 + ], + "spans": [ + { + "bbox": [ + 314, + 621, + 390, + 633 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "In recent years, significant progress has been made in text-to-image (T2I) generation [10, 37, 40-42] due to the development of diffusion models [9, 21, 48, 49]. These T2I diffusion models can generate images that align with natural language descriptions while satisfying human perception and preferences. Consequently, numerous image editing" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 693, + 223, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 693, + 223, + 703 + ], + "spans": [ + { + "bbox": [ + 69, + 693, + 223, + 703 + ], + "type": "text", + "content": "* Corresponding author, sijiezhu@bytedance.com" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 69, + 703, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 703, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 69, + 703, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 69, + 703, + 295, + 713 + ], + "type": "text", + "content": " This work was done during the first author's internship at ByteDance USA" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19206" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "type": "text", + "content": "methods [4, 6, 19, 31] based on these models have been proposed to achieve various editing effects. Instruction-based methods [4, 12, 24] have become increasingly popular as they allow users to conveniently and easily modify images using language instructions without the need to provide masks, as required by mask-based methods [18, 28, 47, 53]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 147, + 298, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 147, + 298, + 422 + ], + "spans": [ + { + "bbox": [ + 56, + 147, + 298, + 422 + ], + "type": "text", + "content": "The training of instruction-based editing models requires the original-edited image pairs and corresponding editing instruction, making it difficult to manually create or collect a large amount of relevant data [58]. To address the issue of scarce training data, existing efforts [13, 19, 25, 62] have attempted to develop various automated pipelines to synthesize large datasets. Specifically, most methods first use large language models (LLMs) to modify the text descriptions of original images. The original images and modified texts are then input into various pre-trained diffusion models to automatically generate edited images. However, current text-to-image diffusion models struggle to fully correspond to input text prompts [15, 59]. This often changes parts of the original images that do not require editing, leading to misaligned editing instructions and original-edited image pairs, thus resulting in noisy supervision signals. To mitigate the potential issues of noisy supervision in image editing models, existing work has attempted to introduce additional recognition pre-training tasks for U-Net [43] such as semantic segmentation [14, 45], or replace CLIP [38] text encoder with vision-language models (VLMs) [12, 24] to better understand editing instructions from noisy supervision signals. However, these methods not only introduce significant computational overhead but also overlook the issue of noisy supervision signals." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 424, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 424, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 296, + 568 + ], + "type": "text", + "content": "In this paper, we focus on addressing the fundamental challenge by introducing more effective editing instructions, as demonstrated in Fig. 2. Our data-oriented method explores a different research question: how much performance improvement can be achieved solely by focusing on supervision signal quality and optimization in image editing? Surprisingly, SuperEdit outperforms existing methods in both GPT-4o and human evaluations, despite using less data and requiring no additional modules or pretraining as shown in Fig.1. This demonstrates that high-quality supervision signals can significantly compensate for architectural simplicity, achieving results comparable to or better than methods with more complex requirements." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 570, + 298, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 298, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 298, + 714 + ], + "type": "text", + "content": "Specifically, to enhance the effectiveness of supervision signals for instruction-based image editing methods, we propose using VLMs to rectify editing instructions, creating better-aligned instructions for the original-edited image pairs. However, determining which VLM to use for this task and how to establish a unified rectification method for various editing instructions remain unexplored problems. To address this, we first analyze the ability of different VLMs to understand the differences between original and edited images, showing that GPT-4o [1] is the most capable of rectifying editing instructions. Additionally, we observe that both editing models and text-to-image diffusion models share a similar prior, as shown in Fig. 4: different inference" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 328, + 69, + 431, + 224 + ], + "blocks": [ + { + "bbox": [ + 328, + 69, + 431, + 224 + ], + "lines": [ + { + "bbox": [ + 328, + 69, + 431, + 224 + ], + "spans": [ + { + "bbox": [ + 328, + 69, + 431, + 224 + ], + "type": "image", + "image_path": "588b8b2639c6ce4ed4512fc8e5f3bd26b3b1ce7fe508459c2ed64d91aa84f442.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 327, + 225, + 432, + 234 + ], + "lines": [ + { + "bbox": [ + 327, + 225, + 432, + 234 + ], + "spans": [ + { + "bbox": [ + 327, + 225, + 432, + 234 + ], + "type": "text", + "content": "(c) InstructDiffusion; Emu-Edit" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 236, + 555, + 291 + ], + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 291 + ], + "type": "text", + "content": "Figure 2. Unlike existing efforts that attempt to (a) scale up edited images with noisy supervision [4, 25], (b) introduce massive VLMs into editing model architecture [12, 24], and (c) perform additional pre-training tasks [14, 45], (d) we focus on improving the effectiveness of supervision signals, which is the fundamental issue of image editing." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 433, + 70, + 539, + 224 + ], + "blocks": [ + { + "bbox": [ + 433, + 70, + 539, + 224 + ], + "lines": [ + { + "bbox": [ + 433, + 70, + 539, + 224 + ], + "spans": [ + { + "bbox": [ + 433, + 70, + 539, + 224 + ], + "type": "image", + "image_path": "39c8d4fcbb602f3d1c7c3d7dbf54a8dd3d16cdeb40b2d534de7497da4668aa2a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 225, + 500, + 234 + ], + "lines": [ + { + "bbox": [ + 468, + 225, + 500, + 234 + ], + "spans": [ + { + "bbox": [ + 468, + 225, + 500, + 234 + ], + "type": "text", + "content": "(d) Ours" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 294, + 555, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 555, + 354 + ], + "type": "text", + "content": "stages correspond to the generation of different image attributes, independent of the input text prompt [2, 17, 19, 36, 46, 56, 61] or editing instructions. Inspired by this, we guide VLMs based on these attributes to establish a unified rectification guideline for various editing instructions, as demonstrated in Fig. 3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 355, + 556, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 556, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 556, + 533 + ], + "type": "text", + "content": "When training with only rectified editing instructions, we find that the editing model can better understand the editing commands but still faces challenges in handling complex scenarios. For example, when the original image contains multiple objects, the edit model struggles to perform an accurate editing function if the instructions modify only one of these objects. Additionally, inherent issues present in pre-trained text-to-image diffusion models [15, 22-24], such as difficulty in understanding quantity, position, or object relationships, persist in the editing models. To address these issues, we propose using contrastive supervision signals to further optimize the editing model. Specifically, we first construct incorrect editing instructions based on the rectified instructions to generate positive and negative samples. We then introduce a triplet loss to guide the model, thereby enhancing the effectiveness of supervision, as shown in Fig. 5." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 329, + 535, + 550, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 535, + 550, + 545 + ], + "spans": [ + { + "bbox": [ + 329, + 535, + 550, + 545 + ], + "type": "text", + "content": "In summary, our contributions are summarized as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 546, + 556, + 714 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 314, + 546, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 546, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 314, + 546, + 555, + 594 + ], + "type": "text", + "content": "- New Insight: We aim to address the noisy supervision problem that arises from the misalignment between editing instructions and original-edited image pairs, which is a fundamental issue overlooked by previous work, as shown in Fig. 2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 594, + 555, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 594, + 555, + 630 + ], + "spans": [ + { + "bbox": [ + 314, + 594, + 555, + 630 + ], + "type": "text", + "content": "- Rectifying Supervision: We leverage diffusion generation priors to guide the vision-language model to generate better-aligned editing instructions for original-edited image pairs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 631, + 555, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 631, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 314, + 631, + 555, + 666 + ], + "type": "text", + "content": "- Facilitating Supervision: We introduce contrastive supervision using triplet loss, enabling the editing model to learn from both positive and negative editing instructions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 666, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 666, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 314, + 666, + 556, + 714 + ], + "type": "text", + "content": "- Promising Results: We achieve significant improvements on multiple benchmarks without additional pre-training or VLM. Compared to SmartEdit [24], we achieved a " + }, + { + "bbox": [ + 314, + 666, + 556, + 714 + ], + "type": "inline_equation", + "content": "9.19\\%" + }, + { + "bbox": [ + 314, + 666, + 556, + 714 + ], + "type": "text", + "content": " improvement while reducing " + }, + { + "bbox": [ + 314, + 666, + 556, + 714 + ], + "type": "inline_equation", + "content": "30 \\times" + }, + { + "bbox": [ + 314, + 666, + 556, + 714 + ], + "type": "text", + "content": " data and " + }, + { + "bbox": [ + 314, + 666, + 556, + 714 + ], + "type": "inline_equation", + "content": "13 \\times" + }, + { + "bbox": [ + 314, + 666, + 556, + 714 + ], + "type": "text", + "content": " model parameters." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19207" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 138, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 138, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 138, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 251, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 251, + 104 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 251, + 104 + ], + "type": "text", + "content": "2.1. Image Editing with Diffusion Models" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 109, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 109, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 109, + 296, + 180 + ], + "type": "text", + "content": "Building on advancements in text-to-image (T2I) diffusion models [10, 37, 40-42, 44], recent research has explored them for image editing [4, 19]. Training-free methods [5, 19, 29, 31, 34, 51] typically achieve this by adjusting attentions in pre-trained T2I models, but have limited performance and generalization capabilities on various editing tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 181, + 296, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 181, + 296, + 265 + ], + "spans": [ + { + "bbox": [ + 55, + 181, + 296, + 265 + ], + "type": "text", + "content": "Training-based methods address these limitations with specialized editing models, which can be categorized into mask-based and instruction-based approaches. Mask-based methods [6, 18, 28, 47, 53, 57] enable fine-grained local edits with user-provided or predicted masks and corresponding text descriptions. However, it struggles with global image editing and is constrained by the lack of mask-based editing data [24]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 266, + 296, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 296, + 397 + ], + "type": "text", + "content": "Instruction-based methods directly accept textual commands, such as \"add a dog\", offering better editing flexibility and generalization. InstructPix2Pix [4] pioneered this paradigm by generating instruction-based editing data and modifying the conditions of T2I diffusion models. Building on this framework, subsequent work introduces vision-language models [12, 24, 33] or additional pre-training tasks for the denoising U-Net [14, 24, 43, 45] to enhance the understanding and reasoning of input conditions. However, these methods not only introduce substantial computational overhead but also overlook the fundamental noisy supervision issue." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 405, + 295, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 405, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 295, + 418 + ], + "type": "text", + "content": "2.2. Generating and Improving Editing Supervision" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 423, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 296, + 579 + ], + "type": "text", + "content": "Due to the difficulty of scaling instruction-based image editing data through manual collection, existing efforts [4, 13, 25, 58, 62] aim to automatically modify text descriptions of original images and generate edited images with T2I diffusion models. However, this approach often produces synthesized images that do not align with the editing instructions, as shown in Figure 3, resulting in noisy editing supervision signals [58, 62]. To address this, MagicBrush [58] manually filters out incorrect editing data, but it is hard to scale. Unlike existing methods focusing on edited image quality, we leverage diffusion prior and vision-language model (i.e., GPT-40 [1]) to create better-aligned instructions with original-edited image pairs, providing more accurate supervision." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 588, + 222, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 222, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 222, + 600 + ], + "type": "text", + "content": "2.3. Alignment of Diffusion Models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "The success of alignment training in large language models (LLMs) [26, 32, 39] has been applied to diffusion models for better image generation. This is achieved by maximizing reward scores [8, 27, 54] or the generation probability of the winner image in a pair [11, 52, 55]. In image editing, HIVE [60] and MultiReward [16] attempt to incorporate reward information into the text condition to align the editing model. In contrast, we guide the editing model by rectifying and constructing contrastive editing instructions, achieving more effective alignment." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 71, + 367, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 367, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 367, + 83 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 91, + 555, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 187 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 187 + ], + "type": "text", + "content": "In this section, we first introduce the most general image editing framework in Sec. 3.1. Then, we explain how to use diffusion priors to rectify editing instructions with the multimodal model (i.e., GPT-4o) in Sec. 3.2, thereby enhancing the accuracy of supervision signals. Finally, we describe how to construct contrastive supervision with both correct and incorrect editing instructions and integrate it into the editing model training using triplet loss in Sec. 3.3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 195, + 547, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 195, + 547, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 195, + 547, + 209 + ], + "type": "text", + "content": "3.1. Instruction-based Image Editing Framework" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 213, + 555, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 297 + ], + "type": "text", + "content": "InstructPix2Pix [4] pioneered instruction-based image editing, performing editing tasks by simultaneously taking the original image " + }, + { + "bbox": [ + 313, + 213, + 555, + 297 + ], + "type": "inline_equation", + "content": "C^I" + }, + { + "bbox": [ + 313, + 213, + 555, + 297 + ], + "type": "text", + "content": " and editing instructions " + }, + { + "bbox": [ + 313, + 213, + 555, + 297 + ], + "type": "inline_equation", + "content": "C^T" + }, + { + "bbox": [ + 313, + 213, + 555, + 297 + ], + "type": "text", + "content": " as input conditions to generate the edited image " + }, + { + "bbox": [ + 313, + 213, + 555, + 297 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 213, + 555, + 297 + ], + "type": "text", + "content": " from random noise " + }, + { + "bbox": [ + 313, + 213, + 555, + 297 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 213, + 555, + 297 + ], + "type": "text", + "content": ". Following the definition of DDPM [21], we randomly sample a timestep " + }, + { + "bbox": [ + 313, + 213, + 555, + 297 + ], + "type": "inline_equation", + "content": "t \\in T" + }, + { + "bbox": [ + 313, + 213, + 555, + 297 + ], + "type": "text", + "content": " during training, and then add corresponding noise " + }, + { + "bbox": [ + 313, + 213, + 555, + 297 + ], + "type": "inline_equation", + "content": "\\epsilon_t" + }, + { + "bbox": [ + 313, + 213, + 555, + 297 + ], + "type": "text", + "content": " to the edited image " + }, + { + "bbox": [ + 313, + 213, + 555, + 297 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 213, + 555, + 297 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 359, + 305, + 555, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 305, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 359, + 305, + 555, + 319 + ], + "type": "interline_equation", + "content": "x _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t}, \\quad \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, I), \\tag {1}", + "image_path": "7622d28c3e82d7b96f3a13b3f05df78496938164af40ddb3708ba9d834fedabf.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 328, + 554, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 554, + 398 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 328, + 554, + 398 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 328, + 554, + 398 + ], + "type": "text", + "content": " is a noise map sampled from a Gaussian distribution, and " + }, + { + "bbox": [ + 313, + 328, + 554, + 398 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t\\coloneqq \\prod_{s = 0}^t\\alpha_s" + }, + { + "bbox": [ + 313, + 328, + 554, + 398 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 328, + 554, + 398 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = 1 - \\beta_{t}" + }, + { + "bbox": [ + 313, + 328, + 554, + 398 + ], + "type": "text", + "content": " is a differentiable function of timestep " + }, + { + "bbox": [ + 313, + 328, + 554, + 398 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 328, + 554, + 398 + ], + "type": "text", + "content": ", which is determined by the denoising sampler such as DDPM [21]. Then the training objective of the editing model " + }, + { + "bbox": [ + 313, + 328, + 554, + 398 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 328, + 554, + 398 + ], + "type": "text", + "content": " is predicting the added noise at timestep " + }, + { + "bbox": [ + 313, + 328, + 554, + 398 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 328, + 554, + 398 + ], + "type": "text", + "content": ", which can be written as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 403, + 555, + 424 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 403, + 555, + 424 + ], + "spans": [ + { + "bbox": [ + 321, + 403, + 555, + 424 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t r a i n}} = \\mathbb {E} _ {x _ {t}, t, C ^ {I}, C ^ {T}, \\epsilon} \\left[ \\left\\| \\epsilon_ {\\theta} \\left(\\operatorname {c o n c a t} \\left(x _ {t}, C ^ {I}\\right), t, C ^ {T}\\right) - \\epsilon_ {t} \\right\\| _ {2} ^ {2} \\right], \\tag {2}", + "image_path": "534d8ccd93749947efec697e11de2936fe5b513c73558c9c37a82c9c46c38d78.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 432, + 555, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 555, + 455 + ], + "type": "text", + "content": "where concat refers to concatenating the image latents of noised edited image " + }, + { + "bbox": [ + 313, + 432, + 555, + 455 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 432, + 555, + 455 + ], + "type": "text", + "content": " and original image " + }, + { + "bbox": [ + 313, + 432, + 555, + 455 + ], + "type": "inline_equation", + "content": "c_{I}" + }, + { + "bbox": [ + 313, + 432, + 555, + 455 + ], + "type": "text", + "content": " in the channel dimension." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 463, + 544, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 544, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 544, + 475 + ], + "type": "text", + "content": "3.2. Rectifying Supervision with Diffusion Priors" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 480, + 556, + 720 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 556, + 720 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 556, + 720 + ], + "type": "text", + "content": "As shown in Fig. 3, existing image editing datasets [4, 13, 58] typically use only Steps 1 and 2: LLMs construct editing prompts and captions, and then text-to-image diffusion models synthesize edited images. However, diffusion models often fail to accurately follow prompts while maintaining image layout, creating mismatches between original-edited pairs and editing instructions, resulting in inaccurate supervision. While better supervision signals for text-to-image diffusion models are common in image generation [3, 50], this approach remains underexplored in image editing due to two challenges: (1) VLMs trained on single-image data struggle with multi-image inputs, and (2) editing instructions vary widely, making unified rectification guidelines difficult. To address these issues, we: (1) analyzed different VLMs' capabilities with multi-image inputs, finding GPT-4o most effective, and (2) discovered that timestep-specific roles in image generation also apply to editing, providing a foundation for a unified rectification method across various instructions (Fig. 3 and 4). Due to page limitations, our VLM analysis is in the Supplementary Material, while this section focuses on Diffusion Prior and Editing Instruction Rectification." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19208" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 246, + 217 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 246, + 217 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 246, + 217 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 246, + 217 + ], + "type": "image", + "image_path": "fdde02f13f77e9774b108c3b58180ea004f09ab08d1e62c594fabfdff486e951.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 220, + 555, + 265 + ], + "lines": [ + { + "bbox": [ + 55, + 220, + 555, + 265 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 555, + 265 + ], + "type": "text", + "content": "Figure 3. (a) Existing work primarily uses LLMs and diffusion models to automatically generate edited images. However, current diffusion models often fail to accurately follow text prompts while maintaining the input image's layout, resulting in mismatches between the original-edited image pairs and the editing instructions. (b) We perform instruction rectification (Step 3) based on the images constructed in Steps 1 and 2. We show VLMs can understand the differences between the images, enabling them to rectify editing instructions to be better aligned with original-edited image pairs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 250, + 70, + 445, + 217 + ], + "blocks": [ + { + "bbox": [ + 250, + 70, + 445, + 217 + ], + "lines": [ + { + "bbox": [ + 250, + 70, + 445, + 217 + ], + "spans": [ + { + "bbox": [ + 250, + 70, + 445, + 217 + ], + "type": "image", + "image_path": "f23c3efcdc9fefb30d130072c56bac743446b7a251c68c1d7ceb3b08c5514d80.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 443, + 76, + 550, + 205 + ], + "blocks": [ + { + "bbox": [ + 443, + 76, + 550, + 205 + ], + "lines": [ + { + "bbox": [ + 443, + 76, + 550, + 205 + ], + "spans": [ + { + "bbox": [ + 443, + 76, + 550, + 205 + ], + "type": "image", + "image_path": "1645bbc98dc6936728b9ff771ba16479949c73578c30bf58456f7f36312800d6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 56, + 284, + 293, + 524 + ], + "blocks": [ + { + "bbox": [ + 56, + 284, + 293, + 524 + ], + "lines": [ + { + "bbox": [ + 56, + 284, + 293, + 524 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 293, + 524 + ], + "type": "image", + "image_path": "5e45d22c74c34dc577fc4e7888b3561686cb3741c730bf1df9aaf04e785d8fd3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 525, + 297, + 604 + ], + "lines": [ + { + "bbox": [ + 55, + 525, + 297, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 297, + 604 + ], + "type": "text", + "content": "Figure 4. We show that the editing model follows consistent generation attributes at different sampling stages, independent of the editing instructions. The early, middle, and late sampling stages correspond to global, local, and detail changes, respectively, while style changes occur at all stages. All the generated images here are DDIM 30-step sampled final images. The orange progress bar and the grid progress bar represent the sampling stages with and without the editing instructions, respectively." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": "Diffusion Generation Priors. Previous work has shown that different timesteps play distinct roles in image generation for text-to-image diffusion models, regardless of the text prompt [2, 17, 19, 36, 46, 56, 61]. We find that this phenomenon also exists in instruction-based editing models and present examples based on pre-trained InstructPix2Pix [4], as shown in Fig. 4. Specifically, diffusion models focus on global layout in the early stages, local object attributes in the mid stages, and image details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 285, + 556, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 556, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 556, + 334 + ], + "type": "text", + "content": "in the late stages of sampling. This finding inspires us to guide VLMs based on these four generation attributes, establishing a unified rectification method for various editing instructions. We provide more analysis and details in the Supplementary Material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 349, + 556, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 556, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 556, + 482 + ], + "type": "text", + "content": "Editing Instruction Rectification. As demonstrated in Fig. 3, we extend the existing editing data generation pipeline by introducing our instruction rectification (Step 3). This process relies on the original edited image pairs obtained through Steps 1 and 2 from previous work. Specifically, we input original-edited image pairs into the vision-language model (i.e., GPT-4o) and instruct it to describe the changes in the edited image compared to the original image according to the above diffusion prior generation attributes. Finally, we use VLM to summarize the instructions and ensure that its length is less than the maximum length of CLIP text encoder, which is 77 tokens." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 491, + 555, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 555, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 555, + 516 + ], + "type": "text", + "content": "3.3. Facilitating Supervision with Contrastive Instructions" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 522, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 715 + ], + "type": "text", + "content": "Although using rectified editing instructions can significantly improve performance across various editing tasks, we find that editing models still struggle with closely related text instructions. For example, \"add a cat on the left side of the image\" and \"add two cats on the right side of the image\" might produce the same edited image. This indicates that inherent biases in pre-trained text-to-image diffusion models [15, 22], such as difficulties in understanding quantity, position, and spatial relationships, persist in editing models. More importantly, our experiments show that training models with rectified editing instructions does not resolve these challenges. To further facilitate supervision signal effectiveness, we drew on successful alignment experiences from large language models [1, 32, 39] and text-to-image diffusion models [7, 52, 54]: constructing positive and negative sample pairs and guiding the model to assign a higher generation probability to positive samples compared to negative ones." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19209" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 253, + 220 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 253, + 220 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 253, + 220 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 253, + 220 + ], + "type": "image", + "image_path": "162a52a8a1dcc266331549d208f7f177bbf2aa21ad63ea7dbd4fdce26317e909.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 224, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 55, + 224, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 555, + 281 + ], + "type": "text", + "content": "Figure 5. (a) Based on the rectified editing instruction and original-edited image pair, we utilize the Vision-Language Models (VLM) to generate various image-related wrong instructions. These involve random substitutions of quantities, spatial locations, and objects within the rectified editing instructions according to the original-edited images context; (b) During each training iteration, we randomly select one wrong instruction " + }, + { + "bbox": [ + 55, + 224, + 555, + 281 + ], + "type": "inline_equation", + "content": "c_{\\text{neg}}^T" + }, + { + "bbox": [ + 55, + 224, + 555, + 281 + ], + "type": "text", + "content": " and input it along with the rectified instruction " + }, + { + "bbox": [ + 55, + 224, + 555, + 281 + ], + "type": "inline_equation", + "content": "c_{pos}^T" + }, + { + "bbox": [ + 55, + 224, + 555, + 281 + ], + "type": "text", + "content": " into the editing model to obtain predicted noises. The goal is to make the rectified instruction's predicted noise " + }, + { + "bbox": [ + 55, + 224, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\epsilon_{pos}" + }, + { + "bbox": [ + 55, + 224, + 555, + 281 + ], + "type": "text", + "content": " closer to the sampled training diffusion noise " + }, + { + "bbox": [ + 55, + 224, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 224, + 555, + 281 + ], + "type": "text", + "content": ", while ensuring the noise from incorrect instructions " + }, + { + "bbox": [ + 55, + 224, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\text{neg}}" + }, + { + "bbox": [ + 55, + 224, + 555, + 281 + ], + "type": "text", + "content": " is farther. Best viewed in color." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 256, + 73, + 555, + 220 + ], + "blocks": [ + { + "bbox": [ + 256, + 73, + 555, + 220 + ], + "lines": [ + { + "bbox": [ + 256, + 73, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 256, + 73, + 555, + 220 + ], + "type": "image", + "image_path": "d2df2c9f86bef1789db5a2f08610bce45bd67df87c7d4970f19964fcf01d8296.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 295, + 296, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 295, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 295, + 296, + 547 + ], + "type": "text", + "content": "Constructing Contrastive Instructions. Unlike the standard alignment process for large language models or text-to-image diffusion models, it is challenging to generate different editing results from the same instruction to create positive and negative sample pairs for image editing tasks. To address this, we construct positive and negative editing instructions for alignment, thereby generating relatively positive and negative edited images. As shown in Fig. 5 (a), we use the original image, edited image, and rectified editing instruction as input. The VLM (i.e., GPT-4o) is used to modify attributes in the rectified editing instruction, such as quantity, spatial relationships, and object types, to create different wrong instructions. Here, we require VLM to modify only a single attribute from the rectified editing instruction in each wrong instruction, keeping most of the editing text unchanged. Since only a few words are replaced between the rectified instruction and the wrong instruction, the text embeddings produced by the CLIP text encoder that serve as input to the denoising model will also be similar. This ensures the task's learning difficulty, helping the model understand how subtle differences between the two editing instructions result in significantly different editing results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": "Facilitating Editing Models with Contrastive Instructions. Our key insight is that enhancing the effectiveness of supervision signals can improve various editing tasks without introducing additional model architectures or pre-training tasks. Therefore, we adhere strictly to the InstructPix2Pix [4] model architecture and training pipeline. To be specific, the inputs including the original image " + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "c^I" + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": ", edited image " + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": ", the rectified instruction " + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "c_{pos}^T" + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": ", and wrong editing instruction " + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "c_{neg}^T" + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": ". During training, we will add a sampled timestep " + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "t \\in T" + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": " to obtain the noised edited image " + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": " with Equation 1. Both the rectified and wrong editing instructions are fed into the denoising model to predict the final noises " + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{pos}" + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{neg}" + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": ", which are then used to" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 295, + 518, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 518, + 307 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 518, + 307 + ], + "type": "text", + "content": "construct positive and negative samples, respectively:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 369, + 314, + 555, + 329 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 314, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 369, + 314, + 555, + 329 + ], + "type": "interline_equation", + "content": "\\epsilon_ {p o s} = \\epsilon_ {\\theta} \\left(\\operatorname {c o n c a t} \\left(x _ {t}, c ^ {I}\\right), t, c _ {p o s} ^ {T}\\right), \\tag {3}", + "image_path": "01838e224f188a2c92f57c56067c8def5d841662bed31a6350c510664095d7d7.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 369, + 331, + 555, + 347 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 331, + 555, + 347 + ], + "spans": [ + { + "bbox": [ + 369, + 331, + 555, + 347 + ], + "type": "interline_equation", + "content": "\\epsilon_ {n e g} = \\epsilon_ {\\theta} \\big (\\operatorname {c o n c a t} \\left(x _ {t}, c ^ {I}\\right), t, c _ {n e g} ^ {T} \\big). \\tag {4}", + "image_path": "962ca0df29e4c303c081c817e7a73d8feb2c67239e6927c6d086330e314f6584.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 354, + 555, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 415 + ], + "type": "text", + "content": "After constructing the positive and negative sample pairs, we aim for the noise predicted by the positive editing instruction " + }, + { + "bbox": [ + 313, + 354, + 555, + 415 + ], + "type": "inline_equation", + "content": "\\epsilon_{pos}" + }, + { + "bbox": [ + 313, + 354, + 555, + 415 + ], + "type": "text", + "content": " to be closer to the true noise " + }, + { + "bbox": [ + 313, + 354, + 555, + 415 + ], + "type": "inline_equation", + "content": "\\epsilon_t" + }, + { + "bbox": [ + 313, + 354, + 555, + 415 + ], + "type": "text", + "content": " sampled during training, compared to the noise predicted by the wrong editing instruction " + }, + { + "bbox": [ + 313, + 354, + 555, + 415 + ], + "type": "inline_equation", + "content": "\\epsilon_{neg}" + }, + { + "bbox": [ + 313, + 354, + 555, + 415 + ], + "type": "text", + "content": ". This goal can be achieved through a triplet loss function:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 346, + 422, + 555, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 422, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 346, + 422, + 555, + 436 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t r i p l e t}} = \\max \\left\\{d \\left(\\epsilon_ {t}, \\epsilon_ {\\text {p o s}}\\right) - d \\left(\\epsilon_ {t}, \\epsilon_ {\\text {n e g}}\\right) + m, 0 \\right\\}, \\tag {5}", + "image_path": "335c462ed4cc554bc5449bac194029f3e04e61cac5418b1d9ebb46e3dcf71815.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 444, + 555, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 555, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 555, + 481 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 444, + 555, + 481 + ], + "type": "inline_equation", + "content": "d(x,y) = \\|\\mathbf{x} - \\mathbf{y}\\|_2^2" + }, + { + "bbox": [ + 313, + 444, + 555, + 481 + ], + "type": "text", + "content": " and margin " + }, + { + "bbox": [ + 313, + 444, + 555, + 481 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 444, + 555, + 481 + ], + "type": "text", + "content": " is a hyper-parameter. The final training loss is the combination of the original diffusion training loss and the triplet loss:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 339, + 490, + 555, + 503 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 490, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 339, + 490, + 555, + 503 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {t r a i n}} + \\lambda \\cdot \\mathcal {L} _ {\\text {t r i p l e t}}, \\text {w h e r e} \\mathcal {L} _ {\\text {t r a i n}} = d \\left(\\epsilon_ {t}, \\epsilon_ {\\text {p o s}}\\right). \\tag {6}", + "image_path": "4493ba2a46af4cee9f01dd3d38ead224bdab0a1660c4438ffecfcd56dea7b3cc.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 510, + 555, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 547 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 547 + ], + "type": "text", + "content": "Please note that the contrastive supervision signals are only used during the training phase. During inference, the editing model only requires one single input editing instruction." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 556, + 388, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 388, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 388, + 570 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 576, + 495, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 495, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 495, + 588 + ], + "type": "text", + "content": "4.1. Data Collection and Construction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": "To build a diverse dataset with various types of editing instructions, we need original and edited images from different data domains, as well as a wide variety of editing instructions. To achieve this, we sampled data from different public editing datasets to construct rectified and contrastive supervision signals. Specifically, we extracted 10,177, 8,807, and 21,016 editing pairs from InstructPix2Pix [4], MagicBrush [58], and Seed-Data-Edit [13], respectively, resulting in a total of 40,000 training samples. During extraction, we strive to ensure that the data for different types of editing tasks is as balanced as possible." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19210" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 69, + 539, + 674 + ], + "blocks": [ + { + "bbox": [ + 71, + 69, + 539, + 674 + ], + "lines": [ + { + "bbox": [ + 71, + 69, + 539, + 674 + ], + "spans": [ + { + "bbox": [ + 71, + 69, + 539, + 674 + ], + "type": "image", + "image_path": "35dbcee4f093c470ccb517a3cce40883521fa0057500e611f1e7679b3bc1a05b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 677, + 555, + 711 + ], + "lines": [ + { + "bbox": [ + 55, + 677, + 555, + 711 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 555, + 711 + ], + "type": "text", + "content": "Figure 6. Visual comparison with existing methods and the corresponding human-aligned GPT-4o evaluation scores (Following, Preserving, Quality Scores from left to right). We achieve better results while preserving the layout, quality, and details of the original image. Please note that we do not claim that our editing results are flawless. We provide more visual comparison results in the supplementary material." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19211" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 64, + 526, + 192 + ], + "blocks": [ + { + "bbox": [ + 82, + 64, + 526, + 192 + ], + "lines": [ + { + "bbox": [ + 82, + 64, + 526, + 192 + ], + "spans": [ + { + "bbox": [ + 82, + 64, + 526, + 192 + ], + "type": "table", + "html": "
MethodExtra ModulePretrain TasksEdit DataModel SizeFollowing ↑Preserving ↑Quality ↑Overall ↑
AccScoreAccScoreAccScoreAccScore
KOSMOS-G [33]✓✓9.0M1.9B51%2.829%1.4327%3.2029.0%2.48
MGIE [12]✓✓1.0M8.1B40%2.4345%2.7938%3.3541.0%2.86
SmartEdit [24]✓✓1.2M14.1B64%3.5066%3.7045%3.5658.3%3.59
MultiReward [16]✓✓320K1.2B63%3.3958%3.4354%3.8058.3%3.54
InstructDiffusion [14]X✓860K1.1B52%2.8754%3.1745%3.5850.3%3.21
InstructPix2Pix [4]XX300K1.1B52%2.9453%3.3150%3.6951.7%3.31
MagicBrush [58]XX310K1.1B51%2.9070%3.8550%3.6757.0%3.47
HIVE [60]XX1.1M1.1B54%2.9356%3.3653%3.7254.3%3.34
HQ-Edit [25]XX500K1.1B51%2.8416%1.6354%3.8440.3%2.77
SuperEdit (Ours)XX40K1.1B67%3.5977%4.1465%4.0169.7%3.91
", + "image_path": "0a37ff70ed3a6ccb3cbd08df08ba02fedc7c167a4cb8c5294066674c27a0a950.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 239, + 220, + 300 + ], + "blocks": [ + { + "bbox": [ + 57, + 239, + 220, + 300 + ], + "lines": [ + { + "bbox": [ + 57, + 239, + 220, + 300 + ], + "spans": [ + { + "bbox": [ + 57, + 239, + 220, + 300 + ], + "type": "image", + "image_path": "508db1f1a9247e288ae28c88a641cb6d98e7eb900f9d1be3a63b0af34aa08f3a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 115, + 300, + 163, + 308 + ], + "lines": [ + { + "bbox": [ + 115, + 300, + 163, + 308 + ], + "spans": [ + { + "bbox": [ + 115, + 300, + 163, + 308 + ], + "type": "text", + "content": "(a) Following Score" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 223, + 239, + 386, + 299 + ], + "blocks": [ + { + "bbox": [ + 223, + 239, + 386, + 299 + ], + "lines": [ + { + "bbox": [ + 223, + 239, + 386, + 299 + ], + "spans": [ + { + "bbox": [ + 223, + 239, + 386, + 299 + ], + "type": "image", + "image_path": "e1c861c97676cba85a10ded11521df6c0aaeb41c54fcf90de3ec9dd7ae60237d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 279, + 300, + 329, + 308 + ], + "lines": [ + { + "bbox": [ + 279, + 300, + 329, + 308 + ], + "spans": [ + { + "bbox": [ + 279, + 300, + 329, + 308 + ], + "type": "text", + "content": "(b) Preserving Score" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 309, + 555, + 343 + ], + "lines": [ + { + "bbox": [ + 55, + 309, + 555, + 343 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 555, + 343 + ], + "type": "text", + "content": "Figure 7. Human evaluation on three evaluation criteria for image editing effects. (a) Following: whether the edited image adheres to the editing instructions; (b) Preserving: whether the image structure outside of the editing instructions has been preserved; (c) Quality: whether the overall quality/aesthetics of the edited image has been degraded compared to the input image. Our SuperEdit achieves the best results on all of these metrics." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 389, + 239, + 550, + 299 + ], + "blocks": [ + { + "bbox": [ + 389, + 239, + 550, + 299 + ], + "lines": [ + { + "bbox": [ + 389, + 239, + 550, + 299 + ], + "spans": [ + { + "bbox": [ + 389, + 239, + 550, + 299 + ], + "type": "image", + "image_path": "d9cd5943e7f01b703b88967325315bdf68d820d296c8d04fdbef5ecaae428227.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 300, + 487, + 308 + ], + "lines": [ + { + "bbox": [ + 446, + 300, + 487, + 308 + ], + "spans": [ + { + "bbox": [ + 446, + 300, + 487, + 308 + ], + "type": "text", + "content": "(c) Quality Score" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 58, + 355, + 293, + 395 + ], + "blocks": [ + { + "bbox": [ + 55, + 195, + 555, + 231 + ], + "lines": [ + { + "bbox": [ + 55, + 195, + 555, + 231 + ], + "spans": [ + { + "bbox": [ + 55, + 195, + 555, + 231 + ], + "type": "text", + "content": "Table 1. Comparison with instruction-based image editing methods on Real-Edit benchmark [16]. Compared to existing work, our method achieves state-of-the-art performance across all metrics using a small amount of high-quality editing data without introducing additional models or pre-training tasks. Please note that the scores range from 0 to 5. " + }, + { + "bbox": [ + 55, + 195, + 555, + 231 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 55, + 195, + 555, + 231 + ], + "type": "text", + "content": " denotes a higher result is better. All baseline results are cited from the MultiReward [16] paper." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 355, + 293, + 395 + ], + "lines": [ + { + "bbox": [ + 58, + 355, + 293, + 395 + ], + "spans": [ + { + "bbox": [ + 58, + 355, + 293, + 395 + ], + "type": "table", + "html": "
Following ↑Preserving ↑Quality ↑Overall ↑
InstructPix2Pix [4]2.412.622.442.49
SmartEdit-13B [24]3.093.062.632.93
SuperEdit3.18+1.80%3.86+16.00%3.37+14.80%3.47+10.80%
", + "image_path": "9b93bcc462f1625d125bf5a8f2dca2f5ebaac0d69bb164fe37a768d9e0bb4580.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 396, + 295, + 431 + ], + "lines": [ + { + "bbox": [ + 55, + 396, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 295, + 431 + ], + "type": "text", + "content": "Table 2. Human evaluation results on Real-Edit [16] benchmark. All the human-evaluated scores range from 0 to 5. Overall represents the average score of Following, Preserving, and Quality scores." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 441, + 296, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 296, + 525 + ], + "type": "text", + "content": "We then applied our proposed methods in Sec. 3 to rectify and construct contrastive editing instructions for these training samples. Since the MagicBrush data has been manually verified, we skip the rectification step for this dataset and directly construct contrastive supervision based on the original editing instructions. For Seed-Data-Edit dataset, we only sample images from the first part of data without human editing instructions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 538, + 182, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 182, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 182, + 552 + ], + "type": "text", + "content": "4.2. Experimental Settings" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 557, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 297, + 714 + ], + "type": "text", + "content": "Evaluation Benchmarks and Metrics. To more accurately assess the effectiveness of various editing models, we conducted assessments on the Real-Edit benchmark [16], which is a human-aligned evaluation benchmark with GPT-4o scoring. Specifically, MultiReward [16] uses high-resolution images from the Unsplash community as a test dataset and combines them with GPT-4o [1] to create an automated evaluation method for single-turn editing. It assesses edited images in terms of accuracy (\\%) and scores (from 0 to 5), evaluating whether they adhere to the editing instructions (Following), whether the image structure outside of the editing instructions has been preserved (Preserving), and whether the overall quality/aesthetics of the edited image has been degraded compared to the original one (Quality)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 357, + 440, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 440, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 440, + 370 + ], + "type": "text", + "content": "4.3. Experimental Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 375, + 555, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 555, + 543 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 555, + 543 + ], + "type": "text", + "content": "Comparison on Real-Edit Benchmark. In Tab. 1, we present the quantitative results of editing effectiveness on the Real-Edit benchmark [16]. Without introducing additional parameters or pre-training stages, our method achieves the best results in the three GPT-4o automated evaluation metrics: Following, Preserving, and Quality, each of which includes percentage accuracy (Acc) and scores (from 0 to 5). For example, compared to SmartEdit [24], which introduces an additional 13B vision-language model (i.e., LLaVA [30]) to the 1.1B InstructPix2Pix [4] framework, we achieved improvements of " + }, + { + "bbox": [ + 313, + 375, + 555, + 543 + ], + "type": "inline_equation", + "content": "11.4\\%" + }, + { + "bbox": [ + 313, + 375, + 555, + 543 + ], + "type": "text", + "content": " Overall Score. This suggests that given accurate and effective supervision signals, the trained editing model can understand and successfully execute the editing instructions, without the need for additional vision-language models." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 544, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 556, + 665 + ], + "type": "text", + "content": "It is worth noting that unlike existing image editing methods, which often show improvement in a single metric while others remain unchanged or worsen, our method achieves comprehensive and significant advancements across all three metrics. This indicates that improving the effectiveness of supervision signals can accurately execute editing instructions while reducing disruption to other non-edited parts of the image, and preserving the quality and aesthetics of the original images. Specifically, we surpassed the current best methods by " + }, + { + "bbox": [ + 313, + 544, + 556, + 665 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 313, + 544, + 556, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 544, + 556, + 665 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 313, + 544, + 556, + 665 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 544, + 556, + 665 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 313, + 544, + 556, + 665 + ], + "type": "text", + "content": " Acc results in Following, Preserving, and Quality, respectively." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Human Evaluation We also conduct a comprehensive human evaluation on Real-Edit benchmarks [16] in Tab. 2 and Fig. 7. The assessment involved 15 experienced evaluators" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19212" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 252 + ], + "type": "text", + "content": "who rated edited images based on three critical metrics: instruction faithfulness (Following), preservation of irrelevant content (Preserving), and visual quality (Quality). The results of this manual evaluation demonstrate strong consistency with the GPT-4o scoring results shown in Tab. 1, with Cohen's Kappa=0.89, Krippendorff's Alpha=0.87. This high alignment thoroughly validates that our proposed SuperEdit significantly outperforms existing methods across all evaluation criteria. Specifically, our SuperEdit surpasses the previous state-of-the-art method SmartEdit [24] by " + }, + { + "bbox": [ + 55, + 72, + 297, + 252 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 55, + 72, + 297, + 252 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 297, + 252 + ], + "type": "inline_equation", + "content": "16\\%" + }, + { + "bbox": [ + 55, + 72, + 297, + 252 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 297, + 252 + ], + "type": "inline_equation", + "content": "14.8\\%" + }, + { + "bbox": [ + 55, + 72, + 297, + 252 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 72, + 297, + 252 + ], + "type": "inline_equation", + "content": "10.8\\%" + }, + { + "bbox": [ + 55, + 72, + 297, + 252 + ], + "type": "text", + "content": " on Following, Preserving, Quality, and Overall scores, respectively. These substantial improvements not only confirm the effectiveness of our approach but also establish SuperEdit as a new benchmark in instruction-guided image editing, achieving superior performance while requiring significantly less training data and cost." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 259, + 298, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 259, + 298, + 559 + ], + "spans": [ + { + "bbox": [ + 56, + 259, + 298, + 559 + ], + "type": "text", + "content": "Visual Comparison with State-of-the-art Methods. We show the visual comparison with existing image editing methods in Fig. 6. Compared to existing instruction-based editing methods, our approach not only better understands and executes editing instructions but also preserves the original image's layout and quality more effectively, thereby significantly outperforming previous methods. For example, with the instruction \"Replace the tiger with a lion, maintaining the same position in the water\" our SuperEdit method achieved superior results (4.8/4.8/4.8) compared to SmartEdit (4.8/4.8/2.5) and other methods. Additionally, our method improves the model's comprehension of editing instructions. For the instruction \"Change the background to a sandy beach with the ocean in the distance\" our method received perfect scores (4.8/5.0/5.0) while SmartEdit only achieved (5.0/4.8/4.0). Similarly, for style transformation instructions like \"Change the image style to look like an impressionist painting style\" SuperEdit significantly outperformed SmartEdit with scores of (4.8/4.8/4.8) versus (1.0/4.8/4.8), demonstrating our method's superior ability to handle complex artistic transformations. Even more impressively, for scene transformation tasks like \"Transform the entire scene to a winter setting with snow covering the houses, trees, and boat\", our SuperEdit achieved (5.0/4.8/4.8) while SmartEdit only obtained (2.0/4.5/4.5). We provide more visual comparisons with other instruction-based image editing methods in the Supplementary Material." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 564, + 149, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 149, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 149, + 577 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "text", + "content": "Ablation on the Rectified and Contrastive Instructions. Considering that the Real-Edit [16] benchmark is evaluated by GPT-4o [1], and its evaluation results closely align with human ratings [16], we choose this benchmark to conduct ablation experiments in Tab. 3. Compared to the original 40K training data, our training data with rectified editing instructions significantly improves all the performance of the editing model. Specifically, our approach improves scores by 0.95, 0.79, and 0.11, and accuracy by " + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "inline_equation", + "content": "21\\%" + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "inline_equation", + "content": "22\\%" + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "text", + "content": " in these three metrics, respectively. In addition, editing performance is further enhanced by incorporating contrastive supervision signals. Compared to using only" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 317, + 71, + 553, + 121 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 553, + 121 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 553, + 121 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 553, + 121 + ], + "type": "table", + "html": "
Rectified InstructionContrastive InstructionsFollowing↑Preserving↑Quality↑
AccScoreAccScoreAccScore
XX41%2.4553%3.2761%3.90
✓X62%3.4075%4.0665%4.01
✓✓67%3.5977%4.1465%4.01
", + "image_path": "7d06eb71ef8ba6b2484c13c9d645cf740a18b4e0589373fcb60be2732f31430d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 154, + 556, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 154, + 556, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 556, + 228 + ], + "type": "text", + "content": "rectified editing instructions, the introduction of contrastive supervision signals improves the following and preserving scores by 0.19 and 0.08, and accuracy by " + }, + { + "bbox": [ + 313, + 154, + 556, + 228 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 154, + 556, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 154, + 556, + 228 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 313, + 154, + 556, + 228 + ], + "type": "text", + "content": ", while maintaining the quality accuracy and score. In summary, both the introduction of rectified editing instructions and contrastive editing instructions improve the overall performance of the editing model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 234, + 556, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 556, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 556, + 390 + ], + "type": "text", + "content": "Ablation on Data Scaling. We investigated the impact of training data volume on model performance by experimenting with datasets ranging from 5k to 40k samples. Tab. 4 shows consistent improvements across all metrics as training data increases. With just 5k samples, our model achieves reasonable performance (54.7% accuracy, 3.42 overall score), but scaling to 40k samples yields substantial gains (69.7% accuracy, 3.91 overall score). The most significant improvements appear in the Preserving and Quality metrics, with 10% and 15%, respectively. This upward trend across all data points demonstrates that SuperEdit effectively leverages additional training examples without performance saturation, suggesting potential for further gains with larger datasets." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 394, + 552, + 456 + ], + "blocks": [ + { + "bbox": [ + 313, + 125, + 555, + 147 + ], + "lines": [ + { + "bbox": [ + 313, + 125, + 555, + 147 + ], + "spans": [ + { + "bbox": [ + 313, + 125, + 555, + 147 + ], + "type": "text", + "content": "Table 3. Ablation study on our methods. Both rectified and contrastive editing instructions achieved improvements across all metrics." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 394, + 552, + 456 + ], + "lines": [ + { + "bbox": [ + 317, + 394, + 552, + 456 + ], + "spans": [ + { + "bbox": [ + 317, + 394, + 552, + 456 + ], + "type": "table", + "html": "
Data SizeFollowing ↑Preserving ↑Quality ↑Overall ↑
AccScoreAccScoreAccScoreAccScore
5k49%2.8760%3.7155%3.6954.7%3.42
10k57%3.2671%3.7658%3.8762.0%3.63
20k64%3.4072%4.0263%3.9466.3%3.79
40k67%3.5977%4.1465%4.0169.7%3.91
", + "image_path": "b17783e3234fd5a6fa44150a75cb6b7d3b8e3f988e3eaaccd0fe8d4a0e1cc5fc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 321, + 456, + 545, + 466 + ], + "lines": [ + { + "bbox": [ + 321, + 456, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 321, + 456, + 545, + 466 + ], + "type": "text", + "content": "Table 4. Ablation study on data scaling results on Real-Edit [16]." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 478, + 384, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 384, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 384, + 491 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "type": "text", + "content": "In this paper, we re-examine image editing models from the perspective of enhancing supervision signals, finding that existing models have not adequately addressed this challenge, resulting in suboptimal performance. We introduce a unified editing instruction rectification guideline based on diffusion priors that better aligns instructions with original-edited image pairs, thereby enhancing supervision effectiveness. We also construct contrastive editing instructions allowing models to learn from both positive and negative examples. Our data-oriented approach explores an important but overlooked research question: What level of performance can be achieved with minimal architectural modifications by primarily focusing on supervision quality and optimization? Remarkably, under both GPT-4o and human evaluation, our method outperforms existing approaches despite using less data and requiring no architectural modifications or additional pretraining. This shows high-quality supervision signals can effectively compensate for architectural simplicity, offering valuable new perspectives for image editing research." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19213" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 112, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 112, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 112, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "type": "text", + "content": "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 191, + 296, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 296, + 235 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 296, + 235 + ], + "type": "text", + "content": "[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 236, + 296, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 296, + 267 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 296, + 267 + ], + "type": "text", + "content": "[4] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 2, 3, 4, 5, 7, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 270, + 295, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 295, + 312 + ], + "type": "text", + "content": "[5] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In ICCV, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 313, + 295, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 313, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 313, + 295, + 346 + ], + "type": "text", + "content": "[6] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. In ICLR, 2023. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 346, + 295, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 295, + 378 + ], + "type": "text", + "content": "[7] Yusuf Dalva and Pinar Yanardag. Noiseclar: A contrastive learning approach for unsupervised discovery of interpretable directions in diffusion models. In CVPR, 2024. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 380, + 295, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 295, + 422 + ], + "type": "text", + "content": "[8] Fei Deng, Qifei Wang, Wei Wei, Tingbo Hou, and Matthias Grundmann. Prdp: Proximal reward difference prediction for large-scale reward finetuning of diffusion models. In CVPR, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 425, + 295, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 425, + 295, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 425, + 295, + 446 + ], + "type": "text", + "content": "[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 447, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 447, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 57, + 447, + 295, + 491 + ], + "type": "text", + "content": "[10] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 491, + 295, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 491, + 295, + 535 + ], + "spans": [ + { + "bbox": [ + 57, + 491, + 295, + 535 + ], + "type": "text", + "content": "[11] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Reinforcement learning for fine-tuning text-to-image diffusion models. NeurIPS, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 536, + 295, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 295, + 568 + ], + "type": "text", + "content": "[12] Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding instruction-based image editing via multimodal large language models. In ICLR, 2024. 2, 3, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 569, + 295, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 569, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 57, + 569, + 295, + 602 + ], + "type": "text", + "content": "[13] Yuying Ge, Sijie Zhao, Chen Li, Yixiao Ge, and Ying Shan. Seed-data-edit technical report: A hybrid dataset for instructional image editing. arXiv preprint arXiv:2405.04007, 2024. 2, 3, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 603, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 295, + 646 + ], + "type": "text", + "content": "[14] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, et al. Instructdiffusion: A generalist modeling interface for vision tasks. In CVPR, 2024. 2, 3, 7, 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 647, + 295, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 680 + ], + "type": "text", + "content": "[15] Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. NeurIPS, 2024. 2, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[16] Xin Gu, Ming Li, Libo Zhang, Fan Chen, Longyin Wen, Tiejian Luo, and Sijie Zhu. Multi-reward as condition for instruction-based image editing. In ICLR, 2025. 3, 7, 8, 1" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "type": "text", + "content": "[17] Qin Guo and Tianwei Lin. Focus on your instruction: Fine-grained and multi-instruction image editing by attention modulation. In CVPR, 2024. 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 108, + 555, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 555, + 140 + ], + "type": "text", + "content": "[18] Qin Guo and Tianwei Lin. Focus on your instruction: Fine-grained and multi-instruction image editing by attention modulation. In CVPR, 2024. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 142, + 554, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 142, + 554, + 175 + ], + "spans": [ + { + "bbox": [ + 317, + 142, + 554, + 175 + ], + "type": "text", + "content": "[19] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In ICLR, 2023. 2, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 177, + 554, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 177, + 554, + 198 + ], + "spans": [ + { + "bbox": [ + 316, + 177, + 554, + 198 + ], + "type": "text", + "content": "[20] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 201, + 554, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 201, + 554, + 221 + ], + "spans": [ + { + "bbox": [ + 316, + 201, + 554, + 221 + ], + "type": "text", + "content": "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 224, + 554, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 224, + 554, + 257 + ], + "spans": [ + { + "bbox": [ + 316, + 224, + 554, + 257 + ], + "type": "text", + "content": "[22] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv preprint arXiv:2403.05135, 2024. 2, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 258, + 555, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 258, + 555, + 292 + ], + "spans": [ + { + "bbox": [ + 316, + 258, + 555, + 292 + ], + "type": "text", + "content": "[23] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. NeurIPS, 2023." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 293, + 554, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 293, + 554, + 346 + ], + "spans": [ + { + "bbox": [ + 316, + 293, + 554, + 346 + ], + "type": "text", + "content": "[24] Yuzhou Huang, Liangbin Xie, Xintao Wang, Ziyang Yuan, Xiaodong Cun, Yixiao Ge, Jiantao Zhou, Chao Dong, Rui Huang, Ruimao Zhang, et al. Smartedit: Exploring complex instruction-based image editing with multimodal large language models. In CVPR, 2024. 2, 3, 7, 8, 1, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 350, + 554, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 554, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 554, + 392 + ], + "type": "text", + "content": "[25] Mude Hui, Siwei Yang, Bingchen Zhao, Yichun Shi, Heng Wang, Peng Wang, Yuyin Zhou, and Cihang Xie. HQ-edit: A high-quality dataset for instruction-based image editing. arXiv preprint arXiv:2404.09990, 2024. 2, 3, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 395, + 555, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 395, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 395, + 555, + 449 + ], + "type": "text", + "content": "[26] Harrison Lee, Samrat Phatale, Hassan Mansoor, Thomas Mesnard, Johan Ferret, Kellie Lu, Colton Bishop, Ethan Hall, Victor Carbune, Abhinav Rastogi, et al. Rlaif: Scaling reinforcement learning from human feedback with ai feedback. arXiv preprint arXiv:2309.00267, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 452, + 554, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 452, + 554, + 494 + ], + "spans": [ + { + "bbox": [ + 316, + 452, + 554, + 494 + ], + "type": "text", + "content": "[27] Ming Li, Taojiannan Yang, Huafeng Kuang, Jie Wu, Zhaoning Wang, Xuefeng Xiao, and Chen Chen. Controlnet++: Improving conditional controls with efficient consistency feedback. In ECCV, 2024. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 497, + 554, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 497, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 497, + 554, + 529 + ], + "type": "text", + "content": "[28] Yuanze Lin, Yi-Wen Chen, Yi-Hsuan Tsai, Lu Jiang, and Ming-Hsuan Yang. Text-driven image editing via learnable regions. In CVPR, 2024. 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 532, + 554, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 532, + 554, + 564 + ], + "spans": [ + { + "bbox": [ + 316, + 532, + 554, + 564 + ], + "type": "text", + "content": "[29] Bingyan Liu, Chengyu Wang, Tingfeng Cao, Kui Jia, and Jun Huang. Towards understanding cross and self-attention in stable diffusion for text-guided image editing. In CVPR, 2024. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 566, + 555, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 566, + 555, + 587 + ], + "spans": [ + { + "bbox": [ + 316, + 566, + 555, + 587 + ], + "type": "text", + "content": "[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2024. 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 590, + 554, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 554, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 554, + 632 + ], + "type": "text", + "content": "[31] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 2, 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 635, + 554, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 554, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 554, + 678 + ], + "type": "text", + "content": "[32] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. NeurIPS, 2022. 3, 4" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "text", + "content": "[33] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-g: Generating images in context with multimodal large language models. In ICLR, 2024. 3, 7" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19214" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "text", + "content": "[34] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 149 + ], + "type": "text", + "content": "[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 295, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 295, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 295, + 183 + ], + "type": "text", + "content": "[36] Or Patashnik, Daniel Garibi, Idan Azuri, Hadar Averbuch-Elor, and Daniel Cohen-Or. Localizing object-level shape variations with text-to-image diffusion models. In ICCV, 2023. 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 295, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 295, + 227 + ], + "type": "text", + "content": "[37] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2023. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 228, + 295, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 295, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 295, + 271 + ], + "type": "text", + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 272, + 295, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 295, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 295, + 315 + ], + "type": "text", + "content": "[39] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. NeurIPS, 2024. 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 316, + 295, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 295, + 348 + ], + "type": "text", + "content": "[40] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 350, + 295, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 295, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 295, + 381 + ], + "type": "text", + "content": "[41] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 383, + 295, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 295, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 295, + 414 + ], + "type": "text", + "content": "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 415, + 295, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 295, + 446 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 295, + 446 + ], + "type": "text", + "content": "[43] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 449, + 295, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 295, + 502 + ], + "type": "text", + "content": "[44] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurlPS, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 504, + 295, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 295, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 295, + 547 + ], + "type": "text", + "content": "[45] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and generation tasks. In CVPR, 2024. 2, 3, 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 548, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 295, + 590 + ], + "type": "text", + "content": "[46] Enis Simsar, Alessio Tonioni, Yongqin Xian, Thomas Hofmann, and Federico Tombari. Lime: localized image editing via attention regularization in diffusion models. arXiv preprint arXiv:2312.09256, 2023. 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 592, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 295, + 635 + ], + "type": "text", + "content": "[47] Jaskirat Singh, Jianming Zhang, Qing Liu, Cameron Smith, Zhe Lin, and Liang Zheng. Smartmask: Context aware high-fidelity mask generation for fine-grained object insertion and layout control. In CVPR, 2024. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 636, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 295, + 668 + ], + "type": "text", + "content": "[48] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 670, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 690 + ], + "type": "text", + "content": "[49] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "text", + "content": "[50] Kcolors Team. Kcolors: Effective training of diffusion model for photorealistic text-to-image synthesis. arXiv preprint, 2024. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 555, + 544 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 315, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 555, + 106 + ], + "type": "text", + "content": "[51] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In CVPR, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 107, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 107, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 315, + 107, + 555, + 150 + ], + "type": "text", + "content": "[52] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In CVPR, 2024. 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 152, + 554, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 152, + 554, + 184 + ], + "spans": [ + { + "bbox": [ + 315, + 152, + 554, + 184 + ], + "type": "text", + "content": "[53] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In CVPR, 2023. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 186, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 186, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 315, + 186, + 555, + 228 + ], + "type": "text", + "content": "[54] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. NeurlPS, 2023. 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 230, + 555, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 230, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 315, + 230, + 555, + 273 + ], + "type": "text", + "content": "[55] Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. In CVPR, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 274, + 554, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 274, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 315, + 274, + 554, + 308 + ], + "type": "text", + "content": "[56] Mingyang Yi, Aoxue Li, Yi Xin, and Zhenguo Li. Towards understanding the working mechanism of text-to-image diffusion model. arXiv preprint arXiv:2405.15330, 2024. 2, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 309, + 555, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 309, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 315, + 309, + 555, + 352 + ], + "type": "text", + "content": "[57] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 354, + 554, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 354, + 554, + 387 + ], + "spans": [ + { + "bbox": [ + 315, + 354, + 554, + 387 + ], + "type": "text", + "content": "[58] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. NeurIPS, 2024. 2, 3, 5, 7, 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 388, + 555, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 388, + 555, + 410 + ], + "spans": [ + { + "bbox": [ + 315, + 388, + 555, + 410 + ], + "type": "text", + "content": "[59] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 411, + 555, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 411, + 555, + 454 + ], + "spans": [ + { + "bbox": [ + 315, + 411, + 555, + 454 + ], + "type": "text", + "content": "[60] Shu Zhang, Xinyi Yang, Yihao Feng, Can Qin, Chia-Chih Chen, Ning Yu, Zeyuan Chen, Huan Wang, Silvio Savarese, Stefano Ermon, et al. Hive: Harnessing human feedback for instructional visual editing. In CVPR, 2024. 3, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 315, + 456, + 555, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 456, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 315, + 456, + 555, + 499 + ], + "type": "text", + "content": "[61] Wentian Zhang, Haozhe Liu, Jinheng Xie, Francesco Faccio, Mike Zheng Shou, and Jurgen Schmidhuber. Cross-attention makes inference cumbersome in text-to-image diffusion models. arXiv preprint arXiv:2404.02747, 2024. 2, 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 315, + 501, + 555, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 501, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 315, + 501, + 555, + 544 + ], + "type": "text", + "content": "[62] Haozhe Zhao, Xiaojian Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. Ultraedit: Instruction-based fine-grained image editing at scale. arXiv preprint arXiv:2407.05282, 2024. 2, 3" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19215" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SuperEvent_ Cross-Modal Learning of Event-based Keypoint Detection for SLAM/2e684548-d342-4d0d-9d82-edd690c021a1_content_list.json b/2025/SuperEvent_ Cross-Modal Learning of Event-based Keypoint Detection for SLAM/2e684548-d342-4d0d-9d82-edd690c021a1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..69b37a5b85c4fa4505a0411f4e3c40e996442a71 --- /dev/null +++ b/2025/SuperEvent_ Cross-Modal Learning of Event-based Keypoint Detection for SLAM/2e684548-d342-4d0d-9d82-edd690c021a1_content_list.json @@ -0,0 +1,1616 @@ +[ + { + "type": "text", + "text": "SuperEvent: Cross-Modal Learning of Event-based Keypoint Detection for SLAM", + "text_level": 1, + "bbox": [ + 89, + 130, + 906, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yannick Burkhardt $^{1,2,3}$", + "bbox": [ + 197, + 179, + 380, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Simon Schaefer $^{1,2,3}$", + "bbox": [ + 401, + 179, + 557, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Stefan Leutenegger $^{1,2,3}$", + "bbox": [ + 598, + 179, + 782, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 233, + 326, + 248 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Event-based keypoint detection and matching holds significant potential, enabling the integration of event sensors into highly optimized Visual SLAM systems developed for frame cameras over decades of research. Unfortunately, existing approaches struggle with the motion-dependent appearance of keypoints and the complex noise prevalent in event streams, resulting in severely limited feature matching capabilities and poor performance on downstream tasks. To mitigate this problem, we propose SuperEvent, a data-driven approach to predict stable keypoints with expressive descriptors. Due to the absence of event datasets with ground truth keypoints labels, we leverage existing frame-based keypoint detectors on readily available event-aligned and synchronized gray-scale frames for self-supervision: we generate temporally sparse keypoint pseudo-labels considering that events are a product of both scene appearance and camera motion. Combined with our novel, information-rich event representation, we enable SuperEvent to effectively learn robust keypoint detection and description in event streams. Finally, we demonstrate the usefulness of SuperEvent by its integration into a modern sparse keypoint and descriptor-based SLAM framework originally developed for traditional cameras, surpassing the state-of-the-art in event-based SLAM by a wide margin. Source code is available at ethz-mrl.github.io/SuperEvent.", + "bbox": [ + 88, + 266, + 485, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 675, + 220, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Event cameras offer exciting advantages over their frame-based counterparts, such as an increased temporal resolution, little motion blur, and a high dynamic range. These properties promise to enhance robustness for robot estimation and perception tasks under fast motion and uncontrollable lighting conditions. However, processing the sparse and asynchronous output of event cameras requires fundamentally different algorithms than traditional frames. The decades-long research edge in frame-based vision resulted", + "bbox": [ + 89, + 700, + 483, + 838 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/45605cbf0dc83051639d4ccbd9c140befef04e3d84f828eeb40cc650d1134b5e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 232, + 905, + 345 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1c2f2ba40ef732dfce855fd108e8099eb6d1e6e20f1c654471e2aeebdc7ae1e5.jpg", + "image_caption": [ + "Figure 1. Detections on the sequence rec1499023756 of the DDD20 [31] dataset (not used for training). Top: Pseudolabels from SuperPoint [15] and SuperGlue [59] on the gray-scale frames. Bottom: Matched keypoints from SuperEvent in the event stream at the corresponding time stamps." + ], + "image_footnote": [], + "bbox": [ + 514, + 348, + 905, + 463 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in most hardware and software being highly optimized for frame processing. A common approach for tasks like Structure-from-Motion, Visual Odometry (VO), Simultaneous Localization and Mapping (SLAM), or Place Recognition is to detect and match keypoints across frames to gain a geometric scene understanding [7, 40]. Consequently, a robust keypoint detector and descriptor working on the event stream would unlock the potential of integrating event cameras into extensive and highly-developed algorithms developed initially for frame-based vision.", + "bbox": [ + 509, + 566, + 906, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While several authors already proposed event-based approaches to detect keypoints [9, 10, 19, 32, 33, 47], their matching abilities still show limitations. Most employ nearest-neighbor matching in pixel space, which restricts their usability to visually simple scenes. In an attempt to overcome this issue, more recent approaches learn descriptors in a data-driven fashion. To generate ground truth labels - inspired by frame-based methods [15] - they warp their event representations with a random homography and penalize non-consistent model predictions. This approach forces the models to learn descriptors to match visually similar keypoints. However, it fundamentally ignores the true", + "bbox": [ + 509, + 719, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Technical University of Munich, {yannick.burkhardt, simon.k.schaefer}@tum.de \n $^{2}$ ETH Zürich, lestefan@ethz.ch \n $^{3}$ Munich Center for Machine Learning (MCML)", + "bbox": [ + 114, + 849, + 460, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8918", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a8e02256b314e4bc787d1586037a71542fa42ba50c41215b629540d8dd726854.jpg", + "image_caption": [ + "Figure 2. Data processing pipeline of SuperEvent: For training, pseudo-labels are generated using a frame-based detector and matcher on the gray-scale frame pairs. We generate spatially synchronized MCTS at the same timestamps and feed them to SuperEvent. The network predictions are compared to the pseudo-labels and the network weights are optimized using backpropagation. During inference, the network predictions can be used to detect keypoints and match their descriptors on the event stream only." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 906, + 345 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "nature of the event generation process that conflates camera motion with scene appearance and geometry. In essence, matches are generated only from one original set of events per keypoint, thereby producing only limited and potentially also unrealistic training data. We hypothesize that this simplification ultimately results in unstable matching.", + "bbox": [ + 88, + 431, + 480, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, we propose training our model to predict descriptors for the keypoint correspondences at different points in time – therefore training with different and actually recorded rather than warped events. Since generating keypoint and match labels manually is intractable, we design a scalable method to exploit the great capabilities of frame-based vision models: generating pseudo-labels in aligned frames and then using them to train our model's predictions. An example is shown in Figure 1.", + "bbox": [ + 88, + 522, + 482, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To provide the model with rich input data, we propose the novel event representation Multi-Channel Time Surfaces (MCTS) which generalizes the commonly used time surface representation [38] to an $n$ -channel tensor of time surfaces for both polarities and different time window sizes. Since an optimal window size depends inversely on the motion magnitude, this approach further increases the model's robustness to both slow and fast motion. Additionally, we revisit network architecture choices of frame- and event-based vision to optimize our model's performance. Figure 2 shows the complete data processing pipeline.", + "bbox": [ + 88, + 657, + 482, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In absence of evaluation procedures to test event-based keypoint and descriptor stability, we develop a benchmark to assess the matched keypoint's quality for pose estimation, similar to benchmarks commonly used in frame-based vision [22, 59, 62, 64]. On two datasets recorded with dif", + "bbox": [ + 88, + 825, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ferent sensors, settings, resolutions, and scenes, we outperform other data-driven and handcrafted approaches in terms of pose estimation by a large margin, demonstrating the robustness of our approach. We carry out ablation studies to confirm the effectiveness of our design choices.", + "bbox": [ + 511, + 431, + 903, + 506 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, we integrate our model into a state-of-the-art (SOTA) stereo visual inertial SLAM (VI-SLAM) framework [40], replacing the frame-based keypoint detector and descriptor extractor with SuperEvent. Our results surpass the current SOTA on event-based Visual Odometry and SLAM. This demonstrates the great potential of an event-sensor integration into existing, well-developed algorithms for frame-based vision.", + "bbox": [ + 511, + 507, + 905, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We name our method SuperEvent in honor of the great influence of the works SuperPoint [15] and SuperGlue [59] on the research field of keypoint detection as well as on our model directly: ultimately, their frame-based detections and matches are used as pseudo-labels to teach keypoint detection and description to SuperEvent.", + "bbox": [ + 511, + 628, + 905, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions: (1) We propose a scalable method to generate training data for interest point detection and matching, leveraging robust existing frame-based keypoint detectors. The resulting data incorporates the real temporal dependence of event generation using real event camera recordings. This enables data-driven models to effectively learn robust keypoint detection and description. (2) We develop the model for event-based vision through key design choices: Our novel MCTS event representation increases the performance of data-driven models by reducing their dependence on motion speed of the scene. A transformer backbone shows to further improve the model's feature ex", + "bbox": [ + 511, + 719, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8919", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "traction. In our conducted experiments, our approach outperforms SOTA detection and matching by a large margin. We publicly release our code to support future event-based vision research. (3) Showcasing the practicality of our method, we integrate the trained model into an existing stereo VI-SLAM approach developed for frame-based vision. The resulting event inertial SLAM system outperforms SOTA approaches on the commonly used TUMVIE [35] and VECtor [18] datasets by a large margin.", + "bbox": [ + 89, + 90, + 480, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 239, + 232, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We begin with a short summary of keypoint detectors for frame-based vision since our work employs some of these findings for model design and training data generation. We then group the event-based works into handcrafted and data-driven keypoint detection, as well as tracking methods. Since our focus in this work lies on keypoint detection and matching based on event data exclusively, we do not consider approaches that rely on combination with other sensors, e.g., frame cameras. For an extensive overview of event-based vision, we refer the interested reader to [17].", + "bbox": [ + 89, + 266, + 482, + 416 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Frame-based keypoint detection: there exists a vast variety of approaches to detecting and describing keypoints in frames. Traditional, handcrafted methods [5, 6, 28, 41, 45, 58, 60] are still widely used, since they are well studied and usually work robustly with little overhead. More recent, data-driven methods outperform the traditional keypoint detectors with impressive results. SuperPoint [15] is the first method trained with homographic adaptation, a self-supervised training technique of warping input images and their detected labels by randomized homographies. SuperPoint's architecture consists of a shared VGG [61]-backbone with a detector and descriptor head, predicting descriptors in a discrete grid which is interpolated to full resolution. The authors later publish the extension SuperGlue [59], a graph neural network that improves the descriptor matching. Following works propose different architectures trained with homographic adaptation [22, 62, 64].", + "bbox": [ + 89, + 417, + 482, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Event-based keypoint detection: handcrafted approaches can be broadly divided into filters and frame reconstruction. Filtering methods work directly on the event stream, attempting to find events originating from corners [1, 11]. While their asynchronous nature offers little processing overhead, downstream tasks are required to process asynchronous data. Existing well-developed downstream solutions for frame-based vision therefore require major modifications.", + "bbox": [ + 89, + 674, + 482, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Frame reconstruction methods collect events to construct a frame representation, such as binary event frames or time surfaces [38]. Adopting methods from frame-based vision, such as the Harris corner detector [28] in [23, 65] or the FAST detector [57] in [49], these approaches detect corners in these event frames. However, they suffer from the in", + "bbox": [ + 89, + 810, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "herent differences between reconstructed event frames and traditional frames, resulting in limited robustness of the algorithms developed for frame-based vision.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Additionally, these handcrafted methods usually employ simple nearest-neighbor matching in pixel space to track keypoints over time. This approach is only reliable for simple scenes since cluttered geometry, high variance in depth, and fast motions cause keypoints to become occluded, tightly clustered, or to jump. Also, the developed heuristics struggle with the ubiquitous and complex noise inherent to event cameras. Usually, they require manual parameter tuning for different camera models, datasets, or sometimes even scenes.", + "bbox": [ + 511, + 137, + 903, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Event-based keypoint detection: data-driven approaches promise to cope better with the event data and its noise. A series of works [9, 10, 47] train data-driven corner detectors using labels by detecting keypoints on corresponding gray-scale frames that are synthetic or from an HVGA ATIS sensor. While these approaches learn data-driven detection, they only track keypoints by nearest-neighbor matching in pixel space, yielding the aforementioned drawback of limited robustness in complex scenes. Additionally, the employed corner detector for ground truth generation as well as the machine-learning models have limited performance due to their simplicity.", + "bbox": [ + 511, + 291, + 903, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For these reasons, EventPoint [32] and SD2Event [19] employ more complex neural networks that additionally predict descriptors. While EventPoint uses the same architecture as SuperPoint [15] and finetunes its pre-trained weights, SD2Event employs agent-based attention to learn detection and description. Both approaches are trained with homographic adaptation and varying time window sizes. While this results in rotation-, distortion-, and scale-invariant models, the main challenge of keypoint description remains largely unsolved: Because most events result from scene or camera motion, corresponding keypoints vary in visual appearance. Since robust keypoint matching requires similar descriptor pairs, this motion dependence must be considered. However, homographic adaptation creates static frame pairs with little to no variance in scene motion within the samples. This contradicts the dynamic and motion-dependent nature of events, resulting in unstable keypoint matching between different timestamps.", + "bbox": [ + 511, + 474, + 903, + 746 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Additionally, the authors of EventPoint propose the event representation Tencode, extending time surfaces with two additional channels to encode the events' polarities. However, to match SuperPoint's architecture with single-channel input, the Tencode tensor is then converted to one channel, disregarding the previously introduced separation of polarities.", + "bbox": [ + 511, + 747, + 903, + 852 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method enforces robust descriptor matching by employing training data containing pseudo-labels for keypoints and descriptors under varying motion. We avoid", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8920", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "synthetic event data since current simulators struggle to model realistic noise. Furthermore, our MCTS representation strictly separates the polarities and provides a range of time window sizes, providing additional information and omitting the time window parameter choice.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Event-based keypoint tracking: while early works fall back on batching events and processing events in a discretized way [37, 52, 68], a promising idea is to exploit their spatial continuous visual flow [2, 13, 30]. Among these approaches is HASTE [3] which calculates the most likely motion hypothesis from a fixed set of translations and rotations for every incoming event. Since HASTE relies on external keypoint initialization, RATE [33] employs a Shi-Tomasi [60] corner detector on the binary event frame, initializing several HASTE instances to track multiple keypoints in parallel.", + "bbox": [ + 89, + 169, + 482, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A major drawback of the described tracking approaches is their computational overhead when processing each event individually. Also, most methods only track one keypoint, resulting in an almost linear growth of computational cost with increasing number of tracker instances. E.g., RATE tracks only 20 features in real-time for an event stream of $180 \\times 40$ pixels resolution – which is hardly sufficient for downstream applications such as SLAM. Additionally, these handcrafted approaches struggle with environmental and hardware changes, as well as the event cameras' noise.", + "bbox": [ + 89, + 335, + 482, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our method SuperEvent works reliably in real-time with adjustable frequency and without re-training or excessive parameter tuning for different camera models. It's dense keypoint heatmap predictions barely result in overhead for more detections. In contrast to the tracking approaches, SuperEvent can be directly integrated into downstream algorithms relying on synchronous output and descriptors, e.g., for SLAM with place recognition, as we demonstrate.", + "bbox": [ + 89, + 489, + 482, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Event-based Odometry: existing Event Odometry (EO) approaches are developed specifically for event processing. While some approaches combine events with frames [8, 25, 26, 46, 66], event-only approaches can be classified as monocular EO [36, 55], monocular EO with IMU [24, 27], stereo EO [21, 67], and stereo EO with IMU [44, 53, 54]. Due to the short history of event cameras, these systems require extensive research and development efforts to work reliably in practice.", + "bbox": [ + 89, + 611, + 482, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast, SuperEvent can be integrated into existing frame-based VO and SLAM systems. We thereby achieve event-based VI-SLAM leveraging highly accurate frame-based systems, emerging from years of research [7, 40].", + "bbox": [ + 89, + 750, + 482, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 828, + 181, + 843 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we describe how we train SuperEvent, introduce the advantages and construction of the MCTS event representation, and detail the network design choices.", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Training Data Generation", + "text_level": 1, + "bbox": [ + 513, + 90, + 750, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To train SuperEvent, we exclusively employ real event camera recordings. Thus, our model learns to cope with the complex event stream properties, e.g., artifacts and noise, which are difficult to simulate realistically.", + "bbox": [ + 511, + 112, + 903, + 172 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, manually annotating keypoints in the asynchronous event data stream is very time-consuming and does not scale. Therefore, we propose an approach to generate ground truth data automatically. We take advantage of the pixel-synchronized events and frames of iniVation DAVIS1 event cameras. Employing a highly robust frame-based keypoint detector and matcher such as SuperPoint [15] + SuperGlue [59] on the gray-scale frames, we can obtain temporally sparse pseudo-labels. To generate a diverse set of training data with reliable pseudo-labels, we process every frame $i$ of a sequence in the following way:", + "bbox": [ + 511, + 172, + 903, + 339 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Firstly, we compute the distances in pixel-space of all key points to their respective matches in frame $i + 1$ . We only consider frames $i$ for training where the distances' median surpasses a threshold $c_{\\mathrm{m}}$ . A small median distance indicates that the scene is nearly static, resulting in insufficient event data for a useful prediction.", + "bbox": [ + 513, + 339, + 903, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- To expose the model to diverse camera pose changes and motion directions, we generate pseudo-labels by matching the descriptors with a subsequent frame $i + j$ . We increase $j$ recursively, resulting in multiple pseudo-labels with increasing time interval between the frames:", + "bbox": [ + 513, + 429, + 903, + 503 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nj \\leftarrow j + j _ {\\mathrm {s}}, j _ {\\mathrm {s}} \\sim \\mathcal {U} (1, j _ {\\max }, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 512, + 903, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with the increment $j_{\\mathrm{s}}$ drawn from the uniform distribution $\\mathcal{U}(1, j_{\\max})$ . $j$ is initialized with 0 before its first update. The maximum step distance $j_{\\max}$ is manually adjusted depending on the properties and texture of the sequence, e.g., for sequences with fast motion, $j_{\\max}$ is decreased to yield sufficient matches.", + "bbox": [ + 524, + 534, + 903, + 623 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- If the number of detected matches falls below the threshold $c_{\\mathrm{m}}$ , these pseudo-labels are discarded, and the next frame $i + 1$ becomes the new reference frame. This procedure ensures sufficient visual overlap and enough texture in the two frames.", + "bbox": [ + 513, + 625, + 903, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Frames with under- or overexposure, or strong motion blur harm the prediction quality of the frame-based model. While these frames are usually filtered out since there are few matches (if $c_{\\mathrm{d}}$ is chosen adequately), we also manually remove scenes that mostly contain too dark, too bright, or blurred frames.", + "bbox": [ + 511, + 700, + 903, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Multi-channel Time Surface", + "text_level": 1, + "bbox": [ + 513, + 797, + 766, + 811 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To allow for efficient batch processing of event data, the asynchronous stream is converted to a tensor with fixed dimensions. In the literature, various variants have been proposed to encode the event stream. An ablation study in [48]", + "bbox": [ + 511, + 819, + 903, + 880 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1https://inivation.com/", + "bbox": [ + 531, + 886, + 653, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8921", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8cda978ae46b17b8ade1a00f9e2ca10b8fb7b681e0a1cf3c83375af6e8cc9a0d.jpg", + "image_caption": [ + "Figure 3. Schematic depiction of the MCTS generation process." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 486, + 391 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "compares some of the most commonly used event representations: binary event frames, time surfaces [38], and voxel grids [70]. For their data-driven feature tracking approach, time surfaces yield the best model performance. The time surface variant Tencode with two more channels encoding the event polarity is proposed in [32]. It is used as model input after conversion to a single-channel frame.", + "bbox": [ + 89, + 430, + 483, + 535 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Time surfaces and its variant Tencode have proven to be an efficient input representation for keypoint detection and tracking algorithms [19, 32, 48], since they assign the highest values to the pixel locations of the most recent events that are usually caused by visual edges. This facilitates finding keypoints along these edges. However, these encodings have shortcomings:", + "bbox": [ + 89, + 537, + 483, + 642 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Time surfaces are one-channel frames that ignore the event polarity and thereby discard potentially useful information. The Tencode representation is an attempt to overcome this issue. However, since it is converted back to a single channel tensor, there is still no clear separation between positive and negative events.", + "- Time surfaces and the Tencode representation depend on the time window length $\\Delta t$ . Increasing this parameter not only results in more events being considered but also decreases the contrast between the most recent events and past events. For high optical flow, this negatively impacts the detection of edges. Conversely, a smaller time window increases this contrast but includes fewer events, which can fail to sufficiently capture edges in areas with little optical flow. Choosing a suitable $\\Delta t$ is challenging, as it depends on the camera and scene motion which might be unknown a priori. The bottom part of Figure 3" + ], + "bbox": [ + 89, + 643, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "shows the effect of varying $\\Delta t$ : Only the time surfaces in the middle clearly display the net on the top right.", + "bbox": [ + 526, + 90, + 903, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our proposed MCTS representation mitigates these problems: It combines time surfaces for each polarity and for various time windows of temporal size $\\Delta t_{n}$ , $n\\in \\{1,\\ldots ,N\\}$ . We distribute the time window sizes logarithmically to trade off data preprocessing and coverage of a wide variety of visual motion speeds. Events are encoded by their timestamp $t$ , their pixel coordinates $x$ and $y$ , and their polarity $p$ . Mathematically, the MCTS tensor at time $\\tau$ considering all $i$ events $\\mathbf{e}_i = (t_i,x_i,y_i,p_i)$ between time $\\tau -\\Delta t_{n}$ and $\\tau$ is defined as", + "bbox": [ + 511, + 121, + 906, + 272 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {M C T S} = \\left(\\mathbf {T S} _ {- 1, \\Delta t _ {1}}, \\dots , \\mathbf {T S} _ {- 1, \\Delta t _ {N}}\\right), \\\\ \\left. \\mathbf {T S} _ {+ 1, \\Delta t _ {1}}, \\dots , \\mathbf {T S} _ {+ 1, \\Delta t _ {N}}\\right), \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 287, + 903, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with each time surface $\\mathbf{TS}$ , initialized with zeros, defined by polarity $p\\in \\{-1, + 1\\}$ and $\\Delta t_{n}$", + "bbox": [ + 511, + 335, + 903, + 367 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T S} _ {p, \\Delta t _ {n}} \\left(x _ {i}, y _ {i}\\right) = \\max _ {i, p _ {i} = p} \\left(1 - \\frac {\\tau - t _ {i}}{\\Delta t _ {n}}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 377, + 903, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The MCTS generation process is visualized in Figure 3.", + "bbox": [ + 511, + 420, + 880, + 435 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Network Architecture & Loss", + "text_level": 1, + "bbox": [ + 511, + 445, + 777, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As commonly employed in frame-based keypoint detection models [15, 22, 64], we combine a shared backbone with a detector and a descriptor head. Starting with the basic SuperPoint [15] architecture with adjusted loss functions for our temporal matching with pseudo-labels, we train various models with different backbones and hyperparameters and compare their performance. As an alternative approach, we replace the grid-based descriptor head with full resolution prediction and the respective loss functions as in [22, 64]. Please refer to the full results of this ablation study in the Supplementary Material.", + "bbox": [ + 511, + 467, + 906, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our resulting architecture consists of a 3-layer MaxViT [63] backbone with a Feature Pyramid Network (FPN) [43] which is also employed in a highly efficient, event-based detector [20]. Combined with grid-based VGG [61] detector and descriptor heads, SuperEvents fully convolutional components can process various input resolutions without retraining. Figure 4 shows the final network architecture.", + "bbox": [ + 511, + 633, + 905, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The loss function evaluates the network prediction for two corresponding training tensors $\\{0,1\\}$ . It combines the individual detector losses $L_{\\mathrm{s},0}$ and $L_{\\mathrm{s},1}$ with the joint descriptor loss $L_{\\mathrm{d}}$ weighted with the constant $c_{\\lambda}$", + "bbox": [ + 511, + 753, + 905, + 814 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {\\mathrm {s}, 0} + L _ {\\mathrm {s}, 1} + c _ {\\lambda} L _ {\\mathrm {d}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 827, + 903, + 843 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following [15], the detection is implemented as a classification problem of the pixels in image patches, i.e., the pixel indices represent the \"classes\". The detector loss $L_{\\mathrm{CE}}$", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8922", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "is thus defined as cross-entropy between a predicted patch of keypoint scores $\\mathbf{y}_{s_h,w}$ (dimension $8\\times 8$ plus no-keypoint-dustbin, thus 65) per grid cell (indices $h,w$ ) and the pseudolabels $\\hat{\\mathbf{y}}_{s_h,w}$ in one-hot encoding", + "bbox": [ + 89, + 90, + 483, + 154 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {s}} = \\frac {1}{H _ {\\mathrm {c}} W _ {\\mathrm {c}}} \\sum_ {\\substack {h = 1 \\\\ w = 1}} ^ {H _ {\\mathrm {c}}, W _ {\\mathrm {c}}} L _ {\\mathrm {CE}} \\left(\\mathbf {y} _ {\\mathrm {s} _ {h}, w}, \\hat {\\mathbf {y}} _ {\\mathrm {s} _ {h}, w}\\right), \\tag{5}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 159, + 482, + 209 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for $H_{\\mathrm{c}} \\times W_{\\mathrm{c}}$ grid cells. In the case of multiple keypoint labels in a grid cell, one is randomly selected.", + "bbox": [ + 89, + 214, + 482, + 244 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Unlike other models trained with homographic adaptation, our sparse pseudo-labels do not allow for dense descriptor training for every pixel or grid cell. As in [15], we only consider a single descriptor (256 channels) per grid cell, $\\mathbf{y}_{\\mathrm{d}_{h_0},w_0}$ and $\\mathbf{y}_{\\mathrm{d}_{h_1},w_1}$ (normalized to length one across the channels) with $(h_0,w_0)$ and $(h_1,w_1)$ denoting the grid indices of tensor 0 and 1. Successful keypoint matching requires similar descriptors for correspondences and distinct descriptors for unrelated keypoints. We measure descriptor similarity with the dot product $d_{h_1,w_1}^{h_0,w_0} = \\mathbf{y}_{\\mathrm{d}_{h_0,w_0}}\\cdot \\mathbf{y}_{\\mathrm{d}_{h_1,w_1}}$ . Our pseudo label $\\hat{y}_{h_1,w_1}^{h_0,w_0} = 1$ indicates that two descriptors at the respective grid cell locations correspond, otherwise $\\hat{y}_{h_1,w_1}^{h_0,w_0} = 0$ . Since the majority of the descriptors do not correspond, the partial loss of descriptor matches is weighted with the constant $c_{\\mathrm{d}}$ . Mathematically, we formulate the descriptor loss as", + "bbox": [ + 89, + 244, + 483, + 489 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {d}} = \\sum_ {\\substack {h _ {0} = 1 \\\\ w _ {0} = 1}} ^ {H _ {\\mathrm {c}}, W _ {\\mathrm {c}}} \\sum_ {\\substack {h _ {1} = 1 \\\\ w _ {1} = 1}} ^ {H _ {\\mathrm {c}}, W _ {\\mathrm {c}}} \\left\\{ \\begin{array}{l l} c _ {\\mathrm {d}} \\max (0, c _ {\\mathrm {p}} - d _ {h _ {1}, w _ {1}} ^ {h _ {0}, w _ {0}}), & \\hat {y} _ {h _ {1}, w _ {1}} ^ {h _ {0}, w _ {0}} = 1, \\\\ \\max (0, d _ {h _ {1}, w _ {1}} ^ {h _ {0}, w _ {0}} - c _ {\\mathrm {n}}), & \\hat {y} _ {h _ {1}, w _ {1}} ^ {h _ {0}, w _ {0}} = 0. \\end{array} \\right. \\tag{6}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 496, + 478, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The constants $c_{\\mathrm{p}}$ and $c_{\\mathrm{n}}$ lead to saturation for maximal desired descriptor (dis-)similarity. All descriptors of grid cells without a keypoint label index $(\\mathbf{y}_{\\mathrm{d}_{h_0},w_0} = 0$ or $\\mathbf{y}_{\\mathrm{d}_{h_1},w_1} = 0)$ do not contribute to the loss.", + "bbox": [ + 89, + 561, + 482, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During inference, we employ a simplified, yet more effective Non-Minimum-Suppression (NMS) [51] approach in the context of keypoint detection. Instead of the common approach to model keypoints as fixed-size boxes and then apply classical NMS as in [15, 32], we find local maxima in the predicted keypoint score heatmap. In contrast to classical NMS, our method only considers local maxima and not all keypoints candidates sufficiently far from other points with higher scores. Additionally, it employs parallelizable operations instead of consecutive updates of the keypoints sorted by score. Every predicted score $s$ at the pixel coordinates $(u, v)$ is considered for a potential keypoint if", + "bbox": [ + 89, + 621, + 482, + 803 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} s [ u, v ] > s [ u + \\Delta u, v + \\Delta v ] \\\\ \\forall \\Delta u, \\Delta v \\in \\left[ - c _ {\\mathrm {b}}, \\dots , - 1, 1, \\dots , c _ {\\mathrm {b}} \\right]. \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 811, + 480, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with $c_{\\mathrm{b}}$ determining the size of the local neighborhood. The keypoint candidates are further filtered by a minimal required score $c_{\\mathrm{s}}$ , or the $n_{\\mathrm{s}}$ highest scores are selected.", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/868c56bea998a1a90bce823b64f7881688517b3d0248e358f3e323bf8ab3e9b1.jpg", + "image_caption": [ + "Figure 4. SuperEvent network architecture and tensor dimensions: a shared transformer backbone is combined with a detector and a descriptor head. The components of the Convolution-Attention blocks $(Conv. + Attn.)$ and the VGG blocks are displayed on the bottom left. Activations are omitted for simplicity." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 903, + 271 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/55f7e88000138737fcdb8a7d259e27a8a0225fa878df51d517a1c3dd79d1c1f4.jpg", + "table_caption": [ + "Table 1. Inference times of our compiled PyTorch model on GPU for common event-stream resolutions (cropped to compatible shape) using float16 precision (same accuracy as float32 in our experiments). NMS requires an additional $0.4\\mathrm{ms}$ on average." + ], + "table_footnote": [], + "table_body": "
160 × 240240 × 320480 × 640720 × 1280
1.3 ms2.6 ms7.8 ms23.2 ms
", + "bbox": [ + 514, + 430, + 906, + 476 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 511, + 643, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate the effectiveness of the presented approach, we conduct a series of experiments and ablation studies. Firstly, we evaluate the detection and description of keypoints by estimating the camera pose changes based on the matched points. We outperform other keypoint detectors and trackers [10, 32, 33] by a large margin.", + "bbox": [ + 511, + 537, + 905, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our second experiment, we integrate SuperEvent into the SLAM framework OKVIS2 [40] originally developed for frame-based cameras. We achieve convincing results on the widely used event-SLAM benchmarks [18, 35] without extensive tuning, showcasing SuperEvent's versatility.", + "bbox": [ + 511, + 628, + 905, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For ablation studies, more qualitative examples, a detailed description of the pose estimation benchmark, and a comparison to frame-based matching for high-speed/ HDR scenarios, please refer to our Supplementary Material.", + "bbox": [ + 511, + 703, + 905, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 773, + 730, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We generated training data as introduced in Section 3.1. To expose the model to various scenes and camera intrinsics, we combine multiple datasets [14, 31, 39, 56, 69], obtaining more than $100\\mathrm{k}$ training sample pairs with pseudolabels. We implement SuperEvent in PyTorch [4] and train it with the Adam [34] optimizer (learning rate $1e^{-4}$ , betas 0.9, 0.999) for 10 epochs.", + "bbox": [ + 511, + 794, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8923", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f174b04f5661e2cd23d65907996d492fca696d4bf6293b0aea8773f2f7b63a83.jpg", + "table_caption": [ + "Table 2. Pose estimation on the Event Camera dataset." + ], + "table_footnote": [], + "table_body": "
MethodECD: Pose Estimation AUC in %
@5°@10°@20°
LLAK [10]0.71.42.1
RATE [33]3.38.418.0
EventPoint [32]1.63.05.4
SuperEvent (ours)22.735.846.7
", + "bbox": [ + 91, + 114, + 480, + 220 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We employ the network architecture as in Section 3.3 with the loss function hyperparameters $c_{\\lambda} = 10$ , $c_{\\mathrm{d}} = 0.5$ , $c_{\\mathrm{p}} = 1.0$ , and $c_{\\mathrm{n}} = 0.2$ . We encode events as a 10-channel MCTS with 5 time window sizes $\\Delta t$ logarithmically distributed from $0.001 \\, \\mathrm{s}$ to $0.1 \\, \\mathrm{s}$ . The detection threshold is set to $c_{\\mathrm{s}} = 0.01$ and the NMS neighborhood $c_{\\mathrm{b}} = 2$ pixels.", + "bbox": [ + 89, + 234, + 480, + 325 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate on a desktop PC with an Intel Core i5-13600 processor, an NVIDIA GeForce GTX 4070 GPU, and 32 GB RAM. Inference times are shown in Table 1.", + "bbox": [ + 89, + 325, + 480, + 371 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Pose Estimation", + "text_level": 1, + "bbox": [ + 89, + 380, + 250, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A common benchmark for evaluating frame-based keypoint detectors is pose estimation [22, 59, 62, 64]. These works use datasets such as ScanNet [12] and MegaDepth [42] containing various images of the same scene with the associated ground truth camera poses. To create a similar event-based benchmark, we employ sequences with ground truth poses from real-world dynamic datasets. We search for uniformly distributed pose changes between $1^{\\circ}$ and $45^{\\circ}$ . The upper bound is chosen to increase the probability of visual overlap. However, since we do not know the ground truth geometry of the scene, it cannot be guaranteed that every sample contains sufficient information to recover the ground truth pose. Next, we process the event stream with various approaches to detect and match keypoints. After outlier removal with Random Sample Consensus (RANSAC) [16], we estimate the camera pose and calculate the error with respect to the ground truth. Finally, we report the area-under-curve (AUC) metric for different thresholds.", + "bbox": [ + 88, + 401, + 482, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare the pose estimation capability of SuperEvent to the 3 approaches LLAK [10], RATE [33], and EventPoint [32]. While we use the open-source implementations of LLAK and RATE, we re-implemented and trained EventPoint by carefully following the steps outlined in the paper. We were unable to evaluate the approach SD2Event [19] due to unpublished code and missing implementation details in the respective work.", + "bbox": [ + 89, + 674, + 482, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The experiment is conducted on two diverse datasets: Firstly, the Event Camera Dataset (ECD) [50] is widely used to evaluate event-based approaches, as it was one of the first publicly available datasets. It is challenging due to the low resolution $(180\\times 240)$ , the noisy output of the initiation DAVIS240C event camera, and fast motion changes. Secondly, we use the more recent dataset Event-aided Di", + "bbox": [ + 89, + 795, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0400c5f873d798847ce964a9e92eef45b185cab54ada1363c6c2f65a1cde414f.jpg", + "table_caption": [ + "Table 3. Pose estimation on Event-aided Direct Sparse Odometry." + ], + "table_footnote": [], + "table_body": "
MethodEDS: Pose Estimation AUC in %
@5°@10°@20°
LLAK [10]0.50.71.0
RATE [33]2.15.110.3
EventPoint [32]1.62.85.2
SuperEvent (ours)15.226.440.1
", + "bbox": [ + 514, + 114, + 903, + 220 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "rect Sparse Odometry (EDS) [29]. In contrast to our training data, this dataset is recorded with a Prophesee Gen $3.1^{2}$ event camera with a higher resolution of $480 \\times 640$ pixels and an over-proportionally increased event rate. Fast, inconsistent camera motions and a high dynamic range make pose estimation challenging.", + "bbox": [ + 509, + 238, + 903, + 328 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 shows that SuperEvent outperforms all baselines by a large margin. On the ECD dataset, its advantage is the greatest for the high-precision estimations below $5^{\\circ}$ error, confirming the effectiveness of our method. In contrast, we found that LLAK detects many keypoints but cannot track the vast majority of them. EventPoint's matching capability is severely limited, possibly due to their static assumptions during training inherently contradicting the dynamic nature of event cameras. Finally, RATE achieves decent results for estimations with up to $20^{\\circ}$ error, but since it cannot recover from tracking loss, it depends on consistent and precise tracking of most keypoints. As seen in the results, this requirement cannot be stably fulfilled.", + "bbox": [ + 509, + 329, + 905, + 525 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3 indicates a slight decrease in SuperEvent's pose estimation performance on the EDS dataset compared to ECD. Since all approaches perform worse, we attribute this effect to the dataset being more challenging. Interestingly, the performance of SuperEvent for $20^{\\circ}$ precision is still in the same region as for the ECD dataset. Considering the different properties of the Prophesee camera compared to the DAVIS camera employed in the training sequences, these results confirm the generalization capability of SuperEvent.", + "bbox": [ + 511, + 526, + 905, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Stereo Event-Visual Inertial SLAM", + "text_level": 1, + "bbox": [ + 511, + 672, + 820, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We demonstrate the usefulness of SuperEvent with a downstream experiment. To date, frame-based visual odometry and SLAM systems largely outperform their event-based counterparts, attributed to their decades-long lead in development. By integrating SuperEvent into the frame-based VI-SLAM system OKVIS2 [40], we enable this framework to process the keypoints detected in the event stream. We compare the performance of the resulting event-only system to other recent event-based Odometry and SLAM approaches on the TUM-VIE dataset [35] in Table 4. Moreover, note that the next best results of the data-driven Mono E(I)O systems DEVO and DEIO are achieved only", + "bbox": [ + 511, + 695, + 905, + 876 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://www_prophesee.ai/", + "bbox": [ + 531, + 886, + 679, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8924", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5032fd5b04f7db556e2e73003d0e3fecc073ae2bf683b5ec8bc28b624e34eb98.jpg", + "table_caption": [ + "Table 4. Results on the small scale mocap sequences of TUM-VIE dataset, ATE in cm. The first three gray results require scale alignment with the ground truth; all other approaches estimate the absolute scale. Baseline numbers taken from [27, 53]." + ], + "table_footnote": [], + "table_body": "
MethodModality1d-trans3d-trans6dofdeskdesk2Average
EVO [55] w/ scale alignmentMono E7.5012.5085.5054.1075.2046.96
DEVO [36] w/ scale alignmentMono E0.501.101.601.701.001.18
DEIO [27] w/ scale alignmentMono E + IMU0.421.111.371.360.731.00
DEIO [27]Mono E + IMU1.081.121.391.411.191.24
ESVO [67]Stereo E12.5417.1913.4612.924.4212.11
ES-PTAM [21]Stereo E1.058.5310.252.507.205.91
ICRA'24 [54]Stereo E + IMU3.8518.90failed8.999.47-
ESVO2 [53]Stereo E + IMU3.337.263.216.164.024.78
OKVIS2 [40] + SuperEvent (ours)Stereo E + IMU0.440.890.430.580.410.55
(ours without loop closure)(0.43)(0.89)(0.43)(0.70)(0.40)(0.57)
", + "bbox": [ + 91, + 128, + 906, + 303 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c56d9126ddd0f75e9659ec6e6bc20ca5491e55c1dd240b2269fa481e3034b7ec.jpg", + "table_caption": [ + "Table 5. Results on VECTor [18] large scale sequences. Our results are obtained with online calibration. Baseline numbers are taken from [27] and [36]. DEVO [36] cannot recover the absolute scale. Sequences marked $\\cdot -\\cdot$ are not evaluated in [27]. ATE in cm." + ], + "table_footnote": [], + "table_body": "
Methodcorr.-dollycorr.-walkunits-dollyunits-scooterschool-dollyschool-scooter
Ours33.13133.16122.6159.0569.9639.07
DEIO [27]492.65325.00826.38304.14--
ESVO [67]failedfailedfailedfailed1371.0983.0
", + "bbox": [ + 91, + 392, + 483, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "after aligning the scale of the estimated trajectories with ground truth. However, real-world applications, e.g., VR and robotics, typically require accurate metric scale. Results on the VECtor dataset [18] are listed in Table 5.", + "bbox": [ + 89, + 496, + 483, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "OKVIS2's loop closure capability relies on long-term descriptor matching. While its effect on the small-scale sequences is limited, we evaluate OKVIS2 + SuperEvent on the loop-floor sequences of the TUM-VIE dataset, since these are explicitly designed to test a system's loop-closure capabilities. Thanks to SuperEvent, the OKVIS2 loop-closure feature works reliably with events, as shown in Table 6. Figure 5 qualitatively compares the estimated trajectories of the sequence loop-floor0. Without loop closure, there is significant drift - albeit not out of the ordinary compared to frame-based VI-SLAM. However, with loop closure enabled, the starting room is recognized and the drift is corrected. This demonstrates the robustness of SuperEvent, achieving successful descriptor matching even with minutes of time between observations.", + "bbox": [ + 88, + 556, + 482, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 799, + 210, + 814 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented SuperEvent, a novel keypoint detector and descriptor for event streams. During its training, we consider the time-dependent keypoint appearance and employ pseudo-labels detected on frames. Combined with the flexible MCTS representation and a transformer back", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6fef1e60530def4d3889263365b5b6c6affc19d139e05b433ae2778722be7296.jpg", + "table_caption": [ + "Table 6. Effect of loop closure of OKVIS2 [40] + SuperEvent on TUM-VIE loop-floor sequences, ATE in cm." + ], + "table_footnote": [], + "table_body": "
loop-floor0123
Estimated length349 m316 m279 m303 m
Loop closure4.964.648.924.74
W/o loop closure132.11161.92116.00129.17
", + "bbox": [ + 514, + 364, + 905, + 436 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f97789cb0ed71692f3f0e708eee8dccc484b96bf1a853b387e5a0620b3898672.jpg", + "image_caption": [ + "Figure 5. OKVIS2 + SuperEvent's estimated trajectories of TUMVIE sequence loop-floor0 with and without loop closure." + ], + "image_footnote": [], + "bbox": [ + 521, + 452, + 898, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "bone, it achieves SOTA performance on keypoint-based pose estimation. SuperEvent's integration into frame-based downstream applications enables event vision in existing keypoint-based systems.", + "bbox": [ + 511, + 622, + 905, + 683 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As future work, SuperEvent can be combined with a frame-based detector to realize systems that simultaneously process frames and events, exploiting their complementary advantages. Also, with a similar approach to SuperEvent, an event-based line detector could improve the performance of downstream applications in man-made environments.", + "bbox": [ + 511, + 684, + 905, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, our training data is generated to enable consistent feature description independent of their motion-dependent appearance. However, since we employ real sequences, consecutive samples are biased to having similar motion directions. This is reflected in the performance of SuperEvent, whose descriptor matchability suffers under strong motion changes. While this is an inherent property of event cameras, a different approach to generating training data could mitigate it.", + "bbox": [ + 511, + 773, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8925", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 122, + 187, + 137 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ignacio Alzugaray and Margarita Chli. Asynchronous corner detection and tracking for event cameras in real time. IEEE Robotics and Automation Letters, 3(4):3177-3184, 2018. 3", + "[2] Ignacio Alzugaray and Margarita Chli. Ace: An efficient asynchronous corner tracker for event cameras. In 2018 International Conference on 3D Vision (3DV), pages 653-661. IEEE, 2018. 4", + "[3] Ignacio Alzugaray and Margarita Chli. HASTE: Multi-hypothesis asynchronous speeded-up tracking of events. In 31st British Machine Vision Virtual Conference (BMVC 2020), page 744. ETH Zurich, Institute of Robotics and Intelligent Systems, 2020. 4", + "[4] Jason Ansel, Edward Yang, Horace He, Natalia Gimelshein, Animesh Jain, Michael Voznesensky, Bin Bao, Peter Bell, David Berard, Evgeni Burovski, Geeta Chauhan, Anjali Choudria, Will Constable, Alban Desmaison, Zachary DeVito, Elias Ellison, Will Feng, Jiong Gong, Michael Gschwind, Brian Hirsh, Sherlock Huang, Kshitteej Kalambarkar, Laurent Kirsch, Michael Lazos, Mario Lezcano, Yanbo Liang, Jason Liang, Yinghai Lu, CK Luk, Bert Maher, Yunjie Pan, Christian Puhrsch, Matthias Reso, Mark Saroufim, Marcos Yukio Siraichi, Helen Suk, Michael Suo, Phil Tillet, Eikan Wang, Xiaodong Wang, William Wen, Shunting Zhang, Xu Zhao, Keren Zhou, Richard Zou, Ajit Mathews, Gregory Chanan, Peng Wu, and Soumith Chintala. PyTorch 2: Faster Machine Learning Through Dynamic Python Bytecode Transformation and Graph Compilation. In 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2 (ASPLOS '24). ACM, 2024. 6", + "[5] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-up robust features (SURF). Computer vision and image understanding, 110(3):346-359, 2008. 3", + "[6] Michael Calonder, Vincent Lepetit, Christoph Strecha, and Pascal Fua. Brief: Binary robust independent elementary features. In Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part IV 11, pages 778-792. Springer, 2010. 3", + "[7] Carlos Campos, Richard Elvira, Juan J. Gómez, José M. M. Montiel, and Juan D. Tardós. ORB-SLAM3: An accurate open-source library for visual, visual-inertial and multi-map SLAM. IEEE Transactions on Robotics, 37(6):1874-1890, 2021. 1, 4", + "[8] Peiyu Chen, Weipeng Guan, and Peng Lu. ESVIO: Event-based stereo visual inertial odometry. IEEE Robotics and Automation Letters, 8(6):3661-3668, 2023. 4", + "[9] Philippe Chiberre, Etienne Perot, Amos Sironi, and Vincent Lepetit. Detecting stable keypoints from events through image gradient prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1387-1394, 2021. 1, 3", + "[10] Philippe Chiberre, Etienne Perot, Amos Sironi, and Vincent" + ], + "bbox": [ + 93, + 147, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Lepetit. Long-lived accurate keypoints in event streams. arXiv preprint arXiv:2209.10385, 2022. 1, 3, 6, 7", + "[11] Xavier Clady, Sio-Hoi Ieng, and Ryad Benosman. Asynchronous event-based corner detection and matching. Neural Networks, 66:91-106, 2015. 3", + "[12] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828–5839, 2017. 7", + "[13] Laurent Dardelet, Ryad Benosman, and Sio-Hoi Ieng. An event-by-event feature detection and tracking invariant to motion direction and velocity. Authorea Preprints, 2023. 4", + "[14] Jeffrey Delmerico, Titus Cieslewski, Henri Rebecq, Matthias Faessler, and Davide Scaramuzza. Are we ready for autonomous drone racing? the uzh-fpv drone racing dataset. In 2019 International Conference on Robotics and Automation (ICRA), pages 6713-6719. IEEE, 2019. 6", + "[15] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperPoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224-236, 2018. 1, 2, 3, 4, 5, 6", + "[16] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 7", + "[17] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 3", + "[18] Ling Gao, Yuxuan Liang, Jiaqi Yang, Shaoxun Wu, Chenyu Wang, Jiaben Chen, and Laurent Kneip. Vector: A versatile event-centric benchmark for multi-sensor slam. IEEE Robotics and Automation Letters, 7(3):8217-8224, 2022. 3, 6, 8", + "[19] Yuan Gao, Yuqing Zhu, Xinjun Li, Yimin Du, and Tianzhu Zhang. SD2Event: Self-supervised learning of dynamic detectors and contextual descriptors for event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3055-3064, 2024. 1, 3, 5, 7", + "[20] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13884-13893, 2023. 5", + "[21] Suman Ghosh, Valentina Cavinato, and Guillermo Gallego. ES-PTAM: Event-based stereo parallel tracking and mapping. In European Conference on Computer Vision (ECCV) Workshops, 2024. 4, 8", + "[22] Pierre Gleize, Weiyao Wang, and Matt Feiszli. SiLK: Simple learned keypoints. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 22442-22451, 2023. 2, 3, 5, 7", + "[23] Arren Glover, Aiko Dinale, Leandro De Souza Rosa, Simeon Bamford, and Chiara Bartolozzi. luvHarris: A practical corner detector for event-cameras. IEEE Transactions on" + ], + "bbox": [ + 517, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8926", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Pattern Analysis and Machine Intelligence, 44(12):10087-10098, 2021. 3", + "[24] Weipeng Guan and Peng Lu. Monocular event visual inertial odometry based on event-corner using sliding windows graph-based optimization. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2438-2445, 2022. 4", + "[25] Weipeng Guan, Peiyu Chen, Yuhan Xie, and Peng Lu. PL- EVIO: Robust monocular event-based visual inertial odometry with point and line features. IEEE Transactions on Automation Science and Engineering, 2023. 4", + "[26] Weipeng Guan, Peiyu Chen, Huibin Zhao, Yu Wang, and Peng Lu. EVI-SAM: Robust, real-time, tightly-coupled event-visual-inertial state estimation and 3d dense mapping. Advanced Intelligent Systems, 6(12):2400243, 2024. 4", + "[27] Weipeng Guan, Fuling Lin, Peiyu Chen, and Peng Lu. DEIO: Deep event inertial odometry. arXiv preprint arXiv:2411.03928, 2024. 4, 8", + "[28] Christopher G. Harris and M. J. Stephens. A combined corner and edge detector. In Alvey Vision Conference, 1988. 3", + "[29] Javier Hidalgo-Carrió, Guillermo Gallego, and Davide Scaramuzza. Event-aided direct sparse odometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5781–5790, 2022. 7", + "[30] Sumin Hu, Yeeun Kim, Hyungtae Lim, Alex Junho Lee, and Hyun Myung. eCDT: Event clustering for simultaneous feature detection and tracking. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3808-3815. IEEE, 2022. 4", + "[31] Yuhuang Hu, Jonathan Binas, Daniel Neil, Shih-Chii Liu, and Tobi Delbruck. DDD20 end-to-end event camera driving dataset: Fusing frames and events with deep learning for improved steering prediction. In 2020 IEEE 23rd International Conference on Intelligent Transportation Systems (ITSC), pages 1-6. IEEE, 2020. 1, 6", + "[32] Ze Huang, Li Sun, Cheng Zhao, Song Li, and Songzhi Su. EventPoint: Self-supervised interest point detection and description for event-based camera. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5396-5405, 2023. 1, 3, 5, 6, 7", + "[33] Mikihiro Ikura, Cedric Le Gentil, Marcus G. Müller, Florian Schuler, Atsushi Yamashita, and Wolfgang Stürzl. RATE: Real-time asynchronous feature tracking with event cameras. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 11662-11669, 2024. 1, 4, 6, 7", + "[34] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. 6", + "[35] Simon Klenk, Jason Chui, Nikolaus Demmel, and Daniel Cremers. TUM-VIE: The TUM stereo visual-inertial event dataset. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8601-8608. IEEE, 2021. 3, 6, 7", + "[36] Simon Klenk, Marvin Motzet, Lukas Koestler, and Daniel Cremers. Deep event visual odometry. In 2024 International Conference on 3D Vision (3DV), pages 739-749. IEEE, 2024. 4, 8" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Beat Kueng, Elias Mueggler, Guillermo Gallego, and Davide Scaramuzza. Low-latency visual odometry using event-based feature tracks. In 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 16-23. IEEE, 2016. 4", + "[38] Xavier Lagorce, Garrick Orchard, Francesco Galluppi, Bertram E. Shi, and Ryad B. Benosman. HOTS: A hierarchy of event-based time-surfaces for pattern recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(7):1346-1359, 2017. 2, 3, 5", + "[39] Alex Junho Lee, Younggun Cho, Young-sik Shin, Ayoung Kim, and Hyun Myung. ViViD++: Vision for visibility dataset. IEEE Robotics and Automation Letters, 7(3):6282-6289, 2022. 6", + "[40] Stefan Leutenegger. OKVIS2: Realtime scalable visual-inertial SLAM with loop closure, 2022. 1, 2, 4, 6, 7, 8", + "[41] Stefan Leutenegger, Margarita Chli, and Roland Y. Siegwart. BRISK: Binary robust invariant scalable keypoints. In 2011 International Conference on Computer Vision, pages 2548-2555, 2011. 3", + "[42] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2041-2050, 2018. 7", + "[43] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 5", + "[44] Zhe Liu, Dianxi Shi, Ruihao Li, and Shaowu Yang. ESVIO: event-based stereo visual-inertial odometry. Sensors, 23(4): 1998, 2023. 4", + "[45] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60:91-110, 2004. 3", + "[46] Florian Mahlknecht, Daniel Gehrig, Jeremy Nash, Friedrich M Rockenbauer, Benjamin Morrell, Jeff Delaune, and Davide Scaramuzza. Exploring event camera-based odometry for planetary robots. IEEE Robotics and Automation Letters, 7(4):8651-8658, 2022. 4", + "[47] Jacques Manderscheid, Amos Sironi, Nicolas Bourdis, Davide Migliore, and Vincent Lepetit. Speed invariant time surface for learning to detect corner points with event-based cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10245-10254, 2019. 1, 3", + "[48] Nico Messikommer, Carter Fang, Mathias Gehrig, and Davide Scaramuzza. Data-driven feature tracking for event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5642-5651, 2023. 4, 5", + "[49] Elias Mueggler, Chiara Bartolozzi, and Davide Scaramuzza. Fast event-based corner detection. Proceedings of the British Machine Vision Conference, 2017. 3", + "[50] Elias Mueggler, Henri Rebecq, Guillermo Gallego, Tobi Delbruck, and Davide Scaramuzza. The event-camera dataset and simulator: Event-based data for pose estimation, visual" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8927", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "odometry, and slam. The International Journal of Robotics Research, 36(2):142-149, 2017. 7", + "[51] Alexander Neubeck and Luc Van Gool. Efficient nonmaximum suppression. In 18th international conference on pattern recognition (ICPR'06), pages 850-855. IEEE, 2006. 6", + "[52] Zhenjiang Ni, Aude Bolopion, Joel Agnus, Ryad Benosman, and Stephane Regnier. Asynchronous event-based visual shape tracking for stable haptic feedback in microrobotics. IEEE Transactions on Robotics, 28(5):1081-1089, 2012. 4", + "[53] Junkai Niu, Sheng Zhong, Xiuyuan Lu, Shaojie Shen, Guillermo Gallego, and Yi Zhou. ESVO2: Direct visual-inertial odometry with stereo event cameras. arXiv e-prints, pages arXiv-2410, 2024. 4, 8", + "[54] Junkai Niu, Sheng Zhong, and Yi Zhou. Imu-aided event-based stereo visual odometry. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 11977–11983, 2024. 4, 8", + "[55] Henri Rebecq, Timo Horstschäfer, Guillermo Gallego, and Davide Scaramuzza. EVO: A geometric approach to event-based 6-dof parallel tracking and mapping in real time. IEEE Robotics and Automation Letters, 2(2):593-600, 2016. 4, 8", + "[56] Juan Pablo Rodríguez-Gómez, Raul Tapia, Julio L Paneque, Pedro Grau, Augusto Gómez Eguíluz, Jose Ramiro Martínez-de Dios, and Anibal Ollero. The GRIFFIN perception dataset: Bridging the gap between flapping-wing flight and robotic perception. IEEE Robotics and Automation Letters, 6(2):1066-1073, 2021. 6", + "[57] Edward Rosten and Tom Drummond. Machine learning for high-speed corner detection. In Computer Vision-ECCV 2006: 9th European Conference on Computer Vision, Graz, Austria, May 7-13, 2006. Proceedings, Part I 9, pages 430-443. Springer, 2006. 3", + "[58] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. ORB: An efficient alternative to SIFT or SURF. In 2011 International Conference on Computer Vision, pages 2564-2571, 2011. 3", + "[59] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 1, 2, 3, 4, 7", + "[60] Jianbo Shi and Tomasi. Good features to track. In 1994 Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 593-600, 1994. 3, 4", + "[61] Karen Simonyan. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 3, 5", + "[62] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 2, 3, 7", + "[63] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. MaxViT: Multi-axis vision transformer. In European conference on computer vision, pages 459-479. Springer, 2022. 5" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Michal Tyszkiiewicz, Pascal Fua, and Eduard Trulls. DISK: Learning local features with policy gradient. Advances in Neural Information Processing Systems, 33:14254-14265, 2020. 2, 3, 5, 7", + "[65] Valentina Vasco, Arren Glover, and Chiara Bartolozzi. Fast event-based harris corner detection exploiting the advantages of event-driven cameras. In 2016 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 4144-4149. IEEE, 2016. 3", + "[66] Antoni Rosinol Vidal, Henri Rebecq, Timo Horstschaefer, and Davide Scaramuzza. Ultimate SLAM? Combining events, images, and IMU for robust visual slam in hdr and high-speed scenarios. IEEE Robotics and Automation Letters, 3(2):994-1001, 2018. 4", + "[67] Yi Zhou, Guillermo Gallego, and Shaojie Shen. Event-based stereo visual odometry. IEEE Transactions on Robotics, 37 (5):1433-1450, 2021. 4, 8", + "[68] Alex Zihao Zhu, Nikolay Atanasov, and Kostas Daniilidis. Event-based feature tracking with probabilistic data association. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 4465-4470. IEEE, 2017. 4", + "[69] Alex Zihao Zhu, Dinesh Thakur, Tolga Özaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The multi-vehicle stereo event camera dataset: An event camera dataset for 3d perception. IEEE Robotics and Automation Letters, 3 (3):2032–2039, 2018. 6", + "[70] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based optical flow using motion compensation. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, 2018. 5" + ], + "bbox": [ + 516, + 90, + 903, + 527 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8928", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SuperEvent_ Cross-Modal Learning of Event-based Keypoint Detection for SLAM/2e684548-d342-4d0d-9d82-edd690c021a1_model.json b/2025/SuperEvent_ Cross-Modal Learning of Event-based Keypoint Detection for SLAM/2e684548-d342-4d0d-9d82-edd690c021a1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..088300a9506d88851f262aa6c06cac22d804e518 --- /dev/null +++ b/2025/SuperEvent_ Cross-Modal Learning of Event-based Keypoint Detection for SLAM/2e684548-d342-4d0d-9d82-edd690c021a1_model.json @@ -0,0 +1,2422 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.131, + 0.908, + 0.153 + ], + "angle": 0, + "content": "SuperEvent: Cross-Modal Learning of Event-based Keypoint Detection for SLAM" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.18, + 0.381, + 0.198 + ], + "angle": 0, + "content": "Yannick Burkhardt\\(^{1,2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.402, + 0.18, + 0.558, + 0.198 + ], + "angle": 0, + "content": "Simon Schaefer\\(^{1,2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.599, + 0.18, + 0.783, + 0.199 + ], + "angle": 0, + "content": "Stefan Leutenegger\\(^{1,2,3}\\)" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.234, + 0.327, + 0.249 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.267, + 0.486, + 0.644 + ], + "angle": 0, + "content": "Event-based keypoint detection and matching holds significant potential, enabling the integration of event sensors into highly optimized Visual SLAM systems developed for frame cameras over decades of research. Unfortunately, existing approaches struggle with the motion-dependent appearance of keypoints and the complex noise prevalent in event streams, resulting in severely limited feature matching capabilities and poor performance on downstream tasks. To mitigate this problem, we propose SuperEvent, a data-driven approach to predict stable keypoints with expressive descriptors. Due to the absence of event datasets with ground truth keypoints labels, we leverage existing frame-based keypoint detectors on readily available event-aligned and synchronized gray-scale frames for self-supervision: we generate temporally sparse keypoint pseudo-labels considering that events are a product of both scene appearance and camera motion. Combined with our novel, information-rich event representation, we enable SuperEvent to effectively learn robust keypoint detection and description in event streams. Finally, we demonstrate the usefulness of SuperEvent by its integration into a modern sparse keypoint and descriptor-based SLAM framework originally developed for traditional cameras, surpassing the state-of-the-art in event-based SLAM by a wide margin. Source code is available at ethz-mrl.github.io/SuperEvent." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.676, + 0.222, + 0.691 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.702, + 0.484, + 0.839 + ], + "angle": 0, + "content": "Event cameras offer exciting advantages over their frame-based counterparts, such as an increased temporal resolution, little motion blur, and a high dynamic range. These properties promise to enhance robustness for robot estimation and perception tasks under fast motion and uncontrollable lighting conditions. However, processing the sparse and asynchronous output of event cameras requires fundamentally different algorithms than traditional frames. The decades-long research edge in frame-based vision resulted" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.233, + 0.906, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.349, + 0.906, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.474, + 0.907, + 0.545 + ], + "angle": 0, + "content": "Figure 1. Detections on the sequence rec1499023756 of the DDD20 [31] dataset (not used for training). Top: Pseudolabels from SuperPoint [15] and SuperGlue [59] on the gray-scale frames. Bottom: Matched keypoints from SuperEvent in the event stream at the corresponding time stamps." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.568, + 0.907, + 0.719 + ], + "angle": 0, + "content": "in most hardware and software being highly optimized for frame processing. A common approach for tasks like Structure-from-Motion, Visual Odometry (VO), Simultaneous Localization and Mapping (SLAM), or Place Recognition is to detect and match keypoints across frames to gain a geometric scene understanding [7, 40]. Consequently, a robust keypoint detector and descriptor working on the event stream would unlock the potential of integrating event cameras into extensive and highly-developed algorithms developed initially for frame-based vision." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "While several authors already proposed event-based approaches to detect keypoints [9, 10, 19, 32, 33, 47], their matching abilities still show limitations. Most employ nearest-neighbor matching in pixel space, which restricts their usability to visually simple scenes. In an attempt to overcome this issue, more recent approaches learn descriptors in a data-driven fashion. To generate ground truth labels - inspired by frame-based methods [15] - they warp their event representations with a random homography and penalize non-consistent model predictions. This approach forces the models to learn descriptors to match visually similar keypoints. However, it fundamentally ignores the true" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.851, + 0.462, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\)Technical University of Munich, {yannick.burkhardt, simon.k.schaefer}@tum.de \n\\(^{2}\\)ETH Zürich, lestefan@ethz.ch \n\\(^{3}\\)Munich Center for Machine Learning (MCML)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8918" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.907, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.357, + 0.908, + 0.414 + ], + "angle": 0, + "content": "Figure 2. Data processing pipeline of SuperEvent: For training, pseudo-labels are generated using a frame-based detector and matcher on the gray-scale frame pairs. We generate spatially synchronized MCTS at the same timestamps and feed them to SuperEvent. The network predictions are compared to the pseudo-labels and the network weights are optimized using backpropagation. During inference, the network predictions can be used to detect keypoints and match their descriptors on the event stream only." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.432, + 0.482, + 0.522 + ], + "angle": 0, + "content": "nature of the event generation process that conflates camera motion with scene appearance and geometry. In essence, matches are generated only from one original set of events per keypoint, thereby producing only limited and potentially also unrealistic training data. We hypothesize that this simplification ultimately results in unstable matching." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.523, + 0.483, + 0.659 + ], + "angle": 0, + "content": "In contrast, we propose training our model to predict descriptors for the keypoint correspondences at different points in time – therefore training with different and actually recorded rather than warped events. Since generating keypoint and match labels manually is intractable, we design a scalable method to exploit the great capabilities of frame-based vision models: generating pseudo-labels in aligned frames and then using them to train our model's predictions. An example is shown in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.659, + 0.483, + 0.825 + ], + "angle": 0, + "content": "To provide the model with rich input data, we propose the novel event representation Multi-Channel Time Surfaces (MCTS) which generalizes the commonly used time surface representation [38] to an \\( n \\)-channel tensor of time surfaces for both polarities and different time window sizes. Since an optimal window size depends inversely on the motion magnitude, this approach further increases the model's robustness to both slow and fast motion. Additionally, we revisit network architecture choices of frame- and event-based vision to optimize our model's performance. Figure 2 shows the complete data processing pipeline." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.826, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In absence of evaluation procedures to test event-based keypoint and descriptor stability, we develop a benchmark to assess the matched keypoint's quality for pose estimation, similar to benchmarks commonly used in frame-based vision [22, 59, 62, 64]. On two datasets recorded with dif" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.905, + 0.507 + ], + "angle": 0, + "content": "ferent sensors, settings, resolutions, and scenes, we outperform other data-driven and handcrafted approaches in terms of pose estimation by a large margin, demonstrating the robustness of our approach. We carry out ablation studies to confirm the effectiveness of our design choices." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.906, + 0.627 + ], + "angle": 0, + "content": "Finally, we integrate our model into a state-of-the-art (SOTA) stereo visual inertial SLAM (VI-SLAM) framework [40], replacing the frame-based keypoint detector and descriptor extractor with SuperEvent. Our results surpass the current SOTA on event-based Visual Odometry and SLAM. This demonstrates the great potential of an event-sensor integration into existing, well-developed algorithms for frame-based vision." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.906, + 0.719 + ], + "angle": 0, + "content": "We name our method SuperEvent in honor of the great influence of the works SuperPoint [15] and SuperGlue [59] on the research field of keypoint detection as well as on our model directly: ultimately, their frame-based detections and matches are used as pseudo-labels to teach keypoint detection and description to SuperEvent." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Contributions: (1) We propose a scalable method to generate training data for interest point detection and matching, leveraging robust existing frame-based keypoint detectors. The resulting data incorporates the real temporal dependence of event generation using real event camera recordings. This enables data-driven models to effectively learn robust keypoint detection and description. (2) We develop the model for event-based vision through key design choices: Our novel MCTS event representation increases the performance of data-driven models by reducing their dependence on motion speed of the scene. A transformer backbone shows to further improve the model's feature ex" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8919" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.228 + ], + "angle": 0, + "content": "traction. In our conducted experiments, our approach outperforms SOTA detection and matching by a large margin. We publicly release our code to support future event-based vision research. (3) Showcasing the practicality of our method, we integrate the trained model into an existing stereo VI-SLAM approach developed for frame-based vision. The resulting event inertial SLAM system outperforms SOTA approaches on the commonly used TUMVIE [35] and VECtor [18] datasets by a large margin." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.241, + 0.233, + 0.258 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.267, + 0.483, + 0.417 + ], + "angle": 0, + "content": "We begin with a short summary of keypoint detectors for frame-based vision since our work employs some of these findings for model design and training data generation. We then group the event-based works into handcrafted and data-driven keypoint detection, as well as tracking methods. Since our focus in this work lies on keypoint detection and matching based on event data exclusively, we do not consider approaches that rely on combination with other sensors, e.g., frame cameras. For an extensive overview of event-based vision, we refer the interested reader to [17]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.418, + 0.483, + 0.675 + ], + "angle": 0, + "content": "Frame-based keypoint detection: there exists a vast variety of approaches to detecting and describing keypoints in frames. Traditional, handcrafted methods [5, 6, 28, 41, 45, 58, 60] are still widely used, since they are well studied and usually work robustly with little overhead. More recent, data-driven methods outperform the traditional keypoint detectors with impressive results. SuperPoint [15] is the first method trained with homographic adaptation, a self-supervised training technique of warping input images and their detected labels by randomized homographies. SuperPoint's architecture consists of a shared VGG [61]-backbone with a detector and descriptor head, predicting descriptors in a discrete grid which is interpolated to full resolution. The authors later publish the extension SuperGlue [59], a graph neural network that improves the descriptor matching. Following works propose different architectures trained with homographic adaptation [22, 62, 64]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.81 + ], + "angle": 0, + "content": "Event-based keypoint detection: handcrafted approaches can be broadly divided into filters and frame reconstruction. Filtering methods work directly on the event stream, attempting to find events originating from corners [1, 11]. While their asynchronous nature offers little processing overhead, downstream tasks are required to process asynchronous data. Existing well-developed downstream solutions for frame-based vision therefore require major modifications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Frame reconstruction methods collect events to construct a frame representation, such as binary event frames or time surfaces [38]. Adopting methods from frame-based vision, such as the Harris corner detector [28] in [23, 65] or the FAST detector [57] in [49], these approaches detect corners in these event frames. However, they suffer from the in" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "herent differences between reconstructed event frames and traditional frames, resulting in limited robustness of the algorithms developed for frame-based vision." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.138, + 0.905, + 0.289 + ], + "angle": 0, + "content": "Additionally, these handcrafted methods usually employ simple nearest-neighbor matching in pixel space to track keypoints over time. This approach is only reliable for simple scenes since cluttered geometry, high variance in depth, and fast motions cause keypoints to become occluded, tightly clustered, or to jump. Also, the developed heuristics struggle with the ubiquitous and complex noise inherent to event cameras. Usually, they require manual parameter tuning for different camera models, datasets, or sometimes even scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.292, + 0.905, + 0.473 + ], + "angle": 0, + "content": "Event-based keypoint detection: data-driven approaches promise to cope better with the event data and its noise. A series of works [9, 10, 47] train data-driven corner detectors using labels by detecting keypoints on corresponding gray-scale frames that are synthetic or from an HVGA ATIS sensor. While these approaches learn data-driven detection, they only track keypoints by nearest-neighbor matching in pixel space, yielding the aforementioned drawback of limited robustness in complex scenes. Additionally, the employed corner detector for ground truth generation as well as the machine-learning models have limited performance due to their simplicity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.475, + 0.905, + 0.747 + ], + "angle": 0, + "content": "For these reasons, EventPoint [32] and SD2Event [19] employ more complex neural networks that additionally predict descriptors. While EventPoint uses the same architecture as SuperPoint [15] and finetunes its pre-trained weights, SD2Event employs agent-based attention to learn detection and description. Both approaches are trained with homographic adaptation and varying time window sizes. While this results in rotation-, distortion-, and scale-invariant models, the main challenge of keypoint description remains largely unsolved: Because most events result from scene or camera motion, corresponding keypoints vary in visual appearance. Since robust keypoint matching requires similar descriptor pairs, this motion dependence must be considered. However, homographic adaptation creates static frame pairs with little to no variance in scene motion within the samples. This contradicts the dynamic and motion-dependent nature of events, resulting in unstable keypoint matching between different timestamps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.748, + 0.905, + 0.853 + ], + "angle": 0, + "content": "Additionally, the authors of EventPoint propose the event representation Tencode, extending time surfaces with two additional channels to encode the events' polarities. However, to match SuperPoint's architecture with single-channel input, the Tencode tensor is then converted to one channel, disregarding the previously introduced separation of polarities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Our method enforces robust descriptor matching by employing training data containing pseudo-labels for keypoints and descriptors under varying motion. We avoid" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8920" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "synthetic event data since current simulators struggle to model realistic noise. Furthermore, our MCTS representation strictly separates the polarities and provides a range of time window sizes, providing additional information and omitting the time window parameter choice." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.17, + 0.483, + 0.336 + ], + "angle": 0, + "content": "Event-based keypoint tracking: while early works fall back on batching events and processing events in a discretized way [37, 52, 68], a promising idea is to exploit their spatial continuous visual flow [2, 13, 30]. Among these approaches is HASTE [3] which calculates the most likely motion hypothesis from a fixed set of translations and rotations for every incoming event. Since HASTE relies on external keypoint initialization, RATE [33] employs a Shi-Tomasi [60] corner detector on the binary event frame, initializing several HASTE instances to track multiple keypoints in parallel." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.337, + 0.483, + 0.488 + ], + "angle": 0, + "content": "A major drawback of the described tracking approaches is their computational overhead when processing each event individually. Also, most methods only track one keypoint, resulting in an almost linear growth of computational cost with increasing number of tracker instances. E.g., RATE tracks only 20 features in real-time for an event stream of \\(180 \\times 40\\) pixels resolution – which is hardly sufficient for downstream applications such as SLAM. Additionally, these handcrafted approaches struggle with environmental and hardware changes, as well as the event cameras' noise." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.49, + 0.483, + 0.61 + ], + "angle": 0, + "content": "Our method SuperEvent works reliably in real-time with adjustable frequency and without re-training or excessive parameter tuning for different camera models. It's dense keypoint heatmap predictions barely result in overhead for more detections. In contrast to the tracking approaches, SuperEvent can be directly integrated into downstream algorithms relying on synchronous output and descriptors, e.g., for SLAM with place recognition, as we demonstrate." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.483, + 0.748 + ], + "angle": 0, + "content": "Event-based Odometry: existing Event Odometry (EO) approaches are developed specifically for event processing. While some approaches combine events with frames [8, 25, 26, 46, 66], event-only approaches can be classified as monocular EO [36, 55], monocular EO with IMU [24, 27], stereo EO [21, 67], and stereo EO with IMU [44, 53, 54]. Due to the short history of event cameras, these systems require extensive research and development efforts to work reliably in practice." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.81 + ], + "angle": 0, + "content": "In contrast, SuperEvent can be integrated into existing frame-based VO and SLAM systems. We thereby achieve event-based VI-SLAM leveraging highly accurate frame-based systems, emerging from years of research [7, 40]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.829, + 0.182, + 0.844 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this section, we describe how we train SuperEvent, introduce the advantages and construction of the MCTS event representation, and detail the network design choices." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.092, + 0.75, + 0.107 + ], + "angle": 0, + "content": "3.1. Training Data Generation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.113, + 0.905, + 0.173 + ], + "angle": 0, + "content": "To train SuperEvent, we exclusively employ real event camera recordings. Thus, our model learns to cope with the complex event stream properties, e.g., artifacts and noise, which are difficult to simulate realistically." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.174, + 0.905, + 0.34 + ], + "angle": 0, + "content": "However, manually annotating keypoints in the asynchronous event data stream is very time-consuming and does not scale. Therefore, we propose an approach to generate ground truth data automatically. We take advantage of the pixel-synchronized events and frames of iniVation DAVIS1 event cameras. Employing a highly robust frame-based keypoint detector and matcher such as SuperPoint [15] + SuperGlue [59] on the gray-scale frames, we can obtain temporally sparse pseudo-labels. To generate a diverse set of training data with reliable pseudo-labels, we process every frame \\(i\\) of a sequence in the following way:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.34, + 0.905, + 0.43 + ], + "angle": 0, + "content": "- Firstly, we compute the distances in pixel-space of all key points to their respective matches in frame \\( i + 1 \\). We only consider frames \\( i \\) for training where the distances' median surpasses a threshold \\( c_{\\mathrm{m}} \\). A small median distance indicates that the scene is nearly static, resulting in insufficient event data for a useful prediction." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.43, + 0.905, + 0.505 + ], + "angle": 0, + "content": "- To expose the model to diverse camera pose changes and motion directions, we generate pseudo-labels by matching the descriptors with a subsequent frame \\( i + j \\). We increase \\( j \\) recursively, resulting in multiple pseudo-labels with increasing time interval between the frames:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.513, + 0.905, + 0.529 + ], + "angle": 0, + "content": "\\[\nj \\leftarrow j + j _ {\\mathrm {s}}, j _ {\\mathrm {s}} \\sim \\mathcal {U} (1, j _ {\\max }, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.535, + 0.905, + 0.624 + ], + "angle": 0, + "content": "with the increment \\( j_{\\mathrm{s}} \\) drawn from the uniform distribution \\( \\mathcal{U}(1, j_{\\max}) \\). \\( j \\) is initialized with 0 before its first update. The maximum step distance \\( j_{\\max} \\) is manually adjusted depending on the properties and texture of the sequence, e.g., for sequences with fast motion, \\( j_{\\max} \\) is decreased to yield sufficient matches." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.625, + 0.905, + 0.699 + ], + "angle": 0, + "content": "- If the number of detected matches falls below the threshold \\( c_{\\mathrm{m}} \\), these pseudo-labels are discarded, and the next frame \\( i + 1 \\) becomes the new reference frame. This procedure ensures sufficient visual overlap and enough texture in the two frames." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.701, + 0.905, + 0.79 + ], + "angle": 0, + "content": "Frames with under- or overexposure, or strong motion blur harm the prediction quality of the frame-based model. While these frames are usually filtered out since there are few matches (if \\( c_{\\mathrm{d}} \\) is chosen adequately), we also manually remove scenes that mostly contain too dark, too bright, or blurred frames." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.798, + 0.767, + 0.813 + ], + "angle": 0, + "content": "3.2. Multi-channel Time Surface" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.82, + 0.905, + 0.881 + ], + "angle": 0, + "content": "To allow for efficient batch processing of event data, the asynchronous stream is converted to a tensor with fixed dimensions. In the literature, various variants have been proposed to encode the event stream. An ablation study in [48]" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.887, + 0.655, + 0.9 + ], + "angle": 0, + "content": "1https://inivation.com/" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8921" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.488, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.4, + 0.478, + 0.415 + ], + "angle": 0, + "content": "Figure 3. Schematic depiction of the MCTS generation process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.431, + 0.484, + 0.536 + ], + "angle": 0, + "content": "compares some of the most commonly used event representations: binary event frames, time surfaces [38], and voxel grids [70]. For their data-driven feature tracking approach, time surfaces yield the best model performance. The time surface variant Tencode with two more channels encoding the event polarity is proposed in [32]. It is used as model input after conversion to a single-channel frame." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.484, + 0.643 + ], + "angle": 0, + "content": "Time surfaces and its variant Tencode have proven to be an efficient input representation for keypoint detection and tracking algorithms [19, 32, 48], since they assign the highest values to the pixel locations of the most recent events that are usually caused by visual edges. This facilitates finding keypoints along these edges. However, these encodings have shortcomings:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.645, + 0.483, + 0.735 + ], + "angle": 0, + "content": "- Time surfaces are one-channel frames that ignore the event polarity and thereby discard potentially useful information. The Tencode representation is an attempt to overcome this issue. However, since it is converted back to a single channel tensor, there is still no clear separation between positive and negative events." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.736, + 0.484, + 0.902 + ], + "angle": 0, + "content": "- Time surfaces and the Tencode representation depend on the time window length \\(\\Delta t\\). Increasing this parameter not only results in more events being considered but also decreases the contrast between the most recent events and past events. For high optical flow, this negatively impacts the detection of edges. Conversely, a smaller time window increases this contrast but includes fewer events, which can fail to sufficiently capture edges in areas with little optical flow. Choosing a suitable \\(\\Delta t\\) is challenging, as it depends on the camera and scene motion which might be unknown a priori. The bottom part of Figure 3" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.645, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "shows the effect of varying \\(\\Delta t\\): Only the time surfaces in the middle clearly display the net on the top right." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.122, + 0.907, + 0.273 + ], + "angle": 0, + "content": "Our proposed MCTS representation mitigates these problems: It combines time surfaces for each polarity and for various time windows of temporal size \\(\\Delta t_{n}\\), \\(n\\in \\{1,\\ldots ,N\\}\\). We distribute the time window sizes logarithmically to trade off data preprocessing and coverage of a wide variety of visual motion speeds. Events are encoded by their timestamp \\(t\\), their pixel coordinates \\(x\\) and \\(y\\), and their polarity \\(p\\). Mathematically, the MCTS tensor at time \\(\\tau\\) considering all \\(i\\) events \\(\\mathbf{e}_i = (t_i,x_i,y_i,p_i)\\) between time \\(\\tau -\\Delta t_{n}\\) and \\(\\tau\\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.289, + 0.905, + 0.326 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {M C T S} = \\left(\\mathbf {T S} _ {- 1, \\Delta t _ {1}}, \\dots , \\mathbf {T S} _ {- 1, \\Delta t _ {N}}\\right), \\\\ \\left. \\mathbf {T S} _ {+ 1, \\Delta t _ {1}}, \\dots , \\mathbf {T S} _ {+ 1, \\Delta t _ {N}}\\right), \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.336, + 0.905, + 0.368 + ], + "angle": 0, + "content": "with each time surface \\(\\mathbf{TS}\\), initialized with zeros, defined by polarity \\(p\\in \\{-1, + 1\\}\\) and \\(\\Delta t_{n}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.378, + 0.905, + 0.412 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T S} _ {p, \\Delta t _ {n}} \\left(x _ {i}, y _ {i}\\right) = \\max _ {i, p _ {i} = p} \\left(1 - \\frac {\\tau - t _ {i}}{\\Delta t _ {n}}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.421, + 0.882, + 0.436 + ], + "angle": 0, + "content": "The MCTS generation process is visualized in Figure 3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.446, + 0.779, + 0.461 + ], + "angle": 0, + "content": "3.3. Network Architecture & Loss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.468, + 0.907, + 0.634 + ], + "angle": 0, + "content": "As commonly employed in frame-based keypoint detection models [15, 22, 64], we combine a shared backbone with a detector and a descriptor head. Starting with the basic SuperPoint [15] architecture with adjusted loss functions for our temporal matching with pseudo-labels, we train various models with different backbones and hyperparameters and compare their performance. As an alternative approach, we replace the grid-based descriptor head with full resolution prediction and the respective loss functions as in [22, 64]. Please refer to the full results of this ablation study in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.635, + 0.906, + 0.754 + ], + "angle": 0, + "content": "Our resulting architecture consists of a 3-layer MaxViT [63] backbone with a Feature Pyramid Network (FPN) [43] which is also employed in a highly efficient, event-based detector [20]. Combined with grid-based VGG [61] detector and descriptor heads, SuperEvents fully convolutional components can process various input resolutions without retraining. Figure 4 shows the final network architecture." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.755, + 0.906, + 0.815 + ], + "angle": 0, + "content": "The loss function evaluates the network prediction for two corresponding training tensors \\(\\{0,1\\}\\). It combines the individual detector losses \\(L_{\\mathrm{s},0}\\) and \\(L_{\\mathrm{s},1}\\) with the joint descriptor loss \\(L_{\\mathrm{d}}\\) weighted with the constant \\(c_{\\lambda}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.828, + 0.905, + 0.844 + ], + "angle": 0, + "content": "\\[\nL = L _ {\\mathrm {s}, 0} + L _ {\\mathrm {s}, 1} + c _ {\\lambda} L _ {\\mathrm {d}}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Following [15], the detection is implemented as a classification problem of the pixels in image patches, i.e., the pixel indices represent the \"classes\". The detector loss \\( L_{\\mathrm{CE}} \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8922" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.155 + ], + "angle": 0, + "content": "is thus defined as cross-entropy between a predicted patch of keypoint scores \\(\\mathbf{y}_{s_h,w}\\) (dimension \\(8\\times 8\\) plus no-keypoint-dustbin, thus 65) per grid cell (indices \\(h,w\\)) and the pseudolabels \\(\\hat{\\mathbf{y}}_{s_h,w}\\) in one-hot encoding" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.16, + 0.483, + 0.21 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {s}} = \\frac {1}{H _ {\\mathrm {c}} W _ {\\mathrm {c}}} \\sum_ {\\substack {h = 1 \\\\ w = 1}} ^ {H _ {\\mathrm {c}}, W _ {\\mathrm {c}}} L _ {\\mathrm {CE}} \\left(\\mathbf {y} _ {\\mathrm {s} _ {h}, w}, \\hat {\\mathbf {y}} _ {\\mathrm {s} _ {h}, w}\\right), \\tag{5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.215, + 0.483, + 0.246 + ], + "angle": 0, + "content": "for \\( H_{\\mathrm{c}} \\times W_{\\mathrm{c}} \\) grid cells. In the case of multiple keypoint labels in a grid cell, one is randomly selected." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.246, + 0.484, + 0.491 + ], + "angle": 0, + "content": "Unlike other models trained with homographic adaptation, our sparse pseudo-labels do not allow for dense descriptor training for every pixel or grid cell. As in [15], we only consider a single descriptor (256 channels) per grid cell, \\(\\mathbf{y}_{\\mathrm{d}_{h_0},w_0}\\) and \\(\\mathbf{y}_{\\mathrm{d}_{h_1},w_1}\\) (normalized to length one across the channels) with \\((h_0,w_0)\\) and \\((h_1,w_1)\\) denoting the grid indices of tensor 0 and 1. Successful keypoint matching requires similar descriptors for correspondences and distinct descriptors for unrelated keypoints. We measure descriptor similarity with the dot product \\(d_{h_1,w_1}^{h_0,w_0} = \\mathbf{y}_{\\mathrm{d}_{h_0,w_0}}\\cdot \\mathbf{y}_{\\mathrm{d}_{h_1,w_1}}\\). Our pseudo label \\(\\hat{y}_{h_1,w_1}^{h_0,w_0} = 1\\) indicates that two descriptors at the respective grid cell locations correspond, otherwise \\(\\hat{y}_{h_1,w_1}^{h_0,w_0} = 0\\). Since the majority of the descriptors do not correspond, the partial loss of descriptor matches is weighted with the constant \\(c_{\\mathrm{d}}\\). Mathematically, we formulate the descriptor loss as" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.497, + 0.48, + 0.56 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {d}} = \\sum_ {\\substack {h _ {0} = 1 \\\\ w _ {0} = 1}} ^ {H _ {\\mathrm {c}}, W _ {\\mathrm {c}}} \\sum_ {\\substack {h _ {1} = 1 \\\\ w _ {1} = 1}} ^ {H _ {\\mathrm {c}}, W _ {\\mathrm {c}}} \\left\\{ \\begin{array}{l l} c _ {\\mathrm {d}} \\max (0, c _ {\\mathrm {p}} - d _ {h _ {1}, w _ {1}} ^ {h _ {0}, w _ {0}}), & \\hat {y} _ {h _ {1}, w _ {1}} ^ {h _ {0}, w _ {0}} = 1, \\\\ \\max (0, d _ {h _ {1}, w _ {1}} ^ {h _ {0}, w _ {0}} - c _ {\\mathrm {n}}), & \\hat {y} _ {h _ {1}, w _ {1}} ^ {h _ {0}, w _ {0}} = 0. \\end{array} \\right. \\tag{6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.562, + 0.483, + 0.621 + ], + "angle": 0, + "content": "The constants \\( c_{\\mathrm{p}} \\) and \\( c_{\\mathrm{n}} \\) lead to saturation for maximal desired descriptor (dis-)similarity. All descriptors of grid cells without a keypoint label index \\( (\\mathbf{y}_{\\mathrm{d}_{h_0},w_0} = 0 \\) or \\( \\mathbf{y}_{\\mathrm{d}_{h_1},w_1} = 0) \\) do not contribute to the loss." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.622, + 0.483, + 0.804 + ], + "angle": 0, + "content": "During inference, we employ a simplified, yet more effective Non-Minimum-Suppression (NMS) [51] approach in the context of keypoint detection. Instead of the common approach to model keypoints as fixed-size boxes and then apply classical NMS as in [15, 32], we find local maxima in the predicted keypoint score heatmap. In contrast to classical NMS, our method only considers local maxima and not all keypoints candidates sufficiently far from other points with higher scores. Additionally, it employs parallelizable operations instead of consecutive updates of the keypoints sorted by score. Every predicted score \\( s \\) at the pixel coordinates \\( (u, v) \\) is considered for a potential keypoint if" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.813, + 0.482, + 0.85 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} s [ u, v ] > s [ u + \\Delta u, v + \\Delta v ] \\\\ \\forall \\Delta u, \\Delta v \\in \\left[ - c _ {\\mathrm {b}}, \\dots , - 1, 1, \\dots , c _ {\\mathrm {b}} \\right]. \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "with \\( c_{\\mathrm{b}} \\) determining the size of the local neighborhood. The keypoint candidates are further filtered by a minimal required score \\( c_{\\mathrm{s}} \\), or the \\( n_{\\mathrm{s}} \\) highest scores are selected." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.905, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.28, + 0.907, + 0.351 + ], + "angle": 0, + "content": "Figure 4. SuperEvent network architecture and tensor dimensions: a shared transformer backbone is combined with a detector and a descriptor head. The components of the Convolution-Attention blocks \\((Conv. + Attn.)\\) and the VGG blocks are displayed on the bottom left. Activations are omitted for simplicity." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.365, + 0.907, + 0.421 + ], + "angle": 0, + "content": "Table 1. Inference times of our compiled PyTorch model on GPU for common event-stream resolutions (cropped to compatible shape) using float16 precision (same accuracy as float32 in our experiments). NMS requires an additional \\(0.4\\mathrm{ms}\\) on average." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.431, + 0.907, + 0.477 + ], + "angle": 0, + "content": "
160 × 240240 × 320480 × 640720 × 1280
1.3 ms2.6 ms7.8 ms23.2 ms
" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.512, + 0.645, + 0.529 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.906, + 0.629 + ], + "angle": 0, + "content": "To validate the effectiveness of the presented approach, we conduct a series of experiments and ablation studies. Firstly, we evaluate the detection and description of keypoints by estimating the camera pose changes based on the matched points. We outperform other keypoint detectors and trackers [10, 32, 33] by a large margin." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.906, + 0.703 + ], + "angle": 0, + "content": "In our second experiment, we integrate SuperEvent into the SLAM framework OKVIS2 [40] originally developed for frame-based cameras. We achieve convincing results on the widely used event-SLAM benchmarks [18, 35] without extensive tuning, showcasing SuperEvent's versatility." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.906, + 0.765 + ], + "angle": 0, + "content": "For ablation studies, more qualitative examples, a detailed description of the pose estimation benchmark, and a comparison to frame-based matching for high-speed/ HDR scenarios, please refer to our Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.774, + 0.731, + 0.789 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We generated training data as introduced in Section 3.1. To expose the model to various scenes and camera intrinsics, we combine multiple datasets [14, 31, 39, 56, 69], obtaining more than \\(100\\mathrm{k}\\) training sample pairs with pseudolabels. We implement SuperEvent in PyTorch [4] and train it with the Adam [34] optimizer (learning rate \\(1e^{-4}\\), betas 0.9, 0.999) for 10 epochs." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8923" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.126, + 0.09, + 0.447, + 0.103 + ], + "angle": 0, + "content": "Table 2. Pose estimation on the Event Camera dataset." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.115, + 0.482, + 0.221 + ], + "angle": 0, + "content": "
MethodECD: Pose Estimation AUC in %
@5°@10°@20°
LLAK [10]0.71.42.1
RATE [33]3.38.418.0
EventPoint [32]1.63.05.4
SuperEvent (ours)22.735.846.7
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.236, + 0.482, + 0.327 + ], + "angle": 0, + "content": "We employ the network architecture as in Section 3.3 with the loss function hyperparameters \\( c_{\\lambda} = 10 \\), \\( c_{\\mathrm{d}} = 0.5 \\), \\( c_{\\mathrm{p}} = 1.0 \\), and \\( c_{\\mathrm{n}} = 0.2 \\). We encode events as a 10-channel MCTS with 5 time window sizes \\( \\Delta t \\) logarithmically distributed from \\( 0.001 \\, \\mathrm{s} \\) to \\( 0.1 \\, \\mathrm{s} \\). The detection threshold is set to \\( c_{\\mathrm{s}} = 0.01 \\) and the NMS neighborhood \\( c_{\\mathrm{b}} = 2 \\) pixels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.327, + 0.482, + 0.372 + ], + "angle": 0, + "content": "We evaluate on a desktop PC with an Intel Core i5-13600 processor, an NVIDIA GeForce GTX 4070 GPU, and 32 GB RAM. Inference times are shown in Table 1." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.381, + 0.251, + 0.395 + ], + "angle": 0, + "content": "4.2. Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.402, + 0.483, + 0.674 + ], + "angle": 0, + "content": "A common benchmark for evaluating frame-based keypoint detectors is pose estimation [22, 59, 62, 64]. These works use datasets such as ScanNet [12] and MegaDepth [42] containing various images of the same scene with the associated ground truth camera poses. To create a similar event-based benchmark, we employ sequences with ground truth poses from real-world dynamic datasets. We search for uniformly distributed pose changes between \\(1^{\\circ}\\) and \\(45^{\\circ}\\). The upper bound is chosen to increase the probability of visual overlap. However, since we do not know the ground truth geometry of the scene, it cannot be guaranteed that every sample contains sufficient information to recover the ground truth pose. Next, we process the event stream with various approaches to detect and match keypoints. After outlier removal with Random Sample Consensus (RANSAC) [16], we estimate the camera pose and calculate the error with respect to the ground truth. Finally, we report the area-under-curve (AUC) metric for different thresholds." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.795 + ], + "angle": 0, + "content": "We compare the pose estimation capability of SuperEvent to the 3 approaches LLAK [10], RATE [33], and EventPoint [32]. While we use the open-source implementations of LLAK and RATE, we re-implemented and trained EventPoint by carefully following the steps outlined in the paper. We were unable to evaluate the approach SD2Event [19] due to unpublished code and missing implementation details in the respective work." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The experiment is conducted on two diverse datasets: Firstly, the Event Camera Dataset (ECD) [50] is widely used to evaluate event-based approaches, as it was one of the first publicly available datasets. It is challenging due to the low resolution \\((180\\times 240)\\), the noisy output of the initiation DAVIS240C event camera, and fast motion changes. Secondly, we use the more recent dataset Event-aided Di" + }, + { + "type": "table_caption", + "bbox": [ + 0.515, + 0.09, + 0.905, + 0.104 + ], + "angle": 0, + "content": "Table 3. Pose estimation on Event-aided Direct Sparse Odometry." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.115, + 0.905, + 0.221 + ], + "angle": 0, + "content": "
MethodEDS: Pose Estimation AUC in %
@5°@10°@20°
LLAK [10]0.50.71.0
RATE [33]2.15.110.3
EventPoint [32]1.62.85.2
SuperEvent (ours)15.226.440.1
" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.239, + 0.905, + 0.329 + ], + "angle": 0, + "content": "rect Sparse Odometry (EDS) [29]. In contrast to our training data, this dataset is recorded with a Prophesee Gen \\(3.1^{2}\\) event camera with a higher resolution of \\(480 \\times 640\\) pixels and an over-proportionally increased event rate. Fast, inconsistent camera motions and a high dynamic range make pose estimation challenging." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.33, + 0.906, + 0.526 + ], + "angle": 0, + "content": "Table 2 shows that SuperEvent outperforms all baselines by a large margin. On the ECD dataset, its advantage is the greatest for the high-precision estimations below \\(5^{\\circ}\\) error, confirming the effectiveness of our method. In contrast, we found that LLAK detects many keypoints but cannot track the vast majority of them. EventPoint's matching capability is severely limited, possibly due to their static assumptions during training inherently contradicting the dynamic nature of event cameras. Finally, RATE achieves decent results for estimations with up to \\(20^{\\circ}\\) error, but since it cannot recover from tracking loss, it depends on consistent and precise tracking of most keypoints. As seen in the results, this requirement cannot be stably fulfilled." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.527, + 0.906, + 0.663 + ], + "angle": 0, + "content": "Table 3 indicates a slight decrease in SuperEvent's pose estimation performance on the EDS dataset compared to ECD. Since all approaches perform worse, we attribute this effect to the dataset being more challenging. Interestingly, the performance of SuperEvent for \\(20^{\\circ}\\) precision is still in the same region as for the ECD dataset. Considering the different properties of the Prophesee camera compared to the DAVIS camera employed in the training sequences, these results confirm the generalization capability of SuperEvent." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.673, + 0.821, + 0.687 + ], + "angle": 0, + "content": "4.3. Stereo Event-Visual Inertial SLAM" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.696, + 0.906, + 0.877 + ], + "angle": 0, + "content": "We demonstrate the usefulness of SuperEvent with a downstream experiment. To date, frame-based visual odometry and SLAM systems largely outperform their event-based counterparts, attributed to their decades-long lead in development. By integrating SuperEvent into the frame-based VI-SLAM system OKVIS2 [40], we enable this framework to process the keypoints detected in the event stream. We compare the performance of the resulting event-only system to other recent event-based Odometry and SLAM approaches on the TUM-VIE dataset [35] in Table 4. Moreover, note that the next best results of the data-driven Mono E(I)O systems DEVO and DEIO are achieved only" + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.68, + 0.901 + ], + "angle": 0, + "content": "\\(^{2}\\)https://www_prophesee.ai/" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8924" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 4. Results on the small scale mocap sequences of TUM-VIE dataset, ATE in cm. The first three gray results require scale alignment with the ground truth; all other approaches estimate the absolute scale. Baseline numbers taken from [27, 53]." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.13, + 0.907, + 0.304 + ], + "angle": 0, + "content": "
MethodModality1d-trans3d-trans6dofdeskdesk2Average
EVO [55] w/ scale alignmentMono E7.5012.5085.5054.1075.2046.96
DEVO [36] w/ scale alignmentMono E0.501.101.601.701.001.18
DEIO [27] w/ scale alignmentMono E + IMU0.421.111.371.360.731.00
DEIO [27]Mono E + IMU1.081.121.391.411.191.24
ESVO [67]Stereo E12.5417.1913.4612.924.4212.11
ES-PTAM [21]Stereo E1.058.5310.252.507.205.91
ICRA'24 [54]Stereo E + IMU3.8518.90failed8.999.47-
ESVO2 [53]Stereo E + IMU3.337.263.216.164.024.78
OKVIS2 [40] + SuperEvent (ours)Stereo E + IMU0.440.890.430.580.410.55
(ours without loop closure)(0.43)(0.89)(0.43)(0.70)(0.40)(0.57)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.326, + 0.484, + 0.385 + ], + "angle": 0, + "content": "Table 5. Results on VECTor [18] large scale sequences. Our results are obtained with online calibration. Baseline numbers are taken from [27] and [36]. DEVO [36] cannot recover the absolute scale. Sequences marked \\(\\cdot -\\cdot\\) are not evaluated in [27]. ATE in cm." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.393, + 0.484, + 0.48 + ], + "angle": 0, + "content": "
Methodcorr.-dollycorr.-walkunits-dollyunits-scooterschool-dollyschool-scooter
Ours33.13133.16122.6159.0569.9639.07
DEIO [27]492.65325.00826.38304.14--
ESVO [67]failedfailedfailedfailed1371.0983.0
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.497, + 0.484, + 0.558 + ], + "angle": 0, + "content": "after aligning the scale of the estimated trajectories with ground truth. However, real-world applications, e.g., VR and robotics, typically require accurate metric scale. Results on the VECtor dataset [18] are listed in Table 5." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.558, + 0.483, + 0.784 + ], + "angle": 0, + "content": "OKVIS2's loop closure capability relies on long-term descriptor matching. While its effect on the small-scale sequences is limited, we evaluate OKVIS2 + SuperEvent on the loop-floor sequences of the TUM-VIE dataset, since these are explicitly designed to test a system's loop-closure capabilities. Thanks to SuperEvent, the OKVIS2 loop-closure feature works reliably with events, as shown in Table 6. Figure 5 qualitatively compares the estimated trajectories of the sequence loop-floor0. Without loop closure, there is significant drift - albeit not out of the ordinary compared to frame-based VI-SLAM. However, with loop closure enabled, the starting room is recognized and the drift is corrected. This demonstrates the robustness of SuperEvent, achieving successful descriptor matching even with minutes of time between observations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.8, + 0.212, + 0.815 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We presented SuperEvent, a novel keypoint detector and descriptor for event streams. During its training, we consider the time-dependent keypoint appearance and employ pseudo-labels detected on frames. Combined with the flexible MCTS representation and a transformer back" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.326, + 0.907, + 0.355 + ], + "angle": 0, + "content": "Table 6. Effect of loop closure of OKVIS2 [40] + SuperEvent on TUM-VIE loop-floor sequences, ATE in cm." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.366, + 0.906, + 0.438 + ], + "angle": 0, + "content": "
loop-floor0123
Estimated length349 m316 m279 m303 m
Loop closure4.964.648.924.74
W/o loop closure132.11161.92116.00129.17
" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.453, + 0.9, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.579, + 0.907, + 0.609 + ], + "angle": 0, + "content": "Figure 5. OKVIS2 + SuperEvent's estimated trajectories of TUMVIE sequence loop-floor0 with and without loop closure." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.623, + 0.906, + 0.684 + ], + "angle": 0, + "content": "bone, it achieves SOTA performance on keypoint-based pose estimation. SuperEvent's integration into frame-based downstream applications enables event vision in existing keypoint-based systems." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.685, + 0.906, + 0.775 + ], + "angle": 0, + "content": "As future work, SuperEvent can be combined with a frame-based detector to realize systems that simultaneously process frames and events, exploiting their complementary advantages. Also, with a similar approach to SuperEvent, an event-based line detector could improve the performance of downstream applications in man-made environments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.775, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Finally, our training data is generated to enable consistent feature description independent of their motion-dependent appearance. However, since we employ real sequences, consecutive samples are biased to having similar motion directions. This is reflected in the performance of SuperEvent, whose descriptor matchability suffers under strong motion changes. While this is an inherent property of event cameras, a different approach to generating training data could mitigate it." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8925" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.123, + 0.188, + 0.138 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.148, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[1] Ignacio Alzugaray and Margarita Chli. Asynchronous corner detection and tracking for event cameras in real time. IEEE Robotics and Automation Letters, 3(4):3177-3184, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.192, + 0.483, + 0.246 + ], + "angle": 0, + "content": "[2] Ignacio Alzugaray and Margarita Chli. Ace: An efficient asynchronous corner tracker for event cameras. In 2018 International Conference on 3D Vision (3DV), pages 653-661. IEEE, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.249, + 0.483, + 0.319 + ], + "angle": 0, + "content": "[3] Ignacio Alzugaray and Margarita Chli. HASTE: Multi-hypothesis asynchronous speeded-up tracking of events. In 31st British Machine Vision Virtual Conference (BMVC 2020), page 744. ETH Zurich, Institute of Robotics and Intelligent Systems, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.321, + 0.482, + 0.569 + ], + "angle": 0, + "content": "[4] Jason Ansel, Edward Yang, Horace He, Natalia Gimelshein, Animesh Jain, Michael Voznesensky, Bin Bao, Peter Bell, David Berard, Evgeni Burovski, Geeta Chauhan, Anjali Choudria, Will Constable, Alban Desmaison, Zachary DeVito, Elias Ellison, Will Feng, Jiong Gong, Michael Gschwind, Brian Hirsh, Sherlock Huang, Kshitteej Kalambarkar, Laurent Kirsch, Michael Lazos, Mario Lezcano, Yanbo Liang, Jason Liang, Yinghai Lu, CK Luk, Bert Maher, Yunjie Pan, Christian Puhrsch, Matthias Reso, Mark Saroufim, Marcos Yukio Siraichi, Helen Suk, Michael Suo, Phil Tillet, Eikan Wang, Xiaodong Wang, William Wen, Shunting Zhang, Xu Zhao, Keren Zhou, Richard Zou, Ajit Mathews, Gregory Chanan, Peng Wu, and Soumith Chintala. PyTorch 2: Faster Machine Learning Through Dynamic Python Bytecode Transformation and Graph Compilation. In 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2 (ASPLOS '24). ACM, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.572, + 0.483, + 0.614 + ], + "angle": 0, + "content": "[5] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-up robust features (SURF). Computer vision and image understanding, 110(3):346-359, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.616, + 0.483, + 0.699 + ], + "angle": 0, + "content": "[6] Michael Calonder, Vincent Lepetit, Christoph Strecha, and Pascal Fua. Brief: Binary robust independent elementary features. In Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part IV 11, pages 778-792. Springer, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.701, + 0.483, + 0.769 + ], + "angle": 0, + "content": "[7] Carlos Campos, Richard Elvira, Juan J. Gómez, José M. M. Montiel, and Juan D. Tardós. ORB-SLAM3: An accurate open-source library for visual, visual-inertial and multi-map SLAM. IEEE Transactions on Robotics, 37(6):1874-1890, 2021. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.772, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[8] Peiyu Chen, Weipeng Guan, and Peng Lu. ESVIO: Event-based stereo visual inertial odometry. IEEE Robotics and Automation Letters, 8(6):3661-3668, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.816, + 0.483, + 0.884 + ], + "angle": 0, + "content": "[9] Philippe Chiberre, Etienne Perot, Amos Sironi, and Vincent Lepetit. Detecting stable keypoints from events through image gradient prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1387-1394, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.887, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[10] Philippe Chiberre, Etienne Perot, Amos Sironi, and Vincent" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.148, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Lepetit. Long-lived accurate keypoints in event streams. arXiv preprint arXiv:2209.10385, 2022. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[11] Xavier Clady, Sio-Hoi Ieng, and Ryad Benosman. Asynchronous event-based corner detection and matching. Neural Networks, 66:91-106, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.165, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[12] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828–5839, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.905, + 0.277 + ], + "angle": 0, + "content": "[13] Laurent Dardelet, Ryad Benosman, and Sio-Hoi Ieng. An event-by-event feature detection and tracking invariant to motion direction and velocity. Authorea Preprints, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.279, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[14] Jeffrey Delmerico, Titus Cieslewski, Henri Rebecq, Matthias Faessler, and Davide Scaramuzza. Are we ready for autonomous drone racing? the uzh-fpv drone racing dataset. In 2019 International Conference on Robotics and Automation (ICRA), pages 6713-6719. IEEE, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.35, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[15] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperPoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224-236, 2018. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.42, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[16] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.477, + 0.905, + 0.546 + ], + "angle": 0, + "content": "[17] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.548, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[18] Ling Gao, Yuxuan Liang, Jiaqi Yang, Shaoxun Wu, Chenyu Wang, Jiaben Chen, and Laurent Kneip. Vector: A versatile event-centric benchmark for multi-sensor slam. IEEE Robotics and Automation Letters, 7(3):8217-8224, 2022. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.618, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[19] Yuan Gao, Yuqing Zhu, Xinjun Li, Yimin Du, and Tianzhu Zhang. SD2Event: Self-supervised learning of dynamic detectors and contextual descriptors for event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3055-3064, 2024. 1, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[20] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13884-13893, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[21] Suman Ghosh, Valentina Cavinato, and Guillermo Gallego. ES-PTAM: Event-based stereo parallel tracking and mapping. In European Conference on Computer Vision (ECCV) Workshops, 2024. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[22] Pierre Gleize, Weiyao Wang, and Matt Feiszli. SiLK: Simple learned keypoints. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 22442-22451, 2023. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] Arren Glover, Aiko Dinale, Leandro De Souza Rosa, Simeon Bamford, and Chiara Bartolozzi. luvHarris: A practical corner detector for event-cameras. IEEE Transactions on" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8926" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.483, + 0.12 + ], + "angle": 0, + "content": "Pattern Analysis and Machine Intelligence, 44(12):10087-10098, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[24] Weipeng Guan and Peng Lu. Monocular event visual inertial odometry based on event-corner using sliding windows graph-based optimization. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2438-2445, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[25] Weipeng Guan, Peiyu Chen, Yuhan Xie, and Peng Lu. PL- EVIO: Robust monocular event-based visual inertial odometry with point and line features. IEEE Transactions on Automation Science and Engineering, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.305 + ], + "angle": 0, + "content": "[26] Weipeng Guan, Peiyu Chen, Huibin Zhao, Yu Wang, and Peng Lu. EVI-SAM: Robust, real-time, tightly-coupled event-visual-inertial state estimation and 3d dense mapping. Advanced Intelligent Systems, 6(12):2400243, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[27] Weipeng Guan, Fuling Lin, Peiyu Chen, and Peng Lu. DEIO: Deep event inertial odometry. arXiv preprint arXiv:2411.03928, 2024. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.377 + ], + "angle": 0, + "content": "[28] Christopher G. Harris and M. J. Stephens. A combined corner and edge detector. In Alvey Vision Conference, 1988. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.379, + 0.482, + 0.434 + ], + "angle": 0, + "content": "[29] Javier Hidalgo-Carrió, Guillermo Gallego, and Davide Scaramuzza. Event-aided direct sparse odometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5781–5790, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.436, + 0.482, + 0.504 + ], + "angle": 0, + "content": "[30] Sumin Hu, Yeeun Kim, Hyungtae Lim, Alex Junho Lee, and Hyun Myung. eCDT: Event clustering for simultaneous feature detection and tracking. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3808-3815. IEEE, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.506, + 0.482, + 0.589 + ], + "angle": 0, + "content": "[31] Yuhuang Hu, Jonathan Binas, Daniel Neil, Shih-Chii Liu, and Tobi Delbruck. DDD20 end-to-end event camera driving dataset: Fusing frames and events with deep learning for improved steering prediction. In 2020 IEEE 23rd International Conference on Intelligent Transportation Systems (ITSC), pages 1-6. IEEE, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[32] Ze Huang, Li Sun, Cheng Zhao, Song Li, and Songzhi Su. EventPoint: Self-supervised interest point detection and description for event-based camera. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5396-5405, 2023. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[33] Mikihiro Ikura, Cedric Le Gentil, Marcus G. Müller, Florian Schuler, Atsushi Yamashita, and Wolfgang Stürzl. RATE: Real-time asynchronous feature tracking with event cameras. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 11662-11669, 2024. 1, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[34] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[35] Simon Klenk, Jason Chui, Nikolaus Demmel, and Daniel Cremers. TUM-VIE: The TUM stereo visual-inertial event dataset. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8601-8608. IEEE, 2021. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[36] Simon Klenk, Marvin Motzet, Lukas Koestler, and Daniel Cremers. Deep event visual odometry. In 2024 International Conference on 3D Vision (3DV), pages 739-749. IEEE, 2024. 4, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[37] Beat Kueng, Elias Mueggler, Guillermo Gallego, and Davide Scaramuzza. Low-latency visual odometry using event-based feature tracks. In 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 16-23. IEEE, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[38] Xavier Lagorce, Garrick Orchard, Francesco Galluppi, Bertram E. Shi, and Ryad B. Benosman. HOTS: A hierarchy of event-based time-surfaces for pattern recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(7):1346-1359, 2017. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[39] Alex Junho Lee, Younggun Cho, Young-sik Shin, Ayoung Kim, and Hyun Myung. ViViD++: Vision for visibility dataset. IEEE Robotics and Automation Letters, 7(3):6282-6289, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[40] Stefan Leutenegger. OKVIS2: Realtime scalable visual-inertial SLAM with loop closure, 2022. 1, 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[41] Stefan Leutenegger, Margarita Chli, and Roland Y. Siegwart. BRISK: Binary robust invariant scalable keypoints. In 2011 International Conference on Computer Vision, pages 2548-2555, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[42] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2041-2050, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[43] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[44] Zhe Liu, Dianxi Shi, Ruihao Li, and Shaowu Yang. ESVIO: event-based stereo visual-inertial odometry. Sensors, 23(4): 1998, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.589 + ], + "angle": 0, + "content": "[45] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60:91-110, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[46] Florian Mahlknecht, Daniel Gehrig, Jeremy Nash, Friedrich M Rockenbauer, Benjamin Morrell, Jeff Delaune, and Davide Scaramuzza. Exploring event camera-based odometry for planetary robots. IEEE Robotics and Automation Letters, 7(4):8651-8658, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[47] Jacques Manderscheid, Amos Sironi, Nicolas Bourdis, Davide Migliore, and Vincent Lepetit. Speed invariant time surface for learning to detect corner points with event-based cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10245-10254, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[48] Nico Messikommer, Carter Fang, Mathias Gehrig, and Davide Scaramuzza. Data-driven feature tracking for event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5642-5651, 2023. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[49] Elias Mueggler, Chiara Bartolozzi, and Davide Scaramuzza. Fast event-based corner detection. Proceedings of the British Machine Vision Conference, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[50] Elias Mueggler, Henri Rebecq, Guillermo Gallego, Tobi Delbruck, and Davide Scaramuzza. The event-camera dataset and simulator: Event-based data for pose estimation, visual" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8927" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "odometry, and slam. The International Journal of Robotics Research, 36(2):142-149, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.176 + ], + "angle": 0, + "content": "[51] Alexander Neubeck and Luc Van Gool. Efficient nonmaximum suppression. In 18th international conference on pattern recognition (ICPR'06), pages 850-855. IEEE, 2006. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[52] Zhenjiang Ni, Aude Bolopion, Joel Agnus, Ryad Benosman, and Stephane Regnier. Asynchronous event-based visual shape tracking for stable haptic feedback in microrobotics. IEEE Transactions on Robotics, 28(5):1081-1089, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.482, + 0.291 + ], + "angle": 0, + "content": "[53] Junkai Niu, Sheng Zhong, Xiuyuan Lu, Shaojie Shen, Guillermo Gallego, and Yi Zhou. ESVO2: Direct visual-inertial odometry with stereo event cameras. arXiv e-prints, pages arXiv-2410, 2024. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[54] Junkai Niu, Sheng Zhong, and Yi Zhou. Imu-aided event-based stereo visual odometry. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 11977–11983, 2024. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[55] Henri Rebecq, Timo Horstschäfer, Guillermo Gallego, and Davide Scaramuzza. EVO: A geometric approach to event-based 6-dof parallel tracking and mapping in real time. IEEE Robotics and Automation Letters, 2(2):593-600, 2016. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[56] Juan Pablo Rodríguez-Gómez, Raul Tapia, Julio L Paneque, Pedro Grau, Augusto Gómez Eguíluz, Jose Ramiro Martínez-de Dios, and Anibal Ollero. The GRIFFIN perception dataset: Bridging the gap between flapping-wing flight and robotic perception. IEEE Robotics and Automation Letters, 6(2):1066-1073, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[57] Edward Rosten and Tom Drummond. Machine learning for high-speed corner detection. In Computer Vision-ECCV 2006: 9th European Conference on Computer Vision, Graz, Austria, May 7-13, 2006. Proceedings, Part I 9, pages 430-443. Springer, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[58] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. ORB: An efficient alternative to SIFT or SURF. In 2011 International Conference on Computer Vision, pages 2564-2571, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[59] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 1, 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[60] Jianbo Shi and Tomasi. Good features to track. In 1994 Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 593-600, 1994. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[61] Karen Simonyan. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[62] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[63] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. MaxViT: Multi-axis vision transformer. In European conference on computer vision, pages 459-479. Springer, 2022. 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[64] Michal Tyszkiiewicz, Pascal Fua, and Eduard Trulls. DISK: Learning local features with policy gradient. Advances in Neural Information Processing Systems, 33:14254-14265, 2020. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[65] Valentina Vasco, Arren Glover, and Chiara Bartolozzi. Fast event-based harris corner detection exploiting the advantages of event-driven cameras. In 2016 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 4144-4149. IEEE, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[66] Antoni Rosinol Vidal, Henri Rebecq, Timo Horstschaefer, and Davide Scaramuzza. Ultimate SLAM? Combining events, images, and IMU for robust visual slam in hdr and high-speed scenarios. IEEE Robotics and Automation Letters, 3(2):994-1001, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[67] Yi Zhou, Guillermo Gallego, and Shaojie Shen. Event-based stereo visual odometry. IEEE Transactions on Robotics, 37 (5):1433-1450, 2021. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[68] Alex Zihao Zhu, Nikolay Atanasov, and Kostas Daniilidis. Event-based feature tracking with probabilistic data association. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 4465-4470. IEEE, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[69] Alex Zihao Zhu, Dinesh Thakur, Tolga Özaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The multi-vehicle stereo event camera dataset: An event camera dataset for 3d perception. IEEE Robotics and Automation Letters, 3 (3):2032–2039, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.528 + ], + "angle": 0, + "content": "[70] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based optical flow using motion compensation. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, 2018. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8928" + } + ] +] \ No newline at end of file diff --git a/2025/SuperEvent_ Cross-Modal Learning of Event-based Keypoint Detection for SLAM/2e684548-d342-4d0d-9d82-edd690c021a1_origin.pdf b/2025/SuperEvent_ Cross-Modal Learning of Event-based Keypoint Detection for SLAM/2e684548-d342-4d0d-9d82-edd690c021a1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2215f94eec34c13b58f11a66f67de39e39e9296c --- /dev/null +++ b/2025/SuperEvent_ Cross-Modal Learning of Event-based Keypoint Detection for SLAM/2e684548-d342-4d0d-9d82-edd690c021a1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da0cf40872b1d7c4dc0f224fd83708d41347eeaa0afadc4654600d958821047f +size 4661944 diff --git a/2025/SuperEvent_ Cross-Modal Learning of Event-based Keypoint Detection for SLAM/full.md b/2025/SuperEvent_ Cross-Modal Learning of Event-based Keypoint Detection for SLAM/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8f5fb3c49ecdf496968b6dec7cd4b8795bae6444 --- /dev/null +++ b/2025/SuperEvent_ Cross-Modal Learning of Event-based Keypoint Detection for SLAM/full.md @@ -0,0 +1,336 @@ +# SuperEvent: Cross-Modal Learning of Event-based Keypoint Detection for SLAM + +Yannick Burkhardt $^{1,2,3}$ + +Simon Schaefer $^{1,2,3}$ + +Stefan Leutenegger $^{1,2,3}$ + +# Abstract + +Event-based keypoint detection and matching holds significant potential, enabling the integration of event sensors into highly optimized Visual SLAM systems developed for frame cameras over decades of research. Unfortunately, existing approaches struggle with the motion-dependent appearance of keypoints and the complex noise prevalent in event streams, resulting in severely limited feature matching capabilities and poor performance on downstream tasks. To mitigate this problem, we propose SuperEvent, a data-driven approach to predict stable keypoints with expressive descriptors. Due to the absence of event datasets with ground truth keypoints labels, we leverage existing frame-based keypoint detectors on readily available event-aligned and synchronized gray-scale frames for self-supervision: we generate temporally sparse keypoint pseudo-labels considering that events are a product of both scene appearance and camera motion. Combined with our novel, information-rich event representation, we enable SuperEvent to effectively learn robust keypoint detection and description in event streams. Finally, we demonstrate the usefulness of SuperEvent by its integration into a modern sparse keypoint and descriptor-based SLAM framework originally developed for traditional cameras, surpassing the state-of-the-art in event-based SLAM by a wide margin. Source code is available at ethz-mrl.github.io/SuperEvent. + +# 1. Introduction + +Event cameras offer exciting advantages over their frame-based counterparts, such as an increased temporal resolution, little motion blur, and a high dynamic range. These properties promise to enhance robustness for robot estimation and perception tasks under fast motion and uncontrollable lighting conditions. However, processing the sparse and asynchronous output of event cameras requires fundamentally different algorithms than traditional frames. The decades-long research edge in frame-based vision resulted + +![](images/45605cbf0dc83051639d4ccbd9c140befef04e3d84f828eeb40cc650d1134b5e.jpg) + +![](images/1c2f2ba40ef732dfce855fd108e8099eb6d1e6e20f1c654471e2aeebdc7ae1e5.jpg) +Figure 1. Detections on the sequence rec1499023756 of the DDD20 [31] dataset (not used for training). Top: Pseudolabels from SuperPoint [15] and SuperGlue [59] on the gray-scale frames. Bottom: Matched keypoints from SuperEvent in the event stream at the corresponding time stamps. + +in most hardware and software being highly optimized for frame processing. A common approach for tasks like Structure-from-Motion, Visual Odometry (VO), Simultaneous Localization and Mapping (SLAM), or Place Recognition is to detect and match keypoints across frames to gain a geometric scene understanding [7, 40]. Consequently, a robust keypoint detector and descriptor working on the event stream would unlock the potential of integrating event cameras into extensive and highly-developed algorithms developed initially for frame-based vision. + +While several authors already proposed event-based approaches to detect keypoints [9, 10, 19, 32, 33, 47], their matching abilities still show limitations. Most employ nearest-neighbor matching in pixel space, which restricts their usability to visually simple scenes. In an attempt to overcome this issue, more recent approaches learn descriptors in a data-driven fashion. To generate ground truth labels - inspired by frame-based methods [15] - they warp their event representations with a random homography and penalize non-consistent model predictions. This approach forces the models to learn descriptors to match visually similar keypoints. However, it fundamentally ignores the true + +$^{1}$ Technical University of Munich, {yannick.burkhardt, simon.k.schaefer}@tum.de + $^{2}$ ETH Zürich, lestefan@ethz.ch + $^{3}$ Munich Center for Machine Learning (MCML) + +![](images/a8e02256b314e4bc787d1586037a71542fa42ba50c41215b629540d8dd726854.jpg) +Figure 2. Data processing pipeline of SuperEvent: For training, pseudo-labels are generated using a frame-based detector and matcher on the gray-scale frame pairs. We generate spatially synchronized MCTS at the same timestamps and feed them to SuperEvent. The network predictions are compared to the pseudo-labels and the network weights are optimized using backpropagation. During inference, the network predictions can be used to detect keypoints and match their descriptors on the event stream only. + +nature of the event generation process that conflates camera motion with scene appearance and geometry. In essence, matches are generated only from one original set of events per keypoint, thereby producing only limited and potentially also unrealistic training data. We hypothesize that this simplification ultimately results in unstable matching. + +In contrast, we propose training our model to predict descriptors for the keypoint correspondences at different points in time – therefore training with different and actually recorded rather than warped events. Since generating keypoint and match labels manually is intractable, we design a scalable method to exploit the great capabilities of frame-based vision models: generating pseudo-labels in aligned frames and then using them to train our model's predictions. An example is shown in Figure 1. + +To provide the model with rich input data, we propose the novel event representation Multi-Channel Time Surfaces (MCTS) which generalizes the commonly used time surface representation [38] to an $n$ -channel tensor of time surfaces for both polarities and different time window sizes. Since an optimal window size depends inversely on the motion magnitude, this approach further increases the model's robustness to both slow and fast motion. Additionally, we revisit network architecture choices of frame- and event-based vision to optimize our model's performance. Figure 2 shows the complete data processing pipeline. + +In absence of evaluation procedures to test event-based keypoint and descriptor stability, we develop a benchmark to assess the matched keypoint's quality for pose estimation, similar to benchmarks commonly used in frame-based vision [22, 59, 62, 64]. On two datasets recorded with dif + +ferent sensors, settings, resolutions, and scenes, we outperform other data-driven and handcrafted approaches in terms of pose estimation by a large margin, demonstrating the robustness of our approach. We carry out ablation studies to confirm the effectiveness of our design choices. + +Finally, we integrate our model into a state-of-the-art (SOTA) stereo visual inertial SLAM (VI-SLAM) framework [40], replacing the frame-based keypoint detector and descriptor extractor with SuperEvent. Our results surpass the current SOTA on event-based Visual Odometry and SLAM. This demonstrates the great potential of an event-sensor integration into existing, well-developed algorithms for frame-based vision. + +We name our method SuperEvent in honor of the great influence of the works SuperPoint [15] and SuperGlue [59] on the research field of keypoint detection as well as on our model directly: ultimately, their frame-based detections and matches are used as pseudo-labels to teach keypoint detection and description to SuperEvent. + +Contributions: (1) We propose a scalable method to generate training data for interest point detection and matching, leveraging robust existing frame-based keypoint detectors. The resulting data incorporates the real temporal dependence of event generation using real event camera recordings. This enables data-driven models to effectively learn robust keypoint detection and description. (2) We develop the model for event-based vision through key design choices: Our novel MCTS event representation increases the performance of data-driven models by reducing their dependence on motion speed of the scene. A transformer backbone shows to further improve the model's feature ex + +traction. In our conducted experiments, our approach outperforms SOTA detection and matching by a large margin. We publicly release our code to support future event-based vision research. (3) Showcasing the practicality of our method, we integrate the trained model into an existing stereo VI-SLAM approach developed for frame-based vision. The resulting event inertial SLAM system outperforms SOTA approaches on the commonly used TUMVIE [35] and VECtor [18] datasets by a large margin. + +# 2. Related Work + +We begin with a short summary of keypoint detectors for frame-based vision since our work employs some of these findings for model design and training data generation. We then group the event-based works into handcrafted and data-driven keypoint detection, as well as tracking methods. Since our focus in this work lies on keypoint detection and matching based on event data exclusively, we do not consider approaches that rely on combination with other sensors, e.g., frame cameras. For an extensive overview of event-based vision, we refer the interested reader to [17]. + +Frame-based keypoint detection: there exists a vast variety of approaches to detecting and describing keypoints in frames. Traditional, handcrafted methods [5, 6, 28, 41, 45, 58, 60] are still widely used, since they are well studied and usually work robustly with little overhead. More recent, data-driven methods outperform the traditional keypoint detectors with impressive results. SuperPoint [15] is the first method trained with homographic adaptation, a self-supervised training technique of warping input images and their detected labels by randomized homographies. SuperPoint's architecture consists of a shared VGG [61]-backbone with a detector and descriptor head, predicting descriptors in a discrete grid which is interpolated to full resolution. The authors later publish the extension SuperGlue [59], a graph neural network that improves the descriptor matching. Following works propose different architectures trained with homographic adaptation [22, 62, 64]. + +Event-based keypoint detection: handcrafted approaches can be broadly divided into filters and frame reconstruction. Filtering methods work directly on the event stream, attempting to find events originating from corners [1, 11]. While their asynchronous nature offers little processing overhead, downstream tasks are required to process asynchronous data. Existing well-developed downstream solutions for frame-based vision therefore require major modifications. + +Frame reconstruction methods collect events to construct a frame representation, such as binary event frames or time surfaces [38]. Adopting methods from frame-based vision, such as the Harris corner detector [28] in [23, 65] or the FAST detector [57] in [49], these approaches detect corners in these event frames. However, they suffer from the in + +herent differences between reconstructed event frames and traditional frames, resulting in limited robustness of the algorithms developed for frame-based vision. + +Additionally, these handcrafted methods usually employ simple nearest-neighbor matching in pixel space to track keypoints over time. This approach is only reliable for simple scenes since cluttered geometry, high variance in depth, and fast motions cause keypoints to become occluded, tightly clustered, or to jump. Also, the developed heuristics struggle with the ubiquitous and complex noise inherent to event cameras. Usually, they require manual parameter tuning for different camera models, datasets, or sometimes even scenes. + +Event-based keypoint detection: data-driven approaches promise to cope better with the event data and its noise. A series of works [9, 10, 47] train data-driven corner detectors using labels by detecting keypoints on corresponding gray-scale frames that are synthetic or from an HVGA ATIS sensor. While these approaches learn data-driven detection, they only track keypoints by nearest-neighbor matching in pixel space, yielding the aforementioned drawback of limited robustness in complex scenes. Additionally, the employed corner detector for ground truth generation as well as the machine-learning models have limited performance due to their simplicity. + +For these reasons, EventPoint [32] and SD2Event [19] employ more complex neural networks that additionally predict descriptors. While EventPoint uses the same architecture as SuperPoint [15] and finetunes its pre-trained weights, SD2Event employs agent-based attention to learn detection and description. Both approaches are trained with homographic adaptation and varying time window sizes. While this results in rotation-, distortion-, and scale-invariant models, the main challenge of keypoint description remains largely unsolved: Because most events result from scene or camera motion, corresponding keypoints vary in visual appearance. Since robust keypoint matching requires similar descriptor pairs, this motion dependence must be considered. However, homographic adaptation creates static frame pairs with little to no variance in scene motion within the samples. This contradicts the dynamic and motion-dependent nature of events, resulting in unstable keypoint matching between different timestamps. + +Additionally, the authors of EventPoint propose the event representation Tencode, extending time surfaces with two additional channels to encode the events' polarities. However, to match SuperPoint's architecture with single-channel input, the Tencode tensor is then converted to one channel, disregarding the previously introduced separation of polarities. + +Our method enforces robust descriptor matching by employing training data containing pseudo-labels for keypoints and descriptors under varying motion. We avoid + +synthetic event data since current simulators struggle to model realistic noise. Furthermore, our MCTS representation strictly separates the polarities and provides a range of time window sizes, providing additional information and omitting the time window parameter choice. + +Event-based keypoint tracking: while early works fall back on batching events and processing events in a discretized way [37, 52, 68], a promising idea is to exploit their spatial continuous visual flow [2, 13, 30]. Among these approaches is HASTE [3] which calculates the most likely motion hypothesis from a fixed set of translations and rotations for every incoming event. Since HASTE relies on external keypoint initialization, RATE [33] employs a Shi-Tomasi [60] corner detector on the binary event frame, initializing several HASTE instances to track multiple keypoints in parallel. + +A major drawback of the described tracking approaches is their computational overhead when processing each event individually. Also, most methods only track one keypoint, resulting in an almost linear growth of computational cost with increasing number of tracker instances. E.g., RATE tracks only 20 features in real-time for an event stream of $180 \times 40$ pixels resolution – which is hardly sufficient for downstream applications such as SLAM. Additionally, these handcrafted approaches struggle with environmental and hardware changes, as well as the event cameras' noise. + +Our method SuperEvent works reliably in real-time with adjustable frequency and without re-training or excessive parameter tuning for different camera models. It's dense keypoint heatmap predictions barely result in overhead for more detections. In contrast to the tracking approaches, SuperEvent can be directly integrated into downstream algorithms relying on synchronous output and descriptors, e.g., for SLAM with place recognition, as we demonstrate. + +Event-based Odometry: existing Event Odometry (EO) approaches are developed specifically for event processing. While some approaches combine events with frames [8, 25, 26, 46, 66], event-only approaches can be classified as monocular EO [36, 55], monocular EO with IMU [24, 27], stereo EO [21, 67], and stereo EO with IMU [44, 53, 54]. Due to the short history of event cameras, these systems require extensive research and development efforts to work reliably in practice. + +In contrast, SuperEvent can be integrated into existing frame-based VO and SLAM systems. We thereby achieve event-based VI-SLAM leveraging highly accurate frame-based systems, emerging from years of research [7, 40]. + +# 3. Method + +In this section, we describe how we train SuperEvent, introduce the advantages and construction of the MCTS event representation, and detail the network design choices. + +# 3.1. Training Data Generation + +To train SuperEvent, we exclusively employ real event camera recordings. Thus, our model learns to cope with the complex event stream properties, e.g., artifacts and noise, which are difficult to simulate realistically. + +However, manually annotating keypoints in the asynchronous event data stream is very time-consuming and does not scale. Therefore, we propose an approach to generate ground truth data automatically. We take advantage of the pixel-synchronized events and frames of iniVation DAVIS1 event cameras. Employing a highly robust frame-based keypoint detector and matcher such as SuperPoint [15] + SuperGlue [59] on the gray-scale frames, we can obtain temporally sparse pseudo-labels. To generate a diverse set of training data with reliable pseudo-labels, we process every frame $i$ of a sequence in the following way: + +- Firstly, we compute the distances in pixel-space of all key points to their respective matches in frame $i + 1$ . We only consider frames $i$ for training where the distances' median surpasses a threshold $c_{\mathrm{m}}$ . A small median distance indicates that the scene is nearly static, resulting in insufficient event data for a useful prediction. + +- To expose the model to diverse camera pose changes and motion directions, we generate pseudo-labels by matching the descriptors with a subsequent frame $i + j$ . We increase $j$ recursively, resulting in multiple pseudo-labels with increasing time interval between the frames: + +$$ +j \leftarrow j + j _ {\mathrm {s}}, j _ {\mathrm {s}} \sim \mathcal {U} (1, j _ {\max }, \tag {1} +$$ + +with the increment $j_{\mathrm{s}}$ drawn from the uniform distribution $\mathcal{U}(1, j_{\max})$ . $j$ is initialized with 0 before its first update. The maximum step distance $j_{\max}$ is manually adjusted depending on the properties and texture of the sequence, e.g., for sequences with fast motion, $j_{\max}$ is decreased to yield sufficient matches. + +- If the number of detected matches falls below the threshold $c_{\mathrm{m}}$ , these pseudo-labels are discarded, and the next frame $i + 1$ becomes the new reference frame. This procedure ensures sufficient visual overlap and enough texture in the two frames. + +Frames with under- or overexposure, or strong motion blur harm the prediction quality of the frame-based model. While these frames are usually filtered out since there are few matches (if $c_{\mathrm{d}}$ is chosen adequately), we also manually remove scenes that mostly contain too dark, too bright, or blurred frames. + +# 3.2. Multi-channel Time Surface + +To allow for efficient batch processing of event data, the asynchronous stream is converted to a tensor with fixed dimensions. In the literature, various variants have been proposed to encode the event stream. An ablation study in [48] + +1https://inivation.com/ + +![](images/8cda978ae46b17b8ade1a00f9e2ca10b8fb7b681e0a1cf3c83375af6e8cc9a0d.jpg) +Figure 3. Schematic depiction of the MCTS generation process. + +compares some of the most commonly used event representations: binary event frames, time surfaces [38], and voxel grids [70]. For their data-driven feature tracking approach, time surfaces yield the best model performance. The time surface variant Tencode with two more channels encoding the event polarity is proposed in [32]. It is used as model input after conversion to a single-channel frame. + +Time surfaces and its variant Tencode have proven to be an efficient input representation for keypoint detection and tracking algorithms [19, 32, 48], since they assign the highest values to the pixel locations of the most recent events that are usually caused by visual edges. This facilitates finding keypoints along these edges. However, these encodings have shortcomings: + +- Time surfaces are one-channel frames that ignore the event polarity and thereby discard potentially useful information. The Tencode representation is an attempt to overcome this issue. However, since it is converted back to a single channel tensor, there is still no clear separation between positive and negative events. +- Time surfaces and the Tencode representation depend on the time window length $\Delta t$ . Increasing this parameter not only results in more events being considered but also decreases the contrast between the most recent events and past events. For high optical flow, this negatively impacts the detection of edges. Conversely, a smaller time window increases this contrast but includes fewer events, which can fail to sufficiently capture edges in areas with little optical flow. Choosing a suitable $\Delta t$ is challenging, as it depends on the camera and scene motion which might be unknown a priori. The bottom part of Figure 3 + +shows the effect of varying $\Delta t$ : Only the time surfaces in the middle clearly display the net on the top right. + +Our proposed MCTS representation mitigates these problems: It combines time surfaces for each polarity and for various time windows of temporal size $\Delta t_{n}$ , $n\in \{1,\ldots ,N\}$ . We distribute the time window sizes logarithmically to trade off data preprocessing and coverage of a wide variety of visual motion speeds. Events are encoded by their timestamp $t$ , their pixel coordinates $x$ and $y$ , and their polarity $p$ . Mathematically, the MCTS tensor at time $\tau$ considering all $i$ events $\mathbf{e}_i = (t_i,x_i,y_i,p_i)$ between time $\tau -\Delta t_{n}$ and $\tau$ is defined as + +$$ +\begin{array}{l} \mathbf {M C T S} = \left(\mathbf {T S} _ {- 1, \Delta t _ {1}}, \dots , \mathbf {T S} _ {- 1, \Delta t _ {N}}\right), \\ \left. \mathbf {T S} _ {+ 1, \Delta t _ {1}}, \dots , \mathbf {T S} _ {+ 1, \Delta t _ {N}}\right), \tag {2} \\ \end{array} +$$ + +with each time surface $\mathbf{TS}$ , initialized with zeros, defined by polarity $p\in \{-1, + 1\}$ and $\Delta t_{n}$ + +$$ +\mathbf {T S} _ {p, \Delta t _ {n}} \left(x _ {i}, y _ {i}\right) = \max _ {i, p _ {i} = p} \left(1 - \frac {\tau - t _ {i}}{\Delta t _ {n}}\right). \tag {3} +$$ + +The MCTS generation process is visualized in Figure 3. + +# 3.3. Network Architecture & Loss + +As commonly employed in frame-based keypoint detection models [15, 22, 64], we combine a shared backbone with a detector and a descriptor head. Starting with the basic SuperPoint [15] architecture with adjusted loss functions for our temporal matching with pseudo-labels, we train various models with different backbones and hyperparameters and compare their performance. As an alternative approach, we replace the grid-based descriptor head with full resolution prediction and the respective loss functions as in [22, 64]. Please refer to the full results of this ablation study in the Supplementary Material. + +Our resulting architecture consists of a 3-layer MaxViT [63] backbone with a Feature Pyramid Network (FPN) [43] which is also employed in a highly efficient, event-based detector [20]. Combined with grid-based VGG [61] detector and descriptor heads, SuperEvents fully convolutional components can process various input resolutions without retraining. Figure 4 shows the final network architecture. + +The loss function evaluates the network prediction for two corresponding training tensors $\{0,1\}$ . It combines the individual detector losses $L_{\mathrm{s},0}$ and $L_{\mathrm{s},1}$ with the joint descriptor loss $L_{\mathrm{d}}$ weighted with the constant $c_{\lambda}$ + +$$ +L = L _ {\mathrm {s}, 0} + L _ {\mathrm {s}, 1} + c _ {\lambda} L _ {\mathrm {d}}. \tag {4} +$$ + +Following [15], the detection is implemented as a classification problem of the pixels in image patches, i.e., the pixel indices represent the "classes". The detector loss $L_{\mathrm{CE}}$ + +is thus defined as cross-entropy between a predicted patch of keypoint scores $\mathbf{y}_{s_h,w}$ (dimension $8\times 8$ plus no-keypoint-dustbin, thus 65) per grid cell (indices $h,w$ ) and the pseudolabels $\hat{\mathbf{y}}_{s_h,w}$ in one-hot encoding + +$$ +L _ {\mathrm {s}} = \frac {1}{H _ {\mathrm {c}} W _ {\mathrm {c}}} \sum_ {\substack {h = 1 \\ w = 1}} ^ {H _ {\mathrm {c}}, W _ {\mathrm {c}}} L _ {\mathrm {CE}} \left(\mathbf {y} _ {\mathrm {s} _ {h}, w}, \hat {\mathbf {y}} _ {\mathrm {s} _ {h}, w}\right), \tag{5} +$$ + +for $H_{\mathrm{c}} \times W_{\mathrm{c}}$ grid cells. In the case of multiple keypoint labels in a grid cell, one is randomly selected. + +Unlike other models trained with homographic adaptation, our sparse pseudo-labels do not allow for dense descriptor training for every pixel or grid cell. As in [15], we only consider a single descriptor (256 channels) per grid cell, $\mathbf{y}_{\mathrm{d}_{h_0},w_0}$ and $\mathbf{y}_{\mathrm{d}_{h_1},w_1}$ (normalized to length one across the channels) with $(h_0,w_0)$ and $(h_1,w_1)$ denoting the grid indices of tensor 0 and 1. Successful keypoint matching requires similar descriptors for correspondences and distinct descriptors for unrelated keypoints. We measure descriptor similarity with the dot product $d_{h_1,w_1}^{h_0,w_0} = \mathbf{y}_{\mathrm{d}_{h_0,w_0}}\cdot \mathbf{y}_{\mathrm{d}_{h_1,w_1}}$ . Our pseudo label $\hat{y}_{h_1,w_1}^{h_0,w_0} = 1$ indicates that two descriptors at the respective grid cell locations correspond, otherwise $\hat{y}_{h_1,w_1}^{h_0,w_0} = 0$ . Since the majority of the descriptors do not correspond, the partial loss of descriptor matches is weighted with the constant $c_{\mathrm{d}}$ . Mathematically, we formulate the descriptor loss as + +$$ +L _ {\mathrm {d}} = \sum_ {\substack {h _ {0} = 1 \\ w _ {0} = 1}} ^ {H _ {\mathrm {c}}, W _ {\mathrm {c}}} \sum_ {\substack {h _ {1} = 1 \\ w _ {1} = 1}} ^ {H _ {\mathrm {c}}, W _ {\mathrm {c}}} \left\{ \begin{array}{l l} c _ {\mathrm {d}} \max (0, c _ {\mathrm {p}} - d _ {h _ {1}, w _ {1}} ^ {h _ {0}, w _ {0}}), & \hat {y} _ {h _ {1}, w _ {1}} ^ {h _ {0}, w _ {0}} = 1, \\ \max (0, d _ {h _ {1}, w _ {1}} ^ {h _ {0}, w _ {0}} - c _ {\mathrm {n}}), & \hat {y} _ {h _ {1}, w _ {1}} ^ {h _ {0}, w _ {0}} = 0. \end{array} \right. \tag{6} +$$ + +The constants $c_{\mathrm{p}}$ and $c_{\mathrm{n}}$ lead to saturation for maximal desired descriptor (dis-)similarity. All descriptors of grid cells without a keypoint label index $(\mathbf{y}_{\mathrm{d}_{h_0},w_0} = 0$ or $\mathbf{y}_{\mathrm{d}_{h_1},w_1} = 0)$ do not contribute to the loss. + +During inference, we employ a simplified, yet more effective Non-Minimum-Suppression (NMS) [51] approach in the context of keypoint detection. Instead of the common approach to model keypoints as fixed-size boxes and then apply classical NMS as in [15, 32], we find local maxima in the predicted keypoint score heatmap. In contrast to classical NMS, our method only considers local maxima and not all keypoints candidates sufficiently far from other points with higher scores. Additionally, it employs parallelizable operations instead of consecutive updates of the keypoints sorted by score. Every predicted score $s$ at the pixel coordinates $(u, v)$ is considered for a potential keypoint if + +$$ +\begin{array}{l} s [ u, v ] > s [ u + \Delta u, v + \Delta v ] \\ \forall \Delta u, \Delta v \in \left[ - c _ {\mathrm {b}}, \dots , - 1, 1, \dots , c _ {\mathrm {b}} \right]. \tag {7} \\ \end{array} +$$ + +with $c_{\mathrm{b}}$ determining the size of the local neighborhood. The keypoint candidates are further filtered by a minimal required score $c_{\mathrm{s}}$ , or the $n_{\mathrm{s}}$ highest scores are selected. + +![](images/868c56bea998a1a90bce823b64f7881688517b3d0248e358f3e323bf8ab3e9b1.jpg) +Figure 4. SuperEvent network architecture and tensor dimensions: a shared transformer backbone is combined with a detector and a descriptor head. The components of the Convolution-Attention blocks $(Conv. + Attn.)$ and the VGG blocks are displayed on the bottom left. Activations are omitted for simplicity. + +Table 1. Inference times of our compiled PyTorch model on GPU for common event-stream resolutions (cropped to compatible shape) using float16 precision (same accuracy as float32 in our experiments). NMS requires an additional $0.4\mathrm{ms}$ on average. + +
160 × 240240 × 320480 × 640720 × 1280
1.3 ms2.6 ms7.8 ms23.2 ms
+ +# 4. Experiments + +To validate the effectiveness of the presented approach, we conduct a series of experiments and ablation studies. Firstly, we evaluate the detection and description of keypoints by estimating the camera pose changes based on the matched points. We outperform other keypoint detectors and trackers [10, 32, 33] by a large margin. + +In our second experiment, we integrate SuperEvent into the SLAM framework OKVIS2 [40] originally developed for frame-based cameras. We achieve convincing results on the widely used event-SLAM benchmarks [18, 35] without extensive tuning, showcasing SuperEvent's versatility. + +For ablation studies, more qualitative examples, a detailed description of the pose estimation benchmark, and a comparison to frame-based matching for high-speed/ HDR scenarios, please refer to our Supplementary Material. + +# 4.1. Implementation Details + +We generated training data as introduced in Section 3.1. To expose the model to various scenes and camera intrinsics, we combine multiple datasets [14, 31, 39, 56, 69], obtaining more than $100\mathrm{k}$ training sample pairs with pseudolabels. We implement SuperEvent in PyTorch [4] and train it with the Adam [34] optimizer (learning rate $1e^{-4}$ , betas 0.9, 0.999) for 10 epochs. + +Table 2. Pose estimation on the Event Camera dataset. + +
MethodECD: Pose Estimation AUC in %
@5°@10°@20°
LLAK [10]0.71.42.1
RATE [33]3.38.418.0
EventPoint [32]1.63.05.4
SuperEvent (ours)22.735.846.7
+ +We employ the network architecture as in Section 3.3 with the loss function hyperparameters $c_{\lambda} = 10$ , $c_{\mathrm{d}} = 0.5$ , $c_{\mathrm{p}} = 1.0$ , and $c_{\mathrm{n}} = 0.2$ . We encode events as a 10-channel MCTS with 5 time window sizes $\Delta t$ logarithmically distributed from $0.001 \, \mathrm{s}$ to $0.1 \, \mathrm{s}$ . The detection threshold is set to $c_{\mathrm{s}} = 0.01$ and the NMS neighborhood $c_{\mathrm{b}} = 2$ pixels. + +We evaluate on a desktop PC with an Intel Core i5-13600 processor, an NVIDIA GeForce GTX 4070 GPU, and 32 GB RAM. Inference times are shown in Table 1. + +# 4.2. Pose Estimation + +A common benchmark for evaluating frame-based keypoint detectors is pose estimation [22, 59, 62, 64]. These works use datasets such as ScanNet [12] and MegaDepth [42] containing various images of the same scene with the associated ground truth camera poses. To create a similar event-based benchmark, we employ sequences with ground truth poses from real-world dynamic datasets. We search for uniformly distributed pose changes between $1^{\circ}$ and $45^{\circ}$ . The upper bound is chosen to increase the probability of visual overlap. However, since we do not know the ground truth geometry of the scene, it cannot be guaranteed that every sample contains sufficient information to recover the ground truth pose. Next, we process the event stream with various approaches to detect and match keypoints. After outlier removal with Random Sample Consensus (RANSAC) [16], we estimate the camera pose and calculate the error with respect to the ground truth. Finally, we report the area-under-curve (AUC) metric for different thresholds. + +We compare the pose estimation capability of SuperEvent to the 3 approaches LLAK [10], RATE [33], and EventPoint [32]. While we use the open-source implementations of LLAK and RATE, we re-implemented and trained EventPoint by carefully following the steps outlined in the paper. We were unable to evaluate the approach SD2Event [19] due to unpublished code and missing implementation details in the respective work. + +The experiment is conducted on two diverse datasets: Firstly, the Event Camera Dataset (ECD) [50] is widely used to evaluate event-based approaches, as it was one of the first publicly available datasets. It is challenging due to the low resolution $(180\times 240)$ , the noisy output of the initiation DAVIS240C event camera, and fast motion changes. Secondly, we use the more recent dataset Event-aided Di + +Table 3. Pose estimation on Event-aided Direct Sparse Odometry. + +
MethodEDS: Pose Estimation AUC in %
@5°@10°@20°
LLAK [10]0.50.71.0
RATE [33]2.15.110.3
EventPoint [32]1.62.85.2
SuperEvent (ours)15.226.440.1
+ +rect Sparse Odometry (EDS) [29]. In contrast to our training data, this dataset is recorded with a Prophesee Gen $3.1^{2}$ event camera with a higher resolution of $480 \times 640$ pixels and an over-proportionally increased event rate. Fast, inconsistent camera motions and a high dynamic range make pose estimation challenging. + +Table 2 shows that SuperEvent outperforms all baselines by a large margin. On the ECD dataset, its advantage is the greatest for the high-precision estimations below $5^{\circ}$ error, confirming the effectiveness of our method. In contrast, we found that LLAK detects many keypoints but cannot track the vast majority of them. EventPoint's matching capability is severely limited, possibly due to their static assumptions during training inherently contradicting the dynamic nature of event cameras. Finally, RATE achieves decent results for estimations with up to $20^{\circ}$ error, but since it cannot recover from tracking loss, it depends on consistent and precise tracking of most keypoints. As seen in the results, this requirement cannot be stably fulfilled. + +Table 3 indicates a slight decrease in SuperEvent's pose estimation performance on the EDS dataset compared to ECD. Since all approaches perform worse, we attribute this effect to the dataset being more challenging. Interestingly, the performance of SuperEvent for $20^{\circ}$ precision is still in the same region as for the ECD dataset. Considering the different properties of the Prophesee camera compared to the DAVIS camera employed in the training sequences, these results confirm the generalization capability of SuperEvent. + +# 4.3. Stereo Event-Visual Inertial SLAM + +We demonstrate the usefulness of SuperEvent with a downstream experiment. To date, frame-based visual odometry and SLAM systems largely outperform their event-based counterparts, attributed to their decades-long lead in development. By integrating SuperEvent into the frame-based VI-SLAM system OKVIS2 [40], we enable this framework to process the keypoints detected in the event stream. We compare the performance of the resulting event-only system to other recent event-based Odometry and SLAM approaches on the TUM-VIE dataset [35] in Table 4. Moreover, note that the next best results of the data-driven Mono E(I)O systems DEVO and DEIO are achieved only + +Table 4. Results on the small scale mocap sequences of TUM-VIE dataset, ATE in cm. The first three gray results require scale alignment with the ground truth; all other approaches estimate the absolute scale. Baseline numbers taken from [27, 53]. + +
MethodModality1d-trans3d-trans6dofdeskdesk2Average
EVO [55] w/ scale alignmentMono E7.5012.5085.5054.1075.2046.96
DEVO [36] w/ scale alignmentMono E0.501.101.601.701.001.18
DEIO [27] w/ scale alignmentMono E + IMU0.421.111.371.360.731.00
DEIO [27]Mono E + IMU1.081.121.391.411.191.24
ESVO [67]Stereo E12.5417.1913.4612.924.4212.11
ES-PTAM [21]Stereo E1.058.5310.252.507.205.91
ICRA'24 [54]Stereo E + IMU3.8518.90failed8.999.47-
ESVO2 [53]Stereo E + IMU3.337.263.216.164.024.78
OKVIS2 [40] + SuperEvent (ours)Stereo E + IMU0.440.890.430.580.410.55
(ours without loop closure)(0.43)(0.89)(0.43)(0.70)(0.40)(0.57)
+ +Table 5. Results on VECTor [18] large scale sequences. Our results are obtained with online calibration. Baseline numbers are taken from [27] and [36]. DEVO [36] cannot recover the absolute scale. Sequences marked $\cdot -\cdot$ are not evaluated in [27]. ATE in cm. + +
Methodcorr.-dollycorr.-walkunits-dollyunits-scooterschool-dollyschool-scooter
Ours33.13133.16122.6159.0569.9639.07
DEIO [27]492.65325.00826.38304.14--
ESVO [67]failedfailedfailedfailed1371.0983.0
+ +after aligning the scale of the estimated trajectories with ground truth. However, real-world applications, e.g., VR and robotics, typically require accurate metric scale. Results on the VECtor dataset [18] are listed in Table 5. + +OKVIS2's loop closure capability relies on long-term descriptor matching. While its effect on the small-scale sequences is limited, we evaluate OKVIS2 + SuperEvent on the loop-floor sequences of the TUM-VIE dataset, since these are explicitly designed to test a system's loop-closure capabilities. Thanks to SuperEvent, the OKVIS2 loop-closure feature works reliably with events, as shown in Table 6. Figure 5 qualitatively compares the estimated trajectories of the sequence loop-floor0. Without loop closure, there is significant drift - albeit not out of the ordinary compared to frame-based VI-SLAM. However, with loop closure enabled, the starting room is recognized and the drift is corrected. This demonstrates the robustness of SuperEvent, achieving successful descriptor matching even with minutes of time between observations. + +# 5. Conclusion + +We presented SuperEvent, a novel keypoint detector and descriptor for event streams. During its training, we consider the time-dependent keypoint appearance and employ pseudo-labels detected on frames. Combined with the flexible MCTS representation and a transformer back + +Table 6. Effect of loop closure of OKVIS2 [40] + SuperEvent on TUM-VIE loop-floor sequences, ATE in cm. + +
loop-floor0123
Estimated length349 m316 m279 m303 m
Loop closure4.964.648.924.74
W/o loop closure132.11161.92116.00129.17
+ +![](images/f97789cb0ed71692f3f0e708eee8dccc484b96bf1a853b387e5a0620b3898672.jpg) +Figure 5. OKVIS2 + SuperEvent's estimated trajectories of TUMVIE sequence loop-floor0 with and without loop closure. + +bone, it achieves SOTA performance on keypoint-based pose estimation. SuperEvent's integration into frame-based downstream applications enables event vision in existing keypoint-based systems. + +As future work, SuperEvent can be combined with a frame-based detector to realize systems that simultaneously process frames and events, exploiting their complementary advantages. Also, with a similar approach to SuperEvent, an event-based line detector could improve the performance of downstream applications in man-made environments. + +Finally, our training data is generated to enable consistent feature description independent of their motion-dependent appearance. However, since we employ real sequences, consecutive samples are biased to having similar motion directions. This is reflected in the performance of SuperEvent, whose descriptor matchability suffers under strong motion changes. While this is an inherent property of event cameras, a different approach to generating training data could mitigate it. + +# References + +[1] Ignacio Alzugaray and Margarita Chli. Asynchronous corner detection and tracking for event cameras in real time. IEEE Robotics and Automation Letters, 3(4):3177-3184, 2018. 3 +[2] Ignacio Alzugaray and Margarita Chli. Ace: An efficient asynchronous corner tracker for event cameras. In 2018 International Conference on 3D Vision (3DV), pages 653-661. IEEE, 2018. 4 +[3] Ignacio Alzugaray and Margarita Chli. HASTE: Multi-hypothesis asynchronous speeded-up tracking of events. In 31st British Machine Vision Virtual Conference (BMVC 2020), page 744. ETH Zurich, Institute of Robotics and Intelligent Systems, 2020. 4 +[4] Jason Ansel, Edward Yang, Horace He, Natalia Gimelshein, Animesh Jain, Michael Voznesensky, Bin Bao, Peter Bell, David Berard, Evgeni Burovski, Geeta Chauhan, Anjali Choudria, Will Constable, Alban Desmaison, Zachary DeVito, Elias Ellison, Will Feng, Jiong Gong, Michael Gschwind, Brian Hirsh, Sherlock Huang, Kshitteej Kalambarkar, Laurent Kirsch, Michael Lazos, Mario Lezcano, Yanbo Liang, Jason Liang, Yinghai Lu, CK Luk, Bert Maher, Yunjie Pan, Christian Puhrsch, Matthias Reso, Mark Saroufim, Marcos Yukio Siraichi, Helen Suk, Michael Suo, Phil Tillet, Eikan Wang, Xiaodong Wang, William Wen, Shunting Zhang, Xu Zhao, Keren Zhou, Richard Zou, Ajit Mathews, Gregory Chanan, Peng Wu, and Soumith Chintala. PyTorch 2: Faster Machine Learning Through Dynamic Python Bytecode Transformation and Graph Compilation. In 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2 (ASPLOS '24). ACM, 2024. 6 +[5] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-up robust features (SURF). Computer vision and image understanding, 110(3):346-359, 2008. 3 +[6] Michael Calonder, Vincent Lepetit, Christoph Strecha, and Pascal Fua. Brief: Binary robust independent elementary features. In Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part IV 11, pages 778-792. Springer, 2010. 3 +[7] Carlos Campos, Richard Elvira, Juan J. Gómez, José M. M. Montiel, and Juan D. Tardós. ORB-SLAM3: An accurate open-source library for visual, visual-inertial and multi-map SLAM. IEEE Transactions on Robotics, 37(6):1874-1890, 2021. 1, 4 +[8] Peiyu Chen, Weipeng Guan, and Peng Lu. ESVIO: Event-based stereo visual inertial odometry. IEEE Robotics and Automation Letters, 8(6):3661-3668, 2023. 4 +[9] Philippe Chiberre, Etienne Perot, Amos Sironi, and Vincent Lepetit. Detecting stable keypoints from events through image gradient prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1387-1394, 2021. 1, 3 +[10] Philippe Chiberre, Etienne Perot, Amos Sironi, and Vincent + +Lepetit. Long-lived accurate keypoints in event streams. arXiv preprint arXiv:2209.10385, 2022. 1, 3, 6, 7 +[11] Xavier Clady, Sio-Hoi Ieng, and Ryad Benosman. Asynchronous event-based corner detection and matching. Neural Networks, 66:91-106, 2015. 3 +[12] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828–5839, 2017. 7 +[13] Laurent Dardelet, Ryad Benosman, and Sio-Hoi Ieng. An event-by-event feature detection and tracking invariant to motion direction and velocity. Authorea Preprints, 2023. 4 +[14] Jeffrey Delmerico, Titus Cieslewski, Henri Rebecq, Matthias Faessler, and Davide Scaramuzza. Are we ready for autonomous drone racing? the uzh-fpv drone racing dataset. In 2019 International Conference on Robotics and Automation (ICRA), pages 6713-6719. IEEE, 2019. 6 +[15] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperPoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224-236, 2018. 1, 2, 3, 4, 5, 6 +[16] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 7 +[17] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 3 +[18] Ling Gao, Yuxuan Liang, Jiaqi Yang, Shaoxun Wu, Chenyu Wang, Jiaben Chen, and Laurent Kneip. Vector: A versatile event-centric benchmark for multi-sensor slam. IEEE Robotics and Automation Letters, 7(3):8217-8224, 2022. 3, 6, 8 +[19] Yuan Gao, Yuqing Zhu, Xinjun Li, Yimin Du, and Tianzhu Zhang. SD2Event: Self-supervised learning of dynamic detectors and contextual descriptors for event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3055-3064, 2024. 1, 3, 5, 7 +[20] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13884-13893, 2023. 5 +[21] Suman Ghosh, Valentina Cavinato, and Guillermo Gallego. ES-PTAM: Event-based stereo parallel tracking and mapping. In European Conference on Computer Vision (ECCV) Workshops, 2024. 4, 8 +[22] Pierre Gleize, Weiyao Wang, and Matt Feiszli. SiLK: Simple learned keypoints. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 22442-22451, 2023. 2, 3, 5, 7 +[23] Arren Glover, Aiko Dinale, Leandro De Souza Rosa, Simeon Bamford, and Chiara Bartolozzi. luvHarris: A practical corner detector for event-cameras. IEEE Transactions on + +Pattern Analysis and Machine Intelligence, 44(12):10087-10098, 2021. 3 +[24] Weipeng Guan and Peng Lu. Monocular event visual inertial odometry based on event-corner using sliding windows graph-based optimization. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2438-2445, 2022. 4 +[25] Weipeng Guan, Peiyu Chen, Yuhan Xie, and Peng Lu. PL- EVIO: Robust monocular event-based visual inertial odometry with point and line features. IEEE Transactions on Automation Science and Engineering, 2023. 4 +[26] Weipeng Guan, Peiyu Chen, Huibin Zhao, Yu Wang, and Peng Lu. EVI-SAM: Robust, real-time, tightly-coupled event-visual-inertial state estimation and 3d dense mapping. Advanced Intelligent Systems, 6(12):2400243, 2024. 4 +[27] Weipeng Guan, Fuling Lin, Peiyu Chen, and Peng Lu. DEIO: Deep event inertial odometry. arXiv preprint arXiv:2411.03928, 2024. 4, 8 +[28] Christopher G. Harris and M. J. Stephens. A combined corner and edge detector. In Alvey Vision Conference, 1988. 3 +[29] Javier Hidalgo-Carrió, Guillermo Gallego, and Davide Scaramuzza. Event-aided direct sparse odometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5781–5790, 2022. 7 +[30] Sumin Hu, Yeeun Kim, Hyungtae Lim, Alex Junho Lee, and Hyun Myung. eCDT: Event clustering for simultaneous feature detection and tracking. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3808-3815. IEEE, 2022. 4 +[31] Yuhuang Hu, Jonathan Binas, Daniel Neil, Shih-Chii Liu, and Tobi Delbruck. DDD20 end-to-end event camera driving dataset: Fusing frames and events with deep learning for improved steering prediction. In 2020 IEEE 23rd International Conference on Intelligent Transportation Systems (ITSC), pages 1-6. IEEE, 2020. 1, 6 +[32] Ze Huang, Li Sun, Cheng Zhao, Song Li, and Songzhi Su. EventPoint: Self-supervised interest point detection and description for event-based camera. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5396-5405, 2023. 1, 3, 5, 6, 7 +[33] Mikihiro Ikura, Cedric Le Gentil, Marcus G. Müller, Florian Schuler, Atsushi Yamashita, and Wolfgang Stürzl. RATE: Real-time asynchronous feature tracking with event cameras. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 11662-11669, 2024. 1, 4, 6, 7 +[34] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. 6 +[35] Simon Klenk, Jason Chui, Nikolaus Demmel, and Daniel Cremers. TUM-VIE: The TUM stereo visual-inertial event dataset. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8601-8608. IEEE, 2021. 3, 6, 7 +[36] Simon Klenk, Marvin Motzet, Lukas Koestler, and Daniel Cremers. Deep event visual odometry. In 2024 International Conference on 3D Vision (3DV), pages 739-749. IEEE, 2024. 4, 8 + +[37] Beat Kueng, Elias Mueggler, Guillermo Gallego, and Davide Scaramuzza. Low-latency visual odometry using event-based feature tracks. In 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 16-23. IEEE, 2016. 4 +[38] Xavier Lagorce, Garrick Orchard, Francesco Galluppi, Bertram E. Shi, and Ryad B. Benosman. HOTS: A hierarchy of event-based time-surfaces for pattern recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(7):1346-1359, 2017. 2, 3, 5 +[39] Alex Junho Lee, Younggun Cho, Young-sik Shin, Ayoung Kim, and Hyun Myung. ViViD++: Vision for visibility dataset. IEEE Robotics and Automation Letters, 7(3):6282-6289, 2022. 6 +[40] Stefan Leutenegger. OKVIS2: Realtime scalable visual-inertial SLAM with loop closure, 2022. 1, 2, 4, 6, 7, 8 +[41] Stefan Leutenegger, Margarita Chli, and Roland Y. Siegwart. BRISK: Binary robust invariant scalable keypoints. In 2011 International Conference on Computer Vision, pages 2548-2555, 2011. 3 +[42] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2041-2050, 2018. 7 +[43] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 5 +[44] Zhe Liu, Dianxi Shi, Ruihao Li, and Shaowu Yang. ESVIO: event-based stereo visual-inertial odometry. Sensors, 23(4): 1998, 2023. 4 +[45] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60:91-110, 2004. 3 +[46] Florian Mahlknecht, Daniel Gehrig, Jeremy Nash, Friedrich M Rockenbauer, Benjamin Morrell, Jeff Delaune, and Davide Scaramuzza. Exploring event camera-based odometry for planetary robots. IEEE Robotics and Automation Letters, 7(4):8651-8658, 2022. 4 +[47] Jacques Manderscheid, Amos Sironi, Nicolas Bourdis, Davide Migliore, and Vincent Lepetit. Speed invariant time surface for learning to detect corner points with event-based cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10245-10254, 2019. 1, 3 +[48] Nico Messikommer, Carter Fang, Mathias Gehrig, and Davide Scaramuzza. Data-driven feature tracking for event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5642-5651, 2023. 4, 5 +[49] Elias Mueggler, Chiara Bartolozzi, and Davide Scaramuzza. Fast event-based corner detection. Proceedings of the British Machine Vision Conference, 2017. 3 +[50] Elias Mueggler, Henri Rebecq, Guillermo Gallego, Tobi Delbruck, and Davide Scaramuzza. The event-camera dataset and simulator: Event-based data for pose estimation, visual + +odometry, and slam. The International Journal of Robotics Research, 36(2):142-149, 2017. 7 +[51] Alexander Neubeck and Luc Van Gool. Efficient nonmaximum suppression. In 18th international conference on pattern recognition (ICPR'06), pages 850-855. IEEE, 2006. 6 +[52] Zhenjiang Ni, Aude Bolopion, Joel Agnus, Ryad Benosman, and Stephane Regnier. Asynchronous event-based visual shape tracking for stable haptic feedback in microrobotics. IEEE Transactions on Robotics, 28(5):1081-1089, 2012. 4 +[53] Junkai Niu, Sheng Zhong, Xiuyuan Lu, Shaojie Shen, Guillermo Gallego, and Yi Zhou. ESVO2: Direct visual-inertial odometry with stereo event cameras. arXiv e-prints, pages arXiv-2410, 2024. 4, 8 +[54] Junkai Niu, Sheng Zhong, and Yi Zhou. Imu-aided event-based stereo visual odometry. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 11977–11983, 2024. 4, 8 +[55] Henri Rebecq, Timo Horstschäfer, Guillermo Gallego, and Davide Scaramuzza. EVO: A geometric approach to event-based 6-dof parallel tracking and mapping in real time. IEEE Robotics and Automation Letters, 2(2):593-600, 2016. 4, 8 +[56] Juan Pablo Rodríguez-Gómez, Raul Tapia, Julio L Paneque, Pedro Grau, Augusto Gómez Eguíluz, Jose Ramiro Martínez-de Dios, and Anibal Ollero. The GRIFFIN perception dataset: Bridging the gap between flapping-wing flight and robotic perception. IEEE Robotics and Automation Letters, 6(2):1066-1073, 2021. 6 +[57] Edward Rosten and Tom Drummond. Machine learning for high-speed corner detection. In Computer Vision-ECCV 2006: 9th European Conference on Computer Vision, Graz, Austria, May 7-13, 2006. Proceedings, Part I 9, pages 430-443. Springer, 2006. 3 +[58] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. ORB: An efficient alternative to SIFT or SURF. In 2011 International Conference on Computer Vision, pages 2564-2571, 2011. 3 +[59] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 1, 2, 3, 4, 7 +[60] Jianbo Shi and Tomasi. Good features to track. In 1994 Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 593-600, 1994. 3, 4 +[61] Karen Simonyan. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 3, 5 +[62] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 2, 3, 7 +[63] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. MaxViT: Multi-axis vision transformer. In European conference on computer vision, pages 459-479. Springer, 2022. 5 + +[64] Michal Tyszkiiewicz, Pascal Fua, and Eduard Trulls. DISK: Learning local features with policy gradient. Advances in Neural Information Processing Systems, 33:14254-14265, 2020. 2, 3, 5, 7 +[65] Valentina Vasco, Arren Glover, and Chiara Bartolozzi. Fast event-based harris corner detection exploiting the advantages of event-driven cameras. In 2016 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 4144-4149. IEEE, 2016. 3 +[66] Antoni Rosinol Vidal, Henri Rebecq, Timo Horstschaefer, and Davide Scaramuzza. Ultimate SLAM? Combining events, images, and IMU for robust visual slam in hdr and high-speed scenarios. IEEE Robotics and Automation Letters, 3(2):994-1001, 2018. 4 +[67] Yi Zhou, Guillermo Gallego, and Shaojie Shen. Event-based stereo visual odometry. IEEE Transactions on Robotics, 37 (5):1433-1450, 2021. 4, 8 +[68] Alex Zihao Zhu, Nikolay Atanasov, and Kostas Daniilidis. Event-based feature tracking with probabilistic data association. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 4465-4470. IEEE, 2017. 4 +[69] Alex Zihao Zhu, Dinesh Thakur, Tolga Özaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The multi-vehicle stereo event camera dataset: An event camera dataset for 3d perception. IEEE Robotics and Automation Letters, 3 (3):2032–2039, 2018. 6 +[70] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based optical flow using motion compensation. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, 2018. 5 \ No newline at end of file diff --git a/2025/SuperEvent_ Cross-Modal Learning of Event-based Keypoint Detection for SLAM/images.zip b/2025/SuperEvent_ Cross-Modal Learning of Event-based Keypoint Detection for SLAM/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..705bfbcb220bc0f9590966b65d5569a9dc0de020 --- /dev/null +++ b/2025/SuperEvent_ Cross-Modal Learning of Event-based Keypoint Detection for SLAM/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:805dc4e83c485da61cad09c2eeb5e85a1d17b3cd4beb9be4107a2097c9e8a8ae +size 618103 diff --git a/2025/SuperEvent_ Cross-Modal Learning of Event-based Keypoint Detection for SLAM/layout.json b/2025/SuperEvent_ Cross-Modal Learning of Event-based Keypoint Detection for SLAM/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fad67819eed9244592d0f903300b4d39979afa4e --- /dev/null +++ b/2025/SuperEvent_ Cross-Modal Learning of Event-based Keypoint Detection for SLAM/layout.json @@ -0,0 +1,9065 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 55, + 103, + 555, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 103, + 555, + 121 + ], + "spans": [ + { + "bbox": [ + 55, + 103, + 555, + 121 + ], + "type": "text", + "content": "SuperEvent: Cross-Modal Learning of Event-based Keypoint Detection for SLAM" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 121, + 142, + 233, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 142, + 233, + 156 + ], + "spans": [ + { + "bbox": [ + 121, + 142, + 233, + 156 + ], + "type": "text", + "content": "Yannick Burkhardt" + }, + { + "bbox": [ + 121, + 142, + 233, + 156 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 246, + 142, + 341, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 142, + 341, + 156 + ], + "spans": [ + { + "bbox": [ + 246, + 142, + 341, + 156 + ], + "type": "text", + "content": "Simon Schaefer" + }, + { + "bbox": [ + 246, + 142, + 341, + 156 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 366, + 142, + 479, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 142, + 479, + 157 + ], + "spans": [ + { + "bbox": [ + 366, + 142, + 479, + 157 + ], + "type": "text", + "content": "Stefan Leutenegger" + }, + { + "bbox": [ + 366, + 142, + 479, + 157 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 185, + 200, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 185, + 200, + 197 + ], + "spans": [ + { + "bbox": [ + 151, + 185, + 200, + 197 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 211, + 297, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 211, + 297, + 510 + ], + "spans": [ + { + "bbox": [ + 54, + 211, + 297, + 510 + ], + "type": "text", + "content": "Event-based keypoint detection and matching holds significant potential, enabling the integration of event sensors into highly optimized Visual SLAM systems developed for frame cameras over decades of research. Unfortunately, existing approaches struggle with the motion-dependent appearance of keypoints and the complex noise prevalent in event streams, resulting in severely limited feature matching capabilities and poor performance on downstream tasks. To mitigate this problem, we propose SuperEvent, a data-driven approach to predict stable keypoints with expressive descriptors. Due to the absence of event datasets with ground truth keypoints labels, we leverage existing frame-based keypoint detectors on readily available event-aligned and synchronized gray-scale frames for self-supervision: we generate temporally sparse keypoint pseudo-labels considering that events are a product of both scene appearance and camera motion. Combined with our novel, information-rich event representation, we enable SuperEvent to effectively learn robust keypoint detection and description in event streams. Finally, we demonstrate the usefulness of SuperEvent by its integration into a modern sparse keypoint and descriptor-based SLAM framework originally developed for traditional cameras, surpassing the state-of-the-art in event-based SLAM by a wide margin. Source code is available at ethz-mrl.github.io/SuperEvent." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 535, + 135, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 535, + 135, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 535, + 135, + 547 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 555, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 296, + 664 + ], + "type": "text", + "content": "Event cameras offer exciting advantages over their frame-based counterparts, such as an increased temporal resolution, little motion blur, and a high dynamic range. These properties promise to enhance robustness for robot estimation and perception tasks under fast motion and uncontrollable lighting conditions. However, processing the sparse and asynchronous output of event cameras requires fundamentally different algorithms than traditional frames. The decades-long research edge in frame-based vision resulted" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 184, + 554, + 274 + ], + "blocks": [ + { + "bbox": [ + 315, + 184, + 554, + 274 + ], + "lines": [ + { + "bbox": [ + 315, + 184, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 315, + 184, + 554, + 274 + ], + "type": "image", + "image_path": "45605cbf0dc83051639d4ccbd9c140befef04e3d84f828eeb40cc650d1134b5e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 315, + 276, + 554, + 367 + ], + "blocks": [ + { + "bbox": [ + 315, + 276, + 554, + 367 + ], + "lines": [ + { + "bbox": [ + 315, + 276, + 554, + 367 + ], + "spans": [ + { + "bbox": [ + 315, + 276, + 554, + 367 + ], + "type": "image", + "image_path": "1c2f2ba40ef732dfce855fd108e8099eb6d1e6e20f1c654471e2aeebdc7ae1e5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 375, + 555, + 431 + ], + "lines": [ + { + "bbox": [ + 313, + 375, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 555, + 431 + ], + "type": "text", + "content": "Figure 1. Detections on the sequence rec1499023756 of the DDD20 [31] dataset (not used for training). Top: Pseudolabels from SuperPoint [15] and SuperGlue [59] on the gray-scale frames. Bottom: Matched keypoints from SuperEvent in the event stream at the corresponding time stamps." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 449, + 555, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 449, + 555, + 569 + ], + "spans": [ + { + "bbox": [ + 312, + 449, + 555, + 569 + ], + "type": "text", + "content": "in most hardware and software being highly optimized for frame processing. A common approach for tasks like Structure-from-Motion, Visual Odometry (VO), Simultaneous Localization and Mapping (SLAM), or Place Recognition is to detect and match keypoints across frames to gain a geometric scene understanding [7, 40]. Consequently, a robust keypoint detector and descriptor working on the event stream would unlock the potential of integrating event cameras into extensive and highly-developed algorithms developed initially for frame-based vision." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 570, + 556, + 714 + ], + "type": "text", + "content": "While several authors already proposed event-based approaches to detect keypoints [9, 10, 19, 32, 33, 47], their matching abilities still show limitations. Most employ nearest-neighbor matching in pixel space, which restricts their usability to visually simple scenes. In an attempt to overcome this issue, more recent approaches learn descriptors in a data-driven fashion. To generate ground truth labels - inspired by frame-based methods [15] - they warp their event representations with a random homography and penalize non-consistent model predictions. This approach forces the models to learn descriptors to match visually similar keypoints. However, it fundamentally ignores the true" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 70, + 673, + 282, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 673, + 282, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 673, + 282, + 713 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 673, + 282, + 713 + ], + "type": "text", + "content": "Technical University of Munich, {yannick.burkhardt, simon.k.schaefer}@tum.de \n" + }, + { + "bbox": [ + 70, + 673, + 282, + 713 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 70, + 673, + 282, + 713 + ], + "type": "text", + "content": "ETH Zürich, lestefan@ethz.ch \n" + }, + { + "bbox": [ + 70, + 673, + 282, + 713 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 70, + 673, + 282, + 713 + ], + "type": "text", + "content": "Munich Center for Machine Learning (MCML)" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8918" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 555, + 274 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 555, + 274 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 555, + 274 + ], + "type": "image", + "image_path": "a8e02256b314e4bc787d1586037a71542fa42ba50c41215b629540d8dd726854.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 282, + 555, + 327 + ], + "lines": [ + { + "bbox": [ + 54, + 282, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 54, + 282, + 555, + 327 + ], + "type": "text", + "content": "Figure 2. Data processing pipeline of SuperEvent: For training, pseudo-labels are generated using a frame-based detector and matcher on the gray-scale frame pairs. We generate spatially synchronized MCTS at the same timestamps and feed them to SuperEvent. The network predictions are compared to the pseudo-labels and the network weights are optimized using backpropagation. During inference, the network predictions can be used to detect keypoints and match their descriptors on the event stream only." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 342, + 294, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 342, + 294, + 413 + ], + "spans": [ + { + "bbox": [ + 54, + 342, + 294, + 413 + ], + "type": "text", + "content": "nature of the event generation process that conflates camera motion with scene appearance and geometry. In essence, matches are generated only from one original set of events per keypoint, thereby producing only limited and potentially also unrealistic training data. We hypothesize that this simplification ultimately results in unstable matching." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 414, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 414, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 54, + 414, + 295, + 521 + ], + "type": "text", + "content": "In contrast, we propose training our model to predict descriptors for the keypoint correspondences at different points in time – therefore training with different and actually recorded rather than warped events. Since generating keypoint and match labels manually is intractable, we design a scalable method to exploit the great capabilities of frame-based vision models: generating pseudo-labels in aligned frames and then using them to train our model's predictions. An example is shown in Figure 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 521, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 521, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 54, + 521, + 295, + 653 + ], + "type": "text", + "content": "To provide the model with rich input data, we propose the novel event representation Multi-Channel Time Surfaces (MCTS) which generalizes the commonly used time surface representation [38] to an " + }, + { + "bbox": [ + 54, + 521, + 295, + 653 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 54, + 521, + 295, + 653 + ], + "type": "text", + "content": "-channel tensor of time surfaces for both polarities and different time window sizes. Since an optimal window size depends inversely on the motion magnitude, this approach further increases the model's robustness to both slow and fast motion. Additionally, we revisit network architecture choices of frame- and event-based vision to optimize our model's performance. Figure 2 shows the complete data processing pipeline." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 654, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 654, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 654, + 295, + 713 + ], + "type": "text", + "content": "In absence of evaluation procedures to test event-based keypoint and descriptor stability, we develop a benchmark to assess the matched keypoint's quality for pose estimation, similar to benchmarks commonly used in frame-based vision [22, 59, 62, 64]. On two datasets recorded with dif" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 342, + 553, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 553, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 553, + 401 + ], + "type": "text", + "content": "ferent sensors, settings, resolutions, and scenes, we outperform other data-driven and handcrafted approaches in terms of pose estimation by a large margin, demonstrating the robustness of our approach. We carry out ablation studies to confirm the effectiveness of our design choices." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 402, + 554, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 554, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 554, + 496 + ], + "type": "text", + "content": "Finally, we integrate our model into a state-of-the-art (SOTA) stereo visual inertial SLAM (VI-SLAM) framework [40], replacing the frame-based keypoint detector and descriptor extractor with SuperEvent. Our results surpass the current SOTA on event-based Visual Odometry and SLAM. This demonstrates the great potential of an event-sensor integration into existing, well-developed algorithms for frame-based vision." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 498, + 554, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 569 + ], + "type": "text", + "content": "We name our method SuperEvent in honor of the great influence of the works SuperPoint [15] and SuperGlue [59] on the research field of keypoint detection as well as on our model directly: ultimately, their frame-based detections and matches are used as pseudo-labels to teach keypoint detection and description to SuperEvent." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": "Contributions: (1) We propose a scalable method to generate training data for interest point detection and matching, leveraging robust existing frame-based keypoint detectors. The resulting data incorporates the real temporal dependence of event generation using real event camera recordings. This enables data-driven models to effectively learn robust keypoint detection and description. (2) We develop the model for event-based vision through key design choices: Our novel MCTS event representation increases the performance of data-driven models by reducing their dependence on motion speed of the scene. A transformer backbone shows to further improve the model's feature ex" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8919" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": "traction. In our conducted experiments, our approach outperforms SOTA detection and matching by a large margin. We publicly release our code to support future event-based vision research. (3) Showcasing the practicality of our method, we integrate the trained model into an existing stereo VI-SLAM approach developed for frame-based vision. The resulting event inertial SLAM system outperforms SOTA approaches on the commonly used TUMVIE [35] and VECtor [18] datasets by a large margin." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 190, + 142, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 142, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 142, + 204 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 211, + 295, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 211, + 295, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 295, + 330 + ], + "type": "text", + "content": "We begin with a short summary of keypoint detectors for frame-based vision since our work employs some of these findings for model design and training data generation. We then group the event-based works into handcrafted and data-driven keypoint detection, as well as tracking methods. Since our focus in this work lies on keypoint detection and matching based on event data exclusively, we do not consider approaches that rely on combination with other sensors, e.g., frame cameras. For an extensive overview of event-based vision, we refer the interested reader to [17]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 331, + 295, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 295, + 534 + ], + "type": "text", + "content": "Frame-based keypoint detection: there exists a vast variety of approaches to detecting and describing keypoints in frames. Traditional, handcrafted methods [5, 6, 28, 41, 45, 58, 60] are still widely used, since they are well studied and usually work robustly with little overhead. More recent, data-driven methods outperform the traditional keypoint detectors with impressive results. SuperPoint [15] is the first method trained with homographic adaptation, a self-supervised training technique of warping input images and their detected labels by randomized homographies. SuperPoint's architecture consists of a shared VGG [61]-backbone with a detector and descriptor head, predicting descriptors in a discrete grid which is interpolated to full resolution. The authors later publish the extension SuperGlue [59], a graph neural network that improves the descriptor matching. Following works propose different architectures trained with homographic adaptation [22, 62, 64]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 534, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 641 + ], + "type": "text", + "content": "Event-based keypoint detection: handcrafted approaches can be broadly divided into filters and frame reconstruction. Filtering methods work directly on the event stream, attempting to find events originating from corners [1, 11]. While their asynchronous nature offers little processing overhead, downstream tasks are required to process asynchronous data. Existing well-developed downstream solutions for frame-based vision therefore require major modifications." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "text", + "content": "Frame reconstruction methods collect events to construct a frame representation, such as binary event frames or time surfaces [38]. Adopting methods from frame-based vision, such as the Harris corner detector [28] in [23, 65] or the FAST detector [57] in [49], these approaches detect corners in these event frames. However, they suffer from the in" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "herent differences between reconstructed event frames and traditional frames, resulting in limited robustness of the algorithms developed for frame-based vision." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 109, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 553, + 228 + ], + "type": "text", + "content": "Additionally, these handcrafted methods usually employ simple nearest-neighbor matching in pixel space to track keypoints over time. This approach is only reliable for simple scenes since cluttered geometry, high variance in depth, and fast motions cause keypoints to become occluded, tightly clustered, or to jump. Also, the developed heuristics struggle with the ubiquitous and complex noise inherent to event cameras. Usually, they require manual parameter tuning for different camera models, datasets, or sometimes even scenes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 231, + 553, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 553, + 374 + ], + "type": "text", + "content": "Event-based keypoint detection: data-driven approaches promise to cope better with the event data and its noise. A series of works [9, 10, 47] train data-driven corner detectors using labels by detecting keypoints on corresponding gray-scale frames that are synthetic or from an HVGA ATIS sensor. While these approaches learn data-driven detection, they only track keypoints by nearest-neighbor matching in pixel space, yielding the aforementioned drawback of limited robustness in complex scenes. Additionally, the employed corner detector for ground truth generation as well as the machine-learning models have limited performance due to their simplicity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 376, + 553, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 553, + 591 + ], + "type": "text", + "content": "For these reasons, EventPoint [32] and SD2Event [19] employ more complex neural networks that additionally predict descriptors. While EventPoint uses the same architecture as SuperPoint [15] and finetunes its pre-trained weights, SD2Event employs agent-based attention to learn detection and description. Both approaches are trained with homographic adaptation and varying time window sizes. While this results in rotation-, distortion-, and scale-invariant models, the main challenge of keypoint description remains largely unsolved: Because most events result from scene or camera motion, corresponding keypoints vary in visual appearance. Since robust keypoint matching requires similar descriptor pairs, this motion dependence must be considered. However, homographic adaptation creates static frame pairs with little to no variance in scene motion within the samples. This contradicts the dynamic and motion-dependent nature of events, resulting in unstable keypoint matching between different timestamps." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 592, + 553, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 553, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 553, + 675 + ], + "type": "text", + "content": "Additionally, the authors of EventPoint propose the event representation Tencode, extending time surfaces with two additional channels to encode the events' polarities. However, to match SuperPoint's architecture with single-channel input, the Tencode tensor is then converted to one channel, disregarding the previously introduced separation of polarities." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Our method enforces robust descriptor matching by employing training data containing pseudo-labels for keypoints and descriptors under varying motion. We avoid" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8920" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "synthetic event data since current simulators struggle to model realistic noise. Furthermore, our MCTS representation strictly separates the polarities and provides a range of time window sizes, providing additional information and omitting the time window parameter choice." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 134, + 295, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 134, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 295, + 266 + ], + "type": "text", + "content": "Event-based keypoint tracking: while early works fall back on batching events and processing events in a discretized way [37, 52, 68], a promising idea is to exploit their spatial continuous visual flow [2, 13, 30]. Among these approaches is HASTE [3] which calculates the most likely motion hypothesis from a fixed set of translations and rotations for every incoming event. Since HASTE relies on external keypoint initialization, RATE [33] employs a Shi-Tomasi [60] corner detector on the binary event frame, initializing several HASTE instances to track multiple keypoints in parallel." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 266, + 295, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 295, + 386 + ], + "type": "text", + "content": "A major drawback of the described tracking approaches is their computational overhead when processing each event individually. Also, most methods only track one keypoint, resulting in an almost linear growth of computational cost with increasing number of tracker instances. E.g., RATE tracks only 20 features in real-time for an event stream of " + }, + { + "bbox": [ + 55, + 266, + 295, + 386 + ], + "type": "inline_equation", + "content": "180 \\times 40" + }, + { + "bbox": [ + 55, + 266, + 295, + 386 + ], + "type": "text", + "content": " pixels resolution – which is hardly sufficient for downstream applications such as SLAM. Additionally, these handcrafted approaches struggle with environmental and hardware changes, as well as the event cameras' noise." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 388, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 295, + 483 + ], + "type": "text", + "content": "Our method SuperEvent works reliably in real-time with adjustable frequency and without re-training or excessive parameter tuning for different camera models. It's dense keypoint heatmap predictions barely result in overhead for more detections. In contrast to the tracking approaches, SuperEvent can be directly integrated into downstream algorithms relying on synchronous output and descriptors, e.g., for SLAM with place recognition, as we demonstrate." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 484, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 295, + 592 + ], + "type": "text", + "content": "Event-based Odometry: existing Event Odometry (EO) approaches are developed specifically for event processing. While some approaches combine events with frames [8, 25, 26, 46, 66], event-only approaches can be classified as monocular EO [36, 55], monocular EO with IMU [24, 27], stereo EO [21, 67], and stereo EO with IMU [44, 53, 54]. Due to the short history of event cameras, these systems require extensive research and development efforts to work reliably in practice." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "type": "text", + "content": "In contrast, SuperEvent can be integrated into existing frame-based VO and SLAM systems. We thereby achieve event-based VI-SLAM leveraging highly accurate frame-based systems, emerging from years of research [7, 40]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 656, + 111, + 668 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 656, + 111, + 668 + ], + "spans": [ + { + "bbox": [ + 55, + 656, + 111, + 668 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "In this section, we describe how we train SuperEvent, introduce the advantages and construction of the MCTS event representation, and detail the network design choices." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 72, + 459, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 459, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 459, + 84 + ], + "type": "text", + "content": "3.1. Training Data Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 89, + 553, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 553, + 137 + ], + "type": "text", + "content": "To train SuperEvent, we exclusively employ real event camera recordings. Thus, our model learns to cope with the complex event stream properties, e.g., artifacts and noise, which are difficult to simulate realistically." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 137, + 553, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 137, + 553, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 553, + 269 + ], + "type": "text", + "content": "However, manually annotating keypoints in the asynchronous event data stream is very time-consuming and does not scale. Therefore, we propose an approach to generate ground truth data automatically. We take advantage of the pixel-synchronized events and frames of iniVation DAVIS1 event cameras. Employing a highly robust frame-based keypoint detector and matcher such as SuperPoint [15] + SuperGlue [59] on the gray-scale frames, we can obtain temporally sparse pseudo-labels. To generate a diverse set of training data with reliable pseudo-labels, we process every frame " + }, + { + "bbox": [ + 313, + 137, + 553, + 269 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 137, + 553, + 269 + ], + "type": "text", + "content": " of a sequence in the following way:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 269, + 553, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 269, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 314, + 269, + 553, + 340 + ], + "type": "text", + "content": "- Firstly, we compute the distances in pixel-space of all key points to their respective matches in frame " + }, + { + "bbox": [ + 314, + 269, + 553, + 340 + ], + "type": "inline_equation", + "content": "i + 1" + }, + { + "bbox": [ + 314, + 269, + 553, + 340 + ], + "type": "text", + "content": ". We only consider frames " + }, + { + "bbox": [ + 314, + 269, + 553, + 340 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 314, + 269, + 553, + 340 + ], + "type": "text", + "content": " for training where the distances' median surpasses a threshold " + }, + { + "bbox": [ + 314, + 269, + 553, + 340 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{m}}" + }, + { + "bbox": [ + 314, + 269, + 553, + 340 + ], + "type": "text", + "content": ". A small median distance indicates that the scene is nearly static, resulting in insufficient event data for a useful prediction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 340, + 553, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 340, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 314, + 340, + 553, + 399 + ], + "type": "text", + "content": "- To expose the model to diverse camera pose changes and motion directions, we generate pseudo-labels by matching the descriptors with a subsequent frame " + }, + { + "bbox": [ + 314, + 340, + 553, + 399 + ], + "type": "inline_equation", + "content": "i + j" + }, + { + "bbox": [ + 314, + 340, + 553, + 399 + ], + "type": "text", + "content": ". We increase " + }, + { + "bbox": [ + 314, + 340, + 553, + 399 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 314, + 340, + 553, + 399 + ], + "type": "text", + "content": " recursively, resulting in multiple pseudo-labels with increasing time interval between the frames:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 377, + 406, + 553, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 406, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 377, + 406, + 553, + 418 + ], + "type": "interline_equation", + "content": "j \\leftarrow j + j _ {\\mathrm {s}}, j _ {\\mathrm {s}} \\sim \\mathcal {U} (1, j _ {\\max }, \\tag {1}", + "image_path": "662a4beb480a542ca78e092deeba29460683835aadda81c8a627232163551791.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 423, + 553, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 423, + 553, + 494 + ], + "spans": [ + { + "bbox": [ + 321, + 423, + 553, + 494 + ], + "type": "text", + "content": "with the increment " + }, + { + "bbox": [ + 321, + 423, + 553, + 494 + ], + "type": "inline_equation", + "content": "j_{\\mathrm{s}}" + }, + { + "bbox": [ + 321, + 423, + 553, + 494 + ], + "type": "text", + "content": " drawn from the uniform distribution " + }, + { + "bbox": [ + 321, + 423, + 553, + 494 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(1, j_{\\max})" + }, + { + "bbox": [ + 321, + 423, + 553, + 494 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 321, + 423, + 553, + 494 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 321, + 423, + 553, + 494 + ], + "type": "text", + "content": " is initialized with 0 before its first update. The maximum step distance " + }, + { + "bbox": [ + 321, + 423, + 553, + 494 + ], + "type": "inline_equation", + "content": "j_{\\max}" + }, + { + "bbox": [ + 321, + 423, + 553, + 494 + ], + "type": "text", + "content": " is manually adjusted depending on the properties and texture of the sequence, e.g., for sequences with fast motion, " + }, + { + "bbox": [ + 321, + 423, + 553, + 494 + ], + "type": "inline_equation", + "content": "j_{\\max}" + }, + { + "bbox": [ + 321, + 423, + 553, + 494 + ], + "type": "text", + "content": " is decreased to yield sufficient matches." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 495, + 553, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 495, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 314, + 495, + 553, + 553 + ], + "type": "text", + "content": "- If the number of detected matches falls below the threshold " + }, + { + "bbox": [ + 314, + 495, + 553, + 553 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{m}}" + }, + { + "bbox": [ + 314, + 495, + 553, + 553 + ], + "type": "text", + "content": ", these pseudo-labels are discarded, and the next frame " + }, + { + "bbox": [ + 314, + 495, + 553, + 553 + ], + "type": "inline_equation", + "content": "i + 1" + }, + { + "bbox": [ + 314, + 495, + 553, + 553 + ], + "type": "text", + "content": " becomes the new reference frame. This procedure ensures sufficient visual overlap and enough texture in the two frames." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 555, + 553, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 553, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 553, + 625 + ], + "type": "text", + "content": "Frames with under- or overexposure, or strong motion blur harm the prediction quality of the frame-based model. While these frames are usually filtered out since there are few matches (if " + }, + { + "bbox": [ + 313, + 555, + 553, + 625 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{d}}" + }, + { + "bbox": [ + 313, + 555, + 553, + 625 + ], + "type": "text", + "content": " is chosen adequately), we also manually remove scenes that mostly contain too dark, too bright, or blurred frames." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 632, + 469, + 643 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 632, + 469, + 643 + ], + "spans": [ + { + "bbox": [ + 314, + 632, + 469, + 643 + ], + "type": "text", + "content": "3.2. Multi-channel Time Surface" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 649, + 553, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 649, + 553, + 697 + ], + "spans": [ + { + "bbox": [ + 313, + 649, + 553, + 697 + ], + "type": "text", + "content": "To allow for efficient batch processing of event data, the asynchronous stream is converted to a tensor with fixed dimensions. In the literature, various variants have been proposed to encode the event stream. An ablation study in [48]" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 325, + 702, + 400, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 400, + 712 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 400, + 712 + ], + "type": "text", + "content": "1https://inivation.com/" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8921" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 298, + 310 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 298, + 310 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 298, + 310 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 298, + 310 + ], + "type": "image", + "image_path": "8cda978ae46b17b8ade1a00f9e2ca10b8fb7b681e0a1cf3c83375af6e8cc9a0d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 316, + 292, + 328 + ], + "lines": [ + { + "bbox": [ + 59, + 316, + 292, + 328 + ], + "spans": [ + { + "bbox": [ + 59, + 316, + 292, + 328 + ], + "type": "text", + "content": "Figure 3. Schematic depiction of the MCTS generation process." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 341, + 296, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 341, + 296, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 296, + 424 + ], + "type": "text", + "content": "compares some of the most commonly used event representations: binary event frames, time surfaces [38], and voxel grids [70]. For their data-driven feature tracking approach, time surfaces yield the best model performance. The time surface variant Tencode with two more channels encoding the event polarity is proposed in [32]. It is used as model input after conversion to a single-channel frame." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 426, + 296, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 509 + ], + "type": "text", + "content": "Time surfaces and its variant Tencode have proven to be an efficient input representation for keypoint detection and tracking algorithms [19, 32, 48], since they assign the highest values to the pixel locations of the most recent events that are usually caused by visual edges. This facilitates finding keypoints along these edges. However, these encodings have shortcomings:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 510, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 582 + ], + "type": "text", + "content": "- Time surfaces are one-channel frames that ignore the event polarity and thereby discard potentially useful information. The Tencode representation is an attempt to overcome this issue. However, since it is converted back to a single channel tensor, there is still no clear separation between positive and negative events." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "- Time surfaces and the Tencode representation depend on the time window length " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": ". Increasing this parameter not only results in more events being considered but also decreases the contrast between the most recent events and past events. For high optical flow, this negatively impacts the detection of edges. Conversely, a smaller time window increases this contrast but includes fewer events, which can fail to sufficiently capture edges in areas with little optical flow. Choosing a suitable " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " is challenging, as it depends on the camera and scene motion which might be unknown a priori. The bottom part of Figure 3" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 96 + ], + "type": "text", + "content": "shows the effect of varying " + }, + { + "bbox": [ + 322, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 322, + 72, + 553, + 96 + ], + "type": "text", + "content": ": Only the time surfaces in the middle clearly display the net on the top right." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 96, + 555, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 555, + 216 + ], + "type": "text", + "content": "Our proposed MCTS representation mitigates these problems: It combines time surfaces for each polarity and for various time windows of temporal size " + }, + { + "bbox": [ + 313, + 96, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\Delta t_{n}" + }, + { + "bbox": [ + 313, + 96, + 555, + 216 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 96, + 555, + 216 + ], + "type": "inline_equation", + "content": "n\\in \\{1,\\ldots ,N\\}" + }, + { + "bbox": [ + 313, + 96, + 555, + 216 + ], + "type": "text", + "content": ". We distribute the time window sizes logarithmically to trade off data preprocessing and coverage of a wide variety of visual motion speeds. Events are encoded by their timestamp " + }, + { + "bbox": [ + 313, + 96, + 555, + 216 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 96, + 555, + 216 + ], + "type": "text", + "content": ", their pixel coordinates " + }, + { + "bbox": [ + 313, + 96, + 555, + 216 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 96, + 555, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 96, + 555, + 216 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 96, + 555, + 216 + ], + "type": "text", + "content": ", and their polarity " + }, + { + "bbox": [ + 313, + 96, + 555, + 216 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 96, + 555, + 216 + ], + "type": "text", + "content": ". Mathematically, the MCTS tensor at time " + }, + { + "bbox": [ + 313, + 96, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 96, + 555, + 216 + ], + "type": "text", + "content": " considering all " + }, + { + "bbox": [ + 313, + 96, + 555, + 216 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 96, + 555, + 216 + ], + "type": "text", + "content": " events " + }, + { + "bbox": [ + 313, + 96, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_i = (t_i,x_i,y_i,p_i)" + }, + { + "bbox": [ + 313, + 96, + 555, + 216 + ], + "type": "text", + "content": " between time " + }, + { + "bbox": [ + 313, + 96, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\tau -\\Delta t_{n}" + }, + { + "bbox": [ + 313, + 96, + 555, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 96, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 96, + 555, + 216 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 323, + 228, + 553, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 228, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 323, + 228, + 553, + 258 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {M C T S} = \\left(\\mathbf {T S} _ {- 1, \\Delta t _ {1}}, \\dots , \\mathbf {T S} _ {- 1, \\Delta t _ {N}}\\right), \\\\ \\left. \\mathbf {T S} _ {+ 1, \\Delta t _ {1}}, \\dots , \\mathbf {T S} _ {+ 1, \\Delta t _ {N}}\\right), \\tag {2} \\\\ \\end{array}", + "image_path": "4d1ad4a76990b4a5136f1ab3756a7b24bed859d3e52cbb67c3bbfde5ed8fa12d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 266, + 553, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 553, + 291 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 553, + 291 + ], + "type": "text", + "content": "with each time surface " + }, + { + "bbox": [ + 313, + 266, + 553, + 291 + ], + "type": "inline_equation", + "content": "\\mathbf{TS}" + }, + { + "bbox": [ + 313, + 266, + 553, + 291 + ], + "type": "text", + "content": ", initialized with zeros, defined by polarity " + }, + { + "bbox": [ + 313, + 266, + 553, + 291 + ], + "type": "inline_equation", + "content": "p\\in \\{-1, + 1\\}" + }, + { + "bbox": [ + 313, + 266, + 553, + 291 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 266, + 553, + 291 + ], + "type": "inline_equation", + "content": "\\Delta t_{n}" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 349, + 299, + 553, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 299, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 349, + 299, + 553, + 326 + ], + "type": "interline_equation", + "content": "\\mathbf {T S} _ {p, \\Delta t _ {n}} \\left(x _ {i}, y _ {i}\\right) = \\max _ {i, p _ {i} = p} \\left(1 - \\frac {\\tau - t _ {i}}{\\Delta t _ {n}}\\right). \\tag {3}", + "image_path": "9375414b3ed1b777aa1e39593a32251817003b4f7a91c6c0d6a69f177cef0c87.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 333, + 539, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 539, + 345 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 539, + 345 + ], + "type": "text", + "content": "The MCTS generation process is visualized in Figure 3." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 353, + 476, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 476, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 476, + 365 + ], + "type": "text", + "content": "3.3. Network Architecture & Loss" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 370, + 555, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 555, + 502 + ], + "type": "text", + "content": "As commonly employed in frame-based keypoint detection models [15, 22, 64], we combine a shared backbone with a detector and a descriptor head. Starting with the basic SuperPoint [15] architecture with adjusted loss functions for our temporal matching with pseudo-labels, we train various models with different backbones and hyperparameters and compare their performance. As an alternative approach, we replace the grid-based descriptor head with full resolution prediction and the respective loss functions as in [22, 64]. Please refer to the full results of this ablation study in the Supplementary Material." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 502, + 554, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 554, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 554, + 597 + ], + "type": "text", + "content": "Our resulting architecture consists of a 3-layer MaxViT [63] backbone with a Feature Pyramid Network (FPN) [43] which is also employed in a highly efficient, event-based detector [20]. Combined with grid-based VGG [61] detector and descriptor heads, SuperEvents fully convolutional components can process various input resolutions without retraining. Figure 4 shows the final network architecture." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 597, + 554, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 554, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 554, + 645 + ], + "type": "text", + "content": "The loss function evaluates the network prediction for two corresponding training tensors " + }, + { + "bbox": [ + 313, + 597, + 554, + 645 + ], + "type": "inline_equation", + "content": "\\{0,1\\}" + }, + { + "bbox": [ + 313, + 597, + 554, + 645 + ], + "type": "text", + "content": ". It combines the individual detector losses " + }, + { + "bbox": [ + 313, + 597, + 554, + 645 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{s},0}" + }, + { + "bbox": [ + 313, + 597, + 554, + 645 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 597, + 554, + 645 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{s},1}" + }, + { + "bbox": [ + 313, + 597, + 554, + 645 + ], + "type": "text", + "content": " with the joint descriptor loss " + }, + { + "bbox": [ + 313, + 597, + 554, + 645 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{d}}" + }, + { + "bbox": [ + 313, + 597, + 554, + 645 + ], + "type": "text", + "content": " weighted with the constant " + }, + { + "bbox": [ + 313, + 597, + 554, + 645 + ], + "type": "inline_equation", + "content": "c_{\\lambda}" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 382, + 655, + 553, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 655, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 382, + 655, + 553, + 668 + ], + "type": "interline_equation", + "content": "L = L _ {\\mathrm {s}, 0} + L _ {\\mathrm {s}, 1} + c _ {\\lambda} L _ {\\mathrm {d}}. \\tag {4}", + "image_path": "85a1f3558b5d8028cd49d882a7d5da740da4960c39bfbc5985dc6c8d93f0319b.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "Following [15], the detection is implemented as a classification problem of the pixels in image patches, i.e., the pixel indices represent the \"classes\". The detector loss " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{CE}}" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8922" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": "is thus defined as cross-entropy between a predicted patch of keypoint scores " + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{s_h,w}" + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": " (dimension " + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "inline_equation", + "content": "8\\times 8" + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": " plus no-keypoint-dustbin, thus 65) per grid cell (indices " + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "inline_equation", + "content": "h,w" + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": ") and the pseudolabels " + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_{s_h,w}" + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": " in one-hot encoding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 99, + 126, + 295, + 166 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 126, + 295, + 166 + ], + "spans": [ + { + "bbox": [ + 99, + 126, + 295, + 166 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {s}} = \\frac {1}{H _ {\\mathrm {c}} W _ {\\mathrm {c}}} \\sum_ {\\substack {h = 1 \\\\ w = 1}} ^ {H _ {\\mathrm {c}}, W _ {\\mathrm {c}}} L _ {\\mathrm {CE}} \\left(\\mathbf {y} _ {\\mathrm {s} _ {h}, w}, \\hat {\\mathbf {y}} _ {\\mathrm {s} _ {h}, w}\\right), \\tag{5}", + "image_path": "c66a9391599e7dfcdf9e00cbacff8c12631b605e4c10211d3fa30cf729257cf1.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 170, + 295, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 170, + 295, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 170, + 295, + 194 + ], + "type": "text", + "content": "for " + }, + { + "bbox": [ + 55, + 170, + 295, + 194 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{c}} \\times W_{\\mathrm{c}}" + }, + { + "bbox": [ + 55, + 170, + 295, + 194 + ], + "type": "text", + "content": " grid cells. In the case of multiple keypoint labels in a grid cell, one is randomly selected." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 194, + 296, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 296, + 388 + ], + "type": "text", + "content": "Unlike other models trained with homographic adaptation, our sparse pseudo-labels do not allow for dense descriptor training for every pixel or grid cell. As in [15], we only consider a single descriptor (256 channels) per grid cell, " + }, + { + "bbox": [ + 55, + 194, + 296, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{\\mathrm{d}_{h_0},w_0}" + }, + { + "bbox": [ + 55, + 194, + 296, + 388 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 194, + 296, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{\\mathrm{d}_{h_1},w_1}" + }, + { + "bbox": [ + 55, + 194, + 296, + 388 + ], + "type": "text", + "content": " (normalized to length one across the channels) with " + }, + { + "bbox": [ + 55, + 194, + 296, + 388 + ], + "type": "inline_equation", + "content": "(h_0,w_0)" + }, + { + "bbox": [ + 55, + 194, + 296, + 388 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 194, + 296, + 388 + ], + "type": "inline_equation", + "content": "(h_1,w_1)" + }, + { + "bbox": [ + 55, + 194, + 296, + 388 + ], + "type": "text", + "content": " denoting the grid indices of tensor 0 and 1. Successful keypoint matching requires similar descriptors for correspondences and distinct descriptors for unrelated keypoints. We measure descriptor similarity with the dot product " + }, + { + "bbox": [ + 55, + 194, + 296, + 388 + ], + "type": "inline_equation", + "content": "d_{h_1,w_1}^{h_0,w_0} = \\mathbf{y}_{\\mathrm{d}_{h_0,w_0}}\\cdot \\mathbf{y}_{\\mathrm{d}_{h_1,w_1}}" + }, + { + "bbox": [ + 55, + 194, + 296, + 388 + ], + "type": "text", + "content": ". Our pseudo label " + }, + { + "bbox": [ + 55, + 194, + 296, + 388 + ], + "type": "inline_equation", + "content": "\\hat{y}_{h_1,w_1}^{h_0,w_0} = 1" + }, + { + "bbox": [ + 55, + 194, + 296, + 388 + ], + "type": "text", + "content": " indicates that two descriptors at the respective grid cell locations correspond, otherwise " + }, + { + "bbox": [ + 55, + 194, + 296, + 388 + ], + "type": "inline_equation", + "content": "\\hat{y}_{h_1,w_1}^{h_0,w_0} = 0" + }, + { + "bbox": [ + 55, + 194, + 296, + 388 + ], + "type": "text", + "content": ". Since the majority of the descriptors do not correspond, the partial loss of descriptor matches is weighted with the constant " + }, + { + "bbox": [ + 55, + 194, + 296, + 388 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{d}}" + }, + { + "bbox": [ + 55, + 194, + 296, + 388 + ], + "type": "text", + "content": ". Mathematically, we formulate the descriptor loss as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 393, + 293, + 443 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 393, + 293, + 443 + ], + "spans": [ + { + "bbox": [ + 59, + 393, + 293, + 443 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {d}} = \\sum_ {\\substack {h _ {0} = 1 \\\\ w _ {0} = 1}} ^ {H _ {\\mathrm {c}}, W _ {\\mathrm {c}}} \\sum_ {\\substack {h _ {1} = 1 \\\\ w _ {1} = 1}} ^ {H _ {\\mathrm {c}}, W _ {\\mathrm {c}}} \\left\\{ \\begin{array}{l l} c _ {\\mathrm {d}} \\max (0, c _ {\\mathrm {p}} - d _ {h _ {1}, w _ {1}} ^ {h _ {0}, w _ {0}}), & \\hat {y} _ {h _ {1}, w _ {1}} ^ {h _ {0}, w _ {0}} = 1, \\\\ \\max (0, d _ {h _ {1}, w _ {1}} ^ {h _ {0}, w _ {0}} - c _ {\\mathrm {n}}), & \\hat {y} _ {h _ {1}, w _ {1}} ^ {h _ {0}, w _ {0}} = 0. \\end{array} \\right. \\tag{6}", + "image_path": "4b6e42e87d6005b0c915a57dbb07514bf0314b61869ae4de26cdc9d44ac00ed1.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 445, + 295, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 295, + 491 + ], + "type": "text", + "content": "The constants " + }, + { + "bbox": [ + 55, + 445, + 295, + 491 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{p}}" + }, + { + "bbox": [ + 55, + 445, + 295, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 445, + 295, + 491 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{n}}" + }, + { + "bbox": [ + 55, + 445, + 295, + 491 + ], + "type": "text", + "content": " lead to saturation for maximal desired descriptor (dis-)similarity. All descriptors of grid cells without a keypoint label index " + }, + { + "bbox": [ + 55, + 445, + 295, + 491 + ], + "type": "inline_equation", + "content": "(\\mathbf{y}_{\\mathrm{d}_{h_0},w_0} = 0" + }, + { + "bbox": [ + 55, + 445, + 295, + 491 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 445, + 295, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{\\mathrm{d}_{h_1},w_1} = 0)" + }, + { + "bbox": [ + 55, + 445, + 295, + 491 + ], + "type": "text", + "content": " do not contribute to the loss." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 492, + 295, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 295, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 295, + 636 + ], + "type": "text", + "content": "During inference, we employ a simplified, yet more effective Non-Minimum-Suppression (NMS) [51] approach in the context of keypoint detection. Instead of the common approach to model keypoints as fixed-size boxes and then apply classical NMS as in [15, 32], we find local maxima in the predicted keypoint score heatmap. In contrast to classical NMS, our method only considers local maxima and not all keypoints candidates sufficiently far from other points with higher scores. Additionally, it employs parallelizable operations instead of consecutive updates of the keypoints sorted by score. Every predicted score " + }, + { + "bbox": [ + 55, + 492, + 295, + 636 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 492, + 295, + 636 + ], + "type": "text", + "content": " at the pixel coordinates " + }, + { + "bbox": [ + 55, + 492, + 295, + 636 + ], + "type": "inline_equation", + "content": "(u, v)" + }, + { + "bbox": [ + 55, + 492, + 295, + 636 + ], + "type": "text", + "content": " is considered for a potential keypoint if" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 643, + 294, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 643, + 294, + 673 + ], + "spans": [ + { + "bbox": [ + 66, + 643, + 294, + 673 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} s [ u, v ] > s [ u + \\Delta u, v + \\Delta v ] \\\\ \\forall \\Delta u, \\Delta v \\in \\left[ - c _ {\\mathrm {b}}, \\dots , - 1, 1, \\dots , c _ {\\mathrm {b}} \\right]. \\tag {7} \\\\ \\end{array}", + "image_path": "bdfad38b37030b0ad0d205255ac167ab4f5109265a85769f60fc977967e21e39.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{b}}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " determining the size of the local neighborhood. The keypoint candidates are further filtered by a minimal required score " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{s}}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ", or the " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{s}}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " highest scores are selected." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 553, + 215 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 553, + 215 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 553, + 215 + ], + "type": "image", + "image_path": "868c56bea998a1a90bce823b64f7881688517b3d0248e358f3e323bf8ab3e9b1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 221, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 313, + 221, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 555, + 277 + ], + "type": "text", + "content": "Figure 4. SuperEvent network architecture and tensor dimensions: a shared transformer backbone is combined with a detector and a descriptor head. The components of the Convolution-Attention blocks " + }, + { + "bbox": [ + 313, + 221, + 555, + 277 + ], + "type": "inline_equation", + "content": "(Conv. + Attn.)" + }, + { + "bbox": [ + 313, + 221, + 555, + 277 + ], + "type": "text", + "content": " and the VGG blocks are displayed on the bottom left. Activations are omitted for simplicity." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 341, + 555, + 377 + ], + "blocks": [ + { + "bbox": [ + 313, + 289, + 555, + 333 + ], + "lines": [ + { + "bbox": [ + 313, + 289, + 555, + 333 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 555, + 333 + ], + "type": "text", + "content": "Table 1. Inference times of our compiled PyTorch model on GPU for common event-stream resolutions (cropped to compatible shape) using float16 precision (same accuracy as float32 in our experiments). NMS requires an additional " + }, + { + "bbox": [ + 313, + 289, + 555, + 333 + ], + "type": "inline_equation", + "content": "0.4\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 289, + 555, + 333 + ], + "type": "text", + "content": " on average." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 341, + 555, + 377 + ], + "lines": [ + { + "bbox": [ + 315, + 341, + 555, + 377 + ], + "spans": [ + { + "bbox": [ + 315, + 341, + 555, + 377 + ], + "type": "table", + "html": "
160 × 240240 × 320480 × 640720 × 1280
1.3 ms2.6 ms7.8 ms23.2 ms
", + "image_path": "55f7e88000138737fcdb8a7d259e27a8a0225fa878df51d517a1c3dd79d1c1f4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 405, + 394, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 405, + 394, + 418 + ], + "spans": [ + { + "bbox": [ + 314, + 405, + 394, + 418 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 426, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 554, + 498 + ], + "type": "text", + "content": "To validate the effectiveness of the presented approach, we conduct a series of experiments and ablation studies. Firstly, we evaluate the detection and description of keypoints by estimating the camera pose changes based on the matched points. We outperform other keypoint detectors and trackers [10, 32, 33] by a large margin." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 498, + 554, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 556 + ], + "type": "text", + "content": "In our second experiment, we integrate SuperEvent into the SLAM framework OKVIS2 [40] originally developed for frame-based cameras. We achieve convincing results on the widely used event-SLAM benchmarks [18, 35] without extensive tuning, showcasing SuperEvent's versatility." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 557, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 554, + 605 + ], + "type": "text", + "content": "For ablation studies, more qualitative examples, a detailed description of the pose estimation benchmark, and a comparison to frame-based matching for high-speed/ HDR scenarios, please refer to our Supplementary Material." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 613, + 447, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 447, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 447, + 624 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": "We generated training data as introduced in Section 3.1. To expose the model to various scenes and camera intrinsics, we combine multiple datasets [14, 31, 39, 56, 69], obtaining more than " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "100\\mathrm{k}" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": " training sample pairs with pseudolabels. We implement SuperEvent in PyTorch [4] and train it with the Adam [34] optimizer (learning rate " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": ", betas 0.9, 0.999) for 10 epochs." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8923" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 91, + 294, + 175 + ], + "blocks": [ + { + "bbox": [ + 77, + 71, + 273, + 81 + ], + "lines": [ + { + "bbox": [ + 77, + 71, + 273, + 81 + ], + "spans": [ + { + "bbox": [ + 77, + 71, + 273, + 81 + ], + "type": "text", + "content": "Table 2. Pose estimation on the Event Camera dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 91, + 294, + 175 + ], + "lines": [ + { + "bbox": [ + 56, + 91, + 294, + 175 + ], + "spans": [ + { + "bbox": [ + 56, + 91, + 294, + 175 + ], + "type": "table", + "html": "
MethodECD: Pose Estimation AUC in %
@5°@10°@20°
LLAK [10]0.71.42.1
RATE [33]3.38.418.0
EventPoint [32]1.63.05.4
SuperEvent (ours)22.735.846.7
", + "image_path": "f174b04f5661e2cd23d65907996d492fca696d4bf6293b0aea8773f2f7b63a83.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 186, + 294, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 186, + 294, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 294, + 258 + ], + "type": "text", + "content": "We employ the network architecture as in Section 3.3 with the loss function hyperparameters " + }, + { + "bbox": [ + 55, + 186, + 294, + 258 + ], + "type": "inline_equation", + "content": "c_{\\lambda} = 10" + }, + { + "bbox": [ + 55, + 186, + 294, + 258 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 186, + 294, + 258 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{d}} = 0.5" + }, + { + "bbox": [ + 55, + 186, + 294, + 258 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 186, + 294, + 258 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{p}} = 1.0" + }, + { + "bbox": [ + 55, + 186, + 294, + 258 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 186, + 294, + 258 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{n}} = 0.2" + }, + { + "bbox": [ + 55, + 186, + 294, + 258 + ], + "type": "text", + "content": ". We encode events as a 10-channel MCTS with 5 time window sizes " + }, + { + "bbox": [ + 55, + 186, + 294, + 258 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 55, + 186, + 294, + 258 + ], + "type": "text", + "content": " logarithmically distributed from " + }, + { + "bbox": [ + 55, + 186, + 294, + 258 + ], + "type": "inline_equation", + "content": "0.001 \\, \\mathrm{s}" + }, + { + "bbox": [ + 55, + 186, + 294, + 258 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 186, + 294, + 258 + ], + "type": "inline_equation", + "content": "0.1 \\, \\mathrm{s}" + }, + { + "bbox": [ + 55, + 186, + 294, + 258 + ], + "type": "text", + "content": ". The detection threshold is set to " + }, + { + "bbox": [ + 55, + 186, + 294, + 258 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{s}} = 0.01" + }, + { + "bbox": [ + 55, + 186, + 294, + 258 + ], + "type": "text", + "content": " and the NMS neighborhood " + }, + { + "bbox": [ + 55, + 186, + 294, + 258 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{b}} = 2" + }, + { + "bbox": [ + 55, + 186, + 294, + 258 + ], + "type": "text", + "content": " pixels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 258, + 294, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 294, + 294 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 294, + 294 + ], + "type": "text", + "content": "We evaluate on a desktop PC with an Intel Core i5-13600 processor, an NVIDIA GeForce GTX 4070 GPU, and 32 GB RAM. Inference times are shown in Table 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 301, + 153, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 153, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 153, + 312 + ], + "type": "text", + "content": "4.2. Pose Estimation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 318, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 318, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 54, + 318, + 295, + 533 + ], + "type": "text", + "content": "A common benchmark for evaluating frame-based keypoint detectors is pose estimation [22, 59, 62, 64]. These works use datasets such as ScanNet [12] and MegaDepth [42] containing various images of the same scene with the associated ground truth camera poses. To create a similar event-based benchmark, we employ sequences with ground truth poses from real-world dynamic datasets. We search for uniformly distributed pose changes between " + }, + { + "bbox": [ + 54, + 318, + 295, + 533 + ], + "type": "inline_equation", + "content": "1^{\\circ}" + }, + { + "bbox": [ + 54, + 318, + 295, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 318, + 295, + 533 + ], + "type": "inline_equation", + "content": "45^{\\circ}" + }, + { + "bbox": [ + 54, + 318, + 295, + 533 + ], + "type": "text", + "content": ". The upper bound is chosen to increase the probability of visual overlap. However, since we do not know the ground truth geometry of the scene, it cannot be guaranteed that every sample contains sufficient information to recover the ground truth pose. Next, we process the event stream with various approaches to detect and match keypoints. After outlier removal with Random Sample Consensus (RANSAC) [16], we estimate the camera pose and calculate the error with respect to the ground truth. Finally, we report the area-under-curve (AUC) metric for different thresholds." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "type": "text", + "content": "We compare the pose estimation capability of SuperEvent to the 3 approaches LLAK [10], RATE [33], and EventPoint [32]. While we use the open-source implementations of LLAK and RATE, we re-implemented and trained EventPoint by carefully following the steps outlined in the paper. We were unable to evaluate the approach SD2Event [19] due to unpublished code and missing implementation details in the respective work." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": "The experiment is conducted on two diverse datasets: Firstly, the Event Camera Dataset (ECD) [50] is widely used to evaluate event-based approaches, as it was one of the first publicly available datasets. It is challenging due to the low resolution " + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "(180\\times 240)" + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": ", the noisy output of the initiation DAVIS240C event camera, and fast motion changes. Secondly, we use the more recent dataset Event-aided Di" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 91, + 553, + 175 + ], + "blocks": [ + { + "bbox": [ + 315, + 71, + 553, + 82 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 553, + 82 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 553, + 82 + ], + "type": "text", + "content": "Table 3. Pose estimation on Event-aided Direct Sparse Odometry." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 91, + 553, + 175 + ], + "lines": [ + { + "bbox": [ + 315, + 91, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 315, + 91, + 553, + 175 + ], + "type": "table", + "html": "
MethodEDS: Pose Estimation AUC in %
@5°@10°@20°
LLAK [10]0.50.71.0
RATE [33]2.15.110.3
EventPoint [32]1.62.85.2
SuperEvent (ours)15.226.440.1
", + "image_path": "0400c5f873d798847ce964a9e92eef45b185cab54ada1363c6c2f65a1cde414f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 189, + 553, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 189, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 312, + 189, + 553, + 260 + ], + "type": "text", + "content": "rect Sparse Odometry (EDS) [29]. In contrast to our training data, this dataset is recorded with a Prophesee Gen " + }, + { + "bbox": [ + 312, + 189, + 553, + 260 + ], + "type": "inline_equation", + "content": "3.1^{2}" + }, + { + "bbox": [ + 312, + 189, + 553, + 260 + ], + "type": "text", + "content": " event camera with a higher resolution of " + }, + { + "bbox": [ + 312, + 189, + 553, + 260 + ], + "type": "inline_equation", + "content": "480 \\times 640" + }, + { + "bbox": [ + 312, + 189, + 553, + 260 + ], + "type": "text", + "content": " pixels and an over-proportionally increased event rate. Fast, inconsistent camera motions and a high dynamic range make pose estimation challenging." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 261, + 554, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 261, + 554, + 416 + ], + "spans": [ + { + "bbox": [ + 312, + 261, + 554, + 416 + ], + "type": "text", + "content": "Table 2 shows that SuperEvent outperforms all baselines by a large margin. On the ECD dataset, its advantage is the greatest for the high-precision estimations below " + }, + { + "bbox": [ + 312, + 261, + 554, + 416 + ], + "type": "inline_equation", + "content": "5^{\\circ}" + }, + { + "bbox": [ + 312, + 261, + 554, + 416 + ], + "type": "text", + "content": " error, confirming the effectiveness of our method. In contrast, we found that LLAK detects many keypoints but cannot track the vast majority of them. EventPoint's matching capability is severely limited, possibly due to their static assumptions during training inherently contradicting the dynamic nature of event cameras. Finally, RATE achieves decent results for estimations with up to " + }, + { + "bbox": [ + 312, + 261, + 554, + 416 + ], + "type": "inline_equation", + "content": "20^{\\circ}" + }, + { + "bbox": [ + 312, + 261, + 554, + 416 + ], + "type": "text", + "content": " error, but since it cannot recover from tracking loss, it depends on consistent and precise tracking of most keypoints. As seen in the results, this requirement cannot be stably fulfilled." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 417, + 554, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 554, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 554, + 525 + ], + "type": "text", + "content": "Table 3 indicates a slight decrease in SuperEvent's pose estimation performance on the EDS dataset compared to ECD. Since all approaches perform worse, we attribute this effect to the dataset being more challenging. Interestingly, the performance of SuperEvent for " + }, + { + "bbox": [ + 313, + 417, + 554, + 525 + ], + "type": "inline_equation", + "content": "20^{\\circ}" + }, + { + "bbox": [ + 313, + 417, + 554, + 525 + ], + "type": "text", + "content": " precision is still in the same region as for the ECD dataset. Considering the different properties of the Prophesee camera compared to the DAVIS camera employed in the training sequences, these results confirm the generalization capability of SuperEvent." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 533, + 502, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 502, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 502, + 544 + ], + "type": "text", + "content": "4.3. Stereo Event-Visual Inertial SLAM" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 551, + 554, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 551, + 554, + 694 + ], + "spans": [ + { + "bbox": [ + 313, + 551, + 554, + 694 + ], + "type": "text", + "content": "We demonstrate the usefulness of SuperEvent with a downstream experiment. To date, frame-based visual odometry and SLAM systems largely outperform their event-based counterparts, attributed to their decades-long lead in development. By integrating SuperEvent into the frame-based VI-SLAM system OKVIS2 [40], we enable this framework to process the keypoints detected in the event stream. We compare the performance of the resulting event-only system to other recent event-based Odometry and SLAM approaches on the TUM-VIE dataset [35] in Table 4. Moreover, note that the next best results of the data-driven Mono E(I)O systems DEVO and DEIO are achieved only" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 702, + 416, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 416, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 416, + 713 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 325, + 702, + 416, + 713 + ], + "type": "text", + "content": "https://www_prophesee.ai/" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8924" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 102, + 555, + 240 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 4. Results on the small scale mocap sequences of TUM-VIE dataset, ATE in cm. The first three gray results require scale alignment with the ground truth; all other approaches estimate the absolute scale. Baseline numbers taken from [27, 53]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 102, + 555, + 240 + ], + "lines": [ + { + "bbox": [ + 56, + 102, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 555, + 240 + ], + "type": "table", + "html": "
MethodModality1d-trans3d-trans6dofdeskdesk2Average
EVO [55] w/ scale alignmentMono E7.5012.5085.5054.1075.2046.96
DEVO [36] w/ scale alignmentMono E0.501.101.601.701.001.18
DEIO [27] w/ scale alignmentMono E + IMU0.421.111.371.360.731.00
DEIO [27]Mono E + IMU1.081.121.391.411.191.24
ESVO [67]Stereo E12.5417.1913.4612.924.4212.11
ES-PTAM [21]Stereo E1.058.5310.252.507.205.91
ICRA'24 [54]Stereo E + IMU3.8518.90failed8.999.47-
ESVO2 [53]Stereo E + IMU3.337.263.216.164.024.78
OKVIS2 [40] + SuperEvent (ours)Stereo E + IMU0.440.890.430.580.410.55
(ours without loop closure)(0.43)(0.89)(0.43)(0.70)(0.40)(0.57)
", + "image_path": "5032fd5b04f7db556e2e73003d0e3fecc073ae2bf683b5ec8bc28b624e34eb98.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 311, + 296, + 380 + ], + "blocks": [ + { + "bbox": [ + 55, + 258, + 296, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 304 + ], + "type": "text", + "content": "Table 5. Results on VECTor [18] large scale sequences. Our results are obtained with online calibration. Baseline numbers are taken from [27] and [36]. DEVO [36] cannot recover the absolute scale. Sequences marked " + }, + { + "bbox": [ + 55, + 258, + 296, + 304 + ], + "type": "inline_equation", + "content": "\\cdot -\\cdot" + }, + { + "bbox": [ + 55, + 258, + 296, + 304 + ], + "type": "text", + "content": " are not evaluated in [27]. ATE in cm." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 311, + 296, + 380 + ], + "lines": [ + { + "bbox": [ + 56, + 311, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 311, + 296, + 380 + ], + "type": "table", + "html": "
Methodcorr.-dollycorr.-walkunits-dollyunits-scooterschool-dollyschool-scooter
Ours33.13133.16122.6159.0569.9639.07
DEIO [27]492.65325.00826.38304.14--
ESVO [67]failedfailedfailedfailed1371.0983.0
", + "image_path": "c56d9126ddd0f75e9659ec6e6bc20ca5491e55c1dd240b2269fa481e3034b7ec.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 393, + 296, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 393, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 393, + 296, + 441 + ], + "type": "text", + "content": "after aligning the scale of the estimated trajectories with ground truth. However, real-world applications, e.g., VR and robotics, typically require accurate metric scale. Results on the VECtor dataset [18] are listed in Table 5." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 441, + 295, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 441, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 54, + 441, + 295, + 620 + ], + "type": "text", + "content": "OKVIS2's loop closure capability relies on long-term descriptor matching. While its effect on the small-scale sequences is limited, we evaluate OKVIS2 + SuperEvent on the loop-floor sequences of the TUM-VIE dataset, since these are explicitly designed to test a system's loop-closure capabilities. Thanks to SuperEvent, the OKVIS2 loop-closure feature works reliably with events, as shown in Table 6. Figure 5 qualitatively compares the estimated trajectories of the sequence loop-floor0. Without loop closure, there is significant drift - albeit not out of the ordinary compared to frame-based VI-SLAM. However, with loop closure enabled, the starting room is recognized and the drift is corrected. This demonstrates the robustness of SuperEvent, achieving successful descriptor matching even with minutes of time between observations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 633, + 129, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 129, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 129, + 645 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "We presented SuperEvent, a novel keypoint detector and descriptor for event streams. During its training, we consider the time-dependent keypoint appearance and employ pseudo-labels detected on frames. Combined with the flexible MCTS representation and a transformer back" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 289, + 554, + 346 + ], + "blocks": [ + { + "bbox": [ + 313, + 258, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 281 + ], + "type": "text", + "content": "Table 6. Effect of loop closure of OKVIS2 [40] + SuperEvent on TUM-VIE loop-floor sequences, ATE in cm." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 289, + 554, + 346 + ], + "lines": [ + { + "bbox": [ + 315, + 289, + 554, + 346 + ], + "spans": [ + { + "bbox": [ + 315, + 289, + 554, + 346 + ], + "type": "table", + "html": "
loop-floor0123
Estimated length349 m316 m279 m303 m
Loop closure4.964.648.924.74
W/o loop closure132.11161.92116.00129.17
", + "image_path": "6fef1e60530def4d3889263365b5b6c6affc19d139e05b433ae2778722be7296.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 319, + 358, + 550, + 453 + ], + "blocks": [ + { + "bbox": [ + 319, + 358, + 550, + 453 + ], + "lines": [ + { + "bbox": [ + 319, + 358, + 550, + 453 + ], + "spans": [ + { + "bbox": [ + 319, + 358, + 550, + 453 + ], + "type": "image", + "image_path": "f97789cb0ed71692f3f0e708eee8dccc484b96bf1a853b387e5a0620b3898672.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 458, + 555, + 482 + ], + "lines": [ + { + "bbox": [ + 313, + 458, + 555, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 458, + 555, + 482 + ], + "type": "text", + "content": "Figure 5. OKVIS2 + SuperEvent's estimated trajectories of TUMVIE sequence loop-floor0 with and without loop closure." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 493, + 554, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 554, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 554, + 541 + ], + "type": "text", + "content": "bone, it achieves SOTA performance on keypoint-based pose estimation. SuperEvent's integration into frame-based downstream applications enables event vision in existing keypoint-based systems." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 542, + 554, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 554, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 554, + 613 + ], + "type": "text", + "content": "As future work, SuperEvent can be combined with a frame-based detector to realize systems that simultaneously process frames and events, exploiting their complementary advantages. Also, with a similar approach to SuperEvent, an event-based line detector could improve the performance of downstream applications in man-made environments." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 613, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 555, + 715 + ], + "type": "text", + "content": "Finally, our training data is generated to enable consistent feature description independent of their motion-dependent appearance. However, since we employ real sequences, consecutive samples are biased to having similar motion directions. This is reflected in the performance of SuperEvent, whose descriptor matchability suffers under strong motion changes. While this is an inherent property of event cameras, a different approach to generating training data could mitigate it." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8925" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 97, + 115, + 109 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 97, + 115, + 109 + ], + "spans": [ + { + "bbox": [ + 56, + 97, + 115, + 109 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 117, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 61, + 117, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 117, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 61, + 117, + 295, + 150 + ], + "type": "text", + "content": "[1] Ignacio Alzugaray and Margarita Chli. Asynchronous corner detection and tracking for event cameras in real time. IEEE Robotics and Automation Letters, 3(4):3177-3184, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 152, + 295, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 152, + 295, + 194 + ], + "spans": [ + { + "bbox": [ + 61, + 152, + 295, + 194 + ], + "type": "text", + "content": "[2] Ignacio Alzugaray and Margarita Chli. Ace: An efficient asynchronous corner tracker for event cameras. In 2018 International Conference on 3D Vision (3DV), pages 653-661. IEEE, 2018. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 197, + 295, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 197, + 295, + 252 + ], + "spans": [ + { + "bbox": [ + 62, + 197, + 295, + 252 + ], + "type": "text", + "content": "[3] Ignacio Alzugaray and Margarita Chli. HASTE: Multi-hypothesis asynchronous speeded-up tracking of events. In 31st British Machine Vision Virtual Conference (BMVC 2020), page 744. ETH Zurich, Institute of Robotics and Intelligent Systems, 2020. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 254, + 294, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 254, + 294, + 450 + ], + "spans": [ + { + "bbox": [ + 62, + 254, + 294, + 450 + ], + "type": "text", + "content": "[4] Jason Ansel, Edward Yang, Horace He, Natalia Gimelshein, Animesh Jain, Michael Voznesensky, Bin Bao, Peter Bell, David Berard, Evgeni Burovski, Geeta Chauhan, Anjali Choudria, Will Constable, Alban Desmaison, Zachary DeVito, Elias Ellison, Will Feng, Jiong Gong, Michael Gschwind, Brian Hirsh, Sherlock Huang, Kshitteej Kalambarkar, Laurent Kirsch, Michael Lazos, Mario Lezcano, Yanbo Liang, Jason Liang, Yinghai Lu, CK Luk, Bert Maher, Yunjie Pan, Christian Puhrsch, Matthias Reso, Mark Saroufim, Marcos Yukio Siraichi, Helen Suk, Michael Suo, Phil Tillet, Eikan Wang, Xiaodong Wang, William Wen, Shunting Zhang, Xu Zhao, Keren Zhou, Richard Zou, Ajit Mathews, Gregory Chanan, Peng Wu, and Soumith Chintala. PyTorch 2: Faster Machine Learning Through Dynamic Python Bytecode Transformation and Graph Compilation. In 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2 (ASPLOS '24). ACM, 2024. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 453, + 295, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 453, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 62, + 453, + 295, + 486 + ], + "type": "text", + "content": "[5] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-up robust features (SURF). Computer vision and image understanding, 110(3):346-359, 2008. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 487, + 295, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 487, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 62, + 487, + 295, + 553 + ], + "type": "text", + "content": "[6] Michael Calonder, Vincent Lepetit, Christoph Strecha, and Pascal Fua. Brief: Binary robust independent elementary features. In Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part IV 11, pages 778-792. Springer, 2010. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 555, + 295, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 555, + 295, + 609 + ], + "spans": [ + { + "bbox": [ + 62, + 555, + 295, + 609 + ], + "type": "text", + "content": "[7] Carlos Campos, Richard Elvira, Juan J. Gómez, José M. M. Montiel, and Juan D. Tardós. ORB-SLAM3: An accurate open-source library for visual, visual-inertial and multi-map SLAM. IEEE Transactions on Robotics, 37(6):1874-1890, 2021. 1, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 611, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 611, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 62, + 611, + 295, + 644 + ], + "type": "text", + "content": "[8] Peiyu Chen, Weipeng Guan, and Peng Lu. ESVIO: Event-based stereo visual inertial odometry. IEEE Robotics and Automation Letters, 8(6):3661-3668, 2023. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 646, + 295, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 646, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 62, + 646, + 295, + 700 + ], + "type": "text", + "content": "[9] Philippe Chiberre, Etienne Perot, Amos Sironi, and Vincent Lepetit. Detecting stable keypoints from events through image gradient prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1387-1394, 2021. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 702, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 702, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 702, + 295, + 713 + ], + "type": "text", + "content": "[10] Philippe Chiberre, Etienne Perot, Amos Sironi, and Vincent" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Lepetit. Long-lived accurate keypoints in event streams. arXiv preprint arXiv:2209.10385, 2022. 1, 3, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 97, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 129 + ], + "type": "text", + "content": "[11] Xavier Clady, Sio-Hoi Ieng, and Ryad Benosman. Asynchronous event-based corner detection and matching. Neural Networks, 66:91-106, 2015. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 130, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 130, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 317, + 130, + 553, + 185 + ], + "type": "text", + "content": "[12] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828–5839, 2017. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 186, + 553, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 219 + ], + "type": "text", + "content": "[13] Laurent Dardelet, Ryad Benosman, and Sio-Hoi Ieng. An event-by-event feature detection and tracking invariant to motion direction and velocity. Authorea Preprints, 2023. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 220, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 220, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 317, + 220, + 553, + 275 + ], + "type": "text", + "content": "[14] Jeffrey Delmerico, Titus Cieslewski, Henri Rebecq, Matthias Faessler, and Davide Scaramuzza. Are we ready for autonomous drone racing? the uzh-fpv drone racing dataset. In 2019 International Conference on Robotics and Automation (ICRA), pages 6713-6719. IEEE, 2019. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 277, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 277, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 317, + 277, + 553, + 331 + ], + "type": "text", + "content": "[15] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperPoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224-236, 2018. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 332, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 332, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 317, + 332, + 553, + 376 + ], + "type": "text", + "content": "[16] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 377, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 377, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 317, + 377, + 553, + 432 + ], + "type": "text", + "content": "[17] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 434, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 434, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 434, + 553, + 487 + ], + "type": "text", + "content": "[18] Ling Gao, Yuxuan Liang, Jiaqi Yang, Shaoxun Wu, Chenyu Wang, Jiaben Chen, and Laurent Kneip. Vector: A versatile event-centric benchmark for multi-sensor slam. IEEE Robotics and Automation Letters, 7(3):8217-8224, 2022. 3, 6, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 489, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 489, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 489, + 553, + 544 + ], + "type": "text", + "content": "[19] Yuan Gao, Yuqing Zhu, Xinjun Li, Yimin Du, and Tianzhu Zhang. SD2Event: Self-supervised learning of dynamic detectors and contextual descriptors for event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3055-3064, 2024. 1, 3, 5, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 545, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 553, + 589 + ], + "type": "text", + "content": "[20] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13884-13893, 2023. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 590, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 634 + ], + "type": "text", + "content": "[21] Suman Ghosh, Valentina Cavinato, and Guillermo Gallego. ES-PTAM: Event-based stereo parallel tracking and mapping. In European Conference on Computer Vision (ECCV) Workshops, 2024. 4, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "type": "text", + "content": "[22] Pierre Gleize, Weiyao Wang, and Matt Feiszli. SiLK: Simple learned keypoints. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 22442-22451, 2023. 2, 3, 5, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "text", + "content": "[23] Arren Glover, Aiko Dinale, Leandro De Souza Rosa, Simeon Bamford, and Chiara Bartolozzi. luvHarris: A practical corner detector for event-cameras. IEEE Transactions on" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8926" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 72, + 295, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 295, + 95 + ], + "type": "text", + "content": "Pattern Analysis and Machine Intelligence, 44(12):10087-10098, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "type": "text", + "content": "[24] Weipeng Guan and Peng Lu. Monocular event visual inertial odometry based on event-corner using sliding windows graph-based optimization. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2438-2445, 2022. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "text", + "content": "[25] Weipeng Guan, Peiyu Chen, Yuhan Xie, and Peng Lu. PL- EVIO: Robust monocular event-based visual inertial odometry with point and line features. IEEE Transactions on Automation Science and Engineering, 2023. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "type": "text", + "content": "[26] Weipeng Guan, Peiyu Chen, Huibin Zhao, Yu Wang, and Peng Lu. EVI-SAM: Robust, real-time, tightly-coupled event-visual-inertial state estimation and 3d dense mapping. Advanced Intelligent Systems, 6(12):2400243, 2024. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 274 + ], + "type": "text", + "content": "[27] Weipeng Guan, Fuling Lin, Peiyu Chen, and Peng Lu. DEIO: Deep event inertial odometry. arXiv preprint arXiv:2411.03928, 2024. 4, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 294, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 298 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 298 + ], + "type": "text", + "content": "[28] Christopher G. Harris and M. J. Stephens. A combined corner and edge detector. In Alvey Vision Conference, 1988. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 300, + 294, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 300, + 294, + 343 + ], + "spans": [ + { + "bbox": [ + 56, + 300, + 294, + 343 + ], + "type": "text", + "content": "[29] Javier Hidalgo-Carrió, Guillermo Gallego, and Davide Scaramuzza. Event-aided direct sparse odometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5781–5790, 2022. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 345, + 294, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 345, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 345, + 294, + 399 + ], + "type": "text", + "content": "[30] Sumin Hu, Yeeun Kim, Hyungtae Lim, Alex Junho Lee, and Hyun Myung. eCDT: Event clustering for simultaneous feature detection and tracking. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3808-3815. IEEE, 2022. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 400, + 294, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 400, + 294, + 466 + ], + "spans": [ + { + "bbox": [ + 56, + 400, + 294, + 466 + ], + "type": "text", + "content": "[31] Yuhuang Hu, Jonathan Binas, Daniel Neil, Shih-Chii Liu, and Tobi Delbruck. DDD20 end-to-end event camera driving dataset: Fusing frames and events with deep learning for improved steering prediction. In 2020 IEEE 23rd International Conference on Intelligent Transportation Systems (ITSC), pages 1-6. IEEE, 2020. 1, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 468, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 294, + 521 + ], + "type": "text", + "content": "[32] Ze Huang, Li Sun, Cheng Zhao, Song Li, and Songzhi Su. EventPoint: Self-supervised interest point detection and description for event-based camera. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5396-5405, 2023. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 523, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 589 + ], + "type": "text", + "content": "[33] Mikihiro Ikura, Cedric Le Gentil, Marcus G. Müller, Florian Schuler, Atsushi Yamashita, and Wolfgang Stürzl. RATE: Real-time asynchronous feature tracking with event cameras. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 11662-11669, 2024. 1, 4, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 612 + ], + "type": "text", + "content": "[34] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 613, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 667 + ], + "type": "text", + "content": "[35] Simon Klenk, Jason Chui, Nikolaus Demmel, and Daniel Cremers. TUM-VIE: The TUM stereo visual-inertial event dataset. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8601-8608. IEEE, 2021. 3, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[36] Simon Klenk, Marvin Motzet, Lukas Koestler, and Daniel Cremers. Deep event visual odometry. In 2024 International Conference on 3D Vision (3DV), pages 739-749. IEEE, 2024. 4, 8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "text", + "content": "[37] Beat Kueng, Elias Mueggler, Guillermo Gallego, and Davide Scaramuzza. Low-latency visual odometry using event-based feature tracks. In 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 16-23. IEEE, 2016. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "type": "text", + "content": "[38] Xavier Lagorce, Garrick Orchard, Francesco Galluppi, Bertram E. Shi, and Ryad B. Benosman. HOTS: A hierarchy of event-based time-surfaces for pattern recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(7):1346-1359, 2017. 2, 3, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 186, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 228 + ], + "type": "text", + "content": "[39] Alex Junho Lee, Younggun Cho, Young-sik Shin, Ayoung Kim, and Hyun Myung. ViViD++: Vision for visibility dataset. IEEE Robotics and Automation Letters, 7(3):6282-6289, 2022. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 230, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 252 + ], + "type": "text", + "content": "[40] Stefan Leutenegger. OKVIS2: Realtime scalable visual-inertial SLAM with loop closure, 2022. 1, 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "type": "text", + "content": "[41] Stefan Leutenegger, Margarita Chli, and Roland Y. Siegwart. BRISK: Binary robust invariant scalable keypoints. In 2011 International Conference on Computer Vision, pages 2548-2555, 2011. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "type": "text", + "content": "[42] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2041-2050, 2018. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 344, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 553, + 398 + ], + "type": "text", + "content": "[43] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 399, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 431 + ], + "type": "text", + "content": "[44] Zhe Liu, Dianxi Shi, Ruihao Li, and Shaowu Yang. ESVIO: event-based stereo visual-inertial odometry. Sensors, 23(4): 1998, 2023. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 434, + 553, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 466 + ], + "type": "text", + "content": "[45] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60:91-110, 2004. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 468, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 553, + 521 + ], + "type": "text", + "content": "[46] Florian Mahlknecht, Daniel Gehrig, Jeremy Nash, Friedrich M Rockenbauer, Benjamin Morrell, Jeff Delaune, and Davide Scaramuzza. Exploring event camera-based odometry for planetary robots. IEEE Robotics and Automation Letters, 7(4):8651-8658, 2022. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 523, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 589 + ], + "type": "text", + "content": "[47] Jacques Manderscheid, Amos Sironi, Nicolas Bourdis, Davide Migliore, and Vincent Lepetit. Speed invariant time surface for learning to detect corner points with event-based cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10245-10254, 2019. 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 590, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 644 + ], + "type": "text", + "content": "[48] Nico Messikommer, Carter Fang, Mathias Gehrig, and Davide Scaramuzza. Data-driven feature tracking for event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5642-5651, 2023. 4, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "type": "text", + "content": "[49] Elias Mueggler, Chiara Bartolozzi, and Davide Scaramuzza. Fast event-based corner detection. Proceedings of the British Machine Vision Conference, 2017. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[50] Elias Mueggler, Henri Rebecq, Guillermo Gallego, Tobi Delbruck, and Davide Scaramuzza. The event-camera dataset and simulator: Event-based data for pose estimation, visual" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8927" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "odometry, and slam. The International Journal of Robotics Research, 36(2):142-149, 2017. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 139 + ], + "type": "text", + "content": "[51] Alexander Neubeck and Luc Van Gool. Efficient nonmaximum suppression. In 18th international conference on pattern recognition (ICPR'06), pages 850-855. IEEE, 2006. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "type": "text", + "content": "[52] Zhenjiang Ni, Aude Bolopion, Joel Agnus, Ryad Benosman, and Stephane Regnier. Asynchronous event-based visual shape tracking for stable haptic feedback in microrobotics. IEEE Transactions on Robotics, 28(5):1081-1089, 2012. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 230 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 230 + ], + "type": "text", + "content": "[53] Junkai Niu, Sheng Zhong, Xiuyuan Lu, Shaojie Shen, Guillermo Gallego, and Yi Zhou. ESVO2: Direct visual-inertial odometry with stereo event cameras. arXiv e-prints, pages arXiv-2410, 2024. 4, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 232, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 294, + 274 + ], + "type": "text", + "content": "[54] Junkai Niu, Sheng Zhong, and Yi Zhou. Imu-aided event-based stereo visual odometry. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 11977–11983, 2024. 4, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "text", + "content": "[55] Henri Rebecq, Timo Horstschäfer, Guillermo Gallego, and Davide Scaramuzza. EVO: A geometric approach to event-based 6-dof parallel tracking and mapping in real time. IEEE Robotics and Automation Letters, 2(2):593-600, 2016. 4, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 321, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 387 + ], + "type": "text", + "content": "[56] Juan Pablo Rodríguez-Gómez, Raul Tapia, Julio L Paneque, Pedro Grau, Augusto Gómez Eguíluz, Jose Ramiro Martínez-de Dios, and Anibal Ollero. The GRIFFIN perception dataset: Bridging the gap between flapping-wing flight and robotic perception. IEEE Robotics and Automation Letters, 6(2):1066-1073, 2021. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 443 + ], + "type": "text", + "content": "[57] Edward Rosten and Tom Drummond. Machine learning for high-speed corner detection. In Computer Vision-ECCV 2006: 9th European Conference on Computer Vision, Graz, Austria, May 7-13, 2006. Proceedings, Part I 9, pages 430-443. Springer, 2006. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "type": "text", + "content": "[58] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. ORB: An efficient alternative to SIFT or SURF. In 2011 International Conference on Computer Vision, pages 2564-2571, 2011. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "type": "text", + "content": "[59] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 1, 2, 3, 4, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "type": "text", + "content": "[60] Jianbo Shi and Tomasi. Good features to track. In 1994 Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 593-600, 1994. 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 580, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 294, + 612 + ], + "type": "text", + "content": "[61] Karen Simonyan. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 614, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 667 + ], + "type": "text", + "content": "[62] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 2, 3, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[63] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. MaxViT: Multi-axis vision transformer. In European conference on computer vision, pages 459-479. Springer, 2022. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 418 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "text", + "content": "[64] Michal Tyszkiiewicz, Pascal Fua, and Eduard Trulls. DISK: Learning local features with policy gradient. Advances in Neural Information Processing Systems, 33:14254-14265, 2020. 2, 3, 5, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 172 + ], + "type": "text", + "content": "[65] Valentina Vasco, Arren Glover, and Chiara Bartolozzi. Fast event-based harris corner detection exploiting the advantages of event-driven cameras. In 2016 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 4144-4149. IEEE, 2016. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "text", + "content": "[66] Antoni Rosinol Vidal, Henri Rebecq, Timo Horstschaefer, and Davide Scaramuzza. Ultimate SLAM? Combining events, images, and IMU for robust visual slam in hdr and high-speed scenarios. IEEE Robotics and Automation Letters, 3(2):994-1001, 2018. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 230, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 262 + ], + "type": "text", + "content": "[67] Yi Zhou, Guillermo Gallego, and Shaojie Shen. Event-based stereo visual odometry. IEEE Transactions on Robotics, 37 (5):1433-1450, 2021. 4, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "text", + "content": "[68] Alex Zihao Zhu, Nikolay Atanasov, and Kostas Daniilidis. Event-based feature tracking with probabilistic data association. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 4465-4470. IEEE, 2017. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 309, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 363 + ], + "type": "text", + "content": "[69] Alex Zihao Zhu, Dinesh Thakur, Tolga Özaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The multi-vehicle stereo event camera dataset: An event camera dataset for 3d perception. IEEE Robotics and Automation Letters, 3 (3):2032–2039, 2018. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 365, + 553, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 418 + ], + "type": "text", + "content": "[70] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based optical flow using motion compensation. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, 2018. 5" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8928" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SuperMat_ Physically Consistent PBR Material Estimation at Interactive Rates/4daa8d06-6978-409c-9337-f3e4e925bd23_content_list.json b/2025/SuperMat_ Physically Consistent PBR Material Estimation at Interactive Rates/4daa8d06-6978-409c-9337-f3e4e925bd23_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d02b4551c27e2c246048351e17a8be438291bf03 --- /dev/null +++ b/2025/SuperMat_ Physically Consistent PBR Material Estimation at Interactive Rates/4daa8d06-6978-409c-9337-f3e4e925bd23_content_list.json @@ -0,0 +1,1580 @@ +[ + { + "type": "text", + "text": "SuperMat: Physically Consistent PBR Material Estimation at Interactive Rates", + "text_level": 1, + "bbox": [ + 98, + 130, + 898, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yijia Hong $^{1,2}$ , Yuan-Chen Guo $^{4*}$ , Ran Yi $^{1}$ , Yulong Chen $^{3}$ , Yan-Pei Cao $^{4\\dagger}$ , Lizhuang Ma $^{1\\dagger}$ $^{1}$ Shanghai Jiao Tong University $^{2}$ Shanghai Innovation Institute $^{3}$ Harbin Institute of Technology $^{4}$ VAST", + "bbox": [ + 107, + 179, + 887, + 232 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3e5b4d318e19304306eb1958d0d7d951e93446b05fc63057ec4f4215243a8b0d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 256, + 173, + 335 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ac9cec3ae7531345cd3634bebdbb2ff5f7ace3c03d12bedd0a65707407c8ba06.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 348, + 173, + 425 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4520b402f018699f03a66ba1caa3806275fe221ca91fc783f720ff6d11642c39.jpg", + "image_caption": [ + "Figure 1. We present an efficient method for high-quality material decomposition in both 2D and 3D. Left: Albedo and RM(Roughness, Metallic) predictions of SuperMat on rendered images under unknown illumination. The first row shows inputs with uneven lighting, the second row displays albedo predictions, and the third row presents RM predictions. Each texel in the RM map here is represented as (1, roughness, metallic). Middle: Rendering results of a 3D object with PBR materials obtained by our method under novel lighting. Right: Our method is efficient and produces high-quality results." + ], + "image_footnote": [], + "bbox": [ + 102, + 440, + 173, + 517 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b83e380c2526df665c449a363951bdaff9dabd27d693a821ac70d7311bab05d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 200, + 258, + 320, + 334 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c158dd331366d0f7b9ddd0284510517fa9fb60d09304d968a47b24a8015ac1cb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 199, + 348, + 326, + 425 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a7c53e92dbf3047f3ba8a0b951ebffb9d237c083849097cf153d784e56588639.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 199, + 440, + 326, + 517 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e3bac2274340d9958e82cde9780f7c6c43f866b6e0b5d6131e9f14114a9ef305.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 348, + 260, + 415, + 332 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/017dce4215003c430b1d983730ce04e2689dd74c3b88dd336210e92b6319ccd4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 348, + 348, + 415, + 424 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0efdaa27ba494598c1c885f2c05e0b7740930f1b960fe7679c6cbe2d48106c1c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 348, + 440, + 415, + 516 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e61811d1388ec4840c75b64c1c83166b1507d2e96b90c46fdf9c081b8d18d204.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 260, + 640, + 342 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ae099eae91dbbf5d0853adfdda42b1c2261fe7d706b5383304092d926c55aa05.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 345, + 640, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ba3683dbf0d29127d3655e43ff5b9c6e799083d150a131edd86d7301a837d089.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 430, + 640, + 516 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ab298a554890b80d69831a17f26157075d2040815fbc9a60f9a532a889caeac2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 258, + 893, + 521 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 617, + 326, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Decomposing physically-based materials from images into their constituent properties remains challenging, particularly when maintaining both computational efficiency and physical consistency. While recent diffusion-based approaches have shown promise, they face substantial computational overhead due to multiple denoising steps and separate models for different material properties. We present SuperMat, a single-step framework that achieves high-quality material decomposition with one-step inference. This enables end-to-end training with perceptual and re-render losses while decomposing albedo, metallic, and roughness maps at millisecond-scale speeds. We further extend our framework to 3D objects through a UV refinement network, enabling consistent material estimation across viewpoints", + "bbox": [ + 88, + 648, + 485, + 861 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "while maintaining efficiency. Experiments demonstrate that SuperMat achieves state-of-the-art PBR material decomposition quality while reducing inference time from seconds to milliseconds per image, and completes PBR material estimation for 3D objects in approximately 3 seconds. The project page is at https://hyj542682306.github.io/SuperMat/.", + "bbox": [ + 511, + 618, + 906, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 753, + 645, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Decomposing Physically-Based Rendering (PBR) materials from images into their constituent properties (e.g., albedo, metallic, and roughness maps) remains a central challenge in computer vision and graphics. This task is essential for creating high-fidelity 3D assets across numerous applications, from film production and game development to virtual and augmented reality pipelines. While recent advances have shown promising results [6, 28, 56], the problem re", + "bbox": [ + 509, + 779, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Project lead", + "bbox": [ + 107, + 875, + 181, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 109, + 887, + 233, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25083", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mains challenging due to its ill-posed nature, particularly given unknown lighting conditions and complex interactions between different material properties. Our work addresses two key objectives: 1) efficient decomposition of high-fidelity material maps from 2D images, and 2) extension of this capability to 3D objects, enabling accurate material estimation across complex geometries.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous works in PBR material estimation generally fall into two main categories: optimization-based generation for 3D objects and data-driven decomposition for 2D images. Optimization-based methods [15, 19, 22, 26] employ differentiable rendering to reconstruct 3D objects with material parameters by minimizing the difference between rendered and target images. Instead, data-driven approaches learn the distribution of various materials from large datasets to perform intrinsic decomposition on 2D images, which is the primary focus of our work. Early methods [46, 64] in this domain fail to produce stable and reliable results due to limited data availability and model capacity. Recent works [6, 25, 28, 56] leverage large diffusion models and reframe material decomposition as a conditional image generation task. However, these methods tend to focus solely on individual material properties, overlooking the interactions between different types of materials and the combined impact on rendered appearances. Even minor inaccuracies in individual material estimates can propagate through the rendering pipeline, leading to substantial visual artifacts. Moreover, these diffusion-based methods face considerable computational overhead - they demand multiple denoising steps (30-50) during inference, causing a critical bottleneck for real-time applications.", + "bbox": [ + 91, + 200, + 486, + 564 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose a deterministic one-step framework that achieves both efficient and high-quality material decomposition. First, we introduce and refine structural expert branches within the UNet architecture. These specialized branches enable concurrent processing of multiple material properties while maintaining domain-specific expertise, allowing our model to decompose albedo, metallic, and roughness maps simultaneously with minimal computational overhead. This design facilitates collaborative and physically consistent material prediction while avoiding the linear cost scaling typically required when estimating multiple material attributes. Second, we address the computational inefficiency of traditional diffusion models by reformulating the DDIM sampling process. Through modification of the scheduler, we achieve single-step inference without compromising quality, dramatically reducing inference time. Most importantly, this single-step formulation enables end-to-end training of the diffusion model, allowing us to compute the perceptual and re-render loss on final predicted material textures, which enhances shadow removal and improves the joint performance of decomposed materials under varying lighting", + "bbox": [ + 89, + 569, + 486, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "conditions. These advances lead to SuperMat, a highly efficient image space material decomposition model that estimates high-fidelity albedo, metallic, and roughness maps at millisecond-scale speeds.", + "bbox": [ + 511, + 90, + 906, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For practical 3D applications, we further extend SuperMat's capabilities to handle 3D objects. We upgrade SuperMat to a multi-view model, SuperMatMV, and design a UV refinement network that processes SuperMatMV's predictions in UV space. This approach addresses the challenge of maintaining consistency and completeness across different, limited viewpoints while preserving the efficiency of our original 2D approach. Our pipeline processes a textured 3D model in approximately 3 seconds, producing high-quality material maps that maintain consistency across arbitrary viewpoints and lighting conditions.", + "bbox": [ + 511, + 152, + 908, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are:", + "bbox": [ + 531, + 318, + 764, + 330 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose SuperMat, a novel image space material decomposition model that generates physically consistent albedo, metallic, and roughness maps simultaneously with a single-step inference process. Our method achieves state-of-the-art performance on diverse datasets, significantly outperforming previous diffusion-based works in speed.", + "- With the help of structural expert branches and single-step inference, we develop an end-to-end image space material decomposition training system that generates clean predictions for all material properties in each iteration, allowing for the integration of advanced techniques such as perceptual loss and re-render loss.", + "- We propose an efficient PBR material decomposition method for 3D models. Combining SuperMatMV and our UV refinement network, this pipeline generates high-quality PBR materials for 3D objects from baked textures in approximately 3 seconds." + ], + "bbox": [ + 513, + 334, + 903, + 606 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 618, + 663, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Material Generation and Decomposition. Recognizing the importance of material information in 3D assets, researchers have increasingly focused on material acquisition for objects and scenes. On one hand, some 3D generation approaches [5, 9, 32, 37, 47, 49, 52, 53, 62] integrate material optimization into the appearance creation process, concentrating on generating PBR maps for textured models. On the other hand, inverse rendering methods aim to reconstruct material-rich 3D representations from single or multi-view images. Some of these methods [7, 19, 57] assume a predefined lighting model to enhance estimation accuracy, but this constraint limits their practical applicability. To address this, other works [2, 15, 20, 22, 26, 34, 59-61] attempt to recover both material properties and unknown illumination simultaneously. As reliance on predefined assumptions decreases, researchers have found that beyond controlling material decomposition through rendered out", + "bbox": [ + 511, + 643, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25084", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e5cff5141d57d26bbfc4c82c263aaf326ddf2eceeedb537c90e47a1380ac6e63.jpg", + "image_caption": [ + "Figure 2. Overview of the SuperMat framework. The UNet architecture incorporates structural expert branches for albedo and RM estimation, enabling parallel material property prediction while sharing a common backbone. During training, we leverage a fixed DDIM scheduler and optimize the network end-to-end using both perceptual loss and re-render loss. The marked region shows the inference pipeline, where our model achieves single-step inference for concurrent material map generation. Input geometric information (normal, position) and environment maps are used to compute the re-render loss." + ], + "image_footnote": [], + "bbox": [ + 109, + 93, + 888, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "puts, directly obtaining image space material priors is also feasible. Building on this idea, similar to ours, another body of research [6, 25, 28, 30, 42-44, 46, 51, 56, 64] focuses on material decomposition from 2D images captured or rendered under unknown lighting conditions. However, existing methods still face challenges related to training data, model architecture, and training strategies, leaving significant room for improvement in decomposition quality.", + "bbox": [ + 88, + 395, + 480, + 516 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Models. Recently, researchers have applied diffusion models to a range of tasks [5, 36, 37, 45, 62], demonstrating the effectiveness of casting outcome prediction as a conditional image generation process. By leveraging prior knowledge from large pre-trained models, recent studies [6, 11, 14, 18, 24, 28, 40, 50, 63] have achieved performance improvements over previous methods in tasks like depth, normal, and material estimation. However, incorporating large models introduces efficiency challenges, particularly high memory requirements and the time-intensive, multistep denoising process needed for inference. Inspired by insights from recent work [17, 21, 27, 29], our approach addresses these issues by significantly reducing memory and inference time, improving the feasibility of applying these models in real-world applications.", + "bbox": [ + 89, + 518, + 482, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generation in UV Space. UV maps are advantageous for fully encapsulating the appearance information of 3D models, leading several works to explore 3D object manipulation directly within the UV space. By representing the texture of an object as a learnable image, these methods [3, 4, 16, 33, 38, 55] enable texture generation, inpainting, and editing, providing new avenues for 3D tasks and allowing 2D image-space techniques and insights to be applied to 3D appearances. Building on this approach, recent methods [1, 12, 31, 52-55] have applied diffusion models to refine", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "UV maps with promising results. In this spirit, to extend SuperMat's capability to 3D, we fine-tune a UV-space refinement diffusion model specifically.", + "bbox": [ + 511, + 393, + 906, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 452, + 604, + 468 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we propose SuperMat, an image space material decomposition model fine-tuned from Stable Diffusion that efficiently generates high-quality albedo, metallic, and roughness maps simultaneously. Given a rendered or captured image of an object under an unknown lighting condition from a specific viewpoint, SuperMat generates the corresponding material maps for the model from that same viewpoint. As shown in Figure 2, SuperMat is a 2D imaged-to-image one-step model with structural expert branches that simultaneously generates both albedo and RM maps, supporting single-step inference. Compared to previous diffusion-based methods, this model produces superior results in less time and with reduced memory usage. Our single-step formulation allows direct generation of material maps during training, in contrast to prior diffusion methods that predict noise $(\\epsilon)$ or velocity (v) terms. This enables end-to-end optimization with both perceptual loss on the generated maps and re-render loss from novel lighting conditions, leading to improved material quality and rendering fidelity. Building upon the single-view model, we further introduce a multi-view version, SuperMatMV, to ensure consistency in predictions across different viewpoints.", + "bbox": [ + 509, + 478, + 906, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While SuperMat and SuperMatMV achieve efficient image space material decomposition, extending this capability to 3D objects presents additional challenges. Instead of relying on optimization-based approaches like score distillation sampling (SDS) that are computationally intensive and prone to mode collapse, we propose an efficient two-stage", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25085", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "pipeline for material estimation on 3D objects. Given a 3D model with only RGB texture, we first apply SuperMatMV to decompose materials from multiple viewpoints. We then introduce a UV refinement network that consolidates these view-dependent predictions into complete, high-quality UV texture maps through single-step inference. As shown in Figure 3, this pipeline enables PBR material estimation for 3D objects in approximately 3 seconds per model.", + "bbox": [ + 89, + 90, + 480, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. SuperMat - Image Space Material Decomposition", + "text_level": 1, + "bbox": [ + 89, + 224, + 480, + 255 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Previous works [6, 28, 56] have demonstrated that casting material estimation as a conditional latent diffusion process, leveraging knowledge from large pre-trained diffusion models like Stable Diffusion [39], is effective. However, for our task that requires the simultaneous generation of multiple material maps, three key issues remain: 1) High model overhead: Fine-tuning Stable Diffusion with its original architecture can only estimate a single material type. Even when combining metallic and roughness into an RM map, treating it as a single material still requires two diffusion models, one for albedo and another for RM, to meet the rendering model's requirements. This results in doubled training time and increased inference space overhead. 2) Slow inference speed: With the widely used DDIM scheduler, diffusion models require 30 to 50 denoising steps to produce satisfactory results, which can be inefficient and time-consuming. 3) Insufficient decomposition effect: Due to the high computational cost of backpropagation in multistep inference, prior diffusion-based methods trained on predicted noise from single-step denoising rather than directly supervising the fully denoised material output. This limitation restricts the model's ability to specialize in the decomposition task and impedes the application of more task-specific techniques.", + "bbox": [ + 91, + 265, + 483, + 627 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Structural Expert Branches. Our solution to the first problem is to introduce a structural expert branch architecture [29, 58]. Previous methods are limited to using a single diffusion model to estimate one material type. Although training separate models for different material maps is an intuitive approach, it introduces issues such as complex implementation, high overhead, and inconsistencies between materials. Considering that, despite their different representations, materials fundamentally reflect an understanding of object features, we incorporate structural expert branches into the UNet model, allowing a single diffusion model to predict multiple material maps. Specifically, we design two expert branches, one for predicting albedo, and the other for predicting RM. For each expert branch, we replicate the UNet's last UpBlock, and conv_out, while the rest of the UNet architecture remains shared between the branches. The modified model first extracts and transforms encoded image features through the shared modules. These features", + "bbox": [ + 89, + 628, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "are then separated into 2 copies by duplicating before the last UpBlock, with each branch ultimately predicting results for its respective domain. The expert branches specialize in material-specific features, while the shared modules focus on commonalities. In this way, we only add a small number of parameters (19,313,604 parameters, $2.23\\%$ of UNet's total), achieving the same performance as previous methods that would otherwise require twice the scale.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Single-Step Inference and End-to-End Training. To address the second problem, we analyze the DDIM scheduler of diffusion models. Previous works [17, 21, 27, 48] have highlighted a flaw in the implementation of the DDIM scheduler, which limits the potential of single-step inference in Stable Diffusion: In the default leading timestep setting of DDIM, for single-step prediction, the model is provided with a timestep $(t = 1)$ suggesting the input is essentially denoised, yet in reality, the input is pure noise. The discrepancy between the timestep and the actual noise level causes the single-step result to be flawed. To align the timestep with the noise level received by the model, we adjust the scheduler to use the trailing setting $(t = T$ for single-step), correcting this issue and enabling single-step inference, significantly reducing the time overhead.", + "bbox": [ + 511, + 212, + 903, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "More importantly, after this correction, we can fine-tune the one-step model end-to-end, effectively resolving the third issue. For a diffusion model with the uncorrected timestep setting, fine-tuning the diffusion model end-to-end is impractical, because generating a clean material map typically requires multiple denoising steps, leading to prohibitive computational complexity due to the gradient calculations. As an alternative, previous methods [6, 28, 56] rely on using the noise latent obtained from a single denoising step of the UNet for loss calculation. However, this approach shifts the model's training objective from material decomposition to general image denoising, which hinders specialization for this task. In contrast, our corrected inference pipeline can generate fully denoised material maps in a single denoising step, making end-to-end fine-tuning both feasible and efficient for material decomposition. Specifically, with the clean material estimations, we compute perceptual losses [23] between the predicted albedo map $k_{d}$ , the RM map $k_{rm}$ , and the ground truth maps $\\hat{k}_d, \\hat{k}_{rm}$ :", + "bbox": [ + 511, + 438, + 906, + 726 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {p} (\\hat {y}, y) = \\sum_ {j} \\frac {1}{C _ {j} H _ {j} W _ {j}} | | \\phi_ {j} (\\hat {y}) - \\phi_ {j} (y) | | _ {1}, \\tag {1} \\\\ \\mathcal {L} _ {m} = \\mathcal {L} _ {p} (\\hat {k} _ {d}, k _ {d}) + \\mathcal {L} _ {p} (\\hat {k} _ {r m}, k _ {r m}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 729, + 903, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\phi$ is a 16-layer VGG network, and $\\phi_{j}$ represents the output of the $j$ -th layer, with a shape of $C_j\\times H_j\\times W_j$ .", + "bbox": [ + 511, + 795, + 903, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that, following the approach of previous work [17], we use \"Zeros\" as the noise type during fine-tuning, meaning no random noise is added.", + "bbox": [ + 511, + 825, + 903, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Re-render Loss in DM Fine-tuning. Data-driven decomposition diffusion models typically focus on the distribu", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25086", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cb2e36bb43d648640679bd5d8db2f80b4d494d0fec520f9e542963e1af2b0453.jpg", + "image_caption": [ + "Figure 3. Material Decomposition for 3D objects. Applying SuperMat's powerful decomposition to 3D objects, we can obtain an object with high-quality materials in just three inference steps. Due to implementation differences, the texels in the RM map for image space decomposition are (1, roughness, metallic), while in the UV refinement process, they are (0, roughness, metallic)." + ], + "image_footnote": [], + "bbox": [ + 102, + 95, + 893, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tion of individual material maps, yet they often overlook the interactions between different types of material maps. In other words, previous methods compute the loss between the predicted values and the ground truth without utilizing material information for further operations or calculations, which could provide stronger guidance for the model. As a result, although the predicted materials may appear close to the ground truth, the rendered output may not align with expectations. This limitation is inherent to their training methods, as they either produce meaningless predicted noise, preventing the application of image-level loss, or fail to generate all material predictions simultaneously. In contrast, our design, which incorporates structural expert branches and single-step prediction, allows us to obtain all denoised material maps during the training iterations. Using this information, we introduce the re-render loss [10], which involves rendering the model in a new lighting environment and comparing the result to the ground truth. Specifically, at each training iteration, we obtain $k_{d}$ , $k_{rm} = (k_{r},k_{m})$ , along with the true normal map $\\hat{k}_n$ , position map $\\hat{k}_p$ , camera viewpoint $\\hat{c}$ , and an environment light map. Then we use this information and our rendering model $\\mathcal{R}$ to compute the predicted material's performance under new illumination. Similarly, with the groundtruth albedo $\\hat{k}_d$ and RM $\\hat{k}_{rm}$ , we compute the ground truth of relighting. The perceptual loss between these two renderings defines our re-render loss:", + "bbox": [ + 91, + 329, + 483, + 724 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e} = \\mathcal {L} _ {p} \\left(\\mathcal {R} _ {\\hat {k} _ {n}, \\hat {k} _ {p}, \\hat {c}} \\left(\\hat {k} _ {d}, \\hat {k} _ {m}, \\hat {k} _ {r}\\right), \\mathcal {R} _ {\\hat {k} _ {n}, \\hat {k} _ {p}, \\hat {c}} \\left(k _ {d}, k _ {m}, k _ {r}\\right)\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 741, + 482, + 775 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Extension to Multi-view Material Decomposition. We extend SuperMat to a multi-view model to enhance the consistency of generated results across different viewpoints. Specifically, we use a 3D self-attention layer and incorporate camera extrinsics as conditions, following MVDream [41]. The multi-view model, SuperMatMV, is capable of simultaneously decomposing materials for 6 input views, providing more accurate and consistent material predictions.", + "bbox": [ + 89, + 779, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. UV Refinement Network", + "text_level": 1, + "bbox": [ + 511, + 328, + 740, + 343 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We aim to apply SuperMat or SuperMatMV to decompose material for 3D objects with only RGB textures. Several previous works [5, 36, 45, 62] use techniques like SDS to bridge the gap between dimensions, but such methods are not applicable to our deterministic model, and suffer from slow computational speed and mode seeking issues. Therefore, we explore using backprojection to transfer 2D decomposition results from image space to UV space, enabling their application to 3D models. Notably, both SuperMat and SuperMatMV are finite-view models, meaning that one inference can only decompose the visible region from a specific set of viewpoints. To address this problem, we propose estimating materials from multiple viewpoints and integrating them. With multi-view decomposition results from SuperMatMV, the UV refinement network aims to paint missing areas and improve the quality of the generated materials.", + "bbox": [ + 511, + 349, + 906, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Backprojection and UV Blending. For each viewpoint, we backproject SuperMatMV's decomposition results from image space onto the model's UV texture map, obtaining the partial albedo UV map $C_d^i$ , RM UV map $C_{rm}^i$ , and mask UV map $M^i$ for viewpoint $i$ . We then blend the backprojection results from all viewpoints in an averaged manner.", + "bbox": [ + 511, + 592, + 908, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For albedo, this can be formulated as:", + "bbox": [ + 511, + 683, + 764, + 696 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM = \\sum_ {i} M ^ {i},\n$$\n", + "text_format": "latex", + "bbox": [ + 666, + 705, + 764, + 736 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC _ {d} = \\frac {\\sum_ {i} C _ {d} ^ {i}}{(M + \\epsilon)}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 655, + 732, + 903, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The RM follows the same computation method. The blended partial maps $C_d, C_{rm}$ , and the partial mask $M$ are then input into the UV refinement one-step model.", + "bbox": [ + 511, + 779, + 906, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "UV Refinement One-Step Model. Due to the limited number of viewpoints, the blended partial maps often have missing areas. To address this issue, we leverage diffusion models for UV refinement: we fine-tune Stable Diffusion into a UV refinement model, which takes a flawed partial material", + "bbox": [ + 511, + 825, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25087", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/568234a5f0e359a20b8679bc598c0708da8cb15529d3c08e5498c8ff3c45f51a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AlbedoMetallicRoughnessRelightingTime(s)↓
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Derender3D [46]20.970.86060.1821---------0.16
IIR [64]21.940.87090.155217.950.76190.291119.730.86390.159620.980.88840.11300.04
IID [25]22.560.89080.128217.380.76150.295620.930.87590.147622.450.90440.10831.45
RGB→X [56]22.300.88600.115415.360.75560.336020.400.84730.159121.510.88610.11323.32
IntrinsicAnything [6]20.940.87880.1520---------6.93
StableMaterial [28]23.440.90190.107620.290.81640.245921.010.88800.132722.560.90430.09790.53
StableMaterialMV [28]24.560.90910.094617.280.78860.282920.860.88380.128623.140.90900.09330.55
SuperMat w/o e2e24.260.89640.109520.790.84440.282620.810.86210.158723.900.90850.09643.09
SuperMat w/o re-render26.700.92360.090624.540.82590.226323.520.90160.136226.410.93780.07020.07
SuperMat27.680.92900.082025.480.86850.222124.250.90570.128127.660.94320.06120.07
SuperMatMV27.560.92890.083926.110.87490.219524.840.90760.120827.640.94310.06280.09
", + "bbox": [ + 91, + 88, + 903, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison of our method with others on the image space decomposition task. We highlight the best, second-best, and third-best results for each metric. \"SuperMat w/o e2e\" denotes our method without the scheduler fix and perceptual loss for end-to-end training. \"SuperMat w/o re-render\" denotes our method without the re-render loss. The first two methods are non-diffusion-based, while the middle five methods are diffusion-based.", + "bbox": [ + 89, + 301, + 906, + 362 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "UV map, a mask UV map, and a position UV map as inputs, and outputs completed and refined material UV maps. Specifically, the material maps are encoded by a VAE and concatenated with downsampled mask and position maps, and then processed by UNet with an expanded 8-channel conv_in. We also modify the model's scheduler to support single-step inference and enable end-to-end training, resulting in a deterministic one-step model.", + "bbox": [ + 88, + 369, + 480, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Material Decomposition Pipeline for 3D Objects. As shown in Fig. 3, with the fine-tuned UV refinement network, we propose a complete pipeline for the material decomposition of 3D objects. Given an object with only RGB texture, we first render it from six orthogonal viewpoints. The rendered results are then batch-processed by SuperMatMV for viewpoint-consistent material decomposition. These materials, including albedo and RM, are backprojected into UV space and blended. The blended partial maps are then refined by the UV refinement one-step model and the final material UV maps are generated. This process requires only three inference steps—one for SuperMatMV, one for refining albedo, and one for refining RM—allowing our pipeline to complete the material decomposition of a 3D object in just 3 seconds.", + "bbox": [ + 89, + 491, + 482, + 719 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 737, + 222, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Datasets. In this paper, we collect 3D objects from Objaverse [8] and Digital Twin Catalog (DTC) [13] datasets. For each object, we render material maps, normal maps, and position maps from 6 orthogonal viewpoints and generate PBR results under 16 different environment lighting conditions from the same viewpoints. We then filter high-quality data using methods such as DINO score considerations [35], resulting in a dataset of 75k objects that combines both synthetic and real-captured data for the im", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "age space decomposition task. Following the approach outlined in Sec. 3.2, we further refine this data for the UV refinement network by selecting cases where the missing area in blended partial UV maps is under $50\\%$ . This process results in a final dataset of 40k objects for training the refinement one-step model.", + "bbox": [ + 511, + 369, + 906, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Comparison of Image Space Decomposition. We randomly select 64 unseen items each from the Objaverse, DTC and BlenderVault [28] datasets. Objaverse and BlenderVault primarily contain synthetic data, whereas DTC consists of scanned objects, offering a source of real-world data. Our method is tested against 7 baseline approaches on these 192 objects. For each object, the model processes 18 inputs corresponding to 3 lighting conditions and 6 viewpoints, predicting 3 (or 1) types of materials and generating 54 (or 18) texture maps in total. All outputs are used to compute metrics, with the final averaged results shown in Tab.1. We use PSNR, SSIM, and LPIPS metrics to evaluate albedo, metallic, and roughness. Additionally, for methods that provide all material types, we compute metrics on the rendering under relighting conditions using the ground truth normals and positions and the predicted values. The results demonstrate that our approach significantly outperforms the SOTA in image space decomposition.", + "bbox": [ + 511, + 460, + 908, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Beyond image quality, we compare inference speeds by calculating the total time required for each method to process the test set, normalized by the dataset size and number of predicted material types. All tests are conducted with a batch size of 1. For consistency in speed comparison, IID, StableMaterial and StableMaterialMV, which typically average over 10 predictions per input, are modified to perform a single prediction. The \"Time(s)\" column in Tab.1 reflects the time required by each method to generate a single material prediction on an NVIDIA A100 GPU with a batch size of 1. Traditional, non-diffusion-based methods like Deren", + "bbox": [ + 511, + 734, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25088", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0963d2da006a4748a87353ac7caf8d3c9664f82da3a59190240f283dfe47a27e.jpg", + "image_caption": [ + "Figure 4. Comparison of our method with others on albedo estimation results. Each pair of rows presents results of the same object under identical lighting conditions from different viewpoints. This organizational format is also followed in other qualitative comparison figures." + ], + "image_footnote": [], + "bbox": [ + 99, + 85, + 897, + 439 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d419c02f86c1cdf2746513ce077a0742047cf8a2617fab8ba7d6b4ae44886b70.jpg", + "image_caption": [ + "Figure 5. Comparison of our method with others on metallic and roughness estimation results. We combined both materials into a single image, with the left half representing metallic and the right half showing roughness. It can be observed that these two materials exhibit more variation and are more challenging to estimate compared to albedo." + ], + "image_footnote": [], + "bbox": [ + 96, + 486, + 478, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "der3D and IIR generally show faster speeds. While Stable-Material attempts to predict multiple materials simultaneously by adding extra channels to conv_in and conv_out, its lack of structural design adjustments limits prediction quality and requires multi-step denoising, resulting in slower performance than traditional methods. In contrast, our ap-", + "bbox": [ + 89, + 809, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/caddc01cd0d9a5160dd7e6da61d7796d71564ca1c7759cf110261fb014f262f4.jpg", + "image_caption": [ + "Figure 6. Results of UV refinement one-step model." + ], + "image_footnote": [], + "bbox": [ + 521, + 489, + 898, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "proach achieves high-quality material decomposition with speeds comparable to traditional methods.", + "bbox": [ + 511, + 626, + 906, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Comparison of Image Space Decomposition. Fig.4 compares albedo prediction results across various cases, including artist-designed objects and real-world scans, showcasing the performance of our method against other approaches. SuperMat demonstrates superior ability in removing highlights and shadows, accurately capturing the intrinsic color, even on challenging metallic surfaces. In Fig.5, we present metallic and roughness predictions from our method alongside other baselines. SuperMat outperforms the others with well-defined boundaries between metallic and non-metallic regions and achieves consistency within material types while preserving high-frequency details. Moreover, our results closely align with physical realism, addressing a significant challenge that previous methods struggled to overcome. Additionally, SuperMatMV demonstrates stronger consistency across different views in", + "bbox": [ + 511, + 659, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25089", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fd0c9c4644be9708947a63a51562392785b83d85421a2b8aa104b997b5c8b44c.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 158, + 88, + 200, + 226 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d2308a59b1d9f190207f17567be32b59dae11d12adef051fb6644b7ebdcdd361.jpg", + "image_caption": [ + "w/o e2e" + ], + "image_footnote": [], + "bbox": [ + 245, + 90, + 292, + 228 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6f042577cea02fb38ef6cc9537426081f79117bc0eb14f3dd3d70a16e104bb1d.jpg", + "image_caption": [ + "w/o re-render" + ], + "image_footnote": [], + "bbox": [ + 341, + 90, + 380, + 228 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f0c0c57f66b0181e0f98609554254a5736e65905d9188bdd943006a041b16f50.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 431, + 90, + 472, + 228 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ef1c30d76a7fd70142c7902df3661335c4a83069405c13af5f8ceb935bf0b972.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 104, + 571, + 138 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fb285aa4f44b71d4662d46cb426871bcb341bc6eeb1b940ed60909d5e272a19a.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 516, + 155, + 571, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cc0e17a7dd2eff3316323ec4c6508f27f13746481cdea13483886c0d66a32c59.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 104, + 661, + 140 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0d9b08838ef8b7a4e7db3be82416be5c7c6812124bb5081eb7582b511f33463e.jpg", + "image_caption": [ + "w/o e2e" + ], + "image_footnote": [], + "bbox": [ + 607, + 156, + 661, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ff48193d0efa128b7fb9c487c6eec94a95ab6f4f6b005bfd984519935f9f1e4b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 104, + 753, + 140 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/53b82535cb740fe0fc1db1ebfd0d127e1baee0b48100cd8087bb6ea25f8fe0c8.jpg", + "image_caption": [ + "w/o re-render" + ], + "image_footnote": [], + "bbox": [ + 697, + 156, + 753, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6d3802c777215d8f231e69f4263da890d69692d7f3b55114ea4c5cf21d979e18.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 104, + 844, + 140 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/caf03c59044843896edd2ffb2f1fa8832cfc1d53bb286d40b2f2a4de7946e2e2.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 790, + 156, + 844, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f5715ed0b3fff20980d02c21c468423010f1713de89020dda29cdeb07c73b447.jpg", + "image_caption": [ + "Figure 8. Results of decomposition for 3D objects. We compare the original and decomposed models under the same illumination and show the relighting outcomes of the new models." + ], + "image_footnote": [], + "bbox": [ + 99, + 286, + 473, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "all material predictions, and is still able to generate reasonable decomposition results for inputs from challenging viewpoints with sparse relevant information. More results will be presented in the supplementary materials.", + "bbox": [ + 89, + 523, + 482, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results of UV Refinement and Material Estimation for 3D Objects. Fig.6 illustrates the effectiveness of the UV refinement one-step model in inpainting missing areas in SuperMatMV's outputs. Fig.8 displays the results from our material estimation pipeline for 3D models. The first column shows the rendering of the input models with only RGB texture, where all surfaces exhibit uniform reflectance behavior. After applying our method, the decomposed materials enable physically-based rendering with surface-specific reflectance properties, as demonstrated in the second and third columns under different lighting conditions.", + "bbox": [ + 89, + 583, + 482, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Studies. In Tab.1, we present ablation studies on two core components of SuperMat: end-to-end (e2e) training and re-render loss. E2E with perceptual loss significantly enhances model performance and efficiency. In fact, within the e2e framework, we experiment with three types of loss functions, with perceptual loss achieving the best results. This suggests that direct supervision of the fully denoised predicted material maps, rather than latent noise, plays a critical role in enhancing the model's capability for this specific task. Detailed findings are available in the sup", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/614d9a813e56f4e0d3c6d24c235948a90a895d291f6f4aa49372283bf7abbfcd.jpg", + "image_caption": [ + "o o o o o o o o o o o o o o o o o o o o o o o" + ], + "image_footnote": [], + "bbox": [ + 565, + 287, + 635, + 354 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/deb5622d2540657eeef09443c3ddd2713601455c6859fbb7f739a8712cd72c33.jpg", + "image_caption": [ + "Ours", + "Figure 9. Qualitative ablation results on UV refinement network." + ], + "image_footnote": [], + "bbox": [ + 565, + 354, + 635, + 420 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9becd74f3dc913dfc3d72d1700e5a0c839dae5b197099b2ba294a64528efc3f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 686, + 287, + 792, + 348 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/79f712c9a1557e8195196cea1bf8c045f21cc6f4bf932f72410d450b70007405.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 356, + 792, + 417 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/113c94b69b5af2681749b75397c16c24804295d26a35eef8c8de5c16c324e86e.jpg", + "image_caption": [ + "Figure 7. Qualitative ablation results for SuperMat. The first row shows albedo, while the second row shows metallic and roughness." + ], + "image_footnote": [], + "bbox": [ + 839, + 285, + 874, + 354 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8c64046ca37abec96a70c8f7159776ae951810d7b594daa63ceea5abc1dfc8fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 841, + 354, + 875, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "plementary materials. The addition of re-render loss, facilitated by the e2e framework, further improves prediction quality, particularly under relighting conditions. Qualitative results of these ablations are shown in Fig. 7. The 'w/o e2e' setting, where the model is trained without perceptual and re-render losses, limits its capacity to remove lighting effects effectively. In contrast, training 'w/o re-render' loss leads to minor inaccuracies in fine material details.", + "bbox": [ + 511, + 454, + 903, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For the decomposition pipeline for 3D objects, we conduct additional ablations on the UV refinement network. Due to limitations from only six orthogonal views, achieving complete texture coverage, accurate merging of overlapping regions, and consistent material transitions are challenging. As shown in Fig.9, omitting the UV refinement leads to incomplete textures, visible color-separation lines, and unnatural transitions between materials, underscoring the importance of this component.", + "bbox": [ + 511, + 575, + 905, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 724, + 640, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present SuperMat, an innovative image space material decomposition one-step model finetuned from Stable Diffusion with task-specific losses in an end-to-end framework. SuperMat enables simultaneous multi-material prediction and supports single-step inference, achieving state-of-the-art quality and outperforming previous diffusion-based methods in efficiency. Through the proposed UV refinement network, we extend SuperMat to decomposition tasks for 3D objects, where it demonstrates exceptional performance and applicability.", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25090", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 217, + 187, + 233 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Raphael Bensadoun, Yanir Kleiman, Idan Azuri, Omri Harosh, Andrea Vedaldi, Natalia Neverova, and Oran Gafni. Meta 3d texturegen: Fast and consistent texture generation for 3d objects. arXiv preprint arXiv:2407.02430, 2024. 3", + "[2] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan Barron, and Hendrik Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. Advances in Neural Information Processing Systems, 34: 10691-10704, 2021. 2", + "[3] Mark Boss, Zixuan Huang, Aaryaman Vasishta, and Varun Jampani. Sf3d: Stable fast 3d mesh reconstruction with uv-unwrapping and illumination disentanglement. arXiv preprint arXiv:2408.00653, 2024. 3", + "[4] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18558-18568, 2023. 3", + "[5] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 3, 5", + "[6] Xi Chen, Sida Peng, Dongchen Yang, Yuan Liu, Bowen Pan, Chengfei Lv, and Xiaowei Zhou. Intrinsicanything: Learning diffusion priors for inverse rendering under unknown illumination. arXiv preprint arXiv:2404.11593, 2024. 1, 2, 3, 4, 6", + "[7] Ziang Cheng, Junxuan Li, and Hongdong Li. Wildlight: In-the-wild inverse rendering with a flashlight. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4305-4314, 2023. 2", + "[8] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 6", + "[9] Kangle Deng, Timothy Omernick, Alexander Weiss, Deva Ramanan, Jun-Yan Zhu, Tinghui Zhou, and Maneesh Agrawala. *Flashtex: Fast relightable mesh texturing with lightcontrolnet*. In European Conference on Computer Vision, pages 90–107. Springer, 2024. 2", + "[10] Valentin Deschaintre, Miika Aittala, Fredo Durand, George Drettakis, and Adrien Bousseau. Single-image svbrdf capture with a rendering-aware deep network. ACM Transactions on Graphics (ToG), 37(4):1-15, 2018. 5" + ], + "bbox": [ + 93, + 243, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "This work was supported by National Natural Science Foundation of China (72192821, 62472282, 62302297 and 62302296), Young Elite Scientists Sponsorship Program by CAST (2022QNRC001), the Fundamental Research Funds for the Central Universities (YG2023QNA35) and YuCaiKe [2023] (231111310300).", + "[11] Yiqun Duan, Xianda Guo, and Zheng Zhu. Diffusiondepth: Diffusion denoising approach for monocular depth estimation. arXiv preprint arXiv:2303.05021, 3, 2023. 3", + "[12] Xiang Feng, Chang Yu, Zoubin Bi, Yintong Shang, Feng Gao, Hongzhi Wu, Kun Zhou, Chenfanfu Jiang, and Yin Yang. Arm: Appearance reconstruction model for relightable 3d generation. arXiv preprint arXiv:2411.10825, 2024. 3", + "[13] James Fort. Introducing the digital twin catalog from reality labs research, 2024. https://ai.meta.com/blog/digital-twin-catalog-3d-reconstruction-shopify-reality-labs-research/, Last accessed on 2025-03-07. 6", + "[14] Xiao Fu, Wei Yin, Mu Hu, Kaixuan Wang, Yuexin Ma, Ping Tan, Shaojie Shen, Dahua Lin, and Xiaoxiao Long. Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image. In European Conference on Computer Vision, pages 241-258. Springer, 2025. 3", + "[15] Jian Gao, Chun Gu, Youtian Lin, Hao Zhu, Xun Cao, Li Zhang, and Yao Yao. Relightable 3d gaussian: Real-time point cloud relighting with brdf decomposition and ray tracing. arXiv preprint arXiv:2311.16043, 2023. 2", + "[16] Lin Gao, Tong Wu, Yu-Jie Yuan, Ming-Xian Lin, Yu-Kun Lai, and Hao Zhang. Tm-net: Deep generative networks for textured meshes. ACM Transactions on Graphics (TOG), 40 (6):1-15, 2021. 3", + "[17] Gonzalo Martin Garcia, Karim Abou Zeid, Christian Schmidt, Daan de Geus, Alexander Hermans, and Bastian Leibe. Fine-tuning image-conditional diffusion models is easier than you think. arXiv preprint arXiv:2409.11355, 2024. 3, 4", + "[18] Ming Gui, Johannes S Fischer, Ulrich Prestel, Pingchuan Ma, Dmytro Kotovenko, Olga Grebenkova, Stefan Andreas Baumann, Vincent Tao Hu, and Björn Ommer. Depthfm: Fast monocular depth estimation with flow matching. arXiv preprint arXiv:2403.13788, 2024. 3", + "[19] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18409-18418, 2022. 2", + "[20] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light, and material decomposition from images using monte carlo rendering and denoising. Advances in Neural Information Processing Systems, 35:22856-22869, 2022. 2", + "[21] Jing He, Haodong Li, Wei Yin, Yixun Liang, Leheng Li, Kaiqiang Zhou, Hongbo Zhang, Bingbing Liu, and YingCong Chen. Lotus: Diffusion-based visual foundation model for high-quality dense prediction. arXiv preprint arXiv:2409.18124, 2024. 3, 4", + "[22] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. Tensoir: Tensorial inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 165-174, 2023. 2", + "[23] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution." + ], + "bbox": [ + 93, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25091", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 4", + "[24] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492-9502, 2024. 3", + "[25] Peter Kocsis, Vincent Sitzmann, and Matthias Nießner. Intrinsic image diffusion for indoor single-view material estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5198-5208, 2024. 2, 3, 6", + "[26] Zhihao Liang, Qi Zhang, Ying Feng, Ying Shan, and Kui Jia. Gs-ir: 3d gaussian splatting for inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21644–21653, 2024. 2", + "[27] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 5404-5411, 2024. 3, 4", + "[28] Yehonathan Litman, Or Patashnik, Kangle Deng, Aviral Agrawal, Rushikesh Zawar, Fernando De la Torre, and Shubham Tulsiani. Materialfusion: Enhancing inverse rendering with material diffusion priors. arXiv preprint arXiv:2409.15273, 2024. 1, 2, 3, 4, 6", + "[29] Xian Liu, Jian Ren, Aliaksandr Siarohin, Ivan Skorokhodov, Yanyu Li, Dahua Lin, Xihui Liu, Ziwei Liu, and Sergey Tulyakov. Hyperhuman: Hyper-realistic human generation with latent structural diffusion. arXiv preprint arXiv:2310.08579, 2023. 3, 4", + "[30] Yunfei Liu, Yu Li, Shaodi You, and Feng Lu. Unsupervised learning for intrinsic image decomposition from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3248-3257, 2020. 3", + "[31] Yuxin Liu, Minshan Xie, Hanyuan Liu, and Tien-Tsin Wong. Text-guided texturing by synchronized multi-view diffusion. arXiv preprint arXiv:2311.12891, 2023. 3", + "[32] Zexiang Liu, Yangguang Li, Youtian Lin, Xin Yu, Sida Peng, Yan-Pei Cao, Xiaojuan Qi, Xiaoshui Huang, Ding Liang, and Wanli Ouyang. Unidream: Unifying diffusion priors for relightable text-to-3d generation. In European Conference on Computer Vision, pages 74–91. Springer, 2024. 2", + "[33] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022. 3", + "[34] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8280-8290, 2022. 2", + "[35] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez," + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6", + "[36] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3, 5", + "[37] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng Bo, and Xiaoguang Han. Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9914–9925, 2024. 2, 3", + "[38] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In ACM SIGGRAPH 2023 conference proceedings, pages 1-11, 2023. 3", + "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 4", + "[40] Saurabh Saxena, Abhishek Kar, Mohammad Norouzi, and David J Fleet. Monocular depth estimation using diffusion models. arXiv preprint arXiv:2302.14816, 2023. 3", + "[41] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 5", + "[42] Zeqi Shi, Xiangyu Lin, and Ying Song. An attention-embedded gan for svbrdf recovery from a single image. Computational Visual Media, 9(3):551-561, 2023. 3", + "[43] Giuseppe Vecchio, Rosalie Martin, Arthur Roullier, Adrien Kaiser, Romain Rouffet, Valentin Deschaintre, and Tamy Boubekeur. Controlmat: a controlled generative approach to material capture. ACM Transactions on Graphics, 43(5): 1-17, 2024.", + "[44] Yitong Wang, Xudong Xu, Li Ma, Haoran Wang, and Bo Dai. Boosting 3d object generation through pbr materials. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 3", + "[45] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 3, 5", + "[46] Felix Wimbauer, Shangzhe Wu, and Christian Rupprecht. De-rendering 3d objects in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18490-18499, 2022. 2, 3, 6", + "[47] Bojun Xiong, Jialun Liu, Jiakui Hu, Chenming Wu, Jinbo Wu, Xing Liu, Chen Zhao, Errui Ding, and Zhouhui Lian. Texgaussian: Generating high-quality pbr material via octree-based 3d gaussian splatting. arXiv preprint arXiv:2411.19654, 2024. 2", + "[48] Guangkai Xu, Yongtao Ge, Mingyu Liu, Chengxiang Fan, Kangyang Xie, Zhiyue Zhao, Hao Chen, and Chunhua Shen. Diffusion models trained with large data are transferable visual models. arXiv preprint arXiv:2403.06090, 2024. 4" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25092", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Xudong Xu, Zhaoyang Lyu, Xingang Pan, and Bo Dai. Matlab: Material-aware text-to-3d via latent brdf auto-encoder. arXiv preprint arXiv:2308.09278, 2023. 2", + "[50] Chongjie Ye, Lingteng Qiu, Xiaodong Gu, Qi Zuo, Yushuang Wu, Zilong Dong, Liefeng Bo, Yuliang Xiu, and Xiaoguang Han. Stablenormal: Reducing diffusion variance for stable and sharp normal. arXiv preprint arXiv:2406.16864, 2024. 3", + "[51] Renjiao Yi, Chenyang Zhu, and Kai Xu. Weakly-supervised single-view image relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8402-8411, 2023. 3", + "[52] Kim Youwang, Tae-Hyun Oh, and Gerard Pons-Moll. Paint-it: Text-to-texture synthesis via deep convolutional texture map optimization and physically-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4347-4356, 2024. 2, 3", + "[53] Xin Yu, Peng Dai, Wenbo Li, Lan Ma, Zhengzhe Liu, and Xiaojuan Qi. Texture generation on 3d meshes with point-uv diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4206–4216, 2023. 2", + "[54] Xin Yu, Ze Yuan, Yuan-Chen Guo, Ying-Tian Liu, Jianhui Liu, Yangguang Li, Yan-Pei Cao, Ding Liang, and Xiaojuan Qi. Texgen: a generative diffusion model for mesh textures. ACM Transactions on Graphics (TOG), 43(6):1-14, 2024.", + "[55] Xianfang Zeng, Xin Chen, Zhongqi Qi, Wen Liu, Zibo Zhao, Zhibin Wang, Bin Fu, Yong Liu, and Gang Yu. Paint3d: Paint anything 3d with lighting-less texture diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4252-4262, 2024. 3", + "[56] Zheng Zeng, Valentin Deschaintre, Iliyan Georgiev, Yannick Hold-Geoffroy, Yiwei Hu, Fujun Luan, Ling-Qi Yan, and Miloš Hasan. Rgb ↔ x: Image decomposition and synthesis using material-and lighting-aware diffusion models. In ACM SIGGRAPH 2024 Conference Papers, pages 1–11, 2024. 1, 2, 3, 4, 6", + "[57] Kai Zhang, Fujun Luan, Zhengqi Li, and Noah Snavely. Iron: Inverse rendering by optimizing neural sdfs and materials from photometric images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5565-5574, 2022. 2", + "[58] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1-20, 2024. 4", + "[59] Qian Zhang, Vikas Thamizharasan, and James Tompkin. Learning physically based material and lighting decompositions for face editing. Computational Visual Media, 10(2): 295-308, 2024. 2", + "[60] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Ner-factor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG), 40(6):1-18, 2021.", + "[61] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18643-18652, 2022. 2", + "[62] Yuqing Zhang, Yuan Liu, Zhiyu Xie, Lei Yang, Zhongyuan Liu, Mengzhou Yang, Runze Zhang, Qilong Kou, Cheng Lin, Wenping Wang, et al. Dreammat: High-quality pbr material generation with geometry-and light-aware diffusion models. ACM Transactions on Graphics (TOG), 43(4):1-18, 2024. 2, 3, 5", + "[63] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5729-5739, 2023. 3", + "[64] Jingsen Zhu, Fujun Luan, Yuchi Huo, Zihao Lin, Zhihua Zhong, Dianbing Xi, Rui Wang, Hujun Bao, Jiaxiang Zheng, and Rui Tang. Learning-based inverse rendering of complex indoor scenes with differentiable monte carlo raytracing. In SIGGRAPH Asia 2022 Conference Papers, pages 1-8, 2022. 2, 3, 6" + ], + "bbox": [ + 516, + 92, + 903, + 371 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "25093", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SuperMat_ Physically Consistent PBR Material Estimation at Interactive Rates/4daa8d06-6978-409c-9337-f3e4e925bd23_model.json b/2025/SuperMat_ Physically Consistent PBR Material Estimation at Interactive Rates/4daa8d06-6978-409c-9337-f3e4e925bd23_model.json new file mode 100644 index 0000000000000000000000000000000000000000..39ed3fa60b102859befe45a013d83bf3e5ccf773 --- /dev/null +++ b/2025/SuperMat_ Physically Consistent PBR Material Estimation at Interactive Rates/4daa8d06-6978-409c-9337-f3e4e925bd23_model.json @@ -0,0 +1,2389 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.099, + 0.131, + 0.9, + 0.153 + ], + "angle": 0, + "content": "SuperMat: Physically Consistent PBR Material Estimation at Interactive Rates" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.18, + 0.888, + 0.233 + ], + "angle": 0, + "content": "Yijia Hong\\(^{1,2}\\), Yuan-Chen Guo\\(^{4*}\\), Ran Yi\\(^{1}\\), Yulong Chen\\(^{3}\\), Yan-Pei Cao\\(^{4\\dagger}\\), Lizhuang Ma\\(^{1\\dagger}\\) \n\\(^{1}\\)Shanghai Jiao Tong University \\(^{2}\\)Shanghai Innovation Institute \\(^{3}\\)Harbin Institute of Technology \\(^{4}\\)VAST" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.257, + 0.174, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.349, + 0.174, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.441, + 0.174, + 0.518 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.201, + 0.259, + 0.321, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.2, + 0.349, + 0.327, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.2, + 0.441, + 0.327, + 0.518 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.349, + 0.261, + 0.416, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.349, + 0.349, + 0.416, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.349, + 0.441, + 0.416, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.261, + 0.641, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.346, + 0.641, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.431, + 0.641, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.259, + 0.895, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.535, + 0.907, + 0.606 + ], + "angle": 0, + "content": "Figure 1. We present an efficient method for high-quality material decomposition in both 2D and 3D. Left: Albedo and RM(Roughness, Metallic) predictions of SuperMat on rendered images under unknown illumination. The first row shows inputs with uneven lighting, the second row displays albedo predictions, and the third row presents RM predictions. Each texel in the RM map here is represented as (1, roughness, metallic). Middle: Rendering results of a 3D object with PBR materials obtained by our method under novel lighting. Right: Our method is efficient and produces high-quality results." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.618, + 0.327, + 0.634 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.65, + 0.486, + 0.862 + ], + "angle": 0, + "content": "Decomposing physically-based materials from images into their constituent properties remains challenging, particularly when maintaining both computational efficiency and physical consistency. While recent diffusion-based approaches have shown promise, they face substantial computational overhead due to multiple denoising steps and separate models for different material properties. We present SuperMat, a single-step framework that achieves high-quality material decomposition with one-step inference. This enables end-to-end training with perceptual and re-render losses while decomposing albedo, metallic, and roughness maps at millisecond-scale speeds. We further extend our framework to 3D objects through a UV refinement network, enabling consistent material estimation across viewpoints" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.619, + 0.907, + 0.726 + ], + "angle": 0, + "content": "while maintaining efficiency. Experiments demonstrate that SuperMat achieves state-of-the-art PBR material decomposition quality while reducing inference time from seconds to milliseconds per image, and completes PBR material estimation for 3D objects in approximately 3 seconds. The project page is at https://hyj542682306.github.io/SuperMat/." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.754, + 0.646, + 0.77 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.78, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Decomposing Physically-Based Rendering (PBR) materials from images into their constituent properties (e.g., albedo, metallic, and roughness maps) remains a central challenge in computer vision and graphics. This task is essential for creating high-fidelity 3D assets across numerous applications, from film production and game development to virtual and augmented reality pipelines. While recent advances have shown promising results [6, 28, 56], the problem re" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.183, + 0.888 + ], + "angle": 0, + "content": "*Project lead" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.234, + 0.901 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.234, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25083" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "mains challenging due to its ill-posed nature, particularly given unknown lighting conditions and complex interactions between different material properties. Our work addresses two key objectives: 1) efficient decomposition of high-fidelity material maps from 2D images, and 2) extension of this capability to 3D objects, enabling accurate material estimation across complex geometries." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.202, + 0.487, + 0.565 + ], + "angle": 0, + "content": "Previous works in PBR material estimation generally fall into two main categories: optimization-based generation for 3D objects and data-driven decomposition for 2D images. Optimization-based methods [15, 19, 22, 26] employ differentiable rendering to reconstruct 3D objects with material parameters by minimizing the difference between rendered and target images. Instead, data-driven approaches learn the distribution of various materials from large datasets to perform intrinsic decomposition on 2D images, which is the primary focus of our work. Early methods [46, 64] in this domain fail to produce stable and reliable results due to limited data availability and model capacity. Recent works [6, 25, 28, 56] leverage large diffusion models and reframe material decomposition as a conditional image generation task. However, these methods tend to focus solely on individual material properties, overlooking the interactions between different types of materials and the combined impact on rendered appearances. Even minor inaccuracies in individual material estimates can propagate through the rendering pipeline, leading to substantial visual artifacts. Moreover, these diffusion-based methods face considerable computational overhead - they demand multiple denoising steps (30-50) during inference, causing a critical bottleneck for real-time applications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.487, + 0.904 + ], + "angle": 0, + "content": "To address these challenges, we propose a deterministic one-step framework that achieves both efficient and high-quality material decomposition. First, we introduce and refine structural expert branches within the UNet architecture. These specialized branches enable concurrent processing of multiple material properties while maintaining domain-specific expertise, allowing our model to decompose albedo, metallic, and roughness maps simultaneously with minimal computational overhead. This design facilitates collaborative and physically consistent material prediction while avoiding the linear cost scaling typically required when estimating multiple material attributes. Second, we address the computational inefficiency of traditional diffusion models by reformulating the DDIM sampling process. Through modification of the scheduler, we achieve single-step inference without compromising quality, dramatically reducing inference time. Most importantly, this single-step formulation enables end-to-end training of the diffusion model, allowing us to compute the perceptual and re-render loss on final predicted material textures, which enhances shadow removal and improves the joint performance of decomposed materials under varying lighting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.152 + ], + "angle": 0, + "content": "conditions. These advances lead to SuperMat, a highly efficient image space material decomposition model that estimates high-fidelity albedo, metallic, and roughness maps at millisecond-scale speeds." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.909, + 0.319 + ], + "angle": 0, + "content": "For practical 3D applications, we further extend SuperMat's capabilities to handle 3D objects. We upgrade SuperMat to a multi-view model, SuperMatMV, and design a UV refinement network that processes SuperMatMV's predictions in UV space. This approach addresses the challenge of maintaining consistency and completeness across different, limited viewpoints while preserving the efficiency of our original 2D approach. Our pipeline processes a textured 3D model in approximately 3 seconds, producing high-quality material maps that maintain consistency across arbitrary viewpoints and lighting conditions." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.319, + 0.766, + 0.332 + ], + "angle": 0, + "content": "In summary, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.335, + 0.905, + 0.44 + ], + "angle": 0, + "content": "- We propose SuperMat, a novel image space material decomposition model that generates physically consistent albedo, metallic, and roughness maps simultaneously with a single-step inference process. Our method achieves state-of-the-art performance on diverse datasets, significantly outperforming previous diffusion-based works in speed." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.441, + 0.905, + 0.531 + ], + "angle": 0, + "content": "- With the help of structural expert branches and single-step inference, we develop an end-to-end image space material decomposition training system that generates clean predictions for all material properties in each iteration, allowing for the integration of advanced techniques such as perceptual loss and re-render loss." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.532, + 0.905, + 0.607 + ], + "angle": 0, + "content": "- We propose an efficient PBR material decomposition method for 3D models. Combining SuperMatMV and our UV refinement network, this pipeline generates high-quality PBR materials for 3D objects from baked textures in approximately 3 seconds." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.335, + 0.905, + 0.607 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.619, + 0.665, + 0.635 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Material Generation and Decomposition. Recognizing the importance of material information in 3D assets, researchers have increasingly focused on material acquisition for objects and scenes. On one hand, some 3D generation approaches [5, 9, 32, 37, 47, 49, 52, 53, 62] integrate material optimization into the appearance creation process, concentrating on generating PBR maps for textured models. On the other hand, inverse rendering methods aim to reconstruct material-rich 3D representations from single or multi-view images. Some of these methods [7, 19, 57] assume a predefined lighting model to enhance estimation accuracy, but this constraint limits their practical applicability. To address this, other works [2, 15, 20, 22, 26, 34, 59-61] attempt to recover both material properties and unknown illumination simultaneously. As reliance on predefined assumptions decreases, researchers have found that beyond controlling material decomposition through rendered out" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25084" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.094, + 0.889, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.316, + 0.908, + 0.388 + ], + "angle": 0, + "content": "Figure 2. Overview of the SuperMat framework. The UNet architecture incorporates structural expert branches for albedo and RM estimation, enabling parallel material property prediction while sharing a common backbone. During training, we leverage a fixed DDIM scheduler and optimize the network end-to-end using both perceptual loss and re-render loss. The marked region shows the inference pipeline, where our model achieves single-step inference for concurrent material map generation. Input geometric information (normal, position) and environment maps are used to compute the re-render loss." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.396, + 0.482, + 0.517 + ], + "angle": 0, + "content": "puts, directly obtaining image space material priors is also feasible. Building on this idea, similar to ours, another body of research [6, 25, 28, 30, 42-44, 46, 51, 56, 64] focuses on material decomposition from 2D images captured or rendered under unknown lighting conditions. However, existing methods still face challenges related to training data, model architecture, and training strategies, leaving significant room for improvement in decomposition quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.483, + 0.746 + ], + "angle": 0, + "content": "Diffusion Models. Recently, researchers have applied diffusion models to a range of tasks [5, 36, 37, 45, 62], demonstrating the effectiveness of casting outcome prediction as a conditional image generation process. By leveraging prior knowledge from large pre-trained models, recent studies [6, 11, 14, 18, 24, 28, 40, 50, 63] have achieved performance improvements over previous methods in tasks like depth, normal, and material estimation. However, incorporating large models introduces efficiency challenges, particularly high memory requirements and the time-intensive, multistep denoising process needed for inference. Inspired by insights from recent work [17, 21, 27, 29], our approach addresses these issues by significantly reducing memory and inference time, improving the feasibility of applying these models in real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Generation in UV Space. UV maps are advantageous for fully encapsulating the appearance information of 3D models, leading several works to explore 3D object manipulation directly within the UV space. By representing the texture of an object as a learnable image, these methods [3, 4, 16, 33, 38, 55] enable texture generation, inpainting, and editing, providing new avenues for 3D tasks and allowing 2D image-space techniques and insights to be applied to 3D appearances. Building on this approach, recent methods [1, 12, 31, 52-55] have applied diffusion models to refine" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.395, + 0.907, + 0.442 + ], + "angle": 0, + "content": "UV maps with promising results. In this spirit, to extend SuperMat's capability to 3D, we fine-tune a UV-space refinement diffusion model specifically." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.453, + 0.606, + 0.469 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.479, + 0.907, + 0.81 + ], + "angle": 0, + "content": "In this paper, we propose SuperMat, an image space material decomposition model fine-tuned from Stable Diffusion that efficiently generates high-quality albedo, metallic, and roughness maps simultaneously. Given a rendered or captured image of an object under an unknown lighting condition from a specific viewpoint, SuperMat generates the corresponding material maps for the model from that same viewpoint. As shown in Figure 2, SuperMat is a 2D imaged-to-image one-step model with structural expert branches that simultaneously generates both albedo and RM maps, supporting single-step inference. Compared to previous diffusion-based methods, this model produces superior results in less time and with reduced memory usage. Our single-step formulation allows direct generation of material maps during training, in contrast to prior diffusion methods that predict noise \\((\\epsilon)\\) or velocity (v) terms. This enables end-to-end optimization with both perceptual loss on the generated maps and re-render loss from novel lighting conditions, leading to improved material quality and rendering fidelity. Building upon the single-view model, we further introduce a multi-view version, SuperMatMV, to ensure consistency in predictions across different viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "While SuperMat and SuperMatMV achieve efficient image space material decomposition, extending this capability to 3D objects presents additional challenges. Instead of relying on optimization-based approaches like score distillation sampling (SDS) that are computationally intensive and prone to mode collapse, we propose an efficient two-stage" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25085" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.213 + ], + "angle": 0, + "content": "pipeline for material estimation on 3D objects. Given a 3D model with only RGB texture, we first apply SuperMatMV to decompose materials from multiple viewpoints. We then introduce a UV refinement network that consolidates these view-dependent predictions into complete, high-quality UV texture maps through single-step inference. As shown in Figure 3, this pipeline enables PBR material estimation for 3D objects in approximately 3 seconds per model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.226, + 0.482, + 0.256 + ], + "angle": 0, + "content": "3.1. SuperMat - Image Space Material Decomposition" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.266, + 0.485, + 0.628 + ], + "angle": 0, + "content": "Previous works [6, 28, 56] have demonstrated that casting material estimation as a conditional latent diffusion process, leveraging knowledge from large pre-trained diffusion models like Stable Diffusion [39], is effective. However, for our task that requires the simultaneous generation of multiple material maps, three key issues remain: 1) High model overhead: Fine-tuning Stable Diffusion with its original architecture can only estimate a single material type. Even when combining metallic and roughness into an RM map, treating it as a single material still requires two diffusion models, one for albedo and another for RM, to meet the rendering model's requirements. This results in doubled training time and increased inference space overhead. 2) Slow inference speed: With the widely used DDIM scheduler, diffusion models require 30 to 50 denoising steps to produce satisfactory results, which can be inefficient and time-consuming. 3) Insufficient decomposition effect: Due to the high computational cost of backpropagation in multistep inference, prior diffusion-based methods trained on predicted noise from single-step denoising rather than directly supervising the fully denoised material output. This limitation restricts the model's ability to specialize in the decomposition task and impedes the application of more task-specific techniques." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Structural Expert Branches. Our solution to the first problem is to introduce a structural expert branch architecture [29, 58]. Previous methods are limited to using a single diffusion model to estimate one material type. Although training separate models for different material maps is an intuitive approach, it introduces issues such as complex implementation, high overhead, and inconsistencies between materials. Considering that, despite their different representations, materials fundamentally reflect an understanding of object features, we incorporate structural expert branches into the UNet model, allowing a single diffusion model to predict multiple material maps. Specifically, we design two expert branches, one for predicting albedo, and the other for predicting RM. For each expert branch, we replicate the UNet's last UpBlock, and conv_out, while the rest of the UNet architecture remains shared between the branches. The modified model first extracts and transforms encoded image features through the shared modules. These features" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "are then separated into 2 copies by duplicating before the last UpBlock, with each branch ultimately predicting results for its respective domain. The expert branches specialize in material-specific features, while the shared modules focus on commonalities. In this way, we only add a small number of parameters (19,313,604 parameters, \\(2.23\\%\\) of UNet's total), achieving the same performance as previous methods that would otherwise require twice the scale." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.905, + 0.438 + ], + "angle": 0, + "content": "Single-Step Inference and End-to-End Training. To address the second problem, we analyze the DDIM scheduler of diffusion models. Previous works [17, 21, 27, 48] have highlighted a flaw in the implementation of the DDIM scheduler, which limits the potential of single-step inference in Stable Diffusion: In the default leading timestep setting of DDIM, for single-step prediction, the model is provided with a timestep \\((t = 1)\\) suggesting the input is essentially denoised, yet in reality, the input is pure noise. The discrepancy between the timestep and the actual noise level causes the single-step result to be flawed. To align the timestep with the noise level received by the model, we adjust the scheduler to use the trailing setting \\((t = T\\) for single-step), correcting this issue and enabling single-step inference, significantly reducing the time overhead." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.439, + 0.907, + 0.727 + ], + "angle": 0, + "content": "More importantly, after this correction, we can fine-tune the one-step model end-to-end, effectively resolving the third issue. For a diffusion model with the uncorrected timestep setting, fine-tuning the diffusion model end-to-end is impractical, because generating a clean material map typically requires multiple denoising steps, leading to prohibitive computational complexity due to the gradient calculations. As an alternative, previous methods [6, 28, 56] rely on using the noise latent obtained from a single denoising step of the UNet for loss calculation. However, this approach shifts the model's training objective from material decomposition to general image denoising, which hinders specialization for this task. In contrast, our corrected inference pipeline can generate fully denoised material maps in a single denoising step, making end-to-end fine-tuning both feasible and efficient for material decomposition. Specifically, with the clean material estimations, we compute perceptual losses [23] between the predicted albedo map \\( k_{d} \\), the RM map \\( k_{rm} \\), and the ground truth maps \\( \\hat{k}_d, \\hat{k}_{rm} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.731, + 0.905, + 0.79 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {p} (\\hat {y}, y) = \\sum_ {j} \\frac {1}{C _ {j} H _ {j} W _ {j}} | | \\phi_ {j} (\\hat {y}) - \\phi_ {j} (y) | | _ {1}, \\tag {1} \\\\ \\mathcal {L} _ {m} = \\mathcal {L} _ {p} (\\hat {k} _ {d}, k _ {d}) + \\mathcal {L} _ {p} (\\hat {k} _ {r m}, k _ {r m}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.796, + 0.905, + 0.826 + ], + "angle": 0, + "content": "where \\(\\phi\\) is a 16-layer VGG network, and \\(\\phi_{j}\\) represents the output of the \\(j\\)-th layer, with a shape of \\(C_j\\times H_j\\times W_j\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.905, + 0.87 + ], + "angle": 0, + "content": "Note that, following the approach of previous work [17], we use \"Zeros\" as the noise type during fine-tuning, meaning no random noise is added." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Re-render Loss in DM Fine-tuning. Data-driven decomposition diffusion models typically focus on the distribu" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25086" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.103, + 0.096, + 0.895, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.279, + 0.908, + 0.325 + ], + "angle": 0, + "content": "Figure 3. Material Decomposition for 3D objects. Applying SuperMat's powerful decomposition to 3D objects, we can obtain an object with high-quality materials in just three inference steps. Due to implementation differences, the texels in the RM map for image space decomposition are (1, roughness, metallic), while in the UV refinement process, they are (0, roughness, metallic)." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.33, + 0.485, + 0.725 + ], + "angle": 0, + "content": "tion of individual material maps, yet they often overlook the interactions between different types of material maps. In other words, previous methods compute the loss between the predicted values and the ground truth without utilizing material information for further operations or calculations, which could provide stronger guidance for the model. As a result, although the predicted materials may appear close to the ground truth, the rendered output may not align with expectations. This limitation is inherent to their training methods, as they either produce meaningless predicted noise, preventing the application of image-level loss, or fail to generate all material predictions simultaneously. In contrast, our design, which incorporates structural expert branches and single-step prediction, allows us to obtain all denoised material maps during the training iterations. Using this information, we introduce the re-render loss [10], which involves rendering the model in a new lighting environment and comparing the result to the ground truth. Specifically, at each training iteration, we obtain \\( k_{d} \\), \\( k_{rm} = (k_{r},k_{m}) \\), along with the true normal map \\( \\hat{k}_n \\), position map \\( \\hat{k}_p \\), camera viewpoint \\( \\hat{c} \\), and an environment light map. Then we use this information and our rendering model \\( \\mathcal{R} \\) to compute the predicted material's performance under new illumination. Similarly, with the groundtruth albedo \\( \\hat{k}_d \\) and RM \\( \\hat{k}_{rm} \\), we compute the ground truth of relighting. The perceptual loss between these two renderings defines our re-render loss:" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.742, + 0.483, + 0.776 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e} = \\mathcal {L} _ {p} \\left(\\mathcal {R} _ {\\hat {k} _ {n}, \\hat {k} _ {p}, \\hat {c}} \\left(\\hat {k} _ {d}, \\hat {k} _ {m}, \\hat {k} _ {r}\\right), \\mathcal {R} _ {\\hat {k} _ {n}, \\hat {k} _ {p}, \\hat {c}} \\left(k _ {d}, k _ {m}, k _ {r}\\right)\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Extension to Multi-view Material Decomposition. We extend SuperMat to a multi-view model to enhance the consistency of generated results across different viewpoints. Specifically, we use a 3D self-attention layer and incorporate camera extrinsics as conditions, following MVDream [41]. The multi-view model, SuperMatMV, is capable of simultaneously decomposing materials for 6 input views, providing more accurate and consistent material predictions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.329, + 0.741, + 0.344 + ], + "angle": 0, + "content": "3.2. UV Refinement Network" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.351, + 0.907, + 0.593 + ], + "angle": 0, + "content": "We aim to apply SuperMat or SuperMatMV to decompose material for 3D objects with only RGB textures. Several previous works [5, 36, 45, 62] use techniques like SDS to bridge the gap between dimensions, but such methods are not applicable to our deterministic model, and suffer from slow computational speed and mode seeking issues. Therefore, we explore using backprojection to transfer 2D decomposition results from image space to UV space, enabling their application to 3D models. Notably, both SuperMat and SuperMatMV are finite-view models, meaning that one inference can only decompose the visible region from a specific set of viewpoints. To address this problem, we propose estimating materials from multiple viewpoints and integrating them. With multi-view decomposition results from SuperMatMV, the UV refinement network aims to paint missing areas and improve the quality of the generated materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.593, + 0.909, + 0.683 + ], + "angle": 0, + "content": "Backprojection and UV Blending. For each viewpoint, we backproject SuperMatMV's decomposition results from image space onto the model's UV texture map, obtaining the partial albedo UV map \\( C_d^i \\), RM UV map \\( C_{rm}^i \\), and mask UV map \\( M^i \\) for viewpoint \\( i \\). We then blend the backprojection results from all viewpoints in an averaged manner." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.684, + 0.766, + 0.698 + ], + "angle": 0, + "content": "For albedo, this can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.668, + 0.706, + 0.765, + 0.737 + ], + "angle": 0, + "content": "\\[\nM = \\sum_ {i} M ^ {i},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.656, + 0.733, + 0.905, + 0.773 + ], + "angle": 0, + "content": "\\[\nC _ {d} = \\frac {\\sum_ {i} C _ {d} ^ {i}}{(M + \\epsilon)}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.908, + 0.825 + ], + "angle": 0, + "content": "The RM follows the same computation method. The blended partial maps \\( C_d, C_{rm} \\), and the partial mask \\( M \\) are then input into the UV refinement one-step model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.901 + ], + "angle": 0, + "content": "UV Refinement One-Step Model. Due to the limited number of viewpoints, the blended partial maps often have missing areas. To address this issue, we leverage diffusion models for UV refinement: we fine-tune Stable Diffusion into a UV refinement model, which takes a flawed partial material" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25087" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.905, + 0.293 + ], + "angle": 0, + "content": "
AlbedoMetallicRoughnessRelightingTime(s)↓
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Derender3D [46]20.970.86060.1821---------0.16
IIR [64]21.940.87090.155217.950.76190.291119.730.86390.159620.980.88840.11300.04
IID [25]22.560.89080.128217.380.76150.295620.930.87590.147622.450.90440.10831.45
RGB→X [56]22.300.88600.115415.360.75560.336020.400.84730.159121.510.88610.11323.32
IntrinsicAnything [6]20.940.87880.1520---------6.93
StableMaterial [28]23.440.90190.107620.290.81640.245921.010.88800.132722.560.90430.09790.53
StableMaterialMV [28]24.560.90910.094617.280.78860.282920.860.88380.128623.140.90900.09330.55
SuperMat w/o e2e24.260.89640.109520.790.84440.282620.810.86210.158723.900.90850.09643.09
SuperMat w/o re-render26.700.92360.090624.540.82590.226323.520.90160.136226.410.93780.07020.07
SuperMat27.680.92900.082025.480.86850.222124.250.90570.128127.660.94320.06120.07
SuperMatMV27.560.92890.083926.110.87490.219524.840.90760.120827.640.94310.06280.09
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.303, + 0.907, + 0.363 + ], + "angle": 0, + "content": "Table 1. Comparison of our method with others on the image space decomposition task. We highlight the best, second-best, and third-best results for each metric. \"SuperMat w/o e2e\" denotes our method without the scheduler fix and perceptual loss for end-to-end training. \"SuperMat w/o re-render\" denotes our method without the re-render loss. The first two methods are non-diffusion-based, while the middle five methods are diffusion-based." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.37, + 0.482, + 0.491 + ], + "angle": 0, + "content": "UV map, a mask UV map, and a position UV map as inputs, and outputs completed and refined material UV maps. Specifically, the material maps are encoded by a VAE and concatenated with downsampled mask and position maps, and then processed by UNet with an expanded 8-channel conv_in. We also modify the model's scheduler to support single-step inference and enable end-to-end training, resulting in a deterministic one-step model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.483, + 0.72 + ], + "angle": 0, + "content": "Material Decomposition Pipeline for 3D Objects. As shown in Fig. 3, with the fine-tuned UV refinement network, we propose a complete pipeline for the material decomposition of 3D objects. Given an object with only RGB texture, we first render it from six orthogonal viewpoints. The rendered results are then batch-processed by SuperMatMV for viewpoint-consistent material decomposition. These materials, including albedo and RM, are backprojected into UV space and blended. The blended partial maps are then refined by the UV refinement one-step model and the final material UV maps are generated. This process requires only three inference steps—one for SuperMatMV, one for refining albedo, and one for refining RM—allowing our pipeline to complete the material decomposition of a 3D object in just 3 seconds." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.738, + 0.223, + 0.755 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Training Datasets. In this paper, we collect 3D objects from Objaverse [8] and Digital Twin Catalog (DTC) [13] datasets. For each object, we render material maps, normal maps, and position maps from 6 orthogonal viewpoints and generate PBR results under 16 different environment lighting conditions from the same viewpoints. We then filter high-quality data using methods such as DINO score considerations [35], resulting in a dataset of 75k objects that combines both synthetic and real-captured data for the im" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.37, + 0.907, + 0.461 + ], + "angle": 0, + "content": "age space decomposition task. Following the approach outlined in Sec. 3.2, we further refine this data for the UV refinement network by selecting cases where the missing area in blended partial UV maps is under \\(50\\%\\). This process results in a final dataset of 40k objects for training the refinement one-step model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.909, + 0.734 + ], + "angle": 0, + "content": "Quantitative Comparison of Image Space Decomposition. We randomly select 64 unseen items each from the Objaverse, DTC and BlenderVault [28] datasets. Objaverse and BlenderVault primarily contain synthetic data, whereas DTC consists of scanned objects, offering a source of real-world data. Our method is tested against 7 baseline approaches on these 192 objects. For each object, the model processes 18 inputs corresponding to 3 lighting conditions and 6 viewpoints, predicting 3 (or 1) types of materials and generating 54 (or 18) texture maps in total. All outputs are used to compute metrics, with the final averaged results shown in Tab.1. We use PSNR, SSIM, and LPIPS metrics to evaluate albedo, metallic, and roughness. Additionally, for methods that provide all material types, we compute metrics on the rendering under relighting conditions using the ground truth normals and positions and the predicted values. The results demonstrate that our approach significantly outperforms the SOTA in image space decomposition." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Beyond image quality, we compare inference speeds by calculating the total time required for each method to process the test set, normalized by the dataset size and number of predicted material types. All tests are conducted with a batch size of 1. For consistency in speed comparison, IID, StableMaterial and StableMaterialMV, which typically average over 10 predictions per input, are modified to perform a single prediction. The \"Time(s)\" column in Tab.1 reflects the time required by each method to generate a single material prediction on an NVIDIA A100 GPU with a batch size of 1. Traditional, non-diffusion-based methods like Deren" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25088" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.086, + 0.898, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.449, + 0.908, + 0.478 + ], + "angle": 0, + "content": "Figure 4. Comparison of our method with others on albedo estimation results. Each pair of rows presents results of the same object under identical lighting conditions from different viewpoints. This organizational format is also followed in other qualitative comparison figures." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.487, + 0.48, + 0.702 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.718, + 0.484, + 0.802 + ], + "angle": 0, + "content": "Figure 5. Comparison of our method with others on metallic and roughness estimation results. We combined both materials into a single image, with the left half representing metallic and the right half showing roughness. It can be observed that these two materials exhibit more variation and are more challenging to estimate compared to albedo." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.903 + ], + "angle": 0, + "content": "der3D and IIR generally show faster speeds. While Stable-Material attempts to predict multiple materials simultaneously by adding extra channels to conv_in and conv_out, its lack of structural design adjustments limits prediction quality and requires multi-step denoising, resulting in slower performance than traditional methods. In contrast, our ap-" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.491, + 0.9, + 0.596 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.608, + 0.866, + 0.622 + ], + "angle": 0, + "content": "Figure 6. Results of UV refinement one-step model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.907, + 0.657 + ], + "angle": 0, + "content": "proach achieves high-quality material decomposition with speeds comparable to traditional methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Qualitative Comparison of Image Space Decomposition. Fig.4 compares albedo prediction results across various cases, including artist-designed objects and real-world scans, showcasing the performance of our method against other approaches. SuperMat demonstrates superior ability in removing highlights and shadows, accurately capturing the intrinsic color, even on challenging metallic surfaces. In Fig.5, we present metallic and roughness predictions from our method alongside other baselines. SuperMat outperforms the others with well-defined boundaries between metallic and non-metallic regions and achieves consistency within material types while preserving high-frequency details. Moreover, our results closely align with physical realism, addressing a significant challenge that previous methods struggled to overcome. Additionally, SuperMatMV demonstrates stronger consistency across different views in" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25089" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.159, + 0.089, + 0.201, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.169, + 0.232, + 0.189, + 0.243 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.091, + 0.294, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.249, + 0.232, + 0.293, + 0.243 + ], + "angle": 0, + "content": "w/o e2e" + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.091, + 0.382, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.232, + 0.398, + 0.243 + ], + "angle": 0, + "content": "w/o re-render" + }, + { + "type": "image", + "bbox": [ + 0.432, + 0.091, + 0.473, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.439, + 0.232, + 0.467, + 0.243 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.105, + 0.572, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.156, + 0.572, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.232, + 0.553, + 0.243 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.106, + 0.663, + 0.141 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.157, + 0.663, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.232, + 0.658, + 0.243 + ], + "angle": 0, + "content": "w/o e2e" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.106, + 0.754, + 0.141 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.157, + 0.754, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.688, + 0.232, + 0.762, + 0.243 + ], + "angle": 0, + "content": "w/o re-render" + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.106, + 0.845, + 0.141 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.157, + 0.845, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.803, + 0.232, + 0.832, + 0.243 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.259, + 0.89, + 0.275 + ], + "angle": 0, + "content": "Figure 7. Qualitative ablation results for SuperMat. The first row shows albedo, while the second row shows metallic and roughness." + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.287, + 0.475, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.473, + 0.483, + 0.515 + ], + "angle": 0, + "content": "Figure 8. Results of decomposition for 3D objects. We compare the original and decomposed models under the same illumination and show the relighting outcomes of the new models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.483, + 0.584 + ], + "angle": 0, + "content": "all material predictions, and is still able to generate reasonable decomposition results for inputs from challenging viewpoints with sparse relevant information. More results will be presented in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.75 + ], + "angle": 0, + "content": "Results of UV Refinement and Material Estimation for 3D Objects. Fig.6 illustrates the effectiveness of the UV refinement one-step model in inpainting missing areas in SuperMatMV's outputs. Fig.8 displays the results from our material estimation pipeline for 3D models. The first column shows the rendering of the input models with only RGB texture, where all surfaces exhibit uniform reflectance behavior. After applying our method, the decomposed materials enable physically-based rendering with surface-specific reflectance properties, as demonstrated in the second and third columns under different lighting conditions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Ablation Studies. In Tab.1, we present ablation studies on two core components of SuperMat: end-to-end (e2e) training and re-render loss. E2E with perceptual loss significantly enhances model performance and efficiency. In fact, within the e2e framework, we experiment with three types of loss functions, with perceptual loss achieving the best results. This suggests that direct supervision of the fully denoised predicted material maps, rather than latent noise, plays a critical role in enhancing the model's capability for this specific task. Detailed findings are available in the sup" + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.294, + 0.541, + 0.348 + ], + "angle": 0, + "content": "o o o o o o o o o o o o o o o o o o o o o o o" + }, + { + "type": "image", + "bbox": [ + 0.566, + 0.289, + 0.637, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.387, + 0.54, + 0.4 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.566, + 0.356, + 0.637, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.288, + 0.794, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.357, + 0.794, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.84, + 0.286, + 0.875, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.842, + 0.356, + 0.876, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.433, + 0.902, + 0.447 + ], + "angle": 0, + "content": "Figure 9. Qualitative ablation results on UV refinement network." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.455, + 0.905, + 0.575 + ], + "angle": 0, + "content": "plementary materials. The addition of re-render loss, facilitated by the e2e framework, further improves prediction quality, particularly under relighting conditions. Qualitative results of these ablations are shown in Fig. 7. The 'w/o e2e' setting, where the model is trained without perceptual and re-render losses, limits its capacity to remove lighting effects effectively. In contrast, training 'w/o re-render' loss leads to minor inaccuracies in fine material details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.576, + 0.906, + 0.712 + ], + "angle": 0, + "content": "For the decomposition pipeline for 3D objects, we conduct additional ablations on the UV refinement network. Due to limitations from only six orthogonal views, achieving complete texture coverage, accurate merging of overlapping regions, and consistent material transitions are challenging. As shown in Fig.9, omitting the UV refinement leads to incomplete textures, visible color-separation lines, and unnatural transitions between materials, underscoring the importance of this component." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.725, + 0.641, + 0.74 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this paper, we present SuperMat, an innovative image space material decomposition one-step model finetuned from Stable Diffusion with task-specific losses in an end-to-end framework. SuperMat enables simultaneous multi-material prediction and supports single-step inference, achieving state-of-the-art quality and outperforming previous diffusion-based methods in efficiency. Through the proposed UV refinement network, we extend SuperMat to decomposition tasks for 3D objects, where it demonstrates exceptional performance and applicability." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25090" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.207 + ], + "angle": 0, + "content": "This work was supported by National Natural Science Foundation of China (72192821, 62472282, 62302297 and 62302296), Young Elite Scientists Sponsorship Program by CAST (2022QNRC001), the Fundamental Research Funds for the Central Universities (YG2023QNA35) and YuCaiKe [2023] (231111310300)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.218, + 0.188, + 0.234 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.244, + 0.483, + 0.3 + ], + "angle": 0, + "content": "[1] Raphael Bensadoun, Yanir Kleiman, Idan Azuri, Omri Harosh, Andrea Vedaldi, Natalia Neverova, and Oran Gafni. Meta 3d texturegen: Fast and consistent texture generation for 3d objects. arXiv preprint arXiv:2407.02430, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.3, + 0.483, + 0.369 + ], + "angle": 0, + "content": "[2] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan Barron, and Hendrik Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. Advances in Neural Information Processing Systems, 34: 10691-10704, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.37, + 0.483, + 0.426 + ], + "angle": 0, + "content": "[3] Mark Boss, Zixuan Huang, Aaryaman Vasishta, and Varun Jampani. Sf3d: Stable fast 3d mesh reconstruction with uv-unwrapping and illumination disentanglement. arXiv preprint arXiv:2408.00653, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.426, + 0.483, + 0.495 + ], + "angle": 0, + "content": "[4] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18558-18568, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.496, + 0.484, + 0.564 + ], + "angle": 0, + "content": "[5] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.566, + 0.484, + 0.633 + ], + "angle": 0, + "content": "[6] Xi Chen, Sida Peng, Dongchen Yang, Yuan Liu, Bowen Pan, Chengfei Lv, and Xiaowei Zhou. Intrinsicanything: Learning diffusion priors for inverse rendering under unknown illumination. arXiv preprint arXiv:2404.11593, 2024. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.483, + 0.691 + ], + "angle": 0, + "content": "[7] Ziang Cheng, Junxuan Li, and Hongdong Li. Wildlight: In-the-wild inverse rendering with a flashlight. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4305-4314, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.692, + 0.484, + 0.775 + ], + "angle": 0, + "content": "[8] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.776, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[9] Kangle Deng, Timothy Omernick, Alexander Weiss, Deva Ramanan, Jun-Yan Zhu, Tinghui Zhou, and Maneesh Agrawala. *Flashtex: Fast relightable mesh texturing with lightcontrolnet*. In European Conference on Computer Vision, pages 90–107. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[10] Valentin Deschaintre, Miika Aittala, Fredo Durand, George Drettakis, and Adrien Bousseau. Single-image svbrdf capture with a rendering-aware deep network. ACM Transactions on Graphics (ToG), 37(4):1-15, 2018. 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.244, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[11] Yiqun Duan, Xianda Guo, and Zheng Zhu. Diffusiondepth: Diffusion denoising approach for monocular depth estimation. arXiv preprint arXiv:2303.05021, 3, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[12] Xiang Feng, Chang Yu, Zoubin Bi, Yintong Shang, Feng Gao, Hongzhi Wu, Kun Zhou, Chenfanfu Jiang, and Yin Yang. Arm: Appearance reconstruction model for relightable 3d generation. arXiv preprint arXiv:2411.10825, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[13] James Fort. Introducing the digital twin catalog from reality labs research, 2024. https://ai.meta.com/blog/digital-twin-catalog-3d-reconstruction-shopify-reality-labs-research/, Last accessed on 2025-03-07. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[14] Xiao Fu, Wei Yin, Mu Hu, Kaixuan Wang, Yuexin Ma, Ping Tan, Shaojie Shen, Dahua Lin, and Xiaoxiao Long. Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image. In European Conference on Computer Vision, pages 241-258. Springer, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[15] Jian Gao, Chun Gu, Youtian Lin, Hao Zhu, Xun Cao, Li Zhang, and Yao Yao. Relightable 3d gaussian: Real-time point cloud relighting with brdf decomposition and ray tracing. arXiv preprint arXiv:2311.16043, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.461 + ], + "angle": 0, + "content": "[16] Lin Gao, Tong Wu, Yu-Jie Yuan, Ming-Xian Lin, Yu-Kun Lai, and Hao Zhang. Tm-net: Deep generative networks for textured meshes. ACM Transactions on Graphics (TOG), 40 (6):1-15, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[17] Gonzalo Martin Garcia, Karim Abou Zeid, Christian Schmidt, Daan de Geus, Alexander Hermans, and Bastian Leibe. Fine-tuning image-conditional diffusion models is easier than you think. arXiv preprint arXiv:2409.11355, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[18] Ming Gui, Johannes S Fischer, Ulrich Prestel, Pingchuan Ma, Dmytro Kotovenko, Olga Grebenkova, Stefan Andreas Baumann, Vincent Tao Hu, and Björn Ommer. Depthfm: Fast monocular depth estimation with flow matching. arXiv preprint arXiv:2403.13788, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.671 + ], + "angle": 0, + "content": "[19] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18409-18418, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[20] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light, and material decomposition from images using monte carlo rendering and denoising. Advances in Neural Information Processing Systems, 35:22856-22869, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.731, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[21] Jing He, Haodong Li, Wei Yin, Yixun Liang, Leheng Li, Kaiqiang Zhou, Hongbo Zhang, Bingbing Liu, and YingCong Chen. Lotus: Diffusion-based visual foundation model for high-quality dense prediction. arXiv preprint arXiv:2409.18124, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[22] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. Tensoir: Tensorial inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 165-174, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25091" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.135 + ], + "angle": 0, + "content": "In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.217 + ], + "angle": 0, + "content": "[24] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492-9502, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.483, + 0.286 + ], + "angle": 0, + "content": "[25] Peter Kocsis, Vincent Sitzmann, and Matthias Nießner. Intrinsic image diffusion for indoor single-view material estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5198-5208, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.343 + ], + "angle": 0, + "content": "[26] Zhihao Liang, Qi Zhang, Ying Feng, Ying Shan, and Kui Jia. Gs-ir: 3d gaussian splatting for inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21644–21653, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.483, + 0.412 + ], + "angle": 0, + "content": "[27] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 5404-5411, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.483, + 0.482 + ], + "angle": 0, + "content": "[28] Yehonathan Litman, Or Patashnik, Kangle Deng, Aviral Agrawal, Rushikesh Zawar, Fernando De la Torre, and Shubham Tulsiani. Materialfusion: Enhancing inverse rendering with material diffusion priors. arXiv preprint arXiv:2409.15273, 2024. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.483, + 0.552 + ], + "angle": 0, + "content": "[29] Xian Liu, Jian Ren, Aliaksandr Siarohin, Ivan Skorokhodov, Yanyu Li, Dahua Lin, Xihui Liu, Ziwei Liu, and Sergey Tulyakov. Hyperhuman: Hyper-realistic human generation with latent structural diffusion. arXiv preprint arXiv:2310.08579, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.483, + 0.62 + ], + "angle": 0, + "content": "[30] Yunfei Liu, Yu Li, Shaodi You, and Feng Lu. Unsupervised learning for intrinsic image decomposition from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3248-3257, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.483, + 0.663 + ], + "angle": 0, + "content": "[31] Yuxin Liu, Minshan Xie, Hanyuan Liu, and Tien-Tsin Wong. Text-guided texturing by synchronized multi-view diffusion. arXiv preprint arXiv:2311.12891, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.483, + 0.734 + ], + "angle": 0, + "content": "[32] Zexiang Liu, Yangguang Li, Youtian Lin, Xin Yu, Sida Peng, Yan-Pei Cao, Xiaojuan Qi, Xiaoshui Huang, Ding Liang, and Wanli Ouyang. Unidream: Unifying diffusion priors for relightable text-to-3d generation. In European Conference on Computer Vision, pages 74–91. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[33] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[34] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8280-8290, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.872, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[35] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.135 + ], + "angle": 0, + "content": "Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[36] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[37] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng Bo, and Xiaoguang Han. Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9914–9925, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[38] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In ACM SIGGRAPH 2023 conference proceedings, pages 1-11, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.907, + 0.399 + ], + "angle": 0, + "content": "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[40] Saurabh Saxena, Abhishek Kar, Mohammad Norouzi, and David J Fleet. Monocular depth estimation using diffusion models. arXiv preprint arXiv:2302.14816, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.441, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[41] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[42] Zeqi Shi, Xiangyu Lin, and Ying Song. An attention-embedded gan for svbrdf recovery from a single image. Computational Visual Media, 9(3):551-561, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[43] Giuseppe Vecchio, Rosalie Martin, Arthur Roullier, Adrien Kaiser, Romain Rouffet, Valentin Deschaintre, and Tamy Boubekeur. Controlmat: a controlled generative approach to material capture. ACM Transactions on Graphics, 43(5): 1-17, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[44] Yitong Wang, Xudong Xu, Li Ma, Haoran Wang, and Bo Dai. Boosting 3d object generation through pbr materials. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.65, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[45] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.907, + 0.775 + ], + "angle": 0, + "content": "[46] Felix Wimbauer, Shangzhe Wu, and Christian Rupprecht. De-rendering 3d objects in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18490-18499, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[47] Bojun Xiong, Jialun Liu, Jiakui Hu, Chenming Wu, Jinbo Wu, Xing Liu, Chen Zhao, Errui Ding, and Zhouhui Lian. Texgaussian: Generating high-quality pbr material via octree-based 3d gaussian splatting. arXiv preprint arXiv:2411.19654, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[48] Guangkai Xu, Yongtao Ge, Mingyu Liu, Chengxiang Fan, Kangyang Xie, Zhiyue Zhao, Hao Chen, and Chunhua Shen. Diffusion models trained with large data are transferable visual models. arXiv preprint arXiv:2403.06090, 2024. 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25092" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[49] Xudong Xu, Zhaoyang Lyu, Xingang Pan, and Bo Dai. Matlab: Material-aware text-to-3d via latent brdf auto-encoder. arXiv preprint arXiv:2308.09278, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[50] Chongjie Ye, Lingteng Qiu, Xiaodong Gu, Qi Zuo, Yushuang Wu, Zilong Dong, Liefeng Bo, Yuliang Xiu, and Xiaoguang Han. Stablenormal: Reducing diffusion variance for stable and sharp normal. arXiv preprint arXiv:2406.16864, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[51] Renjiao Yi, Chenyang Zhu, and Kai Xu. Weakly-supervised single-view image relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8402-8411, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[52] Kim Youwang, Tae-Hyun Oh, and Gerard Pons-Moll. Paint-it: Text-to-texture synthesis via deep convolutional texture map optimization and physically-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4347-4356, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.39 + ], + "angle": 0, + "content": "[53] Xin Yu, Peng Dai, Wenbo Li, Lan Ma, Zhengzhe Liu, and Xiaojuan Qi. Texture generation on 3d meshes with point-uv diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4206–4216, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.482, + 0.447 + ], + "angle": 0, + "content": "[54] Xin Yu, Ze Yuan, Yuan-Chen Guo, Ying-Tian Liu, Jianhui Liu, Yangguang Li, Yan-Pei Cao, Ding Liang, and Xiaojuan Qi. Texgen: a generative diffusion model for mesh textures. ACM Transactions on Graphics (TOG), 43(6):1-14, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.482, + 0.518 + ], + "angle": 0, + "content": "[55] Xianfang Zeng, Xin Chen, Zhongqi Qi, Wen Liu, Zibo Zhao, Zhibin Wang, Bin Fu, Yong Liu, and Gang Yu. Paint3d: Paint anything 3d with lighting-less texture diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4252-4262, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.52, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[56] Zheng Zeng, Valentin Deschaintre, Iliyan Georgiev, Yannick Hold-Geoffroy, Yiwei Hu, Fujun Luan, Ling-Qi Yan, and Miloš Hasan. Rgb ↔ x: Image decomposition and synthesis using material-and lighting-aware diffusion models. In ACM SIGGRAPH 2024 Conference Papers, pages 1–11, 2024. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[57] Kai Zhang, Fujun Luan, Zhengqi Li, and Noah Snavely. Iron: Inverse rendering by optimizing neural sdfs and materials from photometric images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5565-5574, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[58] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1-20, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[59] Qian Zhang, Vikas Thamizharasan, and James Tompkin. Learning physically based material and lighting decompositions for face editing. Computational Visual Media, 10(2): 295-308, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[60] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Ner-factor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG), 40(6):1-18, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[61] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18643-18652, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[62] Yuqing Zhang, Yuan Liu, Zhiyu Xie, Lei Yang, Zhongyuan Liu, Mengzhou Yang, Runze Zhang, Qilong Kou, Cheng Lin, Wenping Wang, et al. Dreammat: High-quality pbr material generation with geometry-and light-aware diffusion models. ACM Transactions on Graphics (TOG), 43(4):1-18, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[63] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5729-5739, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.372 + ], + "angle": 0, + "content": "[64] Jingsen Zhu, Fujun Luan, Yuchi Huo, Zihao Lin, Zhihua Zhong, Dianbing Xi, Rui Wang, Hujun Bao, Jiaxiang Zheng, and Rui Tang. Learning-based inverse rendering of complex indoor scenes with differentiable monte carlo raytracing. In SIGGRAPH Asia 2022 Conference Papers, pages 1-8, 2022. 2, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25093" + } + ] +] \ No newline at end of file diff --git a/2025/SuperMat_ Physically Consistent PBR Material Estimation at Interactive Rates/4daa8d06-6978-409c-9337-f3e4e925bd23_origin.pdf b/2025/SuperMat_ Physically Consistent PBR Material Estimation at Interactive Rates/4daa8d06-6978-409c-9337-f3e4e925bd23_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..939c8888d297429aa6df824b4fcf1e3148738b1a --- /dev/null +++ b/2025/SuperMat_ Physically Consistent PBR Material Estimation at Interactive Rates/4daa8d06-6978-409c-9337-f3e4e925bd23_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:94cada6c6afb2411d7371ee09ced2aa65cb58dbf5e265850d7f6bfc62451ed03 +size 8763930 diff --git a/2025/SuperMat_ Physically Consistent PBR Material Estimation at Interactive Rates/full.md b/2025/SuperMat_ Physically Consistent PBR Material Estimation at Interactive Rates/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9b72fb226a230491c2daab0694813c64515ab589 --- /dev/null +++ b/2025/SuperMat_ Physically Consistent PBR Material Estimation at Interactive Rates/full.md @@ -0,0 +1,308 @@ +# SuperMat: Physically Consistent PBR Material Estimation at Interactive Rates + +Yijia Hong $^{1,2}$ , Yuan-Chen Guo $^{4*}$ , Ran Yi $^{1}$ , Yulong Chen $^{3}$ , Yan-Pei Cao $^{4\dagger}$ , Lizhuang Ma $^{1\dagger}$ $^{1}$ Shanghai Jiao Tong University $^{2}$ Shanghai Innovation Institute $^{3}$ Harbin Institute of Technology $^{4}$ VAST + +![](images/3e5b4d318e19304306eb1958d0d7d951e93446b05fc63057ec4f4215243a8b0d.jpg) + +![](images/ac9cec3ae7531345cd3634bebdbb2ff5f7ace3c03d12bedd0a65707407c8ba06.jpg) + +![](images/4520b402f018699f03a66ba1caa3806275fe221ca91fc783f720ff6d11642c39.jpg) +Figure 1. We present an efficient method for high-quality material decomposition in both 2D and 3D. Left: Albedo and RM(Roughness, Metallic) predictions of SuperMat on rendered images under unknown illumination. The first row shows inputs with uneven lighting, the second row displays albedo predictions, and the third row presents RM predictions. Each texel in the RM map here is represented as (1, roughness, metallic). Middle: Rendering results of a 3D object with PBR materials obtained by our method under novel lighting. Right: Our method is efficient and produces high-quality results. + +![](images/b83e380c2526df665c449a363951bdaff9dabd27d693a821ac70d7311bab05d0.jpg) + +![](images/c158dd331366d0f7b9ddd0284510517fa9fb60d09304d968a47b24a8015ac1cb.jpg) + +![](images/a7c53e92dbf3047f3ba8a0b951ebffb9d237c083849097cf153d784e56588639.jpg) + +![](images/e3bac2274340d9958e82cde9780f7c6c43f866b6e0b5d6131e9f14114a9ef305.jpg) + +![](images/017dce4215003c430b1d983730ce04e2689dd74c3b88dd336210e92b6319ccd4.jpg) + +![](images/0efdaa27ba494598c1c885f2c05e0b7740930f1b960fe7679c6cbe2d48106c1c.jpg) + +![](images/e61811d1388ec4840c75b64c1c83166b1507d2e96b90c46fdf9c081b8d18d204.jpg) + +![](images/ae099eae91dbbf5d0853adfdda42b1c2261fe7d706b5383304092d926c55aa05.jpg) + +![](images/ba3683dbf0d29127d3655e43ff5b9c6e799083d150a131edd86d7301a837d089.jpg) + +![](images/ab298a554890b80d69831a17f26157075d2040815fbc9a60f9a532a889caeac2.jpg) + +# Abstract + +Decomposing physically-based materials from images into their constituent properties remains challenging, particularly when maintaining both computational efficiency and physical consistency. While recent diffusion-based approaches have shown promise, they face substantial computational overhead due to multiple denoising steps and separate models for different material properties. We present SuperMat, a single-step framework that achieves high-quality material decomposition with one-step inference. This enables end-to-end training with perceptual and re-render losses while decomposing albedo, metallic, and roughness maps at millisecond-scale speeds. We further extend our framework to 3D objects through a UV refinement network, enabling consistent material estimation across viewpoints + +while maintaining efficiency. Experiments demonstrate that SuperMat achieves state-of-the-art PBR material decomposition quality while reducing inference time from seconds to milliseconds per image, and completes PBR material estimation for 3D objects in approximately 3 seconds. The project page is at https://hyj542682306.github.io/SuperMat/. + +# 1. Introduction + +Decomposing Physically-Based Rendering (PBR) materials from images into their constituent properties (e.g., albedo, metallic, and roughness maps) remains a central challenge in computer vision and graphics. This task is essential for creating high-fidelity 3D assets across numerous applications, from film production and game development to virtual and augmented reality pipelines. While recent advances have shown promising results [6, 28, 56], the problem re + +mains challenging due to its ill-posed nature, particularly given unknown lighting conditions and complex interactions between different material properties. Our work addresses two key objectives: 1) efficient decomposition of high-fidelity material maps from 2D images, and 2) extension of this capability to 3D objects, enabling accurate material estimation across complex geometries. + +Previous works in PBR material estimation generally fall into two main categories: optimization-based generation for 3D objects and data-driven decomposition for 2D images. Optimization-based methods [15, 19, 22, 26] employ differentiable rendering to reconstruct 3D objects with material parameters by minimizing the difference between rendered and target images. Instead, data-driven approaches learn the distribution of various materials from large datasets to perform intrinsic decomposition on 2D images, which is the primary focus of our work. Early methods [46, 64] in this domain fail to produce stable and reliable results due to limited data availability and model capacity. Recent works [6, 25, 28, 56] leverage large diffusion models and reframe material decomposition as a conditional image generation task. However, these methods tend to focus solely on individual material properties, overlooking the interactions between different types of materials and the combined impact on rendered appearances. Even minor inaccuracies in individual material estimates can propagate through the rendering pipeline, leading to substantial visual artifacts. Moreover, these diffusion-based methods face considerable computational overhead - they demand multiple denoising steps (30-50) during inference, causing a critical bottleneck for real-time applications. + +To address these challenges, we propose a deterministic one-step framework that achieves both efficient and high-quality material decomposition. First, we introduce and refine structural expert branches within the UNet architecture. These specialized branches enable concurrent processing of multiple material properties while maintaining domain-specific expertise, allowing our model to decompose albedo, metallic, and roughness maps simultaneously with minimal computational overhead. This design facilitates collaborative and physically consistent material prediction while avoiding the linear cost scaling typically required when estimating multiple material attributes. Second, we address the computational inefficiency of traditional diffusion models by reformulating the DDIM sampling process. Through modification of the scheduler, we achieve single-step inference without compromising quality, dramatically reducing inference time. Most importantly, this single-step formulation enables end-to-end training of the diffusion model, allowing us to compute the perceptual and re-render loss on final predicted material textures, which enhances shadow removal and improves the joint performance of decomposed materials under varying lighting + +conditions. These advances lead to SuperMat, a highly efficient image space material decomposition model that estimates high-fidelity albedo, metallic, and roughness maps at millisecond-scale speeds. + +For practical 3D applications, we further extend SuperMat's capabilities to handle 3D objects. We upgrade SuperMat to a multi-view model, SuperMatMV, and design a UV refinement network that processes SuperMatMV's predictions in UV space. This approach addresses the challenge of maintaining consistency and completeness across different, limited viewpoints while preserving the efficiency of our original 2D approach. Our pipeline processes a textured 3D model in approximately 3 seconds, producing high-quality material maps that maintain consistency across arbitrary viewpoints and lighting conditions. + +In summary, our contributions are: + +- We propose SuperMat, a novel image space material decomposition model that generates physically consistent albedo, metallic, and roughness maps simultaneously with a single-step inference process. Our method achieves state-of-the-art performance on diverse datasets, significantly outperforming previous diffusion-based works in speed. +- With the help of structural expert branches and single-step inference, we develop an end-to-end image space material decomposition training system that generates clean predictions for all material properties in each iteration, allowing for the integration of advanced techniques such as perceptual loss and re-render loss. +- We propose an efficient PBR material decomposition method for 3D models. Combining SuperMatMV and our UV refinement network, this pipeline generates high-quality PBR materials for 3D objects from baked textures in approximately 3 seconds. + +# 2. Related Works + +Material Generation and Decomposition. Recognizing the importance of material information in 3D assets, researchers have increasingly focused on material acquisition for objects and scenes. On one hand, some 3D generation approaches [5, 9, 32, 37, 47, 49, 52, 53, 62] integrate material optimization into the appearance creation process, concentrating on generating PBR maps for textured models. On the other hand, inverse rendering methods aim to reconstruct material-rich 3D representations from single or multi-view images. Some of these methods [7, 19, 57] assume a predefined lighting model to enhance estimation accuracy, but this constraint limits their practical applicability. To address this, other works [2, 15, 20, 22, 26, 34, 59-61] attempt to recover both material properties and unknown illumination simultaneously. As reliance on predefined assumptions decreases, researchers have found that beyond controlling material decomposition through rendered out + +![](images/e5cff5141d57d26bbfc4c82c263aaf326ddf2eceeedb537c90e47a1380ac6e63.jpg) +Figure 2. Overview of the SuperMat framework. The UNet architecture incorporates structural expert branches for albedo and RM estimation, enabling parallel material property prediction while sharing a common backbone. During training, we leverage a fixed DDIM scheduler and optimize the network end-to-end using both perceptual loss and re-render loss. The marked region shows the inference pipeline, where our model achieves single-step inference for concurrent material map generation. Input geometric information (normal, position) and environment maps are used to compute the re-render loss. + +puts, directly obtaining image space material priors is also feasible. Building on this idea, similar to ours, another body of research [6, 25, 28, 30, 42-44, 46, 51, 56, 64] focuses on material decomposition from 2D images captured or rendered under unknown lighting conditions. However, existing methods still face challenges related to training data, model architecture, and training strategies, leaving significant room for improvement in decomposition quality. + +Diffusion Models. Recently, researchers have applied diffusion models to a range of tasks [5, 36, 37, 45, 62], demonstrating the effectiveness of casting outcome prediction as a conditional image generation process. By leveraging prior knowledge from large pre-trained models, recent studies [6, 11, 14, 18, 24, 28, 40, 50, 63] have achieved performance improvements over previous methods in tasks like depth, normal, and material estimation. However, incorporating large models introduces efficiency challenges, particularly high memory requirements and the time-intensive, multistep denoising process needed for inference. Inspired by insights from recent work [17, 21, 27, 29], our approach addresses these issues by significantly reducing memory and inference time, improving the feasibility of applying these models in real-world applications. + +Generation in UV Space. UV maps are advantageous for fully encapsulating the appearance information of 3D models, leading several works to explore 3D object manipulation directly within the UV space. By representing the texture of an object as a learnable image, these methods [3, 4, 16, 33, 38, 55] enable texture generation, inpainting, and editing, providing new avenues for 3D tasks and allowing 2D image-space techniques and insights to be applied to 3D appearances. Building on this approach, recent methods [1, 12, 31, 52-55] have applied diffusion models to refine + +UV maps with promising results. In this spirit, to extend SuperMat's capability to 3D, we fine-tune a UV-space refinement diffusion model specifically. + +# 3. Method + +In this paper, we propose SuperMat, an image space material decomposition model fine-tuned from Stable Diffusion that efficiently generates high-quality albedo, metallic, and roughness maps simultaneously. Given a rendered or captured image of an object under an unknown lighting condition from a specific viewpoint, SuperMat generates the corresponding material maps for the model from that same viewpoint. As shown in Figure 2, SuperMat is a 2D imaged-to-image one-step model with structural expert branches that simultaneously generates both albedo and RM maps, supporting single-step inference. Compared to previous diffusion-based methods, this model produces superior results in less time and with reduced memory usage. Our single-step formulation allows direct generation of material maps during training, in contrast to prior diffusion methods that predict noise $(\epsilon)$ or velocity (v) terms. This enables end-to-end optimization with both perceptual loss on the generated maps and re-render loss from novel lighting conditions, leading to improved material quality and rendering fidelity. Building upon the single-view model, we further introduce a multi-view version, SuperMatMV, to ensure consistency in predictions across different viewpoints. + +While SuperMat and SuperMatMV achieve efficient image space material decomposition, extending this capability to 3D objects presents additional challenges. Instead of relying on optimization-based approaches like score distillation sampling (SDS) that are computationally intensive and prone to mode collapse, we propose an efficient two-stage + +pipeline for material estimation on 3D objects. Given a 3D model with only RGB texture, we first apply SuperMatMV to decompose materials from multiple viewpoints. We then introduce a UV refinement network that consolidates these view-dependent predictions into complete, high-quality UV texture maps through single-step inference. As shown in Figure 3, this pipeline enables PBR material estimation for 3D objects in approximately 3 seconds per model. + +# 3.1. SuperMat - Image Space Material Decomposition + +Previous works [6, 28, 56] have demonstrated that casting material estimation as a conditional latent diffusion process, leveraging knowledge from large pre-trained diffusion models like Stable Diffusion [39], is effective. However, for our task that requires the simultaneous generation of multiple material maps, three key issues remain: 1) High model overhead: Fine-tuning Stable Diffusion with its original architecture can only estimate a single material type. Even when combining metallic and roughness into an RM map, treating it as a single material still requires two diffusion models, one for albedo and another for RM, to meet the rendering model's requirements. This results in doubled training time and increased inference space overhead. 2) Slow inference speed: With the widely used DDIM scheduler, diffusion models require 30 to 50 denoising steps to produce satisfactory results, which can be inefficient and time-consuming. 3) Insufficient decomposition effect: Due to the high computational cost of backpropagation in multistep inference, prior diffusion-based methods trained on predicted noise from single-step denoising rather than directly supervising the fully denoised material output. This limitation restricts the model's ability to specialize in the decomposition task and impedes the application of more task-specific techniques. + +Structural Expert Branches. Our solution to the first problem is to introduce a structural expert branch architecture [29, 58]. Previous methods are limited to using a single diffusion model to estimate one material type. Although training separate models for different material maps is an intuitive approach, it introduces issues such as complex implementation, high overhead, and inconsistencies between materials. Considering that, despite their different representations, materials fundamentally reflect an understanding of object features, we incorporate structural expert branches into the UNet model, allowing a single diffusion model to predict multiple material maps. Specifically, we design two expert branches, one for predicting albedo, and the other for predicting RM. For each expert branch, we replicate the UNet's last UpBlock, and conv_out, while the rest of the UNet architecture remains shared between the branches. The modified model first extracts and transforms encoded image features through the shared modules. These features + +are then separated into 2 copies by duplicating before the last UpBlock, with each branch ultimately predicting results for its respective domain. The expert branches specialize in material-specific features, while the shared modules focus on commonalities. In this way, we only add a small number of parameters (19,313,604 parameters, $2.23\%$ of UNet's total), achieving the same performance as previous methods that would otherwise require twice the scale. + +Single-Step Inference and End-to-End Training. To address the second problem, we analyze the DDIM scheduler of diffusion models. Previous works [17, 21, 27, 48] have highlighted a flaw in the implementation of the DDIM scheduler, which limits the potential of single-step inference in Stable Diffusion: In the default leading timestep setting of DDIM, for single-step prediction, the model is provided with a timestep $(t = 1)$ suggesting the input is essentially denoised, yet in reality, the input is pure noise. The discrepancy between the timestep and the actual noise level causes the single-step result to be flawed. To align the timestep with the noise level received by the model, we adjust the scheduler to use the trailing setting $(t = T$ for single-step), correcting this issue and enabling single-step inference, significantly reducing the time overhead. + +More importantly, after this correction, we can fine-tune the one-step model end-to-end, effectively resolving the third issue. For a diffusion model with the uncorrected timestep setting, fine-tuning the diffusion model end-to-end is impractical, because generating a clean material map typically requires multiple denoising steps, leading to prohibitive computational complexity due to the gradient calculations. As an alternative, previous methods [6, 28, 56] rely on using the noise latent obtained from a single denoising step of the UNet for loss calculation. However, this approach shifts the model's training objective from material decomposition to general image denoising, which hinders specialization for this task. In contrast, our corrected inference pipeline can generate fully denoised material maps in a single denoising step, making end-to-end fine-tuning both feasible and efficient for material decomposition. Specifically, with the clean material estimations, we compute perceptual losses [23] between the predicted albedo map $k_{d}$ , the RM map $k_{rm}$ , and the ground truth maps $\hat{k}_d, \hat{k}_{rm}$ : + +$$ +\begin{array}{l} \mathcal {L} _ {p} (\hat {y}, y) = \sum_ {j} \frac {1}{C _ {j} H _ {j} W _ {j}} | | \phi_ {j} (\hat {y}) - \phi_ {j} (y) | | _ {1}, \tag {1} \\ \mathcal {L} _ {m} = \mathcal {L} _ {p} (\hat {k} _ {d}, k _ {d}) + \mathcal {L} _ {p} (\hat {k} _ {r m}, k _ {r m}), \\ \end{array} +$$ + +where $\phi$ is a 16-layer VGG network, and $\phi_{j}$ represents the output of the $j$ -th layer, with a shape of $C_j\times H_j\times W_j$ . + +Note that, following the approach of previous work [17], we use "Zeros" as the noise type during fine-tuning, meaning no random noise is added. + +Re-render Loss in DM Fine-tuning. Data-driven decomposition diffusion models typically focus on the distribu + +![](images/cb2e36bb43d648640679bd5d8db2f80b4d494d0fec520f9e542963e1af2b0453.jpg) +Figure 3. Material Decomposition for 3D objects. Applying SuperMat's powerful decomposition to 3D objects, we can obtain an object with high-quality materials in just three inference steps. Due to implementation differences, the texels in the RM map for image space decomposition are (1, roughness, metallic), while in the UV refinement process, they are (0, roughness, metallic). + +tion of individual material maps, yet they often overlook the interactions between different types of material maps. In other words, previous methods compute the loss between the predicted values and the ground truth without utilizing material information for further operations or calculations, which could provide stronger guidance for the model. As a result, although the predicted materials may appear close to the ground truth, the rendered output may not align with expectations. This limitation is inherent to their training methods, as they either produce meaningless predicted noise, preventing the application of image-level loss, or fail to generate all material predictions simultaneously. In contrast, our design, which incorporates structural expert branches and single-step prediction, allows us to obtain all denoised material maps during the training iterations. Using this information, we introduce the re-render loss [10], which involves rendering the model in a new lighting environment and comparing the result to the ground truth. Specifically, at each training iteration, we obtain $k_{d}$ , $k_{rm} = (k_{r},k_{m})$ , along with the true normal map $\hat{k}_n$ , position map $\hat{k}_p$ , camera viewpoint $\hat{c}$ , and an environment light map. Then we use this information and our rendering model $\mathcal{R}$ to compute the predicted material's performance under new illumination. Similarly, with the groundtruth albedo $\hat{k}_d$ and RM $\hat{k}_{rm}$ , we compute the ground truth of relighting. The perceptual loss between these two renderings defines our re-render loss: + +$$ +\mathcal {L} _ {r e} = \mathcal {L} _ {p} \left(\mathcal {R} _ {\hat {k} _ {n}, \hat {k} _ {p}, \hat {c}} \left(\hat {k} _ {d}, \hat {k} _ {m}, \hat {k} _ {r}\right), \mathcal {R} _ {\hat {k} _ {n}, \hat {k} _ {p}, \hat {c}} \left(k _ {d}, k _ {m}, k _ {r}\right)\right). \tag {2} +$$ + +Extension to Multi-view Material Decomposition. We extend SuperMat to a multi-view model to enhance the consistency of generated results across different viewpoints. Specifically, we use a 3D self-attention layer and incorporate camera extrinsics as conditions, following MVDream [41]. The multi-view model, SuperMatMV, is capable of simultaneously decomposing materials for 6 input views, providing more accurate and consistent material predictions. + +# 3.2. UV Refinement Network + +We aim to apply SuperMat or SuperMatMV to decompose material for 3D objects with only RGB textures. Several previous works [5, 36, 45, 62] use techniques like SDS to bridge the gap between dimensions, but such methods are not applicable to our deterministic model, and suffer from slow computational speed and mode seeking issues. Therefore, we explore using backprojection to transfer 2D decomposition results from image space to UV space, enabling their application to 3D models. Notably, both SuperMat and SuperMatMV are finite-view models, meaning that one inference can only decompose the visible region from a specific set of viewpoints. To address this problem, we propose estimating materials from multiple viewpoints and integrating them. With multi-view decomposition results from SuperMatMV, the UV refinement network aims to paint missing areas and improve the quality of the generated materials. + +Backprojection and UV Blending. For each viewpoint, we backproject SuperMatMV's decomposition results from image space onto the model's UV texture map, obtaining the partial albedo UV map $C_d^i$ , RM UV map $C_{rm}^i$ , and mask UV map $M^i$ for viewpoint $i$ . We then blend the backprojection results from all viewpoints in an averaged manner. + +For albedo, this can be formulated as: + +$$ +M = \sum_ {i} M ^ {i}, +$$ + +$$ +C _ {d} = \frac {\sum_ {i} C _ {d} ^ {i}}{(M + \epsilon)}. \tag {3} +$$ + +The RM follows the same computation method. The blended partial maps $C_d, C_{rm}$ , and the partial mask $M$ are then input into the UV refinement one-step model. + +UV Refinement One-Step Model. Due to the limited number of viewpoints, the blended partial maps often have missing areas. To address this issue, we leverage diffusion models for UV refinement: we fine-tune Stable Diffusion into a UV refinement model, which takes a flawed partial material + +
AlbedoMetallicRoughnessRelightingTime(s)↓
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Derender3D [46]20.970.86060.1821---------0.16
IIR [64]21.940.87090.155217.950.76190.291119.730.86390.159620.980.88840.11300.04
IID [25]22.560.89080.128217.380.76150.295620.930.87590.147622.450.90440.10831.45
RGB→X [56]22.300.88600.115415.360.75560.336020.400.84730.159121.510.88610.11323.32
IntrinsicAnything [6]20.940.87880.1520---------6.93
StableMaterial [28]23.440.90190.107620.290.81640.245921.010.88800.132722.560.90430.09790.53
StableMaterialMV [28]24.560.90910.094617.280.78860.282920.860.88380.128623.140.90900.09330.55
SuperMat w/o e2e24.260.89640.109520.790.84440.282620.810.86210.158723.900.90850.09643.09
SuperMat w/o re-render26.700.92360.090624.540.82590.226323.520.90160.136226.410.93780.07020.07
SuperMat27.680.92900.082025.480.86850.222124.250.90570.128127.660.94320.06120.07
SuperMatMV27.560.92890.083926.110.87490.219524.840.90760.120827.640.94310.06280.09
+ +Table 1. Comparison of our method with others on the image space decomposition task. We highlight the best, second-best, and third-best results for each metric. "SuperMat w/o e2e" denotes our method without the scheduler fix and perceptual loss for end-to-end training. "SuperMat w/o re-render" denotes our method without the re-render loss. The first two methods are non-diffusion-based, while the middle five methods are diffusion-based. + +UV map, a mask UV map, and a position UV map as inputs, and outputs completed and refined material UV maps. Specifically, the material maps are encoded by a VAE and concatenated with downsampled mask and position maps, and then processed by UNet with an expanded 8-channel conv_in. We also modify the model's scheduler to support single-step inference and enable end-to-end training, resulting in a deterministic one-step model. + +Material Decomposition Pipeline for 3D Objects. As shown in Fig. 3, with the fine-tuned UV refinement network, we propose a complete pipeline for the material decomposition of 3D objects. Given an object with only RGB texture, we first render it from six orthogonal viewpoints. The rendered results are then batch-processed by SuperMatMV for viewpoint-consistent material decomposition. These materials, including albedo and RM, are backprojected into UV space and blended. The blended partial maps are then refined by the UV refinement one-step model and the final material UV maps are generated. This process requires only three inference steps—one for SuperMatMV, one for refining albedo, and one for refining RM—allowing our pipeline to complete the material decomposition of a 3D object in just 3 seconds. + +# 4. Experiments + +Training Datasets. In this paper, we collect 3D objects from Objaverse [8] and Digital Twin Catalog (DTC) [13] datasets. For each object, we render material maps, normal maps, and position maps from 6 orthogonal viewpoints and generate PBR results under 16 different environment lighting conditions from the same viewpoints. We then filter high-quality data using methods such as DINO score considerations [35], resulting in a dataset of 75k objects that combines both synthetic and real-captured data for the im + +age space decomposition task. Following the approach outlined in Sec. 3.2, we further refine this data for the UV refinement network by selecting cases where the missing area in blended partial UV maps is under $50\%$ . This process results in a final dataset of 40k objects for training the refinement one-step model. + +Quantitative Comparison of Image Space Decomposition. We randomly select 64 unseen items each from the Objaverse, DTC and BlenderVault [28] datasets. Objaverse and BlenderVault primarily contain synthetic data, whereas DTC consists of scanned objects, offering a source of real-world data. Our method is tested against 7 baseline approaches on these 192 objects. For each object, the model processes 18 inputs corresponding to 3 lighting conditions and 6 viewpoints, predicting 3 (or 1) types of materials and generating 54 (or 18) texture maps in total. All outputs are used to compute metrics, with the final averaged results shown in Tab.1. We use PSNR, SSIM, and LPIPS metrics to evaluate albedo, metallic, and roughness. Additionally, for methods that provide all material types, we compute metrics on the rendering under relighting conditions using the ground truth normals and positions and the predicted values. The results demonstrate that our approach significantly outperforms the SOTA in image space decomposition. + +Beyond image quality, we compare inference speeds by calculating the total time required for each method to process the test set, normalized by the dataset size and number of predicted material types. All tests are conducted with a batch size of 1. For consistency in speed comparison, IID, StableMaterial and StableMaterialMV, which typically average over 10 predictions per input, are modified to perform a single prediction. The "Time(s)" column in Tab.1 reflects the time required by each method to generate a single material prediction on an NVIDIA A100 GPU with a batch size of 1. Traditional, non-diffusion-based methods like Deren + +![](images/0963d2da006a4748a87353ac7caf8d3c9664f82da3a59190240f283dfe47a27e.jpg) +Figure 4. Comparison of our method with others on albedo estimation results. Each pair of rows presents results of the same object under identical lighting conditions from different viewpoints. This organizational format is also followed in other qualitative comparison figures. + +![](images/d419c02f86c1cdf2746513ce077a0742047cf8a2617fab8ba7d6b4ae44886b70.jpg) +Figure 5. Comparison of our method with others on metallic and roughness estimation results. We combined both materials into a single image, with the left half representing metallic and the right half showing roughness. It can be observed that these two materials exhibit more variation and are more challenging to estimate compared to albedo. + +der3D and IIR generally show faster speeds. While Stable-Material attempts to predict multiple materials simultaneously by adding extra channels to conv_in and conv_out, its lack of structural design adjustments limits prediction quality and requires multi-step denoising, resulting in slower performance than traditional methods. In contrast, our ap- + +![](images/caddc01cd0d9a5160dd7e6da61d7796d71564ca1c7759cf110261fb014f262f4.jpg) +Figure 6. Results of UV refinement one-step model. + +proach achieves high-quality material decomposition with speeds comparable to traditional methods. + +Qualitative Comparison of Image Space Decomposition. Fig.4 compares albedo prediction results across various cases, including artist-designed objects and real-world scans, showcasing the performance of our method against other approaches. SuperMat demonstrates superior ability in removing highlights and shadows, accurately capturing the intrinsic color, even on challenging metallic surfaces. In Fig.5, we present metallic and roughness predictions from our method alongside other baselines. SuperMat outperforms the others with well-defined boundaries between metallic and non-metallic regions and achieves consistency within material types while preserving high-frequency details. Moreover, our results closely align with physical realism, addressing a significant challenge that previous methods struggled to overcome. Additionally, SuperMatMV demonstrates stronger consistency across different views in + +![](images/fd0c9c4644be9708947a63a51562392785b83d85421a2b8aa104b997b5c8b44c.jpg) +GT + +![](images/d2308a59b1d9f190207f17567be32b59dae11d12adef051fb6644b7ebdcdd361.jpg) +w/o e2e + +![](images/6f042577cea02fb38ef6cc9537426081f79117bc0eb14f3dd3d70a16e104bb1d.jpg) +w/o re-render + +![](images/f0c0c57f66b0181e0f98609554254a5736e65905d9188bdd943006a041b16f50.jpg) +Ours + +![](images/ef1c30d76a7fd70142c7902df3661335c4a83069405c13af5f8ceb935bf0b972.jpg) + +![](images/fb285aa4f44b71d4662d46cb426871bcb341bc6eeb1b940ed60909d5e272a19a.jpg) +GT + +![](images/cc0e17a7dd2eff3316323ec4c6508f27f13746481cdea13483886c0d66a32c59.jpg) + +![](images/0d9b08838ef8b7a4e7db3be82416be5c7c6812124bb5081eb7582b511f33463e.jpg) +w/o e2e + +![](images/ff48193d0efa128b7fb9c487c6eec94a95ab6f4f6b005bfd984519935f9f1e4b.jpg) + +![](images/53b82535cb740fe0fc1db1ebfd0d127e1baee0b48100cd8087bb6ea25f8fe0c8.jpg) +w/o re-render + +![](images/6d3802c777215d8f231e69f4263da890d69692d7f3b55114ea4c5cf21d979e18.jpg) + +![](images/caf03c59044843896edd2ffb2f1fa8832cfc1d53bb286d40b2f2a4de7946e2e2.jpg) +Ours + +![](images/f5715ed0b3fff20980d02c21c468423010f1713de89020dda29cdeb07c73b447.jpg) +Figure 8. Results of decomposition for 3D objects. We compare the original and decomposed models under the same illumination and show the relighting outcomes of the new models. + +all material predictions, and is still able to generate reasonable decomposition results for inputs from challenging viewpoints with sparse relevant information. More results will be presented in the supplementary materials. + +Results of UV Refinement and Material Estimation for 3D Objects. Fig.6 illustrates the effectiveness of the UV refinement one-step model in inpainting missing areas in SuperMatMV's outputs. Fig.8 displays the results from our material estimation pipeline for 3D models. The first column shows the rendering of the input models with only RGB texture, where all surfaces exhibit uniform reflectance behavior. After applying our method, the decomposed materials enable physically-based rendering with surface-specific reflectance properties, as demonstrated in the second and third columns under different lighting conditions. + +Ablation Studies. In Tab.1, we present ablation studies on two core components of SuperMat: end-to-end (e2e) training and re-render loss. E2E with perceptual loss significantly enhances model performance and efficiency. In fact, within the e2e framework, we experiment with three types of loss functions, with perceptual loss achieving the best results. This suggests that direct supervision of the fully denoised predicted material maps, rather than latent noise, plays a critical role in enhancing the model's capability for this specific task. Detailed findings are available in the sup + +![](images/614d9a813e56f4e0d3c6d24c235948a90a895d291f6f4aa49372283bf7abbfcd.jpg) +o o o o o o o o o o o o o o o o o o o o o o o + +![](images/deb5622d2540657eeef09443c3ddd2713601455c6859fbb7f739a8712cd72c33.jpg) +Ours +Figure 9. Qualitative ablation results on UV refinement network. + +![](images/9becd74f3dc913dfc3d72d1700e5a0c839dae5b197099b2ba294a64528efc3f2.jpg) + +![](images/79f712c9a1557e8195196cea1bf8c045f21cc6f4bf932f72410d450b70007405.jpg) + +![](images/113c94b69b5af2681749b75397c16c24804295d26a35eef8c8de5c16c324e86e.jpg) +Figure 7. Qualitative ablation results for SuperMat. The first row shows albedo, while the second row shows metallic and roughness. + +![](images/8c64046ca37abec96a70c8f7159776ae951810d7b594daa63ceea5abc1dfc8fb.jpg) + +plementary materials. The addition of re-render loss, facilitated by the e2e framework, further improves prediction quality, particularly under relighting conditions. Qualitative results of these ablations are shown in Fig. 7. The 'w/o e2e' setting, where the model is trained without perceptual and re-render losses, limits its capacity to remove lighting effects effectively. In contrast, training 'w/o re-render' loss leads to minor inaccuracies in fine material details. + +For the decomposition pipeline for 3D objects, we conduct additional ablations on the UV refinement network. Due to limitations from only six orthogonal views, achieving complete texture coverage, accurate merging of overlapping regions, and consistent material transitions are challenging. As shown in Fig.9, omitting the UV refinement leads to incomplete textures, visible color-separation lines, and unnatural transitions between materials, underscoring the importance of this component. + +# 5. Conclusions + +In this paper, we present SuperMat, an innovative image space material decomposition one-step model finetuned from Stable Diffusion with task-specific losses in an end-to-end framework. SuperMat enables simultaneous multi-material prediction and supports single-step inference, achieving state-of-the-art quality and outperforming previous diffusion-based methods in efficiency. Through the proposed UV refinement network, we extend SuperMat to decomposition tasks for 3D objects, where it demonstrates exceptional performance and applicability. + +# Acknowledgments + +# References + +[1] Raphael Bensadoun, Yanir Kleiman, Idan Azuri, Omri Harosh, Andrea Vedaldi, Natalia Neverova, and Oran Gafni. Meta 3d texturegen: Fast and consistent texture generation for 3d objects. arXiv preprint arXiv:2407.02430, 2024. 3 +[2] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan Barron, and Hendrik Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. Advances in Neural Information Processing Systems, 34: 10691-10704, 2021. 2 +[3] Mark Boss, Zixuan Huang, Aaryaman Vasishta, and Varun Jampani. Sf3d: Stable fast 3d mesh reconstruction with uv-unwrapping and illumination disentanglement. arXiv preprint arXiv:2408.00653, 2024. 3 +[4] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18558-18568, 2023. 3 +[5] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 3, 5 +[6] Xi Chen, Sida Peng, Dongchen Yang, Yuan Liu, Bowen Pan, Chengfei Lv, and Xiaowei Zhou. Intrinsicanything: Learning diffusion priors for inverse rendering under unknown illumination. arXiv preprint arXiv:2404.11593, 2024. 1, 2, 3, 4, 6 +[7] Ziang Cheng, Junxuan Li, and Hongdong Li. Wildlight: In-the-wild inverse rendering with a flashlight. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4305-4314, 2023. 2 +[8] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 6 +[9] Kangle Deng, Timothy Omernick, Alexander Weiss, Deva Ramanan, Jun-Yan Zhu, Tinghui Zhou, and Maneesh Agrawala. *Flashtex: Fast relightable mesh texturing with lightcontrolnet*. In European Conference on Computer Vision, pages 90–107. Springer, 2024. 2 +[10] Valentin Deschaintre, Miika Aittala, Fredo Durand, George Drettakis, and Adrien Bousseau. Single-image svbrdf capture with a rendering-aware deep network. ACM Transactions on Graphics (ToG), 37(4):1-15, 2018. 5 + +This work was supported by National Natural Science Foundation of China (72192821, 62472282, 62302297 and 62302296), Young Elite Scientists Sponsorship Program by CAST (2022QNRC001), the Fundamental Research Funds for the Central Universities (YG2023QNA35) and YuCaiKe [2023] (231111310300). +[11] Yiqun Duan, Xianda Guo, and Zheng Zhu. Diffusiondepth: Diffusion denoising approach for monocular depth estimation. arXiv preprint arXiv:2303.05021, 3, 2023. 3 +[12] Xiang Feng, Chang Yu, Zoubin Bi, Yintong Shang, Feng Gao, Hongzhi Wu, Kun Zhou, Chenfanfu Jiang, and Yin Yang. Arm: Appearance reconstruction model for relightable 3d generation. arXiv preprint arXiv:2411.10825, 2024. 3 +[13] James Fort. Introducing the digital twin catalog from reality labs research, 2024. https://ai.meta.com/blog/digital-twin-catalog-3d-reconstruction-shopify-reality-labs-research/, Last accessed on 2025-03-07. 6 +[14] Xiao Fu, Wei Yin, Mu Hu, Kaixuan Wang, Yuexin Ma, Ping Tan, Shaojie Shen, Dahua Lin, and Xiaoxiao Long. Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image. In European Conference on Computer Vision, pages 241-258. Springer, 2025. 3 +[15] Jian Gao, Chun Gu, Youtian Lin, Hao Zhu, Xun Cao, Li Zhang, and Yao Yao. Relightable 3d gaussian: Real-time point cloud relighting with brdf decomposition and ray tracing. arXiv preprint arXiv:2311.16043, 2023. 2 +[16] Lin Gao, Tong Wu, Yu-Jie Yuan, Ming-Xian Lin, Yu-Kun Lai, and Hao Zhang. Tm-net: Deep generative networks for textured meshes. ACM Transactions on Graphics (TOG), 40 (6):1-15, 2021. 3 +[17] Gonzalo Martin Garcia, Karim Abou Zeid, Christian Schmidt, Daan de Geus, Alexander Hermans, and Bastian Leibe. Fine-tuning image-conditional diffusion models is easier than you think. arXiv preprint arXiv:2409.11355, 2024. 3, 4 +[18] Ming Gui, Johannes S Fischer, Ulrich Prestel, Pingchuan Ma, Dmytro Kotovenko, Olga Grebenkova, Stefan Andreas Baumann, Vincent Tao Hu, and Björn Ommer. Depthfm: Fast monocular depth estimation with flow matching. arXiv preprint arXiv:2403.13788, 2024. 3 +[19] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18409-18418, 2022. 2 +[20] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light, and material decomposition from images using monte carlo rendering and denoising. Advances in Neural Information Processing Systems, 35:22856-22869, 2022. 2 +[21] Jing He, Haodong Li, Wei Yin, Yixun Liang, Leheng Li, Kaiqiang Zhou, Hongbo Zhang, Bingbing Liu, and YingCong Chen. Lotus: Diffusion-based visual foundation model for high-quality dense prediction. arXiv preprint arXiv:2409.18124, 2024. 3, 4 +[22] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. Tensoir: Tensorial inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 165-174, 2023. 2 +[23] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. + +In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 4 +[24] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492-9502, 2024. 3 +[25] Peter Kocsis, Vincent Sitzmann, and Matthias Nießner. Intrinsic image diffusion for indoor single-view material estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5198-5208, 2024. 2, 3, 6 +[26] Zhihao Liang, Qi Zhang, Ying Feng, Ying Shan, and Kui Jia. Gs-ir: 3d gaussian splatting for inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21644–21653, 2024. 2 +[27] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 5404-5411, 2024. 3, 4 +[28] Yehonathan Litman, Or Patashnik, Kangle Deng, Aviral Agrawal, Rushikesh Zawar, Fernando De la Torre, and Shubham Tulsiani. Materialfusion: Enhancing inverse rendering with material diffusion priors. arXiv preprint arXiv:2409.15273, 2024. 1, 2, 3, 4, 6 +[29] Xian Liu, Jian Ren, Aliaksandr Siarohin, Ivan Skorokhodov, Yanyu Li, Dahua Lin, Xihui Liu, Ziwei Liu, and Sergey Tulyakov. Hyperhuman: Hyper-realistic human generation with latent structural diffusion. arXiv preprint arXiv:2310.08579, 2023. 3, 4 +[30] Yunfei Liu, Yu Li, Shaodi You, and Feng Lu. Unsupervised learning for intrinsic image decomposition from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3248-3257, 2020. 3 +[31] Yuxin Liu, Minshan Xie, Hanyuan Liu, and Tien-Tsin Wong. Text-guided texturing by synchronized multi-view diffusion. arXiv preprint arXiv:2311.12891, 2023. 3 +[32] Zexiang Liu, Yangguang Li, Youtian Lin, Xin Yu, Sida Peng, Yan-Pei Cao, Xiaojuan Qi, Xiaoshui Huang, Ding Liang, and Wanli Ouyang. Unidream: Unifying diffusion priors for relightable text-to-3d generation. In European Conference on Computer Vision, pages 74–91. Springer, 2024. 2 +[33] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022. 3 +[34] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8280-8290, 2022. 2 +[35] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, + +Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6 +[36] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3, 5 +[37] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng Bo, and Xiaoguang Han. Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9914–9925, 2024. 2, 3 +[38] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In ACM SIGGRAPH 2023 conference proceedings, pages 1-11, 2023. 3 +[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 4 +[40] Saurabh Saxena, Abhishek Kar, Mohammad Norouzi, and David J Fleet. Monocular depth estimation using diffusion models. arXiv preprint arXiv:2302.14816, 2023. 3 +[41] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 5 +[42] Zeqi Shi, Xiangyu Lin, and Ying Song. An attention-embedded gan for svbrdf recovery from a single image. Computational Visual Media, 9(3):551-561, 2023. 3 +[43] Giuseppe Vecchio, Rosalie Martin, Arthur Roullier, Adrien Kaiser, Romain Rouffet, Valentin Deschaintre, and Tamy Boubekeur. Controlmat: a controlled generative approach to material capture. ACM Transactions on Graphics, 43(5): 1-17, 2024. +[44] Yitong Wang, Xudong Xu, Li Ma, Haoran Wang, and Bo Dai. Boosting 3d object generation through pbr materials. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 3 +[45] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 3, 5 +[46] Felix Wimbauer, Shangzhe Wu, and Christian Rupprecht. De-rendering 3d objects in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18490-18499, 2022. 2, 3, 6 +[47] Bojun Xiong, Jialun Liu, Jiakui Hu, Chenming Wu, Jinbo Wu, Xing Liu, Chen Zhao, Errui Ding, and Zhouhui Lian. Texgaussian: Generating high-quality pbr material via octree-based 3d gaussian splatting. arXiv preprint arXiv:2411.19654, 2024. 2 +[48] Guangkai Xu, Yongtao Ge, Mingyu Liu, Chengxiang Fan, Kangyang Xie, Zhiyue Zhao, Hao Chen, and Chunhua Shen. Diffusion models trained with large data are transferable visual models. arXiv preprint arXiv:2403.06090, 2024. 4 + +[49] Xudong Xu, Zhaoyang Lyu, Xingang Pan, and Bo Dai. Matlab: Material-aware text-to-3d via latent brdf auto-encoder. arXiv preprint arXiv:2308.09278, 2023. 2 +[50] Chongjie Ye, Lingteng Qiu, Xiaodong Gu, Qi Zuo, Yushuang Wu, Zilong Dong, Liefeng Bo, Yuliang Xiu, and Xiaoguang Han. Stablenormal: Reducing diffusion variance for stable and sharp normal. arXiv preprint arXiv:2406.16864, 2024. 3 +[51] Renjiao Yi, Chenyang Zhu, and Kai Xu. Weakly-supervised single-view image relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8402-8411, 2023. 3 +[52] Kim Youwang, Tae-Hyun Oh, and Gerard Pons-Moll. Paint-it: Text-to-texture synthesis via deep convolutional texture map optimization and physically-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4347-4356, 2024. 2, 3 +[53] Xin Yu, Peng Dai, Wenbo Li, Lan Ma, Zhengzhe Liu, and Xiaojuan Qi. Texture generation on 3d meshes with point-uv diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4206–4216, 2023. 2 +[54] Xin Yu, Ze Yuan, Yuan-Chen Guo, Ying-Tian Liu, Jianhui Liu, Yangguang Li, Yan-Pei Cao, Ding Liang, and Xiaojuan Qi. Texgen: a generative diffusion model for mesh textures. ACM Transactions on Graphics (TOG), 43(6):1-14, 2024. +[55] Xianfang Zeng, Xin Chen, Zhongqi Qi, Wen Liu, Zibo Zhao, Zhibin Wang, Bin Fu, Yong Liu, and Gang Yu. Paint3d: Paint anything 3d with lighting-less texture diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4252-4262, 2024. 3 +[56] Zheng Zeng, Valentin Deschaintre, Iliyan Georgiev, Yannick Hold-Geoffroy, Yiwei Hu, Fujun Luan, Ling-Qi Yan, and Miloš Hasan. Rgb ↔ x: Image decomposition and synthesis using material-and lighting-aware diffusion models. In ACM SIGGRAPH 2024 Conference Papers, pages 1–11, 2024. 1, 2, 3, 4, 6 +[57] Kai Zhang, Fujun Luan, Zhengqi Li, and Noah Snavely. Iron: Inverse rendering by optimizing neural sdfs and materials from photometric images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5565-5574, 2022. 2 +[58] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1-20, 2024. 4 +[59] Qian Zhang, Vikas Thamizharasan, and James Tompkin. Learning physically based material and lighting decompositions for face editing. Computational Visual Media, 10(2): 295-308, 2024. 2 +[60] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Ner-factor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG), 40(6):1-18, 2021. +[61] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for + +inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18643-18652, 2022. 2 +[62] Yuqing Zhang, Yuan Liu, Zhiyu Xie, Lei Yang, Zhongyuan Liu, Mengzhou Yang, Runze Zhang, Qilong Kou, Cheng Lin, Wenping Wang, et al. Dreammat: High-quality pbr material generation with geometry-and light-aware diffusion models. ACM Transactions on Graphics (TOG), 43(4):1-18, 2024. 2, 3, 5 +[63] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5729-5739, 2023. 3 +[64] Jingsen Zhu, Fujun Luan, Yuchi Huo, Zihao Lin, Zhihua Zhong, Dianbing Xi, Rui Wang, Hujun Bao, Jiaxiang Zheng, and Rui Tang. Learning-based inverse rendering of complex indoor scenes with differentiable monte carlo raytracing. In SIGGRAPH Asia 2022 Conference Papers, pages 1-8, 2022. 2, 3, 6 \ No newline at end of file diff --git a/2025/SuperMat_ Physically Consistent PBR Material Estimation at Interactive Rates/images.zip b/2025/SuperMat_ Physically Consistent PBR Material Estimation at Interactive Rates/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..441692f8d4720ea593048cf9d0add8316b8ad585 --- /dev/null +++ b/2025/SuperMat_ Physically Consistent PBR Material Estimation at Interactive Rates/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5095c78633128bc4c4438ed62a50c7f38e7416192d5db15694ac26f92400b00 +size 658315 diff --git a/2025/SuperMat_ Physically Consistent PBR Material Estimation at Interactive Rates/layout.json b/2025/SuperMat_ Physically Consistent PBR Material Estimation at Interactive Rates/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..61887fbe781dd6b8cd11e0cdccda8f6be1ce57f2 --- /dev/null +++ b/2025/SuperMat_ Physically Consistent PBR Material Estimation at Interactive Rates/layout.json @@ -0,0 +1,8227 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 60, + 103, + 550, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 103, + 550, + 121 + ], + "spans": [ + { + "bbox": [ + 60, + 103, + 550, + 121 + ], + "type": "text", + "content": "SuperMat: Physically Consistent PBR Material Estimation at Interactive Rates" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 142, + 543, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 142, + 543, + 184 + ], + "spans": [ + { + "bbox": [ + 66, + 142, + 543, + 184 + ], + "type": "text", + "content": "Yijia Hong" + }, + { + "bbox": [ + 66, + 142, + 543, + 184 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 66, + 142, + 543, + 184 + ], + "type": "text", + "content": ", Yuan-Chen Guo" + }, + { + "bbox": [ + 66, + 142, + 543, + 184 + ], + "type": "inline_equation", + "content": "^{4*}" + }, + { + "bbox": [ + 66, + 142, + 543, + 184 + ], + "type": "text", + "content": ", Ran Yi" + }, + { + "bbox": [ + 66, + 142, + 543, + 184 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 66, + 142, + 543, + 184 + ], + "type": "text", + "content": ", Yulong Chen" + }, + { + "bbox": [ + 66, + 142, + 543, + 184 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 66, + 142, + 543, + 184 + ], + "type": "text", + "content": ", Yan-Pei Cao" + }, + { + "bbox": [ + 66, + 142, + 543, + 184 + ], + "type": "inline_equation", + "content": "^{4\\dagger}" + }, + { + "bbox": [ + 66, + 142, + 543, + 184 + ], + "type": "text", + "content": ", Lizhuang Ma" + }, + { + "bbox": [ + 66, + 142, + 543, + 184 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 66, + 142, + 543, + 184 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 66, + 142, + 543, + 184 + ], + "type": "text", + "content": "Shanghai Jiao Tong University " + }, + { + "bbox": [ + 66, + 142, + 543, + 184 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 66, + 142, + 543, + 184 + ], + "type": "text", + "content": "Shanghai Innovation Institute " + }, + { + "bbox": [ + 66, + 142, + 543, + 184 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 66, + 142, + 543, + 184 + ], + "type": "text", + "content": "Harbin Institute of Technology " + }, + { + "bbox": [ + 66, + 142, + 543, + 184 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 66, + 142, + 543, + 184 + ], + "type": "text", + "content": "VAST" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 63, + 203, + 106, + 266 + ], + "blocks": [ + { + "bbox": [ + 63, + 203, + 106, + 266 + ], + "lines": [ + { + "bbox": [ + 63, + 203, + 106, + 266 + ], + "spans": [ + { + "bbox": [ + 63, + 203, + 106, + 266 + ], + "type": "image", + "image_path": "3e5b4d318e19304306eb1958d0d7d951e93446b05fc63057ec4f4215243a8b0d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 63, + 276, + 106, + 337 + ], + "blocks": [ + { + "bbox": [ + 63, + 276, + 106, + 337 + ], + "lines": [ + { + "bbox": [ + 63, + 276, + 106, + 337 + ], + "spans": [ + { + "bbox": [ + 63, + 276, + 106, + 337 + ], + "type": "image", + "image_path": "ac9cec3ae7531345cd3634bebdbb2ff5f7ace3c03d12bedd0a65707407c8ba06.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 63, + 349, + 106, + 410 + ], + "blocks": [ + { + "bbox": [ + 63, + 349, + 106, + 410 + ], + "lines": [ + { + "bbox": [ + 63, + 349, + 106, + 410 + ], + "spans": [ + { + "bbox": [ + 63, + 349, + 106, + 410 + ], + "type": "image", + "image_path": "4520b402f018699f03a66ba1caa3806275fe221ca91fc783f720ff6d11642c39.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 423, + 555, + 479 + ], + "lines": [ + { + "bbox": [ + 55, + 423, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 555, + 479 + ], + "type": "text", + "content": "Figure 1. We present an efficient method for high-quality material decomposition in both 2D and 3D. Left: Albedo and RM(Roughness, Metallic) predictions of SuperMat on rendered images under unknown illumination. The first row shows inputs with uneven lighting, the second row displays albedo predictions, and the third row presents RM predictions. Each texel in the RM map here is represented as (1, roughness, metallic). Middle: Rendering results of a 3D object with PBR materials obtained by our method under novel lighting. Right: Our method is efficient and produces high-quality results." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 123, + 205, + 196, + 265 + ], + "blocks": [ + { + "bbox": [ + 123, + 205, + 196, + 265 + ], + "lines": [ + { + "bbox": [ + 123, + 205, + 196, + 265 + ], + "spans": [ + { + "bbox": [ + 123, + 205, + 196, + 265 + ], + "type": "image", + "image_path": "b83e380c2526df665c449a363951bdaff9dabd27d693a821ac70d7311bab05d0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 122, + 276, + 200, + 337 + ], + "blocks": [ + { + "bbox": [ + 122, + 276, + 200, + 337 + ], + "lines": [ + { + "bbox": [ + 122, + 276, + 200, + 337 + ], + "spans": [ + { + "bbox": [ + 122, + 276, + 200, + 337 + ], + "type": "image", + "image_path": "c158dd331366d0f7b9ddd0284510517fa9fb60d09304d968a47b24a8015ac1cb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 122, + 349, + 200, + 410 + ], + "blocks": [ + { + "bbox": [ + 122, + 349, + 200, + 410 + ], + "lines": [ + { + "bbox": [ + 122, + 349, + 200, + 410 + ], + "spans": [ + { + "bbox": [ + 122, + 349, + 200, + 410 + ], + "type": "image", + "image_path": "a7c53e92dbf3047f3ba8a0b951ebffb9d237c083849097cf153d784e56588639.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 213, + 206, + 254, + 263 + ], + "blocks": [ + { + "bbox": [ + 213, + 206, + 254, + 263 + ], + "lines": [ + { + "bbox": [ + 213, + 206, + 254, + 263 + ], + "spans": [ + { + "bbox": [ + 213, + 206, + 254, + 263 + ], + "type": "image", + "image_path": "e3bac2274340d9958e82cde9780f7c6c43f866b6e0b5d6131e9f14114a9ef305.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 213, + 276, + 254, + 336 + ], + "blocks": [ + { + "bbox": [ + 213, + 276, + 254, + 336 + ], + "lines": [ + { + "bbox": [ + 213, + 276, + 254, + 336 + ], + "spans": [ + { + "bbox": [ + 213, + 276, + 254, + 336 + ], + "type": "image", + "image_path": "017dce4215003c430b1d983730ce04e2689dd74c3b88dd336210e92b6319ccd4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 213, + 349, + 254, + 409 + ], + "blocks": [ + { + "bbox": [ + 213, + 349, + 254, + 409 + ], + "lines": [ + { + "bbox": [ + 213, + 349, + 254, + 409 + ], + "spans": [ + { + "bbox": [ + 213, + 349, + 254, + 409 + ], + "type": "image", + "image_path": "0efdaa27ba494598c1c885f2c05e0b7740930f1b960fe7679c6cbe2d48106c1c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 271, + 206, + 392, + 271 + ], + "blocks": [ + { + "bbox": [ + 271, + 206, + 392, + 271 + ], + "lines": [ + { + "bbox": [ + 271, + 206, + 392, + 271 + ], + "spans": [ + { + "bbox": [ + 271, + 206, + 392, + 271 + ], + "type": "image", + "image_path": "e61811d1388ec4840c75b64c1c83166b1507d2e96b90c46fdf9c081b8d18d204.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 271, + 274, + 392, + 342 + ], + "blocks": [ + { + "bbox": [ + 271, + 274, + 392, + 342 + ], + "lines": [ + { + "bbox": [ + 271, + 274, + 392, + 342 + ], + "spans": [ + { + "bbox": [ + 271, + 274, + 392, + 342 + ], + "type": "image", + "image_path": "ae099eae91dbbf5d0853adfdda42b1c2261fe7d706b5383304092d926c55aa05.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 271, + 341, + 392, + 409 + ], + "blocks": [ + { + "bbox": [ + 271, + 341, + 392, + 409 + ], + "lines": [ + { + "bbox": [ + 271, + 341, + 392, + 409 + ], + "spans": [ + { + "bbox": [ + 271, + 341, + 392, + 409 + ], + "type": "image", + "image_path": "ba3683dbf0d29127d3655e43ff5b9c6e799083d150a131edd86d7301a837d089.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 399, + 205, + 547, + 413 + ], + "blocks": [ + { + "bbox": [ + 399, + 205, + 547, + 413 + ], + "lines": [ + { + "bbox": [ + 399, + 205, + 547, + 413 + ], + "spans": [ + { + "bbox": [ + 399, + 205, + 547, + 413 + ], + "type": "image", + "image_path": "ab298a554890b80d69831a17f26157075d2040815fbc9a60f9a532a889caeac2.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 151, + 489, + 200, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 489, + 200, + 502 + ], + "spans": [ + { + "bbox": [ + 151, + 489, + 200, + 502 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 54, + 514, + 297, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 514, + 297, + 682 + ], + "spans": [ + { + "bbox": [ + 54, + 514, + 297, + 682 + ], + "type": "text", + "content": "Decomposing physically-based materials from images into their constituent properties remains challenging, particularly when maintaining both computational efficiency and physical consistency. While recent diffusion-based approaches have shown promise, they face substantial computational overhead due to multiple denoising steps and separate models for different material properties. We present SuperMat, a single-step framework that achieves high-quality material decomposition with one-step inference. This enables end-to-end training with perceptual and re-render losses while decomposing albedo, metallic, and roughness maps at millisecond-scale speeds. We further extend our framework to 3D objects through a UV refinement network, enabling consistent material estimation across viewpoints" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 490, + 555, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 490, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 490, + 555, + 574 + ], + "type": "text", + "content": "while maintaining efficiency. Experiments demonstrate that SuperMat achieves state-of-the-art PBR material decomposition quality while reducing inference time from seconds to milliseconds per image, and completes PBR material estimation for 3D objects in approximately 3 seconds. The project page is at https://hyj542682306.github.io/SuperMat/." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 597, + 395, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 597, + 395, + 609 + ], + "spans": [ + { + "bbox": [ + 314, + 597, + 395, + 609 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 617, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 617, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 617, + 555, + 715 + ], + "type": "text", + "content": "Decomposing Physically-Based Rendering (PBR) materials from images into their constituent properties (e.g., albedo, metallic, and roughness maps) remains a central challenge in computer vision and graphics. This task is essential for creating high-fidelity 3D assets across numerous applications, from film production and game development to virtual and augmented reality pipelines. While recent advances have shown promising results [6, 28, 56], the problem re" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 111, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 111, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 111, + 703 + ], + "type": "text", + "content": "*Project lead" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25083" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "mains challenging due to its ill-posed nature, particularly given unknown lighting conditions and complex interactions between different material properties. Our work addresses two key objectives: 1) efficient decomposition of high-fidelity material maps from 2D images, and 2) extension of this capability to 3D objects, enabling accurate material estimation across complex geometries." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 159, + 298, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 159, + 298, + 447 + ], + "spans": [ + { + "bbox": [ + 56, + 159, + 298, + 447 + ], + "type": "text", + "content": "Previous works in PBR material estimation generally fall into two main categories: optimization-based generation for 3D objects and data-driven decomposition for 2D images. Optimization-based methods [15, 19, 22, 26] employ differentiable rendering to reconstruct 3D objects with material parameters by minimizing the difference between rendered and target images. Instead, data-driven approaches learn the distribution of various materials from large datasets to perform intrinsic decomposition on 2D images, which is the primary focus of our work. Early methods [46, 64] in this domain fail to produce stable and reliable results due to limited data availability and model capacity. Recent works [6, 25, 28, 56] leverage large diffusion models and reframe material decomposition as a conditional image generation task. However, these methods tend to focus solely on individual material properties, overlooking the interactions between different types of materials and the combined impact on rendered appearances. Even minor inaccuracies in individual material estimates can propagate through the rendering pipeline, leading to substantial visual artifacts. Moreover, these diffusion-based methods face considerable computational overhead - they demand multiple denoising steps (30-50) during inference, causing a critical bottleneck for real-time applications." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 451, + 298, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 298, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 298, + 715 + ], + "type": "text", + "content": "To address these challenges, we propose a deterministic one-step framework that achieves both efficient and high-quality material decomposition. First, we introduce and refine structural expert branches within the UNet architecture. These specialized branches enable concurrent processing of multiple material properties while maintaining domain-specific expertise, allowing our model to decompose albedo, metallic, and roughness maps simultaneously with minimal computational overhead. This design facilitates collaborative and physically consistent material prediction while avoiding the linear cost scaling typically required when estimating multiple material attributes. Second, we address the computational inefficiency of traditional diffusion models by reformulating the DDIM sampling process. Through modification of the scheduler, we achieve single-step inference without compromising quality, dramatically reducing inference time. Most importantly, this single-step formulation enables end-to-end training of the diffusion model, allowing us to compute the perceptual and re-render loss on final predicted material textures, which enhances shadow removal and improves the joint performance of decomposed materials under varying lighting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "type": "text", + "content": "conditions. These advances lead to SuperMat, a highly efficient image space material decomposition model that estimates high-fidelity albedo, metallic, and roughness maps at millisecond-scale speeds." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 121, + 556, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 556, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 556, + 252 + ], + "type": "text", + "content": "For practical 3D applications, we further extend SuperMat's capabilities to handle 3D objects. We upgrade SuperMat to a multi-view model, SuperMatMV, and design a UV refinement network that processes SuperMatMV's predictions in UV space. This approach addresses the challenge of maintaining consistency and completeness across different, limited viewpoints while preserving the efficiency of our original 2D approach. Our pipeline processes a textured 3D model in approximately 3 seconds, producing high-quality material maps that maintain consistency across arbitrary viewpoints and lighting conditions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 325, + 252, + 468, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 252, + 468, + 262 + ], + "spans": [ + { + "bbox": [ + 325, + 252, + 468, + 262 + ], + "type": "text", + "content": "In summary, our contributions are:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 265, + 553, + 480 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 265, + 553, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 265, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 314, + 265, + 553, + 348 + ], + "type": "text", + "content": "- We propose SuperMat, a novel image space material decomposition model that generates physically consistent albedo, metallic, and roughness maps simultaneously with a single-step inference process. Our method achieves state-of-the-art performance on diverse datasets, significantly outperforming previous diffusion-based works in speed." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 349, + 553, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 349, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 314, + 349, + 553, + 420 + ], + "type": "text", + "content": "- With the help of structural expert branches and single-step inference, we develop an end-to-end image space material decomposition training system that generates clean predictions for all material properties in each iteration, allowing for the integration of advanced techniques such as perceptual loss and re-render loss." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 421, + 553, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 421, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 314, + 421, + 553, + 480 + ], + "type": "text", + "content": "- We propose an efficient PBR material decomposition method for 3D models. Combining SuperMatMV and our UV refinement network, this pipeline generates high-quality PBR materials for 3D objects from baked textures in approximately 3 seconds." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 490, + 406, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 490, + 406, + 502 + ], + "spans": [ + { + "bbox": [ + 314, + 490, + 406, + 502 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": "Material Generation and Decomposition. Recognizing the importance of material information in 3D assets, researchers have increasingly focused on material acquisition for objects and scenes. On one hand, some 3D generation approaches [5, 9, 32, 37, 47, 49, 52, 53, 62] integrate material optimization into the appearance creation process, concentrating on generating PBR maps for textured models. On the other hand, inverse rendering methods aim to reconstruct material-rich 3D representations from single or multi-view images. Some of these methods [7, 19, 57] assume a predefined lighting model to enhance estimation accuracy, but this constraint limits their practical applicability. To address this, other works [2, 15, 20, 22, 26, 34, 59-61] attempt to recover both material properties and unknown illumination simultaneously. As reliance on predefined assumptions decreases, researchers have found that beyond controlling material decomposition through rendered out" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25084" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 74, + 544, + 240 + ], + "blocks": [ + { + "bbox": [ + 67, + 74, + 544, + 240 + ], + "lines": [ + { + "bbox": [ + 67, + 74, + 544, + 240 + ], + "spans": [ + { + "bbox": [ + 67, + 74, + 544, + 240 + ], + "type": "image", + "image_path": "e5cff5141d57d26bbfc4c82c263aaf326ddf2eceeedb537c90e47a1380ac6e63.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 250, + 555, + 307 + ], + "lines": [ + { + "bbox": [ + 54, + 250, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 54, + 250, + 555, + 307 + ], + "type": "text", + "content": "Figure 2. Overview of the SuperMat framework. The UNet architecture incorporates structural expert branches for albedo and RM estimation, enabling parallel material property prediction while sharing a common backbone. During training, we leverage a fixed DDIM scheduler and optimize the network end-to-end using both perceptual loss and re-render loss. The marked region shows the inference pipeline, where our model achieves single-step inference for concurrent material map generation. Input geometric information (normal, position) and environment maps are used to compute the re-render loss." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 313, + 294, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 313, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 54, + 313, + 294, + 409 + ], + "type": "text", + "content": "puts, directly obtaining image space material priors is also feasible. Building on this idea, similar to ours, another body of research [6, 25, 28, 30, 42-44, 46, 51, 56, 64] focuses on material decomposition from 2D images captured or rendered under unknown lighting conditions. However, existing methods still face challenges related to training data, model architecture, and training strategies, leaving significant room for improvement in decomposition quality." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 411, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 295, + 590 + ], + "type": "text", + "content": "Diffusion Models. Recently, researchers have applied diffusion models to a range of tasks [5, 36, 37, 45, 62], demonstrating the effectiveness of casting outcome prediction as a conditional image generation process. By leveraging prior knowledge from large pre-trained models, recent studies [6, 11, 14, 18, 24, 28, 40, 50, 63] have achieved performance improvements over previous methods in tasks like depth, normal, and material estimation. However, incorporating large models introduces efficiency challenges, particularly high memory requirements and the time-intensive, multistep denoising process needed for inference. Inspired by insights from recent work [17, 21, 27, 29], our approach addresses these issues by significantly reducing memory and inference time, improving the feasibility of applying these models in real-world applications." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "Generation in UV Space. UV maps are advantageous for fully encapsulating the appearance information of 3D models, leading several works to explore 3D object manipulation directly within the UV space. By representing the texture of an object as a learnable image, these methods [3, 4, 16, 33, 38, 55] enable texture generation, inpainting, and editing, providing new avenues for 3D tasks and allowing 2D image-space techniques and insights to be applied to 3D appearances. Building on this approach, recent methods [1, 12, 31, 52-55] have applied diffusion models to refine" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 312, + 555, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 350 + ], + "type": "text", + "content": "UV maps with promising results. In this spirit, to extend SuperMat's capability to 3D, we fine-tune a UV-space refinement diffusion model specifically." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 358, + 370, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 370, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 370, + 371 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 379, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 379, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 312, + 379, + 555, + 641 + ], + "type": "text", + "content": "In this paper, we propose SuperMat, an image space material decomposition model fine-tuned from Stable Diffusion that efficiently generates high-quality albedo, metallic, and roughness maps simultaneously. Given a rendered or captured image of an object under an unknown lighting condition from a specific viewpoint, SuperMat generates the corresponding material maps for the model from that same viewpoint. As shown in Figure 2, SuperMat is a 2D imaged-to-image one-step model with structural expert branches that simultaneously generates both albedo and RM maps, supporting single-step inference. Compared to previous diffusion-based methods, this model produces superior results in less time and with reduced memory usage. Our single-step formulation allows direct generation of material maps during training, in contrast to prior diffusion methods that predict noise " + }, + { + "bbox": [ + 312, + 379, + 555, + 641 + ], + "type": "inline_equation", + "content": "(\\epsilon)" + }, + { + "bbox": [ + 312, + 379, + 555, + 641 + ], + "type": "text", + "content": " or velocity (v) terms. This enables end-to-end optimization with both perceptual loss on the generated maps and re-render loss from novel lighting conditions, leading to improved material quality and rendering fidelity. Building upon the single-view model, we further introduce a multi-view version, SuperMatMV, to ensure consistency in predictions across different viewpoints." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "While SuperMat and SuperMatMV achieve efficient image space material decomposition, extending this capability to 3D objects presents additional challenges. Instead of relying on optimization-based approaches like score distillation sampling (SDS) that are computationally intensive and prone to mode collapse, we propose an efficient two-stage" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25085" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "content": "pipeline for material estimation on 3D objects. Given a 3D model with only RGB texture, we first apply SuperMatMV to decompose materials from multiple viewpoints. We then introduce a UV refinement network that consolidates these view-dependent predictions into complete, high-quality UV texture maps through single-step inference. As shown in Figure 3, this pipeline enables PBR material estimation for 3D objects in approximately 3 seconds per model." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 178, + 294, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 178, + 294, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 294, + 202 + ], + "type": "text", + "content": "3.1. SuperMat - Image Space Material Decomposition" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 210, + 296, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 210, + 296, + 497 + ], + "spans": [ + { + "bbox": [ + 56, + 210, + 296, + 497 + ], + "type": "text", + "content": "Previous works [6, 28, 56] have demonstrated that casting material estimation as a conditional latent diffusion process, leveraging knowledge from large pre-trained diffusion models like Stable Diffusion [39], is effective. However, for our task that requires the simultaneous generation of multiple material maps, three key issues remain: 1) High model overhead: Fine-tuning Stable Diffusion with its original architecture can only estimate a single material type. Even when combining metallic and roughness into an RM map, treating it as a single material still requires two diffusion models, one for albedo and another for RM, to meet the rendering model's requirements. This results in doubled training time and increased inference space overhead. 2) Slow inference speed: With the widely used DDIM scheduler, diffusion models require 30 to 50 denoising steps to produce satisfactory results, which can be inefficient and time-consuming. 3) Insufficient decomposition effect: Due to the high computational cost of backpropagation in multistep inference, prior diffusion-based methods trained on predicted noise from single-step denoising rather than directly supervising the fully denoised material output. This limitation restricts the model's ability to specialize in the decomposition task and impedes the application of more task-specific techniques." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "content": "Structural Expert Branches. Our solution to the first problem is to introduce a structural expert branch architecture [29, 58]. Previous methods are limited to using a single diffusion model to estimate one material type. Although training separate models for different material maps is an intuitive approach, it introduces issues such as complex implementation, high overhead, and inconsistencies between materials. Considering that, despite their different representations, materials fundamentally reflect an understanding of object features, we incorporate structural expert branches into the UNet model, allowing a single diffusion model to predict multiple material maps. Specifically, we design two expert branches, one for predicting albedo, and the other for predicting RM. For each expert branch, we replicate the UNet's last UpBlock, and conv_out, while the rest of the UNet architecture remains shared between the branches. The modified model first extracts and transforms encoded image features through the shared modules. These features" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "are then separated into 2 copies by duplicating before the last UpBlock, with each branch ultimately predicting results for its respective domain. The expert branches specialize in material-specific features, while the shared modules focus on commonalities. In this way, we only add a small number of parameters (19,313,604 parameters, " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "2.23\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": " of UNet's total), achieving the same performance as previous methods that would otherwise require twice the scale." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 168, + 553, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 346 + ], + "type": "text", + "content": "Single-Step Inference and End-to-End Training. To address the second problem, we analyze the DDIM scheduler of diffusion models. Previous works [17, 21, 27, 48] have highlighted a flaw in the implementation of the DDIM scheduler, which limits the potential of single-step inference in Stable Diffusion: In the default leading timestep setting of DDIM, for single-step prediction, the model is provided with a timestep " + }, + { + "bbox": [ + 313, + 168, + 553, + 346 + ], + "type": "inline_equation", + "content": "(t = 1)" + }, + { + "bbox": [ + 313, + 168, + 553, + 346 + ], + "type": "text", + "content": " suggesting the input is essentially denoised, yet in reality, the input is pure noise. The discrepancy between the timestep and the actual noise level causes the single-step result to be flawed. To align the timestep with the noise level received by the model, we adjust the scheduler to use the trailing setting " + }, + { + "bbox": [ + 313, + 168, + 553, + 346 + ], + "type": "inline_equation", + "content": "(t = T" + }, + { + "bbox": [ + 313, + 168, + 553, + 346 + ], + "type": "text", + "content": " for single-step), correcting this issue and enabling single-step inference, significantly reducing the time overhead." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 347, + 555, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 555, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 555, + 575 + ], + "type": "text", + "content": "More importantly, after this correction, we can fine-tune the one-step model end-to-end, effectively resolving the third issue. For a diffusion model with the uncorrected timestep setting, fine-tuning the diffusion model end-to-end is impractical, because generating a clean material map typically requires multiple denoising steps, leading to prohibitive computational complexity due to the gradient calculations. As an alternative, previous methods [6, 28, 56] rely on using the noise latent obtained from a single denoising step of the UNet for loss calculation. However, this approach shifts the model's training objective from material decomposition to general image denoising, which hinders specialization for this task. In contrast, our corrected inference pipeline can generate fully denoised material maps in a single denoising step, making end-to-end fine-tuning both feasible and efficient for material decomposition. Specifically, with the clean material estimations, we compute perceptual losses [23] between the predicted albedo map " + }, + { + "bbox": [ + 313, + 347, + 555, + 575 + ], + "type": "inline_equation", + "content": "k_{d}" + }, + { + "bbox": [ + 313, + 347, + 555, + 575 + ], + "type": "text", + "content": ", the RM map " + }, + { + "bbox": [ + 313, + 347, + 555, + 575 + ], + "type": "inline_equation", + "content": "k_{rm}" + }, + { + "bbox": [ + 313, + 347, + 555, + 575 + ], + "type": "text", + "content": ", and the ground truth maps " + }, + { + "bbox": [ + 313, + 347, + 555, + 575 + ], + "type": "inline_equation", + "content": "\\hat{k}_d, \\hat{k}_{rm}" + }, + { + "bbox": [ + 313, + 347, + 555, + 575 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 342, + 578, + 553, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 578, + 553, + 625 + ], + "spans": [ + { + "bbox": [ + 342, + 578, + 553, + 625 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {p} (\\hat {y}, y) = \\sum_ {j} \\frac {1}{C _ {j} H _ {j} W _ {j}} | | \\phi_ {j} (\\hat {y}) - \\phi_ {j} (y) | | _ {1}, \\tag {1} \\\\ \\mathcal {L} _ {m} = \\mathcal {L} _ {p} (\\hat {k} _ {d}, k _ {d}) + \\mathcal {L} _ {p} (\\hat {k} _ {r m}, k _ {r m}), \\\\ \\end{array}", + "image_path": "580ff6f2f79b6201138743f51c2888badc76f18da18601edfbfcab84f3381433.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 630, + 553, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 654 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 630, + 553, + 654 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 630, + 553, + 654 + ], + "type": "text", + "content": " is a 16-layer VGG network, and " + }, + { + "bbox": [ + 313, + 630, + 553, + 654 + ], + "type": "inline_equation", + "content": "\\phi_{j}" + }, + { + "bbox": [ + 313, + 630, + 553, + 654 + ], + "type": "text", + "content": " represents the output of the " + }, + { + "bbox": [ + 313, + 630, + 553, + 654 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 630, + 553, + 654 + ], + "type": "text", + "content": "-th layer, with a shape of " + }, + { + "bbox": [ + 313, + 630, + 553, + 654 + ], + "type": "inline_equation", + "content": "C_j\\times H_j\\times W_j" + }, + { + "bbox": [ + 313, + 630, + 553, + 654 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "type": "text", + "content": "Note that, following the approach of previous work [17], we use \"Zeros\" as the noise type during fine-tuning, meaning no random noise is added." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Re-render Loss in DM Fine-tuning. Data-driven decomposition diffusion models typically focus on the distribu" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25086" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 76, + 547, + 210 + ], + "blocks": [ + { + "bbox": [ + 63, + 76, + 547, + 210 + ], + "lines": [ + { + "bbox": [ + 63, + 76, + 547, + 210 + ], + "spans": [ + { + "bbox": [ + 63, + 76, + 547, + 210 + ], + "type": "image", + "image_path": "cb2e36bb43d648640679bd5d8db2f80b4d494d0fec520f9e542963e1af2b0453.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 220, + 555, + 257 + ], + "lines": [ + { + "bbox": [ + 54, + 220, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 54, + 220, + 555, + 257 + ], + "type": "text", + "content": "Figure 3. Material Decomposition for 3D objects. Applying SuperMat's powerful decomposition to 3D objects, we can obtain an object with high-quality materials in just three inference steps. Due to implementation differences, the texels in the RM map for image space decomposition are (1, roughness, metallic), while in the UV refinement process, they are (0, roughness, metallic)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 261, + 296, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 296, + 574 + ], + "type": "text", + "content": "tion of individual material maps, yet they often overlook the interactions between different types of material maps. In other words, previous methods compute the loss between the predicted values and the ground truth without utilizing material information for further operations or calculations, which could provide stronger guidance for the model. As a result, although the predicted materials may appear close to the ground truth, the rendered output may not align with expectations. This limitation is inherent to their training methods, as they either produce meaningless predicted noise, preventing the application of image-level loss, or fail to generate all material predictions simultaneously. In contrast, our design, which incorporates structural expert branches and single-step prediction, allows us to obtain all denoised material maps during the training iterations. Using this information, we introduce the re-render loss [10], which involves rendering the model in a new lighting environment and comparing the result to the ground truth. Specifically, at each training iteration, we obtain " + }, + { + "bbox": [ + 56, + 261, + 296, + 574 + ], + "type": "inline_equation", + "content": "k_{d}" + }, + { + "bbox": [ + 56, + 261, + 296, + 574 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 261, + 296, + 574 + ], + "type": "inline_equation", + "content": "k_{rm} = (k_{r},k_{m})" + }, + { + "bbox": [ + 56, + 261, + 296, + 574 + ], + "type": "text", + "content": ", along with the true normal map " + }, + { + "bbox": [ + 56, + 261, + 296, + 574 + ], + "type": "inline_equation", + "content": "\\hat{k}_n" + }, + { + "bbox": [ + 56, + 261, + 296, + 574 + ], + "type": "text", + "content": ", position map " + }, + { + "bbox": [ + 56, + 261, + 296, + 574 + ], + "type": "inline_equation", + "content": "\\hat{k}_p" + }, + { + "bbox": [ + 56, + 261, + 296, + 574 + ], + "type": "text", + "content": ", camera viewpoint " + }, + { + "bbox": [ + 56, + 261, + 296, + 574 + ], + "type": "inline_equation", + "content": "\\hat{c}" + }, + { + "bbox": [ + 56, + 261, + 296, + 574 + ], + "type": "text", + "content": ", and an environment light map. Then we use this information and our rendering model " + }, + { + "bbox": [ + 56, + 261, + 296, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 56, + 261, + 296, + 574 + ], + "type": "text", + "content": " to compute the predicted material's performance under new illumination. Similarly, with the groundtruth albedo " + }, + { + "bbox": [ + 56, + 261, + 296, + 574 + ], + "type": "inline_equation", + "content": "\\hat{k}_d" + }, + { + "bbox": [ + 56, + 261, + 296, + 574 + ], + "type": "text", + "content": " and RM " + }, + { + "bbox": [ + 56, + 261, + 296, + 574 + ], + "type": "inline_equation", + "content": "\\hat{k}_{rm}" + }, + { + "bbox": [ + 56, + 261, + 296, + 574 + ], + "type": "text", + "content": ", we compute the ground truth of relighting. The perceptual loss between these two renderings defines our re-render loss:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 587, + 295, + 614 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 587, + 295, + 614 + ], + "spans": [ + { + "bbox": [ + 63, + 587, + 295, + 614 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e} = \\mathcal {L} _ {p} \\left(\\mathcal {R} _ {\\hat {k} _ {n}, \\hat {k} _ {p}, \\hat {c}} \\left(\\hat {k} _ {d}, \\hat {k} _ {m}, \\hat {k} _ {r}\\right), \\mathcal {R} _ {\\hat {k} _ {n}, \\hat {k} _ {p}, \\hat {c}} \\left(k _ {d}, k _ {m}, k _ {r}\\right)\\right). \\tag {2}", + "image_path": "0532a452054348fb2870136bfaeac36110bffc5611eea0424caa8ca7efee5caa.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": "Extension to Multi-view Material Decomposition. We extend SuperMat to a multi-view model to enhance the consistency of generated results across different viewpoints. Specifically, we use a 3D self-attention layer and incorporate camera extrinsics as conditions, following MVDream [41]. The multi-view model, SuperMatMV, is capable of simultaneously decomposing materials for 6 input views, providing more accurate and consistent material predictions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 260, + 453, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 453, + 272 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 453, + 272 + ], + "type": "text", + "content": "3.2. UV Refinement Network" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 277, + 555, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 469 + ], + "type": "text", + "content": "We aim to apply SuperMat or SuperMatMV to decompose material for 3D objects with only RGB textures. Several previous works [5, 36, 45, 62] use techniques like SDS to bridge the gap between dimensions, but such methods are not applicable to our deterministic model, and suffer from slow computational speed and mode seeking issues. Therefore, we explore using backprojection to transfer 2D decomposition results from image space to UV space, enabling their application to 3D models. Notably, both SuperMat and SuperMatMV are finite-view models, meaning that one inference can only decompose the visible region from a specific set of viewpoints. To address this problem, we propose estimating materials from multiple viewpoints and integrating them. With multi-view decomposition results from SuperMatMV, the UV refinement network aims to paint missing areas and improve the quality of the generated materials." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 469, + 556, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 469, + 556, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 469, + 556, + 540 + ], + "type": "text", + "content": "Backprojection and UV Blending. For each viewpoint, we backproject SuperMatMV's decomposition results from image space onto the model's UV texture map, obtaining the partial albedo UV map " + }, + { + "bbox": [ + 313, + 469, + 556, + 540 + ], + "type": "inline_equation", + "content": "C_d^i" + }, + { + "bbox": [ + 313, + 469, + 556, + 540 + ], + "type": "text", + "content": ", RM UV map " + }, + { + "bbox": [ + 313, + 469, + 556, + 540 + ], + "type": "inline_equation", + "content": "C_{rm}^i" + }, + { + "bbox": [ + 313, + 469, + 556, + 540 + ], + "type": "text", + "content": ", and mask UV map " + }, + { + "bbox": [ + 313, + 469, + 556, + 540 + ], + "type": "inline_equation", + "content": "M^i" + }, + { + "bbox": [ + 313, + 469, + 556, + 540 + ], + "type": "text", + "content": " for viewpoint " + }, + { + "bbox": [ + 313, + 469, + 556, + 540 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 469, + 556, + 540 + ], + "type": "text", + "content": ". We then blend the backprojection results from all viewpoints in an averaged manner." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 541, + 468, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 541, + 468, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 541, + 468, + 552 + ], + "type": "text", + "content": "For albedo, this can be formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 408, + 559, + 468, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 559, + 468, + 583 + ], + "spans": [ + { + "bbox": [ + 408, + 559, + 468, + 583 + ], + "type": "interline_equation", + "content": "M = \\sum_ {i} M ^ {i},", + "image_path": "65eb5e69dcd35a3698975d192f355864a4db545ef062b126f9ec5759e0b4069f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 401, + 580, + 553, + 612 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 401, + 580, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 401, + 580, + 553, + 612 + ], + "type": "interline_equation", + "content": "C _ {d} = \\frac {\\sum_ {i} C _ {d} ^ {i}}{(M + \\epsilon)}. \\tag {3}", + "image_path": "4324e1ddc7276aa94aa8f4991a2f545f1310752df5cdd10da7fa312ee3e2fb82.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 617, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 653 + ], + "type": "text", + "content": "The RM follows the same computation method. The blended partial maps " + }, + { + "bbox": [ + 313, + 617, + 555, + 653 + ], + "type": "inline_equation", + "content": "C_d, C_{rm}" + }, + { + "bbox": [ + 313, + 617, + 555, + 653 + ], + "type": "text", + "content": ", and the partial mask " + }, + { + "bbox": [ + 313, + 617, + 555, + 653 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 617, + 555, + 653 + ], + "type": "text", + "content": " are then input into the UV refinement one-step model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "content": "UV Refinement One-Step Model. Due to the limited number of viewpoints, the blended partial maps often have missing areas. To address this issue, we leverage diffusion models for UV refinement: we fine-tune Stable Diffusion into a UV refinement model, which takes a flawed partial material" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25087" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 553, + 232 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 553, + 232 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 553, + 232 + ], + "type": "table", + "html": "
AlbedoMetallicRoughnessRelightingTime(s)↓
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Derender3D [46]20.970.86060.1821---------0.16
IIR [64]21.940.87090.155217.950.76190.291119.730.86390.159620.980.88840.11300.04
IID [25]22.560.89080.128217.380.76150.295620.930.87590.147622.450.90440.10831.45
RGB→X [56]22.300.88600.115415.360.75560.336020.400.84730.159121.510.88610.11323.32
IntrinsicAnything [6]20.940.87880.1520---------6.93
StableMaterial [28]23.440.90190.107620.290.81640.245921.010.88800.132722.560.90430.09790.53
StableMaterialMV [28]24.560.90910.094617.280.78860.282920.860.88380.128623.140.90900.09330.55
SuperMat w/o e2e24.260.89640.109520.790.84440.282620.810.86210.158723.900.90850.09643.09
SuperMat w/o re-render26.700.92360.090624.540.82590.226323.520.90160.136226.410.93780.07020.07
SuperMat27.680.92900.082025.480.86850.222124.250.90570.128127.660.94320.06120.07
SuperMatMV27.560.92890.083926.110.87490.219524.840.90760.120827.640.94310.06280.09
", + "image_path": "568234a5f0e359a20b8679bc598c0708da8cb15529d3c08e5498c8ff3c45f51a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 239, + 555, + 287 + ], + "lines": [ + { + "bbox": [ + 55, + 239, + 555, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 555, + 287 + ], + "type": "text", + "content": "Table 1. Comparison of our method with others on the image space decomposition task. We highlight the best, second-best, and third-best results for each metric. \"SuperMat w/o e2e\" denotes our method without the scheduler fix and perceptual loss for end-to-end training. \"SuperMat w/o re-render\" denotes our method without the re-render loss. The first two methods are non-diffusion-based, while the middle five methods are diffusion-based." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 293, + 294, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 293, + 294, + 388 + ], + "spans": [ + { + "bbox": [ + 54, + 293, + 294, + 388 + ], + "type": "text", + "content": "UV map, a mask UV map, and a position UV map as inputs, and outputs completed and refined material UV maps. Specifically, the material maps are encoded by a VAE and concatenated with downsampled mask and position maps, and then processed by UNet with an expanded 8-channel conv_in. We also modify the model's scheduler to support single-step inference and enable end-to-end training, resulting in a deterministic one-step model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 389, + 295, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 295, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 295, + 570 + ], + "type": "text", + "content": "Material Decomposition Pipeline for 3D Objects. As shown in Fig. 3, with the fine-tuned UV refinement network, we propose a complete pipeline for the material decomposition of 3D objects. Given an object with only RGB texture, we first render it from six orthogonal viewpoints. The rendered results are then batch-processed by SuperMatMV for viewpoint-consistent material decomposition. These materials, including albedo and RM, are backprojected into UV space and blended. The blended partial maps are then refined by the UV refinement one-step model and the final material UV maps are generated. This process requires only three inference steps—one for SuperMatMV, one for refining albedo, and one for refining RM—allowing our pipeline to complete the material decomposition of a 3D object in just 3 seconds." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 584, + 136, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 584, + 136, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 584, + 136, + 597 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "Training Datasets. In this paper, we collect 3D objects from Objaverse [8] and Digital Twin Catalog (DTC) [13] datasets. For each object, we render material maps, normal maps, and position maps from 6 orthogonal viewpoints and generate PBR results under 16 different environment lighting conditions from the same viewpoints. We then filter high-quality data using methods such as DINO score considerations [35], resulting in a dataset of 75k objects that combines both synthetic and real-captured data for the im" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 293, + 555, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 365 + ], + "type": "text", + "content": "age space decomposition task. Following the approach outlined in Sec. 3.2, we further refine this data for the UV refinement network by selecting cases where the missing area in blended partial UV maps is under " + }, + { + "bbox": [ + 313, + 293, + 555, + 365 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 293, + 555, + 365 + ], + "type": "text", + "content": ". This process results in a final dataset of 40k objects for training the refinement one-step model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 365, + 556, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 556, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 556, + 581 + ], + "type": "text", + "content": "Quantitative Comparison of Image Space Decomposition. We randomly select 64 unseen items each from the Objaverse, DTC and BlenderVault [28] datasets. Objaverse and BlenderVault primarily contain synthetic data, whereas DTC consists of scanned objects, offering a source of real-world data. Our method is tested against 7 baseline approaches on these 192 objects. For each object, the model processes 18 inputs corresponding to 3 lighting conditions and 6 viewpoints, predicting 3 (or 1) types of materials and generating 54 (or 18) texture maps in total. All outputs are used to compute metrics, with the final averaged results shown in Tab.1. We use PSNR, SSIM, and LPIPS metrics to evaluate albedo, metallic, and roughness. Additionally, for methods that provide all material types, we compute metrics on the rendering under relighting conditions using the ground truth normals and positions and the predicted values. The results demonstrate that our approach significantly outperforms the SOTA in image space decomposition." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": "Beyond image quality, we compare inference speeds by calculating the total time required for each method to process the test set, normalized by the dataset size and number of predicted material types. All tests are conducted with a batch size of 1. For consistency in speed comparison, IID, StableMaterial and StableMaterialMV, which typically average over 10 predictions per input, are modified to perform a single prediction. The \"Time(s)\" column in Tab.1 reflects the time required by each method to generate a single material prediction on an NVIDIA A100 GPU with a batch size of 1. Traditional, non-diffusion-based methods like Deren" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25088" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 68, + 549, + 348 + ], + "blocks": [ + { + "bbox": [ + 61, + 68, + 549, + 348 + ], + "lines": [ + { + "bbox": [ + 61, + 68, + 549, + 348 + ], + "spans": [ + { + "bbox": [ + 61, + 68, + 549, + 348 + ], + "type": "image", + "image_path": "0963d2da006a4748a87353ac7caf8d3c9664f82da3a59190240f283dfe47a27e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 355, + 555, + 378 + ], + "lines": [ + { + "bbox": [ + 55, + 355, + 555, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 355, + 555, + 378 + ], + "type": "text", + "content": "Figure 4. Comparison of our method with others on albedo estimation results. Each pair of rows presents results of the same object under identical lighting conditions from different viewpoints. This organizational format is also followed in other qualitative comparison figures." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 385, + 293, + 555 + ], + "blocks": [ + { + "bbox": [ + 59, + 385, + 293, + 555 + ], + "lines": [ + { + "bbox": [ + 59, + 385, + 293, + 555 + ], + "spans": [ + { + "bbox": [ + 59, + 385, + 293, + 555 + ], + "type": "image", + "image_path": "d419c02f86c1cdf2746513ce077a0742047cf8a2617fab8ba7d6b4ae44886b70.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 568, + 296, + 635 + ], + "lines": [ + { + "bbox": [ + 55, + 568, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 296, + 635 + ], + "type": "text", + "content": "Figure 5. Comparison of our method with others on metallic and roughness estimation results. We combined both materials into a single image, with the left half representing metallic and the right half showing roughness. It can be observed that these two materials exhibit more variation and are more challenging to estimate compared to albedo." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "content": "der3D and IIR generally show faster speeds. While Stable-Material attempts to predict multiple materials simultaneously by adding extra channels to conv_in and conv_out, its lack of structural design adjustments limits prediction quality and requires multi-step denoising, resulting in slower performance than traditional methods. In contrast, our ap-" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 319, + 388, + 550, + 472 + ], + "blocks": [ + { + "bbox": [ + 319, + 388, + 550, + 472 + ], + "lines": [ + { + "bbox": [ + 319, + 388, + 550, + 472 + ], + "spans": [ + { + "bbox": [ + 319, + 388, + 550, + 472 + ], + "type": "image", + "image_path": "caddc01cd0d9a5160dd7e6da61d7796d71564ca1c7759cf110261fb014f262f4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 481, + 529, + 492 + ], + "lines": [ + { + "bbox": [ + 339, + 481, + 529, + 492 + ], + "spans": [ + { + "bbox": [ + 339, + 481, + 529, + 492 + ], + "type": "text", + "content": "Figure 6. Results of UV refinement one-step model." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 496, + 555, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 555, + 520 + ], + "type": "text", + "content": "proach achieves high-quality material decomposition with speeds comparable to traditional methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 522, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 713 + ], + "type": "text", + "content": "Qualitative Comparison of Image Space Decomposition. Fig.4 compares albedo prediction results across various cases, including artist-designed objects and real-world scans, showcasing the performance of our method against other approaches. SuperMat demonstrates superior ability in removing highlights and shadows, accurately capturing the intrinsic color, even on challenging metallic surfaces. In Fig.5, we present metallic and roughness predictions from our method alongside other baselines. SuperMat outperforms the others with well-defined boundaries between metallic and non-metallic regions and achieves consistency within material types while preserving high-frequency details. Moreover, our results closely align with physical realism, addressing a significant challenge that previous methods struggled to overcome. Additionally, SuperMatMV demonstrates stronger consistency across different views in" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25089" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 70, + 123, + 179 + ], + "blocks": [ + { + "bbox": [ + 97, + 70, + 123, + 179 + ], + "lines": [ + { + "bbox": [ + 97, + 70, + 123, + 179 + ], + "spans": [ + { + "bbox": [ + 97, + 70, + 123, + 179 + ], + "type": "image", + "image_path": "fd0c9c4644be9708947a63a51562392785b83d85421a2b8aa104b997b5c8b44c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 103, + 183, + 115, + 192 + ], + "lines": [ + { + "bbox": [ + 103, + 183, + 115, + 192 + ], + "spans": [ + { + "bbox": [ + 103, + 183, + 115, + 192 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 150, + 72, + 179, + 181 + ], + "blocks": [ + { + "bbox": [ + 150, + 72, + 179, + 181 + ], + "lines": [ + { + "bbox": [ + 150, + 72, + 179, + 181 + ], + "spans": [ + { + "bbox": [ + 150, + 72, + 179, + 181 + ], + "type": "image", + "image_path": "d2308a59b1d9f190207f17567be32b59dae11d12adef051fb6644b7ebdcdd361.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 152, + 183, + 179, + 192 + ], + "lines": [ + { + "bbox": [ + 152, + 183, + 179, + 192 + ], + "spans": [ + { + "bbox": [ + 152, + 183, + 179, + 192 + ], + "type": "text", + "content": "w/o e2e" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 209, + 72, + 233, + 181 + ], + "blocks": [ + { + "bbox": [ + 209, + 72, + 233, + 181 + ], + "lines": [ + { + "bbox": [ + 209, + 72, + 233, + 181 + ], + "spans": [ + { + "bbox": [ + 209, + 72, + 233, + 181 + ], + "type": "image", + "image_path": "6f042577cea02fb38ef6cc9537426081f79117bc0eb14f3dd3d70a16e104bb1d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 183, + 243, + 192 + ], + "lines": [ + { + "bbox": [ + 198, + 183, + 243, + 192 + ], + "spans": [ + { + "bbox": [ + 198, + 183, + 243, + 192 + ], + "type": "text", + "content": "w/o re-render" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 264, + 72, + 289, + 181 + ], + "blocks": [ + { + "bbox": [ + 264, + 72, + 289, + 181 + ], + "lines": [ + { + "bbox": [ + 264, + 72, + 289, + 181 + ], + "spans": [ + { + "bbox": [ + 264, + 72, + 289, + 181 + ], + "type": "image", + "image_path": "f0c0c57f66b0181e0f98609554254a5736e65905d9188bdd943006a041b16f50.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 268, + 183, + 285, + 192 + ], + "lines": [ + { + "bbox": [ + 268, + 183, + 285, + 192 + ], + "spans": [ + { + "bbox": [ + 268, + 183, + 285, + 192 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 83, + 350, + 110 + ], + "blocks": [ + { + "bbox": [ + 316, + 83, + 350, + 110 + ], + "lines": [ + { + "bbox": [ + 316, + 83, + 350, + 110 + ], + "spans": [ + { + "bbox": [ + 316, + 83, + 350, + 110 + ], + "type": "image", + "image_path": "ef1c30d76a7fd70142c7902df3661335c4a83069405c13af5f8ceb935bf0b972.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 123, + 350, + 182 + ], + "blocks": [ + { + "bbox": [ + 316, + 123, + 350, + 182 + ], + "lines": [ + { + "bbox": [ + 316, + 123, + 350, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 123, + 350, + 182 + ], + "type": "image", + "image_path": "fb285aa4f44b71d4662d46cb426871bcb341bc6eeb1b940ed60909d5e272a19a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 183, + 338, + 192 + ], + "lines": [ + { + "bbox": [ + 326, + 183, + 338, + 192 + ], + "spans": [ + { + "bbox": [ + 326, + 183, + 338, + 192 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 372, + 83, + 405, + 111 + ], + "blocks": [ + { + "bbox": [ + 372, + 83, + 405, + 111 + ], + "lines": [ + { + "bbox": [ + 372, + 83, + 405, + 111 + ], + "spans": [ + { + "bbox": [ + 372, + 83, + 405, + 111 + ], + "type": "image", + "image_path": "cc0e17a7dd2eff3316323ec4c6508f27f13746481cdea13483886c0d66a32c59.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 372, + 124, + 405, + 182 + ], + "blocks": [ + { + "bbox": [ + 372, + 124, + 405, + 182 + ], + "lines": [ + { + "bbox": [ + 372, + 124, + 405, + 182 + ], + "spans": [ + { + "bbox": [ + 372, + 124, + 405, + 182 + ], + "type": "image", + "image_path": "0d9b08838ef8b7a4e7db3be82416be5c7c6812124bb5081eb7582b511f33463e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 183, + 402, + 192 + ], + "lines": [ + { + "bbox": [ + 375, + 183, + 402, + 192 + ], + "spans": [ + { + "bbox": [ + 375, + 183, + 402, + 192 + ], + "type": "text", + "content": "w/o e2e" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 427, + 83, + 461, + 111 + ], + "blocks": [ + { + "bbox": [ + 427, + 83, + 461, + 111 + ], + "lines": [ + { + "bbox": [ + 427, + 83, + 461, + 111 + ], + "spans": [ + { + "bbox": [ + 427, + 83, + 461, + 111 + ], + "type": "image", + "image_path": "ff48193d0efa128b7fb9c487c6eec94a95ab6f4f6b005bfd984519935f9f1e4b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 427, + 124, + 461, + 182 + ], + "blocks": [ + { + "bbox": [ + 427, + 124, + 461, + 182 + ], + "lines": [ + { + "bbox": [ + 427, + 124, + 461, + 182 + ], + "spans": [ + { + "bbox": [ + 427, + 124, + 461, + 182 + ], + "type": "image", + "image_path": "53b82535cb740fe0fc1db1ebfd0d127e1baee0b48100cd8087bb6ea25f8fe0c8.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 183, + 466, + 192 + ], + "lines": [ + { + "bbox": [ + 421, + 183, + 466, + 192 + ], + "spans": [ + { + "bbox": [ + 421, + 183, + 466, + 192 + ], + "type": "text", + "content": "w/o re-render" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 484, + 83, + 517, + 111 + ], + "blocks": [ + { + "bbox": [ + 484, + 83, + 517, + 111 + ], + "lines": [ + { + "bbox": [ + 484, + 83, + 517, + 111 + ], + "spans": [ + { + "bbox": [ + 484, + 83, + 517, + 111 + ], + "type": "image", + "image_path": "6d3802c777215d8f231e69f4263da890d69692d7f3b55114ea4c5cf21d979e18.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 484, + 124, + 517, + 182 + ], + "blocks": [ + { + "bbox": [ + 484, + 124, + 517, + 182 + ], + "lines": [ + { + "bbox": [ + 484, + 124, + 517, + 182 + ], + "spans": [ + { + "bbox": [ + 484, + 124, + 517, + 182 + ], + "type": "image", + "image_path": "caf03c59044843896edd2ffb2f1fa8832cfc1d53bb286d40b2f2a4de7946e2e2.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 491, + 183, + 509, + 192 + ], + "lines": [ + { + "bbox": [ + 491, + 183, + 509, + 192 + ], + "spans": [ + { + "bbox": [ + 491, + 183, + 509, + 192 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 61, + 227, + 290, + 369 + ], + "blocks": [ + { + "bbox": [ + 61, + 227, + 290, + 369 + ], + "lines": [ + { + "bbox": [ + 61, + 227, + 290, + 369 + ], + "spans": [ + { + "bbox": [ + 61, + 227, + 290, + 369 + ], + "type": "image", + "image_path": "f5715ed0b3fff20980d02c21c468423010f1713de89020dda29cdeb07c73b447.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 374, + 295, + 407 + ], + "lines": [ + { + "bbox": [ + 55, + 374, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 295, + 407 + ], + "type": "text", + "content": "Figure 8. Results of decomposition for 3D objects. We compare the original and decomposed models under the same illumination and show the relighting outcomes of the new models." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 415, + 295, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 295, + 462 + ], + "type": "text", + "content": "all material predictions, and is still able to generate reasonable decomposition results for inputs from challenging viewpoints with sparse relevant information. More results will be presented in the supplementary materials." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "type": "text", + "content": "Results of UV Refinement and Material Estimation for 3D Objects. Fig.6 illustrates the effectiveness of the UV refinement one-step model in inpainting missing areas in SuperMatMV's outputs. Fig.8 displays the results from our material estimation pipeline for 3D models. The first column shows the rendering of the input models with only RGB texture, where all surfaces exhibit uniform reflectance behavior. After applying our method, the decomposed materials enable physically-based rendering with surface-specific reflectance properties, as demonstrated in the second and third columns under different lighting conditions." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "Ablation Studies. In Tab.1, we present ablation studies on two core components of SuperMat: end-to-end (e2e) training and re-render loss. E2E with perceptual loss significantly enhances model performance and efficiency. In fact, within the e2e framework, we experiment with three types of loss functions, with perceptual loss achieving the best results. This suggests that direct supervision of the fully denoised predicted material maps, rather than latent noise, plays a critical role in enhancing the model's capability for this specific task. Detailed findings are available in the sup" + } + ] + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 346, + 228, + 389, + 281 + ], + "blocks": [ + { + "bbox": [ + 321, + 232, + 331, + 275 + ], + "lines": [ + { + "bbox": [ + 321, + 232, + 331, + 275 + ], + "spans": [ + { + "bbox": [ + 321, + 232, + 331, + 275 + ], + "type": "text", + "content": "o o o o o o o o o o o o o o o o o o o o o o o" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 346, + 228, + 389, + 281 + ], + "lines": [ + { + "bbox": [ + 346, + 228, + 389, + 281 + ], + "spans": [ + { + "bbox": [ + 346, + 228, + 389, + 281 + ], + "type": "image", + "image_path": "614d9a813e56f4e0d3c6d24c235948a90a895d291f6f4aa49372283bf7abbfcd.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 346, + 281, + 389, + 333 + ], + "blocks": [ + { + "bbox": [ + 323, + 306, + 330, + 316 + ], + "lines": [ + { + "bbox": [ + 323, + 306, + 330, + 316 + ], + "spans": [ + { + "bbox": [ + 323, + 306, + 330, + 316 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 346, + 281, + 389, + 333 + ], + "lines": [ + { + "bbox": [ + 346, + 281, + 389, + 333 + ], + "spans": [ + { + "bbox": [ + 346, + 281, + 389, + 333 + ], + "type": "image", + "image_path": "deb5622d2540657eeef09443c3ddd2713601455c6859fbb7f739a8712cd72c33.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 342, + 552, + 354 + ], + "lines": [ + { + "bbox": [ + 315, + 342, + 552, + 354 + ], + "spans": [ + { + "bbox": [ + 315, + 342, + 552, + 354 + ], + "type": "text", + "content": "Figure 9. Qualitative ablation results on UV refinement network." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 420, + 228, + 485, + 276 + ], + "blocks": [ + { + "bbox": [ + 420, + 228, + 485, + 276 + ], + "lines": [ + { + "bbox": [ + 420, + 228, + 485, + 276 + ], + "spans": [ + { + "bbox": [ + 420, + 228, + 485, + 276 + ], + "type": "image", + "image_path": "9becd74f3dc913dfc3d72d1700e5a0c839dae5b197099b2ba294a64528efc3f2.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 421, + 282, + 485, + 331 + ], + "blocks": [ + { + "bbox": [ + 421, + 282, + 485, + 331 + ], + "lines": [ + { + "bbox": [ + 421, + 282, + 485, + 331 + ], + "spans": [ + { + "bbox": [ + 421, + 282, + 485, + 331 + ], + "type": "image", + "image_path": "79f712c9a1557e8195196cea1bf8c045f21cc6f4bf932f72410d450b70007405.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 514, + 226, + 535, + 281 + ], + "blocks": [ + { + "bbox": [ + 64, + 205, + 544, + 217 + ], + "lines": [ + { + "bbox": [ + 64, + 205, + 544, + 217 + ], + "spans": [ + { + "bbox": [ + 64, + 205, + 544, + 217 + ], + "type": "text", + "content": "Figure 7. Qualitative ablation results for SuperMat. The first row shows albedo, while the second row shows metallic and roughness." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 514, + 226, + 535, + 281 + ], + "lines": [ + { + "bbox": [ + 514, + 226, + 535, + 281 + ], + "spans": [ + { + "bbox": [ + 514, + 226, + 535, + 281 + ], + "type": "image", + "image_path": "113c94b69b5af2681749b75397c16c24804295d26a35eef8c8de5c16c324e86e.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 515, + 281, + 536, + 334 + ], + "blocks": [ + { + "bbox": [ + 515, + 281, + 536, + 334 + ], + "lines": [ + { + "bbox": [ + 515, + 281, + 536, + 334 + ], + "spans": [ + { + "bbox": [ + 515, + 281, + 536, + 334 + ], + "type": "image", + "image_path": "8c64046ca37abec96a70c8f7159776ae951810d7b594daa63ceea5abc1dfc8fb.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 360, + 553, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 553, + 455 + ], + "type": "text", + "content": "plementary materials. The addition of re-render loss, facilitated by the e2e framework, further improves prediction quality, particularly under relighting conditions. Qualitative results of these ablations are shown in Fig. 7. The 'w/o e2e' setting, where the model is trained without perceptual and re-render losses, limits its capacity to remove lighting effects effectively. In contrast, training 'w/o re-render' loss leads to minor inaccuracies in fine material details." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 456, + 554, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 554, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 554, + 563 + ], + "type": "text", + "content": "For the decomposition pipeline for 3D objects, we conduct additional ablations on the UV refinement network. Due to limitations from only six orthogonal views, achieving complete texture coverage, accurate merging of overlapping regions, and consistent material transitions are challenging. As shown in Fig.9, omitting the UV refinement leads to incomplete textures, visible color-separation lines, and unnatural transitions between materials, underscoring the importance of this component." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 574, + 392, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 392, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 392, + 586 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "In this paper, we present SuperMat, an innovative image space material decomposition one-step model finetuned from Stable Diffusion with task-specific losses in an end-to-end framework. SuperMat enables simultaneous multi-material prediction and supports single-step inference, achieving state-of-the-art quality and outperforming previous diffusion-based methods in efficiency. Through the proposed UV refinement network, we extend SuperMat to decomposition tasks for 3D objects, where it demonstrates exceptional performance and applicability." + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25090" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 193, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 193, + 295, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 193, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 61, + 193, + 295, + 237 + ], + "type": "text", + "content": "[1] Raphael Bensadoun, Yanir Kleiman, Idan Azuri, Omri Harosh, Andrea Vedaldi, Natalia Neverova, and Oran Gafni. Meta 3d texturegen: Fast and consistent texture generation for 3d objects. arXiv preprint arXiv:2407.02430, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 237, + 295, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 237, + 295, + 292 + ], + "spans": [ + { + "bbox": [ + 61, + 237, + 295, + 292 + ], + "type": "text", + "content": "[2] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan Barron, and Hendrik Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. Advances in Neural Information Processing Systems, 34: 10691-10704, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 293, + 295, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 293, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 61, + 293, + 295, + 337 + ], + "type": "text", + "content": "[3] Mark Boss, Zixuan Huang, Aaryaman Vasishta, and Varun Jampani. Sf3d: Stable fast 3d mesh reconstruction with uv-unwrapping and illumination disentanglement. arXiv preprint arXiv:2408.00653, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 337, + 295, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 337, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 62, + 337, + 295, + 392 + ], + "type": "text", + "content": "[4] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18558-18568, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 392, + 296, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 296, + 446 + ], + "type": "text", + "content": "[5] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 448, + 296, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 448, + 296, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 448, + 296, + 501 + ], + "type": "text", + "content": "[6] Xi Chen, Sida Peng, Dongchen Yang, Yuan Liu, Bowen Pan, Chengfei Lv, and Xiaowei Zhou. Intrinsicanything: Learning diffusion priors for inverse rendering under unknown illumination. arXiv preprint arXiv:2404.11593, 2024. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 502, + 295, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 295, + 547 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 295, + 547 + ], + "type": "text", + "content": "[7] Ziang Cheng, Junxuan Li, and Hongdong Li. Wildlight: In-the-wild inverse rendering with a flashlight. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4305-4314, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 548, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 548, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 62, + 548, + 296, + 613 + ], + "type": "text", + "content": "[8] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 614, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 614, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 62, + 614, + 296, + 669 + ], + "type": "text", + "content": "[9] Kangle Deng, Timothy Omernick, Alexander Weiss, Deva Ramanan, Jun-Yan Zhu, Tinghui Zhou, and Maneesh Agrawala. *Flashtex: Fast relightable mesh texturing with lightcontrolnet*. In European Conference on Computer Vision, pages 90–107. Springer, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[10] Valentin Deschaintre, Miika Aittala, Fredo Durand, George Drettakis, and Adrien Bousseau. Single-image svbrdf capture with a rendering-aware deep network. ACM Transactions on Graphics (ToG), 37(4):1-15, 2018. 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "This work was supported by National Natural Science Foundation of China (72192821, 62472282, 62302297 and 62302296), Young Elite Scientists Sponsorship Program by CAST (2022QNRC001), the Fundamental Research Funds for the Central Universities (YG2023QNA35) and YuCaiKe [2023] (231111310300)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[11] Yiqun Duan, Xianda Guo, and Zheng Zhu. Diffusiondepth: Diffusion denoising approach for monocular depth estimation. arXiv preprint arXiv:2303.05021, 3, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 108, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 162 + ], + "type": "text", + "content": "[12] Xiang Feng, Chang Yu, Zoubin Bi, Yintong Shang, Feng Gao, Hongzhi Wu, Kun Zhou, Chenfanfu Jiang, and Yin Yang. Arm: Appearance reconstruction model for relightable 3d generation. arXiv preprint arXiv:2411.10825, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "text", + "content": "[13] James Fort. Introducing the digital twin catalog from reality labs research, 2024. https://ai.meta.com/blog/digital-twin-catalog-3d-reconstruction-shopify-reality-labs-research/, Last accessed on 2025-03-07. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 219, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 275 + ], + "type": "text", + "content": "[14] Xiao Fu, Wei Yin, Mu Hu, Kaixuan Wang, Yuexin Ma, Ping Tan, Shaojie Shen, Dahua Lin, and Xiaoxiao Long. Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image. In European Conference on Computer Vision, pages 241-258. Springer, 2025. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "text", + "content": "[15] Jian Gao, Chun Gu, Youtian Lin, Hao Zhu, Xun Cao, Li Zhang, and Yao Yao. Relightable 3d gaussian: Real-time point cloud relighting with brdf decomposition and ray tracing. arXiv preprint arXiv:2311.16043, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 320, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 365 + ], + "type": "text", + "content": "[16] Lin Gao, Tong Wu, Yu-Jie Yuan, Ming-Xian Lin, Yu-Kun Lai, and Hao Zhang. Tm-net: Deep generative networks for textured meshes. ACM Transactions on Graphics (TOG), 40 (6):1-15, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 365, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 421 + ], + "type": "text", + "content": "[17] Gonzalo Martin Garcia, Karim Abou Zeid, Christian Schmidt, Daan de Geus, Alexander Hermans, and Bastian Leibe. Fine-tuning image-conditional diffusion models is easier than you think. arXiv preprint arXiv:2409.11355, 2024. 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 422, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 477 + ], + "type": "text", + "content": "[18] Ming Gui, Johannes S Fischer, Ulrich Prestel, Pingchuan Ma, Dmytro Kotovenko, Olga Grebenkova, Stefan Andreas Baumann, Vincent Tao Hu, and Björn Ommer. Depthfm: Fast monocular depth estimation with flow matching. arXiv preprint arXiv:2403.13788, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 478, + 553, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 531 + ], + "type": "text", + "content": "[19] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18409-18418, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 534, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 553, + 578 + ], + "type": "text", + "content": "[20] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light, and material decomposition from images using monte carlo rendering and denoising. Advances in Neural Information Processing Systems, 35:22856-22869, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 578, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 553, + 633 + ], + "type": "text", + "content": "[21] Jing He, Haodong Li, Wei Yin, Yixun Liang, Leheng Li, Kaiqiang Zhou, Hongbo Zhang, Bingbing Liu, and YingCong Chen. Lotus: Diffusion-based visual foundation model for high-quality dense prediction. arXiv preprint arXiv:2409.18124, 2024. 3, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 635, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 690 + ], + "type": "text", + "content": "[22] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. Tensoir: Tensorial inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 165-174, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[23] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25091" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 171 + ], + "type": "text", + "content": "[24] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492-9502, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 295, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 295, + 226 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 295, + 226 + ], + "type": "text", + "content": "[25] Peter Kocsis, Vincent Sitzmann, and Matthias Nießner. Intrinsic image diffusion for indoor single-view material estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5198-5208, 2024. 2, 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 294, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 271 + ], + "type": "text", + "content": "[26] Zhihao Liang, Qi Zhang, Ying Feng, Ying Shan, and Kui Jia. Gs-ir: 3d gaussian splatting for inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21644–21653, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 295, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 295, + 326 + ], + "type": "text", + "content": "[27] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 5404-5411, 2024. 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 295, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 295, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 295, + 381 + ], + "type": "text", + "content": "[28] Yehonathan Litman, Or Patashnik, Kangle Deng, Aviral Agrawal, Rushikesh Zawar, Fernando De la Torre, and Shubham Tulsiani. Materialfusion: Enhancing inverse rendering with material diffusion priors. arXiv preprint arXiv:2409.15273, 2024. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 382, + 295, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 295, + 437 + ], + "type": "text", + "content": "[29] Xian Liu, Jian Ren, Aliaksandr Siarohin, Ivan Skorokhodov, Yanyu Li, Dahua Lin, Xihui Liu, Ziwei Liu, and Sergey Tulyakov. Hyperhuman: Hyper-realistic human generation with latent structural diffusion. arXiv preprint arXiv:2310.08579, 2023. 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 295, + 491 + ], + "type": "text", + "content": "[30] Yunfei Liu, Yu Li, Shaodi You, and Feng Lu. Unsupervised learning for intrinsic image decomposition from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3248-3257, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 492, + 295, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 295, + 525 + ], + "type": "text", + "content": "[31] Yuxin Liu, Minshan Xie, Hanyuan Liu, and Tien-Tsin Wong. Text-guided texturing by synchronized multi-view diffusion. arXiv preprint arXiv:2311.12891, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 525, + 295, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 295, + 581 + ], + "type": "text", + "content": "[32] Zexiang Liu, Yangguang Li, Youtian Lin, Xin Yu, Sida Peng, Yan-Pei Cao, Xiaojuan Qi, Xiaoshui Huang, Ding Liang, and Wanli Ouyang. Unidream: Unifying diffusion priors for relightable text-to-3d generation. In European Conference on Computer Vision, pages 74–91. Springer, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 582, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 295, + 624 + ], + "type": "text", + "content": "[33] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 625, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 295, + 689 + ], + "type": "text", + "content": "[34] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8280-8290, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 690, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 690, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 690, + 295, + 713 + ], + "type": "text", + "content": "[35] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez," + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "type": "text", + "content": "Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "type": "text", + "content": "[36] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 215 + ], + "type": "text", + "content": "[37] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng Bo, and Xiaoguang Han. Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9914–9925, 2024. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "type": "text", + "content": "[38] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In ACM SIGGRAPH 2023 conference proceedings, pages 1-11, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 261, + 555, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 555, + 316 + ], + "type": "text", + "content": "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 316, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 348 + ], + "type": "text", + "content": "[40] Saurabh Saxena, Abhishek Kar, Mohammad Norouzi, and David J Fleet. Monocular depth estimation using diffusion models. arXiv preprint arXiv:2302.14816, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 349, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 349, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 349, + 553, + 381 + ], + "type": "text", + "content": "[41] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 383, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 415 + ], + "type": "text", + "content": "[42] Zeqi Shi, Xiangyu Lin, and Ying Song. An attention-embedded gan for svbrdf recovery from a single image. Computational Visual Media, 9(3):551-561, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 415, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 469 + ], + "type": "text", + "content": "[43] Giuseppe Vecchio, Rosalie Martin, Arthur Roullier, Adrien Kaiser, Romain Rouffet, Valentin Deschaintre, and Tamy Boubekeur. Controlmat: a controlled generative approach to material capture. ACM Transactions on Graphics, 43(5): 1-17, 2024." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "type": "text", + "content": "[44] Yitong Wang, Xudong Xu, Li Ma, Haoran Wang, and Bo Dai. Boosting 3d object generation through pbr materials. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 514, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 553, + 568 + ], + "type": "text", + "content": "[45] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 3, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 570, + 555, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 555, + 613 + ], + "type": "text", + "content": "[46] Felix Wimbauer, Shangzhe Wu, and Christian Rupprecht. De-rendering 3d objects in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18490-18499, 2022. 2, 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 614, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 668 + ], + "type": "text", + "content": "[47] Bojun Xiong, Jialun Liu, Jiakui Hu, Chenming Wu, Jinbo Wu, Xing Liu, Chen Zhao, Errui Ding, and Zhouhui Lian. Texgaussian: Generating high-quality pbr material via octree-based 3d gaussian splatting. arXiv preprint arXiv:2411.19654, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "text", + "content": "[48] Guangkai Xu, Yongtao Ge, Mingyu Liu, Chengxiang Fan, Kangyang Xie, Zhiyue Zhao, Hao Chen, and Chunhua Shen. Diffusion models trained with large data are transferable visual models. arXiv preprint arXiv:2403.06090, 2024. 4" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25092" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[49] Xudong Xu, Zhaoyang Lyu, Xingang Pan, and Bo Dai. Matlab: Material-aware text-to-3d via latent brdf auto-encoder. arXiv preprint arXiv:2308.09278, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 162 + ], + "type": "text", + "content": "[50] Chongjie Ye, Lingteng Qiu, Xiaodong Gu, Qi Zuo, Yushuang Wu, Zilong Dong, Liefeng Bo, Yuliang Xiu, and Xiaoguang Han. Stablenormal: Reducing diffusion variance for stable and sharp normal. arXiv preprint arXiv:2406.16864, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 207 + ], + "type": "text", + "content": "[51] Renjiao Yi, Chenyang Zhu, and Kai Xu. Weakly-supervised single-view image relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8402-8411, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "type": "text", + "content": "[52] Kim Youwang, Tae-Hyun Oh, and Gerard Pons-Moll. Paint-it: Text-to-texture synthesis via deep convolutional texture map optimization and physically-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4347-4356, 2024. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "type": "text", + "content": "[53] Xin Yu, Peng Dai, Wenbo Li, Lan Ma, Zhengzhe Liu, and Xiaojuan Qi. Texture generation on 3d meshes with point-uv diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4206–4216, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 310, + 294, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 294, + 354 + ], + "type": "text", + "content": "[54] Xin Yu, Ze Yuan, Yuan-Chen Guo, Ying-Tian Liu, Jianhui Liu, Yangguang Li, Yan-Pei Cao, Ding Liang, and Xiaojuan Qi. Texgen: a generative diffusion model for mesh textures. ACM Transactions on Graphics (TOG), 43(6):1-14, 2024." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 355, + 294, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 294, + 410 + ], + "type": "text", + "content": "[55] Xianfang Zeng, Xin Chen, Zhongqi Qi, Wen Liu, Zibo Zhao, Zhibin Wang, Bin Fu, Yong Liu, and Gang Yu. Paint3d: Paint anything 3d with lighting-less texture diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4252-4262, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 476 + ], + "type": "text", + "content": "[56] Zheng Zeng, Valentin Deschaintre, Iliyan Georgiev, Yannick Hold-Geoffroy, Yiwei Hu, Fujun Luan, Ling-Qi Yan, and Miloš Hasan. Rgb ↔ x: Image decomposition and synthesis using material-and lighting-aware diffusion models. In ACM SIGGRAPH 2024 Conference Papers, pages 1–11, 2024. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 478, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 533 + ], + "type": "text", + "content": "[57] Kai Zhang, Fujun Luan, Zhengqi Li, and Noah Snavely. Iron: Inverse rendering by optimizing neural sdfs and materials from photometric images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5565-5574, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "type": "text", + "content": "[58] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1-20, 2024. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "type": "text", + "content": "[59] Qian Zhang, Vikas Thamizharasan, and James Tompkin. Learning physically based material and lighting decompositions for face editing. Computational Visual Media, 10(2): 295-308, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[60] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Ner-factor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG), 40(6):1-18, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[61] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 294 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18643-18652, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 172 + ], + "type": "text", + "content": "[62] Yuqing Zhang, Yuan Liu, Zhiyu Xie, Lei Yang, Zhongyuan Liu, Mengzhou Yang, Runze Zhang, Qilong Kou, Cheng Lin, Wenping Wang, et al. Dreammat: High-quality pbr material generation with geometry-and light-aware diffusion models. ACM Transactions on Graphics (TOG), 43(4):1-18, 2024. 2, 3, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 174, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 228 + ], + "type": "text", + "content": "[63] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5729-5739, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 230, + 553, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 294 + ], + "type": "text", + "content": "[64] Jingsen Zhu, Fujun Luan, Yuchi Huo, Zihao Lin, Zhihua Zhong, Dianbing Xi, Rui Wang, Hujun Bao, Jiaxiang Zheng, and Rui Tang. Learning-based inverse rendering of complex indoor scenes with differentiable monte carlo raytracing. In SIGGRAPH Asia 2022 Conference Papers, pages 1-8, 2022. 2, 3, 6" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25093" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Supercharged One-step Text-to-Image Diffusion Models with Negative Prompts/1564dd9e-386a-48aa-bdfb-0018fc932f68_content_list.json b/2025/Supercharged One-step Text-to-Image Diffusion Models with Negative Prompts/1564dd9e-386a-48aa-bdfb-0018fc932f68_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5e0b36602a1fbb794f13911adf929b7deb94e11f --- /dev/null +++ b/2025/Supercharged One-step Text-to-Image Diffusion Models with Negative Prompts/1564dd9e-386a-48aa-bdfb-0018fc932f68_content_list.json @@ -0,0 +1,1442 @@ +[ + { + "type": "text", + "text": "Supercharged One-step Text-to-Image Diffusion Models with Negative Prompts", + "text_level": 1, + "bbox": [ + 98, + 130, + 898, + 155 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Viet Nguyen $^{1\\star}$ , Anh Nguyen $^{1\\star}$ , Trung Dao $^{1}$ , Khoi Nguyen $^{1}$ , Cuong Pham $^{1,2}$ , Toan Tran $^{1}$ , Anh Tran $^{1}$", + "bbox": [ + 102, + 179, + 890, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Qualcomm AI Research† $^{2}$ Posts & Telecom. Inst. of Tech.", + "bbox": [ + 254, + 215, + 743, + 234 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/28f5f00cf99608049d3527edd4096bd0f1698bb8cf333253acd97907f5f819a3.jpg", + "image_caption": [ + "Figure 1. Negative prompting are a key technique for improving controllability and image quality in text-to-image diffusion models. However, classifier-free guidance (CFG) [19] only supports negative prompt integration in multi-step models due to its iterative nature. Our method, NASA, is the first to enable negative prompting in one-step models, allowing for the suppression of unwanted features and providing greater control over image attributes." + ], + "image_footnote": [], + "bbox": [ + 94, + 253, + 906, + 488 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 561, + 326, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The escalating demand for real-time image synthesis has driven significant advancements in one-step diffusion models, which inherently offer expedited generation speeds compared to traditional multi-step methods. However, this enhanced efficiency is frequently accompanied by a compromise in the controllability of image attributes. While negative prompting, typically implemented via classifier-free guidance (CFG), has proven effective for fine-grained control in multi-step models, its application to one-step generators remains largely unaddressed. Due to the lack of iterative refinement, as in multi-step diffusion, directly applying CFG to one-step generation leads to blending artifacts and diminished output quality. To fill this gap, we introduce Negative-Away Steer Attention (NASA), an efficient method that integrates negative prompts into one-step diffusion models. NASA operates within the intermediate representation space by leveraging cross-attention mechanisms to suppress undesired visual attributes. This strategy", + "bbox": [ + 88, + 592, + 485, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "avoids the blending artifacts inherent in output-space guidance and achieves high efficiency, incurring only a minimal $1.89\\%$ increase in FLOPs compared to the computational doubling of CFG. Furthermore, NASA can be seamlessly integrated into existing timestep distillation frameworks, enhancing the student's output quality. Experimental results demonstrate that NASA substantially improves controllability and output quality, achieving an HPSv2 score of 31.21, setting a new state-of-the-art benchmark for one-step diffusion models.", + "bbox": [ + 511, + 561, + 908, + 712 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 751, + 645, + 766 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models have recently become popular due to their capacity to produce high-quality, diverse outputs in image [3, 7, 36, 40], 3D [26, 37, 45], audio [14, 21] and video [5, 15, 49, 50] synthesis. Unlike other generative models, such as Generative Adversarial Networks (GANs) [16, 23], diffusion models gradually refine their output through a series of steps, allowing them to achieve diverse yet impressively detailed and realistic outputs. However, this iterative", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "\\* Equal Contribution", + "bbox": [ + 91, + 875, + 209, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Qualcomm AI Research is an initiative of Qualcomm Technologies, Inc.", + "bbox": [ + 91, + 887, + 480, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18004", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "refinement is both time-consuming and computationally demanding, which limits their practicality in real-world applications. As a result, there is an increasing interest in techniques to speed up diffusion models while maintaining output quality. Recent distillation methods [12, 34, 51, 52] compress the entire multi-step generation process into a single step, a breakthrough that enables real-time image synthesis. This speedup makes diffusion models practical for applications like content creation and interactive media, where their slow performance was previously a major drawback. Despite these impressive speed advancements in one-step generators, we have identified a significant limitation that affects all these models: the inability to support negative prompting [4, 47]. This feature, which allows users to explicitly exclude unwanted elements, is widely available in multi-step diffusion models but remains notably absent in their single-step alternatives.", + "bbox": [ + 89, + 90, + 480, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The most widely adopted approach for negative prompting, classifier-free guidance (CFG) [19] works by generating two parallel noise predictions during each denoising step - one incorporating the positive prompt and another using the negative prompt. The system then subtracts the negative prediction from the positive one, effectively steering generation away from unwanted attributes. This process relies on the iterative nature of multi-step diffusion, where corrections can be applied and refined over dozens of steps.", + "bbox": [ + 89, + 349, + 480, + 486 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One-step generators, however, collapse this entire process into a single forward pass, eliminating the iterative refinement cycles that make CFG possible. Without these multiple correction opportunities, there is no straightforward way to incorporate negative guidance into the generation pipeline. Consequently, users face an unappealing choice between the instant generation that might include unwanted elements or waiting longer for more precisely controlled results. This speed-versus-control dilemma represents the most significant practical barrier to widespread adoption of one-step generators in professional creative workflows, highlighting the urgent need for novel approaches that can bridge this gap.", + "bbox": [ + 89, + 488, + 480, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this issue, we introduce a novel method called Negative-Away Steer Attention (NASA) used in inference, which incorporates negative prompt guidance directly into one-step diffusion models. NASA works by adjusting cross-attention layers, effectively reducing unwanted features in generated images. By steering attention within intermediate representation space, NASA offers image control capabilities that were previously achievable only in multi-step diffusion models.", + "bbox": [ + 89, + 686, + 480, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With our newly developed capability, we have explored integrating negative prompting directly into one-step distillation frameworks, focusing on the SwiftBrush (SB) family of models [12, 34]. The SB approach is compelling for its unique, image-free distillation methodology. By match-", + "bbox": [ + 89, + 824, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing teacher model probability distributions and applying sophisticated training techniques, SBv2 can outperform its own teacher on standard image synthesis benchmarks. Despite these technical achievements, SB models still fall short in human preference evaluations. The core challenge lies in the distillation process: without real image supervision, the system lacks clear signals about which generation patterns should be avoided. This is precisely where negative prompting and our NASA framework come in.", + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Committed to maintaining SB's valuable image-free distillation approach while incorporating guidance sampling to align with human preference, we provide the critical missing component, a way to steer the model away from undesirable outputs during training. NASA enables something previously impossible: guidance sampling within a one-step generator. Through this integration, our model achieves an HPSv2 [48] score of 31.21, establishing a new state-of-the-art on human preference metrics for one-step models.", + "bbox": [ + 511, + 227, + 903, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are twofold:", + "bbox": [ + 532, + 364, + 816, + 378 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce Negative-Away Steer Attention (NASA), the first method to integrate negative prompt guidance into one-step and few-step diffusion models in both inference and distillation settings.", + "- By integrating NASA into the SwiftBrush (SB) distillation framework, we achieve a new state-of-the-art for one-step diffusion models on human preference metrics. This marks the first successful integration of negative prompt guidance directly into one-step model training." + ], + "bbox": [ + 513, + 383, + 903, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 534, + 653, + 549 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Distribution Matching Distillation", + "text_level": 1, + "bbox": [ + 513, + 559, + 808, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent work [12, 34, 51, 52] has explored text-to-image diffusion distillation through distribution matching techniques. These methods directly align the output distributions of teacher and student models, particularly through Variational Score Distillation (VSD). These techniques can be broadly categorized into two approaches: image-dependent training and image-free training.", + "bbox": [ + 511, + 582, + 903, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image-dependent methods rely on real images for adversarial loss [41, 51] or synthetic ones generated by the teacher models for reconstruction loss [52]. Despite their effectiveness in preserving image quality, these approaches face significant practical limitations: they require substantial storage capacity, consume high memory during training, demand extensive computational resources, and risk mode collapse due to intensive image supervision.", + "bbox": [ + 511, + 688, + 903, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, image-free approaches [12, 34] have demonstrated promising results by relying solely on prompt text for training, significantly reducing resource requirements while maintaining competitive performance. However, image-free techniques often encounter training instability, particularly within VSD-based frameworks. Existing solu", + "bbox": [ + 511, + 810, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18005", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tions to stabilize training [52] introduce significant memory and computational overhead, limiting their practical application. Our work addresses this gap by introducing a simple yet effective method that enhances training stability without additional computational burden, making image-free variational score distillation produce better models.", + "bbox": [ + 89, + 90, + 483, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Sampling Guidance with Negative Prompts", + "text_level": 1, + "bbox": [ + 89, + 188, + 457, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Classifier-free guidance (CFG) [19] and negative prompting [4, 47] are key techniques for content control during denoising. In multi-step diffusion models, these methods enhance desired image features and suppress undesired ones, improving visual fidelity and reducing artifacts [1, 13, 24]. Score-based approaches like NFSD [24] employ negative prompting to steer generated samples toward real-image distributions, while SDS-Bridge [33] uses negative prompts to model source distributions more accurately. Despite their success, negative prompting has yet to be integrated into one-step or few-step diffusion models. To bridge this gap, we introduce NASA, the first method to enable negative prompting in one-step and few-step diffusion models, enhancing control in faster generation settings.", + "bbox": [ + 89, + 210, + 483, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 433, + 228, + 449 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Models [20, 39, 43] produce high-quality images by progressively denoising inputs. This technique involves a forward process that gradually adds noise to the data and a reverse process that reconstructs the original data by removing noise. While the original diffusion models act on the image space [20, 43], LDM [39] processes on the latent space produced by a pretrained VAE for more efficient computation. In particular, starting with a data point $\\mathbf{x}_0$ sampled from an unknown distribution $q(\\mathbf{x}_0)$ , the forward process gradually diffuses $\\mathbf{x}_0$ into a standard Gaussian noise $\\mathbf{x}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$ through $T$ consecutive timesteps, where $\\mathbf{I}$ is the identity matrix. At each time-step $t$ , a noisier version of $\\mathbf{x}_0$ , denoted as $\\mathbf{x}_t$ , is drawn from $q_t(\\mathbf{x}_t | \\mathbf{x}_0) = \\mathcal{N}(\\alpha_t \\mathbf{x}_0, \\sigma_t^2 \\mathbf{I})$ using a standard Gaussian noise $\\epsilon \\sim \\mathcal{N}(0, \\mathbf{I})$ :", + "bbox": [ + 89, + 458, + 483, + 671 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {t} = \\alpha_ {t} \\mathbf {x} _ {0} + \\sigma_ {t} \\boldsymbol {\\epsilon}, \\quad \\forall t \\in \\{0, \\dots , T \\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 676, + 480, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\{(\\alpha_t,\\sigma_t)\\}_{t = 1}^T$ defines the noise schedule, with boundary conditions $(\\alpha_0,\\sigma_0) = (1,0)$ for the clean sample and $(\\alpha_{T},\\sigma_{T}) = (0,1)$ for pure noise.", + "bbox": [ + 89, + 698, + 482, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Conversely, the reverse process aims to gradually reconstruct the original data by denoising the input over $T$ steps, starting from an initial random Gaussian noise $\\mathbf{x}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ . The model is trained by minimizing the difference between the noise estimated by the model, $\\epsilon_{\\phi}$ , parameterized by $\\phi$ , and the actual noise in Eq. (1):", + "bbox": [ + 89, + 744, + 483, + 835 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\phi} \\mathbb {E} _ {t \\sim \\mathcal {U} (0, T), \\epsilon \\sim \\mathcal {N} (0, I)} \\| \\epsilon_ {\\phi} (\\mathbf {x} _ {t}, t) - \\boldsymbol {\\epsilon} \\| _ {2} ^ {2}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 840, + 482, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-to-Image Diffusion Models guide the sampling process using an additional text prompt y to produce outputs", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "that are both photorealistic and aligned with the provided textual descriptions. The training objective, slightly modified from the unconditional loss in Eq. (2), is defined as:", + "bbox": [ + 511, + 90, + 906, + 137 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\phi} \\mathbb {E} _ {t \\sim \\mathcal {U} (0, T), \\epsilon \\sim \\mathcal {N} (0, I), \\mathbf {y}} \\| \\epsilon_ {\\phi} (\\mathbf {x} _ {t}, t, \\mathbf {y}) - \\epsilon \\| _ {2} ^ {2}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 150, + 906, + 174 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Classifier-free Guidance (CFG) [19] is an inference technique designed to enhance the quality of generated images by blending the predictions from both a conditional and an unconditional model. At each sampling step, CFG adjusts the denoiser's output using a control parameter $\\kappa > 1$ , allowing controlled guidance that aligns more closely with the desired conditions:", + "bbox": [ + 511, + 181, + 906, + 287 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {\\epsilon}} _ {\\phi} (\\mathbf {x} _ {t}, t, \\mathbf {y}, \\kappa) = (1 - \\kappa) \\boldsymbol {\\epsilon} _ {\\phi} (\\mathbf {x} _ {t}, t) + \\kappa \\boldsymbol {\\epsilon} _ {\\phi} (\\mathbf {x} _ {t}, t, \\mathbf {y}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 299, + 906, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Negative Prompting [4, 47] provides enhanced control by suppressing unwanted features in the generated content. Instead of producing an unconditional output, the model generates an output conditioned on the negative prompt $\\mathbf{y}^{-}$ , as follows:", + "bbox": [ + 511, + 328, + 905, + 402 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {\\epsilon}} _ {\\phi} \\left(\\mathbf {x} _ {t}, t, \\mathbf {y}, \\mathbf {y} ^ {-}, \\kappa\\right) = (1 - \\kappa) \\boldsymbol {\\epsilon} _ {\\phi} \\left(\\mathbf {x} _ {t}, t, \\mathbf {y} ^ {-}\\right) + \\kappa \\boldsymbol {\\epsilon} _ {\\phi} \\left(\\mathbf {x} _ {t}, t, \\mathbf {y}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 416, + 905, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Variational Score Distillation (VSD) is a powerful framework that utilizes pretrained diffusion-based text-to-image models to enhance text-based generation across both 3D [45] and 2D [12, 34, 51, 52] domains. The central aim of VSD is to align the renderings of a differentiable generator with the probability density of plausible images as guided by the 2D diffusion model. To accomplish this, VSD employs a two-teacher approach that uses a fixed pretrained diffusion model $\\epsilon_{\\psi}$ and an adaptive LoRA-based teacher $\\epsilon_{\\pi}$ . While training, the student model $f_{\\theta}$ produces 2D images $\\hat{\\mathbf{x}}_0 = f_{\\theta}(\\mathbf{z},\\mathbf{y})$ using an input noise $\\mathbf{z} \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ and a text prompt $\\mathbf{y}$ . Noisy images $\\hat{\\mathbf{x}}_t = \\alpha_t\\hat{\\mathbf{x}}_0 + \\sigma_t\\epsilon$ are then fed into both teacher models. The LoRA teacher model $\\epsilon_{\\pi}$ aligns with the student distribution by minimizing a denoising L2 loss on single-step samples. This arrangement supports robust and adaptive guidance suitable for a variety of generative architectures. The gradient of the learning objective with respect to the student's parameters $\\theta$ is defined as:", + "bbox": [ + 511, + 446, + 906, + 719 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {V S D}} = \\mathbb {E} _ {t, \\mathbf {y}, \\mathbf {z}, \\kappa = C} \\left[ w (t) \\left(\\hat {\\epsilon} _ {\\psi} (\\hat {\\mathbf {x}} _ {t}, t, \\mathbf {y}, \\kappa) \\right. \\right. \\tag {6} \\\\ \\left. - \\hat {\\epsilon} _ {\\pi} (\\hat {\\mathbf {x}} _ {t}, t, \\mathbf {y}, \\kappa) \\frac {\\partial \\pmb {f} _ {\\theta} (\\mathbf {z} , \\mathbf {y})}{\\partial \\theta} \\right], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 729, + 905, + 799 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $C$ is a constant and $w(t)$ is the time-dependent weight that adjusts the contribution of each timestep, aligning the student's outputs with the diffusion model's predictions. Through alternating updates of the student and LoRA teacher, VSD enables efficient, high-fidelity generation across both text-to-image and text-to-3D tasks.", + "bbox": [ + 511, + 809, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18006", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/45e97e4177142f5a39114c7ddda68d95cd238911bc11629dff624be765090099.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
NPAnimePhotoCAPaintingsAverage
29.6229.1728.7928.6929.07
✓31.6629.1730.5830.3830.42
", + "bbox": [ + 106, + 89, + 467, + 146 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. The effect of negative prompts (NP) on the HPSv2 score of PixArt- $\\alpha$ [9]. Using generic negative prompts to remove visual artifacts (e.g., \"worst quality\", \"blurry\", \"bad anatomy\") leads to a significant improvement. Higher scores indicate better alignment with human preferences.", + "bbox": [ + 89, + 152, + 482, + 223 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f591d916009fa213f3d60703514f68ce8f4417e072ea7a5bc79ffb06a3ea11ff.jpg", + "image_caption": [ + "Figure 2. The inherent single-pass nature of one-step diffusion models renders them incompatible with direct CFG, resulting in artifact-laden outputs." + ], + "image_footnote": [], + "bbox": [ + 93, + 232, + 480, + 367 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Observations", + "text_level": 1, + "bbox": [ + 89, + 426, + 225, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Motivation", + "text_level": 1, + "bbox": [ + 89, + 452, + 212, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Positive prompt is not enough. While vision generative models like diffusion models excel at incorporating positive attributes, they fundamentally struggle with understanding negation. Consider requesting \"A portrait of a friendly monster, not scary\" (Fig. 1), the model typically fails to exclude the specified negative attribute \"scary\". Negative prompting addresses this limitation by creating a separate, dedicated input for exclusions – completely removing negation terms from the positive prompt and processing them independently. This clear separation allows the model to accurately distinguish what to avoid generating, providing significantly better control over the final output.", + "bbox": [ + 88, + 474, + 482, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Community Need. The growing use of negative prompting in the generative AI community further highlights its importance. Platforms like Civit AI [10] and various open-source repositories [2, 11, 29] have shown that using both types of prompts together results in more refined, higher-quality model outputs that align more closely with user specifications. Users can clearly specify both desired and unwanted attributes in a prompt pair, which is particularly useful when working with models that generate complex or highly detailed content. Experimentally, Table 1 and Figure 1 confirm that using negative prompts effectively enhances image quality of the model.", + "bbox": [ + 88, + 657, + 482, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Critical Gap. Despite the established need and widespread adoption of negative prompts in multi-step diffusion models and the rapidly growing prominence of one-step approaches, a significant limitation persists: no existing one", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "step methods effectively support negative prompting. Our proposed method fills this gap by providing a simple, efficient, and training-free mechanism that integrates negative prompt into one-step diffusion models to (1) improve controllability and (2) enhance the quality of the output image.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Challenges", + "text_level": 1, + "bbox": [ + 511, + 174, + 633, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A widely used method for implementing negative prompting is the CFG mechanism. However, applying CFG directly to one-step diffusion models presents significant challenges. CFG was designed for multi-step models, where the output is iteratively refined, allowing for gradual adjustments using both negative and unconditional outputs. In contrast, one-step models generate images in a single pass without the iterative denoising process, making them inherently unsuitable for this method. Direct application of CFG to these models leads to undesirable image blending, resulting in unnatural artifacts, as shown in Fig. 2. This highlights a crucial limitation: the CFG mechanism, though effective in multi-step models, is not directly translatable to one-step models without significant modifications.", + "bbox": [ + 511, + 196, + 906, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Proposed Method", + "text_level": 1, + "bbox": [ + 511, + 421, + 686, + 438 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our proposed methodology is detailed in Sec. 5.1 for the core mechanism. We present two primary implementation strategies for NASA: a training-free approach (NASA-I) and a distillation-based approach (NASA-T), which are presented in Sec. 5.2 and Sec. 5.3, respectively.", + "bbox": [ + 511, + 446, + 905, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.1. NASA: Negative-Away Steer Attention", + "text_level": 1, + "bbox": [ + 511, + 531, + 841, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast to CFG implementations that operate within the output space (synthesized image), as commonly practiced in multi-step diffusion models, we propose a novel approach that strategically relocates the guidance mechanism to the intermediate representation space of the diffusion model.", + "bbox": [ + 511, + 551, + 905, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, cross-attention layers in these models capture semantic connections between image and text features [17], making them ideal for controlling the feature alignment. We hypothesize that cross-attention layer features are particularly well-suited for this approach. We selectively manipulate these cross-attention maps to attenuate attention responses associated with negative prompts relative to those elicited by positive prompts, thereby filtering out unwanted semantic attributes.", + "bbox": [ + 511, + 628, + 905, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This representation-space guidance strategy offers a crucial advantage: it enables the exclusion of undesirable elements prior to the final image synthesis stage. This removal mechanism significantly mitigates the risk of generating blending artifacts, which are common byproducts of image-space manipulation techniques.", + "bbox": [ + 511, + 765, + 905, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our representation-space guidance also extends to architectures like FLUX [25], which replace traditional cross-attention with joint self-attention blocks. [46] show that", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18007", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c1efb25a93a4ba5d98b6406d9914dc0106f9d8ba62cde33b93f4071faefac6a7.jpg", + "image_caption": [ + "Figure 3. Left: An overview of the Negative-Away Steer Attention (NASA) pipeline. Positive (green) and negative (red) prompts are fed into a text encoder to generate positive and negative text features. The NASA module then processes these features, which adjusts the one-step diffusion model to steer the output image away from the negative features, refining it based on the positive features. Right: The details of the NASA module. It processes queries $(\\mathbf{Q}_l)$ in layer $l$ , note we will omit the subscript $l$ in subsequent notations to improve readability, with positive $(\\mathbf{V}^{+},\\mathbf{K}^{+})$ and negative $(\\mathbf{V}^{-},\\mathbf{K}^{-})$ key-value pairs to create positive $(\\mathbf{Z}^{+})$ and negative $(\\mathbf{Z}^{-})$ attention outputs. The final output $(\\mathbf{Z}^{\\mathrm{NASA}})$ is calculated by subtracting the weighted negative features $(\\mathbf{Z}^{-})$ from the positive features $(\\mathbf{Z}^{+})$ ." + ], + "image_footnote": [], + "bbox": [ + 102, + 85, + 893, + 305 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ac933e6e1c62f6066b1b984ad875dff2251cb27e44c23ac827d08f6165e9170c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFLUX.1-schnellSDXL-LCMSDXL-DMD2
4 steps1 step4 steps1 step4 steps1 step
None23%44%43%-27%25%
CFG30%0%14%-25%0%
NASA-I100%99%97%-100%100%
", + "bbox": [ + 91, + 407, + 496, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Comparison of success rate of unwanted feature removal in generated images. Note that SDXL-LCM [31] does not support one-step inference.", + "bbox": [ + 89, + 503, + 483, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "these blocks still embed distinct image-text cross-attention operations. This allows NASA to be seamlessly applied, proving its robustness across different backbones.", + "bbox": [ + 89, + 550, + 482, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall pipeline of the NASA mechanism is illustrated in Fig. 3. Given the latent feature representation at a given layer $l$ , denoted as $\\mathbf{Z}_l$ , and positive text features $\\mathbf{c}_p$ extracted from CLIP text encoder, we define the query, key, and value matrices as $\\mathbf{Q}_l = \\mathbf{Z}_l\\mathbf{W}_{q,l}$ , $\\mathbf{K}_l^+ = \\mathbf{c}_p\\mathbf{W}_{k,l}$ , and $\\mathbf{V}_l^+ = \\mathbf{c}_p\\mathbf{W}_{v,l}$ , where $\\mathbf{W}_{q,l}$ , $\\mathbf{W}_{k,l}$ , and $\\mathbf{W}_{v,l}$ are projection matrices at current layer $l$ . Analogously, for negative text features $\\mathbf{c}_n$ , we use the same projection matrices to define the corresponding key and value matrices for negative prompt conditioning: $\\mathbf{K}_l^- = \\mathbf{c}_n\\mathbf{W}_{k,l}$ and $\\mathbf{V}_l^- = \\mathbf{c}_n\\mathbf{W}_{v,l}$ . The NASA mechanism is formulated as follows:", + "bbox": [ + 89, + 595, + 483, + 762 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} _ {l} ^ {+} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} _ {l} \\left(\\mathbf {K} _ {l} ^ {+}\\right) ^ {\\top}}{\\sqrt {d}}\\right) \\mathbf {V} _ {l} ^ {+}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 768, + 482, + 805 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} _ {l} ^ {-} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} _ {l} \\left(\\mathbf {K} _ {l} ^ {-}\\right) ^ {\\top}}{\\sqrt {d}}\\right) \\mathbf {V} _ {l} ^ {-}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 805, + 482, + 840 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} _ {l} ^ {\\mathrm {N A S A}} = \\mathbf {Z} _ {l} ^ {+} - \\alpha \\cdot \\mathbf {Z} _ {l} ^ {-}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 840, + 482, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ is the scale factor to control the degree of negative feature removal.", + "bbox": [ + 89, + 869, + 482, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. NASA-I: Training-free Setting", + "text_level": 1, + "bbox": [ + 511, + 410, + 785, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Applicable to One-/Few-Step Models. To assess the effectiveness of NASA-I, we carry out experiments with pretrained one-step and few-step models, such as LCM [31], DMD2 [51] and FLUX.1-schnell [25]. We design a small set of positive and negative prompts and generate 100 images per model, both with and without negative prompts. We then calculate the percentage of images that successfully exclude the feature specified by the negative prompt. For example, given the positive prompt \"A photo of a person\" and the negative prompt \"male\", we measure the percentage of generated images that do not depict a male person. The average results across all prompt sets are summarized in Tab. 2. Our findings show that while traditional CFG is ineffective for one-step and few-step models, NASA-I provides precise control over unwanted features in generated images. Additionally, this capability is also qualitatively demonstrated in Fig. 1, Fig. 4 and Fig. 5.", + "bbox": [ + 511, + 439, + 906, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Efficiency Considerations. As described in Sec. 5.1, we reuse the projection matrices of the positive prompt for the negative prompt, ensuring that no additional model parameters are introduced and keeping the model size unchanged. Furthermore, applying NASA-I to all one-step and few-step models introduces only a minor computational overhead, increasing FLOPs by just $1.89\\%$ . This represents a significant efficiency improvement over CFG, which doubles the computational cost due to the need for separate forward passes for positive and negative prompts. Consequently, NASA-I enables negative prompt guidance in one-step diffusion models without introducing substantial latency, making it a practical solution for real-time applications.", + "bbox": [ + 511, + 703, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18008", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.3. NASA-T: Distillation Setting", + "text_level": 1, + "bbox": [ + 89, + 90, + 349, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given their effectiveness in improving image quality during inference, their potential utility in model distillation is worth exploring. In this section, we introduce a method that integrates the negative prompt into one-step diffusion distillation. While our approach is applicable to various distillation methods, we select SwiftBrush (SB) [12, 34], an image-free method, as the baseline for improvement due to its efficiency. As described in Eq. (6), SB updates the student model using the gradient of the VSD loss. To incorporate the negative prompt $\\mathbf{y}^{-}$ , we modify the existing formulation by replacing $\\hat{\\epsilon}_{\\psi}(\\hat{\\mathbf{x}}_t,t,\\mathbf{y},\\kappa)$ (see Eq. (4)) with $\\hat{\\epsilon}_{\\psi}(\\hat{\\mathbf{x}}_t,t,\\mathbf{y},\\mathbf{y}^{-},\\kappa)$ (see Eq. (5)). In this way, the teacher model provides explicit guidance to the student model, helping it learn to steer away from undesirable features. Moreover, we incorporate the NASA mechanism into the student model, enabling it to process negative prompts during training. This allows the student model to better recognize and suppress unnatural artifacts, leading to improved learning and enhanced image quality. Note that with the LoRA teacher model $\\hat{\\epsilon}_{\\pi}$ , we can choose whether to incorporate negative prompts. Empirically, we found that not using them yields better results. Detailed results and discussions are provided in Sec. 6.3.", + "bbox": [ + 89, + 112, + 485, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 89, + 468, + 223, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 493, + 279, + 510 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics and Benchmark. In diffusion model evaluation, Fréchet Inception Distance (FID) [18] on the zero-shot MS-COCO2014 [27] benchmark has been the standard metric. However, recent studies [8, 38, 51] suggest that FID in this benchmark does not always correlate with actual image quality and adopt human-focused alignment metrics like Human Preference Score v2 (HPSv2) [48] for benchmarking. Our experiments also follow this conclusion, so we choose HPSv2 as the primary metric for evaluating the models. Moreover, due to the absence of negative prompts in the MS-COCO2014 benchmark, we utilize the NegOpt dataset from [35], which contains pairs of positive and negative prompts. From this dataset, we randomly sample 30K prompt pairs to construct a subset for evaluation. Text alignment is then assessed using CLIP scores, where $\\mathrm{CLIP^{+}}$ evaluates alignment with the positive prompt and $\\mathrm{CLIP^{-}}$ evaluates alignment with the negative prompt.", + "bbox": [ + 89, + 521, + 483, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Datasets. Following SBv2 [12], we use total of 3.3M prompts from JourneyDB [44] and a subset of LAION $[42]^1$ . For negative prompts in the distillation setting (Sec. 5.3), we construct a dataset of 10K negative prompts derived from commonly used terms like \"bad photo, duplicate, low resolution, low quality\".", + "bbox": [ + 89, + 784, + 483, + 877 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We base our method on SBv2 [12] with our proposed modifications in Sec. 5.3. We utilize three backbones SDv1.5 [39], SDv2.1 [39] and PixArt- $\\alpha$ [9] as the frozen teachers and LoRA teachers are initialized with rank $r = 64$ and scaling $\\gamma = 128$ . The learning rates are set to $1e^{-6}$ for the student model and $1e^{-3}$ for the LoRA teacher, using the AdamW optimizer [30]. All our training is conducted on four NVIDIA A100 40GB GPUs with the total batch size of 64 for the SD models and 32 for PixArt- $\\alpha$ . Moreover, [6] found that instead of using a fixed CFG scale $\\kappa$ during training, sampling $\\kappa$ from a uniform distribution within a given range improves training robustness. Based on this insight, we set the teacher models' CFG scales as follows: $\\kappa \\sim \\mathcal{U}(0.5,4)$ for SDv1.5 and SDv2.1, and $\\kappa \\sim \\mathcal{U}(0.5,3)$ for PixArt- $\\alpha$ . For NASA-T, in each iteration, we uniformly sample the scale $\\alpha \\sim \\mathcal{U}(0,1)$ to help the student model adapt to different conditions and enhance generalization. For NASA-I, unless explicitly specified, we set $\\alpha = 0.1$ for SDv1.5-based models, $\\alpha = 0.2$ for SDv2.1-based models, and $\\alpha = 0.5$ for PixArt- $\\alpha$ -based models. Further details can be found in the Appendix.", + "bbox": [ + 511, + 90, + 906, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.2. Main Results", + "text_level": 1, + "bbox": [ + 511, + 415, + 651, + 430 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Results. Table 3 presents a comparison between our approach and prior distilled text-to-image diffusion models. Using our reimplementation of SBv2 [12] as the baseline, NASA-T consistently enhances image quality across all tested backbones. Compared to other methods, our model achieves the highest HPSv2 scores on average, demonstrating its clear superiority in image generation quality. This suggests that integrating complex loss functions like DMD2 [51] may not always be necessary for achieving state-of-the-art results. Moreover, while applying CFG directly degrades one-step model outputs, NASA-I further enhances image quality across all models when using negative prompts like \"worst quality, low quality, ugly, duplicate, out of frame, deformed, blurry, bad anatomy, watermark\". Notably, our method on the PixArt- $\\alpha$ -based backbone achieves a record-breaking HPSv2 score of 31.21, surpassing all other one-step diffusion models baselines. Regarding the NegOpt dataset, with the support of NASA-I, all models demonstrate a reduction in the $\\mathrm{CLIP}^{-}$ score while maintaining the $\\mathrm{CLIP}^{+}$ score. This indicates our method effectively eliminates unwanted features in generated outputs while preserving desired traits.", + "bbox": [ + 511, + 441, + 908, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Results. As discussed in Sec. 5.2, NASA-I is a versatile, training-free method for both one-step and few-step diffusion models. Figure 4 and Figure 5 present qualitative results of integrating NASA-I into the SDXL-DMD2 [51] and FLUX.1-schnell [25] models, respectively. When using a conditional negative prompt, the generated images retain key semantic features while effectively suppressing unwanted attributes specified by the negative prompt.", + "bbox": [ + 511, + 779, + 910, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1All datasets were downloaded and evaluated at MovianAI.", + "bbox": [ + 107, + 886, + 423, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18009", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/94bc3fc383891011e919360e3f806cbdc17999ac91af90247b146ef7508c4004.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetNegOptHPSv2
MethodCLIP+↑CLIP-↓Anime ↑Photo ↑CA ↑Paintings ↑Average ↑
PixArt-α-based backbone
PixArt-α [9] [Teacher]0.350.0529.6229.1728.7928.6929.07
YOSO [32]0.360.0828.7528.0628.5228.5728.48
+ NASA-I0.360.0628.7428.0528.5628.6028.49 (+0.01)
DMD [52]0.350.0829.3128.6728.4628.4128.71
+ CFG = 1.50.340.0930.0227.0728.3628.0728.38 (-0.33)
+ CFG = 2.50.310.1326.7423.8625.1324.6625.10 (-3.61)
+ NASA-I0.350.0529.3328.7128.4928.5328.77 (+0.06)
SBv2*0.360.0932.1929.0930.3929.6930.34
+ NASA-I0.360.0632.6029.5831.0930.6530.98 (+0.64)
+ NASA-T0.350.0832.3329.2630.7530.1030.61 (+0.27)
+ NASA-T + CFG = 1.50.340.1029.4726.5028.2227.6827.97 (-2.37)
+ NASA-T + NASA-I0.350.0632.6629.6731.3130.7131.09 (+0.75)
+ NASA-T + NASA-I (α = 1)0.350.0532.6529.6531.4531.0631.21 (+0.87)
Stable Diffusion 1.5-based backbone
SDv1.5 [39] [Teacher]0.310.0626.5127.1926.0626.1226.47
InstaFlow-0.9B [28]0.330.1126.1026.6225.9225.9526.15
+ NASA-I0.330.1026.1526.6825.9325.9826.19 (+0.04)
DMD2 [51]0.310.1026.3927.0025.8025.8226.25
+ NASA-I0.310.0926.3927.0225.8025.8326.26 (+0.01)
SBv2*0.320.0927.1827.5826.6926.6227.02
+ NASA-I0.320.0827.1927.5926.7126.6327.03 (+0.01)
+ NASA-T0.350.1027.6427.9427.1927.0327.45 (+0.43)
+ NASA-T + NASA-I0.350.0927.6527.9727.1827.0427.46 (+0.44)
Stable Diffusion 2.1-based backbone
SDv2.1 [39] [Teacher]0.330.0627.4826.8926.8627.4627.17
SBv2 [12]0.340.1027.2527.6226.8626.7727.13
+ NASA-I0.350.0827.4427.8426.9127.0227.30 (+0.17)
SBv2*0.370.1027.5627.8426.9727.0327.35
+ NASA-I0.370.0827.7028.0027.1227.2227.51 (+0.16)
+ NASA-T0.360.0928.0028.6527.4427.2627.84 (+0.49)
+ NASA-T + NASA-I0.360.0728.0428.6827.5127.5027.93 (+0.58)
", + "bbox": [ + 174, + 88, + 823, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Quantitative comparisons between our method and previous works. NASA-I refers to our training-free method introduced in Sec. 5.2, while NASA-T denotes our distillation-based approach described in Sec. 5.3. SBv2* refers to our reimplementation of SBv2 with a randomly sampled CFG scale $\\kappa$ .", + "bbox": [ + 89, + 537, + 906, + 580 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cd87ca5eb6940b8ea62f3dbc8670a0a27ee17e081844b9f092029ba30e85a70c.jpg", + "image_caption": [ + "Figure 4. Effect of different scale values (0.0 to 1.0) in NASA-I with SDXL-DMD2 [51] for both one-step and few-step settings, illustrating the progressive influence on visual details and composition." + ], + "image_footnote": [], + "bbox": [ + 151, + 590, + 846, + 845 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18010", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/393710f27e24d69dcb9494f23cf0cd7e5d701b4345160768a0599150ef649410.jpg", + "image_caption": [ + "Figure 5. Qualitative results of NASA-I in FLUX.1-schnell [25]." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 903, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ce7533494e5f6c15882df42521eeabdbdd5c06b5836bb9f0fe2cdaaf4654f19b.jpg", + "image_caption": [ + "Figure 6. Qualitative results of NASA-I for text-to-video generation in CausVid [53]." + ], + "image_footnote": [], + "bbox": [ + 93, + 243, + 390, + 329 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5933f6c822b2d2568a91d31bf760a8312a61062190b15958f6ff4462019d29ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 392, + 243, + 687, + 329 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0c9cdb25b630e8198754b173f1ef9249f3e01efe444a1205264f6b7033e9f4b0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodVBenchmark-Long benchmark
Aesthetic Quality ↑Imaging Quality ↑
None61.9867.12
NASA-I63.3367.36
", + "bbox": [ + 696, + 257, + 900, + 328 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1fd9a75d397df439a764e6ed6d639b67c7f76330d4f364d8783adc4c26b44dc7.jpg", + "table_caption": [ + "Table 4. CausVid quality metrics." + ], + "table_footnote": [], + "table_body": "
NASA✓✓
Teacher✓✓✓✓
LoRA Teacher✓✓
HPSv2 Avg. ↑30.3429.8429.8030.6129.89
", + "bbox": [ + 94, + 369, + 475, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation studies on the integration of negative prompts in model distillation using the PixArt- $\\alpha$ backbone.", + "bbox": [ + 89, + 449, + 482, + 477 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/57fae8adf7a27b7fb899130db0f95ddf7097e631a9594cdfe0137c0a5709a7c5.jpg", + "image_caption": [ + "Figure 7. Ablation study on the NASA scale using our distilled model based on the PixArt- $\\alpha$ backbone." + ], + "image_footnote": [], + "bbox": [ + 106, + 491, + 464, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Application to Text-to-Video Models. To demonstrate the versatility of our approach, we extended NASA-I to a few-step video diffusion model, CausVid [53]. As shown in Fig. 6, our method effectively suppresses unwanted visual attributes while preserving key semantic content and temporal consistency. Quantitative evaluation on the VBenchmarkLong [22] benchmark further validates this. Table 4 shows that using NASA-I with common negative prompts like \"low quality, bad anatomy, blurry, distortion, ugly, low resolution, unclear details\" improves both the aesthetic and imaging quality scores of the generated videos. This highlights NASA-I's potential as a general-purpose tool for enhancing guided generation across different modalities.", + "bbox": [ + 89, + 703, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 372, + 679, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5 presents an ablation on integrating negative prompts during model distillation with the PixArt- $\\alpha$ [9] backbone. The baseline is our reimplementation of SBv2 with a randomly sampled CFG scale $\\kappa$ . Simply applying negative prompts only in the teacher and/or the LoRA teacher models slightly degrades the HPSv2 score, as the student model is not explicitly trained on this negative conditioning. In contrast, incorporating NASA with negative prompts during training allows the student model to better capture and utilize this conditioning, leading to improved performance. Furthermore, we find that integrating negative prompts solely into the teacher model yields better results than applying them to both the teacher and LoRA teacher models.", + "bbox": [ + 509, + 393, + 906, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Using our PixArt- $\\alpha$ -based distilled model as the base, we varied the NASA scale $\\alpha$ and evaluated the results with HPSv2 on the benchmark dataset. Figure 7 demonstrates that NASA-I generally enhances image synthesis, resulting in outputs that better align with human preferences. However, if the NASA scale $\\alpha$ is set too high, the negative attributes may overly dominate the positive attributes, adversely affecting the final image quality.", + "bbox": [ + 511, + 590, + 906, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Discussion and Conclusion", + "text_level": 1, + "bbox": [ + 511, + 723, + 761, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces NASA, the first method to integrate negative prompt guidance into one-step and few-step diffusion models for both inference and distillation. By leveraging cross-attention adjustments, NASA effectively suppresses unwanted features in generated images. Experimental results show that NASA enhances image quality and outperforms strong baselines, establishing a new state-of-the-art in one-step model distillation based on human preference metrics. While NASA's effectiveness relies on selecting an appropriate scale for suppressing negative features, this parameter does not require extensive tuning.", + "bbox": [ + 509, + 747, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18011", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 118, + 187, + 133 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mohammadreza Armandpour, Ali Sadeghian, Huangjie Zheng, Amir Sadeghian, and Mingyuan Zhou. Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. ArXiv, abs/2304.04968, 2023. 3", + "[2] AUTOMATIC1111. Automatic1111/stable-diffusion-webui, 2025. 4", + "[3] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 1", + "[4] Yuanhao Ban, Ruochen Wang, Tianyi Zhou, Minhao Cheng, Boqing Gong, and Cho-Jui Hsieh. Understanding the impact of negative prompts: When and how do they take effect? In European conference on computer vision, pages 190–206. Springer, 2024. 2, 3", + "[5] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Yuanzhen Li, Tomer Michaeli, Oliver Wang, Deqing Sun, Tali Dekel, and Inbar Mosseri. Lumiere: A space-time diffusion model for video generation. ArXiv, abs/2401.12945, 2024. 1", + "[6] Clement Chadebec, Onur Tasar, Eyal Benaroche, and Benjamin Aubin. Flash diffusion: Accelerating any conditional diffusion model for few steps image generation. arXiv preprint arXiv:2406.02347, 2024. 6", + "[7] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James T. Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. ArXiv, abs/2310.00426, 2023. 1", + "[8] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- $\\sigma$ : Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv: 2403.04692, 2024. 6", + "[9] Junsong Chen, Jincheng YU, Chongjian GE, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 4, 6, 7, 8", + "[10] Civitai Platform. https://civitai.com.4", + "[11] comfyanonymous. comfyanonymous/comfyui, 2025. 4", + "[12] Trung Dao, Thuan Hoang Nguyen, Thanh Le, Duc Vu, Khoi Nguyen, Cuong Pham, and Anh Tran. Swiftbrush v2: Make your one-step diffusion model better than its teacher. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2, 3, 6, 7", + "[13] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 3" + ], + "bbox": [ + 93, + 143, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Zach Evans, CJ Carr, Josiah Taylor, Scott H. Hawley, and Jordi Pons. Fast timing-conditioned latent audio diffusion. ArXiv, abs/2402.04825, 2024. 1", + "[15] Peng Gao, Le Zhuo, Ziyi Lin, Chris Liu, Junsong Chen, Ruoyi Du, Enze Xie, Xu Luo, Longtian Qiu, Yuhang Zhang, Chen Lin, Rongjie Huang, Shijie Geng, Renrui Zhang, Junlin Xi, Wenqi Shao, Zhengkai Jiang, Tianshuo Yang, Weicai Ye, He Tong, Jingwen He, Yu Jiao Qiao, and Hongsheng Li. Lumina-t2x: Transforming text into any modality, resolution, and duration via flow-based large diffusion transformers. ArXiv, abs/2405.05945, 2024. 1", + "[16] Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1", + "[17] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control.(2022). In International Conference on Learning Representations, 2023. 4", + "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Neural Information Processing Systems, 2017. 6", + "[19] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 1, 2, 3", + "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. Advances in Neural Information Processing Systems, 33, 2020. 3", + "[21] Rongjie Huang, Jia-Bin Huang, Dongchao Yang, Yi Ren, Luping Liu, Mingze Li, Zhenhui Ye, Jinglin Liu, Xiaoyue Yin, and Zhou Zhao. Make-an-audio: Text-to-audio generation with prompt-enhanced diffusion models. ArXiv, abs/2301.12661, 2023. 1", + "[22] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 8", + "[23] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1", + "[24] Oren Katzir, Or Patashnik, Daniel Cohen-Or, and Dani Lischinski. Noise-free score distillation. arXiv preprint arXiv: 2310.17590, 2023. 3", + "[25] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024. 4, 5, 6, 8", + "[26] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 1", + "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18012", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6", + "[28] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, and Qiang Liu. InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation. arXiv preprint arXiv:2309.06380, 2023. 7", + "[29] Illyasviel. Illyasviel/stable-diffusion-webui-forge, 2025. 4", + "[30] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 6", + "[31] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv: 2310.04378, 2023. 5", + "[32] Yihong Luo, Xiaolong Chen, Xinghua Qu, and Jing Tang. You only sample once: Taming one-step text-to-image synthesis by self-cooperative diffusion gans. arXiv preprint arXiv: 2403.12931, 2024. 7", + "[33] David McAllister, Songwei Ge, Jia-Bin Huang, David W. Jacobs, Alexei A. Efros, Aleksander Holynski, and Angjoo Kanazawa. Rethinking score distillation as a bridge between image distributions. ArXiv, abs/2406.09417, 2024. 3", + "[34] Thuan Hoang Nguyen and Anh Tran. Swiftbrush: One-step text-to-image diffusion model with variational score distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 6", + "[35] Michael Ogezi and Ning Shi. Optimizing negative prompts for enhanced aesthetics and fidelity in text-to-image generation. arXiv preprint arXiv:2403.07605, 2024. 6", + "[36] Dustin Podell, Zion English, Kyle Lacey, A. Blattmann, Tim Dockhorn, Jonas Muller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. International Conference on Learning Representations, 2023. 1", + "[37] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D diffusion. *ArXiv*, abs/2209.14988, 2022. 1", + "[38] Yuxi Ren, Xin Xia, Yanzuo Lu, Jiacheng Zhang, Jie Wu, Pan Xie, Xing Wang, and Xuefeng Xiao. Hyper-sd: Trajectory segmented consistency model for efficient image synthesis. arXiv preprint arXiv: 2404.13686, 2024. 6", + "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 3, 6, 7", + "[40] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1", + "[41] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 2" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. LAION-5B: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 6", + "[43] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 3", + "[44] Keqiang Sun, Junting Pan, Yuying Ge, Hao Li, Haodong Duan, Xiaoshi Wu, Renrui Zhang, Aojun Zhou, Zipeng Qin, Yi Wang, Jifeng Dai, Yu Qiao, Limin Wang, and Hongsheng Li. JourneyDB: A benchmark for generative image understanding. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 6", + "[45] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 1, 3", + "[46] Tianyi Wei, Dongdong Chen, Yifan Zhou, and Xingang Pan. Enhancing mmdit-based text-to-image models for similar subject generation. arXiv preprint arXiv:2411.18301, 2024. 4", + "[47] Max Woolf. Stable diffusion 2.0 and the importance of negative prompts for good results. https://minimaxir.com/2022/11/stable-diffusion-negative-prompt/, 2023.2,3", + "[48] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. ArXiv, abs/2306.09341, 2023. 2, 6", + "[49] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190, 2023. 1", + "[50] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Xiaotao Gu, Yuxuan Zhang, Weihan Wang, Yean Cheng, Ting Liu, Bin Xu, Yuxiao Dong, and Jie Tang. Cogvideox: Text-to-video diffusion models with an expert transformer. ArXiv, abs/2408.06072, 2024. 1", + "[51] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. In NeurIPS, 2024. 2, 3, 5, 6, 7", + "[52] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In CVPR, 2024. 2, 3, 7", + "[53] Tianwei Yin, Qiang Zhang, Richard Zhang, William T Freeman, Fredo Durand, Eli Shechtman, and Xun Huang. From slow bidirectional to fast autoregressive video diffusion models. 2025. 8" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18013", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Supercharged One-step Text-to-Image Diffusion Models with Negative Prompts/1564dd9e-386a-48aa-bdfb-0018fc932f68_model.json b/2025/Supercharged One-step Text-to-Image Diffusion Models with Negative Prompts/1564dd9e-386a-48aa-bdfb-0018fc932f68_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7aba0cb75a6aace1b85ad1e06cea5e3cb390fb2e --- /dev/null +++ b/2025/Supercharged One-step Text-to-Image Diffusion Models with Negative Prompts/1564dd9e-386a-48aa-bdfb-0018fc932f68_model.json @@ -0,0 +1,2035 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.099, + 0.131, + 0.9, + 0.156 + ], + "angle": 0, + "content": "Supercharged One-step Text-to-Image Diffusion Models with Negative Prompts" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.18, + 0.892, + 0.201 + ], + "angle": 0, + "content": "Viet Nguyen\\(^{1\\star}\\), Anh Nguyen\\(^{1\\star}\\), Trung Dao\\(^{1}\\), Khoi Nguyen\\(^{1}\\), Cuong Pham\\(^{1,2}\\), Toan Tran\\(^{1}\\), Anh Tran\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.216, + 0.744, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\)Qualcomm AI Research† \\(^{2}\\)Posts & Telecom. Inst. of Tech." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.255, + 0.907, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.498, + 0.908, + 0.555 + ], + "angle": 0, + "content": "Figure 1. Negative prompting are a key technique for improving controllability and image quality in text-to-image diffusion models. However, classifier-free guidance (CFG) [19] only supports negative prompt integration in multi-step models due to its iterative nature. Our method, NASA, is the first to enable negative prompting in one-step models, allowing for the suppression of unwanted features and providing greater control over image attributes." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.562, + 0.327, + 0.577 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.593, + 0.486, + 0.866 + ], + "angle": 0, + "content": "The escalating demand for real-time image synthesis has driven significant advancements in one-step diffusion models, which inherently offer expedited generation speeds compared to traditional multi-step methods. However, this enhanced efficiency is frequently accompanied by a compromise in the controllability of image attributes. While negative prompting, typically implemented via classifier-free guidance (CFG), has proven effective for fine-grained control in multi-step models, its application to one-step generators remains largely unaddressed. Due to the lack of iterative refinement, as in multi-step diffusion, directly applying CFG to one-step generation leads to blending artifacts and diminished output quality. To fill this gap, we introduce Negative-Away Steer Attention (NASA), an efficient method that integrates negative prompts into one-step diffusion models. NASA operates within the intermediate representation space by leveraging cross-attention mechanisms to suppress undesired visual attributes. This strategy" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.563, + 0.909, + 0.713 + ], + "angle": 0, + "content": "avoids the blending artifacts inherent in output-space guidance and achieves high efficiency, incurring only a minimal \\(1.89\\%\\) increase in FLOPs compared to the computational doubling of CFG. Furthermore, NASA can be seamlessly integrated into existing timestep distillation frameworks, enhancing the student's output quality. Experimental results demonstrate that NASA substantially improves controllability and output quality, achieving an HPSv2 score of 31.21, setting a new state-of-the-art benchmark for one-step diffusion models." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.752, + 0.646, + 0.767 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Diffusion models have recently become popular due to their capacity to produce high-quality, diverse outputs in image [3, 7, 36, 40], 3D [26, 37, 45], audio [14, 21] and video [5, 15, 49, 50] synthesis. Unlike other generative models, such as Generative Adversarial Networks (GANs) [16, 23], diffusion models gradually refine their output through a series of steps, allowing them to achieve diverse yet impressively detailed and realistic outputs. However, this iterative" + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.876, + 0.21, + 0.888 + ], + "angle": 0, + "content": "\\* Equal Contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.888, + 0.482, + 0.901 + ], + "angle": 0, + "content": "† Qualcomm AI Research is an initiative of Qualcomm Technologies, Inc." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.876, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "18004" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.348 + ], + "angle": 0, + "content": "refinement is both time-consuming and computationally demanding, which limits their practicality in real-world applications. As a result, there is an increasing interest in techniques to speed up diffusion models while maintaining output quality. Recent distillation methods [12, 34, 51, 52] compress the entire multi-step generation process into a single step, a breakthrough that enables real-time image synthesis. This speedup makes diffusion models practical for applications like content creation and interactive media, where their slow performance was previously a major drawback. Despite these impressive speed advancements in one-step generators, we have identified a significant limitation that affects all these models: the inability to support negative prompting [4, 47]. This feature, which allows users to explicitly exclude unwanted elements, is widely available in multi-step diffusion models but remains notably absent in their single-step alternatives." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.351, + 0.482, + 0.487 + ], + "angle": 0, + "content": "The most widely adopted approach for negative prompting, classifier-free guidance (CFG) [19] works by generating two parallel noise predictions during each denoising step - one incorporating the positive prompt and another using the negative prompt. The system then subtracts the negative prediction from the positive one, effectively steering generation away from unwanted attributes. This process relies on the iterative nature of multi-step diffusion, where corrections can be applied and refined over dozens of steps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.482, + 0.685 + ], + "angle": 0, + "content": "One-step generators, however, collapse this entire process into a single forward pass, eliminating the iterative refinement cycles that make CFG possible. Without these multiple correction opportunities, there is no straightforward way to incorporate negative guidance into the generation pipeline. Consequently, users face an unappealing choice between the instant generation that might include unwanted elements or waiting longer for more precisely controlled results. This speed-versus-control dilemma represents the most significant practical barrier to widespread adoption of one-step generators in professional creative workflows, highlighting the urgent need for novel approaches that can bridge this gap." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.482, + 0.822 + ], + "angle": 0, + "content": "To address this issue, we introduce a novel method called Negative-Away Steer Attention (NASA) used in inference, which incorporates negative prompt guidance directly into one-step diffusion models. NASA works by adjusting cross-attention layers, effectively reducing unwanted features in generated images. By steering attention within intermediate representation space, NASA offers image control capabilities that were previously achievable only in multi-step diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.482, + 0.901 + ], + "angle": 0, + "content": "With our newly developed capability, we have explored integrating negative prompting directly into one-step distillation frameworks, focusing on the SwiftBrush (SB) family of models [12, 34]. The SB approach is compelling for its unique, image-free distillation methodology. By match-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": "ing teacher model probability distributions and applying sophisticated training techniques, SBv2 can outperform its own teacher on standard image synthesis benchmarks. Despite these technical achievements, SB models still fall short in human preference evaluations. The core challenge lies in the distillation process: without real image supervision, the system lacks clear signals about which generation patterns should be avoided. This is precisely where negative prompting and our NASA framework come in." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.228, + 0.905, + 0.364 + ], + "angle": 0, + "content": "Committed to maintaining SB's valuable image-free distillation approach while incorporating guidance sampling to align with human preference, we provide the critical missing component, a way to steer the model away from undesirable outputs during training. NASA enables something previously impossible: guidance sampling within a one-step generator. Through this integration, our model achieves an HPSv2 [48] score of 31.21, establishing a new state-of-the-art on human preference metrics for one-step models." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.365, + 0.818, + 0.379 + ], + "angle": 0, + "content": "In summary, our contributions are twofold:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.384, + 0.905, + 0.442 + ], + "angle": 0, + "content": "- We introduce Negative-Away Steer Attention (NASA), the first method to integrate negative prompt guidance into one-step and few-step diffusion models in both inference and distillation settings." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.444, + 0.905, + 0.519 + ], + "angle": 0, + "content": "- By integrating NASA into the SwiftBrush (SB) distillation framework, we achieve a new state-of-the-art for one-step diffusion models on human preference metrics. This marks the first successful integration of negative prompt guidance directly into one-step model training." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.384, + 0.905, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.535, + 0.655, + 0.55 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.56, + 0.81, + 0.575 + ], + "angle": 0, + "content": "2.1. Distribution Matching Distillation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.905, + 0.688 + ], + "angle": 0, + "content": "Recent work [12, 34, 51, 52] has explored text-to-image diffusion distillation through distribution matching techniques. These methods directly align the output distributions of teacher and student models, particularly through Variational Score Distillation (VSD). These techniques can be broadly categorized into two approaches: image-dependent training and image-free training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.905, + 0.809 + ], + "angle": 0, + "content": "Image-dependent methods rely on real images for adversarial loss [41, 51] or synthetic ones generated by the teacher models for reconstruction loss [52]. Despite their effectiveness in preserving image quality, these approaches face significant practical limitations: they require substantial storage capacity, consume high memory during training, demand extensive computational resources, and risk mode collapse due to intensive image supervision." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In contrast, image-free approaches [12, 34] have demonstrated promising results by relying solely on prompt text for training, significantly reducing resource requirements while maintaining competitive performance. However, image-free techniques often encounter training instability, particularly within VSD-based frameworks. Existing solu" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "18005" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.182 + ], + "angle": 0, + "content": "tions to stabilize training [52] introduce significant memory and computational overhead, limiting their practical application. Our work addresses this gap by introducing a simple yet effective method that enhances training stability without additional computational burden, making image-free variational score distillation produce better models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.189, + 0.459, + 0.206 + ], + "angle": 0, + "content": "2.2. Sampling Guidance with Negative Prompts" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.211, + 0.485, + 0.424 + ], + "angle": 0, + "content": "Classifier-free guidance (CFG) [19] and negative prompting [4, 47] are key techniques for content control during denoising. In multi-step diffusion models, these methods enhance desired image features and suppress undesired ones, improving visual fidelity and reducing artifacts [1, 13, 24]. Score-based approaches like NFSD [24] employ negative prompting to steer generated samples toward real-image distributions, while SDS-Bridge [33] uses negative prompts to model source distributions more accurately. Despite their success, negative prompting has yet to be integrated into one-step or few-step diffusion models. To bridge this gap, we introduce NASA, the first method to enable negative prompting in one-step and few-step diffusion models, enhancing control in faster generation settings." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.434, + 0.23, + 0.45 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.484, + 0.672 + ], + "angle": 0, + "content": "Diffusion Models [20, 39, 43] produce high-quality images by progressively denoising inputs. This technique involves a forward process that gradually adds noise to the data and a reverse process that reconstructs the original data by removing noise. While the original diffusion models act on the image space [20, 43], LDM [39] processes on the latent space produced by a pretrained VAE for more efficient computation. In particular, starting with a data point \\(\\mathbf{x}_0\\) sampled from an unknown distribution \\(q(\\mathbf{x}_0)\\), the forward process gradually diffuses \\(\\mathbf{x}_0\\) into a standard Gaussian noise \\(\\mathbf{x}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})\\) through \\(T\\) consecutive timesteps, where \\(\\mathbf{I}\\) is the identity matrix. At each time-step \\(t\\), a noisier version of \\(\\mathbf{x}_0\\), denoted as \\(\\mathbf{x}_t\\), is drawn from \\(q_t(\\mathbf{x}_t | \\mathbf{x}_0) = \\mathcal{N}(\\alpha_t \\mathbf{x}_0, \\sigma_t^2 \\mathbf{I})\\) using a standard Gaussian noise \\(\\epsilon \\sim \\mathcal{N}(0, \\mathbf{I})\\):" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.678, + 0.482, + 0.695 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {t} = \\alpha_ {t} \\mathbf {x} _ {0} + \\sigma_ {t} \\boldsymbol {\\epsilon}, \\quad \\forall t \\in \\{0, \\dots , T \\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.699, + 0.483, + 0.746 + ], + "angle": 0, + "content": "where \\(\\{(\\alpha_t,\\sigma_t)\\}_{t = 1}^T\\) defines the noise schedule, with boundary conditions \\((\\alpha_0,\\sigma_0) = (1,0)\\) for the clean sample and \\((\\alpha_{T},\\sigma_{T}) = (0,1)\\) for pure noise." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.746, + 0.484, + 0.837 + ], + "angle": 0, + "content": "Conversely, the reverse process aims to gradually reconstruct the original data by denoising the input over \\(T\\) steps, starting from an initial random Gaussian noise \\(\\mathbf{x}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\). The model is trained by minimizing the difference between the noise estimated by the model, \\(\\epsilon_{\\phi}\\), parameterized by \\(\\phi\\), and the actual noise in Eq. (1):" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.842, + 0.483, + 0.866 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\phi} \\mathbb {E} _ {t \\sim \\mathcal {U} (0, T), \\epsilon \\sim \\mathcal {N} (0, I)} \\| \\epsilon_ {\\phi} (\\mathbf {x} _ {t}, t) - \\boldsymbol {\\epsilon} \\| _ {2} ^ {2}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Text-to-Image Diffusion Models guide the sampling process using an additional text prompt y to produce outputs" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.138 + ], + "angle": 0, + "content": "that are both photorealistic and aligned with the provided textual descriptions. The training objective, slightly modified from the unconditional loss in Eq. (2), is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.151, + 0.907, + 0.175 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\phi} \\mathbb {E} _ {t \\sim \\mathcal {U} (0, T), \\epsilon \\sim \\mathcal {N} (0, I), \\mathbf {y}} \\| \\epsilon_ {\\phi} (\\mathbf {x} _ {t}, t, \\mathbf {y}) - \\epsilon \\| _ {2} ^ {2}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.182, + 0.907, + 0.288 + ], + "angle": 0, + "content": "Classifier-free Guidance (CFG) [19] is an inference technique designed to enhance the quality of generated images by blending the predictions from both a conditional and an unconditional model. At each sampling step, CFG adjusts the denoiser's output using a control parameter \\(\\kappa > 1\\), allowing controlled guidance that aligns more closely with the desired conditions:" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.3, + 0.907, + 0.318 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {\\epsilon}} _ {\\phi} (\\mathbf {x} _ {t}, t, \\mathbf {y}, \\kappa) = (1 - \\kappa) \\boldsymbol {\\epsilon} _ {\\phi} (\\mathbf {x} _ {t}, t) + \\kappa \\boldsymbol {\\epsilon} _ {\\phi} (\\mathbf {x} _ {t}, t, \\mathbf {y}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.329, + 0.906, + 0.403 + ], + "angle": 0, + "content": "Negative Prompting [4, 47] provides enhanced control by suppressing unwanted features in the generated content. Instead of producing an unconditional output, the model generates an output conditioned on the negative prompt \\(\\mathbf{y}^{-}\\), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.417, + 0.906, + 0.448 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {\\epsilon}} _ {\\phi} \\left(\\mathbf {x} _ {t}, t, \\mathbf {y}, \\mathbf {y} ^ {-}, \\kappa\\right) = (1 - \\kappa) \\boldsymbol {\\epsilon} _ {\\phi} \\left(\\mathbf {x} _ {t}, t, \\mathbf {y} ^ {-}\\right) + \\kappa \\boldsymbol {\\epsilon} _ {\\phi} \\left(\\mathbf {x} _ {t}, t, \\mathbf {y}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.907, + 0.72 + ], + "angle": 0, + "content": "Variational Score Distillation (VSD) is a powerful framework that utilizes pretrained diffusion-based text-to-image models to enhance text-based generation across both 3D [45] and 2D [12, 34, 51, 52] domains. The central aim of VSD is to align the renderings of a differentiable generator with the probability density of plausible images as guided by the 2D diffusion model. To accomplish this, VSD employs a two-teacher approach that uses a fixed pretrained diffusion model \\(\\epsilon_{\\psi}\\) and an adaptive LoRA-based teacher \\(\\epsilon_{\\pi}\\). While training, the student model \\(f_{\\theta}\\) produces 2D images \\(\\hat{\\mathbf{x}}_0 = f_{\\theta}(\\mathbf{z},\\mathbf{y})\\) using an input noise \\(\\mathbf{z} \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) and a text prompt \\(\\mathbf{y}\\). Noisy images \\(\\hat{\\mathbf{x}}_t = \\alpha_t\\hat{\\mathbf{x}}_0 + \\sigma_t\\epsilon\\) are then fed into both teacher models. The LoRA teacher model \\(\\epsilon_{\\pi}\\) aligns with the student distribution by minimizing a denoising L2 loss on single-step samples. This arrangement supports robust and adaptive guidance suitable for a variety of generative architectures. The gradient of the learning objective with respect to the student's parameters \\(\\theta\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.731, + 0.906, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {V S D}} = \\mathbb {E} _ {t, \\mathbf {y}, \\mathbf {z}, \\kappa = C} \\left[ w (t) \\left(\\hat {\\epsilon} _ {\\psi} (\\hat {\\mathbf {x}} _ {t}, t, \\mathbf {y}, \\kappa) \\right. \\right. \\tag {6} \\\\ \\left. - \\hat {\\epsilon} _ {\\pi} (\\hat {\\mathbf {x}} _ {t}, t, \\mathbf {y}, \\kappa) \\frac {\\partial \\pmb {f} _ {\\theta} (\\mathbf {z} , \\mathbf {y})}{\\partial \\theta} \\right], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\( C \\) is a constant and \\( w(t) \\) is the time-dependent weight that adjusts the contribution of each timestep, aligning the student's outputs with the diffusion model's predictions. Through alternating updates of the student and LoRA teacher, VSD enables efficient, high-fidelity generation across both text-to-image and text-to-3D tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18006" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.09, + 0.468, + 0.147 + ], + "angle": 0, + "content": "
NPAnimePhotoCAPaintingsAverage
29.6229.1728.7928.6929.07
✓31.6629.1730.5830.3830.42
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.154, + 0.483, + 0.224 + ], + "angle": 0, + "content": "Table 1. The effect of negative prompts (NP) on the HPSv2 score of PixArt-\\(\\alpha\\) [9]. Using generic negative prompts to remove visual artifacts (e.g., \"worst quality\", \"blurry\", \"bad anatomy\") leads to a significant improvement. Higher scores indicate better alignment with human preferences." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.233, + 0.482, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.376, + 0.483, + 0.418 + ], + "angle": 0, + "content": "Figure 2. The inherent single-pass nature of one-step diffusion models renders them incompatible with direct CFG, resulting in artifact-laden outputs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.428, + 0.227, + 0.443 + ], + "angle": 0, + "content": "4. Observations" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.453, + 0.214, + 0.468 + ], + "angle": 0, + "content": "4.1. Motivation" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.476, + 0.483, + 0.657 + ], + "angle": 0, + "content": "Positive prompt is not enough. While vision generative models like diffusion models excel at incorporating positive attributes, they fundamentally struggle with understanding negation. Consider requesting \"A portrait of a friendly monster, not scary\" (Fig. 1), the model typically fails to exclude the specified negative attribute \"scary\". Negative prompting addresses this limitation by creating a separate, dedicated input for exclusions – completely removing negation terms from the positive prompt and processing them independently. This clear separation allows the model to accurately distinguish what to avoid generating, providing significantly better control over the final output." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.659, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Community Need. The growing use of negative prompting in the generative AI community further highlights its importance. Platforms like Civit AI [10] and various open-source repositories [2, 11, 29] have shown that using both types of prompts together results in more refined, higher-quality model outputs that align more closely with user specifications. Users can clearly specify both desired and unwanted attributes in a prompt pair, which is particularly useful when working with models that generate complex or highly detailed content. Experimentally, Table 1 and Figure 1 confirm that using negative prompts effectively enhances image quality of the model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Critical Gap. Despite the established need and widespread adoption of negative prompts in multi-step diffusion models and the rapidly growing prominence of one-step approaches, a significant limitation persists: no existing one" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "step methods effectively support negative prompting. Our proposed method fills this gap by providing a simple, efficient, and training-free mechanism that integrates negative prompt into one-step diffusion models to (1) improve controllability and (2) enhance the quality of the output image." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.175, + 0.635, + 0.192 + ], + "angle": 0, + "content": "4.2. Challenges" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.198, + 0.907, + 0.41 + ], + "angle": 0, + "content": "A widely used method for implementing negative prompting is the CFG mechanism. However, applying CFG directly to one-step diffusion models presents significant challenges. CFG was designed for multi-step models, where the output is iteratively refined, allowing for gradual adjustments using both negative and unconditional outputs. In contrast, one-step models generate images in a single pass without the iterative denoising process, making them inherently unsuitable for this method. Direct application of CFG to these models leads to undesirable image blending, resulting in unnatural artifacts, as shown in Fig. 2. This highlights a crucial limitation: the CFG mechanism, though effective in multi-step models, is not directly translatable to one-step models without significant modifications." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.422, + 0.687, + 0.439 + ], + "angle": 0, + "content": "5. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.906, + 0.524 + ], + "angle": 0, + "content": "Our proposed methodology is detailed in Sec. 5.1 for the core mechanism. We present two primary implementation strategies for NASA: a training-free approach (NASA-I) and a distillation-based approach (NASA-T), which are presented in Sec. 5.2 and Sec. 5.3, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.532, + 0.843, + 0.547 + ], + "angle": 0, + "content": "5.1. NASA: Negative-Away Steer Attention" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.629 + ], + "angle": 0, + "content": "In contrast to CFG implementations that operate within the output space (synthesized image), as commonly practiced in multi-step diffusion models, we propose a novel approach that strategically relocates the guidance mechanism to the intermediate representation space of the diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.906, + 0.764 + ], + "angle": 0, + "content": "Specifically, cross-attention layers in these models capture semantic connections between image and text features [17], making them ideal for controlling the feature alignment. We hypothesize that cross-attention layer features are particularly well-suited for this approach. We selectively manipulate these cross-attention maps to attenuate attention responses associated with negative prompts relative to those elicited by positive prompts, thereby filtering out unwanted semantic attributes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.906, + 0.856 + ], + "angle": 0, + "content": "This representation-space guidance strategy offers a crucial advantage: it enables the exclusion of undesirable elements prior to the final image synthesis stage. This removal mechanism significantly mitigates the risk of generating blending artifacts, which are common byproducts of image-space manipulation techniques." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Our representation-space guidance also extends to architectures like FLUX [25], which replace traditional cross-attention with joint self-attention blocks. [46] show that" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18007" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.086, + 0.895, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.316, + 0.908, + 0.402 + ], + "angle": 0, + "content": "Figure 3. Left: An overview of the Negative-Away Steer Attention (NASA) pipeline. Positive (green) and negative (red) prompts are fed into a text encoder to generate positive and negative text features. The NASA module then processes these features, which adjusts the one-step diffusion model to steer the output image away from the negative features, refining it based on the positive features. Right: The details of the NASA module. It processes queries \\((\\mathbf{Q}_l)\\) in layer \\(l\\), note we will omit the subscript \\(l\\) in subsequent notations to improve readability, with positive \\((\\mathbf{V}^{+},\\mathbf{K}^{+})\\) and negative \\((\\mathbf{V}^{-},\\mathbf{K}^{-})\\) key-value pairs to create positive \\((\\mathbf{Z}^{+})\\) and negative \\((\\mathbf{Z}^{-})\\) attention outputs. The final output \\((\\mathbf{Z}^{\\mathrm{NASA}})\\) is calculated by subtracting the weighted negative features \\((\\mathbf{Z}^{-})\\) from the positive features \\((\\mathbf{Z}^{+})\\)." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.409, + 0.498, + 0.5 + ], + "angle": 0, + "content": "
MethodFLUX.1-schnellSDXL-LCMSDXL-DMD2
4 steps1 step4 steps1 step4 steps1 step
None23%44%43%-27%25%
CFG30%0%14%-25%0%
NASA-I100%99%97%-100%100%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.505, + 0.484, + 0.547 + ], + "angle": 0, + "content": "Table 2. Comparison of success rate of unwanted feature removal in generated images. Note that SDXL-LCM [31] does not support one-step inference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.483, + 0.596 + ], + "angle": 0, + "content": "these blocks still embed distinct image-text cross-attention operations. This allows NASA to be seamlessly applied, proving its robustness across different backbones." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.484, + 0.763 + ], + "angle": 0, + "content": "The overall pipeline of the NASA mechanism is illustrated in Fig. 3. Given the latent feature representation at a given layer \\(l\\), denoted as \\(\\mathbf{Z}_l\\), and positive text features \\(\\mathbf{c}_p\\) extracted from CLIP text encoder, we define the query, key, and value matrices as \\(\\mathbf{Q}_l = \\mathbf{Z}_l\\mathbf{W}_{q,l}\\), \\(\\mathbf{K}_l^+ = \\mathbf{c}_p\\mathbf{W}_{k,l}\\), and \\(\\mathbf{V}_l^+ = \\mathbf{c}_p\\mathbf{W}_{v,l}\\), where \\(\\mathbf{W}_{q,l}\\), \\(\\mathbf{W}_{k,l}\\), and \\(\\mathbf{W}_{v,l}\\) are projection matrices at current layer \\(l\\). Analogously, for negative text features \\(\\mathbf{c}_n\\), we use the same projection matrices to define the corresponding key and value matrices for negative prompt conditioning: \\(\\mathbf{K}_l^- = \\mathbf{c}_n\\mathbf{W}_{k,l}\\) and \\(\\mathbf{V}_l^- = \\mathbf{c}_n\\mathbf{W}_{v,l}\\). The NASA mechanism is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.769, + 0.483, + 0.806 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} _ {l} ^ {+} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} _ {l} \\left(\\mathbf {K} _ {l} ^ {+}\\right) ^ {\\top}}{\\sqrt {d}}\\right) \\mathbf {V} _ {l} ^ {+}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.806, + 0.483, + 0.841 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} _ {l} ^ {-} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} _ {l} \\left(\\mathbf {K} _ {l} ^ {-}\\right) ^ {\\top}}{\\sqrt {d}}\\right) \\mathbf {V} _ {l} ^ {-}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.842, + 0.483, + 0.862 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} _ {l} ^ {\\mathrm {N A S A}} = \\mathbf {Z} _ {l} ^ {+} - \\alpha \\cdot \\mathbf {Z} _ {l} ^ {-}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.9 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is the scale factor to control the degree of negative feature removal." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.411, + 0.787, + 0.427 + ], + "angle": 0, + "content": "5.2. NASA-I: Training-free Setting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.44, + 0.907, + 0.699 + ], + "angle": 0, + "content": "Applicable to One-/Few-Step Models. To assess the effectiveness of NASA-I, we carry out experiments with pretrained one-step and few-step models, such as LCM [31], DMD2 [51] and FLUX.1-schnell [25]. We design a small set of positive and negative prompts and generate 100 images per model, both with and without negative prompts. We then calculate the percentage of images that successfully exclude the feature specified by the negative prompt. For example, given the positive prompt \"A photo of a person\" and the negative prompt \"male\", we measure the percentage of generated images that do not depict a male person. The average results across all prompt sets are summarized in Tab. 2. Our findings show that while traditional CFG is ineffective for one-step and few-step models, NASA-I provides precise control over unwanted features in generated images. Additionally, this capability is also qualitatively demonstrated in Fig. 1, Fig. 4 and Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Efficiency Considerations. As described in Sec. 5.1, we reuse the projection matrices of the positive prompt for the negative prompt, ensuring that no additional model parameters are introduced and keeping the model size unchanged. Furthermore, applying NASA-I to all one-step and few-step models introduces only a minor computational overhead, increasing FLOPs by just \\(1.89\\%\\). This represents a significant efficiency improvement over CFG, which doubles the computational cost due to the need for separate forward passes for positive and negative prompts. Consequently, NASA-I enables negative prompt guidance in one-step diffusion models without introducing substantial latency, making it a practical solution for real-time applications." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18008" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.35, + 0.108 + ], + "angle": 0, + "content": "5.3. NASA-T: Distillation Setting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.486, + 0.462 + ], + "angle": 0, + "content": "Given their effectiveness in improving image quality during inference, their potential utility in model distillation is worth exploring. In this section, we introduce a method that integrates the negative prompt into one-step diffusion distillation. While our approach is applicable to various distillation methods, we select SwiftBrush (SB) [12, 34], an image-free method, as the baseline for improvement due to its efficiency. As described in Eq. (6), SB updates the student model using the gradient of the VSD loss. To incorporate the negative prompt \\(\\mathbf{y}^{-}\\), we modify the existing formulation by replacing \\(\\hat{\\epsilon}_{\\psi}(\\hat{\\mathbf{x}}_t,t,\\mathbf{y},\\kappa)\\) (see Eq. (4)) with \\(\\hat{\\epsilon}_{\\psi}(\\hat{\\mathbf{x}}_t,t,\\mathbf{y},\\mathbf{y}^{-},\\kappa)\\) (see Eq. (5)). In this way, the teacher model provides explicit guidance to the student model, helping it learn to steer away from undesirable features. Moreover, we incorporate the NASA mechanism into the student model, enabling it to process negative prompts during training. This allows the student model to better recognize and suppress unnatural artifacts, leading to improved learning and enhanced image quality. Note that with the LoRA teacher model \\(\\hat{\\epsilon}_{\\pi}\\), we can choose whether to incorporate negative prompts. Empirically, we found that not using them yields better results. Detailed results and discussions are provided in Sec. 6.3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.469, + 0.224, + 0.485 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.494, + 0.281, + 0.511 + ], + "angle": 0, + "content": "6.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.485, + 0.78 + ], + "angle": 0, + "content": "Metrics and Benchmark. In diffusion model evaluation, Fréchet Inception Distance (FID) [18] on the zero-shot MS-COCO2014 [27] benchmark has been the standard metric. However, recent studies [8, 38, 51] suggest that FID in this benchmark does not always correlate with actual image quality and adopt human-focused alignment metrics like Human Preference Score v2 (HPSv2) [48] for benchmarking. Our experiments also follow this conclusion, so we choose HPSv2 as the primary metric for evaluating the models. Moreover, due to the absence of negative prompts in the MS-COCO2014 benchmark, we utilize the NegOpt dataset from [35], which contains pairs of positive and negative prompts. From this dataset, we randomly sample 30K prompt pairs to construct a subset for evaluation. Text alignment is then assessed using CLIP scores, where \\(\\mathrm{CLIP^{+}}\\) evaluates alignment with the positive prompt and \\(\\mathrm{CLIP^{-}}\\) evaluates alignment with the negative prompt." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.785, + 0.485, + 0.878 + ], + "angle": 0, + "content": "Training Datasets. Following SBv2 [12], we use total of 3.3M prompts from JourneyDB [44] and a subset of LAION \\([42]^1\\). For negative prompts in the distillation setting (Sec. 5.3), we construct a dataset of 10K negative prompts derived from commonly used terms like \"bad photo, duplicate, low resolution, low quality\"." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.41 + ], + "angle": 0, + "content": "Implementation Details. We base our method on SBv2 [12] with our proposed modifications in Sec. 5.3. We utilize three backbones SDv1.5 [39], SDv2.1 [39] and PixArt- \\(\\alpha\\) [9] as the frozen teachers and LoRA teachers are initialized with rank \\(r = 64\\) and scaling \\(\\gamma = 128\\). The learning rates are set to \\(1e^{-6}\\) for the student model and \\(1e^{-3}\\) for the LoRA teacher, using the AdamW optimizer [30]. All our training is conducted on four NVIDIA A100 40GB GPUs with the total batch size of 64 for the SD models and 32 for PixArt-\\(\\alpha\\). Moreover, [6] found that instead of using a fixed CFG scale \\(\\kappa\\) during training, sampling \\(\\kappa\\) from a uniform distribution within a given range improves training robustness. Based on this insight, we set the teacher models' CFG scales as follows: \\(\\kappa \\sim \\mathcal{U}(0.5,4)\\) for SDv1.5 and SDv2.1, and \\(\\kappa \\sim \\mathcal{U}(0.5,3)\\) for PixArt-\\(\\alpha\\). For NASA-T, in each iteration, we uniformly sample the scale \\(\\alpha \\sim \\mathcal{U}(0,1)\\) to help the student model adapt to different conditions and enhance generalization. For NASA-I, unless explicitly specified, we set \\(\\alpha = 0.1\\) for SDv1.5-based models, \\(\\alpha = 0.2\\) for SDv2.1-based models, and \\(\\alpha = 0.5\\) for PixArt-\\(\\alpha\\)-based models. Further details can be found in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.416, + 0.653, + 0.431 + ], + "angle": 0, + "content": "6.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.442, + 0.909, + 0.775 + ], + "angle": 0, + "content": "Quantitative Results. Table 3 presents a comparison between our approach and prior distilled text-to-image diffusion models. Using our reimplementation of SBv2 [12] as the baseline, NASA-T consistently enhances image quality across all tested backbones. Compared to other methods, our model achieves the highest HPSv2 scores on average, demonstrating its clear superiority in image generation quality. This suggests that integrating complex loss functions like DMD2 [51] may not always be necessary for achieving state-of-the-art results. Moreover, while applying CFG directly degrades one-step model outputs, NASA-I further enhances image quality across all models when using negative prompts like \"worst quality, low quality, ugly, duplicate, out of frame, deformed, blurry, bad anatomy, watermark\". Notably, our method on the PixArt-\\(\\alpha\\)-based backbone achieves a record-breaking HPSv2 score of 31.21, surpassing all other one-step diffusion models baselines. Regarding the NegOpt dataset, with the support of NASA-I, all models demonstrate a reduction in the \\(\\mathrm{CLIP}^{-}\\) score while maintaining the \\(\\mathrm{CLIP}^{+}\\) score. This indicates our method effectively eliminates unwanted features in generated outputs while preserving desired traits." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Qualitative Results. As discussed in Sec. 5.2, NASA-I is a versatile, training-free method for both one-step and few-step diffusion models. Figure 4 and Figure 5 present qualitative results of integrating NASA-I into the SDXL-DMD2 [51] and FLUX.1-schnell [25] models, respectively. When using a conditional negative prompt, the generated images retain key semantic features while effectively suppressing unwanted attributes specified by the negative prompt." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.424, + 0.9 + ], + "angle": 0, + "content": "1All datasets were downloaded and evaluated at MovianAI." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "18009" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.176, + 0.089, + 0.825, + 0.53 + ], + "angle": 0, + "content": "
DatasetNegOptHPSv2
MethodCLIP+↑CLIP-↓Anime ↑Photo ↑CA ↑Paintings ↑Average ↑
PixArt-α-based backbone
PixArt-α [9] [Teacher]0.350.0529.6229.1728.7928.6929.07
YOSO [32]0.360.0828.7528.0628.5228.5728.48
+ NASA-I0.360.0628.7428.0528.5628.6028.49 (+0.01)
DMD [52]0.350.0829.3128.6728.4628.4128.71
+ CFG = 1.50.340.0930.0227.0728.3628.0728.38 (-0.33)
+ CFG = 2.50.310.1326.7423.8625.1324.6625.10 (-3.61)
+ NASA-I0.350.0529.3328.7128.4928.5328.77 (+0.06)
SBv2*0.360.0932.1929.0930.3929.6930.34
+ NASA-I0.360.0632.6029.5831.0930.6530.98 (+0.64)
+ NASA-T0.350.0832.3329.2630.7530.1030.61 (+0.27)
+ NASA-T + CFG = 1.50.340.1029.4726.5028.2227.6827.97 (-2.37)
+ NASA-T + NASA-I0.350.0632.6629.6731.3130.7131.09 (+0.75)
+ NASA-T + NASA-I (α = 1)0.350.0532.6529.6531.4531.0631.21 (+0.87)
Stable Diffusion 1.5-based backbone
SDv1.5 [39] [Teacher]0.310.0626.5127.1926.0626.1226.47
InstaFlow-0.9B [28]0.330.1126.1026.6225.9225.9526.15
+ NASA-I0.330.1026.1526.6825.9325.9826.19 (+0.04)
DMD2 [51]0.310.1026.3927.0025.8025.8226.25
+ NASA-I0.310.0926.3927.0225.8025.8326.26 (+0.01)
SBv2*0.320.0927.1827.5826.6926.6227.02
+ NASA-I0.320.0827.1927.5926.7126.6327.03 (+0.01)
+ NASA-T0.350.1027.6427.9427.1927.0327.45 (+0.43)
+ NASA-T + NASA-I0.350.0927.6527.9727.1827.0427.46 (+0.44)
Stable Diffusion 2.1-based backbone
SDv2.1 [39] [Teacher]0.330.0627.4826.8926.8627.4627.17
SBv2 [12]0.340.1027.2527.6226.8626.7727.13
+ NASA-I0.350.0827.4427.8426.9127.0227.30 (+0.17)
SBv2*0.370.1027.5627.8426.9727.0327.35
+ NASA-I0.370.0827.7028.0027.1227.2227.51 (+0.16)
+ NASA-T0.360.0928.0028.6527.4427.2627.84 (+0.49)
+ NASA-T + NASA-I0.360.0728.0428.6827.5127.5027.93 (+0.58)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.538, + 0.908, + 0.581 + ], + "angle": 0, + "content": "Table 3. Quantitative comparisons between our method and previous works. NASA-I refers to our training-free method introduced in Sec. 5.2, while NASA-T denotes our distillation-based approach described in Sec. 5.3. SBv2* refers to our reimplementation of SBv2 with a randomly sampled CFG scale \\(\\kappa\\)." + }, + { + "type": "image", + "bbox": [ + 0.153, + 0.592, + 0.847, + 0.846 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.855, + 0.908, + 0.884 + ], + "angle": 0, + "content": "Figure 4. Effect of different scale values (0.0 to 1.0) in NASA-I with SDXL-DMD2 [51] for both one-step and few-step settings, illustrating the progressive influence on visual details and composition." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18010" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.905, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.306, + 0.223, + 0.692, + 0.238 + ], + "angle": 0, + "content": "Figure 5. Qualitative results of NASA-I in FLUX.1-schnell [25]." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.244, + 0.391, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.393, + 0.244, + 0.688, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.697, + 0.258, + 0.901, + 0.329 + ], + "angle": 0, + "content": "
MethodVBenchmark-Long benchmark
Aesthetic Quality ↑Imaging Quality ↑
None61.9867.12
NASA-I63.3367.36
" + }, + { + "type": "image_caption", + "bbox": [ + 0.137, + 0.339, + 0.641, + 0.355 + ], + "angle": 0, + "content": "Figure 6. Qualitative results of NASA-I for text-to-video generation in CausVid [53]." + }, + { + "type": "table_caption", + "bbox": [ + 0.697, + 0.34, + 0.898, + 0.354 + ], + "angle": 0, + "content": "Table 4. CausVid quality metrics." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.37, + 0.477, + 0.443 + ], + "angle": 0, + "content": "
NASA✓✓
Teacher✓✓✓✓
LoRA Teacher✓✓
HPSv2 Avg. ↑30.3429.8429.8030.6129.89
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.45, + 0.483, + 0.478 + ], + "angle": 0, + "content": "Table 5. Ablation studies on the integration of negative prompts in model distillation using the PixArt- \\(\\alpha\\) backbone." + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.492, + 0.465, + 0.659 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.665, + 0.483, + 0.693 + ], + "angle": 0, + "content": "Figure 7. Ablation study on the NASA scale using our distilled model based on the PixArt-\\(\\alpha\\) backbone." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Application to Text-to-Video Models. To demonstrate the versatility of our approach, we extended NASA-I to a few-step video diffusion model, CausVid [53]. As shown in Fig. 6, our method effectively suppresses unwanted visual attributes while preserving key semantic content and temporal consistency. Quantitative evaluation on the VBenchmarkLong [22] benchmark further validates this. Table 4 shows that using NASA-I with common negative prompts like \"low quality, bad anatomy, blurry, distortion, ugly, low resolution, unclear details\" improves both the aesthetic and imaging quality scores of the generated videos. This highlights NASA-I's potential as a general-purpose tool for enhancing guided generation across different modalities." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.373, + 0.68, + 0.387 + ], + "angle": 0, + "content": "6.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.395, + 0.907, + 0.591 + ], + "angle": 0, + "content": "Table 5 presents an ablation on integrating negative prompts during model distillation with the PixArt-\\(\\alpha\\) [9] backbone. The baseline is our reimplementation of SBv2 with a randomly sampled CFG scale \\(\\kappa\\). Simply applying negative prompts only in the teacher and/or the LoRA teacher models slightly degrades the HPSv2 score, as the student model is not explicitly trained on this negative conditioning. In contrast, incorporating NASA with negative prompts during training allows the student model to better capture and utilize this conditioning, leading to improved performance. Furthermore, we find that integrating negative prompts solely into the teacher model yields better results than applying them to both the teacher and LoRA teacher models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.591, + 0.907, + 0.713 + ], + "angle": 0, + "content": "Using our PixArt-\\(\\alpha\\)-based distilled model as the base, we varied the NASA scale \\(\\alpha\\) and evaluated the results with HPSv2 on the benchmark dataset. Figure 7 demonstrates that NASA-I generally enhances image synthesis, resulting in outputs that better align with human preferences. However, if the NASA scale \\(\\alpha\\) is set too high, the negative attributes may overly dominate the positive attributes, adversely affecting the final image quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.724, + 0.763, + 0.74 + ], + "angle": 0, + "content": "7. Discussion and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.748, + 0.907, + 0.903 + ], + "angle": 0, + "content": "This paper introduces NASA, the first method to integrate negative prompt guidance into one-step and few-step diffusion models for both inference and distillation. By leveraging cross-attention adjustments, NASA effectively suppresses unwanted features in generated images. Experimental results show that NASA enhances image quality and outperforms strong baselines, establishing a new state-of-the-art in one-step model distillation based on human preference metrics. While NASA's effectiveness relies on selecting an appropriate scale for suppressing negative features, this parameter does not require extensive tuning." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18011" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.119, + 0.188, + 0.135 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.145, + 0.483, + 0.213 + ], + "angle": 0, + "content": "[1] Mohammadreza Armandpour, Ali Sadeghian, Huangjie Zheng, Amir Sadeghian, and Mingyuan Zhou. Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. ArXiv, abs/2304.04968, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.215, + 0.482, + 0.241 + ], + "angle": 0, + "content": "[2] AUTOMATIC1111. Automatic1111/stable-diffusion-webui, 2025. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.243, + 0.483, + 0.325 + ], + "angle": 0, + "content": "[3] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.327, + 0.483, + 0.395 + ], + "angle": 0, + "content": "[4] Yuanhao Ban, Ruochen Wang, Tianyi Zhou, Minhao Cheng, Boqing Gong, and Cho-Jui Hsieh. Understanding the impact of negative prompts: When and how do they take effect? In European conference on computer vision, pages 190–206. Springer, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.397, + 0.482, + 0.478 + ], + "angle": 0, + "content": "[5] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Yuanzhen Li, Tomer Michaeli, Oliver Wang, Deqing Sun, Tali Dekel, and Inbar Mosseri. Lumiere: A space-time diffusion model for video generation. ArXiv, abs/2401.12945, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.482, + 0.535 + ], + "angle": 0, + "content": "[6] Clement Chadebec, Onur Tasar, Eyal Benaroche, and Benjamin Aubin. Flash diffusion: Accelerating any conditional diffusion model for few steps image generation. arXiv preprint arXiv:2406.02347, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.537, + 0.482, + 0.605 + ], + "angle": 0, + "content": "[7] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James T. Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. ArXiv, abs/2310.00426, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.482, + 0.675 + ], + "angle": 0, + "content": "[8] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- \\(\\sigma\\) : Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv: 2403.04692, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.677, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[9] Junsong Chen, Jincheng YU, Chongjian GE, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.416, + 0.774 + ], + "angle": 0, + "content": "[10] Civitai Platform. https://civitai.com.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.453, + 0.789 + ], + "angle": 0, + "content": "[11] comfyanonymous. comfyanonymous/comfyui, 2025. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[12] Trung Dao, Thuan Hoang Nguyen, Thanh Le, Duc Vu, Khoi Nguyen, Cuong Pham, and Anh Tran. Swiftbrush v2: Make your one-step diffusion model better than its teacher. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[13] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.145, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[14] Zach Evans, CJ Carr, Josiah Taylor, Scott H. Hawley, and Jordi Pons. Fast timing-conditioned latent audio diffusion. ArXiv, abs/2402.04825, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.135, + 0.905, + 0.244 + ], + "angle": 0, + "content": "[15] Peng Gao, Le Zhuo, Ziyi Lin, Chris Liu, Junsong Chen, Ruoyi Du, Enze Xie, Xu Luo, Longtian Qiu, Yuhang Zhang, Chen Lin, Rongjie Huang, Shijie Geng, Renrui Zhang, Junlin Xi, Wenqi Shao, Zhengkai Jiang, Tianshuo Yang, Weicai Ye, He Tong, Jingwen He, Yu Jiao Qiao, and Hongsheng Li. Lumina-t2x: Transforming text into any modality, resolution, and duration via flow-based large diffusion transformers. ArXiv, abs/2405.05945, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.246, + 0.905, + 0.301 + ], + "angle": 0, + "content": "[16] Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.302, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[17] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control.(2022). In International Conference on Learning Representations, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.904, + 0.413 + ], + "angle": 0, + "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Neural Information Processing Systems, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.904, + 0.441 + ], + "angle": 0, + "content": "[19] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. Advances in Neural Information Processing Systems, 33, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[21] Rongjie Huang, Jia-Bin Huang, Dongchao Yang, Yi Ren, Luping Liu, Mingze Li, Zhenhui Ye, Jinglin Liu, Xiaoyue Yin, and Zhou Zhao. Make-an-audio: Text-to-audio generation with prompt-enhanced diffusion models. ArXiv, abs/2301.12661, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.553, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[22] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.651, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[23] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.72, + 0.905, + 0.76 + ], + "angle": 0, + "content": "[24] Oren Katzir, Or Patashnik, Daniel Cohen-Or, and Dani Lischinski. Noise-free score distillation. arXiv preprint arXiv: 2310.17590, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.904, + 0.789 + ], + "angle": 0, + "content": "[25] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024. 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[26] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18012" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[28] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, and Qiang Liu. InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation. arXiv preprint arXiv:2309.06380, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.471, + 0.207 + ], + "angle": 0, + "content": "[29] Illyasviel. Illyasviel/stable-diffusion-webui-forge, 2025. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[30] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.251, + 0.482, + 0.305 + ], + "angle": 0, + "content": "[31] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv: 2310.04378, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.308, + 0.482, + 0.362 + ], + "angle": 0, + "content": "[32] Yihong Luo, Xiaolong Chen, Xinghua Qu, and Jing Tang. You only sample once: Taming one-step text-to-image synthesis by self-cooperative diffusion gans. arXiv preprint arXiv: 2403.12931, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.364, + 0.482, + 0.419 + ], + "angle": 0, + "content": "[33] David McAllister, Songwei Ge, Jia-Bin Huang, David W. Jacobs, Alexei A. Efros, Aleksander Holynski, and Angjoo Kanazawa. Rethinking score distillation as a bridge between image distributions. ArXiv, abs/2406.09417, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.421, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[34] Thuan Hoang Nguyen and Anh Tran. Swiftbrush: One-step text-to-image diffusion model with variational score distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.533 + ], + "angle": 0, + "content": "[35] Michael Ogezi and Ning Shi. Optimizing negative prompts for enhanced aesthetics and fidelity in text-to-image generation. arXiv preprint arXiv:2403.07605, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.535, + 0.482, + 0.603 + ], + "angle": 0, + "content": "[36] Dustin Podell, Zion English, Kyle Lacey, A. Blattmann, Tim Dockhorn, Jonas Muller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. International Conference on Learning Representations, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.605, + 0.482, + 0.646 + ], + "angle": 0, + "content": "[37] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D diffusion. *ArXiv*, abs/2209.14988, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.648, + 0.482, + 0.703 + ], + "angle": 0, + "content": "[38] Yuxi Ren, Xin Xia, Yanzuo Lu, Jiacheng Zhang, Jie Wu, Pan Xie, Xing Wang, and Xuefeng Xiao. Hyper-sd: Trajectory segmented consistency model for efficient image synthesis. arXiv preprint arXiv: 2404.13686, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.705, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[40] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[41] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[42] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. LAION-5B: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.905, + 0.244 + ], + "angle": 0, + "content": "[43] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.246, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[44] Keqiang Sun, Junting Pan, Yuying Ge, Hao Li, Haodong Duan, Xiaoshi Wu, Renrui Zhang, Aojun Zhou, Zipeng Qin, Yi Wang, Jifeng Dai, Yu Qiao, Limin Wang, and Hongsheng Li. JourneyDB: A benchmark for generative image understanding. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[45] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.467 + ], + "angle": 0, + "content": "[46] Tianyi Wei, Dongdong Chen, Yifan Zhou, and Xingang Pan. Enhancing mmdit-based text-to-image models for similar subject generation. arXiv preprint arXiv:2411.18301, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.904, + 0.524 + ], + "angle": 0, + "content": "[47] Max Woolf. Stable diffusion 2.0 and the importance of negative prompts for good results. https://minimaxir.com/2022/11/stable-diffusion-negative-prompt/, 2023.2,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[48] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. ArXiv, abs/2306.09341, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[49] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.65, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[50] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Xiaotao Gu, Yuxuan Zhang, Weihan Wang, Yean Cheng, Ting Liu, Bin Xu, Yuxiao Dong, and Jie Tang. Cogvideox: Text-to-video diffusion models with an expert transformer. ArXiv, abs/2408.06072, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[51] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. In NeurIPS, 2024. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[52] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In CVPR, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[53] Tianwei Yin, Qiang Zhang, Richard Zhang, William T Freeman, Fredo Durand, Eli Shechtman, and Xun Huang. From slow bidirectional to fast autoregressive video diffusion models. 2025. 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18013" + } + ] +] \ No newline at end of file diff --git a/2025/Supercharged One-step Text-to-Image Diffusion Models with Negative Prompts/1564dd9e-386a-48aa-bdfb-0018fc932f68_origin.pdf b/2025/Supercharged One-step Text-to-Image Diffusion Models with Negative Prompts/1564dd9e-386a-48aa-bdfb-0018fc932f68_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..129809fb751e9ab933260f967438d430850962d8 --- /dev/null +++ b/2025/Supercharged One-step Text-to-Image Diffusion Models with Negative Prompts/1564dd9e-386a-48aa-bdfb-0018fc932f68_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:142afe5dc7361aa5e7826487a28cfabfbbfbc9727850cc0bbf3f638d166d605d +size 8113359 diff --git a/2025/Supercharged One-step Text-to-Image Diffusion Models with Negative Prompts/full.md b/2025/Supercharged One-step Text-to-Image Diffusion Models with Negative Prompts/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ab470b1b9162d7ea51d4e3e89cb5013aba7f216e --- /dev/null +++ b/2025/Supercharged One-step Text-to-Image Diffusion Models with Negative Prompts/full.md @@ -0,0 +1,283 @@ +# Supercharged One-step Text-to-Image Diffusion Models with Negative Prompts + +Viet Nguyen $^{1\star}$ , Anh Nguyen $^{1\star}$ , Trung Dao $^{1}$ , Khoi Nguyen $^{1}$ , Cuong Pham $^{1,2}$ , Toan Tran $^{1}$ , Anh Tran $^{1}$ + +$^{1}$ Qualcomm AI Research† $^{2}$ Posts & Telecom. Inst. of Tech. + +![](images/28f5f00cf99608049d3527edd4096bd0f1698bb8cf333253acd97907f5f819a3.jpg) +Figure 1. Negative prompting are a key technique for improving controllability and image quality in text-to-image diffusion models. However, classifier-free guidance (CFG) [19] only supports negative prompt integration in multi-step models due to its iterative nature. Our method, NASA, is the first to enable negative prompting in one-step models, allowing for the suppression of unwanted features and providing greater control over image attributes. + +# Abstract + +The escalating demand for real-time image synthesis has driven significant advancements in one-step diffusion models, which inherently offer expedited generation speeds compared to traditional multi-step methods. However, this enhanced efficiency is frequently accompanied by a compromise in the controllability of image attributes. While negative prompting, typically implemented via classifier-free guidance (CFG), has proven effective for fine-grained control in multi-step models, its application to one-step generators remains largely unaddressed. Due to the lack of iterative refinement, as in multi-step diffusion, directly applying CFG to one-step generation leads to blending artifacts and diminished output quality. To fill this gap, we introduce Negative-Away Steer Attention (NASA), an efficient method that integrates negative prompts into one-step diffusion models. NASA operates within the intermediate representation space by leveraging cross-attention mechanisms to suppress undesired visual attributes. This strategy + +avoids the blending artifacts inherent in output-space guidance and achieves high efficiency, incurring only a minimal $1.89\%$ increase in FLOPs compared to the computational doubling of CFG. Furthermore, NASA can be seamlessly integrated into existing timestep distillation frameworks, enhancing the student's output quality. Experimental results demonstrate that NASA substantially improves controllability and output quality, achieving an HPSv2 score of 31.21, setting a new state-of-the-art benchmark for one-step diffusion models. + +# 1. Introduction + +Diffusion models have recently become popular due to their capacity to produce high-quality, diverse outputs in image [3, 7, 36, 40], 3D [26, 37, 45], audio [14, 21] and video [5, 15, 49, 50] synthesis. Unlike other generative models, such as Generative Adversarial Networks (GANs) [16, 23], diffusion models gradually refine their output through a series of steps, allowing them to achieve diverse yet impressively detailed and realistic outputs. However, this iterative + +refinement is both time-consuming and computationally demanding, which limits their practicality in real-world applications. As a result, there is an increasing interest in techniques to speed up diffusion models while maintaining output quality. Recent distillation methods [12, 34, 51, 52] compress the entire multi-step generation process into a single step, a breakthrough that enables real-time image synthesis. This speedup makes diffusion models practical for applications like content creation and interactive media, where their slow performance was previously a major drawback. Despite these impressive speed advancements in one-step generators, we have identified a significant limitation that affects all these models: the inability to support negative prompting [4, 47]. This feature, which allows users to explicitly exclude unwanted elements, is widely available in multi-step diffusion models but remains notably absent in their single-step alternatives. + +The most widely adopted approach for negative prompting, classifier-free guidance (CFG) [19] works by generating two parallel noise predictions during each denoising step - one incorporating the positive prompt and another using the negative prompt. The system then subtracts the negative prediction from the positive one, effectively steering generation away from unwanted attributes. This process relies on the iterative nature of multi-step diffusion, where corrections can be applied and refined over dozens of steps. + +One-step generators, however, collapse this entire process into a single forward pass, eliminating the iterative refinement cycles that make CFG possible. Without these multiple correction opportunities, there is no straightforward way to incorporate negative guidance into the generation pipeline. Consequently, users face an unappealing choice between the instant generation that might include unwanted elements or waiting longer for more precisely controlled results. This speed-versus-control dilemma represents the most significant practical barrier to widespread adoption of one-step generators in professional creative workflows, highlighting the urgent need for novel approaches that can bridge this gap. + +To address this issue, we introduce a novel method called Negative-Away Steer Attention (NASA) used in inference, which incorporates negative prompt guidance directly into one-step diffusion models. NASA works by adjusting cross-attention layers, effectively reducing unwanted features in generated images. By steering attention within intermediate representation space, NASA offers image control capabilities that were previously achievable only in multi-step diffusion models. + +With our newly developed capability, we have explored integrating negative prompting directly into one-step distillation frameworks, focusing on the SwiftBrush (SB) family of models [12, 34]. The SB approach is compelling for its unique, image-free distillation methodology. By match- + +ing teacher model probability distributions and applying sophisticated training techniques, SBv2 can outperform its own teacher on standard image synthesis benchmarks. Despite these technical achievements, SB models still fall short in human preference evaluations. The core challenge lies in the distillation process: without real image supervision, the system lacks clear signals about which generation patterns should be avoided. This is precisely where negative prompting and our NASA framework come in. + +Committed to maintaining SB's valuable image-free distillation approach while incorporating guidance sampling to align with human preference, we provide the critical missing component, a way to steer the model away from undesirable outputs during training. NASA enables something previously impossible: guidance sampling within a one-step generator. Through this integration, our model achieves an HPSv2 [48] score of 31.21, establishing a new state-of-the-art on human preference metrics for one-step models. + +In summary, our contributions are twofold: + +- We introduce Negative-Away Steer Attention (NASA), the first method to integrate negative prompt guidance into one-step and few-step diffusion models in both inference and distillation settings. +- By integrating NASA into the SwiftBrush (SB) distillation framework, we achieve a new state-of-the-art for one-step diffusion models on human preference metrics. This marks the first successful integration of negative prompt guidance directly into one-step model training. + +# 2. Related Work + +# 2.1. Distribution Matching Distillation + +Recent work [12, 34, 51, 52] has explored text-to-image diffusion distillation through distribution matching techniques. These methods directly align the output distributions of teacher and student models, particularly through Variational Score Distillation (VSD). These techniques can be broadly categorized into two approaches: image-dependent training and image-free training. + +Image-dependent methods rely on real images for adversarial loss [41, 51] or synthetic ones generated by the teacher models for reconstruction loss [52]. Despite their effectiveness in preserving image quality, these approaches face significant practical limitations: they require substantial storage capacity, consume high memory during training, demand extensive computational resources, and risk mode collapse due to intensive image supervision. + +In contrast, image-free approaches [12, 34] have demonstrated promising results by relying solely on prompt text for training, significantly reducing resource requirements while maintaining competitive performance. However, image-free techniques often encounter training instability, particularly within VSD-based frameworks. Existing solu + +tions to stabilize training [52] introduce significant memory and computational overhead, limiting their practical application. Our work addresses this gap by introducing a simple yet effective method that enhances training stability without additional computational burden, making image-free variational score distillation produce better models. + +# 2.2. Sampling Guidance with Negative Prompts + +Classifier-free guidance (CFG) [19] and negative prompting [4, 47] are key techniques for content control during denoising. In multi-step diffusion models, these methods enhance desired image features and suppress undesired ones, improving visual fidelity and reducing artifacts [1, 13, 24]. Score-based approaches like NFSD [24] employ negative prompting to steer generated samples toward real-image distributions, while SDS-Bridge [33] uses negative prompts to model source distributions more accurately. Despite their success, negative prompting has yet to be integrated into one-step or few-step diffusion models. To bridge this gap, we introduce NASA, the first method to enable negative prompting in one-step and few-step diffusion models, enhancing control in faster generation settings. + +# 3. Preliminaries + +Diffusion Models [20, 39, 43] produce high-quality images by progressively denoising inputs. This technique involves a forward process that gradually adds noise to the data and a reverse process that reconstructs the original data by removing noise. While the original diffusion models act on the image space [20, 43], LDM [39] processes on the latent space produced by a pretrained VAE for more efficient computation. In particular, starting with a data point $\mathbf{x}_0$ sampled from an unknown distribution $q(\mathbf{x}_0)$ , the forward process gradually diffuses $\mathbf{x}_0$ into a standard Gaussian noise $\mathbf{x}_T \sim \mathcal{N}(\mathbf{0}, \mathbf{I})$ through $T$ consecutive timesteps, where $\mathbf{I}$ is the identity matrix. At each time-step $t$ , a noisier version of $\mathbf{x}_0$ , denoted as $\mathbf{x}_t$ , is drawn from $q_t(\mathbf{x}_t | \mathbf{x}_0) = \mathcal{N}(\alpha_t \mathbf{x}_0, \sigma_t^2 \mathbf{I})$ using a standard Gaussian noise $\epsilon \sim \mathcal{N}(0, \mathbf{I})$ : + +$$ +\mathbf {x} _ {t} = \alpha_ {t} \mathbf {x} _ {0} + \sigma_ {t} \boldsymbol {\epsilon}, \quad \forall t \in \{0, \dots , T \}, \tag {1} +$$ + +where $\{(\alpha_t,\sigma_t)\}_{t = 1}^T$ defines the noise schedule, with boundary conditions $(\alpha_0,\sigma_0) = (1,0)$ for the clean sample and $(\alpha_{T},\sigma_{T}) = (0,1)$ for pure noise. + +Conversely, the reverse process aims to gradually reconstruct the original data by denoising the input over $T$ steps, starting from an initial random Gaussian noise $\mathbf{x}_T\sim \mathcal{N}(\mathbf{0},\mathbf{I})$ . The model is trained by minimizing the difference between the noise estimated by the model, $\epsilon_{\phi}$ , parameterized by $\phi$ , and the actual noise in Eq. (1): + +$$ +\min _ {\phi} \mathbb {E} _ {t \sim \mathcal {U} (0, T), \epsilon \sim \mathcal {N} (0, I)} \| \epsilon_ {\phi} (\mathbf {x} _ {t}, t) - \boldsymbol {\epsilon} \| _ {2} ^ {2}. \tag {2} +$$ + +Text-to-Image Diffusion Models guide the sampling process using an additional text prompt y to produce outputs + +that are both photorealistic and aligned with the provided textual descriptions. The training objective, slightly modified from the unconditional loss in Eq. (2), is defined as: + +$$ +\min _ {\phi} \mathbb {E} _ {t \sim \mathcal {U} (0, T), \epsilon \sim \mathcal {N} (0, I), \mathbf {y}} \| \epsilon_ {\phi} (\mathbf {x} _ {t}, t, \mathbf {y}) - \epsilon \| _ {2} ^ {2}. \tag {3} +$$ + +Classifier-free Guidance (CFG) [19] is an inference technique designed to enhance the quality of generated images by blending the predictions from both a conditional and an unconditional model. At each sampling step, CFG adjusts the denoiser's output using a control parameter $\kappa > 1$ , allowing controlled guidance that aligns more closely with the desired conditions: + +$$ +\hat {\boldsymbol {\epsilon}} _ {\phi} (\mathbf {x} _ {t}, t, \mathbf {y}, \kappa) = (1 - \kappa) \boldsymbol {\epsilon} _ {\phi} (\mathbf {x} _ {t}, t) + \kappa \boldsymbol {\epsilon} _ {\phi} (\mathbf {x} _ {t}, t, \mathbf {y}). \tag {4} +$$ + +Negative Prompting [4, 47] provides enhanced control by suppressing unwanted features in the generated content. Instead of producing an unconditional output, the model generates an output conditioned on the negative prompt $\mathbf{y}^{-}$ , as follows: + +$$ +\hat {\boldsymbol {\epsilon}} _ {\phi} \left(\mathbf {x} _ {t}, t, \mathbf {y}, \mathbf {y} ^ {-}, \kappa\right) = (1 - \kappa) \boldsymbol {\epsilon} _ {\phi} \left(\mathbf {x} _ {t}, t, \mathbf {y} ^ {-}\right) + \kappa \boldsymbol {\epsilon} _ {\phi} \left(\mathbf {x} _ {t}, t, \mathbf {y}\right). \tag {5} +$$ + +Variational Score Distillation (VSD) is a powerful framework that utilizes pretrained diffusion-based text-to-image models to enhance text-based generation across both 3D [45] and 2D [12, 34, 51, 52] domains. The central aim of VSD is to align the renderings of a differentiable generator with the probability density of plausible images as guided by the 2D diffusion model. To accomplish this, VSD employs a two-teacher approach that uses a fixed pretrained diffusion model $\epsilon_{\psi}$ and an adaptive LoRA-based teacher $\epsilon_{\pi}$ . While training, the student model $f_{\theta}$ produces 2D images $\hat{\mathbf{x}}_0 = f_{\theta}(\mathbf{z},\mathbf{y})$ using an input noise $\mathbf{z} \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ and a text prompt $\mathbf{y}$ . Noisy images $\hat{\mathbf{x}}_t = \alpha_t\hat{\mathbf{x}}_0 + \sigma_t\epsilon$ are then fed into both teacher models. The LoRA teacher model $\epsilon_{\pi}$ aligns with the student distribution by minimizing a denoising L2 loss on single-step samples. This arrangement supports robust and adaptive guidance suitable for a variety of generative architectures. The gradient of the learning objective with respect to the student's parameters $\theta$ is defined as: + +$$ +\begin{array}{l} \nabla_ {\theta} \mathcal {L} _ {\mathrm {V S D}} = \mathbb {E} _ {t, \mathbf {y}, \mathbf {z}, \kappa = C} \left[ w (t) \left(\hat {\epsilon} _ {\psi} (\hat {\mathbf {x}} _ {t}, t, \mathbf {y}, \kappa) \right. \right. \tag {6} \\ \left. - \hat {\epsilon} _ {\pi} (\hat {\mathbf {x}} _ {t}, t, \mathbf {y}, \kappa) \frac {\partial \pmb {f} _ {\theta} (\mathbf {z} , \mathbf {y})}{\partial \theta} \right], \\ \end{array} +$$ + +where $C$ is a constant and $w(t)$ is the time-dependent weight that adjusts the contribution of each timestep, aligning the student's outputs with the diffusion model's predictions. Through alternating updates of the student and LoRA teacher, VSD enables efficient, high-fidelity generation across both text-to-image and text-to-3D tasks. + +
NPAnimePhotoCAPaintingsAverage
29.6229.1728.7928.6929.07
✓31.6629.1730.5830.3830.42
+ +Table 1. The effect of negative prompts (NP) on the HPSv2 score of PixArt- $\alpha$ [9]. Using generic negative prompts to remove visual artifacts (e.g., "worst quality", "blurry", "bad anatomy") leads to a significant improvement. Higher scores indicate better alignment with human preferences. + +![](images/f591d916009fa213f3d60703514f68ce8f4417e072ea7a5bc79ffb06a3ea11ff.jpg) +Figure 2. The inherent single-pass nature of one-step diffusion models renders them incompatible with direct CFG, resulting in artifact-laden outputs. + +# 4. Observations + +# 4.1. Motivation + +Positive prompt is not enough. While vision generative models like diffusion models excel at incorporating positive attributes, they fundamentally struggle with understanding negation. Consider requesting "A portrait of a friendly monster, not scary" (Fig. 1), the model typically fails to exclude the specified negative attribute "scary". Negative prompting addresses this limitation by creating a separate, dedicated input for exclusions – completely removing negation terms from the positive prompt and processing them independently. This clear separation allows the model to accurately distinguish what to avoid generating, providing significantly better control over the final output. + +Community Need. The growing use of negative prompting in the generative AI community further highlights its importance. Platforms like Civit AI [10] and various open-source repositories [2, 11, 29] have shown that using both types of prompts together results in more refined, higher-quality model outputs that align more closely with user specifications. Users can clearly specify both desired and unwanted attributes in a prompt pair, which is particularly useful when working with models that generate complex or highly detailed content. Experimentally, Table 1 and Figure 1 confirm that using negative prompts effectively enhances image quality of the model. + +Critical Gap. Despite the established need and widespread adoption of negative prompts in multi-step diffusion models and the rapidly growing prominence of one-step approaches, a significant limitation persists: no existing one + +step methods effectively support negative prompting. Our proposed method fills this gap by providing a simple, efficient, and training-free mechanism that integrates negative prompt into one-step diffusion models to (1) improve controllability and (2) enhance the quality of the output image. + +# 4.2. Challenges + +A widely used method for implementing negative prompting is the CFG mechanism. However, applying CFG directly to one-step diffusion models presents significant challenges. CFG was designed for multi-step models, where the output is iteratively refined, allowing for gradual adjustments using both negative and unconditional outputs. In contrast, one-step models generate images in a single pass without the iterative denoising process, making them inherently unsuitable for this method. Direct application of CFG to these models leads to undesirable image blending, resulting in unnatural artifacts, as shown in Fig. 2. This highlights a crucial limitation: the CFG mechanism, though effective in multi-step models, is not directly translatable to one-step models without significant modifications. + +# 5. Proposed Method + +Our proposed methodology is detailed in Sec. 5.1 for the core mechanism. We present two primary implementation strategies for NASA: a training-free approach (NASA-I) and a distillation-based approach (NASA-T), which are presented in Sec. 5.2 and Sec. 5.3, respectively. + +# 5.1. NASA: Negative-Away Steer Attention + +In contrast to CFG implementations that operate within the output space (synthesized image), as commonly practiced in multi-step diffusion models, we propose a novel approach that strategically relocates the guidance mechanism to the intermediate representation space of the diffusion model. + +Specifically, cross-attention layers in these models capture semantic connections between image and text features [17], making them ideal for controlling the feature alignment. We hypothesize that cross-attention layer features are particularly well-suited for this approach. We selectively manipulate these cross-attention maps to attenuate attention responses associated with negative prompts relative to those elicited by positive prompts, thereby filtering out unwanted semantic attributes. + +This representation-space guidance strategy offers a crucial advantage: it enables the exclusion of undesirable elements prior to the final image synthesis stage. This removal mechanism significantly mitigates the risk of generating blending artifacts, which are common byproducts of image-space manipulation techniques. + +Our representation-space guidance also extends to architectures like FLUX [25], which replace traditional cross-attention with joint self-attention blocks. [46] show that + +![](images/c1efb25a93a4ba5d98b6406d9914dc0106f9d8ba62cde33b93f4071faefac6a7.jpg) +Figure 3. Left: An overview of the Negative-Away Steer Attention (NASA) pipeline. Positive (green) and negative (red) prompts are fed into a text encoder to generate positive and negative text features. The NASA module then processes these features, which adjusts the one-step diffusion model to steer the output image away from the negative features, refining it based on the positive features. Right: The details of the NASA module. It processes queries $(\mathbf{Q}_l)$ in layer $l$ , note we will omit the subscript $l$ in subsequent notations to improve readability, with positive $(\mathbf{V}^{+},\mathbf{K}^{+})$ and negative $(\mathbf{V}^{-},\mathbf{K}^{-})$ key-value pairs to create positive $(\mathbf{Z}^{+})$ and negative $(\mathbf{Z}^{-})$ attention outputs. The final output $(\mathbf{Z}^{\mathrm{NASA}})$ is calculated by subtracting the weighted negative features $(\mathbf{Z}^{-})$ from the positive features $(\mathbf{Z}^{+})$ . + +
MethodFLUX.1-schnellSDXL-LCMSDXL-DMD2
4 steps1 step4 steps1 step4 steps1 step
None23%44%43%-27%25%
CFG30%0%14%-25%0%
NASA-I100%99%97%-100%100%
+ +Table 2. Comparison of success rate of unwanted feature removal in generated images. Note that SDXL-LCM [31] does not support one-step inference. + +these blocks still embed distinct image-text cross-attention operations. This allows NASA to be seamlessly applied, proving its robustness across different backbones. + +The overall pipeline of the NASA mechanism is illustrated in Fig. 3. Given the latent feature representation at a given layer $l$ , denoted as $\mathbf{Z}_l$ , and positive text features $\mathbf{c}_p$ extracted from CLIP text encoder, we define the query, key, and value matrices as $\mathbf{Q}_l = \mathbf{Z}_l\mathbf{W}_{q,l}$ , $\mathbf{K}_l^+ = \mathbf{c}_p\mathbf{W}_{k,l}$ , and $\mathbf{V}_l^+ = \mathbf{c}_p\mathbf{W}_{v,l}$ , where $\mathbf{W}_{q,l}$ , $\mathbf{W}_{k,l}$ , and $\mathbf{W}_{v,l}$ are projection matrices at current layer $l$ . Analogously, for negative text features $\mathbf{c}_n$ , we use the same projection matrices to define the corresponding key and value matrices for negative prompt conditioning: $\mathbf{K}_l^- = \mathbf{c}_n\mathbf{W}_{k,l}$ and $\mathbf{V}_l^- = \mathbf{c}_n\mathbf{W}_{v,l}$ . The NASA mechanism is formulated as follows: + +$$ +\mathbf {Z} _ {l} ^ {+} = \operatorname {S o f t m a x} \left(\frac {\mathbf {Q} _ {l} \left(\mathbf {K} _ {l} ^ {+}\right) ^ {\top}}{\sqrt {d}}\right) \mathbf {V} _ {l} ^ {+}, \tag {7} +$$ + +$$ +\mathbf {Z} _ {l} ^ {-} = \operatorname {S o f t m a x} \left(\frac {\mathbf {Q} _ {l} \left(\mathbf {K} _ {l} ^ {-}\right) ^ {\top}}{\sqrt {d}}\right) \mathbf {V} _ {l} ^ {-}, \tag {8} +$$ + +$$ +\mathbf {Z} _ {l} ^ {\mathrm {N A S A}} = \mathbf {Z} _ {l} ^ {+} - \alpha \cdot \mathbf {Z} _ {l} ^ {-}, \tag {9} +$$ + +where $\alpha$ is the scale factor to control the degree of negative feature removal. + +# 5.2. NASA-I: Training-free Setting + +Applicable to One-/Few-Step Models. To assess the effectiveness of NASA-I, we carry out experiments with pretrained one-step and few-step models, such as LCM [31], DMD2 [51] and FLUX.1-schnell [25]. We design a small set of positive and negative prompts and generate 100 images per model, both with and without negative prompts. We then calculate the percentage of images that successfully exclude the feature specified by the negative prompt. For example, given the positive prompt "A photo of a person" and the negative prompt "male", we measure the percentage of generated images that do not depict a male person. The average results across all prompt sets are summarized in Tab. 2. Our findings show that while traditional CFG is ineffective for one-step and few-step models, NASA-I provides precise control over unwanted features in generated images. Additionally, this capability is also qualitatively demonstrated in Fig. 1, Fig. 4 and Fig. 5. + +Efficiency Considerations. As described in Sec. 5.1, we reuse the projection matrices of the positive prompt for the negative prompt, ensuring that no additional model parameters are introduced and keeping the model size unchanged. Furthermore, applying NASA-I to all one-step and few-step models introduces only a minor computational overhead, increasing FLOPs by just $1.89\%$ . This represents a significant efficiency improvement over CFG, which doubles the computational cost due to the need for separate forward passes for positive and negative prompts. Consequently, NASA-I enables negative prompt guidance in one-step diffusion models without introducing substantial latency, making it a practical solution for real-time applications. + +# 5.3. NASA-T: Distillation Setting + +Given their effectiveness in improving image quality during inference, their potential utility in model distillation is worth exploring. In this section, we introduce a method that integrates the negative prompt into one-step diffusion distillation. While our approach is applicable to various distillation methods, we select SwiftBrush (SB) [12, 34], an image-free method, as the baseline for improvement due to its efficiency. As described in Eq. (6), SB updates the student model using the gradient of the VSD loss. To incorporate the negative prompt $\mathbf{y}^{-}$ , we modify the existing formulation by replacing $\hat{\epsilon}_{\psi}(\hat{\mathbf{x}}_t,t,\mathbf{y},\kappa)$ (see Eq. (4)) with $\hat{\epsilon}_{\psi}(\hat{\mathbf{x}}_t,t,\mathbf{y},\mathbf{y}^{-},\kappa)$ (see Eq. (5)). In this way, the teacher model provides explicit guidance to the student model, helping it learn to steer away from undesirable features. Moreover, we incorporate the NASA mechanism into the student model, enabling it to process negative prompts during training. This allows the student model to better recognize and suppress unnatural artifacts, leading to improved learning and enhanced image quality. Note that with the LoRA teacher model $\hat{\epsilon}_{\pi}$ , we can choose whether to incorporate negative prompts. Empirically, we found that not using them yields better results. Detailed results and discussions are provided in Sec. 6.3. + +# 6. Experiments + +# 6.1. Experimental Setup + +Metrics and Benchmark. In diffusion model evaluation, Fréchet Inception Distance (FID) [18] on the zero-shot MS-COCO2014 [27] benchmark has been the standard metric. However, recent studies [8, 38, 51] suggest that FID in this benchmark does not always correlate with actual image quality and adopt human-focused alignment metrics like Human Preference Score v2 (HPSv2) [48] for benchmarking. Our experiments also follow this conclusion, so we choose HPSv2 as the primary metric for evaluating the models. Moreover, due to the absence of negative prompts in the MS-COCO2014 benchmark, we utilize the NegOpt dataset from [35], which contains pairs of positive and negative prompts. From this dataset, we randomly sample 30K prompt pairs to construct a subset for evaluation. Text alignment is then assessed using CLIP scores, where $\mathrm{CLIP^{+}}$ evaluates alignment with the positive prompt and $\mathrm{CLIP^{-}}$ evaluates alignment with the negative prompt. + +Training Datasets. Following SBv2 [12], we use total of 3.3M prompts from JourneyDB [44] and a subset of LAION $[42]^1$ . For negative prompts in the distillation setting (Sec. 5.3), we construct a dataset of 10K negative prompts derived from commonly used terms like "bad photo, duplicate, low resolution, low quality". + +Implementation Details. We base our method on SBv2 [12] with our proposed modifications in Sec. 5.3. We utilize three backbones SDv1.5 [39], SDv2.1 [39] and PixArt- $\alpha$ [9] as the frozen teachers and LoRA teachers are initialized with rank $r = 64$ and scaling $\gamma = 128$ . The learning rates are set to $1e^{-6}$ for the student model and $1e^{-3}$ for the LoRA teacher, using the AdamW optimizer [30]. All our training is conducted on four NVIDIA A100 40GB GPUs with the total batch size of 64 for the SD models and 32 for PixArt- $\alpha$ . Moreover, [6] found that instead of using a fixed CFG scale $\kappa$ during training, sampling $\kappa$ from a uniform distribution within a given range improves training robustness. Based on this insight, we set the teacher models' CFG scales as follows: $\kappa \sim \mathcal{U}(0.5,4)$ for SDv1.5 and SDv2.1, and $\kappa \sim \mathcal{U}(0.5,3)$ for PixArt- $\alpha$ . For NASA-T, in each iteration, we uniformly sample the scale $\alpha \sim \mathcal{U}(0,1)$ to help the student model adapt to different conditions and enhance generalization. For NASA-I, unless explicitly specified, we set $\alpha = 0.1$ for SDv1.5-based models, $\alpha = 0.2$ for SDv2.1-based models, and $\alpha = 0.5$ for PixArt- $\alpha$ -based models. Further details can be found in the Appendix. + +# 6.2. Main Results + +Quantitative Results. Table 3 presents a comparison between our approach and prior distilled text-to-image diffusion models. Using our reimplementation of SBv2 [12] as the baseline, NASA-T consistently enhances image quality across all tested backbones. Compared to other methods, our model achieves the highest HPSv2 scores on average, demonstrating its clear superiority in image generation quality. This suggests that integrating complex loss functions like DMD2 [51] may not always be necessary for achieving state-of-the-art results. Moreover, while applying CFG directly degrades one-step model outputs, NASA-I further enhances image quality across all models when using negative prompts like "worst quality, low quality, ugly, duplicate, out of frame, deformed, blurry, bad anatomy, watermark". Notably, our method on the PixArt- $\alpha$ -based backbone achieves a record-breaking HPSv2 score of 31.21, surpassing all other one-step diffusion models baselines. Regarding the NegOpt dataset, with the support of NASA-I, all models demonstrate a reduction in the $\mathrm{CLIP}^{-}$ score while maintaining the $\mathrm{CLIP}^{+}$ score. This indicates our method effectively eliminates unwanted features in generated outputs while preserving desired traits. + +Qualitative Results. As discussed in Sec. 5.2, NASA-I is a versatile, training-free method for both one-step and few-step diffusion models. Figure 4 and Figure 5 present qualitative results of integrating NASA-I into the SDXL-DMD2 [51] and FLUX.1-schnell [25] models, respectively. When using a conditional negative prompt, the generated images retain key semantic features while effectively suppressing unwanted attributes specified by the negative prompt. + +
DatasetNegOptHPSv2
MethodCLIP+↑CLIP-↓Anime ↑Photo ↑CA ↑Paintings ↑Average ↑
PixArt-α-based backbone
PixArt-α [9] [Teacher]0.350.0529.6229.1728.7928.6929.07
YOSO [32]0.360.0828.7528.0628.5228.5728.48
+ NASA-I0.360.0628.7428.0528.5628.6028.49 (+0.01)
DMD [52]0.350.0829.3128.6728.4628.4128.71
+ CFG = 1.50.340.0930.0227.0728.3628.0728.38 (-0.33)
+ CFG = 2.50.310.1326.7423.8625.1324.6625.10 (-3.61)
+ NASA-I0.350.0529.3328.7128.4928.5328.77 (+0.06)
SBv2*0.360.0932.1929.0930.3929.6930.34
+ NASA-I0.360.0632.6029.5831.0930.6530.98 (+0.64)
+ NASA-T0.350.0832.3329.2630.7530.1030.61 (+0.27)
+ NASA-T + CFG = 1.50.340.1029.4726.5028.2227.6827.97 (-2.37)
+ NASA-T + NASA-I0.350.0632.6629.6731.3130.7131.09 (+0.75)
+ NASA-T + NASA-I (α = 1)0.350.0532.6529.6531.4531.0631.21 (+0.87)
Stable Diffusion 1.5-based backbone
SDv1.5 [39] [Teacher]0.310.0626.5127.1926.0626.1226.47
InstaFlow-0.9B [28]0.330.1126.1026.6225.9225.9526.15
+ NASA-I0.330.1026.1526.6825.9325.9826.19 (+0.04)
DMD2 [51]0.310.1026.3927.0025.8025.8226.25
+ NASA-I0.310.0926.3927.0225.8025.8326.26 (+0.01)
SBv2*0.320.0927.1827.5826.6926.6227.02
+ NASA-I0.320.0827.1927.5926.7126.6327.03 (+0.01)
+ NASA-T0.350.1027.6427.9427.1927.0327.45 (+0.43)
+ NASA-T + NASA-I0.350.0927.6527.9727.1827.0427.46 (+0.44)
Stable Diffusion 2.1-based backbone
SDv2.1 [39] [Teacher]0.330.0627.4826.8926.8627.4627.17
SBv2 [12]0.340.1027.2527.6226.8626.7727.13
+ NASA-I0.350.0827.4427.8426.9127.0227.30 (+0.17)
SBv2*0.370.1027.5627.8426.9727.0327.35
+ NASA-I0.370.0827.7028.0027.1227.2227.51 (+0.16)
+ NASA-T0.360.0928.0028.6527.4427.2627.84 (+0.49)
+ NASA-T + NASA-I0.360.0728.0428.6827.5127.5027.93 (+0.58)
+ +Table 3. Quantitative comparisons between our method and previous works. NASA-I refers to our training-free method introduced in Sec. 5.2, while NASA-T denotes our distillation-based approach described in Sec. 5.3. SBv2* refers to our reimplementation of SBv2 with a randomly sampled CFG scale $\kappa$ . + +![](images/cd87ca5eb6940b8ea62f3dbc8670a0a27ee17e081844b9f092029ba30e85a70c.jpg) +Figure 4. Effect of different scale values (0.0 to 1.0) in NASA-I with SDXL-DMD2 [51] for both one-step and few-step settings, illustrating the progressive influence on visual details and composition. + +![](images/393710f27e24d69dcb9494f23cf0cd7e5d701b4345160768a0599150ef649410.jpg) +Figure 5. Qualitative results of NASA-I in FLUX.1-schnell [25]. + +![](images/ce7533494e5f6c15882df42521eeabdbdd5c06b5836bb9f0fe2cdaaf4654f19b.jpg) +Figure 6. Qualitative results of NASA-I for text-to-video generation in CausVid [53]. + +![](images/5933f6c822b2d2568a91d31bf760a8312a61062190b15958f6ff4462019d29ac.jpg) + +
MethodVBenchmark-Long benchmark
Aesthetic Quality ↑Imaging Quality ↑
None61.9867.12
NASA-I63.3367.36
+ +Table 4. CausVid quality metrics. + +
NASA✓✓
Teacher✓✓✓✓
LoRA Teacher✓✓
HPSv2 Avg. ↑30.3429.8429.8030.6129.89
+ +Table 5. Ablation studies on the integration of negative prompts in model distillation using the PixArt- $\alpha$ backbone. + +![](images/57fae8adf7a27b7fb899130db0f95ddf7097e631a9594cdfe0137c0a5709a7c5.jpg) +Figure 7. Ablation study on the NASA scale using our distilled model based on the PixArt- $\alpha$ backbone. + +Application to Text-to-Video Models. To demonstrate the versatility of our approach, we extended NASA-I to a few-step video diffusion model, CausVid [53]. As shown in Fig. 6, our method effectively suppresses unwanted visual attributes while preserving key semantic content and temporal consistency. Quantitative evaluation on the VBenchmarkLong [22] benchmark further validates this. Table 4 shows that using NASA-I with common negative prompts like "low quality, bad anatomy, blurry, distortion, ugly, low resolution, unclear details" improves both the aesthetic and imaging quality scores of the generated videos. This highlights NASA-I's potential as a general-purpose tool for enhancing guided generation across different modalities. + +# 6.3. Ablation Studies + +Table 5 presents an ablation on integrating negative prompts during model distillation with the PixArt- $\alpha$ [9] backbone. The baseline is our reimplementation of SBv2 with a randomly sampled CFG scale $\kappa$ . Simply applying negative prompts only in the teacher and/or the LoRA teacher models slightly degrades the HPSv2 score, as the student model is not explicitly trained on this negative conditioning. In contrast, incorporating NASA with negative prompts during training allows the student model to better capture and utilize this conditioning, leading to improved performance. Furthermore, we find that integrating negative prompts solely into the teacher model yields better results than applying them to both the teacher and LoRA teacher models. + +Using our PixArt- $\alpha$ -based distilled model as the base, we varied the NASA scale $\alpha$ and evaluated the results with HPSv2 on the benchmark dataset. Figure 7 demonstrates that NASA-I generally enhances image synthesis, resulting in outputs that better align with human preferences. However, if the NASA scale $\alpha$ is set too high, the negative attributes may overly dominate the positive attributes, adversely affecting the final image quality. + +# 7. Discussion and Conclusion + +This paper introduces NASA, the first method to integrate negative prompt guidance into one-step and few-step diffusion models for both inference and distillation. By leveraging cross-attention adjustments, NASA effectively suppresses unwanted features in generated images. Experimental results show that NASA enhances image quality and outperforms strong baselines, establishing a new state-of-the-art in one-step model distillation based on human preference metrics. While NASA's effectiveness relies on selecting an appropriate scale for suppressing negative features, this parameter does not require extensive tuning. + +# References + +[1] Mohammadreza Armandpour, Ali Sadeghian, Huangjie Zheng, Amir Sadeghian, and Mingyuan Zhou. Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. ArXiv, abs/2304.04968, 2023. 3 +[2] AUTOMATIC1111. Automatic1111/stable-diffusion-webui, 2025. 4 +[3] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 1 +[4] Yuanhao Ban, Ruochen Wang, Tianyi Zhou, Minhao Cheng, Boqing Gong, and Cho-Jui Hsieh. Understanding the impact of negative prompts: When and how do they take effect? In European conference on computer vision, pages 190–206. Springer, 2024. 2, 3 +[5] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Yuanzhen Li, Tomer Michaeli, Oliver Wang, Deqing Sun, Tali Dekel, and Inbar Mosseri. Lumiere: A space-time diffusion model for video generation. ArXiv, abs/2401.12945, 2024. 1 +[6] Clement Chadebec, Onur Tasar, Eyal Benaroche, and Benjamin Aubin. Flash diffusion: Accelerating any conditional diffusion model for few steps image generation. arXiv preprint arXiv:2406.02347, 2024. 6 +[7] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James T. Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. ArXiv, abs/2310.00426, 2023. 1 +[8] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- $\sigma$ : Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv: 2403.04692, 2024. 6 +[9] Junsong Chen, Jincheng YU, Chongjian GE, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 4, 6, 7, 8 +[10] Civitai Platform. https://civitai.com.4 +[11] comfyanonymous. comfyanonymous/comfyui, 2025. 4 +[12] Trung Dao, Thuan Hoang Nguyen, Thanh Le, Duc Vu, Khoi Nguyen, Cuong Pham, and Anh Tran. Swiftbrush v2: Make your one-step diffusion model better than its teacher. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2, 3, 6, 7 +[13] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 3 + +[14] Zach Evans, CJ Carr, Josiah Taylor, Scott H. Hawley, and Jordi Pons. Fast timing-conditioned latent audio diffusion. ArXiv, abs/2402.04825, 2024. 1 +[15] Peng Gao, Le Zhuo, Ziyi Lin, Chris Liu, Junsong Chen, Ruoyi Du, Enze Xie, Xu Luo, Longtian Qiu, Yuhang Zhang, Chen Lin, Rongjie Huang, Shijie Geng, Renrui Zhang, Junlin Xi, Wenqi Shao, Zhengkai Jiang, Tianshuo Yang, Weicai Ye, He Tong, Jingwen He, Yu Jiao Qiao, and Hongsheng Li. Lumina-t2x: Transforming text into any modality, resolution, and duration via flow-based large diffusion transformers. ArXiv, abs/2405.05945, 2024. 1 +[16] Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1 +[17] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control.(2022). In International Conference on Learning Representations, 2023. 4 +[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Neural Information Processing Systems, 2017. 6 +[19] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 1, 2, 3 +[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. Advances in Neural Information Processing Systems, 33, 2020. 3 +[21] Rongjie Huang, Jia-Bin Huang, Dongchao Yang, Yi Ren, Luping Liu, Mingze Li, Zhenhui Ye, Jinglin Liu, Xiaoyue Yin, and Zhou Zhao. Make-an-audio: Text-to-audio generation with prompt-enhanced diffusion models. ArXiv, abs/2301.12661, 2023. 1 +[22] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 8 +[23] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1 +[24] Oren Katzir, Or Patashnik, Daniel Cohen-Or, and Dani Lischinski. Noise-free score distillation. arXiv preprint arXiv: 2310.17590, 2023. 3 +[25] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024. 4, 5, 6, 8 +[26] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 1 +[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence + +Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6 +[28] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, and Qiang Liu. InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation. arXiv preprint arXiv:2309.06380, 2023. 7 +[29] Illyasviel. Illyasviel/stable-diffusion-webui-forge, 2025. 4 +[30] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 6 +[31] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv: 2310.04378, 2023. 5 +[32] Yihong Luo, Xiaolong Chen, Xinghua Qu, and Jing Tang. You only sample once: Taming one-step text-to-image synthesis by self-cooperative diffusion gans. arXiv preprint arXiv: 2403.12931, 2024. 7 +[33] David McAllister, Songwei Ge, Jia-Bin Huang, David W. Jacobs, Alexei A. Efros, Aleksander Holynski, and Angjoo Kanazawa. Rethinking score distillation as a bridge between image distributions. ArXiv, abs/2406.09417, 2024. 3 +[34] Thuan Hoang Nguyen and Anh Tran. Swiftbrush: One-step text-to-image diffusion model with variational score distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 6 +[35] Michael Ogezi and Ning Shi. Optimizing negative prompts for enhanced aesthetics and fidelity in text-to-image generation. arXiv preprint arXiv:2403.07605, 2024. 6 +[36] Dustin Podell, Zion English, Kyle Lacey, A. Blattmann, Tim Dockhorn, Jonas Muller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. International Conference on Learning Representations, 2023. 1 +[37] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D diffusion. *ArXiv*, abs/2209.14988, 2022. 1 +[38] Yuxi Ren, Xin Xia, Yanzuo Lu, Jiacheng Zhang, Jie Wu, Pan Xie, Xing Wang, and Xuefeng Xiao. Hyper-sd: Trajectory segmented consistency model for efficient image synthesis. arXiv preprint arXiv: 2404.13686, 2024. 6 +[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 3, 6, 7 +[40] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1 +[41] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 2 + +[42] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. LAION-5B: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 6 +[43] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 3 +[44] Keqiang Sun, Junting Pan, Yuying Ge, Hao Li, Haodong Duan, Xiaoshi Wu, Renrui Zhang, Aojun Zhou, Zipeng Qin, Yi Wang, Jifeng Dai, Yu Qiao, Limin Wang, and Hongsheng Li. JourneyDB: A benchmark for generative image understanding. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 6 +[45] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 1, 3 +[46] Tianyi Wei, Dongdong Chen, Yifan Zhou, and Xingang Pan. Enhancing mmdit-based text-to-image models for similar subject generation. arXiv preprint arXiv:2411.18301, 2024. 4 +[47] Max Woolf. Stable diffusion 2.0 and the importance of negative prompts for good results. https://minimaxir.com/2022/11/stable-diffusion-negative-prompt/, 2023.2,3 +[48] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. ArXiv, abs/2306.09341, 2023. 2, 6 +[49] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190, 2023. 1 +[50] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Xiaotao Gu, Yuxuan Zhang, Weihan Wang, Yean Cheng, Ting Liu, Bin Xu, Yuxiao Dong, and Jie Tang. Cogvideox: Text-to-video diffusion models with an expert transformer. ArXiv, abs/2408.06072, 2024. 1 +[51] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. In NeurIPS, 2024. 2, 3, 5, 6, 7 +[52] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In CVPR, 2024. 2, 3, 7 +[53] Tianwei Yin, Qiang Zhang, Richard Zhang, William T Freeman, Fredo Durand, Eli Shechtman, and Xun Huang. From slow bidirectional to fast autoregressive video diffusion models. 2025. 8 \ No newline at end of file diff --git a/2025/Supercharged One-step Text-to-Image Diffusion Models with Negative Prompts/images.zip b/2025/Supercharged One-step Text-to-Image Diffusion Models with Negative Prompts/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8dea02973bea62d946effd2e7dffbed6c79189b5 --- /dev/null +++ b/2025/Supercharged One-step Text-to-Image Diffusion Models with Negative Prompts/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d7b5c973ef0503e4e3bb291d7e2ea13d5f0e482badcdf464f614eda90175ad4 +size 809870 diff --git a/2025/Supercharged One-step Text-to-Image Diffusion Models with Negative Prompts/layout.json b/2025/Supercharged One-step Text-to-Image Diffusion Models with Negative Prompts/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..76914083d0fe9801159406c03bbb5abe92848cf9 --- /dev/null +++ b/2025/Supercharged One-step Text-to-Image Diffusion Models with Negative Prompts/layout.json @@ -0,0 +1,8217 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 60, + 103, + 550, + 123 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 103, + 550, + 123 + ], + "spans": [ + { + "bbox": [ + 60, + 103, + 550, + 123 + ], + "type": "text", + "content": "Supercharged One-step Text-to-Image Diffusion Models with Negative Prompts" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 142, + 545, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 142, + 545, + 159 + ], + "spans": [ + { + "bbox": [ + 63, + 142, + 545, + 159 + ], + "type": "text", + "content": "Viet Nguyen" + }, + { + "bbox": [ + 63, + 142, + 545, + 159 + ], + "type": "inline_equation", + "content": "^{1\\star}" + }, + { + "bbox": [ + 63, + 142, + 545, + 159 + ], + "type": "text", + "content": ", Anh Nguyen" + }, + { + "bbox": [ + 63, + 142, + 545, + 159 + ], + "type": "inline_equation", + "content": "^{1\\star}" + }, + { + "bbox": [ + 63, + 142, + 545, + 159 + ], + "type": "text", + "content": ", Trung Dao" + }, + { + "bbox": [ + 63, + 142, + 545, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 63, + 142, + 545, + 159 + ], + "type": "text", + "content": ", Khoi Nguyen" + }, + { + "bbox": [ + 63, + 142, + 545, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 63, + 142, + 545, + 159 + ], + "type": "text", + "content": ", Cuong Pham" + }, + { + "bbox": [ + 63, + 142, + 545, + 159 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 63, + 142, + 545, + 159 + ], + "type": "text", + "content": ", Toan Tran" + }, + { + "bbox": [ + 63, + 142, + 545, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 63, + 142, + 545, + 159 + ], + "type": "text", + "content": ", Anh Tran" + }, + { + "bbox": [ + 63, + 142, + 545, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 156, + 171, + 455, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 171, + 455, + 186 + ], + "spans": [ + { + "bbox": [ + 156, + 171, + 455, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 156, + 171, + 455, + 186 + ], + "type": "text", + "content": "Qualcomm AI Research† " + }, + { + "bbox": [ + 156, + 171, + 455, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 156, + 171, + 455, + 186 + ], + "type": "text", + "content": "Posts & Telecom. Inst. of Tech." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 58, + 201, + 555, + 387 + ], + "blocks": [ + { + "bbox": [ + 58, + 201, + 555, + 387 + ], + "lines": [ + { + "bbox": [ + 58, + 201, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 58, + 201, + 555, + 387 + ], + "type": "image", + "image_path": "28f5f00cf99608049d3527edd4096bd0f1698bb8cf333253acd97907f5f819a3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 394, + 555, + 439 + ], + "lines": [ + { + "bbox": [ + 54, + 394, + 555, + 439 + ], + "spans": [ + { + "bbox": [ + 54, + 394, + 555, + 439 + ], + "type": "text", + "content": "Figure 1. Negative prompting are a key technique for improving controllability and image quality in text-to-image diffusion models. However, classifier-free guidance (CFG) [19] only supports negative prompt integration in multi-step models due to its iterative nature. Our method, NASA, is the first to enable negative prompting in one-step models, allowing for the suppression of unwanted features and providing greater control over image attributes." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 445, + 200, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 445, + 200, + 456 + ], + "spans": [ + { + "bbox": [ + 151, + 445, + 200, + 456 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 469, + 297, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 469, + 297, + 685 + ], + "spans": [ + { + "bbox": [ + 54, + 469, + 297, + 685 + ], + "type": "text", + "content": "The escalating demand for real-time image synthesis has driven significant advancements in one-step diffusion models, which inherently offer expedited generation speeds compared to traditional multi-step methods. However, this enhanced efficiency is frequently accompanied by a compromise in the controllability of image attributes. While negative prompting, typically implemented via classifier-free guidance (CFG), has proven effective for fine-grained control in multi-step models, its application to one-step generators remains largely unaddressed. Due to the lack of iterative refinement, as in multi-step diffusion, directly applying CFG to one-step generation leads to blending artifacts and diminished output quality. To fill this gap, we introduce Negative-Away Steer Attention (NASA), an efficient method that integrates negative prompts into one-step diffusion models. NASA operates within the intermediate representation space by leveraging cross-attention mechanisms to suppress undesired visual attributes. This strategy" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 445, + 556, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 556, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 556, + 564 + ], + "type": "text", + "content": "avoids the blending artifacts inherent in output-space guidance and achieves high efficiency, incurring only a minimal " + }, + { + "bbox": [ + 313, + 445, + 556, + 564 + ], + "type": "inline_equation", + "content": "1.89\\%" + }, + { + "bbox": [ + 313, + 445, + 556, + 564 + ], + "type": "text", + "content": " increase in FLOPs compared to the computational doubling of CFG. Furthermore, NASA can be seamlessly integrated into existing timestep distillation frameworks, enhancing the student's output quality. Experimental results demonstrate that NASA substantially improves controllability and output quality, achieving an HPSv2 score of 31.21, setting a new state-of-the-art benchmark for one-step diffusion models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 595, + 395, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 595, + 395, + 607 + ], + "spans": [ + { + "bbox": [ + 314, + 595, + 395, + 607 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Diffusion models have recently become popular due to their capacity to produce high-quality, diverse outputs in image [3, 7, 36, 40], 3D [26, 37, 45], audio [14, 21] and video [5, 15, 49, 50] synthesis. Unlike other generative models, such as Generative Adversarial Networks (GANs) [16, 23], diffusion models gradually refine their output through a series of steps, allowing them to achieve diverse yet impressively detailed and realistic outputs. However, this iterative" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 693, + 128, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 693, + 128, + 703 + ], + "spans": [ + { + "bbox": [ + 56, + 693, + 128, + 703 + ], + "type": "text", + "content": "\\* Equal Contribution" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 703, + 294, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 703, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 703, + 294, + 713 + ], + "type": "text", + "content": "† Qualcomm AI Research is an initiative of Qualcomm Technologies, Inc." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18004" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "type": "text", + "content": "refinement is both time-consuming and computationally demanding, which limits their practicality in real-world applications. As a result, there is an increasing interest in techniques to speed up diffusion models while maintaining output quality. Recent distillation methods [12, 34, 51, 52] compress the entire multi-step generation process into a single step, a breakthrough that enables real-time image synthesis. This speedup makes diffusion models practical for applications like content creation and interactive media, where their slow performance was previously a major drawback. Despite these impressive speed advancements in one-step generators, we have identified a significant limitation that affects all these models: the inability to support negative prompting [4, 47]. This feature, which allows users to explicitly exclude unwanted elements, is widely available in multi-step diffusion models but remains notably absent in their single-step alternatives." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 277, + 294, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 294, + 385 + ], + "type": "text", + "content": "The most widely adopted approach for negative prompting, classifier-free guidance (CFG) [19] works by generating two parallel noise predictions during each denoising step - one incorporating the positive prompt and another using the negative prompt. The system then subtracts the negative prediction from the positive one, effectively steering generation away from unwanted attributes. This process relies on the iterative nature of multi-step diffusion, where corrections can be applied and refined over dozens of steps." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 387, + 294, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 294, + 542 + ], + "type": "text", + "content": "One-step generators, however, collapse this entire process into a single forward pass, eliminating the iterative refinement cycles that make CFG possible. Without these multiple correction opportunities, there is no straightforward way to incorporate negative guidance into the generation pipeline. Consequently, users face an unappealing choice between the instant generation that might include unwanted elements or waiting longer for more precisely controlled results. This speed-versus-control dilemma represents the most significant practical barrier to widespread adoption of one-step generators in professional creative workflows, highlighting the urgent need for novel approaches that can bridge this gap." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 544, + 294, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 294, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 294, + 651 + ], + "type": "text", + "content": "To address this issue, we introduce a novel method called Negative-Away Steer Attention (NASA) used in inference, which incorporates negative prompt guidance directly into one-step diffusion models. NASA works by adjusting cross-attention layers, effectively reducing unwanted features in generated images. By steering attention within intermediate representation space, NASA offers image control capabilities that were previously achievable only in multi-step diffusion models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 653, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 294, + 713 + ], + "type": "text", + "content": "With our newly developed capability, we have explored integrating negative prompting directly into one-step distillation frameworks, focusing on the SwiftBrush (SB) family of models [12, 34]. The SB approach is compelling for its unique, image-free distillation methodology. By match-" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": "ing teacher model probability distributions and applying sophisticated training techniques, SBv2 can outperform its own teacher on standard image synthesis benchmarks. Despite these technical achievements, SB models still fall short in human preference evaluations. The core challenge lies in the distillation process: without real image supervision, the system lacks clear signals about which generation patterns should be avoided. This is precisely where negative prompting and our NASA framework come in." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 180, + 553, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 553, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 553, + 288 + ], + "type": "text", + "content": "Committed to maintaining SB's valuable image-free distillation approach while incorporating guidance sampling to align with human preference, we provide the critical missing component, a way to steer the model away from undesirable outputs during training. NASA enables something previously impossible: guidance sampling within a one-step generator. Through this integration, our model achieves an HPSv2 [48] score of 31.21, establishing a new state-of-the-art on human preference metrics for one-step models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 326, + 289, + 500, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 289, + 500, + 300 + ], + "spans": [ + { + "bbox": [ + 326, + 289, + 500, + 300 + ], + "type": "text", + "content": "In summary, our contributions are twofold:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 304, + 553, + 411 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 304, + 553, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 304, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 314, + 304, + 553, + 350 + ], + "type": "text", + "content": "- We introduce Negative-Away Steer Attention (NASA), the first method to integrate negative prompt guidance into one-step and few-step diffusion models in both inference and distillation settings." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 351, + 553, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 351, + 553, + 411 + ], + "spans": [ + { + "bbox": [ + 315, + 351, + 553, + 411 + ], + "type": "text", + "content": "- By integrating NASA into the SwiftBrush (SB) distillation framework, we achieve a new state-of-the-art for one-step diffusion models on human preference metrics. This marks the first successful integration of negative prompt guidance directly into one-step model training." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 423, + 400, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 423, + 400, + 435 + ], + "spans": [ + { + "bbox": [ + 314, + 423, + 400, + 435 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 443, + 495, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 443, + 495, + 455 + ], + "spans": [ + { + "bbox": [ + 314, + 443, + 495, + 455 + ], + "type": "text", + "content": "2.1. Distribution Matching Distillation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 461, + 553, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 553, + 544 + ], + "type": "text", + "content": "Recent work [12, 34, 51, 52] has explored text-to-image diffusion distillation through distribution matching techniques. These methods directly align the output distributions of teacher and student models, particularly through Variational Score Distillation (VSD). These techniques can be broadly categorized into two approaches: image-dependent training and image-free training." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 545, + 553, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 553, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 553, + 640 + ], + "type": "text", + "content": "Image-dependent methods rely on real images for adversarial loss [41, 51] or synthetic ones generated by the teacher models for reconstruction loss [52]. Despite their effectiveness in preserving image quality, these approaches face significant practical limitations: they require substantial storage capacity, consume high memory during training, demand extensive computational resources, and risk mode collapse due to intensive image supervision." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "content": "In contrast, image-free approaches [12, 34] have demonstrated promising results by relying solely on prompt text for training, significantly reducing resource requirements while maintaining competitive performance. However, image-free techniques often encounter training instability, particularly within VSD-based frameworks. Existing solu" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18005" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": "tions to stabilize training [52] introduce significant memory and computational overhead, limiting their practical application. Our work addresses this gap by introducing a simple yet effective method that enhances training stability without additional computational burden, making image-free variational score distillation produce better models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 149, + 280, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 149, + 280, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 149, + 280, + 163 + ], + "type": "text", + "content": "2.2. Sampling Guidance with Negative Prompts" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 167, + 296, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 167, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 296, + 335 + ], + "type": "text", + "content": "Classifier-free guidance (CFG) [19] and negative prompting [4, 47] are key techniques for content control during denoising. In multi-step diffusion models, these methods enhance desired image features and suppress undesired ones, improving visual fidelity and reducing artifacts [1, 13, 24]. Score-based approaches like NFSD [24] employ negative prompting to steer generated samples toward real-image distributions, while SDS-Bridge [33] uses negative prompts to model source distributions more accurately. Despite their success, negative prompting has yet to be integrated into one-step or few-step diffusion models. To bridge this gap, we introduce NASA, the first method to enable negative prompting in one-step and few-step diffusion models, enhancing control in faster generation settings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 343, + 140, + 356 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 140, + 356 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 140, + 356 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 363, + 296, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 296, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 296, + 532 + ], + "type": "text", + "content": "Diffusion Models [20, 39, 43] produce high-quality images by progressively denoising inputs. This technique involves a forward process that gradually adds noise to the data and a reverse process that reconstructs the original data by removing noise. While the original diffusion models act on the image space [20, 43], LDM [39] processes on the latent space produced by a pretrained VAE for more efficient computation. In particular, starting with a data point " + }, + { + "bbox": [ + 55, + 363, + 296, + 532 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 55, + 363, + 296, + 532 + ], + "type": "text", + "content": " sampled from an unknown distribution " + }, + { + "bbox": [ + 55, + 363, + 296, + 532 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_0)" + }, + { + "bbox": [ + 55, + 363, + 296, + 532 + ], + "type": "text", + "content": ", the forward process gradually diffuses " + }, + { + "bbox": [ + 55, + 363, + 296, + 532 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 55, + 363, + 296, + 532 + ], + "type": "text", + "content": " into a standard Gaussian noise " + }, + { + "bbox": [ + 55, + 363, + 296, + 532 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 55, + 363, + 296, + 532 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 55, + 363, + 296, + 532 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 363, + 296, + 532 + ], + "type": "text", + "content": " consecutive timesteps, where " + }, + { + "bbox": [ + 55, + 363, + 296, + 532 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 55, + 363, + 296, + 532 + ], + "type": "text", + "content": " is the identity matrix. At each time-step " + }, + { + "bbox": [ + 55, + 363, + 296, + 532 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 363, + 296, + 532 + ], + "type": "text", + "content": ", a noisier version of " + }, + { + "bbox": [ + 55, + 363, + 296, + 532 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 55, + 363, + 296, + 532 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 55, + 363, + 296, + 532 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 55, + 363, + 296, + 532 + ], + "type": "text", + "content": ", is drawn from " + }, + { + "bbox": [ + 55, + 363, + 296, + 532 + ], + "type": "inline_equation", + "content": "q_t(\\mathbf{x}_t | \\mathbf{x}_0) = \\mathcal{N}(\\alpha_t \\mathbf{x}_0, \\sigma_t^2 \\mathbf{I})" + }, + { + "bbox": [ + 55, + 363, + 296, + 532 + ], + "type": "text", + "content": " using a standard Gaussian noise " + }, + { + "bbox": [ + 55, + 363, + 296, + 532 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 55, + 363, + 296, + 532 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 97, + 536, + 294, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 536, + 294, + 550 + ], + "spans": [ + { + "bbox": [ + 97, + 536, + 294, + 550 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {t} = \\alpha_ {t} \\mathbf {x} _ {0} + \\sigma_ {t} \\boldsymbol {\\epsilon}, \\quad \\forall t \\in \\{0, \\dots , T \\}, \\tag {1}", + "image_path": "abe85bd28eef26dce3dbb70a5eeeda23baa689f0f856889b2958778668ed9204.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 553, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 553, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 553, + 295, + 590 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 553, + 295, + 590 + ], + "type": "inline_equation", + "content": "\\{(\\alpha_t,\\sigma_t)\\}_{t = 1}^T" + }, + { + "bbox": [ + 55, + 553, + 295, + 590 + ], + "type": "text", + "content": " defines the noise schedule, with boundary conditions " + }, + { + "bbox": [ + 55, + 553, + 295, + 590 + ], + "type": "inline_equation", + "content": "(\\alpha_0,\\sigma_0) = (1,0)" + }, + { + "bbox": [ + 55, + 553, + 295, + 590 + ], + "type": "text", + "content": " for the clean sample and " + }, + { + "bbox": [ + 55, + 553, + 295, + 590 + ], + "type": "inline_equation", + "content": "(\\alpha_{T},\\sigma_{T}) = (0,1)" + }, + { + "bbox": [ + 55, + 553, + 295, + 590 + ], + "type": "text", + "content": " for pure noise." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 590, + 296, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 296, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 296, + 662 + ], + "type": "text", + "content": "Conversely, the reverse process aims to gradually reconstruct the original data by denoising the input over " + }, + { + "bbox": [ + 55, + 590, + 296, + 662 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 590, + 296, + 662 + ], + "type": "text", + "content": " steps, starting from an initial random Gaussian noise " + }, + { + "bbox": [ + 55, + 590, + 296, + 662 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 55, + 590, + 296, + 662 + ], + "type": "text", + "content": ". The model is trained by minimizing the difference between the noise estimated by the model, " + }, + { + "bbox": [ + 55, + 590, + 296, + 662 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}" + }, + { + "bbox": [ + 55, + 590, + 296, + 662 + ], + "type": "text", + "content": ", parameterized by " + }, + { + "bbox": [ + 55, + 590, + 296, + 662 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 590, + 296, + 662 + ], + "type": "text", + "content": ", and the actual noise in Eq. (1):" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 91, + 666, + 295, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 666, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 91, + 666, + 295, + 685 + ], + "type": "interline_equation", + "content": "\\min _ {\\phi} \\mathbb {E} _ {t \\sim \\mathcal {U} (0, T), \\epsilon \\sim \\mathcal {N} (0, I)} \\| \\epsilon_ {\\phi} (\\mathbf {x} _ {t}, t) - \\boldsymbol {\\epsilon} \\| _ {2} ^ {2}. \\tag {2}", + "image_path": "275a2ca49f0f5e5fbb1bf3564637d836d011907839ca4a96b42981dbbb004526.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Text-to-Image Diffusion Models guide the sampling process using an additional text prompt y to produce outputs" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "content": "that are both photorealistic and aligned with the provided textual descriptions. The training objective, slightly modified from the unconditional loss in Eq. (2), is defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 340, + 119, + 555, + 138 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 119, + 555, + 138 + ], + "spans": [ + { + "bbox": [ + 340, + 119, + 555, + 138 + ], + "type": "interline_equation", + "content": "\\min _ {\\phi} \\mathbb {E} _ {t \\sim \\mathcal {U} (0, T), \\epsilon \\sim \\mathcal {N} (0, I), \\mathbf {y}} \\| \\epsilon_ {\\phi} (\\mathbf {x} _ {t}, t, \\mathbf {y}) - \\epsilon \\| _ {2} ^ {2}. \\tag {3}", + "image_path": "5a069a53103c367d9eed7a35b2e071ea2e78d108e73b2cbf35e4e6cca336d71f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 144, + 555, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 228 + ], + "type": "text", + "content": "Classifier-free Guidance (CFG) [19] is an inference technique designed to enhance the quality of generated images by blending the predictions from both a conditional and an unconditional model. At each sampling step, CFG adjusts the denoiser's output using a control parameter " + }, + { + "bbox": [ + 313, + 144, + 555, + 228 + ], + "type": "inline_equation", + "content": "\\kappa > 1" + }, + { + "bbox": [ + 313, + 144, + 555, + 228 + ], + "type": "text", + "content": ", allowing controlled guidance that aligns more closely with the desired conditions:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 323, + 237, + 555, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 237, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 323, + 237, + 555, + 251 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {\\epsilon}} _ {\\phi} (\\mathbf {x} _ {t}, t, \\mathbf {y}, \\kappa) = (1 - \\kappa) \\boldsymbol {\\epsilon} _ {\\phi} (\\mathbf {x} _ {t}, t) + \\kappa \\boldsymbol {\\epsilon} _ {\\phi} (\\mathbf {x} _ {t}, t, \\mathbf {y}). \\tag {4}", + "image_path": "a5de797430bd92a227038f1c15fada669f41c7f2e7f00f02bbb986f04cbb9cd9.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 260, + 554, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 554, + 319 + ], + "type": "text", + "content": "Negative Prompting [4, 47] provides enhanced control by suppressing unwanted features in the generated content. Instead of producing an unconditional output, the model generates an output conditioned on the negative prompt " + }, + { + "bbox": [ + 313, + 260, + 554, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^{-}" + }, + { + "bbox": [ + 313, + 260, + 554, + 319 + ], + "type": "text", + "content": ", as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 330, + 554, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 330, + 554, + 354 + ], + "spans": [ + { + "bbox": [ + 314, + 330, + 554, + 354 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {\\epsilon}} _ {\\phi} \\left(\\mathbf {x} _ {t}, t, \\mathbf {y}, \\mathbf {y} ^ {-}, \\kappa\\right) = (1 - \\kappa) \\boldsymbol {\\epsilon} _ {\\phi} \\left(\\mathbf {x} _ {t}, t, \\mathbf {y} ^ {-}\\right) + \\kappa \\boldsymbol {\\epsilon} _ {\\phi} \\left(\\mathbf {x} _ {t}, t, \\mathbf {y}\\right). \\tag {5}", + "image_path": "8a9a11ce784b2318faa5bd712d333760466689f698edac3c366c9af0243bac9a.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 354, + 555, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 570 + ], + "type": "text", + "content": "Variational Score Distillation (VSD) is a powerful framework that utilizes pretrained diffusion-based text-to-image models to enhance text-based generation across both 3D [45] and 2D [12, 34, 51, 52] domains. The central aim of VSD is to align the renderings of a differentiable generator with the probability density of plausible images as guided by the 2D diffusion model. To accomplish this, VSD employs a two-teacher approach that uses a fixed pretrained diffusion model " + }, + { + "bbox": [ + 313, + 354, + 555, + 570 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\psi}" + }, + { + "bbox": [ + 313, + 354, + 555, + 570 + ], + "type": "text", + "content": " and an adaptive LoRA-based teacher " + }, + { + "bbox": [ + 313, + 354, + 555, + 570 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\pi}" + }, + { + "bbox": [ + 313, + 354, + 555, + 570 + ], + "type": "text", + "content": ". While training, the student model " + }, + { + "bbox": [ + 313, + 354, + 555, + 570 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 313, + 354, + 555, + 570 + ], + "type": "text", + "content": " produces 2D images " + }, + { + "bbox": [ + 313, + 354, + 555, + 570 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0 = f_{\\theta}(\\mathbf{z},\\mathbf{y})" + }, + { + "bbox": [ + 313, + 354, + 555, + 570 + ], + "type": "text", + "content": " using an input noise " + }, + { + "bbox": [ + 313, + 354, + 555, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 313, + 354, + 555, + 570 + ], + "type": "text", + "content": " and a text prompt " + }, + { + "bbox": [ + 313, + 354, + 555, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 313, + 354, + 555, + 570 + ], + "type": "text", + "content": ". Noisy images " + }, + { + "bbox": [ + 313, + 354, + 555, + 570 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_t = \\alpha_t\\hat{\\mathbf{x}}_0 + \\sigma_t\\epsilon" + }, + { + "bbox": [ + 313, + 354, + 555, + 570 + ], + "type": "text", + "content": " are then fed into both teacher models. The LoRA teacher model " + }, + { + "bbox": [ + 313, + 354, + 555, + 570 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\pi}" + }, + { + "bbox": [ + 313, + 354, + 555, + 570 + ], + "type": "text", + "content": " aligns with the student distribution by minimizing a denoising L2 loss on single-step samples. This arrangement supports robust and adaptive guidance suitable for a variety of generative architectures. The gradient of the learning objective with respect to the student's parameters " + }, + { + "bbox": [ + 313, + 354, + 555, + 570 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 354, + 555, + 570 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 578, + 554, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 578, + 554, + 633 + ], + "spans": [ + { + "bbox": [ + 321, + 578, + 554, + 633 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {V S D}} = \\mathbb {E} _ {t, \\mathbf {y}, \\mathbf {z}, \\kappa = C} \\left[ w (t) \\left(\\hat {\\epsilon} _ {\\psi} (\\hat {\\mathbf {x}} _ {t}, t, \\mathbf {y}, \\kappa) \\right. \\right. \\tag {6} \\\\ \\left. - \\hat {\\epsilon} _ {\\pi} (\\hat {\\mathbf {x}} _ {t}, t, \\mathbf {y}, \\kappa) \\frac {\\partial \\pmb {f} _ {\\theta} (\\mathbf {z} , \\mathbf {y})}{\\partial \\theta} \\right], \\\\ \\end{array}", + "image_path": "023d361d96ab619acdec9dda92c2c0cafe421dfdbc31ca2374d4e1f347fe2838.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": " is a constant and " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": " is the time-dependent weight that adjusts the contribution of each timestep, aligning the student's outputs with the diffusion model's predictions. Through alternating updates of the student and LoRA teacher, VSD enables efficient, high-fidelity generation across both text-to-image and text-to-3D tasks." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18006" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 71, + 286, + 116 + ], + "blocks": [ + { + "bbox": [ + 65, + 71, + 286, + 116 + ], + "lines": [ + { + "bbox": [ + 65, + 71, + 286, + 116 + ], + "spans": [ + { + "bbox": [ + 65, + 71, + 286, + 116 + ], + "type": "table", + "html": "
NPAnimePhotoCAPaintingsAverage
29.6229.1728.7928.6929.07
✓31.6629.1730.5830.3830.42
", + "image_path": "45e97e4177142f5a39114c7ddda68d95cd238911bc11629dff624be765090099.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 295, + 177 + ], + "lines": [ + { + "bbox": [ + 55, + 121, + 295, + 177 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 295, + 177 + ], + "type": "text", + "content": "Table 1. The effect of negative prompts (NP) on the HPSv2 score of PixArt-" + }, + { + "bbox": [ + 55, + 121, + 295, + 177 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 121, + 295, + 177 + ], + "type": "text", + "content": " [9]. Using generic negative prompts to remove visual artifacts (e.g., \"worst quality\", \"blurry\", \"bad anatomy\") leads to a significant improvement. Higher scores indicate better alignment with human preferences." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 57, + 184, + 294, + 291 + ], + "blocks": [ + { + "bbox": [ + 57, + 184, + 294, + 291 + ], + "lines": [ + { + "bbox": [ + 57, + 184, + 294, + 291 + ], + "spans": [ + { + "bbox": [ + 57, + 184, + 294, + 291 + ], + "type": "image", + "image_path": "f591d916009fa213f3d60703514f68ce8f4417e072ea7a5bc79ffb06a3ea11ff.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 297, + 295, + 331 + ], + "lines": [ + { + "bbox": [ + 55, + 297, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 295, + 331 + ], + "type": "text", + "content": "Figure 2. The inherent single-pass nature of one-step diffusion models renders them incompatible with direct CFG, resulting in artifact-laden outputs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 338, + 138, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 138, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 138, + 350 + ], + "type": "text", + "content": "4. Observations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 358, + 130, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 130, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 130, + 370 + ], + "type": "text", + "content": "4.1. Motivation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 376, + 295, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 376, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 54, + 376, + 295, + 520 + ], + "type": "text", + "content": "Positive prompt is not enough. While vision generative models like diffusion models excel at incorporating positive attributes, they fundamentally struggle with understanding negation. Consider requesting \"A portrait of a friendly monster, not scary\" (Fig. 1), the model typically fails to exclude the specified negative attribute \"scary\". Negative prompting addresses this limitation by creating a separate, dedicated input for exclusions – completely removing negation terms from the positive prompt and processing them independently. This clear separation allows the model to accurately distinguish what to avoid generating, providing significantly better control over the final output." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 521, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 521, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 54, + 521, + 295, + 664 + ], + "type": "text", + "content": "Community Need. The growing use of negative prompting in the generative AI community further highlights its importance. Platforms like Civit AI [10] and various open-source repositories [2, 11, 29] have shown that using both types of prompts together results in more refined, higher-quality model outputs that align more closely with user specifications. Users can clearly specify both desired and unwanted attributes in a prompt pair, which is particularly useful when working with models that generate complex or highly detailed content. Experimentally, Table 1 and Figure 1 confirm that using negative prompts effectively enhances image quality of the model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Critical Gap. Despite the established need and widespread adoption of negative prompts in multi-step diffusion models and the rapidly growing prominence of one-step approaches, a significant limitation persists: no existing one" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "step methods effectively support negative prompting. Our proposed method fills this gap by providing a simple, efficient, and training-free mechanism that integrates negative prompt into one-step diffusion models to (1) improve controllability and (2) enhance the quality of the output image." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 138, + 388, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 138, + 388, + 152 + ], + "spans": [ + { + "bbox": [ + 313, + 138, + 388, + 152 + ], + "type": "text", + "content": "4.2. Challenges" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 156, + 555, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 555, + 324 + ], + "type": "text", + "content": "A widely used method for implementing negative prompting is the CFG mechanism. However, applying CFG directly to one-step diffusion models presents significant challenges. CFG was designed for multi-step models, where the output is iteratively refined, allowing for gradual adjustments using both negative and unconditional outputs. In contrast, one-step models generate images in a single pass without the iterative denoising process, making them inherently unsuitable for this method. Direct application of CFG to these models leads to undesirable image blending, resulting in unnatural artifacts, as shown in Fig. 2. This highlights a crucial limitation: the CFG mechanism, though effective in multi-step models, is not directly translatable to one-step models without significant modifications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 334, + 420, + 347 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 420, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 420, + 347 + ], + "type": "text", + "content": "5. Proposed Method" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 354, + 554, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 554, + 415 + ], + "type": "text", + "content": "Our proposed methodology is detailed in Sec. 5.1 for the core mechanism. We present two primary implementation strategies for NASA: a training-free approach (NASA-I) and a distillation-based approach (NASA-T), which are presented in Sec. 5.2 and Sec. 5.3, respectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 421, + 515, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 515, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 515, + 433 + ], + "type": "text", + "content": "5.1. NASA: Negative-Away Steer Attention" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 437, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 498 + ], + "type": "text", + "content": "In contrast to CFG implementations that operate within the output space (synthesized image), as commonly practiced in multi-step diffusion models, we propose a novel approach that strategically relocates the guidance mechanism to the intermediate representation space of the diffusion model." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 498, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 605 + ], + "type": "text", + "content": "Specifically, cross-attention layers in these models capture semantic connections between image and text features [17], making them ideal for controlling the feature alignment. We hypothesize that cross-attention layer features are particularly well-suited for this approach. We selectively manipulate these cross-attention maps to attenuate attention responses associated with negative prompts relative to those elicited by positive prompts, thereby filtering out unwanted semantic attributes." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 606, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 554, + 677 + ], + "type": "text", + "content": "This representation-space guidance strategy offers a crucial advantage: it enables the exclusion of undesirable elements prior to the final image synthesis stage. This removal mechanism significantly mitigates the risk of generating blending artifacts, which are common byproducts of image-space manipulation techniques." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Our representation-space guidance also extends to architectures like FLUX [25], which replace traditional cross-attention with joint self-attention blocks. [46] show that" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18007" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 68, + 547, + 242 + ], + "blocks": [ + { + "bbox": [ + 63, + 68, + 547, + 242 + ], + "lines": [ + { + "bbox": [ + 63, + 68, + 547, + 242 + ], + "spans": [ + { + "bbox": [ + 63, + 68, + 547, + 242 + ], + "type": "image", + "image_path": "c1efb25a93a4ba5d98b6406d9914dc0106f9d8ba62cde33b93f4071faefac6a7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 250, + 555, + 318 + ], + "lines": [ + { + "bbox": [ + 54, + 250, + 555, + 318 + ], + "spans": [ + { + "bbox": [ + 54, + 250, + 555, + 318 + ], + "type": "text", + "content": "Figure 3. Left: An overview of the Negative-Away Steer Attention (NASA) pipeline. Positive (green) and negative (red) prompts are fed into a text encoder to generate positive and negative text features. The NASA module then processes these features, which adjusts the one-step diffusion model to steer the output image away from the negative features, refining it based on the positive features. Right: The details of the NASA module. It processes queries " + }, + { + "bbox": [ + 54, + 250, + 555, + 318 + ], + "type": "inline_equation", + "content": "(\\mathbf{Q}_l)" + }, + { + "bbox": [ + 54, + 250, + 555, + 318 + ], + "type": "text", + "content": " in layer " + }, + { + "bbox": [ + 54, + 250, + 555, + 318 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 54, + 250, + 555, + 318 + ], + "type": "text", + "content": ", note we will omit the subscript " + }, + { + "bbox": [ + 54, + 250, + 555, + 318 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 54, + 250, + 555, + 318 + ], + "type": "text", + "content": " in subsequent notations to improve readability, with positive " + }, + { + "bbox": [ + 54, + 250, + 555, + 318 + ], + "type": "inline_equation", + "content": "(\\mathbf{V}^{+},\\mathbf{K}^{+})" + }, + { + "bbox": [ + 54, + 250, + 555, + 318 + ], + "type": "text", + "content": " and negative " + }, + { + "bbox": [ + 54, + 250, + 555, + 318 + ], + "type": "inline_equation", + "content": "(\\mathbf{V}^{-},\\mathbf{K}^{-})" + }, + { + "bbox": [ + 54, + 250, + 555, + 318 + ], + "type": "text", + "content": " key-value pairs to create positive " + }, + { + "bbox": [ + 54, + 250, + 555, + 318 + ], + "type": "inline_equation", + "content": "(\\mathbf{Z}^{+})" + }, + { + "bbox": [ + 54, + 250, + 555, + 318 + ], + "type": "text", + "content": " and negative " + }, + { + "bbox": [ + 54, + 250, + 555, + 318 + ], + "type": "inline_equation", + "content": "(\\mathbf{Z}^{-})" + }, + { + "bbox": [ + 54, + 250, + 555, + 318 + ], + "type": "text", + "content": " attention outputs. The final output " + }, + { + "bbox": [ + 54, + 250, + 555, + 318 + ], + "type": "inline_equation", + "content": "(\\mathbf{Z}^{\\mathrm{NASA}})" + }, + { + "bbox": [ + 54, + 250, + 555, + 318 + ], + "type": "text", + "content": " is calculated by subtracting the weighted negative features " + }, + { + "bbox": [ + 54, + 250, + 555, + 318 + ], + "type": "inline_equation", + "content": "(\\mathbf{Z}^{-})" + }, + { + "bbox": [ + 54, + 250, + 555, + 318 + ], + "type": "text", + "content": " from the positive features " + }, + { + "bbox": [ + 54, + 250, + 555, + 318 + ], + "type": "inline_equation", + "content": "(\\mathbf{Z}^{+})" + }, + { + "bbox": [ + 54, + 250, + 555, + 318 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 323, + 304, + 396 + ], + "blocks": [ + { + "bbox": [ + 56, + 323, + 304, + 396 + ], + "lines": [ + { + "bbox": [ + 56, + 323, + 304, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 323, + 304, + 396 + ], + "type": "table", + "html": "
MethodFLUX.1-schnellSDXL-LCMSDXL-DMD2
4 steps1 step4 steps1 step4 steps1 step
None23%44%43%-27%25%
CFG30%0%14%-25%0%
NASA-I100%99%97%-100%100%
", + "image_path": "ac933e6e1c62f6066b1b984ad875dff2251cb27e44c23ac827d08f6165e9170c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 399, + 296, + 433 + ], + "lines": [ + { + "bbox": [ + 55, + 399, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 296, + 433 + ], + "type": "text", + "content": "Table 2. Comparison of success rate of unwanted feature removal in generated images. Note that SDXL-LCM [31] does not support one-step inference." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 436, + 295, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 295, + 472 + ], + "type": "text", + "content": "these blocks still embed distinct image-text cross-attention operations. This allows NASA to be seamlessly applied, proving its robustness across different backbones." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 472, + 296, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 604 + ], + "type": "text", + "content": "The overall pipeline of the NASA mechanism is illustrated in Fig. 3. Given the latent feature representation at a given layer " + }, + { + "bbox": [ + 55, + 472, + 296, + 604 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 472, + 296, + 604 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 55, + 472, + 296, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_l" + }, + { + "bbox": [ + 55, + 472, + 296, + 604 + ], + "type": "text", + "content": ", and positive text features " + }, + { + "bbox": [ + 55, + 472, + 296, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_p" + }, + { + "bbox": [ + 55, + 472, + 296, + 604 + ], + "type": "text", + "content": " extracted from CLIP text encoder, we define the query, key, and value matrices as " + }, + { + "bbox": [ + 55, + 472, + 296, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_l = \\mathbf{Z}_l\\mathbf{W}_{q,l}" + }, + { + "bbox": [ + 55, + 472, + 296, + 604 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 472, + 296, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_l^+ = \\mathbf{c}_p\\mathbf{W}_{k,l}" + }, + { + "bbox": [ + 55, + 472, + 296, + 604 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 472, + 296, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_l^+ = \\mathbf{c}_p\\mathbf{W}_{v,l}" + }, + { + "bbox": [ + 55, + 472, + 296, + 604 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 472, + 296, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{q,l}" + }, + { + "bbox": [ + 55, + 472, + 296, + 604 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 472, + 296, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{k,l}" + }, + { + "bbox": [ + 55, + 472, + 296, + 604 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 472, + 296, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{v,l}" + }, + { + "bbox": [ + 55, + 472, + 296, + 604 + ], + "type": "text", + "content": " are projection matrices at current layer " + }, + { + "bbox": [ + 55, + 472, + 296, + 604 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 472, + 296, + 604 + ], + "type": "text", + "content": ". Analogously, for negative text features " + }, + { + "bbox": [ + 55, + 472, + 296, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_n" + }, + { + "bbox": [ + 55, + 472, + 296, + 604 + ], + "type": "text", + "content": ", we use the same projection matrices to define the corresponding key and value matrices for negative prompt conditioning: " + }, + { + "bbox": [ + 55, + 472, + 296, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_l^- = \\mathbf{c}_n\\mathbf{W}_{k,l}" + }, + { + "bbox": [ + 55, + 472, + 296, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 472, + 296, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_l^- = \\mathbf{c}_n\\mathbf{W}_{v,l}" + }, + { + "bbox": [ + 55, + 472, + 296, + 604 + ], + "type": "text", + "content": ". The NASA mechanism is formulated as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 102, + 609, + 295, + 638 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 609, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 102, + 609, + 295, + 638 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} _ {l} ^ {+} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} _ {l} \\left(\\mathbf {K} _ {l} ^ {+}\\right) ^ {\\top}}{\\sqrt {d}}\\right) \\mathbf {V} _ {l} ^ {+}, \\tag {7}", + "image_path": "3953b38d20117ccf806821711c7ccc62e10e7d98dda893e0f690df1e4beeee72.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 638, + 295, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 638, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 102, + 638, + 295, + 666 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} _ {l} ^ {-} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} _ {l} \\left(\\mathbf {K} _ {l} ^ {-}\\right) ^ {\\top}}{\\sqrt {d}}\\right) \\mathbf {V} _ {l} ^ {-}, \\tag {8}", + "image_path": "d9d1d7f65443626be8ede37cd09b8e03da0755ed052d9ca79a350f76619757b7.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 666, + 295, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 666, + 295, + 682 + ], + "spans": [ + { + "bbox": [ + 116, + 666, + 295, + 682 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} _ {l} ^ {\\mathrm {N A S A}} = \\mathbf {Z} _ {l} ^ {+} - \\alpha \\cdot \\mathbf {Z} _ {l} ^ {-}, \\tag {9}", + "image_path": "4904c145a6d89b5dbff9d9f7d20bcaf1bfaac5f08d98555168cfde0532bf641b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "type": "text", + "content": " is the scale factor to control the degree of negative feature removal." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 325, + 481, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 481, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 481, + 338 + ], + "type": "text", + "content": "5.2. NASA-I: Training-free Setting" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 348, + 555, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 555, + 553 + ], + "type": "text", + "content": "Applicable to One-/Few-Step Models. To assess the effectiveness of NASA-I, we carry out experiments with pretrained one-step and few-step models, such as LCM [31], DMD2 [51] and FLUX.1-schnell [25]. We design a small set of positive and negative prompts and generate 100 images per model, both with and without negative prompts. We then calculate the percentage of images that successfully exclude the feature specified by the negative prompt. For example, given the positive prompt \"A photo of a person\" and the negative prompt \"male\", we measure the percentage of generated images that do not depict a male person. The average results across all prompt sets are summarized in Tab. 2. Our findings show that while traditional CFG is ineffective for one-step and few-step models, NASA-I provides precise control over unwanted features in generated images. Additionally, this capability is also qualitatively demonstrated in Fig. 1, Fig. 4 and Fig. 5." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "content": "Efficiency Considerations. As described in Sec. 5.1, we reuse the projection matrices of the positive prompt for the negative prompt, ensuring that no additional model parameters are introduced and keeping the model size unchanged. Furthermore, applying NASA-I to all one-step and few-step models introduces only a minor computational overhead, increasing FLOPs by just " + }, + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "inline_equation", + "content": "1.89\\%" + }, + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "content": ". This represents a significant efficiency improvement over CFG, which doubles the computational cost due to the need for separate forward passes for positive and negative prompts. Consequently, NASA-I enables negative prompt guidance in one-step diffusion models without introducing substantial latency, making it a practical solution for real-time applications." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18008" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 214, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 214, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 214, + 85 + ], + "type": "text", + "content": "5.3. NASA-T: Distillation Setting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 297, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 297, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 297, + 365 + ], + "type": "text", + "content": "Given their effectiveness in improving image quality during inference, their potential utility in model distillation is worth exploring. In this section, we introduce a method that integrates the negative prompt into one-step diffusion distillation. While our approach is applicable to various distillation methods, we select SwiftBrush (SB) [12, 34], an image-free method, as the baseline for improvement due to its efficiency. As described in Eq. (6), SB updates the student model using the gradient of the VSD loss. To incorporate the negative prompt " + }, + { + "bbox": [ + 55, + 89, + 297, + 365 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^{-}" + }, + { + "bbox": [ + 55, + 89, + 297, + 365 + ], + "type": "text", + "content": ", we modify the existing formulation by replacing " + }, + { + "bbox": [ + 55, + 89, + 297, + 365 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\psi}(\\hat{\\mathbf{x}}_t,t,\\mathbf{y},\\kappa)" + }, + { + "bbox": [ + 55, + 89, + 297, + 365 + ], + "type": "text", + "content": " (see Eq. (4)) with " + }, + { + "bbox": [ + 55, + 89, + 297, + 365 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\psi}(\\hat{\\mathbf{x}}_t,t,\\mathbf{y},\\mathbf{y}^{-},\\kappa)" + }, + { + "bbox": [ + 55, + 89, + 297, + 365 + ], + "type": "text", + "content": " (see Eq. (5)). In this way, the teacher model provides explicit guidance to the student model, helping it learn to steer away from undesirable features. Moreover, we incorporate the NASA mechanism into the student model, enabling it to process negative prompts during training. This allows the student model to better recognize and suppress unnatural artifacts, leading to improved learning and enhanced image quality. Note that with the LoRA teacher model " + }, + { + "bbox": [ + 55, + 89, + 297, + 365 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\pi}" + }, + { + "bbox": [ + 55, + 89, + 297, + 365 + ], + "type": "text", + "content": ", we can choose whether to incorporate negative prompts. Empirically, we found that not using them yields better results. Detailed results and discussions are provided in Sec. 6.3." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 371, + 137, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 137, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 137, + 384 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 391, + 171, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 171, + 404 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 171, + 404 + ], + "type": "text", + "content": "6.1. Experimental Setup" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 413, + 296, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 296, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 296, + 617 + ], + "type": "text", + "content": "Metrics and Benchmark. In diffusion model evaluation, Fréchet Inception Distance (FID) [18] on the zero-shot MS-COCO2014 [27] benchmark has been the standard metric. However, recent studies [8, 38, 51] suggest that FID in this benchmark does not always correlate with actual image quality and adopt human-focused alignment metrics like Human Preference Score v2 (HPSv2) [48] for benchmarking. Our experiments also follow this conclusion, so we choose HPSv2 as the primary metric for evaluating the models. Moreover, due to the absence of negative prompts in the MS-COCO2014 benchmark, we utilize the NegOpt dataset from [35], which contains pairs of positive and negative prompts. From this dataset, we randomly sample 30K prompt pairs to construct a subset for evaluation. Text alignment is then assessed using CLIP scores, where " + }, + { + "bbox": [ + 55, + 413, + 296, + 617 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP^{+}}" + }, + { + "bbox": [ + 55, + 413, + 296, + 617 + ], + "type": "text", + "content": " evaluates alignment with the positive prompt and " + }, + { + "bbox": [ + 55, + 413, + 296, + 617 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP^{-}}" + }, + { + "bbox": [ + 55, + 413, + 296, + 617 + ], + "type": "text", + "content": " evaluates alignment with the negative prompt." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 621, + 296, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 296, + 695 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 296, + 695 + ], + "type": "text", + "content": "Training Datasets. Following SBv2 [12], we use total of 3.3M prompts from JourneyDB [44] and a subset of LAION " + }, + { + "bbox": [ + 55, + 621, + 296, + 695 + ], + "type": "inline_equation", + "content": "[42]^1" + }, + { + "bbox": [ + 55, + 621, + 296, + 695 + ], + "type": "text", + "content": ". For negative prompts in the distillation setting (Sec. 5.3), we construct a dataset of 10K negative prompts derived from commonly used terms like \"bad photo, duplicate, low resolution, low quality\"." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": "Implementation Details. We base our method on SBv2 [12] with our proposed modifications in Sec. 5.3. We utilize three backbones SDv1.5 [39], SDv2.1 [39] and PixArt- " + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": " [9] as the frozen teachers and LoRA teachers are initialized with rank " + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "inline_equation", + "content": "r = 64" + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": " and scaling " + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\gamma = 128" + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": ". The learning rates are set to " + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "inline_equation", + "content": "1e^{-6}" + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": " for the student model and " + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "inline_equation", + "content": "1e^{-3}" + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": " for the LoRA teacher, using the AdamW optimizer [30]. All our training is conducted on four NVIDIA A100 40GB GPUs with the total batch size of 64 for the SD models and 32 for PixArt-" + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": ". Moreover, [6] found that instead of using a fixed CFG scale " + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": " during training, sampling " + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": " from a uniform distribution within a given range improves training robustness. Based on this insight, we set the teacher models' CFG scales as follows: " + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\kappa \\sim \\mathcal{U}(0.5,4)" + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": " for SDv1.5 and SDv2.1, and " + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\kappa \\sim \\mathcal{U}(0.5,3)" + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": " for PixArt-" + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": ". For NASA-T, in each iteration, we uniformly sample the scale " + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\alpha \\sim \\mathcal{U}(0,1)" + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": " to help the student model adapt to different conditions and enhance generalization. For NASA-I, unless explicitly specified, we set " + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\alpha = 0.1" + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": " for SDv1.5-based models, " + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\alpha = 0.2" + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": " for SDv2.1-based models, and " + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\alpha = 0.5" + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": " for PixArt-" + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": "-based models. Further details can be found in the Appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 329, + 399, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 399, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 399, + 341 + ], + "type": "text", + "content": "6.2. Main Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 350, + 556, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 556, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 556, + 613 + ], + "type": "text", + "content": "Quantitative Results. Table 3 presents a comparison between our approach and prior distilled text-to-image diffusion models. Using our reimplementation of SBv2 [12] as the baseline, NASA-T consistently enhances image quality across all tested backbones. Compared to other methods, our model achieves the highest HPSv2 scores on average, demonstrating its clear superiority in image generation quality. This suggests that integrating complex loss functions like DMD2 [51] may not always be necessary for achieving state-of-the-art results. Moreover, while applying CFG directly degrades one-step model outputs, NASA-I further enhances image quality across all models when using negative prompts like \"worst quality, low quality, ugly, duplicate, out of frame, deformed, blurry, bad anatomy, watermark\". Notably, our method on the PixArt-" + }, + { + "bbox": [ + 313, + 350, + 556, + 613 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 350, + 556, + 613 + ], + "type": "text", + "content": "-based backbone achieves a record-breaking HPSv2 score of 31.21, surpassing all other one-step diffusion models baselines. Regarding the NegOpt dataset, with the support of NASA-I, all models demonstrate a reduction in the " + }, + { + "bbox": [ + 313, + 350, + 556, + 613 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^{-}" + }, + { + "bbox": [ + 313, + 350, + 556, + 613 + ], + "type": "text", + "content": " score while maintaining the " + }, + { + "bbox": [ + 313, + 350, + 556, + 613 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^{+}" + }, + { + "bbox": [ + 313, + 350, + 556, + 613 + ], + "type": "text", + "content": " score. This indicates our method effectively eliminates unwanted features in generated outputs while preserving desired traits." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "type": "text", + "content": "Qualitative Results. As discussed in Sec. 5.2, NASA-I is a versatile, training-free method for both one-step and few-step diffusion models. Figure 4 and Figure 5 present qualitative results of integrating NASA-I into the SDXL-DMD2 [51] and FLUX.1-schnell [25] models, respectively. When using a conditional negative prompt, the generated images retain key semantic features while effectively suppressing unwanted attributes specified by the negative prompt." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 259, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 259, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 259, + 712 + ], + "type": "text", + "content": "1All datasets were downloaded and evaluated at MovianAI." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18009" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 107, + 70, + 504, + 419 + ], + "blocks": [ + { + "bbox": [ + 107, + 70, + 504, + 419 + ], + "lines": [ + { + "bbox": [ + 107, + 70, + 504, + 419 + ], + "spans": [ + { + "bbox": [ + 107, + 70, + 504, + 419 + ], + "type": "table", + "html": "
DatasetNegOptHPSv2
MethodCLIP+↑CLIP-↓Anime ↑Photo ↑CA ↑Paintings ↑Average ↑
PixArt-α-based backbone
PixArt-α [9] [Teacher]0.350.0529.6229.1728.7928.6929.07
YOSO [32]0.360.0828.7528.0628.5228.5728.48
+ NASA-I0.360.0628.7428.0528.5628.6028.49 (+0.01)
DMD [52]0.350.0829.3128.6728.4628.4128.71
+ CFG = 1.50.340.0930.0227.0728.3628.0728.38 (-0.33)
+ CFG = 2.50.310.1326.7423.8625.1324.6625.10 (-3.61)
+ NASA-I0.350.0529.3328.7128.4928.5328.77 (+0.06)
SBv2*0.360.0932.1929.0930.3929.6930.34
+ NASA-I0.360.0632.6029.5831.0930.6530.98 (+0.64)
+ NASA-T0.350.0832.3329.2630.7530.1030.61 (+0.27)
+ NASA-T + CFG = 1.50.340.1029.4726.5028.2227.6827.97 (-2.37)
+ NASA-T + NASA-I0.350.0632.6629.6731.3130.7131.09 (+0.75)
+ NASA-T + NASA-I (α = 1)0.350.0532.6529.6531.4531.0631.21 (+0.87)
Stable Diffusion 1.5-based backbone
SDv1.5 [39] [Teacher]0.310.0626.5127.1926.0626.1226.47
InstaFlow-0.9B [28]0.330.1126.1026.6225.9225.9526.15
+ NASA-I0.330.1026.1526.6825.9325.9826.19 (+0.04)
DMD2 [51]0.310.1026.3927.0025.8025.8226.25
+ NASA-I0.310.0926.3927.0225.8025.8326.26 (+0.01)
SBv2*0.320.0927.1827.5826.6926.6227.02
+ NASA-I0.320.0827.1927.5926.7126.6327.03 (+0.01)
+ NASA-T0.350.1027.6427.9427.1927.0327.45 (+0.43)
+ NASA-T + NASA-I0.350.0927.6527.9727.1827.0427.46 (+0.44)
Stable Diffusion 2.1-based backbone
SDv2.1 [39] [Teacher]0.330.0627.4826.8926.8627.4627.17
SBv2 [12]0.340.1027.2527.6226.8626.7727.13
+ NASA-I0.350.0827.4427.8426.9127.0227.30 (+0.17)
SBv2*0.370.1027.5627.8426.9727.0327.35
+ NASA-I0.370.0827.7028.0027.1227.2227.51 (+0.16)
+ NASA-T0.360.0928.0028.6527.4427.2627.84 (+0.49)
+ NASA-T + NASA-I0.360.0728.0428.6827.5127.5027.93 (+0.58)
", + "image_path": "94bc3fc383891011e919360e3f806cbdc17999ac91af90247b146ef7508c4004.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 426, + 555, + 460 + ], + "lines": [ + { + "bbox": [ + 55, + 426, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 555, + 460 + ], + "type": "text", + "content": "Table 3. Quantitative comparisons between our method and previous works. NASA-I refers to our training-free method introduced in Sec. 5.2, while NASA-T denotes our distillation-based approach described in Sec. 5.3. SBv2* refers to our reimplementation of SBv2 with a randomly sampled CFG scale " + }, + { + "bbox": [ + 55, + 426, + 555, + 460 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 55, + 426, + 555, + 460 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 93, + 468, + 518, + 670 + ], + "blocks": [ + { + "bbox": [ + 93, + 468, + 518, + 670 + ], + "lines": [ + { + "bbox": [ + 93, + 468, + 518, + 670 + ], + "spans": [ + { + "bbox": [ + 93, + 468, + 518, + 670 + ], + "type": "image", + "image_path": "cd87ca5eb6940b8ea62f3dbc8670a0a27ee17e081844b9f092029ba30e85a70c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 677, + 555, + 700 + ], + "lines": [ + { + "bbox": [ + 55, + 677, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 555, + 700 + ], + "type": "text", + "content": "Figure 4. Effect of different scale values (0.0 to 1.0) in NASA-I with SDXL-DMD2 [51] for both one-step and few-step settings, illustrating the progressive influence on visual details and composition." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18010" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 553, + 171 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 553, + 171 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 553, + 171 + ], + "type": "image", + "image_path": "393710f27e24d69dcb9494f23cf0cd7e5d701b4345160768a0599150ef649410.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 176, + 423, + 188 + ], + "lines": [ + { + "bbox": [ + 187, + 176, + 423, + 188 + ], + "spans": [ + { + "bbox": [ + 187, + 176, + 423, + 188 + ], + "type": "text", + "content": "Figure 5. Qualitative results of NASA-I in FLUX.1-schnell [25]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 193, + 239, + 261 + ], + "blocks": [ + { + "bbox": [ + 57, + 193, + 239, + 261 + ], + "lines": [ + { + "bbox": [ + 57, + 193, + 239, + 261 + ], + "spans": [ + { + "bbox": [ + 57, + 193, + 239, + 261 + ], + "type": "image", + "image_path": "ce7533494e5f6c15882df42521eeabdbdd5c06b5836bb9f0fe2cdaaf4654f19b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 268, + 392, + 281 + ], + "lines": [ + { + "bbox": [ + 83, + 268, + 392, + 281 + ], + "spans": [ + { + "bbox": [ + 83, + 268, + 392, + 281 + ], + "type": "text", + "content": "Figure 6. Qualitative results of NASA-I for text-to-video generation in CausVid [53]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 240, + 193, + 421, + 261 + ], + "blocks": [ + { + "bbox": [ + 240, + 193, + 421, + 261 + ], + "lines": [ + { + "bbox": [ + 240, + 193, + 421, + 261 + ], + "spans": [ + { + "bbox": [ + 240, + 193, + 421, + 261 + ], + "type": "image", + "image_path": "5933f6c822b2d2568a91d31bf760a8312a61062190b15958f6ff4462019d29ac.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 426, + 204, + 551, + 260 + ], + "blocks": [ + { + "bbox": [ + 426, + 204, + 551, + 260 + ], + "lines": [ + { + "bbox": [ + 426, + 204, + 551, + 260 + ], + "spans": [ + { + "bbox": [ + 426, + 204, + 551, + 260 + ], + "type": "table", + "html": "
MethodVBenchmark-Long benchmark
Aesthetic Quality ↑Imaging Quality ↑
None61.9867.12
NASA-I63.3367.36
", + "image_path": "0c9cdb25b630e8198754b173f1ef9249f3e01efe444a1205264f6b7033e9f4b0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 58, + 293, + 291, + 350 + ], + "blocks": [ + { + "bbox": [ + 426, + 269, + 549, + 280 + ], + "lines": [ + { + "bbox": [ + 426, + 269, + 549, + 280 + ], + "spans": [ + { + "bbox": [ + 426, + 269, + 549, + 280 + ], + "type": "text", + "content": "Table 4. CausVid quality metrics." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 293, + 291, + 350 + ], + "lines": [ + { + "bbox": [ + 58, + 293, + 291, + 350 + ], + "spans": [ + { + "bbox": [ + 58, + 293, + 291, + 350 + ], + "type": "table", + "html": "
NASA✓✓
Teacher✓✓✓✓
LoRA Teacher✓✓
HPSv2 Avg. ↑30.3429.8429.8030.6129.89
", + "image_path": "1fd9a75d397df439a764e6ed6d639b67c7f76330d4f364d8783adc4c26b44dc7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 356, + 295, + 378 + ], + "lines": [ + { + "bbox": [ + 55, + 356, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 295, + 378 + ], + "type": "text", + "content": "Table 5. Ablation studies on the integration of negative prompts in model distillation using the PixArt- " + }, + { + "bbox": [ + 55, + 356, + 295, + 378 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 356, + 295, + 378 + ], + "type": "text", + "content": " backbone." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 65, + 389, + 284, + 521 + ], + "blocks": [ + { + "bbox": [ + 65, + 389, + 284, + 521 + ], + "lines": [ + { + "bbox": [ + 65, + 389, + 284, + 521 + ], + "spans": [ + { + "bbox": [ + 65, + 389, + 284, + 521 + ], + "type": "image", + "image_path": "57fae8adf7a27b7fb899130db0f95ddf7097e631a9594cdfe0137c0a5709a7c5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 526, + 295, + 548 + ], + "lines": [ + { + "bbox": [ + 55, + 526, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 295, + 548 + ], + "type": "text", + "content": "Figure 7. Ablation study on the NASA scale using our distilled model based on the PixArt-" + }, + { + "bbox": [ + 55, + 526, + 295, + 548 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 526, + 295, + 548 + ], + "type": "text", + "content": " backbone." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": "Application to Text-to-Video Models. To demonstrate the versatility of our approach, we extended NASA-I to a few-step video diffusion model, CausVid [53]. As shown in Fig. 6, our method effectively suppresses unwanted visual attributes while preserving key semantic content and temporal consistency. Quantitative evaluation on the VBenchmarkLong [22] benchmark further validates this. Table 4 shows that using NASA-I with common negative prompts like \"low quality, bad anatomy, blurry, distortion, ugly, low resolution, unclear details\" improves both the aesthetic and imaging quality scores of the generated videos. This highlights NASA-I's potential as a general-purpose tool for enhancing guided generation across different modalities." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 295, + 416, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 416, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 416, + 306 + ], + "type": "text", + "content": "6.3. Ablation Studies" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 312, + 555, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 312, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 312, + 312, + 555, + 468 + ], + "type": "text", + "content": "Table 5 presents an ablation on integrating negative prompts during model distillation with the PixArt-" + }, + { + "bbox": [ + 312, + 312, + 555, + 468 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 312, + 312, + 555, + 468 + ], + "type": "text", + "content": " [9] backbone. The baseline is our reimplementation of SBv2 with a randomly sampled CFG scale " + }, + { + "bbox": [ + 312, + 312, + 555, + 468 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 312, + 312, + 555, + 468 + ], + "type": "text", + "content": ". Simply applying negative prompts only in the teacher and/or the LoRA teacher models slightly degrades the HPSv2 score, as the student model is not explicitly trained on this negative conditioning. In contrast, incorporating NASA with negative prompts during training allows the student model to better capture and utilize this conditioning, leading to improved performance. Furthermore, we find that integrating negative prompts solely into the teacher model yields better results than applying them to both the teacher and LoRA teacher models." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 468, + 555, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 555, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 555, + 564 + ], + "type": "text", + "content": "Using our PixArt-" + }, + { + "bbox": [ + 313, + 468, + 555, + 564 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 468, + 555, + 564 + ], + "type": "text", + "content": "-based distilled model as the base, we varied the NASA scale " + }, + { + "bbox": [ + 313, + 468, + 555, + 564 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 468, + 555, + 564 + ], + "type": "text", + "content": " and evaluated the results with HPSv2 on the benchmark dataset. Figure 7 demonstrates that NASA-I generally enhances image synthesis, resulting in outputs that better align with human preferences. However, if the NASA scale " + }, + { + "bbox": [ + 313, + 468, + 555, + 564 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 468, + 555, + 564 + ], + "type": "text", + "content": " is set too high, the negative attributes may overly dominate the positive attributes, adversely affecting the final image quality." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 573, + 466, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 573, + 466, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 573, + 466, + 586 + ], + "type": "text", + "content": "7. Discussion and Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 592, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 592, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 592, + 555, + 715 + ], + "type": "text", + "content": "This paper introduces NASA, the first method to integrate negative prompt guidance into one-step and few-step diffusion models for both inference and distillation. By leveraging cross-attention adjustments, NASA effectively suppresses unwanted features in generated images. Experimental results show that NASA enhances image quality and outperforms strong baselines, establishing a new state-of-the-art in one-step model distillation based on human preference metrics. While NASA's effectiveness relies on selecting an appropriate scale for suppressing negative features, this parameter does not require extensive tuning." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18011" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 94, + 115, + 106 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 94, + 115, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 94, + 115, + 106 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 114, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 114, + 295, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 114, + 295, + 168 + ], + "spans": [ + { + "bbox": [ + 61, + 114, + 295, + 168 + ], + "type": "text", + "content": "[1] Mohammadreza Armandpour, Ali Sadeghian, Huangjie Zheng, Amir Sadeghian, and Mingyuan Zhou. Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. ArXiv, abs/2304.04968, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 170, + 294, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 170, + 294, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 170, + 294, + 190 + ], + "type": "text", + "content": "[2] AUTOMATIC1111. Automatic1111/stable-diffusion-webui, 2025. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 192, + 295, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 192, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 192, + 295, + 257 + ], + "type": "text", + "content": "[3] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "type": "text", + "content": "[4] Yuanhao Ban, Ruochen Wang, Tianyi Zhou, Minhao Cheng, Boqing Gong, and Cho-Jui Hsieh. Understanding the impact of negative prompts: When and how do they take effect? In European conference on computer vision, pages 190–206. Springer, 2024. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 314, + 294, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 314, + 294, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 314, + 294, + 378 + ], + "type": "text", + "content": "[5] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Yuanzhen Li, Tomer Michaeli, Oliver Wang, Deqing Sun, Tali Dekel, and Inbar Mosseri. Lumiere: A space-time diffusion model for video generation. ArXiv, abs/2401.12945, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 380, + 294, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 294, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 294, + 423 + ], + "type": "text", + "content": "[6] Clement Chadebec, Onur Tasar, Eyal Benaroche, and Benjamin Aubin. Flash diffusion: Accelerating any conditional diffusion model for few steps image generation. arXiv preprint arXiv:2406.02347, 2024. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 425, + 294, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 425, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 425, + 294, + 479 + ], + "type": "text", + "content": "[7] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James T. Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. ArXiv, abs/2310.00426, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 480, + 294, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 294, + 534 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 294, + 534 + ], + "type": "text", + "content": "[8] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- " + }, + { + "bbox": [ + 62, + 480, + 294, + 534 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 62, + 480, + 294, + 534 + ], + "type": "text", + "content": " : Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv: 2403.04692, 2024. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 536, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 294, + 601 + ], + "type": "text", + "content": "[9] Junsong Chen, Jincheng YU, Chongjian GE, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 4, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 602, + 254, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 254, + 613 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 254, + 613 + ], + "type": "text", + "content": "[10] Civitai Platform. https://civitai.com.4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 614, + 277, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 277, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 277, + 624 + ], + "type": "text", + "content": "[11] comfyanonymous. comfyanonymous/comfyui, 2025. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 625, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 679 + ], + "type": "text", + "content": "[12] Trung Dao, Thuan Hoang Nguyen, Thanh Le, Duc Vu, Khoi Nguyen, Cuong Pham, and Anh Tran. Swiftbrush v2: Make your one-step diffusion model better than its teacher. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "text", + "content": "[13] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 317, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 317, + 73, + 553, + 105 + ], + "type": "text", + "content": "[14] Zach Evans, CJ Carr, Josiah Taylor, Scott H. Hawley, and Jordi Pons. Fast timing-conditioned latent audio diffusion. ArXiv, abs/2402.04825, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 106, + 553, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 106, + 553, + 193 + ], + "spans": [ + { + "bbox": [ + 317, + 106, + 553, + 193 + ], + "type": "text", + "content": "[15] Peng Gao, Le Zhuo, Ziyi Lin, Chris Liu, Junsong Chen, Ruoyi Du, Enze Xie, Xu Luo, Longtian Qiu, Yuhang Zhang, Chen Lin, Rongjie Huang, Shijie Geng, Renrui Zhang, Junlin Xi, Wenqi Shao, Zhengkai Jiang, Tianshuo Yang, Weicai Ye, He Tong, Jingwen He, Yu Jiao Qiao, and Hongsheng Li. Lumina-t2x: Transforming text into any modality, resolution, and duration via flow-based large diffusion transformers. ArXiv, abs/2405.05945, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 194, + 553, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 194, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 317, + 194, + 553, + 238 + ], + "type": "text", + "content": "[16] Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 239, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 239, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 317, + 239, + 553, + 282 + ], + "type": "text", + "content": "[17] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control.(2022). In International Conference on Learning Representations, 2023. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 283, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 327 + ], + "type": "text", + "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Neural Information Processing Systems, 2017. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 327, + 553, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 349 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 349 + ], + "type": "text", + "content": "[19] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 1, 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "type": "text", + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. Advances in Neural Information Processing Systems, 33, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 383, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 553, + 436 + ], + "type": "text", + "content": "[21] Rongjie Huang, Jia-Bin Huang, Dongchao Yang, Yi Ren, Luping Liu, Mingze Li, Zhenhui Ye, Jinglin Liu, Xiaoyue Yin, and Zhou Zhao. Make-an-audio: Text-to-audio generation with prompt-enhanced diffusion models. ArXiv, abs/2301.12661, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 437, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 437, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 437, + 553, + 514 + ], + "type": "text", + "content": "[22] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 515, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 515, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 553, + 568 + ], + "type": "text", + "content": "[23] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 570, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 570, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 317, + 570, + 553, + 601 + ], + "type": "text", + "content": "[24] Oren Katzir, Or Patashnik, Daniel Cohen-Or, and Dani Lischinski. Noise-free score distillation. arXiv preprint arXiv: 2310.17590, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 602, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 624 + ], + "type": "text", + "content": "[25] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024. 4, 5, 6, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 625, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 553, + 690 + ], + "type": "text", + "content": "[26] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18012" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "text", + "content": "[28] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, and Qiang Liu. InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation. arXiv preprint arXiv:2309.06380, 2023. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 288, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 288, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 288, + 163 + ], + "type": "text", + "content": "[29] Illyasviel. Illyasviel/stable-diffusion-webui-forge, 2025. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 164, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 294, + 196 + ], + "type": "text", + "content": "[30] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "type": "text", + "content": "[31] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv: 2310.04378, 2023. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 243, + 294, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 286 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 286 + ], + "type": "text", + "content": "[32] Yihong Luo, Xiaolong Chen, Xinghua Qu, and Jing Tang. You only sample once: Taming one-step text-to-image synthesis by self-cooperative diffusion gans. arXiv preprint arXiv: 2403.12931, 2024. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 288, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 288, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 288, + 294, + 331 + ], + "type": "text", + "content": "[33] David McAllister, Songwei Ge, Jia-Bin Huang, David W. Jacobs, Alexei A. Efros, Aleksander Holynski, and Angjoo Kanazawa. Rethinking score distillation as a bridge between image distributions. ArXiv, abs/2406.09417, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 333, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 294, + 387 + ], + "type": "text", + "content": "[34] Thuan Hoang Nguyen and Anh Tran. Swiftbrush: One-step text-to-image diffusion model with variational score distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 388, + 294, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 422 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 422 + ], + "type": "text", + "content": "[35] Michael Ogezi and Ning Shi. Optimizing negative prompts for enhanced aesthetics and fidelity in text-to-image generation. arXiv preprint arXiv:2403.07605, 2024. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 423, + 294, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 423, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 423, + 294, + 477 + ], + "type": "text", + "content": "[36] Dustin Podell, Zion English, Kyle Lacey, A. Blattmann, Tim Dockhorn, Jonas Muller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. International Conference on Learning Representations, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 479, + 294, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 511 + ], + "type": "text", + "content": "[37] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D diffusion. *ArXiv*, abs/2209.14988, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 513, + 294, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 513, + 294, + 556 + ], + "spans": [ + { + "bbox": [ + 56, + 513, + 294, + 556 + ], + "type": "text", + "content": "[38] Yuxi Ren, Xin Xia, Yanzuo Lu, Jiacheng Zhang, Jie Wu, Pan Xie, Xing Wang, and Xuefeng Xiao. Hyper-sd: Trajectory segmented consistency model for efficient image synthesis. arXiv preprint arXiv: 2404.13686, 2024. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 558, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 558, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 558, + 295, + 612 + ], + "type": "text", + "content": "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 3, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 614, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 295, + 679 + ], + "type": "text", + "content": "[40] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 681, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 295, + 712 + ], + "type": "text", + "content": "[41] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "type": "text", + "content": "[42] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. LAION-5B: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 140, + 553, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 193 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 193 + ], + "type": "text", + "content": "[43] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 194, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 194, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 194, + 553, + 270 + ], + "type": "text", + "content": "[44] Keqiang Sun, Junting Pan, Yuying Ge, Hao Li, Haodong Duan, Xiaoshi Wu, Renrui Zhang, Aojun Zhou, Zipeng Qin, Yi Wang, Jifeng Dai, Yu Qiao, Limin Wang, and Hongsheng Li. JourneyDB: A benchmark for generative image understanding. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 272, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 326 + ], + "type": "text", + "content": "[45] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 327, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 369 + ], + "type": "text", + "content": "[46] Tianyi Wei, Dongdong Chen, Yifan Zhou, and Xingang Pan. Enhancing mmdit-based text-to-image models for similar subject generation. arXiv preprint arXiv:2411.18301, 2024. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 371, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 415 + ], + "type": "text", + "content": "[47] Max Woolf. Stable diffusion 2.0 and the importance of negative prompts for good results. https://minimaxir.com/2022/11/stable-diffusion-negative-prompt/, 2023.2,3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 415, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 459 + ], + "type": "text", + "content": "[48] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. ArXiv, abs/2306.09341, 2023. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "type": "text", + "content": "[49] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 514, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 553, + 581 + ], + "type": "text", + "content": "[50] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Xiaotao Gu, Yuxuan Zhang, Weihan Wang, Yean Cheng, Ting Liu, Bin Xu, Yuxiao Dong, and Jie Tang. Cogvideox: Text-to-video diffusion models with an expert transformer. ArXiv, abs/2408.06072, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "type": "text", + "content": "[51] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. In NeurIPS, 2024. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "type": "text", + "content": "[52] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In CVPR, 2024. 2, 3, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[53] Tianwei Yin, Qiang Zhang, Richard Zhang, William T Freeman, Fredo Durand, Eli Shechtman, and Xun Huang. From slow bidirectional to fast autoregressive video diffusion models. 2025. 8" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18013" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Supercharging Floorplan Localization with Semantic Rays/834ff1eb-78d1-464a-8d20-1e5cbed96504_content_list.json b/2025/Supercharging Floorplan Localization with Semantic Rays/834ff1eb-78d1-464a-8d20-1e5cbed96504_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ed3fadd0da3ca95d6e0869123fae1cc7d7c7fec3 --- /dev/null +++ b/2025/Supercharging Floorplan Localization with Semantic Rays/834ff1eb-78d1-464a-8d20-1e5cbed96504_content_list.json @@ -0,0 +1,1399 @@ +[ + { + "type": "text", + "text": "Supercharging Floorplan Localization with Semantic Rays", + "text_level": 1, + "bbox": [ + 202, + 130, + 794, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuval Grader \nTel Aviv University", + "bbox": [ + 300, + 181, + 459, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hadar Averbuch-Elor Cornell University", + "bbox": [ + 524, + 181, + 696, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://tau-vailab.github.io/SemRayLoc/", + "bbox": [ + 323, + 234, + 671, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 259, + 272, + 336, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Floorplans provide a compact representation of the building's structure, revealing not only layout information but also detailed semantics such as the locations of windows and doors. However, contemporary floorplan localization techniques mostly focus on matching depth-based structural cues, ignoring the rich semantics communicated within floorplans. In this work, we introduce a semantic-aware localization framework that jointly estimates depth and semantic rays, consolidating over both for predicting a structural-semantic probability volume. Our probability volume is constructed in a coarse-to-fine manner: We first sample a small set of rays to obtain an initial low-resolution probability volume. We then refine these probabilities by performing a denser sampling only in high-probability regions and process the refined values for predicting a 2D location and orientation angle. We conduct an evaluation on two standard floorplan localization benchmarks. Our experiments demonstrate that our approach substantially outperforms state-of-the-art methods, achieving significant improvements in recall metrics compared to prior works. Moreover, we show that our framework can easily incorporate additional metadata such as room labels, enabling additional gains in both accuracy and efficiency.", + "bbox": [ + 109, + 306, + 483, + 670 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 112, + 718, + 243, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Camera localization is a longstanding problem in computer vision, with significant applications in 3D reconstruction [19, 25-28], augmented reality [11, 16, 23, 29, 31], and navigation [20, 24, 30, 32, 36]. Localization within indoor environments is especially challenging due to the absence of reliable GPS signals and the complexity of reasoning across multiple floors and layers. Hence, to bypass complicated 3D model-based localization techniques, prior work [7, 8, 12, 13, 15, 35] has explored the problem of localizing camera observations within a provided 2D floorplan map by matching", + "bbox": [ + 109, + 744, + 482, + 912 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/77ef471a7a1601e41515129379e2e22703659419519e68e969f5e6eddefc7a41.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 270, + 714, + 342 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/858b3aaecf03cba21dc1abcc2b6f98c495992a3f6724f89654ed9f55cb160005.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 351, + 715, + 478 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6cd75802e283904d983688ac00f6389d1e1bddeecdf915c83ef41fca98fe8618.jpg", + "image_caption": [ + "Input & GT" + ], + "image_footnote": [], + "bbox": [ + 522, + 479, + 715, + 609 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/14fc60037bf8e390c93cfefa04e4bacfe3e63beec82af7bef376240fe45a3769.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 348, + 880, + 476 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1c8b5dad1f5d9f5c90b4e1ea7987df7e3438757c72b5d6306ba7ed55a817c077.jpg", + "image_caption": [ + "Output Localization", + "Figure 1. Floorplan localization using a raw binary floorplan (middle row) often yields ambiguous predictions. In this work, we utilize a richer, yet readily available, representation: a floorplan enhanced with semantic labels (bottom row). We present an approach that supercharges floorplan localization with semantic rays, enabling for resolving localization ambiguities, as illustrated by the comparison on the right." + ], + "image_footnote": [], + "bbox": [ + 728, + 481, + 880, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "depth-based structural cues.", + "bbox": [ + 511, + 772, + 699, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, while floorplans offer a compact and lightweight scene representation, structural cues within floorplans often correlate with multiple candidate locations, particularly for environments with repetitive or symmetric layouts. Consider the example in Figure 1. Can you localize the input image within the floorplan? Provided with just a raw (walls only) floorplan, room corners are indistinguishable and hence localization is", + "bbox": [ + 511, + 791, + 883, + 912 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27116", + "bbox": [ + 478, + 950, + 519, + 963 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "highly ambiguous, as can also be observed by the probabilities predicted by the state-of-the-art F3Loc [7] technique (middle row, right). To resolve such ambiguities, we are interested in utilizing a slightly different representation: a semantics-aware floorplan, such as the one illustrated on the bottom row of Figure 1.", + "bbox": [ + 109, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Accordingly, we introduce a semantic-aware ray-based localization framework that integrates semantic cues with depth-based predictions. In particular, we propose a semantic prediction network that predicts accurate semantic ray representations along with optional additional metadata (such as room labels) from a single RGB image with a limited field-of-view. We process these rays to compute a semantic probability volume, which is then fused with depth information for constructing a structural-semantic probability volume.", + "bbox": [ + 109, + 183, + 480, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our framework follows a coarse-to-fine localization strategy. We first operate over a low-resolution image for an efficient floorplan search. This initial search yields the Top- $k$ candidate locations. Finally, we refine these candidates and select the best match using a high-resolution ray representation. By adopting this coarse-to-fine approach, our method effectively constrains the localization search to the most promising regions while keeping the computation cost feasible.", + "bbox": [ + 109, + 335, + 480, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our experiments demonstrate that our approach yields improvements by factors ranging from two to three across most metrics compared to the state-of-the-art technique [7], upon which our method is built. This substantial gain underscores the effectiveness of fusing semantic and depth ray predictions into a unified probability volume. We further show that our coarse-to-fine strategy offers a flexible tradeoff between accuracy and computational cost, with performance consistently improving as larger candidate sets (Top- $k$ ) are evaluated—making our method adaptable to diverse task requirements. Moreover, incorporating additional metadata further enhances precision, leading to significantly improved localization accuracies.", + "bbox": [ + 109, + 473, + 482, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Explicitly stated, our contributions are:", + "bbox": [ + 130, + 685, + 392, + 699 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a semantic ray prediction network that receives a single RGB image as input.", + "- We propose an efficient and unified framework that fuses semantic and depth ray predictions into a structural-semantic probability volume, which effectively resolves localization ambiguities.", + "- Results that demonstrate that significant improvements over state-of-the-art methods." + ], + "bbox": [ + 112, + 702, + 482, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 112, + 839, + 254, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual Localization. The task of visual localization has received ongoing attention throughout the past several decades. Traditional approaches often rely on image re", + "bbox": [ + 109, + 866, + 482, + 912 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "trieval or on a 3D Structure-from-Motion (SfM) model of the environment. In the image retrieval paradigm, methods such as NetVLAD [1] or RelocNet [2] compare a query image against a database of labeled images. Once the closest match is found, the query pose is approximated by the retrieved image's pose. Other methods explicitly construct a 3D SfM model of a scene to establish 2D-3D correspondences between the query image and the reconstructed 3D structure [19, 25-28]. After matching local image descriptors to 3D points, robust solvers estimate the 6-DoF pose.", + "bbox": [ + 511, + 90, + 883, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent learning-based pipelines deviate from classical 2D-3D matching. Scene-coordinate regression methods predict dense 3D coordinates for every pixel in the query image [6, 29, 31], whereas pose regression methods directly estimate the 6-DoF camera pose via neural networks [16, 32, 36]. Although promising for single-scene scenarios, these methods must be retrained or fine-tuned to handle new environments.", + "bbox": [ + 511, + 257, + 883, + 377 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Floorplan Localization. Prior work addressing the task of floorplan localization primarily focused on depth-based cues, leveraging image-derived depth predictions or sensors, to match depth obtained from floorplans. In particular, LiDAR-based methods [3, 4, 18, 34] utilize precise laser scans but restrict usability on most mobile devices. Alternative sources of geometric cues, including semi dense visual odometry (SDVO) [8], or point clouds from depth cameras [14], can circumvent heavy LiDAR hardware.", + "bbox": [ + 511, + 383, + 883, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Earlier works compare extracted room edges directly to the 2D layout [5], often assuming knowledge of camera or room height [5, 8]. Other approaches, are embedding RGB images and floorplans into a shared metric space in order to do the localization. For instance, LaLaLoc [13] uses panoramic depth layout image which is rendered at known heights at different locations within the floorplan, where the localization is done by doing similarity in the embedded space. LaLaLoc++ [12] removes the hight assumption by embedding the entire floorplan into the feature space. However, these approaches often require an upright camera pose [12, 13], making them less flexible for hand-held or head-mounted devices. F3Loc [7] localizes by predicting depth rays from a given image and generating probability volumes that indicate the likelihood of each the depth prediction to a particular location on the floorplan. Although effectively leverages geometric cues from depth maps, it does not incorporate semantic information, which may reduce its robustness in environments with repetitive structures or occlusions.", + "bbox": [ + 511, + 534, + 883, + 849 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semantic-based cues are less commonly used for indoor localization, but several works have utilized them for this task. Wang et al.[33] extracts scene texts from images in large indoor spaces and performs localiza", + "bbox": [ + 511, + 851, + 880, + 912 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27117", + "bbox": [ + 478, + 950, + 517, + 962 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion by matching this text to the floorplan. SeDAR [21] uses a CNN to extract semantic labels from an image and perform Monte Carlo localization. By contrast to our single-image localization framework, it depends on sequential images and depth sensors. PF-Net [15] applies a differentiable particle filter with a learned observation model, relying on a computationally-intensive embedding process that yields limited performance for the single-image scenario. SPVLoc [10] matches captured images to panoramic renderings to estimate location, leveraging additional height information present in these renderings. Similarly, Kim et al. [17] perform localization by using a pre-captured 3D map of the environment. Our approach does not assume the availability of these additional cues. LASER [22] treats the floorplan as a set of points and synthesizes view-dependent features, including the semantic label of each point, for matching purposes. They embed images into circular features, which are then compared to pose features in the same embedded space. By contrast, we model the semantics as fine-grained ray predictions. Furthermore, unlike these prior works, our approach can operate over images with non-zero pitch and roll.", + "bbox": [ + 109, + 90, + 480, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 112, + 474, + 202, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we propose a floorplan localization framework that jointly estimates semantic and depth rays to infer the 2D camera location and orientation relative to a given floorplan. Specifically, we assume that we are provided with an RGB Image $I \\in \\mathbb{R}^{h \\times w \\times 3}$ , where $h$ and $w$ denote the height and width of the image, respectively, and a 2D floorplan map $F \\in \\{0,1,2,\\dots,C\\}^{H \\times W}$ , represented as a matrix of dimensions $H \\times W$ , where each element is assigned a semantic label from $C$ unique semantic categories, with zero denoting empty space. The semantic categories we consider in our work are wall, window and door, but our framework could easily incorporate additional categories (e.g., staircases, columns).", + "bbox": [ + 109, + 507, + 482, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our objective is to predict the camera's 2D location $(x,y)$ and orientation angle $\\theta$ at which the image $I$ was captured. That is, given the observation $O_{I,F} = (I,F)$ , our goal is to infer the location parameters $S_{I,F} = (x,y,\\theta)$ . To this end, we adopt a probabilistic framework by modeling the posterior distribution $p(S_{I,F}\\mid O_{I,F})$ . We discretize the camera pose space as $\\mathcal{S} = \\{S_i\\}$ and define a probability volume $P\\in \\mathbb{R}^{\\hat{H}\\times \\hat{W}\\times O}$ where each element $P(S_{i})$ represents the posterior probability $p(S_i\\mid O_{I,F})$ for a candidate pose $S_{i}$ . Here, $\\hat{H}$ and $\\hat{W}$ denote the number of discretized cells in the $x$ and $y$ dimensions, respectively, and $O$ represents the number of orientation bins. The", + "bbox": [ + 109, + 710, + 482, + 912 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "final predicted camera pose is then given by", + "bbox": [ + 513, + 90, + 805, + 107 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat{S}_{I,F} = \\operatorname *{arg max}_{S_{i}\\in \\mathcal{S}}p(S_{i}\\mid O_{I,F}).\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 114, + 803, + 142 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We proceed to provide background (Section 3.1), prior to introducing our semantic prediction network (Section 3.2), which constructs a semantic probability volume. We then describe how it is fused with depth cues to perform floorplan localization (Section 3.3). Finally, we provide training and implementation details (Section 3.4); additional details can be found in the supplementary material. An overview of our approach is provided in Figure 2.", + "bbox": [ + 511, + 152, + 883, + 289 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Background: F3Loc", + "text_level": 1, + "bbox": [ + 511, + 297, + 705, + 314 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our work builds upon F3Loc [7], a recent technique that estimates depth rays for performing floorplan localization given a single image or image sequence. We briefly outline several key components from their work that provide background for our framework. For additional details, we refer readers to their work.", + "bbox": [ + 511, + 320, + 883, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Depth Rays Prediction. Given a query image, a depth prediction network estimates per-column depth values that capture the distance from the camera to the nearest wall along specific angles. These values are then linearly interpolated to produce a fixed set of equiangular depth rays $\\hat{r}_d\\in \\mathbb{R}^l$ that represent the floorplan depth, with $l$ denoting the number of predicted rays.", + "bbox": [ + 511, + 417, + 883, + 525 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Estimating Depth Probability Volume. For every candidate location $(x,y)$ on the floorplan and each discrete orientation $\\theta$ , a corresponding set of reference rays is generated based on the floorplan's geometry. The predicted depth rays are compared with these reference rays to compute a likelihood score for each grid cell and orientation, resulting in a three-dimensional probability volume $P_{d} \\in [0,1]^{\\left[\\hat{H},\\hat{W},O\\right]}$ . For instance, given a $10\\mathrm{m} \\times 7\\mathrm{m}$ floorplan discretized at $0.1\\mathrm{m}$ with $10^{\\circ}$ increments in orientation yields a probability volume $P_{d} \\in [0,1]^{[100,70,36]}$ .", + "bbox": [ + 511, + 531, + 883, + 699 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The final camera 2D location and orientation are determined by selecting the grid cell with the highest likelihood in the probability volume. This process maximizes the posterior probability, thereby estimating the camera's $x$ and $y$ coordinates as well as its orientation.", + "bbox": [ + 511, + 699, + 880, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Adding a Semantic Prediction Network", + "text_level": 1, + "bbox": [ + 511, + 784, + 849, + 799 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To utilize semantic cues for performing floorplan localization, we propose to add a semantic prediction network that first predicts semantic rays, and then processes these for constructing a semantic probability volume. We provide additional details in what follows, and then present an optional room type prediction component, which can be utilized if room labels are available.", + "bbox": [ + 511, + 806, + 883, + 911 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27118", + "bbox": [ + 478, + 950, + 517, + 962 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f38ed0d871f009b8ab30b71b5f53881bff7c9ffcd4a6b084b349294da124e2b3.jpg", + "image_caption": [ + "Figure 2. Overview of our pipeline. The input image is processed to generate depth rays, semantic rays, and optionally additional metadata (e.g., room type prediction). We interpolate the ray predictions to a low-resolution representation and generate the depth probability volume $P_{d}$ and the semantic probability volume $P_{s}$ (optionally masked according to the room type prediction). These probability volumes are then fused to form the structural-semantic probability volume $P_{c}$ for efficient coarse localization. Finally, we refine the candidate poses using high-resolution ray predictions and predict the final 2D camera location and orientation, visualized with an arrow on the right." + ], + "image_footnote": [], + "bbox": [ + 112, + 98, + 874, + 262 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/50938694901a08d9a0703247fc0202c72c1c2e8939d5799855a233394475f580.jpg", + "image_caption": [ + "Figure 3. Overview of our semantic prediction network that predicts a set of semantic rays through the Semantic Ray Branch (top) and an optional room type value—e.g., Living Room—through the Room Classification Branch (bottom). The room type is used for extracting the mask $M_{\\mathrm{room}}$ , as visualized on the bottom right." + ], + "image_footnote": [], + "bbox": [ + 112, + 388, + 473, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantic Rays Prediction. Unlike the continuous depth values estimated in prior work, the semantic rays should correspond to semantic categories, which are represented as a set of discrete classes. Therefore, we construct a network that produces a semantic ray representation $\\hat{r}_s \\in \\{1, \\dots, C\\}^l$ from the image, where each ray is classified into one of $C$ semantic categories. We provide an overview of our semantic ray prediction network in Figure 3.", + "bbox": [ + 109, + 623, + 482, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in the figure, our semantic network architecture leverages a pretrained ResNet50 backbone to extract robust, high-level features from an input RGB image $I$ . After reducing the feature channels using a CNN and projecting them into a lower-dimensional subspace, positional encodings are computed to preserve spatial information. Two sets of learnable tokens are introduced: a set of $l$ ray tokens responsible for predicting the semantic ray representation $\\hat{r}_s$ and a single (optional) CLS token dedicated for representing global room clas", + "bbox": [ + 109, + 761, + 482, + 912 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sification information.", + "bbox": [ + 513, + 388, + 663, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A single-head cross attention module integrates these tokens with the flattened spatial features, yielding refined tokens that capture both global context and local details. In the ray branch, the refined ray tokens are first processed by a self-attention block that enables each token to interact with all others, thereby aggregating complementary contextual information. The enriched tokens are then passed through an MLP to produce per-token semantic logits, which after normalization form the final semantic ray vector $\\hat{r}_s$ . If room labels are available in the dataset, a similar network processes the CLS token for room type prediction, as we further detail later.", + "bbox": [ + 511, + 407, + 883, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Estimating Semantic Probability Volume. To obtain the semantic probability map, $P_{s} \\in [0,1]^{[\\hat{H},\\hat{W},O]}$ , we first need to interpolate the $l$ predicted semantic rays. Regular linear interpolation—which prior work used for depth estimation—is unsuitable in the context of discrete labeling since interpolating between class labels can produce non-valid or semantically meaningless results. Instead, we propose a voting-based interpolation scheme: We reduce the original equiangular rays to the desired count by applying a majority vote within a small neighborhood. We use a window of three rays, assigning the label that appears most frequently in that window to the center target ray; see the supplementary material for the full algorithm. Next, we compute the score for each set of rays by taking the $L_{1}$ difference between the predicted semantic labels and the reference labels. The score is then exponentiated and normalized to form the semantic probability volume $P_{s}$ , which quantifies the likelihood of each candidate pose based on the alignment between the semantic rays and the candidate pose.", + "bbox": [ + 511, + 608, + 883, + 912 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27119", + "bbox": [ + 478, + 950, + 517, + 962 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Room Type Prediction. In addition to predicting semantic rays, our semantic network can optionally also predict the room type, which corresponds to the room from which the input image was taken. This is achieved by processing the CLS token in the semantic network (see Figure 3). If the predicted room probability exceeds a threshold $T_{\\mathrm{room}}$ , the predicted room type is then used to construct a mask $M_{\\mathrm{room}}$ from the polygons associated with that room type. For example, if the model predicts a Living Room label with high confidence, the mask $M_{\\mathrm{room}}$ retains only the regions corresponding to the living room by setting all other areas to zero. Let $\\tilde{P}$ denote the original probability volume, then the masked probability volume $P$ is computed as:", + "bbox": [ + 111, + 90, + 483, + 303 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP = M _ {\\mathrm {r o o m}} \\odot \\tilde {P},\n$$\n", + "text_format": "latex", + "bbox": [ + 238, + 316, + 357, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\odot$ denotes element-wise multiplication, thereby filtering out all the probabilities which are not in the living room and substantially narrowing down the search space. A detailed analysis of room type distributions and the model's prediction accuracy is provided in the supplementary material.", + "bbox": [ + 111, + 340, + 482, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Floorplan Localization via a Structural-Semantic Probability Volume", + "text_level": 1, + "bbox": [ + 112, + 439, + 483, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We obtain the final probability map by generating the depth probability map $P_{d}$ , following the procedure described in Section 3.1, and the semantic probability map $P_{s}$ , according to the approach detailed in Section 3.2. To leverage both semantic and geometric cues, we fuse the two probability maps using a weighted combination:", + "bbox": [ + 111, + 476, + 483, + 568 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {c} = w _ {s} \\cdot P _ {s} + w _ {d} \\cdot P _ {d},\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 579, + 380, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w_{s}$ denotes the weight given to the semantic cues, while $w_{d} = 1 - w_{s}$ represents the weight of the depth cues. These weights are determined using a held-out validation set, as further detailed in Section 4. Note that as both $P_{s}$ and $P_{d}$ are generated from interpolated rays, which we refer to as low-resolution rays (see Figure 4, predicted), and hence the initial probability volume $P_{c}$ is also in low resolution.", + "bbox": [ + 111, + 606, + 483, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Location Extraction. Our approach follows a coarse-to-fine strategy to achieve precise localization while maintaining computational efficiency. Given $P_{c}$ , we first extract the Top- $k$ candidate poses from the structural-semantic probability volume based on their scores, ensuring that each candidate is separated by at least $\\delta_{\\mathrm{res}}$ in translation. Next, for each candidate, we generate an augmented set of orientations by perturbing its original angle in increments of $\\pm \\delta_{\\mathrm{ang}}$ up to a maximum deviation of $\\Delta_{\\mathrm{max}}$ , resulting in the final augmented set:", + "bbox": [ + 111, + 734, + 483, + 885 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n[ 0, \\pm \\delta_ {\\text {a n g}}, \\pm 2 \\delta_ {\\text {a n g}}, \\dots , \\pm \\Delta_ {\\max} ].\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 896, + 401, + 912 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a651a3996ada6479ed5a323afd0cf046076b6e13ee7d32c53e35d5047ae6c4ac.jpg", + "image_caption": [ + "Figure 4. Comparing low-resolution and high-resolution rays with our coarse-to-fine approach. Given an input image (left), we construct a structural-semantic probability volume by comparing low resolution ground-truth and predicted rays (center). Location extraction from this coarse volume directly often yields significant errors, as illustrated with the yellow arrows. Results are refined only for the Top-k candidate poses by comparing the high resolution ground-truth and predicted rays (right). This allows for efficiently extracting more accurate predictions, as further detailed in Section 3.3." + ], + "image_footnote": [], + "bbox": [ + 516, + 90, + 880, + 209 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At each candidate location, we compute the corresponding high-resolution ground-truth depth and semantic ray representations, yielding $l$ rays per candidate location. These ground-truth rays are then compared against the original predicted $l$ rays using a similarity metric (see supplementary material for metric details). For a visual comparison of the high-resolution ground-truth and predicted rays, please refer to Figure 4. The candidate with the highest similarity score is selected as the final predicted location.", + "bbox": [ + 511, + 378, + 883, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This coarse-to-fine refinement process effectively leverages the full resolution of the predictions, which were initially interpolated for runtime efficiency, to achieve more precise localization. Note that there is a tradeoff between accuracy and computation time: finer resolutions yield improved precision at the expense of increased computational load. This tradeoff is illustrated in Figure 4. As can be observed, comparing low resolution rays often yields inaccurate predictions. For instance, environments with multiple semantic objects can suffer from loss of critical details (e.g., the left door is omitted in the middle row) in the low-resolution prediction, which can lead to misclassification, further motivating the refinement step. We present a quantitative analysis of this tradeoff in the supplementary material.", + "bbox": [ + 511, + 530, + 883, + 756 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training and Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 768, + 834, + 784 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Both networks are trained in an end-to-end manner. For depth prediction, an $L_{1}$ loss supervises the predicted depths against ground-truth depth maps. For semantic prediction, a cross-entropy loss is applied to the predicted semantic labels. If room labels $R$ are available in the dataset, an additional cross-entropy loss is used for the room label, and the network is trained to optimize both objectives jointly. As in prior work, data aug", + "bbox": [ + 511, + 791, + 883, + 912 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27120", + "bbox": [ + 478, + 950, + 519, + 963 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mentation techniques—including virtual roll-pitch augmentation—are employed to improve robustness to nonupright camera poses. The networks are optimized using the Adam optimizer with an initial learning rate of $1 \\times 10^{-3}$ . We use a floorplan resolution of $0.1\\mathrm{m}$ and an angular granularity of $10^{\\circ}$ . We set the number of predicted rays to $l = 40$ and interpolate these to 7 rays during the coarse stage of localization. For the Location Extraction module, we use $\\delta_{\\mathrm{res}} = 0.1\\mathrm{m}$ , $\\delta_{\\mathrm{ang}} = 5^{\\circ}$ , and $\\Delta_{\\mathrm{max}} = 5^{\\circ}$ , and refine our results using Top-5 poses.", + "bbox": [ + 111, + 90, + 480, + 243 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Results", + "text_level": 1, + "bbox": [ + 112, + 256, + 199, + 272 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we present a comprehensive evaluation of our localization method. We begin by introducing the datasets we use in our experiments, followed by a discussion of the baselines we compare against and the evaluation metrics. The main results are reported in Section 4.1. We conduct an ablation study to assess the contributions of various components of our approach in Section 4.2. We report results for our approach under two conditions: one in which room labels are not utilized (denoted as $\\mathrm{Ours}_s$ ) and another in which room labels are employed to further refine the predictions (denoted as $\\mathrm{Ours}_r$ ). Additional experiments and qualitative results are provided in the supplementary material.", + "bbox": [ + 111, + 282, + 480, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We conduct experiments on two popular datasets: Structured 3D (S3D) [37] and ZInD [9]. S3D is a synthetic dataset containing realistic 3D renders of 3,296 houses. We use the fully furnished version of S3D, as employed in previous works. ZInD consists of 1,575 unfurnished homes containing only panoramic images. We crop these panoramas to perspective views with an $80^{\\circ}$ field of view (as was done in S3D), and generate a fixed-size dataset from the resulting images. For both datasets, we follow their official train and test splits.", + "bbox": [ + 111, + 487, + 480, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our approach against two baselines: F3Loc [7] (considering the single-image localization component) and LASER [22]. We also report performance using an ORACLE ray prediction. This oracle ray prediction simulates the best possible performance achievable by our pipeline using ground truth depth and semantic rays. Note that the Oracle ray prediction baseline does not incorporate room-aware features. For F3Loc, results on the ZIND dataset are obtained from our experiments by running their publicly available code on the dataset, as the original paper does not include results on this dataset. We also use the publicly available implementation of LASER. Additional details are provided in the supplementary material.", + "bbox": [ + 111, + 646, + 480, + 859 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. Following prior work [7, 22], we report recall metrics computed at distance thresholds of $0.1\\mathrm{m}$ , $0.5\\mathrm{m}$ , and $1\\mathrm{m}$ . We also report recall for predictions with an", + "bbox": [ + 111, + 866, + 480, + 912 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c665924910f6f0dc5c2eff4e5cc9426eadcf15f39e049037b6c4f35b20e4e44c.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Recall performance on the S3D and ZInD datasets. The table reports recall at thresholds of $0.1\\mathrm{m}$ , $0.5\\mathrm{m}$ , $1\\mathrm{m}$ , and $1\\mathrm{m}$ with a $30^{\\circ}$ orientation tolerance for LASER, F3Loc, our approach without room labels $(\\mathrm{Ours}_s)$ , our approach with room labels $(\\mathrm{Ours}_r)$ , and an Oracle ray prediction." + ], + "table_body": "
S3D R@
Method0.1m0.5m1m1m 30°
LASER0.76.410.48.7
F3Loc1.514.622.421.3
Ourss5.4241.8753.5252.61
Oursr5.7045.5358.7857.49
ORACLE61.0093.8494.8794.57
ZInD R@
Method0.1m0.5m1m1m 30°
LASER1.3811.0617.5513.64
F3Loc0.677.9015.0711.46
Ourss2.9824.0033.9629.30
Oursr3.3126.6038.0131.86
ORACLE26.4260.8567.6965.13
", + "bbox": [ + 535, + 88, + 859, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "orientation error bounded to less than $30^{\\circ}$ (at the $1\\mathrm{m}$ threshold). Recall is defined as the percentage of predictions that fall within these thresholds.", + "bbox": [ + 511, + 440, + 883, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 511, + 496, + 733, + 512 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results are reported in Table 1. On both S3D and ZinD, our method more than doubles F3Loc's and LASER performance across all thresholds. Notably, considering S3D over the $\\mathrm{R@1m30^{\\circ}}$ metric—which reflects the quality of matches between the predicted and actual camera views—in comparison to F3Loc, our method improves by more than three times. Room type predictions yield improvements of $9.2\\%$ in $\\mathrm{R@1m30^{\\circ}}$ on the S3D dataset and $8.7\\%$ on the ZInD dataset. We include an additional experiment in the supplementary material that evaluates F3Loc with our refinement module, demonstrating that both the semantic rays and our refinement module provide significant performance gains.", + "bbox": [ + 511, + 518, + 883, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We observe that significant performance improvements are also achieved for a very fine-grained recall metric of $0.1\\mathrm{m}$ , boosting performance from $1.5\\%$ (F3Loc) to $5.7\\%$ with our approach on S3D. We attribute this to our coarse-to-fine strategy, which effectively refines the coarse location estimates into precise predictions, as further validated in our ablation study.", + "bbox": [ + 511, + 715, + 882, + 820 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 5 presents qualitative examples that illustrate how the integration of floorplan semantics with precise depth cues enables our pipeline to effectively resolve localization ambiguities. For instance, in the third row, we can see that F3Loc, which does not use semantics, interprets this image as a blank wall, while LASER misinter", + "bbox": [ + 511, + 821, + 883, + 912 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27121", + "bbox": [ + 478, + 950, + 517, + 962 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/95882f81b1b62fdf7506fdbf047cbb5cf655a42fff36b39c99d5aed1dcf943be.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison of our method with F3Loc and LASER. Warmer colors correspond to regions with higher probabilities. Below each map we report the localization error in meters and degrees. We use arrows to visualize the ground truth location (magenta) and the predicted location (white)." + ], + "image_footnote": [], + "bbox": [ + 116, + 85, + 478, + 294 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "prets the window size and predicts an incorrect location.", + "bbox": [ + 112, + 395, + 480, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 112, + 420, + 264, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We demonstrate the effect of incorporating each component of our method on overall localization performance. Specifically, we conduct the following ablations: (1) Base, which corresponds to computing the structural-semantic probability volume and selecting the maximum probability without any additional refinement. (2) Removing semantic interpolation (denoted as -Interpolation), where we replace our majority voting interpolation with a simple linear interpolation followed by rounding. (3) Adding room predictions (denoted as +Room), where we assess the effect of integrating room type predictions into our localization pipeline. (4) Adding refinement (denoted as +Refine), which assesses our coarse-to-fine approach, which refines our localization extraction via the Top-K candidate poses. (5) Adding room predictions and refinement (denoted as +Room&Refine), which combines both components.", + "bbox": [ + 111, + 443, + 482, + 699 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From the results in Table 2, we can see that on the $1\\mathrm{m}30^{\\circ}$ metric, the addition of our refinement module improves performance by $8.6\\%$ relative to the baseline. This indicates that a substantial amount of information is lost during the initial interpolation process if not properly addressed, thereby strongly motivating the use of coarse-to-fine strategies in our Location Extraction module. We also observe a significant improvement of $11.5\\%$ from the room prediction component, which is discussed in further detail in the supplementary material and visualized in Figure 6. Finally, our majority voting interpolation approach contributes an additional gain of $0.9\\%$ compared to a simple interpolation strategy. When combining all these improvements, our method achieves", + "bbox": [ + 111, + 700, + 483, + 912 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/10a493b75ee35579609dd4b728f19874a6442cfc0f9f9772bab6c39c90d0bb91.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method0.1m0.5m1m1m 30°
Base4.6538.3549.4048.44
- Interpolation4.7338.4448.9147.99
+ Room5.1242.9255.5754.04
+ Refine5.4241.8753.5252.61
+ Room&Refine5.7045.5358.7857.49
", + "bbox": [ + 516, + 87, + 882, + 183 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation study, evaluating the effect of incorporating the different components in our pipeline on the S3D dataset.", + "bbox": [ + 511, + 184, + 882, + 213 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/29b668141624274f1c51c29320cb2b933f770b5b1e8861e9b87f091d4c1df815.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison of using room predictions and our coarse-to-fine refinement approach. Below each map we report the localization error in meters and degrees. Warmer colors correspond to regions with higher probabilities. Overlaid on the estimated probabilities, we show the ground truth location (magenta) and the predicted location." + ], + "image_footnote": [], + "bbox": [ + 514, + 229, + 877, + 368 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "an overall enhancement of $18.6\\%$ on the 1m $30^{\\circ}$ metric relative to our base model.", + "bbox": [ + 511, + 484, + 882, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Figure 6, we illustrate the impact of incorporating room type predictions alongside the location extraction module. The figure clearly demonstrates how these components refine the probability volume by narrowing down the candidate poses, which in turn improves overall localization accuracy.", + "bbox": [ + 511, + 516, + 883, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We analyze the impact of different combinations of predicted depth and semantic features on the S3D dataset. Figure 7 summarizes the recall performance for various depth and semantic weight configurations used to compute the structural-semantic probability volume. In Figure 8 we visualize examples from two extreme scenarios (depth only and semantic only) and from the configuration adopted in our work $\\left([w_{s}, w_{d}] = [0.4, 0.6]\\right)$ , selected according to the validation set.", + "bbox": [ + 511, + 608, + 883, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As can be observed from Figure 8 (and also reflected in Figure 7), relying solely on semantic rays tends to produce a more diffused probability volume with multiple ambiguous candidate locations. This occurs because most images contain a single semantic object of a standard size, allowing an accurate ray pattern to be fitted in more than one location. Similarly, depth cues alone also suffer from ambiguity, particularly in repetitive environments or when the image captures three or fewer walls, which leads to uncertainty in the localization estimate. However, when these semantic cues are com", + "bbox": [ + 511, + 746, + 883, + 910 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27122", + "bbox": [ + 478, + 950, + 517, + 962 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c2dfec3c57efe79de669b4ba1270029e8c3f6e09f854cb967d0b2f1cdb4e7d82.jpg", + "image_caption": [ + "Figure 7. Recall vs. weight combinations on the S3D test set. The plot shows the recall metrics for four different thresholds: $0.1m$ , $0.5m$ , $1m$ , and $1m30^{\\circ}$ . The x-axis displays the weight combinations in the order $(w_{s}, w_{d})$ . A vertical dashed line at $w_{s} = 0.4$ , $w_{d} = 0.6$ highlights the weight combination selected using the validation set." + ], + "image_footnote": [], + "bbox": [ + 114, + 88, + 480, + 289 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5dc6920b295726ffc7ed155da72789d9b30d5f087949db643d7f4207bff66beb.jpg", + "image_caption": [ + "Figure 8. Qualitative Comparison of Depth-Only, Semantic-Only, and Fused Structural-Semantic Probability Volumes. Below each map we report the localization error in meters and degrees. Warmer colors correspond to regions with higher probabilities." + ], + "image_footnote": [], + "bbox": [ + 116, + 400, + 478, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "bined with depth rays, the resulting probability volume becomes significantly more concentrated. This integration effectively filters out spurious candidates and sharpens the localization estimate.", + "bbox": [ + 111, + 715, + 480, + 775 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additionally, the supplementary material offers a detailed analysis of how varying the Top-K candidates influences localization refinement, along with additional experiments that motivate various design choices, such as using hard thresholds in multiple steps of our pipeline.", + "bbox": [ + 111, + 777, + 480, + 852 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Runtime Analysis", + "text_level": 1, + "bbox": [ + 112, + 862, + 241, + 877 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We report runtime performance breakdown in Table 3, using the same parameters employed in Table 1 with", + "bbox": [ + 112, + 881, + 480, + 912 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c4ff0c1de34ba0cda5cb5eade46e3c8ff35eea79f3b999ca06da73d775f6f102.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
KPredictionLocRefinTotal
10.038 ± 0.1190.174 ± 0.0330.141 ± 0.0730.364 ± 0.142
30.033 ± 0.0930.154 ± 0.0260.356 ± 0.1190.554 ± 0.157
50.034 ± 0.1030.155 ± 0.0290.577 ± 0.1850.778 ± 0.218
", + "bbox": [ + 517, + 89, + 879, + 171 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Performance breakdown over different Top-K values. Each entry is mean ± std (s). Prediction denotes the ray predictions, Loc refers to the localization process, and Refin represents the refinement stage, during which candidate locations are evaluated by computing the ground truth rays.", + "bbox": [ + 511, + 181, + 883, + 252 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "varying K values. As shown in the table, the per-image inference time increases with the number of candidate refinements, $K$ , reflecting the additional computations required during refinement. All experiments were conducted on a single CPU without multithreading to avoid introducing bias. Future work could explore parallelizing the refinement stage by computing all ground-truth rays simultaneously for further speed-up. Importantly, even at $K = 5$ the setting we adopt in our work, the inference time remains reasonable, striking a practical balance between computational cost and improvements in localization accuracy. We further observe that the prediction and localization steps require similar amounts of time, while the refinement step grows monotonically as $K$ increases.", + "bbox": [ + 511, + 280, + 883, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 522, + 632, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we presented a semantic-aware localization framework that extends floorplan-based camera localization by fusing semantic labels with geometric depth cues. Our approach leverages a novel semantic ray prediction network alongside an established depth estimation method to generate a semantic-structural probability volume, which significantly improves localization accuracy, especially in environments with repetitive or ambiguous structural patterns.", + "bbox": [ + 511, + 547, + 883, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our extensive experiments on the S3D and ZInD datasets demonstrate that integrating semantic cues effectively resolves depth-based ambiguities and consistently outperforms state-of-the-art methods such as F3Loc and LASER. Ablation studies confirm that a balanced combination of depth and semantic information, coupled with a coarse-to-fine localization strategy and the use of room labels, yields optimal performance.", + "bbox": [ + 511, + 685, + 882, + 805 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Looking forward, extending our framework to incorporate additional semantic labels and other modalities, such as textual information, promises to further enhance localization robustness in challenging indoor settings. In general, our approach represents an important step towards accurate and reliable indoor localization systems by effectively leveraging semantic and geometric cues.", + "bbox": [ + 511, + 806, + 883, + 912 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27123", + "bbox": [ + 478, + 950, + 517, + 962 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 114, + 89, + 209, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Relja Arandjelovic, Petr Gronat, Akihiko Torii, Tomas Pajdla, and Josef Sivic. NetVLAD: CNN architecture for weakly supervised place recognition. In CVPR, pages 5297-5307, 2016. 2", + "[2] Vassileios Balntas, Shuda Li, and Victor Prisacariu. Relocnet: Continuous metric learning relocalisation using neural nets. In ECCV, pages 751-767, 2018. 2", + "[3] Federico Boniardi, Tim Caselitz, Rainer Kummerle, and Wolfram Burgard. Robust lidar-based localization in architectural floor plans. In 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3318-3324. IEEE, 2017. 2", + "[4] Federico Boniardi, Tim Caselitz, Rainer Kummerle, and Wolfram Burgard. A pose graph-based localization system for long-term navigation in cad floor plans. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 84-97. IEEE, 2019. 2", + "[5] Federico Boniardi, Abhinav Valada, Rohit Mohan, Tim Caselitz, and Wolfram Burgard. Robot localization in floor plans using a room layout edge extraction network. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5291-5297. IEEE, 2019. 2", + "[6] Eric Brachmann, Alexander Krull, Sebastian Nowozin, Jamie Shotton, Frank Michel, Stefan Gumhold, and Carsten Rother. DSAC-differentiable RANSAC for camera localization. In CVPR, pages 6684-6692, 2017. 2", + "[7] Changan Chen, Rui Wang, Christoph Vogel, and Marc Pollefeys. F3Loc: Fusion and filtering for floorplan localization. arXiv preprint arXiv:2403.03370, 2024. 1, 2, 3, 6", + "[8] Hang Chu, Dong Ki Kim, and Tsuhan Chen. You are here: Mimicking the human thinking process in reading floorplans. In ICCV, pages 2210-2218, 2015. 1, 2", + "[9] Steve Cruz, Will Hutchcroft, Yuguang Li, Naji Khosravan, Ivaylo Boyadzhiev, and Sing Bing Kang. Zillow indoor dataset: Annotated floor plans with $360^{\\circ}$ panoramas and 3d room layouts. In CVPR, pages 2133-2143, 2021. 6", + "[10] Niklas Gard, Anna Hilsmann, and Peter Eisert. Spvloc: Semantic panoramic viewpoint matching for 6d camera localization in unseen environments. arXiv preprint arXiv:2404.10527, 2024.3", + "[11] B. Glocker, S. Izadi, J. Shotton, and A. Criminisi. Realtime RGB-D camera relocalization. In Proceedings of the Mixed and Augmented Reality (ISMAR) Conference. IEEE, 2013. 1", + "[12] Henry Howard-Jenkins and Victor Adrian Prisacariu. LaLaLoc++: Global floor plan comprehension for layout localisation in unvisited environments. In ECCV, pages 693-709, 2022. 1, 2", + "[13] Henry Howard-Jenkins, Jose-Raul Ruiz-Sarmiento, and Victor Adrian Prisacariu. LaLaLoc: Latent layout localisation in dynamic, unvisited environments. arXiv preprint arXiv:2104.09169, 2021. 1, 2" + ], + "bbox": [ + 114, + 114, + 483, + 912 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Seigo Ito, Felix Endres, Markus Kuderer, Gian Diego Tipaldi, Cyril Stachniss, and Wolfram Burgard. W-RGBD: Floor-plan-based indoor global localization using a depth camera and wifi. In 2014 IEEE International Conference on Robotics and Automation (ICRA), pages 417-422. IEEE, 2014. 2", + "[15] Peter Karkus, David Hsu, and Wee Sun Lee. Particle filter networks with application to visual localization. In Proceedings of the 2nd Conference on Robot Learning (CoRL), pages 169-178. PMLR, 2018. 1, 3", + "[16] Alex Kendall, Matthew Grimes, and Roberto Cipolla. PoseNet: A convolutional network for real-time 6-DOF camera relocalization. In ICCV, pages 2938-2946, 2015. 1, 2", + "[17] Junho Kim, Jiwon Jeong, and Young Min Kim. Fully geometric panoramic localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3", + "[18] Z. Li, M. H. Ang, and D. Rus. Online localization with imprecise floor space maps using stochastic gradient descent. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8571-8578. IEEE, 2020. 2", + "[19] Liu Liu, Hongdong Li, and Yuchao Dai. Efficient global 2D-3D matching for camera localization in a large-scale 3D map. In ICCV, pages 2372-2381, 2017. 1, 2", + "[20] Pranay Mathur, Rajesh Kumar, and Sarthak Upadhyay. Sparse image-based navigation architecture to mitigate the need of precise localization in mobile robots. arXiv preprint arXiv:2203.15272, 2022. 1", + "[21] Oscar Mendez, Simon Hadfield, Nicolas Pugeault, and Richard Bowden. SeDAR: Reading floorplans like a human—using deep learning to enable human-inspired localisation. *IJCV*, 128:1286–1310, 2020. 3", + "[22] Zhixiang Min, Naji Khosravan, Zachary Bessinger, Manjunath Narayana, Sing Bing Kang, Enrique Dunn, and Ivaylo Boyadzhiev. LASER: Latent space rendering for 2D visual localization. In CVPR, pages 11122-11131, 2022. 3, 6", + "[23] R. A. Newcombe, A. J. Davison, S. Izadi, P. Kohli, O. Hilliges, J. Shotton, D. Molyneaux, S. Hodges, D. Kim, and A. Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In Proceedings of the Mixed and Augmented Reality (ISMAR) Conference. IEEE, 2011. 1", + "[24] Takahiro Niwa, Shun Taguchi, and Noriaki Hirose. Spatio-temporal graph localization networks for image-based navigation. arXiv preprint arXiv:2204.13237, 2022.1", + "[25] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: robust hierarchical localization at large scale. In CVPR, pages 12716-12725, 2019. 1, 2", + "[26] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Fast image-based localization using direct 2d-to-3d matching. In ICCV, pages 667-674, 2011.", + "[27] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Improving image-based localization by active correspondence search. In ECCV, pages 752-765, 2012." + ], + "bbox": [ + 516, + 92, + 883, + 912 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27124", + "bbox": [ + 478, + 950, + 519, + 963 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Efficient & effective prioritized matching for large-scale image-based localization. IEEE TPAMI, 39(9):1744-1756, 2016. 1, 2", + "[29] Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene coordinate regression forests for camera relocalization in rgb-d images. In CVPR, pages 2930-2937, 2013. 1, 2", + "[30] Janine Thoma, Danda Pani Paudel, Ajad Chhatkuli, Thomas Probst, and Luc Van Gool. Mapping, localization and path planning for image-based navigation using visual features and map. In CVPR. IEEE, 2019. 1", + "[31] Julien Valentin, Matthias Nießner, Jamie Shotton, Andrew Fitzgibbon, Shahram Izadi, and Philip HS Torr. Exploiting uncertainty in regression forests for accurate camera relocalization. In CVPR, pages 4400-4408, 2015. 1, 2", + "[32] Florian Walch, Caner Hazirbas, Laura Leal-Taixe, Torsten Sattler, Sebastian Hilsenbeck, and Daniel Cremers. Image-based localization using lstms for structured feature correlation. In ICCV, pages 627-637, 2017. 1, 2", + "[33] Shenlong Wang, Sanja Fidler, and Raquel Urtasun. Lost shopping! monocular localization in large indoor spaces. In ICCV, pages 2695-2703, 2015. 2", + "[34] Xipeng Wang, Ryan J Marcotte, and Edwin Olson. GLFP: Global localization from a floor plan. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1627-1632. IEEE, 2019. 2", + "[35] Wera Winterhalter, Freya Fleckenstein, Bastian Steder, Luciano Spinello, and Wolfram Burgard. Accurate indoor localization for RGB-D smartphones and tablets given 2d floor plans. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3138-3143. IEEE, 2015. 1", + "[36] Jian Wu, Liwei Ma, and Xiaolin Hu. Delving deeper into convolutional neural networks for camera relocalization. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 5644-5651. IEEE, 2017. 1, 2", + "[37] Jia Zheng, Junfei Zhang, Jing Li, Rui Tang, Shenghua Gao, and Zihan Zhou. Structured3d: A large photorealistic dataset for structured 3d modeling. In ECCV, 2020.6" + ], + "bbox": [ + 114, + 90, + 483, + 696 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27125", + "bbox": [ + 478, + 950, + 517, + 962 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Supercharging Floorplan Localization with Semantic Rays/834ff1eb-78d1-464a-8d20-1e5cbed96504_model.json b/2025/Supercharging Floorplan Localization with Semantic Rays/834ff1eb-78d1-464a-8d20-1e5cbed96504_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c4fe2278542046921c0cf4c6b1ff71833da1b2cd --- /dev/null +++ b/2025/Supercharging Floorplan Localization with Semantic Rays/834ff1eb-78d1-464a-8d20-1e5cbed96504_model.json @@ -0,0 +1,1859 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.203, + 0.131, + 0.795, + 0.154 + ], + "angle": 0, + "content": "Supercharging Floorplan Localization with Semantic Rays" + }, + { + "type": "text", + "bbox": [ + 0.301, + 0.182, + 0.46, + 0.218 + ], + "angle": 0, + "content": "Yuval Grader \nTel Aviv University" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.182, + 0.697, + 0.218 + ], + "angle": 0, + "content": "Hadar Averbuch-Elor Cornell University" + }, + { + "type": "text", + "bbox": [ + 0.324, + 0.236, + 0.673, + 0.25 + ], + "angle": 0, + "content": "https://tau-vailab.github.io/SemRayLoc/" + }, + { + "type": "title", + "bbox": [ + 0.26, + 0.273, + 0.338, + 0.289 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.307, + 0.485, + 0.671 + ], + "angle": 0, + "content": "Floorplans provide a compact representation of the building's structure, revealing not only layout information but also detailed semantics such as the locations of windows and doors. However, contemporary floorplan localization techniques mostly focus on matching depth-based structural cues, ignoring the rich semantics communicated within floorplans. In this work, we introduce a semantic-aware localization framework that jointly estimates depth and semantic rays, consolidating over both for predicting a structural-semantic probability volume. Our probability volume is constructed in a coarse-to-fine manner: We first sample a small set of rays to obtain an initial low-resolution probability volume. We then refine these probabilities by performing a denser sampling only in high-probability regions and process the refined values for predicting a 2D location and orientation angle. We conduct an evaluation on two standard floorplan localization benchmarks. Our experiments demonstrate that our approach substantially outperforms state-of-the-art methods, achieving significant improvements in recall metrics compared to prior works. Moreover, we show that our framework can easily incorporate additional metadata such as room labels, enabling additional gains in both accuracy and efficiency." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.719, + 0.245, + 0.735 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.746, + 0.483, + 0.913 + ], + "angle": 0, + "content": "Camera localization is a longstanding problem in computer vision, with significant applications in 3D reconstruction [19, 25-28], augmented reality [11, 16, 23, 29, 31], and navigation [20, 24, 30, 32, 36]. Localization within indoor environments is especially challenging due to the absence of reliable GPS signals and the complexity of reasoning across multiple floors and layers. Hence, to bypass complicated 3D model-based localization techniques, prior work [7, 8, 12, 13, 15, 35] has explored the problem of localizing camera observations within a provided 2D floorplan map by matching" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.271, + 0.715, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.352, + 0.716, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.481, + 0.716, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.613, + 0.657, + 0.632 + ], + "angle": 0, + "content": "Input & GT" + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.349, + 0.882, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.482, + 0.882, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.728, + 0.615, + 0.861, + 0.631 + ], + "angle": 0, + "content": "Output Localization" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.642, + 0.885, + 0.74 + ], + "angle": 0, + "content": "Figure 1. Floorplan localization using a raw binary floorplan (middle row) often yields ambiguous predictions. In this work, we utilize a richer, yet readily available, representation: a floorplan enhanced with semantic labels (bottom row). We present an approach that supercharges floorplan localization with semantic rays, enabling for resolving localization ambiguities, as illustrated by the comparison on the right." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.773, + 0.7, + 0.788 + ], + "angle": 0, + "content": "depth-based structural cues." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.792, + 0.885, + 0.913 + ], + "angle": 0, + "content": "However, while floorplans offer a compact and lightweight scene representation, structural cues within floorplans often correlate with multiple candidate locations, particularly for environments with repetitive or symmetric layouts. Consider the example in Figure 1. Can you localize the input image within the floorplan? Provided with just a raw (walls only) floorplan, room corners are indistinguishable and hence localization is" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.952, + 0.52, + 0.964 + ], + "angle": 0, + "content": "27116" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.111, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "highly ambiguous, as can also be observed by the probabilities predicted by the state-of-the-art F3Loc [7] technique (middle row, right). To resolve such ambiguities, we are interested in utilizing a slightly different representation: a semantics-aware floorplan, such as the one illustrated on the bottom row of Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.184, + 0.482, + 0.334 + ], + "angle": 0, + "content": "Accordingly, we introduce a semantic-aware ray-based localization framework that integrates semantic cues with depth-based predictions. In particular, we propose a semantic prediction network that predicts accurate semantic ray representations along with optional additional metadata (such as room labels) from a single RGB image with a limited field-of-view. We process these rays to compute a semantic probability volume, which is then fused with depth information for constructing a structural-semantic probability volume." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.336, + 0.482, + 0.472 + ], + "angle": 0, + "content": "Our framework follows a coarse-to-fine localization strategy. We first operate over a low-resolution image for an efficient floorplan search. This initial search yields the Top-\\(k\\) candidate locations. Finally, we refine these candidates and select the best match using a high-resolution ray representation. By adopting this coarse-to-fine approach, our method effectively constrains the localization search to the most promising regions while keeping the computation cost feasible." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.474, + 0.483, + 0.683 + ], + "angle": 0, + "content": "Our experiments demonstrate that our approach yields improvements by factors ranging from two to three across most metrics compared to the state-of-the-art technique [7], upon which our method is built. This substantial gain underscores the effectiveness of fusing semantic and depth ray predictions into a unified probability volume. We further show that our coarse-to-fine strategy offers a flexible tradeoff between accuracy and computational cost, with performance consistently improving as larger candidate sets (Top- \\(k\\)) are evaluated—making our method adaptable to diverse task requirements. Moreover, incorporating additional metadata further enhances precision, leading to significantly improved localization accuracies." + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.686, + 0.393, + 0.7 + ], + "angle": 0, + "content": "Explicitly stated, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.703, + 0.483, + 0.733 + ], + "angle": 0, + "content": "- We introduce a semantic ray prediction network that receives a single RGB image as input." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.733, + 0.483, + 0.792 + ], + "angle": 0, + "content": "- We propose an efficient and unified framework that fuses semantic and depth ray predictions into a structural-semantic probability volume, which effectively resolves localization ambiguities." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.793, + 0.482, + 0.822 + ], + "angle": 0, + "content": "- Results that demonstrate that significant improvements over state-of-the-art methods." + }, + { + "type": "list", + "bbox": [ + 0.113, + 0.703, + 0.483, + 0.822 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.84, + 0.255, + 0.856 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.867, + 0.483, + 0.913 + ], + "angle": 0, + "content": "Visual Localization. The task of visual localization has received ongoing attention throughout the past several decades. Traditional approaches often rely on image re" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.884, + 0.258 + ], + "angle": 0, + "content": "trieval or on a 3D Structure-from-Motion (SfM) model of the environment. In the image retrieval paradigm, methods such as NetVLAD [1] or RelocNet [2] compare a query image against a database of labeled images. Once the closest match is found, the query pose is approximated by the retrieved image's pose. Other methods explicitly construct a 3D SfM model of a scene to establish 2D-3D correspondences between the query image and the reconstructed 3D structure [19, 25-28]. After matching local image descriptors to 3D points, robust solvers estimate the 6-DoF pose." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.884, + 0.378 + ], + "angle": 0, + "content": "Recent learning-based pipelines deviate from classical 2D-3D matching. Scene-coordinate regression methods predict dense 3D coordinates for every pixel in the query image [6, 29, 31], whereas pose regression methods directly estimate the 6-DoF camera pose via neural networks [16, 32, 36]. Although promising for single-scene scenarios, these methods must be retrained or fine-tuned to handle new environments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.384, + 0.884, + 0.534 + ], + "angle": 0, + "content": "Floorplan Localization. Prior work addressing the task of floorplan localization primarily focused on depth-based cues, leveraging image-derived depth predictions or sensors, to match depth obtained from floorplans. In particular, LiDAR-based methods [3, 4, 18, 34] utilize precise laser scans but restrict usability on most mobile devices. Alternative sources of geometric cues, including semi dense visual odometry (SDVO) [8], or point clouds from depth cameras [14], can circumvent heavy LiDAR hardware." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.884, + 0.851 + ], + "angle": 0, + "content": "Earlier works compare extracted room edges directly to the 2D layout [5], often assuming knowledge of camera or room height [5, 8]. Other approaches, are embedding RGB images and floorplans into a shared metric space in order to do the localization. For instance, LaLaLoc [13] uses panoramic depth layout image which is rendered at known heights at different locations within the floorplan, where the localization is done by doing similarity in the embedded space. LaLaLoc++ [12] removes the hight assumption by embedding the entire floorplan into the feature space. However, these approaches often require an upright camera pose [12, 13], making them less flexible for hand-held or head-mounted devices. F3Loc [7] localizes by predicting depth rays from a given image and generating probability volumes that indicate the likelihood of each the depth prediction to a particular location on the floorplan. Although effectively leverages geometric cues from depth maps, it does not incorporate semantic information, which may reduce its robustness in environments with repetitive structures or occlusions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.852, + 0.882, + 0.913 + ], + "angle": 0, + "content": "Semantic-based cues are less commonly used for indoor localization, but several works have utilized them for this task. Wang et al.[33] extracts scene texts from images in large indoor spaces and performs localiza" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.951, + 0.519, + 0.963 + ], + "angle": 0, + "content": "27117" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.111, + 0.092, + 0.482, + 0.44 + ], + "angle": 0, + "content": "tion by matching this text to the floorplan. SeDAR [21] uses a CNN to extract semantic labels from an image and perform Monte Carlo localization. By contrast to our single-image localization framework, it depends on sequential images and depth sensors. PF-Net [15] applies a differentiable particle filter with a learned observation model, relying on a computationally-intensive embedding process that yields limited performance for the single-image scenario. SPVLoc [10] matches captured images to panoramic renderings to estimate location, leveraging additional height information present in these renderings. Similarly, Kim et al. [17] perform localization by using a pre-captured 3D map of the environment. Our approach does not assume the availability of these additional cues. LASER [22] treats the floorplan as a set of points and synthesizes view-dependent features, including the semantic label of each point, for matching purposes. They embed images into circular features, which are then compared to pose features in the same embedded space. By contrast, we model the semantics as fine-grained ray predictions. Furthermore, unlike these prior works, our approach can operate over images with non-zero pitch and roll." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.475, + 0.204, + 0.491 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.508, + 0.483, + 0.706 + ], + "angle": 0, + "content": "In this work, we propose a floorplan localization framework that jointly estimates semantic and depth rays to infer the 2D camera location and orientation relative to a given floorplan. Specifically, we assume that we are provided with an RGB Image \\( I \\in \\mathbb{R}^{h \\times w \\times 3} \\), where \\( h \\) and \\( w \\) denote the height and width of the image, respectively, and a 2D floorplan map \\( F \\in \\{0,1,2,\\dots,C\\}^{H \\times W} \\), represented as a matrix of dimensions \\( H \\times W \\), where each element is assigned a semantic label from \\( C \\) unique semantic categories, with zero denoting empty space. The semantic categories we consider in our work are wall, window and door, but our framework could easily incorporate additional categories (e.g., staircases, columns)." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.712, + 0.483, + 0.913 + ], + "angle": 0, + "content": "Our objective is to predict the camera's 2D location \\((x,y)\\) and orientation angle \\(\\theta\\) at which the image \\(I\\) was captured. That is, given the observation \\(O_{I,F} = (I,F)\\), our goal is to infer the location parameters \\(S_{I,F} = (x,y,\\theta)\\). To this end, we adopt a probabilistic framework by modeling the posterior distribution \\(p(S_{I,F}\\mid O_{I,F})\\). We discretize the camera pose space as \\(\\mathcal{S} = \\{S_i\\}\\) and define a probability volume \\(P\\in \\mathbb{R}^{\\hat{H}\\times \\hat{W}\\times O}\\) where each element \\(P(S_{i})\\) represents the posterior probability \\(p(S_i\\mid O_{I,F})\\) for a candidate pose \\(S_{i}\\). Here, \\(\\hat{H}\\) and \\(\\hat{W}\\) denote the number of discretized cells in the \\(x\\) and \\(y\\) dimensions, respectively, and \\(O\\) represents the number of orientation bins. The" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.806, + 0.108 + ], + "angle": 0, + "content": "final predicted camera pose is then given by" + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.116, + 0.804, + 0.143 + ], + "angle": 0, + "content": "\\[\n\\hat{S}_{I,F} = \\operatorname *{arg max}_{S_{i}\\in \\mathcal{S}}p(S_{i}\\mid O_{I,F}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.154, + 0.885, + 0.29 + ], + "angle": 0, + "content": "We proceed to provide background (Section 3.1), prior to introducing our semantic prediction network (Section 3.2), which constructs a semantic probability volume. We then describe how it is fused with depth cues to perform floorplan localization (Section 3.3). Finally, we provide training and implementation details (Section 3.4); additional details can be found in the supplementary material. An overview of our approach is provided in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.299, + 0.707, + 0.315 + ], + "angle": 0, + "content": "3.1. Background: F3Loc" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.321, + 0.884, + 0.412 + ], + "angle": 0, + "content": "Our work builds upon F3Loc [7], a recent technique that estimates depth rays for performing floorplan localization given a single image or image sequence. We briefly outline several key components from their work that provide background for our framework. For additional details, we refer readers to their work." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.419, + 0.884, + 0.526 + ], + "angle": 0, + "content": "Depth Rays Prediction. Given a query image, a depth prediction network estimates per-column depth values that capture the distance from the camera to the nearest wall along specific angles. These values are then linearly interpolated to produce a fixed set of equiangular depth rays \\(\\hat{r}_d\\in \\mathbb{R}^l\\) that represent the floorplan depth, with \\(l\\) denoting the number of predicted rays." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.532, + 0.884, + 0.7 + ], + "angle": 0, + "content": "Estimating Depth Probability Volume. For every candidate location \\((x,y)\\) on the floorplan and each discrete orientation \\(\\theta\\), a corresponding set of reference rays is generated based on the floorplan's geometry. The predicted depth rays are compared with these reference rays to compute a likelihood score for each grid cell and orientation, resulting in a three-dimensional probability volume \\(P_{d} \\in [0,1]^{\\left[\\hat{H},\\hat{W},O\\right]}\\). For instance, given a \\(10\\mathrm{m} \\times 7\\mathrm{m}\\) floorplan discretized at \\(0.1\\mathrm{m}\\) with \\(10^{\\circ}\\) increments in orientation yields a probability volume \\(P_{d} \\in [0,1]^{[100,70,36]}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.7, + 0.882, + 0.775 + ], + "angle": 0, + "content": "The final camera 2D location and orientation are determined by selecting the grid cell with the highest likelihood in the probability volume. This process maximizes the posterior probability, thereby estimating the camera's \\(x\\) and \\(y\\) coordinates as well as its orientation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.785, + 0.851, + 0.8 + ], + "angle": 0, + "content": "3.2. Adding a Semantic Prediction Network" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.807, + 0.884, + 0.912 + ], + "angle": 0, + "content": "To utilize semantic cues for performing floorplan localization, we propose to add a semantic prediction network that first predicts semantic rays, and then processes these for constructing a semantic probability volume. We provide additional details in what follows, and then present an optional room type prediction component, which can be utilized if room labels are available." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.952, + 0.519, + 0.963 + ], + "angle": 0, + "content": "27118" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.099, + 0.875, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.28, + 0.885, + 0.366 + ], + "angle": 0, + "content": "Figure 2. Overview of our pipeline. The input image is processed to generate depth rays, semantic rays, and optionally additional metadata (e.g., room type prediction). We interpolate the ray predictions to a low-resolution representation and generate the depth probability volume \\( P_{d} \\) and the semantic probability volume \\( P_{s} \\) (optionally masked according to the room type prediction). These probability volumes are then fused to form the structural-semantic probability volume \\( P_{c} \\) for efficient coarse localization. Finally, we refine the candidate poses using high-resolution ray predictions and predict the final 2D camera location and orientation, visualized with an arrow on the right." + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.389, + 0.475, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.112, + 0.513, + 0.484, + 0.598 + ], + "angle": 0, + "content": "Figure 3. Overview of our semantic prediction network that predicts a set of semantic rays through the Semantic Ray Branch (top) and an optional room type value—e.g., Living Room—through the Room Classification Branch (bottom). The room type is used for extracting the mask \\( M_{\\mathrm{room}} \\), as visualized on the bottom right." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.624, + 0.483, + 0.76 + ], + "angle": 0, + "content": "Semantic Rays Prediction. Unlike the continuous depth values estimated in prior work, the semantic rays should correspond to semantic categories, which are represented as a set of discrete classes. Therefore, we construct a network that produces a semantic ray representation \\(\\hat{r}_s \\in \\{1, \\dots, C\\}^l\\) from the image, where each ray is classified into one of \\(C\\) semantic categories. We provide an overview of our semantic ray prediction network in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.762, + 0.483, + 0.913 + ], + "angle": 0, + "content": "As illustrated in the figure, our semantic network architecture leverages a pretrained ResNet50 backbone to extract robust, high-level features from an input RGB image \\(I\\). After reducing the feature channels using a CNN and projecting them into a lower-dimensional subspace, positional encodings are computed to preserve spatial information. Two sets of learnable tokens are introduced: a set of \\(l\\) ray tokens responsible for predicting the semantic ray representation \\(\\hat{r}_s\\) and a single (optional) CLS token dedicated for representing global room clas" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.39, + 0.664, + 0.404 + ], + "angle": 0, + "content": "sification information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.409, + 0.885, + 0.591 + ], + "angle": 0, + "content": "A single-head cross attention module integrates these tokens with the flattened spatial features, yielding refined tokens that capture both global context and local details. In the ray branch, the refined ray tokens are first processed by a self-attention block that enables each token to interact with all others, thereby aggregating complementary contextual information. The enriched tokens are then passed through an MLP to produce per-token semantic logits, which after normalization form the final semantic ray vector \\(\\hat{r}_s\\). If room labels are available in the dataset, a similar network processes the CLS token for room type prediction, as we further detail later." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.884, + 0.913 + ], + "angle": 0, + "content": "Estimating Semantic Probability Volume. To obtain the semantic probability map, \\( P_{s} \\in [0,1]^{[\\hat{H},\\hat{W},O]} \\), we first need to interpolate the \\( l \\) predicted semantic rays. Regular linear interpolation—which prior work used for depth estimation—is unsuitable in the context of discrete labeling since interpolating between class labels can produce non-valid or semantically meaningless results. Instead, we propose a voting-based interpolation scheme: We reduce the original equiangular rays to the desired count by applying a majority vote within a small neighborhood. We use a window of three rays, assigning the label that appears most frequently in that window to the center target ray; see the supplementary material for the full algorithm. Next, we compute the score for each set of rays by taking the \\( L_{1} \\) difference between the predicted semantic labels and the reference labels. The score is then exponentiated and normalized to form the semantic probability volume \\( P_{s} \\), which quantifies the likelihood of each candidate pose based on the alignment between the semantic rays and the candidate pose." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.952, + 0.519, + 0.963 + ], + "angle": 0, + "content": "27119" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.112, + 0.092, + 0.485, + 0.304 + ], + "angle": 0, + "content": "Room Type Prediction. In addition to predicting semantic rays, our semantic network can optionally also predict the room type, which corresponds to the room from which the input image was taken. This is achieved by processing the CLS token in the semantic network (see Figure 3). If the predicted room probability exceeds a threshold \\( T_{\\mathrm{room}} \\), the predicted room type is then used to construct a mask \\( M_{\\mathrm{room}} \\) from the polygons associated with that room type. For example, if the model predicts a Living Room label with high confidence, the mask \\( M_{\\mathrm{room}} \\) retains only the regions corresponding to the living room by setting all other areas to zero. Let \\( \\tilde{P} \\) denote the original probability volume, then the masked probability volume \\( P \\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.239, + 0.317, + 0.359, + 0.334 + ], + "angle": 0, + "content": "\\[\nP = M _ {\\mathrm {r o o m}} \\odot \\tilde {P},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.341, + 0.483, + 0.432 + ], + "angle": 0, + "content": "where \\(\\odot\\) denotes element-wise multiplication, thereby filtering out all the probabilities which are not in the living room and substantially narrowing down the search space. A detailed analysis of room type distributions and the model's prediction accuracy is provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.44, + 0.484, + 0.471 + ], + "angle": 0, + "content": "3.3. Floorplan Localization via a Structural-Semantic Probability Volume" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.477, + 0.484, + 0.569 + ], + "angle": 0, + "content": "We obtain the final probability map by generating the depth probability map \\( P_{d} \\), following the procedure described in Section 3.1, and the semantic probability map \\( P_{s} \\), according to the approach detailed in Section 3.2. To leverage both semantic and geometric cues, we fuse the two probability maps using a weighted combination:" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.58, + 0.382, + 0.597 + ], + "angle": 0, + "content": "\\[\nP _ {c} = w _ {s} \\cdot P _ {s} + w _ {d} \\cdot P _ {d},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.607, + 0.484, + 0.727 + ], + "angle": 0, + "content": "where \\( w_{s} \\) denotes the weight given to the semantic cues, while \\( w_{d} = 1 - w_{s} \\) represents the weight of the depth cues. These weights are determined using a held-out validation set, as further detailed in Section 4. Note that as both \\( P_{s} \\) and \\( P_{d} \\) are generated from interpolated rays, which we refer to as low-resolution rays (see Figure 4, predicted), and hence the initial probability volume \\( P_{c} \\) is also in low resolution." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.735, + 0.484, + 0.886 + ], + "angle": 0, + "content": "Location Extraction. Our approach follows a coarse-to-fine strategy to achieve precise localization while maintaining computational efficiency. Given \\(P_{c}\\), we first extract the Top-\\(k\\) candidate poses from the structural-semantic probability volume based on their scores, ensuring that each candidate is separated by at least \\(\\delta_{\\mathrm{res}}\\) in translation. Next, for each candidate, we generate an augmented set of orientations by perturbing its original angle in increments of \\(\\pm \\delta_{\\mathrm{ang}}\\) up to a maximum deviation of \\(\\Delta_{\\mathrm{max}}\\), resulting in the final augmented set:" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.897, + 0.402, + 0.914 + ], + "angle": 0, + "content": "\\[\n[ 0, \\pm \\delta_ {\\text {a n g}}, \\pm 2 \\delta_ {\\text {a n g}}, \\dots , \\pm \\Delta_ {\\max} ].\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.091, + 0.882, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.212, + 0.885, + 0.351 + ], + "angle": 0, + "content": "Figure 4. Comparing low-resolution and high-resolution rays with our coarse-to-fine approach. Given an input image (left), we construct a structural-semantic probability volume by comparing low resolution ground-truth and predicted rays (center). Location extraction from this coarse volume directly often yields significant errors, as illustrated with the yellow arrows. Results are refined only for the Top-k candidate poses by comparing the high resolution ground-truth and predicted rays (right). This allows for efficiently extracting more accurate predictions, as further detailed in Section 3.3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.379, + 0.884, + 0.528 + ], + "angle": 0, + "content": "At each candidate location, we compute the corresponding high-resolution ground-truth depth and semantic ray representations, yielding \\( l \\) rays per candidate location. These ground-truth rays are then compared against the original predicted \\( l \\) rays using a similarity metric (see supplementary material for metric details). For a visual comparison of the high-resolution ground-truth and predicted rays, please refer to Figure 4. The candidate with the highest similarity score is selected as the final predicted location." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.531, + 0.884, + 0.757 + ], + "angle": 0, + "content": "This coarse-to-fine refinement process effectively leverages the full resolution of the predictions, which were initially interpolated for runtime efficiency, to achieve more precise localization. Note that there is a tradeoff between accuracy and computation time: finer resolutions yield improved precision at the expense of increased computational load. This tradeoff is illustrated in Figure 4. As can be observed, comparing low resolution rays often yields inaccurate predictions. For instance, environments with multiple semantic objects can suffer from loss of critical details (e.g., the left door is omitted in the middle row) in the low-resolution prediction, which can lead to misclassification, further motivating the refinement step. We present a quantitative analysis of this tradeoff in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.769, + 0.836, + 0.785 + ], + "angle": 0, + "content": "3.4. Training and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.792, + 0.885, + 0.913 + ], + "angle": 0, + "content": "Both networks are trained in an end-to-end manner. For depth prediction, an \\( L_{1} \\) loss supervises the predicted depths against ground-truth depth maps. For semantic prediction, a cross-entropy loss is applied to the predicted semantic labels. If room labels \\( R \\) are available in the dataset, an additional cross-entropy loss is used for the room label, and the network is trained to optimize both objectives jointly. As in prior work, data aug" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.952, + 0.521, + 0.964 + ], + "angle": 0, + "content": "27120" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.112, + 0.092, + 0.482, + 0.244 + ], + "angle": 0, + "content": "mentation techniques—including virtual roll-pitch augmentation—are employed to improve robustness to nonupright camera poses. The networks are optimized using the Adam optimizer with an initial learning rate of \\(1 \\times 10^{-3}\\). We use a floorplan resolution of \\(0.1\\mathrm{m}\\) and an angular granularity of \\(10^{\\circ}\\). We set the number of predicted rays to \\(l = 40\\) and interpolate these to 7 rays during the coarse stage of localization. For the Location Extraction module, we use \\(\\delta_{\\mathrm{res}} = 0.1\\mathrm{m}\\), \\(\\delta_{\\mathrm{ang}} = 5^{\\circ}\\), and \\(\\Delta_{\\mathrm{max}} = 5^{\\circ}\\), and refine our results using Top-5 poses." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.257, + 0.2, + 0.273 + ], + "angle": 0, + "content": "4. Results" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.283, + 0.482, + 0.48 + ], + "angle": 0, + "content": "In this section, we present a comprehensive evaluation of our localization method. We begin by introducing the datasets we use in our experiments, followed by a discussion of the baselines we compare against and the evaluation metrics. The main results are reported in Section 4.1. We conduct an ablation study to assess the contributions of various components of our approach in Section 4.2. We report results for our approach under two conditions: one in which room labels are not utilized (denoted as \\( \\mathrm{Ours}_s \\)) and another in which room labels are employed to further refine the predictions (denoted as \\( \\mathrm{Ours}_r \\)). Additional experiments and qualitative results are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.488, + 0.482, + 0.639 + ], + "angle": 0, + "content": "Datasets. We conduct experiments on two popular datasets: Structured 3D (S3D) [37] and ZInD [9]. S3D is a synthetic dataset containing realistic 3D renders of 3,296 houses. We use the fully furnished version of S3D, as employed in previous works. ZInD consists of 1,575 unfurnished homes containing only panoramic images. We crop these panoramas to perspective views with an \\(80^{\\circ}\\) field of view (as was done in S3D), and generate a fixed-size dataset from the resulting images. For both datasets, we follow their official train and test splits." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.647, + 0.482, + 0.86 + ], + "angle": 0, + "content": "Baselines. We compare our approach against two baselines: F3Loc [7] (considering the single-image localization component) and LASER [22]. We also report performance using an ORACLE ray prediction. This oracle ray prediction simulates the best possible performance achievable by our pipeline using ground truth depth and semantic rays. Note that the Oracle ray prediction baseline does not incorporate room-aware features. For F3Loc, results on the ZIND dataset are obtained from our experiments by running their publicly available code on the dataset, as the original paper does not include results on this dataset. We also use the publicly available implementation of LASER. Additional details are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.867, + 0.482, + 0.913 + ], + "angle": 0, + "content": "Metrics. Following prior work [7, 22], we report recall metrics computed at distance thresholds of \\(0.1\\mathrm{m}\\), \\(0.5\\mathrm{m}\\), and \\(1\\mathrm{m}\\). We also report recall for predictions with an" + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.089, + 0.86, + 0.34 + ], + "angle": 0, + "content": "
S3D R@
Method0.1m0.5m1m1m 30°
LASER0.76.410.48.7
F3Loc1.514.622.421.3
Ourss5.4241.8753.5252.61
Oursr5.7045.5358.7857.49
ORACLE61.0093.8494.8794.57
ZInD R@
Method0.1m0.5m1m1m 30°
LASER1.3811.0617.5513.64
F3Loc0.677.9015.0711.46
Ourss2.9824.0033.9629.30
Oursr3.3126.6038.0131.86
ORACLE26.4260.8567.6965.13
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.343, + 0.884, + 0.414 + ], + "angle": 0, + "content": "Table 1. Recall performance on the S3D and ZInD datasets. The table reports recall at thresholds of \\(0.1\\mathrm{m}\\), \\(0.5\\mathrm{m}\\), \\(1\\mathrm{m}\\), and \\(1\\mathrm{m}\\) with a \\(30^{\\circ}\\) orientation tolerance for LASER, F3Loc, our approach without room labels \\((\\mathrm{Ours}_s)\\), our approach with room labels \\((\\mathrm{Ours}_r)\\), and an Oracle ray prediction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.884, + 0.486 + ], + "angle": 0, + "content": "orientation error bounded to less than \\(30^{\\circ}\\) (at the \\(1\\mathrm{m}\\) threshold). Recall is defined as the percentage of predictions that fall within these thresholds." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.497, + 0.735, + 0.513 + ], + "angle": 0, + "content": "4.1.Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.519, + 0.884, + 0.716 + ], + "angle": 0, + "content": "Results are reported in Table 1. On both S3D and ZinD, our method more than doubles F3Loc's and LASER performance across all thresholds. Notably, considering S3D over the \\(\\mathrm{R@1m30^{\\circ}}\\) metric—which reflects the quality of matches between the predicted and actual camera views—in comparison to F3Loc, our method improves by more than three times. Room type predictions yield improvements of \\(9.2\\%\\) in \\(\\mathrm{R@1m30^{\\circ}}\\) on the S3D dataset and \\(8.7\\%\\) on the ZInD dataset. We include an additional experiment in the supplementary material that evaluates F3Loc with our refinement module, demonstrating that both the semantic rays and our refinement module provide significant performance gains." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.716, + 0.883, + 0.821 + ], + "angle": 0, + "content": "We observe that significant performance improvements are also achieved for a very fine-grained recall metric of \\(0.1\\mathrm{m}\\), boosting performance from \\(1.5\\%\\) (F3Loc) to \\(5.7\\%\\) with our approach on S3D. We attribute this to our coarse-to-fine strategy, which effectively refines the coarse location estimates into precise predictions, as further validated in our ablation study." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.822, + 0.884, + 0.913 + ], + "angle": 0, + "content": "Figure 5 presents qualitative examples that illustrate how the integration of floorplan semantics with precise depth cues enables our pipeline to effectively resolve localization ambiguities. For instance, in the third row, we can see that F3Loc, which does not use semantics, interprets this image as a blank wall, while LASER misinter" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.952, + 0.518, + 0.963 + ], + "angle": 0, + "content": "27121" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.087, + 0.48, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.112, + 0.299, + 0.484, + 0.368 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison of our method with F3Loc and LASER. Warmer colors correspond to regions with higher probabilities. Below each map we report the localization error in meters and degrees. We use arrows to visualize the ground truth location (magenta) and the predicted location (white)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.396, + 0.482, + 0.412 + ], + "angle": 0, + "content": "prets the window size and predicts an incorrect location." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.421, + 0.266, + 0.437 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.444, + 0.483, + 0.7 + ], + "angle": 0, + "content": "We demonstrate the effect of incorporating each component of our method on overall localization performance. Specifically, we conduct the following ablations: (1) Base, which corresponds to computing the structural-semantic probability volume and selecting the maximum probability without any additional refinement. (2) Removing semantic interpolation (denoted as -Interpolation), where we replace our majority voting interpolation with a simple linear interpolation followed by rounding. (3) Adding room predictions (denoted as +Room), where we assess the effect of integrating room type predictions into our localization pipeline. (4) Adding refinement (denoted as +Refine), which assesses our coarse-to-fine approach, which refines our localization extraction via the Top-K candidate poses. (5) Adding room predictions and refinement (denoted as +Room&Refine), which combines both components." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.701, + 0.484, + 0.913 + ], + "angle": 0, + "content": "From the results in Table 2, we can see that on the \\(1\\mathrm{m}30^{\\circ}\\) metric, the addition of our refinement module improves performance by \\(8.6\\%\\) relative to the baseline. This indicates that a substantial amount of information is lost during the initial interpolation process if not properly addressed, thereby strongly motivating the use of coarse-to-fine strategies in our Location Extraction module. We also observe a significant improvement of \\(11.5\\%\\) from the room prediction component, which is discussed in further detail in the supplementary material and visualized in Figure 6. Finally, our majority voting interpolation approach contributes an additional gain of \\(0.9\\%\\) compared to a simple interpolation strategy. When combining all these improvements, our method achieves" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.088, + 0.883, + 0.184 + ], + "angle": 0, + "content": "
Method0.1m0.5m1m1m 30°
Base4.6538.3549.4048.44
- Interpolation4.7338.4448.9147.99
+ Room5.1242.9255.5754.04
+ Refine5.4241.8753.5252.61
+ Room&Refine5.7045.5358.7857.49
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.185, + 0.883, + 0.214 + ], + "angle": 0, + "content": "Table 2. Ablation study, evaluating the effect of incorporating the different components in our pipeline on the S3D dataset." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.23, + 0.879, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.373, + 0.884, + 0.456 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison of using room predictions and our coarse-to-fine refinement approach. Below each map we report the localization error in meters and degrees. Warmer colors correspond to regions with higher probabilities. Overlaid on the estimated probabilities, we show the ground truth location (magenta) and the predicted location." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.883, + 0.514 + ], + "angle": 0, + "content": "an overall enhancement of \\(18.6\\%\\) on the 1m \\(30^{\\circ}\\) metric relative to our base model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.517, + 0.884, + 0.607 + ], + "angle": 0, + "content": "In Figure 6, we illustrate the impact of incorporating room type predictions alongside the location extraction module. The figure clearly demonstrates how these components refine the probability volume by narrowing down the candidate poses, which in turn improves overall localization accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.884, + 0.745 + ], + "angle": 0, + "content": "We analyze the impact of different combinations of predicted depth and semantic features on the S3D dataset. Figure 7 summarizes the recall performance for various depth and semantic weight configurations used to compute the structural-semantic probability volume. In Figure 8 we visualize examples from two extreme scenarios (depth only and semantic only) and from the configuration adopted in our work \\(\\left([w_{s}, w_{d}] = [0.4, 0.6]\\right)\\), selected according to the validation set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.747, + 0.885, + 0.911 + ], + "angle": 0, + "content": "As can be observed from Figure 8 (and also reflected in Figure 7), relying solely on semantic rays tends to produce a more diffused probability volume with multiple ambiguous candidate locations. This occurs because most images contain a single semantic object of a standard size, allowing an accurate ray pattern to be fitted in more than one location. Similarly, depth cues alone also suffer from ambiguity, particularly in repetitive environments or when the image captures three or fewer walls, which leads to uncertainty in the localization estimate. However, when these semantic cues are com" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.952, + 0.519, + 0.963 + ], + "angle": 0, + "content": "27122" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.089, + 0.482, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.112, + 0.303, + 0.484, + 0.387 + ], + "angle": 0, + "content": "Figure 7. Recall vs. weight combinations on the S3D test set. The plot shows the recall metrics for four different thresholds: \\(0.1m\\), \\(0.5m\\), \\(1m\\), and \\(1m30^{\\circ}\\). The x-axis displays the weight combinations in the order \\((w_{s}, w_{d})\\). A vertical dashed line at \\(w_{s} = 0.4\\), \\(w_{d} = 0.6\\) highlights the weight combination selected using the validation set." + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.401, + 0.48, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.112, + 0.619, + 0.484, + 0.687 + ], + "angle": 0, + "content": "Figure 8. Qualitative Comparison of Depth-Only, Semantic-Only, and Fused Structural-Semantic Probability Volumes. Below each map we report the localization error in meters and degrees. Warmer colors correspond to regions with higher probabilities." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.716, + 0.482, + 0.776 + ], + "angle": 0, + "content": "bined with depth rays, the resulting probability volume becomes significantly more concentrated. This integration effectively filters out spurious candidates and sharpens the localization estimate." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.778, + 0.482, + 0.853 + ], + "angle": 0, + "content": "Additionally, the supplementary material offers a detailed analysis of how varying the Top-K candidates influences localization refinement, along with additional experiments that motivate various design choices, such as using hard thresholds in multiple steps of our pipeline." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.863, + 0.242, + 0.878 + ], + "angle": 0, + "content": "Runtime Analysis" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.882, + 0.482, + 0.913 + ], + "angle": 0, + "content": "We report runtime performance breakdown in Table 3, using the same parameters employed in Table 1 with" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.09, + 0.88, + 0.172 + ], + "angle": 0, + "content": "
KPredictionLocRefinTotal
10.038 ± 0.1190.174 ± 0.0330.141 ± 0.0730.364 ± 0.142
30.033 ± 0.0930.154 ± 0.0260.356 ± 0.1190.554 ± 0.157
50.034 ± 0.1030.155 ± 0.0290.577 ± 0.1850.778 ± 0.218
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.182, + 0.884, + 0.253 + ], + "angle": 0, + "content": "Table 3. Performance breakdown over different Top-K values. Each entry is mean ± std (s). Prediction denotes the ray predictions, Loc refers to the localization process, and Refin represents the refinement stage, during which candidate locations are evaluated by computing the ground truth rays." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.281, + 0.884, + 0.506 + ], + "angle": 0, + "content": "varying K values. As shown in the table, the per-image inference time increases with the number of candidate refinements, \\(K\\), reflecting the additional computations required during refinement. All experiments were conducted on a single CPU without multithreading to avoid introducing bias. Future work could explore parallelizing the refinement stage by computing all ground-truth rays simultaneously for further speed-up. Importantly, even at \\(K = 5\\) the setting we adopt in our work, the inference time remains reasonable, striking a practical balance between computational cost and improvements in localization accuracy. We further observe that the prediction and localization steps require similar amounts of time, while the refinement step grows monotonically as \\(K\\) increases." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.523, + 0.633, + 0.538 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.549, + 0.884, + 0.685 + ], + "angle": 0, + "content": "In this work, we presented a semantic-aware localization framework that extends floorplan-based camera localization by fusing semantic labels with geometric depth cues. Our approach leverages a novel semantic ray prediction network alongside an established depth estimation method to generate a semantic-structural probability volume, which significantly improves localization accuracy, especially in environments with repetitive or ambiguous structural patterns." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.686, + 0.883, + 0.806 + ], + "angle": 0, + "content": "Our extensive experiments on the S3D and ZInD datasets demonstrate that integrating semantic cues effectively resolves depth-based ambiguities and consistently outperforms state-of-the-art methods such as F3Loc and LASER. Ablation studies confirm that a balanced combination of depth and semantic information, coupled with a coarse-to-fine localization strategy and the use of room labels, yields optimal performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.807, + 0.884, + 0.913 + ], + "angle": 0, + "content": "Looking forward, extending our framework to incorporate additional semantic labels and other modalities, such as textual information, promises to further enhance localization robustness in challenging indoor settings. In general, our approach represents an important step towards accurate and reliable indoor localization systems by effectively leveraging semantic and geometric cues." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.952, + 0.518, + 0.963 + ], + "angle": 0, + "content": "27123" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.09, + 0.21, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.115, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Relja Arandjelovic, Petr Gronat, Akihiko Torii, Tomas Pajdla, and Josef Sivic. NetVLAD: CNN architecture for weakly supervised place recognition. In CVPR, pages 5297-5307, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.173, + 0.484, + 0.214 + ], + "angle": 0, + "content": "[2] Vassileios Balntas, Shuda Li, and Victor Prisacariu. Relocnet: Continuous metric learning relocalisation using neural nets. In ECCV, pages 751-767, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.217, + 0.484, + 0.285 + ], + "angle": 0, + "content": "[3] Federico Boniardi, Tim Caselitz, Rainer Kummerle, and Wolfram Burgard. Robust lidar-based localization in architectural floor plans. In 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3318-3324. IEEE, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.287, + 0.484, + 0.356 + ], + "angle": 0, + "content": "[4] Federico Boniardi, Tim Caselitz, Rainer Kummerle, and Wolfram Burgard. A pose graph-based localization system for long-term navigation in cad floor plans. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 84-97. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.358, + 0.484, + 0.44 + ], + "angle": 0, + "content": "[5] Federico Boniardi, Abhinav Valada, Rohit Mohan, Tim Caselitz, and Wolfram Burgard. Robot localization in floor plans using a room layout edge extraction network. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5291-5297. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.443, + 0.484, + 0.511 + ], + "angle": 0, + "content": "[6] Eric Brachmann, Alexander Krull, Sebastian Nowozin, Jamie Shotton, Frank Michel, Stefan Gumhold, and Carsten Rother. DSAC-differentiable RANSAC for camera localization. In CVPR, pages 6684-6692, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.514, + 0.484, + 0.568 + ], + "angle": 0, + "content": "[7] Changan Chen, Rui Wang, Christoph Vogel, and Marc Pollefeys. F3Loc: Fusion and filtering for floorplan localization. arXiv preprint arXiv:2403.03370, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.571, + 0.484, + 0.613 + ], + "angle": 0, + "content": "[8] Hang Chu, Dong Ki Kim, and Tsuhan Chen. You are here: Mimicking the human thinking process in reading floorplans. In ICCV, pages 2210-2218, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.615, + 0.484, + 0.682 + ], + "angle": 0, + "content": "[9] Steve Cruz, Will Hutchcroft, Yuguang Li, Naji Khosravan, Ivaylo Boyadzhiev, and Sing Bing Kang. Zillow indoor dataset: Annotated floor plans with \\(360^{\\circ}\\) panoramas and 3d room layouts. In CVPR, pages 2133-2143, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.686, + 0.484, + 0.74 + ], + "angle": 0, + "content": "[10] Niklas Gard, Anna Hilsmann, and Peter Eisert. Spvloc: Semantic panoramic viewpoint matching for 6d camera localization in unseen environments. arXiv preprint arXiv:2404.10527, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.743, + 0.484, + 0.797 + ], + "angle": 0, + "content": "[11] B. Glocker, S. Izadi, J. Shotton, and A. Criminisi. Realtime RGB-D camera relocalization. In Proceedings of the Mixed and Augmented Reality (ISMAR) Conference. IEEE, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.8, + 0.484, + 0.854 + ], + "angle": 0, + "content": "[12] Henry Howard-Jenkins and Victor Adrian Prisacariu. LaLaLoc++: Global floor plan comprehension for layout localisation in unvisited environments. In ECCV, pages 693-709, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.857, + 0.484, + 0.913 + ], + "angle": 0, + "content": "[13] Henry Howard-Jenkins, Jose-Raul Ruiz-Sarmiento, and Victor Adrian Prisacariu. LaLaLoc: Latent layout localisation in dynamic, unvisited environments. arXiv preprint arXiv:2104.09169, 2021. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.115, + 0.115, + 0.484, + 0.913 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.885, + 0.175 + ], + "angle": 0, + "content": "[14] Seigo Ito, Felix Endres, Markus Kuderer, Gian Diego Tipaldi, Cyril Stachniss, and Wolfram Burgard. W-RGBD: Floor-plan-based indoor global localization using a depth camera and wifi. In 2014 IEEE International Conference on Robotics and Automation (ICRA), pages 417-422. IEEE, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.885, + 0.232 + ], + "angle": 0, + "content": "[15] Peter Karkus, David Hsu, and Wee Sun Lee. Particle filter networks with application to visual localization. In Proceedings of the 2nd Conference on Robot Learning (CoRL), pages 169-178. PMLR, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.885, + 0.285 + ], + "angle": 0, + "content": "[16] Alex Kendall, Matthew Grimes, and Roberto Cipolla. PoseNet: A convolutional network for real-time 6-DOF camera relocalization. In ICCV, pages 2938-2946, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.287, + 0.885, + 0.342 + ], + "angle": 0, + "content": "[17] Junho Kim, Jiwon Jeong, and Young Min Kim. Fully geometric panoramic localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.343, + 0.885, + 0.411 + ], + "angle": 0, + "content": "[18] Z. Li, M. H. Ang, and D. Rus. Online localization with imprecise floor space maps using stochastic gradient descent. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8571-8578. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.413, + 0.885, + 0.453 + ], + "angle": 0, + "content": "[19] Liu Liu, Hongdong Li, and Yuchao Dai. Efficient global 2D-3D matching for camera localization in a large-scale 3D map. In ICCV, pages 2372-2381, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.455, + 0.885, + 0.509 + ], + "angle": 0, + "content": "[20] Pranay Mathur, Rajesh Kumar, and Sarthak Upadhyay. Sparse image-based navigation architecture to mitigate the need of precise localization in mobile robots. arXiv preprint arXiv:2203.15272, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.885, + 0.564 + ], + "angle": 0, + "content": "[21] Oscar Mendez, Simon Hadfield, Nicolas Pugeault, and Richard Bowden. SeDAR: Reading floorplans like a human—using deep learning to enable human-inspired localisation. *IJCV*, 128:1286–1310, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.566, + 0.885, + 0.633 + ], + "angle": 0, + "content": "[22] Zhixiang Min, Naji Khosravan, Zachary Bessinger, Manjunath Narayana, Sing Bing Kang, Enrique Dunn, and Ivaylo Boyadzhiev. LASER: Latent space rendering for 2D visual localization. In CVPR, pages 11122-11131, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.635, + 0.885, + 0.717 + ], + "angle": 0, + "content": "[23] R. A. Newcombe, A. J. Davison, S. Izadi, P. Kohli, O. Hilliges, J. Shotton, D. Molyneaux, S. Hodges, D. Kim, and A. Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In Proceedings of the Mixed and Augmented Reality (ISMAR) Conference. IEEE, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.719, + 0.885, + 0.772 + ], + "angle": 0, + "content": "[24] Takahiro Niwa, Shun Taguchi, and Noriaki Hirose. Spatio-temporal graph localization networks for image-based navigation. arXiv preprint arXiv:2204.13237, 2022.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.885, + 0.828 + ], + "angle": 0, + "content": "[25] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: robust hierarchical localization at large scale. In CVPR, pages 12716-12725, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.829, + 0.885, + 0.87 + ], + "angle": 0, + "content": "[26] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Fast image-based localization using direct 2d-to-3d matching. In ICCV, pages 667-674, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.871, + 0.885, + 0.913 + ], + "angle": 0, + "content": "[27] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Improving image-based localization by active correspondence search. In ECCV, pages 752-765, 2012." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.885, + 0.913 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.952, + 0.52, + 0.964 + ], + "angle": 0, + "content": "27124" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.092, + 0.484, + 0.147 + ], + "angle": 0, + "content": "[28] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Efficient & effective prioritized matching for large-scale image-based localization. IEEE TPAMI, 39(9):1744-1756, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.149, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[29] Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene coordinate regression forests for camera relocalization in rgb-d images. In CVPR, pages 2930-2937, 2013. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.207, + 0.483, + 0.261 + ], + "angle": 0, + "content": "[30] Janine Thoma, Danda Pani Paudel, Ajad Chhatkuli, Thomas Probst, and Luc Van Gool. Mapping, localization and path planning for image-based navigation using visual features and map. In CVPR. IEEE, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.263, + 0.482, + 0.33 + ], + "angle": 0, + "content": "[31] Julien Valentin, Matthias Nießner, Jamie Shotton, Andrew Fitzgibbon, Shahram Izadi, and Philip HS Torr. Exploiting uncertainty in regression forests for accurate camera relocalization. In CVPR, pages 4400-4408, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.333, + 0.483, + 0.389 + ], + "angle": 0, + "content": "[32] Florian Walch, Caner Hazirbas, Laura Leal-Taixe, Torsten Sattler, Sebastian Hilsenbeck, and Daniel Cremers. Image-based localization using lstms for structured feature correlation. In ICCV, pages 627-637, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.391, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[33] Shenlong Wang, Sanja Fidler, and Raquel Urtasun. Lost shopping! monocular localization in large indoor spaces. In ICCV, pages 2695-2703, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.433, + 0.483, + 0.5 + ], + "angle": 0, + "content": "[34] Xipeng Wang, Ryan J Marcotte, and Edwin Olson. GLFP: Global localization from a floor plan. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1627-1632. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.503, + 0.483, + 0.585 + ], + "angle": 0, + "content": "[35] Wera Winterhalter, Freya Fleckenstein, Bastian Steder, Luciano Spinello, and Wolfram Burgard. Accurate indoor localization for RGB-D smartphones and tablets given 2d floor plans. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3138-3143. IEEE, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.588, + 0.483, + 0.642 + ], + "angle": 0, + "content": "[36] Jian Wu, Liwei Ma, and Xiaolin Hu. Delving deeper into convolutional neural networks for camera relocalization. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 5644-5651. IEEE, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.644, + 0.483, + 0.697 + ], + "angle": 0, + "content": "[37] Jia Zheng, Junfei Zhang, Jing Li, Rui Tang, Shenghua Gao, and Zihan Zhou. Structured3d: A large photorealistic dataset for structured 3d modeling. In ECCV, 2020.6" + }, + { + "type": "list", + "bbox": [ + 0.115, + 0.092, + 0.484, + 0.697 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.952, + 0.519, + 0.963 + ], + "angle": 0, + "content": "27125" + } + ] +] \ No newline at end of file diff --git a/2025/Supercharging Floorplan Localization with Semantic Rays/834ff1eb-78d1-464a-8d20-1e5cbed96504_origin.pdf b/2025/Supercharging Floorplan Localization with Semantic Rays/834ff1eb-78d1-464a-8d20-1e5cbed96504_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d940c011290dd331f664ff6e7739148112f08e33 --- /dev/null +++ b/2025/Supercharging Floorplan Localization with Semantic Rays/834ff1eb-78d1-464a-8d20-1e5cbed96504_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1dacdf38337fec0d5ef9ee8c1ccd4184db684ed4dec4f29aab275dec93792593 +size 21722016 diff --git a/2025/Supercharging Floorplan Localization with Semantic Rays/full.md b/2025/Supercharging Floorplan Localization with Semantic Rays/full.md new file mode 100644 index 0000000000000000000000000000000000000000..36a700a64af7945a5439edfee1373cb65c6815e2 --- /dev/null +++ b/2025/Supercharging Floorplan Localization with Semantic Rays/full.md @@ -0,0 +1,264 @@ +# Supercharging Floorplan Localization with Semantic Rays + +Yuval Grader +Tel Aviv University + +Hadar Averbuch-Elor Cornell University + +https://tau-vailab.github.io/SemRayLoc/ + +# Abstract + +Floorplans provide a compact representation of the building's structure, revealing not only layout information but also detailed semantics such as the locations of windows and doors. However, contemporary floorplan localization techniques mostly focus on matching depth-based structural cues, ignoring the rich semantics communicated within floorplans. In this work, we introduce a semantic-aware localization framework that jointly estimates depth and semantic rays, consolidating over both for predicting a structural-semantic probability volume. Our probability volume is constructed in a coarse-to-fine manner: We first sample a small set of rays to obtain an initial low-resolution probability volume. We then refine these probabilities by performing a denser sampling only in high-probability regions and process the refined values for predicting a 2D location and orientation angle. We conduct an evaluation on two standard floorplan localization benchmarks. Our experiments demonstrate that our approach substantially outperforms state-of-the-art methods, achieving significant improvements in recall metrics compared to prior works. Moreover, we show that our framework can easily incorporate additional metadata such as room labels, enabling additional gains in both accuracy and efficiency. + +# 1. Introduction + +Camera localization is a longstanding problem in computer vision, with significant applications in 3D reconstruction [19, 25-28], augmented reality [11, 16, 23, 29, 31], and navigation [20, 24, 30, 32, 36]. Localization within indoor environments is especially challenging due to the absence of reliable GPS signals and the complexity of reasoning across multiple floors and layers. Hence, to bypass complicated 3D model-based localization techniques, prior work [7, 8, 12, 13, 15, 35] has explored the problem of localizing camera observations within a provided 2D floorplan map by matching + +![](images/77ef471a7a1601e41515129379e2e22703659419519e68e969f5e6eddefc7a41.jpg) + +![](images/858b3aaecf03cba21dc1abcc2b6f98c495992a3f6724f89654ed9f55cb160005.jpg) + +![](images/6cd75802e283904d983688ac00f6389d1e1bddeecdf915c83ef41fca98fe8618.jpg) +Input & GT + +![](images/14fc60037bf8e390c93cfefa04e4bacfe3e63beec82af7bef376240fe45a3769.jpg) + +![](images/1c8b5dad1f5d9f5c90b4e1ea7987df7e3438757c72b5d6306ba7ed55a817c077.jpg) +Output Localization +Figure 1. Floorplan localization using a raw binary floorplan (middle row) often yields ambiguous predictions. In this work, we utilize a richer, yet readily available, representation: a floorplan enhanced with semantic labels (bottom row). We present an approach that supercharges floorplan localization with semantic rays, enabling for resolving localization ambiguities, as illustrated by the comparison on the right. + +depth-based structural cues. + +However, while floorplans offer a compact and lightweight scene representation, structural cues within floorplans often correlate with multiple candidate locations, particularly for environments with repetitive or symmetric layouts. Consider the example in Figure 1. Can you localize the input image within the floorplan? Provided with just a raw (walls only) floorplan, room corners are indistinguishable and hence localization is + +highly ambiguous, as can also be observed by the probabilities predicted by the state-of-the-art F3Loc [7] technique (middle row, right). To resolve such ambiguities, we are interested in utilizing a slightly different representation: a semantics-aware floorplan, such as the one illustrated on the bottom row of Figure 1. + +Accordingly, we introduce a semantic-aware ray-based localization framework that integrates semantic cues with depth-based predictions. In particular, we propose a semantic prediction network that predicts accurate semantic ray representations along with optional additional metadata (such as room labels) from a single RGB image with a limited field-of-view. We process these rays to compute a semantic probability volume, which is then fused with depth information for constructing a structural-semantic probability volume. + +Our framework follows a coarse-to-fine localization strategy. We first operate over a low-resolution image for an efficient floorplan search. This initial search yields the Top- $k$ candidate locations. Finally, we refine these candidates and select the best match using a high-resolution ray representation. By adopting this coarse-to-fine approach, our method effectively constrains the localization search to the most promising regions while keeping the computation cost feasible. + +Our experiments demonstrate that our approach yields improvements by factors ranging from two to three across most metrics compared to the state-of-the-art technique [7], upon which our method is built. This substantial gain underscores the effectiveness of fusing semantic and depth ray predictions into a unified probability volume. We further show that our coarse-to-fine strategy offers a flexible tradeoff between accuracy and computational cost, with performance consistently improving as larger candidate sets (Top- $k$ ) are evaluated—making our method adaptable to diverse task requirements. Moreover, incorporating additional metadata further enhances precision, leading to significantly improved localization accuracies. + +Explicitly stated, our contributions are: + +- We introduce a semantic ray prediction network that receives a single RGB image as input. +- We propose an efficient and unified framework that fuses semantic and depth ray predictions into a structural-semantic probability volume, which effectively resolves localization ambiguities. +- Results that demonstrate that significant improvements over state-of-the-art methods. + +# 2. Related Work + +Visual Localization. The task of visual localization has received ongoing attention throughout the past several decades. Traditional approaches often rely on image re + +trieval or on a 3D Structure-from-Motion (SfM) model of the environment. In the image retrieval paradigm, methods such as NetVLAD [1] or RelocNet [2] compare a query image against a database of labeled images. Once the closest match is found, the query pose is approximated by the retrieved image's pose. Other methods explicitly construct a 3D SfM model of a scene to establish 2D-3D correspondences between the query image and the reconstructed 3D structure [19, 25-28]. After matching local image descriptors to 3D points, robust solvers estimate the 6-DoF pose. + +Recent learning-based pipelines deviate from classical 2D-3D matching. Scene-coordinate regression methods predict dense 3D coordinates for every pixel in the query image [6, 29, 31], whereas pose regression methods directly estimate the 6-DoF camera pose via neural networks [16, 32, 36]. Although promising for single-scene scenarios, these methods must be retrained or fine-tuned to handle new environments. + +Floorplan Localization. Prior work addressing the task of floorplan localization primarily focused on depth-based cues, leveraging image-derived depth predictions or sensors, to match depth obtained from floorplans. In particular, LiDAR-based methods [3, 4, 18, 34] utilize precise laser scans but restrict usability on most mobile devices. Alternative sources of geometric cues, including semi dense visual odometry (SDVO) [8], or point clouds from depth cameras [14], can circumvent heavy LiDAR hardware. + +Earlier works compare extracted room edges directly to the 2D layout [5], often assuming knowledge of camera or room height [5, 8]. Other approaches, are embedding RGB images and floorplans into a shared metric space in order to do the localization. For instance, LaLaLoc [13] uses panoramic depth layout image which is rendered at known heights at different locations within the floorplan, where the localization is done by doing similarity in the embedded space. LaLaLoc++ [12] removes the hight assumption by embedding the entire floorplan into the feature space. However, these approaches often require an upright camera pose [12, 13], making them less flexible for hand-held or head-mounted devices. F3Loc [7] localizes by predicting depth rays from a given image and generating probability volumes that indicate the likelihood of each the depth prediction to a particular location on the floorplan. Although effectively leverages geometric cues from depth maps, it does not incorporate semantic information, which may reduce its robustness in environments with repetitive structures or occlusions. + +Semantic-based cues are less commonly used for indoor localization, but several works have utilized them for this task. Wang et al.[33] extracts scene texts from images in large indoor spaces and performs localiza + +tion by matching this text to the floorplan. SeDAR [21] uses a CNN to extract semantic labels from an image and perform Monte Carlo localization. By contrast to our single-image localization framework, it depends on sequential images and depth sensors. PF-Net [15] applies a differentiable particle filter with a learned observation model, relying on a computationally-intensive embedding process that yields limited performance for the single-image scenario. SPVLoc [10] matches captured images to panoramic renderings to estimate location, leveraging additional height information present in these renderings. Similarly, Kim et al. [17] perform localization by using a pre-captured 3D map of the environment. Our approach does not assume the availability of these additional cues. LASER [22] treats the floorplan as a set of points and synthesizes view-dependent features, including the semantic label of each point, for matching purposes. They embed images into circular features, which are then compared to pose features in the same embedded space. By contrast, we model the semantics as fine-grained ray predictions. Furthermore, unlike these prior works, our approach can operate over images with non-zero pitch and roll. + +# 3. Method + +In this work, we propose a floorplan localization framework that jointly estimates semantic and depth rays to infer the 2D camera location and orientation relative to a given floorplan. Specifically, we assume that we are provided with an RGB Image $I \in \mathbb{R}^{h \times w \times 3}$ , where $h$ and $w$ denote the height and width of the image, respectively, and a 2D floorplan map $F \in \{0,1,2,\dots,C\}^{H \times W}$ , represented as a matrix of dimensions $H \times W$ , where each element is assigned a semantic label from $C$ unique semantic categories, with zero denoting empty space. The semantic categories we consider in our work are wall, window and door, but our framework could easily incorporate additional categories (e.g., staircases, columns). + +Our objective is to predict the camera's 2D location $(x,y)$ and orientation angle $\theta$ at which the image $I$ was captured. That is, given the observation $O_{I,F} = (I,F)$ , our goal is to infer the location parameters $S_{I,F} = (x,y,\theta)$ . To this end, we adopt a probabilistic framework by modeling the posterior distribution $p(S_{I,F}\mid O_{I,F})$ . We discretize the camera pose space as $\mathcal{S} = \{S_i\}$ and define a probability volume $P\in \mathbb{R}^{\hat{H}\times \hat{W}\times O}$ where each element $P(S_{i})$ represents the posterior probability $p(S_i\mid O_{I,F})$ for a candidate pose $S_{i}$ . Here, $\hat{H}$ and $\hat{W}$ denote the number of discretized cells in the $x$ and $y$ dimensions, respectively, and $O$ represents the number of orientation bins. The + +final predicted camera pose is then given by + +$$ +\hat{S}_{I,F} = \operatorname *{arg max}_{S_{i}\in \mathcal{S}}p(S_{i}\mid O_{I,F}). +$$ + +We proceed to provide background (Section 3.1), prior to introducing our semantic prediction network (Section 3.2), which constructs a semantic probability volume. We then describe how it is fused with depth cues to perform floorplan localization (Section 3.3). Finally, we provide training and implementation details (Section 3.4); additional details can be found in the supplementary material. An overview of our approach is provided in Figure 2. + +# 3.1. Background: F3Loc + +Our work builds upon F3Loc [7], a recent technique that estimates depth rays for performing floorplan localization given a single image or image sequence. We briefly outline several key components from their work that provide background for our framework. For additional details, we refer readers to their work. + +Depth Rays Prediction. Given a query image, a depth prediction network estimates per-column depth values that capture the distance from the camera to the nearest wall along specific angles. These values are then linearly interpolated to produce a fixed set of equiangular depth rays $\hat{r}_d\in \mathbb{R}^l$ that represent the floorplan depth, with $l$ denoting the number of predicted rays. + +Estimating Depth Probability Volume. For every candidate location $(x,y)$ on the floorplan and each discrete orientation $\theta$ , a corresponding set of reference rays is generated based on the floorplan's geometry. The predicted depth rays are compared with these reference rays to compute a likelihood score for each grid cell and orientation, resulting in a three-dimensional probability volume $P_{d} \in [0,1]^{\left[\hat{H},\hat{W},O\right]}$ . For instance, given a $10\mathrm{m} \times 7\mathrm{m}$ floorplan discretized at $0.1\mathrm{m}$ with $10^{\circ}$ increments in orientation yields a probability volume $P_{d} \in [0,1]^{[100,70,36]}$ . + +The final camera 2D location and orientation are determined by selecting the grid cell with the highest likelihood in the probability volume. This process maximizes the posterior probability, thereby estimating the camera's $x$ and $y$ coordinates as well as its orientation. + +# 3.2. Adding a Semantic Prediction Network + +To utilize semantic cues for performing floorplan localization, we propose to add a semantic prediction network that first predicts semantic rays, and then processes these for constructing a semantic probability volume. We provide additional details in what follows, and then present an optional room type prediction component, which can be utilized if room labels are available. + +![](images/f38ed0d871f009b8ab30b71b5f53881bff7c9ffcd4a6b084b349294da124e2b3.jpg) +Figure 2. Overview of our pipeline. The input image is processed to generate depth rays, semantic rays, and optionally additional metadata (e.g., room type prediction). We interpolate the ray predictions to a low-resolution representation and generate the depth probability volume $P_{d}$ and the semantic probability volume $P_{s}$ (optionally masked according to the room type prediction). These probability volumes are then fused to form the structural-semantic probability volume $P_{c}$ for efficient coarse localization. Finally, we refine the candidate poses using high-resolution ray predictions and predict the final 2D camera location and orientation, visualized with an arrow on the right. + +![](images/50938694901a08d9a0703247fc0202c72c1c2e8939d5799855a233394475f580.jpg) +Figure 3. Overview of our semantic prediction network that predicts a set of semantic rays through the Semantic Ray Branch (top) and an optional room type value—e.g., Living Room—through the Room Classification Branch (bottom). The room type is used for extracting the mask $M_{\mathrm{room}}$ , as visualized on the bottom right. + +Semantic Rays Prediction. Unlike the continuous depth values estimated in prior work, the semantic rays should correspond to semantic categories, which are represented as a set of discrete classes. Therefore, we construct a network that produces a semantic ray representation $\hat{r}_s \in \{1, \dots, C\}^l$ from the image, where each ray is classified into one of $C$ semantic categories. We provide an overview of our semantic ray prediction network in Figure 3. + +As illustrated in the figure, our semantic network architecture leverages a pretrained ResNet50 backbone to extract robust, high-level features from an input RGB image $I$ . After reducing the feature channels using a CNN and projecting them into a lower-dimensional subspace, positional encodings are computed to preserve spatial information. Two sets of learnable tokens are introduced: a set of $l$ ray tokens responsible for predicting the semantic ray representation $\hat{r}_s$ and a single (optional) CLS token dedicated for representing global room clas + +sification information. + +A single-head cross attention module integrates these tokens with the flattened spatial features, yielding refined tokens that capture both global context and local details. In the ray branch, the refined ray tokens are first processed by a self-attention block that enables each token to interact with all others, thereby aggregating complementary contextual information. The enriched tokens are then passed through an MLP to produce per-token semantic logits, which after normalization form the final semantic ray vector $\hat{r}_s$ . If room labels are available in the dataset, a similar network processes the CLS token for room type prediction, as we further detail later. + +Estimating Semantic Probability Volume. To obtain the semantic probability map, $P_{s} \in [0,1]^{[\hat{H},\hat{W},O]}$ , we first need to interpolate the $l$ predicted semantic rays. Regular linear interpolation—which prior work used for depth estimation—is unsuitable in the context of discrete labeling since interpolating between class labels can produce non-valid or semantically meaningless results. Instead, we propose a voting-based interpolation scheme: We reduce the original equiangular rays to the desired count by applying a majority vote within a small neighborhood. We use a window of three rays, assigning the label that appears most frequently in that window to the center target ray; see the supplementary material for the full algorithm. Next, we compute the score for each set of rays by taking the $L_{1}$ difference between the predicted semantic labels and the reference labels. The score is then exponentiated and normalized to form the semantic probability volume $P_{s}$ , which quantifies the likelihood of each candidate pose based on the alignment between the semantic rays and the candidate pose. + +Room Type Prediction. In addition to predicting semantic rays, our semantic network can optionally also predict the room type, which corresponds to the room from which the input image was taken. This is achieved by processing the CLS token in the semantic network (see Figure 3). If the predicted room probability exceeds a threshold $T_{\mathrm{room}}$ , the predicted room type is then used to construct a mask $M_{\mathrm{room}}$ from the polygons associated with that room type. For example, if the model predicts a Living Room label with high confidence, the mask $M_{\mathrm{room}}$ retains only the regions corresponding to the living room by setting all other areas to zero. Let $\tilde{P}$ denote the original probability volume, then the masked probability volume $P$ is computed as: + +$$ +P = M _ {\mathrm {r o o m}} \odot \tilde {P}, +$$ + +where $\odot$ denotes element-wise multiplication, thereby filtering out all the probabilities which are not in the living room and substantially narrowing down the search space. A detailed analysis of room type distributions and the model's prediction accuracy is provided in the supplementary material. + +# 3.3. Floorplan Localization via a Structural-Semantic Probability Volume + +We obtain the final probability map by generating the depth probability map $P_{d}$ , following the procedure described in Section 3.1, and the semantic probability map $P_{s}$ , according to the approach detailed in Section 3.2. To leverage both semantic and geometric cues, we fuse the two probability maps using a weighted combination: + +$$ +P _ {c} = w _ {s} \cdot P _ {s} + w _ {d} \cdot P _ {d}, +$$ + +where $w_{s}$ denotes the weight given to the semantic cues, while $w_{d} = 1 - w_{s}$ represents the weight of the depth cues. These weights are determined using a held-out validation set, as further detailed in Section 4. Note that as both $P_{s}$ and $P_{d}$ are generated from interpolated rays, which we refer to as low-resolution rays (see Figure 4, predicted), and hence the initial probability volume $P_{c}$ is also in low resolution. + +Location Extraction. Our approach follows a coarse-to-fine strategy to achieve precise localization while maintaining computational efficiency. Given $P_{c}$ , we first extract the Top- $k$ candidate poses from the structural-semantic probability volume based on their scores, ensuring that each candidate is separated by at least $\delta_{\mathrm{res}}$ in translation. Next, for each candidate, we generate an augmented set of orientations by perturbing its original angle in increments of $\pm \delta_{\mathrm{ang}}$ up to a maximum deviation of $\Delta_{\mathrm{max}}$ , resulting in the final augmented set: + +$$ +[ 0, \pm \delta_ {\text {a n g}}, \pm 2 \delta_ {\text {a n g}}, \dots , \pm \Delta_ {\max} ]. +$$ + +![](images/a651a3996ada6479ed5a323afd0cf046076b6e13ee7d32c53e35d5047ae6c4ac.jpg) +Figure 4. Comparing low-resolution and high-resolution rays with our coarse-to-fine approach. Given an input image (left), we construct a structural-semantic probability volume by comparing low resolution ground-truth and predicted rays (center). Location extraction from this coarse volume directly often yields significant errors, as illustrated with the yellow arrows. Results are refined only for the Top-k candidate poses by comparing the high resolution ground-truth and predicted rays (right). This allows for efficiently extracting more accurate predictions, as further detailed in Section 3.3. + +At each candidate location, we compute the corresponding high-resolution ground-truth depth and semantic ray representations, yielding $l$ rays per candidate location. These ground-truth rays are then compared against the original predicted $l$ rays using a similarity metric (see supplementary material for metric details). For a visual comparison of the high-resolution ground-truth and predicted rays, please refer to Figure 4. The candidate with the highest similarity score is selected as the final predicted location. + +This coarse-to-fine refinement process effectively leverages the full resolution of the predictions, which were initially interpolated for runtime efficiency, to achieve more precise localization. Note that there is a tradeoff between accuracy and computation time: finer resolutions yield improved precision at the expense of increased computational load. This tradeoff is illustrated in Figure 4. As can be observed, comparing low resolution rays often yields inaccurate predictions. For instance, environments with multiple semantic objects can suffer from loss of critical details (e.g., the left door is omitted in the middle row) in the low-resolution prediction, which can lead to misclassification, further motivating the refinement step. We present a quantitative analysis of this tradeoff in the supplementary material. + +# 3.4. Training and Implementation Details + +Both networks are trained in an end-to-end manner. For depth prediction, an $L_{1}$ loss supervises the predicted depths against ground-truth depth maps. For semantic prediction, a cross-entropy loss is applied to the predicted semantic labels. If room labels $R$ are available in the dataset, an additional cross-entropy loss is used for the room label, and the network is trained to optimize both objectives jointly. As in prior work, data aug + +mentation techniques—including virtual roll-pitch augmentation—are employed to improve robustness to nonupright camera poses. The networks are optimized using the Adam optimizer with an initial learning rate of $1 \times 10^{-3}$ . We use a floorplan resolution of $0.1\mathrm{m}$ and an angular granularity of $10^{\circ}$ . We set the number of predicted rays to $l = 40$ and interpolate these to 7 rays during the coarse stage of localization. For the Location Extraction module, we use $\delta_{\mathrm{res}} = 0.1\mathrm{m}$ , $\delta_{\mathrm{ang}} = 5^{\circ}$ , and $\Delta_{\mathrm{max}} = 5^{\circ}$ , and refine our results using Top-5 poses. + +# 4. Results + +In this section, we present a comprehensive evaluation of our localization method. We begin by introducing the datasets we use in our experiments, followed by a discussion of the baselines we compare against and the evaluation metrics. The main results are reported in Section 4.1. We conduct an ablation study to assess the contributions of various components of our approach in Section 4.2. We report results for our approach under two conditions: one in which room labels are not utilized (denoted as $\mathrm{Ours}_s$ ) and another in which room labels are employed to further refine the predictions (denoted as $\mathrm{Ours}_r$ ). Additional experiments and qualitative results are provided in the supplementary material. + +Datasets. We conduct experiments on two popular datasets: Structured 3D (S3D) [37] and ZInD [9]. S3D is a synthetic dataset containing realistic 3D renders of 3,296 houses. We use the fully furnished version of S3D, as employed in previous works. ZInD consists of 1,575 unfurnished homes containing only panoramic images. We crop these panoramas to perspective views with an $80^{\circ}$ field of view (as was done in S3D), and generate a fixed-size dataset from the resulting images. For both datasets, we follow their official train and test splits. + +Baselines. We compare our approach against two baselines: F3Loc [7] (considering the single-image localization component) and LASER [22]. We also report performance using an ORACLE ray prediction. This oracle ray prediction simulates the best possible performance achievable by our pipeline using ground truth depth and semantic rays. Note that the Oracle ray prediction baseline does not incorporate room-aware features. For F3Loc, results on the ZIND dataset are obtained from our experiments by running their publicly available code on the dataset, as the original paper does not include results on this dataset. We also use the publicly available implementation of LASER. Additional details are provided in the supplementary material. + +Metrics. Following prior work [7, 22], we report recall metrics computed at distance thresholds of $0.1\mathrm{m}$ , $0.5\mathrm{m}$ , and $1\mathrm{m}$ . We also report recall for predictions with an + +
S3D R@
Method0.1m0.5m1m1m 30°
LASER0.76.410.48.7
F3Loc1.514.622.421.3
Ourss5.4241.8753.5252.61
Oursr5.7045.5358.7857.49
ORACLE61.0093.8494.8794.57
ZInD R@
Method0.1m0.5m1m1m 30°
LASER1.3811.0617.5513.64
F3Loc0.677.9015.0711.46
Ourss2.9824.0033.9629.30
Oursr3.3126.6038.0131.86
ORACLE26.4260.8567.6965.13
+ +Table 1. Recall performance on the S3D and ZInD datasets. The table reports recall at thresholds of $0.1\mathrm{m}$ , $0.5\mathrm{m}$ , $1\mathrm{m}$ , and $1\mathrm{m}$ with a $30^{\circ}$ orientation tolerance for LASER, F3Loc, our approach without room labels $(\mathrm{Ours}_s)$ , our approach with room labels $(\mathrm{Ours}_r)$ , and an Oracle ray prediction. + +orientation error bounded to less than $30^{\circ}$ (at the $1\mathrm{m}$ threshold). Recall is defined as the percentage of predictions that fall within these thresholds. + +# 4.1.Quantitative Evaluation + +Results are reported in Table 1. On both S3D and ZinD, our method more than doubles F3Loc's and LASER performance across all thresholds. Notably, considering S3D over the $\mathrm{R@1m30^{\circ}}$ metric—which reflects the quality of matches between the predicted and actual camera views—in comparison to F3Loc, our method improves by more than three times. Room type predictions yield improvements of $9.2\%$ in $\mathrm{R@1m30^{\circ}}$ on the S3D dataset and $8.7\%$ on the ZInD dataset. We include an additional experiment in the supplementary material that evaluates F3Loc with our refinement module, demonstrating that both the semantic rays and our refinement module provide significant performance gains. + +We observe that significant performance improvements are also achieved for a very fine-grained recall metric of $0.1\mathrm{m}$ , boosting performance from $1.5\%$ (F3Loc) to $5.7\%$ with our approach on S3D. We attribute this to our coarse-to-fine strategy, which effectively refines the coarse location estimates into precise predictions, as further validated in our ablation study. + +Figure 5 presents qualitative examples that illustrate how the integration of floorplan semantics with precise depth cues enables our pipeline to effectively resolve localization ambiguities. For instance, in the third row, we can see that F3Loc, which does not use semantics, interprets this image as a blank wall, while LASER misinter + +![](images/95882f81b1b62fdf7506fdbf047cbb5cf655a42fff36b39c99d5aed1dcf943be.jpg) +Figure 5. Qualitative comparison of our method with F3Loc and LASER. Warmer colors correspond to regions with higher probabilities. Below each map we report the localization error in meters and degrees. We use arrows to visualize the ground truth location (magenta) and the predicted location (white). + +prets the window size and predicts an incorrect location. + +# 4.2. Ablation Study + +We demonstrate the effect of incorporating each component of our method on overall localization performance. Specifically, we conduct the following ablations: (1) Base, which corresponds to computing the structural-semantic probability volume and selecting the maximum probability without any additional refinement. (2) Removing semantic interpolation (denoted as -Interpolation), where we replace our majority voting interpolation with a simple linear interpolation followed by rounding. (3) Adding room predictions (denoted as +Room), where we assess the effect of integrating room type predictions into our localization pipeline. (4) Adding refinement (denoted as +Refine), which assesses our coarse-to-fine approach, which refines our localization extraction via the Top-K candidate poses. (5) Adding room predictions and refinement (denoted as +Room&Refine), which combines both components. + +From the results in Table 2, we can see that on the $1\mathrm{m}30^{\circ}$ metric, the addition of our refinement module improves performance by $8.6\%$ relative to the baseline. This indicates that a substantial amount of information is lost during the initial interpolation process if not properly addressed, thereby strongly motivating the use of coarse-to-fine strategies in our Location Extraction module. We also observe a significant improvement of $11.5\%$ from the room prediction component, which is discussed in further detail in the supplementary material and visualized in Figure 6. Finally, our majority voting interpolation approach contributes an additional gain of $0.9\%$ compared to a simple interpolation strategy. When combining all these improvements, our method achieves + +
Method0.1m0.5m1m1m 30°
Base4.6538.3549.4048.44
- Interpolation4.7338.4448.9147.99
+ Room5.1242.9255.5754.04
+ Refine5.4241.8753.5252.61
+ Room&Refine5.7045.5358.7857.49
+ +Table 2. Ablation study, evaluating the effect of incorporating the different components in our pipeline on the S3D dataset. + +![](images/29b668141624274f1c51c29320cb2b933f770b5b1e8861e9b87f091d4c1df815.jpg) +Figure 6. Qualitative comparison of using room predictions and our coarse-to-fine refinement approach. Below each map we report the localization error in meters and degrees. Warmer colors correspond to regions with higher probabilities. Overlaid on the estimated probabilities, we show the ground truth location (magenta) and the predicted location. + +an overall enhancement of $18.6\%$ on the 1m $30^{\circ}$ metric relative to our base model. + +In Figure 6, we illustrate the impact of incorporating room type predictions alongside the location extraction module. The figure clearly demonstrates how these components refine the probability volume by narrowing down the candidate poses, which in turn improves overall localization accuracy. + +We analyze the impact of different combinations of predicted depth and semantic features on the S3D dataset. Figure 7 summarizes the recall performance for various depth and semantic weight configurations used to compute the structural-semantic probability volume. In Figure 8 we visualize examples from two extreme scenarios (depth only and semantic only) and from the configuration adopted in our work $\left([w_{s}, w_{d}] = [0.4, 0.6]\right)$ , selected according to the validation set. + +As can be observed from Figure 8 (and also reflected in Figure 7), relying solely on semantic rays tends to produce a more diffused probability volume with multiple ambiguous candidate locations. This occurs because most images contain a single semantic object of a standard size, allowing an accurate ray pattern to be fitted in more than one location. Similarly, depth cues alone also suffer from ambiguity, particularly in repetitive environments or when the image captures three or fewer walls, which leads to uncertainty in the localization estimate. However, when these semantic cues are com + +![](images/c2dfec3c57efe79de669b4ba1270029e8c3f6e09f854cb967d0b2f1cdb4e7d82.jpg) +Figure 7. Recall vs. weight combinations on the S3D test set. The plot shows the recall metrics for four different thresholds: $0.1m$ , $0.5m$ , $1m$ , and $1m30^{\circ}$ . The x-axis displays the weight combinations in the order $(w_{s}, w_{d})$ . A vertical dashed line at $w_{s} = 0.4$ , $w_{d} = 0.6$ highlights the weight combination selected using the validation set. + +![](images/5dc6920b295726ffc7ed155da72789d9b30d5f087949db643d7f4207bff66beb.jpg) +Figure 8. Qualitative Comparison of Depth-Only, Semantic-Only, and Fused Structural-Semantic Probability Volumes. Below each map we report the localization error in meters and degrees. Warmer colors correspond to regions with higher probabilities. + +bined with depth rays, the resulting probability volume becomes significantly more concentrated. This integration effectively filters out spurious candidates and sharpens the localization estimate. + +Additionally, the supplementary material offers a detailed analysis of how varying the Top-K candidates influences localization refinement, along with additional experiments that motivate various design choices, such as using hard thresholds in multiple steps of our pipeline. + +# Runtime Analysis + +We report runtime performance breakdown in Table 3, using the same parameters employed in Table 1 with + +
KPredictionLocRefinTotal
10.038 ± 0.1190.174 ± 0.0330.141 ± 0.0730.364 ± 0.142
30.033 ± 0.0930.154 ± 0.0260.356 ± 0.1190.554 ± 0.157
50.034 ± 0.1030.155 ± 0.0290.577 ± 0.1850.778 ± 0.218
+ +Table 3. Performance breakdown over different Top-K values. Each entry is mean ± std (s). Prediction denotes the ray predictions, Loc refers to the localization process, and Refin represents the refinement stage, during which candidate locations are evaluated by computing the ground truth rays. + +varying K values. As shown in the table, the per-image inference time increases with the number of candidate refinements, $K$ , reflecting the additional computations required during refinement. All experiments were conducted on a single CPU without multithreading to avoid introducing bias. Future work could explore parallelizing the refinement stage by computing all ground-truth rays simultaneously for further speed-up. Importantly, even at $K = 5$ the setting we adopt in our work, the inference time remains reasonable, striking a practical balance between computational cost and improvements in localization accuracy. We further observe that the prediction and localization steps require similar amounts of time, while the refinement step grows monotonically as $K$ increases. + +# 5. Conclusion + +In this work, we presented a semantic-aware localization framework that extends floorplan-based camera localization by fusing semantic labels with geometric depth cues. Our approach leverages a novel semantic ray prediction network alongside an established depth estimation method to generate a semantic-structural probability volume, which significantly improves localization accuracy, especially in environments with repetitive or ambiguous structural patterns. + +Our extensive experiments on the S3D and ZInD datasets demonstrate that integrating semantic cues effectively resolves depth-based ambiguities and consistently outperforms state-of-the-art methods such as F3Loc and LASER. Ablation studies confirm that a balanced combination of depth and semantic information, coupled with a coarse-to-fine localization strategy and the use of room labels, yields optimal performance. + +Looking forward, extending our framework to incorporate additional semantic labels and other modalities, such as textual information, promises to further enhance localization robustness in challenging indoor settings. In general, our approach represents an important step towards accurate and reliable indoor localization systems by effectively leveraging semantic and geometric cues. + +# References + +[1] Relja Arandjelovic, Petr Gronat, Akihiko Torii, Tomas Pajdla, and Josef Sivic. NetVLAD: CNN architecture for weakly supervised place recognition. In CVPR, pages 5297-5307, 2016. 2 +[2] Vassileios Balntas, Shuda Li, and Victor Prisacariu. Relocnet: Continuous metric learning relocalisation using neural nets. In ECCV, pages 751-767, 2018. 2 +[3] Federico Boniardi, Tim Caselitz, Rainer Kummerle, and Wolfram Burgard. Robust lidar-based localization in architectural floor plans. In 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3318-3324. IEEE, 2017. 2 +[4] Federico Boniardi, Tim Caselitz, Rainer Kummerle, and Wolfram Burgard. A pose graph-based localization system for long-term navigation in cad floor plans. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 84-97. IEEE, 2019. 2 +[5] Federico Boniardi, Abhinav Valada, Rohit Mohan, Tim Caselitz, and Wolfram Burgard. Robot localization in floor plans using a room layout edge extraction network. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5291-5297. IEEE, 2019. 2 +[6] Eric Brachmann, Alexander Krull, Sebastian Nowozin, Jamie Shotton, Frank Michel, Stefan Gumhold, and Carsten Rother. DSAC-differentiable RANSAC for camera localization. In CVPR, pages 6684-6692, 2017. 2 +[7] Changan Chen, Rui Wang, Christoph Vogel, and Marc Pollefeys. F3Loc: Fusion and filtering for floorplan localization. arXiv preprint arXiv:2403.03370, 2024. 1, 2, 3, 6 +[8] Hang Chu, Dong Ki Kim, and Tsuhan Chen. You are here: Mimicking the human thinking process in reading floorplans. In ICCV, pages 2210-2218, 2015. 1, 2 +[9] Steve Cruz, Will Hutchcroft, Yuguang Li, Naji Khosravan, Ivaylo Boyadzhiev, and Sing Bing Kang. Zillow indoor dataset: Annotated floor plans with $360^{\circ}$ panoramas and 3d room layouts. In CVPR, pages 2133-2143, 2021. 6 +[10] Niklas Gard, Anna Hilsmann, and Peter Eisert. Spvloc: Semantic panoramic viewpoint matching for 6d camera localization in unseen environments. arXiv preprint arXiv:2404.10527, 2024.3 +[11] B. Glocker, S. Izadi, J. Shotton, and A. Criminisi. Realtime RGB-D camera relocalization. In Proceedings of the Mixed and Augmented Reality (ISMAR) Conference. IEEE, 2013. 1 +[12] Henry Howard-Jenkins and Victor Adrian Prisacariu. LaLaLoc++: Global floor plan comprehension for layout localisation in unvisited environments. In ECCV, pages 693-709, 2022. 1, 2 +[13] Henry Howard-Jenkins, Jose-Raul Ruiz-Sarmiento, and Victor Adrian Prisacariu. LaLaLoc: Latent layout localisation in dynamic, unvisited environments. arXiv preprint arXiv:2104.09169, 2021. 1, 2 + +[14] Seigo Ito, Felix Endres, Markus Kuderer, Gian Diego Tipaldi, Cyril Stachniss, and Wolfram Burgard. W-RGBD: Floor-plan-based indoor global localization using a depth camera and wifi. In 2014 IEEE International Conference on Robotics and Automation (ICRA), pages 417-422. IEEE, 2014. 2 +[15] Peter Karkus, David Hsu, and Wee Sun Lee. Particle filter networks with application to visual localization. In Proceedings of the 2nd Conference on Robot Learning (CoRL), pages 169-178. PMLR, 2018. 1, 3 +[16] Alex Kendall, Matthew Grimes, and Roberto Cipolla. PoseNet: A convolutional network for real-time 6-DOF camera relocalization. In ICCV, pages 2938-2946, 2015. 1, 2 +[17] Junho Kim, Jiwon Jeong, and Young Min Kim. Fully geometric panoramic localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3 +[18] Z. Li, M. H. Ang, and D. Rus. Online localization with imprecise floor space maps using stochastic gradient descent. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8571-8578. IEEE, 2020. 2 +[19] Liu Liu, Hongdong Li, and Yuchao Dai. Efficient global 2D-3D matching for camera localization in a large-scale 3D map. In ICCV, pages 2372-2381, 2017. 1, 2 +[20] Pranay Mathur, Rajesh Kumar, and Sarthak Upadhyay. Sparse image-based navigation architecture to mitigate the need of precise localization in mobile robots. arXiv preprint arXiv:2203.15272, 2022. 1 +[21] Oscar Mendez, Simon Hadfield, Nicolas Pugeault, and Richard Bowden. SeDAR: Reading floorplans like a human—using deep learning to enable human-inspired localisation. *IJCV*, 128:1286–1310, 2020. 3 +[22] Zhixiang Min, Naji Khosravan, Zachary Bessinger, Manjunath Narayana, Sing Bing Kang, Enrique Dunn, and Ivaylo Boyadzhiev. LASER: Latent space rendering for 2D visual localization. In CVPR, pages 11122-11131, 2022. 3, 6 +[23] R. A. Newcombe, A. J. Davison, S. Izadi, P. Kohli, O. Hilliges, J. Shotton, D. Molyneaux, S. Hodges, D. Kim, and A. Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In Proceedings of the Mixed and Augmented Reality (ISMAR) Conference. IEEE, 2011. 1 +[24] Takahiro Niwa, Shun Taguchi, and Noriaki Hirose. Spatio-temporal graph localization networks for image-based navigation. arXiv preprint arXiv:2204.13237, 2022.1 +[25] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: robust hierarchical localization at large scale. In CVPR, pages 12716-12725, 2019. 1, 2 +[26] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Fast image-based localization using direct 2d-to-3d matching. In ICCV, pages 667-674, 2011. +[27] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Improving image-based localization by active correspondence search. In ECCV, pages 752-765, 2012. + +[28] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Efficient & effective prioritized matching for large-scale image-based localization. IEEE TPAMI, 39(9):1744-1756, 2016. 1, 2 +[29] Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene coordinate regression forests for camera relocalization in rgb-d images. In CVPR, pages 2930-2937, 2013. 1, 2 +[30] Janine Thoma, Danda Pani Paudel, Ajad Chhatkuli, Thomas Probst, and Luc Van Gool. Mapping, localization and path planning for image-based navigation using visual features and map. In CVPR. IEEE, 2019. 1 +[31] Julien Valentin, Matthias Nießner, Jamie Shotton, Andrew Fitzgibbon, Shahram Izadi, and Philip HS Torr. Exploiting uncertainty in regression forests for accurate camera relocalization. In CVPR, pages 4400-4408, 2015. 1, 2 +[32] Florian Walch, Caner Hazirbas, Laura Leal-Taixe, Torsten Sattler, Sebastian Hilsenbeck, and Daniel Cremers. Image-based localization using lstms for structured feature correlation. In ICCV, pages 627-637, 2017. 1, 2 +[33] Shenlong Wang, Sanja Fidler, and Raquel Urtasun. Lost shopping! monocular localization in large indoor spaces. In ICCV, pages 2695-2703, 2015. 2 +[34] Xipeng Wang, Ryan J Marcotte, and Edwin Olson. GLFP: Global localization from a floor plan. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1627-1632. IEEE, 2019. 2 +[35] Wera Winterhalter, Freya Fleckenstein, Bastian Steder, Luciano Spinello, and Wolfram Burgard. Accurate indoor localization for RGB-D smartphones and tablets given 2d floor plans. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3138-3143. IEEE, 2015. 1 +[36] Jian Wu, Liwei Ma, and Xiaolin Hu. Delving deeper into convolutional neural networks for camera relocalization. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 5644-5651. IEEE, 2017. 1, 2 +[37] Jia Zheng, Junfei Zhang, Jing Li, Rui Tang, Shenghua Gao, and Zihan Zhou. Structured3d: A large photorealistic dataset for structured 3d modeling. In ECCV, 2020.6 \ No newline at end of file diff --git a/2025/Supercharging Floorplan Localization with Semantic Rays/images.zip b/2025/Supercharging Floorplan Localization with Semantic Rays/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f890dc0abd18324bd3cfb63acd5d740f1e4235bb --- /dev/null +++ b/2025/Supercharging Floorplan Localization with Semantic Rays/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:09413247d47ef4d0a00d22551582fb5cc59d83206d2aed74005619a439b6437e +size 439789 diff --git a/2025/Supercharging Floorplan Localization with Semantic Rays/layout.json b/2025/Supercharging Floorplan Localization with Semantic Rays/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..76dc6b400d227660d5aec89d23290f3710185a70 --- /dev/null +++ b/2025/Supercharging Floorplan Localization with Semantic Rays/layout.json @@ -0,0 +1,8098 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 124, + 103, + 486, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 103, + 486, + 121 + ], + "spans": [ + { + "bbox": [ + 124, + 103, + 486, + 121 + ], + "type": "text", + "content": "Supercharging Floorplan Localization with Semantic Rays" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 184, + 144, + 281, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 144, + 281, + 172 + ], + "spans": [ + { + "bbox": [ + 184, + 144, + 281, + 172 + ], + "type": "text", + "content": "Yuval Grader \nTel Aviv University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 321, + 144, + 426, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 144, + 426, + 172 + ], + "spans": [ + { + "bbox": [ + 321, + 144, + 426, + 172 + ], + "type": "text", + "content": "Hadar Averbuch-Elor Cornell University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 198, + 186, + 411, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 186, + 411, + 198 + ], + "spans": [ + { + "bbox": [ + 198, + 186, + 411, + 198 + ], + "type": "text", + "content": "https://tau-vailab.github.io/SemRayLoc/" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 216, + 206, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 216, + 206, + 228 + ], + "spans": [ + { + "bbox": [ + 159, + 216, + 206, + 228 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 243, + 296, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 243, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 67, + 243, + 296, + 531 + ], + "type": "text", + "content": "Floorplans provide a compact representation of the building's structure, revealing not only layout information but also detailed semantics such as the locations of windows and doors. However, contemporary floorplan localization techniques mostly focus on matching depth-based structural cues, ignoring the rich semantics communicated within floorplans. In this work, we introduce a semantic-aware localization framework that jointly estimates depth and semantic rays, consolidating over both for predicting a structural-semantic probability volume. Our probability volume is constructed in a coarse-to-fine manner: We first sample a small set of rays to obtain an initial low-resolution probability volume. We then refine these probabilities by performing a denser sampling only in high-probability regions and process the refined values for predicting a 2D location and orientation angle. We conduct an evaluation on two standard floorplan localization benchmarks. Our experiments demonstrate that our approach substantially outperforms state-of-the-art methods, achieving significant improvements in recall metrics compared to prior works. Moreover, we show that our framework can easily incorporate additional metadata such as room labels, enabling additional gains in both accuracy and efficiency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 569, + 149, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 569, + 149, + 582 + ], + "spans": [ + { + "bbox": [ + 69, + 569, + 149, + 582 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 590, + 295, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 590, + 295, + 723 + ], + "spans": [ + { + "bbox": [ + 67, + 590, + 295, + 723 + ], + "type": "text", + "content": "Camera localization is a longstanding problem in computer vision, with significant applications in 3D reconstruction [19, 25-28], augmented reality [11, 16, 23, 29, 31], and navigation [20, 24, 30, 32, 36]. Localization within indoor environments is especially challenging due to the absence of reliable GPS signals and the complexity of reasoning across multiple floors and layers. Hence, to bypass complicated 3D model-based localization techniques, prior work [7, 8, 12, 13, 15, 35] has explored the problem of localizing camera observations within a provided 2D floorplan map by matching" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 319, + 214, + 437, + 271 + ], + "blocks": [ + { + "bbox": [ + 319, + 214, + 437, + 271 + ], + "lines": [ + { + "bbox": [ + 319, + 214, + 437, + 271 + ], + "spans": [ + { + "bbox": [ + 319, + 214, + 437, + 271 + ], + "type": "image", + "image_path": "77ef471a7a1601e41515129379e2e22703659419519e68e969f5e6eddefc7a41.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 320, + 278, + 438, + 379 + ], + "blocks": [ + { + "bbox": [ + 320, + 278, + 438, + 379 + ], + "lines": [ + { + "bbox": [ + 320, + 278, + 438, + 379 + ], + "spans": [ + { + "bbox": [ + 320, + 278, + 438, + 379 + ], + "type": "image", + "image_path": "858b3aaecf03cba21dc1abcc2b6f98c495992a3f6724f89654ed9f55cb160005.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 320, + 380, + 438, + 483 + ], + "blocks": [ + { + "bbox": [ + 320, + 380, + 438, + 483 + ], + "lines": [ + { + "bbox": [ + 320, + 380, + 438, + 483 + ], + "spans": [ + { + "bbox": [ + 320, + 380, + 438, + 483 + ], + "type": "image", + "image_path": "6cd75802e283904d983688ac00f6389d1e1bddeecdf915c83ef41fca98fe8618.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 485, + 402, + 500 + ], + "lines": [ + { + "bbox": [ + 347, + 485, + 402, + 500 + ], + "spans": [ + { + "bbox": [ + 347, + 485, + 402, + 500 + ], + "type": "text", + "content": "Input & GT" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 446, + 276, + 539, + 377 + ], + "blocks": [ + { + "bbox": [ + 446, + 276, + 539, + 377 + ], + "lines": [ + { + "bbox": [ + 446, + 276, + 539, + 377 + ], + "spans": [ + { + "bbox": [ + 446, + 276, + 539, + 377 + ], + "type": "image", + "image_path": "14fc60037bf8e390c93cfefa04e4bacfe3e63beec82af7bef376240fe45a3769.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 446, + 381, + 539, + 483 + ], + "blocks": [ + { + "bbox": [ + 446, + 381, + 539, + 483 + ], + "lines": [ + { + "bbox": [ + 446, + 381, + 539, + 483 + ], + "spans": [ + { + "bbox": [ + 446, + 381, + 539, + 483 + ], + "type": "image", + "image_path": "1c8b5dad1f5d9f5c90b4e1ea7987df7e3438757c72b5d6306ba7ed55a817c077.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 445, + 487, + 526, + 499 + ], + "lines": [ + { + "bbox": [ + 445, + 487, + 526, + 499 + ], + "spans": [ + { + "bbox": [ + 445, + 487, + 526, + 499 + ], + "type": "text", + "content": "Output Localization" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 508, + 541, + 586 + ], + "lines": [ + { + "bbox": [ + 313, + 508, + 541, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 541, + 586 + ], + "type": "text", + "content": "Figure 1. Floorplan localization using a raw binary floorplan (middle row) often yields ambiguous predictions. In this work, we utilize a richer, yet readily available, representation: a floorplan enhanced with semantic labels (bottom row). We present an approach that supercharges floorplan localization with semantic rays, enabling for resolving localization ambiguities, as illustrated by the comparison on the right." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 612, + 428, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 428, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 428, + 624 + ], + "type": "text", + "content": "depth-based structural cues." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 627, + 541, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 627, + 541, + 723 + ], + "spans": [ + { + "bbox": [ + 313, + 627, + 541, + 723 + ], + "type": "text", + "content": "However, while floorplans offer a compact and lightweight scene representation, structural cues within floorplans often correlate with multiple candidate locations, particularly for environments with repetitive or symmetric layouts. Consider the example in Figure 1. Can you localize the input image within the floorplan? Provided with just a raw (walls only) floorplan, room corners are indistinguishable and hence localization is" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 753, + 318, + 763 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 318, + 763 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 318, + 763 + ], + "type": "text", + "content": "27116" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 294, + 144 + ], + "type": "text", + "content": "highly ambiguous, as can also be observed by the probabilities predicted by the state-of-the-art F3Loc [7] technique (middle row, right). To resolve such ambiguities, we are interested in utilizing a slightly different representation: a semantics-aware floorplan, such as the one illustrated on the bottom row of Figure 1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 145, + 294, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 145, + 294, + 264 + ], + "spans": [ + { + "bbox": [ + 67, + 145, + 294, + 264 + ], + "type": "text", + "content": "Accordingly, we introduce a semantic-aware ray-based localization framework that integrates semantic cues with depth-based predictions. In particular, we propose a semantic prediction network that predicts accurate semantic ray representations along with optional additional metadata (such as room labels) from a single RGB image with a limited field-of-view. We process these rays to compute a semantic probability volume, which is then fused with depth information for constructing a structural-semantic probability volume." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 266, + 294, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 266, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 67, + 266, + 294, + 373 + ], + "type": "text", + "content": "Our framework follows a coarse-to-fine localization strategy. We first operate over a low-resolution image for an efficient floorplan search. This initial search yields the Top-" + }, + { + "bbox": [ + 67, + 266, + 294, + 373 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 266, + 294, + 373 + ], + "type": "text", + "content": " candidate locations. Finally, we refine these candidates and select the best match using a high-resolution ray representation. By adopting this coarse-to-fine approach, our method effectively constrains the localization search to the most promising regions while keeping the computation cost feasible." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 375, + 295, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 375, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 67, + 375, + 295, + 540 + ], + "type": "text", + "content": "Our experiments demonstrate that our approach yields improvements by factors ranging from two to three across most metrics compared to the state-of-the-art technique [7], upon which our method is built. This substantial gain underscores the effectiveness of fusing semantic and depth ray predictions into a unified probability volume. We further show that our coarse-to-fine strategy offers a flexible tradeoff between accuracy and computational cost, with performance consistently improving as larger candidate sets (Top- " + }, + { + "bbox": [ + 67, + 375, + 295, + 540 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 375, + 295, + 540 + ], + "type": "text", + "content": ") are evaluated—making our method adaptable to diverse task requirements. Moreover, incorporating additional metadata further enhances precision, leading to significantly improved localization accuracies." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 80, + 543, + 240, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 543, + 240, + 554 + ], + "spans": [ + { + "bbox": [ + 80, + 543, + 240, + 554 + ], + "type": "text", + "content": "Explicitly stated, our contributions are:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 556, + 295, + 651 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 69, + 556, + 295, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 556, + 295, + 580 + ], + "spans": [ + { + "bbox": [ + 69, + 556, + 295, + 580 + ], + "type": "text", + "content": "- We introduce a semantic ray prediction network that receives a single RGB image as input." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 580, + 295, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 580, + 295, + 627 + ], + "spans": [ + { + "bbox": [ + 69, + 580, + 295, + 627 + ], + "type": "text", + "content": "- We propose an efficient and unified framework that fuses semantic and depth ray predictions into a structural-semantic probability volume, which effectively resolves localization ambiguities." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 628, + 294, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 628, + 294, + 651 + ], + "spans": [ + { + "bbox": [ + 69, + 628, + 294, + 651 + ], + "type": "text", + "content": "- Results that demonstrate that significant improvements over state-of-the-art methods." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 69, + 665, + 156, + 677 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 665, + 156, + 677 + ], + "spans": [ + { + "bbox": [ + 69, + 665, + 156, + 677 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 686, + 295, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 686, + 295, + 723 + ], + "spans": [ + { + "bbox": [ + 67, + 686, + 295, + 723 + ], + "type": "text", + "content": "Visual Localization. The task of visual localization has received ongoing attention throughout the past several decades. Traditional approaches often rely on image re" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 541, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 541, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 541, + 204 + ], + "type": "text", + "content": "trieval or on a 3D Structure-from-Motion (SfM) model of the environment. In the image retrieval paradigm, methods such as NetVLAD [1] or RelocNet [2] compare a query image against a database of labeled images. Once the closest match is found, the query pose is approximated by the retrieved image's pose. Other methods explicitly construct a 3D SfM model of a scene to establish 2D-3D correspondences between the query image and the reconstructed 3D structure [19, 25-28]. After matching local image descriptors to 3D points, robust solvers estimate the 6-DoF pose." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 204, + 541, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 541, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 541, + 299 + ], + "type": "text", + "content": "Recent learning-based pipelines deviate from classical 2D-3D matching. Scene-coordinate regression methods predict dense 3D coordinates for every pixel in the query image [6, 29, 31], whereas pose regression methods directly estimate the 6-DoF camera pose via neural networks [16, 32, 36]. Although promising for single-scene scenarios, these methods must be retrained or fine-tuned to handle new environments." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 304, + 541, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 541, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 541, + 422 + ], + "type": "text", + "content": "Floorplan Localization. Prior work addressing the task of floorplan localization primarily focused on depth-based cues, leveraging image-derived depth predictions or sensors, to match depth obtained from floorplans. In particular, LiDAR-based methods [3, 4, 18, 34] utilize precise laser scans but restrict usability on most mobile devices. Alternative sources of geometric cues, including semi dense visual odometry (SDVO) [8], or point clouds from depth cameras [14], can circumvent heavy LiDAR hardware." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 423, + 541, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 541, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 541, + 673 + ], + "type": "text", + "content": "Earlier works compare extracted room edges directly to the 2D layout [5], often assuming knowledge of camera or room height [5, 8]. Other approaches, are embedding RGB images and floorplans into a shared metric space in order to do the localization. For instance, LaLaLoc [13] uses panoramic depth layout image which is rendered at known heights at different locations within the floorplan, where the localization is done by doing similarity in the embedded space. LaLaLoc++ [12] removes the hight assumption by embedding the entire floorplan into the feature space. However, these approaches often require an upright camera pose [12, 13], making them less flexible for hand-held or head-mounted devices. F3Loc [7] localizes by predicting depth rays from a given image and generating probability volumes that indicate the likelihood of each the depth prediction to a particular location on the floorplan. Although effectively leverages geometric cues from depth maps, it does not incorporate semantic information, which may reduce its robustness in environments with repetitive structures or occlusions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 674, + 539, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 674, + 539, + 723 + ], + "spans": [ + { + "bbox": [ + 313, + 674, + 539, + 723 + ], + "type": "text", + "content": "Semantic-based cues are less commonly used for indoor localization, but several works have utilized them for this task. Wang et al.[33] extracts scene texts from images in large indoor spaces and performs localiza" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "text", + "content": "27117" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 72, + 294, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 294, + 348 + ], + "type": "text", + "content": "tion by matching this text to the floorplan. SeDAR [21] uses a CNN to extract semantic labels from an image and perform Monte Carlo localization. By contrast to our single-image localization framework, it depends on sequential images and depth sensors. PF-Net [15] applies a differentiable particle filter with a learned observation model, relying on a computationally-intensive embedding process that yields limited performance for the single-image scenario. SPVLoc [10] matches captured images to panoramic renderings to estimate location, leveraging additional height information present in these renderings. Similarly, Kim et al. [17] perform localization by using a pre-captured 3D map of the environment. Our approach does not assume the availability of these additional cues. LASER [22] treats the floorplan as a set of points and synthesizes view-dependent features, including the semantic label of each point, for matching purposes. They embed images into circular features, which are then compared to pose features in the same embedded space. By contrast, we model the semantics as fine-grained ray predictions. Furthermore, unlike these prior works, our approach can operate over images with non-zero pitch and roll." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 376, + 124, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 376, + 124, + 388 + ], + "spans": [ + { + "bbox": [ + 69, + 376, + 124, + 388 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 402, + 295, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 402, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 67, + 402, + 295, + 559 + ], + "type": "text", + "content": "In this work, we propose a floorplan localization framework that jointly estimates semantic and depth rays to infer the 2D camera location and orientation relative to a given floorplan. Specifically, we assume that we are provided with an RGB Image " + }, + { + "bbox": [ + 67, + 402, + 295, + 559 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{h \\times w \\times 3}" + }, + { + "bbox": [ + 67, + 402, + 295, + 559 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 67, + 402, + 295, + 559 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 67, + 402, + 295, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 402, + 295, + 559 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 67, + 402, + 295, + 559 + ], + "type": "text", + "content": " denote the height and width of the image, respectively, and a 2D floorplan map " + }, + { + "bbox": [ + 67, + 402, + 295, + 559 + ], + "type": "inline_equation", + "content": "F \\in \\{0,1,2,\\dots,C\\}^{H \\times W}" + }, + { + "bbox": [ + 67, + 402, + 295, + 559 + ], + "type": "text", + "content": ", represented as a matrix of dimensions " + }, + { + "bbox": [ + 67, + 402, + 295, + 559 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 67, + 402, + 295, + 559 + ], + "type": "text", + "content": ", where each element is assigned a semantic label from " + }, + { + "bbox": [ + 67, + 402, + 295, + 559 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 67, + 402, + 295, + 559 + ], + "type": "text", + "content": " unique semantic categories, with zero denoting empty space. The semantic categories we consider in our work are wall, window and door, but our framework could easily incorporate additional categories (e.g., staircases, columns)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "spans": [ + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "type": "text", + "content": "Our objective is to predict the camera's 2D location " + }, + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "type": "text", + "content": " and orientation angle " + }, + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "type": "text", + "content": " at which the image " + }, + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "type": "text", + "content": " was captured. That is, given the observation " + }, + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "type": "inline_equation", + "content": "O_{I,F} = (I,F)" + }, + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "type": "text", + "content": ", our goal is to infer the location parameters " + }, + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "type": "inline_equation", + "content": "S_{I,F} = (x,y,\\theta)" + }, + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "type": "text", + "content": ". To this end, we adopt a probabilistic framework by modeling the posterior distribution " + }, + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "type": "inline_equation", + "content": "p(S_{I,F}\\mid O_{I,F})" + }, + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "type": "text", + "content": ". We discretize the camera pose space as " + }, + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "type": "inline_equation", + "content": "\\mathcal{S} = \\{S_i\\}" + }, + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "type": "text", + "content": " and define a probability volume " + }, + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "type": "inline_equation", + "content": "P\\in \\mathbb{R}^{\\hat{H}\\times \\hat{W}\\times O}" + }, + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "type": "text", + "content": " where each element " + }, + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "type": "inline_equation", + "content": "P(S_{i})" + }, + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "type": "text", + "content": " represents the posterior probability " + }, + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "type": "inline_equation", + "content": "p(S_i\\mid O_{I,F})" + }, + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "type": "text", + "content": " for a candidate pose " + }, + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "type": "inline_equation", + "content": "\\hat{H}" + }, + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "type": "inline_equation", + "content": "\\hat{W}" + }, + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "type": "text", + "content": " denote the number of discretized cells in the " + }, + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "type": "text", + "content": " dimensions, respectively, and " + }, + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 67, + 563, + 295, + 723 + ], + "type": "text", + "content": " represents the number of orientation bins. The" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 72, + 493, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 493, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 493, + 85 + ], + "type": "text", + "content": "final predicted camera pose is then given by" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 362, + 91, + 492, + 113 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 91, + 492, + 113 + ], + "spans": [ + { + "bbox": [ + 362, + 91, + 492, + 113 + ], + "type": "interline_equation", + "content": "\\hat{S}_{I,F} = \\operatorname *{arg max}_{S_{i}\\in \\mathcal{S}}p(S_{i}\\mid O_{I,F}).", + "image_path": "11b9a51512045430b2d798b00bfe89ee2474806e7418e801d1e7455396c98ae5.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 121, + 541, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 541, + 229 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 541, + 229 + ], + "type": "text", + "content": "We proceed to provide background (Section 3.1), prior to introducing our semantic prediction network (Section 3.2), which constructs a semantic probability volume. We then describe how it is fused with depth cues to perform floorplan localization (Section 3.3). Finally, we provide training and implementation details (Section 3.4); additional details can be found in the supplementary material. An overview of our approach is provided in Figure 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 236, + 432, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 432, + 249 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 432, + 249 + ], + "type": "text", + "content": "3.1. Background: F3Loc" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 254, + 541, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 254, + 541, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 541, + 326 + ], + "type": "text", + "content": "Our work builds upon F3Loc [7], a recent technique that estimates depth rays for performing floorplan localization given a single image or image sequence. We briefly outline several key components from their work that provide background for our framework. For additional details, we refer readers to their work." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 331, + 541, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 541, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 541, + 416 + ], + "type": "text", + "content": "Depth Rays Prediction. Given a query image, a depth prediction network estimates per-column depth values that capture the distance from the camera to the nearest wall along specific angles. These values are then linearly interpolated to produce a fixed set of equiangular depth rays " + }, + { + "bbox": [ + 313, + 331, + 541, + 416 + ], + "type": "inline_equation", + "content": "\\hat{r}_d\\in \\mathbb{R}^l" + }, + { + "bbox": [ + 313, + 331, + 541, + 416 + ], + "type": "text", + "content": " that represent the floorplan depth, with " + }, + { + "bbox": [ + 313, + 331, + 541, + 416 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 331, + 541, + 416 + ], + "type": "text", + "content": " denoting the number of predicted rays." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 421, + 541, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 541, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 541, + 554 + ], + "type": "text", + "content": "Estimating Depth Probability Volume. For every candidate location " + }, + { + "bbox": [ + 313, + 421, + 541, + 554 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 313, + 421, + 541, + 554 + ], + "type": "text", + "content": " on the floorplan and each discrete orientation " + }, + { + "bbox": [ + 313, + 421, + 541, + 554 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 421, + 541, + 554 + ], + "type": "text", + "content": ", a corresponding set of reference rays is generated based on the floorplan's geometry. The predicted depth rays are compared with these reference rays to compute a likelihood score for each grid cell and orientation, resulting in a three-dimensional probability volume " + }, + { + "bbox": [ + 313, + 421, + 541, + 554 + ], + "type": "inline_equation", + "content": "P_{d} \\in [0,1]^{\\left[\\hat{H},\\hat{W},O\\right]}" + }, + { + "bbox": [ + 313, + 421, + 541, + 554 + ], + "type": "text", + "content": ". For instance, given a " + }, + { + "bbox": [ + 313, + 421, + 541, + 554 + ], + "type": "inline_equation", + "content": "10\\mathrm{m} \\times 7\\mathrm{m}" + }, + { + "bbox": [ + 313, + 421, + 541, + 554 + ], + "type": "text", + "content": " floorplan discretized at " + }, + { + "bbox": [ + 313, + 421, + 541, + 554 + ], + "type": "inline_equation", + "content": "0.1\\mathrm{m}" + }, + { + "bbox": [ + 313, + 421, + 541, + 554 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 421, + 541, + 554 + ], + "type": "inline_equation", + "content": "10^{\\circ}" + }, + { + "bbox": [ + 313, + 421, + 541, + 554 + ], + "type": "text", + "content": " increments in orientation yields a probability volume " + }, + { + "bbox": [ + 313, + 421, + 541, + 554 + ], + "type": "inline_equation", + "content": "P_{d} \\in [0,1]^{[100,70,36]}" + }, + { + "bbox": [ + 313, + 421, + 541, + 554 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 554, + 539, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 539, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 539, + 613 + ], + "type": "text", + "content": "The final camera 2D location and orientation are determined by selecting the grid cell with the highest likelihood in the probability volume. This process maximizes the posterior probability, thereby estimating the camera's " + }, + { + "bbox": [ + 313, + 554, + 539, + 613 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 554, + 539, + 613 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 554, + 539, + 613 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 554, + 539, + 613 + ], + "type": "text", + "content": " coordinates as well as its orientation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 621, + 520, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 520, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 520, + 633 + ], + "type": "text", + "content": "3.2. Adding a Semantic Prediction Network" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 639, + 541, + 722 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 541, + 722 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 541, + 722 + ], + "type": "text", + "content": "To utilize semantic cues for performing floorplan localization, we propose to add a semantic prediction network that first predicts semantic rays, and then processes these for constructing a semantic probability volume. We provide additional details in what follows, and then present an optional room type prediction component, which can be utilized if room labels are available." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "text", + "content": "27118" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 78, + 535, + 208 + ], + "blocks": [ + { + "bbox": [ + 69, + 78, + 535, + 208 + ], + "lines": [ + { + "bbox": [ + 69, + 78, + 535, + 208 + ], + "spans": [ + { + "bbox": [ + 69, + 78, + 535, + 208 + ], + "type": "image", + "image_path": "f38ed0d871f009b8ab30b71b5f53881bff7c9ffcd4a6b084b349294da124e2b3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 221, + 541, + 289 + ], + "lines": [ + { + "bbox": [ + 67, + 221, + 541, + 289 + ], + "spans": [ + { + "bbox": [ + 67, + 221, + 541, + 289 + ], + "type": "text", + "content": "Figure 2. Overview of our pipeline. The input image is processed to generate depth rays, semantic rays, and optionally additional metadata (e.g., room type prediction). We interpolate the ray predictions to a low-resolution representation and generate the depth probability volume " + }, + { + "bbox": [ + 67, + 221, + 541, + 289 + ], + "type": "inline_equation", + "content": "P_{d}" + }, + { + "bbox": [ + 67, + 221, + 541, + 289 + ], + "type": "text", + "content": " and the semantic probability volume " + }, + { + "bbox": [ + 67, + 221, + 541, + 289 + ], + "type": "inline_equation", + "content": "P_{s}" + }, + { + "bbox": [ + 67, + 221, + 541, + 289 + ], + "type": "text", + "content": " (optionally masked according to the room type prediction). These probability volumes are then fused to form the structural-semantic probability volume " + }, + { + "bbox": [ + 67, + 221, + 541, + 289 + ], + "type": "inline_equation", + "content": "P_{c}" + }, + { + "bbox": [ + 67, + 221, + 541, + 289 + ], + "type": "text", + "content": " for efficient coarse localization. Finally, we refine the candidate poses using high-resolution ray predictions and predict the final 2D camera location and orientation, visualized with an arrow on the right." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 69, + 308, + 290, + 397 + ], + "blocks": [ + { + "bbox": [ + 69, + 308, + 290, + 397 + ], + "lines": [ + { + "bbox": [ + 69, + 308, + 290, + 397 + ], + "spans": [ + { + "bbox": [ + 69, + 308, + 290, + 397 + ], + "type": "image", + "image_path": "50938694901a08d9a0703247fc0202c72c1c2e8939d5799855a233394475f580.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 68, + 406, + 296, + 473 + ], + "lines": [ + { + "bbox": [ + 68, + 406, + 296, + 473 + ], + "spans": [ + { + "bbox": [ + 68, + 406, + 296, + 473 + ], + "type": "text", + "content": "Figure 3. Overview of our semantic prediction network that predicts a set of semantic rays through the Semantic Ray Branch (top) and an optional room type value—e.g., Living Room—through the Room Classification Branch (bottom). The room type is used for extracting the mask " + }, + { + "bbox": [ + 68, + 406, + 296, + 473 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{room}}" + }, + { + "bbox": [ + 68, + 406, + 296, + 473 + ], + "type": "text", + "content": ", as visualized on the bottom right." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 494, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 494, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 67, + 494, + 295, + 601 + ], + "type": "text", + "content": "Semantic Rays Prediction. Unlike the continuous depth values estimated in prior work, the semantic rays should correspond to semantic categories, which are represented as a set of discrete classes. Therefore, we construct a network that produces a semantic ray representation " + }, + { + "bbox": [ + 67, + 494, + 295, + 601 + ], + "type": "inline_equation", + "content": "\\hat{r}_s \\in \\{1, \\dots, C\\}^l" + }, + { + "bbox": [ + 67, + 494, + 295, + 601 + ], + "type": "text", + "content": " from the image, where each ray is classified into one of " + }, + { + "bbox": [ + 67, + 494, + 295, + 601 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 67, + 494, + 295, + 601 + ], + "type": "text", + "content": " semantic categories. We provide an overview of our semantic ray prediction network in Figure 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 603, + 295, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 603, + 295, + 723 + ], + "spans": [ + { + "bbox": [ + 67, + 603, + 295, + 723 + ], + "type": "text", + "content": "As illustrated in the figure, our semantic network architecture leverages a pretrained ResNet50 backbone to extract robust, high-level features from an input RGB image " + }, + { + "bbox": [ + 67, + 603, + 295, + 723 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 67, + 603, + 295, + 723 + ], + "type": "text", + "content": ". After reducing the feature channels using a CNN and projecting them into a lower-dimensional subspace, positional encodings are computed to preserve spatial information. Two sets of learnable tokens are introduced: a set of " + }, + { + "bbox": [ + 67, + 603, + 295, + 723 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 67, + 603, + 295, + 723 + ], + "type": "text", + "content": " ray tokens responsible for predicting the semantic ray representation " + }, + { + "bbox": [ + 67, + 603, + 295, + 723 + ], + "type": "inline_equation", + "content": "\\hat{r}_s" + }, + { + "bbox": [ + 67, + 603, + 295, + 723 + ], + "type": "text", + "content": " and a single (optional) CLS token dedicated for representing global room clas" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 308, + 406, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 308, + 406, + 319 + ], + "spans": [ + { + "bbox": [ + 314, + 308, + 406, + 319 + ], + "type": "text", + "content": "sification information." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 323, + 541, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 541, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 541, + 468 + ], + "type": "text", + "content": "A single-head cross attention module integrates these tokens with the flattened spatial features, yielding refined tokens that capture both global context and local details. In the ray branch, the refined ray tokens are first processed by a self-attention block that enables each token to interact with all others, thereby aggregating complementary contextual information. The enriched tokens are then passed through an MLP to produce per-token semantic logits, which after normalization form the final semantic ray vector " + }, + { + "bbox": [ + 313, + 323, + 541, + 468 + ], + "type": "inline_equation", + "content": "\\hat{r}_s" + }, + { + "bbox": [ + 313, + 323, + 541, + 468 + ], + "type": "text", + "content": ". If room labels are available in the dataset, a similar network processes the CLS token for room type prediction, as we further detail later." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 482, + 541, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 541, + 723 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 541, + 723 + ], + "type": "text", + "content": "Estimating Semantic Probability Volume. To obtain the semantic probability map, " + }, + { + "bbox": [ + 313, + 482, + 541, + 723 + ], + "type": "inline_equation", + "content": "P_{s} \\in [0,1]^{[\\hat{H},\\hat{W},O]}" + }, + { + "bbox": [ + 313, + 482, + 541, + 723 + ], + "type": "text", + "content": ", we first need to interpolate the " + }, + { + "bbox": [ + 313, + 482, + 541, + 723 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 482, + 541, + 723 + ], + "type": "text", + "content": " predicted semantic rays. Regular linear interpolation—which prior work used for depth estimation—is unsuitable in the context of discrete labeling since interpolating between class labels can produce non-valid or semantically meaningless results. Instead, we propose a voting-based interpolation scheme: We reduce the original equiangular rays to the desired count by applying a majority vote within a small neighborhood. We use a window of three rays, assigning the label that appears most frequently in that window to the center target ray; see the supplementary material for the full algorithm. Next, we compute the score for each set of rays by taking the " + }, + { + "bbox": [ + 313, + 482, + 541, + 723 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 313, + 482, + 541, + 723 + ], + "type": "text", + "content": " difference between the predicted semantic labels and the reference labels. The score is then exponentiated and normalized to form the semantic probability volume " + }, + { + "bbox": [ + 313, + 482, + 541, + 723 + ], + "type": "inline_equation", + "content": "P_{s}" + }, + { + "bbox": [ + 313, + 482, + 541, + 723 + ], + "type": "text", + "content": ", which quantifies the likelihood of each candidate pose based on the alignment between the semantic rays and the candidate pose." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "text", + "content": "27119" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 72, + 296, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 72, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 68, + 72, + 296, + 240 + ], + "type": "text", + "content": "Room Type Prediction. In addition to predicting semantic rays, our semantic network can optionally also predict the room type, which corresponds to the room from which the input image was taken. This is achieved by processing the CLS token in the semantic network (see Figure 3). If the predicted room probability exceeds a threshold " + }, + { + "bbox": [ + 68, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{room}}" + }, + { + "bbox": [ + 68, + 72, + 296, + 240 + ], + "type": "text", + "content": ", the predicted room type is then used to construct a mask " + }, + { + "bbox": [ + 68, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{room}}" + }, + { + "bbox": [ + 68, + 72, + 296, + 240 + ], + "type": "text", + "content": " from the polygons associated with that room type. For example, if the model predicts a Living Room label with high confidence, the mask " + }, + { + "bbox": [ + 68, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{room}}" + }, + { + "bbox": [ + 68, + 72, + 296, + 240 + ], + "type": "text", + "content": " retains only the regions corresponding to the living room by setting all other areas to zero. Let " + }, + { + "bbox": [ + 68, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\tilde{P}" + }, + { + "bbox": [ + 68, + 72, + 296, + 240 + ], + "type": "text", + "content": " denote the original probability volume, then the masked probability volume " + }, + { + "bbox": [ + 68, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 68, + 72, + 296, + 240 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 251, + 219, + 264 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 251, + 219, + 264 + ], + "spans": [ + { + "bbox": [ + 146, + 251, + 219, + 264 + ], + "type": "interline_equation", + "content": "P = M _ {\\mathrm {r o o m}} \\odot \\tilde {P},", + "image_path": "12f58b72c9ed08e19bdd1ddad4ba34cdb854f37655456da323e64de5133d0c77.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 270, + 295, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 270, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 68, + 270, + 295, + 342 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 68, + 270, + 295, + 342 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 68, + 270, + 295, + 342 + ], + "type": "text", + "content": " denotes element-wise multiplication, thereby filtering out all the probabilities which are not in the living room and substantially narrowing down the search space. A detailed analysis of room type distributions and the model's prediction accuracy is provided in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 348, + 296, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 348, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 69, + 348, + 296, + 373 + ], + "type": "text", + "content": "3.3. Floorplan Localization via a Structural-Semantic Probability Volume" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 377, + 296, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 377, + 296, + 450 + ], + "spans": [ + { + "bbox": [ + 68, + 377, + 296, + 450 + ], + "type": "text", + "content": "We obtain the final probability map by generating the depth probability map " + }, + { + "bbox": [ + 68, + 377, + 296, + 450 + ], + "type": "inline_equation", + "content": "P_{d}" + }, + { + "bbox": [ + 68, + 377, + 296, + 450 + ], + "type": "text", + "content": ", following the procedure described in Section 3.1, and the semantic probability map " + }, + { + "bbox": [ + 68, + 377, + 296, + 450 + ], + "type": "inline_equation", + "content": "P_{s}" + }, + { + "bbox": [ + 68, + 377, + 296, + 450 + ], + "type": "text", + "content": ", according to the approach detailed in Section 3.2. To leverage both semantic and geometric cues, we fuse the two probability maps using a weighted combination:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 459, + 233, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 459, + 233, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 459, + 233, + 472 + ], + "type": "interline_equation", + "content": "P _ {c} = w _ {s} \\cdot P _ {s} + w _ {d} \\cdot P _ {d},", + "image_path": "0a1064e41e626e92c524d0230159098fc933b5427d7d0924a53e874a6c2e6783.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 480, + 296, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 480, + 296, + 575 + ], + "spans": [ + { + "bbox": [ + 68, + 480, + 296, + 575 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 68, + 480, + 296, + 575 + ], + "type": "inline_equation", + "content": "w_{s}" + }, + { + "bbox": [ + 68, + 480, + 296, + 575 + ], + "type": "text", + "content": " denotes the weight given to the semantic cues, while " + }, + { + "bbox": [ + 68, + 480, + 296, + 575 + ], + "type": "inline_equation", + "content": "w_{d} = 1 - w_{s}" + }, + { + "bbox": [ + 68, + 480, + 296, + 575 + ], + "type": "text", + "content": " represents the weight of the depth cues. These weights are determined using a held-out validation set, as further detailed in Section 4. Note that as both " + }, + { + "bbox": [ + 68, + 480, + 296, + 575 + ], + "type": "inline_equation", + "content": "P_{s}" + }, + { + "bbox": [ + 68, + 480, + 296, + 575 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 68, + 480, + 296, + 575 + ], + "type": "inline_equation", + "content": "P_{d}" + }, + { + "bbox": [ + 68, + 480, + 296, + 575 + ], + "type": "text", + "content": " are generated from interpolated rays, which we refer to as low-resolution rays (see Figure 4, predicted), and hence the initial probability volume " + }, + { + "bbox": [ + 68, + 480, + 296, + 575 + ], + "type": "inline_equation", + "content": "P_{c}" + }, + { + "bbox": [ + 68, + 480, + 296, + 575 + ], + "type": "text", + "content": " is also in low resolution." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 582, + 296, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 582, + 296, + 701 + ], + "spans": [ + { + "bbox": [ + 68, + 582, + 296, + 701 + ], + "type": "text", + "content": "Location Extraction. Our approach follows a coarse-to-fine strategy to achieve precise localization while maintaining computational efficiency. Given " + }, + { + "bbox": [ + 68, + 582, + 296, + 701 + ], + "type": "inline_equation", + "content": "P_{c}" + }, + { + "bbox": [ + 68, + 582, + 296, + 701 + ], + "type": "text", + "content": ", we first extract the Top-" + }, + { + "bbox": [ + 68, + 582, + 296, + 701 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 68, + 582, + 296, + 701 + ], + "type": "text", + "content": " candidate poses from the structural-semantic probability volume based on their scores, ensuring that each candidate is separated by at least " + }, + { + "bbox": [ + 68, + 582, + 296, + 701 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{res}}" + }, + { + "bbox": [ + 68, + 582, + 296, + 701 + ], + "type": "text", + "content": " in translation. Next, for each candidate, we generate an augmented set of orientations by perturbing its original angle in increments of " + }, + { + "bbox": [ + 68, + 582, + 296, + 701 + ], + "type": "inline_equation", + "content": "\\pm \\delta_{\\mathrm{ang}}" + }, + { + "bbox": [ + 68, + 582, + 296, + 701 + ], + "type": "text", + "content": " up to a maximum deviation of " + }, + { + "bbox": [ + 68, + 582, + 296, + 701 + ], + "type": "inline_equation", + "content": "\\Delta_{\\mathrm{max}}" + }, + { + "bbox": [ + 68, + 582, + 296, + 701 + ], + "type": "text", + "content": ", resulting in the final augmented set:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 119, + 710, + 246, + 723 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 710, + 246, + 723 + ], + "spans": [ + { + "bbox": [ + 119, + 710, + 246, + 723 + ], + "type": "interline_equation", + "content": "[ 0, \\pm \\delta_ {\\text {a n g}}, \\pm 2 \\delta_ {\\text {a n g}}, \\dots , \\pm \\Delta_ {\\max} ].", + "image_path": "d58cc73998d79d4bf1daf3e350dd7d361039b263894e98c8ec9778c84f4fadb4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 72, + 539, + 166 + ], + "blocks": [ + { + "bbox": [ + 316, + 72, + 539, + 166 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 539, + 166 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 539, + 166 + ], + "type": "image", + "image_path": "a651a3996ada6479ed5a323afd0cf046076b6e13ee7d32c53e35d5047ae6c4ac.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 167, + 541, + 277 + ], + "lines": [ + { + "bbox": [ + 313, + 167, + 541, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 541, + 277 + ], + "type": "text", + "content": "Figure 4. Comparing low-resolution and high-resolution rays with our coarse-to-fine approach. Given an input image (left), we construct a structural-semantic probability volume by comparing low resolution ground-truth and predicted rays (center). Location extraction from this coarse volume directly often yields significant errors, as illustrated with the yellow arrows. Results are refined only for the Top-k candidate poses by comparing the high resolution ground-truth and predicted rays (right). This allows for efficiently extracting more accurate predictions, as further detailed in Section 3.3." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 300, + 541, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 541, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 541, + 418 + ], + "type": "text", + "content": "At each candidate location, we compute the corresponding high-resolution ground-truth depth and semantic ray representations, yielding " + }, + { + "bbox": [ + 313, + 300, + 541, + 418 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 300, + 541, + 418 + ], + "type": "text", + "content": " rays per candidate location. These ground-truth rays are then compared against the original predicted " + }, + { + "bbox": [ + 313, + 300, + 541, + 418 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 300, + 541, + 418 + ], + "type": "text", + "content": " rays using a similarity metric (see supplementary material for metric details). For a visual comparison of the high-resolution ground-truth and predicted rays, please refer to Figure 4. The candidate with the highest similarity score is selected as the final predicted location." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 420, + 541, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 541, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 541, + 599 + ], + "type": "text", + "content": "This coarse-to-fine refinement process effectively leverages the full resolution of the predictions, which were initially interpolated for runtime efficiency, to achieve more precise localization. Note that there is a tradeoff between accuracy and computation time: finer resolutions yield improved precision at the expense of increased computational load. This tradeoff is illustrated in Figure 4. As can be observed, comparing low resolution rays often yields inaccurate predictions. For instance, environments with multiple semantic objects can suffer from loss of critical details (e.g., the left door is omitted in the middle row) in the low-resolution prediction, which can lead to misclassification, further motivating the refinement step. We present a quantitative analysis of this tradeoff in the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 609, + 511, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 511, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 511, + 621 + ], + "type": "text", + "content": "3.4. Training and Implementation Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 627, + 541, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 627, + 541, + 723 + ], + "spans": [ + { + "bbox": [ + 313, + 627, + 541, + 723 + ], + "type": "text", + "content": "Both networks are trained in an end-to-end manner. For depth prediction, an " + }, + { + "bbox": [ + 313, + 627, + 541, + 723 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 313, + 627, + 541, + 723 + ], + "type": "text", + "content": " loss supervises the predicted depths against ground-truth depth maps. For semantic prediction, a cross-entropy loss is applied to the predicted semantic labels. If room labels " + }, + { + "bbox": [ + 313, + 627, + 541, + 723 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 627, + 541, + 723 + ], + "type": "text", + "content": " are available in the dataset, an additional cross-entropy loss is used for the room label, and the network is trained to optimize both objectives jointly. As in prior work, data aug" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 318, + 763 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 318, + 763 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 318, + 763 + ], + "type": "text", + "content": "27120" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 72, + 294, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 72, + 294, + 193 + ], + "spans": [ + { + "bbox": [ + 68, + 72, + 294, + 193 + ], + "type": "text", + "content": "mentation techniques—including virtual roll-pitch augmentation—are employed to improve robustness to nonupright camera poses. The networks are optimized using the Adam optimizer with an initial learning rate of " + }, + { + "bbox": [ + 68, + 72, + 294, + 193 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-3}" + }, + { + "bbox": [ + 68, + 72, + 294, + 193 + ], + "type": "text", + "content": ". We use a floorplan resolution of " + }, + { + "bbox": [ + 68, + 72, + 294, + 193 + ], + "type": "inline_equation", + "content": "0.1\\mathrm{m}" + }, + { + "bbox": [ + 68, + 72, + 294, + 193 + ], + "type": "text", + "content": " and an angular granularity of " + }, + { + "bbox": [ + 68, + 72, + 294, + 193 + ], + "type": "inline_equation", + "content": "10^{\\circ}" + }, + { + "bbox": [ + 68, + 72, + 294, + 193 + ], + "type": "text", + "content": ". We set the number of predicted rays to " + }, + { + "bbox": [ + 68, + 72, + 294, + 193 + ], + "type": "inline_equation", + "content": "l = 40" + }, + { + "bbox": [ + 68, + 72, + 294, + 193 + ], + "type": "text", + "content": " and interpolate these to 7 rays during the coarse stage of localization. For the Location Extraction module, we use " + }, + { + "bbox": [ + 68, + 72, + 294, + 193 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{res}} = 0.1\\mathrm{m}" + }, + { + "bbox": [ + 68, + 72, + 294, + 193 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 68, + 72, + 294, + 193 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{ang}} = 5^{\\circ}" + }, + { + "bbox": [ + 68, + 72, + 294, + 193 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 68, + 72, + 294, + 193 + ], + "type": "inline_equation", + "content": "\\Delta_{\\mathrm{max}} = 5^{\\circ}" + }, + { + "bbox": [ + 68, + 72, + 294, + 193 + ], + "type": "text", + "content": ", and refine our results using Top-5 poses." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 203, + 122, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 203, + 122, + 216 + ], + "spans": [ + { + "bbox": [ + 69, + 203, + 122, + 216 + ], + "type": "text", + "content": "4. Results" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 224, + 294, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 224, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 68, + 224, + 294, + 380 + ], + "type": "text", + "content": "In this section, we present a comprehensive evaluation of our localization method. We begin by introducing the datasets we use in our experiments, followed by a discussion of the baselines we compare against and the evaluation metrics. The main results are reported in Section 4.1. We conduct an ablation study to assess the contributions of various components of our approach in Section 4.2. We report results for our approach under two conditions: one in which room labels are not utilized (denoted as " + }, + { + "bbox": [ + 68, + 224, + 294, + 380 + ], + "type": "inline_equation", + "content": "\\mathrm{Ours}_s" + }, + { + "bbox": [ + 68, + 224, + 294, + 380 + ], + "type": "text", + "content": ") and another in which room labels are employed to further refine the predictions (denoted as " + }, + { + "bbox": [ + 68, + 224, + 294, + 380 + ], + "type": "inline_equation", + "content": "\\mathrm{Ours}_r" + }, + { + "bbox": [ + 68, + 224, + 294, + 380 + ], + "type": "text", + "content": "). Additional experiments and qualitative results are provided in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 386, + 294, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 386, + 294, + 506 + ], + "spans": [ + { + "bbox": [ + 68, + 386, + 294, + 506 + ], + "type": "text", + "content": "Datasets. We conduct experiments on two popular datasets: Structured 3D (S3D) [37] and ZInD [9]. S3D is a synthetic dataset containing realistic 3D renders of 3,296 houses. We use the fully furnished version of S3D, as employed in previous works. ZInD consists of 1,575 unfurnished homes containing only panoramic images. We crop these panoramas to perspective views with an " + }, + { + "bbox": [ + 68, + 386, + 294, + 506 + ], + "type": "inline_equation", + "content": "80^{\\circ}" + }, + { + "bbox": [ + 68, + 386, + 294, + 506 + ], + "type": "text", + "content": " field of view (as was done in S3D), and generate a fixed-size dataset from the resulting images. For both datasets, we follow their official train and test splits." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 512, + 294, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 512, + 294, + 681 + ], + "spans": [ + { + "bbox": [ + 68, + 512, + 294, + 681 + ], + "type": "text", + "content": "Baselines. We compare our approach against two baselines: F3Loc [7] (considering the single-image localization component) and LASER [22]. We also report performance using an ORACLE ray prediction. This oracle ray prediction simulates the best possible performance achievable by our pipeline using ground truth depth and semantic rays. Note that the Oracle ray prediction baseline does not incorporate room-aware features. For F3Loc, results on the ZIND dataset are obtained from our experiments by running their publicly available code on the dataset, as the original paper does not include results on this dataset. We also use the publicly available implementation of LASER. Additional details are provided in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 686, + 294, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 686, + 294, + 723 + ], + "spans": [ + { + "bbox": [ + 68, + 686, + 294, + 723 + ], + "type": "text", + "content": "Metrics. Following prior work [7, 22], we report recall metrics computed at distance thresholds of " + }, + { + "bbox": [ + 68, + 686, + 294, + 723 + ], + "type": "inline_equation", + "content": "0.1\\mathrm{m}" + }, + { + "bbox": [ + 68, + 686, + 294, + 723 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 68, + 686, + 294, + 723 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{m}" + }, + { + "bbox": [ + 68, + 686, + 294, + 723 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 68, + 686, + 294, + 723 + ], + "type": "inline_equation", + "content": "1\\mathrm{m}" + }, + { + "bbox": [ + 68, + 686, + 294, + 723 + ], + "type": "text", + "content": ". We also report recall for predictions with an" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 328, + 70, + 526, + 269 + ], + "blocks": [ + { + "bbox": [ + 328, + 70, + 526, + 269 + ], + "lines": [ + { + "bbox": [ + 328, + 70, + 526, + 269 + ], + "spans": [ + { + "bbox": [ + 328, + 70, + 526, + 269 + ], + "type": "table", + "html": "
S3D R@
Method0.1m0.5m1m1m 30°
LASER0.76.410.48.7
F3Loc1.514.622.421.3
Ourss5.4241.8753.5252.61
Oursr5.7045.5358.7857.49
ORACLE61.0093.8494.8794.57
ZInD R@
Method0.1m0.5m1m1m 30°
LASER1.3811.0617.5513.64
F3Loc0.677.9015.0711.46
Ourss2.9824.0033.9629.30
Oursr3.3126.6038.0131.86
ORACLE26.4260.8567.6965.13
", + "image_path": "c665924910f6f0dc5c2eff4e5cc9426eadcf15f39e049037b6c4f35b20e4e44c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 271, + 541, + 327 + ], + "lines": [ + { + "bbox": [ + 313, + 271, + 541, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 541, + 327 + ], + "type": "text", + "content": "Table 1. Recall performance on the S3D and ZInD datasets. The table reports recall at thresholds of " + }, + { + "bbox": [ + 313, + 271, + 541, + 327 + ], + "type": "inline_equation", + "content": "0.1\\mathrm{m}" + }, + { + "bbox": [ + 313, + 271, + 541, + 327 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 271, + 541, + 327 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{m}" + }, + { + "bbox": [ + 313, + 271, + 541, + 327 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 271, + 541, + 327 + ], + "type": "inline_equation", + "content": "1\\mathrm{m}" + }, + { + "bbox": [ + 313, + 271, + 541, + 327 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 271, + 541, + 327 + ], + "type": "inline_equation", + "content": "1\\mathrm{m}" + }, + { + "bbox": [ + 313, + 271, + 541, + 327 + ], + "type": "text", + "content": " with a " + }, + { + "bbox": [ + 313, + 271, + 541, + 327 + ], + "type": "inline_equation", + "content": "30^{\\circ}" + }, + { + "bbox": [ + 313, + 271, + 541, + 327 + ], + "type": "text", + "content": " orientation tolerance for LASER, F3Loc, our approach without room labels " + }, + { + "bbox": [ + 313, + 271, + 541, + 327 + ], + "type": "inline_equation", + "content": "(\\mathrm{Ours}_s)" + }, + { + "bbox": [ + 313, + 271, + 541, + 327 + ], + "type": "text", + "content": ", our approach with room labels " + }, + { + "bbox": [ + 313, + 271, + 541, + 327 + ], + "type": "inline_equation", + "content": "(\\mathrm{Ours}_r)" + }, + { + "bbox": [ + 313, + 271, + 541, + 327 + ], + "type": "text", + "content": ", and an Oracle ray prediction." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 349, + 541, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 541, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 541, + 384 + ], + "type": "text", + "content": "orientation error bounded to less than " + }, + { + "bbox": [ + 313, + 349, + 541, + 384 + ], + "type": "inline_equation", + "content": "30^{\\circ}" + }, + { + "bbox": [ + 313, + 349, + 541, + 384 + ], + "type": "text", + "content": " (at the " + }, + { + "bbox": [ + 313, + 349, + 541, + 384 + ], + "type": "inline_equation", + "content": "1\\mathrm{m}" + }, + { + "bbox": [ + 313, + 349, + 541, + 384 + ], + "type": "text", + "content": " threshold). Recall is defined as the percentage of predictions that fall within these thresholds." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 393, + 449, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 449, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 449, + 406 + ], + "type": "text", + "content": "4.1.Quantitative Evaluation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 411, + 541, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 541, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 541, + 567 + ], + "type": "text", + "content": "Results are reported in Table 1. On both S3D and ZinD, our method more than doubles F3Loc's and LASER performance across all thresholds. Notably, considering S3D over the " + }, + { + "bbox": [ + 313, + 411, + 541, + 567 + ], + "type": "inline_equation", + "content": "\\mathrm{R@1m30^{\\circ}}" + }, + { + "bbox": [ + 313, + 411, + 541, + 567 + ], + "type": "text", + "content": " metric—which reflects the quality of matches between the predicted and actual camera views—in comparison to F3Loc, our method improves by more than three times. Room type predictions yield improvements of " + }, + { + "bbox": [ + 313, + 411, + 541, + 567 + ], + "type": "inline_equation", + "content": "9.2\\%" + }, + { + "bbox": [ + 313, + 411, + 541, + 567 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 411, + 541, + 567 + ], + "type": "inline_equation", + "content": "\\mathrm{R@1m30^{\\circ}}" + }, + { + "bbox": [ + 313, + 411, + 541, + 567 + ], + "type": "text", + "content": " on the S3D dataset and " + }, + { + "bbox": [ + 313, + 411, + 541, + 567 + ], + "type": "inline_equation", + "content": "8.7\\%" + }, + { + "bbox": [ + 313, + 411, + 541, + 567 + ], + "type": "text", + "content": " on the ZInD dataset. We include an additional experiment in the supplementary material that evaluates F3Loc with our refinement module, demonstrating that both the semantic rays and our refinement module provide significant performance gains." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 567, + 540, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 540, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 540, + 650 + ], + "type": "text", + "content": "We observe that significant performance improvements are also achieved for a very fine-grained recall metric of " + }, + { + "bbox": [ + 313, + 567, + 540, + 650 + ], + "type": "inline_equation", + "content": "0.1\\mathrm{m}" + }, + { + "bbox": [ + 313, + 567, + 540, + 650 + ], + "type": "text", + "content": ", boosting performance from " + }, + { + "bbox": [ + 313, + 567, + 540, + 650 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 313, + 567, + 540, + 650 + ], + "type": "text", + "content": " (F3Loc) to " + }, + { + "bbox": [ + 313, + 567, + 540, + 650 + ], + "type": "inline_equation", + "content": "5.7\\%" + }, + { + "bbox": [ + 313, + 567, + 540, + 650 + ], + "type": "text", + "content": " with our approach on S3D. We attribute this to our coarse-to-fine strategy, which effectively refines the coarse location estimates into precise predictions, as further validated in our ablation study." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 651, + 541, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 651, + 541, + 723 + ], + "spans": [ + { + "bbox": [ + 313, + 651, + 541, + 723 + ], + "type": "text", + "content": "Figure 5 presents qualitative examples that illustrate how the integration of floorplan semantics with precise depth cues enables our pipeline to effectively resolve localization ambiguities. For instance, in the third row, we can see that F3Loc, which does not use semantics, interprets this image as a blank wall, while LASER misinter" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "text", + "content": "27121" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 68, + 293, + 233 + ], + "blocks": [ + { + "bbox": [ + 71, + 68, + 293, + 233 + ], + "lines": [ + { + "bbox": [ + 71, + 68, + 293, + 233 + ], + "spans": [ + { + "bbox": [ + 71, + 68, + 293, + 233 + ], + "type": "image", + "image_path": "95882f81b1b62fdf7506fdbf047cbb5cf655a42fff36b39c99d5aed1dcf943be.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 68, + 236, + 296, + 291 + ], + "lines": [ + { + "bbox": [ + 68, + 236, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 68, + 236, + 296, + 291 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison of our method with F3Loc and LASER. Warmer colors correspond to regions with higher probabilities. Below each map we report the localization error in meters and degrees. We use arrows to visualize the ground truth location (magenta) and the predicted location (white)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 313, + 294, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 313, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 69, + 313, + 294, + 326 + ], + "type": "text", + "content": "prets the window size and predicts an incorrect location." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 333, + 162, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 333, + 162, + 346 + ], + "spans": [ + { + "bbox": [ + 69, + 333, + 162, + 346 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 351, + 295, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 351, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 68, + 351, + 295, + 554 + ], + "type": "text", + "content": "We demonstrate the effect of incorporating each component of our method on overall localization performance. Specifically, we conduct the following ablations: (1) Base, which corresponds to computing the structural-semantic probability volume and selecting the maximum probability without any additional refinement. (2) Removing semantic interpolation (denoted as -Interpolation), where we replace our majority voting interpolation with a simple linear interpolation followed by rounding. (3) Adding room predictions (denoted as +Room), where we assess the effect of integrating room type predictions into our localization pipeline. (4) Adding refinement (denoted as +Refine), which assesses our coarse-to-fine approach, which refines our localization extraction via the Top-K candidate poses. (5) Adding room predictions and refinement (denoted as +Room&Refine), which combines both components." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 555, + 296, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 555, + 296, + 723 + ], + "spans": [ + { + "bbox": [ + 68, + 555, + 296, + 723 + ], + "type": "text", + "content": "From the results in Table 2, we can see that on the " + }, + { + "bbox": [ + 68, + 555, + 296, + 723 + ], + "type": "inline_equation", + "content": "1\\mathrm{m}30^{\\circ}" + }, + { + "bbox": [ + 68, + 555, + 296, + 723 + ], + "type": "text", + "content": " metric, the addition of our refinement module improves performance by " + }, + { + "bbox": [ + 68, + 555, + 296, + 723 + ], + "type": "inline_equation", + "content": "8.6\\%" + }, + { + "bbox": [ + 68, + 555, + 296, + 723 + ], + "type": "text", + "content": " relative to the baseline. This indicates that a substantial amount of information is lost during the initial interpolation process if not properly addressed, thereby strongly motivating the use of coarse-to-fine strategies in our Location Extraction module. We also observe a significant improvement of " + }, + { + "bbox": [ + 68, + 555, + 296, + 723 + ], + "type": "inline_equation", + "content": "11.5\\%" + }, + { + "bbox": [ + 68, + 555, + 296, + 723 + ], + "type": "text", + "content": " from the room prediction component, which is discussed in further detail in the supplementary material and visualized in Figure 6. Finally, our majority voting interpolation approach contributes an additional gain of " + }, + { + "bbox": [ + 68, + 555, + 296, + 723 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 68, + 555, + 296, + 723 + ], + "type": "text", + "content": " compared to a simple interpolation strategy. When combining all these improvements, our method achieves" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 316, + 69, + 540, + 145 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 540, + 145 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 540, + 145 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 540, + 145 + ], + "type": "table", + "html": "
Method0.1m0.5m1m1m 30°
Base4.6538.3549.4048.44
- Interpolation4.7338.4448.9147.99
+ Room5.1242.9255.5754.04
+ Refine5.4241.8753.5252.61
+ Room&Refine5.7045.5358.7857.49
", + "image_path": "10a493b75ee35579609dd4b728f19874a6442cfc0f9f9772bab6c39c90d0bb91.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 146, + 540, + 169 + ], + "lines": [ + { + "bbox": [ + 313, + 146, + 540, + 169 + ], + "spans": [ + { + "bbox": [ + 313, + 146, + 540, + 169 + ], + "type": "text", + "content": "Table 2. Ablation study, evaluating the effect of incorporating the different components in our pipeline on the S3D dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 315, + 182, + 537, + 292 + ], + "blocks": [ + { + "bbox": [ + 315, + 182, + 537, + 292 + ], + "lines": [ + { + "bbox": [ + 315, + 182, + 537, + 292 + ], + "spans": [ + { + "bbox": [ + 315, + 182, + 537, + 292 + ], + "type": "image", + "image_path": "29b668141624274f1c51c29320cb2b933f770b5b1e8861e9b87f091d4c1df815.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 295, + 541, + 361 + ], + "lines": [ + { + "bbox": [ + 313, + 295, + 541, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 541, + 361 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison of using room predictions and our coarse-to-fine refinement approach. Below each map we report the localization error in meters and degrees. Warmer colors correspond to regions with higher probabilities. Overlaid on the estimated probabilities, we show the ground truth location (magenta) and the predicted location." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 384, + 540, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 540, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 540, + 407 + ], + "type": "text", + "content": "an overall enhancement of " + }, + { + "bbox": [ + 313, + 384, + 540, + 407 + ], + "type": "inline_equation", + "content": "18.6\\%" + }, + { + "bbox": [ + 313, + 384, + 540, + 407 + ], + "type": "text", + "content": " on the 1m " + }, + { + "bbox": [ + 313, + 384, + 540, + 407 + ], + "type": "inline_equation", + "content": "30^{\\circ}" + }, + { + "bbox": [ + 313, + 384, + 540, + 407 + ], + "type": "text", + "content": " metric relative to our base model." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 409, + 541, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 541, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 541, + 480 + ], + "type": "text", + "content": "In Figure 6, we illustrate the impact of incorporating room type predictions alongside the location extraction module. The figure clearly demonstrates how these components refine the probability volume by narrowing down the candidate poses, which in turn improves overall localization accuracy." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 482, + 541, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 541, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 541, + 590 + ], + "type": "text", + "content": "We analyze the impact of different combinations of predicted depth and semantic features on the S3D dataset. Figure 7 summarizes the recall performance for various depth and semantic weight configurations used to compute the structural-semantic probability volume. In Figure 8 we visualize examples from two extreme scenarios (depth only and semantic only) and from the configuration adopted in our work " + }, + { + "bbox": [ + 313, + 482, + 541, + 590 + ], + "type": "inline_equation", + "content": "\\left([w_{s}, w_{d}] = [0.4, 0.6]\\right)" + }, + { + "bbox": [ + 313, + 482, + 541, + 590 + ], + "type": "text", + "content": ", selected according to the validation set." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 591, + 541, + 721 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 591, + 541, + 721 + ], + "spans": [ + { + "bbox": [ + 313, + 591, + 541, + 721 + ], + "type": "text", + "content": "As can be observed from Figure 8 (and also reflected in Figure 7), relying solely on semantic rays tends to produce a more diffused probability volume with multiple ambiguous candidate locations. This occurs because most images contain a single semantic object of a standard size, allowing an accurate ray pattern to be fitted in more than one location. Similarly, depth cues alone also suffer from ambiguity, particularly in repetitive environments or when the image captures three or fewer walls, which leads to uncertainty in the localization estimate. However, when these semantic cues are com" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "text", + "content": "27122" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 70, + 294, + 229 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 294, + 229 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 294, + 229 + ], + "type": "image", + "image_path": "c2dfec3c57efe79de669b4ba1270029e8c3f6e09f854cb967d0b2f1cdb4e7d82.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 68, + 239, + 296, + 306 + ], + "lines": [ + { + "bbox": [ + 68, + 239, + 296, + 306 + ], + "spans": [ + { + "bbox": [ + 68, + 239, + 296, + 306 + ], + "type": "text", + "content": "Figure 7. Recall vs. weight combinations on the S3D test set. The plot shows the recall metrics for four different thresholds: " + }, + { + "bbox": [ + 68, + 239, + 296, + 306 + ], + "type": "inline_equation", + "content": "0.1m" + }, + { + "bbox": [ + 68, + 239, + 296, + 306 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 68, + 239, + 296, + 306 + ], + "type": "inline_equation", + "content": "0.5m" + }, + { + "bbox": [ + 68, + 239, + 296, + 306 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 68, + 239, + 296, + 306 + ], + "type": "inline_equation", + "content": "1m" + }, + { + "bbox": [ + 68, + 239, + 296, + 306 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 68, + 239, + 296, + 306 + ], + "type": "inline_equation", + "content": "1m30^{\\circ}" + }, + { + "bbox": [ + 68, + 239, + 296, + 306 + ], + "type": "text", + "content": ". The x-axis displays the weight combinations in the order " + }, + { + "bbox": [ + 68, + 239, + 296, + 306 + ], + "type": "inline_equation", + "content": "(w_{s}, w_{d})" + }, + { + "bbox": [ + 68, + 239, + 296, + 306 + ], + "type": "text", + "content": ". A vertical dashed line at " + }, + { + "bbox": [ + 68, + 239, + 296, + 306 + ], + "type": "inline_equation", + "content": "w_{s} = 0.4" + }, + { + "bbox": [ + 68, + 239, + 296, + 306 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 68, + 239, + 296, + 306 + ], + "type": "inline_equation", + "content": "w_{d} = 0.6" + }, + { + "bbox": [ + 68, + 239, + 296, + 306 + ], + "type": "text", + "content": " highlights the weight combination selected using the validation set." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 71, + 317, + 293, + 487 + ], + "blocks": [ + { + "bbox": [ + 71, + 317, + 293, + 487 + ], + "lines": [ + { + "bbox": [ + 71, + 317, + 293, + 487 + ], + "spans": [ + { + "bbox": [ + 71, + 317, + 293, + 487 + ], + "type": "image", + "image_path": "5dc6920b295726ffc7ed155da72789d9b30d5f087949db643d7f4207bff66beb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 68, + 490, + 296, + 544 + ], + "lines": [ + { + "bbox": [ + 68, + 490, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 68, + 490, + 296, + 544 + ], + "type": "text", + "content": "Figure 8. Qualitative Comparison of Depth-Only, Semantic-Only, and Fused Structural-Semantic Probability Volumes. Below each map we report the localization error in meters and degrees. Warmer colors correspond to regions with higher probabilities." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 567, + 294, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 567, + 294, + 614 + ], + "spans": [ + { + "bbox": [ + 68, + 567, + 294, + 614 + ], + "type": "text", + "content": "bined with depth rays, the resulting probability volume becomes significantly more concentrated. This integration effectively filters out spurious candidates and sharpens the localization estimate." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 616, + 294, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 616, + 294, + 675 + ], + "spans": [ + { + "bbox": [ + 68, + 616, + 294, + 675 + ], + "type": "text", + "content": "Additionally, the supplementary material offers a detailed analysis of how varying the Top-K candidates influences localization refinement, along with additional experiments that motivate various design choices, such as using hard thresholds in multiple steps of our pipeline." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 683, + 148, + 695 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 683, + 148, + 695 + ], + "spans": [ + { + "bbox": [ + 69, + 683, + 148, + 695 + ], + "type": "text", + "content": "Runtime Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 698, + 294, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 698, + 294, + 723 + ], + "spans": [ + { + "bbox": [ + 69, + 698, + 294, + 723 + ], + "type": "text", + "content": "We report runtime performance breakdown in Table 3, using the same parameters employed in Table 1 with" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 71, + 538, + 136 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 538, + 136 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 538, + 136 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 538, + 136 + ], + "type": "table", + "html": "
KPredictionLocRefinTotal
10.038 ± 0.1190.174 ± 0.0330.141 ± 0.0730.364 ± 0.142
30.033 ± 0.0930.154 ± 0.0260.356 ± 0.1190.554 ± 0.157
50.034 ± 0.1030.155 ± 0.0290.577 ± 0.1850.778 ± 0.218
", + "image_path": "c4ff0c1de34ba0cda5cb5eade46e3c8ff35eea79f3b999ca06da73d775f6f102.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 144, + 541, + 200 + ], + "lines": [ + { + "bbox": [ + 313, + 144, + 541, + 200 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 541, + 200 + ], + "type": "text", + "content": "Table 3. Performance breakdown over different Top-K values. Each entry is mean ± std (s). Prediction denotes the ray predictions, Loc refers to the localization process, and Refin represents the refinement stage, during which candidate locations are evaluated by computing the ground truth rays." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 222, + 541, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 222, + 541, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 222, + 541, + 400 + ], + "type": "text", + "content": "varying K values. As shown in the table, the per-image inference time increases with the number of candidate refinements, " + }, + { + "bbox": [ + 313, + 222, + 541, + 400 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 222, + 541, + 400 + ], + "type": "text", + "content": ", reflecting the additional computations required during refinement. All experiments were conducted on a single CPU without multithreading to avoid introducing bias. Future work could explore parallelizing the refinement stage by computing all ground-truth rays simultaneously for further speed-up. Importantly, even at " + }, + { + "bbox": [ + 313, + 222, + 541, + 400 + ], + "type": "inline_equation", + "content": "K = 5" + }, + { + "bbox": [ + 313, + 222, + 541, + 400 + ], + "type": "text", + "content": " the setting we adopt in our work, the inference time remains reasonable, striking a practical balance between computational cost and improvements in localization accuracy. We further observe that the prediction and localization steps require similar amounts of time, while the refinement step grows monotonically as " + }, + { + "bbox": [ + 313, + 222, + 541, + 400 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 222, + 541, + 400 + ], + "type": "text", + "content": " increases." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 414, + 387, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 414, + 387, + 426 + ], + "spans": [ + { + "bbox": [ + 314, + 414, + 387, + 426 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 434, + 541, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 541, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 541, + 542 + ], + "type": "text", + "content": "In this work, we presented a semantic-aware localization framework that extends floorplan-based camera localization by fusing semantic labels with geometric depth cues. Our approach leverages a novel semantic ray prediction network alongside an established depth estimation method to generate a semantic-structural probability volume, which significantly improves localization accuracy, especially in environments with repetitive or ambiguous structural patterns." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 543, + 540, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 540, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 540, + 638 + ], + "type": "text", + "content": "Our extensive experiments on the S3D and ZInD datasets demonstrate that integrating semantic cues effectively resolves depth-based ambiguities and consistently outperforms state-of-the-art methods such as F3Loc and LASER. Ablation studies confirm that a balanced combination of depth and semantic information, coupled with a coarse-to-fine localization strategy and the use of room labels, yields optimal performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 639, + 541, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 541, + 723 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 541, + 723 + ], + "type": "text", + "content": "Looking forward, extending our framework to incorporate additional semantic labels and other modalities, such as textual information, promises to further enhance localization robustness in challenging indoor settings. In general, our approach represents an important step towards accurate and reliable indoor localization systems by effectively leveraging semantic and geometric cues." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "text", + "content": "27123" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 70, + 71, + 128, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 71, + 128, + 83 + ], + "spans": [ + { + "bbox": [ + 70, + 71, + 128, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 70, + 91, + 296, + 723 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 75, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 75, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Relja Arandjelovic, Petr Gronat, Akihiko Torii, Tomas Pajdla, and Josef Sivic. NetVLAD: CNN architecture for weakly supervised place recognition. In CVPR, pages 5297-5307, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 75, + 137, + 296, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 137, + 296, + 169 + ], + "spans": [ + { + "bbox": [ + 75, + 137, + 296, + 169 + ], + "type": "text", + "content": "[2] Vassileios Balntas, Shuda Li, and Victor Prisacariu. Relocnet: Continuous metric learning relocalisation using neural nets. In ECCV, pages 751-767, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 171, + 296, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 171, + 296, + 225 + ], + "spans": [ + { + "bbox": [ + 75, + 171, + 296, + 225 + ], + "type": "text", + "content": "[3] Federico Boniardi, Tim Caselitz, Rainer Kummerle, and Wolfram Burgard. Robust lidar-based localization in architectural floor plans. In 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3318-3324. IEEE, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 75, + 227, + 296, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 227, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 75, + 227, + 296, + 281 + ], + "type": "text", + "content": "[4] Federico Boniardi, Tim Caselitz, Rainer Kummerle, and Wolfram Burgard. A pose graph-based localization system for long-term navigation in cad floor plans. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 84-97. IEEE, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 75, + 283, + 296, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 283, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 75, + 283, + 296, + 348 + ], + "type": "text", + "content": "[5] Federico Boniardi, Abhinav Valada, Rohit Mohan, Tim Caselitz, and Wolfram Burgard. Robot localization in floor plans using a room layout edge extraction network. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5291-5297. IEEE, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 75, + 350, + 296, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 350, + 296, + 404 + ], + "spans": [ + { + "bbox": [ + 75, + 350, + 296, + 404 + ], + "type": "text", + "content": "[6] Eric Brachmann, Alexander Krull, Sebastian Nowozin, Jamie Shotton, Frank Michel, Stefan Gumhold, and Carsten Rother. DSAC-differentiable RANSAC for camera localization. In CVPR, pages 6684-6692, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 75, + 407, + 296, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 407, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 75, + 407, + 296, + 449 + ], + "type": "text", + "content": "[7] Changan Chen, Rui Wang, Christoph Vogel, and Marc Pollefeys. F3Loc: Fusion and filtering for floorplan localization. arXiv preprint arXiv:2403.03370, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 75, + 452, + 296, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 452, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 75, + 452, + 296, + 485 + ], + "type": "text", + "content": "[8] Hang Chu, Dong Ki Kim, and Tsuhan Chen. You are here: Mimicking the human thinking process in reading floorplans. In ICCV, pages 2210-2218, 2015. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 75, + 487, + 296, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 487, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 75, + 487, + 296, + 540 + ], + "type": "text", + "content": "[9] Steve Cruz, Will Hutchcroft, Yuguang Li, Naji Khosravan, Ivaylo Boyadzhiev, and Sing Bing Kang. Zillow indoor dataset: Annotated floor plans with " + }, + { + "bbox": [ + 75, + 487, + 296, + 540 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 75, + 487, + 296, + 540 + ], + "type": "text", + "content": " panoramas and 3d room layouts. In CVPR, pages 2133-2143, 2021. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 70, + 543, + 296, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 543, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 70, + 543, + 296, + 586 + ], + "type": "text", + "content": "[10] Niklas Gard, Anna Hilsmann, and Peter Eisert. Spvloc: Semantic panoramic viewpoint matching for 6d camera localization in unseen environments. arXiv preprint arXiv:2404.10527, 2024.3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 70, + 588, + 296, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 588, + 296, + 631 + ], + "spans": [ + { + "bbox": [ + 70, + 588, + 296, + 631 + ], + "type": "text", + "content": "[11] B. Glocker, S. Izadi, J. Shotton, and A. Criminisi. Realtime RGB-D camera relocalization. In Proceedings of the Mixed and Augmented Reality (ISMAR) Conference. IEEE, 2013. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 70, + 633, + 296, + 676 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 633, + 296, + 676 + ], + "spans": [ + { + "bbox": [ + 70, + 633, + 296, + 676 + ], + "type": "text", + "content": "[12] Henry Howard-Jenkins and Victor Adrian Prisacariu. LaLaLoc++: Global floor plan comprehension for layout localisation in unvisited environments. In ECCV, pages 693-709, 2022. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 70, + 678, + 296, + 723 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 678, + 296, + 723 + ], + "spans": [ + { + "bbox": [ + 70, + 678, + 296, + 723 + ], + "type": "text", + "content": "[13] Henry Howard-Jenkins, Jose-Raul Ruiz-Sarmiento, and Victor Adrian Prisacariu. LaLaLoc: Latent layout localisation in dynamic, unvisited environments. arXiv preprint arXiv:2104.09169, 2021. 1, 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 541, + 723 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 541, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 541, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 541, + 138 + ], + "type": "text", + "content": "[14] Seigo Ito, Felix Endres, Markus Kuderer, Gian Diego Tipaldi, Cyril Stachniss, and Wolfram Burgard. W-RGBD: Floor-plan-based indoor global localization using a depth camera and wifi. In 2014 IEEE International Conference on Robotics and Automation (ICRA), pages 417-422. IEEE, 2014. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 541, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 541, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 541, + 183 + ], + "type": "text", + "content": "[15] Peter Karkus, David Hsu, and Wee Sun Lee. Particle filter networks with application to visual localization. In Proceedings of the 2nd Conference on Robot Learning (CoRL), pages 169-178. PMLR, 2018. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 184, + 541, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 541, + 225 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 541, + 225 + ], + "type": "text", + "content": "[16] Alex Kendall, Matthew Grimes, and Roberto Cipolla. PoseNet: A convolutional network for real-time 6-DOF camera relocalization. In ICCV, pages 2938-2946, 2015. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 227, + 541, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 227, + 541, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 227, + 541, + 270 + ], + "type": "text", + "content": "[17] Junho Kim, Jiwon Jeong, and Young Min Kim. Fully geometric panoramic localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 271, + 541, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 271, + 541, + 325 + ], + "spans": [ + { + "bbox": [ + 316, + 271, + 541, + 325 + ], + "type": "text", + "content": "[18] Z. Li, M. H. Ang, and D. Rus. Online localization with imprecise floor space maps using stochastic gradient descent. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8571-8578. IEEE, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 327, + 541, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 541, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 541, + 358 + ], + "type": "text", + "content": "[19] Liu Liu, Hongdong Li, and Yuchao Dai. Efficient global 2D-3D matching for camera localization in a large-scale 3D map. In ICCV, pages 2372-2381, 2017. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 360, + 541, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 360, + 541, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 541, + 403 + ], + "type": "text", + "content": "[20] Pranay Mathur, Rajesh Kumar, and Sarthak Upadhyay. Sparse image-based navigation architecture to mitigate the need of precise localization in mobile robots. arXiv preprint arXiv:2203.15272, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 404, + 541, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 541, + 446 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 541, + 446 + ], + "type": "text", + "content": "[21] Oscar Mendez, Simon Hadfield, Nicolas Pugeault, and Richard Bowden. SeDAR: Reading floorplans like a human—using deep learning to enable human-inspired localisation. *IJCV*, 128:1286–1310, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 448, + 541, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 448, + 541, + 501 + ], + "spans": [ + { + "bbox": [ + 316, + 448, + 541, + 501 + ], + "type": "text", + "content": "[22] Zhixiang Min, Naji Khosravan, Zachary Bessinger, Manjunath Narayana, Sing Bing Kang, Enrique Dunn, and Ivaylo Boyadzhiev. LASER: Latent space rendering for 2D visual localization. In CVPR, pages 11122-11131, 2022. 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 502, + 541, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 502, + 541, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 502, + 541, + 567 + ], + "type": "text", + "content": "[23] R. A. Newcombe, A. J. Davison, S. Izadi, P. Kohli, O. Hilliges, J. Shotton, D. Molyneaux, S. Hodges, D. Kim, and A. Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In Proceedings of the Mixed and Augmented Reality (ISMAR) Conference. IEEE, 2011. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 569, + 541, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 569, + 541, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 569, + 541, + 611 + ], + "type": "text", + "content": "[24] Takahiro Niwa, Shun Taguchi, and Noriaki Hirose. Spatio-temporal graph localization networks for image-based navigation. arXiv preprint arXiv:2204.13237, 2022.1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 613, + 541, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 541, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 541, + 655 + ], + "type": "text", + "content": "[25] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: robust hierarchical localization at large scale. In CVPR, pages 12716-12725, 2019. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 656, + 541, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 656, + 541, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 656, + 541, + 689 + ], + "type": "text", + "content": "[26] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Fast image-based localization using direct 2d-to-3d matching. In ICCV, pages 667-674, 2011." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 689, + 541, + 723 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 689, + 541, + 723 + ], + "spans": [ + { + "bbox": [ + 316, + 689, + 541, + 723 + ], + "type": "text", + "content": "[27] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Improving image-based localization by active correspondence search. In ECCV, pages 752-765, 2012." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 318, + 763 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 318, + 763 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 318, + 763 + ], + "type": "text", + "content": "27124" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 70, + 72, + 296, + 552 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 70, + 72, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 72, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 70, + 72, + 296, + 116 + ], + "type": "text", + "content": "[28] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Efficient & effective prioritized matching for large-scale image-based localization. IEEE TPAMI, 39(9):1744-1756, 2016. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 71, + 118, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 118, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 71, + 118, + 295, + 162 + ], + "type": "text", + "content": "[29] Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene coordinate regression forests for camera relocalization in rgb-d images. In CVPR, pages 2930-2937, 2013. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 163, + 295, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 163, + 295, + 206 + ], + "spans": [ + { + "bbox": [ + 70, + 163, + 295, + 206 + ], + "type": "text", + "content": "[30] Janine Thoma, Danda Pani Paudel, Ajad Chhatkuli, Thomas Probst, and Luc Van Gool. Mapping, localization and path planning for image-based navigation using visual features and map. In CVPR. IEEE, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 71, + 208, + 294, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 208, + 294, + 261 + ], + "spans": [ + { + "bbox": [ + 71, + 208, + 294, + 261 + ], + "type": "text", + "content": "[31] Julien Valentin, Matthias Nießner, Jamie Shotton, Andrew Fitzgibbon, Shahram Izadi, and Philip HS Torr. Exploiting uncertainty in regression forests for accurate camera relocalization. In CVPR, pages 4400-4408, 2015. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 71, + 263, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 263, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 71, + 263, + 295, + 308 + ], + "type": "text", + "content": "[32] Florian Walch, Caner Hazirbas, Laura Leal-Taixe, Torsten Sattler, Sebastian Hilsenbeck, and Daniel Cremers. Image-based localization using lstms for structured feature correlation. In ICCV, pages 627-637, 2017. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 71, + 309, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 309, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 71, + 309, + 294, + 341 + ], + "type": "text", + "content": "[33] Shenlong Wang, Sanja Fidler, and Raquel Urtasun. Lost shopping! monocular localization in large indoor spaces. In ICCV, pages 2695-2703, 2015. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 342, + 295, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 342, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 70, + 342, + 295, + 396 + ], + "type": "text", + "content": "[34] Xipeng Wang, Ryan J Marcotte, and Edwin Olson. GLFP: Global localization from a floor plan. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1627-1632. IEEE, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 70, + 398, + 295, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 398, + 295, + 463 + ], + "spans": [ + { + "bbox": [ + 70, + 398, + 295, + 463 + ], + "type": "text", + "content": "[35] Wera Winterhalter, Freya Fleckenstein, Bastian Steder, Luciano Spinello, and Wolfram Burgard. Accurate indoor localization for RGB-D smartphones and tablets given 2d floor plans. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3138-3143. IEEE, 2015. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 70, + 465, + 295, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 465, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 70, + 465, + 295, + 508 + ], + "type": "text", + "content": "[36] Jian Wu, Liwei Ma, and Xiaolin Hu. Delving deeper into convolutional neural networks for camera relocalization. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 5644-5651. IEEE, 2017. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 70, + 510, + 295, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 510, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 70, + 510, + 295, + 552 + ], + "type": "text", + "content": "[37] Jia Zheng, Junfei Zhang, Jing Li, Rui Tang, Shenghua Gao, and Zihan Zhou. Structured3d: A large photorealistic dataset for structured 3d modeling. In ECCV, 2020.6" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "text", + "content": "27125" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Superpowering Open-Vocabulary Object Detectors for X-ray Vision/60987d09-4da2-484e-8cc1-bea5f240eaf8_content_list.json b/2025/Superpowering Open-Vocabulary Object Detectors for X-ray Vision/60987d09-4da2-484e-8cc1-bea5f240eaf8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c5484906058000ca358d979ff9a12af75c0c1b10 --- /dev/null +++ b/2025/Superpowering Open-Vocabulary Object Detectors for X-ray Vision/60987d09-4da2-484e-8cc1-bea5f240eaf8_content_list.json @@ -0,0 +1,1423 @@ +[ + { + "type": "text", + "text": "Superpowering Open-Vocabulary Object Detectors for X-ray Vision", + "text_level": 1, + "bbox": [ + 156, + 130, + 839, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pablo Garcia-Fernandez $^{1*}$ Lorenzo Vaquero $^{1,2}$ Mingxuan Liu $^{3}$ Feng Xue $^{3*}$ Daniel Cores $^{1}$ Nicu Sebe $^{3}$ Manuel Mucientes $^{1}$ Elisa Ricci $^{2,3}$", + "bbox": [ + 173, + 160, + 821, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Santiago de Compostela, Spain 2Fondazione Bruno Kessler, Italy 3University of Trento, Italy", + "bbox": [ + 173, + 202, + 821, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{pablogarcia.fernandez, daniel.cores, manuel.mucientes}@usc.es lvaquerootal@fbk.eu", + "bbox": [ + 129, + 241, + 867, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{mingxuan.liu, feng.xue, niculae.sebe, e.ricci}@unitn.it *Corresponding authors", + "bbox": [ + 173, + 255, + 821, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 291, + 325, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Open-vocabulary object detection (OvOD) is set to revolutionize security screening by enabling systems to recognize any item in X-ray scans. However, developing effective OvOD models for X-ray imaging presents unique challenges due to data scarcity and the modality gap that prevents direct adoption of RGB-based solutions. To overcome these limitations, we propose RAXO, a training-free framework that repurposes off-the-shelf RGB OvOD detectors for robust X-ray detection. RAXO builds high-quality X-ray class descriptors using a dual-source retrieval strategy. It gathers relevant RGB images from the web and enriches them via a novel X-ray material transfer mechanism, eliminating the need for labeled databases. These visual descriptors replace text-based classification in OvOD, leveraging intra-modal feature distances for robust detection. Extensive experiments demonstrate that RAXO consistently improves OvOD performance, providing an average mAP increase of up to 17.0 points over base detectors. To further support research in this emerging field, we also introduce DET-COMPASS, a new benchmark featuring bounding box annotations for over 300 object categories, enabling large-scale evaluation of OvOD in X-ray. Code and dataset available at: https://pagf188.github.io/RAXO/.", + "bbox": [ + 88, + 321, + 483, + 670 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 693, + 220, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Automated object detection technologies for X-ray imaging are essential to maintain public safety, enabling the identification of prohibited items at checkpoints in high-risk environments such as airports, train stations, museums and stadiums [29]. These systems improve security while simultaneously reducing the workload of human inspectors.", + "bbox": [ + 89, + 719, + 482, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Conventional X-ray object detectors rely on supervised learning [2, 14, 20] and are inherently limited by the object categories present in their training datasets (see Fig. 1a). This limitation is exacerbated by the high cost of X-ray machinery and the requirement for expert annotation, often restricting these systems to fewer than 20 object classes [32],", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/25cf5ef969acd77759db2ca01adf2e5caba7178cbccd2f42808b19029b75cc47.jpg", + "image_caption": [ + "Figure 1. (a) Traditional X-ray object detectors are constrained by the limited categories in their training datasets. (b) We introduce the task of open-vocabulary object detection (OvOD) for X-ray imaging and propose RAXO, a training-free method that adapts off-the-shelf RGB OvOD models to X-ray data. (c) RAXO greatly improves detection performance across multiple benchmarks." + ], + "image_footnote": [], + "bbox": [ + 532, + 287, + 883, + 545 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "thereby hindering broader real-world applications.", + "bbox": [ + 511, + 643, + 844, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In light of the expanding diversity of man-made objects and evolving security demands, an open-vocabulary object detection (OvOD) framework capable of recognizing arbitrary X-ray object categories defined by the user is imperative. Despite its importance, open-vocabulary detection in the X-ray domain remains largely unexplored. Concurrent works aiming to extend beyond base categories have managed to generalize to at most four unseen classes [16], thereby highlighting the inherent challenges of the task.", + "bbox": [ + 511, + 657, + 905, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in OvOD for conventional RGB images have been driven by large-scale annotated datasets, which facilitate effective alignment between visual and textual features [15, 18, 44]. However, these advancements do not directly transfer to X-ray imagery. In practice, applying state-of-the-art RGB-based OvOD models to X-ray scans leads to significant performance degradation, as illustrated", + "bbox": [ + 511, + 795, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20770", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in Fig. 1c. Moreover, retraining these models on X-ray data is often impractical given the scarcity of large-scale annotated X-ray datasets.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by these challenges, this work opens up open-vocabulary object detection for X-ray imaging by repurposing robust, off-the-shelf RGB-based detectors. To this end, we propose RAXO (tRAining-free adaptation for X-ray Open-vocabulary detection), a training-free method that seamlessly adapts RGB OvOD models to X-ray. (Fig. 1b).", + "bbox": [ + 89, + 136, + 480, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We find that the main reason RGB-based OvOD detectors fail in the X-ray domain is the disruption of text-visual feature alignment. The appearance disparity between the same object in RGB and X-ray modalities causes textual embeddings, aligned with RGB features, to mismatch with X-ray visual features. RAXO addresses this issue on three simple steps: (1) visual sample acquisition, which involves obtaining X-ray images that represent user-defined categories; (2) class descriptor modeling, which uses these images to construct descriptors that effectively encode the visual appearance of each category within the X-ray domain; and (3) classifier construction, where the computed descriptors are used to build a visual classifier that replaces the inter-modal (i.e., text-to-visual) classification of conventional OvODs. This new classifier exploits the fact that, despite the modality shift, intra-modal (e.g., visual-to-visual) feature distances remain reliable indicators for object identification. In this way RAXO enables any OvOD to successfully detect X-ray objects based on their inherent visual properties.", + "bbox": [ + 91, + 227, + 482, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To facilitate comprehensive evaluation, we also introduce DET-COMPASS, a detection dataset comprising 370 distinct object classes with paired X-ray and RGB annotations. Extensive experiments demonstrate that our approach consistently improves detection performance by an average of $\\uparrow 8.4$ AP across multiple benchmarks and OvOD models, with gains that scale as the underlying detectors become more powerful (Tab. 2). In summary, our contributions are:", + "bbox": [ + 89, + 513, + 482, + 633 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We formally introduce the problem of open-vocabulary object detection for X-ray imagery without training, addressing a critical need for security in real-world scenarios.", + "- We introduce DET-COMPASS, a novel benchmark with bounding box annotations across 370 object categories, enabling standardized evaluation of OvOD methods in the X-ray modality.", + "- We propose RAXO, a novel training-free approach that repurposes off-the-shelf RGB-based OvOD models for X-ray detection by constructing robust visual descriptors. RAXO achieves new state-of-the-art results across multiple benchmarks." + ], + "bbox": [ + 89, + 635, + 482, + 815 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 829, + 232, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Open-vocabulary object detection has advanced rapidly with the advent of Vision-Language Models (VLMs), enabling detectors to generalize beyond fixed categories.", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "OvOD detectors leverage weak supervision signals to improve detection accuracy. Based on the type of weak supervision utilized, OvOD methods can be categorized into (i) region-aware training, (ii) pseudo-labeling, (iii) knowledge distillation, and (iv) transfer learning.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Region-aware training methods aim to improve localization and feature representation by refining the alignment between image regions and their corresponding textual descriptions. Approaches such as DetCLIP [40], DetCLIPv2 [41], CORA [39], and VLDet [15] adopt this strategy. Pseudo-labeling methods, rely on large pretrained VLMs to generate pseudo-labels, effectively expanding the training set. Methods like RegionCLIP [43], PromptDet [6], CoDET [21], GLIP [13], Detric [44], and Grounding DINO [18] follow this approach. Knowledge distillation techniques, such as BARON [37], DK-DETR [12], CLIPSelf [38], and SIC-CADS [5], employ VLMs as teachers in a teacher-student framework, transferring knowledge from the VLM image encoder to enhance the detector backbone. In contrast, transfer learning approaches integrate the VLM encoder directly, either through fine-tuning, as seen in OWL-ViT [23], or by freezing the encoder, as in F-VLM [11].", + "bbox": [ + 511, + 169, + 903, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Regardless of the training strategy, OvOD methods heavily depend on VLMs trained on large-scale datasets. However, such datasets are unavailable in the X-ray modality, making this approach impractical. To address this, RAXO seamlessly adapts existing RGB-based OvOD methods for X-ray object detection without requiring additional training. X-ray object detection. Several datasets have been developed to tackle the challenge of object detection in X-ray imagery [1, 33]. SIXray [22], OPIXray [36], and CLCXray [42] focused on detecting occluded objects, introducing strategies to identify deliberately hidden prohibited object. To enhance dataset scale and diversity, PIDray [34] and HiXray [30] provided larger benchmarks, significantly increasing the number of annotated images. Meanwhile, PIXray [19] pioneered the introduction of an X-ray segmentation dataset and proposed a real-time framework for segmenting prohibited objects, further advancing automated threat detection in security screening applications.", + "bbox": [ + 511, + 428, + 903, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous X-ray object detectors [28, 31, 35] considered mainly a closed-set paradigm, thus being limited to localize objects of a small set of predefined categories. More recently, Lin et al. [16] proposed fine-tuning a CLIP adapter to bridge the modality gap between CLIP's training data and X-ray images, thereby achieving OvOD in X-ray data. They demonstrated generalization on four novel object classes, but both training and evaluation were limited by the lack of largescale, well-annotated X-ray data. To address these issues, RAXO leverages data retrieval from the web to expand detection capabilities to a broader range of common objects. Furthermore, our proposed re-labeled DET-COMPASS dataset enables evaluation across a wider variety of classes.", + "bbox": [ + 511, + 704, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20771", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/9dd510c4920a4e50714e1ea457008cff6abbc7e110eae7593002eac5edc629f4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
VenueImagesClassesModality
DvXray [20]TIFS’2432,00015X-ray
PIXray [19]TMM’225,04615X-ray
CLCXray [42]TIFS’229,56512X-ray
FSOD [32]ACMMM’2212,33320X-ray
EDS [31]CVPR’2214,21910X-ray
PIDray [34]ICCV’2147,67712X-ray
HiXray [30]ICCV’2145,3658X-ray
DET-COMPASS (Ours)-1,928370X-ray+RGB
", + "bbox": [ + 91, + 88, + 480, + 208 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Existing X-ray detection datasets. DET-COMPASS is the dataset with the highest number of categories and the only one providing pixel-level alignment between X-ray and RGB data.", + "bbox": [ + 89, + 210, + 482, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. DET-COMPASS", + "text_level": 1, + "bbox": [ + 89, + 268, + 254, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Object detection in security X-ray scans has advanced significantly in recent years. However, evaluating OvOD detectors in this modality remains challenging due to the limited number of annotated object categories in existing X-ray benchmarks. For instance, the largest X-ray detection dataset, FSOD [32], includes annotations for only 20 classes (see Tab. 1). This limitation severely constrains the comprehensive evaluation of OvOD methods, which require a broad and diverse category set (or say vocabulary) to assess generalization to unseen object semantics.", + "bbox": [ + 88, + 294, + 482, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address this gap, we introduce DET-COMPASS, a novel benchmark that repurposes the COMPASS-XP classification dataset [9] for object detection through meticulous manual bounding box annotation. DET-COMPASS comprises 370 distinct object classes, offering an order-of-magnitude increase in vocabulary size over previous X-ray detection benchmarks. Additionally, it provides pixel-aligned RGB images, ensuring precise spatial correspondence across modalities and facilitating the development of multimodal models. Each object is also labeled with a visibility attribute, indicating whether it produces a discernible signature in the X-ray spectrum. Further details are provided in Appendix C.", + "bbox": [ + 88, + 446, + 482, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As summarized in Tab. 1, DET-COMPASS sets a new standard in class diversity and uniquely integrates multimodal, pixel-aligned X-ray and RGB data. The dataset will be released under an open license, serving as a valuable resource for advancing OvOD research in security screening and industrial inspection.", + "bbox": [ + 89, + 643, + 482, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. OvOD for X-ray Imaging", + "text_level": 1, + "bbox": [ + 89, + 750, + 326, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preliminaries: OvOD in the RGB domain. Most RGB OvOD detectors [44, 45] follow a two-stage pipeline. During training, a region proposal network (RPN) is learned to yield a set of $M$ proposals by $\\{\\mathbf{z}_m\\}_{m=1}^M = \\Phi_{\\mathrm{RPN}}(\\mathbf{I}^{RGB})$ , where $\\mathbf{z}_m \\in \\mathbb{R}^D$ is a $D$ -dimensional region-of-interest (RoI) feature embedding. Then, a bounding box regressor predicts coordinates for each proposed region via $\\hat{\\mathbf{b}}_m = \\Phi_{\\mathrm{REG}}(\\mathbf{z}_m)$ . A set of text-based classifiers $\\mathbf{W} = \\{\\mathbf{w}_c | \\mathbf{w}_c \\in \\mathbb{R}^D\\}_{c=1}^{|C_{\\text{train}}|}$ are used", + "bbox": [ + 89, + 776, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to compute classification scores for each region as $\\langle \\mathbf{w}_c, \\mathbf{z}_m \\rangle$ , where $\\langle \\cdot, \\cdot \\rangle$ is the cosine similarity function and $C^{\\mathrm{train}}$ denotes the training vocabulary. In this way, each region's class is determined by the class with the highest score. Here, the classifier $\\mathbf{W}$ is constructed by encoding class names in $C^{\\mathrm{train}}$ using a pre-trained VLM text encoder, e.g., CLIP [26]. During training, OvOD models update all parameters while keeping $\\mathbf{W}$ frozen. This enables RGB-region-class alignment by leveraging large-scale RGB data and the pre-aligned vision-language semantic space of VLMs, facilitating open-vocabulary inference with any test vocabulary $C^{\\mathrm{test}}$ . The vocabularies $C^{\\mathrm{train}}$ and $C^{\\mathrm{test}}$ may be disjoint or overlapping.", + "bbox": [ + 511, + 90, + 903, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem formulation. In this work, we study open-vocabulary object detection (OvOD) in X-ray modality. Specifically, given an input X-ray image $\\mathbf{I}$ and a vocabulary $C^{\\mathrm{test}}$ defined by users at test time, an X-ray OvOD detector $\\mathcal{F}_{\\mathrm{X - ray}}$ aims to detect objects specified in $C^{\\mathrm{test}}$ from $\\mathbf{I}$ (e.g., a \"Power bank\" in a passenger's backpack scan at an airport security checkpoint). Theoretically, this detection process can be formulated as $\\mathcal{F}_{\\mathrm{X - ray}}: \\mathbf{I} \\to \\{(\\mathbf{b}_m, c_m)\\}_{m=1}^{M}$ , where $\\mathbf{b}_m \\in \\mathbb{R}^4$ denotes the coordinates of each bounding box, and $c_m \\in C^{\\mathrm{test}}$ denotes the class label of each bounding box.", + "bbox": [ + 511, + 277, + 903, + 430 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "X-ray OvOD presents significant challenges due to the following reasons: i) Directly applying a pre-trained RGB OvOD detectors to X-ray images leads to suboptimal performance due to the visual modality gap, as shown in Fig. 1; ii) The scarcity of large-scale security X-ray datasets limits the application of RGB OvOD training techniques. These techniques typically rely on extensive image-text annotated data for strong supervision. To tackle these challenges, we introduce RAXO, a plug-and-play module that adapts any off-the-shelf RGB OvOD detectors to X-ray modalities. RAXO requires no training or in-domain detection annotations. Next, we present our approach.", + "bbox": [ + 511, + 431, + 906, + 613 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "5. RAXO: Training-free Modality Adaptation", + "text_level": 1, + "bbox": [ + 511, + 631, + 897, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2, RAXO enables X-ray OvOD by constructing high-quality visual descriptors, $X_{c}$ , for each class $c$ in the user-defined vocabulary, $C^{\\mathrm{test}}$ . To achieve this, RAXO first acquires X-ray samples in an open-ended manner, leveraging both in-house and web-based retrieval sources (Sec. 5.1). Then, it extracts the features of these samples and segments the relevant information, modeling the visual descriptor $X_{c}$ (Sec. 5.2). Once the visual descriptors are constructed offline, they can be directly applied to any off-the-shelf OvOD detector by replacing the conventional text-based classifier $\\mathbf{W}$ , with our visual-based classifier $\\mathcal{X}$ (Sec. 5.3). Thus, RAXO effectively overcomes the misalignment between X-ray features and text semantics, enabling OvOD detectors pre-trained with RGB data to accurately identify X-ray objects based on their intrinsic visual characteristics.", + "bbox": [ + 511, + 657, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20772", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5a5b7f18c2a2e2a5e1ac03b8167e8cfb2316ed1a2d351756ee24b2a9607f162f.jpg", + "image_caption": [ + "Figure 2. Architecture of RAXO. For a given user-defined class $c \\in C^{\\mathrm{test}}$ , RAXO first retrieves its corresponding X-ray images $\\mathcal{G}_c^{XRAY}$ from in-house and web sources, using its (1) Visual Samples Acquisition pipeline (Sec. 5.1). Following this, RAXO extracts the features of the images and segments them with its (2) Class Descriptor Modeling module (Sec. 5.2), creating ensemble visual descriptors for the class $\\mathcal{X}_c$ and the background $\\mathcal{X}_{\\mathrm{bg}}$ . Finally, the text-based classifier from the baseline RGB OvOD detector is replaced with our (3) Visual-based Classifier (Sec. 5.3) $\\mathcal{X}$ , which yields accurate predictions on the X-ray modality." + ], + "image_footnote": [], + "bbox": [ + 109, + 88, + 883, + 361 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.1. Visual Sample Acquisition", + "text_level": 1, + "bbox": [ + 89, + 459, + 328, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Obtaining informative and representative images is crucial for generating robust visual descriptors. To this end, we propose a dual-source acquisition pipeline that retrieves a gallery of relevant X-ray samples $\\mathcal{G}_c^{\\mathrm{XRAY}}$ for each user-specified class $c\\in C^{\\mathrm{test}}$ . This pipeline consists of two modules: an in-house retrieval module and a web-powered retrieval module.", + "bbox": [ + 89, + 481, + 483, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In-house retrieval. Given an in-domain X-ray dataset, $\\mathcal{D}_{\\mathrm{XRAY}}^{\\mathrm{in - house}}$ containing categories $C^{\\mathrm{in - house}}$ , our sample acquisition pipeline first attempts to retrieve the $K$ most relevant images for the user-specified class $c$ based on class name matching: $\\mathcal{G}_c^{\\mathrm{XRAY}} = \\bigl \\{\\mathbf{I}|\\bigl (\\mathbf{I},c\\bigr)\\in \\mathcal{D}_{\\mathrm{XRAY}}^{\\mathrm{in - house}}\\bigr \\} ^K$", + "bbox": [ + 89, + 590, + 483, + 670 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, since the user-defined vocabulary $C^{\\mathrm{test}}$ is open-ended, we cannot assume that every user-specified category will be present in $C^{\\mathrm{in-house}}$ . Consequently, for some categories, the retrieved set $\\mathcal{G}_c^{\\mathrm{XRAY}}$ may be empty. To overcome this limitation and fully support open-vocabulary user input, RAXO further incorporates a novel web-powered retrieval module. This module retrieves RGB images from the web and applies a material transfer mechanism to synthesize X-ray-style representations, which we introduce next.", + "bbox": [ + 89, + 670, + 483, + 805 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Web-powered retrieval. To obtain high-quality visual samples for a category $c$ that is not present in the in-house vocabulary $C^{\\mathrm{in - house}}$ , we leverage the vast availability of web-based RGB image data $\\mathcal{D}_{\\mathrm{RGB}}^{\\mathrm{web}}$ as an open-ended auxiliary source. Specifically, we perform text-based web retrieval using the class name $c$ as a search", + "bbox": [ + 89, + 809, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "query, retrieving the top $K$ results from Google Images as $\\widetilde{\\mathcal{G}}_c^{\\mathrm{web}} = \\left\\{\\mathbf{I}^{\\mathrm{RGB}} \\mid (\\mathbf{I}^{\\mathrm{RGB}}, c) \\in \\mathcal{D}_{\\mathrm{RGB}}^{\\mathrm{web}}\\right\\}^K$ .", + "bbox": [ + 511, + 460, + 906, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The raw web-retrieved results $\\widetilde{\\mathcal{G}}_c^{\\mathrm{web}}$ are often noisy and may not always contain clear instances of the target class $c$ . To refine these results, we apply a filtering step using an RGB OvOD detector $\\mathcal{F}_{\\mathrm{RGB}}$ . Specifically, we discard images in the raw web-retrieved results where class $c$ is not confidently detected, retaining only those where the detection confidence exceeds a threshold $\\tau$ as $\\mathcal{G}_c^{\\mathrm{web}} = \\mathrm{Filter}(\\widetilde{\\mathcal{G}}_c^{\\mathrm{web}},\\mathcal{F}_{\\mathrm{RGB}},c,\\tau)$", + "bbox": [ + 511, + 497, + 905, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Material-transfer mechanism for web-retrieved images. The substantial visual disparity between RGB and X-ray modalities prevents direct use of web-retrieved RGB samples $\\mathcal{G}_c^{\\mathrm{web}}$ for constructing X-ray class descriptors. Style transfer methods [7] fall short in bridging this gap, as they fail to capture the underlying material properties of objects (as shown in Tab. 4). To address this, we introduce a novel material-transfer mechanism for generating synthetic X-ray samples from web-retrieved images. As shown in Fig. 3, our approach consists of two key steps: $i)$ constructing an offline material database $\\mathcal{M}$ that encapsulates the expected X-ray appearance of various materials, and $ii)$ adapting RGB samples $\\mathcal{G}_c^{\\mathrm{web}}$ to X-ray style by applying the corresponding material properties from $\\mathcal{M}$ .", + "bbox": [ + 511, + 611, + 905, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To construct $\\mathcal{M}$ , we employ a Large Language Model (LLM) to cluster the class names of $C^{\\mathrm{in - house}}$ into subsets $C_m\\subset C^{\\mathrm{in - house}}$ , where $C_m$ contains classes that share the same $m$ material - i.e., \" $C_{\\mathrm{metal}} = \\{\\mathrm{gun},\\mathrm{knife},\\mathrm{fork}\\}$ \" or \" $C_{\\mathrm{leather}} = \\{\\mathrm{boot},\\mathrm{belt}\\}$ \". Each material $m$ is then associated", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20773", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a88c4c7f2a5b725a52a3a73abfc50ad5408cb2bd816543e140661b8258b0ac43.jpg", + "image_caption": [ + "Figure 3. Web-powered retrieval and material-transfer mechanism for the class \"violin\". We retrieve violin samples from the web, filter them using $\\mathcal{F}_{\\mathrm{RGB}}$ , and inpaint the retrieved appearance into the object masks to generate synthetic X-ray samples." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 480, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with the corresponding set of images $\\mathcal{K}_m = \\{\\mathbf{I} \\mid (\\mathbf{I}, c_k) \\in \\mathcal{D}_{\\mathrm{XRAY}}^{\\mathrm{in-house}}, c_k \\in C_m\\}$ . These images are used to compute the appearance, $\\mathcal{A}_m \\in \\mathbb{R}^3$ , of the material as:", + "bbox": [ + 89, + 412, + 483, + 460 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} _ {m} = \\frac {1}{| \\mathcal {K} _ {m} |} \\sum_ {\\mathbf {I} \\in \\mathcal {K} _ {m}} \\left(\\frac {\\sum_ {i , j} ^ {H , W} \\left(\\mathbf {I} _ {i , j} \\odot \\Omega \\left(\\mathbf {I} _ {i , j}\\right)\\right)}{\\sum \\Omega (\\mathbf {I})}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 465, + 482, + 510 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\odot$ denotes element-wise multiplication and $\\Omega : \\mathbf{I} \\to \\{0,1\\}^{H \\times W}$ is a segmentation function that produces a binary mask, to isolate the object of interest in $\\mathbf{I}$ . The resulting material-appearance pairs $(m, \\mathcal{A}_m)$ collectively form our material database $\\mathcal{M}$ .", + "bbox": [ + 89, + 518, + 483, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Once $\\mathcal{M}$ is computed offline, it can be used to adapt the web-retrieved RGB samples, $\\mathcal{G}_c^{\\mathrm{web}}$ , to the X-ray modality. To achieve this, we first prompt the LLM to link the class $c$ with its corresponding material $m_c \\in \\mathcal{M}$ (e.g., $c = \\text{\"violin\"}$ is linked with the material $m_c = \\text{\"wood\"}$ ). Subsequently, the X-ray appearance $\\mathcal{A}_m^c$ of the linked material is used to render the characteristic X-ray style on each web-retrieved RGB samples as:", + "bbox": [ + 89, + 594, + 483, + 715 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {c} ^ {\\mathrm {X R A Y}} = \\left\\{\\Omega \\left(\\mathbf {I} ^ {\\mathrm {R G B}}\\right) \\odot \\left(\\mathcal {A} _ {m} ^ {c} \\cdot \\mathbf {1}\\right) \\mid \\mathbf {I} ^ {\\mathrm {R G B}} \\in \\mathcal {G} _ {c} ^ {\\mathrm {w e b}} \\right\\}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 722, + 482, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the X-ray visual gallery for every class freely specified by the user is built as $\\mathcal{G} = \\bigcup_{c\\in C^{\\mathrm{test}}}\\mathcal{G}_c^{\\mathrm{XRAY}}$ through our dual-source retrieval pipeline, seamlessly integrating high-quality in-house X-ray samples with synthetic images derived from web data.", + "bbox": [ + 89, + 750, + 483, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Class Descriptor Modeling", + "text_level": 1, + "bbox": [ + 89, + 833, + 333, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The effectiveness of RAXO hinges on leveraging its visual gallery $\\mathcal{G}$ to construct a robust visual descriptor $\\chi_{c}$ that accurately captures the visual X-ray properties of class $c$ . A", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "naive approach, such as averaging the feature representations of all $\\mathcal{G}_c^{\\mathrm{XRAY}}$ instances, fails to account for intra-class variability (e.g., distinct shapes of utility knives versus chef knives) and does not differentiate between foreground and background regions. To address these limitations, we propose a novel class descriptor modeling strategy shown in Fig. 2(2).", + "bbox": [ + 511, + 90, + 903, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, we process each sample $\\mathbf{I} \\in \\mathcal{G}_c^{XRAY}$ with $\\Omega$ to separate the X-ray object from its background. Simultaneously, a feature extractor $\\phi$ extracts per-patch embeddings $\\phi(\\mathbf{I}) \\in \\mathbb{R}^{H' \\times W' \\times D}$ , where $H' \\times W'$ represents the number of spatial tokens, and $D$ is the embedding dimension. Following [10], we resize (denoted by $\\zeta$ ) the segmentation mask to match the spatial dimensions of the feature map, and subsequently we use it to compute both a positive and a negative prototype. The positive prototype, designed to capture the significant visual features of the object without background interference, is computed as the average embedding over the foreground region as:", + "bbox": [ + 511, + 196, + 906, + 378 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {\\mathbf {I}} ^ {\\text {p o s}} = \\frac {\\sum_ {i , j} ^ {H ^ {\\prime} , W ^ {\\prime}} \\zeta (\\Omega (\\mathbf {I} _ {i , j})) \\odot \\phi (\\mathbf {I} _ {i , j})}{\\sum \\zeta (\\Omega (\\mathbf {I}))}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 386, + 903, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "On the other hand, the negative prototype is obtained as the average embedding over the complementary region (background) as:", + "bbox": [ + 511, + 431, + 903, + 477 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {\\mathbf {I}} ^ {\\text {n e g}} = \\frac {\\sum_ {i , j} ^ {H ^ {\\prime} , W ^ {\\prime}} \\left(1 - \\zeta \\left(\\Omega \\left(\\mathbf {I} _ {\\mathbf {i} , \\mathbf {j}}\\right)\\right)\\right) \\odot \\phi \\left(\\mathbf {I} _ {\\mathbf {i} , \\mathbf {j}}\\right)}{\\sum \\left(1 - \\zeta \\left(\\Omega (\\mathbf {I})\\right)\\right)}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 484, + 903, + 525 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Subsequently, to construct the final visual descriptor $\\mathcal{X}_c$ of class $c$ , we compute the average positive prototype $\\overline{\\mathbf{x}}_c^{\\mathrm{pos}}$ from $\\mathcal{G}_c^{XRAY}$ and unite it with all the individual positive prototypes as:", + "bbox": [ + 511, + 532, + 903, + 593 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {X} _ {c} = \\left[ \\bar {\\mathbf {x}} _ {c} ^ {\\text {p o s}}, \\left\\{\\mathbf {x} _ {\\mathbf {I}} ^ {\\text {p o s}} \\mid \\mathbf {I} \\in \\mathcal {G} _ {c} \\right\\} \\right]. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 601, + 903, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This formulation effectively captures both fine-grained object details and a holistic class-level representation, handling intra-class variability. Additionally, we also construct a global background descriptor by using the negative prototypes across the entire visual sample gallery as $\\mathcal{X}_{bg} = [\\overline{\\mathbf{x}}^{\\mathrm{neg}},\\{\\mathbf{x}_{\\mathbf{I}}^{\\mathrm{neg}}\\mid \\mathbf{I}\\in \\mathcal{G}\\} ]$ . This negative prototype is used to further improve detection reliability by filtering out low-quality proposals (e.g., a region proposed on the background) during inference. A key advantage of this approach is its modularity: visual descriptors $\\mathcal{X}_c$ are computed off-line and can be incrementally expanded with new object categories, seamlessly integrating with the OvOD paradigm without training or requiring X-ray detection data.", + "bbox": [ + 511, + 628, + 906, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.3. Classification is All You Need", + "text_level": 1, + "bbox": [ + 511, + 833, + 774, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 2(3), once the visual descriptors are constructed for each class in the user-specified vocabulary, along with the background class, the RAXO classifier", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20774", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathcal{X} = \\{\\left[\\mathcal{X}_c,\\mathcal{X}_{bg}\\right]\\mid c\\in C^{\\mathrm{test}}\\}$ can be directly applied to any OvOD detector to classify proposals $\\mathbf{z_m}$ as:", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {c} = \\underset {c \\in C ^ {\\text {t e s t}} ^ {\\prime}} {\\arg \\max } \\underset {\\mathbf {x} \\in \\mathcal {X}} {\\max } \\langle \\mathbf {z} _ {m}, \\mathbf {x} \\rangle , \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 137, + 482, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $C^{\\mathrm{test}}$ denotes the test-time vocabulary $C^{\\mathrm{test}}$ extended with the additional \"background\" class. If a proposal matches the background, it is removed. This simple-yet-effective strategy enables any pre-trained OvOD detectors to achieve strong performance on X-ray object detection with minimal modifications and no need for re-training.", + "bbox": [ + 89, + 171, + 483, + 263 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Descriptor consistency criterion. To further reduce incorrect proposals from the OvOD detector, RAXO introduces a novel Descriptor Consistency Criterion (DCC) to enforce class consistency in predictions. DCC evaluates how well a proposal aligns with its predicted class relative to others, suppressing weakly aligned proposals. For a proposal $\\mathbf{z}_m$ , we measure its similarity to the closest prototype as $s_1 = \\max_{\\mathbf{x} \\in \\mathcal{X}_{\\hat{c}}} \\langle \\mathbf{z}_m, \\mathbf{x} \\rangle$ . We also compute its mean similarity to the average prototypes of all other classes as $s_2 = \\mathrm{avg}_{c \\in C^{\\mathrm{test}}, c \\neq \\hat{c}} \\langle \\mathbf{z}_m, \\bar{\\mathbf{x}}_c^{\\mathrm{pos}} \\rangle$ . The difference $\\Delta = s_1 - s_2$ serves as a confidence measure, where a higher value indicates greater alignment with its predicted class than with any other. Proposals with $\\Delta$ below a threshold $\\sigma$ are discarded.", + "bbox": [ + 89, + 267, + 483, + 464 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 89, + 479, + 222, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation protocol. Conventional OvOD models are typically assessed using a Cross-Dataset Transfer Evaluation (CDTE) protocol, where the model is trained on one dataset and evaluated on a different dataset in a zero-shot manner [45]. In contrast, our setting involves a more challenging scenario where both the datasets and their underlying modalities differ. To address this, we introduce a novel Cross-Modality Transfer Evaluation (CMTE) protocol. Under CMTE, an OvOD model is trained on a source RGB dataset and subsequently evaluated on target $X$ -ray datasets without any additional training or fine-tuning. For performance assessment, we employ the standard MS COCO [17] metrics: AP, AP50, and AP75.", + "bbox": [ + 89, + 503, + 483, + 699 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. Unless stated otherwise, RAXO's default configuration employs SAM 2 [27] as the object segmentation module $\\Omega$ , ViT-B/14 [3] pretrained with DINOv2 [25] as the visual encoder $\\phi$ , and GPT-4 [24] as the LLM. We set the consistency threshold to $\\sigma = 0.15$ , the similarity threshold to $\\tau = 0.5$ , and use $K = 30$ samples to construct the visual descriptors. $\\mathcal{D}_{\\mathrm{XRAY}}^{in - house}$ is composed of hold-out samples sourced from the evaluation X-ray datasets and is completely disjoint from the test split. $\\mathcal{D}_{\\mathrm{RGB}}^{web}$ is constructed by collecting images from the web using the public Google Custom Search API [8]. A detailed analysis of the impact of hyperparameter choices is provided in Sec. 6.2. Further implementation details can be found in Supp. D.", + "bbox": [ + 89, + 704, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets and baselines. We evaluate RAXO by integrating it with four different state-of-the-art OvOD detectors that were exclusively trained on RGB images: GroundingDINO [18], Detic [44], VLDet [15], and CoDet [21]. Our CMTE evaluation is performed across six diverse X-ray datasets: PIXray [19], PIDray [34], CLCXray [42], DvXray [20], HiXray [30], and our proposed DET-COMPASS, which together comprise 343 visible unique classes and over 140k images. A detailed description of these datasets is provided in Tab. 1.", + "bbox": [ + 511, + 90, + 906, + 243 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Open-Vocabulary Detection Results", + "text_level": 1, + "bbox": [ + 511, + 253, + 823, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "RAXO equips OvOD models with X-ray vision. We evaluate the X-ray adaptation capabilities of RAXO on the aforementioned datasets and baselines in Tab. 2. Since RAXO is training-free, we report the results directly on the test splits, accounting for all available categories. Furthermore, we analyze the influence of the gallery composition by varying $\\mathcal{G}$ from a $0/100$ configuration (i.e., all samples in $\\mathcal{G}$ are retrieved from the web, as $C^{\\mathrm{test}} \\cap C^{\\mathrm{in-house}} = \\emptyset$ ) to a $100/0$ configuration (i.e., all samples in $\\mathcal{G}$ are retrieved from the in-house database, as $C^{\\mathrm{test}} \\subset C^{\\mathrm{in-house}}$ ). Each experiment is repeated three times with different random distributions of in-domain and web categories, and we report the averaged results.", + "bbox": [ + 511, + 276, + 906, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Tab. 2, off-the-shelf OvOD detectors trained on RGB images perform poorly on X-ray data. In contrast, even without access to any in-domain data from $C^{\\mathrm{test}}$ (0/100 setting), RAXO enhances baseline OvOD methods, yielding an average improvement of $\\uparrow 2.1$ points. This gain is further amplified when in-domain samples are incorporated into $\\mathcal{G}$ , which facilitates the construction of more robust materials in $\\mathcal{M}$ and yields higher-quality visual descriptors. Notably, even a limited number of in-domain samples (20/80 setting) results in an average improvement of $\\uparrow 4.8$ points. Moreover, increasing the in-domain samples (50/50 setting) yields even more significant gains, particularly in datasets with a large number of classes, such as DET-COMPASS, where RAXO offers an average boost of $\\uparrow 14.9$ points. Finally, when $\\mathcal{G}$ consists entirely of in-domain samples, RAXO delivers an average improvement of $\\uparrow 14.7$ points across all OvOD methods and datasets, all in a training-free manner.", + "bbox": [ + 511, + 473, + 908, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "RAXO works with any OvOD detector. The consistent gains of RAXO across all evaluated detectors in Tab. 2, along with its simple integration (Sec. 5.3), confirm the generalization of our approach to any RGB OvOD model. Moreover, RAXO scales with the baseline method, yielding larger gains when integrated with more robust detectors (e.g., RAXO enhances G-DINO $\\uparrow$ 9.4 points on average across all $G$ settings, compared to $\\uparrow$ 7.5 points for Detic). This highlights the advantage of our training-free method, which can leverage the rapid advancements in RGB OvOD detectors without needing large amounts of labeled X-ray data.", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20775", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f924d6734f0dc9db20923945d47dc655339b9d8f9813858a535f3a233578b053.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
GMethodD-COMPASSPIXrayPIDrayCLCXrayDvXrayHiXrayAvg.
G-DINO [18]13.412.910.96.710.07.010.2
100/80/2050/5020/800/100+ RAXO47.9↑34.536.9↑24.016.5↑5.622.2↑15.522.6↑12.617.1↑10.127.2↑17.0
41.0↑27.633.8↑20.915.4↑4.518.0↑11.321.0↑11.014.5↑7.524.0↑13.8
31.4↑18.025.4↑12.515.5↑4.617.0↑10.316.1↑6.113.4↑6.419.8↑9.6
20.5↑7.121.6↑8.713.9↑3.010.0↑3.315.0↑5.09.8↑2.815.1↑4.9
14.0↑0.616.1↑3.213.4↑2.57.1↑0.412.4↑2.47.9↑0.911.8↑1.6
VLDet [15]10.69.86.94.47.45.17.4
100/80/2050/5020/800/100+ RAXO36.4↑25.832.3↑22.511.7↑4.815.4↑11.020.1↑12.714.8↑9.721.8↑14.4
31.8↑21.229.2↑19.411.0↑4.112.7↑8.316.8↑9.413.1↑8.019.1↑11.7
23.7↑13.124.0↑14.210.4↑3.511.1↑6.712.1↑4.711.2↑6.115.4↑8.0
16.2↑5.621.6↑11.89.4↑2.55.2↑0.810.6↑3.29.3↑4.212.1↑4.7
11.1↑0.514.1↑4.38.9↑2.04.4↑0.09.0↑1.68.3↑3.29.3↑1.9
Detic [44]11.59.37.14.77.04.87.4
100/80/2050/5020/800/100+ RAXO35.3↑23.827.3↑18.011.3↑4.214.0↑9.319.4↑12.414.2↑9.420.3↑12.9
30.7↑19.223.9↑14.610.8↑3.712.3↑7.618.0↑11.012.1↑7.318.0↑10.6
24.4↑12.919.5↑10.210.3↑3.29.2↑4.514.6↑7.611.0↑6.214.8↑7.4
16.4↑4.915.2↑5.99.6↑2.58.0↑3.312.7↑5.79.9↑5.112.0↑4.6
11.9↑0.413.4↑4.19.1↑2.05.2↑0.59.4↑2.47.9↑3.19.5↑2.1
CoDet [21]8.47.35.73.15.63.45.6
100/80/2050/5020/800/100+ RAXO35.8↑27.427.9↑20.610.3↑4.614.8↑11.717.6↑12.013.2↑9.819.9↑14.3
32.2↑23.825.1↑17.89.5↑3.812.0↑8.915.4↑9.811.7↑8.317.7↑12.1
24.0↑15.620.0↑12.79.5↑3.89.2↑6.111.5↑5.99.9↑6.514.0↑8.4
17.8↑9.414.8↑7.58.5↑2.85.1↑2.09.4↑3.88.1↑4.710.6↑5.0
12.2↑3.811.5↑4.28.1↑2.44.0↑0.96.9↑1.36.5↑3.18.2↑2.6
", + "bbox": [ + 117, + 88, + 872, + 436 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/03b905b2d0b4d9c92434e9c8765ed0f0bcde604cc3691514810ab2a2e68aabab.jpg", + "table_caption": [ + "Table 2. X-ray OvOD performance under the Cross-Modality Transfer Evaluation (CMTE) setting on DET-COMPASS (ours), PINray [19], PIDray [34], CLCXray [42], DvXray [20], and HiXray [30] datasets. We integrate RAXO into different baselines using different gallery $\\mathcal{G}$ compositions, from using only ${\\mathcal{D}}_{\\mathrm{{XRAY}}}^{\\text{in-house }}$ data $\\left( {{100}/0}\\right)$ to exclusively ${\\mathcal{D}}_{\\mathrm{{RGB}}}^{\\text{web }}$ samples (0/100). RAXO consistently improves the performance of all baseline OvOD detectors across every dataset. AP is used." + ], + "table_footnote": [], + "table_body": "
PIXray
ScissorsWrenchBatteryPliersAP50
OVXD†16.946.214.66.421.0
BARON + RAXO18.239.48.322.022.0
CoDet + RAXO43.652.69.018.630.9
Detic + RAXO46.360.57.936.637.8
VLDet + RAXO48.362.710.638.540.0
G-DINO + RAXO49.761.611.351.543.5
", + "bbox": [ + 94, + 508, + 473, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "RAXO surpasses training-based approaches. We also compare RAXO with OVXD [16], the only concurrent work that extends X-ray detection beyond base categories. OVXD is a fully-supervised method trained directly on X-ray data, and its generalization capabilities are evaluated on a hold-out set comprising only four categories (i.e., Scissors, Wrench, Battery, and Pliers). In Tab. 3, we follow this setting and evaluate RAXO on PINray [19], excluding the aforementioned classes from the in-house dataset $\\mathcal{D}_{\\mathrm{XRAY}}^{\\mathrm{in - house}}$ . Our results demonstrate that RAXO outperforms OVXD even when both approaches utilize the same BARON detector [37]. Notably, unlike OVXD, our method does not retrain BARON or any other component. Yet, our training-free RAXO performs remarkably well, achieving an AP50", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/31a5a8dcc10d14f7c17acb8b5511a7eebc02466e8d9830fa99f67ddac27d4de1.jpg", + "table_caption": [ + "Table 3. Comparison under OVXD [16] setting, where the methods do not have access to in-domain data from the displayed categories. $\\dagger$ OVXD is a supervised method explicitly trained on X-ray images for alignment. AP50 is used." + ], + "table_footnote": [], + "table_body": "
Mod. Filt. Trans.APAP50AP75
G-DINO [18]12.914.913.4
+ RAXOX-ray RGBS7.9↓5.09.8↓5.18.3↓5.1
13.9↑1.017.4↑2.514.3↑0.9
RGB✓14.3↑1.417.6↑2.714.7↑1.3
RGB✓18.3↑3.415.1↑1.7
RGB✓M16.1↑3.219.8↑4.9
", + "bbox": [ + 521, + 508, + 890, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation of web-based retrieval on PIXray [19] using only web-based samples (setting $0/100$ ). Mod. indicates retrieval of a specific modality from the web, Filt. refers to filtering retrieved images with $\\mathcal{F}_{\\mathrm{RGB}}$ , and Trans. specifies style-transfer via: $S$ a diffusion-based method [7] or $M$ our material-transfer mechanism.", + "bbox": [ + 511, + 616, + 906, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of 43.5 when paired with G-DINO [18].", + "bbox": [ + 511, + 698, + 777, + 713 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 723, + 666, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section we study the core components of RAXO. We conduct our experiments on PixRay [19], using GroundingDINO [13] as our baseline OvOD. Consistent findings are reported in Supp. E and H.", + "bbox": [ + 511, + 744, + 905, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Web-based samples. Tab. 4 presents an ablation study on the components of our web-based visual sample acquisition pipeline (Sec. 5.1). Notably, directly retrieving X-ray samples from the web $(\\mathrm{Mod.} = \\mathrm{X}$ -ray) leads to a $\\downarrow 5.0$ point decrease in AP. In contrast, retrieving RGB images from the web $(\\mathrm{Mod.} = \\mathrm{RGB})$ provides a modest AP improvement of", + "bbox": [ + 511, + 810, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20776", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4670a4ed58b61ce1a19f01619452f9219468d48659362688d3213dba52a1db38.jpg", + "table_caption": [], + "table_footnote": [ + "Table 5. Ablation of class representation construction on PIXray [19] using in-domain samples (setting $100/0$ ). $\\Omega$ indicates building positive prototypes using segmentation masks. $\\mathbf{x}^{\\mathrm{neg}}$ indicates use of negative prototypes. $\\mathcal{X}$ indicates ensemble multiple prototypes in class descriptors. $DCC$ indicates use of prototype consistency criterion." + ], + "table_body": "
Ω x neg X DCCAPAP50AP75
G-DINO [18]12.914.913.4
+ RAXO25.2↑12.331.4↑16.526.2↑12.8
✓27.1↑14.233.6↑18.728.4↑15.0
✓ ✓27.8↑14.934.5↑19.629.3↑15.9
✓27.5↑14.634.0↑19.128.8↑15.4
✓ ✓ ✓28.5↑15.635.3↑20.429.9↑16.5
✓ ✓ ✓36.9↑24.045.0↑30.139.0↑25.6
", + "bbox": [ + 99, + 88, + 470, + 199 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\uparrow 1.0$ point. We hypothesize that directly using X-ray examples obtained from the web is ineffective due to the limited availability of high-quality X-ray images online. Filtering the retrieved RGB images using $\\mathcal{F}_{\\mathrm{RGB}}$ increases AP by an additional $\\uparrow 0.4$ points. Nonetheless, they still exhibit a significant visual gap compared to their X-ray counterparts. A style-transfer approach using StyleShot [7] (Trans. $= S$ ) provides a marginal AP gain of $\\uparrow 0.3$ points, likely due to its limited capacity to understand the intrinsic material properties of the objects. Conversely, our proposed material-transfer mechanism (Trans. $= M$ ) yields a substantial AP improvement of $\\uparrow 1.8$ points, underscoring its effectiveness in bridging the X-ray modality gap.", + "bbox": [ + 89, + 300, + 483, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Class descriptor modeling. Tab. 5 presents an ablation study analyzing the impact of the various components used to construct class representations. As shown, using a simple per-class average of all the features from each sample already results in a significant AP improvement of $\\uparrow 12.3$ points over the baseline OvOD. Refining the representation by applying an object segmentation method to isolate the foreground features $(\\Omega)$ further boosts the AP by $\\uparrow 1.9$ points. Moreover, leveraging background features to construct negative prototypes $(\\mathbf{x}^{\\mathrm{neg}})$ adds an additional $\\uparrow 0.7$ points.", + "bbox": [ + 89, + 501, + 483, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Replacing the averaged class prototypes with our proposed visual descriptor $(\\mathcal{X})$ yields another $\\uparrow 0.7$ point improvement, demonstrating its enhanced capability to capture intra-class visual variability. Finally, incorporating our descriptor consistency criterion to suppress low-quality proposals $(DCC)$ provides a further AP increase of $\\uparrow 8.4$ points, underscoring the effectiveness of our proposed framework.", + "bbox": [ + 89, + 652, + 482, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hyperparameter study. Fig. 4 analyzes the effect of varying the number of samples $K$ used to construct the visual descriptors. Even with just a single sample per class, RAXO improves AP by $\\uparrow 6.5$ points over the baseline. As $K$ increases, performance continues to improve, reaching saturation at $K = 30$ with a total AP gain of $\\uparrow 24.0$ points.", + "bbox": [ + 89, + 763, + 482, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Furthermore, Fig. 5 explores the influence of the threshold $\\sigma$ in the Descriptor Consistency Module. Higher values of $\\sigma$ impose stricter consistency requirements, de", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4b74a760a4e378c70dc0de82ea55681b64fd98e75c5e79308d65d0aadc3210e0.jpg", + "image_caption": [ + "Figure 4. Impact of $K$ in class representations evaluated on PINray [19] using a G-DINO [18] baseline in the 100/0 setting." + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 901, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4b282f175fcce032c9c8714b863deafad0e5af70cb1264185bc000316d26dbe8.jpg", + "image_caption": [ + "Figure 5. Impact of $\\sigma$ in Descriptor Consistency Module on PIxray [19] using a G-DINO [18] baseline in the $100/0$ setting." + ], + "image_footnote": [], + "bbox": [ + 517, + 282, + 901, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "manding greater separation between the predicted class and other classes before accepting a proposal. We observe that $\\sigma = 0.15$ provides an optimal balance between false positives and false negatives, yielding an AP of 36.9 points.", + "bbox": [ + 511, + 494, + 906, + 555 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 571, + 640, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we pioneered the task of open-vocabulary object detection (OvOD) for X-ray imagery, a challenge shaped by the unique characteristics of X-ray data and the scarcity of annotated examples. To address this, we introduced RAXO, a training-free method that repurposes off-the-shelf RGB-based OvOD detectors for the X-ray domain. By leveraging intra-modal feature distances, a novel material-transfer mechanism, and robust class descriptor modeling, RAXO effectively bridges the modality gap between RGB and X-ray imagery. Our extensive experiments, conducted across multiple benchmarks and the newly proposed DET-COMPASS dataset, demonstrate that RAXO consistently enhances detection performance, achieving average mAP improvements of up to 17.0 points over baseline methods.", + "bbox": [ + 511, + 597, + 906, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Looking ahead, the modularity of RAXO opens promising avenues for future research, including further refinement of visual descriptors and adaptation to additional modalities. Overall, our work not only sets a new state-of-the-art for X-ray open-vocabulary detection but also lays the groundwork for a thriving research direction in this emerging domain.", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20777", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 256, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 375, + 187, + 391 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Samet Akcay and Toby Breckon. Towards automatic threat detection: A survey of advances of deep learning within x-ray security imaging. Pattern Recognition, 122:108245, 2022. 2, 3", + "[2] An Chang, Yu Zhang, Shunli Zhang, Leisheng Zhong, and Li Zhang. Detecting prohibited objects with physical size constraint from cluttered x-ray baggage images. Knowledge-Based Systems, 2022. 1", + "[3] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 6", + "[4] Penghui Du, Yu Wang, Yifan Sun, Luting Wang, Yue Liao, Gang Zhang, Errui Ding, Yan Wang, Jingdong Wang, and Si Liu. Lami-detr: Open-vocabulary detection with language model instruction. In ECCV, 2024. 4, 5", + "[5] Ruohuan Fang, Guansong Pang, and Xiao Bai. Simple image-level classification improves open-vocabulary object detection. In AAAI, 2024. 2", + "[6] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated images. In ECCV, 2022. 2", + "[7] Junyao Gao, Yanchen Liu, Yanan Sun, Yinhao Tang, Yanhong Zeng, Kai Chen, and Cairong Zhao. Styleshot: A snapshot on any style. arXiv:2407.01414, 2024. 4, 7, 8, 5", + "[8] Google LLC. Custom search JSON API reference. https: // developers.google.com/custom-search/v1, 2024. Accessed: 2025-03-07. 6, 3", + "[9] Lewis D. Griffin, Matthew Caldwell, and Jerone T. A. Andrews. COMPASS-XP. Zenodo, 2019. 3, 1", + "[10] Laurynas Karazija, Iro Laina, Andrea Vedaldi, and Christian Rupprecht. Diffusion models for open-vocabulary segmentation. In ECCV, 2024. 5" + ], + "bbox": [ + 94, + 400, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "We thank CINECA and the ISCRA initiative for the availability of high-performance computing resources. This work was partially supported by the EU HORIZON IAMI (HORIZON-CL3-2023-FCT-01-04-101168272) project, the EU HORIZON ELIAS (HORIZON-CL4-2022-HUMAN-02-101120237) project, the EU HORIZON ELLIOT (HORIZON-CL4-2024-HUMAN-03-101214398) project, the MUR PNRR FAIR (PE00000013) project funded by the NextGenerationEU, the Spanish Ministerio de Ciencia e Innovacion (grant numbers PID2020-112623GB-I00, PID2023-149549NB-I00), and the Galician Conselleria de Cultura, Educacion e Ordenacion Universitaria (2024-2027 ED431G-2023/04). Some of these grants are co-funded by the European Regional Development Fund (ERDF). Pablo Garcia-Fernandez is supported by the Spanish Ministerio de Universidades under the FPU national plan (grant number FPU21/05581).", + "[11] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models. arXiv:2209.15639, 2022. 2", + "[12] Liangqi Li, Jiaxu Miao, Dahu Shi, Wenming Tan, Ye Ren, Yi Yang, and Shiliang Pu. Distilling detr with visual-linguistic knowledge for open-vocabulary object detection. In ICCV, 2023. 2", + "[13] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, 2022. 2, 7", + "[14] Mingyuan Li, Tong Jia, Hao Wang, Bowen Ma, Hui Lu, Shuyang Lin, Da Cai, and Dongyue Chen. Ao-detr: Anti-overlapping detr for x-ray prohibited items detection. IEEE Transactions on Neural Networks and Learning Systems, 2024. 1", + "[15] Chuang Lin, Peize Sun, Yi Jiang, Ping Luo, Lizhen Qu, Gholamreza Haffari, Zehuan Yuan, and Jianfei Cai. Learning object-language alignments for open-vocabulary object detection. In ICLR, 2023. 1, 2, 6, 7, 5", + "[16] Shuyang Lin, Tong Jia, Hao Wang, Bowen Ma, Mingyuan Li, and Dongyue Chen. Detection of novel prohibited item categories for real-world security inspection. Eng. Appl. Artif. Intell., 2025. 1, 2, 7", + "[17] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 6", + "[18] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding DINO: marrying DINO with grounded pre-training for open-set object detection. In ECCV, 2024. 1, 2, 6, 7, 8, 4, 5", + "[19] Bowen Ma, Tong Jia, Min Su, Xiaodong Jia, Dongyue Chen, and Yichun Zhang. Automated segmentation of prohibited items in x-ray baggage images using dense de-overlap attention snake. TMM, 2022. 2, 3, 6, 7, 8, 4, 5", + "[20] Bowen Ma, Tong Jia, Mingyuan Li, Songsheng Wu, Hao Wang, and Dongyue Chen. Towards dual-view x-ray baggage inspection: A large-scale benchmark and adaptive hierarchical cross refinement for prohibited item discovery. IEEE TIFS, 2024. 1, 3, 6, 7, 5", + "[21] Chuofan Ma, Yi Jiang, Xin Wen, Zehuan Yuan, and Xiaojuan Qi. CoDet: Co-occurrence guided region-word alignment for open-vocabulary object detection. In NeurIPS, 2023. 2, 6, 7, 5", + "[22] Caijing Miao, Lingxi Xie, Fang Wan, Chi Su, Hongye Liu, Jianbin Jiao, and Qixiang Ye. Sixray: A large-scale security inspection x-ray benchmark for prohibited item discovery in overlapping images. In CVPR, 2019. 2", + "[23] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, et al. Simple open-vocabulary object detection. In ECCV, 2022. 2", + "[24] OpenAI. GPT-4 technical report. CoRR, abs/2303.08774, 2023.6" + ], + "bbox": [ + 94, + 92, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20778", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv:2304.07193, 2023. 6", + "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3", + "[27] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dólar, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv:2408.00714, 2024. 6", + "[28] Haifeng Sima, Bailiang Chen, Chaosheng Tang, Yudong Zhang, and Junding Sun. Multi-scale feature attention-detection transformer: Multi-scale feature attention for security check object detection. IET Computer Vision, 2024. 2", + "[29] Archana Singh and Dhiraj. Advancements in machine learning techniques for threat item detection in x-ray images: a comprehensive survey. Int. J. Multim. Inf. Retr., 2024. 1", + "[30] Renshuai Tao, Yanlu Wei, Xiangjian Jiang, Hainan Li, Haotong Qin, Jiakai Wang, Yuqing Ma, Libo Zhang, and Xi-anglong Liu. Towards real-world x-ray security inspection: A high-quality benchmark and lateral inhibition module for prohibited items detection. In ICCV, 2021. 2, 3, 6, 7, 5", + "[31] Renshuai Tao, Hainan Li, Tianbo Wang, Yanlu Wei, Yifu Ding, Bowei Jin, Hongping Zhi, Xianglong Liu, and Aishan Liu. Exploring endogenous shift for cross-domain detection: A large-scale benchmark and perturbation suppression network. In CVPR, 2022. 2, 3", + "[32] Renshuai Tao, Tianbo Wang, Ziyang Wu, Cong Liu, Aishan Liu, and Xianglong Liu. Few-shot x-ray prohibited item detection: A benchmark and weak-feature enhancement network. In ACMMM, 2022. 1, 3", + "[33] Divya Velayudhan, Taimur Hassan, Ernesto Damiani, and Naoufel Werghi. Recent advances in baggage threat detection: A comprehensive and systematic survey. ACM Computing Surveys, 55(8):1-38, 2022. 2, 3", + "[34] Boying Wang, Libo Zhang, Longyin Wen, Xianglong Liu, and Yanjun Wu. Towards real-world prohibited item detection: A large-scale x-ray benchmark. In ICCV, 2021. 2, 3, 6, 7, 5", + "[35] Ruxue Wang, Yuliang Shi, and Mingyu Cai. Optimization and research of suspicious object detection algorithm in x-ray image. In 2023 IEEE 6th Information Technology, Networking, Electronic and Automation Control Conference (ITNEC), 2023. 2", + "[36] Yanlu Wei, Renshuai Tao, Zhangjie Wu, Yuqing Ma, Libo Zhang, and Xianglong Liu. Occluded prohibited items detection: An x-ray security inspection benchmark and de-occlusion attention module. In ACMMM, 2020. 2", + "[37] Size Wu, Wenwei Zhang, Sheng Jin, Wentao Liu, and" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Chen Change Loy. Aligning bag of regions for open-vocabulary object detection. In CVPR, 2023. 2, 7", + "[38] Size Wu, Wenwei Zhang, Lumin Xu, Sheng Jin, Xiangtai Li, Wentao Liu, and Chen Change Loy. Clipself: Vision transformer distills itself for open-vocabulary dense prediction. In ICLR, 2024. 2", + "[39] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In CVPR, 2023. 2", + "[40] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. Detclip: Dictionary-enriched visual-concept paralleled pretraining for open-world detection. In NeurIPS, 2022. 2", + "[41] Lewei Yao, Jianhua Han, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, and Hang Xu. Detclipv2: Scalable open-vocabulary object detection pre-training via word-region alignment. In CVPR, 2023. 2", + "[42] Cairong Zhao, Liang Zhu, Shuguang Dou, Weihong Deng, and Liang Wang. Detecting overlapped objects in x-ray security imagery by a label-aware mechanism. IEEE TIFS, 2022, 2, 3, 6, 7, 5", + "[43] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In CVPR, 2022. 2", + "[44] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, 2022. 1, 2, 3, 6, 7, 5", + "[45] Chaoyang Zhu and Long Chen. A survey on open-vocabulary detection and segmentation: Past, present, and future. IEEE TPAMI, 2024. 3, 6" + ], + "bbox": [ + 516, + 92, + 903, + 542 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20779", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Superpowering Open-Vocabulary Object Detectors for X-ray Vision/60987d09-4da2-484e-8cc1-bea5f240eaf8_model.json b/2025/Superpowering Open-Vocabulary Object Detectors for X-ray Vision/60987d09-4da2-484e-8cc1-bea5f240eaf8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..411cd385808118484c36b3c0b7751b029181edbf --- /dev/null +++ b/2025/Superpowering Open-Vocabulary Object Detectors for X-ray Vision/60987d09-4da2-484e-8cc1-bea5f240eaf8_model.json @@ -0,0 +1,1958 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.158, + 0.131, + 0.84, + 0.154 + ], + "angle": 0, + "content": "Superpowering Open-Vocabulary Object Detectors for X-ray Vision" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.161, + 0.823, + 0.198 + ], + "angle": 0, + "content": "Pablo Garcia-Fernandez\\(^{1*}\\) Lorenzo Vaquero\\(^{1,2}\\) Mingxuan Liu\\(^{3}\\) Feng Xue\\(^{3*}\\) Daniel Cores\\(^{1}\\) Nicu Sebe\\(^{3}\\) Manuel Mucientes\\(^{1}\\) Elisa Ricci\\(^{2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.203, + 0.823, + 0.24 + ], + "angle": 0, + "content": "1University of Santiago de Compostela, Spain 2Fondazione Bruno Kessler, Italy 3University of Trento, Italy" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.242, + 0.868, + 0.256 + ], + "angle": 0, + "content": "{pablogarcia.fernandez, daniel.cores, manuel.mucientes}@usc.es lvaquerootal@fbk.eu" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.256, + 0.823, + 0.27 + ], + "angle": 0, + "content": "{mingxuan.liu, feng.xue, niculae.sebe, e.ricci}@unitn.it *Corresponding authors" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.292, + 0.326, + 0.307 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.323, + 0.485, + 0.671 + ], + "angle": 0, + "content": "Open-vocabulary object detection (OvOD) is set to revolutionize security screening by enabling systems to recognize any item in X-ray scans. However, developing effective OvOD models for X-ray imaging presents unique challenges due to data scarcity and the modality gap that prevents direct adoption of RGB-based solutions. To overcome these limitations, we propose RAXO, a training-free framework that repurposes off-the-shelf RGB OvOD detectors for robust X-ray detection. RAXO builds high-quality X-ray class descriptors using a dual-source retrieval strategy. It gathers relevant RGB images from the web and enriches them via a novel X-ray material transfer mechanism, eliminating the need for labeled databases. These visual descriptors replace text-based classification in OvOD, leveraging intra-modal feature distances for robust detection. Extensive experiments demonstrate that RAXO consistently improves OvOD performance, providing an average mAP increase of up to 17.0 points over base detectors. To further support research in this emerging field, we also introduce DET-COMPASS, a new benchmark featuring bounding box annotations for over 300 object categories, enabling large-scale evaluation of OvOD in X-ray. Code and dataset available at: https://pagf188.github.io/RAXO/." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.694, + 0.222, + 0.71 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.81 + ], + "angle": 0, + "content": "Automated object detection technologies for X-ray imaging are essential to maintain public safety, enabling the identification of prohibited items at checkpoints in high-risk environments such as airports, train stations, museums and stadiums [29]. These systems improve security while simultaneously reducing the workload of human inspectors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Conventional X-ray object detectors rely on supervised learning [2, 14, 20] and are inherently limited by the object categories present in their training datasets (see Fig. 1a). This limitation is exacerbated by the high cost of X-ray machinery and the requirement for expert annotation, often restricting these systems to fewer than 20 object classes [32]," + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.289, + 0.885, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.547, + 0.907, + 0.631 + ], + "angle": 0, + "content": "Figure 1. (a) Traditional X-ray object detectors are constrained by the limited categories in their training datasets. (b) We introduce the task of open-vocabulary object detection (OvOD) for X-ray imaging and propose RAXO, a training-free method that adapts off-the-shelf RGB OvOD models to X-ray data. (c) RAXO greatly improves detection performance across multiple benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.644, + 0.845, + 0.659 + ], + "angle": 0, + "content": "thereby hindering broader real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.906, + 0.795 + ], + "angle": 0, + "content": "In light of the expanding diversity of man-made objects and evolving security demands, an open-vocabulary object detection (OvOD) framework capable of recognizing arbitrary X-ray object categories defined by the user is imperative. Despite its importance, open-vocabulary detection in the X-ray domain remains largely unexplored. Concurrent works aiming to extend beyond base categories have managed to generalize to at most four unseen classes [16], thereby highlighting the inherent challenges of the task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Recent advances in OvOD for conventional RGB images have been driven by large-scale annotated datasets, which facilitate effective alignment between visual and textual features [15, 18, 44]. However, these advancements do not directly transfer to X-ray imagery. In practice, applying state-of-the-art RGB-based OvOD models to X-ray scans leads to significant performance degradation, as illustrated" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20770" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "in Fig. 1c. Moreover, retraining these models on X-ray data is often impractical given the scarcity of large-scale annotated X-ray datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.137, + 0.482, + 0.227 + ], + "angle": 0, + "content": "Motivated by these challenges, this work opens up open-vocabulary object detection for X-ray imaging by repurposing robust, off-the-shelf RGB-based detectors. To this end, we propose RAXO (tRAining-free adaptation for X-ray Open-vocabulary detection), a training-free method that seamlessly adapts RGB OvOD models to X-ray. (Fig. 1b)." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.228, + 0.483, + 0.514 + ], + "angle": 0, + "content": "We find that the main reason RGB-based OvOD detectors fail in the X-ray domain is the disruption of text-visual feature alignment. The appearance disparity between the same object in RGB and X-ray modalities causes textual embeddings, aligned with RGB features, to mismatch with X-ray visual features. RAXO addresses this issue on three simple steps: (1) visual sample acquisition, which involves obtaining X-ray images that represent user-defined categories; (2) class descriptor modeling, which uses these images to construct descriptors that effectively encode the visual appearance of each category within the X-ray domain; and (3) classifier construction, where the computed descriptors are used to build a visual classifier that replaces the inter-modal (i.e., text-to-visual) classification of conventional OvODs. This new classifier exploits the fact that, despite the modality shift, intra-modal (e.g., visual-to-visual) feature distances remain reliable indicators for object identification. In this way RAXO enables any OvOD to successfully detect X-ray objects based on their inherent visual properties." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.515, + 0.483, + 0.635 + ], + "angle": 0, + "content": "To facilitate comprehensive evaluation, we also introduce DET-COMPASS, a detection dataset comprising 370 distinct object classes with paired X-ray and RGB annotations. Extensive experiments demonstrate that our approach consistently improves detection performance by an average of \\(\\uparrow 8.4\\) AP across multiple benchmarks and OvOD models, with gains that scale as the underlying detectors become more powerful (Tab. 2). In summary, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.636, + 0.483, + 0.68 + ], + "angle": 0, + "content": "- We formally introduce the problem of open-vocabulary object detection for X-ray imagery without training, addressing a critical need for security in real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.681, + 0.483, + 0.741 + ], + "angle": 0, + "content": "- We introduce DET-COMPASS, a novel benchmark with bounding box annotations across 370 object categories, enabling standardized evaluation of OvOD methods in the X-ray modality." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.741, + 0.483, + 0.816 + ], + "angle": 0, + "content": "- We propose RAXO, a novel training-free approach that repurposes off-the-shelf RGB-based OvOD models for X-ray detection by constructing robust visual descriptors. RAXO achieves new state-of-the-art results across multiple benchmarks." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.636, + 0.483, + 0.816 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.83, + 0.233, + 0.846 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Open-vocabulary object detection has advanced rapidly with the advent of Vision-Language Models (VLMs), enabling detectors to generalize beyond fixed categories." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "OvOD detectors leverage weak supervision signals to improve detection accuracy. Based on the type of weak supervision utilized, OvOD methods can be categorized into (i) region-aware training, (ii) pseudo-labeling, (iii) knowledge distillation, and (iv) transfer learning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.17, + 0.905, + 0.426 + ], + "angle": 0, + "content": "Region-aware training methods aim to improve localization and feature representation by refining the alignment between image regions and their corresponding textual descriptions. Approaches such as DetCLIP [40], DetCLIPv2 [41], CORA [39], and VLDet [15] adopt this strategy. Pseudo-labeling methods, rely on large pretrained VLMs to generate pseudo-labels, effectively expanding the training set. Methods like RegionCLIP [43], PromptDet [6], CoDET [21], GLIP [13], Detric [44], and Grounding DINO [18] follow this approach. Knowledge distillation techniques, such as BARON [37], DK-DETR [12], CLIPSelf [38], and SIC-CADS [5], employ VLMs as teachers in a teacher-student framework, transferring knowledge from the VLM image encoder to enhance the detector backbone. In contrast, transfer learning approaches integrate the VLM encoder directly, either through fine-tuning, as seen in OWL-ViT [23], or by freezing the encoder, as in F-VLM [11]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.905, + 0.702 + ], + "angle": 0, + "content": "Regardless of the training strategy, OvOD methods heavily depend on VLMs trained on large-scale datasets. However, such datasets are unavailable in the X-ray modality, making this approach impractical. To address this, RAXO seamlessly adapts existing RGB-based OvOD methods for X-ray object detection without requiring additional training. X-ray object detection. Several datasets have been developed to tackle the challenge of object detection in X-ray imagery [1, 33]. SIXray [22], OPIXray [36], and CLCXray [42] focused on detecting occluded objects, introducing strategies to identify deliberately hidden prohibited object. To enhance dataset scale and diversity, PIDray [34] and HiXray [30] provided larger benchmarks, significantly increasing the number of annotated images. Meanwhile, PIXray [19] pioneered the introduction of an X-ray segmentation dataset and proposed a real-time framework for segmenting prohibited objects, further advancing automated threat detection in security screening applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Previous X-ray object detectors [28, 31, 35] considered mainly a closed-set paradigm, thus being limited to localize objects of a small set of predefined categories. More recently, Lin et al. [16] proposed fine-tuning a CLIP adapter to bridge the modality gap between CLIP's training data and X-ray images, thereby achieving OvOD in X-ray data. They demonstrated generalization on four novel object classes, but both training and evaluation were limited by the lack of largescale, well-annotated X-ray data. To address these issues, RAXO leverages data retrieval from the web to expand detection capabilities to a broader range of common objects. Furthermore, our proposed re-labeled DET-COMPASS dataset enables evaluation across a wider variety of classes." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20771" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.482, + 0.209 + ], + "angle": 0, + "content": "
VenueImagesClassesModality
DvXray [20]TIFS’2432,00015X-ray
PIXray [19]TMM’225,04615X-ray
CLCXray [42]TIFS’229,56512X-ray
FSOD [32]ACMMM’2212,33320X-ray
EDS [31]CVPR’2214,21910X-ray
PIDray [34]ICCV’2147,67712X-ray
HiXray [30]ICCV’2145,3658X-ray
DET-COMPASS (Ours)-1,928370X-ray+RGB
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.211, + 0.483, + 0.253 + ], + "angle": 0, + "content": "Table 1. Existing X-ray detection datasets. DET-COMPASS is the dataset with the highest number of categories and the only one providing pixel-level alignment between X-ray and RGB data." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.269, + 0.255, + 0.284 + ], + "angle": 0, + "content": "3. DET-COMPASS" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.295, + 0.483, + 0.446 + ], + "angle": 0, + "content": "Object detection in security X-ray scans has advanced significantly in recent years. However, evaluating OvOD detectors in this modality remains challenging due to the limited number of annotated object categories in existing X-ray benchmarks. For instance, the largest X-ray detection dataset, FSOD [32], includes annotations for only 20 classes (see Tab. 1). This limitation severely constrains the comprehensive evaluation of OvOD methods, which require a broad and diverse category set (or say vocabulary) to assess generalization to unseen object semantics." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.447, + 0.483, + 0.644 + ], + "angle": 0, + "content": "To address this gap, we introduce DET-COMPASS, a novel benchmark that repurposes the COMPASS-XP classification dataset [9] for object detection through meticulous manual bounding box annotation. DET-COMPASS comprises 370 distinct object classes, offering an order-of-magnitude increase in vocabulary size over previous X-ray detection benchmarks. Additionally, it provides pixel-aligned RGB images, ensuring precise spatial correspondence across modalities and facilitating the development of multimodal models. Each object is also labeled with a visibility attribute, indicating whether it produces a discernible signature in the X-ray spectrum. Further details are provided in Appendix C." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.735 + ], + "angle": 0, + "content": "As summarized in Tab. 1, DET-COMPASS sets a new standard in class diversity and uniquely integrates multimodal, pixel-aligned X-ray and RGB data. The dataset will be released under an open license, serving as a valuable resource for advancing OvOD research in security screening and industrial inspection." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.751, + 0.327, + 0.768 + ], + "angle": 0, + "content": "4. OvOD for X-ray Imaging" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.777, + 0.483, + 0.904 + ], + "angle": 0, + "content": "Preliminaries: OvOD in the RGB domain. Most RGB OvOD detectors [44, 45] follow a two-stage pipeline. During training, a region proposal network (RPN) is learned to yield a set of \\(M\\) proposals by \\(\\{\\mathbf{z}_m\\}_{m=1}^M = \\Phi_{\\mathrm{RPN}}(\\mathbf{I}^{RGB})\\), where \\(\\mathbf{z}_m \\in \\mathbb{R}^D\\) is a \\(D\\)-dimensional region-of-interest (RoI) feature embedding. Then, a bounding box regressor predicts coordinates for each proposed region via \\(\\hat{\\mathbf{b}}_m = \\Phi_{\\mathrm{REG}}(\\mathbf{z}_m)\\). A set of text-based classifiers \\(\\mathbf{W} = \\{\\mathbf{w}_c | \\mathbf{w}_c \\in \\mathbb{R}^D\\}_{c=1}^{|C_{\\text{train}}|}\\) are used" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.273 + ], + "angle": 0, + "content": "to compute classification scores for each region as \\(\\langle \\mathbf{w}_c, \\mathbf{z}_m \\rangle\\), where \\(\\langle \\cdot, \\cdot \\rangle\\) is the cosine similarity function and \\(C^{\\mathrm{train}}\\) denotes the training vocabulary. In this way, each region's class is determined by the class with the highest score. Here, the classifier \\(\\mathbf{W}\\) is constructed by encoding class names in \\(C^{\\mathrm{train}}\\) using a pre-trained VLM text encoder, e.g., CLIP [26]. During training, OvOD models update all parameters while keeping \\(\\mathbf{W}\\) frozen. This enables RGB-region-class alignment by leveraging large-scale RGB data and the pre-aligned vision-language semantic space of VLMs, facilitating open-vocabulary inference with any test vocabulary \\(C^{\\mathrm{test}}\\). The vocabularies \\(C^{\\mathrm{train}}\\) and \\(C^{\\mathrm{test}}\\) may be disjoint or overlapping." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.278, + 0.905, + 0.431 + ], + "angle": 0, + "content": "Problem formulation. In this work, we study open-vocabulary object detection (OvOD) in X-ray modality. Specifically, given an input X-ray image \\(\\mathbf{I}\\) and a vocabulary \\(C^{\\mathrm{test}}\\) defined by users at test time, an X-ray OvOD detector \\(\\mathcal{F}_{\\mathrm{X - ray}}\\) aims to detect objects specified in \\(C^{\\mathrm{test}}\\) from \\(\\mathbf{I}\\) (e.g., a \"Power bank\" in a passenger's backpack scan at an airport security checkpoint). Theoretically, this detection process can be formulated as \\(\\mathcal{F}_{\\mathrm{X - ray}}: \\mathbf{I} \\to \\{(\\mathbf{b}_m, c_m)\\}_{m=1}^{M}\\), where \\(\\mathbf{b}_m \\in \\mathbb{R}^4\\) denotes the coordinates of each bounding box, and \\(c_m \\in C^{\\mathrm{test}}\\) denotes the class label of each bounding box." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.907, + 0.614 + ], + "angle": 0, + "content": "X-ray OvOD presents significant challenges due to the following reasons: i) Directly applying a pre-trained RGB OvOD detectors to X-ray images leads to suboptimal performance due to the visual modality gap, as shown in Fig. 1; ii) The scarcity of large-scale security X-ray datasets limits the application of RGB OvOD training techniques. These techniques typically rely on extensive image-text annotated data for strong supervision. To tackle these challenges, we introduce RAXO, a plug-and-play module that adapts any off-the-shelf RGB OvOD detectors to X-ray modalities. RAXO requires no training or in-domain detection annotations. Next, we present our approach." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.632, + 0.898, + 0.65 + ], + "angle": 0, + "content": "5. RAXO: Training-free Modality Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.901 + ], + "angle": 0, + "content": "As illustrated in Fig. 2, RAXO enables X-ray OvOD by constructing high-quality visual descriptors, \\(X_{c}\\), for each class \\(c\\) in the user-defined vocabulary, \\(C^{\\mathrm{test}}\\). To achieve this, RAXO first acquires X-ray samples in an open-ended manner, leveraging both in-house and web-based retrieval sources (Sec. 5.1). Then, it extracts the features of these samples and segments the relevant information, modeling the visual descriptor \\(X_{c}\\) (Sec. 5.2). Once the visual descriptors are constructed offline, they can be directly applied to any off-the-shelf OvOD detector by replacing the conventional text-based classifier \\(\\mathbf{W}\\), with our visual-based classifier \\(\\mathcal{X}\\) (Sec. 5.3). Thus, RAXO effectively overcomes the misalignment between X-ray features and text semantics, enabling OvOD detectors pre-trained with RGB data to accurately identify X-ray objects based on their intrinsic visual characteristics." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20772" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.089, + 0.885, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.364, + 0.907, + 0.435 + ], + "angle": 0, + "content": "Figure 2. Architecture of RAXO. For a given user-defined class \\( c \\in C^{\\mathrm{test}} \\), RAXO first retrieves its corresponding X-ray images \\( \\mathcal{G}_c^{XRAY} \\) from in-house and web sources, using its (1) Visual Samples Acquisition pipeline (Sec. 5.1). Following this, RAXO extracts the features of the images and segments them with its (2) Class Descriptor Modeling module (Sec. 5.2), creating ensemble visual descriptors for the class \\( \\mathcal{X}_c \\) and the background \\( \\mathcal{X}_{\\mathrm{bg}} \\). Finally, the text-based classifier from the baseline RGB OvOD detector is replaced with our (3) Visual-based Classifier (Sec. 5.3) \\( \\mathcal{X} \\), which yields accurate predictions on the X-ray modality." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.46, + 0.33, + 0.476 + ], + "angle": 0, + "content": "5.1. Visual Sample Acquisition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.482, + 0.484, + 0.587 + ], + "angle": 0, + "content": "Obtaining informative and representative images is crucial for generating robust visual descriptors. To this end, we propose a dual-source acquisition pipeline that retrieves a gallery of relevant X-ray samples \\(\\mathcal{G}_c^{\\mathrm{XRAY}}\\) for each user-specified class \\(c\\in C^{\\mathrm{test}}\\). This pipeline consists of two modules: an in-house retrieval module and a web-powered retrieval module." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.592, + 0.484, + 0.671 + ], + "angle": 0, + "content": "In-house retrieval. Given an in-domain X-ray dataset, \\(\\mathcal{D}_{\\mathrm{XRAY}}^{\\mathrm{in - house}}\\) containing categories \\(C^{\\mathrm{in - house}}\\) , our sample acquisition pipeline first attempts to retrieve the \\(K\\) most relevant images for the user-specified class \\(c\\) based on class name matching: \\(\\mathcal{G}_c^{\\mathrm{XRAY}} = \\bigl \\{\\mathbf{I}|\\bigl (\\mathbf{I},c\\bigr)\\in \\mathcal{D}_{\\mathrm{XRAY}}^{\\mathrm{in - house}}\\bigr \\} ^K\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.671, + 0.484, + 0.806 + ], + "angle": 0, + "content": "However, since the user-defined vocabulary \\( C^{\\mathrm{test}} \\) is open-ended, we cannot assume that every user-specified category will be present in \\( C^{\\mathrm{in-house}} \\). Consequently, for some categories, the retrieved set \\( \\mathcal{G}_c^{\\mathrm{XRAY}} \\) may be empty. To overcome this limitation and fully support open-vocabulary user input, RAXO further incorporates a novel web-powered retrieval module. This module retrieves RGB images from the web and applies a material transfer mechanism to synthesize X-ray-style representations, which we introduce next." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Web-powered retrieval. To obtain high-quality visual samples for a category \\( c \\) that is not present in the in-house vocabulary \\( C^{\\mathrm{in - house}} \\), we leverage the vast availability of web-based RGB image data \\( \\mathcal{D}_{\\mathrm{RGB}}^{\\mathrm{web}} \\) as an open-ended auxiliary source. Specifically, we perform text-based web retrieval using the class name \\( c \\) as a search" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.461, + 0.907, + 0.496 + ], + "angle": 0, + "content": "query, retrieving the top \\(K\\) results from Google Images as \\(\\widetilde{\\mathcal{G}}_c^{\\mathrm{web}} = \\left\\{\\mathbf{I}^{\\mathrm{RGB}} \\mid (\\mathbf{I}^{\\mathrm{RGB}}, c) \\in \\mathcal{D}_{\\mathrm{RGB}}^{\\mathrm{web}}\\right\\}^K\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.498, + 0.906, + 0.606 + ], + "angle": 0, + "content": "The raw web-retrieved results \\(\\widetilde{\\mathcal{G}}_c^{\\mathrm{web}}\\) are often noisy and may not always contain clear instances of the target class \\(c\\). To refine these results, we apply a filtering step using an RGB OvOD detector \\(\\mathcal{F}_{\\mathrm{RGB}}\\). Specifically, we discard images in the raw web-retrieved results where class \\(c\\) is not confidently detected, retaining only those where the detection confidence exceeds a threshold \\(\\tau\\) as \\(\\mathcal{G}_c^{\\mathrm{web}} = \\mathrm{Filter}(\\widetilde{\\mathcal{G}}_c^{\\mathrm{web}},\\mathcal{F}_{\\mathrm{RGB}},c,\\tau)\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.906, + 0.822 + ], + "angle": 0, + "content": "Material-transfer mechanism for web-retrieved images. The substantial visual disparity between RGB and X-ray modalities prevents direct use of web-retrieved RGB samples \\(\\mathcal{G}_c^{\\mathrm{web}}\\) for constructing X-ray class descriptors. Style transfer methods [7] fall short in bridging this gap, as they fail to capture the underlying material properties of objects (as shown in Tab. 4). To address this, we introduce a novel material-transfer mechanism for generating synthetic X-ray samples from web-retrieved images. As shown in Fig. 3, our approach consists of two key steps: \\(i)\\) constructing an offline material database \\(\\mathcal{M}\\) that encapsulates the expected X-ray appearance of various materials, and \\(ii)\\) adapting RGB samples \\(\\mathcal{G}_c^{\\mathrm{web}}\\) to X-ray style by applying the corresponding material properties from \\(\\mathcal{M}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To construct \\(\\mathcal{M}\\), we employ a Large Language Model (LLM) to cluster the class names of \\(C^{\\mathrm{in - house}}\\) into subsets \\(C_m\\subset C^{\\mathrm{in - house}}\\), where \\(C_m\\) contains classes that share the same \\(m\\) material - i.e., \" \\(C_{\\mathrm{metal}} = \\{\\mathrm{gun},\\mathrm{knife},\\mathrm{fork}\\}\\)\" or \" \\(C_{\\mathrm{leather}} = \\{\\mathrm{boot},\\mathrm{belt}\\}\\)\". Each material \\(m\\) is then associated" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20773" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.482, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.333, + 0.485, + 0.39 + ], + "angle": 0, + "content": "Figure 3. Web-powered retrieval and material-transfer mechanism for the class \"violin\". We retrieve violin samples from the web, filter them using \\(\\mathcal{F}_{\\mathrm{RGB}}\\), and inpaint the retrieved appearance into the object masks to generate synthetic X-ray samples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.484, + 0.462 + ], + "angle": 0, + "content": "with the corresponding set of images \\(\\mathcal{K}_m = \\{\\mathbf{I} \\mid (\\mathbf{I}, c_k) \\in \\mathcal{D}_{\\mathrm{XRAY}}^{\\mathrm{in-house}}, c_k \\in C_m\\}\\). These images are used to compute the appearance, \\(\\mathcal{A}_m \\in \\mathbb{R}^3\\), of the material as:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.467, + 0.483, + 0.511 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} _ {m} = \\frac {1}{| \\mathcal {K} _ {m} |} \\sum_ {\\mathbf {I} \\in \\mathcal {K} _ {m}} \\left(\\frac {\\sum_ {i , j} ^ {H , W} \\left(\\mathbf {I} _ {i , j} \\odot \\Omega \\left(\\mathbf {I} _ {i , j}\\right)\\right)}{\\sum \\Omega (\\mathbf {I})}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.519, + 0.484, + 0.594 + ], + "angle": 0, + "content": "where \\(\\odot\\) denotes element-wise multiplication and \\(\\Omega : \\mathbf{I} \\to \\{0,1\\}^{H \\times W}\\) is a segmentation function that produces a binary mask, to isolate the object of interest in \\(\\mathbf{I}\\). The resulting material-appearance pairs \\((m, \\mathcal{A}_m)\\) collectively form our material database \\(\\mathcal{M}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.595, + 0.484, + 0.716 + ], + "angle": 0, + "content": "Once \\(\\mathcal{M}\\) is computed offline, it can be used to adapt the web-retrieved RGB samples, \\(\\mathcal{G}_c^{\\mathrm{web}}\\), to the X-ray modality. To achieve this, we first prompt the LLM to link the class \\(c\\) with its corresponding material \\(m_c \\in \\mathcal{M}\\) (e.g., \\(c = \\text{\"violin\"}\\) is linked with the material \\(m_c = \\text{\"wood\"}\\)). Subsequently, the X-ray appearance \\(\\mathcal{A}_m^c\\) of the linked material is used to render the characteristic X-ray style on each web-retrieved RGB samples as:" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.723, + 0.483, + 0.743 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {c} ^ {\\mathrm {X R A Y}} = \\left\\{\\Omega \\left(\\mathbf {I} ^ {\\mathrm {R G B}}\\right) \\odot \\left(\\mathcal {A} _ {m} ^ {c} \\cdot \\mathbf {1}\\right) \\mid \\mathbf {I} ^ {\\mathrm {R G B}} \\in \\mathcal {G} _ {c} ^ {\\mathrm {w e b}} \\right\\}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.826 + ], + "angle": 0, + "content": "Finally, the X-ray visual gallery for every class freely specified by the user is built as \\(\\mathcal{G} = \\bigcup_{c\\in C^{\\mathrm{test}}}\\mathcal{G}_c^{\\mathrm{XRAY}}\\) through our dual-source retrieval pipeline, seamlessly integrating high-quality in-house X-ray samples with synthetic images derived from web data." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.334, + 0.851 + ], + "angle": 0, + "content": "5.2. Class Descriptor Modeling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The effectiveness of RAXO hinges on leveraging its visual gallery \\(\\mathcal{G}\\) to construct a robust visual descriptor \\(\\chi_{c}\\) that accurately captures the visual X-ray properties of class \\(c\\). A" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.196 + ], + "angle": 0, + "content": "naive approach, such as averaging the feature representations of all \\(\\mathcal{G}_c^{\\mathrm{XRAY}}\\) instances, fails to account for intra-class variability (e.g., distinct shapes of utility knives versus chef knives) and does not differentiate between foreground and background regions. To address these limitations, we propose a novel class descriptor modeling strategy shown in Fig. 2(2)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.197, + 0.907, + 0.379 + ], + "angle": 0, + "content": "First, we process each sample \\(\\mathbf{I} \\in \\mathcal{G}_c^{XRAY}\\) with \\(\\Omega\\) to separate the X-ray object from its background. Simultaneously, a feature extractor \\(\\phi\\) extracts per-patch embeddings \\(\\phi(\\mathbf{I}) \\in \\mathbb{R}^{H' \\times W' \\times D}\\), where \\(H' \\times W'\\) represents the number of spatial tokens, and \\(D\\) is the embedding dimension. Following [10], we resize (denoted by \\(\\zeta\\)) the segmentation mask to match the spatial dimensions of the feature map, and subsequently we use it to compute both a positive and a negative prototype. The positive prototype, designed to capture the significant visual features of the object without background interference, is computed as the average embedding over the foreground region as:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.387, + 0.905, + 0.427 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {\\mathbf {I}} ^ {\\text {p o s}} = \\frac {\\sum_ {i , j} ^ {H ^ {\\prime} , W ^ {\\prime}} \\zeta (\\Omega (\\mathbf {I} _ {i , j})) \\odot \\phi (\\mathbf {I} _ {i , j})}{\\sum \\zeta (\\Omega (\\mathbf {I}))}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.433, + 0.905, + 0.478 + ], + "angle": 0, + "content": "On the other hand, the negative prototype is obtained as the average embedding over the complementary region (background) as:" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.486, + 0.905, + 0.526 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {\\mathbf {I}} ^ {\\text {n e g}} = \\frac {\\sum_ {i , j} ^ {H ^ {\\prime} , W ^ {\\prime}} \\left(1 - \\zeta \\left(\\Omega \\left(\\mathbf {I} _ {\\mathbf {i} , \\mathbf {j}}\\right)\\right)\\right) \\odot \\phi \\left(\\mathbf {I} _ {\\mathbf {i} , \\mathbf {j}}\\right)}{\\sum \\left(1 - \\zeta \\left(\\Omega (\\mathbf {I})\\right)\\right)}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.533, + 0.905, + 0.594 + ], + "angle": 0, + "content": "Subsequently, to construct the final visual descriptor \\(\\mathcal{X}_c\\) of class \\(c\\), we compute the average positive prototype \\(\\overline{\\mathbf{x}}_c^{\\mathrm{pos}}\\) from \\(\\mathcal{G}_c^{XRAY}\\) and unite it with all the individual positive prototypes as:" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.602, + 0.905, + 0.621 + ], + "angle": 0, + "content": "\\[\n\\mathcal {X} _ {c} = \\left[ \\bar {\\mathbf {x}} _ {c} ^ {\\text {p o s}}, \\left\\{\\mathbf {x} _ {\\mathbf {I}} ^ {\\text {p o s}} \\mid \\mathbf {I} \\in \\mathcal {G} _ {c} \\right\\} \\right]. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.826 + ], + "angle": 0, + "content": "This formulation effectively captures both fine-grained object details and a holistic class-level representation, handling intra-class variability. Additionally, we also construct a global background descriptor by using the negative prototypes across the entire visual sample gallery as \\(\\mathcal{X}_{bg} = [\\overline{\\mathbf{x}}^{\\mathrm{neg}},\\{\\mathbf{x}_{\\mathbf{I}}^{\\mathrm{neg}}\\mid \\mathbf{I}\\in \\mathcal{G}\\} ]\\). This negative prototype is used to further improve detection reliability by filtering out low-quality proposals (e.g., a region proposed on the background) during inference. A key advantage of this approach is its modularity: visual descriptors \\(\\mathcal{X}_c\\) are computed off-line and can be incrementally expanded with new object categories, seamlessly integrating with the OvOD paradigm without training or requiring X-ray detection data." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.776, + 0.849 + ], + "angle": 0, + "content": "5.3. Classification is All You Need" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig. 2(3), once the visual descriptors are constructed for each class in the user-specified vocabulary, along with the background class, the RAXO classifier" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20774" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.483, + 0.123 + ], + "angle": 0, + "content": "\\(\\mathcal{X} = \\{\\left[\\mathcal{X}_c,\\mathcal{X}_{bg}\\right]\\mid c\\in C^{\\mathrm{test}}\\}\\) can be directly applied to any OvOD detector to classify proposals \\(\\mathbf{z_m}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.138, + 0.483, + 0.164 + ], + "angle": 0, + "content": "\\[\n\\hat {c} = \\underset {c \\in C ^ {\\text {t e s t}} ^ {\\prime}} {\\arg \\max } \\underset {\\mathbf {x} \\in \\mathcal {X}} {\\max } \\langle \\mathbf {z} _ {m}, \\mathbf {x} \\rangle , \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.172, + 0.484, + 0.264 + ], + "angle": 0, + "content": "where \\( C^{\\mathrm{test}} \\) denotes the test-time vocabulary \\( C^{\\mathrm{test}} \\) extended with the additional \"background\" class. If a proposal matches the background, it is removed. This simple-yet-effective strategy enables any pre-trained OvOD detectors to achieve strong performance on X-ray object detection with minimal modifications and no need for re-training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.268, + 0.484, + 0.465 + ], + "angle": 0, + "content": "Descriptor consistency criterion. To further reduce incorrect proposals from the OvOD detector, RAXO introduces a novel Descriptor Consistency Criterion (DCC) to enforce class consistency in predictions. DCC evaluates how well a proposal aligns with its predicted class relative to others, suppressing weakly aligned proposals. For a proposal \\(\\mathbf{z}_m\\), we measure its similarity to the closest prototype as \\(s_1 = \\max_{\\mathbf{x} \\in \\mathcal{X}_{\\hat{c}}} \\langle \\mathbf{z}_m, \\mathbf{x} \\rangle\\). We also compute its mean similarity to the average prototypes of all other classes as \\(s_2 = \\mathrm{avg}_{c \\in C^{\\mathrm{test}}, c \\neq \\hat{c}} \\langle \\mathbf{z}_m, \\bar{\\mathbf{x}}_c^{\\mathrm{pos}} \\rangle\\). The difference \\(\\Delta = s_1 - s_2\\) serves as a confidence measure, where a higher value indicates greater alignment with its predicted class than with any other. Proposals with \\(\\Delta\\) below a threshold \\(\\sigma\\) are discarded." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.48, + 0.223, + 0.496 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.505, + 0.484, + 0.7 + ], + "angle": 0, + "content": "Evaluation protocol. Conventional OvOD models are typically assessed using a Cross-Dataset Transfer Evaluation (CDTE) protocol, where the model is trained on one dataset and evaluated on a different dataset in a zero-shot manner [45]. In contrast, our setting involves a more challenging scenario where both the datasets and their underlying modalities differ. To address this, we introduce a novel Cross-Modality Transfer Evaluation (CMTE) protocol. Under CMTE, an OvOD model is trained on a source RGB dataset and subsequently evaluated on target \\(X\\)-ray datasets without any additional training or fine-tuning. For performance assessment, we employ the standard MS COCO [17] metrics: AP, AP50, and AP75." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Implementation details. Unless stated otherwise, RAXO's default configuration employs SAM 2 [27] as the object segmentation module \\(\\Omega\\), ViT-B/14 [3] pretrained with DINOv2 [25] as the visual encoder \\(\\phi\\), and GPT-4 [24] as the LLM. We set the consistency threshold to \\(\\sigma = 0.15\\), the similarity threshold to \\(\\tau = 0.5\\), and use \\(K = 30\\) samples to construct the visual descriptors. \\(\\mathcal{D}_{\\mathrm{XRAY}}^{in - house}\\) is composed of hold-out samples sourced from the evaluation X-ray datasets and is completely disjoint from the test split. \\(\\mathcal{D}_{\\mathrm{RGB}}^{web}\\) is constructed by collecting images from the web using the public Google Custom Search API [8]. A detailed analysis of the impact of hyperparameter choices is provided in Sec. 6.2. Further implementation details can be found in Supp. D." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.244 + ], + "angle": 0, + "content": "Datasets and baselines. We evaluate RAXO by integrating it with four different state-of-the-art OvOD detectors that were exclusively trained on RGB images: GroundingDINO [18], Detic [44], VLDet [15], and CoDet [21]. Our CMTE evaluation is performed across six diverse X-ray datasets: PIXray [19], PIDray [34], CLCXray [42], DvXray [20], HiXray [30], and our proposed DET-COMPASS, which together comprise 343 visible unique classes and over 140k images. A detailed description of these datasets is provided in Tab. 1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.255, + 0.825, + 0.27 + ], + "angle": 0, + "content": "6.1. Open-Vocabulary Detection Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.277, + 0.907, + 0.472 + ], + "angle": 0, + "content": "RAXO equips OvOD models with X-ray vision. We evaluate the X-ray adaptation capabilities of RAXO on the aforementioned datasets and baselines in Tab. 2. Since RAXO is training-free, we report the results directly on the test splits, accounting for all available categories. Furthermore, we analyze the influence of the gallery composition by varying \\(\\mathcal{G}\\) from a \\(0/100\\) configuration (i.e., all samples in \\(\\mathcal{G}\\) are retrieved from the web, as \\(C^{\\mathrm{test}} \\cap C^{\\mathrm{in-house}} = \\emptyset\\)) to a \\(100/0\\) configuration (i.e., all samples in \\(\\mathcal{G}\\) are retrieved from the in-house database, as \\(C^{\\mathrm{test}} \\subset C^{\\mathrm{in-house}}\\)). Each experiment is repeated three times with different random distributions of in-domain and web categories, and we report the averaged results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.474, + 0.909, + 0.732 + ], + "angle": 0, + "content": "As shown in Tab. 2, off-the-shelf OvOD detectors trained on RGB images perform poorly on X-ray data. In contrast, even without access to any in-domain data from \\( C^{\\mathrm{test}} \\) (0/100 setting), RAXO enhances baseline OvOD methods, yielding an average improvement of \\( \\uparrow 2.1 \\) points. This gain is further amplified when in-domain samples are incorporated into \\( \\mathcal{G} \\), which facilitates the construction of more robust materials in \\( \\mathcal{M} \\) and yields higher-quality visual descriptors. Notably, even a limited number of in-domain samples (20/80 setting) results in an average improvement of \\( \\uparrow 4.8 \\) points. Moreover, increasing the in-domain samples (50/50 setting) yields even more significant gains, particularly in datasets with a large number of classes, such as DET-COMPASS, where RAXO offers an average boost of \\( \\uparrow 14.9 \\) points. Finally, when \\( \\mathcal{G} \\) consists entirely of in-domain samples, RAXO delivers an average improvement of \\( \\uparrow 14.7 \\) points across all OvOD methods and datasets, all in a training-free manner." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "RAXO works with any OvOD detector. The consistent gains of RAXO across all evaluated detectors in Tab. 2, along with its simple integration (Sec. 5.3), confirm the generalization of our approach to any RGB OvOD model. Moreover, RAXO scales with the baseline method, yielding larger gains when integrated with more robust detectors (e.g., RAXO enhances G-DINO \\(\\uparrow\\)9.4 points on average across all \\(G\\) settings, compared to \\(\\uparrow\\)7.5 points for Detic). This highlights the advantage of our training-free method, which can leverage the rapid advancements in RGB OvOD detectors without needing large amounts of labeled X-ray data." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20775" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.119, + 0.089, + 0.874, + 0.438 + ], + "angle": 0, + "content": "
GMethodD-COMPASSPIXrayPIDrayCLCXrayDvXrayHiXrayAvg.
G-DINO [18]13.412.910.96.710.07.010.2
100/80/2050/5020/800/100+ RAXO47.9↑34.536.9↑24.016.5↑5.622.2↑15.522.6↑12.617.1↑10.127.2↑17.0
41.0↑27.633.8↑20.915.4↑4.518.0↑11.321.0↑11.014.5↑7.524.0↑13.8
31.4↑18.025.4↑12.515.5↑4.617.0↑10.316.1↑6.113.4↑6.419.8↑9.6
20.5↑7.121.6↑8.713.9↑3.010.0↑3.315.0↑5.09.8↑2.815.1↑4.9
14.0↑0.616.1↑3.213.4↑2.57.1↑0.412.4↑2.47.9↑0.911.8↑1.6
VLDet [15]10.69.86.94.47.45.17.4
100/80/2050/5020/800/100+ RAXO36.4↑25.832.3↑22.511.7↑4.815.4↑11.020.1↑12.714.8↑9.721.8↑14.4
31.8↑21.229.2↑19.411.0↑4.112.7↑8.316.8↑9.413.1↑8.019.1↑11.7
23.7↑13.124.0↑14.210.4↑3.511.1↑6.712.1↑4.711.2↑6.115.4↑8.0
16.2↑5.621.6↑11.89.4↑2.55.2↑0.810.6↑3.29.3↑4.212.1↑4.7
11.1↑0.514.1↑4.38.9↑2.04.4↑0.09.0↑1.68.3↑3.29.3↑1.9
Detic [44]11.59.37.14.77.04.87.4
100/80/2050/5020/800/100+ RAXO35.3↑23.827.3↑18.011.3↑4.214.0↑9.319.4↑12.414.2↑9.420.3↑12.9
30.7↑19.223.9↑14.610.8↑3.712.3↑7.618.0↑11.012.1↑7.318.0↑10.6
24.4↑12.919.5↑10.210.3↑3.29.2↑4.514.6↑7.611.0↑6.214.8↑7.4
16.4↑4.915.2↑5.99.6↑2.58.0↑3.312.7↑5.79.9↑5.112.0↑4.6
11.9↑0.413.4↑4.19.1↑2.05.2↑0.59.4↑2.47.9↑3.19.5↑2.1
CoDet [21]8.47.35.73.15.63.45.6
100/80/2050/5020/800/100+ RAXO35.8↑27.427.9↑20.610.3↑4.614.8↑11.717.6↑12.013.2↑9.819.9↑14.3
32.2↑23.825.1↑17.89.5↑3.812.0↑8.915.4↑9.811.7↑8.317.7↑12.1
24.0↑15.620.0↑12.79.5↑3.89.2↑6.111.5↑5.99.9↑6.514.0↑8.4
17.8↑9.414.8↑7.58.5↑2.85.1↑2.09.4↑3.88.1↑4.710.6↑5.0
12.2↑3.811.5↑4.28.1↑2.44.0↑0.96.9↑1.36.5↑3.18.2↑2.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.441, + 0.907, + 0.497 + ], + "angle": 0, + "content": "Table 2. X-ray OvOD performance under the Cross-Modality Transfer Evaluation (CMTE) setting on DET-COMPASS (ours), PINray [19], PIDray [34], CLCXray [42], DvXray [20], and HiXray [30] datasets. We integrate RAXO into different baselines using different gallery \\( \\mathcal{G} \\) compositions, from using only \\( {\\mathcal{D}}_{\\mathrm{{XRAY}}}^{\\text{in-house }} \\) data \\( \\left( {{100}/0}\\right) \\) to exclusively \\( {\\mathcal{D}}_{\\mathrm{{RGB}}}^{\\text{web }} \\) samples (0/100). RAXO consistently improves the performance of all baseline OvOD detectors across every dataset. AP is used." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.509, + 0.475, + 0.617 + ], + "angle": 0, + "content": "
PIXray
ScissorsWrenchBatteryPliersAP50
OVXD†16.946.214.66.421.0
BARON + RAXO18.239.48.322.022.0
CoDet + RAXO43.652.69.018.630.9
Detic + RAXO46.360.57.936.637.8
VLDet + RAXO48.362.710.638.540.0
G-DINO + RAXO49.761.611.351.543.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.622, + 0.483, + 0.678 + ], + "angle": 0, + "content": "Table 3. Comparison under OVXD [16] setting, where the methods do not have access to in-domain data from the displayed categories. \\(\\dagger\\)OVXD is a supervised method explicitly trained on X-ray images for alignment. AP50 is used." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "RAXO surpasses training-based approaches. We also compare RAXO with OVXD [16], the only concurrent work that extends X-ray detection beyond base categories. OVXD is a fully-supervised method trained directly on X-ray data, and its generalization capabilities are evaluated on a hold-out set comprising only four categories (i.e., Scissors, Wrench, Battery, and Pliers). In Tab. 3, we follow this setting and evaluate RAXO on PINray [19], excluding the aforementioned classes from the in-house dataset \\(\\mathcal{D}_{\\mathrm{XRAY}}^{\\mathrm{in - house}}\\). Our results demonstrate that RAXO outperforms OVXD even when both approaches utilize the same BARON detector [37]. Notably, unlike OVXD, our method does not retrain BARON or any other component. Yet, our training-free RAXO performs remarkably well, achieving an AP50" + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.51, + 0.892, + 0.611 + ], + "angle": 0, + "content": "
Mod. Filt. Trans.APAP50AP75
G-DINO [18]12.914.913.4
+ RAXOX-ray RGBS7.9↓5.09.8↓5.18.3↓5.1
13.9↑1.017.4↑2.514.3↑0.9
RGB✓14.3↑1.417.6↑2.714.7↑1.3
RGB✓18.3↑3.415.1↑1.7
RGB✓M16.1↑3.219.8↑4.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.617, + 0.907, + 0.687 + ], + "angle": 0, + "content": "Table 4. Ablation of web-based retrieval on PIXray [19] using only web-based samples (setting \\(0/100\\)). Mod. indicates retrieval of a specific modality from the web, Filt. refers to filtering retrieved images with \\(\\mathcal{F}_{\\mathrm{RGB}}\\), and Trans. specifies style-transfer via: \\(S\\) a diffusion-based method [7] or \\(M\\) our material-transfer mechanism." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.699, + 0.779, + 0.714 + ], + "angle": 0, + "content": "of 43.5 when paired with G-DINO [18]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.724, + 0.667, + 0.74 + ], + "angle": 0, + "content": "6.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.746, + 0.906, + 0.807 + ], + "angle": 0, + "content": "In this section we study the core components of RAXO. We conduct our experiments on PixRay [19], using GroundingDINO [13] as our baseline OvOD. Consistent findings are reported in Supp. E and H." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Web-based samples. Tab. 4 presents an ablation study on the components of our web-based visual sample acquisition pipeline (Sec. 5.1). Notably, directly retrieving X-ray samples from the web \\((\\mathrm{Mod.} = \\mathrm{X}\\) -ray) leads to a \\(\\downarrow 5.0\\) point decrease in AP. In contrast, retrieving RGB images from the web \\((\\mathrm{Mod.} = \\mathrm{RGB})\\) provides a modest AP improvement of" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20776" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.472, + 0.2 + ], + "angle": 0, + "content": "
Ω x neg X DCCAPAP50AP75
G-DINO [18]12.914.913.4
+ RAXO25.2↑12.331.4↑16.526.2↑12.8
✓27.1↑14.233.6↑18.728.4↑15.0
✓ ✓27.8↑14.934.5↑19.629.3↑15.9
✓27.5↑14.634.0↑19.128.8↑15.4
✓ ✓ ✓28.5↑15.635.3↑20.429.9↑16.5
✓ ✓ ✓36.9↑24.045.0↑30.139.0↑25.6
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.203, + 0.483, + 0.287 + ], + "angle": 0, + "content": "Table 5. Ablation of class representation construction on PIXray [19] using in-domain samples (setting \\(100/0\\)). \\(\\Omega\\) indicates building positive prototypes using segmentation masks. \\(\\mathbf{x}^{\\mathrm{neg}}\\) indicates use of negative prototypes. \\(\\mathcal{X}\\) indicates ensemble multiple prototypes in class descriptors. \\(DCC\\) indicates use of prototype consistency criterion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.484, + 0.497 + ], + "angle": 0, + "content": "\\(\\uparrow 1.0\\) point. We hypothesize that directly using X-ray examples obtained from the web is ineffective due to the limited availability of high-quality X-ray images online. Filtering the retrieved RGB images using \\(\\mathcal{F}_{\\mathrm{RGB}}\\) increases AP by an additional \\(\\uparrow 0.4\\) points. Nonetheless, they still exhibit a significant visual gap compared to their X-ray counterparts. A style-transfer approach using StyleShot [7] (Trans. \\(= S\\)) provides a marginal AP gain of \\(\\uparrow 0.3\\) points, likely due to its limited capacity to understand the intrinsic material properties of the objects. Conversely, our proposed material-transfer mechanism (Trans. \\(= M\\)) yields a substantial AP improvement of \\(\\uparrow 1.8\\) points, underscoring its effectiveness in bridging the X-ray modality gap." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.502, + 0.484, + 0.653 + ], + "angle": 0, + "content": "Class descriptor modeling. Tab. 5 presents an ablation study analyzing the impact of the various components used to construct class representations. As shown, using a simple per-class average of all the features from each sample already results in a significant AP improvement of \\(\\uparrow 12.3\\) points over the baseline OvOD. Refining the representation by applying an object segmentation method to isolate the foreground features \\((\\Omega)\\) further boosts the AP by \\(\\uparrow 1.9\\) points. Moreover, leveraging background features to construct negative prototypes \\((\\mathbf{x}^{\\mathrm{neg}})\\) adds an additional \\(\\uparrow 0.7\\) points." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.654, + 0.483, + 0.76 + ], + "angle": 0, + "content": "Replacing the averaged class prototypes with our proposed visual descriptor \\((\\mathcal{X})\\) yields another \\(\\uparrow 0.7\\) point improvement, demonstrating its enhanced capability to capture intra-class visual variability. Finally, incorporating our descriptor consistency criterion to suppress low-quality proposals \\((DCC)\\) provides a further AP increase of \\(\\uparrow 8.4\\) points, underscoring the effectiveness of our proposed framework." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.764, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Hyperparameter study. Fig. 4 analyzes the effect of varying the number of samples \\( K \\) used to construct the visual descriptors. Even with just a single sample per class, RAXO improves AP by \\( \\uparrow 6.5 \\) points over the baseline. As \\( K \\) increases, performance continues to improve, reaching saturation at \\( K = 30 \\) with a total AP gain of \\( \\uparrow 24.0 \\) points." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Furthermore, Fig. 5 explores the influence of the threshold \\(\\sigma\\) in the Descriptor Consistency Module. Higher values of \\(\\sigma\\) impose stricter consistency requirements, de" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.903, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.243, + 0.907, + 0.272 + ], + "angle": 0, + "content": "Figure 4. Impact of \\( K \\) in class representations evaluated on PINray [19] using a G-DINO [18] baseline in the 100/0 setting." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.284, + 0.903, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.45, + 0.906, + 0.478 + ], + "angle": 0, + "content": "Figure 5. Impact of \\(\\sigma\\) in Descriptor Consistency Module on PIxray [19] using a G-DINO [18] baseline in the \\(100/0\\) setting." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.495, + 0.907, + 0.556 + ], + "angle": 0, + "content": "manding greater separation between the predicted class and other classes before accepting a proposal. We observe that \\(\\sigma = 0.15\\) provides an optimal balance between false positives and false negatives, yielding an AP of 36.9 points." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.572, + 0.642, + 0.589 + ], + "angle": 0, + "content": "7. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.907, + 0.809 + ], + "angle": 0, + "content": "In this work, we pioneered the task of open-vocabulary object detection (OvOD) for X-ray imagery, a challenge shaped by the unique characteristics of X-ray data and the scarcity of annotated examples. To address this, we introduced RAXO, a training-free method that repurposes off-the-shelf RGB-based OvOD detectors for the X-ray domain. By leveraging intra-modal feature distances, a novel material-transfer mechanism, and robust class descriptor modeling, RAXO effectively bridges the modality gap between RGB and X-ray imagery. Our extensive experiments, conducted across multiple benchmarks and the newly proposed DET-COMPASS dataset, demonstrate that RAXO consistently enhances detection performance, achieving average mAP improvements of up to 17.0 points over baseline methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Looking ahead, the modularity of RAXO opens promising avenues for future research, including further refinement of visual descriptors and adaptation to additional modalities. Overall, our work not only sets a new state-of-the-art for X-ray open-vocabulary detection but also lays the groundwork for a thriving research direction in this emerging domain." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20777" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.258, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.114, + 0.486, + 0.352 + ], + "angle": 0, + "content": "We thank CINECA and the ISCRA initiative for the availability of high-performance computing resources. This work was partially supported by the EU HORIZON IAMI (HORIZON-CL3-2023-FCT-01-04-101168272) project, the EU HORIZON ELIAS (HORIZON-CL4-2022-HUMAN-02-101120237) project, the EU HORIZON ELLIOT (HORIZON-CL4-2024-HUMAN-03-101214398) project, the MUR PNRR FAIR (PE00000013) project funded by the NextGenerationEU, the Spanish Ministerio de Ciencia e Innovacion (grant numbers PID2020-112623GB-I00, PID2023-149549NB-I00), and the Galician Conselleria de Cultura, Educacion e Ordenacion Universitaria (2024-2027 ED431G-2023/04). Some of these grants are co-funded by the European Regional Development Fund (ERDF). Pablo Garcia-Fernandez is supported by the Spanish Ministerio de Universidades under the FPU national plan (grant number FPU21/05581)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.376, + 0.188, + 0.392 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.401, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[1] Samet Akcay and Toby Breckon. Towards automatic threat detection: A survey of advances of deep learning within x-ray security imaging. Pattern Recognition, 122:108245, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.457, + 0.483, + 0.512 + ], + "angle": 0, + "content": "[2] An Chang, Yu Zhang, Shunli Zhang, Leisheng Zhong, and Li Zhang. Detecting prohibited objects with physical size constraint from cluttered x-ray baggage images. Knowledge-Based Systems, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.513, + 0.483, + 0.594 + ], + "angle": 0, + "content": "[3] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.596, + 0.482, + 0.65 + ], + "angle": 0, + "content": "[4] Penghui Du, Yu Wang, Yifan Sun, Luting Wang, Yue Liao, Gang Zhang, Errui Ding, Yan Wang, Jingdong Wang, and Si Liu. Lami-detr: Open-vocabulary detection with language model instruction. In ECCV, 2024. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.652, + 0.482, + 0.692 + ], + "angle": 0, + "content": "[5] Ruohuan Fang, Guansong Pang, and Xiao Bai. Simple image-level classification improves open-vocabulary object detection. In AAAI, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.693, + 0.483, + 0.748 + ], + "angle": 0, + "content": "[6] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated images. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.749, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[7] Junyao Gao, Yanchen Liu, Yanan Sun, Yinhao Tang, Yanhong Zeng, Kai Chen, and Cairong Zhao. Styleshot: A snapshot on any style. arXiv:2407.01414, 2024. 4, 7, 8, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.79, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[8] Google LLC. Custom search JSON API reference. https: // developers.google.com/custom-search/v1, 2024. Accessed: 2025-03-07. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.483, + 0.859 + ], + "angle": 0, + "content": "[9] Lewis D. Griffin, Matthew Caldwell, and Jerone T. A. Andrews. COMPASS-XP. Zenodo, 2019. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[10] Laurynas Karazija, Iro Laina, Andrea Vedaldi, and Christian Rupprecht. Diffusion models for open-vocabulary segmentation. In ECCV, 2024. 5" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.401, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[11] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models. arXiv:2209.15639, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.906, + 0.203 + ], + "angle": 0, + "content": "[12] Liangqi Li, Jiaxu Miao, Dahu Shi, Wenming Tan, Ye Ren, Yi Yang, and Shiliang Pu. Distilling detr with visual-linguistic knowledge for open-vocabulary object detection. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[13] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.906, + 0.327 + ], + "angle": 0, + "content": "[14] Mingyuan Li, Tong Jia, Hao Wang, Bowen Ma, Hui Lu, Shuyang Lin, Da Cai, and Dongyue Chen. Ao-detr: Anti-overlapping detr for x-ray prohibited items detection. IEEE Transactions on Neural Networks and Learning Systems, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.384 + ], + "angle": 0, + "content": "[15] Chuang Lin, Peize Sun, Yi Jiang, Ping Luo, Lizhen Qu, Gholamreza Haffari, Zehuan Yuan, and Jianfei Cai. Learning object-language alignments for open-vocabulary object detection. In ICLR, 2023. 1, 2, 6, 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[16] Shuyang Lin, Tong Jia, Hao Wang, Bowen Ma, Mingyuan Li, and Dongyue Chen. Detection of novel prohibited item categories for real-world security inspection. Eng. Appl. Artif. Intell., 2025. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[17] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[18] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding DINO: marrying DINO with grounded pre-training for open-set object detection. In ECCV, 2024. 1, 2, 6, 7, 8, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[19] Bowen Ma, Tong Jia, Min Su, Xiaodong Jia, Dongyue Chen, and Yichun Zhang. Automated segmentation of prohibited items in x-ray baggage images using dense de-overlap attention snake. TMM, 2022. 2, 3, 6, 7, 8, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.906, + 0.691 + ], + "angle": 0, + "content": "[20] Bowen Ma, Tong Jia, Mingyuan Li, Songsheng Wu, Hao Wang, and Dongyue Chen. Towards dual-view x-ray baggage inspection: A large-scale benchmark and adaptive hierarchical cross refinement for prohibited item discovery. IEEE TIFS, 2024. 1, 3, 6, 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.906, + 0.746 + ], + "angle": 0, + "content": "[21] Chuofan Ma, Yi Jiang, Xin Wen, Zehuan Yuan, and Xiaojuan Qi. CoDet: Co-occurrence guided region-word alignment for open-vocabulary object detection. In NeurIPS, 2023. 2, 6, 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[22] Caijing Miao, Lingxi Xie, Fang Wan, Chi Su, Hongye Liu, Jianbin Jiao, and Qixiang Ye. Sixray: A large-scale security inspection x-ray benchmark for prohibited item discovery in overlapping images. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[23] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, et al. Simple open-vocabulary object detection. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[24] OpenAI. GPT-4 technical report. CoRR, abs/2303.08774, 2023.6" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.093, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20778" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[25] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv:2304.07193, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[27] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dólar, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv:2408.00714, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.401 + ], + "angle": 0, + "content": "[28] Haifeng Sima, Bailiang Chen, Chaosheng Tang, Yudong Zhang, and Junding Sun. Multi-scale feature attention-detection transformer: Multi-scale feature attention for security check object detection. IET Computer Vision, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.404, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[29] Archana Singh and Dhiraj. Advancements in machine learning techniques for threat item detection in x-ray images: a comprehensive survey. Int. J. Multim. Inf. Retr., 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.447, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[30] Renshuai Tao, Yanlu Wei, Xiangjian Jiang, Hainan Li, Haotong Qin, Jiakai Wang, Yuqing Ma, Libo Zhang, and Xi-anglong Liu. Towards real-world x-ray security inspection: A high-quality benchmark and lateral inhibition module for prohibited items detection. In ICCV, 2021. 2, 3, 6, 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.518, + 0.482, + 0.586 + ], + "angle": 0, + "content": "[31] Renshuai Tao, Hainan Li, Tianbo Wang, Yanlu Wei, Yifu Ding, Bowei Jin, Hongping Zhi, Xianglong Liu, and Aishan Liu. Exploring endogenous shift for cross-domain detection: A large-scale benchmark and perturbation suppression network. In CVPR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.589, + 0.482, + 0.643 + ], + "angle": 0, + "content": "[32] Renshuai Tao, Tianbo Wang, Ziyang Wu, Cong Liu, Aishan Liu, and Xianglong Liu. Few-shot x-ray prohibited item detection: A benchmark and weak-feature enhancement network. In ACMMM, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.482, + 0.7 + ], + "angle": 0, + "content": "[33] Divya Velayudhan, Taimur Hassan, Ernesto Damiani, and Naoufel Werghi. Recent advances in baggage threat detection: A comprehensive and systematic survey. ACM Computing Surveys, 55(8):1-38, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[34] Boying Wang, Libo Zhang, Longyin Wen, Xianglong Liu, and Yanjun Wu. Towards real-world prohibited item detection: A large-scale x-ray benchmark. In ICCV, 2021. 2, 3, 6, 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.482, + 0.828 + ], + "angle": 0, + "content": "[35] Ruxue Wang, Yuliang Shi, and Mingyu Cai. Optimization and research of suspicious object detection algorithm in x-ray image. In 2023 IEEE 6th Information Technology, Networking, Electronic and Automation Control Conference (ITNEC), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.884 + ], + "angle": 0, + "content": "[36] Yanlu Wei, Renshuai Tao, Zhangjie Wu, Yuqing Ma, Libo Zhang, and Xianglong Liu. Occluded prohibited items detection: An x-ray security inspection benchmark and de-occlusion attention module. In ACMMM, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.887, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] Size Wu, Wenwei Zhang, Sheng Jin, Wentao Liu, and" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "Chen Change Loy. Aligning bag of regions for open-vocabulary object detection. In CVPR, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[38] Size Wu, Wenwei Zhang, Lumin Xu, Sheng Jin, Xiangtai Li, Wentao Liu, and Chen Change Loy. Clipself: Vision transformer distills itself for open-vocabulary dense prediction. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.221 + ], + "angle": 0, + "content": "[39] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[40] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. Detclip: Dictionary-enriched visual-concept paralleled pretraining for open-world detection. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[41] Lewei Yao, Jianhua Han, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, and Hang Xu. Detclipv2: Scalable open-vocabulary object detection pre-training via word-region alignment. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[42] Cairong Zhao, Liang Zhu, Shuguang Dou, Weihong Deng, and Liang Wang. Detecting overlapped objects in x-ray security imagery by a label-aware mechanism. IEEE TIFS, 2022, 2, 3, 6, 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[43] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[44] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, 2022. 1, 2, 3, 6, 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.544 + ], + "angle": 0, + "content": "[45] Chaoyang Zhu and Long Chen. A survey on open-vocabulary detection and segmentation: Past, present, and future. IEEE TPAMI, 2024. 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20779" + } + ] +] \ No newline at end of file diff --git a/2025/Superpowering Open-Vocabulary Object Detectors for X-ray Vision/60987d09-4da2-484e-8cc1-bea5f240eaf8_origin.pdf b/2025/Superpowering Open-Vocabulary Object Detectors for X-ray Vision/60987d09-4da2-484e-8cc1-bea5f240eaf8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..14013eeb1073788800470c8a919cd393a49f5805 --- /dev/null +++ b/2025/Superpowering Open-Vocabulary Object Detectors for X-ray Vision/60987d09-4da2-484e-8cc1-bea5f240eaf8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88e427598c8175df392c6a9a24c446e86cf0583f90f718dd7fe89c37a237dfe0 +size 1167726 diff --git a/2025/Superpowering Open-Vocabulary Object Detectors for X-ray Vision/full.md b/2025/Superpowering Open-Vocabulary Object Detectors for X-ray Vision/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2cca2bb792ee0e742638fa044a68db2fbf95cb3e --- /dev/null +++ b/2025/Superpowering Open-Vocabulary Object Detectors for X-ray Vision/full.md @@ -0,0 +1,279 @@ +# Superpowering Open-Vocabulary Object Detectors for X-ray Vision + +Pablo Garcia-Fernandez $^{1*}$ Lorenzo Vaquero $^{1,2}$ Mingxuan Liu $^{3}$ Feng Xue $^{3*}$ Daniel Cores $^{1}$ Nicu Sebe $^{3}$ Manuel Mucientes $^{1}$ Elisa Ricci $^{2,3}$ + +1University of Santiago de Compostela, Spain 2Fondazione Bruno Kessler, Italy 3University of Trento, Italy + +{pablogarcia.fernandez, daniel.cores, manuel.mucientes}@usc.es lvaquerootal@fbk.eu + +{mingxuan.liu, feng.xue, niculae.sebe, e.ricci}@unitn.it *Corresponding authors + +# Abstract + +Open-vocabulary object detection (OvOD) is set to revolutionize security screening by enabling systems to recognize any item in X-ray scans. However, developing effective OvOD models for X-ray imaging presents unique challenges due to data scarcity and the modality gap that prevents direct adoption of RGB-based solutions. To overcome these limitations, we propose RAXO, a training-free framework that repurposes off-the-shelf RGB OvOD detectors for robust X-ray detection. RAXO builds high-quality X-ray class descriptors using a dual-source retrieval strategy. It gathers relevant RGB images from the web and enriches them via a novel X-ray material transfer mechanism, eliminating the need for labeled databases. These visual descriptors replace text-based classification in OvOD, leveraging intra-modal feature distances for robust detection. Extensive experiments demonstrate that RAXO consistently improves OvOD performance, providing an average mAP increase of up to 17.0 points over base detectors. To further support research in this emerging field, we also introduce DET-COMPASS, a new benchmark featuring bounding box annotations for over 300 object categories, enabling large-scale evaluation of OvOD in X-ray. Code and dataset available at: https://pagf188.github.io/RAXO/. + +# 1. Introduction + +Automated object detection technologies for X-ray imaging are essential to maintain public safety, enabling the identification of prohibited items at checkpoints in high-risk environments such as airports, train stations, museums and stadiums [29]. These systems improve security while simultaneously reducing the workload of human inspectors. + +Conventional X-ray object detectors rely on supervised learning [2, 14, 20] and are inherently limited by the object categories present in their training datasets (see Fig. 1a). This limitation is exacerbated by the high cost of X-ray machinery and the requirement for expert annotation, often restricting these systems to fewer than 20 object classes [32], + +![](images/25cf5ef969acd77759db2ca01adf2e5caba7178cbccd2f42808b19029b75cc47.jpg) +Figure 1. (a) Traditional X-ray object detectors are constrained by the limited categories in their training datasets. (b) We introduce the task of open-vocabulary object detection (OvOD) for X-ray imaging and propose RAXO, a training-free method that adapts off-the-shelf RGB OvOD models to X-ray data. (c) RAXO greatly improves detection performance across multiple benchmarks. + +thereby hindering broader real-world applications. + +In light of the expanding diversity of man-made objects and evolving security demands, an open-vocabulary object detection (OvOD) framework capable of recognizing arbitrary X-ray object categories defined by the user is imperative. Despite its importance, open-vocabulary detection in the X-ray domain remains largely unexplored. Concurrent works aiming to extend beyond base categories have managed to generalize to at most four unseen classes [16], thereby highlighting the inherent challenges of the task. + +Recent advances in OvOD for conventional RGB images have been driven by large-scale annotated datasets, which facilitate effective alignment between visual and textual features [15, 18, 44]. However, these advancements do not directly transfer to X-ray imagery. In practice, applying state-of-the-art RGB-based OvOD models to X-ray scans leads to significant performance degradation, as illustrated + +in Fig. 1c. Moreover, retraining these models on X-ray data is often impractical given the scarcity of large-scale annotated X-ray datasets. + +Motivated by these challenges, this work opens up open-vocabulary object detection for X-ray imaging by repurposing robust, off-the-shelf RGB-based detectors. To this end, we propose RAXO (tRAining-free adaptation for X-ray Open-vocabulary detection), a training-free method that seamlessly adapts RGB OvOD models to X-ray. (Fig. 1b). + +We find that the main reason RGB-based OvOD detectors fail in the X-ray domain is the disruption of text-visual feature alignment. The appearance disparity between the same object in RGB and X-ray modalities causes textual embeddings, aligned with RGB features, to mismatch with X-ray visual features. RAXO addresses this issue on three simple steps: (1) visual sample acquisition, which involves obtaining X-ray images that represent user-defined categories; (2) class descriptor modeling, which uses these images to construct descriptors that effectively encode the visual appearance of each category within the X-ray domain; and (3) classifier construction, where the computed descriptors are used to build a visual classifier that replaces the inter-modal (i.e., text-to-visual) classification of conventional OvODs. This new classifier exploits the fact that, despite the modality shift, intra-modal (e.g., visual-to-visual) feature distances remain reliable indicators for object identification. In this way RAXO enables any OvOD to successfully detect X-ray objects based on their inherent visual properties. + +To facilitate comprehensive evaluation, we also introduce DET-COMPASS, a detection dataset comprising 370 distinct object classes with paired X-ray and RGB annotations. Extensive experiments demonstrate that our approach consistently improves detection performance by an average of $\uparrow 8.4$ AP across multiple benchmarks and OvOD models, with gains that scale as the underlying detectors become more powerful (Tab. 2). In summary, our contributions are: + +- We formally introduce the problem of open-vocabulary object detection for X-ray imagery without training, addressing a critical need for security in real-world scenarios. +- We introduce DET-COMPASS, a novel benchmark with bounding box annotations across 370 object categories, enabling standardized evaluation of OvOD methods in the X-ray modality. +- We propose RAXO, a novel training-free approach that repurposes off-the-shelf RGB-based OvOD models for X-ray detection by constructing robust visual descriptors. RAXO achieves new state-of-the-art results across multiple benchmarks. + +# 2. Related Work + +Open-vocabulary object detection has advanced rapidly with the advent of Vision-Language Models (VLMs), enabling detectors to generalize beyond fixed categories. + +OvOD detectors leverage weak supervision signals to improve detection accuracy. Based on the type of weak supervision utilized, OvOD methods can be categorized into (i) region-aware training, (ii) pseudo-labeling, (iii) knowledge distillation, and (iv) transfer learning. + +Region-aware training methods aim to improve localization and feature representation by refining the alignment between image regions and their corresponding textual descriptions. Approaches such as DetCLIP [40], DetCLIPv2 [41], CORA [39], and VLDet [15] adopt this strategy. Pseudo-labeling methods, rely on large pretrained VLMs to generate pseudo-labels, effectively expanding the training set. Methods like RegionCLIP [43], PromptDet [6], CoDET [21], GLIP [13], Detric [44], and Grounding DINO [18] follow this approach. Knowledge distillation techniques, such as BARON [37], DK-DETR [12], CLIPSelf [38], and SIC-CADS [5], employ VLMs as teachers in a teacher-student framework, transferring knowledge from the VLM image encoder to enhance the detector backbone. In contrast, transfer learning approaches integrate the VLM encoder directly, either through fine-tuning, as seen in OWL-ViT [23], or by freezing the encoder, as in F-VLM [11]. + +Regardless of the training strategy, OvOD methods heavily depend on VLMs trained on large-scale datasets. However, such datasets are unavailable in the X-ray modality, making this approach impractical. To address this, RAXO seamlessly adapts existing RGB-based OvOD methods for X-ray object detection without requiring additional training. X-ray object detection. Several datasets have been developed to tackle the challenge of object detection in X-ray imagery [1, 33]. SIXray [22], OPIXray [36], and CLCXray [42] focused on detecting occluded objects, introducing strategies to identify deliberately hidden prohibited object. To enhance dataset scale and diversity, PIDray [34] and HiXray [30] provided larger benchmarks, significantly increasing the number of annotated images. Meanwhile, PIXray [19] pioneered the introduction of an X-ray segmentation dataset and proposed a real-time framework for segmenting prohibited objects, further advancing automated threat detection in security screening applications. + +Previous X-ray object detectors [28, 31, 35] considered mainly a closed-set paradigm, thus being limited to localize objects of a small set of predefined categories. More recently, Lin et al. [16] proposed fine-tuning a CLIP adapter to bridge the modality gap between CLIP's training data and X-ray images, thereby achieving OvOD in X-ray data. They demonstrated generalization on four novel object classes, but both training and evaluation were limited by the lack of largescale, well-annotated X-ray data. To address these issues, RAXO leverages data retrieval from the web to expand detection capabilities to a broader range of common objects. Furthermore, our proposed re-labeled DET-COMPASS dataset enables evaluation across a wider variety of classes. + +
VenueImagesClassesModality
DvXray [20]TIFS’2432,00015X-ray
PIXray [19]TMM’225,04615X-ray
CLCXray [42]TIFS’229,56512X-ray
FSOD [32]ACMMM’2212,33320X-ray
EDS [31]CVPR’2214,21910X-ray
PIDray [34]ICCV’2147,67712X-ray
HiXray [30]ICCV’2145,3658X-ray
DET-COMPASS (Ours)-1,928370X-ray+RGB
+ +Table 1. Existing X-ray detection datasets. DET-COMPASS is the dataset with the highest number of categories and the only one providing pixel-level alignment between X-ray and RGB data. + +# 3. DET-COMPASS + +Object detection in security X-ray scans has advanced significantly in recent years. However, evaluating OvOD detectors in this modality remains challenging due to the limited number of annotated object categories in existing X-ray benchmarks. For instance, the largest X-ray detection dataset, FSOD [32], includes annotations for only 20 classes (see Tab. 1). This limitation severely constrains the comprehensive evaluation of OvOD methods, which require a broad and diverse category set (or say vocabulary) to assess generalization to unseen object semantics. + +To address this gap, we introduce DET-COMPASS, a novel benchmark that repurposes the COMPASS-XP classification dataset [9] for object detection through meticulous manual bounding box annotation. DET-COMPASS comprises 370 distinct object classes, offering an order-of-magnitude increase in vocabulary size over previous X-ray detection benchmarks. Additionally, it provides pixel-aligned RGB images, ensuring precise spatial correspondence across modalities and facilitating the development of multimodal models. Each object is also labeled with a visibility attribute, indicating whether it produces a discernible signature in the X-ray spectrum. Further details are provided in Appendix C. + +As summarized in Tab. 1, DET-COMPASS sets a new standard in class diversity and uniquely integrates multimodal, pixel-aligned X-ray and RGB data. The dataset will be released under an open license, serving as a valuable resource for advancing OvOD research in security screening and industrial inspection. + +# 4. OvOD for X-ray Imaging + +Preliminaries: OvOD in the RGB domain. Most RGB OvOD detectors [44, 45] follow a two-stage pipeline. During training, a region proposal network (RPN) is learned to yield a set of $M$ proposals by $\{\mathbf{z}_m\}_{m=1}^M = \Phi_{\mathrm{RPN}}(\mathbf{I}^{RGB})$ , where $\mathbf{z}_m \in \mathbb{R}^D$ is a $D$ -dimensional region-of-interest (RoI) feature embedding. Then, a bounding box regressor predicts coordinates for each proposed region via $\hat{\mathbf{b}}_m = \Phi_{\mathrm{REG}}(\mathbf{z}_m)$ . A set of text-based classifiers $\mathbf{W} = \{\mathbf{w}_c | \mathbf{w}_c \in \mathbb{R}^D\}_{c=1}^{|C_{\text{train}}|}$ are used + +to compute classification scores for each region as $\langle \mathbf{w}_c, \mathbf{z}_m \rangle$ , where $\langle \cdot, \cdot \rangle$ is the cosine similarity function and $C^{\mathrm{train}}$ denotes the training vocabulary. In this way, each region's class is determined by the class with the highest score. Here, the classifier $\mathbf{W}$ is constructed by encoding class names in $C^{\mathrm{train}}$ using a pre-trained VLM text encoder, e.g., CLIP [26]. During training, OvOD models update all parameters while keeping $\mathbf{W}$ frozen. This enables RGB-region-class alignment by leveraging large-scale RGB data and the pre-aligned vision-language semantic space of VLMs, facilitating open-vocabulary inference with any test vocabulary $C^{\mathrm{test}}$ . The vocabularies $C^{\mathrm{train}}$ and $C^{\mathrm{test}}$ may be disjoint or overlapping. + +Problem formulation. In this work, we study open-vocabulary object detection (OvOD) in X-ray modality. Specifically, given an input X-ray image $\mathbf{I}$ and a vocabulary $C^{\mathrm{test}}$ defined by users at test time, an X-ray OvOD detector $\mathcal{F}_{\mathrm{X - ray}}$ aims to detect objects specified in $C^{\mathrm{test}}$ from $\mathbf{I}$ (e.g., a "Power bank" in a passenger's backpack scan at an airport security checkpoint). Theoretically, this detection process can be formulated as $\mathcal{F}_{\mathrm{X - ray}}: \mathbf{I} \to \{(\mathbf{b}_m, c_m)\}_{m=1}^{M}$ , where $\mathbf{b}_m \in \mathbb{R}^4$ denotes the coordinates of each bounding box, and $c_m \in C^{\mathrm{test}}$ denotes the class label of each bounding box. + +X-ray OvOD presents significant challenges due to the following reasons: i) Directly applying a pre-trained RGB OvOD detectors to X-ray images leads to suboptimal performance due to the visual modality gap, as shown in Fig. 1; ii) The scarcity of large-scale security X-ray datasets limits the application of RGB OvOD training techniques. These techniques typically rely on extensive image-text annotated data for strong supervision. To tackle these challenges, we introduce RAXO, a plug-and-play module that adapts any off-the-shelf RGB OvOD detectors to X-ray modalities. RAXO requires no training or in-domain detection annotations. Next, we present our approach. + +# 5. RAXO: Training-free Modality Adaptation + +As illustrated in Fig. 2, RAXO enables X-ray OvOD by constructing high-quality visual descriptors, $X_{c}$ , for each class $c$ in the user-defined vocabulary, $C^{\mathrm{test}}$ . To achieve this, RAXO first acquires X-ray samples in an open-ended manner, leveraging both in-house and web-based retrieval sources (Sec. 5.1). Then, it extracts the features of these samples and segments the relevant information, modeling the visual descriptor $X_{c}$ (Sec. 5.2). Once the visual descriptors are constructed offline, they can be directly applied to any off-the-shelf OvOD detector by replacing the conventional text-based classifier $\mathbf{W}$ , with our visual-based classifier $\mathcal{X}$ (Sec. 5.3). Thus, RAXO effectively overcomes the misalignment between X-ray features and text semantics, enabling OvOD detectors pre-trained with RGB data to accurately identify X-ray objects based on their intrinsic visual characteristics. + +![](images/5a5b7f18c2a2e2a5e1ac03b8167e8cfb2316ed1a2d351756ee24b2a9607f162f.jpg) +Figure 2. Architecture of RAXO. For a given user-defined class $c \in C^{\mathrm{test}}$ , RAXO first retrieves its corresponding X-ray images $\mathcal{G}_c^{XRAY}$ from in-house and web sources, using its (1) Visual Samples Acquisition pipeline (Sec. 5.1). Following this, RAXO extracts the features of the images and segments them with its (2) Class Descriptor Modeling module (Sec. 5.2), creating ensemble visual descriptors for the class $\mathcal{X}_c$ and the background $\mathcal{X}_{\mathrm{bg}}$ . Finally, the text-based classifier from the baseline RGB OvOD detector is replaced with our (3) Visual-based Classifier (Sec. 5.3) $\mathcal{X}$ , which yields accurate predictions on the X-ray modality. + +# 5.1. Visual Sample Acquisition + +Obtaining informative and representative images is crucial for generating robust visual descriptors. To this end, we propose a dual-source acquisition pipeline that retrieves a gallery of relevant X-ray samples $\mathcal{G}_c^{\mathrm{XRAY}}$ for each user-specified class $c\in C^{\mathrm{test}}$ . This pipeline consists of two modules: an in-house retrieval module and a web-powered retrieval module. + +In-house retrieval. Given an in-domain X-ray dataset, $\mathcal{D}_{\mathrm{XRAY}}^{\mathrm{in - house}}$ containing categories $C^{\mathrm{in - house}}$ , our sample acquisition pipeline first attempts to retrieve the $K$ most relevant images for the user-specified class $c$ based on class name matching: $\mathcal{G}_c^{\mathrm{XRAY}} = \bigl \{\mathbf{I}|\bigl (\mathbf{I},c\bigr)\in \mathcal{D}_{\mathrm{XRAY}}^{\mathrm{in - house}}\bigr \} ^K$ + +However, since the user-defined vocabulary $C^{\mathrm{test}}$ is open-ended, we cannot assume that every user-specified category will be present in $C^{\mathrm{in-house}}$ . Consequently, for some categories, the retrieved set $\mathcal{G}_c^{\mathrm{XRAY}}$ may be empty. To overcome this limitation and fully support open-vocabulary user input, RAXO further incorporates a novel web-powered retrieval module. This module retrieves RGB images from the web and applies a material transfer mechanism to synthesize X-ray-style representations, which we introduce next. + +Web-powered retrieval. To obtain high-quality visual samples for a category $c$ that is not present in the in-house vocabulary $C^{\mathrm{in - house}}$ , we leverage the vast availability of web-based RGB image data $\mathcal{D}_{\mathrm{RGB}}^{\mathrm{web}}$ as an open-ended auxiliary source. Specifically, we perform text-based web retrieval using the class name $c$ as a search + +query, retrieving the top $K$ results from Google Images as $\widetilde{\mathcal{G}}_c^{\mathrm{web}} = \left\{\mathbf{I}^{\mathrm{RGB}} \mid (\mathbf{I}^{\mathrm{RGB}}, c) \in \mathcal{D}_{\mathrm{RGB}}^{\mathrm{web}}\right\}^K$ . + +The raw web-retrieved results $\widetilde{\mathcal{G}}_c^{\mathrm{web}}$ are often noisy and may not always contain clear instances of the target class $c$ . To refine these results, we apply a filtering step using an RGB OvOD detector $\mathcal{F}_{\mathrm{RGB}}$ . Specifically, we discard images in the raw web-retrieved results where class $c$ is not confidently detected, retaining only those where the detection confidence exceeds a threshold $\tau$ as $\mathcal{G}_c^{\mathrm{web}} = \mathrm{Filter}(\widetilde{\mathcal{G}}_c^{\mathrm{web}},\mathcal{F}_{\mathrm{RGB}},c,\tau)$ + +Material-transfer mechanism for web-retrieved images. The substantial visual disparity between RGB and X-ray modalities prevents direct use of web-retrieved RGB samples $\mathcal{G}_c^{\mathrm{web}}$ for constructing X-ray class descriptors. Style transfer methods [7] fall short in bridging this gap, as they fail to capture the underlying material properties of objects (as shown in Tab. 4). To address this, we introduce a novel material-transfer mechanism for generating synthetic X-ray samples from web-retrieved images. As shown in Fig. 3, our approach consists of two key steps: $i)$ constructing an offline material database $\mathcal{M}$ that encapsulates the expected X-ray appearance of various materials, and $ii)$ adapting RGB samples $\mathcal{G}_c^{\mathrm{web}}$ to X-ray style by applying the corresponding material properties from $\mathcal{M}$ . + +To construct $\mathcal{M}$ , we employ a Large Language Model (LLM) to cluster the class names of $C^{\mathrm{in - house}}$ into subsets $C_m\subset C^{\mathrm{in - house}}$ , where $C_m$ contains classes that share the same $m$ material - i.e., " $C_{\mathrm{metal}} = \{\mathrm{gun},\mathrm{knife},\mathrm{fork}\}$ " or " $C_{\mathrm{leather}} = \{\mathrm{boot},\mathrm{belt}\}$ ". Each material $m$ is then associated + +![](images/a88c4c7f2a5b725a52a3a73abfc50ad5408cb2bd816543e140661b8258b0ac43.jpg) +Figure 3. Web-powered retrieval and material-transfer mechanism for the class "violin". We retrieve violin samples from the web, filter them using $\mathcal{F}_{\mathrm{RGB}}$ , and inpaint the retrieved appearance into the object masks to generate synthetic X-ray samples. + +with the corresponding set of images $\mathcal{K}_m = \{\mathbf{I} \mid (\mathbf{I}, c_k) \in \mathcal{D}_{\mathrm{XRAY}}^{\mathrm{in-house}}, c_k \in C_m\}$ . These images are used to compute the appearance, $\mathcal{A}_m \in \mathbb{R}^3$ , of the material as: + +$$ +\mathcal {A} _ {m} = \frac {1}{| \mathcal {K} _ {m} |} \sum_ {\mathbf {I} \in \mathcal {K} _ {m}} \left(\frac {\sum_ {i , j} ^ {H , W} \left(\mathbf {I} _ {i , j} \odot \Omega \left(\mathbf {I} _ {i , j}\right)\right)}{\sum \Omega (\mathbf {I})}\right), \tag {1} +$$ + +where $\odot$ denotes element-wise multiplication and $\Omega : \mathbf{I} \to \{0,1\}^{H \times W}$ is a segmentation function that produces a binary mask, to isolate the object of interest in $\mathbf{I}$ . The resulting material-appearance pairs $(m, \mathcal{A}_m)$ collectively form our material database $\mathcal{M}$ . + +Once $\mathcal{M}$ is computed offline, it can be used to adapt the web-retrieved RGB samples, $\mathcal{G}_c^{\mathrm{web}}$ , to the X-ray modality. To achieve this, we first prompt the LLM to link the class $c$ with its corresponding material $m_c \in \mathcal{M}$ (e.g., $c = \text{"violin"}$ is linked with the material $m_c = \text{"wood"}$ ). Subsequently, the X-ray appearance $\mathcal{A}_m^c$ of the linked material is used to render the characteristic X-ray style on each web-retrieved RGB samples as: + +$$ +\mathcal {G} _ {c} ^ {\mathrm {X R A Y}} = \left\{\Omega \left(\mathbf {I} ^ {\mathrm {R G B}}\right) \odot \left(\mathcal {A} _ {m} ^ {c} \cdot \mathbf {1}\right) \mid \mathbf {I} ^ {\mathrm {R G B}} \in \mathcal {G} _ {c} ^ {\mathrm {w e b}} \right\}, \tag {2} +$$ + +Finally, the X-ray visual gallery for every class freely specified by the user is built as $\mathcal{G} = \bigcup_{c\in C^{\mathrm{test}}}\mathcal{G}_c^{\mathrm{XRAY}}$ through our dual-source retrieval pipeline, seamlessly integrating high-quality in-house X-ray samples with synthetic images derived from web data. + +# 5.2. Class Descriptor Modeling + +The effectiveness of RAXO hinges on leveraging its visual gallery $\mathcal{G}$ to construct a robust visual descriptor $\chi_{c}$ that accurately captures the visual X-ray properties of class $c$ . A + +naive approach, such as averaging the feature representations of all $\mathcal{G}_c^{\mathrm{XRAY}}$ instances, fails to account for intra-class variability (e.g., distinct shapes of utility knives versus chef knives) and does not differentiate between foreground and background regions. To address these limitations, we propose a novel class descriptor modeling strategy shown in Fig. 2(2). + +First, we process each sample $\mathbf{I} \in \mathcal{G}_c^{XRAY}$ with $\Omega$ to separate the X-ray object from its background. Simultaneously, a feature extractor $\phi$ extracts per-patch embeddings $\phi(\mathbf{I}) \in \mathbb{R}^{H' \times W' \times D}$ , where $H' \times W'$ represents the number of spatial tokens, and $D$ is the embedding dimension. Following [10], we resize (denoted by $\zeta$ ) the segmentation mask to match the spatial dimensions of the feature map, and subsequently we use it to compute both a positive and a negative prototype. The positive prototype, designed to capture the significant visual features of the object without background interference, is computed as the average embedding over the foreground region as: + +$$ +\mathbf {x} _ {\mathbf {I}} ^ {\text {p o s}} = \frac {\sum_ {i , j} ^ {H ^ {\prime} , W ^ {\prime}} \zeta (\Omega (\mathbf {I} _ {i , j})) \odot \phi (\mathbf {I} _ {i , j})}{\sum \zeta (\Omega (\mathbf {I}))}, \tag {3} +$$ + +On the other hand, the negative prototype is obtained as the average embedding over the complementary region (background) as: + +$$ +\mathbf {x} _ {\mathbf {I}} ^ {\text {n e g}} = \frac {\sum_ {i , j} ^ {H ^ {\prime} , W ^ {\prime}} \left(1 - \zeta \left(\Omega \left(\mathbf {I} _ {\mathbf {i} , \mathbf {j}}\right)\right)\right) \odot \phi \left(\mathbf {I} _ {\mathbf {i} , \mathbf {j}}\right)}{\sum \left(1 - \zeta \left(\Omega (\mathbf {I})\right)\right)}. \tag {4} +$$ + +Subsequently, to construct the final visual descriptor $\mathcal{X}_c$ of class $c$ , we compute the average positive prototype $\overline{\mathbf{x}}_c^{\mathrm{pos}}$ from $\mathcal{G}_c^{XRAY}$ and unite it with all the individual positive prototypes as: + +$$ +\mathcal {X} _ {c} = \left[ \bar {\mathbf {x}} _ {c} ^ {\text {p o s}}, \left\{\mathbf {x} _ {\mathbf {I}} ^ {\text {p o s}} \mid \mathbf {I} \in \mathcal {G} _ {c} \right\} \right]. \tag {5} +$$ + +This formulation effectively captures both fine-grained object details and a holistic class-level representation, handling intra-class variability. Additionally, we also construct a global background descriptor by using the negative prototypes across the entire visual sample gallery as $\mathcal{X}_{bg} = [\overline{\mathbf{x}}^{\mathrm{neg}},\{\mathbf{x}_{\mathbf{I}}^{\mathrm{neg}}\mid \mathbf{I}\in \mathcal{G}\} ]$ . This negative prototype is used to further improve detection reliability by filtering out low-quality proposals (e.g., a region proposed on the background) during inference. A key advantage of this approach is its modularity: visual descriptors $\mathcal{X}_c$ are computed off-line and can be incrementally expanded with new object categories, seamlessly integrating with the OvOD paradigm without training or requiring X-ray detection data. + +# 5.3. Classification is All You Need + +As shown in Fig. 2(3), once the visual descriptors are constructed for each class in the user-specified vocabulary, along with the background class, the RAXO classifier + +$\mathcal{X} = \{\left[\mathcal{X}_c,\mathcal{X}_{bg}\right]\mid c\in C^{\mathrm{test}}\}$ can be directly applied to any OvOD detector to classify proposals $\mathbf{z_m}$ as: + +$$ +\hat {c} = \underset {c \in C ^ {\text {t e s t}} ^ {\prime}} {\arg \max } \underset {\mathbf {x} \in \mathcal {X}} {\max } \langle \mathbf {z} _ {m}, \mathbf {x} \rangle , \tag {6} +$$ + +where $C^{\mathrm{test}}$ denotes the test-time vocabulary $C^{\mathrm{test}}$ extended with the additional "background" class. If a proposal matches the background, it is removed. This simple-yet-effective strategy enables any pre-trained OvOD detectors to achieve strong performance on X-ray object detection with minimal modifications and no need for re-training. + +Descriptor consistency criterion. To further reduce incorrect proposals from the OvOD detector, RAXO introduces a novel Descriptor Consistency Criterion (DCC) to enforce class consistency in predictions. DCC evaluates how well a proposal aligns with its predicted class relative to others, suppressing weakly aligned proposals. For a proposal $\mathbf{z}_m$ , we measure its similarity to the closest prototype as $s_1 = \max_{\mathbf{x} \in \mathcal{X}_{\hat{c}}} \langle \mathbf{z}_m, \mathbf{x} \rangle$ . We also compute its mean similarity to the average prototypes of all other classes as $s_2 = \mathrm{avg}_{c \in C^{\mathrm{test}}, c \neq \hat{c}} \langle \mathbf{z}_m, \bar{\mathbf{x}}_c^{\mathrm{pos}} \rangle$ . The difference $\Delta = s_1 - s_2$ serves as a confidence measure, where a higher value indicates greater alignment with its predicted class than with any other. Proposals with $\Delta$ below a threshold $\sigma$ are discarded. + +# 6. Experiments + +Evaluation protocol. Conventional OvOD models are typically assessed using a Cross-Dataset Transfer Evaluation (CDTE) protocol, where the model is trained on one dataset and evaluated on a different dataset in a zero-shot manner [45]. In contrast, our setting involves a more challenging scenario where both the datasets and their underlying modalities differ. To address this, we introduce a novel Cross-Modality Transfer Evaluation (CMTE) protocol. Under CMTE, an OvOD model is trained on a source RGB dataset and subsequently evaluated on target $X$ -ray datasets without any additional training or fine-tuning. For performance assessment, we employ the standard MS COCO [17] metrics: AP, AP50, and AP75. + +Implementation details. Unless stated otherwise, RAXO's default configuration employs SAM 2 [27] as the object segmentation module $\Omega$ , ViT-B/14 [3] pretrained with DINOv2 [25] as the visual encoder $\phi$ , and GPT-4 [24] as the LLM. We set the consistency threshold to $\sigma = 0.15$ , the similarity threshold to $\tau = 0.5$ , and use $K = 30$ samples to construct the visual descriptors. $\mathcal{D}_{\mathrm{XRAY}}^{in - house}$ is composed of hold-out samples sourced from the evaluation X-ray datasets and is completely disjoint from the test split. $\mathcal{D}_{\mathrm{RGB}}^{web}$ is constructed by collecting images from the web using the public Google Custom Search API [8]. A detailed analysis of the impact of hyperparameter choices is provided in Sec. 6.2. Further implementation details can be found in Supp. D. + +Datasets and baselines. We evaluate RAXO by integrating it with four different state-of-the-art OvOD detectors that were exclusively trained on RGB images: GroundingDINO [18], Detic [44], VLDet [15], and CoDet [21]. Our CMTE evaluation is performed across six diverse X-ray datasets: PIXray [19], PIDray [34], CLCXray [42], DvXray [20], HiXray [30], and our proposed DET-COMPASS, which together comprise 343 visible unique classes and over 140k images. A detailed description of these datasets is provided in Tab. 1. + +# 6.1. Open-Vocabulary Detection Results + +RAXO equips OvOD models with X-ray vision. We evaluate the X-ray adaptation capabilities of RAXO on the aforementioned datasets and baselines in Tab. 2. Since RAXO is training-free, we report the results directly on the test splits, accounting for all available categories. Furthermore, we analyze the influence of the gallery composition by varying $\mathcal{G}$ from a $0/100$ configuration (i.e., all samples in $\mathcal{G}$ are retrieved from the web, as $C^{\mathrm{test}} \cap C^{\mathrm{in-house}} = \emptyset$ ) to a $100/0$ configuration (i.e., all samples in $\mathcal{G}$ are retrieved from the in-house database, as $C^{\mathrm{test}} \subset C^{\mathrm{in-house}}$ ). Each experiment is repeated three times with different random distributions of in-domain and web categories, and we report the averaged results. + +As shown in Tab. 2, off-the-shelf OvOD detectors trained on RGB images perform poorly on X-ray data. In contrast, even without access to any in-domain data from $C^{\mathrm{test}}$ (0/100 setting), RAXO enhances baseline OvOD methods, yielding an average improvement of $\uparrow 2.1$ points. This gain is further amplified when in-domain samples are incorporated into $\mathcal{G}$ , which facilitates the construction of more robust materials in $\mathcal{M}$ and yields higher-quality visual descriptors. Notably, even a limited number of in-domain samples (20/80 setting) results in an average improvement of $\uparrow 4.8$ points. Moreover, increasing the in-domain samples (50/50 setting) yields even more significant gains, particularly in datasets with a large number of classes, such as DET-COMPASS, where RAXO offers an average boost of $\uparrow 14.9$ points. Finally, when $\mathcal{G}$ consists entirely of in-domain samples, RAXO delivers an average improvement of $\uparrow 14.7$ points across all OvOD methods and datasets, all in a training-free manner. + +RAXO works with any OvOD detector. The consistent gains of RAXO across all evaluated detectors in Tab. 2, along with its simple integration (Sec. 5.3), confirm the generalization of our approach to any RGB OvOD model. Moreover, RAXO scales with the baseline method, yielding larger gains when integrated with more robust detectors (e.g., RAXO enhances G-DINO $\uparrow$ 9.4 points on average across all $G$ settings, compared to $\uparrow$ 7.5 points for Detic). This highlights the advantage of our training-free method, which can leverage the rapid advancements in RGB OvOD detectors without needing large amounts of labeled X-ray data. + +
GMethodD-COMPASSPIXrayPIDrayCLCXrayDvXrayHiXrayAvg.
G-DINO [18]13.412.910.96.710.07.010.2
100/80/2050/5020/800/100+ RAXO47.9↑34.536.9↑24.016.5↑5.622.2↑15.522.6↑12.617.1↑10.127.2↑17.0
41.0↑27.633.8↑20.915.4↑4.518.0↑11.321.0↑11.014.5↑7.524.0↑13.8
31.4↑18.025.4↑12.515.5↑4.617.0↑10.316.1↑6.113.4↑6.419.8↑9.6
20.5↑7.121.6↑8.713.9↑3.010.0↑3.315.0↑5.09.8↑2.815.1↑4.9
14.0↑0.616.1↑3.213.4↑2.57.1↑0.412.4↑2.47.9↑0.911.8↑1.6
VLDet [15]10.69.86.94.47.45.17.4
100/80/2050/5020/800/100+ RAXO36.4↑25.832.3↑22.511.7↑4.815.4↑11.020.1↑12.714.8↑9.721.8↑14.4
31.8↑21.229.2↑19.411.0↑4.112.7↑8.316.8↑9.413.1↑8.019.1↑11.7
23.7↑13.124.0↑14.210.4↑3.511.1↑6.712.1↑4.711.2↑6.115.4↑8.0
16.2↑5.621.6↑11.89.4↑2.55.2↑0.810.6↑3.29.3↑4.212.1↑4.7
11.1↑0.514.1↑4.38.9↑2.04.4↑0.09.0↑1.68.3↑3.29.3↑1.9
Detic [44]11.59.37.14.77.04.87.4
100/80/2050/5020/800/100+ RAXO35.3↑23.827.3↑18.011.3↑4.214.0↑9.319.4↑12.414.2↑9.420.3↑12.9
30.7↑19.223.9↑14.610.8↑3.712.3↑7.618.0↑11.012.1↑7.318.0↑10.6
24.4↑12.919.5↑10.210.3↑3.29.2↑4.514.6↑7.611.0↑6.214.8↑7.4
16.4↑4.915.2↑5.99.6↑2.58.0↑3.312.7↑5.79.9↑5.112.0↑4.6
11.9↑0.413.4↑4.19.1↑2.05.2↑0.59.4↑2.47.9↑3.19.5↑2.1
CoDet [21]8.47.35.73.15.63.45.6
100/80/2050/5020/800/100+ RAXO35.8↑27.427.9↑20.610.3↑4.614.8↑11.717.6↑12.013.2↑9.819.9↑14.3
32.2↑23.825.1↑17.89.5↑3.812.0↑8.915.4↑9.811.7↑8.317.7↑12.1
24.0↑15.620.0↑12.79.5↑3.89.2↑6.111.5↑5.99.9↑6.514.0↑8.4
17.8↑9.414.8↑7.58.5↑2.85.1↑2.09.4↑3.88.1↑4.710.6↑5.0
12.2↑3.811.5↑4.28.1↑2.44.0↑0.96.9↑1.36.5↑3.18.2↑2.6
+ +Table 2. X-ray OvOD performance under the Cross-Modality Transfer Evaluation (CMTE) setting on DET-COMPASS (ours), PINray [19], PIDray [34], CLCXray [42], DvXray [20], and HiXray [30] datasets. We integrate RAXO into different baselines using different gallery $\mathcal{G}$ compositions, from using only ${\mathcal{D}}_{\mathrm{{XRAY}}}^{\text{in-house }}$ data $\left( {{100}/0}\right)$ to exclusively ${\mathcal{D}}_{\mathrm{{RGB}}}^{\text{web }}$ samples (0/100). RAXO consistently improves the performance of all baseline OvOD detectors across every dataset. AP is used. + +
PIXray
ScissorsWrenchBatteryPliersAP50
OVXD†16.946.214.66.421.0
BARON + RAXO18.239.48.322.022.0
CoDet + RAXO43.652.69.018.630.9
Detic + RAXO46.360.57.936.637.8
VLDet + RAXO48.362.710.638.540.0
G-DINO + RAXO49.761.611.351.543.5
+ +RAXO surpasses training-based approaches. We also compare RAXO with OVXD [16], the only concurrent work that extends X-ray detection beyond base categories. OVXD is a fully-supervised method trained directly on X-ray data, and its generalization capabilities are evaluated on a hold-out set comprising only four categories (i.e., Scissors, Wrench, Battery, and Pliers). In Tab. 3, we follow this setting and evaluate RAXO on PINray [19], excluding the aforementioned classes from the in-house dataset $\mathcal{D}_{\mathrm{XRAY}}^{\mathrm{in - house}}$ . Our results demonstrate that RAXO outperforms OVXD even when both approaches utilize the same BARON detector [37]. Notably, unlike OVXD, our method does not retrain BARON or any other component. Yet, our training-free RAXO performs remarkably well, achieving an AP50 + +Table 3. Comparison under OVXD [16] setting, where the methods do not have access to in-domain data from the displayed categories. $\dagger$ OVXD is a supervised method explicitly trained on X-ray images for alignment. AP50 is used. + +
Mod. Filt. Trans.APAP50AP75
G-DINO [18]12.914.913.4
+ RAXOX-ray RGBS7.9↓5.09.8↓5.18.3↓5.1
13.9↑1.017.4↑2.514.3↑0.9
RGB✓14.3↑1.417.6↑2.714.7↑1.3
RGB✓18.3↑3.415.1↑1.7
RGB✓M16.1↑3.219.8↑4.9
+ +Table 4. Ablation of web-based retrieval on PIXray [19] using only web-based samples (setting $0/100$ ). Mod. indicates retrieval of a specific modality from the web, Filt. refers to filtering retrieved images with $\mathcal{F}_{\mathrm{RGB}}$ , and Trans. specifies style-transfer via: $S$ a diffusion-based method [7] or $M$ our material-transfer mechanism. + +of 43.5 when paired with G-DINO [18]. + +# 6.2. Ablation Study + +In this section we study the core components of RAXO. We conduct our experiments on PixRay [19], using GroundingDINO [13] as our baseline OvOD. Consistent findings are reported in Supp. E and H. + +Web-based samples. Tab. 4 presents an ablation study on the components of our web-based visual sample acquisition pipeline (Sec. 5.1). Notably, directly retrieving X-ray samples from the web $(\mathrm{Mod.} = \mathrm{X}$ -ray) leads to a $\downarrow 5.0$ point decrease in AP. In contrast, retrieving RGB images from the web $(\mathrm{Mod.} = \mathrm{RGB})$ provides a modest AP improvement of + +
Ω x neg X DCCAPAP50AP75
G-DINO [18]12.914.913.4
+ RAXO25.2↑12.331.4↑16.526.2↑12.8
✓27.1↑14.233.6↑18.728.4↑15.0
✓ ✓27.8↑14.934.5↑19.629.3↑15.9
✓27.5↑14.634.0↑19.128.8↑15.4
✓ ✓ ✓28.5↑15.635.3↑20.429.9↑16.5
✓ ✓ ✓36.9↑24.045.0↑30.139.0↑25.6
+ +Table 5. Ablation of class representation construction on PIXray [19] using in-domain samples (setting $100/0$ ). $\Omega$ indicates building positive prototypes using segmentation masks. $\mathbf{x}^{\mathrm{neg}}$ indicates use of negative prototypes. $\mathcal{X}$ indicates ensemble multiple prototypes in class descriptors. $DCC$ indicates use of prototype consistency criterion. + +$\uparrow 1.0$ point. We hypothesize that directly using X-ray examples obtained from the web is ineffective due to the limited availability of high-quality X-ray images online. Filtering the retrieved RGB images using $\mathcal{F}_{\mathrm{RGB}}$ increases AP by an additional $\uparrow 0.4$ points. Nonetheless, they still exhibit a significant visual gap compared to their X-ray counterparts. A style-transfer approach using StyleShot [7] (Trans. $= S$ ) provides a marginal AP gain of $\uparrow 0.3$ points, likely due to its limited capacity to understand the intrinsic material properties of the objects. Conversely, our proposed material-transfer mechanism (Trans. $= M$ ) yields a substantial AP improvement of $\uparrow 1.8$ points, underscoring its effectiveness in bridging the X-ray modality gap. + +Class descriptor modeling. Tab. 5 presents an ablation study analyzing the impact of the various components used to construct class representations. As shown, using a simple per-class average of all the features from each sample already results in a significant AP improvement of $\uparrow 12.3$ points over the baseline OvOD. Refining the representation by applying an object segmentation method to isolate the foreground features $(\Omega)$ further boosts the AP by $\uparrow 1.9$ points. Moreover, leveraging background features to construct negative prototypes $(\mathbf{x}^{\mathrm{neg}})$ adds an additional $\uparrow 0.7$ points. + +Replacing the averaged class prototypes with our proposed visual descriptor $(\mathcal{X})$ yields another $\uparrow 0.7$ point improvement, demonstrating its enhanced capability to capture intra-class visual variability. Finally, incorporating our descriptor consistency criterion to suppress low-quality proposals $(DCC)$ provides a further AP increase of $\uparrow 8.4$ points, underscoring the effectiveness of our proposed framework. + +Hyperparameter study. Fig. 4 analyzes the effect of varying the number of samples $K$ used to construct the visual descriptors. Even with just a single sample per class, RAXO improves AP by $\uparrow 6.5$ points over the baseline. As $K$ increases, performance continues to improve, reaching saturation at $K = 30$ with a total AP gain of $\uparrow 24.0$ points. + +Furthermore, Fig. 5 explores the influence of the threshold $\sigma$ in the Descriptor Consistency Module. Higher values of $\sigma$ impose stricter consistency requirements, de + +![](images/4b74a760a4e378c70dc0de82ea55681b64fd98e75c5e79308d65d0aadc3210e0.jpg) +Figure 4. Impact of $K$ in class representations evaluated on PINray [19] using a G-DINO [18] baseline in the 100/0 setting. + +![](images/4b282f175fcce032c9c8714b863deafad0e5af70cb1264185bc000316d26dbe8.jpg) +Figure 5. Impact of $\sigma$ in Descriptor Consistency Module on PIxray [19] using a G-DINO [18] baseline in the $100/0$ setting. + +manding greater separation between the predicted class and other classes before accepting a proposal. We observe that $\sigma = 0.15$ provides an optimal balance between false positives and false negatives, yielding an AP of 36.9 points. + +# 7. Conclusions + +In this work, we pioneered the task of open-vocabulary object detection (OvOD) for X-ray imagery, a challenge shaped by the unique characteristics of X-ray data and the scarcity of annotated examples. To address this, we introduced RAXO, a training-free method that repurposes off-the-shelf RGB-based OvOD detectors for the X-ray domain. By leveraging intra-modal feature distances, a novel material-transfer mechanism, and robust class descriptor modeling, RAXO effectively bridges the modality gap between RGB and X-ray imagery. Our extensive experiments, conducted across multiple benchmarks and the newly proposed DET-COMPASS dataset, demonstrate that RAXO consistently enhances detection performance, achieving average mAP improvements of up to 17.0 points over baseline methods. + +Looking ahead, the modularity of RAXO opens promising avenues for future research, including further refinement of visual descriptors and adaptation to additional modalities. Overall, our work not only sets a new state-of-the-art for X-ray open-vocabulary detection but also lays the groundwork for a thriving research direction in this emerging domain. + +# Acknowledgements + +# References + +[1] Samet Akcay and Toby Breckon. Towards automatic threat detection: A survey of advances of deep learning within x-ray security imaging. Pattern Recognition, 122:108245, 2022. 2, 3 +[2] An Chang, Yu Zhang, Shunli Zhang, Leisheng Zhong, and Li Zhang. Detecting prohibited objects with physical size constraint from cluttered x-ray baggage images. Knowledge-Based Systems, 2022. 1 +[3] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 6 +[4] Penghui Du, Yu Wang, Yifan Sun, Luting Wang, Yue Liao, Gang Zhang, Errui Ding, Yan Wang, Jingdong Wang, and Si Liu. Lami-detr: Open-vocabulary detection with language model instruction. In ECCV, 2024. 4, 5 +[5] Ruohuan Fang, Guansong Pang, and Xiao Bai. Simple image-level classification improves open-vocabulary object detection. In AAAI, 2024. 2 +[6] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated images. In ECCV, 2022. 2 +[7] Junyao Gao, Yanchen Liu, Yanan Sun, Yinhao Tang, Yanhong Zeng, Kai Chen, and Cairong Zhao. Styleshot: A snapshot on any style. arXiv:2407.01414, 2024. 4, 7, 8, 5 +[8] Google LLC. Custom search JSON API reference. https: // developers.google.com/custom-search/v1, 2024. Accessed: 2025-03-07. 6, 3 +[9] Lewis D. Griffin, Matthew Caldwell, and Jerone T. A. Andrews. COMPASS-XP. Zenodo, 2019. 3, 1 +[10] Laurynas Karazija, Iro Laina, Andrea Vedaldi, and Christian Rupprecht. Diffusion models for open-vocabulary segmentation. In ECCV, 2024. 5 + +We thank CINECA and the ISCRA initiative for the availability of high-performance computing resources. This work was partially supported by the EU HORIZON IAMI (HORIZON-CL3-2023-FCT-01-04-101168272) project, the EU HORIZON ELIAS (HORIZON-CL4-2022-HUMAN-02-101120237) project, the EU HORIZON ELLIOT (HORIZON-CL4-2024-HUMAN-03-101214398) project, the MUR PNRR FAIR (PE00000013) project funded by the NextGenerationEU, the Spanish Ministerio de Ciencia e Innovacion (grant numbers PID2020-112623GB-I00, PID2023-149549NB-I00), and the Galician Conselleria de Cultura, Educacion e Ordenacion Universitaria (2024-2027 ED431G-2023/04). Some of these grants are co-funded by the European Regional Development Fund (ERDF). Pablo Garcia-Fernandez is supported by the Spanish Ministerio de Universidades under the FPU national plan (grant number FPU21/05581). +[11] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models. arXiv:2209.15639, 2022. 2 +[12] Liangqi Li, Jiaxu Miao, Dahu Shi, Wenming Tan, Ye Ren, Yi Yang, and Shiliang Pu. Distilling detr with visual-linguistic knowledge for open-vocabulary object detection. In ICCV, 2023. 2 +[13] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, 2022. 2, 7 +[14] Mingyuan Li, Tong Jia, Hao Wang, Bowen Ma, Hui Lu, Shuyang Lin, Da Cai, and Dongyue Chen. Ao-detr: Anti-overlapping detr for x-ray prohibited items detection. IEEE Transactions on Neural Networks and Learning Systems, 2024. 1 +[15] Chuang Lin, Peize Sun, Yi Jiang, Ping Luo, Lizhen Qu, Gholamreza Haffari, Zehuan Yuan, and Jianfei Cai. Learning object-language alignments for open-vocabulary object detection. In ICLR, 2023. 1, 2, 6, 7, 5 +[16] Shuyang Lin, Tong Jia, Hao Wang, Bowen Ma, Mingyuan Li, and Dongyue Chen. Detection of novel prohibited item categories for real-world security inspection. Eng. Appl. Artif. Intell., 2025. 1, 2, 7 +[17] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 6 +[18] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding DINO: marrying DINO with grounded pre-training for open-set object detection. In ECCV, 2024. 1, 2, 6, 7, 8, 4, 5 +[19] Bowen Ma, Tong Jia, Min Su, Xiaodong Jia, Dongyue Chen, and Yichun Zhang. Automated segmentation of prohibited items in x-ray baggage images using dense de-overlap attention snake. TMM, 2022. 2, 3, 6, 7, 8, 4, 5 +[20] Bowen Ma, Tong Jia, Mingyuan Li, Songsheng Wu, Hao Wang, and Dongyue Chen. Towards dual-view x-ray baggage inspection: A large-scale benchmark and adaptive hierarchical cross refinement for prohibited item discovery. IEEE TIFS, 2024. 1, 3, 6, 7, 5 +[21] Chuofan Ma, Yi Jiang, Xin Wen, Zehuan Yuan, and Xiaojuan Qi. CoDet: Co-occurrence guided region-word alignment for open-vocabulary object detection. In NeurIPS, 2023. 2, 6, 7, 5 +[22] Caijing Miao, Lingxi Xie, Fang Wan, Chi Su, Hongye Liu, Jianbin Jiao, and Qixiang Ye. Sixray: A large-scale security inspection x-ray benchmark for prohibited item discovery in overlapping images. In CVPR, 2019. 2 +[23] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, et al. Simple open-vocabulary object detection. In ECCV, 2022. 2 +[24] OpenAI. GPT-4 technical report. CoRR, abs/2303.08774, 2023.6 + +[25] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv:2304.07193, 2023. 6 +[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3 +[27] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dólar, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv:2408.00714, 2024. 6 +[28] Haifeng Sima, Bailiang Chen, Chaosheng Tang, Yudong Zhang, and Junding Sun. Multi-scale feature attention-detection transformer: Multi-scale feature attention for security check object detection. IET Computer Vision, 2024. 2 +[29] Archana Singh and Dhiraj. Advancements in machine learning techniques for threat item detection in x-ray images: a comprehensive survey. Int. J. Multim. Inf. Retr., 2024. 1 +[30] Renshuai Tao, Yanlu Wei, Xiangjian Jiang, Hainan Li, Haotong Qin, Jiakai Wang, Yuqing Ma, Libo Zhang, and Xi-anglong Liu. Towards real-world x-ray security inspection: A high-quality benchmark and lateral inhibition module for prohibited items detection. In ICCV, 2021. 2, 3, 6, 7, 5 +[31] Renshuai Tao, Hainan Li, Tianbo Wang, Yanlu Wei, Yifu Ding, Bowei Jin, Hongping Zhi, Xianglong Liu, and Aishan Liu. Exploring endogenous shift for cross-domain detection: A large-scale benchmark and perturbation suppression network. In CVPR, 2022. 2, 3 +[32] Renshuai Tao, Tianbo Wang, Ziyang Wu, Cong Liu, Aishan Liu, and Xianglong Liu. Few-shot x-ray prohibited item detection: A benchmark and weak-feature enhancement network. In ACMMM, 2022. 1, 3 +[33] Divya Velayudhan, Taimur Hassan, Ernesto Damiani, and Naoufel Werghi. Recent advances in baggage threat detection: A comprehensive and systematic survey. ACM Computing Surveys, 55(8):1-38, 2022. 2, 3 +[34] Boying Wang, Libo Zhang, Longyin Wen, Xianglong Liu, and Yanjun Wu. Towards real-world prohibited item detection: A large-scale x-ray benchmark. In ICCV, 2021. 2, 3, 6, 7, 5 +[35] Ruxue Wang, Yuliang Shi, and Mingyu Cai. Optimization and research of suspicious object detection algorithm in x-ray image. In 2023 IEEE 6th Information Technology, Networking, Electronic and Automation Control Conference (ITNEC), 2023. 2 +[36] Yanlu Wei, Renshuai Tao, Zhangjie Wu, Yuqing Ma, Libo Zhang, and Xianglong Liu. Occluded prohibited items detection: An x-ray security inspection benchmark and de-occlusion attention module. In ACMMM, 2020. 2 +[37] Size Wu, Wenwei Zhang, Sheng Jin, Wentao Liu, and + +Chen Change Loy. Aligning bag of regions for open-vocabulary object detection. In CVPR, 2023. 2, 7 +[38] Size Wu, Wenwei Zhang, Lumin Xu, Sheng Jin, Xiangtai Li, Wentao Liu, and Chen Change Loy. Clipself: Vision transformer distills itself for open-vocabulary dense prediction. In ICLR, 2024. 2 +[39] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In CVPR, 2023. 2 +[40] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. Detclip: Dictionary-enriched visual-concept paralleled pretraining for open-world detection. In NeurIPS, 2022. 2 +[41] Lewei Yao, Jianhua Han, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, and Hang Xu. Detclipv2: Scalable open-vocabulary object detection pre-training via word-region alignment. In CVPR, 2023. 2 +[42] Cairong Zhao, Liang Zhu, Shuguang Dou, Weihong Deng, and Liang Wang. Detecting overlapped objects in x-ray security imagery by a label-aware mechanism. IEEE TIFS, 2022, 2, 3, 6, 7, 5 +[43] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In CVPR, 2022. 2 +[44] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, 2022. 1, 2, 3, 6, 7, 5 +[45] Chaoyang Zhu and Long Chen. A survey on open-vocabulary detection and segmentation: Past, present, and future. IEEE TPAMI, 2024. 3, 6 \ No newline at end of file diff --git a/2025/Superpowering Open-Vocabulary Object Detectors for X-ray Vision/images.zip b/2025/Superpowering Open-Vocabulary Object Detectors for X-ray Vision/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..815af8ae7f7d504cec3e8c7265d9c74bfb753faf --- /dev/null +++ b/2025/Superpowering Open-Vocabulary Object Detectors for X-ray Vision/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:491182f70a4efe16831cf32e6d511a300b2a24df59ddb7a771ce3f94c09442f8 +size 646727 diff --git a/2025/Superpowering Open-Vocabulary Object Detectors for X-ray Vision/layout.json b/2025/Superpowering Open-Vocabulary Object Detectors for X-ray Vision/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..58909907130a31012fcdf5e6c2992ed373eca730 --- /dev/null +++ b/2025/Superpowering Open-Vocabulary Object Detectors for X-ray Vision/layout.json @@ -0,0 +1,9770 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 96, + 103, + 514, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 103, + 514, + 121 + ], + "spans": [ + { + "bbox": [ + 96, + 103, + 514, + 121 + ], + "type": "text", + "content": "Superpowering Open-Vocabulary Object Detectors for X-ray Vision" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 127, + 503, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 127, + 503, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 503, + 156 + ], + "type": "text", + "content": "Pablo Garcia-Fernandez" + }, + { + "bbox": [ + 106, + 127, + 503, + 156 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 106, + 127, + 503, + 156 + ], + "type": "text", + "content": " Lorenzo Vaquero" + }, + { + "bbox": [ + 106, + 127, + 503, + 156 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 106, + 127, + 503, + 156 + ], + "type": "text", + "content": " Mingxuan Liu" + }, + { + "bbox": [ + 106, + 127, + 503, + 156 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 106, + 127, + 503, + 156 + ], + "type": "text", + "content": " Feng Xue" + }, + { + "bbox": [ + 106, + 127, + 503, + 156 + ], + "type": "inline_equation", + "content": "^{3*}" + }, + { + "bbox": [ + 106, + 127, + 503, + 156 + ], + "type": "text", + "content": " Daniel Cores" + }, + { + "bbox": [ + 106, + 127, + 503, + 156 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 106, + 127, + 503, + 156 + ], + "type": "text", + "content": " Nicu Sebe" + }, + { + "bbox": [ + 106, + 127, + 503, + 156 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 106, + 127, + 503, + 156 + ], + "type": "text", + "content": " Manuel Mucientes" + }, + { + "bbox": [ + 106, + 127, + 503, + 156 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 106, + 127, + 503, + 156 + ], + "type": "text", + "content": " Elisa Ricci" + }, + { + "bbox": [ + 106, + 127, + 503, + 156 + ], + "type": "inline_equation", + "content": "^{2,3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 160, + 503, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 160, + 503, + 190 + ], + "spans": [ + { + "bbox": [ + 106, + 160, + 503, + 190 + ], + "type": "text", + "content": "1University of Santiago de Compostela, Spain 2Fondazione Bruno Kessler, Italy 3University of Trento, Italy" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 79, + 191, + 531, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 191, + 531, + 202 + ], + "spans": [ + { + "bbox": [ + 79, + 191, + 531, + 202 + ], + "type": "text", + "content": "{pablogarcia.fernandez, daniel.cores, manuel.mucientes}@usc.es lvaquerootal@fbk.eu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 202, + 503, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 202, + 503, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 202, + 503, + 213 + ], + "type": "text", + "content": "{mingxuan.liu, feng.xue, niculae.sebe, e.ricci}@unitn.it *Corresponding authors" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 152, + 231, + 199, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 231, + 199, + 243 + ], + "spans": [ + { + "bbox": [ + 152, + 231, + 199, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 255, + 296, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 255, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 54, + 255, + 296, + 531 + ], + "type": "text", + "content": "Open-vocabulary object detection (OvOD) is set to revolutionize security screening by enabling systems to recognize any item in X-ray scans. However, developing effective OvOD models for X-ray imaging presents unique challenges due to data scarcity and the modality gap that prevents direct adoption of RGB-based solutions. To overcome these limitations, we propose RAXO, a training-free framework that repurposes off-the-shelf RGB OvOD detectors for robust X-ray detection. RAXO builds high-quality X-ray class descriptors using a dual-source retrieval strategy. It gathers relevant RGB images from the web and enriches them via a novel X-ray material transfer mechanism, eliminating the need for labeled databases. These visual descriptors replace text-based classification in OvOD, leveraging intra-modal feature distances for robust detection. Extensive experiments demonstrate that RAXO consistently improves OvOD performance, providing an average mAP increase of up to 17.0 points over base detectors. To further support research in this emerging field, we also introduce DET-COMPASS, a new benchmark featuring bounding box annotations for over 300 object categories, enabling large-scale evaluation of OvOD in X-ray. Code and dataset available at: https://pagf188.github.io/RAXO/." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 549, + 135, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 549, + 135, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 549, + 135, + 562 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 570, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 641 + ], + "type": "text", + "content": "Automated object detection technologies for X-ray imaging are essential to maintain public safety, enabling the identification of prohibited items at checkpoints in high-risk environments such as airports, train stations, museums and stadiums [29]. These systems improve security while simultaneously reducing the workload of human inspectors." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Conventional X-ray object detectors rely on supervised learning [2, 14, 20] and are inherently limited by the object categories present in their training datasets (see Fig. 1a). This limitation is exacerbated by the high cost of X-ray machinery and the requirement for expert annotation, often restricting these systems to fewer than 20 object classes [32]," + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 326, + 228, + 541, + 432 + ], + "blocks": [ + { + "bbox": [ + 326, + 228, + 541, + 432 + ], + "lines": [ + { + "bbox": [ + 326, + 228, + 541, + 432 + ], + "spans": [ + { + "bbox": [ + 326, + 228, + 541, + 432 + ], + "type": "image", + "image_path": "25cf5ef969acd77759db2ca01adf2e5caba7178cbccd2f42808b19029b75cc47.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 433, + 555, + 499 + ], + "lines": [ + { + "bbox": [ + 313, + 433, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 555, + 499 + ], + "type": "text", + "content": "Figure 1. (a) Traditional X-ray object detectors are constrained by the limited categories in their training datasets. (b) We introduce the task of open-vocabulary object detection (OvOD) for X-ray imaging and propose RAXO, a training-free method that adapts off-the-shelf RGB OvOD models to X-ray data. (c) RAXO greatly improves detection performance across multiple benchmarks." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 510, + 517, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 517, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 517, + 521 + ], + "type": "text", + "content": "thereby hindering broader real-world applications." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 521, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 629 + ], + "type": "text", + "content": "In light of the expanding diversity of man-made objects and evolving security demands, an open-vocabulary object detection (OvOD) framework capable of recognizing arbitrary X-ray object categories defined by the user is imperative. Despite its importance, open-vocabulary detection in the X-ray domain remains largely unexplored. Concurrent works aiming to extend beyond base categories have managed to generalize to at most four unseen classes [16], thereby highlighting the inherent challenges of the task." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "content": "Recent advances in OvOD for conventional RGB images have been driven by large-scale annotated datasets, which facilitate effective alignment between visual and textual features [15, 18, 44]. However, these advancements do not directly transfer to X-ray imagery. In practice, applying state-of-the-art RGB-based OvOD models to X-ray scans leads to significant performance degradation, as illustrated" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20770" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "in Fig. 1c. Moreover, retraining these models on X-ray data is often impractical given the scarcity of large-scale annotated X-ray datasets." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 108, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 294, + 179 + ], + "type": "text", + "content": "Motivated by these challenges, this work opens up open-vocabulary object detection for X-ray imaging by repurposing robust, off-the-shelf RGB-based detectors. To this end, we propose RAXO (tRAining-free adaptation for X-ray Open-vocabulary detection), a training-free method that seamlessly adapts RGB OvOD models to X-ray. (Fig. 1b)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 180, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 180, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 180, + 295, + 407 + ], + "type": "text", + "content": "We find that the main reason RGB-based OvOD detectors fail in the X-ray domain is the disruption of text-visual feature alignment. The appearance disparity between the same object in RGB and X-ray modalities causes textual embeddings, aligned with RGB features, to mismatch with X-ray visual features. RAXO addresses this issue on three simple steps: (1) visual sample acquisition, which involves obtaining X-ray images that represent user-defined categories; (2) class descriptor modeling, which uses these images to construct descriptors that effectively encode the visual appearance of each category within the X-ray domain; and (3) classifier construction, where the computed descriptors are used to build a visual classifier that replaces the inter-modal (i.e., text-to-visual) classification of conventional OvODs. This new classifier exploits the fact that, despite the modality shift, intra-modal (e.g., visual-to-visual) feature distances remain reliable indicators for object identification. In this way RAXO enables any OvOD to successfully detect X-ray objects based on their inherent visual properties." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 407, + 295, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 295, + 502 + ], + "type": "text", + "content": "To facilitate comprehensive evaluation, we also introduce DET-COMPASS, a detection dataset comprising 370 distinct object classes with paired X-ray and RGB annotations. Extensive experiments demonstrate that our approach consistently improves detection performance by an average of " + }, + { + "bbox": [ + 55, + 407, + 295, + 502 + ], + "type": "inline_equation", + "content": "\\uparrow 8.4" + }, + { + "bbox": [ + 55, + 407, + 295, + 502 + ], + "type": "text", + "content": " AP across multiple benchmarks and OvOD models, with gains that scale as the underlying detectors become more powerful (Tab. 2). In summary, our contributions are:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 503, + 295, + 646 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 503, + 295, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 295, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 295, + 538 + ], + "type": "text", + "content": "- We formally introduce the problem of open-vocabulary object detection for X-ray imagery without training, addressing a critical need for security in real-world scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 539, + 295, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 539, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 539, + 295, + 586 + ], + "type": "text", + "content": "- We introduce DET-COMPASS, a novel benchmark with bounding box annotations across 370 object categories, enabling standardized evaluation of OvOD methods in the X-ray modality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 586, + 295, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 586, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 586, + 295, + 646 + ], + "type": "text", + "content": "- We propose RAXO, a novel training-free approach that repurposes off-the-shelf RGB-based OvOD models for X-ray detection by constructing robust visual descriptors. RAXO achieves new state-of-the-art results across multiple benchmarks." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 657, + 142, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 142, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 142, + 670 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Open-vocabulary object detection has advanced rapidly with the advent of Vision-Language Models (VLMs), enabling detectors to generalize beyond fixed categories." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "OvOD detectors leverage weak supervision signals to improve detection accuracy. Based on the type of weak supervision utilized, OvOD methods can be categorized into (i) region-aware training, (ii) pseudo-labeling, (iii) knowledge distillation, and (iv) transfer learning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 134, + 553, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 134, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 134, + 553, + 337 + ], + "type": "text", + "content": "Region-aware training methods aim to improve localization and feature representation by refining the alignment between image regions and their corresponding textual descriptions. Approaches such as DetCLIP [40], DetCLIPv2 [41], CORA [39], and VLDet [15] adopt this strategy. Pseudo-labeling methods, rely on large pretrained VLMs to generate pseudo-labels, effectively expanding the training set. Methods like RegionCLIP [43], PromptDet [6], CoDET [21], GLIP [13], Detric [44], and Grounding DINO [18] follow this approach. Knowledge distillation techniques, such as BARON [37], DK-DETR [12], CLIPSelf [38], and SIC-CADS [5], employ VLMs as teachers in a teacher-student framework, transferring knowledge from the VLM image encoder to enhance the detector backbone. In contrast, transfer learning approaches integrate the VLM encoder directly, either through fine-tuning, as seen in OWL-ViT [23], or by freezing the encoder, as in F-VLM [11]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 339, + 553, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 553, + 555 + ], + "type": "text", + "content": "Regardless of the training strategy, OvOD methods heavily depend on VLMs trained on large-scale datasets. However, such datasets are unavailable in the X-ray modality, making this approach impractical. To address this, RAXO seamlessly adapts existing RGB-based OvOD methods for X-ray object detection without requiring additional training. X-ray object detection. Several datasets have been developed to tackle the challenge of object detection in X-ray imagery [1, 33]. SIXray [22], OPIXray [36], and CLCXray [42] focused on detecting occluded objects, introducing strategies to identify deliberately hidden prohibited object. To enhance dataset scale and diversity, PIDray [34] and HiXray [30] provided larger benchmarks, significantly increasing the number of annotated images. Meanwhile, PIXray [19] pioneered the introduction of an X-ray segmentation dataset and proposed a real-time framework for segmenting prohibited objects, further advancing automated threat detection in security screening applications." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "text", + "content": "Previous X-ray object detectors [28, 31, 35] considered mainly a closed-set paradigm, thus being limited to localize objects of a small set of predefined categories. More recently, Lin et al. [16] proposed fine-tuning a CLIP adapter to bridge the modality gap between CLIP's training data and X-ray images, thereby achieving OvOD in X-ray data. They demonstrated generalization on four novel object classes, but both training and evaluation were limited by the lack of largescale, well-annotated X-ray data. To address these issues, RAXO leverages data retrieval from the web to expand detection capabilities to a broader range of common objects. Furthermore, our proposed re-labeled DET-COMPASS dataset enables evaluation across a wider variety of classes." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20771" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 294, + 165 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 294, + 165 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 294, + 165 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 294, + 165 + ], + "type": "table", + "html": "
VenueImagesClassesModality
DvXray [20]TIFS’2432,00015X-ray
PIXray [19]TMM’225,04615X-ray
CLCXray [42]TIFS’229,56512X-ray
FSOD [32]ACMMM’2212,33320X-ray
EDS [31]CVPR’2214,21910X-ray
PIDray [34]ICCV’2147,67712X-ray
HiXray [30]ICCV’2145,3658X-ray
DET-COMPASS (Ours)-1,928370X-ray+RGB
", + "image_path": "9dd510c4920a4e50714e1ea457008cff6abbc7e110eae7593002eac5edc629f4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 167, + 295, + 200 + ], + "lines": [ + { + "bbox": [ + 55, + 167, + 295, + 200 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 295, + 200 + ], + "type": "text", + "content": "Table 1. Existing X-ray detection datasets. DET-COMPASS is the dataset with the highest number of categories and the only one providing pixel-level alignment between X-ray and RGB data." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 213, + 156, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 213, + 156, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 156, + 224 + ], + "type": "text", + "content": "3. DET-COMPASS" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 233, + 295, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 233, + 295, + 353 + ], + "spans": [ + { + "bbox": [ + 54, + 233, + 295, + 353 + ], + "type": "text", + "content": "Object detection in security X-ray scans has advanced significantly in recent years. However, evaluating OvOD detectors in this modality remains challenging due to the limited number of annotated object categories in existing X-ray benchmarks. For instance, the largest X-ray detection dataset, FSOD [32], includes annotations for only 20 classes (see Tab. 1). This limitation severely constrains the comprehensive evaluation of OvOD methods, which require a broad and diverse category set (or say vocabulary) to assess generalization to unseen object semantics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 354, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 354, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 54, + 354, + 295, + 510 + ], + "type": "text", + "content": "To address this gap, we introduce DET-COMPASS, a novel benchmark that repurposes the COMPASS-XP classification dataset [9] for object detection through meticulous manual bounding box annotation. DET-COMPASS comprises 370 distinct object classes, offering an order-of-magnitude increase in vocabulary size over previous X-ray detection benchmarks. Additionally, it provides pixel-aligned RGB images, ensuring precise spatial correspondence across modalities and facilitating the development of multimodal models. Each object is also labeled with a visibility attribute, indicating whether it produces a discernible signature in the X-ray spectrum. Further details are provided in Appendix C." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 510, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 582 + ], + "type": "text", + "content": "As summarized in Tab. 1, DET-COMPASS sets a new standard in class diversity and uniquely integrates multimodal, pixel-aligned X-ray and RGB data. The dataset will be released under an open license, serving as a valuable resource for advancing OvOD research in security screening and industrial inspection." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 200, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 200, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 200, + 608 + ], + "type": "text", + "content": "4. OvOD for X-ray Imaging" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 615, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 615, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 615, + 295, + 715 + ], + "type": "text", + "content": "Preliminaries: OvOD in the RGB domain. Most RGB OvOD detectors [44, 45] follow a two-stage pipeline. During training, a region proposal network (RPN) is learned to yield a set of " + }, + { + "bbox": [ + 55, + 615, + 295, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 615, + 295, + 715 + ], + "type": "text", + "content": " proposals by " + }, + { + "bbox": [ + 55, + 615, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}_m\\}_{m=1}^M = \\Phi_{\\mathrm{RPN}}(\\mathbf{I}^{RGB})" + }, + { + "bbox": [ + 55, + 615, + 295, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 615, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_m \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 55, + 615, + 295, + 715 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 55, + 615, + 295, + 715 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 615, + 295, + 715 + ], + "type": "text", + "content": "-dimensional region-of-interest (RoI) feature embedding. Then, a bounding box regressor predicts coordinates for each proposed region via " + }, + { + "bbox": [ + 55, + 615, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{b}}_m = \\Phi_{\\mathrm{REG}}(\\mathbf{z}_m)" + }, + { + "bbox": [ + 55, + 615, + 295, + 715 + ], + "type": "text", + "content": ". A set of text-based classifiers " + }, + { + "bbox": [ + 55, + 615, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{W} = \\{\\mathbf{w}_c | \\mathbf{w}_c \\in \\mathbb{R}^D\\}_{c=1}^{|C_{\\text{train}}|}" + }, + { + "bbox": [ + 55, + 615, + 295, + 715 + ], + "type": "text", + "content": " are used" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": "to compute classification scores for each region as " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "\\langle \\mathbf{w}_c, \\mathbf{z}_m \\rangle" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "\\langle \\cdot, \\cdot \\rangle" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " is the cosine similarity function and " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "C^{\\mathrm{train}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " denotes the training vocabulary. In this way, each region's class is determined by the class with the highest score. Here, the classifier " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " is constructed by encoding class names in " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "C^{\\mathrm{train}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " using a pre-trained VLM text encoder, e.g., CLIP [26]. During training, OvOD models update all parameters while keeping " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " frozen. This enables RGB-region-class alignment by leveraging large-scale RGB data and the pre-aligned vision-language semantic space of VLMs, facilitating open-vocabulary inference with any test vocabulary " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "C^{\\mathrm{test}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": ". The vocabularies " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "C^{\\mathrm{train}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "C^{\\mathrm{test}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " may be disjoint or overlapping." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 220, + 553, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 553, + 341 + ], + "type": "text", + "content": "Problem formulation. In this work, we study open-vocabulary object detection (OvOD) in X-ray modality. Specifically, given an input X-ray image " + }, + { + "bbox": [ + 313, + 220, + 553, + 341 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 313, + 220, + 553, + 341 + ], + "type": "text", + "content": " and a vocabulary " + }, + { + "bbox": [ + 313, + 220, + 553, + 341 + ], + "type": "inline_equation", + "content": "C^{\\mathrm{test}}" + }, + { + "bbox": [ + 313, + 220, + 553, + 341 + ], + "type": "text", + "content": " defined by users at test time, an X-ray OvOD detector " + }, + { + "bbox": [ + 313, + 220, + 553, + 341 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{X - ray}}" + }, + { + "bbox": [ + 313, + 220, + 553, + 341 + ], + "type": "text", + "content": " aims to detect objects specified in " + }, + { + "bbox": [ + 313, + 220, + 553, + 341 + ], + "type": "inline_equation", + "content": "C^{\\mathrm{test}}" + }, + { + "bbox": [ + 313, + 220, + 553, + 341 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 220, + 553, + 341 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 313, + 220, + 553, + 341 + ], + "type": "text", + "content": " (e.g., a \"Power bank\" in a passenger's backpack scan at an airport security checkpoint). Theoretically, this detection process can be formulated as " + }, + { + "bbox": [ + 313, + 220, + 553, + 341 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{X - ray}}: \\mathbf{I} \\to \\{(\\mathbf{b}_m, c_m)\\}_{m=1}^{M}" + }, + { + "bbox": [ + 313, + 220, + 553, + 341 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 220, + 553, + 341 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_m \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 313, + 220, + 553, + 341 + ], + "type": "text", + "content": " denotes the coordinates of each bounding box, and " + }, + { + "bbox": [ + 313, + 220, + 553, + 341 + ], + "type": "inline_equation", + "content": "c_m \\in C^{\\mathrm{test}}" + }, + { + "bbox": [ + 313, + 220, + 553, + 341 + ], + "type": "text", + "content": " denotes the class label of each bounding box." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 342, + 555, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 486 + ], + "type": "text", + "content": "X-ray OvOD presents significant challenges due to the following reasons: i) Directly applying a pre-trained RGB OvOD detectors to X-ray images leads to suboptimal performance due to the visual modality gap, as shown in Fig. 1; ii) The scarcity of large-scale security X-ray datasets limits the application of RGB OvOD training techniques. These techniques typically rely on extensive image-text annotated data for strong supervision. To tackle these challenges, we introduce RAXO, a plug-and-play module that adapts any off-the-shelf RGB OvOD detectors to X-ray modalities. RAXO requires no training or in-domain detection annotations. Next, we present our approach." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 500, + 549, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 500, + 549, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 500, + 549, + 514 + ], + "type": "text", + "content": "5. RAXO: Training-free Modality Adaptation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": "As illustrated in Fig. 2, RAXO enables X-ray OvOD by constructing high-quality visual descriptors, " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "X_{c}" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": ", for each class " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " in the user-defined vocabulary, " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "C^{\\mathrm{test}}" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": ". To achieve this, RAXO first acquires X-ray samples in an open-ended manner, leveraging both in-house and web-based retrieval sources (Sec. 5.1). Then, it extracts the features of these samples and segments the relevant information, modeling the visual descriptor " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "X_{c}" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " (Sec. 5.2). Once the visual descriptors are constructed offline, they can be directly applied to any off-the-shelf OvOD detector by replacing the conventional text-based classifier " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": ", with our visual-based classifier " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " (Sec. 5.3). Thus, RAXO effectively overcomes the misalignment between X-ray features and text semantics, enabling OvOD detectors pre-trained with RGB data to accurately identify X-ray objects based on their intrinsic visual characteristics." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20772" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 70, + 541, + 286 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 541, + 286 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 541, + 286 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 541, + 286 + ], + "type": "image", + "image_path": "5a5b7f18c2a2e2a5e1ac03b8167e8cfb2316ed1a2d351756ee24b2a9607f162f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 288, + 555, + 344 + ], + "lines": [ + { + "bbox": [ + 55, + 288, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 555, + 344 + ], + "type": "text", + "content": "Figure 2. Architecture of RAXO. For a given user-defined class " + }, + { + "bbox": [ + 55, + 288, + 555, + 344 + ], + "type": "inline_equation", + "content": "c \\in C^{\\mathrm{test}}" + }, + { + "bbox": [ + 55, + 288, + 555, + 344 + ], + "type": "text", + "content": ", RAXO first retrieves its corresponding X-ray images " + }, + { + "bbox": [ + 55, + 288, + 555, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_c^{XRAY}" + }, + { + "bbox": [ + 55, + 288, + 555, + 344 + ], + "type": "text", + "content": " from in-house and web sources, using its (1) Visual Samples Acquisition pipeline (Sec. 5.1). Following this, RAXO extracts the features of the images and segments them with its (2) Class Descriptor Modeling module (Sec. 5.2), creating ensemble visual descriptors for the class " + }, + { + "bbox": [ + 55, + 288, + 555, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_c" + }, + { + "bbox": [ + 55, + 288, + 555, + 344 + ], + "type": "text", + "content": " and the background " + }, + { + "bbox": [ + 55, + 288, + 555, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{bg}}" + }, + { + "bbox": [ + 55, + 288, + 555, + 344 + ], + "type": "text", + "content": ". Finally, the text-based classifier from the baseline RGB OvOD detector is replaced with our (3) Visual-based Classifier (Sec. 5.3) " + }, + { + "bbox": [ + 55, + 288, + 555, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 55, + 288, + 555, + 344 + ], + "type": "text", + "content": ", which yields accurate predictions on the X-ray modality." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 364, + 201, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 364, + 201, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 201, + 376 + ], + "type": "text", + "content": "5.1. Visual Sample Acquisition" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 381, + 296, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 381, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 381, + 296, + 464 + ], + "type": "text", + "content": "Obtaining informative and representative images is crucial for generating robust visual descriptors. To this end, we propose a dual-source acquisition pipeline that retrieves a gallery of relevant X-ray samples " + }, + { + "bbox": [ + 55, + 381, + 296, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_c^{\\mathrm{XRAY}}" + }, + { + "bbox": [ + 55, + 381, + 296, + 464 + ], + "type": "text", + "content": " for each user-specified class " + }, + { + "bbox": [ + 55, + 381, + 296, + 464 + ], + "type": "inline_equation", + "content": "c\\in C^{\\mathrm{test}}" + }, + { + "bbox": [ + 55, + 381, + 296, + 464 + ], + "type": "text", + "content": ". This pipeline consists of two modules: an in-house retrieval module and a web-powered retrieval module." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 468, + 296, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 296, + 531 + ], + "type": "text", + "content": "In-house retrieval. Given an in-domain X-ray dataset, " + }, + { + "bbox": [ + 55, + 468, + 296, + 531 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{XRAY}}^{\\mathrm{in - house}}" + }, + { + "bbox": [ + 55, + 468, + 296, + 531 + ], + "type": "text", + "content": " containing categories " + }, + { + "bbox": [ + 55, + 468, + 296, + 531 + ], + "type": "inline_equation", + "content": "C^{\\mathrm{in - house}}" + }, + { + "bbox": [ + 55, + 468, + 296, + 531 + ], + "type": "text", + "content": " , our sample acquisition pipeline first attempts to retrieve the " + }, + { + "bbox": [ + 55, + 468, + 296, + 531 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 468, + 296, + 531 + ], + "type": "text", + "content": " most relevant images for the user-specified class " + }, + { + "bbox": [ + 55, + 468, + 296, + 531 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 468, + 296, + 531 + ], + "type": "text", + "content": " based on class name matching: " + }, + { + "bbox": [ + 55, + 468, + 296, + 531 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_c^{\\mathrm{XRAY}} = \\bigl \\{\\mathbf{I}|\\bigl (\\mathbf{I},c\\bigr)\\in \\mathcal{D}_{\\mathrm{XRAY}}^{\\mathrm{in - house}}\\bigr \\} ^K" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 531, + 296, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 296, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 296, + 638 + ], + "type": "text", + "content": "However, since the user-defined vocabulary " + }, + { + "bbox": [ + 55, + 531, + 296, + 638 + ], + "type": "inline_equation", + "content": "C^{\\mathrm{test}}" + }, + { + "bbox": [ + 55, + 531, + 296, + 638 + ], + "type": "text", + "content": " is open-ended, we cannot assume that every user-specified category will be present in " + }, + { + "bbox": [ + 55, + 531, + 296, + 638 + ], + "type": "inline_equation", + "content": "C^{\\mathrm{in-house}}" + }, + { + "bbox": [ + 55, + 531, + 296, + 638 + ], + "type": "text", + "content": ". Consequently, for some categories, the retrieved set " + }, + { + "bbox": [ + 55, + 531, + 296, + 638 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_c^{\\mathrm{XRAY}}" + }, + { + "bbox": [ + 55, + 531, + 296, + 638 + ], + "type": "text", + "content": " may be empty. To overcome this limitation and fully support open-vocabulary user input, RAXO further incorporates a novel web-powered retrieval module. This module retrieves RGB images from the web and applies a material transfer mechanism to synthesize X-ray-style representations, which we introduce next." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": "Web-powered retrieval. To obtain high-quality visual samples for a category " + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": " that is not present in the in-house vocabulary " + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "inline_equation", + "content": "C^{\\mathrm{in - house}}" + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": ", we leverage the vast availability of web-based RGB image data " + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{RGB}}^{\\mathrm{web}}" + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": " as an open-ended auxiliary source. Specifically, we perform text-based web retrieval using the class name " + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": " as a search" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 365, + 555, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 392 + ], + "type": "text", + "content": "query, retrieving the top " + }, + { + "bbox": [ + 313, + 365, + 555, + 392 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 365, + 555, + 392 + ], + "type": "text", + "content": " results from Google Images as " + }, + { + "bbox": [ + 313, + 365, + 555, + 392 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{G}}_c^{\\mathrm{web}} = \\left\\{\\mathbf{I}^{\\mathrm{RGB}} \\mid (\\mathbf{I}^{\\mathrm{RGB}}, c) \\in \\mathcal{D}_{\\mathrm{RGB}}^{\\mathrm{web}}\\right\\}^K" + }, + { + "bbox": [ + 313, + 365, + 555, + 392 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 394, + 554, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 554, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 554, + 479 + ], + "type": "text", + "content": "The raw web-retrieved results " + }, + { + "bbox": [ + 313, + 394, + 554, + 479 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{G}}_c^{\\mathrm{web}}" + }, + { + "bbox": [ + 313, + 394, + 554, + 479 + ], + "type": "text", + "content": " are often noisy and may not always contain clear instances of the target class " + }, + { + "bbox": [ + 313, + 394, + 554, + 479 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 394, + 554, + 479 + ], + "type": "text", + "content": ". To refine these results, we apply a filtering step using an RGB OvOD detector " + }, + { + "bbox": [ + 313, + 394, + 554, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{RGB}}" + }, + { + "bbox": [ + 313, + 394, + 554, + 479 + ], + "type": "text", + "content": ". Specifically, we discard images in the raw web-retrieved results where class " + }, + { + "bbox": [ + 313, + 394, + 554, + 479 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 394, + 554, + 479 + ], + "type": "text", + "content": " is not confidently detected, retaining only those where the detection confidence exceeds a threshold " + }, + { + "bbox": [ + 313, + 394, + 554, + 479 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 394, + 554, + 479 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 394, + 554, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_c^{\\mathrm{web}} = \\mathrm{Filter}(\\widetilde{\\mathcal{G}}_c^{\\mathrm{web}},\\mathcal{F}_{\\mathrm{RGB}},c,\\tau)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 484, + 554, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 554, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 554, + 651 + ], + "type": "text", + "content": "Material-transfer mechanism for web-retrieved images. The substantial visual disparity between RGB and X-ray modalities prevents direct use of web-retrieved RGB samples " + }, + { + "bbox": [ + 313, + 484, + 554, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_c^{\\mathrm{web}}" + }, + { + "bbox": [ + 313, + 484, + 554, + 651 + ], + "type": "text", + "content": " for constructing X-ray class descriptors. Style transfer methods [7] fall short in bridging this gap, as they fail to capture the underlying material properties of objects (as shown in Tab. 4). To address this, we introduce a novel material-transfer mechanism for generating synthetic X-ray samples from web-retrieved images. As shown in Fig. 3, our approach consists of two key steps: " + }, + { + "bbox": [ + 313, + 484, + 554, + 651 + ], + "type": "inline_equation", + "content": "i)" + }, + { + "bbox": [ + 313, + 484, + 554, + 651 + ], + "type": "text", + "content": " constructing an offline material database " + }, + { + "bbox": [ + 313, + 484, + 554, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 484, + 554, + 651 + ], + "type": "text", + "content": " that encapsulates the expected X-ray appearance of various materials, and " + }, + { + "bbox": [ + 313, + 484, + 554, + 651 + ], + "type": "inline_equation", + "content": "ii)" + }, + { + "bbox": [ + 313, + 484, + 554, + 651 + ], + "type": "text", + "content": " adapting RGB samples " + }, + { + "bbox": [ + 313, + 484, + 554, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_c^{\\mathrm{web}}" + }, + { + "bbox": [ + 313, + 484, + 554, + 651 + ], + "type": "text", + "content": " to X-ray style by applying the corresponding material properties from " + }, + { + "bbox": [ + 313, + 484, + 554, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 484, + 554, + 651 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "To construct " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": ", we employ a Large Language Model (LLM) to cluster the class names of " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "C^{\\mathrm{in - house}}" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": " into subsets " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "C_m\\subset C^{\\mathrm{in - house}}" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "C_m" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": " contains classes that share the same " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": " material - i.e., \" " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{metal}} = \\{\\mathrm{gun},\\mathrm{knife},\\mathrm{fork}\\}" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "\" or \" " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{leather}} = \\{\\mathrm{boot},\\mathrm{belt}\\}" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "\". Each material " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": " is then associated" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20773" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 294, + 258 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 294, + 258 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 294, + 258 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 294, + 258 + ], + "type": "image", + "image_path": "a88c4c7f2a5b725a52a3a73abfc50ad5408cb2bd816543e140661b8258b0ac43.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 263, + 296, + 308 + ], + "lines": [ + { + "bbox": [ + 55, + 263, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 296, + 308 + ], + "type": "text", + "content": "Figure 3. Web-powered retrieval and material-transfer mechanism for the class \"violin\". We retrieve violin samples from the web, filter them using " + }, + { + "bbox": [ + 55, + 263, + 296, + 308 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{RGB}}" + }, + { + "bbox": [ + 55, + 263, + 296, + 308 + ], + "type": "text", + "content": ", and inpaint the retrieved appearance into the object masks to generate synthetic X-ray samples." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 327, + 296, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 365 + ], + "type": "text", + "content": "with the corresponding set of images " + }, + { + "bbox": [ + 55, + 327, + 296, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_m = \\{\\mathbf{I} \\mid (\\mathbf{I}, c_k) \\in \\mathcal{D}_{\\mathrm{XRAY}}^{\\mathrm{in-house}}, c_k \\in C_m\\}" + }, + { + "bbox": [ + 55, + 327, + 296, + 365 + ], + "type": "text", + "content": ". These images are used to compute the appearance, " + }, + { + "bbox": [ + 55, + 327, + 296, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_m \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 327, + 296, + 365 + ], + "type": "text", + "content": ", of the material as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 369, + 295, + 404 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 369, + 295, + 404 + ], + "spans": [ + { + "bbox": [ + 85, + 369, + 295, + 404 + ], + "type": "interline_equation", + "content": "\\mathcal {A} _ {m} = \\frac {1}{| \\mathcal {K} _ {m} |} \\sum_ {\\mathbf {I} \\in \\mathcal {K} _ {m}} \\left(\\frac {\\sum_ {i , j} ^ {H , W} \\left(\\mathbf {I} _ {i , j} \\odot \\Omega \\left(\\mathbf {I} _ {i , j}\\right)\\right)}{\\sum \\Omega (\\mathbf {I})}\\right), \\tag {1}", + "image_path": "b47ddf251bab5ff5f51e2c834b3955fec7ff82dd13585cdec1385d7070a1ca62.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 411, + 296, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 296, + 470 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 411, + 296, + 470 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 411, + 296, + 470 + ], + "type": "text", + "content": " denotes element-wise multiplication and " + }, + { + "bbox": [ + 55, + 411, + 296, + 470 + ], + "type": "inline_equation", + "content": "\\Omega : \\mathbf{I} \\to \\{0,1\\}^{H \\times W}" + }, + { + "bbox": [ + 55, + 411, + 296, + 470 + ], + "type": "text", + "content": " is a segmentation function that produces a binary mask, to isolate the object of interest in " + }, + { + "bbox": [ + 55, + 411, + 296, + 470 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 55, + 411, + 296, + 470 + ], + "type": "text", + "content": ". The resulting material-appearance pairs " + }, + { + "bbox": [ + 55, + 411, + 296, + 470 + ], + "type": "inline_equation", + "content": "(m, \\mathcal{A}_m)" + }, + { + "bbox": [ + 55, + 411, + 296, + 470 + ], + "type": "text", + "content": " collectively form our material database " + }, + { + "bbox": [ + 55, + 411, + 296, + 470 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 411, + 296, + 470 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "type": "text", + "content": "Once " + }, + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "type": "text", + "content": " is computed offline, it can be used to adapt the web-retrieved RGB samples, " + }, + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_c^{\\mathrm{web}}" + }, + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "type": "text", + "content": ", to the X-ray modality. To achieve this, we first prompt the LLM to link the class " + }, + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "type": "text", + "content": " with its corresponding material " + }, + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "type": "inline_equation", + "content": "m_c \\in \\mathcal{M}" + }, + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "type": "text", + "content": " (e.g., " + }, + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "type": "inline_equation", + "content": "c = \\text{\"violin\"}" + }, + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "type": "text", + "content": " is linked with the material " + }, + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "type": "inline_equation", + "content": "m_c = \\text{\"wood\"}" + }, + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "type": "text", + "content": "). Subsequently, the X-ray appearance " + }, + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_m^c" + }, + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "type": "text", + "content": " of the linked material is used to render the characteristic X-ray style on each web-retrieved RGB samples as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 572, + 295, + 588 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 572, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 70, + 572, + 295, + 588 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {c} ^ {\\mathrm {X R A Y}} = \\left\\{\\Omega \\left(\\mathbf {I} ^ {\\mathrm {R G B}}\\right) \\odot \\left(\\mathcal {A} _ {m} ^ {c} \\cdot \\mathbf {1}\\right) \\mid \\mathbf {I} ^ {\\mathrm {R G B}} \\in \\mathcal {G} _ {c} ^ {\\mathrm {w e b}} \\right\\}, \\tag {2}", + "image_path": "2edaa4a01caf53375385d244788ef68729bbfbe23c9e4a5bd9aa9e6c3c9bc98d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 296, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 654 + ], + "type": "text", + "content": "Finally, the X-ray visual gallery for every class freely specified by the user is built as " + }, + { + "bbox": [ + 55, + 594, + 296, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = \\bigcup_{c\\in C^{\\mathrm{test}}}\\mathcal{G}_c^{\\mathrm{XRAY}}" + }, + { + "bbox": [ + 55, + 594, + 296, + 654 + ], + "type": "text", + "content": " through our dual-source retrieval pipeline, seamlessly integrating high-quality in-house X-ray samples with synthetic images derived from web data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 660, + 204, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 204, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 204, + 673 + ], + "type": "text", + "content": "5.2. Class Descriptor Modeling" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "The effectiveness of RAXO hinges on leveraging its visual gallery " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " to construct a robust visual descriptor " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\chi_{c}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " that accurately captures the visual X-ray properties of class " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ". A" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 155 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 155 + ], + "type": "text", + "content": "naive approach, such as averaging the feature representations of all " + }, + { + "bbox": [ + 313, + 72, + 553, + 155 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_c^{\\mathrm{XRAY}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 155 + ], + "type": "text", + "content": " instances, fails to account for intra-class variability (e.g., distinct shapes of utility knives versus chef knives) and does not differentiate between foreground and background regions. To address these limitations, we propose a novel class descriptor modeling strategy shown in Fig. 2(2)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 156, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 555, + 300 + ], + "type": "text", + "content": "First, we process each sample " + }, + { + "bbox": [ + 313, + 156, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathcal{G}_c^{XRAY}" + }, + { + "bbox": [ + 313, + 156, + 555, + 300 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 156, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 156, + 555, + 300 + ], + "type": "text", + "content": " to separate the X-ray object from its background. Simultaneously, a feature extractor " + }, + { + "bbox": [ + 313, + 156, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 156, + 555, + 300 + ], + "type": "text", + "content": " extracts per-patch embeddings " + }, + { + "bbox": [ + 313, + 156, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\phi(\\mathbf{I}) \\in \\mathbb{R}^{H' \\times W' \\times D}" + }, + { + "bbox": [ + 313, + 156, + 555, + 300 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 156, + 555, + 300 + ], + "type": "inline_equation", + "content": "H' \\times W'" + }, + { + "bbox": [ + 313, + 156, + 555, + 300 + ], + "type": "text", + "content": " represents the number of spatial tokens, and " + }, + { + "bbox": [ + 313, + 156, + 555, + 300 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 156, + 555, + 300 + ], + "type": "text", + "content": " is the embedding dimension. Following [10], we resize (denoted by " + }, + { + "bbox": [ + 313, + 156, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 313, + 156, + 555, + 300 + ], + "type": "text", + "content": ") the segmentation mask to match the spatial dimensions of the feature map, and subsequently we use it to compute both a positive and a negative prototype. The positive prototype, designed to capture the significant visual features of the object without background interference, is computed as the average embedding over the foreground region as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 359, + 306, + 553, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 306, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 359, + 306, + 553, + 338 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {\\mathbf {I}} ^ {\\text {p o s}} = \\frac {\\sum_ {i , j} ^ {H ^ {\\prime} , W ^ {\\prime}} \\zeta (\\Omega (\\mathbf {I} _ {i , j})) \\odot \\phi (\\mathbf {I} _ {i , j})}{\\sum \\zeta (\\Omega (\\mathbf {I}))}, \\tag {3}", + "image_path": "4ddb61d4a0e0b6b36e2da36827a145cfbda65bcf01dc438be2933dd646c283a1.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 342, + 553, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 553, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 553, + 378 + ], + "type": "text", + "content": "On the other hand, the negative prototype is obtained as the average embedding over the complementary region (background) as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 350, + 384, + 553, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 384, + 553, + 416 + ], + "spans": [ + { + "bbox": [ + 350, + 384, + 553, + 416 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {\\mathbf {I}} ^ {\\text {n e g}} = \\frac {\\sum_ {i , j} ^ {H ^ {\\prime} , W ^ {\\prime}} \\left(1 - \\zeta \\left(\\Omega \\left(\\mathbf {I} _ {\\mathbf {i} , \\mathbf {j}}\\right)\\right)\\right) \\odot \\phi \\left(\\mathbf {I} _ {\\mathbf {i} , \\mathbf {j}}\\right)}{\\sum \\left(1 - \\zeta \\left(\\Omega (\\mathbf {I})\\right)\\right)}. \\tag {4}", + "image_path": "dac46e8a3c7013146fbbe5fbf3e8a178b413b68e354ed7afff6271dfc291c9a0.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 422, + 553, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 553, + 470 + ], + "type": "text", + "content": "Subsequently, to construct the final visual descriptor " + }, + { + "bbox": [ + 313, + 422, + 553, + 470 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_c" + }, + { + "bbox": [ + 313, + 422, + 553, + 470 + ], + "type": "text", + "content": " of class " + }, + { + "bbox": [ + 313, + 422, + 553, + 470 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 422, + 553, + 470 + ], + "type": "text", + "content": ", we compute the average positive prototype " + }, + { + "bbox": [ + 313, + 422, + 553, + 470 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{x}}_c^{\\mathrm{pos}}" + }, + { + "bbox": [ + 313, + 422, + 553, + 470 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 422, + 553, + 470 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_c^{XRAY}" + }, + { + "bbox": [ + 313, + 422, + 553, + 470 + ], + "type": "text", + "content": " and unite it with all the individual positive prototypes as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 375, + 476, + 553, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 476, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 375, + 476, + 553, + 491 + ], + "type": "interline_equation", + "content": "\\mathcal {X} _ {c} = \\left[ \\bar {\\mathbf {x}} _ {c} ^ {\\text {p o s}}, \\left\\{\\mathbf {x} _ {\\mathbf {I}} ^ {\\text {p o s}} \\mid \\mathbf {I} \\in \\mathcal {G} _ {c} \\right\\} \\right]. \\tag {5}", + "image_path": "8a7de2ad3b69e3dd27febfee4d337214e87dc491dcc0b33f31059b068208b302.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 498, + 555, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 654 + ], + "type": "text", + "content": "This formulation effectively captures both fine-grained object details and a holistic class-level representation, handling intra-class variability. Additionally, we also construct a global background descriptor by using the negative prototypes across the entire visual sample gallery as " + }, + { + "bbox": [ + 313, + 498, + 555, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{bg} = [\\overline{\\mathbf{x}}^{\\mathrm{neg}},\\{\\mathbf{x}_{\\mathbf{I}}^{\\mathrm{neg}}\\mid \\mathbf{I}\\in \\mathcal{G}\\} ]" + }, + { + "bbox": [ + 313, + 498, + 555, + 654 + ], + "type": "text", + "content": ". This negative prototype is used to further improve detection reliability by filtering out low-quality proposals (e.g., a region proposed on the background) during inference. A key advantage of this approach is its modularity: visual descriptors " + }, + { + "bbox": [ + 313, + 498, + 555, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_c" + }, + { + "bbox": [ + 313, + 498, + 555, + 654 + ], + "type": "text", + "content": " are computed off-line and can be incrementally expanded with new object categories, seamlessly integrating with the OvOD paradigm without training or requiring X-ray detection data." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 660, + 474, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 474, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 474, + 672 + ], + "type": "text", + "content": "5.3. Classification is All You Need" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "As shown in Fig. 2(3), once the visual descriptors are constructed for each class in the user-specified vocabulary, along with the background class, the RAXO classifier" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20774" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\{\\left[\\mathcal{X}_c,\\mathcal{X}_{bg}\\right]\\mid c\\in C^{\\mathrm{test}}\\}" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": " can be directly applied to any OvOD detector to classify proposals " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "\\mathbf{z_m}" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 123, + 109, + 295, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 109, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 123, + 109, + 295, + 129 + ], + "type": "interline_equation", + "content": "\\hat {c} = \\underset {c \\in C ^ {\\text {t e s t}} ^ {\\prime}} {\\arg \\max } \\underset {\\mathbf {x} \\in \\mathcal {X}} {\\max } \\langle \\mathbf {z} _ {m}, \\mathbf {x} \\rangle , \\tag {6}", + "image_path": "d23dd93a7fbb0419f67b448b2baff146372f04de329446cfa42765c88852ce6f.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 136, + 296, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 136, + 296, + 209 + ], + "spans": [ + { + "bbox": [ + 55, + 136, + 296, + 209 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 136, + 296, + 209 + ], + "type": "inline_equation", + "content": "C^{\\mathrm{test}}" + }, + { + "bbox": [ + 55, + 136, + 296, + 209 + ], + "type": "text", + "content": " denotes the test-time vocabulary " + }, + { + "bbox": [ + 55, + 136, + 296, + 209 + ], + "type": "inline_equation", + "content": "C^{\\mathrm{test}}" + }, + { + "bbox": [ + 55, + 136, + 296, + 209 + ], + "type": "text", + "content": " extended with the additional \"background\" class. If a proposal matches the background, it is removed. This simple-yet-effective strategy enables any pre-trained OvOD detectors to achieve strong performance on X-ray object detection with minimal modifications and no need for re-training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 212, + 296, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 212, + 296, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 296, + 368 + ], + "type": "text", + "content": "Descriptor consistency criterion. To further reduce incorrect proposals from the OvOD detector, RAXO introduces a novel Descriptor Consistency Criterion (DCC) to enforce class consistency in predictions. DCC evaluates how well a proposal aligns with its predicted class relative to others, suppressing weakly aligned proposals. For a proposal " + }, + { + "bbox": [ + 55, + 212, + 296, + 368 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_m" + }, + { + "bbox": [ + 55, + 212, + 296, + 368 + ], + "type": "text", + "content": ", we measure its similarity to the closest prototype as " + }, + { + "bbox": [ + 55, + 212, + 296, + 368 + ], + "type": "inline_equation", + "content": "s_1 = \\max_{\\mathbf{x} \\in \\mathcal{X}_{\\hat{c}}} \\langle \\mathbf{z}_m, \\mathbf{x} \\rangle" + }, + { + "bbox": [ + 55, + 212, + 296, + 368 + ], + "type": "text", + "content": ". We also compute its mean similarity to the average prototypes of all other classes as " + }, + { + "bbox": [ + 55, + 212, + 296, + 368 + ], + "type": "inline_equation", + "content": "s_2 = \\mathrm{avg}_{c \\in C^{\\mathrm{test}}, c \\neq \\hat{c}} \\langle \\mathbf{z}_m, \\bar{\\mathbf{x}}_c^{\\mathrm{pos}} \\rangle" + }, + { + "bbox": [ + 55, + 212, + 296, + 368 + ], + "type": "text", + "content": ". The difference " + }, + { + "bbox": [ + 55, + 212, + 296, + 368 + ], + "type": "inline_equation", + "content": "\\Delta = s_1 - s_2" + }, + { + "bbox": [ + 55, + 212, + 296, + 368 + ], + "type": "text", + "content": " serves as a confidence measure, where a higher value indicates greater alignment with its predicted class than with any other. Proposals with " + }, + { + "bbox": [ + 55, + 212, + 296, + 368 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 55, + 212, + 296, + 368 + ], + "type": "text", + "content": " below a threshold " + }, + { + "bbox": [ + 55, + 212, + 296, + 368 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 212, + 296, + 368 + ], + "type": "text", + "content": " are discarded." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 380, + 136, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 136, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 136, + 392 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 399, + 296, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 296, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 296, + 554 + ], + "type": "text", + "content": "Evaluation protocol. Conventional OvOD models are typically assessed using a Cross-Dataset Transfer Evaluation (CDTE) protocol, where the model is trained on one dataset and evaluated on a different dataset in a zero-shot manner [45]. In contrast, our setting involves a more challenging scenario where both the datasets and their underlying modalities differ. To address this, we introduce a novel Cross-Modality Transfer Evaluation (CMTE) protocol. Under CMTE, an OvOD model is trained on a source RGB dataset and subsequently evaluated on target " + }, + { + "bbox": [ + 55, + 399, + 296, + 554 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 399, + 296, + 554 + ], + "type": "text", + "content": "-ray datasets without any additional training or fine-tuning. For performance assessment, we employ the standard MS COCO [17] metrics: AP, AP50, and AP75." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": "Implementation details. Unless stated otherwise, RAXO's default configuration employs SAM 2 [27] as the object segmentation module " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": ", ViT-B/14 [3] pretrained with DINOv2 [25] as the visual encoder " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": ", and GPT-4 [24] as the LLM. We set the consistency threshold to " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\sigma = 0.15" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": ", the similarity threshold to " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\tau = 0.5" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": ", and use " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "K = 30" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " samples to construct the visual descriptors. " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{XRAY}}^{in - house}" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " is composed of hold-out samples sourced from the evaluation X-ray datasets and is completely disjoint from the test split. " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{RGB}}^{web}" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " is constructed by collecting images from the web using the public Google Custom Search API [8]. A detailed analysis of the impact of hyperparameter choices is provided in Sec. 6.2. Further implementation details can be found in Supp. D." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": "Datasets and baselines. We evaluate RAXO by integrating it with four different state-of-the-art OvOD detectors that were exclusively trained on RGB images: GroundingDINO [18], Detic [44], VLDet [15], and CoDet [21]. Our CMTE evaluation is performed across six diverse X-ray datasets: PIXray [19], PIDray [34], CLCXray [42], DvXray [20], HiXray [30], and our proposed DET-COMPASS, which together comprise 343 visible unique classes and over 140k images. A detailed description of these datasets is provided in Tab. 1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 201, + 504, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 504, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 504, + 213 + ], + "type": "text", + "content": "6.1. Open-Vocabulary Detection Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 219, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 219, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 219, + 555, + 373 + ], + "type": "text", + "content": "RAXO equips OvOD models with X-ray vision. We evaluate the X-ray adaptation capabilities of RAXO on the aforementioned datasets and baselines in Tab. 2. Since RAXO is training-free, we report the results directly on the test splits, accounting for all available categories. Furthermore, we analyze the influence of the gallery composition by varying " + }, + { + "bbox": [ + 313, + 219, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 219, + 555, + 373 + ], + "type": "text", + "content": " from a " + }, + { + "bbox": [ + 313, + 219, + 555, + 373 + ], + "type": "inline_equation", + "content": "0/100" + }, + { + "bbox": [ + 313, + 219, + 555, + 373 + ], + "type": "text", + "content": " configuration (i.e., all samples in " + }, + { + "bbox": [ + 313, + 219, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 219, + 555, + 373 + ], + "type": "text", + "content": " are retrieved from the web, as " + }, + { + "bbox": [ + 313, + 219, + 555, + 373 + ], + "type": "inline_equation", + "content": "C^{\\mathrm{test}} \\cap C^{\\mathrm{in-house}} = \\emptyset" + }, + { + "bbox": [ + 313, + 219, + 555, + 373 + ], + "type": "text", + "content": ") to a " + }, + { + "bbox": [ + 313, + 219, + 555, + 373 + ], + "type": "inline_equation", + "content": "100/0" + }, + { + "bbox": [ + 313, + 219, + 555, + 373 + ], + "type": "text", + "content": " configuration (i.e., all samples in " + }, + { + "bbox": [ + 313, + 219, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 219, + 555, + 373 + ], + "type": "text", + "content": " are retrieved from the in-house database, as " + }, + { + "bbox": [ + 313, + 219, + 555, + 373 + ], + "type": "inline_equation", + "content": "C^{\\mathrm{test}} \\subset C^{\\mathrm{in-house}}" + }, + { + "bbox": [ + 313, + 219, + 555, + 373 + ], + "type": "text", + "content": "). Each experiment is repeated three times with different random distributions of in-domain and web categories, and we report the averaged results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 375, + 556, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 556, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 556, + 579 + ], + "type": "text", + "content": "As shown in Tab. 2, off-the-shelf OvOD detectors trained on RGB images perform poorly on X-ray data. In contrast, even without access to any in-domain data from " + }, + { + "bbox": [ + 313, + 375, + 556, + 579 + ], + "type": "inline_equation", + "content": "C^{\\mathrm{test}}" + }, + { + "bbox": [ + 313, + 375, + 556, + 579 + ], + "type": "text", + "content": " (0/100 setting), RAXO enhances baseline OvOD methods, yielding an average improvement of " + }, + { + "bbox": [ + 313, + 375, + 556, + 579 + ], + "type": "inline_equation", + "content": "\\uparrow 2.1" + }, + { + "bbox": [ + 313, + 375, + 556, + 579 + ], + "type": "text", + "content": " points. This gain is further amplified when in-domain samples are incorporated into " + }, + { + "bbox": [ + 313, + 375, + 556, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 375, + 556, + 579 + ], + "type": "text", + "content": ", which facilitates the construction of more robust materials in " + }, + { + "bbox": [ + 313, + 375, + 556, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 375, + 556, + 579 + ], + "type": "text", + "content": " and yields higher-quality visual descriptors. Notably, even a limited number of in-domain samples (20/80 setting) results in an average improvement of " + }, + { + "bbox": [ + 313, + 375, + 556, + 579 + ], + "type": "inline_equation", + "content": "\\uparrow 4.8" + }, + { + "bbox": [ + 313, + 375, + 556, + 579 + ], + "type": "text", + "content": " points. Moreover, increasing the in-domain samples (50/50 setting) yields even more significant gains, particularly in datasets with a large number of classes, such as DET-COMPASS, where RAXO offers an average boost of " + }, + { + "bbox": [ + 313, + 375, + 556, + 579 + ], + "type": "inline_equation", + "content": "\\uparrow 14.9" + }, + { + "bbox": [ + 313, + 375, + 556, + 579 + ], + "type": "text", + "content": " points. Finally, when " + }, + { + "bbox": [ + 313, + 375, + 556, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 375, + 556, + 579 + ], + "type": "text", + "content": " consists entirely of in-domain samples, RAXO delivers an average improvement of " + }, + { + "bbox": [ + 313, + 375, + 556, + 579 + ], + "type": "inline_equation", + "content": "\\uparrow 14.7" + }, + { + "bbox": [ + 313, + 375, + 556, + 579 + ], + "type": "text", + "content": " points across all OvOD methods and datasets, all in a training-free manner." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "RAXO works with any OvOD detector. The consistent gains of RAXO across all evaluated detectors in Tab. 2, along with its simple integration (Sec. 5.3), confirm the generalization of our approach to any RGB OvOD model. Moreover, RAXO scales with the baseline method, yielding larger gains when integrated with more robust detectors (e.g., RAXO enhances G-DINO " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "9.4 points on average across all " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": " settings, compared to " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "7.5 points for Detic). This highlights the advantage of our training-free method, which can leverage the rapid advancements in RGB OvOD detectors without needing large amounts of labeled X-ray data." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20775" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 70, + 534, + 346 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 534, + 346 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 534, + 346 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 534, + 346 + ], + "type": "table", + "html": "
GMethodD-COMPASSPIXrayPIDrayCLCXrayDvXrayHiXrayAvg.
G-DINO [18]13.412.910.96.710.07.010.2
100/80/2050/5020/800/100+ RAXO47.9↑34.536.9↑24.016.5↑5.622.2↑15.522.6↑12.617.1↑10.127.2↑17.0
41.0↑27.633.8↑20.915.4↑4.518.0↑11.321.0↑11.014.5↑7.524.0↑13.8
31.4↑18.025.4↑12.515.5↑4.617.0↑10.316.1↑6.113.4↑6.419.8↑9.6
20.5↑7.121.6↑8.713.9↑3.010.0↑3.315.0↑5.09.8↑2.815.1↑4.9
14.0↑0.616.1↑3.213.4↑2.57.1↑0.412.4↑2.47.9↑0.911.8↑1.6
VLDet [15]10.69.86.94.47.45.17.4
100/80/2050/5020/800/100+ RAXO36.4↑25.832.3↑22.511.7↑4.815.4↑11.020.1↑12.714.8↑9.721.8↑14.4
31.8↑21.229.2↑19.411.0↑4.112.7↑8.316.8↑9.413.1↑8.019.1↑11.7
23.7↑13.124.0↑14.210.4↑3.511.1↑6.712.1↑4.711.2↑6.115.4↑8.0
16.2↑5.621.6↑11.89.4↑2.55.2↑0.810.6↑3.29.3↑4.212.1↑4.7
11.1↑0.514.1↑4.38.9↑2.04.4↑0.09.0↑1.68.3↑3.29.3↑1.9
Detic [44]11.59.37.14.77.04.87.4
100/80/2050/5020/800/100+ RAXO35.3↑23.827.3↑18.011.3↑4.214.0↑9.319.4↑12.414.2↑9.420.3↑12.9
30.7↑19.223.9↑14.610.8↑3.712.3↑7.618.0↑11.012.1↑7.318.0↑10.6
24.4↑12.919.5↑10.210.3↑3.29.2↑4.514.6↑7.611.0↑6.214.8↑7.4
16.4↑4.915.2↑5.99.6↑2.58.0↑3.312.7↑5.79.9↑5.112.0↑4.6
11.9↑0.413.4↑4.19.1↑2.05.2↑0.59.4↑2.47.9↑3.19.5↑2.1
CoDet [21]8.47.35.73.15.63.45.6
100/80/2050/5020/800/100+ RAXO35.8↑27.427.9↑20.610.3↑4.614.8↑11.717.6↑12.013.2↑9.819.9↑14.3
32.2↑23.825.1↑17.89.5↑3.812.0↑8.915.4↑9.811.7↑8.317.7↑12.1
24.0↑15.620.0↑12.79.5↑3.89.2↑6.111.5↑5.99.9↑6.514.0↑8.4
17.8↑9.414.8↑7.58.5↑2.85.1↑2.09.4↑3.88.1↑4.710.6↑5.0
12.2↑3.811.5↑4.28.1↑2.44.0↑0.96.9↑1.36.5↑3.18.2↑2.6
", + "image_path": "f924d6734f0dc9db20923945d47dc655339b9d8f9813858a535f3a233578b053.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 403, + 290, + 488 + ], + "blocks": [ + { + "bbox": [ + 55, + 349, + 555, + 393 + ], + "lines": [ + { + "bbox": [ + 55, + 349, + 555, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 555, + 393 + ], + "type": "text", + "content": "Table 2. X-ray OvOD performance under the Cross-Modality Transfer Evaluation (CMTE) setting on DET-COMPASS (ours), PINray [19], PIDray [34], CLCXray [42], DvXray [20], and HiXray [30] datasets. We integrate RAXO into different baselines using different gallery " + }, + { + "bbox": [ + 55, + 349, + 555, + 393 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 349, + 555, + 393 + ], + "type": "text", + "content": " compositions, from using only " + }, + { + "bbox": [ + 55, + 349, + 555, + 393 + ], + "type": "inline_equation", + "content": "{\\mathcal{D}}_{\\mathrm{{XRAY}}}^{\\text{in-house }}" + }, + { + "bbox": [ + 55, + 349, + 555, + 393 + ], + "type": "text", + "content": " data " + }, + { + "bbox": [ + 55, + 349, + 555, + 393 + ], + "type": "inline_equation", + "content": "\\left( {{100}/0}\\right)" + }, + { + "bbox": [ + 55, + 349, + 555, + 393 + ], + "type": "text", + "content": " to exclusively " + }, + { + "bbox": [ + 55, + 349, + 555, + 393 + ], + "type": "inline_equation", + "content": "{\\mathcal{D}}_{\\mathrm{{RGB}}}^{\\text{web }}" + }, + { + "bbox": [ + 55, + 349, + 555, + 393 + ], + "type": "text", + "content": " samples (0/100). RAXO consistently improves the performance of all baseline OvOD detectors across every dataset. AP is used." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 403, + 290, + 488 + ], + "lines": [ + { + "bbox": [ + 58, + 403, + 290, + 488 + ], + "spans": [ + { + "bbox": [ + 58, + 403, + 290, + 488 + ], + "type": "table", + "html": "
PIXray
ScissorsWrenchBatteryPliersAP50
OVXD†16.946.214.66.421.0
BARON + RAXO18.239.48.322.022.0
CoDet + RAXO43.652.69.018.630.9
Detic + RAXO46.360.57.936.637.8
VLDet + RAXO48.362.710.638.540.0
G-DINO + RAXO49.761.611.351.543.5
", + "image_path": "03b905b2d0b4d9c92434e9c8765ed0f0bcde604cc3691514810ab2a2e68aabab.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "RAXO surpasses training-based approaches. We also compare RAXO with OVXD [16], the only concurrent work that extends X-ray detection beyond base categories. OVXD is a fully-supervised method trained directly on X-ray data, and its generalization capabilities are evaluated on a hold-out set comprising only four categories (i.e., Scissors, Wrench, Battery, and Pliers). In Tab. 3, we follow this setting and evaluate RAXO on PINray [19], excluding the aforementioned classes from the in-house dataset " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{XRAY}}^{\\mathrm{in - house}}" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": ". Our results demonstrate that RAXO outperforms OVXD even when both approaches utilize the same BARON detector [37]. Notably, unlike OVXD, our method does not retrain BARON or any other component. Yet, our training-free RAXO performs remarkably well, achieving an AP50" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 319, + 403, + 545, + 483 + ], + "blocks": [ + { + "bbox": [ + 55, + 492, + 295, + 536 + ], + "lines": [ + { + "bbox": [ + 55, + 492, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 295, + 536 + ], + "type": "text", + "content": "Table 3. Comparison under OVXD [16] setting, where the methods do not have access to in-domain data from the displayed categories. " + }, + { + "bbox": [ + 55, + 492, + 295, + 536 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 492, + 295, + 536 + ], + "type": "text", + "content": "OVXD is a supervised method explicitly trained on X-ray images for alignment. AP50 is used." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 403, + 545, + 483 + ], + "lines": [ + { + "bbox": [ + 319, + 403, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 319, + 403, + 545, + 483 + ], + "type": "table", + "html": "
Mod. Filt. Trans.APAP50AP75
G-DINO [18]12.914.913.4
+ RAXOX-ray RGBS7.9↓5.09.8↓5.18.3↓5.1
13.9↑1.017.4↑2.514.3↑0.9
RGB✓14.3↑1.417.6↑2.714.7↑1.3
RGB✓18.3↑3.415.1↑1.7
RGB✓M16.1↑3.219.8↑4.9
", + "image_path": "31a5a8dcc10d14f7c17acb8b5511a7eebc02466e8d9830fa99f67ddac27d4de1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 488, + 555, + 544 + ], + "lines": [ + { + "bbox": [ + 313, + 488, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 488, + 555, + 544 + ], + "type": "text", + "content": "Table 4. Ablation of web-based retrieval on PIXray [19] using only web-based samples (setting " + }, + { + "bbox": [ + 313, + 488, + 555, + 544 + ], + "type": "inline_equation", + "content": "0/100" + }, + { + "bbox": [ + 313, + 488, + 555, + 544 + ], + "type": "text", + "content": "). Mod. indicates retrieval of a specific modality from the web, Filt. refers to filtering retrieved images with " + }, + { + "bbox": [ + 313, + 488, + 555, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{RGB}}" + }, + { + "bbox": [ + 313, + 488, + 555, + 544 + ], + "type": "text", + "content": ", and Trans. specifies style-transfer via: " + }, + { + "bbox": [ + 313, + 488, + 555, + 544 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 488, + 555, + 544 + ], + "type": "text", + "content": " a diffusion-based method [7] or " + }, + { + "bbox": [ + 313, + 488, + 555, + 544 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 488, + 555, + 544 + ], + "type": "text", + "content": " our material-transfer mechanism." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 553, + 476, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 553, + 476, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 553, + 476, + 565 + ], + "type": "text", + "content": "of 43.5 when paired with G-DINO [18]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 573, + 408, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 573, + 408, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 573, + 408, + 586 + ], + "type": "text", + "content": "6.2. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 590, + 554, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 554, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 554, + 639 + ], + "type": "text", + "content": "In this section we study the core components of RAXO. We conduct our experiments on PixRay [19], using GroundingDINO [13] as our baseline OvOD. Consistent findings are reported in Supp. E and H." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": "Web-based samples. Tab. 4 presents an ablation study on the components of our web-based visual sample acquisition pipeline (Sec. 5.1). Notably, directly retrieving X-ray samples from the web " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "(\\mathrm{Mod.} = \\mathrm{X}" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": " -ray) leads to a " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\downarrow 5.0" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": " point decrease in AP. In contrast, retrieving RGB images from the web " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "(\\mathrm{Mod.} = \\mathrm{RGB})" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": " provides a modest AP improvement of" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20776" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 288, + 158 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 288, + 158 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 288, + 158 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 288, + 158 + ], + "type": "table", + "html": "
Ω x neg X DCCAPAP50AP75
G-DINO [18]12.914.913.4
+ RAXO25.2↑12.331.4↑16.526.2↑12.8
✓27.1↑14.233.6↑18.728.4↑15.0
✓ ✓27.8↑14.934.5↑19.629.3↑15.9
✓27.5↑14.634.0↑19.128.8↑15.4
✓ ✓ ✓28.5↑15.635.3↑20.429.9↑16.5
✓ ✓ ✓36.9↑24.045.0↑30.139.0↑25.6
", + "image_path": "4670a4ed58b61ce1a19f01619452f9219468d48659362688d3213dba52a1db38.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 160, + 295, + 227 + ], + "lines": [ + { + "bbox": [ + 55, + 160, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 160, + 295, + 227 + ], + "type": "text", + "content": "Table 5. Ablation of class representation construction on PIXray [19] using in-domain samples (setting " + }, + { + "bbox": [ + 55, + 160, + 295, + 227 + ], + "type": "inline_equation", + "content": "100/0" + }, + { + "bbox": [ + 55, + 160, + 295, + 227 + ], + "type": "text", + "content": "). " + }, + { + "bbox": [ + 55, + 160, + 295, + 227 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 160, + 295, + 227 + ], + "type": "text", + "content": " indicates building positive prototypes using segmentation masks. " + }, + { + "bbox": [ + 55, + 160, + 295, + 227 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathrm{neg}}" + }, + { + "bbox": [ + 55, + 160, + 295, + 227 + ], + "type": "text", + "content": " indicates use of negative prototypes. " + }, + { + "bbox": [ + 55, + 160, + 295, + 227 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 55, + 160, + 295, + 227 + ], + "type": "text", + "content": " indicates ensemble multiple prototypes in class descriptors. " + }, + { + "bbox": [ + 55, + 160, + 295, + 227 + ], + "type": "inline_equation", + "content": "DCC" + }, + { + "bbox": [ + 55, + 160, + 295, + 227 + ], + "type": "text", + "content": " indicates use of prototype consistency criterion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 238, + 296, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 296, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 296, + 393 + ], + "type": "inline_equation", + "content": "\\uparrow 1.0" + }, + { + "bbox": [ + 55, + 238, + 296, + 393 + ], + "type": "text", + "content": " point. We hypothesize that directly using X-ray examples obtained from the web is ineffective due to the limited availability of high-quality X-ray images online. Filtering the retrieved RGB images using " + }, + { + "bbox": [ + 55, + 238, + 296, + 393 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{RGB}}" + }, + { + "bbox": [ + 55, + 238, + 296, + 393 + ], + "type": "text", + "content": " increases AP by an additional " + }, + { + "bbox": [ + 55, + 238, + 296, + 393 + ], + "type": "inline_equation", + "content": "\\uparrow 0.4" + }, + { + "bbox": [ + 55, + 238, + 296, + 393 + ], + "type": "text", + "content": " points. Nonetheless, they still exhibit a significant visual gap compared to their X-ray counterparts. A style-transfer approach using StyleShot [7] (Trans. " + }, + { + "bbox": [ + 55, + 238, + 296, + 393 + ], + "type": "inline_equation", + "content": "= S" + }, + { + "bbox": [ + 55, + 238, + 296, + 393 + ], + "type": "text", + "content": ") provides a marginal AP gain of " + }, + { + "bbox": [ + 55, + 238, + 296, + 393 + ], + "type": "inline_equation", + "content": "\\uparrow 0.3" + }, + { + "bbox": [ + 55, + 238, + 296, + 393 + ], + "type": "text", + "content": " points, likely due to its limited capacity to understand the intrinsic material properties of the objects. Conversely, our proposed material-transfer mechanism (Trans. " + }, + { + "bbox": [ + 55, + 238, + 296, + 393 + ], + "type": "inline_equation", + "content": "= M" + }, + { + "bbox": [ + 55, + 238, + 296, + 393 + ], + "type": "text", + "content": ") yields a substantial AP improvement of " + }, + { + "bbox": [ + 55, + 238, + 296, + 393 + ], + "type": "inline_equation", + "content": "\\uparrow 1.8" + }, + { + "bbox": [ + 55, + 238, + 296, + 393 + ], + "type": "text", + "content": " points, underscoring its effectiveness in bridging the X-ray modality gap." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 397, + 296, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 397, + 296, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 296, + 517 + ], + "type": "text", + "content": "Class descriptor modeling. Tab. 5 presents an ablation study analyzing the impact of the various components used to construct class representations. As shown, using a simple per-class average of all the features from each sample already results in a significant AP improvement of " + }, + { + "bbox": [ + 55, + 397, + 296, + 517 + ], + "type": "inline_equation", + "content": "\\uparrow 12.3" + }, + { + "bbox": [ + 55, + 397, + 296, + 517 + ], + "type": "text", + "content": " points over the baseline OvOD. Refining the representation by applying an object segmentation method to isolate the foreground features " + }, + { + "bbox": [ + 55, + 397, + 296, + 517 + ], + "type": "inline_equation", + "content": "(\\Omega)" + }, + { + "bbox": [ + 55, + 397, + 296, + 517 + ], + "type": "text", + "content": " further boosts the AP by " + }, + { + "bbox": [ + 55, + 397, + 296, + 517 + ], + "type": "inline_equation", + "content": "\\uparrow 1.9" + }, + { + "bbox": [ + 55, + 397, + 296, + 517 + ], + "type": "text", + "content": " points. Moreover, leveraging background features to construct negative prototypes " + }, + { + "bbox": [ + 55, + 397, + 296, + 517 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}^{\\mathrm{neg}})" + }, + { + "bbox": [ + 55, + 397, + 296, + 517 + ], + "type": "text", + "content": " adds an additional " + }, + { + "bbox": [ + 55, + 397, + 296, + 517 + ], + "type": "inline_equation", + "content": "\\uparrow 0.7" + }, + { + "bbox": [ + 55, + 397, + 296, + 517 + ], + "type": "text", + "content": " points." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 517, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 601 + ], + "type": "text", + "content": "Replacing the averaged class prototypes with our proposed visual descriptor " + }, + { + "bbox": [ + 55, + 517, + 295, + 601 + ], + "type": "inline_equation", + "content": "(\\mathcal{X})" + }, + { + "bbox": [ + 55, + 517, + 295, + 601 + ], + "type": "text", + "content": " yields another " + }, + { + "bbox": [ + 55, + 517, + 295, + 601 + ], + "type": "inline_equation", + "content": "\\uparrow 0.7" + }, + { + "bbox": [ + 55, + 517, + 295, + 601 + ], + "type": "text", + "content": " point improvement, demonstrating its enhanced capability to capture intra-class visual variability. Finally, incorporating our descriptor consistency criterion to suppress low-quality proposals " + }, + { + "bbox": [ + 55, + 517, + 295, + 601 + ], + "type": "inline_equation", + "content": "(DCC)" + }, + { + "bbox": [ + 55, + 517, + 295, + 601 + ], + "type": "text", + "content": " provides a further AP increase of " + }, + { + "bbox": [ + 55, + 517, + 295, + 601 + ], + "type": "inline_equation", + "content": "\\uparrow 8.4" + }, + { + "bbox": [ + 55, + 517, + 295, + 601 + ], + "type": "text", + "content": " points, underscoring the effectiveness of our proposed framework." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 676 + ], + "type": "text", + "content": "Hyperparameter study. Fig. 4 analyzes the effect of varying the number of samples " + }, + { + "bbox": [ + 55, + 605, + 295, + 676 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 605, + 295, + 676 + ], + "type": "text", + "content": " used to construct the visual descriptors. Even with just a single sample per class, RAXO improves AP by " + }, + { + "bbox": [ + 55, + 605, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\uparrow 6.5" + }, + { + "bbox": [ + 55, + 605, + 295, + 676 + ], + "type": "text", + "content": " points over the baseline. As " + }, + { + "bbox": [ + 55, + 605, + 295, + 676 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 605, + 295, + 676 + ], + "type": "text", + "content": " increases, performance continues to improve, reaching saturation at " + }, + { + "bbox": [ + 55, + 605, + 295, + 676 + ], + "type": "inline_equation", + "content": "K = 30" + }, + { + "bbox": [ + 55, + 605, + 295, + 676 + ], + "type": "text", + "content": " with a total AP gain of " + }, + { + "bbox": [ + 55, + 605, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\uparrow 24.0" + }, + { + "bbox": [ + 55, + 605, + 295, + 676 + ], + "type": "text", + "content": " points." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Furthermore, Fig. 5 explores the influence of the threshold " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " in the Descriptor Consistency Module. Higher values of " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " impose stricter consistency requirements, de" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 552, + 190 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 552, + 190 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 552, + 190 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 552, + 190 + ], + "type": "image", + "image_path": "4b74a760a4e378c70dc0de82ea55681b64fd98e75c5e79308d65d0aadc3210e0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 192, + 555, + 215 + ], + "lines": [ + { + "bbox": [ + 313, + 192, + 555, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 555, + 215 + ], + "type": "text", + "content": "Figure 4. Impact of " + }, + { + "bbox": [ + 313, + 192, + 555, + 215 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 192, + 555, + 215 + ], + "type": "text", + "content": " in class representations evaluated on PINray [19] using a G-DINO [18] baseline in the 100/0 setting." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 224, + 552, + 353 + ], + "blocks": [ + { + "bbox": [ + 317, + 224, + 552, + 353 + ], + "lines": [ + { + "bbox": [ + 317, + 224, + 552, + 353 + ], + "spans": [ + { + "bbox": [ + 317, + 224, + 552, + 353 + ], + "type": "image", + "image_path": "4b282f175fcce032c9c8714b863deafad0e5af70cb1264185bc000316d26dbe8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 356, + 554, + 378 + ], + "lines": [ + { + "bbox": [ + 313, + 356, + 554, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 554, + 378 + ], + "type": "text", + "content": "Figure 5. Impact of " + }, + { + "bbox": [ + 313, + 356, + 554, + 378 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 356, + 554, + 378 + ], + "type": "text", + "content": " in Descriptor Consistency Module on PIxray [19] using a G-DINO [18] baseline in the " + }, + { + "bbox": [ + 313, + 356, + 554, + 378 + ], + "type": "inline_equation", + "content": "100/0" + }, + { + "bbox": [ + 313, + 356, + 554, + 378 + ], + "type": "text", + "content": " setting." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 392, + 555, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 555, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 555, + 440 + ], + "type": "text", + "content": "manding greater separation between the predicted class and other classes before accepting a proposal. We observe that " + }, + { + "bbox": [ + 313, + 392, + 555, + 440 + ], + "type": "inline_equation", + "content": "\\sigma = 0.15" + }, + { + "bbox": [ + 313, + 392, + 555, + 440 + ], + "type": "text", + "content": " provides an optimal balance between false positives and false negatives, yielding an AP of 36.9 points." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 453, + 392, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 392, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 392, + 466 + ], + "type": "text", + "content": "7. Conclusions" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 473, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 555, + 640 + ], + "type": "text", + "content": "In this work, we pioneered the task of open-vocabulary object detection (OvOD) for X-ray imagery, a challenge shaped by the unique characteristics of X-ray data and the scarcity of annotated examples. To address this, we introduced RAXO, a training-free method that repurposes off-the-shelf RGB-based OvOD detectors for the X-ray domain. By leveraging intra-modal feature distances, a novel material-transfer mechanism, and robust class descriptor modeling, RAXO effectively bridges the modality gap between RGB and X-ray imagery. Our extensive experiments, conducted across multiple benchmarks and the newly proposed DET-COMPASS dataset, demonstrate that RAXO consistently enhances detection performance, achieving average mAP improvements of up to 17.0 points over baseline methods." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Looking ahead, the modularity of RAXO opens promising avenues for future research, including further refinement of visual descriptors and adaptation to additional modalities. Overall, our work not only sets a new state-of-the-art for X-ray open-vocabulary detection but also lays the groundwork for a thriving research direction in this emerging domain." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20777" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 157, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 157, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 157, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 297, + 115, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 115, + 310 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 115, + 310 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 317, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 317, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 317, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 61, + 317, + 294, + 360 + ], + "type": "text", + "content": "[1] Samet Akcay and Toby Breckon. Towards automatic threat detection: A survey of advances of deep learning within x-ray security imaging. Pattern Recognition, 122:108245, 2022. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 361, + 295, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 361, + 295, + 405 + ], + "spans": [ + { + "bbox": [ + 61, + 361, + 295, + 405 + ], + "type": "text", + "content": "[2] An Chang, Yu Zhang, Shunli Zhang, Leisheng Zhong, and Li Zhang. Detecting prohibited objects with physical size constraint from cluttered x-ray baggage images. Knowledge-Based Systems, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 406, + 295, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 406, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 61, + 406, + 295, + 470 + ], + "type": "text", + "content": "[3] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 472, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 472, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 62, + 472, + 294, + 514 + ], + "type": "text", + "content": "[4] Penghui Du, Yu Wang, Yifan Sun, Luting Wang, Yue Liao, Gang Zhang, Errui Ding, Yan Wang, Jingdong Wang, and Si Liu. Lami-detr: Open-vocabulary detection with language model instruction. In ECCV, 2024. 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 516, + 294, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 516, + 294, + 548 + ], + "spans": [ + { + "bbox": [ + 62, + 516, + 294, + 548 + ], + "type": "text", + "content": "[5] Ruohuan Fang, Guansong Pang, and Xiao Bai. Simple image-level classification improves open-vocabulary object detection. In AAAI, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 548, + 295, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 548, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 62, + 548, + 295, + 592 + ], + "type": "text", + "content": "[6] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated images. In ECCV, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 593, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 593, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 62, + 593, + 295, + 624 + ], + "type": "text", + "content": "[7] Junyao Gao, Yanchen Liu, Yanan Sun, Yinhao Tang, Yanhong Zeng, Kai Chen, and Cairong Zhao. Styleshot: A snapshot on any style. arXiv:2407.01414, 2024. 4, 7, 8, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 625, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 625, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 62, + 625, + 294, + 658 + ], + "type": "text", + "content": "[8] Google LLC. Custom search JSON API reference. https: // developers.google.com/custom-search/v1, 2024. Accessed: 2025-03-07. 6, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 658, + 295, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 295, + 680 + ], + "type": "text", + "content": "[9] Lewis D. Griffin, Matthew Caldwell, and Jerone T. A. Andrews. COMPASS-XP. Zenodo, 2019. 3, 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 58, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 681, + 294, + 712 + ], + "type": "text", + "content": "[10] Laurynas Karazija, Iro Laina, Andrea Vedaldi, and Christian Rupprecht. Diffusion models for open-vocabulary segmentation. In ECCV, 2024. 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 58, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 55, + 90, + 297, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 278 + ], + "type": "text", + "content": "We thank CINECA and the ISCRA initiative for the availability of high-performance computing resources. This work was partially supported by the EU HORIZON IAMI (HORIZON-CL3-2023-FCT-01-04-101168272) project, the EU HORIZON ELIAS (HORIZON-CL4-2022-HUMAN-02-101120237) project, the EU HORIZON ELLIOT (HORIZON-CL4-2024-HUMAN-03-101214398) project, the MUR PNRR FAIR (PE00000013) project funded by the NextGenerationEU, the Spanish Ministerio de Ciencia e Innovacion (grant numbers PID2020-112623GB-I00, PID2023-149549NB-I00), and the Galician Conselleria de Cultura, Educacion e Ordenacion Universitaria (2024-2027 ED431G-2023/04). Some of these grants are co-funded by the European Regional Development Fund (ERDF). Pablo Garcia-Fernandez is supported by the Spanish Ministerio de Universidades under the FPU national plan (grant number FPU21/05581)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[11] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models. arXiv:2209.15639, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 118, + 554, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 160 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 160 + ], + "type": "text", + "content": "[12] Liangqi Li, Jiaxu Miao, Dahu Shi, Wenming Tan, Ye Ren, Yi Yang, and Shiliang Pu. Distilling detr with visual-linguistic knowledge for open-vocabulary object detection. In ICCV, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "text", + "content": "[13] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, 2022. 2, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 206, + 554, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 554, + 258 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 554, + 258 + ], + "type": "text", + "content": "[14] Mingyuan Li, Tong Jia, Hao Wang, Bowen Ma, Hui Lu, Shuyang Lin, Da Cai, and Dongyue Chen. Ao-detr: Anti-overlapping detr for x-ray prohibited items detection. IEEE Transactions on Neural Networks and Learning Systems, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "type": "text", + "content": "[15] Chuang Lin, Peize Sun, Yi Jiang, Ping Luo, Lizhen Qu, Gholamreza Haffari, Zehuan Yuan, and Jianfei Cai. Learning object-language alignments for open-vocabulary object detection. In ICLR, 2023. 1, 2, 6, 7, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 305, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 348 + ], + "type": "text", + "content": "[16] Shuyang Lin, Tong Jia, Hao Wang, Bowen Ma, Mingyuan Li, and Dongyue Chen. Detection of novel prohibited item categories for real-world security inspection. Eng. Appl. Artif. Intell., 2025. 1, 2, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "text", + "content": "[17] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 394, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 448 + ], + "type": "text", + "content": "[18] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding DINO: marrying DINO with grounded pre-training for open-set object detection. In ECCV, 2024. 1, 2, 6, 7, 8, 4, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "type": "text", + "content": "[19] Bowen Ma, Tong Jia, Min Su, Xiaodong Jia, Dongyue Chen, and Yichun Zhang. Automated segmentation of prohibited items in x-ray baggage images using dense de-overlap attention snake. TMM, 2022. 2, 3, 6, 7, 8, 4, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 493, + 554, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 554, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 554, + 547 + ], + "type": "text", + "content": "[20] Bowen Ma, Tong Jia, Mingyuan Li, Songsheng Wu, Hao Wang, and Dongyue Chen. Towards dual-view x-ray baggage inspection: A large-scale benchmark and adaptive hierarchical cross refinement for prohibited item discovery. IEEE TIFS, 2024. 1, 3, 6, 7, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 548, + 554, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 554, + 590 + ], + "type": "text", + "content": "[21] Chuofan Ma, Yi Jiang, Xin Wen, Zehuan Yuan, and Xiaojuan Qi. CoDet: Co-occurrence guided region-word alignment for open-vocabulary object detection. In NeurIPS, 2023. 2, 6, 7, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "type": "text", + "content": "[22] Caijing Miao, Lingxi Xie, Fang Wan, Chi Su, Hongye Liu, Jianbin Jiao, and Qixiang Ye. Sixray: A large-scale security inspection x-ray benchmark for prohibited item discovery in overlapping images. In CVPR, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 636, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 689 + ], + "type": "text", + "content": "[23] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, et al. Simple open-vocabulary object detection. In ECCV, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 712 + ], + "type": "text", + "content": "[24] OpenAI. GPT-4 technical report. CoRR, abs/2303.08774, 2023.6" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20778" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[25] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv:2304.07193, 2023. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "text", + "content": "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 262 + ], + "type": "text", + "content": "[27] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dólar, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv:2408.00714, 2024. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 263, + 294, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 317 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 317 + ], + "type": "text", + "content": "[28] Haifeng Sima, Bailiang Chen, Chaosheng Tang, Yudong Zhang, and Junding Sun. Multi-scale feature attention-detection transformer: Multi-scale feature attention for security check object detection. IET Computer Vision, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 319, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 352 + ], + "type": "text", + "content": "[29] Archana Singh and Dhiraj. Advancements in machine learning techniques for threat item detection in x-ray images: a comprehensive survey. Int. J. Multim. Inf. Retr., 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "type": "text", + "content": "[30] Renshuai Tao, Yanlu Wei, Xiangjian Jiang, Hainan Li, Haotong Qin, Jiakai Wang, Yuqing Ma, Libo Zhang, and Xi-anglong Liu. Towards real-world x-ray security inspection: A high-quality benchmark and lateral inhibition module for prohibited items detection. In ICCV, 2021. 2, 3, 6, 7, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 410, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 410, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 294, + 464 + ], + "type": "text", + "content": "[31] Renshuai Tao, Hainan Li, Tianbo Wang, Yanlu Wei, Yifu Ding, Bowei Jin, Hongping Zhi, Xianglong Liu, and Aishan Liu. Exploring endogenous shift for cross-domain detection: A large-scale benchmark and perturbation suppression network. In CVPR, 2022. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 466, + 294, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 294, + 509 + ], + "type": "text", + "content": "[32] Renshuai Tao, Tianbo Wang, Ziyang Wu, Cong Liu, Aishan Liu, and Xianglong Liu. Few-shot x-ray prohibited item detection: A benchmark and weak-feature enhancement network. In ACMMM, 2022. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 511, + 294, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 294, + 554 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 294, + 554 + ], + "type": "text", + "content": "[33] Divya Velayudhan, Taimur Hassan, Ernesto Damiani, and Naoufel Werghi. Recent advances in baggage threat detection: A comprehensive and systematic survey. ACM Computing Surveys, 55(8):1-38, 2022. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 556, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 294, + 599 + ], + "type": "text", + "content": "[34] Boying Wang, Libo Zhang, Longyin Wen, Xianglong Liu, and Yanjun Wu. Towards real-world prohibited item detection: A large-scale x-ray benchmark. In ICCV, 2021. 2, 3, 6, 7, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 601, + 294, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 655 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 655 + ], + "type": "text", + "content": "[35] Ruxue Wang, Yuliang Shi, and Mingyu Cai. Optimization and research of suspicious object detection algorithm in x-ray image. In 2023 IEEE 6th Information Technology, Networking, Electronic and Automation Control Conference (ITNEC), 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 294, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 700 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 700 + ], + "type": "text", + "content": "[36] Yanlu Wei, Renshuai Tao, Zhangjie Wu, Yuqing Ma, Libo Zhang, and Xianglong Liu. Occluded prohibited items detection: An x-ray security inspection benchmark and de-occlusion attention module. In ACMMM, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 702, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 702, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 702, + 294, + 713 + ], + "type": "text", + "content": "[37] Size Wu, Wenwei Zhang, Sheng Jin, Wentao Liu, and" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 430 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Chen Change Loy. Aligning bag of regions for open-vocabulary object detection. In CVPR, 2023. 2, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "type": "text", + "content": "[38] Size Wu, Wenwei Zhang, Lumin Xu, Sheng Jin, Xiangtai Li, Wentao Liu, and Chen Change Loy. Clipself: Vision transformer distills itself for open-vocabulary dense prediction. In ICLR, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 553, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 175 + ], + "type": "text", + "content": "[39] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In CVPR, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 175, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 218 + ], + "type": "text", + "content": "[40] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. Detclip: Dictionary-enriched visual-concept paralleled pretraining for open-world detection. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "text", + "content": "[41] Lewei Yao, Jianhua Han, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, and Hang Xu. Detclipv2: Scalable open-vocabulary object detection pre-training via word-region alignment. In CVPR, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "text", + "content": "[42] Cairong Zhao, Liang Zhu, Shuguang Dou, Weihong Deng, and Liang Wang. Detecting overlapped objects in x-ray security imagery by a label-aware mechanism. IEEE TIFS, 2022, 2, 3, 6, 7, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 309, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 353 + ], + "type": "text", + "content": "[43] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In CVPR, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 354, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 397 + ], + "type": "text", + "content": "[44] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, 2022. 1, 2, 3, 6, 7, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 399, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 430 + ], + "type": "text", + "content": "[45] Chaoyang Zhu and Long Chen. A survey on open-vocabulary detection and segmentation: Past, present, and future. IEEE TPAMI, 2024. 3, 6" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20779" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Supervised Exploratory Learning for Long-Tailed Visual Recognition/3fb70b61-418e-4890-be11-1e54f33e6902_content_list.json b/2025/Supervised Exploratory Learning for Long-Tailed Visual Recognition/3fb70b61-418e-4890-be11-1e54f33e6902_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..31f0933dea421a4805fade2a4c25d197e2a81789 --- /dev/null +++ b/2025/Supervised Exploratory Learning for Long-Tailed Visual Recognition/3fb70b61-418e-4890-be11-1e54f33e6902_content_list.json @@ -0,0 +1,1908 @@ +[ + { + "type": "text", + "text": "Supervised Exploratory Learning for Long-Tailed Visual Recognition", + "text_level": 1, + "bbox": [ + 148, + 130, + 849, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhongquan Jian $^{1,3,\\dagger}$ , Yanhao Chen $^{2,\\dagger}$ , Yancheng Wang $^{4}$ , Junfeng Yao $^{2,3,1,*}$ , Meihong Wang $^{3,*}$ , Qingqiang Wu $^{2,3,*}$", + "bbox": [ + 202, + 179, + 792, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer and Data Science, Minjiang University, Fuzhou, China", + "bbox": [ + 197, + 215, + 799, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Film, Xiamen University, Xiamen, China", + "bbox": [ + 290, + 233, + 705, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ School of Informatics, Xiamen University, Xiamen, China", + "bbox": [ + 264, + 251, + 732, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ School of Computing and Data Science, Xiamen University Malaysia, Malaysia", + "bbox": [ + 181, + 268, + 816, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yao0010, wangmh, wuqq}@xmu.edu.cn", + "bbox": [ + 344, + 289, + 645, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 339, + 326, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Long-tailed data poses a significant challenge for deep learning models, which tend to prioritize accurate classification of head classes while largely neglecting tail classes. Existing techniques, such as class re-balancing, logit adjustment, and data augmentation, aim to enlarge decision regions of tail classes or achieve clear decision boundaries, leaving the robustness of decision regions underconsidered. This paper proposes a simple yet effective Supervised Exploratory Learning (SEL) framework to achieve these goals simultaneously from space exploration perspectives. SEL employs the adaptive Optimal Foraging Algorithm (OFA) to generate diverse exploratory examples, integrating Class-biased Complement (CbC) for balanced class distribution and Fitness-weighted Sampling (FwS) for space exploration. Both theoretical analysis and empirical results demonstrate that SEL enhances class balance, sharpens decision boundaries, and strengthens decision regions. SEL is a plug-and-play training framework that can be seamlessly integrated into model training or classifier adjustment stages, making it highly adaptable and compatible with existing methods and facilitating further performance improvements. Extensive experiments on various long-tailed benchmarks demonstrate SEL's superiority.", + "bbox": [ + 88, + 369, + 483, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 746, + 220, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the success of deep learning in various Computer Vision (CV) tasks, real-world applications of automated systems have become more prevalent, such as autonomous driving [49], medical diagnosis [11], security surveillance [14], and so on. This success is largely due to the availability of large-scale labeled datasets, such as ImageNet [30], and the development of deep learning models that can learn", + "bbox": [ + 89, + 771, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d682e41c9907346c428374f717708c4048919ba1fbd5fab3faf845311d005759.jpg", + "image_caption": [ + "Training Samples", + "(a) Existing methods." + ], + "image_footnote": [], + "bbox": [ + 519, + 349, + 707, + 433 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e6a7ec404eeb79a47528cb4c447808ee26ceac88404adfcdfd53a3ce5de24fe9.jpg", + "image_caption": [ + "Test Samples", + "Test Samples" + ], + "image_footnote": [ + "$\\bigcirc$ Test sample" + ], + "bbox": [ + 712, + 351, + 900, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f1f5bded2fef2c3b95d8429afa9d6eab4feafea71c125260c8612a6d5ba1abfc.jpg", + "image_caption": [ + "(b) Existing methods trained with SEL.", + "Figure 1. Comparison of decision regions and boundaries produced by (a) existing methods and (b) proposed SEL." + ], + "image_footnote": [ + "$\\bullet$ Training sample $\\triangle$ Exploratory example Class boun" + ], + "bbox": [ + 521, + 449, + 707, + 540 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "complex patterns from balanced data [1]. However, real-world data often exhibit long-tailed distributions with heavy class imbalance, where head classes have a massive number of samples, while tail classes are associated with only a few samples [46]. Such data imbalance poses a significant challenge for developing unbiased automated recognition systems, as deep learning models tend to be biased towards head classes and disregard tail classes, which are often the most important ones in practice [28].", + "bbox": [ + 509, + 638, + 906, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To solve this challenge, a multitude of studies have been proposed in recent years [46], including re-sampling [44, 47], logit adjustment [23, 38], data augmentation [27], classifier design [40], representation learning [22, 52], class-sensitive learning [13, 28], transfer learning [25, 29], ensemble learning [8, 39, 48] and decoupled training [17, 18, 41] methods. Despite their differences in principle, they all enhance the model's representation capabilities and de", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors; †Equal contributions.", + "bbox": [ + 107, + 886, + 357, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1870", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "velop an optimal classifier with clear decision boundaries to separate different classes. On the one hand, effective representation learning can alleviate the issue of imbalanced data distribution and create a balanced embedding space; on the other hand, robust classifiers can improve the recognition capability of tail classes.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thanks to recent advances in self-supervised learning and pre-training, foundational models now learn representations much more effectively. In the pre-trained era, imbalanced data primarily affects the classifier, where sufficient samples of head classes occupy decision space, while tail classes suffer from sparse, obscure decision regions. As a result, the potential of pre-trained foundational models is not fully released. During the inference stage, while informative representations of tail classes are obtained, they often differ from those in the training set. This discrepancy erodes the well-defined margins of the classifier [25], leading to the misclassification of samples from tail classes as head classes, as shown in Fig. 1(a). Recently, numerous methods have been developed to optimize the classifier. Representatively, BCL [52] and TSC [28] use Supervised Contrastive Learning (SCL) to tighten intra-class spaces and enlarge inter-class margins, significantly improving decision boundaries between classes. Furthermore, MixUp [45] and its variants [23, 42] are good at generating diversity samples that are beneficial for improving decision regions of tail classes. However, their goal of smoothing the decision space increases boundary ambiguity, which hinders classification since errors often occur at boundaries.", + "bbox": [ + 91, + 186, + 486, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by evolutionary algorithms like the Optimal Foraging Algorithm (OFA) [51], we attempt to tackle biased decision regions and boundaries in long-tailed recognition through a space exploration perspective. To this end, we propose Supervised Exploratory Learning (SEL), a simple yet effective framework that seamlessly integrates with existing methods to enhance their performance. SEL leverages an adaptive OFA operator to synthesize exploratory examples, compensating for insufficient tail-class data and enabling balanced decision regions and boundaries, as shown in Fig. 1(b). To be specific, SEL is composed of two modules: Class-biased Complement (CbC) and Fitness-weighted Sampling (FwS). The former is responsible for sampling more tail-class samples as target samples to balance the data distribution, while the latter selects more confident samples from adjacent classes as candidate samples to facilitate the formation of clear decision boundaries between classes. Hence, the synthesized exploratory examples are diverse and informative in representing the distribution of unseen samples in reality. SEL is a plug-and-play training framework compatible with existing long-tailed recognition methods, requiring no backbone modifications or additional parameters. The main contributions of this paper are summarized as follows:", + "bbox": [ + 91, + 537, + 486, + 902 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We attempt to address the challenge of long-tailed data from the perspective of space exploration. An adaptive OFA operator is employed to synthesize exploratory examples, which are informative and diverse to represent the distribution of unseen samples in reality.", + "- We propose a novel SEL framework that enables existing long-tailed recognition models to achieve balanced decision regions and clear decision boundaries, thereby improving their performance without altering network structure or increasing model parameters.", + "- Rational analysis proves the effectiveness of SEL in improving the classification of tail classes. Considerable performance improvements of existing long-tailed recognition methods verify the effectiveness of SEL." + ], + "bbox": [ + 513, + 90, + 906, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 319, + 653, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Data Augmentation (DA) [50] is an effective method in improving model robustness, particularly in scenarios of data scarcity or imbalance. In this section, we review recent advances in DA for long-tailed recognition, categorized into explicit and implicit methods.", + "bbox": [ + 511, + 345, + 906, + 421 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Explicit Data Augmentation", + "text_level": 1, + "bbox": [ + 513, + 433, + 764, + 449 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional DA techniques, like flipping, rotating, cropping, and padding, effectively improve robustness on long-tailed datasets and mitigate overfitting by creating new image-label pairs consistent with the training data [9, 34]. More recently, Population-Based Augmentation [10], AutoAugment [2], and Randaugment [3] have emerged to automatically choose optimal DA strategies according to characteristics of datasets, which have demonstrated their efficacy in long-tailed data [4, 21, 33]. Similarly, Wang et al. [37] generated images for tail classes using encoder variation information learned from head classes. Zada et al. [43] utilizes pure noise images as samples for tail classes.", + "bbox": [ + 511, + 455, + 905, + 637 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although generating new images is an effective data augmentation method, it is time-consuming and computationally expensive. Moreover, the diversity of the generated images is limited, as they are created through simple transformations of the original training data.", + "bbox": [ + 511, + 638, + 903, + 714 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Implicit Data Augmentation", + "text_level": 1, + "bbox": [ + 513, + 726, + 764, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Different from explicit DA, implicit DA generates new samples by interpolating existing samples in the feature space, which can effectively improve the model's generalization ability. Representatively, MixUp [45] introduced complexity control to unexplored regions in the data space by linearly interpolating discrete sample points, effectively reducing the generalization error of the model. Subsequently, CutMix [42], an extension of MixUp, augmented data by inserting a patch from one image into another, with labels blended according to the patch areas. This method can", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1871", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "transfer context-rich background information from dominant classes to less frequent ones by combining foreground patches from the latter with background images from the former [32]. Manifold MixUp [36] further developed this concept by applying linear interpolation to the features of input images, which proved to be more effective for long-tail learning. MetaSAug [26] learns class covariance matrices from a small, balanced validation set to augment tail classes semantically. Recently, H2T [25] improved diversity by replacing part of the tail class feature map with that of the head classes.", + "bbox": [ + 89, + 90, + 483, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "These findings highlight the superiority of implicit DA methods in enhancing model generalization on long-tailed data. Following this line, we propose a novel implicit DA technique, from the perspective of space exploration, to augment the diversity of training data, especially for tail classes, and improve the robustness of the long-tailed recognition model's performance.", + "bbox": [ + 89, + 257, + 483, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 378, + 225, + 396 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Definition", + "text_level": 1, + "bbox": [ + 89, + 404, + 202, + 419 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For image classification, $\\{x,y\\}$ represents a sample consisting of an image $x$ and its corresponding label $y\\in \\mathcal{V}$ , where $\\mathcal{V} = \\{1,2,\\dots,C\\}$ is the set of classes. In the long-tailed scenario, different classes have varying sample sizes. The imbalance ratio $\\rho = \\frac{n_{\\mathrm{max}}}{n_{\\mathrm{min}}}$ measures the skewness of the data distribution, where $n_{\\mathrm{max}}$ and $n_{\\mathrm{min}}$ represent the maximum and minimum number of training samples across classes. Typically, training samples for head classes are significantly greater than those for tail classes.", + "bbox": [ + 89, + 426, + 483, + 563 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing long-tailed recognition methods usually learn a complex function mapping from the input space $\\mathcal{X}$ to the target space $\\mathcal{Y}$ . This function is usually implemented as the composition of an encoder $f: \\mathcal{X} \\to \\mathcal{Z} \\in \\mathbb{R}^d$ and a linear classifier $w: \\mathcal{Z} \\to \\mathcal{Y}$ . Both are crucial for the final classification accuracy [46].", + "bbox": [ + 89, + 564, + 483, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Motivation", + "text_level": 1, + "bbox": [ + 89, + 666, + 212, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Firstly, the poor performance of the tail classes is primarily due to their lack of training samples. A straightforward solution is to increase the number of tail class samples through methods such as re-sampling [44] or to emphasize their importance through re-weighting [47]. Although these methods are effective, they do not generate new, diverse samples for tail classes different from the training data, limiting the enhancement of generalization. Secondly, MixUp [45] and its variants [23, 36, 42] are good at generating diverse samples that are beneficial to improving the decision regions of the tail classes. However, on the one hand, head-class samples are more likely to be randomly selected to construct mixed samples, worsening the imbalance of training data. On the other hand, their goal of smoothing the de", + "bbox": [ + 89, + 688, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/97d63d4038a2da45930095677aa2a619d1ffe2dee1d69c5b20acc64174a9a0c7.jpg", + "image_caption": [ + "(a) MixUp" + ], + "image_footnote": [], + "bbox": [ + 519, + 90, + 666, + 157 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bcb92c6d87e4e3b7b1e2d1496346d32a03f02a9f3f27ef471ebf97eabc85d040.jpg", + "image_caption": [ + "(b) OFA", + "Figure 2. The differences and relationships among MixUp, OFA and SEL. MixUp linearly interpolates two random samples (typically from different classes, such as $z_{a}$ and $z_{c}$ ) to create a new sample, while OFA leverages the better individual $(z_{s'})$ to guide weaker individuals $(z_{s})$ toward better evolution. Combining their strengths, SEL gradually expands and reinforces minority decision regions by synthesizing adjacent exploratory examples." + ], + "image_footnote": [], + "bbox": [ + 666, + 90, + 750, + 157 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a02a9029b7da64dbdab100a1e32cb259f7075caef87d57e3fa37e7f4a726b97f.jpg", + "image_caption": [ + "(c) SEL" + ], + "image_footnote": [], + "bbox": [ + 753, + 90, + 897, + 157 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cision space intensifies the boundary ambiguity, as shown in Fig. 2(a), where the mixed sample $z_{e}$ may fall into another class's region. Finally, methods like BCL [52] and TSC [28] use SCL [19] to tighten intra-class spaces and expand inter-class margins, significantly improving decision boundaries but reducing decision region robustness while increasing false positives risk.", + "bbox": [ + 511, + 306, + 905, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address these issues simultaneously, we propose to synthesize exploratory examples to enhance the diversity of the training data and improve the robustness of the decision regions and boundaries. As shown in Fig. 2(b), OFA [51] is good at exploring the sample space by leveraging the better individual (evaluated by the task-specific fitness function) to guide weaker ones, promoting more effective evolution and optimization. Combining the strengths of MixUp and OFA, SEL views samples of the same class as a population and expands its decision region by synthesizing exploratory examples in uncharted spaces, as illustrated in Fig. 2(c).", + "bbox": [ + 511, + 412, + 905, + 578 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that MixUp is an intuitive and effective enhancement method applied at the input layer [7]. However, due to varying model structures and input patterns, unifying these methods is challenging. Consequently, most enhanced training frameworks [22, 25, 36] operate in the feature space, as most classification models first derive feature representations before classifying images. Similarly, SEL is also operated in the feature space, making it compatible with various models and methods.", + "bbox": [ + 511, + 579, + 905, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Supervised Exploratory Learning", + "text_level": 1, + "bbox": [ + 511, + 727, + 808, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SEL is a compatible framework that can be seamlessly integrated into existing methods. In SEL, an adaptive OFA operator is employed to synthesize exploratory examples to supplement and balance the original training samples.", + "bbox": [ + 511, + 750, + 905, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3.1. Exploratory Example Synthesization", + "text_level": 1, + "bbox": [ + 511, + 820, + 815, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by the space exploration process of evolutionary algorithms, we propose to synthesize exploratory examples using an adaptive OFA operator. The OFA [51], a metaheuristic algorithm introduced in 2017, is characterized by", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1872", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "its simple structure, strong global search efficiency, and proven affine invariance [15]. It simulates animals' foraging behavior, where each individual follows better ones to explore the environment. In OFA and its variants, the operator is defined in dimension. Here, we extend its formulation to a vector operation. Specifically, we treat samples in the batch as populations, and synthesize exploratory examples by using the following formula:", + "bbox": [ + 89, + 90, + 483, + 212 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {e} = z _ {s} + r \\left(z _ {s ^ {\\prime}} - z _ {s}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 219, + 482, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $z_{s} \\in \\mathcal{Z}$ and $z_{s'} \\in \\mathcal{Z}$ denote the target and candidate samples, respectively. $r \\in (0,1)$ is a coefficient that controls the exploration distance, which is randomly sampled during model training. Generally, OFA selects the candidate sample $z_{s'}$ with higher fitness to guide weaker individuals ( $z_{s}$ ) toward better evolution. To adapt OFA to the requirement of long-tailed recognition, we introduce two modules, CbC and FwS, to regulate the selection of target and candidate samples, respectively.", + "bbox": [ + 89, + 239, + 483, + 377 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2. Class-biased Complement (CbC)", + "text_level": 1, + "bbox": [ + 89, + 381, + 366, + 396 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To alleviate the class imbalance problem, diverse exploratory examples are synthesized to complement the tail classes. We first count the number of samples in each class and then identify the additional samples required (i.e., exploratory examples) to balance the class distribution. As illustrated in Fig. 2(c), the target sample's label is assigned to the exploratory example, eliminating the decision boundary ambiguity issue present in MixUp. Hence, synthesizing varying numbers of exploratory examples for different classes becomes a task of selecting target samples. In practice, training samples with the target class are allowed to be selected as the target samples multiple times. Specifically, for a sample $z_{s}$ in the batch, it will be utilized as the target sample $k_{c}$ times if it belongs to class $c$ .", + "bbox": [ + 89, + 400, + 483, + 611 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nk _ {c} = \\operatorname {C e i l} \\left(\\frac {n _ {\\max } - n _ {c}}{n _ {c}}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 618, + 480, + 646 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $n_c$ denotes the sample number of class $c$ , and $n_{max}$ represents the sample size of the largest head class. $\\text{Ceil}(\\ast)$ is the ceiling function.", + "bbox": [ + 89, + 650, + 482, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this way, the issue of class imbalance is addressed by synthesizing exploratory examples with different classes, ensuring that the number of samples in each class is balanced. The key challenges are guaranteeing the diversity of synthesized exploratory examples and the rationality of assigned labels, a crucial difference that distinguishes SEL from MixUp and its variants. As shown in Fig. 3, our solution is to identify neighboring classes and place exploratory examples in the regions between the target and adjacent classes, ensuring rational label assignment.", + "bbox": [ + 89, + 696, + 483, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.3. Neighboring Class Recognition", + "text_level": 1, + "bbox": [ + 89, + 851, + 352, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In MixUp and its variants, the label of the synthesized sample is determined by linear interpolation of the target and", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9d6540a18a942a52d8410d52e6c54ddcd393f8933f800cf372c471452129a3b0.jpg", + "image_caption": [ + "Figure 3. Synthesize exploratory examples by the adaptive OFA operator. More exploratory examples are generated for tail classes." + ], + "image_footnote": [], + "bbox": [ + 519, + 87, + 901, + 186 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "candidate samples. This approach results in label misassignment if the synthesized sample falls within the decision region of another class. To avoid this situation, we propose selecting candidate samples from neighboring classes, and placing synthesized samples between the target and adjacent class regions. Specifically, for samples in the batch $\\mathcal{B}$ , we calculate the center of each class and then determine the Euclidean distances between the target and the others.", + "bbox": [ + 511, + 250, + 906, + 371 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd \\left(c, c ^ {\\prime}\\right) = \\| \\frac {1}{\\left| \\mathcal {B} ^ {c} \\right|} \\sum_ {z _ {i} \\in \\mathcal {B}} ^ {| \\mathcal {B} |} z _ {i} \\cdot \\mathbb {I} _ {y _ {i} = c} - \\frac {1}{\\left| \\mathcal {B} ^ {c ^ {\\prime}} \\right|} \\sum_ {z _ {i} \\in \\mathcal {B}} ^ {| \\mathcal {B} |} z _ {i} \\cdot \\mathbb {I} _ {y _ {i} = c ^ {\\prime}} \\| _ {2} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 380, + 903, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{B}^c$ and $\\mathcal{B}^{c'}$ are sets of samples belong to class $c$ and $c'$ in the batch $\\mathcal{B}$ , respectively. $\\mathbb{I}_*$ is the indicator function, and $\\| * \\|_2$ denotes the L2 norm. Therefore, $d(c, c')$ denotes the distance between the centers of class $c$ and class $c'$ . The nearest $k_c'$ classes are chosen as neighboring classes for class $c$ , from which a neighbor is randomly selected, such as class $c'$ , and the candidate sample will be picked from $\\mathcal{B}^{c'}$ based on their fitnesses, as done in OFA.", + "bbox": [ + 511, + 436, + 906, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.4. Fitness-weighted Sampling (FwS)", + "text_level": 1, + "bbox": [ + 511, + 566, + 792, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following the principle of OFA, the prediction probability of the correct class is considered the fitness of the sample, and the sample with higher fitness is selected as the candidate sample. To reduce the operation complexity and increase the exploration randomness, we employ probability-based sampling to select the candidate sample. Specifically, the sampling probabilities are obtained by:", + "bbox": [ + 511, + 584, + 905, + 691 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {s ^ {\\prime}} = \\operatorname {N o r m} \\left(\\left\\{\\hat {y} _ {i} ^ {c ^ {\\prime}} \\right\\} _ {i = 1} ^ {\\left| \\mathcal {B} ^ {c ^ {\\prime}} \\right|}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 700, + 903, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c'$ represents the selected neighboring class, $\\mathcal{B}^{c'}$ is the set of samples belong to class $c'$ in the batch $\\mathcal{B}$ . $\\hat{y}_i^{c'}$ is the predicted probability of the $i$ -th sample belonging to its correct class $c'$ , and $\\text{Norm}(*)$ is the normalization function that ensures the sum of $p_{s'}$ equals 1.", + "bbox": [ + 511, + 733, + 905, + 810 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this way, for each target sample $z_{s}$ , a candidate sample $z_{s'}$ with higher confidence is sampled from the batch, thereby achieving exploratory examples with diversity. FwS is essential for the validity of SEL, which is why we use the adaptive OFA operator to synthesize exploratory examples. The rational analysis is elaborated later.", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1873", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.5. Training with SEL Framework", + "text_level": 1, + "bbox": [ + 89, + 90, + 354, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SEL is a plug-and-play training framework that can be seamlessly integrated into the training process of the original model. The optimization objective of SEL is to correctly classify the exploratory examples, which is equivalent to maximizing the predicted probability of the correct class.", + "bbox": [ + 89, + 109, + 483, + 185 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S E L} = - \\frac {1}{C} \\sum_ {c = 1} ^ {C} \\left(\\frac {1}{k _ {c}} \\sum_ {i = 1} ^ {k _ {c}} \\log \\left(\\hat {y} _ {i} ^ {c}\\right)\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 191, + 482, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{y}_i^c$ denotes the predicted probability of the $i$ -th exploratory example belonging to the $c$ -th class, and $k_c$ is the number of exploratory examples synthesized for class $c$ .", + "bbox": [ + 89, + 239, + 483, + 285 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the overall optimization objective is defined as:", + "bbox": [ + 109, + 285, + 477, + 300 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = (1 - \\lambda) \\mathcal {L} _ {\\text {T a s k}} + \\lambda \\mathcal {L} _ {\\text {S E L}} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 306, + 482, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{Task}$ is the optimization objective of the original method, and $\\lambda$ is a weighting coefficient that controls the impact of SEL loss.", + "bbox": [ + 89, + 330, + 483, + 377 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Rationale Analysis", + "text_level": 1, + "bbox": [ + 89, + 383, + 271, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The exploitation process of OFA has proved to be effective and rational [51]. In this section, we provide a rational analysis of SEL. For clarity, we denote the weights of the linear classifier as $w = \\{w_{c}\\}_{c = 1}^{C}$ , where $C$ is the number of classes, and $w_{c} \\in \\mathbb{R}^{d \\times 1}$ denotes the weights of class $c$ , $d$ is the feature dimension. For a sample $z \\in \\mathbb{R}^{d \\times 1}$ , the predicted probability of class $c$ can be expressed as:", + "bbox": [ + 89, + 406, + 483, + 512 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} ^ {c} = \\frac {e ^ {w _ {c} ^ {T} z}}{\\sum_ {c ^ {\\prime} = 1} ^ {C} e ^ {w _ {c ^ {\\prime}} ^ {T} z}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 517, + 482, + 558 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hence, $\\hat{y}^c\\propto w_c^T z$ . The classifier attempts to assign a higher probability to the target class, i.e., $\\hat{y}^c >\\hat{y}^{c'}$ , where $c$ and $c^\\prime$ denote the target and other classes, respectively.", + "bbox": [ + 89, + 564, + 483, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Without loss of generality, the selection of the target sample $(z_{s})$ and candidate sample $(z_{s}^{\\prime})$ can be divided into two cases, as summarized in Tab. 1. We use the subscripts $t$ (tail) and $h$ (head) to represent their relative relation. Generally, the first case typically dominates, as the CbC module selects more tail-class targets, while the long-tailed distribution increases the chance of candidate samples being from head classes. Additionally, since exploratory examples share labels with targets, the optimization aims to challenge the model's bias toward head classes.", + "bbox": [ + 89, + 609, + 483, + 760 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the first case, the goal of $\\hat{y}_e^t >\\hat{y}_e^h$ is equivalent to:", + "bbox": [ + 109, + 760, + 467, + 777 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {t} ^ {T} z _ {e} - w _ {h} ^ {T} z _ {e} > 0 \\Rightarrow \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 781, + 482, + 806 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {t} ^ {T} \\left(z _ {t} + r \\left(z _ {h} - z _ {t}\\right)\\right) - w _ {h} ^ {T} \\left(z _ {t} + r \\left(z _ {h} - z _ {t}\\right)\\right) > 0\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 803, + 439, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After expanding and transforming the inequality, we have:", + "bbox": [ + 89, + 828, + 477, + 843 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} w _ {t} ^ {T} z _ {t} - w _ {h} ^ {T} z _ {t} > \\frac {r}{1 - r} \\left(w _ {h} ^ {T} z _ {h} - w _ {t} ^ {T} z _ {h}\\right) \\tag {9} \\\\ \\Rightarrow \\hat {y} _ {t} ^ {t} - \\hat {y} _ {t} ^ {h} > \\frac {r}{1 - r} (\\hat {y} _ {h} ^ {h} - \\hat {y} _ {h} ^ {t}) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 848, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5bbc068156c4c08ed5395c2ea76faf73f4b4dd9717ca5016857f34fb13ee8616.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Casezszs'zeGoal
1ztzhzt + r(zh - zt)ˆy e t >ˆy e h
2zhztzh + r(zt - zh)ˆy e h >ˆy e t
", + "bbox": [ + 519, + 89, + 905, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Two types of synthesized exploratory examples.", + "bbox": [ + 539, + 172, + 877, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{y}_h^h$ and $\\hat{y}_h^t$ represent probabilities of the candidate sample $(z_h)$ belonging to the head class and tail classes, respectively. The classifier favors head classes $(\\hat{y}_h^h - \\hat{y}_h^t > 0)$ , while the FwS module reinforces this by selecting high-confidence candidates. Therefore, the right item is tend to be positive, and $\\hat{y}_t^t > \\hat{y}_t^h$ is achieved, meaning that optimizing tail-class exploratory examples effectively improves the classification accuracy of original tail-class samples.", + "bbox": [ + 511, + 212, + 903, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarly, for the second case, the optimization goal of SEL is to make $\\hat{y}_e^h >\\hat{y}_e^t$ yielding:", + "bbox": [ + 511, + 334, + 905, + 364 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} _ {h} ^ {h} - \\hat {y} _ {h} ^ {t} > \\frac {r}{1 - r} \\left(\\hat {y} _ {t} ^ {t} - \\hat {y} _ {t} ^ {h}\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 375, + 903, + 401 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{y}_t^t$ and $\\hat{y}_t^h$ represent probabilities of the candidate sample $(z_t)$ belonging to the tail class and head class, respectively. Although the classifier tends to misclassify tail-class samples as head classes, the FwS mitigates this by selecting better candidates, which possess higher probabilities in the correct classes. Hence, optimizing the head-class exploratory examples does not cause significant damage to the original head sample.", + "bbox": [ + 511, + 412, + 905, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 547, + 645, + 564 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Dataset", + "text_level": 1, + "bbox": [ + 511, + 571, + 609, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CIFAR-10/100-LT [5] are the subsets of CIFAR-10/100 [20]. Both contain 50,000 training images and 10,000 validation images, each of size $32 \\times 32$ , with 10 and 100 classes, respectively. The imbalance factors used in the experiment are set to 100, 50, and 10.", + "bbox": [ + 511, + 593, + 905, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ImageNet-LT [30] is a long-tailed version of ImageNet [6] by sampling a subset following the Pareto distribution with power value $\\alpha = 0.64$ . It consists of $115.8\\mathrm{K}$ images of 1000 classes in total, with 1280 to 5 images per class.", + "bbox": [ + 511, + 669, + 905, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "iNaturalist 2018 [12] is a large-scale dataset containing 437.5K images from 8,142 classes. It is long-tailed by nature, with an extremely imbalanced distribution.", + "bbox": [ + 511, + 729, + 905, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Basic Setting", + "text_level": 1, + "bbox": [ + 511, + 784, + 650, + 800 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1. Compared Methods", + "text_level": 1, + "bbox": [ + 511, + 806, + 697, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of our proposed SEL, we integrated it into several existing methods introduced in recent years, including re-balanced techniques (IB [31], GCL [23], KPS [24]), contrastive learning (BCL [52]), feature compression (FCC [22]), and data augmentation strategies", + "bbox": [ + 511, + 825, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1874", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5b1d042552082832fcbed615fd8691bd343e9befa1abebca57f06d65db027233.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsCIFAR-100-LTCIFAR-10-LT
AvenueImbalance Ratio ρ10050101005010
-ResNet-32 [9]39.9646.0156.6671.9675.9585.92
+ MixUp [45]41.55↑1.5947.11↑1.1057.18↑0.5273.43↑1.4777.19↑1.2486.37↑0.45
+ SEL44.53↑4.5749.64↑3.8559.17↑2.5175.84↑3.8879.81↑3.8686.78↑0.86
ICCV 2021IB [31]40.5046.3156.7575.1179.9588.01
+ SEL42.34↑1.8447.95↑1.6457.57↑0.8275.04↓0.0779.81↓0.1488.13↑0.12
CVPR 2022BCL [52]51.7656.5167.9083.6186.1390.10
+ SEL52.30↑0.5457.25↑0.7468.43↑0.5384.44↑0.8386.31↑0.1890.26↑0.16
CVPR 2022GCL [23]46.5051.7261.7980.5684.7489.65
+ SEL47.89↑1.3952.74↑1.0262.41↑0.6281.86↑1.3085.13↑0.3989.78↑0.13
CVPR 2023FCC+CE [22]40.2045.9357.8073.8079.5787.75
+ SEL42.33↑2.1348.05↑2.1258.75↑0.9577.88↑4.0882.10↑2.5388.80↑1.05
CVPR 2023GLMC [7]53.9158.8768.0783.6886.9091.16
+ SEL56.48↑2.5761.13↑2.2670.75↑2.6885.40↑1.7288.57↑1.6792.83↑1.67
TPAMI 2023KPS [24]41.9747.9259.5982.3284.2989.10
+ SEL44.01↑2.0447.80↓0.1260.03↑0.4483.48↑1.1684.49↑0.2089.34↑0.24
AAAI 2024H2T+CE [25]42.2747.5858.2479.9182.8088.77
+ SEL42.45↑0.1847.80↑0.2258.66↑0.4280.43↑0.5282.80↓0.0088.79↑0.02
", + "bbox": [ + 109, + 88, + 887, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparisons between raw methods and their SEL-enhanced counterparts on CIFAR-LT datasets.", + "bbox": [ + 184, + 452, + 810, + 467 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(H2T [25], GLMC [7]). These methods were chosen not only for their representativeness but also for their publicly available code and configurations, ensuring ease of reproducibility. Most employ a two-stage training strategy, with the first stage dedicated to representation learning and the second to classification. Additionally, we use ResNet models as baselines to directly compare the performance improvements achieved by MixUp and SEL.", + "bbox": [ + 89, + 493, + 482, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.2. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 625, + 302, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We reproduce the compared methods using their official codes and configurations, setting the random seed to 2024 for fair comparison. The base ResNet models are configured identically to those in Li et al. [23]. To simplify application, we first determine the optimal way to apply SEL on different datasets by using the base model, i.e., ResNet-32 for CIFAR-10/100-LT and ResNet-50 for others, then replace it with corresponding models to assess SEL's effectiveness on these models. Specifically, for CIFAR-100-LT, SEL is applied during the final 20 training epochs. For other datasets, SEL is applied by further fine-tuning the classifier for 20 epochs on CIFAR-10-LT and 10 epochs on ImageNet-LT and iNaturalist 2018. The weight coefficient $\\lambda$ in Eq. (6) is set to 0.9, with the number of neighboring classes $k_c'$ set to 2, 5, 10, and 50 for CIFAR-10-LT, CIFAR-100-LT, ImageNet-LT, and iNaturalist 2018, respectively. Top-1 accuracy serves as the evaluation metric.", + "bbox": [ + 89, + 643, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Improvement of Existing Methods", + "text_level": 1, + "bbox": [ + 511, + 493, + 812, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.1. Results on CIFAR-10/100-LT", + "text_level": 1, + "bbox": [ + 511, + 516, + 761, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For CIFAR-10/100-LT, the results with and without SEL are presented in Tab. 2. Across 48 experiment groups, our proposed SEL significantly improves 44 of them, with an average increase of $1.37\\%$ (4.57% max and $0.02\\%$ min). Extensive experimental results fully verify the effectiveness of our method on long-tailed recognition, particularly for scenarios with high imbalance ratios. Generally, integrating MixUp into the vanilla ResNet-32 model significantly enhances performance, highlighting the effectiveness of data augmentation. However, our proposed SEL achieves even greater improvements across all scenarios, and even surpasses existing methods. For example, ResNet-32+SEL outperforms IB on both datasets except for CIFAR-10-LT-10, and surpasses GCL, KPS, and H2T on CIFAR-100-LT, though it underperforms on CIFAR-10-LT.", + "bbox": [ + 509, + 536, + 906, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, SEL significantly enhances the performance of existing methods, with particularly pronounced improvements in high-imbalance scenarios. For example, SEL is notably effective on KPS at an imbalance ratio of 100 but shows limited impact at other imbalance ratios. In addition, GLMC+SEL and FCC+SEL significantly outperform their original counterparts on both datasets, with average improvements of $2.10\\%$ and $2.14\\%$ , respectively. For the re-balanced methods IB, GCL, and KPS, SEL yields", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1875", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5ea8bc60f24e497370a209bb2ef91ebd9fd4d8329d75eb829399a89192adb6cf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodManyMediumFewAll
ResNet-5063.3833.7426.0444.80
+ Mixup62.9735.4428.5845.71↑0.91
+ SEL64.1438.4131.2747.86↑3.06
BCL65.8353.1636.2755.67
+ SEL65.4554.8236.5356.27↑0.60
GCL65.0452.7735.6955.51
+ SEL65.1353.9439.1856.12↑0.61
FCC+CE65.5252.1833.8354.68
+ SEL64.7454.0035.6655.60↑0.92
GLMC69.4252.4930.4156.17
+ SEL68.6754.3738.2857.24↑1.07
", + "bbox": [ + 99, + 88, + 475, + 311 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3dc1d0f5a9227fff87e1df30552a7f96d94f56b7f40b1b23e9b7a716da5a435b.jpg", + "table_caption": [ + "Table 3. Comparisons on the ImageNet-LT dataset." + ], + "table_footnote": [], + "table_body": "
MethodManyMediumFewAll
ResNet-5074.8968.2461.4265.95
+ Mixup71.3469.5363.8867.29↑1.34
+ SEL73.0470.3866.7469.10↑3.15
BCL73.4771.9968.8270.80
+ SEL71.9973.0469.2171.32↑0.52
GCL73.7572.2868.3470.76
+ SEL72.6373.5870.0572.00↑1.24
FCC+CE72.7871.0967.4169.70
+ SEL71.1772.4169.5571.09↑1.39
GLMC76.0473.6971.7773.14
+ SEL74.6875.1473.6674.51↑1.37
", + "bbox": [ + 99, + 351, + 475, + 577 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "greater improvements on CIFAR-100-LT than on CIFAR-10-LT, demonstrating its superiority in scenarios with more classes. However, SEL is less effective on BCL and H2T. BCL uses contrastive learning to reinforce balanced decision regions, aligning with SEL's objective. The difference is that SEL further enhances the robustness of decision regions, resulting in additional improvements. H2T is also an evolutionary algorithm-inspired method that generates diverse samples through cross-mutation, following the principles of Genetic Algorithm (GA), thus indicating the fundamental difference between SEL and H2T stems from the distinction between OFA and GA. The slight improvement of H2T+SEL further validates SEL's effectiveness in generating diverse samples, aligning with H2T's principles.", + "bbox": [ + 89, + 630, + 482, + 843 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.2. Results on ImageNet-LT and iNaturalist 2018", + "text_level": 1, + "bbox": [ + 89, + 849, + 454, + 866 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For ImageNet-LT and iNaturalist 2018, experimental results are listed in Tab. 3 and Tab. 4, respectively. Following pre", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a896ff6d025599d29191620f9ea697c8c300214195da8fc8bd6be734f949be78.jpg", + "table_caption": [ + "Table 4. Comparisons on the iNaturalist 2018 dataset." + ], + "table_footnote": [], + "table_body": "
CbCFwSTop-1 Accuracy
ρ =100ρ =10
××39.9656.66
×✓40.26↑0.3056.93↑0.27
✓×43.44↑3.4857.84↑1.18
✓✓44.53↑4.5759.17↑2.51
", + "bbox": [ + 545, + 88, + 874, + 200 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Ablations of key modules.", + "bbox": [ + 602, + 212, + 813, + 226 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "vious works [7, 16], also report accuracy across three class splits: Many $(n_c \\geq 100)$ , Medium $(20 < n_c < 100)$ , and Few $(n_c \\leq 20)$ . Overall, SEL outperforms MixUp on both datasets, and consistently enhances the compared methods, achieving an average improvement of $1.25\\%$ (max $3.06\\%$ , min $0.61\\%$ ) on ImageNet-LT and $1.53\\%$ (max $3.15\\%$ , min $0.52\\%$ ) on iNaturalist 2018. As seen, training with SEL typically improves the performance of Few classes while slightly reducing that of Many classes, aligning with the rationale proved in Sec. 3.4. However, the greater gain in Few classes outweighs the slight decline in Many classes, leading to an overall improvement in the All classes.", + "bbox": [ + 511, + 253, + 906, + 435 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Further Analysis", + "text_level": 1, + "bbox": [ + 511, + 445, + 679, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further analyze the principles of SEL quantitatively, we conducted a series of ablation studies on the CIFAR-10/100-LT datasets using ResNet-32 [9] as the base model.", + "bbox": [ + 511, + 467, + 905, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.1. Visualization of Decision Regions", + "text_level": 1, + "bbox": [ + 511, + 521, + 790, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 4 visualizes the decision regions and boundaries of different classes using t-SNE [35] across various methods. These models are trained on the CIFAR-100-LT with an imbalance ratio of 100 and tested on a balanced test set. For comparison, we selected four classes from CIFAR-100 test set for visualization: where apple and aquarium fish are the head classes with the most training samples, and willow tree and wolf are the tail classes with the fewest training samples. The decision boundaries are derived using the SVM algorithm. From Fig. 4, we can observe that SEL results in more balanced decision regions for both head and tail classes, aligning with our motivation.", + "bbox": [ + 511, + 540, + 906, + 722 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.2. Role of Key Modules", + "text_level": 1, + "bbox": [ + 511, + 729, + 705, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The CbC and FwS modules are the core components of SEL. To examine their roles, we conduct ablation experiments on CIFAR-100-LT with imbalance ratios of 100 and 10. Tab. 5 presented the experimental results, showing that the CbC module significantly enhances the baseline method, with improvements of $3.48\\%$ $(39.96\\% \\rightarrow 43.44\\%)$ and $1.18\\%$ $(56.66\\% \\rightarrow 57.84\\%)$ for imbalance ratios of 100 and 10, respectively. These results highlight the CbC module's crucial role in balancing data distribution, enabling the model to learn more effectively from tail classes. However,", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1876", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/716691e4944b523ec5b595f6875a736a441d44aa5fe965579e1542e1fccd8a35.jpg", + "image_caption": [ + "(a) CE" + ], + "image_footnote": [ + "apple aquarium fish willow tree wolf" + ], + "bbox": [ + 102, + 87, + 367, + 223 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8eab3c4522ffdf4793ba371a8356e68ec5cba34e59fe4c0ccc9c3880429927fd.jpg", + "image_caption": [ + "(b) MixUp" + ], + "image_footnote": [], + "bbox": [ + 367, + 88, + 630, + 223 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8d700e333e0d6188682ac844bd4b2f8788258555d41a092de5028b5a3bb9c4f6.jpg", + "image_caption": [ + "(c) SEL" + ], + "image_footnote": [], + "bbox": [ + 630, + 88, + 893, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1b4246f60e2705712b128da47d2aa8a22f9e635de2fec3633c1840ca93c4884d.jpg", + "image_caption": [ + "Figure 4. Visualization of decision regions and boundaries on the CIFAR-100 test set (each class contains 100 test samples), where apple and aquarium fish are head classes with the most training samples, and willow tree and wolf are tail classes with the fewest training samples.", + "(a) CIFAR-100-LT" + ], + "image_footnote": [], + "bbox": [ + 91, + 316, + 284, + 402 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4c2c44f2c6f7026de5a3a7d334556998300dedc4211bfc5d1bd98e4bcfb48945.jpg", + "image_caption": [ + "(b) CIFAR-10-LT" + ], + "image_footnote": [], + "bbox": [ + 290, + 319, + 483, + 402 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9f4076fa1f5e32e733a4046011f0c2730e0b25bc3987b2e6576f6ba263458504.jpg", + "image_caption": [ + "Figure 5. Changes with different neighboring class numbers.", + "(a) start epochs", + "Figure 6. Changes with different parameters." + ], + "image_footnote": [], + "bbox": [ + 91, + 460, + 287, + 549 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ba8e8d79274170b70c7a7f9f52d3b7bb01bb264004f60b7fee9f941e8db10d9f.jpg", + "image_caption": [ + "(b) weight $(\\lambda)$" + ], + "image_footnote": [], + "bbox": [ + 292, + 460, + 483, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the FwS module has a minimal impact without the CbC module, yielding only $0.30\\%$ and $0.27\\%$ improvements for imbalance ratios of 100 and 10, respectively. Its effectiveness is fully realized when combined with the CbC module, leading to a significant boost in model performance.", + "bbox": [ + 89, + 622, + 483, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.3. Impact of different parameters", + "text_level": 1, + "bbox": [ + 89, + 713, + 352, + 728 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we conduct exploratory experiments on the impact of the number of neighboring classes $(k_c^{\\prime})$ in the FwS module. Experimental results are summarized in Fig. 5, where $k_{c}^{\\prime}$ is varied from 1 to 10 and relative accuracy (values are adjusted by subtracting their mean) is utilized for comparison. As seen, the model performs better with a higher $k^{\\prime}_{c}$ on CIFAR-100-LT but with a lower $k^{\\prime}_{c}$ on CIFAR-10LT. This difference arises from the number of classes, as CIFAR-100-LT has 100 classes, whereas CIFAR-10-LT has only 10. Based on these observations, we empirically set $k^{\\prime}_{c}$ to 5 for CIFAR-100-LT and 2 for CIFAR-10-LT.", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.4. When to Start SEL Training", + "text_level": 1, + "bbox": [ + 511, + 320, + 758, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we explore the optimal starting point for SEL training to maximize computational efficiency, as its synthesis of additional exploratory examples increases computational costs. Specifically, we integrate SEL into the training process at different epochs, ranging from 100 to 200, as shown in Fig. 6(a). Here, starting at 200 epochs indicates a two-stage training mode, where SEL is applied with an additional 20 epochs after the original training process. Similarly, relative accuracy is used as the ordinate for clearly showing the parameter's impact. It is obvious that the model's performance steadily improves as the SEL starting point increases, with most datasets peaking at 180 epochs, except for CIFAR-10-LT-100 and CIFAR-10-LT-10. Therefore, we set the starting point to 180 epochs for CIFAR-100-LT and 200 epochs for CIFAR-10-LT.", + "bbox": [ + 511, + 340, + 906, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Building on the optimal starting points, we further examine the impact of parameter $\\lambda$ in Eq. (6). Fig. 6(b) illustrates the performance variation with $\\lambda$ ranges from 0.0 to 1.0. Note that $\\lambda = 0.0$ indicates the absence of SEL, while $\\lambda = 1.0$ signifies its full integration. For most datasets, as $\\lambda$ increases, the model's performance gradually improves, peaking at $\\lambda = 0.9$ before declining. However, for CIFAR-10-LT-10, the simplest dataset, performance remains relatively stable, with a peak at $\\lambda = 0.6$ .", + "bbox": [ + 511, + 569, + 908, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 722, + 633, + 738 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose SEL, a simple yet effective training framework that addresses long-tailed challenges through the space exploration perspective. SEL is a plug-and-play method that boosts performance without modifying the model or adding parameters. It uses an adaptive OFA operator to generate exploratory examples that capture unseen tail-class features, enhancing decision regions and boundaries. The effectiveness of SEL is validated mathematically through rational analysis and empirically through experiments on various long-tailed datasets and methods.", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1877", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the Solfeggio ear training intelligent robot and cloud platform research and development project for music education (No.2024CXY0102), the 3D visualization digital twin integrated control system (No.2023CXY0111), and the public technology service platform project of Xiamen City (No.3502Z20231043).", + "bbox": [ + 89, + 114, + 480, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 220, + 187, + 236 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yoshua Bengio, Yann Lecun, and Geoffrey Hinton. Deep learning for ai. Communications of the ACM, 64(7):58-65, 2021. 1", + "[2] Ekin D. Cubuk, Barret Zoph, Dandelion Mané, Vijay Vasudevan, and Quoc V. Le. Autoaugment: Learning augmentation strategies from data. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 113-123, 2019. 2", + "[3] Ekin Dogus Cubuk, Barret Zoph, Jon Shlens, and Quoc Le. Randaugment: Practical automated data augmentation with a reduced search space. In Advances in Neural Information Processing Systems, pages 18613-18624. Curran Associates, Inc., 2020. 2", + "[4] Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, and Jiaya Jia. Parametric contrastive learning. In 2021 IEEE/CVF International Conference on Computer Vision, pages 695-704, 2021. 2", + "[5] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge J. Belongie. Class-balanced loss based on effective number of samples. In IEEE Conference on Computer Vision and Pattern Recognition, pages 9268-9277, 2019. 5", + "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 5", + "[7] Fei Du, Peng Yang, Qi Jia, Fengtao Nan, Xiaoting Chen, and Yun Yang. Global and local mixture consistency cumulative learning for long-tailed visual recognitions. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15814-15823, 2023. 3, 6, 7", + "[8] Hao Guo and Song Wang. Long-tailed multi-label visual recognition by collaborative training on uniform and rebalanced samplings. In IEEE Conference on Computer Vision and Pattern Recognition, pages 15089-15098, 2021. 1", + "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2, 6, 7", + "[10] Daniel Ho, Eric Liang, Xi Chen, Ion Stoica, and Pieter Abbeel. Population based augmentation: Efficient learning of augmentation policy schedules. In Proceedings of the 36th International Conference on Machine Learning, pages 2731-2741, 2019. 2", + "[11] Gregory Holste, Yiliang Zhou, Song Wang, Ajay Jaiswal, Mingquan Lin, Sherry Zhuge, Yuzhe Yang, Dongkyun Kim," + ], + "bbox": [ + 93, + 246, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Trong-Hieu Nguyen-Mau, Minh-Triet Tran, Jaehyup Jeong, Wongi Park, Jongbin Ryu, Feng Hong, Arsh Verma, Yosuke Yamagishi, Changhyun Kim, Hyeryeong Seo, Myungjoo Kang, Leo Anthony Celi, Zhiyong Lu, Ronald M. Summers, George Shih, Zhangyang Wang, and Yifan Peng. Towards long-tailed, multi-label disease classification from chest x-ray: Overview of the cxr-It challenge. Medical Image Analysis, 97:103224, 2024. 1", + "[12] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alexander Shepard, Hartwig Adam, Pietro Perona, and Serge J. Belongie. The inaturalist species classification and detection dataset. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, pages 8769-8778, 2018. 5", + "[13] Chengkai Hou, Jieyu Zhang, Haonan Wang, and Tianyi Zhou. Subclass-balancing contrastive learning for long-tailed recognition. In 2023 IEEE/CVF International Conference on Computer Vision, pages 5372-5384, 2023. 1", + "[14] Shuyan Hu, Wei Ni, Xin Wang, and Abbas Jamalipour. Disguised tailing and video surveillance with solar-powered fixed-wing unmanned aerial vehicle. IEEE Transactions on Vehicular Technology, 71(5):5507-5518, 2022. 1", + "[15] ZhongQuan Jian and GuangYu Zhu. Affine invariance of meta-heuristic algorithms. Information Sciences, 576:37-53, 2021. 4", + "[16] Yan Jin, Mengke Li, Yang Lu, Yiu-ming Cheung, and Hanzi Wang. Long-tailed visual recognition via self-heterogeneous integration with knowledge excavation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23695-23704, 2023. 7", + "[17] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. In 8th International Conference on Learning Representations, 2020. 1", + "[18] Bingyi Kang, Yu Li, Sa Xie, Zehuan Yuan, and Jiashi Feng. Exploring balanced feature spaces for representation learning. In 9th International Conference on Learning Representations, 2021. 1", + "[19] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673, 2020. 3", + "[20] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5", + "[21] Jun Li, Zichang Tan, Jun Wan, Zhen Lei, and Guodong Guo. Nested collaborative learning for long-tailed visual recognition. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6939-6948, 2022. 2", + "[22] Jian Li, Ziyao Meng, Daqian Shi, Rui Song, Xiaolei Dao, Jingwen Wang, and Hao Xu. FCC: feature clusters compression for long-tailed visual recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24080-24089, 2023. 1, 3, 5, 6", + "[23] Mengke Li, Yiu-Ming Cheung, and Yang Lu. Long-tailed visual recognition via gaussian clouded logit adjustment. In" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1878", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6919-6928, 2022. 1, 2, 3, 5, 6", + "[24] Mengke Li, Yiu-Ming Cheung, and Zhikai Hu. Key point sensitive loss for long-tailed visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4): 4812-4825, 2023. 5, 6", + "[25] Mengke Li, Zhikai Hu, Yang Lu, Weichao Lan, Yiu-ming Cheung, and Hui Huang. Feature fusion from head to tail for long-tailed visual recognition. In Thirty-Eighth AAAI Conference on Artificial Intelligence, Thirty-Sixth Conference on Innovative Applications of Artificial Intelligence, Fourteenth Symposium on Educational Advances in Artificial Intelligence, pages 13581-13589, 2024. 1, 2, 3, 6", + "[26] Shuang Li, Kaixiong Gong, Chi Harold Liu, Yulin Wang, Feng Qiao, and Xinjing Cheng. Metasaug: Meta semantic augmentation for long-tailed visual recognition. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5212-5221, 2021. 3", + "[27] Shuang Li, Kaixiong Gong, Chi Harold Liu, Yulin Wang, Feng Qiao, and Xinjing Cheng. Metasaug: Meta semantic augmentation for long-tailed visual recognition. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5212-5221, 2021. 1", + "[28] Tianhong Li, Peng Cao, Yuan Yuan, Lijie Fan, Yuzhe Yang, Rogerio Feris, Piotr Indyk, and Dina Katabi. Targeted supervised contrastive learning for long-tailed recognition. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6918, 2022. 1, 2, 3", + "[29] Bo Liu, Haoxiang Li, Hao Kang, Gang Hua, and Nuno Vasconcelos. Gistnet: a geometric structure transfer network for long-tailed recognition. In 2021 IEEE/CVF International Conference on Computer Vision, pages 8189-8198, 2021. 1", + "[30] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X. Yu. Large-scale long-tailed recognition in an open world. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2537-2546, 2019. 1, 5", + "[31] Seulki Park, Jongin Lim, Younghan Jeon, and Jin Young Choi. Influence-balanced loss for imbalanced visual classification. In 2021 IEEE/CVF International Conference on Computer Vision, pages 715-724, 2021. 5, 6", + "[32] Seulki Park, Youngkyu Hong, Byeongho Heo, Sangdoo Yun, and Jin Young Choi. The majority can help the minority: Context-rich minority oversampling for long-tailed classification. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6877-6886, 2022. 3", + "[33] Jiawei Ren, Cunjun Yu, Shunan Sheng, Xiao Ma, Haiyu Zhao, Shuai Yi, and Hongsheng Li. Balanced meta-softmax for long-tailed visual recognition. In Advances in Neural Information Processing Systems 33, 2020. 2", + "[34] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott E. Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In 2015 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-9, 2015. 2", + "[35] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 7" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Vikas Verma, Alex Lamb, Christopher Beckham, Amir Najafi, Ioannis Mitlagkas, David Lopez-Paz, and Yoshua Bengio. Manifold mixup: Better representations by interpolating hidden states. In Proceedings of the 36th International Conference on Machine Learning, pages 6438-6447, 2019. 3", + "[37] Jianfeng Wang, Thomas Lukasiewicz, Xiaolin Hu, Jianfei Cai, and Zhenghua Xu. Rsg: A simple but effective module for learning imbalanced datasets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3784-3793, 2021. 2", + "[38] Peng Wang, Kai Han, Xiu-Shen Wei, Lei Zhang, and Lei Wang. Contrastive learning based hybrid networks for long-tailed image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 943-952, 2021. 1", + "[39] Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu, and Stella X. Yu. Long-tailed recognition by routing diverse distribution-aware experts. In 9th International Conference on Learning Representations, 2021. 1", + "[40] Tong Wu, Ziwei Liu, Qingqiu Huang, Yu Wang, and Dahua Lin. Adversarial robustness under long-tailed distribution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 8659-8668, 2021. 1", + "[41] Shiyu Xuan and Shiliang Zhang. Decoupled contrastive learning for long-tailed recognition. In Thirty-Eighth AAAI Conference on Artificial Intelligence, Thirty-Sixth Conference on Innovative Applications of Artificial Intelligence, Fourteenth Symposium on Educational Advances in Artificial Intelligence, pages 6396-6403, 2024. 1", + "[42] Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Seong Joon Oh, Youngjoon Yoo, and Junsuk Choe. Cutmix: Regularization strategy to train strong classifiers with localizable features. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6022-6031, 2019. 2, 3", + "[43] Shiran Zada, Itay Benou, and Michal Irani. Pure noise to the rescue of insufficient data: Improving imbalanced classification by training on random noise images. In International Conference on Machine Learning, pages 25817-25833, 2022. 2", + "[44] Yuhang Zang, Chen Huang, and Chen Change Loy. Fasa: Feature augmentation and sampling adaptation for long-tailed instance segmentation. In 2021 IEEE/CVF International Conference on Computer Vision, pages 3437-3446, 2021. 1, 3", + "[45] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In 6th International Conference on Learning Representations, 2018. 2, 3, 6", + "[46] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):10795-10816, 2023. 1, 3", + "[47] Zizhao Zhang and Tomas Pfister. Learning fast sample reweighting without reward data. In 2021 IEEE/CVF International Conference on Computer Vision, pages 705-714, 2021. 1, 3", + "[48] Boyan Zhou, Quan Cui, Xiu-Shen Wei, and Zhao-Min Chen. BBN: bilateral-branch network with cumulative learning for" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1879", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "long-tailed visual recognition. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9716-9725, 2020. 1", + "[49] Weitao Zhou, Zhong Cao, Nanshan Deng, Xiaoyu Liu, Kun Jiang, and Diange Yang. Dynamically conservative self-driving planner for long-tail cases. IEEE Transactions on Intelligent Transportation Systems, 24(3):3476-3488, 2023. 1", + "[50] Chenguang Zhu, Yichong Xu, Xiang Ren, Bill Yuchen Lin, Meng Jiang, and Wenhao Yu. Knowledge-augmented methods for natural language processing. In Proceedings of the Sixteenth ACM International Conference on Web Search and Data Mining, pages 1228-1231, 2023. 2", + "[51] Guang-Yu Zhu and Wei-Bo Zhang. Optimal foraging algorithm for global optimization. Applied Soft Computing, 51: 294–313, 2017. 2, 3, 5", + "[52] Jianggang Zhu, Zheng Wang, Jingjing Chen, Yi-Ping Phoebe Chen, and Yu-Gang Jiang. Balanced contrastive learning for long-tailed visual recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6898-6907, 2022. 1, 2, 3, 5, 6" + ], + "bbox": [ + 91, + 90, + 482, + 387 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1880", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Supervised Exploratory Learning for Long-Tailed Visual Recognition/3fb70b61-418e-4890-be11-1e54f33e6902_model.json b/2025/Supervised Exploratory Learning for Long-Tailed Visual Recognition/3fb70b61-418e-4890-be11-1e54f33e6902_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8c86a1630b08ee2331723ad87eafd4b6871613a6 --- /dev/null +++ b/2025/Supervised Exploratory Learning for Long-Tailed Visual Recognition/3fb70b61-418e-4890-be11-1e54f33e6902_model.json @@ -0,0 +1,2653 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.149, + 0.131, + 0.85, + 0.154 + ], + "angle": 0, + "content": "Supervised Exploratory Learning for Long-Tailed Visual Recognition" + }, + { + "type": "text", + "bbox": [ + 0.204, + 0.18, + 0.794, + 0.217 + ], + "angle": 0, + "content": "Zhongquan Jian\\(^{1,3,\\dagger}\\), Yanhao Chen\\(^{2,\\dagger}\\), Yancheng Wang\\(^{4}\\), Junfeng Yao\\(^{2,3,1,*}\\), Meihong Wang\\(^{3,*}\\), Qingqiang Wu\\(^{2,3,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.217, + 0.8, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer and Data Science, Minjiang University, Fuzhou, China" + }, + { + "type": "text", + "bbox": [ + 0.291, + 0.234, + 0.707, + 0.251 + ], + "angle": 0, + "content": "\\(^{2}\\)School of Film, Xiamen University, Xiamen, China" + }, + { + "type": "text", + "bbox": [ + 0.265, + 0.252, + 0.733, + 0.269 + ], + "angle": 0, + "content": "\\(^{3}\\)School of Informatics, Xiamen University, Xiamen, China" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.27, + 0.817, + 0.288 + ], + "angle": 0, + "content": "\\(^{4}\\)School of Computing and Data Science, Xiamen University Malaysia, Malaysia" + }, + { + "type": "text", + "bbox": [ + 0.346, + 0.29, + 0.646, + 0.305 + ], + "angle": 0, + "content": "{yao0010, wangmh, wuqq}@xmu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.34, + 0.327, + 0.354 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.371, + 0.485, + 0.719 + ], + "angle": 0, + "content": "Long-tailed data poses a significant challenge for deep learning models, which tend to prioritize accurate classification of head classes while largely neglecting tail classes. Existing techniques, such as class re-balancing, logit adjustment, and data augmentation, aim to enlarge decision regions of tail classes or achieve clear decision boundaries, leaving the robustness of decision regions underconsidered. This paper proposes a simple yet effective Supervised Exploratory Learning (SEL) framework to achieve these goals simultaneously from space exploration perspectives. SEL employs the adaptive Optimal Foraging Algorithm (OFA) to generate diverse exploratory examples, integrating Class-biased Complement (CbC) for balanced class distribution and Fitness-weighted Sampling (FwS) for space exploration. Both theoretical analysis and empirical results demonstrate that SEL enhances class balance, sharpens decision boundaries, and strengthens decision regions. SEL is a plug-and-play training framework that can be seamlessly integrated into model training or classifier adjustment stages, making it highly adaptable and compatible with existing methods and facilitating further performance improvements. Extensive experiments on various long-tailed benchmarks demonstrate SEL's superiority." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.747, + 0.222, + 0.763 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.772, + 0.483, + 0.879 + ], + "angle": 0, + "content": "With the success of deep learning in various Computer Vision (CV) tasks, real-world applications of automated systems have become more prevalent, such as autonomous driving [49], medical diagnosis [11], security surveillance [14], and so on. This success is largely due to the availability of large-scale labeled datasets, such as ImageNet [30], and the development of deep learning models that can learn" + }, + { + "type": "image_caption", + "bbox": [ + 0.578, + 0.342, + 0.65, + 0.351 + ], + "angle": 0, + "content": "Training Samples" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.351, + 0.709, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.66, + 0.435, + 0.761, + 0.446 + ], + "angle": 0, + "content": "(a) Existing methods." + }, + { + "type": "image_caption", + "bbox": [ + 0.78, + 0.342, + 0.834, + 0.351 + ], + "angle": 0, + "content": "Test Samples" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.352, + 0.901, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.778, + 0.45, + 0.83, + 0.458 + ], + "angle": 0, + "content": "Test Samples" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.45, + 0.709, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.619, + 0.542, + 0.801, + 0.553 + ], + "angle": 0, + "content": "(b) Existing methods trained with SEL." + }, + { + "type": "image_footnote", + "bbox": [ + 0.522, + 0.555, + 0.8, + 0.567 + ], + "angle": 0, + "content": "\\(\\bullet\\) Training sample \\(\\triangle\\) Exploratory example Class boun" + }, + { + "type": "image_footnote", + "bbox": [ + 0.833, + 0.555, + 0.899, + 0.566 + ], + "angle": 0, + "content": "\\(\\bigcirc\\) Test sample" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.578, + 0.907, + 0.607 + ], + "angle": 0, + "content": "Figure 1. Comparison of decision regions and boundaries produced by (a) existing methods and (b) proposed SEL." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.64, + 0.907, + 0.777 + ], + "angle": 0, + "content": "complex patterns from balanced data [1]. However, real-world data often exhibit long-tailed distributions with heavy class imbalance, where head classes have a massive number of samples, while tail classes are associated with only a few samples [46]. Such data imbalance poses a significant challenge for developing unbiased automated recognition systems, as deep learning models tend to be biased towards head classes and disregard tail classes, which are often the most important ones in practice [28]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "To solve this challenge, a multitude of studies have been proposed in recent years [46], including re-sampling [44, 47], logit adjustment [23, 38], data augmentation [27], classifier design [40], representation learning [22, 52], class-sensitive learning [13, 28], transfer learning [25, 29], ensemble learning [8, 39, 48] and decoupled training [17, 18, 41] methods. Despite their differences in principle, they all enhance the model's representation capabilities and de" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.358, + 0.901 + ], + "angle": 0, + "content": "*Corresponding authors; †Equal contributions." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1870" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "velop an optimal classifier with clear decision boundaries to separate different classes. On the one hand, effective representation learning can alleviate the issue of imbalanced data distribution and create a balanced embedding space; on the other hand, robust classifiers can improve the recognition capability of tail classes." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.187, + 0.487, + 0.534 + ], + "angle": 0, + "content": "Thanks to recent advances in self-supervised learning and pre-training, foundational models now learn representations much more effectively. In the pre-trained era, imbalanced data primarily affects the classifier, where sufficient samples of head classes occupy decision space, while tail classes suffer from sparse, obscure decision regions. As a result, the potential of pre-trained foundational models is not fully released. During the inference stage, while informative representations of tail classes are obtained, they often differ from those in the training set. This discrepancy erodes the well-defined margins of the classifier [25], leading to the misclassification of samples from tail classes as head classes, as shown in Fig. 1(a). Recently, numerous methods have been developed to optimize the classifier. Representatively, BCL [52] and TSC [28] use Supervised Contrastive Learning (SCL) to tighten intra-class spaces and enlarge inter-class margins, significantly improving decision boundaries between classes. Furthermore, MixUp [45] and its variants [23, 42] are good at generating diversity samples that are beneficial for improving decision regions of tail classes. However, their goal of smoothing the decision space increases boundary ambiguity, which hinders classification since errors often occur at boundaries." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.539, + 0.487, + 0.903 + ], + "angle": 0, + "content": "Inspired by evolutionary algorithms like the Optimal Foraging Algorithm (OFA) [51], we attempt to tackle biased decision regions and boundaries in long-tailed recognition through a space exploration perspective. To this end, we propose Supervised Exploratory Learning (SEL), a simple yet effective framework that seamlessly integrates with existing methods to enhance their performance. SEL leverages an adaptive OFA operator to synthesize exploratory examples, compensating for insufficient tail-class data and enabling balanced decision regions and boundaries, as shown in Fig. 1(b). To be specific, SEL is composed of two modules: Class-biased Complement (CbC) and Fitness-weighted Sampling (FwS). The former is responsible for sampling more tail-class samples as target samples to balance the data distribution, while the latter selects more confident samples from adjacent classes as candidate samples to facilitate the formation of clear decision boundaries between classes. Hence, the synthesized exploratory examples are diverse and informative in representing the distribution of unseen samples in reality. SEL is a plug-and-play training framework compatible with existing long-tailed recognition methods, requiring no backbone modifications or additional parameters. The main contributions of this paper are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.905, + 0.167 + ], + "angle": 0, + "content": "- We attempt to address the challenge of long-tailed data from the perspective of space exploration. An adaptive OFA operator is employed to synthesize exploratory examples, which are informative and diverse to represent the distribution of unseen samples in reality." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.168, + 0.907, + 0.243 + ], + "angle": 0, + "content": "- We propose a novel SEL framework that enables existing long-tailed recognition models to achieve balanced decision regions and clear decision boundaries, thereby improving their performance without altering network structure or increasing model parameters." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.243, + 0.905, + 0.303 + ], + "angle": 0, + "content": "- Rational analysis proves the effectiveness of SEL in improving the classification of tail classes. Considerable performance improvements of existing long-tailed recognition methods verify the effectiveness of SEL." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.092, + 0.907, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.32, + 0.655, + 0.336 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.346, + 0.907, + 0.422 + ], + "angle": 0, + "content": "Data Augmentation (DA) [50] is an effective method in improving model robustness, particularly in scenarios of data scarcity or imbalance. In this section, we review recent advances in DA for long-tailed recognition, categorized into explicit and implicit methods." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.434, + 0.766, + 0.45 + ], + "angle": 0, + "content": "2.1. Explicit Data Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.457, + 0.906, + 0.638 + ], + "angle": 0, + "content": "Traditional DA techniques, like flipping, rotating, cropping, and padding, effectively improve robustness on long-tailed datasets and mitigate overfitting by creating new image-label pairs consistent with the training data [9, 34]. More recently, Population-Based Augmentation [10], AutoAugment [2], and Randaugment [3] have emerged to automatically choose optimal DA strategies according to characteristics of datasets, which have demonstrated their efficacy in long-tailed data [4, 21, 33]. Similarly, Wang et al. [37] generated images for tail classes using encoder variation information learned from head classes. Zada et al. [43] utilizes pure noise images as samples for tail classes." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.64, + 0.905, + 0.715 + ], + "angle": 0, + "content": "Although generating new images is an effective data augmentation method, it is time-consuming and computationally expensive. Moreover, the diversity of the generated images is limited, as they are created through simple transformations of the original training data." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.727, + 0.766, + 0.743 + ], + "angle": 0, + "content": "2.2. Implicit Data Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Different from explicit DA, implicit DA generates new samples by interpolating existing samples in the feature space, which can effectively improve the model's generalization ability. Representatively, MixUp [45] introduced complexity control to unexplored regions in the data space by linearly interpolating discrete sample points, effectively reducing the generalization error of the model. Subsequently, CutMix [42], an extension of MixUp, augmented data by inserting a patch from one image into another, with labels blended according to the patch areas. This method can" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1871" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.257 + ], + "angle": 0, + "content": "transfer context-rich background information from dominant classes to less frequent ones by combining foreground patches from the latter with background images from the former [32]. Manifold MixUp [36] further developed this concept by applying linear interpolation to the features of input images, which proved to be more effective for long-tail learning. MetaSAug [26] learns class covariance matrices from a small, balanced validation set to augment tail classes semantically. Recently, H2T [25] improved diversity by replacing part of the tail class feature map with that of the head classes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.258, + 0.484, + 0.366 + ], + "angle": 0, + "content": "These findings highlight the superiority of implicit DA methods in enhancing model generalization on long-tailed data. Following this line, we propose a novel implicit DA technique, from the perspective of space exploration, to augment the diversity of training data, especially for tail classes, and improve the robustness of the long-tailed recognition model's performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.38, + 0.226, + 0.397 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.405, + 0.204, + 0.42 + ], + "angle": 0, + "content": "3.1. Definition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.428, + 0.484, + 0.564 + ], + "angle": 0, + "content": "For image classification, \\(\\{x,y\\}\\) represents a sample consisting of an image \\(x\\) and its corresponding label \\(y\\in \\mathcal{V}\\), where \\(\\mathcal{V} = \\{1,2,\\dots,C\\}\\) is the set of classes. In the long-tailed scenario, different classes have varying sample sizes. The imbalance ratio \\(\\rho = \\frac{n_{\\mathrm{max}}}{n_{\\mathrm{min}}}\\) measures the skewness of the data distribution, where \\(n_{\\mathrm{max}}\\) and \\(n_{\\mathrm{min}}\\) represent the maximum and minimum number of training samples across classes. Typically, training samples for head classes are significantly greater than those for tail classes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.484, + 0.656 + ], + "angle": 0, + "content": "Existing long-tailed recognition methods usually learn a complex function mapping from the input space \\(\\mathcal{X}\\) to the target space \\(\\mathcal{Y}\\). This function is usually implemented as the composition of an encoder \\(f: \\mathcal{X} \\to \\mathcal{Z} \\in \\mathbb{R}^d\\) and a linear classifier \\(w: \\mathcal{Z} \\to \\mathcal{Y}\\). Both are crucial for the final classification accuracy [46]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.667, + 0.214, + 0.682 + ], + "angle": 0, + "content": "3.2. Motivation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Firstly, the poor performance of the tail classes is primarily due to their lack of training samples. A straightforward solution is to increase the number of tail class samples through methods such as re-sampling [44] or to emphasize their importance through re-weighting [47]. Although these methods are effective, they do not generate new, diverse samples for tail classes different from the training data, limiting the enhancement of generalization. Secondly, MixUp [45] and its variants [23, 36, 42] are good at generating diverse samples that are beneficial to improving the decision regions of the tail classes. However, on the one hand, head-class samples are more likely to be randomly selected to construct mixed samples, worsening the imbalance of training data. On the other hand, their goal of smoothing the de" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.091, + 0.667, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.16, + 0.617, + 0.17 + ], + "angle": 0, + "content": "(a) MixUp" + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.091, + 0.751, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.691, + 0.16, + 0.728, + 0.17 + ], + "angle": 0, + "content": "(b) OFA" + }, + { + "type": "image", + "bbox": [ + 0.754, + 0.091, + 0.898, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.807, + 0.159, + 0.843, + 0.169 + ], + "angle": 0, + "content": "(c) SEL" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.181, + 0.907, + 0.279 + ], + "angle": 0, + "content": "Figure 2. The differences and relationships among MixUp, OFA and SEL. MixUp linearly interpolates two random samples (typically from different classes, such as \\( z_{a} \\) and \\( z_{c} \\)) to create a new sample, while OFA leverages the better individual \\( (z_{s'}) \\) to guide weaker individuals \\( (z_{s}) \\) toward better evolution. Combining their strengths, SEL gradually expands and reinforces minority decision regions by synthesizing adjacent exploratory examples." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.307, + 0.906, + 0.412 + ], + "angle": 0, + "content": "cision space intensifies the boundary ambiguity, as shown in Fig. 2(a), where the mixed sample \\( z_{e} \\) may fall into another class's region. Finally, methods like BCL [52] and TSC [28] use SCL [19] to tighten intra-class spaces and expand inter-class margins, significantly improving decision boundaries but reducing decision region robustness while increasing false positives risk." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.413, + 0.906, + 0.579 + ], + "angle": 0, + "content": "To address these issues simultaneously, we propose to synthesize exploratory examples to enhance the diversity of the training data and improve the robustness of the decision regions and boundaries. As shown in Fig. 2(b), OFA [51] is good at exploring the sample space by leveraging the better individual (evaluated by the task-specific fitness function) to guide weaker ones, promoting more effective evolution and optimization. Combining the strengths of MixUp and OFA, SEL views samples of the same class as a population and expands its decision region by synthesizing exploratory examples in uncharted spaces, as illustrated in Fig. 2(c)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.906, + 0.715 + ], + "angle": 0, + "content": "Note that MixUp is an intuitive and effective enhancement method applied at the input layer [7]. However, due to varying model structures and input patterns, unifying these methods is challenging. Consequently, most enhanced training frameworks [22, 25, 36] operate in the feature space, as most classification models first derive feature representations before classifying images. Similarly, SEL is also operated in the feature space, making it compatible with various models and methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.728, + 0.809, + 0.744 + ], + "angle": 0, + "content": "3.3. Supervised Exploratory Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.812 + ], + "angle": 0, + "content": "SEL is a compatible framework that can be seamlessly integrated into existing methods. In SEL, an adaptive OFA operator is employed to synthesize exploratory examples to supplement and balance the original training samples." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.821, + 0.816, + 0.836 + ], + "angle": 0, + "content": "3.3.1. Exploratory Example Synthesization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Inspired by the space exploration process of evolutionary algorithms, we propose to synthesize exploratory examples using an adaptive OFA operator. The OFA [51], a metaheuristic algorithm introduced in 2017, is characterized by" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1872" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.213 + ], + "angle": 0, + "content": "its simple structure, strong global search efficiency, and proven affine invariance [15]. It simulates animals' foraging behavior, where each individual follows better ones to explore the environment. In OFA and its variants, the operator is defined in dimension. Here, we extend its formulation to a vector operation. Specifically, we treat samples in the batch as populations, and synthesize exploratory examples by using the following formula:" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.22, + 0.483, + 0.235 + ], + "angle": 0, + "content": "\\[\nz _ {e} = z _ {s} + r \\left(z _ {s ^ {\\prime}} - z _ {s}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.241, + 0.484, + 0.378 + ], + "angle": 0, + "content": "where \\( z_{s} \\in \\mathcal{Z} \\) and \\( z_{s'} \\in \\mathcal{Z} \\) denote the target and candidate samples, respectively. \\( r \\in (0,1) \\) is a coefficient that controls the exploration distance, which is randomly sampled during model training. Generally, OFA selects the candidate sample \\( z_{s'} \\) with higher fitness to guide weaker individuals (\\( z_{s} \\)) toward better evolution. To adapt OFA to the requirement of long-tailed recognition, we introduce two modules, CbC and FwS, to regulate the selection of target and candidate samples, respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.382, + 0.367, + 0.397 + ], + "angle": 0, + "content": "3.3.2. Class-biased Complement (CbC)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.401, + 0.484, + 0.612 + ], + "angle": 0, + "content": "To alleviate the class imbalance problem, diverse exploratory examples are synthesized to complement the tail classes. We first count the number of samples in each class and then identify the additional samples required (i.e., exploratory examples) to balance the class distribution. As illustrated in Fig. 2(c), the target sample's label is assigned to the exploratory example, eliminating the decision boundary ambiguity issue present in MixUp. Hence, synthesizing varying numbers of exploratory examples for different classes becomes a task of selecting target samples. In practice, training samples with the target class are allowed to be selected as the target samples multiple times. Specifically, for a sample \\( z_{s} \\) in the batch, it will be utilized as the target sample \\( k_{c} \\) times if it belongs to class \\( c \\)." + }, + { + "type": "equation", + "bbox": [ + 0.207, + 0.619, + 0.482, + 0.647 + ], + "angle": 0, + "content": "\\[\nk _ {c} = \\operatorname {C e i l} \\left(\\frac {n _ {\\max } - n _ {c}}{n _ {c}}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.651, + 0.483, + 0.696 + ], + "angle": 0, + "content": "where \\( n_c \\) denotes the sample number of class \\( c \\), and \\( n_{max} \\) represents the sample size of the largest head class. \\( \\text{Ceil}(\\ast) \\) is the ceiling function." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.697, + 0.484, + 0.848 + ], + "angle": 0, + "content": "In this way, the issue of class imbalance is addressed by synthesizing exploratory examples with different classes, ensuring that the number of samples in each class is balanced. The key challenges are guaranteeing the diversity of synthesized exploratory examples and the rationality of assigned labels, a crucial difference that distinguishes SEL from MixUp and its variants. As shown in Fig. 3, our solution is to identify neighboring classes and place exploratory examples in the regions between the target and adjacent classes, ensuring rational label assignment." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.852, + 0.353, + 0.867 + ], + "angle": 0, + "content": "3.3.3. Neighboring Class Recognition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In MixUp and its variants, the label of the synthesized sample is determined by linear interpolation of the target and" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.088, + 0.903, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.197, + 0.906, + 0.226 + ], + "angle": 0, + "content": "Figure 3. Synthesize exploratory examples by the adaptive OFA operator. More exploratory examples are generated for tail classes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.251, + 0.907, + 0.372 + ], + "angle": 0, + "content": "candidate samples. This approach results in label misassignment if the synthesized sample falls within the decision region of another class. To avoid this situation, we propose selecting candidate samples from neighboring classes, and placing synthesized samples between the target and adjacent class regions. Specifically, for samples in the batch \\(\\mathcal{B}\\), we calculate the center of each class and then determine the Euclidean distances between the target and the others." + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.381, + 0.905, + 0.437 + ], + "angle": 0, + "content": "\\[\nd \\left(c, c ^ {\\prime}\\right) = \\| \\frac {1}{\\left| \\mathcal {B} ^ {c} \\right|} \\sum_ {z _ {i} \\in \\mathcal {B}} ^ {| \\mathcal {B} |} z _ {i} \\cdot \\mathbb {I} _ {y _ {i} = c} - \\frac {1}{\\left| \\mathcal {B} ^ {c ^ {\\prime}} \\right|} \\sum_ {z _ {i} \\in \\mathcal {B}} ^ {| \\mathcal {B} |} z _ {i} \\cdot \\mathbb {I} _ {y _ {i} = c ^ {\\prime}} \\| _ {2} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.438, + 0.907, + 0.558 + ], + "angle": 0, + "content": "where \\(\\mathcal{B}^c\\) and \\(\\mathcal{B}^{c'}\\) are sets of samples belong to class \\(c\\) and \\(c'\\) in the batch \\(\\mathcal{B}\\), respectively. \\(\\mathbb{I}_*\\) is the indicator function, and \\(\\| * \\|_2\\) denotes the L2 norm. Therefore, \\(d(c, c')\\) denotes the distance between the centers of class \\(c\\) and class \\(c'\\). The nearest \\(k_c'\\) classes are chosen as neighboring classes for class \\(c\\), from which a neighbor is randomly selected, such as class \\(c'\\), and the candidate sample will be picked from \\(\\mathcal{B}^{c'}\\) based on their fitnesses, as done in OFA." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.567, + 0.794, + 0.582 + ], + "angle": 0, + "content": "3.3.4. Fitness-weighted Sampling (FwS)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.906, + 0.692 + ], + "angle": 0, + "content": "Following the principle of OFA, the prediction probability of the correct class is considered the fitness of the sample, and the sample with higher fitness is selected as the candidate sample. To reduce the operation complexity and increase the exploration randomness, we employ probability-based sampling to select the candidate sample. Specifically, the sampling probabilities are obtained by:" + }, + { + "type": "equation", + "bbox": [ + 0.628, + 0.701, + 0.905, + 0.726 + ], + "angle": 0, + "content": "\\[\np _ {s ^ {\\prime}} = \\operatorname {N o r m} \\left(\\left\\{\\hat {y} _ {i} ^ {c ^ {\\prime}} \\right\\} _ {i = 1} ^ {\\left| \\mathcal {B} ^ {c ^ {\\prime}} \\right|}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.906, + 0.811 + ], + "angle": 0, + "content": "where \\( c' \\) represents the selected neighboring class, \\( \\mathcal{B}^{c'} \\) is the set of samples belong to class \\( c' \\) in the batch \\( \\mathcal{B} \\). \\( \\hat{y}_i^{c'} \\) is the predicted probability of the \\( i \\)-th sample belonging to its correct class \\( c' \\), and \\( \\text{Norm}(*) \\) is the normalization function that ensures the sum of \\( p_{s'} \\) equals 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In this way, for each target sample \\( z_{s} \\), a candidate sample \\( z_{s'} \\) with higher confidence is sampled from the batch, thereby achieving exploratory examples with diversity. FwS is essential for the validity of SEL, which is why we use the adaptive OFA operator to synthesize exploratory examples. The rational analysis is elaborated later." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1873" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.092, + 0.355, + 0.106 + ], + "angle": 0, + "content": "3.3.5. Training with SEL Framework" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.111, + 0.484, + 0.186 + ], + "angle": 0, + "content": "SEL is a plug-and-play training framework that can be seamlessly integrated into the training process of the original model. The optimization objective of SEL is to correctly classify the exploratory examples, which is equivalent to maximizing the predicted probability of the correct class." + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.193, + 0.483, + 0.235 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S E L} = - \\frac {1}{C} \\sum_ {c = 1} ^ {C} \\left(\\frac {1}{k _ {c}} \\sum_ {i = 1} ^ {k _ {c}} \\log \\left(\\hat {y} _ {i} ^ {c}\\right)\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.24, + 0.484, + 0.286 + ], + "angle": 0, + "content": "where \\(\\hat{y}_i^c\\) denotes the predicted probability of the \\(i\\)-th exploratory example belonging to the \\(c\\)-th class, and \\(k_c\\) is the number of exploratory examples synthesized for class \\(c\\)." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.286, + 0.478, + 0.301 + ], + "angle": 0, + "content": "Finally, the overall optimization objective is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.308, + 0.483, + 0.326 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = (1 - \\lambda) \\mathcal {L} _ {\\text {T a s k}} + \\lambda \\mathcal {L} _ {\\text {S E L}} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.332, + 0.484, + 0.378 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{Task}\\) is the optimization objective of the original method, and \\(\\lambda\\) is a weighting coefficient that controls the impact of SEL loss." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.385, + 0.272, + 0.4 + ], + "angle": 0, + "content": "3.4. Rationale Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.407, + 0.484, + 0.513 + ], + "angle": 0, + "content": "The exploitation process of OFA has proved to be effective and rational [51]. In this section, we provide a rational analysis of SEL. For clarity, we denote the weights of the linear classifier as \\( w = \\{w_{c}\\}_{c = 1}^{C} \\), where \\( C \\) is the number of classes, and \\( w_{c} \\in \\mathbb{R}^{d \\times 1} \\) denotes the weights of class \\( c \\), \\( d \\) is the feature dimension. For a sample \\( z \\in \\mathbb{R}^{d \\times 1} \\), the predicted probability of class \\( c \\) can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.221, + 0.518, + 0.483, + 0.559 + ], + "angle": 0, + "content": "\\[\n\\hat {y} ^ {c} = \\frac {e ^ {w _ {c} ^ {T} z}}{\\sum_ {c ^ {\\prime} = 1} ^ {C} e ^ {w _ {c ^ {\\prime}} ^ {T} z}} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.484, + 0.61 + ], + "angle": 0, + "content": "Hence, \\(\\hat{y}^c\\propto w_c^T z\\) . The classifier attempts to assign a higher probability to the target class, i.e., \\(\\hat{y}^c >\\hat{y}^{c'}\\) , where \\(c\\) and \\(c^\\prime\\) denote the target and other classes, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.484, + 0.761 + ], + "angle": 0, + "content": "Without loss of generality, the selection of the target sample \\((z_{s})\\) and candidate sample \\((z_{s}^{\\prime})\\) can be divided into two cases, as summarized in Tab. 1. We use the subscripts \\(t\\) (tail) and \\(h\\) (head) to represent their relative relation. Generally, the first case typically dominates, as the CbC module selects more tail-class targets, while the long-tailed distribution increases the chance of candidate samples being from head classes. Additionally, since exploratory examples share labels with targets, the optimization aims to challenge the model's bias toward head classes." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.761, + 0.468, + 0.778 + ], + "angle": 0, + "content": "For the first case, the goal of \\(\\hat{y}_e^t >\\hat{y}_e^h\\) is equivalent to:" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.782, + 0.483, + 0.808 + ], + "angle": 0, + "content": "\\[\nw _ {t} ^ {T} z _ {e} - w _ {h} ^ {T} z _ {e} > 0 \\Rightarrow \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.804, + 0.44, + 0.822 + ], + "angle": 0, + "content": "\\[\nw _ {t} ^ {T} \\left(z _ {t} + r \\left(z _ {h} - z _ {t}\\right)\\right) - w _ {h} ^ {T} \\left(z _ {t} + r \\left(z _ {h} - z _ {t}\\right)\\right) > 0\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.829, + 0.478, + 0.844 + ], + "angle": 0, + "content": "After expanding and transforming the inequality, we have:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.849, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} w _ {t} ^ {T} z _ {t} - w _ {h} ^ {T} z _ {t} > \\frac {r}{1 - r} \\left(w _ {h} ^ {T} z _ {h} - w _ {t} ^ {T} z _ {h}\\right) \\tag {9} \\\\ \\Rightarrow \\hat {y} _ {t} ^ {t} - \\hat {y} _ {t} ^ {h} > \\frac {r}{1 - r} (\\hat {y} _ {h} ^ {h} - \\hat {y} _ {h} ^ {t}) \\\\ \\end{array}\n\\]" + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.09, + 0.906, + 0.164 + ], + "angle": 0, + "content": "
Casezszs'zeGoal
1ztzhzt + r(zh - zt)ˆy e t >ˆy e h
2zhztzh + r(zt - zh)ˆy e h >ˆy e t
" + }, + { + "type": "table_caption", + "bbox": [ + 0.54, + 0.173, + 0.879, + 0.188 + ], + "angle": 0, + "content": "Table 1. Two types of synthesized exploratory examples." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.905, + 0.334 + ], + "angle": 0, + "content": "where \\(\\hat{y}_h^h\\) and \\(\\hat{y}_h^t\\) represent probabilities of the candidate sample \\((z_h)\\) belonging to the head class and tail classes, respectively. The classifier favors head classes \\((\\hat{y}_h^h - \\hat{y}_h^t > 0)\\), while the FwS module reinforces this by selecting high-confidence candidates. Therefore, the right item is tend to be positive, and \\(\\hat{y}_t^t > \\hat{y}_t^h\\) is achieved, meaning that optimizing tail-class exploratory examples effectively improves the classification accuracy of original tail-class samples." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.335, + 0.906, + 0.366 + ], + "angle": 0, + "content": "Similarly, for the second case, the optimization goal of SEL is to make \\(\\hat{y}_e^h >\\hat{y}_e^t\\) yielding:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.375, + 0.905, + 0.402 + ], + "angle": 0, + "content": "\\[\n\\hat {y} _ {h} ^ {h} - \\hat {y} _ {h} ^ {t} > \\frac {r}{1 - r} \\left(\\hat {y} _ {t} ^ {t} - \\hat {y} _ {t} ^ {h}\\right) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.413, + 0.906, + 0.535 + ], + "angle": 0, + "content": "where \\(\\hat{y}_t^t\\) and \\(\\hat{y}_t^h\\) represent probabilities of the candidate sample \\((z_t)\\) belonging to the tail class and head class, respectively. Although the classifier tends to misclassify tail-class samples as head classes, the FwS mitigates this by selecting better candidates, which possess higher probabilities in the correct classes. Hence, optimizing the head-class exploratory examples does not cause significant damage to the original head sample." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.548, + 0.646, + 0.565 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.573, + 0.61, + 0.587 + ], + "angle": 0, + "content": "4.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.906, + 0.669 + ], + "angle": 0, + "content": "CIFAR-10/100-LT [5] are the subsets of CIFAR-10/100 [20]. Both contain 50,000 training images and 10,000 validation images, each of size \\(32 \\times 32\\), with 10 and 100 classes, respectively. The imbalance factors used in the experiment are set to 100, 50, and 10." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.67, + 0.906, + 0.731 + ], + "angle": 0, + "content": "ImageNet-LT [30] is a long-tailed version of ImageNet [6] by sampling a subset following the Pareto distribution with power value \\(\\alpha = 0.64\\). It consists of \\(115.8\\mathrm{K}\\) images of 1000 classes in total, with 1280 to 5 images per class." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.731, + 0.906, + 0.775 + ], + "angle": 0, + "content": "iNaturalist 2018 [12] is a large-scale dataset containing 437.5K images from 8,142 classes. It is long-tailed by nature, with an extremely imbalanced distribution." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.785, + 0.651, + 0.801 + ], + "angle": 0, + "content": "4.2. Basic Setting" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.807, + 0.699, + 0.822 + ], + "angle": 0, + "content": "4.2.1. Compared Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.902 + ], + "angle": 0, + "content": "To evaluate the effectiveness of our proposed SEL, we integrated it into several existing methods introduced in recent years, including re-balanced techniques (IB [31], GCL [23], KPS [24]), contrastive learning (BCL [52]), feature compression (FCC [22]), and data augmentation strategies" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1874" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.111, + 0.089, + 0.888, + 0.442 + ], + "angle": 0, + "content": "
MethodsCIFAR-100-LTCIFAR-10-LT
AvenueImbalance Ratio ρ10050101005010
-ResNet-32 [9]39.9646.0156.6671.9675.9585.92
+ MixUp [45]41.55↑1.5947.11↑1.1057.18↑0.5273.43↑1.4777.19↑1.2486.37↑0.45
+ SEL44.53↑4.5749.64↑3.8559.17↑2.5175.84↑3.8879.81↑3.8686.78↑0.86
ICCV 2021IB [31]40.5046.3156.7575.1179.9588.01
+ SEL42.34↑1.8447.95↑1.6457.57↑0.8275.04↓0.0779.81↓0.1488.13↑0.12
CVPR 2022BCL [52]51.7656.5167.9083.6186.1390.10
+ SEL52.30↑0.5457.25↑0.7468.43↑0.5384.44↑0.8386.31↑0.1890.26↑0.16
CVPR 2022GCL [23]46.5051.7261.7980.5684.7489.65
+ SEL47.89↑1.3952.74↑1.0262.41↑0.6281.86↑1.3085.13↑0.3989.78↑0.13
CVPR 2023FCC+CE [22]40.2045.9357.8073.8079.5787.75
+ SEL42.33↑2.1348.05↑2.1258.75↑0.9577.88↑4.0882.10↑2.5388.80↑1.05
CVPR 2023GLMC [7]53.9158.8768.0783.6886.9091.16
+ SEL56.48↑2.5761.13↑2.2670.75↑2.6885.40↑1.7288.57↑1.6792.83↑1.67
TPAMI 2023KPS [24]41.9747.9259.5982.3284.2989.10
+ SEL44.01↑2.0447.80↓0.1260.03↑0.4483.48↑1.1684.49↑0.2089.34↑0.24
AAAI 2024H2T+CE [25]42.2747.5858.2479.9182.8088.77
+ SEL42.45↑0.1847.80↑0.2258.66↑0.4280.43↑0.5282.80↓0.0088.79↑0.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.185, + 0.453, + 0.812, + 0.468 + ], + "angle": 0, + "content": "Table 2. Comparisons between raw methods and their SEL-enhanced counterparts on CIFAR-LT datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.494, + 0.483, + 0.616 + ], + "angle": 0, + "content": "(H2T [25], GLMC [7]). These methods were chosen not only for their representativeness but also for their publicly available code and configurations, ensuring ease of reproducibility. Most employ a two-stage training strategy, with the first stage dedicated to representation learning and the second to classification. Additionally, we use ResNet models as baselines to directly compare the performance improvements achieved by MixUp and SEL." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.625, + 0.303, + 0.64 + ], + "angle": 0, + "content": "4.2.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We reproduce the compared methods using their official codes and configurations, setting the random seed to 2024 for fair comparison. The base ResNet models are configured identically to those in Li et al. [23]. To simplify application, we first determine the optimal way to apply SEL on different datasets by using the base model, i.e., ResNet-32 for CIFAR-10/100-LT and ResNet-50 for others, then replace it with corresponding models to assess SEL's effectiveness on these models. Specifically, for CIFAR-100-LT, SEL is applied during the final 20 training epochs. For other datasets, SEL is applied by further fine-tuning the classifier for 20 epochs on CIFAR-10-LT and 10 epochs on ImageNet-LT and iNaturalist 2018. The weight coefficient \\(\\lambda\\) in Eq. (6) is set to 0.9, with the number of neighboring classes \\(k_c'\\) set to 2, 5, 10, and 50 for CIFAR-10-LT, CIFAR-100-LT, ImageNet-LT, and iNaturalist 2018, respectively. Top-1 accuracy serves as the evaluation metric." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.494, + 0.813, + 0.51 + ], + "angle": 0, + "content": "4.3. Improvement of Existing Methods" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.517, + 0.763, + 0.531 + ], + "angle": 0, + "content": "4.3.1. Results on CIFAR-10/100-LT" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.537, + 0.907, + 0.763 + ], + "angle": 0, + "content": "For CIFAR-10/100-LT, the results with and without SEL are presented in Tab. 2. Across 48 experiment groups, our proposed SEL significantly improves 44 of them, with an average increase of \\(1.37\\%\\) (4.57% max and \\(0.02\\%\\) min). Extensive experimental results fully verify the effectiveness of our method on long-tailed recognition, particularly for scenarios with high imbalance ratios. Generally, integrating MixUp into the vanilla ResNet-32 model significantly enhances performance, highlighting the effectiveness of data augmentation. However, our proposed SEL achieves even greater improvements across all scenarios, and even surpasses existing methods. For example, ResNet-32+SEL outperforms IB on both datasets except for CIFAR-10-LT-10, and surpasses GCL, KPS, and H2T on CIFAR-100-LT, though it underperforms on CIFAR-10-LT." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Additionally, SEL significantly enhances the performance of existing methods, with particularly pronounced improvements in high-imbalance scenarios. For example, SEL is notably effective on KPS at an imbalance ratio of 100 but shows limited impact at other imbalance ratios. In addition, GLMC+SEL and FCC+SEL significantly outperform their original counterparts on both datasets, with average improvements of \\(2.10\\%\\) and \\(2.14\\%\\), respectively. For the re-balanced methods IB, GCL, and KPS, SEL yields" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1875" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.477, + 0.312 + ], + "angle": 0, + "content": "
MethodManyMediumFewAll
ResNet-5063.3833.7426.0444.80
+ Mixup62.9735.4428.5845.71↑0.91
+ SEL64.1438.4131.2747.86↑3.06
BCL65.8353.1636.2755.67
+ SEL65.4554.8236.5356.27↑0.60
GCL65.0452.7735.6955.51
+ SEL65.1353.9439.1856.12↑0.61
FCC+CE65.5252.1833.8354.68
+ SEL64.7454.0035.6655.60↑0.92
GLMC69.4252.4930.4156.17
+ SEL68.6754.3738.2857.24↑1.07
" + }, + { + "type": "table_caption", + "bbox": [ + 0.135, + 0.323, + 0.44, + 0.337 + ], + "angle": 0, + "content": "Table 3. Comparisons on the ImageNet-LT dataset." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.352, + 0.476, + 0.578 + ], + "angle": 0, + "content": "
MethodManyMediumFewAll
ResNet-5074.8968.2461.4265.95
+ Mixup71.3469.5363.8867.29↑1.34
+ SEL73.0470.3866.7469.10↑3.15
BCL73.4771.9968.8270.80
+ SEL71.9973.0469.2171.32↑0.52
GCL73.7572.2868.3470.76
+ SEL72.6373.5870.0572.00↑1.24
FCC+CE72.7871.0967.4169.70
+ SEL71.1772.4169.5571.09↑1.39
GLMC76.0473.6971.7773.14
+ SEL74.6875.1473.6674.51↑1.37
" + }, + { + "type": "table_caption", + "bbox": [ + 0.127, + 0.589, + 0.446, + 0.602 + ], + "angle": 0, + "content": "Table 4. Comparisons on the iNaturalist 2018 dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.631, + 0.483, + 0.844 + ], + "angle": 0, + "content": "greater improvements on CIFAR-100-LT than on CIFAR-10-LT, demonstrating its superiority in scenarios with more classes. However, SEL is less effective on BCL and H2T. BCL uses contrastive learning to reinforce balanced decision regions, aligning with SEL's objective. The difference is that SEL further enhances the robustness of decision regions, resulting in additional improvements. H2T is also an evolutionary algorithm-inspired method that generates diverse samples through cross-mutation, following the principles of Genetic Algorithm (GA), thus indicating the fundamental difference between SEL and H2T stems from the distinction between OFA and GA. The slight improvement of H2T+SEL further validates SEL's effectiveness in generating diverse samples, aligning with H2T's principles." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.851, + 0.455, + 0.867 + ], + "angle": 0, + "content": "4.3.2. Results on ImageNet-LT and iNaturalist 2018" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.87, + 0.483, + 0.902 + ], + "angle": 0, + "content": "For ImageNet-LT and iNaturalist 2018, experimental results are listed in Tab. 3 and Tab. 4, respectively. Following pre" + }, + { + "type": "table", + "bbox": [ + 0.547, + 0.089, + 0.875, + 0.202 + ], + "angle": 0, + "content": "
CbCFwSTop-1 Accuracy
ρ =100ρ =10
××39.9656.66
×✓40.26↑0.3056.93↑0.27
✓×43.44↑3.4857.84↑1.18
✓✓44.53↑4.5759.17↑2.51
" + }, + { + "type": "table_caption", + "bbox": [ + 0.604, + 0.213, + 0.815, + 0.227 + ], + "angle": 0, + "content": "Table 5. Ablations of key modules." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.255, + 0.907, + 0.436 + ], + "angle": 0, + "content": "vious works [7, 16], also report accuracy across three class splits: Many \\((n_c \\geq 100)\\), Medium \\((20 < n_c < 100)\\), and Few \\((n_c \\leq 20)\\). Overall, SEL outperforms MixUp on both datasets, and consistently enhances the compared methods, achieving an average improvement of \\(1.25\\%\\) (max \\(3.06\\%\\), min \\(0.61\\%\\)) on ImageNet-LT and \\(1.53\\%\\) (max \\(3.15\\%\\), min \\(0.52\\%\\)) on iNaturalist 2018. As seen, training with SEL typically improves the performance of Few classes while slightly reducing that of Many classes, aligning with the rationale proved in Sec. 3.4. However, the greater gain in Few classes outweighs the slight decline in Many classes, leading to an overall improvement in the All classes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.446, + 0.681, + 0.462 + ], + "angle": 0, + "content": "4.4. Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.468, + 0.906, + 0.514 + ], + "angle": 0, + "content": "To further analyze the principles of SEL quantitatively, we conducted a series of ablation studies on the CIFAR-10/100-LT datasets using ResNet-32 [9] as the base model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.522, + 0.792, + 0.538 + ], + "angle": 0, + "content": "4.4.1. Visualization of Decision Regions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.541, + 0.907, + 0.723 + ], + "angle": 0, + "content": "Fig. 4 visualizes the decision regions and boundaries of different classes using t-SNE [35] across various methods. These models are trained on the CIFAR-100-LT with an imbalance ratio of 100 and tested on a balanced test set. For comparison, we selected four classes from CIFAR-100 test set for visualization: where apple and aquarium fish are the head classes with the most training samples, and willow tree and wolf are the tail classes with the fewest training samples. The decision boundaries are derived using the SVM algorithm. From Fig. 4, we can observe that SEL results in more balanced decision regions for both head and tail classes, aligning with our motivation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.731, + 0.706, + 0.746 + ], + "angle": 0, + "content": "4.4.2. Role of Key Modules" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The CbC and FwS modules are the core components of SEL. To examine their roles, we conduct ablation experiments on CIFAR-100-LT with imbalance ratios of 100 and 10. Tab. 5 presented the experimental results, showing that the CbC module significantly enhances the baseline method, with improvements of \\(3.48\\%\\) \\((39.96\\% \\rightarrow 43.44\\%)\\) and \\(1.18\\%\\) \\((56.66\\% \\rightarrow 57.84\\%)\\) for imbalance ratios of 100 and 10, respectively. These results highlight the CbC module's crucial role in balancing data distribution, enabling the model to learn more effectively from tail classes. However," + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1876" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.088, + 0.368, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.227, + 0.255, + 0.239 + ], + "angle": 0, + "content": "(a) CE" + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.089, + 0.631, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.469, + 0.227, + 0.529, + 0.24 + ], + "angle": 0, + "content": "(b) MixUp" + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.089, + 0.895, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.74, + 0.227, + 0.784, + 0.239 + ], + "angle": 0, + "content": "(c) SEL" + }, + { + "type": "image_footnote", + "bbox": [ + 0.348, + 0.242, + 0.652, + 0.256 + ], + "angle": 0, + "content": "apple aquarium fish willow tree wolf" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.267, + 0.908, + 0.296 + ], + "angle": 0, + "content": "Figure 4. Visualization of decision regions and boundaries on the CIFAR-100 test set (each class contains 100 test samples), where apple and aquarium fish are head classes with the most training samples, and willow tree and wolf are tail classes with the fewest training samples." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.317, + 0.285, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.406, + 0.238, + 0.417 + ], + "angle": 0, + "content": "(a) CIFAR-100-LT" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.32, + 0.484, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.341, + 0.406, + 0.435, + 0.417 + ], + "angle": 0, + "content": "(b) CIFAR-10-LT" + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.431, + 0.467, + 0.445 + ], + "angle": 0, + "content": "Figure 5. Changes with different neighboring class numbers." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.462, + 0.289, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.15, + 0.552, + 0.231, + 0.564 + ], + "angle": 0, + "content": "(a) start epochs" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.462, + 0.484, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.351, + 0.552, + 0.426, + 0.564 + ], + "angle": 0, + "content": "(b) weight \\((\\lambda)\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.577, + 0.42, + 0.591 + ], + "angle": 0, + "content": "Figure 6. Changes with different parameters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.623, + 0.484, + 0.699 + ], + "angle": 0, + "content": "the FwS module has a minimal impact without the CbC module, yielding only \\(0.30\\%\\) and \\(0.27\\%\\) improvements for imbalance ratios of 100 and 10, respectively. Its effectiveness is fully realized when combined with the CbC module, leading to a significant boost in model performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.714, + 0.353, + 0.729 + ], + "angle": 0, + "content": "4.4.3. Impact of different parameters" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this section, we conduct exploratory experiments on the impact of the number of neighboring classes \\((k_c^{\\prime})\\) in the FwS module. Experimental results are summarized in Fig. 5, where \\(k_{c}^{\\prime}\\) is varied from 1 to 10 and relative accuracy (values are adjusted by subtracting their mean) is utilized for comparison. As seen, the model performs better with a higher \\(k^{\\prime}_{c}\\) on CIFAR-100-LT but with a lower \\(k^{\\prime}_{c}\\) on CIFAR-10LT. This difference arises from the number of classes, as CIFAR-100-LT has 100 classes, whereas CIFAR-10-LT has only 10. Based on these observations, we empirically set \\(k^{\\prime}_{c}\\) to 5 for CIFAR-100-LT and 2 for CIFAR-10-LT." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.321, + 0.759, + 0.337 + ], + "angle": 0, + "content": "4.4.4. When to Start SEL Training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.907, + 0.568 + ], + "angle": 0, + "content": "In this section, we explore the optimal starting point for SEL training to maximize computational efficiency, as its synthesis of additional exploratory examples increases computational costs. Specifically, we integrate SEL into the training process at different epochs, ranging from 100 to 200, as shown in Fig. 6(a). Here, starting at 200 epochs indicates a two-stage training mode, where SEL is applied with an additional 20 epochs after the original training process. Similarly, relative accuracy is used as the ordinate for clearly showing the parameter's impact. It is obvious that the model's performance steadily improves as the SEL starting point increases, with most datasets peaking at 180 epochs, except for CIFAR-10-LT-100 and CIFAR-10-LT-10. Therefore, we set the starting point to 180 epochs for CIFAR-100-LT and 200 epochs for CIFAR-10-LT." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.57, + 0.909, + 0.706 + ], + "angle": 0, + "content": "Building on the optimal starting points, we further examine the impact of parameter \\(\\lambda\\) in Eq. (6). Fig. 6(b) illustrates the performance variation with \\(\\lambda\\) ranges from 0.0 to 1.0. Note that \\(\\lambda = 0.0\\) indicates the absence of SEL, while \\(\\lambda = 1.0\\) signifies its full integration. For most datasets, as \\(\\lambda\\) increases, the model's performance gradually improves, peaking at \\(\\lambda = 0.9\\) before declining. However, for CIFAR-10-LT-10, the simplest dataset, performance remains relatively stable, with a peak at \\(\\lambda = 0.6\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.723, + 0.634, + 0.739 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose SEL, a simple yet effective training framework that addresses long-tailed challenges through the space exploration perspective. SEL is a plug-and-play method that boosts performance without modifying the model or adding parameters. It uses an adaptive OFA operator to generate exploratory examples that capture unseen tail-class features, enhancing decision regions and boundaries. The effectiveness of SEL is validated mathematically through rational analysis and empirically through experiments on various long-tailed datasets and methods." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1877" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.482, + 0.207 + ], + "angle": 0, + "content": "This work is supported by the Solfeggio ear training intelligent robot and cloud platform research and development project for music education (No.2024CXY0102), the 3D visualization digital twin integrated control system (No.2023CXY0111), and the public technology service platform project of Xiamen City (No.3502Z20231043)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.222, + 0.188, + 0.237 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.247, + 0.483, + 0.288 + ], + "angle": 0, + "content": "[1] Yoshua Bengio, Yann Lecun, and Geoffrey Hinton. Deep learning for ai. Communications of the ACM, 64(7):58-65, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.29, + 0.484, + 0.359 + ], + "angle": 0, + "content": "[2] Ekin D. Cubuk, Barret Zoph, Dandelion Mané, Vijay Vasudevan, and Quoc V. Le. Autoaugment: Learning augmentation strategies from data. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 113-123, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.361, + 0.483, + 0.43 + ], + "angle": 0, + "content": "[3] Ekin Dogus Cubuk, Barret Zoph, Jon Shlens, and Quoc Le. Randaugment: Practical automated data augmentation with a reduced search space. In Advances in Neural Information Processing Systems, pages 18613-18624. Curran Associates, Inc., 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.432, + 0.483, + 0.487 + ], + "angle": 0, + "content": "[4] Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, and Jiaya Jia. Parametric contrastive learning. In 2021 IEEE/CVF International Conference on Computer Vision, pages 695-704, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.489, + 0.484, + 0.545 + ], + "angle": 0, + "content": "[5] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge J. Belongie. Class-balanced loss based on effective number of samples. In IEEE Conference on Computer Vision and Pattern Recognition, pages 9268-9277, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.547, + 0.483, + 0.614 + ], + "angle": 0, + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.617, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[7] Fei Du, Peng Yang, Qi Jia, Fengtao Nan, Xiaoting Chen, and Yun Yang. Global and local mixture consistency cumulative learning for long-tailed visual recognitions. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15814-15823, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.688, + 0.484, + 0.744 + ], + "angle": 0, + "content": "[8] Hao Guo and Song Wang. Long-tailed multi-label visual recognition by collaborative training on uniform and rebalanced samplings. In IEEE Conference on Computer Vision and Pattern Recognition, pages 15089-15098, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.746, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[10] Daniel Ho, Eric Liang, Xi Chen, Ion Stoica, and Pieter Abbeel. Population based augmentation: Efficient learning of augmentation policy schedules. In Proceedings of the 36th International Conference on Machine Learning, pages 2731-2741, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[11] Gregory Holste, Yiliang Zhou, Song Wang, Ajay Jaiswal, Mingquan Lin, Sherry Zhuge, Yuzhe Yang, Dongkyun Kim," + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.247, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.203 + ], + "angle": 0, + "content": "Trong-Hieu Nguyen-Mau, Minh-Triet Tran, Jaehyup Jeong, Wongi Park, Jongbin Ryu, Feng Hong, Arsh Verma, Yosuke Yamagishi, Changhyun Kim, Hyeryeong Seo, Myungjoo Kang, Leo Anthony Celi, Zhiyong Lu, Ronald M. Summers, George Shih, Zhangyang Wang, and Yifan Peng. Towards long-tailed, multi-label disease classification from chest x-ray: Overview of the cxr-It challenge. Medical Image Analysis, 97:103224, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[12] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alexander Shepard, Hartwig Adam, Pietro Perona, and Serge J. Belongie. The inaturalist species classification and detection dataset. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, pages 8769-8778, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.29, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[13] Chengkai Hou, Jieyu Zhang, Haonan Wang, and Tianyi Zhou. Subclass-balancing contrastive learning for long-tailed recognition. In 2023 IEEE/CVF International Conference on Computer Vision, pages 5372-5384, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.347, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[14] Shuyan Hu, Wei Ni, Xin Wang, and Abbas Jamalipour. Disguised tailing and video surveillance with solar-powered fixed-wing unmanned aerial vehicle. IEEE Transactions on Vehicular Technology, 71(5):5507-5518, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[15] ZhongQuan Jian and GuangYu Zhu. Affine invariance of meta-heuristic algorithms. Information Sciences, 576:37-53, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.447, + 0.905, + 0.515 + ], + "angle": 0, + "content": "[16] Yan Jin, Mengke Li, Yang Lu, Yiu-ming Cheung, and Hanzi Wang. Long-tailed visual recognition via self-heterogeneous integration with knowledge excavation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23695-23704, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.517, + 0.905, + 0.586 + ], + "angle": 0, + "content": "[17] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. In 8th International Conference on Learning Representations, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.588, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[18] Bingyi Kang, Yu Li, Sa Xie, Zehuan Yuan, and Jiashi Feng. Exploring balanced feature spaces for representation learning. In 9th International Conference on Learning Representations, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.645, + 0.905, + 0.713 + ], + "angle": 0, + "content": "[19] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.716, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[20] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[21] Jun Li, Zichang Tan, Jun Wan, Zhen Lei, and Guodong Guo. Nested collaborative learning for long-tailed visual recognition. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6939-6948, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.802, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[22] Jian Li, Ziyao Meng, Daqian Shi, Rui Song, Xiaolei Dao, Jingwen Wang, and Hao Xu. FCC: feature clusters compression for long-tailed visual recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24080-24089, 2023. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[23] Mengke Li, Yiu-Ming Cheung, and Yang Lu. Long-tailed visual recognition via gaussian clouded logit adjustment. In" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1878" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6919-6928, 2022. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[24] Mengke Li, Yiu-Ming Cheung, and Zhikai Hu. Key point sensitive loss for long-tailed visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4): 4812-4825, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.483, + 0.274 + ], + "angle": 0, + "content": "[25] Mengke Li, Zhikai Hu, Yang Lu, Weichao Lan, Yiu-ming Cheung, and Hui Huang. Feature fusion from head to tail for long-tailed visual recognition. In Thirty-Eighth AAAI Conference on Artificial Intelligence, Thirty-Sixth Conference on Innovative Applications of Artificial Intelligence, Fourteenth Symposium on Educational Advances in Artificial Intelligence, pages 13581-13589, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.275, + 0.482, + 0.343 + ], + "angle": 0, + "content": "[26] Shuang Li, Kaixiong Gong, Chi Harold Liu, Yulin Wang, Feng Qiao, and Xinjing Cheng. Metasaug: Meta semantic augmentation for long-tailed visual recognition. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5212-5221, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.483, + 0.412 + ], + "angle": 0, + "content": "[27] Shuang Li, Kaixiong Gong, Chi Harold Liu, Yulin Wang, Feng Qiao, and Xinjing Cheng. Metasaug: Meta semantic augmentation for long-tailed visual recognition. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5212-5221, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.483, + 0.483 + ], + "angle": 0, + "content": "[28] Tianhong Li, Peng Cao, Yuan Yuan, Lijie Fan, Yuzhe Yang, Rogerio Feris, Piotr Indyk, and Dina Katabi. Targeted supervised contrastive learning for long-tailed recognition. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6918, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.483, + 0.538 + ], + "angle": 0, + "content": "[29] Bo Liu, Haoxiang Li, Hao Kang, Gang Hua, and Nuno Vasconcelos. Gistnet: a geometric structure transfer network for long-tailed recognition. In 2021 IEEE/CVF International Conference on Computer Vision, pages 8189-8198, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.483, + 0.607 + ], + "angle": 0, + "content": "[30] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X. Yu. Large-scale long-tailed recognition in an open world. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2537-2546, 2019. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.483, + 0.664 + ], + "angle": 0, + "content": "[31] Seulki Park, Jongin Lim, Younghan Jeon, and Jin Young Choi. Influence-balanced loss for imbalanced visual classification. In 2021 IEEE/CVF International Conference on Computer Vision, pages 715-724, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.483, + 0.734 + ], + "angle": 0, + "content": "[32] Seulki Park, Youngkyu Hong, Byeongho Heo, Sangdoo Yun, and Jin Young Choi. The majority can help the minority: Context-rich minority oversampling for long-tailed classification. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6877-6886, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[33] Jiawei Ren, Cunjun Yu, Shunan Sheng, Xiao Ma, Haiyu Zhao, Shuai Yi, and Hongsheng Li. Balanced meta-softmax for long-tailed visual recognition. In Advances in Neural Information Processing Systems 33, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.483, + 0.859 + ], + "angle": 0, + "content": "[34] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott E. Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In 2015 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-9, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[35] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[36] Vikas Verma, Alex Lamb, Christopher Beckham, Amir Najafi, Ioannis Mitlagkas, David Lopez-Paz, and Yoshua Bengio. Manifold mixup: Better representations by interpolating hidden states. In Proceedings of the 36th International Conference on Machine Learning, pages 6438-6447, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.907, + 0.232 + ], + "angle": 0, + "content": "[37] Jianfeng Wang, Thomas Lukasiewicz, Xiaolin Hu, Jianfei Cai, and Zhenghua Xu. Rsg: A simple but effective module for learning imbalanced datasets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3784-3793, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[38] Peng Wang, Kai Han, Xiu-Shen Wei, Lei Zhang, and Lei Wang. Contrastive learning based hybrid networks for long-tailed image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 943-952, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.301, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[39] Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu, and Stella X. Yu. Long-tailed recognition by routing diverse distribution-aware experts. In 9th International Conference on Learning Representations, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[40] Tong Wu, Ziwei Liu, Qingqiu Huang, Yu Wang, and Dahua Lin. Adversarial robustness under long-tailed distribution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 8659-8668, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.497 + ], + "angle": 0, + "content": "[41] Shiyu Xuan and Shiliang Zhang. Decoupled contrastive learning for long-tailed recognition. In Thirty-Eighth AAAI Conference on Artificial Intelligence, Thirty-Sixth Conference on Innovative Applications of Artificial Intelligence, Fourteenth Symposium on Educational Advances in Artificial Intelligence, pages 6396-6403, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.567 + ], + "angle": 0, + "content": "[42] Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Seong Joon Oh, Youngjoon Yoo, and Junsuk Choe. Cutmix: Regularization strategy to train strong classifiers with localizable features. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6022-6031, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[43] Shiran Zada, Itay Benou, and Michal Irani. Pure noise to the rescue of insufficient data: Improving imbalanced classification by training on random noise images. In International Conference on Machine Learning, pages 25817-25833, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.636, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[44] Yuhang Zang, Chen Huang, and Chen Change Loy. Fasa: Feature augmentation and sampling adaptation for long-tailed instance segmentation. In 2021 IEEE/CVF International Conference on Computer Vision, pages 3437-3446, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.705, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[45] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In 6th International Conference on Learning Representations, 2018. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[46] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):10795-10816, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[47] Zizhao Zhang and Tomas Pfister. Learning fast sample reweighting without reward data. In 2021 IEEE/CVF International Conference on Computer Vision, pages 705-714, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[48] Boyan Zhou, Quan Cui, Xiu-Shen Wei, and Zhao-Min Chen. BBN: bilateral-branch network with cumulative learning for" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1879" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "long-tailed visual recognition. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9716-9725, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[49] Weitao Zhou, Zhong Cao, Nanshan Deng, Xiaoyu Liu, Kun Jiang, and Diange Yang. Dynamically conservative self-driving planner for long-tail cases. IEEE Transactions on Intelligent Transportation Systems, 24(3):3476-3488, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.483, + 0.275 + ], + "angle": 0, + "content": "[50] Chenguang Zhu, Yichong Xu, Xiang Ren, Bill Yuchen Lin, Meng Jiang, and Wenhao Yu. Knowledge-augmented methods for natural language processing. In Proceedings of the Sixteenth ACM International Conference on Web Search and Data Mining, pages 1228-1231, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[51] Guang-Yu Zhu and Wei-Bo Zhang. Optimal foraging algorithm for global optimization. Applied Soft Computing, 51: 294–313, 2017. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.483, + 0.388 + ], + "angle": 0, + "content": "[52] Jianggang Zhu, Zheng Wang, Jingjing Chen, Yi-Ping Phoebe Chen, and Yu-Gang Jiang. Balanced contrastive learning for long-tailed visual recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6898-6907, 2022. 1, 2, 3, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1880" + } + ] +] \ No newline at end of file diff --git a/2025/Supervised Exploratory Learning for Long-Tailed Visual Recognition/3fb70b61-418e-4890-be11-1e54f33e6902_origin.pdf b/2025/Supervised Exploratory Learning for Long-Tailed Visual Recognition/3fb70b61-418e-4890-be11-1e54f33e6902_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ec8f94d09fff1f7a03c955f70b03d06d656e3a2c --- /dev/null +++ b/2025/Supervised Exploratory Learning for Long-Tailed Visual Recognition/3fb70b61-418e-4890-be11-1e54f33e6902_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b9ffa526f08600e5052cfd4dbf5ebbab1ef0ae6b03d27b79b56d20b7d17c7270 +size 2777344 diff --git a/2025/Supervised Exploratory Learning for Long-Tailed Visual Recognition/full.md b/2025/Supervised Exploratory Learning for Long-Tailed Visual Recognition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8eb7c5ef65d869981fd1ee63d0de786f14e32129 --- /dev/null +++ b/2025/Supervised Exploratory Learning for Long-Tailed Visual Recognition/full.md @@ -0,0 +1,383 @@ +# Supervised Exploratory Learning for Long-Tailed Visual Recognition + +Zhongquan Jian $^{1,3,\dagger}$ , Yanhao Chen $^{2,\dagger}$ , Yancheng Wang $^{4}$ , Junfeng Yao $^{2,3,1,*}$ , Meihong Wang $^{3,*}$ , Qingqiang Wu $^{2,3,*}$ + +$^{1}$ School of Computer and Data Science, Minjiang University, Fuzhou, China + +$^{2}$ School of Film, Xiamen University, Xiamen, China + +$^{3}$ School of Informatics, Xiamen University, Xiamen, China + +$^{4}$ School of Computing and Data Science, Xiamen University Malaysia, Malaysia + +{yao0010, wangmh, wuqq}@xmu.edu.cn + +# Abstract + +Long-tailed data poses a significant challenge for deep learning models, which tend to prioritize accurate classification of head classes while largely neglecting tail classes. Existing techniques, such as class re-balancing, logit adjustment, and data augmentation, aim to enlarge decision regions of tail classes or achieve clear decision boundaries, leaving the robustness of decision regions underconsidered. This paper proposes a simple yet effective Supervised Exploratory Learning (SEL) framework to achieve these goals simultaneously from space exploration perspectives. SEL employs the adaptive Optimal Foraging Algorithm (OFA) to generate diverse exploratory examples, integrating Class-biased Complement (CbC) for balanced class distribution and Fitness-weighted Sampling (FwS) for space exploration. Both theoretical analysis and empirical results demonstrate that SEL enhances class balance, sharpens decision boundaries, and strengthens decision regions. SEL is a plug-and-play training framework that can be seamlessly integrated into model training or classifier adjustment stages, making it highly adaptable and compatible with existing methods and facilitating further performance improvements. Extensive experiments on various long-tailed benchmarks demonstrate SEL's superiority. + +# 1. Introduction + +With the success of deep learning in various Computer Vision (CV) tasks, real-world applications of automated systems have become more prevalent, such as autonomous driving [49], medical diagnosis [11], security surveillance [14], and so on. This success is largely due to the availability of large-scale labeled datasets, such as ImageNet [30], and the development of deep learning models that can learn + +![](images/d682e41c9907346c428374f717708c4048919ba1fbd5fab3faf845311d005759.jpg) +Training Samples +(a) Existing methods. + +Test Samples +Test Samples +![](images/e6a7ec404eeb79a47528cb4c447808ee26ceac88404adfcdfd53a3ce5de24fe9.jpg) +$\bigcirc$ Test sample + +(b) Existing methods trained with SEL. +Figure 1. Comparison of decision regions and boundaries produced by (a) existing methods and (b) proposed SEL. +![](images/f1f5bded2fef2c3b95d8429afa9d6eab4feafea71c125260c8612a6d5ba1abfc.jpg) +$\bullet$ Training sample $\triangle$ Exploratory example Class boun + +complex patterns from balanced data [1]. However, real-world data often exhibit long-tailed distributions with heavy class imbalance, where head classes have a massive number of samples, while tail classes are associated with only a few samples [46]. Such data imbalance poses a significant challenge for developing unbiased automated recognition systems, as deep learning models tend to be biased towards head classes and disregard tail classes, which are often the most important ones in practice [28]. + +To solve this challenge, a multitude of studies have been proposed in recent years [46], including re-sampling [44, 47], logit adjustment [23, 38], data augmentation [27], classifier design [40], representation learning [22, 52], class-sensitive learning [13, 28], transfer learning [25, 29], ensemble learning [8, 39, 48] and decoupled training [17, 18, 41] methods. Despite their differences in principle, they all enhance the model's representation capabilities and de + +velop an optimal classifier with clear decision boundaries to separate different classes. On the one hand, effective representation learning can alleviate the issue of imbalanced data distribution and create a balanced embedding space; on the other hand, robust classifiers can improve the recognition capability of tail classes. + +Thanks to recent advances in self-supervised learning and pre-training, foundational models now learn representations much more effectively. In the pre-trained era, imbalanced data primarily affects the classifier, where sufficient samples of head classes occupy decision space, while tail classes suffer from sparse, obscure decision regions. As a result, the potential of pre-trained foundational models is not fully released. During the inference stage, while informative representations of tail classes are obtained, they often differ from those in the training set. This discrepancy erodes the well-defined margins of the classifier [25], leading to the misclassification of samples from tail classes as head classes, as shown in Fig. 1(a). Recently, numerous methods have been developed to optimize the classifier. Representatively, BCL [52] and TSC [28] use Supervised Contrastive Learning (SCL) to tighten intra-class spaces and enlarge inter-class margins, significantly improving decision boundaries between classes. Furthermore, MixUp [45] and its variants [23, 42] are good at generating diversity samples that are beneficial for improving decision regions of tail classes. However, their goal of smoothing the decision space increases boundary ambiguity, which hinders classification since errors often occur at boundaries. + +Inspired by evolutionary algorithms like the Optimal Foraging Algorithm (OFA) [51], we attempt to tackle biased decision regions and boundaries in long-tailed recognition through a space exploration perspective. To this end, we propose Supervised Exploratory Learning (SEL), a simple yet effective framework that seamlessly integrates with existing methods to enhance their performance. SEL leverages an adaptive OFA operator to synthesize exploratory examples, compensating for insufficient tail-class data and enabling balanced decision regions and boundaries, as shown in Fig. 1(b). To be specific, SEL is composed of two modules: Class-biased Complement (CbC) and Fitness-weighted Sampling (FwS). The former is responsible for sampling more tail-class samples as target samples to balance the data distribution, while the latter selects more confident samples from adjacent classes as candidate samples to facilitate the formation of clear decision boundaries between classes. Hence, the synthesized exploratory examples are diverse and informative in representing the distribution of unseen samples in reality. SEL is a plug-and-play training framework compatible with existing long-tailed recognition methods, requiring no backbone modifications or additional parameters. The main contributions of this paper are summarized as follows: + +- We attempt to address the challenge of long-tailed data from the perspective of space exploration. An adaptive OFA operator is employed to synthesize exploratory examples, which are informative and diverse to represent the distribution of unseen samples in reality. +- We propose a novel SEL framework that enables existing long-tailed recognition models to achieve balanced decision regions and clear decision boundaries, thereby improving their performance without altering network structure or increasing model parameters. +- Rational analysis proves the effectiveness of SEL in improving the classification of tail classes. Considerable performance improvements of existing long-tailed recognition methods verify the effectiveness of SEL. + +# 2. Related Work + +Data Augmentation (DA) [50] is an effective method in improving model robustness, particularly in scenarios of data scarcity or imbalance. In this section, we review recent advances in DA for long-tailed recognition, categorized into explicit and implicit methods. + +# 2.1. Explicit Data Augmentation + +Traditional DA techniques, like flipping, rotating, cropping, and padding, effectively improve robustness on long-tailed datasets and mitigate overfitting by creating new image-label pairs consistent with the training data [9, 34]. More recently, Population-Based Augmentation [10], AutoAugment [2], and Randaugment [3] have emerged to automatically choose optimal DA strategies according to characteristics of datasets, which have demonstrated their efficacy in long-tailed data [4, 21, 33]. Similarly, Wang et al. [37] generated images for tail classes using encoder variation information learned from head classes. Zada et al. [43] utilizes pure noise images as samples for tail classes. + +Although generating new images is an effective data augmentation method, it is time-consuming and computationally expensive. Moreover, the diversity of the generated images is limited, as they are created through simple transformations of the original training data. + +# 2.2. Implicit Data Augmentation + +Different from explicit DA, implicit DA generates new samples by interpolating existing samples in the feature space, which can effectively improve the model's generalization ability. Representatively, MixUp [45] introduced complexity control to unexplored regions in the data space by linearly interpolating discrete sample points, effectively reducing the generalization error of the model. Subsequently, CutMix [42], an extension of MixUp, augmented data by inserting a patch from one image into another, with labels blended according to the patch areas. This method can + +transfer context-rich background information from dominant classes to less frequent ones by combining foreground patches from the latter with background images from the former [32]. Manifold MixUp [36] further developed this concept by applying linear interpolation to the features of input images, which proved to be more effective for long-tail learning. MetaSAug [26] learns class covariance matrices from a small, balanced validation set to augment tail classes semantically. Recently, H2T [25] improved diversity by replacing part of the tail class feature map with that of the head classes. + +These findings highlight the superiority of implicit DA methods in enhancing model generalization on long-tailed data. Following this line, we propose a novel implicit DA technique, from the perspective of space exploration, to augment the diversity of training data, especially for tail classes, and improve the robustness of the long-tailed recognition model's performance. + +# 3. Methodology + +# 3.1. Definition + +For image classification, $\{x,y\}$ represents a sample consisting of an image $x$ and its corresponding label $y\in \mathcal{V}$ , where $\mathcal{V} = \{1,2,\dots,C\}$ is the set of classes. In the long-tailed scenario, different classes have varying sample sizes. The imbalance ratio $\rho = \frac{n_{\mathrm{max}}}{n_{\mathrm{min}}}$ measures the skewness of the data distribution, where $n_{\mathrm{max}}$ and $n_{\mathrm{min}}$ represent the maximum and minimum number of training samples across classes. Typically, training samples for head classes are significantly greater than those for tail classes. + +Existing long-tailed recognition methods usually learn a complex function mapping from the input space $\mathcal{X}$ to the target space $\mathcal{Y}$ . This function is usually implemented as the composition of an encoder $f: \mathcal{X} \to \mathcal{Z} \in \mathbb{R}^d$ and a linear classifier $w: \mathcal{Z} \to \mathcal{Y}$ . Both are crucial for the final classification accuracy [46]. + +# 3.2. Motivation + +Firstly, the poor performance of the tail classes is primarily due to their lack of training samples. A straightforward solution is to increase the number of tail class samples through methods such as re-sampling [44] or to emphasize their importance through re-weighting [47]. Although these methods are effective, they do not generate new, diverse samples for tail classes different from the training data, limiting the enhancement of generalization. Secondly, MixUp [45] and its variants [23, 36, 42] are good at generating diverse samples that are beneficial to improving the decision regions of the tail classes. However, on the one hand, head-class samples are more likely to be randomly selected to construct mixed samples, worsening the imbalance of training data. On the other hand, their goal of smoothing the de + +![](images/97d63d4038a2da45930095677aa2a619d1ffe2dee1d69c5b20acc64174a9a0c7.jpg) +(a) MixUp + +![](images/bcb92c6d87e4e3b7b1e2d1496346d32a03f02a9f3f27ef471ebf97eabc85d040.jpg) +(b) OFA +Figure 2. The differences and relationships among MixUp, OFA and SEL. MixUp linearly interpolates two random samples (typically from different classes, such as $z_{a}$ and $z_{c}$ ) to create a new sample, while OFA leverages the better individual $(z_{s'})$ to guide weaker individuals $(z_{s})$ toward better evolution. Combining their strengths, SEL gradually expands and reinforces minority decision regions by synthesizing adjacent exploratory examples. + +![](images/a02a9029b7da64dbdab100a1e32cb259f7075caef87d57e3fa37e7f4a726b97f.jpg) +(c) SEL + +cision space intensifies the boundary ambiguity, as shown in Fig. 2(a), where the mixed sample $z_{e}$ may fall into another class's region. Finally, methods like BCL [52] and TSC [28] use SCL [19] to tighten intra-class spaces and expand inter-class margins, significantly improving decision boundaries but reducing decision region robustness while increasing false positives risk. + +To address these issues simultaneously, we propose to synthesize exploratory examples to enhance the diversity of the training data and improve the robustness of the decision regions and boundaries. As shown in Fig. 2(b), OFA [51] is good at exploring the sample space by leveraging the better individual (evaluated by the task-specific fitness function) to guide weaker ones, promoting more effective evolution and optimization. Combining the strengths of MixUp and OFA, SEL views samples of the same class as a population and expands its decision region by synthesizing exploratory examples in uncharted spaces, as illustrated in Fig. 2(c). + +Note that MixUp is an intuitive and effective enhancement method applied at the input layer [7]. However, due to varying model structures and input patterns, unifying these methods is challenging. Consequently, most enhanced training frameworks [22, 25, 36] operate in the feature space, as most classification models first derive feature representations before classifying images. Similarly, SEL is also operated in the feature space, making it compatible with various models and methods. + +# 3.3. Supervised Exploratory Learning + +SEL is a compatible framework that can be seamlessly integrated into existing methods. In SEL, an adaptive OFA operator is employed to synthesize exploratory examples to supplement and balance the original training samples. + +# 3.3.1. Exploratory Example Synthesization + +Inspired by the space exploration process of evolutionary algorithms, we propose to synthesize exploratory examples using an adaptive OFA operator. The OFA [51], a metaheuristic algorithm introduced in 2017, is characterized by + +its simple structure, strong global search efficiency, and proven affine invariance [15]. It simulates animals' foraging behavior, where each individual follows better ones to explore the environment. In OFA and its variants, the operator is defined in dimension. Here, we extend its formulation to a vector operation. Specifically, we treat samples in the batch as populations, and synthesize exploratory examples by using the following formula: + +$$ +z _ {e} = z _ {s} + r \left(z _ {s ^ {\prime}} - z _ {s}\right) \tag {1} +$$ + +where $z_{s} \in \mathcal{Z}$ and $z_{s'} \in \mathcal{Z}$ denote the target and candidate samples, respectively. $r \in (0,1)$ is a coefficient that controls the exploration distance, which is randomly sampled during model training. Generally, OFA selects the candidate sample $z_{s'}$ with higher fitness to guide weaker individuals ( $z_{s}$ ) toward better evolution. To adapt OFA to the requirement of long-tailed recognition, we introduce two modules, CbC and FwS, to regulate the selection of target and candidate samples, respectively. + +# 3.3.2. Class-biased Complement (CbC) + +To alleviate the class imbalance problem, diverse exploratory examples are synthesized to complement the tail classes. We first count the number of samples in each class and then identify the additional samples required (i.e., exploratory examples) to balance the class distribution. As illustrated in Fig. 2(c), the target sample's label is assigned to the exploratory example, eliminating the decision boundary ambiguity issue present in MixUp. Hence, synthesizing varying numbers of exploratory examples for different classes becomes a task of selecting target samples. In practice, training samples with the target class are allowed to be selected as the target samples multiple times. Specifically, for a sample $z_{s}$ in the batch, it will be utilized as the target sample $k_{c}$ times if it belongs to class $c$ . + +$$ +k _ {c} = \operatorname {C e i l} \left(\frac {n _ {\max } - n _ {c}}{n _ {c}}\right) \tag {2} +$$ + +where $n_c$ denotes the sample number of class $c$ , and $n_{max}$ represents the sample size of the largest head class. $\text{Ceil}(\ast)$ is the ceiling function. + +In this way, the issue of class imbalance is addressed by synthesizing exploratory examples with different classes, ensuring that the number of samples in each class is balanced. The key challenges are guaranteeing the diversity of synthesized exploratory examples and the rationality of assigned labels, a crucial difference that distinguishes SEL from MixUp and its variants. As shown in Fig. 3, our solution is to identify neighboring classes and place exploratory examples in the regions between the target and adjacent classes, ensuring rational label assignment. + +# 3.3.3. Neighboring Class Recognition + +In MixUp and its variants, the label of the synthesized sample is determined by linear interpolation of the target and + +![](images/9d6540a18a942a52d8410d52e6c54ddcd393f8933f800cf372c471452129a3b0.jpg) +Figure 3. Synthesize exploratory examples by the adaptive OFA operator. More exploratory examples are generated for tail classes. + +candidate samples. This approach results in label misassignment if the synthesized sample falls within the decision region of another class. To avoid this situation, we propose selecting candidate samples from neighboring classes, and placing synthesized samples between the target and adjacent class regions. Specifically, for samples in the batch $\mathcal{B}$ , we calculate the center of each class and then determine the Euclidean distances between the target and the others. + +$$ +d \left(c, c ^ {\prime}\right) = \| \frac {1}{\left| \mathcal {B} ^ {c} \right|} \sum_ {z _ {i} \in \mathcal {B}} ^ {| \mathcal {B} |} z _ {i} \cdot \mathbb {I} _ {y _ {i} = c} - \frac {1}{\left| \mathcal {B} ^ {c ^ {\prime}} \right|} \sum_ {z _ {i} \in \mathcal {B}} ^ {| \mathcal {B} |} z _ {i} \cdot \mathbb {I} _ {y _ {i} = c ^ {\prime}} \| _ {2} \tag {3} +$$ + +where $\mathcal{B}^c$ and $\mathcal{B}^{c'}$ are sets of samples belong to class $c$ and $c'$ in the batch $\mathcal{B}$ , respectively. $\mathbb{I}_*$ is the indicator function, and $\| * \|_2$ denotes the L2 norm. Therefore, $d(c, c')$ denotes the distance between the centers of class $c$ and class $c'$ . The nearest $k_c'$ classes are chosen as neighboring classes for class $c$ , from which a neighbor is randomly selected, such as class $c'$ , and the candidate sample will be picked from $\mathcal{B}^{c'}$ based on their fitnesses, as done in OFA. + +# 3.3.4. Fitness-weighted Sampling (FwS) + +Following the principle of OFA, the prediction probability of the correct class is considered the fitness of the sample, and the sample with higher fitness is selected as the candidate sample. To reduce the operation complexity and increase the exploration randomness, we employ probability-based sampling to select the candidate sample. Specifically, the sampling probabilities are obtained by: + +$$ +p _ {s ^ {\prime}} = \operatorname {N o r m} \left(\left\{\hat {y} _ {i} ^ {c ^ {\prime}} \right\} _ {i = 1} ^ {\left| \mathcal {B} ^ {c ^ {\prime}} \right|}\right) \tag {4} +$$ + +where $c'$ represents the selected neighboring class, $\mathcal{B}^{c'}$ is the set of samples belong to class $c'$ in the batch $\mathcal{B}$ . $\hat{y}_i^{c'}$ is the predicted probability of the $i$ -th sample belonging to its correct class $c'$ , and $\text{Norm}(*)$ is the normalization function that ensures the sum of $p_{s'}$ equals 1. + +In this way, for each target sample $z_{s}$ , a candidate sample $z_{s'}$ with higher confidence is sampled from the batch, thereby achieving exploratory examples with diversity. FwS is essential for the validity of SEL, which is why we use the adaptive OFA operator to synthesize exploratory examples. The rational analysis is elaborated later. + +# 3.3.5. Training with SEL Framework + +SEL is a plug-and-play training framework that can be seamlessly integrated into the training process of the original model. The optimization objective of SEL is to correctly classify the exploratory examples, which is equivalent to maximizing the predicted probability of the correct class. + +$$ +\mathcal {L} _ {S E L} = - \frac {1}{C} \sum_ {c = 1} ^ {C} \left(\frac {1}{k _ {c}} \sum_ {i = 1} ^ {k _ {c}} \log \left(\hat {y} _ {i} ^ {c}\right)\right) \tag {5} +$$ + +where $\hat{y}_i^c$ denotes the predicted probability of the $i$ -th exploratory example belonging to the $c$ -th class, and $k_c$ is the number of exploratory examples synthesized for class $c$ . + +Finally, the overall optimization objective is defined as: + +$$ +\mathcal {L} = (1 - \lambda) \mathcal {L} _ {\text {T a s k}} + \lambda \mathcal {L} _ {\text {S E L}} \tag {6} +$$ + +where $\mathcal{L}_{Task}$ is the optimization objective of the original method, and $\lambda$ is a weighting coefficient that controls the impact of SEL loss. + +# 3.4. Rationale Analysis + +The exploitation process of OFA has proved to be effective and rational [51]. In this section, we provide a rational analysis of SEL. For clarity, we denote the weights of the linear classifier as $w = \{w_{c}\}_{c = 1}^{C}$ , where $C$ is the number of classes, and $w_{c} \in \mathbb{R}^{d \times 1}$ denotes the weights of class $c$ , $d$ is the feature dimension. For a sample $z \in \mathbb{R}^{d \times 1}$ , the predicted probability of class $c$ can be expressed as: + +$$ +\hat {y} ^ {c} = \frac {e ^ {w _ {c} ^ {T} z}}{\sum_ {c ^ {\prime} = 1} ^ {C} e ^ {w _ {c ^ {\prime}} ^ {T} z}} \tag {7} +$$ + +Hence, $\hat{y}^c\propto w_c^T z$ . The classifier attempts to assign a higher probability to the target class, i.e., $\hat{y}^c >\hat{y}^{c'}$ , where $c$ and $c^\prime$ denote the target and other classes, respectively. + +Without loss of generality, the selection of the target sample $(z_{s})$ and candidate sample $(z_{s}^{\prime})$ can be divided into two cases, as summarized in Tab. 1. We use the subscripts $t$ (tail) and $h$ (head) to represent their relative relation. Generally, the first case typically dominates, as the CbC module selects more tail-class targets, while the long-tailed distribution increases the chance of candidate samples being from head classes. Additionally, since exploratory examples share labels with targets, the optimization aims to challenge the model's bias toward head classes. + +For the first case, the goal of $\hat{y}_e^t >\hat{y}_e^h$ is equivalent to: + +$$ +w _ {t} ^ {T} z _ {e} - w _ {h} ^ {T} z _ {e} > 0 \Rightarrow \tag {8} +$$ + +$$ +w _ {t} ^ {T} \left(z _ {t} + r \left(z _ {h} - z _ {t}\right)\right) - w _ {h} ^ {T} \left(z _ {t} + r \left(z _ {h} - z _ {t}\right)\right) > 0 +$$ + +After expanding and transforming the inequality, we have: + +$$ +\begin{array}{l} w _ {t} ^ {T} z _ {t} - w _ {h} ^ {T} z _ {t} > \frac {r}{1 - r} \left(w _ {h} ^ {T} z _ {h} - w _ {t} ^ {T} z _ {h}\right) \tag {9} \\ \Rightarrow \hat {y} _ {t} ^ {t} - \hat {y} _ {t} ^ {h} > \frac {r}{1 - r} (\hat {y} _ {h} ^ {h} - \hat {y} _ {h} ^ {t}) \\ \end{array} +$$ + +
Casezszs'zeGoal
1ztzhzt + r(zh - zt)ˆy e t >ˆy e h
2zhztzh + r(zt - zh)ˆy e h >ˆy e t
+ +Table 1. Two types of synthesized exploratory examples. + +where $\hat{y}_h^h$ and $\hat{y}_h^t$ represent probabilities of the candidate sample $(z_h)$ belonging to the head class and tail classes, respectively. The classifier favors head classes $(\hat{y}_h^h - \hat{y}_h^t > 0)$ , while the FwS module reinforces this by selecting high-confidence candidates. Therefore, the right item is tend to be positive, and $\hat{y}_t^t > \hat{y}_t^h$ is achieved, meaning that optimizing tail-class exploratory examples effectively improves the classification accuracy of original tail-class samples. + +Similarly, for the second case, the optimization goal of SEL is to make $\hat{y}_e^h >\hat{y}_e^t$ yielding: + +$$ +\hat {y} _ {h} ^ {h} - \hat {y} _ {h} ^ {t} > \frac {r}{1 - r} \left(\hat {y} _ {t} ^ {t} - \hat {y} _ {t} ^ {h}\right) \tag {10} +$$ + +where $\hat{y}_t^t$ and $\hat{y}_t^h$ represent probabilities of the candidate sample $(z_t)$ belonging to the tail class and head class, respectively. Although the classifier tends to misclassify tail-class samples as head classes, the FwS mitigates this by selecting better candidates, which possess higher probabilities in the correct classes. Hence, optimizing the head-class exploratory examples does not cause significant damage to the original head sample. + +# 4. Experiments + +# 4.1. Dataset + +CIFAR-10/100-LT [5] are the subsets of CIFAR-10/100 [20]. Both contain 50,000 training images and 10,000 validation images, each of size $32 \times 32$ , with 10 and 100 classes, respectively. The imbalance factors used in the experiment are set to 100, 50, and 10. + +ImageNet-LT [30] is a long-tailed version of ImageNet [6] by sampling a subset following the Pareto distribution with power value $\alpha = 0.64$ . It consists of $115.8\mathrm{K}$ images of 1000 classes in total, with 1280 to 5 images per class. + +iNaturalist 2018 [12] is a large-scale dataset containing 437.5K images from 8,142 classes. It is long-tailed by nature, with an extremely imbalanced distribution. + +# 4.2. Basic Setting + +# 4.2.1. Compared Methods + +To evaluate the effectiveness of our proposed SEL, we integrated it into several existing methods introduced in recent years, including re-balanced techniques (IB [31], GCL [23], KPS [24]), contrastive learning (BCL [52]), feature compression (FCC [22]), and data augmentation strategies + +
MethodsCIFAR-100-LTCIFAR-10-LT
AvenueImbalance Ratio ρ10050101005010
-ResNet-32 [9]39.9646.0156.6671.9675.9585.92
+ MixUp [45]41.55↑1.5947.11↑1.1057.18↑0.5273.43↑1.4777.19↑1.2486.37↑0.45
+ SEL44.53↑4.5749.64↑3.8559.17↑2.5175.84↑3.8879.81↑3.8686.78↑0.86
ICCV 2021IB [31]40.5046.3156.7575.1179.9588.01
+ SEL42.34↑1.8447.95↑1.6457.57↑0.8275.04↓0.0779.81↓0.1488.13↑0.12
CVPR 2022BCL [52]51.7656.5167.9083.6186.1390.10
+ SEL52.30↑0.5457.25↑0.7468.43↑0.5384.44↑0.8386.31↑0.1890.26↑0.16
CVPR 2022GCL [23]46.5051.7261.7980.5684.7489.65
+ SEL47.89↑1.3952.74↑1.0262.41↑0.6281.86↑1.3085.13↑0.3989.78↑0.13
CVPR 2023FCC+CE [22]40.2045.9357.8073.8079.5787.75
+ SEL42.33↑2.1348.05↑2.1258.75↑0.9577.88↑4.0882.10↑2.5388.80↑1.05
CVPR 2023GLMC [7]53.9158.8768.0783.6886.9091.16
+ SEL56.48↑2.5761.13↑2.2670.75↑2.6885.40↑1.7288.57↑1.6792.83↑1.67
TPAMI 2023KPS [24]41.9747.9259.5982.3284.2989.10
+ SEL44.01↑2.0447.80↓0.1260.03↑0.4483.48↑1.1684.49↑0.2089.34↑0.24
AAAI 2024H2T+CE [25]42.2747.5858.2479.9182.8088.77
+ SEL42.45↑0.1847.80↑0.2258.66↑0.4280.43↑0.5282.80↓0.0088.79↑0.02
+ +Table 2. Comparisons between raw methods and their SEL-enhanced counterparts on CIFAR-LT datasets. + +(H2T [25], GLMC [7]). These methods were chosen not only for their representativeness but also for their publicly available code and configurations, ensuring ease of reproducibility. Most employ a two-stage training strategy, with the first stage dedicated to representation learning and the second to classification. Additionally, we use ResNet models as baselines to directly compare the performance improvements achieved by MixUp and SEL. + +# 4.2.2. Implementation Details + +We reproduce the compared methods using their official codes and configurations, setting the random seed to 2024 for fair comparison. The base ResNet models are configured identically to those in Li et al. [23]. To simplify application, we first determine the optimal way to apply SEL on different datasets by using the base model, i.e., ResNet-32 for CIFAR-10/100-LT and ResNet-50 for others, then replace it with corresponding models to assess SEL's effectiveness on these models. Specifically, for CIFAR-100-LT, SEL is applied during the final 20 training epochs. For other datasets, SEL is applied by further fine-tuning the classifier for 20 epochs on CIFAR-10-LT and 10 epochs on ImageNet-LT and iNaturalist 2018. The weight coefficient $\lambda$ in Eq. (6) is set to 0.9, with the number of neighboring classes $k_c'$ set to 2, 5, 10, and 50 for CIFAR-10-LT, CIFAR-100-LT, ImageNet-LT, and iNaturalist 2018, respectively. Top-1 accuracy serves as the evaluation metric. + +# 4.3. Improvement of Existing Methods + +# 4.3.1. Results on CIFAR-10/100-LT + +For CIFAR-10/100-LT, the results with and without SEL are presented in Tab. 2. Across 48 experiment groups, our proposed SEL significantly improves 44 of them, with an average increase of $1.37\%$ (4.57% max and $0.02\%$ min). Extensive experimental results fully verify the effectiveness of our method on long-tailed recognition, particularly for scenarios with high imbalance ratios. Generally, integrating MixUp into the vanilla ResNet-32 model significantly enhances performance, highlighting the effectiveness of data augmentation. However, our proposed SEL achieves even greater improvements across all scenarios, and even surpasses existing methods. For example, ResNet-32+SEL outperforms IB on both datasets except for CIFAR-10-LT-10, and surpasses GCL, KPS, and H2T on CIFAR-100-LT, though it underperforms on CIFAR-10-LT. + +Additionally, SEL significantly enhances the performance of existing methods, with particularly pronounced improvements in high-imbalance scenarios. For example, SEL is notably effective on KPS at an imbalance ratio of 100 but shows limited impact at other imbalance ratios. In addition, GLMC+SEL and FCC+SEL significantly outperform their original counterparts on both datasets, with average improvements of $2.10\%$ and $2.14\%$ , respectively. For the re-balanced methods IB, GCL, and KPS, SEL yields + +
MethodManyMediumFewAll
ResNet-5063.3833.7426.0444.80
+ Mixup62.9735.4428.5845.71↑0.91
+ SEL64.1438.4131.2747.86↑3.06
BCL65.8353.1636.2755.67
+ SEL65.4554.8236.5356.27↑0.60
GCL65.0452.7735.6955.51
+ SEL65.1353.9439.1856.12↑0.61
FCC+CE65.5252.1833.8354.68
+ SEL64.7454.0035.6655.60↑0.92
GLMC69.4252.4930.4156.17
+ SEL68.6754.3738.2857.24↑1.07
+ +Table 3. Comparisons on the ImageNet-LT dataset. + +
MethodManyMediumFewAll
ResNet-5074.8968.2461.4265.95
+ Mixup71.3469.5363.8867.29↑1.34
+ SEL73.0470.3866.7469.10↑3.15
BCL73.4771.9968.8270.80
+ SEL71.9973.0469.2171.32↑0.52
GCL73.7572.2868.3470.76
+ SEL72.6373.5870.0572.00↑1.24
FCC+CE72.7871.0967.4169.70
+ SEL71.1772.4169.5571.09↑1.39
GLMC76.0473.6971.7773.14
+ SEL74.6875.1473.6674.51↑1.37
+ +greater improvements on CIFAR-100-LT than on CIFAR-10-LT, demonstrating its superiority in scenarios with more classes. However, SEL is less effective on BCL and H2T. BCL uses contrastive learning to reinforce balanced decision regions, aligning with SEL's objective. The difference is that SEL further enhances the robustness of decision regions, resulting in additional improvements. H2T is also an evolutionary algorithm-inspired method that generates diverse samples through cross-mutation, following the principles of Genetic Algorithm (GA), thus indicating the fundamental difference between SEL and H2T stems from the distinction between OFA and GA. The slight improvement of H2T+SEL further validates SEL's effectiveness in generating diverse samples, aligning with H2T's principles. + +# 4.3.2. Results on ImageNet-LT and iNaturalist 2018 + +For ImageNet-LT and iNaturalist 2018, experimental results are listed in Tab. 3 and Tab. 4, respectively. Following pre + +Table 4. Comparisons on the iNaturalist 2018 dataset. + +
CbCFwSTop-1 Accuracy
ρ =100ρ =10
××39.9656.66
×✓40.26↑0.3056.93↑0.27
✓×43.44↑3.4857.84↑1.18
✓✓44.53↑4.5759.17↑2.51
+ +Table 5. Ablations of key modules. + +vious works [7, 16], also report accuracy across three class splits: Many $(n_c \geq 100)$ , Medium $(20 < n_c < 100)$ , and Few $(n_c \leq 20)$ . Overall, SEL outperforms MixUp on both datasets, and consistently enhances the compared methods, achieving an average improvement of $1.25\%$ (max $3.06\%$ , min $0.61\%$ ) on ImageNet-LT and $1.53\%$ (max $3.15\%$ , min $0.52\%$ ) on iNaturalist 2018. As seen, training with SEL typically improves the performance of Few classes while slightly reducing that of Many classes, aligning with the rationale proved in Sec. 3.4. However, the greater gain in Few classes outweighs the slight decline in Many classes, leading to an overall improvement in the All classes. + +# 4.4. Further Analysis + +To further analyze the principles of SEL quantitatively, we conducted a series of ablation studies on the CIFAR-10/100-LT datasets using ResNet-32 [9] as the base model. + +# 4.4.1. Visualization of Decision Regions + +Fig. 4 visualizes the decision regions and boundaries of different classes using t-SNE [35] across various methods. These models are trained on the CIFAR-100-LT with an imbalance ratio of 100 and tested on a balanced test set. For comparison, we selected four classes from CIFAR-100 test set for visualization: where apple and aquarium fish are the head classes with the most training samples, and willow tree and wolf are the tail classes with the fewest training samples. The decision boundaries are derived using the SVM algorithm. From Fig. 4, we can observe that SEL results in more balanced decision regions for both head and tail classes, aligning with our motivation. + +# 4.4.2. Role of Key Modules + +The CbC and FwS modules are the core components of SEL. To examine their roles, we conduct ablation experiments on CIFAR-100-LT with imbalance ratios of 100 and 10. Tab. 5 presented the experimental results, showing that the CbC module significantly enhances the baseline method, with improvements of $3.48\%$ $(39.96\% \rightarrow 43.44\%)$ and $1.18\%$ $(56.66\% \rightarrow 57.84\%)$ for imbalance ratios of 100 and 10, respectively. These results highlight the CbC module's crucial role in balancing data distribution, enabling the model to learn more effectively from tail classes. However, + +(a) CE +![](images/716691e4944b523ec5b595f6875a736a441d44aa5fe965579e1542e1fccd8a35.jpg) +apple aquarium fish willow tree wolf + +![](images/8eab3c4522ffdf4793ba371a8356e68ec5cba34e59fe4c0ccc9c3880429927fd.jpg) +(b) MixUp + +![](images/8d700e333e0d6188682ac844bd4b2f8788258555d41a092de5028b5a3bb9c4f6.jpg) +(c) SEL + +![](images/1b4246f60e2705712b128da47d2aa8a22f9e635de2fec3633c1840ca93c4884d.jpg) +Figure 4. Visualization of decision regions and boundaries on the CIFAR-100 test set (each class contains 100 test samples), where apple and aquarium fish are head classes with the most training samples, and willow tree and wolf are tail classes with the fewest training samples. +(a) CIFAR-100-LT + +![](images/4c2c44f2c6f7026de5a3a7d334556998300dedc4211bfc5d1bd98e4bcfb48945.jpg) +(b) CIFAR-10-LT + +![](images/9f4076fa1f5e32e733a4046011f0c2730e0b25bc3987b2e6576f6ba263458504.jpg) +Figure 5. Changes with different neighboring class numbers. +(a) start epochs +Figure 6. Changes with different parameters. + +![](images/ba8e8d79274170b70c7a7f9f52d3b7bb01bb264004f60b7fee9f941e8db10d9f.jpg) +(b) weight $(\lambda)$ + +the FwS module has a minimal impact without the CbC module, yielding only $0.30\%$ and $0.27\%$ improvements for imbalance ratios of 100 and 10, respectively. Its effectiveness is fully realized when combined with the CbC module, leading to a significant boost in model performance. + +# 4.4.3. Impact of different parameters + +In this section, we conduct exploratory experiments on the impact of the number of neighboring classes $(k_c^{\prime})$ in the FwS module. Experimental results are summarized in Fig. 5, where $k_{c}^{\prime}$ is varied from 1 to 10 and relative accuracy (values are adjusted by subtracting their mean) is utilized for comparison. As seen, the model performs better with a higher $k^{\prime}_{c}$ on CIFAR-100-LT but with a lower $k^{\prime}_{c}$ on CIFAR-10LT. This difference arises from the number of classes, as CIFAR-100-LT has 100 classes, whereas CIFAR-10-LT has only 10. Based on these observations, we empirically set $k^{\prime}_{c}$ to 5 for CIFAR-100-LT and 2 for CIFAR-10-LT. + +# 4.4.4. When to Start SEL Training + +In this section, we explore the optimal starting point for SEL training to maximize computational efficiency, as its synthesis of additional exploratory examples increases computational costs. Specifically, we integrate SEL into the training process at different epochs, ranging from 100 to 200, as shown in Fig. 6(a). Here, starting at 200 epochs indicates a two-stage training mode, where SEL is applied with an additional 20 epochs after the original training process. Similarly, relative accuracy is used as the ordinate for clearly showing the parameter's impact. It is obvious that the model's performance steadily improves as the SEL starting point increases, with most datasets peaking at 180 epochs, except for CIFAR-10-LT-100 and CIFAR-10-LT-10. Therefore, we set the starting point to 180 epochs for CIFAR-100-LT and 200 epochs for CIFAR-10-LT. + +Building on the optimal starting points, we further examine the impact of parameter $\lambda$ in Eq. (6). Fig. 6(b) illustrates the performance variation with $\lambda$ ranges from 0.0 to 1.0. Note that $\lambda = 0.0$ indicates the absence of SEL, while $\lambda = 1.0$ signifies its full integration. For most datasets, as $\lambda$ increases, the model's performance gradually improves, peaking at $\lambda = 0.9$ before declining. However, for CIFAR-10-LT-10, the simplest dataset, performance remains relatively stable, with a peak at $\lambda = 0.6$ . + +# 5. Conclusion + +In this paper, we propose SEL, a simple yet effective training framework that addresses long-tailed challenges through the space exploration perspective. SEL is a plug-and-play method that boosts performance without modifying the model or adding parameters. It uses an adaptive OFA operator to generate exploratory examples that capture unseen tail-class features, enhancing decision regions and boundaries. The effectiveness of SEL is validated mathematically through rational analysis and empirically through experiments on various long-tailed datasets and methods. + +# Acknowledgment + +This work is supported by the Solfeggio ear training intelligent robot and cloud platform research and development project for music education (No.2024CXY0102), the 3D visualization digital twin integrated control system (No.2023CXY0111), and the public technology service platform project of Xiamen City (No.3502Z20231043). + +# References + +[1] Yoshua Bengio, Yann Lecun, and Geoffrey Hinton. Deep learning for ai. Communications of the ACM, 64(7):58-65, 2021. 1 +[2] Ekin D. Cubuk, Barret Zoph, Dandelion Mané, Vijay Vasudevan, and Quoc V. Le. Autoaugment: Learning augmentation strategies from data. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 113-123, 2019. 2 +[3] Ekin Dogus Cubuk, Barret Zoph, Jon Shlens, and Quoc Le. Randaugment: Practical automated data augmentation with a reduced search space. In Advances in Neural Information Processing Systems, pages 18613-18624. Curran Associates, Inc., 2020. 2 +[4] Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, and Jiaya Jia. Parametric contrastive learning. In 2021 IEEE/CVF International Conference on Computer Vision, pages 695-704, 2021. 2 +[5] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge J. Belongie. Class-balanced loss based on effective number of samples. In IEEE Conference on Computer Vision and Pattern Recognition, pages 9268-9277, 2019. 5 +[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 5 +[7] Fei Du, Peng Yang, Qi Jia, Fengtao Nan, Xiaoting Chen, and Yun Yang. Global and local mixture consistency cumulative learning for long-tailed visual recognitions. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15814-15823, 2023. 3, 6, 7 +[8] Hao Guo and Song Wang. Long-tailed multi-label visual recognition by collaborative training on uniform and rebalanced samplings. In IEEE Conference on Computer Vision and Pattern Recognition, pages 15089-15098, 2021. 1 +[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2, 6, 7 +[10] Daniel Ho, Eric Liang, Xi Chen, Ion Stoica, and Pieter Abbeel. Population based augmentation: Efficient learning of augmentation policy schedules. In Proceedings of the 36th International Conference on Machine Learning, pages 2731-2741, 2019. 2 +[11] Gregory Holste, Yiliang Zhou, Song Wang, Ajay Jaiswal, Mingquan Lin, Sherry Zhuge, Yuzhe Yang, Dongkyun Kim, + +Trong-Hieu Nguyen-Mau, Minh-Triet Tran, Jaehyup Jeong, Wongi Park, Jongbin Ryu, Feng Hong, Arsh Verma, Yosuke Yamagishi, Changhyun Kim, Hyeryeong Seo, Myungjoo Kang, Leo Anthony Celi, Zhiyong Lu, Ronald M. Summers, George Shih, Zhangyang Wang, and Yifan Peng. Towards long-tailed, multi-label disease classification from chest x-ray: Overview of the cxr-It challenge. Medical Image Analysis, 97:103224, 2024. 1 +[12] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alexander Shepard, Hartwig Adam, Pietro Perona, and Serge J. Belongie. The inaturalist species classification and detection dataset. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, pages 8769-8778, 2018. 5 +[13] Chengkai Hou, Jieyu Zhang, Haonan Wang, and Tianyi Zhou. Subclass-balancing contrastive learning for long-tailed recognition. In 2023 IEEE/CVF International Conference on Computer Vision, pages 5372-5384, 2023. 1 +[14] Shuyan Hu, Wei Ni, Xin Wang, and Abbas Jamalipour. Disguised tailing and video surveillance with solar-powered fixed-wing unmanned aerial vehicle. IEEE Transactions on Vehicular Technology, 71(5):5507-5518, 2022. 1 +[15] ZhongQuan Jian and GuangYu Zhu. Affine invariance of meta-heuristic algorithms. Information Sciences, 576:37-53, 2021. 4 +[16] Yan Jin, Mengke Li, Yang Lu, Yiu-ming Cheung, and Hanzi Wang. Long-tailed visual recognition via self-heterogeneous integration with knowledge excavation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23695-23704, 2023. 7 +[17] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. In 8th International Conference on Learning Representations, 2020. 1 +[18] Bingyi Kang, Yu Li, Sa Xie, Zehuan Yuan, and Jiashi Feng. Exploring balanced feature spaces for representation learning. In 9th International Conference on Learning Representations, 2021. 1 +[19] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673, 2020. 3 +[20] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5 +[21] Jun Li, Zichang Tan, Jun Wan, Zhen Lei, and Guodong Guo. Nested collaborative learning for long-tailed visual recognition. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6939-6948, 2022. 2 +[22] Jian Li, Ziyao Meng, Daqian Shi, Rui Song, Xiaolei Dao, Jingwen Wang, and Hao Xu. FCC: feature clusters compression for long-tailed visual recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24080-24089, 2023. 1, 3, 5, 6 +[23] Mengke Li, Yiu-Ming Cheung, and Yang Lu. Long-tailed visual recognition via gaussian clouded logit adjustment. In + +2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6919-6928, 2022. 1, 2, 3, 5, 6 +[24] Mengke Li, Yiu-Ming Cheung, and Zhikai Hu. Key point sensitive loss for long-tailed visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4): 4812-4825, 2023. 5, 6 +[25] Mengke Li, Zhikai Hu, Yang Lu, Weichao Lan, Yiu-ming Cheung, and Hui Huang. Feature fusion from head to tail for long-tailed visual recognition. In Thirty-Eighth AAAI Conference on Artificial Intelligence, Thirty-Sixth Conference on Innovative Applications of Artificial Intelligence, Fourteenth Symposium on Educational Advances in Artificial Intelligence, pages 13581-13589, 2024. 1, 2, 3, 6 +[26] Shuang Li, Kaixiong Gong, Chi Harold Liu, Yulin Wang, Feng Qiao, and Xinjing Cheng. Metasaug: Meta semantic augmentation for long-tailed visual recognition. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5212-5221, 2021. 3 +[27] Shuang Li, Kaixiong Gong, Chi Harold Liu, Yulin Wang, Feng Qiao, and Xinjing Cheng. Metasaug: Meta semantic augmentation for long-tailed visual recognition. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5212-5221, 2021. 1 +[28] Tianhong Li, Peng Cao, Yuan Yuan, Lijie Fan, Yuzhe Yang, Rogerio Feris, Piotr Indyk, and Dina Katabi. Targeted supervised contrastive learning for long-tailed recognition. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6918, 2022. 1, 2, 3 +[29] Bo Liu, Haoxiang Li, Hao Kang, Gang Hua, and Nuno Vasconcelos. Gistnet: a geometric structure transfer network for long-tailed recognition. In 2021 IEEE/CVF International Conference on Computer Vision, pages 8189-8198, 2021. 1 +[30] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X. Yu. Large-scale long-tailed recognition in an open world. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2537-2546, 2019. 1, 5 +[31] Seulki Park, Jongin Lim, Younghan Jeon, and Jin Young Choi. Influence-balanced loss for imbalanced visual classification. In 2021 IEEE/CVF International Conference on Computer Vision, pages 715-724, 2021. 5, 6 +[32] Seulki Park, Youngkyu Hong, Byeongho Heo, Sangdoo Yun, and Jin Young Choi. The majority can help the minority: Context-rich minority oversampling for long-tailed classification. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6877-6886, 2022. 3 +[33] Jiawei Ren, Cunjun Yu, Shunan Sheng, Xiao Ma, Haiyu Zhao, Shuai Yi, and Hongsheng Li. Balanced meta-softmax for long-tailed visual recognition. In Advances in Neural Information Processing Systems 33, 2020. 2 +[34] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott E. Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In 2015 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-9, 2015. 2 +[35] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 7 + +[36] Vikas Verma, Alex Lamb, Christopher Beckham, Amir Najafi, Ioannis Mitlagkas, David Lopez-Paz, and Yoshua Bengio. Manifold mixup: Better representations by interpolating hidden states. In Proceedings of the 36th International Conference on Machine Learning, pages 6438-6447, 2019. 3 +[37] Jianfeng Wang, Thomas Lukasiewicz, Xiaolin Hu, Jianfei Cai, and Zhenghua Xu. Rsg: A simple but effective module for learning imbalanced datasets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3784-3793, 2021. 2 +[38] Peng Wang, Kai Han, Xiu-Shen Wei, Lei Zhang, and Lei Wang. Contrastive learning based hybrid networks for long-tailed image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 943-952, 2021. 1 +[39] Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu, and Stella X. Yu. Long-tailed recognition by routing diverse distribution-aware experts. In 9th International Conference on Learning Representations, 2021. 1 +[40] Tong Wu, Ziwei Liu, Qingqiu Huang, Yu Wang, and Dahua Lin. Adversarial robustness under long-tailed distribution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 8659-8668, 2021. 1 +[41] Shiyu Xuan and Shiliang Zhang. Decoupled contrastive learning for long-tailed recognition. In Thirty-Eighth AAAI Conference on Artificial Intelligence, Thirty-Sixth Conference on Innovative Applications of Artificial Intelligence, Fourteenth Symposium on Educational Advances in Artificial Intelligence, pages 6396-6403, 2024. 1 +[42] Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Seong Joon Oh, Youngjoon Yoo, and Junsuk Choe. Cutmix: Regularization strategy to train strong classifiers with localizable features. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6022-6031, 2019. 2, 3 +[43] Shiran Zada, Itay Benou, and Michal Irani. Pure noise to the rescue of insufficient data: Improving imbalanced classification by training on random noise images. In International Conference on Machine Learning, pages 25817-25833, 2022. 2 +[44] Yuhang Zang, Chen Huang, and Chen Change Loy. Fasa: Feature augmentation and sampling adaptation for long-tailed instance segmentation. In 2021 IEEE/CVF International Conference on Computer Vision, pages 3437-3446, 2021. 1, 3 +[45] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In 6th International Conference on Learning Representations, 2018. 2, 3, 6 +[46] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):10795-10816, 2023. 1, 3 +[47] Zizhao Zhang and Tomas Pfister. Learning fast sample reweighting without reward data. In 2021 IEEE/CVF International Conference on Computer Vision, pages 705-714, 2021. 1, 3 +[48] Boyan Zhou, Quan Cui, Xiu-Shen Wei, and Zhao-Min Chen. BBN: bilateral-branch network with cumulative learning for + +long-tailed visual recognition. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9716-9725, 2020. 1 +[49] Weitao Zhou, Zhong Cao, Nanshan Deng, Xiaoyu Liu, Kun Jiang, and Diange Yang. Dynamically conservative self-driving planner for long-tail cases. IEEE Transactions on Intelligent Transportation Systems, 24(3):3476-3488, 2023. 1 +[50] Chenguang Zhu, Yichong Xu, Xiang Ren, Bill Yuchen Lin, Meng Jiang, and Wenhao Yu. Knowledge-augmented methods for natural language processing. In Proceedings of the Sixteenth ACM International Conference on Web Search and Data Mining, pages 1228-1231, 2023. 2 +[51] Guang-Yu Zhu and Wei-Bo Zhang. Optimal foraging algorithm for global optimization. Applied Soft Computing, 51: 294–313, 2017. 2, 3, 5 +[52] Jianggang Zhu, Zheng Wang, Jingjing Chen, Yi-Ping Phoebe Chen, and Yu-Gang Jiang. Balanced contrastive learning for long-tailed visual recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6898-6907, 2022. 1, 2, 3, 5, 6 \ No newline at end of file diff --git a/2025/Supervised Exploratory Learning for Long-Tailed Visual Recognition/images.zip b/2025/Supervised Exploratory Learning for Long-Tailed Visual Recognition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6e0b4295fc567ce546b3eb6a5e0a4289c977c2b0 --- /dev/null +++ b/2025/Supervised Exploratory Learning for Long-Tailed Visual Recognition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:537687bcf1cc2bd5c52c8e4d64060854df4f338c3326f4748f7a340bc65b9934 +size 527462 diff --git a/2025/Supervised Exploratory Learning for Long-Tailed Visual Recognition/layout.json b/2025/Supervised Exploratory Learning for Long-Tailed Visual Recognition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5d14d5061f973f2dcd76f22f188bbf8574472ff8 --- /dev/null +++ b/2025/Supervised Exploratory Learning for Long-Tailed Visual Recognition/layout.json @@ -0,0 +1,10764 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 91, + 103, + 520, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 103, + 520, + 121 + ], + "spans": [ + { + "bbox": [ + 91, + 103, + 520, + 121 + ], + "type": "text", + "content": "Supervised Exploratory Learning for Long-Tailed Visual Recognition" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 124, + 142, + 485, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 142, + 485, + 171 + ], + "spans": [ + { + "bbox": [ + 124, + 142, + 485, + 171 + ], + "type": "text", + "content": "Zhongquan Jian" + }, + { + "bbox": [ + 124, + 142, + 485, + 171 + ], + "type": "inline_equation", + "content": "^{1,3,\\dagger}" + }, + { + "bbox": [ + 124, + 142, + 485, + 171 + ], + "type": "text", + "content": ", Yanhao Chen" + }, + { + "bbox": [ + 124, + 142, + 485, + 171 + ], + "type": "inline_equation", + "content": "^{2,\\dagger}" + }, + { + "bbox": [ + 124, + 142, + 485, + 171 + ], + "type": "text", + "content": ", Yancheng Wang" + }, + { + "bbox": [ + 124, + 142, + 485, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 124, + 142, + 485, + 171 + ], + "type": "text", + "content": ", Junfeng Yao" + }, + { + "bbox": [ + 124, + 142, + 485, + 171 + ], + "type": "inline_equation", + "content": "^{2,3,1,*}" + }, + { + "bbox": [ + 124, + 142, + 485, + 171 + ], + "type": "text", + "content": ", Meihong Wang" + }, + { + "bbox": [ + 124, + 142, + 485, + 171 + ], + "type": "inline_equation", + "content": "^{3,*}" + }, + { + "bbox": [ + 124, + 142, + 485, + 171 + ], + "type": "text", + "content": ", Qingqiang Wu" + }, + { + "bbox": [ + 124, + 142, + 485, + 171 + ], + "type": "inline_equation", + "content": "^{2,3,*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 121, + 171, + 489, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 171, + 489, + 185 + ], + "spans": [ + { + "bbox": [ + 121, + 171, + 489, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 121, + 171, + 489, + 185 + ], + "type": "text", + "content": "School of Computer and Data Science, Minjiang University, Fuzhou, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 178, + 185, + 432, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 185, + 432, + 198 + ], + "spans": [ + { + "bbox": [ + 178, + 185, + 432, + 198 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 178, + 185, + 432, + 198 + ], + "type": "text", + "content": "School of Film, Xiamen University, Xiamen, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 162, + 199, + 448, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 199, + 448, + 213 + ], + "spans": [ + { + "bbox": [ + 162, + 199, + 448, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 162, + 199, + 448, + 213 + ], + "type": "text", + "content": "School of Informatics, Xiamen University, Xiamen, China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 111, + 213, + 500, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 213, + 500, + 228 + ], + "spans": [ + { + "bbox": [ + 111, + 213, + 500, + 228 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 111, + 213, + 500, + 228 + ], + "type": "text", + "content": "School of Computing and Data Science, Xiamen University Malaysia, Malaysia" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 211, + 229, + 395, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 229, + 395, + 241 + ], + "spans": [ + { + "bbox": [ + 211, + 229, + 395, + 241 + ], + "type": "text", + "content": "{yao0010, wangmh, wuqq}@xmu.edu.cn" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 152, + 269, + 200, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 269, + 200, + 280 + ], + "spans": [ + { + "bbox": [ + 152, + 269, + 200, + 280 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 293, + 296, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 293, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 54, + 293, + 296, + 569 + ], + "type": "text", + "content": "Long-tailed data poses a significant challenge for deep learning models, which tend to prioritize accurate classification of head classes while largely neglecting tail classes. Existing techniques, such as class re-balancing, logit adjustment, and data augmentation, aim to enlarge decision regions of tail classes or achieve clear decision boundaries, leaving the robustness of decision regions underconsidered. This paper proposes a simple yet effective Supervised Exploratory Learning (SEL) framework to achieve these goals simultaneously from space exploration perspectives. SEL employs the adaptive Optimal Foraging Algorithm (OFA) to generate diverse exploratory examples, integrating Class-biased Complement (CbC) for balanced class distribution and Fitness-weighted Sampling (FwS) for space exploration. Both theoretical analysis and empirical results demonstrate that SEL enhances class balance, sharpens decision boundaries, and strengthens decision regions. SEL is a plug-and-play training framework that can be seamlessly integrated into model training or classifier adjustment stages, making it highly adaptable and compatible with existing methods and facilitating further performance improvements. Extensive experiments on various long-tailed benchmarks demonstrate SEL's superiority." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 591, + 135, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 135, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 135, + 604 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 611, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 295, + 696 + ], + "type": "text", + "content": "With the success of deep learning in various Computer Vision (CV) tasks, real-world applications of automated systems have become more prevalent, such as autonomous driving [49], medical diagnosis [11], security surveillance [14], and so on. This success is largely due to the availability of large-scale labeled datasets, such as ImageNet [30], and the development of deep learning models that can learn" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 318, + 277, + 433, + 343 + ], + "blocks": [ + { + "bbox": [ + 353, + 270, + 397, + 277 + ], + "lines": [ + { + "bbox": [ + 353, + 270, + 397, + 277 + ], + "spans": [ + { + "bbox": [ + 353, + 270, + 397, + 277 + ], + "type": "text", + "content": "Training Samples" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 318, + 277, + 433, + 343 + ], + "lines": [ + { + "bbox": [ + 318, + 277, + 433, + 343 + ], + "spans": [ + { + "bbox": [ + 318, + 277, + 433, + 343 + ], + "type": "image", + "image_path": "d682e41c9907346c428374f717708c4048919ba1fbd5fab3faf845311d005759.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 344, + 465, + 353 + ], + "lines": [ + { + "bbox": [ + 403, + 344, + 465, + 353 + ], + "spans": [ + { + "bbox": [ + 403, + 344, + 465, + 353 + ], + "type": "text", + "content": "(a) Existing methods." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 436, + 278, + 551, + 342 + ], + "blocks": [ + { + "bbox": [ + 477, + 270, + 510, + 277 + ], + "lines": [ + { + "bbox": [ + 477, + 270, + 510, + 277 + ], + "spans": [ + { + "bbox": [ + 477, + 270, + 510, + 277 + ], + "type": "text", + "content": "Test Samples" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 436, + 278, + 551, + 342 + ], + "lines": [ + { + "bbox": [ + 436, + 278, + 551, + 342 + ], + "spans": [ + { + "bbox": [ + 436, + 278, + 551, + 342 + ], + "type": "image", + "image_path": "e6a7ec404eeb79a47528cb4c447808ee26ceac88404adfcdfd53a3ce5de24fe9.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 476, + 356, + 507, + 362 + ], + "lines": [ + { + "bbox": [ + 476, + 356, + 507, + 362 + ], + "spans": [ + { + "bbox": [ + 476, + 356, + 507, + 362 + ], + "type": "text", + "content": "Test Samples" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 509, + 439, + 550, + 448 + ], + "lines": [ + { + "bbox": [ + 509, + 439, + 550, + 448 + ], + "spans": [ + { + "bbox": [ + 509, + 439, + 550, + 448 + ], + "type": "inline_equation", + "content": "\\bigcirc" + }, + { + "bbox": [ + 509, + 439, + 550, + 448 + ], + "type": "text", + "content": " Test sample" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 319, + 356, + 433, + 428 + ], + "blocks": [ + { + "bbox": [ + 319, + 356, + 433, + 428 + ], + "lines": [ + { + "bbox": [ + 319, + 356, + 433, + 428 + ], + "spans": [ + { + "bbox": [ + 319, + 356, + 433, + 428 + ], + "type": "image", + "image_path": "f1f5bded2fef2c3b95d8429afa9d6eab4feafea71c125260c8612a6d5ba1abfc.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 429, + 490, + 437 + ], + "lines": [ + { + "bbox": [ + 378, + 429, + 490, + 437 + ], + "spans": [ + { + "bbox": [ + 378, + 429, + 490, + 437 + ], + "type": "text", + "content": "(b) Existing methods trained with SEL." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 319, + 439, + 489, + 449 + ], + "lines": [ + { + "bbox": [ + 319, + 439, + 489, + 449 + ], + "spans": [ + { + "bbox": [ + 319, + 439, + 489, + 449 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 319, + 439, + 489, + 449 + ], + "type": "text", + "content": " Training sample " + }, + { + "bbox": [ + 319, + 439, + 489, + 449 + ], + "type": "inline_equation", + "content": "\\triangle" + }, + { + "bbox": [ + 319, + 439, + 489, + 449 + ], + "type": "text", + "content": " Exploratory example Class boun" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 313, + 457, + 555, + 480 + ], + "lines": [ + { + "bbox": [ + 313, + 457, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 555, + 480 + ], + "type": "text", + "content": "Figure 1. Comparison of decision regions and boundaries produced by (a) existing methods and (b) proposed SEL." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 506, + 555, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 506, + 555, + 615 + ], + "spans": [ + { + "bbox": [ + 312, + 506, + 555, + 615 + ], + "type": "text", + "content": "complex patterns from balanced data [1]. However, real-world data often exhibit long-tailed distributions with heavy class imbalance, where head classes have a massive number of samples, while tail classes are associated with only a few samples [46]. Such data imbalance poses a significant challenge for developing unbiased automated recognition systems, as deep learning models tend to be biased towards head classes and disregard tail classes, which are often the most important ones in practice [28]." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "To solve this challenge, a multitude of studies have been proposed in recent years [46], including re-sampling [44, 47], logit adjustment [23, 38], data augmentation [27], classifier design [40], representation learning [22, 52], class-sensitive learning [13, 28], transfer learning [25, 29], ensemble learning [8, 39, 48] and decoupled training [17, 18, 41] methods. Despite their differences in principle, they all enhance the model's representation capabilities and de" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 219, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 219, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 219, + 713 + ], + "type": "text", + "content": "*Corresponding authors; †Equal contributions." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1870" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "velop an optimal classifier with clear decision boundaries to separate different classes. On the one hand, effective representation learning can alleviate the issue of imbalanced data distribution and create a balanced embedding space; on the other hand, robust classifiers can improve the recognition capability of tail classes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 148, + 298, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 148, + 298, + 422 + ], + "spans": [ + { + "bbox": [ + 56, + 148, + 298, + 422 + ], + "type": "text", + "content": "Thanks to recent advances in self-supervised learning and pre-training, foundational models now learn representations much more effectively. In the pre-trained era, imbalanced data primarily affects the classifier, where sufficient samples of head classes occupy decision space, while tail classes suffer from sparse, obscure decision regions. As a result, the potential of pre-trained foundational models is not fully released. During the inference stage, while informative representations of tail classes are obtained, they often differ from those in the training set. This discrepancy erodes the well-defined margins of the classifier [25], leading to the misclassification of samples from tail classes as head classes, as shown in Fig. 1(a). Recently, numerous methods have been developed to optimize the classifier. Representatively, BCL [52] and TSC [28] use Supervised Contrastive Learning (SCL) to tighten intra-class spaces and enlarge inter-class margins, significantly improving decision boundaries between classes. Furthermore, MixUp [45] and its variants [23, 42] are good at generating diversity samples that are beneficial for improving decision regions of tail classes. However, their goal of smoothing the decision space increases boundary ambiguity, which hinders classification since errors often occur at boundaries." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 426, + 298, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 298, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 298, + 715 + ], + "type": "text", + "content": "Inspired by evolutionary algorithms like the Optimal Foraging Algorithm (OFA) [51], we attempt to tackle biased decision regions and boundaries in long-tailed recognition through a space exploration perspective. To this end, we propose Supervised Exploratory Learning (SEL), a simple yet effective framework that seamlessly integrates with existing methods to enhance their performance. SEL leverages an adaptive OFA operator to synthesize exploratory examples, compensating for insufficient tail-class data and enabling balanced decision regions and boundaries, as shown in Fig. 1(b). To be specific, SEL is composed of two modules: Class-biased Complement (CbC) and Fitness-weighted Sampling (FwS). The former is responsible for sampling more tail-class samples as target samples to balance the data distribution, while the latter selects more confident samples from adjacent classes as candidate samples to facilitate the formation of clear decision boundaries between classes. Hence, the synthesized exploratory examples are diverse and informative in representing the distribution of unseen samples in reality. SEL is a plug-and-play training framework compatible with existing long-tailed recognition methods, requiring no backbone modifications or additional parameters. The main contributions of this paper are summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 314, + 72, + 555, + 239 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 314, + 72, + 553, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 553, + 132 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 553, + 132 + ], + "type": "text", + "content": "- We attempt to address the challenge of long-tailed data from the perspective of space exploration. An adaptive OFA operator is employed to synthesize exploratory examples, which are informative and diverse to represent the distribution of unseen samples in reality." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 133, + 555, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 133, + 555, + 192 + ], + "spans": [ + { + "bbox": [ + 314, + 133, + 555, + 192 + ], + "type": "text", + "content": "- We propose a novel SEL framework that enables existing long-tailed recognition models to achieve balanced decision regions and clear decision boundaries, thereby improving their performance without altering network structure or increasing model parameters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 192, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 192, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 314, + 192, + 553, + 239 + ], + "type": "text", + "content": "- Rational analysis proves the effectiveness of SEL in improving the classification of tail classes. Considerable performance improvements of existing long-tailed recognition methods verify the effectiveness of SEL." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 253, + 400, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 253, + 400, + 266 + ], + "spans": [ + { + "bbox": [ + 314, + 253, + 400, + 266 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 274, + 555, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 555, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 555, + 334 + ], + "type": "text", + "content": "Data Augmentation (DA) [50] is an effective method in improving model robustness, particularly in scenarios of data scarcity or imbalance. In this section, we review recent advances in DA for long-tailed recognition, categorized into explicit and implicit methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 343, + 468, + 356 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 343, + 468, + 356 + ], + "spans": [ + { + "bbox": [ + 314, + 343, + 468, + 356 + ], + "type": "text", + "content": "2.1. Explicit Data Augmentation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 361, + 554, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 554, + 505 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 554, + 505 + ], + "type": "text", + "content": "Traditional DA techniques, like flipping, rotating, cropping, and padding, effectively improve robustness on long-tailed datasets and mitigate overfitting by creating new image-label pairs consistent with the training data [9, 34]. More recently, Population-Based Augmentation [10], AutoAugment [2], and Randaugment [3] have emerged to automatically choose optimal DA strategies according to characteristics of datasets, which have demonstrated their efficacy in long-tailed data [4, 21, 33]. Similarly, Wang et al. [37] generated images for tail classes using encoder variation information learned from head classes. Zada et al. [43] utilizes pure noise images as samples for tail classes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 506, + 553, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 553, + 566 + ], + "type": "text", + "content": "Although generating new images is an effective data augmentation method, it is time-consuming and computationally expensive. Moreover, the diversity of the generated images is limited, as they are created through simple transformations of the original training data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 575, + 468, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 575, + 468, + 588 + ], + "spans": [ + { + "bbox": [ + 314, + 575, + 468, + 588 + ], + "type": "text", + "content": "2.2. Implicit Data Augmentation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "Different from explicit DA, implicit DA generates new samples by interpolating existing samples in the feature space, which can effectively improve the model's generalization ability. Representatively, MixUp [45] introduced complexity control to unexplored regions in the data space by linearly interpolating discrete sample points, effectively reducing the generalization error of the model. Subsequently, CutMix [42], an extension of MixUp, augmented data by inserting a patch from one image into another, with labels blended according to the patch areas. This method can" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1871" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 203 + ], + "type": "text", + "content": "transfer context-rich background information from dominant classes to less frequent ones by combining foreground patches from the latter with background images from the former [32]. Manifold MixUp [36] further developed this concept by applying linear interpolation to the features of input images, which proved to be more effective for long-tail learning. MetaSAug [26] learns class covariance matrices from a small, balanced validation set to augment tail classes semantically. Recently, H2T [25] improved diversity by replacing part of the tail class feature map with that of the head classes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 204, + 296, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 296, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 296, + 289 + ], + "type": "text", + "content": "These findings highlight the superiority of implicit DA methods in enhancing model generalization on long-tailed data. Following this line, we propose a novel implicit DA technique, from the perspective of space exploration, to augment the diversity of training data, especially for tail classes, and improve the robustness of the long-tailed recognition model's performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 300, + 138, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 138, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 138, + 314 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 320, + 124, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 320, + 124, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 320, + 124, + 332 + ], + "type": "text", + "content": "3.1. Definition" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 338, + 296, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 296, + 446 + ], + "type": "text", + "content": "For image classification, " + }, + { + "bbox": [ + 55, + 338, + 296, + 446 + ], + "type": "inline_equation", + "content": "\\{x,y\\}" + }, + { + "bbox": [ + 55, + 338, + 296, + 446 + ], + "type": "text", + "content": " represents a sample consisting of an image " + }, + { + "bbox": [ + 55, + 338, + 296, + 446 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 338, + 296, + 446 + ], + "type": "text", + "content": " and its corresponding label " + }, + { + "bbox": [ + 55, + 338, + 296, + 446 + ], + "type": "inline_equation", + "content": "y\\in \\mathcal{V}" + }, + { + "bbox": [ + 55, + 338, + 296, + 446 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 338, + 296, + 446 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{1,2,\\dots,C\\}" + }, + { + "bbox": [ + 55, + 338, + 296, + 446 + ], + "type": "text", + "content": " is the set of classes. In the long-tailed scenario, different classes have varying sample sizes. The imbalance ratio " + }, + { + "bbox": [ + 55, + 338, + 296, + 446 + ], + "type": "inline_equation", + "content": "\\rho = \\frac{n_{\\mathrm{max}}}{n_{\\mathrm{min}}}" + }, + { + "bbox": [ + 55, + 338, + 296, + 446 + ], + "type": "text", + "content": " measures the skewness of the data distribution, where " + }, + { + "bbox": [ + 55, + 338, + 296, + 446 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{max}}" + }, + { + "bbox": [ + 55, + 338, + 296, + 446 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 338, + 296, + 446 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{min}}" + }, + { + "bbox": [ + 55, + 338, + 296, + 446 + ], + "type": "text", + "content": " represent the maximum and minimum number of training samples across classes. Typically, training samples for head classes are significantly greater than those for tail classes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 447, + 296, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 296, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 296, + 519 + ], + "type": "text", + "content": "Existing long-tailed recognition methods usually learn a complex function mapping from the input space " + }, + { + "bbox": [ + 55, + 447, + 296, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 55, + 447, + 296, + 519 + ], + "type": "text", + "content": " to the target space " + }, + { + "bbox": [ + 55, + 447, + 296, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 55, + 447, + 296, + 519 + ], + "type": "text", + "content": ". This function is usually implemented as the composition of an encoder " + }, + { + "bbox": [ + 55, + 447, + 296, + 519 + ], + "type": "inline_equation", + "content": "f: \\mathcal{X} \\to \\mathcal{Z} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 447, + 296, + 519 + ], + "type": "text", + "content": " and a linear classifier " + }, + { + "bbox": [ + 55, + 447, + 296, + 519 + ], + "type": "inline_equation", + "content": "w: \\mathcal{Z} \\to \\mathcal{Y}" + }, + { + "bbox": [ + 55, + 447, + 296, + 519 + ], + "type": "text", + "content": ". Both are crucial for the final classification accuracy [46]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 528, + 130, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 130, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 130, + 540 + ], + "type": "text", + "content": "3.2. Motivation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "Firstly, the poor performance of the tail classes is primarily due to their lack of training samples. A straightforward solution is to increase the number of tail class samples through methods such as re-sampling [44] or to emphasize their importance through re-weighting [47]. Although these methods are effective, they do not generate new, diverse samples for tail classes different from the training data, limiting the enhancement of generalization. Secondly, MixUp [45] and its variants [23, 36, 42] are good at generating diverse samples that are beneficial to improving the decision regions of the tail classes. However, on the one hand, head-class samples are more likely to be randomly selected to construct mixed samples, worsening the imbalance of training data. On the other hand, their goal of smoothing the de" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 72, + 408, + 125 + ], + "blocks": [ + { + "bbox": [ + 318, + 72, + 408, + 125 + ], + "lines": [ + { + "bbox": [ + 318, + 72, + 408, + 125 + ], + "spans": [ + { + "bbox": [ + 318, + 72, + 408, + 125 + ], + "type": "image", + "image_path": "97d63d4038a2da45930095677aa2a619d1ffe2dee1d69c5b20acc64174a9a0c7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 126, + 377, + 134 + ], + "lines": [ + { + "bbox": [ + 348, + 126, + 377, + 134 + ], + "spans": [ + { + "bbox": [ + 348, + 126, + 377, + 134 + ], + "type": "text", + "content": "(a) MixUp" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 408, + 72, + 459, + 125 + ], + "blocks": [ + { + "bbox": [ + 408, + 72, + 459, + 125 + ], + "lines": [ + { + "bbox": [ + 408, + 72, + 459, + 125 + ], + "spans": [ + { + "bbox": [ + 408, + 72, + 459, + 125 + ], + "type": "image", + "image_path": "bcb92c6d87e4e3b7b1e2d1496346d32a03f02a9f3f27ef471ebf97eabc85d040.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 422, + 126, + 445, + 134 + ], + "lines": [ + { + "bbox": [ + 422, + 126, + 445, + 134 + ], + "spans": [ + { + "bbox": [ + 422, + 126, + 445, + 134 + ], + "type": "text", + "content": "(b) OFA" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 143, + 555, + 220 + ], + "lines": [ + { + "bbox": [ + 313, + 143, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 143, + 555, + 220 + ], + "type": "text", + "content": "Figure 2. The differences and relationships among MixUp, OFA and SEL. MixUp linearly interpolates two random samples (typically from different classes, such as " + }, + { + "bbox": [ + 313, + 143, + 555, + 220 + ], + "type": "inline_equation", + "content": "z_{a}" + }, + { + "bbox": [ + 313, + 143, + 555, + 220 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 143, + 555, + 220 + ], + "type": "inline_equation", + "content": "z_{c}" + }, + { + "bbox": [ + 313, + 143, + 555, + 220 + ], + "type": "text", + "content": ") to create a new sample, while OFA leverages the better individual " + }, + { + "bbox": [ + 313, + 143, + 555, + 220 + ], + "type": "inline_equation", + "content": "(z_{s'})" + }, + { + "bbox": [ + 313, + 143, + 555, + 220 + ], + "type": "text", + "content": " to guide weaker individuals " + }, + { + "bbox": [ + 313, + 143, + 555, + 220 + ], + "type": "inline_equation", + "content": "(z_{s})" + }, + { + "bbox": [ + 313, + 143, + 555, + 220 + ], + "type": "text", + "content": " toward better evolution. Combining their strengths, SEL gradually expands and reinforces minority decision regions by synthesizing adjacent exploratory examples." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 461, + 72, + 549, + 125 + ], + "blocks": [ + { + "bbox": [ + 461, + 72, + 549, + 125 + ], + "lines": [ + { + "bbox": [ + 461, + 72, + 549, + 125 + ], + "spans": [ + { + "bbox": [ + 461, + 72, + 549, + 125 + ], + "type": "image", + "image_path": "a02a9029b7da64dbdab100a1e32cb259f7075caef87d57e3fa37e7f4a726b97f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 125, + 515, + 133 + ], + "lines": [ + { + "bbox": [ + 493, + 125, + 515, + 133 + ], + "spans": [ + { + "bbox": [ + 493, + 125, + 515, + 133 + ], + "type": "text", + "content": "(c) SEL" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 243, + 554, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 554, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 554, + 326 + ], + "type": "text", + "content": "cision space intensifies the boundary ambiguity, as shown in Fig. 2(a), where the mixed sample " + }, + { + "bbox": [ + 313, + 243, + 554, + 326 + ], + "type": "inline_equation", + "content": "z_{e}" + }, + { + "bbox": [ + 313, + 243, + 554, + 326 + ], + "type": "text", + "content": " may fall into another class's region. Finally, methods like BCL [52] and TSC [28] use SCL [19] to tighten intra-class spaces and expand inter-class margins, significantly improving decision boundaries but reducing decision region robustness while increasing false positives risk." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 327, + 554, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 554, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 554, + 458 + ], + "type": "text", + "content": "To address these issues simultaneously, we propose to synthesize exploratory examples to enhance the diversity of the training data and improve the robustness of the decision regions and boundaries. As shown in Fig. 2(b), OFA [51] is good at exploring the sample space by leveraging the better individual (evaluated by the task-specific fitness function) to guide weaker ones, promoting more effective evolution and optimization. Combining the strengths of MixUp and OFA, SEL views samples of the same class as a population and expands its decision region by synthesizing exploratory examples in uncharted spaces, as illustrated in Fig. 2(c)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 459, + 554, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 554, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 554, + 566 + ], + "type": "text", + "content": "Note that MixUp is an intuitive and effective enhancement method applied at the input layer [7]. However, due to varying model structures and input patterns, unifying these methods is challenging. Consequently, most enhanced training frameworks [22, 25, 36] operate in the feature space, as most classification models first derive feature representations before classifying images. Similarly, SEL is also operated in the feature space, making it compatible with various models and methods." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 576, + 495, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 495, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 495, + 589 + ], + "type": "text", + "content": "3.3. Supervised Exploratory Learning" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 594, + 554, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 643 + ], + "type": "text", + "content": "SEL is a compatible framework that can be seamlessly integrated into existing methods. In SEL, an adaptive OFA operator is employed to synthesize exploratory examples to supplement and balance the original training samples." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 650, + 499, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 650, + 499, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 650, + 499, + 662 + ], + "type": "text", + "content": "3.3.1. Exploratory Example Synthesization" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "Inspired by the space exploration process of evolutionary algorithms, we propose to synthesize exploratory examples using an adaptive OFA operator. The OFA [51], a metaheuristic algorithm introduced in 2017, is characterized by" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1872" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": "its simple structure, strong global search efficiency, and proven affine invariance [15]. It simulates animals' foraging behavior, where each individual follows better ones to explore the environment. In OFA and its variants, the operator is defined in dimension. Here, we extend its formulation to a vector operation. Specifically, we treat samples in the batch as populations, and synthesize exploratory examples by using the following formula:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 129, + 174, + 295, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 174, + 295, + 186 + ], + "spans": [ + { + "bbox": [ + 129, + 174, + 295, + 186 + ], + "type": "interline_equation", + "content": "z _ {e} = z _ {s} + r \\left(z _ {s ^ {\\prime}} - z _ {s}\\right) \\tag {1}", + "image_path": "7383f362c5ef6c0595dc292d3306152bf061a52811ed527a852d08ae4063b3c2.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 190, + 296, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 296, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 296, + 299 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 190, + 296, + 299 + ], + "type": "inline_equation", + "content": "z_{s} \\in \\mathcal{Z}" + }, + { + "bbox": [ + 55, + 190, + 296, + 299 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 190, + 296, + 299 + ], + "type": "inline_equation", + "content": "z_{s'} \\in \\mathcal{Z}" + }, + { + "bbox": [ + 55, + 190, + 296, + 299 + ], + "type": "text", + "content": " denote the target and candidate samples, respectively. " + }, + { + "bbox": [ + 55, + 190, + 296, + 299 + ], + "type": "inline_equation", + "content": "r \\in (0,1)" + }, + { + "bbox": [ + 55, + 190, + 296, + 299 + ], + "type": "text", + "content": " is a coefficient that controls the exploration distance, which is randomly sampled during model training. Generally, OFA selects the candidate sample " + }, + { + "bbox": [ + 55, + 190, + 296, + 299 + ], + "type": "inline_equation", + "content": "z_{s'}" + }, + { + "bbox": [ + 55, + 190, + 296, + 299 + ], + "type": "text", + "content": " with higher fitness to guide weaker individuals (" + }, + { + "bbox": [ + 55, + 190, + 296, + 299 + ], + "type": "inline_equation", + "content": "z_{s}" + }, + { + "bbox": [ + 55, + 190, + 296, + 299 + ], + "type": "text", + "content": ") toward better evolution. To adapt OFA to the requirement of long-tailed recognition, we introduce two modules, CbC and FwS, to regulate the selection of target and candidate samples, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 302, + 224, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 302, + 224, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 302, + 224, + 314 + ], + "type": "text", + "content": "3.3.2. Class-biased Complement (CbC)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 317, + 296, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 317, + 296, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 296, + 484 + ], + "type": "text", + "content": "To alleviate the class imbalance problem, diverse exploratory examples are synthesized to complement the tail classes. We first count the number of samples in each class and then identify the additional samples required (i.e., exploratory examples) to balance the class distribution. As illustrated in Fig. 2(c), the target sample's label is assigned to the exploratory example, eliminating the decision boundary ambiguity issue present in MixUp. Hence, synthesizing varying numbers of exploratory examples for different classes becomes a task of selecting target samples. In practice, training samples with the target class are allowed to be selected as the target samples multiple times. Specifically, for a sample " + }, + { + "bbox": [ + 55, + 317, + 296, + 484 + ], + "type": "inline_equation", + "content": "z_{s}" + }, + { + "bbox": [ + 55, + 317, + 296, + 484 + ], + "type": "text", + "content": " in the batch, it will be utilized as the target sample " + }, + { + "bbox": [ + 55, + 317, + 296, + 484 + ], + "type": "inline_equation", + "content": "k_{c}" + }, + { + "bbox": [ + 55, + 317, + 296, + 484 + ], + "type": "text", + "content": " times if it belongs to class " + }, + { + "bbox": [ + 55, + 317, + 296, + 484 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 317, + 296, + 484 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 126, + 490, + 294, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 490, + 294, + 512 + ], + "spans": [ + { + "bbox": [ + 126, + 490, + 294, + 512 + ], + "type": "interline_equation", + "content": "k _ {c} = \\operatorname {C e i l} \\left(\\frac {n _ {\\max } - n _ {c}}{n _ {c}}\\right) \\tag {2}", + "image_path": "0b740ccd9396ccaed2904ab4668e7716d23f80036a3cb0e225f898326c007638.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 515, + 295, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 295, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 295, + 551 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 515, + 295, + 551 + ], + "type": "inline_equation", + "content": "n_c" + }, + { + "bbox": [ + 55, + 515, + 295, + 551 + ], + "type": "text", + "content": " denotes the sample number of class " + }, + { + "bbox": [ + 55, + 515, + 295, + 551 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 515, + 295, + 551 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 515, + 295, + 551 + ], + "type": "inline_equation", + "content": "n_{max}" + }, + { + "bbox": [ + 55, + 515, + 295, + 551 + ], + "type": "text", + "content": " represents the sample size of the largest head class. " + }, + { + "bbox": [ + 55, + 515, + 295, + 551 + ], + "type": "inline_equation", + "content": "\\text{Ceil}(\\ast)" + }, + { + "bbox": [ + 55, + 515, + 295, + 551 + ], + "type": "text", + "content": " is the ceiling function." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 552, + 296, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 296, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 296, + 671 + ], + "type": "text", + "content": "In this way, the issue of class imbalance is addressed by synthesizing exploratory examples with different classes, ensuring that the number of samples in each class is balanced. The key challenges are guaranteeing the diversity of synthesized exploratory examples and the rationality of assigned labels, a crucial difference that distinguishes SEL from MixUp and its variants. As shown in Fig. 3, our solution is to identify neighboring classes and place exploratory examples in the regions between the target and adjacent classes, ensuring rational label assignment." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 674, + 216, + 686 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 674, + 216, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 674, + 216, + 686 + ], + "type": "text", + "content": "3.3.3. Neighboring Class Recognition" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "In MixUp and its variants, the label of the synthesized sample is determined by linear interpolation of the target and" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 69, + 552, + 148 + ], + "blocks": [ + { + "bbox": [ + 318, + 69, + 552, + 148 + ], + "lines": [ + { + "bbox": [ + 318, + 69, + 552, + 148 + ], + "spans": [ + { + "bbox": [ + 318, + 69, + 552, + 148 + ], + "type": "image", + "image_path": "9d6540a18a942a52d8410d52e6c54ddcd393f8933f800cf372c471452129a3b0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 156, + 554, + 178 + ], + "lines": [ + { + "bbox": [ + 313, + 156, + 554, + 178 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 554, + 178 + ], + "type": "text", + "content": "Figure 3. Synthesize exploratory examples by the adaptive OFA operator. More exploratory examples are generated for tail classes." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 198, + 555, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 555, + 294 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 555, + 294 + ], + "type": "text", + "content": "candidate samples. This approach results in label misassignment if the synthesized sample falls within the decision region of another class. To avoid this situation, we propose selecting candidate samples from neighboring classes, and placing synthesized samples between the target and adjacent class regions. Specifically, for samples in the batch " + }, + { + "bbox": [ + 313, + 198, + 555, + 294 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 198, + 555, + 294 + ], + "type": "text", + "content": ", we calculate the center of each class and then determine the Euclidean distances between the target and the others." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 301, + 553, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 301, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 317, + 301, + 553, + 346 + ], + "type": "interline_equation", + "content": "d \\left(c, c ^ {\\prime}\\right) = \\| \\frac {1}{\\left| \\mathcal {B} ^ {c} \\right|} \\sum_ {z _ {i} \\in \\mathcal {B}} ^ {| \\mathcal {B} |} z _ {i} \\cdot \\mathbb {I} _ {y _ {i} = c} - \\frac {1}{\\left| \\mathcal {B} ^ {c ^ {\\prime}} \\right|} \\sum_ {z _ {i} \\in \\mathcal {B}} ^ {| \\mathcal {B} |} z _ {i} \\cdot \\mathbb {I} _ {y _ {i} = c ^ {\\prime}} \\| _ {2} \\tag {3}", + "image_path": "5bbf04afe6d10fa9c566ce6fe185824ede26321b6585eb3488157c64e27659c3.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^c" + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^{c'}" + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "text", + "content": " are sets of samples belong to class " + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "text", + "content": " in the batch " + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "inline_equation", + "content": "\\mathbb{I}_*" + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "text", + "content": " is the indicator function, and " + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "inline_equation", + "content": "\\| * \\|_2" + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "text", + "content": " denotes the L2 norm. Therefore, " + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "inline_equation", + "content": "d(c, c')" + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "text", + "content": " denotes the distance between the centers of class " + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "text", + "content": " and class " + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "text", + "content": ". The nearest " + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "inline_equation", + "content": "k_c'" + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "text", + "content": " classes are chosen as neighboring classes for class " + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "text", + "content": ", from which a neighbor is randomly selected, such as class " + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "text", + "content": ", and the candidate sample will be picked from " + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^{c'}" + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "text", + "content": " based on their fitnesses, as done in OFA." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 449, + 485, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 485, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 485, + 460 + ], + "type": "text", + "content": "3.3.4. Fitness-weighted Sampling (FwS)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 463, + 554, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 554, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 554, + 548 + ], + "type": "text", + "content": "Following the principle of OFA, the prediction probability of the correct class is considered the fitness of the sample, and the sample with higher fitness is selected as the candidate sample. To reduce the operation complexity and increase the exploration randomness, we employ probability-based sampling to select the candidate sample. Specifically, the sampling probabilities are obtained by:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 384, + 555, + 553, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 555, + 553, + 574 + ], + "spans": [ + { + "bbox": [ + 384, + 555, + 553, + 574 + ], + "type": "interline_equation", + "content": "p _ {s ^ {\\prime}} = \\operatorname {N o r m} \\left(\\left\\{\\hat {y} _ {i} ^ {c ^ {\\prime}} \\right\\} _ {i = 1} ^ {\\left| \\mathcal {B} ^ {c ^ {\\prime}} \\right|}\\right) \\tag {4}", + "image_path": "ff9227be4b309c2d883cace1982094743b6815c6b28ecc577a1d2db88dc6fe7e.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 581, + 554, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 554, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 554, + 642 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 581, + 554, + 642 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 313, + 581, + 554, + 642 + ], + "type": "text", + "content": " represents the selected neighboring class, " + }, + { + "bbox": [ + 313, + 581, + 554, + 642 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^{c'}" + }, + { + "bbox": [ + 313, + 581, + 554, + 642 + ], + "type": "text", + "content": " is the set of samples belong to class " + }, + { + "bbox": [ + 313, + 581, + 554, + 642 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 313, + 581, + 554, + 642 + ], + "type": "text", + "content": " in the batch " + }, + { + "bbox": [ + 313, + 581, + 554, + 642 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 581, + 554, + 642 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 581, + 554, + 642 + ], + "type": "inline_equation", + "content": "\\hat{y}_i^{c'}" + }, + { + "bbox": [ + 313, + 581, + 554, + 642 + ], + "type": "text", + "content": " is the predicted probability of the " + }, + { + "bbox": [ + 313, + 581, + 554, + 642 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 581, + 554, + 642 + ], + "type": "text", + "content": "-th sample belonging to its correct class " + }, + { + "bbox": [ + 313, + 581, + 554, + 642 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 313, + 581, + 554, + 642 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 581, + 554, + 642 + ], + "type": "inline_equation", + "content": "\\text{Norm}(*)" + }, + { + "bbox": [ + 313, + 581, + 554, + 642 + ], + "type": "text", + "content": " is the normalization function that ensures the sum of " + }, + { + "bbox": [ + 313, + 581, + 554, + 642 + ], + "type": "inline_equation", + "content": "p_{s'}" + }, + { + "bbox": [ + 313, + 581, + 554, + 642 + ], + "type": "text", + "content": " equals 1." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "In this way, for each target sample " + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "inline_equation", + "content": "z_{s}" + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": ", a candidate sample " + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "inline_equation", + "content": "z_{s'}" + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": " with higher confidence is sampled from the batch, thereby achieving exploratory examples with diversity. FwS is essential for the validity of SEL, which is why we use the adaptive OFA operator to synthesize exploratory examples. The rational analysis is elaborated later." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1873" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 217, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 217, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 217, + 83 + ], + "type": "text", + "content": "3.3.5. Training with SEL Framework" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 87, + 296, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 87, + 296, + 147 + ], + "spans": [ + { + "bbox": [ + 55, + 87, + 296, + 147 + ], + "type": "text", + "content": "SEL is a plug-and-play training framework that can be seamlessly integrated into the training process of the original model. The optimization objective of SEL is to correctly classify the exploratory examples, which is equivalent to maximizing the predicted probability of the correct class." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 102, + 152, + 295, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 152, + 295, + 186 + ], + "spans": [ + { + "bbox": [ + 102, + 152, + 295, + 186 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S E L} = - \\frac {1}{C} \\sum_ {c = 1} ^ {C} \\left(\\frac {1}{k _ {c}} \\sum_ {i = 1} ^ {k _ {c}} \\log \\left(\\hat {y} _ {i} ^ {c}\\right)\\right) \\tag {5}", + "image_path": "96dab35cd3bd3484d8093af35d89b21b6291dbe72ad7cf8512e03823b5624c80.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 190, + 296, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 296, + 226 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 296, + 226 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 190, + 296, + 226 + ], + "type": "inline_equation", + "content": "\\hat{y}_i^c" + }, + { + "bbox": [ + 55, + 190, + 296, + 226 + ], + "type": "text", + "content": " denotes the predicted probability of the " + }, + { + "bbox": [ + 55, + 190, + 296, + 226 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 190, + 296, + 226 + ], + "type": "text", + "content": "-th exploratory example belonging to the " + }, + { + "bbox": [ + 55, + 190, + 296, + 226 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 190, + 296, + 226 + ], + "type": "text", + "content": "-th class, and " + }, + { + "bbox": [ + 55, + 190, + 296, + 226 + ], + "type": "inline_equation", + "content": "k_c" + }, + { + "bbox": [ + 55, + 190, + 296, + 226 + ], + "type": "text", + "content": " is the number of exploratory examples synthesized for class " + }, + { + "bbox": [ + 55, + 190, + 296, + 226 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 190, + 296, + 226 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 226, + 292, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 226, + 292, + 238 + ], + "spans": [ + { + "bbox": [ + 67, + 226, + 292, + 238 + ], + "type": "text", + "content": "Finally, the overall optimization objective is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 119, + 243, + 295, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 243, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 119, + 243, + 295, + 258 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = (1 - \\lambda) \\mathcal {L} _ {\\text {T a s k}} + \\lambda \\mathcal {L} _ {\\text {S E L}} \\tag {6}", + "image_path": "d846c04f74e0252eced8fde068b1b845c0bdecbc41f84d576f3ae19f13022c90.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 262, + 296, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 296, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 296, + 299 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 262, + 296, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Task}" + }, + { + "bbox": [ + 55, + 262, + 296, + 299 + ], + "type": "text", + "content": " is the optimization objective of the original method, and " + }, + { + "bbox": [ + 55, + 262, + 296, + 299 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 262, + 296, + 299 + ], + "type": "text", + "content": " is a weighting coefficient that controls the impact of SEL loss." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 304, + 166, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 166, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 166, + 316 + ], + "type": "text", + "content": "3.4. Rationale Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 322, + 296, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 322, + 296, + 406 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 296, + 406 + ], + "type": "text", + "content": "The exploitation process of OFA has proved to be effective and rational [51]. In this section, we provide a rational analysis of SEL. For clarity, we denote the weights of the linear classifier as " + }, + { + "bbox": [ + 55, + 322, + 296, + 406 + ], + "type": "inline_equation", + "content": "w = \\{w_{c}\\}_{c = 1}^{C}" + }, + { + "bbox": [ + 55, + 322, + 296, + 406 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 322, + 296, + 406 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 322, + 296, + 406 + ], + "type": "text", + "content": " is the number of classes, and " + }, + { + "bbox": [ + 55, + 322, + 296, + 406 + ], + "type": "inline_equation", + "content": "w_{c} \\in \\mathbb{R}^{d \\times 1}" + }, + { + "bbox": [ + 55, + 322, + 296, + 406 + ], + "type": "text", + "content": " denotes the weights of class " + }, + { + "bbox": [ + 55, + 322, + 296, + 406 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 322, + 296, + 406 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 322, + 296, + 406 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 322, + 296, + 406 + ], + "type": "text", + "content": " is the feature dimension. For a sample " + }, + { + "bbox": [ + 55, + 322, + 296, + 406 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{d \\times 1}" + }, + { + "bbox": [ + 55, + 322, + 296, + 406 + ], + "type": "text", + "content": ", the predicted probability of class " + }, + { + "bbox": [ + 55, + 322, + 296, + 406 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 322, + 296, + 406 + ], + "type": "text", + "content": " can be expressed as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 135, + 410, + 295, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 410, + 295, + 442 + ], + "spans": [ + { + "bbox": [ + 135, + 410, + 295, + 442 + ], + "type": "interline_equation", + "content": "\\hat {y} ^ {c} = \\frac {e ^ {w _ {c} ^ {T} z}}{\\sum_ {c ^ {\\prime} = 1} ^ {C} e ^ {w _ {c ^ {\\prime}} ^ {T} z}} \\tag {7}", + "image_path": "6568b6b776f0468a67ff9e71cfbfbebebf8305b9dbdbc4829c9ad58ca117bd7e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 447, + 296, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 296, + 483 + ], + "type": "text", + "content": "Hence, " + }, + { + "bbox": [ + 55, + 447, + 296, + 483 + ], + "type": "inline_equation", + "content": "\\hat{y}^c\\propto w_c^T z" + }, + { + "bbox": [ + 55, + 447, + 296, + 483 + ], + "type": "text", + "content": " . The classifier attempts to assign a higher probability to the target class, i.e., " + }, + { + "bbox": [ + 55, + 447, + 296, + 483 + ], + "type": "inline_equation", + "content": "\\hat{y}^c >\\hat{y}^{c'}" + }, + { + "bbox": [ + 55, + 447, + 296, + 483 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 55, + 447, + 296, + 483 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 447, + 296, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 447, + 296, + 483 + ], + "type": "inline_equation", + "content": "c^\\prime" + }, + { + "bbox": [ + 55, + 447, + 296, + 483 + ], + "type": "text", + "content": " denote the target and other classes, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 483, + 296, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 602 + ], + "type": "text", + "content": "Without loss of generality, the selection of the target sample " + }, + { + "bbox": [ + 55, + 483, + 296, + 602 + ], + "type": "inline_equation", + "content": "(z_{s})" + }, + { + "bbox": [ + 55, + 483, + 296, + 602 + ], + "type": "text", + "content": " and candidate sample " + }, + { + "bbox": [ + 55, + 483, + 296, + 602 + ], + "type": "inline_equation", + "content": "(z_{s}^{\\prime})" + }, + { + "bbox": [ + 55, + 483, + 296, + 602 + ], + "type": "text", + "content": " can be divided into two cases, as summarized in Tab. 1. We use the subscripts " + }, + { + "bbox": [ + 55, + 483, + 296, + 602 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 483, + 296, + 602 + ], + "type": "text", + "content": " (tail) and " + }, + { + "bbox": [ + 55, + 483, + 296, + 602 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 483, + 296, + 602 + ], + "type": "text", + "content": " (head) to represent their relative relation. Generally, the first case typically dominates, as the CbC module selects more tail-class targets, while the long-tailed distribution increases the chance of candidate samples being from head classes. Additionally, since exploratory examples share labels with targets, the optimization aims to challenge the model's bias toward head classes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 602, + 286, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 602, + 286, + 616 + ], + "spans": [ + { + "bbox": [ + 67, + 602, + 286, + 616 + ], + "type": "text", + "content": "For the first case, the goal of " + }, + { + "bbox": [ + 67, + 602, + 286, + 616 + ], + "type": "inline_equation", + "content": "\\hat{y}_e^t >\\hat{y}_e^h" + }, + { + "bbox": [ + 67, + 602, + 286, + 616 + ], + "type": "text", + "content": " is equivalent to:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 619, + 295, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 619, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 132, + 619, + 295, + 639 + ], + "type": "interline_equation", + "content": "w _ {t} ^ {T} z _ {e} - w _ {h} ^ {T} z _ {e} > 0 \\Rightarrow \\tag {8}", + "image_path": "f196d6a71ea2810c41b43367ff640e3571f1f8183c6cbb9fd4b955b6cbe31740.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 71, + 636, + 269, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 636, + 269, + 651 + ], + "spans": [ + { + "bbox": [ + 71, + 636, + 269, + 651 + ], + "type": "interline_equation", + "content": "w _ {t} ^ {T} \\left(z _ {t} + r \\left(z _ {h} - z _ {t}\\right)\\right) - w _ {h} ^ {T} \\left(z _ {t} + r \\left(z _ {h} - z _ {t}\\right)\\right) > 0", + "image_path": "bcdfdaef8bd652c2973b11b6d6b283baab02f8e13975af42f07ca9ac5364860e.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 656, + 292, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 656, + 292, + 668 + ], + "spans": [ + { + "bbox": [ + 55, + 656, + 292, + 668 + ], + "type": "text", + "content": "After expanding and transforming the inequality, we have:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 93, + 672, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 672, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 93, + 672, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} w _ {t} ^ {T} z _ {t} - w _ {h} ^ {T} z _ {t} > \\frac {r}{1 - r} \\left(w _ {h} ^ {T} z _ {h} - w _ {t} ^ {T} z _ {h}\\right) \\tag {9} \\\\ \\Rightarrow \\hat {y} _ {t} ^ {t} - \\hat {y} _ {t} ^ {h} > \\frac {r}{1 - r} (\\hat {y} _ {h} ^ {h} - \\hat {y} _ {h} ^ {t}) \\\\ \\end{array}", + "image_path": "c5f20f292816c6cd4980957c50938deb5cb4cc086fde942cd58c87f5b37b833f.jpg" + } + ] + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 318, + 71, + 554, + 129 + ], + "blocks": [ + { + "bbox": [ + 318, + 71, + 554, + 129 + ], + "lines": [ + { + "bbox": [ + 318, + 71, + 554, + 129 + ], + "spans": [ + { + "bbox": [ + 318, + 71, + 554, + 129 + ], + "type": "table", + "html": "
Casezszs'zeGoal
1ztzhzt + r(zh - zt)ˆy e t >ˆy e h
2zhztzh + r(zt - zh)ˆy e h >ˆy e t
", + "image_path": "5bbc068156c4c08ed5395c2ea76faf73f4b4dd9717ca5016857f34fb13ee8616.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 330, + 137, + 537, + 148 + ], + "lines": [ + { + "bbox": [ + 330, + 137, + 537, + 148 + ], + "spans": [ + { + "bbox": [ + 330, + 137, + 537, + 148 + ], + "type": "text", + "content": "Table 1. Two types of synthesized exploratory examples." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 168, + 553, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 264 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 168, + 553, + 264 + ], + "type": "inline_equation", + "content": "\\hat{y}_h^h" + }, + { + "bbox": [ + 313, + 168, + 553, + 264 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 168, + 553, + 264 + ], + "type": "inline_equation", + "content": "\\hat{y}_h^t" + }, + { + "bbox": [ + 313, + 168, + 553, + 264 + ], + "type": "text", + "content": " represent probabilities of the candidate sample " + }, + { + "bbox": [ + 313, + 168, + 553, + 264 + ], + "type": "inline_equation", + "content": "(z_h)" + }, + { + "bbox": [ + 313, + 168, + 553, + 264 + ], + "type": "text", + "content": " belonging to the head class and tail classes, respectively. The classifier favors head classes " + }, + { + "bbox": [ + 313, + 168, + 553, + 264 + ], + "type": "inline_equation", + "content": "(\\hat{y}_h^h - \\hat{y}_h^t > 0)" + }, + { + "bbox": [ + 313, + 168, + 553, + 264 + ], + "type": "text", + "content": ", while the FwS module reinforces this by selecting high-confidence candidates. Therefore, the right item is tend to be positive, and " + }, + { + "bbox": [ + 313, + 168, + 553, + 264 + ], + "type": "inline_equation", + "content": "\\hat{y}_t^t > \\hat{y}_t^h" + }, + { + "bbox": [ + 313, + 168, + 553, + 264 + ], + "type": "text", + "content": " is achieved, meaning that optimizing tail-class exploratory examples effectively improves the classification accuracy of original tail-class samples." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 265, + 554, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 554, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 554, + 289 + ], + "type": "text", + "content": "Similarly, for the second case, the optimization goal of SEL is to make " + }, + { + "bbox": [ + 313, + 265, + 554, + 289 + ], + "type": "inline_equation", + "content": "\\hat{y}_e^h >\\hat{y}_e^t" + }, + { + "bbox": [ + 313, + 265, + 554, + 289 + ], + "type": "text", + "content": " yielding:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 377, + 297, + 553, + 318 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 297, + 553, + 318 + ], + "spans": [ + { + "bbox": [ + 377, + 297, + 553, + 318 + ], + "type": "interline_equation", + "content": "\\hat {y} _ {h} ^ {h} - \\hat {y} _ {h} ^ {t} > \\frac {r}{1 - r} \\left(\\hat {y} _ {t} ^ {t} - \\hat {y} _ {t} ^ {h}\\right) \\tag {10}", + "image_path": "553788961182620e44e7809104bf8c86eefd792e39c481dace8ae11da99d8519.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 327, + 554, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 554, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 554, + 423 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 327, + 554, + 423 + ], + "type": "inline_equation", + "content": "\\hat{y}_t^t" + }, + { + "bbox": [ + 313, + 327, + 554, + 423 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 327, + 554, + 423 + ], + "type": "inline_equation", + "content": "\\hat{y}_t^h" + }, + { + "bbox": [ + 313, + 327, + 554, + 423 + ], + "type": "text", + "content": " represent probabilities of the candidate sample " + }, + { + "bbox": [ + 313, + 327, + 554, + 423 + ], + "type": "inline_equation", + "content": "(z_t)" + }, + { + "bbox": [ + 313, + 327, + 554, + 423 + ], + "type": "text", + "content": " belonging to the tail class and head class, respectively. Although the classifier tends to misclassify tail-class samples as head classes, the FwS mitigates this by selecting better candidates, which possess higher probabilities in the correct classes. Hence, optimizing the head-class exploratory examples does not cause significant damage to the original head sample." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 434, + 395, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 395, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 395, + 447 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 453, + 373, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 373, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 373, + 464 + ], + "type": "text", + "content": "4.1. Dataset" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 470, + 554, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 554, + 529 + ], + "type": "text", + "content": "CIFAR-10/100-LT [5] are the subsets of CIFAR-10/100 [20]. Both contain 50,000 training images and 10,000 validation images, each of size " + }, + { + "bbox": [ + 313, + 470, + 554, + 529 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 313, + 470, + 554, + 529 + ], + "type": "text", + "content": ", with 10 and 100 classes, respectively. The imbalance factors used in the experiment are set to 100, 50, and 10." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 530, + 554, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 554, + 578 + ], + "type": "text", + "content": "ImageNet-LT [30] is a long-tailed version of ImageNet [6] by sampling a subset following the Pareto distribution with power value " + }, + { + "bbox": [ + 313, + 530, + 554, + 578 + ], + "type": "inline_equation", + "content": "\\alpha = 0.64" + }, + { + "bbox": [ + 313, + 530, + 554, + 578 + ], + "type": "text", + "content": ". It consists of " + }, + { + "bbox": [ + 313, + 530, + 554, + 578 + ], + "type": "inline_equation", + "content": "115.8\\mathrm{K}" + }, + { + "bbox": [ + 313, + 530, + 554, + 578 + ], + "type": "text", + "content": " images of 1000 classes in total, with 1280 to 5 images per class." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 578, + 554, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 554, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 554, + 613 + ], + "type": "text", + "content": "iNaturalist 2018 [12] is a large-scale dataset containing 437.5K images from 8,142 classes. It is long-tailed by nature, with an extremely imbalanced distribution." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 621, + 398, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 398, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 398, + 634 + ], + "type": "text", + "content": "4.2. Basic Setting" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 639, + 427, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 427, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 427, + 651 + ], + "type": "text", + "content": "4.2.1. Compared Methods" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": "To evaluate the effectiveness of our proposed SEL, we integrated it into several existing methods introduced in recent years, including re-balanced techniques (IB [31], GCL [23], KPS [24]), contrastive learning (BCL [52]), feature compression (FCC [22]), and data augmentation strategies" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1874" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 70, + 543, + 350 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 543, + 350 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 543, + 350 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 543, + 350 + ], + "type": "table", + "html": "
MethodsCIFAR-100-LTCIFAR-10-LT
AvenueImbalance Ratio ρ10050101005010
-ResNet-32 [9]39.9646.0156.6671.9675.9585.92
+ MixUp [45]41.55↑1.5947.11↑1.1057.18↑0.5273.43↑1.4777.19↑1.2486.37↑0.45
+ SEL44.53↑4.5749.64↑3.8559.17↑2.5175.84↑3.8879.81↑3.8686.78↑0.86
ICCV 2021IB [31]40.5046.3156.7575.1179.9588.01
+ SEL42.34↑1.8447.95↑1.6457.57↑0.8275.04↓0.0779.81↓0.1488.13↑0.12
CVPR 2022BCL [52]51.7656.5167.9083.6186.1390.10
+ SEL52.30↑0.5457.25↑0.7468.43↑0.5384.44↑0.8386.31↑0.1890.26↑0.16
CVPR 2022GCL [23]46.5051.7261.7980.5684.7489.65
+ SEL47.89↑1.3952.74↑1.0262.41↑0.6281.86↑1.3085.13↑0.3989.78↑0.13
CVPR 2023FCC+CE [22]40.2045.9357.8073.8079.5787.75
+ SEL42.33↑2.1348.05↑2.1258.75↑0.9577.88↑4.0882.10↑2.5388.80↑1.05
CVPR 2023GLMC [7]53.9158.8768.0783.6886.9091.16
+ SEL56.48↑2.5761.13↑2.2670.75↑2.6885.40↑1.7288.57↑1.6792.83↑1.67
TPAMI 2023KPS [24]41.9747.9259.5982.3284.2989.10
+ SEL44.01↑2.0447.80↓0.1260.03↑0.4483.48↑1.1684.49↑0.2089.34↑0.24
AAAI 2024H2T+CE [25]42.2747.5858.2479.9182.8088.77
+ SEL42.45↑0.1847.80↑0.2258.66↑0.4280.43↑0.5282.80↓0.0088.79↑0.02
", + "image_path": "5b1d042552082832fcbed615fd8691bd343e9befa1abebca57f06d65db027233.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 113, + 358, + 496, + 370 + ], + "lines": [ + { + "bbox": [ + 113, + 358, + 496, + 370 + ], + "spans": [ + { + "bbox": [ + 113, + 358, + 496, + 370 + ], + "type": "text", + "content": "Table 2. Comparisons between raw methods and their SEL-enhanced counterparts on CIFAR-LT datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 391, + 295, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 295, + 487 + ], + "type": "text", + "content": "(H2T [25], GLMC [7]). These methods were chosen not only for their representativeness but also for their publicly available code and configurations, ensuring ease of reproducibility. Most employ a two-stage training strategy, with the first stage dedicated to representation learning and the second to classification. Additionally, we use ResNet models as baselines to directly compare the performance improvements achieved by MixUp and SEL." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 495, + 185, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 185, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 185, + 506 + ], + "type": "text", + "content": "4.2.2. Implementation Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": "We reproduce the compared methods using their official codes and configurations, setting the random seed to 2024 for fair comparison. The base ResNet models are configured identically to those in Li et al. [23]. To simplify application, we first determine the optimal way to apply SEL on different datasets by using the base model, i.e., ResNet-32 for CIFAR-10/100-LT and ResNet-50 for others, then replace it with corresponding models to assess SEL's effectiveness on these models. Specifically, for CIFAR-100-LT, SEL is applied during the final 20 training epochs. For other datasets, SEL is applied by further fine-tuning the classifier for 20 epochs on CIFAR-10-LT and 10 epochs on ImageNet-LT and iNaturalist 2018. The weight coefficient " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": " in Eq. (6) is set to 0.9, with the number of neighboring classes " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "k_c'" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": " set to 2, 5, 10, and 50 for CIFAR-10-LT, CIFAR-100-LT, ImageNet-LT, and iNaturalist 2018, respectively. Top-1 accuracy serves as the evaluation metric." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 391, + 497, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 497, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 497, + 403 + ], + "type": "text", + "content": "4.3. Improvement of Existing Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 409, + 466, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 466, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 466, + 420 + ], + "type": "text", + "content": "4.3.1. Results on CIFAR-10/100-LT" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 425, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 425, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 312, + 425, + 555, + 604 + ], + "type": "text", + "content": "For CIFAR-10/100-LT, the results with and without SEL are presented in Tab. 2. Across 48 experiment groups, our proposed SEL significantly improves 44 of them, with an average increase of " + }, + { + "bbox": [ + 312, + 425, + 555, + 604 + ], + "type": "inline_equation", + "content": "1.37\\%" + }, + { + "bbox": [ + 312, + 425, + 555, + 604 + ], + "type": "text", + "content": " (4.57% max and " + }, + { + "bbox": [ + 312, + 425, + 555, + 604 + ], + "type": "inline_equation", + "content": "0.02\\%" + }, + { + "bbox": [ + 312, + 425, + 555, + 604 + ], + "type": "text", + "content": " min). Extensive experimental results fully verify the effectiveness of our method on long-tailed recognition, particularly for scenarios with high imbalance ratios. Generally, integrating MixUp into the vanilla ResNet-32 model significantly enhances performance, highlighting the effectiveness of data augmentation. However, our proposed SEL achieves even greater improvements across all scenarios, and even surpasses existing methods. For example, ResNet-32+SEL outperforms IB on both datasets except for CIFAR-10-LT-10, and surpasses GCL, KPS, and H2T on CIFAR-100-LT, though it underperforms on CIFAR-10-LT." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Additionally, SEL significantly enhances the performance of existing methods, with particularly pronounced improvements in high-imbalance scenarios. For example, SEL is notably effective on KPS at an imbalance ratio of 100 but shows limited impact at other imbalance ratios. In addition, GLMC+SEL and FCC+SEL significantly outperform their original counterparts on both datasets, with average improvements of " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "2.10\\%" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "2.14\\%" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": ", respectively. For the re-balanced methods IB, GCL, and KPS, SEL yields" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1875" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 291, + 247 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 291, + 247 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 291, + 247 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 291, + 247 + ], + "type": "table", + "html": "
MethodManyMediumFewAll
ResNet-5063.3833.7426.0444.80
+ Mixup62.9735.4428.5845.71↑0.91
+ SEL64.1438.4131.2747.86↑3.06
BCL65.8353.1636.2755.67
+ SEL65.4554.8236.5356.27↑0.60
GCL65.0452.7735.6955.51
+ SEL65.1353.9439.1856.12↑0.61
FCC+CE65.5252.1833.8354.68
+ SEL64.7454.0035.6655.60↑0.92
GLMC69.4252.4930.4156.17
+ SEL68.6754.3738.2857.24↑1.07
", + "image_path": "5ea8bc60f24e497370a209bb2ef91ebd9fd4d8329d75eb829399a89192adb6cf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 61, + 278, + 291, + 457 + ], + "blocks": [ + { + "bbox": [ + 82, + 255, + 269, + 266 + ], + "lines": [ + { + "bbox": [ + 82, + 255, + 269, + 266 + ], + "spans": [ + { + "bbox": [ + 82, + 255, + 269, + 266 + ], + "type": "text", + "content": "Table 3. Comparisons on the ImageNet-LT dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 278, + 291, + 457 + ], + "lines": [ + { + "bbox": [ + 61, + 278, + 291, + 457 + ], + "spans": [ + { + "bbox": [ + 61, + 278, + 291, + 457 + ], + "type": "table", + "html": "
MethodManyMediumFewAll
ResNet-5074.8968.2461.4265.95
+ Mixup71.3469.5363.8867.29↑1.34
+ SEL73.0470.3866.7469.10↑3.15
BCL73.4771.9968.8270.80
+ SEL71.9973.0469.2171.32↑0.52
GCL73.7572.2868.3470.76
+ SEL72.6373.5870.0572.00↑1.24
FCC+CE72.7871.0967.4169.70
+ SEL71.1772.4169.5571.09↑1.39
GLMC76.0473.6971.7773.14
+ SEL74.6875.1473.6674.51↑1.37
", + "image_path": "3dc1d0f5a9227fff87e1df30552a7f96d94f56b7f40b1b23e9b7a716da5a435b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 499, + 295, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 499, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 55, + 499, + 295, + 668 + ], + "type": "text", + "content": "greater improvements on CIFAR-100-LT than on CIFAR-10-LT, demonstrating its superiority in scenarios with more classes. However, SEL is less effective on BCL and H2T. BCL uses contrastive learning to reinforce balanced decision regions, aligning with SEL's objective. The difference is that SEL further enhances the robustness of decision regions, resulting in additional improvements. H2T is also an evolutionary algorithm-inspired method that generates diverse samples through cross-mutation, following the principles of Genetic Algorithm (GA), thus indicating the fundamental difference between SEL and H2T stems from the distinction between OFA and GA. The slight improvement of H2T+SEL further validates SEL's effectiveness in generating diverse samples, aligning with H2T's principles." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 673, + 278, + 686 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 673, + 278, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 673, + 278, + 686 + ], + "type": "text", + "content": "4.3.2. Results on ImageNet-LT and iNaturalist 2018" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "For ImageNet-LT and iNaturalist 2018, experimental results are listed in Tab. 3 and Tab. 4, respectively. Following pre" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 334, + 70, + 535, + 159 + ], + "blocks": [ + { + "bbox": [ + 77, + 466, + 272, + 476 + ], + "lines": [ + { + "bbox": [ + 77, + 466, + 272, + 476 + ], + "spans": [ + { + "bbox": [ + 77, + 466, + 272, + 476 + ], + "type": "text", + "content": "Table 4. Comparisons on the iNaturalist 2018 dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 334, + 70, + 535, + 159 + ], + "lines": [ + { + "bbox": [ + 334, + 70, + 535, + 159 + ], + "spans": [ + { + "bbox": [ + 334, + 70, + 535, + 159 + ], + "type": "table", + "html": "
CbCFwSTop-1 Accuracy
ρ =100ρ =10
××39.9656.66
×✓40.26↑0.3056.93↑0.27
✓×43.44↑3.4857.84↑1.18
✓✓44.53↑4.5759.17↑2.51
", + "image_path": "a896ff6d025599d29191620f9ea697c8c300214195da8fc8bd6be734f949be78.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 369, + 168, + 498, + 179 + ], + "lines": [ + { + "bbox": [ + 369, + 168, + 498, + 179 + ], + "spans": [ + { + "bbox": [ + 369, + 168, + 498, + 179 + ], + "type": "text", + "content": "Table 5. Ablations of key modules." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "text", + "content": "vious works [7, 16], also report accuracy across three class splits: Many " + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "inline_equation", + "content": "(n_c \\geq 100)" + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "text", + "content": ", Medium " + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "inline_equation", + "content": "(20 < n_c < 100)" + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "text", + "content": ", and Few " + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "inline_equation", + "content": "(n_c \\leq 20)" + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "text", + "content": ". Overall, SEL outperforms MixUp on both datasets, and consistently enhances the compared methods, achieving an average improvement of " + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "inline_equation", + "content": "1.25\\%" + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "text", + "content": " (max " + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "inline_equation", + "content": "3.06\\%" + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "text", + "content": ", min " + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "inline_equation", + "content": "0.61\\%" + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "text", + "content": ") on ImageNet-LT and " + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "inline_equation", + "content": "1.53\\%" + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "text", + "content": " (max " + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "inline_equation", + "content": "3.15\\%" + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "text", + "content": ", min " + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "inline_equation", + "content": "0.52\\%" + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "text", + "content": ") on iNaturalist 2018. As seen, training with SEL typically improves the performance of Few classes while slightly reducing that of Many classes, aligning with the rationale proved in Sec. 3.4. However, the greater gain in Few classes outweighs the slight decline in Many classes, leading to an overall improvement in the All classes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 353, + 416, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 416, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 416, + 365 + ], + "type": "text", + "content": "4.4. Further Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 370, + 554, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 554, + 407 + ], + "type": "text", + "content": "To further analyze the principles of SEL quantitatively, we conducted a series of ablation studies on the CIFAR-10/100-LT datasets using ResNet-32 [9] as the base model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 413, + 484, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 484, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 484, + 426 + ], + "type": "text", + "content": "4.4.1. Visualization of Decision Regions" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 428, + 555, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 428, + 555, + 572 + ], + "spans": [ + { + "bbox": [ + 313, + 428, + 555, + 572 + ], + "type": "text", + "content": "Fig. 4 visualizes the decision regions and boundaries of different classes using t-SNE [35] across various methods. These models are trained on the CIFAR-100-LT with an imbalance ratio of 100 and tested on a balanced test set. For comparison, we selected four classes from CIFAR-100 test set for visualization: where apple and aquarium fish are the head classes with the most training samples, and willow tree and wolf are the tail classes with the fewest training samples. The decision boundaries are derived using the SVM algorithm. From Fig. 4, we can observe that SEL results in more balanced decision regions for both head and tail classes, aligning with our motivation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 578, + 432, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 432, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 432, + 590 + ], + "type": "text", + "content": "4.4.2. Role of Key Modules" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "The CbC and FwS modules are the core components of SEL. To examine their roles, we conduct ablation experiments on CIFAR-100-LT with imbalance ratios of 100 and 10. Tab. 5 presented the experimental results, showing that the CbC module significantly enhances the baseline method, with improvements of " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "3.48\\%" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "(39.96\\% \\rightarrow 43.44\\%)" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "1.18\\%" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "(56.66\\% \\rightarrow 57.84\\%)" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " for imbalance ratios of 100 and 10, respectively. These results highlight the CbC module's crucial role in balancing data distribution, enabling the model to learn more effectively from tail classes. However," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1876" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 69, + 225, + 177 + ], + "blocks": [ + { + "bbox": [ + 63, + 69, + 225, + 177 + ], + "lines": [ + { + "bbox": [ + 63, + 69, + 225, + 177 + ], + "spans": [ + { + "bbox": [ + 63, + 69, + 225, + 177 + ], + "type": "image", + "image_path": "716691e4944b523ec5b595f6875a736a441d44aa5fe965579e1542e1fccd8a35.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 179, + 156, + 189 + ], + "lines": [ + { + "bbox": [ + 133, + 179, + 156, + 189 + ], + "spans": [ + { + "bbox": [ + 133, + 179, + 156, + 189 + ], + "type": "text", + "content": "(a) CE" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 212, + 191, + 399, + 202 + ], + "lines": [ + { + "bbox": [ + 212, + 191, + 399, + 202 + ], + "spans": [ + { + "bbox": [ + 212, + 191, + 399, + 202 + ], + "type": "text", + "content": "apple aquarium fish willow tree wolf" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 225, + 70, + 386, + 177 + ], + "blocks": [ + { + "bbox": [ + 225, + 70, + 386, + 177 + ], + "lines": [ + { + "bbox": [ + 225, + 70, + 386, + 177 + ], + "spans": [ + { + "bbox": [ + 225, + 70, + 386, + 177 + ], + "type": "image", + "image_path": "8eab3c4522ffdf4793ba371a8356e68ec5cba34e59fe4c0ccc9c3880429927fd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 287, + 179, + 323, + 190 + ], + "lines": [ + { + "bbox": [ + 287, + 179, + 323, + 190 + ], + "spans": [ + { + "bbox": [ + 287, + 179, + 323, + 190 + ], + "type": "text", + "content": "(b) MixUp" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 386, + 70, + 547, + 178 + ], + "blocks": [ + { + "bbox": [ + 386, + 70, + 547, + 178 + ], + "lines": [ + { + "bbox": [ + 386, + 70, + 547, + 178 + ], + "spans": [ + { + "bbox": [ + 386, + 70, + 547, + 178 + ], + "type": "image", + "image_path": "8d700e333e0d6188682ac844bd4b2f8788258555d41a092de5028b5a3bb9c4f6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 452, + 179, + 479, + 189 + ], + "lines": [ + { + "bbox": [ + 452, + 179, + 479, + 189 + ], + "spans": [ + { + "bbox": [ + 452, + 179, + 479, + 189 + ], + "type": "text", + "content": "(c) SEL" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 56, + 251, + 174, + 319 + ], + "blocks": [ + { + "bbox": [ + 55, + 211, + 555, + 234 + ], + "lines": [ + { + "bbox": [ + 55, + 211, + 555, + 234 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 555, + 234 + ], + "type": "text", + "content": "Figure 4. Visualization of decision regions and boundaries on the CIFAR-100 test set (each class contains 100 test samples), where apple and aquarium fish are head classes with the most training samples, and willow tree and wolf are tail classes with the fewest training samples." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 251, + 174, + 319 + ], + "lines": [ + { + "bbox": [ + 56, + 251, + 174, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 174, + 319 + ], + "type": "image", + "image_path": "1b4246f60e2705712b128da47d2aa8a22f9e635de2fec3633c1840ca93c4884d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 84, + 321, + 145, + 330 + ], + "lines": [ + { + "bbox": [ + 84, + 321, + 145, + 330 + ], + "spans": [ + { + "bbox": [ + 84, + 321, + 145, + 330 + ], + "type": "text", + "content": "(a) CIFAR-100-LT" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 178, + 253, + 296, + 319 + ], + "blocks": [ + { + "bbox": [ + 178, + 253, + 296, + 319 + ], + "lines": [ + { + "bbox": [ + 178, + 253, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 178, + 253, + 296, + 319 + ], + "type": "image", + "image_path": "4c2c44f2c6f7026de5a3a7d334556998300dedc4211bfc5d1bd98e4bcfb48945.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 321, + 266, + 330 + ], + "lines": [ + { + "bbox": [ + 208, + 321, + 266, + 330 + ], + "spans": [ + { + "bbox": [ + 208, + 321, + 266, + 330 + ], + "type": "text", + "content": "(b) CIFAR-10-LT" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 56, + 365, + 176, + 435 + ], + "blocks": [ + { + "bbox": [ + 64, + 341, + 285, + 352 + ], + "lines": [ + { + "bbox": [ + 64, + 341, + 285, + 352 + ], + "spans": [ + { + "bbox": [ + 64, + 341, + 285, + 352 + ], + "type": "text", + "content": "Figure 5. Changes with different neighboring class numbers." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 365, + 176, + 435 + ], + "lines": [ + { + "bbox": [ + 56, + 365, + 176, + 435 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 176, + 435 + ], + "type": "image", + "image_path": "9f4076fa1f5e32e733a4046011f0c2730e0b25bc3987b2e6576f6ba263458504.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 437, + 141, + 446 + ], + "lines": [ + { + "bbox": [ + 91, + 437, + 141, + 446 + ], + "spans": [ + { + "bbox": [ + 91, + 437, + 141, + 446 + ], + "type": "text", + "content": "(a) start epochs" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 93, + 456, + 257, + 468 + ], + "lines": [ + { + "bbox": [ + 93, + 456, + 257, + 468 + ], + "spans": [ + { + "bbox": [ + 93, + 456, + 257, + 468 + ], + "type": "text", + "content": "Figure 6. Changes with different parameters." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 179, + 365, + 296, + 435 + ], + "blocks": [ + { + "bbox": [ + 179, + 365, + 296, + 435 + ], + "lines": [ + { + "bbox": [ + 179, + 365, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 179, + 365, + 296, + 435 + ], + "type": "image", + "image_path": "ba8e8d79274170b70c7a7f9f52d3b7bb01bb264004f60b7fee9f941e8db10d9f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 437, + 260, + 446 + ], + "lines": [ + { + "bbox": [ + 214, + 437, + 260, + 446 + ], + "spans": [ + { + "bbox": [ + 214, + 437, + 260, + 446 + ], + "type": "text", + "content": "(b) weight " + }, + { + "bbox": [ + 214, + 437, + 260, + 446 + ], + "type": "inline_equation", + "content": "(\\lambda)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 493, + 296, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 296, + 553 + ], + "type": "text", + "content": "the FwS module has a minimal impact without the CbC module, yielding only " + }, + { + "bbox": [ + 55, + 493, + 296, + 553 + ], + "type": "inline_equation", + "content": "0.30\\%" + }, + { + "bbox": [ + 55, + 493, + 296, + 553 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 493, + 296, + 553 + ], + "type": "inline_equation", + "content": "0.27\\%" + }, + { + "bbox": [ + 55, + 493, + 296, + 553 + ], + "type": "text", + "content": " improvements for imbalance ratios of 100 and 10, respectively. Its effectiveness is fully realized when combined with the CbC module, leading to a significant boost in model performance." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 565, + 216, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 565, + 216, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 565, + 216, + 577 + ], + "type": "text", + "content": "4.4.3. Impact of different parameters" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "In this section, we conduct exploratory experiments on the impact of the number of neighboring classes " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "(k_c^{\\prime})" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " in the FwS module. Experimental results are summarized in Fig. 5, where " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "k_{c}^{\\prime}" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " is varied from 1 to 10 and relative accuracy (values are adjusted by subtracting their mean) is utilized for comparison. As seen, the model performs better with a higher " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "k^{\\prime}_{c}" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " on CIFAR-100-LT but with a lower " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "k^{\\prime}_{c}" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " on CIFAR-10LT. This difference arises from the number of classes, as CIFAR-100-LT has 100 classes, whereas CIFAR-10-LT has only 10. Based on these observations, we empirically set " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "k^{\\prime}_{c}" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " to 5 for CIFAR-100-LT and 2 for CIFAR-10-LT." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 254, + 464, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 254, + 464, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 464, + 266 + ], + "type": "text", + "content": "4.4.4. When to Start SEL Training" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 270, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 449 + ], + "type": "text", + "content": "In this section, we explore the optimal starting point for SEL training to maximize computational efficiency, as its synthesis of additional exploratory examples increases computational costs. Specifically, we integrate SEL into the training process at different epochs, ranging from 100 to 200, as shown in Fig. 6(a). Here, starting at 200 epochs indicates a two-stage training mode, where SEL is applied with an additional 20 epochs after the original training process. Similarly, relative accuracy is used as the ordinate for clearly showing the parameter's impact. It is obvious that the model's performance steadily improves as the SEL starting point increases, with most datasets peaking at 180 epochs, except for CIFAR-10-LT-100 and CIFAR-10-LT-10. Therefore, we set the starting point to 180 epochs for CIFAR-100-LT and 200 epochs for CIFAR-10-LT." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 451, + 556, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 556, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 556, + 559 + ], + "type": "text", + "content": "Building on the optimal starting points, we further examine the impact of parameter " + }, + { + "bbox": [ + 313, + 451, + 556, + 559 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 451, + 556, + 559 + ], + "type": "text", + "content": " in Eq. (6). Fig. 6(b) illustrates the performance variation with " + }, + { + "bbox": [ + 313, + 451, + 556, + 559 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 451, + 556, + 559 + ], + "type": "text", + "content": " ranges from 0.0 to 1.0. Note that " + }, + { + "bbox": [ + 313, + 451, + 556, + 559 + ], + "type": "inline_equation", + "content": "\\lambda = 0.0" + }, + { + "bbox": [ + 313, + 451, + 556, + 559 + ], + "type": "text", + "content": " indicates the absence of SEL, while " + }, + { + "bbox": [ + 313, + 451, + 556, + 559 + ], + "type": "inline_equation", + "content": "\\lambda = 1.0" + }, + { + "bbox": [ + 313, + 451, + 556, + 559 + ], + "type": "text", + "content": " signifies its full integration. For most datasets, as " + }, + { + "bbox": [ + 313, + 451, + 556, + 559 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 451, + 556, + 559 + ], + "type": "text", + "content": " increases, the model's performance gradually improves, peaking at " + }, + { + "bbox": [ + 313, + 451, + 556, + 559 + ], + "type": "inline_equation", + "content": "\\lambda = 0.9" + }, + { + "bbox": [ + 313, + 451, + 556, + 559 + ], + "type": "text", + "content": " before declining. However, for CIFAR-10-LT-10, the simplest dataset, performance remains relatively stable, with a peak at " + }, + { + "bbox": [ + 313, + 451, + 556, + 559 + ], + "type": "inline_equation", + "content": "\\lambda = 0.6" + }, + { + "bbox": [ + 313, + 451, + 556, + 559 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 572, + 388, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 572, + 388, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 572, + 388, + 585 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we propose SEL, a simple yet effective training framework that addresses long-tailed challenges through the space exploration perspective. SEL is a plug-and-play method that boosts performance without modifying the model or adding parameters. It uses an adaptive OFA operator to generate exploratory examples that capture unseen tail-class features, enhancing decision regions and boundaries. The effectiveness of SEL is validated mathematically through rational analysis and empirically through experiments on various long-tailed datasets and methods." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1877" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 294, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 294, + 163 + ], + "type": "text", + "content": "This work is supported by the Solfeggio ear training intelligent robot and cloud platform research and development project for music education (No.2024CXY0102), the 3D visualization digital twin integrated control system (No.2023CXY0111), and the public technology service platform project of Xiamen City (No.3502Z20231043)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 195, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 195, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 195, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 61, + 195, + 295, + 228 + ], + "type": "text", + "content": "[1] Yoshua Bengio, Yann Lecun, and Geoffrey Hinton. Deep learning for ai. Communications of the ACM, 64(7):58-65, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 229, + 296, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 229, + 296, + 284 + ], + "spans": [ + { + "bbox": [ + 61, + 229, + 296, + 284 + ], + "type": "text", + "content": "[2] Ekin D. Cubuk, Barret Zoph, Dandelion Mané, Vijay Vasudevan, and Quoc V. Le. Autoaugment: Learning augmentation strategies from data. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 113-123, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 285, + 295, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 285, + 295, + 340 + ], + "spans": [ + { + "bbox": [ + 61, + 285, + 295, + 340 + ], + "type": "text", + "content": "[3] Ekin Dogus Cubuk, Barret Zoph, Jon Shlens, and Quoc Le. Randaugment: Practical automated data augmentation with a reduced search space. In Advances in Neural Information Processing Systems, pages 18613-18624. Curran Associates, Inc., 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 342, + 295, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 295, + 385 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 295, + 385 + ], + "type": "text", + "content": "[4] Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, and Jiaya Jia. Parametric contrastive learning. In 2021 IEEE/CVF International Conference on Computer Vision, pages 695-704, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 387, + 296, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 387, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 62, + 387, + 296, + 431 + ], + "type": "text", + "content": "[5] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge J. Belongie. Class-balanced loss based on effective number of samples. In IEEE Conference on Computer Vision and Pattern Recognition, pages 9268-9277, 2019. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 433, + 295, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 433, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 62, + 433, + 295, + 486 + ], + "type": "text", + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 488, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 488, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 488, + 295, + 544 + ], + "type": "text", + "content": "[7] Fei Du, Peng Yang, Qi Jia, Fengtao Nan, Xiaoting Chen, and Yun Yang. Global and local mixture consistency cumulative learning for long-tailed visual recognitions. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15814-15823, 2023. 3, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 544, + 296, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 544, + 296, + 589 + ], + "spans": [ + { + "bbox": [ + 62, + 544, + 296, + 589 + ], + "type": "text", + "content": "[8] Hao Guo and Song Wang. Long-tailed multi-label visual recognition by collaborative training on uniform and rebalanced samplings. In IEEE Conference on Computer Vision and Pattern Recognition, pages 15089-15098, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 590, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 590, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 62, + 590, + 295, + 634 + ], + "type": "text", + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "text", + "content": "[10] Daniel Ho, Eric Liang, Xi Chen, Ion Stoica, and Pieter Abbeel. Population based augmentation: Efficient learning of augmentation policy schedules. In Proceedings of the 36th International Conference on Machine Learning, pages 2731-2741, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "text", + "content": "[11] Gregory Holste, Yiliang Zhou, Song Wang, Ajay Jaiswal, Mingquan Lin, Sherry Zhuge, Yuzhe Yang, Dongkyun Kim," + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 160 + ], + "type": "text", + "content": "Trong-Hieu Nguyen-Mau, Minh-Triet Tran, Jaehyup Jeong, Wongi Park, Jongbin Ryu, Feng Hong, Arsh Verma, Yosuke Yamagishi, Changhyun Kim, Hyeryeong Seo, Myungjoo Kang, Leo Anthony Celi, Zhiyong Lu, Ronald M. Summers, George Shih, Zhangyang Wang, and Yifan Peng. Towards long-tailed, multi-label disease classification from chest x-ray: Overview of the cxr-It challenge. Medical Image Analysis, 97:103224, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 228 + ], + "type": "text", + "content": "[12] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alexander Shepard, Hartwig Adam, Pietro Perona, and Serge J. Belongie. The inaturalist species classification and detection dataset. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, pages 8769-8778, 2018. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 229, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 229, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 229, + 553, + 274 + ], + "type": "text", + "content": "[13] Chengkai Hou, Jieyu Zhang, Haonan Wang, and Tianyi Zhou. Subclass-balancing contrastive learning for long-tailed recognition. In 2023 IEEE/CVF International Conference on Computer Vision, pages 5372-5384, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 274, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 553, + 319 + ], + "type": "text", + "content": "[14] Shuyan Hu, Wei Ni, Xin Wang, and Abbas Jamalipour. Disguised tailing and video surveillance with solar-powered fixed-wing unmanned aerial vehicle. IEEE Transactions on Vehicular Technology, 71(5):5507-5518, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 319, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 352 + ], + "type": "text", + "content": "[15] ZhongQuan Jian and GuangYu Zhu. Affine invariance of meta-heuristic algorithms. Information Sciences, 576:37-53, 2021. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 354, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 553, + 407 + ], + "type": "text", + "content": "[16] Yan Jin, Mengke Li, Yang Lu, Yiu-ming Cheung, and Hanzi Wang. Long-tailed visual recognition via self-heterogeneous integration with knowledge excavation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23695-23704, 2023. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 409, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 409, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 317, + 409, + 553, + 464 + ], + "type": "text", + "content": "[17] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. In 8th International Conference on Learning Representations, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 465, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 465, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 465, + 553, + 510 + ], + "type": "text", + "content": "[18] Bingyi Kang, Yu Li, Sa Xie, Zehuan Yuan, and Jiashi Feng. Exploring balanced feature spaces for representation learning. In 9th International Conference on Learning Representations, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 510, + 553, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 510, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 317, + 510, + 553, + 564 + ], + "type": "text", + "content": "[19] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673, 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 567, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 589 + ], + "type": "text", + "content": "[20] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 590, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 634 + ], + "type": "text", + "content": "[21] Jun Li, Zichang Tan, Jun Wan, Zhen Lei, and Guodong Guo. Nested collaborative learning for long-tailed visual recognition. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6939-6948, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "text", + "content": "[22] Jian Li, Ziyao Meng, Daqian Shi, Rui Song, Xiaolei Dao, Jingwen Wang, and Hao Xu. FCC: feature clusters compression for long-tailed visual recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24080-24089, 2023. 1, 3, 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[23] Mengke Li, Yiu-Ming Cheung, and Yang Lu. Long-tailed visual recognition via gaussian clouded logit adjustment. In" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1878" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "text", + "content": "2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6919-6928, 2022. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 139 + ], + "type": "text", + "content": "[24] Mengke Li, Yiu-Ming Cheung, and Zhikai Hu. Key point sensitive loss for long-tailed visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4): 4812-4825, 2023. 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 295, + 217 + ], + "type": "text", + "content": "[25] Mengke Li, Zhikai Hu, Yang Lu, Weichao Lan, Yiu-ming Cheung, and Hui Huang. Feature fusion from head to tail for long-tailed visual recognition. In Thirty-Eighth AAAI Conference on Artificial Intelligence, Thirty-Sixth Conference on Innovative Applications of Artificial Intelligence, Fourteenth Symposium on Educational Advances in Artificial Intelligence, pages 13581-13589, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 217, + 294, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 217, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 57, + 217, + 294, + 271 + ], + "type": "text", + "content": "[26] Shuang Li, Kaixiong Gong, Chi Harold Liu, Yulin Wang, Feng Qiao, and Xinjing Cheng. Metasaug: Meta semantic augmentation for long-tailed visual recognition. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5212-5221, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 295, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 295, + 326 + ], + "type": "text", + "content": "[27] Shuang Li, Kaixiong Gong, Chi Harold Liu, Yulin Wang, Feng Qiao, and Xinjing Cheng. Metasaug: Meta semantic augmentation for long-tailed visual recognition. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5212-5221, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 295, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 295, + 382 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 295, + 382 + ], + "type": "text", + "content": "[28] Tianhong Li, Peng Cao, Yuan Yuan, Lijie Fan, Yuzhe Yang, Rogerio Feris, Piotr Indyk, and Dina Katabi. Targeted supervised contrastive learning for long-tailed recognition. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6918, 2022. 1, 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 383, + 295, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 295, + 426 + ], + "type": "text", + "content": "[29] Bo Liu, Haoxiang Li, Hao Kang, Gang Hua, and Nuno Vasconcelos. Gistnet: a geometric structure transfer network for long-tailed recognition. In 2021 IEEE/CVF International Conference on Computer Vision, pages 8189-8198, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 295, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 295, + 480 + ], + "type": "text", + "content": "[30] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X. Yu. Large-scale long-tailed recognition in an open world. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2537-2546, 2019. 1, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 481, + 295, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 295, + 525 + ], + "type": "text", + "content": "[31] Seulki Park, Jongin Lim, Younghan Jeon, and Jin Young Choi. Influence-balanced loss for imbalanced visual classification. In 2021 IEEE/CVF International Conference on Computer Vision, pages 715-724, 2021. 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 526, + 295, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 295, + 581 + ], + "type": "text", + "content": "[32] Seulki Park, Youngkyu Hong, Byeongho Heo, Sangdoo Yun, and Jin Young Choi. The majority can help the minority: Context-rich minority oversampling for long-tailed classification. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6877-6886, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 582, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 295, + 624 + ], + "type": "text", + "content": "[33] Jiawei Ren, Cunjun Yu, Shunan Sheng, Xiao Ma, Haiyu Zhao, Shuai Yi, and Hongsheng Li. Balanced meta-softmax for long-tailed visual recognition. In Advances in Neural Information Processing Systems 33, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 625, + 295, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 295, + 680 + ], + "type": "text", + "content": "[34] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott E. Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In 2015 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-9, 2015. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "text", + "content": "[35] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[36] Vikas Verma, Alex Lamb, Christopher Beckham, Amir Najafi, Ioannis Mitlagkas, David Lopez-Paz, and Yoshua Bengio. Manifold mixup: Better representations by interpolating hidden states. In Proceedings of the 36th International Conference on Machine Learning, pages 6438-6447, 2019. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 183 + ], + "type": "text", + "content": "[37] Jianfeng Wang, Thomas Lukasiewicz, Xiaolin Hu, Jianfei Cai, and Zhenghua Xu. Rsg: A simple but effective module for learning imbalanced datasets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3784-3793, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 184, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 237 + ], + "type": "text", + "content": "[38] Peng Wang, Kai Han, Xiu-Shen Wei, Lei Zhang, and Lei Wang. Contrastive learning based hybrid networks for long-tailed image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 943-952, 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 238, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 238, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 238, + 553, + 282 + ], + "type": "text", + "content": "[39] Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu, and Stella X. Yu. Long-tailed recognition by routing diverse distribution-aware experts. In 9th International Conference on Learning Representations, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 283, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 327 + ], + "type": "text", + "content": "[40] Tong Wu, Ziwei Liu, Qingqiu Huang, Yu Wang, and Dahua Lin. Adversarial robustness under long-tailed distribution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 8659-8668, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 327, + 553, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 393 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 393 + ], + "type": "text", + "content": "[41] Shiyu Xuan and Shiliang Zhang. Decoupled contrastive learning for long-tailed recognition. In Thirty-Eighth AAAI Conference on Artificial Intelligence, Thirty-Sixth Conference on Innovative Applications of Artificial Intelligence, Fourteenth Symposium on Educational Advances in Artificial Intelligence, pages 6396-6403, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 394, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 449 + ], + "type": "text", + "content": "[42] Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Seong Joon Oh, Youngjoon Yoo, and Junsuk Choe. Cutmix: Regularization strategy to train strong classifiers with localizable features. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6022-6031, 2019. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "type": "text", + "content": "[43] Shiran Zada, Itay Benou, and Michal Irani. Pure noise to the rescue of insufficient data: Improving imbalanced classification by training on random noise images. In International Conference on Machine Learning, pages 25817-25833, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 503, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 503, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 503, + 553, + 557 + ], + "type": "text", + "content": "[44] Yuhang Zang, Chen Huang, and Chen Change Loy. Fasa: Feature augmentation and sampling adaptation for long-tailed instance segmentation. In 2021 IEEE/CVF International Conference on Computer Vision, pages 3437-3446, 2021. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 558, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 558, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 558, + 553, + 602 + ], + "type": "text", + "content": "[45] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In 6th International Conference on Learning Representations, 2018. 2, 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 603, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 646 + ], + "type": "text", + "content": "[46] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):10795-10816, 2023. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "text", + "content": "[47] Zizhao Zhang and Tomas Pfister. Learning fast sample reweighting without reward data. In 2021 IEEE/CVF International Conference on Computer Vision, pages 705-714, 2021. 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[48] Boyan Zhou, Quan Cui, Xiu-Shen Wei, and Zhao-Min Chen. BBN: bilateral-branch network with cumulative learning for" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1879" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 307 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "type": "text", + "content": "long-tailed visual recognition. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9716-9725, 2020. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "text", + "content": "[49] Weitao Zhou, Zhong Cao, Nanshan Deng, Xiaoyu Liu, Kun Jiang, and Diange Yang. Dynamically conservative self-driving planner for long-tail cases. IEEE Transactions on Intelligent Transportation Systems, 24(3):3476-3488, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 217 + ], + "type": "text", + "content": "[50] Chenguang Zhu, Yichong Xu, Xiang Ren, Bill Yuchen Lin, Meng Jiang, and Wenhao Yu. Knowledge-augmented methods for natural language processing. In Proceedings of the Sixteenth ACM International Conference on Web Search and Data Mining, pages 1228-1231, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 251 + ], + "type": "text", + "content": "[51] Guang-Yu Zhu and Wei-Bo Zhang. Optimal foraging algorithm for global optimization. Applied Soft Computing, 51: 294–313, 2017. 2, 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 295, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 295, + 307 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 295, + 307 + ], + "type": "text", + "content": "[52] Jianggang Zhu, Zheng Wang, Jingjing Chen, Yi-Ping Phoebe Chen, and Yu-Gang Jiang. Balanced contrastive learning for long-tailed visual recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6898-6907, 2022. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1880" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file